Uploaded by mcmaster2015

0.0 (0)

**System and Control Theory
**

SIAM Studies in Applied Mathematics

This series of monographs focuses on mathematics and its applications to problems

of current concern to industry, government, and society. These monographs will be of

interest to applied mathematicians, numerical analysts, statisticians, engineers, and

scientists who have an active need to learn useful methodology.

Series List

Vol. 1 Lie-B¨ acklund Transformations in Applications

Robert L. Anderson and Nail H. Ibragimov

Vol. 2 Methods and Applications of Interval Analysis

Ramon E. Moore

Vol. 3 Ill-Posed Problems for Integrodiﬀerential Equations in Mechanics and

Electromagnetic Theory

Frederick Bloom

Vol. 4 Solitons and the Inverse Scattering Transform

Mark J. Ablowitz and Harvey Segur

Vol. 5 Fourier Analysis of Numerical Approximations of Hyperbolic Equations

Robert Vichnevetsky and John B. Bowles

Vol. 6 Numerical Solution of Elliptic Problems

Garrett Birkhoﬀ and Robert E. Lynch

Vol. 7 Analytical and Numerical Methods for Volterra Equations

Peter Linz

Vol. 8 Contact Problems in Elasticity: A Study of Variational Inequalities and

Finite Element Methods

N. Kikuchi and J. T. Oden

Vol. 9 Augmented Lagrangian and Operator-Splitting Methods in Nonlinear

Mechanics

Roland Glowinski and P. Le Tallec

Vol. 10 Boundary Stabilization of Thin Plate Splines

John E. Lagnese

Vol. 11 Electro-Diﬀusion of Ions

Isaak Rubinstein

Vol. 12 Mathematical Problems in Linear Viscoelasticity

Mauro Fabrizio and Angelo Morro

Vol. 13 Interior-Point Polynomial Algorithms in Convex Programming

Yurii Nesterov and Arkadii Nemirovskii

Vol. 14 The Boundary Function Method for Singular Perturbation Problems

Adelaida B. Vasil’eva, Valentin F. Butuzov, and Leonid V. Kalachev

Vol. 15 Linear Matrix Inequalities in System and Control Theory

Stephen Boyd, Laurent El Ghaoui, Eric Feron, and Venkataramanan

Balakrishnan

Stephen Boyd, Laurent El Ghaoui,

Eric Feron, and Venkataramanan Balakrishnan

Linear Matrix Inequalities in

System and Control Theory

Society for Industrial and Applied Mathematics

®

Philadelphia

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

All rights reserved. No part of this book may be reproduced, stored, or transmitted in any

manner without the written permission of the Publisher. For information, write to the Society

for Industrial and Applied Mathematics, 3600 University City Science Center, Philadelphia,

Pennsylvania 19104-2688.

The royalties from the sales of this book are being placed in a fund to help students

attend SIAM meetings and other SIAM related activities. This fund is administered by

SIAM and qualiﬁed individuals are encouraged to write directly to SIAM for guidelines.

Library of Congress Cataloging-in-Publication Data

Linear matrix inequalities in system and control theory / Stephen Boyd

. . . [et al.].

p. cm. -- (SIAM studies in applied mathematics ; vol. 15)

Includes bibliographical references and index.

ISBN 0-89871-334-X

1. Control theory. 2. Matrix inequalities. 3. Mathematical

optimization. I. Boyd, Stephen P. II. Series: SIAM studies in

applied mathematics : 15.

QA402.3.L489 1994

515’.64--dc20 94-10477

Contents

Preface vii

Acknowledgments ix

1 Introduction 1

1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 A Brief History of LMIs in Control Theory . . . . . . . . . . . . . . . 2

1.3 Notes on the Style of the Book . . . . . . . . . . . . . . . . . . . . . . 4

1.4 Origin of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Some Standard Problems Involving LMIs 7

2.1 Linear Matrix Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 Some Standard Problems . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.3 Ellipsoid Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.4 Interior-Point Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.5 Strict and Nonstrict LMIs . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.6 Miscellaneous Results on Matrix Inequalities . . . . . . . . . . . . . . 22

2.7 Some LMI Problems with Analytic Solutions . . . . . . . . . . . . . . 24

Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3 Some Matrix Problems 37

3.1 Minimizing Condition Number by Scaling . . . . . . . . . . . . . . . . 37

3.2 Minimizing Condition Number of a Positive-Deﬁnite Matrix . . . . . . 38

3.3 Minimizing Norm by Scaling . . . . . . . . . . . . . . . . . . . . . . . 38

3.4 Rescaling a Matrix Positive-Deﬁnite . . . . . . . . . . . . . . . . . . . 39

3.5 Matrix Completion Problems . . . . . . . . . . . . . . . . . . . . . . . 40

3.6 Quadratic Approximation of a Polytopic Norm . . . . . . . . . . . . . 41

3.7 Ellipsoidal Approximation . . . . . . . . . . . . . . . . . . . . . . . . . 42

Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4 Linear Diﬀerential Inclusions 51

4.1 Diﬀerential Inclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4.2 Some Speciﬁc LDIs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

4.3 Nonlinear System Analysis via LDIs . . . . . . . . . . . . . . . . . . . 54

Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

5 Analysis of LDIs: State Properties 61

5.1 Quadratic Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5.2 Invariant Ellipsoids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

v

vi Contents

Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

6 Analysis of LDIs: Input/Output Properties 77

6.1 Input-to-State Properties . . . . . . . . . . . . . . . . . . . . . . . . . 77

6.2 State-to-Output Properties . . . . . . . . . . . . . . . . . . . . . . . . 84

6.3 Input-to-Output Properties . . . . . . . . . . . . . . . . . . . . . . . . 89

Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

7 State-Feedback Synthesis for LDIs 99

7.1 Static State-Feedback Controllers . . . . . . . . . . . . . . . . . . . . . 99

7.2 State Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

7.3 Input-to-State Properties . . . . . . . . . . . . . . . . . . . . . . . . . 104

7.4 State-to-Output Properties . . . . . . . . . . . . . . . . . . . . . . . . 107

7.5 Input-to-Output Properties . . . . . . . . . . . . . . . . . . . . . . . . 109

7.6 Observer-Based Controllers for Nonlinear Systems . . . . . . . . . . . 111

Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

8 Lur’e and Multiplier Methods 119

8.1 Analysis of Lur’e Systems . . . . . . . . . . . . . . . . . . . . . . . . . 119

8.2 Integral Quadratic Constraints . . . . . . . . . . . . . . . . . . . . . . 122

8.3 Multipliers for Systems with Unknown Parameters . . . . . . . . . . . 124

Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

9 Systems with Multiplicative Noise 131

9.1 Analysis of Systems with Multiplicative Noise . . . . . . . . . . . . . . 131

9.2 State-Feedback Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . 134

Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

10 Miscellaneous Problems 141

10.1 Optimization over an Aﬃne Family of Linear Systems . . . . . . . . . 141

10.2 Analysis of Systems with LTI Perturbations . . . . . . . . . . . . . . . 143

10.3 Positive Orthant Stabilizability . . . . . . . . . . . . . . . . . . . . . . 144

10.4 Linear Systems with Delays . . . . . . . . . . . . . . . . . . . . . . . . 144

10.5 Interpolation Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

10.6 The Inverse Problem of Optimal Control . . . . . . . . . . . . . . . . . 147

10.7 System Realization Problems . . . . . . . . . . . . . . . . . . . . . . . 148

10.8 Multi-Criterion LQG . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

10.9 Nonconvex Multi-Criterion Quadratic Problems . . . . . . . . . . . . . 151

Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

Notation 157

List of Acronyms 159

Bibliography 161

Index 187

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Preface

The basic topic of this book is solving problems from system and control theory using

convex optimization. We show that a wide variety of problems arising in system

and control theory can be reduced to a handful of standard convex and quasiconvex

optimization problems that involve matrix inequalities. For a few special cases there

are “analytic solutions” to these problems, but our main point is that they can be

solved numerically in all cases. These standard problems can be solved in polynomial-

time (by, e.g., the ellipsoid algorithm of Shor, Nemirovskii, and Yudin), and so are

tractable, at least in a theoretical sense. Recently developed interior-point methods

for these standard problems have been found to be extremely eﬃcient in practice.

Therefore, we consider the original problems from system and control theory as solved.

This book is primarily intended for the researcher in system and control theory,

but can also serve as a source of application problems for researchers in convex op-

timization. Although we believe that the methods described in this book have great

practical value, we should warn the reader whose primary interest is applied control

engineering. This is a research monograph: We present no speciﬁc examples or nu-

merical results, and we make only brief comments about the implications of the results

for practical control engineering. To put it in a more positive light, we hope that this

book will later be considered as the ﬁrst book on the topic, not the most readable or

accessible.

The background required of the reader is knowledge of basic system and control

theory and an exposure to optimization. Sontag’s book Mathematical Control The-

ory [Son90] is an excellent survey. Further background material is covered in the

texts Linear Systems [Kai80] by Kailath, Nonlinear Systems Analysis [Vid92] by

Vidyasagar, Optimal Control: Linear Quadratic Methods [AM90] by Anderson and

Moore, and Convex Analysis and Minimization Algorithms I [HUL93] by Hiriart–

Urruty and Lemar´echal.

We also highly recommend the book Interior-point Polynomial Algorithms in Con-

vex Programming [NN94] by Nesterov and Nemirovskii as a companion to this book.

The reader will soon see that their ideas and methods play a critical role in the basic

idea presented in this book.

vii

Acknowledgments

We thank A. Nemirovskii for encouraging us to write this book. We are grateful to S.

Hall, M. Gevers, and A. Packard for introducing us to multiplier methods, realization

theory, and state-feedback synthesis methods, respectively. This book was greatly

improved by the suggestions of J. Abedor, P. Dankoski, J. Doyle, G. Franklin, T.

Kailath, R. Kosut, I. Petersen, E. Pyatnitskii, M. Rotea, M. Safonov, S. Savastiouk,

A. Tits, L. Vandenberghe, and J. C. Willems. It is a special pleasure to thank V.

Yakubovich for his comments and suggestions.

The research reported in this book was supported in part by AFOSR (under

F49620-92-J-0013), NSF (under ECS-9222391), and ARPA (under F49620-93-1-0085).

L. El Ghaoui and E. Feron were supported in part by the D´el´egation G´en´erale pour

l’Armement. E. Feron was supported in part by the Charles Stark Draper Career

Development Chair at MIT. V. Balakrishnan was supported in part by the Control

and Dynamical Systems Group at Caltech and by NSF (under NSFD CDR-8803012).

This book was typeset by the authors using L

A

T

E

X, and many macros originally

written by Craig Barratt.

Stephen Boyd Stanford, California

Laurent El Ghaoui Paris, France

Eric Feron Cambridge, Massachusetts

Venkataramanan Balakrishnan College Park, Maryland

ix

Linear Matrix Inequalities in

System and Control Theory

Chapter 1

Introduction

1.1 Overview

The aim of this book is to show that we can reduce a very wide variety of prob-

lems arising in system and control theory to a few standard convex or quasiconvex

optimization problems involving linear matrix inequalities (LMIs). Since these result-

ing optimization problems can be solved numerically very eﬃciently using recently

developed interior-point methods, our reduction constitutes a solution to the original

problem, certainly in a practical sense, but also in several other senses as well. In com-

parison, the more conventional approach is to seek an analytic or frequency-domain

solution to the matrix inequalities.

The types of problems we consider include:

• matrix scaling problems, e.g., minimizing condition number by diagonal scaling

• construction of quadratic Lyapunov functions for stability and performance anal-

ysis of linear diﬀerential inclusions

• joint synthesis of state-feedback and quadratic Lyapunov functions for linear

diﬀerential inclusions

• synthesis of state-feedback and quadratic Lyapunov functions for stochastic and

delay systems

• synthesis of Lur’e-type Lyapunov functions for nonlinear systems

• optimization over an aﬃne family of transfer matrices, including synthesis of

multipliers for analysis of linear systems with unknown parameters

• positive orthant stability and state-feedback synthesis

• optimal system realization

• interpolation problems, including scaling

• multicriterion LQG/LQR

• inverse problem of optimal control

In some cases, we are describing known, published results; in others, we are extending

known results. In many cases, however, it seems that the results are new.

By scanning the list above or the table of contents, the reader will see that Lya-

punov’s methods will be our main focus. Here we have a secondary goal, beyond

showing that many problems from Lyapunov theory can be cast as convex or quasi-

convex problems. This is to show that Lyapunov’s methods, which are traditionally

1

2 Chapter 1 Introduction

applied to the analysis of system stability, can just as well be used to ﬁnd bounds on

system performance, provided we do not insist on an “analytic solution”.

1.2 A Brief History of LMIs in Control Theory

The history of LMIs in the analysis of dynamical systems goes back more than 100

years. The story begins in about 1890, when Lyapunov published his seminal work

introducing what we now call Lyapunov theory. He showed that the diﬀerential equa-

tion

d

dt

x(t) = Ax(t) (1.1)

is stable (i.e., all trajectories converge to zero) if and only if there exists a positive-

deﬁnite matrix P such that

A

T

P +PA < 0. (1.2)

The requirement P > 0, A

T

P +PA < 0 is what we now call a Lyapunov inequality on

P, which is a special form of an LMI. Lyapunov also showed that this ﬁrst LMI could

be explicitly solved. Indeed, we can pick any Q = Q

T

> 0 and then solve the linear

equation A

T

P+PA = −Q for the matrix P, which is guaranteed to be positive-deﬁnite

if the system (1.1) is stable. In summary, the ﬁrst LMI used to analyze stability of a

dynamical system was the Lyapunov inequality (1.2), which can be solved analytically

(by solving a set of linear equations).

The next major milestone occurs in the 1940’s. Lur’e, Postnikov, and others in

the Soviet Union applied Lyapunov’s methods to some speciﬁc practical problems

in control engineering, especially, the problem of stability of a control system with a

nonlinearity in the actuator. Although they did not explicitly form matrix inequalities,

their stability criteria have the form of LMIs. These inequalities were reduced to

polynomial inequalities which were then checked “by hand” (for, needless to say, small

systems). Nevertheless they were justiﬁably excited by the idea that Lyapunov’s

theory could be applied to important (and diﬃcult) practical problems in control

engineering. From the introduction of Lur’e’s 1951 book [Lur57] we ﬁnd:

This book represents the ﬁrst attempt to demonstrate that the ideas ex-

pressed 60 years ago by Lyapunov, which even comparatively recently ap-

peared to be remote from practical application, are now about to become a

real medium for the examination of the urgent problems of contemporary

engineering.

