Control Engineering

Series Editor
William S. Levine
Department of Electrical and Computer Engineering
University of Maryland
College Park, MD 20742-3285
USA

Editorial Advisory Board

Okko Bosgra William Powers
Delft University Ford Motor Company (retired)
The Netherlands USA

Graham Goodwin Mark Spong

University of Newcastle University of Illinois
Australia Urbana-Champaign
USA
Petar Kokotovic
University of California lori Hashimoto
Santa Barbara Kyoto University
USA Kyoto, Japan

Manfred Morari
ETH
ZUrich, Switzerland
Keqin Gu
Vladimir L. Kharitonov
Jie ehen

Stability of Time-Delay Systems

Springer Science+Business Media, LLC

KeqinGu Vladimir L. Kharitonov
Southem illinois University at Edwardsville CINVESTAV-IPN
Department of Mechanical and Department of Automatic Control
Industrial Engineering A.P. 14-740, Mexico, D.F.
Edwardsville, IL 62026-1805 Mexico
USA

Jie Chen
University of California at Riverside
Department of Electrical Engineering
Riverside, CA 92521
USA

Library of Congress Cataloging-in-Publication Data

Gu, Keqin, 1957-

Stability of time-delay systems / Keqin Gu, Jie Chen, Vladimir L. Kharitonov.
p. cm.- (Control engineering)
Includes bibliographical references and index.
ISBN 978-1-4612-6584-9 ISBN 978-1-4612-0039-0 (eBook)
DOI 10.1007/978-1-4612-0039-0
1. Automatic control. 2. Time-delay systems. 3. Stability. 1. Chen, Jie, 1963- II.
Kharitonov, Vladimir, L. 1950- III. TitIe.

TJ213.G83 2003
629.~c21
2003041932
CIP

Printed on acid-free paper.

@2003Springer Science+Business Media New York
OriginaIly published by Birkhăuser Boston in 2003
Softcover reprint of the hardcover 1st edition 2003

AII rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC), except for brief excerpts in connec-
tion with reviews or scholarly analysis.Use in connection with any form ofinformation storage and retrieval,
electronic adaptation, computersoftware, or by similar or dissimilar methodology now known or hereafter
developed is forbidden .
The use of general descriptive names, trade names, trademarks , etc., in this publication, even if the former
are not especially identified, is not to be taken as a sign that such names, as understood by the Trade Marks
and Merchandise Marks Act, may accordingly be used freely by anyone.

ISBN 978-1-4612-6584-9 SPIN 10782604

Typeset by the authors.

987 6 5 4 3 2 1
KG dedicates the book to his parents Lijan Gu and Jieping Jiang for
their love and support.

VK dedicates the book to his teacher, the late Professor Vladimir

Ivanovich Zubov (1930-2000) for his friendship and mentorship.

JC dedicates the book to his wife Yan and sons Christopher and
Stephen for their love and trust.
Contents

Preface xiii

Notation xix

1 Introduction to Time-Delay Systems 1

1.1 Introduction . . . . . . . . . . . . 1
1.2 A simple time-delay system . . . 5
1.3 Functional differential equations 8
1.4 Stability of time-delay systems 10
1.4.1 Stability concept . . . . . 10
1.4.2 Lyapunov-Krasovskii Theorem 11
1.4.3 Razumikhin Theorem 13
1.5 Linear systems . . . . . . . . 15
1.6 Linear time-invariant systems 17
1. 7 Neutral time-delay systems 20
1.8 Outline of the text . . . . . 22
1.9 Notes 25
1.9.1 A brief historic note 25
1.9.2 Application examples 26
1.9.3 Analysis of time-delay systems 27

1 Frequency Domain Approach 29

2 Systems with Commensurate Delays 31

2.1 Introduction . . . . . . . . . . 31
2.2 Some classical stability tests . 34
2.2.1 2-D stability tests .. 34
2.2.2 Pseudo-delay methods 36
2.2.3 Direct method .. 38
2.3 Frequency-sweeping tests .. 44
Vlll Contents

2.4 Constant matrix tests 55

2.5 Notes 67
2.5.1 Classical results. 67
2.5.2 Frequency-sweeping and constant matrix tests 68

3 Systems with Incommensurate Delays 69

3.1 Introduction . . . . . . 69
3.2 Small gain/ J.L theorem 70
3.2.1 Small gain theorem. 70
3.2.2 Structured singular value 77
3.3 Frequency-sweeping conditions 79
3.4 Computational complexity analysis 87
3.4.1 Basic complexity concepts. 88
3.4.2 Proof of NP-hardness 90
3.5 Sufficient stability conditions 96
3.5.1 Systems of one delay. 96
3.5.2 Systems of multiple delays. 103
3.6 Neutral delay systems 106
3.7 Summary . . . . . . . 113
3.8 Notes ........ . 114
3.8.1 Small gain theorem and J.L 114
3.8.2 Stability of systems with incommensurate delays 114
3.8.3 Complexity issues . . . . . . . . . . . . . 115
3.8.4 Sufficient conditions and neutral systems 115

4 Robust Stability Analysis 117

4.1 Uncertain systems . . . 117
4.2 Characteristic quasipolynomial 117
4.3 Zeros of a quasipolynomial. . 119
4.3.1 Exponential diagram . 120
4.3.2 Potential diagram 122
4.4 Uncertain quasipolynornial . 125
4.4.1 Value set . . . . . . 126
4.4.2 Zero exclusion principle 126
4.5 Edge Theorem . . . . . . . . . 129
4.5.1 Stability of an edge subfamily. 132
 Contents ix

4.5.2 Interval quasipolynomial. . 134

4.6 Multivariate polynomial approach. 135
4.6.1 Multivariate polynomials . 136
4.6.2 Stable polynomials . . . . . 137
4.6.3 Stability of an interval multivariate polynomial 140
4.6.4 Stability of a diamond family of multivariate
polynomials . 142
4.7 Notes 144

II Time Domain Approach 145

5 Systems with Single Delay 147

5.1 Introduction........................... 147
5.2 Delay-independent stability criteria based on the Razumikhin
Theorem. . . . . . . . . 151
5.2.1 Single delay case . . . . . . . . . . . . . 151
5.2.2 Distributed delay case . . . . . . . . . . 154
5.3 Simple delay-dependent stability criteria based
on the Razumikhin Theorem . . . . . . . . . . 156
5.3.1 Model transformation . . . . . . . . . . 157
5.3.2 Simple delay-dependent stability criteria using
explicit model transformation . . . . . . . . . . 158
5.3.3 Additional dynamics . . . . . . . . . . . . . . . 160
5.3.4 Simple delay-dependent stability criteria using
implicit model transformation. . . . . . 165
5.4 Delay-independent stability criteria based
on the Lyapunov-Krasovskii Stability Theorem 169
5.4.1 Systems with single delay . . . . . . . . 169
5.4.2 Systems with distributed delays. . . . . 171
5.5 Delay-dependent stability criteria using a simple
Lyapunov-Krasovskii functional. . . . . . . . . . . . . . . . 172
5.5.1 Stability criteria using explicit model transformation 172
5.5.2 Stability criteria using implicit model transformation 173
5.6 Complete quadratic Lyapunov-Krasovskii functional 175
5.6.1 Introduction ............... . 175
5.6.2 Fundamental solution and matrix UW(T) 177
5.6.3 Lyapunov-Krasovskii functionals . . . . . 179
x Contents

5.7 Discretized Lyapunov functional method for systems with

single delay . . . . . . 182
5.7.1 Introduction . 182
5.7.2 Discretization . 183
5.7.3 Lyapunov-Krasovskii functional condition 185
5.7.4 Lyapunov-Krasovskii derivative condition 188
5.7.5 Stability criterion and examples. . . . . . 191
5.8 Notes 193
5.8.1 Results based on the Razumikhin Theorem 193
5.8.2 Model transformation and additional dynamics 194
5.8.3 Lyapunov-Krasovskii method 194

6 Robust Stability Analysis 197

6.1 Introduction . . . . . . . 197
6.2 Uncertainty characterization. 197
6.2.1 Polytopic uncertainty 198
6.2.2 Subpolytopic uncertainty 199
6.2.3 Norm-bounded uncertainty 201
6.2.4 Block-diagonal uncertainty 205
6.3 Robust stability criteria based on the Razumikhin Theorem 207
6.3.1 Delay-independent stability for systems with single
delay . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
6.3.2 Delay-independent stability criteria for systems with
distributed delays . . . . . . . . . . . . . . . . . . . 209
6.3.3 Delay-dependent stability criteria with explicit model
transformation . . . . . . . . . . . . . . . . . . . . . 211
6.4 Delay-independent stability criteria using the
Lyapunov-Krasovskii functional . . . . . 213
6.4.1 Systems with single delay . . . . . . . 213
6.4.2 Systems with distributed delays. . . . 218
6.5 Delay-dependent stability criteria using simple
Lyapunov-Krasovskii functional . . . . . . 219
6.6 Complete quadratic Lyapunov-Krasovskii
functional approach . . . . . . . . . . . . 224
6.7 Discretized Lyapunov functional method for systems with
single delay . . . . . . . . . . . . . 227
6.7.1 General case . . . . . . . . 227
6.7.2 Norm-bounded uncertainty 229
 Contents xi

6.8 Notes 231

6.8.1 Uncertainty characterization . . . . . . . . . . . . . 231
6.8.2 Stability results based on the Razumikhin Theorem
and Lyapunov-Krasovskii functional . . . . . . . . . 232

7 Systems with Multiple and Distributed Delays 233

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . 233
7.2 Delay-independent stability criteria of systems with multiple
delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
7.3 Simple delay-dependent stability criteria of systems with
multiple delays . . . . . . . . . . . . . . . . . . . . . . . . . 235
7.4 Complete quadratic functional for general linear systems . . 238
7.5 Discretized Lyapunov functional method for systems with
multiple delays . . . . 241
7.5.1 Problem setup . . . . . . . . . . . . . . . 242
7.5.2 Discretization................ 244
7.5.3 Lyapunov-Krasovskii functional condition 246
7.5.4 Lyapunov-Krasovskii derivative condition 251
7.5.5 Stability condition and examples . . . . . 258
7.6 Discretized Lyapunov functional method for systems with
distributed delays. . . . . 260
7.6.1 Problem statement 261
7.6.2 Discretization . . . 262
7.6.3 Lyapunov-Krasovskii functional condition 263
7.6.4 Lyapunov-Krasovskii derivative condition 265
7.6.5 Stability criterion and examples. 270
7.7 Notes 271

III Input-Output Approach 273

8 Input-Output Stability 275

8.1 Introduction . . . . . . 275
8.2 Input-output stability . . . . . 276
8.3 Method of comparison systems 279
8.3.1 Problem setup . . . . . 279
8.3.2 Frequency domain approach. 281
8.3.3 Time domain approach ... 284
xii Contents

8.4 Scaled small gain problem . . . . . . . . . . . 287

8.5 Robust stability under dynamical uncertainty 292
8.5.1 Problem setup . . . . . . . . 292
8.5.2 Uncertainty characterization 293
8.5.3 Robust small gain problem . 295
8.6 Approximation of delay elements .. 296
8.6.1 Approximation of time-varying delay. 296
8.6.2 Approximation of distributed delays 299
8.6.3 Approximation by multiple delays 303
8.7 Notes 307

A Matrix Facts 309

A.l Notation. 309
A.2 Determinant . 310
A.3 Eigenvalue problems 310
A.4 Singular value decomposition 312
A.5 Norms 312
A.6 Matrix measure . 314
A.7 Kronecker product and sum 314
A.8 Notes 314

B Linear Matrix Inequalities and Quadratic Integral

Inequalities 315
B.l Basic LMI problem. . . . . . . . . . . . . . . . . . . 315
B.2 GEVP . . . . . . . . . . . . . . . . . . . . . . . . . . 316
B.3 Transforming nonlinear matrix inequalities to LMI form 318
B.4 S-procedure. . . . . . . . . . . 319
B.5 Elimination of matrix variables 319
B.6 Quadratic integral inequalities. 322
B.7 Notes 324

Bibliography 327

Index 351
Preface
This book is devoted to the study of stability of time-delay systems, an
area long pursued by mathematicians, physical and life scientists, engineers,
and economists. Given the history and maturity of the discipline, one may
ask what new material could be presented beyond the extensive existing
body of work. The answer lies in the enduring and surprising vitality of the
subject: in particular, over the last decade, research has yielded a number of
notable results. The study of time-delay systems has therefore undergone
a significant conceptual and practical leap. Our aim in this book is to
present some of the highlights of these advances and to further develop the
techniques and tools along recent trends.
By definition, the future evolution of a time-delay system depends not
only on its present state but also on its history. This particular cause and
effect relationship can be most succinctly captured and has been tradi-
tionally modeled by differential-difference equations, or more generally, by
functional differential equations. While in practice many dynamical systems
may be satisfactorily described by ordinary differential equations alone, for
which the system's future evolution depends solely on its current state,
there are times when the delay effect cannot be neglected. In short, time
delay is hardly a matter of rarity; numerous examples presented in this
book and elsewhere attest to its prevalence. And, because of their ubiqui-
tous presence, time-delay systems have been an active area of scientific re-
search in a wide range of natural and social sciences: they are often known
as hereditary systems, systems with after-effect, systems with time-lag,
and more generally, as a subclass of functional differential equations and
infinite-dimensional systems.
The field of time-delay systems had its origins in the 18th century, and
it received substantial attention in the early 20th century in works devoted
to the modeling of biological, ecological, and engineering systems. Stability
of time-delay systems became a formal subject of study in the 1940s, with
contributions from such towering figures as Pontryagin and Bellman. Over
the years its interest and popularity have grown steadily. In particular, in
the last 10 to 15 years there has been a surge in research and a prolifera-
tion of new techniques and results. A cursory glance at the large numbers
of articles published in international conferences, organized workshops, and
archive journals during this period indicates the scale and magnitude of this
progress. Advances in numerical methods and control theory, especially the
theory of robust stability and control, have had substantial impact on the
XIV Preface

field. We observe that the small gain theorem made it possible to formulate
the stability problem of time-delay systems as one of robust stability and
of the structured singular value. Techniques in robust stability analysis of
uncertain polynomials find generalizations to uncertain quasipolynomials,
which serve as models for uncertain time-delay systems. Efficient numeri-
cal algorithms for solving linear matrix inequalities (LMI) have generated
substantial interest to pose stability problems as LMI conditions.
Recognizing these advances, we concluded that a book project would be
the natural outcome of more than a decade's creative work which we hope
will contribute to future studies of the subject. Our primary objective is to
present a sufficiently thorough yet focused treatment of the key methods
and results of the past 10 years or so, and to systematically organize and
discuss in sufficient technical depth many of the results that are otherwise
scattered in various journals and conference proceedings. Since the volume
of such publications can be overwhelming, we hope this book will be a
convenient reference and will provide quick access to important issues and
developments in time-delay systems.
Our particular viewpoint is centered on computability, robust stability
and robust control. We emphasize conceptual significance over technical
details, without sacrificing mathematical rigor. We stress fundamental, in-
tuitive observations behind methods and results, and we try to build links
and relationships among concepts that, at first glance, may seem unre-
lated. In other words, we try to provide an "insider's story" based on our
experience. This perspective has guided us in both the selection and presen-
tation of material: while the facts and results are certainly worthwhile, we
also hope that the ideas and intuition will be helpful in understanding the
nitty-gritty technical details, especially to new readers. On the other hand,
we face the unpleasant yet inevitable task of limiting ourselves to some
selected topics, excluding materials that to some may be equally or even
more important. Examples include classical results, recent developments
on stabilization, control, and filtering of time-delay systems, and those on
stochastic time-delay systems, to name just several. Of course, this selec-
tion merely reflects the authors' viewpoints and preferences rather than
the value of the particular issues. Fortunately, we have been able to refer
readers to several classic and recently-published works on these topics.
Despite widespread interest, it is still uncommon to find time-delay sys-
tems adopted as a regular course at most institutions. For many, the first
and perhaps only encounter with the subject is likely to take place in an
undergraduate course on systems modeling or automatic control. For this
reason, our book has been conceived and written as a monograph aimed
at researchers, practicing engineers, and graduate students. We do not pre-
sume, although we do not preclude the possibility, that it can serve as a
textbook for an advanced graduate course. It is nevertheless appropriate as
a self-study text. Most of the book is easily accessible to a typical second-
 Preface xv

year graduate student in engineering with basic knowledge of state-space

and transfer function descriptions and stability concepts of dynamical sys-
tems. A few specialized topics will require more advanced mathematical
background. Readers with less preparation may opt to skip the proofs of
such results, but quickly move on to the theorem statements. The book can
be read according to the reader's interests: for those interested mainly in
frequency-domain results, the first four chapters can be read in order. For
time-domain techniques, the reader may begin with Chapter 1 and then
move directly to Chapters 5 to 7. Most of the material in Chapter 8 can be
understood after Chapters 5 and 6. The book has been made sufficiently
self-contained, with necessary technical preliminaries integrated into the
respective chapters and included in two appendices.
The contents are centered on the theme of stability and robust stability
of time-delay systems. Chapter 1 begins with a number of practical exam-
ples in which time delays play an important role. It then continues with
an introductory exposition of some basic concepts and results essential to
stability analysis, such as functional differential equation representation,
characteristic quasipolynomials, the Lyapunov-Krasovskii Stability Theo-
rem, and the Razumikhin Theorem.
The next three chapters develop frequency-domain criteria for stability
and robust stability of linear time-invariant systems. Chapter 2 focuses
on systems with commensurate delays only. It presents both frequency-
sweeping and constant matrix tests, which are both necessary and suffi-
cient conditions for delay-dependent and delay-independent stability, and
both require computing the generalized eigenvalues of matrix pairs. Chap-
ter 3 studies systems with incommensurate delays. The development is built
upon a small gain approach, and the necessary and sufficient condition for
stability independent of delay is shown to be equivalent to a problem of
computing the structured singular value. The chapter opens with a brief ex-
posure to key concepts found in robust stability analysis, such as the small
gain theorem and the structured singular value, and ends with a formal
analysis of the computational complexity inherent in the stability prob-
lem. Both Chapters 2 and 3 treat systems modeled either by state-space
descriptions or quasipolynomials, but with no consideration of system un-
certainty. Uncertain time-delay systems are addressed in Chapter 4. More
specifically, this chapter examines uncertain quasipolynomials of systems
with incommensurate delays, that is, families of multivariate polynomials
whose coefficients are permitted to vary in a prescribed set; notable exam-
ples in this class include interval and diamond quasipolynomials. In much
the same spirit as in robust control, this chapter develops robust stability
conditions that ensure the stability of the entire family of quasipolynomi-
als, by checking the so-called edge, and in some instances only the vertices
of the quasipolynomial family.
xvi Preface

Chapters 5 to 7 are devoted to time-domain methods. In this setting,

fundamental tools such as the Lyapunov-Krasovskii Theorem and Razu-
mikhin Theorem play a critical role. Chapter 5 concentrates on systems
with a single delay. While providing various stability conditions, this chap-
ter also introduces a number of techniques prevailing in the analysis of
time-delay systems, including model transformation, discretized Lyapunov
functional methods, and LMI conditions. This paves the way for extensions
to uncertain systems with memoryless uncertainty and systems with mul-
tiple incommensurate delays, pursued in Chapters 6 and 7, respectively. It
is worth noting that unlike in frequency-domain approaches, time-domain
methods are more advantageous in accommodating nonlinear, time-varying
systems. Furthermore, because of the nature of time-domain techniques,
these chapters deal exclusively with state-space descriptions, and the sta-
bility tests amount to sufficient conditions that are posed as solutions of
LMI programs.
The final chapter, Chapter 8, introduces an input-output stability for-
mulation. The main purpose here is to consider robust stability problems
under dynamical uncertainty. Also discussed is the method of comparison
systems and approximation of time-delay elements.
The book concludes with two appendices. Appendix A summarizes the
key matrix facts and identities used throughout the book, and Appendix
B provides basic concepts and techniques of LMIs.
The writing of any book invariably involves the participation of many
people. We begin by gratefully acknowledging the contributions of our fam-
ily. KG would like to express his heartfelt thanks to Xinxin, Siyao, and
Patrick; VK to Olga; and JC to Yan, Christopher, and Stephen. In a way,
this work is a tribute to them, because they had to tolerate the countless
hours which otherwise would have been better spent with them. Indeed,
had it not been for the support and understanding of our families, we
would have long since abandoned the project. Furthermore, KG dedicates
the book to his parents Lijian Gu and Jieping Jiang for their love and
support, VK to his teacher the late Professor Vladimir Ivanovich Zubov
(1930-2000) for his friendship and mentorship, and JC to his wife Yan and
sons Christopher and Stephen for their love and trust. The authors would
also like to thank their personal friends, who never cease to be a source of
inspiration. These fine people have often urged us to clear the hidden hur-
dles of the writing process. We hope the book measures up to their vision
and expectations.
Of the friends and colleagues with whom we had the fortune to work
on problems directly pertinent to this book, we would like to acknowledge
Guoxiang Gu, Qing-Long Han, Silviu-Iulian Niculescu, Bahram Shafai, An-
dre Tits, Demin Xu, and Alexey Zhabko. One of the more heartening
fringe benefits of this project has been the opportunity to revisit some
of our collaborative work with these individuals; a substantial part of the
 Preface xvii

book is based on such work. We are equally grateful to Chaouki Abdal-

lah, Jean-Michel Dion, Luc Dugard, Michael Fan, Didier Georges, Jacob
Kogan, Brad Lehman, James Louisell, Olivier Sename, Gilead Tadmor,
Sophie Tarboriech, Onur Toker, Li Qiu, Erik Verriest, and Kemin Zhou.
We benefited from their ideas and wisdom through numerous discussions.
They contributed either directly to the book via joint work, or indirectly by
making suggestions and imparting to us their discerning taste. We thank
Birkhauser for the opportunity to publish the book, and its fine editorial
staff for their patience and technical support. They have leniently accom-
modated several "delays" which we hope have paid off. We thank Rex Pierce
and Gang Chen, who each helped prepare a number of computer-generated
graphs. Finally, we gratefully acknowledge the US National Science Foun-
dation for its financial support during the period of this project, under the
direction of Kishan Baheti and Rose Gombay, respectively.
In closing we ask the reader for the privilege of a moment of reflection
on our collaborative process. It may seem somewhat odd that the authors
never had a record of working together before initiating such a major,
time-consuming project. The idea was first initiated by KG at the 1997
American Control Conference held at Albuquerque, New Mexico, during a
casual conversation between KG and JC. Writing a book together seemed
at first a remote possibility. It remained so until the summer of 1998, when
at a robust control workshop held in Sienna, Italy, VK and JC rekindled
the idea. The picturesque setting of Tuscany's rolling hills and stunning
sunsets seemed to promise a happy ending, and the project then took off.
The three authors had discussions via emails and phone calls, conceiving
the contents and carving out a game plan. It would take another four
years to implement the plan, and after numerous drafts and iterations, the
project was finally brought to its present form. We have attempted in good
faith to communicate our views and passion for the subject. The book, in
retrospect, is a triumph of will and friendship. We thus thank each other
as well.

Keqin Gu
Southern Illinois University at Edwardsville

Vladimir L. Kharitonov
CINVESTAV-IPN, Mexico

Jie Chen
University of California at Riverside

October 2002
Notation
Sets and Operators

The sets of real numbers, n component real

vectors, and n by n real matrices
The sets of complex numbers, n component
complex vectors, and n by n complex matrices
The open unit disk {s I sEC, lsi < I}
The set of positive real numbers
The unit circle {s I sEC, lsi = I}
The set of integers
The closure, interior, and boundary of §, where § is
any set. For example, iiil is the closed unit disk.
§n {(Sl, S2, •.. , sn) I Si E §, i = 1,2, ... , n}, § any set
c+ The set {w E C I Re(w) > O}
L2 The set of real square integrable functions
£2 The set of of functions of jw integrable for w
?-leo The set oftransfer functions H(s), with IIHlleo < 00
Re(w),Im(w) The real and imaginary parts of w E C
w* The complex conjugate of w
11·11 Vector or matrix norm.
C[a,b] The set of JRn-valued continuous functions on [a, b]
C C[-r, OJ
1Ict>lI c The continuous norm max 11ct>(~)11 for ct> E C[a, bj
a~e9
X Derivative of x with respect to time t, ~~
£(.),£-1(-) Laplace transform and the inverse Laplace transform

Matrix-related notation
Thanspose, component-wise complex conjugate,
and Hermitian (conjugate) transpose of matrix A.
IAI Matrix formed by taking absolute value of each
entry of A

°
A> 0, A 2:': 0 The matrix A is positive (semi)definite « or ::; similar)
A>B,A2:':B A - B > 0, A - B 2:':
A(A), Ai(A) An eigenvalue and the ith eigenvalue of matrix A
xx Notation

A(A,B) A generalized eigenvalue of matrix pair (A, B)

Amax(A) , Amin(A) The maximum and the minimum eigenvalue of
Hermitian matrix A
p(A) Spectral radius of matrix A, m~ IAi(A)1
t
a(A) Spectrum (the set of all the eigenvalues) of
matrix A
a(A,B) Spectrum of matrix pair (A, B)
ai(A) The ith singular value of A
amax(A) or o-(A) The maximum singular value of A
amin(A) or Q:(A) The minimum singular value of A
v(A) Matrix measure of matrix A
A®B Kronecker product of A and B
AEBB Kronecker sum of A and B, A ® I + I ® B
1

Introduction to Time-Delay
Systems
1.1 Introduction
We begin this chapter by presenting a number of practical examples of time-
delay systems. This should give the reader a glimpse into how widely time
delays may occur in practice. To make the presentation more accessible,
we then analyze a simple time-delay system, which illustrates some of the
important concepts in a concrete manner. This analysis sets the stage for a
formal introduction to functional differential equations as a representation
of time-delay systems, and to such stability analysis tools as the Lyapunov-
Krasovskii Theorem and Razumikhin Theorem. Linear delay systems are
discussed in greater length with frequency domain descriptions. The chap-
ter is concluded with a brief outline of subsequent content.
Ordinary differential equations in the form of

i:(t) = f(t, x(t)) (1.1)

have been a prevalent model description for dynamical systems. In this

description, the variables x(t) E ]Rn are known as the state variables, and
the differential equations characterize the evolution of the state variables
with respect to time. A fundamental presumption on a system modeled as
such is that the future evolution of the system is completely determined
by the current value of the state variables. In other words, the value of the
state variables x(t), to :S t < 00, for any to, can be found once the initial
condition
x(to) = Xo (1.2)
is known. Needless to say, ordinary differential equations in general, and
stability and control of dynamical systems so modelled in particular, have
been an extensively developed subject of scientific learning.
In practice, however, many dynamical systems cannot be satisfactorily
modeled by an ordinary differential equation. In particular, for many sys-
tems, the future evolution of the state variables x(t) not only depends on
their current value x(to), but also on their past values, say x(~), to - r :S
~ :S to· Such a system is called a time-delay system. Time-delay systems
may arise in practice for many reasons.
2 1. Introduction to Time-Delay Systems

FIGURE 1.1. Geometry of turning

Example 1.1 Regenerative chatter in metal cutting. Shown in Fig-

ure 1.1 is the metal cutting process in a typical machine tool such as a
lathe. A cylindrical workpiece rotates with constant angular velocity wand
the cutting tool translates along the axis of the workpiece with constant
linear velocity wf /27r, where f is the feed rate in length per revolution cor-
responding to the normal thickness of the chip removed. The tool generates
a surface as the material is removed, shown as shaded, and any vibration
of the tool is reflected on this surface. In regenerative chatter, the surface
generated by the previous pass becomes the upper surface of the chip on the
subsequent pass. A model often used in studying such a process is shown
in Figure 1.2. This time-delay system can be described by the equation

my(t) + ciJ(t) + ky(t) = -FtU + yet) - yet - T)), (1.3)

where m, c, and k reflect the inertia, damping, and stiffness characteristics

of the machine tool, the delay time T = 27r / w corresponds to the time
for the workpiece to complete one revolution, and FtO is the thrust force
depending on the instantaneous chip thickness f +y( t) - y( t - T). It is often
sufficient to consider Ft (.) to be linear, and techniques for linear time-delay
systems are often used. 0

Notice that in the above example the system is autonomous (without

feedback control), and the delay is the result of an intrinsic property of the
system. In response to recent demands on high speed machining, the study
of the time-delay and nonlinear nature of the system is receiving renewed
interest.

Example 1.2 Internal combustion engine. In the control of internal

combustion engines, the Mean Torque Production Model is often used. In
 1.1 Introduction 3

FIGURE 1.2. Model of regenerative chatter

this model, the crankshaft rotation is modeled by the motion equation

(1.4)
where Ti is the indicated torque generated by the engine, which is delayed
by Ti seconds due to engine cycle delays, resulted from, e.g., fuel-air mixing,
ignition delay, and cylinder pressure force propagation; Tf represents the
friction, Tload the load, J the moment of inertia, and w the angular velocity
of the crankshaft. Feedback control action is applied to manipulate the
indicated torque T i , through the controller dynamics governed by

x(t) f(x(t),w(t», (1.5)

Ti(t) = h(x(t),w(t». (1.6)
Using (1.4) to (1.6), the closed-loop system becomes
1
wet) }[h(x(t - Ti),W(t - Ti» - Tf(t) - T1oad(t)],
x(t) f(x(t),W(t».
o
Note that unlike in Example 1.1, herein the delay results from feedback
control action, instead of from the system itself; that is, due to the nature of
the system, the feedback is delayed. Another type of delay of similar effects
may be incurred due to delayed measurements. In both delayed control and
delayed measurement, the delay is usually considered undesirable since it
has the tendency to deteriorate the system performance or even destabilize
the system.

Example 1.3 Delayed resonator. A delay resonator is a modification

of the classical vibration absorber shown in Figure 1.3. The main structure
4 1. Introduction to Time-Delay Systems

FIGURE 1.3. Classical vibration absorber

consists of an object with mass M, a spring with constant k, and a damper

with damping factor c, subject to a harmonic excitation f(t). The classi-
cal vibration absorber consists of m a, Ca, and ka, and is to be attached
to a main structure. With the parameters designed appropriately, this can
substantially reduce the amplitude of vibration in the structure. However,
the classical absorber suffers from a number of shortcomings, such as its
sensitivity to the excitation frequency. To enhance the performance, an ad-
ditional force proportional to the delayed displacement of M is introduced.
Such a structure is known as a delayed resonator and is shown in Figure
1.4. The equation of motion for the entire system can be written as

maxa(t) + ca(xa(t) - x(t)) + ka(xa(t) - x(t))

-gx(t - r) = 0,
Mx(t) + (c + ca)x(t) - CaXa(t) + (k + ka)x(t)
-kaxa(t) + gx(t - r) = f(t).

It is known that with a proper design, the delayed resonator can signif-
icantly enhance the system performance, in reducing sensitivity to the
excitation frequency, eliminating the main structure's vibration with the
presence of damping, and simultaneously suppressing two excitation fre-
quencies.D

In this example, the time delay also occurs by way of feedback control
action. However, in contrast to the delay effect in Example 1.2, the delay
at present is intentionally introduced to enhance the system performance.
Time-delay systems may also arise from simplification of some partial
differential equations. Furthermore, time-delay systems are used to model
dynamical systems in many other scientific disciplines including engineer-
ing, biology, ecology, and economics. The Notes section at the end of this
chapter will give a brief historic overview to discuss some related literature.
 1.2 A simple time-delay system 5

FIGURE 1.4. Delayed resonator

1.2 A simple time-delay system

We now acquaint the reader with some of the basic concepts of time-delay
systems by analyzing the following simple system:
x(t) = aox(t) + alX(t - r) + h(t). (1.7)
°
Here ao, ab and r are real numbers, and r > is the delay time. We want
to solve the equation from the time instant t = 0. In order to calculate X(O),
so as to advance the solution beyond the time instant t = 0, we need the

t = ~, °: ;
values of x(O) and x( -r). Similarly, in order to calculate x at the instant
~ < r, we need x(~) and x(~ - r). For such a ~, since -r ::;
~ -r < 0, x( ~ -r) cannot be generated in the solution process. It is therefore
clear that in order for the solution to be uniquely defined, it is necessary to
specify the value of x(t) for the entire interval -r ::; t ::; 0; in other words,
we need to specify the initial condition
x(t) = ¢(t), t E [-r, 0], (1.8)
with some given continuous function ¢ : [-r, 0] ~ R
Once the initial condition (1.8) is given, the solution is well defined.
Indeed, for t E [0, r], since x(t-r) is already known in this interval, we can
treat (1.7) as an ordinary differential equation, so as to obtain

x(t) = eaotx(O) + lot eao(t-u) [alx(u- r) + h(u)]du.

Upon obtaining x(t) for t E [0, r], we can calculate x(t) for t E [r,2r]
similarly, that is,
6 1. Introduction to Time-Delay Systems

Continuing this process will allow us to obtain the solution x(t), -r ::; t <
00. Such a process is known as the method of steps.

In the following development, we shall assume that the forcing function

h(t) is exponentially bounded, so that
Ih(t)1 ::; Ke ct for some K > 0, c E lR,
and that its Laplace transform

H(s) = C[h(t)]
exists. It can be shown that the solution of (1.7) satisfies the exponential

lt
bound
Ix(t)1 ::; ae bt (11¢llc + Ih(u)ldU)

for some a > 0 and b > 0, where

11¢llc = -r:S(I:SO
max I¢(O)I

is the continuous norm of ¢. This implies that the Laplace transform of x(t)
exists as well, and as such it can be used to solve the equation. Indeed, if
we take the Laplace transform of (1.7), together with the initial condition
(1.8), we obtain

sX(s) - ¢(O) = aoX(s) + al [e- sr X(s) + lOr e-S(u+rl¢(U)dU] + H(s),

where X(s) is the Laplace transform of x(t). Solving for X(s), we obtain

where
~(s) = s - ao - ale- rs (1.10)
is the characteristic quasipolynomial of the system. Let <I>(t) denote the
inverse Laplace transform of l/~(s),

<I>(t) = C-l[l/~(s)]. (1.11)

Then <I>(t) is known as the fundamental solution of (1.7). It can be seen

from (1.9) that <I>(t) is the solution of (1.7) with

h(t) = 0
and the initial condition

x(O) = { ~ 0=0,
-r ::; 0 < O.
 1.2 A simple time-delay system 7

Given the fundamental solution (1.11), it is easy to see, using the convolu-
tion theorem, that the solution of (1.7) under an arbitrary initial condition
¢(t) and forcing function h(t) can be expressed as

x(t) = q>(t)¢(O) + a1 lOr q>(t - u - r)¢(u)du + lot q>(t - u)h(u)du. (1.12)

Equation (1.12) is called the Cauchy formula (also known as the variation-
of-constant formula) for the system, which explicitly expresses the solution
in terms of the initial condition ¢ and the forcing function h.
From (1.12) it is clear that the growth of the general solution for any
bounded h is intimately related to the rate of exponential growth of the fun-
damental solution q>, which is determined by the poles (also known as the
characteristic roots) of the system, i.e., the solutions of the characteristic
equation
~(s) = O. (1.13)
Unlike for systems without delay, generally (1.13) has an infinite number
of solutions. According to complex variables theory, however, since ~(s) is
an entire function, it cannot have an infinite number of zeros within any
compact set lsi::; R, for any finite R > O. Therefore, "most" of the system's
poles go to infinity.
To understand the distribution of the poles, notice that (1.10) and (1.13)
imply that
(1.14)
When s ----> 00, the left hand side of (1.14) approaches infinity, and therefore
the right hand side must also approach infinity. But this means that
lim Re(s) = -00.
s---oo
Therefore, for any given real scalar a, there are only a finite number of
poles of the system with real parts greater than a. With this in mind, let
Si, i = 1, 2, ... be the poles of the system, and any a such that
i = 1, 2, ....
Using the residual theorem from complex variables theory, we may write
(1.11) as

2:
['JI.+jOO
q>(t) = Jo . e- st ~ -l(s)ds + Res[e sit ~ -l(Si)]. (1.15)
a.-JOO Re(si»a.

It can be shown that the first term satisfies the condition

for t > O.
For the particular time-delay system (1. 7), we may now summarize the
above discussion in the following proposition.
8 1. Introduction to Time-Delay Systems

Proposition 1.1 For any a E JR, there are only a finite number of poles
with real parts greater than a. Let h(t) == 0 and Si, i = 1, 2, "', be the
poles of the system (1. 7), and let

ao = m?J(Re(si) .
•
Then for any a > ao, there exists a L > 0 such that the solution of (1.7)
with the initial condition (1.8) satisfies the inequality

Thus, in order for the solution to approach zero under an arbitrary initial
condition, it is sufficient that all the poles of the system have negative real
parts; in fact, this is also neccesary. This conclusion is the same as that for
systems without delay.

1.3 Functional differential equations

We can use functional differential equations to describe time-delay systems.
To formally introduce the concept of functional differential equations, let
C([a,b],JRn) be the set of continuous functions mapping the interval [a,b]
to JRn. In many situations, one may wish to identify a maximum time delay
r of a system. In this case, we are often interested in the set of continuous
functions mapping [-r, 0] to JRn, for which we simplify the notation to
C = C( [-r, 0], JRn). For any A > 0 and any continuous function of time
'If; E C([to - r, to + A], JR n ), and to ::; t ::; to + A, let 'If;t E C be a segment
of the function 'If; defined as 'If;t((}) = 'If;(t + (}), -r ::; (} ::; O. The general
form of a retarded functional differential equation (RFDE) (or functional
differential equation of retarded type) is

x(t) = f(t, Xt), (1.16)

where x(t) E JRn and f : JRxC --+ JRn. Equation (1.16) indicates that the
derivative of the state variables x at time t depends on t and x(~) for
t - r ::; ~ ::; t. As such, to determine the future evolution of the state, it
is necessary to specify the initial state variables x(t) in a time interval of
length r, say, from to - r to to, i.e.,

Xto = ¢, (1.17)

where ¢ E C is given. In other words, x(to + tJ) = ¢(tJ), -r ::; tJ ::; O.

 1.3 Functional differential equations 9

For examples of RFDEs, consider

x(t) ax(t) + b(x - r) + coswt, (1.18)

x(t) (2 + sinwt)x(t), (1.19)

x(t) = lOr c(B)x(t + B)dB, (1.20)

x(t) = ax(t)x(t - r). (1.21)

Here (1.18) represents a nonhomogeneous linear time-invariant system with

a single delay. Equation (1.19) is a linear time-varying ordinary differential
equation, which is a special functional differential equation. Equation (1.20)
represents a system with distributed delay and (1.21) represents a nonlinear
system. An RFDE may also involve higher order derivatives, known as
a higher order RFDE. As with differential equations without delay, we
may introduce additional variables to transform a higher order RFDE to a
standard first order functional differential equation of the form (1.16). For
example,
y(t) + 2y(t -1) + y(t) = 0
can be written as

with the variables

Xl (t) y(t),
X2(t) y(t).
It is important to point out that in an RFDE, the highest order derivative
does not contain any delayed variables. When such a term does appear, then
we have a functional differential equation of neutral type. For example,

5x(t) + 2x(t - r) + x(t) - x(t - r) = 0

is a neutral functional differential equation (NFDE). In this book, we will

mainly focus on RFDEs, and thus for simplicity will often refer to an RFDE
as a functional differential equation. We will also use the terms time-delay
system and functional differential equation interchangeably. Delay systems
of neutral type, i.e., systems described by NFDEs, will only be discussed
briefly, and will be spelled out where they appear.
For an A > 0, a function X is said to be a solution of (1.16) on the
interval [to - r, to + A) if within this interval X is continuous and satisfies
the RFDE (1.16). Here the derivative with respect to time t should be
interpreted as a one-sided derivative in the forward direction. Of course a
solution also implies that (t, xd is within the domain of the definition of
10 1. Introduction to Time-Delay Systems

I. If the solution also satisfies the initial condition (1.17), we say that it
is a solution of the equation with the initial condition (1.17), or simply a
solution through (to, ¢). We denote it as x(to, ¢, I) when it is important
to specify the particular RFDE with the given initial condition. The value
of x(to, ¢, I) at t is denoted as x(t; to, ¢, I). We will omit I to write x(to, ¢)
or x(t;to,¢) when I is obvious from the context.
A fundamental issue in the study of both ordinary differential equations
and functional differential equations is the existence and uniqueness of so-
lution. It is not our aim in this book, however, to delve into this issue for
RFDEs. Instead, we state the following theorem without proof.

Theorem 1.2 Suppose that 0 is an open set in lRxC, I : 0 --+ lRn is

continuous, and I(t, ¢) is Lipschitzian in ¢ in each compact set in 0, that
is, lor each given compact set 00 C 0, there exists a constant L, such that

1I/(t'¢l) - l(t,¢2)11 :::; LII¢l - ¢211

lor any (t'¢l) E 00 and (t,¢2) E 00. II (to,¢) E 0, then there exists a
unique solution 01 (1.16) through (to, ¢ ) .

In nearly all cases the systems treated in this book satisfy the conditions
in this theorem, and hence the existence and uniqueness of the solutions
are ensured. Exceptions will be pointed out explicitly.

1.4 Stability of time-delay systems

1.4.1 Stability concept

Let y(t) be a solution of the RFDE (1.16). The stability of the solution
concerns the system's behavior when the system trajectory x(t) deviates
from y(t). Throughout this book, without loss of generality, we will assume
that the functional differential equation (1.16) admits the solution x(t) = 0,
which will be referred to as the trivial solution. Indeed, if it is desirable to
study the stability of a nontrivial solution y(t), then we may resort to the
variable transformation z(t) = x(t) - y(t), so that the new system

i(t) = I(t, Zt + Yt) - I(t, Yt) (1.22)

has the trivial solution z(t) = O.
For a function ¢ E C([a,bj,lR n ), define the continuous norm 11·ll eby
11¢lle = a::::;9::;b
max II¢(O)II·

In this definition, the vector norm 11·11 represents the 2-norm 11.11 2 , It should
be pointed out that other norms such as I-norm and oo-norm can also be
 1.4 Stability of time-delay systems 11

used in most of our subsequent development, as long as the same norm is

used consistently.

x(t) = °
Definition 1.1 For the system described by (1.16), the trivial solution

°
is said to be stable if for any to E JR. and any E > 0, there
exists a 0 = o(to, E) > such that Ilxto lie < 0 implies Ilx(t)11 < E for t ::::: to·

°
It is said to be asymptotically stable if it is stable, and for any to E JR. and
any E > 0, there exists a oa = Oa(to, E) > such that Ilxtolle < Oa implies
lim x( t) = 0. It is said to be uniformly stable if it is stable and 0(to, E)
t->oo

°
can be chosen independently of to. It is uniformly asymptotically stable if
it is uniformly stable and there exists a Oa > such that for any TJ > 0,
there exists aT = T(Oa,TJ), such that Ilxtolle < 0 implies Ilx(t)11 < TJ for
t ::::: to + T and to E JR.. It is globally (uniformly) asymptotically stable if
it is (uniformly) asymptotically stable and Oa can be an arbitrarily large,
finite number.

One should note that the stability notions herein are not at all different
from their counterparts for systems without delay, modulo to the different
assumptions on the initial conditions. Note also that in this book, we shall
only be concerned with asymptotic stability. Thus, while the Lyapunov-
Krasovskii theorem and Razumikhin theorem will be stated in complete-
ness in the next two sections, inclusive of both stability and asymptotic
stability, elsewhere an asymptotically stable system is often simply said to
be "stable." Correspondingly, we will sometimes refer to a system that is
not asymptotically stable as "unstable."

1.4.2 Lyapunov-Krasovskii Theorem

As in the study of systems without delay, an effective method for deter-
mining the stability of a time-delay system is the Lyapunov method. For
a system without delay, this requires the construction of a Lyapunov func-
tion V(t, x(t)), which in some sense is a potential measure quantifying the
deviation of the state x(t) from the trivial solution 0. Since for a delay-free
system x(t) is needed to specify the system's future evolution beyond t, and
since in a time-delay system the "state" at time t required for the same
purpose is the value of x(t) in the interval [t - r, t], i.e., Xt, it is natural to
expect that for a time-delay system, the corresponding Lyapunov function
be a functional V(t, Xt) depending on Xt, which also should measure the
deviation of Xt from the trivial solution 0. Such a functional is known as a
Lyapunov-K rasovskii functional.
More specifically, let V (t, ¢) be differential, and let Xt ( T, ¢) be the solu-
tion of (1.16) at time t with the initial condition X-r = ¢. We may calculate
the derivative of V(t, xt} with respect to t and evaluate it at t = T. This
12 1. Introduction to Time-Delay Systems

gives rise to

V(T,¢) ~V(t,Xt)lt=T, Xt=¢

= limsup! [V(T+At,XT+~t(T,¢)) - V(T,¢)).
~t-O I.J.t

Intuitively, a nonpositive V indicates that Xt does not grow with t, which

in turn means that the system under consideration is stable in light of De-
finition 1.1. The more precise statement of this observation is the following
theorem.

Theorem 1.3 (Lyapunov-Krasovskii Stability Theorem) Suppose

f : IRxC --+ IRn in (1.16) maps IR x (bounded sets in C) into a bounded
sets in IR n , and that u, v, w : i+ --+ i+ are continuous nondecreasing
functions, where additionally u(s) and v(s) are positive for s > 0, and
u(O) = v(O) = O. If there exists a continuous differentiable functional
V : IR x C --+ IR such that

u(II¢(O)ID :::; V(t, ¢) :::; v(II¢lle)

and
V(t, ¢) :::; -w(II¢(O)II),
then the trivial solution of (1.16) is uniformly stable. Ifw(s) > 0 for s > 0,
then it is uniformly asymptotically stable. If, in addition, lim s _ oo u(s) = 00,
then it is globally uniformly asymptotically stable.

Proof. For any c > 0, we can find a sufficiently small 8 = 8(c) > 0 such
that v(8) < u(c). Hence, for any initial time to and any initial condition
Xto = ¢ with 11¢lle < 8, we have V(t,Xt) :::; 0, and therefore V(t,Xt) :::;
V(t, ¢), for any t ~ to. This implies that

u(llx(t)11) :::; V(t, Xt) :::; V(to, ¢) :::; v(II¢lie) :::; v(8) < u(c),
which in turn implies that Ilx(t)11 < c for t ~ to. This proves the uniform
stability.
To prove uniform asymptotic stability, let c > 0 and 8a > 0 such that
v(8 a ) < u(c). Then IIxtolle :::; 8a implies that IIx(t)1I < c for t ~ to. For this
8a and any TJ > 0, we need to show that there exists aT = T(8 a , TJ), such
that IIx(t)1I < TJ for t ~ to + T. Let 8b > 0 satisfy v(8 b ) < u(TJ). Then, it
suffices to show that IIXto+Tlle < 8b , which implies u(lIx(t)11) :::; V(t,Xt) <
v(8 b ) < u(TJ) for t ~ to+T and therefore IIx(t)1I < TJ. We establish this fact
by contradiction. Suppose that this is not true; in other words, no such T
exists. It follows that IIxtlle ~ 8b for all t ~ to. This means that a sequence
{tk} exists such that
to + (2k - l)r :::; tk :::; to + 2kr, k = 1,2"" ,
 1.4 Stability of time-delay systems 13

and
Ilx(tk)112 Db.
For a sufficiently large L, due to the assumption on f, we have 1I±(t)1I =
Ilf(t,xt)11 < L for all t 2 to. Therefore, for t E h = [tk - ft,tk + due ft],
to the Mean Value Theorem, for some B E [0,1],

IIx(t) II = Ilx(tk) + X(tk + B(t - tk))(t - tk)1I

> IIX(tk)II-II±(tk + B(t - tk))II'I(t - tk)1
Db
> Db - L· 2L
Db
2'
Therefore,
V(t, Xt) :::; -W(Db/2), for t E Ik
and V(t, Xt) :::; 0 otherwise. By increasing L if necessary, we can assume
that these intervals do not overlap, and hence

But this implies that V(tk, Xtk) < 0 for a sufficiently large k, which is a
contradiction. This proves the uniform asymptotic stability.
Finally, if lims-->CXl u(s) = 00, then Da above may be arbitrarily large,
and E can be chosen after oa is given to satisfy v(oa) < U(E), and therefore
global uniform asymptotic stability can be concluded. •
It is obvious from the above proof that u, V, W, and V (t, .) need only be
defined in a neighborhood of zero except for the case of global stability.
Notice also that the lower bound of V need only to be a positive function
of q';(0).

1.4.3 Razumikhin Theorem

That the Lyapunov-Krasovskii functional requires the state variable x(t)
in the interval [t - r, t] necessitates the manipulation of functionals, which
consequently makes the application of the Lyapunov-Krasovskii Theorem
rather difficult. This difficulty may sometimes be circumvented using the
Razumikhin Theorem, an alternative result involving essentially only func-
tions rather than functionals.
The key idea behind the Razumikhin Theorem also focuses on a function
V(x) representative of the size of x(t). For such a function,

V(Xt) = max V(x(t + B))

IIE[-r,OJ
14 1. Introduction to Time-Delay Systems

serves to measure the size of Xt. IfV(x(t)) < V(Xt), then Vex) > 0 does not
make V(Xt) grow. Indeed, for V(Xt) to not grow, it is only necessary that
V(x(t)) is not positive whenever V(x(t)) = V(Xt). The precise statement
is as follows.

Theorem 1.4 (Razumikhin Theorem) Suppose f : lRxC --+ lRn in

(1.16) takes lR x (bounded sets of C) into bounded sets of lRn , and u, v, w :
R+ --+ R+ are continuous nondecreasing functions, u(s) and v(s) are posi-
tive for s > 0, and u(o) = v(o) = 0, v strictly increasing. If there exists.a
continuously differentiable function V : lR x lRn --+ lR such that

u(llxlD ::; Vet, x) ::; v(llxlD, for tE lR and x E lRn , (1.23)

and the derivative of V along the solution x(t) of (1.16) satisfies

Vet, x(t)) ::; -w(llx(t)ID whenever Vet + e, x(t + e)) ::; Vet, x(t)) (1.24)

eE
° °
for [-r, OJ, then the system (1.16) is uniformly stable.
If, in addition, w(s) > for s > 0, and there exists a continuous nonde-
creasing function pes) > s for s > such that condition (1.24) is strength-
ened to

Vet, x(t)) ::; -w(llx(t)ID if Vet + e, x(t + e)) ::; p(V(t, x(t))) (1.25)

e
for E [-r, 0], then the system (1.16) is uniformly asymptotically stable.
If in addition lim s _ co u(s) = 00, then the system (1.16) is globally
uniformly asymptotically stable.

Proof. To prove uniform stability, for any given c > 0, let °

< 6 <
v- 1 (u(c)). Then for any given to and ¢, II¢II < 6, we have Veto +e, ¢(e)) ::;
v(6) < uk) for e E [-r,Oj. Let x be the solution of (1.16) with initial
condition Xto = ¢. According to (1.24), as t increases, whenever Vet, x(t)) =
v(6) and Vet + e, x(t + e)) ::; v(6) for e E [-r, OJ, Vet, x(t)) ::; 0. Due to the
continuity of Vet, x(t)), it is therefore impossible for Vet, x(t)) to exceed
v(6). In other words, we have Vet, x(t)) ::; v(8) < u(c) for t ~ to - r. But
this implies Ilx(t)11 ::; c for t ~ to - r.
To prove uniform asymptotic stability under the strengthened condition

°
(1.25), we need to show that for any sufficiently small 8a (such that there
exists an c > to satisfy v(8 a ) = uk)), and any arbitrarily given ""
", < c, there exists aT = T(8 a ,,,,), such that Ilxtll < ", for t ~ to + T,
°
<

II¢II < 8a and to E R

Let Vi satisfy ° < Vi < u(",) (say, Vi = O.gu(",)). The continuity of
function p implies that there exists an a > 0, such that a < pes) - s for
Vi ::; s < v(8 a ). Let N be the smallest positive integer to satisfy Vi +Na ~
v(8 a ), and let 'Y = min w(s). Let tk = to + k(ah + r). We will show
V-I (Vi):$;s:$;e
 1.5 Linear systems 15

that

V(t, x(t)) ~ min{Vi + (N - k)a, v(8 a )} for tk - r ~ t ~ tk (1.26)

implies
V(t, x(t)) ~ Vi + (N - k - l)a for t ~ tk + ah, (1.27)
especially, it implies (1.27) for tk+l - r ~ t ~ tk+!' An induction then
shows that V(t,x(t)) ~ Vi < u(1]) for t ~ tN = to + N(ah + r). This
implies jjx(t)jj ~ 1] for t ~ to +T, T = N(ah+r), which will complete the
proof.
Assume that V(tk' X(tk)) > Vi+(N -k-l)a. Then, according to condition
(1.25), there exists ad> 0 such that Vet, x(t)) ~ -'Y and

V(tk' X(tk)) ~ Vet, x(t)) > V(tk + d, X(tk + d)) = VI + (N - k - l)a

for tk ~ t < tk + d for some d > O. Clearly,

d ~ V(tk' X(tk)) - V(tk + d, X(tk + d)) ~a / T

'Y
Also, it is easy to see that the inequality in (1.27) holds for t ~ tk + d
(and therefore also holds for t ~ tk + ah) since Vet, x(t)) ~ 0 whenever
V(t,x(t)) = VI + (N - k - l)a. Thus the uniform asymptotic stability is
proven.
If lim s _ co u(s) = 00, then for arbitrary large 8a > 0 there exists an
c > 0 to satisfy v(8 a ) = u(c). This allows us to conclude global uniform
asymptotic stability. •

1.5 Linear systems

Of particular interest among RFDEs is the instance when f is linear with
respect to Xt. Systems in this category are linear retarded delay systems.
Linear delay systems are amenable to more analysis tools and hence enable
more in-depth studies. We shall first give a brief review of general linear
time-varying delay systems and next focus on linear time-invariant systems.
A general linear time-delay system can be described by the RFDE

x(t) = L(t)xt + h(t), (1.28)

where L(t) is a time-varying linear operator acting on Xt. In this case, it

is always possible to find a matrix function F : lR x [-r, 0] ---t lR nxn of
bounded variation, such that

F(t,O) = 0
16 1. Introduction to Time-Delay Systems

and
L(t)¢ = lOr do [F(t, O)]¢(O). (1.29)

Here in (1.29), the Stieltjes integral is required in general, and the subscript
o is used to indicate that () (rather than t) is the integration variable. As
such, a general linear RFDE can be represented as

x(t) = lOr do [F(t, O)]x(t + 0) + h(t). (1.30)

In particular, many linear RFDEs can be further specialized to

x(t) = t; Adt)x(t - rk(t)) + lr°A(t, O)x(t + O)dO + h(t),

K
(1.31)

where Ak(t) and A(t,O) are given n x n real continuous matrix functions,
and rk(t) are given continuous functions representing time-varying delays,
which can be ordered with no loss of generality, so that

Under such circumstances,

F(t, 0) = -1° max(II,-r)

A(t, r)dr - L
-rdt»11
Ak(t).

The fundamental solution <p(t, to) of the RFDE (1.30) is the n x n matrix
function satisfying the homogeneous equation

together with the initial condition

<p(to + 0, to) = {oJ 0-r:::;

=00< 0 (1.32)

Note that although this initial condition does not satisfy the existence
condition in Theorem 1.2, the existence and uniqueness of the solution can
nevertheless be established. In fact, with the fundamental solution so given,
the solution of the linear RFDE (1.30) with the initial condition Xto = ¢
can be expressed as

x(t;to,¢) = <p(t,to)¢(O) + it to
<p(t,a)da[G(a;to,¢, h)], (1.33)
 1.6 Linear time-invariant systems 17

where

G(t; to, ¢, h) = ¢(O) + it jto-s

to -r
de[F(s, O)J¢(s - to + O)ds + it
to
h(a)da.

Note that herein we have extended the domain of definition of F(t, 0) to

-00 < 0 < +00, with the understanding that
F(s,O)=O 0>0
F(s,O) = F(s, -r) 0 < -r.
It is clear that for a linear system, and for any of its solutions y(t), z(t) =
x(t) - y(t) satisfies the equation

i(t) = lOr de [F(t, O)Jz(t + 0). (1.34)

Thus, for a linear RFDE, the stability of every solution is equivalent to the
stability of its trivial solution. For this reason, we will simply say that a
system is stable, instead that any particular solution is stable. It should
also be clear that a linear system is stable if and only if it is globally stable.

1.6 Linear time-invariant systems

If the function F in (1.30) is independent of time t, then the system de-
scribed by (1.30) is linear time-invariant (LTI). An LTI RFDE can be
written as
x(t) = lOr dF(O)x(t + 0) + h(t). (1.35)

In this case, the fundamental solution <l>(t, to) only depends on t - to. As a
result, without loss of generality, we only need to consider to = O. We write
<l>(t) for <l>(t, 0).
Like their counterparts without delay, LTI time-delay systems can be
studied effectively using frequency domain methods. This distinguishing
advantage is made available by such frequency domain analysis tools as
Laplace transforms. The case study given in Section 1.2 has shed light on
how this may proceed, and is now extended below to general LTI systems.
Let
xo = ¢ (1.36)
be the initial condition. Taking the Laplace transform of (1.35) with the
initial condition (1.36), we obtain
18 1. Introduction to Time-Delay Systems

where X(s) and H(s) are the Laplace transform of x(t) and h(t), respec-
tively,

X(s) C[x(t)] = 100

x(t)e-stdt,

H(s) C[x(t)] = 100

h(t)e-stdt.

Solving for X (s) yields

X(s) = ~ -l(S) [4>(0) + lOr eOSdF(B) 10° e- aS 4>(a)da + H(S)] , (1.37)

where
~(s) = sf - lOr eOSdF(B) (1.38)

is called the characteristic matrix. The equation

det[~(s)] =0 (1.39)

is called the characteristic equation, and the expression det[~(s)] the char-
acteristic function, or alternatively, the characteristic quasipolynomial. The
solutions to (1.39) are called the characteristic roots or poles of the system.
Let h(t) = 0 and choose the initial condition (1.32). Then the right hand
side of (1.37) becomes ~ -1 (s). From this and the definition offundamental
solution, it follows that
C[<I>] = ~-l(s),
i.e., ~ -1 (s) is the Laplace transform of the fundamental solution. This
shows that the fundamental solution also satisfies the equation

~(t) = lOr <I>(t + 8)dF(8),

together with the same initial condition (1.32).
The key result that enables frequency domain analysis of stability for
time-delay systems is the following theorem, which generalizes Proposition
1.1. This result is given herein without proof. One may, however, go through
an analysis analogous to that in Section 1.2 for an intuitive understanding.

Theorem 1.5 For any real scalar ,,(, the number of the solutions, count-
ing their multiplicities, to the characteristic equation (1.39) with real parts
greater than "( is finite. Define

aO = max {Re(s)1 det[~(s)l = O}. (1.40)

The following statements are true.

 1.6 Linear time-invariant systems 19

(i) The LTI delay system (1.35) is stable if and only if ao < o.
(ii) FOT any a > ao, there exists an L > 1 such that any solution x(t) of
(1.35) with h(t) = 0 and the initial condition (1.36) is bounded by

(1.41)

(iii) ao is continuous with respect to Tk, fOT all Tk ~ 0, k = 1, 2, ... , K.

The number ao is known as the stability exponent of the system. The

above theorem states that an LTI delay system is stable if and only if its
stability exponent is strictly negative. This is equivalent to saying that all
the poles of the system have negative real parts, a fact that forms the very
basis of frequency domain stability analysis, in the same spirit as for delay-
free LTI systems. Indeed, this fact renders the study of the stability of an
LTI delay system into the study of the zeros of the system's characteristic
quasipolynomial. Moreover, as a fundamental fact, while unlike polynomials
the zeros of a quasi polynomial are not necessarily continuous with respect
to the delays Tk > 0, k = 1, 2,··· , K as Tk --+ 0, the theorem makes it
clear that for an LTI RFDE, the stability exponent ao defines a continuous
function of Tk, for all Tk ~ 0, k = 1,2,··· , K. This continuity property is
reassuring and will playa crucial role in our subsequent stability studies.
One particular class of LTI delay systems under (1.35) is those with
pointwise (or concentrated) delays, which can be further simplified to the
description
K
x(t) = LAkX(t - Tk).
k=O
Here Ak are given n x n real constant matrices, and Tk are given real
constants, ordered such that
o= TO < Tl < ... < r K = r.
For such systems, the characteristic quasipolynomial Do (s) takes the form

m
= Po(s) + LPk(s)e- hkS ,

k=l

where Pk(S), k = 0, 1, ... , m are polynomials, and h k , k = 1, 2, ... , m

are sums of the delay parameters rk. It is evident that this class of systems
can also be described by the scalar differential-difference equation

+ L LPkiy(i)(t -
n-l m
y(n)(t) h k ) = 0,
i=O k=O
20 1. Introduction to Time-Delay Systems

which admits the same quasipolynomial with

n-l
Po(s) sn + LPOiSi,
i=O
n-l

Pk(S) = LPki Si , k = 1,2, m.

i=O

Thus, both the state-space and differential-difference equations can be used

to describe the same system, and the two representations are mutually ex-
changeable. In general, the ratios between the delays, rdrj, may be irra-
tional numbers, in which case the delays are said to be incommensurate.
When all such ratios are rational numbers, on the other hand, we say that
the delays are commensurate. In the latter case, the delays rk (and hence
hi) become· integer multiples of a certain positive T. The characteristic
quasipolynomial A( s) of systems with commensurate delays can then be
written as
q

a(s, e- T) = Lak(s)e-kTS.
k=O
It is worth noting that for systems with incommensurate delays the char-
acteristic quasipolynomial is in effect a multivariate polynomial of several
variables, while for those with commensurate delays, it may be viewed
as a bivariate polynomial, or even further, a polynomial with coefficients
themselves as polynomials. This will prove to be a fundamental distinction.
Indeed, it will be shown in the subsequent chapters that for systems with
commensurate delays, the characteristic quasipolynomial is much easier to
analyze.

1. 7 Neutral time-delay systems

We will only consider linear time-invariant systems with pointwise delays
in the form of
K
L[Akx(t - rk) + Bkx(t - rk)] = 0, (1.42)
k=O
where Ao is nonsingular. We may assume Ao = I without loss of generality.
The initial condition can again be expressed as

xo = ¢. (1.43)

Obviously, the function ¢ needs to be differentiable for the solutions to be

well defined, although a relaxation to discontinuous solutions is possible.
 1.7 Neutral time-delay systems 21

Let the characteristic quasipolynomial b.( s) be defined as

The solutions of the characteristic equation

D.(s) = 0 (1.44)

are referred to as the poles or characteristic roots of the system. We can

state the following theorem.

Theorem 1.6 Consider the system described by (1.42). Define

ao = sup{Re(s) I det(D.(s)) = O}. (1.45)

The following statements are true.

(i) System (1.42) is stable if ao < O.
(ii) For any a > ao, there exists an L > 0 such that any solution x(t) of
(1.42) with the initial condition (1.43) is bounded by
(1.46)

where
me/> = -r::;t::;O
max (11¢(t)11 + 11¢(t)ll)·

We will not give proof here. Interested readers are referred to the litera-
ture discussed in the Notes section.
It is interesting to compare the above theorem for neutral delay systems
to Theorem 1.5 for retarded delay systems. First, while there can be only a
finite number of characteristic roots on the right hand side of any vertical
line Re( s) = 'Y in a retarded delay system, this is not necessarily the case for
a neutral delay system. Second, in the definition of exponential exponent,
"max" in (1.40) is replaced by "sup" for a sufficient condition for stability
of neutral delay systems. In other words, in order to satisfy condition (i)
in the above theorem, all the characteristic roots need to be on the left
hand side of some vertical line Re( s) = -a for some a > O. Having all the
characteristic roots on the left hand side of the imaginary axis is a necessary
but not sufficient condition for the stability of neutral delay systems. There
exist neutral delay systems with all the roots on the left hand side of (but
may have roots arbitrarily close to) the imaginary axis, and the systems are
not asymptotically stable as defined in Definition 1.1. For retarded time-
delay systems, this complication does not occur since there are only a finite
number of characteristic roots on the right hand side of any given vertical
line. Third, for neutral time-delay systems, the stability exponent ao is not
necessarily continuous with respect to delays as they approach zero.
22 1. Introduction to Time-Delay Systems

We will briefly discuss the distribution of the characteristic roots of the

system described by (1.42). For the special case Ak = 0, k = 1,2, ... , K, this
system reduces to a retarded delay system. First, there can only be a finite
number of characteristic roots within any finite region because .6.(s) is an
entire junction. In other words, the set of characteristic roots cannot have
any accumulation points. Second, all the characteristic roots of a sufficiently
large modulus belong to a finite number of root strips

where
J.Ll > J.L2 > ... > J.Lp ;::: O.
A strip with J.Lk > 0 is of retarded type, and a strip with J.Lk = 0 is of
neutml type. Clearly, within a retarded root strip, lim Re(s) = -00. For
8-+00
a neutral root strip, as s - 00, it may approach a vertical line, including
the imaginary axis. For a retarded delay system, all the root strips are of
retarded type. A neutral delay system may (but does not always) possess
a neutral root strip. The proof of this property can be carried out by a
rather sophisticated study of the relationship of .6.( s) and its comparison
function, and will not be discussed here.

1.8 Outline of the text

The focus of this book is the study of stability and robust stability of time-
delay systems. In particular, we concentrate largely on linear delay systems,
or systems that can be analyzed essentially using tools developed for linear
systems. Within this scope, the rest of the book proceeds to develop time
and frequency domain stability and robust stability conditions. The results
can be grossly divided into three categories: Chapters 2, 3, and 4 contain
stability conditions based on frequency domain approaches, Chapters 5, 6,
and 7 present time domain approaches, and Chapter 8 discusses input-
output approaches. Our aim is to report the main advances of last fifteen
to twenty years on the stability and robust stability studies of time-delay
systems, and to develop these results in a systematic as well as coherent
manner. We intend to present a focused view rather than an exhaustive
coverage.
Chapter 2 studies LTI systems with commensurate delays, and focuses
on delay-dependent and delay-independent stability conditions. Here delay-
independent stability is referred to as the property whereby a system is
stable for all nonnegative delay values, while delay-dependent stability is
characterized by the so-called delay margin. The chapter opens with a close
examination of several prevalent classical results known collectively as two-
variable criteria, which serves two useful purposes. First, the study empha-
 1.8 Outline of the text 23

sizes the critical role that the continuity of the stability exponent plays in
stability analysis. Second, it helps reveal the weakness of classical stability
results. These two aspects thus provide the guide for the ensuing devel-
opment, leading to the frequency-sweeping criteria of Section 2.3 and the
constant matrix tests of Section 2.4, both of which are necessary and suffi-
cient conditions for delay-dependent and delay-independent stability alike,
and both require only the computations of matrix pencils. The frequency-
sweeping tests are readily obtained based upon simple algebraic manipula-
tions, and are reminiscent of small gain conditions. On the other hand, the
derivation of the constant matrix tests draws heavily upon the theory of
polynomials, notably the results concerning the Schur-Cohn stability test.
While frequency-sweeping criteria are to be performed on the basis of fre-
quency gridding, the constant matrix tests require only the computation of
constant matrices.
Chapter 3 studies systems with incommensurate delays, which is both
motivated by and built upon a small gain approach. This chapter begins
with a concise introduction of an operator theoretic system description as
well as a statement of the small gain theorem, the rudiments needed in
a small gain approach. It continues with a brief narrative of the struc-
tured singular value, which is a key measure in robust control theory and
in the characterization of stability for LTI systems with incommensurate
delays. Indeed, it is shown that the necessary and sufficient condition for
delay-independent stability requires computing a structured singular value;
similarly, a sufficient condition is also obtained in the same manner for
delay-dependent stability. While these provide useful conceptual results,
the structured singular value is known to be difficult to compute, and in
fact poses an intractable computational problem in general. Consequently,
one is led to the contemplation on the inherent computational complexity
of the stability problem, for which we give a formal analysis drawing upon
concepts and techniques typically found in computing theory and operation
research, with such notions as NP-hard and NP-complete problems. The
analysis is preceded by a brief summary of necessary materials drawn from
complexity theory, and it led to the classification of the stability problem
as NP-hard. In light of this difficulty, we then develop a number of readily
computable sufficient conditions. The chapter ends with an extension to
neutral LTI delay systems, with both commensurate and incommensurate
delays.
Chapter 4 examines the robust stability of LTI delay systems described
by uncertain quasipolynomials with incommensurate delays. By robust sta-
bility we mean that the stability is maintained for the entire family of qua-
sipolynomials as the coefficients vary within a prescribed set. Fundamental
robust stability concepts and results such as the zero exclusion principle,
the edge theorem, and the phase growth condition are extended to and fur-
ther developed for uncertain quasipolynomials. More specifically, the chap-
24 1. Introduction to Time-Delay Systems

ter first presents the zero exclusion principle, which provides a general,
geometrical characterization on the zeros of uncertain quasipolynomials.
This leads to the first main result of the chapter, namely, the edge theorem.
The result is analogous to its counterpart in robust stability theory, and it
states that for a polytopic family of quasipolynomials, robust stability is
achieved if the edge members of the family are all stable. It is worth noting
that in the robust stability literature, the polytopic family remains the most
general uncertain polynomial for which a tractable necessary and sufficient
stability condition is available. As such, the edge theorem for the polytopic
quasipolynomial presents an equally strong result. The chapter continues
to identify some classes of poly topic quasipolynomial families whose sta-
bility is implied by the stability of a finite number of quasipolynomials in
this family. Furthermore, it presents a multivariate polynomial approach
for the interval and diamond multivariate polynomials, which in turn pro-
vide robust stability conditions for the interval and diamond families of
quasipolynomials.
Beginning with Chapter 5, we present time domain stability and robust
stability conditions. This is where the Lyapunov-Krasovskii Theorem and
the Razumikhin Theorem come to play their roles. Chapter 5 treats sys-
tems with a single delay, but without uncertainty. The chapter discusses
various Lyapunov-Krasovskii functionals as well as Razumikhin-type sta-
bility criteria, resulting in various delay-independent and delay-dependent
stability conditions, mostly posed as solutions to LMIs. The relationships
between various results are also explored. It is shown that the use of model
transformation in deriving some of the delay-dependent stability results
will inherently lead to conservatism, because of the resultant additional
dynamics. It is also shown that for necessary and sufficient stability con-
ditions, a complete quadratic Lyapunov-Krasovskii functional needs to be
used. A discretized Lyapunov functional method is used to reduce such
a Lyapunov-Krasovskii functional to a finite-dimensional LMI. Numerical
results show that conservatism is usually very small even for rather coarse
discretization.
Chapter 6 presents extensions to uncertain delay systems. Unlike Chap-
ter 4, the system uncertainty description herein is given on the system's
state matrices, which are either poly topic or norm-bounded. Both these
uncertainty descriptions are widely found in state-space models. Robust
delay-dependent and delay-independent stability conditions that also ad-
mit the form of LMIs are developed in parallel to Chapter 5. In particular,
for polytopic uncertainty, these LMIs are defined at the vertices of the
uncertain family.
Chapter 7 extends the results of Chapters 5 and 6 to systems with multi-
ple delays and distributed delays with piecewise constant coefficients, which
mayor may not have uncertainty. The ideas are very similar to the previous
two chapters but are technically more delicate.
 1.9 Notes 25

Chapter 8 discusses the input-output stability. It starts with a brief dis-

cussion of input-output stability and small gain theorem similar to that in
Chapter 3 but with more emphasis on the state space formulation. Then
it considers the special case of a system with delay as the nominal system
and time delay modeled as uncertainty. The result shows that a number of
stability and robust stability conditions presented in Chapters 5 and 6 ob-
tained by using the Lyapunov-Krasovskii functional method can be derived
and further extended using the small gain formulation. It further discusses
the general setting with time-delay systems as the nominal system with dy-
namical block-diagonal feedback uncertainty. As an application, it is shown
how systems with time-varying delay and general distributed delays can be
approximated as systems with constant pointwise delays or distributed de-
lays with piecewise constant coefficient, with the errors modeled as dynamic
feedback uncertainty.
To make the book adequately self-contained, we have also included two
appendices. Appendix A summarizes some key matrix facts and identities
used throughout the book. Appendix B is a quick reference to basic concepts
and techniques of LMI and quadratic integral inequalities.
At the end of each chapter a section of Notes is provided. In addition
to pointing to the most pertinent references, the section also summarizes
and comments on the results of that chapter. We will mainly limit our
references to those most closely related, than to provide a broad survey.
Nevertheless, the bibliography contains most of the literature needed.

1.9 Notes
1.9.1 A brief historic note
Time-delay systems are also known as hereditary systems, systems with
aftereffects, or systems with time lags. The first functional differential
equations were considered by, among others, such great mathematicians
as Euler, Bernoulli, Lagrange, Laplace, and Poisson in the eighteenth cen-
tury, arising from various geometry problems. In the early twentieth cen-
tury, a number of practical problems were also modelled using FDEs. These
include viscoelasticity problems studied in 1909, the predator-prey model
used in population dynamics in 1928-1931, both by Volterra, mathemat-
ical biology problems studied in 1934 by Kostyzin, and ship stabilization
problems in 1942 by Minorsky.
On the stability of time-delay systems, Pontryagin obtained some fun-
damental results concerning the zeros of quasipolynomials in 1942, and
Chebotarev published a number of papers devoted to the Routh-Hurwitz
criterion for quasipolynomials in early 1940s. A paper of Myshkis in 1949
26 1. Introduction to Time-Delay Systems

formulated for the first time the initial value problem. Those attempting to
extend Lyapunov's theory to time-delay systems initially encountered some
difficulties, which were not resolved until the work of Krasovskii. Krasovskii
was the first to emphasize the importance of adopting Xt rather than x(t)
as the state, and to develop further the idea to fruitation. This took place in
1956. The idea of bypassing functionals was due to Razumikhin, which led
directly to the Razumikhin Theorem. Further extensions along these lines
were made by Kolmanovskii and Nosov to neutral functional differential
equations. For a more comprehensive account of these early developments,
we refer to the book by Kolmanovskii and Nosov [152]. Other treatises of
time-delay systems include the books by Bellman and Cooke [9], G6recki
et al. [78], and Hale and Verduyn Lunel [105], who, like Kolmanovskii and
Nosov, detail the developments of their times and additionally cover other
aspects of time-delay systems. The recent monograph by Niculescu [204]
also documents briefly the historical developments.
Since 1990s there has been a substantial increase of research activities
on time-delay systems in the systems and control community. The main
achievements in this period are characterized by computational improve-
ment and in the study of robust stability. These advances form the core of
the later chapters, where brief reviews and references to the latest progress
will be provided accordingly.

1.9.2 Application examples

The regenerative chatter system in Example 1.1 is discussed in Moon and
Johnson [193], where a nice overview of the topic is given. For some linear
analysis of such a system, see Stepan [252]. Other manufacturing processes
also provide rich examples of time-delay systems. See Moon [192] for an
overview, and Dorf and Kusiak [56] for additional examples.
The modeling of internal combustion engines similar to that in Example
1.2 is discussed by Kao and Moskwa [135]. Other parts of internal combus-
tion engines, e.g., exhaust manifold, also involve time delays. A comprehen-
sive survey of internal combustion engine modeling is provided by Cook and
Powell [41]. Steel rolling mill control is an example of measurement delay,
which is found in Sbarbaro-Hofer [240].
The classical vibration absorber is found in most undergraduate text-
books on vibration. Comprehensive coverage on this subject can be found
in Soong [249] and Tongue [264]. The delayed resonator was proposed by
Olgac and Holm-Hansen in [218]. Analysis of the delayed resonator as well
as its applications are reported in [216, 217, 63, 125]. There are other ex-
amples of delays introduced intentionally in a system, e.g., finite difference
approximation to differentiation, and works on stabilization of chaotic at-
tractors. The former is a very common industrial practice; for example, it
 1.9 Notes 27

is used in the adaptive control of robot manipulators where the angular

acceleration is approximated by the finite difference of measured angular
velocity [44]. The latter can be found in the experimental work by Pyra-
gas [232], for example, which seeks to use delayed feedback to stabilize an
unstable periodic trajectory embedded in a chaotic attractor.
Time-delay systems may also arise from the simplification of partial dif-
ferential equations. Such time-delay systems are usually of neutral type. For
an example of this kind see Brayton [24].
Time-delay systems are encountered in many different disciplines includ-
ing engineering, biology, economics, and ecology. Examples in these dis-
ciplines are documented in such books as Hale and Verduyn Lunel [105],
Kolmanovskii and Myshkis [153], Kolmanovskii and Nosov [152], Niculescu
[204]' Bellman and Cooke [9], and Gorecki et al. [78].

