Communications and Control Engineering

Series Editors
E.D. Sontag • M. Thoma • A. Isidori • J.H. van Schuppen

Published titles include:

Stability and Stabilization of Infinite Dimensional Randomized Algorithms for Analysis and Control
Systems with Applications of Uncertain Systems
Zheng-Hua Luo, Bao-Zhu Guo and Omer Morgul Roberto Tempo, Giuseppe Calafiore and Fabrizio
Dabbene
Nonsmooth Mechanics (Second edition)
Bernard Brogliato Switched Linear Systems
Zhendong Sun and Shuzhi S. Ge
Nonlinear Control Systems II
Alberto Isidori Subspace Methods for System Identification
Tohru Katayama
L2-Gain and Passivity Techniques in Nonlinear Control
Arjan van der Schaft Digital Control Systems
Ioan D. Landau and Gianluca Zito
Control of Linear Systems with Regulation and Input
Constraints Multivariable Computer-controlled Systems
Ali Saberi, Anton A. Stoorvogel and Peddapullaiah Efim N. Rosenwasser and Bernhard P. Lampe
Sannuti
Dissipative Systems Analysis and Control
Robust and H∞ Control (Second edition)
Ben M. Chen Bernard Brogliato, Rogelio Lozano, Bernhard Maschke
and Olav Egeland
Computer Controlled Systems
Efim N. Rosenwasser and Bernhard P. Lampe Algebraic Methods for Nonlinear Control Systems
Giuseppe Conte, Claude H. Moog and Anna M. Perdon
Control of Complex and Uncertain Systems
Stanislav V. Emelyanov and Sergey K. Korovin Polynomial and Rational Matrices
Tadeusz Kaczorek
Robust Control Design Using H∞ Methods
Simulation-based Algorithms for Markov Decision
Ian R. Petersen, Valery A. Ugrinovski and Andrey V.
Processes
Savkin
Hyeong Soo Chang, Michael C. Fu, Jiaqiao Hu
Model Reduction for Control System Design and Steven I. Marcus
Goro Obinata and Brian D.O. Anderson Iterative Learning Control
Control Theory for Linear Systems Hyo-Sung Ahn, Kevin L. Moore and YangQuan Chen
Harry L. Trentelman, Anton Stoorvogel and Malo Hautus Distributed Consensus in Multi-vehicle Cooperative
Functional Adaptive Control Control
Simon G. Fabri and Visakan Kadirkamanathan Wei Ren and Randal W. Beard

Control of Singular Systems with Random Abrupt

Positive 1D and 2D Systems
Changes
Tadeusz Kaczorek
El-Kébir Boukas
Identification and Control Using Volterra Models
Nonlinear and Adaptive Control with Applications
Francis J. Doyle III, Ronald K. Pearson
Alessandro Astolfi, Dimitrios Karagiannis and Romeo
and Babatunde A. Ogunnaike
Ortega
Non-linear Control for Underactuated Mechanical
Stabilization, Optimal and Robust Control
Systems
Aziz Belmiloudi
Isabelle Fantoni and Rogelio Lozano
Control of Nonlinear Dynamical Systems
Robust Control (Second edition)
Felix L. Chernous’ko, Igor M. Ananievski and Sergey
Jürgen Ackermann
A. Reshmin
Flow Control by Feedback
Periodic Systems
Ole Morten Aamo and Miroslav Krstić
Sergio Bittanti and Patrizio Colaneri
Learning and Generalization (Second edition)
Mathukumalli Vidyasagar Discontinuous Systems
Yury V. Orlov
Constrained Control and Estimation
Graham C. Goodwin, María M. Seron and José A. De
Doná
Huaguang Zhang • Derong Liu • Zhiliang Wang

Controlling Chaos
Suppression, Synchronization and
Chaotification

123
Huaguang Zhang, PhD Derong Liu, PhD
Zhiliang Wang, PhD Chinese Academy of Sciences
Institute of Automation
Northeastern University Beijing 100190
College of Information Science PR China
and Engineering and
Shenyang 110004 University of Illinois at Chicago
PR China Department of Electrical
[email protected] and Computer Engineering
Chicago
Illinois 60607
USA
[email protected]

ISBN 978-1-84882-522-2 e-ISBN 978-1-84882-523-9

DOI 10.1007/978-1-84882-523-9
Springer Dordrecht Heidelberg London New York

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

Library of Congress Control Number: 2009930031


c Springer-Verlag London Limited 2009

MATLAB® and Simulink® are registered trademarks of The MathWorks, Inc., 3 Apple Hill Drive,
Natick, MA 01760-2098, U.S.A., www.mathworks.com

Apart from any fair dealing for the purposes of research or private study, or criticism or review, as per-
mitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced,
stored or transmitted, in any form or by any means, with the prior permission in writing of the publis-
hers, or in the case of reprographic reproduction in accordance with the terms of licenses issued by the
Copyright Licensing Agency. Enquiries concerning reproduction outside those terms should be sent to
the publishers.
The use of registered names, trademarks, etc., in this publication does not imply, even in the absence of a
specific statement, that such names are exempt from the relevant laws and regulations and therefore free
for general use.
The publisher makes no representation, express or implied, with regard to the accuracy of the information
contained in this book and cannot accept any legal responsibility or liability for any errors or omissions
that may be made.

Cover design: le-tex publishing services GmbH, Leipzig, Germany

Printed on acid-free paper

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

 To
Leon O. Chua
The father of nonlinear circuit theory and cellular neural networks
The inventor of a five-element electronic circuit for generating
chaotic signals – the Chua circuit
Foreword

Today, chaos is no longer a stranger to many. Instead, chaos theory and technology
have gradually become well known as a promising research field with significant
impacts on an increasing number of novel, potentially attractive, time- and energy-
critical engineering applications.
There are sufficient scientific evidence and practical reasons for studying and uti-
lizing chaos control and synchronization. In a system where irregular responses are
undesirable or even harmful, chaos should be reduced as much as possible or even
completely suppressed, while if complex dynamics is beneficial and useful, chaos
synchronization and generation become desirable. The scope of chaos control for
applications, as it has turned out, covers such widespread areas as secure wired and
wireless communications and encryption of information data (e.g., texts, images,
and videos), biological systems (e.g., understanding of the functioning of human
brain and heart, so as to suppress deadly epilepsy and cardiac arrhythmias), liquid
mixing (e.g., chemical reactions and medical drug production), crisis management
(e.g., critical prevention of cascading fatal voltage collapses of power grids), high-
performance circuits and devices (e.g., cellular neural networks, multi-coupled sig-
nal modulators, and power converters), and artificial intelligence in decision making
and organization of complex multi-agent networks and systems in industries, eco-
nomics, as well as military alike.
Chaos control, even by linguistic definition, involves both basic chaos theory and
automatic control techniques. More than ten years ago, we pointed out that1 “We
subscribe to the idea that any viable theory should be one that is constantly evolving
and developing. When merged, the twin virtues of classical control theory and ad-
vanced nonlinear science may bring about unexpected benefits. For chaotic systems,
if we can incorporate control mechanisms that exploit certain defining characteris-
tics of chaos, the approach becomes more interesting and, hopefully, more efficient
and effective. In addition, due to the inherent association of nonlinearity and chaos
to various issues, the scope of chaos control is much more diverse. When compared

1 Preface, G. Chen and X. Dong, From Chaos to Order: Methodologies, Perspectives and Applica-

tions. World Scientific, Singapore, 1998.

vii
viii Foreword

to other conventional approaches, chaos control has its unique features with regard
to such aspects as objectives, perspectives, problem formulations, and performance
measures.” Today, we are especially pleased to welcome the present monograph
Controlling Chaos that continues to address the important subject of chaos control,
including in particular the key topics of chaos suppression, synchronization, and
generation (also named chaotification).
This exciting and yet challenging research and development area has continu-
ously been a scientific interdiscipline involving systems and control engineers, the-
oretical and experimental physicists, computational and applied mathematicians, bi-
ologists and physiologists, and electronics specialists, among others. The progress
has been very promising and encouraging to date. Nevertheless, achievements
notwithstanding,2 “Chaos control calls for new efforts and endeavors for the coming
millennium. In this new era, perhaps today’s contemporary concepts of nonlinear
dynamics and controls will undergo yet another cycle of rethinking and reorganiz-
ing; perhaps the chaos control theories, methodologies, and perspectives that have
been drawn together in this book will unleash some other new ideas and valuable
applications; perhaps real breakthroughs will begin to take place, bringing enhance-
ment, improvement, and sustainability to the complex living nature – it is only the
beginning ... .”

Hong Kong Guanrong Chen

February 2009

2Epilogue, G. Chen and X. Dong, From Chaos to Order: Methodologies, Perspectives and Appli-
cations. World Scientific, Singapore, 1998.
Preface

Background of This Book

Chaos theory, once considered to be the third revolution in physics following rela-
tivity theory and quantum mechanics, has been studied extensively in the past thirty
years. A lot of chaotic phenomena have been found and enormous mathematical
strides have been taken. Nowadays, it has been agreed by scientists and engineers
that chaos is ubiquitous in natural sciences and social sciences, such as in physics,
chemistry, mathematics, biology, ecology, physiology, economics, and so on. Wher-
ever nonlinearity exists, chaos may be found. For a long time, chaos was thought of
as a harmful behavior that could decrease the performance of a system and there-
fore should be avoided when the system is running. One remarkable feature of a
chaotic system distinguishing itself from other nonchaotic systems is that the sys-
tem is extremely sensitive to initial conditions. Any tiny perturbation of the initial
conditions will significantly alter the long-term dynamics of the system. This fact
means that when one wants to control a chaotic system one must make sure that the
measurement of the needed signals is absolutely precise. Otherwise any attempt of
controlling chaos would make the dynamics of the system go to an unexpected state.
With the development of chaos theory and practice in engineering, more and more
people want to know the answers to the following questions:
(1) Can chaos be controlled?
(2) Can chaos be utilized?
(3) Can two chaotic systems be in resonance as in the case of periodic ones?
(4) If the answer to the second question is positive, then how to generate chaos in a
nonchaotic system?
These questions have been partly answered by Ott, Pecora, and Chen in the
1990s, which has led a surge in the application study of chaos. From then on, a
new research area, chaos control, including suppression, utilization, and generation
of chaotic phenomena, came into being. Among these studies, three aspects attract

ix
x Preface

more attention; that is, stabilization of chaos, synchronization of chaos, and anti-
control of chaos.3

Why This Book?

Although several monographs on controlling chaos have been published, the present
book has unique features which distinguish it from others.
First, the types of chaotic systems studied in this monograph are rather extensive.
From the point of view of physics, readers can find not only well-known chaotic
systems, such as the Lorenz system, the Rössler system, and the Hénon map, but
also some new chaotic systems which appeared in recent years, such as the Liu
hyperchaotic system, the Liao chaotic system, the Chen chaotic system, and the Lü
chaotic system. From the point of view of models, one can find difference equations,
ordinary differential equations, and time-delayed differential equations in this book,
which are the main mathematical models describing chaos.
Second, since the monograph is a summary of the authors’ previous research,
the methods proposed here for stabilizing, synchronizing, and generating chaos in a
great degree benefit from the theory of nonlinear control systems, and are more ad-
vanced than that appear in other introductory books. One example is that in order to
stabilize a chaotic system to one of its equilibria, an inverse optimal control method
is developed in this book. The controller designed according to this method not only
stabilizes the system but also optimizes a meaningful cost functional. Therefore, the
difficulty of solving the Hamilton–Jacobi–Bellman (HJB) equation is avoided. An-
other example is that in order to synchronize two discrete-time chaotic systems, the
exact linearization method is used which provides a unified framework for controller
design for both continuous-time and discrete-time chaotic systems. Yet a third ex-
ample is that in order to chaotify a continuous-time nonchaotic system, a kind of
impulsive control method is developed. A mathematical proof shows that the chaos
induced by this method satisfies Devaney’s definition of chaos.
Last but not least, some rather unique contributions are included in this mono-
graph. One notable feature is the combination of fuzzy logic and chaos. Besides
the famous Takagi–Sugeno (T–S) fuzzy model, a novel model, the fuzzy hyperbolic
model (FHM), which was initially proposed by one of the authors and whose merits
in modeling and control have been illustrated in our book4 earlier, is also included
in this book. In this monograph we combine chaos and fuzzy logic in many aspects:
the T–S fuzzy model is used in Chaps. 4 and 8 for suppressing, modeling, and syn-
chronizing chaotic systems, respectively; and the chaotification of the discrete-time
FHM and the continuous-time FHM is studied in Chap. 9. Another notable feature
is that the methods proposed in this monograph can be applied to a wide class of
chaotic systems rather than a specific chaotic system. For example, in Chap. 4 a
systematic method is proposed for stabilizing discrete-time chaotic systems and in

3 Anticontrol of chaos is also known as chaotification. In this book, they are synonymous.
4 H. Zhang and D. Liu, Fuzzy Modeling and Fuzzy Control. Birkhäuser, Boston, 2006.
Preface xi

Chap. 5 a method based on nonlinear geometric control theory is proposed which

provides a systematic procedure to synchronize two or more chaotic systems.

The Content of This Book

The whole book involves nine chapters. As indicated by the title of the book, the
main content of the book is composed of three parts: suppressing chaos, synchro-
nizing chaos, and generating chaos. To make the book self contained, additional
materials are added to provide readers with a brief review of the history of chaos
control and some necessary mathematical preliminaries on dynamical systems.
In Chap. 1, we briefly review the history of chaos theory and chaos control.
We first review the history of chaos by following the important events in the de-
velopment of chaos theory. We start the review from the last decade of the 19th
century to the 1980s in the 20th century. The work of many distinguished scientists,
such as Poincaré, Birkhoff, van de Pol, Littlewood, Andoronov, Lorenz, Smale, Kol-
mogorov, Arnol’d, Feigenbaum, Li, Yorke, and May, is summarized. After that, we
review the development of chaos control from three different aspects, i.e., from the
points of view of suppression, synchronization, and chaotification. For each aspect,
not only the main methods are introduced but also the ideas behind those meth-
ods are mentioned. Some representative methods are introduced, such as the Ott–
Grebogi–Yorke (OGY) method and its extensions, the entrainment and migration
method, the time-delay feedback method, and some state feedback methods. Chaos
synchronization is introduced according to different synchronization patterns, such
as complete synchronization, phase synchronization, lag synchronization, and gen-
eralized synchronization. Chaotification was proposed by Chen in 1997 and has
attracted a lot of attention since then. Methods for chaotification will be reviewed,
including the state feedback method, the state delay feedback method, the impulsive
control method, and the Smale horseshoe method.
In Chap. 2, necessary mathematical background materials on nonlinear dynam-
ics and chaos are introduced. Dynamical system theory is a powerful tool for chaos
study. For the completeness of the book we provide a brief introduction to nonlin-
ear dynamical systems. This chapter is rather difficult for readers with engineering
background since many mathematical concepts, definitions, and theorems are in-
volved. The content of this chapter includes two parts. Some concepts and defini-
tions about nonlinear ordinary equations and dynamical systems are introduced first,
such as the concepts of flow, fixed point, equilibrium state, invariant set, attractor,
stable (unstable) manifold, Floquet index, Lyapunov exponents, and Smale horse-
shoe. We also state some important theorems, such as the theorem about existence
and uniqueness of solutions, the Hartman-Grobman theorems, and the Lyapunov
stability theorems. Some concepts and theorems about retarded functional differen-
tial equations (RFDEs) are introduced next, such as the definitions of solutions and
the initial problem, existence and uniqueness of solutions, and stability of solutions.
After that, some stability criteria for RFDEs are introduced.
xii Preface

In Chap. 3, the entrainment and migration control of chaos is introduced. It is usu-

ally the case that several attractors coexist in a complex dynamical system. These
attractors can be stable or unstable. When a key parameter is changed the attractor
may be altered either in appearance or in spatial position, or in both. One of the goals
of chaos control is to steer the system’s trajectory to the expected state. The back-
ground of entrainment and migration control is based on two facts: a multi-attractor
chaotic system is sensitive to both initial conditions and parameters and each sta-
ble attractor has its basin of attraction. After introducing the basics of entrainment
and migration control, a modified entrainment and migration control method: the
open-plus-closed-loop (OPCL) control method, is introduced. The OPCL method is
an effective method for polynomial chaotic systems when the order of the polyno-
mial is less than two. To overcome this restriction, based on the OPCL method, we
propose an improved method, the open-plus-nonlinear-closed-loop (OPNCL) con-
trol method. The OPNCL method can be applied to polynomial chaotic systems
with arbitrary order and greatly improves the control accuracy. Finally, the OP-
NCL method is employed for controlling continuous-time polynomial chaotic sys-
tems and discrete-time polynomial chaotic systems. By simulating the Chua circuit,
the logistic map, and the Hénon map, we validate the effectiveness of the OPNCL
method.
The parameters of a chaotic system play an important role, whose variation will
lead to completely different dynamics. Sometimes, we want to design a controller
which is optimal in a certain sense. However, it is a difficult task to solve the HJB
equation when one designs an optimal controller directly along the conventional
route. To solve the aforementioned problems, in Chap. 4, we focus on two kinds
of methods of suppressing chaos: the adaptive control method and the inverse op-
timal control method. We develop two new methods of parametric adaptive control
for a class of discrete-time chaotic systems and a class of continuous-time chaotic
systems with multiple parameters, respectively. The systems are assumed to be lin-
early parameterized in the adaptive control algorithm. Then, systems with nonlinear
distributed parameters and uncertain noise are considered. We apply the inverse op-
timal control method to stabilize a new four-dimensional chaotic system. The merit
of this approach is that it does not need to solve the complicated HJB equation.
In Chap. 5, the synchronization problems of both continuous-time systems and
discrete-time chaotic systems are studied. After introducing some necessary prelim-
inaries, a method is proposed to make the single output signal of the response system
synchronized with that of the drive system with a scalar controller. The method is
based on nonlinear geometric control theory and an exact linearization technique.
Then, this method is generalized to the case of multiple output signals. An adap-
tive method is also proposed to synchronize two different continuous-time chaotic
systems. Finally, for problems of synchronizing discrete-time chaotic systems with
parametric perturbations, an adaptive controller is designed. This method is based
on the theory of exact linearization of discrete-time systems. The methods devel-
oped in this chapter have a systematic procedure and can be used with a rather wide
class of chaotic systems.
Preface xiii

In Chap. 6, we study how to synchronize two identical or different chaotic sys-

tems by impulsive control methods. Impulsive control is an efficient method to deal
with dynamical systems which cannot be controlled by continuous control. Addi-
tionally, in the synchronization process, the response system receives information
from the drive system only at discrete time instants, which drastically reduces the
amount of synchronization information transmitted from the drive system to the re-
sponse system and makes this method more efficient in a great number of real-life
applications. We first study the complete synchronization of a class of chaotic sys-
tems and, after that, we develop synchronization methods for the unified systems
with channel time delay in the sense of practical stability. Then, robust synchroniza-
tion schemes are studied for the chaotic systems with parametric uncertainty and
parametric mismatch. Our goal is to develop practical impulsive control methods
for different synchronization schemes.
In engineering applications, time delays always exist, and parameters of the sys-
tem are inevitably perturbed by external disturbances. Moreover, the values of de-
lays and parameters are often unknown in advance. In some special cases, the struc-
ture or parameters of the drive system are even unknown in advance. In Chap. 7, we
study how to synchronize chaotic systems when time delay exists and the synchro-
nized systems have different structures. Synchronization methods are developed for
a class of delayed chaotic systems when the drive system and the response system
have the same structures but different parameters. After that, the problem of syn-
chronizing different chaotic systems is studied. Some concrete examples are pre-
sented to show how to design the controller. Based on that, a more general case,
synchronizing two different delayed chaotic neural networks with unknown param-
eters, is studied.
In recent years, fuzzy logic systems have received much attention from control
theorists as a powerful tool for nonlinear control. A motivation for using fuzzy sys-
tems and fuzzy control stems in part from the fact that they are particularly suitable
in industrial processes when the physical systems or qualitative criteria are too com-
plex to model and they have provided an efficient and effective way in the control of
complex uncertain nonlinear or ill-defined systems. Among various kinds of fuzzy
control or system methods, the T–S fuzzy system is widely accepted as a power-
ful tool for fuzzy control. Chap. 8 focuses on the modeling and synchronizing of
chaotic and hyperchaotic systems. We first introduce fuzzy modeling methods for
some classical chaotic systems via the T–S fuzzy models. Next, we model some hy-
perchaotic systems with T–S fuzzy models and then, based on these fuzzy models,
we develop an H∞ synchronization method for two different hyperchaotic systems.
Finally, the problem of synchronizing a class of time-delay chaotic systems based
on the T–S fuzzy model is considered.
As a reverse process of suppressing or eliminating chaotic behaviors in order to
reduce the complexity of an individual system or a coupled system, chaotification
aims at creating or enhancing the system’s complexity for some special applica-
tions. In recent years, many conventional control methods and special techniques
were applied to the chaotification of discrete-time dynamical systems or continuous-
time dynamical systems. However, most of them are based on the assumption that
xiv Preface

the analytical representations of the nonlinear dynamical systems to be chaotified

are known exactly. For an unknown or uncertain nonlinear dynamical system, the
above methods are ineffective. Chap. 9 is devoted to the chaotification of dynam-
ical systems which are originally nonchaotic. We develop a simple nonlinear state
feedback controller to chaotify the discrete-time FHM with uncertain parameters.
After that, we use an impulsive and nonlinear feedback control method to chaotify
the continuous-time FHM. We believe that if a nonchaotic system can be approxi-
mated by a fuzzy model with adequate accuracy, then the chaotifying controller of
the fuzzy model will also make the nonchaotic system chaotic. Finally, we chaotify
two classes of continuous-time linear systems via a sampled data control approach.

Acknowledgments

The authors would like to acknowledge the help and encouragement they have re-
ceived during the course of writing this book. A great deal of the materials presented
in this book is based on the research that we conducted with several colleagues and
former students, including Y. H. Xie, D. S. Yang, Z. S. Wang, W. Huang, J. Wang,
Y. B. Quan, Z. B. Liu, and H. X. Guan. We wish to acknowledge especially Dr. T.
D. Ma and Dr. Y. Zhao for their hard work on this book. The authors also wish to
thank Prof. S. S. Wiggins, Prof. A. Medio, Prof. M. Lines, Prof. M. W. Hirsch, Prof.
S. Smale, Prof. R. L. Devaney, and Prof. Y. A. Kuznetsov for their excellent books
on the theory of nonlinear dynamics and chaos. We are very grateful to the National
Natural Science Foundation of China (Grant Nos. 60521003, 60534010, 60572070,
60621001, 60728307, 60774048, and 60804006), the Program for Innovative Re-
search Team in Universities (Grant No. IRT0421), the Program for Changjiang
Scholars, the National High Technology Research and Development Program of
China (863 Program) (Grant No. 2006AA04Z183), and the 111 Project of the Min-
istry of Education of China (B08015), which provided necessary financial support
for writing this book.

Shenyang, China Huaguang Zhang

Beijing, China Derong Liu
Chicago, USA Zhiliang Wang
February 2009
Contents

1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 The Origin and Development of Chaos Theory . . . . . . . . . . . . . . . . . . 1
1.2 Control of Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Suppression of Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Synchronization of Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 Control and Synchronization of Spatiotemporal Chaos . . . . . 9
1.3 Anticontrol of Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2 Preliminaries of Nonlinear Dynamics and Chaos . . . . . . . . . . . . . . . . . . . 17

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Existence, Uniqueness, Flow, and Dynamical Systems . . . . . . . . . . . . 20
2.3.1 Existence and Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.2 Flow and Dynamical Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Equilibrium, Periodic Orbit, Quasiperiodic Orbit, and Poincaré Map 24
2.4.1 Equilibrium of Continuous-Time Systems . . . . . . . . . . . . . . . . 24
2.4.2 Periodic and Quasiperiodic Orbits . . . . . . . . . . . . . . . . . . . . . . 25
2.4.3 Poincaré Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5 Invariant and Attracting Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.6 Continuous-Time Systems in the Plane . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.7 General Solutions of Discrete-Time Linear Systems . . . . . . . . . . . . . 37
2.8 Discrete-Time Systems in the Plane . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.9 Stabilities of Trajectories I: The Lyapunov First Method . . . . . . . . . . 43
2.9.1 The Definition of Lyapunov Stability . . . . . . . . . . . . . . . . . . . . 43
2.9.2 Floquet Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.10 Stabilities of Trajectories II: The Lyapunov Second Method . . . . . . . 50
2.11 Chaotic Sets and Chaotic Attractors . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.12 Symbolic Dynamics and the Shift Map . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.13 Lyapunov Exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

xv
xvi Contents

2.14 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.14.1 Tent Map and Logistic Map . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.14.2 Smale Horseshoe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.14.3 The Lorenz System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2.15 Basics of Functional Differential Equations Theory . . . . . . . . . . . . . . 71
2.16 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

3 Entrainment and Migration Control of Chaotic Systems . . . . . . . . . . . . 77

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.2 Basics on Entrainment and Migration . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2.1 Entrainment Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2.2 Attractor Selection by Migration Control . . . . . . . . . . . . . . . . 80
3.3 OPCL Control Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.4 Global Control of a Class of Continuous-Time Polynomial
Chaotic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.4.1 OPNCL Control Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.4.2 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.5 Global Control of a Class of Discrete-Time Systems . . . . . . . . . . . . . 95
3.5.1 OPNCL Control for Discrete-Time Systems . . . . . . . . . . . . . . 95
3.5.2 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

4 Feedback Control of Chaotic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.2 Model-Reference Adaptive Control of a Class of Discrete-Time
Chaotic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.2.1 Basic Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.2.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.2.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.3 Model-Reference Adaptive Control of a Class of Continuous-Time
Chaotic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.3.1 Basic Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.3.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.3.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.4 Control of a Class of Chaotic Systems Based on Inverse Optimal
Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.4.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.4.2 Inverse Optimal Controller Design . . . . . . . . . . . . . . . . . . . . . . 126
4.4.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Contents xvii

5 Synchronizing Chaotic Systems Based on Feedback Control . . . . . . . . 133

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.2 Synchronization of Continuous-Time Chaotic Systems with a
Single Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.2.1 Basic Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.2.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
5.2.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.3 Synchronization of Multi-Signals in Continuous-Time Chaotic
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.3.1 Basic Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.3.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.3.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
5.4 Synchronization of Different Continuous-Time Chaotic Systems . . . 150
5.4.1 Synchronization of Different Chaotic Systems with
Parameter Perturbations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.4.2 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.5 Synchronization of Discrete-Time Chaotic Systems . . . . . . . . . . . . . . 157
5.5.1 Basic Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
5.5.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.5.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

6 Synchronizing Chaotic Systems via Impulsive Control . . . . . . . . . . . . . . 169

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
6.2 Complete Synchronization of a Class of Chaotic Systems via
Impulsive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
6.2.1 System Description and the Synchronization Problem . . . . . . 170
6.2.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
6.2.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
6.3 Lag-Synchronization of the Unified Systems via Impulsive Control . 180
6.3.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
6.3.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
6.3.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
6.4 Impulsive Synchronization of Different Chaotic Systems . . . . . . . . . 187
6.4.1 System Description and Synchronization Problem . . . . . . . . . 188
6.4.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
6.4.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
6.5 Impulsive Synchronization of a Class of Chaotic Delayed Neural
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
6.5.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
6.5.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
6.5.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
xviii Contents

7 Synchronization of Chaotic Systems with Time Delay . . . . . . . . . . . . . . 209

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
7.2 Adaptive Synchronization of a Class of Delayed Chaotic Systems . . 210
7.2.1 Adaptive Synchronization of the Delayed Chen Chaotic
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
7.2.2 Adaptive Synchronization of the Delayed Lorenz Systems . . 216
7.3 Adaptive Synchronization Between Two Different Delayed
Chaotic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
7.3.1 Adaptive Synchronization Between Delayed Chen Chaotic
Systems and Delayed Lorenz Chaotic Systems . . . . . . . . . . . . 222
7.3.2 Adaptive Synchronization Between Two Different
Delayed Ikeda Chaotic Systems . . . . . . . . . . . . . . . . . . . . . . . . 229
7.3.3 Adaptive Synchronization Between Two Different
Delayed Liao Chaotic Systems . . . . . . . . . . . . . . . . . . . . . . . . . 233
7.4 Synchronization of Chaotic Delayed Neural Networks . . . . . . . . . . . . 238
7.4.1 Synchronization of a Class of Chaotic Delayed Neural
Networks Based on Inverse Optimal Control Theory . . . . . . 238
7.4.2 Adaptive Synchronization Between Two Different Chaotic
Delayed Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267

8 Synchronizing Chaotic Systems Based on Fuzzy Models . . . . . . . . . . . . 269

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
8.2 Modeling Chaotic Systems via T–S Fuzzy Models . . . . . . . . . . . . . . . 270
8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models . . 277
8.3.1 Fuzzy Modeling of Hyperchaotic Systems . . . . . . . . . . . . . . . 277
8.3.2 Synchronization of Two Different Hyperchaotic Systems . . . 289
8.3.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
8.4 Synchronizing Fuzzy Chaotic Systems with Time-Varying Delays . . 298
8.4.1 Basic Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
8.4.2 Main Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
8.4.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306

9 Chaotification of Nonchaotic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
9.2 Chaotification of Discrete-Time Fuzzy Hyperbolic Model with
Uncertain Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
9.2.1 Basic Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
9.2.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
9.2.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
9.3 Chaotification of Continuous-Time Fuzzy Hyperbolic Model
Using Impulsive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
Contents xix

9.3.1 Basic Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318

9.3.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
9.3.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
9.4 Chaotification of Linear Systems Using Sampled Data Control . . . . . 327
9.4.1 Basic Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
9.4.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
9.4.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
9.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
Chapter 1
Overview

Abstract In this chapter, we will provide a brief introduction to the history and de-
velopment of chaos and chaos control. We begin our introduction with chaos theory.
Some important events are narrated in chronological order. After that, we turn our
attention to the field of chaos control. We present this subject in three parts; that is,
suppressing chaos, synchronizing chaos, and chaotification. We will mention some
main approaches used in research and point out open problems in each part.

1.1 The Origin and Development of Chaos Theory

The word ‘chaos’ generally refers to a phenomenon that is disordered and irregu-
lar. In modern scientific terminology, however, ‘chaos’ refers to a pseudo-random
phenomenon generated in a deterministic system. Henri Poincaré is acknowledged
as the first person to glimpse the possibility of chaos. When he studied the stabil-
ity properties of the solar system at the end of the 19th century, Poincaré found
that even in the case of three masses moving under Newton’s law of attraction they
could still exhibit very complicated behavior [56]. This kind of motion depends
sensitively on the initial conditions, thereby rendering long-term prediction impos-
sible [15]. Birkhoff developed geometric methods created by Poincaré and found
many different types of long-term limiting behaviors, such as ω -limit set and α -
limit set. The term ‘dynamical systems’ comes from the work [3] but chaos has
always been in the background. During the first half of the 20th century, nonlinear
oscillators were mostly studied due to their vital role in the development of such
technologies as radio, radar, phase-locked loops, and lasers. Success in technology
stimulates the invention of new mathematical tools. The pioneers in this area include
van der Pol, Andronov, Littlewood, Cartwright, Levinson, and Smale. Especially, in
the 1950s, a type of complexity unknown previously was revealed by Cartwright,
Littlewood, and Levinson to show that a certain forced nonlinear oscillator had an
infinite number of different periods. Smale extended the result of Cartwright, Lit-
tlewood, and Levinson in a general framework and illustrated the phenomenon with

1
2 1 Overview

his ‘horseshoe’ mapping [64]. Meanwhile, in a separate development, Poincaré’s

geometric methods were extended to yield a much deeper understanding of celes-
tial mechanics. One of the most important mathematical achievements, the KAM
theorem, named by taking the first letters of Kolmogorov, Arnol’d, and Moser, was
proved in 1963. This theorem tells us that a Hamiltonian system will still be Hamil-
tonian when it is subjected to tiny perturbations, which gives an answer that the solar
system is stable in some degree. This theorem also implies that in energy preserva-
tion systems complicated behavior would still arise.
Prior to Smale, E. Lorenz, an American meteorologist, made an important con-
tribution in 1963 when he used a computer to study a group of ordinary differ-
ential equations. These equations were reduced from the partial differential equa-
tions which described the turbulent motion of the atmosphere. Lorenz found that a
small change in initial conditions led to very different outcomes in a relatively short
time; this property is called sensitive dependence on initial conditions. This is re-
ally a great discovery since almost all traditional scientists at that time believed that
two trajectories emitted from close initial points would always be close when time
evolves. Laplace’s famous assertion is an extreme reflection of this idea: if there is
an omnipotent spirit who can distinguish any tiny difference of initial conditions
and can discriminate all the forces in the universe, then he can know all the history
and all the future about everything in the universe. From the uncertainty principle
in quantum mechanics we know that one cannot exactly determine the position and
velocity of a particle at the same time. This means that if a system has the property
of sensitive dependence on initial conditions, then one cannot predict its long-term
behavior since we cannot avoid the error of measurement on the initial conditions.
Lorenz used the phrase ‘butterfly effect’ to refer to the phenomenon of sensitive
dependence on initial conditions. He explained that the phenomenon means that a
butterfly flapping its wings in Australia today could affect the weather in the United
States a month later. It was not until the 1970s that Lorenz’s work became known to
the more theoretical mathematical community.
The 1970s were the boom years for chaos. In 1971 Ruelle and Takens proposed
a new theory for the onset of turbulence in fluids, based on abstract considerations
about strange attractors [58]. The word ‘chaos’ was first introduced by T. Y. Li and J.
Yorke in 1975 [37] to designate systems that have aperiodic behavior more compli-
cated than equilibrium, periodic, or quasiperodic motion. But, in fact, their work is
a special case of the theorem obtained by Sharkovskii in 1964 [61], which, because
of political reasons, was not known by western mathematicians for a long time. In
1976, May showed how chaos arises in iterated mappings in population dynamics,
and wrote an influential review article that stressed the pedagogical importance of
studying simple nonlinear systems, to counterbalance the often misleading linear
intuition fostered by traditional education [41]. Next came the most surprising dis-
covery of all, from the physicist Feigenbaum. He discovered that there are certain
universal laws governing the transition from regular to chaotic behavior; roughly
speaking, completely different systems can go chaotic in the same way. His work
established a link between chaos and phase transitions, and enticed a generation of
physicists to the study of dynamics.
1.2 Control of Chaos 3

Entering into the 1980s, computers become a powerful tool, used to help re-
searchers visualize the complicated structures of strange attractors, calculate char-
acteristic indices of the chaotic systems, and provide evidence required by proofs
[19, 69]. Mandelbrot constructed the theory of fractal geometry at the end of the
1970s, and drew the first picture of a Mandelbrot set [39]. The theory of fractal
geometry generalizes the notion of dimension from integers to real numbers and
has become a powerful tool for characterizing the complicated structures of strange
attractors. From the middle of the 1980s, more and more researchers have paid at-
tention to how to control chaos, including suppression, synchronization, and chao-
tification.

1.2 Control of Chaos

There are many practical reasons for controlling chaos. In some situations, such as in
a distributed artificial intelligence system, chaotic behavior is undesirable and will
degrade the performance of the entire system. Naturally, chaos should be reduced
as much as possible, or totally suppressed.
However, recent research has shown that chaos may be useful under certain
circumstances. In the following, we will consider three examples where chaos is
needed.
Since a chaotic attractor is usually embedded in a dense set of unstable limit
cycles, if any of these limit cycles can be stabilized, it may be desirable to stabilize
one that characterizes a certain maximal system performance. This character can be
used to design a multi-purpose system.
Fluid mixing is another typical example in which chaos is not only useful but
also very important. When two fluids are to be thoroughly mixed and the required
energy is to be minimized, one will hope that the particle motion of the fluids is
strongly chaotic, since otherwise it is hard to obtain rigorous mixing properties, due
to the possibility of invariant 2-tori in the flow.
The sensitivity of chaotic systems to small perturbations can be used to steer sys-
tem trajectories to a desired target quickly. NASA scientists, utilizing the sensitivity
of the three-body problem of celestial mechanics to small perturbations, have used
small amounts of residual hydrazine fuel to send the spacecraft ISEE-3/IEC more
than 50 million miles across the solar system, which achieved the first scientific
cometary encounter [62]. This would not have been possible in a nonchaotic system
since a large control effort would be required to quickly direct the system trajectory
to reach distant targets.
For a long time chaotic systems were thought of as unpredictable and uncontrol-
lable: unpredictable, because a small disturbance in the initial conditions or param-
eters of a chaotic system produces changes of the original motion that grow expo-
nentially; uncontrollable, because small disturbances usually lead to other chaotic
states but not to any stable and regular motion [17]. However, the requirement for
controlling chaos, which is either the demand of industrial applications or the aca-
4 1 Overview

demic interest, becomes stronger and stronger than ever before. The first paper on
controlling chaos was published in 1989 [25], but it is the paper by Ott, Grebogi,
and Yorke [46] that attracted more attention and their method is named the OGY
method. In 1990, Pecora and Carroll opened another important application field,
synchronization of chaos [49]. Several years later, Chen and Lai proposed a method
for generating chaos in a nonchaotic discrete-time system [10]. From then on, a
new research aspect, anticontrol of chaos, or chaotification, came into being. Up to
now, numerous control methods have been proposed, developed, and tested. Many
experiments have demonstrated that chaotic physical systems respond quite well
to those newly developed, conventional or novel, simple or sophisticated, control
strategies. Applications of chaos control have been proposed in such diverse areas
of research as biology, medicine, physiology, epidemiology, chemical engineering,
laser physics, electric power systems, fluid mechanics, aerodynamics, electronics,
communications, and so on. Chaos control has become a process that manages the
dynamics of a nonlinear system on a wider scale, with the hope that more benefits
may thus be derived.

1.2.1 Suppression of Chaos

In the beginning, the goal of controlling chaos was to eliminate the harmful chaotic
motion. A remarkable feature which distinguishes chaotic systems from other non-
linear systems is that usually a chaotic system has at least one chaotic attractor.1
Thus, the key of the problem is how to utilize the attractor in the procedure of con-
troller design. The OGY method is based on such an idea: for a neighborhood of
a point located on the chaotic attractor there is always a periodic orbit, which may
be stable or unstable, passing through the neighborhood. Thus, if one tunes the pa-
rameters of the chaotic system slightly such that the trajectory is displaced on the
periodic orbit, then, from that time on, the trajectory will always be along the peri-
odic orbit. Because of the limitation of the OGY method, that it is only adapted to
the low-dimensional case in which the equilibrium points of the considered chaotic
system are saddles and the observed state variables are not contaminated with noise,
some modified OGY methods were proposed again [16, 18, 20, 27, 76]. In nature,
the OGY method is a local method. One may wait for a long time before the trajec-
tory goes into the neighborhood in which the controller will be turned on. However,
because of the ergodicity of the chaotic motion on the attractor, there is always an
instant when the trajectory will be in the neighborhood. To shorten the waiting time,
a targeting method was proposed which can direct the trajectory of the chaotic sys-
tem into the neighborhood in a very short time [62, 63]. But sometimes it is difficult

1 In some literature, a strange attractor has a different meaning to a chaotic attractor. The former
emphasizes the geometrical structure of the attractor and sometimes has a fractional dimension.
The latter emphasizes the sensitivity to initial conditions. Two trajectories starting from differ-
ent but close initial points will be divergent exponentially. Therefore, a system having a strange
attractor may not be a chaotic one while a chaotic attractor may have integer dimension [47].
1.2 Control of Chaos 5

to obtain the parameters of a chaotic system. Even if the parameters can be ob-
tained, it is difficult to adjust them as control inputs. So, the continuous feedback
control method is considered. There is a problem which should be answered when
continuous feedback is used: will the topological structure of the chaotic attractor
be destroyed by the continuous feedback? At present, we can answer the question
as follows: if all of the invariant sets of a chaotic system are hyperbolic invari-
ant sets, then the small perturbation will not destroy the topological structure of its
state space [80]. We know that many chaotic systems have hyperbolic invariant sets.
Therefore, at least for these systems, the continuous feedback control method is fea-
sible. The studies over the last decade show that almost all the methods that do well
for conventional nonlinear control problems can be used to suppress chaos (see [8]
and references therein).
One main goal of controlling chaos is to steer the trajectory of a chaotic system to
a periodic one. So, the first step of the control procedure is to identify these periodic
orbits.
Suppose that the mathematical model is known. For discrete-time systems, the
periodic orbits can be obtained by calculation according to its definition, no mat-
ter whether stable or unstable. For continuous-time systems, the problem can be
changed to the case of discrete time by the Poincaré section method. Thus, the key
problem is how to identify the periodic orbits of discrete-time systems. In practice,
one wants to use computers to find the periodic orbits of the discrete-time systems.
The iterations of a discrete-time system can only get stable periodic orbits. For un-
stable periodic orbits (UPOs) the iteration method loses its effectiveness because
even if a point is located on a UPO at present, at the next instant of the evolution
it will deviate from the UPO due to small errors caused either by calculation or by
measurement.
Four methods are known which work well for identifying UPOs. The first method
is called the word-lifting technique which was proposed by Hao [22]. Using this
method an analytical expression of periodic orbits can be obtained. The second one
was proposed in [45] in which both symbol dynamics and graph theory are used to
determine a periodic orbit. The main idea of this method is the construction of a
directed graph which represents the structure of the state space for the dynamical
system under investigation. The third one is to transform the original chaotic system
by a homeomorphism (the term will be introduced in the following chapter). With-
out changing the topological structure of the state space, this transformation only
changes the stabilities of the orbits. Thus, UPOs become stable periodic orbits of
the new system. Once these stable periodic orbits are identified by some numerical
algorithms [13, 35, 52, 53], we can get the UPOs of the original chaotic system by
inverse transformation. The fourth method is very ingenious, in which the retarded
state variables are taken as inputs [54]. Considering the fact that a chaotic trajectory
can approach any periodic orbit, if we take an expected period as the delayed time
and design a stabilizing controller, then the trajectory is just on the expected peri-
odic orbit when the controlled chaotic system is stabilized. Only the structure of the
system and period of orbit are needed in the delayed feedback method. The delayed
feedback method is a two-in-one method. The procedure of identifying the UPOs
6 1 Overview

is combined with the procedure of stabilization. The periodic orbit will be obtained
provided that the controlled system is stabilized. This merit makes the method em-
ployed extensively [4, 5, 6, 7, 23, 33, 34, 66]. However, with deeper study a fatal
problem is found with the delayed feedback method; that is, the delayed feedback
method cannot stabilize the UPOs which have real eigenvalues of odd number. This
is called the odd number limitation. Fortunately, there are some modified delayed
feedback methods which can go round the obstacle of the odd number limitation
[43, 44, 70].
If the mathematical model is unknown and only some measurable signals can be
obtained, then, at first, one needs to utilize the technique of space reconstruction to
get a model. Thus, the four methods discussed above can be used again [9, 65].
A merit of the strategy of feedback control is that the energy of input is small
and will not destroy the structure of the state space and, when the control goal is
reached, the controller can be turned off. However, the strategy requires the con-
tinuous monitoring of the state, which is difficult in some cases. For example, the
extremely high temperature in plasma will make it very difficult to precisely mea-
sure some physical quantities in the plasma [24]. Non-feedback control methods for
suppressing chaos can overcome this difficulty. There are many different kinds of
non-feedback control methods. Some of them do not require the understanding of
the chaotic system dynamics, such as the method with a periodic signal to suppress
chaos [42] or the method with a noise signal to suppress chaos [55]. When imple-
menting a control strategy, there are fewer limitations required by non-feedback con-
trol methods compared with that of feedback control methods. But, the cost is that
with the non-feedback control strategy one can only achieve some general control
goals, such as eliminating chaos, realizing periodic motion, etc. Some non-feedback
control strategies can steer the trajectories of a chaotic system to some predesigned
orbits, which, especially, may not be the inherent orbits of the original chaotic sys-
tem [12, 29, 30, 31]. But, in that situation, one must have an exact understanding
about the dynamics of the original system. When a non-feedback control method is
used, the control signal would not reduce to zero even if the control goal has been
realized, which will produce new problems in practice.

1.2.2 Synchronization of Chaos

The common meaning of synchronization is that the phases of two or more oscilla-
tors change according to some patterns. C. Huygens was the first person who found
the phenomenon of synchronization of coupled pendulums. Actually, synchroniza-
tion phenomena are abundant in science, nature, engineering, and social life. Sys-
tems as diverse as clocks, singing crickets, cardiac pacemakers, firing neurons, and
applauding audiences exhibit a tendency to operate in synchrony [50]. These phe-
nomena are universal and can be understood within a common framework based
on modern nonlinear dynamics. Existing studies of suppressing chaos show that a
periodic input signal can make a chaotic system periodic with the same frequency
1.2 Control of Chaos 7

as that of the input signal. This situation can be looked at as a pattern of synchro-
nization in which a periodic signal synchronizes with a chaotic signal (it can be
thought of as a periodic motion with period of infinite length). Furthermore, if the
input signal is replaced with a chaotic signal, what will happen? Will there be the
phenomenon of synchronization? Intuitively, even if the two chaotic signals are gen-
erated from the chaotic systems with the same structure, this synchronization would
yet not happen because of the sensitivity of chaotic systems. However, Pecora and
Carroll firstly found the synchronization of chaos in the experiment of circuits. They
took the Lorenz system as the drive system, and replicated the same system as the
response system. When they replaced the first state variable of the response sys-
tem with the first state variable of the drive system, the synchronization happened.
This synchronization scheme is called the P–C scheme. They found that the syn-
chronization in the P–C scheme is not only globally asymptotically stable but also
structurally stable. By a strict analysis they pointed out that two chaotic systems in
the P–C scheme can get synchronized only if the conditional Lyapunov exponents
of the response system are all negative.2
Although the problems considered by Ott and Pecora seemed different and inde-
pendent from each other, soon researchers realized the relationship between them.
In fact, many control methods developed in suppressing chaos can also be used in
synchronizing chaos. In this section, rather than talking more about the synchro-
nization methods, we pay more attention to the the patterns of synchronization that
can be obtained among chaotic systems. For convenience, suppose that we have the
following chaotic systems:3

ẋ = f (t, x), y = h(x), x ∈ Rn , h : Rn → Rm . (1.1)

Taking (1.1) as the drive system, we construct the response system as

x̂˙ = fˆ(t, x̂, y), ŷ = ĥ(x̂), x̂ ∈ R p , ĥ : R p → Rq , (1.2)

where x and x̂ are state variables and y and ŷ are outputs. Suppose that the following
relation holds when (1.1) and (1.2) get synchronized:

ŷ = ϕ (y), ϕ : Rm → Rq , (1.3)
where ϕ is continuous but not necessarily differentiable [26]. (y, ŷ) is called the
manifold of synchronization.
Considering the dynamics of drive–response systems and the structure of the
manifold of synchronization, we give the following categories of synchronization
of chaos.

2 It should be pointed out that this condition is not necessary for synchronization of chaos. Some
chaotic systems cannot get synchronized as in P–C schemes, but they indeed can get synchronized
[21, 48].
3 We take a continuous-time system as an example. The same illustration can also be done for

discrete-time systems.
8 1 Overview

(i) Complete synchronization. In this case, x̂(t) = x(t); that is, m = q and ϕ is an
identical mapping. If fˆ = f , the relation is called identically complete synchro-
nization. If fˆ = f , the relation is called nonidentically complete synchroniza-
tion. The manifold of complete synchronization is a diagonal hyperplane in the
phase space composed of x − x̂.
(ii) Generalized synchronization [59]. If the state of the response system is uniquely
determined by the state of the drive system, that is, (1.3) holds with ϕ a certain
mapping, then the drive system and the response system are thought of as in the
pattern of generalized synchronization. No conditions are imposed on m and q.
In this case, the manifold of synchronization is no longer a diagonal hyperplane
of the combined phase space x − x̂.4
(iii) Lag synchronization [57]. In this case, x̂(t) ≈ x(t − τ ), where τ is a small posi-
tive number.
(iv) Anticipating synchronization [72]. Analogous to lag synchronization, when an-
ticipating synchronization happens, the state variables between the drive system
and the response system have the following relationship:

x̂(t) = x(t + τ ), τ > 0.

(v) Phase synchronization. Let φ1 and φ2 be the phases of system (1.1) and system
(1.2), respectively. The phase synchronization means that for rational numbers
m and n, there is a small positive number ε such that

|mφ1 − nφ2| < ε .

The key in phase synchronization is the determination of phase φ1 and phase φ2

[51]. There is no unified definition in high-dimensional chaotic systems. There
are three common methods for the low-dimensional chaotic systems [24]:
a. Choose a state variable, such as x(t), and find the instants, t1 ,t2 , . . . ,tn , . . . , at
which the state variable gets its local maxima. Then, the phase of a chaotic
trajectory can be defined as
t − tn
ϕM = 2π + 2π n, tn+1 > t > tn . (1.4)
tn+1 − tn
b. If the flow of chaotic oscillators has a proper rotation around a certain refer-
ence point, the phase can be defined in a more intuitive and straightforward
manner. Project the trajectory on to a two-dimensional plane, for example
the x–y plane, with the origin at the reference point. The phase ϕP can sim-
ply be defined by

4 A numerical method to verify whether a stable generalized synchronization happens or not is

to construct an auxiliary system with the same structure as (1.2). Provide the same input signals
derived from the drive system to the response system and the auxiliary system. Choose the initial
point of the auxiliary system in the neighborhood of the initial point of the response system. If a
stable generalized synchronization happens, then the trajectories of the auxiliary system and the
response system will converge to each other. Otherwise, they will be different [1].
1.2 Control of Chaos 9

tan ϕP = y/x. (1.5)

c. Using Hilbert transformation. Choose a scalar s(t) and define a complex
variable as
ξ (t) = s(t) + iŝ(t) = A(t)eiϕH (t) , (1.6)
 ∞
−1 s(τ )
ŝ(t) = π P dτ ,
−∞ t −τ
∞
where P −∞ means the Cauchy principal value of the integral.
Remark 1.1. The phenomena of lag synchronization and generalized synchroniza-
tion need to be considered when a synchronization system is designed. First, the
perturbations caused by the noise of the environment and measurement cannot be
ignored. Therefore, the drive system and the response system are almost never iden-
tical. Second, no matter how fast the signal transmission is in the channel, time is
always needed. Thus, at the port of the response system, the input signal is not at
the same time instant as that of the drive system. In fact, the drive system and the
response system are as follows:
⎧
⎪ dx(t)
⎪
⎨ = f (t, x(t)), t = t − τ ,
dt
⎪
⎩ dx̂(t) = fˆ(t, x̂(t), x(t)).
⎪
dt

Remark 1.2. The phenomenon of phase synchronization is completely different
from the preceding cases. In general, when synchronization occurs, the conditional
Lyapunov exponents of the response system are all negative. At that time, the re-
sponse system is a nonchaotic system with a chaotic output. When phase synchro-
nization takes place, since the relation is on the phases of the synchronized systems,
the conditional Lyapunov exponents may be positive. 
The synchronizations mentioned above are all schemes of unidirectional couples.
There is much more content to be studied if the coupling scheme is bidirectional
and the number of coupled systems is greater than two. No matter which scheme
of synchronization, the above five kinds of phenomena of synchronization are still
encountered the most.

1.2.3 Control and Synchronization of Spatiotemporal Chaos

All of the chaotic systems mentioned above are described by ordinary differential
equations or difference equations. This means that for these chaotic systems, the
variation rate of the state variables with respect to spatial position is ignored. We
call these chaotic systems temporal chaotic systems. But, sometimes, a chaotic sys-
tem is not isotropic and must be described with partial differential equations. For
10 1 Overview

these chaotic systems, the synchronization can take place either in time or in space
position. Since the state of a partial differential equation belongs to a function space
whose dimension is infinite, a spatiotemporal chaotic system can provide more pat-
terns than those in a temporal chaotic system. Thus, the considered problems of
synchronizing two or more spatiotemporal chaotic systems are much more compli-
cated than those considered for temporal chaotic systems. At present, research on
control and synchronization of spatiotemporal chaos mainly follows two directions
[24, 32, 79]:
(i) The studies of coupled map lattices. Each lattice is a discrete mapping and
evolves with time. The adjacent lattices are joined with specified boundary con-
ditions. There are two reasons for studying coupled map lattices. First, it can
be obtained by discretizing a partial differential equation. Therefore, many im-
portant characteristics that belong to partial differential equations can be found
in the systems of coupled map lattices. Second, mathematical tools required by
coupled map lattices are much simpler than what is needed when one directly
studies partial differential equations.
(ii) The studies of the underlying partial differential equations. Many of the im-
portant properties of a partial differential equation cannot be presented by a
coupled map lattice. Therefore, it is necessary to study the original partial dif-
ferential equations.
There is a marked difference between temporal chaotic systems and spatiotemporal
chaotic systems when one wants to control them. A temporal chaotic system can be
controlled only by altering its vector fields through a feedback controller. But even
for a coupled map lattice this method is impractical since lots of controllers would
be required. A feasible and effective method is to alter the state of some lattices
and diffuse the alteration through coupling relations so that the control goal can be
achieved [24].
Nowadays, more and more researchers have realized that the study of spatiotem-
poral chaos can provide instructive inspiration and necessary ways for solving the
problem of turbulence.

1.3 Anticontrol of Chaos

Since lots of applications of chaos are found in many areas, a new notion, anticontrol
of chaos (or chaotification), was proposed [10]. Later, the concept and method of
synthesis of chaos were proposed [71]. Anticontrol of chaos became a new direction
in the study of chaos control. It seems that based on the well-developed theory of
chaos there should be no difficulties in chaotification. However, it is not so. At the
beginning, the difficulties mainly exist in the following aspects:
(i) How to define chaos? The first definition of chaos was proposed by Li and
Yorke in 1975 for the mappings on intervals of one-dimensional space [37].
1.3 Anticontrol of Chaos 11

Later, Marotto generalized this definition to the differential mappings on high-

dimensional space [40]. Another definition in common use was proposed by De-
vaney from the topological point of view [14]. Based on Devaney’s definition,
Wiggins proposed a more concise one which has been widely used [28, 75].
However, these definitions are all for discrete-time systems. For continuous-
time systems, there is no unified criterion for what is chaos. A practical method
is to compute the Lyapunov exponents of the controlled system. If there is a
positive Lyapunov exponent, then it is believed that the controlled system be-
comes chaotic [2]. But this method is not the ideal one since, in a great degree,
the computation of Lyapunov exponents depends on the numerical algorithm as
well as the precision of computation.
(ii) How to design a chaotifying controller? If the Lyapunov exponents are used as
the criterion, how does one design a chaotifying controller? We may chaotify
some systems by tuning certain parameters. But there are also many systems
which are not sensitive to the variation of parameters and, therefore, would not
be chaotified no matter how the parameters are tuned. Furthermore, for some
complicated systems, such as a coupled map lattice or a partial differential
equation, we cannot determine what method can really make them chaotified
because of the limited understanding about their dynamics.
(iii) How to utilize the existing theoretical results? Some theoretical results, such as
the Melnikov method and the Shilnikov method [75], although useful in theo-
retical analysis, are difficult to apply in practice.
A nice controller of chaotification should not only be easy to implement but also
have a solid mathematical foundation. From 1996, Chen and Wang developed a se-
ries of feedback methods which can make a wide class of nonchaotic systems chao-
tified. By the inspiration of their works, many chaotifying methods were proposed
[11].
According to the criteria of chaos judgement, the existing methods of chaotifica-
tion can be categorized as follows:
(i) Boundedness + positive Lyapunov exponents + Marotto theorem. Since the
motion of chaos is in a bounded region and is sensitively dependent on initial
conditions, the key in the procedure of chaotification is to make the controlled
system have a positive Lyapunov exponent while its trajectory remains in a
bounded region. For discrete-time systems, by using, modulo operation, Chen
and Lai first proposed a chaotifying method which can guarantee that all the
Lyapunov exponents are positive. Later, Chen and Wang improved the result
by replacing the modulo operation by a small amplitude controller. Moreover,
by using the Marotto theorem, Chen and Lai provided a proof that the chaos
generated by their methods satisfies the definition of Li–Yorke [10].
(ii) Introducing chaotic dynamics. This class includes three main methods:
a. Using continuous feedback control. First, transform some well-known chaotic
systems, such as the Lorenz system, the Rössler system, etc., into a normal
format, and then extract their nonlinear terms as one part of the controller.
12 1 Overview

Another part of the controller is designed to stabilize the original system.

These two parts constitute a total chaotification controller [73].
b. Using impulsive control. The system is excited by impulses at discrete in-
stants. The impulsive signal is composed of two parts. One part is to elimi-
nate the original motion, the other part is to make the system chaotic. Usu-
ally, this part is composed of chaotic signals [78].
c. Introducing time delay. This method is based on the following idea: the
introduction of time delay makes a finite-dimensional system become an
infinite-dimensional system. Thus, it is possible to produce more compli-
cated behavior within it, especially chaos. Along this route, Wang studied
further the problem of chaotification of continuous-time systems and he
proposed a method which makes the outputs of the controlled system ap-
proximate the outputs of a chaotic system [74].
(iii) Proving the existence of a Smale horseshoe. In the literature [68], a piecewise
controller was designed to make a Duffing oscillator generate a Smale horse-
shoe. This method is difficult to generalize to other systems. Recently, Yang et
al. proposed a computer auxiliary proving method to show the existence of a
Smale horseshoe. When the method is applied, Poincaré mapping needs to be
constructed first. The merit of the method is that it makes the Melnikov method
better operated [77].
(iv) Switching mechanism. Lü et al. proposed a method in 2002, which can generate
a chaotic system with a complex attractor. The method is simple and effective,
but no rigorous proof is given to explain why chaos occurs [38].
Over the past ten years, many important results have been obtained. But there are
still a lot of problems remaining unsolved, such as generating hyperchaotic sys-
tems, chaotifying a partial differential equation, etc. Undoubtedly, these problems
are great challenges for researchers.

1.4 Summary

In this chapter, we briefly introduced the history and presented the current status
of chaos and chaos control. Due to the focus of the book, we do not list all the
methods developed in chaos control. Our attention is to give an introduction to the
development of theory so as to provide a rough description of chaos control for a
new researcher in this area. Readers who are interested in how the synchronization
of chaos is utilized in applications are referred to, e.g., [36] and [67].
References 13

References

1. Abarbanel HDI, Rulkov NF, Sushchik MM (1996) Generalized synchronization of chaos: the
auxiliary system approach. Phys Rev E 53:4528–4535
2. Alligood KT, Sauer TD, Yorke JA (1997) Chaos – An Introduction to Dynamical Aystems.
Springer, New York
3. Birkhoff GD (1966) Dynamical Systems (revised edtion). American Mathematical Society,
Providence, RI
4. Bielawski S, Derozier D, Glorieux P (1993) Experimental characterization of unstable peri-
odic orbits by controlling chaos. Phys Rev A 47:2492–2495
5. Bleich ME, Hochheiser D, Moloney JV, Socolar J (1997) Controlling extended systems with
spatially filtered, time-delayed feedback. Phys Rev E 55:2119–2126
6. Bleich ME, Socolar J (2000) Delayed feedback control of a paced excitable oscillator. Int J
Bifurc Chaos 10:603–609
7. Brandt ME, Chen G (2000) Time-delay feedback control of complex pathological rhythms in
an atrioventricular conduction model. Int J Bifurc Chaos 10:2781–2784
8. Chen G, Dong X (1998) From Chaos to Order: Methodologies, Perspectives and Applica-
tions. World Scientific, Singapore
9. Chen G, Chen G, de Figueiredo RJP (1999) Feedback control of unknown chaotic dynamical
systems based on time-series data. IEEE Trans Circuits Syst I 46:640–644
10. Chen G, Lai D (1996) Feedback control of Lyapunov exponents for discrete-time dynamical
systems. Int J Bifurc Chaos 6:1341–1349
11. Chen G, Wang XF (2006) Chaotification of Dynamical Systems: Theories, Methodologies,
and Applications. Shanghai Jiaotong University Press, Shanghai
12. Chen L (2001) An open-plus-closed-loop control for discrete chaos and hyperchaos. Phys
Lett A 281:327–333
13. Davidchack RL, Lai Y (1999) Efficient algorithm for detecting unstable periodic orbits in
chaotic systems. Phys Rev E 60:6172–6175
14. Devaney RL (1989) An Introduction to Chaotic Dynamical Systems, 2nd edn. Addison-
Wesley, New York
15. Diacu F, Holmes P (1996) Celestial Encounters: The Origins of Chaos and Stability. Prince-
ton University Press, Princeton, NJ
16. Ding M, Ott E, Grebogi C (1994) Controlling chaos in a temporally irregular environment.
Phys D 74:386–394
17. Dyson F (1988) Infinite in All Directions. Harper and Row, New York
18. Epureanu BI, Dowell EH (2000) Optimal multi-dimensional OGY controller. Physica D
139:87–96
19. Enns R, McGuire G (1997) Nonlinear Physics with Maple for Scientists and Engineers.
Birkhäuser, Boston
20. Grebogi C, Lai Y (1997) Controlling chaos in high dimensions. IEEE Trans Circuits Syst I
44:971–975
21. Gutiérrez JM, Iglesias A (1998) Synchronizing chaotic systems with positive conditional Lya-
punov exponents by using convex combinations of the drive and response systems. Phys Lett
A 239:174–180
22. Hao B (1989) Elementary Symbolic Dynamics and Chaos in Dissipative Systems. World Sci-
entific, Singapore
23. Hikihara T, Kawagoshi T (1996) An experimental study on stabilization of unstable periodic
motion in magneto-elastic chaos. Phys Lett A 211:29–36
24. Hu G, Xiao J, Zheng Z (2000) Chaos Control. Shanghai Scientific and Technological Educa-
tion Publishing House, Shanghai
25. Hübler A (1989) Adaptive control of chaotic systems. Helv Phys Acta 62:343–347
26. Hunt BR, Ott E, Yorke JA (1997) Differentiable generalized synchronization of chaos. Phys
Rev E 55:4029–4034
14 1 Overview

27. Iplikci S, Denizhan Y (2001) Control of chaotic systems using targeting by extended control
regions method. Physica D 150:163–176
28. Ingraham R (1992) A Survey of Nonlinear Dynamics (‘Chaos Theory’). World Scientific,
Singapore
29. Jackson EA (1991) On the control of complex dynamic systems. Physica D 50:341–366
30. Jackson EA, Grosu I (1995) An open-plus-closed-loop (OPCL) control of complex dynamic
systems. Physica D 85:1–9
31. Jackson EA, Hübler A (1990) Periodic entrainment of chaotic logistic map dynamics. Physica
D 44:407–420
32. Kahan S, Montagne R (2001) Controlling spatio-temporal chaos. Physica A 290:87–91
33. Konishi K, Hirai M, Kokame H (1998) Decentralized delayed-feedback control of a coupled
map model for open flow. Phys Rev E 58:3055–3059
34. Konishi K, Kokame H, Hirata K (1999) Coupled map car-following model and its delayed-
feedback control. Phys Rev E 60:4000–4007
35. Lan Y, Cvitanović P (2004) Variational method for finding periodic orbits in a general flow.
Phys Rev E 69:061217(1–10)
36. Larson LE, Liu JM, Tsimring LS (2006) Digital Communications using Chaos and Nonlinear
Dynamics. Springer, New York
37. Li TY, Yorke JA (1975) Period three implies chaos. Am Math Mon 82:985–992
38. Lü J, Zhou TS, Chen G, Yang X (2002) Generating chaos with a switching piecewise-linear
controller. Chaos 12:344–349
39. Mandelbrot B (1983) The Fractal Geometry of Nature. Freeman, Paris
40. Marotto FR (1978) Snap-back repellers imply chaos in Rn . J Math Anal Appl 63:199–223
41. May RM (1976) Simple mathematical model with very complicated dynamics. Nature
261:459–467
42. Mirus KA, Sprott JC (1999) Controlling chaos in a high dimensional system with periodic
parametric perturbations. Phys Lett A 254:275–278
43. Nakajima H (1997) On analytical properties of delayed feedback control of chaos. Phys Lett
A 232:207–210
44. Nakajima H, Ueda Y (1998) Limitation of generalized delayed feedback control. Physica D
111:143–150
45. Osipanko G (2007) Dynamical Systems, Graphs, and Algorithms. Springer, Berlin
46. Ott E, Grebogi C, Yorke JA (1990) Controlling chaos. Phys Rev Lett 64:1196–1199
47. Ott E (2002) Chaos in Dynamical Systems. Cambridge University Press, London
48. Pan S, Yin F (1997) Using chaos to control chaotic systems. Phys Lett A 231:173–178
49. Pecora LM, Carroll TL (1990) Synchronization in chaotic systems. Phys Rev Lett 64:821–824
50. Pilovsky AS, Rosenblum MG, Kurths J (2001) Synchronization: A Universal Concepts in
Nonlinear Science. Cambridge University Press, London
51. Pilovsky AS, Rosenblum MG, Osipov GV, Kurths J (1997) Phase synchronization of chaotic
oscillators by external driving. Physica D 104:219–238
52. Pingel D, Schmelcher P, Diakonos FK, Biham O (2000) Theory and application of systematic
detection of unstable periodic orbits in dynamical systems. Phys Rev E 62:2119–2134
53. Pingel D, Schmelcher P, Diakonos FK (2001) Detecting unstable periodic orbits in chaotic
continuous-time dynamical systems. Phys Rev E 64:026214
54. Pyragas K (1992) Continuous control of chaos by self-controlling feedback. Phys Lett A
170:421–428
55. Ramesh M, Narayanan S (1999) Chaos control by nonfeedback methods in the presence of
noise. Chaos Solitons Fractals 10:1473–1489
56. Robinson RC (2004) An Introduction to Dynamical Systems: Continuous and Discrete. Pren-
tice Hall, New York
57. Rosenblum MG, Pikovsky AS, Kurths J (1997) From phase to lag synchronization in coupled
chaotic oscillators. Phys Rev Lett 78:4193–4196
58. Ruelle D, Takens F (1971) On the nature of turbulence. Commun Math Phys 20:167–192
59. Rulkov NF, Sushchik MM, Tsimring LS, Abarband HDI (1995) Generalized synchronization
of chaos in directionally coupled chaotic systems. Phys Rev E 51:980–994
References 15

60. Schiff SJ, Jerger K, Duong DH, Chang T, Spano ML, Ditto WL (1994) Controlling chaos in
the brain. Nature 370:615–620
61. Sharkovskii AN (1964) Coexistence of cycles of a continuous map of a line into itself. J Ukr
Math 16:61–71
62. Shinbort T, Grebogi C, Yorke JA, Ott E (1993) Using small perturbations to control chaos.
Nature 363:411–417
63. Shinbort T, Ott E, Grebogi C, Yorke JA (1990) Using chaos to direct orbits to targets. Phys
Rev Lett 65:3215–3218
64. Smale S (1967) Differentiable dynamical systems. Bull Amer Math Society 73:747–817
65. So P, Ott E, Sauer T, Gluckman BJ, Grebogi C, Schiff SJ (1997) Extracting unstable periodic
orbits from chaotic time series data. Phys Rev E 55:5398–5417
66. Socolar JES, Sukow DW, Gauthier DJ (1994) Stabilizing unstable periodic orbits in fast dy-
namical systems. Phys Rev E 50:3245–3248
67. Stavroulakis P (2006) Chaos Applications in Telecommunications. CRC Press, New York
68. Tang KS, Man KF, Zhong GQ, Chen G (2001) Generating chaos via x|x|. IEEE Trans Circuits
Syst I 48:636–641
69. Tucker W (2002) A rigorous ODE solver and Smale’s 14th problem. Found Comput Math
2:53–117
70. Ushio T (1996) Limitation of delayed feedback control in nonlinear discrete-time systems.
IEEE Trans Circuits Syst I 43:815–816
71. Vanecek A, Celikovsky S (1996) Control Systems: From Linear Analysis to Synthesis of
Chaos. Prentice Hall, Upper Saddle River, NJ
72. Voss H (2000) Anticipating chaotic synchronization. Phys Rev E 61:5115–5119
73. Wang XF, Chen G (2003) Generating topologically conjugate chaotic systems via feedback
control. IEEE Trans Circuits Syst I 50:812–817
74. Wang XF, Chen G, Yu X (2000) Anticontrol of chaos in continuous-time systems via time-
delay feedback. Chaos 10:771–779
75. Wiggins S (2003) Introduction to Applied Nonlinear Dynamical Systems and Chaos, 2nd edn.
Springer, New York
76. Yagasaki K, Uozumi T (1998) New approach for controlling chaotic dynamical systems. Phys
Lett A 238:349–357
77. Yang XS, Li QD (2007) Chaotic Systems and Chaotic Circuits. Science Press, Beijing (in
Chinese)
78. Zhang HG, Wang ZL, Liu D (2005) Chaotifying fuzzy hyperbolic model using impulsive and
nonlinear feedback control approaches. Int J Bifurc Chaos 15:2603–2610
79. Zhao Y (2003) Introduction to some methods of chaos analysis and control for PDEs. In:
Chen G (ed) Chaos Control. Springer, New York
80. Zhang Z (1997) The Principle of Differential Dynamical Systems. Science Press, Beijing
Chapter 2
Preliminaries of Nonlinear Dynamics and Chaos

Abstract This chapter provides a brief review of some concepts and tools related to
the subject of the monograph – chaos suppression, chaos synchronization, and chao-
tification. After a quick review of the history of ‘dynamical systems,’ we provide a
summary of important definitions and theorems, including equilibrium points, pe-
riodic orbits, quasiperiodic orbits, stable and unstable manifolds, attractors, chaotic
attractors, Lyapunov stability, orbital stability, and symbolic dynamics, which are
all from the theory of ordinary differential equations and ordinary difference equa-
tions. The results are summarized both for continuous time and for discrete time.
Then, we present three examples for chaotic attractors including the logistic map,
the Lorenz attractor, and the Smale horseshoe. At the end of the chapter, we pro-
vide some necessary definitions and theorems of functional differential equations
(PDEs).

2.1 Introduction

Roughly speaking, a dynamical system consists of two ingredients: a rule or ‘dy-

namics,’ which is described by a set of equations (difference, differential, integral,
functional, or abstract operator equations, or a combination of some of them) and
specifies how a system evolves, and an initial condition or ‘state’ from which the
system starts. A nonlinear dynamical system is a dynamical system described by a
set of nonlinear equations; that is, the dynamical variables describing the properties
of the system (for example, position, velocity, acceleration, pressure, etc.) appear in
the equations in a nonlinear form. The most successful class of rules for describ-
ing natural phenomena are differential equations. All the major theories of physics
are stated in terms of differential equations. This observation led the mathematician
V. I. Arnol’d to comment, ‘consequently, differential equations lie at the basis of
scientific mathematical philosophy’ [2]. This scientific philosophy began with the
discovery of calculus by Newton and Leibniz and continues to the present day. The
theory of dynamical systems grew out of the qualitative study of differential equa-

17
18 2 Preliminaries of Nonlinear Dynamics and Chaos

tions, which in turn began as an attempt to understand and predict the motions that
surround us such as the orbits of the planets, the vibrations of a string, the ripples on
the surface of a pond, and the forever evolving patterns of the weather. The first two
hundred years of this scientific philosophy, from Newton and Euler to Hamilton and
Maxwell, produced many stunning successes in formulating the ‘rules of the world,’
but only limited results in finding their solutions.
By the end of the 19th century, researchers had realized that many nonlinear dif-
ferential equations did not have explicit solutions. Even the case of three masses
moving under the laws of Newtonian attraction could exhibit very complicated be-
havior and its explicit solution was not possible to obtain (e.g., the motion of the
sun, the earth, and the moon cannot be given explicitly in terms of known func-
tions). Short-term solutions could be given by power series, but these were not use-
ful in determining long-term behavior. The modern theory of nonlinear dynamical
systems began with Poincaré at the end of the 19th century with fundamental ques-
tions concerning the stability and evolution of the solar system. Poincaré shifted
the focus from finding explicit solutions to discovering geometric properties of so-
lutions. He introduced many ideas in specific examples. In particular, he realized
that a deterministic system in which the outside forces are not varying and are not
random can exhibit behavior that is apparently random (i.e., chaotic). Poincaré’s
point of view was enthusiastically adopted and developed by G. D. Birkhoff. He
found many different types of long-term limiting behavior. His work resulted in the
book [4] from which the term ‘dynamical systems’ came. Other people, such as Lya-
punov, Pontryagin, Andronov, Morser, Smale, Peixoto, Kolmogorov, Arnol’d, Sinai,
Lorenz, May, Yorke, Feigenbaum, Ruelle, and Takens, all made important contribu-
tions to the theory of dynamical systems. The field of nonlinear dynamical systems
and especially the study of chaotic systems has been hailed as one of the important
breakthroughs in science in the 20th century. Today, nonlinear dynamical systems
are used to describe a vast variety of scientific and engineering phenomena and have
been applied to a broad spectrum of problems in physics, chemistry, mathematics,
biology, medicine, economics, and various engineering disciplines.
This chapter is a brief review of some concepts and tools related to the subject
of the monograph – chaos suppression, chaos synchronization, and chaotification.
The goal of the chapter is to provide readers with some necessary background on
nonlinear dynamical systems and chaos so as to ease the difficulty when they read
subsequent chapters of this book. Readers interested in the complete theory of dy-
namical systems are recommended to refer to [1, 7, 11, 13, 17].

2.2 Background

Two types of models are extensively studied in the field of dynamical systems: the
continuous-time model and the discrete-time model. Most continuous-time nonlin-
ear dynamical systems are described by a differential equation of the form
2.2 Background 19

ẋ = f (x,t; μ ), (2.1)

with x ∈ U ⊂ Rn , t ∈ R, and μ ∈ V ⊂ R p , where U and V are open sets in Rn and

R p . Meanwhile, most discrete-time nonlinear dynamical systems are described by
an equation of the form

x(k + 1) = f (x(k), k; μ ), k = 0, 1, 2, . . . . (2.2)

We refer to (2.1) as a vector field or ordinary differential equation and to (2.2) as

a map or difference equation. By a solution of (2.1) we mean a map, x, from some
interval I ⊂ R into Rn , which is denoted by

x : I → Rn ,
t → x(t),

such that x(t) satisfies (2.1), i.e.,

ẋ(t) = f (x(t),t; μ ).

The map x has the geometrical interpretation of a curve in Rn , and (2.1) gives the
tangent vector at each point of the curve, hence the reason for referring to (2.1) as
a vector field. We will refer to the space of independent variables of (2.1) (i.e., Rn )
as the phase space or state space. One goal of the study of dynamical systems is to
understand the geometry of solution curves in the phase space. It is useful to distin-
guish a solution curve which passes through a particular point in the phase space at
a specific time, i.e., for a solution x(t) with x(t0 ) = x0 . We refer to this as specifying
an initial condition or initial value. This is often included in the expression for a
solution by x(t,t0 , x0 ). In some situations explicitly displaying the initial condition
may be unimportant, in which case we will denote the solution merely as x(t). In
other situations the initial time may be always understood to be a specific value,
say t0 = 0; in this case we would denote the solution as x(t, x0 ). Similarly, it may
be useful to explicitly display the parametric dependence of solutions. In this case
we would write x(t,t0 , x0 ; μ ) or, if we were not interested in the initial condition,
x(t; μ ).
Ordinary differential equations that depend explicitly on time (i.e., ẋ = f (x,t; μ ))
are referred to as nonautonomous or time-dependent ordinary differential equations
or vector fields, and ordinary differential equations that do not depend explicitly
on time (i.e., ẋ = f (x; μ )) are referred to as autonomous or time-independent ordi-
nary differential equations or vector fields. The same terminology may be used in
the same way for discrete-time systems. It should be noted that a nonautonomous
vector field or map can always be made autonomous by redefining time as a new in-
dependent variable. This is done as follows. For a vector field ẋ = f (x,t), by writing
it as
dx f (x,t)
= (2.3)
dt 1
20 2 Preliminaries of Nonlinear Dynamics and Chaos

and using the chain rule, we can introduce a new independent variable s so that (2.3)
becomes ⎧
⎪ dx
⎪
⎨ ≡ ẋ = f (x,t),
ds
(2.4)
⎪
⎩ dt ≡ t˙ = 1.
⎪
ds
If we define y = (x,t) and f˜(y) = ( f (x,t), 1)T , we see that (2.4) becomes
T

dy
= f˜(y), y ∈ Rn+1 .
ds
For the map x(k + 1) = f (x(k), k), if we define y(k) = (x(k), k)T and f˜(y) =
( f (x(k), k), k + 1)T , we get the autonomous system under the new phase space

y(k + 1) = f˜(y(k)), y ∈ Rn+1 .

So, it is generally sufficient to consider autonomous systems

ẋ = f (x), x ∈ Rn , (2.5)

and
x(k + 1) = g(x(k)), x ∈ Rn . (2.6)

2.3 Existence, Uniqueness, Flow, and Dynamical Systems

2.3.1 Existence and Uniqueness

Consider the autonomous vector field (2.5). Geometrically, x(t) is a curve in Rn

whose tangent vector ẋ(t) exists for all t in its dominant field J and equals f (x(t)).
For simplicity, we usually take initial time t0 = 0. The main problem in differential
equations is to find the solution for any initial value problem; that is, to determine the
solution of the system that satisfies the initial condition x(t0 ) = x0 for each x0 ∈ Rn .
Unfortunately, nonlinear differential equations may have no solutions satisfying
certain initial conditions.
Example 2.1 ([9]). Consider the following simple first-order differential equation:

1, if x < 0,
ẋ =
−1, if x ≥ 0.

This vector field on R points to the left when x ≥ 0 and to the right if x < 0. Conse-
quently, there is no solution that satisfies the initial condition x(0) = 0. Indeed, such
a solution must initially decrease since ẋ(0) = −1, but, for all negative values of x,
solutions must increase. This cannot happen. Note further that solutions are never
2.3 Existence, Uniqueness, Flow, and Dynamical Systems 21

defined for all time. For example, if x0 > 0, then the solution through x0 is given by
x(t) = x0 − t, but this solution is only valid for −∞ < t < x0 for the same reason as
above.
The problem in this example is that the vector field is not continuous at 0; when-
ever a vector field is discontinuous we face the possibility that nearby vectors may
point in ‘opposing’ directions, thereby causing solutions to halt at these points. 

Example 2.2. Consider the following differential equation:

ẋ = 3x2/3 .

The identically zero function u : R → R given by u(t) ≡ 0 is clearly a solution with

initial condition u(0) = 0. But u0 (t) = t 3 is also a solution satisfying this initial
condition. Moreover, for any τ > 0, the function given by

0, if t ≤ τ ,
uτ (t) =
(t − τ )3 , if t > τ

is also a solution satisfying the initial condition uτ (0) = 0. While the differential
equation in this example is continuous at x0 = 0, the problems arise because 3x2/3
is not differentiable at this point. 

From these two examples it is clear that, to ensure the existence and uniqueness
of solutions, certain conditions must be imposed on the function f . In the first ex-
ample, f is not continuous at the point 0, while, in the second example, f fails to be
differentiable at 0. It turns out that the assumption that f is continuously differen-
tiable is sufficient to guarantee both existence and uniqueness of the solution. In fact,
we can furthermore guarantee the existence and uniqueness under a weaker condi-
tion, called the Lipschitz condition, on f . We now state several qualitative theorems
about the solutions of system (2.5) [7].

Theorem 2.1 (Local Existence and Uniqueness). Let U ⊂ Rn be an open subset

of real Euclidean space (or of a differentiable manifold M 1 ), let x0 ∈ U, and let
f : U → Rn be a (locally) Lipschitzian map, i.e.,

f (y) − f (x) ≤ K x − y

for some K < ∞. Then, there are some constant c > 0 and a unique solution
x(·, x0 ) : (−c, c) → U satisfying the differential equation described by (2.5) with
initial condition x(0) = x0 . 

1 Roughly speaking, a manifold is a set which locally has the structure of Euclidean space. In ap-
plications, manifolds are most often met as m-dimensional surfaces embedded in Rn . If the surface
has no singular points, i.e., the derivative of the function representing the surface has maximal
rank, then by the implicit function theorem it can locally be represented as a graph. The surface is
a Cr manifold if the (local) coordinate charts representing it are Cr .
22 2 Preliminaries of Nonlinear Dynamics and Chaos

The local existence theorem becomes global in all cases when we work on com-
pact manifolds2 M instead of open spaces like Rn .
Theorem 2.2 (Global Existence). The differential equation ẋ = f (x), x ∈ M, with
M compact, and f ∈ C1 , has solution curves defined for all t ∈ R. 
The local theorem can be extended to show that solutions depend in a ‘nice’
manner on initial conditions.
Theorem 2.3 (Dependence on Initial Value). Let U ∈ Rn be open and suppose that
f : U → Rn has a Lipschitz constant K. Let y(t), z(t) be solutions of ẋ = f (x) on the
closed interval [t0 ,t1 ]. Then, for all t ∈ [t0 ,t1 ],

y(t) − z(t) ≤ y(t0 ) − z(t0 ) eK(t−t0 ) . 

2.3.2 Flow and Dynamical Systems

If x(t) is a solution of (2.5), then x(t + τ ) is also a solution for any τ ∈ R. So, it
suffices to choose a fixed initial time, say, t0 = 0, which is understood and therefore
often omitted from the notation. If we denote by φt (x) = φ (t, x) the state in Rm
reached by the system at time t starting from x, then the totality of solutions of (2.5)
can be represented by a one-parameter family of maps φ t : U → Rm satisfying

d
[φ (t, x)] = f [φ (τ , x)]
dt t=τ

for all x ∈ U and for all τ ∈ I for which the solution is defined. The family of
maps φt (x) = φ (t, x) is called the flow (or the flow map) generated by the vector
field f . The set of points {φ (t, x0 ) : t ∈ I} defines an orbit of (2.5), starting from
a given point x0 . It is a solution curve in the state space, parameterized by time.
The set {[t, φ (t, x0 )] : t ∈ I} is a trajectory of (2.5) and it evolves in the space of
motions. However, in applications, the terms ‘orbit’ and ‘trajectory’ are often used
as synonyms. A simple example of a trajectory in the space of motions R × R2
and the corresponding orbit in the state space R2 are given in Fig. 2.1. Clearly,
the orbit is obtained by projecting the trajectory on to the state space. The flows
generated by vector fields form a very important subset of a more general class of
maps, characterized by the following definition.
Definition 2.1. A flow is a map φ : I ⊂ R × X → X where X is a metric space, that
is, a space endowed with a distance function, and φ has the following properties:
2 A compact manifold is a manifold that is compact as a topological space, such as the circle
(the only one-dimensional compact manifold) and the n-dimensional sphere and torus. For many
problems in topology and geometry, it is convenient to study compact manifolds because of their
‘nice’ behavior. Among the properties making compact manifolds ‘nice’ are the facts that they
can be covered finitely by many coordinate charts and that any continuous real-valued function is
bounded on a compact manifold.
2.3 Existence, Uniqueness, Flow, and Dynamical Systems 23

(i) φ (0, x) = x for every x ∈ X (identity axiom);

(ii) φ (t + s, x) = φ (s, φ (t, x)) = φ (t, φ (s, x)) = φ (s + t, x), that is, time-translated
solutions remain solutions;
(iii) for fixed t, φ t := φ (t, ·) is a homeomorphism3 of the phase space on X. 