In summary, Lur’e and others were the ﬁrst to apply Lyapunov’s methods to practical

control engineering problems. The LMIs that resulted were solved analytically, by

hand. Of course this limited their application to small (second, third order) systems.

The next major breakthrough came in the early 1960’s, when Yakubovich, Popov,

Kalman, and other researchers succeeded in reducing the solution of the LMIs that

arose in the problem of Lur’e to simple graphical criteria, using what we now call

the positive-real (PR) lemma (see §2.7.2). This resulted in the celebrated Popov

criterion, circle criterion, Tsypkin criterion, and many variations. These criteria could

be applied to higher order systems, but did not gracefully or usefully extend to systems

containing more than one nonlinearity. From the point of view of our story (LMIs in

control theory), the contribution was to show how to solve a certain family of LMIs

by graphical methods.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

1.2 A Brief History of LMIs in Control Theory 3

The important role of LMIs in control theory was already recognized in the early

1960’s, especially by Yakubovich [Yak62, Yak64, Yak67]. This is clear simply from

the titles of some of his papers from 1962–5, e.g., The solution of certain matrix

inequalities in automatic control theory (1962), and The method of matrix inequalities

in the stability theory of nonlinear control systems (1965; English translation 1967).

The PR lemma and extensions were intensively studied in the latter half of the

1960s, and were found to be related to the ideas of passivity, the small-gain criteria

introduced by Zames and Sandberg, and quadratic optimal control. By 1970, it was

known that the LMI appearing in the PR lemma could be solved not only by graphical

means, but also by solving a certain algebraic Riccati equation (ARE). In a 1971

paper [Wil71b] on quadratic optimal control, J. C. Willems is led to the LMI

_

A

T

P +PA+Q PB +C

T

B

T

P +C R

_

≥ 0, (1.3)

and points out that it can be solved by studying the symmetric solutions of the ARE

A

T

P +PA−(PB +C

T

)R

−1

(B

T

P +C) +Q = 0,

which in turn can be found by an eigendecomposition of a related Hamiltonian matrix.

(See §2.7.2 for details.) This connection had been observed earlier in the Soviet Union,

where the ARE was called the Lur’e resolving equation (see [Yak88]).

So by 1971, researchers knew several methods for solving special types of LMIs:

direct (for small systems), graphical methods, and by solving Lyapunov or Riccati

equations. From our point of view, these methods are all “closed-form” or “analytic”

solutions that can be used to solve special forms of LMIs. (Most control researchers

and engineers consider the Riccati equation to have an “analytic” solution, since the

standard algorithms that solve it are very predictable in terms of the eﬀort required,

which depends almost entirely on the problem size and not the particular problem

data. Of course it cannot be solved exactly in a ﬁnite number of arithmetic steps for

systems of ﬁfth and higher order.)

In Willems’ 1971 paper we ﬁnd the following striking quote:

The basic importance of the LMI seems to be largely unappreciated. It

would be interesting to see whether or not it can be exploited in compu-

tational algorithms, for example.

Here Willems refers to the speciﬁc LMI (1.3), and not the more general form that

we consider in this book. Still, Willems’ suggestion that LMIs might have some

advantages in computational algorithms (as compared to the corresponding Riccati

equations) foreshadows the next chapter in the story.

The next major advance (in our view) was the simple observation that:

The LMIs that arise in system and control theory can be formulated as

convex optimization problems that are amenable to computer solution.

Although this is a simple observation, it has some important consequences, the most

important of which is that we can reliably solve many LMIs for which no “analytic

solution” has been found (or is likely to be found).

This observation was made explicitly by several researchers. Pyatnitskii and Sko-

rodinskii [PS82] were perhaps the ﬁrst researchers to make this point, clearly and

completely. They reduced the original problem of Lur’e (extended to the case of mul-

tiple nonlinearities) to a convex optimization problem involving LMIs, which they

This electronic version is for personal use and may not be duplicated or distributed.

4 Chapter 1 Introduction

then solved using the ellipsoid algorithm. (This problem had been studied before, but

the “solutions” involved an arbitrary scaling matrix.) Pyatnitskii and Skorodinskii

were the ﬁrst, as far as we know, to formulate the search for a Lyapunov function as

a convex optimization problem, and then apply an algorithm guaranteed to solve the

optimization problem.

We should also mention several precursors. In a 1976 paper, Horisberger and Be-

langer [HB76] had remarked that the existence of a quadratic Lyapunov function that

simultaneously proves stability of a collection of linear systems is a convex problem

involving LMIs. And of course, the idea of having a computer search for a Lya-

punov function was not new—it appears, for example, in a 1965 paper by Schultz et

al. [SSHJ65].

The ﬁnal chapter in our story is quite recent and of great practical importance: the

development of powerful and eﬃcient interior-point methods to solve the LMIs that

arise in system and control theory. In 1984, N. Karmarkar introduced a new linear

programming algorithm that solves linear programs in polynomial-time, like the ellip-

soid method, but in contrast to the ellipsoid method, is also very eﬃcient in practice.

Karmarkar’s work spurred an enormous amount of work in the area of interior-point

methods for linear programming (including the rediscovery of eﬃcient methods that

were developed in the 1960s but ignored). Essentially all of this research activity con-

centrated on algorithms for linear and (convex) quadratic programs. Then in 1988,

Nesterov and Nemirovskii developed interior-point methods that apply directly to con-

vex problems involving LMIs, and in particular, to the problems we encounter in this

book. Although there remains much to be done in this area, several interior-point

algorithms for LMI problems have been implemented and tested on speciﬁc families

of LMIs that arise in control theory, and found to be extremely eﬃcient.

A summary of key events in the history of LMIs in control theory is then:

• 1890: First LMI appears; analytic solution of the Lyapunov LMI via Lyapunov

equation.

• 1940’s: Application of Lyapunov’s methods to real control engineering prob-

lems. Small LMIs solved “by hand”.

• Early 1960’s: PR lemma gives graphical techniques for solving another family

of LMIs.

• Late 1960’s: Observation that the same family of LMIs can be solved by solving

an ARE.

• Early 1980’s: Recognition that many LMIs can be solved by computer via

convex programming.

• Late 1980’s: Development of interior-point algorithms for LMIs.

It is fair to say that Yakubovich is the father of the ﬁeld, and Lyapunov the grandfather

of the ﬁeld. The new development is the ability to directly solve (general) LMIs.

1.3 Notes on the Style of the Book

We use a very informal mathematical style, e.g., we often fail to mention regularity or

other technical conditions. Every statement is to be interpreted as being true modulo

appropriate technical conditions (that in most cases are trivial to ﬁgure out).

We are very informal, perhaps even cavalier, in our reduction of a problem to

an optimization problem. We sometimes skip “details” that would be important if

the optimization problem were to be solved numerically. As an example, it may be

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

1.4 Origin of the Book 5

necessary to add constraints to the optimization problem for normalization or to ensure

boundedness. We do not discuss initial guesses for the optimization problems, even

though good ones may be available. Therefore, the reader who wishes to implement

an algorithm that solves a problem considered in this book should be prepared to

make a few modiﬁcations or additions to our description of the “solution”.

In a similar way, we do not pursue any theoretical aspects of reducing a problem

to a convex problem involving matrix inequalities. For example, for each reduced

problem we could state, probably simplify, and then interpret in system or control

theoretic terms the optimality conditions for the resulting convex problem. Another

fascinating topic that could be explored is the relation between system and control

theory duality and convex programming duality. Once we reduce a problem arising

in control theory to a convex program, we can consider various dual optimization

problems, lower bounds for the problem, and so on. Presumably these dual problems

and lower bounds can be given interesting system-theoretic interpretations.

We mostly consider continuous-time systems, and assume that the reader can

translate the results from the continuous-time case to the discrete-time case. We

switch to discrete-time systems when we consider system realization problems (which

almost always arise in this form) and also when we consider stochastic systems (to

avoid the technical details of stochastic diﬀerential equations).

The list of problems that we consider is meant only to be representative, and

certainly not exhaustive. To avoid excessive repetition, our treatment of problems

becomes more terse as the book progresses. In the ﬁrst chapter on analysis of lin-

ear diﬀerential inclusions, we describe many variations on problems (e.g., computing

bounds on margins and decay rates); in later chapters, we describe fewer and fewer

variations, assuming that the reader could work out the extensions.

Each chapter concludes with a section entitled Notes and References, in which we

hide proofs, precise statements, elaborations, and bibliography and historical notes.

The completeness of the bibliography should not be overestimated, despite its size

(over 500 entries). The appendix contains a list of notation and a list of acronyms

used in the book. We apologize to the reader for the seven new acronyms we introduce.

To lighten the notation, we use the standard convention of dropping the time

argument from the variables in diﬀerential equations. Thus, ˙ x = Ax is our short

form for dx/dt = Ax(t). Here A is a constant matrix; when we encounter time-

varying coeﬃcients, we will explicitly show the time dependence, as in ˙ x = A(t)x.

Similarly, we drop the dummy variable from deﬁnite integrals, writing for example,

_

T

0

u

T

ydt for

_

T

0

u(t)

T

y(t)dt. To reduce the number of parentheses required, we adopt

the convention that the operators Tr (trace of a matrix) and E (expected value) have

lower precedence than multiplication, transpose, etc. Thus, Tr A

T

B means Tr

_

A

T

B

_

.

1.4 Origin of the Book

This book started out as a section of the paper Method of Centers for Minimizing

Generalized Eigenvalues, by Boyd and El Ghaoui [BE93], but grew too large to be

a section. For a few months it was a manuscript (that presumably would have been

submitted for publication as a paper) entitled Generalized Eigenvalue Problems Aris-

ing in Control Theory. Then Feron, and later Balakrishnan, started adding material,

and soon it was clear that we were writing a book, not a paper. The order of the

authors’ names reﬂects this history.

This electronic version is for personal use and may not be duplicated or distributed.

Chapter 2

Some Standard Problems

Involving LMIs

2.1 Linear Matrix Inequalities

A linear matrix inequality (LMI) has the form

F(x)

∆

= F

0

+

m

i=1

x

i

F

i

> 0, (2.1)

where x ∈ R

m

is the variable and the symmetric matrices F

i

= F

T

i

∈ R

n×n

, i =

0, . . . , m, are given. The inequality symbol in (2.1) means that F(x) is positive-

deﬁnite, i.e., u

T

F(x)u > 0 for all nonzero u ∈ R

n

. Of course, the LMI (2.1) is

equivalent to a set of n polynomial inequalities in x, i.e., the leading principal minors

of F(x) must be positive.

We will also encounter nonstrict LMIs, which have the form

F(x) ≥ 0. (2.2)

The strict LMI (2.1) and the nonstrict LMI (2.2) are closely related, but a precise

statement of the relation is a bit involved, so we defer it to §2.5. In the next few

sections we consider strict LMIs.

The LMI (2.1) is a convex constraint on x, i.e., the set {x | F(x) > 0} is convex.

Although the LMI (2.1) may seem to have a specialized form, it can represent a

wide variety of convex constraints on x. In particular, linear inequalities, (convex)

quadratic inequalities, matrix norm inequalities, and constraints that arise in control

theory, such as Lyapunov and convex quadratic matrix inequalities, can all be cast in

the form of an LMI.

Multiple LMIs F

(1)

(x) > 0, . . . , F

(p)

(x) > 0 can be expressed as the single LMI

diag(F

(1)

(x), . . . , F

(p)

(x)) > 0. Therefore we will make no distinction between a set

of LMIs and a single LMI, i.e., “the LMI F

(1)

(x) > 0, . . . , F

(p)

(x) > 0” will mean “the

LMI diag(F

(1)

(x), . . . , F

(p)

(x)) > 0”.

When the matrices F

i

are diagonal, the LMI F(x) > 0 is just a set of linear

inequalities. Nonlinear (convex) inequalities are converted to LMI form using Schur

complements. The basic idea is as follows: the LMI

_

Q(x) S(x)

S(x)

T

R(x)

_

> 0, (2.3)

7

8 Chapter 2 Some Standard Problems Involving LMIs

where Q(x) = Q(x)

T

, R(x) = R(x)

T

, and S(x) depend aﬃnely on x, is equivalent to

R(x) > 0, Q(x) −S(x)R(x)

−1

S(x)

T

> 0. (2.4)

In other words, the set of nonlinear inequalities (2.4) can be represented as the

LMI (2.3).

As an example, the (maximum singular value) matrix norm constraint Z(x) < 1,

where Z(x) ∈ R

p×q

and depends aﬃnely on x, is represented as the LMI

_

I Z(x)

Z(x)

T

I

_

> 0

(since Z < 1 is equivalent to I −ZZ

T

> 0). Note that the case q = 1 reduces to a

general convex quadratic inequality on x.

The constraint c(x)

T

P(x)

−1

c(x) < 1, P(x) > 0, where c(x) ∈ R

n

and P(x) =

P(x)

T

∈ R

n×n

depend aﬃnely on x, is expressed as the LMI

_

P(x) c(x)

c(x)

T

1

_

> 0.

More generally, the constraint

Tr S(x)

T

P(x)

−1

S(x) < 1, P(x) > 0,

where P(x) = P(x)

T

∈ R

n×n

and S(x) ∈ R

n×p

depend aﬃnely on x, is handled by

introducing a new (slack) matrix variable X = X

T

∈ R

p×p

, and the LMI (in x and

X):

Tr X < 1,

_

X S(x)

T

S(x) P(x)

_

> 0.

Many other convex constraints on x can be expressed in the form of an LMI; see the

Notes and References.

2.1.1 Matrices as variables

We will often encounter problems in which the variables are matrices, e.g., the Lya-

punov inequality

A

T

P +PA < 0, (2.5)

where A ∈ R

n×n

is given and P = P

T

is the variable. In this case we will not write out

the LMI explicitly in the form F(x) > 0, but instead make clear which matrices are

the variables. The phrase “the LMI A

T

P +PA < 0 in P” means that the matrix P is

a variable. (Of course, the Lyapunov inequality (2.5) is readily put in the form (2.1),

as follows. Let P

1

, . . . , P

m

be a basis for symmetric n ×n matrices (m = n(n +1)/2).