1.9.3 Analysis of time-delay systems

We have chosen to bypass the existence and uniqueness issues concern-
ing the solutions of functional differential equations for two reasons. First,
these issues by themselves merit a comprehensive study and are evidently
beyond the present scope. Second, for the systems under consideration in
this book, which are mainly linear systems, the existence and uniqueness
are essentially ensured and thus rendered moot. Nevertheless, the inter-
ested reader is referred to Hale and Verduyn Lunel [105] for an extensive
treatment of these issues.
We are mainly concerned with asymptotic stability. For this reason other
stability concepts and definitions have not been introduced. Among them,
exponential stability provides a useful measure for decaying rate, or con-
vergence speed. For this we refer to Kharitonov [137] and the references
therein. We note that more elaborate notions, such as uniform bounded-
ness and attraction to a set rather than a point, can also be useful; see,
e.g., [105J. For an LTI system, however, the notions of asymptotic and
exponential stability are equivalent.
The Lyapunov-Krasovskii and Razumikhin Theorems are the corner-
stones in time domain stability anaysis of delay systems. The use of Lya-
punov functionals was first discussed by Krasovskii in his pioneering work
[160]. Razumikhin theory originated in [233, 234]. A systematic presenta-
tion of these techniques is given by Hale and Verduyn Lunel [105], together
with further extensions.
Frequency domain methods are a routine theme in stability analysis,
which enjoy a vast literature. An extensive coverage of quasipolynomials
can be found in Stepan [251], Hale and Verduyn Lunel [105], and Kogan
[148]. The continuity of the stability exponent was discovered by Datko
[51]' who also pointed out that the property does not hold necessarily for
28 1. Introduction to Time-Delay Systems

NFDEs. The distribution of characteristic roots are covered in Bellman and

Cooke [9]. Frequency domain analysis and quasipolynomials will constitute
a primary topic in the subsequent chapters and pertinent references will be
provided therein.
From a broader perspective, RFDE is a subset of infinite-dimensional
systems whose growth is "exponentially determined," i.e., the grow rate is
determined by the maximum real part of the poles, with an exponential
bound. Infinite-dimensional systems are a well-studied branch of math-
ematics and systems science; for a glimpse of this subject, see, e.g., Luo,
Guo, and Morgul [184], Curtain and Pritchard [46], and Curtain and Zwart
[47]. Other areas developed in this context, to name just a few, include con-
trol synthesis and stochastic delay systems; for the former we refer to Foias,
Ozbay, and Tannenbaum [65], and for the latter we refer to Boukas and
Liu [18].
The book is written under the presumption that the reader is familiar
with the theory of ordinary (delay-free) dynamical systems, for which many
excellent books are available; we recommend Bryson and Ho [27], Kailath
[129], Vidyasagar [277], and Zhou, Doyle, and Glover [294].
 Part I

Frequency Domain
Approach
2

Systems with Commensurate

Delays
2.1 Introduction
Stability criteria based on a frequency domain representation are time-
honored tools in the study of dynamical systems. Classical examples of
frequency domain stability criteria include such results as the Nyquist test
and the root-locus method. With the aid of the small gain theorem, fre-
quency domain tests have become increasingly more prevalent in stability
analysis, and have played an especially central role in the theory of robust
control. More generally, while frequency domain methods are used predom-
inantly in the analysis of linear systems, they have also found utilities in the
study of nonlinear systems, with such tools as describing functions, Popov
and circle criterion, and the small gain theorem as well. Various frequency-
sweeping tests are now commonplace. Generally, frequency domain tests are
often favored for their conceptual simplicity and computational ease, which
typically can be checked in an efficient manner by plotting graphically a
certain frequency-dependent measure.
This chapter develops frequency domain stability tests for LTI systems
with pointwise commensurate delays. From Chapter 1 we know that the
stability of an LTI delay system can be completely determined by the
solutions to its characteristic equation. Two stability notions, known as
delay-independent stability (or stability independent of delay) and delay-
dependent stability (or stability dependent of delay), respectively, will be
studied. The chapter opens with a statement of these two notions and pro-
ceeds to examine a number of classical stability tests. Frequency domain
delay-independent and delay-dependent necessary and sufficient stability
conditions will then be developed. Specifically, for systems with commen-
surate delays, a frequency-sweeping test will be provided, which requires
computing the eigenvalues and generalized eigenvalues of certain frequency-
dependent matrices. This result is further developed to yield tests requiring
only the computation of constant matrices. Systems with incommensurate
delays will be discussed in Chapter 3.
Consider the LTI delay systems described by the state-space equation
K
x(t) = Ao x(t) + L Ak x(t - rk), (2.1)
k=l
32 2. Systems with Commensurate Delays

where A o, Ak E IR nxn are given system matrices, and rk are delay times.
As discussed in Chapter 1, the stability of this system is fully determined
by its characteristic function

p (Sj e- r1S , "', e- rKS ) = det (SI - Ao - t

k=l
Ake-rkS) . (2.2)

Specifically, the system is stable if and only if p (Sj e- r1S , " ' , e- rKS ) has
no zero, or root, in the closed right half plane C+. We state this formally
below, as a definition.

Definition 2.1 The characteristic function (2.2) is said to be stable if

(2.3)
It is said to be stable independent of delay if (2.3) holds for all rk :?: 0,
k = 1, 2, "', K. The system (2.1) is said to be stable if its charac-
teristic function (2.2) is stable, and is stable independent of delay if its
characteristic function is stable independent of delay.

Hence, the delay system (2.1) is stable independent of delay if the stability
persists with respect to all possible nonnegative delays. On the other hand,
if it is stable only for a subset of nonnegative delays, then we say that the
stability is delay-dependent.
It is important to recognize that the characteristic function of a time-
delay system defines a real quasipolynomial of s. This feature makes it
possible to extend a number of key properties of polynomials to the study
of time-delay systems. Indeed, a critical ingredient in our subsequent de-
velopment is based on the continuity of the stability exponent with respect
to delays, a fact we have stated in Chapter 1. For a glimpse of its impor-
tance at the outset, let the system (2.1) be stable at the delay values r k,
k = 1, 2, "', K. Then by the continuity property of the stability ex-
ponent, the system will remain stable within a neighborhood of r k. Much
of our effort in determining delay-independent and delay-dependent sta-
bility then amounts to expanding that neighborhood, so that the stability
exponent remains negative. In turn, this amounts to finding the critical
delay values at which the characteristic roots intersect the stability bound-
ary, i.e., the imaginary axis, thus rendering the system unstable. In other
words, when the delay values deviate from r k, we want to determine the
smallest deviation of rk from r k, such that

(2.4)

Evidently, if (2.4) admits no solution, i.e., p (jWj e- jr1w , , e- jrKW ) # 0

for all WEIR, then the system is stable independent of delay. This obser-
vation, albeit a simple continuity argument, underlies the central idea in
 2.1 Introduction 33

our frequency domain stability analysis of LTI delay systems, in both this
chapter and Chapter 3. We shall further elaborate this point in the next
section, using systems with commensurate delays.
For the rest of this chapter, we shall consider LTI delay systems with
commensurate delays. Systems in this class can be represented by
m

±(t) = Ao x(t) + LAk x(t - kT), T;:::: O. (2.5)

k=l
Its characteristic quasipolynomial possesses the form of
q

a (s, e- TS ) = Lak(s)e-kTS, (2.6)

k=O
where
n-l n-l
ao(s) = sn +L aOi si , ak(s) = L aki si , k = 1, q. (2.7)
i=O i=O

We note that a (s, e- TS ) may also be the characteristic quasipolynomial of

the system
n-l q
y(n)(t) +L Lakiy(i)(t - kT) = O. (2.8)
i=O k=O
The distinction between the two representations is thus inessential. It is
convenient to introduce the variable z = e- TS , and write (2.6) as a bivariate
polynomial
q
a(s, z) = Lak(S)z-k, (2.9)
k=O
Notice that the order of ao(s), often known as the "principal term," is higher
than the order of any ak(s), k = 1, 2, "', q. Throughout this chapter,
we assume that the system (2.5) is stable at T = 0, or equivalently, a(s, 1)
is stable. Following the above continuity argument, the smallest deviation
of T from T = 0 such that the system becomes unstable can be determined
as
T := min {T ;: : 0 I a (Jw, e- jTW ) = 0 for some w E lR} . (2.10)
We call T the delay margin of the system. For any T E [0, T), the system
is stable, and whenever T = 00, the system is stable independent of delay.
Note that for any finite T, the frequency at which a (Jw, e- jTW ) = 0 repre-
sents the first contact or crossing of the characteristic roots from the stable
region to the unstable one. Note also that multiple crossings may exist.
From Chapter 1, however, since only finitely many unstable roots may be
in the right half plane, there are only a finite number of zero crossings.
34 2. Systems with Commensurate Delays

Moreover, since a (s, e- TS ) is a real quasipolynomial, all its complex roots

appear in complex conjugate pairs; that is, it satisfies the conjugate sym-
metry property. Consequently, it suffices to consider only the zero crossings
at positive frequencies. Let
a (jWi' e- jOi ) = 0, Wi > 0, Bi E [0, 27r], i = 1, 2, "', N.

Furthermore, define "Ii = Bi/Wi. It is clear that

(2.11)

This gives a general formula for computing the delay margin. More gener-
ally, we may assume, with no loss of generality, that
"11 < 'TJ2 < ... < 'TJN'
It follows that the system is stable for all T E ('TJi' 'TJi+l) whenever it is
stable for some T* E ('TJi' 'TJi+l)' This then allows us to ascertain a system's

°
stability in the full range of delay values, beyond the interval determined
by T. It also indicates that the stability at T = can be made without loss
of generality.
With the bivariate polynomial representation (2.9), it is evident that T
can be found by solving the imaginary roots s E ac+ and the unitary roots
z E allJ) of a(s, z), giving rise to a stability criterion commonly referred
to as a two-variable criterion. The two-variable criterion appears to be
the origin of many classical stability tests for systems with commensurate
delays, which attempt to solve the bivariate polynomial (2.9) in one way or
another. In particular, most of the classical tests attempt to accomplish this
by eliminating one variable, thus converting the stability problem to one
free of delay, and seeking the solution of polynomials of one single variable.
In the next section we discuss a number of tests in this spirit. These sample
tests are representative of classical results and should give the reader the
essential flavor of the two-variable criterion.

2.2 Some classical stability tests

2. 2.1 2-D stability tests

The representation of the characteristic quasipolynomial via a bivariate
polynomial lends the handy recognition that it may be treated as the char-
acteristic polynomial of a 2-D system, and that the stability of a time-delay
system may then be analyzed as in the case of a 2-D polynomial. Indeed,
consider the bilinear transformation
1+).
s=l_).' (2.12)
 2.2 Some classical stability tests 35

which maps 8 from the open right half complex plane C+ to A in the open
unit disk D. Construct the 2-D polynomial

It is evident that a(8, z) = 0 for some (8, z) E ac+ x aD if and only

if b(,X, z) = 0 for some (,X, z) E aD x aD. In addition, for T > 0, the
quasipolynomial a (8, e- rs ) has no root in ( \ if and only if b(,X, z) is
stable; here by stability of a 2-D polynomial, we mean that all its roots
lie outside the closed region liD x liD. Hence, under the assumption that the
system is stable at T = 0, to verify whether the system is stable independent
of delay, it suffices to check whether the 2-D polynomial b(,X, z) is stable.
To determine whether the stability is delay-dependent, it is necessary to
calculate the roots of b(A, z). Stability of 2-D polynomials and 2-D systems
is a well-studied subject in the theory of signal processing. The equivalence
noted herein between the stability of a time-delay system and that of a
2-D polynomial enables us to draw upon the existing analysis techniques
developed extensively for the latter, although 2-D stability tests themselves
generally pose a formidable computational task.
In broader terms, one may view the bivariate polynomial a(8, z) as a 2-D
polynomial and tackle the stability problem directly. Define the conjugate
polynomial
.- z qa ( -8, z -1) .
a-( 8, z )._
By the conjugate symmetry of a(8, z), it follows that (8, z) E ac+ x aD is
a root of a(8, z) if and only if it is also a root of a(8, z). Thus, in order to
find the roots of a(8, z) on ac+ x aD, it suffices to solve the simultaneous
polynomial equations

a(8, z) 0, (2.13)
0,(8, z) O. (2.14)

When no solution exists, and when the system is stable in the delay-free
case, it must also be stable independent of delay. Otherwise, when the
two equations do admit a common solution, it is possible to eliminate one
variable, resulting in a polynomial of one single variable. For example, we
may eliminate 8 and obtain a polynomial in z, say b(z). If b(z) has no
unitary root, we may again conclude that the system is stable independent
of delay. Otherwise, we may proceed to find all the unitary roots Zi of b( z).
There are only a finite number of such roots since b( z) is a polynomial.
For each Zi, a(8, Zi) is a polynomial of the variable 8, which too admits
only a finite number of possible roots 8i E ac+. Thus, we find all the roots
(8i' Zi) E ac+ x aD such that a(8i' Zi) = O. Since the bivariate polynomial
satisfies the conjugate symmetry property, only those 8i on the positive
imaginary axis need to be considered. Thus, consider 8i = jWi, Zi = e-()i,
36 2. Systems with Commensurate Delays

where Wi > 0, and Bi E [0, 211'J. The delay margin can then be determined
using (2.11).

Example 2.1 Consider the scalar delay system

x(t) = -x(t - 7) - x(t - 27), 7;::: O. (2.15)

Evidently, the system is stable at 7 = O. Its characteristic quasipolynomial

is given by

The corresponding bivariate polynomials are

a(s, z)
o'(s, z)
Eliminating s yields the polynomial equation

z4 + z3 + z + 1 = (z + 1)2(z2 - Z + 1) = 0,
which has the solutions Zl = -1, Z2 = (1 + j.;3)/2, and Z3 = (1- j.;3)/2.
Correspondingly, Sl = 0, S2 = -j.;3, and S3 = j.;3. As a result, T is found
from S3 and Z3 = e- j 'lr/3, as T = 11'/(3.;3).0

2.2.2 Pseudo-delay methods

One of the standard techniques in the stability analysis of discrete-time
systems is to transform the stability problem into one of continuous-time
systems. This is accomplished by mapping conformally the unit disc to the
left half of the complex plane. The fact that the stability of a time-delay
system is characterized by the roots of a bivariate polynomial raises the
possibility that the same conformal map may be utilized on the variable z,
which may facilitate the stability analysis. This idea leads us to consider
the bilinear transformation
l-sT
Z=--- T>O, (2.16)
l+sT'
and the polynomial

c(T, s) := (1
1-
+ sT)qa ( s, 1 + sT .
ST)
Since c(O, s) = a(s, 1), the polynomial c(O, s) is stable by assumption.
Furthermore, for any 0 ::; T < 00, the factor (1 + sT)q has no effect on the
imaginary roots of a(s, ~+~~). Thus, for all 0 ::; T ::; 00, the imaginary
roots of a (s, ~+~~) consist of those of c(T, s), 0 < T < 00, and those
 2.2 Some classical stability tests 37

of a(s, - 1); the latter polynomial accounts for the case T = 00. It then
suffices to find the imaginary roots of a( s, -1), and for finite values of T to
calculate the roots of c(T, s), which is a polynomial in s whose coefficients
are parameterized by T. Clearly, if neither of the two polynomials has any
imaginary root, then a(s, ~+:~) has no imaginary zero for all ~ T ~ °
00. As a result, a(s, z) has no root in 8C+ x 8][J), and hence the delay
system is stable independent of delay. However, if for some T = Ti > 0,
the polynomial c(1i, s) has an imaginary root Si = jWi, Wi > 0, then there
exists a unitary Zi = e- jOi E 8][J), ()i E [0, 27rJ, where

Zi = 1- jWiTi or ()i = 2 tan-1 (WiTi),

1 + jWiTi
such that a(si, Zi) = 0. These roots furnish an estimate on the delay
margin, given by

1'1 := min { ~i tan- 1(Wi Ti)! Wi > °}.

Additionally, a(s, -1) may contain imaginary roots as well. At these roots,
2l + 1
T = --7r, l = 0, 1,
Wi
They too yield an estimate on the delay margin, determined as

1'2 := min { :i! Wi > °}.

Consequently, the delay margin is found as

l' = min{1'b 1'2}.

The significance of the above technique, known broadly as the pseudo-

delay method, is quite self-evident. In essence, it reduces the stability prob-
lem effectively to one free of delay, which in turn requires calculating only
roots of a single-variable polynomial. While this is accomplished at the ex-
pense of a parameter-dependent polynomial, it nonetheless allows the use
of various classical stability tests such as the Routh-Hurwitz test and the
root-locus method.

Example 2.2 A second order delay system is described by

T ?: 0. (2.17)

The corresponding characteristic quasipolynomial is

a (s, e- TS ) = s2 + 4s + 4 - 0.25e- TS •
38 2. Systems with Commensurate Delays

The system is stable at T = O. Form the polynomial c(T, s) as

c(T, s) = T s3 + (1 + 4T)S2 + (4 + 4.25T)s + 3.75.

To use the Routh-Hurwitz criterion for c(T, s), we construct the Routh
array
S3 : T 4+ 4.25T
s2 : 1+4T 3.75
sl : 4±16.5T±17T 2
1+4T
so: 3.75
It follows that c(T, s) is stable for all 0 < T < 00. Moreover, since the
polynomial a(s, - 1) = s2 + 4s + 4.25 is also stable, we conclude that
a (s, e- rs ) has no imaginary root for any T ;::: O. That is, the system is
stable independent of delay. 0

Example 2.3 The following system can be easily verified to be stable at

T= 0:

x(t) = (~1 !1) x(t) + (~ ~1) x(t - T), (2.18)

The characteristic quasipolynomial is obtained as

and the polynomial c(T, s) is found as

c(T, s)=Ts3+ S2+(T+2)s+1.
The Routh array is generated as
s3: T T +2
s2: 1 1
Sl: 2
so: 1

Hence for any 0 < T < 00, c(T, s) is stable. Consider next a(s, - 1) =
s2+ 1, which has a pair of imaginary roots s = ±j. This leads to "7 = 11". 0

2.2.3 Direct method

It is further possible to find the zero-crossing frequencies directly based on
the conjugate symmetry property of the quasipolynomial, without using
transformation of any kind. To illustrate, it is instructive to begin with
systems with a single delay. In this situation, the quasipolynomial (2.6)
assumes the simpler form

(2.19)
 2.2 Some classical stability tests 39

By conjugate symmetry, if a (s, e- rs ) = 0 for some s = jw, it is necessary

that
ao(jw) + al(jw)e-j'TW 0,
ao(-jw)+al(-jw)ej'TW o.
Eliminating e}'TW yields

(2.20)

This defines a polynomial equation in w2 • As a consequence, a finite number

of zero-crossing frequencies Wi can be determined by solving the equation.
We may then find the phase angle of al(jwi)/ao(jwi). By equating ()i to

the delay margin can be duly computed according to (2.11). When the
quasipolynomial is stable for T = 0 and (2.20) admits no solution, the
system is stable independent of delay, corresponding to the conditions

. ) I< 1,
Iaoal(jw)
(JW Vw>O, . (2.21)

I>
or

I:~g:~ 1,
Vw>O.

The latter, however, is not possible since ao(s) is a higher order polynomial
than al(s). Thus, we conclude that the necessary and sufficient condition
for stability independent of delay is that (2.21) holds, and that a(s, 1) =
ao(s) + ales) is stable. In this case, ao(s) is also stable, corresponding to
T= 00.

A closely related sufficient condition for the quasipolynomial (2.19) to

be stable independent of delay is that ao(s) is stable and (2.21) holds for
all w ~ O. This condition is of special interest for a number of reasons.
First, it can be held as a special case of the well-known Roche's Theorem,
a result routinely found in the complex variables theory. Indeed, Roche's
Theorem states that for any function f(s) analytic in C+, the functions
ao(s) and ao(s) + f(s) will have the same number of zeros in C+ whenever
If(jw)/ao(jw) I < 1 for all wE R The sufficient condition alluded to above
corresponds to the special case f(s) = al(s)e-'Ts.
Second, this sufficient condition also has a linkage to a classical result
known as Tsypkin's Theorem. Consider a unity feedback system with an
open-loop transfer function
40 2. Systems with Commensurate Delays

Tsypkin's Theorem states that under the condition that G(s), or equiv-
alently ao(s), is stable, then the closed-loop system will be stable for all
T ~ 0 whenever IG(jw)1 < 1 for all w E JR., which coincides with the
above sufficient condition. Finally, we note that the condition provides a
frequency-sweeping test in the form of a small gain theorem, which can
be checked either numerically by inspecting the equation (2.20), or graph-
ically by plotting the frequency-dependent measure lal(jw)/ao(jw)l. The
small gain theorem will play a predominant role in our development of
frequency-sweeping tests and will be discussed in Chapter 3.

Example 2.4 A well-studied example in classical stability analysis of time-

delay systems is the first-order delay system

±(t) = -ax(t) - bx(t - T), (2.22)

where a and b are real constants. The quasipolynomial for this system is

Under the assumption that a+b > 0, the system is stable at T = O. Suppose
that this is the case. Then, the equation (2.20) is given by

w 2 + a 2 - b2 = 0,

which may have a nontrivial solution only when lal < Ibl, yielding the zero-
crossing frequency WI = )b2 - a 2. Clearly, this is possible only when either
a > 0, b > 0, or a < 0, b > 0. In the former case,

(}1 = L. . b -11" = 11" - tan- 1 (WI)

- .
JWl +a a
Correspondingly,

In the latter case,

(}1 . b
= L. JWl +a - 11" = tan -1 (WI)
-I
a
I '
and thus
cos- 1 (¥)
T = -v'---r:b2~_=a=;;2:-

On the other hand, the system is stable independent of delay if and only if
a + b > 0, and a ~ Ibl. Indeed, the condition a ~ Ibl guarantees that

1 (jw) 1= I_b 1<1 Vw>O.

laao(jw) jw + a '
 2.2 Some classical stability tests 41

Note that when a = b > 0, lal(jW)/ao(jw)1 < 1 for all w > 0, but
lal(jw)/ao(jw)1 =
1 at w = 0. In this case, the system remains to be
stable independent of delay.
To further illustrate, consider the system given in Example 2.3, whose
quasipolynomial is in the form of (2.19), with

ao(s) = s2 + s + 1, al(s) = s.
The equation (2.20) is simplified to

w4 - 2w 2 + 1 = 0,
which yields the zero-crossing frequency WI = 1. In addition,

L. a1 (j) = 0.
ao(j)
Hence, 01 = 1r. This leads to the same T = 71', as obtained in Example 2.3.
o
In the case of multiple commensurate delays, the following iterative calcu-
lation can be employed to extend the above test. With the quasipolynomial
(2.6), define the new quasipolynomial
a(l) (s, e- rs ) :=ao(s)a(s, e- rs ) -aq(s)e-qTSa(-s, eTS ).
A simple calculation yields
q-l
L a~I)(s)e-kTS
k=O
q-l
L [ao( -s)ak(s) - aq(s)aq_k( -s)] e- kTS .
k=O
Furthermore,
q-l
a(l) (-s, eTS ) = L [ao(s)ak(-s) - aq(-s)aq_k(s)] ekTS .
k=O
Note that a(1) (s, e- rs ) and a(1) (-s, eTS ) may be considered as a frequency-
dependent linear transform of a (s, e- rs ) and a (-s, eTS ),

ao(s) -aq(s)e- qTS

-aq(-s)e qTS ao( -s)
Thus, whenever

a (jw, e- jTW ) 0,
a (-jw, e jTW ) = 0,
42 2. Systems with Commensurate Delays

it must be true that

a(1) (jw, e- iTW ) 0,

a(1) (-jw, eiTW ) O.

In other words, the imaginary roots of a (s, e- TS ) constitute a subset of

the imaginary roots of a(1) (s, e-TS). Hence, the zero-crossing frequencies
for the former can be obtained by finding the imaginary roots of the latter.
In particular, under the condition that

laq(jw)
ao(jw)
1< 1, Vw E JR.,

the two sets of roots coincide. Nevertheless, the order of a(1) (s, e- TS ) in
e- TS has been reduced to q-1. We may continue this procedure, by defining

That is,
q-i
a(i+l) (s, e- TS ) = :~:::>~i)(s)e-kTS,
k=O

with

ak(s),
a(i)
o
(-s)a(i)
k (s) _ a(i)
q-t.(s)a(i).
q-t-k (s).
For i =q- 1, the procedure generates the function

This effectively reduces the problem to that with one single delay, from
which all zero-crossing frequencies of a(q) (s, e- TS ), and hence those of
a (s, e- TS ), can be computed.

Example 2.5 Let us invoke the iterative procedure to test the stability of
the system (2.15). Note that ao(s) = s, al(s) = a2(s) = 1. Construct

ab1)(s) ao( -s)ao(s) - a2(s)a2( -s) = _s2 - 1,

aP)(s) ao(-s)al(s) - a2(s)al(-s) = -s-1.

Next, solve the equation

 2.2 Some classical stability tests 43

The positive zero-crossing frequencies are found to be WI = V3. We then

calculate
81 = L ap )(jV3) _ 7r = L -jV3 -1 -7r = ~.
a~I)(jV3) 2 3

This gives the same value T = 7r /(3V3). 0

In conclusion, it is critical to find the zero-crossing frequencies of the char-
acteristic quasipolynomial. This indeed is a defining feature of frequency
domain stability tests in general and will be a key step in our study of sys-
tems with commensurate delays in particular. The zero-crossing frequencies
enable us to assert both delay-dependent and delay-independent stability.
Generally, if a system is stable in the delay-free case, then it will be stable
independent of delay whenever its characteristic quasipolynomial exhibits
no zero crossing, while if such crossing does occur, one may compute a
delay margin below which the system remains stable. This thread is com-
mon to and has been well exhibited in the three classical stability tests
alluded to above, which all seek to find the zero-crossing frequencies via an
elimination procedure, reducing in essence a bivariate polynomial to that
of a single variable. Two properties of quasipolynomials have been essen-
tial in these tests. The first is the continuity of the stability exponent with
respect to delays, which is the fundamental cause motivating the computa-
tion of zero-crossing frequencies. This nature of frequency domain tests will
continue to dominate our subsequent development. Second, the conjugate
symmetry property of a real quasi polynomial facilitates the computation
of zero-crossing frequencies. Nevertheless, a careful reader may notice that
the latter property is not necessary, and that a similar reduction procedure
can still be executed for complex quasipolynomials that may not satisfy
the conjugate symmetric property.
Useful notwithstanding, it appears that the classical tests are largely
effective for simple systems of a low order and with a few delays; in fact, the
past success appears to be largely limited to examples in this category, as
we demonstrate in this section. This does not come as a surprise, however.
A little more thought reveals that the limitation is rooted in the elimination
procedure. Indeed, for the three tests examined herein, we come to note
the following aspects.

1. The 2-D tests require that one variable be eliminated and a resultant
polynomial generated from two bivariate polynomials. This can be
difficult for high order systems with many delays, and there exists no
systematic procedure for performing this elimination.

2. The pseudo-delay method requires first a bilinear transformation and

next, the generation and solving of a sequence of polynomials in the
parameter T, that is, the polynomials forming the first column of the
44 2. Systems with Commensurate Delays

Routh array. With the free parameter T, generating the Routh array
can be a cumbersome task.
3. Likewise, the direct method requires a complicated iterative proce-
dure which involves algebraic manipulations of polynomials.

Most notably, the elimination step in each of the three tests requires sym-
bolic computation, or calculation "by hand," which, from a computational
perspective, is highly undesirable. Thus, for high order systems with many
delays, their implementation may encounter a varying degree of difficulty,
rendering the tests inefficient and less applicable.

2.3 Frequency-sweeping tests

Our examination of the classical tests should have sufficiently demonstrated
the need for more efficient, computationally oriented stability criteria. This
section develops frequency-sweeping tests for systems of commensurate de-
lays, including delay-dependent and delay-independent criteria. The next
section develops tests that are based on the computation of constant ma-
trices. Both types of tests emphasize implementational ease and computa-
tional efficiency.
To highlight the main ideas in our derivation, it will be instructive to
begin with systems of a single delay. Thus, consider the system

x(t) = Aox(t) + AIX(t - T), T ~ 0. (2.23)

The quasipolynomial for the system is given by

(2.24)

We first provide a necessary and sufficient condition for stability indepen-

dent of delay.

Theorem 2.1 The system (2.23) is stable independent of delay if and only
if

(i) Ao is stable,
(ii) Ao + Al is stable, and
(iii)
't/w > 0, (2.25)
where p(.) denotes the spectral radius of a matrix.

Proof. For the system to be stable independent of delay, it is necessary

that it be stable for T = 00 and T = 0, which mandates the conditions (i)
 2.3 Frequency-sweeping tests 45

and (ii), respectively. Assume then that Ao is stable, so that (jw1 -AO)-I Al
is well defined for all w E R Suppose that (2.25) holds. This means that
for any eigenvalue Ai ((jw1 - Ao)-IA I ) of the matrix (jw1 - AO)-IA I ,
i = 1, ... , n,

Vw > 0,

which in turn means that

Vw > 0

for any T E JR+. Since

n
det (I - (jw1 - AO)-I Ale- jrw ) = II [1 - Ai ((jw1 - AO)-I Ale- jrw )] ,
i=1

it follows that

Vw>O,

or equivalently,

Vw > O.

Moreover, note that the condition (ii) precludes the possibility of det(Ao +
AI) = O. Therefore, we have shown that for all wE JR,

that is, the quasipolynomial of the system does not intersect the imaginary
axis. Hence, the system is stable independent of delay. This proves the
sufficiency.
To establish the necessity, it suffices to show that (iii) is necessary. To-
ward this end, assume first that p ((jwo1 - AO)-I AI) = 1 for some Wo > O.
This implies that the matrix (jwoI - AO)-I Al has an eigenvalue ejOo , for
some eo E [0, 21l'J. Let TO = eo/woo It is clear that

or
det (jwo1 - Ao - Ale-jrowo) = O.
As such, the system becomes unstable at T = TO, and hence cannot be
stable independent of delay. Next, suppose that p ((jw1 - AO)-I AI) > 1
for some w > O. Since p ((jw1 - AO)-I AI) is a continuous function of w,
and since
46 2. Systems with Commensurate Delays

there exists some Wo E (w, 00), such that p ((jwof - AO)-1 AI) = 1. This,
as we have shown above, implies that the system cannot be stable inde-
pendent of delay. Consequently, we conclude that the conditions (i-iii) are
necessary for the system to be stable independent of delay, thus completing
the proof. _
The frequency-sweeping test given in Theorem 2.1 may be considered an
extended small gain condition, to be discussed later in Chapter 3. Stability
conditions of this type are routinely found in robust control theory and
are generally held as efficient measures in stability analysis; a subsequent
example will attest to this point. Evidently, our derivation of this test still
draws fundamentally upon the continuity property of the stability exponent
of quasipolynomials. Yet unlike the classical results, our test gets rid of the
variable elimination procedure and lends a readily implementable criterion.
The test can be easily checked by computing essentially the frequency-
dependent measure p ((jwf - AO)-1 AI)' which is amenable to computation
because the ease of computing the spectral radius.
While this frequency-sweeping test is more numerically inclined and in-
deed has a justifiable advantage in computation, as a small gain type con-
dition it may also be used for analysis purposes. This will be shown in
forthcoming examples. Moreover, it is of interest to point out that since
(sf - AO)-1 Al defines an analytic function in ~\, the spectral radius
p ( (sf - Ao) -1 AI) enjoys stronger properties than merely the continuity
required in establishing stability. In fact, it will be shown in Chapter 3
that, whenever Ao is stable, p ((sf - A O)-IA 1) is a sub harmonic function
in C+, and so is the function p ((sf - AO)-1 A 1e- rs ). One direct conse-
quence of this fact is that not only is p ((sf - AO)-1 A 1e- rs ) continuous in
C+, for any T :2: 0, but also it satisfies the well-known maximum modulus
principle; that is, it achieves its maximum on the boundary 8C+ of C+. In
other words,

sup p ((sf - AO)-1 A 1e- rs ) sup p ((jwf - AO)-1 A 1e- jrw )

sEC+ w~O

This observation thus provides an alternative explanation to the condition

(2.25). More generally, it has more far-reaching implications and may be
used to advantage for other purposes.

Example 2.6 Let A E Rnxn be a stable matrix and consider the delay
system
x(t) = A x(t) +,BA x(t - T), T :2: 0, ,B E R (2.26)
We want to determine the set of ,B and A such that the system is stable
independent of delay. Note first that the condition (ii) dictates that,B > -1.
 2.3 Frequency-sweeping tests 47

Let A be the Jordan form of A and let Ai be its eigenvalues. Then,

p ((jw1 - A)-1 A) = p ((jw1 - A)-1A) = max

,
vi (ReAi)2+(w-ImAi)2
IAil

Hence, if Ai E R for all i, we have p ((3(jw1 - A)-1 A) < 1 for all w > 0 and
all (3 E (-1, 1]. In other words, the system is stable independent of delay
for all (3 E (-1, 1]. On the other hand, if A has complex eigenvalues, then

sup p ((3(jw1 - A)-1 A) = 1(31 m~x 1 RA\ .1·

w>o ' e A,

This suggests that the system will be stable independent of delay if

It is worth noting that this example has been studied elsewhere using clas-
sical stability tests, but the analysis was, understandably, restricted to 2 x 2
matrices. In general, for a matrix A of an arbitrary dimension, the classical
analysis becomes highly nontrivial. 0

Example 2.7 Consider Example 2.2, in which

Ao = ( -2 0)
0.5 -2 '

It is trivial to verify that Ao and Ao + A1 are both stable. In addition,

(81 - Ao) -1 A1 = (8 +1 2)2 ( 8 + 2 ) ( 0 0.5 ) .

0.5

As a result,

Evidently, p ((jwI - Ao)-1 A 1) < 1 for all w > O. This leads to the same
conclusion that the system (2.17) is stable independent of delay. 0

When the system is not stable independent of delay, it remains possible to

extend the above technique to compute the delay margin. In fact, the proof
of Theorem 2.1 already shed some light on how this may be accomplished.
Under the condition that Ao has no eigenvalue on 8C+, i.e., when (jw1 -Ao)
is invertible, the zero crossings of the characteristic quasipolynomial occur
at the frequencies w where
48 2. Systems with Commensurate Delays

that is, when certain eigenvalues of (jwI - AO)-I Al have a unit modulus.
In general, however, if only delay-dependent stability can be ensured, Ao
may have eigenvalues on ac+. This leads us to the consideration of the
frequency-dependent matrix pencil (jwI - Ao) - AA I . For a matrix pair
(A, B), denote its ith generalized eigenvalue by Ai(A, B). Furthermore,
define
e(A, B) := min {IAII det(A - AB) = O}.
It is a well-known fact that the number of finite generalized eigenvalues for
(A, B) is at most equal to the rank of B. Also, if the rank of B is constant,
then Ai(A, B) is continuous with respect to the elements of A and B.
The idea is then to use the frequency-dependent generalized eigenvalues
Ai(jwI - A o, AI) to replace the eigenvalues Ai ((jwI - AO)-I AI)' which
may not exist at certain frequencies.

Theorem 2.2 Suppose that the system (2.23) is stable at T = O. Let

Furthermore, define

if Ai (jwV - A o, AI) = e-je~

for some w~ E (0, 00), e~ E [0, 27r]

if e(jwI - A o, AI) > 1, Vw E (0, 00).

Then,

That is, the system (2.23) is stable for all T E [0, T), but becomes unstable
at T = T.

Proof. Consider first the case T = 00. This corresponds to the condition

for all T ~ 0 and all w > O. Hence, it implies that for any T ~ 0,

Vw E (0, 00).

In addition, under the assumption that the system is stable at T 0,

det(Ao + AI) =I- O. This suggests that for any other T,

det (jwI - Ao - AIe- jTW ) =I- 0

at w = 0 as well. It thus follows by the continuity of the stability exponent

that
 2.3 Frequency-sweeping tests 49

Ai:, a result, the system is stable for all T E [0, 00); in other words, the
system is stable independent of delay. Suppose now that T < 00. For any
T E [0, T), we claim that

Vw E [0, 00).
This is clearly the case for w =/:: wL for at these frequencies

for all i = 1, "', n. Additionally, at w = w~, since for any T E [0, T),
TW~ =/:: (h by definition, we have det (jwV - Ao - Ale-j7'W~) =/:: 0. This
proves the claim. Again, by an appeal to the continuity argument, we con-
clude that the system is stable for all T E [0, T). On the other hand, if
T = T, then there exists a pair (w~, O~) such that T = 01/wL and

det (jw~I - Ao - Ale-i'i'wt) = det (jw~I - Ao - Ale-j9~) = 0.

That is, the system is unstable. The proof is now completed. •

It follows instantly that the delay-independent stability condition, Theo-
e
rem 2.1, can be equivalently stated in terms ofthe measure (jwI - A o, At).

Corollary 2.3 The system (2.23) is stable independent of delay if and only
if
(i) Ao is stable,
(ii) Ao + Al is stable, and
(iii)
Vw>o. (2.27)

In light of Theorem 2.2, an algorithm can be devised at once to com-

pute the delay margin. First we compute Ai (jwI - A o, Al)' This is to
be done by gridding the frequency axis, and at each gridding point, com-
puting the generalized eigenvalues using standard numerical algorithms.
If p(jwI-Ao, A l ) > 1 for all wE (0,00), we conclude that the sys-
tem is stable for all T E [0, 00). Otherwise, the computation generates a
pair (wi, O~), which always yields a nonzero estimate of the delay mar-

°
gin. This can be seen by noting that whenever det (jwV - Ao - Al) =/:: 0,
the inequality < 01 < 21l' holds, and hence Ti > 0. One should note,
unsurprisingly, that the computation of the delay margin is more demand-
ing than the determination of the delay-independent stability. Note also
that the computation needs to be performed only over a finite frequency
interval, since the moduli of the generalized eigenvalues will inevitably ex-
ceed one beyond a certain frequency. This, together with the fact that the
generalized eigenvalues are continuous functions of w, is reassuring from a
numerical standpoint.
50 2. Systems with Commensurate Delays

We are now ready to tackle systems of multiple commensurate delays.

Specifically, consider
m

x(t) = Ao x(t) + LAk x(t - kT), 7 ~ O. (2.28)

k=1

In light of Theorems 2.1 and 2.2, it is natural to seek to transform the

stability problem to one with a single delay. This is readily accomplished
by the use of the Schur complement. We call the reader's attention to the
following matrix identity, which is a straightforward consequence of the
Schur's complement for determinant, given in Appendix A.

Lemma 2.4 For any z E C and matrices Pk E cnxn , k = 0, 1, m,

det (~PkZk)

C pJ-( )}
0 1 0

dcl{Z 1 0 0 1
-Po -PI -Pm - I

Theorem 2.5 The system (2.28) is stable independent of delay if and only
if

(i) Ao is stable,
m
(ii) L:: Ak is stable, and
k=O
(iii)
p(Mm(jw)) < 1, 'v'w > 0, (2.29)
where

(s1 - ~O)-I Al

Mm(s):= ( .

o 1

Theorem 2.6 Suppose that the system (2.28) is stable at 7 = 0, and let
q = rank(Am). Furthermore, define

(Ji if Ai (G(jwU, H(jw~)) = e-j(Ji

min
-7' ._ {
t .-
.::.1>.
I::;k::;n w k for some w~ E (0, (0), oi
E [0, 27l']

00 if e.(G(jw), H(jw)) > 1, 'v'w E (0, (0)

 2.3 Frequency-sweeping tests 51

where
0 1 0

G(s),~ (
0
-(s1 - Ao)
0
Al
1
Am- I
).
H(s) := diag ( 1, ... , 1, -Am ) .
Then,
T:= min ri.
1:::;i:5q+n(m-l)
The system (2.28) is stable for all T E [0, r), but becomes unstable at T = r.

Proof. Let z = e-iTW , Po = jw1 - Ao, and Pk = -Ak. It follows from

Lemma 2.4 that

det(jW1 - "fAke-ikTW) = (-ltmdet(G(jw) - e-iTWH(jw)).

k=O
This effectively reduces the stability problem to one with a single delay.
Thus the proof for Theorem 2.6 follows as in that for Theorem 2.2. The
proof for Theorem 2.5 may proceed by recognizing that the condition for
stability independent of delay is given by Theorem 2.5 (i-ii), and addition-
ally,
e(G(jw), H(jw)) > 1, 'v'wE(O,oo).
Under the condition that G(jw} is invertible,
1
e(G(jw), H(jw)) = p(G-I(jw)H(jw))
In view of the matrix identity

r
0 1 0

( 0
-Po
0
-PI
1
-Pm- 1

-r}
_PO-l Pm-l

~CTP' 0

1
(2.30)

we conclude that Mm(jw) = G-I(jw)H(jw). The prooffor Theorem 2.5 is

thus completed. _
The following examples serve to illustrate the computational advantage
of these results, demonstrating that the frequency-sweeping tests obtained
herein can be fairly effective for systems with high dimensions and many
delays. Such systems are not easily handled using classical stability tests.
52 2. Systems with Commensurate Delays

Example 2.8 Consider the system (2.28), with four states and three com-
mensurate delays, given by

0, ~ ).
1 0
0 1
Ao = 0 0
-3 -5 -2
( -0.00 0.005 0.25
Al = 0.~05 0.005
0
0
0
~).
(T
-1 0 -0.5
0.0025 0
A2 =
-1
0
0
0.05

-0.5 -0.5
0 O~ ).

0 0.075
( Oog75 0.05 0.05 0~25 )
A3 0.05 0.05 0 o .
0 -2.5 0 -1

A simple computation shows that the eigenvalues of Ao are located at

-0.6887 ± 1. 7636j
-0.3113 ± 0.6790j.

Hence, Ao is stable. Similarly, ~~=o Ak is also found stable. As such, we

need to examine the condition (2.29) to check whether the system is stable
independent of delay. The spectral radius p(MmUw)) is shown in Figure
2.1. From this plot, it is clear that the condition (2.29) is not satisfied.
Therefore, we conclude that the system is not stable independent of delay.
o
Example 2.9 Consider the same system in Example 2.8, with A k , k =

n,
1, 2, 3 modified to

(I ~} A2~P
0 0 0 0
0 0 0 0
Al 0 0 0 0

q
0 -1 -2 -1 -1
0 0

U
0 0
A3 =
0 0
-5 0 -2

It can be easily verified that the system is stable at T = O. Let us com-

pute T based on Theorem 2.6. First, perform an initial computation of
 2.3 Frequency-sweeping tests 53

1.4

1.3

1.2

1.1

~ 1
;3
'2
Ii'
::00.9

0.8

0.7

0.6

0.5
0 0.5 1.5 2 2.5 3 3.5 4 4.5 5
Frequency

FIGURE 2.1. The spectral radius p(Mm(jw))

Ai(G(jW), H(jw)) on a crude frequency grid, to determine an approximate

interval in which IAi(G(jW), H(jw))1 cross the value of 1. This interval
was found confined within the interval [0, 2]. The computation is then per-
formed using a much finer grid over [0, 2] with 1000 grid points for the
purpose of achieving a higher precision. This yields the critical frequency-
angle pairs (wi, et):

(0.2743, 3.5952), (0.9149, 2.9014), (0.9169, 2.9001), (1.3033, 0.4931),

from which we obtain T = 0.4931/1.3033 = 0.3783. 0

We end this section with a comment on systems described by the scalar

differential-difference equation

LL
n-l q
y(n)(t) + akiy(i)(t - kr) = 0, r 2: O. (2.31)
i=O k=O

The characteristic quasipolynomial for such a system is given by (2.6) and

(2.7). Clearly, both Theorems 2.5 and 2.6 can be immediately applied to
54 2. Systems with Commensurate Delays

(2.31) by defining

1 o
Ao
o 1
-aO,n-l
).
)
o o
o o
-ak,n-l

The tests, however, can be further simplified computationally. Indeed, we

recognize with ease the following conditions for the stability of such a sys-
tem.

Corollary 2.7 The system (2.31) is stable independent of delay if and only
if

(i) ao(s) is stable,

q
(ii) I: ak (s) is stable, and
k=O
(iii)
p(Ma(jw)) < 1, 'Vw > 0, (2.32)
where

)
-~ ... _aq_l{s) _ aq{s)

M.(s) ,~ ( ao{s)
~ ...
ao{s)
0
ao{s)
0

1 0

Corollary 2.8 Suppose that the system (2.31) is stable at T = O. Define

if Ai (G a(jw1), Ha(jw1)) = e-jll~

for some w~ E (0, 00), B~ E [0, 27r]
if e(Ga(jw), Ha(jw)) > 1, 'Vw E (0, 00)

where
o 1

G.(s) ,~ (

Ha(s) := diag (1, 1,

 2.4 Constant matrix tests 55

Then,

The system (2.31) is stable for all T E [0, r), but becomes unstable at T = r.

Note that the dimensions of these matrices are equal to q, which may be
considerably lower than that of the matrices in Theorems 2.5 and 2.6. As
such, it is possible to check both conditions more efficiently.

2.4 Constant matrix tests

In spite of their implementational simplicity, frequency-sweeping tests by
nature may not be executed in finite computation, and the computational
accuracy hinges on the fineness of the frequency grids. Thus they are likely
to be at a disadvantage if high computational precision is sought. This
issue can be more acute in the computation of the delay margin. This
consideration thus leads us to search for alternative stability tests, tests
that are both readily implementable and can be performed via finite step
algorithms. In particular, for numerical precision it will be highly desirable
to eliminate any frequency sweep, while retaining the merits of computing
eigenvalues and generalized eigenvalues, for the computational ease of the
latter. The stability tests to be developed below seek to combine these
advantageous features, which will require only the computation of constant
matrices.
We begin with a reexamination of the quasipolynomial (2.6), while leav-
ing aside its interpretation either as one for the state-space description
(2.5), or that for the differential-difference equation model (2.31). We have
seen that the stability of the quasipolynomial, either independent or de-
pendent of delay, can be assessed by finding its zero-crossing frequencies.
More specifically, it amounts to finding the roots (s, z) E 8C+ x 8ID>, so
that the bivariate polynomial a(s, z) = 0, or equivalently, all such w > 0
that a(jw, z) =0 for some z E 8ID>. A key, albeit straightforward, observa-
tion is that for each wE (0, 00), a(jw, z) defines a polynomial in z, whose
root locations can be analyzed using established tools. Our development is
built upon this recognition. With the unit disk ID> being the stability region
of concern, we come to note the well-known Schur-Cohn criterion, which
gives a necessary and sufficient condition for the roots of a polynomial to
lie in ID>. For ease of reference, we make a slight digression to this subject,
but choose to leave the proofs to the relevant literature cited in the Notes
section at the end of this chapter.
56 2. Systems with Commensurate Delays

Consider a complex polynomial

q
p(z) = I>kZk.
k=O

Define its conjugate polynomial p(z) by

q

p(z) := zq I)kZ-k.
k=O

Furthermore, define the Schur-Cohn-Fujiwara matrix as

W := pH (S)p(S) - pH (S)p(S),

where p(S) is the matrix polynomial

and S is the shift matrix

The Schur-Cohn-Fujiwara matrix W = [Wij] is Hermitian and consists of

elements
i-I

Wij = I: (Pn-i-r-IPn-j-r-l - Pi+r+IPHr+l) , i ::; j.

r=O

It is known that the polynomial p( z) has all of its roots in lID if and only
if W is positive definite. In particular, the roots of p(z), Zi, i = 1, ... , q,
are related to W via the following Orlando formula.

Lemma 2.9 (Orlando Formula) Let Zi, i = 1, ,q, be the roots

of the complex polynomial p(z) = L:k=OPkZk, whose corresponding Schur-
Cohn-Fujiwara matrix is defined by W. Then
q
det (W) = Ipi q II (1 - ZiZj) .
i,j=1

We next define the matrix

 2.4 Constant matrix tests 57

Note that

(I P:,
pq Pq-I ... PI)
:).
o Pq ... P2
(
··· ..
. ' .
o 0 pq

Note also that C is the test matrix in the Schur-Cohn criterion. The fol-
lowing fact establishes a relationship between C and W.

Lemma 2.10
det(W) = (-I)q det(C).

Proof. Suppose first that Po =/: O. Then by definition p(S) is invertible.

Since the polynomial matrices p(S) and p(S) commute, we have

Based on this fact and the Schur determinant formula, we find that

det(W) det(pH (S)p(S) - pH (S)p(S))

= det (pH (S)p(S)p-l (S) - pH (S)) det(p(S))
= (-I)q det(pH (S) - pH (S)p-l(S)p(S)) det(p(S))
q (P(S) P(S))
= (-1) det pH (S) pH (S)

= ( )q (P(S) p(S))T
-1 det pH(S) pH(S)

= (-1)q det (pT(S) (pT(S))H)

pT(S) (pT(S))H .

Hence, det(W) = (-I)q det(C). Additionally, in case Po = 0, the equality

remains valid; this is seen by an appeal to the limit Po ---+ O. •
It follows from Lemma 2.9 that p(z) has a root on the unit circle ali)
if and only if det(W) = 0, and from Lemma 2.10 that p(z) has a uni-
tary root if and only if det( C) = O. The latter sheds light on how zero-
crossing frequencies may be found. Essentially, for each w E 1R+, one may
viewa(jw, z) as a complex polynomial in z, and construct accordingly its
Schur-Cohn-Fujiwara matrix or Schur-Cohn test matrix, both of which are
frequency dependent. It then follows that at that w, a(jw, z) has a root
58 2. Systems with Commensurate Delays

z E aD whenever the determinant of these matrices is equal to zero. This

consequently ensures the existence of a pair (jw, z) E ac+ x aD such that
a(jw, z) = 0. In other words, if w is a zero-crossing frequency, then it can be
found from the condition that the determinant of the frequency-dependent
Schur-Cohn-Fujiwara matrix or Schur-Cohn test matrix is equal to zero.
More specifically, for any fixed s consider the polynomial
q
a(z) := a(s, z) = I>k(S)zk.
k=O

Define its conjugate polynomial as

q
a(z) := zq L)Zk(S)z-k.
k=O

Furthermore, define the matrices

)
0 0
( a,(,)
aleS) ao(s) 0
~l(S) :=

aq_~(s) aq-2(S) ao(s)

).
aq_l(s) aleS)

r
( a,(')
aq(s) a2(s)
E,(s) ,~ (2.33)

0 aq(s)
and
~(s) := (
~l(S)
~!j (s)
~2(S)
~{l (s) ).
In view of Lemmas 2.9 and 2.10, we assert that that for any s,
n
det(~(s)) = (-1)q laq(s)1 2q II (1 - ZiZj) , (2.34)
i,j=l

where Zi, i = 1, "', q, are the roots of a(z) for fixed s. Therefore, for
any w > 0, there is a z E aD such that a(jw, z) = a(z) = whenever °
°
det (~(jw)) = 0. The implication is then fairly clear: By solving the equa-
tion det(~(jw)) = 0, we find all such w > that a(z) = a(jw, z) has a
root on aD. Since det(~(jw)) defines a polynomial in w, the solutions can
be found by solving the eigenvalues of a constant matrix. Clearly, there are
only a finite number of solutions, and among them the positive solutions
constitute the zero-crossing frequencies. Once the zero-crossing frequencies
are found, the unitary roots of a(jw, z) can also be determined at each zero-
crossing frequency by solving the complex polynomial a(z) = a(jw, z). The
 2.4 Constant matrix tests 59

delay margin T can then be computed immediately from the solutions of

these two problems. The following result implements this solution proce-
dure.

Theorem 2.11 Suppose that the quasipolynomial (2.6) is stable at T = O.

Let Hn := 0, Tn := I, and

aq-l,i
aqi
i = 0,1, ... , n - 1,

~ ),
ao, o
ah aOi
(
i = 0, 1, n-l,

aq~l'i a q -2,i aOi

(j)iTi (j)i Hi )
(_j)i HT (-j)iTT '
i = 0,1, n.

Furthermore, define

)
o I o
p~ (

Then, T = 00 if o-(P) n IR+ = 0 or o-(P) n IR+ = {O}. Additionally, let

).
1 0

F(s) ,~( r 0 1
-ao(s) -al(s) -aq_l(s)

G(s) := diag( 1, 1, ... , aq(s) ).

Then T = 00 if o-(F(jWk), G(jWk)) n all]) = 0 for all 0 -=I Wk E o-(P) n IR+.
In these cases the quasipolynomial (2.6) is stable independent of delay.
Otherwise,
(h
T= min -,
1:::;k:::;2nq Wk

where 0 -=I Wk E o-(P) n IR+ and (h E [0, 27rJ satisfy the relation e- jOk E
o-(F(jWk)' G(jWk))' The quasipolynomial (2.6) is stable for all T E [0, T),
but is unstable at T = T.
60 2. Systems with Commensurate Delays

Proof. Suppose first that a(P) n R+ = 0. This implies that for any
wE R+, det(wI - P) =f:. 0. In view of Lemma 2.4,

det(wI - P) =

L;2(jW) )
L;¥(jw) .

Hence, the condition that a(P) n R+ = 0 is equivalent to det(L;(jw)) =f:. °

°
for all w E R+. This implies that a(jw, z) has no root on 8D, that is,
a (jw, e-irw ) =f:. for all wE R+ and T E [0, (0). Since a (s, e- TS ) is stable

°
then a (jw, e-irw ) =f:. °
at T = 0, it is stable for all T E [0, (0). Note also that if a(P) nR+ = {O},
for all w E R+ and T E [0, (0), despite that
a(O, z) = may hold for some z on 8D. This follows from the fact that
a(jw, z) has no root on 8D for any w =f:. 0, and that a (jw, e-irw ) =f:. °
°
at w = 0, by the assumption that a (s, e- rs ) is stable at T = 0; the latter
implies that a(jw, 1) =f:. for all w E R+. Hence, again, we conclude that
a(s, e- TS ) is stable for all T E [0, (0) if a(P) nR+ = {O}.

°
Suppose now that a(P) nR+ =f:. 0 and a(P) nR+ =f:. {O}. In other words,
P does have positive real roots. Let =f:. Wk E a(P) n R+. Since according
to Lemma 2.4,
q
det(zG(s) - F(s)) = Lai(S)zi = a(s, z),
i=O

it follows that a(jwk, z) has no root on 8D if a(F(jwk), G(jWk)) n8D = 0.

°°
As a result, we may conclude that T = 00 if a(F(jwk), G(jWk)) n 8D = 0
for all =f:. Wk E a(P) n R+. If, however, a(F(jwk), G(jWk)) n 8D =f:. 0
for some =f:. Wk E a(P) n R+, then there exist some Zk E 8D such that
Zk E a(F(jwk), G(jWk))' Consequently, there exist some (h E [0, 21r]

°
such that Zk= e- ifh and a (jWk' e- ifh ) = 0, or equivalently, e- j1h E
a(F(jwk), G(jwk))' It follows that a (jWk' e- jrwk ) = for T = (h/Wk,
and hence the system is unstable at T = T. However, for any T E [0, T),

°
since TWk =f:. Ok for all (Wk' Ok) such that e- iIJk E a(F(jwk), G(jWk)),
we have a (jWk' e- jrwk ) =f:. 0. This implies that a (jw, e- jrw ) =f:. for all
wE R+. Therefore, a (s, e- rs ) is stable for all T E [0, T). The proof is now
completed . •
Theorem 2.11 suggests that a two-step procedure can be employed to test
the stability of the quasipolynomial (2.6). First compute the eigenvalues of
the 2nq x 2nq matrix P. If P has no real eigenvalue or only one real eigen-
value at zero, we conclude that the quasipolynomial is stable independent
of delay. If this is not the case, compute next the generalized eigenvalues of
the q x q matrix pair (F(jwk), G(jWk)), with respect to each positive real
eigenvalue Wk of P. If for all such eigenvalues the pair (F(jwk), G(jwk)) has
 2.4 Constant matrix tests 61

no generalized eigenvalue on the unit circle, we again conclude that the qua-
sipolynomial is stable independent of delay. Otherwise, we obtain the delay
margin T. Note that if aq(jw) -# 0 for all wE lR+, the pair (F(jw), G(jw))
will always have generalized eigenvalues on oJ[]) whenever (T(P) n lR+ -# 0.
Indeed, this implies that det(E(jwk)) = 0 for some Wk E (T(p)nlR+. Accord-
ing to (2.34), we may conclude that a(jwk' z) will always have roots on or
symmetric about oJ[]). Since the roots of a(jw, z) are continuous functions
of w, there must exist also some WI such that a(jwI, z) has roots on oJ[]).
Thus, under this circumstance, we may assert without further computing
the generalized eigenvalues of F(jw) and G(jw), that the quasipolynomial
cannot be stable independent of delay.
The following result gives an alternative test in the same spirit. It may
be executed by computing sequentially the eigenvalues of two matrices. A
minor restriction, however, is that the assumption ao (jw) -# 0 needs to be
imposed for all wE lR+.

Theorem 2.12 Suppose that ao(jw) -# 0 for all wE lR+. Suppose also that
the quasipolynomial (2.6) is stable at T = O. Let (A, B, C) be a minimal
realization ofEl1(s)En-s), where El(S) and E2(S) are given by (2. 33}.
Furthermore, define the Hamiltonian matrix

._ (
H .- A BBT) lD>2nqx2nq
-CTC _AT E~ .

Then, T = 00 if(T(H)noC+ = 0 or (T(H)noC+ = {O}. Additionally, define

_aq_l(s)

Ma(s),~ (
-~)
-~
ao(s) ao(s) ao(s)
o o
~
1 o
Then, if (T (Ma(jWk)) n oJ[]) = 0 for all 0 -# Wk E lR+ such that jWk E (T(H),
T = 00. In these cases the quasipolynomial (2.6) is stable independent of
delay. Otherwise,

where jWk E (T(H), Wk E lR+, Wk -# 0, and Ok E [0, 27r] satisfy the relation
e- j1h E (T(Ma(jWk)).

Proof. First, since ao(jw) -# 0 for all wE lR+, it follows that

det(jwI - A) ao(jw) -# 0,
det(jwI + AT) (-ltao(-jw) -# 0,
62 2. Systems with Commensurate Delays

and ~ -1 (jw) is well defined, for all w E R Consider the determinant

det(jwI - H). A straightforward calculation shows that

det(jwI - H)
jWI-A -BBT)
det ( eTc jw + AT
det(jwI - A)det (I - BT(_jwI - A)-TeTe(jwI - A)-IB)
det(jwI + AT)
det(jwI - A)det (I - E2(jw)EIH(jw)Ell(jw)Ef(jw))
det(jwI + AT)
= (_1)n lao(jw)1 2 det (I - E2(jw)EIH(jw)Ell(jw)Ef(jw))
= (-It lao(jw)12 det (EIH (jw)) det (E{f - Ef (jw)Ell(jw)E2(jw))
1-
( _l)nl ao (.JW )1 2 ao 1 (.JW )1 2 dt(
e
El(jW)
Ef (jw)
E2(jW))
E{f (jw)
= (-It det(E(jw)).

In this derivation, we have used repeatedly the Schur determinant for-

mula and relied on the facts that El(S) and E2(S) commute and that
det (El(S)) = ao(s). Hence, det(jwI -H) = 0 if and only if det(E(jw)) = O.
Let F(s) and G(s) be defined as in Theorem 2.11. Since ao(jw) i= 0, F(jw)
is invertible. In light of the fact that Ma(s) = F- 1 (s)G(s), the result follows
at once. •
Similarly, Theorem 2.12 indicates that the stability of the quasipolyno-
mial (2.6) can be checked by first computing the eigenvalues of the 2qn x 2qn
Hamiltonian matrix H, and next the eigenvalues of the constant q x q ma-
trix Ma(jWk), for each Wk E R+ such that jWk is a pure imaginary eigen-
value of H. It is clear that the imaginary eigenvalues of H correspond to
exactly the real eigenvalues of P. It is also useful to note that the com-
putation of the imaginary eigenvalues of a Hamiltonian matrix is closely
related to that of computing the HOC! norm of a certain transfer function
matrix. In the present setting, the transfer function matrix in question is
~ll(s)~f( -s). This connection has additional implications, which will be
revisited in Chapter 3.
Thus a notable feature with the stability tests in both Theorems 2.11
and 2.12 is that they require only the computation of constant matrices. As
a result, while retaining the computational ease of the frequency-sweeping
tests developed in the preceding section, both tests enjoy superior numerical
precision, though the benefit is gained at the expense of increased matrix
dimensions. The following example serves to demonstrate the numerical
advantage.
 2.4 Constant matrix tests 63

Example 2.10 The following quasipolynomial, with n = 4 and q = 3,

corresponds to the system in Example 2.9, which is known to be stable at
T = 0, but not stable independent of delay:

(8 4 + 28 3 + 58 2 + 38 + 2) + (8 2 + 2) e- rs
+ (8 2 + 8 + 2) e- 2rs + (28 3 + 58) e- 3rs .
We may compute T based on Theorems 2.11 and 2.12. Construct first the
matrices P and H, both of which are 24 x 24 matrices. Using Theorem
2.11, we found the positive real eigenvalues of P to be
0.27440341977960 - O.OOOOOOOOOOOOOOj,
0.91548178048665 - O.OOOOOOOOOOOOOOj,
1.30259246725682 + O.OOOOOOOOOOOOOOj,
and the corresponding unitary roots are

-0.89898090347872 + 0.43798782537941j,
-0.97120708373425 - 0.23823685798887j,
0.88081464844625 - 0.47346124982147j,
respectively. The reader may want to verify that the modulus of all the roots
is 1.00000000000000. From these results, we found T = 0.37864202327266.
The computation results based on Theorem 2.12 turn out to be identical
up to the indicated precision (16 digits), and thus the two sets of results
demonstrate for one another the numerical accuracy of the tests. 0

The final part of this section is devoted to systems described by the

state-space representation (2.5). At the outset, we note that the stability
of such a system can always be analyzed by studying its characteristic qua-
sipolynomial, whereas the latter can be computed routinely using, e.g., the
Leverrier-Faddeeva algorithm. The constant matrix tests developed above
are then applicable. Often, however, it may be numerically cumbersome
to compute the characteristic quasipolynomial, and hence it may be more
advantageous to tackle the stability problem based on the state-space de-
scription directly. In what follows we develop one such test, in which the
computation of the characteristic quasipolynomial is rendered unnecessary.
We recall that for the delay system of commensurate delays given by the
state-space form (2.5), its stability is fully determined by the root locations
of the bivariate polynomial

a(8, z) = det(81 - fAkZk) .

k=O

Assume that the system is stable at T = O. Then by fixing z E allJJ, the sta-
bility problem amounts to determining whether the matrix (E:;;'=o AkZk)
64 2. Systems with Commensurate Delays

has eigenvalues on 8«:+. Indeed, if it has no imaginary eigenvalue for all

z E 81ID, the system is stable independent of delay. Suppose otherwise that
it does have imaginary eigenvalues for some Zi E 8 lID, Zi = e- j9i . That is,
there exist some jWi E OC+ such that

In this case we may compute the delay margin once finding the pairs
(Wi, ()i)' Consequently, we are led to the problem of finding all such criti-
cal Z E 81ID that u (2:;;'=0 AkZk) has an imaginary eigenvalue, and that of
solving the imaginary eigenvalues for all such Z E 81ID. The Kronecker sum
of matrices discussed in Appendix A comes to our aid in this task.
Recall that for square matrices A and B,

u(A ED B) = P + IL: oX E u(A), IL E u(B)}.

Hence, whenever u (2:;;'=0 AkZk) n 8C+ =I 0, it is necessary that

(2.35)

This underlies the main idea in our derivation.

Theorem 2.13 Suppose that the system (2.5) is stable at T = 0. Define

the matrices Bk E ]Rn2, k = 0, 1, "', 2m by Bm := Ao ED Air, and

Bm - k := I ® Af, k= 1, "', m.
Furthermore, define

)
o I o

o I
-Bo -B2m-l

Then, T = 00 if u(V, U) n 81ID = 0. If, however, u(V, U) n 81ID =I 0 and

u (2:;;'=0 AkZf) = {O} for all Zi E u(V, U) n 81ID, then T = 00 as well. In
these cases the system (2.5) is stable independent of delay. Otherwise,
()i
T:= min -,
1:5i::52n 2 m Wi

where ()i E [0, 271"J, Wi E ]R+, Wi =I 0, and e- j9 , E u(V, U) satisfy the

relation jWi E u (2:;;'=0 Ake-jk9i). The system (2.5) is stable for all T E
[0, T), but is unstable at T = T.
 2.4 Constant matrix tests 65

Proof. In view of Lemma 2.4, it is straightfoward to verify that

det(zU - V) = det(~Bkzk)
= zn 2 m det (f
k=l
Bm+kZk + Bm + f
k=l
Bm_k z - k ) .

Thus, for any z E 8lIJJ,

det(zU - V)

zn'm det (~(Ak ® I)zk + An <Il Air + ~(I ® Anz-' )

zn'm det (t,(A'Zk ® J) + t,(J ® Ai Z-k))

zn'm det [ (t, A,Z') (t, A,z') H].

<Il

Suppose that o-(V, U) n 8lIJJ = 0. Then, for any z E 8lIJJ,

This implies that for any z E 8lIJJ and any W E lR.+, det(jwI - L:;;'=o AkZk) =1=
O. Thus, for any T E [0, (0) and any w E lR.+,

det (jWI - f
k=O
Ake-jkTW) =1= O. (2.36)

Consequently, the system is stable independent of delay, i.e., T = 00. Sup-

pose next that o-(V, U) n 8lIJJ =1= 0, but 0- (L:;;'=o AkZf) = {O} for all
Zi E o-(V, U) n 8lIJJ. This means that (2.36) holds for all w > O. By the
assumption that the system is stable at T = 0, the condition must also
hold at w = 0; namely, it holds for all w E lR.+. Thus, the system remains
stable independent of delay. On the other hand, let ()i E [0, 27r] such that
Zi = e- jOi E o-(V, U), and let Wi > 0 such that jWi E 0- (L:;;'=o Akzf). It
follows that
det(jWJ - fAke-jkTWi) =0
k=O
at T = ()dWi.Thus, the system is unstable at T = T. In contrast, for any
T E [0, T), one can show similarly that the system is stable, as in the proof
for Theorem 2.11. •
66 2. Systems with Commensurate Delays

As a consequence of Theorem 2.13, we see that the stability of a system

given in a state-space description, such as that in (2.5), can also be deter-
mined by computing first the generalized eigenvalues Zi of the 2n 2 m x 2n2 m
matrices V and U, and next the eigenvalues ofthe n x n matrix 2:;;'=0 AkZf ,
without resorting to the system's quasipolynomial. The computation is to
be performed on constant matrices as well, and thus a high computational
precision can be ensured. The downside of this result lies in the significant
increase in the size of the matrices involved in the computation.

Example 2.11 While mainly of computational significance, the constant

matrix tests developed in this section may also be used as analytical tools
for simple systems. We demonstrate this point by revisiting Example 2.1:

:ret) = -x(t - r) - x(t - 2r), r;::: O.

In order to apply Theorem 2.13, we first obtain

~ ), V-(~
1 0 0
U= ( 0 1 0 1 00)
010
001 - 0 o 0 1 '
000 -1 1 101
which in turn yields IT(V, U) = {-I, O.5±O.866j}. Clearly, IT(V, U) CaD.
For Z = -1, IT (2:~=0 AkZk) = {O}. However, for Z = 0.5±0.866j = e±i'Tr/3,
IT (2::~=oAkZk) = {:rJ3 j}. Hence, (7r/3)/J3 = 7r/(3J3). This value
T=
is identical to that obtained in Example 2.1. Similarly, if we are to invoke
Theorem 2.11, we need to construct

0 0
P=' ( 1 0
J -1 0
-1 -1

This results in £T(P) = {O, ± 1.732j}. Furthermore, we have

F(s) = (~s !1)' G(s) = I.

It follows that IT(F(1.732j), I) = {0.5 - 0.866j, - 1.5 + 0.866j}. This,

again, yields T = 7r/(3V3). 0

In summary, the development in this section should have demonstrated

sufficiently well how classical stability criteria for polynomials of one vari-
able can be applied to lead to efficient stability tests in the study of time-
delay systems. Theorems 2.11 and 2.12 are just two of such applications.
Likewise, matrix analysis techniques, which also play an important role in
 2.5 Notes 67

classical stability analysis, can be used to advantage; Theorem 2.13 pro-

vides one such example. In light of the wide variety of classical results, one
anticipates that more stability tests for time-delay systems can be made
available in this manner. Indeed, as an example, one may consider treating
the bivariate polynomial a(s, z) as a complex polynomial in s with coeffi-
cients as functions of z. This makes it possible to analyze the stability of
a(s, z) as a Hurwitz stability problem, one that can be tackled by using
such classical stability results as Hurwitz and Hermite criteria. It is then
possible to develop, in parallel to Theorem 2.11, stability tests of a similar
nature. Nevertheless, as a cautionary note, one should exercise care to as-
certain that the criteria in use be applicable to complex polynomials. For
example, in checking the Hurwitz stability of a complex polynomial, the
Hurwitz test must be replaced by the more general Bilherz criterion. We
leave further deliberations of these aspects to the reader.
In closing this chapter, we emphasize that the classical methods and the
more recent stability tests are rooted in the same fundamental cause and
both seek to compute the zero-crossing frequencies of characteristic func-
tions. Indeed, they only differ in methods of computation. While in the
classical methods the key lies in a variable elimination procedure which
by nature requires symbolic computation and hence can be difficult, the
results developed in the present chapter bypass such a procedure and seek
to compute matrix measures whose computation can be systematically im-
plemented. For systems with commensurate delays, this comes down to
the computation of eigenvalues and generalized eigenvalues of frequency-
dependent or constant matrices, resulting in numerically oriented, more
computable delay-independent and delay-dependent stability tests. The
main asset of these results lies in their computational efficiency. This can be
significant for high order systems with many delays, for which the classical
methods appear to be ill-equipped.

2.5 Notes

2.5.1 Classical results

Frequency domain stability criteria have long been in existence for time-
delay systems. Notable classical examples include results such as the Tsyp-
kin's test and the Pontryagin criterion; a concise elaboration of these and
other classical results can be found in the monographs by Gorecki et al.
[78], Stepan [251]' Niculescu [204], and the survey article by Niculescu et al.
[214]. Formally introduced by Kamen [132]' the term "stability independent
of delay" has since become routinely used in the literature, though delay-
independent and delay-dependent stability problems had been heavily stud-
ied much earlier. Kamen [131]-[133] also advocated the use of two-variable
68 2. Systems with Commensurate Delays

criterion. Further development of the two-variable criterion led to a variety

of stability tests for systems with commensurate delays, among which we
note the polynomial elimination methods by Brierley et al. [25J, Chiasson
[40], Chiasson, Brierley, and Lee [38, 39J, Hertz, Jury, and Zeheb [110], the
pseudo-delay methods advanced by Rekasius [235J, Thowsen [257J-[259J,
Hertz, Jury, and Zeheb [111J, and the iterative method developed in Wal-
ton and Marshall [282J and Marshall et al. [190J. Other closely related
techniques include an optimization algorithm suggested by Lewis and An-
derson [172J and graphical criteria by Mori and Kokame [197J and Mori,
Kokame, and Kuwahara [198J. For an examination of Roche's theorem, we
recommend Levinson and Redheffer [171 J. Two-dimensional systems and
the stability of 2-D systems are extensively documented in Bose [17J.

2.5.2 Frequency-sweeping and constant matrix tests

The idea of using spectral radius and eigenvalues as stability tests were
advanced in Chen [32J, Chen and Latchman [35J, and Chen, Gu, and Nett
[34J. Section 2.3 draws essentially upon these references. More specifically,
frequency-sweeping conditions for stability independent of delay and those
for computing the delay margin are taken from [32, 35J, with the constant
matrix tests from [34J. Similar results have been subsequently pursued by
SU [253J, and Niculescu and Ionescu [211J, which share the essential spirit
of [34J. These efforts together bring the stability problem in the case of
commensurate delays to a satisfactory closure. The matrix background for
this chapter is readily found in many texts; we refer to Golub and Van
Loan [76] and Horn and Johnson [114J. For a comprehensive coverage of
classical stability results such as Schur-Cohn criterion and Orlando formula,
we recommend Barnett [7] and Marden [189].
3

Systems with Incommensurate

Delays
3.1 Introduction
Unlike for those systems with commensurate delays, stability analysis for
systems with incommensurate delays proves far more difficult. In particular,
most of the results known to date are sufficient conditions, and seldom
are they nonconservative. Necessary and sufficient conditions are scarce
and are hardly computable; indeed, many such results merely fall into a
restatement-in one form or another--of the definition of stability, thus
rendering the conditions of little use. The reader will quickly find out that
this lack of strong results is by no means coincidental.
In this chapter we study delay systems with multiple pointwise incom-
mensurate delays. In particular, we focus on the LTI systems described by
the state-space equation
m

x(t) = Aox(t) + LAkX(t - rk), (3.1)

k=l

where the delays rk are assumed to be independent of each other. A special

case in this class is the system

T ~ 0, (3.2)

that is, the system contains only one delay. We shall first develop small gain
type, frequency-sweeping stability conditions for general LTI systems in the
form of (3.1). The results employ a matrix measure known as the struc-
tured singular value, a measure used widely in robust control analyis and
design. The structured singular value approach constitutes a step forward
in the study of systems with incommensurate delays, in that it provides a
necessary and sufficient condition by means of a well-developed theoretical
notion as well as a computational tool, for which standard numerical al-
gorithms and commercial software programs are available. It thus enables
a systematic analysis of the stability problem from a computational per-
spective, which can be accomplished at least approximately. On the other
hand, the exact computation of the structured singular value is also known
to be difficult, and in fact is known to be intractable in general. This raises
the speculation that the stability analysis for systems of incommensurate
70 3. Systems with Incommensurate Delays

delays may be intrinsically difficult. The conjecture consequently leads to

a formal analysis of the computational complexity of the stability problem,
drawing upon concepts and techniques found in decision theory. We shall
attempt to provide a formal proof that the exact stability check would
require solving a so-called NP-hard decision problem. In the language of
computing theory and operations research, such a problem is essentially
equated to one whose computational complexity increases exponentially
with the problem size. Problems in this category are commonly believed to
be and hence are regarded as computationally intractable, especially when
the problem size is large. The NP-hardness proof reveals that the compu-
tation required for determining stability is likely to increase exponentially
with the number of incommensurate delays and thus amounts to a com-
putationally intractable problem in general. For this intrinsic difficulty, we
then develop a number of sufficient stability conditions. The chapter ends
with a brief discussion on neutral delay systems, together with a summary
note on frequency domain methods.

3.2 Small gain/ J-l theorem

We first provide a brief narrative on the classical small gain theorem and
its extensions. The proof of this result together with other relevant details
are relegated to references given at the end of the chapter. The exposure
commences with an input-output description of systems. The subject will
be further discussed in Chapter 8.