Remark 2.1. A distance on a space X (or, a metric on X) is a function d(·, ·) : X ×

X → R+ satisfying the following properties for all x, y ∈ X:
(i) d(x, y) ≥ 0 and d(x, y) = 0 if and only if x = y;
(ii) d(x, y) = d(y, x) (symmetry);
(iii) d(x, y) ≤ d(x, z) + d(z, y) (triangle inequality).
Notice that there also exist notions of distance which are perfectly meaningful but
do not satisfy the definition above and therefore do not define a metric, for example
the distance between a point and a set A

d(x, A) = inf d(x, y)

y∈A

and the distance between two sets A and B

d(A, B) = inf inf d(x, y). 

x∈A y∈B

In the following, we give a formal definition of a ‘dynamical system.’

Definition 2.2. A dynamical system is a triplet {T, X, φ t } where T is a time set, X

is a state space, and φ t : X → X is a flow parameterized by t ∈ T . 

y y

time

x
(a) (b) x

Fig. 2.1 A damped oscillator in R2 : (a) space of motions; (b) state space

3h : U ⊂ Rn → V ⊂ Rm is said to be a Cr diffeomorphism if both h and h−1 are Cr . h is called a

homeomorphism if r = 0.
24 2 Preliminaries of Nonlinear Dynamics and Chaos

2.4 Equilibrium, Periodic Orbit, Quasiperiodic Orbit, and

Poincaré Map

2.4.1 Equilibrium of Continuous-Time Systems

Definition 2.3 (Equilibrium of Autonomous Systems). An equilibrium solution of

(2.5) is a point x¯ ∈ Rn such that
f (x)
¯ = 0,
i.e., a solution which does not change in time. Other terms often substituted for the
term ‘equilibrium solution’ are ‘fixed point,’ ‘stationary point,’ ‘rest point,’ ‘singu-
larity,’ ‘critical point,’ or ‘steady state.’ 
Remark 2.2. What about the notion of equilibria of nonautonomous vector fields?
We should note that ideas developed for autonomous systems can lead to incor-
rect results for nonautonomous systems. For example, consider the following one-
dimensional nonautonomous field:

ẋ = −x + t. (2.7)

The solution through the point x0 at t = 0 is given by

x(t) = t − 1 + e−t (x0 + 1),

from which it is clear that all solutions asymptotically approach the solution t − 1 as
t → ∞. The frozen time or ‘instantaneous’ fixed points for (2.7) are given by

x = t.

At a fixed t, this is the unique point where the vector field is zero. However, x = t is
not a solution of (2.7). This is different from the case of an autonomous vector field
where a fixed point is a solution of the vector field. 
Definition 2.4. Consider the following nonautonomous system:

ẋ = f (t, x), (2.8)

where f : [0, ∞) × D → Rn is piecewise continuous in t and locally Lipschitz in x on

[0, ∞) × D, and D ⊂ Rn is a domain that contains the origin x = 0. The origin is an
equilibrium point of (2.8) at t = 0 if

f (t, 0) = 0, ∀t ≥ 0. 

For the discrete-time system (2.6), an equilibrium solution is the point x¯ ∈ Rn

such that
x¯ = g(x).
¯
An equilibrium solution of a discrete-time system is usually called a fixed point.
2.4 Equilibrium, Periodic Orbit, Quasiperiodic Orbit, and Poincaré Map 25

2.4.2 Periodic and Quasiperiodic Orbits

Consider again the basic system of differential equation

ẋ = f (x) (2.9)

and the derived flow φ . A solution φ (t, x∗ ) of system (2.9) through a point x∗ is said
to be periodic with period T > 0 if φ (T, x∗ ) = x∗ . The set L0 = {φ (t, x∗ ) : t ∈ [0, T )}
is a closed curve in the state space and is called a periodic orbit or cycle. T is called
the period of the cycle and measures its time length. It should be emphasized that
isolated periodic solutions are possible only for nonlinear differential equations.
Moreover, a limit cycle4 can be structurally stable in the sense that, if it exists for
a given system of differential equations, it will persist under a slight perturbation
of the system in the parameter space. On the contrary, linear systems of differential
equations in Rm (m ≥ 2) may have a continuum of periodic solutions characterized
by a pair of purely imaginary eigenvalues (the case of a center, which will be in-
troduced later). But, these periodic solutions can be destroyed by arbitrarily small
perturbations of the coefficients. In other words, these periodic solutions are not
structurally stable.
For the discrete-time system of xk+1 = G(xk ), an n-periodic orbit is defined as
the set of points L0 = {x0 , x1 , . . . , xn−1 } with xi = x j (i = j) such that

x1 = G(x0 ), x2 = G(x1 ), . . . , xn−1 = G(xn−2 ), x0 = G(xn−1 ).

It should be noted that each point in an n-periodic orbit is an n-periodic point since,
for k = 0, . . . , n − 1,

xk = Gn (xk ) and G j (xk ) = xk for 0 < j < n.

Periodic orbits of continuous-time systems and discrete-time systems are illus-

trated in Fig. 2.2.
To illustrate what quasiperiodic orbits are we will consider two examples, one
for discrete time and one for continuous time.
Example 2.3. Consider the following unit-circle map:

Mc : S1 → S1 , zn+1 = Mc (zn ) = czn , (2.10)

where zn = ei2πθn , θn ∈ R, α is a positive constant, and c = ei2πα . The map (2.10)

describes an anticlockwise jump of a particle on the unit circle S1 . The length of
the circular arc between two adjacent jump points is α . If α is rational, that is,
α = p/q with p and q integers, then any (initial) point on S1 is a q-periodic point
of the map Mc . If α is irrational, at each iteration a new point is added on the unit

4 A limit cycle is an isolated periodic solution of an autonomous system. The points on the limit
cycle constitute the limit set, which is the set of points in the state space that a trajectory repeatedly
visits. A limit set is only defined for discrete-time or continuous-time autonomous systems.
26 2 Preliminaries of Nonlinear Dynamics and Chaos

G ( x0 )
2

G( x0 )
x0 x0

n -1
G ( x0 )
L0 L0

(a) (b)

Fig. 2.2 Periodic orbits in: (a) a continuous-time system; (b) a discrete-time system

circle. For any jump point there exist other jump points arbitrarily close to it, but no
point is revisited in finite time. This means that the map is topologically transitive
(see Definition 2.12). However, points that are close to each other will remain close
under the iteration. 

Example 2.4. The simplest example of quasiperiodic motion in continuous time is a

system defined by a pair of oscillators of the form

ẍ + ω12 x = 0, ÿ + ω22 y = 0,

where x, y ∈ R and ω1 and ω2 are real constants. The above system can be rewritten
in the form of first-order linear differential equations in R4 :
⎧
⎪
⎪ ẋ1 = −ω1 x2 ,
⎨
ẋ2 = ω1 x1 ,
(2.11)
⎪
⎪ ẏ1 = −ω2 y2 ,
⎩
ẏ2 = ω2 y1 ,

where x2 = x and y2 = y. Transforming the variables x1 , x2 and y1 , y2 into polar

coordinates, system (2.11) can be written as
⎧
⎪
⎪ θ̇ = ω1 ,
⎨ 1
ṙ1 = 0,
(2.12)
⎪
⎪ θ̇ = ω2 ,
⎩ 2
ṙ2 = 0,

where θi and ri (i = 1, 2) denote the angle and the radius, respectively. We can see
that the above equations describe a particle rotating on a two-dimensional torus for
a given pair (r1 , r2 ), ri > 0 (i = 1, 2) (see Fig. 2.3). There are two basic possibilities
for the motion:
(i) ω1 /ω2 is a rational number, in which case there exists a continuum of periodic
orbits of period q;
2.4 Equilibrium, Periodic Orbit, Quasiperiodic Orbit, and Poincaré Map 27

(ii) ω1 /ω2 is an irrational number, in which case the orbit starting from any initial
point on the torus wanders on it, getting arbitrarily near any other point, with-
out ever returning to that exact initial point. The flow generated by (2.12) is
topologically transitive on the torus (see Fig. 2.4).
In both cases, points that are close to each other remain close under the action of the
flow. 

A general definition of quasiperiodicity of an orbit as a function of time can be

given as follows:
Definition 2.5 ([11], p. 128). A function h : R → Rm is called quasiperiodic if it can
be written in the form of h(t) = H(ω1t, ω2t, . . . , ωm t), where H is periodic of period
2π in each of its arguments, and two or more of the m (positive) frequencies are
incommensurable. 

θ1
θ2

Fig. 2.3 The motion on two-dimensional torus

(a) (b)

Fig. 2.4 The quasiperiodic motions with different evolution times

28 2 Preliminaries of Nonlinear Dynamics and Chaos

2.4.3 Poincaré Map

A Poincaré map is a classical technique for analyzing dynamical systems. It re-

places the flow of an nth-order continuous-time system with an (n − 1)st-order
discrete-time system. The definition of the Poincaré map ensures that the limit sets
of the discrete-time system correspond to the limit sets of the underlying flow. The
Poincaré map’s usefulness lies in the reduction of order and the fact that it bridges
the gap between continuous-time and discrete-time systems.
The definitions of a Poincaré map are different for autonomous systems and
nonautonomous systems. We present the two cases separately.
Case 1: In this case we consider an nth-order time-periodic nonautonomous sys-
tem ẋ = f (t, x), with period T . We can convert it into an (n + 1)st-order au-
tonomous system by appending an extra state θ := 2π t/T. Then, the autonomous
system is given by 
ẋ = f (x, θ T /2π ), x(t0 ) = x0 ,
(2.13)
θ̇ = 2π /T, θ (t0 ) = 2π t0 /T.
Since f is time periodic with period T , system (2.13) is periodic in θ with period
2π . Hence, the planes θ = 0 and θ = 2π may be identified and the state space
transformed from the Euclidean space Rn+1 to the cylindrical space Rn × S1 ,
where S1 is the unit circle. The solution of (2.13) in the cylindrical state space is

x(t) φt (x0 ,t0 )

= , (2.14)
θ (t) 2π t/T mod 2π

where the modulo function restricts to 0 ≤ θ < 2π . Consider the n-dimensional

hyperplane Σ ∈ Rn × S1 defined by

Σ := {(x, θ ) ∈ Rn × S1 : θ = θ0 }.

Every T seconds, the trajectory of (2.14) intersects Σ (see Fig. 2.5). The resulting
map PN : Σ → Σ (Rn → Rn ) is defined by

PN (x) := φt0 +T (x,t0 ).

PN is called the Poincaré map of the nonautonomous system.

Case 2: Consider an nth-order autonomous system with a limit cycle Γ shown in
Fig. 2.6. Let x0 be a point on the limit cycle and let Σ be an (n − 1)-dimensional
hyperplane transversal to Γ at x0 . The trajectory emanating from x0 will hit Σ
at x0 in T seconds, where T is the minimum period of the limit cycle. Due to
the continuity of φt with respect to the initial condition, the trajectory starting on
Σ in a sufficiently small neighborhood of x0 will, in approximately T seconds,
intersect Σ in the vicinity of x0 . Hence, φt and Σ define a mapping PA of some
neighborhood U ⊂ Σ of x0 on to another neighborhood V ⊂ Σ of x0 . PA is a
Poincaré map of the autonomous system.
2.5 Invariant and Attracting Sets 29

θ0

PN (x)

Fig. 2.5 The Poincaré map of a first-order nonautonomous system

x
PA(x ) x0
Σ

Fig. 2.6 The Poincaré map of a third-order autonomous system

Remark 2.3.
(i) PA is defined locally, i.e., in a neighborhood of x0 . Unlike the nonautonomous
case, it is not guaranteed that the trajectory emanating from any point on Σ will
intersect Σ .
(ii) For a Euclidean state space, the point PA (x) is not the first point where φt (x)
intersects Σ ; φt (x) must pass through Σ at least once before returning to V . This
is in contrast with the cylindrical state space in Fig. 2.5.
(iii) PA is a diffeomorphism and is, therefore, invertible and differentiable [12]. 

2.5 Invariant and Attracting Sets

Definition 2.6 (Invariant Set). Let S ⊂ Rn be a set. Then,

(i) (Continuous time) S is said to be invariant under the vector field ẋ = f (x) if for
any x0 ∈ S we have x(t, 0, x0 ) ∈ S for all t ∈ R, where x(0, 0, x0 ) = x0 .
30 2 Preliminaries of Nonlinear Dynamics and Chaos

(ii) (Discrete time) S is said to be invariant under the map xk+1 = g(xk ) if for any
x0 ∈ S we have gn (x0 ) ∈ S for all n ∈ Z.
If we restrict ourselves to positive time (i.e., t ≥ 0, and n ≥ 0), then we refer to S as
a positively invariant set, while, for negative time, as a negatively invariant set. 
The definition means that trajectories starting in the invariant set remain in the
invariant set, for all of their future and all of their past.
Definition 2.7. An invariant set S ⊂ Rn is said to be a Cr (r ≥ 1) invariant manifold
if S has the structure of a Cr differentiable manifold. Similarly, a positively (nega-
tively) invariant set S ⊂ Rn is said to be a Cr (r ≥ 1) positively (negatively) invariant
manifold if S has the structure of a Cr differentiable manifold. 
Definition 2.8. Let φ (t, x) be a flow on a metric space M. Then, a point y ∈ M is
called an ω -limit point of x ∈ M for φ (t, x) if there exists an infinitely increasing
sequence {ti } such that
lim d(φ (ti , x), y) = 0.
i→∞

The set of all ω -limit points of x for φ (t, x) is called the ω -limit set and is denoted
by ω (x). 
The definitions of α -limit point and α -limit set of a point x ∈ M are obtained just
by taking sequences ti decreasing in i to −∞. The α -limit set of x is denoted as α (x).
Definition 2.9. A point x0 is called nonwandering if the following condition holds.
Flows: for any neighborhood U of x0 and T > 0, there exists some |t| > T such that

φ (t,U) ∩U = 0;
/

Maps: for any neighborhood U of x0 , there exists some n = 0 such that

gn (U) ∩U = 0.
/

The set of all nonwandering points of a flow or map is called the nonwandering set
of that particular flow or map. 
Definition 2.10. A closed invariant set A ⊂ Rn is called an attracting set if there is
some neighborhood U of A such that
Flows: for any t ≥ 0, φ (t,U) ⊂ U and t>0 φ (t,U) = A;
Maps: for any n ≥ 0, gn (U) ⊂ U and n>0 gn (U) = A. 
Definition

2.11. The basin of attraction of an attracting set A is given by
Flows: t≤0 φ (t,U);

Maps: n≤0 gn (U);
where U is any open set satisfying Definition 2.10. 
Definition 2.12. A closed invariant set A is said to be topologically transitive if, for
any two open sets U,V ⊂ A,
Flows: there exists a t ∈ R such that φ (t,U) ∩V = 0;/
Maps: there exists an n ∈ Z such that gn (U) ∩V = 0./ 
2.6 Continuous-Time Systems in the Plane 31

Definition 2.13. An attractor is a topologically transitive attracting set. 

2.6 Continuous-Time Systems in the Plane

In this section and the next two sections we will discuss the types of equilibrium
points of planar systems of continuous time and discrete time, respectively. In ap-
plications, we very often encounter linear systems described by two first-order dif-
ferential equations (or a differential equation of second order), either because the
underlying model is linear or because it is linearized around an equilibrium point.
Systems in two-dimensional space are particularly easy to discuss in full detail and
give rise to a number of interesting basic dynamic configurations. Moreover, in prac-
tice, it is very difficult or impossible to determine the exact values of the eigenvalues
and eigenvectors for matrices of order greater than two. Thus, one can draw inspi-
ration from the discussion about planar systems when studying high-dimensional
systems.
The general form of a continuous-time planar system can be written as

ẋ x a11 a12 x
=A = , (2.15)
ẏ y a21 a22 y

where x, y ∈ R and ai j are real constants. If det(A) = 0, the unique equilibrium, for
which ẋ = ẏ = 0, is x = y = 0. The characteristic equation is

λ 2 − tr(A)λ + det(A) = 0,

and the eigenvalues are

1 √ 
λ1,2 = tr(A) ± Δ ,
2
where Δ ≡ (tr(A))2 − 4det(A) is called the discriminant. For system (2.15) the dis-
criminant is
Δ = (a11 − a22)2 + 4a12a21 .
The different types of dynamical behavior of (2.15) can be described in terms of
the two eigenvalues of the matrix A, which in planar systems can be completely
characterized by the trace and determinant of A. In the following we consider non-
degenerate equilibria for which λ1 and λ2 are both nonzero, when there is no explicit
claim. We distinguish behaviors according to the sign of the discriminant.
Case 1: Δ > 0. Eigenvalues and eigenvectors are real. Solutions have the form

x(t) = c1 eλ1t u1 + c2 eλ2t u2 ,

(1) (1)

y(t) = c1 eλ1t u1 + c2 eλ2t u2 ,

(2) (2)
32 2 Preliminaries of Nonlinear Dynamics and Chaos
y y

x x

(a) (b)

y y

x x

(c) (d)
y
y

x x

(e) (f)

Fig. 2.7 Equilibrium types in the plane

   
(1) (2) T (1) (2) T
where u1 = u1 , u1 and u2 = u2 , u2 are eigenvectors corresponding
to the eigenvalues λ1 and λ2 , respectively. We have three basic subcases corre-
sponding to Fig. 2.7 (a), (b), and (e), respectively (eigenvalues are plotted in the
complex plane).
(i) tr(A) < 0, det(A) > 0. In this case, eigenvalues and eigenvectors are real
and both eigenvalues are negative (say, 0 > λ1 > λ2 ). The two-dimensional
2.6 Continuous-Time Systems in the Plane 33

y y

x x

(a) (b)

Fig. 2.8 Equilibrium types in the plane with repeated eigenvalue: (a) bicritical node; (b) Jordan
node

state space coincides with the stable eigenspace.5 The equilibrium is called
a stable node, and the term ‘node’ refers to the characteristic shape of the
ensemble of orbits around the equilibrium.
(ii) tr(A) > 0, det(A) > 0. In this case, eigenvalues and eigenvectors are real, both
eigenvalues are positive (say, λ1 > λ2 > 0), and the state space coincides with
the unstable eigenspace. The equilibrium is called an unstable node.
(iii) det(A) = 0. In this case, Δ > 0 independent of the sign of the trace of A. One
eigenvalue is positive, and the other is negative (say, λ1 > 0 > λ2 ). There are,
then, a one-dimensional stable eigenspace and a one-dimensional unstable
eigenspace and the equilibrium is known as a saddle point.
Case 2: Δ < 0. The eigenvalues and eigenvectors are complex conjugate pairs and
we have
(λ1 , λ2 ) = (λ , λ̄ ) = α ± iβ
with
1 1√
α = tr(A), β = −Δ .
2 2
The solutions have the form

x(t) = Ceα t cos(β t + φ ),
y(t) = Ceα t sin(β t + φ ),

and the motion is oscillatory. If α = 0 there is no strict periodicity in the sense

that there exists no τ such that x(t) = x(t + τ ). However, a conditional period can
be defined as the length of time between two successive maxima of a variable,
which is equal to 2π /β . The frequency is simply the number of oscillations per
5 An eigenspace is spanned by eigenvectors. A stable eigenspace is spanned by the eigenvectors
corresponding to negative eigenvalues, and an unstable eigenspace is spanned by the eigenvectors
corresponding to positive eigenvalues.
34 2 Preliminaries of Nonlinear Dynamics and Chaos

time unit, that is, β /2π . The amplitude or size of the oscillations depends on
the initial condition and eα t (more on this point below). There are three subcases
depending on the sign of tr(A) and therefore of Re(λ ) = α ; see the corresponding
illustrations in Figs. 2.7 (c), (d), and (f), respectively.
(i) tr(A) < 0, Re(λ ) = α < 0. The oscillations are dampened and the system
converges to the equilibrium. The equilibrium point is known as a focus or,
sometimes, a vortex, due to the characteristic shape of the orbits around
the equilibrium. In this case the focus or vortex is stable and the stable
eigenspace coincides with the state space.
(ii) tr(A) > 0, Re(λ ) = α > 0. The amplitude of the oscillations gets larger with
time and the system diverges from the equilibrium. The unstable eigenspace
coincides with the state space and the equilibrium point is called an unstable
focus or vortex.
(iii) tr(A) = 0, Re(λ ) = α = 0. In this special case we have a pair of purely imag-
inary eigenvalues. Orbits neither converge to, nor diverge from, the equilib-
rium point, but they oscillate regularly around it with a constant amplitude
that depends only on initial conditions and the equilibrium point is called a
center.
Case 3: Δ = 0. The eigenvalues are real and equal to each other, λ1 = λ2 = λ .
In this case, if A = λ I, only one eigenvector can be determined, say u =
 T
u(1) , u(2) , defining a single straight line through the origin. We can write the
general solution as

x(t) = (c1 u(1) + c2 v(1) )eλ t + tc2 u(1) eλ t ,
y(t) = (c1 u(2) + c2 v(2) )eλ t + tc2u(2) eλ t ,

with 
x(0) = c1 u(1) + c2v(1) ,
y(0) = c1 u(2) + c2 v(2) .
The equilibrium type is again a node, sometimes called a Jordan node. An ex-
ample of this type is provided in Fig. 2.8 (b), where it is obvious that there is
a single eigenvector. If A = λ I the equilibrium is still a node, sometimes called
a bicritical node. However, all half-lines from the origin are solutions, giving a
star-shaped form (see Fig. 2.8 (a)).
Fig. 2.9 provides a very useful geometric representation in the (tr(A), det(A))
plane of the various cases discussed above. Quadrants III and IV of the plane cor-
respond to saddle points, quadrant II to stable nodes and foci, and quadrant I to
unstable nodes and foci. The parabola divides quadrants I and II into nodes and
foci (the former below the parabola and the latter above it). The positive part of the
det(A) axis corresponds to centers.
The analysis of systems with n > 2 variables can be developed along the same
lines although geometric insight will fail when the dimension of the state space is
larger than three. In order to give the reader a broad idea of common situations we
2.6 Continuous-Time Systems in the Plane 35

det (A)

tr (A)

Fig. 2.9 Continuous-time dynamics in R2

depict sample orbits of three-dimensional systems in Fig. 2.10, which indicates the
corresponding eigenvalues in the complex plane. The system in Fig. 2.10 (a) has two
positive real eigenvalues and one negative real eigenvalue. The equilibrium point is
an unstable saddle. In this case the plane associates with the positive real eigen-
values. All orbits eventually converge to the unstable eigenspace and are captured
by the expanding dynamics. The only exceptions are those orbits initiating on the
stable eigenspace (defined by the eigenvector associated with the negative eigen-
value) which converge to the plane at the equilibrium point. When A is an m × m
matrix, we divide the eigenvectors (or, in the complex case, the vectors equal to the
real and imaginary parts of them) into three groups, according to whether the cor-
responding eigenvalues have negative, positive, or zero real parts. Then, the subsets
of the state space spanned (or generated) by each group of vectors are known as the
stable, unstable, and center eigenspaces, respectively, and denoted by E s , E u , and
E c . Notice that the term saddle in Rm refers to all cases in which there exist some
eigenvalues with positive and some with negative real parts. We use the term saddle
node when eigenvalues are all real, saddle focus when some of the eigenvalues are
complex. An example of the latter is presented in Fig. 2.10 (b). The two real vectors
associated with the complex conjugate pair of eigenvalues, with negative real parts,
span the stable eigenspace. Orbits approach asymptotically the unstable eigenspace
defined by the eigenvector associated with the positive eigenvalue, along which the
dynamics is explosive.
36 2 Preliminaries of Nonlinear Dynamics and Chaos

(a)
z

(b)

Fig. 2.10 Continuous-time dynamics in R3 : (a) saddle node; (b) saddle focus
2.7 General Solutions of Discrete-Time Linear Systems 37

2.7 General Solutions of Discrete-Time Linear Systems

Now consider the following linear, discrete-time system described by

x(n + 1) = Bx(n), x ∈ Rm . (2.16)

It is easy to see that x = 0 is the only equilibrium solution. If κi is a real, distinct

eigenvalue of the m × m matrix B and vi is the corresponding real eigenvector so that
Bvi = κi vi , it can be verified that

x(n) = κin vi (2.17)

is a solution of (2.16). Suppose that we have a pair of eigenvalues of B

(κ j , κ j+1 ) = (κ j , κ̄ j ) = σ j ± iθ j

with a corresponding pair of eigenvectors

(v j , v j+1 ) = (v j , v¯j ) = p j ± iq j ,

where p j and q j are m-dimensional vectors. Then, the pair of functions

⎧
⎪ 1 n 
⎨ x j (n) = κ v j + κ̄ njv¯j = rnj [p j cos(ω j n) − q j sin(ω j n)],
2 j (2.18)
 
⎩ x j+1 (n) = − i κ n v j − κ̄ nv¯j = rn [p j sin(ω j n) + q j cos(ω j n)]
⎪
j j j
2
are the jth and ( j + 1)st solutions of (2.16), respectively. Here, we have used the
polar coordinate transformations

σ j = r j cos ω j ,
(2.19)
θ j = r j sin ω j ,

and a well-known result

(cos ω ± i sin ω )n = cos(ω n) ± isin(ω n).


From (2.19) we have r j = σ 2j + θ j2 . Then, r j is simply the modulus of the complex
(l) (l)
eigenvalues. If p j and q j denote the lth elements of p j and q j , respectively, then
in polar coordinates we have
(l) (l) (l)
p j = C j cos(φ j ),
(l) (l) (l)
q j = C j sin(φ j ),

where l = 1, 2, . . . , m. (2.18) can be rewritten as m-dimensional vectors whose lth

elements have the form
38 2 Preliminaries of Nonlinear Dynamics and Chaos

(l) (l) (l)

x j (n) = C j rnj cos(ω j n + φ j ),
(l) (l) (l)
x j+1 (n) = C j rnj sin(ω j n + φ j ).

Assuming that we have m linearly independent solutions defined by (2.17) and

(2.18), by the superposition principle the general solution of (2.16) can be written
as a linear combination of the individual solutions, namely

x(n) = c1 x1 (n) + c2 x2 (n) + · · · + cm xm (n),

where ci are constants depending on the initial conditions.

When eigenvalues are repeated, the general solution becomes

h n j −1
x(n) = ∑ ∑ k jl nl κ nj,
j=1 l=0

where n j ≥ 1 is the multiplicity of the jth eigenvalue, h ≤ m is the number of distinct

eigenvalues, and k jl are independent vectors whose values depend on the m arbitrary
initial conditions.
Inspection of (2.17) and (2.18) indicates that if the modulus of any of the eigen-
values is greater than 1, solutions tend to +∞ or −∞ as time goes to +∞. On the
contrary, if all eigenvalues have modulus smaller than 1, solutions converge asymp-
totically to the equilibrium point. Analogous to the situation of continuous time, we
call the space spanned by the eigenvectors whose corresponding eigenvalues have
modulus less than 1 (greater than 1) a stable eigenspace (unstable eigenspace) and
call the space spanned by the eigenvectors whose corresponding eigenvalues have
modulus equal to 1 a center eigenspace. We also use the same symbols, E s , E u , and
E c , to denote them.

2.8 Discrete-Time Systems in the Plane

The discrete-time autonomous system that is analogous to the continuous-time sys-

tem (2.15) is described by

xn+1 xn b11 b12 xn

=B = . (2.20)
yn+1 yn b21 b22 yn

We assume that the matrix I − B is nonsingular. Thus, the origin is the unique equi-
librium point of (2.20). The characteristic equation is analogous to the continuous
case as well,
κ 2 − tr(B)κ + det(B) = 0,
and the eigenvalues are given by
2.8 Discrete-Time Systems in the Plane 39

1
κ1,2 = tr(B) ± tr(B)2 − 4det(B) .
2

We also assume that the equilibria of (2.20) are nondegenerate, i.e., |κ1 |, |κ2 | = 1.
We will discuss the dynamics of the discrete-time system (2.20) for the following
three cases.
Case 1: Δ > 0. The eigenvalues are real and solutions take the form
⎧
⎨ x(n) = c1 κ n v(1) + c2 κ n v(1) ,
1 1 2 2
⎩ y(n) = c κ n v(2) + c κ n v(2) .
1 1 1 2 2 2

(i) If |κ1 | < 1 and |κ2 | < 1 the fixed point is a stable node. This means that
solutions are sequences of points approaching the equilibrium as n → +∞.
If κ1 , κ2 > 0 the approach is monotonic; otherwise, there are improper os-

1
1
2 2
3 3
1 4
5
3 2 6
7 1
2
8
3

(a) (b)

1 2 12
1 4 6 8 10
3
6
2 5 3
4 4
6 1 5
7 2
3 9
4
2 11
1
3
5

(c) (d)

Fig. 2.11 Phase diagrams for real eigenvalues: (a) and (c) stable nodes; (b) and (d) saddle points
40 2 Preliminaries of Nonlinear Dynamics and Chaos

cillations6 (see Figs. 2.11 (a) and (c), respectively). In this case, the stable
eigenspace coincides with the state space.
(ii) If |κ1 | > 1 and |κ2 | > 1 the fixed point is an unstable node. In this case,
solutions are sequences of points approaching equilibrium as n → −∞. If
κ1 , κ2 > 0 the approach is monotonic; otherwise, there are improper oscil-
lations (as in Figs. 2.11 (a) and (c), respectively, but arrows point in the
opposite direction and the time order of points is reversed). In this case, the
unstable eigenspace coincides with the state space.
(iii) If |κ1 | > 1 and |κ2 | < 1 the fixed point is a saddle point. No sequences of
points approach the equilibrium for n → ±∞ except for those originating
from points on the eigenvectors associated with κ2 . Again, if κ1 , κ2 > 0
orbits move monotonically (see Fig. 2.11 (b)); otherwise they oscillate im-
properly (see Fig. 2.11 (d)). The stable and unstable eigenspaces are one
dimensional.
Case 2: Δ < 0. In this case, det(B) > 0. Eigenvalues are a complex conjugate pair
given by
(κ1 , κ2 ) = (κ , κ̄ ) = σ ± iθ
and solutions are sequences of points situated on a spiral whose amplitude in-
creases or decreases in time according to the factor rn , where

r = |σ ± iθ | = σ 2 + θ 2 = det(B)

is the modulus of the complex eigenvalue pair. Solutions are of the form

x(n) = Crn cos(ω n + φ ),
y(n) = Crn sin(ω n + φ ).

(i) If r < 1 solutions converge to equilibrium and the equilibrium point is a

stable focus (see Fig. 2.12 (a)).
(ii) If r > 1 solutions diverge and the equilibrium point is an unstable focus (as
in Fig. 2.12 (a), but arrows point in the opposite direction and the time order
of points is reversed).
(iii) If r = 1 the eigenvalues lie exactly on the unit circle, an exceptional case.
There are two subcases which depend on the frequency of the oscillation
ω /2π , ω = arccos[tr(B)/2]:
a. ω /2π is rational and the orbit in the state space is a periodic sequence
of points situated on a circle, the radius of which depends on initial con-
ditions (see Fig. 2.12 (b));
b. ω /2π is irrational and the sequence is nonperiodic or quasiperiodic, that
is, starting from any point on the circle, orbits stay on the circle but no
6 First-order, discrete-time equations (where the order is determined as the difference between the

extreme time indices) can also have fluctuating behavior, called improper oscillations, owing to the
fact that if their eigenvalue β < 0, β n will be positive or negative according to whether n is even or
odd. The term improper refers to the fact that in this case oscillations of variables have a ‘kinky’
form that does not properly describe the smoother ups and downs of real variables.
2.8 Discrete-Time Systems in the Plane 41

1
5

1
3 1

2 5

6 4
4 5
2
2 3 4

3
(a) (b)
5
10 9
4

1
6
8
3

2 7
(c)

Fig. 2.12 Phase diagrams for complex eigenvalues: (a) a stable focus; (b) periodic cycles; (c) a
quasiperiodic solution

sequence returns to the initial point in finite time. Therefore, solutions

wander on the circle filling it up, without ever becoming periodic (see
Fig. 2.12 (c)).
Case 3: Δ = 0. There is a repeated real eigenvalue κ1 = tr(B)/2. The general form
of solutions with a repeated eigenvalue κ is as follows:

x(n) = (c1 v(1) + c2u(1) )κ n + nc2 v(1) κ n ,
y(n) = (c1 v(2) + c2 u(2) )κ n + nc2 v(2) κ n .

If |κ | < 1, limn→∞ nκ n = 0. If the repeated eigenvalue is equal to 1 in absolute

value, the equilibrium is unstable (with improper oscillations for κ1 = −1). How-
ever, the divergence is linear not exponential.
The dynamics of the discrete case can be conveniently summarized by the di-
agram in Fig. 2.13. (For the sake of simplicity, we represent orbits as continuous
rather than dotted curves.) If we replace the greater-than sign with the equal sign
in conditions (i)–(iii) of Case 2, we obtain three lines intersecting each other in the
42 2 Preliminaries of Nonlinear Dynamics and Chaos

(tr(B), det(B)) plane, defining a triangle. Points inside the triangle correspond to
stable combinations of the trace and determinant of B.7 The parabola defined by

tr(B)2 = 4det(B)

divides the plane into two regions corresponding to real eigenvalues (below the
parabola) and complex eigenvalues (above the parabola). Combinations of trace and
determinant above the parabola but in the triangle lead to stable foci, combinations
below the parabola but in the triangle are stable nodes. All other combinations lead
to unstable equilibria.
Fig. 2.14 is an example of a system in R3 . There are a complex conjugate pair
with modulus less than one, and one dominant real eigenvalue greater than one. The
equilibrium point is a saddle focus.

det (B)

–2 –1 1 2
tr (B)

Fig. 2.13 Discrete-time dynamics in R2

7 In Case 2, if 1 + tr(B) + det(B) = 0 while (ii) and (iii) hold, one eigenvalue is equal to −1; if

1 − tr(B) + det(B) = 0 while (i) and (iii) hold, one eigenvalue is equal to +1; and if det(B) = 1
while (i) and (ii) hold, the two eigenvalues are a complex conjugate pair with modulus equal to 1.
2.9 Stabilities of Trajectories I: The Lyapunov First Method 43

7
6

5
2
4

3
1

Fig. 2.14 A discrete-time dynamics in R3

2.9 Stabilities of Trajectories I: The Lyapunov First Method

Stability theory plays a central role in systems theory and engineering. The so-called
Lyapunov first method or Lyapunov indirect method is used to study the stability of
a system’s trajectories through calculating the eigenvalues of a linearized system at
the objective trajectories. This means that the Lyapunov first method is essentially a
local method which can only be used in a neighborhood of the objective trajectory.

2.9.1 The Definition of Lyapunov Stability

Let x(t)
¯ be any solution of (2.5). Roughly speaking, x(t)
¯ is stable if solutions starting
‘close’ to x(t)
¯ at a given time remain close to x(t)¯ for all later times. It is asymp-
totically stable if nearby solutions not only stay close, but also converge to x(t)
¯ as
t → ∞.
44 2 Preliminaries of Nonlinear Dynamics and Chaos

Definition 2.14 (Lyapunov Stability). x(t) ¯ is said to be stable (or Lyapunov stable)
if, for any given ε > 0, there exists a δ = δ (ε ) > 0 such that, for any other solution,
¯ 0 ) − y(t0 ) < δ , we have x(t)
y(t), of (2.5) satisfying x(t ¯ − y(t) < ε for t > t0 ,
t0 ∈ R. 
A solution which is not stable is said to be unstable.
Definition 2.15 (Asymptotic Stability). x(t)
¯ is said to be asymptotically stable if it
is Lyapunov stable and, for any other solution, y(t), of (2.5), there exists a constant
¯ 0 ) − y(t0 ) < b, then limt→∞ x(t)
b > 0 such that, if x(t ¯ − y(t) = 0. 
A new stability definition which is different from Lyapunov’s definitions is given
as follows.
Definition 2.16. An orbit generated by system ẋ = f (x) (x ∈ Rn ), with initial condi-
tion x0 on a compact, φ -invariant subset A of the state space (i.e., φ (A) ⊂ A), is said
to be orbitally stable (asymptotically orbitally stable) if the invariant set

Γ = {φ (t, x0 ) : x0 ∈ A, t ≥ 0}

(the forward orbit of x0 ) is stable (asymptotically stable) according to Definition

2.14 (Definition 2.15). 
The analogous definitions of stability for autonomous dynamical systems in dis-
crete time with the general form

x(n + 1) = G(x(n)), x ∈ Rm (2.21)

are as follows.
Definition 2.17. The equilibrium point x¯ is Lyapunov stable (or, simply, stable) if,
for every ε > 0, there exists δ (ε ) such that

x0 − x¯ < δ (ε ) ⇒ Gn (x0 ) − x¯ < ε , ∀n > 0. 

Definition 2.18. The equilibrium point x¯ is asymptotically stable if

(i) it is stable and
(ii) ∃b > 0 such that

x0 − x¯ < b ⇒ lim Gn (x0 ) − x¯ = 0. 

n→∞

Property (ii) can be replaced by the following equivalent property:

(ii’) there exists b > 0 and, for each ε > 0, there exists an integer T = T (b, ε ) > 0
such that
x0 − x¯ < b ⇒ Gn (x0 ) − x¯ < ε , ∀n ≥ T.
Figs. 2.15 and 2.16 are the visualization of Definitions 2.14 and 2.15. Notice
that these two definitions imply that we have information on the infinite-time exis-
tence of solutions. This is obvious for equilibrium solutions but is not necessary for
2.9 Stabilities of Trajectories I: The Lyapunov First Method 45

y (0) y(0)
δ
x x
y(t) y(t)
b
ε

(a) (b)
Fig. 2.15 (a) Lyapunov stability; (b) asymptotic stability

x (t ) y(t )

(a)

t t0
x (t ) y(t )

(b)
t

Fig. 2.16 (a) Lyapunov stability; (b) asymptotic stability

nearby solutions. Also, these definitions are for autonomous systems, since in the
nonautonomous case it may be that δ and b depend explicitly on t0 .
In order to determine the stability of x(t)
¯ we must understand the nature of solu-
tions near x(t).
¯ Let
x(t) = x(t)
¯ + y. (2.22)
Substituting (2.22) into (2.5) and performing Taylor expansion about x(t)
¯ gives

˙¯ + ẏ = f (x(t))
ẋ = x(t) ¯ + D f (x(t))y
¯ + o( y 2), (2.23)

where D f is the derivative of f called the Jacobian matrix of f , · denotes a norm

on Rn , and o( y 2) denotes the higher-order infinitesimal term of y 2 . Using the
46 2 Preliminaries of Nonlinear Dynamics and Chaos

˙¯ = f (x(t)),
fact that x(t) ¯ (2.23) becomes

ẏ = D f (x(t))y
¯ + o( y 2). (2.24)

Equation (2.24) describes the evolution of orbits near x(t).

¯ For stability questions
we are concerned with the behavior of solutions arbitrarily close to x(t),
¯ so it seems
reasonable that this question could be answered by studying the associated linear
system
ẏ = D f (x(t))y.
¯ (2.25)
Usually it is difficult to determine the stability of x(t) by (2.25) since there are no
general analytical methods for finding the solution of linear ordinary differential
equations with time-dependent coefficients. However, if x¯ is an equilibrium solu-
tion, i.e., x(t)
¯ = x,¯ then D f (x(t))
¯ = D f (x)
¯ is a matrix with constant entries, and the
solution of (2.25) through the point y0 ∈ Rn at t = 0 can immediately be written as

y(t) = eD f (x)t
¯
y0 .

Thus, y(t) is asymptotically stable if all eigenvalues of D f (x)

¯ have negative real
parts.
Theorem 2.4. Suppose that all of the eigenvalues of D f (x)
¯ have negative real parts.
Then, the equilibrium solution x = x¯ of the nonlinear vector field (2.5) is asymptot-
ically stable. 
Definition 2.19. Let x = x¯ be a fixed point of ẋ = f (x), x ∈ Rn . Then, x¯ is called a
hyperbolic fixed point if none of the eigenvalues of D f (x)
¯ has zero real part. 
The eigenvalues of the Jacobian matrix D f (x)
¯ are also referred to as the eigen-
values of the fixed point x.
¯
Theorem 2.5 (Hartman–Grobman). If x¯ is a hyperbolic fixed point of (2.5), then
there is a homeomorphism h defined on some neighborhood N of x¯ in Rn , locally

Ws
Es

h(N)
N

u
W
Eu

Fig. 2.17 The Hartman–Grobman theorem

2.9 Stabilities of Trajectories I: The Lyapunov First Method 47

s
Ws E
u
E
s
E
u
W s
Ws E E
s

s Ws
E Ws

u u
E W Ws s
Ws E

(a) (b)

Fig. 2.18 Stable and unstable eigenspaces and manifolds in R2

mapping orbits of the nonlinear system (2.5) to those of the linear system (2.25).
The map h preserves the sense of orbits and can also be chosen so as to preserve
parameterization by time. 

If h is a homeomorphism, then from Theorem 2.5 we can deduce that asymp-

totic stability (or the lack of it) for the linear system (2.25) implies local asymptotic
stability of the nonlinear system (2.5) (or the lack of it). However, homeomorphic
equivalence does not preserve all the interesting geometric features of a dynamical
system. For example, a linear system characterized by an asymptotically stable node
is topologically conjugate to another linear system characterized by an asymptoti-
cally stable focus.
If the equilibrium point is not hyperbolic, that is to say, if there exists at least one
eigenvalue with real part exactly equal to 0, the Hartman–Grobman theorem cannot
be applied. The reason is that the linearized system is not sufficiently informative.
In particular, the stability properties of the system depend on the higher-order terms
of the expansion which have been ignored in the approximation (2.25).
In the above discussion of linear systems we emphasized the importance of cer-
tain invariant subspaces, i.e., the eigenspaces, defined by the eigenvectors of the
Jacobian matrix. If the nonlinear system (2.5) has an isolated,8 hyperbolic equilib-
rium x, ¯ in the neighborhood of x¯ there exist certain invariant surfaces, called stable
and unstable manifolds, which are the nonlinear counterparts of the stable and un-
stable eigenspaces. Locally, these manifolds are continuous deformations, respec-
tively, of the stable and unstable eigenspaces of the linear system (2.25) (because x¯
is hyperbolic, there is no center eigenspace for (2.25)) and they are tangents to the
eigenspaces of the linear system (2.25) at x. ¯ We denote stable manifolds, unstable
manifolds, and center manifolds by W s , W u , and W c . Some simple examples of the
phase diagrams of nonlinear systems and the corresponding linearized systems in
R2 and R3 are provided in Figs. 2.17, 2.18, and 2.19.

8 An equilibrium point is isolated if it has a surrounding neighborhood containing no other equi-

librium point.
48 2 Preliminaries of Nonlinear Dynamics and Chaos

Es Ws

Eu

u
W

Es Ws

(a)
Es
Ws

Es

Ws
Es Ws

(b)
Es Ws

Eu

u
W
Es Ws

(c)
Fig. 2.19 Stable and unstable eigenspaces and manifolds in R3
2.9 Stabilities of Trajectories I: The Lyapunov First Method 49

The method of linear approximation can be applied in a perfectly analogous man-

ner to nonlinear systems described by difference equations. Consider system (2.21),
with a fixed point x,
¯ and assume that G is differentiable. A local linear approxima-
tion of (2.21) near x¯ is
ξ (n + 1) = DG(x)¯ ξ (n), (2.26)
where ξ = x − x¯ and DG(x)¯ is the Jacobian matrix of partial derivatives of G, eval-
uated at x.
¯
The discrete-time version of the Hartman–Grobman theorem for x(n + 1) =
G(x(n)) is perfectly analogous to that for flows except for the following important
differences:
(i) For discrete-time systems, fixed points are hyperbolic if none of the eigenvalues
of the Jacobian matrix, evaluated at the equilibrium, is equal to 1 in modulus.
(ii) The map h of the Hartman–Grobman theorem defining the local relationship
between the nonlinear system (2.21) and the linearized system (2.26) is a dif-
feomorphism if the eigenvalues of DG(x) ¯ satisfy a nonresonance condition. In
the case of maps this condition requires that no eigenvalue κk of DG(x)¯ satisfies
m
κk = ∏ κici
i=1

for any choice of ci ≥ 0 with ∑i ci ≥ 2.

2.9.2 Floquet Theory

Local stability of a periodic solution of the system ẋ = f (x) (x ∈ Rn ) can be dis-

cussed in terms of eigenvalues of certain matrices. Suppose that the system has a
periodic orbit Γ = {x∗ (t) : t ∈ [0, T ), x∗ (t) = x∗ (t + T )}. Define ξ := x(t) − x∗ (t).
Linearizing ξ̇ about ξ = 0, i.e., about the periodic orbit Γ , we obtain

ξ̇ = A(t)ξ , (2.27)

where the matrix A(t) := D f (x∗ (t)) has periodic coefficients of period T , so that
A(t) = A(t + T ). Solutions of (2.27) take the general form of

B(t)eλ t ,

where the vector B(t) is periodic in time with period T , B(t) = B(t + T ). Denote the
fundamental matrix of (2.27) as Φ (t), that is, the m × m time-varying matrix whose
m columns are solutions of (2.27). Thus, Φ (t) can be written as

Φ (t) = Z(t)etR ,
50 2 Preliminaries of Nonlinear Dynamics and Chaos

where Z(t) is an m × m, T -periodic matrix and R is a constant m × m matrix. More-

over, we can always set Φ (0) = Z(0) = I, from which we get

Φ (T ) = eT R .

Therefore, the dynamics of orbits near the cycle Γ are determined by the eigenvalues
(λ1 , . . . , λm ) of the matrix eT R which are uniquely determined by (2.27).9 The λ s are
called characteristic (Floquet) multipliers of (2.27), whereas the eigenvalues of R,
(κ1 , . . . , κm ), are called characteristic (Floquet) exponents.
One of the roots (multipliers), say λ1 , is always equal to 1, so that one of the
characteristic exponents, say κ1 , is always equal to 0, which implies that one of
the solutions of (2.27) must have the form B(t) = B(t + T ). This can be verified
by putting B(t) = x∗ (t) and differentiating it with respect to time. The presence of
a characteristic multiplier equal to 1 (a characteristic exponent equal to 0) can be
interpreted as that if, starting from a point on the periodic orbit Γ , the system is
perturbed by a small displacement in the direction of the flow, it will remain on Γ .
What happens for small, random displacements off Γ depends only on the remaining
m − 1 multipliers λ j ( j = 2, . . . , m) (or the remaining κ j exponents, j = 2, . . . , m),
provided none of the other moduli is equal to 1 (provided none of them is equal to
0). In particular, we have
(i) If all the characteristic multipliers λ j ( j = 2, . . . , m) satisfy the conditions
|λ j | < 1, then the periodic orbit is asymptotically (in fact, exponentially) or-
bitally stable.
(ii) If for at least one of the multipliers, say λk , |λk | > 1, then the periodic orbit is
unstable.

2.10 Stabilities of Trajectories II: The Lyapunov Second Method

The so-called second or direct method of Lyapunov is one of the greatest landmarks
in the theory of dynamical systems and has proved to be an immensely fruitful tool
for analysis. The basic idea of the method is as follows. Suppose that there is a vector
field in the plane with a fixed point x,
¯ and we want to determine whether it is stable
or not. Roughly speaking, according to our previous definitions of stability it would
be sufficient to find a neighborhood U of x¯ for which orbits starting in U remain
in U for all positive time (for the moment we do not distinguish between stability
and asymptotic stability). This condition would be satisfied if we could show that
the vector field is either tangent to the boundary of U or pointing inward toward x¯
(see Fig. 2.20). This situation should remain true even as we shrink U down to x. ¯
Now, Lyapunov’s method gives us a way of making this precise; we will show this
for vector fields in the plane and then generalize our results to Rn .
9 The matrix eT R itself is uniquely determined but for a similarity transformation, that is, we can
substitute eT R with P−1 eT R P where P is a nonsingular m × m matrix. This transformation leaves
eigenvalues unchanged.
2.10 Stabilities of Trajectories II: The Lyapunov Second Method 51

Suppose that we have the vector field


ẋ = f (x, y),
(2.28)
ẏ = g(x, y), (x, y) ∈ R2 ,

which has a fixed point at (x, ¯ y)

¯ (assume that it is stable). We want to show that in
any neighborhood of (x, ¯ y)
¯ the above situation holds. Let V (x, y) be a scalar-valued
function on R2 , i.e., V : R2 → R (and at least C1 ), with V (x,¯ y)
¯ = 0, and such that
the loci of points satisfying V (x, y) = C = constant form closed curves for different
values of C encircling (x, ¯ y)
¯ with V (x, y) > 0 in a neighborhood of (x, ¯ y)
¯ (see Fig.
2.21).
Recall that the gradient of V , ∇V , is a vector perpendicular to the tangent vector
along each curve V = C which points in the direction of increasing V . So, if the
vector field were always to be either tangent or pointing inward for each of these
curves surrounding (x, ¯ y),
¯ we would have

∇V (x, y) · (ẋ, ẏ) ≤ 0,

Fig. 2.20 The vector field on the boundary of U

∇V ∇V

∇V
∇V
(x , y )

V = constant
∇V
∇V
Fig. 2.21 Level set of V and ∇V denotes gradient vector of V at various points on the boundary
52 2 Preliminaries of Nonlinear Dynamics and Chaos

where the ‘·’ represents the usual vector scalar product. (This is simply the derivative
of V along orbits of (2.28), and is sometimes referred to as the orbital derivative.)
We now state the general theorem which makes these ideas precise.
Theorem 2.6. Consider the following vector field:

ẋ = f (x), x ∈ Rn . (2.29)

Let x¯ be a fixed point of (2.29) and let V : U → R be a C1 function defined on some

neighborhood U of x¯ such that
¯ = 0 and V (x) > 0 if x = x.
(i) V (x) ¯
(ii) V̇ ≤ 0 in U − {x}.
¯
Then, x¯ is stable. Moreover, if
(iii) V̇ < 0 in U − {x},
¯
then x¯ is asymptotically stable. 
We refer to V as a Lyapunov function. If U can be chosen to be all of Rn , then x¯
is said to be globally asymptotically stable, if (i) and (iii) hold.
Sometimes it is possible to prove asymptotic stability of a fixed point even when
the Lyapunov function V in the relevant neighborhood of the point implies that
V̇ ≤ 0, but not necessarily V̇ < 0. For that case we have the following theorem.
Theorem 2.7 (Invariance Principle of LaSalle). Let x¯ = 0 be a fixed point of ẋ =
f (x) and V a Lyapunov function such that V̇ ≤ 0 on some neighborhood N of x¯= 0.
If x0 ∈ N has its forward orbit, γ + (x0 ) = {φ (t, x0 ) : t ≥ 0}, bounded with limit points
in N, and M is the largest invariant subset of E = {x ∈ N : V̇ (x) = 0}, then

φ (t, x0 ) → M as t → ∞. 

According to Theorem 2.7, if a Lyapunov function V (x) can be found such that
V̇ (x) ≤ 0 for x ∈ N, among the sets of points with forward orbits in N there exist
sets of points defined by
Vk = {x : V (x ≤ k)}
(k is a finite and positive scalar) which lie entirely in N. Since V̇ ≤ 0, the sets Vk are
invariant in the sense that no orbit starting in a Vk can ever move outside of it. If,
in addition, it could be shown that the fixed point x¯ = 0 is the largest (or, for that
matter, the only) invariant subset of E, Theorem 2.7 would guarantee its asymptotic
stability.
The direct method can also be extended to discrete-time systems. We only state
a result analogous to Theorem 2.6 in the following. A discrete-time version of the
invariance principle of LaSalle will be introduced in Chap. 5.
Theorem 2.8. Consider the system described by the difference equation given in
(2.21). Let x¯ = 0 again be an isolated equilibrium point at the origin. If there exists
a C1 function V (xn ) : N → R, defined on some neighborhood N ⊂ Rm of x¯= 0, such
that
2.11 Chaotic Sets and Chaotic Attractors 53

(i) V (0) = 0,
(ii) V (x) > 0 in N − {0},
(iii) Δ V (xn ) := V [G(xn )] − V (xn ) ≤ 0 in N − {0},
then x¯= 0 is stable (in the sense of Lyapunov). Moreover, if
(iv) Δ V (x) < 0 in N − {0},
then x¯= 0 is asymptotically stable. 

2.11 Chaotic Sets and Chaotic Attractors

More complicated invariant, attracting sets and attractors in structure than that of
periodic or quasiperiodic sets are called chaotic. A dynamical system (discrete-time
or continuous-time) is called chaotic if its typical orbits are aperiodic, bounded, and
such that nearby orbits separate fast in time. Chaotic orbits never converge to a stable
fixed or periodic point, but exhibit sustained instability, while remaining forever in
a bounded region of the state space.
Definition 2.20. A flow φ (a continuous map G) on a metric space M is said to
possess sensitive dependence on initial conditions on M if there exists a real number
δ > 0 such that for all x ∈ M and for all ε > 0, there exist y ∈ M (y = x) and T > 0 (an
integer n > 0) such that d(x, y) < ε and d[φ (T, x), φ (T, y)] > δ (d[Gn (x), Gn (y)] >
δ ). 

Definition 2.21. A flow φ (a continuous map G) is said to be chaotic on a compact

invariant set10 A if:
(i) it is topologically transitive on A (Definition 2.12);
(ii) it has sensitive dependence on initial conditions on A. 

Remark 2.4. There is something that should be pointed out.

(i) Condition (i) of Definition 2.21 guarantees that the invariant set is single and
indecomposable.
(ii) Condition (ii) of Definition 2.21 can be made sharper (and more restrictive) in
two ways. First, the divergence of nearby points taking place at an exponential
rate is required. This property can be made more precise by means of Lyapunov
exponents which will be introduced later. Second, we may require that the di-
vergence (exponential or otherwise) occurs for each pair x, y ∈ A. In this case,
the flow φ or map G is called expansive on A.
(iii) The requirement that A is compact is necessary. Consider the following differ-
ential equation:
ẋ = ax, x ∈ R, a > 0,

10 Roughly speaking, a subset D of Rn is said to be compact if it can be covered by a finite

collection of open sets {U j }lj=1 .
54 2 Preliminaries of Nonlinear Dynamics and Chaos

which is linear and its solution is φ (t, x) = x0 eat . Therefore, the flow map φ
is topologically transitive on the open, unbounded (and therefore noncompact)
invariant sets (−∞, 0) and (0, ∞). Also, for any two points x1 , x2 ∈ R and x1 = x2
we have
|φ (t, x1 ) − φ (t, x2 )| = eat |x1 − x2 |
and φ has sensitive dependence on initial conditions on R. However, the orbits
generated by φ are not chaotic.
(iv) This definition refers to a ‘chaotic flow (or map) on a set A’ or, for short, a
‘chaotic set A.’ It does not imply that all orbits of a chaotic flow (or map) on A
are chaotic. In fact, there are many nonchaotic orbits on chaotic sets, in particu-
lar, many unstable periodic orbits. They are so important that some researchers
add a third condition for chaos, that periodic orbits are dense on A [5]. This is
an interesting property and it is automatically satisfied if the chaotic invariant
set is hyperbolic [17].
(v) Two quite general results can be used to confirm the close relationship between
chaos, as characterized in Definition 2.21, and dense periodic sets. The first re-
sult [3] states that for any continuous map on a metric space, transitivity and the
presence of a dense set of periodic orbits imply sensitive dependence on initial
conditions, that is, chaos. The second result [16] states that for any continuous
map on an interval of R, transitivity alone implies the presence of a dense set
of periodic orbits and, therefore, in view of the first result, it implies sensitive
dependence on initial conditions, and therefore chaos.
(vi) There are several other different definitions of chaos based on orbits rather than
sets. For example, in [1] (p. 196, Definition 5.2; p. 235, Definition 6.2; pp. 385–
386, Definition 9.6), a chaotic set is defined as the ω -limit set of a chaotic orbit
Gn (x0 ) which itself is contained in the ω -limit set. In this case, the presence
of sensitive dependence on initial conditions (or a positive Lyapunov charac-
teristic exponent) is not enough to characterize chaotic properties of orbits and
additional conditions must be added to exclude unstable periodic or quasiperi-
odic orbits. 

2.12 Symbolic Dynamics and the Shift Map

Symbolic dynamics is a powerful tool for understanding the orbit structure of a large
class of dynamical systems. In this section we only provide a brief introduction to
this tool.
To establish the tool, three steps are needed. First, we define an auxiliary system
characterized by a map, called a shift map, acting on a space of infinite sequences
called the symbol space. Next, we prove some properties of the shift map. Finally,
we establish a certain equivalence relation between a map we want to study and the
shift map, and show that the relationship preserves the properties in question.
2.12 Symbolic Dynamics and the Shift Map 55

We begin by defining the symbol space and the shift map. Let S be a collection
of symbols. In a physical interpretation, the elements of S could be anything, for
example letters of an alphabet or discrete readings of some measuring device for the
observation of a given dynamical system. To make ideas more clear, we assume here
that S consists of only two symbols; let them be 0 and 1. Then, we have S = {0, 1}.
Next, we want to construct the space of all possible bi-infinite sequences of 0 and 1,
defined as
Σ2 := · · · S × S × S × · · · .
A point in Σ2 , s, is therefore represented as a bi-infinity-tuple of elements of S, that
is, s ∈ Σ2 means
s = {. . . s−n . . . s−1 s0 s1 . . . sn . . . },
where ∀i, si ∈ S (i.e., si = 0 or 1). For example, s = {. . . 00010100111 . . .}.
We can define a distance function d¯in the space Σ2
+∞
d(si , s¯i )
¯ s)
d(s, ¯= ∑ 2|i|
, (2.30)
i=−∞

where d is the discrete distance in S = {0, 1}, that is


0 if si = s¯i ;
d(si , s¯i ) =
1 if si = s¯i .

This means that two points of Σ2 are close to each other if their central elements
are close, i.e., if the elements whose indexes have small absolute values are close.
¯ i , s¯i ), the infinite sum on the right-hand side of
Notice that, from the definition of d(s
(2.30) is less than 3, and, therefore, converges.
Next, we define the shift map on Σ2 as

T : Σ2 → Σ2 , T (s) = s and si = si+1 .

The map T maps each entry of a sequence from one place to the left. Similarly, the
one-sided shift map T+ can be defined on the space of one-sided infinite sequences,
Σ2+ , that is, s ∈ Σ2+ , where s = {s0 s1 . . . sn . . . }. In this case, we have

T+ : Σ2+ → Σ2+ , T+ (s) = s and si = si+1 ,

so that
T+ (s0 s1 s2 . . . ) = (s0 s1 s2 . . . ) = (s1 s2 s3 . . . ).
It is obvious that the T+ map shifts a one-sided sequence from one place to the left
and drops its first element. Although maps T and T+ have very similar properties, T
is invertible whereas T+ is not. The distance on Σ2+ is essentially the same as (2.30)
with the difference that the infinite sum will now run from zero to ∞. The map T+
can be used to prove chaotic properties of certain noninvertible, one-dimensional
maps frequently employed in applications.
56 2 Preliminaries of Nonlinear Dynamics and Chaos

Theorem 2.9. The shift map T+ on Σ2+ is chaotic according to Definition 2.21. 
Remark 2.5. The shift map T+ on Σ2+ has a property that is stronger than topo-
logical transitivity called topological mixing. In general, we say that a map G is
topologically mixing on a set A if for any two open subsets U and V of A there
exists a positive integer N0 such that Gn (U) ∩ V = 0/ for all n ≥ N0 . If a map G is
topologically mixing, then for any integer n the map Gn is topologically transitive.

The importance of the fact that the shift map is chaotic in a precise sense lies
in that the chaotic properties of invariant sets of certain one- and two-dimensional
maps and three-dimensional flows may sometimes be proved by showing that the
dynamics on these sets are topologically conjugate to that of a shift map on a sym-
bol space. This indirect argument is often the only available strategy for investi-
gating nonlinear maps (or flows). We have encountered the concept of topological
conjugacy in the Hartman–Grobman theorem (Theorem 2.5, which we called home-
omorphic equivalence) between a nonlinear map (or flow) and its linearization in a
neighborhood of a fixed point. We now provide some formal definitions.
Definition 2.22. Let X and Y be topological spaces, and let f : X → X and g : Y →
Y be continuous functions. We say that f is topologically semiconjugate to g if
there exists a continuous surjection11 h : Y → X such that f ◦ h = h ◦ g. If h is a
homeomorphism, then we say that f and g are topologically conjugate, and we call
h a topological conjugation between f and g.
Similarly, a flow φ on X is topologically semiconjugate to a flow ψ on Y if there
is a continuous surjection h : Y → X such that φ (h(y),t) = h(ψ (y,t)) for each y ∈ Y ,
t ∈ R. If h is a homeomorphism, then φ and ψ are topologically conjugate. 
Remark 2.6. Topological conjugation defines an equivalence relation in the space
of all continuous surjections of a topological space to itself, by declaring f and g to
be related if they are topologically conjugate. This equivalence relationship is very
useful in the theory of dynamical systems, since each class contains all functions
which share the same dynamics from the topological viewpoint. In fact, orbits of
g are mapped to homeomorphic orbits of f through the conjugation. Writing g =
h−1 ◦ f ◦ h makes this fact evident: gn = h−1 ◦ f n ◦ h. Roughly speaking, topological
conjugation is a ‘change of coordinates’ in the topological sense. 
However, the analogous definition for flows is somewhat restrictive. In fact, we
require the maps φ (·,t) and ψ (·,t) to be topologically conjugate for each t, which
requires more than simply that orbits of φ be mapped to orbits of ψ homeomorphi-
cally. This motivates the definition of topological equivalence, which also partitions
the set of all flows in X into classes of flows sharing the same dynamics, again from
the topological viewpoint.
Definition 2.23. We say that two flows ψ and φ of a compact manifold M are topo-
logically equivalent if there is an homeomorphism h : Y → X, mapping orbits of
11 A function f : X → Y is a surjection if, for every y ∈ Y , there is an x ∈ X such that f (x) = y.
2.13 Lyapunov Exponent 57

ψ to orbits of φ homeomorphically, and preserving orientation of the orbits. This

means that
(i) {h(ψ (y,t)) : t ∈ R} = {φ (h(y),t) : t ∈ R} for each y ∈ Y ;
(ii) for each y ∈ Y , there is δ > 0 such that, if 0 < |s| < t < δ , and if s satisfies
φ (h(y), s) = h(ψ (y,t)), then s > 0. 

2.13 Lyapunov Exponent

Although sensitive dependence on initial conditions can be verified in some cases,

it is not easy to verify for many systems. The Lyapunov exponent is a generalization
of the eigenvalues at an equilibrium point, and it is used as a measure of expo-
nential divergence of orbits. Suppose that φ (t, x0 ) and φ (t, y0 ) are solutions of an
autonomous vector field ẋ = f (x) starting from x0 and y0 , respectively. By using the
linear approximation for fixed t, we get

φ (t, y0 ) − φ (t, x0 ) ≈ Dx φ (t, x0 )(y0 − x0 ).

For any curve starting from initial condition xs , letting

∂ ∂ xs
v(t) := φ (t, xs ) = Dx φ (t, x0 ) = Dx φ (t, x0 )v0
∂s s=0 ∂s

∂ xs
and v0 := , then v(t) satisfies the first variation equation
∂s
d
v(t) = D f(φ (t,x0 )) v(t).
dt
If v0 = y0 − x0 , then v(t) would give the infinitesimal displacement at time t.
The growth rate of v(t) is a number χ such that

v(t) ≈ Ceχ t ,

where C is a constant. Taking the logarithm, we have

ln v(t) ln(C)
≈ + χ;
t t
therefore,
ln v(t)
χ = lim .
t→∞ t
Definition 2.24. Let v(t) be the solution of the first variation equation, starting from
x0 with v(0) = v0 . The Lyapunov exponent for initial condition x0 and initial in-
finitesimal displacement v0 is defined as
58 2 Preliminaries of Nonlinear Dynamics and Chaos

ln v(t)
χ (x0 , v0 ) = lim ,
t→∞ t
whenever this limit exists. 

Remark 2.7. For most initial conditions x0 for which the forward orbit is bounded,
the Lyapunov exponents exist for all vectors v. In n-dimensional state space, there
are at most n distinct values for χ (x0 , v) as v varies. If we count multiplicities, then
there are exactly n values,

χ1 (x0 ) = χ (x0 , v1 ), χ2 (x0 ) = χ2 (x0 , v2 ), . . . , χn (x0 ) = χ (x0 , vn ).

We can order these so that

χ1 (x0 ) ≥ χ2 (x0 ) ≥ · · · ≥ χn (x0 ). 

Several results on Lyapunov exponents are listed as follows. For detailed proofs
please refer to [13].
Theorem 2.10. Assume that x0 is a fixed point of the differential equation ẋ = f (x).
Then, the Lyapunov exponents at the fixed point are the real parts of the eigenvalues
of the fixed point. 

Theorem 2.11. Let x0 be an initial condition such that φ (t, x0 ) is bounded and ω (x0 )
does not contain any fixed points. Then,

χ (x0 , f (x0 )) = 0. 

Remark 2.8. The above theorem means that there is no growth or decay in the direc-
tion of the vector field, v = f (x0 ). 

Theorem 2.12. Let x0 be an initial condition on a periodic orbit of period T . Then,

the principal (n − 1) Lyapunov exponents are given by (ln |λ j |)/T , where λ j are the
characteristic multipliers of the periodic orbit and the eigenvalues of the Poincaré
map. 

Theorem 2.13. Assume that φ (t, x0 ) and φ (t, y0 ) are two orbits for the same dif-
ferential equation, which are bounded and converge exponentially (i.e., there are
constants a > 0 and C ≥ 1 such that φ (t, x0 ) − φ (t, y0 ) ≤ Ce−at for t ≥ 0). Then,
the Lyapunov exponents for x0 and y0 are the same. So, if the limits defining the Lya-
punov exponents exist for one of the points, they exist for the other point. The vec-
tors which give the various Lyapunov exponents can be different at the two points.


Theorem 2.14. Consider the system of ẋ = f (x) in Rn . Assume that x0 is a point

such that the Lyapunov exponents χ1 (x0 ), . . . , χn (x0 ) exist.
(i) Then, the sum of Lyapunov exponents is the limit of the average of the diver-
gence along the trajectory,
2.14 Examples 59
n 
1 T
∑ χ j (x0) = Tlim
→∞ T 0
∇ · fφ (t,x0 ) dt.
j=1

(ii) In particular, if the system has constant divergence δ , then the sum of the Lya-
punov exponents at any point must equal δ .
(iii) In the three-dimensional case, assume that the divergence is a constant δ and
that x0 is a point for which the positive orbit is bounded and ω (x0 ) does not
contain any fixed points. If χ1 (x0 ) is a nonzero Lyapunov exponent at x0 , then
the other two Lyapunov exponents are 0 and δ − χ1. 

The definition of the Lyapunov exponent for discrete-time systems is similar to

the case of continuous time.
Definition 2.25. The Lyapunov exponent for the maps xn+1 = G(xn ) is defined by

Dn G(x0 )w
χ (x0 , w) = lim ln ,
n→∞ w

where
Dn G(x0 ) = DG(x0 )DG(x1 ) · · · G(xn−1 )
and w is a vector in the tangent space at x0 . 

2.14 Examples

In this section, we will explore how chaos appears by investigating some examples.

2.14.1 Tent Map and Logistic Map

The tent map has the form of

⎧ 1
⎪
⎨ 2y, if 0 ≤ y ≤ ,
2
GΛ (y) = (2.31)
⎪
⎩ 2(1 − y), if 1 < y ≤ 1,
2
which is shown in Fig. 2.22.
Proposition 2.1. The tent map (2.31) is chaotic on [0, 1].

Proof. Consider that the graph of the nth iteration of GΛ consists of 2n linear pieces,
each with slope ±2n . Each of these linear pieces of the graph is defined on a subin-
terval of [0, 1] of length 2−n . Then, for any open subinterval J of [0, 1], we can find
a subinterval K of J of length 2−n , such that the image of K under GΛn covers the
60 2 Preliminaries of Nonlinear Dynamics and Chaos

entire interval [0, 1]. Therefore, GΛ is topologically transitive on [0, 1]. This fact,
and the discussion in point (v) of Remark 2.4 [16], proves the proposition. 

Remark 2.9. From the geometry of the iterated map GΛm , it appears that the graph of
GΛm on J intersects the bisector and therefore GΛm has a fixed point in J. This proves
that periodic points are dense in [0, 1]. Also, for any x ∈ J there exists a y ∈ J such
that |GΛn (x) − GΛn (y)| ≥ 1/2 and, therefore, GΛ has sensitive dependence on initial
conditions. 

This result can be used to show that the logistic map

G4 : [0, 1] → [0, 1], G4 (x) = 4x(1 − x)

(see Fig. 2.23) is also chaotic. Consider the map h(y) = sin2 (π y/2). The map h is
continuous and, restricted to [0, 1], is also one-to-one and onto. Its inverse is contin-
uous and h is thus a homeomorphism. Consider now the diagram
G
[0, 1] −−−Λ−→ [0, 1]
⏐ ⏐
⏐
h(y)
⏐h(y)

[0, 1] −−−−→ [0, 1]
G4

where GΛ is the tent map. Recalling the trigonometric relations:

(i) sin2 (θ ) + cos2 (θ ) = 1;
(ii) 4 sin2 (θ ) cos2 (θ ) = sin2 (2θ );
(iii) sin(π − θ ) = sin(θ );
we can see that the diagram is commutative. Hence, the map G4 is topologically
conjugate to GΛ and, therefore, its dynamics on [0, 1] is chaotic.

GΛ (x) GΛ3 (x)

1 1

1 1
1 x 1 x
2 2
(a) (b)

Fig. 2.22 The tent map: (a) GΛ ; (b) GΛ3

2.14 Examples 61

1
G4 (x)
0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.
x

Fig. 2.23 The logistic map G4

Both the tent map and the logistic map are all from an interval of R to itself.
There are also many other one-dimensional mappings presenting chaotic dynamics.
In fact, for one-dimensional mappings from R to itself we have the following so-
called Li–Yorke theorem [13], famous for the phrase ‘period three implying chaos,’
and Sarkovskii’s theorem [13], the generalization of Li–Yorke’s result. Since many
references include the proof of these two theorems, we will only state the content of
the theorems.
Theorem 2.15 (Li–Yorke [13]). Assume that f is a continuous function from R to
itself.
(i) If f has a period-3 point, then it has points of all periods.
(ii) Assume that there is a point x0 such that either
a. f 3 (x0 ) ≤ x0 < f (x0 ) < f 2 (x0 ) or
b. f 3 (x0 ) ≥ x0 > f (x0 ) > f 2 (x0 ).
Then, f has points of all periods. 
This theorem was obtained by Li and Yorke in 1975 and soon after it was shown
that Li–Yorke theorem is a special case of Sharkovskii’s theorem. Before introduc-
ing Sharkovskii’s theorem, we first define a new order for natural numbers as fol-
lows:
3  5  7 · · ·  2 · 3  2 · 5  2 · 7  · · · 2k · 3  2k · 5  2k · 7  · · ·  2k  2k−1  · · ·  22  2  1.

Theorem 2.16 (Sharkovskii’s theorem [13]). Let f be a continuous function from

R to itself. Suppose that f has period-n points and n  k. Then, f has period-k
points. 
62 2 Preliminaries of Nonlinear Dynamics and Chaos

2.14.2 Smale Horseshoe

The Smale horseshoe is the prototypical map possessing a chaotic invariant set.
Therefore, a thorough understanding of the Smale horseshoe is absolutely essential
for understanding what is meant by the term ‘chaos’ as it is applied to the dynamics
of specific physical systems [10].
Consider the geometrical construction in Fig. 2.24. Take a square S on the plane
(Fig. 2.24 (a)). Contract it in the horizontal direction and expand it in the verti-
cal direction (Fig. 2.24 (b)). Fold it in the middle (Fig. 2.24 (c)) and place it so
that it intersects the original square S along two vertical strips (Fig. 2.24 (d)). This
procedure defines a map f : R2 → R2 . The image f (S) of the square S under this
transformation resembles a horseshoe. That is why it is called a horseshoe map.
The exact shape of the image f (S) is irrelevant; however, for simplicity we assume
that both the contraction and expansion are linear and that the vertical strips in the
intersection are rectangle. The map f can be made invertible and smooth together
with its inverse. The inverse map f −1 transforms the horseshoe f (S) back into the
square S through stages (d)–(a). This inverse transformation maps the dotted square
S shown in Fig. 2.24 (d) into the dotted horizontal horseshoe in Fig. 2.24 (a), which
is assumed to intersect the original square S along two horizontal rectangles.
Denote the vertical strips in the intersection S ∩ f (S) by V1 and V2 ,

S ∩ f (S) = V1 ∪V2

(see Fig. 2.25 (a)). Now make the most important step: perform the second iteration
of the map f . Under this iteration, the vertical strips V1 and V2 will be transformed

B C f
B
C

B C

S S

A D
A D C B
-1
f
A D
A D
(a) (b) (c) (d)

Fig. 2.24 Construction of the horseshoe map

2.14 Examples 63

into two ‘thin horseshoes’ that intersect the square S along four narrow vertical
strips: V11 ,V21 ,V22 , and V12 (see Fig. 2.25 (b)). We write this as

S ∩ f (S) ∩ f 2 (S) = V11 ∪V21 ∪V22 ∪V12 .

Similarly,
S ∩ f −1 (S) = H1 ∪ H2 ,
where H1 and H2 are the horizontal strips shown in Fig. 2.25 (c), and

S ∩ f −1 (S) ∩ f −2 (S) = H11 ∪ H12 ∪ H22 ∪ H21 ,

with four narrow horizontal strips Hi j (Fig. 2.25 (d)). Notice that f (Hi ) = Vi , i = 1, 2,
as well as f 2 (Hi j ) = Vi j , i, j = 1.2 (see Fig. 2.26).
Iterating the map f further, we obtain 2k vertical strips in the intersection S ∩
f k (S), k ∈ N. Similarly, iteration of f −1 gives 2k horizontal strips in the intersection
S ∩ f −k (S), k ∈ N.
Most points leave the square S under iterations of f or f −1 . We consider all
remaining points in the square under all iterations of f and f −1 :

Γ = {x ∈ S : f k (x) ∈ S, ∀k ∈ Z}.

Clearly, if the set Γ is nonempty, it is an invariant set of the discrete-time dynamical

system defined by f . This set can be alternatively presented as an infinite intersec-
tion,
Γ = · · · ∩ f −k (S) ∩ · · · ∩ f −1 (S) ∩ S ∩ f (S) ∩ · · · f k (S) ∩ · · · .
It is clear from this representation that the set Γ has a peculiar shape. Indeed, it
should be located within
f −1 (S) ∩ S ∩ f (S),
which is formed by four small squares (see Fig. 2.27 (a)). Furthermore, it should be
located inside
f −2 (S) ∩ f −1 (S) ∩ S ∩ f (S) ∩ f 2(S),

V1 V2 V 11 V 21 V 22 V 12
H 21
H2
H 22

H 12
H1
H 11
U U 2 U -1 U U
S f ( S) f ( S)
U
S f ( S) S f ( S) S f -1
( S) f -2
( S)

(a) (b) (c) (d)

Fig. 2.25 Vertical and horizontal strips

64 2 Preliminaries of Nonlinear Dynamics and Chaos

H 21
f (H 12) f (H 22)
H 22

H 12
f (H 11) f (H 21)
H 11
V 11 V 21 V 22 V 12
f f

(a) (b) (c)

Fig. 2.26 Transformation f 2 (Hi j ) = Vi j , i, j = 1, 2

which is the union of sixteen smaller squares (see Fig. 2.27 (b)), and so forth. In the
limit, we get a Cantor set. About the horseshoe map, we have the following lemma.

U U -1
f ( S)
-1 -2 U U
f ( S) S f ( S) f ( S)
U U 2
S f ( S) f ( S)

(a) (b)

Fig. 2.27 Location of the invariant set

Lemma 2.1. There is a one-to-one correspondence h : Γ → Σ2 between points of Γ

and all bi-infinite sequences of two symbols.

Proof. For any point x ∈ Γ , define a sequence of the two symbols {1, 2} by

ω = {. . . , ω−2 , ω−1 , ω0 , ω1 , ω2 , . . . }

by the formula 
1, if f k (x) ∈ H1 ,
ωk = (2.32)
2, if f k (x) ∈ H2 ,
for k ∈ Z. Here, f 0 = id, the identity map. Clearly, this formula defines a map
h : Γ → Σ2 , which assigns a sequence to each point of the invariant set. To ver-
ify that this map is invertible, take a sequence ω ∈ Σ2 , fix m > 0, and consider a set
Rm (ω ) of all points x ∈ S, not necessarily belonging to Γ , such that
2.14 Examples 65

f k (x) ∈ Hωk ,

for −m ≤ k ≤ m− 1. For example, if m = 1, the set R1 is one of the four intersections

V j ∩ Hk . In general, Rm belongs to the intersection of a vertical and a horizontal strip.
These strips are getting thinner and thinner as m → ∞, approaching in the limit a
vertical and a horizontal segment, respectively. Such segments obviously intersect
at a single point x with h(x) = ω . Thus, h : Γ → Σ2 is a one-to-one map. It implies
that Γ is nonempty. 

Remark 2.10. The map h : Γ → Σ2 is continuous together with its inverse (a home-
omorphism) if we use the standard Euclidean metric in S ⊂ R2 and the metric given
by (2.30) in Σ2 . 

Consider now a point x ∈ Γ and its corresponding sequence ω = h(x), where h is

the map previously constructed. Next, consider a point y = f (x), that is, the image
of x under the horseshoe map f . Since y ∈ Γ by definition, there is a sequence that
corresponds to y : θ = h(y). As one can easily see from (2.32), there is a simple
relationship between these sequences ω and θ . Namely,

θk = ωk+1 , k ∈ Z,

since
f k ( f (x)) = f k+1 (x).
In other words, the sequence θ can be obtained from the sequence ω by the shift
map σ :
θ = σ (ω ).
Therefore, the restriction of f to its invariant set Γ ⊂ R2 is equivalent to the shift
map σ on the set of sequences Σ2 . This result can be formulated as the following
lemma.

Lemma 2.2. h( f (x)) = σ (h(x)), for all x ∈ Γ . 

This lemma can be written as an even shorter one:

f |Γ = h−1 ◦ σ ◦ h.

Combining Lemmas 2.1 and 2.2 with obvious properties of the shift dynamics
on Σ2 , we get a theorem giving a rather complete description of the behavior of the
horseshoe map.
Theorem 2.17. The horseshoe map f has a closed invariant set Γ that contains a
countable set of periodic orbits of arbitrarily long period, and an uncountable set
of nonperiodic orbits, among which there are orbits passing arbitrarily close to any
point of Γ . 

Remark 2.11. The limit set Γ of a Smale horseshoe map is unstable and, therefore,
not attracting. It follows that the existence of a Smale horseshoe does not imply
66 2 Preliminaries of Nonlinear Dynamics and Chaos

the existence of a chaotic attractor. The existence of a Smale horseshoe does imply,
however, that there is a region in state space that experiences sensitive dependence
on initial conditions. Thus, even when there is no strange attractor in the flow, the
dynamics of the system can appear chaotic until the steady state is reached. 

Remark 2.12. We can slightly perturb the constructed map f without qualitative
changes to its dynamics. Clearly, Smale’s construction is based on a sufficiently
strong contraction or expansion, combined with a fold. Thus, a (smooth) perturba-
tion f˜ will have similar vertical and horizontal strips, which are no longer rectangles
but curvilinear regions. However, provided that the perturbation is sufficiently small,
these strips will shrink to curves that deviate only slightly from vertical and hori-
zontal lines. Thus, the construction can be carried through word for word, and the
perturbed map f˜ will have an invariant set Γ on which the dynamics is completely
described by the shift map σ on the sequence space Σ2 . This is an example of struc-
turally stable behavior. 

2.14.3 The Lorenz System

Although plenty of numerical evidence of chaotic behavior arising from a variety of

problems in different fields of applications has been provided, apart from the cases
of one-dimensional, noninvertible maps, there are few rigorous proofs that specific
mathematical models possess chaotic attractors as characterized in one or more of
the above definitions, and those proofs are mostly restricted to artificial examples
unlikely to arise in typical applications. We now turn to an example of a chaotic
attractor derived from a celebrated model first discussed in 1963 by E. Lorenz to
provide a mathematical description of atmospheric turbulence. Lorenz’s investiga-
tion was enormously influential and stimulated a vast literature in the years that
followed. Its extraordinary success was in part due to the fact that it showed how
computer technology could be effectively used to study nonlinear dynamical sys-
tems. Lorenz’s work provided strong numerical evidence that a low-dimensional
system of differential equations with simple nonlinearities could generate extremely
complicated orbits. The original Lorenz model is defined by the following three dif-
ferential equations: ⎧
⎨ ẋ = −σ x + σ y,
ẏ = −xz + rx − y, (2.33)
⎩
ż = xy − bz,
where x, y, z ∈ R and σ , r, b > 0. System (2.33) is symmetrical under the transforma-
tion (x, y, z) → (−x, −y, z). The three equilibria are

E1 : 0, 0),
(0, 
E2 : (  b(r − 1), b(r
− 1), r − 1),
E3 : (− b(r − 1), − b(r − 1), r − 1).
2.14 Examples 67

In numerical analysis of the Lorenz model, the typical parameter configuration is

σ = 10 and b = 8/3. When r < rH = 24.74 there are two symmetric unstable pe-
riodic orbits with the Lorenz system. When r > rH , the celebrated Lorenz attractor
(the so-called ‘butterfly’) is observed numerically (see Fig. 2.28). The three Lya-
punov exponents are 1.497, 0, and −22.46, which imply that the Lorenz system has
sensitive dependence on initial conditions.

50

45

40

35

30

25
z

20

15

10

0
−20 −15 −10 −5 0 5 10 15 20
x

Fig. 2.28 The Lorenz attractor

Remark 2.13. When we refer to a computer to visualize the numerical solution of

a chaotic system, an important issue arises. If we take into account the combined
influence of round-off errors in numerical computations and the property of diver-
gence of nearby trajectories for chaotic behavior, how can we trust numerical com-
putations of trajectories to give us reliable results? (We note that the same problem
arises in experimental measurements in which ‘noise’ plays the role of round-off
errors.) If the system’s behavior is chaotic, then even small numerical errors are
amplified exponentially in time. Perhaps all of our results for chaotic systems are
artifacts of the numerical computation procedure. Even if they are not artifacts, per-
haps the numerical values of the properties depend critically on the computational
procedures. If that is true, how general are our results? Although it is difficult to
answer these questions once and for all, it is comforting to know that while it is true
that the details of a particular trajectory do depend on the round-off errors in the
numerical computation, the trajectory actually calculated does follow very closely
68 2 Preliminaries of Nonlinear Dynamics and Chaos

some trajectory of the system. That is, the trajectory one calculates might not be
exactly the one he thinks, but it is very close to one of the possible trajectories of
the system. In more technical terms, we say that the computed trajectory shadows
some possible trajectories of the system. (A proof of this shadowing property for
chaotic systems is given in [6] and [14].) In general, we are most often interested
in properties that are averaged over a trajectory; in many cases those average values
are independent of the particular trajectory we follow. So, as long as we follow some
possible trajectory for the system, we can have confidence that our results are a good
characterization of the system’s behavior. Recently, W. Tucker’s work [15] strength-
ened the above discussion, in which he has shown, using a computer-assisted proof,
that the Lorenz system not only has sensitive dependence on initial conditions, but
also has a chaotic attractor. 
Although a full mathematical analysis of the observed attractor is still lacking,
some of the attractor’s properties have been established through a combination of
numerical evidence and theoretical arguments. Before presenting the analysis we
first consider the following three facts about system (2.33).
(i) The trace of the Jacobian matrix

∂ ẋ ∂ ẏ ∂ ż
tr[D f (x, y, z)] = + + = −(b + σ + 1) < 0
∂x ∂y ∂z
is constant and negative along orbits. Thus, any three-dimensional volume of
initial conditions is contracted along orbits at a rate equal to

γ = −(b + σ + 1) < 0;

that is to say, the system is dissipative.