Then take F

0

= 0 and F

i

= −A

T

P

i

− P

i

A.) Leaving LMIs in a condensed form

such as (2.5), in addition to saving notation, may lead to more eﬃcient computation;

see §2.4.4.

As another related example, consider the quadratic matrix inequality

A

T

P +PA+PBR

−1

B

T

P +Q < 0, (2.6)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

2.2 Some Standard Problems 9

where A, B, Q = Q

T

, R = R

T

> 0 are given matrices of appropriate sizes, and

P = P

T

is the variable. Note that this is a quadratic matrix inequality in the variable

P. It can be expressed as the linear matrix inequality

_

−A

T

P −PA−Q PB

B

T

P R

_

> 0.

This representation also clearly shows that the quadratic matrix inequality (2.6) is

convex in P, which is not obvious.

2.1.2 Linear equality constraints

In some problems we will encounter linear equality constraints on the variables, e.g.

P > 0, A

T

P +PA < 0, Tr P = 1, (2.7)

where P ∈ R

k×k

is the variable. Of course we can eliminate the equality constraint to

write (2.7) in the formF(x) > 0. Let P

1

, . . . , P

m

be a basis for symmetric k×k matrices

with trace zero (m = (k(k + 1)/2) −1) and let P

0

be a symmetric k ×k matrix with

Tr P

0

= 1. Then take F

0

= diag(P

0

, −A

T

P

0

−P

0

A) and F

i

= diag(P

i

, −A

T

P

i

−P

i

A)

for i = 1, . . . , m.

We will refer to constraints such as (2.7) as LMIs, leaving any required elimination

of equality constraints to the reader.

2.2 Some Standard Problems

Here we list some common convex and quasiconvex problems that we will encounter

in the sequel.

2.2.1 LMI problems

Given an LMI F(x) > 0, the corresponding LMI Problem (LMIP) is to ﬁnd x

feas

such

that F(x

feas

) > 0 or determine that the LMI is infeasible. (By duality, this means:

Find a nonzero G ≥ 0 such that Tr GF

i

= 0 for i = 1, . . . , m and Tr GF

0

≤ 0; see the

Notes and References.) Of course, this is a convex feasibility problem. We will say

“solving the LMI F(x) > 0” to mean solving the corresponding LMIP.

As an example of an LMIP, consider the “simultaneous Lyapunov stability prob-

lem” (which we will see in §5.1): We are given A

i

∈ R

n×n

, i = 1, . . . , L, and need to

ﬁnd P satisfying the LMI

P > 0, A

T

i

P +PA

i

< 0, i = 1, . . . , L,

or determine that no such P exists. Determining that no such P exists is equivalent

to ﬁnding Q

0

, . . . , Q

L

, not all zero, such that

Q

0

≥ 0, . . . , Q

L

≥ 0, Q

0

=

L

i=1

_

Q

i

A

T

i

+A

i

Q

i

_

, (2.8)

which is another (nonstrict) LMIP.

This electronic version is for personal use and may not be duplicated or distributed.

10 Chapter 2 Some Standard Problems Involving LMIs

2.2.2 Eigenvalue problems

The eigenvalue problem (EVP) is to minimize the maximum eigenvalue of a matrix

that depends aﬃnely on a variable, subject to an LMI constraint (or determine that

the constraint is infeasible), i.e.,

minimize λ

subject to λI −A(x) > 0, B(x) > 0

where A and B are symmetric matrices that depend aﬃnely on the optimization

variable x. This is a convex optimization problem.

EVPs can appear in the equivalent form of minimizing a linear function subject

to an LMI, i.e.,

minimize c

T

x

subject to F(x) > 0

(2.9)

with F an aﬃne function of x. In the special case when the matrices F

i

are all diagonal,

this problem reduces to the general linear programming problem: minimizing the

linear function c

T

x subject to a set of linear inequalities on x.

Another equivalent form for the EVP is:

minimize λ

subject to A(x, λ) > 0

where A is aﬃne in (x, λ). We leave it to the reader to verify that these forms are

equivalent, i.e., any can be transformed into any other.

As an example of an EVP, consider the problem (which appears in §6.3.2):

minimize γ

subject to

_

−A

T

P −PA−C

T

C PB

B

T

P γI

_

> 0, P > 0

where the matrices A ∈ R

n×n

, B ∈ R

n×p

, and C ∈ R

m×n

are given, and P and γ are

the optimization variables. From our remarks above, this EVP can also be expressed

in terms of the associated quadratic matrix inequality:

minimize γ

subject to A

T

P +PA+C

T

C +γ

−1

PBB

T

P < 0, P > 0

2.2.3 Generalized eigenvalue problems

The generalized eigenvalue problem (GEVP) is to minimize the maximum generalized

eigenvalue of a pair of matrices that depend aﬃnely on a variable, subject to an LMI

constraint. The general form of a GEVP is:

minimize λ

subject to λB(x) −A(x) > 0, B(x) > 0, C(x) > 0

where A, B and C are symmetric matrices that are aﬃne functions of x. We can

express this as

minimize λ

max

(A(x), B(x))

subject to B(x) > 0, C(x) > 0

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

2.2 Some Standard Problems 11

where λ

max

(X, Y ) denotes the largest generalized eigenvalue of the pencil λY − X

with Y > 0, i.e., the largest eigenvalue of the matrix Y

−1/2

XY

−1/2

. This GEVP is

a quasiconvex optimization problem since the constraint is convex and the objective,

λ

max

(A(x), B(x)), is quasiconvex. This means that for feasible x, ˜ x and 0 ≤ θ ≤ 1,

λ

max

(A(θx + (1 −θ)˜ x), B(θx + (1 −θ)˜ x)) ≤

max {λ

max

(A(x), B(x)), λ

max

(A(˜ x), B(˜ x))} .

Note that when the matrices are all diagonal and A(x) and B(x) are scalar, this

problem reduces to the general linear-fractional programming problem, i.e., minimiz-

ing a linear-fractional function subject to a set of linear inequalities. In addition,

many nonlinear quasiconvex functions can be represented in the form of a GEVP with

appropriate A, B, and C; see the Notes and References.

An equivalent alternate form for a GEVP is

minimize λ

subject to A(x, λ) > 0

where A(x, λ) is aﬃne in x for ﬁxed λ and aﬃne in λ for ﬁxed x, and satisﬁes the

monotonicity condition λ > µ =⇒ A(x, λ) ≥ A(x, µ). As an example of a GEVP,

consider the problem

maximize α

subject to −A

T

P −PA−2αP > 0, P > 0

where the matrix A is given, and the optimization variables are the symmetric matrix

P and the scalar α. (This problem arises in §5.1.)

2.2.4 A convex problem

Although we will be concerned mostly with LMIPs, EVPs, and GEVPs, we will also

encounter the following convex problem, which we will abbreviate CP:

minimize log det A(x)

−1

subject to A(x) > 0, B(x) > 0

(2.10)

where A and B are symmetric matrices that depend aﬃnely on x. (Note that when

A > 0, log det A

−1

is a convex function of A.)

Remark: Problem CP can be transformed into an EVP, since det A(x) > λ can

be represented as an LMI in x and λ; see [NN94, ¸6.4.3]. As we do with variables

which are matrices, we leave these problems in the more natural form.

As an example of CP, consider the problem:

minimize log det P

−1

subject to P > 0, v

T

i

Pv

i

≤ 1, i = 1, . . . , L

(2.11)

here v

i

∈ R

n

are given and P = P

T

∈ R

n×n

is the variable.

We will encounter several variations of this problem, which has the following

interpretation. Let E denote the ellipsoid centered at the origin determined by P,

This electronic version is for personal use and may not be duplicated or distributed.

12 Chapter 2 Some Standard Problems Involving LMIs

E

∆

=

_

z

¸

¸

z

T

Pz ≤ 1

_

. The constraints are simply v

i

∈ E. Since the volume of E is pro-

portional to (det P)

−1/2

, minimizing log det P

−1

is the same as minimizing the volume

of E. So by solving (2.11), we ﬁnd the minimum volume ellipsoid, centered at the ori-

gin, that contains the points v

1

, . . . , v

L

, or equivalently, the polytope Co{v

1

, . . . , v

L

},

where Co denotes the convex hull.

2.2.5 Solving these problems

The standard problems (LMIPs, EVPs, GEVPs, and CPs) are tractable, from both

theoretical and practical viewpoints:

• They can be solved in polynomial-time (indeed with a variety of interpretations

for the term “polynomial-time”).

• They can be solved in practice very eﬃciently.

By “solve the problem” we mean: Determine whether or not the problem is fea-

sible, and if it is, compute a feasible point with an objective value that exceeds the

global minimum by less than some prespeciﬁed accuracy.

2.3 Ellipsoid Algorithm

We ﬁrst describe a simple ellipsoid algorithm that, roughly speaking, is guaranteed

to solve the standard problems. We describe it here because it is very simple and,

from a theoretical point of view, eﬃcient (polynomial-time). In practice, however, the

interior-point algorithms described in the next section are much more eﬃcient.

Although more sophisticated versions of the ellipsoid algorithm can detect in-

feasible constraints, we will assume that the problem we are solving has at least one

optimal point, i.e., the constraints are feasible. (In the feasibility problem, we consider

any feasible point as being optimal.) The basic idea of the algorithm is as follows. We

start with an ellipsoid E

(0)

that is guaranteed to contain an optimal point. We then

compute a cutting plane for our problem that passes through the center x

(0)

of E

(0)

.

This means that we ﬁnd a nonzero vector g

(0)

such that an optimal point lies in the

half-space

_

z

¸

¸

g

(0)T

(z −x

(0)

) ≤ 0

_

(or in the half-space

_

z

¸

¸

g

(0)T

(z −x

(0)

) < 0

_

,

depending on the situation). (We will explain how to do this for each of our problems

later.) We then know that the sliced half-ellipsoid

E

(0)

∩

_

z

¸

¸

¸ g

(0)T

(z −x

(0)

) ≤ 0

_

contains an optimal point. Now we compute the ellipsoid E

(1)

of minimum volume

that contains this sliced half-ellipsoid; E

(1)

is then guaranteed to contain an optimal

point. The process is then repeated.

We now describe the algorithm more explicitly. An ellipsoid E can be described

as

E =

_

z

¸

¸

(z −a)

T

A

−1

(z −a) ≤ 1

_

where A = A

T

> 0. The minimum volume ellipsoid that contains the half-ellipsoid

_

z

¸

¸

(z −a)

T

A

−1

(z −a) ≤ 1, g

T

(z −a) ≤ 0

_

is given by

˜

E =

_

z

¸

¸

¸ (z − ˜ a)

T

˜

A

−1

(z − ˜ a) ≤ 1

_

,

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

2.3 Ellipsoid Algorithm 13

where

˜ a = a −

A˜ g

m+ 1

,

˜

A =

m

2

m

2

−1

_

A−

2

m+ 1

A˜ g˜ g

T

A

_

,

and ˜ g = g/

_

g

T

Ag. (We note that these formulas hold only for m ≥ 2. In the case of

one variable, the minimum length interval containing a half-interval is the half-interval

itself; the ellipsoid algorithm, in this case, reduces to the familiar bisection algorithm.)

The ellipsoid algorithm is initialized with x

(0)

and A

(0)

such that the corresponding

ellipsoid contains an optimal point. The algorithm then proceeds as follows: for

k = 1, 2, . . .

compute a g

(k)

that deﬁnes a cutting plane at x

(k)

˜ g :=

_

g

(k)T

A

(k)

g

(k)

_

−1/2

g

(k)

x

(k+1)

:= x

(k)

−

1

m+1

A

(k)

˜ g

A

(k+1)

:=

m

2

m

2

−1

_

A

(k)

−

2

m+1

A

(k)

˜ g˜ g

T

A

(k)

_

This recursion generates a sequence of ellipsoids that are guaranteed to contain an

optimal point. It turns out that the volume of these ellipsoids decreases geometrically.

We have

vol(E

(k)

) ≤ e

−

k

2m

vol(E

(0)

),

and this fact can be used to prove polynomial-time convergence of the algorithm. (We

refer the reader to the papers cited in the Notes and References for precise statements

of what we mean by “polynomial-time” and indeed, “convergence,” as well as proofs.)

We now show how to compute cutting planes for each of our standard problems.

LMIPs: Consider the LMI

F(x) = F

0

+

m

i=1

x

i

F

i

> 0.

If x does not satisfy this LMI, there exists a nonzero u such that

u

T

F(x)u = u

T

_

F

0

+

m

i=1

x

i

F

i

_

u ≤ 0.

Deﬁne g by g

i

= −u

T

F

i

u, i = 1, . . . , m. Then for any z satisfying g

T

(z − x) ≥ 0 we

have

u

T

F(z)u = u

T

_

F

0

+

m

i=1

z

i

F

i

_

u = u

T

F(x)u −g

T

(z −x) ≤ 0.

It follows that every feasible point lies in the half-space {z | g

T

(z −x) < 0}, i.e., this

g deﬁnes a cutting plane for the LMIP at the point x.

EVPs: Consider the EVP

minimize c

T

x

subject to F(x) > 0

This electronic version is for personal use and may not be duplicated or distributed.

14 Chapter 2 Some Standard Problems Involving LMIs

Suppose ﬁrst that the given point x is infeasible, i.e., F(x) > 0. Then we can con-

struct a cutting plane for this problem using the method described above for LMIPs.

In this case, we are discarding the half-space {z | g

T

(z − x) ≥ 0} because all such

points are infeasible.

Now assume that the given point x is feasible, i.e., F(x) > 0. In this case, the

vector g = c deﬁnes a cutting plane for the EVP at the point x. Here, we are discarding

the half-space {z | g

T

(z − x) > 0} because all such points (whether feasible or not)

have an objective value larger than x, and hence cannot be optimal.

GEVPs: We consider the formulation

minimize λ

max

(A(x), B(x))

subject to B(x) > 0, C(x) > 0

Here, A(x) = A

0

+

m

i=1

x

i

A

i

, B(x) = B

0

+

m

i=1

x

i

B

i

and C(x) = C

0

+

m

i=1

x

i

C

i

.

Now suppose we are given a point x. If the constraints are violated, we use the method

described for LMIPs to generate a cutting plane at x.

Suppose now that x is feasible. Pick any u = 0 such that

(λ

max

(A(x), B(x))B(x) −A(x)) u = 0.