3.2.1 Small gain theorem

In studying signals and systems, it is useful to identify signals with function
spaces. Of particular pertinence here is the class of energy-bounded signals,
which, in mathematical terms, are defined as square integrable functions.
More specifically, n-dimensional energy-bounded signals form the function
space

L2 := {J: IR ~ IR n I f(t) is measurable, IIfll£2 < ()()} ,

(1: IIf(t)II~dt)
where
IIfll£2 := 1/2,

and 11·112 is the Holder £2 norm,

n ) 1/2
IIf(t)II2:= ( ~ Ih(t)12
 3.2 Small gain/ J.£ theorem 71

i:
Define the Fourier transform of ! (t) by

I(jw):= !(t)e-jwtdt,

and the space of Lebesgue integrable functions

£2 := {I: lR ---+ en I l(jw) is measurable, 111111:.2 < oo},

where
111111:.2 := (2~ i:
The well-known Parseval identity states that
III(jw)lI~dw) 1/2

(3.3)
In turn, the function spaces L2 and £2 form an isometrical isomorphism.
The Fourier transform thus provides an equivalent characterization of a
signal via its frequency domain representation.
By so identifying the input and output signals with elements of function
spaces, a system can be viewed as an operator mapping the input space to
the output space, as shown in Figure 3.1. In particular, a linear system can
be viewed as a linear operator. Let H: L2 ---+ L2 be a linear system. Then
its induced norm can be defined as

where u E L2 denotes the system's input and w E L2 the output. When

this induced norm is well defined and finite, the system is said to be BIBO
(bounded-input bounded-output) L 2-stable, whose induced norm IIHII2,2 is
finite. For an LTI system, the induced norm can be quantified by computing
the system's frequency response. Let H(s) be the transfer function matrix
of an LTI BIBO L 2 -stable system H. In light of the Parseval identity (3.3),
the induced norm IIHII2,2 can be alternatively characterized as

FIGURE 3.1. Input-output system description

72 3. Systems with Incommensurate Delays

In other words, the induced norm coincides with the peak magnitude of
the system's frequency response, whereas the magnitude is measured by
the largest singular value and has the interpretation as the system's gain.
It follows that an LTI system is BIBO L 2 -stable if and only if its frequency
response is essentially bounded on OC+. If in addition an LTI BIBO L 2 -
stable system is also causal, then its transfer function matrix H(s) will
be analytic in C+ and essentially bounded on 8C+. It is known that the
collection of all such functions forms another function space, known as the
Hardy space of Hoc functions, endowed with the Hoc norm and defined as

Hoc := {H I H(s) analytic in C+, IIHlloc:= sup a(H(s)) < oo} .

sEIC+

Thus, LTI causal BIBO L 2-stable systems correspond to functions in Hoc,

and the maximum gain of an LTI causal BIBO L 2-stable system is equal
to the Hoc norm of its transfer function matrix. This consequently gives an
operator theoretic interpretation of the familiar notion of gain defined in
the frequency domain. One should note that in the present book only causal
systems are considered, and thus for an LTI causal system, the notion of
stability is identified with that of BIBO L 2-stability; that is, a causal LTI
system is BIBO L 2 -stable if and only if its transfer function has no pole in
C+.
Functions in the Hoc space constitute a well-studied mathematical sub-
ject of deep and rich results. The linkage between system analysis and Hoc
theory has an immediate yet profound impact: it furnishes a well-developed
mathematical theory, a powerful tool for the study of causal BIBO L 2 -stable
systems. The stage is thus set to usher in the small gain theorem. Consider
the linear feedback system given in Figure 3.2. The configuration, often
referred to as M-tl loop, is the standard feedback interconnection of two
systems M and tl.

Lemma 3.1 (Small Gain Theorem) Suppose that M and tl are both
causal, linear, and BlBO L 2 -stable. Then the M-tlloop is BlBO L 2 -stable

tl

FIGURE 3.2. Feedback interconnection

 3.2 Small gain/ /l. theorem 73

if
IIM~1I2,2 < 1. (3.4)

Hence, the small gain theorem provides a sufficient condition that governs
the stability of feedback interconnections. One should note that in its full
generality, modulo appropriate modifications, the small gain theorem in
fact is applicable in a much broader sense. Nevertheless, the statement given
in Lemma 3.1 is sufficiently general for our present purpose. In particular,
with a more far-reaching implication, it allows the study of time-varying
systems and can therefore be applied to systems containing time-varying
delays. This will be illustrated shortly via examples in the present chapter
and be investigated in greater depth in Chapter 8.
The small gain theorem has played an especially prominent role in the ro-
bust control of uncertain LTI systems. In robust stability analysis, one typi-
cally models a system's uncertainty by deterministic, unknown but bounded
perturbations to its parameters or transfer function matrix. A specific un-
certainty model, known as unstructured uncertainty, describes unknown
perturbations as stable transfer function matrices with a prescribed 11.00
norm bound, by means of the set description

In other words, for an unstructured uncertainty, i.e., an element of 1311.00 , no

other information is known than the fact that its transfer function matrix
is bounded in the 11.00 norm; without loss of generality, the 11.00 norm
bound can be taken as one. While simplistic, this uncertainty description
nevertheless enables us to strengthen the small gain theorem to a necessary
and sufficient condition for the robust stability of the M-~ loop, when ~
is assumed to be an unstructured uncertainty.

Lemma 3.2 (Small Gain Theorem-Robust Stability) Let ME 11.00 '

Then the M -~ loop is BlBO L 2 -stable for all A E 1311.00 if and only if

(3.5)

In particular, if M(jw) ~ 1 for some w, then it is possible to construct a

"destabilizing" uncertainty A such that A E 1311.00 , and

so that the M-~ loop becomes unstable.

Example 3.1 Consider the LTI delay system

x(t) = Ao x(t) + Al x(t - r).

74 3. Systems with Incommensurate Delays

This system can be realized as in Figure 3.3, in which .6.: L2 ---- L2

is a linear operator such that .6.x(t) = x(t - T). Evidently the system
can be remodeled by the M-.6. loop in Figure 3.3, by taking M so that
£1(s) = (sf - AO)-l AI, and treating fictitiously the delay operator .6. as
an uncertainty. Since 11.6.112,2 = 11b.1100 = 1, and

IIM.6.112,2::; IIMI12,211.6.112,2 = 11£1110011.6.112,2 = 11£11100,

it follows from Lemma 3.1 that the system will be stable whenever Ao is
stable and
II(sf - AO)-l Ailioo < 1.
Thus the small gain theorem gives rise to a sufficient condition for stabil-
ity independent of delay. Similarly, this analysis extends to systems with
multiple delays. Consider the LTI system
m

x(t) = Ao x(t) + I: Ak x (t - rk) .

k=l
The system can be realized as in Figure 3.4, and in turn by the M-.6. loop.

l)
In this case, M is constructed so that

M(s) ~( (sf _AO)-I ( Al

and the delay operators are collectively represented by the block diagonal
uncertainty

Each of the nonzero, diagonal blocks in .6. is a delay operator: .6. kx(t) =
x(t - rk). Since II.6. k I1 2 ,2 = 1, it follows that 11.6.112,2 = 1. Consequently, a
sufficient, delay-independent condition for the system to be stable is that
Ao is stable and 11£11100 < 1.
More generally, the idea of representing a delay by a fictitious uncertainty
can be generalized to systems with time-varying delays. We illustrate this
point briefly below; the idea will be developed in details in Chapter 8.
Consider the linear delay system

x(t) = Aox(t) + AIX(t - T(t)),

where T : R ---- R+ is time-varying. Assume that T is continuously dif-
ferentiable, bounded, and i(t) ::; 0 for all t. Then for any x(t) E L 2 , we
have
 3.2 Small gain/ J.L theorem 75

M
dx/dt I (1/s)1 I X
,J I I~--~--~I~
I
I
I

'-------I
I A0 I I
I
I
I I
I
J
L _________________ - - __ I

I I
I I

FIGURE 3.3. Realization of systems with one delay

where it = 1 - T > 1. Thus,

1~ x 2(t - r(t))dt ~ 1: 00
x2(u)du,

and the linear operator ~x(t) = x(t-r(t)) satisfies the inequality 11~112.2 ~
1. As a result, a sufficient condition for the system to be stable is that Ao
is stable and lI(sf - Ao)-t At 1100 < 1. 0

Example 3.2 Tsypkin's test is concerned with closed-loop stability of

the system depicted in Figure 3.5, where P( s) is a stable, scalar transfer
function, and r ~ 0 is a delay constant. Suppose that IP(O)I < 1. Then
Tsypkin's test states that the system will be stable independent of delay if
and only if IP(jw)1 < 1, for all w > O. This condition, however, is equivalent
to 11P1i00 < 1, which, in light of Lemma 3.2, is necessary and sufficient
for the M-~ loop to be stable, with M(s) = P(s) and'&' E BHoo. The
implication then is that under the assumption IP(O)I < 1, the system can
be represented equivalently by the M-D. loop and its stability be given
by the small gain condition (3.5), despite that the delay may not be fully
equated to an uncertainty. Indeed,though we may represent the delay by an
LTI uncertainty .&.(s) = e-"T"S E BHoo , .&. cannot be completely uncertain
as required, for .&.(0) = 1. Otherwise, at any other w, .&.(jw) can fully
replicate an unstructured uncertainty, since r can be chosen arbitrarily. In
summary, the small gain condition (3.5) will remain necessary and sufficient
when the additional assumption IP(O)I < 1 is imposed at w = 0, but the
necessity of the condition will in general vanish otherwise. This suggests
that in modeling delay operators by uncertainties, caution must be paid to
the zero frequency w = 0. 0
76 3. Systems with Incommensurate Delays

r- - - - - - - - -- - --- - - - - - - - - - - - ----I
1 M 1
1 1
1 X 1
1 1
1 1
1 1
1 1
1 1
1 _______________________ I

1 1

:___________
A 1 1

FIGURE 3.4. Realization of systems with multiple delays

,/ P(s)
-

e- rs

FIGURE 3.5. Closed loop delay system

The examples given above underscore the role that the small gain type
results may play in stability analysis of time-delay systems. They demon-
strate that the small gain theorem can be applied effectively to yield a
sufficient condition for delay-independent stability; sometimes, this is true
even for systems with time-varying delays. In particular, Example 3.2 shows
that in certain situations it is possible to obtain a necessary and sufficient
condition, specifically when the delay can be modeled as an unstructured
uncertainty modulo a side constraint at s = 0, as in the case of the Tsypkin
criterion. This, however, ceases to be true when the system contains multi-
ple delays. Example 3.1 exhibits that in the latter case, one unstructured
uncertainty can no longer represent fully several delays without undue con-
servatism. It thus calls for a more refined description that can incorporate
the structural information of the uncertainty that results when several such
 3.2 Small gain/ Ji, theorem 77

delays are lumped together. For this purpose, the so-called structured un-
certainty provides a remedy.

3.2.2 Structured singular value

In general, structured uncertainty is a broad description and may refer
to any uncertainty that is not unstructured and is meant to provide such
additional, more specific information that a crude description such as un-
structured uncertainty fails to capture. Thus, let ~ be a strict subset in
the vector space C 1x1 . Any .60 E ~ may then be considered a structured
uncertainty of dimension l. The structured singular value, or Ji" provides a
matrix measure contingent on the structure defined by ~.

Definition 3.1 The structured singular value ilL!!. (M) of a matrix M E

C 1x1 with respect to the block structure ~ is defined to be zero if there is
no .60 E ~ such that det(I - M D.) = 0, and otherwise

1lL!!.(M) = (min {O'(.6o) I det(I - M.6o) =

Ll.EL!!.
O})-l
When applied to the M-.6o loop of LTI systems, the structured singular
value defines a precise frequency-dependent robust stability measure with
respect to uncertainties of the structure ~, for any structure ~. The exact
robust stability condition is furnished by an extended form of the small
gain theorem, sometimes dubbed as "small Ji, theorem."

Lemma 3.3 (Small Il Theorem) Let if E Hoo and define

~oo := {Li E Hool IILilloo : : : 1, Li(jw) E ~, Vw E 1R}.

Then the M -.60 loop is EIBO L 2 -stable for all Li E ~oo if and only if

(3.6)

where

A canonical uncertainty structure is that of block diagonal matrices; it is

known that an uncertainty of any arbitrary structure can be transformed
readily into a block diagonal one by simple algebraic manipulation. For
a block diagonal uncertainty, every .60 E ~ may contain nonzero entries
only on its main diagonal, while all off-diagonal blocks are zero. Example
3.1 shows that an even more specified diagonal structure is particularly
78 3. Systems with Incommensurate Delays

pertinent in dealing with systems of multiple delays. More specifically, the

block diagonal uncertainties possess the form of
m
(3.7)

Uncertainties of this structure, known as "repeated complex scalars" in the

robust control literature, will be considered in the remainder of this chapter.
One should note that other types of uncertainty structures such as "full
complex," "repeated real scalar," and "full real" blocks of uncertainty can
also be augmented in Xm whenever needed. While of no direct relevance
herein, the latter classes of uncertainties find utilities in studying robust
stability of time-delay systems.
It follows from Lemma 3.3 that in the presence of structured uncertain-
ties, the stability of the M-fl loop can be ascertained by computing point-
wise in frequency the structured singular value J.L a (if (jw)). Unfortunately,
this computation is known to be difficult; in general, the computational
problem has been found to be as difficult as one of NP-hard decision prob-
lems, and hence is unlikely to be tractable computationally. The standard
practice in robust control then is to resort to approximating computation
via upper and lower bounds. Several such bounds are available and are
summarized below together with other relevant properties. We note that
most of these properties are handy and are thus provided without proof.

Lemma 3.4 The following properties hold for J.La(M).

(i) If d = {8I1 8 E C}, then J.La(M) = p(M).

(ii) If d = C 1X1 , then J.La(M) = 7J(M).
(iii) For any block diagonal d, p(M)::; J.La(M) ::;7J(M).
(iv) Let if E 'Hoc. Then for any block diagonal d, J.L a (if (s)) is a sub-
harmonic function in C+ and its maximum is achieved on 8C+.
(v) For any block diagonal d, define

Bd:= {fl Edl 7J(fl) ::; I},

U:= {U E dl UU* = I} .
In particular, denote U by Um if d = X m , i.e.,
Um : = {U E Xml UU* = I}.
Then J.La(M) = J.La(UM) = J.La(MU). Furthermore,

J.La(M) = m~ p(Mfl) = maxp(MU). (3.8)

aEBa UEU
 3.3 Frequency-sweeping conditions 79

(vi) Let ~ = X m . Define

Vm: = {diag(Dl' D m ) I D·E""kiXki
t \l.,.
D·-D
,1 . -
H }•
1.

Then, f.La(M) = f.La(DMD- 1 ) fOT any DE V m . In addition,

(3.9)

(vii) Let ~ = X m . Then infDEv", a(DMD-l) :::; 1 if and only if there

exists some D E Vm such that

D-MHDM2:0. (3.10)

(viii) Let~ = X m . Suppose that M = baH with a, b E C 1, and partition

a, b compatibly with Xm • Then,
m

f.La(M) =L la[ibil· (3.11)

i=1

Hence, f.L a (M) can be approximately computed by solving the minimiza-

tion problem in (3.9), or the maximization problem in (3.8). The property
(vii) indicates that the former is equivalent to an LMI problem, which
has a unique minimum and can be solved systematically using available
software programs. The latter, while having multiple local maxima, can
be approached by means of a power algorithm. The property (iv) reveals
that f.La(lvf(s)) still possesses some desirable analytic properties that aid
in checking robust stability. For example, it is sub harmonic and hence a
continuous function, wherever M(s) is analytic. In particular, since every
subharmonic function satisfies the maximum modulus principle, f.La(M(s))
achieves its maximum on 8C+, i.e., the boundary of C+, whenever M(s) is
stable. Finally, the properties (i), (ii), and (viii) show that in some special
cases closed-form expressions can be found for f.L a ( M) .

3.3 Frequency-sweeping conditions

Several leads now point to the possibility that the structured singular value
may be used as a measure for delay-independent stability of systems with
multiple incommensurate delays. Indeed, Theorem 2.1 indicates that in the
more specialized case of a single delay, the structured singular value offers
a necessary and sufficient condition, for as demonstrated in Example 3.2,
a single delay can be treated as a repeated complex scalar uncertainty, for
which f.L reduces to the spectral radius. Example 3.1 also shows that it is
possible to model multiple delays by repeated complex scalar uncertainties,
80 3. Systems with Incommensurate Delays

collected in a structured uncertainty of the block diagonal form described

by Xm • That the small f. L theorem may serve as a necessary and sufficient
condition in general then becomes intuitively clear, though the condition
must be addressed separately at w = 0, as we have been so alerted by
Example 3.2. In fact, based on Example 3.1 and Lemma 3.3, the reader
may already come to recognize the following result.

Theorem 3.5 Let rk ~ 0, k = 1, 2, "', m be independent, incommen-

surate delays. Then the system (3.1) is stable independent of delay if and
only if

(i) Ao is stable,
m
(ii) L Ak is stable, and
k=O
(iii)
f..Lx", (M(jw)) < 1, Vw > 0, (3.12)

where

M(s) ,~ ( 1)(s[ - Ao)-I ( A, (3.13)

Proof. The conditions (i) and (ii) are required to ensure that the system
be stable for rk = 00 and rk = 0, for all k = 1, "', m, and hence are
necessary. Under the condition (i), the system is stable independent of delay
if and only if for all s E C+

(3.14)

Let
~(s) := diag (e- r1S I, "', e-r",s 1) .

It follows that (3.14) is satisfied for all s E C+ if and only if

det(I - M(s)~(s)) =1= 0,

Since M(s) and ~(s) are analytic in C+, the spectral radius p(M(s)~(s)),
either in its own right or as a special case of f..Lx", (M(s)~(s)), defines a
subharmonic function in C+. As a result, it achieves its maximum on ac+.
That is,
sup p(M(s)~(s)) = supp(M(jw)~(jw)).
SEC+ wER
 3.3 Frequency-sweeping conditions 81

Since rk are arbitrary, independent real constants, we may replace il(jw)

by an arbitrary constant matrix U E Urn, for all w > O. Consequently,

sup sup p (M(jw)il(jw)) sup sup p (M(jw)U)

rk2:0 w>O w>OUEU",
sup P'x", (M(jw)).
w>O

Therefore, whenever (3.12) is true, the condition (3.14) holds for all s E «:\
except at s = O. However, under the condition (ii), it is necessary that

This implies that (3.14) also holds at s = 0, and hence it does for all s E C+.
In other words, the system is stable independent of delay. The proof for
the sufficiency part is completed. To show the necessity, assume first that
J-tx", (M(jwo)) = 1 for some wo > O. This means that a matrix U E Urn
exists such that

U = diag(e- jIl1 I, ... , e- jllm 1), (h E [0, 27f], k = 1, ... , m,

and that det (I - M(jwo)U) = O. A set of delay constants can be con-
structed as rk = (h/wo, so that the condition (3.14) is violated at s = jwo.
Thus, the system cannot be stable independent of delay. Additionally, if
J-tx", (M(jw)) > 1 for some w > 0, then by the continuity of J-txm (M(jw)),
and in addition the fact that

lim J-tx (M(jw))

w-+oo Tn
= 0,
there must exist some Wo E (w, 00) such that J-tXm (M(jwo)) = 1. As such,
we may also conclude that the system is not stable independent of delay.
The proof is now completed. _
Just as the structured singular value serves as a nonstandard matrix
measure generalizing the common notion of singular values, when applied
to a transfer function matrix, it extends the concept of gain defined on the
basis of singular values. Theorem 3.5 then provides a small gain type con-
dition for delay-independent stability. There is, however, a subtle difference
between the condition (3.12) and the small J-t condition, which concerns the
frequency w = O. We have repeatedly warned against this difference. For
a further clarification, we state the following alternative to Theorem 3.5.
The proof is similar and hence omitted.

Theorem 3.6 Let rk ~ 0, k = 1, 2, ... , m be independent, incommen-

surate delays. Then the system (3.1) is stable independent of delay if and
only if
82 3. Systems with Incommensurate Delays

(i) Ao is stable;
(ii) J.Lxm (M(O)) < 1, or J.Lxm (M(O)) = 1 but det (I - M(O)) =J 0;
(iii) J.Lxm (M(jw)) < 1, Vw> O.

Consequently, while the small J.L theorem requires that J.Lxm (M(O)) < 1,
it is possible that J.Lxm (M(O)) ::; 1 while the system may still be stable
independent of delay.
It is then clear that in general the condition for delay-independent stabil-
ity does not match exactly that for robust stability; the latter is stronger
and in fact provides only a sufficient condition for stability independent
of delay. The reason, evidently, points to the fact that the delays cannot
be taken exactly as uncertainties, a phenomenon we have already observed
in Example 3.2. This difference helps explain a discrepancy between any
necessary and sufficient condition and the classical two-variable criterion,
which was perceived for some time as a necessary and sufficient condition
for stability independent of delay. To further illustrate, consider the case
of one delay. The following corollary is a counterpart to Theorem 2.1 and
follows immediately from Theorem 3.6.

Corollary 3.7 System (3.2) is stable independent of delay if and only if

(i) Ao is stable;
(ii) p (Ai) I AI) < 1, or p (Ai}l AI) = 1 but det(Ao + AI) =J 0;
(iii) P (jwI - AO)-I AI) < 1, Vw > O.
Suppose that Ao is stable. A sufficient condition for stability independent
of delay is then given by

Vw~O. (3.15)

This, however, coincides with the condition

Vw ~ 0,

for .6. = {811 8 E C}. It thus follows that the condition (3.15) is equivalent
to
det(1 - (sl - Ao)-I AIZ) =J 0,
which is equivalent to the standard two-variable criterion

det(sl - Ao - AIZ) =J 0, (3.16)

In other words, the two-variable criterion, which coincides with a small J.L
condition, is only a sufficient condition. One notes that for such a condition
to be also necessary, there must exist an "uncertainty" ~(s) E .6. 00 with
which the matrix I - M(s)~(s) loses rank. Since T ~ 0 can be selected
 3.3 Frequency-sweeping conditions 83

arbitrarily, for s :/= 0, such an uncertainty can be constructed as ~(s) =

e- rs I, so that the matrix 1 - (sl - AO)-l Al~(S) loses rank. However, the
construction fails at s = 0, for e- rs 1 = 1 for any T. This is precisely where
the two-variable criterion (3.16), or more generally the small f.l condition,
loses the necessity; the fictitious uncertainty we use to model the delay is
fixed, rather than uncertain, at s = 0.
We emphasize, in light of the difficulty in computing the structured sin-
gular value, that Theorems 3.5 and 3.6 can only provide in effect sufficient
or necessary conditions for stability independent of delay, for in general
only upper and lower bounds of the structured singular value may be ef-
ficiently computed. Nevertheless, while computationally less encouraging,
the results do provide an exact characterization that can be analyzed sys-
tematically and computed approximately. As a direct benefit, it will enable
the derivation of many sufficient or necessary conditions; a series of such
sufficient conditions will be presented later in this chapter. Moreover, for
certain special cases where the structured singular value admits closed-form
expressions, the characterization leads to exact, closed-form stability crite-
ria. The case of a single delay is one such example. Another case of interest
arises when the transfer function matrix M(s) is rank-one. The scalar delay
system described by the differential-difference equation
n-I m
y(n)(t) +L
I>kiy(i)(t - rk) = 0, (3.17)
k=Oi=O

belongs to this category. Define the polynomials

n-I
ao(s) Sn + "'"'
~aOis,
i (3.18)
i=O
n-I
Laki si , k = 1, 2, .'., m. (3.19)
i=O

Corollary 3.8 Let rk ;::: 0, k = 1, 2, .", m be independent, incommen-

surate delays. Then the following statements are true.
(1) Suppose that for some u E ]Rn and v = ( vi T
vm )T E ]Rl ,
( Al Am ) = uv T .
Then the system (3.1) is stable independent of delay if and only if

(i) Ao is stable;
m

(ii) L Ak is stable;
k=O
m

(iii) L IV[(jwl - Ao)-Iul < 1, \/w > 0.

k=I
84 3. Systems with Incommensurate Delays

(2) Let ak(s), k = 0, 1, ... , m, be defined in (3.18) and (3.19). Then the
system (3.17) is stable independent of delay if and only if

(i) ao(s) is stable;

m

(ii) I>k(S) is stable;

k=O
(iii)

"iw > O.

(3) The first-order scalar delay system

q
y(t) + a y(t) + L ak y(t - rk) = 0,
k=l
is stable independent of delay if and only if

(i) a > 0;
q q q

(ii) a - L lak I > 0, or a - L lak I = 0 but a + L ak =I O.

k=l k=l k=l

Proof. The statement (1) follows from Lemma 3.4 (viii), which leads to
m
fLx,,,(M(jw)) =L IVk(jwI - AO)-lul·
k=l
To prove the statement (2), note that it constitutes a special case of (1),
with
o
Ao
1
-aO,n-l
)
,

This gives rise to

 3.3 Frequency-sweeping conditions 85

Finally, the statement (3) is easily recognized as a special case of (2), with
q
L lakl
/1. x..{M(jw)) = ;:~ + a2
The proof is then completed. •
We conclude this section by presenting a small gain type sufficient condi-
tion for delay-dependent stability, using also the structured singular value.
Before proceeding, we note that in principle a delay-dependent stability
condition can be obtained in the process of computing /1.xm (M(jw)) , in
verifying (3.12). Alternatively, it may also be obtained by solving a maxi-
mization problem of the form described by (3.8). In either case, the solution,
were it attainable, would yield certain critical, "destabilizing" .6. E Urn, at
the frequencies w where /1.x,jM(jw)) = 1. The critical delay values may
then be estimated from .6. and w. Evidently, while the estimates would
provide an exact range of delay values for which the system is stable, the
computation of such critical .6. is, unfortunately, as difficult as that of
/1.xm (M(jw)) , if not more, and hence from a computational standpoint is
unlikely to be tractable in general. The condition to be developed below
inherits the same complexity in computing /1., but is more amenable to
analysis.
Consider the system (3.1) with incommensurate delays rk. Using

x(t - ri) = x(t) -l~ri x(u)du = x(t) -l~ri (~AkX(U - rk)) du,
we may rewrite the state-space equation (3.1) as

(3.20)

The process of transforming the original equation (3.1) to the new repre-
sentation (3.20) is known as model transformation, which is widely used in
both frequency and time domain stability analysis of time-delay systems.
It should be pointed out that although the stability of (3.20) implies that
of (3.1), the reverse is not necessarily true due to the presence of additional
dynamics. This subtle point will be illustrated in Chapter 5. The Laplace
transform of the integral term is given by

Hence, for any rk, k = 1, ... , m, the system is stable if

#0,
86 3. Systems with Incommensurate Delays

Under the assumption that A := E~o Ai is stable, this condition is equiv-

alent to

det [ 1 - (s1 - A)-I ~Ai

m
e-r'S
's - 1 ( ~Ake-rkS
m ) 1i 0, 'Vs E C+.

(3.21)
This consequently leads to a delay-dependent stability condition character-
ized by the structured singular value.

Theorem 3.9 Let rk ~ 0, k = 1, 2, ,m be independent, incom-

mensurate delays. Then the system (3.1) is stable for all rk E [0, Tk),
k = 1, ... ,,'.I
m iF
m
(i) E Ak is stable,
k=O
(ii)
JLa (M(jw; T)) < 1, 'Vw>O, (3.22)

M(s; T):= C)(Sf - (~~)r (r,A,C ... rmAmC),

C := (Ao Al ... Am).

Proof. For any Tk > 0, k = 1, ... , m, we may first rewrite the condition
(3.21) as

or equivalently

det(1 - M(s; T)~(S)) i 0, 'Vs E C+,

where

~(s) := diag(~I(s), ~2(S), .", ~m(s)),

~i(S) := oi(s)diag (1, e- 1, ... , e-r",s 1) ,
r18
e- ri8 - 1
Oi(S) :=

It thus suffices to show that Oi (s) E Bl£ oo , for all i = 1, , m, and to

invoke the small JL theorem. Evidently, Oi(S) E 1£00. Furthermore, since for
 3.4 Computational complexity analysis 87

any rk E [0, rk), k = 1, "', m,

we have 118i l oo ::; 1. This completes the proof. •

The idea behind the above derivation is rather clear and familiar; it also
seeks to approximate delay effects by a structured uncertainty, though in a
different way. It should be clear that the model transformation technique
can be applied in further depth to yield more sophisticated representations
of the delay system, hence leading to other forms of uncertainty to model
the delay effects and consequently to additional delay-dependent stability
conditions of a similar nature. More generally, such approximations need
not be restricted to model transformation alone; in fact, the delay effects
may be modeled by uncertainties in a variety of ways. Nevertheless, fun-
damental computational difficulty remains. Clearly, the barrier lies at the
computation of the structured singular value, thus necessitating a trade-
off between the conservatism of a test and its computational complexity.
Moreover, while Theorem 3.9 does provide a delay-dependent stability con-
dition, it mayor may not be less conservative than the delay-independent
condition given in Theorem 3.5.

3.4 Computational complexity analysis

The computation of structured singular values is known to be difficult. That
a stability test is equated to a structured singular value problem thus casts
doubts on that test's practical feasibility. For systems with incommensurate
delays, we have seen, in light of Theorem 3.5, that the exact verification of
delay-independent stability must bear the same degree of difficulty as that
of computing a specific j.L, corresponding to a specific M(s) and the specific
uncertainty structure Xm . Delay-dependent stability will only prove more
difficult. Can the computation of that particular j.L, as a special case of the
general structured singular value problem, be tractable? Or is the stability
problem inherently difficult? While it would be clearly invalid to conclude
anything based on the known computational difficulty of the general j.L
problem, we provide a formal, conclusive proof to show that the delay-
independent stability problem is in general equivalent to one of NP-hard
decision problems, whereas the notion of NP-hardness will be defined in a
precise manner. The proof, which requires concepts and techniques found in
computational complexity theory, thus translates into the statement that
the stability problem is indeed fundamentally difficult, and in general is
unlikely to be tractable computationally.
88 3. Systems with Incommensurate Delays

3.4.1 Basic complexity concepts

The computational complexity theory we refer to here is a paradigm con-
cerning the study of the intrinsic computational difficulty, and therefore
the classification of problems into classes that are tractable or intractable.
Whether a problem is tractable or not is measured by the computational
requirement in computing a solution, that is, the demand in computer time
and space for solving the problem. An intractable problem may be solvable
in theory provided that unlimited resources of computation are available,
but is unlikely to be solvable in practice in any realistic sense. The the-
ory has become an extensively developed branch of computing science and
operation research, built on languages and concepts that seem unfamiliar,
though not uncommon, to the field of systems and control, and as such
are rather removed from the general subject of stability analysis. We will
therefore attempt to provide only a minimal glimpse into the theory re-
quired in the present context and focus exclusively on its pertinence to our
stability problem. The exposure, for this reason, will be mainly informative
though deliberately informal, in order to avoid overloading the reader with
additional, unnecessary complications.
The central issue in the complexity theory dwells on the notions of ef-
ficient algorithms and tractable problems solvable by efficient algorithms.
An efficient algorithm, also known as a polynomial-time algorithm, is one
that can perform the computation in polynomial time; in other words,
the required computation time, or the number of basic operations, can
be bounded by a polynomial function of the problem size, or the length
of input data. A problem is deemed tractable if it can be solved by a
polynomial-time algorithm in the worst case. More specifically, let ¢>(x)
denote the number of basic operations required for computing a solution
from the available data x. Let also n be the size ofthe data length, or more
generally, the problem size, which can be thought of as the dimension of a
problem or the number of unknowns. Under some common, standard mea-
sure of computation efficiency (known as the Turing machine in computer
science), the problem is said to be polynomial-time solvable if there exists
a positive integer q such that for all possible data x,

that is, asymptotically, the computation cost will not grow faster than
at a polynomial rate, in the worst case and hence in all instances of the
problem. It is customary to refer to this class of problems as the class P.
For a problem not in the class P, there exists at least one instance for
which the required computation time cannot be bounded by a polynomial
function. Problems in the latter category are generally held intractable.
One should be cautious in interpreting the practical difficulty in computing
problems in P. The formalism gives no indication and indeed de-emphasizes
the importance of nand q. It could well be possible that for all practical
 3.4 Computational complexity analysis 89

purposes, a problem of a complexity O(n lOO ) cannot be realistically solved

even for a moderate n, though it belongs to P and therefore is considered
tractable.
There is yet another class of decision problems whose membership in P
cannot be--in fact, has never been-ascertained, but whose solution can
be verified via a polynomial-time algorithm, if such a solution has been
postulated a priori in some manner. This class of problems are referred to
as the class NP. Intuitively, it is easier to verify than to find a solution.
Indeed, it is an immediate fact that

PcNP.

An analogy can be made here between finding a difficult proof for a mathe-
matical theorem versus checking an existing proof. A longstanding problem
in the complexity theory, one that remains unresolved today and seemingly
will be open in the foreseeable future, questions whether

P=lNP.

While no conclusion has been reached either way, it has been generally
believed that this is likely to be true, a de facto proposition supported by
ample empirical evidence. The fact, under the premise that it does hold,
commands fundamental importance in complexity studies, due to a strong
property of problems in NP. It turns out that NP contains a subclass in
which every problem can be transformed via a polynomial-time algorithm
to another problem in the same subclass. Such a problem is said to be N P-
complete. More specifically, we say that a problem P is NP-hard if every
problem in NP is polynomially reducible to P, and NP-complete if it is
NP-hard and P E NP. Here by the polynomial reducibility of a problem
PI to another problem P2, we mean that every instance PI of PI can be
transformed by a polynomial-time algorithm to an instance P2 of P2, so that
PI admits a solution if and only if P2 does, and PI does not admit a solution
if and only if P2 does not. The implication then is that it is unlikely to find a
polynomial-time algorithm for an NP-hard problem P, for otherwise every
problem in NP can be polynomially reduced to P and consequently solved
via a polynomial-time algorithm, but this would necessarily contradict the
fact that P =I NP. Similarly, since an NP-complete problem is also NP-
hard, NP-complete problems are unlikely to be tractable as well.
That every problem in NP can be polynomially reduced to an NP-hard
problem suggests a proof for NP-hardness. It follows that if a problem
PI is NP-hard and polynomially reducible to some problem P2, then P2
must also be NP-hard. In other words, if one is to prove that P2 is NP-
hard, one may attempt to reduce polynomially a known NP-hard problem
PI to P2· In particular, it suffices to polynomially reduce PI to just one
instance of P2, since that particular instance of P2 is as difficult as any
90 3. Systems with Incommensurate Delays

instance of PI, and therefore in the worst case, P2 must be as difficult

as Pl. A vast number of NP-hard and NP-complete problems have been
discovered in this manner, among them notably are such classical problems
as the satisfiability problem and the traveling salesman problem. The same
strategy is adopted in our proof of the NP-hardness of the stability problem.

3·4·2 Proof of NP-hardness

In proving that the stability problem is NP-hard, we attempt to reduce
polynomially a well-established NP-hard problem, known as the Knapsack
problem, to a sequence of NP-hard problems leading finally to the stability
problem. In the reduction procedure, we construct one specific instance for
each problem and show that the Knapsack problem reduces polynomially
to that instance.
We begin with the statement of the Knapsack problem.
Knapsack Problem Given a nonzero vector of integers c E zn, determine
if there exists a vector X E {-I, l}n such that cT X = O. This problem is
called the Knapsack problem.
The Knapsack problem is a classical integer programming problem that is
known to be NP-complete. We need to expand it to complex numbers of
unit modulus, a problem called the T-Knapsack problem.
T-Knapsack Problem Given a sequence of nonzero vectors ai E zn,
i = 1, ..., q, determine if there exists a vector Z E Tn such that Z = 0, aT
for all i = 1, ... , q. This problem is called the T-Knapsack problem.

Lemma 3.10 The T-Knapsack problem is NP-hard.

Proof. Consider first the equation

1 + Zl + Z2 = 0, (3.23)
for some Zl = ej91 , Z2 = e j92 . Evidently, it admits a solution if and only if

cos (h + cos O2 -1,

sin 0 1 + sin O2 O.
By solving the latter equations, we obtain the solution for (3.23):
Zl = e±j (271" /3) ,
Next, consider the equations
Zo + Zi,1 + Zi,2 = 0, i = 1, n, (3.24)
and n
L Ci (Zi,l - Zi,2) = 0, (3.25)
i=l
 3.4 Computational complexity analysis 91

where C = ( Cl Cn ) T E zn is a nonzero vector, and

Z := (zo Zl,l Zl,2 Zn,l Zn,2) T E 1l'2n+l.

Construct the vector sequence al, ... , an+l E z2n+l as

ai (1 0 0 1 1 0 o )T, i = 1, n,

an+l ( 0 Cl -Cl Cn -Cn )T.

It follows that (3.24-3.25) hold if and only if

i = 1, ... , n + 1. (3.26)

Suppose that the Knapsack problem admits a solution X E {-I, l}n for
some nonzero C E zn. Let Zo = 1, and

if Xi = 1
i = 1, ... , n.
if Xi =-1

In view of the solution to (3.23), it is clear that Z satisfies equations (3.24-

3.25) and hence is a solution to (3.26). Conversely, suppose that (3.26)
admits a solution Z E 1l'2n+l. This implies that
Z· 1 z· 2
1 + -'-' + -'-' = 0, i = 1, n,
Zo Zo
and in turn
Zi,l = e±j (211" /3) ,
Zo
We may then find

X 2. -_ Zi,l - Zi,2 ±1 = 1, ... ,

Zo (e j (211" /3) _ e- j (21r /3)) = , i n,

and hence an X E {-I, l}n, so that cT X = O. In other words, we have

shown that the Knapsack problem can be polynomially reduced to an in-
stance of the 1l'-Knapsack problem. Based on the NP-hardness of the former,
we conclude that the 1l'-Knapsack problem is NP-hard . •
The next result shows that the 1r-Knapsack problem is polynomially
reducible to a number of optimization problems, known respectively as
sesquilinear and bilinear programming problems.

Lemma 3.11 For a given positive definite matrix A E lR nxn , the following
problems are NP-hard.

(i) Sesquilinear programming: Determine if sup IZH AZI < 1.

ZEW n
92 3. Systems with Incommensurate Delays

(ii) Bilinear programming: Determine if sup IZ[ AZ21 < 1.

Zl,Z2E1iii n

Proof. For a positive definite A, we may write A = QH Q. Furthermore,

for any Zb Z2 E iIDn ,

Hence,
sup IZr AZ21 S; sup IIQZII~ = sup IZ H AZI.
Zl,Z2E1iii n ZE1iii n zE1iii n

Additionally, it is evident that

sup IZr AZ21 ~ sup IZ H AZI.

Zl,Z2E1iii n ZE1iii n

This establishes the fact that

sup IZr AZ21 = sup IZ H AZI.

Zl,Z2E1iii n ZE1iii n

Similarly,
sup IZr AZ21 = sup IZ H AZI.
Zl,Z2E1['n ZE1['n

We claim that

sup IZr AZ21 = sup IZr AZ21.

Zl,Z2E1iii n Zl,Z2E1['n

This can be seen either as a consequence of the maximum modulus princi-

ple, whereas Z[ AZ2 is viewed as a multivariate complex function, or more
simply by invoking the properties of the structured singular value. Indeed,
define

1 ).

Then in light of Lemma 3.4 (v), it is easy to observe that with kl = ... =
k2n = 1,

sup IZr AZ21 = sup P(A~lB~2)

Zl,Z2E1iii n ~1'~2EB Xn

sup p2 (M~)

sup p2 (MU)
UEU2n

= sup p (AU1 BU2 )

U j ,U2EUn

sup IZr AZ21.

Zl,Z2E1['n
 3.4 Computational complexity analysis 93

We have thus shown that

sup IZ HAZI = sup IZr AZ2 1

zEliiin Zl,Z2Eliiin

= sup IZr AZ2 1 (3.27)

Zl,Z2ETn

sup IZ HAZI.
ZETn

Next, we construct a specific positive definite matrix A. Let c E be a zn

nonzero vector, and let ai be given as in the proof of Lemma 3.10. Define

1
(3.28)

It follows from (3.27) that

sup IZ HAZI = sup IZ HAZI =1- - a21

z~ Z~ n+

Hence, if the 'lI'-Knapsack problem has a solution, then sup IZH AZI
zEliiin
1. Otherwise, sup IZH AZI < 1. Consequently, the 'lI'-Knapsack problem
ZEliii n
has been polynomially reduced to a special instance of the sesquilinear
programming problem, and similarly the bilinear programming problem.
We then conclude that both problems are NP-hard . •
The above development finally culminates at the main result of this sec-
tion. We are now ready to construct one specific system with incommen-
surate delays, for which we show that it is stable independent of delay if
and only if the 'lI'-Knapsack problem admits a solution. The implication
then is that the 'lI'-Knapsack problem is polynomially reducible to a par-
ticular instance of the stability problem, thereby enabling us to conclude
the NP-hardness of the latter. Consequently, we conclude that the compu-
tational complexity in checking the stability will grow exponentially with
the number of incommensurate delays.

Theorem 3.12 Let rk ~ 0, k = 1, 2, "', m be independent, incommen-

surate delays. Then the following problems are NP-hard.

(i) Is the system (3.1) stable independent of delay'?

(ii) Is the system (3.1) stable for all rk E [0, Tk), where Tk > 0, k =
1, "', m?
94 3. Systems with Incommensurate Delays

Proof. Let Wo > 0 and A E IR nxn be the positive definite matrix in

(3.28). Define

Ao:=
CA-I 0f
-Wo
0
-1
0
wof
0
_A-l
W~Io ) ' (3.29)
0 -wof 0 -1

and
eke~+l k = 1, ... , n
e n+1 eI_n k = n + 1, ... , 2n
(3.30)
ek-n+lern+2 k = 2n + 1, ... , 3n
e2n+2eI_2n+l k = 3n + 1, ... , 4n
where ek E 1R2n+2 is the kth Euclidean coordinate. Denote m = 2n + 2,
m
and 4> .. := Ao + L: Akzk . It is clear that for any Zk E iiii, k = 1, ... , m,
k=l

4> .. =
(
_A-l
ZT
2 f
wof
o 0)
wof
-Wo _A-l Z3 '
o zI -1
for some Zi E F, i = 1, 2, 3, 4. Suppose that sup IZH AZI < 1. This
ZEliiR
implies that

It then follows that

-A-l Za o
4>.. + 4>~ _ (
2 -
Z!j
0
o
-1
0
0
o
-A-l
ZH
1. )
-1
< 0,
b

where
Za = Zl +Z2
2
-n
ElI),

Since for any >. E 0'(4) .. ),

we conclude that 0'(4) .. ) n C+ = 0 for all Zi E iiii n , i = 1, 2, 3, 4. In other

words, 4>.. has no eigenvalue in the closed right half plane, or equivalently,

't/Zk E iiii, k = 1, ... , m.

 3.4 Computational complexity analysis 95

Evidently, the condition is equivalent to

for all rk ~ O. Consequently, with the matrices A k , k = 0, 1, "', m

constructed in (3.29) and (3.30), the system (3.1) is stable independent of
delay. On the other hand, suppose that sup IZH AZI = 1. Define
ZED"

Since, in light of (3.27),

sup IZ HAZI = sup IZ HAZI = 1,

ZEY" ZED"

there exists a Z E Tn such that det(W z) = O. This is easily seen by an

appeal to the Schur determinant formula. Using the same formula, it is
straightforward to verify that

det (jWoI - [-~~I ~:]) = O.

Let Zl = Z3 = Z, and Z2 = Z3 = Z. Then,

~ _ ( Wz woI)
z- -woI Wz .

Consequently, there exist Ok E [0, 211'], k = 1, "', n, such that

e-j8k k = 1, "', n
z '- { e- j (27r- 8 k) k = n + 1, . . . , 2n
k .- e-j8k k = 2n + 1, "', 3n
e- j (27r-8 k ) k = 3n + 1, "', 4n

and det (jwoI - ~z) = 0, or equivalently,

det (jWOI - Ao - f
k=l
AkZk) = O.

By equating

Ok/WO k= 1, "', n
(211' - Ok)/WO k= n+ 1, "', 2n
Ok/WO k= 2n + 1, "', 3n
(211' - Ok)/WO k= 3n + 1, "', 4n
96 3. Systems with Incommensurate Delays

we have
det (jWOI - Ao - f
k=l
Ake-jrkWo) = O.

That is, with the matrices Ak in (3.29) and (3.30), the system (3.1) can-
not be stable independent of delay. To summarize, we have shown that
if the 'lI'-Knapsack problem admits a solution, then we may construct via
a polynomial-time algorithm the matrix A in (3.28), and in turn a delay
system (3.1) with Ak given by (3.29) and (3.30), so that it is stable inde-
pendent of delay. Conversely, if the 'lI'-Knapsack problem has no solution,
then the same delay system cannot be stable independent of delay. In other
words, we have reduced polynomially the NP-hard 'lI'-Knapsack problem to
a specific instance of the problem of determining whether the system (3.1)
is stable independent of delay. Consequently, we have proven Theorem 3.12
(i); that is, for systems with independent, incommensurate delays, the prob-
lem of checking delay-independent stability is NP-hard. Since the case (i) is
a special case of (ii), it follows that Theorem 3.12 (ii) is also true: the prob-
lem of checking delay-dependent stability, for any delay region bounded by
rk, is NP-hard as well . •

3.5 Sufficient stability conditions

The computational requirement of necessary and sufficient conditions-
which can be demanding even for systems with commensurate delays and
has been proven impractical for systems with incommensurate delays-
makes efficient, less demanding sufficient conditions an attractive alterna-
tive. Such conditions provide a complementary tool in general, and indeed
appear to be the sole means for systems of incommensurate delays in partic-
ular. With these incentives, we now develop sufficient stability conditions.
For the time being, we shall be mainly interested in simple, readily com-
putable conditions that require only a moderate amount of computation.
Many sufficient conditions the reader will come across in the later chapters
will be more complex and will generally require solving an LMI optimiza-
tion problem. These conditions will be developed based on time domain
techniques. The reader will see that the small gain type stability criteria
developed herein provide a potent vehicle in the development of sufficient
conditions; with a variety of matrix norms and measures coming to aid, it
leads readily to many sufficient conditions of interest.

3.5.1 Systems of one delay

We begin our development with systems of one delay. The following result
provides a sequence of sufficient conditions for delay-independent stability.
 3.5 Sufficient stability conditions 97

Theorem 3.13 Suppose that Ao is stable. Then the system (3.2) is stable
independent of delay if one of the following conditions holds.

(i) p ((jwI - AO)-l AI) < 1, Vw ~ 0.

(ii) For any induced matrix norm II· II,

II(jwI - AO)-l Alii < 1, Vw ~ 0.

(iii) (j ((jwI - AO)-l AI) < 1, Vw ~ 0, or equivalently,

lI(sI - AO)-l Al 1100 < 1.

(iv) For any induced matrix norm 11·11 and the corresponding matrix mea-
sure v(·), IIAIII < -v(Ao).
(v) For any induced matrix norm 11·11 such that IleAotll :::; (e-T/t, Vt ~ 0
for some ( ~ 1 and 7] > 0, IIAdl < 7]/(.

Proof. The condition (i) is a restatement of the sufficient condition

(3.15), while (ii) and (iii) are a direct consequence of (i), in view of the'
fact that pO :::; II . II for any induced matrix norm II . II. To establish (iv),
we first invoke the properties of matrix measures, which give rise to

and
-v(jwI) + v(Ao) :::; v( -jwI + Ao) :::; v( -jwI) + v(Ao).
We claim that v(jwI) = v( -jwI) = 0, and hence v( -jwI + Ao) = I/(Ao).
Indeed, by definition, we have

. I)
v (JW = 1m
1· III + a"jwIIi - 1
= 1·1m 11 + j(Jwl- 1 =
0
.
<1-40+ (J <1-40+ (J
Similarly, 1/( -jwI) = O. We are thus led to the inequality

Since by (ii) the system is stable whenever

1
IIA111 < II(jwI - Ao)-lll'
the condition (iv) follows. Finally, to prove (v), we use the Laplace trans-
form
98 3. Systems with Incommensurate Delays

For any s E C+, it follows that

This completes the proof. •

We note that of all the conditions in Theorem 3.13, the conservatism rises
in ascending order, with (i) the least conservative and (iv-v) the most. A
payoff, however, is that the conditions (iii), (iv) , and (v) do not require
frequency sweep. In particular, the Hoo norm of (sl - AO)-l Al can be
computed using a bisection method, which amounts to checking whether
the Hamiltonian matrix

has an eigenvalue on the imaginary axis, including the origin. The relation
between the Hoo norm and the Hamiltonian matrix H is well known, and
the bisection method is well documented in the Hoo control literature. The
connection between the Hoo norm and other time-domain measures will be
further explored in Chapter 8.
The next result also draws upon frequency-sweeping tests, which give
conditions based on the solution of the Lyapunov equation

A6P + PAo = -21. (3.31 )

Theorem 3.14 Suppose that Ao is stable, and let P be the unique positive
definite solution of {3.31}. Then the system {3.2} is stable independent of
delay if one of the following conditions holds.

{i} O'(PAd < 1.

{ii}
(3.32)

Proof. The existence and uniqueness of the positive definite solution

P of (3.31) is guaranteed by the fact that Ao is stable. Let <1>( s) := (sl -
AO)-I. Then, by adding and subtracting jwP on the left hand side of
(3.31) followed by pre- and post-multiplication of the resulting equation by
<l>H (jw) and <I>(jw), it follows that

<l>H (jw)P + P<I>(jw) = 2<1>H (jw)<I>(jw).

This in turn leads to

(<I>(jW)Al)H (PAl) + (PAJ)H (<I>(jW)Al) = 2 (<I>(jW)Al)H (<I>(jw)A I ).

(3.33)
 3.5 Sufficient stability conditions 99

Therefore, by applying the triangle inequality, we have

a (<I>(jw) Ad ~ a(PAd·

The condition (i) then follows from Theorem 3.13 (iii). To establish (ii),
let x E en be the eigenvector of <I>(jW)AI associated with the eigenvalue >.
such that 1>'1 = p (<I>(jw)A I ). Then, pre- and post-multiplication of (3.33)
by x H and x leads to

X
H ('XPAI + 2('XPAI)H) x = 1\12 H
A X X.

Since

X
H ('XP Al + ('XP At) H) X <
2 -

we have

However,

p CPA,+ fPA1)H) < p C~PA'+ ~~PAIlHI)

< p (,'XIIPAII +21>'IIPAIIT)

1>'lp (,PAl IT : IPAII).

This gives
p(<I>(jw)At} = 1>'1 ~ p (,PAti T 2+ IPAII).

The condition (ii) then follows from Theorem 3.13 (i) . •

The conditions in Theorem 3.14 are a consequence of Theorem 3.13 (i),
(ii), and (iii), but may improve the other conditions. For example, it is
straightforward to show that

(3.34)

indicating that Theorem 3.14 (i) is less conservative than Theorem 3.13 (iv),
when the norm in question is II . 112. Indeed, let x E ]Rn be the eigenvector
100 3. Systems with Incommensurate Delays

of P associated with the largest eigenvalue Amax(P) = u(P). Pre- and

post-multiplying (3.31) by x T and x yields

_ T (Ai' + Ao) _ xT X
x 2 x - u(P)"
Since

x T (Ai' +
2 Ao) x <
_ \
Amax
(Ai' +
2 Ao) x T x -_ V2 (A) T
0 xx,

the inequality (3.34) follows. Thus, whenever Theorem 3.13 (iv) is true,
so is Theorem 3.14 (i). The example below demonstrates the merits and
weaknesses of these sufficient conditions.

Example 3.3 The following delay system is modified after Example 2.8:

x(t) = Ao x(t) + .8A1 x(t - r),

where

AO=(~ ~ ~ ~ ),
-0.05 0.005 0.25 0)
A _ ( 0.005 0.005 0 0
1- 0 0 0 0 '
-2 -3 -5 -2 -1 0 -0.5 0

and .8 E R is a parameter whose range is to be determined so that system

is stable independent of delay. It is known from Example 2.8 that Ao is
stable. Figure 3.6 gives the plot of p ((jwI - Ao)-l A 1 ), which allows the
calculation of the maximal permissible range of .8,

/3 := sup {1.81: 1.8lp ((jwI - AO)-l AI) < I} = 1.2196.

w;:::O

Figure 3.7 plots the bound lI(jwI - AO)-l AlII, for Holder f1 f2, and foo
induced matrix norms, respectively. The reader may compare these plots
to assess the conservatism of Theorem 3.13 (ii-iii), with the three different
norms in use. A simple calculation shows that in all three cases, v(Ao) > O.
Thus, Theorem 3.13 (iv) ceases to be useful. The estimates of/3 based upon
other sufficient conditions are summarized in the following table. 0

p(.) II ·111 11·112 11·1100 u(PA 1 ) p (IPAIIJ2+IPA11)

.8 1.2196 0.5279 0.7983 0.7453 0.256 0.3357

Example 3.4 Consider the scalar differential-difference equation

n-1 n-1
y(n)(t) + L aiy(i)(t) + L biy(i)(t - r) = 0,
i=O i=O
 3.5 Sufficient stability conditions 101

0.9

0.8

0.7

0.6

~0.5
.a
'6,
~04

0.3

02

O.t

0
0 0.5 t.5 2.5 3.5 4.5
Frequency

FIGURE 3.6. The spectral radius p((jwI - Ao)-l Ad

_H,I
__ H,
t8 .... H.

t.6

II>
'Ot.2
.a
'2
Cl
~t

0.4

0.2L...-.....- ......- _ - ' - _......_ _"-_.....t.._ _L - _....._ ......_---I

o 05 t.5 2.5 3.5 4.5
Frequency

FIGURE 3.7. The norms II(jwI - Ao)-l At/I

102 3. Systems with Incommensurate Delays

which, when posed in the form of (3.2), corresponds to

o
-bn - I
)
The conditions in Theorem 3.13 and Theorem 3.14 can be further simplified
for this class of systems. The following table lists the relevant quantities,

Al cb'l ' B
p(<P(S)AI) IbT<p(s)CI 1 bn_ls n +"+bo 'I
Sn+an_IS n I+".+ao
a(<p(s)At} 11<p( s)c11211b112 Isn+a n ISn
')'{s)
I+"'+aol
IIAIIIt m~ Ibil E ICil m~lbil
• i •
a(PA I ) IIPcI1211bl12 {3
p (IPAII'l~+IPAII) IW'IPcl+IIPcIl21IbI12 a+{3
2 2
n-I
where <p(s) := (s1 - A)-I, P = [Pij], a:= E IbiPinl, and
i=O

{3 .-
.- (~lbiI2) (~IPinI2) ,

')'(s) := (~lsI2i) (~lbiI2) .

The second column in the table corresponds to a general rank-one matrix
Al = cbT , and the third to the specific Al = B; correspondingly, B = cbT
with

b = (-b o -b i

It is useful to observe that when Al is a rank-one matrix, the Hex> norm

of the scalar transfer function bT<p(s)c actually gives the exact delay-
independent stability condition. 0
 3.5 Sufficient stability conditions 103

3.5.2 Systems of multiple delays

It is straightforward to extend the above sufficient conditions to systems
with multiple incommensurate delays. The following theorem provides the
counterparts to Theorem 3.13 and Theorem 3.14.

Theorem 3.15 Let rk 20, k = 1, 2, ... , m be independent, incommen-

surate delays. Suppose that Ao is stable. Then the system (3.1) is stable
independent of delay if one of the following conditions holds.

(i) ILx,,.(M(jw)) < 1, Vw 2 0.

(ii) For any absolute or unitarily invariant induced matrix norm II . II,
IIM(jw)11 < 1, Vw 2 0.

(iii) II(sI - AO)-I ( Al Am )11= < 1/01i.

(iv) 0' ([ Al Am ]) < -v2(Ao)/01i.
(v) If for some ( 2 1 and "1 > 0, IIe Aot l12 :S (e-1]t, Vt 2 0,

Am ]) < "1/( 01i ().

(vi) 0' (p [ Al Am ]) < 1/01i.
(vii)

p ( IRIT 2+ IRI) <.

1

Here M(s) is given by (3.13), P is the solution to (3.31), and

Proof. The condition (i) is a direct consequence of Theorem 3.5. To

prove (ii), it suffices to invoke Lemma 3.4 (v), and to note that for any
absolute or unitarily invariant norm,

p (M(jw)U) :S IIM(jw)UII = IIM(jw)11

holds for any U E Um . The proofs for (iii-vi) are similar to those found in
the proofs for Theorem 3.13 and Theorem 3.14, and hence are omitted. To
establish (vii), we first show that

MH (jw)R + RT M(jw) = 2MH (jw)M(jw).

104 3. Systems with Incommensurate Delays

This can be done in the same manner as in the proof for Theorem 3.14 (ii).
Accordingly, it follows that for any U E Um,

(M(jW)U)H (RU) + (RU)H (M(jw)U) = 2(M(jw)U)H (M(jw)U).

Let U be so chosen that f.lx m (M(jw)) = p(M(jw)U). Moreover, let x be
the eigenvector of M(jw)U corresponding to the eigenvalue>. such that
1>'1 = p(M(jw)U). Then, as in the prooffor Theorem 3.14 (ii), we have

X
H (>.RU + 2(5:. RU t) x= I>. 12 x Hx

and
1.11',; p CRU + fRU)H) .
Since lUI = [IUij 11 = I and

p CRU + f>JlU)H) < p C.lIIRIIUI ~1.lIIRITIUI)

1>'lp (IRI ~IRIT),

the result follows. •

Much of Theorem 3.15 is built upon the somewhat crude bound J-L~ (-) $
0'(.), which is valid for any block structure ~. Clearly, the D-scaled upper
bound for f.lx m (.) can be used to further tighten the conditions. For this
purpose, it is of interest to consider positive diagonal scaling matrices in
the set
v!:= {diag( d11kll ... , dm1km): d k > o}.
Let d := Jf 4·
k=l
It is straightforward to verify that for any D E V!,

The following conditions are then clear. Note that by setting dk = 1/ y'm,
k = 1, ... , m, these conditions replicate some of those given in Theorem
3.15.

Corollary 3.16 Let rk ?: 0, k = 1, 2, ... , m be independent, incommen-

surate delays. Suppose that Ao is stable. Then the system (3.1) is stable
independent of delay provided that

Vw?: o.
 3.5 Sufficient stability conditions 105

Furthermore, it is stable independent of delay if there exist d k > 0, k =

m
1, "', m, l: d~ = 1, such that one of the following conditions holds.
k=l

{i} II{sI - AO)-l ({l/dd A I ... (l/dm )Am) 11.:>0 < 1.

{ii} 0' ([ (l/d l )A I ... (l/dm )Am]) < -v2(Ao)·
{iii} If for some ( :2: 1 and TJ > 0, lIeAot l12 ~ (e-7)t, Yt:2: 0,

0' ([ (1/d1 )A I ... (l/d m )Am]) < TJ/(.

{ivy 0' (P ((l/ddAI ... (l/dm )Am)) < 1.

Here P is the solution to {3.31}.

In light of Theorem 3.9, simple sufficient conditions can also be obtained

for delay-dependent stability. We note that similarity scaling can be em-
ployed as well to improve further the following conditions. This is straight-
forward and thus is not pursued herein.

Theorem 3.17 Let rk :2: 0, k = 1, 2, "', m be independent, incommen-

m
surate delays. Suppose that l: Ak
is stable. Then the system {3.1} is stable
k=O
for all rk E [0, Tk), k = 1, "', m, if one of the following conditions holds.

{i} For any absolute or unitarily invariant induced matrix norm II . II,
/lM(jw; T)II < 1, Yw :2: O.

{iii} 0' ([ TIAIG ... TmAmG]) < -V2 (~Ak) /vm(m + 1).

(iv) If for ,orne ( 20 1 and ~ > 0, 0' [exp (t. A,t)1 (e~"', Vt
<; 20 0,
106 3. Systems with Incommensurate Delays

(vi)

Here M(s; r) and C are given as in Theorem 3.9, P is the solution to

the Lyapunov equation

and

The proof of this result resembles that for Theorem 3.15 and hence is
omitted.

3.6 Neutral delay systems

Neutral delay systems constitute a more general class than those of the
retarded type. An LTI neutral delay system with pointwise delays can be
generally described by the differential-difference equation
m m

rk ::::: O. (3.35)
k=l k=l

We note that the two sides of (3.35) need not contain an equal number of de-
lays, but this can nevertheless be assumed with no loss of generality. More-
over, as in retarded systems, the delays can be commensurate or incom-
mensurate. For neutral systems of commensurate delays, the differential-
difference equation is specified to
m m

x(t) - I: Bk x(t - kT) = Ao x(t) + I: Ak x(t - kT), T ::::: O. (3.36)

k=l k=l

We show in this section how the preceding results may be extended to neu-
tral delay systems, with both commensurate and incommensurate delays.
Stability of neutral delay systems proves to be a more complex issue
because the system involves the derivative of the delayed state. In fact, the
stability notion itself needs to be modified in the first place. Define the
 3.6 Neutral delay systems 107

characteristic function by

p (s; e- r1S , "', e-r",S) := det (s (I - t

k=l
Bk e- rkS ) - t
k=O
Ake- rkS ) .

(3.37)
Based on Theorem 1.6 in Chapter 1, we give the following stability defini-
tion.

Definition 3.2 The chamcteristic function (3.37) is said to be stable if

there exists an a < 0 such that
(3.38)
where C a + := {s: Re( s) > a}. It is said to be stable independent of delay
if (3.38) holds for all rk 2: 0, k = 1, "', m. The neutml system (3.35) is
said to be stable if its chamcteristic function (3.37) is stable, and is stable
independent of delay if its chamcteristic function is stable independent of
delay.

As pointed out in Chapter 1, the stability of neutral delay systems re-

quires a stronger condition, which manifests itself through (3.38). This
stronger condition is needed to ensure that the characteristic roots of a
neutral system exhibit the same continuity property as that of retarded
systems. More specifically, while the stability exponent of a retarded sys-
tem is continuous with respect to all Tk 2: 0, k = 1,2, ... ,m, for a neutral
system this continuity may not hold in general at rl = ... = Tm = 0,
though it is continuous at other delay values. The condition (3.38) will
guarantee that in the neutral case the stability exponent is also continuous
at this point, and thus for all rk 2: 0, k = 1,2,··· , m. We note, however,
that in the retarded case, the condition (3.38) reduces to the same stability
definition for the quasipolynomial of a retarded system.
One requirement to ensure stability of neutral systems concerns the dif-
ference equation
m

x(t) - LBk x(t - rk) = 0, (3.39)

k=l

For the system (3.35) to be stable, it is necessary that the difference equa-
tion (3.39) be stable, whereas the stability is defined in the same sense as
in Definition 3.2.

Definition 3.3 The difference equation (3.39) is said to be stable if there

exists an a < 0 such that

(3.40)
108 3. Systems with Incommensurate Delays

An important fact states that the stability of (3.39) is a global property

with respect to Tk, k = 1,2, ... ,mj it will be stable independent of delay
whenever it is stable for some delay values. In the remainder of this section,
when referring to stability, we mean the stability in the sense of Definitions
3.2 and 3.3.
With the above necessary modifications, we now proceed to derive sta-
bility conditions. We shall first consider the system (3.36), i.e., the class
of neutral systems with commensurate delays. The following result gives a
simple necessary and sufficient condition for the stability of the difference
equation (3.39) in this case.

°
Lemma 3.18 Let rk = kT, k = 1, "', m. Then the system (3.39) is
stable for all T ~ if and only if p(B) < 1, where
Bm - 1

°
I

Proof. Using Lemma 2.4 and (2.30), we obtain

det(I - fBke-hS) = (-1t m det (I - Be- TS ) .

°
k=l

Hence, if p(B) < 1, we may find an 0: < such that (3.40) holds. Otherwise,
assume that p(B) ~ 1. Then, B has an eigenvalue >..(B) = p(B)e j8 for
some B E [0, 211"]. Consequently, for any T > 0, one may find s = 0: +
jw E C+, where 0: = (logp(B)IT) ~
det (l - Be- TS ) = 0, and hence
°
and w = (BIT) ~ 0, such that

det (I -f
k=l
Bke- hS ) = 0.

Therefore, the system (3.39) is not stable. This completes the proof. _
In much the same manner, we may then obtain similar delay-dependent
and delay-independent stability conditions. With no loss of generality, we
shall assume that the system (3.36) is stable at T = 0. This enforces the
condition that

det(I - ~B}' 0
The following results resemble the frequency-sweeping criteria developed
for retarded systems, presented in Theorems 2.5 and 2.6, respectively. The
proofs are also similar and are left to the reader. We remind the reader of
the definition of delay margin, which can be generalized to neutral delay
systems in the same manner.
 3.6 Neutral delay systems 109

Theorem 3.19 Suppose that the system (3.36) is stable at T = 0, and that
p(B) < 1. Let rank (Am) = q. Furthermore, define

if Ad G(jw~), H(jw~)) = e- j9l

for some w~ E (0, 00), o~ E [0, 211"1
if e.(G(jw), H(jw» > 1, Vw E (0, 00)

where
1

G(s) :=

H(s) :=
( -(SI~~) °
Fl(S)
diag (1, ... , 1, - Fm(s» ,
Fk(S) := k=l, ... , m.

Then,
T:= min Ti.
l~i~q+n(m-l)

System (3.36) is stable for all T E [0, T), but becomes unstable at T = T.

Theorem 3.20 System (3.36) is stable independent of delay if and only if

(i) p(B) < 1,

(ii) Ao is stable,

(iii) (1 - f k=l
Bk) -1 f
k=O
Ak is stable, and

(iv)
p(Mm(jw)) < 1, Vw > 0, (3.41)
where

~(s).rl(S)
Mm(s) := ( .

with ~(s) := (s1 -

°
Ao)-l.
1

Consider next systems described by the neutral differential-difference

equation
q q
+L +L L akiy(i)(t -
n-l
y(n)(t) bky(n)(t - kT) kT) = 0. (3.42)
k=l i=O k=O
110 3. Systems with Incommensurate Delays

Ai> with retarded systems, stability conditions in this case can be simplified.
Define the polynomials
q
b(z) := Lbkz k
k=l
bks n + ak(s), k = 1, ... , q,

where ao(s) and ak(s), k = 1, ... , q, are defined as in (3.18) and (3.19).
Based on Theorems 3.19 and 3.20, it is not difficult to extend Corollaries
2.7 and 2.8 to the neutral system (3.42).

Corollary 3.21 System (3.42) is stable independent of delay if and only

if

(i) b(z) is Schur stable; that is, b(z) has all its zeros in the open unit
disc;

(ii) ao(s) is stable;

q
(iii) ao (s) + 2: Pk (s) is stable; and
k=l
(iv) p(Ma(jw)) < 1, Vw > 0,

where
_E.!..W. ...
(
ao(s)
1 ... -~
ao(s)
o -~)
ao(s)
0
Ma(s) := b ·· ..
· .
1 0

Corollary 3.22 Suppose that the system (3.42) is stable at T = 0, and

that b(z) is Schur stable. Furthermore, define

if Ai (Ga(jwU, H a (jw1)) = e-j8~

for some w1 E (0, 00), e~ E [0, 27r]
e.
if (Ga(jw), Ha(jw)) > 1, Vw E (0, 00)

where

1 0

Ga(s) :=

(-aL 0
-Pl(S)
1
-Pq-l(S)
).
Ha(s) := diag (1, ... , 1, Pq(s)).
 3.6 Neutral delay systems 111

Then,

System {3.42} is stable for all 7 E [0, r), but becomes unstable at 7 = r.

Example 3.5 A scalar neutral delay system is given as

y(t) + 0.8Y(t - 7) - (3 y(t - 27) + 2y(t) + y(t - 7) = O.

In light of Lemma 3.18 and Corollary 3.21,

B = (-~.8 g)
and
k)
o .
s+2

For (3 = 0.2625, p(B) = 1.05. Under this circumstance the system is unsta-
ble for any 7 2': O. For (3 = 0.09, p(B) = 0.9. Figure 3.8 plots the spectral
radius p(Ma(jw)), which enables us to conclude that the system is stable
independent of delay. 0

0.75

0.65

0.6

0.55

0.5'--~-~-~----'-~---'--~~-~---.J
o w g • ~ ~ ~ ~ • w
Frequency

FIGURE 3.8. The spectral radius p(Ma(jw))

Finally, extensions may also be pursued for neutral systems with in-
commensurate delays. In particular, by mimicking the proof of Theorem
3.5, we obtain the following necessary and sufficient condition for delay-
independent stability.
112 3. Systems with Incommensurate Delays

Theorem 3.23 Let rk ;::: 0, k = 1, 2, "', m be independent, incommen-

surate delays. Then the system {3.35} is stable independent of delay if and
only if

{i} /-Lx", (B) < 1,

{ii} Ao is stable,

{iii} (1 - ~ k=l
Bk) -1 ~ Ak is stable,
k=O
and

(ivy /-Lx", (M(jw)) < 1, Vw > 0,

where

B .-
~·-

M(s) :=

Likewise, in light of Corollary 3.8 and its proof, it is straightforward to

find a simplified condition for scalar differential-difference equations with
incommensurate delays.

Corollary 3.24 Let rk ;::: 0, k = 1, 2, "', m be independent, incom-

mensurate delays. Then the scalar differential-difference equation
q n-1 q
y(n)(t) + Lbky(n)(t - rk) +L Lakiy(i)(t - rk) = 0,
k=l i=O k=O
is stable independent of delay if and only if

{ii} ao (s) is stable,

q

{iii} ao(s) + LPk(S) is stable, and

k=l
{ivY

Vw > 0.
 3.7 Summary 113

In conclusion, whether for computing the delay margin or to determine

delay-independent stability, the frequency-sweeping conditions obtained
herein bear no essential difference from their counterparts for retarded sys-
tems. Nevertheless, we point out that while it is possible to extend the
preceding constant matrix tests also to neutral systems, such extensions
will be more complex and computationally more intensive. We leave these
extensions to the reader.

3.7 Summary
Our development of the frequency domain tests in this and the preceding
chapter was heavily influenced by robust control theory. Whether for sys-
tems with commensurate delays or those with incommensurate delays, the
stability problem has been interpreted as one of robust stability. Indeed,
from a conceptual standpoint, delay-independent and delay-dependent sta-
bility can both be held as notions of robust stability, whereas the delay
values are uncertain within the specified intervals. The application of the
continuity property of the stability exponent is reminiscent of the well-
known zero exclusion principle, which has played a critical role in robust
stability analysis and will continue to be an enabling tool in Chapter 4 in
our study of robust stability of time-delay systems. The use of the struc-
tured singular values as stability criteria, as well as the use of the spectral
radius for systems of commensurate delays, thus come more as a necessary
outcome than a sheer coincidence. More broadly, both can be interpreted
as generalized notions of gain, and the stability criteria as small gain con-
ditions. Small gain type results can be suitably applied to time-varying
systems; this has been demonstrated in the present chapter using examples
involving time-varying delays and will be further discussed in Chapter 8.
They can also be extended to address uncertain time-delay systems and
used as a synthesis tool in control design for time-delay systems.
While it appears justified to contend that the stability problem is largely
resolved in the case of commensurate delays, the problem remains open
for systems with incommensurate delays. The proof of NP-hardness in this
case leads to, on a firm theoretical basis, the somewhat pessimistic con-
clusion that it is unlikely to be tractable in general. One should therefore
resort to computable, albeit approximate, stability tests. The structured
singular values provide yet again relatively less conservative sufficient con-
ditions, by means of its bounds. Various relaxations and derivatives, with
varying degrees of conservatism versus computational ease, can be obtained
from the bounds, among which the least conservative is the D-scaled up-
per bound. This bound can be computed efficiently using commercial soft-
ware programs and yet still provide an acceptable answer. Amid the gener-
ally pessimistic prospect, on the other hand, it should be understood that
114 3. Systems with Incommensurate Delays

the NP-hardness is a worst-case, extreme measure of complexity. In other

words, while the stability problem may be intractable in the worst case,
the NP-hardness does not rule out the possibility that certain more spe-
cialized cases may still be benign enough to yield computable solutions.
The rank-one case provides one such example.
In summary, frequency domain stability analysis of time-delay systems
is a natural progression of classical stability results, both of which are built
upon the fact that an LTI system, with or without delay, can be fully
represented by its frequency domain description, and the system's stabil-
ity be determined by the location of its characteristic roots. A frequency
domain approach thus opens the door for many well-developed frequency
domain tools. The continuity property of characteristic functions proves
fundamental and indeed constitutes the cornerstone of all frequency do-
main methods, both for systems with commensurate delays and for those
with incommensurate delays. In both cases, it reduces effectively the delay-
independent and delay-dependent stability problems into one of computing
the zero-crossing frequencies of the characteristic function. The key thus
lies in whether and how the zero-crossing frequencies can be computed.

3.8 Notes
3.8.1 Small gain theorem and J-l
Robust control is an immensely rich subject of deep mathematical sophisti-
cation. There have been many well-written texts on the subject. The recent
survey article by Chen and Tits [36] provides a concise summary of key
progress in robust stability analysis. The structured singular value is cred-
ited to Doyle [57] and Safonov [238] and has found widespread applications
in robust control design. An in-depth study of this notion can be found in
Packard and Doyle [219], and Zhou, Doyle, and Glover [294]. A generalized
J.L with both complex- and real-valued uncertainties, suggested by Fan, Tits,
and Doyle [62], can be useful in studying robust stability of time-delay sys-
tems. The small gain theorem is credited to Zames [290, 291], whose more
recent treatise can be found in the texts by Desoer and Vidyasagar [54],
Zhou, Doyle, and Glover [294], and Chen and Gu [33]. The same references
contain detailed discussions on norms and spaces relevant to the control
context.

3.8.2 Stability of systems with incommensumte delays

The stability of systems with incommensurate delays is the main subject
of Datko [49], Cooke and Ferreira [42], and Hale, Infante, and Tsen [102],
 3.8 Notes 115

which all studied retarded as well as neutral delay systems. In fact, it was
due to [49] and [42] that the continuity property was firmly established.
Hale, Infante, and Tsen [102] investigated in meticulous detail the stability
problem with the general state-space formulation and gave the sweeping
conditions for scalar differential-difference equations. The approach via the
structured singular value, and the suggestion that the stability problem
be treated as one of robust stability, were initiated by Chen and Latch-
man [35], which led to the f-L characterization of conditions for stability
independent of delay. The idea was later extended in development of delay-
dependent conditions by Niculescu and Chen [208] and Huang and Zhou
[119]. The results in Section 3.3 are mainly based on [35, 208] and [208].
Chen and Latchman [35] and Huang and Zhou [118, 119] also addressed
robust stability and stabilization of delay systems in this framework.

3.8.3 Complexity issues

Computational complexity theory is a routine course of study in computer
science and operation research. Among numerous texts on the subject,
Davis and Weyuker [52] offers a thorough classical treatment tuned to a
typical theoretical computer science readership, while Blum et al. [14] fo-
cuses on the development of the theory in numerical analysis. Narratives of
the theory aimed at a control readership are found in Vidyasagar [279] and
Chen and Gu [33]. The exposure in Subsection 3.4.1 is adapted from [33].
In the recent years, complexity studies have had a significant impact on
control theoretical research, notably on problems found in robust stability
analysis. A large number of control problems, both reasonably formulated
and practically meaningful, have been found to be NP-hard. Pertinent cases
include the computation of the so-called real, complex, and mixed real and
complex f-L [23, 262]' the computation of bounds for f-L [261, 69], and cer-
tainly the stability problem for systems with incommensurate delays [263].
The NP-hardness proof given in Subsection 3.4.2 is essentially based on
Toker and Ozbay [263].

3.8.4 Sufficient conditions and neutral systems

Awide variety of sufficient conditions exists for stability independent of de-
lay. Some of the results most pertinent to the present chapter, to name a
few, are obtained by Mori and co-workers [199, 194, 195, 196], Noldus [215],
Wang, Chen, and Lin [283], Hmamed [112, 113], Bourles [19], Luo and van
den Bosch [183], Zhou [293], Gu and Lee [80], and Verriest, Fan, and Kull-
starn [274]. Among these results we note in particular the 1iCXl -norm, small
gain type conditions in [293] and [80]. Section 3.5, however, is largely based
on Chen, Xu, and Shafai [37], which expands on the necessary and sufficient
small gain conditions developed in [35], and provides a unified treatment
116 3. Systems with Incommensurate Delays

for many of the aforementioned sufficient conditions. Many additional suffi-

cient, delay-independent and delay-dependent conditions have been devel-
oped in time domain, based on an Lyapunov-Krasovskii-Razumikhin ap-
proach. These conditions are typically posed as solutions to LMI problems,
a subject to be undertaken in the subsequent chapters.
While retarded delay systems have been extensively studied, fewer results
are available for neutral systems, though several of the classical tests, such
as [111] and [282], remain applicable. Section 3.6 is adapted from Chen [32],
and draws additionally upon [102]. For more discussions on neutral delay
systems and additional references, we refer to Niculescu [204].
4

Robust Stability Analysis

4.1 Uncertain systems
In the previous chapters the stability property of a given time-delay system
was studied. In engineering applications, it is now very common that one
does not know exactly the system under investigation; that is, the system
contains some elements (blocks) that are uncertain. Usually it is known that
these uncertain elements belong to some specific admissible domains, which
in turn depend on the nature of the elements and also on the information
available about the system. In other words, it is known only that the system
belongs to the family of systems that arises when the uncertain elements
(blocks) range over the admissible domains and therefore one may treat the
family as a new object for analysis. This family is referred to as an uncertain
system. When it is possible to show that all systems of the family are stable,
the stability of the original system that is a particular member of the family
is guaranteed.
In this chapter several basic results on the stability of uncertain time-
delay systems will be presented. We start with a linear time-invariant delay
system whose description contains parametric uncertainty.