(ii) It is possible to define a trapping region such that all orbits outside of it tend to
it, and no orbits ever leave it. To see this, consider the function

V (x, y, z) = x2 + y2 + (z − r − σ )2 = K 2 (r + σ )2 (2.34)

defining a sphere with center at (x = y = 0; z = σ + r) and radius K(σ + r). The

time derivative of (2.34) along the solution of (2.33) is
2
r+σ (r + σ )2
V̇ (x, y, z) = −2σ x2 − 2y2 − 2b z − +b .
2 2

V̇ = 0 defines an ellipsoid outside of which V̇ < 0. For a sufficiently large value

of the radius (for sufficiently large K, given r and σ ), the sphere (2.34) will
contain all three fixed points and all orbits on the boundary of the sphere will
point inward. Consequently, system (2.33) is ‘trapped’ inside the sphere.
(iii) Numerical evidence indicates that for r ∈ (13.8, 14), there exist two symmetric
homoclinic orbits (that is, orbits that connect a fixed point to itself) asymptot-
ically approaching the origin for t → ±∞, tangentially to the z axis; see the
sketch in Fig. 2.29.
2.14 Examples 69

Keeping the three facts in mind, we can investigate the Lorenz system by con-
structing a geometric model which, under a certain hypothesis, provides a reason-
able approximation of the dynamics of the original model with ‘canonical’ parame-
ter values σ = 10, b = 8/3, and r > rH .
We first consider a system of differential equations in R3 depending on a param-
eter μ with the following properties:
(i) for a certain value μh of the parameter, there exists a pair of symmetrical ho-
moclinic orbits, asymptotically converging to the origin, and tangential to the
positive z axis;
(ii) the origin is a saddle-point equilibrium and the dynamics in a neighborhood
N of the equilibrium, for μ in a neighborhood of μh , is approximated by the
system ⎧
⎨ ẋ = λ1 x,
ẏ = λ2 y,
⎩
ż = λ3 z,
where λ2 < λ3 < 0 < λ1 and −λ3 /λ1 < 1;
(iii) the system is invariant under the change of coordinates (x, y, z) → (−x, −y, z).
Under these conditions, for (x, y, z) ∈ N and μ near μh , it is possible to construct
a two-dimensional cross section Σ , such that the transversal intersections of orbits
with Σ define a two-dimensional Poincaré map P : Σ → Σ . For values of |x| and
|μ − μh | sufficiently small, the dynamics of P can further be approximated by a one-
dimensional, noninvertible map Gμ [−a, a] − {0} → R defined on an interval of the
x axis, but not at x = 0.
A typical formulation of the map Gμ is

a μ + cxδ , if x > 0;
Gμ (x) =
−a μ − c|x|δ , if x < 0;

Fig. 2.29 Homoclinic orbits in the Lorenz model

70 2 Preliminaries of Nonlinear Dynamics and Chaos

–α 0 α
Fig. 2.30 One-dimensional map for the Lorenz model

where a < 0, c > 0, δ = −λ3 /λ1 , and 0 < δ < 1.

Assuming that the one-dimensional approximation remains valid for values of the
parameter μ outside the neighborhood of μ = μh (that is, outside the neighborhood
of the homoclinic orbit), values of μ > μh can be chosen so that there exists a closed
interval [−α , α ] with α > 0 such that Gμ [−α , α ]\{0} → [−α , α ] and

lim Gμ (x) = α > 0, lim Gμ (x) = −α < 0.

x→0− x→0+

Then, Gμ (x) > 1 for x ∈ [−α , α ] (x = 0), and limx→0± Gμ (x) = +∞. The map Gμ
on the interval is depicted in Fig. 2.30.
Because Gμ (x) > 1 for all x ∈ [−α , α ] (x = 0), Gμ is a piecewise-expanding
map and has therefore sensitive dependence on initial conditions. There are no fixed
points or stable periodic points and most orbits on the interval [−α , 0) ∪ (0, α ] are
attracted to a chaotic invariant set. Although increasing μ beyond the homoclinic
value μh leads to stable chaotic motion, if we take μ very large, the system reverts
to simpler dynamical behavior and stable periodic orbits reappear.

Remark 2.14. The idea that some essential aspects of the dynamics of the original
system (2.33) could be described by a one-dimensional map was first put forward
by Lorenz himself. In order to ascertain whether the numerically observed attractor
could be periodic rather than chaotic, he plotted successive maxima of the variable
z along an orbit on the numerically observed attractor. In doing so, he discovered
that the plot of zn+1 against zn has a simple shape, as illustrated in Fig. 2.31. The
points of the plot lie almost exactly on a curve whose form changes as the parameter
r varies. Setting σ = 10, b = 8/3, and r = 28 (the traditional ‘chaotic values’), we
2.15 Basics of Functional Differential Equations Theory 71

48

46

44

42

40
zn+1

38

36

34

32

30
30 32 34 36 38 40 42 44 46 48
zn

Fig. 2.31 Successive maxima of z for the Lorenz attractor

obtain a curve resembling a distorted tent map. It has slope everywhere greater than
1 in absolute value so, again, it approximates a piecewise-expanding map. For such
a map there cannot be stable fixed points or stable periodic orbits and, for randomly
chosen initial values, orbits converge to a chaotic attractor. 

2.15 Basics of Functional Differential Equations Theory

In many applications, the system under consideration is not governed by a princi-

ple of causality; that is, the future state of the system is dependent not only on the
present state but also on the past states. The theory about these systems has been
extensively developed, which formed the framework of the subject – functional dif-
ferential equations (FDEs). In this section, we only provide some necessary notions
and theorems about FDEs so that this book can be read smoothly.
Let R = (−∞, ∞) and R+ = [0, ∞). Let C = C([−τ , 0], Rn ) denote the space of
continuous functions mapping the interval [−τ , 0] into R. We designate the norm of
an element φ in C by φ τ = sup−τ ≤θ ≤0 φ (θ ) . Each x ∈ Rn can also be looked as
an element in C: x(θ ) = x, θ ∈ [−τ , 0]. If σ ∈ R, α ≥ 0, and x ∈ C[σ − τ , σ + α ],
then, for any t ∈ [σ , σ + α ], we let xt ∈ C be defined by xt (θ ) = x(t + θ ), −τ ≤ θ ≤ 0.
If D is a subset of R ×C, f : D → Rn is a given function, and ‘·’ represents the right-
hand derivative, we say that the relation
72 2 Preliminaries of Nonlinear Dynamics and Chaos

ẋ = f (t, xt ) (2.35)

is a retarded functional differential equation on D and will denote this equation by

RFDE. We write RFDE( f ) if we wish to emphasize that the equation is defined by
f.
Definition 2.26. A function x is said to be a solution of (2.35) on [σ − τ , σ + α ] if
there are σ ∈ R and α > 0 such that x ∈ C([σ − τ , σ + α ], Rn ), (t, xt ) ∈ D, and x(t)
satisfies (2.35) for t ∈ [σ , σ + α ). For given σ ∈ R, φ ∈ C, we say that x(σ , φ , f ) is
a solution of (2.35) with initial value φ at σ or simply a solution through (σ , φ ) if
there is an α > 0 such that x(σ , φ , f ) is a solution of (2.35) on [σ − τ , σ + α ] and
xσ (σ , φ , f ) = φ . 
In the following we only state the main results; the proofs can be found in [8].
Theorem 2.18 (Existence). Suppose that Ω is an open subset in R × C and f 0 ∈
C(Ω , Rn ). If (σ , φ ) ∈ Ω , then there is a solution of the RFDE( f 0 ) passing through
(σ , φ ). More generally, if W ⊆ Ω is compact and f 0 ∈ C(Ω , Rn ) is given, then there
is a neighborhood V ⊆ Ω of W such that f 0 ∈ C0 (V, Rn ), there is a neighborhood
U ⊆ C0 (V, Rn ) of f 0 , and an α > 0 such that, for any (σ , φ ) ∈ W , f ∈ U, there is a
solution x(σ , φ , f ) of the RFDE( f ) through (σ , φ ) which exists on [σ − τ , σ + α ].


Theorem 2.19 (Continuous Dependence). Suppose that Ω ⊆ R×C is open, (σ 0 , φ 0 ) ∈

σ , f 0 ∈ C(Ω , Rn ), and x0 is a solution of the RFDE( f 0 ) through (σ 0 , φ 0 ) which ex-
ists and is unique on [σ 0 − τ , b]. Let W 0 ⊆ Ω be the compact set defined by

W 0 = {(t, xt0 )|t ∈ [σ 0 , b]}

and let V 0 be a neighborhood of W 0 on which f 0 is bounded. If (σ k , φ k , f k ), k =

1, 2, . . . , satisfies σ k → σ 0 , φ k → φ 0 , and f k − f 0 V 0 → 0 as k → ∞, then there is a
k0 such that the RFDE( f k ) for k ≥ k0 is such that each solution xk = xk (σ k , φ k , f k )
through (σ k , φ k ) exists on [σ k − τ , b] and xk → x0 uniformly on [σ 0 − τ , b]. Since
all xk may not be defined on [σ 0 − τ , b], by xk → x0 uniformly on [σ 0 − r, b], we
mean that, for any ε > 0, there is a k1 (ε ) such that xk (t), k ≥ k1 (ε ), is defined on
[σ 0 − τ + ε , b], and xk → x0 uniformly on [σ 0 − τ + ε , b]. 

Theorem 2.20. Suppose that Ω is an open set in R × C, f : Ω → Rn is continuous,

and f (t, φ ) is Lipschitzian in φ in each compact set in Ω , i.e.,

f (t, φ ) − f (t, ψ ) ≤ K φ − ψ ,

for arbitrary φ , ψ ∈ C, where K is a constant. If (σ , φ ) ∈ Ω , then there is a unique

solution of (2.35) through (σ , φ ). 

Definition 2.27. If, for all (σ , φ ) ∈ R × C, the solution of (2.35) through (σ , φ ),

x(σ , φ )(t), exists on [σ − τ , ∞), then we say that (2.35) has global solutions. 
2.15 Basics of Functional Differential Equations Theory 73

Theorem 2.21. If there exist continuous functions M, N : R → R+ such that

f (t, φ ) ≤ M(t) + N(t) φ , (t, φ ) ∈ R × C,

then (2.35) has global solutions. 

In the following, we introduce the stability theory for (2.35). Analogous to what
we have done with ordinary differential equations, the stability of a general solution
x(t) of (2.35) is equivalent to the stability of the zero solution of a new RFDE.
Therefore, without loss of generality, we assume that f (t, 0) = 0 for all t ∈ R.
Definition 2.28.
(i) The solution x = 0 of (2.35) is said to be stable if, for any σ ∈ R and ε > 0,
there exists a δ = δ (ε , σ ) > 0 such that φ τ < δ implies that x(t, σ , φ ) < ε
for t ≥ σ .
(ii) The solution x = 0 of (2.35) is said to be uniformly stable if the δ in (i) is
independent of σ .
(iii) The solution x = 0 of (2.35) is said to be attractive if, for any σ ∈ R, there exists
a b = b(σ ) such that φ τ ≤ b implies that x(t, σ , φ ) → 0 (t → ∞). That is,
for any ε > 0 and φ τ ≤ b, there exists a T (σ , ε , φ ) such that x(t, σ , φ ) < ε
whenever t ≥ σ + T (σ , ε , φ ). If b = +∞, the solution x = 0 of (2.35) is said to
be globally attractive.
(iv) The solution x = 0 of (2.35) is said to be asymptotically stable if it is both stable
and attractive.
(v) The solution x = 0 of (2.35) is said to be uniformly attractive if in (iii) b is
independent of σ and T only depends on ε .
(vi) The solution x = 0 of (2.35) is said to be uniformly asymptotically stable if it is
both uniformly stable and uniformly attractive.
(vii) The solution x = 0 of (2.35) is said to be globally asymptotically stable if it is
stable and globally attractive. 
Definition 2.29. The solution x = 0 of (2.35) is said to be exponentially stable if
there exists a β > 0 and, for any ε > 0, there is a δ = δ (ε ) > 0 such that, for any
σ ∈ R, φ τ < δ implies that

x(t, σ , φ ) ≤ ε exp[−β (t − σ )]

for t ≥ σ . The solution x = 0 of (2.35) is said to be globally exponentially stable if

there exist a β > 0 and an η > 0 such that, for any (σ , φ ) ∈ R × C, the following
inequality holds:
x(t, σ , φ ) ≤ η φ τ exp[−β (t − σ )]
for t ≥ σ . 
Definition 2.30. A solution x(t, σ , φ ) of (2.35) is bounded if there is a β (σ , φ ) such
that x(t, σ , φ ) < β (σ , φ ) for t ≥ σ − τ . The solution is uniformly bounded if, for
any α > 0, there is a β = β (α ) > 0 such that, for all σ ∈ R, φ ∈ C, and φ ≤ α ,
we have x(t, σ , φ ) ≤ β (α ) for all t ≥ σ . 
74 2 Preliminaries of Nonlinear Dynamics and Chaos

In the following, we introduce some sufficient conditions for the stability of the
solution x = 0 of (2.35) which generalize the second method of Lyapunov for ordi-
nary differential equations.
If the functional V : R × C → R+ is continuous and x(t, σ , φ ) is the solution of
(2.35) through (σ , φ ), we define

1
V̇ (t, φ ) = lim [V (t + h, xt+h (t, φ )) − V (t, φ )].
h→0+ h

The function V̇ (t, φ ) is the upper right-hand derivative of V (t, φ ) along the solution
of (2.35).
If the function V : R × R → R+ is continuous, we define
1
V̇ (t, φ (0)) = lim [V (t + h, x(t, φ )(t + h)) − V(t, φ (0))].
h→0+ h

The function V̇ (t, φ (0)) is the upper right-hand derivative of V (t, x) along the solu-
tion of (2.35).
Sometimes we write V̇(2.35) (t, φ ) and V̇(2.35) (t, φ (0)) to emphasize the dependence
on (2.35), respectively.
Theorem 2.22. Suppose that f : R × C → Rn takes R× (bounded set of C) into
bounded sets of Rn , and u, v, w : R+ → R+ are continuous nondecreasing functions,
u(s) and v(s) are positive for s > 0, and u(0) = v(0) = 0. If there is a continuous
functional V : R × C → R+ such that

u( φ (0) ) ≤ V (t, φ ) ≤ v( φ τ ),

V̇(2.35) (t, φ ) ≤ −w( φ (0) ),

then the solution x = 0 of (2.35) is uniformly stable. If u(s) → ∞ as s → ∞, the
solutions of (2.35) are uniformly bounded. If w(s) > 0 for s > 0, then the solution
x = 0 is uniformly asymptotically stable. 

Corollary 2.1. Suppose that f : R × C → Rn takes R× (bounded set of C) into

bounded sets of Rn , and u, w : R+ → R+ are continuous nondecreasing functions,
u(s) and w(s) are positive for s > 0, u(0) = w(0) = 0, and u(s) → ∞ (s → ∞). If
there is a continuous functional V : R × C → R+ such that

u( φ (0) ) ≤ V (t, φ ), V (t, 0) = 0,

V̇(2.35) (t, φ ) ≤ −w( φ (0) ),

then the solution x = 0 of (2.35) is globally asymptotically stable. 

Theorem 2.23. Suppose that f : R × C → Rn takes R× (bounded set of C) into

bounded sets of Rn , and u, v, w : R+ → R+ are continuous nondecreasing functions,
u(s), v(s), and w(s) are positive for s > 0, and u(0) = v(0) = 0. Suppose that P(s) is
2.16 Summary 75

a continuous and nondecreasing function satisfying P(s) > s for s > 0. If there is a
continuous function V : R × Rn → R+ such that

u( x ) ≤ V (t, φ ) ≤ v( x )

and

V̇(2.35) (t, φ (0)) ≤ −w( φ (0) ) when V (t + θ , φ (θ )) < P(V (t, φ (0))), θ ∈ [−τ , 0],

then, the solution x = 0 of (2.35) is uniformly asymptotically stable. Moreover, if

u(s) → ∞ as s → ∞, then the solution x = 0 of (2.35) is globally asymptotically
stable. 

Theorem 2.24. Suppose that α , β , p, and μ are positive constants. If there is a

continuous function V : R × Rn → R+ such that
p p
α x ≤ V (t, x) ≤ β x

and

V̇(2.35) (t, φ (0)) ≤ −μ V (t, φ (0)) when sup [eμθ V (t + θ , x(t + θ ))] = V (t, x(t)),
−τ ≤θ ≤0

then, the solution x = 0 of (2.35) is globally exponentially stable. 

2.16 Summary

One basic goal in studying dynamical systems is to explore how the trajectories of a
system evolve as time proceeds. So, in this chapter, we began with the theorems on
existence and uniqueness of solutions of ordinary differential equations. In the sub-
sequent sections we focused on a special class of solutions: equilibrium solutions.
Besides that notion, we introduced other definitions such as fixed point, periodic
orbit, quasiperiodic orbit, ω -limit set, invariant set, etc., and, furthermore, the key
concepts of the book, chaos and chaotic attractors, were introduced. Two power-
ful tools in studying chaotic systems, Lyapunov exponents and symbolic dynamics,
were discussed briefly. The stability issue was discussed and a detailed category of
the types of fixed point of planar systems for both continuous time and discrete time
was provided. Three famous examples on chaos were carefully presented through
which we wanted to give a concrete understanding about chaos. Finally, we pro-
vided some necessary preliminaries on retarded functional differential equations for
the purpose of self-containment of the book.
76 2 Preliminaries of Nonlinear Dynamics and Chaos

References

1. Alligood KT, Sauer TD, Yorke JA (1997) Chaos – An Introduction to Dynamical Systems.
Springer, New York
2. Arnold VI (1988) Geometrical Methods in the Theory of Ordinary Differential Equations,
2nd edn. Springer, New York
3. Banks J, Brooks J, Cairns G, Davis G, Stacey P (1992) On Devaney’s definition of chaos. Am
Math Mon 99:332–334
4. Birkhoff GD (1927) Dynamical Systems. American Mathematical Society, Providence, RI
5. Devaney RL (1989) An Introduction to Chaotic Dynamical Systems, 2nd edn. Addison-
Wesley, New York
6. Grebogi C, Hammel SM, Yorke JA, Sauer T (1990) Shadowing of physical trajectories in
chaotic dynamics: containment and refinement. Phys Rev Lett 65:1527–1530
7. Guckenheimer J, Holmes P (1983) Nonlinear Oscillations, Dynamical Systems, and Bifurca-
tions of Vector Fields. Springer, New York
8. Hale J (1977) Theory of Functional Differential Equations. Springer, New York
9. Hirsch MW, Smale S (1974) Differential Equations, Dynamical Aystems and Linear Algebra.
Academic Press, New York
10. Kuznetsov YA (1998) Elements of Applied Bifurcation Theory, 2nd edn. Springer, New York
11. Medio A, Lines M (2001) Nonlinear Dynamics: A Primer. Cambridge University Press, Lon-
don
12. Parker TS, Chua LO (1989) Practical Numerical Algorithms for Chaotic Systems. Springer,
New York
13. Robinson RC (2004) An Introduction to Dynamical Systems: Continuous and Discrete. Pren-
tice Hall, Upper Saddle River, NJ
14. Sauer T, Grebogi C, Yorke JA (1997) How long do numerical chaotic solutions remain valid?
Phys Rev Lett 79:59–62
15. Tucker W (2002) A rigorous ODE solver and Smale’s 14th problem. Found Comput Math
2:53–117
16. Vellekoop M, Berglund R (1994) On intervals: transitivity → chaos. Am Math Mon 101:353–
355
17. Wiggins S (2003) Introduction to Applied Nonlinear Dynamical Systems and Chaos, 2nd edn.
Springer, New York
Chapter 3
Entrainment and Migration Control of Chaotic
Systems

Abstract In this chapter, the entrainment and the migration methods of chaos con-
trol are discussed. The background of entrainment and migration control is based
on two facts: a multi-attractor chaotic system is sensible to both initial values and
parameters and each stable attractor has its basin of attraction. A common goal of
chaos control is to steer the trajectory of a chaotic system to a periodic orbit which
itself is a solution of the chaotic system. Sometimes, one wants to steer the trajec-
tory to a predesigned orbit which is not the solution of the chaotic system. For some
multiple attractor chaotic systems, one would like to direct the trajectory from one
attractor to the others. To achieve these goals, entrainment and migration methods
are developed. We first introduce these methods and their extension, the open-plus-
closed-loop (OPCL) method. Based on the OPCL method, a new control scheme,
the open-plus-nonlinear-closed-loop (OPNCL) method, is developed. In addition,
we apply the method of OPNCL to a class of dynamical systems, both continuous
time and discrete time.

3.1 Introduction

Usually, several attractors can coexist in a complex dynamical system. These attrac-
tors can be stable or unstable. When a key parameter is changed the attractor may
be altered in either appearance or spatial position, or in both. The background of
entrainment and migration control is based on two facts: a multi-attractor chaotic
system is sensible to both initial values and parameters and each stable attractor has
its basin of attraction. A common goal of chaos control is to steer the trajectory of
a chaotic system to a periodic orbit which itself is a solution of the chaotic sys-
tem. Sometimes, one wants to steer the trajectory to a predesigned orbit which is
not the solution of the chaotic system. For some multiple attractor chaotic systems,
one would like to direct the trajectory from one attractor to others. To achieve these
goals, entrainment and migration methods are proposed. In this chapter, we will in-
troduce these methods and their expansion, the OPCL method. Based on the OPCL

77
78 3 Entrainment and Migration Control of Chaotic Systems

method, a new control scheme, the OPNCL method, is proposed. Besides that, we
will apply the method of OPNCL to a class of dynamical systems, both continuous
time and discrete time.

3.2 Basics on Entrainment and Migration

3.2.1 Entrainment Control

Consider the following dynamical systems:

xn+1 = f (xn ) (3.1)

or
ẋ = f (x), (3.2)
where x ∈ Rm and f : Rm → Rm is smooth. Sometimes, it is difficult to measure and
feed-back the state variables. Moreover, sometimes one wishes to steer the trajectory
of systems (3.1) and (3.2) to a certain region in the phase space, or even to some
predesigned orbits. For example, the target orbits can be given by

xn = g n (3.3)

or
x(t) = g(t). (3.4)
These target orbits may be fixed points, periodic orbits, nonperiodic orbits, or even
chaotic orbits. We would like to achieve the specified target, for example, (3.3), by
a control signal Ŝn . That is to say, the following controlled discrete-time system:

xn+1 = f (xn ) + Ŝn

takes (3.3) as its solution. Thus, we must have that

xn+1 = f (xn ) + Ŝn,

(3.5)
Ŝn = gn+1 − f (gn ).

For the case of continuous time, we have analogous equations

ẋ = f (x) + Ŝ(t),
(3.6)
Ŝ(t) = ġ(t) − f (g(t)).

It should be noted that Ŝn and Ŝ(t) are independent of the state variables of sys-
tems (3.1) and (3.2). This control method belongs to an open-loop control scheme.
Thus, the targets (3.3) and (3.4) are not arbitrary. They should be stable solutions of
3.2 Basics on Entrainment and Migration 79

the controlled systems. We take (3.6) as an example. If g(t) is a periodic orbit, its
stability can be determined by the vector field

δ ẋ = D fx δ x, (3.7)

where D fx = (∂ f /∂ x)|x=g(t) and δ x = x(t)− g(t). Because g(t) and f (x) are known,
the conditions to hold for stability are that all Floquet characteristic multipliers are
negative. Thus, we can select a certain region C, of the state space, so that the dis-
tance between any couple of points in this region is contracted by (3.7). This kind of
region is called a convergence region. In this region, for any distance, Δ x, it always
holds that
Δ xT D fx Δ x < 0. (3.8)
Therefore, if g(t) is in this region, it will be stable. For any bounded dissipative
system this region always exists. In the discrete-time case, the variation equation
given in (3.7) becomes

δ xn+1 = D fx δ xn , δ xn = xn − gn .

The condition (3.8) becomes

|Δ xT D fx Δ x| < Δ xT Δ x. (3.9)

Once the target has been selected, the next step is to determine BE(g), i.e., the basin
of entrainment (the range of initial conditions of x(t) that will satisfy the condition
limt→∞ x(t) − g(t) = 0). This step is far more difficult than determining a stable
target. A point in a contractive region must converge to the target orbit but, compared
with the basin of entrainment, this contractive region is very small. In general, we
cannot get an analytical expression for a basin of entrainment, but we can get a
rough understanding of a basin of entrainment by means of numerical simulations
or experiments.
Now, it is clear how the entrainment method is implemented. First, a target orbit
g(t) (or gn ) should be selected according to the requirements of the stability in prac-
tice. The basin of entrainment of the target orbit must cover a part of the attractor of
the chaotic system to be controlled. Then, the controlled systems are as follows:
(i) for the discrete-time case, we have

xn+1 = f (xn ) + (gn+1 − f (gn ))Sn ; (3.10)

(ii) for the continuous-time case, we have

ẋ = f (x) + (ġ(t) − f (g(t)))S(t). (3.11)

Here, S(t) = 0 when the controller does not work and S(t) = 1 when the controller
begins to work. The same definition is for Sn as well. The controller is turned on
when the trajectory of the chaotic system goes into the overlapping region of the
80 3 Entrainment and Migration Control of Chaotic Systems

basin of entrainment and the chaotic attractor. Since these kinds of controllers were
first proposed by Hübler [3], we will refer to them as Hübler open-loop controllers.
There are some distinguishable merits of the entrainment method. The selection
of the target orbits is very flexible. The target orbits can be ones which are not solu-
tions of the chaotic systems, which enhances the ability of control and, meanwhile,
avoids real-time monitoring and feedback. Therefore, the controller would be sim-
plified significantly. However, there are some problems with this method. On one
hand, a total understanding about the original chaotic system is required when one
wants to determine the control signals. This is not always achievable. On the other
hand, the control signals usually have large amplitude and the original dynamics
changes in a large range. It should be emphasized that both methods (3.10) and
(3.11) cannot stabilize unstable periodic orbits. From (3.7) and (3.9) it is easy to see
that the stability of the target orbits is not changed by the controller.

3.2.2 Attractor Selection by Migration Control

One interesting point with entrainment control is that g(t) and gn are not required to
be periodic orbits. They may be a migration orbit from one point in the state space
to another. Thus, we can use entrainment control to achieve an important control
task, the migration of state.
One of the features of nonlinear systems is that there usually exist multiple attrac-
tors corresponding to some parametric conditions. Since each attractor has its own
basin of attraction, different initial conditions will result in different final states.
If one wants to migrate an orbit from a chaotic attractor to a periodic orbit, or to
migrate a periodic orbit to another periodic orbit, then the migration orbit between
different points in the state space is a reasonable selection.
Suppose that g(t) is the selected orbit which is used to migrate g(0) = xS to
g(t → ∞) = xT . We require that

lim x(t) − g(t) = 0.

t→∞

The control equation that should be obeyed by the target is (3.11). Since the starting
point of the migration orbit can be located at the chaotic attractor, the whole under-
standing about the distribution of the basin of entrainment is not required. The only
requirement is the stability of g(t). After the starting attractor and the terminal at-
tractor are determined, we should select the starting point xS and the terminal point
xT of the migration orbit. The basin of attraction may be quite big, which makes the
selection of xT very flexible. Since there is much freedom in the selection of g(t), it
is not difficult to guarantee local stability.
We take the Lorenz system as an example. The Lorenz system is of the following
form:
3.2 Basics on Entrainment and Migration 81
⎧
⎨ ẋ1 = σ (x2 − x1 ),
ẋ = ρ x1 − x1 x3 − x2 , (3.12)
⎩ 2
ẋ3 = x1 x2 − β x3.
Let g = (g1 (t), g2 (t), g3 (t))T be a migration orbit. Adding the Hübler open-loop
control, (3.12) becomes
⎧
⎨ u̇1 = σ (u2 − u1 ),
u̇ = ρ u1 − u2 − u1u3 − u1g3 − g1 u3 , (3.13)
⎩ 2
u̇3 = u1 u2 − β u3 + u1 g2 + u2g1 ,

where ui = xi (t) − gi (t), i = 1, 2, 3. The stability of g(t) can be guaranteed by requir-

ing that there is a positive-definite function V (u) such that V̇ (u) is negative definite.
One such function can be chosen as
1
V (u) = (γ u21 + u22 + u23),
2
where γ is a positive constant to be specified. If we set

a2 + β b2 = γσ , 2a = γσ + ρ − g3 (t), 2β b = g2 (t), γ̇ = 0, (3.14)

then we have that

V̇ (u) = (γσ + ρ − g3 (t))u1 u2 − γσ u21 − u22 − β u23 + g2(t)u1 u3

= −(au1 − u2 )2 − β (bu1 − u3 )2 . (3.15)

From (3.14) we get

⎧
⎪  1
⎪
⎪ g22 (t) 2
⎪
⎪ a(t) = 1 + 1 + g3(t) − ρ − ,
⎪
⎪ 4β
⎪
⎨
g2 (t)
⎪
⎪ b(t) = ,
⎪
⎪ 2β
⎪
⎪
⎪
⎪ a2 (t) + β b2(t)
⎩ γ (t) = .
σ
In order to make (3.15) negative, a(t) must be a real number, which leads to

g22 (t)
g3 (t) > ρ − 1 + . (3.16)
4β

Noting that γ̇ = 0, we get another restriction on g(t):

a(t)ġ3 (t) − g2(t)ġ2 (t)/2β = 0. (3.17)

Thus, any orbit which satisfies conditions (3.16) and (3.17) must be stable. Fur-
thermore, its basin of entrainment is the whole state space. Conditions (3.16) and
82 3 Entrainment and Migration Control of Chaotic Systems

(3.17) are just sufficient conditions, not necessary ones. When σ = 10, β = 8/3,
and 14 < ρ < 24, the Lorenz system has two attractors, and they are
 
A+ : ( β (ρ − 1), β (ρ − 1), ρ − 1),
 
A− : (− β (ρ − 1), − β (ρ − 1), ρ − 1).

Suppose that the Lorenz system starts from A+ . We want to migrate the trajectory
of the Lorenz system to A− along the migration orbit g(t). We select the following
migration orbit:

g1 (t) = g2 (t) = [β (ρ − 1)]1/2(2e−t/5 − 1),

(3.18)
g3 (t) = (ρ − 1){0.05 sin[π (1 − e−t/5 )] + 1}.

The orbit (3.18) does not satisfy condition (3.17), but it is stable. The initial point is
selected in the basin of attraction of A+ . The controller is turned on at t = 0. There-
fore, the trajectory of the Lorenz system is entrained to the basin of attraction of A− .
Then, the controller is turned off. The trajectory goes to A− under the domination
of the dynamics of the Lorenz system. This procedure is shown in Fig. 3.1.
If the values of σ and β are fixed but the value of ρ varies in the range
[24.1, 24.7], a chaotic attractor appears besides the two fixed points. We take two
steps to entrain the trajectory of the Lorenz system from the chaotic attractor to A− .

A+
6

2 The migration orbit

x2

−2

−4

A−
−6

−8
−8 −6 −4 −2 0 2 4 6 8
x1

Fig. 3.1 Migrating the trajectory of the Lorenz system from A+ to A−

3.2 Basics on Entrainment and Migration 83

First, we choose a fixed point

1
g∗1 = g∗2 = [β (ρ − 1)] 2 , g∗3 = (ρ − 1) + 50

in the convergent region as the goal of entrainment. From (3.15) to (3.17) we know
that this point is the unique fixed point of system (3.13). So, whenever the controller
is turned on, the trajectory of the Lorenz system will be entrained to this point.
Therefore, the difficulties encountered in the selection of the target orbit, which are
caused by the complex structure of the basin of entrainment of each attractor, can
be avoided. Second, we select the following migration orbit [2]:

g1 (t) = g2 (t) = [β (ρ − 1)]1/2[2e−(t−t0 )/5 − 1],

g3 (t) = (ρ − 1) + 50e−(t−t0 )/5 , t ≥ t0 ,

where t0 is the instant at which the entrainment to g∗ is finished and the migration
from g∗ to A− is started. The simulation result is shown in Fig. 3.2.

The migration orbit

g∗
The migration orbit

80

60
A− 40
40
x3

20 20

0 0
−20
−15
−10 −20
−5
0
5
10 x2
15
20 −40
25
x1

Fig. 3.2 Migrating the trajectory of the Lorenz system from the chaotic attractor to A−
84 3 Entrainment and Migration Control of Chaotic Systems

3.3 OPCL Control Scheme

Although Hübler’s controller is successfully used in many applications, there are

cases in which this method cannot be implemented. It was believed that all of the
strange attractors had the property that these attractors intersected some convergent
region in the phase space. Therefore, it was conjectured that this might be a general
feature of strange attractors. This would have been nice (but not essential), since the
initial state of a system on a strange attractor could then be selected arbitrarily close
to g(t), and hence certainly in its basin of entrainment. However, the Chua oscillator
is a notable exception whose strange attractor can lie entirely in a nonconvergent
region. Therefore, the Chua oscillator cannot be entrained near the strange attractor
using the entrainment method (however, it can be entrained outside the attractor’s
region). Moreover, linear feedback can only stabilize the oscillator to one of the
unstable periodic orbits. To overcome the limitations of both the open-loop and the
closed-loop control methods, the OPCL method was proposed.
Theorem 3.1. Let g(t) be an arbitrary smooth function, and x(t) be a solution of

ẋ = F(x,t) + S(t)K(g, x,t) (x ∈ Rn ), (3.19)

where S(t) = 0 (if t < t0 ) and S(t) = 1 (if t ≥ t0 ), and F(x,t) is everywhere Lips-
chitzian. There exists a function K(g, x,t), linear in x(t), such that none of the basins
of entrainment (associated with g(t) and t0 ), defined by

BE(g|t0 ) = {x(t0 ) : lim x(t) − g(t) = 0}, (3.20)

t→∞

are empty sets.

Proof. Consider a special kind of control, K(g,t), which is the sum of Hübler’s
open-loop control and a particular form of a linear closed-loop control, namely

K(g,t) = ġ − F(g,t) + C(g,t)(g(t) − x(t)), (3.21)

where the matrix C(g,t) is given by

∂ F(g,t)
C(g,t) = −A (3.22)
∂g

and A = (ai j ) is any constant matrix whose eigenvalues all have negative real parts.
The control (3.21) and (3.22) is referred to as the OPCL control.
Set x(t) = g(t) + u(t), where u(t0 ) is arbitrarily small. Substitute x(t) into (3.19)
when t ≥ t0 , and expand F(g(t) + u(t),t) for small u(t). This yields the following
variational equation:

du ∂ F(g,t)
= u(t) − C(g,t)u(t) = Au(t). (3.23)
dt ∂g
3.4 Global Control of a Class of Continuous-Time Polynomial Chaotic Systems 85

Since all eigenvalues of A have negative real parts, the asymptotic stability of (3.23)
is established. Since this asymptotic stability also holds for the nonlinear system
(3.19), BE(g) is not an empty set. 

Remark 3.1. This is a remarkably general theorem, regardless of whether g(t) is or

is not a solution of ġ = F(g,t), and it does not restrict g(t) to any region of the
phase space. This latter point is particularly important for the possible application
of this kind of control for the purpose of transferring a system from one attractor to
another (‘migration control’). What is not established by the theorem is the range of
the basin of entrainment, BE(g), in the phase space. It is to be expected that BE(g)
will depend on both the particular dynamical system and some characteristics of
g(t). Indeed, substituting x(t) = g(t) + u(t) into (3.19) and expanding F(g + u,t),
we get
∞
du 1 ∂ k F(g,t) k
= Au(t) + ∑ u.
dt k=2 k! ∂ gk
The above discussion can easily be seen from this formula. 

3.4 Global Control of a Class of Continuous-Time Polynomial

Chaotic Systems

3.4.1 OPNCL Control Scheme

The OPCL control is very effective for systems whose right-hand sides of differen-
tial equations are mth-order polynomials, where m ≤ 2. When m > 2, it is difficult
to determine the basin of entrainment, which will be very complex with the increase
of m. To overcome this difficulty, an improved OPCL method, called the OPNCL
method, was proposed [5].
Consider the following mth-order polynomial systems:

ẋ = F(x,t) (x ∈ Rn ) (3.24)

or
m
∑ j j
a j1 j2 ... jn (t)x11 x22 · · · xnjn (i = 1, 2, . . . , n).
(i)
ẋi = Fi (x,t) = (3.25)
j1 + j2 +···+ jn =0

Without loss of generality, we suppose that there is at least one nonzero parameter,
(i)
for example, a j1 j2 ... jn (t) = 0 (1 ≤ i ≤ n). For system (3.24) we add the following
OPNCL controller:

K(g, x,t) = ġ − F(g,t) + D(g, x,t)(g(t) − x(t)), (3.26)

86 3 Entrainment and Migration Control of Chaotic Systems

where D(g, x,t) is a nonlinear function of g(t) − x(t); its concrete form will be given
later. Adding the control term K(g, x,t) to system (3.24) through a switching func-
tion S(t), we get
ẋ = F(x,t) + S(t)K(g, x,t), (3.27)
where K(g, x,t) is specified by (3.26). For an arbitrary target g(t), we will prove that
the basin of entrainment of (3.24) is not an empty set. Furthermore, the entrainment
is global. Set S(t) = 1 and let x(t) = g(t) + u(t), where u(t) ∈ Rn . Substituting x(t)
into (3.27), we get
ġ + u̇ = F(g + u,t) + K(g, g + u,t).
Substituting (3.26) into the above equation, we get
m
1 ∂ k F(g,t) k
u̇ = F(g + u,t) − F(g,t) − D(g, x,t)u = ∑ k! ∂ gk
u − D(g, x,t)u. (3.28)
k=1

∂ k F(g,t) k
To calculate u , we note that F = (F1 , F2 , . . . , Fn )T and Fi = Fi (x1 , x2 , . . . ,
∂ gk
xn ). Thus, we have

Fi (x1 + h1, x2 + h2, . . . , xn + hn ) − Fi(x1 , x2 , . . . , xn )

n n n
1 ∂ Fi 1 ∂ 2 Fi 1 ∂ 3 Fi
= ∑ ∂ x j h j + 2! ∑ ∂ x j ∂ xk h j hk + 3! ∑ h j hk hl + · · · .
1! j=1 j,k=1 j,k,l=1 ∂ x j ∂ xk ∂ xl

Note that
n
∂ 2 Fi ∂ 2 Fi 2 ∂ 2 Fi ∂ 2 Fi
∑ h j hk = h1 + h1 h2 + · · · + h1 hn
j,k=1 ∂ x j ∂ xk ∂ x1 ∂ x1 ∂ x1 ∂ x2 ∂ x1 ∂ xn
∂ 2 Fi ∂ 2 Fi 2 ∂ 2 Fi
+ h2 h1 + h2 + · · · + h2 hn
∂ x2 ∂ x1 ∂ x2 ∂ x2 ∂ x2 ∂ xn
+ ···
∂ 2 Fi ∂ 2 Fi ∂ 2 Fi 2
+ hn h1 + hn h2 + · · · + h
∂ xn ∂ x1 ∂ xn ∂ x2 ∂ xn ∂ xn n
∂ 2 Fi ∂ 2 Fi ∂ 2 Fi
= h1 + h2 + · · · + hn ,
∂ x1 ∂ x1 ∂ x1 ∂ x2 ∂ x1 ∂ xn
∂ 2 Fi ∂ 2 Fi ∂ 2 Fi
h1 + h2 + · · · + hn ,
∂ x2 ∂ x1 ∂ x2 ∂ x2 ∂ x2 ∂ xn
⎛ ⎞
h1
∂ 2 Fi ∂ 2 Fi ∂ 2 Fi ⎜ h2 ⎟
⎜ ⎟
..., h1 + h2 + · · · + hn ⎜ .. ⎟
∂ xn ∂ x1 ∂ xn ∂ x2 ∂ xn ∂ xn ⎝ . ⎠
hn
3.4 Global Control of a Class of Continuous-Time Polynomial Chaotic Systems 87
⎡ ⎛ ⎞
h1
⎢ ∂ 2F ∂ 2 Fi ∂ 2 Fi ⎜ h2 ⎟
⎢ i ⎜ ⎟
=⎢ , ,..., ⎜ .. ⎟ ,
⎣ ∂ x1 ∂ x1 ∂ x1 ∂ x2 ∂ x1 ∂ xn ⎝ . ⎠
hn
⎛ ⎞⎤ ⎛ ⎞
h1 h1
⎜ ⎟ ⎥ ⎜ ⎟
2
∂ Fi 2
∂ Fi 2
∂ Fi ⎜ 2 ⎟⎥ ⎜ h 2 ⎟
h
··· , , ,..., ⎜ .. ⎟⎥ ⎜ .. ⎟ .
∂ xn ∂ x1 ∂ xn ∂ x2 ∂ xn ∂ xn ⎝ . ⎠ ⎦ ⎝ . ⎠
hn hn

∂ 3 Fi
Analogously, we get the expressions of h j hk hl , . . . .
∂ x j ∂ xk ∂ xl
To entrain the solution of system (3.24) to the target orbit g(t), we require that
the condition
lim (x(t) − g(t)) = 0
t→∞

holds. This will require that limt→∞ u(t) = 0 holds. So, we set
m
1 ∂ k F(g,t)
D(g, x,t) = ∑ k! ∂ gk
(x(t) − g(t))k−1 − A, (3.29)
k=1

where A = (ai j ) is an n × n constant matrix with all its eigenvalues having negative
real parts. Substituting (3.29) into (3.28), we get a linear differential equation with
respect to u(t):
u̇(t) = Au(t). (3.30)
The zero solution of this system is asymptotically stable, which implies that the
solution of (3.27) is asymptotically stable. According to (3.20), the definition of
BE(g), we know that BE(g) is not empty. Since the stability of u(t) is independent
of initial conditions, BE(g) is global, which means that, for any starting point x0 ,
the relation limt→∞ x(t) − g(t) = 0 always holds.
By summarizing the above analysis, we obtain the following theorem.
Theorem 3.2 ([5]). Suppose that g(t) ∈ Cr (r ≥ 1) and x(t) is the solution of system
(3.27). Then, there exists an OPNCL controller K(g, x,t), which is nonlinear with
respect to x(t), such that the basin of entrainment (3.20) is not empty. Furthermore,
by this control, the entrainment is global. 
In the rest of this section, we will study the ability of entrainment with respect to
controllers of open-loop, linear open-loop, OPCL, and OPNCL.
Consider the controlled system (3.27). For convenience, we suppose that x(t) =
g(t) + u(t), where g(t) is the target orbit, and set S(t) = 1.
(i) K(g, x,t) is the Hübler open-loop action. In this case

K(g, x,t) = H(ġ, g,t) = ġ − F(g,t).

Substituting it into (3.27), we have

88 3 Entrainment and Migration Control of Chaotic Systems
m
1 ∂ k F(g,t) k
u̇ = F(u + g,t) − F(g,t) = ∑ k! ∂ gk
u. (3.31)
k=1

This equation is a complicated nonautonomous differential equation. In general,

it is difficult to judge whether the zero solution of the above equation is stable.
The basin of entrainment usually exists in the neighborhood of u(t) = 0 even if
the basin of entrainment is not an empty set.
(ii) K(g, x,t) is a linear feedback action. In this case

K(g, x,t) = −A(g(t) − x(t)),

where A is a constant matrix with all its eigenvalues having negative real parts.
Substituting the above equation into (3.27), we have

u̇ = Au + F(u + g,t) − ġ. (3.32)

Expanding F(u + g,t) with respect to u + g at u = 0, we obtain

m
1 ∂ k F(g,t) k
F(u + g,t) = F(g,t) + ∑ u.
k=1 k! ∂ gk

If g(t) is a solution of (3.24), then (3.32) can be changed to

m
∂ F(g,t) 1 ∂ k F(g,t) k
u̇ = A + u+ ∑ u. (3.33)
∂g k=2 k! ∂ gk

The discussion about u = 0 is analogous to (i). If g(t) is not a solution of (3.24),

we must have ġ − F(g,t) → 0 so that u → 0 when t → ∞. That is to say, g(t)
is restricted to asymptotically approximate a solution of (3.24). So, when linear
feedback is used, the target orbit is usually selected as an unstable periodic orbit
embedded in a chaotic attractor.
(iii) K(g, x,t) is an OPCL action. In this case

K(g, x,t) = ġ − F(g,t) + C(g,t)(g(t) − x(t)), (3.34)

where C(g,t) has the form

∂ F(g,t)
C(g,t) = −B
∂g

and B = (bi j )n×n is a constant matrix with all its eigenvalues having negative
real parts. Substituting (3.34) into (3.27), we have
m
1 ∂ k F(g,t) k
u̇ = Bu + ∑ u. (3.35)
k=2 k! ∂ gk
3.4 Global Control of a Class of Continuous-Time Polynomial Chaotic Systems 89

The judgement of stability of u = 0 for system (3.35) is easier than that of (3.32)
because the linear part of (3.35) is a constant matrix. When m = 2, system (3.25)
has the form
n n n
ẋi = ai (t) + ∑ bi j (t)x j + ∑ ∑ c jk (t)x j xk
(i)
(i = 1, 2, . . . , n). (3.36)
j=1 j=1 k=1

Adding OPCL action to (3.36) and transforming it into the equation about u, we
get
n n n
∑ bi j (t)u j + ∑ ∑ c jk (t)u j uk
(i)
u̇i = (i = 1, 2, . . . , n). (3.37)
j=1 j=1 k=1

It is clear that (3.37) is independent of the target orbit g(t).1 The stability of
u = 0 of (3.37) can be discussed by the Lyapunov direct method. However,
when m ≥ 3, the determination of the basin of entrainment is more complicated
than the case of m = 2, because the right-hand side of (3.35) may include g(t)
as its parameters.
From the above discussion and Theorem 3.2 we can find the superiority of the OP-
NCL method. First, the OPNCL method is independent of the target orbit g(t); sec-
ond, the basin of entrainment of the OPNCL method is global; that is to say, the
starting point may be arbitrarily selected in phase space.

3.4.2 Simulations

For validating the effectiveness of the OPNCL method, we take Chua’s circuit as an
example [1]. The diagram of Chua’s circuit is very simple and is shown in Fig. 3.3.
Chua’s circuit is composed of an inductor L, two capacitors C1 and C2 , a linear
resistor G, and a nonlinear resistor g. The system equations of Chua’s circuit are as
follows: ⎧
⎪ dVC1
⎪
⎪ C1 dt = G(VC2 − VC1 ) − g(VC1 ),
⎪
⎪
⎪
⎨
dVC2
C2 = G(VC1 − VC2 ) + iL , (3.38)
⎪
⎪ dt
⎪
⎪
⎪
⎪
⎩ L diL = −V .
C2
dt
Here, VC1 and VC2 are voltages of C1 and C2 , respectively, iL is the current through
the inductor L, and the nonlinear resistor g has the following feature:
1
g(VC1 ) = m0VC1 + (m1 − m0 )(|VC1 + 1| − |VC1 − 1|),
2

1 Both the Rössler system and the Lorenz system belong to this case whose linear parts are con-
stants.
90 3 Entrainment and Migration Control of Chaotic Systems

G iR

L
V2 VR
V1
C2 C1

iL
g

Fig. 3.3 The diagram of Chua’s circuit

where m0 , m1 < 0 are parameters. For the convenience of discussion, we implement

a transformation
iL tG
x1 = VC1 , x2 = VC2 , x3 = , τ = ,
G C2
and replace the piecewise-linear nonlinear function by a third-order polynomial

g(VC1 ) = a0VC1 [2(1 − G)VC21 − 1],

where a0 is a suitable selected parameter. If we set

C2 1 C2 C2
p= , r= −1 a0 , q = − 2 ,
C1 G C1 LG

and still use the symbol t to denote the transformed time τ , then system (3.38)
becomes a dimensionless one:
⎧
⎨ ẋ1 = px2 + r(x1 − 2x31 ),
ẋ = x1 − x2 + x3 , (3.39)
⎩ 2
ẋ3 = qx2 .

When p = 10, r = 10/7, and q = −100/7, system (3.39) has a chaotic attractor,
which is shown in Fig. 3.4. Suppose that we want to steer the solution of system
(3.39) to a target g(t) = (g1 (t), g2 (t), g3 (t))T . Adding an OPCL control action to
(3.39), we get a group of equations about u as follows:
3.4 Global Control of a Class of Continuous-Time Polynomial Chaotic Systems 91

0.2

0.15

0.1

0.05

0
x3

−0.05

−0.1

−0.15

−0.2
1
0.5 1.5
1
0 0.5
0
−0.5 −0.5
−1
x2 −1 −1.5
x1

Fig. 3.4 The attractor of Chua’s circuit

3 3 3 3 3 3
1 ∂ 2 Fi (g,t) 1 ∂ 3 Fi (g,t)
u̇i = ∑ bi j u j + 2 ∑ ∑ ∂ g j ∂ g k
u j uk +
6 ∑ ∑ ∑ ∂ g j ∂ gk ∂ gl u j uk ul
j=1 j=1 k=1 j=1 k=1 l=1

(i = 1, 2, 3). (3.40)

From (3.39) and (3.40), we get

⎧
⎨ u̇1 = b11 u1 − 6rg1(t)u21 − 2ru31,
u̇ = b22 u2 , (3.41)
⎩ 2
u̇3 = b33 u3 ,

where (bi j )3×3 is a diagonal matrix with all-negative elements. According to the
notation used in [4], the first equation of (3.40) can be written as

u̇1 = −(a + b(t)u1 + cu21)u1 ,

where a = −b11, b(t) = 6rg1 (t), and c = 2r. When b2 (t) < 4ac, the BE(g) of Chua’s
circuit (3.39) is global. When b2 (t) > 4ac, the cases of ġ1 (t) = 0 and ġ1 = 0 should
be considered, respectively, in order to determine the BE(g). However, for some
target g(t), it becomes more difficult to determine the basin of entrainment because
of the uncertainty on the sign of b2 (t) − 4ac [4]. For example, we choose the target
as follows:
g(t) = (g1 , g2 , g3 )T = (1 + 0.1 sint,t cost, −10)T .
92 3 Entrainment and Migration Control of Chaotic Systems

According to the OPCL design procedure, we have a = −b11 , b(t) = 6r(1 +

0.1 sint), and c = 2r. Thus, the sign of b2 (t) − 4ac is undetermined, which results
in the difficulty of getting the basin of entrainment.
If the OPNCL action is applied to system (3.39), the control terms are as follows:

K(g, x,t) = ġ − F(g,t) + D(g, x,t)(g − x)

⎛ ⎞ ⎛ ⎞
ġ1 − F1(g,t) r(1 − 6g21)(x1 − g1 ) + p(x2 − g2)
= ⎝ ġ2 − F2(g,t) ⎠ − ⎝ (x1 − g1 ) − (x2 − g2 ) + (x3 − g3) ⎠
ġ3 − F3(g,t) −q(x2 − g2 )
⎛ ⎞
−6rg1 u21 − 2ru31 + b11 u1
−⎝ b22 u2 ⎠.
b33 u3

Substituting the above equation into (3.39) and setting x(t) = g(t) + u(t), we get

u̇i = bii ui (i = 1, 2, 3). (3.42)

Choose bii = −1 (i = 1, 2, 3). It is easy to see that u will tend to zero quickly
without determining the basin of entrainment. The effects of methods of open-loop,
linear closed-loop, OPCL, and OPNCL are shown in Figs. 3.5–3.8, in which the
controllers, ui , are those of (3.31), (3.32), (3.35), and (3.42). From these figures
we can see that it is impossible for the methods of open-loop and linear feedback

0
u1

−2

−4
0 2 4 6 8 10 12 14 16 18 20

5
u2

−5
0 2 4 6 8 10 12 14 16 18 20

10

5
u3

−5

−10
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Fig. 3.5 Error curves under the Hübler open-loop action

3.4 Global Control of a Class of Continuous-Time Polynomial Chaotic Systems 93

0
u1

−2

−4
0 2 4 6 8 10 12 14 16 18 20

20

10
u2

−10

−20
0 2 4 6 8 10 12 14 16 18 20

40

20
u3

−20
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Fig. 3.6 Error curves under the linear feedback action

Fig. 3.7 Error curves under the OPCL action

94 3 Entrainment and Migration Control of Chaotic Systems

0.5

u1
0

−0.5
0 2 4 6 8 10 12 14 16 18 20

0.5
u2

−0.5

−1
0 2 4 6 8 10 12 14 16 18 20

10

5
u3

−5

−10
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Fig. 3.8 Error curves under the OPNCL action

0.3

0.2

0.1
x3

−0.1

−0.2
1
0.5 1.5
1
0 0.5
0
−0.5 −0.5
−1
x2 −1 −1.5
x1

Fig. 3.9 The solution of Chua’s circuit system approaches the target orbit quickly under the OP-
NCL action
3.5 Global Control of a Class of Discrete-Time Systems 95

to achieve the goal. The OPCL method can approximately achieve the aim but the
error curves are piecewise convergent. However, the OPNCL method can achieve
the goal and the basin of entrainment is global. We choose another target g(t) =
(0.5 sint, 0.3 + e−t , 0.8 cost)T . Under the action of the OPNCL method, the solution
of Chua’s circuit system is entrained to the target orbit quickly, which is shown in
Fig. 3.9.

3.5 Global Control of a Class of Discrete-Time Systems

3.5.1 OPNCL Control for Discrete-Time Systems

In this section, we will study the OPNCL control for a class of discrete-time systems.
For the given mapping

xk+1 = F(xk ) (x ∈ R p ), (3.43)

by adding a Hübler open-loop control action on the right-hand side of (3.43), we get

xk+1 = F(xk ) + [gk+1 − F(gk )]Sk (gk ∈ R p ), (3.44)

where the {gk } are some target to be reached and Sk is the switching function (Sk =
0, k < 0; 0 ≤ Sk ≤ 1, k ≥ 0). This kind of control (when Sk = 1) was first proposed
by Hübler and was used in the study of the logistic map and nonlinear damped
oscillations. Jackson and Hübler studied the global control scale of the logistic map
and proposed the concept of the basin of entrainment, denoted by BE({gk }); that is,

x0 ∈ BE({gk }),

if the solution of (3.44) satisfies the following condition:

lim xk − gk = 0.
t→∞

The determination of BE({gk }) is highly dependent on the given target orbit. In

some cases, BE({gk }) even has a fractal structure, which makes it difficult to use in
practical applications.
If the target orbit is a point g0 in the phase space, then this point must be in a
convergent region C, which is defined as

C = {x : ∂ Fi /∂ x j − δi j μ (x) = 0, ∀|μ (x)| < 1},

where
1, i = j;
δi j =
 j.
0, i =
96 3 Entrainment and Migration Control of Chaotic Systems

If the target orbit is not a point, then it is not easy to determine BE({gk }).
By adding a linear closed-loop control action to the right-hand side of system
(3.43), we have
xk+1 = F(xk ) + Dk (gk − xk ),
where Dk is a suitable selected matrix. For chaotic systems, the target orbit is usually
required to be an unstable periodic solution embedded in the chaotic attractor, which
satisfies
gk+1 = F(gk ).
Now suppose that there is a discrete-time chaotic system
 
(i) (1) (2) (p)
xk+1 = Fi xk , xk , . . . , xk (i = 1, 2, . . . , p), (3.45)

(1) (2)
where Fi (xk ) (x%k ∈ R p , i = 1, 2,
&. . . , p) is an mth-order polynomial about xk , xk , . . . ,
(p) (1) (p)
xk . If {gk } = g1 , . . . , gk is the target orbit to be achieved for (3.45), by adding
a combination control, K(gk , xk ), composed of a Hübler open-loop and a nonlinear
feedback, to the right-hand side of (3.45), we get
(i)
xk+1 = Fi (xk ) + Sk Ki (gk , xk ) (i = 1, 2, . . . , p), (3.46)

where
p  
Ki (gk , xk ) = gk+1 − Fi (gk ) − ∑ Ci j xk − gk ,
(i) ( j) ( j)

j=1

1 ∂ Fi (gk ) 1 p ∂ 2 Fi (gk ) (l)

+ ∑ ( j) (l) (xk − gk ) + · · ·
(l)
Ci j =
1! ∂ g( j) 2! l=1 ∂ g ∂ g
k k k
p
1 ∂ m Fi (gk )
+ ∑
(l) (l) (n) (n)
( j)
(x − gk ) · · · (xk − gk ) − ai j ,
(n) k
l,...,n=1 m! (l)
∂g ∂g ···∂g
k k k

and ai j is the element of a suitably selected matrix A. Thus, when Sk = 1 the set
BE({gk }) = {x0 : limn→∞ xn − gn = 0} is not empty, but is global. That is to say,
from any initial point the solution of system (3.45) can be entrained to the target
orbit {gk } by adding an OPNCL control action.
In fact, when Sk = 1, setting xk = uk + gk and noting that F(xk ) is an mth-order
polynomial, we know that the expansion of Fi (gk + uk ) is finite. Substituting xk =
uk + gk into (3.46), we get
3.5 Global Control of a Class of Discrete-Time Systems 97

(i) (i)
uk+1 + gk+1 = Fi (uk + gk ) + Ki (gk , xk )
p
1 ∂ Fi (gk )
∑
( j)
= Fi (gk ) + ( j)
uk + · · ·
1! j=1 ∂ gk
p
1 ∂ m Fi (gk ) ( j)
∑
(l)
+ ( j) (l) k
u · · · uk + Ki (gk , xk ).
m! j,...,l=1 ∂ gk · · · ∂ gk

Substituting the expression of Ki (gk , xk ) into the above equation, we get a group of
equations with respect to uk :
p
∑ ai j uk
(i) ( j)
uk+1 = . (3.47)
j=1

There are many ways to select the matrix A. For simplicity, we can choose ai j =
(i)
0 (i = j) and |aii | < 1. By this choice, the uk (i = 1, . . . , p) of (3.47) will be conver-
gent to 0, which means that the solution of system (3.45) is entrained to the target
orbit {gk } after Sk = 1.

3.5.2 Simulations

To present the effectiveness of the OPNCL control action we take the logistic map,

xk+1 = λ xk (1 − xk ), (3.48)

as the first example. The bifurcation diagram2 (it displays how the periodic points
of the logistic map vary with λ ) and the Lyapunov exponent of (3.48) are shown in
Fig. 3.10 and Fig. 3.11. When λ = 3.7 the logistic map is in chaotic state. Set the
target orbit as gk = 0.65 + 0.15(−1)k and a11 = −0.95. When k = 2000, Sk = 1. The
control effect is shown in Fig. 3.12.
The second example is to consider the Hénon map:

xk+1 = yk + 1 − ax2k ,
(3.49)
yk+1 = bxk .

2 Consider an nth-order continuous-time system

ẋ = f (x, α )

with a parameter α ∈ R. As α changes, the limit sets of the system also change. Typically, a small
change in α produces small quantitative changes in a limit set. For instance, perturbing α could
change the position of a limit set slightly and, if the limit set is not an equilibrium point, its shape
or size could also change. There is also the possibility that a small change in α can cause a limit
set to undergo a qualitative change. Such a qualitative change is called a bifurcation and the value
of α at which a bifurcation occurs is called a bifurcation value.
98 3 Entrainment and Migration Control of Chaotic Systems

Fig. 3.10 The bifurcation diagram of logistic map

Fig. 3.11 The Lyapunov exponent of logistic map

3.5 Global Control of a Class of Discrete-Time Systems 99

Fig. 3.12 The logistic map under the action of OPNCL control

Fig. 3.13 The bifurcation diagram of Hénon map

100 3 Entrainment and Migration Control of Chaotic Systems

0.5

0
Lyapunov Exponents

−0.5

−1

−1.5

−2
0 0.2 0.4 0.6 0.8 1 1.2 1.4
a

Fig. 3.14 The Lyapunov exponents of Hénon map

1
x(k)

−1

−2
0 500 1000 1500 2000 2500 3000 3500 4000
k
1.5

1
y(k)

0.5

−0.5
0 500 1000 1500 2000 2500 3000 3500 4000
k

Fig. 3.15 The Hénon map under the action of OPNCL control
3.6 Summary 101

1.4

1.2

0.8

0.6
y(k)

0.4

0.2

−0.2

−0.4
−1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
x(k)

Fig. 3.16 The phase diagram of Hénon map under the action of OPNCL control

When a = 1.75 and b = 0.3 the Hénon map is in chaotic state. The bifurcation
diagram and Lyapunov exponents are shown in Fig. 3.13 and Fig. 3.14. The target
orbit is chosen as
(1) (2)
gk = 2 + 0.2 sin(0.02kπ ), gk = 1 + 0.25 cos(0.04kπ ).

When k = 2000, Sk = 1, the OPNCL controller begins to work. The control effect is
shown in Fig. 3.15 and Fig. 3.16, respectively.

Remark 3.2. In theory, the OPNCL method can entrain the solution of the controlled
system to any target orbit, even if the target is unbounded. But, in practice, when the
target orbit is unbounded it is impossible to design the controller since the output
energy of the controller is infinite. 

3.6 Summary

In this chapter, we studied methods for chaos control, entrainment, and migration.
At the beginning, we introduced an open-loop entrainment method, and then a mod-
ified entrainment method, the OPCL control method, was introduced. Based on the
OPCL method, we proposed the OPNCL control method. It can entrain the solu-
tion of the controlled system to any bounded target orbit. We applied the OPNCL
102 3 Entrainment and Migration Control of Chaotic Systems

method to both continuous-time systems and discrete-time systems. Simulation re-

sults validated the theoretical analysis.

References

1. Chua LO, Komura M, Matsumoto T (1986) The double scroll family: parts I and II. IEEE
Trans Circuits Syst I 33:1073–1118
2. Hu G, Xiao J, Zheng Z (2000) Chaos Control. Shanghai Scientific and Technological Educa-
tion Publishing House, Shanghai
3. Hübler A (1989) Adaptive control of chaotic systems. Helv Phys Acta 62:343–346
4. Jackson EA, Grosu I (1995) An open-plus-closed-loop (OPCL) control of complex dynamic
systems. Physica D 85:1–9
5. Wang J, Tian P, Chen C (2000) Global control of continuous polynomial chaotic systems.
Control Decis 15:309–313 (in Chinese)
Chapter 4
Feedback Control of Chaotic Systems

Abstract The parameters of a chaotic system play an important role, whose vari-
ation will lead to completely different dynamics. Sometimes, we want to design a
controller which is optimal in a certain sense. In this chapter, we focus on two kinds
of methods of suppressing chaos: the adaptive control method and the inverse op-
timal control method. We develop two new methods of parametric adaptive control
for a class of discrete-time chaotic systems and a class of continuous-time chaotic
systems with multiple parameters, respectively. The systems are first assumed to
be linear with respect to parameters. Then, systems with nonlinear distributed pa-
rameters and uncertain noise are considered. Finally, we apply the inverse optimal
control method to stabilize a four-dimensional chaotic system.

4.1 Introduction

In recent years, many conventional and novel control methods have been applied to
the suppression of chaotic behaviors, such as adaptive control [14], fuzzy control
[8], linear feedback control [13], impulsive control [3], time-delay feedback control
[9], optimal control [6], etc.
The parameters of a chaotic system play an important role, whose variation will
lead to completely different dynamics. The dynamical mechanism of the system
may change qualitatively during a small change of parameters. Such abrupt tran-
sitions between dynamical regimes are usually modeled in terms of bifurcations,
catastrophes, and crises [2, 4]. In other cases the transitions are not unwanted, but
are useful for switching between related phenomena (bistability and hysteresis). Re-
cently, there has been much work done in the area of parametric adaptive control of
chaotic systems, most of which usually assumed the parameters to be constants or
to be varying very slowly compared to the dynamical time scales of the systems. In
real applications the parameters are not always linearly distributed. Sometimes, we
want to design a controller which is optimal in a certain sense. However, it is a dif-
ficult task to solve the Hamilton–Jacobi–Bellman (HJB) equation when one designs

103
104 4 Feedback Control of Chaotic Systems

an optimal controller directly along the conventional route. To solve the aforemen-
tioned problems, in this chapter, we focus on two kinds of methods of suppressing
chaos: the adaptive control method and the inverse optimal control method. We will
develop two new methods of parametric adaptive control for a class of discrete-
time chaotic systems and a class of continuous-time chaotic systems with multiple
parameters, respectively. The systems are first assumed to be linear with respect
to parameters. Then, systems with nonlinear distributed parameters and uncertain
noise are considered.
Finally, we will apply the inverse optimal control method to stabilize a four-
dimensional chaotic system. In the 1990s, Freeman and Kokotovic [1] proposed
the inverse optimal control method, in which an optimal controller was constructed
through a Lyapunov function. As long as we can find such a Lyapunov function, it
can be inferred that the designed controller is optimal with a specified performance
index. The merit of this approach is that we do not need to solve the complicated
HJB equation.

4.2 Model-Reference Adaptive Control of a Class of

Discrete-Time Chaotic Systems

4.2.1 Basic Formulation

Consider a class of discrete-time nonlinear systems as follows:

x(k + 1) = F(x(k), μ (k)), k = 1, 2, . . . , (4.1)

where x(k) ∈ Rn is the state vector, μ (k) = (μ1 (k), . . . , μm (k))T is the parameter
vector which determines the nature of dynamics and is time dependent under the
influence of both disturbances and control, and F : Rn × Rm → Rn is a nonlinear
map. First, suppose that F is linearly parameterized. Many typical discrete-time
chaotic systems belong to this case, such as the Hénon map and the logistic map.
Then, (4.1) can be rewritten as
m
x(k + 1) = A(x(k)) + ∑ Bl (x(k))μl (k)
l=1

or
m
x p (k + 1) = A p (x(k)) + ∑ B pq (x(k))μq (k), p = 1, 2, . . . , n,
q=1

where A, Bl : Rn → Rn(l = 1, 2, . . . , m), A p , B pq : Rn → R (p = 1, 2, . . . , n; q =

1, 2, . . . , m).
Consider the following reference model:

y(k + 1) = FM (y(k), μg ), (4.2)

4.2 Model-Reference Adaptive Control of a Class of Discrete-Time Chaotic Systems 105

where FM : Rn × Rm → Rn , and μg is the desired value (target value).

Our goal is to design an appropriate adaptive controller such that the parameter
μ (k) converges to μg as soon as possible.

4.2.2 Main Results

In light of the basic idea of adaptive control, we propose a new parametric adaptive
algorithm as follows:

μ (k + 1) = μ (k) + α (x(k))G(e(k + 1)), (4.3)

where α = diag {αl } (l = 1, 2, . . . , m), αl : Rn → R, G : Rn → Rm , and e(k + 1) =

x(k + 1) − y(k + 1). G is an arbitrary nonlinear vector function with the error signal
e(k + 1) as its argument. Noting that there is no coupling between the adjustable
system (4.1) and the reference model (4.2), when x(k) or y(k) presents high periodic
or even quasiperiodic chaotic behavior, the change of the value of e(k) = x(k) − y(k)
will become extremely complicated. In order to effectively control the complicated
system, in light of the idea of coupling systems, we redirect the state vector x(k) of
(4.1) to the reference model (4.2), i.e.,

y(k + 1) = FM (x(k), μg ). (4.4)

Therefore, we can let the dynamics of the adjustable system redirect to the reference
model to predict the dynamics of y(k + 1). Let FM = F, and denote Gl (z) = ∑nk=1 zk
(l = 1, 2, . . . , m), where z = (z1 , z2 , . . . , zn )T ∈ Rn . Then, combining (4.1), (4.3), and
(4.4) yields
⎧ m
⎪
⎪ x p (k + 1) = A p (x(k)) + ∑ B pq (x(k))μq (k),
⎪
⎪
⎪
⎪ q=1
⎨ m
y p (k + 1) = A p (x(k)) + ∑ B pq (x(k))μqg , (4.5)
⎪
⎪ q=1  
⎪
⎪ n m
⎪
⎪
⎩ μl (k + 1) = μl (k) + αl (x(k)) ∑ ∑ B pq (x(k))(μq (k)−μqg ) ,
p=1 q=1

where A p , B pq (p = 1, 2, . . . , n; q = 1, 2, . . . , m) are nonlinear functions of x(k) and

l = 1, 2, . . . , m. For any q, there is at least one of B pq (p = 1, 2, . . . , n) which is not
equal to 0. μqg (q = 1, 2, . . . , m) are target values of μq (k). By choosing αl (x(k))
appropriately, it can be proven that the above parametric adaptive control equations
make the parameters converge asymptotically to the desired values for any disturbed
parameter initial values. The relevant theorem is presented as follows.

Theorem 4.1 ([11]). For the parametric adaptive control equations (4.5), if we
choose
αl (x(k)) = −K(x(k))Wl (x(k)), (4.6)
106 4 Feedback Control of Chaotic Systems
n
where Wl (x(k)) = ∑ B pl (x(k)),
p=1

' (−1
m
K(x(k)) = 2 d + ∑ Wl2 (x(k)) , (4.7)
l=1

and d > 0 is an arbitrary constant, then the parameters of (4.5) will converge asymp-
totically to the desired values.

Proof. First, we choose a Lyapunov function as

m  2
V (k) = ∑ μl (k) − μlg , (4.8)
l=1

and then we have

Δ V (k) = V (k + 1) − V(k)
m m
= ∑ (μl (k + 1) − μlg)2 − ∑ (μl (k) − μlg)2 . (4.9)
l=1 l=1

Substituting (4.5) into (4.9), and denoting μ j (k)− μ jg = μk, jg (1 ≤ j ≤ m), we derive
' ' ((2
m n m m
Δ V (k) = ∑ μk,lg + αl (x(k)) ∑ ∑ B pq(x(k))μk,qg − ∑ μk,lg
2
. (4.10)
l=1 p=1 q=1 l=1

Substituting (4.6) into (4.10), we get

' (' (2
m m
Δ V (k) = K(x(k)) K(x(k)) ∑ Wl2 (x(k)) − 2 ∑ Wq (x(k))μk,qg . (4.11)
l=1 q=1

Substituting (4.7) into (4.11) yields

' ' ((2
m
Δ V (k) = −d K(x(k)) ∑ Wq(x(k))(μq (k) − μqg) ≤ 0.
q=1

Since Δ V (k) is negative semidefinite, V (k) is a nonincreasing function with re-

spect to time t and bounded from below. Therefore, V (k) must converge to a finite
value V (∞). Noting that V (1) is bounded, we have
s s
lim
s→∞
∑ Δ V (k) = s→∞
lim ∑ (V (k + 1) − V(k))
k=1 k=1
= V (∞) − V (1)
= finite value.
4.2 Model-Reference Adaptive Control of a Class of Discrete-Time Chaotic Systems 107

Noting that Δ V (k) ≤ 0, we can derive

lim (μq (k) − μqg) = 0 (q = 1, 2, . . . , m).

k→∞

It is clear that system (4.5) is asymptotically stable with respect to μq (k). In other
words, the parameters μq (k) will asymptotically converge to the desired values μqg ,
q = 1, 2, . . . , m, respectively. 

Next, we will discuss the case that the system parameters are nonlinearly dis-
tributed. Moreover, the disturbances of the uncertain noise are taken into account.
Consider a class of discrete-time nonlinear systems as follows:

x(k + 1) = F(x(k), μ (k)) + Δ W (k),

where Δ W (k) is the uncertain noise. Consider the same reference model as in (4.4),
i.e.,
y(k + 1) = FM (x(k), μg ).
Furthermore, assume that F = FM , and the adaptive control algorithm is designed
the same as (4.3), i.e.,

μ (k + 1) = μ (k) + α (x(k))G(e(k + 1)),

where
e(k + 1) = x(k + 1) − y(k + 1)
= F(x(k), μ (k)) − F(x(k), μg ) + Δ W (k).
In the above, α (x(k)) is defined as
' (−1
m
αl (x(k)) = −2Rl (x(k)) d + ∑ R2l (x(k)) , (4.12)
l=1

n )
where Rl (x(k)) = ∑ ∂ Fi (x(k), μ (k)) ∂ μlg and d > 0 is an arbitrary positive con-
i=1
stant.
Using Taylor expansion of F(x(k), μ (k)) − F(x(k), μg ) around the neighborhood
of μ (k) = μg , we can derive
' (
n m
∂ Fi
μl (k + 1) = μl (k) + αl (x(k)) ∑ ∑ (μ j (k) − μ jg )
i=1 j=1 ∂ μ jg

+ h2l (x(k), μ (k) − μg ), (4.13)

where
n
h2l (x(k), μ (k) − μg ) = ∑ (αl (x(k))h1i (x(k), μ (k) − μg ) + Δ Wi )
i=1
108 4 Feedback Control of Chaotic Systems

and h1i (x(k), μ (k) − μg ) = o(μ (k) − μg ). Now, consider the stability of solutions of
the linear part of (4.13), i.e.,
' (
n m
∂ Fi
μl (k + 1) = μl (k) + αl (x(k)) ∑ ∑ (μ j (k)−μ jg ) . (4.14)
i=1 j=1 ∂ μ jg

We choose a Lyapunov function as

m
V (k) = ∑ (μl (k) − μlg )2 ,
l=1

and then we have

Δ V = V (k + 1) − V(k)
m m
= ∑ (μl (k + 1) − μlg)2 − ∑ (μl (k) − μlg)2 . (4.15)
l=1 l=1

Substituting (4.14) into (4.15), we get

⎧ ⎫
⎪' (⎛ ' (2 ⎞−1 ⎪2
⎨ n m
∂ Fi m n
∂ Fi ⎬
Δ V = −4d ∑ ∑ (μ j (k) − μ jg ) ⎝d + ∑ ∑ ⎠
⎩ i=1 j=1 ∂ μ jg
⎪ l=1 i=1 ∂ μlg
⎪
⎭

≤ 0. (4.16)

From (4.16) we can infer that (4.14) is asymptotically stable. As for the nonlinear
equation given in (4.13), as long as the term h2l is small enough, it is stable, too.

4.2.3 Simulations

An illustrative example is used here to demonstrate the effectiveness of the proposed

parametric adaptive control method. Consider the Hénon map:

x1 (k + 1) = −px21 (k) + x2 (k) + 1,
k = 0, 1, . . . , (4.17)
x2 (k + 1) = qx1 (k),

where p and q are real parameters. When p = 1.4 and q = 0.3, system (4.17) presents
chaotic behaviors. When p = 0.85 and q = 0.55, system (4.17) is in the normal state.
Suppose that the reference model is given as follows:

y1 (k + 1) = −0.85x21(k) + x2 (k) + 1,
y2 (k + 1) = 0.55x1(k).
4.2 Model-Reference Adaptive Control of a Class of Discrete-Time Chaotic Systems 109

1
x1 (k)
0

−1

−2
0 500 1000 1500
k
1.5

1
x2 (k)

0.5

−0.5

−1
0 500 1000 1500
k

Fig. 4.1 The process of controlling the Hénon map (i)

1.6

1.4
p(k)

1.2

0.8
0 500 1000 1500
k
0.8

0.6
q(k)

0.4

0.2

0
0 500 1000 1500
k

Fig. 4.2 The parameter change process of controlling the Hénon map (i)
110 4 Feedback Control of Chaotic Systems

1
x1 (k)
0

−1

−2
0 500 1000 1500
k
1

0.5
x2 (k)

−0.5

−1
0 500 1000 1500
k

Fig. 4.3 The process of controlling the Hénon map (ii)

1.8

1.6

1.4
p(k)

1.2

0.8
0 500 1000 1500
k
0.8

0.6
q(k)

0.4

0.2

0
0 500 1000 1500
k

Fig. 4.4 The parameter change process of controlling the Hénon map (ii)
4.3 Model-Reference Adaptive Control of a Class of Continuous-Time Chaotic Systems 111

Assume that the parameters p and q of system (4.17) deviate from the desired
values (pg = 0.85, qg = 0.55) by a sudden disturbance. The parameters p and q
change to p = 1.4 and q = 0.3 for k = 501, 502, . . ., 850. The disturbed system (4.17)
presents chaotic behavior with the changed parameters. The control goal is to make
the changed parameters return to the desired values such that the disturbed system
(4.17) returns to the normal state. Apply the parametric adaptive control law (4.5),
i.e.,

p(k + 1) = p(k) + 2x21 (k)Q [x1 (k + 1) − y1(k + 1) + x2(k + 1) − y2(k + 1)],

q(k + 1) = q(k) − 2x1(k)Q [x1 (k + 1) − y1(k + 1) + x2(k + 1) − y2(k + 1)],

where Q = (d + x41 (k) + x21 (k))−1 , d = 2, and k = 851, . . ., 1500. The simulation
results are given in Figs. 4.1 and 4.2. From the two figures, we can see that the
disturbed parameters converge to the desired values rapidly. This control method
can be applied to many practical discrete-time systems.
On the other hand, assume that the normal state of system (4.17) is chaotic. The
system presents nonchaotic behavior with external disturbances. Now the control
goal is to make the disturbed system return to the chaotic state. Still consider the
Hénon map, but the disturbed parameters are given as p = 0.85, q = 0.55, and k =
501, . . ., 850. Suppose that the reference model is given as follows:

y1 (k + 1) = −1.4x21(k) + x2 (k) + 1,
y2 (k + 1) = 0.3x1 (k).

Apply the parametric adaptive control law (4.5), where d = 2 and k = 851, . . . , 1500.
The simulation results are given in Figs. 4.3 and 4.4. From the two figures, we can
see that the disturbed parameters converge to the desired values rapidly.

4.3 Model-Reference Adaptive Control of a Class of

Continuous-Time Chaotic Systems

4.3.1 Basic Formulation

Consider a class of continuous-time nonlinear dynamical systems:

ẋ = F(x,t, μ ), (4.18)

where x ∈ Rn is the state vector, μ ∈ Rm is the parameter vector which determines

the nature of the dynamics, and F : Rn × R × Rm → Rn . First, suppose that F is
linear with respect to parameters, and then (4.18) can be described by the following
equation:
m
ẋ = φ (x,t) + ∑ ψk (x,t)μk , (4.19)
k=1
112 4 Feedback Control of Chaotic Systems

where φ , ψk : Rn × R → Rn (k = 1, 2, . . . , m), and μ = (μ1 , μ2 , . . . , μm )T is the pa-

rameter vector. The parameter vector determines the asymptotic behavior of a typ-
ical trajectory. In spite of any sources of noise or accidental fluctuations, the pur-
pose of the parametric adaptive control is to maintain the parameter vector μ at a
predetermined desired value. The disturbed parameters are updated by the control
algorithm. Under the influence of the fluctuations as well as the control, the param-
eter μ becomes time-dependent quantity μ (t). The action of the control draws them
toward a desired value μg . One of the demands on the control method is that the
restoration should be as brief as possible. A further requirement is the ability of the
control method to control the system for any initial condition x(0) in the state space
with respect to any desired parameter value μg , i.e., the parametric adaptive control
is global.
If the instantaneous values of the parameters are known, then their deviations
from the desired values could readily be calculated. It would then be trivial to con-
trol the system by updating the parameters with their desired values. In reality, re-
searchers do not generally know the changed value of the parameters at any instant.
Without any information about parameter changes, the problem of detecting devia-
tions from the desired values becomes much more difficult. The parameters can only
be identified by the properties of the global behavior of the system. To detect devi-
ations in the parameter setting, one can use the vector x, the deviation between the
vector x of and the reference vector y, to estimate the values of system parameters.

4.3.2 Main Results

First, consider a reference model ẏ = FM (y,t, μ ), which has the same dynamics as
the system (4.19) at the desired value μ = μg , i.e.,

ẏ = FM (y,t, μg ). (4.20)

The present adaptive control law is given as follows:

μ̇ = β (x,t)G(ė), (4.21)

where β (x,t) = diag {βl (x,t)} (l = 1, 2, . . . , m) is a diagonal matrix, e = x − y is the

error vector, and G : Rn → Rm is a continuous vector function with the error signal
as its argument. The difficulty lies in how to choose an appropriate G.
The systems (4.19)–(4.21) constitute a set of control equations. If the parame-
ter μ of system (4.19) is equal to μg , and system behavior tends to an equilibrium
or low-periodic behavior, the reference model (4.20) can be considered as a simple
pattern of behavior. Thus, the control strategy containing (4.19), (4.20), and (4.21)
is a successful set of parametric adaptive control equations [7, 10]. However, when
μ = μg , if system (4.19) presents high periodic, quasiperiodic, or even chaotic be-
havior, then the error signal e or ė becomes an extremely complicated varying value.
In order to effectively control more complicated systems, we will redirect the out-
4.3 Model-Reference Adaptive Control of a Class of Continuous-Time Chaotic Systems 113

put x(t) of system (4.19) to system (4.20). We then derive the coupling equation
between the system and the model:

y = FM (x,t, μg ).

Therefore, the proposed method of control consists of a combination of equations

(4.19), (4.20), and (4.21), or
⎧
⎨ ẋk = Fk (x,t, μ ),
ẏ = FMk (x,t, μg ), (4.22)
⎩ k
μ̇l = βl (x,t)Gl (ė),

where ė = ẋ − ẏ, k = 1, 2, . . . , n, and l = 1, 2, . . . , m. We assume that the model dy-

namics is the same as that of the system’s, i.e., FMk = Fk (k = 1, 2, . . . , n). Then,
the control equations (4.22) are globally stable with respect to μl if we choose Gl
appropriately. The relevant theorem is given as follows.
Theorem 4.2 ([11]). The parameters of the parametric adaptive control equations
(4.22) can converge asymptotically to the desired values for any disturbed parameter
initial values if we choose
n
βl (x,t) = − ∑ ψlk (x,t)
k=1

and
n
Gl (ė) = ∑ ėk , (4.23)
k=1

where l = 1, 2, . . . , m.
Proof. Suppose that F(x,t, μ ) = φ (x,t) + ∑m l=1 ψl (x,t)μl , in which φ , ψl : R ×
n

R → Rn , φ (x,t) = (φ1 (x,t), φ2 (x,t), . . . , φn (x,t))T , ψl (x,t) = (ψl1 (x,t), ψl2 (x,t), . . . ,
ψln (x,t))T , and ψlk (x,t) (l = 1, 2, . . . , m; k = 1, 2, · · · , n) are uniformly continuous
real functions. Then, the adaptive control algorithm (4.22) can be transformed to the
following:
⎧ m
⎪
⎪ ẋk = φk (x,t) + ∑ ψlk (x,t)μl ,
⎪
⎪
⎪
⎨ l=1
m
ẏk = φk (x,t) + ∑ ψlk (x,t)μlg , (4.24)
⎪
⎪ l=1
⎪
⎪ n m n
⎪
⎩ μ̇l = − ∑ ψlk (x,t) ∑ ∑ ψ pq (x,t) (μ p − μ pg) .
k=1 p=1 q=1

Now, we choose a Lyapunov function as

1 m
V (t) = ∑ (μl − μlg)2 ,
2 l=1

and then we have

114 4 Feedback Control of Chaotic Systems
m
V̇ (t) = ∑ (μl − μlg )μ̇l
l=1
' (' (-
m n m n
= ∑ (μl − μlg ) − ∑ ψlk (x,t) ∑ ∑ ψ pq(x,t)(μ p − μ pg)
l=1 k=1 p=1 q=1
' (
m n n
= ∑ (μl − μlg ) − ∑ ψlk (x,t) ∑ ψlq (x,t)(μl − μlg)
l=1 k=1 q=1
' ( -
n m n
− ∑ ψlk (x,t) ∑ ∑ ψ pq(x,t)(μ p − μ pg)
k=1 p=l q=1
⎧' (2
m ⎨ n
=− ∑ ∑ ψlk (x,t) (μl − μlg )2
l=1
⎩ k=1
' ( -
n m n
+ ∑ ψlk (x,t) (μl − μlg) ∑ ∑ ψ pq(x,t)(μ p − μ pg)
k=1 p=l q=1
-2
m n
=− ∑ ∑ ψlk (x,t)(μl − μlg) ≤ 0.
l=1 k=1

It is shown that V̇ (t) is a negative semidefinite function. Therefore, V (t) is a non-

increasing function of time which is also bounded from below and hence converges
to a finite value V (∞). Since the initial value μ (0) is bounded, V (0) is bounded. In
such a case we have  t
lim V̇ (t)dt = V (∞) − V (0),
t→∞ 0

which is a finite value. Since ψlk (x,t) (l = 1, 2, . . . , m; k = 1, 2, . . . , n) are uniformly

continuous, V̇ (t) is also uniformly continuous. Then, we find that

lim V̇ (t) = 0
t→∞

or
m n
lim
x→∞
∑ ∑ ψlk (x,t)(μl − μlg) = 0.
l=1 k=1

Note that ∑nk=1 ψlk (x,t) (l = 1, 2, . . . , m) do not approach constants as t → ∞.