Deﬁne g by

g

i

= −u

T

(λ

max

(A(x), B(x))B

i

−A

i

)u, i = 1, . . . , m.

We claim g deﬁnes a cutting plane for this GEVP at the point x. To see this, we note

that for any z,

u

T

(λ

max

(A(x), B(x))B(z) −A(z)) u = −g

T

(z −x).

Hence, if g

T

(z −x) ≥ 0 we ﬁnd that

λ

max

(A(z), B(z)) ≥ λ

max

(A(x), B(x))

which establishes our claim.

CPs: We now consider our standard CP (2.10). When the given point x is infeasible

we already know how to generate a cutting plane. So we assume that x is feasible. In

this case, a cutting plane is given by the gradient of the objective

−log det A(x) = −log det

_

A

0

+

m

i=1

x

i

A

i

_

at x, i.e., g

i

= −Tr A

i

A(x)

−1

. Since the objective function is convex we have for all

z,

log det A(z)

−1

≥ log det A(x)

−1

+g

T

(z −x).

In particular, g

T

(z − x) > 0 implies log det A(z)

−1

> log det A(x)

−1

, and hence z

cannot be optimal.

2.4 Interior-Point Methods

Since 1988, interior-point methods have been developed for the standard problems.

In this section we describe a simple interior-point method for solving an EVP, given

a feasible starting point. The references describe more sophisticated interior-point

methods for the other problems (including the problem of computing a feasible starting

point for an EVP).

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

2.4 Interior-Point Methods 15

2.4.1 Analytic center of an LMI

The notion of the analytic center of an LMI plays an important role in interior-point

methods, and is important in its own right. Consider the LMI

F(x) = F

0

+

m

i=1

x

i

F

i

> 0,

where F

i

= F

T

i

∈ R

n×n

. The function

φ(x)

∆

=

_

log det F(x)

−1

F(x) > 0

∞ otherwise,

(2.12)

is ﬁnite if and only if F(x) > 0, and becomes inﬁnite as x approaches the boundary

of the feasible set {x|F(x) > 0}, i.e., it is a barrier function for the feasible set. (We

have already encountered this function in our standard problem CP.)

We suppose now that the feasible set is nonempty and bounded, which implies

that the matrices F

1

, . . . , F

m

are linearly independent (otherwise the feasible set will

contain a line). It can be shown that φ is strictly convex on the feasible set, so it has

a unique minimizer, which we denote x

:

x

∆

= arg min

x

φ(x). (2.13)

We refer to x

**as the analytic center of the LMI F(x) > 0. Equivalently,
**

x

= arg max

F(x) > 0

det F(x), (2.14)

that is, F(x

**) has maximum determinant among all positive-deﬁnite matrices of the
**

form F(x).

Remark: The analytic center depends on the LMI (i.e., the data F0, . . . , Fm)

and not its feasible set: We can have two LMIs with diﬀerent analytic centers but

identical feasible sets. The analytic center is, however, invariant under congruence

of the LMI: F(x) > 0 and T

T

F(x)T > 0 have the same analytic center provided

T is nonsingular.

We now turn to the problem of computing the analytic center of an LMI. (This

is a special form of our problem CP.) Newton’s method, with appropriate step length

selection, can be used to eﬃciently compute x

**, starting from a feasible initial point.
**

We consider the algorithm:

x

(k+1)

:= x

(k)

−α

(k)

H(x

(k)

)

−1

g(x

(k)

), (2.15)

where α

(k)

is the damping factor of the kth iteration, and g(x

(k)

) and H(x

(k)

) denote

the gradient and Hessian of φ, respectively, at x

(k)

. In [NN94, §2.2], Nesterov and

Nemirovskii give a simple step length rule appropriate for a general class of barrier

functions (called self-concordant), and in particular, the function φ. Their damping

factor is:

α

(k)

:=

_

1 if δ(x

(k)

) ≤ 1/4,

1/(1 +δ(x

(k)

)) otherwise,

(2.16)

This electronic version is for personal use and may not be duplicated or distributed.

16 Chapter 2 Some Standard Problems Involving LMIs

where

δ(x)

∆

=

_

g(x)

T

H(x)

−1

g(x)

is called the Newton decrement of φ at x. Nesterov and Nemirovskii show that this

step length always results in x

(k+1)

feasible, i.e., F(x

(k+1)

) > 0, and convergence of

x

(k)

to x

.

Indeed, they give sharp bounds on the number of steps required to compute x

**to a given accuracy using Newton’s method with the step length (2.16). They prove
**

that φ(x

(k)

) −φ(x

) ≤ whenever

k ≥ c

1

+c

2

log log(1/) +c

3

_

φ(x

(0)

) −φ(x

)

_

, (2.17)

where c

1

, c

2

, and c

3

are three absolute constants, i.e., speciﬁc numbers. The ﬁrst

and second terms on the right-hand side do not depend on any problem data, i.e.,

the matrices F

0

, . . . , F

m

, and the numbers m and n. The second term grows so slowly

with required accuracy that for all practical purposes it can be lumped together with

the ﬁrst and considered an absolute constant. The last term on the right-hand side

of (2.17) depends on how “centered” the initial point is.

2.4.2 The path of centers

Now consider the standard EVP:

minimize c

T

x

subject to F(x) > 0

Let λ

opt

denote its optimal value, so for each λ > λ

opt

the LMI

F(x) > 0, c

T

x < λ (2.18)

is feasible. We will also assume that the LMI (2.18) has a bounded feasible set, and

therefore has an analytic center which we denote x

(λ):

x

(λ)

∆

= arg min

x

_

log det F(x)

−1

+ log

1

λ −c

T

x

_

.

The curve given by x

(λ) for λ > λ

opt

is called the path of centers for the EVP.

It can be shown that it is analytic and has a limit as λ →λ

opt

, which we denote x

opt

.

The point x

opt

is optimal (or more precisely, the limit of a minimizing sequence) since

for λ > λ

opt

, x

**(λ) is feasible and satisﬁes c
**

T

x

(λ) < λ.

The point x

(λ) is characterized by

∂

∂x

i

¸

¸

¸

¸

x

(λ)

_

log det F(x)

−1

+ log

1

λ −c

T

x

_

= −Tr F(x

(λ))

−1

F

i

+

c

i

λ −c

T

x

(λ)

= 0, i = 1, . . . , m.

(2.19)

2.4.3 Method of centers

The method of centers is a simple interior-point algorithm that solves an EVP, given a

feasible starting point. The algorithm is initialized with λ

(0)

and x

(0)

, with F(x

(0)

) > 0

and c

T

x

(0)

< λ

(0)

, and proceeds as follows:

λ

(k+1)

:= (1 −θ)c

T

x

(k)

+θλ

(k)

x

(k+1)

:= x

(λ

(k+1)

)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

2.4 Interior-Point Methods 17

where θ is a parameter with 0 < θ < 1.

The classic method of centers is obtained with θ = 0. In this case, however, x

(k)

does not (quite) satisfy the new inequality c

T

x < λ

(k+1)

. With θ > 0, however, the

current iterate x

(k)

is feasible for the inequality c

T

x < λ

(k+1)

, F(x) > 0, and therefore

can be used as the initial point for computing the next iterate x

(λ

(k+1)

) via Newton’s

method. Since computing an analytic center is itself a special type of CP, we can view

the method of centers as a way of solving an EVP by solving a sequence of CPs (which

is done using Newton’s method).

We now give a simple proof of convergence. Multiplying (2.19) by (x

(k)

i

− x

opt

i

)

and summing over i yields:

Tr F(x

(k)

)

−1

_

F(x

(k)

) −F(x

opt

)

_

=

1

λ

(k)

−c

T

x

(k)

c

T

(x

(k)

−x

opt

).

Since Tr F(x

(k)

)

−1

F(x

opt

) ≥ 0, we conclude that

n ≥

1

λ

(k)

−c

T

x

(k)

(c

T

x

(k)

−λ

opt

).

Replacing c

T

x

(k)

by (λ

(k+1)

−θλ

(k)

)/(1 −θ) yields

(λ

(k+1)

−λ

opt

) ≤

n +θ

n + 1

(λ

(k)

−λ

opt

),

which proves that λ

(k)

approaches λ

opt

with at least geometric convergence. Note that

we can also express the inequality above in the form

c

T

x

(k)

−λ

opt

≤ n(λ

(k)

−c

T

x

(k)

),

which shows that the stopping criterion

until

_

λ

(k)

−c

T

x

(k)

< /n

_

guarantees that on exit, the optimal value has been found within .

We make a few comments here, and refer the reader to the Notes and References for

further elaboration. First, this variation on the method of centers is not polynomial-

time, but more sophisticated versions are. Second, and perhaps more important, we

note that two simple modiﬁcations of the method of centers as described above yield

an algorithm that is fairly eﬃcient in practice. The modiﬁcations are:

• Instead of the Nesterov–Nemirovskii step length, a step length chosen to mini-

mize φ along the Newton search direction (i.e., an exact line-search) will yield

faster convergence to the analytic center.

• Instead of deﬁning x

**(λ) to be the analytic center of the LMI F(x) > 0, c
**

T

x < λ,

we deﬁne it to be the analytic center of the LMI F(x) > 0 along with q copies

of c

T

x < λ, where q > 1 is an integer. In other words we use

x

(λ)

∆

= arg min

x

_

log det F(x)

−1

+q log

1

λ −c

T

x

_

.

Using q > 1, say q ≈ n where F(x) ∈ R

n×n

, yields much faster reduction of λ

per iteration.

Among interior-point methods for the standard problems, the method of centers

is not the most eﬃcient. The most eﬃcient algorithms developed so far appear to be

primal-dual algorithms (and variations) and the projective method of Nemirovskii; see

the Notes and References.

This electronic version is for personal use and may not be duplicated or distributed.

18 Chapter 2 Some Standard Problems Involving LMIs

2.4.4 Interior-point methods and problem structure

An important feature of interior-point methods is that problem structure can be ex-

ploited to increase eﬃciency. The idea is very roughly as follows. In interior-point

methods most of the computational eﬀort is devoted to computing the Newton direc-

tion of a barrier or similar function. It turns out that this Newton direction can be

expressed as the solution of a weighted least-squares problem of the same size as the

original problem. Using conjugate-gradient and other related methods to solve these

least-squares systems gives two advantages. First, by exploiting problem structure in

the conjugate-gradient iterations, the computational eﬀort required to solve the least-

squares problems is much smaller than by standard “direct” methods such as QR or

Cholesky factorization. Second, it is possible to terminate the conjugate-gradient iter-

ations before convergence, and still obtain an approximation of the Newton direction

suitable for interior-point methods. See the Notes and References for more discussion.

An example will demonstrate the eﬃciencies that can be obtained using the tech-

niques sketched above. The problem is an EVP that we will encounter in §6.2.1.

We are given matrices A

1

, . . . , A

L

∈ R

n×n

, symmetric matrices D

1

, . . . , D

L

, E ∈

R

n×n

. We consider the EVP

minimize Tr EP

subject to A

T

i

P +PA

i

+D

i

< 0, i = 1, . . . , L

(2.20)

In this problem the variable is the matrix P, so the dimension of the optimization

variable is m = n(n + 1)/2. When the Lyapunov inequalities are combined into one

large LMI F(x) > 0, we ﬁnd that F(x) ∈ R

N×N

with N = Ln. This LMI has much

structure: It is block-diagonal with each block a Lyapunov inequality.

Vandenberghe and Boyd have developed a (primal-dual) interior-point method

that solves (2.20), exploiting the problem structure. They prove the worst-case esti-

mate of O(m

2.75

L

1.5

) arithmetic operations to solve the problem to a given accuracy.

In comparison, the ellipsoid method solves (2.20) to a given accuracy in O(m

3.5

L)

arithmetic operations (moreover, the constant hidden in the O(·) notation is much

larger for the ellipsoid algorithm).

Numerical experiments on families of problems with randomly generated data

and families of problems arising in system and control theory show that the actual

performance of the interior-point method is much better than the worst-case estimate:

O(m

α

L

β

) arithmetic operations, with α ≈ 2.1 and β ≈ 1.2. This complexity estimate

is remarkably small. For example, the cost of solving L Lyapunov equations of the

same size is O(m

1.5

L). Therefore, the relative cost of solving L coupled Lyapunov

inequalities, compared to the cost of solving L independent Lyapunov inequalities

is O(m

0.6

L

0.2

). This example illustrates one of our points: The computational cost

of solving one of the standard problems (which has no “analytic solution”) can be

comparable to the computational cost of evaluating the solution of a similar problem

that has an “analytic solution”.

2.5 Strict and Nonstrict LMIs

We have so far assumed that the optimization problems LMIP, EVP, GEVP, and CP

involve strict LMIs. We will also encounter these problems with nonstrict LMIs, or

more generally, with a mixture of strict and nonstrict LMIs.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

2.5 Strict and Nonstrict LMIs 19

As an example consider the nonstrict version of the EVP (2.9), i.e.

minimize c

T

x

subject to F(x) ≥ 0

Intuition suggests that we could simply solve the strict EVP (by, say, an interior-point

method) to obtain the solution of the nonstrict EVP. This is correct in most but not

all cases.

If the strict LMI F(x) > 0 is feasible, then we have

{x ∈ R

m

| F(x) ≥ 0} = {x ∈ R

m

| F(x) > 0}, (2.21)

i.e., the feasible set of the nonstrict LMI is the closure of the feasible set of the strict

LMI. It follows that

inf

_

c

T

x | F(x) ≥ 0

_

= inf

_

c

T

x | F(x) > 0

_

. (2.22)

So in this case, we can solve the strict EVP to obtain a solution of the nonstrict EVP.

This is true for the problems GEVP and CP as well.

We will say that the LMI F(x) ≥ 0 is strictly feasible if its strict version is feasible,

i.e., if there is some x

0

∈ R

m

such that F(x

0

) > 0. We have just seen that when an

LMI is strictly feasible, we can replace nonstrict inequality with strict inequality in the

problems EVP, GEVP, and CP in order to solve them. In the language of optimization

theory, the requirement of strict feasibility is a (very strong) constraint qualiﬁcation.