4.2 Characteristic quasipolynomial

It has already been mentioned that a linear time-invariant system with
concentrated delays
N
L [AkX(t - Tk) + BkX(t - Tk)] = 0, det(Ao) =J 0, (4.1)
k=O
°
where = TO < T1 < ... < TN, is exponentially stable if and only if there
exists a positive E such that the real part of every zero of the characteristic
function,

(4.2)

of the system is less than -E. In this case f(s) is called Hurwitz stable. In
(4.2) coefficients aki are real or complex numbers and exponent coefficients
118 4. Robust Stability Analysis

are real and ordered as follows 0 = rO < rl < ... < rm. Function /(8) is
known as a qua8ipolynomial and may be written in different forms
m n
/(8) = LPi(8)e- riS
=L 1/Jk(8)8 n -k, (4.3)
i=O

where polynomials

and quasipolynomial

k = 0, 1, ... ,n.
Example 4.1 Let us consider the scalar equation

x(t) + x(t - r) + x(t) = O. (4.4)

Its characteristic quasipolynomial is

/(8) = 8 + 8e- rs + 1.
All zeros of the quasipolynomial have negative real part. To check this we
assume first that it has a zero 80 = a + j/3, with positive real part, a > O.
Then the real and the imaginary parts of /(80) are equal to zero

{ a[l + e- ro cos(r/3)] + /3e- r o: sin(r~) + 1 = 0,

/3[1 + e- ro cos(r/3)]- ae- ro sm(r/3) = O.

From these equations we conclude that

a
1 + e- ro cos(r/3) =- 2'
a2 + /3
Our assumption, a > 0, implies that the left hand side of the last equality is
positive while the right hand side is negative. Therefore /(8) has no zeros
with positive real part. Assume now that 8 = j/3, then /3 i= 0, because
/(0) = 1 i= O. In this case

/3sin(r/3) + 1 = 0, (4.5)
/3[1 + cos(r/3)] O. (4.6)

From (4.6), we may conclude that

cos(r/3) = -1.
This implies that sin(r/3) = 0, which contradicts (4.5). In other words /(8)
has no zeros in the closed right half-plane of the complex plane. Therefore
 4.3 Zeros of a quasipolynomial 119

all zeros of the quasipolynomial have negative real part. On the other hand
the equation (4.4) is not exponentially stable because
sup {Res I f(s) = O} = O.
This statement can be easily checked if we observe that if So is a zero of
f(s) then

Therefore, as So -+ 00, it becomes asymptotically close to the imaginary

axis. 0

The above example shows that for a neutral time-delay system to be

exponentially stable, it is not enough to check that all zeros of its charac-
teristic function lie on the open right half complex plane.

Remark 4.1 There are also examples of neutml systems where all the ze-
ros are on the open left half complex plane, but the system is not stable in
the sense of Lyapunov in addition to not exponentially stable.

4.3 Zeros of a quasipolynomial

In this subsection several important results about zeros of a quasipolyno-
mial function will be presented.
In order to make our analysis a little bit more general we consider a
quasipolynomial
n m
f(s) = L L(aki + jbki )sn-k e(Oi+if3i)S, (4.7)
k=Oi=O

where j is the imaginary unit and aki, bki , ai, /3i are real numbers. This
function may be written in two forms
m n
f(s) = LPi(s)e(Oi+if3i)8 = L 1Pk(S)sn-k,
i=O k=O
where
n
Pi(S) = L(aki + jbki)Sn-k, i = 0,1, ... , m,
k=O
and
m
1Pk(s) = L(aki + jbkde(Oi+if3i)S, k = 0, 1, ... , n.
i=O
In the following we assume that the
120 4. Robust Stability Analysis

• exponent coefficients (ai + j{3i), i = 0,1, ... ,m, are distinct complex
numbers;
• polynomials Pi (8), i = 0, 1, ... , m, are not trivial.

Under these conditions f( 8) may have a finite number of zeros only in the
case when m = 0. In the following we assume that m > 0, if not explicitly
stated otherwise.

4.3.1 Exponential diagram

Let us mark on the complex plane points that are complex conjugate num-
bers to the exponent coefficients of f (8 ), (ai - j (3i), i = 0, 1, ... , m, and form
then the convex hull of these points.
The convex hull is in general a convex polygon whose boundary consists
of a finite number of segments (see Figure 4.1).
1m

FIGURE 4.1. Exponential diagram

The boundary polygon of the convex hull is known as exponential diagram

of quasipolynomial f(8). Now with each boundary segment we associate a
ray emanating from the origin and in the direction of the outer normal to
the boundary segment. In the case when the exponential diagram degen-
erates into a segment of the complex plane, we consider the segment as a
two-sided one, and therefore two opposite rays should be associated with
the segment.
Let these rays be
 4.3 Zeros of a quasipolynomial 121

Then a sufficiently small E > 0 may be chosen such that the E-sectors

ri(E) = {s=PeiI"IPE[O,OO)''PE['Pi-~E''Pi+~E]}'
i=1,2, ... ,N

have no common point except the origin.

Theorem 4.1 For every E > 0 there is R(E) > 0, such that all zeros of
f(s) with magnitudes greater than R(E) lie inside one of the E-sectors ri(E),
i=1,2, ... ,N.

Proof. Let us consider a vertex of the exponential diagram; for simplicity

we assume that the vertex is ao - j{3o. The vertex connects two consecutive
boundary segments of the diagram. Let r i, r i+ 1 be the rays associated with
these boundary segments. Let cI? be the pointed sector of the complex plane
between these rays. If the angular size of sector cI? is 'I' ('I' E (0,71']), then
for any E E (0,'1') sectors ri(E) and ri+l(E) have no common points except
s = O. Define by cI?(E) the part of sector cI? which lies between sectors ri(E)
and riH(E). Now we calculate

Direct geometric calculations shows that for any s E cI?( E) we have

Re {[(ai - ao) + j({3i - {30)] s} ::::: -p lsi sin G) , for i = 1,2, ... , m. (4.8)

Now we select R > 0 such that all zeros of all polynomials Pi(S), i =
0,1, ... , m, lie inside the disc

DR = { s I lsi < R } .

Let IsI ~ R then f (s) can be factorized as

f(s) = Po (s)e(ao+i,Bo)s [1 + fi=l

Pi(S) e[(ai-aO)+i(,Bi-,BO)]sj.
Po(s)
(4.9)

If additionally s E cI?(E), then by inequality (4.8)

::~:~ e[(a i -a o )+i(!3i-!30)]S I ::::: I::~ :~ I e- plsl sin( ~),

I for i = 1,2, ... , m,

and we can conclude that

122 4. Robust Stability Analysis

Hence there exists R(c) ~ R such that for all s E <I>(c) with magnitude
greater than R(c) the expression in square brackets in (4.9) is not zero.
The first two factors in (4.9) are also nonzero for such values of s. It means
that f(s) has no zeros in <I>(c) with magnitude greater than R(c). Repeating
these arguments for all other vertex points of the exponential diagram we
arrive to the theorem statement. _

Remark 4.2 A more detailed analysis, based on the argument principle,

shows that each c-sector fi(c) contains an infinite (countable) number of
zeros of f(s).

From the above theorem one can easily derive the following consequence
about zeros of f(s).

Corollary 4.2 If among exponential coefficients (Oi + j!3i), i = 0, 1, ... , m,

there are at least two with distinct imaginary parts, then f(s) has zeros
with arbitmrily large positive real parts.

Proof. It follows from geometric consideration that in this case one of the
rays points toward the open right half complex plane. So the corresponding
c-sector lies entirely in this half plane. According to Remark 4.2, inside the
sector f(s) has an infinite number of zeros. _
This corollary implies that the only case when f(s) may have all zeros
in the open left half complex plane is when all exponent coefficients of the
quasipolynomial have the same imaginary part. In this case, without any
loss of generality one may assume that all exponent coefficients are real
numbers.

4.3.2 Potential diagram

As has been shown in the previous subsection, only a quasipolynomial with
real exponent coefficients may have all zeros in the open left half complex
plane. Let
n m

f(s) = L L(aki + jbki)Sn-keQ;;s. (4.10)

k=Oi=O
Here 00 < 01 < ... < Om. The quasipolynomial may also be written as

= L cv s kv el'v s ,
N
f(s) (4.11)
v=O
where all complex coefficients Cv are supposed to be nonzero, and addi-
tionally we assume that there are no two terms in the sum for which both
kv and Iv are the same; that is, the terms have either different potential
factors or different exponential factors.
 4.3 Zeros of a quasipolynomial 123

Let us superpose the b, k)-plane with the complex plane in such a way
that the ,-axis coincides with the real axis of the complex plane and in
turn the k-axis coincides with the imaginary one. On this joint plane we
mark points bv, kv) corresponding to all terms in (4.11).
Then we construct the upper part of the envelope of these points (see
Figure 4.2).

Im(k)

Re(y)

FIGURE 4.2. Potential diagram

This upper part is known as the potential diagram of f(s), and it consists
of a finite number of segments. Let M be the number of these segments.
With each segment of the diagram we associate the logarithmic curve
AK = {s=x+iylx=ILKln(y), YE[l,oo)},
K=1,2, ... ,M.
Here ILK is a real number such that vector (ILK' 1) is along the direction of
the outer normal to the corresponding segment. If ILK > 0, then AK lies
in the right half complex plane, while the curve corresponding to ILK < 0
belongs to the left half complex plane. For ILk = 0 the corresponding AK
coincides with the imaginary axis.
For sufficiently small E > 0 the logarithmic E-sectors
_ +' I x E [(ILK -
{ s-x !E) In(y), (ILK + !E) In(y)], }
zy YE[l,oo) ,
K=1,2, ... ,M
have no common points except (0,1).
The following theorem describes distribution of zeros of quasipolynomial
(4.11).

Theorem 4.3 For every E > 0 there exists R1(E) > 0, such that all zeros
of f (s) in the upper half complex plane with magnitudes greater than Rl (E)
lie in the union of logarithmic E-sectors AK(E), K = 1,2, ... , M.
124 4. Robust Stability Analysis

Proof. Let us consider an interior vertex of the potential diagram. For

the sake of definiteness let this vertex be ("f 0, k o). There are two neighboring
segments of the diagram which have ho, k o) as the common vertex. Let AI<.
and AI<.+l be the logarithmic curves associated with these segments. Denote
by W the part of the upper half complex plane between the curves. Each
point s of W can be represented as

s = ILln(y) + jy (4.12)

for some IL E (ILk' ILk+1)' and y > 1. Here ILk and ILk+l are coefficients ofthe
corresponding logarithmic curves Ak and Ak +1 (ILk < ILk+l)' If c: is positive
and
< ILk+l - ILk
c: 2'
then logarithmic sectors Ak (c:) Ak+ 1 (c:) have no common points except
(0,1). In this case we denote by w(c:) the part of the upper half com-
plex plane between the sectors. In the set w(c:) the quasipolynomial (4.11)
can be factorized as follows

f(s) = Coskoe'Yos [1 + t Cv sk"-koebv-'YO)sj. (4.13)

v=l Co

Let us first define the value

and the value

Then simple geometric manipulations show that for all IL E [ILk + c:, ILk - c:]
the following inequality holds

(4.14)

Now we can see that along the logarithmic line s = ILln(y) + jy, y ~ 1,
where IL E [ILk + c:, ILk - c:J,

I~ sk,,-koeb,,-'Yo)S I ::; I~: 12Ik,,-ko'y(k,,-kO)+J.'b,,-'Yo).

Taking into account inequality (4.14) we arrive at the conclusion that

The right hand side of the last inequality approaches zero as y --4 00.
Therefore the sum in the square brackets in (4.13) also approaches zero as
 4.4 Uncertain quasipolynomial 125

y -+ 00. In other words, for sufficiently large values of y, function I(s) has
no zeros along any logarithmic curve (4.12) where IL E [ILk -c:, ILk +c:J. This
observation proves the fact that there exists R > 0 such that no zero of I(s)
with magnitude greater than R lies in the domain 'If (c:). These arguments
can be applied for all other interior vertices of the potential diagram.
A slightly modified argument can also be applied for the two extreme
vertices of the diagram. _

Remark 4.3 The zeros oll(s) with large magnitudes in the lower hall
complex plane lie in the union 01 logarithmic c:-sectors obtained by the mir-
ror image 01 AI\: (c:) , K. = 1,2, ... ,M, with respect to the real axis.

Remark 4.4 Applying the principle 01 argument it may be shown that I(s)
has infinite (countable) number 01 zeros in every logarithmic c:-sector AI\:(C:)'

It follows from the theorem that I(s) has zeros with arbitrarily large
positive real parts when at least one of the values ILl\: is positive. Using the
potential diagram of I(s), one can conclude that such positive ILl\: exists
only if the outer normal of one of the segments forming the diagram points
toward the open right half complex plane. In order to guarantee the absence
of zeros of I(s) with arbitrary large positive real parts, one has to assume
that one of the terms in (4.11), for example, Cos ko e"YoB , satisfies the following
two conditions

• ko ~ k j , j = 1,2, ... ,N;

• "10 ~ "Ij, j = 1,2, ... ,N.

When such a term exists it is called the principal term of quasipolynomial

(4.11).

Corollary 4.4 Quasipolynomial (4.11) may have all zeros in the open left
hall complex plane only il it has a principal term.

4.4 Uncertain quasipolynomial

It was shown in the previous section that quasipolynomial (4.2) may have
all zeros in the open left half-plane only if it has a principal term. This
condition may be expressed as

deg(po) ~ deg(Pi), i = 1,2, ... , m. (4.15)

Without any loss of generality one may assume that aoo =I 0, then aoosne- rOB

is the principal term of I(s).

126 4. Robust Stability Analysis

With quasi polynomial (4.2) one may associate two vectors, the coefficient
vector

and the exponent coefficient vector

When it is convenient we will write !(s, a, r) instead of !(s) in order to show

these vectors explicitly. Now every quasipolynomial of the form (4.2) may
be uniquely interpreted as the corresponding point (a, r) of the coefficient
space L. Assume now that there is a family F of quasipolynomials of the
form (4.2). With this family one may associate the corresponding set

QF = { (a,r) I !(s,a,r) E F } C L

in the coefficient space. In turn QF defines uniquely the family

F = { !(s,a,r) I (a,r) E QF }

of quasi polynomials. Because of this one-to-one correspondence we will

sometimes treat QF and F as the same object when it will not produce
any ambiguity. The family F is often referred to as an uncertain quasipoly-
nomial. We say that F is robustly stable if all members of the family are
Hurwitz stable.

4.4.1 Value set

Let us introduce the concept of the value set for an uncertain quasi poly-
nomial F. The value set of F is defined for a given complex number So as
the following set
VF(so) = { !(so) I! E F },

of the complex plane.

4.4-2 Zero exclusion principle

There are several interesting results in the robust stability analysis which
are based on the zero exclusion principle. In some sense this principle serves
as a fundamental statement that allows us, under some mild assumptions
on uncertain quasi polynomials, to obtain a very general robust stability
criterion for F.
The assumptions one needs to formulate the principle are as follows:
 4.4 Uncertain quasipolynomial 127

• AI: Every member of the uncertain quasipolynomial F has a nonzero

principal term. According to our previous discussion this assumption
may be stated as
deg(po) = n ~ deg(pi)' i = 1,2, ... , m,
for all f(s) E F.
• A2: The exponent coefficient vector of every member of the uncertain
quasipolynomial has only positive components, that is, ri > 0, i =
1,2, ... , m, for every f(s) E F.
• A3: There exist R > 0 and c > 0 such that for every f(s) in the
uncertain quasipolynomial, the corresponding quasipolynomial '1f;o(s)
(see decomposition (4.3)) has no zeros of magnitude greater than R
(if any) with real part greater than -c.
• A4: The set QF is compact and pathwise connected.

Remark 4.5 The assumptions Al-A3 are necessary conditions for the
robust stability of F.

Remark 4.6 Assumption A3 holds automatically when all quasipolynomi-

als from F satisfy
deg(po) = n > deg(Pi), i = 1,2, ... ,m.
Now everything is ready to state the zero exclusion principle.

Theorem 4.5 Let the uncertain quasipolynomial F satisfy assumptions

Al-A4. Then all members of F are Hurwitz stable if and only if
• at least one member of F is Hurwitz stable,
• for every point s = jw on the imaginary axis, the value set VF(jw)
computed at this point does not contain the origin of the complex
plane.

Proof. The necessity part is nearly obvious. In fact, the first condition
of the theorem holds because all quasipolynomials in F are Hurwitz stable.
To check the second condition let us assume that VF(jwo) contains the
origin. There exists a quasipolynomial in the family that has jwo as a zero
and therefore the quasipolynomial is not Hurwitz stable. This contradicts
the Hurwitz stability of all members of F.
Now let us address the sufficiency part. It follows directly from assump-
tion A3 that there exists a positive value Ro such that no one member of
F has zeros in the complex half-plane

I
{ s Re( s) ~ - ~c }
128 4. Robust Stability Analysis

with magnitude greater or equal than Ro. According to the second condition
of the theorem there are no quasipolynomials in F that have zeros on the
imaginary axis.
Assume by contradiction that there exists f(s, 80, ro) in F that is not
Hurwitz stable. It means that the quasipolynomial has at least one zero
So with positive real part. By the first condition of theorem, there is
f(s, a1. rl) E F which is Hurwitz stable, that is, all zeros of f(s, a1. rl)
have negative real parts. According to assumption A4 there is a path in
QF:
a = aI" r = r JL , 11- E [0,1) ,
connecting points (ao, ro) and (a 1. rJ). Let us define

11-0 = inf {11- > 0 Isuch that f(s, aI" rJL) is Hurwitz stable}.
It follows directly from the definition of the value that 11-0 E (0,1). Poly-
nomial f(s, a JLo ' r JLo ) cannot be Hurwitz stable; otherwise by a continuity
argument, assumptions A1-A3 imply that there should be a positive II
such that for every 11- E (11-0 -II, 11-0 +11), quasipolynomial f(s, aI" rJL) is also
Hurwitz stable, which contradicts the definition of 11-0' On the other hand,
f(s,aJLo,r JLo ) cannot have a zero with positive real part because otherwise
for 11- a little bit smaller than 11-0, quasipolynomial f(s, aI" rJL) will also have
a zero in the open right half complex plane which contradicts again to the
choice of 11-0' It also cannot have zeros on the imaginary axis by the sec-
ond condition of the theorem. So f (s, a!Joo' r!Joo) cannot be unstable, either.
This contradiction has its origin in our assumption that f(s, 80, ro) is not
a Hurwitz stable quasipolynomial. This conclusion completes the proof of
the sufficiency part of the theorem. _
The second condition of the theorem admits certain relaxation.

Remark 4.7 The second condition of theorem 4.5 is equivalent to the

following two conditions:

• at least for one point, jwo, of the imaginary axis 0 fj. VF(jWo)i

• 0 fj. aVF(jw) for all other points of the imaginary axis. Here aVF(jw)
stands for the boundary of the value set VF (jw).

Proof. The statement follows directly from the property that the value
set changes continuously with respect to w. _
It is a difficult task in some applications to check the second condi-
tion of the theorem because there are no effective algorithms for the exact
construction of the value set. Also in many cases one cannot check this con-
dition for all points on the imaginary axis but only for a finite set of such
points. Having in mind this observation it seems quite natural to look for
 4.5 Edge Theorem 129

particular cases when this condition may be efficiently checked. Certainly

in such cases some additional restrictions should be imposed on the set QF.
One of such particular cases is given by the Edge Theorem.

4.5 Edge Theorem

In this section the class of polytopic families of quasi polynomials will be
studied. For these families there is no uncertainty in the exponent coeffi-
cients, that is, all ri , i = 1,2, ... , m, are fixed numbers. Also, the uncertainty
in the coefficient vector has a poly topic nature. In order to define formally
such a family one needs a finite set

L L a~~} 8n-ke-ris,
n m
f(v}(8) = V = 1,2, ... , N,
k=O i=O

of quasipolynomials, where as before, 0 = TO < T1 < ... < Tm. The poly-
topic family is defined as the convex hull of the quasi polynomials

Quasipolynomials f(v}(8), v = 1,2, ... , N, are called generators of the poly-

topic family. The set of generators is called the minimal one if after exclu-
sion of anyone of the generators, the resulting polytopic family does not
coincide with P. Let the coefficient vector of f(v}(8) be a(v}, v = 1,2, ... , N.
The corresponding set Qp for the family is the direct product of the
polytope

A= {a t =
v=1
I ~ 0, v = 1,2, ... , Nj
JLva(v} JL v t
v=l
JLv = 1} ,

and the point r = (r1, r2, ... , rm):

Qp = A x r.
It is clear that the vertices of A are coefficient vectors of the generator
quasipolynomials. When the set of generators is the minimal one then
every a(v} is a vertex of A. It explains why quasipolynomials f(v}(8),
V = 1,2, ... , N, are called vertex quasipolynomials of P.

It is a necessary condition for the Hurwitz stability of a given quasi poly-

nomial f(8) that it has a principal term. To prevent unnecessary technical
complications associated with a possible "degree dropping," the following
technical condition is assumed to hold.
130 4. Robust Stability Analysis

Condition 4.1 The principal coefficient of every generator f(v)(s) is pos-

itive, a~~) > 0, 1/ = 1,2, ... , N.

This condition guarantees that all members of P have a principal term,

because now for arbitrary f(s) E P the coefficient

N
aOO = "'"
~ (v)
J.Lvaoo > 0.
v=l

Generator f(v)(s) can be written in two different forms as

m n
f(v)(s) = LP~V)(s)e-riS = L 1/J}:)(s)sn-k,
i=O k=O
where
n m
p~v)(s) = Lat)sn-k, and 1/Jkv )(s) = Lat)e- riS .
k=O i=O

Now one may associate with P the auxiliary family

which also has the polytopic structure.

The following condition is needed to guarantee that all members of P
are Hurwitz stable. Actually this condition is a consequence of assumption
A3 discussed in subsection 4.4.2.

Condition 4.2 There exist c > 0 and R > 0 such that members of 1}! have
no zeros in the half-plane

{s IRe( s) 2:: -c }

with magnitude greater than R.

Remark 4.8 Condition 4.2 holds automatically when for all generators
deg(p~v») = n > deg(p~v»), i = 1,2, ... , m.

Remark 4.9 Let rI, r2, ... , rm be such numbers that ri = kir, where all ki'
i = 1,2, ... , m, are natural numbers. Then Condition 4.2 holds if and only
if all polynomials from
 4.5 Edge Theorem 131

are anti-Schur, that is, all zeros of every polynomial from II lie outside the
closed unit disc of the complex plane. Here
m

1f (v) ( Z ) -_ '~
"' (v) Z ki , v -- 1 , 2 , ... , N .
a Oi
i=O

For the rest of this section a key role is played by the edges of A. An
edge is a segment connecting two vertices of A such that the set A without
the segment remains convex. In other words, an edge is a one-dimensional
face of polytope A. Every edge can be written as the convex hull of two
vertices
JLa(a) + (1 - JL)a(f3), where JL E [0, 1J .

One may associate the corresponding edge subfamily of quasipolynomials

with the edge
JLf(a)(s) + (1 - JL)f(f3) (s), where JL E [0, 1J. (4.16)

Let us denote by Ep the union of all edge subfamilies of P. We can state

the following Edge Theorem.

Theorem 4.6 Let family P satisfy Conditions 4.1 and 4.2. All members
of the family are Hurwitz stable if and only if all quasipolynomials of the
edge subfamilies from Ep are Hurwitz stable.

Proof. Necessity is obvious, because all quasipolynomials of the edge

subfamilies are also members of P.
To check sufficiency, we observe first that the value set Vp(so) computed
in an arbitrary point So of the complex plane is the polygon

Vp(SO) = {t, JLvf(v)(so) JLv I ~ 0, v = 1,2, ... , N; t, JL v = 1 } ,

whose boundary segments are composed of images of some edge subfamilies.

That is, for every point z E 8Vp (so) there exists an edge subfamily (4.16)
such that z = JLof(a) (so) + (1 - JLo)f(f3) (so) for some JLo E [O,lJ.
Conditions 4.1 and 4.2 imply that there exists R > such that no one
of the quasipolynomials from P may have zeros in the closed complex half-
°
plane

with magnitude greater than R. Assume by contradiction that there is an

°
unstable quasipolynomial II (s) E P. Then it should have a zero, say Sl,
with Re(sl) 2:: 0. It means that E Vp(Sl)' Observe that

f(s) --+ 00, as Re(s) --+ 00

132 4. Robust Stability Analysis

for every f(s) E P. Hence, there exists an So with sufficiently large real part
such that no member of P has So as a zero; in other words, 0 ~ Vp(so).
Define now s-y = (1 - ,)so + ,SI and compute the value set Vp(s-y). For
, = 0 this set does not contain the origin, while for , = 1 it does. By
continuity arguments there exists '0
E (0,1], such that 0 E aVp(s-yo)'
It is clear that Re(s-yo) 2: O. According to our previous observation there
should be an edge family (4.16) such that 0 = ILof(e» (s-y) + (1- ILo)f({3) (s-y)
for some ILo E [0,1]. In other words, the edge quasipolynomial fo(s) =
ILof(e»(s) + (1 - ILo)f({3)(s) is not stable. This conclusion contradicts the
second condition of the theorem. The contradiction is a direct consequence
of our assumption that family P has at least one unstable quasipolynomial.

Theorem 4.6 reduces the robust stability analysis of a poly topic family to
•
a simpler problem of Hurwitz stability analysis of a finite set of one parame-
ter subfamilies of the type (4.16). Subfamilies (4.16) serve as a testing set
for the robust stability analysis of P. The number of such subfamilies may
in general be very high; the upper bound for this number is C'j. = N(~-I).
Analyzing the proof one can observe that not all edge subfamilies need to
be taken into account in the analysis. We only need to check those edges
whose images appear on the boundary of the value set for at least one point
jw. The number of such subfamilies may be significantly smaller than that
of all edge subfamilies. This observation sometimes allows one to obtain a
simpler testing set. But one question remains open: How can one check the
stability of an edge subfamily?

4.5.1 Stability of an edge subfamily

Here stability analysis of a simple one-parameter family of the form

f/1-(s) = (1- IL)fo(s) + ILh(s), where IL E [0,1]' (4.17)

is treated. This is a simple polytopic family with two generators: fo(s),

h(s). It is assumed that (4.17) satisfies Conditions 4.1 and 4.2.
Our principal goal now is to derive some tractable stability conditions
for the family (4.17).

Claim 4.1 Let generator quasipolynomials fo (s) and h (s) be Hurwitz sta-
ble. Then all quasipolynomials f/1-(s), IL E [0,1] are Hurwitz stable if and
only if the complex curve

fo(jw)
z = -(--)'
h jw wE (-00, +00),

does not touch the negative real semi-axis of the complex plane.
 4.5 Edge Theorem 133

Proof. The statement follows directly from the fact that the instability
of one of the quasipolynomials from (4.17) means that in (4.17) there is a
quasipolynomial with at least one zero on the imaginary axis. •
One interesting question is the following: Under what conditions does the
Hurwitz stability of vertex quasipolynomials fo(s) and h(s) imply that of
f/.L(s) for all J.L E [0, I]? This question leads us to the concept of convex
directions. First we rewrite (4.17) as

f/.L(s) = fo(s) + J.Lg(s), J.L E [0,1]'

where g(s) = h(s) - fo(s).

Definition 4.1 A quasipolynomial

n m
g(s) = LLokisn-ke-riS,
k=Oi=O
is called a convex direction for the set of Hurwitz stable quasipolynomials
of the form (4.2) if for every Hurwitz stable quasipolynomial fO(8) of this
form, Hurwitz stability of fo(s) + g(s) implies that of (1 - J.L)fo(s) + J.Lg(s)
for all J.L E [0,1J.

The next theorem gives necessary and sufficient conditions for g(s) to be
a convex direction.

Theorem 4.7 A quasipolynomial g( s) is a convex direction for the set of

Hurwitz stable quasipolynomials of the form (4.2) if and only if for all w > 0
such that g(jw) -I 0, the following inequality holds

aarg(g(jw)) < _rm ISin(2ar g (g(jw))+r m w)/

ow - 2 + 2u..> •

Example 4.2 Quasipolynomial q(s)e- rS , where r > 0 and q(s) is a poly-

nomial of degree :s n, is a convex direction for the set of Hurwitz stable
quasi polynomials of the form (4.2) if

1. 2r:S rm;

2. for all w > 0 such that q(jw) -:J. 0, the following inequality holds

aarg( q(jw)) < Isin(2 arg(g(jw))) I.

oW - 2w

o
134 4. Robust Stability Analysis

4.5.2 Interval quasipolynomial

In this section a special polytope of quasipolynomials is treated. Namely, we
consider an interval quasipolynomial that is a family of quasipolynomials
of the form

where 0 = ro < rl < r2 < ... < rm'

It can also be written as

where

are interval polynomials. It will be convenient sometimes to associate with

IIo the exponential factor e-ros.
With each interval polynomial IIi one may associate four special vertex
polynomials

p~l)(S)
p~2)(s)

pP)(s)
p~4)(s)

These polynomials are such that the value set VIIj (jw) is a rectangle whose
corner points are p~II)(jw), lJ = 1,2,3,4:

The value set of IIie- ris at the point jw can be obtained by a rotation of
VIIi (jw) about the origin through the angle -riW. This observation allows
us to conclude that the value set V1(jw) of the interval quasipolynomial I
is a sum of such rotated rectangles:

In geometrical terms, the value set is a polygon. The number of corner

points of the polygon cannot exceed 4m + 1 , and all of them are images of
 4.6 Multivariate polynomial approach 135

vertex quasi polynomials of the form

m
L>~v;)(s)e-"'s ,
i=O

where Vi E {I, 2, 3, 4}, i = 0, ... , m. We remind the readers that here TO = O.

On the other hand, at every given frequency, the boundary segments of
the polygon are included in the union of value sets of the following one-
parameter families

LP~v;)(s)e-riS + [JlPk")(s) + (1 - Jl)p~)(s)] e- rks , (4.18)

if.k

where, as before, k = 0, I, ... ,m, Vi E {1,2,3,4}, and

(a,{3) E {(1,2),(2,3), (3,4),(4, I)}.

Simple calculations show that there are exactly (m + 1)4m + I families of
this form.

Theorem 4.8 The interval quasipolynomial I is Hurwitz stable if and only

if

• there are E > 0 and R > 0 such that in the interval family

{i~ aOie-r;s I aOi E [goi,aoi], i = 0, l, ... ,m },

there is no quasipolynomial which has at least one zero with magnitude

greater than R and real part greater than -c;
• all one-parameter families of the fOT77L (4.18) with 2rk > Tm are Hur-
witz stable.

Proof. This follows directly from the Edge Theorem and convex direc-
tion concept. •

4.6 Multivariate polynomial approach

Let us consider the system

Aox(t) + Box(t) + AIX(t - rd + BIX(t - Td = O. (4.19)

The characteristic function of the system

(4.20)
136 4. Robust Stability Analysis

may be written as a polynomial of two variables, sand z = e- T1S :

p(s, z) = det (sAo + Bo + SZA1 + zBd.

It is possible then to use the polynomial for stability analysis of the system
(4.19).

Lemma 4.9 Quasipolynomial (4.20) has no zeros with nonnegative real

parts if p( s, z) has no zeros in the domain

~={ (s,z) 1 Re(s)~O, Izl~l}.

Proof. In fact, if f(s) has a zero, so, for which Re(so) ~ 0, then p(s, z)
has zero (so, zo) = (so, e- T1SO ) E ~ •
It is convenient to introduce the following change of variables to achieve
a more uniform formulation:
1-z
Sl = 8, and 82 = --.
l+z
Let n2 be the partial degree of p(8, z) with respect to z, then polynomial

has no zeros in the domain

if and only if p(8, z) has no zeros in the domain ~.

Having in mind this observation and Lemma 4.9, one may apply known
stability results for multivariate polynomials to the stability analysis of
time-delay systems. Certainly one may consider also the case when the
time-delay system has more than one incommensurate delay. In this case
polynomials with more than two variables will arise.

4.6.1 Multivariate polynomials

A multivariate polynomial is a finite sum of the following form
_ ~ .. . i l i2 im
p ( Sl,82,· .. ,8 m ) - L...,a'I'2''''m 8 1 82 ... Sm· (4.21)

Here 81,82, ... ,8 m are independent variables, and coefficients ail i2 ... i m are
real or complex constants.
Two polynomials are called relatively prime if they have no common
divisors except constant polynomials. Two primes are either coinciding up
to a nonzero constant multiplier, or are relatively prime.
 4.6 Multivariate polynomial approach 137

A vector (S~o), S~o), ... , 8~)) with complex elements is called a zero of
(4. 21) , 1'f P((0) (0)) -- 0 •
8 1 ,8 2(0) , ... , 8 m

In contrast to the case of univariate polynomials, multivariate poly-

nomials have an infinite number of zeros. Let the partial degree nm of
p(Sl, 82, ... , sm) with respect to Sm be positive. Consider the power expan-
sion of the polynomial with respect to this variable

(4.22)

The main coefficient a~':) (S1, S2, ... , sm-d is a nonzero polynomial with
respect to the rest of the variables. Therefore, one can always fix these
variables such that the main coefficient is nonzero to arrive at an nm th
order polynomial of single variable Sm. This polynomial has exactly nm
roots. Each such root, together with the fixed variables, defines a zero of
p(S1, S2, ... , sm). Roughly speaking, there are a finite number of (m - 1)-
dimensional zero manifolds in the m-dimensional complex space.
One more essential difference between univariate and multivariate poly-
nomials is the following: Two multivariate polynomials may be relatively
prime and still possess common zeros.

4.6.2 Stable polynomials

In this subsection we introduce a class of stable multivariate polynomi-
als. To this end we first define the set of polynomials of a given degree.
Given a vector of partial degrees (nl, n2, .. . , n m ), one may define a set of
polynomials

(4.23)

With every polynomial P(Sl, 82, ... , Sm) from the set, one may associate
the real coefficient vector

from the N-dimensional coefficient space, where N = (nl + 1)(n2+ 1) ... (nm
+ 1).
In the following we will often denote the vector (s 1 , 82, ... , 8 m ) as s. The
following lemma is the standard extension of the well-known continuity
property of zeros of univariate polynomials with respect to small coefficient
variations.

Lemma 4.10 Let polynomial Po(s) E Pnl,n2, ... ,n", have a zero s(O) =
(s~O), s~O), ... , s~)). Denote by a(O) the coefficient vector of thi8 polynomial.
138 4. Robust Stability Analysis

For arbitrary c > 0 there exists 8 > 0 such that every polynomial p(s) with
the coefficient vector a in the 8-neighborhood of a(O):

has a zero in the c-neighborhood of s(O).

According to this lemma, the zeros of multivariate polynomials possess

continuity property with respect to coefficient variations.
Define polydomain

(4.24)

Definition 4.2 A polynomial p(s) E Pnl,n2, ... ,n", is said to be strict sense
stable (SSS) if
p(s) =1=0, "is E r~).

The principal deficiency of this stability concept is that SSS polynomials

do not always preserve the stability property under arbitrary small coeffi-
cient variations. The property becomes very fragile when polynomials have
large zeros close to the essential boundary
m
n= U{ (S1, S2, ... , sm) I Re(si) = O}, (4.25)
i=1

of the polydomain r~).

A new class of stable multivariate polynomials will be defined inductively
with respect to the number of variables.

Definition 4.3

1. For m = 1: given n1 ~ 0, polynomial P(S1) of degree n1 is called

stable if it is Hurwitz stable;

2. for m > 1: given vector (n1, n2, ... , n m ), polynomial p( S1 , S2, ... , sm)
of degree (n1, n2, ... , n m ) is called stable if it satisfies the following
conditions:

• polynomial p( 81, S2, ... , 8 m ) is SSS;

• in decomposition (4.22) with respect to every one of the vari-
ables, the main coefficient (the coefficient of the highest order
term) is a stable polynomial of m - 1 variables whose par-
tial degrees coincide with the corresponding partial degrees of
p(S1' S2, ... , sm).
 4.6 Multivariate polynomial approach 139

The principal advantage of the stability concept is that the class of stable
polynomials is the biggest one for which polynomials preserve the stability
property under small coefficient variations.

Lemma 4.11 Let a polynomial p(s) E Pnl,n2, ... ,n", be stable, and assume
that nm > O. Define the transformed polynomial

jJ(s) = (smtmp (Sl, ... ,Sm-1, s~),

then p(s) is stable and deg(p) = deg(p).
It is worth noting that the statement is not true in general for SSS (see
the following example).

Example 4.3 Polynomial P(Sl,S2) = SlS2 + Sl + 1 E P1,1 is SSS, but the

transformed one p(s}, S2) = S1 + SlS2 + S2 has the trivial zero and therefore
is not SSS. 0

Corollary 4.12 Successively applying Lemma 4.11 to several variables one

can get from a stable polynomial p(s) a set of stable polynomials of the same
degree.

Theorem 4.13 Let P(S1, S2, ... , sm) E P nl ,n2, ... ,n m be a stable polynomial.
There exists a E: > 0 such that every polynomial with a coefficient vector
from the E:-neighborhood of the coefficient vector ofp(s}, S2, ... , sm) is stable.

One more useful property of stable polynomials is the invariance of the

stability under differentiation.

Theorem 4.14 Let p(s}, S2, ... , sm) E P nl ,n2, ... ,n m be a stable polynomial.
If nm > 0, then derivative
(1)( )
8 p( Sl , S2, ... , sm) nl g
'""' v a k S2, ... , sm k_
8s m L.t 8s Sl - ...
k=O m
nm-l (m-1)(
g )
'""' v a i Sl, ... , sm i
L.t g sm-1
i=O vSm

(4.26)

is a stable polynomial from P nl ,n2, ... ,n m -1.

Corollary 4.15 Let p(s}, S2, ... , sm) E P nl ,n2, ... ,n m be a stable polynomial.
Then all coefficients in decomposition (4.22) are stable (m - I)-variate
polynomials from P nl ,n2 , ... ,nm-l .
140 4. Robust Stability Analysis

As it follows from the example below, for SSS polynomials this statement
is not true in general.

Example 4.4 Polynomial p( Sl, S2) = Sl S2 + Sl + 1 belongs to Pl,l, and it

is SSS, but the derivative with respect to the second argument is not a SSS
polynomial of Pl,o. 0

Theorem 4.16 All coefficients of a real stable polynomial

have the same sign.

It was already mentioned that the main reason for the enormous sensitiv-
ity of the strict sense stability to small coefficient variations is the existence
of zeros close to the essential boundary (4.25) of the set r~). The following
theorem shows that stable polynomials have no such zeros.

Theorem 4.17 For every stable polynomial p(s) E Pnl,n2, ... ,n", there ex-
ists c > 0, such that it has no zeros in the c-neighborhood of the essential
boundary n.

4.6.3 Stability of an interval multivariate polynomial

Consider a real multivariate interval polynomial

Of course we assume that I c P n1 ,n2, ... ,n"" that is, all members of the
family have the same degree n = (nl, n2, ... , n m ). Moreover we assume that
all lower bounds are positive, that is Q:il ... i", > 0.

Theorem 4.18 The interval polynomial is stable if and only if 4 . 2m - l

special members of the family with extreme coefficients are stable.

Let us define explicitly these testing polynomials. To this end we intro-

duce the set of m-dimensional sign vectors

2= {(VI,1I2, ... ,lIm)lvi E {-1,+1}, i= 1,2, ... ,m}.

This set consists of 2m elements. With every vector from the set, one can
associate four polynomials as follows:
 4.6 Multivariate polynomial approach 141

Given a sign vector (vt, V2, ... , v m ), the coefficients of the first polyno-
mial p~Vl.V2'''''V''') (SI, S2, ... , sm) are defined by the following rule: When the
index sum is even (il + i2 + ... + im = 2l) then

when the index sum is odd (il + i2 + ... + im = 2l + 1), then

Coefficients of the second polynomial

are defined by the rule: When the index sum is even (il + i2 + ... + im = 2l),
then

when the index sum is odd (il + i2 + ... + im = 2l + 1), then

Coefficients of the third polynomial

are defined by the rule: When the index sum is even (il +i2 + ... +i m = 2l),
then

when the index sum is odd (il + i2 + ... + im = 2l + 1), then

And the coefficients of the last one

P4(Vl.V2 ..... tn)( SI,S2,,,,,Sm )

V

are defined by the rule: When the index sum is even (il + i2 + ... + im = 2l)
then
142 4. Robust Stability Analysis

Proceeding in a similar way for all the elements from 3, one finally ob-
tains 4 . 2m polynomials. Taking into account that all polynomials in I
have only real coefficients, we may reduce the set by half and consider only
those of them for which Vm = 1. After this reduction, we obtain the desired
testing set of 4· 2m - 1 polynomials which is referred to in Theorem 4.18.

4· 6·4 Stability of a diamond family of multivariate

polynomials
Consider a m-variate polynomial whose real coefficient vector, a, varies in
the N-dimensional diamond

.~. '" la.. . - a(O)

'l.1'l.2···'ttn i 1 i2 ... i rn I< r
-
} ,
(4.28)
111.2 ... 't 77t

where a(O) denotes the center point and p(O) (s) = p(s, a(O») is the center
polynomial; r > 0 is the radius of the diamond. From here on we denote a
diamond family of multivariate polynomials by 'Dm = { p (', a) I a ED} .
According to the coefficient positivity assumption, all coefficients of sta-
ble polynomials have the same sign. Therefore, without loss of generality,
..
we may assume a~~12 i", > 0 and

(4.29)

It is clear that the last condition implies that 'Dm C Pnl,n2, ... ,n",.

Theorem 4.19 The diamond family of multivariate polynomials, 'D m , is

stable if and only if (m + 1) 2m +1 extreme polynomials of the family are
stable.

As in the previous section, let us define explicitly this testing set of

polynomials. To this end we divide the essential boundary n(m) into 2m
subregions in such a way that for every subregion there exist exactly 4m
one-parametric families that describe the boundary of the value set
 4.6 Multivariate polynomial approach 143

of the diamond family. Here w = (Wl,W2, oo.,w m ) belongs to the selected

subregion.
The first subregion

00 = {jw I Wi E (0,1) ; i = 1,00', m }.

Here the 4· m one-parametric families covering the boundary of the value
set are the following ones:

(4.30)

Let us fix one of the indices, say a, then for subregion

O",={ jwl w",E(1,OO), wkE(O,l); kE{1,oo.,m}, k#a}.

Here the boundary of the value set of the diamond family is defined by the
following one-parametric families:

{ p(O) (8) ± r [J.L ± (1 - J.L) SkJ s~a I J.L E [0, 1J}, k = 1,00', m. (4.31)

Here
l", = { n", - 1, if a = k, (4.32)
n"" if a # k,
where a E {I, 00', m}. There are exactly C;' = m subregions of this type.
And for everyone of them there are defined exactly 4m one-parametric
families.
Now fixing two of the indices, say a and f3 (a # (3), we define subregion

o {. I w'" E (1,00), wf3 E (1,00), }.

""f3 = JW Wk E (0,1); k E {I, 00', m}, k =I a, k =I j3
Here we have the following one-parametric families:

{p(O) (s) ± r [J.L ± (1 - J.L) SkJ s~as~ IJ.L E [0, 1J}, (4.33)
k=l,oo.,m; k#a, k#j3,
where l", and lf3 are defined in (4.32), and a,j3 E {l,oo.,m}, a # f3. The
number of such subregions is C~ = m( m - 1) /2.
Now fixing v < m of the indices, say a, f3, 00., 'Y (they are all distinct), we
can define subregion

0""f3, ... ,7 ={ jw I Wk EW~,E1?'~~ {~~oo~~\',ooJ ¥'~, ~7: J~~.~~ # 'Y }.

Here we have again 4m one-parametric families:

{p(O)(s)±r[J.L±(l-J.L)SkJs~s~oo.s~'Y I J.LE[O,lJ}, k=l,oo.,m,

(4.34)
144 4. Robust Stability Analysis

where la, l{3, ... , l"( are defined as above. In this way we find C:;" subregions
of this type.
Continuing this process, we define the last subregion

Ol,2, ... ,m={jw IWkE(1,oo)j k=1, ... ,m},

where the boundary of the value set is defined by the following 4m one-
parametric families:

{ p(O) (8) ± r [It ± (1 - It) Sk] SilS~2 ... s~ I It E [0,1] }, k = 1, ... , m.

(4.35)
Finally, we obtain "'~
~.=O
C:n
= 2m subregions and associate 4m one-
°
parametric families with each one of them. Now, substituting in everyone
of these families It = and It = 1 we obtain the testing polynomials.

4.7 Notes
For the case of time-invariant systems, there is a strong connection between
the exponential stability of a system and the location on the complex plane
of the zeros of its characteristic function. For the case of time-delay sys-
tems of concentrated delays, the characteristic function is a finite sum of
potential and exponential factors.
The book [9] remains one of the best references for those who are in-
terested in the study of zeros of such functions. It contains an exhaustive
analysis of the location on the complex plane of zeros of quasipolynomials
from several interesting classes. A much more general and profound study
of zeros of quasipolynomial functions may be found in the Ph.D. disser-
tation of E. Schwengeler, see [241]. In our presentation of the exponential
and potential diagrams, we followed this dissertation. Example 4.1 is taken
from [152].
The Edge Theorem has been proved first for the case of retarded type
quasi polynomials in [70], and then has been extended to the case of neutral
type quasipolynornials in [71].
Stability analysis of an edge subfamily of quasipolynomials, including
conditions for convex directions, Theorem 4.7, may be found in [138]. Sta-
bility conditions for an interval quasipolynomial are also given in [138].
Basic results on applications of multivariate polynomials for stability
analysis of quasipolynomials are given in [131]-[133].
In presenting of stability conditions for multivariate polynomials, we fol-
lowed the publications [141]-[143].
 Part II

Time Domain Approach

5

Systems with Single Delay

5.1 Introduction
In this chapter we will explore the time domain approaches of stability
analysis. An advantage of time domain methods is the ease of handling
nonlinearity and time-varying uncertainties. However, in order to illustrate
the basic ideas, in this chapter we will concentrate on the stability problem
of linear time-invariant systems with single delay

x(t) = Aox(t) + AlX(t - r), (5.1)

where Ao and Al are given n x n real matrices. The usual initial condition
is in the form of
Xo = rjJ. (5.2)
We will defer the discussions on the uncertainties and systems with multiple
delays as well as distributed delays to later chapters.
We will use the Razumikhin Theorem (Theorem 1.4 in Chapter 1) and
the Lyapunov-Krasovskii Stability Theorem (Theorem 1.3 in Chapter 1) to
discuss the stability of the system. We will restrict ourselves to using the
bounded quadratic Lyapunov function or Lyapunov-Krasovskii functional,
and aim at arriving at stability criteria that can be written in the form
of Linear Matrix Inequalities (LMI) or a closely related form. Efficient
numerical methods are available to solve LMIs. The readers are encouraged
to read Appendix B to become familiar with the basic concepts of LMI
and a number of useful topics especially relevant to time-delay systems,
notably the variable elimination technique in LMI and quadratic integral
inequalities (Jensen Inequality). We will frequently refer to them in this
chapter.
The first part of this chapter will mainly use the Razumikhin Theorem.
Section 5.2 discusses delay-independent stability criteria. The first part dis-
cusses systems with a single delay expressed by (5.1), which can be derived
using the simplest form of Lyapunov function. The delay-independent sta-
bility criteria test if a system is asymptotically stable for arbitrary delay. It
is intuitively obvious that such a criterion would be very conservative if the
delay r is already known and small. It turns out that delay-independent
stability criteria effectively consider the delayed term, i.e., second term on
the right hand side of (5.1), to be always detrimental to the stability of
the system. Therefore, such criteria are useful when the system behavior is
148 5. Systems with Single Delay

dominated by the first term on the right hand side of (5.1) (i.e., Al is rather
small) and when we want to obtain a quick assurance that the delayed term
will not destabilize the system. Section 5.2 continues with a discussions on
the delay-independent stability of systems with distributed delays, which
is useful for deriving the delay-dependent stability criteria in Section 5.3.
Section 5.3 derives simple delay-dependent stability criteria using the
Razumikhin Theorem. In practice, as indicated in Example 1.2 in Chapter
1, there are systems that require feedback to improve the system perfor-
mance, and the feedback channel involves delays. In these cases, the delay-
independent stability criteria are clearly insufficient. In many cases, it is
more desirable that the system does not have a delay, and the presence
of delay is indeed detrimental to system stability and performance. It is
therefore natural to transform a delayed system into a system without de-
lay plus distributed delays, and to treat the term with distributed delay as
disturbance. This process is known as model transformation. We can use
the delay-independent stability criterion for distributed delay discussed in
Section 5.2 to obtain a stability criterion, which turns out to depend on
the size of delay. This is because the magnitude of the disturbance term is
closely related to the size of time delay. An important observation for such
a criterion is that the delay is always considered detrimental to stability.
Indeed, a necessary condition for such a criterion to be satisfied is that it
is satisfied for any smaller delay.
The delay-dependent stability criteria so arrived are still conservative.
In addition to the potential conservatism caused by applying the delay-
independent stability criterion to the system with distributed delays result-
ing from the model transformation, we will show that the model transfor-
mation process introduces spurious poles not present in the original system.
The dynamics represented by such spurious poles are known as additional
dynamics. As delay increases from zero, it is possible for the additional
dynamics to become unstable before the original system does. Therefore,
the model transformation process itself may introduce conservatism.
We will also show that it is possible to derive a simple delay-dependent
stability criterion without explicit model transformation. It turns out that
the resulting stability criteria include as special cases the delay-independent
and delay-dependent stability criteria derived in the earlier part of Section
5.3.
Sections 5.4 and 5.5 uses the Lyapunov-Krasovskii method to derive
delay-independent and simple delay-dependent stability criteria, parallel
to the results obtained by using Razumikhin Theorem. It is interesting to
notice that some resulting stability criteria take a very similar form as the
ones obtained using Razumikhin Theorem. The Razumikhin results can be
obtained from the Lyapunov-Krasovskii results by imposing additional con-
straints. While this may give the impression that the Razumikhin results
are more conservative, it will be shown in Chapter 6 that the Razumikhin
 5.1 Introduction 149

Theorem-based results are also applicable to systems with uncertain time-

varying delay. From this point of view, the results are not surprising; for the
stability of a wider class of systems, the criteria need to be more stringent.
There are, however, many practical cases where both delay-independent
and simple delay-dependent stability criteria discussed in Sections 5.2 to
5.5 are unsatisfactory. This is so not only from the quantitative point of
view (i.e., they may be overly conservative), but also from the qualitative
point of view. Indeed, in both types of criteria, the delay is always regarded
as detrimental to stability. Therefore, such criteria are clearly not satisfac-
tory if we want to explore the possibility of using time delay (as opposed
to the case of reluctantly accepting delay in the delayed feedback) to sta-
bilize the system or to enhance the system performance. AB discussed in
Chapter 1, time delays are indeed introduced intentionally in practice. It is
not uncommon that a system is unstable without delay and is stable with
certain delay. It is therefore natural to discuss the stability criteria that can
handle this type of problem. In Section 5.6, we will show that the existence
of a more general quadratic Lyapunov-Krasovskii functional is necessary
and sufficient for stability.
Unfortunately, it is generally very difficult to check the existence of
the quadratic Lyapunov-Krasovskii functional discussed in Section 5.6. For
practical computation, it is shown in Section 5.7 that it is possible to choose
a piecewise linear kernel to approximate the continuous one with a resulting
guaranteed stability limit approaching the continuous case. This method is
known as the discretized Lyapunov functional method. Numerical examples
show that in general, a somewhat coarse discretization is often sufficient to
arrive at a very accurate stability limit in practice.
In this chapter, because only the bounded quadratic Lyapunov function
is used, we only need to use the following restricted form of the Razumikhin
Theorem.

Proposition 5.1 A time-delay system with maximum time delay r is as-

ymptotically stable if there exists a bounded quadratic Lyapunov function
V such that for some c; > 0, it satisfies

(5.3)

and its derivative along the system trajectory V(x( t)) satisfies

(5.4)

whenever
V(x(t +~)) ~ pV(x(t)), - r ~ ~ ~ ° (5.5)

for some constant p > 1.

150 5. Systems with Single Delay

Proof. Compared with the Razumikhin Theorem (Theorem 1.4 in Chap-

ter 1), the only necessary condition missing is that

V(x) :S v(x).

Since V (x) is bounded quadratic, for a sufficiently large K > 0, we have

Therefore, all the Razumikhin Theorem conditions are satisfied with

This completes the proof. _

We will refer to (5.3) as the Lyapunov function condition, and (5.4)
subject to (5.5) as the Razumikhin derivative condition.
Similarly, we can state a restricted version of Lyapunov-Krasovskii The-
orem.

Proposition 5.2 A time-delay system is asymptotically stable if there ex-

ists a bounded quadratic Lyapunov-K rasovskii functional V (¢) such that for
some c > 0, it satisfies
(5.6)

and its derivative along the system trajectory,

satisfies
(5.7)

Proof. The only condition omitted from the Lyapunov-Krasovskii sta-

bility condition is
V(¢) :S v(II¢llc)·
But since V is a bounded quadratic functional of ¢, the above is satisfied
for
v(I\¢l\c) = KI\¢I\~

°
when K > is sufficiently large. _
We will refer to (5.6) as the Lyapunov-Krasovskii functional condition,
and (5.7) as the Lyapunov-Krasovskii derivative condition.
 5.2 Delay-independent-Razumikhin Theorem 151

5.2 Delay-independent stability criteria based on

the Razumikhin Theorem
5.2.1 Single delay case
Consider the system with single delay described by (5.1). We will use the
Razumikhin Theorem to obtain a simple stability condition using the Lya-
punov function
(5.8)

Proposition 5.3 The system described by (5.1) is asymptotically stable if

there exist a scalar

and a real symmetric matrix P such that

( P Ao +
Aip
Air P + oP PAl)
-oP <.
° (5.9)

Proof. We will use Proposition 5.1. Choose Lyapunov function V as in

(5.8). Since (5.9) implies P > 0, we can conclude that for some sufficiently
small t: > 0, the Lyapunov function condition

(5.10)
is satisfied. Now consider the derivative of V along the trajectory of the
system (5.1),
. d
V(x(t)) = dt V(x(t)) = 2xT (t)P[A ox(t) + AIX(t - r)]. (5.11)

Whenever Xt satisfies

V(x(t + 0)) < pV(x(t)) for all - r ~ 0 ~ ° (5.12)

for some p > 1, we can conclude that for any 0 > 0,
V(x(t)) ~ 2xT(t)P[Aox(t) + AIX(t - r)]
+o[pxT(t)Px(t) - xT(t - r)Px(t - r)]
"J,T ( P Ao + Air P + opP PAl )"J, (5.13)
'POr Aip -op 'POr'

where
rPOr = (xT(t) xT(t - r) )T.
The inequality (5.9) implies that for some sufficiently small 8 > 0, p = 1+8,
PAo+AirP+opP
( Aip PAl)
-oP <,
°
152 5. Systems with Single Delay

which, according to (5.13), implies that the Razumikhin derivative condi-

tion
V(x(t)) S -Ellx(t)112 (5.14)
is satisfied. According to Proposition 5.1, the system is asymptotically sta-
ble . •
The first important point to be noted is that the stability criterion stated
in Proposition 5.3 does not depend on the delay r. Therefore, it is obvious
that if the stability criterion is satisfied, then the system is asymptotically
stable for arbitrary delay r. As also discussed in Chapter 2, this type of sta-
bility criteria are known as delay-independent stability criteria. If the delay
of the system is small, such criteria often give very conservative stability
assessments.
In essense, the criterion in Proposition 5.3 considers the term AIx(t-r) to
be always detrimental to stability. To see this, using the Schur complement,
under the assumption P > 0, (5.9) is equivalent to
T 1 I T
PAo+AoP+aP+-PAIP- AIP<O,
a
or, with Q = p-I,
T 1 T
AoQ + QAo + aQ + -AIQAI < 0.
a
Since the last two terms aQ and iAIQA[ are both positive semi-definite,
the above also implies that Ao is Hurwitz. In fact, it can be said roughly
that the larger Al is, the more difficult it is to satisfy the above criterion.
Although (5.9) is linear with respect to P for fixed a, and linear with
respect to a for fixed P, it is not linear with the combined variables (P, a),
and therefore, it is not an LMI. This makes computation more difficult. Let
a be the solution of the following generalized eigenvalue problem (in the
LMI sense, often abbreviated as GEVP, see Appendix B):
a = supa, (5.15)

subject to

PAo +A6'P+aP < 0, (5.16)

P > 0. (5.17)

Then, since the (1,1) entry of the matrix in (5.9) has to be negative definite,
any a satisfying (5.9) must satisfy
0< a < a.
For the least conservative result using this formulation, it is necessary to
solve (5.9) for all a in the interval (0, a). The sweeping of a may be compu-
tationally costly. Alternatively, we may choose to fix a = a/2 and accept
 5.2 Delay-independent-Razumikhin Theorem 153

the additional conservatism. Then, the computation consists of solving the

GEVP to obtain a and solving LMI problem (5.9) with a = aj2 fixed.

Example 5.1 Consider the system

. (-20 -0.90) x(t) + (-1-1 -10) x(t - r),

x(t) = (3

where (3 ~ 0 and r > O. For each given (3, we choose a = aj2 and check the
feasibility of (5.9). A bisection process is used to find (3max, the greatest (3
for (5.9) to be feasible. It was found that (3max = 0.9. 0

In feedback synthesis, sometimes it is preferable to use the following

equivalent dual form of Proposition 5.3.

Proposition 5.4 The system described by (5.1) is asymptotically stable if

there exist a scalar
a>O
and a symmetric matrix Q such that

( AoQ + QA&, +aQ AIQ ) < O. (5.18)

QAf -aQ

Proof. Applying Proposition 5.3, the system is asymptotically stable if

(5.9) is satisfied. Left-multiply (5.9) by

diag (P-l P-I),

and right-multiply by its transpose. We can obtain (5.18) after a variable

transformation

Based on the above dual form of stability criterion, state feedback sta-
•
bilization design is more convenient. For example, given the system

x(t) = Aox(t) + AIX(t - r) + Bu(t),

we are to design a feedback control

u(t) = Kox(t) + KIX(t - r),

such that the closed loop system is stable independent of delay. Since the
closed loop system is

x(t) = (Ao + BKo)x(t) + (AI + BKI)X(t - r)

154 5. Systems with Single Delay

using Proposition 5.4, the condition we need to satisfy is

( (Ao + BKo)Q T+ Q(Ao + BKof + aQ (AI + BKI)Q) 0

Q(AI + BKd -aQ < .
With a variable transformation

the above becomes

( AoQ+BVo+QA6+VlBT+aQ AIQ+BVI) 0
(AIQ + Bvif -aQ < ,
which is an LMI if a is fixed. Such a process is not directly possible based
on Proposition 5.3.
Most of the stability criteria discussed in this chapter have a dual form,
which can be obtained by a similar procedure. We will not discuss these
dual forms individually to avoid repetition.

5.2.2 Distributed delay case

Now consider the system with distributed delays

x(t) = Aox(t) + lOr A(B)x(t + B)dB, (5.19)

where Ao is a given constant matrix and A( B) is a given matrix valued

function of BE [-r, 0]. We can again use V(x) in (5.8) to study the stability
of the system and conclude the following.

Proposition 5.5 The system with distributed delays described by (5.19) is

asymptotically stable if there exist a symmetric matrix P, a scalar function
a(B) > 0, for 0::; B::; r, (5.20)
and a symmetric matrix function R( B) such that

PAo + A6 P + lOr R(B)dB < 0, (5.21)

( a(B)P - R(B) PA(B) ) < 0 for 0::; B ::; r. (5.22)

AT(B)P -a(B)P

Proof. Use the Razumikhin Theorem in a similar way to the proof of

Proposition 5.3. Since (5.22) implies P > 0, we can conclude that the
Lyapunov function V(x) = x T Px satisfies

V(x) .:::: Ellxl1 2

 5.2 Delay-independent-Razumikhin Theorem 155

for sufficiently small E: > O. Also, let p > 1. Whenever

V(x(t + 0)) < pV(x(t)) for all - r ::; 0::; 0

is satisfied, we can calculate

V(x(t)) = 2xT (t)P [Aox(t) + lOr A(O)x(t + O)dO]

lOr A(O)x(t + O)dO]

l:
< 2xT (t)P [Aox(t) +

+ a(O) [px T (t)Px(t) - xT(t + O)Px(t + O)]dO

= xT(t) [PA o + A~ P + lOr R(O)dO] x(t)

+ lOr (xT(t) xT(t + 0) )

( pa(O)P
AT(O)p
- R(O) PA(O) ) ( x(t)
-a(O)P x(t + 0)
) dO
.

With the above expression, and the fact that p > 1 can be arbitrarily close
to 1, (5.21) and (5.22) imply

V(x(t)) ::; -E:llx(t)112

for some sufficiently small E: > O. Therefore, the system is asymptotically

stable according to Proposition 5.1. •
An especially interesting case is when A( 0) is piecewise constant, in which
case we may choose a( 0) and R( 0) to be also piecewise constant, reducing
(5.20)-(5.22) to a finite set of matrix inequalities. This is the case when we
apply Proposition 5.5 to the simple delay-dependent stability case discussed
in the next section.
The stability criterion described in Proposition 5.5 may be regarded as
delay-independent in the following sense: if the distributed delay -0 in
x(t+O) in (5.19) is changed to -(rnew/r)O, and thus the maximum delay of
the system is changed from r to r new , the stability condition in Proposition
5.5 remains the same.

Proposition 5.6 The conditions in Proposition 5.5 are satisfied for the

l:
system (5.19) if and only if they are also satisfied for the following system

x(t) = Aox(t) + A(O)x (t + rn;w 0) dO. (5.23)

156 5. Systems with Single Delay

Proof. With a transformation of integration variable, the system (5.23)

can be written as

(5.24)

The stability conditions in Proposition 5.5 for this new system are therefore

anew (B) > 0, for 0 ~ B ~ r new ,

P Ao + A~ P + iOrnew Rnew (B)dB < 0,

( anew (B)P - Rnew (B) rn:w P A( rn:w B) )

_r_AT(_r_B)p -anew (B)P < 0, BE [-rnew,OJ.
Tnew Tnew

It is then obvious that the satisfaction of (5.20)-(5.22) is equivalent to the

satisfaction of the above three inequalities with the following correspon-
dence

anew (B) r (rrnew ) '

--a --B
rnew

Rnew(B) _r R(_r B)
ruew rnew
This completes the proof. •

5.3 Simple delay-dependent stability criteria based

on the Razumikhin Theorem
The delay-independent stability problem of the system (5.1) discussed in
the last section takes the following point of view. Consider an asymptoti-
cally stable system without time delay

x(t) = Aox(t). (5.25)

With arbitrary time delay, as the coefficient matrix Al of the delayed term
grows from zero, the system performance deteriorates and eventually the
system is in danger of losing stability.
However, in practice, one often encounters nominal open-loop systems of
the form (5.25) which are unstable. Feedback with small delay (due either
to measurement or control delay) is needed to stabilize the system. It is
therefore useful to take another point of view. Rewrite the system (5.1) in
the following form

x(t) = (Ao + AI)X(t) + AI(X(t - r) - x(t)).

 5.3 Simple delay-dependent-Razumikhin Theorem 157

We consider the term

Al(X(t - r) - x(t»
to be a disturbance to the nominal stable system

x(t) = (Ao + Al)X(t).

The nominal system is stable with satisfactory performance. AB the delay
r increases from zero, the performance of the system may deteriorate and
eventually is in danger of losing stability. We need some way to reflect
the fact that the disturbance grows from zero as r increases from zero.
This can be achieved by model transformation, which, in combination to
the delay-independent stability for distributed delay problem, allows us
to derive a simple delay-dependent stability sufficient condition. While this
formulation still gives a conservative stability estimate, it is still very useful
because of its simplicity.

5.3.1 Model transformation

Consider again the system (5.1) with initial condition (5.2), where the
initial function ¢ E C([-r, 0], an). With the observation that

x(t - r) = x(t) -[Or x(t + O)dO

= x(t) _[o}AoX(t + 0) + AIX(t - r + O»)dO,

for t ? r, we can write the system (5.1) as

x(t) = [Ao

+ i:+ Al)X(t)
[-AIAox(t + 0) - AIAIX(t - r + O))dO, (5.26)

with initial condition

x(O) = 1/;(0), - r 5: 0 5: r (5.21)

where

</J(O) , -r 5: 0 5: 0,
1/;( 0) = {
solution of (5.1) with initial condition (5.2), 0< 0 5: r.
(5.28)
Therefore, the system described by (5.1) and (5.2) is embedded in the sys-
tem described by (5.26) and (5.21) without the initial condition constraint
158 5. Systems with Single Delay

(5.28). Since this is a time-invariant system, in studying the stability prob-

lem, we can shift the initial time to write it in a more standard form

y(t) = Aoy(t) + l:r A(B)y(t + B)dB, (5.29)

where

(5.30)

with initial condition

y(B) = 1f;(B), - 2r ~ B ~ O. (5.31)

Here, we have used y instead of x to represent the state variable to empha-

size that these two are indeed different systems.
The process of transforming a system represented by (5.1) to one repre-
sented by (5.29) and (5.30) is known as a model transformation. Clearly, the
stability of the system represented by (5.29) to (5.31) implies the stability
of the original system. However, the reverse is not necessarily true because
of the lifting of the initial condition constraint (5.28). As will be seen later
in this section, the effect of lifting the initial condition constraint results
in additional dynamics characterized by spurious poles of the transformed
systems not present in the original system.

5.3.2 Simple delay-dependent stability criteria using explicit

model transformation
With the model transformation, we arrived at a system described by (5.29)
to (5.31), which includes the original system described by (5.1) and (5.2)
as a special case. Since the transformed system is one with distributed
delay, we can use Proposition 5.5 to derive the stability condition, which,
of course, is sufficient for the stability of the original system.

Proposition 5.7 The system described by (5.1) is asymptotically stable if

there exist real scalars Qo > 0, Ql > 0 and real symmetric matrices P > 0,
Ro, R 1 , such that
0, (5.32)

< 0, k = 0,1. (5.33)

Proof. We only need to prove the stability of the transformed system

described by (5.29) and (5.30). Using Proposition 5.5, it can be concluded
 5.3 Simple delay-dependent-Razumikhin Theorem 159

that the system is asymptotically stable if there exist a( 0), P, and R( 0) to

satisfy

pAo + A5' P + i 0
2r R(O)dO < 0, (5.34)

( ~(O)P
AT(O)P
- R{O) pA{o)
-a{O)P
)
< 0, - 2r ~ 0 < 0. (5.35)

Choosing the following piecewise constant (matrix) functions

-r ~ 0 < 0,
a{O)
-2r ~ 0 < -r,

R(O)
-r ~ 0 < 0,
-2r ~ 0 <-r
completes the proof. _
The stability criterion in Proposition 5.7 depends on the time delay r
and is therefore delay-dependent. We can eliminate the arbitrary matrices
Ro and Rl among the three matrix inequalities (5.32) to (5.33) to arrive
at the following equivalent form.

°
Corollary 5.8 The system described by (5.1) is asymptotically stable if
there exist real symmetric matrix P > and real scalars ao > 0, al >
such that
°
(5.36)

where

Proof. Divide (5.32) by r, and use Proposition B.6 in Appendix B to

eliminate Ro among the resulting matrix inequality and (5.33) for k = 0,
then eliminate Rl among the resulting matrix inequality and (5.33) for
k = 1. _
Again, due to the multiplication of ao and aI, the stability condition
described by (5.32) and (5.33) or (5.36) is not LM!. Let 0: be the solution
of the GEVP
0: = sup a, (5.38)
subject to the constraints

P > 0, (5.39)
!r [P{Ao + Ad + {Ao + Alf P] + aP < 0. (5.40)
160 5. Systems with Single Delay

Then al + a2 < 0 since M < O. Also, it is clear that 0 is proportional to

1/r. In order to check the satisfaction of the conditions, it is necessary to
search through the triangular region

ao > 0,
al > 0,
ao +al < 0,

for the existence of fixed ao and al such that the LMI (5.36) is feasible.
We can, of course, choose ao and al a priori within this region and accept
the conservatism. A reasonable choice seems to be

ao = (5.41)
2(IIAIAoll + II AlIl2)'
ollAlll2
(5.42)

Example 5.2 Consider the system

x(t) = (~2 -~.9) x(t) + (=~ ~1) x(t - r).

Applying Proposition 5.7, and using fixed ao and al calculated by (5.41)

and (5.42), we use a bisection process to calculate rmax such that (5.36)
is feasible for r < rmax. The calculation results in rmax = 0.9041. For
comparison, the analytical delay limit for stability is also calculated to be
r~~ytical = 6.1725. The analytical delay limit is calculated by the smallest
r such that there exists an imaginary pole. It can be seen that the criterion
presented here can be very conservative as compared to the analytical limit.
o
It should be pointed out that although delay-independent stability im-
plies delay-dependent stability for any given delay, because of the conser-
vatism of the above criterion, there indeed exist delay-independent systems
that do not satisfy the above delay-dependent stability criterion, as will be
shown in Example 5.4 later in this chapter.

5.3.3 Additional dynamics

In this subsection we will present the basic idea of additional dynamics
phenomena. Most of the proofs are omitted. Interested readers are referred
to references mentioned in the Notes section at the end of this chapter. It is
important to realize that model transformation may introduce significant
conservatism that is independent of the method used to judge the stability
of the transformed system.
 5.3 Simple delay-dependent-Razumikhin Theorem 161

In order to more explicitly express the additional dynamics and their

effects on the system behaviors, let us look at the transformed system
(5.29) from a slightly different angle.

Theorem 5.9 The transformed system described by {5.29} and {5.30} with
initial condition {5.31} is equivalent to the following system

iJ(t) = Aoy(t) + Aty(t - r) + z(t),

{ (5.43)
z(t) = At J~r z(t + B)dB
and the initial condition

{ y(B) = cp(B), for - r ~ B~ o.

z(B) = ¢(B),
The initial condition constraint {5.28} is equivalent to

Notice, instead of 'IjJ(O), -2r ~ B ~ 0 we may consider (cp(B), ¢(B)),

-r ~ 0 ~ 0 as independent initial condition.
Proof. We will show that given anyone system, we can derive the other.
From the transformed system (5.29) and (5.30), we can write

iJ(t) = Aoy(t) + Aty(t - r) + z(t),

I:
where

z(t) At [y(t) - y(t - r) - (Aoy(t + B) + Aly(t - r + B»dB]

At [ : [iJ(t + B) - Aoy(t + B) + A1y(t - r + B)] dB

At I: z(t + B)dB.

We have therefore obtained (5.43).

On the other hand, from (5.43) we can also obtain (5.29) and (5.30).
From the first equation of (5.43), we obtain

z(t) = iJ(t) - Aoy(t) - A1y(t - r).

Using the above in the second equation of (5.43), we arrive at

iJ(t) - Aoy(t) - Aty(t - r)

= At [Or [iJ(t + 0) - Aoy(t + B) + Aty(t - r + B)] dB.

162 5. Systems with Single Delay

Integrating the y(t + 0) term, we obtain

yet) = (Ao + AI) yet) - Al lOr [Aoy(t + B) + AIy(t - r + B)] dB,

which is (5.29) and (5.30).
The initial condition relation can also be easily verified. •
The theorem above explicitly expressed the additional dynamics, i.e., the
dynamics of the variable z(t), which are governed by the second equation
of (5.43).
In order to further characterize the additional dynamics, notice that the
characteristic quasipolynomial of (5.43) is

det ( S1 - A o 0- e- Al
rs
.6.t (8) =
.6.a (8).6. 0 (8),
where
.6. 0 (8) = det(81 - A o - e- rs Ad (5.44)
is the characteristic quasipolynomial of the original system (5.1), and
1 -rs )
.6.a ( 8) = det ( 1 - - : AI.

Therefore, the poles of the transformed system consist of the poles of the
original system satisfying .6. 0 (8) = 0, and additional poles, which are the
solutions of
1 -rs )
.6. a (8) = det ( 1 - -: Al = O. (5.45)

It is clear that although the stability of the transformed system implies the
stability of the original system, the reverse is not necessarily true. As the
delay r increases, it is possible that one of the additional poles may cross
the imaginary axis before any of the poles of the original system do. It is
therefore possible that the original system is stable but the transformed
system is unstable. This is the stability implication of the additional dy-
namics.
To further characterize the additional dynamics, we can conclude from
(5.45) that

.6. a (8) =det ( 1 - 1

- :
-rs Al ) =}]n (1 - Ai
1
-:
-rs) '

where Ai, is the ith eigenvalue of matrix AI. Let 8 = 8ik, k = 1,2,3, ... be
all the solutions of the equation
 5.3 Simple delay-dependent-Razumikhin Theorem 163

Then Sik, i = 1,2, ... , n; k = 1,2,3, ... are all the additional poles of the
system described by (5.29) and (5.30). We will refer to Sik as the kth
additional pole corresponding to Ai (or corresponding to the ith eigenvalue
of AI)'
It is of interest to consider the trend of the additional poles Sik as the
delay r ~ 0+. We will state the following result without proof.

Proposition 5.10 For any given AI, all the additional poles satisfy

lim
r-O+
Re(sik) = -00.
As a result of this proposition, all the additional poles have negative real
parts for sufficiently small r. As r increases, some of the additional poles Sik
may cross the imaginary axis. It turns out that the exact value when that
happens can be analytically calculated. We will state the follow theorem,
again without proof.

Theorem 5.11 Corresponding to an eigenvalue Ai of AI, Im(Ai) "# 0,

there is an additional pole Sik on the imaginary axis if and only if the time
delay satisfies

k7r+ L(Ai)
r = rik = Im(Ai) > 0, k = 0, ±1, ±2, ....
Corresponding to a positive real eigenvalue Ai of AI, there is an additional
pole on the imaginary axis if and only if
1
r = Ai'
No additional poles corresponding to a negative real eigenvalue Ai of Al
will reach the imaginary axis for any finite delay.

Given a time delay r, we can use the above theorem to find the region
r r, such that all the additional poles (zeros of ~a (s))have strictly negative
real parts if and only if all the eigenvalues of Al lie in r r:

rr = {z = x + iy I x < yltan(ry), -nlr < y < nlr}.

The region r r is shown in Figure 5.1. It is worth mentioning that the set
r r satisfies the following relations:
1
rr = -rl = {zlr I z E rt}.
r
In other words, the shape of the set r 1 remains the same for any r > 0,
but its linear dimension is inversely proportional to the delay r. Obviously,
r
as r ~ +0, the region r approaches the whole complex plane C.
164 5. Systems with Single Delay

'm

~; f··~····~····~··~····~····~·,~·
····~····~·~~···~····.~
.. ~~~
. , ! 'r'
Ao:z
., E------~----~-c~-"----_+----_+1rh_ "
-~:~~~~~~~~
-f. ~· · · · · ·~·'· · · · · ~· · · · ~l=
•

The following examples are presented to illustrate a number of possibil-

ities.

Example 5.3 Consider the system

x(t) = (-;,2 -~.9) x(t) + (=: ~I) x(t - r).

Since all the eigenvalues of Al are real and negative, all the additional poles
of the transformed system are on the left half plane for any r. Therefore
the stability of this system and the transformed system are equivalent.
The conservatism shown in Example 5.2 is due to the application of the
Razumikhin Theorem, not due to the model transformation. 0

Example 5 .4 Consider

.
x(t) =
( -6
0.2 -~.8 ) x(t) + (~8 !8) x(t - r).