Then, we obtain
lim (μl − μlg ) = 0 (l = 1, 2, . . . , m).
t→∞

Therefore, we have proven that the parametric adaptive control equations (4.24)
are efficient. In other words, the parameter μl is asymptotically stable, which
achieves the stability match between the adjustable system and the reference model.

Next, we will discuss the case in which the system is nonlinearly parameterized.
Here, we cannot adopt the same adaptive control method as (4.24). Consider the
4.3 Model-Reference Adaptive Control of a Class of Continuous-Time Chaotic Systems 115

following equations: ⎧
⎨ ẋk = Fk (x,t, μ ) + Δ Wk (t),
ẏ = FMk (x,t, μg ), (4.25)
⎩ k
μ̇l = βl (x,t)Gl (ė),
where l = 1, 2, . . . , m and Δ Wk (t) (k = 1, 2, . . . , n) are disturbed terms of external un-
certain noise. Suppose that there is no modeling error, i.e., Fk = FMk (k = 1, 2, . . . , n),
and Gl is defined by (4.23). Then, the third equation of (4.25) can be rewritten as
' (
n
μ̇l = βl (x,t) ∑ (Fk (x,t, μ ) − Fk (x,t, μg )) + Δ Wk (t) . (4.26)
k=1

Expanding (Fk (x,t, μ ) − Fk (x,t, μg )) around a neighborhood of μ = μg , one ob-

tains
m
∂ Fk (x,t, μ )
Fk (x,t, μ ) − Fk (x,t, μg ) = ∑ (μl − μlg ) + h1k (x,t, μ − μg ), (4.27)
l=1 ∂ μlg

where ∂ Fk (x,t, μ )/∂ μlg = (∂ Fk (x,t, μ )/∂ μ )|μ =μg and h1k (x,t, μ − μg ) is the higher-
order infinitesimal term of μ − μg . Substituting (4.27) into (4.26), we get
-
n m
∂ Fk (x,t, μ )
μ̇l = βl (x,t) ∑ ∑ (μl − μlg ) + h1k (x,t, μ − μg ) + Δ Wk (t)
k=1 l=1 ∂ μlg
m n
∂ Fk (x,t, μ )
= βl (x,t) ∑ ∑ (μl − μlg ) + h2(x,t, μ − μg), (4.28)
l=1 k=1 ∂ μlg

where h2 (x,t, μ − μg ) = ∑nk=1 βl (x,t)(h1k (x,t, μ − μg ) + Δ Wk (t)). Let

n
∂ Fk (x,t, μ )
βl (x,t) = − ∑ . (4.29)
k=1 ∂ μlg

Substituting (4.29) into (4.28), we get

' (' (
n m n
∂ Fk (x,t, μ ) ∂ Fq (x,t, μ )
μ̇l = − ∑ ∑ ∑ ∂ μ pg (μ p − μ pg)
k=1 ∂ μlg p=1 q=1 (4.30)
+ h2 (x,t, μ − μg).

Since the term h2 (x,t, μ − μg ) includes the information of the unknown distur-
bance Δ Wk (t), the parametric adaptive law (4.30) is not feasible. Therefore, let us
consider the following parametric adaptive law instead of (4.30):
' (' (
n m n
∂ Fk (x,t, μ ) ∂ Fq (x,t, μ )
μ̇l = − ∑ ∑ ∑ ∂ μ pg (μ p − μ pg) . (4.31)
k=1 ∂ μlg p=1 q=1
116 4 Feedback Control of Chaotic Systems

Now we analyze the stability of the solution of (4.31). We still choose a Lyapunov
function as
1 m
V (t) = ∑ (μl − μlg )2 ,
2 l=1
and then we get
m  
V̇ (t) = ∑ μl − μlg μ̇l
l=1
' (' (-
m   n
∂ Fk (x,t, μ ) m n
∂ Fq (x,t, μ )
= ∑ μl − μlg − ∑ ∑ ∑ ∂ μ pg (μ p − μ pg)
l=1 k=1 ∂ μlg p=1 q=1
' (
m n
∂ Fk (x,t, μ ) n ∂ Fq (x,t, μ )
= ∑ (μl − μlg ) − ∑ ∑ ∂ μlg (μl − μlg)
l=1 k=1 ∂ μlg q=1
' (' (-
n
∂ Fk (x,t, μ )   m n
∂ Fq (x,t, μ )
− ∑ μl − μlg ∑ ∑ ∂ μlg (μ p − μ pg)
k=1 ∂ μlg p=l q=1
⎧' (2
m ⎨ n
∂ Fk (x,t, μ )  2
=− ∑ ∑ μl − μlg
l=1
⎩ k=1 ∂ μlg
' ( -
n
∂ Fk (x,t, μ )   m n ∂ Fq (x,t, μ )
+ ∑ μl − μlg ∑ ∑ (μ p − μ pg)
k=1 ∂ μlg p=l q=1
∂ μlg
-2
m n
∂ Fk (x,t, μ )  
=− ∑ ∑ μl − μlg
l=1 k=1 ∂ μlg
≤ 0.

Similar to the discussion of V (t) earlier, one obtains that limt→∞ V̇ (t) = 0 and

lim (μl − μlg ) = 0 (l = 1, 2, . . . , m).

t→∞

We have proven that (4.31) is asymptotically stable. For the nonlinear equation
given in (4.30), if the term h2 (x,t, μ − μg ) is sufficiently small, then it is also stable.

4.3.3 Simulations

To visualize the effectiveness of the proposed parametric adaptive control method,

two illustrative examples are presented below. First, consider the Duffing equation:

ẋ1 = x2 ,
(4.32)
ẋ2 = −p2 x1 − x31 − p1 x2 + q cos(1.8t),
4.3 Model-Reference Adaptive Control of a Class of Continuous-Time Chaotic Systems 117

2.5

1.5

0.5
x2

−0.5

−1

−1.5

−2

−2.5
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x1

Fig. 4.5 The chaotic attractor of the Duffing equation described by (4.32)

where p1 , p2 , and q are system parameters. When p1 = 0.4, p2 = −1.1, q = 1.8, and
the initial state is given as x(0) = (1, 1)T , system (4.32) presents chaotic behavior,
which is illustrated in Fig. 4.5.
Suppose that the reference model is

ẋ1 = x2 ,
ẋ2 = 0.8x1 − x31 − 0.7x2 + 0.6 cos(1.8t),

i.e., the desired values of the system parameters are p1g = 0.7, p2g = −0.8, and
qg = 0.6. The reference model corresponds to a normal state. Suppose that the pa-
rameters are disturbed to p1 = 0.4, p2 = −1.1, and q = 1.8, which make system
(4.32) chaotic. Our goal is to make the disturbed system return to the normal state.
Applying the parametric adaptive control law (4.24), we get
⎧
⎨ ṗ1 = −(−x2 ) [−x2 (p1 − 0.7) − x1(p2 + 0.8) + (q − 0.6) cos(1.8t)] ,
ṗ2 = −(−x1 ) [−x2 (p1 − 0.7) − x1(p2 + 0.8) + (q − 0.6) cos(1.8t)] ,
⎩
q̇ = − cos(1.8t) [−x2 (p1 − 0.7) − x1(p2 + 0.8) + (q − 0.6) cos(1.8t)] .

The simulation results are shown in Figs. 4.6–4.10 which show that the disturbed
parameters converge quickly to their desired values after the transient. Figs. 4.6 and
4.7 show the moving trajectory of the undisturbed normal state, disturbed chaotic
state, and controlled desired state of system (4.32). From Figs. 4.8–4.10 we can
see that parametric adaptive control law (4.24) is very powerful in stabilizing the
system. The control method can be applied to electric power systems in a chaotic
118 4 Feedback Control of Chaotic Systems

1.5

0.5
x1

−0.5

−1

−1.5

−2
−60 −40 −20 0 20 40 60 80
t (sec)

Fig. 4.6 The changed state response curve of x1 (t), which is disturbed to be chaotic at t = −30,
and approaches the desired state when the controller is applied at t = 0

state after the external disturbance, which automatically adjusts to a normal state
and guarantees the security of the electric power network.
Now, we assume that an external disturbance Δ W (t) acts on the Duffing equation
when t = 0. The parametric adaptive law remains unchanged. Choosing Δ W (t) as
a random background noise and maxt {|Δ W (t)|} = 0.1, the adaptive control law is
remarkably robust (see Figs. 4.11–4.15).
Next, consider the Rössler system:
⎧
⎨ ẋ1 = −x2 − x3 ,
ẋ2 = x1 + μ1 x2 , (4.33)
⎩
ẋ3 = 0.2 + μ2x3 (x1 − 5.7),

where μ1 and μ2 are system parameters. When the desired parameter values are
μ1g = 0.2 and μ2g = 1, system (4.33) corresponds to a chaotic state (see Fig. 4.16).
Suppose that the reference model of system (4.33) is chosen as
⎧
⎨ ẋ1 = −x2 − x3 ,
ẋ2 = x1 + 0.2x2,
⎩
ẋ3 = 0.2 + x3(x1 − 5.7).

Suppose that the parameters μ1 and μ2 of system (4.33) are disturbed from the target
values (μ1g = 0.2, μ2g = 1), and set to μ1 = −0.8, μ2 = 2.8, which corresponds to a
nonchaotic state. The control automatically changes the parameters until the system
4.3 Model-Reference Adaptive Control of a Class of Continuous-Time Chaotic Systems 119

2.5

1.5

0.5
x2

−0.5

−1

−1.5

−2

−2.5
−60 −40 −20 0 20 40 60 80
t (sec)

Fig. 4.7 The changed state response curve of x2 (t), which is disturbed to be chaotic at t = −30,
and approaches the desired state when the controller is applied at t = 0

0.85

0.8

0.75

0.7

0.65
p1

0.6

0.55

0.5

0.45

0.4

0.35
0 10 20 30 40 50 60 70 80
t (sec)

Fig. 4.8 Parameter p1 (disturbed initial value p1 (0) = 0.4) approaches the desired value p1g = 0.7
120 4 Feedback Control of Chaotic Systems

−0.6

−0.7

−0.8

−0.9
p2

−1

−1.1

−1.2

−1.3
0 10 20 30 40 50 60 70 80
t (sec)

Fig. 4.9 Parameter p2 (disturbed initial value p2 (0) = −1.1) approaches the desired value p2g =
−0.8

1.8

1.6

1.4

1.2
q

0.8

0.6

0.4
0 10 20 30 40 50 60 70 80
t (sec)

Fig. 4.10 Parameter q (disturbed initial value q(0) = 1.8) approaches the desired value qg = 0.6
4.3 Model-Reference Adaptive Control of a Class of Continuous-Time Chaotic Systems 121

1.5

0.5
x1

−0.5

−1

−1.5

−2
−60 −40 −20 0 20 40 60 80
t (sec)

Fig. 4.11 External disturbance on the Duffing equation. The changed state response curve of x1 (t),
which is disturbed to be chaotic at t = −30, and approaches the desired state when the controller
is applied at t = 0

2.5

1.5

0.5
x2

−0.5

−1

−1.5

−2

−2.5
−60 −40 −20 0 20 40 60 80
t (sec)

Fig. 4.12 External disturbance on the Duffing equation. The changed state response curve of x2 (t),
which is disturbed to be chaotic at t = −30, and approaches the desired state when the controller
is applied at t = 0
122 4 Feedback Control of Chaotic Systems

0.8

0.75

0.7

0.65

0.6
p1

0.55

0.5

0.45

0.4

0.35
0 10 20 30 40 50 60 70 80
t (sec)

Fig. 4.13 External disturbance on the Duffing equation. Parameter p1 (disturbed initial value
p1 (0) = 0.4) approaches the desired value p1g = 0.7

−0.6

−0.7

−0.8

−0.9
p2

−1

−1.1

−1.2

−1.3
0 10 20 30 40 50 60 70 80
t (sec)

Fig. 4.14 External disturbance on the Duffing equation. Parameter p2 (disturbed initial value
p2 (0) = −1.1) approaches the desired value p2g = −0.8
4.3 Model-Reference Adaptive Control of a Class of Continuous-Time Chaotic Systems 123

1.8

1.6

1.4

1.2
q

0.8

0.6

0.4
0 10 20 30 40 50 60 70 80
t (sec)

Fig. 4.15 External disturbance on the Duffing equation. Parameter q (disturbed initial value q(0) =
1.8) approaches the desired value qg = 0.6

25

20

15
x3

10

0
10
15
0 10
5
−10 0
−5
x2 −20 −10 x1

Fig. 4.16 The chaotic attractor of the Rössler system

124 4 Feedback Control of Chaotic Systems

1.2

0.8

0.6

0.4
μ1

0.2

−0.2

−0.4

−0.6

−0.8
0 10 20 30 40 50 60 70 80
t (sec)

Fig. 4.17 Parameter μ1 (disturbed initial value μ1 (0) = −0.8) approaches the desired value μ1g =
0.2

2.5

2
μ2

1.5

0.5
0 10 20 30 40 50 60 70 80
t (sec)

Fig. 4.18 Parameter μ2 (disturbed initial value μ2 (0) = 2.8) approaches the desired value μ2g = 1
4.4 Control of a Class of Chaotic Systems Based on Inverse Optimal Control 125

returns to the initial target state. Using the parametric adaptive law (4.24), we get

μ̇1 = −x2 [x2 (μ1 − 0.2) + x3(x1 − 5.7)(μ2 − 1)],
(4.34)
μ̇2 = −x3 (x1 − 5.7)[x2(μ1 − 0.2) + x3(x1 − 5.7)(μ2 − 1)].

The simulation results are shown in Figs. 4.17 and 4.18. The figures show that
the disturbed parameters converge quickly to the desired parameter values after the
transient, which indicates that the parametric adaptive control law (4.24) is very
powerful in stabilizing the system. This example makes it clear that the control
method can be applied to control a nonchaotic state to a chaotic state. For example, a
chaotic diode resonator should be kept in a chaotic state under external disturbances.

4.4 Control of a Class of Chaotic Systems Based on Inverse

Optimal Control

4.4.1 Problem Description

Consider a four-dimensional chaotic system [5] as follows:

⎧
⎪
⎪ ẋ1 = a (x2 − x1) + x2 x3 x4 ,
⎨
ẋ2 = b (x1 + x2) − x1 x3 x4 ,
(4.35)
⎪
⎪ ẋ3 = −cx3 + x1 x2 x4 ,
⎩
ẋ4 = −dx4 + x1 x2 x3 ,

where x = (x1 , x2 , x3 , x4 )T ∈ R4 is the state vector a, b, c, and d are known parame-

ters of the system (4.35), which satisfy the following conditions: a > 0, b > 0, c > 0,
d > 0, and b − (a + c + d) < 0.
By choosing various parameters, system (4.35) has 65 unstable equilibria: S0 : (0,
0, 0, 0)T , Sk : (xi1 , xi2 , x3j , x4j )T , k = 1, . . . , 64, i = 1, . . . , 8, j = 1, . . . , 8, where

1  1 
x11 = √ a (a + b + p) q, x21 = − √ a (a + b + p) q ,
2a 2a
1 1
x31 = −2a (a + b − p) q, x41 = − −2a (a + b − p) q ,
2a 2a
1  1 
x51 = √ a (a + b − p) q, x61 = − √ a (a + b − p) q ,
2a 2a
1 1
x71 = −2a (a + b + p) q, x18 = − −2a (a + b + p) q ,
2a .   2a
q j 1 j 2 j c j
xi2 = ± i , x3 = ± ab x1 − adq, x4 = ± x3 ,
x1 q q
√ 
q = cd, p = a2 + 6ab + b2.
126 4 Feedback Control of Chaotic Systems

20

10

0
x3

−10

−20

20
10
0
x2
−10
−20 10 20
−10 0
−20 x1

Fig. 4.19 The projection of the chaotic attractor of system (4.35) in the space of x1 –x2 –x3 with
a = 30, b = 10, c = 37, and d = 10

The chaotic behavior of the system (4.35) is shown in Fig. 4.19, when the parameters
of the system (4.35) are chosen as a = 30, b = 10, c = 37, and d = 10 and the initial
condition of the system (4.35) is chosen as x(0) = (−10, 10, 10, 20)T .

4.4.2 Inverse Optimal Controller Design

Suppose that the controlled chaotic system is

⎧
⎪ ẋ1 = a (x2 − x1 ) + x2 x3 x4 ,
⎪
⎪
⎪
⎪
⎪
⎪
⎨ ẋ2 = b (x1 + x2 ) − x1 x3 x4 + 1 u,
⎪
3 (4.36)
⎪
⎪
⎪
⎪ ẋ3 = −cx3 + x1 x2 x4 ,
⎪
⎪
⎪
⎪
⎩ ẋ = −dx + x x x ,
4 4 1 2 3

where u is the control input. We have the following theorem.

Theorem 4.3 ([12]). If the linear state feedback controller is designed as

 
(a + 3b)2
u = − 6b + + 1 x2 , (4.37)
2a
4.4 Control of a Class of Chaotic Systems Based on Inverse Optimal Control 127

then the zero equilibrium of the controlled four-dimensional chaotic system (4.36)
is globally asymptotically stable.
Proof. Construct a Lyapunov function as
1 2 
V= x1 + 3x22 + x23 + x24 .
2
Then, the time derivative of V along the trajectory of system (4.36) is given as
follows:

V̇ = x1 ẋ1 + 3x2ẋ2 + x3 ẋ3 + x4 ẋ4

= a (x2 − x1 ) x1 + x1 x2 x3 x4 + 3b (x1 + x2 ) x2 − 3x1x2 x3 x4
− cx23 + x1 x2 x3 x4 − dx24 + x1x2 x3 x4 + x2 u
2
(a + 3b)
= − a x1 − x2 − cx23 − dx24
2a
' (
(a + 3b)2
+ 3b + x22 + x2 u
4a
= L f V + (LgV ) u, (4.38)

where
⎧ ' (
2
⎪
⎨ (a + 3b) (a + 3b)2
L f V = −a x1 − x2 − cx23 − dx24 + 3b + x22 ,
2a 4a (4.39)
⎪
⎩
LgV = x2 .

From (4.38), we can easily obtain L f V ≤ 0 if LgV = 0.

Next, we design a simple linear state feedback controller as follows:
' (
−1 3 (a + 3b)2
u = −β R (x) (LgV ) = − 3b + + k0 x2 , (4.40)
4a
' (
2
1 3 (a + 3b)
where β > 0 is a constant, R−1 (x) = 3b + + k0 , and k0 > 0 is a
β 4a
constant to be determined later.
Substituting (4.40) into (4.38), we obtain
' (
2
(a + 3b) 2 2 (a + 3b)2
V̇ = −a x1 − x2 − cx3 − dx4 − k0 + x22 ,
2a 2a

which implies that V̇ < 0 for all x = 0, i.e., the controller (4.40) can globally asymp-
totically stabilize the closed-loop system (4.36) .
Based on the idea of inverse optimal control, in order to determine the value of
k0 , we define the following performance index:
128 4 Feedback Control of Chaotic Systems
%  t  &
J(u) = lim 2β V (x(t)) + l(x(τ )) +uT (x (τ )) R (x (τ )) u (x (τ )) dτ , (4.41)
t→∞ 0

where
l (x) = −2β L f V + β 2R−1 (x) (LgV )2 > 0 . (4.42)
Substituting (4.39) into (4.42), we can derive
2
(a + 3b)
l (x) = 2aβ x1 − x2 + 2cβ x23 + 2d β x24
2a
' ( ' (
(a + 3b)2 2 3 (a + 3b)2
− 2β 3b + x2 + β 3b + + k0 x22 . (4.43)
4a 4a

Let l (x) > 0 and

(a + 3b)2 (4.44)
k0 = 3b − +1 .
4a
From (4.43), we obtain
2
(a + 3b)
l (x) = 2aβ x1 − x2 + 2cβ x23 + 2d β x24 + β x22 > 0.
2a

Substituting (4.40) into (4.38), we get

V̇ = L f V − β R−1 (x) (LgV )2 . (4.45)

Multiplying −2β on both sides of (4.45), and noting (4.42), we have

−2β V̇ = l (x) + β 2 R−1 (x) (LgV )2 . (4.46)

From (4.40) we have

uT R (x) u = β 2 R−1 (x) (LgV )2 . (4.47)
Substituting (4.47) into (4.46), we derive

l (x) + uTR (x) u = −2β V̇ . (4.48)

Substituting (4.48) into (4.41), we have

  t /

J (u) = lim 2β V (x (t)) + −2β V̇ dτ = 2β V (x (0)) .
t→∞ 0

Therefore, the controller (4.40) is optimal with performance index (4.41). Substitut-
ing (4.44) into (4.40), we get the optimal controller (4.37). 
4.4 Control of a Class of Chaotic Systems Based on Inverse Optimal Control 129

30

25

20

15
x3

10

0
30
20 30
10 20
0 10
x2 0
−10 −10 x1
−20 −20

Fig. 4.20 Globally asymptotically stable system with a = 30, b = 10, c = 37, and d = 10

10

5
x3

0
−5

−6
−4
0
−2
x2 0
2 x1
4
5 6

Fig. 4.21 The projection of the chaotic attractor of system (4.35) in the space of x1 –x2 –x3 with
a = 35, b = 10, c = 1, and d = 10
130 4 Feedback Control of Chaotic Systems

30

25

20
x3

15

10

0
30
20 30
10 20
0 10
x2 0
−10 −10 x1
−20 −20

Fig. 4.22 Globally asymptotically stable system with a = 35, b = 10, c = 1, and d = 10

4.4.3 Simulations

The following illustrative example is used to demonstrate the effectiveness of the

above method.
According to Theorem 4.3, we get u = −121x2. As shown in Fig. 4.20, the
chaotic trajectory of the system can be controlled rapidly to its zero equilibrium
under different initial conditions. If the parameters of system (4.35) are chosen as
a = 35, b = 10, c = 1, and d = 10, and the initial condition of the system (4.35) is
selected as x(0) = (3, 3, 3, 3)T , the chaotic behavior of the system (4.35) is shown
in Fig. 4.21. According to Theorem 4.3, we get u = −1699x2 /14. As shown in
Fig. 4.22, the chaotic trajectory of the system can be controlled rapidly to its zero
equilibrium under different initial conditions, too. The above numerical simulation
results show the effectiveness of the controller.

4.5 Summary

In this chapter, we investigated the problem of feedback control of discrete-time

or continuous-time chaotic systems via different control methods. In Sect. 4.2, we
developed a parametric adaptive controller to control discrete-time chaotic systems
with multiple parameters. In Sect. 4.3, another parametric adaptive control method
was developed to control continuous-time chaotic systems. The considered systems
References 131

are both linearly parameterized and nonlinearly parameterized, respectively. In Sect.

4.4, we studied how to stabilize a chaotic system by inverse optimal control theory
in detail. This method can avoid the difficulty caused by solving the HJB equation,
and hence is very efficient in practice when the controller is demanded to be optimal
in a certain sense.

References

1. Freeman RA, Kokotovic PV (1996) Inverse optimality in robust stabilization. J Control Optim
34:1365–1391
2. Guckenheimer J, Holmes P (1983) Nonlinear Oscillations, Dynamical Systems and Bifurca-
tions of vector Fields. Spinger, New York
3. Khadra A, Liu XZ, Shen XM (2005) Impulsive control and synchronization of spatiotemporal
chaos. Chaos Solitons Fractals 26:615–636
4. Poston T, Steward I (1978) Catastrophe Theory and its Application. Pitman, London
5. Qi GY, Du SZ, Chen GR, Chen ZQ, Yuan ZZ (2005) On a four-dimensional chaotic system.
Chaos Solitons Fractals 23:1671–1682
6. Rafikov M, Balthazar JM (2004) On an optimal control design for Rössler system. Phys Lett
A 333:241–245
7. Sinha S, Ramaswamy R, Rao JS (1990) Adaptive control in nonlinear dynamics. Physica D
43:118–128
8. Tanaka K, Ikeda T, Wang HO (1998) A unified approach to controlling chaos via an LMI-
based fuzzy control system design. IEEE Trans Circuits Syst I 45:1021–1040
9. Tian YP (2005) Delayed feedback control of chaos in a switched arrival system. Phys Lett A
339:446–454
10. Vassiliadis D (1994) Parametric adaptive control and parameter identification of low-
dimensional chaotic systems. Physica D 71:319–341
11. Wang J, Wang X (1998) Parametric adaptive control in nonlinear dynamical systems. Int J
Bifurc Chaos 8:2215–2223
12. Xie YH (2007) Researches on control and synchronization of several kinds of nonlinear
chaotic systems with time delay. Doctoral dissertation, Northeastern University, Shenyang
13. Yassen MT (2005) Controlling chaos and synchronization for new chaotic system using linear
feedback control. Chaos Solitons Fractals 26:913–920
14. Yau HT, Chen CL (2007) Chaos control of Lorenz systems using adaptive controller with
input saturation. Chaos Solitons Fractals 34:1567–1574
Chapter 5
Synchronizing Chaotic Systems Based on
Feedback Control

Abstract The goal of this chapter is to present some methods that can be used
to synchronize a rather wide class of chaotic systems. Considering that a chaotic
system in nature is a nonlinear system, the theories and methods developed for con-
trolling nonlinear systems could be utilized for synchronizing chaotic systems. The
content of this chapter is about the synchronization of chaotic systems which can be
described by ordinary differential equations and ordinary difference equations. The
method is mainly based on the technology of feedback linearization. In this chapter
we first develop synchronization methods for continuous-time chaotic systems, and
after that we study the problem of synchronizing discrete-time chaotic systems. We
also study the situation of synchronizing two different chaotic systems with adaptive
and nonadaptive controllers.

5.1 Introduction

Although many methods have been proposed [1, 3] for synchronizing two or more
chaotic systems since the landmark work by Pecora and Carroll [2], a systematic
framework under which the synchronizing controller can be designed is still not
constructed. It is usually the case that one method is only applicable to some speci-
fied chaotic systems. Our goal of this chapter is to present some methods that work
for a rather wide class of chaotic systems. Considering that a chaotic system in
nature is a nonlinear system, the theories and methods developed for controlling
nonlinear systems could be utilized for the synchronization of chaotic systems. The
content of this chapter is about the synchronization of chaotic systems which can
be described by ordinary differential equations and ordinary difference equations.
The method is based on the technology of feedback linearization. In this chapter
we first develop synchronization methods for continuous-time chaotic systems, and
after that we study the problem of synchronizing discrete-time chaotic systems.

133
134 5 Synchronizing Chaotic Systems Based on Feedback Control

5.2 Synchronization of Continuous-Time Chaotic Systems with a

Single Input

5.2.1 Basic Formulation

Consider the following single-input single-output (SISO) nonlinear system:


ẋ = f (x) + g(x)u,
(5.1)
y = h(x),

where x ∈ Rn represents the state variables, f , g : Rn → Rn are vector fields, u ∈ R

is the input, y ∈ R is the output, and h : Rn → R is the output function. Denote
Bρ (x0 ) = {x ∈ Rn : x − x0 < ρ , ρ > 0}, the neighborhood of x0 in Rn . We need
the following definition.
Definition 5.1. System (5.1) is said to have relative degree r if
(i) Lg Lkf h(x) = 0, for 0 ≤ k < r − 1 and for all x ∈ Bρ (x0 );

(ii) Lg Lr−1
f h(x) = 0.

∂ h(x) ∂ Lk−1
f h(x)
Here, L f h(x) := f (x), Lkf h(x) := L f Lk−1
f h(x) = f (x), Lg Lkf h(x) :=
∂x ∂x
∂ Lkf h(x)
g(x), and L0f h(x) := h(x). 
∂x
Theorem 5.1 ([9]). If system (5.1) has relative degree r at x0 , then r ≤ n. Set
⎧
⎪
⎪ φ1 (x) = h(x),
⎪
⎨ φ2 (x) = L f h(x),
.. (5.2)
⎪
⎪ .
⎪
⎩
φr (x) = Lr−1
f h(x).

If r < n, it is always possible to find n − r additional functions φr+1 (x), . . . , φn (x)

such that the mapping
⎛ ⎞
φ1 (x)
⎜ ⎟
Φ (x) = ⎝ ... ⎠
φn (x)

has a Jacobian matrix which is nonsingular at x0 and therefore qualifies as a local

coordinate transformation in a neighborhood of x0 . The value of these additional
functions at x0 can be fixed arbitrarily. Moreover, it is always possible to choose
φr+1 (x), . . . , φn (x) in such a way that

Lg φi (x) = 0 for all r + 1 ≤ i ≤ n and all x around x0 . 

5.2 Synchronization of Continuous-Time Chaotic Systems with a Single Input 135

Theorem 5.2 ([9]). If system (5.1) has relative degree r at x0 , then by transformation
z = Φ (x) system (5.1) can be transformed to
⎧
⎪ ż1 = z2 ,
⎪
⎪
⎪
⎪ ż2 = z3 ,
⎪
⎪
⎪
⎪ ..
⎪
⎪ .
⎪
⎪
⎪
⎨ żr−1 = zr ,
żr = b(z) + a(z)u, (5.3)
⎪
⎪
⎪
⎪ żr+1 = q r+1 (z),
⎪
⎪ ..
⎪
⎪
⎪
⎪ .
⎪
⎪
⎪
⎪ żn = qn (z),
⎩
y = z1 ,

where a(z) = Lg Lr−1 r −1

f h(x), b(z) = L f h(x), and qi (z) = L f φi (Φ (z)) for all r + 1 ≤
i ≤ n. 

System (5.3) can be written in a compact form:

⎧ 1 1
⎨ ż = Λ z + E(b(z) + a(z)u),
⎪
ż2 = Q(z1 , z2 ), (5.4)
⎪
⎩
y = z1 ,

where z1 = (z1 , z2 , . . . , zr )T , z2 = (zr+1 , zr+2 , . . . , zn )T ,

⎛ ⎞
0 1 0 ··· 0
⎜ 0 0 1 ··· 0 ⎟
⎜ ⎟
⎜ ⎟
Λ = ⎜ ... ... ... . . . ... ⎟
⎜ ⎟
⎝ 0 0 0 ··· 1 ⎠
0 0 0 ··· 0

is an r × r matrix, E = (0, 0, . . . , 0, 1)Tr×1 , and Q(z1 , z2 ) = (qr+1 (z), . . . , qn (z))T .

When z1 = 0, the second equation of (5.4) becomes

ż2 = Q(0, z2 ). (5.5)

System (5.5) is called the zero dynamics which reflects the internal stability of sys-
tem (5.4).

Remark 5.1. If r = n, then system (5.3) becomes

136 5 Synchronizing Chaotic Systems Based on Feedback Control
⎧
⎪ ż1 = z2 ,
⎪
⎪
⎪
⎪ ż2 = z3 ,
⎪
⎪
⎪
⎨ ..
. (5.6)
⎪
⎪ żn−1 = zn ,
⎪
⎪
⎪
⎪ żn = b(z) + a(z)u,
⎪
⎪
⎩
y = z1 .

If a(z) = 0 for any Φ −1 (z) ∈ Bρ (x0 ), then by setting

1
u= (−b(z) + v),
a(z)

the nonlinear control system (5.6) becomes a linear control system given by
⎧
⎪ ż1 = z2 ,
⎪
⎪
⎪
⎪ ż = z3 ,
⎪
⎪ 2
⎪
⎨ ..
.
⎪
⎪ ż = zn ,
⎪
⎪
n−1
⎪
⎪ żn = v,
⎪
⎪
⎩
y = z1 .

5.2.2 Main Results

Suppose that we have the following chaotic system:

ẋ = f (x),
(5.7)
y = h(x).

We construct the following controlled response system:

x̂˙ = f (x̂) + g(x̂)u,

(5.8)
ŷ = h(x̂),

such that system (5.8) has relative degree r (r ≤ n). Set ẑ = Φ (x̂), where Φ is defined
as (5.2). Thus, system (5.8) can be transformed to
⎧ 1
˙ 1
⎨ ẑ = Λ ẑ + E(b(ẑ) + a(ẑ)u),
⎪
ẑ˙2 = Q(ẑ1 , ẑ2 ),
⎪
⎩
ŷ = ẑ1 .
5.2 Synchronization of Continuous-Time Chaotic Systems with a Single Input 137

Similarly, by z = Φ (x) system (5.7) is transformed to

⎧ 1 1
⎨ ż = Λ z + Eb(z),
⎪
ż2 = Q(z1 , z2 ),
⎪
⎩
y = z1 .

Defining e = (e1 , e2 )T = (ẑ1 −z1 , ẑ2 −z2 )T , we can obtain the following error system:

ė1 = Λ e1 + E(b(ẑ1 , ẑ2 ) + a(ẑ1 , ẑ2 )u − b(z1, z2 )),

(5.9)
ė2 = Q(ẑ1 , ẑ2 ) − Q(z1 , z2 ).

Theorem 5.3 ([15]). Denote C = (c0 , c1 , . . . , cr−1 ), where c0 , . . . , cr−1 are coeffi-
cients of the Hurwitz polynomial P(s) = sr + cr−1 sr−1 + · · · + c1 s + c0 . If a(ẑ) = 0
for all z ∈ Bρ (Φ (x0 )) and the zero dynamics of system (5.8) is asymptotically stable,
then with the controller
1
u=− (b(ẑ1 , ẑ2 ) − b(z1 , z2 ) + Ce) (5.10)
a(ẑ1 , ẑ2 )

the origin of system (5.9) is asymptotically stable.

Proof. Since Φ is nonsingular and the zero dynamics of system (5.9) is asymptoti-
cally stable, z and ẑ are both bounded. Therefore, the output signal of the controller
is also bounded. Substituting (5.10) into the first equation of (5.9), we have

ė1 = Λ̃ e1 , (5.11)
⎛ ⎞
0 1 0 ··· 0
⎜ 0 0 1 ··· 0 ⎟
⎜ ⎟
⎜ .. ⎟ . By the selection of C we know that all
where Λ̃ = ⎜ ... ..
.
.. . .
. . . ⎟
⎜ ⎟
⎝ 0 0 ··· 0 1 ⎠
−c0 −c1 −c2 · · · −cr−1
eigenvalues of Λ̃ have real parts which are less than 0, and therefore the origin of
(5.11) is asymptotically stable. The proof is completed. 

Remark 5.2. The above theorem guarantees that outputs of system (5.7) and system
(5.8) can be synchronized asymptotically, i.e., lim |y − ŷ| = 0. 
t→∞

5.2.3 Simulations

We take both the Rössler system and the Lorenz system as examples to show how
the method developed above is used.
138 5 Synchronizing Chaotic Systems Based on Feedback Control

5.2.3.1 Synchronization of the Rössler System

The state equations of the Rössler system are as follows:

⎧
⎨ ẋ = −y − z,
⎪
ẏ = x + ay, (5.12)
⎪
⎩
ż = b + z(x − c).

When a = 0.2, b = 0.2, and c = 5.7, this system has a chaotic attractor, which
is shown in Fig. 5.1. Taking system (5.12) as the drive system, we construct the
following response system:
⎧
˙
⎨ x̂ = −ŷ − ẑ,
⎪
ŷ˙ = x̂ + aŷ, (5.13)
⎪
⎩˙
ẑ = b + ẑ(x̂ − c) + u,

where u is the controller and g(x̂) = (0, 0, 1)T . Choosing the second state of system
(5.12) as the output, we can easily verify that the system (5.13) has relative degree
3. The transformation can be chosen as (ŷ, x̂ + aŷ, ax̂ + (a2 − 1)ŷ − ẑ)T . Select C =

25

20

15
x3

10

0
10
5 15
0 10
−5 5
0
−10
−5
x2 −15 −10 x1

Fig. 5.1 The attractor of Rössler system

5.2 Synchronization of Continuous-Time Chaotic Systems with a Single Input 139

40

20
x̂1 − x1

−20
0 10 20 30 40 50 60 70 80 90 100

20

10
x̂2 − x2

−10

−20
0 10 20 30 40 50 60 70 80 90 100

200

100
x̂3 − x3

−100

−200
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 5.2 The error curves of system (5.12) and system (5.13)

(27, 27, 9). Initial values of system (5.12) and system (5.13) are set as (30, 10, 10)T
and (20, 12, 10)T, respectively. When t = 40 sec the controller is turned on. Fig. 5.2
presents the error curves of the synchronization between systems (5.12) and (5.13).
Since the coordinate transformation is linear, all three state variables of the system
(5.13) will converge to those of the system (5.12).

5.2.3.2 Synchronization of the Lorenz System

The Lorenz system is described by the following differential equations:

⎧
⎨ ẋ = σ (y − x),
⎪
ẏ = ρ x − y − xz, (5.14)
⎪
⎩
ż = xy − β z.

When σ = 10, β = 8/3, and ρ = 28, the Lorenz system has a chaotic attractor,
which is the well-known butterfly attractor and is shown in Fig. 5.3.
We take system (5.14) as the drive system and construct the controlled response
system as follows:
140 5 Synchronizing Chaotic Systems Based on Feedback Control

50

40

30
x3

20

10

0
40

20

0
15 20
−20 10
0 5
−10 −5
x2 −40 −15
−20
x1

Fig. 5.3 The attractor of Lorenz system

⎧
˙
⎨ x̂ = σ (ŷ − x̂),
⎪
ŷ˙ = ρ x̂ − ŷ − x̂ẑ + u, (5.15)
⎪
⎩˙
ẑ = x̂ŷ − β ẑ.
Here, the output is chosen as h(x̂) = x̂ and g = [0, 1, 0]T . By simple calculation we
have Lg h(x̂) = 0, L f h(x̂) = σ (ŷ − x̂), and Lg L f h(x̂) = σ = 0. Therefore, the relative
degree of system (5.15) is 2. Choosing the coordinate transformation as

ẑ = (ẑ1 , ẑ2 , ẑ3 )T = (h(x̂), L f h(x̂), ẑ)T , (5.16)

we can transform system (5.15) into the following form:

⎧
⎪
⎪ ẑ˙1 = ẑ2 ,
⎪
⎨
ẑ˙2 = L2f h(ẑ) + Lg L f h(ẑ)u,
⎪
⎪
⎩ ẑ˙3 = 1 (σ ẑ21 + ẑ1 ẑ2 ) − β ẑ3 .
⎪
σ
Similarly, system (5.14) can be changed to
5.2 Synchronization of Continuous-Time Chaotic Systems with a Single Input 141
⎧
⎪
⎪ ż1 = z2 ,
⎪
⎨
ż2 = L2f h(z),
⎪
⎪ 1
⎪
⎩ ż3 = (σ z21 + z1 z2 ) − β z3 .
σ
Here, L2f h(z) = (σ 2 + σ ρ )x−(σ 2 + σ )y− σ xz, Lg L f h(z) = −σ , and z = (z1 , z2 , z3 )T .
Set e = ẑ − z. The error system can be obtained as
⎧
⎪
⎪ ė1 = e2 ,
⎪
⎨
ė2 = L2f h(ẑ) − L2f h(z) + Lg L f h(ẑ)u, (5.17)
⎪
⎪ 1
⎪
⎩ ė3 = e1 (ẑ1 + z1 ) + (ẑ2 e1 + z1 e2 ) − β e3.
σ
The zero dynamics equation of system (5.17) is

ė3 = −ρ e3,

20

0
x̂1 − x1

−20

−40
0 5 10 15 20 25 30
t (sec)
50
x̂2 − x2

−50
0 5 10 15 20 25 30
t (sec)
50
x̂3 − x3

−50
0 5 10 15 20 25 30
t (sec)

Fig. 5.4 The error curves of system (5.14) and system (5.15)
142 5 Synchronizing Chaotic Systems Based on Feedback Control

which is asymptotically stable. Select C = (4, 4). Under the action of controller
1
u=− (L2 h(ẑ) − L2f h(z) +Ce) the origin of system (5.17) is asymptotically
Lg L f h(ẑ) f
stable. Furthermore, for any x̂ ∈ R3 it is always true that Lg L f h(x̂) = 0, and the origin
e = 0 is also globally and asymptotically stable. Since the controller is added on the
second equation of (5.15), if x̂ tends to x it implies that ŷ tends to y. Recall that
e3 = ẑ − z. The three state variables of system (5.15) will be synchronized to those
of system (5.14). Fig. 5.4 validates the above analysis.

5.3 Synchronization of Multi-Signals in Continuous-Time

Chaotic Systems

The method proposed in the last section can only synchronize a single signal be-
tween the drive system and the response system. In this section we study how to
synchronize multi-signals simultaneously, which is especially useful in secret com-
munications when channel resources are not enough.

5.3.1 Basic Formulation

To synchronize multi-output signals we first generalize the notion of relative degree

so that the multi-input multi-output (MIMO) nonlinear control theory can be used
to design the controller.
Definition 5.2. Suppose that f1 and f2 are smooth vector fields of Rn . The Lie
bracket of f1 and f2 is defined as

∂ f2 ∂ f1
[ f1 , f2 ](x) = f1 (x) − f2 (x),
∂x ∂x
where x is in the domain of f1 and f2 . 

Definition 5.3. Given k smooth vector fields f1 , . . . , fk defined in U ∈ Rn , the notion

Δ = span{ f1 , . . . , fk }

is called the distribution spanned by f1 , . . . , fk . For any x ∈ U,

Δ (x) = span{ f1 (x), . . . , fk (x)}. 

Remark 5.3. For each x ∈ U, Δ (x) is actually a vector space in Rn . 

Definition 5.4. A distribution Δ is involutive if the Lie bracket [ f1 , f2 ] of any pair

of vector fields f1 and f2 belonging to Δ is a vector field which belongs to Δ , i.e.,
5.3 Synchronization of Multi-Signals in Continuous-Time Chaotic Systems 143

f1 ∈ Δ , f2 ∈ Δ ⇒ [ f1 , f2 ] ∈ Δ . 

Consider the following MIMO system:

ẋ = f (x) + g(x)u,
(5.18)
y = h(x),

where f : Rn → Rn , f ∈ C∞ , u = (u1 , . . . , um )T , y = (y1 , . . . , ym )T , h = (h1 , . . . , hm )T ,

g(x) = (g1 (x), . . . , gm (x)), gi : Rn → Rn , hi : Rn → R, hi ∈ C∞ , and ui ∈ R, i =
1, . . . , m, 1 ≤ m ≤ n.
Definition 5.5. MIMO system (5.18) has a vector relative degree {r1 , . . . , rm } at x0
if
(i) Lg j Lkf hi (x) = 0, 1 ≤ ∀i, j ≤ m, ∀k ≤ ri − 1, ∀x ∈ Bρ (x0 );
(ii) the m × m matrix
⎛ ⎞
Lg1 Lrf1 −1 h1 (x) · · · Lgm Lrf1 −1 h1 (x)
⎜ r −1 r −1 ⎟
⎜ Lg1 L f2 h2 (x) · · · Lgm L f2 h2 (x) ⎟
⎜ ⎟
⎜ .. .. .. ⎟
⎝ . . . ⎠
rm −1 rm −1
Lg1 L f hm (x) · · · Lgm L f hm (x)

is nonsingular at x = x0 . 
Theorem 5.4 ([9]). Suppose that system (5.18) has a vector relative degree {r1 , . . . ,
rm } at x0 . Then,
r1 + · · · + rm ≤ n.
For 1 ≤ i ≤ m, set ⎧ i
⎪
⎪
φ1 (x) = hi (x),
⎪
⎪
⎪
⎨ φ2i (x) = L f hi (x),
⎪ ..
⎪
⎪ .
⎪
⎪
⎩ i
φri (x) = Lrfi −1 hi (x).
If r = r1 + · · · + rm is strictly less than n, it is always possible to find n − r additional
functions φr+1 (x), . . . , φn (x) such that the mapping

Φ (x) = [(φ11 (x), . . . , φr11 , . . . , φ1m (x), . . . , φrmm (x), φr+1 (x), . . . , φn (x))T (5.19)

has a Jacobian matrix which is nonsingular at x0 and therefore qualifies as a local

coordinate transformation in a neighborhood of x0 . Moreover, if the distribution

G = span{g1 , . . . , gm }

is involutive near x0 , it is always possible to choose φr+1 (x), . . . , φn (x) in such a way
that
144 5 Synchronizing Chaotic Systems Based on Feedback Control

Lg j φi (x) = 0
for all r + 1 ≤ i ≤ n, all 1 ≤ j ≤ m, and all x around x0 . 
Remark 5.4. By transformation Φ system (5.18) is changed to
⎧
⎨ ξ̇ = Λ ξ + E(A(ξ , η )u + B(ξ , η )),
⎪
η̇ = Q(ξ , η ), (5.20)
⎪
⎩
y = ξ,

where ξ ∈ Rr , ξ = (ξ 1 , . . . , ξ m )T , ξ i = (ξ1i , . . . , ξrii )T = (φ1i , . . . , φrii )T , η ∈ Rn−r , and

η = (φr+1 , . . . , φn )T . Λ and E are block diagonal matrices, i.e.,

Λ = diag{Λi } and E = diag{Ei },

where ⎛ ⎞
0 1 0 ···0
⎜0 0 1 ···0⎟
⎜ ⎟
⎜ .. ⎟
Λi = ⎜ ... ..
.
.. . .
. . .⎟
⎜ ⎟
⎝0 0 0 ··· 1 ⎠
0 0 0 · · · 0 r ×r
i i

and
Ei = (0, . . . , 0, 1)Tri ×1
for all 1 ≤ i ≤ m,
⎛ ⎞
Lg1 Lrf1 −1 h1 (ξ , η ) · · · Lgm Lrf1 −1 h1 (ξ , η )
⎜ r −1 r −1 ⎟
⎜ Lg1 L f2 h2 (ξ , η ) · · · Lgm L f2 h2 (ξ , η ) ⎟
A(ξ , η ) = ⎜
⎜ .. ..
⎟
⎟
⎝ . ··· . ⎠
Lg1 Lrfm −1 hm (ξ , η ) · · · Lgm Lrfm −1 hm (ξ , η )
m×m,

and
B(ξ , η ) = (Lrf1 h1 (ξ , η ), Lrf2 h2 (ξ , η ), . . . , Lrfm hm (ξ , η ))T .
When ξ = 0, the second equation of (5.20) becomes

η̇ = Q(0, η )

and it is called the zero dynamics. 

Theorem 5.5. Suppose that system (5.18) has relative degree r and asymptotically
stable zero dynamics. Then, with the controller
1
u=− (B(ξ , η ) + Cξ )
A−1(ξ , η )

the origin of system (5.20) is asymptotically stable. Here, C = diag{Ci } and

5.3 Synchronization of Multi-Signals in Continuous-Time Chaotic Systems 145

Ci = (ci0 , ci1 , . . . , ciri −1 )

are the coefficients of the Hurwitz polynomial

Pi (s) = sri −1 + ciri −1 sri −1 + ciri −2 sri −2 + · · · + ci0. 

5.3.2 Main Results

Suppose that the drive chaotic system has the following form:

ẋ = f (x),
(5.21)
y = h(x),

where f and h are smooth vector fields satisfying f : Rn → Rn , f (0) = 0, and

h : Rn → Rm . We construct the controlled response system as

x̂˙ = f (x̂) + g(x̂)u,

(5.22)
ŷ = h(x̂).

If the system (5.22) has relative degree r, then by nonsingular coordinate transfor-
mation (5.19), system (5.22) is changed to
⎧
⎪ ˙
⎨ ξ̂ = Λ ξ̂ + E(A(ξ̂ , η̂ )u + B(ξ̂ , η̂ )),
⎪

⎪ η̂˙ = Q(ξ̂ , η̂ ), (5.23)

⎪
⎩
ŷ = ξ̂ .

Similarly, system (5.21) is changed to

⎧
⎨ ξ̇ = Λ ξ + EB(ξ , η ),
⎪
η̇ = Q(ξ , η ),
⎪
⎩
y = ξ,

where the definitions of Λ , E, A(·, ·), B(·, ·), and C are the same as those in Remark
5.4. Denote e = ξ̂ − ξ and ε = η̂ − η . The error system can then be obtained as

ė = Λ e + E(A(ξ̂ , η̂ )u + B(ξ̂ , η̂ ) − B(ξ , η )),

(5.24)
ε̇ = Q(ξ̂ , η̂ ) − Q(ξ , η ) := F(e, ε , ξ , η ).

Theorem 5.6 ([15]). If MIMO system (5.22) has relative degree r in U ⊂ Rn as well
as asymptotically stable zero dynamics, then the following controller:
146 5 Synchronizing Chaotic Systems Based on Feedback Control

1
u=− (B(x̂) − B(x) + Ce) (5.25)
A−1 (x̂)

stabilizes the first equation of system (5.24) at e = 0, where e = (e1 , e2 , . . . , em )T ,

ei = (hi (x̂) − hi (x), L f hi (x̂) − L f hi (x), . . . , Lrfi −1 hi (x̂) − Lrfi −1 hi (x))T ,

B(x̂) = B(Φ −1 (ξ̂ , η̂ )), and B(x) = B(Φ −1 (ξ , η ))

Proof. Since the zero dynamics of (5.22) is stable, x̂ is bounded. Because (5.21) is
a chaotic system, x is also bounded. By the smoothness of f and h we know that
for each 0 ≤ k ≤ r and 1 ≤ i ≤ m, Lkf hi (x) is also bounded. Therefore, the controller
(5.25) is realizable. Substituting the controller into the first equation of (5.24), we
get
ė = Λ e, (5.26)
where Λ = diag(Λi ) and
⎛ ⎞
0 0 ···
1 0
⎜ 0 1 ···
0 0 ⎟
⎜ ⎟
⎜ .. ..
.. . . .. ⎟
Λi = ⎜ . .. . . ⎟
⎜ ⎟
⎝ 0 0 0 ··· 1 ⎠
−ci0 −ci1 −ci2 · · · −ciri −1 r ×r .
i i

Therefore, we have
m
det(λ I − Λ ) = ∏ det(λ Ii − Λi ).
i=1

From the selection of C we know that the real parts of all eigenvalues of Λ are less
than 0. Thus, the proof is completed. 

5.3.3 Simulations

We take the Rössler hyperchaotic system [14] as an example to validate the effec-
tiveness of the proposed method. The system equation is as follows:
⎧
⎪ ẋ1 = −(x2 + x3),
⎪
⎪
⎨ ẋ = x + 0.25x + x ,
2 1 2 4
(5.27)
⎪ ẋ
⎪ 3 = 3.0 + x x
1 3 ,
⎪
⎩
ẋ4 = 0.05x4 − 0.5x3.

The projection into R3 of the chaotic attractor is presented in Fig. 5.5. This sys-
tem has two positive Lyapunov exponents, i.e., λ1 = 0.16 and λ2 = 0.03, which
means that the trajectory of the system has two divergent directions. This fact im-
5.3 Synchronization of Multi-Signals in Continuous-Time Chaotic Systems 147

60
200

40
x3

x4
100

0 20
50 100 50
0 0 0 50
−50 −50 0
−100 −50
x2 −100 −200 x2 −100 −100
x1 x1

60 60

40 40
x4

x4

20 20
300 300
200 50 200 50
100 0 100 0
−50 −50
0 −100 x3 0 −100 x2
x3 x1

Fig. 5.5 The chaotic attractor of system (5.27)

plies that one cannot even stabilize the hyperchaotic system to the origin with only
one scalar controller. In the following, we will design a vector controller to synchro-
nize two Rössler hyperchaotic systems.
Taking system (5.27) as the drive system, we construct the following controlled
response system: ⎧
⎪
⎪ x̂˙1 = −(x̂2 + x̂3),
⎪
⎪
⎨ x̂˙ = x̂ + 0.25x̂ + x̂ + u ,
2 1 2 4 1
(5.28)
⎪
⎪ ˙
x̂3 = 3.0 + x̂1x̂3 + u2 ,
⎪
⎪
⎩˙
x̂4 = 0.05x̂4 − 0.5x̂3.
This means that g1 = (0, 1, 0, 0)T and g2 = (0, 0, 1, 0)T . Select outputs as h1 (x̂) = x̂1
and h2 (x̂) = x̂4 . By simple calculation, we have

Lg1 h1 (x̂) = Lg1 h2 (x̂) = Lg2 h1 (x̂) = Lg2 h2 (x̂) = 0

and
Lg1 L f h1 (x̂) Lg2 L f h1 (x̂) −1 −1
A(x̂) = = , (5.29)
Lg1 L f h2 (x̂) Lg2 L f h2 (x̂) 0 −0.5
148 5 Synchronizing Chaotic Systems Based on Feedback Control

150

100

50
x̂1 − x1

−50

−100
0 50 100 150 200 250 300
t (sec)

40

20
x̂4 − x4

−20

−40
0 50 100 150 200 250 300
t (sec)

Fig. 5.6 The error curves of the synchronized Rössler hyperchaotic systems

which is nonsingular. Therefore, system (5.28) has relative degree (2, 2). Choose a
coordinate transformation as

Φ (x̂) = (ẑ1 , ẑ2 , ẑ3 , ẑ4 )T = (h1 (x̂), L f h1 (x̂), h2 (x̂), L f h2 (x̂))T .

System (5.28) is changed to

ẑ˙ = Λ ẑ + E(B(ẑ) + A(ẑ)u),

where
⎛ ⎞ ⎛ ⎞
0 1 0 0 0 0 ' 2 (
⎜0 0 1 0⎟ ⎜1 0⎟ L f h1 (ẑ) u1
Λ =⎜
⎝0
⎟, E = ⎜ ⎟ , B(ẑ) = , and u = .
0 0 1⎠ ⎝0 0⎠ L2f h2 (ẑ) u2
0 0 0 0 0 1

A(ẑ) is equal to (5.29). Applying the same transformation to system (5.27), we get

ż = Λ z + EB(z).

Defining e = ẑ − z, we get the error system as

5.3 Synchronization of Multi-Signals in Continuous-Time Chaotic Systems 149

300

200
x̂2 − x2

100

−100
0 50 100 150 200 250 300
t (sec)

400

200
x̂3 − x3

−200

−400
0 50 100 150 200 250 300
t (sec)

Fig. 5.7 The error curves of the synchronized Rössler hyperchaotic systems

ė = Λ e + E(B(ẑ) − B(z) + A(ẑ)u).

Choose
625 50 0 0
C = diag{C1 ,C2 } = .
0 0 625 50
Initial values are set to

(x1 , x2 , x3 , x4 )T = (−50, −15, 70, 35)T,

(x̂1 , x̂2 , x̂3 , x̂4 )T = (−45, −10, 68, 34)T.

Simulation results are shown in Figs. 5.6 and 5.7. When t = 150 sec, the con-
troller is turned on. From the two figures we find that while the two controlled states
go to synchronization with the respective states of system (5.27), so do the other
two states. This fact can be explained as follows: since the controllers are added on
both the second and the third equations of system (5.28), from the forth equation of
the system we can see that if x̂4 tends to x4 it must imply that x̂3 tends to x3 . With
this in mind and from the first equation of system (5.28) we see that if x̂1 tends to x1
then x̂2 must tend to x2 .
150 5 Synchronizing Chaotic Systems Based on Feedback Control

5.4 Synchronization of Different Continuous-Time Chaotic

Systems

In the last section we assumed that the synchronized chaotic systems have the iden-
tical structure and the same parameters, but in practice the structures of the drive
system and the response system may be different and effects of noise (such as
parameters’ shift and measurement errors, etc.) should not be ignored. So, in this
section, we consider the problem of synchronizing chaotic systems with different
structures and parameter perturbations.

5.4.1 Synchronization of Different Chaotic Systems with

Parameter Perturbations

Suppose that the drive system has the following form:

ẋ = f (x) + F(x)α , (5.30)

where x ∈ U1 ⊂ Rn is the state, α ∈ Rm represents the known parameters, f (x) ∈ Rn ,

and F(x) ∈ Rn×m . The controlled response system is as follows:

x̃˙ = g(x̃) + G(x̃)β + u, (5.31)

where x̃ ∈ U2 ⊂ Rn is the state, β ∈ Rq represents the known parameters, g(x̃) ∈ Rn ,

G(x̃) ∈ Rn×q , and u ∈ Rn is the controller. Denote e = x̃ − x. From (5.30) and (5.31)
we get the error system

ė = g(x̃) + G(x̃)β − f (x) − F(x)α + u. (5.32)

If lim e = lim x̃(t, x̃0 ) − x(t, x0 ) = 0 we say that system (5.30) and system
t→∞ t→∞
(5.31) are synchronized. In the rest of this section we suppose that each element of
F(x) and G(x̃) is in L∞ . This is a reasonable assumption since the systems consid-
ered here are chaotic.
Theorem 5.7 ([17]). With the controller

u = f (x) + F(x)α̃ − g(x̃) − G(x̃)β̃ − ke (5.33)

and parameter update laws

α̃˙ = − δ F T (x)e,
(5.34)
β̃˙ =γ GT (x̃)e,
system (5.30) and system (5.31) can be synchronized globally and asymptotically.
Here, k > 0, δ > 0, and γ > 0 are constants and α̃ and β̃ are estimates of α and β .
Proof. From (5.32) and (5.33), we have
5.4 Synchronization of Different Continuous-Time Chaotic Systems 151

ė = F(x)(α̃ − α ) − G(x̃)(β̃ − β ) − ke. (5.35)

Let α̂ = α̃ − α and β̂ = β̃ − β . A Lyapunov function is chosen as

 
1 T 1 T 1 T
V (e, α̂ , β̂ ) = e e + α̂ α̂ + β̂ β̂ ,
2 δ γ

and its derivative along system (5.35) is

1 1 ˙
V̇ (e, α̂ , β̂ ) = eT ė + α̂ T α̂˙ + β̂ T β̂
δ γ
= [F(x)α̂ − G(x̃)β̂ − ke]Te − α̂ T F T (x)e + β̂ T GT (x̃)e
= −keTe
≤ 0.

Since e, α̂ , β̂ ∈ L∞ , from V̇ ≤ −keTe we have

 t
1 V (0)
e 2dt = (V (0) − V(t)) ≤ ,
0 k k
which means that e ∈ L2 . Because we have assumed that the elements of F(x) and
G(x̃) belong to L∞ , from (5.35) we have ė ∈ L∞ . Therefore, according to the Barbalat
lemma [10] we know that lim e = 0; that is, lim e = lim x̃(t, x̃0 ) − x(t, x0 ) = 0.
t→∞ t→∞ t→∞
The proof is completed. 

Remark 5.5. Many chaotic systems have the form of (5.30), such as Lorenz, Chen,
Lü, Rössler, Chua, Van der Pol, and Duffing, etc. 

5.4.2 Simulations

In this section we will show the effectiveness of the presented method by simulation.
Suppose that the drive system is the Lorenz system
⎧
⎨ ẋ1 = α1 (x2 − x1),
⎪
ẋ2 = α2 x1 − x1 x3 − x2 ,
⎪
⎩
ẋ3 = x1 x2 − α x3 .

Take the controlled Lü system as the response system

⎧
˙
⎨ x̃1 = β1 (x̃2 − x̃1) + u1 ,
⎪
x̃˙2 = −x̃1 x̃3 + β2 x̃2 + u2,
⎪
⎩˙
x̃3 = x̃1 x̃2 − β3x̃3 + u3 .
152 5 Synchronizing Chaotic Systems Based on Feedback Control

In this case x = (x1 , x2 , x3 )T , α = (α1 , α2 , α3 )T , x̃ = (x̃1 , x̃2 , x̃3 )T , β = (β1 , β2 , β3 )T ,

u = (u1 , u2 , u3 )T ,
⎛ ⎞ ⎛ ⎞
0 x2 − x1 0 0
f (x) = ⎝ −x1 x3 − x2 ⎠ , F(x) = ⎝ 0 x1 0 ⎠ ,
x1 x2 0 0 −x3
⎛ ⎞ ⎛ ⎞
0 x̃2 − x̃1 0 0
g(x̃) = ⎝ −x̃1 x̃3 ⎠ , and G(x̃) = ⎝ 0 x̃2 0 ⎠ .
x̃1 x̃2 0 0 −x̃3
The controller and parameter update laws are
⎧ ⎛ ⎞
⎪
⎪ α̃1 (x2 − x1 ) − β̃1(x̃2 − x̃1 ) − ke1
⎪
⎪
⎪
⎪ u = ⎝ α̃2 x1 − x1 x3 − x2 + x̃1 x̃3 − β̃2 x̃2 − ke2 ⎠ ,
⎨
x1 x2 − α̃3 x3 − x̃1 x̃2 + β̃3 x̃3 − ke3
⎪
⎪
⎪ α̃˙ = (δ (x1 − x2 )e1 , −δ x1 e2 , δ x3 e3 )T ,
⎪
⎪
⎪
⎩˙
β̃ = (γ (x̃2 − x̃1 )e1 , γ x̃2 e2 , −γ x̃3 e3 )T .

In the simulation, ‘unknown’ parameters are chosen as

α = (10, 28, 8/3)T and β = (36, 20, 3)T.

20

0
x̃1 − x1

−20

−40
0 10 20 30 40 50 60 70 80 90 100
t (sec)
20

0
x̃2 − x2

−20

−40
0 10 20 30 40 50 60 70 80 90 100
t (sec)
10

5
x̃3 − x3

−5

−10
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 5.8 The error curves of the synchronized Lorenz system and Lü system when k = 1 and
δ =γ =1
5.4 Synchronization of Different Continuous-Time Chaotic Systems 153

20

10
α̃1
0

−10

−20
0 10 20 30 40 50 60 70 80 90 100
t (sec)
60

40
α̃2

20

0
0 10 20 30 40 50 60 70 80 90 100
t (sec)
10

5
α̃3

−5
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 5.9 The α̃ curves when k = 1 and δ = γ = 1

15

10
β̃1

0
0 10 20 30 40 50 60 70 80 90 100
t (sec)
60

40
β̃2

20

0
0 10 20 30 40 50 60 70 80 90 100
t (sec)
10

5
β̃3

−5
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 5.10 The β̃ curves when k = 1 and δ = γ = 1

154 5 Synchronizing Chaotic Systems Based on Feedback Control

20

10
x̃1 − x1
0

−10
0 10 20 30 40 50 60 70 80 90 100
t (sec)
20

10
x̃2 − x2

−10

−20
0 10 20 30 40 50 60 70 80 90 100
t (sec)
10

5
x̃3 − x3

−5

−10
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 5.11 The error curves of the synchronized Lorenz system and Lü system when k = 1 and
δ =γ =4

20

10
α̃1

−10

−20
0 10 20 30 40 50 60 70 80 90 100
t (sec)
60

40
α̃2

20

0
0 10 20 30 40 50 60 70 80 90 100
t (sec)
10

0
α̃3

−10

−20
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 5.12 The α̃ curves when k = 1 and δ = γ = 4

5.4 Synchronization of Different Continuous-Time Chaotic Systems 155

20

β̃1 10

−10
0 10 20 30 40 50 60 70 80 90 100
t (sec)
50

0
β̃2

−50
0 10 20 30 40 50 60 70 80 90 100
t (sec)
10

0
β̃3

−10

−20
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 5.13 The β̃ curves when k = 1 and δ = γ = 4

20

10
x̃1 − x1

−10

−20
0 10 20 30 40 50 60 70 80 90 100
t (sec)
50
x̃2 − x2

−50
0 10 20 30 40 50 60 70 80 90 100
t (sec)
10

5
x̃3 − x3

−5

−10
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 5.14 The error curves of the synchronized Lorenz system and Lü system when k = 3 and
δ =γ =1
156 5 Synchronizing Chaotic Systems Based on Feedback Control

20

10
α̃1 0

−10

−20
0 10 20 30 40 50 60 70 80 90 100
t (sec)
60

40
α̃2

20

0
0 10 20 30 40 50 60 70 80 90 100
t (sec)
5
α̃3

−5
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 5.15 The α̃ curves when k = 3 and δ = γ = 1

20

15

10
β̃1

0
0 10 20 30 40 50 60 70 80 90 100
t (sec)
40

30

20
β̃2

10

0
0 10 20 30 40 50 60 70 80 90 100
t (sec)
10

5
β̃3

−5
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 5.16 The β̃ curves when k = 3 and δ = γ = 1

5.5 Synchronization of Discrete-Time Chaotic Systems 157

The initial values of the systems’ states are

x(0) = (10, 10, 10)T and x̃(0) = (2, 2, 2)T .

Parameters’ initial values are α̃ (0) = β̃ (0) = (3, 3, 3)T , k = 1, and δ = γ = 1. Simu-
lation results are shown in Figs. 5.8–5.16. From Figs. 5.11–5.13, it is easy to see that
increasing δ and γ does not shorten the transient time. On the contrary, the ampli-
tude of the oscillation becomes bigger. However, from Figs. 5.14–5.16, we find that
increasing k can shorten the transient time and decrease the amplitude of oscillation.

5.5 Synchronization of Discrete-Time Chaotic Systems

With the popularization of personal computers more and more people have real-
ized the importance of data security. Compared to continuous-time chaotic systems,
discrete-time chaotic systems are more suitable for digital computers in applications
of secure communications. Two reasons can be used to explain why discrete-time
chaotic systems have more advantages than continuous-time chaotic systems. First,
chaos can be found in very simple discrete-time systems even if the dimension of
the dynamical equation is only one. But, in a continuous-time autonomous system,
the dimension of the phase space must be greater than two in order to see chaos.
Second, a discrete-time chaotic system can be directly utilized using a computer
program while a continuous-time chaotic system needs to be discretized before it
is used. In the process of discretization one must pay careful attention to the cho-
sen method since false discretization may cause deviation dynamics between the
discretized system and the original continuous-time system.
Although there are already some works on the control and synchronization of
discrete-time chaotic systems [4, 5, 6, 12], most of them assume that parameters
of the drive and the response systems are identical. In this section we first propose
a control method to synchronize two chaotic systems with the same parameters,
after that an adaptive algorithm is designed to tune the controller’s parameters when
parameter disturbance exists. To show the effectiveness of the proposed methods, a
computer simulation is provided in the last part of the section.

5.5.1 Basic Formulation

Consider the following SISO discrete-time nonlinear control systems:

x(k + 1) = f (x(k), u(k)),

(5.36)
y(k) = h(x(k)),
158 5 Synchronizing Chaotic Systems Based on Feedback Control

where x ∈ Rn , u ∈ R, y ∈ R, f : U → U and h : U → L are smooth mappings satis-

fying f (0, 0) = 0 and h(0) = 0, and U and L are subsets of Rn and R, respectively.
Denote h ◦ f = h( f (x, u)) and f s = f ◦ f ◦ · · · ◦ f , i.e.,
0 12 3
s times

f s = f ( f (· · · f (x, us ) · · · , u2 ), u1 ).

Definition 5.6. If there exists a minimal positive integer r such that for any 0 ≤ i ≤
r − 1 the following conditions are satisfied:
∂ (h ◦ f i )
(i) = 0,
∂u
∂ (h ◦ f r )
(ii) = 0,
∂u
then we say that system (5.36) has relative degree r. 

Remark 5.6. System (5.36) having relative degree r means that inputs at k = 0 do
not affect the system’s output until the instant k = r. 

Theorem 5.8 ([11]). If system (5.36) has relative degree r = n, then there is a coor-
dinate transformation z = Φ (x) such that, in the new coordinates, system (5.36) has
the following form:
z(k + 1) = Az(k) + Bv(k),
y(k) = Cz(k),
where

z = (z1 , . . . , zn )T
= (h(x), h ◦ f (x, u), . . . , h ◦ f n−1(x, u))T , (5.37)
n
v = h◦ f (x, u), (5.38)
⎛ ⎞
0 1 0 ··· 0 0
⎜0 0 1 ··· 0 0⎟
⎜ ⎟
⎜ .. ⎟
A = ⎜ ... ..
.
.. . .
. .
..
. .⎟ (5.39)
⎜ ⎟
⎝0 0 0 ··· 0 1⎠
0 0 0 ··· 0 0 n×n,
B = (0, . . . , 0, 1)Tn×1 ,
C = (1, 0, . . . , 0)1×n. (5.40)


5.5 Synchronization of Discrete-Time Chaotic Systems 159

5.5.2 Main Results

5.5.2.1 Synchronization of Discrete-Time Chaotic Systems with the Same

Parameters

Before stating the main results we need the following lemma.

Lemma 5.1. If system (5.36) has relative degree n and is of the following form:

x(k + 1) = f (x(k), u(k)) = F(x(k)) + gu(k),

(5.41)
y(k) = h(x(k)) = d T x(k),

then it can be transformed to

z(k + 1) = Az(k) + B(a(z(k)) + bu(k)),

(5.42)
y(k) = Cz(k),

where g, d are constant vectors, A, B,C are defined as those in (5.39) and (5.40),
a(·) : Rn → R is a continuous function, and b is a nonzero constant.

Proof. Define mappings ψ = F(x) and φu = x + gu. Then, f (x, u) = φu ◦ ψ (x). Since
system (5.41) has relative degree n, we have

∂ ∂ ∂h ∂ (φu ◦ ψ )i
(h ◦ f i ) = (h ◦ (φu ◦ ψ )i ) = = 0, 0 ≤ i ≤ n − 1,
∂u ∂u ∂ (φu ◦ ψ )i ∂u
∂ ∂ ∂h ∂ (φu ◦ ψ )n
(h ◦ f n) = (h ◦ (φu ◦ ψ )n) = = 0,
∂u ∂u ∂ (φu ◦ ψ )n ∂u

which means that f n = F n + gu. Noticing that h(x) = d T x, we have h ◦ f n = h ◦ F n +

h(gu) = d T F +d T gu. Let a(x) = d T F(x) and b = d T g. We have h◦ f n (x) = a(x)+bu.
Thus, the proof is completed. 

Suppose that the drive system is as follows:

x(k + 1) = f (x(k), 0),

(5.43)
y(k) = h(x(k)),

where x ∈ Rn , y ∈ R is the output, and f : Rn × R → Rn and h : Rn → R are smooth

mappings satisfying f (0, 0) = 0 and h(0) = 0. Construct the following controlled
response system:
x̂(k + 1) = f (x̂(k), u),
(5.44)
ŷ(k) = h(x̂(k)),
where u is a controller. If
lim ŷ(k) − y(k) = 0,
k→∞
160 5 Synchronizing Chaotic Systems Based on Feedback Control

then system (5.43) and system (5.44) are said to be in synchronization. Before de-
signing the controller, we assume that the following two conditions hold for system
(5.44):
(i) system (5.44) has relative degree n in the neighborhood of zero Bρ (0);
(ii) h ◦ f n (x, u) = a(x) + bu.
For some mechanical systems [12], the first condition can be satisfied. At least
for affine control systems the second condition can also be satisfied.
Considering the above assumptions and Lemma 5.1, we know that through the
nonlinear coordinate transformation

ẑ = Φ (x̂) = (h(x̂), h ◦ f (x̂, u), . . . , h ◦ f n−1(x̂, u))T

system (5.44) can be transformed into

ẑ(k + 1) = Aẑ(k) + B(a(ẑ(k)) + bu(k)),

ŷ(k) = Cẑ(k).

With the same transformation, system (5.43) becomes

z(k + 1) = Az(k) + Ba(z(k)),

y(k) = Cz(k).

Set e = ẑ − z. Then, we get the error system as

e(k + 1) = Ae(k) + B(a(ẑ) − a(z) + bu). (5.45)

Theorem 5.9 ([16]). With the controller

u = −b−1 (a(ẑ) − a(z) + Qe), (5.46)

the origin of system (5.45) is asymptotically stable, where Q = (q0 , q1 , . . . , qn )T are

coefficients of a polynomial P(s) = sn + qn−1sn−1 + · · · + q1 s + q0 such that the
absolute values of all roots of P(s) are less than 1.

Proof. Substituting controller (5.46) into system (5.45), we get

e(k + 1) = Λ e(k),

where ⎛ ⎞
0 0 ··· 0
1 0
⎜ 0 1 ··· 0
0 0 ⎟
⎜ ⎟
⎜ ... ..
.. . . . . .. ⎟ .
Λ =⎜ . . .. . ⎟
⎜ ⎟
⎝ 0 0 0 ··· 0 1 ⎠
−q0 −q1 −q2 · · · −qn−2 −qn−1
5.5 Synchronization of Discrete-Time Chaotic Systems 161

By the selection of Q we know that all eigenvalues of Λ have absolute values less
than 1. The proof is completed. 

5.5.2.2 Synchronization of Discrete-Time Chaotic Systems with Parameter

Perturbation

In this section we will study how to synchronize two discrete-time chaotic systems
whose parameters are not identical. Before stating the main results we need two
lemmas.
For a general discrete-time system

y(k) = φ T (k − 1)β0, (5.47)

where y(k) ∈ R and φ (k − 1) is a linear or nonlinear function of output Y = (y(k −

1), . . . , y(k − p))T and input U = (u(k − 1), . . . , u(k − m))T . ŷ(k), the estimate of y(k),
can be obtained by
ŷ(k) = φ T (k − 1)β (k), (5.48)
where β (k) is the estimate of β0 .
Lemma 5.2 ([7]). Calculate β (k) according to the following algorithm:
⎧
⎪ P(k − 2)φ (k − 1)
⎪
⎪ β (k) = β (k − 1) + T
⎪ [y(k) − φ T (k − 1)β (k − 1)],
⎪
⎪ φ (k − 1)P(k − 2)φ (k − 1)
⎪
⎪
⎪
⎪ if φ T (k − 1)P(k − 2)φ (k − 1) = 0,
⎪
⎪
⎪
⎪
⎪ T
⎨ β (k) = β (k − 1), if φ (k − 1)P(k − 2)φ (k − 1) = 0,
⎪
P(k − 2)φ (k − 1)φ T(k − 1)P(k − 2)
⎪
⎪ P(k − 1) = P(k − 2) − ,
⎪
⎪ φ T (k − 1)P(k − 2)φ (k − 1)
⎪
⎪
⎪
⎪ if φ T (k − 1)P(k − 2)φ (k − 1) = 0,
⎪
⎪
⎪
⎪
⎪
⎪ P(k − 1) = P(k − 2), if φ T (k − 1)P(k − 2)φ (k − 1) = 0,
⎪
⎪
⎩
P(0) = I.

If the rank of (φ (1), . . . , φ (s)) equals the dimension of β0 , then β (k) will converge
to β0 within s steps. 
Setting G ⊂ R, we state the discrete version of the LaSalle invariance principle.
Lemma 5.3 ([13]). For a discrete-time system x(k + 1) = f (x(k)), where f : Rn →
Rn is a smooth mapping, if the following two conditions are satisfied:
(i) there is a function V ∈ C[G ⊂ Rn , R] such that along the solutions of the system,
we have
Δ V (x(k)) = V (x(k + 1)) − V(x(k)) ≤ 0;
(ii) there exists a bounded solution x(k) such that x(k) ⊂ G for all k ≥ 0,
162 5 Synchronizing Chaotic Systems Based on Feedback Control

then there is a constant c such that x(k) → M ∩V −1 (c)when k → ∞, where M denotes

the maximal invariance set in E = {x : Δ V (x) = 0, x ∈ G}, V −1 (c) = {x : V (x) =
c, x ∈ Rm }, and G is the closure of G. 
In the following we will design an adaptive controller to synchronize two
discrete-time chaotic systems when parameter perturbation exists. Suppose that the
drive system has the following form:

x(k + 1) = f T (x(k))θ0 ,
(5.49)
y(k) = Dx(k),

where x ∈ Rn , f : Rn → Rq×n is a smooth mapping, D ∈ R1×n , and θ0 ∈ Rq is the

unknown parameter matrix. We construct the following controlled response system:

x̂(k + 1) = f T (x̂(k))θ (k) + gu,

(5.50)
ŷ(k) = Dx̂(k),

where g ∈ Rn , u ∈ R, and θ (k) is the estimate of θ0 . Suppose that system (5.50)

has relative degree n in U ⊂ Rn . Then, the input–output form of system (5.43) is as
follows:
y(k) = D f T (x(k − 1))θ0 . (5.51)
We construct an auxiliary system

ỹ(k) = D f T (x(k − 1))θ (k − 1).

Since system (5.50) has relative degree n in U, we can find a local transformation
ẑ = Φ (x) such that, in the new coordinates, system (5.50) becomes

ẑ(k + 1) = Aẑ(k) + B(a(ẑ(k), θ (k)) + bu(k)),

ŷ(k) = Cẑ(k).

With the same transformation, system (5.49) becomes

z(k + 1) = Az(k) + Ba(z(k), θ0 ),

y(k) = Cz(k),

where A, B, and C have the same form as in (5.39) and (5.40). Setting e(k) = ẑ(k) −
z(k), we get the error system as

e(k + 1) = Ae(k) + B(a(ẑ(k), θ (k)) − a(z(k), θ0 ) + bu). (5.52)

Theorem 5.10 ([16]). With the following controller:

u = −b−1(a(ẑ(k), θ (k)) − a(z(k), θ (k)) + ce) (5.53)

and parameter update laws

5.5 Synchronization of Discrete-Time Chaotic Systems 163
⎧
⎪ P(k − 2)D f T (k − 1)[y(k) − D f T(k − 1)θ (k − 1)]
⎪
⎪ θ (k) = θ (k − 1) + ,
⎪
⎪ D f T (k − 1)P(k − 2) f (k − 1)DT
⎪
⎪
⎪
⎪
⎪
⎪ if D f T (k − 1)P(k − 2) f (k − 1)DT = 0,
⎪
⎪
⎪
⎪ θ (k) = θ (k − 1), if D f T (k − 1)P(k − 2) f (k − 1)DT = 0,
⎪
⎨
P(k − 2) f (k − 1)DTD f T (k − 1)P(k − 2)
⎪
⎪ P(k − 1) = P(k − 2) − ,
⎪
⎪ D f T (k − 1)P(k − 2) f (k − 1)DT
⎪
⎪
⎪
⎪
⎪
⎪ if D f T (k − 1)P(k − 2) f (k − 1)DT = 0,
⎪
⎪
⎪ P(k − 1) = P(k − 2), if D f T (k − 1)P(k − 2) f (k − 1)DT = 0,
⎪
⎪
⎪
⎩
P(0) = I,

the origin of system (5.52) is asymptotically stable in U ∈ Rn , where Q = (q0 , . . . ,

qn−1 )T are coefficients of a polynomial P(s) = sn + qn−1sn−1 + · · · + q1 s + q0 whose
roots have absolute values less than 1.
Proof. Substituting (5.53) into (5.52), we have

e(k + 1) = Ae(k) + B(a(ẑ(k), θ (k)) − a(z(k), θ0 )

− a(ẑ(k), θ (k)) + a(z(k), θ (k)) − Qe)
= Λ e(k) + B(a(z(k), θ (k)) − a(z(k), θ0 )), (5.54)

where ⎛ ⎞
0 0 ··· 0
1 0
⎜ 0 1 ··· 0
0 0 ⎟
⎜ ⎟
⎜ .. ..
.. . . .. .. ⎟
Λ =⎜ . . . . . . ⎟
⎜ ⎟
⎝ 0 0 0 ··· 0 1 ⎠
−q0 −q1 −q2 · · · qn−2 −qn−1 n×n.
For system (5.54) we choose the following Lyapunov function:
n
V (e(k)) = ∑ |ei (k)|,
i=1

where ei (k) is the ith element of e. Thus, we have

Δ V = V (e(k + 1)) − V(e(k)) = −|e1 (k)| + |a(z(k), θ (k)) − a(z(k), θ0 )|.

Since system (5.49) is a chaotic system, there must exist a large N > 0 such that
the condition of Lemma 5.2 is satisfied. Therefore, θ (k) will equal θ0 at most in N
steps. Thus, we have
Δ V |k>N = −|e1 (k)| ≤ 0.
When k > N the error system becomes e(k + 1) = Λ e(k). If Δ V |k>N = 0, it means
that e1 (k) = 0 (k > N). From the error system equations we have ei (k) = 0, 2 ≤ i ≤ n.
By Lemma 5.3 we know that system (5.52) is asymptotically stable in U. The proof
is completed. 
164 5 Synchronizing Chaotic Systems Based on Feedback Control

5.5.3 Simulations

In this section we take the Hénon system [8] as an example to illustrate the proposed
methods. The difference equations of the Hénon system are as follows:

x1 (k + 1) =1 − p0x21 (k) + x2 (k),

(5.55)
x2 (k + 1) =q0 x1 (k).

When p = 1.4 and q = 0.3, the Hénon system is in chaotic state and its attractor is
shown in Fig. 5.17. The system has two fixed points:
'   (T
(q0 − 1)2 + 4p0
q0 − 1 + q0 − 1 + (q0 − 1)2 + 4p0
X f1 = , q0 ,
2p0 2p0
'   (T
q0 − 1 − (q0 − 1)2 + 4p0 q0 − 1 + (q0 − 1)2 + 4p0
X f2 = , q0 .
2p0 2p0

0.4

0.3
Xf1

0.2

0.1

x2 0

−0.1

−0.2 Xf2

−0.3

−0.4
−1.5 −1 −0.5 0 0.5 1 1.5
x1

Fig. 5.17 The chaotic attractor of Hénon system (5.55)

5.5 Synchronization of Discrete-Time Chaotic Systems 165

From Fig. 5.17 we can see that X f1 is in the attractor. To apply the proposed
method we should first displace the origin of system (5.55) to X f1 . For simplicity
we will use the same symbols to denote the transformed system. Using the new
coordinates, the Hénon system is as follows:

x1 (k + 1) = −p0 x21 (k) − 2X f1 p0 x1 (k) + x2 (k),

x2 (k + 1) = q0 x1 (k).

Let x1 be the output. Then, the drive system has the following form:
⎧ 2
⎨ x1 (k + 1) = −p0 x1 (k) − 2X f1 p0 x1 (k) + x2 (k),
⎪
x2 (k + 1) = q0 x1 (k), (5.56)
⎪
⎩
y(k) = x1 (k).