When an LMI is feasible but not strictly feasible, (2.21) need not hold, and the

EVPs with the strict and nonstrict LMIs can be very diﬀerent. As a simple example,

consider F(x) = diag(x, −x) with x ∈ R. The right-hand side of (2.22) is +∞ since

the strict LMI F(x) > 0 is infeasible. The left-hand side, however, is always 0, since

the LMI F(x) ≥ 0 has the single feasible point x = 0. This example shows one of the

two pathologies that can occur: The nonstrict inequality contains an implicit equality

constraint (in contrast with an explicit equality constraint as in §2.1.2).

The other pathology is demonstrated by the example F(x) = diag(x, 0) with

x ∈ R and c = −1. Once again, the strict LMI is infeasible so the right-hand side

of (2.22) is +∞. The feasible set for the nonstrict LMI is the interval [0, ∞) so the

right-hand side is −∞. The problem here is that F(x) is always singular. Of course,

the nonstrict LMI F(x) ≥ 0 is equivalent (in the sense of deﬁning equal feasible sets) to

the “reduced” LMI

˜

F(x) = x ≥ 0. Note that this reduced LMI satisﬁes the constraint

qualiﬁcation, i.e., is strictly feasible.

2.5.1 Reduction to a strictly feasible LMI

It turns out that any feasible nonstrict LMI can be reduced to an equivalent LMI that

is strictly feasible, by eliminating implicit equality constraints and then reducing the

resulting LMI by removing any constant nullspace.

The precise statement is: Let F

0

, . . . , F

m

∈ R

n×n

be symmetric. Then there

is a matrix A ∈ R

m×p

with p ≤ m, a vector b ∈ R

m

, and symmetric matrices

˜

F

0

, . . . ,

˜

F

p

∈ R

q×q

with q ≤ n such that:

F(x) ≥ 0 ⇐⇒

x = Az +b for some z ∈ R

p

with

˜

F(z)

∆

=

˜

F

0

+

p

i=1

z

i

˜

F

i

≥ 0

This electronic version is for personal use and may not be duplicated or distributed.

20 Chapter 2 Some Standard Problems Involving LMIs

where the LMI

˜

F(z) ≥ 0 is either infeasible or strictly feasible. See the Notes and

References for a proof.

The matrix A and vector b describe the implicit equality constraints for the LMI

F(x) ≥ 0. Similarly, the LMI

˜

F(z) ≥ 0 can be interpreted as the original LMI with

its constant nullspace removed (see the Notes and References). In most of the prob-

lems encountered in this book, there are no implicit equality constraints or nontrivial

common nullspace for F, so we can just take A = I, b = 0, and

˜

F = F.

Using this reduction we can, at least in principle, always deal with strictly feasible

LMIs. For example we have

inf

_

c

T

x

¸

¸

F(x) ≥ 0

_

= inf

_

c

T

(Az +b)

¸

¸

¸

˜

F(z) ≥ 0

_

= inf

_

c

T

(Az +b)

¸

¸

¸

˜

F(z) > 0

_

since the LMI

˜

F(z) ≥ 0 is either infeasible or strictly feasible.

2.5.2 Example: Lyapunov inequality

To illustrate the previous ideas we consider the simplest LMI arising in control theory,

the Lyapunov inequality:

A

T

P +PA ≤ 0, P > 0, (2.23)

where A ∈ R

k×k

is given, and the symmetric matrix P is the variable. Note that this

LMI contains a strict inequality as well as a nonstrict inequality.

We know from system theory that the LMI (2.23) is feasible if and only if all

trajectories of ˙ x = Ax are bounded, or equivalently, if the eigenvalues of A have

nonpositive real part, and those with zero real part are nondefective, i.e., correspond

to Jordan blocks of size one. We also know from system theory that the LMI (2.23) is

strictly feasible if and only if all trajectories of ˙ x = Ax converge to zero, or equivalently,

all the eigenvalues of A have negative real part.

Consider the interesting case where the LMI (2.23) is feasible but not strictly

feasible. From the remarks above, we see that by a change of coordinates we can put

A in the form

˜

A

∆

= T

−1

AT

= diag

__

0 ω

1

I

k1

−ω

1

I

k1

0

_

, . . . ,

_

0 ω

r

I

kr

−ω

r

I

kr

0

_

, 0

kr+1

, A

stab

_

where 0 < ω

1

< · · · < ω

r

, 0

kr+1

denotes the zero matrix in R

kr+1×kr+1

, and all the

eigenvalues of the matrix A

stab

∈ R

s×s

have negative real part. Roughly speaking,

we have separated out the stable part of A, the part corresponding to each imaginary

axis eigenvalue, and the part associated with eigenvalue zero.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

2.5 Strict and Nonstrict LMIs 21

Using standard methods of Lyapunov theory it can be shown that

_

P

¸

¸

A

T

P +PA ≤ 0, P > 0

_

=

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

T

−T

˜

PT

−1

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

˜

P = diag

__

P

1

Q

1

Q

T

1

P

1

_

, . . . ,

_

P

r

Q

r

Q

T

r

P

r

_

, P

r+1

, P

stab

_

P

i

∈ R

ki×ki

, i = 1, . . . , r + 1, Q

T

i

= −Q

i

, i = 1, . . . , r

_

P

i

Q

i

Q

T

i

P

i

_

> 0, i = 1, . . . , r, P

r+1

> 0, P

stab

> 0

A

T

stab

P

stab

+P

stab

A

stab

≤ 0

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

.

From this characterization we can ﬁnd a reduced LMI that is strictly feasible. We

can take the symmetric matrices P

1

, . . . , P

r+1

, P

stab

and the skew-symmetric matrices

Q

1

, . . . , Q

r

as the “free” variable z; the aﬃne mapping from z into x simply maps

these matrices into

P = T

−T

diag

__

P

1

Q

1

Q

T

1

P

1

_

, . . . ,

_

P

r

Q

r

Q

T

r

P

r

_

, P

r+1

, P

stab

_

T

−1

. (2.24)

Put another way, the equality constraints implicit in the LMI (2.23) are that P must

have this special structure.

Now we substitute P in the form (2.24) back into the original LMI (2.23). We

ﬁnd that

A

T

P +PA = T

T

diag

_

0, A

T

stab

P

stab

+P

stab

A

stab

_

T

where the zero matrix has size 2k

1

+· · ·+2k

r

+k

r+1

. We remove the constant nullspace

to obtain the reduced version, i.e., A

T

stab

P

stab

+ P

stab

A

stab

≤ 0. Thus, the reduced

LMI corresponding to (2.23) is

_

P

i

Q

i

Q

T

i

P

i

_

> 0, i = 1, . . . , r, P

r+1

> 0,

P

stab

> 0, A

T

stab

P

stab

+P

stab

A

stab

≤ 0.

(2.25)

This reduced LMI is strictly feasible, since we can take P

1

, . . . , P

r+1

as identity matri-

ces, Q

1

, . . . , Q

r

as zero matrices, and P

stab

as the solution of the Lyapunov equation

A

T

stab

P

stab

+P

stab

A

stab

+I = 0.

In summary, the original LMI (2.23) has one symmetric matrix of size k as variable

(i.e., the dimension of the original variable x is m = k(k + 1)/2). The reduced

LMI (2.25) has as variable the symmetric matrices P

1

, . . . , P

r+1

, P

stab

and the skew-

symmetric matrices Q

1

, . . . , Q

r

, so the dimension of the variable z in the reduced LMI

is

p =

r

i=1

k

2

i

+

k

r+1

(k

r+1

+ 1)

2

+

s(s + 1)

2

< m.

The original LMI (2.23) involves a matrix inequality of size n = 2k (i.e., two inequali-

ties of size k). The reduced LMI (2.25) involves a matrix inequality of size k +s < n.

This electronic version is for personal use and may not be duplicated or distributed.

22 Chapter 2 Some Standard Problems Involving LMIs

2.6 Miscellaneous Results on Matrix Inequalities

2.6.1 Elimination of semideﬁnite terms

In the LMIP

ﬁnd x such that F(x) > 0, (2.26)

we can eliminate any variables for which the corresponding coeﬃcient in F is semidef-

inite. Suppose for example that F

m

≥ 0 and has rank r < n. Roughly speaking, by

taking x

m

extremely large, we can ensure that F is positive-deﬁnite on the range of

F

m

, so the LMIP reduces to making F positive-deﬁnite on the nullspace of F

m

.

More precisely, let F

m

= UU

T

with U full-rank, and let

˜

U be an orthogonal

complement of U, e.g., U

T

˜

U = 0 and [U

˜

U] is of maximum rank (which in this case

just means that [U

˜

U] is nonsingular). Of course, we have

F(x) > 0 ⇐⇒

_

˜

U U

_

T

F(x)

_

˜

U U

_

> 0.

Since

_

˜

U U

_

T

F(x)

_

˜

U U

_

=

_

˜

U

T

˜

F(˜ x)

˜

U

˜

U

T

˜

F(˜ x)U

U

T

˜

F(˜ x)

˜

U U

T

˜

F(˜ x)U +x

m

(U

T

U)

2

_

,

where ˜ x = [x

1

· · · x

m−1

]

T

and

˜

F(˜ x)

∆

= F

0

+ x

1

F

1

+ · · · + x

m−1

F

m−1

, we see that

F(x) > 0 if and only if

˜

U

T

˜

F

˜

U(˜ x) > 0, and x

m

is large enough that

U

T

˜

F(˜ x)U +x

m

(U

T

U)

2

> U

T

˜

F(˜ x)

˜

U

_

˜

U

T

˜

F(˜ x)

˜

U

_

−1

˜

U

T

˜

F(˜ x)U.

Therefore, the LMIP (2.26) is equivalent to

ﬁnd x

1

, . . . , x

m−1

such that

˜

U

T

˜

F(x

1

, . . . , x

m−1

)

˜

U > 0.

2.6.2 Elimination of matrix variables

When a matrix inequality has some variables that appear in a certain form, we can

derive an equivalent inequality without those variables. Consider

G(z) +U(z)XV (z)

T

+V (z)X

T

U(z)

T

> 0, (2.27)

where the vector z and the matrix X are (independent) variables, and G(z) ∈ R

n×n

,

U(z) and V (z) do not depend on X. Matrix inequalities of the form (2.27) arise in

the controller synthesis problems described in Chapter 7.

Suppose that for every z,

˜

U(z) and

˜

V (z) are orthogonal complements of U(z)

and V (z) respectively. Then (2.27) holds for some X and z = z

0

if and only if the

inequalities

˜

U(z)

T

G(z)

˜

U(z) > 0,

˜

V (z)

T

G(z)

˜

V (z) > 0 (2.28)

hold with z = z

0

. In other words, feasibility of the matrix inequality (2.27) with

variables X and z is equivalent to the feasibility of (2.28) with variable z; we have

eliminated the matrix variable X from (2.27) to form (2.28). Note that if U(z) or

V (z) has rank n for all z, then the ﬁrst or second inequality in (2.28) disappears. We

prove this lemma in the Notes and References.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

2.6 Miscellaneous Results on Matrix Inequalities 23

We can express (2.28) in another form using Finsler’s lemma (see the Notes and

References):

G(z) −σU(z)U(z)

T

> 0, G(z) −σV (z)V (z)

T

> 0

for some σ ∈ R.

As an example, we will encounter in §7.2.1 the LMIP with LMI

Q > 0, AQ+QA

T

+BY +Y

T

B

T

< 0, (2.29)

where Q and Y are the variables. This LMIP is equivalent to the LMIP with LMI

Q > 0, AQ+QA

T

< σBB

T

,

where the variables are Q and σ ∈ R. It is also equivalent to the LMIP

Q > 0,

˜

B

T

_

AQ+QA

T

_

˜

B < 0,

with variable Q, where

˜

B is any matrix of maximum rank such that

˜

B

T

B = 0. Thus

we have eliminated the variable Y from (2.29) and reduced the size of the matrices in

the LMI.

2.6.3 The S-procedure

We will often encounter the constraint that some quadratic function (or quadratic

form) be negative whenever some other quadratic functions (or quadratic forms) are

all negative. In some cases, this constraint can be expressed as an LMI in the data

deﬁning the quadratic functions or forms; in other cases, we can form an LMI that is

a conservative but often useful approximation of the constraint.

The S-procedure for quadratic functions and nonstrict inequalities

Let F

0

, . . . , F

p

be quadratic functions of the variable ζ ∈ R

n

:

F

i

(ζ)

∆

= ζ

T

T

i

ζ + 2u

T

i

ζ +v

i

, i = 0, . . . , p,

where T

i

= T

T

i

. We consider the following condition on F

0

, . . . , F

p

:

F

0

(ζ) ≥ 0 for all ζ such that F

i

(ζ) ≥ 0, i = 1, . . . , p. (2.30)

Obviously if

there exist τ

1

≥ 0, . . . , τ

p

≥ 0 such that

for all ζ, F

0

(ζ) −

p

i=1

τ

i

F

i

(ζ) ≥ 0,

(2.31)

then (2.30) holds. It is a nontrivial fact that when p = 1, the converse holds, provided

that there is some ζ

0

such that F

1

(ζ

0

) > 0.

Remark: If the functions Fi are aﬃne, then (2.31) and (2.30) are equivalent;

this is the Farkas lemma.

Note that (2.31) can be written as

_

T

0

u

0

u

T

0

v

0

_

−

p

i=1

τ

i

_

T

i

u

i

u

T

i

v

i

_

≥ 0.

This electronic version is for personal use and may not be duplicated or distributed.

24 Chapter 2 Some Standard Problems Involving LMIs

The S-procedure for quadratic forms and strict inequalities

We will use another variation of the S-procedure, which involves quadratic forms and

strict inequalities. Let T

0

, . . . , T

p

∈ R

n×n

be symmetric matrices. We consider the

following condition on T

0

, . . . , T

p

:

ζ

T

T

0

ζ > 0 for all ζ = 0 such that ζ

T

T

i

ζ ≥ 0, i = 1, . . . , p. (2.32)

It is obvious that if

there exists τ

1

≥ 0, . . . , τ

p

≥ 0 such that T

0

−

p

i=1

τ

i

T

i

> 0, (2.33)

then (2.32) holds. It is a nontrivial fact that when p = 1, the converse holds, provided

that there is some ζ

0

such that ζ

T

0

T

1

ζ

0

> 0. Note that (2.33) is an LMI in the variables

T

0

and τ

1

, . . . , τ

p

.

Remark: The ﬁrst version of the S-procedure deals with nonstrict inequalities

and quadratic functions that may include constant and linear terms. The second

version deals with strict inequalities and quadratic forms only, i.e., quadratic

functions without constant or linear terms.