Al has a pair of complex conjugate eigenvalues, ).1,2 = - 4±j4. The small-

est positive delay for an additional pole to reach imaginary axis can be
calculated as
3~
r=-=0589
16 . .
Therefore, the stability of the system and that of the transformed system
are equivalent for r < ~~. It can be checked using Proposition 5.14 in
the next section that the system is stable independent of delay. However,
the delay r = ~~ destabilizes the transformed system. The transformation
therefore introduces considerable conservatism. 0
 5.3 Simple delay-dependent-Razumikhin Theorem 165

Example 5.5 Consider

. ( ) (-3 -2.5) () (1.5

x t = 1 0.5 x t + -0.5
Al has two real eigenvalues, Al = 1 and A2 = -1, respectively. The small-
est positive delay for an additional pole to reach imaginary axis can be
calculated as r = 1. Therefore, the stability of the system and that of the
transformed system are equivalent for r < 1. However, it can be calculated
that the minimum delay to destabilize the original system is known to be
r = 2.4184 (using, for example, the method covered in Chapter 2), while
the minimum delay to destabilize the corresponding transformed system is
only r = 1. Again, as in the previous example, significant conservatism will
be introduced by the transformation. 0

5.3.4 Simple delay-dependent stability criteria using implicit

model transformation
Consider again the system
x(t) = Aox(t) + AIX(t - r) (5.46)
before model transformation. It is possible to not use model transformation
directly and still obtain a relatively simple stability criterion. During the
bounding in deriving stability criteria, information x(t + B), -2r ~ B ~ 0
is used. The constraint
x(t + B) = Aox(t + B) + AIX(t - r + B), - r ~ B~ 0 (5.47)
is partially but not fully accounted for. We will call such methods implicit
model transformation based methods. We will show that such a scheme can
indeed obtain stability results which include as special cases both delay-
independent and delay-dependent stability criteria obtained earlier in this
chapter, and therefore is less conservative.
To illustrate the idea, for the Lyapunov function Vex) = x T Px and real
matrices X T = X, Y, ZT = Z satisfying

({r ~) >0. (5.48)

We can write
Vex) 2x T (t)P[A ox(t) + AIX(t - r)]
< 2x T (t)P[A ox(t) + AIX(t - r)]
+ lOr ( xT(t) ~xT(t+~) )

({r ~)(
x(t)
t~x(t + 0
) d~. (5.49)
166 5. Systems with Single Delay

Using (5.47), we can write the term involving Z above as

(5.50)

where
x(t+e) )
Xtr(e)= ( x(t+e-r) . (5.51)

When for a p > 1,

V(x(t + 0)) < pV(x(t)) for all - 2r ~ 0 ~ 0,

we have

o:[pxT(t)Px(t) - xT(t - r)Px(t - r)]

+0:0 lOr [px T (t)Px(t) - x T (t + e)Px(t + e)]de

+0:1 lOr [pxT(t)Px(t) - xT(t + e- r)Px(t + e- r)]de 2': 0, (5.52)

for 0: 2': 0,0:0 2': 0,0:1 2': 0.

Using (5.50) and (5.52) in (5.49), carrying out integration for terms in
(5.49) whenever possible, we arrive at

V(x(t))
<_ T (0) ( ~ + (0: + O:Or + 0:1 r )pP PAl - Y) (0)
x tr (PAl _ y)T -o:P Xtr

it T() ( AfizAo - O:oP AfizAI

+ t-r Xtr e
Ai ZAo
) )
Ai ZA 1 - O:lP Xtr(e de,
where
~ = PAD + Ar P + rX + y + yT.
In view of the fact that p > 1 can be arbitrary close to 1, the above
expression implies that the system is asymptotically stable if there exist
real scalars 0: 2': 0,0:0 2': 0,0:1 2': 0, real symmetric matrices pT = P > 0,
X T = X, y, ZT = Z such that (5.48) and

( ~+(0:+0:0r+0:1r)P PA1-y)
{PAl - y)T -o:P < 0, (5.53)

( Air ZAo - O:oP

Ai ZAo
Air ZA I
Ai ZAI - O:IP
) < ° (5.54)
 5.3 Simple delay-dependent-Razumikhin Theorem 167

are satisfied.
We can eliminate the matrix variable Z in (5.48) and (5.54), then elimi-
nate X in the resulting matrix inequality and (5.53) to obtain the following
equivalent form.

Theorem 5.12 The system described by (5.46) is asymptotically stable if

there exist real scalars a > 0, ao > 0, a1 > 0, and real matrix Y and real
symmetric matrix P such that

-YA o
-aoP
(5.55)
°
°
where

It is interesting to examine (5.55) closely. If we set Y = 0, then ob-

viously the optimal selection is ao -4 0+, a1 -4 0+, and we recover the
delay-independent stability criterion discussed in Proposition 5.3. On the
other hand, if we set Y = PAl, then the optimal selection is a -4 0+,
and we recover the delay-dependent stability criterion derived by explicit
model transformation discussed in Corollary 5.8. It is possible to eliminate
the matrix variable Y using Corollary B.5 in Appendix B to obtain the
following.

Corollary 5.13 The system described by (5.46) is asymptotically stable if

there exist real scalars n > 0, no > 0, a1 > 0, and real symmetric matrix
P such that
P(aI - At}Ao P(aI - A 1)A 1
-aoP - ar Air P Ao -arAirPA 1 ) < 0,
-arA[PAo -alP - arAj PAl
(5.56)
where
1
M = -[P(Ao + AI) + (Ao + At}T Pj + (ao + at}P. (5.57)
r

To discuss the region of the scaling factors a, ao, and aI, first notice
that the (1,1) entry of (5.56) defined in (5.57) implies that ao and a1 must
lie in the triangular region ao > 0, a1 > 0, ao + a1 < ct, where ct is defined
in (5.38) to (5.40).
To find a bound for a, assume the matrix w = (Ao AI) has full row
rank, let w= (Air AT) T be a right inverse of w. We can left-multiply
168 5. Systems with Single Delay

(5.56) by the matrix

( 1 ~WT)
o 1 '
and right-multiply by its transpose to obtain

M -PA1w- -W 1 TR )
'T < 0,
A

(
TTl
-w Al P - ;:IW -R - o:rw Pw
A

where
R = diag (o:oP O:lP),
and
1
M = -(PAo + Ao P)
A
r
T
+ (0:-r + 0:0 + 0:1 ) r
1
P - 2 (o:oAo PA o + 0:1A1 PAt}.
AT A AT A

(5.58)
Therefore, given 0:0,0:1, let &(0:0,0:1) be the solution of the following GEVP

&(0:0,0:1) = max 0:, (5.59)

subject to !VI < 0, P > O. (5.60)

Then, 0: must satisfy 0 < 0: < &(0:0,0:1) if &(0:0, o:d > O.

With the above discussions, a reasonable choice for the scaling factors is
to choose 0:0 and 0:1 as

all(o:l - A1)Aoil
(5.61)
3(11(0:1 - A1)Aoil + 11(0:1 - A 1)A1ID'
all(o:l - A 1 )A1 11
(5.62)

and a reasonable choice of 0: is

(5.63)

We can first set 0: = 0 in calculating 0:0 and 0:1 using (5.61) and (5.62),
then calculate 0: using 0:0 and 0:1 just calculated in (5.63), and update 0:0
and 0:1 using the newly calculated 0: (we can, of course, also update 0: using
the newly calculated 0:0 and 0:1).

Example 5.6 Consider the same system discussed in Example 5.2. Using
the method discussed here with 0:,0:0, and 0:1 calculated by (5.61) to (5.63),
with 0:0, 0:1, and 0: all updated, the calculation indicates that the the
system is asymptotically stable for r < rmax = 0.9476. For r = r max , the
scaling factors are 0: = 0.4813, 0:0 = 0.7160, and 0:1 = 0.6208. This is an
improvement over the method using explicit model transformation. 0
 5.4 Delay-independent-Lyapunov-Krasovskii Theorem 169

5.4 Delay-independent stability criteria based on

the Lyapunov-Krasovskii Stability Theorem
In this section, we will discuss the stability of the same system (5.1) using
some simple Lyapunov-Krasovskii functional method. The results parallel
those obtained by Razumikhin Theorem.

5.4.1 Systems with single delay

Consider again the system (5.1). Probably the simplest stability criterion
can be obtained by using the following Lyapunov-Krasovskii functional

where the matrices P and S are symmetric and positive definite. We will
write the above as

It can be easily calculated that the derivative of V along the system tra-
jectory is

V(xt)=(xT(t) xT(t-r))(PATIApo+A(fP+S PAI)(X(t) )

-S x(t-r) '

or we can write Xt as ¢ to obtain

V(¢) = (¢T(O) ¢T(_r)) (PA¢lpO+A(fP+S PAl) (¢(O) )

-S ¢(-r)·
(5.65)
It is clear that V(xt) ~ -cllx(t)112 for some sufficiently small c > 0 if the
matrix in expression (5.65) is negative definite. Thus we can conclude the
following.

Proposition 5.14 System (5.1) is asymptotically stable if there exist real

symmetric matrices
P>O (5.66)
and S, such that

( P Ao + A(f P + S PAl) < 0 (5.67)

Aip -S
is satisfied.
170 5. Systems with Single Delay

Proof. Use Proposition 5.2, and choose Lyapunov-Krasovskii functional

(5.64). Notice that (5.67) implies
S> 0, (5.68)
which together with (5.66) implies
V(<!» ~ cll<!>(0)11 2
for some sufficiently small c > O. The Lyapunov-Krasovskii functional con-
dition is satisfied. Also, (5.67) implies that
V(<!» :::; -cll<!>(0)11 2
in view of (5.65), and the Lyapunov-Krasovskii derivative condition is also
satisfied. Therefore, the system is asymptotically stable according to Propo-
sition 5.2. •
It is interesting to compare the stability criterion in Proposition 5.14
obtained using the Lyapunov-Krasovkii Stability Theorem with the corre-
sponding criterion in Proposition 5.3 obtained using the Razumikhin The-
orem. It can be seen that the conditions in Proposition 5.3 can be obtained
from those of Proposition 5.14 by introducing the additional constraint
S= aP. (5.69)
This constraint, of course, makes the criterion in Proposition 5.3 obtained
using the Razumikhin Theorem more conservative than the corresponding
criterion in Proposition 5.14 obtained using the Lyapunov-Krasovskii Sta-
bility Theorem. Also notice that the criterion in Proposition 5.14 is in the
form of LMI. Therefore, it is easier to check using the readily available LMI
solver. For this type of system, it is better to use the Lyapunov-Krasovskii
Stability Theorem. On the other hand, if the system has a time-varying
delay, then Proposition 5.14 is no longer applicable, but the criterion in
Proposition 5.3 is still applicable as will be shown in Chapter 6.
To illustrate the criterion, we present the following numerical example.

Example 5.7 Consider the system

x(t) = (~2 -~.9) x(t) + f3 (=~ ~1) x(t - r),

where f3 ~ 0 and r > O. The parameter f3 to some extent measures the
size of the delayed term. A bisection scheme is used to calculate f3 max , the
maximum value of f3 the system can tolerate without violating the delay-
independent stability criterion in Proposition 5.14. The calculation results
in f3max = 0.9. This is the same as the result obtained using the Razumikhin
Theorem in Example 5.1. However, this is not always the case. For example,
for the system discussed in Example 5.5, the Razumikhin Theorem-based
method yields f3 max = 1.0425, while the Lyapunov-Krasovskii functional-
based method yields a slightly large bound f3 max = 1.0428. 0
 5.4 Delay-independent-Lyapunov-Krasovskii Theorem 171

5.4.2 Systems with distributed delays

Consider the system with distributed delays

x(t) = Aox(t) + lOr A(O)x(t + O)dO, (5.70)

where Ao is a given constant matrix and A( 0) is a continuous matrix valued

function given for 0 E [-r,Oj.
The Lyapunov-Krasovskii functional for the system can be chosen as

V(¢) = ¢T(O)p¢(O) + lOr [1°¢T(~)8(0)¢(~)d~] dO. (5.71)

The derivative of V can be calculated as

V(¢) = ¢T(O) [PAo+A5P+ lOr 8(0)dO] ¢(O)

+2¢T (0) l : P A( O)¢( O)dO

- lOr ¢T (0)8(0)¢(0)dO.

To facilitate further development, add and subtract a term involving the

relaxation matrix function R( 0) in the above, resulting in

V(¢) = ¢T(O) [PAo+A5P+lOrR(0)dO]¢(0)

+ lOr (¢T(O) ¢T(O)) ( ~~(O)pR(O) PA(O) )

-8(0)

( :~~~ ) dO. (5.72)

The derivative condition will be satisfied if the two matrices in the above
are negative definite.

Proposition 5.15 The system described by (5.70) is asymptotically stable

if there exist a real symmetric matrix

P > ° (5.73)
and real symmetric matrix functions R(O) and 8(0), such that

PAo + A5 P + lOr R(O)dO < 0, (5.74)

and
8(0) - R(O)
( AT(O)p PA(O))
-8(0) <,
°o E [-r, OJ. (5.75)
172 5. Systems with Single Delay

Proof. Use Proposition 5.2 and Lyapunov-Krasovskii functional (5.71).

Since (5.75) implies S(B) > 0, it is clear that V(q'» 2: cllq'>(0)112 for some
sufficiently small c > 0, the Lyapunov-Krasovskii functional condition (5.6)
is satisfied. In expression (5.72), the second term is always less than or equal
to zero due to (5.75). Also, (5.74) implies the existence of a sufficiently small
c > 0 such that
PA o + A6' P + lOr R(B)dB :s; -d.
Therefore, V(q'» :s; -cllq'>(0)1I 2, the Lyapunov-Krasovskii derivative condi-
tion is also satisfied. Therefore, the system is asymptotically stable accord-
ing to Proposition 5.2. •
Similar to Proposition 5.6, we can show that the above stability criterion
is delay-independent. Also, the corresponding result using the Razumikhin
Theorem (Proposition 5.5) can be obtained from the above result by intro-
ducing additional constraint

S( B) = a( B)P.

5.5 Delay-dependent stability criteria using a

simple Lyapunov-Krasovskii functional
5.5.1 Stability criteria using explicit model transformation
Simple delay-dependent stability criterion can also be derived with the
Lyapunov-Krasovskii Theorem, in parallel to the Razumikhin Theorem for-
mulation. Consider a system described by (5.1). Recall that we can use a
model transformation to obtain a system with distributed delays repre-
sented by (5.29) and (5.30). The stability of the system described by (5.29)
and (5.30) implies that of (5.1). We can apply Proposition 5.15 to obtain
the following Proposition.

Proposition 5.16 The system described by (5.1) is asymptotically stable

if there exist real symmetric matrices P, R o, R 1 , So, and SI such that

P > 0, (5.76)
-PA 1 A o -PA~)
-So 0 < 0, (5.77)
o -SI

where
1 T
M = -[P(A o + AI) + (Ao + AI) PJ + So + SI. (5.78)
r
 5.5 Delay-dependent-simple Lyapunov-Krasovskii functional 173

Proof. As is discussed above, it is sufficient to prove the stability of the

transformed system described by (5.29) and (5.30). Apply Proposition 5.15
and choose

R(O) { Ro,
Rb
-r ~ 0 < 0,
-2r ~ 0 < -r,

8(0) {80,81, -r ~ 0 < 0,

-2r ~ 0 < -r,

to obtain the sufficient condition consisting of (5.76) and

(5.79)

Divide (5.79) by r, and eliminate Ro and Rl from the resulting matrix

inequality and (5.80) to obtain (5.77) . •
Once again, the corresponding result in Corollary 5.8 can be obtained by
applying additional constraints

If one is to consider the stability of only the system (5.1), the above cri-
terion is clearly better than Proposition 5.8 since it is less conservative
and computationally more convenient because of the linearity of parame-
ters. The main value of Corollary 5.8 is again because it is also valid for
time-varying delay, as will be discussed in Chapter 6.

Example 5.8 Consider the same system discussed in Example 5.2. Using
Proposition 5.16, we can find that r max , the maximum delay for the system
to satisfy the condition, is r max = 1.000. This is less conservative than the
result in Example 5.2 obtained using the Razumikhin Theorem. However,
it is still very far from the true delay limit for stability r~~ytic = 6.1725.
o

5.5.2 Stability criteria using implicit model transformation

Similar to the case of Razumikhin Theorem-based methods, it is also possi-
ble to obtain less conservative stability criteria using implicit model trans-
formation. Consider again the system with single delay described by (5.1).
For real matrices XT = X, Y and ZT = Z satisfying

(5.81)
174 5. Systems with Single Delay

consider the quadratic Lyapunov-Krasovskii functional

where

cf>(~ + 0), - r ~ 0 ~ 0, - r ~ ~ ~ 0,
Aocf>(O) + A1cf>(-r).

Realizing

we can calculate

where

M = PAo+A~P+rA~ZAo +8,
cf>~r = (cf>T(O) cf>T(-r»).

Letting
T
T ·T
cf>oiJ = ( cf> (0) cf> (0) ) ,
we have

o ~ 1 0
-r
T(X Y)
cf>oiJ yT Z cf>oiJ dO
= rcf>T(O)Xcf>(O) + 2cf>T(0)y(cf>(0) - cf>(-r»
+ [Or ;l (O)Z¢(O)dO. (5.83)

Adding (5.83) and (5.82) using the constraint

(5.84)

we obtain

. T ( N PAl+rA~ZAl-y)A.
V(cf»~cf>or symmetric -8 + rA[ZAl 'l'Or'

where
N = P Ao + A~ P + r A~ ZAo + 8 + r X + y + yT,
 5.6 Complete quadratic Lyapunov-Krasovskii functional 175

from which we conclude that the system (5.1) is asymptotically stable if

there exist real matrices P = pT > 0, X T = X, Y, and ZT = Z satisfying

( N -Y)
(5.81) and
PAl +rA6'ZAI 0 (5.85)
symmetric -S+rAiZAl <.
To illustrate the connection between this result and the previous results
based on Lyapunov-Krasovskii functional methods, we eliminate the matrix
variable Z in (5.85) and (5.81) using Proposition B.6 in Appendix B to
obtain the following.

Proposition 5.17 The system described by (5.1) is asymptotically stable

if there exist real matrices XT = X, Y and
P>O,
such that
if
( Aip- yT (5.86)
-YAo
where
if = PAo +AoP+S+rX + Y + yT.
If we set Y = 0, we obtain the delay-independent stability criterion
discussed in Proposition 5.14.
On the other hand, if we choose Y = P AI, it can be easily shown that the
resulting criterion is equivalent to the delay-dependent stability criterion
discussed in Proposition 5.16 with the following variable transformation
from (S, X) to (So, SI):
S rSI,
X -~[P(Ao
r
+ AI) + (Ao + Alf Pj- So - SI·

Example 5.9 Consider the system discussed in Example 5.8. Using the
above stability criterion, it can be concluded that the maximum delay for
stability is r max = 4.3588. This is much less conservative compared to the
corresponding result obtained using explicit model transformation. 0

5.6 Complete quadratic Lyapunov-Krasovskii

functional

5.6.1 Introduction
From the numerical examples presented so far in this chapter, it is obvious
that there is still substantial conservatism for the simple delay-dependent
176 5. Systems with Single Delay

stability criteria discussed so far. Another serious limitation of these delay-

dependent stability criteria is that the system is always regarded as "less
stable" than the corresponding system with the delay set to zero. As is
discussed in Chapter 1, there are indeed practical cases where time delay
is intentionally introduced to stabilize the system or enhance the system
performance. A striking example is the following system

x(t) - O.l±(t) + x(t) = u(t). (5.87)

If the control u(t) is set to zero, the system is clearly unstable because of
"negative damping." Classic control theory would suggest using a deriva-
tive feedback u(t) = -kx(t). Indeed, a control gain k > 0.1 will make the
system stable by achieving positive damping. It is well known that deriva-
tive feedback is usually not easy to achieve. It seems reasonable to use the
following finite difference

x(t) - x(t - r)
u(t) = x(t - r) - x(t) = -r--'--'--'----'- (5.88)
r

to approximate -rx(t). Such an approximation seems reasonable as long

as r is sufficiently small. However, none of the methods discussed so far
in this chapter will be able to affirm the stability of such a system, since
setting the delay r = 0 voids the effect of the control u(t). It will be shown
later in the next section of this chapter that there is indeed a large interval
[r min, r max] with r min ~ 0.1 such that the system is asymptotically stable
if r falls into this interval.
In this section, we will enlarge the class of Lyapunov-Krasovskii func-
tional to complete quadratic form. It will be shown that the existence of
such a functional, which satisfies the Lyapunov-Krasovskii functional condi-
tion and derivative condition, is indeed a necessary and sufficient condition
for asymptotic stability of the system.
The development in this section, while much more complicated, has some
similarity to the Lyapunov equation

(5.89)

corresponding to the ordinary differential equation

x(t) = Ax(t). (5.90)

Recall that (5.89) has a positive definite solution for any given positive
definite matrix W if and only if the system (5.90) is asymptotically stable.
Indeed, if both P and Ware positive definite, we can choose Lyapunov
function
V(x) = xTpx,
 5.6 Complete quadratic Lyapunov-Krasovskii functional 177

and verify that its derivative along the trajectory of the system (5.90) is
V(x) = xT(PA+ATp)x
= -xTWx.
Furthermore, the solution in this case can be explicitly expressed as

P= 1 00
cpT (t)WCP(t)dt,

where
cp(t) = etA

is the fundamental solution of the system (5.90), which satisfies the equa-
tion
ci>(t) = Acp(t),
and the initial condition
cp(O) = I.
Indeed, it will be shown that this idea can be extended to time-delay sys-
tems in that it is still possible to prescribe the derivative expression, and
"solve for" the Lyapunov-Krasovskii functional.

5.6.2 Fundamental solution and matrix UW(T)

We will return to the discussion of the system with single defay
dx(t)
---;It = Aox(t) + AIX(t - r), (5.91)

where Ao and Al are constant n x n matrices and r is a positive delay.

We will assume that the system is asymptotically stable and try to obtain
a quadratic Lyapunov-Krasovskii functional. The solution of the system
under the initial condition
Xo =¢ (5.92)
will be denoted as x(t,¢). Recall from Chapter 1 that the fundamental
solution cp(t) satisfies the equation
d
dt cp(t) = Aocp(t) + AlCP(t - r), t ~ 0, (5.93)

and the initial condition

cp(O) = I, and cp(t) = 0 for t < O. (5.94)
Any solution of (5.91) under initial condition (5.92), x(t, ¢), can be ex-
pressed in terms of cp(t)

x(t, ¢) = cp(t)¢(O) + [Or cp(t - r - (})Al¢((})d(), for t ~ 0, (5.95)

178 5. Systems with Single Delay

which is a special case of (1.33) in Chapter 1. The fundamental solution

q,(t) can also be expressed as the inverse Laplace transform of 6,-1(8),

From (5.96), it can be easily concluded that q,{t) also satisfies the equation

d
dt q,(t) = q,(t)Ao + q,(t - r)A1' t ~ 0 (5.97)

and initial condition (5.94). Since the system (5.91) is asymptotically stable,
the stability exponent ao is negative. Therefore, we can find an a, ao <
a < 0, such that
11q,(t)11 ~ KeO'.t, for all t ~ 0 (5.98)
for some K > 1. In other words, the fundamental solution approaches zero
exponentially.
With the fundamental solution q,{t), for a symmetric matrix Wand real
scalar r, we can define the matrix Uw{r) as

Uw{r) = 1 00
q,T{t)Wq,{t+r)dt. (5.99)

This integration is well defined because q,(t) vanishes for t < 0 and ap-
proaches zero exponentially as t -+ +00. We will omit the subscript W
when no confusion may arise. For r ~ 0, with a transformation of integra-
tion variable ~ = t + r, and considering the symmetry of W, we write

UT(r) 10 00
q,T(t + r)Wq,(t)dt

1 q,T(~)Wq,{~
00
- r)~
= 10 00
q,T(~)Wq,(~ - r)d~.
In the last step, we have used the fact that q,(~ -r) vanishes when 0 < ~ ~
r. Thus we have arrived at the useful fact
U(-r) = UT(r). (5.100)

In particular U(O) is a symmetric matrix.

Differentiating ~T(t)Wq,(t) and using (5.97), we can write

:t (q,T(t)Wq,(t))
= q,T(t)Wq,(t)Ao + q,T(t)Wq,(t - r)A 1
+A6q,T(t)Wq,(t) + Afq,T(t - r)Wq,(t).
 5.6 Complete quadratic Lyapunov-Krasovskii functional 179

Integrating both sides of the above from 0 to 00, we arrive at another useful
fact
W + U(O)Ao + A5U(0) + UT(r)Al + AfU(r) = O. (5.101)
Differentiating (5.99) with respect to r using (5.97) and (5.100), we can
also write
d~ U(r) = U(r)Ao + UT(r - r)Al' r E [0, r]. (5.102)

With the above discussions, we can easily show that U(r), r E [0, r] can be
written as the solution of a two point boundary value problem as stated in
the following theorem.

Theorem 5.18 Matrix U(r) is the solution of the second order ordinary
differential equation

d2U(r) = dU(r) A _ ATdU(r) ATU()A _ ATU( )A (5.103)

dr2 dr 0 0 dr + 0 r 0 1 r 1,
with the boundary conditions

dU(r) 1 + dUT(r) 1 = -W, (5.104)

dr T=O dr T=O
d~~r) IT=o = U(O)Ao + UT(r)A 1 . (5.105)

Proof. Differentiating (5.102) and using (5.102) in the transposed form,

we can write
ddT
dr U(r)Ao + dr U (r - r)Al
d T T T
dr U(r)Ao - Ao U (r - r)Al - Al U(r)A 1 .

Substituting UT(r - r)Al in the second term on the right by (5.102), we

reach (5.103). (5.105) is (5.102) for r = O. (5.104) is obtained by using
(5.105) in (5.101) . •
If the system is asymptotically stable, such a two point boundary value
problem has a unique solution. Indeed, the definition of U(r) can be ex-
tended to unstable systems by considering the solution as the definition if
it exists and is unique.

5.6.3 Lyapunov-K rasovskii functionals

For a stable system (5.91), we will first try to find a quadratic Lyapunov-
Krasovskii functional V (¢) = Vw ( ¢) such that

vw(xt) = -xT(t)Wx(t), (5.106)

180 5. Systems with Single Delay

where
(5.107)

Let x(t,cjJ) be the solution of (5.91) with initial condition Xo = cjJ, then
it can be easily confirmed that

(5.108)

satisfies (5.106). We can express vw(cjJ) as an explicit quadratic functional

of cjJ using (5.95) and (5.99)

vw(cjJ)
cjJT (O)Uw (O)cjJ(O)

+2cjJT (0) lOr Uw( -r - B)Al cjJ( B)dB

+ lOr lOr cjJT(BdAfUw(Bl - (2)AlcjJ(B2)dBldB2. (5.109)

Based on vw(cjJ), we can easily construct a Lyapunov-Krasovskii func-

tional to satisfy
iJ(xd = -xT(t + r)Wx(t + r) (5.110)
for -r S r S O. This is accomplished by

(5.111)

With (5.106) and (5.109) to (5.111), we can take a linear combination to

achieve

V(Xt) = -xT(t)WIX(t) - xT(t - r)W3x(t - r)

- lOr xT(t + r)W2x(t + r)dr, (5.112)

by

v(Xt)

= VW t+r W2+ W3(Xt) + IOriO xT(t+B)W2X(t+B)dBdr

+ lOr xT(t + B)W3 X(t + B)dB

VWt+rW2+W3(Xt) + I: xT(t + B)[(r + B)W2 + W3lx(t + B)dB.

 5.6 Complete quadratic Lyapunov-Krasovskii functional 181

Or more explicitly,
v( ¢) = ¢T (O)U(O)¢(O)
+2¢T(0) lOr U(-r - O)AI¢(O)dO

+ l : lOr ¢T(OI)AfU(OI - (2)A I¢(02)dOId02

+ lOr ¢T(O)[(r + O)W2 + W3]¢(O)dO, (5.113)

where
U(O) = UWl+rW2+W3(O).
Indeed we can prove the following.

Theorem 5.19 If the system described by {5.91} is asymptotically stable,

then for any real symmetric matrices WI > 0, W 2 ::::: 0, and W3 > 0, the
Lyapunov-Krasovskii functional V(¢) = v(¢), with v(¢) defined in {5.113},
satisfies both the Lyapunov-Krasovskii functional condition
v(¢) ::::: cll¢(O)W (5.114)
and the Lyapunov-K rasovskii derivative condition
(5.115)

Proof. Condition (5.115) is obviously satisfied. To show (5.114), con-

sider, for sufficiently small c > 0 such that

H - ( WI + dAo + Al') CAl) 0

- cAr W3 >,
and let
V(¢) = v(¢) - c¢T(O)¢(O).
We have

! v(Xt) v(Xt) - 2cxT (t) [Aox(t) + AIX(t - r)]

-XT(t)WIx(t) - l : xT(t + ~)W2X(t + ~)d~

-xT(t - r)W3x(t - r) - 2cx T (t)[Aox(t) + AIx(t - r)]

-(xT(t) xT(t-r))H( x(t) )
x(t - r)

- lOr xT(t + ~)W2X(t + ~)d~

< -8I1x(t)1I2
182 5. Systems with Single Delay

for 0 < 8 < Amin(H). Integrate from 0 to 00, noticing lim x(t) = 0, we
t->oo
obtain

Therefore

This completes the proof. _

5.7 Discretized Lyapunov functional method for

systems with single delay
5.7.1 Introduction
From the discussions in the last section, in order to minimize the conser-
vatism in the stability problem of the system

±(t) = Aox(t) + AIX(t - r), (5.116)

we should consider a complete quadratic Lyapunov-Krasovskii functional

V(1)) = 1>T(O)P1>(O) + 21>T(0) lOr Q(~)1>(~)d~

+ lOr [lOr ¢,r(~)R(~,7])1>(7])d7]] d~
+ lOr ¢T(~)S(~)¢(~)d~, (5.117)

where
p = pT E IRnxn , (5.118)
and for all -r ::::: ~ ::::: 0 and -r ::::: 7] ::::: 0,

Q(~) E IR nxn , (5.119)

R(~, 7]) = RT(7],~) E IR nxn , (5.120)
S(~) ST(~) E IR nxn , (5.121)

and Q, R, and S are continuous. In practice, to numerically check the

existence of such a quadratic functional is not an easy task. In this sec-
tion, we will choose Q, R, and S to be piecewise linear functions. This
will allow us to write both the Lyapunov-Krasovskii functional condition
and Lyapunov-Krasovskii derivative condition in LMI form, allowing con-
venient and efficient numerical implementation. This method is known as
the discretized Lyapunov functional (DLF) method.
 5.7 Discretized Lyapunov functional 183

Considering (5.120) and (5.121), the derivative of Lyapunov-Krasovskii

functional along the trajectory of the system (5.116) may be calculated as
V(¢) = 2¢T(0)P[Ao¢(0) + A1¢(-r)]
+2[Ao¢(0) + A1¢( _r)]T lOr Q(~)¢(~)d~

+2¢T (0) lOr Q(~)¢(~)d~

+2 lOr d~ lOr ¢T(OR(~,,,,)¢(,,,)d,,,

+2 lOr ¢T(OS(~)¢(~)d~.

An integration by parts, considering (5.120) and (5.121), allows us to write

V(¢) in a standard quadratic form as follows:
V(¢) = -¢T(O)[-PAo-A~P-Q(O)-QT(O)-S(O)]¢(O)
_¢T (-r)S( -r)¢( -r)

- lOr ¢T(~)S(~)¢(~)d~
- lOr d~ lOr ¢T(~) [:~R(~,,,,) + :",R(~,,,,)] ¢(",)d",
+2¢T(0)[PAI - Q(-r)]¢(-r)

+2¢T (0) lOr [A~ Q(~) - Q(~) + R(O, ~)]¢(~)d~

+2¢T( -r) [o}AfQ(~) - R(-r, ~)]¢(~)d~. (5.122)

Recall that the system is asymptotically stable if and only if there exists a
Lyapunov-Krasovskii functional (5.117) such that the Lyapunov-Krasovskii
functional condition
(5.123)
and its derivative (5.122) satisfies the Lyapunov-Krasovskii derivative con-
dition
(5.124)
for some E: > O.

5. 7.2 Discretization
The basic idea of the DLF method is to divide the domain of definition of
matrix functions Q, R, and S into smaller regions, and choose these ma-
trix functions to be continuous piecewise linear, thus reducing the choice of
184 5. Systems with Single Delay

Lyapunov-Krasovskii functional V into choosing a finite number of para-

meters. Furthermore, it is possible to write the Lyapunov-Krasovskii func-
tional condition and Lyapunov-Krasovskii derivative condition in the form
of LMls. The techniques discussed in the Appendix B, especially variable
elimination and quadratic integral inequality, are especially useful.
Divide the delay interval L = [-r, OJ into N segments Lp = [Op,Op-IJ,
p = 1,2, ... , N
of equal length

h = rlN.
Then
p=0,1,2, ... ,N.

This also divides the square S = [-r, OJ x [-r, OJ into N x N small squares
Spq = [Op, Op-IJ x [Oq, Oq-IJ. Each square is further divided into two triangles

{(Op+ah,Oq+,8h)
o S; ,8 S; 1,
I OS;aS;,8 }
,

~q {(Op+ah,Oq+,8h) I }.
The continuous matrix functions Q(~) and S(~) are chosen to be linear
within each segment Lp , and the continuous matrix function R( ~, ".,) is cho-
sen to be linear within each triangular region ~~ or T),q. Let

Qp Q(Op),
Sp S(Op),
Rpq = R(Op,Oq).
Then, since these functions are piecewise linear, they can be expressed in
terms of their values at the dividing points using a linear interpolation
formula, i.e., for 0 S; a S; 1, P = 1,2, ... , N,

Q(Op + ah) Q(p)(a) = (1 - a)Qp + aQp-l, (5.125)

S(Op + ah) S(p)(a) = (1 - a)Sp + aSp_I, (5.126)

and for 0 S; a S; 1,0 S; ,8 S; 1, P = 1,2, ... , N, q = 1,2, ... , N,

R(Op + ah, Oq + ,8h)

R(pq) (a,,8)

{ (1 - a)Rpq + ,8Rp-I,q-1 + (a - ,8)Rp-I,q, a 2: ,8, (5.127)

(1 - ,8)Rpq + aRp-I,q-1 + (,8 - a)Rp,q_I, a <,8.

Thus the Lyapunov-Krasovskii functional is completely determined by P,

Qp, Sp, Rpq , p,q = 0,1, ... ,N.
 5.7 Discretized Lyapunov functional 185

5. 7.3 Lyapunov-Krasovskii functional condition

With the choice of piecewise linear Q, R, and S as (5.125) to (5.127), we
may divide the integration in (5.117) into segments [Op, Op-l], thus rewriting
V(¢) as

N N N N
V( ¢) = ¢T (O)P¢(O) + 2¢T (0) L VQP +L L VRpq + L Vsp, (5.128)
p=1 p=lq=1 p=1

where

VQP = 11 Q(p) (a)¢(p) (a)hda, (5.129)

VRpq = 11 [11 ¢(P)T(a)R(pq)(a,.B)¢(q)(.B)hd.B] hda, (5.130)

VsP = 11 ¢(p)T(a)S(p) (a)¢(p) (a)hda (5.131)

and
¢(P)(a) = ¢(Op + ah). (5.132)

We can show the following.

Proposition 5.20 The Lyapunov-Krasovskii functional V(¢) in (5.117)

to (5. 121}, with Q, S, and R piecewise linear as in (5.125) to (5. 127},
satisfies

V(¢) = 10r (¢T(O)

1
wT(a») (PQT Q)
R (¢(O») ~ VsP.
W(a) da + ~

(5.133)
Furthermore, it also satisfies

if
Sp > 0, p= 0,1, ... ,N. (5.135)

Therefore the Lyapunov-Kmsovskii functional condition (5.123) is satisfied

if (5. 135} and

( ~T
Q
q
R+S
~) >0 (5.136)
186 5. Systems with Single Delay

are satisfied. In the above

(5.137)

(5.138)

(5.139)

and

w(O)(a) "p(I) (a)

w(1)(a) "p(2) (a) + "p(1)(a)
W(2)(a) "p(3) (a) + "p(2) (a)
w(a) = (5.140)

W(N-l)(a)
W(N)(a)

"p(p) (a) h 10! q/p) ((3) d(3, (5.141)

"p(p) (a) h luI ¢(p) ((3)d(3, (5.142)

p= 1,2, ... ,N.

Proof. Once (5.134) are established, the sufficiency of (5.136) is easily

seen. Therefore we will proceed to prove (5.134). With the expression of
V in (5.128) to (5.132), for VQP we may carry out integration by parts to
obtain

VQP 11 Q(p) (a)¢(p)(a)hda

Q(p)(l)"p(P>(l) _11 [! Q(p) (a)] "p(p)(a)da

= Qp-l"p(p)(l) -1 1
(Qp-l - Qp)"p(p) (a)da

= 11 [Qp-l"p(p) (a) + Qp"p(p) (a)]da. (5.143)

 5.7 Discretized Lyapunov functional 187

Similarly, for VRpq, integration by parts with respect to {3 yields

VRpq 11 [1 1
¢(p)T(a)R(pq) (a, (3)¢(q) ({3)hd{3] hda

= 11 ¢(p)T(a) [R(Pq) (a, l)1/J(q)(l)

-1 1 (m(p;~a,{3)1/J(q)({3)d{3]hda

11 ¢(p)T(a) ([aRp_1,q-1 + (1- a)Rp,q-1)1/J(q) (1)

-1<> (Rp-1,q-1 - Rp-1,q)1/J(q) ({3)d{3

-i 1
(Rp,q-1 - Rpq)1/J(q) ({3)d{3 ) hda.

Further integrate by parts with respect to a

VRpq = 1/J(p)T(l) [Rp-1,q-11/J(q) (1) -1 1

(Rp- 1,q-1 - Rp-1,q)1/J(q) ({3)d{3]

-1 1
1/J(p)T(a)[(Rp_1,q_1 - R p,q-1)1/J(q) (1)

-(Rp-1,q-1 - Rp- 1,q)1/J(q) (a)

+(Rp,q-1 - Rpq)1/J(q) (a)]da

11 [1/J&,) (a)Rp-1,q-11/J(q) (a) + 1/J&,)(a)Rp-1,q1/J(q)(a)

+,¢(p)T(a)Rp,q_l'¢(q) (a) + '¢(P)T(a)Rpq'¢(q) (a)]da. (5.144)
Using (5.143) and (5.144) in (5.128) yields (5.133). Also, given (5.135), we
have

p=l
N
~ 1 1
¢(p)T(a)S(p)(a)¢(p)(a)hda

~1
N 1
= ¢(p)T(a)[(l_ a)Sp + aSp-1]¢(P) (a)hda

11 a¢(1)T(a)So¢(1) (a)hda + 11 (1- a)¢(N)T(a)SN¢(N)(a)hda

+~
N-1
1 1
[a¢(p+1)T(a)Sp¢(p+1) (a) + (1- a)¢(p)T(a)Sp¢(p) (a)]hda.
188 5. Systems with Single Delay

Using quadratic integral inequality (or Jensen Inequality, Corollary B.9 in

Appendix B) in each term above we obtain

N
~ l' [([
r
LVsP > ¢(1)({3)hd{3) T 80 ( [ ¢(I)({3)hd{3)
p=1

+ ( [ ¢IN)T ({3) hdli 8N ( [ ¢IN)T (Ii) hd{3») 1do

1 N-1 {1
+'h ~ 10 [10:1 ¢(p+1l(a)hda +
r]
10 ¢(pl(a)hda
T

Sp [1,1 ¢(p+ll(a)hda + 10: ¢(Pl(a)hda] da

= ~ 11 [1/J~l(a)So1/J(ll(a) + 1/J(NlT(a)SN1/J(Nl(a)

+~ }; (1)1>>+1) (0) + 1>lp) (O»T8 p (1>11>+1)(0) + 1>IP) (0»1do

11 'lIT(a)8'l1{a)da. (5.145)

Using (5.145) in (5.133) yields (5.134) . •

5.7.4 Lyapunov-K rasovskii derivative condition

With piecewise linear Q, R, and S as in (5.125) to (5.127), we have, for
(}p < ~ < (}p-b (}q < ", < (}q-b

1
S(~) = 'h(Sp-1 - Sp),
1
Q(~) = 'h{Qp-1 - Qp),
8R{~,,,,) 8R(~, 7]) 1
8~ + 87] 'h{Rp-1,q-1 - Rpq).

Therefore, after dividing the integration interval [-r,O] into discretization

segments [(}p, (}p-1], p = 1,2, ... , N, through tedious calculation, we can
 5.7 Discretized Lyapunov functional 189

verify that V(</» in (5.122) may be written as

V(</» = -</>5r b.</>or - ~

N
11
</>(p)T(a)Sdp</>(P)(a)da

-LL 11 </>(p)T(a)Rdpq</>(q)((3)dad(3
N N 1 1

p=1q=1 0 0

L 1[(1 - a)(D; + D;) + a(D; - D;)] </>(p) (a)da,

N 1
+2</>5r
p=1 0
or

V(</» -</>5r b.</>or -11¢t (a)Sd¢(a)da

-11[11 ¢T (a)Rd¢U3)da] d(3

11 +
+2</>5r [D S (1 - 2a)Da]¢(a)hda, (5.146)

where

</>Or = (</>(0)
</>(-r)
)
, (5.147)

_
</>(a) =.
C'(a) )</>2 (a)
, (5.148)

</>N (a)

b. ( b.oo b.01 ) (5.149)

b.61 b. l l '

~oo = -PAo - A6P - Qo - QT; - So, (5.150)

~01 = QN - PAl, (5.151)
~11 = SN, (5.152)

Sd = diag ( Sd1 Sd2 00. SdN ), (5.153)

Sdp Sp-1 - Sp, (5.154)

~ll
Rd12
( Rd21 Rd22 ~'N
Rd2N )
Rd = . , (5.155)

RdN1 RdN2 00' RdNN

Rdpq = h(Rp- 1,q-1 - Rpq), (5.156)
190 5. Systems with Single Delay

DS Div ), (5.157)

DpS (5.158)

Dop =
(5.159)

DIp = (5.160)

Da (5.161)
vap (5.162)

Vop (5.163)

Dfp = (5.164)

Proposition 5.21 For the Lyapunov-KrasovskiifunctionalV(¢) in (5.117)

to (5.121), with Q, S, and R piecewise linear as in (5.125) to (5.127), its
derivative V(¢) along the trajectory of the system (5.116) satisfies

for arbitrary matrix U satisfying

(5.166)

As a result, the Lyapunov-Krasovskii derivative condition (5.124) is satis-

fied if

(5.167)
 5.7 Discretized Lyapunov functional 191

Proof. From (5.146), it is easy to verify that

V(¢) = -1 1
( ¢6r[D S + (1- 2a)Da] Jt (a) )

( U -1) ( lD s + (1- 2a)Da]T ¢Or ) da

-1 Sd ¢(a)

-¢6r [~- DSUD sT - ~DaUDaT] ¢Or

_(1 ¢(a)da) T (1 ¢(a)da)

1
Rd
1
(5.168)

for arbitrary U. If (5.166) is satisfied, then use the Jensen Inequality (Propo-
sition B.8 in Appendix B) to the first term, we have

11 ( ¢6r[DS + (1 _ 2a)Da] ¢T (a) )

( U
-1
-1) ( ¢(a)
Sd
lDs + (1- 2a)DajT ¢Or ) da

> 11 ( ¢6r[DS + (1 _ 2a)Da] ¢T (a) ) da

( U
-1
-1) Jo{1 ( ¢(a)
Sd
lDs + (1 - 2a)Da]T ¢Or ) da

( ¢6r 1-T )(DSUDS T -DS)(¢or )

= Jo ¢ (a)da _DsT Sd J01¢(a)da .
Using the above in (5.168) yields (5.165). Therefore, if (5.166) and

(
~_lDaUDaT
_D s :;'
_Ds
R+Sd
)
> °
are satisfied, the Lyapunov-Krasovskii derivative condition (5.124) is sat-
isfied. The arbitrary matrix function U can be eliminated from the above
inequality and (5.166) using Proposition B.6 in Appendix B to yield (5.167) .
•
5.7.5 Stability criterion and examples
We can summarize from the above discussion the following.

Proposition 5.22 The system with single time delay described by (5.116)
is asymptotically stable if there exist n x n matrices P = pT; Qp, Sp = SJ,
p = 0,1, ... ,N; Rpq = Rfp, p = 0,1, ... , N, q = 0,1, ... , N such that (5.136)
and (5.167) are satisfied, with notations defined in (5.137) to (5.139) and
(5.149) to (5. 164}.
192 5. Systems with Single Delay

Proof. Choose Lyapunov-Krasovskii functional (5.117) to· (5.121), with

Q, S, and R piecewise linear as in (5.125) to (5.127). Then (5.167) implies
that the Lyapunov-Krasovskii derivative condition is satisfied according to
Proposition 5.21.
To show that the Lyapunov-Krasovskii functional condition is satisfied,
we only need to show that (5.135) is satisfied, which is the only condition
needed in addition to (5.136) according to Proposition 5.20. Since SN =
Ll1t. Sp-1 - Sp = Sdp, P = 1,2, ... , N are all principal minors of the left
hand side of (5.167), we can conclude

SN > 0,
Sp-1 - Sp > 0, p = 1,2, ... ,N.
This implies
So> Sl > ... > SN > 0,
which implies (5.135).
Since both the Lyapunov-Krasovskii functional condition and Lyapunov-
Krasovskii derivative condition are satisfied, the system is asymptotically
stable. _
Clearly the criterion is in the form of LMI since all the system parameters
appear linearly. The following examples are provided to show the effective-
ness of the DLF method. The first example illustrates the convergence to
the analytical limit as N increases.

Example 5.10 Consider the system

. (-20 -0.90) x(t) + (-1-1 -10) x(t - r).

x(t) =

We want to calculate the maximum time delay rmax the system can tolerate
and still retain stability. Proposition 5.22 can be used to check the stability
of the system for any given r. A bisection process is used to estimate r max
for different N. The analytical stability limit is also listed for comparison.
The results are very close to the analytical limit. Even the result for N = 1
is sufficient for most practical applications, and is much better than the
ones obtained in earlier sections. The convergent trend to the analytical
limit is obvious. 0

Analytical
6.17258

Example 5.11 Consider the system

x(t) - O.l±(t) + 2x(t) - x(t - r) = O.

 5.8 Notes 193

This is obtained from the system (5.87) by applying feedback control (5.88).
This system is clearly unstable for r = O. To apply Proposition 5.22, the
system is written in the standard form

d ( x(t) ) ( 0 1 ) ( x(t) )
dt x(t) -2 0.1 x(t)

+ ( 01 0)
0
( x(tx(t -- r)r) ) .
The system is stable for r E (rmin' rmax). Proposition 5.22 with different N
is used to estimate rmin and rmax. The computation consists of sweeping
through a range of r with relatively large step size, and a bisection process
near the lower limit rmin and the upper limit rmax. The result is listed in
the following table along with the analytical limit. 0
N 1 2 3 Analytical
rmin 0.1006 0.1003 0.1003 0.1002
rmax 1.4272 1.6921 1.7161 1.7178

5.8 Notes
5.B.l Results based on the Razumikhin Theorem
An example of the results prior to the modern LMI approach is by Thowsen
[260]. There is a large amount of work done on this topic. The delay-
independent formulation in Proposition 5.3 and the dual form is equivalent
to Proposition 18 in Niculescu et al. [214]. The delay-dependent formulation
in Proposition 5.7 is equivalent to Proposition 19 in Niculescu et al. [214].
Some studies try to avoid the scaling factor and write the stability criteria
in LMI form. For example, it can be shown, using variable elimination
of LMI, that Li and de Souza [173] effectively choose the scaling factors
0:1 = 0:2 = 1 when specialized to the case of Proposition 5.7. It seems
clear that fixing the scaling factors are essential if one is to restrict the
formulation to LMI form. However, it is also essential that these scaling
factors should be proportional to the magnitude of the system matrices Ao
and AI. For example, if Ao and Al are multiplied by 1000 and the delay
r is divided by 1000 (such as choosing a different unit, say second versus
millisecond), this accelerates the evolution of the system without affecting
the stability of the system. A stability criterion with the scaling factor not
proportional to Ao and Al would conclude that the stability of these two
system may be different. The scaling scheme proposed here is from Gu and
Han [94].
The Razumikhin Theorem-based formulation with implicit model trans-
formation described in (5.53) and (5.54) was first proposed in Park [221] for
194 5. Systems with Single Delay

the special case of a = ao = a1. The forms presented here with connections
to other criteria is from Gu [92].

5.8.2 Model transformation and additional dynamics

A number of other model transformations are used in practice, see Kol-
manovskii and Richard [158].
The nonequivalence of transformed system and original system in general
was pointed out by Verriest [271] in the context of reducing a class of sys-
tems with distributed delays to pointwise delays. The concept of additional
dynamics was proposed by Gu and Niculescu [97] based on decomposition
of a characteristic quasipolynomial, where the proofs not included here can
be found. Kharitonov and Melchor-Aguilar [140] extended the additional
dynamics to systems with multiple delays. Additional dynamics in other
types of model transformations was discussed in Gu and Niculescu [98].
Kharitonov and Melchor-Aguilar [144] wrote the additional dynamics as
an explicit dynamic equation as in (5.43) and extended to systems with
time-varying delays and coefficients.

5.8.3 Lyapunov-Krasovskii method

It should be noted that the delay-independent criterion using the Lyapunov-
Krasovskii functional method presented here is not a necessary and suffi-
cient condition, see Bliman [13]. There is also a large amount of work done
on delay-independent results and delay-dependent stability based on sim-
ple Lyapunov-Krasovskii functional (see Niculescu et al. [214]). The delay-
dependent stability result here can be found in Niculescu et al. [212]. The
particular derivation for simple delay-dependent stability based on explicit
model transformation is from Gu and Han [94].
A stability criterion based on implicit model transformation was pro-
posed by Park [220], which is the special case of (5.81) and (5.85) with

( yTX y)
Z =
(P~ArV
+ WT) rV -1 (
P+ W
1
TV
A)
1 .

The formulation without such a constraint was given in Park [221] with
extensions to considering uncertainty and cost function by Lee, Moon, and
Kwon [165]. The elimination of variable with connections to other delay-
independent and delay-dependent stability criteria was given by Gu [92].
The idea of a complete quadratic Lyapunov-Krasovskii functional was
proposed by Repin [236]. The special case of x being scalar was solved by
Datko [48]. The case of W 2 = W3 = 0 was solved by Infante and Castelan
[122]. Kharitonov and Zhabko [145] proposed the case W2 =I 0 and W3 =I 0
in studying the robust stability.
 5.8 Notes 195

The complete quadratic Lyapunov-Krasovskii functional is also used in,

e.g., Fridman and Shaked [66,67].
The idea of the DLF method was proposed by Gu [83]. Refinements
using variable elimination [87] and integral inequality (Jensen Inequality)
[88] were proposed to improve the method. The formulation presented is
based on [89], where an alternative condition for the Lyapunov-Krasovskii
functional condition was also discussed in addition to the one discussed
here.
6

Robust Stability Analysis

6.1 Introduction
In this chapter we will discuss the robust stability problem of time-delay
systems using the time-domain approach. The robust stability problem
considers the stability problem of systems that contain some uncertainties.
As is well known, it is usually impossible to describe a practical sys-
tem exactly. First, there are often parameters or parasitic processes that
are not completely known. Second, due to the limitation of mathematical
tools available, we usually try to use a relatively simple model to approxi-
mate a practical system. As a result, some aspects of the system dynamics
(known as unmodeled dynamics) are ignored. Third, some control systems
are required to operate within a range of different operating conditions. To
capture these uncertain factors, it is often possible to identify a bounding
set such that all the possible uncertainties fall within this set and yet it is
not too difficult to analyze mathematically.
In this chapter we will first discuss some useful ways of characterizing un-
certainties, including subpolytopic uncertainty and norm-bounded uncer-
tainty. It is often desirable to refine the norm-bounded uncertainty to be of
block-diagonal structure to reduce conservatism. Then we will discuss the
treatment of robust stability problem with either generic uncertainty form
or specific uncertainty form. The Razumikhin Theorem, simple Lyapunov-
Krasovskii functional, or complete quadratic Lyapunov-Krasovskii func-
tional will be used to derive delay-independent or delay-dependent stability
results.

6.2 Uncertainty characterization

We will discuss the uncertainty characterization for the systems with single
delay. Consider the system
i:(t) = Ao(xt, t)x(t) + A1(xt, t)x(t - r),
where Ao E lR. nxn , Al E lR. nxn are uncertain coefficient matrices not known
completely, except that they are within a compact set n which we will refer
to as the uncertainty set,
(Ao(xt, t), A1(xt, t)) En for all t ~ o.
198 6. Robust Stability Analysis

The uncertainty set characterizes the uncertainties and serves as basic in-
formation needed to carry out robust stability analysis. Notice also that the
coefficients may depend on the time t as well as the current and previous
state variable x(t+~), -r ::; ~ ::; O. For the sake of convenience, we will not
explicitly show these dependences or only show the dependence on time t
when no confusion may arise.
A good choice of uncertainty set is a compromise between minimizing
conservatism (and therefore, it is desirable to make the uncertainty set
"small") and the mathematical tractibility (and therefore, it is desirable
to make the uncertainty set structurally simple). In this section, we will
discuss some common uncertainty structures.
In addition to uncertainties in the coefficient matrices, the time delay r
may also involve uncertainties. In this chapter we will mainly discuss the
case where r = r(t) may be time-varying and is known to be within the
interval r(t) E (0, rmax]. However, we will reserve the term "uncertainty"
to coefficient uncertainty.

6.2.1 Polytopic uncertainty

The first class of uncertainty frequently encountered in practice is the poly-
topic uncertainty. In this case, there exist, say, nv elements of the uncer-
tainty set 0,
W(k) = (A~k), A~k)), k = 1,2, ... , n v ,
known as vertices, such that 0 can be expressed as the convex hull of these
vertices
0= CO{w(k) I k = 1,2, ... ,nv }'
In other words, the uncertainty set 0 consists of all the convex linear com-
bination of the vertices

For example, in practice, there are often some uncertain parameters in

the system. Each uncertain parameter may vary between a lower limit and
an upper limit, and these uncertain parameters often appear linearly in
the system matrices. In this case, the collection of all the possible system
matrices form a poly topic set. The vertices of this set can be calculated
by setting the parameters to either lower or upper limit. If there are np
uncertain parameters, it is easy to see that there are nv = 2np vertices.
To see this idea more clearly, assume that there are two uncertain para-
meters, a and {3, varying between lower and upper bounds
 6.2 Uncertainty characterization 199

They appear linearly in the system matrices

Ao Aon + aAOa + f3 A o,e,

Al A ln + aAla + f3A 1,e,

or in the abbreviated form

W = Wn + aWa + f3w,e.
Then, the uncertainty set possesses four vertices

W(I)
= Wn + aminWa + f3 mi nw ,e,
w(2)
= Wn + amaxWa + f3 minW,e,
w(3)
= Wn + aminWa + f3 max w,e,
w(4) Wn + amaxWa + f3 maxW,e.

6.2.2 Subpolytopic uncertainty

The class of subpolytopic uncertainty is more general than polytopic un-
certainty. In this case the uncertainty set also possesses, say, nv vertices,
and the uncertainty set n is contained in (but not necessarily equal to) the
convex hull of the vertices

He
r\ co {W( i ),2, -- 1, 2 , ... , nv }.

In other words, any element wEn can be expressed as a convex linear

combination of the vertices, i.e.,
nv
W = Lf3iw(i),
i=1

for some scalars f3 i satisfying

nv
f3i ~ 0, 'Lf3i = 1.
i=l

Clearly, polytopic uncertainty is a special case of subpolytopic uncertainty.

An important class of subpolytopic uncertainties arises from uncertain
parameters that appear multilinearly in the system matrices, as the follow-
ing proposition shows. In such cases, the products of different parameters
may appear in the system matrices. Each parameter appears linearly if all
the other parameters are held constant.
200 6. Robust Stability Analysis

Proposition 6.1 The set 0 formed by np parameters, ai, i = 1,2, ... , np,
appearing multilinearly, with each parameter varying between a lower limit
and an upper limit

is subpolytopic. The nv = 2np vertices Clm be calculated by setting each

parameter ai to either the lower limit ai min or the upper limit ai max for
i=1,2, ... ,np.

Proof. We will use induction. For np = 1, it is obvious. Assume the

proposition is valid for np = m. Then for np = m + 1,

o -- {W
m,n
+ a m+l Wm,v I a(wm +l,min:$ a m+l :$ a m+l,maJ"
w ) E 0m
m,n, m,v

where Om is a set of multilinear expression of m parameters ai, i =

1,2, ... , m with each parameter varying between the lower limit and the
upper limit. According to the inductive assumption, Om is subpolytopic,
and all the vertices can be obtained by letting each of the m parameters as-
sume the lower and the upper limit. Denote these vertices as (W~n,W~v),
i = 1,2,3, ... ,2m. Then, for any (Wm,n,wm,v) E Om, we can write

for some
2m
f3i ~ 0, :L:f3i = 1.
i=l

This indicates that any element W of 0 can be expressed as

But a m +1,min :$ a m +l :$ a m +l,max can be written as

 6.2 Uncertainty characterization 201

Therefore,

i=l

Since the 2m +1 coefficients Di = f3i'Y ~ 0, D2"'+i = f3 i (1 - ')') > 0, i

1,2, ... ,2m , sum up to be 1
2",+1 2m 2"' 2'"

L Di = Lf3i(l- ')') + Lf3i'Y = Lf3i = 1,

i=1 i=1 i=1 i=1

we can conclude that n is subpolytopic with 2m +1 vertices

i)
(Wm,n + (lm+l,mmWm,v,
. (i») (Wm,n
(i) + (lm+l,maxWm,v,
(i) ) Z. -- 1, 2 , ... , 2m ,

which can be obtained by letting each parameter assume its lower and
upper limit. •
As will be seen later, subpolytopic uncertainties are easy to treat in
time-domain stability analysis, since most stability criteria are in the form
of matrix inequalities with system matrices appearing linearly. Such ma-
trix inequalities are satisfied for all the possible system matrices in the
uncertainty set if and only if they are satisfied at the vertices.
There are, however, still practical difficulties. For example, for the case of
multilinear uncertain parameters, there are nv = 2np vertices of uncertainty
set if there are np uncertain parameters. The number nv can be very large
for a rather modest np. For example, np = 10 uncertain parameters will
result in nv = 1024. This phenomenon, known as combinatorial explosion,
limits the practical applications of the subpolytopic uncertainty model to
only a few uncertain parameters. There are also other uncertainties that
cannot be reasonably modeled by a subpolytopic uncertainty set with a
reasonable number of vertices. In such cases, norm-bounded uncertainties
are often used.

6.2.3 Norm-bounded uncertainty

In terms of norm-bounded uncertainty, we decompose the system matrices
W = (Ao, AI) into two parts: the nominal part wn = (A on , A In ) and the
uncertain part t::.w = (t::.Ao, t::.AI):

W = Wn + t::.w,
202 6. Robust Stability Analysis

i.e.,

Ao Aon + .6.Ao,
Al = Aln + .6.AI.
The uncertain part is written as

(6.1)

where E, Go, and G I are known constant matrices. F is an uncertain matrix

satisfying
IIFII ~ 1. (6.2)
In other words, the uncertainty set n can be expressed as

A slightly more general uncertainty is the linear fractional norm-bounded

uncertainty (or LF norm-bounded uncertainty), in which the uncertain
matrices in (6.1) are replaced by

and F satisfies (6.2). Clearly, the norm-bounded uncertainty is a special

case of LF norm-bounded uncertainty when D = o. Of course, for the LF
norm-bounded uncertainty to be well defined, the matrix I - F D has to
be invertible for arbitrary F satisfying (6.2). It is not difficult to see that
this is equivalent to requiring

I _DTD > O. (6.4)

We will say that the uncertainty description (6.3) is well posed if (6.4)
is satisfied. The LF norm-bounded uncertainty can be motivated by the
following system

x(t) Aonx(t) + AInX(t - r) + Eu(t), (6.5)

y(t) Gox(t) + GIx(t - r) + Du(t), (6.6)

subject to uncertain feedback

u(t) = F(t)y(t), (6.7)

or equivalently, in view of (6.2),

Ilu(t)112 ~ Ily(t)112.
 6.2 Uncertainty characterization 203

From this setting, it may appear at first sight that the bounding (6.2) is
rather restrictive. For example, it is often desirable to model the uncertain
matrix F in (6.3) to satisfy

FPF T :::;Q (6.8)

instead of (6.2) where P and Q are symmetric positive definite matrices

introduced to reflect different scales of the uncertainty components. How-
ever, this can be easily transformed into the standard setting as follows.
Let U and V be matrices satisfying

VVT P,
UTU Q-l,

and let

F* UFV, (6.9)
E* EU- l , (6.10)
Go = V-lGO, (6.11)
Gi V-lGl, (6.12)
D* V-lDU- l , (6.13)

then (6.8) and (6.3) are equivalent to

IIF*II = IluFVl1 :::; 1,

and
(~Ao ~Al) = E*(I - F* D*)-I F* (Go Gt).
On the other hand, we can, of course, consider another possible situation
where the uncertainties ~Ao and ~AI are independent of each other, i.e.,

~Ao = EoFoGo, (6.14)

~Al = ElFIG!, (6.15)
with

IlFoll < 1, (6.16)

11F111 < 1. (6.17)

In handling norm-bounded or LF norm-bounded uncertainty, the follow-

ing lemmas prove very useful.

Lemma 6.2 For any real matrices E, G and real symmetric positive defi-
nite matrix P, with compatible dimensions,
204 6. Robust Stability Analysis

Proof. Since
(EPE T + GT P-1G) - (EG + GT ET)
= (EP - GT)p-l(EP _ GT)T
> 0,
the conclusion is easy to see. _

Lemma 6.3 Let DE IRpxq , E E IRnxm , G E IRPxn , and P = pT E IRnxn

and D satisfies
(6.18)
The inequality
-P + E(I - XD)-l XG + [E(1 - XD)-l XGl T < 0 (6.19)
is satisfied for any real matrix X E IRqxp satisfying

IIXII ~ 1 (6.20)
if and only if there exists a real scalar). > 0 such that
P + )'GT G
( -ET E + )'GT D )
+ )'DTG -).(I - DT D) < 0, (6.21)

or equivalently,

Proof. In view of (6.18), the inequality (6.19) is well posed and is equiv-
alent to
-ep~ +2eE(I _XD)-lXG~ < 0 (6.23)
for all ~ =f O. The inequality (6.23) can be written as

-e P~ + 2e Ery < 0, (6.24)

where
ry = (I - XD)-l XG~, (6.25)
or
ry = X(Dry + G~). (6.26)
Equation (6.26) is valid for all X satisfying (6.20) if and only if
ryT ry ~ (Dry + G~f(Dry + G~). (6.27)
Therefore, (6.19) subject to (6.20) is equivalent to (6.24) subject to (6.27).
Using S-procedure (Proposition B.3 in Appendix B), this is equivalent to
the existence of a). > 0 such that (6.21) is satisfied. This is equivalent to
(6.22) according to the Schur complement. _
For the special case of D = 0, we have the following.
 6.2 Uncertainty characterization 205

Lemma 6.4 For E E jRnxm, G E jRPxn, and P = pT E jRnxn, the in-

equality
-p + EXG + [EXG]T < 0 (6.28)
is satisfied for all the real matrices X E jRqxp satisfying

IIXII ::; 1 (6.29)

if and only if there exists a real scalar oX > 0 such that

_p+oXGTG E )
( ET -AI < 0, (6.30)

or equivalently
(6.31)

In the above two lemmas, if the matrices P, E, G, and D depend on a

parameter p, then oX also depends on p. We may restrict oX to be a constant
independent of p, in which case, (6.21), (6.22), (6.31), and (6.30) become
sufficient but not necessary conditions.

6.2.4 Block-diagonal uncertainty

Block-diagonal uncertainty represents a very general class of uncertainty.
In our context, a system with block-diagonal uncertainty can be expressed
as
m

x(t) Aonx(t) + AlnX(t - r) +L Ejuj(t), (6.32)

j=l
m
Yi(t) = GOix(t) + Glix(t - r) +L DijUij(t), i = 1,2, ... , m,(6.33)
j=l

subject to uncertain feedback

Ui = FiYi, i = 1,2, ... ,m. (6.34)

We can properly scale the uncertain matrices Fi, i = 1,2, ... , m similar to
(6.9) to (6.13) such that they satisfy

IlFill ::; 1, i = 1,2, ... , m. (6.35)

Indeed it is often possible to extract all the uncertain elements from the
system and consider them as feedback. Then the system with all the un-
certain elements removed can be regarded as a multi-input-multi-output
206 6. Robust Stability Analysis

system. This process allows us to write the uncertain system in the stan-
dard block-diagonal form (6.32) to (6.35) and is known as "pulling out
uncertainties."
Equations (6.32) to (6.34) can be written as (6.5) to (6.7), with

(6.36)

which justifies the name "block-diagonal uncertainty." Notice also that the
block-diagonal uncertainty includes as a special case the independent un-
certainty case described by (6.14) to (6.17), since (6.14) and (6.15) can be
written as

Due to (6.35), F clearly satisfies (6.2). However, the analysis based on

(6.2) does not take advantage of the special structure of the uncertainty
and may lead to overly conservative stability estimate. Let

KI = diag (klIll k2121 (6.37)

Kr diag (klhr k212r (6.38)

where Iii and Iir are the identity matrices with the dimensions the same
as the number of rows and columns of F i , respectively. Then an important
observation is that
KIFK;l = F
for arbitrary ki' i = 1,2, ... , m. Therefore, we can rewrite (6.7) as
U = KIFK;ly,
or
(6.39)
where

Therefore, the system can be written as

x(t) Aonx(t) + AlnX(t - r) + EK1uK(t), (6.40)

YK(t) K;lGox(t) + K;lG1X(t - r) + K;l DK1uK(t), (6.41)
and (6.39). We can therefore apply the norm-bounded or LF norm-bounded
uncertainty criteria to the system described by (6.40), (6.41), and (6.39)
and
IIFI! ::; 1.
 6.3 Robust stability-Razumikhin Theorem 207

If the stability criterion is satisfied for some ki' i = 1,2, ... , m, then the
system is stable. Many stability criteria can be written as LMI with either
k? or 11k? appearing linearly. The possibility of searching through all the
possible scaling factors significantly reduces the conservatism. It has been
argued that this scaling process allows one to reach "very close to the
optimum" in terms of reducing conservatism.

6.3 Robust stability criteria based on the

Razumikhin Theorem
In this section we will discuss the robust stability problem using the Razu-
mikhin Theorem. Similar to the systems without uncertainty, the condi-
tions will be closely related to the conditions obtained using the Lyapunov-
Krasovskii functional discussed later in this chapter. To avoid repetition,
we will only discuss the generic uncertainty set. The development on spe-
cific uncertainty form (such as polytopic uncertainty and norm-bounded
uncertainty) will be carried out in the context of the Lyapunov-Krasovskii
functional method. The corresponding results based on the Razumikhin
Theorem can be obtained by a simple substitution from the corresponding
results based on the Lyapunov-Krasovskii functional method.

6.3.1 Delay-independent stability for systems with single delay

Consider the system
±(t) = Ao(t)x(t) + A 1 (t)x(t - r(t», (6.42)
where
(Ao(t), Al (t)) E n for all t > 0, (6.43)
and the time delay is uncertain and possibly time-varying, within a known
interval
0< r(t) ::; rmax. (6.44)
The matrices A o, Al and delay r may also depend on the state Xt. How-
ever, we will suppress the notation for the dependence on Xt for the sake
of convenience. Sometimes we will also suppress the notation for the de-
pendence on time t. The uncertainty set n is compact (which is equivalent
to closed and bounded in this finite dimensional context). We will use the
Razumikhin Theorem to study the stability of the system. Let
V(X) = xTpx, (6.45)
then
V(x(t) = 2xT(t)P[Aox(t) + AIX(t - r)].
208 6. Robust Stability Analysis

The Razumikhin Theorem indicates that the system is asymptotically sta-

ble if there exists a scalar p > 1 and matrix P > 0 such that

V(x(t)) :::; -cllx(t)112 (6.46)

whenever

V(x(t + 0)) < pV(x(t)) for all - rmax :::; 0 :::; O. (6.47)

With (6.47), then for any a ;::: 0,

V(x(t)) < 2xT(t)P[Aox(t) + A1X(t - r)]

+a(pxT(t)Px(t) - xT(t - r)Px(t - r)]
= ",T ( P Ao + Air P + apP PAl ) '" (6.48)
'f'Or Ai P -aP 'f'Or'

where
¢Or = (xT(t) xT(t - r) )T.
We can therefore conclude the following.

Proposition 6.5 The system described by (6.42) to (6.44) is asymptoti-

cally stable if there exist a scalar

a> 0,

and a symmetric matrix P such that

( P Ao + Air P + aP PAl ) < 0 (6.49)

Aip -aP

for all

Proof. Similar to the case without uncertainty discussed in Chapter 5 .

•
The stability condition above is delay-independent since the stability
criterion is independent of the delay bound rmax. Similar to the case of
systems without uncertainty, we will see that the condition (6.49) can be
obtained from the corresponding condition using the Lyapunov-Krasovskii
functional by imposing a constraint. Notice that (6.49) generally represents
an infinite number of matrix inequalities since n contains an infinite number
of points. It is therefore unrealistic to use Proposition 6.5 directly to check
stability. There are many cases, such as polytopic uncertainty and norm-
bounded uncertainty, where we can reduce to a finite number of matrix
inequalities. We will illustrate these ideas in the section dealing with the
Lyapunov-Krasovskii functional methods.
 6.3 Robust stability-Razumikhin Theorem 209

6.3.2 Delay-independent stability criteria for systems with

distributed delays
Consider a system with distributed delays

±(t) = Ao(t)x(t) + 10
-r(t)
A(t, B)x(t + r(t,B))dB, (6.50)

with
(Ao(t),A(t,·)) E Or for all t 2: 0, (6.51)
°< r(t, B) :::; rmax for - r :::; £I < 0, (6.52)
where Or is compact. For the sake of convenience, we will often suppress
the notation of the explicit dependence of time t when no confusion may
arise. It is possible to write the system in a simpler form without sacrificing
the generality. For example, by reparameterizing £I, one may easily change
the integration limit in (6.50) to [-1,0]. However, since the main purpose
of studying such a system with distributed delay here is to derive the delay-
dependent stability criteria for systems with a single delay, we choose to
keep the formulation as close to the original form as possible. The study of
distributed delay in its own right will be conducted in Chapter 7.
We will study this system again using Lyapunov function V(x) as in
(6.45). The system is asymptotically stable if (6.46) is satisfied whenever
(6.47) is satisfied. For any a(t, B) 2: 0, -r(t) :::; £I :::; 0, when (6.47) is
satisfied,

V(x(t)) 2x T (t)P [AoX(t) + 10

-r(t)
A(B)x(t + r(t, B))dB]

< 2X T (t)P[AoX(t)+jO A(B)x(t + r(t,B))dB]

-r(t)

+10 a(t,B)[pxT(t)Px(t)
-r(t)
_x T (t + r(t, B))Px(t + r(t, £I))] dB
0 T ( M(t,B) PA(t,B))
= 1-r(t) ¢O(J AT(t, B)P -a(t, B)P ¢O(JdB, (6.53)

where

M(t, B) = ~[PAo(t)
r
+ A6(t)P] + a(t, B)pP + R(B, Ao(t),A(t,')) (6.54)

and
_ ( x(t) )
¢0(J- x(t+r(t,B)) .
210 6. Robust Stability Analysis

The matrix function R((), Ao(t), A(t,·)) (which we will write R(()) for the
sake of convenience) is constrained to satisfy

/
0 R( ())d() = O. (6.55)
-r(t)

From the expression of V(x(t)) we may conclude the following.

Proposition 6.6 The system with distributed delays described by (6.50) to

(6.52) is asymptotically stable if there exist a symmetric matrix P, scalar
function
a(t, ()) > 0, - r(t) ~ () ~ 0,
and symmetric matrix function R(()) = R((), Ao(t), A(t, .)) satisfying the
constraint (6.55), such that

( N(()) PA(()) ) (6.56)

AT(())P -a(t,())P < 0

for all-r(t) ~ 0 ~ 0 and (Ao,A(·)) E Or, where

N(O) = _l_[PAo + A6' PJ + a(t, O)P + R(O).

rmax

Proof. If (6.56) is satisfied, the (2,2) entry ofthe matrix in (6.56) implies

P>O, (6.57)

which in turn implies

(6.58)
Also, the (1,1) entry implies

N(O) < O.
Using (6.55), the above implies

/
0 N(O)dO
-r(t)

r(t) [PAo + A6' PJ + r(t)a(t, O)P

rmax
< O.
In view of (6.57), the above implies
PA o +A6'P < O.
Therefore, if we set p = 1, we have
N(O) ~ M(O).
 6.3 Robust stability-Razumikhin Theorem 211

The above and (6.56) imply

( M(O) PA(t,O) ) <0

AT(t,O)P -a(O)P
for p = 1. Due to continuity and the compactness of nr, it is also valid for
p = 1 + {j > 0 if {j > 0 is sufficiently small. Therefore,
V(¢(O)) ~ -cll¢(0)11 2 (6.59)

in view of (6.53). We can conclude that the system is asymptotically stable

according the Razumikhin Theorem in view of (6.58) and (6.59) . •
It is not difficult to see that condition (6.55) can be relaxed to

1 0
-r(t)
R( O)dO "? 0 (6.60)

and R can be chosen to be independent of the uncertain matrices Ao(t)

and matrix function A(t, .). However, it will be seen that keeping R as a
general matrix function of the uncertainty is essential in deriving a simple
delay-dependent stability criterion.

6.3.3 Delay-dependent stability criteria with explicit model

transformation
Consider again an uncertain system described by (6.42) to (6.52). Here we
will attempt to derive a simple delay-dependent stability criterion using
the Razumikhin Theorem. Using

x(t - r(t)) = x(t) -1 0

-r(t)
x(t + O)dO,

with x(t + 0) replaced by the right hand side of the system equation (6.42)
with appropriate time shift, we can obtain the transformed system

x(t) = [Ao(t) + Al (t)]x(t)

-AI(t) 1 0

-r(t)
[Ao(t + O)x(t + 0)
+AI(t + O)x(t + 0 - r(t + O))]dO,
which can be written as

x(t) = Ao(t)x(t) + 1 0

-2r(t)
A(t, O)x(t + T(t, O))dO, (6.61)

(Ao(t), A(t, .)) E 02r, t "? 0,

212 6. Robust Stability Analysis

where
Ao = Ao +Al }
A(O) = -AlAoo, A( -r + 0) = -AlAlO, -r ~ 0 < 0 ,
for (Ao,Ad E n and (Aoo,AlO) En, 0 E [-r,O]
(6.62)
and
0, -r(t) ~ 0 < 0,
r(t,O) = { 0 _ r(t + B), -2r(t) ~ 0 < r(t).
This model transformation, of course, also introduces additional dynamics.
These additional dynamics are more complicated and will not be discussed
here. Applying Proposition 6.6, we can obtain the following.

Lemma 6.7 The system described by {6.42} to {6.52} is asymptotically

stable if there exist scalars ak, k = 0, 1, symmetric matrix P and symmetric
matrix function R = R(Ao, AI, A oo , A lO ), such that

( Nk
-(PAlAkO)T
-PAlAkO )
-ak P < 0, k = 0,1 (6.63)

for (Ao, A 1) En and (Aoo, A 1o) En, where

No = 1 [P(Ao + A 1) + (Ao
-2 + Ad T P] + aoP + R,
rmax
1 T
N1 -2-[P(Ao + A 1) + (Ao + Ad P] + alP - R.
rmax

Proof. Apply Proposition 6.6 to the transformed system (6.61) and un-
certainty set (6.62) with the selection
0.0 -r(t) < B ~ 0,
a(t, B) ={ 0.1 -2r(t) < B ~ -r(t),
and for -r < 0 ~ 0
R(B) R(Ao, A 1, Aoo, A 10 ),
R(O - r) -R(Ao, A 1, Aoo, A 10 ),
where Ako = Ak(t + 0), k = 0,1. •

Proposition 6.8 The system described by {6.42} to {6.52} is asymptoti-

cally stable if there exist symmetric matrix P, and scalars 0.0 and 0.1 such
that
M -PA1Aoo -PAlAlO)
( -A~oAfp -aoP 0 <0 (6.64)
-AfoAf P O - a l P
for all (Ao, Ad E nand (Aoo, Alo) E n, where

M = _1_[P(A o + Ad + (Ao + Alf P] + (0.0 + al)P.

rmax
 6.4 Delay-independent-Lyapunov-Krasovskii functional 213

Proof. Apply Lemma 6.7 and use Proposition B.6 in Appendix B to

eliminate the matrix function R in (6.63) for k = 0 and 1. •
Note that it is essential to allow R to be an arbitrary function of A o,
A}, A oo , and Alo in Lemma 6.7 in order for its elimination to be valid
in the above proof. Of course, we may choose R to be independent of the
uncertain matrices Ao,AI,Aoo , and A lO in Lemma 6.7 and accept it as
the final stability criterion. This will obviously be more conservative than
Proposition 6.8 because of the additional restriction on R.