Similarly, for the controlled response system

⎧ 2
⎨ x̂1 (k + 1) = 1 − p(k)x̂1 (k) + x̂2 (k),
⎪
x̂2 (k + 1) = q(k)x̂1 (k) + u,
⎪
⎩
ŷ(k) = x̂1 (k)

we can transform it into the following form:

⎧ 2
⎨ x̂1 (k + 1) = −p(k)x̂1 (k) − 2X f1 p(k)x̂1 (k) + x̂2 (k),
⎪
x̂2 (k + 1) = q(k)x̂1 (k) + u, (5.57)
⎪
⎩
ŷ(k) = x̂1 (k).

Notice that system (5.55) can be transformed into the input–output form as

y1 (k + 1) = 1 − p0x21 (k) + q0 x1 (k − 1) = φ T (k)θ0 .

Thus, we can construct an auxiliary system

ỹ(k + 1) = 1 − p(k)x21(k) + q(k)x1 (k − 1) = φ T (k)θ (k).

So, parameter update laws (5.34) can be used. It is easy to validate that system (5.57)
has relative degree 2. By the following coordinate transformation:

Φ (x̂) = (ẑ1 , ẑ2 ) = (h(x̂), h ◦ f (x̂))T = (x̂, −p(k)x̂21 (k) − 2X f1 p(k)x̂1 (k) + x̂2 (k))T ,

system (5.57) is changed to

⎧
⎨ ẑ1 (k + 1) = ẑ2 (k),
⎪
ẑ2 (k + 1) = −p(k)z22 (k) − 2X f1 (k)p(k)ẑ2 (k) + q(k)ẑ1 (k) + u,
⎪
⎩
ŷ(k) = ẑ1 (k).
166 5 Synchronizing Chaotic Systems Based on Feedback Control

With the same transformation, system (5.56) is changed to

⎧
⎨ z1 (k + 1) = z2 (k),
⎪
z2 (k + 1) = −p0 z22 (k) − 2X f1 p0 z2 (k) + q0z1 (k),
⎪
⎩
y(k + 1) = z1 (k).

Setting e1 (k) = ẑ1 (k) − z1 (k) and e2 (k) = ẑ2 (k) − z2 (k), we get the following error
system:
⎧
⎨ e1 (k + 1) = e1 (k),
e (k + 1) = −p(k)z22 (k) − 2X f1 (k)p(k)ẑ2 (k) + q(k)ẑ1 (k)
⎩ 2
+p0 z22 (k) + 2X f1 p0 z2 (k) − q0 z1 (k) + u.

The controller is designed according to (5.53) with Q = (0.25, 1)T. Choose p0 = 1.4,
q0 = 0.3, p(0) = 1.4, and q(0) = 0.3. Initial values of the drive system and the
response system are (0.3, 0.1)T and (1.3, 0.5)T. Simulation results are shown in Figs.
5.18 and 5.19. Before k = 1000 the two systems have the same parameters. Within
0 ≤ k < 500 no controller is added on the response system. When 500 ≤ k < 1000
the controller is turned on and the two Hénon systems become synchronized very

2
x̂1 − x1

−2

−4
0 200 400 600 800 1000 1200 1400 1600 1800 2000
k

0
x̂2 − x2

−1

−2

−3
0 200 400 600 800 1000 1200 1400 1600 1800 2000
k

Fig. 5.18 The errors of synchronized Hénon systems

5.6 Summary 167

1.4

1.35

1.3
p(k)

1.25

1.2

1.15
1500 1550 1600 1650 1700 1750 1800 1850 1900 1950 2000
k

0.4

0.35

0.3
q(k)

0.25

0.2

1500 1550 1600 1650 1700 1750 1800 1850 1900 1950 2000
k

Fig. 5.19 Parameter curves of p(k) and q(k)

quickly. When k = 1000 the parameters of the response system are disturbed to
1.2 and 0.2, and the previous controller cannot synchronize the Hénon systems any
more. At k = 1500, parameter update laws are added and the two Hénon systems
are synchronized again. From Fig. 5.19 we can see that parameters of the response
system converge to those of the drive system when synchronization is achieved.

5.6 Summary

We have studied the problems of synchronization of both continuous-time and

discrete-time chaotic systems in this chapter. The methods developed in Sects. 5.2,
5.3, and 5.5 are mainly based on geometric control theory of nonlinear systems. In
these methods there are some common steps. First, outputs should be determined
such that the response system has a relative degree as large as possible. After that
the original system is transformed to normal form so that the controller can easily
be designed. The controller designed in Sect. 5.4 is much more complicated than
those in other sections. The reason is that the problem considered in this section is
also more difficult than those in the other sections. This controller aims for synchro-
168 5 Synchronizing Chaotic Systems Based on Feedback Control

nizing general chaotic systems. We do not exclude the special cases in which the
controller can be simplified.

References

1. Boccaletti S, Kurths J, Osipov G, Valladares DL, Zhou CS (2002) The synchronization of

chaotic systems. Phys Rep 366:1–101
2. Pecora LM, Carroll TL (1990) Synchronization in chaotic systems. Phys Rev Lett 64:821–824
3. Pecora LM, Carroll TL, Johnson GA, Mar DJ (1997) Fundamentals of synchronization in
chaotic systems, concepts, and applications. Chaos 7:520–543
4. Chen G, Dong X (1998) From Chaos to Order: Methodoligies, Perspectives and Applications.
World Scientific, Singapore
5. Cruz C, Nijmeijer H (2000) Synchronization through filtering. Int J Bifurc Chaos 10:763–775
6. Fuh CC, Tsai HH (2002) Control of discrete-time chaotic systems with feedback linearization.
Chaos Solitons Fractals 13:285–294
7. Goodwin GC, Sin KS (1984) Adaptive Filtering, Prediction and Control. Prentice Hall, New
Jersey
8. Hénon M (1976) A two-dimensional mapping with a strange attractor. Commun Math Phys
50:69–77
9. Isidori A (1989) Nonlinear Control Systems: An Introduction, 2nd edn. Springer, New York
10. Khalil HK (2002) Nonlinear Systems, 3rd edn. Prentice Hall, New Jersey
11. Lee HG, Arapostathis A, Marcus SI (1987) Linearization of discrete-time systems. Int J Con-
trol 45:1803–1822
12. Liao TL, Huang NG (1997) Control and synchronization of discrete-time chaotic systems via
variable structure control technique. Phys Lett A 234:262–268
13. Liao XX (2000) Theory and Applications of Stability for Dynamical Systems. National De-
fence Industry Press, Beijing
14. Rössler OE (1979) An equation for hyperchaos. Phys Lett A 71:155–157
15. Wang ZL, Zhang HG (2002) Synchronization of high dimensional chaotic systems based on
nonlinear feedback control. In: Proc 15th IFAC Conf, Barcelona, Spain, pp. 347–351
16. Wang ZL, Zhang HG (2002) Adaptive synchronization of discrete chaotic systems. In: Proc
1st Int Conf Machine Learning and Cybernetics, Beijing, China, pp. 336–339
17. Zhang HG, Huang W, Wang ZL, Chai T (2006) Adaptive synchronization between two dif-
ferent chaotic systems with unknown parameters. Phys Lett A 350:363–366
Chapter 6
Synchronizing Chaotic Systems via Impulsive
Control

Abstract Impulsive control is an efficient method to deal with dynamical systems

which cannot be controlled by continuous control. In addition, in the synchroniza-
tion process, the response system receives information from the drive system only
at discrete time instants, which drastically reduces the amount of synchronization
information transmitted from the drive system to the response system and makes
this method more efficient in a great number of real-life applications. In this chap-
ter, we will study how to synchronize two identical or different chaotic systems by
impulsive control methods. We first study the complete synchronization of a class
of chaotic systems, and after that we develop synchronization methods for unified
systems with channel time delay in the sense of practical stability. Then, robust syn-
chronization schemes are studied for chaotic systems with parametric uncertainty
and parametric mismatch. The aim is to present some practical impulsive control
methods for different synchronization schemes.

6.1 Introduction

In this chapter, we will study how to synchronize two identical or different chaotic
systems by impulsive control methods. Impulsive control is an efficient method to
deal with dynamical systems which cannot be controlled by continuous control [18].
In addition, in the synchronization process, the response system receives informa-
tion from the drive system only at discrete time instants, which drastically reduces
the amount of synchronization information transmitted from the drive system to the
response system and makes this method more efficient in a great number of real-life
applications.
As is well known, for many applications, parametric uncertainties of systems
are inevitable, and the effects of these uncertainties will destroy the synchroniza-
tion. In the past few years, the analysis of synchronization for chaotic systems with
parametric uncertainties has gained much research attention [4, 8, 11]. Most of the
researches only concern the synchronization between two identical chaotic systems

169
170 6 Synchronizing Chaotic Systems via Impulsive Control

with known parameters or identical unknown parameters. However, in practice, it is

difficult to ensure that the structures of the drive and the response systems are iden-
tical. Moreover, the parametric uncertainties of the drive and response systems are
always different and time varying. Therefore, it is significant to investigate synchro-
nization between two different chaotic systems in the presence of different time-
varying parametric uncertainties. Furthermore, most of the existing literature has
only considered chaotic systems described by impulsive ordinary differential equa-
tions (ODEs) [11, 20]. Due to the difficulty in investigating the stability of impulsive
delayed or functional differential equations (DDEs or FDEs), there is little literature
dealing with the synchronization problems of delayed chaotic systems via impulsive
control [7, 8].
In this chapter we first study the complete synchronization of a class of chaotic
systems, and after that we develop synchronization methods for unified systems with
channel time delay in the sense of practical stability. Then, robust synchronization
schemes are studied for chaotic systems with parametric uncertainties and paramet-
ric mismatch. Our goal is to present some practical impulsive control methods for
different synchronization schemes.

6.2 Complete Synchronization of a Class of Chaotic Systems via

Impulsive Control

6.2.1 System Description and the Synchronization Problem

Consider the drive system as follows:

ẋ(t) = Ax(t) + h(x(t)), (6.1)

where x ∈ Rn is the state vector, A ∈ Rn×n is a constant matrix, and h(x(t)) ∈ Rn is

a continuous nonlinear function. Assume that

h(x(t)) − h(x̃(t)) = N(x(t), x̃(t))(x(t) − x̃(t)), (6.2)

where N(x(t), x̃(t)) ∈ Rn×n is a bounded matrix with elements depending on x(t)
and x̃(t). Most chaotic systems can be described by (6.1) and (6.2), such as the
Lorenz system, the Rössler system, the Chen system, the Lü system, the unified
system, several variants of Chua’s circuit, etc.
Suppose that a set of discrete instants {tk } satisfies

0 < t1 < t2 < · · · < tk < tk+1 < · · · , lim tk = ∞ (k ∈ N), 0 ≤ t0 < t1 .
k→∞

At discrete time instant tk , the state variables of the drive system are transmitted
to the response system and the state variables of the response system are suddenly
changed at these instants. Therefore, the response system can be written in the fol-
6.2 Complete Synchronization of a Class of Chaotic Systems via Impulsive Control 171

lowing form:
⎧
˙
⎨ x̃(t) = Ax̃(t) + h(x̃(t)), t = tk ,
⎪
Δ x̃(tk ) = x̃(tk+ ) − x̃(tk− ) = x̃(tk+ ) − x̃(tk ) = −Bk e(tk ), t = tk , k ∈ N, (6.3)
⎪
⎩
x̃(t0+ ) = x̃0 ,

where

e(t) = x(t) − x̃(t) = (x1 (t) − x̃1(t), x2 (t) − x̃2(t), . . . , xn (t) − x̃n (t))T

is the synchronization error. Let x̃(tk+ ) = lim x̃(t), x̃(tk− ) = lim x̃(t), i.e., tk+ and
t→tk+ t→tk−
tk− denote the times after and before tk , respectively. x̃(tk− )
= x̃(tk ) implies that x̃(t)
is left-continuous for t ≥ t0 . Bk ∈ Rn×n are the impulsive control gains. Δ x(tk ) =
x(tk+ )−x(tk− ) = 0. Then, from (6.1) and (6.3), the following error system is obtained:
⎧
⎨ ė(t) = (A + N(x(t), x̃(t)))e(t), t = tk ,
⎪
Δ e|t=tk = e(tk+ ) − e(tk ) = Bk e(tk ), t = tk , k ∈ N, (6.4)
⎪
⎩ +
e(t0 ) = e0 .

Here, the objective is to find the conditions on the control gains Bk and the im-
pulsive distances δ¯k = tk − tk−1 (k ∈ N) such that the error system (6.4) is asymptot-
ically stable, which implies that the impulsively controlled response system (6.3) is
asymptotically synchronized with the drive system (6.1) for arbitrary initial condi-
tions.

6.2.2 Main Results

Theorem 6.1 ([20]). Let βk , λA , and λN be the largest eigenvalues of (I + Bk )T (I +

Bk ), A + AT , and N(x, x̃) + N T (x, x̃), respectively. If there exists a constant ρ > 1
such that
ln(ρβk ) + (λA + λN )δ¯k ≤ 0, k ∈ N, (6.5)
then the impulsively controlled response system (6.3) is asymptotically synchro-
nized with the drive system (6.1).

Proof. Let the Lyapunov function be in the form of

V (e) = eT e.

For t ∈ (tk−1 ,tk ], k ∈ N, its derivative along the solution of (6.4) is

172 6 Synchronizing Chaotic Systems via Impulsive Control

V̇ (e) = ėT e + eTė

= [(A + N(x, x̃))e]T e + eT[(A + N(x, x̃))e]
= eT (A + AT)e + eT[N(x, x̃) + N T (x, x̃)]e
≤ (λA + λN )V (e).

Therefore,

V (e(t)) ≤ V (e(tk−1
+
)) exp[(λA + λN )(t − tk−1 )], t ∈ (tk−1 ,tk ], k ∈ N. (6.6)

On the other hand, when t = tk , we have

V (e(tk+ )) = [(I + Bk )e(tk )]T (I + Bk )e(tk )

= eT (tk )(I + Bk )T (I + Bk )e(tk )
≤ βkV (e(tk )), k ∈ N. (6.7)

From inequalities (6.6) and (6.7), we know for any t ∈ (t0 ,t1 ],

V (e(t)) ≤ V (e(t0+ )) exp[(λA + λN )(t − t0 )],

and this leads to

V (e(t1 )) ≤ V (e(t0+ )) exp[(λA + λN )(t1 − t0 )]

and
V (e(t1+ )) ≤ β1V (e(t1 )) ≤ V (e(t0+ ))β1 exp[(λA + λN )(t1 − t0 )].
Thus, for any t ∈ (t1 ,t2 ],

V (e(t)) ≤ V (e(t1+ )) exp[(λA + λN )(t − t1 )] ≤ V (e(t0+ ))β1 exp[(λA + λN )(t − t0 )].

In general, for t ∈ (tk ,tk+1 ],

V (e(t)) ≤ V (e(t0+ ))β1 β2 · · · βk exp[(λA + λN )(t − t0 )]

= V (e0 )β1 exp[(λA + λN )δ¯1 ] · · · βk exp[(λA + λN )δ¯k ]
× exp[(λA + λN )(t − tk )]
1
≤ V (e0 ) k exp[(λA + λN )(t − tk )]
ρ
⎧
⎪ 1 ¯
⎨ V (e0 ) ρ k exp[(λA + λN )δk+1 ], for λA + λN > 0,
⎪
≤
⎪
⎪ 1
⎩ V (e0 ) , for λA + λN ≤ 0.
ρk

This means that V (e(t)) → 0 and e(t) → 0 as t → ∞. Thus, the origin of system (6.4)
is asymptotically stable, which implies that the impulsively controlled response sys-
6.2 Complete Synchronization of a Class of Chaotic Systems via Impulsive Control 173

tem (6.3) is asymptotically synchronized with the drive system (6.1). This completes
the proof. 
4 4
Based on the matrix theory, we can obtain λN ≤ 4N(x, x̃) + N T(x, x̃)4, where ·
refers to any induced matrix norm in this section. Then, we can derive the following
corollary.

Corollary 6.1. Let βk and λA be the largest eigenvalues of (I + Bk )T (I + Bk ) and

A + AT, respectively. If there exists a constant ρ > 1 such that
 4 4
ln(ρβk ) + λA + 4N(x, x̃) + N T (x, x̃)4 δ¯k ≤ 0, k ∈ N, (6.8)

then the impulsively controlled response system (6.3) is asymptotically synchro-

nized with the drive system (6.1). 

Remark 6.1. Since the trajectory of a chaotic system is bounded, inequalities (6.5)
and (6.8) hold for suitable values of βk and δ¯k (k ∈ N). 

Remark 6.2. Generally speaking, the estimate of a matrix norm is easier than that of
eigenvalues of a matrix. Therefore, it is more convenient to apply Corollary 6.1 than
Theorem 6.1 in practice. 

Remark 6.3. In Theorem 6.1 and Corollary 6.1, if δ¯k = δ¯ > 0 and Bk = B (k ∈ N),
then the conditions of Theorem 6.1 and Corollary 6.1 can be reduced to

ln(ρβ ) + (λA + λN )δ¯ ≤ 0

and  4 4
ln(ρβ ) + λA + 4N(x, x̃) + N T (x, x̃)4 δ¯ ≤ 0,
respectively, where β is the largest eigenvalue of (I + B)T(I + B). 

Theorem 6.2 ([12]). Let βk , λA , and λN be the largest eigenvalues of (I + Bk )T (I +

Bk ), A + AT, and N(x, x̃) + N T (x, x̃), respectively. If one of the following conditions
holds:
(i) λA + λN ≥ 0, and there exists a constant ρ > 1 such that

ln(ρβ2k−1 β2k ) + (λA + λN )(t2k+1 − t2k−1 ) ≤ 0 (6.9)

and
sup{βk exp[(λA + λN )(tk+1 − tk )]} = ϒ < ∞; (6.10)
k

(ii) λA + λN < 0, and there exists a constant ρ (0 ≤ ρ < −(λA + λN )) such that

ln(βk ) − ρ (tk − tk−1 ) ≤ 0, k ∈ N; (6.11)

then the impulsively controlled response system (6.3) is asymptotically synchro-

nized with the drive system (6.1).
174 6 Synchronizing Chaotic Systems via Impulsive Control

Proof. Choose the following Lyapunov function:

V (e) = eT e.

For t ∈ (tk−1 ,tk ] (k ∈ N), the time derivative of V (e) along the solution of (6.4) is

V̇ (e(t)) = ėT e + eTė

= [(A + N(x, x̃))e]T e + eT[(A + N(x, x̃))e]
= eT (A + AT)e + eT[N(x, x̃) + N T (x, x̃)]e
≤ (λA + λN )V (e(t)),

which implies that

V (e(t)) ≤ V (e(tk−1
+
)) exp[(λA + λN )(t − tk−1 )]. (6.12)

On the other hand, when t = tk , we obtain

V (e(tk+ )) = [(I + Bk )e(tk )]T (I + Bk )e(tk )

= eT (tk )(I + Bk )T (I + Bk )e(tk )
≤ βkV (e(tk )). (6.13)

For t ∈ (tk ,tk+1 ], k ∈ N, from (6.12) and (6.13), we get

V (e(t)) ≤ V (e(t0+ ))β1 β2 · · · βk exp[(λA + λN )(t − t0 )]. (6.14)

If λA + λN ≥ 0, there exists a constant ρ > 1. From (6.9), (6.10), and (6.14), when
(i) t ∈ (t2k−1 ,t2k ], we have
2k−1
V (e(t)) ≤ V (e(t0+ )) ∏ βi exp[(λA + λN )(t − t0)]
i=1
2k−1
≤ V (e0 ) ∏ βi exp[(λA + λN )(t2k − t0)]
i=1
= V (e0 )β1 β2 exp[(λA + λN )(t3 − t1 )] · · ·
× β2k−3β2k−2 exp[(λA + λN )(t2k−1 − t2k−3 )]
× β2k−1 exp[(λA + λN )(t2k − t2k−1)]
× exp[(λA + λN )(t1 − t0 )]
V (e0 )
≤ϒ exp[(λA + λN )(t1 − t0 )], (6.15)
ρ k−1

(ii) t ∈ (t2k ,t2k+1 ], we have

6.2 Complete Synchronization of a Class of Chaotic Systems via Impulsive Control 175

2k
V (e(t)) ≤ V (e(t0+ )) ∏ βi exp[(λA + λN )(t − t0 )]
i=1
2k
≤ V (e0 ) ∏ βi exp[(λA + λN )(t2k+1 − t0)]
i=1
= V (e0 )β1 β2 exp[(λA + λN )(t3 − t1 )] · · ·
× β2k−1 β2k exp[(λA + λN )(t2k+1 − t2k−1 )]
× exp[(λA + λN )(t1 − t0 )]
V (e0 )
≤ exp[(λA + λN )(t1 − t0 )]. (6.16)
ρk

If λA + λN < 0, and there exists a constant ρ (0 ≤ ρ < −(λA + λN )) such that

(6.11) holds, i.e., βk ≤ exp[ρ (tk − tk−1 )], then

V (e(t)) ≤ V (e(t0+ ))β1 β2 · · · βk exp[(λA + λN )(t − t0 )]

≤ V (e0 ) exp[ρ (tk − t0 )] exp[(λA + λN )(t − t0 )]
≤ V (e0 ) exp[(λA + λN + ρ )(t − t0 )]. (6.17)

Thus, from (6.15)–(6.17), we have V (e(t)) → 0 and e(t) → 0 as t → ∞. Then, the

origin of the error system (6.4) is asymptotically stable. This completes the proof.

Corollary 6.2. Let βk be the largest eigenvalue of (I + Bk )T (I + Bk ) (k ∈ N). If there

exists a constant ρ > 1 such that
4 4 4 4
ln(ρβ2k−1 β2k ) + 4A + AT4 + 4N(x, x̃) + N T(x, x̃)4 (t2k+1 − t2k−1 ) ≤ 0,

then the impulsively controlled response system (6.3) is asymptotically synchro-

nized with the drive system (6.1). 

Remark 6.4. From (6.9) in Theorem 6.2, it can be seen that we need only to choose
the odd switching sequence {t2k−1 } instead of the whole switching sequence {tk } as
in Theorem 6.1. 

Remark 6.5. The inequality (6.9) can be generalized to the following condition.
There exist a finite integer n0 > 0 and a constant ρ > 1 such that

ln(ρβn0 (k−1)+1 · · · βn0 k ) + (λA + λN )(tn0 k+1 − tn0 (k−1)+1 ) ≤ 0, k ∈ N. (6.18)

The choice of n0 in (6.18) depends on the actual system to be considered. Especially,

n0 = 1 corresponds to the case of the whole switching sequence {tk } and n0 = 2
corresponds to the case of the odd switching sequence {t2k−1 }. 

Remark 6.6. In Theorem 6.2 and Corollary 6.2, if t2k+1 − t2k−1 = Δ > 0 and Bk = B
(k ∈ N), then the conditions of Theorem 6.2 and Corollary 6.2 can be reduced to

ln(ρβ 2 ) + (λA + λN )Δ ≤ 0
176 6 Synchronizing Chaotic Systems via Impulsive Control

and 4 4 4 4
ln(ρβ 2 ) + 4A + AT4 + 4N(x, x̃) + N T (x, x̃)4 Δ ≤ 0,
respectively, where β is the largest eigenvalue of (I + B)T(I + B). 

6.2.3 Simulations

We take both the Lorenz system and the unified system as examples to show how to
use the proposed methods to synchronize chaotic systems.

6.2.3.1 Synchronizing the Lorenz System

The Lorenz system is described by the following differential equations:

⎧
⎨ ẋ1 = α1 (x2 − x1),
⎪
ẋ2 = α2 x1 − x2 − x1x3 , (6.19)
⎪
⎩
ẋ3 = x1 x2 − α3 x3 .

First, we decompose the linear and nonlinear parts of the Lorenz system (6.19), and
rewrite it as follows:
ẋ = Ax + h(x), (6.20)
⎛ ⎞ ⎛ ⎞
−α1 α1 0 0
where x(t) = (x1 , x2 , x3 )T , A = ⎝ α2 −1 0 ⎠, and h(x) = ⎝ −x1 x3 ⎠. The im-
0 0 −α3 x1 x2
pulsively controlled response system is given by
⎧
⎨ x̃˙ = Ax̃ + h(x̃), t = tk ,
Δ x̃ = −Bk e(tk ), t = tk , k ∈ N, (6.21)
⎩ +
x̃(t0 ) = x̃0 ,

where x̃ = (x̃1 , x̃2 , x̃3 )T . From (6.20) and (6.21), we have

⎧
⎨ ė(t) = Ae + h(x) − h(x̃), t = tk ,
Δ e = B e(t ), t = tk , k ∈ N,
⎩ + k k
e(t0 ) = e0 ,

where e = (e1 , e2 , e3 )T = (x1 − x̃1 , x2 − x̃2 , x3 − x̃3 )T is the synchronization error and
⎛ ⎞ ⎛ ⎞⎛ ⎞
0 0 0 0 e1
h(x) − h(x̃) = ⎝ x̃1 x̃3 − x1x3 ⎠ = ⎝ −x3 0 −x̃1 ⎠ ⎝ e2 ⎠ .
x1 x2 − x̃1x̃2 x2 x̃1 0 e3

Therefore, we obtain
6.2 Complete Synchronization of a Class of Chaotic Systems via Impulsive Control 177
⎛ ⎞
0 0 0
N(x, x̃) = ⎝ −x3 0 −x̃1 ⎠ ,
x2 x̃1 0
⎛ ⎞
0 −x3 x2
N(x, x̃) + N T (x, x̃) = ⎝ −x3 0 0 ⎠ ,
x2 0 0
and 4 4
4N(x, x̃) + N T (x, x̃)4 = max{|x3 | + |x2|, |x3 |, |x2 |}.
∞
In this simulation, we take α1 = 10, α2 = 28, and α3 = 8/3. Then, λA = 28.0512.
4 system, we find4−20 ≤ x1 ≤ 20, −30 ≤ x2 ≤
From the chaotic attractor of the Lorenz
30, and 0 ≤ x3 ≤ 50, which leads to 4N(x, x̃) + N T(x, x̃)4∞ ≤ 80. Let δ¯k = δ¯ (k ∈ N)
and Bk = B = diag{σ , σ , σ }. Then, βk = β = (σ + 1)2 . According to Corollary 6.1
and Remark 6.3, the impulsively controlled response system (6.21) is asymptotically
synchronized with the drive system (6.20) if the following condition is satisfied:
2
ln ρ + ln(σ + 1)
0 < δ¯ ≤ − .
108.0512

Letting σ = −0.7 and ρ = 1.1, we have 0 < δ¯ ≤ 0.0214. The numerical simulation
results with σ = −0.7 and δ¯ = 0.02 are shown in Fig. 6.1. The initial conditions
of the drive and response systems are taken as x0 = (6, 2, 4)T and x̃0 = (4, 1, 5)T ,
respectively.

1
e1

−1
0 0.05 0.1 0.15 0.2 0.25 0.3

1
e2

−1
0 0.05 0.1 0.15 0.2 0.25 0.3

0.5

0
e3

−0.5

−1
0 0.05 0.1 0.15 0.2 0.25 0.3
t (sec)
Fig. 6.1 Synchronization error curves of the Lorenz system with δ = 0.02
178 6 Synchronizing Chaotic Systems via Impulsive Control

6.2.3.2 Synchronizing the Unified System

The unified system is described by the following differential equations:

⎧
⎪
⎪ ẋ1 = (25a + 10)(x2 − x1 ),
⎨
ẋ2 = (28 − 35a)x1 − x1 x3 + (29a − 1)x2, (6.22)
⎪
⎩ ẋ3 = x1 x2 − 1 (a + 8)x3.
⎪
3
When a ∈ [0, 1], the unified system has a chaotic attractor, which is shown in Fig.
6.2. First, we decompose the linear and nonlinear parts of the unified system (6.22),
and rewrite it as follows:
ẋ = Ax + h(x), (6.23)
where x(t) = (x1 , x2 , x3 )T ,
⎛ ⎞ ⎛ ⎞
−(25a + 10) 25a + 10 0 0
⎜ 0 ⎟ , and h(x) = ⎝ −x x ⎠ .
A = ⎝ 28 − 35a 29a − 1 ⎠ 1 3
a+8
0 0 − x1 x2
3

30

20

10

0
x1

−10

−20

−30
40
20 60
50
0 40
30
−20 20
10
x2 −40 0 x3

Fig. 6.2 The attractor of the unified system with a = 1

6.2 Complete Synchronization of a Class of Chaotic Systems via Impulsive Control 179

The impulsively controlled response system is given by

⎧
⎨ x̃˙ = Ax̃ + h(x̃), t = tk ,
Δ x̃ = −Bk e(tk ), t = tk , k ∈ N, (6.24)
⎩ +
x̃(t0 ) = x̃0 ,

where x̃ = (x̃1 , x̃2 , x̃3 )T . From (6.23) and (6.24), we have

⎧
⎨ ė(t) = Ae + h(x) − h(x̃), t = tk ,
Δ e = B e(t ), t = tk , k ∈ N,
⎩ + k k
e(t0 ) = e0 ,

where e = (e1 , e2 , e3 )T = (x1 − x̃1 , x2 − x̃2 , x3 − x̃3 )T is the synchronization error and
⎛ ⎞ ⎛ ⎞⎛ ⎞
0 0 0 0 e1
h(x) − h(x̃) = ⎝ x̃1 x̃3 − x1x3 ⎠ = ⎝ −x3 0 −x̃1 ⎠ ⎝ e2 ⎠ .
x1 x2 − x̃1x̃2 x2 x̃1 0 e3

Therefore, we obtain ⎛ ⎞
0 0 0
N(x, x̃) = ⎝ −x3 0 −x̃1 ⎠ ,
x2 x̃1 0

1
e1

−1
0 0.05 0.1 0.15 0.2 0.25 0.3

1
e2

−1
0 0.05 0.1 0.15 0.2 0.25 0.3

0.5

0
e3

−0.5

−1
0 0.05 0.1 0.15 0.2 0.25 0.3
t (sec)
Fig. 6.3 Synchronization error curves of the unified system with Δ = 0.03
180 6 Synchronizing Chaotic Systems via Impulsive Control
⎛ ⎞
0 −x3 x2
N(x, x̃) + N T (x, x̃) = ⎝ −x3 0 0 ⎠ ,
x2 0 0
and 4 4
4N(x, x̃) + N T (x, x̃)4 = max{|x3 | + |x2|, |x3 |, |x2 |}.
∞
4 4
In this simulation, we choose a = 1. Then, 4A + AT4∞ = 98. From Fig. 6.2, we find
−30 ≤ x1 ≤ 30, −30 ≤ x2 ≤ 35, and 0 ≤ x3 ≤ 55. Therefore,
4 4
4N(x, x̃) + N T (x, x̃)4 ≤ 90.
∞

Let t2k+1 −t2k−1 = Δ (k ∈ N) and Bk = B = diag{σ , σ , σ }. Then, βk = β = (σ + 1)2 .

According to Corollary 6.2 and Remark 6.6, the impulsively controlled response
system (6.24) is asymptotically synchronized with the drive system (6.23) if the
following condition is satisfied:

ln ρ + 2 ln(σ + 1)2
0<Δ ≤− .
188
Let σ = −0.8 and ρ = 1.1. Then, 0 < Δ ≤ 0.0338. The numerical simulation results
with σ = −0.8 and Δ = 0.03 are shown in Fig. 6.3. The initial conditions of the drive
and response systems are taken as x0 = (6, 2, 4)T and x̃0 = (4, 1, 5)T , respectively.

6.3 Lag-Synchronization of the Unified Systems via Impulsive

Control

Strictly speaking, in the practical environment with signal propagation delays, it is

not reasonable to require the response system to synchronize with the drive system at
exactly the same time. For this reason, Jiang et al. [5] redefined the chaotic synchro-
nization in such a way that the state of the drive system at time t − η asymptotically
synchronizes with that of the response system at time t, namely,

lim x(t − η ) − x̃(t) = 0,

t→∞

where x and x̃ are the state vectors of the drive system and the response system,
respectively, and η is a finite channel time delay, which is often unknown.
In this section, we study how to achieve the synchronization of the unified sys-
tems with channel time delay and parametric uncertainties based on the practical
stability theory. Meanwhile, a new definition of chaos synchronization is described,
namely, the state of the drive system at time t − η synchronizes with that of the re-
sponse system at time t in the sense of practical stability, i.e., for any different initial
conditions, x0 (t − η ) and x̃0 (t),

lim x(t − η ) − x̃(t) < ζ , 0 < ζ < +∞,

t→∞
6.3 Lag-Synchronization of the Unified Systems via Impulsive Control 181

where · refers to the Euclidean vector norm or the induced matrix 2-norm.

6.3.1 Preliminaries

Let a general nonautonomous chaotic system be

ẋ = f (t, x, u(t)),

where f : R+ ×Rn ×Rm → Rn is continuous, x ∈ Rn is the state vector, and u : R+ →

Rm is the external force which is independent of the system. Suppose that a set of
discrete instants {tk } satisfies 0 < t1 < t2 < · · · < tk < tk+1 < · · · , lim tk = ∞ (k ∈ N),
k→∞
and 0 ≤ t0 < t1 .
Let U(k, x) = Δ x|t=tk = x(tk+ ) − x(tk− ) = x(tk+ ) − x(tk ) be the change of the state
vector at the instant tk . Then, the impulsively controlled chaotic system is given by
⎧
⎨ ẋ = f (t, x, u(t)), t = tk ,
Δ x = U(k, x), t = tk , k ∈ N, (6.25)
⎩ +
x(t0 ) = x0 , t0 ≥ 0.

To study the practical stability of impulsive differential equation given in (6.25),

we introduce the following definitions and lemmas.

Definition 6.1 ([17]). Let V : R+ × Rn → R+ . Then, V is said to belong to class V0

if
(i) V is continuous in (tk−1 ,tk ] × Rn and

lim V (t, y) = V (tk+ , x).

(t,y)→(tk+ ,x)

(ii) V is locally Lipschitz in x. 

Definition 6.2 ([17]). Let V ∈ V0 . For any (t, x) ∈ (tk−1 ,tk ] × Rn , the upper and right
(Dini) derivative of V (t, x(t)), along the solution x(t) of (6.25), is defined as

1
D+V (t, x) := lim sup [V (t + h, x + h f (t, x, u(t))) − V(t, x)]. 
h→0 + h
Instead of studying the stability of the nth-order impulsive differential equation
given in (6.25), it is convenient to study that of a scalar impulsive differential equa-
tion, whose stability properties are related to (6.25).
Definition 6.3 ([19]). Comparison system. Let V ∈ V0 and assume that

D+V (t, x) ≤ g(t,V (t, x), v(t)), t = tk ,

V (t, x + U(k, x)) ≤ ψk (V (t, x)), t = tk ,

 182 6 Synchronizing Chaotic Systems via Impulsive Control

where g : R+ × R+ × R+ → R is continuous and ψk : R+ → R+ is nondecreas-

ing. ω , v ∈ Θ := {σ : R+ → R+ , σ (s) is continuous for s ∈ (tk−1 ,tk ], lim σ (t) =
t→tk+
σ (tk+ )}. Then, the system
⎧
⎨ ω̇ = g(t, ω , v(t)), t = tk ,
ω (tk+ ) = ψk (ω (tk )), t = tk , k ∈ N, (6.26)
⎩
ω (t0+ ) = ω0 ≥ 0

is the comparison system of system (6.25). 

Definition 6.4 ([19]). PC(R+ ×Rn , R+ ) denotes the set of functions Φ : R+ ×Rn →
R+ which are continuous for t ∈ R+ , t = tk , and are left-continuous for every t ∈ R+ .
At points tk , Φ has a discontinuity of the first kind. 
Let a set Ω be

Ω := {u ∈ Rm : Γ (t, u) ≤ r(t), t ≥ t0 },

where Γ ∈ C(R+ × Rm , R+ ) and r(t) is the maximal solution of the comparison

system (6.26).
Definition 6.5 ([6]). Practical stability. System (6.25) is said to be practically stable
if, given (λ , ζ ), λ > 0, and ζ > 0, we have that x0 < λ implies that x(t) < ζ ,
t ≥ t0 , for some t0 ∈ R+ and every u ∈ Ω . 
Definition 6.6 ([17]). A function α is said to belong to class K if α ∈ C(R+ , R+ ),
α (0) = 0, and α (x) is strictly increasing in x. 
Lemma 6.1 ([13]). Assume that
(i) g ∈ PC(R+ × R+ × R+ , R+ ) and g(t, u, v) is nondecreasing in u for each (t, v)
and nondecreasing in v for each (t, u).
(ii) ψk is nondecreasing for each k.
(iii) 0 < λ < ζ is given.
(iv) V ∈ PC(R+ × Rn , R+ ) , V (t, x) is locally Lipschitz in x, and there exist α , β ∈
K such that

β ( x ) ≤ V (t, x) ≤ α ( x ), (t, x) ∈ R+ × S(ρ ), ρ > ζ .

(v) For (t, x) ∈ (tk ,tk+1 ] × S(ρ ) and u(t) ∈ Ω ,

D+V (t, x) ≤ g(t,V (t, x), Γ (t, u)).

(vi) x ∈ S(ζ ) implies that x + U(k, x) ∈ S(ρ ) and V (tk+ , x + U(k, x)) ≤ ψk (V (t, x)).
(vii) α (λ ) < β (ζ ).
Then, the practical stability properties of the comparison system (6.26) with respect
to (α (λ ), β (ζ )) imply the corresponding practical stability properties of system
(6.25) with respect to (λ , ζ ) for every u(t) ∈ Ω , where S(ρ ) := {x ∈ Rn : x <
ρ }. 
6.3 Lag-Synchronization of the Unified Systems via Impulsive Control 183

Remark 6.7. In Lemma 6.1, if the conditions in (iii) and (vii) are changed to 0 <
ζ < λ and α (λ ) > β (ζ ), respectively, the conclusion is still valid. 

Lemma 6.2 ([19]). Let the comparison system (6.26) be described by the following
equations:
⎧
⎪ g(t, ω , v(t)) = ϕω + θ , ϕ > 0, θ > 0, t = tk ,
⎪
⎨
⎪ ψk (ω ) = γk ω , γk > 0, t = tk , k ∈ N, (6.27)
⎪
⎩
ω (t0+ ) = ω0 ≥ 0.

Let γk = γ > 0 be a constant. For given (λ , ζ ), λ > 0, ζ > 0, if

1
ln(γ ) + ϕ < 0
δ
and
θ 1 − eϕδ
< ζ,
ϕ (1 − γ eϕδ )
where δ = tk+1 − tk , 0 < δ < +∞, then system (6.25) is practically stable with re-
spect to (λ , ζ ) for any λ < +∞. 

6.3.2 Main Results

Consider the following unified system with parametric uncertainty:

ẋ(t) = (A + Δ A)x(t) + h(x(t)), (6.28)

where x(t) ∈ Rn is the state vector, A ∈ Rn×n is a constant matrix, h(x(t)) : Rn → Rn

is a continuous nonlinear function, and Δ A ∈ Rn×n is a time-varying parametric
uncertainty, which satisfies the following assumptions:
(i) Δ A = μ F(t), μ > 0, F(t) = [ fi j (t)]n×n , fi j (t) ≤ 1.
(ii) Δ A does not destroy the chaotic behavior of the chaotic system (6.28).
Taking system (6.28) as the drive system, the response system can be described
by
x̃˙(t) = (A + Δ Ã)x̃(t) + h(x̃(t)). (6.29)
Considering channel time delay η , we replace t with t − η in (6.28), and write
the drive system at time t − η as follows:

ẋ(t − η ) = (A + Δ Ā)x(t − η ) + h(x(t − η )). (6.30)

In the above, Δ Ã, Δ Ā, and Δ A satisfy the same assumptions, Δ Ã = μ F̃(t) =
μ [ f˜i j (t)]n×n , and Δ Ā = μ F(t − η ) = μ [ fi j (t − η )]n×n .
184 6 Synchronizing Chaotic Systems via Impulsive Control

If the linear error feedback composed of the state variables of the drive system
(6.30) and the response system (6.29) is used as impulsive control signal, the con-
trolled response system can be written in the following form:
⎧
⎨ x̃˙(t) = (A + Δ Ã)x̃(t) + h(x̃(t)), t = tk ,
Δ x̃(t) = Bk (x̃(t) − x(t − η )), t = tk , k ∈ N, (6.31)
⎩ +
x̃(t0 ) = x̃0 .

From (6.30) and (6.31), the following error system can be obtained:
⎧
⎪
⎪ ė(t) = Ae(t) + h(x(t − η )) − h(x̃(t))
⎪
⎪
⎨ + Δ Āx(t − η ) − Δ Ãx̃(t), t = tk ,
(6.32)
⎪
⎪ Δ e|t=tk = Bk e(tk ), t = tk , k ∈ N,
⎪
⎪
⎩ e(t + ) = e ,
0 0

where e(t) = x(t − η ) − x̃(t) = (x1 (t − η ) − x̃1 (t), x2 (t − η ) − x̃2 (t), . . . , xn (t − η ) −

x̃n (t))T is the synchronization error. Here, the objective is to design the impulsive
control gains Bk ∈ Rn×n (k ∈ N) such that the error system (6.32) is practically
stable, which implies that the impulsively controlled system (6.31) is synchronized
with the drive system (6.30) for any initial conditions.
Let Bk = diag{γk − 1, γk − 1, γk − 1} (k ∈ N). From system (6.22), system (6.32)
can be described by
⎧ 3 3
⎪
⎪ ė1 = −(25a + 10)e1 + (25a + 10)e2 + μ ∑ f1i xi − ∑ f˜1i x̃i ,
⎪
⎪
⎪
⎪ i=1 i=1
⎪
⎪
⎪
⎪
3 3
⎨ ė2 = (28 − 35a)e1 + (29a − 1)e2 − x3e1 − x̃1e3 + μ ∑ f2i xi − ∑ f˜2i x̃i ,
⎪
i=1 i=1
⎪
⎪ 1 3 3
⎪
⎪
⎪
ė3 = − (a + 8)e3 + x2e1 + x̃1e2 + μ ∑ f3i xi − ∑ f˜3i x̃i , t = tk ,
⎪
⎪ 3 i=1 i=1
⎪
⎪ Δ e|t=tk = Bk e, t = tk , k ∈ N,
⎪
⎪
⎩ +
e(t0 ) = e0 .
(6.33)

For simplicity, in (6.33) and the rest of this section, e(t), x̃(t), x(t − η ), F̃(t),
F(t − η ), f˜i j (t), and fi j (t − η ) are denoted by e, x̃, x, F̃, F, f˜i j , and fi j , respectively.
Remark 6.8. Due to the boundedness of the chaotic signals, there exist constants M j
such that x̃ j ≤ M j , x j ≤ M j , j = 1, 2, 3. 
Theorem 6.3 ([11]). The practical stability of the synchronization error system
(6.32) with respect to (λ , ζ ) is equivalent to that of the comparison system (6.27)
√ 3
with respect to ( 3λ , ζ ) , where ϕ = Hmax = max{H1 , H2 , H3 }, θ = H f = 6μ ∑ M j ,
j=1
H1 = |28 − 35a| − (25a + 10) + M2 + M3 + 3μ , H2 = |25a + 10| + (29a − 1) + M1 +
1
3μ , and H3 = − (a + 8) + M1 + 3μ .
3
6.3 Lag-Synchronization of the Unified Systems via Impulsive Control 185

Proof. Choose a Lyapunov function as

V (t, e) = |e1 | + |e2 | + |e3 | .

Then, we have

D+V (t, e) = ė1 sgn(e1 ) + ė2 sgn(e2 ) + ė3 sgn(e3 )

5 ' (6
3 3
= −(25a + 10)e1 + (25a + 10)e2 + μ ∑ f1i xi − ∑ f˜1i x̃i sgn(e1 )
i=1 i=1
5
+ (28 − 35a)e1 + (29a − 1)e2 − x3 e1 − x̃1 e3
' (6
3 3
+μ ∑ f2i xi − ∑ f˜2i x̃i sgn(e2 )
i=1 i=1
5 ' (6
3 3
1
+ − (a + 8)e3 + x2 e1 + x̃1 e2 + μ
3 ∑ f3i xi − ∑ f˜3i x̃i sgn(e3 )
i=1 i=1

5 6
3
≤ |28 − 35a| − (25a + 10) + |x2 | + |x3 | + μ ∑ | fi1 | |e1 |
i=1
5 6
3
+ |25a + 10| + (29a − 1) + |x̃1 | + μ ∑ | fi2 | |e2 |
i=1
5 6
3
1
+ − (a + 8) + |x̃1 | + μ ∑ | fi3 | |e3 |
3 i=1
3 3
+μ ∑∑ fi j − f˜i j x̃ j
j=1 i=1
3 3
≤ H1 |e1 | + H2 |e2 | + H3 |e3 | + μ ∑∑ fi j − f˜i j M j
j=1 i=1
3
≤ max{H1 , H2 , H3 }V (t, e) + 6μ ∑ Mj.
j=1

Hence, g(t, ω , v) = ϕω + θ , where ϕ = Hmax = max{H1 , H2 , H3 } and θ = H f =

3
6μ ∑ M j .
j=1
Since V (tk+ , e(tk+ )) = V (tk , (I + Bk )e(tk )) = γkV (tk , e(tk )), we have ψk (ω ) = γk ω .

As is well known, |e1 | + |e2 | + |e3 | ≥ e21 + e22 + e23, and from
186 6 Synchronizing Chaotic Systems via Impulsive Control

(|e1 | + |e2 | + |e3 |)2 = e21 + e22 + e23 + 2 |e1 | |e2 | + 2 |e2 | |e3 | + 2 |e1 | |e3 |
≤ e21 + e22 + e23 + (e21 + e22 ) + (e22 + e23 ) + (e21 + e23 )
= 3(e21 + e22 + e23 ),
√  √
we get |e1 | + |e2 | + |e3 | ≤ 3 e21 + e22 + e23 , i.e., α (s) = 3s and β (s) = s.
From e ∈ S(ζ ), or e < ζ , we have

e + U(k, e) = e + Bk e = γk e < γk ζ < ρ .

If there exists γk ∈ (0, 1) such that the last inequality holds, then e +U(k, e) ∈ S(ρ ).
Other conditions in Theorem 6.3 can easily be proved, which are omitted here. This
completes the proof of the theorem. 

Based on the results of Lemma 6.2 and Theorem 6.3, we will give the following
practical stability criterion of the error system (6.32).

Theorem 6.4 ([11]). Let γk = γ > 0 be a constant, δ = tk+1 −tk , 0 < δ < +∞, k ∈ N,
(λ , ζ ) be given, λ > 0, and ζ > 0. If

1
ln(γ ) + Hmax < 0, (6.34)
δ
H f 1 − eHmax δ
< ζ, (6.35)
Hmax (1 − γ eHmax δ )
then the synchronization error system (6.32) is practically stable with respect to
(λ , ζ ) for any λ < +∞, where Hmax and H f are defined in Theorem 6.3.

Proof. Replace ϕ and θ in Lemma 6.2 with Hmax and H f in Theorem 6.3, respec-
tively. Then, the conclusion can easily be obtained. For details, which are omitted
here, please refer to [19]. This completes the proof. 

6.3.3 Simulations

Let a = 1. From the attractor of the unified system, we get −30 ≤ x1 ≤ 30,
−30 ≤ x2 ≤ 35, 0 ≤ x3 ≤ 55, and then M ⎛1 = 30, M2 = 35, ⎞ M3 = 55. Assume
sint cost sint
that the parametric uncertainty Δ A = 0.03 ⎝ cost cost cost ⎠, i.e., μ = 0.03 and
sint cost sint
⎛ ⎞
sint cost sint
F(t) = [ fi j (t)]3×3 = ⎝ cost cost cost ⎠. By computation, we have H1 = 62.09,
sint cost sint
H2 = 93.09, H3 = 27.09, and H f = 21.6. So, Hmax = 93.09. The initial conditions
6.4 Impulsive Synchronization of Different Chaotic Systems 187

0.4

0.2
e1

−0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.4

0.2
e2

−0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.2

0
e3

−0.2

−0.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t (sec)

Fig. 6.4 Synchronization error curves with δ = 0.005, γ = 0.1

of the drive and response systems are x0 = (4, 1, 5)T and x̃0 = (3.6, 0.8, 5.4)T, re-
spectively. Let δ = 0.005. From (6.34) it gives γ < 0.6279. We choose γ = 0.1, and
from (6.35) it gives ζ ≈ 0.1636. It means that under this condition, the maximum
amplitudes of e1 , e2 , and e3 are less than 0.1636. We choose the ‘unknown’ channel
time delay η = 3 sec. The simulation result is shown in Fig. 6.4. Obviously, the syn-
chronization errors fluctuate around zero with small amplitudes that are estimated
by Theorem 6.4.

Remark 6.9. ζ will decrease as the parametric uncertainty Δ A decreases. Specifi-

cally, ζ will converge to zero when Δ A = 0; that is, the response system (6.31) is
asymptotically synchronized with the drive system (6.30). 

6.4 Impulsive Synchronization of Different Chaotic Systems

In practice, it is difficult to ensure the structures of the drive and the response sys-
tems to be identical. Moreover, the parametric uncertainties of the drive and the
response systems are always different and time varying. Therefore, it is significant
188 6 Synchronizing Chaotic Systems via Impulsive Control

to investigate synchronization schemes between two different chaotic systems in the

presence of different time-varying parametric uncertainties.

6.4.1 System Description and Synchronization Problem

Consider the drive system as follows:

ẋ(t) = (A + Δ A(t))x(t) + f (x(t)), (6.36)

where x ∈ Rn is the state vector, A ∈ Rn×n is a known constant matrix, and f (x(t)) ∈
Rn is a continuous nonlinear function. Assume that

f (x(t)) − f (x̃(t)) = N(x(t), x̃(t))(x(t) − x̃(t)), (6.37)

where N(x(t), x̃(t)) ∈ Rn×n is a matrix with bounded elements depending on x(t) ∈
Rn and x̃(t) ∈ Rn .
The parametric uncertainty Δ A(t) is said to be admissible if Δ A(t) satisfies the
following assumptions:
(i) Δ A(t) = EF(t)H, where E and H are known real constant matrices with appro-
priate dimensions, and the uncertain matrix F(t) ∈ Rn×n satisfies F(t) ≤ 1.
(ii) The parametric uncertainty Δ A(t) does not destroy the chaotic behavior of the
chaotic system (6.36).
Remark 6.10. The parametric uncertainty Δ A(t) is a time-varying perturbation,
which is often used in other papers to deal with the scheme of robust stabilization
for uncertain systems [2, 14].
Consider the channel time delay η ; we replace t with t − η in (6.36), and write
the drive system at time t − η as follows:

ẋ(t − η ) = (A + Δ A(t − η ))x(t − η ) + f (x(t − η )). (6.38)

The response system is as follows:

x̃˙(t) = (Ã + Δ Ã(t))x̃(t) + g(x̃(t)), (6.39)

where x̃ ∈ Rn is the state vector, Ã ∈ Rn×n is a known constant matrix, and g(x(t)) ∈
Rn is a continuous nonlinear function. The parametric uncertainty Δ Ã(t) = Ẽ F̃(t)H̃
satisfies the same assumptions as that for Δ A(t). Suppose that the following condi-
tion holds:
f (x̃(t)) − g(x̃(t)) = U(x̃(t))x̃(t), (6.40)
where U(x̃(t)) ∈ Rn×n is a matrix with each element bounded depending on x̃(t).
Suppose that a set of discrete instants {tk } satisfies

0 < t1 < t2 < · · · < tk < tk+1 < · · · , lim tk = ∞ (k ∈ N), 0 ≤ t0 < t1 .
k→∞
6.4 Impulsive Synchronization of Different Chaotic Systems 189

At the discrete time tk , the state variables of the drive system are transmitted to
the response system as part of the control inputs such that the state variables of
the response system are suddenly changed at these instants. Therefore, the response
system can be written in the following form:
⎧
⎨ x̃˙(t) = (Ã + Δ Ã(t))x̃(t) + g(x̃(t)), t = tk ,
Δ x̃(t) = x̃(tk+ ) − x̃(tk− ) = x̃(tk+ ) − x̃(tk ) = −Bk e(t), t = tk , k ∈ N, (6.41)
⎩ +
x̃(t0 ) = x̃0 ,

where

e(t) = x(t − η ) − x̃(t) = [x1 (t − η ) − x̃1 (t), x2 (t − η ) − x̃2 (t), . . . , xn (t − η ) − x̃n (t)]T

is the synchronization error. x̃(t) is left-continuous at t = tk and Bk ∈ Rn×n are the

impulsive control gains. Then, from (6.38) and (6.41), the following error system is
obtained:
⎧
⎪
⎪ ė(t) = [A + Δ A(t − η ) + N(x(t − η ), x̃(t))] e(t)
⎪
⎪ 7 8
⎨ + A − Ã + Δ A(t − η ) − Δ Ã(t) + U(x̃(t)) x̃(t), t = tk ,
(6.42)
⎪
⎪ Δ e|t=tk = e(tk+ ) − e(tk ) = Bk e(tk ), t = tk , k ∈ N,
⎪
⎪
⎩ +
e(t0 ) = e0 .

For simplicity, in the rest of the section, x̃(t), x(t − η ), N(x(t − η ), x̃(t)), U(x̃(t)),
Δ A(t − η ), and Δ Ã(t) are denoted by x̃, x, N, U, Δ A, and Δ Ã, respectively.
Remark 6.11. Due to the boundedness of the chaotic signals, there exist positive
constants ς and ς̃ such that x ≤ ς and x̃ ≤ ς̃ . 

Definition 6.7. The synchronization of systems (6.38) and (6.41) is said to have
been achieved if, for arbitrary initial conditions x0 and x̃0 , the trivial solution of the
error system (6.42) converges to a predetermined neighborhood of the origin for any
admissible parametric uncertainties. 

Here, the objective is to find the conditions on the control gains Bk and the im-
pulsive distances δ¯k = tk − tk−1 (k ∈ N) such that the error magnitude, i.e., e ,
converges to below some constant ξ , which implies that the impulsively controlled
response system (6.41) is synchronized with the drive system (6.38) for arbitrary
initial conditions.

6.4.2 Main Results

The following theorem gives sufficient conditions for robust stability of the error
system (6.42), which imply that the impulsively controlled response system (6.41)
is synchronized with the drive system (6.38).
190 6 Synchronizing Chaotic Systems via Impulsive Control

Theorem 6.5. Let βk , λA , and λN be the largest eigenvalues of (I + Bk )T (I + Bk ),

A + AT, and N + N T , respectively. The response system (6.41) is synchronized with
the drive system (6.38) if one of the following conditions holds:
(i) when θ1 ≥ 0, there exists a constant ρ > 1 such that

ln(ρβ2k−1 β2k ) + θ1 (t2k+1 − t2k−1 ) ≤ 0 (6.43)

and
sup{βk exp[θ1 (tk+1 − tk )]} = ϒ < ∞; (6.44)
k

(ii) when θ1 < 0, there exists a constant ρ (0 ≤ ρ < −θ1 ) such that

ln(βk ) − ρ (tk − tk−1 ) ≤ 0, k ∈ N, (6.45)

where ξ > 0 is the bound of the error magnitude e and can be chosen small
enough, where
ς̃  4 4 4 44 4 4 4
θ1 = λA + λN +2 E H + 2 4A − Ã4 +2 E H +2 4Ẽ 4 4H̃ 4 + 4U + U T4 .
ξ
Proof. Choose the following Lyapunov function:

V (e) = eT e.

For t ∈ (tk−1 ,tk ], k ∈ N, the time derivative of V (e) along the solution of (6.42) is

V̇ (e(t)) = ėT e + eTė

7 8T
= (A + Δ A + N)e + (A − Ã+ Δ A − Δ Ã + U)x̃ e
7 8
+ eT (A + Δ A + N)e + (A − Ã+ Δ A − Δ Ã + U)x̃
= eT (A + AT)e + eT(N + N T )e + eT (Δ A + Δ AT)e
+ 2eT(A − Ã + Δ A − Δ Ã + U)x̃
≤ (λA + λN + 2 E H ) eT e
 4 4 4 44 4 4 4
+ 2 4A − Ã4 + 2 E H + 2 4Ẽ 4 4H̃ 4 + 4U + U T4 ς̃ e .
2
For e ≥ ξ one can conclude that e ≤ e /ξ and, therefore,
6.4 Impulsive Synchronization of Different Chaotic Systems 191

V̇ (e(t)) ≤ (λA + λN + 2 E H )eT e

ς̃  4 4 4 44 4 4 4
+ 2 4A − Ã4 + 2 E H + 2 4Ẽ 4 4H̃ 4 + 4U + U T4 e 2
ξ

≤ λA + λN + 2 E H

ς̃  4 4 4 44 4 4 4
+ 2 4A − Ã4 + 2 E H + 2 4Ẽ 4 4H̃ 4 + 4U + U T4 V (e(t))
ξ
= θ1V (e(t)), (6.46)

which implies that

V (e(t)) ≤ V (e(tk−1
+
)) exp[θ1 (t − tk−1 )], t ∈ (tk−1 ,tk ], k ∈ N. (6.47)

On the other hand, when t = tk , we obtain

V (e(tk+ )) = [(I + Bk )e(tk )]T (I + Bk )e(tk )

= eT (tk )(I + Bk )T (I + Bk )e(tk )
≤ βkV (e(tk )).

Let k = 1 in (6.47). Then, we have

V (e(t)) ≤ V (e(t0+ )) exp[θ1 (t − t0 )], t ∈ (t0 ,t1 ],

which leads to
V (e(t1 )) ≤ V (e(t0+ )) exp[θ1 (t1 − t0 )]
and
V (e(t1+ )) ≤ β1V (e(t1 )) ≤ V (e(t0+ ))β1 exp[θ1 (t1 − t0 )].
Similarly, for t ∈ (t1 ,t2 ], we have

V (e(t)) ≤ V (e(t1+ )) exp[θ1 (t − t1 )] ≤ V (e(t0+ ))β1 exp[θ1 (t − t0)].

In general, for t ∈ (tk ,tk+1 ],

V (e(t)) ≤ V (e(t0+ ))β1 β2 · · · βk exp[θ1 (t − t0 )]. (6.48)

If θ1 ≥ 0, there exists a constant ρ > 1 and, from (6.43), (6.44), and (6.48), when
(i) t ∈ (t2k−1 ,t2k ], we have
192 6 Synchronizing Chaotic Systems via Impulsive Control

2k−1
V (e(t)) ≤ V (e(t0+ )) ∏ βi exp[θ1 (t − t0)]
i=1
2k−1
≤ V (e0 ) ∏ βi exp[θ1 (t2k − t0)]
i=1
= V (e0 )β1 β2 exp[θ1 (t3 − t1 )] · · · β2k−3 β2k−2 exp[θ1 (t2k−1 − t2k−3 )]
× β2k−1 exp[θ1 (t2k − t2k−1 )] exp[θ1 (t1 − t0 )]
V (e0 )
≤ϒ exp[θ1 (t1 − t0 )]; (6.49)
ρ k−1

(ii) t ∈ (t2k ,t2k+1 ], we have

2k
V (e(t)) ≤ V (e(t0+ )) ∏ βi exp[θ1 (t − t0 )]
i=1
2k
≤ V (e0 ) ∏ βi exp[θ1 (t2k+1 − t0 )]
i=1
= V (e0 )β1 β2 exp[θ1 (t3 − t1 )] · · · β2k−1 β2k exp[θ1 (t2k+1 − t2k−1 )]
× exp[θ1 (t1 − t0 )]
V (e0 )
≤ exp[θ1 (t1 − t0 )]. (6.50)
ρk

If θ1 < 0, and there exists a constant ρ (0 ≤ ρ < −θ1 ) such that (6.45) holds, i.e.,
βk ≤ exp[ρ (tk − tk−1 )], then

V (e(t)) ≤ V (e(t0+ ))β1 β2 · · · βk exp[θ1 (t − t0 )]

≤ V (e0 ) exp[ρ (tk − t0 )] exp[θ1 (t − t0 )]
≤ V (e0 ) exp[(θ1 + ρ )(t − t0 )]. (6.51)

It follows from (6.49)–(6.51) that the error magnitude e will converge to below
the constant ξ if the error started from e > ξ . This completes the proof. 
Remark 6.12. Most typical chaotic systems with parametric uncertainties can be de-
scribed by (6.36) or (6.39), and the continuous nonlinear functions can satisfy (6.37)
and (6.40), such as the Lorenz system, the Chen system, the Rössler system, the uni-
fied system, the Lü system, several variants of Chua’s circuit, etc. 
Remark 6.13. If A = Ã and f (·) = g(·), then the structures of system (6.38) and sys-
tem (6.39) are identical. Therefore, Theorem 6.5 is also applicable to the impulsive
synchronization between two identical chaotic systems with or without parametric
uncertainties. 
Remark 6.14. In (6.41) or (6.42), the value of the channel time delay η is not re-
quired to be known for executing the impulsive control, since we can obtain the
time-delay signal x(t − η ) at the discrete time tk easily and use it in the response
system blindly when it is received. 
6.4 Impulsive Synchronization of Different Chaotic Systems 193

Based
4 on the 4 matrix 4 theory 4and the boundedness
4 4 of chaotic signals, we obtain
λA ≤ 4A + AT4, λN ≤ 4N + N T 4 ≤ φ , and 4U + U T 4 ≤ ϕ , where φ and ϕ are posi-
tive real constants that can be obtained for different chaotic systems. Then, we have
the following corollary.
Corollary 6.3. Let βk be the largest eigenvalue of (I + Bk )T (I + Bk ). If there exists
a constant ρ > 1 such that

ln(ρβ2k−1 β2k ) + θ2 (t2k+1 − t2k−1) ≤ 0, k ∈ N, (6.52)

and
sup{βk exp[θ2 (tk+1 − tk )]} = ϒ < ∞,
k

then the response system (6.41) is synchronized with the drive system (6.38), where
ξ > 0 is the bound of the error magnitude e and can be chosen small enough,
where
4 4 ς̃ 4 4 4 44 4
θ2 = 4A + AT4 + φ + 2 E H + 2 4A − Ã4 + 2 E H + 2 4Ẽ 4 4H̃ 4 + ϕ .
ξ

Remark 6.15. Though 4 Corollary
4 6.3 4 will be4more conservative than Theorem 6.5,
the computation of 4A + AT4 and 4N + N T 4 will be easier than that of λA and λN
in general. Therefore, the result in Corollary 6.3 will be more convenient for use in
practice. 
4 4
Remark46.16. From
4 the boundedness of chaotic systems, we know that λA , 4A + AT4,
λN , φ , 4A − Ã4, and ϕ are bounded, and the inequalities (6.43) and (6.52) can be
satisfied by choosing appropriate βk and Δk = t2k+1 − t2k−1 (k ∈ N). 
In practice, for purpose of convenience, the gains Bk are always selected as a
constant matrix and the impulsive distances Δk are set to be a positive constant.
Then, we have the following corollary.
Corollary 6.4. Assume that t2k+1 −t2k−1 = Δ > 0 and Bk = B (k ∈ N). If there exists
a constant ρ > 1 such that

ln(ρβ 2 ) + Δ θ2 ≤ 0 and β exp(Δ θ2 ) = ϒ < ∞,

then the response system (6.41) is synchronized with the drive system (6.38), where
ξ > 0 is the bound of the error magnitude e and can be chosen small enough, β
is the largest eigenvalue of (I + B)T(I + B), and θ2 is defined in Corollary 6.3. 
Remark 6.17. It is worth pointing out that the topic of synchronization between dif-
ferent chaotic systems with different parametric uncertainties has attracted a great
deal of attention, and the scheme considered here is more general in a practical en-
vironment. In this section, to compensate for the different structures of the drive
and the response systems, stronger impulsive strength and higher impulsive fre-
quency are necessary. On the other hand, to save the energy of impulsive control
194 6 Synchronizing Chaotic Systems via Impulsive Control

and improve its performance, one can introduce an additional feedback controller to
compensate for the different structures, and this will lead to a more complicated con-
troller. Hence, to achieve synchronization, we can choose one of the two schemes
mentioned above by considering implementation simplicity and performance im-
provement, which depend on the practical situation.

6.4.3 Simulations

In this simulation, the impulsive distances Δk are set to be a positive constant Δ , and
the gains Bk are selected as a constant matrix, i.e., Bk = B = diag{d, d, d}. Then,
β = (d + 1)2.
In order to observe the synchronization behavior of different chaotic systems, we
assume that the Lorenz system drives the Chen system. Therefore, the drive and the
response systems are as follows:
⎧
⎨ ẋ1 = α1 (x2 − x1 ),
ẋ2 = α2 x1 − x1 x3 − x2 , (6.53)
⎩
ẋ3 = x1 x2 − α3 x3

and ⎧
⎨ x̃˙1 = α̃1 (x̃2 − x̃1),
x̃˙2 = (α̃2 − α̃1 )x̃1 − x̃1 x̃3 + α̃2 x̃2 , (6.54)
⎩˙
x̃3 = x̃1 x̃2 − α̃3 x̃3 .
From (6.42), (6.53), and (6.54), the following compact form of the error system is
obtained:
⎧
⎨ ė(t) = (A + Δ A + N)e(t) + (A − Ã+ Δ A − Δ Ã + U)x̃(t), t = tk ,
Δ e|t=t = e(tk+ ) − e(tk ) = Bk e(tk ), t = tk , k ∈ N,
⎩ + k
e(t0 ) = e0 .

Let α1 = 10, α2 = 28, α3 = 8/3, α̃1 = 35, α̃2 = 28, and α̃3 = 3. Systems (6.53)
and (6.54) are chaotic with the above parameters. Then,
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
−α1 α1 0 −10 10 0 −20 38 0
A = ⎝ α2 −1 0 ⎠ = ⎝ 28 −1 0 ⎠, A + AT = ⎝ 38 −2 0 ⎠,
0 0 −α3 0 0 −8/3 0 0 −16/3
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
−α̃1 α̃1 0 −35 35 0 25 −25 0
à = ⎝ α̃2 − α̃1 α̃2 0 ⎠ = ⎝ −7 28 0 ⎠, and A − à = ⎝ 35 −29 0 ⎠.
0 0 −α̃3 0 0 −3 0 0 1/3
6.4 Impulsive Synchronization of Different Chaotic Systems 195

−3
x 10
1.4

1.2

0.8
Δ

0.6

ρ=1 1
0.4
3 3
10 10
Stable region
0.2 30 30

0
−1.5 −1.4 −1.3 −1.2 −1.1 −1 −0.9 −0.8 −0.7 −0.6 −0.5
d

Fig. 6.5 The boundaries of stable region with different ρ

−3
x 10
4

3.5

2.5
Δ

2
ξ =2 2

1.5 1.5 1.5

1 1 1

0.5 0.5
0.5
Stable region

0
−1.5 −1.4 −1.3 −1.2 −1.1 −1 −0.9 −0.8 −0.7 −0.6 −0.5
d

Fig. 6.6 The boundaries of stable region with different ξ

196 6 Synchronizing Chaotic Systems via Impulsive Control

0.2

0
e1

−0.2
0 10 20 30 40 50 60 70 80

0.2

0
e2

−0.2
0 10 20 30 40 50 60 70 80

0.02
e3

−0.02
0 10 20 30 40 50 60 70 80
t (sec)

(a)

0.7

0.6

0.5

0.4
e

0.3

0.2

0.1

0
0 10 20 30 40 50 60 70 80
t (sec)

(b)

Fig. 6.7 Simulation results for synchronization between the Lorenz system and the Chen system
with Δ = 0.001. (a) Synchronization errors, (b) synchronization error magnitude e
6.4 Impulsive Synchronization of Different Chaotic Systems 197

50

0
e1

−50
0 10 20 30 40 50 60 70 80

50
e2

−50
0 10 20 30 40 50 60 70 80

50
e3

−50
0 10 20 30 40 50 60 70 80
t (sec)

Fig. 6.8 Synchronization errors without impulsive control between the Lorenz system and the
Chen system with Δ = 0.001

⎛ ⎞
sin t 0 0
Let Δ A = −Δ Ã = 0.04 ⎝ 0 cost 0 ⎠, where E = H = Ẽ = H̃ = 0.2I and
0 0 sint
⎛ ⎞
sint 0 0
F(t) = −F̃(t) = ⎝ 0 cost 0 ⎠. We have
0 0 sint
⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎛ ⎞
0 0 0 0 0 e1
Ne = f (x) − f (x̃) = ⎝ −x1 x3 ⎠ − ⎝ −x̃1 x̃3 ⎠ = ⎝ −x3 0 −x̃1 ⎠ ⎝ e2 ⎠ ,
x1 x2 x̃1 x̃2 x2 x̃1 0 e3
⎛ ⎞ ⎛ ⎞
0 0 0 0 −x3 x2
i.e., N = ⎝ −x3 0 −x̃1 ⎠. Then, N + N T = ⎝ −x3 0 0 ⎠ and
x2 x̃1 0 x2 0 0
4 4  
4 N + N T 4 = x2 + x2 ≤ x2 + x2 + x2 = x ≤ ς = φ .
2 3 1 2 3
⎛ ⎞
0
From f (x̃) = g(x̃) = ⎝ −x̃1 x̃3 ⎠, we have U = 0 and ϕ = 0.
x̃1 x̃2
198 6 Synchronizing Chaotic Systems via Impulsive Control

4 In this4 simulation, 4we choose4 ς = 55 and ς̃ = 60.4 By 4 4computation,

4 we get
4A + AT4 = 50.0512, 4A − Ã4 = 57.5257, E H = 4Ẽ 4 4H̃ 4 = 0.04, and φ =
ς = 55. Fig. 6.5 shows the stable region for different ρ . The whole region under the
curve of ρ = 1 is the stable region. When ρ → ∞, the stable region approaches a
vertical line d = −1. Similarly, Fig. 6.6 shows the stable region for different ξ .
We choose ρ = 1.1, ξ = 0.5, d = −0.98, and the ‘unknown’ channel time delay
η = 3 sec. The initial conditions of the drive and response systems are taken as
x0 = (4, 1, 5)T and x̃0 = (3.6, 0.8, 5.4)T, respectively. Hence, from Corollary 6.4, the
estimate of the bound of the stable region is given by

ln ρ + ln β 2
0<Δ ≤− = 0.0011.
13930
Using these parameters, conditions in Corollary 6.4 are satisfied for Δ ≤ 0.0011.
Robust impulsive synchronization between the Lorenz system and the Chen system
with impulsive distance Δ = 0.001 is given in Fig. 6.7. We can see that the syn-
chronization has been achieved practically and e is smaller than ξ = 0.5. Fig. 6.8
shows the synchronization errors without impulsive control.

6.5 Impulsive Synchronization of a Class of Chaotic Delayed

Neural Networks

Recently, there have been many efforts for the study of dynamical properties of de-
layed neural networks (DNNs) [1, 9]. Furthermore, it has been shown that these
networks can exhibit some complicated dynamics and even chaotic behavior if the
parameters and time delays are appropriately chosen [3, 10, 23]. Thus, the problem
of stabilization and synchronization of chaotic DNNs has received extensive consid-
eration [15, 21, 22]. This section addresses a practical issue of using an impulsive
control method to synchronize a class of chaotic DNNs with different time-varying
parametric uncertainties. This class of chaotic DNNs includes several well-known
chaotic DNNs, such as chaotic delayed Hopfield neural networks (DHNNs) and
chaotic delayed cellular neural networks (DCNNs), etc. Based on the theory of Lya-
punov stability and impulsive FDEs, some new sufficient conditions are derived,
which are expressed in terms of matrix norm inequalities. Furthermore, the syn-
chronization error magnitude can be reduced arbitrarily as long as some specific
conditions hold.

6.5.1 Preliminaries

A class of chaotic DNNs considered in this section is described by the following

delayed differential equations:
6.5 Impulsive Synchronization of a Class of Chaotic Delayed Neural Networks 199
5 6
n n
u̇i (t) = −di ci ui (t) − ∑ ai j fˆj (u j (t)) − ∑ bi j fˆj (u j (t − τ )) + Ji ,
j=1 j=1 (6.55)
i = 1, 2, . . . , n,
or 7 8
u̇(t) = −D Cu(t) − A fˆ(u(t)) − B fˆ(u(t − τ )) + J , (6.56)
where n ≥ 2 denotes the number of neurons, u(t) = (u1 (t), u2 (t), . . . , un (t))T denotes
the state vector associated with the neurons, D = diag{d1 , d2 , . . . , dn } > 0 represents
the amplification gain matrix, and C = diag{c1 , c2 , . . . , cn } > 0 is an appropriate co-
efficient matrix such that the solution of the model given in (6.55) or (6.56) remains
bounded. The feedback matrix A = (ai j )n×n indicates the strength of neuron inter-
connections, while the delayed feedback matrix B = (bi j )n×n indicates the strength
of neuron interconnections with constant delay parameter τ . The activation func-
tion, fˆ(u) = ( fˆ1 (u1 ), fˆ2 (u2 ), . . . , fˆn (un ))T , describes the behavior of the neurons’
response, and J = (J1 , J2 , . . . , Jn )T is a real constant input vector.
The activation function, fˆi (xi ), satisfies the following assumption.

Assumption 6.1. Each function fˆi : R → R (i = 1, 2, . . . , n) is bounded, and satis-

fies the Lipschitz condition with a Lipschitz constant σi > 0, i.e., fˆi (u) − fˆi (v) ≤
σi |u − v| for all u, v ∈ R. 

Remark 6.18. Equation (6.55) unifies several well-known neural networks such as
Hopfield neural networks with or without delays and cellular neural networks with
or without delays. In particular, if di = 1 and the activation function fˆi (ui ) is sigmoid,
then (6.55) describes the dynamics of Hopfield neural networks. Similarly, if di = 1
and the activation function fˆi (ui ) = (|ui + 1| − |ui − 1|)/2, then (6.55) describes the
dynamics of cellular neural networks. 

Under Assumption 6.1, it is not difficult to prove the existence of an equilib-

rium point of system (6.56) by using Brouwer’s fixed point theorem. Let u∗ =
(u∗1 , u∗2 , . . . , u∗n )T be the equilibrium point of system (6.56). For convenience, we shift
the equilibrium point u∗ of (6.56) to the origin by letting x(t) = u(t) − u∗ , and then
system (6.56) can be transformed into

ẋ(t) = −D [Cx(t) − A f (x(t)) − B f (x(t − τ ))] , (6.57)

where x(t) = (x1 (t), x2 (t), . . . , xn (t))T is the state vector of the new system (6.57),
and f (x(t)) = ( f1 (x1 ), f2 (x2 ), . . . , fn (xn ))T with fi (xi ) = fˆi (xi + u∗i ) − fˆi (xi ), i =
1, 2, . . . , n. It is easy to see that by Assumption 6.1, fi (xi ) satisfies the following
assumption.

Assumption 6.2. Each function fi : R → R (i = 1, 2, . . . , n) is bounded, and satis-

fies the Lipschitz condition with a Lipschitz constant σi > 0, i.e., | fi (u) − fi (v)| ≤
σi |u − v| for all u, v ∈ R. 

To derive our main results, the following lemma is necessary.

200 6 Synchronizing Chaotic Systems via Impulsive Control

First, some notation about impulsive FDEs is given. We denote

K ∗ := {ψ ∈ K : ψ (s) < s, ∀s > 0}

and
Ξ := {H ∈ C(R+ , R+ ) : H(0) = 0, H(s) > 0, ∀s > 0}.
Consider the following impulsive FDEs:

ẋ(t) = f (t, xt ), t ≥ t0 ,
(6.58)
x(tk+ ) = Jk (x(tk− )), k ∈ N,

where f : [t0 , ∞) × PC([−τ , 0], Rn ) → Rn and Jk (x) : S(ρ ) → Rn for each k ∈ N. For
any t ≥ t0 , xt ∈ PC([−τ , 0], Rn ) is defined by xt (s) = x(t + s), −τ ≤ s ≤ 0. Suppose
that the set of discrete instants {tk } satisfies

0 < t1 < t2 < · · · < tk < tk+1 < · · · , lim tk = ∞, 0 ≤ t0 < t1 , k ∈ N.

k→∞

We denote lim x(t) = x(tk+ ) and lim x(t) = x(tk− ). Then, x(tk ) = x(tk− ) implies that
t→tk+ t→tk−
x(t) is left-continuous for t ≥ t0 .

Lemma 6.3 (Lyapunov-like Stability Theorem [7, 16]). The zero solution of
(6.58) is uniformly asymptotically stable if there exist V ∈ V0 , ω1 , ω2 ∈ K , ψ ∈ K ∗
and H ∈ Ξ such that
(i) ω1 ( x ) ≤ V (t, x) ≤ ω2 ( x ) for (t, x) ∈ [t0 , ∞) × S(ρ ).
(ii) For all x ∈ S(ρ1 ), 0 < ρ1 ≤ ρ , and k ∈ N, V (tk , Jk (x)) ≤ ψ (V (tk− , x)).
(iii) For any solution x(t) of (6.58), V (t + s, x(t + s)) ≤ ψ −1 (V (t, x)), − τ ≤ s ≤ 0,
implies that D+V (t, x(t)) ≤ g(t)H(V (t, x(t))), where g : [t0 , ∞) → R+ is locally
integrable and ψ −1 is the inverse function of ψ .
(iv) H is nondecreasing and there exist constants λ2 ≥ λ1 > 0 and A0 > 0 such that,
for all k ∈ N and μ > 0, λ1 ≤ tk − tk−1 ≤ λ2 and
 μ  tk
du
− g(s)ds ≥ A0 . 
ψ ( μ ) H(u) tk−1

6.5.2 Main Results

In this section, we study the robust synchronization scheme for two identical chaotic
DNNs with parametric uncertainties via impulsive control. Consider the following
drive system described by:

ẋ(t) = −D [Cx(t) − (A + Δ A(t)) f (x(t)) − (B + Δ B(t)) f (x(t − τ ))] , (6.59)

6.5 Impulsive Synchronization of a Class of Chaotic Delayed Neural Networks 201

where x(t) = (x1 (t), x2 (t), . . . , xn (t))T is the state vector, f (x) = ( f1 (x1 ), f2 (x2 ),
. . . , fn (xn ))T is the nonlinear vector-value function, D = diag{d1 , d2 , . . . , dn } > 0,
C = diag{c1 , c2 , . . . , cn } > 0, and A = (ai j ) ∈ Rn×n and B = (bi j ) ∈ Rn×n are two
real constant matrices.
The parametric uncertainties Δ A(t) ∈ Rn×n and Δ B(t) ∈ Rn×n satisfy the follow-
ing assumptions.
(i) Δ A(t) = E1 F1 (t)H1 , Δ B(t) = E2 F2 (t)H2 , where Ei and Hi (i = 1, 2) are known
real constant matrices with appropriate dimensions, and the uncertain matrices
Fi (t) satisfy Fi (t) ≤ 1.
(ii) The parametric uncertainties Δ A(t) and Δ B(t) do not destroy the chaotic be-
havior of the chaotic neural networks (6.59).
The response system can be described by
9 :
x̃˙(t) = −D Cx̃(t) − (A + Δ Ã(t)) f (x̃(t)) − (B + Δ B̃(t)) f (x̃(t − τ )) ,

where x̃(t) is the state vector of the response system Δ Ã(t) = Ẽ1 F̃1 (t)H̃1 and
Δ B̃(t) = Ẽ2 F̃2 (t)H̃2 satisfy the same assumptions as Δ A(t) and Δ B(t).
At the discrete time tk , the state variables of the drive system are transmitted to
the response system as the control inputs such that the state variables of the response
system are suddenly changed at these instants. Therefore, the response system can
be written in the following form:
⎧ 7
˙
⎨ x̃(t) = −D Cx̃(t) − (A + Δ Ã(t)) f8(x̃(t))
⎪
− (B + Δ B̃(t)) f (x̃(t − τ )) , t ≥ 0, t = tk , (6.60)
⎪
⎩ +
x̃(tk ) = x̃(tk ) + (Wk − I)e(tk ), t = tk , k ∈ N,

where Wk ∈ Rn×n , I is an n × n identity matrix, and e(t) = x̃(t) − x(t) is the synchro-
nization error. Then, from (6.59) and (6.60), the following error system is obtained:
⎧ 7
⎪
⎪ ė(t) = − D Ce(t) − (A + Δ Ã(t))h(e(t), x(t))
⎪
⎪
⎨ − (B + Δ B̃(t))h(e(t − τ ), x(t − τ )) − (Δ Ã(t) − Δ A(t)) f (x(t))
8 (6.61)
⎪
⎪ − (Δ B̃(t) − Δ B(t)) f (x(t − τ )) , t ≥ 0, t = tk ,
⎪
⎪
⎩ +
e(tk ) = Wk e(tk ), t = tk , k ∈ N,

where h(e(t), x(t)) = f (e(t) + x(t)) − f (x(t)) = f (x̃(t)) − f (x(t)). Using Assump-
tion 6.2, we have
n n
h(e(t), x(t)) 2
= ∑ h2i (ei (t), xi (t)) ≤ ∑ σi2 (e2i (t)) ≤ σ 2 e(t) 2

i=1 i=1

and
2 2
h(e(t − τ ), x(t − τ )) ≤ σ 2 e(t − τ ) ,
where σ = max {σi }.
1≤i≤n
202 6 Synchronizing Chaotic Systems via Impulsive Control

Remark 6.19. Due to the boundedness of the chaotic signals, there exists a positive
constant χ such that x ≤ χ and x̃ ≤ χ . 

Remark 6.20. According to Lemma 6.3, for investigating the stability of the zero
solution of the error system (6.61), we construct such an appropriate Lyapunov-like
function V ∈ V0 that the first two conditions of Lemma 6.3 hold. Then, by calculat-
ing and estimating the Dini derivative of V along the trajectory of the error system,
one can obtain the candidates of g and H. Then, based on condition (iv), we derive
a sufficient condition to ensure robust stability of the error system. Furthermore, to
obtain an appropriate construction of V , the estimate of D+V (t, x(t)) often requires
some necessary inequality techniques such as the matrix norm inequality. 

Definition 6.8. The synchronization of systems (6.59) and (6.60) is said to have
been achieved if, for arbitrary initial conditions x0 and x̃0 , the trivial solution of the
error system (6.61) converges to a predetermined neighborhood of the origin for any
admissible parametric uncertainty. 

Here, the objective is to find conditions on the control gains Wk and the impulsive
distances δ¯k = tk − tk−1 (k ∈ N) such that the error magnitude, i.e., e , converges
to below some constant ξ , which implies that the impulsively controlled response
system (6.60) is synchronized with the drive system (6.59) for arbitrary initial con-
ditions. Similarly, for convenience of analysis, we choose a constant control gain
and the same impulsive distance, i.e., Wk = wI and δ¯k = δ¯.

Theorem 6.6. Assume that δ¯k = δ¯ and Wk = wI for k ∈ N. If the following condi-
tions hold:
(i) |w| < 1,
(ii) 2 ln |w| + vδ¯ < 0,
then the response system (6.60) is synchronized with the drive system (6.59), where
ξ > 0 is the bound of the error magnitude e and can be chosen small enough,

4 44 4  4 4 4 4
v = − 2 min {di ci } + σ d 2 4Ẽ1 4 4H̃1 4 + (1 + w−2) B + 4Ẽ2 4 4H̃2 4
¯
1≤i≤n

2χ  4 44 4 4 4 4 4
+2 A + 4 4 4 4
E1 H1 + Ẽ1 H̃1 + E2 H2 + Ẽ2 H̃2 4 4 4 4
ξ
> 0,

and d¯= max {di }.