Remark: Suppose that T0, u0 and v0 depend aﬃnely on some parameter ν. Then

the condition (2.30) is convex in ν. This does not, however, mean that the problem

of checking whether there exists ν such that (2.30) holds has low complexity. On

the other hand, checking whether (2.31) holds for some ν is an LMIP in the

variables ν and τ1, . . . , τp. Therefore, this problem has low complexity. See the

Notes and References for further discussion.

S-procedure example

In Chapter 5 we will encounter the following constraint on the variable P:

for all ξ = 0 and π satisfying π

T

π ≤ ξ

T

C

T

Cξ,

_

ξ

π

_

T

_

A

T

P +PA PB

B

T

P 0

__

ξ

π

_

< 0.

(2.34)

Applying the second version of the S-procedure, (2.34) is equivalent to the existence

of τ ≥ 0 such that

_

_

A

T

P +PA+τC

T

C PB

B

T

P −τI

_

_

< 0.

Thus the problem of ﬁnding P > 0 such that (2.34) holds can be expressed as an

LMIP (in P and the scalar variable τ).

2.7 Some LMI Problems with Analytic Solutions

There are analytic solutions to several LMI problems of special form, often with im-

portant system and control theoretic interpretations. We brieﬂy describe some of these

results in this section. At the same time we introduce and deﬁne several important

terms from system and control theory.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

2.7 Some LMI Problems with Analytic Solutions 25

2.7.1 Lyapunov’s inequality

We have already mentioned the LMIP associated with Lyapunov’s inequality, i.e.

P > 0, A

T

P +PA < 0

where P is variable and A ∈ R

n×n

is given. Lyapunov showed that this LMI is feasible

if and only the matrix A is stable, i.e., all trajectories of ˙ x = Ax converge to zero as

t →∞, or equivalently, all eigenvalues of A must have negative real part. To solve this

LMIP, we pick any Q > 0 and solve the Lyapunov equation A

T

P +PA = −Q, which

is nothing but a set of n(n + 1)/2 linear equations for the n(n + 1)/2 scalar variables

in P. This set of linear equations will be solvable and result in P > 0 if and only if

the LMI is feasible. In fact this procedure not only ﬁnds a solution when the LMI is

feasible; it parametrizes all solutions as Q varies over the positive-deﬁnite cone.

2.7.2 The positive-real lemma

Another important example is given by the positive-real (PR) lemma, which yields

a “frequency-domain” interpretation for a certain LMIP, and under some additional

assumptions, a numerical solution procedure via Riccati equations as well. We give a

simpliﬁed discussion here, and refer the reader to the References for more complete

statements.

The LMI considered is:

P > 0,

_

A

T

P +PA PB −C

T

B

T

P −C −D

T

−D

_

≤ 0, (2.35)

where A ∈ R

n×n

, B ∈ R

n×p

, C ∈ R

p×n

, and D ∈ R

p×p

are given, and the matrix

P = P

T

∈ R

n×n

is the variable. (We will encounter a variation on this LMI in §6.3.3.)

Note that if D + D

T

> 0, the LMI (2.35) is equivalent to the quadratic matrix

inequality

A

T

P +PA+ (PB −C

T

)(D +D

T

)

−1

(PB −C

T

)

T

≤ 0. (2.36)

We assume for simplicity that A is stable and the system (A, B, C) is minimal.

The link with system and control theory is given by the following result. The

LMI (2.35) is feasible if and only if the linear system

˙ x = Ax +Bu, y = Cx +Du (2.37)

is passive, i.e.,

_

T

0

u(t)

T

y(t) dt ≥ 0

for all solutions of (2.37) with x(0) = 0.

Passivity can also be expressed in terms of the transfer matrix of the linear sys-

tem (2.37), deﬁned as

H(s)

∆

= C(sI −A)

−1

B +D

for s ∈ C. Passivity is equivalent to the transfer matrix H being positive-real, which

means that

H(s) +H(s)

∗

≥ 0 for all Re s > 0.

This electronic version is for personal use and may not be duplicated or distributed.

26 Chapter 2 Some Standard Problems Involving LMIs

When p = 1 this condition can be checked by various graphical means, e.g., plotting

the curve given by the real and imaginary parts of H(iω) for ω ∈ R (called the Nyquist

plot of the linear system). Thus we have a graphical, frequency-domain condition for

feasibility of the LMI (2.35). This approach was used in much of the work in the 1960s

and 1970s described in §1.2.

With a few further technical assumptions, including D+D

T

> 0, the LMI (2.35)

can be solved by a method based on Riccati equations and Hamiltonian matrices.

With these assumptions the LMI (2.35) is feasible if and only if there exists a real

matrix P = P

T

satisfying the ARE

A

T

P +PA+ (PB −C

T

)(D +D

T

)

−1

(PB −C

T

)

T

= 0, (2.38)

which is just the quadratic matrix inequality (2.36) with equality substituted for in-

equality. Note that P > 0.

To solve the ARE (2.38) we ﬁrst form the associated Hamiltonian matrix

M =

_

A−B(D +D

T

)

−1

C B(D +D

T

)

−1

B

T

−C

T

(D +D

T

)

−1

C −A

T

+C

T

(D +D

T

)

−1

B

T

_

.

Then the system (2.37) is passive, or equivalently, the LMI (2.35) is feasible, if and

only if M has no pure imaginary eigenvalues. This fact can be used to form a Routh–

Hurwitz (Sturm) type test for passivity as a set of polynomial inequalities in the data

A, B, C, and D.

When M has no pure imaginary eigenvalues we can construct a solution P

are

as

follows. Pick V ∈ R

2n×n

so that its range is a basis for the stable eigenspace of

M, e.g., V = [v

1

· · · v

n

] where v

1

, . . . , v

n

are a set of independent eigenvectors of M

associated with its n eigenvalues with negative real part. Partition V as

V =

_

V

1

V

2

_

,

where V

1

and V

2

are square; then set P

are

= V

2

V

−1

1

. The solution P

are

thus obtained

is the minimal element among the set of solutions of (2.38): if P = P

T

satisﬁes (2.38),

then P ≥ P

are

. Much more discussion of this method, including the precise technical

conditions, can be found in the References.

2.7.3 The bounded-real lemma

The same results appear in another important form, the bounded-real lemma. Here

we consider the LMI

P > 0,

_

A

T

P +PA+C

T

C PB +C

T

D

B

T

P +D

T

C D

T

D −I

_

≤ 0. (2.39)

where A ∈ R

n×n

, B ∈ R

n×p

, C ∈ R

p×n

, and D ∈ R

p×p

are given, and the matrix

P = P

T

∈ R

n×n

is the variable. For simplicity we assume that A is stable and

(A, B, C) is minimal.

This LMI is feasible if and only the linear system (2.37) is nonexpansive, i.e.,

_

T

0

y(t)

T

y(t) dt ≤

_

T

0

u(t)

T

u(t) dt

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 27

for all solutions of (2.37) with x(0) = 0, This condition can also be expressed in

terms of the transfer matrix H. Nonexpansivity is equivalent to the transfer matrix

H satisfying the bounded-real condition, i.e.,

H(s)

∗

H(s) ≤ I for all Re s > 0.

This is sometimes expressed as H

∞

≤ 1 where

H

∞

∆

= sup{ H(s) | Re s > 0}

is called the H

∞

norm of the transfer matrix H.

This condition is easily checked graphically, e.g., by plotting H(iω) versus ω ∈ R

(called a singular value plot for p > 1 and a Bode magnitude plot for p = 1).

Once again we can relate the LMI (2.39) to an ARE. With some appropriate tech-

nical conditions (see the Notes and References) including D

T

D < I, the LMI (2.39)

is feasible if and only the ARE

A

T

P +PA+C

T

C + (PB +C

T

D)(I −D

T

D)

−1

(PB +C

T

D)

T

= 0 (2.40)

has a real solution P = P

T

. Once again this is the quadratic matrix inequality

associated with the LMI (2.39), with equality instead of the inequality.

We can solve this equation by forming the Hamiltonian matrix

M =

_

A+B(I −D

T

D)

−1

D

T

C B(I −D

T

D)

−1

B

T

−C

T

(I −DD

T

)

−1

C −A

T

−C

T

D(I −D

T

D)

−1

B

T

_

.

Then the system (2.37) is nonexpansive, or equivalently, the LMI (2.39) is feasible,

if and only if M has no imaginary eigenvalues. In this case we can ﬁnd the minimal

solution to (2.40) as described in §2.7.2.

2.7.4 Others

Several other LMI problems have analytic solutions, e.g., the ones that arise in the

synthesis of state-feedback for linear systems, or estimator gains for observers. As a

simple example consider the LMI

P > 0, AP +PA

T

< BB

T

where A ∈ R

n×n

and B ∈ R

n×p

are given and P is the variable. This LMI is feasible

if and only if the pair (A, B) is stabilizable, i.e., there exists a matrix K ∈ R

p×n

such

that A+BK is stable. There are simple methods for determining whether this is the

case, and when it is, of constructing an appropriate P. We will describe this and other

results in Chapter 7.

Several standard results from linear algebra can also be interpreted as analytic

solutions to certain LMI problems of special form. For example, it is well-known

that among all symmetric positive-deﬁnite matrices with given diagonal elements,

the diagonal matrix has maximum determinant. This gives an analytic solution to a

special (and fairly trivial) CP (see §3.5).

Notes and References

LMIs

The term “Linear Matrix Inequality” was coined by J. C. Willems and is widely used now.

As we mentioned in ¸1.2, the term was used in several papers from the 1970s to refer to the

This electronic version is for personal use and may not be duplicated or distributed.

28 Chapter 2 Some Standard Problems Involving LMIs

speciﬁc LMI (1.3). The term is also consistent with the title of an early paper by Bellman

and Ky Fan: On Systems of Linear Inequalities in Hermitian Matrix Variables [BF63] (see

below).

A more accurate term might be “Aﬃne Matrix Inequality”, since the matrix is an aﬃne

and not linear function of the variable. We think of the term “Linear Matrix Inequality” as

analogous to the term “Linear Inequalities” used to describe a

T

i

x ≤ bi, i = 1, . . . , p.

When m = 1, an LMI is nothing but a matrix pencil. The theory of matrix pencils is very old

and quite complete, and there is an extensive literature (see, e.g., Golub and Van Loan [GL89]

and the references therein). For the case of m = 1 all of our standard problems are readily

solved, e.g., by simultaneous diagonalization of F0 and F1 by an appropriate congruence.

The case m = 1 does not arise on its own in any interesting problem from system or control

theory, but it does arise in the “line-search (sub)problem”, i.e., when we restrict one of our

standard problems to a speciﬁc line. In this context the variable represents the “step length”

to be taken in an iteration of an algorithm, and it useful to know that such problems can be

solved extremely eﬃciently.

Schur complements for nonstrict inequalities

The Schur complement result of section ¸2.1 can be generalized to nonstrict inequalities as

follows. Suppose Q and R are symmetric. The condition

_

Q S

S

T

R

_

≥ 0 (2.41)

is equivalent to

R ≥ 0, Q−SR

†

S

T

≥ 0, S(I −RR

†

) = 0,

where R

†

denotes the Moore–Penrose inverse of R.

To see this, let U be an orthogonal matrix that diagonalizes R, so that

U

T

RU =

_

Σ 0

0 0

_

,

where Σ > 0 and diagonal. Inequality (2.41) holds if and only if

_

I 0

0 U

T

__

Q S

S

T

R

__

I 0

0 U

_

=

_

¸

_

Q S1 S2

S

T

1

Σ 0

S

T

2

0 0

_

¸

_

≥ 0,

where [S1 S2] = SU, with appropriate partitioning. We must then have S2 = 0, which holds

if and only if S(I −RR

†

) = 0, and

_

Q S1

S

T

1

Σ

_

≥ 0,

which holds if and only if Q−SR

†

S

T

≥ 0.

Formulating convex problems in terms of LMIs

The idea that LMIs can be used to represent a wide variety of convex constraints can be

found in Nesterov and Nemirovskii’s report [NN90b] (which is really a preliminary version

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 29

of their book [NN94]) and software manual [NN90a]. They formalize the idea of a “positive-

deﬁnite representable” function; see ¸5.3, ¸5.4, and ¸6.4 of their book [NN94]. The idea

is also discussed in the article [Ali92b] and thesis [Ali91] of Alizadeh. In [BE93], Boyd

and El Ghaoui give a list of quasiconvex functions that can be represented as generalized

eigenvalues of matrices that depend aﬃnely on the variable.

Infeasibility criterion for LMIs

The LMI F(x) > 0 is infeasible means the aﬃne set |F(x) [ x ∈ R

m

¦ does not intersect the

positive-deﬁnite cone. From convex analysis, this is equivalent to the existence of a linear

functional ψ that is positive on the positive-deﬁnite cone and nonpositive on the aﬃne set

of matrices. The linear functionals that are positive on the positive-deﬁnite cone are of the

form ψ(F) = Tr GF, where G ≥ 0 and G ,= 0. From the fact that ψ is nonpositive on the

aﬃne set |F(x)[x ∈ R

m

¦, we can conclude that Tr GFi = 0, i = 1, . . . , m and Tr GF0 ≤ 0.

These are precisely the conditions for infeasibility of an LMI that we mentioned in ¸2.2.1.

For the special case of multiple Lyapunov inequalities, these conditions are given Bellman

and Ky Fan [BF63] and Kamenetskii and Pyatnitskii [KP87a, KP87b].

It is straightforward to derive optimality criteria for the other problems, using convex anal-

ysis. Some general references for convex analysis are the books [Roc70, Roc82] and survey

article [Roc93] by Rockafellar. The recent text [HUL93] gives a good overview of convex

analysis; LMIs are used as examples in several places.

Complexity of convex optimization

The important role of convexity in optimization is fairly widely known, but perhaps not

well enough appreciated, at least outside the former Soviet Union. In a standard (Western)

treatment of optimization, our standard problems LMIP, EVP, GEVP, and CP would be

considered very diﬃcult since they are nondiﬀerentiable and nonlinear. Their convexity

properties, however, make them tractable, both in theory and in practice.