6.4 Delay-independent stability criteria using the

Lyapunov-Krasovskii functional

6.4.1 Systems with single delay

Consider the system
x(t) = Ao(t)x(t) + Al(t)X(t - r). (6.65)
The system matrices Ao and Al may be dependent on time and not known,
except that they are within a compact set n:
(Ao(t), Al(t)) En for all t ~ o. (6.66)
The time delay r is assumed to be constant.
Similar to the systems without uncertainty, we may choose the Lyapunov-

I:
Krasovskii functional

V(¢) = ¢T(O)pX(O) + ¢T(O)S¢(O)dO. (6.67)

The derivative along the system trajectory is

We can therefore arrive at the following.

Proposition 6.9 The system described by {6.65} and {6.66} is asymptot-

ically stable if there exist symmetric matrices
p>o, (6.69)
and S, such that

II - ( PAo +A5 P
1- Aip
+S P_ASl) <0 (6.70)

is satisfied for all the possible system matrices (Ao, AI) E n.

214 6. Robust Stability Analysis

It is clear that the stability criterion in Proposition 6.9 is independent

of delay.
Since (6.70) is applicable to all the points of the uncertainty set D, one
needs to check an infinite number of LMls, which is unrealistic. The first
simple case where the problem can be reduced to a finite number of LMls
is subpolytopic uncertainty.

Proposition 6.10 If the uncertainty set D is subpolytopic with vertices

W
(i) _
-
(A 0
(i) A(i»
' 1
. - 1, 2 , ... , n v ,
, 2-

then the system described by (6.65) and (6.66) is asymptotically stable if

there exist symmetric matrices P and 8, such that (6.69) is satisfied, and
(6.70) is satisfied for all

(Ao, AI) = (A~i), A~i», i = 1,2, ... , nv.

Proof. Since
(i) ' A(i»
(A 0 1
r. . -
EH, 2 -
1 , 2 , ... ,nv ,
we only need to prove that if (6.70) is satisfied at all the vertices, it is also
satisfied at arbitrary (Ao, AI) E D. To see this, since for any (Ao, AI) E D,
we can write
nv
Ai L/3jA~j), i = 0, 1
j=1
nv
for some /3j ~ 0, L/3j = 1,
j=1

(6.70) can be written as

~ (PA~j) +A~)Tp+8 PA(j)

1
~/3j A(j)Tp ) <0.
j=1 1 -8
Since /3j ~ 0, the above is implied by

( PA(.)~ +A~)Tp+8 PA~j)) < 0, j = 1,2, ... ,nv. (6.71)

A/ P -8

•
Example 6.1 Consider the following uncertain system,

x(t) = ( -2 + p(t) p(t) ) x(t)

p(t) -0.9 + p(t)
-1 + p(t) 0 )
+a ( -1 -1 _ p(t) x(t - r),
 6.4 Delay-independent-Lyapunov-Krasovskii functional 215

where
Ip(t)1 ::; 0.1.
This is a system with subpolytopic uncertainty. The uncertainty set n has
.
two vertIces (A(i)
0 ' A(i»).
1 ,z = 1 , 2 ,were
h

( -2.1
-1.1
-0.1
-0.1 )
-1.0 '
AP) = 0: (
-1 ~0.9 ),
A(2)
o
= (
-1.9 0.1 ) A~2) = 0: ( -0.9
~1.1 ) .
0.1 -0.8 ' -1.0

The calculation using Proposition 6.10 indicates that the system is stable
independent of delays for 0 ::; 0: < O:max = 0.6730. 0

Next, we will discuss LF norm-bounded uncertainty case. In this case,

Ao(t) Aon + ~Ao(t), (6.72)

Al(t) A in + ~Al(t), (6.73)
where

(~Ao(t) ~AI (t) ) = E(1 - F(t)D)-1 F(t) (Go GI ), (6.74)

and
IIFII ::; 1. (6.75)
We have the following.

Proposition 6.11 The system described by (6.65) and (6.72) to (6.75) is

asymptotically stable if there exist symmetric matrices

P>O
and 8 such that
PAon + AInP + 8 + GIGo PAIn + GIG I PE + GID )
( AinP+GiGo -8 + GiG I GiD < O.
ETp+DTG o DTG I -I +DTD
(6.76)

Proof. Use Proposition 6.9; the system is stable if there exist a P' > 0
and a 8' such that

P' Ao + Alp' + 8'

( AiP' P' Al )
-8' < O.
In view of (6.72) to (6.75), the above can be written as
216 6. Robust Stability Analysis

where

p = -
( P' A On
AT
In
P'
+ ATOn P' + 8' P'A In
-8' ), (6.78)

E ( ~'E ) , (6.79)

G = ( Go GI ). (6.80)
Use Lemma 6.3, (6.77) is satisfied for allllFll :::; 1 if and only ifthere exists
a >. > 0 such that

which is (6.76) with

P = !p'
>. '
8 !8'
>. .

At this point, it is interesting to revisit the Razumikhin Theorem method

•
to study a closely related system

x(t) = Ao(t)x(t) + AI(t)X(t - r(t)), (6.81)

with r(t) uncertain and possibly time-varying. The uncertain coefficient
matrices are described by (6.72) to (6.75). First notice that with a generic
uncertainty, the Razumikhin result in Proposition 6.5 is very similar to the
result in Proposition 6.9. In fact, the conditions in Proposition 6.5 can be
obtained by restricting 8 = aP for a > 0 (and P > 0 is no longer needed
since it is already implied by the other LMI). With this in mind, we can
immediately obtain the following from Proposition 6.11.

Corollary 6.12 The system described by (6.81) with the coefficient uncer-
tainty described by (6.72) to (6.75) is asymptotically stable if there exists
a symmetric matrix P and scalar a > 0 such that (6.76) is satisfied with
8=aP.

Proof. Since Proposition 6.11 is derived from Proposition 6.9, the result
is immediately clear with the observation of the connection between Propo-
sition 6.5 and Proposition 6.9, and the fact that P and 8 in Proposition
6.11 is proportional to P and 8 in Proposition 6.9 . •
There are other cases in this chapter that the Razumikhin Theorem-
based results can be easily read out from the corresponding Lyapunov-
Krasovskii functional-based results by a simple substitution. The readers
should not have difficulty identifying these cases.
 6.4 Delay-independent-Lyapunov-Krasovskii functional 217

Finally we consider the system with block-diagonal uncertainty described

by (6.65) and (6.72) to (6.75), with F having a diagonal structure

(6.82)

To be specific, let
m m
Fi E jRPiXqi, P = LPi' q = L qi· (6.83)
i=l i=l

With this structure, (6.75) is equivalent to IlFill ~ 1, i = 1,2, ... , m. Then

with the scaling discussed in Subsection 6.2.4, we can arrive at the following
stability criterion.

Corollary 6.13 The uncertain system described by (6.65), (6.72) to (6.75),

and (6.82) to (6.83) is asymptotically stable if there exist real symmetric
matrices P > 0, S and scalars Ai > 0, i = 1,2, ... , m such that

PAIn + Gt; ArG I

-S+GfArG I
DTArG I

where

Ar = diag (AIIql A2Iq2 AmIq", ),

Al diag (AIIPI A2Ip2 AmIp", ),
M PAon+A5nP+S+G5ArGo,

and h is the k x k identity matrix.

Proof. As discussed in Subsection 6.2.4, it is sufficient to consider the

stability of the system by the one discribed by (6.40), (6.41), (6.39), (6.75),
and (6.37) to (6.38), with ki' i = 1,2, ... ,m arbitrary constants. Using
Proposition 6.11 to this system, and left-multiply the third row of (6.76)
by KII and right-multiply the third column by K l- I , define Ai = Ijkt. •
Note that this stability criterion is still in the form of LMI. It is reduced
to Proposition 6.11 if we restrict Ai = 1, i = 1,2, ... , m.

Example 6.2 Consider the system with LF norm-bounded uncertainty

described by (6.65) and (6.72) to (6.75) with

AOn ( -2 0 )
o -0.9 '

( -0.5 0 )
-0.5 -0.5 '
218 6. Robust Stability Analysis

E = /31, D = 0.51, Go = GI = 1.
Using Proposition 6.11, we can conclude that the system is asymptotically
stable independent of delays for /3 < /3 max = 0.3378.
If D = 0 and all the other parameters remain the same, then the uncer-
tainty is norm-bounded. We can calculate /3 max = 0.6073. 0

6.4.2 Systems with distributed delays

Consider the system with distributed delays

x(t) = Ao(t)x(t) + lOr A(t, O)x(t + O)dO, (6.84)

with
(Ao(t),A(t,.)) En for all t 2: 0, (6.85)
where n is compact, and

We will supress the explicit notation for the dependence of A o and Al on

time t where no confusion may arise. The Lyapunov-Krasovskii functional
for the system can be chosen as

The derivative of V can be calculated as

V(¢) = ¢T(O)[PAo+A~P+lorS(O)dO]¢(O)
+2¢T(0) l : PA(O)¢(O)dO

- lOr ¢T(O)S(O)¢(O)dO,

or for arbitrary R( 0, Ao, A(0)) satisfying

(6.86)

we have
 6.5 Delay-dependent-simple Lyapunov-Krasovskii functional 219

where
M(O) = !(PAo + A5' P) + S(O) + R(O, Ao, A(·». (6.88)
r
We can therefore conclude the following.

Proposition 6.14 The distributed delay system described by (6.84) and

(6.85) is asymptotically stable if there exists a symmetric matrix
P>O,
symmetric matrix function R( 0, A o, A(·» satisfying (6.86), and symmetric
matrix function S(O) such that

M(O)
( AT(O)p PA(O) )
-S(O) < 0, - r ~ 0 ~ 0
for all
(Ao,A(·» En,
where M(O) is defined in (6. 88}.

Notice in the above that M(O) also depends on the uncertain system
matrices Ao and A( 0) due to R. In practice, we may choose R to be inde-
pendent of the uncertain matrices and accept the additional conservatism.
As in the case of the Razumikhin Theorem-based formulation, an impor-
tant application of this delay-independent stability criterion is to derive a
delay-dependent stability criteria for systems with discrete delays. In such
case, it is essential to allow R to be dependent on the uncertainty such that
it can be analytically eliminated to result in a simple formulation.

6.5 Delay-dependent stability criteria using simple

Lyapunov-Krasovskii functional
Consider again the uncertain system with single delay
x(t) = Ao(t)x(t) + Al (t)x(t - r), (6.89)
and uncertainty expression

(Ao(t), Al(t» En for all t ~ O. (6.90)

Here, we assume the delay r to be a known constant. Similar to the case
of systems without uncertainty, a model transformation can be carried out
to transform the system to one with distributed delays

x(t) = Ao(t)x(t) + i:r A(t, O)x(t + O)dO, (6.91)

220 6. Robust Stability Analysis

where

Ao(t) = Ao(t) + A1 (t),

A(t,o) -A1(t)Ao(t + 0),
A(-r + 0) -A1(t)A1(t + 0),
-r < 0 ~ O.

Therefore, the uncertainty set

(Ao(t), A(t,.)) EO, t 2:: 0

can be described as

k=~+~ }
A(O) = -A1AolJ,A(-r + 0) = -A1A11J, -r ~ 0 < 0 .
for (Ao,A1) E 0 and (AolJ,A11J) EO, 0 E [-r,OJ
(6.92)
As discussed earlier, the stability of the transformed system implies the
stability of the original system. However, the reverse is not necessarily true
due to the presence of additional dynamics.
Applying Proposition 6.14 to the transformed system, we can conclude
the following.

Lemma 6.15 The system described by (6.89) and (6.90) is asymptotically

stable if there exist symmetric matrices P, So, Sl, and matrix function
R(Ao,Al,AolJ,AllJ) such that

( Nk
-(PA1AklJ)T
-PAlAklJ )
-Sk < 0, k = 0,1 (6.93)

for all (Ao, A 1) E 0 and (AolJ, AllJ) E 0, where

1 T
No 2r (P(Ao + AI) + (Ao + A 1) P) + So + R,
1 T
N1 = 2r (P(Ao + A 1) + (Ao + A 1) P) + Sl - R.

Proof. Apply Proposition 6.14 to the transformed system described by

(6.91) to (6.92), and choose

-r < 0 ~ 0,
-2r < 0 ~ -r,
and for -r < 0 ~ 0
 6.5 Delay-dependent-simple Lyapunov-Krasovskii functional 221

We can eliminate the arbitrary matrix function R in (6.93) for k = 0 and

•
1 using Proposition B.6 of Appendix B to obtain the following proposition.

Proposition 6.16 The system described by (6.89) and (6.90) is asymp-

totically stable if there exist symmetric matrices

P>O,

8 0 and 8 1 such that

M -PAl A08 -PAIAIO)

( -A6oAi P -80 0 <0 (6.94)
-AIOAlP 0 -81
for all (Ao,AI) En and (A0 9,A I 9) En, where
M = ![P(A
r
o + Ad + (Ao + Al)T PJ + 8 0 + 8 1 .

Similar to the delay-independent stability case, it is interesting to com-

pare Proposition 6.16 with Proposition 6.8. The result in Proposition 6.8
can be obtained from Proposition 6.16 by the following substitutions

8k {=: ok P ,ok>Ofork=O,l,
r {=: Tmax.

Recall that Proposition 6.8 is applicable to time-varying delay and is de-

rived using the Razumikhin Theorem.
As compared to delay-independent stability discussed earlier and the
discretized Lyapunov functional method to be discussed later, a significant
complicating factor in using this formulation for robust stability analysis
is that some entries in (6.94) involve products of system matrices. This is
the result of model transformation. For subpolytopic uncertainty, let

W
(i)
= (A(i) A(i». 1, 2, ... , nv
0' 1 , 1 =

be the vertices of the uncertainty set n. Since (Ao,AI) and (A o8 ,A I 9)

appear bilinearly in (6.94), an argument similar to Proposition 6.1 allows
us to conclude that the set of the matrices on the left hand side of (6.94)
for (Ao, AI) En and (Aoo, Alo) En is a subpolytopic set. We can therefore
conclude that it is only necessary to check (6.94) at n~ points:

(Ao, At> = (A~/), A~i»,

(Aoo,A]o) = (~),A~j»,
= 1,2, ... ,nv;j = 1,2, ... ,nv o
222 6. Robust Stability Analysis

We will next discuss the norm-bounded uncertainty described by

Ao(t) = AOn + ~Ao(t), (6.95)

Al (t) Aln + ~AI (t), (6.96)

where
( ~Ao(t) ~AI(t) ) = EF(t) ( Go G1 ), (6.97)
and
IIF(t)11 ~ l. (6.98)
We can state the following.

Proposition 6.17 The system described by (6.89) with uncertainty de-

scribed by (6.95) to (6.98) is asymptotically stable if there exist symmetric
matrices
p>o,
So and Sl and scalar Il such that

Mn -PA1nAon -PA1nA ln
-A6'n A fn P -So + IlG6'Go IlG6' Gl
-AlnAtn P IlGiGo -St + IlGiGt
ETp 0 0
~(Go + G 1) -G1AOn -GlAln
-(PAInE)T 0 0
PE ~(Go + GI)T -PAInE
0 -(GtAOn)T 0
0 -(GtAIn)T 0
-I 0 0 < 0, (6.99)
0 -I -GtE
0 -(GIE)T -Ill

where
Mn = ![P(Aon
r
+ Aln ) + (Aon + Alnf P] + So + Sl'

Proof. According to Proposition 6.16, the system is asymptotically sta-

ble if (6.94) is satisfied for all (Ao, AI) satisfying (6.95) to (6.97), and all
(A08, A 18 ) satisfying

A08(t) = Aon + ~A08(t), (6.100)

A18(t) = Aln + ~A18(t), (6.101)

where

(~A08(t) ~A18(t)) = EFo(t) (Go G1 ), (6.102)

 6.5 Delay-dependent·-sirnple Lyapunov-Krasovskii functional 223

and
IlFoll ~ 1. (6.103)
Using (6.95) to (6.97) in (6.94), we obtain

P + EF(t)G < 0, (6.104)

where

P -
( M.
-A~oATnP
-AlOAln P
-PAlnAoo
-So
0
-PAlnAlo
0
-Sl
),
E (fE),
G = ( ~(Go + Gd -GlAOO -GlAlO ).

According to Lemma 6.4, a sufficient condition for (6.104) is

for some A > O. Dividing A and using the Schur complement, the above is
equivalent to
_!P !E GT )
( t~T ~J 0 < O.
G 0 -J
Redefine
1
~P as P,
1
~So as So,
1
~Sl as Sl'

Then

Mn -PAlnAoo -PAlnA lO PE ~(Go + Gl)T

-A~oATnP -So 0 0 -(GlAOO)T
-AlOAln P 0 -Sl 0 -(GlAlO)T < O.
ETp 0 0 -J 0
~(Go + G 1) -GlAOO -GlAlO 0 -J
Using (6.100) to (6.103) in the above, we obtain

-P + EFo(t)G < 0,
224 6. Robust Stability Analysis

where
-PA1nA1n
0
PE
0
~(Go +
-(G1AOn)T
Gt)']
-81 0 -(G1 Alnf ,
0 -f 0
-G 1Aln 0 -f

Using Lemma 6.4, the above is equivalent to the existence of a real scalar
> 0 such that
f.1.

ET ( i!.
+ f.1.GT G E ) 0
-f.1. f <,
which is (6.99). Notice, f.1. > 0 is implied by (6.99) . •
Example 6.3 Consider the system (6.89) with norm-bounded uncertainty
described by (6.95) to (6.98) with

Aon = (~2 -~.9)' A 1n = (=~ ~1)'

and
E = 0.2f, Go = G1 = f.
Using Proposition 6.17, we can obtain that the system is asymptotically
stable for delays r < rmax = 0.5935. We will see later that this result is still
substantially more conservative than the one obtained by the discretized
Lyapunov functional method. 0

6.6 Complete quadratic Lyapunov-Krasovskii

functional approach
Recall from Chapter 5 that for a system without uncertainty
x(t) = Aonx(t) + AlnX(t - r), (6.105)
we may prescribe the derivative of a quadratic Lyapunov-Krasovskii func-
tional as
.
V(¢)1(6.105)
T
= -¢ (O)Wl¢(O) - 1 0
-r
T T
¢ (T)W2¢(T)dT - ¢ (-r)W3¢( -r).
(6.106)
 6.6 Complete quadratic Lyapunov-Krasovskii functional approach 225

The corresponding Lyapunov-Krasovskii functional is

V(¢» = ¢>T(O)U(O)¢>(O)

+2¢>T(0) lOr U(-r - O)AIn¢>(O)dO

+ lOr lOr ¢>T (OI)AinU(OI - (2)AIn¢>(02)dOId02

+ lOr ¢>T(O)W(O)¢>(O)dO, (6.107)

where
W(O) = W3 + (r + 0)W2'
U(O) = UWl (0) + lOr UW2(0)dr + UW3(0)

= 1 00
<pT(t)[WI + rW2 + W3]<P(t + O)dt,
and <p(t) is the fundamental solution of the system (6.105). When WI > 0,
W2 ~ 0, W3 ~ 0, we may easily conclude that V(¢» ~ -c:l1¢>(0)11 2 and
V(¢» ~ O. A slight modification of V(¢» allows us to satisfy the condition
V(¢» ~ £11¢>(0)11 2 .
In this section we will explore applying the Lyapunov-Krasovskii func-
tional (6.107) to the uncertain system
i:(t) = [Aon + ~Ao(t)]x(t) + [Aln + ~AI(t)]X(t - r). (6.108)
We will show that although we may set W2 = W3 = 0 in systems without
uncertainty, these two matrices play important roles in robust analysis. The
purpose is to illustrate the idea rather than obtain the least conservative
condition. We will also assume the uncertainty satisfies
(6.109)
where Po > 0 and PI > 0 are real scalars, and R.n, R I , and S are positive
definite matrices. As indicated earlier, this can also be reformulated in the
standard norm-bound uncertainty. The uncertainties ~Ao and ~AI are
assumed to be independent.
We still use <p(t) to represent the fundamental solution of the nominal
system without uncertainty (6.105). Taking the derivative of the Lyapunov-
Krasovskii functional along the trajectory of the uncertain system (6.108)
yields

l:
. T
V(¢»I(6.108) = V(¢>)I(6.105) + 2[~Ao¢>(0) + ~A1¢>( -r)] .

. [U(O)¢>(O) + U(-r - O)AIn¢>(O)dO] . (6.110)

226 6. Robust Stability Analysis

Using Lemma 6.4, we can obtain

2[~Ao¢(0)]TU(0)¢(0)

< ¢T(O)~A~~ ~Ao¢(O) + ¢T(O)UT(O) ( ~ ) -1 U(O)¢(O)

2[~Al ¢( -r) ]T[U(O)¢(O)]

< ¢T (-r )~Ai :1 ~Al ¢( -r) + ¢T (O)U T (0) ( :1) -1 U(O)¢(O)

2[~Ao¢(0)]T lOr U( -r - O)Aln¢(O)dO

< r¢T (O)~A~ RoJL ~Ao¢(O)

+ 10
-r [U( -r - O)Aln¢(O)]T --;;
(Ro)-1 [U(-r - O)Aln¢(O)]dO
2[~Al¢(-r)]T lOr U(-r - O)Aln¢(O)dO

< r¢T (-r)~AlTRI

-~Al¢( -r)
JL
o
+ l)U( -r - O)A1n¢(O)r -;
(R )-1 [U( -r - O)Aln¢(O)]dO.

Using the above and (6.106) in (6.110), considering (6.109), we obtain

V( ¢) 1(6108)
< _¢T(O) [WI - (1 + r)~ S - JLU T (O)(RC;1 + R 11)U(0)] ¢(O)

- lOr ¢T(r)[W2 - JLAinUT(-r - O)(RC;1 + Rl1)U(-r - O)AlnJ

¢(r)dr

_¢T(_r) [W3 - (1 + r)~ s] ¢(-r).

From which we can conclude the following.

Theorem 6.18 If the nominal system (6.105) is stable, then the uncertain
system (6.108) to (6.109) is asymptotically stable if we can find a scalar
JL > 0, matrices WI > 0, W 2 2': 0, and W3 2': 0 such that
WI > (1 + r) Po S + JLU T (O)(RQl + Rll )U(O),
JL
W2 > JLAinUT(-r-O)(RQl +Rl1)U(-r-O)Aln,
W3 > (1 + r) PI S.
JL
 6.7 Discretized Lyapunov functional 227

Proof. From the above discussion, the Lyapunov-Krasovskii functional

V(¢» ~ 0, and V(¢» ~ -c:11¢>(0)11 2 for some sufficiently small E > O. A slight
modification similar to the case without uncertainty allows us to conclude
V(¢» ~ c:II¢>(O)II2. Therefore, the stability of the system is established . •

6.7 Discretized Lyapunov functional method for

systems with single delay
6. 7.1 General case
Consider again the uncertain system with single delay,

x(t) = Ao(t)x(t) + A}(t)x(t - r), (6.111)

where the system matrices are bounded by a known compact set n in the
following manner,

(Ao(t),A}(t)) E n for all t ~ O. (6.112)

Use the process almost identical to the case with systems without uncer-
tainty described in Section 5.7 of Chapter 5: using Lyapunov functional

V(¢» = q.7(O)P¢>(O) + 2¢>T(0) 1: Q(O¢>(O~

+ lOr [lOr ¢>T(OR(~, 1J)¢>(1J)d1J] d~
+ 1: ¢>T(OS(~)¢>(~)d~, (6.113)

with a discretization process for the Lyapunov functional and derivative,

the following can be concluded.

Proposition 6.19 The system described by (6.111) and (6.112) is asymp-

totically stable if there exist n x n matrices P = p T ; Qp, Sp = S;;,
p = 0, 1, ... , N; Rpq = R~p, p = 0,1, ... , N, q = 0,1, ... , N such that

( ~T
Q
q -) > 0
R+S
(6.114)

and

(6.115)
228 6. Robust Stability Analysis

are satisfied for all (Ao, AI) En, where

Q ( Qo Q2 ... QN ), (6.116)

~o ~N
~l
( Rll
R ~10 ~IN
)
, (6.117)

RNO RNl RNN

S diag( kSo kSI kSN ), (6.118)

~ = ( ~oo ~Ol ) (6.119)

~51 ~11 '

~oo - PAo-A5P -Qo-Q5 -So, (6.120)

~Ol QN -PAl, (6.121)
~11 = SN, (6.122)

Sd diag ( Sdl Sd2 ... SdN ), (6.123)

Sdp = Sp-l - Sp, (6.124)

~1I
Rdl2

Rd
( Rd21 Rd22 ~>N
Rd2N
.
)
, (6.125)

RdNI RdN2 RdNN

Rdpq = h(Rp-l,q-1 - Rpq), (6.126)

(6.127)

(6.128)

(6.129)

(6.130)

(6.131)

(6.132)

(6.133)

(6.134)
 6.7 Discretized Lyapunov functional 229

Proof. The process is almost identical to the case without uncertainty

discussed in Section 5.7 of Chapter 5. The only subtlety is that the U ma-
trix in deriving the Lyapunov-Krasovskii derivative condition needs to be
replaced by an arbitrary matrix function of the uncertain system matrices
Ao and Al in order for the elimination of U to be valid. _
For subpolytopic uncertainty, let the vertices of the uncertainty set n be

(i) A(i)) .
(A 0 ' I ,Z = 1, 2 , ... ,nv ·

Then the inequality (6.115) only needs to be checked for

(Ao, Ad = (Ao(i) ,AI(i) ),

reducing the problem into a finite set of LMIs.

6.7.2 Norm-bounded uncertainty

We will consider again the norm-bounded uncertainty

Aon + ~Ao(t), (6.135)

Aln + ~AI(t), (6.136)

where

(~Ao(t) ~AI(t)) = EF(t) (Go G I ), (6.137)

and

IIFII ~ 1. (6.138)

In spite of the apparent complexity, it can be seen that (6.114) does not
involve system parameters, and (6.115) has a structure very similar to
the delay-independent stability case, and much simpler than the delay-
dependent case using model transformation. Taking advantage of this struc-
ture, we can write (6.115) as

P + EF(t)G + (EF(t)Gf < 0, (6.139)

230 6. Robust Stability Analysis

where

(6.140)

(6.141)

(6.142)

where the subscript n indicates that in the corresponding expressions, the

matrices Ao and At should be replaced by the corresponding nominal value
Aon and Atn, respectively. Using Lemma 6.4, we can conclude that (6.139)
is satisfied for all IIFII ::; 1 if there exists a .x > 0 such that
P+IGTGE)
( ET A _II <0. (6.143)
A

To summarize, we have the following.

Proposition 6.20 The system described by (6.111) and (6.135) to (6.138)

is asymptotically stable if there exist n x n matrices P = pT; Qp, Sp = S-:,
p = 0, 1, ... ,N; Rpq = R~p, p = 0, 1, ... ,N, q = 0, 1, ... ,N such that

I!.T
(Q q -) > 0
R+S
(6.144)

and
Lln - LlG -D~
( _D~T Rd+Sd
_DaT 0 (6.145)
n
EJ, E;
are satisfied. In the above,

A (G'{;Go G'{;Gt)
~G = GiGo erG t '
 6.8 Notes 231

and other notation are defined in (6.116) to (6.134) and (6.140) to (6.142),
with the subscript "n" indicating that Ao and Al should be replaced by the
nominal values Aon and A In , respectively, in the corresponding expressions.

Proof. It has already been shown above that (6.143) and (6.144) are
sufficient for stability. Divide both by .x, and redefine ±P, ±Qp, ±Sp, and
± Rpq as P, Qp, Sp, and Rpq , respectively. •

Example 6.4 Consider the system (6.111) with norm-bounded uncertainty

described by (6.135) to (6.138) with

-1
Aon = ( -2
0
0 )
-0.9 ' A ln = ( -1 o)
-1

and
E = 0.21, Go = GI = I.
Using Proposition 6.20, we estimate the maximum delay r max such that
the system remains asymptotically stable for r < rmax. The computational
results using different discretization N are listed in the following table.
Similar to the case without uncertainty, even for the coarsest discretization
of N = 1, the result is much less conservative than the corresponding
result obtained using the simple Lyapunov-Krasovskii functional method
in Example 6.3 obtained using Proposition 6.17. 0

6.8 Notes
6.8.1 Uncertainty characterization
The uncertainty characterization has been extensively studied in the liter-
ature on robust stability and control. Earlier literature include the study
by Yakubovich [287)-[289], which is equivalent to the norm-bounded un-
certainty discussed here. Indeed, an equivalent of Lemma 6.3 was obtained
in these papers. The special case Lemma 6.4 was obtained by Petersen and
Hollot [224). The generality of block-diagonal uncertainty was shown in the
"pulling out the uncertainty" process by Doyle, Wall, and Stein [58). The
poly topic uncertainty and subpolytopic uncertainty were discussed in Ho-
risberger and Belanger [115]' Boyd and Yang [22), and Gu [82). Most of the
results on manipulating these uncertainties in the context of LMI can be
found in the book by Boyd et al. [21 J.
232 6. Robust Stability Analysis

6.8.2 Stability results based on the Razumikhin Theorem and

Lyapunov-K rasovskii functional
Because effective computational methods were unavailable at the time, the
earlier works are limited to estimations of bounds of uncertainty and delay-
coefficient using simple matrix norm or matrix measures. For examples
using the Razumikhin Theorem, see Wu and Mizukami [285] for a delay-
independent result and Su and Huang [254] for a delay-dependent result.
For examples using the Lyapunov-Krasovskii functional method, see Kim
[147] and Sun, Hsieh, and Yang [255] for delay-independent results.
Later on, due to the advancement of numerical techniques for solving
Riccati equations, most attempts are made to formulate such problems in
the form of Riccati equations or Riccati inequality; see Kwon and Pearson
[163], Lee, Kim, and Kwon [168], and Shen, Chen, and Kung [244] for delay-
independent results. In fact, some of the results are indeed equivalent to
the LMI formulations in view of the Schur complement.
The particular derivation used here is based on Gu and Han [94]. The
variable elimination (or introduction) makes it easier to compare differ-
ent formulations. It is also possible to pursue results using implicit model
transformation, see, for example, Lee, Moon, and Kwon [165], which can
be improved based on the same principle as in Chapter 5.
With the advancement of computational methods to solve LMI, interests
are shifting to much more flexible formulations of LM!. See, for example,
Kokame, Kobayashi, and Mori [151] for an example of delay-independent
stability using the Lyapunov-Krasovskii functional method. For the case
of delay-dependent robust stability with norm-bounded uncertainty, see Li
and de Souza [174] for Lyapunov-Krasovskii functional approach, and Li
and de Souza [173] for Razumikhin Theorem approach. As mentioned in
Chapter 5, scaling factors are usually not included in the LMI formulation,
which represents a major shortcoming in applications.
The additional dynamics of systems with time-varying delay and coeffi-
cients are treated in Kharitonov and Melchor-Aguilar [144]. The treatment
of norm-bounded uncertainty for discretized Lyapunov functional is based
on the idea of Han and Gu [107]' but using the more recent formulation in
discretized Lyapunov functional method proposed in Gu [89]. The impor-
tance of using W2 and W3 in the complete quadratic Lyapunov-Krasovskii
functional for robust stability was suggested by Kharitonov and Zhabko
[145].
For time-varying delays with bounded derivative, it is also possible to
use a Lyapunov-Krasovskii functional approach; see Chapter 8. In many
situations, it is also possible to directly estimate a bound of exponential
decay rate without using the Razumikhin Theorem, see Wu and Mizukami
[286], Hou and Qian [116], and Lehman and Shujaee [169].
7

Systems with Multiple and

Distributed Delays
7.1 Introduction
In this chapter we will discuss the stability and robust stability of systems
with multiple delays and distributed delays using time domain methods.
We will not treat specific uncertainty type and only give results in the form
of "matrix inequalities need to be satisfied for every point in the uncertainty
set." Of course, such a form cannot be directly implemented numerically.
However, readers should not have difficulty applying the same principle
discussed in Chapter 6 to reduce the problem to a finite number of matrix
inequalities for some special type of uncertainties, such as subpolytopic
uncertainty and norm-bounded uncertainty. Also, in view of the similarities
between Razumikhin Theorem-based results and the ones based on simple
Lyapunov-Krasovskii functional results, we will not discuss Razumikhin
Theorem-based methods in this chapter.
Section 7.2 discusses the delay-independent stability criteria for systems
with multiple delays. Section 7.3 discusses delay-dependent stability us-
ing a simple Lyapunov-Krasovskii functional. Similar to the single delay
case, these results are very conservative. Section 7.4 discusses the complete
quadratic Lyapunov-Krasovskii functional for the general case including
systems with multiple delays and distributed delays. Similar to the sin-
gle delay case, the existence of a complete quadratic Lyapunov-Krasovskii
functional is necessary and sufficient for asymptotic stability. Section 7.5
discusses the discretized Lyapunov functional method for systems with
multiple delays. Section 7.6 discusses the discretized Lyapunov functional
method for systems with distributed delays. Based on the principles dis-
cussed, the readers should not have difficulty obtaining stability results for
systems with both discrete and distributed delays.
The delay-independent stability problem for systems with distributed
delays has already been discussed in Chapter 5. Delay-dependent stability
criteria using a simple Lyapunov-Krasovskii functional can be derived using
model transformation with a similar procedure as discrete delays and will
not be pursued here.
234 7. Systems with Multiple and Distributed Delays

7.2 Delay-independent stability criteria of systems

with multiple delays
Consider the uncertain system with multiple delays

L Ai(t)X(t - ri),
K
x(t) = (7.1)
i=O

where
o= ro < r1 < ... < r K = r, (7.2)
and
w(t) En, for all t 2: 0, (7.3)
and w represents all the coefficient matrices:
(7.4)
and n is a compact set. Using the Lyapunov-Krasovskii functional

V(¢) = ¢T(O)p¢(O) + ~
K
iTi
0
¢T(O)Si¢(O)dO,

where
pT >0, (7.5)
sf> 0, i = 1,2, ... ,K, (7.6)
we can easily calculate

V(¢) = ¢T(O) [PA o+A5'P+t,Si]¢(O)

K
+2¢T (O)P L Ai¢( -ri)
i=1

- L ¢T (-ri)Si¢( -ri).
K

i=1

Let

and
K
PAo +A5'P+ LSi PAl PA 2 PA K
i=l
Aip -S1 0 0
IlK = AIp 0 -S2 0 (7.7)

AT
KP 0 0 -SK
 7.3 Simple delay-dependent-multiple delays 235

Then . -T-
V(¢» = ¢>KilK¢>K'
We can therefore conclude the following.

Proposition 7.1 The system with multiple delays described by {7.1} to

{7.4} is asymptotically stable if there exist real symmetric matrices P, 8 1 ,
8 2 , ... , 8 K such that
P>O, (7.8)
and
ilK < 0 for all wEn, (7.9)
where ilK is defined in {7.7} and w is defined in (7.4).

Notice that (7.6) is already implied by (7.9). The stability conditions

are clearly independent of delays. Therefore, Proposition 7.1 is a delay-
independent stability criterion.
The conditions in the above proposition are equivalent to the following
statement: there exist symmetric matrices P, 81, 82 , ... , 8 K and matrix
functions Ri(Ao, ... ,AK)' i = 1,2, ... ,K, such that (7.8) and the following
LMls are satisfied for all possible uncertainties wEn:
K

P Ao + A~ P + ~)8i + R i ) < 0, (7.10)

i=1

( -Ri
At PAi)
-8i < 0, i = 1,2, ... ,K. (7.11)

The equivalence can be established by eliminating R i , i = 1,2, ... , K in

(7.10) and (7.11). We can, of course, restrict ~ to constant matrices and
accept the additional conservatism.

7.3 Simple delay-dependent stability criteria of

systems with multiple delays
To derive a simple delay-dependent stability criteria, we apply model trans-
formation. Utilizing

x(t - ri) = x(t) - l:i x(t + O)dO

we can transform (7.1) to

x(t) =
KKK
2:= Ai(t)X(t) - 2:= 2:=
i=O i=1 j=O
10

-ri
A(t)Aj(t + O)x(t + 0 - rj)dB.
236 7. Systems with Multiple and Distributed Delays

Of course, this transformation introduces additional dynamics similar to

the case of systems with single delay. Using the simple Lyapunov-Krasovskii
functional V{xt} of the following form:

where P and Sij, i = 1,2, ... , K; j = 0,1, ... , K are symmetric positive
definite matrices. We can easily obtain

V( ¢) ~ ¢T (0) [p ~ A;(t) + ~ Af(t)P + t. t. 1

r;S;j ¢(O)

-247 (0)PLL
K

i=l j=o
K
1 0

-ri
Ai {t)Aj {t + €)¢{€ - rj)~
K K 0
- LLI ¢T(€-rj)Sij¢{€-rj)~,
i=1 j=O -ri

or for arbitrary matrix functions ~j{w{t),w{t + €)), i = 1,2, ... , K; j =

1,2, ... , K and Ui{w{t),w{t + €)), i = 2,3, ... , K

where

~j, i = 1,2, ... ,K; j = 1,2, ... ,K, (7.12)

K 1
" ~j + -U
= -~ 1"
i , i = 2,3, ... ,K, (7.13)
j=1 t

K 1 K
- L R 1j - ;- LUi + M, (7.14)
j=1 1 i=2

and
l(K K KK)
M = - rl Pt;Ai + t;ATP+ t;{;riSij . (7.15)

From the above, we can obtain the following simple delay-dependent sta-
bility criterion.
 7.3 Simple delay-dependent-multiple delays 237

Proposition 7.2 The system with multiple delays described by (7.1) to

(7.4) is asymptotically stable if there exist symmetric matrices P, Sij, i =
1,2, ... ,K, and symmetric matrix functions R;,j(w,we), i = 1,2, ... ,K; j =
1,2, ... , K; Ui(W,We), i = 2,3, ... , K, such that
P>O, (7.16)
and
( Wij
(7.17)
[PAiAje]T
for all wEn, We E n, where
W = ... AK)' (7.18)
We ... AKe), (7.19)
Wij = Wij(W,We) are defined in (7.12) to (7.15).

Note that Sij > 0 has already been implied by (7.17). We can restrict
Rij and Ui to be constant matrices and accept the additional conservatism.
We can also eliminate these matrix functions to arrive at the following
equivalent stability condition.

Corollary 7.3 The system with multiple delays described by (7.1) and
(7.4) is asymptotically stable if there exist symmetric matrices P, Sij,
i = 1,2, ... , K; j = 0, 1,2, ... , K, such that (7.16) is satisfied, and
M PA1We PA2We PAKWe
(PA1Wef *Sl 0 0
(PA2Wef 0 /2S2 0 >0

o o
is satisfied for all wEn and we E n, where
Si = diag (SiQ Si1 SiK ),
wand We are defined in (7.18) and (7.19), respectively.

Proof. For a given i, eliminate Rij, j = 1, ... ,K in (7.17) to obtain

where
1 .
-Ui, 2> 1,
ri
238 7. Systems with Multiple and Distributed Delays

This is equivalent to

Then eliminate Ui , i = 2,3, ... , K from the above. •

It it important to realize that the conditions in Proposition 7.2 or Corol-
lary 7.3, although structurally appealing because of their simplicity, are
very conservative, not only because of the additional dynamics introduced
in the model transformation and the usage of the particular Lyapunov-
Krasovskii functional, but also because we ignored the overlapping of dis-
tributed delay intervals of the transformed system.

7.4 Complete quadratic functional for general

linear systems
To overcome the conservatism of the stability criteria discussed so far in this
chapter, we will in the following discuss the complete quadratic Lyapunov-
Krasovskii functional. Similar to the case of single delay, the existence of
such a Lyapunov-Krasovskii functional is necessary and sufficient for sta-
bility. Consider a general linear time-invariant time-delay system

x(t) = lOr d[F(B)]x(t + B). (7.20)

This system includes, as special cases, the systems with multiple discrete de-
lays and distributed delays. The complete quadratic Lyapunov-Krasovskii
functional is very similar to the case with single delay. However, the deriva-
tion is far more sophisticated. In this section we will present the main idea.
Interested readers are referred to the literature discussed in the Notes sec-
tion for technical details.
Let <I>(t) denote the fundamental solution of the system (7.20), i.e., it
satisfies
<i>(t) = lOr d[F(B)]<I>(t + B),
with initial condition
<I>(t) = {I,0, t = 0,
t < 0.
For a real positive definite symmetric matrix W, let
 7.4 Complete quadratic functional for general linear systems 239

Using the Parseval Theorem, we can alternatively write

where

Obviously
U&(r) = Uw(-r).
Based on Uw(r), for a asymptotically stable system, we can again prescribe
(7.21)
using the Lyapunov-Krasovskii functional

vw(¢) = 1 00
xT(t, ¢)Wx(t, ¢)dt,

where x(t, ¢) is the solution of (7.20) with initial condition

Xo = ¢.
As shown by Huang [117], vw(¢) can be explicitly expressed in terms of ¢,

vw(¢) = ¢T(O)Uw(O)¢(O) + 2¢T(0) lOr 1°Uw(u - B)du[F(u)J¢(B)dB

+ lOr 1° dv¢T(v)d7J[F(7])f

[[Or 1°Uw(v -7] +u - ~)du[F(U)l¢(~)d~] . (7.22)

Also notice that, for -r ::; T < 0, we can achieve

by the Lyapunov-Krasovskii functional

With the above discussion, we can construct a complete quadratic Lyapunov-

Krasovskii functional. We will, however, state the following proposition for
the following special class of systems

x(t) =
K
~AiX(t - ri) + lr°A(B)x(t + B)dB, (7.23)

which contain only pointwise and distributed delays.

240 7. Systems with Multiple and Distributed Delays

Proposition 7.4 If the system described by (7.23) is asymptotically stable,

then there exists a Lyapunov-Krasovskii functional

V(¢) = ¢T(O)p¢(O) + 2¢T(0) lOr Q(~)¢(~)d~

+ lOr ¢T (~)S(~)¢(~)d~

+ lOr [lOr ¢T(~)R(~'11)¢(11)d11]~' (7.24)

where pT = P E lRnxn , and the matrix functions Q(~) E lRnxn , S(~) =

ST(~) E lRnxn , RT(~,11) = R(11,~) E lRnxn, such that

V(¢) ~ ell¢112, (7.25)

and
(7.26)
are satisfied for some e > O. Furthermore, the matrix functions Q(~), S(~),
and R( ~ , 11) are continuous everywhere except at pointwise delays ~, 11 = r i,
i=I,2, ... ,K-1.

Proof. We will construct a complete quadratic Lyapunov-Krasovskii

functional for the general system described by (7.20), which reduces to the
form (7.24) with appropriate continuity for the special type of the system
(7.23). For arbitrary continuous function {3((}) ~ 0, -r ~ () ~ 0, denote

lr°{3((}) IdF((}) I ~ ~~
K
~{3((}i)IIF((}i) - F((}i-l)II·
-r=8o<82< . ·<8K=0 -
Since F has bounded variation, the above is well defined. Especially for

l:
{3((}) == 1, the above becomes the total variation, which we denote as J.t

p. = IdF((})I·

Consider for some small 8 > 0,

V(¢) = V(l+6I')I(¢) + 8 lOr [1 ¢T((})¢((})d(}] IdF(r)l.

0
(7.27)

Since

lOr [II¢(O)112 -
l:
V(¢) = -(1 + 8p.)II¢(0)W + 8 II¢(r)II 2)1dF(r)1

= -II¢(O)W - 8 II¢(r) 112 IdF(r) I

< -II¢(0)II2,
 7.5 Discretized Lyapunov functional-multiple delays 241

inequality (7.26) is satisfied. To show (7.25), consider, for 0 < c < y'8/j.l,
v(¢) = V(¢) - cll¢(O)112.
Direct calculation using (7.21) yields

v(¢) = -11¢(O)11 2- 8 [Or 11¢(r)11 2IdF(r)l- 2c¢T(O) [Or dF(B)¢(B)

< -11¢(O)11 2 - 8 [Or 11¢(r)11 2 IdF(r)1 + 2cll¢(O)11 [Or 11¢(B)llldF(B)1

_[Or (~II¢(O)II- v8 11 ¢(r)ll) 2 IdF(r)l_ (1- c:) 11¢(O)112

< O.
Since the above is valid for any ¢, it is also valid for Xt; we therefore have
V(Xt) ::::; O.
Integrating the above from t = 0 to 00, using the fact that lim x(t)
t-+oo
= 0,
we have
0- v(xo) ::::; 0
or
V(¢) = V(xo) ~ cll¢(O)11 2 ,
which is (7.25). Finally we notice that when the system is of the special case
of (7.23), F(B) is continuously differentiable everywhere except at B = Ti,
i = 0,1, ... , K, in which case the constructed Lyapunov-Krasovskii func-
tional (7.27) can be written as (7.24) with appropriate continuity. _
It should be noted that the proof above actually constructed a com-
plete quadratic Lyapunov-Krasovskii functional for the general form of the
system (7.20).

7.5 Discretized Lyapunov functional method for

systems with multiple delays
The basic ideas of the discretized Lyapunov functional method for systems
with multiple delays are very similar to the single delay case. The formu-
lations are much more involved. The main new technical issues unique to
the multiple delay case are

1. The discretization mesh needs to be nonuniform in general to accom-

modate the possibility of incommensurate delays or commensurate
delay with small common factor in order to keep computational cost
reasonable.
242 7. Systems with Multiple and Distributed Delays

2. Many quantities involve two indices: one related to the delays, which
we usually use i or j; the other related to the discretization, which
we usually use p or q. We often need to manipulate these two indices
such as changing the order of summation.
3. The appropriate application of the quadratic integral inequality for
the S terms to render as small as possible the conservatism due to
overlapping of integration interval.

7.5.1 Problem setup

Consider the system with multiple delays
K
x(t) = L A(t)x(t - Ti), (7.28)
i=O
where
o = TO < TI < T2 < ... < T K = T
are the constant delays. It is possible that the delays are incommensurate.
The system matrices may be uncertain, a compact bounding set 0 is known:
w(t) = (Ao(t), Al (t), ... , AK(t)) E 0, for all t ~ O. (7.29)
As is discussed in the last section, the existence of a quadratic Lyapunov-
Krasovskii functional

V(¢) = ¢T(O)p¢(O) + 2¢T(O) [Or Q(~)¢(Od~

+ [Or ¢T (~)S(~)¢(~)d~

+ [Or [[Or ¢T(~)R(~'17)¢(17)d17] d~

is a necessary and sufficient condition for the stability of such a system
without uncertainty. Generally, the matrix functions Q, S, and R have
discontinuities for ~ and 17 at -Ti, i = 1,2, ... , K - 1. To avoid dealing
with discontinuous functions, we choose an equivalent form of Lyapunov-
Krasovskii functional involving only continuous functions:

V(¢) = ¢T(O)P¢(O)+2L¢T(O)
K

i=l
1° Qi(~)¢(~)d~
iri°¢T(~)Si(~)¢(~)d~
-Ti

K
+~

+ ~t, L[.t ¢T«)R'j((,~)¢(~)d~l

j
d(, (7.30)
 7.5 Discretized Lyapunov functional-multiple delays 243

where Qi, Si, and Rij are continuous matrix functions, and

P pT E jRnxn, (7.31)
Qi(~) E jRnxn, (7.32)
SiT(~) Si(~) E jRnxn, (7.33)
RijT(~,'fJ) Rji('fJ,~) E jRnxn, (7.34)
fori=1,2, ... ,K; j=1,2, ... ,K.

Since the additional functions Qi(~), Si(~), Rij (~, 'fJ), and RiK (~, 'fJ), i, j =
1,2, ... , K - 1 are introduced only to account for discontinuities, we will
constrain these functions to the following special forms without loss of
generality:

Qi(~) Qi = constant, (7.35)

Si(~) Si = constant, (7.36)
Rij(~, 'fJ) Rij = constant, (7.37)
RiK(~,'fJ) RiK ('fJ) = RK iT ('fJ) independent of ~, (7.38)
for i = 1,2, ... , K - 1; j = 1,2, ... , K -1.

Taking the derivative of V in (7.30) with respect to time along the tra-
jectory of the system (7.28), carrying out integration by parts similar to
the single delay case, we can again write V(cp) as a quadratic expression of
cp:

K K
V(cp) - LLcpT(-ri)~ijcp(-rj)
i=O j=O

K
+2LLcpT(-ri)
i=O j=l
K
1° IIij(~)cp(~)d~
-ri

- [Or cpT (~)SK (Ocp(~)d~

-2 ~1 [:i cpT(~) [I: kK ('fJ) cp('fJ) d'fJ] d~
-fo cpT(O [fO (aRKK(~,'fJ)
-r -r a~

+ aRK;'fJ(~' 7])) CP(7])d7]] d~, (7.39)

244 7. Systems with Multiple and Distributed Delays

where

D.oo -[PAO+A~P+ ~(Qi+QiT +Si)

+QK (0) + QKT (0) + SK (0)] , (7.40)

D.Oi = Qi - P Ai, D.ii = Si, 1:::; i :::; K - 1,

D.KK = SK(-r), D.OK = QK(_r) - PA K ,
0, i =I j, 1 :::; i :::; K, 1:::; j :::; K,
D.J;, 0:::; i :::; K, 0:::; j :::; K, (7.41)

and
K-1
rrOj(~) = A~Qj +L Rij + RKj(O), (7.42)
i=1
K-1
rr OK (~) = A~ QK (~) +LRiK (~) + RKK (O,~) _ QK (~),
i=1
rr ij (~) AT Qj - Rij, 1:::; i :::; K - 1, 1:::; j :::; K - 1,
rr iK (~) = ATQK(~) - RiK(~), 1:::; i:::; K -1,
rrKj(O AIQj - RKj(-r), 1:::; j:::; K -1,
l1KK (~) AkQK(~) - RKK(_r,~). (7.43)

Recall that the system is asymptotically stable if there exist an c: > 0 and a
Lyapunov-Krasovskii functional (7.30) such that the Lyapunov-Krasovskii
functional condition
V(¢» ~ c:11¢>(O)W (7.44)
and the Lyapunov-Krasovskii derivative condition

(7.45)

are satisfied. The conditions are necessary and sufficient if there is no un-
certainty.

7.5.2 Discretization
For systems with multiple delays, it is essential that the division of the delay
interval [-r, 0] is compatible with the delays such that -ri, i = 1,2, ... , K-1
are among the division points. In other words, let ()p, P = 0,1, ... , N be the
dividing points
 7.5 Discretized Lyapunov functional-multiple delays 245

then
-ri = {}N" i = 1,2, ... , K.
Thus each interval [-ri,Oj is divided into Ni intervals. Let the length of
pth segment be hp
hp = {}p-l - (}p.
For the sake of convenience, define

No 0, (7.46)
ho = 0, (7.47)
hN+1 0. (7.48)
We have
°= No < Nl < ... < N K = N,
and
Ni
ri = Lhp, i = 1,2, ... ,K.
p=l

°
The matrix functions QK (~), SK (~), RiK (~), and RKK (~,,,,) are again
chosen to be piecewise linear as follows: for ~ a ~ 1, p = 1,2, ... , N,

QK({}p + ahp) QK(p) (a) = (1- a)Q{; + aQ{;_l, (7.49)

SK({}p + ahp) = SK(p) (a)= (1- a)S: + aSi!-l, (7.50)
RiK ({}p + ahp) RiK(p)(a) = (1 - a)R~K + aR~~l' (7.51)
i = 1,2, ... , K -1,

and for °
~ a ~ 1, ° /3 :::;
~ 1, p = 1,2, ... , N, q = 1,2, ... , N,
RKK ({}p + ahp, (}q + /3hq)

!
= R KK (pq)(a,/3)
(1 - a)R{;/ + /3R{;5.,q-l + (a - /3)R{;_~,q,
a'?:. /3;
(7.52)
(1- /3)R!// + aR!/_~,q_l + (/3 - a)R:':-l'
a < /3.
Thus, the Lyapunov-Krasovskii functional V is completely determined by
P , Q i , QK
P' Si 'P'
SK Rij ' P
RiK' RKK . ,')-- 1
pq , Z "
2.. "K - l',p,q-
- "1 .. " N ' °
As will be seen later, we often need to manipulate terms involving two
types of indices: one related to the delay, which we usually use i or j, the
other related to the discretization, which we usually use p or q. For this
purpose, it is useful to introduce the notation

(7,53)
246 7. Systems with Multiple and Distributed Delays

In other words, Mp may assume K + 1 distinct values 0,1, ... , K according

to the following rule

i if N i - 1 < p :::; N i ,
o if p = O.

It is not difficult to see that

MN; i, (7.54)
N Mp > P > N Mp - 1' (7.55)

Therefore Mp may be regarded as the inverse of Ni in some sense.

For the sake of convenience, we will also adopt the convention that for
any Uk,

whenever i > j. (7.56)

Then, for any indexed expressions U;, i = 1,2, ... , Kj p = 1,2, ... , N, it is
often useful to change the order of summation as follows

K N; N K
LLU;
i=l p=l
LLU;,
p=l i=Mp
(7.57)

K-l N; NK_l K-l N K-l

LLU;
i=l p=l
L L U;=L L U;, p=l i=Mp
(7.58)

K-l Ni-l NK-l-l K-l N K-l

L LU;
i=l p=l
L L U;=L L U;.
p=l i=Mp+l p=l i=Mp+ 1
(7.59)

In the last step of (7.58) and (7.59), we have used the convention (7.56)
and the fact that Mp = K > K -1 for p > N K - 1 .

7.5.3 Lyapunov-Krasovskii functional condition

With the discretization, the integrations over [-r,O] in the expression of
V in (7.30) may be divided into integrations over fOp, Op-l], p = 1,2, ... , N.
With some manipulations, all the terms except those involving Si, i =
1,2, ... , K can be collected to arrive at the following result.

Lemma 7.5 The Lyapunov-Krasovskii functional V{¢) as expressed in

(7.30) to (7.38), with QK, SK, R iK , and RKK piecewise linear as ex-
 7.5 Discretized Lyapunov functional-multiple delays 247

pressed in (7.49) to (7.52), may be written as

V(¢)

11 (¢T(O) <1>T wT(a))

(
~T ~ j;) (:(0) ) da
QKT il KT RKK w(a)

+L
K-1

i=l
1 0

-ri
¢T (~)Si¢(~)d~

+L
N

p=1 0
1 1
¢(p)T(a)SK(p) (a)¢(p) (a)hpda, (7.60)

where

(Q1 Q2 ... QK-1), (7.61)

(Qf! Qf ... Qf$),

e"
R12 R1,K-1

).
R21 R22 R2,K-1
R
~K-1,1 R K - 1,2 RK-I,K-I
RfK N
ilK (m
K
R~K R~K R'K
R2K
.
N ),
~fJ-I'K RK-I,K
I
RZ-I,K

~K
( RKK
10
RKK
01
RKK
11
RKK)
ON
RKK
IN
RKK . , (7.62)
RKK RKK RKK
NO NI NN

and

( J~", ¢(,)d<
f~r2 ¢(~)d~
)
<1>= (7.63)

~~rK-l ¢(~)d~ ,
248 7. Systems with Multiple and Distributed Delays

w(O)(a) ""(I)(a)
W(I)(a) ""(2) (a) + ",,(I)(a)
W(2) (a) ""(3) (a) + ",,(2) (a)
W(a) , (7.64)

W(N-l)(a) ""fN~(a) + ",,(N-l)(a)

W(N)(a)
"" N (a)
",,(p) (a) hp 10: ¢(p) (f3) df3 , p=1,2, ... ,N, (7.65)

""(p) (a) = hp i 1
¢(p) (f3)df3, p= 1,2, ... ,N, (7.66)

¢(p) (a) ¢(Bp + ahp), p=1,2, ... ,N. (7.67)

Proof. Similar to the single delay case, divide the integration interval
[-r,O] to the segments [Bp ,Bp _ 1 ], we have
K-l ° N
V(¢) = ¢T(0)P¢(0)+2L¢T(0)Qij ¢(~)d~+2L¢T(0)VQK(P)
~1 -~ p=1
K-l N
+ L Vs; +L VSK(p)
i=1 p=1

+ };}; lc, [.t, ~T({)Rij~(")d"l d{

K-l N O N N
+2 L L j ¢T(~)VR;K(P)d~
+ LLVRKK(Pq),
i=1 p=1 -ri p=1 q=1
where

11 QK(p) (a)¢(p) (a)hpda,

[Or; ¢T(~)Si¢(~)d~,
11 ¢(p)T(a)SK(P) (a)¢(p) (a)hpda,

11 RiK(p) (a)¢(p) (a)hpda,

11 [11 ¢(p)T(a)RKK(pq) (a, 1])¢(q) (f3)hqd,B] hpda.

Integrating by parts for VQK(P)' VRiK(p), VRKK(pq) similar to the single delay
case, and collecting terms, we may reach (7.60) . •
 7.5 Discretized Lyapunov functional-multiple delays 249

If the last two terms involving Si and SK(p) in (7.30) do not exist, the
positive definiteness of the matrix in the first term of (7.60) is clearly a
necessary and sufficient condition for the Lyapunov-Krasovksii functional
condition (7.44). The extent of conservatism depends on how the terms
involving Si and SK(p) are incorporated to the rest of the quadratic ex-
pression.

Proposition 7.6 For the Lyapunov-Krasovskii functional V (¢) expressed

in (7.30) to (7.38), with QK, SK, R iK , and RKK piecewise linear as ex-
pressed in (7.49) to (7.52), the Lyapunov-Krasovskii functional condition
(7.44) is satisfied for some sufficiently small c > 0 if
> 0, i = 1,2, ... , K - 1,
Si (7.68)
S: > 0, p=O,I, ... ,N, (7.69)
and
Q
R+S (7.70)
CRK -SF)T
are satisfied, where
- = diag (1
S hNl
Sl (7.71)

fJ fl
F = (
f6 R
~:-1 ff-1
!pi = {I, p:S; Ni - 1 (or equivalently i 2 M p +1 ),
0, otherwise,

S' = diag( lo Sb II S~ ... r;Sj.),

hp max{hp,hp+1 }, p= 1,2, ... ,N -1,

K-1
Sf
P S: + LSi, p=O,I, ... ,N. (7.72)
i=Mp +1

Proof. Let
250 7. Systems with Multiple and Distributed Delays

Since the terms in the first summation of (7.73) can be written as

LOri ¢T(~)Si¢(~)d~
= E1
p=1 Jo
[1 ¢(p)T(a) [aS i + (1- a)Si]¢(P) (a)hpda

+ 11 ¢(N;)T(a)[aS i + (1- a)O]¢(N;) (a)hNida

+ 11 ¢(Ni)T(a)[(I_ a)Si]¢(N;) (a)hNida,

where the first two terms are of similar form to the expression of SK(p), we
can combine the first two terms of the above with the last term in (7.73)
to arrive at
Vs = Vs' + V,§, (7.74)

where

L 1¢(P)T(a)[aS~_1 + (1- a)S~]¢(p)(a)hpda,

N 1
Vs'
p=1 0
K-l 1
Vs = ~ 10 ¢(NdT(a)(l - a)Si</>(Ni ) (a)hN;da.

In view of (7.68) and (7.69), we can use the quadratic integral inequality
(Corollary B.9 of Appendix B) to obtain

N 1
Vs' = ~1 [a¢(p+l)T(a)S~</>(P+1)(a)hp+l
+(1- a)¢(p)T(a)S~¢(p)(a)hp]da

~ t(
p=oJo
[a[h p+ 1¢(P+1) (a)]T .!-S~[hp+l¢(P+1)(a)]
hp

+(1- a)[hp¢(p) (a)]T .!-S~[hp¢(p)(a)Jl da

hp

> LN 11 ["p(P+l) (a) + "p(p) (a)JT it1 S~["p(p+1)(a) + "p(p) (a)Jda

p=o 0 p

11 wT(a)8w(a)da. (7.75)
 7.5 Discretized Lyapunov functional-multiple delays 251

Similarly,

L
K-l 1
Vs > f t/J(N;)T(a)_I-Sit/J(N;)da
i=l 10 hN ;

11 E1 E1
T

= ~ [eI>i - wp(a)] h~. Si [eI>i - wp(a)] da

11
t=l 0 p=O ' FO

= [eI> - Fw(a)JT S[eI> - Fw(a)Jda. (7.76)

Using (7.73), (7.74), (7.75), and (7.76) in (7.60), we have

V(¢)

> 11 (¢T(O) eI>T wT(a))

( ~T ~+S
QKT
-SF ~;
(ilK _ SF)T R,KK + 8' + FTSF
) (:(0) )
w(a)
da.

In view of the above and (7.70), the Lyapunov-Krasovskii functional con-

dition (7.44) is satisfied. •
It is easily seen that if Proposition 7.6 is applied to the case of single delay
with uniform mesh (h p = h, p = 1,2, ... , N), it reduces to Proposition 5.20
in Chapter 5. Although the Lyapunov-Krasovskii functional condition for
the nonuniform mesh case is no more complicated than the uniform mesh
case, it is not the case for the Lyapunov-Krasovskii derivative condition as
will be seen in the next subsection.

7.5.4 Lyapunov-Krasovskii derivative condition

After discretization, the derivative V expression is rather complicated.
However, after tedious algebra, it is possible to consolidate to the form
stated in the following lemma.

Lemma 7.7 Let the Lyapunov-K rasovskii functional V be expressed in

(7.30) to (7.38), with QK, SK, RiK , and RKK piecewise linear as ex-
pressed in (7.49) to (7.52). The derivative V along the tmjectory of the
252 7. Systems with Multiple and Distributed Delays

system described by (7.28) to (7.29) satisfies

V(¢) = -Jt ll.¢ + 2¢T 11 [DB + (1- 2o:)Da]¢(o:)do:

-1 1
¢T(o:)S:f¢(o:)do:

- [11¢(O:)do:]T (R:a K +R;lf) [11¢(O:)dO:]

-1 [11 a
( ¢T(o:) ¢T ((3) )

( ~KK R:faK ) (
o
~(o:)
¢((3)
) d{3] do:
,
(7.77)
da
where
ll.O1
ll.= COO ll.1O
.
ll.11 aOK
ll.1K
.
)
, (7.78)

ll.KO ll.K1 ll.KK

[ K-1
ll.oo = - P Ao + A5 P + ~ (Qi + QiT + Si)

+Qf[ + Qf[T + Scf} (7.79)

ll.Oi = Qi _ P Ai, ll.ii = Si, 1 ~i ~ K -1,

ll.KK S~, ll.oK = Q~ - PA K,
ll.ij = 0, i =1= j, 1 ~ i ~ K, 1 ~ j ~ K,
ll.ij = ll.~, o ~ i ~ K, 0 ~ j ~ K, (7.80)

~I D~N
D82
( DB Df2 DfN )
DB = : 11 . ,
Dh Db D1<N

D;' = ,};. h" ( AJ'Qi + ~ R" + ~KT )

+~ [AJ' (Q: + Q::-l) + ~ (l¢K + ~I) + (~K + Rf.:-I)1

-(Q;;-1 - Q;;),
 7.5 Discretized Lyapunov functional-multiple delays 253

K-l
D ip
S
L..J hP (A!Qj
'"
t
_ Rij) + hp t p + QK
2 [A!(QK p + RiK
p-l ) _ (RiK )]
p-l'
j=Mp
K-l
D Kp L hp(AkQj - RWT)

+ ~ [Ak(Q: + Q:-l) - (Ri5: + Ri5!-I)],

for 1 $ i $ K - 1, 1 $ p $ N,

DON)
D'lN
. ,
D'h . 1YkN

~ [ A~(Q: - Q:-1) + L (R;K - R;!:I) + ~/ - Rfi.:-l1,

K-l
k=1

hp [A!(QK _ QK ) _ (RiK _ RiK )]

2 • p p-l p p-l'

h; [Ak(Q: - Q:-1) - (Ri5: - Ri5!-l)] ,

for 1 $ i $ K - 1, 1 $ p $ N,

s!f = diag( S!/t S/h ... S!fN ),

s!£, S:-l - Sf:, 1 $p$N,

RK
ds12 dslN
Rt"
( Rlf21 RK RK )
RK
ds
ds22 ~lf2N ,

RK
dsNl RK
dsN2 RK
dsNN
K-1 K-1
RK
dspq = L hp(R iK q +
q-1 - RiK) L hq(RiKT
p-l _ RiKT)
p ,
i=Mp i=Mq

R:fsIft R:fs~ R:fsfN )

KK
R ds = ( ~:fslft R:fs~ ~:fslfN,
RKK
dsN1 RKK
dsN2 RKK
dsN N
RKK
dspq ~[(hp + hq)(R:_~,q_l - R:/) + (hp - hq)(R::-1 - Rf:_~,q)],
254 7. Systems with Multiple and Distributed Delays

Proof. Dividing all integrations on the delay intervals [-ri,O] into inte-
gration on segments [Op, Op-l], with a change of variable such as

p=1,2, ... ,N,

within each segment, one obtains

where
AT A

V~ ¢ !:l¢, (7.83)

Vn
K
L¢T(-ri)
i=O
K Nj

j=lp=1 0
1
LL
IIij(Op 1 + ahp)¢(p) (a)hpda, (7.84)

VSK L
N

p=1 0
1
1
¢(p)T(a)S~¢(p)(a)da, (7.85)

VnK
K-l N,
L L
i=1 p=1
1 0
1
¢(p)T(a)

[~l (R;~, - R;K)¢(')(/l)d/l] hpda, (7.86)

VRKK LL
N

p=lq=1
N
1 1
¢(p)T(a)

[1 (1 8+ 18(38)
0

1
o hp 8a hq R KK (Op + ahp, Oq + (3h q)
¢(q) ((3)h qd(3] hpda. (7.87)
 705 Discretized Lyapunov functional-multiple delays 255

Exchanging the order of summation between indices j and p, then direct

calculation gives

Yn ~ ~ ~T (-ri) /.' t, [;t, 11,.0'; (6p +ahp) 1~(p)(,,)do

=
K
t;~</>T(-ri)
N
1 1
[Dip + (1- 2a)Dfpl</>(p) (a)dao (7088)

Similarly, exchanging the order of summation between indices i and p, we

obtain

YRK ~ t, i};, /.' ~(P)T [t. /.' (,,) (E,":, - E,K)~(q) (fJ)dfJ 1 hpda

~ t,t. [/.' ~(P)(")dar.};, hp(R;":, - R;K) [/.' ~(q)(")dOl

[1 1
4>(a)da]T Rl [1 1
4>(a)da] ,

where
KO
Rifl
RK
Rd12
RK
RK
dIN
RK
o

(
o o o

d21 d22 d2N

RK o
RK o
RK o

dNl dN2 dNN

R dpq
o K
=

Using R"l; = ~[RdK + RdKT ], we can further write

VRK =
o r </>(a)da
21 [10
1
A ] T
Ri; [ 10r </>(a)da
1
A ]

0
(7.89)

For VRKK, we can break the integration over

{(a,,8) 10 ::; a ::; 1, °: ; ,8 ::; I}

into two triangular regions

{(a,,8)IO < a::; 1,0::;,8::; a},

{(a,,8)IO < a::;l,a::;,8::;I},
256 7. Systems with Multiple and Distributed Delays

to write

VR.K ~ 2~ t, l ~(P)T 1" (,,) [hq(R""-K,,q - R;;,n

+hp(R:_If,q_1 - R:_If,q)]¢(q)({3)d{3da

= 2101Jt (a) [loa (R:fsK + R:aK)1>({3)d{3] da

= VR~K + VR!<"K, (7.90)
where

VR~K = 210 [loa ¢T (a)R:fs K¢({3)d{3] da,

1
(7.91)

210 [loa ¢T(a)R:aK¢({3)d{3] da.

1
VR!<"K

Exchange the order of integration

VR~K = 210 1 [l~fj ¢t (a)R:fs K¢({3)da] d{3.

Since R:fs K is symmetric, take transpose yields

VR~K = 210
1
[l~Jj ¢T({3)R:fsK ¢(a)da] d{3
210
1
[l~a ¢T(a)R:fsK¢({3)d{3] da. (7.92)

Adding (7.91) and (7.92) yields

VR~K = 1 1
10 [10 ¢T(a)R:fsK ¢({3)d{3] da.

For VR!<" K, one can write

(7.93)
 7.5 Discretized Lyapunov functional-multiple delays 257

Use (7.83), (7.88), (7.85), (7.90), and (7.93) in (7.82) . •

Notice that the last term in (7.77) arises due to the nonuniform mesh.
If a uniform mesh is used (which is not possible for the incommensurate
delay case, and is not practical in the case of commensurate delays with
small common factor), it is easily seen that R:a
K = 0, and the last term in
(7.77) vanishes.
With the above lemma, we can obtain the Lyapunov-Krasovskii deriva-
tive condition in a similar manner to the single delay case.

Proposition 7.8 Let the Lyapunov-Krasovskii functional V be expressed

in {7.30} to {7.38}, with QK, SK, RiK , and RKK piecewise linear as ex-
pressed in {7.49} to {7.52}. The derivative V along the trajectory of the sys-
tem described by {7.28} to {7.29} satisfies the Lyapunov-Krasovskii deriv-
ative condition {7.45} if there exists a real matrix W = W T such that

> 0 (7.94)

E 0,

and

(7.95)

Proof. From Lemma 7.7, it is not difficult to verify

V(¢)

-1 1
( ¢T[DS + (1- 2a)Da] ¢t (a) )

( U -~ ) ( lDs + (1 - 2a) D a]T¢ ) da

-/ Sd - W ¢(a)

+¢T (_~ + DSUDsT + ~DaUDaT) ¢

-[1 ¢(a)da]T (R~K +

1
[1 ¢(a)da]
Rlf)
1

-1 [1
1 Q
(¢T(a) ¢T((3)) (irffaK~K) ( ~~~~ )d(3] da.

Given

( U
-/
-/
S{f - W
)
> 0, (7.96)
258 7. Systems with Multiple and Distributed Delays

one may use the Jensen Inequality (Proposition B.8 of Appendix B) to

obtain

V(¢)

< -1 1
( Jt[D + (1 -
S 2a)Da] Jt (a) ) da

( U
-J
-~
Sd - W
) r
10
1
( lDs
¢(a)
+ (1- 2a)Da]T;P ) da

+;pT (-D. + D'UDBT + ~DaUDaT);P

-[1 ;P(a)dar (R:f.K + Rd~) [1 ;P(a)da]
1 1

-11 [1°( ;pT (a) ;pT ~:aK WaK ) ( %~~~ )d13] da

(fi) ) (

= - (;pT J; ;pT(a)da )
(~;;JDaUDaT SF~W+R:f.K +Rd~ ) (~;;P(a)da)
_11 [1° ( ;pT (a) J? (13) ) (~:aK WaK ) ( %~~~ )d13J da.
Therefore, the Lyapunov-Krasovskii derivative condition is satisfied if (7.96),
(7.95), and

D. - IDaUD aT _DB )
( (7.97)
-D.1- SK-W+RKK+R'K
d d. ds
>0

are satisfied. We may eliminate the variable U from (7.95) and (7.97) using
Proposition B.6 of Appendix B to obtain (7.94) . •
For the case of uniform mesh, since R:aK = 0, one can choose W = 0,
and (7.95) can be omitted, in which case (7.94) reduces to (5.167) if applied
to the single delay case.

7.5.5 Stability condition and examples

Combining Propositions 7.6 and 7.8 we can obtain the following stability
condition.

Proposition 7.9 The system described by (7.28) to (7.29) is asymptoti-

cally stable if there exist n by n real matrices P = pT, Qi, Q:, Si = SiT,
S pK -- SKp ' Rij - - RjiT' p RiK' RK pq K -
- RKqp KT., .;, J' -- 1
, 2
, ...K
, - l',
 7.5 Discretized Lyapunov functional-multiple delays 259

p,q = 0,1, ... ,N;

and Nn by Nn real matrix W = W T , such that (7.70)
and (7.95) are satisfied and (7.94) is satisfied/or all (Ao,Al, ... ,AK ) En,
with the notation defined in (7.61) to (7.62), (7.71) to (7.72), and (7.78)
to (7. 81}.

Proof. According to Propositions 7.6 and 7.8, the system satisfies both
the Lyapunov-Krasovskii functional condition and Lyapunov-Krasovskii
derivative condition, and therefore is asymptotically stable, if (7.68), (7.69),
(7.70), (7.94), and (7.95) are satisfied. It remains to be shown that (7.68)
and (7.69) are not needed.
From (7.95), one concludes that W > 0, which together with the (3,3)
entries ofthe matrix in (7.94) implies (7.69). Examining the diagonal entries
of ..6. of the matrix in (7.94), one may conclude that (7.68) is implied by
(7.94) . •
To illustrate the effectiveness of the method, the following examples are
presented. The first example gives a sense of how close to the analytical
limit can be reached with a rather small N.

Example 7.1 Consider the following uncertainty-free time-delay system

. = (-2 0 )
x(t) 0 -0.9 x(t)+ (-1 0 )
-1 -1 [0.05x(t - 0.5r) + 0.95x(t - r)].

The maximum time delay for stability can be analytically calculated as

rmax = 8.5976.
Using Proposition 7.9, the maximum time delay for stability is estimated
using different Ndl and Nd2 as follows.
Nd2
Ndl 1 2
1 8.585 8.586
2 8.594 8.596
It is clear that the results are very close to the analytical limit even for
rather small N. 0

The next example illustrates the case of polytopic uncertainty and in-
commensurate delays.

Example 7.2 This example considers the uncertain system

x(t) = (-2p(t)+ p(t) p(t)
-0.9 + p(t)
) x(t)

+0.15 ( =~ + p(t) ~1- p(t) ) x (t - ~)

-1 - p(t) p(t) )
+0.85 ( -1 _ p(t) -1 _ p(t) x(t - r),
260 7. Systems with Multiple and Distributed Delays

where p is a time-varying parameter satisfying

Ip(t)1 :::; 0.05, for all t.

Using Proposition 7.9 with Ndl = Nd2 = 1, it is concluded that the system
is stable for r :::; rmax = 5.70. For Ndl = Nd2 = 2, the result is improved to
rmax = 5.75.0

The next example considers the case where the system is unstable when
the delay vanishes.