1≤i≤n

Proof. Choose the following Lyapunov-like function:

V (e) = eT e. (6.62)

For all e ∈ S(ρ1 ), 0 < ρ1 ≤ ρ , and k ∈ N, we have

6.5 Impulsive Synchronization of a Class of Chaotic Delayed Neural Networks 203

V (tk , e(tk+ )) = [Wk e(tk )]TWk e(tk )

= eT (tk )WkTWk e(tk )
= w2 eT (tk )e(tk )
= w2V (tk , e(tk )).

Let ψ (s) = w2 s. Then, ψ ∈ K ∗ . For any solution of (6.61), if

V (t − τ , e(t − τ )) ≤ ψ −1 (V (t, e(t))),

then we have
eT (t − τ )e(t − τ ) ≤ w−2 eT (t)e(t).
For e ≥ ξ one can conclude that e ≤ e 2 /ξ . Subsequently, calculating the
Dini derivative of V (e) along the trajectory of (6.61) yields
7
D+V (t, e(t)) = 2eT (t)D −Ce(t) + (A + Δ Ã(t))h(e(t), x(t))
+ (B + Δ B̃(t))h(e(t − τ ), x(t − τ ))
8
+(Δ Ã(t) − Δ A(t)) f (x(t)) + (Δ B̃(t) − Δ B(t)) f (x(t − τ ))
n  4 4 4 4
≤ − 2 ∑i=1 di ci e2i (t) + 2σ D A + 4Ẽ1 4 4H̃1 4 e(t) 2
 4 4 4 4
+ 2σ D B + 4Ẽ2 4 4H̃2 4 e(t) e(t − τ )
 4 4 4 4
+ 2σ D E1 H1 + 4Ẽ1 4 4H̃1 4 x(t) e(t)
 4 4 4 4
+ 2σ D E2 H2 + 4Ẽ2 4 4H̃2 4 x(t − τ ) e(t)
 4 4 4 4
≤ − 2 min {di ci } e(t) 2 + 2σ d¯ A + 4Ẽ1 4 4H̃1 4 e(t) 2
1≤i≤n
 4 4 4 4  
+ σ d¯ B + 4Ẽ2 4 4H̃2 4 e(t) 2
+ e(t − τ ) 2

2σ χ d¯ 4 4 4 4
+ E1 H1 + 4Ẽ1 4 4H̃1 4 e(t) 2
ξ
2σ χ d¯ 4 4 4 4
+ E2 H2 + 4Ẽ2 4 4H̃2 4 e(t) 2
ξ

 4 4 4 4
≤ −2 min {di ci } + 2σ d¯ A + 4Ẽ1 4 4H̃1 4
1≤i≤n
  4 4 4 4
+ σ d¯ 1 + w−2 B + 4Ẽ2 4 4H̃2 4
2σ χ d¯ 4 44 4
+ E1 H1 + 4Ẽ1 4 4H̃1 4
ξ

4 4 4 4
4
+ E2 H2 + Ẽ2 H̃2 4 4 4 e(t) 2

= vV (t, e(t)).

Let g(t) = 1 and H(s) = vs. Condition (iv) of Lemma 6.3 implies that
204 6 Synchronizing Chaotic Systems via Impulsive Control
 μ  tk  μ
du du
− g(s)ds = − (tk − tk−1 )
ψ ( μ ) H(u) tk−1 w2 μ vu
2
= − ln |w| − (tk − tk−1 )
v
1
= − (2 ln |w| + vδ¯)
v
= A0
> 0.

It is easy to see that (6.62) satisfies condition (i) of Lemma 6.3. Thus, the condi-
tions of Lemma 6.3 are all satisfied. Therefore, we can conclude that the error mag-
nitude e can converge to below the constant ξ if the error started from e > ξ ,
which implies that the response system (6.60) is synchronized with the drive system
(6.59). This completes the proof. 

6.5.3 Simulations

In order to observe the synchronization behavior in two chaotic DNNs, we define

the drive system and the response system as

ẋ(t) = −D [Cx(t) − (A + Δ A(t)) f (x(t)) − (B + Δ B(t)) f (x(t − τ ))]

and
7 8
x̃˙(t) = −D Cx̃(t) − (A + Δ Ã(t)) f (x̃(t)) − (B + Δ B̃(t)) f (x̃(t − τ )) ,

where
√
1 + π /4 20 −1.3 2π /4 0.1
√
A= , B= ,
0.1 1 + π /4 0.1 −1.3 2π /4

D = C = I, f (x) = ( f1 (x1 ), f2 (x2 ))T , and fi (xi ) = (|xi + 1| − |xi − 1|)/2. Let F1 (t) =
sint 0 cost 0
−F̃1 (t) = , F2 (t) = −F̃2 (t) = , and Ei = Ẽi = Hi = H̃i =
0 sint 0 cost
0.2I (i = 1, 2).
From (6.59), (6.60), and (6.61), we get the error system as follows:
⎧ 7
⎪
⎪ ė(t) = − D Ce(t) − (A + Δ Ã(t))h(e(t), x(t))
⎪
⎪
⎨ − (B + Δ B̃(t))h(e(t − τ ), x(t − τ )) − (Δ Ã(t) − Δ A(t)) f (x(t))
8
⎪
⎪ −(Δ B̃(t) − Δ B(t)) f (x(t − τ )) , t ≥ 0, t = tk ,
⎪
⎪
⎩ +
e(tk ) = Wk e(tk ), t = tk , k ∈ N.
6.5 Impulsive Synchronization of a Class of Chaotic Delayed Neural Networks 205

0.04

0.035
ξ = 0.5 0.5
0.03

0.2
0.2
0.025

0.02 0.1
δ̄

0.1

0.015 0.05
0.05

0.01

Stable region Stable region

0.005

0
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
w

Fig. 6.9 The boundaries of stable region with different ξ

0.05

0
e1

−0.05
0 1 2 3 4 5 6 7 8 9 10

0.05

0
e2

−0.05
0 1 2 3 4 5 6 7 8 9 10

0.05
e

0
0 1 2 3 4 5 6 7 8 9 10
t (sec)

Fig. 6.10 Synchronization errors and synchronization error magnitude with δ¯ = 0.015
206 6 Synchronizing Chaotic Systems via Impulsive Control

We choose χ = 16. From Theorem 6.6, we get the stable region for different ξ ,
which is shown in Fig. 6.9. Let w = 0.2, ξ = 0.05, and σ1 = σ2 = 1. Then, the
estimate of the bound of the stable region is given by
2 2
δ¯ < − ln |w| = − ln |0.2| = 0.0177.
v 181.9792

Using these parameters, conditions in Theorem 6.6 are satisfied for δ¯ < 0.0177.
Synchronization errors and synchronization error magnitude with impulsive dis-
tance δ¯ = 0.015 are given in Fig. 6.10. As we can see, synchronization has been
achieved practically and e is smaller than ξ = 0.05. The initial conditions of the
drive and response systems are taken as x(s) = (0.1, 0.1)T and x̃(s) = (0.2, −0.2)T,
respectively, for −1 ≤ s ≤ 0.

6.6 Summary

In this chapter we have studied synchronization schemes for both identical and dif-
ferent chaotic systems. Sect. 6.2 studies the complete synchronization of a class
of chaotic systems. Sects. 6.3–6.5 mainly focus on robust synchronization, and the
synchronization error will converge to a predetermined level, which can be reduced
arbitrarily as long as some specific conditions hold. It should be pointed out that
the conditions we give for synchronization are only sufficient ones, and continued
research is desirable since less conservative conditions still need to be developed,
which are interesting and important for secure communications.

References

1. Cao J (2001) Global stability conditions for delayed CNNs. IEEE Trans Circuits Syst I
48:1330–1333
2. Gau RS, Lien CH, Hsieh JG (2007) Global exponential stability for uncertain cellular neu-
ral networks with multiple time-varying delays via LMI approach. Chaos Solitons Fractals
32:1258–1267
3. Gilli M (1993) Strange attractors in delayed cellular neural networks. IEEE Trans Circuits
Syst I 40:849–853
4. Huang H, Li HX, Zhong J (2006) Master–slave synchronization of general Lur’e systems
with time-varying delay and parameter uncertainty. Int J Bifurc Chaos 16:281–294
5. Jiang GP, Zheng WX, Chen GR (2004) Global chaos synchronization with channel time-
delay. Chaos Solitons Fractals 20:267–275
6. Lakshmikantham V, Leela S, Martynyuk AA (1990) Practical Stability of Nonlinear Systems.
World Scientific, Singapore
7. Li CD, Liao XF, Yang XF (2005) Impulsive stabilization and synchronization of a class of
chaotic delay systems. Chaos 15:043103
8. Li P, Cao JD, Wang ZD (2007) Robust impulsive synchronization of coupled delayed neural
networks with uncertainties. Physica A 373:261–272
References 207

9. Liu QS, Cao JD, Xia YS (2005) A delayed neural network for solving linear projection equa-
tions and its analysis. IEEE Trans Neural Networks 16:834–843
10. Lu HT (2002) Chaotic attractors in delayed neural networks. Phys Lett A 298:109–116
11. Ma TD, Zhang HG, Wang ZL (2007) Impulsive synchronization for unified chaotic systems
with channel time-delay and parameter uncertainty. Acta Phys Sinica 56:3796–3802
12. Ma TD, Zhang HG (2008) Impulsive synchronization of a class of unified chaotic systems
with parameter uncertainty. J Syst Simu , 18:4923–4926
13. McRae FA (1994) Practical stability of impulsive control systems. J Math Anal Appl
181:656–672
14. Moon YS, Park P, Kwon WH, Lee YS (2001) Delay-dependent robust stabilization of uncer-
tain state-delayed systems. Int J Control 74:1447–1455
15. Wang ZS, Zhang HG (2006) Global synchronization of a class of chaotic neural networks.
Acta Phys Sinica 55:2687–2693
16. Yan JR, Shen JH (1999) Impulsive stabilization and synchronization of functional differential
equations by Lyapunov-Razumikhin functions. Nonlinear Anal 37:245–255
17. Yang T (1999) Impulsive control. IEEE Trans Autom Control 44:1081–1083
18. Yang T (2001) Impulsive Control Theory. Springer, New York
19. Yang T, Chua LO (2000) Practical stability of impulsive synchronization between two nonau-
tonomous chaotic systems. Int J Bifurc Chaos 10:859-867
20. Zhang HG, Huang W, Wang ZL (2005) Impulsive control for synchronization of a class of
chaotic systems. Dyn Contin Discrete Impuls Syst B 12:153–161
21. Zhang HG, Xie YH, Liu D (2006) Synchronization of a class of delayed chaotic neural net-
works with fully unknown parameters. Dyn Contin Discrete Impuls Syst B 13:297–308
22. Zhang HG, Xie YH, Wang ZL, Zheng CD (2007) Adaptive synchronization between two
different chaotic neural networks with time delay. IEEE Trans Neural Networks 18:1841–
1845
23. Zou F, Nossek JA (1993) Bifurcation and chaos in cellular neural networks. IEEE Trans Cir-
cuits Syst I 40:166–173
Chapter 7
Synchronization of Chaotic Systems with Time
Delay

Abstract In many physical, industrial, and engineering systems, delays occur due
to the finite capabilities of information processing and data transmission among
various parts of the system. Delays could arise as well from inherent physical phe-
nomena like mass transport flow or recycling. Also, they could be by-products of
computational delays or could intentionally be introduced for some design consid-
erations. Such delays could be constant or time varying, known or unknown, deter-
ministic or stochastic depending on the system under consideration. In all of these
cases, the time-delay factors have counteracting effects on the system behavior and
most of the time lead to poor performance. Therefore, the subject of time-delay sys-
tems has been investigated as functional differential equations over the past three
decades. In this chapter, we study how to synchronize chaotic systems when time
delay exists and the synchronized systems have different structures. We first develop
synchronization methods for a class of delayed chaotic systems when the drive sys-
tem and the response system have the same structure but different parameters. After
that, the problem of synchronizing different chaotic systems is studied. Some con-
crete examples are presented to show how to design the controller. Based on that,
a more general case, synchronizing two different delayed chaotic neural networks
with known and unknown parameters, is considered.

7.1 Introduction

In the last two chapters, we have studied the synchronization of chaotic systems
which are described by ordinary differential equations. However, in many physical,
industrial, and engineering systems, delays occur due to the finite capabilities of
information processing and data transmission among various parts of the system.
Delays could arise as well from inherent physical phenomena like mass transport
flow or recycling. Also, they could be by-products of computational delays or could
intentionally be introduced for some design considerations. Such delays could be
constant or time varying, known or unknown, deterministic or stochastic depending

209
210 7 Synchronization of Chaotic Systems with Time Delay

on the system under consideration. In all of these cases, the time-delay factors have,
by and large, counteracting effects on the system behavior and most of the time
lead to poor performance. Therefore, the subject of time-delay systems has been
investigated as functional differential equations over the past three decades. This has
occupied a separate discipline in mathematical sciences falling between differential
and difference equations.
It is well known that chaos can easily be generated in a time-delay system. When
synchronizing two time-delay chaotic systems, the values of delays and parameters
of the systems are often unknown in advance. In some special cases, the structure
or parameters of the drive system are even unknown in advance. Therefore, in this
chapter, we study the synchronization of chaotic systems with time delays and with
different structures. In Sect. 7.2, we first develop the synchronization methods for
a class of delayed chaotic systems when the drive system and the response system
have the same structure but different parameters. After that, in Sect. 7.3, the problem
of synchronizing different chaotic systems is studied. Some examples are presented
to show how to design the controller. Based on that, in Sect. 7.4, a more general
case, synchronizing two different delayed chaotic neural networks with known and
unknown parameters, is considered.

7.2 Adaptive Synchronization of a Class of Delayed Chaotic

Systems

In this section, the synchronization problem of two identical delayed chaotic sys-
tems with unknown parameters is studied. We take the delayed Chen system and
the delayed Lorenz system as examples to show the effectiveness of the proposed
method.

7.2.1 Adaptive Synchronization of the Delayed Chen Chaotic

Systems

The drive delayed Chen chaotic system [14] in this section is described by the fol-
lowing differential equations:
⎧
⎨ ẋ1 (t) = a(x2 (t) − x1(t)),
ẋ2 (t) = (c − a)x1(t) − x1(t)x3 (t) + (c − d) x2 (t) + dx2(t − τ ), (7.1)
⎩
ẋ3 (t) = x1 (t)x2 (t) − bx3(t),

where a, b, c, and d are the unknown parameters of system (7.1), d < η , and η is
a positive constant. The chaotic behavior of the drive system (7.1) is shown in Fig.
7.1, when the initial values of the parameters of the drive system (7.1) are selected
as a = 35, b = 3, c = 28, d = 3, and τ = 0.1.
7.2 Adaptive Synchronization of a Class of Delayed Chaotic Systems 211

40

30

20
x3

10

0
40
20 −40
−20
0
0
x1 −20
20 x2
−40 40

Fig. 7.1 Chaotic behavior of the delayed Chen chaotic system

The response delayed Chen chaotic system is described by the following equa-
tions:
⎧
⎨ ẏ1 (t) = a(y2 (t) − y1(t)) + u1,
ẏ (t) = (c − a)y1(t) − y1 (t)y3 (t) + (c − d) y2 (t) + dy2(t − τ ) + u2, (7.2)
⎩ 2
ẏ3 (t) = y1 (t)y2 (t) − by3(t) + u3,

where u1 , u2 , and u3 denote the designed controllers that will realize the synchro-
nization of system (7.1) and system (7.2).
Let us define the synchronization error signal as ei (t) = yi (t)− xi (t) and eiτ (t) =
yi (t − τ ) − xi (t − τ ) (i = 1, 2, 3). The fact that ei (t) → 0 as t → ∞ means that the
drive system and the response system are synchronized. The goal of the controllers’
design is to obtain ui (t) so that lim ei (t) = lim yi (t) − xi (t) = 0, where ·
t→∞ t→∞
denotes the Euclidean norm.
Therefore, the error system between (7.1) and (7.2) can be expressed as
⎧
⎨ ė1 = a(e2 − e1 ) + u1,
ė2 = (c − a)e1 + (c − d) e2 − e1 e3 − e3 x1 − e1 x3 + de2τ + u2 , (7.3)
⎩
ė3 = e1 e2 + e1 x2 + e2 x1 − be3 + u3.

Then, we have the following theorem.

Theorem 7.1 ([16]). By the following controller:
212 7 Synchronization of Chaotic Systems with Time Delay
⎧
⎨ u1 = â(e1 − e2) − e1 ,  
u2 = e1 x3 − (ĉ− â) e1 + dˆ − ĉ − k e2 , (7.4)
⎩
u3 = −e1 x2 + b̂ − 1 e3

and the parameter update law

⎧
⎪
⎪ â˙ = −e21 ,
⎪
⎨˙
b̂ = −e23 ,
(7.5)
⎪
⎪ ĉ˙ = e1 e2 + e22 ,
⎪
⎩ ˙ˆ
d = −e22 ,

the drive system (7.1) and the response system (7.2) will be globally asymptotically
synchronized. Here, â, b̂, ĉ, and dˆ are the estimated values of unknown parameters
a, b, c, and d, respectively, and k > (η 2 + 1)/2.

Proof. Substituting (7.4) into (7.3), the error system (7.3) can be rewritten as
⎧
⎨ ė1 = (a − â) (e2 − e1 ) − e1 , 
ė2 = (c − ĉ − a + â)e ˆ
1 + c − ĉ − d + d − k e2 − e1 e3 − e3 x1 + de2τ , (7.6)
⎩ 
ė3 = e1 e2 + e2x1 − b − b̂ e3 − e3 .

Let ea = a − â, eb = b − b̂, ec = c − ĉ, and ed = d − dˆ be the estimated errors

of the unknown parameters a, b, c, and d, respectively. Now construct a Lyapunov
function in the form of
 t
1 2
V (e) = e + e22 + e23 + e2a + e2b + e2c + e2d + e22 (s) ds .
2 1 t−τ

Then, the time derivative of V (e) along the trajectory of the error dynamical system
(7.6) is as follows:
1 1
V̇ (e) = e1 ė1 + e2ė2 + e3ė3 + eaėa + ebėb + ec ėc + ed ėd + e22 − e22τ
2 2
= e1 (ea (e2 − e1 ) − e1) + e3 (e1 e2 + e2 x1 − eb e3 − e3 )
+ e2 ((ec − ea)e1 + (ec − ed − k) e2 − e1 e3 − e3 x1 + de2τ )
  1 1
− ea â˙ + ebb̂˙ + ecĉ˙ + ed d˙ˆ + e22 − e22τ . (7.7)
2 2
Substituting (7.5) into (7.7), we obtain
1 1
V̇ (e) = −e21 − e23 − ke22 + e2 de2τ + e22 − e22τ
2 2
d2 1 1 1
≤ −e21 − e23 − ke22 + e22 + e22τ + e22 − e22τ
2 2 2 2
η 2 1
≤ −e21 − e23 − k − − e2 .
2 2 2
7.2 Adaptive Synchronization of a Class of Delayed Chaotic Systems 213
 
Setting Q = k − (η 2 + 1)/2, we obtain V̇ (e) ≤ − e21 + e23 + Qe22 .
When k > (η 2 + 1)/2, i.e., Q > 0, we get V̇ (e) ≤ 0, which implies that V̇ (e) < 0
for all e(t) = 0. Since V (e) is positive definite and V̇ (e) is negative semidefinite, it
follows that e1 ∈ L∞ , e2 ∈ L∞ , and e3 ∈ L∞ . From the fact that
 t  t
−(e21 + Qe22 + e23 )ds ≤ − V̇ ds ≤ V (e1 (0), e2 (0), e3 (0)),
0 0

we can easily show that e1 ∈ L2 , e2 ∈ L2 , and e3 ∈ L2 . From (7.6) we have ė1 ∈ L∞ ,

ė2 ∈ L∞ , and ė3 ∈ L∞ . Thus, by Barbalat’s lemma, we have

lim ei (t) = lim yi (t) − xi (t) = 0.

t→∞ t→∞

This means that the proposed controller (7.4) and (7.5) can globally asymptotically
synchronize the system (7.1) and the system (7.2). This completes the proof. 

The following illustrative example is used to demonstrate the effectiveness of the

above method.
According to Theorem 7.1, the initial condition of the drive delayed Chen chaotic
system is x0 = (5, 5, 15)T , and the initial condition of the response delayed Chen
chaotic system is y0 = (3, 3, 12)T . The initial values for the estimation of parameters
ˆ = 4. The synchro-
are selected as follows: â(0) = 32, b̂(0) = 4, ĉ(0) = 25, and d(0)

Fig. 7.2 The synchronization error curves of the drive system (7.1) and the response system (7.2)
214 7 Synchronization of Chaotic Systems with Time Delay

Fig. 7.3 The curve of the adaptive parameter â(t)

Fig. 7.4 The curve of the adaptive parameter b̂(t)

7.2 Adaptive Synchronization of a Class of Delayed Chaotic Systems 215

Fig. 7.5 The curve of the adaptive parameter ĉ(t)

ˆ
Fig. 7.6 The curve of the adaptive parameter d(t)
216 7 Synchronization of Chaotic Systems with Time Delay

nization error curves of the drive system (7.1) and the response system (7.2) with the
controller u(t) = (u1 (t), u2 (t), u3 (t))T are shown in Fig. 7.2. The curves of the esti-
mated parameters â, b̂, ĉ, and dˆ are shown in Figs. 7.3–7.6, respectively. From simu-
lation results we can see that the synchronization errors converge asymptotically to
zero, i.e., the trajectory of the response delayed Chen chaotic system can synchro-
nize asymptotically with the trajectory of the drive delayed Chen chaotic system.
The estimated parameters converge asymptotically to some constants, which shows
the effectiveness of the adaptive synchronization scheme.

7.2.2 Adaptive Synchronization of the Delayed Lorenz Systems

The drive delayed Lorenz chaotic system [13] is described by the following differ-
ential equations:
⎧
⎨ ẋ1 (t) = a(x2 (t) − x1 (t)),
ẋ2 (t) = cx1 (t) − x1(t)x3 (t) + (d − 1) x2 (t) − dx2(t − τ ), (7.8)
⎩
ẋ3 (t) = x1 (t)x2 (t) − bx3(t),

where a, b, c, and d are the unknown parameters of the system (7.8), and d < η .
The chaotic behavior of the drive system (7.8) is shown in Fig. 7.7, when the initial
values of the parameters of the drive system (7.8) are selected as a = 10, b = 8/3,
c = 5, d = 6.5, and τ = 0.5.
The response delayed Lorenz chaotic system is described by the following equa-
tions:
⎧
⎨ ẏ1 (t) = a(y2 (t) − y1 (t)) + u1,
ẏ2 (t) = cy1 (t) − y1(t)y3 (t) + (d − 1) y2 (t) − dy2(t − τ ) + u2, (7.9)
⎩
ẏ3 (t) = y1 (t)y2 (t) − by3(t) + u3,

where u1 , u2 , and u3 denote the designed controllers that will realize the synchro-
nization between system (7.1) and system (7.2).
Therefore, the error system between (7.8) and (7.9) can be expressed as
⎧
⎨ ė1 = a(e2 − e1) + u1,
ė2 = ce1 + (d − 1) e2 − e1e3 − e3x1 − e1x3 − de2τ + u2, (7.10)
⎩
ė3 = e1 e2 + e1x2 + e2x1 − be3 + u3 .

Then, we have the following theorem.

Theorem 7.2 ([16]). By the controller

⎧
⎨ u1 = â(e1 − e2 ) − e1 , 
u2 = e1 x3 − ĉe1 − dˆ+ k e2 , (7.11)
⎩
u3 = −e1 x2 + b̂ − 1 e3
7.2 Adaptive Synchronization of a Class of Delayed Chaotic Systems 217

40

30

20
x3

10

0
−20 40
−10 20
0 0
x1 10 −20 x2
20 −40

Fig. 7.7 Chaotic behavior of the delayed Lorenz chaotic system

and the parameter update law

⎧
⎪ ˙ 2
⎪ â = e1 e2 − e1 ,
⎪
⎨˙ 2
b̂ = −e3 ,
(7.12)
⎪
⎪ ĉ˙ = e1 e2 ,
⎪
⎩ d˙ˆ = e2 ,
2

the drive system (7.8) and the response system (7.9) will be globally asymptotically
synchronized. Here, â, b̂, ĉ, and dˆ are the estimated values of unknown parameters
a, b, c, and d, respectively, and k > (η 2 − 1)/2.

Proof. Substituting (7.11) into (7.10), the error system (7.10) can be rewritten by
⎧
⎨ ė1 = (a − â) (e2 − e1) − e1 , 
ė2 = (c − ĉ)e1 + d− dˆ −k − 1 e2 − e1e3 − e3x1 − de2τ , (7.13)
⎩
ė3 = e1 e2 + e2 x1 − b − b̂ e3 − e3 .

Let ea = a − â, eb = b − b̂, ec = c − ĉ, and ed = d − dˆ be the estimated errors of

unknown parameters a, b, c, and d, respectively. Now construct a Lyapunov function
in the form of
218 7 Synchronization of Chaotic Systems with Time Delay
 t
1 2
V (e) = e + e22 + e23 + e2a + e2b + e2c + e2d + e22 (s) ds .
2 1 t−τ

Then, the time derivative of V (e) along the trajectory of the error dynamical system
(7.13) is as follows:
1 1
V̇ (e) = e1 ė1 + e2ė2 + e3ė3 + eaėa + ebėb + ec ėc + ed ėd + e22 − e22τ
2 2
= e1 (ea (e2 − e1 ) − e1) + e3 (e1 e2 + e2 x1 − eb e3 − e3 )
+ e2 (ec e1 + (ed − k − 1) e2 − e1 e3 − e3 x1 − de2τ )
  1 1
− ea â˙ + ebb̂˙ + ecĉ˙ + ed d˙ˆ + e22 − e22τ . (7.14)
2 2
Substituting (7.12) into (7.14), we obtain
1 1
V̇ (e) = − e21 − e23 − ke22 − e2de2τ − e22 − e22τ
2 2
d 2 1 1 1
≤ − e21 − e23 − ke22 + e22 + e22τ − e22 − e22τ
2 2 2 2
η2 1 2
≤ − e21 − e23 − k − + e . (7.15)
2 2 2
 
Setting Q = k − (η 2 − 1)/2, we obtain V̇ (e) ≤ − e21 + e23 + Qe22 .
When k > (η 2 − 1)/2, i.e., Q > 0, we get V̇ (e) ≤ 0, which implies that V̇ (e) < 0
for all e(t) = 0. Since V (e) is positive definite and V̇ (e) is negative semidefinite, it
follows that e1 ∈ L∞ , e2 ∈ L∞ , and e3 ∈ L∞ . From the fact that
 t  t
−(e21 + Qe22 + e23 )ds ≤ − V̇ ds ≤ V (e1 (0), e2 (0), e3 (0)),
0 0

we can easily show that e1 ∈ L2 , e2 ∈ L2 , and e3 ∈ L2 . From (7.10) we have

ė1 ∈ L∞ , ė2 ∈ L∞ , and ė3 ∈ L∞ . Thus, by Barbalat’s lemma, we have lim ei (t) =
t→∞
lim yi (t) − xi (t) = 0. This means that the proposed controller (7.11) with param-
t→∞
eter update law (7.12) can globally asymptotically synchronize the system (7.8) and
the system (7.9). This completes the proof. 

The following illustrative example is used to demonstrate the effectiveness of the

above method.
According to Theorem 7.2, the initial condition of the drive delayed Lorenz
chaotic system is x0 = (−0.1, −0.1, 0.2)T , and the initial condition of the response
delayed Lorenz chaotic system is y0 = (3.1, −0.2, 0.3)T . The initial values for the
estimation of parameters are selected as follows: â(0) = 9, b̂(0) = 3, ĉ(0) = 6, and
ˆ = 7. The synchronization error curves of the drive system (7.8) and the re-
d(0)
sponse system (7.9) with the controller u(t) = (u1 (t), u2 (t), u3 (t))T are shown in
Fig. 7.8. The curves of the estimated parameters â, b̂, ĉ, and dˆ are shown in Figs.
7.9–7.12, respectively. From simulation results we can see that the synchronization
7.2 Adaptive Synchronization of a Class of Delayed Chaotic Systems 219

Fig. 7.8 The synchronization error curves of the drive system (7.8) and the response system (7.9)

Fig. 7.9 The curve of the adaptive parameter â(t)

220 7 Synchronization of Chaotic Systems with Time Delay

Fig. 7.10 The curve of the adaptive parameter b̂(t)

Fig. 7.11 The curve of the adaptive parameter ĉ(t)

7.3 Adaptive Synchronization Between Two Different Delayed Chaotic Systems 221

ˆ
Fig. 7.12 The curve of the adaptive parameter d(t)

errors converge asymptotically to zero, i.e., the trajectory of the response delayed
Lorenz chaotic system can synchronize asymptotically with the trajectory of the
drive delayed Lorenz chaotic system. The estimated parameters converge asymptot-
ically to some constants, which shows the effectiveness of the adaptive synchroniza-
tion scheme.

7.3 Adaptive Synchronization Between Two Different Delayed

Chaotic Systems

In the last section, the underlying assumption is that the drive system and the re-
sponse system have identical dynamic structures. In this section, we consider the
problem of synchronizing two different delayed chaotic systems with unknown pa-
rameters. Three examples are used to illustrate the design procedure of the proposed
synchronization methods.
222 7 Synchronization of Chaotic Systems with Time Delay

7.3.1 Adaptive Synchronization Between Delayed Chen Chaotic

Systems and Delayed Lorenz Chaotic Systems

The drive delayed Chen system in this section is described by differential equations
(7.1).
The response delayed Lorenz system is described by the following equations:
⎧
⎨ ẏ1 (t) = α (y2 (t) − y1(t)) + u1,
ẏ2 (t) = γ y1 (t) − y1 (t)y3 (t) + (θ − 1) y2 (t) − θ y2 (t − τ ) + u2, (7.16)
⎩
ẏ3 (t) = y1 (t)y2 (t) − β y3(t) + u3,

where α , β , γ , and θ are the unknown parameters of system (7.16) and u1 , u2 , and
u3 denote the designed controllers that will realize the synchronization of system
(7.1) and system (7.16).
Therefore, the error system between (7.1) and (7.16) can be expressed as
⎧
⎪ ė1 = α (y2 − y1 ) − a(x2 − x1) + u1 ,
⎪
⎪
⎨ ė = γ y − (c − a)x − e e − e x − e x
2 1 1 1 3 3 1 1 3
(7.17)
⎪
⎪ + ( θ − 1) y 2 − (c − d) x 2 − θ y 2 τ − dx2τ + u2,
⎪
⎩
ė3 = e1 e2 + e1 x2 + e2 x1 − β y3 + bx3 + u3 .

Thus, we have the following theorem.

Theorem 7.3 ([16]). By the following controller:

⎧
⎪
⎪ u1 = α̂ (e1 − e2 ) + (â − α̂ )(x2 − x1 ) − e1,
⎪
⎪
⎨ u2 = (ĉ − â)x1 + (ĉ − d)xˆ 2 − γ̂ y1 + (1 − θ̂ )y2
(7.18)
⎪
⎪ + e1 x3 + θ̂ y2τ + dxˆ 2 τ − e2 ,
⎪
⎪
⎩
u3 =(β̂ − 1)e3 + (β̂ − b̂)x3 − e1 x2

and the parameter update law

⎧
⎪
⎪ â˙ = e2 x1 − e1(x2 − x1 ),
⎪
⎪˙
⎪
⎪ b̂ = e3 x3 ,
⎪
⎪
⎪
⎪ ˙ = −e2 x1 − e x2 ,
ĉ
⎪
⎪ 2
⎨ d˙ˆ = e x − e x ,
2 2 2 2τ
(7.19)
⎪
⎪ α̂˙ = e1 (y2 − y1),
⎪
⎪ ˙
⎪
⎪ β̂ = −e3 y3 ,
⎪
⎪
⎪
⎪ γ̂˙ = e2 y1 ,
⎪
⎪
⎩˙
θ̂ = e2 y2 − e2 y2τ ,

the drive system (7.1) and the response system (7.16) will be globally asymptotically
synchronized. Here, â, b̂, ĉ, d,ˆ α̂ , β̂ , γ̂ , and θ̂ are the estimated values of unknown
parameters a, b, c, d, α , β , γ , and θ , respectively.
7.3 Adaptive Synchronization Between Two Different Delayed Chaotic Systems 223

Proof. Substituting (7.18) into (7.17), the error system (7.17) can be rewritten as
⎧
⎪
⎪ ė1 = (α − α̂ ) (y2 − y1 ) − (a − â) (x2 − x1 ) − e1 ,
⎪
⎨ ė2 = (γ − γ̂ ) y1 − (c − ĉ − a + â)x1 − c − ĉ − d + dˆ x2
⎪
      (7.20)
⎪
⎪ + θ − θ̂ y2 − θ − θ̂ y2τ − d − dˆ x2τ − e2 − e1 e3 − e3 x1 ,
⎪
⎪  
⎩
ė3 = e1 e2 + e2 x1 − (β − β̂ )y3 + b − b̂ x3 − e3 .

Let ea = a − â, eb = b − b̂, ec = c − ĉ, ed = d − d,ˆ eα = α − α̂ , eβ = β − β̂ ,

eγ = γ − γ̂ , and eθ = θ − θ̂ be the estimated errors of unknown parameters a, b, c,
d, α , β , γ , and θ , respectively. Now construct a Lyapunov function in the form of
1 2 
V (e) = e1 + e22 + e23 + e2a + e2b + e2c + e2d + e2α + e2β + e2γ + e2θ .
2
Then, the time derivative of V (e) along the trajectory of the error dynamical system
(7.20) is as follows:

V̇ (e) = e1 ė1 + e2ė2 + e3ė3 + eaėa + ebėb + ec ėc

+ ed ėd + eα ėα + eβ ėβ + eγ ėγ + eθ ėθ
= e1 (eα (y2 − y1 ) − ea(x2 − x1 ) − e1)
 
+ e3 e1 e2 + e2 x1 − eβ y3 + eb x3 − e3
 
+ e2 eγ y1 − (ec − ea )x1 − (ec − ed ) x2
+ e2 (eθ y2 − eθ y2τ − ed x2τ − e2 − e1 e3 − e3 x1 )
 
˙
− ea â˙ + ebb̂˙ + ec ĉ˙ + ed dˆ˙ + eα α̂˙ + eβ β̂ + eγ γ̂˙ + eθ θ̂˙
= e1 eα (y2 − y1) − ea e1 (x2 − x1 ) − e21
+ e3 e1 e2 + e3 e2 x1 − eβ e3 y3 + eb e3 x3 − e23
+ e2 eγ y1 − e2 ec x1 + e2 ea x1 − ec e2 x2
+ ed e2 x2 + eθ e2 y2 − eθ e2 y2 τ − ed e2 x2 τ
 
˙
− ea â˙ + ebb̂˙ + ec ĉ˙ + ed d˙ˆ + eα α̂˙ + eβ β̂ + eγ γ̂˙ + eθ θ̂˙
− e22 − e1e3 e2 − e2 e3 x1 . (7.21)

Substituting (7.19) into (7.21), we obtain

 
V̇ (e) = − e21 + e22 + e23 ≤ 0,

which implies that V̇ (e) < 0 for all e(t) = 0. It is clear that e1 ∈ L∞ , e2 ∈ L∞ , and
e3 ∈ L∞ . From the fact
 t  t
−(e21 + e22 + e23 )ds ≤ − V̇ ds ≤ V (e1 (0), e2 (0), e3 (0)),
0 0
224 7 Synchronization of Chaotic Systems with Time Delay

we can easily show that e1 ∈ L2 , e2 ∈ L2 , and e3 ∈ L2 . From (7.17) we have

ė1 ∈ L∞ , ė2 ∈ L∞ , and ė3 ∈ L∞ . Thus, by Barbalat’s lemma, we have lim ei (t) =
t→∞
lim yi (t) − xi (t) = 0. This means that the proposed controller (7.18) and (7.19)
t→∞
can globally asymptotically synchronize the drive system (7.1) and the response
system (7.16). This completes the proof. 

The following illustrative example is used to demonstrate the effectiveness of the

above method.
According to Theorem 7.3, the initial condition of the drive delayed Chen chaotic
system is x0 = (5, 5, 15)T , and the initial condition of the response delayed Lorenz
chaotic system is y0 = (−0.1, −0.1, 0.2)T . The ‘unknown’ parameters are chosen
as α = 10, β = 8/3, γ = 5, and θ = 6.5. The initial values of the estimation of
parameters are selected as follows: â(0) = 32, b̂(0) = 4, ĉ(0) = 25, d(0) ˆ = 2, α̂ (0) =
9, β̂ (0) = 3, γ̂ (0) = 4, and θ̂ (0) = 5. The synchronization error curves of the drive
system (7.1) and the response system (7.16) are shown in Fig. 7.13. The curves of
the estimated parameters â, b̂, ĉ, d,ˆ α̂ , β̂ , γ̂ , and θ̂ are shown in Figs. 7.14–7.21,
respectively.
From simulation results we can see that the synchronization errors converge
asymptotically to zero, i.e., the trajectory of the response delayed Lorenz chaotic
system can synchronize asymptotically with the trajectory of the drive delayed Chen

Fig. 7.13 The synchronization error curves of the drive system (7.1) and the response system
(7.16)
7.3 Adaptive Synchronization Between Two Different Delayed Chaotic Systems 225

Fig. 7.14 The curve of the adaptive parameter â(t)

Fig. 7.15 The curve of the adaptive parameter b̂(t)

226 7 Synchronization of Chaotic Systems with Time Delay

Fig. 7.16 The curve of the adaptive parameter ĉ(t)

ˆ
Fig. 7.17 The curve of the adaptive parameter d(t)
7.3 Adaptive Synchronization Between Two Different Delayed Chaotic Systems 227

Fig. 7.18 The curve of the adaptive parameter α̂ (t)

Fig. 7.19 The curve of the adaptive parameter β̂ (t)

228 7 Synchronization of Chaotic Systems with Time Delay

Fig. 7.20 The curve of the adaptive parameter γ̂ (t)

Fig. 7.21 The curve of the adaptive parameter θ̂ (t)

7.3 Adaptive Synchronization Between Two Different Delayed Chaotic Systems 229

chaotic system. The estimated parameters converge asymptotically to some con-

stants, which shows the effectiveness of the adaptive synchronization scheme.

7.3.2 Adaptive Synchronization Between Two Different Delayed

Ikeda Chaotic Systems

The drive delayed Ikeda chaotic system [4] in this section is described by the fol-
lowing differential equation:

ẋ (t) = −cx (t) + a f (x (t − τ )) , (7.22)

where τ > 0 denotes the bounded delay, a > 0, and c > 0. We assume that c is the un-
known parameter of system (7.22), and a is the known parameter of system (7.22).
The initial condition of system (7.22) is given by x(t) = ϕx (t) ∈ C([−τ , 0], R),
where C ([−τ , 0] , R) denotes the set of all continuous functions from [−τ , 0] to R.
The function f (x) is a one-dimensional and continuous nonlinear function.
The response delayed Ikeda chaotic system is described by the following equa-
tion:
ẏ (t) = −dy (t) + bg (y (t − τ )) + u (t) , (7.23)
where τ > 0 denotes the bounded delay, b > 0, and d > 0. We assume that d is
the unknown parameters of the system (7.23), and b is the known parameter of the
system (7.23). The initial conditions of system (7.23) are given by y(t) = ϕy (t) ∈
C([−τ , 0], R), where C ([−τ , 0] , R) denotes the set of all continuous functions from
[−τ , 0] to R. The function g (y) is a one-dimensional and continuous nonlinear func-
tion, and u(t) denotes the designed controller that will achieve the goal of synchro-
nizing system (7.22) and system (7.23), i.e., lim e (t) = lim y (t) − x (t) = 0.
t→∞ t→∞
For this problem, we have the following theorem.
Theorem 7.4 ([16]). By the following controller:
ˆ
u (t) = −ĉx(t) + a f (x (t − τ )) + dy(t) − bg (y (t − τ )) − ke (t) (7.24)

and the parameter update law

ĉ˙ = eT (t)x(t),
(7.25)
d˙ˆ = −eT (t)y(t),

the drive system (7.22) and the response system (7.23) will be globally asymptot-
ically synchronized. Here, k > 0 and ĉ and dˆ are the estimated values of unknown
parameters c and d, respectively.
Proof. From (7.22)–(7.24), we get the error system as follows:
 
ė (t) = − d − dˆ y (t) + (c − ĉ) x (t) − ke (t) . (7.26)
230 7 Synchronization of Chaotic Systems with Time Delay
 T
Define ε := eT (t), ĉ, dˆ . If a Lyapunov function candidate is chosen as

1 T  2 
V (ε ) = e (t) e (t) + (c − ĉ)2 + d − dˆ ,
2
then the time derivative of V (ε ) along the trajectory of the error system (7.26) is as
follows:
   
V̇ (ε ) = eT (t) ė (t) + (c − ĉ) −ĉ˙ + d − dˆ (−dˆ˙ )
   
= eT (t) − d − dˆ y (t) + (c − ĉ) x (t)
 
− (c − ĉ) ĉ˙ − d − dˆ dˆ˙ − keT (t) e (t) . (7.27)

Substituting (7.25) into (7.27), we obtain

V̇ (ε ) = −keT (t) e (t) ≤ 0,

which implies that V̇ (ε ) < 0 for all e(t) = 0. It is clear that e(t) ∈ L∞ , ĉ ∈ L∞ , and
dˆ ∈ L∞ . From the fact that
 t
[V (e(0)) − V (e(t))] V (e(0))
e(t) 2 ds = ≤ ,
0 k k
we can easily show that e(t) ∈ L2 . From (7.26) we have ė(t) ∈ L∞ . Thus, by Bar-
balat’s lemma, we have lim e(t) = 0. This means that the proposed controller (7.24)
t→∞
can globally asymptotically synchronize the system (7.22) and the system (7.23).
This completes the proof. 
The following illustrative example is used to demonstrate the effectiveness of the
above method.
Assume that the following delayed Ikeda chaotic system [7] is the drive system:

ẋ (t) = −cx (t) + a f (x (t − τ )) , (7.28)

where the activation function f (x (t − τ )) = sin (x (t − τ )), and the following de-
layed Ikeda chaotic system [6] is the response system:

ẏ (t) = −dy (t) + bg (y (t − τ )) , (7.29)

where the activation function g (y (t − τ )) = cos (my (t − τ )).

The initial conditions of the drive system (7.28) and the response system (7.29)
are taken as x(s) = 1 and y(s) = 1.5 for −2 ≤ s ≤ 0, respectively. The parameters of
the drive system (7.28) and the response system (7.29) are selected as c = 1, a = 20,
d = 1, b = 2, m = 2, and τ = 2. The state curve of the drive system (7.28) is shown
in Fig. 7.22, and the state curve of the response system (7.29) is shown in Fig. 7.23.
According to Theorem 7.4, the initial values of the drive system and the response
system are x (0) = 0.5 and y (0) = 1. We choose ĉ(0) = 1.5, d(0) ˆ = 1.1, and k = 1.
The synchronization error curve of the drive system (7.28) and the response sys-
7.3 Adaptive Synchronization Between Two Different Delayed Chaotic Systems 231

15

10

5
x

−5

−10
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 7.22 The state curve of the drive chaotic system (7.28)

1.5

0.5

0
y

−0.5

−1

−1.5

−2
0 10 20 30 40 50 60 70
t (sec)

Fig. 7.23 The state curve of the response chaotic system (7.29)
232 7 Synchronization of Chaotic Systems with Time Delay

0.6

0.4

0.2

0
e

−0.2

−0.4

−0.6

−0.8
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Fig. 7.24 The synchronization error curve of the drive system (7.28) and the response system
(7.29)

1.8

1.7

1.6

1.5

1.4
â

1.3

1.2

1.1

0.9
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Fig. 7.25 The curve of the adaptive parameter ĉ(t)

7.3 Adaptive Synchronization Between Two Different Delayed Chaotic Systems 233

1.6

1.5

1.4

1.3
b̂

1.2

1.1

0.9
0 2 4 6 8 10 12 14 16 18 20
t (sec)

ˆ
Fig. 7.26 The curve of the adaptive parameter d(t)

tem (7.29) is shown in Fig. 7.24. The curves of the estimated parameters ĉ and
dˆ are shown in Figs. 7.25 and 7.26, respectively. From simulation results we can
see that the synchronization errors converge asymptotically to zero, and the esti-
mated parameters converge asymptotically to some constants, i.e., the trajectory of
the response system can synchronize asymptotically with the trajectory of the drive
system.

7.3.3 Adaptive Synchronization Between Two Different Delayed

Liao Chaotic Systems

The drive delayed Liao chaotic system [3] in this section is described by the follow-
ing differential equation:

ẋ(t) = −α x(t) + a f (x (t) − bx (t − τ ) + c) , (7.30)

where τ > 0 denotes the bounded delay, α , a, and b are the parameters of the sys-
tem (7.30), and c is a constant. We assume that α is the unknown parameter of
system (7.30), and a > 0 and b > 0 are the known parameters of the system. The
initial conditions of system (7.30) are given by x(t) = ϕx (t) ∈ C([−τ , 0], R), where
C ([−τ , 0] , R) denotes the set of all continuous functions from [−τ , 0] to R. The
function f (x) is a one-dimensional and continuous nonlinear function.
234 7 Synchronization of Chaotic Systems with Time Delay

The response delayed Liao chaotic system is described by the following equation:

ẏ(t) = −β y(t) + mg (y (t) − ny (t − τ ) + c) + u(t), (7.31)

where τ > 0 denotes the bounded delay and β , m, and n are the parameters of
system (7.31). We assume that β is the unknown parameter of system (7.31), and
m > 0 and n > 0 are the known parameters of the system. The initial conditions of
system (7.31) are given by y(t) = ϕy (t) ∈ C([−ρ , 0], R). The function g (y) is a
one-dimensional and continuous nonlinear function, and u(t) denotes the designed
controller that will realize the synchronization of system (7.30) and system (7.31),
i.e., lim e (t) = lim y (t) − x (t) = 0.
t→∞ t→∞
Theorem 7.5 ([16]). By the following controller:

u (t) = − α̂ x(t) + a f (x (t) − bx (t − τ ) + c)

+ β̂ y(t) − mg (y (t) − ny (t − τ ) + c) − ke (t) (7.32)

and the parameter update law

α̂˙ = eT (t)x(t),
˙ (7.33)
β̂ = −eT (t)y(t),

the drive system (7.30) and the response system (7.31) will be globally asymptoti-
cally synchronized. Here, k > 0 and α̂ and β̂ are the estimated values of unknown
parameters α and β , respectively.
Proof. From (7.30)–(7.32), we get the error system as follows:
 
ė (t) = − β − β̂ y(t) + (α − α̂ ) x(t) − ke (t) . (7.34)

Define

ε := (eT (t) , α̂ , β̂ )T .

If a Lyapunov function candidate is chosen as

1 T  2
V (ε ) = e (t) e (t) + (α − α̂ )2 + β − β̂ ,
2

then the time derivative of V (ε ) along the trajectory of the error system (7.34) is as
follows:
    
˙
V̇ (ε ) = eT (t) ė (t) + (α − α̂ ) −α̂˙ + β − β̂ −β̂
   
= eT (t) − β − β̂ y(t) + (α − α̂ ) x(t)
 
˙
− (α − α̂ ) α̂˙ − β − β̂ β̂ − keT (t) e (t) . (7.35)
7.3 Adaptive Synchronization Between Two Different Delayed Chaotic Systems 235

Substituting (7.33) into (7.35), we obtain

V̇ (ε ) = −keT (t) e (t) ≤ 0, (7.36)

which implies that V̇ (ε ) < 0 for all e(t) = 0. It is clear that e(t) ∈ L∞ , α̂ ∈ L∞ , and
β̂ ∈ L∞ . From the fact that
 t
2 [V (e(0)) − V (e(t))] V (e(0))
e(s) ds = ≤ ,
0 k k
we can easily show that e(t) ∈ L2 . From (7.34) we have ė(t) ∈ L∞ . Thus, by Bar-
balat’s lemma, we have lim e(t) = 0. This means that the proposed controller (7.32)
t→∞
can globally asymptotically synchronize the system (7.30) and the system (7.31).
This completes the proof. 

Assume that the following delayed Liao chaotic system [8] is the drive system:

ẋ = −α x (t) + a f (x (t) − bx (t − τ ) + c) , (7.37)

where the activation function is

2
f (x) = ∑ ai [tanh (x + ki) − tanh(x − ki)] .
i=1

0.8

0.6

0.4
x

0.2

−0.2

−0.4
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 7.27 The state curve of the drive chaotic system (7.37)
236 7 Synchronization of Chaotic Systems with Time Delay

1.2

0.8

0.6
y

0.4

0.2

−0.2
0 10 20 30 40 50 60 70 80 90 100
t (sec)

Fig. 7.28 The state curve of the response chaotic system (7.38)

The initial conditions of the drive system (7.37) are taken as x (s) = 0.5, s ∈
[−1, 0], α = 1, a = 3, b = 4.5, a1 = 2, a2 = −1.5, k1 = 1, k2 = 4/3, c = 0, and
τ = 1. The state curve of the drive system (7.37) is shown in Fig. 7.27.
Take the following delayed Liao chaotic system [12] as the response system:

ẏ(t) = −β y(t) + mg (y (t) − ny (t − τ ) + c) , (7.38)

where the activation function

2
g (y) = ∑ ai [arctan (y + ki) − arctan (y − ki)] .
i=1

The initial conditions of the response system (7.38) are taken as y (s) = 1, s ∈ [−1, 0],
β = 1, m = 3, n = 4.5, a1 = 2, a2 = −1.5, k1 = 1, k2 = 4/3, c = 0, and τ = 1. The
state curve of the response system (7.38) is shown in Fig. 7.28.
According to Theorem 7.5, the initial values of the drive system and the response
system are x (0) = 0.7 and y (0) = 1.2. Choose α̂ (0) = 0.7, β̂ (0) = 1.1, and k = 1.
The synchronization error curve of the drive system (7.37) and the response sys-
tem (7.38) is shown in Fig. 7.29. The curves of the estimated parameters α̂ and
β̂ are shown in Figs. 7.30 and 7.31, respectively. From simulation results we can
see that the synchronization errors converge asymptotically to zero, and the esti-
mated parameters converge asymptotically to some constants, i.e., the trajectory of
7.3 Adaptive Synchronization Between Two Different Delayed Chaotic Systems 237

Fig. 7.29 The synchronization error curve of the drive system (7.37) and the response system
(7.38)

Fig. 7.30 The curve of the adaptive parameter α̂ (t)

238 7 Synchronization of Chaotic Systems with Time Delay

Fig. 7.31 The curve of the adaptive parameter β̂ (t)

the response system can synchronize asymptotically with the trajectory of the drive
system.

7.4 Synchronization of Chaotic Delayed Neural Networks

7.4.1 Synchronization of a Class of Chaotic Delayed Neural

Networks Based on Inverse Optimal Control Theory

Recently, there is increasing interest in the study of dynamical properties of de-

layed neural networks (DNNs). Most previous studies have predominantly con-
centrated on the stability analysis and periodic oscillations of this kind of network
[1, 20, 21, 22]. It has been shown that such networks can exhibit some complicated
dynamics and even chaotic behaviors. In particular, by appropriately choosing the
network parameters and time delays, the dynamical behavior of these networks can
be made very complicated [2, 11]. Synchronization of chaotic DNNs has potential
applications in several areas such as secure communications, associative memory,
combinatorial optimization, etc. [9, 10]. So, synchronization of this class of chaotic
DNNs [15, 17] is essential and useful in real-world applications.
7.4 Synchronization of Chaotic Delayed Neural Networks 239

In this section, the problem of synchronizing two identical chaotic DNNs is stud-
ied. Under the framework of the inverse optimal control approach, synchronization
controllers and adaptive updating laws of parameters with known and unknown pa-
rameters are designed, respectively [18].

7.4.1.1 Problem Description

A class of delayed chaotic neural networks in this section is described by the fol-
lowing differential equation:
n n
ẋi (t) = −ci xi (t) + ∑ ai j g j (x j (t)) + ∑ bi j g j (x j (t − τ j )) + Ui , (7.39)
j=1 j=1

where n denotes the number of neurons in the network, xi denotes the state variable ; <
associated with the ith neuron, τ j ≥ 0 denotes bounded delay and ρ = max τ j ,
ci > 0, ai j indicates the strength of the neuron interconnections within the network,
bi j indicates the strength of the neuron interconnections within the network with
constant delay parameter τ j , i, j = 1, . . . , n, and Ui is a constant input vector function.
The activation function g j (x j ) satisfies

g j (ξ ) − g j (ζ )
0≤ ≤ δ j,
ξ −ζ

for all ξ , ζ ∈ R, and ξ = ζ . The initial conditions of system (7.39) are given by
xi (t) = φi (t) ∈ C ([−ρ , 0] , R).
Chaotic dynamics is extremely sensitive to initial conditions. Even infinitesimal
changes in the initial conditions will lead to an exponential divergence of orbits.
In order to observe the synchronization behavior in this class of chaotic DNNs,
we study two chaotic DNNs where the drive system and the response system have
identical dynamical structures. The drive system’s state variables are denoted by xi
and the response system’s state variables are denoted by yi .

7.4.1.2 Synchronization of a Class of Chaotic DNNs with Known Parameters

Suppose that the drive system has the form of (7.39). The response system is de-
scribed by the following equation:
n n
ẏi (t) = −ci yi (t) + ∑ ai j g j (y j (t)) + ∑ bi j g j (y j (t − τ j )) + Ui + ui (t), (7.40)
j=1 j=1

where the initial conditions of system (7.40) are given by yi (t) = ϕi (t) ∈ C([−ρ , 0] ,
R), and ui (t) denotes the designed controller that will realize the synchronization
of system (7.39) and system (7.40).
240 7 Synchronization of Chaotic Systems with Time Delay

Let us define the synchronization error signal as ei (t) = yi (t) − xi (t), where xi (t)
and yi (t) are the ith state variables of the drive and response neural networks, re-
spectively. The fact that ei (t) → 0 as t → ∞ means that the drive neural network and
the response neural network are synchronized. The goal of the controller design is
to obtain ui (t) so that lim ei (t) = lim yi (t) − xi (t) = 0, where · denotes the
t→∞ t→∞
Euclidean vector norm.
Therefore, the error system between (7.39) and (7.40) can be expressed as
n n
ėi (t) = −ci ei (t) + ∑ ai j f j (e j (t)) + ∑ bi j f j (e j (t − τ j )) + ui(t), (7.41)
j=1 j=1

where f j (e j (t)) = g j (e j (t) + x j (t)) − g j (x j (t)) and f j (0) = 0, or by the following

compact form:

ė(t) = −Ce(t) + A f (e(t)) + B f (e(t − τ )) + u(t), (7.42)

where C = diag{c1 , . . . , cn }, A = (ai j )n×n , B = (bi j )n×n , τ = (τ1 , . . . , τn )T , u(t) =

(u1 (t), . . . , un (t))T , f (e(t)) = ( f1 (e1 (t)), . . . , fn (en (t)))T , and

f (e(t − τ )) = ( f1 (e1 (t − τ1 )), . . . , fn (en (t − τn )))T .

Before stating the main result, the following lemma is needed.

Lemma 7.1. For any matrices X, Y ∈ Rn×k and Q ∈ Rn×n with Q = QT > 0, the
following inequality holds:

X TY + Y T X ≤ X T QX + Y T Q−1Y. 

About the synchronization of systems (7.39) and (7.40), we have the following
theorem.
Theorem 7.6 ([18]). If the controller is designed as follows:
 
u(t) = − AAT + BBT + 2Σ 2 e (t) , (7.43)

systems (7.39) and (7.40) will be globally asymptotically synchronized, where Σ =

diag{δ1 , . . . , δn }.

Proof. Now construct a Lyapunov function of the form

n  t
1 1
V (e) = eT (t) e (t) +
2 2 ∑ t−τ j
δ j2 e2j (s) ds.
j=1

Then, the time derivative of V (e) along the trajectory of the error system (7.42) is
as follows:
7.4 Synchronization of Chaotic Delayed Neural Networks 241

1 1
V̇ (e) = eT (t) ė (t) + eT (t) Σ 2 e (t) − eT (t − τ ) Σ 2 e (t − τ )
2 2
= − eT (t)Ce(t) + eT (t) (A f (e(t)) + B f (e(t − τ )))
1 1
+ eT (t) Σ 2 e (t) − eT (t − τ ) Σ 2 e (t − τ ) + eT (t) u(t)
2 2
:= L f¯V + (LgV ) u (t) , (7.44)

where

L f¯V := − eT (t)Ce(t) + eT (t) (A f (e(t)) + B f (e(t − τ )))

1 1
+ eT (t) Σ 2 e (t) − eT (t − τ ) Σ 2 e (t − τ ) , (7.45)
2 2

LgV := eT (t) . (7.46)

Applying Lemma 7.1 with Q = I, we get
1 1
V̇ (e) ≤ − eT (t)Ce(t) + eT (t) AAT e(t) + f T (e(t)) f (e(t))
2 2
1 T T 1 T
+ e (t) BB e(t) + f (e(t − τ )) f (e(t − τ ))
2 2
1 T 1 T
+ e (t) Σ e (t) − e (t − τ ) Σ 2 e (t − τ ) + eT (t) u(t)
2
2 2
1 T 1
≤ − e (t)Ce(t) + e (t) AAT e(t) + eT (t) Σ 2 e(t)
T
2 2
1 T T 1 T 2
+ e (t) BB e(t) + e (t − τ ) Σ e(t − τ )
2 2
1 T 1 T
+ e (t) Σ e (t) − e (t − τ ) Σ 2 e (t − τ ) + eT (t) u(t)
2
2 2
1 1
≤ eT (t) −C + AAT + BBT + Σ 2 e(t) + eT (t) u(t). (7.47)
2 2

We design a linear state feedback controller as follows:

u = −β R−1 (e) (LgV )T , (7.48)

where β ≥ 2 is a constant. Denoting

1 T 
R−1 (e) = AA + BBT + 2Σ 2 (7.49)
β
and substituting (7.46) and (7.49) into (7.48), we can derive
 
u = −β R−1 (e) (LgV )T = − AAT + BBT + 2Σ 2 e (t) . (7.50)

Substituting (7.50) into (7.47), we obtain

242 7 Synchronization of Chaotic Systems with Time Delay

1 1
V̇ (e) ≤ −eT (t) C + AAT + BBT + Σ 2 e(t)
2 2
1 T 1 T
≤ −λmin C + AA + BB + Σ 2 e 2
2 2
≤ 0,

which implies that V̇ (e) < 0 for all e(t) = 0. Since V (ε ) is positive definite and V̇ (ε )
is negative semidefinite, it follows that e(t) ∈ L∞ . From the fact that
 t
2 V (e(0)) − V (e(t))
e (s) ds =
0 1 1
λmin C + AAT + BBT + Σ 2
2 2
V (e(0))
≤ ,
1 T 1 T
λmin C + AA + BB + Σ 2
2 2

we can easily show that e(t) ∈ L2 . From (7.41) we have ė(t) ∈ L∞ . Thus, by Bar-
balat’s lemma, we have lim e(t) = 0, i.e., lim e (t) = lim y (t) − x (t) = 0. This
t→∞ t→∞ t→∞
means that the proposed controller (7.50) can globally asymptotically synchronize
the system (7.39) and the system (7.40). This completes the proof. 

Under the framework of the inverse optimal control approach, the controller
(7.43) can be proved to optimize a cost functional of the following form:
  t /

J(u) = lim E(x(t)) + l(x(τ )) + uT (τ )R(x(τ ))u(τ ) dτ , (7.51)
t→∞ 0

where R = RT > 0, u (0) = 0, and l(x(τ )) and E(x(t)) are positive-definite and radi-
ally unbounded functions [5].
Theorem 7.7 ([18]). The following cost functional for system (7.42):
  t /

J (u) = lim 2β V (e (t)) + l (e (τ )) + uT(τ )R (e (τ )) u(τ ) dτ (7.52)
t→∞ 0

will be minimized by controller (7.43), where

l (e) = −2β L f¯V + β 2 (LgV ) R−1 (e) (LgV )T (7.53)

and  −1
R (e) = β AAT + BBT + 2Σ 2 , β ≥ 2. (7.54)

Proof. First, E(e(t)) = 2β V (e(t)) is a positive-definite and radially unbounded

function. We will prove that l(e) is a radially unbounded function. Using (7.45),
(7.46), and (7.54), we obtain
7.4 Synchronization of Chaotic Delayed Neural Networks 243

1 1
l (e) ≥ − 2β eT (t) −C + AAT + BBT + Σ 2 e (t)
2 2
 
+ β eT (t) AAT + BBT + 2Σ 2 e (t)
= 2β eT (t)Ce (t) .

This means that l(e) is radially unbounded. Substituting (7.48) into (7.44), we get

V̇ (e) = L f¯V − β (LgV ) R−1 (e) (LgV )T .

Multiplying both sides of the above equation by −2β , we obtain

−2β V̇ (e) = −2β L f¯V + 2β 2 (LgV ) R−1 (e) (LgV )T . (7.55)

Considering (7.50) and (7.53), from (7.55) we get

l (e) + uT R (e) u = −2β V̇ (e) . (7.56)

Substituting (7.56) into (7.52), we have

  t /

J (u) = lim 2β V (e (t)) + −2β V̇ (e (τ )) dτ
t→∞ 0
= lim {2β V (e (t)) + (−2β V (e (t)) + 2β V (e (0)))}
t→∞
= 2β V (e (0)) .

Thus, the minimum of the cost functional is Jmin (u) = 2β V (e (0)) for the optimal
controller (7.43). The proof is completed. 

The following illustrative example is used to demonstrate the effectiveness of the

above method.
We consider the delayed chaotic cellular neural network [2] as follows:
⎛ π ⎞
ẋ1 (t) 10 x1 (t) 1 + 20 g1 (x1 (t))
=− +⎝ 4 π ⎠ g2 (x2 (t))
ẋ2 (t) 01 x2 (t) 0.1 1 +
√ 4
⎛ ⎞
1.3 2π
⎜− 4 0.1
√ ⎟ g1 (x1 (t − 1))
+⎝ ⎠ , (7.57)
1.3 2π g2 (x2 (t − 1))
0.1 −
4
where gi (xi (t)) = 0.5(|xi (t) + 1| − |xi (t) − 1|), i = 1, 2.
The chaotic behavior of the drive system is shown in Fig. 7.32, which is plotted
with the initial condition (x1 (s), x2 (s))T = (0.01, 0.1)T for −1 ≤ s ≤ 0. The curves
of the drive system’s state variables x1 and x2 are shown in Fig. 7.33. The chaotic
behavior of the response system is shown in Fig. 7.34, which is plotted with the
initial condition (y1 (s), y2 (s))T = (2, −0.2)T for −1 ≤ s ≤ 0.
244 7 Synchronization of Chaotic Systems with Time Delay

0.8

0.6

0.4

0.2

0
x2

−0.2

−0.4

−0.6

−0.8

−1
−20 −15 −10 −5 0 5 10 15 20
x1

Fig. 7.32 Chaotic behavior of the drive delayed chaotic cellular neural network

Fig. 7.33 The curves of the drive system’s state variables x1 and x2
7.4 Synchronization of Chaotic Delayed Neural Networks 245

0.8

0.6

0.4

0.2

0
y2

−0.2

−0.4

−0.6

−0.8

−1
−20 −15 −10 −5 0 5 10 15 20
y1

Fig. 7.34 Chaotic behavior of the response delayed chaotic cellular neural network

Fig. 7.35 The error curves of the delayed chaotic cellular neural network
246 7 Synchronization of Chaotic Systems with Time Delay

The synchronization error curves of the response system and the drive system
are shown in Fig. 7.35. From simulation results we can see that the synchronization
errors converge asymptotically to zero, i.e., the trajectory of the response delayed
chaotic cellular neural network can synchronize asymptotically with the trajectory
of the drive delayed chaotic cellular neural network.

7.4.1.3 Synchronization of a Class of Nonidentical Chaotic DNNs

In the above subsection, the parameters of the two neural networks to be synchro-
nized are known and are identical. Because the synchronized systems are inevitably
perturbed by external factors and cannot be exactly identical, synchronizing two
nonidentical delayed chaotic neural networks is more essential and useful in real-
world applications. In this section, we consider the synchronization problem of a
class of chaotic DNNs in which the parameters of the response system are tunable.
Suppose that the drive system has the form of (7.39), and the response system is
described by the following equation:
n n
ẏi (t) = −ĉi yi (t) + ∑ âi j g j (y j (t)) + ∑ b̂i j g j (y j (t − τ j )) + Ui + ui (t), (7.58)
j=1 j=1

where ĉi , âi j , and b̂i j are tunable parameters of the response system whose updating
laws need to be designed, and ui (t) denotes the designed controller that will achieve
the synchronization of system (7.39) and system (7.58).
Let us define the synchronization error signal as ei (t) = yi (t) − xi (t). The fact
that ei (t) → 0 as t → ∞ means that the drive neural network and the response neural
network are synchronized. Therefore, the error system between (7.39) and (7.58)
can be expressed as
n n
ėi (t) = − ci ei (t) + ∑ ai j f j (e j (t)) + ∑ bi j f j (e j (t − τ j ))
j=1 j=1
n
+ (ci − ĉi ) yi (t) + ∑ (âi j − ai j ) g j (y j (t))
j=1
n
+ ∑ (b̂i j − bi j )g j (y j (t − τ j )) + ui(t), (7.59)
j=1

where f j (e j (t)) = g j (e j (t) + x j (t)) − g j (x j (t)), and f j (0) = 0, or by the following

compact form:

ė(t) = − Ce(t) + A f (e(t)) + B f (e(t − τ )) + C̃y (t) + Ãg (y (t))

+ B̃g (y (t − τ )) + u(t), (7.60)
7.4 Synchronization of Chaotic Delayed Neural Networks 247

where C = diag{c1 , . . . , cn }, A = (ai j )n×n , B = (bi j )n×n , τ = (τ1 , . . . , τn )T , C̃(t) =

diag{c̃1 (t), c̃2 (t), . . . , c̃n (t)}, Ã(t) = (ãi j (t))n×n , B̃(t) = (b̃i j (t))n×n , c̃i = ci − ĉi , in
which ãi j = âi j − ai j and b̃i j = b̂i j − bi j are the estimated errors of unknown param-
eters ci , ai j , and bi j , respectively, f (e(t)) = ( f1 (e1 (t)), . . . , fn (en (t)))T , f (e(t − τ )) =
( f1 (e1 (t − τ1 )), . . . , fn (en (t − τn )))T , and u(t) = (u1 (t), . . . , un (t))T .

Theorem 7.8 ([16]). By the following controller:

 
u(t) = − AAT + BBT + 2Σ 2 e (t) (7.61)

and the parameter update law

⎧
⎨ ĉ˙i = ei (t) yi (t) ,
â˙i j = −ei (t) g j (y j (t)) , (7.62)
⎩˙
b̂i j = −ei (t) g j (y j (t − τ j )) ,

the systems (7.39) and (7.58) will be globally asymptotically synchronized. Here,
i, j = 1, . . . , n and Σ = diag{δ1 , . . . , δn }.

Proof. Define

ε := eT(t), c̃1 , . . . , c̃n , ã11 , . . . , ã1n , . . . , ãn1 , . . . , ãnn ,
T
b̃11 , . . . , b̃1n , . . . , b̃n1 , . . . , b̃nn .

We choose  t
n
1 1
V (ε ) = eT (t) e (t) +
2 2 ∑ t−τ j
δ j2 e2j (s) ds
j=1
1 n 1 n 1 n
+ ∑ c̃2i + ∑ ã2i j + ∑ b̃2i j .
2 i=1 2 i, j=1 2 i, j=1
Its time derivative can be derived as follows:
1 1
V̇ (ε ) = eT (t) ė (t) + eT (t) Σ 2 e (t) − eT (t − τ ) Σ 2 e (t − τ )
2 2
n n n
∑
+ c̃ c̃˙ + i i ∑
ã ã˙ + b̃ b̃˙
ij ij ∑ ij ij
i=1 i, j=1 i, j=1
 
= − eT (t)Ce(t) + eT(t) A f (e(t)) + B f (e(t − τ )) + eT (t) u(t)
  1
+ eT (t) C̃y (t) + Ãg (y (t)) + B̃g (y (t − τ )) + eT (t) Σ 2 e (t)
2
n
1 T
− e (t − τ ) Σ e (t − τ ) − ∑ c̃i ĉ˙i
2
2 i=1
n n
+ ∑ ãi j â˙i j + ∑ b̃i j b̂˙ i j
i, j=1 i, j=1

:= L f¯V + (LgV ) u (t) ,

248 7 Synchronization of Chaotic Systems with Time Delay

where
 
L f¯V := − eT (t)Ce(t) + eT (t) A f (e(t)) + B f (e(t − τ ))
  1
+ eT (t) C̃y (t) + Ãg (y (t)) + B̃g (y (t − τ )) + eT (t) Σ 2 e (t)
2
n
1 T
− e (t − τ ) Σ 2 e (t − τ ) − ∑ c̃i ĉ˙i
2 i=1
n n
+ ∑ ãi j â˙i j + ∑ b̃i j b̂˙ i j
i, j=1 i, j=1

and
LgV := eT (t) .
Applying Lemma 7.1 with Q = I, we get
1 1
V̇ (ε ) ≤ − eT (t)Ce(t) + eT (t) AAT e(t) + f T (e(t)) f (e(t))
2 2
1 T T 1 T
+ e (t) BB e(t) + f (e(t − τ )) f (e(t − τ ))
2 2
1 T 1 T
+ e (t) Σ e (t) − e (t − τ ) Σ 2 e (t − τ ) + eT (t) u(t)
2
2  2 
+ eT (t) C̃y (t) + Ãg (y (t)) + B̃g (y (t − τ ))
n n n
− ∑ c̃i ĉ˙i + ∑ ãi j â˙i j + ∑ b̃i j b̂˙ i j
i=1 i, j=1 i, j=1
1 1
≤ − eT (t)Ce(t) + eT (t) AAT e(t) + eT (t) Σ 2 e(t)
2 2
1 1
+ eT (t) BBT e(t) + eT (t − τ ) Σ 2 e(t − τ )
2 2
1 T 1 T
+ e (t) Σ e (t) − e (t − τ ) Σ 2 e (t − τ ) + eT (t) u(t)
2
2 2
n   n  
+ ∑ c̃i ei (t) yi (t) − ĉ˙i + ∑ ãi j â˙i j + ei (t) g j (y j (t))
i=1 i, j=1
n  
+ ∑ b̃i j b̂˙ i j + ei (t) g j (y j (t − τ j )) .
i, j=1

Substituting (7.62) into the above inequality, we obtain

1 1
V̇ (ε ) ≤ eT (t) −C + AAT + BBT + Σ 2 e(t) + eT (t) u(t). (7.63)
2 2

Substituting (7.61) into (7.63), we obtain

7.4 Synchronization of Chaotic Delayed Neural Networks 249

1 1
V̇ (ε ) ≤ −eT (t) C + AAT + BBT + Σ 2 e(t)
2 2
1 1
≤ −λmin C + AAT + BBT + Σ 2 e 2
2 2
≤ 0,

which implies that V̇ (ε ) < 0 for all e(t) = 0. Since V (ε ) is positive definite and
V̇ (ε ) is negative semidefinite, it follows that e(t) ∈ L∞ , c̃i (t) ∈ L∞ , ãi j (t) ∈ L∞ , and
b̃i (t) ∈ L∞ . From the fact that
 t
2 V (e(0)) − V (e(t))
e (s) ds =
0 1 1
λmin C + AAT + BBT + Σ 2
2 2
V (e(0))
≤ ,
1 T 1 T
λmin C + AA + BB + Σ 2
2 2

we can easily show that e(t) ∈ L2 . From (7.59) we have ė(t) ∈ L∞ . Thus, by Bar-
balat’s lemma, we have lim e(t) = 0, i.e., lim e (t) = lim y (t) − x (t) = 0. This
t→∞ t→∞ t→∞
means that the proposed controller (7.61) can globally asymptotically synchronize
the system (7.39) and the system (7.58). This completes the proof. 
Remark 7.1. Theorem 7.8 can similarly be proved by showing that the cost func-
tional (7.52) is minimized by the optimal control laws (7.61) and (7.62). 
The following illustrative example is used to demonstrate the effectiveness of the
above method.
We consider that the drive delayed chaotic cellular neural network has the form
of (7.39). When the parameters and the initial conditions are the same as those in
(7.57), the chaotic behavior of the drive system is shown in Fig. 7.32.
The response delayed chaotic cellular neural network is chosen in the form of
(7.58). In this simulation, the initial values of ‘unknown’ parameter vectors of the
response system are selected as

1.1 0 1.1 + π /4 20.1

Ĉ(0) = , Â(0) = ,
0 1.1 0.11 1.1 + π /4
√
−1.29 2π /4 0.11
√
B̂(0) = .
0.11 −1.29 2π /4
The chaotic behavior of the response system is shown in Fig. 7.34, which is
plotted with the initial condition (y1 (s), y2 (s))T = (2, −0.2)T for −1 ≤ s ≤ 0.
The synchronization error curves of the response system and the drive system
are shown in Fig. 7.36. The curves of Ĉ, Â, and B̂ are shown in Figs. 7.37–7.39,
respectively. Obviously, the synchronization errors converge asymptotically to zero
and the adaptive parameters converge asymptotically to some constants.
250 7 Synchronization of Chaotic Systems with Time Delay

Fig. 7.36 The error curves of the delayed chaotic cellular neural network

Fig. 7.37 The curves of parameter Ĉ

7.4 Synchronization of Chaotic Delayed Neural Networks 251

Fig. 7.38 The curves of parameter Â

Fig. 7.39 The curves of parameter B̂

252 7 Synchronization of Chaotic Systems with Time Delay

7.4.2 Adaptive Synchronization Between Two Different Chaotic

Delayed Neural Networks

In this section, the problem of synchronizing two different chaotic DNNs with un-
known parameters is studied. By Lyapunov stability theory, the delay-independent
and delay-dependent adaptive synchronization controllers and adaptive updating
laws of parameters are designed, respectively [19].

7.4.2.1 Problem Description

A class of delayed chaotic neural networks is described by the following time-delay

differential equation:
n n
ẋi (t) = −ci xi (t) + ∑ ai j g j (x j (t)) + ∑ bi j g j (x j (t − τ j )) + Hi , (7.64)
j=1 j=1

where n ≥ 2 denotes the number of neurons in the network, xi (t) denotes the
; <associated with the ith neuron, τ j ≥ 0 denotes the bounded delay,
state variable
ρ = max τ j , ci > 0 denotes the self connection of neurons, ai j indicates the
strength of the neuron interconnections within the network, bi j indicates the strength
of the neuron interconnections within the network with constant delay τ j , and Hi
is an external constant input, i, j = 1, . . . , n. The activation function g j : R → R,
j ∈ {1, 2, . . . , n}, is bounded, and satisfies the condition g j (0) = 0 and the Lips-
chitz condition with a Lipschitz constant Lx > 0, i.e., g j (ξ ) − g j (ς ) ≤ Lx ξ − ς
for all ξ , ς ∈ R and ξ = ς . The initial conditions of system (7.64) are given by
xi (t) = φi (t) ∈ C ([−ρ , 0] , R), where C ([−ρ , 0] , R) denotes the set of all continu-
ous functions from [−ρ , 0] to R.
Suppose that the drive system has the form of (7.64). The response system is
described by the following equation:
n n
ẏi (t) = −di yi (t) + ∑ pi j f j (y j (t)) + ∑ qi j f j (y j (t − τ j )) + Wi + ui (t), (7.65)
j=1 j=1

where n ≥ 2 denotes the number of neurons in the network, yi (t) denotes the state
variable associated with the ith neuron, di > 0 denotes the self connection of neu-
rons, pi j and qi j indicate the interconnection strengths among neurons without and
with the bounded delay τ j , Wi is an external constant input, and ui (t) denotes the
external control input and will be appropriately designed to obtain a certain con-
trol objective, i, j = 1, . . . , n. The activation function f j : R → R ( j ∈ {1, 2, . . . , n})
is bounded, and satisfies the condition f j (0) = 0 and the Lipschitz condition with
a Lipschitz constant Ly > 0, i.e., f j (ξ ) − f j (ς ) ≤ Ly ξ − ς for all ξ , ς ∈ R,
and ξ = ς . The initial conditions of system (7.65) are given by yi (t) = ϕi (t) ∈
C ([−ρ , 0] , R).
7.4 Synchronization of Chaotic Delayed Neural Networks 253

Define the synchronization error signal ei (t) = yi (t) − xi (t). Our goal is to de-
sign a controller u(t) = (u1 (t), u2 (t), . . . , un (t))T such that the trajectory of the
response delayed neural network (7.65) can synchronize asymptotically with the
trajectory of the drive delayed neural network (7.64), i.e.,lim e(t) = 0, where
t→∞
e(t) = (e1 (t), e2 (t), . . . , en (t))T .

7.4.2.2 The Delay-independent Controller Design

In this section, we assume that ci and ai j are the unknown parameters of system
(7.64), and bi j is the known parameter of system (7.64). We assume that di and pi j
are the unknown parameters of the response system (7.65), and qi j is the known
parameter of the response system (7.65).
We will need the following assumption and definition.
Assumption 7.1. Both the states of the drive system and the response system are
bounded, i.e., x(t) ≤ Bd and y(t) ≤ Br . 
Definition 7.1. A vector function Φ (x) is defined by

x −1 x, if x = 0,
Φ (x) =
0, if x = 0.


Clearly, we have Φ (0) = 0 from Definition 7.1.
Theorem 7.9 ([19]). By the following controller:

u (t) = − Ĉ(t)x(t) + Â(t)g(x(t)) + D̂(t)y(t) − P̂(t) f (y(t))

− (m1 + m2 )Φ (e) + H − W − Ke (t) (7.66)

and the parameter update law

⎧˙
⎪ ĉi (t) = ei (t)xi (t),
⎪
⎨ â˙i j (t) = −ei (t)g j (x j (t)),
(7.67)
⎪
⎪ d˙ˆ (t) = −ei (t)yi (t),
⎩ ˙i
p̂i j (t) = ei (t) f j (y j (t)),

i, j = 1, . . . , n, the response system (7.65) can synchronize asymptotically with the

drive system (7.64), where the positive diagonal matrix K = diag{k1 , k2 , . . . , kn } is
specified in advance, m1 = Lx Bd B , m2 = Ly Br Q , B = (bi j )n×n , Q = (qi j )n×n ,
H = (H1 , H2 , . . . , Hn )T , W = (W1 ,W2 , . . . ,Wn )T , Ĉ(t) = diag{ĉ1 (t), ĉ2 (t), . . . , ĉn (t)},
Â(t) = (âi j (t))n×n , D̂(t) = diag{dˆ1 (t), dˆ2 (t), . . . , dˆn (t)}, P̂(t) = ( p̂i j (t))n×n , ĉi (t),
âi j (t), dˆi (t), and p̂i j (t) are the estimated values of unknown parameters ci , ai j , di ,
and pi j , respectively, g(x(t)) = (g1 (x1 ), . . . , gn (xn ))T , f (y(t)) = ( f1 (y1 ), . . . , fn (yn ))T ,
and Φ (e) is the same vector function as defined in Definition 7.1.
254 7 Synchronization of Chaotic Systems with Time Delay

Proof. From (7.64)–(7.66), we get the error system as follows:

n
ėi (t) = − (ĉi (t) − ci) xi (t) + ∑ (âi j (t) − ai j ) g j (x j (t))
j=1
n  
− ∑ bi j g j (x j (t − τ j )) + dˆi (t) − di yi (t)
j=1
n n
− ∑ ( p̂i j (t) − pi j ) f j (y j (t)) + ∑ qi j f j (y j (t − τ j ))
j=1 j=1

− (m1 + m2 )Φ (e) − ki ei (t) . (7.68)

Let c̃i (t) = ĉi (t) − ci , ãi j (t) = âi j (t) − ai j , d˜i (t) = dˆi (t) − di, and p̃i j (t) = p̂i j (t) − pi j
be the estimated errors of unknown parameters ci , ai j , di , and pi j , respectively. We
obtain the following compact form:

ė (t) = − C̃(t)x(t) + Ã(t)g(x(t)) − Bg(x(t − τ ))

+ D̃(t)y(t) − P̃(t) f (y(t)) + Q f (y(t − τ ))
− (m1 + m2 )Φ (e) − Ke (t) , (7.69)

where τ = (τ1 , . . . , τn )T , C̃(t) = diag{c̃1 (t), c̃2 (t), . . . , c̃n (t)}, D̃(t) = diag{d˜1 (t), d˜2 (t),
. . . , d˜n (t)}, Ã(t) = (ãi j (t))n×n , P̃(t) = ( p̃i j (t))n×n , g(x(t − τ )) = (g1 (x1 (t − τ1 )), . . . ,
gn (xn (t − τn )))T , and f (y(t − τ )) = ( f1 (y1 (t − τ1 )), . . . , fn (yn (t − τn )))T .
Define

ε := (eT (t) , c̃1 (t), . . . , c̃n (t), ã11 (t), . . . , ã1n (t), . . . , ãn1 (t), . . . , ãnn (t),
d˜1 (t), . . . , d˜n (t), p̃11 (t), . . . , p̃1n (t), . . . , p̃n1 (t), . . . , p̃nn (t))T .

If a Lyapunov function candidate is chosen as

'
n n
1 T
V (ε ) = e (t) e (t) + ∑ c̃2i (t) + ∑ ã2i j (t)
2 i=1 i, j=1
(
n n
+ ∑ d˜i2 (t) + ∑ p̃2i j (t) , (7.70)
i=1 i, j=1

then the time derivative of V (ε ) along the trajectory of the error system (7.69) is as
follows:
n n
V̇ (ε ) = eT (t) ė (t) + ∑ c̃i (t)c̃˙i (t) + ∑ ãi j (t)ã˙i j (t)
i=1 i, j=1
n n
+ ∑ d˜i (t)d˙˜i (t) + ∑ p̃i j (t) p̃˙i j (t)
i=1 i, j=1
7.4 Synchronization of Chaotic Delayed Neural Networks 255

= eT (t) (−C̃(t)x(t) + Ã(t)g(x(t)) + D̃(t)y(t) − P̃(t) f (y(t)))

+ eT (t) (−Bg(x(t − τ )) + Q f (y(t − τ )) − (m1 + m2 )Φ (e))
n n
+ ∑ c̃i (t)ĉ˙i (t) + ∑ ãi j (t)â˙i j (t) − eT (t) Ke (t)
i=1 i, j=1
n n
+ ∑ d˜i (t)d˙ˆi (t) + ∑ p̃i j (t) p̂˙i j (t)
i=1 i, j=1
n   n  
= ∑ c̃i (t) −ei (t) xi (t) + ĉ˙i (t) + ∑ d˜i (t) ei (t) yi (t) + d˙ˆi(t)
i=1 i=1
n  
+ ∑ ãi j (t) ei (t) g j (x j (t)) + â˙i j (t)
i, j=1
n  
+ ∑ p̃i j (t) −ei (t) f j (y j (t)) + p̂˙ i j (t)
i, j=1

+ eT (t) (−Bg(x(t − τ )) + Q f (y(t − τ ))

− (m1 + m2 )Φ (e)) − eT (t) Ke (t) . (7.71)

Substituting (7.67) into (7.71), we obtain

V̇ (ε ) ≤ e (t) (Lx −B x (t − τ ) + Ly Q y (t − τ ) )
T T T
− e (t) m1 Φ (e) − e (t) m2 Φ (e) − e (t) Ke (t)
≤ e (t) (Lx B Bd + Ly Q Br ) − m1 e (t)
− m2 e (t) − eT (t) Ke (t)
= − eT (t) Ke (t) ≤ 0,

which implies that V̇ (ε ) < 0 for all e(t) = 0. Since V (ε ) is positive definite and V̇ (ε )
is negative semidefinite, it follows that e(t) ∈ L∞ , c̃i (t) ∈ L∞ , ãi j (t) ∈ L∞ , d˜i (t) ∈ L∞ ,
and p̃i j (t) ∈ L∞ . From the fact that
 t
2 V (e(0)) − V (e(t)) V (e(0))
e(s) ds = ≤ ,
0 min{ki } min{ki }

we can easily show that e(t) ∈ L2 . From (7.69) we have ė(t) ∈ L∞ . Thus, by Bar-
balat’s lemma, we have lim e(t) = 0, i.e., lim e (t) = lim y (t) − x (t) = 0. This
t→∞ t→∞ t→∞
means that the proposed controller (7.66) can globally asymptotically synchronize
the system (7.64) and the system (7.65). This completes the proof. 