In [Roc93, p194], Rockafellar makes the important point:

One distinguishing idea which dominates many issues in optimization theory is

convexity . . . An important reason is the fact that when a convex function is

minimized over a convex set every locally optimal solution is global. Also, ﬁrst-

order necessary conditions turn out to be suﬃcient. A variety of other properties

conducive to computation and interpretation of solutions ride on convexity as

well. In fact the great watershed in optimization isn’t between linearity and

nonlinearity, but convexity and nonconvexity.

Detailed discussion of the (low) complexity of convex optimization problems can be found

in the books by Gr¨ otschel, Lov´ asz, Schrijver [GLS88], Nemirovskii and Yudin [NY83], and

Vavasis [Vav91].

Complete and detailed worst-case complexity analyses of several algorithms for our standard

problems can be found in Chapters 3, 4, and 6 of Nesterov and Nemirovskii [NN94].

Ellipsoid algorithm

The ellipsoid algorithm was developed by Shor, Nemirovskii, and Yudin in the 1970s [Sho85,

NY83]. It was used by Khachiyan in 1979 to prove that linear programs can be solved in

polynomial-time [Kha79, GL81, GLS88]. Discussion of the ellipsoid method, as well as

several extensions and variations, can be found in [BGT81] and [BB91, ch14]. A detailed

history of its development, including English and Russian references, appears in chapter 3 of

Akg¨ ul [Akg84].

This electronic version is for personal use and may not be duplicated or distributed.

30 Chapter 2 Some Standard Problems Involving LMIs

Optimization problems involving LMIs

One of the earliest papers on LMIs is also, in our opinion, one of the best: On systems of

linear inequalities in Hermitian matrix variables, by Bellman and Ky Fan [BF63]. In this

paper we ﬁnd many results, e.g., a duality theory for the multiple Lyapunov inequality EVP

and a theorem of the alternative for the multiple Lyapunov inequality LMIP. The paper

concentrates on the associated mathematics: there is no discussion of how to solve such

problems, or any potential applications, although they do comment that such inequality

systems arise in Lyapunov theory. Given Bellman’s legendary knowledge of system and

control theory, and especially Lyapunov theory, it is tempting to conjecture that Bellman

was aware of the potential applications. But so far we have found no evidence for this.

Relevant work includes Cullum et al. [CDW75], Craven and Mond [CM81], Polak and

Wardi [PW82], Fletcher [Fle85], Shapiro [Sha85], Friedland et al. [FNO87], Goh and

Teo [GT88], Panier [Pan89], Allwright [All89], Overton and Womersley [Ove88, Ove92,

OW93, OW92], Ringertz [Rin91], Fan and Nekooie [FN92, Fan93] and Hiriart–Urruty and

Ye [HUY92].

LMIPs are solved using various algorithms for convex optimization in Boyd and Yang [BY89],

Pyatnitskii and Skorodinskii [PS83], Kamenetskii and Pyatnitskii [KP87a, KP87b].

A survey of methods for solving problems involving LMIs used by researchers in control theory

can be found in the paper by Beck [Bec91]. The software packages [BDG

+

91] and [CS92a]

use convex programming to solve many robust control problems. Boyd [Boy94] outlines how

interior-point convex optimization algorithms can be used to build robust control software

tools. Convex optimization problems involving LMIs have been used in control theory since

about 1985: Gilbert’s method [Gil66] was used by Doyle [Doy82] to solve a diagonal scaling

problem; another example is the musol program of Fan and Tits [FT86].

Among other methods used in control to solve problems involving multiple LMIs is the

“method of alternating convex projections” described in the article by Grigoriadis and Skel-

ton [GS92]. This method is essentially a relaxation method, and requires an analytic ex-

pression for the projection onto the feasible set of each LMI. It is not a polynomial-time

algorithm; its complexity depends on the problem data, unlike the ellipsoid method or the

interior-point methods described in [NN94], whose complexities depend on the problem size

only. However, the alternating projections method is reported to work well in many cases.

Interior-point methods for LMI problems

Interior-point methods for various LMI problems have recently been developed by several re-

searchers. The ﬁrst were Nesterov and Nemirovskii [NN88, NN90b, NN90a, NN91a, NN94,

NN93]; others include Alizadeh [Ali92b, Ali91, Ali92a], Jarre [Jar93c], Vandenberghe and

Boyd [VB93b], Rendl, Vanderbei, and Wolkowocz [RVW93], and Yoshise [Yos94].

Of course, general interior-point methods (and the method of centers in particular) have a

long history. Early work includes the book by Fiacco and McCormick [FM68], the method

of centers described by Huard et al. [LH66, Hua67], and Dikin’s interior-point method

for linear programming [Dik67]. Interest in interior-point methods surged in 1984 when

Karmarkar [Kar84] gave his interior-point method for solving linear programs, which appears

to have very good practical performance as well as a good worst-case complexity bound. Since

the publication of Karmarkar’s paper, many researchers have studied interior-point methods

for linear and quadratic programming. These methods are often described in such a way

that extensions to more general (convex) constraints and objectives are not clear. However,

Nesterov and Nemirovskii have developed a theory of interior-point methods that applies to

more general convex programming problems, and in particular, every problem that arises

in this book; see [NN94]. In particular, they derive complexity bounds for many diﬀerent

interior-point algorithms, including the method of centers, Nemirovskii’s projective algorithm

and primal-dual methods. The most eﬃcient algorithms seem to be Nemirovskii’s projective

algorithm and primal-dual methods (for the case of linear programs, see [Meh91]).

Other recent articles that consider interior-point methods for more general convex program-

ming include Sonnevend [Son88], Jarre [Jar91, Jar93b], Kortanek et al. [KPY91], Den

Hertog, Roos, and Terlaky [DRT92], and the survey by Wright [Wri92].

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 31

Interior-point methods for GEVPs are described in Boyd and El Ghaoui [BE93] (variation

on the method of centers), and Nesterov and Nemirovskii [NN91b] and [NN94, ¸4.4] (a

variation on Nemirovskii’s projective algorithm). Since GEVPs are not convex problems,

devising a reliable stopping criterion is more challenging than for the convex problems LMIP,

EVP, and CP. A detailed complexity analysis (in particular, a statement and proof of the

polynomial-time complexity of GEVPs) is given in [NN91b, NN94]. See also [Jar93a].

Several researchers have recently studied the possibility of switching from an interior-point

method to a quadratically convergent local method in order to improve on the ﬁnal conver-

gence; see [Ove92, OW93, OW92, FN92, NF92].

Interior-point methods for CPs and other related extremal ellipsoid volume problems can be

found in [NN94, ¸6.5].

Interior-point methods and problem structure

Many researchers have developed algorithms that take advantage of the special structure

of the least-squares problems arising in interior-point methods for linear programming. As

far as we know, the ﬁrst (and so far, only) interior-point algorithm that takes advantage

of the special (Lyapunov) structure of an LMI problem arising in control is described in

Vandenberghe and Boyd [VB93b, VB93a]. Nemirovskii’s projective method can also take

advantage of such structure; these two algorithms appear to be the most eﬃcient algorithms

developed so far for solving the LMIs that arise in control theory.

Software for solving LMI problems

Gahinet and Nemirovskii have recently developed a software package called LMI-Lab [GN93]

based on an earlier FORTRAN code [NN90a], which allows the user to describe an LMI

problem in a high-level symbolic form (not unlike the formulas that appear throughout this

book!). LMI-Lab then solves the problem using Nemirovskii’s projective algorithm, taking

advantage of some of the problem structure (e.g., block structure, diagonal structure of some

of the matrix variables).

Recently, El Ghaoui has developed another software package for solving LMI problems. This

noncommercial package, called LMI-tool, can be used with matlab. It is available via anony-

mous ftp (for more information, send mail to elghaoui@ensta.fr). Another version of LMI-

tool, developed by Nikoukhah and Delebecque, is available for use with the matlab-like

freeware package scilab; in this version, LMI-tool has been interfaced with Nemirovskii’s

projective code. scilab can be obtained via anonymous ftp (for more information, send mail

to Scilab@inria.fr).

A commercial software package that solves a few specialized control system analysis and

design problems via LMI formulation, called optin, was recently developed by Olas and

Associates; see [OS93, Ola94].

Reduction to a strictly feasible LMI

In this section we prove the following statement. Let F0, . . . , Fm ∈ R

n×n

be symmetric

matrices. Then there is a matrix A ∈ R

m×p

with p ≤ m, a vector b ∈ R

m

, and symmetric

matrices

˜

F0, . . . ,

˜

Fp ∈ R

q×q

with q ≤ n such that:

F(x) ≥ 0 if and only if x = Az +b for some z ∈ R

p

and

˜

F(z)

∆

=

˜

F0 +

p

i=1

zi

˜

Fi ≥ 0.

In addition, if the LMI F(x) ≥ 0 is feasible, then the LMI

˜

F(z) ≥ 0 is strictly feasible.

This electronic version is for personal use and may not be duplicated or distributed.

32 Chapter 2 Some Standard Problems Involving LMIs

Consider the LMI

F(x) = F0 +

m

i=1

xiFi ≥ 0, (2.42)

where Fi ∈ R

n×n

, i = 0, . . . , m. Let X denote the feasible set |x [ F(x) ≥ 0¦.

If X is empty, there is nothing to prove; we can take p = m, A = I, b = 0,

˜

F0 = F0, . . . ,

˜

Fp =

Fp.

Henceforth we assume that X is nonempty. If X is a singleton, say X = |x0¦, then with

p = 1, A = 0, b = x0,

˜

F0 = 1, and

˜

F1 = 0, the statement follows.

Now consider the case when X is neither empty nor a singleton. Then, there is an aﬃne

subspace / of minimal dimension p ≥ 1 that contains X. Let a1, . . . ap a basis for the linear

part of /. Then every x ∈ / can be written as x = Az +b where A = [a1 ap] is full-rank

and b ∈ R

m

. Deﬁning G0 = F(b), Gi = F(ai)−F0, i = 1, . . . , p, and G(z) = G0 +

p

i=1

ziGi,

we see that F(x) ≥ 0 if and only if there exists z ∈ R

p

satisfying x = Az +b and G(z) ≥ 0.

Let Z

∆

= |z ∈ R

p

[ G(z) ≥ 0¦. By construction, Z has nonempty interior. Let z0 a point in

the interior of Z and let v lie in the nullspace of G(z0). Now, v

T

G(z)v is a nonnegative aﬃne

function of z ∈ Z and is zero at an interior point of Z. Therefore, it is identically zero over

Z, and hence over R

p

. Therefore, v belongs to the intersection B of the nullspaces of the

Gi, i = 0, . . . , p. Conversely, any v belonging to B will satisfy v

T

G(z)v = 0 for any z ∈ R

p

.

B may therefore be interpreted as the “constant” nullspace of the LMI (2.42). Let q be the

dimension of B.

If q = n (i.e., G0 = = Gp = 0), then obviously F(x) ≥ 0 if and only if x = Az +b. In this

case, the statement is satisﬁed with

˜

F0 = I ∈ R

p×p

,

˜

Fi = 0, i = 1, . . . , p.

If q < n, let v1, . . . , vq a basis for B. Complete the basis v1, . . . , vq by vq+1, . . . , vn to

obtain a basis of R

n

. Deﬁne U = [vq+1 vn],

˜

Fi = U

T

GiU, i = 1, . . . , p and

˜

F(z) =

˜

F0 +

p

i=1

zi

˜

Fi. Then, the LMI G(z) ≥ 0 is equivalent to the LMI

˜

F(z) ≥ 0, and by

construction, there exists z0 such that

˜

F(z0) > 0. This concludes the proof.

Elimination procedure for matrix variables

We now prove the matrix elimination result stated in ¸2.6.2: Given G, U, V , there exists X

such that

G+UXV

T

+V X

T

U

T

> 0 (2.43)

if and only if

˜

U

T

G

˜

U > 0,

˜

V

T

G

˜

V > 0 (2.44)

holds, where

˜

U and

˜

V are orthogonal complements of U and V respectively.

It is obvious that if (2.43) holds for some X, so do inequalities (2.44). Let us now prove the

converse.

Suppose that inequalities (2.44) are feasible. Suppose now that inequality (2.43) is not feasible

for any X. In other words, suppose that

G+UXV

T

+V X

T

U

T

,> 0,

for every X. By duality, this is equivalent to the condition

there exists Z ,= 0 with Z ≥ 0, V

T

ZU = 0, and Tr GZ ≤ 0. (2.45)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 33

Now let us show that Z ≥ 0 and V

T

ZU = 0 imply that

Z =

˜

V HH

T

˜

V

T

+

˜

UKK

T

˜

U

T

(2.46)

for some matrices H, K (at least one of which is nonzero, since otherwise Z = 0). This

will ﬁnish our proof, since (2.44) implies that Tr GZ > 0 for Z of the form (2.46), which

contradicts Tr GZ ≤ 0 in (2.45).

Let R be a Cholesky factor of Z, i.e. Z = RR

T

. The condition V

T

ZU = 0 is equivalent to

V

T

RR

T

U = (V

T

R)(U

T

R)

T

= 0.

This means that there exists a unitary matrix T, and matrices M and N such that

V

T

RT = [0 M] and U

T

RT = [N 0] ,

where the number of columns of M and N add up to that of R. In other words, RT can be

written as

RT = [A B]

for some matrices A, B such that V

T

A = 0, U

T

B = 0. From the deﬁnition of

˜

V and

˜

U,

matrices A, B can be written A =

˜

V H, B =

˜

UK for some matrices H, K, and we have

Z = RR

T

= (RT)(RT)

T

= AA

T

+BB

T

=

˜

V HH

T

˜

V

T

+

˜

UKK

T

˜

U

T

,

which is the desired result.

The equivalence between the conditions

˜

U(z)

T

G(z)

˜

U(z) > 0

and

G(z) −σU(z)U(z)

T

> 0 for some real scalar σ

is through Finsler’s lemma [Fin37] (see also ¸3.2.6 of [Sch73]), which states that if x

T

Qx > 0

for all nonzero x such that x

T

Ax = 0, where Q and A are symmetric, real matrices, then

there exists a real scalar such that Q−σA > 0.

Finsler’s lemma has also been directly used to eliminate variables in certain matrix inequali-

ties (see for example [PH86, KR88, BPG89b]); it is closely related to the S-procedure. We

also note another result on elimination of matrix variables, due to Parrott [Par78].

The elimination lemma is related to a matrix dilation problem considered in [DKW82], which

was used in control problems in [PZPB91, PZP

+

92, Gah92, Pac94, IS93a, Iwa93] (see

also the Notes and References of Chapter 7).