Example 7.3 Consider the system

x(t) - O.l±(t) + x(t) + x(t - r/2) - x(t - r) = O.
This system is unstable for r = O. For small r, the last two terms can
be viewed as a finite difference, approximating dJt) , and therefore, as r
increases, it may improve stability. However, this approximation may not
be valid for large r. Write the system in the state space form

1 )(Xl(t))+(O
( 0-1 0.1 0)(x 1(t-r/2))
X2(t) -1 0 X2(t - r/2)
00)(x1(t-r))
+( 1 0 X2(t - r) .
The stability criterion in Proposition 7.9 is used to find an interval of
r, [rmin' rmax], for which the system is guaranteed to be stable. A simple
search with uniform increment of 0.1 is conducted. Then a bisection near
the upper and lower bound is conducted to find rmin and rmax to within
0.001 for a given Ndl and N d2 . Using Ndl = Nd2 = 1, it was estimated
that [rmin, rmax] = [0.204,1.350]. For Ndl = Nd2 = 2, the interval can
be enlarged to [0.203,1.372]. It can be verified that for r = 0.2025, the
system has an imaginary pole at 1.0077i, and for r = 1.3723, the system
has an imaginary pole at 1. 3786i. Therefore, the exact stability interval is
(0.2025,1.3723) and these estimates are very accurate. 0

7.6 Discretized Lyapunov functional method for

systems with distributed delays
We will only consider the case where the coefficient matrix for the distrib-
uted delay is piecewise constant. The Lyapunov functional used is identical
to the case of single delay. However, similar to the case of multiple delays,
it is important that the discretization mesh is compatible to the delay, i.e.,
the points where the coefficient matrix changes have to be selected as the
dividing points. This again often dictates the usage of nonuniform mesh.
 7.6 Discretized Lyapunov functional-distributed delays 261

7.6.1 Problem statement

Consider the system with distributed delays

x(t) = Ao(t)x(t) + lOr A(t, O)x(t + O)dO, (7.98)

where the coefficient matrix A is piecewise independent of 0

A(t, 0) = Ai(t), - ri S 0 < -ri-J, i = 1,2, ... , K, (7.99)

with ris satisfying

o= ro < rl < r2 < ... < rK = r. (7.100)

In other words,

x(t) = Ao(t)x(t) + t, Ai(t) l~:i-l x(t + O)dO. (7.101)

The system matrices are uncertain and are bounded by a known compact
set
(7.102)
For the sake of simplicity, we will often suppress the time dependence of
the system matrices and write Ao, Ai, and A(O) instead of Ao(t), Ai(t) and
A(t,O).
A general quadratic Lyapunov functional will again be used:

V(¢) = ¢T (O)P¢(O) + 2¢T (0) lOr Q(~)¢(Od~

+ lOr [lOr ¢T(~)R(~,1])¢(1])d1]] d~
+ l: ¢T(~)S(~)¢(~)d~, (7.103)

where
P = pT E jRnxn , (7.104)
and for all -r S ~ S 0 and -r S 1] S 0,

Q(~) E jRnxn, (7.105)

S(~) ST(~) E jRnxn, (7.106)
R(~, 1]) RT(1],~) E jRnxn. (7.107)
262 7. Systems with Multiple and Distributed Delays

Taking the derivative of V with respect to time along the trajectory of the
system (7.98), integration by parts when necessary similar to the single
delay case, one can write Vasa quadratic functional of ¢ as follows:
V(¢)

- (¢T(O) ¢T(_r)) (~~O(_r) ~«~;?) ( :~~r) )