The following illustrative example is used to demonstrate the effectiveness of the

above method.
Assume that the following delayed chaotic cellular neural network [2] is the drive
system:
256 7 Synchronization of Chaotic Systems with Time Delay

ẋ1 (t) x1 (t) g1 (x1 (t)) g1 (x1 (t − τ1 ))

= −C +A +B , (7.72)
ẋ2 (t) x2 (t) g2 (x2 (t)) g2 (x2 (t − τ2 ))

where g j (x j (t)) = 0.5( x j (t) + 1 − x j (t) − 1 ), j = 1, 2, and the following delayed

chaotic Hopfield neural network [11] is the response system:

ẏ1 (t) y1 (t) f1 (y1 (t)) f1 (y1 (t − τ3 ))

= −D +P +Q + u (t) , (7.73)
ẏ2 (t) y2 (t) f2 (y2 (t)) f2 (y2 (t − τ4 ))

where f j (y j (t)) = tanh (y j (t)) , j = 1, 2.

For numerical simulations, we use the fourth-order Runge–Kutta method to solve
the systems of differential equations (7.66), (7.67), (7.72), and (7.73). The ‘un-
known’ parameters of the drive system (7.72) and the response system (7.73) with-
out controller u(t) are selected as C = diag[1, 1], D = diag{1, 1},
√
1 + π /4 20 −1.3 2π /4 0.1
√
A= , B= ,
0.1 1 + π /4 0.1 −1.3 2π /4

2 −0.1 −1.5 −0.1

P= , Q= ,
−5 4.5 −0.2 −4
and τ1 = τ2 = τ3 = τ4 = 1. The initial conditions of the drive system (7.72)
and the response system (7.73) are taken as (x1 (s), x2 (s))T = (0.01, 0.1)T and
(y1 (s), y2 (s))T = (0.4, 0.6)T for −1 ≤ s ≤ 0, respectively. The chaotic behavior
of the drive system (7.72) is shown in Fig. 7.32, and the response system (7.73)
without the controller u(t) is shown in Fig. 7.40.
In the example, we choose Bd = 16.0312 and Br = 6.0828 from Fig. 7.32 and Fig.
7.40, respectively. According to Theorem 7.9, the initial values of the controller and
the parameter update law are selected as follows: Ĉ(0) = diag{1.1, 1.1}, D̂(0) =
diag{1.1, 1.1}, K = diag{1, 1}, Lx = Ly = 1,

1.885 20.1 2.1 −0.09

Â(0) = , and P̂(0) = .
0.11 1.885 −4.9 4.5

The synchronization error curves of the drive system (7.72) and the response
system (7.73) with the controller u(t) are shown in Fig. 7.41. The curves of the es-
timated parameters Ĉ, Â, D̂, and P̂ are shown in Figs. 7.42–7.45, respectively. From
simulation results we can see that the synchronization errors converge asymptot-
ically to zero, i.e., the trajectory of the response delayed chaotic Hopfield neural
network can synchronize asymptotically with the trajectory of the drive delayed
chaotic cellular neural network. The estimated parameters converge asymptotically
to some constants, which shows the effectiveness of the adaptive synchronization
scheme.
7.4 Synchronization of Chaotic Delayed Neural Networks 257

2
y2

−2

−4

−6
−1.2 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6
y1

Fig. 7.40 Chaotic behavior of the response system (7.73)

Fig. 7.41 The synchronization error curves of the drive system (7.72) and the response system
(7.73)
258 7 Synchronization of Chaotic Systems with Time Delay

Fig. 7.42 The curves of parameter Ĉ(t)

Fig. 7.43 The curves of parameter Â(t)

7.4 Synchronization of Chaotic Delayed Neural Networks 259

Fig. 7.44 The curves of parameter D̂(t)

Fig. 7.45 The curves of parameter P̂(t)

260 7 Synchronization of Chaotic Systems with Time Delay

7.4.2.3 The Delay-dependent Controller Design

In this section, we assume that ci , ai j , and bi j are the unknown parameters of system
(7.64) and di , pi j , and qi j are the unknown parameters of the response system (7.65).
Theorem 7.10 ([16]). By the following controller:

u (t) = − Ĉx(t) + Âg(x(t)) + B̂g(x(t − τ )) + D̂y(t)

− P̂ f (y(t)) − Q̂ f (y(t − τ )) + H − W − Ke (t) (7.74)

and the parameter update law

⎧
⎪
⎪ ĉ˙i = ei (t)xi (t),
⎪
⎪
⎪
⎪ â˙i j = −ei (t)g j (x j (t)),
⎪
⎨˙
b̂i j = −ei (t)g j (x j (t − τ j )),
(7.75)
⎪
⎪
⎪ d˙ˆi = −ei (t)yi (t),
⎪
⎪ p̂˙i j = ei (t) f j (y j (t)),
⎪
⎪
⎩˙
q̂i j = ei (t) f j (y j (t − τ j )),

i, j = 1, . . . , n, the response system (7.65) can synchronize asymptotically with the

drive system (7.64), where the positive diagonal matrix K = diag{k1 , k2 , . . . , kn } is
specified in advance, τ = (τ1 , τ2 , . . . , τn )T , H = (H1 , H2 , . . . , Hn )T , W = (W1 ,W2 , . . . ,
Wn )T , Ĉ(t) = diag{ĉ1 (t), ĉ2 (t), . . . , ĉn (t)}, Â(t) = (âi j (t))n×n , B̂(t) = (b̂i j (t))n×n ,
D̂(t) = diag{dˆ1(t), dˆ2 (t), . . . , dˆn (t)}, P̂(t) = ( p̂i j (t))n×n , Q̂(t) = (q̂i j (t))n×n , ĉi (t),
âi j (t), b̂i j (t), dˆi (t), p̂i j (t), and q̂i j (t) are the estimated values of unknown param-
eters ci , ai j , bi j , di , pi j , and qi j , respectively, g(x(t)) = (g1 (x1 (t)), . . . , gn (xn (t)))T ,
f (y(t)) = ( f1 (y1 (t)), . . . , fn (yn (t)))T , g(x(t − τ )) = (g1 (x1 (t − τ1 )), . . . , gn (xn (t −
τn )))T , and f (y(t − τ )) = ( f1 (y1 (t − τ1 )), . . . , fn (yn (t − τn )))T.
Proof. From (7.64), (7.65), and (7.74), we get the error system as follows:
n
ėi (t) = − (ĉi − ci ) xi (t) + ∑ (âi j − ai j ) g j (x j (t)) − ki ei (t)
j=1
n   
+ ∑ b̂i j − bi j g j (x j (t − τ j )) + dˆi − di yi (t)
j=1
n n
− ∑ ( p̂i j − pi j ) f j (y j (t)) − ∑ (q̂i j − qi j ) f j (y j (t − τ j )).
j=1 j=1

Let c̃i (t) = ĉi (t) − ci , ãi j (t) = âi j (t) − ai j , b̃i j (t) = b̂i j (t) − bi j , d˜i (t) = dˆi (t) − di ,
p̃i j (t) = p̂i j (t) − pi j , and q̃i j (t) = q̂i j (t) − qi j be the estimated errors of unknown
parameters ci , ai j , bi j , di , pi j , and qi j , respectively.
We obtain the following compact form:

ė (t) = − C̃(t)x(t) + Ã(t)g(x(t)) + B̃(t)g(x(t − τ ))

+ D̃(t)y(t) − P̃(t) f (y(t)) − Q̃(t) f (y(t − τ )) − Ke (t) , (7.76)
7.4 Synchronization of Chaotic Delayed Neural Networks 261

where C̃(t) = diag{c̃1 (t), c̃2 (t), . . . , c̃n (t)}, D̃(t) = diag{d˜1 (t), d˜2 (t), . . . , d˜n (t)}, Ã(t)
= (ãi j (t))n×n , B̃(t) = (b̃i j (t))n×n , P̃(t) = ( p̃i j (t))n×n , and Q̃(t) = (q̃i j (t))n×n . Define

ε := [ eT (t), c̃1 (t), . . . , c̃n (t), ã11 (t), . . . , ã1n (t), . . . , ãn1 (t), . . . , ãnn (t),
b̃11 (t), . . . , b̃1n (t), . . . , b̃n1 (t), . . . , b̃nn (t), d˜1 (t), . . . , d˜n (t),
p̃11 (t), . . . , p̃1n (t), . . . , p̃n1 (t), . . . , p̃nn (t),
q̃11 (t), . . . , q̃1n (t), . . . , q̃n1 (t), . . . , q̃nn (t)]T .

If a Lyapunov function candidate is chosen as

'
n n n n
1 T
V (ε ) = e (t) e (t) + ∑ c̃2i (t) + ∑ ã2i j (t) + ∑ b̃2i j (t) + ∑ d˜i2 (t)
2 i=1 i, j=1 i, j=1 i=1
(
n n
+ ∑ q̃2i j (t) + ∑ p̃2i j (t) ,
i, j=1 i, j=1

then the time derivative of V (ε ) along the trajectory of the error system (7.76) is as
follows:
n n
V̇ (ε ) = eT (t) ė (t) + ∑ c̃i (t)c̃˙i (t) + ∑ ãi j (t)ã˙i j (t)
i=1 i, j=1
n n
+ ∑ b̃i j (t)b̃˙ i j (t) + ∑ d˜i (t)d˙˜i (t)
i, j=1 i=1
n n
+ ∑ p̃i j (t) p̃˙i j (t) + ∑ q̃i j (t)q̃˙i j (t)
i, j=1 i, j=1

= eT (t) (−C̃(t)x(t) + Ã(t)g(x(t)) + D̃(t)y(t) − P̃(t) f (y(t))

+ B̃g(x(t − τ )) − Q̃ f (y(t − τ )))
n n
+ ∑ c̃i (t)ĉ˙i (t) + ∑ ãi j (t)â˙i j (t) − eT (t) Ke (t)
i=1 i, j=1
n n
+ ∑ b̃i j (t)b̂˙ i j (t) + ∑ d˜i (t)d˙ˆi (t)
i, j=1 i=1
n n
+ ∑ p̃i j (t) p̂˙i j (t) + ∑ q̃i j (t)q̂˙i j (t)
i, j=1 i, j=1
n  
= ∑ c̃i (t) −ei (t) xi (t) + ĉ˙i (t)
i=1
n  
+ ∑ d˜i (t) ei (t) yi (t) + d˙ˆi(t)
i=1
262 7 Synchronization of Chaotic Systems with Time Delay
n  
+ ∑ ãi j (t) ei (t) g j (x j (t)) + â˙i j (t)
i, j=1
n  
+ ∑ b̃i j (t) ei (t) g j (x j (t − τ j )) + b̂˙ i j (t)
i, j=1
n  
+ ∑ p̃i j (t) −ei (t) f j (y j (t)) + p̂˙ i j (t)
i, j=1
n  
+ ∑ q̃i j (t) −ei (t) f j (y j (t − τ j )) + q̂˙i j (t) − eT (t) Ke (t) . (7.77)
i, j=1

Substituting (7.75) into (7.77), we obtain

V̇ (ε ) = −eT (t) Ke (t) ≤ 0,

which implies that V̇ (ε ) < 0 for all e(t) = 0. Since V (ε ) is positive definite and V̇ (ε )
is negative semidefinite, it follows that e(t) ∈ L∞ , c̃i (t) ∈ L∞ , ãi j (t) ∈ L∞ , b̃i j (t) ∈ L∞ ,
d˜i (t) ∈ L∞ , p̃i j (t) ∈ L∞ , and q̃i j (t) ∈ L∞ . From the fact that
 t
2 V (e(0)) − V (e(t)) V (e(0))
e(s) ds = ≤ ,
0 min{ki } min{ki }

we can easily show that e(t) ∈ L2 . From (7.76) we have ė(t) ∈ L∞ . Thus, by Bar-
balat’s lemma, we have lim e(t) = 0, i.e., lim e (t) = lim y (t) − x (t) = 0 . This
t→∞ t→∞ t→∞
means that the proposed controller (7.74) and (7.75) can globally asymptotically
synchronize the system (7.64) and the system (7.65). This completes the proof. 

The following illustrative example is used to demonstrate the effectiveness of the

above method.
Suppose that the drive delayed chaotic cellular neural network has the form of
(7.72). We consider that the response delayed chaotic Hopfield neural network has
the form of (7.73).
According to Theorem 7.10, the initial values of the controller and the parameter
update law are selected as follows: Ĉ(0) = diag{1.1, 1.1}, D̂(0) = diag{1.1, 1.1},
K = diag{1, 1},

1.885 20.1 −1.432 0.11

Â(0) = , B̃ (0) = ,
0.11 1.885 0.11 − 1.432

2.1 −0.09 −1.56 − 0.09

P̂(0) = , and Q̃ (0) = .
−4.9 4.5 −0.19 − 4.1
The synchronization error curves of the drive system (7.72) and the response
system (7.73) with the controller u(t) are shown in Fig. 7.46. The curves of the es-
timated parameters Ĉ, Â, B̂, D̂, P̂, and Q̂ are shown in Figs. 7.47–7.52, respectively.
From simulation results we can see that the synchronization errors converge asymp-
totically to zero, i.e., the trajectory of the response delayed chaotic Hopfield neural
7.4 Synchronization of Chaotic Delayed Neural Networks 263

Fig. 7.46 The synchronization error curves of the drive system (7.72) and the response system
(7.73)

Fig. 7.47 The curves of parameter Ĉ(t)

264 7 Synchronization of Chaotic Systems with Time Delay

Fig. 7.48 The curves of parameter Â(t)

Fig. 7.49 The curves of parameter B̂(t)

7.4 Synchronization of Chaotic Delayed Neural Networks 265

Fig. 7.50 The curves of parameter D̂(t)

Fig. 7.51 The curves of parameter P̂(t)

266 7 Synchronization of Chaotic Systems with Time Delay

Fig. 7.52 The curves of parameter Q̂(t)

network can synchronize asymptotically with the trajectory of the drive delayed
chaotic cellular neural network. The estimated parameters converge asymptotically
to some constants, which shows the effectiveness of the adaptive synchronization
scheme.

7.5 Summary

In this chapter, the problem of adaptive synchronization of delayed chaotic systems

is studied. Many examples are presented to show how to design the synchronization
controllers when the systems to be synchronized have different structures and differ-
ent parameters. The method proposed in Sect. 7.2 can adaptively synchronize two
identical delay chaotic systems with unknown parameters. The methods developed
in Sects. 7.3 and 7.4.2 can adaptively synchronize different chaotic systems. In Sect.
7.4.1, a controller based on inverse optimal control theory is designed, which not
only can synchronize two delayed chaotic systems, but also can optimize a mean-
ingful performance index. The numerical simulations have shown that the proposed
methods are effective and can be used for a quite wide class of delayed chaotic
systems.
References 267

References

1. Cao J (2000) Periodic solution and exponential stability of delayed CNNs. Phys Lett A
270:157–163
2. Gilli M (1993) Strange attractors in delayed cellular neural networks. IEEE Trans Circuits
Syst I 40:849–853
3. Gopalsmay K, Issic LKC (1994) Convergence under dynamical thresholds with delays. IEEE
Trans Neural Networks 8:341–348
4. Ikeda K (1979) Multiple-valued stationary state and its instability of the transmitted light by
a ring cavity system. Opt Commun 30:257–261
5. Krstic M, Li ZH (1998) Inverse optimal design of input-to-state stabilizing nonlinear con-
trollers. IEEE Trans Autom Control 43:336–351
6. Li CD, Liao XF, Wong KW (2004) Chaotic lag synchronization of coupled time-delayed
systems and its applications in secure communication. Physica D 194:187–202
7. Li CD, Liao XF, Zhang R (2005) A unified approach for impulsive lag synchronization of
chaotic systems with time delay. Chaos Solitons Fractals 23:1177–1184
8. Liao XF, Wong KW, Leung CS, Wu Z (2001) Hopf bifurcation and chaos in a single delayed
neuron equation with non-monotonic activation function. Chaos Solitons Fractals 12:1535–
1547
9. Liu D, Zhang Y, Zhang HG (2005) Self-learning call admission control for CDMA cellular
networks. IEEE Trans Neural Networks 16:1219–1228
10. Liu D, Xiong X, DasGupta B, Zhang HG (2006) Motif discoveries in unaligned molecular
sequences using self-organizing neural networks. IEEE Trans Neural Networks 17:919–928
11. Lu HT (2002) Chaotic attractors in delayed neural networks. Phys Lett A 298:109–116
12. Peng J, Liao XF (2003) Synchronization of a coupled time-delay chaotic system and its ap-
plication to secure communications. Comput Res Dev 40:263–268
13. Ren HP, Liu D, Han CZ (2006) Anticontrol of chaos via direct time delay feedback. Acta
Phys Sinica 55:2694–2708
14. Song Y, Wei J (2004) Bifurcation analysis for Chen’s system with delayed feedback and its
application to control of chaos. Chaos Solitons Fractals 22:75–91
15. Wang ZS, Zhang HG, Wang ZL (2006) Global synchronization of a class of chaotic neural
networks. Acta Phys Sinica 55:2687–2693
16. Xie YH (2007) Researches on control and synchronization of several kinds of nonlinear
chaotic systems with time delay. Ph.D. dissertation, Northeastern University, Shenyang
17. Zhang HG, Guan HX, Wang ZS (2007) Adaptive synchronization of neural networks with
different attractors. Prog Nat Sci 17:687–695
18. Zhang HG, Xie Y, Liu D (2006) Synchronization of a class of delayed chaotic neural networks
with fully unknown parameters. Dyn Contin Discrete Impuls Syst B 13:297–308
19. Zhang HG, Xie Y, Wang ZL, Zheng CD (2007) Adaptive synchronization between two dif-
ferent chaotic neural networks with time-delay. IEEE Trans Neural Networks 18:1841–1845
20. Zhang HG, Wang ZS, Liu D (2007) Robust exponential stability of cellular neural networks
with multiple time varying delays. IEEE Trans Circuits Syst II 54:730–734
21. Zhang HG, Wang ZS (2007) Global asymptotic stability of delayed cellular neural networks.
IEEE Trans Neural Networks 18:947–950
22. Zhang Y (2002) Absolute periodicity and absolute stability of delayed neural networks. IEEE
Trans Circuits Syst I 49:256–261
Chapter 8
Synchronizing Chaotic Systems Based on Fuzzy
Models

Abstract A motivation for using fuzzy systems and fuzzy control stems in part from
the fact that they are particularly suitable for industrial processes when the physical
systems or qualitative criteria are too complex to model and they have provided
an efficient and effective way in the control of complex uncertain nonlinear or ill-
defined systems. In recent years, fuzzy logic systems have received much attention
from control theorists as a powerful tool for nonlinear control. In this chapter, we
first introduce fuzzy modeling methods for some classical chaotic systems via the
Takagi–Sugeno (T–S) fuzzy model. Next, we model some hyperchaotic systems
using the T–S fuzzy model and then, based on these fuzzy models, we develop
an H∞ synchronization method for two different hyperchaotic systems. Finally, the
problem of synchronizing a class of time-delayed chaotic systems based on the T–S
fuzzy model is considered.

8.1 Introduction

In recent years, fuzzy logic systems have received much attention from the control
community as a powerful tool for nonlinear control. A motivation for using fuzzy
systems and fuzzy control stems in part from the fact that they are particularly suit-
able for industrial processes when the physical systems or qualitative criteria are
too complex to model and they have provided an efficient and effective way in the
control of complex uncertain nonlinear or ill-defined systems.
Among various kinds of fuzzy control or system methods, the T–S fuzzy system
is widely accepted as a powerful tool for the design of fuzzy control [12, 19, 20,
21, 24]. The T–S fuzzy model is frequently used for mathematical simplicity of
analysis. In this type of fuzzy model, local dynamics in different state–space regions
is represented by linear models [13]. The overall model of the system is achieved by
fuzzy blending of these linear models.
It is well known that, since the pioneering work of Pecora and Carroll [11], syn-
chronization of two chaotic dynamical systems is one of the most important applica-

269
270 8 Synchronizing Chaotic Systems Based on Fuzzy Models

tions of chaos and has been paid much attention. Various control methods have been
applied to synchronize two chaotic or hyperchaotic systems, such as linear feedback
control, nonlinear feedback control, and impulsive control [1, 22, 23]. Among these
synchronization methods, some must use high gains in designing parameters, others
need Lipschitz conditions of nonlinear terms to be satisfied. Moreover, many results
aforementioned dealt with one or two kinds of specified chaotic systems. There are
few unified methods suitable for synchronization of a variety of chaotic systems.
In particular, for chaotic systems that evolve within a bounded region of the state
space, the T–S fuzzy model can represent the nonlinear dynamics by a small set of
linear subsystems coupled with linguistic variables. The idea of controlling nonlin-
ear dynamical systems based on the T–S fuzzy model can always be divided into
two steps: (i) represent the nonlinear system using the T–S fuzzy system and (ii)
design a controller for the fuzzy system. Based on the parallel distributed compen-
sation (PDC) scheme, a lot of research on the synchronization of fuzzy systems has
been carried out.
In this chapter, we first introduce fuzzy modeling methods for some classical
chaotic systems via the T–S fuzzy model. Next, we model some hyperchaotic sys-
tems with the T–S fuzzy model and then, based on the fuzzy model, we develop
an H∞ synchronization method for two different hyperchaotic systems. Finally, the
problem of synchronizing a class of time-delayed chaotic systems based on the T–S
fuzzy model is considered.

8.2 Modeling Chaotic Systems via T–S Fuzzy Models

The T–S fuzzy dynamical model, which originated from Takagi and Sugeno, is de-
scribed by fuzzy IF–THEN rules in which the consequent parts represent local linear
models. To realize a fuzzy model-based design, chaotic systems should first be rep-
resented exactly by T–S fuzzy models. From the investigation of many well-known
continuous-time and discrete-time chaotic systems, we find that nonlinear terms
have a common variable or depend only on one variable. If we take it as the premise
variable of fuzzy rules, a simple fuzzy dynamical model can be obtained and will
represent exactly chaotic systems. We observe that all the well-known chaotic sys-
tems can be applied to synchronization and secure communications either by a fuzzy
driving signal or by a crisp driving signal. In the following, we will introduce how
to represent many well-known chaotic systems by the T–S fuzzy model [9].
Consider a general chaotic dynamical system as follows:

sx(t) = f (x(t)) + g(x(t))u(t) + C, (8.1)

where sx(t) is ẋ(t) in continuous-time or x(t + 1) in discrete-time systems, respec-

tively; x(t) and u(t) are the state and control input vectors, respectively; C is a con-
stant vector; and f (·) and g(·) are nonlinear functions with appropriate dimensions.
Then, the fuzzy model is composed of the following rules:
8.2 Modeling Chaotic Systems via T–S Fuzzy Models 271

Ri : IF z(t) is Mi ,
(8.2)
THEN sx(t) = Ai x(t) + Bi u(t) + bi,

where x(t) ∈ Rn and u(t) ∈ Rm denote the state vector and the input vector, respec-
tively, Ai ∈ Rn×n and Bi ∈ Rn×m are the known system matrix and input matrix with
appropriate dimensions, respectively, z(t) is a premise variable, Mi is a fuzzy set, bi
is a constant vector, i = 1, 2, . . . , r, and r is the number of fuzzy rules.
By taking a standard fuzzy inference strategy, i.e., using a singleton fuzzifier,
product fuzzy inference, and center average defuzzifier, the final continuous-time
fuzzy T–S system is inferred as follows:

∑ri=1 ωi (z(t))[Ai x(t) + Bi u(t) + bi]

sx(t) = , (8.3)
∑ri=1 ωi (z(t))
where ωi (z(t)) = Mi (z(t)), in which Mi (z(t)) is the degree of membership of z(t) in
Mi , with ⎧
⎨ r ω (z(t)) > 0,
∑ i
⎩ i=1
ω (z(t)) ≥ 0, i = 1, 2, . . . , r.
i

By using μi (z(t)) = ωi (z(t))/ ∑ri=1 ωi (z(t)) instead of ωi (z(t)), system (8.3) is

rewritten as
r
sx(t) = ∑ μi (z(t))[Ai x(t) + Bi u(t) + bi].
i=1

The premise variables are assumed to be independent of the input vector u(t) in
this chapter. However, it should be noted that all the stability conditions derived in
this chapter can be applied even to the case that the premise variables depend on the
input vector. This assumption is employed to avoid a complicated defuzzification
process of fuzzy controllers.

Assumption 8.1. Considering the boundedness of chaotic systems, the fuzzy set is
chosen as the region of the system trajectory in the set

Ω := {x(t) ∈ Rn : x(t)  ε }.

For chaotic systems, the existence of the parameter ε is natural. Therefore, nonlinear
systems can be represented exactly by the T–S fuzzy model. 

In the process of constructing a T–S fuzzy model (8.2) which represents exactly
the nonlinear system (8.1), we will focus on nonlinear terms of the nonlinear system.
The consistency of nonlinear terms in the system and its associated fuzzy represen-
tation is emphasized here. Without loss of generality, fuzzy modeling methods are
proposed for three kinds of nonlinear terms [9]; that is (i) only one variable in a
nonlinear term; (ii) multi-variables in a nonlinear term; and (iii) multiple nonlinear
terms in a system.
Case 1: The Hénon Map (Only One Variable in a Nonlinear Term) is given by
272 8 Synchronizing Chaotic Systems Based on Fuzzy Models

x1 (t + 1) = 1.4 − x21(t) + 0.3x2(t),

x2 (t + 1) = x1 (t).

Consider a single scalar nonlinear function f (x(t)) = x21 (t) which depends only on
one state variable xk in the Hénon map (xk = x1 ). Let the nonlinear term f (xk ) take
the form φ (xk )x̃k , where
⎧
⎨ xk , if lim f (xk ) ∈ L∞ ,
⎪
x̃k = x∈Ω xk
⎪ xk →0
⎩
1, otherwise.

Take xk , which forms the function φ (xk ), as the premise variable. Then, the fuzzy
representation is composed of the following fuzzy rules:

IF xk is Fi ,
THEN fˆ = di x̃k ,

for i = 1, 2, . . . , r, where Fi is the fuzzy set, fˆ(xk ) is a fuzzy representation of f (xk ),

di is a constant coefficient to be determined, and r is the number of fuzzy rules. The
fuzzy inferred output is written as
r
∑ μi (xk )di
i=1
fˆ(xk ) = r x̃k .
∑ μi (xk )
i=1

r
Without loss of generality, it is required that ∑ μi (xk ) = 1, μi (xk ) ∈ [0, 1], which
i=1
r
further yields φ (xk ) = ∑ μi (xk )di . We let r = 2 and specify the membership func-
i=1
tions. From μ1 (xk ) + μ2 (xk ) = 1 and μ1 (xk )d1 + μ2 (xk )d2 = φ (xk ), we have

−d2 1
μ1 = + φ (xk ),
d1 − d2 d1 − d2
μ2 = 1 − μ1 .

For instance, let d = d1 = −d2 in which d is the upper bound of φ (xk ), i.e., d =
supx∈Ω |φ (xk )|. The fuzzy sets can be chosen as

1 φ (xk )
F1 = 1+ ,
2 d
1 φ (xk )
F2 = 1− .
2 d
8.2 Modeling Chaotic Systems via T–S Fuzzy Models 273

0.4

0.3
Xf1

0.2

0.1

x2 0

−0.1

−0.2 Xf2

−0.3

−0.4
−1.5 −1 −0.5 0 0.5 1 1.5
x1

Fig. 8.1 The chaotic attractor of the fuzzy Hénon map

Let x1 be the premise variable. Then, the equivalent fuzzy model of the Hénon
map can be constructed as

IF x1 is Fi ,
THEN ẋ(t) = Ai x(t) + bi,

−2 0.3 2 0.3
i = 1, 2, where d = 2, φ (x1 ) = x1 , A1 = , A2 = , b1 = b2 =
1 0 1 0
1.4
, and the fuzzy sets are
0

1 x1 
F1 = 1+ ,
2 2
1  x1 
F2 = 1− .
2 2
The chaotic attractor of the fuzzy Hénon map is shown in Fig. 8.1.
Case 2: The Rössler System (Multi-variables in a Nonlinear Term) is given by
⎧
⎨ ẋ1 = −x2 − x3 ,
⎪
ẋ2 = x1 + 0.2x2,
⎪
⎩
ẋ3 = 0.2 + x3(x1 − 5).
274 8 Synchronizing Chaotic Systems Based on Fuzzy Models

Consider a single scalar nonlinear function f (x) in which x = (x1 , x2 , . . . , xn )T , where

many state variables appear in it. Assume that the nonlinear term f (x) can be ex-
pressed as f (x) = φ1 (x1 )φ2 (x2 ) · · · φn (xn )x̃k , where
⎧
⎨ xk , if lim f (xk ) ∈ L∞ for some k = 1, 2, . . . , n,
⎪
x̃k = x∈Ω xk
⎪ xk →0
⎩
1, otherwise.

Let the variables xk in φk (·), for k = 1, 2, . . . , n, be premise variables. Then, the ith
rule of the fuzzy system will be of the following form:

IF x1 is F1i and, . . . , and xn is Fni ,

THEN fˆ = di x̃k ,

i = 1, 2, . . . , r, where Fki (k = 1, 2, . . . , n) is the fuzzy set, fˆ is a fuzzy representation

of f (x), and di is to be determined later. The final output of the fuzzy system is
inferred as follows:
r
∑ μi (x)di
i=1
fˆ(x) = r x̃k ,
∑ μi (x)
i=1

where μi (x) = ∏nk=1 Fki (x)/ ∑ri=1 ∏nk=1 Fki (x) and Fki (xk ) is the degree of member-
ship of xk in f (x). The membership functions and coefficients can be chosen such
that Fki (xk ) ∈ [0, 1], ∑ri=1 μi (x) = 1, and φ1 (x1 )φ2 (x2 ) · · · φn (xn ) = ∑ri=1 μi (x)di .
The nonlinear term x1 x3 can have the extracted variable as x1 or x3 . Here, we
choose φ (x1 ) = x1 and let x1 be the premise variable of fuzzy rules, which satisfies
x1 (t) ∈ [−d, d]. The fuzzy model which represents exactly the Rössler system is

IF x1 is Fi ,
THEN ẋ(t) = Ai x(t) + bi,

for i = 1, 2, where x(t) = (x1 (t), x2 (t), x3 (t))T , d = 10.5,

⎛ ⎞ ⎛ ⎞
0 −1 −1 0 −1 −1
A1 = ⎝ 1 0.2 0 ⎠ , A2 = ⎝ 1 0.2 0 ⎠,
0 0 5.5 0 0 − 15.5
⎛
⎞
0
b1 = b2 = ⎝ 0 ⎠ ,
0.2
φ (x1 ) = x1 , and the fuzzy sets are chosen as
1 x1  1 x1 
F1 = 1+ , F2 = 1− .
2 10.5 2 10.5
8.2 Modeling Chaotic Systems via T–S Fuzzy Models 275

The chaotic attractor of the fuzzy Rössler system is shown in Fig. 8.2.
Case 3: The Transformed Rössler System (Multiple Nonlinear Terms in a Sys-
tem) is given by ⎧
⎨ ẋ1 = −x2 − exp(x3 ),
⎪
ẋ2 = x1 + ax2,
⎪
⎩
ẋ3 = x1 + c exp(−x3 ) − b,
where a = 0.2, b = 5, and c = 0.2. More than one nonlinear term would be si-
multaneously considered in the system. When a nonlinear m × 1 vector f (x) =
( f1 (x), f2 (x), . . . , fm (x))T is considered, each element of f (x) is assumed to satisfy
f j (x) = φ1 j (x1 )φ2 j (x2 ) · · · φn j (xn )x̃ j , j = 1, 2, . . . , m. According to Cases 1 and 2, the
fuzzy system representing the nonlinear terms is described as follows:

IF x1 is F1i and, . . . , and xn is Fni ,

THEN fˆ = (dk1 x̃k1 , dk2 x̃k2 , . . . , dkm x̃km )T (k = 1, 2, . . . , n),

i = 1, 2, . . . , r. The final output of the fuzzy system is inferred as follows:

r
⎛ r r
⎞T
∑
⎜ i=1 μ i (x)d k1 ∑ μ i (x)d k2 ∑ μ i (x)d km ⎟
fˆ(x) = ⎜
⎝ r
i=1
x̃k1 , r x̃k2 , . . . ,
i=1
r x̃km ⎟
⎠,
∑ μi (x) ∑ μi (x) ∑ μi (x)
i=1 i=1 i=1

25

20

15
x3

10

0
10
5 15
0 10
−5 5
0
−10 −5
x2 −15 −10 x1

Fig. 8.2 The attractor of the fuzzy Rössler system

276 8 Synchronizing Chaotic Systems Based on Fuzzy Models

with μi (x) = ∏nk=1 Fki (xk )/ ∑ri=1 ∏nk=1 Fki (xk ). The remaining procedure is the same
as that for Case 2. It is noted that if the nonlinear terms f j (x), for j = 1, 2, . . . , m,
have a common factor, the number of fuzzy rules may be reduced. Since the Rössler
system depends on the common variable x3 (t), the premise variable is set as x3 (t).
However, the variable x3 (t) of two nonlinear terms cannot be extracted due to
exp(x3 ) exp(−x3 )
lim ∈
/ L∞ and lim ∈/ L∞ , and thus bias terms will appear in
x∈Ω x3 x∈Ω x3
x3 →0 x3 →0
the fuzzy model.
The fuzzy model which represents exactly the transformed Rössler system is

IF x3 is Fi ,
THEN ẋ(t) = Ai x(t) + bi,

i = 1, 2, 3, 4, where x(t) = (x1 (t), x2 (t), x3 (t))T . The transformed fuzzy Rössler sys-
tem is represented exactly by the fuzzy model with r = 4 and
⎛ ⎞
0 −1 0
Ai = ⎝ 1 0.2 0 ⎠ , i = 1, 2, 3, 4,
1 0 0

b1 = (−75, 0, 10)T, b2 = (−75, 0, −20)T,

b3 = (75, 0, 10)T, and b4 = (75, 0, −20)T.
The fuzzy sets are chosen as

0
x3

−2

−4

−6
10
5 15
10
0 5
−5 0
−5
x2 −10 −10 x1

Fig. 8.3 The attractor of the fuzzy transformed Rössler’s system

8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 277

1 exp(x3 ) exp(−x3 )
F1 = 1+ 1+ ,
22 75 75
1 exp(x3 ) exp(−x3 )
F2 = 2 1+ 1− ,
2 75 75
1 exp(x3 ) exp(−x3 )
F3 = 2 1− 1+ ,
2 75 75
1 exp(x3 ) exp(−x3 )
F4 = 2 1− 1− .
2 75 75

The chaotic attractor of the fuzzy Rössler system is shown in Fig. 8.3.

8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy

Models

8.3.1 Fuzzy Modeling of Hyperchaotic Systems

In this subsection, we investigate the T–S fuzzy modeling of many typical hyper-
chaotic systems. In the following hyperchaotic systems, x1 , x2 , x3 , and x4 are state
variables, α , β , θ , σ , e, l, k, h, β1 , and β2 are system parameters. It has been proven
that all the following systems are hyperchaotic if their parameters are chosen ap-
propriately. Assume that x1 ∈ [c̃1 − d˜1 , c̃1 + d˜1], x2 ∈ [c̃2 − d˜2, c̃2 + d˜2 ], d˜1 > 0, and
d˜2 > 0. According to the property of boundedness of hyperchaotic systems, this
assumption is reasonable. The exact values of c̃1 , d˜1 , c̃2 , and d˜2 can be obtained
by simulation. Since there is no input in the autonomous hyperchaotic systems, the
term Bi u(t) in the T–S fuzzy model is omitted in the process of fuzzy modeling of
hyperchaotic systems.
System 1: The Chen Hyperchaotic System [3]
⎧
⎪
⎪ ẋ1 = α (x2 − x1) + x4 ,
⎨
ẋ2 = σ x1 − x1 x3 + θ x2 ,
⎪
⎪ ẋ3 = x1 x2 − β x3,
⎩
ẋ4 = x2 x3 + lx4 .

For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F11 and x2 (t) is F12 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F21 and x2 (t) is F22 , THEN ẋ(t) = A2 x(t) + b2,
R3 : IF x1 (t) is F31 and x2 (t) is F32 , THEN ẋ(t) = A3 x(t) + b3,
R4 : IF x1 (t) is F41 and x2 (t) is F42 , THEN ẋ(t) = A4 x(t) + b4,
278 8 Synchronizing Chaotic Systems Based on Fuzzy Models

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞
−α α 0 1
⎜ σ θ 2(− c̃ ˜
1 − d1 ) 0⎟
A1 = ⎜ ⎝ 0 2(c̃1 + d˜1 )
⎟,
−β 0⎠
0 0 0 l
⎛ ⎞
−α α 0 1
⎜ σ θ 2(− c̃ ˜
1 + d1 ) 0⎟
A2 = ⎜
⎝ 0 2(c̃1 − d˜1 )
⎟,
−β 0⎠
0 0 0 l
⎛ ⎞ ⎛ ⎞
−α α 0 1 −α α 0 1
⎜ σ θ 0 0⎟ ⎜ σ θ 0 0⎟
A3 = ⎜
⎝ 0
⎟ , A4 = ⎜ ⎟,
0 −β 0⎠ ⎝ 0 0 −β 0⎠
0 0 2(c̃2 + d˜2) l 0 0 2(c̃2 − d˜2 ) l
and
b1 = b2 = b3 = b4 = (0 0 0 0)T .
The membership functions are chosen as

1 c̃1 − x1 (t) 1 c̃1 − x1(t)

F11 (x1 (t)) = 1− , F21 (x1 (t)) = 1+ ,
2 ˜
d1 2 d˜1

1 c̃2 − x2(t) 1 c̃2 − x2 (t)

F32 (x2 (t)) = 1− , F42 (x2 (t)) = 1+ ,
2 d˜2 2 d˜2
F12 (x2 (t)) = F22 (x2 (t)) = F31 (x1 (t)) = F41 (x1 (t)) = 1.
System 2: The Liu Hyperchaotic System [7, 14]
⎧
⎪
⎪ ẋ1 = α (x2 − x1),
⎨
ẋ2 = β x1 − kx1 x3 + x4 ,
⎪
⎪ ẋ = −θ x3 + hx21 ,
⎩ 3
ẋ4 = −σ x1 .

For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F1 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F2 , THEN ẋ(t) = A2 x(t) + b2,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 279
⎛ ⎞
−α α 0 0
⎜ β 0 k(−c̃1 − d˜1) 1⎟
A1 = ⎜
⎝ h(c̃1 + d˜1)
⎟,
0 −θ 0⎠
−σ 0 0 0
⎛ ⎞
−α α 0 0
⎜ β 0 k(−c̃1 + d˜1) 1⎟
A2 = ⎜
⎝ h(c̃1 − d˜1)
⎟,
0 −θ 0⎠
−σ 0 0 0
and
b1 = b2 = (0 0 0 0)T .
The membership functions are chosen as

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)

F1 (x1 (t)) = 1− , F2 (x1 (t)) = 1+ .
2 d˜1 2 d˜1
System 3: The Lü Hyperchaotic System (i) [18]
⎧
⎪
⎪ ẋ1 = α (x2 − x1) + x4 ,
⎨
ẋ2 = −x1 x3 + θ x2 ,
⎪ ẋ3 = x1 x2 − β x3,
⎪
⎩
ẋ4 = x1 x3 + lx4 .

For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F1 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F2 , THEN ẋ(t) = A2 x(t) + b2,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞ ⎛ ⎞
−α α 0 1 −α α 0 1
⎜ 0 θ − c̃ − d˜1 0 ⎟ ⎜ 0 θ − c̃ + d˜1 0⎟
A1 = ⎜⎝ 0 c̃1 + d˜1
1 ⎟ , A2 = ⎜
⎠ ⎝
1 ⎟,
−β 0 ˜
0 c̃1 − d1 −β 0⎠
0 0 c̃1 + d˜1 l 0 0 c̃1 − d˜1 l

and
b1 = b2 = (0 0 0 0)T .
The membership functions are chosen as

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)

F1 (x1 (t)) = 1− , F2 (x1 (t)) = 1+ .
2 ˜
d1 2 d˜1
280 8 Synchronizing Chaotic Systems Based on Fuzzy Models

System 4: The Lü Hyperchaotic System (ii) [4]

⎧
⎪
⎪ ẋ1 = α (x2 − x1 ),
⎨
ẋ2 = −x1 x3 + θ x2 + x4 ,
⎪
⎪ ẋ3 = x1 x2 − β x3 ,
⎩
ẋ4 = x3 − σ x4 .

For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F1 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F2 , THEN ẋ(t) = A2 x(t) + b2,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞ ⎛ ⎞
−α α 0 0 −α α 0 0
⎜ 0 θ −c̃1 − d˜1 1 ⎟ ⎜ θ −c̃1 + d˜1 1 ⎟
A1 = ⎜ ⎟ , A2 = ⎜ 0 ⎟,
⎝ 0 c̃1 + d˜1 −β 0 ⎠ ⎝ ˜
0 c̃1 − d1 −β 0 ⎠
0 0 1 −σ 0 0 1 −σ

and
b1 = b2 = (0 0 0 0)T .
The membership functions are chosen as

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)

F1 (x1 (t)) = 1− , F2 (x1 (t)) = 1+ .
2 d˜1 2 d˜1
System 5: The Modified Lü Hyperchaotic System [15]
⎧
⎪
⎪ ẋ1 = α (x2 − x1 + x2 x3 ),
⎨
ẋ2 = −x1 x3 + β x2 + x4,
⎪
⎪ ẋ3 = x1 x2 − θ x3 ,
⎩
ẋ4 = −kx1 .
For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F11 and x2 (t) is F12 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F21 and x2 (t) is F22 , THEN ẋ(t) = A2 x(t) + b2,
(8.4)
R3 : IF x1 (t) is F31 and x2 (t) is F32 , THEN ẋ(t) = A3 x(t) + b3,
R4 : IF x1 (t) is F41 and x2 (t) is F42 , THEN ẋ(t) = A4 x(t) + b4,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 281
⎛ ⎞
−α α 0 0
⎜ 0 β 2(− c̃ − d˜1) 1⎟
A1 = ⎜
⎝ 0 2(c̃1 + d˜1 )
1 ⎟,
−θ 0⎠
−k 0 0 0
⎛ ⎞
−α α 0 0
⎜ 0 β 2(− c̃ + d˜1) 1⎟
A2 = ⎜
⎝ 0 2(c̃1 − d˜1 )
1 ⎟,
−θ 0⎠
−k 0 0 0
⎛ ⎞ ⎛ ⎞
−α α 2α (c̃2 + d˜2) 0 −α α 2α (c̃2 − d˜2) 0
⎜ 0 β 0 1 ⎟ ⎜ 0 β 0 1⎟
A3 = ⎜
⎝ 0 0
⎟ , A4 = ⎜ ⎟,
−θ 0⎠ ⎝ 0 0 −θ 0⎠
−k 0 0 0 −k 0 0 0
and
b1 = b2 = b3 = b4 = (0 0 0 0)T .
The membership functions are chosen as

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)

F11 (x1 (t)) = 1− , F21 (x1 (t)) = 1+ ,
2 ˜
d1 2 d˜1

1 c̃2 − x2 (t) 1 c̃2 − x2 (t)

F32 (x2 (t)) = 1− , F42 (x2 (t)) = 1+ ,
2 ˜
d2 2 d˜2
F12 (x2 (t)) = F22 (x2 (t)) = F31 (x1 (t)) = F41 (x1 (t)) = 1.
System 6: The Lorenz Hyperchaotic System [6, 7]
⎧
⎪ ẋ1 = −α (x1 − x2 ) + x4,
⎪
⎨
ẋ2 = lx1 − x2 − x1 x3 ,
⎪
⎪ ẋ 3 = x1 x2 − β x3 ,
⎩
ẋ4 = −x1 x3 + σ x4 .

For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F1 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F2 , THEN ẋ(t) = A2 x(t) + b2,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞ ⎛ ⎞
−α α 0 1 −α α 0 1
⎜ l −1 − c̃ − d˜1 0 ⎟ ⎜ l −1 − c̃ + d˜1 0⎟
A1 = ⎜
⎝ 0 c̃1 + d˜1
1 ⎟ , A2 = ⎜
⎠ ⎝
1 ⎟,
−β 0 ˜
0 c̃1 − d1 −β 0⎠
0 0 −c̃1 − d˜1 σ 0 0 −c̃1 + d˜1 σ
282 8 Synchronizing Chaotic Systems Based on Fuzzy Models

and
b1 = b2 = (0 0 0 0)T .
The membership functions are chosen as

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)

F1 (x1 (t)) = 1− , F2 (x1 (t)) = 1+ .
2 d˜1 2 d˜1
System 7: A New Hyperchaotic System [2]
⎧
⎪
⎪ ẋ1 = α (x2 − x1 ) + ex2 x3 ,
⎨
ẋ2 = θ x1 + x2 − σ x1x3 + x4,
⎪ ẋ3 = x1 x2 − β x3 ,
⎪
⎩
ẋ4 = −kx2 .

For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F11 and x2 (t) is F12 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F21 and x2 (t) is F22 , THEN ẋ(t) = A2 x(t) + b2,
(8.5)
R3 : IF x1 (t) is F31 and x2 (t) is F32 , THEN ẋ(t) = A3 x(t) + b3,
R4 : IF x1 (t) is F41 and x2 (t) is F42 , THEN ẋ(t) = A4 x(t) + b4,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞
−α α 0 0
⎜ θ 1 2 σ (− c̃ ˜
1 − d1 ) 1⎟
A1 = ⎜ ⎝ 0 2(c̃1 + d˜1)
⎟,
−β 0⎠
0 −k 0 0
⎛ ⎞
−α α 0 0
⎜ θ 1 2σ (−c̃1 + d˜1 ) 1 ⎟
A2 = ⎜
⎝ 0 2(c̃1 − d˜1)
⎟,
−β 0⎠
0 −k 0 0
⎛ ⎞
−α α 2e(c̃2 + d˜2) 0
⎜ θ 1 0 1⎟
A3 = ⎜
⎝ 0
⎟,
0 −β 0⎠
0 −k 0 0
⎛ ⎞
−α α 2e(c̃2 − d˜2) 0
⎜ θ 1 0 1⎟
A4 = ⎜
⎝ 0
⎟,
0 −β 0⎠
0 −k 0 0
and
b1 = b2 = b3 = b4 = (0 0 0 0)T .
8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 283

The membership functions are chosen as

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)

F11 (x1 (t)) = 1− , F21 (x1 (t)) = 1+ ,
2 ˜
d1 2 d˜1

1 c̃2 − x2 (t) 1 c̃2 − x2 (t)

F32 (x2 (t)) = 1− , F42 (x2 (t)) = 1+ ,
2 ˜
d2 2 d˜2
F12 (x2 (t)) = F22 (x2 (t)) = F31 (x1 (t)) = F41 (x1 (t)) = 1.
System 8: Another New Hyperchaotic System [5]
⎧
⎪
⎪ ẋ1 = α (x2 − x1 ),
⎨
ẋ2 = σ x1 + θ x2 − x1x3 − x4,
⎪ ẋ3 = x1 x2 − β x3 ,
⎪
⎩
ẋ4 = x1 + k.

For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F1 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F2 , THEN ẋ(t) = A2 x(t) + b2,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞
−α α 0 0
⎜ σ θ −c̃1 − d˜1 −1 ⎟
A1 = ⎜ ⎝ 0 c̃1 + d˜1
⎟,
−β 0 ⎠
1 0 0 0
⎛ ⎞
−α α 0 0
⎜ σ θ −c̃1 + d˜1 −1 ⎟
A2 = ⎜
⎝ 0 c̃1 − d˜1
⎟,
−β 0 ⎠
1 0 0 0
and
b1 = b2 = (0 0 0 k)T .
The membership functions are chosen as:

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)

F1 (x1 (t)) = 1− , F2 (x1 (t)) = 1+ .
2 d˜1 2 d˜1
284 8 Synchronizing Chaotic Systems Based on Fuzzy Models

System 9: The Generalized Hénon–Heiles System [16]

⎧
⎪
⎪ ẋ1 = x3 ,
⎨
ẋ2 = x4 ,
⎪
⎪ ẋ3 = −x1 − 2x1x2 ,
⎩
ẋ4 = −x21 − x2 + x22 .

For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F11 and x2 (t) is F12 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F21 and x2 (t) is F22 , THEN ẋ(t) = A2 x(t) + b2,
R3 : IF x1 (t) is F31 and x2 (t) is F32 , THEN ẋ(t) = A3 x(t) + b3,
R4 : IF x1 (t) is F41 and x2 (t) is F42 , THEN ẋ(t) = A4 x(t) + b4,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞
0 0 1 0
⎜ 0 0 0 1⎟
A1 = ⎜ ⎝
⎟,
−1 −4(c̃1 + d˜1 ) 0 0⎠
−2(c̃1 + d˜1) −1 0 0
⎛ ⎞
0 0 1 0
⎜ 0 0 0 1⎟
A2 = ⎜
⎝
⎟,
−1 −4(c̃1 − d1 ) 0 0 ⎠
˜
−2(c̃1 − d˜1) −1 0 0
⎛ ⎞
0 0 1 0
⎜ 0 0 0 1⎟
A3 = ⎜
⎝ −1
⎟,
0 0 0⎠
0 −1 + 2(c̃2 + d˜2) 0 0
⎛ ⎞
0 0 1 0
⎜ 0 0 0 1⎟
A4 = ⎜
⎝ −1
⎟,
0 0 0⎠
0 −1 + 2(c̃2 − d˜2) 0 0
and
b1 = b2 = b3 = b4 = (0 0 0 0)T .
The membership functions are chosen as

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)

F11 (x1 (t)) = 1− , F21 (x1 (t)) = 1+ ,
2 ˜
d1 2 d˜1

1 c̃2 − x2 (t) 1 c̃2 − x2 (t)

F32 (x2 (t)) = 1− , F42 (x2 (t)) = 1+ ,
2 ˜
d2 2 d˜2
8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 285

F12 (x2 (t)) = F22 (x2 (t)) = F31 (x1 (t)) = F41 (x1 (t)) = 1.
System 10: The Modified Rössler Hyperchaotic System (i) [10]
⎧
⎪
⎪ ẋ1 = −x2 − x3 ,
⎨
ẋ2 = x1 + α x2 + x4 ,
(8.6)
⎪
⎪ ẋ3 = β + β1x1 + x1x3 ,
⎩
ẋ4 = −θ x3 + σ x4 .

For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F1 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F2 , THEN ẋ(t) = A2 x(t) + b2,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞
0 −1 −1 0
⎜1 α 0 1⎟
A1 = ⎜ ⎝ β1 0 c̃1 + d˜1
⎟,
0⎠
0 0 −θ σ
⎛ ⎞
0 −1 −1 0
⎜1 α 0 1⎟
A2 = ⎜
⎝ β1 0 c̃1 − d˜1
⎟,
0⎠
0 0 −θ σ
and
b1 = b2 = (0 0 β 0)T .
The membership functions are chosen as

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)

F1 (x1 (t)) = 1− , F2 (x1 (t)) = 1+ .
2 ˜
d1 2 d˜1
System 11: The Modified Rössler Hyperchaotic System (ii) [10]
⎧
⎪ ẋ1 = −x2 − x3 ,
⎪
⎨
ẋ2 = x1 + α x2 + x4 ,
⎪
⎪ ẋ3 = β + β2x3 + x1x3 ,
⎩
ẋ4 = −θ x3 + σ x4 .

For the above hyperchaotic system, we derive the following T–S fuzzy model to
represent it exactly:

R1 : IF x1 (t) is F1 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F2 , THEN ẋ(t) = A2 x(t) + b2,
286 8 Synchronizing Chaotic Systems Based on Fuzzy Models

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞ ⎛ ⎞
0 −1 −1 0 0 −1 −1 0
⎜1 α 0 1 ⎟ ⎜1 α 0 1⎟
A1 = ⎜ ⎟ ⎜
⎝ 0 0 β2 + c̃1 + d˜1 0 ⎠ , A2 = ⎝ 0 0 β2 + c̃1 − d˜1
⎟,
0⎠
0 0 −θ σ 0 0 −θ σ

and
b1 = b2 = (0 0 β 0)T .
The membership functions are chosen as

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)

F1 (x1 (t)) = 1− , F2 (x1 (t)) = 1+ .
2 ˜
d1 2 d˜1
To summarize, all the aforementioned hyperchaotic systems can be represented
exactly by the following T–S fuzzy models:

Ri : IF z(t) is Fi , THEN ẋ(t) = Ai x(t) + bi,

i = 1, 2, . . . , r, where the premise variable z(t) is a proper state variable, bi is a con-

stant vector, and r is the number of fuzzy IF–THEN rules. Then, applying the prod-
uct inference rule, a singleton fuzzifier, and the center of gravity defuzzifier to the
above fuzzy rule base, the overall fuzzy hyperchaotic systems can be inferred as
r
ẋ(t) = ∑ μi (z(t))[Ai x(t) + bi], (8.7)
i=1

where
Fi (z(t))
μi (z(t)) = r .
∑i=1 Fi (z(t))
Next, we give a systematic modeling method to represent exactly the above hy-
perchaotic systems by T–S fuzzy models. Note that all the aforementioned hyper-
chaotic systems can be expressed as follows:

ẋ(t) = Ax(t) + f (x) + b, (8.8)

where Ax(t) and f (x) denote the linear part and the nonlinear part of hyperchaotic
systems, respectively, and b is the constant vector of the hyperchaotic systems. Next,
fuzzy modeling methods are summarized for several cases of the nonlinear part f (x),
as follows.
Case 1: Only one nonlinear term in f (x). Systems 10 and 11 belong to this
case. For this case, we can choose any state variable in the nonlinear term as premise
variable. For example, in system 10, the unique nonlinear term is x1 x3 . Then, we can
choose either x1 or x3 as premise variable. If we choose x1 as premise variable, then
the membership functions are defined as
8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 287

1 c̃1 − x1 (t)
F1 (x1 (t)) = 1−
2 d˜1
and
1 c̃1 − x1 (t)
F2 (x1 (t)) = 1+ ,
2 d˜1
and the fuzzy IF–THEN rules are defined as:

R1 : IF x1 (t) is F1 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F2 , THEN ẋ(t) = A2 x(t) + b2,
where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T . For the above fuzzy rule base, by taking
a standard fuzzy inference strategy, i.e., using a singleton fuzzifier, product fuzzy
inference, and center average defuzzifier, we get the following T–S fuzzy system:

∑2i=1 Fi (x1 (t))(Ai x(t) + bi )

ẋ(t) = . (8.9)
∑2i=1 Fi (x1 (t))
The rest of the task is to choose appropriate A1 , A2 , b1 , and b2 such that (8.9) can be
equivalent to (8.8). For system 10, each part of (8.8) is presented as follows:
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 −1 −1 0 0 0
⎜1 α 0 1⎟ ⎜ 0 ⎟ ⎜0 ⎟
A=⎜ ⎝ β1 0
⎟ , f (x) = ⎜ ⎟ ⎜ ⎟
⎝ x1 x3 ⎠ , and b = ⎝ β ⎠ .
0 0⎠
0 0 −θ σ 0 0

Denote A1 = A1 + A1 , A2 = A2 + A2 , and let A1 = A2 = A and b1 = b2 = β . Substi-
tuting them into (8.9), and making (8.9) equal to (8.8), we get the values of A1 and
A2 as follows:
⎛ ⎞ ⎛ ⎞
0 0 0 0 0 0 0 0
⎜0 0 0 0⎟ ⎜0 0 0 0⎟
A1 = ⎜ ⎟  ⎜
⎝ 0 0 c̃1 + d˜1 0 ⎠ , A2 = ⎝ 0 0 c̃1 − d˜1 0 ⎠ .
⎟

0 0 0 0 0 0 0 0

Consequently, we derive the exact T–S fuzzy models for this class of hyperchaotic
systems.
Case 2: More than one nonlinear term in f (x), but there is a common factor in
these nonlinear terms. Systems 2, 3, 4, 6, and 8 belong to this case. For this case,
we can choose the common factor in the nonlinear terms as premise variable. The
remaining process of T–S fuzzy modeling of this class of hyperchaotic systems is
similar to Case 1, and therefore it is omitted.
Case 3: More than one nonlinear term in f (x), and there is no common factor
in these nonlinear terms. Systems 1, 5, 7, and 9 belong to this case. For this case, we
have to choose two state variables occurring in f (x) as premise variables. For exam-
ple, in system 1, we can choose x1 and x2 as premise variables and the membership
288 8 Synchronizing Chaotic Systems Based on Fuzzy Models

functions are defined as

1 c̃1 − x1 (t) 1 c̃1 − x1 (t)
F11 (x1 (t)) = 1− , F21 (x1 (t)) = 1+ ,
2 d˜1 2 d˜1

1 c̃2 − x2 (t) 1 c̃2 − x2 (t)

F32 (x2 (t)) = 1− , F42 (x2 (t)) = 1+ ,
2 ˜
d2 2 d˜2
F12 (x2 (t)) = F22 (x2 (t)) = F31 (x1 (t)) = F41 (x1 (t)) = 1.
Furthermore, the following fuzzy IF–THEN rules are presented:

R1 : IF x1 (t) is F11 and x2 (t) is F12 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F21 and x2 (t) is F22 , THEN ẋ(t) = A2 x(t) + b2,
R3 : IF x1 (t) is F31 and x2 (t) is F32 , THEN ẋ(t) = A3 x(t) + b3,
R4 : IF x1 (t) is F41 and x2 (t) is F42 , THEN ẋ(t) = A4 x(t) + b4,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T .

For the above fuzzy rule base, by taking a standard fuzzy inference strategy, i.e.,
using a singleton fuzzifier, product fuzzy inference, and center average defuzzifier,
we get the following T–S fuzzy system:

∑4i=1 ∏2j=1 Fi j (x j (t)) (Ai x(t) + b)

ẋ(t) = . (8.10)
∑4i=1 ∏2j=1 Fi j (x j (t))

The remaining task is to choose appropriate Ai and bi (i = 1, 2, 3, 4) such that

(8.10) can be equivalent to (8.8). For system 1, each part of (8.8) is represented as
follows:
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
−α α 0 1 0 0
⎜ σ θ 0 0 ⎟ ⎜ −x1 x3 ⎟ ⎜0⎟
A=⎜ ⎟ ⎜ ⎟
⎝ 0 0 − β 0 ⎠ , f (x) = ⎝ x1 x2 ⎠ , and b = ⎝ 0 ⎠ .
⎜ ⎟

0 0 0 l x2 x3 0

Denote Ai = Ai + Ai , and let Ai = A, bi = b (i = 1, 2, 3, 4). Substituting them into
(8.10) and making (8.10) equal to (8.8), we get the values of Ai as follows:
⎛ ⎞
0 0 0 0
⎜0 0 2(−c̃1 − d˜1 ) 0 ⎟
A1 = ⎜
⎝ 0 2(c̃1 + d˜1)
⎟,
0 0⎠
0 0 0 0
⎛ ⎞
0 0 0 0
⎜ 0 0 2(− c̃ ˜
1 + d1 ) 0⎟
A2 = ⎜
⎝ 0 2(c̃1 − d˜1)
⎟,
0 0⎠
0 0 0 0
8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 289
⎛ ⎞ ⎛ ⎞
0 0 0 0 0 0 0 0
⎜0 0 0 0⎟ ⎜0 0 0 0⎟
A3 =⎜
⎝0
⎟ , and A4 = ⎜
 ⎟.
0 0 0⎠ ⎝0 0 0 0⎠
0 0 2(c̃2 + d˜2 ) 0 0 0 2(c̃2 − d˜2) 0
Consequently, we derive the exact T–S fuzzy model of this class of hyperchaotic
systems.

8.3.2 Synchronization of Two Different Hyperchaotic Systems

In this section, we will investigate the synchronization of two different hyperchaotic

systems based on the T–S fuzzy models [17]. First, an assumption is given as fol-
lows.
Consider the fuzzy hyperchaotic system (8.7) as the drive system, and the fol-
lowing fuzzy hyperchaotic system as the response system:
r
˙ = ∑ hi (ẑ(t))[Âi x̂(t) + b̂i] + u(t).
x̂(t) (8.11)
i=1

Suppose in (8.11) b̂i = b̂ and in (8.7) bi = b. Based on the PDC technique, a fuzzy
controller can be designed to realize the synchronization as follows:
Subcontroller u1 (t):

Ri : IF z(t) is Fi , THEN u1 (t) = Ki x(t), i = 1, 2, . . . , r,

Subcontroller u2 (t):

Ri : IF ẑ(t) is Fi , THEN u2 (t) = −Hi x̂(t), i = 1, 2, . . . , r.

The overall fuzzy controller is constructed by a parallel connection, i.e.,

r r
u(t) = u1 (t) + u2(t) = ∑ μi (z(t))Ki x(t) − ∑ hi (ẑ(t))Hi x̂(t). (8.12)
i=1 i=1

Denote the error signal as e(t) = x(t) − x̂(t). Substituting (8.12) into (8.11), we can
derive the closed-loop synchronization error system as follows:
r r
ė(t) = ∑ μi (z(t))(Ai − Ki )e(t) + ∑ hi (ẑ(t))(Âi − Hi )e(t) + ϖ (t), (8.13)
i=1 i=1

where
r r
ϖ (t) = ∑ μi (z(t))(Ai − Ki )x̂(t) − ∑ hi (ẑ(t))(Âi − Hi )x(t) + (b − b̂).
i=1 i=1
290 8 Synchronizing Chaotic Systems Based on Fuzzy Models

For the synchronization error system (8.13), consider the following H∞ perfor-
mance index:
 tf  tf
1 γ 2 + γ22
e (t)e(t)dt < eT (0)Pe(0) + 1
T
ϖ T (t)ϖ (t)dt, (8.14)
0 2 2 0

where γ1 > 0 and γ2 > 0 are prescribed attenuation levels, and t f is the terminal
time.

Theorem 8.1. Consider the fuzzy error system (8.13). For the given constants γ1 > 0
and γ2 > 0, if there exist P = PT > 0, Xi , and Yi with appropriate dimensions, such
that the following linear matrix inequalities (LMIs) hold:

PAi + ATi P − XiT − Xi + I 12 P

1 <0 (8.15)
2P −γ12 I

and
PÂi + ÂTi P − YiT − Yi + I 12 P
1 < 0, (8.16)
2P −γ22 I
then we can choose Ki = P−1 Xi and Hi = P−1Yi in (8.12), such that the robust per-
formance (8.14) is guaranteed for i = 1, 2, . . . , r.

Proof. Now, we select the Lyapunov function as

V (t) = eT (t)Pe(t),

where P is a positive-definite matrix. The time derivative of V (t) along the trajectory
of (8.13) is given by

V̇ (t) = ėT (t)Pe(t) + eT(t)Pė(t)

= eT (t) (Ξ + Π )T Pe(t) + eT(t)P (Ξ + Π ) e(t)
+ ϖ T (t)Pe(t) + eT(t)Pϖ (t)
' (
T
e(t) (Ξ + Π )T P + P(Ξ + Π ) P e(t)
=
ϖ (t) P 0 ϖ (t)
' (
T
e(t) Ξ T P + PΞ + Π T P + PΠ P e(t)
= ,
ϖ (t) P 0 ϖ (t)

where
r
Ξ = ∑ μi (z(t))(Ai − Ki )
i=1

and
r
Π = ∑ hi (ẑ(t))(Âi − Hi ).
i=1
8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 291

Based on (8.14), define

 tf  tf
J= [2eT (t)e(t) − (γ12 + γ22 )ϖ T (t)ϖ (t) + V̇ (t)]dt − V̇ (t)dt.
0 0

We have
 tf
J= [2eT (t)e(t) − (γ12 + γ22 )ϖ T (t)ϖ (t) + V̇ (t)]dt − V (t f ) + V (0)
0
 tf r T 1
e(t) Ω 2P e(t)
≤ ∑ μi (z(t)) ϖ (t) 1
−γ12 I ϖ (t)
0 i=1 2P
-
r T 1
e(t) Θ 2P e(t)
+ ∑ hi(ẑ(t)) ϖ (t) 1
−γ22 I ϖ (t)
dt + V (0), (8.17)
i=1 2P

where
Ω = (Ai − Ki )T P + P(Ai − Ki ) + I
and
Θ = (Âi − Hi )T P + P(Âi − Hi ) + I.
It is easy to see that if the following inequalities hold:
1
Ω 2P
1 <0 (8.18)
2P −γ12 I

and
1
Θ 2P
1 < 0, (8.19)
2P −γ22 I
then we have
 tf  
J= 2eT (t)e(t) − (γ12 + γ22 )ϖ T (t)ϖ (t) dt ≤ V (0). (8.20)
0

From (8.20), it is obvious that the H∞ performance index (8.14) is guaranteed. De-
note Xi = PKi and Yi = PHi . Then, (8.18) and (8.19) are equivalent to the LMIs
(8.15) and (8.16), respectively. This completes the proof. 

8.3.3 Simulations

To illustrate the effectiveness of the theoretical analysis and design method, some
examples are included for illustration. First, we investigate the T–S fuzzy modeling
of system 2 (the Liu hyperchaotic system) and system 6 (the Lorenz hyperchaotic
system). When the parameters are chosen as α = 10, β = 40, θ = 2.5, σ = 10.6,
k = 1, and h = 4, system 2 has two positive Lyapunov exponents λ1 = 1.1491 and
λ2 = 0.12688. Thus, system 2 is hyperchaotic in this case, whose attractors are
292 8 Synchronizing Chaotic Systems Based on Fuzzy Models

150

100

50
x3

−50
40
20 20
0 10
0
−20
−10
x2 −40 −20 x1

Fig. 8.4 Projection of the Liu hyperchaotic attractor on x1 − x2 − x3

0.5

0
x4

−0.5

−1
40
20 20
0 10
0
−20
−10
x2 −40 −20 x1

Fig. 8.5 Projection of the Liu hyperchaotic attractor on x1 − x2 − x4

8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 293

shown in Figs. 8.4 and 8.5. The initial state vector is chosen as x(0) = (1, 1, −1, 0)T.
From the simulation we get x1 ∈ [c̃1 − d˜1 , c̃1 + d˜1 ], where c̃1 = 1.2347 and d˜1 =
18.6833. Then, we can derive the T–S fuzzy models of system 2 as follows:

R1 : IF x1 (t) is F1 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F2 , THEN ẋ(t) = A2 x(t) + b2,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞ ⎛ ⎞
−10 10 0 0 −10 10 0 0
⎜ 40 0 −19.9180 1 ⎟ ⎜ 40 0 17.4486 1 ⎟
A1 = ⎜
⎝ 79.6720 0
⎟ , A2 = ⎜ ⎟,
−2.5 0⎠ ⎝ −69.7944 0 −2.5 0 ⎠
−10.6 0 0 0 −10.6 0 0 0

and
b1 = b2 = (0 0 0 0)T .
The membership functions are chosen as

1 1.2347 − x1(t) 1 1.2347 − x1(t)

F1 (x1 (t)) = 1− , F2 (x1 (t)) = 1+ .
2 18.6833 2 18.6833

When the parameters are chosen as α = 10, l = 28, β = 8/3, and σ = 1.3, system
6 has two positive Lyapunov exponents λ1 = 0.39854 and λ2 = 0.24805. Thus,
system 6 is hyperchaotic in this case, which is shown in Figs. 8.6 and 8.7. The
initial state vector is chosen as x(0) = (1, 2, 1, 2)T . From the simulation we get x1 ∈
[c̃1 − d˜1 , c̃1 + d˜1 ], where c̃1 = −1.4366 and d˜1 = 28.4241. Then, we can derive the
T–S fuzzy models of systems 6 as follows:

R1 : IF x1 (t) is F1 , THEN ẋ(t) = A1 x(t) + b1,

R2 : IF x1 (t) is F2 , THEN ẋ(t) = A2 x(t) + b2,

where x(t) = (x1 (t), x2 (t), x3 (t), x4 (t))T ,

⎛ ⎞
−10 10 0 1
⎜ 28 −1 −26.9875 0 ⎟
A1 = ⎜ ⎝ 0 26.9875
⎟,
−8/3 0 ⎠
0 0 −26.9875 1.3
⎛ ⎞
−10 10 0 1
⎜ 28 −1 29.8607 0 ⎟
A2 = ⎜
⎝ 0 −29.8607 −8/3
⎟,
0 ⎠
0 0 29.8607 1.3
and
b1 = b2 = (0 0 0 0)T .
294 8 Synchronizing Chaotic Systems Based on Fuzzy Models

50

40

30
x3

20

10

0
40
20 40
0 20
0
−20
−20
x2 −40 −40 x1

Fig. 8.6 Projection of the Lorenz hyperchaotic attractor on x1 − x2 − x3

200

100

0
x4

−100

−200
40
20 40
0 20
0
−20
−20
x2 −40 −40 x1

Fig. 8.7 Projection of the Lorenz hyperchaotic attractor on x1 − x2 − x4

8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 295

The membership functions are chosen as

1 −1.4366 − x1(t)
F1 (x1 (t)) = 1− ,
2 28.4241

1 −1.4366 − x1(t)
F2 (x1 (t)) = 1+ .
2 28.4241
In the following, based on the above T–S fuzzy hyperchaotic models, we in-
vestigate the synchronization between the Liu hyperchaotic system and the Lorenz
hyperchaotic system. We consider the Liu hyperchaotic system as the drive sys-
tem and the Lorenz hyperchaotic system as the response system. The initial con-
ditions of the drive system and the response system are x(0) = (15, 25, 100, 0)T
and x̂(0) = (−25, −20, 30, 100)T, respectively. According to Theorem 8.1, choos-
ing γ1 = γ2 = 0.1, we get the positive-definite matrix P and feedback gains Ki and
Hi (i = 1, 2) computed by LMI Control Toolbox in Matlab R
7.0 as follows:
⎛ ⎞
1.6093 −0.0410 −0.0022 −0.3284
⎜ −0.0410 0.3789 −0.0044 0.0092 ⎟
P=⎜ ⎝ −0.0022 −0.0044 0.1059 −0.0003 ⎠ ,
⎟

−0.3284 0.0092 −0.0003 0.2155

⎛ ⎞
39.3794 13.5798 −0.2381 76.9472
⎜ 40.6736 142.2773 −1.6448 −6.2164 ⎟
K1 = ⎜
⎝ 29.4006 −22.1059 516.3327
⎟,
−25.0879 ⎠
68.0431 5.3914 −13.0605 371.6015
⎛ ⎞
38.9603 14.2497 2.9000 77.1718
⎜ 40.6733 143.8961 14.6389 −5.0789 ⎟
K2 = ⎜
⎝ −27.7512 37.8918 514.4715
⎟,
31.3370 ⎠
66.1391 7.3065 18.7104 372.9590
⎛ ⎞
35.7096 13.3922 2.1372 72.7527
⎜ 39.6357 134.9720 −3.9529 −5.4449 ⎟
H1 = ⎜
⎝ 51.7616 −29.9640 487.2763
⎟,
−22.7680 ⎠
62.5319 6.6327 −8.1879 350.9440
and ⎛ ⎞
35.3859 13.8039 0.4756 72.9124
⎜ 38.7566 136.4886 16.1391 −4.6764 ⎟
H2 = ⎜ ⎟
⎝ −50.9872 44.6941 485.6173 29.9071 ⎠ .
61.7018 7.5811 14.2738 351.4143
The synchronization error curves are presented in Figs. 8.8–8.11, which show
that H∞ synchronization between two different hyperchaotic systems is achieved.
296 8 Synchronizing Chaotic Systems Based on Fuzzy Models

Fig. 8.8 Synchronization error e1 (t) between two different hyperchaotic systems

Fig. 8.9 Synchronization error e2 (t) between two different hyperchaotic systems
8.3 Synchronization of Hyperchaotic Systems via T–S Fuzzy Models 297

Fig. 8.10 Synchronization error e3 (t) between two different hyperchaotic systems

Fig. 8.11 Synchronization error e4 (t) between two different hyperchaotic systems
298 8 Synchronizing Chaotic Systems Based on Fuzzy Models

8.4 Synchronizing Fuzzy Chaotic Systems with Time-Varying

Delays

This section deals with the synchronization problem of chaotic systems with time-
varying delays and unknown parameters. A fuzzy controller is developed and pre-
sented via T–S fuzzy models. By using an appropriate Lyapunov–Krasovskii func-
tional, we derive synchronization conditions in the form of an LMI. The advantages
of the approach are that the structurer of the fuzzy controller is simple and this
method can be applied to most of the synchronization control of chaotic systems.
A numerical example is given to demonstrate the validity of the proposed synchro-
nization approach.

8.4.1 Basic Formulation

A T–S fuzzy system is described by a set of fuzzy IF–THEN rules where each rule
locally represents a linear input–output realization of the system over a certain re-
gion of the state space. For the synchronization problem, the drive and response
systems are designed such that the internal states of the drive and response systems
are to be the same. Let the chaotic response system be described by a T–S fuzzy
model with the following rule-base:

Ri : IF z1 (t) is Mi1 , . . . , and zg (t) is Mig ,

THEN ẋ(t) = Ai x(t) + Bi x(t − τ (t)) + ε (t) + u(t),

i = 1, 2, . . . , l, where x(t) ∈ Rn is the state vector of the response system, z(t) =

(z1 (t), z2 (t), . . . , zg (t))T is a vector of the measurable premise variables, Mir (r =
1, 2, . . . , g) are fuzzy sets, l is the number of fuzzy rules, Ai and Bi are some constant
matrices of compatible dimensions, u(t) ∈ Rn is the control input vector, ε (t) is the
oscillated force or constant value in the chaotic dynamics, and τ (t) is a time-varying
delay satisfying 0 < τ (t) ≤ τ0 and 0 ≤ τ̇ (t) ≤ h < 1.
Using the singleton fuzzifier, product fuzzy inference, and weighted average de-
fuzzifier, the output of the above fuzzy response system is inferred as follows:
l
ẋ(t) = ∑ vi (z(t))[Ai x(t) + Bix(t − τ (t))] + ε (t) + u(t), (8.21)
i=1

where
l
ωi (z(t))
vi (z(t)) = l
, vi (z(t)) ≥ 0, ∑ vi (z(t)) = 1,
∑ ωi (z(t)) i=1
i=1
g l
ωi (z(t)) = ∏ Mi j (z(t)), ∑ ωi (z(t)) > 0, ωi (z(t)) ≥ 0,
j=1 i=1
8.4 Synchronizing Fuzzy Chaotic Systems with Time-Varying Delays 299

and Mi j (z(t)) is the degree of membership of z(t) in Mi j .

Similar to the fuzzy response system, the fuzzy drive system can be described by
the following rules:

Ri : IF ẑ1 (t) is Mi1 , . . . , and ẑg (t) is Mig ,

THEN ẏ(t) = Âi y(t) + B̂i y(t − τ (t)) + ε (t),

i = 1, . . . , l, where y(t) ∈ Rn is the state vector of the drive system, ẑ(t) = [ẑ1 (t), ẑ2 (t),
. . . , ẑg (t)]T is a vector of the premise variables, and Âi and B̂i are some unknown
constant matrices of compatible dimensions. The overall fuzzy drive system can be
inferred as
l
ẏ(t) = ∑ λi (ẑ(t))[Âi y(t) + B̂i y(t − τ (t))] + ε (t). (8.22)
i=1

Denote the synchronization error signal by e(t) = x(t) − y(t). From (8.21) and
(8.22), the error dynamics of e(t) is arranged as

ė(t) =ẋ(t) − ẏ(t)

l
= ∑ vi (z(t) [Ai e(t) + Bi e(t − τ (t))]
i=1
' (
l l 7 8
+ ∑ vi (z(t)) − ∑ λi(ẑ(t)) Âi y(t) + B̂iy(t − τ (t))
i=1 i=1
l (8.23)
+ ∑ vi (z(t))[(Ai − Âi )y(t) + (Bi − B̂i )y(t − τ (t))] + u(t)
i=1
l
= ∑ vi (z(t))[Ai e(t) + Bie(t − τ (t))]
i=l
l
+ ∑ vi (z(t))[(Ai − Âi )y(t) + (Bi − B̂i )y(t − τ (t))] + ϖ (t) + u(t),
i=1

where
' (
l l
ϖ (t) = ∑ vi (z(t)) − ∑ λi(ẑ(t)) [Âi y(t) + B̂iy(t − τ (t))].
i=1 i=1

The purpose of this section is to design adaptive control law such that the drive
and response systems are synchronized; that is,

lim e(t) = 0.
t→∞
300 8 Synchronizing Chaotic Systems Based on Fuzzy Models

8.4.2 Main Result

In order to achieve the synchronization of the two systems, let us design the follow-
ing fuzzy controller based on the PDC technique,

Ri : IF z1 (t) is Mi1 , . . . , and zg (t) is Mig ,

THEN u(t) = −Ki e(t) + (−Ai + Ci (t))y(t) + (−Bi + Di (t))y(t − τ (t)),

i = 1, 2, . . . , l, where Ki are the feedback gains to be designed later and Ci (t) and
Di (t) are adjustable control gains. The inferred control law is represented as
l
u(t) = ∑ vi (z(t))[−Ki e(t) + (−Ai + Ci (t))y(t) + (−Bi + Di (t))y(t − τ (t))].
i=1
(8.24)
Substituting (8.24) into (8.23), we obtain the following error system:

ė(t) =ẋ(t) − ẏ(t)

l
= ∑ vi (z(t)) [(Ai − Ki )e(t) + Bi e(t − τ (t))]
i=1
l 7 8
+ ∑ vi (z(t)) (Ci (t) − Âi )y(t) + (Di (t) − B̂i )y(t − τ (t)) + ϖ (t). (8.25)
i=1

Theorem 8.2. For the error system (8.25), if there exist symmetric positive-definite
matrices P and Q, a positive constant σ , feedback gain matrices Ki , and the para-
metric adaptive laws
⎧
⎪
⎪ ċ1 jk (t) = −vk (z(t))α1 m j yk (t),
⎨
.. (8.26)
⎪ .
⎪
⎩
ċl jk (t) = −vk (z(t))αl m j yk (t),
⎧ ˙
⎪ d1 jk (t) = −vk (z(t))β1 m j yk (t − τ (t)),
⎪
⎨
.. (8.27)
⎪ .
⎪
⎩ ˙
dl jk (t) = −vk (z(t))βl m j yk (t − τ (t)),
where αi and βi are some given positive constants, m j is the elements of 2eT P, ci jk
and di jk are elements of Ci and Di , i = 1, 2, . . . , l, j, k = 1, 2, . . . , n, such that
⎛ ⎞
∏ PBi P I
⎜ BT P − (1 − h)Q 0 0 ⎟
⎜ i ⎟ < 0, (8.28)
⎝ P 0 − σ 2I 0 ⎠
I 0 0 −I
where ∏ = (Ai − Ki )T P + P(Ai − Ki ) + Q, then we obtain H∞ performance given by
8.4 Synchronizing Fuzzy Chaotic Systems with Time-Varying Delays 301
 tf  tf
eT (t)e(t)dt ≤ eT (0)e(0) + σ 2 ϖ T (t)ϖ (t)dt.
0 0

Proof. Let us select the Lyapunov-Krasovskii functional as

V (t) = V1 (t) + V2(t),

where
t
V1 (t) = eT (t)Pe(t) + eT (s)Qe(s)ds,
t−τ (t)
 
l n n
1  2 1
V2 (t) = ∑ ∑ ∑ ci jk (t) − âi jk + (di jk (t) − b̂i jk )2 ,
i=1 j=1 k=1 2αi 2βi

where âi jk and b̂i jk are elements of unknown parameters matrices Âi and B̂i , respec-
tively.
The derivative of V (t) along the trajectory of (8.25) is

V̇1 (t) = ėT (t)Pe(t) + eT(t)Pė(t) + eT (t)Qe(t)

− (1 − τ̇ (t)eT (t − τ (t))Qe(t − τ (t)))
l ;
≤ ∑ vi (z(t)) eT (t)[(Ai − Ki )T P + P(Ai − Ki ) + Q]e(t)
i=1
<
+ eT (t − τ (t))BTi Pe(t) + eT(t)PBi e(t − τ (t))
− (1 − h)eT(t − τ (t))Qe(t − τ (t)) + ϖ T(t)Pe(t) + eT(t)Pϖ (t)
l ;
+ ∑ vi (z(t)) 2eT (t)P(Ci (t) − Âi)y(t)
i=1
T
<
+ 2e (t)P(Di (t) − B̂i )y(t − τ (t))

and
l n n  
1 1
V̇2 (t) = ∑ ∑ ∑ (ci jk (t) − âi jk )ċi jk (t) + (di jk (t) − b̂i jk )d˙i jk (t)
i=1 j=1 k=1 αi βi
l n n 9
=− ∑ ∑ ∑ vi (z(t)) (ci jk (t) − âi jk )m j yk (t)
i=1 j=1 k=1
:
+ (di jk (t) − b̂i jk )m j yk (t − τ (t))
r 7 8
= − 2eT(t)P ∑ vi (z(t)) (Ci (t) − Âi )y(t) + (Di (t) − B̂i)y(t − τ (t)) . (8.29)
i=1

Using (8.26) and (8.29), we obtain

302 8 Synchronizing Chaotic Systems Based on Fuzzy Models

V̇ (t) = V̇1 (t) + V̇2(t)

l ;
≤ ∑ vi (z(t)) eT (t)[(Ai − Ki )T P + P(Ai − Ki ) + Q]e(t)
i=1
<
+ eT (t − τ (t))BTi Pe(t) + eT(t)PBi e(t − τ (t))
− (1 − h)eT(t − τ (t))Qe(t − τ (t)) + ϖ T(t)Pe(t) + eT(t)Pϖ (t). (8.30)

The H∞ performance is satisfied as follows:

 tf
J= [eT (t)e(t) − σ 2 ϖ T (t)ϖ (t)]dt, (8.31)
0

where t f denotes the terminal time of the control and σ is a disturbance attenuation
value which denotes the effect of ϖ (t) on e(t). Using (8.30) and (8.31), we obtain
 tf 7 8  tf
T 2 T
J= e (t)e(t) − σ ϖ (t)ϖ (t) + V̇ (t) dt − V̇ (t)dt
0 0
 tf 7 8
= eT (t)e(t) − σ 2ϖ T (t)ϖ (t) + V̇ (t) dt − V (t f ) + V (0)
0
 tf l
≤
0
∑ vi (z(t))η T (t)Φη (t)dt + V (0),
i=1
⎛ ⎞
∏ +I PBi P
where η (t) = (e(t)T , e(t − τ (t))T , ϖ (t)T )T , Φ = ⎝ BTi P −(1 − h)Q 0 ⎠,
P 0 −σ 2 I
T
and ∏ = (Ai − Ki ) P + P(Ai − Ki ) + Q.
If there exist symmetric positive-definite matrices P > 0 and Q > 0, a positive
constant σ , and feedback gain matrices Ki such that
⎛ ⎞
∏ +I PBi P
⎝ BTi P −(1 − h)Q 0 ⎠ < 0, (8.32)
P 0 −σ 2 I

we get  tf
J= [eT (t)e(t) − σ 2 ϖ T (t)ϖ (t)]dt < V (0).
0
The matrix inequality (8.32) can be converted into (8.28) via the Schur complement.
The proof is complete. 

8.4.3 Simulations

A numerical example is given to show the validity of the proposed synchronization

method of T–S fuzzy chaotic systems with time-varying delays.
8.4 Synchronizing Fuzzy Chaotic Systems with Time-Varying Delays 303

The fuzzy drive system [8] can be described as

Ri : IF z1 (t) is Mi1 , . . . , zg (t) is Mig ,

(8.33)
THEN ẏ(t) = Âi y(t) + B̂iy(t − τ (t)) + ε (t),

i = 1, 2.
In this fuzzy model, the state vector y = (y1 , y2 )T are taken as the premise vari-
ables. The membership functions of fuzzy sets are
R1 : IF y1 (t) is M11 , THEN ẏ(t) = Â1 y(t) + B̂1 y(t − τ (t)) + ε (t),
R2 : IF y1 (t) is M21 , THEN ẏ(t) = Â2 y(t) + B̂2 y(t − τ (t)) + ε (t),
where y(t) = (y1 (t), y2 (t))T . Let parameter matrices Â1 , Â2 , B̂1 , B̂2 , and ε (t) be
respectively selected as follows

0 1 0 1 0 0
Â1 = , Â2 = , B̂1 = B̂2 = ,
−47.6 −0.18 1.2 −0.18 0.01 0.01

and
0
ε (t) = .
25 cos(1.29t)
Let τ (t) = 0.125 sin(2t), the membership functions are selected as follows

y21 y2
M11 (y1 (t)) = , M21 (y1 (t)) = 1 − 1 .
100 100
Here, we suppose that the response chaotic system to be synchronized has the
same structure as the drive system. The fuzzy response system can be described by
the following rules:
R1 : IF x1 (t) is F11 , THEN ẋ(t) = A1 x(t) + B1x(t − τ (t)) + ε (t),
R2 : IF x1 (t) is F21 , THEN ẋ(t) = A2 x(t) + B2x(t − τ (t)) + ε (t),
where x(t) = (x1 (t), x2 (t))T ,

0 1 0 1 0 0
A1 = , A2 = , B1 = B2 = ,
−39.6 −0.1 0.4 −0.1 0.01 0.01

ε (t) and τ (t) are the same definition as drive systems, and the membership functions
are selected as follows:
x21 x2
M11 (x1 (t)) = , M21 (x1 (t)) = 1 − 1 .
100 100
We illustrate each of these synthesis procedures, in which the LMI Control Tool-
box is used to compute the solutions of LMIs. From the assumption, we obtain the
symmetric positive-definite matrices P and Q and the feedback gains K1 and K2 as
follows:
0.2299 0.0000 1.5325 0.0000
P= , Q= ,
0.0000 0.2298 0.0000 1.5325
304 8 Synchronizing Chaotic Systems Based on Fuzzy Models

8.0084 − 19.2968 8.0084 0.7000

K1 = , K2 = .
−19.2968 7.9097 0.7001 7.9097
Given the initial state values of the drive system and the response system as
y(0) = (1, − 1)T and x(0) = (2, 3)T and σ = 0.5 as the prescribed value. The
parametric adaptive laws are chosen as α1 = 50, α2 = 10, β1 = 2, and β2 = 1.
Let us assume that
0 0 0 0
C1 = , C2 = ,
c21
1 c22
1 c21
2 c22
2

where c21 22 21 22
1 , c1 , c2 , and c2 are uncertain parameters. The initial value of C1 (t) and
C2 (t) are selected as follows

0 0 0 0
C1 (0) = = ,
c21
1 (0) c 22 (0)
1 −0.4 − 2

0 0 0 0
C2 (0) = =
c21 22
2 (0) c2 (0) −2 − 8
The synchronization errors and the estimated errors are illustrated as follows. The
synchronization errors of each state variable are shown in Figs. 8.12 and 8.13.
Changing parameters can be clearly seen in Fig. 8.14 by using the scheme pro-
posed in Theorem 8.2. The adaptive synchronization is more rapidly achieved for
uncertain drive system with the response system.

Fig. 8.12 The synchronization error e1

8.4 Synchronizing Fuzzy Chaotic Systems with Time-Varying Delays 305

Fig. 8.13 The synchronization error e2

Fig. 8.14 Changing parameter curves of c21 22 21 22

1 , c1 , c2 , and c2
306 8 Synchronizing Chaotic Systems Based on Fuzzy Models

8.5 Summary

In this chapter, first, we represented the typical chaotic and hyperchaotic systems via
the T–S fuzzy models. Based on the fuzzy hyperchaotic models, simple fuzzy con-
trollers were designed for synchronizing hyperchaotic systems. After that, a fuzzy
controller was presented for the synchronization of chaotic systems based on the T–
S fuzzy models with uncertain parameters and time-varying delays, which is based
on the PDC technique, the T–S fuzzy models, and the adaptive control method. The
theoretical analysis and simulation study show the validity of the proposed methods.

References

1. Chen G, Dong X (1998) From Chaos to Order: Methodologies, Perspectives and Applica-
tions. World Scientific, Singapore
2. Chen ZQ, Yang Y, Qi GY, Yuan ZZ (2007) A novel hyperchaos only with one equilibrium.
Phys Lett A 360:696–701
3. EI-Dessoky MM (2007) Synchronization and anti-synchronization of a hyperchaotic Chen
system. Chaos Solitons Fractals 39:1790–1797
4. Elabbasy EM, Agiza HN, EI-Dessoky MM (2006) Adaptive synchronization of a hyper-
chaotic system with uncertain parameter. Chaos Solitons Fractals 30:1133–1142
5. Gao TG, Chen ZQ, Yuan ZZ, Yu DC (2007) Adaptive synchronization of a new hyperchaotic
system with uncertain parameters. Chaos Solitons Fractals 33:922–928
6. Jia Q (2007) Hyperchaos generated from the Lorenz chaotic system and its control. Phys Lett
A 366:217–222
7. Jia Q (2007) Adaptive control and synchronization of a new hyperchaotic system with un-
known parameters. Phys Lett A 362:424–429
8. Kim JH, Shin H, Kim E, Park M (2005) Synchronization of time-delayed TS fuzzy chaotic
systems via scalar output variable. Int J Bifurc Chaos 15:2593–2601
9. Lian KY, Chiang TS, Chiu CS, Liu P (2001) Synthesis of fuzzy model-based designs to syn-
chronization and secure communications for chaotic systems. IEEE Trans Syst Man Cybern
B 31:66–83
10. Nikolov S, Clodong S (2006) Hyperchaos–chaos–hyperchaos transition in modified Rőssler
systems. Chaos Solitons Fractals 28:252–263
11. Pecora LM, Carroll TL (1990) Synchronization in chaotic systems. Phys Rev Lett 64:821–824
12. Takagi T, Sugeno M (1985) Fuzzy identification of systems and its applications to modeling
and control. IEEE Trans Syst Man Cybern 15:116–132
13. Tanaka K, Ikeda T, Wang HO (1998) A unified approach to controlling chaos via an LMI-
based fuzzy control system design. IEEE Trans Circuits Syst I 45:1021–1040
14. Wang FQ, Liu CX (2006) Hyperchaos evolved from the Liu chaotic system. Chinese Phys
15:963–968
15. Wang GY, Zhang X, Zheng Y, Li YX (2006) A new modified hyperchaotic Lü system. Physica
A 371:260–272
16. Wu XY, Guan ZH, Wu ZP (2008) Adaptive synchronization between two hyperchaotic sys-
tems. Nonlinear Anal 68:1346–1351
17. Yang DS, Zhang HG, Li AP, Meng ZY (2007) Generalized synchronization of two non–
identical chaotic systems based on fuzzy model. Acta Phys Sinica 56:3121–3126
18. Yassen MT (2007) On hyperchaos synchronization of a hyperchaotic Lü system. Nonlinear
Anal 68:3592–3600
19. Zhang HG, Cai LL, Bien Z (2000) A fuzzy basis function vector-based multivariable adaptive
fuzzy controller for nonlinear systems. IEEE Trans Syst Man Cybern B 30:210–217
References 307

20. Zhang HB, Liao XF, Yu JB (2005) Fuzzy modeling and synchronization of hyperchaotic
systems. Chaos Solitons Fractals 26:835–843
21. Zhang HG, Guan HX, Wang ZS (2007) Adaptive synchronization of neural networks with
different attractors. Prog Nat Sci 17:687–695
22. Zhang HG, Huang W, Wang ZL, Chai TY (2006) Adaptive synchronization between different
chaotic systems with unknown parameters. Phys Lett A 350:363–366
23. Zhang HG, Xie YH, Liu D (2006) Synchronization of a class of delayed chaotic neural net-
works with fully unknown parameters. Dyn Contin Discrete Impuls Syst B 13:297–308
24. Zhang HG, Yang DD, Chai TY (2007) Guaranteed cost networked control for T–S fuzzy
systems with time delay. IEEE Trans Syst Man Cybern C 37:160–172
Chapter 9
Chaotification of Nonchaotic Systems

Abstract As a reverse process of suppressing or eliminating chaotic behaviors in

order to reduce the complexity of an individual system or a coupled system, chao-
tification aims at creating or enhancing the system’s complexity for some special
applications. More precisely, chaotification is to generate some chaotic behaviors
from a given system, which is nonchaotic or even is stable originally. In this chap-
ter, we will investigate how to chaotify dynamical systems which are nonchaotic
originally. In this chapter, first, we develop a simple nonlinear state feedback con-
troller to chaotify the discrete-time fuzzy hyperbolic model (FHM) with uncertain
parameters. Second, we use aan impulsive and nonlinear feedback control method to
chaotify the continuous-time FHM. Finally, we chaotify two classes of continuous-
time linear systems via a sampled data control approach.

9.1 Introduction

In this chapter, we will investigate how to chaotify dynamical systems which are
nonchaotic originally. As a reverse process of suppressing or eliminating chaotic
behaviors in order to reduce the complexity of an individual system or a coupled
system, chaotification aims at creating or enhancing the system’s complexity for
some special applications. More precisely, chaotification is to generate some chaotic
behaviors from a given system, which is nonchaotic or even is stable originally [22].
It is well known that chaotification of dynamical systems has many potential appli-
cations in electronic, mechanical, biological, and medical systems, etc. [9, 16, 20].
In recent years, many conventional control methods and special techniques were
applied to the chaotification of discrete-time dynamical systems or continuous-time
dynamical systems, such as linear feedback control [9, 12], time-delay feedback
control [17, 19], nonlinear feedback control [11], adaptive inverse optimal control
[25], impulsive control [21, 27], saturating control [13], switching control [1], and
other control methods [22], etc. However, most of them are based on the assumption
that the analytical representations of nonlinear dynamical systems to be chaotified

309
310 9 Chaotification of Nonchaotic Systems

are known exactly. For an unknown or uncertain nonlinear dynamical system, the
above methods are ineffective. Recently, Zhang and Quan [24] and Margaliot et
al. [14] successively proposed a novel fuzzy model named the hyperbolic tangent
model (FHM). The FHM is a nonlinear model in nature which is very suitable for
expressing nonlinear dynamical properties, and it is easy to design a stable and op-
timal controller based on the FHM than other models such as the fuzzy dynamic
model, T–S model, etc. Moreover, the FHM has been shown to be a universal ap-
proximator, i.e., it can approximate an arbitrary nonlinear continuous function on a
compact set with arbitrary accuracy [26]. This provides a theoretical foundation for
using the FHM to represent complex nonlinear systems approximately. Therefore,
the FHM can be regarded as an intermediate step for chaotifying an unknown non-
linear dynamical system. In this chapter, first, we develop a simple nonlinear state
feedback controller to chaotify the discrete-time FHM with uncertain parameters.
Second, we use an impulsive and nonlinear feedback control method to chaotify the
continuous-time FHM. Finally, we chaotify two classes of continuous-time linear
systems via a sampled data control approach.

9.2 Chaotification of Discrete-Time Fuzzy Hyperbolic Model

with Uncertain Parameters

9.2.1 Basic Formulation

First, we define the fuzzy rule base for inferring the discrete-time FHM.
Definition 9.1 ([14, 23]). Given a plant with n state variables x(k) = (x1 (k), . . . ,
xn (k))T , we call the following fuzzy rule base a discrete-time hyperbolic-type fuzzy
rule base if it satisfies the following conditions:
(i) Every fuzzy rule adopts the following form:

Rl : IF x1 (k) is Fx1 and x2 (k) is Fx2 . . . and xn (k) is Fxn ,

THEN xi (k + 1) = ±cx1 ± cx2 ± · · · ± cxn ,

l = 1, 2, . . . , 2n , where Fxi (i = 1, 2, . . . , n) are fuzzy sets of xi (k) which include

two linguistic values P (positive) and N (negative), cxi (i = 1, 2, . . . , n) is a pos-
itive constant corresponding to Fxi , and ‘±’ stands for either the plus or the
minus sign. The actual signs in the THEN part are determined in the following
manner: if the term characterizing Fxi in the IF part is P, then cxi in the THEN
part appears with a plus sign; otherwise cxi appears with a minus sign.
(ii) The constant term cxi in the THEN part corresponds to Fxi in the IF part; that is,
if there is an Fxi term in the IF part, cxi must appear in the THEN part; otherwise
cxi does not appear in the THEN part.
9.2 Chaotification of Discrete-Time Fuzzy Hyperbolic Model with Uncertain Parameters 311

(iii) There are 2n fuzzy rules in the rule base; that is, it includes all possible P and
N combinations of state variables in the IF part and all sign combinations of
constants in the THEN part. 

Theorem 9.1 ([14, 23]). Given a hyperbolic-type rule base, if we define the mem-
bership functions of Pxi and Nxi as
9  2 :
μPx (xi (k)) = exp − xi (k) − sxi /2 ,
i
9  2 :
μNx (xi (k)) = exp − xi (k) + sxi /2 ,
i

where sxi > 0, then we can derive the following model:

x(k + 1) = A tanh(Sx(k)),

where A is a constant matrix, S = diag {sx1 , . . . , sxn }, and

tanh(Sx(k)) := (tanh(sx1 x1 (k)), . . . , tanh(sxn xn (k)))T . 

This model is named the discrete-time FHM. In the sequel, sxi is denoted by si
for simplicity.
Consider a general n-dimensional discrete-time autonomous system with the
form
x(k + 1) = g(x(k)), (9.1)
where g is a C1 nonlinear map. Denote Z̄ = N ∪ {0}. Suppose that t ∈ Z̄ and denote
gt the t times compositions of g with itself. x∗ is a p-periodic point of g (p ∈ N) if
g p (x∗ ) = x∗ and gt (x∗ ) = x∗ for t < p. When p = 1, x∗ is a fixed point. Let g (x) and
det(g (x)) be the Jacobian of g at the point x and its determinant, respectively. With
this notation, we can give the definition of a snap-back repeller.

Definition 9.2 ([15]). Suppose that x∗ is a fixed point of g with all eigenvalues of
g (x∗ ) exceeding 1 in magnitude, and suppose that there exists a point x0 = x∗ in
a neighborhood1 of x∗ such that, for some positive integer m, gm (x0 ) = x∗ and
det((gm ) (x0 )) = 0. Then, x∗ is said to be a snap-back repeller of g. 

Lemma 9.1 ([15]). If system (9.1) has a snap-back repeller then the system is
chaotic in the sense of Li–Yorke, namely,
(i) there exists a positive integer n such that, for every integer p ≥ n, system (9.1)
has p periodic points;
(ii) there exists a scrambled set (an uncountable invariant set S containing no peri-
odic points) such that
a. g(S) ⊂ S,
1 In [15], Marotto pointed out that identifying a repelling neighborhood of x∗ is not in general a
difficult task. In fact, since the local unstable manifold of x∗ includes all points of Rn close to x∗ ,
every closed ball B(x∗ , r) defined by the Euclidean norm must be so for sufficiently small r > 0.
312 9 Chaotification of Nonchaotic Systems

b. for every y ∈ S and any periodic point x of (9.1),

4 4
4 4
lim sup 4gk (x) − gk (y)4 > 0,
k→∞

c. for every x, y ∈ S with x = y,

4 4
4 4
lim sup 4gk (x) − gk (y)4 > 0;
k→∞

(iii) there exists an uncountable subset S0 ⊂ S such that, for any x, y ∈ S0 ,

4 4
4 4
lim inf 4gk (x) − gk (y)4 = 0. 
k→∞

Lemma 9.2 (Gershgorin Disc Theorem [3]). Let A be a complex matrix whose
n
i jth element is ai j . For every i = 1, 2, . . . , n, let ri := ∑ ai j . Let Di be a disc
j=1, j=i
of radius ri centered at aii . Then, all eigenvalues of A lie in the union of Di for
n

i = 1, 2, . . . , n, i.e., λi ∈ D j for i = 1, 2, . . . , n. 
j=1

9.2.2 Main Results

Now, consider the following uncertain discrete-time FHM:

x(k + 1) = (A + Δ A) tanh(Sx(k)), (9.2)

where Δ A represents the uncertain parameters. Assume that Δ A and S are bounded
in the sense of infinite norm, i.e., there exist two positive constants γ and η such that
Δ A ∞ ≤ γ and S ∞ ≤ η . Consequently, it can be inferred that A + Δ A ∞ ≤ ρ ,
where ρ is another positive constant.
Next, it will be proven that the controlled system is chaotic in the sense of Li–
Yorke if we design the parameters of the following controller (9.3) appropriately.
The relevant result is given as follows.
Theorem 9.2 ([28]). For the uncertain discrete-time FHM (9.2), given an arbitrarily
small δ , there exists a sufficiently large constant β such that under the controller

βπ
u(k) = δ sin x(k)
δ
 T
βπ βπ βπ
= δ sin x1 (k) , δ sin x2 (k) , . . . , δ sin xn (k) , (9.3)
δ δ δ

the controlled uncertain discrete-time FHM

x(k + 1) = (A + Δ A) tanh(Sx(k)) + u(k)

9.2 Chaotification of Discrete-Time Fuzzy Hyperbolic Model with Uncertain Parameters 313

is chaotic in the sense of Li–Yorke.

Proof. Define

βπ
x(k + 1) = (A + Δ A) tanh(Sx(k)) + δ sin x(k) := g(x(k)). (9.4)
δ

Obviously, x∗ = 0 is a fixed point of g, which will be proven to be a snap-back

repeller later. Define
 /
1 1
Φ (Sx) := diag ,..., ,
cosh2 (s1 x1 ) cosh2 (sn xn )
 /
βπ βπ βπ
cos x := diag cos x1 , . . . , cos xn .
δ δ δ

βπ ∗
Note that Φ (Sx∗ ) = Φ (0) = I and cos x = cos(0) = I, where I denotes the
δ
identity matrix of appropriate dimensions. Differentiating (9.4) with respect to x at
the fixed point x∗ = 0 yields

g (x∗ ) = (A + Δ A)S + πβ I.

According to Lemma 9.2, one can assert that all eigenvalues λi (g (x∗ )), i = 1, 2, . . ., n,
n
lie in the union of all Di of radius ∑ (ai j + Δ ai j )s j centered at (aii + Δ aii )si +
j=1, j=i
πβ . Therefore, as long as one chooses an appropriate β , such that the following
inequalities are met:
n
(aii + Δ aii )si + πβ − ∑ (ai j + Δ ai j )s j > 1, for i = 1, 2, . . . , n, (9.5)
j=1, j=i

it is guaranteed that all λi (g (x∗ )) exceed 1 in magnitude for i = 1, 2, . . . , n. Indeed,

according to the definition of infinite norm, and noting that S is a diagonal matrix,
one has -
n
(A + Δ A)S ∞ = max
i
∑ (ai j + Δ ai j )s j .
j=1

Therefore, the following inequalities are satisfied:

n
∑ (ai j + Δ ai j )s j ≤ (A + Δ A)S ∞ − |(aii + Δ aii )si | , for i = 1, 2, . . . , n. (9.6)
j=1
j=i

If (aii + Δ aii )si ≥ 0, based on the inequalities (9.6), when

1
β> ( 1 + (A + Δ A)S ∞) ,
π
314 9 Chaotification of Nonchaotic Systems

one gets
n
(aii + Δ aii )si + πβ − ∑ (ai j + Δ ai j )s j > 1 + 2(aii + Δ aii )si , for i = 1, 2, . . . , n.
j=1, j=i

Consequently, the inequalities (9.5) are satisfied.

1
Otherwise, for (aii + Δ aii )si < 0, when β > ( 1 + (A + Δ A)S ∞ ), the inequal-
π
ities (9.5) are satisfied directly. Then, according to the property of infinity norm, one
has
(A + Δ A)S ∞ ≤ A + Δ A ∞ S ∞ = ρη .
Therefore, if β > (1 + ρη )/π , one can guarantee that all eigenvalues of g (x∗ )
exceed 1 in magnitude.
Next, it is shown that there exists a point x0 ∈ B(x∗ , r), with x0 = x∗ , where
B(x∗ , r) := {x : x − x∗ ≤ r} and r is a sufficiently small positive constant (that
is, this r is small enough to guarantee that the closed ball B(x∗ , r) defined by Eu-
clidean norm is a repelling neighborhood of x∗ ), such that g(x0 ) = 0 = x∗ and
det(g (x0 )) = 0. In the sequel, for simplicity, x < y denotes that every component of
x − y is negative and x > y denotes that every component of x − y is positive. First,
δ δ T 3δ 3δ T
we choose two points x1 = ,..., and x2 = ,..., . Obviously,
√ 2β 2β 2β 2β
if β > (3 nδ )/(2r), then it can be guaranteed that x1 ∈ B(x , r) and x2 ∈ B(x∗ , r)
∗

simultaneously.
Noting that s j > 0 (for j = 1, . . . , n), x1 > 0, x2 > 0, and tanh(x) < x for x > 0,
we can derive the following two inequalities:
n    βπ
gi (x1 ) ≥ − ∑ ai j + Δ ai j tanh s j x1 j + δ sin x1i
j=1 δ
δ
≥ − A + ΔA ∞ S ∞ +δ
2β
ρηδ
=− +δ (9.7)
2β
and
n   βπ
gi (x2 ) ≤ ∑ ai j + Δ ai j tanh(s j x2 j ) + δ sin
δ
x2i
j=1
3δ
≤ A + ΔA ∞ S ∞ −δ
2β
3ρηδ
= −δ, (9.8)
2β

where i = 1, 2, . . . , n. From (9.7) we can infer that if β > (ρη )/2 then g(x1 ) > 0
can be guaranteed. From (9.8) we can infer that if β > (3ρη )/2 then g(x2 ) < 0
9.2 Chaotification of Discrete-Time Fuzzy Hyperbolic Model with Uncertain Parameters 315

can be guaranteed. Combining the above two results, it can be concluded that if β >
δ δ T 3δ 3δ T
(3ρη )/2, then there exist two points x1 = ,..., and x2 = ,..., ,
2β 2β 2β 2β
such that g(x1 ) > 0 and g(x2 ) < 0. Therefore, by the mean value theorem of calculus
[10], there exists a point x0 ∈ B(x∗ , r), x1 < x0 < x2 , such that g(x0 ) = 0 = x∗ . Then,
as long as an appropriate β satisfying det(g (x0 )) = 0 is determined, we can con-
clude that x∗ is a snap-back repeller of the map g defined in (9.4). In fact, considering
that β can be chosen sufficiently large, we can always determine an appropriate con-
βπ 0
stant β , such that det(g (x0 )) = det (A + Δ A)SΦ (Sx0) + πβ cos x = 0.
δ
∗
Then, it can be inferred that x is a snap-back repeller of the map g defined in (9.4).
 β is selected, √
To conclude, if an appropriate such / that β > β¯ and det(g (x0 )) = 0
1 + ρη 3ρη 3 nδ
is satisfied, where β¯ = max , , , then x∗ = 0 is a snap-back
π 2 2r
repeller of the map g defined in (9.4). Then, according to Lemma 9.1, it can be con-
cluded that the controlled system is chaotic in the sense of Li–Yorke. This completes
the proof. 
Remark 9.1. On one hand, there exists the possibility that the selected β is not very
appropriate, which makes the equality det(g (x0 )) = 0 met, though the probability of
the event det(g (x0 )) = 0 is very small for a certain β . Then, the controlled system
may be unable to generate chaos in this case. On the other hand, the exact value
of r in B(x∗ , r) is not easy to determine. Based on the two facts above, a feasible
method is presented here. Actually, we can regard β as an adjustable parameter, i.e.,
the value of β is increased gradually from the initial value 0. In this process, we can
see that a series of period-doubling bifurcation behaviors occur continually. In the
end, chaos will be generated in the controlled system. 
Remark 9.2. Assume that the original discrete-time system to be chaotified is fully
unknown. However, if it can be approximated by the uncertain discrete-time FHM
(9.2) with adequate accuracy, then we can believe that the controller for chaotifica-
tion of the uncertain discrete-time FHM (9.2) can chaotify the original system at the
same time. 

9.2.3 Simulations

To visualize the effectiveness of the theoretical analysis and controller design, an

example is included for illustration. Consider an uncertain discrete-time FHM as
follows:
x(k + 1) = (A + Δ A) tanh(Sx(k)) + u(k),
⎛ ⎞
1 2 1.5
where x(k) = (x1 , x2 , x3 )T , A = ⎝ 0 2 1 ⎠,
1.5 1 3
316 9 Chaotification of Nonchaotic Systems

Fig. 9.1 The periodic orbits of x1 with β = 0

Fig. 9.2 The period-doubling bifurcation diagram of x1

9.2 Chaotification of Discrete-Time Fuzzy Hyperbolic Model with Uncertain Parameters 317

Fig. 9.3 The first period-doubling bifurcation diagram of x1

Fig. 9.4 The chaotic attractor of the controlled system with β = 200
318 9 Chaotification of Nonchaotic Systems
⎛ ⎞
kπ kπ
⎜ 0.07 sin 2 0.1 cos
2
0 ⎟
⎜ ⎟
⎜ kπ kπ ⎟
⎜ ⎟
Δ A = ⎜ 0.01 cos 0 0.05 cos ⎟,
⎜ 2 2 ⎟
⎜ ⎟
⎝ kπ kπ ⎠
0.09 cos 0.08 sin 0
2 2
S = diag {s1 , s2 , s3 } = diag {1, 1, 1}, and

βπ
u(k) = δ sin x(k) .
δ

In the simulation, the magnitude of the control input is arbitrarily chosen as

δ = 0.1. The initial value of the controlled system is chosen as x(0) = (0, 1, 2)T .
According to Theorem 9.2, β can be regarded as an adjustable control parameter.
Fig. 9.1 presents the periodic orbits of an uncertain discrete-time FHM without con-
trol. When β takes values from 0 to 1 gradually, the period-doubling bifurcation2
diagram of x1 is shown in Fig. 9.2. For clarity, Fig. 9.3 describes the first period-
doubling bifurcation of x1 . The bifurcation processes of x2 and x3 are similar to that
of x1 , and are therefore omitted. The chaotic attractor of the controlled system is pre-
sented in Fig. 9.4. From these numerical results, it can be verified that the designed
controller realizes the chaotification of the above uncertain discrete-time FHM.

9.3 Chaotification of Continuous-Time Fuzzy Hyperbolic Model

Using Impulsive Control

9.3.1 Basic Formulation

First, we review some necessary preliminaries for Devaney’s definition of chaos and
the continuous-time FHM.
Definition 9.3 ([2]). A map φ : S → S, where S is a set, is chaotic if
(i) φ has sensitive dependence on initial conditions, i.e., for any x ∈ S and any
neighborhood Ñ of x in S, there exists a δ > 0 such that φ m (x) − φ m (y) > δ
for some y ∈ Ñ and m > 0, where φ m is the mth-order iteration of φ , i.e., φ m :=
φ ◦ φ ◦ · · · ◦ φ (m times).
(ii) φ is topologically transitive, i.e., for any pair of subsets U,V ⊂ S, there exists
an integer m > 0 such that φ m (U) ∩V = 0. /
(iii) The periodic points of φ are dense in S. 

2 In mathematics, a period-doubling bifurcation in a dynamical system is a bifurcation in which

the system switches to a new behavior with twice the period of the original system.
9.3 Chaotification of Continuous-Time Fuzzy Hyperbolic Model Using Impulsive Control 319

Remark 9.3. Definition 9.3 is for discrete-time systems. For continuous-time sys-
tems, we need to construct a Poincaré section so as to get a Poincaré map. Details
will be provided later. 

Definition 9.4 ([24]). Given a plant with n input variables x = (x1 (t), . . . , xn (t))T and
an output variable ẋ, we call the fuzzy rule base a continuous-time hyperbolic-type
fuzzy rule base if it satisfies the following conditions:
(i) Every fuzzy rule has the following form:

Rl : IF x1 is Fx1 and x2 is Fx2 . . . and xn is Fxn ,

THEN x˙i = ±cx1 ± cx2 ± · · · ± cxn ,

l = 1, 2, . . . , 2n , where Fxi (i = 1, . . . , n) is a fuzzy set of xi , which includes two

linguistic values P (positive) and N (negative), cxi (i = 1, 2, . . . , n) is a positive
constant corresponding to Fxi , and ‘±’ stands for either the plus or the minus
sign. The actual signs in the THEN part are determined in the following manner:
if in the IF part the term characterizing Fxi is P, then in the THEN part cxi
appears with a plus sign; otherwise, cxi appears with a minus sign.
(ii) The constant term cxi in the THEN part corresponds to Fxi in the IF part; that
is, if there is an Fxi term in the IF part, cxi must appear in the THEN part.
Otherwise, cxi does not appear in the THEN part.
(iii) There are 2n fuzzy rules in the rule base and these are all the possible P and N
combinations of input variables in the IF part, and all the sign combinations of
constants in the THEN part. 

The following theorem explains how a continuous-time FHM is constructed.

Theorem 9.3 ([24]). Given a hyperbolic-type fuzzy rule base, if we define the mem-
bership functions of Pxi and Nxi as
1 2 1 2
μPxi (xi ) = e− 2 (xi − kxi ) , μNxi (xi ) = e− 2 (xi + kxi ) ,

where kxi > 0 (i = 1, . . . , n), then we can derive the following model:

ẋ = A tanh(Kx), (9.9)

where K = diag {kx1 , . . . , kxn }, A is a constant matrix, and

tanh(Kx) := (tanh(k1 x1 ), . . . , tanh(kn xn ))T .

This model is named continuous-time FHM. 

Definition 9.5 ([7]). A matrix A is said to be a Hurwitz diagonally stable matrix if

there exists a diagonal matrix Q = diag{qi } > 0 (i = 1, . . . , n) such that

AT Q + QA < 0. 
320 9 Chaotification of Nonchaotic Systems

9.3.2 Main Results

Now, we consider the following control system:

ẋ = A f (x) + Bu, (9.10)

where x ∈ Rn , f (x) = tanh(Kx) is an n-dimensional vector function, u = u(x(t)) ∈

Rm×1 is a vector function of state x, and A ∈ Rn×n and B ∈ Rn×m are both constant
matrices.
Assumption 9.1. (A, B) is completely controllable. 
Under Assumption 9.1, we have the following lemma.
Lemma 9.3. For system (9.10), if the controller u is chosen as follows:

u(x(t)) = L tanh(Kx(t)), (9.11)

where L ∈ Rm×n is a constant matrix such that (A + BL) is a Hurwitz diagonally

stable matrix, then the origin of the closed-loop system (9.10) is globally asymptot-
ically stable.
Proof. Since (A + BL) is a Hurwitz diagonally stable matrix, there exists a diago-
nal matrix Q = diag{qi } > 0 such that (A + BL)TQ + Q(A + BL) < 0. Choose the
following Lyapunov function:
n  xi
V (x) = 2 ∑ qi fi (ξi )dξi ,
i=1 0

where fi (·) is the ith element of f (x); then V (x) is positive definite and radially
unbounded, namely, V (x) > 0 for all x = 0 and V (x) → +∞ as x → +∞. The time
derivative of V (x) along system (9.10) can be computed as follows:

V̇ (x) = f T (x)Qẋ(t) + ẋT Q f (x)

= f T (x)Q[A f (x) + Bu] + [A f (x) + Bu]TQ f (x). (9.12)

Substituting (9.11) into (9.12) and noting that (A + BL) is a Hurwitz diagonally
stable matrix, we get
7 8
V̇ (x) = f T (x) (A + BL)T Q + Q(A + BL) f (x) < 0, if x = 0.
Moreover, the proof is independent of initial values, so the closed-loop system (9.10)
is globally and asymptotically stable. 
Remark 9.4. For the purpose of deriving our main results, we only require the sta-
bility of closed-loop systems (9.10) and (9.11). It should be noted that there are
other results for the stability of closed-loop systems (9.10) and (9.11). System (9.9)
and the closed-loop systems (9.10) and (9.11) can be regarded as recurrent neural
networks. They are also special cases of Lur’e systems [8]. 
9.3 Chaotification of Continuous-Time Fuzzy Hyperbolic Model Using Impulsive Control 321

Remark 9.5. Under the conditions of Lemma 9.3, solutions of the controlled system
(9.10) are defined in a bounded region D ⊂ Rn . 

Remark 9.6. If we let à = A + BL, system (9.10) with u given in (9.11) becomes

ẋ = Ã f (x). (9.13)

Because of the form of f (x), there exists a constant γ > 0 such that, for all differ-
ent x, y ∈ D, f (y) − f (x) ≤ γ y − x . That is to say, f (x) is Lipschitzian on D.
From the theory of ordinary differential equations, we know that system (9.10) has
a unique continuous solution φ (t,t0 , x0 ) through a given initial point, (t0 , x0 ), which
is also continuously dependent on x0 [5]. 

Now, let us consider the following impulsive control system:

⎧
⎨ ẋ = Ã f (x), t = τk ,
Δ x = I (x(t)), t = τk , k = 1, 2, 3, . . . , (9.14)
⎩ + k
x(t0 ) = x0 , t0 ≥ 0,

where à has been defined in (9.13), Ik (x(t)), k = 1, 2, 3, . . . , are impulsive control

signals defined in D ⊂ Rn , t0 = τ0 < τ1 < · · · < τk , τk − τk−1 = T , lim τk = +∞,
k→+∞
Δ x = x(τk+ ) − x(τk− ), and x(τk+ ) and x(τk− ) are the right and left limits of x(t) at
t = τk .
For system (9.14), we have the following lemma.

Lemma 9.4. If Ik (x(t)) is chosen as follows:

Ik (x(t)) = yk − x(t − ), t = τk , k = 1, 2, 3, . . . , (9.15)

where yk = g(yk−1 ) and g : Y → Y , Y ⊂ D, then, when g(·) is chaotic and satisfies

Devaney’s three criteria, the controlled system (9.14) is also chaotic and satisfies
Devaney’s criteria.

Proof. First, it should be emphasized that for system (9.14) to display chaotic dy-
namics, its phase space must be a finite region; that is to say, there exists an M > 0
such that the trajectory of system (9.14), φ (t,t0 , x0 ), satisfies φ (t,t0 , x0 ) ≤ M for
all t in the domain. In fact, for t ∈ [τ0 , τ1 ), system (9.14) is under the action of
à f (x). By Lemma 9.3, its trajectory, φ (t,t0+ , x0 ), is asymptotically stable. When
t = τ1 , Ã f (x) turns off and I1 (x(τ1 )) acts on system (9.14). The trajectory of sys-
tem (9.14) jumps from φ (τ1− ,t0+ , x0 ) to x10 . We denote the trajectory at each instant
t = τk as xk0 , k = 1, 2, 3, . . . , where x00 = x0 . Since I1 (x(τ1 )) is a finite impulse sig-
nal, x10 is also finite. For t ∈ (τ1 , τ2 ), I1 (x(τ1 )) turns off and à f (x) turns on with
the initial point (τ1+ , x10 ). The trajectory of system (9.14) in this time interval, de-
noted φ (t, τ1+ , x10 ), is also asymptotically stable. The analysis in other time intervals,
(τk , τk+1 ), k = 2, 3, 4, . . . , is the same as that in (τ1 , τ2 ), and situations at other time
instants, τk , k = 2, 3, 4, . . . , are similar to that at t = τ1 . It is easy to see that although
322 9 Chaotification of Nonchaotic Systems

the trajectory may be discontinuous at t = τk , k = 1, 2, 3, . . . , the whole trajectory is

indeed in a bounded region.
In the following we will prove that system (9.14) is chaotic and satisfies De-
vaney’s three criteria.
To prove that a continuous-time system is chaotic, one method is to show that its
Poincaré map is chaotic [4]. Suppose that φ (t,t0+ , x0 ) is a solution of system (9.13)
with an initial value x(t0+ ) = x0 . According to Remark 9.6, the solution φ (t,t0+ , x0 )
is unique and, thus, φ −1 (t,t0+ , x0 ) exists. Define a set of Poincaré sections Sk as

Sk = {(t, x) : x ∈ D,t = τk } , k = 1, 2, 3, . . . .

The Poincaré map is defined as

P : V → V, P = ψ ◦ g ◦ ψ −1,

where V = ψ (Y ) and ψ (x) = φ (τk + T, τk , x) = φ (τ0 + T, τ0 , x). We will first prove

that P is extremely sensitive to initial values, i.e., there exists an ε > 0 and, for any
two points x0 , x¯0 ∈ V , there exists an N > 0, such that
4 N 4
4P (x0 ) − PN (x¯0 )4 > ε . (9.16)

Since
PN = (ψ ◦ g ◦ ψ −1) ◦ · · · ◦ (ψ ◦ g ◦ ψ −1) = ψ ◦ gN ◦ ψ −1, (9.17)
0 12 3
Nth-order iteration of P
to prove that (9.16) holds is equivalent to prove that
4  N   4
4ψ g (y0 ) − ψ gN (y¯0 ) 4 > ε (9.18)

holds, where y0 = ψ −1 (x0 ) and y¯0 = ψ −1 (x¯0 ).

If (9.18) does not hold, i.e., for any ε0 satisfying ε ≥ ε0 > 0 there exists an
n(ε0 ) > 0 such that
ψ (gn (y0 )) − ψ (gn(y¯0 )) ≤ ε0 ,
from the fact that ψ (x) is continuous we know that, for the above ε0 , there exists a
δ0 such that
gn (y0 ) − gn (y¯0 ) < δ0 for all n ≥ n(ε0 ).
But, this contradicts the fact that g is a chaotic map satisfying Devaney’s definition.
Therefore, P satisfies the first criterion of Devaney.
Next, we show that P is topologically transitive, that is, for any pair of subsets
E, F ⊂ V , there exists an integer N > 0 such that

PN (E) ∩ F = 0.
/ (9.19)

It is known that for any two subsets ψ −1 (E) and ψ −1 (F) ⊂ Y , there exists an integer
N > 0 such that  
gN ψ −1 (E) ∩ ψ −1 (F) = 0. / (9.20)
9.3 Chaotification of Continuous-Time Fuzzy Hyperbolic Model Using Impulsive Control 323

Acting on both sides of (9.20) by ψ , we get

  
ψ gN ψ −1 (E) ∩ F = 0.
/

Noticing (9.17), we therefore proved (9.19).

Finally, we will prove that the periodic points of P are dense in V . Denote the set
of periodic points of a map f as Per( f ). Since Per(g) = Y and ψ is one to one, we
have Per(P) = V . This completes the proof. 

Summarizing the results above, we have the following theorem:

Theorem 9.4 ([27]). For the following system:
⎧
⎨ ẋ = A f (x) + Bu, t = τk ,
Δ x = I (x(t)), t = τk , (9.21)
⎩ + k
x(t0 ) = x0 , t0 ≥ 0, k = 1, 2, 3, . . . ,

where A and B satisfy Assumption 9.1, if u is chosen as in (9.11) and Ik (x(t)) is

chosen as in (9.15), the system (9.21) will present chaotic dynamics and satisfy the
three criteria of Devaney. 
Remark 9.7. The conclusion of Theorem 9.4 can also be kept if the impulsive control
is chosen as Ik (x(t)) = h(yk ) − x(t − ), where h : Y → Y is a homeomorphism. A
special case is to choose h(yk ) = Λ yk , where Λ = diag {λ1 , . . . , λn } with λi > 0,
i = 1, . . . , n, is a constant diagonal matrix. This means that we can chaotify the
original system with arbitrarily small impulsive energy. 

9.3.3 Simulations

Suppose that we have the following fuzzy rule base:

R1 : IF x1 is Px1 and x2 is Px2 , THEN ẋ3 = cx1 + cx2 ;
R2 : IF x1 is Nx1 and x2 is Px2 , THEN ẋ3 = −cx1 + cx2 ;
R3 : IF x1 is Px1 and x2 is Nx2 , THEN ẋ3 = cx1 − cx2 ;
R4 : IF x1 is Nx1 and x2 is Nx2 , THEN ẋ3 = −cx1 − cx2 ;
R5 : IF x1 is Px1 and x3 is Px3 , THEN ẋ2 = cx1 + cx3 ;
R6 : IF x1 is Nx1 and x3 is Px3 , THEN ẋ2 = −cx1 + cx3 ;
R7 : IF x1 is Px1 and x3 is Nx3 , THEN ẋ2 = cx1 − cx3 ;
R8 : IF x1 is Nx1 and x3 is Nx3 , THEN ẋ2 = −cx1 − cx3 ;
R9 : IF x2 is Px2 and x3 is Px3 , THEN ẋ1 = cx2 + cx3 ;
R10 : IF x2 is Nx2 and x3 is Px3 , THEN ẋ1 = −cx2 + cx3 ;
R11 : IF x2 is Px2 and x3 is Nx3 , THEN ẋ1 = cx2 − cx3 ;
R12 : IF x2 is Nx2 and x3 is Nx3 , THEN ẋ1 = −cx2 − cx3 .
324 9 Chaotification of Nonchaotic Systems

Here, we choose membership functions Pxi and Nxi as follows:

1 2 1 2
μPxi (x) = e− 2 (xi − ki ) , μNxi (x) = e− 2 (xi + ki ) .

Then, we have the following third-order model:

ẋ = A f (x) = A tanh(Kx),
⎛ ⎞ ⎛ ⎞
0 cx2 cx3 0 1 3
where x = (x1 , x2 , x3 )T , A = ⎝ cx1 0 cx3 ⎠ = ⎝ 2 0 3 ⎠, and K = diag{k1 ,
cx1 cx2 0 2 1 0
k2 , k3 } = diag {2, 3, 1}. According to Theorem 9.4, the controlled system is
⎧
⎨ ẋ = (A + BL) f (x), t = τk ,
Δ x = I (x(t)), t = τk , k = 1, 2, 3, . . . , (9.22)
⎩ + k
x(t0 ) = x0 .

We select B = (1, 1, 1)T and L = (−2, −1, −2). We calculate Q = diag{q1, q2 , q3 } =

diag{17.1186, 25.6779, 19.2584}. It is easy to verify that the matrix A + BL is Hur-
witz diagonally stable. The initial value is x0 = (5, 2, −1)T and the impulsive control
is
Ik (x) = Λ g(yk−1 ) − x(t − ),

Fig. 9.5 State curves of system (9.23)

9.3 Chaotification of Continuous-Time Fuzzy Hyperbolic Model Using Impulsive Control 325

Fig. 9.6 State curves of system (9.22) when Λ = diag {0.8, 0, 4}

3.5

2.5

2
x3

1.5

0.5

0
2
1
0 1.5
1
0.5
−1 0
−0.5
x2 x1

Fig. 9.7 Phase diagram of system (9.22) when Λ = diag {0.8, 0, 4}

326 9 Chaotification of Nonchaotic Systems

Fig. 9.8 State curves of system (9.22) when Λ = diag {0.8, 0.2, 1}

1.2

0.8

0.6
x3

0.4

0.2

−0.2
2

1
1.5 2
0 0.5 1
−0.5 0
x2
x1

Fig. 9.9 Phase diagram of system (9.22) when Λ = diag {0.8, 0.2, 1}
9.4 Chaotification of Linear Systems Using Sampled Data Control 327
      T
where Λ = diag {λ1 , λ2 , λ3 } and g(yk−1 ) = g1 y1k−1 , g2 y2k−1 , g3 y3k−1 with
gi (·) a logistic map:
   
yik = gi yik−1 = ayik−1 1 − yik−1 , i = 1, 2, 3.

When a = 4.0, the logistic map is chaotic.

Without impulsive control, the following system

ẋ = (A + BL) f (x) (9.23)

is globally asymptotically stable, whose trajectories are shown in Fig. 9.5. With
the impulsive control and T = 0.01, from Figs. 9.6–9.9, we see that the controlled
system is in chaos for different Λ . These results verify the claims of Theorem 9.4
and Remark 9.7.

9.4 Chaotification of Linear Systems Using Sampled Data

Control

9.4.1 Basic Formulation

Consider the following two classes of continuous-time linear systems. One class of
systems is given as follows:
ẋ = Ax + Bu, (9.24)
where x ∈ Rn is the state vector, A ∈ Rn×n is a Hurwitz stable matrix whose
eigenvalues are real and nonidentical to one another, i.e., λi < 0, λi = λ j (i = j,
i, j = 1, 2, . . . , n), B ∈ Rn×m , m ≤ n, rank(B) = r, u ∈ Rm satisfies

u ∞ ≤ ε0 , (9.25)

and ε0 is an arbitrary positive constant. The aim is to design a controller satisfying

(9.25) to chaotify system (9.24).
The other class of systems is described as follows:

ẋ = A x + bu, (9.26)

where x ∈ Rn is the state vector, and A ∈ Rn×n is a Hurwitz stable matrix whose
eigenvalues are nonidentical√to one another. Moreover, there is a pair of complex
eigenvalues. i.e., λ1,2  = α ± −1β , λ  ∈ R (i = 3, 4, . . . , n), α < 0, λ  < 0, λ  = λ 
i i i j
(i = j, i, j = 3, 4, . . . , n), b ∈ Rn×1 , and u is an input variable satisfying

|u| ≤ ε0 . (9.27)

The aim is to design a controller satisfying (9.27) to chaotify the system (9.26).
328 9 Chaotification of Nonchaotic Systems

ϕε (x)

4ε
−4ε −2ε 2ε x

Fig. 9.10 The sawtooth function

To chaotify the above two classes of stable linear systems, a lemma is presented
below.
Lemma 9.5 ([18]). Consider the following discrete-time linear systems:

x(k + 1) = Λ x(k) + u(k), (9.28)

where x(k) = (x1 (k), x2 (k), . . . , xn (k))T , and Λ is a Schur stable matrix, i.e., ρ (Λ ) <
1, where ρ (·) denotes the spectral radius. If the controller u(k) is chosen as

u(k) = Φε (σ x(k)) := (ϕε (σ x1 (k)), ϕε (σ x2 (k)), . . . , ϕε (σ xn (k)))T ,

;  <
where σ > max 3 Λ ∞ , Λ ∞ + 1 + 1 , then the discrete-time linear system
(9.28) is chaotic in the sense of Li–Yorke. Moreover, u(k) ∞ ≤ ε (k = 1, 2, . . . ),
where ε is an arbitrarily small positive constant, and ϕε (x) is a sawtooth function
shown in Fig. 9.10, which can be described as follows:

ϕε (x) = (−1)l (x − 2l ε ), (2l − 1)ε ≤ x ≤ (2l + 1)ε , l ∈ Z. (9.29)

9.4.2 Main Results

First, we will develop a sampled data controller based on Lemma 9.5 to make the
linear systems (9.24) chaotic.
Since the eigenvalues of A are nonidentical to one another, there exists a nonsin-
gular matrix Q ∈ Rn×n such that system (9.24) can be transformed into

x˙¯ = Āx¯+ B̄u, (9.30)

where x¯ = Q−1 x, Ā = Q−1 AQ is a diagonal matrix whose diagonal elements are

 T
λi (i = 1, 2, . . . , n), B¯ = Q−1 B := b̄T1 , b̄T2 , . . . , b̄Tn , and b̄1 , b̄2 , . . . , and b̄n are row
¯
vectors of B.
9.4 Chaotification of Linear Systems Using Sampled Data Control 329

Assuming that b̄i1 , b̄i2 , . . . , b̄ir are linearly independent columns of B, (9.30) can
be rewritten as ⎧
˙
⎪ x¯i1 = λi1 x¯i1 + b̄i1 u,
⎪
⎪
⎪ ˙
⎪ i2 = λi2 x¯i2 + b̄i2 u,
⎪
⎨
x
¯
..
. (9.31)
⎪
⎪ ˙¯ir = λir x¯ir + b̄ir u,
⎪
⎪ x
⎪
⎪
⎩ ..
.
According to (9.31), introducing a new vector x̃, we can denote
⎧
⎪ x̃1 = x¯i1 ,
⎪
⎪
⎪
⎪ x̃2 = x¯i2 ,
⎪
⎨ ..
. (9.32)
⎪
⎪
⎪
⎪ x̃ r = x¯ir ,
⎪
⎪
⎩ ..
.

i.e., x̃ = R−1 x,
¯ where R−1 is a nonsingular transformation matrix corresponding to
(9.32). Denoting P = QR, system (9.24) can be rewritten as

x̃˙ = Ãx̃ + B̃u, (9.33)

where x̃ = P−1 x, Ã = P−1 AP = diag{λ̃1 , λ̃2 , . . . , λ̃n }, and B̃ = P−1 B. Moreover, the
first r row vectors of B̃ are linearly independent.
Since rank(B) = rank(B̃) = r, there exists a nonsingular matrix C ∈ Rm×m satis-
fying
Ir×r 0r×(m−r)
B̃C = , (9.34)
D(n−r)×r 0(n−r)×(m−r)
where Ir×r is the identity matrix, 0r×(m−r) and 0(n−r)×(m−r) are zero matrices, and
the matrix D(n−r)×r can be represented as
⎛ ⎞
d11 . . . d1r
⎜ . .. .. ⎟
D(n−r)×r = ⎝ .. . . ⎠. (9.35)
d(n−r)1 . . . d(n−r)r

Choose
u = Cv,
where v ∈ Rm×1 is a redefined control input. Therefore, the controlled system (9.33)
is transformed into
x̃˙ = Ãx̃ + B̃Cv. (9.36)
Next, we will discretize system (9.36). Assume that the sampling interval T > 0 is
sufficiently small. Cascading the sampler with a zeroth-order hold, system (9.36) is
discretized to
330 9 Chaotification of Nonchaotic Systems

x̃(k + 1) = Ã(T )x̃(k) + B̃(T )v(k),

% &
where Ã(T ) = eÃT = diag eλ̃1 T , eλ̃2 T , . . . , eλ̃n T ,
 T
B̃(T ) = eÃ(t) dt B̃C
0
⎛ ⎛ ⎞ ⎞
eλ̃1t··· 0
⎜ T⎜ ⎟ ⎟
=⎜ ⎜ .. ..
. 0 ⎟ ⎟
⎝ 0
⎝ . ⎠dt ⎠ B̃C
0 · · · eλ̃nt
⎛ ⎞
1
(eλ̃1 T − 1) · · · 0
λ̃1
⎜ .. .. ⎟ Ir×r 0r×(m−r)
=⎜
⎝ . . 0
⎟
⎠ D(n−r)×r 0(n−r)×(m−r)
0 ··· 1
λ̃n
(eλ̃n T − 1)
⎛ ⎞
1
λ̃1
(eλ̃1 T − 1) ··· 0 0 ··· 0
⎜ .. .. . . .. ⎟
⎜ .. ⎟
⎜ . . 0 . . .⎟
⎜ ⎟
⎜ 0 ··· 1
(eλ̃r T − 1) 0 ··· 0⎟
⎜ λ̃r ⎟
= ⎜ d11 λ̃r+1 T d1r ⎟,
⎜ (e − 1) · · · (eλ̃r+1 T − 1) 0 ··· 0⎟
⎜ λ̃r+1 λ̃r+1 ⎟
⎜ . .. .. .. . . .. ⎟
⎜ .. . . . . .⎟
⎝ ⎠
d(n−r)1 λ̃ T d(n−r)r
λ̃
(e − 1) · · ·
n
λ̃n
(eλ̃n T − 1) 0 · · · 0
n

x̃(k) = x̃(kT ), and v(k) = v(kT ) (k = 0, 1, 2, . . . ), i.e., the discretized system is

⎧
⎪
⎪ x̃ (k + 1) = eλ̃1 T x̃1 (k) + λ̃1 (eλ̃1 T − 1)v1 (k),
⎪ 1
⎪ 1
⎪
⎪ ..
⎪
⎪ .
⎪
⎪
⎪
⎪
⎪
⎪ x̃r (k + 1) = eλ̃r T x̃r (k) + λ̃1 (eλ̃r T − 1)vr (k),
⎪
⎪ r
⎪
⎨ x̃r+1 (k + 1) = eλ̃r+1 T x̃r+1 (k) + d11 (eλ̃r+1 T − 1)v1 (k) + · · ·
λ̃r+1
d1r λ̃r+1 T − 1)v (k), (9.37)
⎪
⎪ + (e r
⎪
⎪ λ̃ r+1
⎪
⎪ ..
⎪
⎪ .
⎪
⎪
⎪
⎪ x̃ (k + 1) = eλ̃n T x̃ (k) + d(n−r)1 (eλ̃n T − 1)v (k) + · · ·
⎪
⎪ n n 1
⎪
⎪ λ̃n
⎪
⎩ d(n−r)r λ̃ T
+ λ̃ (e − 1)vr (k).
n
n

Choose the control input as

9.4 Chaotification of Linear Systems Using Sampled Data Control 331
⎛ ⎞
λ̃1
ϕε (σ x̃1 (k))
⎜ eλ̃1 T −1 ⎟
⎛ ⎞ ⎜
⎜
.. ⎟
⎟
v1 (k) ⎜ . ⎟
⎜ ⎟ ⎜ λ̃r ⎟
v(k) = ⎝ ... ⎠ = ⎜⎜ eλ̃r T −1
ϕε ( σ x̃ r (k)) ⎟,
⎟ (9.38)
⎜ 0 ⎟
vm (k) ⎜ ⎟
⎜ .. ⎟
⎝ . ⎠
0

where ϕε (x) is a sawtooth function. Then, the subsystem composed of the first r
variables of (9.37) is given by
⎧
⎪
⎪
⎪ x̃1 (k + 1) = eλ̃1 T x̃1 (k) + ϕε (σ x̃1 (k)),
⎪
⎨ x̃2 (k + 1) = eλ̃2 T x̃2 (k) + ϕε (σ x̃2 (k)),
⎪ ..
⎪
⎪ .
⎪
⎩
x̃r (k + 1) = eλ̃r T x̃r (k) + ϕε (σ x̃r (k)),

which is equivalent to the following vector form:

x̃R (k + 1) = ÃR (T )x̃R (k) + Φε (σ x̃R (k)), (9.39)

⎛ ⎞
x̃1 % &
⎜ .. ⎟
where x̃R = ⎝ . ⎠, ÃR (T ) = diag eλ̃1 T , eλ̃2 T , . . . , eλ̃r T , and
x̃r
⎛ ⎞
ϕε (σ x̃1 )
⎜ .. ⎟
Φε (σ x̃R ) = ⎝ . ⎠.
ϕε (σ x̃r )

It is obvious that Φε (σ x̃R ) ≤ ε . To make sure that u

∞ ∞ ≤ ε0 , denoting
-
λ̃1 λ̃r
δ = max ,..., , (9.40)
eλ̃1 T − 1 eλ̃r T − 1

we can then choose

ε0
ε= . (9.41)
δ C ∞
It can be seen that the last n − r state variables x̃r+1 , x̃r+2 , . . . , x̃n have no effect on
the first r state variables. At the same time, it can be proven that x̃r+1 , x̃r+2 , . . . , x̃n
are bounded if the control input variables are bounded. In fact, the solution of the
discrete-time system (9.37) can be given as

 k k−1  k−1−i
x̃(k) = Ã(T ) x̃(0) + ∑ Ã(T ) B̃(T )v(i).
i=1
332 9 Chaotification of Nonchaotic Systems
4 4
Denote 4Ã(T )4∞ = α . Since λ̃i < 0 (i = 1, 2, . . . , n), it can be inferred that α < 1.
Furthermore, it can be seen that v ∞ ≤ εδ = ε0 / C ∞ from (9.38), (9.40), and
(9.41). Therefore,
k−1 4 4
lim x̃(k)
k→∞
∞ ≤ lim α k x̃(0)
k→∞
∞+ lim
k→∞
∑ α k−1−i 4B̃(T )4∞ v(i) ∞
i=1
4 4
ε0 4B̃(T )4∞ k−1
≤ lim ∑ α k−1−i
C ∞ k→∞ i=1
4 4
ε0 4B̃(T )4∞
= . (9.42)
C ∞ (1 − α )

It is obvious that x̃ is chaotic if x̃R is chaotic. So, the next step is to chaotify system
(9.39). Since λ̃i < 0 (i = 1, 2, . . . n), it is clear that eλ̃i T < 1. Therefore, system (9.39)
is stable without control (i.e., v(k) ≡ 0). According to Lemma 9.5, choosing
 /
4 4 4 4
σ > max 3 4ÃR (T )4∞ , 4ÃR (T )4∞ + 1 + 1 , (9.43)

system (9.39) is chaotic in the sense of Li–Yorke, that is, system (9.37) is chaotic in
the sense of Li–Yorke.
The fact that x̃(k) in (9.37) is chaotic means that x̃(t) in (9.33) and x(t) in (9.24)
are chaotic. Therefore, we can derive the following theorem.
Theorem 9.5 ([6]). For the n-dimensional continuous-time linear controlled system
(9.24), if the nonlinear feedback controller is chosen as
⎛  ⎞
λ̃1
ϕε σ (P−1 x(kT ))1
⎜ e 1 −1
λ̃ T
⎟
⎛ ⎞ ⎜ .. ⎟
⎜ . ⎟
v1 (t) ⎜   ⎟
⎜ ⎟ ⎜ λ̃r −1 x(kT )) ⎟
u(t) = Cv(t) = C ⎝ ... ⎠ = C ⎜ ⎜ eλ̃r T −1
ϕε σ (P r ⎟,
⎟ (9.44)
⎜ 0 ⎟
vm (t) ⎜ ⎟
⎜ .. ⎟
⎝ . ⎠
0

where kT ≤ t < (k + 1)T , k = 0, 1, 2, . . . , then the controlled system (9.24) is

chaotic in the sense of Li–Yorke, where P is the transformation matrix in (9.33),
(P−1 x(kT )) j ( j = 1, 2, . . . , r) is the jth component of the vector P−1 x(kT ), the ma-
trix C is decided by (9.34), and ε and σ satisfy (9.41) and (9.43), respectively. 
Remark 9.8. In (9.35), denoting
r
Di = ∑ di j ,
j=1

where i = 1, 2, . . . , n − r, we have
9.4 Chaotification of Linear Systems Using Sampled Data Control 333

4 4 1 λ̃1 T 1 D1 λ̃r+1 T
4B̃(T )4 = max (e − 1), . . . , (eλ̃r T − 1), (e − 1), . . .,
∞
λ̃1 λ̃r λ̃r+1
/
D(n−r) λ̃n T
(e − 1) .
λ̃n
4 4
From the above analysis, it can be seen that 4B̃(T )4∞ will decrease with the
decreasing of the sampling interval T . Therefore, it can be inferred from (9.42) that
the chaotic region will shrink with the decreasing of the sampling interval T and the
control upper bound ε0 . 
Remark 9.9. The control input in Theorem 9.5 satisfies u ∞ ≤ ε0 , where ε0 is an
arbitrarily small positive constant. In fact, if the magnitude of the control input can
be chosen arbitrarily large, then we can chaotify any linear system ẋ = Ax + Bu
by assigning appropriate poles and adopting the control input in the form of (9.44)
as long as (A, B) is controllable completely. In this case, the controller consists of
two parts: one is a linear state feedback controller and the other is a nonlinear state
feedback controller. 
Next, we will investigate the problem of chaotification of the linear systems
(9.26).
Since the eigenvalues of (9.26) are λ1,2 = α ± iβ , λ  ∈ R (i = 3, 4, . . . , n), there
i
exists a nonsingular matrix P such that (9.26) can be transformed into the following
Jordan standard form:
x̃˙ = Ã x̃ + b̃u, (9.45)
Ã1 0 α β
where x̃ = P−1 x, Ã = P−1 A P = is a Jordan matrix, Ã1 = ,
0 Ã2 −β α
; <  T
Ã2 = diag λ3 , . . . , λn , and b̃ = P−1 b := b̃1 , b̃2 , . . . , b̃n .
Discretizing (9.45) by a zeroth-order hold, we can derive the discretized linear
systems as follows:
x̃(k + 1) = Ã (T )x̃(k) + b̃(T )u(k), (9.46)
where T is the sampling interval,

 Ã1 (T ) 0
à (T ) = eà T = ,
0 Ã2 (T )
 T
 eα T Λ1 (T ) 0
b̃(T ) = eà (t) dt b̃ = b̃,
0 α2 + β 2 0 Λ2 (T )
 eα T cos β T −eα T sin β T
Ã1 (T ) = eÃ1 T = ,
eα T sin β T eα T cos β T
%  
&
Ã2 (T ) = diag eλ3 T , . . . eλn T ,
⎛ ⎞
α cos β T + β sin β T − α e−α T −(α sin β T − β cos β T + β e−α T )
Λ1 (T ) = ⎝ ⎠,
α sin β T − β cos β T + β e−α T α cos β T + β sin β T − α e−α T
334 9 Chaotification of Nonchaotic Systems

Λ2 (T ) = diag {σ3 (T ), . . . , σn (T )} ,
and 
(α 2 + β 2)(eλi T − 1)
σi (T ) = (i = 3, . . . , n).
eα T λi
Choosing the sampling interval T as
π
T = T0 = , (9.47)
β
; <
we get Ã1 (T0 ) = diag −eα T0 , −eα T0 . Consequently,
%  
&
à (T0 ) = diag −eα T0 , −eα T0 , eλ3 T0 , . . . , eλn T0 .

Here,
⎛ ⎞ ⎛ ⎞
−α (1 + e−α T0 ) −β (1 + e−α T0 ) 0 b̃1 (T0 )
eα T0
⎝ β (1 + e−α T0 ) −α (1 + e−α T0 ) 0 ⎠ b̃ := ⎜ . ⎟
b̃(T0 ) = 2 ⎝ .. ⎠ ,
α +β2 0 0 Λ2 (T0 ) b̃n (T ) 0

where
−eα T0  
b̃1 (T0 ) = α (1 + e−α T0 )b̃1 + β (1 + e−α T0 )b̃2 ,
α +β
2 2

−eα T0  
b̃2 (T0 ) = α (1 + e−α T0 )b̃2 − β (1 + e−α T0 )b̃1 ,
α2 + β 2
b̃i   
b̃i (T0 ) =  eλi T0 − 1 (i = 3, . . . , n).
λi
Therefore, (9.46) can be rewritten as follows:
⎧
⎪ x̃1 (k + 1) = −eα T0 x̃1 (k) + b̃1(T0 )u(k),
⎪
⎪ α T0 x̃ (k) + b̃ (T )u(k),
⎪
⎨ x̃2 (k + 1) = −e
⎪
λ T
2 2 0
x̃3 (k + 1) = e 3 0 x̃3 (k) + b̃3 (T0 )u(k), (9.48)
⎪
⎪ ..
⎪
⎪ .
⎪
⎩ 
x̃n (k + 1) = eλn T0 x̃n (k) + b̃n (T0 )u(k).

It is obvious that the state variables of (9.48), i.e., x̃1 (k), x̃2 (k), . . . , x̃n (k), are inde-
pendent. Since b is nonsingular, b̃ is nonsingular after a nonsingular transformation.
So, all b̃i (T0 ) are not zero. We consider the following one-dimensional systems:

x̃1 (k + 1) = −eα T0 x̃1 (k) + b̃1(T0 )u(k). (9.49)

According to Lemma 9.5, choosing the input u(k) as

9.4 Chaotification of Linear Systems Using Sampled Data Control 335

ϕε (σ x̃1 (k))
u(k) = , (9.50)
b̃1 (T0 )

where
2ε0 eα T0
ε = ε0 b̃1 (T0 ) = 2 α (1 + e−α T0 )b̃1 + β (1 + e−α T0 )b̃2 ,
α +β2
%  &
σ > max 3eα T0 , eα T0 + 1 + 1 ,

the controlled system (9.49) is chaotic in the sense of Li–Yorke. Moreover, |u(k)| ≤
ε0 .
With the control input u(k) of (9.50), the solution of the closed-loop system
(9.48) is given by

 k k−1  k−1−i
x̃(k) = Ã (T0 ) x̃(0) + ∑ Ã (T0 ) b̃(T0 )u(k).
i=1
4 4
Denote 4Ã (T0 )4∞ = ρ . Since α < 0 and λi < 0 (i = 3, 4, . . . , n), it is obvious that
ρ < 1. Therefore,
k−1 4 4
lim x̃(k)
k→∞
∞ ≤ lim ρ k x̃(0)
k→∞
∞+ lim
k→∞ i=1
∑ ρ (k−1−i) 4b̃(T0 )4∞ |u(k)|
4 4
ε0 4b̃(T0 )4∞
≤ ,
1−ρ

i.e., the state variables of system (9.48) are bounded. For the n-dimensional system
(9.48), the fact that x̃1 (k) is chaotic means that x̃2 (k), . . . , x̃n (k) are chaotic.
The fact that the discretized system (9.48) is chaotic means that the continuous-
time systems (9.45) and (9.26) are chaotic. Therefore, when b̃1 (T0 ) = 0, we can
adopt the following control input:

ϕε (σ x̃1 (k)) ϕε (σ (P−1 x(kT0 ))1 )

u(t) = = ,
b̃1 (T0 ) b̃1 (T0 )

where kT0 ≤ t < (k + 1)T0 , k = 0, 1, 2, . . . , to chaotify the system (9.26).

For x̃2 , . . . , x̃n , we can derive similar conclusions. Therefore, we have the follow-
ing theorem.

Theorem 9.6 ([6]). For the n-dimensional continuous-time linear systems (9.26),
when b̃i (T0 ) = 0 (i = 1, 2, . . . , n), we can adopt the following control input to chao-
tify the continuous-time linear systems (9.26)

ϕε (σ (P−1 x(kT0 ))i )

u(t) = , (9.51)
b̃i (T0 )
336 9 Chaotification of Nonchaotic Systems

where kT0 ≤ t < (k + 1)T0 , k = 0, 1, 2, . . . , |u(t)| ≤ ε0 , P is the nonsingular matrix

which transforms (9.26) into (9.45), (P−1 x(kT ))i (i = 1, 2, . . . , n) denotes the ith
component of the vector P−1 x(kT ), T0 is determined by (9.47), ε = ε0 b̃i (T0 ) ,
   /
 
σ > max 3eα T0 , eα T0 + 1 + 1, 3eλi T0 , eλi T0 + 1 + 1 (i = 3, 4, . . . , n),

b̃i (T0 ) is given by

⎧
⎪
⎪ −eα T0  
⎪
⎪ b̃ 1 (T0 ) = α (1 + e−α T0 )b̃1 + β (1 + e−α T0 )b̃2 ,
⎪
⎪ α 2 +β2
⎨ −eα T0  
b̃2 (T0 ) = 2 α (1 + e−α T0 )b̃2 − β (1 + e−α T0 )b̃1 , (9.52)
⎪
⎪ α + β 2
⎪
⎪
⎪
⎪ b̃i   
⎩ b̃i (T0 ) =  eλi T0 − 1 (i = 3, . . . , n),
λi
 T
and b̃1 , b̃2 , . . . , b̃n = P−1 b. 

Remark 9.10. The control input u(k) in Theorem 9.6 satisfies |u| ≤ ε0 , where ε0 is
an arbitrarily small positive constant. If, without this restraint, i.e., the magnitude of
the control input u(k) can be given arbitrarily large, then we can chaotify any linear
systems ẋ = Ax+ bu as long as (A, b) is completely controllable. Here, we can assign
the poles of the closed-loop system appropriately by state feedback control first, and
then add a nonlinear feedback control (9.51) to the system; then, we can chaotify
any linear systems ẋ = Ax + bu. In this case, the controller also consists of two parts:
one is a linear state feedback controller and the other is a nonlinear state feedback
controller. 

9.4.3 Simulations

In this subsection, two illustrative examples are presented to demonstrate the effec-
tiveness of the proposed chaotification approach. First, let us consider the following
three-dimensional continuous-time linear system:
⎛ ⎞ ⎛ ⎞⎛ ⎞ ⎛ ⎞
ẋ1 −29/18 17/18 1/6 x1 10
⎝ ẋ2 ⎠ = ⎝ 7/18 −19/18 1/6 ⎠ ⎝ x2 ⎠ + ⎝ 1 1 ⎠ u1 . (9.53)
u2
ẋ3 −5/18 11/18 −5/6 x3 01

By calculation, the eigenvalues of the system matrix A is given by⎛λ1 = −1/2, ⎞λ2 =
1 1 2
−1, and λ3 = −2. Choosing the nonsingular matrix P as P = ⎝ 1 1 −1 ⎠, the
1 −2 1
transformation is given by x̃ = P−1 x. System (9.53) is transformed into the following
diagonal form:
9.4 Chaotification of Linear Systems Using Sampled Data Control 337
⎛ ⎞ ⎛ ⎞⎛ ⎞ ⎛ ⎞
x̃˙1 −1/2 0 0 x̃1 2/3 8/9
⎝ x̃˙2 ⎠ = ⎝ 0 −1 0 ⎠ ⎝ x̃2 ⎠ + ⎝ 1/3 −2/9 ⎠ u1 .
u2
x̃˙3 0 0 −2 x̃3 0 −1/3

It can be verified that the first two row vectors of b̃ are linearly independent. So, we
2/3 8/9 1/2 2
get C−1 = , i.e., C = . It can be calculated easily that
1/3 −2/9 3/4 −3/2
C ∞ = 2.5. According to (9.43), we can choose σ = 3. Assume that ε0 = 0.2 and
the initial values are chosen as (0.01, 0.01, 0)T and the sampling intervals are cho-
sen as T = 5, T = 1, and T = 0.2, respectively. According to (9.40), δ is presented
as δ = 1.0068, δ = 1.5820, and δ = 5.5167, respectively. According to (9.41), ε is
presented as ε = 0.0795, ε = 0.0506, and ε = 0.0145. The chaotic response curves
are shown in Figs. 9.11–9.13 corresponding to various sampling intervals, respec-
tively. From the simulation results we can see that the scales of the chaotic region
become smaller and smaller with the decrease of sampling intervals.
Second, let us consider the following three-dimensional continuous-time linear
system:
⎛ ⎞ ⎛ ⎞⎛ ⎞ ⎛ ⎞
ẋ1 −10/9 −1/9 −25/18 x1 1
⎝ ẋ2 ⎠ = ⎝ 5/18 −13/18 −5/18 ⎠ ⎝ x2 ⎠ + ⎝ 1 ⎠ u. (9.54)
ẋ3 2/3 2/3 −1/6 x3 1

0.2

0.2 0.1

0
x1

0
x3

0.2 −0.1
−0.2
0.2 0
0 x1 −0.2
−0.2 −0.2 0 100 200 300 400 500
x2
t

0.2 0.4

0.1 0.2

0 0
x2

x3

−0.1 −0.2

−0.2 −0.4
0 100 200 300 400 500 0 100 200 300 400 500
t t

Fig. 9.11 Chaotic orbit in phase space and state variables of system (9.53) with T = 5
338 9 Chaotification of Nonchaotic Systems

0.1

0.1 0.05

0
x3

x1
−0.1
−0.05
0.2
0.1
0 0
x2 −0.1
−0.2 −0.1 x1 0 100 200 300 400 500
t

0.2 0.2

0.1 0.1

0 0
x2

−0.1 x3 −0.1

−0.2 −0.2
0 100 200 300 400 500 0 100 200 300 400 500
t t

Fig. 9.12 Chaotic orbit in phase space and state variables of system (9.53) with T = 1

0.04

0.05
0.02

0
x1
x3

−0.02
−0.05 0.05
0.05 0
0 −0.04
−0.05 −0.05 x1 0 100 200 300 400 500
x2
t

0.04 0.1

0.02 0.05

0 0
x3
x2

−0.02 −0.05

−0.04 −0.1
0 100 200 300 400 500 0 100 200 300 400 500
t t

Fig. 9.13 Chaotic orbit in phase space and state variables of system (9.53) with T = 0.2
9.5 Summary 339

0.05

0.05

0
x3

x1
−0.05
0.05
0.05
0 0
x2 −0.05
−0.05 −0.05 x1 0 100 200 300 400 500
t

0.04 0.04

0.02
0.02
0

x3
0
x2

−0.02
−0.02
−0.04

−0.04 −0.06
0 100 200 300 400 500 0 100 200 300 400 500
t t

Fig. 9.14 Chaotic attractor and state variables of the controlled system (9.54)

The eigenvalues of the state matrix A are given by λ1,2

 = −1/2 ± i, λ  = −1. Choos-
3
⎛ ⎞
1/9 5/9 1/3
ing a nonsingular matrix P = ⎝ 2/9 1/9 −1/3 ⎠, the linear system (9.54) is trans-
1/3 −1/3 0
formed into
⎛ ⎞ ⎛ ⎞⎛ ⎞ ⎛ ⎞
x̃˙1 −1/2 1 0 x̃1 4
⎝ x̃˙2 ⎠ = ⎝ −1 −1/2 0 ⎠ ⎝ x̃2 ⎠ + ⎝ 1 ⎠ u.
x̃˙3 0 0 −1 x̃3 0

It is easy to see that b̃1 = 4, b̃2 = 1, and b̃3 = 0. According to (9.47), we get T0 = π .
2
According to (9.52), we can derive b̃1 (T0 ) = (e−π /2 + 1) ≈ −0.4832. Assuming
5
that ε0 = 0.2, according to Theorem 9.6, we get ε ≈ 0.0996. Choosing σ = 3 and
T
the initial states x0 = (0.01, 0.01, 0.02) , the linear system (9.54) is chaotified and
is shown by Fig. 9.14.

9.5 Summary

In this chapter, we have investigated the problem of chaotification of both linear

systems and nonchaotic nonlinear systems. In Sect. 9.2, we developed a simple
sinusoidal nonlinear feedback controller to chaotify the discrete-time FHM with
340 9 Chaotification of Nonchaotic Systems

uncertain parameters. In Sect. 9.3, a composite controller based on nonlinear state

feedback control and the impulsive control method was designed to chaotify the
continuous-time FHM. In Sect. 9.4, two classes of continuous-time linear systems
were chaotified based on sampled data control. It is verified that all the chaos gen-
erated by the above methods satisfies either Li–Yorke’s definition or Devaney’s def-
inition through mathematical analysis. We believe that these chaotification methods
can be used not only for FHM and the above two classes of linear systems, but also
for a wide class of nonchaotic systems.

References

1. Chen QF, Zhong QH, Hong YG, Chen GR (2007) Generation and control of spherical circular
attractors using switching schemes. Int J Bifurc Chaos 17:243–253
2. Devaney RL (1989) An Introduction to Chaotic Dynamical Systems, 2nd edn. Addison-
Wesley, New York
3. Golub GH, van Loan CF (1983) Matrix Computations. Johns Hopkins University Press, Bal-
timore
4. Guckenheimer J, Holmes P (1983) Nonlinear Oscillations, Dynamical Systems, and Bifurca-
tions of Vector Fields. Springer, New York
5. Hirsch MW, Smale S (1974) Differential Equations, Dynamical Systems, and Linear Algebra.
Academic Press, New York
6. Huang W (2005) A study of synchronization of chaotic systems and chaotification of non-
chaotic systems. Ph.D. dissertation, Northeastern University, Shenyang
7. Kaszkurewicz E, Bhaya A (2000) Matrix Diagonal Stability in Systems and Computation.
Birkhäuser, Boston
8. Khalil HK (2002) Nonlinear Systems, 3rd edn. Prentice Hall, New Jersey
9. Lai DJ, Chen GR (2005) Chaotification of discrete-time dynamical systems: an extension of
the Chen–Lai algorithm. Int J Bifurc Chaos 15:109–117
10. Lang S (1997) Undergraduate Analysis, 2nd edn. Springer, New York
11. Li Z, Park JB, Chen GR, Joo YH, Choi YH (2002) Generating chaos via feedback control
from a stable TS fuzzy system through a sinusoidal nonlinearity. Int J Bifurc Chaos 12:2283–
2291
12. Lu JG (2005) Generating chaos via decentralized linear state feedback and a class of nonlinear
functions. Chaos Solitons Fractals 25:403–413
13. Lu JG (2006) Chaotic behavior in sampled-data control systems with saturating control.
Chaos Solitons Fractals 30:147–155
14. Margaliot M, Langholz G (2003) A new approach to fuzzy modeling and control of discrete-
time systems. IEEE Trans Fuzzy Syst 11:486–494
15. Marotto FR (2005) On redefining a snap-back repeller. Chaos Solitons Fractals 25:25–28
16. Schiff SJ, Jerger KD, Duong H, Chang T, Spano ML, Ditto WL (1994) Controlling chaos in
the brain. Nature 370:615–620
17. Wang XF, Chen GR, Yu XH (2000) Anticontrol of chaos in continuous-time systems via
time-delay feedback. Chaos 10:1–9
18. Wang XF, Chen GR (2000) Chaotifying a stable LTI system by tiny feedback control. IEEE
Trans Circuits Syst I 47:410–415
19. Wang XF, Chen GR, Man KF (2001) Making a continuous-time minimum-phase system
chaotic by using time-delay feedback. IEEE Trans Circuits Syst I 48:641–645
20. Yang W, Ding M, Mandell AJ, Ott E (1995) Preserving chaos: control strategies to preserve
complex dynamics with potential relevance to biological disorders. Phys Rev E 51:102–110
References 341

21. Yang L, Liu Z, Chen GR (2002) Chaotifying a continuous-time system via impulsive input.
Int J Bifurc Chaos 12:1121–1128
22. Yang RT, Hong YG, Qin HS, Chen GR (2005) Anticontrol of chaos for dynamic systems in
p-normal form: a homogeneity-based approach. Chaos Solitons Fractals 25:687–697
23. Zhang HG, Quan YB (2000) Modeling and control based on fuzzy hyperbolic model. Acta
Autom Sinica 26:729–735
24. Zhang HG, Quan YB (2001) Modeling, identification and control of a class of non-linear
system. IEEE Trans Fuzzy Syst 9:349–354
25. Zhang HG, Wang ZL, Liu D (2004) Chaotifying fuzzy hyperbolic model using adaptive in-
verse optimal control approach. Int J Bifurc Chaos 14:3505–3517
26. Zhang HG, Wang ZL, Li M, Quan YB, Zhang MJ (2004) Generalized fuzzy hyperbolic
model: a universal approximator. Acta Autom Sinica 30:416–422
27. Zhang HG, Wang ZL, Liu D (2005) Chaotifying fuzzy hyperbolic model using impulsive and
nonlinear feedback control approaches. Int J Bifurc Chaos 15:2603–2610
28. Zhao Y, Zhang HG, Zheng CD (2008) Anticontrol of chaos for discrete-time fuzzy hyperbolic
model with uncertain parameters. Chinese Phys B 17:529–535
Index

α -limit point, 30 Duffing equation, 116

α -limit set, 1, 30 dynamical system, 17, 23
n-periodic point, 25
ω -limit point, 30 equilibrium, 2
ω -limit set, 1, 30 equivalence relation, 56
ergodicity, 4
anticontrol of chaos, 10 expansive, 53
attracting set, 30
attractor, 31, 138 FHM, 310, 311
first variation equation, 57
Barbalat lemma, 151 fixed point, 58, 311
basin of attraction, 30 Floquet characteristic multipliers, 50, 79
basin of entrainment, 79 flow, 22
bicritical node, 34 focus, 34
bifurcation, 97 fractal geometry, 3
bi-infinity-tuple, 55 functional differential equations, 71
Brouwer’s fixed point theorem, 199
globally asymptotically stable, 52
Cantor set, 64 Hénon map, 108
cellular neural networks, 199 Hénon system, 164
center, 25, 34 homeomorphism, 5, 23, 56, 65
channel time delay, 180 homoclinic orbit, 68
chaotic, 53 Hopfield neural networks, 199
chaotic delayed neural networks, 198 horseshoe map, 62
chaotification, 10 Hurwitz polynomial, 137, 145
characteristic exponents, 50 hyperbolic fixed point, 46
characteristic multipliers, 58 hyperbolic invariant set, 5
Chen system, 194 hyperchaotic, 146
comparison system, 181
conditional period, 33 improper oscillation, 40
conditional Lyapunov exponents, 7 impulsive functional differential equations,
cycle, 25 198
invariance set, 162
discriminant, 31 inverse optimal control, 104
dissipative, 68 involutive, 142
distance, 23
distribution, 142 Jordan node, 34

343
344 Index

LaSalle invariance principle, 161 saddle focus, 35, 42

Lie bracket, 142 saddle node, 35
limit cycle, 25 saddle point, 33
limit set, 25 sensitive dependence on initial conditions, 53
Lorenz system, 137, 176 shift map, 54, 55, 65
Lü system, 151 SISO, 134
Lyapunov exponent, 11, 53, 57 snap-back repeller, 311
Lyapunov function, 52 stable node, 33
Lyapunov-like stability theorem, 200 structurally stable, 25
superposition principle, 38
MIMO, 142 symbol space, 54
symbolic dynamics, 54
nonwandering set, 30

odd number limitation, 6 topological equivalence, 56

OGY, 4 topological mixing, 56
one-sided shift map, 55 topological transitivity, 56
OPCL, 77 topologically conjugate, 56
OPNCL, 77, 78 topologically equivalent, 56
orbit, 22 topologically semiconjugate, 56
topologically transitive, 30
parametric uncertainty, 183 trajectory, 22
PDC, 270, 289, 300, 306 T–S fuzzy model, 269
periodic, 2
periodic orbit, 25 unified system, 176
Poincaré map, 28, 319 unstable node, 33
Poincaré section, 319
practical stability, 180
vector fields, 142
quasiperiodic, 2, 40 vector relative degree, 143
vortex, 34
relative degree, 134, 158
Rössler system, 137 zero dynamics, 135, 144