+2 (¢T(O) ¢T(_r)) lOr ( )

~~~g ¢(~)d~
_JO ¢T(~) d8(~) ¢(~)d~
-r d~

-Jo d~Jo ¢T(~)[aR(~,,,,) + aR(~,,,,)

-r -r a~ a",
_AT(~)Q(",) - QT(~)A("')] ¢(",)d"" (7.108)

where

~OO -PAo - A~ P - Q(O) - QT(O) - 8(0),

rO(~) A~ Q(~) + P A(O _ d~~~) + R(O, ~),

rl(~) -R(-r,~).

Recall that the system is asymptotically stable if there exists a Lyapunov-

Krasovskii functional (7.103) which satisfies
(7.109)
and its derivative satisfies
(7.110)

for some E > O. The conditions are necessary and sufficient if there is no
uncertainty in the system.

7.6.2 Discretization
Similar to the case of multiple delays, it is again important to allow nonuni-
form mesh in general, such that A(~) is independent of ~ in each segment.
Specifically, let 0 = 00 > 01 > ... > ON = -r be chosen such that
-ri = ON;, i = 0, 1,2, ... , K,
thus the intervals [-ri' OJ is divided into Ni segments [Op, Op-d of length

hp = Op-l - Op,
 7.6 Discretized Lyapunov functional-distributed delays 263

p = 1,2, ... , N i . Adopt the convention

No=O.
Then
o= No < N1 < N2 < ... < N K = N.
The parameters of the Lyapunov-Krasovskii functional are again chosen to
be piecewise linear, i.e., for 0 ~ a ~ 1 and p = 1,2, ... , N

Q(Bp + ahp) = Q(p)(a) = (1 - a)Qp + aQp-1, (7.111)

S(Bp + ah p) S(p)(a) = (1 - a)Sp + aSp_I, (7.112)

and for 0 ~ a ~ 1, 0 ~ 13 ~ 1, P = 1,2, ... , N, q = 1,2, ... , N,

R(Bp + ahp, Bq + f3hq)

R(pq) (a, 13)
(I - a)Rpq + f3Rp-1,q-1 + (a - f3)Rp- 1,q, a ? 13;
{ (7.113)
(1 - f3)Rpq + aRp-1,q-1 + (13 - a)Rp,q-I, a < 13.

7.6.3 Lyapunov-K rasovskii functional condition

Even though the mesh is not uniform, the Lyapunov-Krasovskii functional
condition has a form similar to the uniform mesh case presented in the
single delay case. Specifically, let

(7.114)
We have the following.

Proposition 7.10 The Lyapunov-Krasovskiifunctional V(¢» as expressed

in (7.103) to (7.107), with Q, S, and R piecewise linear as expressed in
(7.111) to (7.113), satisfies

V(¢»

11 (¢>T(O) wT(a)) ( ~T q
R
) (¢>(O)
w(a)
) da

+ ];
N
1 1
¢>(p)T(a)S(p)(a)¢>(p)(a)hpda (7.115)

> 11 (¢>T(O) wT(a)) (~T ~+s) ( ~\~) ) da, (7.116)

if
Sp > 0, p = 0,1, ... ,N. (7.117)
264 7. Systems with Multiple and Distributed Delays

Therefore the Lyapunov-Krasovskiifunctional condition (7.109) is satisfied

if (7.117) and

( ~T q
Q
-)
R+8
>0 (7.118)

are satisfied, where

Q ( Qo Q1 ... QN ), (7.119)
R01
R (~
~10 Rll ~N
~1N
)
, (7.120)

RNO RNl RNN

diag ( lo 8 II 8
0 1 ... h~ 8 N ), (7.121)

max{hp , hp+d, p = 1,2, ... ,N -1, (7.122)

hI, (7.123)
hN , (7.124)

and

111 (0) (a) 'l/J(1) (a)

1l1(1)(a) 'l/J(2) (a) + 'l/J(1) (a)

111 (2) (a) 'l/J(3) (a) + 'l/J(2) (a)
1l1(a) = (7.125)

1l1(N-1)(a)
1l1(N) (a)

'l/J(p) (a) hp lQ ¢(p) ((3)d(3, (7.126)

'l/J(p) (a) hp i 1
¢(p) ((3)d(3, (7.127)

p=1,2, ... ,N.

Proof. The proof for (7.115) is similar to that of (5.133) in Proposition

5.20 for the single delay case. To prove (7.116), we can bound the last term
 7.6 Discretized Lyapunov functional-distributed delays 265

of (7.115) as follows:
N
~VsP g ~
N
11
¢(p)T(a)S(p)(a)¢(p)(a)hpda

~
N
1 1
¢(p)T(a)[(l- a)Sp + aSp-1]¢(P) (a)hpda

11 ah 1¢(1)T(a) ~1 SOh1¢(1)(a)da
+ 11 (1- a)hN¢(N)T(a) h~ SNhN¢(N) (a)da

+ L r
N-1 1 [
ahp+1¢(P+1)T(a)~Sphp+1¢(P+1)(a)
P=l k ~1
+(1- a)hp¢(p)T(a) ~p Sphp¢(p) (a)] da

> r1 ah1¢(1)T(a)lSoh1¢(1)(a)da
Jo ho

+ r1(1- a)h N ¢(N)T(a)2-SN h N ¢(N>(a)da

Jo hN

+~ l' ["",*, ¢U>+ ')T (a) ;p Sphp+! ¢(PH) (a)

+(1- a)hp¢(P)T(a)lSphp¢(p) (a)] da.

hp

In view of (7.117), we can use the quadratic inequalities (Corollary B.9 in

Appendix B) in each term above to obtain

from which (7.116) follows. (7.118) is clearly implied by (7.116) . •

The above proposition can also be considered as the special case of
Lemma 7.5 and Proposition 7.6.

7.6.4 Lyapunov-Krasovskii derivative condition

Similar to the case of multiple delays, it is useful to define

Mp = min{i I Ni ~ p}. (7.128)

266 7. Systems with Multiple and Distributed Delays

We have

A(~) = AMp independent of ~ within ()p ::; ~ < ()p-1.

We will write
(7.129)

for the sake of convenience. Then, after the discretization, dividing the
integration interval [-r,O] into N segments [()p,()p-1], P = 1,2, ... ,N, the
derivative V of the Lyapunov-Krasovskii functional V along the trajectory
of the system described by (7.98) to (7.102) may be written as

V(¢) = -¢5r tl.¢or + 2¢5r 11 [DS + (1- 2a)Da]¢(a)da

-1 1
¢T(a)Sd¢(a)da

+2 (11 ¢(a)da) T 11 [ES + (1- 2a)Ea]¢(a)da

_(1 ¢(a)da) T Rds (1 ¢(a)da)

1 1

-1 [1" (
1
¢T (a) ¢T ((3) )

(°
R~a °
Rda ) ( ~(a) ) dfJ]]da
¢((3) ,
(7.130)

where

(7.131)

DS ( DOl D02 DON

Dl1 Dh DS ) , (7.132)
IN

Da ( Dih D02 DON

= DI1 DI2 Da ) , (7.133)
IN
 7.6 Discretized Lyapunov functional-distributed delays 267

or more explicitly,

Dop = ~[AnQp + Qp-d + 2PAp + (Ii{)p + RO,p-l)]

-(Qp-l - Qp), (7.134)

D~p - h; [RNp + RN,p-l], (7.135)

Dop = h;[A5(Qp - Qp-l) + (Ii{)p - Ii{),p-t)], (7.136)

hp ]
Dfp = 2[RN,P-l - RNp, (7.137)

Sd = diag( Sdl Sd2 ... SdN ), (7.138)

Sdp = Sp-l - Sp, (7.139)

ES =
CII
E~l
Ef2
E~2 ElN
E~N )
· , (7.140)

ENl EN2 ENN

Ea Cfl
E~l
Ef2
E~2 ~N
E~N )
· , (7.141)

ENl EN2 ENN

hphq T )
E;q = -2- Ap (Qq + Qq-l , (7.142)
hphq T )
(7.143)
E;q -2- Ap (Qq - Qq-l ,

Rdsl2
Rds
C~U Rds2l Rds22 ~"N
R ds2N
·
)
, (7.144)
R dsNl RdsN2 R dsNN
1
Rdspq 2[(hp + hq)(Rp-l,q-l - Rpq),
+(hp - hq)(Rp,q-l - Rp-l,q)], (7.145)

R dal2
(~U
Rda21 Rda22 ~'N
Rda2N )
Rda = . , (7.146)
R daNl R daN2RdaNN
1
Rdapq = 2(hp - hq)(Rp-l,q-l - Rp-l,q - Rp,q-l + Rpq), (7.147)
268 7. Systems with Multiple and Distributed Delays

and

( ¢T (0) ¢T ( -r) ) ,
(¢(l)T(O!) ¢(2)T(0!) ... ¢(N)T(O!)).

From the quadratic expression (7.130), we can obtain the Lyapunov-

Krasovskii derivative condition as follows.

Proposition 7.11 The derivative of the Lyapunov-Krasovskii functional

V(¢) expressed in {7.103} to {7.107}, with Q, S, and R piecewise linear as
in {7.111} to {7.113}, along the trajectory of the system expressed by {7.98}
to {7.102}, satisfies

V(¢)
< - (¢~r Io1 ¢T (O!)dO! )
_~DaUDaT + ~ -Ds - ~DaUEaT ) ( ¢Or )
(
_DsT - ~EaUDaT Sd - Esa + Rds - W I01 ¢(O!)dO!

-1 [l
1 CX
(¢T(O!) 4t(~)) (~ra ~a) ( :~~~ )d~] dO!,
(7.148)

provided

( U
-J Sd
-J) > O. (7.149)

Therefore, the Lyapunov-K rasovskii derivative condition {7.11 O} is satisfied

for some sufficiently small E > 0 if

> 0, (7.150)

> 0 (7.151)

are satisfied for all (Ao, AI, A2, ... , AK) E O.

 7.6 Discretized Lyapunov functional-distributed delays 269

Proof. The idea is very similar to the single delay and multiple delay
case. From (7.130), it can be verified that
V(¢)

= -1 (¢hE1
¢T(a)) (~I S:) ( ~f~ ) da
-¢6r fl.¢or + ¢6r [DSUD ST + ~DaUDaT] ¢Or
+2¢6r [DSUE ST + ~DaUEaT] 11 ¢((3)df3

+ 11 ¢T(f3)df3 [ESUW T + ~EaUEaT] 11 ¢(f3)df3

_(1 ¢(a)da) T Rds (1 ¢(a)da)

1 1

- Jor Jor ( ¢AT (a) ¢AT (13) ) ( Rra

1
[ 0 Rda ) ( ¢(a) ) df3 ] d
0 ¢(f3) a

for arbitrary matrix function U(Ao, AI, A2, ... , AK), where

¢hE = ¢6r[D S + (1 - 2a)Da] + 11 ¢T (f3)df3[ES + (1 - 2a)Ea].

In view of (7.149), we can use the Jensen Inequality (Proposition B.8 in
Appendix B) in the first term above, and collect terms to obtain
V(¢)
< - (¢6r fol ¢T (a)da )

( _~DaUDaT + fl. _D S - ~DaUEaT ) ( ¢Or )

_DsT - ~EaUDaT Sd - Esa + Rds - W f; ¢(a)da

-1 [l1 a
(¢T(a) ¢T(f3)) (irra ~a) ( ~~~~ ) df3 ] da
for arbitrary matrix W, where

E sa = E S + EsT + ~EaUEaT.
Therefore, the Lyapunov-Krasovskii derivative condition is satisfied if there
exist a matrix Wand a matrix function U such that (7.149), (7.151), and

(7.152)

are satisfied. Finally, we can use variable elimination discussed in Propo-

sition B.6 of Appendix B to eliminate the matrix function U from (7.149)
and (7.152) to obtain (7.150) . •
270 7. Systems with Multiple and Distributed Delays

7.6.5 Stability criterion and examples

From the above discussion in this section, we can summarize as follows.

Corollary 7.12 The system described by (7.98) to (7.102) is asymptoti-

cally stable if there exist n by n matrices P, Qp, Sp = S~, Rpq = R~p,
p, q = 0,1, ... , N; and Nn by Nn real matrix W = W T , such that (7.118),
(7. 150}, and (7.151) are satisfied, with notations defined in (7.119) to
(7.124) and (7.131) to (7. 147}.

Proof. This follows from Propositions 7.10 and 7.11. Similar to the sin-
gle delay and multiple delay case, condition (7.117) is already implied by
(7.150) . •
To illustrate the effectiveness of the discretized Lyapunov functional
method for systems with distributed delay, the following example is con-
sidered. The example considers a system resulting from a model transfor-
mation from a system with single pointwise delay.

Example 7.4 Consider the system

x(t) = Ao(t)x(t) + 1 0
-r/2 AIX(t + ~)d~ + l-
-r
r2 2
/ A x(t + ~)d~, (7.153)

where

Ao ( 0.5-1.5 0 ),
-1
(7.154)

Al
( 02 2.5-0.5 ), (7.155)

A2
( 0 -1 ).
-1 0
(7.156)

This is the result of model transformation from the system with single delay

x. ()
t = (-3
1 -2.5)
0.5 x ()
t + (1.5
-0.5 2.5) x (t - r /2) ,
-1.5 (7.157)

which was discussed in Example 5.5 in Chapter 5. We have scaled the

time delay r to conform to the notation in this section. In Example 5.5,
it was found that although the system (7.157) is asymptotically stable for
r < 4.8368, the transformed system (7.153) is asymptotically stable only
for r < 2 due to the presence of additional dynamics. Using Corollary
7.12, the stability limit rmax such that the system (7.153) is asymptotically
stable for r :<:; rmax are estimated. For Ndl = Nd2 = 1 (N = 2). The
calculation results in rmax = 1.9725 with only about 1.4% conservatism.
The estimation is improved to rmax = 1.9999 for Ndl = Nd2 = 2 (N = 4).
o
 7.7 Notes 271

7.7 Notes
Similar to the single delay case, there are numerous results available for
both delay-independent and delay-dependent stability results. See [214] for
an overview. For delay-indepent stability using Lyapunov-Krasovskii func-
tional method, see Ge, Frank, and Lin[74].
For a similar result to Corollary 7.3 using the Razumikhin Theorem, see
Cao, Sun, and Cheng [28] (similar to the single delay case, scaling factors
can be introduced). The derivation used here is original.
For systems with multiple delays, in addition to delay-independent and
delay-dependent stability, it is also possible to consider the concept of
mixed delay-independent and delay-dependent stability; see Kolmanovskii,
Niculescu, and Richard [159].
The extension to the general linear time-delay system of the complete
quadratic Lyapunov-Krasovskii functional (7.22) was reached by Huang
[117]. The modification (7.27) is along the same line as [145].
The discretized Lyapunov functional formulation for systems with mul-
tiple delays was first proposed in Gu [86]. The simplification to the current
form discussed here is based on Gu [90]. The expression of V(¢) in the
proof of Proposition 7.8 can be generalized to

V(¢)

-1 1
( 4/[DS + (1- 2a)Da] ;;/(a))

( U -J ) ([Ds+(1-2a)D a jT¢)d
-J S,f-aZ-(I-a)W ¢(a) a

+¢T ( -D. + DSUD sT + ~DaUDaT) ¢

-[1 1
¢(a)da]T (R:fsK + Rlf) [1 1
¢(a)da]

-1 [11 0
(¢T(a) ¢T({3)) (~:aK ~:aK) ( %~;j )df3] da.
The one used in the proof of Proposition 7.8 is the special case of the above
by setting Z = W. For nonuniform mesh with large difference of mesh size,
a stability criterion based on the above expression can be less conservative
but much more complicated.
The discretized Lyapunov functional method for systems with distributed
delay was proposed in Gu et al. [95]. The formulation presented here is
based on Gu [91].
 Part III

Input-Output Approach
8

Input-Output Stability
8.1 Introduction
In this chapter we will discuss the input-output stability of time-delay
systems. The main purpose of presenting this formulation is to consider
the stability problem of systems under dynamical uncertainty. The input-
output stability framework discussed in this chapter is similar in flavor to
Chapter 3. A system is often written in the form of a nominal system and
an uncertain feedback, and stability is concluded based on a small gain
argument. However, the formulation here is more general. For example,
the nominal system may be a time-delay system, and the feedback may
be a more general uncertainty which is not necessarily due to time-delay
elements.
Based on the input-output framework, it is often possible to write the
stability criteria using the Lyapunov-Krasovskii functional formulation or,
in the case of linear time-invariant nominal systems, frequency sweeping
form. We will emphasize the Lyapunov-Krasovskii formulation in this chap-
ter since the frequency-sweeping form is very similar to Chapter 3.
Section 8.2 will discuss the basic concept of input-output stability and
the small gain condition for the preservation of input-output stability in a
feedback configuration. As a simple illustration, Section 8.3 discusses the
method of comparison systems. In this approach, the time-delay system
is embedded in a family of systems with a forward system without time
delay and a dynamical feedback uncertainty. The same LMI is obtained by
either a frequency domain approach using the Bounded Real Lemma, or
directly by a time domain approach. Then we begin a more comprehensive
discussion of the methodology of the input-output framework, with known
constant time delays considered as part of the nominal system rather than
uncertainty. In Section 8.4 we set up the scaled small gain problem. In
Section 8.5 we explain how the robust stability problem with dynamical
feedback uncertainty can be formulated as a scaled small gain problem and
discuss the uncertainty characterization. It will also be shown that some
robust stability conditions discussed in Chapter 6 are also sufficient for
dynamical uncertainty.
As an application of the above framework, in Section 8.6 we discuss the
approximation of time-delay elements in stability analysis. Specifically, we
discuss the approximation of time-varying delays by time-invariant delays,
276 8. Input-Output Stability

the approximation of arbitrary distributed delay by one with piecewise

constant coefficient, and the approximation of arbitrary linear system by
one with multiple delays. The errors of such approximations are modeled
as a dynamical uncertainty, and therefore, the stability of the system can
be analyzed by the method discussed in Sections 8.4 and 8.5.
In this chapter, for matrices and column vectors, we will use 11·11 instead
of 11·112 to represent the 2-norm. However, we will still keep the subscript
for function norms.

8.2 Input-output stability

Let Dl+ represent the set of all the functions 1 : lR+ - t IRn which are square
Jo
integrable, i.e., oo 111(t)11 2dt is well defined and finite. In this case, we can
define the L 2-norm as

111112 = 100
111(t)1I 2 dt.

For any T > 0 and a function 1 : i+ -t IRn , the truncation operator (or
projection operator) P r is defined as

and the shift operator Sr is defined as

t;::: T,
0::; t < T.
Let L 2e + be defined as

L 2e + = U : i+ -t IRn 1 Prl E L 2+ for all T > O}.

The superscript n in L 2+ or L 2e + may be omitted when the dimension n
is not important or understood from the context.
In the input-output approach, a system is a mapping H : L 2e + - t L 2e +.
For a real nonsingular matrix X and a system H, we define a new system
Hx by
Hx 1 = XH(X- 1 J).
If F and G are two systems, then FG is also a system defined as

(FG)I = F(GJ).
A system H is linear if H(al + f3g) = aHI + f3Hg for all a,f3 E IR and
I, g E L 2e+. It is causal if it satisfies P rH = P rHP r for all T > O. In other
 8.2 Input-output stability 277

words, if H is a causal system, and g = HI, then g(t) is independent of

1(0, ~ > t. We will assume that all the systems discussed from here on are
causal unless specifically mentioned. H is time-invariant if 8 T H = H8 T for
any T > O. Clearly, 8 is both causal and time-invariant, and P is causal
T T

but not time-invariant.

We say that a system H is L 2-stable if HI E L 2+ whenever IE L 2+. It is
L2-stable with finite gain if it is L 2-stable and there exist real scalars 1 > 0
and b such that IIH/I12 S 111/112 + b for all I E L 2+. It is L 2-stable with
finite gain and zero bias if it is L 2-stable and there exists a real scalar 1 > 0
such that IIH/I12 S 111/112 for all I E L 2+. For the sake of convenience,
we will say a system is input-output stable if it is L 2 -stable with finite gain
and zero bias. If H is input-output stable, we can define its gain ,o(H) as

In practice, a system H is often described using some internal states. We

will say that a system H is internally stable if the zero state is globally
uniformly asymptotically stable. For example, the functional differential
equations

x(t) I(t, Xt, Ut),

y(t) h(t, Xt, Ut),

satisfying

I(t, 0, 0) = 0,
h(t,O,O) = 0,
describe a system H. The mapping from u to y for zero initial condition

X(t) = 0, u(t) = 0, t S 0,

can be written as y = Hu. System H is internally stable if the zero state

of the functional differential equation

x(t) = I(t, Xt, 0) (8.1)

is globally uniformly asymptotically stable. Notice that we have allowed a
slight abuse of notation by using H to represent both the functional dif-
ferential equation description (with the possibility of nonzero initial con-
ditions) and the input output mapping corresponding to the zero initial
condition. If a system H is represented by a linear time-invariant func-
tional differential equation, then in view of the solution expression (1.33)
in Chapter 1, internal stability implies input-output stability. On the other
hand, if certain controllability and observability conditions are satisfied,
then input-output stability also implies internal stability. A linear time-
invariant system H can also be represented by its transfer function matrix
278 8. Input-Output Stability

H(s) and can be studied by the frequency domain approach. Furthermore,

its gain is equal to its 11.00 norm

1'0(H) = IiH(s)lloo.

For general nonlinear systems, the relationship between internal stability

and input-output stability is more complicated and will not be pursued
here.
Given systems G and F, connect the two systems in a feedback loop,

y = Gu, (8.2)
u = Fy+v. (8.3)

If the on~to-one relationship between input v and output y is well defined

by equations (8.2) and (8.3), then we can write y = Hv, and H can be
regarded as a system, and we will write H=feedback(G,F). An important
conclusion in input-output stability theory is the small gain theorem, which
states that the combined system feedback(G, F) is well defined and input-
output stable if both G and F are input-output stable and satisfy

(8.4)

Condition (8.4) is known as the small gain condition. If the small gain
condition is satisfied, then we can easily find an upper bound of the gain
of the feedback system 1'0 (feedback(G, F)): from (8.2) and (8.3), we can
write

IIyl12 < 1'0(G)lluI12

1'0(G)IIFy + vl12
< 1'0(G)(IIFYI12 + Iiv1i2)
< 1'0(G)C'Yo(F)IIYI12 + IlvI12).
Therefore,

This indicates

1'0 (G)
1'0 (feedback(G, F)) $; (1 -l'o(Gho(F))"

For a feedback system H = feedback( G, F) and a real nonsingular matrix

X, it is easy to confirm Hx = feedback(G x , Fx)· This allows us to con-
clude that feedback(G, F) is input-output stable if and only if
feedback(G x , F x) is input-output stable.
 8.3 Method of comparison systems 279

In practice, the input v in (8.2) and (8.3) is often due to nonzero initial
conditions, in which case, we sometimes omit v and write

y = Gu,
u Fy.

We still say the above system is input-ouput stable to mean that the system
described by (8.2) and (8.3) is input-output stable.

8.3 Method of comparison systems

As a simple illustration of the input-output approach, we will discuss the
method of comparison systems. In this method, a time-delay system is
embedded in a system without uncertainty subject to dynamical feedback
uncertainty known as the comparison system. Of course, the stability of the
comparison system implies the stability of the original time-delay system.
In our development, we will illustrate some features of the input-output
approach of robust stability analysis to be developed in later sections.

8.3.1 Problem setup

Consider a system with single delay

z(t) = Aoz(t) + AIZ(t - r), t ~ 0, (8.5)

where z(t) E ]Rn, A o, Al E lR,nxn. As usual, the initial condition is in the

form of
Zo = ¢ E C.
Then, for an arbitrary matrix C, we can write

1:
z(t) (Ao + C)z(t) + (AI - C)z(t - r) - C[z(t) - z(t - r)]

(Ao + C)z(t) + (AI - C)z(t - r) - C z(t + O)dO.

Using (8.5) in the last term, the above becomes

x(t) = (Ao+C)x(t)+(AI-C)x(t-r)-C lO}Aox(t+O)+AIx(t+o-r)]do.

(8.6)
Notice that we have changed the notation for state variable in (8.6). Similar
to the model transformation discussed in Chapter 5, an initial condition
constraint needs to be imposed for x(t + 0), -r ::; 0 ::; r to make them
280 8. Input-Output Stability

equivalent. The stability analysis disregarding this initial condition con-

straint will introduce additional conservatism due to additional dynamics.
The process of transforming (8.5) to (8.6) without initial condition con-
straint is known as a parameterized model transformation. Indeed, if Cis
chosen as AI, this becomes the model transformation discussed in Chapter
5. If C = 0, then (8.6) reverts back to (8.5).
System (8.6) can be written as

x(t) = (Ao + C)x(t) + (AI - C)UI(t) + rCU2(t) , (8.7)

YI(t) = x(t), (8.8)
Y2(t) Aox(t) + AIUI(t), (8.9)

with feedback

UI(t) = YI(t - r), (8.10)

U2(t) -~ lOr Y2(t + fJ)dfJ. (8.11)

Let uT = (uf uf) and yT = (yf yf). Then we can write the
forward system described by (8.7) to (8.9) as

y=Gu, (8.12)

and the feedback system described by (8.10) and (8.11) as

u=ay. (8.13)

This casts the problem in an input-output form. From the small gain the-
orem, and the equivalence of stability for feedback (G x, F x) and that of
feedback(G, F), we can conclude that the system is input-output stable if
we can find a real nonsingular matrix X such that

< 1, (8.14)
< 1. (8.15)

We will approach the problem in a robust stability framework. First we

identify a family of real nonsingular matrices X, such that (8.15) is satisfied
for all X EX. Then we will check whether there exists an X E X such that
(8.14) is satisfied. The system is declared stable if such an X exists. In this
procedure, we are effectively considering the robust stability problem of the
system by thinking of a as a class of systems, or dynamical uncertainty,
which satisfies (8.15) for all X E X. The explicit stability conditions can
be derived through either a frequency domain approach or time domain
approach. We will illustrate both approaches in the following subsection.
 8.3 Method of comparison systems 281

8.3.2 Frequency domain approach

The forward system G is a linear time-invariant system without delay, with
transfer function matrix

G(s) = 6(s1 - ..4)-113 + b,

where

..4 Ao+C, (8.16)

13 ( Al - C rC), (8.17)

6 = (~:), (8.18)

b (1 1 ~). (8.19)

The feedback system Ll has a "block-diagonal structure"

Ul = Ll 1 Yb (8.20)
U2 = Ll2Y2, (8.21)

where Lll and Ll2 are described by (8.10) and (8.11), respectively. In fact,
both Lll and Ll2 are "linear scalar systems" with Hoo norms bounded by
one. In other words, the transfer function matrices of Llk are

(8.22)

where Ok (s), k = 1, 2 are stable scalar transfer functions satisfying

II0k(S)lloo = sup 10k(S)I::; 1, k = 1,2. (8.23)
Re(s);:::O

Indeed, Ok(S) in (8.22) can be obtained by taking the Laplace transform of

(8.10) and (8.11)
e- rs ,
1 - e- rs
rs
which can be shown to satisfy (8.23). Thus we can embed the system de-
scribed by (8.5) in the uncertain system feedback(G, Ll), consisting of G
described by (8.7) to (8.9) and the dynamical feedback uncertainty Ll de-
scribed by (8.20) and (8.21), with Lll and Ll2 being arbitrary dynamical
systems satisfying (8.22) and (8.23).
The transfer function matrix of Ll can be written in a block diagonal form
as ~(s) = diag (01(s)1 02(s)1). Using the small gain theorem, since
11~lloo ::; 1, a sufficient condition for the system to be input-output stable
282 8. Input-Output Stability

is /'0 (G) =IIG(s)lloo < 1. We can obtain a stronger result by scaling. Let X
consist of nonsingular block-diagonal matrices of appropriate dimensions,

X = {diag (Xl X2 ) I Xl, X 2 E jRnxn and nonsingular} . (8.24)

It is easy to see that for any X EX,

(8.25)

We can conclude that the system is stable if there exists a nonsingular

block-diagonal matrix X such that

(8.26)

A frequency-sweeping method can be used to check the satisfaction of

(8.26). Indeed, the satisfaction of (8.26) for some nonsingular X with block-
diagonal structure expressed in (8.24) is equivalent to the satisfaction of

GT ( -jw)ZG(jw) < Z, os w < 00,

jjTziJ < Z,

for some block-diagonal symmetric positive definite matrix

We can use the following Bounded Real Lemma to write (8.26) in the form
of a frequency-independent LM!.

Lemma 8.1 (Bounded Real Lemma) A linear system H described by

x(t) = Ax(t) + Bu(t),

y(t) Cx(t) + Du(t),
is internally stable and satisfies

/'o(H) < 1, (8.27)

if and only if there exists P E jRnxn such that

pT =P >0 (8.28)

and

are satisfied.
 8.3 Method of comparison systems 283

Indeed, using the Bounded Real Lemma, we can conclude that

is satisfied if and only if there exists aPE jRnxn such that (8.28) and

are satisfied. From the above discussions, we can conclude the following.

Proposition 8.2 The system described by (8.5) is input-output stable if

there exist real symmetric matrices P, Zl, and Z2 such that the following
linear matrix inequalities are satisfied:

P>O,

P(A 1 - C) + Air Z2 A I
-Zl + Aj Z2A I
rpc)
o < 0,
o -Z2
(8.31)
where

Proof. First, notice that (8.31) implies Zl > 0 and Z2 > 0 in view of
the (3,3) and (2,2) entries of the matrix. From the above discussion, the
system is stable if (8.30) is satisfied for some nonsingular X E X described
in (8.24). Left-multiply the second row of (8.30) by XT and right-multiply
the second column by X; then define

Z = diag (Zl Z2) = XT X

and use the definitions (8.16) to (8.19) . •
It is interesting to observe that the above criterion is equivalent to the
delay-dependent stability criterion obtained by using implicit model trans-
formation and the Lyapunov-Krasovskii functional approach discussed in
Chapter 5. Indeed, by eliminating the variable X in (5.81) and (5.85),
changing variables S = ZI, Y = PC, Z = ~Z2' and multiplying -1 in
the third row and column, we obtain (8.31). This also implies that the
delay-independent stability criterion and the delay-dependent stability cri-
terion using explicit model transformation obtained using the Lyapunov-
Krasovskii functional discussed in Chapter 5 are equivalent to setting C = 0
and C = AI, respectively, in (8.31).
As was discussed in Chapter 5, one source of conservatism of stability
condition derived by using the simple Lyapunov-Krasovskii functional is
due to the additional dynamics introduced in the model transformation
284 8. Input-Output Stability

process. The equivalence established above revealed another source of con-

servatism since the stability condition in this section is obtained by treating
some time-delay elements as uncertainty, i.e., embedding it in a class of dy-
namical systems.
It should be noted, however, that if we restrict ourselves to linear time-
invariant systems, it is possible to allow X and Z to depend on the fre-
quency w in many situations to reduce conservatism, as is often done in
the robust stability literature.

8.3.3 'Rime domain approach

The above conclusion can also be obtained by using a time domain ap-
proach. Let Z be the class of positive definite block-diagonal matrices,

Clearly,
Z = {XTX I X EX}.

First we will show that the operator A = diag( Al A 2 ) described by

(8.10) to (8.21) satisfies the following integral quadratic constraint (IQC)

(8.32)

for any t > 0, Y E L 2e +, U = Ay, and Z E Z. For any X E X, (8.32)

implies that
(8.33)

since

lot IIAxY(T)II dT -lot IIY(T)II dT

2 2

lot [AY(T))]T ZAY(T))dT -lot i?(T)ZY(T)dT

< 0,

where
 8.3 Method of comparison systems 285

To show (8.32), let u = (u[ ur f = (41Y[ 42yr )T, and using

zero initial conditions, we have

lot uf(T)ZlUl(T)dT
= I:r yf(T)ZlYl(T)dT

Io yf(T)ZlYl(T)dT
t r
-

~ lot yf(T)ZlYl(T)dT. (8.34)

Also, using Jensen's Inequality and exchanging the order of integration, we

can obtain

lot UnT)Z2U2(T)dT
= :2 lot [lOr Y2 (T + O)dO] T Z2 [lOr Y2 (T + O)dO] dT
< ~ lot [lOr ynT+O)Z2Y2(T+O)dO] dT
= ~ [Or [lot ynT + O)Z2Y2(T + O)dT] dO
~ ;1: [t' yi(T)z"l/2(T)dT] dO

< ~ [Or [lot ynT)Z2Y2(T)dT] dO

= lot ynT)Z2Y2(T)dT. (8.35)

Inequalities (8.34) and (8.35) imply (8.32), which is equivalent to (8.33).

With (8.33) established, the input-output stability can be established if
there exists a X E X such that

'YO(Gx) < 1. (8.36)

To study the condition for (8.36), consider a Lyapunov function

V(x) = xTpx. (8.37)

The following expression is a quadratic expression of x and u:

W(x, u) = V{x) + yT Zy - uT Zu, (8.38)

286 8. Input-Output Stability

where y = (yi yf) T represents the linear expression (8.8) and (8.9) of
x and u. If

V(x) ~ cllxl1 2, (8.39)

W(x, u) S -c(llxI1 2+ IluW), (8.40)
for some c > 0 and arbitrary x and u, then (8.36) is satisfied. To show
this, integrate (8.40) from 0 to t, using zero initial condition (and therefore
V(x(O)) = 0), we obtain

V(x(t)) + Iot[yT(r)Zy(r)-uT(r)ZU(r)]dr S -c Iotlllx(r)112+llu(r)112]dr.

(8.41)
Let

Then (8.41) implies

lot IIXy(r)11 2 dr - (1- 8) lot IIXu(r)11 2 dr

< lot [yT(r)Zy(r) - uT(r)(Z - d}u(r)]dr

< -c lot Ilx(r)11 2dr - V(x(t))

< O.
This shows
'Yo(Gx) S (1 - 8) < 1.
The above discussions indicate that the system is input-output stable if
conditions (8.39) and (8.40) are satisfied for some c > 0 and arbitrary x
and u. Clearly, condition (8.39) is equivalent to
P>O. (8.42)
Also, writing out the explicit expression of W,
W(x,u)
2xT P(Ax + Bu) + (Cx + Du? Z(Cx + Du) - uT Zu
( T T)(PA_+ATPj-CT_ZC PfJ+C}TZD)(X)
x u BTp+DTZC DTZD-Z u'
we can conclude that

( PAj-ATPj-C-z:.ZC PfJ + C}TZD ) 0

(8.43)
BTp+DTZC DTZD-Z <
 8.4 Scaled small gain problem 287

is a sufficient condition for (8.39). Therefore, the system is stable if (8.42)

and (8.43) are satisfied. This is equivalent to Proposition 8.2.

8.4 Scaled small gain problem

We will now consider the more general setting with the possibility of a
time-delay system as the nominal forward system. Let G be described by
the functional differential equations

±(t) = f(t, Xt, u(t)), (8.44)

y(t) h(t, Xt, u(t)), (8.45)

where f and h satisfy

f(t,O,O) 0, (8.46)
h(t, 0, 0) = 0, (8.47)

and x(t) E lR.n, u(t) E lR.m, y(t) E lR. m. Notice that we have constrained
u(t) and y(t) to have the same dimension. This constraint is imposed for
notational convenience and can be relaxed. Let X c lR. mxm be a given
set of real nonsingular matrices. The scaled small gain problem is to check
whether G is internally stable and there exists an X E X such that

l'o(Gx ) < 1. (8.48)

The special case of X = {I}, i.e., check whether G is internally stable and
satisfies
l'o(G) < 1, (8.49)
is known as the small gain problem or 1£00 problem. When G is linear
and time invariant, a frequency domain technique can be used. Indeed, a
frequency-sweeping method is often possible to check (8.48). In the follow-
ing, we will discuss the time domain approach.
Let
z = {XT X I X EX}. (8.50)

Then, the scaled small gain problem is equivalent to finding whether there
exists a Z E Z such that

(8.51)

is satisfied for any t, u and y = Gu. To achieve (8.51), we have the following.
288 8. Input-Output Stability

Proposition 8.3 Let Wk : i+ ~ i+, Wk(O) = 0, Wk(S) > 0, for s > 0,

k = 1,2,3, and lim Wl(S) = 00. Letf: lRxCxlRm ~ lRn takelR x (bounded
8-+00
sets of C x lRm) into bounded sets of lRn , and satisfy {8.46}. Also let h :
lRxCxlRm ~ lRm be continuous and satisfy (8.47). Then system G de-
scribed by {8.44} and {8.45} is internally stable, input-output stable, and
satisfies {8.48} for some X E X if there exists a Z E Z and a Lyapunov-
Kmsovskii functional V(t, xt} that satisfies
(8.52)
and the functional
W(t,Xt,u(t)) = V(t, Xt) + yT(t)Zy(t) - uT(t)Zu(t) (8.53)
satisfies
W(t,Xt,u(t)):::; -w3(llx(t)11) -E'lIu(t)1I 2 , (8.54)
where V denotes the derivative of V along the system tmjectory, and y(t)
denotes the expression {8.45}.

Proof. Setting u = ° (8.54),

in we obtain
V(t,Xt):::; -w3(lIx(t)lI) _yT(t)Zy(t):::; -w3(lIx(t)II),
which together with (8.52) proves the internal stability by invoking the
Lyapunov-Krasovskii Stability Theorem in Chapter 1. It remains to be

°
proven that (8.48) is satisfied, or equivalently, (8.51) is satisfied. Integrating
(8.54) from to t, we obtain

V(t,xt} - V(O,xo) + lot[yT(r)Zy(r) - uT(r)Zu(r)]dr

< -lot w3(lIx(r)lI)dr -lot wT(r)u(r)dr.

Due to zero initial condition Xo

Therefore,
= 0, we have V(O, xo) = °in view of (8.52).

lot [yT(r)Zy(r) - uT(r)Zu(r)]dr

< -lot w3(lIx(r)lI)dr -lot wT(r)u(r)dr - V(t,Xt)

< -lot wT(r)u(r)dr,

i.e.,

. lot yT(r)Zy(r)dr < lot uT(r)(Z - E'I)u(r)dr

< ,.? lot uT(r)Zu(r)dr,

 8.4 Scaled small gain problem 289

where

When the system G is linear, and the functional V(t,Xt) is uniformly

•
bounded quadratic, then W(t, Xt, u(t)) is a quadratic functional of Xt and
u(t). The upper bound of V(t, Xt) can always be satisfied for w2(1lxtllc) =
Kllxtll~ for some sufficiently large K. We can choose

for sufficiently small E > 0, and conditions (8.52) and (8.54) become
V(t, Xt) > Ellx(t)112, (8.55)
W(t,Xt,u(t)) < -Ellx(t)112 -Ellu(t)W. (8.56)
In this case it is often possible to reduce the conditions into an LMI prob-
lem.
As a simple illustration, consider a system G described by the following
equations

x(t) Aox(t) + AlX(t - r) + Eu(t), (8.57)

y(t) Gox(t) + Glx(t - r) + Du(t). (8.58)

Choose a simple Lyapunov-Krasovskii functional

where pT = P > 0, ST = S > o. It can be easily calculated that

W(t, ¢, u) = ¢;,)I¢xu,
where

and

II = ( PAo + Aifp + S +. C;rZGo PAl + GifZGl PE+Gif ZD )

-S+G[ZG l G[ZD .
symmetrIc DTZD-Z
(8.60)
Therefore, we can conclude the following.

Proposition 8.4 Given a set of nonsingular matrices X, let Z = {XT X I

X E X}. There exists an X E X such that 'Yo(G x ) < 1 if there exists a
Z E Z, P > 0 such that II < o.
290 8. Input-Output Stability

Similar to the stability problem, the simple Lyapunov-Krasovskii func-

tional (8.59) yields a delay-independent result and is often too conservative.
A more accurate and meaningful result can be obtained by using a complete
quadratic Lyapunov-Krasovskii functional

V(t, ¢) = ¢T(O)p¢(O) + 2¢T(0) [Or Q(~)¢(~)d~

+ [Or [[Or ¢T(~)R(~, 1])¢(1])d1]] d~

+ [Or ¢T(~)S(~)¢(~)d~, (8.61)

where Rand S satisfy the symmetry as discussed in Section 5.7 of Chapter

5. Then we can calculate

W(t,¢,u)
= -¢T(O)[-PAo - A5 P - Q(O) - QT(O) - S(O)]¢(O)
_¢T (-r)S( -r)¢( -r)

_[Or ¢T (~)S(~)¢(~)d~
- [Or d~ [Or ¢T(~) [:~R.(~,1]) + :1]R(~,1])] ¢(1])d1]
+2¢T(0)[PAl - Q(-r)]¢(-r)

+2¢T(0) [Or [A5Q(~) - Q(~) + R(O,~)] ¢(~)d~

+2¢T(-r) [:[AfQ(~) - R(-r,~)]¢(~)d~
+2¢T(0)PEu + 2u T ET [Or Q(~)¢(~)d~ - u T Zu

+( Go¢(O) + G1 ¢( -r) + Du f Z( Go¢(O) + G1¢( -r) + Du). (8.62)

We can write the conditions in an LMI form by the discretized Lyapunov
functional method. Using the same notation as in Section 5.7, choose Q, S,
and R to be piecewise linear, then, as was already shown in Proposition 5.21
in Chapter 5 that a sufficient condition for (8.55) is Sp > 0, p = 0, I, ... , N
and
(8.63)

Furthermore, it can be shown that W(t, ¢, u) is equal to the expression of

V in (5.122) of Chapter 5 plus the last four terms (three terms involving
 8.4 Scaled small gain problem 291

u and the last term) in (8.62). Therefore, we can easily modify (5.146) in
Chapter 5 to obtain

W{t,r/J,u) = -r/J~rul:::.r/Joru -111t{a)Bd¢{a)da

-1 [1
1 1
¢T {a)Rd¢(I3)da] d{3

+2r/J~ru 11 [DB + {1- 2a)D a ] ¢{a)hda, (8.64)

DB = (~; ),
Da = (~; ),
and

DBu h T( Ql +Qo Q2+Ql

"2E QN+QN-l),

D ua = h T( QI-QO Q2-Ql
-E QN -QN-l ).
2
Then the same manipulation as in Proposition 5.21 of Chapter 5 allows us
to conclude that (8.56) is satisfied if

(8.65)

We can therefore conclude the following.

Proposition 8.5 Consider a system G described by (8.57) and (8.58), a

given set of nonsingular matrices X, and let Z = {XT X I X E X}. There
exists an X E X such that 'Yo{Gx) < 1 if there exist a Z E Z, real
matrices P = pT, Qp, Bp, Rpq = R~p, p = 0,1, ... , N; q = 0,1, ... , N such
that (8.63) and (8.65) are satisfied.

Proof. The only thing that needs to be shown is that Bp > O. But it is
already implied by (8.65) in view of the (3,3) entry of the matrix in the
left-hand side. •
292 8. Input-Output Stability

The above proposition concerns the case with single delays. However, the
same ideas apply to the case where G is a system with multiple pointwise
delays or distributed delays with piecewise constant coefficients. We can
make similar modifications to the results in Chapter 7 to derive sufficient
conditions for 'Yo(Gx) < 1 in an LMI form.

8.5 Robust stability under dynamical uncertainty

8.5.1 Problem setup
Consider a system H formed by a feedback loop consisting of a forward
system G and a feedback uncertain system 4:
y = Gu, (8.66)
u = 4y, (8.67)
where G is described by the functional differential equations (8.44) and
(8.45). System 4 is unknown except that it belongs to a set V of input-
output stable systems. The robust stability problem is to find conditions
such that G is internally stable and the feedback system H so formed is
input-output stable for all 4 E V.
Let X be a set of nonsingular matrices such that
'Yo(4x):51 for all X E X and 4 E V. (8.68)
For our purpose, the set V is characterized by the set X. To obtain a
less conservative result, use a large X while remaining mathematically
tractable. Using the small gain theorem, we conclude the following.

Proposition 8.6 The feedback system H described above is input-output

stable if there exists a X E X such that
(8.69)

Proof. This is a direct consequence of the small gain theorem and the
fact that the input-output stability of feedback(G, 4) is equivalent to the
input-output stability of feedback( G x, 4 x). •
Thus the robust stability problem under dynamical uncertainty is trans-
formed to the scaled small gain problem discussed in the last section.

Example 8.1 Consider a uncertain system H consisting of forward sytem

G described by (8.57) and (8.58) and uncertain dynamical feedback ~. It is
known that ~ is input-output stable with gain no more than one. In other
words, the uncertainty set D can be described by
v= {414input-output stable, 1'0(4) :5 I}.
 8.5 Robust stability under dynamical uncertainty 293

We want to find a robust stability condition for H. We can choose X =

{1£1 I I£i- O}. Clearly, (8.68) is satisfied. Therefore, according to Proposition
8.6, H is input-output stable if there exists an X = 1'1 such that (8.69) is
satisfied. Proposition 8.5 indicates that we only need to satisfy (8.63) and
(8.65) with Z = >.1, oX > O. This is a set of LM!. 0

The above example is a generalization to Proposition 6.20 in Chapter 6,

where the class of systems considered is restricted to D = 0 and memo-
ryless feedback uncertainty U = F(y, t)y with the uncertain matrix F(y, t)
satisfying IIF(y,t)11 ~ 1. Indeed, with D = 0, the stability conditions in
the above example are equivalent to the conditions for Proposition 6.20 in
Chapter 6. To see this, we only need to show the equivalence between (8.65)
and (6.115) in Chapter 6. This can be accomplished by a permutation of
rows and columns, and setting oX = 1 (this is without loss of generality since
we can scale other LMI variables P, Q k, etc., by a factor of oX).

8.5.2 Uncertainty chamcterization

A very large class of uncertain systems can be modeled by (8.66) and
(8.67) with A having a "block-diagonal" structure through a process known
as "pulling out the uncertainty," proposed by Doyle. Such an uncertainty
structure has been discussed extensively in the literature. We will only
briefly discuss some main points and refer the details to the literature
cited.
Let U and y be partitioned as

(U[ ur ui ),
(y[ yf yJ ),
with Uk(t) E lR.m ", Yk(t) E lR. mk , k = 1,2, ... ,£, and A is such that Uk
depends only on Yk:
Uk = AkYk, Ak E'Dk .

Notice that we have constrained Uk and Yk to have the same dimension.

Again, this constraint is imposed for notational convenience and can be
lifted. Symbolically we can write

and
'Yo(Ak) ~ 1 for all Ak E'Dk.
The scaling matrix set X also has a block-diagonal structure:
294 8. Input-Output Stability

Depending on the specific known information about the uncertainty set V k ,

we can specify the corresponding scaling set Xk to satisfy
IO(a kXk ) :::; 1 for all X k E X k and a k E V k ,
where a kXk is defined as

akXJ = Xkak(Xk'l I), f E L 2e+.

If a k has no other constraint except IO(ak) :::; 1, which is known as a

full block, then we can let X k = {ILk1mk I ILk E JR, ILk i= O}. The set X k
can often be enlarged when a k has more known constraints. An important
example is repeated linear blocks. In this case, we can further partition

where Uki(t) E JRmk, mki(t) E JRm\ mk = mk/£k, and each Uki depends
only on Yki with an identical linear mapping
Uki = AkYki, i = 1,2, ... , £k, Ak E ih,
where Ak is linear. The most common such linear blocks are (a) lin-
ear time-invariant system with transfer function matrix ~k(S): Uki(S) =
~k(S)Yki(S), and (b) time-varying memoryless gain matrix ~k(t): Uki(t) =
~k(t)Yki(t). Of course, linear time-varying dynamical systems also fall into
this category. Due to linearity, we can write

Therefore, we can choose X k = {MkQ9Imk I Mk E JR£k xt\ Mk nonsingular},

where Q9 represents Kronecker product of matrices. The special case of
mk = 1 is known as a scalar block, in which case X k consists of all the real
nonsingular matrices.
Another important case is the time-invariant algebraic block

where the Jacobian matrix 8Uk/8Yk is symmetric and satisfies 118uk/8Ykll :::;
1. The symmetry of Jacobian matrix is satisfied by many system compo-
nents due to physical laws. The conservatism can be further reduced by
using multipliers. We will not discuss multipliers here.
When the systems discussed are linear and time invariant, it is also pos-
sible to have the scaling matrix X be frequency dependent, which allows
further reduction of conservatism. The theory of structured singular value
indicates that even with frequency-dependent scaling, there may still be
conservatism except when the number of blocks are sufficiently small.
 8.5 Robust stability under dynamical uncertainty 295

8.5.3 Robust small gain problem

With the framework discussed so far, it is not much more difficult to discuss
the problem of guaranteed gain under block-diagonal uncertainty. Let

where Uk,Yk E ]Rmk. The system is subject to the feedback uncertainty

The resulting feedback system can be written as

Of course, the system H depends on a. The robust small gain problem is

to check whether the resulting feedback system satisfies

/'o(H) < 1 for all a E D. (8.70)

Let "-"2 satisfy

/'0 (ax ) ::; 1 for all X E "-"2 and a ED, (8.71)
and let

x = {diag( 1m X) I XE"-"2},
Z {diag( 1m X T X ) I X E "-"2} .
If system G is described by the functional differential equation (8.44) and
(8.45) satisfying (8.46) and (8.47), then (8.70) is satisfied if the conditions
in Proposition 8.3 are satisfied for the above defined "-" and Z. To see this,
integrate (8.53) from 0 to t, considering zero initial conditions, to obtain

Io (IIYl(r)11 2 - II ul(r)II 2)dr

t

< -lot w3(IIx(r)II) - lot IIul(r)II2dr - lot IIu2(r)II2dr

E E

-V(t,xt} -lo [yf(r)ZY2(r) - ur(r)Zu2(r))dr

t

< -E lot IIul(r)//2dr -lot[yr(r)ZY2(r) - ur(r) Zu2(r)]dr.

But (8.71) guarantees

Io [yr(r)ZY2(r) - ur(r)Zu2(r)]dr ~ o.
t
296 8. Input-Output Stability

It It
Therefore,
IIY1{r)11 2dr ~ (I - c) lIul{r)1I2dr,

which implies (8.70).

8.6 Approximation of delay elements

As an application of the above theory, we will consider the approximation
of time delays in the stability analysis. As has been shown in the pre-
vious chapters, the stability problem of time-delay systems in general is
formidable from the numerical computation point of view, and the class of
systems we can treat directly is very limited. To extend the applicability
of these methods to more general cases, we can consider appropriate ap-
proximations. The errors of such approximation can often be modeled as
dynamical feedback uncertainties. The input-output approach is very con-
venient in analyzing the stability of the resulting feedback system, which
guarantees the stability of the original system. This approach can be re-
garded as an extension to the method of comparison systems by allowing
the nominal system to be a time-delay system. We will only show the gen-
eral idea here. Most formulations discussed can be refined to obtain less
conservative stability criteria.

8.6.1 Approximation of time-varying delay

To illustrate the approximation of time-varying delay, consider the following
system
i{t) = Ao{t)z{t) + Al{t)Z{t - r{t)). (8.72)
In Chapter 6, we used the Razumikhin Theorem to study the stability of
the above system. Here we assume that additional information about the
delay r is available:

rm < r{t) ~ rM, (8.73)

r{t) < p, (8.74)

where rM > rm > 0 and 0 ~ p < 1. It seems reasonable to approximate

the time-varying delay by a time-invariant delay r a, ra E [rm, rMJ. Using
(8.72), we can write

z{t - r{t)) = z{t - ra) _I t ra

- z{r)dr

_I
t-r(t)

t ra
z{t - ra) - [Ao{r)z{r) + A 1{r)z{r - r{r))Jdr.
t-r(t)
 8.6 Approximation of delay elements 297

Substitute (8.72) by the above to obtain

x(t) = Ao(t)x(t) + At (t)x(t - ra)

-At(t) I t r
- ,, [Ao(r)x(r)dr
t-r(t)
+ At (r)x(r - r(r))]dr. (8.75)

Notice that we have changed notation for the state variable in (8.75). Sim-
ilar to the case discussed in Chapter 5, (8.75) is equivalent to (8.72) only
if a constraint is applied to the initial condition. Without the constraint,
the stability of (8.75) implies that of (8.72) but not vice versa because of
additional dynamics, although when rM - rm is sufficiently small, the ad-
ditional dynamics will not be unstable. The integration term in (8.75) is
considered to be the uncertainty.
System (8.75) can be written as

(8.76)

(8.77)

where the forward system G is described as

x(t) = Ao(t)x(t) + At(t)x(t - ra) - rdAt(t)U2(t), (8.78)
1
Yt (t) = y'f"=px(t), (8.79)

Y2(t) = Ao(t)x(t) + At (t)Ut (t), (8.80)

where
rd = max{rM - ra,ra - r m }, (8.81)
and the feedback A is described as
(8.82)

(8.83)

To obtain the smallest coefficient in the uncertainty term (the last term in
(8.78)), we can choose ra = (rM + r m )/2. We will show that
'Yo(A kXk ) :$ 1, for all nonsingular X k E lR nxn , k = 1,2. (8.84)
Once (8.84) is established, we can conclude the following.

Proposition 8.7 The system described by (8.72) is input-output stable if

the scaled small gain problem
'Yo(Gx ) < 1 for some X = diag (Xt X 2 ), Xl, X 2 E lR nxn nonsingular
has a solution, where G is described by (8.78) to (8.81).
298 8. Input-Output Stability

We can use the method discribed in Section 8.4 to obtain an LMI con-
dition. If the coefficient matrices Ao and Al are constant matrices, we can
also use a frequency domain method.
Proof. From the above discussion, we only need to show (8.84), which
is equivalent to

Ik = lt UI(T)ZkUk(T)dT ~ lt yI(T)ZkYk(T)dT (8.85)

for Uk = tlkYk and all Zk E IR nxn , Zk = Z[ > O. For k = 1,

II = lt Ui(T)ZIUI(T)dT = (1- p) lt yf(T - r(T))ZIYI(T - r(T))dT.

Let B = p(T) = T - r(T). Due to the derivative bound, the inverse T =

p-I(B) = q(B) is well defined and differentiable, and satisfies
0+ rm ~ 0 + rM,
q(O) ~ (8.86)
1
o < q'(O) ~ - - .
I-p
(8.87)

Change integration variable 0 = p(T), considering (8.87) and the zero initial
condition, to obtain

(1- p) I t - r (t)

-r(O)
yf(O)ZIYI(O)q'(O)dO
rt-r(t)
(1- p) Jo yf(O)ZIYI(O)q'(B)dB

< lt yf(O)ZIYI(B)dO,

which is (8.85) for k = 1. For k = 2,

I2 = lt Uf(T)Z2U2(T)dT

12
rd
t
Jo
[I T

T-r(T)
-
ra
Y2(B)dB] T Z2 [I T
-
T-r(T)
ra
Y2(O)dB] dT.

lt l
Using Jensen's Inequality, we obtain
1 max {T-r(T),T-r a }
I2 ~ 2' Ir(T) - ral ynB)Z2Y2(O)dBdT.
rd 0 min{ T-r(T),T-r a }

Exchange the order of integration, considering the zero initial condition

[l
(Y2(O) = 0 for all B ~ 0), to obtain
1 It-min{ra,r(t)} min {Q(o),t} ]
I2 ~ 2' Ir(T) - ral dT ynO)Z2Y2(B)dO,
rd 0 min{,2(O),t}
 8.6 Approximation of delay elements 299

where

Q(B) = max{B + r a , q(B)} ,

Q(B) min{B + r a , q(B)}.
Since
Ir(r) - ral ~ rd,
min{max{B + ra, q(B)}, t} - min{min{B + ra, q(B)}, t} < rd,

l
we have
min{ma.x{6+ r",q(II)}.t}
o~ Ir(r) - raldr ~ r~.
min{min{6+r",q(II)},t}
Therefore,

Thus, (8.85) for k = 2 is also established. _

8.6.2 Approximation of distributed delays

To illustrate the dynamical uncertainty introduced by approximating the
coefficient matrix of distributed delays, consider the system

x(t) = lOr A(t, O)x(t + O)dO. (8.88)

We would like to approximate the matrix A(t, 0) by a piecewise constant

matrix in 0. Specifically, let the interval [-r, 0] be divided into N segments,
[Bp, Bp - 1], P = 1,2, ... , N,
-r = BN < BN- 1 < ... < Bo = O.
We will also write the length of each segment hp = Bp - 1 - Bp. With this
division, we can write the system as

x(t) = L In
N rllp - 1

A(t, B)x(t + B)dB.

p=l lip

It seems reasonable to make the following approximation

;,(t) ~ t, 1:'-' Ap(t)x(t + O)dO + u(t), (8.89)

300 8. Input-Output Stability

where
Ap(t) = h1
p
l 1JP

9p
-
1
A(t, fJ)dO, (8.90)

and u( t) represents the error of such an approximation and can be explicitly

expressed as

To estimate the error bound, we first give the following result.

Lemma 8.8 A system u = Fy described by

u(t) lOr U(t, fJ)y(t + O)dO,

y(r) = 0, r ~ 0,
satisfies

Proof. Using Jensen's Inequality, we have

lt uT(r)u(r)dr

lt [lOr [lOr U(r,O)y(r + O)dO] dr

lt l:
U(r,O)y(r + O)dO]T

< [r yT(r+fJ)UT(r,O)U(r,fJ)y(r+O)dO] dr

r lOr [lot yT(r + O)UT(r, O)U(r, O)y(r + O)dr] dO

~ r 1'. [1.'+0 yT(r)UT(T-O,O)U(T-O,O)Y(T)dT] dO

< r lOr [lot yT(r)UT(r - O,O)U(r - O,O)y(r)dr] dO

lot yT(r) [r lOr UT(r - O,O)U(r - fJ,O)dO] y(r)dr

< [r lOr (T:5' IIU(t, 0)11 dO] lot yT(r)y(r)dr.
2)

•
From the above lemma, we can immediately conclude the following.
 8.6 Approximation of delay elements 301

Proposition 8.9 The system described by {8.88} is input-output stable if

the system G described by {8.89} and
y(t) = ax(t) (8.92)
satisfies
'Yo(G) < 1, (8.93)
where a is a constant satisfying

(8.94)

Proof. System (8.89) can be written as

y Gu,
u = /l.y,
where /l. is described by

1 N 18P-1
u(t) = - L (Ap(t) - A(t, B))y(t + B)dB. (8.95)
a p=l 8p

It is easy to conclude from Lemma 8.8 that 'Yo(/l.) :::; 1. Therefore, (8.93)
is sufficient for stability. •
In applying the above proposition, we may choose, for example,
N
a = r Lhpe~, (8.96)
p=l

where
ep = ~~ IIAp(t) - A(t, B)II. (8.97)
8E(8~8p-IJ
As pointed out in Section 8.4, a numerical method can be devised to check
(8.93).
When hp, p = 1,2, ... , N are small, a smaller error bound can often be
obtained by carrying out model transformation: for B E [Bp, Bp - 1 ],

x(t + B) = x(t + Bap) - I t +8ap

t+8
x(T)dT

x(t + Bap) _I [1 A(T,~)X(T

t

t+1I
+
8ap 0
-r
+ ~)d~] dT,(8.98)
where
Bap = (Bp + Bp - 1 )/2. (8.99)
Using the above in (8.91), we may conclude the following.
302 8. Input-Output Stability

Proposition 8.10 The system described by {8.88} is input-output stable

if the system C described by {8.89} and
yet) = (3x(t) (8.100)
satisfies
(8.101)
where

~ ~ ~ [L (rg~rIIA(t,O)II') dot,h;e;]'" (8.102)

Proof. Use (8.98) in (8.91), taking into account (8.90) and (8.100), to
obtain

1 N
u(t)=-73~lop t Jp
-
1
(Ap(t)-A(t,O)) [I +Oap 1_rA(T,~)Y(T+~)d~dT
t
t+9
0
] dO.
(8.103)
Therefore, the system (8.89) can be written as
y = Cu,
u = li y ,
where Ii is defined in (8.103). Define

z(t) = AIY(t) = [Or A(t, ~)y(t + ~)d~. (8.104)

Then Lemma 8.8 indicates that

(8.105)

Using (8.104) in (8.103), we have

u(t) ~ -~ t, f-' (Ap(t) - A(t, 0» [ ( ' . ' Z(r)dr] dO

-~ t, f-' (Ap(t) - A(t, 0» [to z(t +r)dr] dO.

Exchanging the order of integration, we obtain

u(t) ~ ~ t, [{-. t-· (Ap(t) - A(t, O»dOz(t+ r)dr

- rOap
lop lop
r
(Ap(t) - A(t, O))dOz(t + T)dT] ,
 8.6 Approximation of delay elements 303

1
or
0
u(t) = A2Z(t) = (31 -r H(t, r)z(t + r)dr,

where

Since

Lemma 8.8 indicates that

We have
1'0(.~) ~ 1'0(A2 )')'o(A 1 ) ~ l.
Therefore, (8.101) is sufficient for stability. •
To compare the two criteria, assume uniform division hp = h = r/N,
and use (8.96) to calculate 0:, then we have

When h is sufficiently small, (3 < 0:, and therefore, 1'o(G) < 1'o(G). There-
fore, Proposition 8.10 is less conservative than Proposition 8.9 when h is
sufficiently small.

8.6.3 Approximation by multiple delays

Consider
x(t) = i: dF(O)x(t + 0),

where F is of bounded variation, continuous from the left on (-r, 0), and
(8.106)

satisfies F(O) = 0, 0 ~ OJ F(O) = F( -r), 0 < -r. We would like to

approximate it by a system with multiple pointwise delays. Let

h = r/N,
rk kh,
rmk (k - 1/2)h,
rMk (k + 1/2)h.
304 8. Input-Output Stability

We can write

:i:(t) = ~[~~~k dF((})x(t + (}). (8.107)

Using

x(t + (}) = x(t - rk) - i- rk

:i:(t + r)dr

= x(t - rk) - i- rk
[[Or dF((})x(t + r + (})] dr
in (8.107), we obtain
N
:i:(t) = I: AkX(t - rk) + U(t), (8.108)
k=O

where

u(t) = - t l- r
k=O -rMk
"'k dF((})
JIJ
r-r k
Z(t + r)dr, (8.109)

and
z(t) = lOr dF((})x(t + (}).
Let
y(t) = ax(t), (8.110)
where
a = J.L2 yfrh,
and J.L is the total variation of F in [-r, 0]:
K
J.L = sup
K>O
I:
i=l
IIF((}i) - F((}i-l)1I = Var F((}).
[-r,O]
-r=IJ O <1J 1 < ... <IJK=O

Then we can write the system (8.106) as

y - Gu,
u 1l.y,

where G is described by (8.108) and (8.110), and

 8.6 Approximation of delay elements 305

where u = .a1z is described by (8.109), and z = .a2 y is described by

1
z(t) = -;; JO dF(B)y(t + B).
-r

With this framework, we will show in the following proposition that 'YO (.a) ::;

°
1, and therefore the stability is assured if 'Yo(G) < 1. An important obser-
vation is IIul12 ::; allxlb, and a -+ as h -+ 0, representing asymptotically
close approximation, which is intuitively clear.

Proposition 8.11 The system described by (8.106) is input-output stable

if the system G described by (8.108) and (8.110) satisfies 'Yo(G) < 1.

Proof. We only need to show 'Yo(.a) ::; 1. For arbitrary interval [a, b]
and scalar function (3( B) 2:: 0, B E [a, b], define

1 b
(3(B)ldF(B)1 = !~
a=8o<8 1 <···<8K=b
8K
(3(Bi)IIF(Bi) - F(Bi-1)11·

The special case of (3(B) = 1 is the total variation of F(B) in [a, b]:

I a
b
IdF(B)1 = Var F(B),
[a,~

especially

I: IdF(B) I = Jj.

With this definition, we have another form of Jensen's Inequality (see [245]):
306 8. Input-Output Stability

We can write

Therefore

Also, exchange the order of integration in (8.109) to obtain

u(t) = t. l~~~k [l- <nkr

dF(B) - l:Mk dF(B)] z(t + r)dr
= t. l~~~k [F( -rmk) - F(r) - F(r) + F( -rMk)Jz(t + r)dr

lOr T(r)z(t + r)dr,

where

T(r) = [F( -rmk) - F(r)]- [F(r) - F( -rMk)], r E [-rMk' -rmk].

Using Lemma 8.8, we can conclude that

 8.7 Notes 307

or

'Y~(al)
r
= ~i~~~k IIT(r)11 2dr
N h/2
= ~ih/2"T(r - kh)1I 2dr

r
h/2 N
= ih/2 ~ IIT(r - kh)1I 2dr

< ih/2
h/2 C
~ IIT(r - kh)1I dr

jh/2
< J.L 2dr
-h/2
hJ.L 2.

Therefore
'Yo(a 1) S JrJ.L2h.
We obtain

'Yo (a) < 'Yo(al ho(a2)

< JrJ.L2hJ.L/a.
l.

•
8.7 Notes
The input-output approach to stability and the small gain theorem are
discussed in more details in Desoer and Vidyasagar [54J and Vidyasagar
[277J, where more general settings are used.
The comparison system approach were explored by, for example, Halanay
[101], Lakshmikantham and Leela [164J, Driver [59J, and Huang and Zhou
[118J. The material presented here is a generalization of Zhang, Knopse,
and Tsiotras [292J. In [292]' C is constrained to AIM and the frequency
domain method was used. Equivalence was established with the stability
criterion proposed in Park [220J (see also the Notes in Chapter 5).
The scaled small gain problem and related passivity problem for finite-
dimensional linear time-invariant systems have been discussed in detail
in, for example, Boyd and Yang [22J and Packard and Doyle [219], with
308 8. Input-Output Stability

emphasis on robust stability. For linear time-delay systems, see Kokame,

Kobayashi, and Mori [151J, De Souza and Li [250J, Gu [85J, and Han and
Gu [108J for the time domain approach and Huang and Zhou [119J for the
frequency domain approach. The process of ''pulling out uncertainty" was
discussed in Doyle, Wall, and Stein [58J. For the multiplier method used for
nonlinear time-invariant uncertainty, see Popov [228J and Bliman [12J for
the single variable case, and Desoer and WU [53J for the multiple variable
case with symmetric Jacobian matrix (known as reciprocal relations, see
Karnopp, Margolis, and Rosenberg [136]).
The time-varying system described by (8.72) to (8.74) was discussed in
Gu and Han [93J, where a more primitive discretized Lyapunov functional
method was used and ra was set at rm. The special case of rm = 0 was dis-
cussed in, for example, Kim [146J for the delay-dependent case and Cao and
Sun [29J, Phoojaruenchanachai and Furuta [226J for the delay-independent
case. The framework presented as an approximation problem is new. It
should also be noted that it is also possible for time-varying delays to have
stabilizing effects, see Louisell [177J.
The importance of studying the stability problem under approximation
is illustrated by a number of examples in the literature: Louisell [180J gave
a system that is stable independent of delays when delays are constrained
to be commensurate, and becomes unstable with a small deviation from
this constraint. Another example was given by Engelborghs, Dambrine,
and Roose [61 J where a stable system with distributed delays obtained
by some infinite-dimensional control design (such as spectrum assignment)
may become unstable when numerical quadrature is used to approximate
the distributed delays.
Appendix A

Matrix Facts
A.I Notation
In this appendix, we will discuss some useful facts about matrices. We
assume the readers are familiar with the basic matrix algebra. The material
presented here is not intended to be self-contained. Rather, it is intended
to fix the notation and point out some facts that may not be familiar to
some readers. For a systematic presentation of these materials, the readers
are referred to the references discussed in the Notes section at the end of
this appendix.
We use Rn to represent the set of real column vectors with n components,
and en of complex column vectors. Similarly, Rmxn and e mxn represent
the sets of real and complex m by n matrices, respectively. We will use Xi
to denote the ith component of column vector X, and aij to represent the
element of matrix A at the ith row and jth column. AT is the transpose
of A. For complex matrices, the Hermitian transpose defined as A H ~ AT
is often used, where A is the complex conjugate of A. In the following, we
will mainly discuss complex matrices, but with obvious specialization to
real matrices. I denotes an identity matrix. If we need to emphasize the
dimention, we may write, for example, In for the n-dimensional identity
matrix. The rank of A, the number of linearly independent columns, is
denoted as rank(A). If rank(A) = mini m, n}, then we say A is full rank.
We say A is of full row rank if rank (A) = m, and full column rank if
rank(A) = n. For a square matrix A E e nxn , we also use tr(A) to represent
the trace of A, i.e.,

n
tr(A) = L aii·
i=t

A block-diagonal matrix with At, A 2, ... , Ap in the diagonal entries is de-

noted as diag (At A2 . . . Ap).
A square matrix A is Hermitian if AH = A. We use A> 0 to denote the
fact that A is Hermitian positive definite. Similarly, "2::," "<," and "~"
denote positive semi-definiteness, negative definiteness, and negative semi-
definiteness in the matrix context. We also use A > B to mean A - B > 0,
with obvious extension to other inequality signs.
310 Appendix A. Matrix Facts

A.2 Determinant
For a square matrix A E cnxn , its determinant is denoted det (A). If A
and B are both square matrices, then it can be shown that

det(AB) = det(A) det(B). (A.I)

Another well-known fact is that

det (~ ~) = det(A) det(D), (A.2)

whenever A and D are both square. If A is square and nonsingular, then

we can use (A.l), (A.2), and the fact

to obtain
det( ~ ~) = det(A) det(D - CA- l B), (A.3)

which is known as the Schur (determinant) complement. Similarly, if Dis

nonsingular, we can show

det (~ ~) = det(D) det(A - BD-lC). (A.4)

Let A = I and D = I in the Schur complements (A.3) and (A.4), we arrive

at another very useful identity. If C B is square

det(I - CB) = det(I - BC).

Especially if k and g are column vectors, we have

A.3 Eigenvalue problems

For a matrix A E cnxn , any solution ..\ of the following nth order polyno-
mial equation
det(M - A) = 0 (A.5)
is known as an eigenvalue and is often denoted as "\(A). Obviously, there
are always n eigenvalues for an n by n matrix counting multiplicity. The
eigenvalues are often ordered according to some scheme, in which case we
 A.3 Eigenvalue problems 311

often use Ai(A) to denote the ith eigenvalue of A. The set of all the eigen-
values of matrix A is known as the spectrum of A and is denoted O'(A).
The spectrum radius is defined as
p(A) = m!1X IAi(A)I.
l~.~n

We also write, similarly,

The eigenvalues of a matrix satisfy the following simple relations

n
L Ai(A) = tr(A) ,
i=l
n
II Ai(A) = det(A).
i=l

Also, if A E (Cmxn and B E(Cnxm, then AB and BA share the same

nonzero eigenvalues.
e
Corresponding to any eigenvalue A, an eigenvector is any E (Cn, =1= 0 e
satisfying
(A.6)
For an eigenvalue A of multiplicity m, if there do not exist m linearly
independent eigenvectors, then we can always find one or more genemlized
eigenvectors 'fJ satisfying
(A - AI)1J = e, (A.7)
e
where is either an eigenvector or a generalized eigenvector. We can always
find a set of m linearly independent vectors consisting of the eigenvectors
and generalized eigenvectors corresponding to an eigenvalue of multiplicity
m.
If A is Hermitian, then all the eigenvalues are real, and there are always
m independent eigenvectors corresponding to an eigenvalue of multiplicity
m. In which case, we use Amin(A) and Amax(A) to denote the minimum and
maximum eigenvalues.
Given two square matrices A and B, a genemlized eigenvalue A is a
solution of the equation
det(AB - A) = 0,
which is often denoted as A(A, B). The spectrum O'(A, B) is the set of
all the generalized eigenvalues. Similarly, an eigenvector in the generalized
e
eigenvalue problem is any =1= 0 satisfying
Ae = ABe.
Notice the difference between the generalized eigenvector and the eigenvec-
tor in the generalized eigenvalue problem.
312 Appendix A. Matrix Facts

A.4 Singular value decomposition

For a given A E em x n, let p = min {m, n}, then we can always write
A=UEV H , (A.8)
where E is a real diagonal matrix

E diag( 0"1 0"2 •.. O"p),

0"1 > 0"2 ~ ••• ~ 0"p ~ 0,

and U E emxp , V E enxp satisfy
UHU = VHV = Ip.
Expression (A.8) is known as the singular value decomposition of A, and
i = 1,2, ... ,n are its singular values. We write
O"i,

O"max(A) max O"i,

l:::;i:::;p

O"min(A) = min O"i.

l:::;i:::;p

We sometimes write 0" max and 0" min as (j and Q, respectively. Singular values
have the following relations with the eigenvalues

A.5 Norms
A norm II . II is a function defined on en satisfying the following three
conditions

1. °
Ilxll = if and only if x = 0,
2. Ilaxll = lal '11xll for any scalar a,
3. Ilx + yll s; Ilxll + Ilyll·
For x E en, let
x = (Xl X2 ... xn )T .

Then the three most common norms are

1. I-norm:
n
Ilxll! = L IXil.
i=l
 A.5 Norms 313

2. 2-norm:
n

Ilxlb = L
i=l
I iI 2, X

3. oo-norm

The 2-norm is geometrically the most appealing, but I-norm and 00-
norms are sometimes more convenient to use. Obviously, an m by n matrix
can be regarded as a vector with m x n components, and the norms can
be defined accordingly. For example, corresponding to the 2-norm, we can
define the following norm of a matrix A E mxn c
m n
IIAIIF = LL laijl2.
i=l j=l
However, the most useful norms for matrices are the induced norms defined
in terms of the corresponding vector norm

IIAxlik
IIAllk = IIxllk9
max IIAxllk = max - -
IIxllk;6o IIxlik .

It can be shown that the three most common induced norms are

1. I-norm
m

2. 2-norm

3. OQ-norm

Clearly, IIAII2 = O"max{A). An important advantage of induced norms

is that they are compatible with matrix multiplication operation in the
following sense:
IIABII :::; IIAII . IIBII·
314 Appendix A. Matrix Facts

A.6 Matrix measure

For a square matrix A E cnxn , its matrix measure is defined as
v(A) = lim III + aAII- 1,
0-+0+ a
where II . II can be any matrix norm. Obviously, different matrix norms
result in different matrix measures. Unlike norm, the matrix measure may
assume either a positive or a negative value. Matrix measures are often
used in robust stability analysis to obtain less conservative results than if
norms are used.

A.7 Kronecker product and sum

For A E cmxn , B E C pxq , the Kronecker product of A and B, A ® B, is a
mp by nq matrix defined as

allB
a a21 B
A®B= ( .

am1B

For square matrices A E c mxm , B E cnxn , we can also define their Kro-
necker sum as

If Ai, i = 1,2, ... , m are the eigenvalues of A, and Itj' j = 1,2, ... , n are
the eigenvalues of B, then the mn eigenvalues of A®B and AE9B are Ailtj,
and Ai + Itj' i = 1,2, ... , mj j = 1,2, ... , n, respectively.

A.8 Notes
Most of the materials presented here can be found in Golub and Van Loan
[76] and Horn and Johnson [114]. The Kronecker product and sum are
covered in Graham [79].
Appendix B

Linear Matrix Inequalities and

Quadratic Integral Inequalities
B.1 Basic LMI problem

A strict LMI has the general form of

m

F(x) ~ Fo + L:>iFi > 0, (B.l)

i=l

where
x = (Xl X2 ... Xm) TERm

is a vector consisting of m variables, and Fi = Ft E jRnxn, i = 0,1,2, ... , m

are m + 1 given constant matrices. An LMI may also be nonstrict, where
">" is replaced by "~." We will only discuss strict LMI here. Notice that
the variables appear linearly on the left hand side of the inequality. The
basic LMI problem is to find whether or not there exists an X E jRm such
that (B. 1) is satisfied. The importance of LMI lies in two facts: (i) There
exist efficient computational methods to solve LMI; (ii) Many practical
problems may be formulated in the standard LMI form. In this appendix,
we will illustrate some simple cases where it is possible to transform some
common problems to the standard LMI form (B.l).
Obviously,
F(x) < °
may be written as
-F(x) > 0,
which is in the standard form (B.l). Also, a set of LMIs may be viewed as
one LM!. For example,

F(i)(X) > 0, i = 1,2

is equivalent to one LMI

F(1)(x)
( O °
F(2)(x)
316 Appendix B. LMI and Quadratic Integral Inequalities

For the sake of convenience, the variables often appear in matrix form.
For example,

(B.2)
(B.3)

the Lyapunov inequality, is an LMI with symmetric matrix P = pT E lR nxn

as its variable, which is equivalent to ~ n( n + 1) independent scalar variables
and can be written in the standard form of (B.1). With this understanding,
it is common practice to leave an LMI with matrix variables.
Another important fact worth mentioning is that the positive definiteness
of a matrix implies the positive definiteness of all its principle minors, where
a principle minor is a matrix (say, m x m) consisting of the elements in
the i 1,i2, ... ,i m th rows and columns of the original matrix (i 1,i 2, ... ,i m
distinct). For example,

implies
a > 0, d > 0, f > 0,
and

B.2 GEVP
The GEVP is to find
inf ,x, (B.4)
subject to the constraint

,XB(x) - A(x) > 0, (B.5)

B(x) > 0, (B.6)
C(x) > 0, (B.7)

where A(x), B(x), and C(x) are matrices depending on the variable x E lRm
linearly. As a special case, the constraint (B.7) may be absent. The termi-
nology is motivated by the generalized eigenvalue problem in the classical
eigenvalue theory of matrices, which is also very useful in this book. To see
the connection, consider the special case of the above problem. When A
and B are constant matrices, and the additional constraint (B. 7) is absent,
 B.2 GEVP 317

then the solution of the problem is the greatest >. to satisfy the following
generalized eigenvalue problem
A~ = >'B~, ~ # 0.
We will refer to the problem described by (B.4) to (B.7) by the abbreviation
GEVP in order to distinguish it from the classical generalized eigenvalue
problem.
A number of problems encountered in this book may be conveniently
transformed into the standard GEVP. For example, consider the problem
a = sup a, (B.8)
subject to
P(x) + aQ(x) > 0, (B.9)
P(x) > 0, (B.IO)
R(x) < 0. (B.ll)
This problem can be transformed to the standard GEVP

~a = inf,6,
subject to
,6P(x) + Q(x) > 0,
P(x) > 0,
-R(x) > 0.
Another problem that can be easily transformed to the standard GEVP is
a = sup a,
subject to
P(x) + aQ(x) < 0,
Q(x) > 0.
This is equivalent to the standard GEVP

~a = inf,6,
subject to
-,6P(x) - Q(x) > 0,
-P(x) > 0,
Q(x) > 0.
We will simply refer to these problems also as GEVP without further com-
ments.
318 Appendix B. LMI and Quadratic Integral Inequalities

B.3 Transforming nonlinear matrix inequalities to

LMI form
In many practical problems, the parameters may appear nonlinearly in
their most natural form. The fact that many of them can be transformed
into an LMI form is an important observation. Here we only mention some
important techniques.

Proposition B.1 For any nonsingular matrix A of compatible dimension,

a matrix inequality
F>O (B.12)
is satisfied if and only if
AFAT > O. (B.13)

This obvious fact has some important consequences. Especially, we may

choose A to be block version of "elementary matrices," in which case, left-
multiplying F by A is equivalent to one of the following operations on
matrix F:

1. Multiply some row block by a nonsingular matrix,

2. Left-multiply a row block by a matrix and add to another row block,

3. Interchange two row blocks.

Proposition B.I implies that we may perform a series of the above-

mentioned operations on (B.12) and the corresponding operations on the
columns to arrive at an equivalent matrix inequality. A very important
special case of the above is the following Schur complement.

Corollary B.2 For matrices A, B, C, the inequality

(B.14)

is equivalent to the following two inequalities

A > 0, (B.15)
C - BT A-I B > O. (B.16)

Proof. Using Proposition B.l, left-multiply the first row of (B.14) by

_BT A-I and add to the second row, and perform the symmetric operation
on the columns. •
In formulating practical problems to the LMI, one often needs to trans-
form inequalities of the form (B.15) and (B.16) to the form (B.14).
 B.4 S-procedure 319

B.4 S-procedure
S-procedure plays important role in robust stability theory. It can be stated
as follows.

Proposition B.3 Let Fi E IR nxn , i = 0,1,2, ... ,p. Then the following
statement is true

e Fo~ > 0 for any ~ E IR n satisfying e Fi~ ~ 0, (B.I7)

if there exist real scalars Ti ~ 0, i = 1,2, ... ,p such that
p

Fo - LTiFi > O. (B.I8)

i=1

For p = 1, these two statements are equivalent.

There are a number of variations. For example, the proposition is still

valid if ">" in both (B.I7) and (B.I8) are replaced by "~."

B.5 Elimination of matrix variables

A technique especially useful in time-delay systems is the elimination of
matrix variables in LMls. The following proposition states that a matrix
variable appearing in two off-diagonal symmetric entries in a linear matrix
inequality can be eliminated.

Proposition B.4 There exists a matrix X such that

(B.19)

if and only if

(t T ~) > 0, (B.20)

(~T ~) > O. (B.2I)

Notice that (B.20) and (B.2I) are obtained by deleting the row and
column containing X in two different ways.
Proof. Necessity is obvious since the left hand side of both (B.20)
and (B.2I) are principal minors of the left hand side of (B.I9). For suf-
ficiency, left-multiply second row of (B.I9) by - V T R- 1 and add to the
320 Appendix B. LMI and Quadratic Integral Inequalities

third row, with symmetric operation for the columns. Equivalently, left-
multiply (B.19) by

( o~ ~_VTR- ~),
I I

and right-multiply by its transpose, to show that (B.19) is equivalent to

But the above is clearly satisfied for

x = QR-IV (B.22)
if (B.20) and (B.21) are satisfied in view of Schur complement . •
In some cases, matrix variables appearing in multiple entries can also be
eliminated. Here is an example.

n
Corollary B.5 There exists a matrix X such that

( P+XEo+ElxT Q+XE
(Q+xEf R > 0, (B.23)
XT VT
if and only if

( P+E,{;SEo
MT ~) > 0, (B.24)

(~ ~) > 0, (B.25)

where
M Q - Eifv T + Eif SEl,
N R- VE-ETVT +ETSE.

Proof. Left-multiply the third row of the left hand side of (B.23) by
- (Eo E) T and add to the second row, with symmetric operation on
the columns, then use Proposition BA, to realize that (B.23) is equivalent
to (B.24) and
R- VE _ETVT +ETSE V - ETS )
( (V - ET Sf S > O. (B.26)

Since (B.25) can be obtained from (B.26) by left-multiplying the second

row by ET and adding to the first row with symmetric operation on the
columns, (B.26) is equivalent to (B.25) . •
 B.5 Elimination of matrix variables 321

It is also possible to eliminate a common matrix variable in a diagonal

entry of two matrix inequalities, as is shown in the following proposition.

Proposition B.6 There exists a symmetric matrix X such that

( P1 -LXL T Ql
Qr Rl ) > 0, (B.27)

( P2 +X Q2
QrR2 ) >
° (B.28)

if and only if
( PI + LP,LT Ql LQ2

°°
Qr Rl ) >0 (B.29)
QrLT R2

Proof. Using the Schur complement, it is easy to see that (B.29) is

equivalent to
(B.30)
(B.31)

and
.bo = PI + LP2L T - QIRIIQf - LQ2R2 1 Qf LT > 0. (B.32)
Also, (B.27) and (B.28) are equivalent to (B.30), (B.31), and
.bo 1 PI - LXLT - QIRIIQf > 0, (B.33)
.bo2 = P2 + X - Q2R;;lQf > 0. (B.34)
Therefore, we only need to show that, given (B.30) and (B.31), there exists
an X satisfying (B.33) and (B.34) if and only if (B.32) is satisfied.
The existence of X to satisfy (B.33) and (B.34) implies (B.32) since
.bo = .bo 1 + L.bo 2LT. On the other hand, if (B.32) is satisfied, let
X = Q2R21Qr - P2 +d
for some sufficiently small E °
> results in
.bo2 = cI > 0,
and
.bo 1 = .bo - L.bo 2L T = .bo - ELLT > 0,
thus (B.33) and (B.34) are satisfied. •
The above results (Proposition BA, Corollary B.5, and Proposition B.6)
can be extended to the case where the constant matrices are replaced by
continuous matrix functions depending on variables within a compact set.
For example, Proposition BA can be extended to the following.
322 Appendix B. LMI and Quadratic Integral Inequalities

Corollary B.7 For given continuous matrix functions P( 0), Q( 0), R( 0),
V(O), and S(O), with variable 0 varying with a compact set 0 C lR.k, there
exists a continuous matrix function X(O) such that

P(O) Q(O) X(O))

( QT(O) R(O) V(O) >OforallOEO, (B.35)
XT (0) V T (0) S( 0)
if and only if
P(O) Q(O) )
(
QT(O) R(O)
> 0, (B.36)

R(O) V(O) )
(
VT(O) S(O)
> 0, (B.37)

for all 0 E o.
Proof. Based on Proposition B.4 and its proof, it is sufficient to prove
that X as expressed in (B.22) is continuous. From (B.36), R > 0 for all
o E O. This implies that R > cI for some constant c: > 0 since R is
continuous and 0 is compact. Therefore, R- 1 exists and is continuous and
bounded. Since Q and V are also continuous, X is continuous. •

B.6 Quadratic integral inequalities

The following integral inequality is known as the Jensen Inequality, which
plays an important role in the stability problem of time-delay systems.

Proposition B.8 (Jensen Inequality) For any constant matrix M E

lR. mxm , M = MT > 0, scalar, > 0, vector function w : [0,,] _ lR. m such
that the integrations concerned are well defined, then

,l' wT((3)Mw((3)d(3 ~ (1' W((3)d(3) T M (1' W((3)d(3) . (B.38)

Proof. It is easy to see, using the Schur complement, that

for any 0 :S (3 :S ,. Integration of the above inequality from 0 to , yields

( f~ wT ((3)M w((3)d(3 fo' wT ((3)d(3 ) > 0

fo w((3)d(3 ,M- 1 - .

Use the Shur complement in the above to reach (B.38) . •

 B.6 Quadratic integral inequalities 323

Corollary B.9 For any constant matrix M E jRmxm, M = MT > 0 and

vector functions W,Wl,W2 : [0, 1] ~ jRm, such that the integmtions in the
following are well defined, then

11 [(1- a)wf(a)Mwl(a) + awf(a)Mw2(a)]da

> 11 (1 0
WI (f3)df3 + i1 W2(f3)df3 ) T

M (1 0
WI (f3)df3 +i 1 w2(f3)df3 ) da, (B.39)

11 (1 - a)wT(a)Mw(a)da

> 11 (1 0
W(f3)d f3 ) T ±M (1 0
W(f3)d f3 ) da (BAD)

> 11 (1 0
W(f3)df3 ) T M (1 0
w(f3)df3 ) da, (BA1)

11 awT(a)Mw(a)da
> (i
11 1w(f3)df3) T 1 ~ aM (i1 w(f3)df3) da (BA2)

> 11 (i1W(f3)df3 ) T M (i 1w(f3)df3) da. (BA3)

Proof. Use Proposition B.8 for the case 'Y = 1, and

w(f3) = { WI (13), 0:::; 13:::; a,

W2(f3), a <13:::; 1,
to obtain

1 0
wf(f3)Mwl(f3)df3 + i 1
wf(f3)Mw2(f3)df3

> (1 WI (f3)df3 + i w2(f3)df3) T M (1

0 1 0
WI (f3)df3 + i 1
w2(f3)df3 ) da.

Integrate the above from a = 0 to a = 1, and exchange the order of

integration between a and 13 in both terms of the left hand side to arrive
324 Appendix B. LMI and Quadratic Integral Inequalities

at inequality (B.39). To see (BAO),

11 (1- a)wT(a)Mw(a)da

11 11 d(3wT(a)Mw(a)da

11 1f3 wT(a)Mw(a)dad(3
> l (J.' W(a)da) T ~M (J.' W(a)da) d~,
which is (BAO). In the above, the equality in the second step is the result
of exchanging the order of integration, and the inequality in the last step
results from the Jensen Inequality (B.38). Considering the fact that l/a 2': 1
in the integration interval, (BA1) is obvious. The proof of (BA2) and (BA3)
are similar. •

B.7 Notes
The importance of matrix inequalities, especially LMIs, in systems theory
and control has long been recognized since Lyapunov theory was published
around 1890 [185]. Indeed, solutions of some matrix inequalities have ap-
peared as early as early 1960s by Yakubovich [287]-[289], see also the work
by Horisberger and Belanger [115]. Largely because of the unavailability of
efficient numerical algorithms for the general form of LMI, most of the ear-
lier works on problems related to LMIs were reformulated in forms such as
Lyapunov equations and Algebraic Riccati Equations (AREs). The realiza-
tion that LMI is a convex optimization problem [20], and the development
of the efficient interior point method [203] have spurred tremendous inter-
est among researchers in control systems theory to formulate many control
problems in LMI form. See the book by Boyd et al. [21] for more concepts
on LMI, GEVP, and applications in system and control theory. For more
recent progress, and generalization to semidefinite programming, see [266]
and [75]. Commercial software of an LMI solver in MAT LAB is available
[72].
For a more comprehensive and systematic coverage of the materials in
Sections B.l to B.3, see the book by Boyd et al. [21].
The elimination of variables in LMI in Section B.5 was discussed in [87]
and [89]. Some other cases of variable elimination are discussed in [21].
 B.7 Notes 325

The quadratic integral inequalities discussed in Section B.6 were dis-

cussed in [89J in the time-delay stability context. The Jensen Inequality is
covered in many books on probability theory (see, for example, [245]).
Bibliography
A selected list of literature entries are included in this bibliography. Most
entries are cited in the text. Some additional entries are also included for
additional readings. Additional literature can be found in a number of
books, such as [204, 105, 154J.

[lJ C. T. Abdallah, P. Dorato, J. Benftez-Read, and R. Byrne. Delayed

positive feedback can stabilize oscillatory systems. Proceedings of
1993 American Control Conference, San Francisco, CA, 3106-3107,
1993.

[2J H. C. Ansell. On certain two-variable generalizations of circuit the-

ory, with applications to network of transmission lines and lumped
reactances. IEEE Trans. Circuit Theory, 11:214-223, 1964.

[3J L. Arnold, H. Crauel, and V. Wihstutz. Stabilization of linear systems

by noise. SIAM J. Control Optim., 21(3):451-461, 1983.

[4J K. Astrom and B. Wittenmark. Adaptive Control, 2nd Ed. Addison-

Wesley, Reading, MA, 1995.

[5J S. P. Banks. A note on non-linear observers. Int. J. Control,

34(1):185-190, 1981.

[6] B. R. Barmish and Z. Shi. Robust stability and perturbed systems

with time delays. Automatica, 25(3):371-381, 1989.

[7J S. Barnett. Polynomials and Linear Control Systems. Marcel Dekker,

New York, 1983.

[8] D. R. Bell and S. E. A. Mohammed. On the solution of stochastic

ordinary differential equations via small delays. Stochastics and Sto-
chastic Reports, 28:293-299, 1989.

[9J R. Bellman and K. L. Cooke. Differential-Difference Equations. Aca-

demic Press, New York, 1963.

[10J R. Bellman, J. Bentsman, and S. M. Meerkov. Vibrational control

of nonlinear systems: vibrational stabilizability. IEEE Trans. Auto.
Control, AC-31 (8):710-716, 1986.
328 Bibliography

[11J F. Blanchini, R. Lo Cigno, and R. Tempo. A robust feedback strat-

egy for rate control in ATM networks. Proceedings of 1998 American
Control Conference, Philadelphia, PA, 2852-2857, 1998.
[12J P.-A. Bliman. Extension of Popov absolute stability criterion to non-
autonomous systems with delays. Int. J. Control, 73(15):1349-1361,
2000.
[13] P.-A. Bliman. LMI characterization of strong delay-independent sta-
bility of linear delay systems via quadratic Lyapunov-Krasovskii func-
tionals. Systems and Control Letters, 43(4):263-274, 200l.
[14J L. Blum, F. Cucker, M. Shub, and S. Smale. Complexity and Real
Computation. Springer-Verlag, Berlin, Germany, 1997.
[15J F. G. Boese. Stability conditions for the general linear difference-
differential equation with constant coefficients and one constant de-
lay. J. Math. Anal. Appl., 140:136--176, 1989.
[16] F. G. Boese. Stability criteria for second-order dynamical systems
involving several time delays. SIAM J. Math. Anal., 26:1306--1330,
1995.
[17] N. K. Bose. Applied Multidimensional Systems Theory. Van Nos-
trand, New York, 1982.
[18] E.-K. Boukas and Z. K. Liu. Deterministic and Stochastic Time-
Delayed Systems. Birkhauser, Boston, 200l.
[19] H. Bourles. a-stability of systems governed by a functional differential
equation-extension of results concerning linear delay systems. Int.
J. Control, 45(6):2233-2234, 1987.
[20] S. Boyd and C. Barratt. Linear Controller Design: Limits of Perfor-
mance. Prentice-Hall, Englewood Cliffs, NJ, 1991.
[21] S. Boyd, L. EI Ghaoui, E. Feron, and V. Balakrishnan. Linear Ma-
trix Inequalities in System and Control Theory. SIAM, Philadelphia,
1994.
[22J S. Boyd and Q. Yang. Structured and simultaneous Lyapunov func-
tions for system stability problems. Int. J. Control, 49:2215-2240,
1989.
[23] R. P. Braatz, P. M. Young, J. C. Doyle, and M. Morari. Computa-
tional complexity of j.L calculation. IEEE Trans. Auto. Control, AC-
39(5):1000-1002, May 1994.
[24] R. Brayton. Nonlinear oscillations in a distributed network. Quart.
Appl. Math., 24:289-301, 1976.
 Bibliography 329

[25J S. D. Brierley, J. N. Chiasson, E. B. Lee, and S. H. Zak. On stability

independent of delay for linear systems. IEEE 7rans. Auto. Control,
AC-27(1):252-254, Feb. 1982.

[26J M. Briggs and R. Vinter. Linear filtering for time-delay systems. IMA
J. Math. Control Info., 6:167-178, 1989.

[27J A. E. Bryson, Jr. and Y. C. Ho. Applied Optimal Control. Hemisphere

Publishing, Washington, DC, 1975.

[28J Y.-Y. Cao, Y.-X. Sun, and C. Cheng. Delay-dependent robust stabi-
lization of uncertain systems with multiple state delays. IEEE 7rans.
Auto. Control, 43(11):1608-1612, 1998.

[29J Y.-Y. Cao and Y.-X. Sun. Robust stabilization of uncertain sys-
tems with time-varying multistate delays. IEEE 7rans. Auto. Con-
trol, 43(10):1484-1488, 1998.

[30J W. B. Castelan and E. F. Infante. On a functional equation arising in

the stability theory of difference-differential equations. Quart. Applied
Math. 35:311-319,1977.

[31J V.-S. Chellaboina. Robust stability and performance for linear and
nonlinear uncertain systems with structured uncertainty. Ph.D. Dis-
sertation, School of Aerospace Engineering, Georgia Institute of Tech-
nology, 1996.

[32J J. Chen. On computing the maximal delay intervals for stability of

linear delay systems. IEEE 7rans. Auto. Control, 40:1087-1093,1995.
[33J J. Chen and G. Gu. Control-Oriented System Identification: An 'Hoc
Approach. Wiley Interscience, New York, 2000.

[34J J. Chen, G. Gu, and C. N. Nett. A new method for computing delay
margins for stability of linear delay systems. Systems and Control
Letters, 26(2):107-117, Sept. 1995.

[35J J. Chen and H. A. Latchman. Frequency sweeping tests for stabil-

ity independent of delay. IEEE 7rans. Auto. Control, 40:1640-1645,
1995.

[36J J. Chen and A. L. Tits. Robust control analysis. In J. G. Webster

(Ed.), Encyclopedia of Electrical and Electronics Engineering. Wiley,
New York, 18:602-616, 1999.

[37J J. Chen, D. Xu, and B. Shafai. On sufficient conditions for stability

independent of delay. IEEE 7rans. Auto. Control, 4O(9):1675-1680,
Sept. 1995.
330 Bibliography

[38] J. N. Chiasson, S. D. Brierley, and E. B. Lee. A simplified derivation

of the Zeheb-Walach 2-D stability test with application to time-delay
systems. IEEE Trans. Auto. Control, AC-30(4):411-414, April 1985.
[39] J. N. Chiasson, S. D. Brierley, and E. B. Lee. Corrections to "A
simplified derivation of the Zeheb-Walach 2-D stability test with ap-
plication to time-delay systems." IEEE Trans. Auto. Control, AC-
31(1):91-92, Jan. 1986.
[40] J. N. Chiasson. A method for computing the interval of delay values
for which a differential-delay system is stable. IEEE Trans. Auto.
Control, AC-33(12):1176-1178, Dec. 1988.
[41] J. A. Cook and B. K. Powell. Modeling of an internal combustion
engine for control analysis. IEEE Control Systems Magazine, 20-25,
August 1988.
[42] K. 1. Cooke and J. M. Ferreira. Stability conditions for linear re-
tarded functional differential equations. J. Math. Anal. Appl., 96:480-
504, 1983.
[43] K. L. Cooke and Z. Grossman. Discrete delay, distributed delay and
stability switches. J. Math. Anal. Appl., 86:592-627, 1982.

[44] J. J. Craig, P. Hsu, and S. S. Sastry. Adaptive control of mechanical

manipulators. Int. J. Robotics Research, 6(2):16-28, 1987.
[45] L. Crocco. Aspects of combustion stability in liquid propellant rocket
motors, part I: fundamentals-low frequency instability with mono-
propellants. J. Amer. Rocket Soc., 21:163-178, 1951.
[46] R. F. Curtain and A. J. Pritchard. Infinite-dimensional linear sys-
tems theory. Lecture Notes in Control and Info. Sciences, 8, Springer-
Verlag, Berlin, 1978.
[47] R. F. Curtain and H. Zwart. An Introduction to Infinite-dimensional
Linear Systems Theory. Texts in Applied Mathematics, 21, Springer-
Verlag, New York, 1995.
[48] R. Datko. An algorithm for computing Liapunov functionals for some
differential difference equations. In L. Weiss (Ed.), Ordinary Differ-
ential Equations. Academic Press, New York, 387-398, 1972.
[49] R. Datko. A procedure for determination of the exponential stabil-
ity of certain differential-difference equations. Quart. Appl. Math.,
36:279--292, 1978.
[50] R. Datko. Lyapunov functionals for certain linear delay-differential
equations in a Hilbert space. J. Math. Anal. Appl., 76:37-57, 1980.
 Bibliography 331

[51J R. Datko. Not all feedback stabilized hyperbolic systems are robust
with respect to small time delays in their feedbacks. SIAM J. Control
Optim., 26:697-713, 1988.

[52] M. D. Davis and E. J. Weyuker, Computability, Complexity, and Lan-

guages. Academic Press, Orlando, FL, 1983.

[53J C. A. Desoer and M. Y. Wu. Stability of a nonlinear time-invariant

feedback system under almost constant inputs. Automatica, 5:231-
233, 1969.

[54J C. A. Desoer and M. Vidyasagar. Feedback Systems: Input-Output

Properties. Academic Press, New York, 1975.

[55J O. Diekmann, A. A. von Gils, S. M. Verduyn Lunel, and H.-O.

Walther. Delay Equations, Functional-, Complex- and Nonlinear
Analysis. Appled Mathematics Sciences Series, 110, Springer-Verlag,
New York, 1995.

[56J R. C. Dorf and A. Kusiak. Handbook of Manufacturing and Automa-

tion. Wiley, New York, 1994.

[57J J. C. Doyle. Analysis of feedback systems with structured uncertain-

ties. Proc. lEE, Part D, 129(6):242-250, Nov. 1982.

[58J J. C. Doyle, J. Wall, and G. Stein. Performance and robustness analy-

sis for structured uncertainty. IEEE Conference on Decision and
Control, 629--636, 1982.

[59J R. D. Driver. Existence and stability of a delay-differential systems.

Arch. Rational Mech. Anal., 10:401-426, 1962.

[60] L. Dugard and E. Verriest (Eds.). Stability and Control of Time-

Delay Systems. Lecture Notes in Control and Information Sciences,
228, Springer-Verlag, New York, 1997.

[61J K. Engelborghs, M. Dambrine, and D. Roose. Limitation of a class of

stabilization methods for delay systems. IEEE Trans. Auto. Control,
46(2):336-339, 2001.

[62] M. K. H. Fan, A. L. Tits, and J. C. Doyle. Robustness in the presence

of mixed parametric uncertainty and unmodeled dynamics. IEEE
Trans. Auto. Control, AC-36(1):25-38, Jan. 1991.

[63] D. Filipovic and N. Olgac. Delayed resonator with speed feedback

including dual frequency-theory and experiment. Proceedings of 36th
IEEE Conference on Decision and Control, San Diego, CA, 2535-
2540, December 1997.
332 Bibliography

[64] P. Florchinger and E.I. Verriest. Stabilization of nonlinear stochastic

systems with delay feedback. Proceedings of 32nd IEEE Conference
Decision and Control, San Antonio, TX, 859-860, December 1993.

[65] C. Foias, H. Ozbay, and A. Tannenbaum. Robust Control of Infinite

Dimensional Systems: Frequency Domain Methods. Lecture Notes in
Control and Information Sciences, 209, Springer-Verlag, New York,
1995.

[66] E. Fridman and U. Shaked. Hoc-norm and invariant manifolds of

systems with state delays. Systems and Control Letters, 36(2):157-
165, 1999.

[67] E. Fridman and U. Shaked. Finite horizon Hoc state-feedback con-

trol of continuous-time systems with state delays. IEEE Trans. Auto.
Control, 45(12):2406-2411, 2000.

[68] M. Fu, H. Li, and S.-I. Niculescu. Robust stability and stabilization
of time-delay systems via integral quadratic constraint approach. In
L. Dugard and E. I. Verriest (Eds.), Stability and Control of Time-
Delay Systems. Lecture Notes in Control and Information Sciences,
228, Springer-Verlag, New York, 101-116, 1997.

[69] M. Fu. The real J.1, is hardly approximable. IEEE Trans. Auto. Con-
trol, AC-42:1286-1288, 1997.

[70] M. Fu, A. W. Olbrot, and M. P. Polis. Robust stability for time-

delay systems: the edge theorem and graphical test. IEEE Trans.
Auto. Control, AC-34:813-820, 1989.

[71] M. Fu, A. W. Olbrot, and M. P. Polis. Edge theorem and graphical

test for robust stability of neutral time-delay systems. Automatica,
27:739-742, 1991.

[72] P. Gahinet, A. Nemirovski, A. Laub, and M. Chilali. LMI Control

Toolbox for Use with MATLAB, Mathworks, Natick, MA, 1995.

[73] M. Garey and D. Johnson. Computers and Intmctability: A Guide

to the Theory of NP-Completeness. W. H. Freeman, San Francisco,
1979.

[74] J.-H. Ge, P. M. Frank, and C.-F. Lin. Hoc control via output feedback
for state delayed systems. Int. J. Control, 64(1):1-8, 1996.

[75] L. El Ghaoui and S.-I. Niculescu (Eds.). Advances in Linear Matrix

Inequality Methods in Control. Advances in Control and Design, 2,
SIAM, Philadelphia, 1999.
 Bibliography 333

[76J G. H. Golub and C. F. van Loan. Matrix Computations. Johns Hop-

kins Univ. Press, Baltimore, MD, 1983.

[77J K. Gopalsamy. Stability and Oscillations in Delay Differential Equa-

tions of Population Dynamics. Mathematics and Its Applications Se-
ries, 74, Kluwer Academic Publishers, Dordrecht, 1992.

[78J H. Gorecki, S. Fuksa, P. Grabowski, and A. Korytowski. Analysis

and Synthesis of Time-Delay Systems. Polish Scientific Publishers,
Warszawa, 1989.

[79J A. Graham. Kronecker Products and Matrix Calculus with Applica-

tions. Ellis Horwood, Ltd., Chichester, UK, 1981.
[80J G. Gu and E. B. Lee. Stability testing of time delay systems. Auto-
matica, 25(5):777-780, 1989.
[81J K. Gu. Hoo control of systems under norm bounded uncertainties in
all system matrices. IEEE Trans. Auto. Control, 39(1):127-131, 1994.

[82J K. Gu. Designing stabilizing control of uncertain systems by quasi-

convex optimization. IEEE Trans. Auto. Control, 39(6):1320-1322,
1994.
[83J K. Gu. Discretized LMI set in the stability problem of linear uncertain
time-delay systems. Int. J. Control, 68:923-934, 1997.

[84J K. Gu. A generalized discretization scheme of Lyapunov functional

in the stability problem of linear uncertain time-delay systems. Int.
J. Robust and Nonlinear Control, 9:1-14, 1998.
[85J K. Gu. Stability of linear time-delay systems with block-diagonal
uncertainty. Proceedings of 1998 American Control Conference,
Philadelphia, PA, 1943-1947, June 1998.
[86J K. Gu. Discretized Lyapunov functional for uncertain systems with
multiple time-delay. Int. J. Control, 72(16):1436-1445, 1999.

[87J K. Gu. Partial solution of LMI in stability problem of time-delay

systems. Proceedings of 38th IEEE Conference on Decision and
Control, Phoenix, AZ, 227-232, December 1999.
[88] K. Gu. An integral inequality in the stability problem of time-delay
systems. Proceedings of 39th IEEE Conference on Decision and Con-
trol, Sidney, Australia, 2805--2810, December 2000.

[89] K. Gu. A further refinement of discretized Lyapunov functional

method for the stability of time-delay systems. Int. J. Control,
74{1O):967-976, 2001.
334 Bibliography

[90] K. Gu. Refined discretized Lyapunov functional method for systems

with multiple delays. 40th Conference on Decision and Control, Or-
lando, FL, December 2001.

[91] K. Gu. An improved stability criterion for systems with distributed

delays. 3rd IFAC Workshop on Time Delay Systems, Santa Fe, NM,
December 8-10, 2001.

[92] K. Gu. Simple improved delay-dependent stability criteria using im-

plicit model transformation. Technical Report, Department of Me-
chanical and Industrial Engineering, Southern Illinois University at
Edwardsville, 2001.

[93] K. Gu and Q.-L. Han. Discretized Lyapunov functional for linear

uncertain systems with time-varying delay. 2000 American Control
Conference, Chicago, IL, 2000.

[94] K. Gu and Q.-L. Han. A revisit of some delay-dependent stability cri-

teria for uncertain time-delay systems. 3rd IFAC Workshop on Time
Delay Systems, Santa Fe, NM, December 8-10, 2001.

[95] K. Gu, Q.-L. Han, A. C. J. Luo, and S.-I. Niculescu. Discretized

Lyapunov functional for systems with distributed delay and piecewise
constant coefficients. Int. J. Control, 74(7):737-744, 2001.

[96] K. Gu, A. C. J. Luo, and S.-I. Niculescu. Discretized Lyapunov func-

tional for systems with distributed delay. 1999 European Control
Conference, Germany.
[97] K. Gu and S.-I. Niculescu. Additional dynamics in transformed time-
delay systems. IEEE Trans. Auto. Control, 45(3):572-575, 2000.

[98] K. Gu and S.-I. Niculescu. Further remarks on additional dynamics in

various model transformations of linear delay systems. IEEE Trans.
Auto. Control, 46(3):497-500, 2001.
[99] W. M. Haddad and D. Bernstein. Explicit construction of quadratic
Lyapunov functions for the small gain and positivity theorems. Pro-
ceedings 30th IEEE Conference Decision and Control, Brighton, Eng-
land, December 1991.

[100] K. P. Hadeler. Delay equations in biology. In H.-O. Peitgen and H.-O.

Walther (Eds.), Functional Differential Equations and Approximation
of Fixed Points. Lecture Notes Mathematics, 730, Springer-Verlag,
Berlin, 136-157, 1979.

[101] A. Halanay. Differential Equations: Stability, Oscillations, Time-lags.

Academic Press, New York, 1966.
 Bibliography 335

[102] J. K. Hale, E. F. Infante, and F. S. P. Tsen. Stability in linear delay

equations. J. Math. Anal. Appl., 105:533-555, 1985.
[103J J. K. Hale. Effects of delays on stability and control, Report CDSN97-
270, Center for Dynamical Systems and Nonlinear Studies, Georgia
Institute of Technology, 1997.
[104] J. K. Hale and W. Huang. Global geometry of the stable regions for
two delay differential equations. J. Math. Anal. Appl., 178:344-362,
1993.
[105] J. K. Hale and S. M. Verduyn Lunel. Introduction to Functional Dif-
ferential Equations. Springer-Verlag, New York, 1993.
[106J J. K. Hale, L. T. Magalhaes, and W. M. Oliva. An Introduction to
Infinite Dynamical Systems-Geometric Theory, Applied Mathemat-
ical Sciences, 47, Springer-Verlag, New York, 1985.
[107J Q.-L. Han and K. Gu. On robust stability of time-delay systems with
norm-bounded uncertainty. 2001 American Control Conference, Ar-
lington, VA, 4644-4649, June 25-27, 2001.
[108] Q.-L. Han and K. Gu. On robust stability of time-delay systems with
block-diagonal uncertainty. 40th Conference on Decision and Control,
Orlando, FL, December 2001.
[109] Q.-L. Han and K. Gu. Robust stability of time-delay systems with
passive uncertainty. 3rd IFAC Workshop on Time Delay Systems,
Santa Fe, NM, December 2001.
[110J D. Hertz, E. J. Jury, and E. Zeheb. Stability independent and de-
pendent of delay for delay differential systems. J. Franklin Inst.,
318(3):143-150, Sept. 1984.
[111J D. Hertz, E. J. Jury, and E. Zeheb. Simplified analytic stability
test for systems with commensurate time delays. Proc. lEE, Part
D, 131(1):52-56, Jan. 1984.
[112J A. Hmamed. On the stability of time-delay systems: new results. Int.
J. Control, 43(1):321-324, 1986.

[113] A. Hmamed. Stability conditions of delay-differential systems. Int. J.

Control, 43(2):455-463, 1986.
[114] R. A. Horn and C. A. Johnson. Matrix Analysis. Cambridge Univer-
sity Press, Cambridge, UK, 1985.
[115] H. P. Horisberger and P. R. Belanger. Regulators for linear, time
invariant plants with uncertain parameters. IEEE Trans. Auto. Con-
trol, AC-21:705-708, 1976.
336 Bibliography

[116) C. Hou and J. Qian. On an estimate of the decay rate for applications
of Razumikhin-type theorem. IEEE Trans. Auto. Control, 43(7):958-
960, 1998.

[117] W. Huang. Generalization of Liapunov's theorem in a linear delay

system. J. Math. Anal. Appl., 142:83-94, 1989.

[118] Y.-P. Huang and K. Zhou. Robust stability of uncertain time-delay

systems. IEEE Trans. Auto. Control, 45(11):2169-2173, 2000.

[119] Y.-P. Huang and K. Zhou. On the robustness of uncertain time-delay

systems with structured uncertainties. Systems and Control Letters,
41:367-376, 2000.

[120] C. Hwang, Y. H. Chuang, and Y. P. Shih. Numeric computation

of quadratic cost functional for linear systems with multiple delays.
Control and Computers, 22(3):90-95, 1995.

[121] E. F. Infante. Some results on the Lyapunov stability of functional

equations. In K. B. Hannsgen, T. L. Herdman, H. W. Stech, and
R. L. Wheeler (Eds), Volterra and Functional Differential Equations.
Lecture Notes in Pure and Applied Mathematics 81, Marcel Dekker,
New York, 51-60, 1982.

[122] E. F. Infante and W. V. Castelan. A Lyapunov functional for a matrix

difference-differential equation. J. Diff. Equations, 29:439-451, 1978.

[123) A. F. Ivanov and A. Sharkovsky. Oscillations in singularly perturbed

delay equations. Dyn. Rep., Expo. Dyn. Syst., New Ser., 1:164-224,
1992.

[124) A. F. Ivanov and E. I. Verriest. Robust stability of delay-difference

equations. In U. Helmke, R. Mennicken, and J. Saurer (Eds.), Systems
and Networks: Mathematical Theory and Applications, University of
Regensburg, 725-726, 1994.

[125] N. Jalili and N. Olgac. Stability analysis of multiple delayed resonator

vibration absorbers. Dynamics, Acoustics and Simulations, ASME
Winter Annual Meeting, DE-98:211-217, 1998.

[126] E. A. Jonckheere. Algebraic and Differential Topology of Robust Sta-

bility, Oxford University Press, Oxford, 1997.

[127] P. T. Kabamba, S. M. Meerkov and E. K. Poh. Closed loop vibra-

tional control: state and output feedback stabilizability. Proceedings
30-th IEEE Conference on Decision and Control, Brighton, England,
2375-2376, Dec. 1991.
 Bibliography 337

[128] K. Kadiman and D. Williamson. Optimal finite word length linear

quadratic regulation. IEEE Trans. Auto. Control, 34(12):1218-1228,
1989.
[129] T. Kailath. Linear Systems. Prentice-Hall, Englewood Cliffs, NJ,
1980.
[130] E. W. Kamen. Lectures on Algebraic System Theory: Linear Systems
over Rings, NASA Contractor Report 3016, 1978.
[131] E. W. Kamen. On the relationship between zero criteria for two-
variable polynomials and asymptotic stability of delay differential
equations. IEEE Trans. Auto. Control, AC-25(5):983-984, Oct. 1980.
[132] E. W. Kamen. Linear systems with commensurate time delays: stabil-
ity and stabilization independent of delay. IEEE Trans. Auto. Con-
trol, AC-27(2):367-375, 1982.
[133] E. W. Kamen. Correction to "Linear systems with commensurate
time delays: stability and stabilization independent of delay." IEEE
Trans. Auto. Control, AC-28(2):248-249, Feb. 1983.

[134] E. W. Kamen, P. P. Khargonekar, and A. Tannenbaum. Stabilization

of time-delay systems using finite-dimensional compensators. IEEE
Trans. Auto. Control, AC-30(1):75-78, 1985.

[135] M. Kao and J. J. Moskwa. Turbocharged diesel engine modeling for

nonlinear engine control and state estimation. ASME J. Dynamic
Systems, Measurement and Control, 117(1):20-30, 1995.
[136] D. C. Karnopp, D. L. Margolis, and R. C. Rosenberg. System
Dynamics-A Unified Approach, 2nd Ed. Wiley, New York, 1990.
[137] V. 1. Kharitonov. Robust stability analysis of time delay systems: a
survey. Annual Reviews in Control, 23:185-196, 1999.
[138] V. L. Kharitonov and A. P. Zhabko. Robust stability of time-delay
systems. IEEE Trans. Auto. Control, AC-39:2388-2397, 1994.
[139] V. L. Kharitonov. Asymptotic stability of an equilibrium position of
a family of system of linear differential equations. Differential Equa-
tions (in Russian), 14:2086-2088, 1978.

[140] V. L. Kharitonov and D. Melchor-Aguilar. On delay-dependent sta-

bility conditions. Systems and Control Letters, 40:71-76, 2000.
[1411 V. 1. Kharitonov and J. A. Torres Munoz. Robust stability of mul-
tivariate polynomials-Part 1: Small coefficient perturbations. Mul-
tidimensional Systems and Signal Processing, 10:7-20, 1999.
338 Bibliography

[142] V. L. Kharitonov, J. A. Torres Munoz, and M. 1. Ramirez Sosa. Ro-

bust stability of multivariate polynomials-Part 2: Polytopic coef-
ficient variations. Multidimensional Systems and Signal Processing,
10:21-32, 1999.

[143] V. L. Kharitonov, M. 1. Ramirez Sosa, and J. A. Torres Munoz. Ro-

bust stability of multivariate polynomials-Part 3: Frequency domain
approach. Multidimensional Systems and Signal Processing, 11:213-
232,2000.

[144] V. L. Kharitonov and D. Melchor-Aguilar. Additional dynamics for

time varying delay systems. 40th IEEE Conference on Decision and
Control, Orlando, FL, December 4-7, 2001.

[145] V. L. Kharitonov and A. P. Zhabko. Lyapunov-Krasovskii approach

to the robust stability analysis of time-delay systems. To appear in
A utomatica.

[146] J.-H. Kim. Delay and its time-derivative dependent robust stability
of time-delayed linear systems with uncertainty. IEEE Trans. Auto.
Control, 46(5):789-792, 2001.

[147] J.-H. Kim. Robust stability of linear systems with delayed perturba-
tions. IEEE Trans. Auto. Control, 41(12):1820-1822, 1996.

[148] J. Kogan. Robust Stability and Convexity: An Introduction. Lecture

notes in Control and Information Sciences, 201, Springer-Verlag, New
York, 1995.

[149] A. Kojima, K. Uchida, and E. Shimemura. Robust stabilization of un-

certain time delay systems via combined internal-external approach.
IEEE Trans. Auto. Control, 38:373-378, 1993.

[150] H. Kokame, K. Hirata, K. Konishi, and T. Mori. State difference

feedback for stabilizing uncertain steady states of non-linear systems.
Int. J. Control, 74(6):537-546, 2001.

[151] H. Kokame, H. Kobayashi, and T. Mori. Robust Hex:> performance for

linear delay-differential systems with time-varying uncertainty. IEEE
Trans. Auto. Control, 43(2):223-226, 1998.

[152] V. B. Kolmanovskii and V. R. Nosov. Stability of Functional Dif-

ferential Equations. Mathematics in Science and Engineering, 180,
Academic Press, New York, 1986.

[153] V. Kolmanovskii and A. Myshkis. Applied Theory of Functional Dif-

ferential Equations. Kluwer, Dordrecht, the Netherlands, 1992.
 Bibliography 339

[154] V. Kolmanovskii and A. Myshkis. Introduction to the Theory and Ap-

plications of Functional Differential Equations. Kluwer, Dordrecht,
the Netherlands, 1999.

[155] V. B. Kolmanovskii, S.-I. Niculescu, and K. Gu. Delay effects on

stability: a survey. Proceedings 38th IEEE Conference on Decision
and Control, Phoenix, AZ, 1993-1998, Dec. 7-10, 1999.

[156] V. B. Kolmanovskii and L. E. Shaikhet. Control of systems with af-

tereffect. Translations of Mathematical Monographs, 157, American
Mathamatical Society, Providence, 1996.

[157] V. B. Kolmanovskii and J.-P. Richard. Stability of some systems with

distributed delays. European Journal of Automation, special issue on
analysis and control of time-delay systems, 31:971-982, 1997.

[158] V. B. Kolmanovskii and J.-P. Richard. Stability of some linear sys-

tems with delays. IEEE Trans. Auto. Control, 44(5):984-989, 1999.

[159] V. B. Kolmanovskii, S.-I. Niculescu, and J.-P. Richard. On the

Liapunov-Krasovskii functionals for stability analysis of linear delay
systems. Int. J. Control, 72(4):374-384, 1999.

[160] N. N. Krasovskii. Stability of Motion [Russian], Moscow, 1959; [Eng-

lish translation] Stanford University Press, Stanford, CA, 1963.

[161] Y. Kuang. Delay Differential Equations with Applications in Popula-

tion Dynamics, Academic Press, Boston, 1993.

[162] W. H. Kwon, G. W. Lee, and S. W. Kim. Performance improvement

using time delays in multivariable controller design. Int. J. Control,
52:1455-1473, 1990.

[163] W. H. Kwon and A. E. Pearson. A note on feedback stabilization of a

difference system. IEEE Trans. Auto. Control, AC-22:468-470, 1977.

[164] V. Lakshmikantham and S. Leela. Differential and Integml Inequali-

ties. Academic Press, New York, 1969.

[165] Y. S. Lee, Y. S. Moon, and W. H. Kwon. Delay-dependent guaran-

teed cost control for uncertain state-delayed systems. 2001 American
Control Conference, Arlington, VA, 3376-3381, June 25-27, 2001.

[166] E. B. Lee, A. W. Olbrot, and S. H. Zak. Cyclicity of polynomial

systems. IEEE Trans. Auto. Control, AC-27(2):451-453, 1982.

[167] E. B. Lee and S. Zak. On spectrum placement for linear time invariant
delay systems. IEEE Trans. Auto. Control, AC-27(2):446-449, 1982.
340 Bibliography

[168] J. H. Lee, S. W. Kim, and W. H. Kwon. Memoryless Hoo controllers

for state delayed systems. IEEE TI-ans. Auto. Control, 39(1):159-162,
1994.
[169] B. Lehman and K. Shujaee. Delay independent stability conditions
and decay estimates for time-varying functional differential equations.
IEEE TI-ans. Auto. Control, 39(8):1673-1676, 1994.
[170] B. Lehman, J. Bentsman, S. Verduyn Lunel and E. I. Verriest. Vi-
brational control of nonlinear time-lag systems with arbitrarily large
but bounded delay: averaging theory, stabilizability, and transient
behavior. IEEE TI-ans. Auto. Control, AC-39(5):898-912, 1994.
[171] N. Levinson and R. M. Redheffer. Complex Variables. Baltimore,
Holden-Day, 1970.
[172] R. M. Lewis and B. D. O. Anderson. Necessary and sufficient condi-
tions for delay-independent stability of autonomous systems. IEEE
TI-ans. Auto. Control, AC-25(4):735-739, Aug. 1980.
[173] X. Li and C. E. de Souza. Criteria for robust stability and stabi-
lization of uncertain linear systems with state delay. Automatica,
33(9):1657-1662, 1997.
[174] X. Li and C. E. de Souza. Delay-dependent robust stability and sta-
bilization of uncertain time-delay systems: a linear matrix inequality
approach. IEEE TI-ans. Auto. Control, 42(8):1144-1148, 1997.
[175] K. Liu. Caratheodory approximate solutions for a class of semilinear
stochastic evolution equations with time delays. J. Math. Anal. Appl.,
220:349-364, 1998.
[176] J. Louisell. A stability analysis for a class of differential-delay equa-
tion having time-varying delay. In S. Busenberg and M. Martelli
(Eds.), Delay Differential Equations and Dynamical Systems. Lec-
ture Notes in Mathematics, 1475, Springer-Verlag, Berlin Heidelberg,
225-242, 1991.
[177] J. Louisell. Growth estimates and asymptotic stability for a class of
differential-delay equation having time-varying delay. J. Math. Anal.
Appl., 164:453-479, 1992.
[178] J. Louisell. A velocity functional for an analysis of stability in delay-
differential equations. SIAM J. Math. Anal., 25:1181-1194, 1994.
[179] J. Louisell. Stability criteria with a symmetric operator occurring
in linear and nonlinear delay-differential equations. In Differential
Equations, Dynamical Systems, and Control Science. Lecture Notes
in Pure and Applied Mathematics, 152, Marcel Dekker, New York,
159-171, 1994.
 Bibliography 341

[180] J. Louisell. Absolute stability in linear delay-differential systems: Ill-

posedness and robustness. IEEE Trans. Auto. Control, 40:1288-1291,
1995.
[181] J. Louisell. Numerics of the stability exponent and eigenvalue ab-
scissas of a matrix delay system. In L. Dugard and E. I. Verriest
(Eds.), Stability and Control of Time-delay Systems. Lecture Notes
in Control and Information Sciences, 228, Springer-Verlag, 140--157,
1997.
[182] J. Louisell. Accurate determination of the stability exponent and
eigenvalue abscissas in a linear delay-differential system. Proceedings
of the 4th European Control Conference, Brussels, 1997.

[183] J. S. Luo and P. P. J. van den Bosch. Independent of delay sta-

bility criteria for uncertain linear state space models. Automatica,
33(2):171-179, Feb. 1997.

[184] Z.-H. Luo, B.-Z. Guo, and O. Morgul. Stability and Stabilization of In-
finite Dimensional Systems with Applications. Springer-Verlag, Lon-
don, 1999.

[185] A. M. Lyapunov. The general problem of the stability of motion. Tay-

lor and Francis, London, 1992.

[186] N. MacDonald. Biological Delay Systems: Linear Stability Theory.

Cambridge University Press, Cambridge, 1989.

[187] M. Malek-Zavarei and M. Jamshidi. Time Delay Systems: Analysis,

Optimization and Applications. North-Holland Systems and Control
Series, 9, North Holland, Amsterdam, 1987.

[188] X. Mao. Comments on "an improved Razumikhin-type theorem and

its applications." IEEE Trans. Auto. Control, 42(3):429-430, 1997.

[189] M. Marden. Geometry of Polynomials. American Mathematical So-

ciety, Providence, RI, 1989.

[190] J. E. Marshall, H. G6recki, K. Walton, and A. Korytowski. Time-

Delay Systems: Stability and Performance Criteria with Applications,
Ellis Horwood, New York, NY, 1992.

[191] S. M. Meerkov, T. Runolfsson, and Z. Schuss. Correlation free forms

for nonlinear stochastic systems. J. Math. Anal. Appl. 139:483-495,
1989.

[192] F. C. Moon (Ed.). Dynamics and Chaos in Manufacturing Process.

Wiley, New York, 1998.
342 Bibliography

[193] F. C. Moon and M. A. Johnson. Nonlinear dynamics and chaos in

manufacturing process. In F. C. Moon (Ed.), Dynamics and Chaos
in Manufacturing Process, Wiley, New York, 3-32, 1998.
[194] T. Mori, N. Fukuma, and M. Kuwahara. Simple stability criteria for
single and composite linear systems with time delays. Int. J. Control,
34(6):1175-1184, 1981.
[195J T. Mori, N. Fukuma, and M. Kuwahara. On an estimate of the decay
rate for stable linear delay systems. Int. J. Control, 36(1):95-97, 1982.
[196J T. Mori, E. Noldus, and M. Kuwahara. A way to stabilize linear
systems with delayed state. Automatica, 19(5):571-573, 1983.
[197] T. Mori and H. Kokame. Stability of x{t) = Ax{t) +Bx{t-r). IEEE
Trans. Auto. Control, AC-34{ 4):460-462, April 1989.
[198] T. Mori, H. Kokame, and M. Kuwahara. Analysis of time-delay sys-
tems: stability and instability. Proceedings 25th IEEE Conference on
Decision and Control, Athens, Greece, 895-898, Dec. 1986.
[199J T. Mori. Criteria for asymptotic stability of linear time-delay systems.
IEEE Trans. Auto. Control, AC-30:158-160, 1985.
[200] A. S. Morse. Ring models for delay differential systems. Automatica,
12:529--531, 1976.
[201] H. Nakajima. On analytical properties of delayed feedback control of
chaos. Physics Letters A, 232:207-210, 1997.
[202] H. Nakajima and Y. Ueda. Limitation of generalized delayed feedback
control. Physica D, 111:143-150, 1998.
[203] Y. Nesterov and A. Nemirovskii. Interior-Point Polynomial Algo-
rithms in Convex Programming. SIAM, Philadelphia, 1994.
[204J S.-1. Niculescu. Delay Effects on Stability: A Robust Control Ap-
proach. Springer-Verlag, Heidelberg, Germany, 200l.
[205] S.-1. Niculescu. Stability and hyperbolicity of linear systems with
delayed state: a matrix pencil approach. IMA J. Math. Control Info.,
15(4):331-347, 1998.
[206J SA. Niculescu. Delay-interval stability and hyperbolicity of linear
time-delay systems: a matrix pencil approach. 4th European Control
Conference, Brussels, Belgium, July 1997.
[207] S.-1. Niculescu. Hoc memoryless control with an a-stability constraint
for time-delay systems: an LMI approach. IEEE Trans. Auto. Con-
trol, 43(5):739--743, 1998.
 Bibliography 343

[208] S. I. Niculescu and J. Chen. Frequency sweeping tests for asymp-

totic stability: a model transformation for multiple delays. Proceed-
ings 38th IEEE Conference on Decision and Control, Phoenix, AZ,
4678-4683, Dec. 1999.

[209] S. I. Niculescu, L. Dugard, and J. M. Dion. Stabilite et stabilisation

robustes des systemes a retard. Proc. Journees Robustesse, Toulouse,
France, 1995.

[210] S.-I. Niculescu, J.-M. Dion, L. Dugard, and H. Li. Stability of linear
systems with delayed state: an L.M.I. approach. European Journal
of Automation, special issue on analysis and control of time-delay
systems, 31:955-970, 1997.

[211] S. I. Niculescu and V. Ionescu. On delay-independent stability crite-

ria: a matrix-pencil approach. IMA J. Math. Control Info., 14(3):299-
306, Sept. 1997.

[212] S.-I. Niculescu, A. T. Neto, J.-M. Dion, and L. Dugard. Delay-

dependent stability of linear systems with delayed state: an LMI ap-
proach. Proceedings 34th IEEE Conference on Decision and Control,
1495-1497, 1995.

[213] S.-I. Niculescu, C. E. de Souza, J.-M. Dion, and L. Dugard. Robust

stability and stabilization of uncertain linear systems with state de-
lay: single delay case (I), and multiple delays case (II). Proceedings
IFAC Workshop Robust Control Design, Rio de Janeiro, Brazil, 469-
474 and 475-480, 1994.

[214] S.-I. Niculescu, E. I. Verriest, L. Dugard, and J.-M. Dion. Stabil-

ity and robust stability of time-delay systems: a guided tour. In L.
Dugard and E. I. Verriest (Eds.), Stability and Control of Time-delay
Systems, Lecture Notes in Control and Information Sciences, 228,
Springer-Verlag, New York, 1-71, 1997.

[215] E. Noldus. Stabilization of a class of distributional convolution equa-

tions. Int. J. Control, 41(4):947-960, 1985.

[216] N. Olgac, H. Elmali and S. Vijayan. Introduction to the dual fre-

quency fixed delayed resonator. J. Sound and Vibration, 189(3):355-
367,1996.

[217] N. Olgac, H. Elmali, M. Hosek, and M. Renzulli. Active vibration con-

trol of distributed systems using delayed resonator with acceleration
feedback. ASME J. Dynamic Systems, Measurement and Control,
119:380-389, 1997.
344 Bibliography

[218J N. Olgac and B. T. Holm-Hansen. A novel active vibration absorption

technique: delayed resonator. J. Sound and Vibration, 176(1):93-104,
1994.

[219J A. Packard and J. Doyle. The complex structured singular value.

Automatica, 29:71-109, 1993.

[220] P. Park. A delay-dependent stability for systems with uncertain time-

invariant delays. IEEE Trans. Auto. Control, 44(4):876-877, 1999.

[221] P. Park. Delay-dependent stability criteria for systems with uncertain

time-varying delays. Private Communication with Keqin Gu, 2000.

[222J J. C. Passelergue, N. Hadjsaid, D. Georges, V. Hanneton, S. Vitet,

and R. Feuillet. An efficient index to deal with interaction phenom-
ena of FACTS devices in power systems. POWERCON'98, Beijing,
August 18-22, 1998.

[223] J. E. Perkins, U. Helmke, and J. B. Moore. Balancing factorization via

gradient flow techniques and the singular value decomposition. Pro-
ceedings 30th IEEE Conference on Decision and Control, Brighton,
England, 551-556, December 1991.

[224J I. Petersen and C. V. HoHot. A Riccati equation approach to the

stabilization of uncertain linear systems. Automatica, 22(4):397-411,
1986.
[225] B. Petterson, R. Robinett, and J. Werner. Lag-stabilized force feed-
back damping. Sandia National Laboratories, SAND91-0194, UC-
406, 1991.

[226] S. Phoojaruenchanachai and K. Furuta. Memoryless stabilization of

uncertain linear systems including time-varying state delays. IEEE
Trans. Auto. Control, 37(7):1022-1026, 1992.

[227] L. S. Pontryagin. On the zeros of some elementary transcendental

functions. [Russian] Izv. Akad. Nauk SSR, Ser. Mat., 6:115-134, 1942.
[English translation] Amer. Math. Soc. Transl. 2, 1:95-110, 1955.

[228] V. M. Popov. Absolute stability of nonlinear systems of automatic

control. Automation and Remote Control, 22:857-875, 1961.

[229] V. M. Popov. Hyperstability of Control Systems. Springer-Verlag,

New York, 1973.

[230] V. M. Popov and A. Halanay. On the stability of nonlinear automatic

control systems with lagging argument. Automatic Remote Control,
23:783-786, 1962.
 Bibliography 345

[231J K. Pyragas. Continuous control of chaos by self-controlling feedback.

Physics Letters A, 170:421--428, 1992.

[232J K. Pyragas. Control of chaos via extended delay feedback. Physics

Letters A, 206:323-330, 1995.

[233J B. S. Razumikhin. On the stability of systems with a delay. [RussianJ

Prikl. Mat. Meh., 20:500-512, 1956.

[234J B. S. Razumikhin. Application of Liapunov's method to problems in

the stability of systems with a delay. [RussianJ Automat. i Telemeh.,
21:740-749, 1960.

[235J Z. V. Rekasius. A stability test for systems with delays. Proceedings

of Joint Automatic Control Conference, Paper No. TP9-A, 1980.

[236J Y. M. Repin. Quadratic Lyapunov functionals for systems with delay.

[RussianJ Prikl. Mat. Meh., 29:564-566, 1965.

[237J N. Rouche and J. Mawhin. Equations Differentielles Ordinaires, Vol.

2. Mason, Paris, 1973.

[238J M. G. Safonov. Stability margins of diagonally perturbed multivari-

able feedback systems. Proc. lEE, Part D, 129(6):251-256, Nov. 1982.

[239J D. Salamon. Control and Observation of Neutral Systems. Pitman,

Boston, 1984.

[240J D. Sbarbaro-Hofer. Control of a steel rolling mill. IEEE Control Sys-

tems Magazine, 69-75, June 1993.

[241] E. Schwengeler. Geometrisches uber die Verteilung der Nullstellen

spezieller ganzer Funktionen (Exponentialsummen), Ph.D. disserta-
tion, Zurich, 1925.

[242] C. Scherer, P. Gahinet, and M. ChilaH. Multiobjective output-

feedback control via LMI optimization. IEEE Trans. Auto. Control,
42(7):896-911, 1997.

[243J U. Shaked, 1. Yaesh, and C. E. de Souza. Bounded real criteria for

linear time-delay systems. IEEE Trans. Auto. Control, 43(7):1016-
1022, 1998.

[244J J.-C. Shen, B.-S. Chen, and F.-C. Kung. Memoryless stabilization of
uncertain dynamic delay systems: Riccati equation approach. IEEE
Trans. Auto. Control, 36(5):638-640, 1991.

[245J A. N. Shiryayev. Probability. Springer-Verlag, New York, 1996.

346 Bibliography

[246] A. Snyder, N. Hadjsaid, D. Georges, L. Mili, A. Phadke, O. Fau-

con, and S. Vitet. Inter-area oscillation damping with power system
stabilizers and synchronized phasor measurements. POWERCON'98,
Beijing, August 18-22, 1998.

[247] E. Sontag, H. J. Sussmann, and Y. Yang. A general result on the

stabilization of linear systems using bounded controls. IEEE Trans.
Auto. Control, 39:2411-2425, 1994.
[248] E. D. Sontag. Linear systems over commutative rings: a survey.
Richerche Automat, 7:1-34, 1976.
[249] T. T. Soong. Active Structural Control; Theory and Practice. Wiley,
New York, 1990.
[250] C. E. de Souza and X. Li. Delay-dependent robust Hoo control of un-
certain linear state-delayed systems. Automatica, 35:1313-1321,1999.
[251] G. Stepan. Retarded Dynamical Systems; Stability and Characteristic
function. Wiley, New York, 1989.
[252] G. Stepan. Delay-differential equation models for machine tool chat-
ter. In F. C. Moon (Ed.), Dynamics and Chaos in Manufacturing
Process. Wiley, New York, 165-192, 1998.
[253] J. H. Suo The asymptotic stability of linear autonomous systems with
commensurate delays. IEEE Trans. Auto. Control, AC-40:1114-1118,
1995.
[254] T.-J. Su and C.-G. Huang. Robust stability of delay-dependence for
linear uncertain systems. IEEE Trans. Auto. Control, 37(10):1656-
1659, 1992.
[255] Y.-J. Sun, J.-G. Hsieh, and H.-C. Yang. On the stability of uncer-
tain systems with multiple time-varying delays. IEEE Trans. Auto.
Control, 42(1):101-105,1997.
[256] F. Takens. Detecting Strange Attractors in Turbulence. Lecture Notes
in Mathematics, 898, Springer-Verlag, Berlin, 1981.

[257] A. Thowsen. An analytical stability test for a class of time-delay

systems. IEEE Trans. Auto. Control, AC-26:735-736, 1981.
[258] A. Thowsen. The Routh-Hurwitz method for stability determination
of linear differential-difference systems. Int. J. Control, 33:991-995,
1981.
[259] A. Thowsen. Delay-independent asymptotic stability of linear sys-
tems. Proc. lEE, Part D, 129(3):73-75, 1982.
 Bibliography 347

[260] A. Thowsen. Uniform ultimate boundedness of the solutions of un-

certain dynamic delay systems with state-dependent and memoryless
feedback control. Int. J. Control, 37{5}:1135-1143, 1983.

[261] O. Toker. On the conservatism of upper bound tests for structured

singular value analysis. Proceedings 35th IEEE Conference on Deci-
sion and Control, Kobe, Japan, 1295-1300, Dec. 1996.

[262] O. Toker and H. Ozbay. On the complexity of purely complex J.L com-
putation and related problems in multidimensional systems. IEEE
Trans. Auto. Control, AC-43:409-414, 1998.

[263] O. Toker and H. Ozbay. Complexity issues in robust stability of linear

delay-differential systems. Math., Control, Signals, Syst., 9:386-400,
1996.

[264] B. H. Tongue. Principles of Vibration. Oxford University Press, New

York, 1996.

[265] H. Trinh and M. Aldeen. On the stability of linear systems with

delayed perturbation. IEEE Trans. Auto. Control, 39{9}:1948-1951,
1994.

[266] L. Vandenberghe and S. Boyd. Semidefinite programming. SIAM Re-

view, 38{1}:49-95, 1996.

[267] V. N. Vapnik. The Nature of Statistical Learning Theory. Springer-

Verlag, Berlin, 1995.

[268] E. 1. Verriest. Stabilization of deterministic and stochastic systems

with uncertain time delays. Proceedings of the 33d IEEE Conference
on Decision and Control, Orlando, FL, 3829-3834, Dec. 1994.

[269] E. 1. Verriest. Robust stability of time-varying systems with unknown

bounded delays. Proceedings of the 33d IEEE Conference on Decision
and Control, Orlando, FL, 417-422, Dec. 1994.

[270] E.1. Verriest. Stability of systems with distributed delays. IFAC Sys-
tem Structure and Control, Nantes, France, 294-299, 1995.

[271] E. 1. Verriest. Linear systems with rational distributed delay: re-

duction and stability. 1999 European Control Conference, Karlsruhe,
Germany, August 31-September 3, 1999.

[272] E. 1. Verriest and P. Florchinger. Stability of stochastic systems

with uncertain time delays. Systems and Control Letters, 24(1):41-47,
1995.
348 Bibliography

[273] E. I. Verriest and M. K. H. Fan. Robust stability of nonlinearly per-

turbed delay systems. Proceedings of the 35th IEEE Conference on
Decision and Control, Kobe, Japan, 2090-2091, December 1996.

[274] E. I. Verriest, M. K. H. Fan, and J. Kullstam. Frequency domain

robust stability criteria for linear delay systems. Proceedings 32nd
IEEE Conference on Decision and Control, San Antonio, TX, 3473-
3478, December 1993.
[275] E. I. Verriest and A. F. Ivanov. Robust stabilization of systems with
delayed feedback. Proceedings of the 2nd International Symposium on
Implicit and Robust Systems, Warsau, Poland, 190-193, July 1991.
[276] E. I. Verriest and A. F. Ivanov. Robust stability of systems
with delayed feedback. Circuits, Systems, and Signal Processing,
13(2/3):213--222, 1994.
[277] M. Vidyasagar. Nonlinear Systems Analysis, 2nd Ed. Prentice-Hall,
Englewood Cliffs, NJ, 1993.
[278] M. Vidyasagar. A Theory of Learning and Generalization. Springer-
Verlag, Berlin, 1997.

[279] M. Vidyasagar. Statistical learning theory and randomized algo-

rithms for control. IEEE Control Systems Magazine, 18:69-85, 1998.

[280] V. Volterra. Sulle equazioni integrodifferenziali della teorie

dell'elasticita. Atti. Accad. Lincei, 18:295, 1909.
[281] V. Volterra. Theorie Mathematique de la lutte pour la vie. Gauthier-
Villars, Paris, 1931.
[282] K. Walton and J. E. Marshall. Direct method for TDS stability analy-
sis. Proc. lEE, Part D, 134(2):101-107, March 1987.
[283] S. S. Wang, B. S. Chen, and T. P. Lin. Robust stability of time-delay
systems. Int. J. Control, 46(3):963-976, 1987.

[284] M. Wirth and U. Brunner. Observer models for real-time traffic con-
trol. Proceedings of the American Control Conference, Baltimore,
MD, 923-926, 1994.
[285] H. Wu and K. Mizukami. Linear and nonlinear stabilizing continuous
controllers of uncertain dynamical systems including state delays.
IEEE Trans. Auto. Control, 41(1):116--121, 1996.
[286] H. Wu and K. Mizukami. Robust stability criteria for dynamical sys-
tems including delayed perturbation. IEEE Trans. Auto. Control,
40(3):487-490, 1995.
 Bibliography 349

[287] V. A. Yakubovich. The solution of certain matrix inequalities en-

countered in automatic control theory. Soviet Math. Dokl., 3:620-623,
1962.
[288] V. A. Yakubovich. Solution of certain matrix inequalities encountered
in nonlinear control theory. Soviet Math. Dokl., 5:652-656, 1964.
[289] V. A. Yakubovich. Frequency conditions for the existence of ab-
solutely stable periodic and almost periodic limiting regimes of
control systems with many nonstationary elements. IFAC World
Congress, London, 1966.
[290] G. Zames. On the input-output stability of nonlinear time-varying
feedback systems, part I. IEEE Trans. Auto. Control, AC-1l:228-
238, 1966.
[291] G. Zames. On the input-output stability of nonlinear time-varying
feedback systems, part II. IEEE Trans. Auto. Control, AC-1l:465-
476, 1966.
[292] J. Zhang, C. R. Knopse, and P. Tsiotras. Stability of time-delay sys-
tems: equivalence between Lyapunov and scaled small-gain condi-
tions. IEEE Trans. Auto. Control, 46(3):482-486, 2001.
[293] K. Zhou, Robust Control of Uncertain Systems and HOC Optimiza-
tion. Ph.D. Dissertation, University of Minnesota, Twin Cities, MN,
1988.
[294] K. Zhou, J. C. Doyle, and K. Glover. Robust and Optimal Control.
Prentice-Hall, Upper Saddle River, NJ, 1996.
Index
II . 112,2 norm, 71 discretized Lyapunov functional,
227, 241, 260
additional dynamics, 162 discretized Lyapunov functional
approximation approach, 296 method, 183
DLF, see discretized Lyapunov
B7too ,73 functional
Bilherz criterion, 67 dual form, 153
bilinear programming, 91
bilinear transformation, 34, 36 Edge Theorem, 131
bivariate polynomial, 33, 63 €-sector, 121
block-diagonal uncertainty exponential diagram, 120
pulling out uncertainty, 231
Bounded Real Lemma, 282 feedback(G, F), 278
Fourier transform, 71
Cauchy formula, 7 functional differential equation, 8
characteristic equation, 7, 18 neutral,9
characteristic function, 18, 107 nonlinear, 9
characteristic matrix, 18 retarded, 8
characteristic quasipolynomial, fundamental solution, 6, 16, 238
18,117
characteristic root, see Pole gain, 72
Class NP, 89 generalized eigenvalue, 48
Class P, 88 generalized eigenvalue problem,
comparison system, 279 311, see also GEVP
computational complexity, 87 GEVP, 316
conjugate polynomial, 35, 56, 58
conjugate symmetry, 35, 39 ?too norm, 62, 72, 98, 278
H x ,276
continuous norm, 6, 10
Hamiltonian matrix, 61, 98
convex direction, 133
Hardy space, 72
delay Holder vector norm, 70, 100
commensurate, 20 Hurwitz stability, 67
concentrated, see pointwise
induced matrix norm, 97
distributed, 9, 154, 209, 218
induced matrix norms, 103
incommensurate, 20
induced norm, 71
multiple, 234
initial condition, 5
pointwise, 19
interval polynomial, 139
time-varying, 207
interval quasipolynomial, 134
delay margin, 33, 108
delay system, see time delay Jensen Inequality, 322
system Jordan decomposition, 47
delayed resonator, 3
discretization, 244, 262 Knapsack problem, 90
352 Index

Kronecker product, 314 parameterized, 280

Kronecker sum, 64, 314 see structured singular value
j.L,
multivariate polynomial, 136
£1 norm, see Holder vector norm diamond family, 142
£2 space, 71
£2 norm, see Holder vector norm neutral systems, 106
L2 space, 70 norm
£ex; norm, see Holder vector induced, 313
norm vector, 312
L 2 -norm,276 NP-complete, 89
L 2-stable, 277 NP-hard, 78, 87, 89
L 2e +,276
L 2+,276 Orlando formula, 56
Laplace transform, 85, 97
P T ,276
linear operator, 71, 74
Parseval identity, 71
LM! problem, 79, 315
pole, 7, 18
variable elimination, 319
polynomial reducibility, 89
logarithmic E-sector, 123
polynomial-time algorithm, 88
Lyapunov equation, 98, 106
potential diagram, 123
Lyapunov function condition,
pseudo-delay method, 37
150
Lyapunov-Krasovskii derivative quasipolynomial, 32
condition, 150, 251,
257,268 Razumikhin derivative
Lyapunov-Krasovskii functional, condition, 150
11 Razumikhin Theorem, 14, 207
complete quadratic, 176, regenerative chatter, 2
224, 238, 241 Roche's theorem, 39
discretization, 183, 227 Routh array, 38, 44
Lyapunov-Krasovskii functional Routh-Hurwitz criterion, 38
condition, 150, 246, 263
Lyapunov-Krasovskii Stability ST, 276
Theorem, 12 S-procedure, 319
Schur complement, 50
M -.Do loop, 72 for determinants, 310
matrix measure, 97 for matrix inequalities, 318
matrix pencil, 48 Schur determinant formula, 50,
maximum modulus principle, 46, 57, 62, 95
79,92 Schur-Cohn criterion, 55
method of steps, 6 Schur-Cohn-Fujiwara matrix, 56
minimal realization, 61 sesquilinear programming, 91
model transformation, 87, 158, shift matrix, 56
212, 219 singular value, 72
explicit, 158 singular value decomposition,
implicit, 165 312
 Index 353

small gain condition, 278 LF norm-bounded, 202

small gain problem, 287 linear fractional
scaled,287 norm-bounded, see LF
small gain theorem, 40, 72 norm-bounded
small J.L theorem, 77 norm-bounded, 201
SSS, see Strict sense stable poly topic, 198
stability, 11 repeated complex scalar, 78
2-D polynomial, 35 structured, 77, 87
asymptotic, 11 subpolytopic, 199
BlBO,71 unstructured, 73
delay-dependent, 32, 159,
212, 235, 236 value set, 126
delay-independent, 32, 152, variation-of-constant formula,
155, 172, 214, 235 see Cauchy formula
global, 11
LTI system, 19 well posed, 202
multivariate polynomial,
zero exclusion principle, 113, 127
138
robust, 73
robust stability, 77
strict sense, 138
uniform, 11
stability exponent, 19
structured singular value, 77, 79,
86, 92
subharmonic function, 46, 78-80

lI'-Knapsack problem, 90
time-delay, see delay
time-delay system, 9
linear, 15
linear time-invariant, 17
LTI, see linear
time-invariant
neutral, 20
triangle inequality, 99
trivial solution, 10
Tsypkin theorem, 40
Tsypkin's test, 75
two-variable criterion, 34, 82

uncertain quasipolynomial, 126

uncertainty
block diagonal, 77, 205
dynamic feedback, 281
Control Engineering

Series Editor
William S. Levine
Department of Electrical and Computer Engineering
University of Maryland
College Park, MD 20742-3285
USA

Aims and Scope

Control engineering is an increasingly diverse subject, whose technologies range from

simple mechanical devices to complex electro-mechanical systems. Applications are seen
in everything from biological control systems to the tracking controllers of CD players.
Some methods, H-infinity design for example, for the analysis and design of control
systems are based on sophisticated mathematics while others, such as PID control, are
understood and implemented through experimentation and empirical analysis.

The Birkhauser series Systems and Control: Foundations and Applications examines
the abstract and theoretical mathematical aspects of control. Control Engineering
complements this effort through a study of the industrial and applied implementation of
control- from techniques for analysis and design to hardware implementation, test, and
evaluation. While recognizing the harmony between abstract theory and physical applica-
tion, these publications emphasize real-world results and concerns. Problems and ex-
amples use the least amount of abstraction required, remaining committed to issues of
consequence, such as cost, tradeoffs, reliability, and power consumption.

The series includes professional expository monographs, advanced textbooks, handbooks,

and thematic compilations of applications/case studies.

Readership

The publications will appeal to a broad interdisciplinary readership of engineers at the

graduate and professional levels. Applied theorists and practitioners in industry and academia
will find the publications accessible across the varied terrain of control engineering research.

Preparation of manuscripts

We encourage the preparation of manuscripts in 1:\1EX for delivery as camera-ready hard

copy, which leads to rapid publication, or on a diskette.

Proposals should be sent directly to the editor or to: Birkhauser Boston,

675 Massachusetts Avenue, Cambridge, MA 02139, U.S.A.
Published Books

Lyapunov-Based Control of Mechanical Systems

M.S. de Queiroz, D.M. Dawson, S.P. Nagarkatti, and F. Zhang

Nonlinear Control and Analytic Mechanics

H.G. Kwatny and G.L Blankenship

Qualitative Theory of Hybrid Dynamical Systems

AS. Matveev and A. V. Savkin

Robust Kalman Filtering for Signals and Systems with

Large Uncertainties
I.R. Peterson and A V. Savkin

Control Systems Theory with Engineering Applications

S.E. Lyshevski

Control Systems with Actuator Saturation:

Analysis and Design
T. Hu and Z Lin

Deterministic and Stochastic Time-Delay Systems

E.K. Boukas and ZK. Uu

Hybrid Dynamical Systems

A V. Savkin and R.J. Evans

Stability and Control of Dynamical Systems with Applications:

A Tribute to Anthony N. Michel
D. Uu and P.J. Antsaklis, editors

Stability of Time-Delay Systems

K. Gu, V.L Kharitonov, and J. Chen

Nonlinear Control of Engineering Systems:

A Lyapunov-Based Approach
W.E. Dixon, A Behal, D.M. Dawson, and S.P. Nagarkatti

Forthcoming books in the Control Engineering series

PID Controllers for Time-Delay Systems

S.P. Bhattacharyya, A Datta, and G.J. Silva

Verification and Synthesis of Hybrid Systems

E. Asarin, T. Dang, and O. Maler

Qualitative Nonlinear Dynamics of Communication Networks

V. Kulkarni