The S-procedure

The problem of determining if a quadratic form is nonnegative when other quadratic forms are

nonnegative has been studied by mathematicians for at least seventy years. For a complete

discussion and references, we refer the reader to the survey article by Uhlig [Uhl79]. See also

the book by Hestenes [Hes81, p354-360] and Horn and Johnson [HJ91, p78-86] for proofs

of various S-procedure results.

The application of the S-procedure to control problems dates back to 1944, when Lur’e

and Postnikov used it to prove the stability of some particular nonlinear systems. The

name “S-procedure” was introduced much later by Aizerman and Gantmacher in their 1964

This electronic version is for personal use and may not be duplicated or distributed.

34 Chapter 2 Some Standard Problems Involving LMIs

book [AG64]; since then Yakubovich has given more general formulations and corrected

errors in some earlier proofs [Yak77, FY79].

The proof of the two S-procedure results described in this chapter can be found in the

articles by Yakubovich [FY79, Yak77, Yak73]. Recent and important developments about

the S-procedure can be found in [Yak92] and references therein.

Complexity of S-procedure condition

Suppose that T0, u0, and v0 depend aﬃnely on some variable ν; the condition (2.30) is

convex in ν. Indeed, for ﬁxed ζ, the constraint F0(ζ) ≥ 0 is a linear constraint on ν and the

constraint (2.30) is simply an inﬁnite number of these linear constraints. We might therefore

imagine that the problem of checking whether there exists ν such that (2.30) holds has low

complexity since it can be cast as a convex feasibility problem. This is wrong. In fact, merely

verifying that the condition (2.30) holds for ﬁxed data Ti, ui, and vi, is as hard as solving

a general indeﬁnite quadratic program, which is NP-complete. We could determine whether

there exists ν such that (2.30) holds with polynomially many steps of the ellipsoid algorithm;

the problem is that ﬁnding a cutting plane (which is required for each step of the ellipsoid

algorithm) is itself an NP-complete problem.

Positive-real and bounded-real lemma

In 1961, Popov gave the famous Popov frequency-domain stability criterion for the absolute

stability problem [Pop62]. Popov’s criterion could be checked via graphical means, by verify-

ing that the Nyquist plot of the “linear part” of the nonlinear system was conﬁned to a speciﬁc

region in the complex plane. Yakubovich [Yak62, Yak64] and Kalman [Kal63a, Kal63b]

established the connection between the Popov criterion and the existence of a positive-

deﬁnite matrix satisfying certain matrix inequalities. The PR lemma is also known by various

names such as the Yakubovich–Kalman–Popov–Anderson Lemma or the Kalman–Yakubovich

Lemma. The PR lemma is now standard material, described in several books on control and

systems theory, e.g., Narendra and Taylor [NT73], Vidyasagar [Vid92, pp474–478], Faurre

and Depeyrot [FD77], Anderson and Vongpanitlerd [AV73, ch5–7], Brockett [Bro70], and

Willems [Wil70].

In its original form, the PR Lemma states that the transfer function c(sI − A)

−1

b of the

single-input single-output minimal system (A, b, c) is positive-real, i.e.

Re c(sI −A)

−1

b ≥ 0 for all Re s > 0 (2.47)

if and only if there exists P > 0 such that A

T

P +PA ≤ 0 and Pb = c

T

. The condition (2.47)

can be checked graphically.

Anderson [And67, And66b, And73] extended the PR lemma to multi-input multi-output

systems, and derived similar results for nonexpansive systems [And66a]. Willems [Wil71b,

Wil74a] described connections between the PR lemma, certain quadratic optimal control

problems and the existence of symmetric solutions to the ARE; it is in [Wil71b] that we ﬁnd

Willems’ quote on the role of LMIs in quadratic optimal control (see page 3).

The connections between passivity, LMIs and AREs can be summarized as follows. We

consider A, B, C, and D such that all eigenvalues of A have negative real part, (A, B) is

controllable and D +D

T

> 0. The following statements are equivalent:

1. The system

˙ x = Ax +Bu, y = Cx +Du, x(0) = 0

is passive, i.e., satisﬁes

_

T

0

u(t)

T

y(t)dt ≥ 0

for all u and T ≥ 0.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 35

2. The transfer matrix H(s) = C(sI −A)

−1

B +D is positive-real, i.e.,

H(s) +H(s)

∗

≥ 0

for all s with Re s ≥ 0.

3. The LMI

_

A

T

P +PA PB −C

T

B

T

P −C −(D +D

T

)

_

≤ 0 (2.48)

in the variable P = P

T

is feasible.

4. There exists P = P

T

satisfying the ARE

A

T

P +PA+ (PB −C

T

)(D +D

T

)

−1

(PB −C

T

)

T

= 0. (2.49)

5. The sizes of the Jordan blocks corresponding to the pure imaginary eigenvalues of the

Hamiltonian matrix

M =

_

A−B(D +D

T

)

−1

C B(D +D

T

)

−1

B

T

−C

T

(D +D

T

)

−1

C −A

T

+C

T

(D +D

T

)

−1

B

T

_

are all even.

The equivalence of 1, 2, 3, and 4 can be found in Theorem 4 of [Wil71b] and also [Wil74a].

The equivalence of 5 is found in, for example, Theorem 1 of [LR80]. Origins of this result can

be traced back to Reid [Rei46] and Levin [Lev59]; it is explicitly stated in Potter [Pot66].

An excellent discussion of this result and its connections with spectral factorization theory

can be found in [AV73]; see also [Fra87, BBK89, Rob89]. Connections between the ex-

tremal points of the set of solutions of the LMI (2.48) and the solutions of the ARE (2.49)

are explored in [Bad82]. The method for constructing Pare from the stable eigenspace of

M, described in ¸2.7.2, is in [BBK89]; this method is an obvious variation of Laub’s algo-

rithm [Lau79, AL84], or Van Dooren’s [Doo81]. A less stable numerical method can be found

in [Rei46, Lev59, Pot66, AV73]. For other discussions of strict passivity, see the articles

by Wen [Wen88], Lozano–Leal and Joshi [LLJ90]. It is also possible to check passivity (or

nonexpansivity) of a system symbolically using Sturm methods, which yields Routh–Hurwitz

like algorithms; see Siljak [Sil71, Sil73] and Boyd, Balakrishnan, and Kabamba [BBK89].

The PR lemma is used in many areas, e.g., interconnected systems [MH78] and stochastic

processes and stochastic realization theory [Fau67, Fau73, AM74].

This electronic version is for personal use and may not be duplicated or distributed.

Chapter 3

Some Matrix Problems

3.1 Minimizing Condition Number by Scaling

The condition number of a matrix M ∈ R

p×q

, with p ≥ q, is the ratio of its largest

and smallest singular values, i.e.,

κ(M)

∆

=

_

λ

max

(M

T

M)

λ

min

(M

T

M)

_

1/2

for M full-rank, and κ(M) = ∞otherwise. For square invertible matrices this reduces

to κ(M) = MM

−1

.

We consider the problem:

minimize κ(LMR)

subject to L ∈ R

p×p

, diagonal and nonsingular

R ∈ R

q×q

, diagonal and nonsingular

(3.1)

where L and R are the optimization variables, and the matrix M ∈ R

p×q

is given.

We will show that this problem can be transformed into a GEVP. We assume without

loss of generality that p ≥ q and M is full-rank.

Let us ﬁx γ > 1. There exist nonsingular, diagonal L ∈ R

p×p

and R ∈ R

q×q

such that κ(LMR) ≤ γ if and only if there are nonsingular, diagonal L ∈ R

p×p

and

R ∈ R

q×q

and µ > 0 such that

µI ≤ (LMR)

T

(LMR) ≤ µγ

2

I.

Since we can absorb the factor 1/

√

µ into L, this is equivalent to the existence of

nonsingular, diagonal L ∈ R

p×p

and R ∈ R

q×q

such that

I ≤ (LMR)

T

(LMR) ≤ γ

2

I,

which is the same as

(RR

T

)

−1

≤ M

T

(L

T

L)M ≤ γ

2

(RR

T

)

−1

. (3.2)

This is equivalent to the existence of diagonal P ∈ R

p×p

, Q ∈ R

q×q

with P > 0,

Q > 0 and

Q ≤ M

T

PM ≤ γ

2

Q. (3.3)

37

38 Chapter 3 Some Matrix Problems

To see this, ﬁrst suppose that L ∈ R

p×p

and R ∈ R

q×q

are nonsingular and diagonal,

and (3.2) holds. Then (3.3) holds with P = L

T

L and Q = (RR

T

)

−1

. Conversely,

suppose that (3.3) holds for diagonal P ∈ R

p×p

and Q ∈ R

q×q

with P > 0, Q > 0.

Then, (3.2) holds for L = P

1/2

and R = Q

−1/2

.

Hence we can solve (3.1) by solving the GEVP:

minimize γ

2

subject to P ∈ R

p×p

and diagonal, P > 0,

Q ∈ R

q×q

and diagonal, Q > 0

Q ≤ M

T

PM ≤ γ

2

Q

Remark: This result is readily extended to handle scaling matrices that have

a given block-diagonal structure, and more generally, the constraint that one or

more blocks are equal. Complex matrices are readily handled by substituting M

∗

(complex-conjugate transpose) for M

T

.

3.2 Minimizing Condition Number of a

Positive-Deﬁnite Matrix

A related problem is minimizing the condition number of a positive-deﬁnite matrix

M that depends aﬃnely on the variable x, subject to the LMI constraint F(x) > 0.

This problem can be reformulated as the GEVP

minimize γ

subject to F(x) > 0, µ > 0, µI < M(x) < γµI

(3.4)

(the variables here are x, µ, and γ).

We can reformulate this GEVP as an EVP as follows. Suppose

M(x) = M

0

+

m

i=1

x

i

M

i

, F(x) = F

0

+

m

i=1

x

i

F

i

.

Deﬁning the new variables ν = 1/µ, ˜ x = x/µ, we can express (3.4) as the EVP with

variables ν = 1/µ, ˜ x = x/µ and γ:

minimize γ

subject to νF

0

+

m

i=1

˜ x

i

F

i

> 0, I < νM

0

+

m

i=1

˜ x

i

M

i

< γI

3.3 Minimizing Norm by Scaling

The optimal diagonally scaled norm of a matrix M ∈ C

n×n

is deﬁned as

ν(M)

∆

= inf

__

_

DMD

−1

_

_

¸

¸

D ∈ C

n×n

is diagonal and nonsingular

_

,

where M is the norm (i.e., largest singular value) of the matrix M. We will show

that ν(M) can be computed by solving a GEVP.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

3.4 Rescaling a Matrix Positive-Definite 39

Note that

ν(M) = inf

_

γ

¸

¸

¸

¸

¸

_

DMD

−1

_

∗

_

DMD

−1

_

< γ

2

I

for some diagonal, nonsingular D

_

= inf

_

γ

¸

¸

M

∗

D

∗

DM ≤ γ

2

D

∗

D for some diagonal, nonsingular D

_

= inf

_

γ

¸

¸

M

∗

PM ≤ γ

2

P for some diagonal P = P

T

> 0

_

.

Therefore ν(M) is the optimal value of the GEVP

minimize γ

subject to P > 0 and diagonal, M

∗

PM < γ

2

P

Remark: This result can be extended in many ways. For example, the more

general case of block-diagonal similarity-scaling, with equality constraints among

the blocks, is readily handled. As another example, we can solve the problem of

simultaneous optimal diagonal scaling of several matrices M1, . . . , Mp: To com-

pute

inf

_

max

i = 1, . . . , p

_

_

DMiD

−1

_

_

¸

¸

¸

¸

¸

D ∈ C

n×n

is diagonal

and nonsingular

_

,

we simply compute the optimal value of the GEVP

minimize γ

subject to P > 0 and diagonal, M

∗

i

PMi < γ

2

P, i = 1, . . . , p

Another closely related quantity, which plays an important role in the stability

analysis of linear systems with uncertain parameters, is

inf

_

γ

¸

¸

¸

¸

¸

P > 0, M

∗

PM +i(M

∗

G−GM) < γ

2

P,

P and G diagonal and real

_

, (3.5)

which can also be computed by solving a GEVP.

3.4 Rescaling a Matrix Positive-Deﬁnite

We are given a matrix M ∈ C

n×n

, and ask whether there is a diagonal matrix D > 0

such that the Hermitian part of DM is positive-deﬁnite. This is true if and only if

the LMI

M

∗

D +DM > 0, D > 0 (3.6)

is feasible. So determining whether a matrix can be rescaled positive-deﬁnite is an

LMIP.

When this LMI is feasible, we can ﬁnd a feasible scaling D with the smallest

condition number by solving the EVP

minimize µ

subject to D diagonal, M

∗

D +DM > 0, I < D < µ

2

I

This electronic version is for personal use and may not be duplicated or distributed.

40 Chapter 3 Some Matrix Problems

Remark: The LMI (3.6) can also be interpreted as the condition for the existence

of a diagonal quadratic Lyapunov function that proves stability of −M; see the

Notes and References and ¸10.3.

3.5 Matrix Completion Problems

In a matrix completion problem some entries of a matrix are ﬁxed and the others are

to be chosen so the resulting (“completed”) matrix has some property, e.g., is Toeplitz

and positive-deﬁnite. Many of these problems can be cast as LMIPs. The simplest

example is positive-deﬁnite completion: Some entries of a symmetric matrix are given,

and we are to complete the matrix to make it positive-deﬁnite. Evidently this is an

LMIP.

We can easily recover some known results about this problem. We can assume that

the diagonal entries are speciﬁed (otherwise we simply make them very large, i.e., elim-

inate them; see §2.6.1). Then the set of all positive-deﬁnite completions, if nonempty,

is bounded. Hence the associated LMI has an analytic center A

(see §2.4.1). The

matrix A

is characterized by Tr A

−1

B = 0 for every B = B

T

that has zero entries

where A has speciﬁed entries (see §2.4.2). Thus the matrix A

−1

has a zero entry

in every location corresponding to an unspeciﬁed entry in the original matrix, i.e.,

it has the same sparsity structure as the original matrix. Depending on the pattern

of speciﬁed entries, this condition can lead to an “analytic solution” of the problem.

But of course an arbitrary positive-deﬁnite completion problem is readily solved as an

LMIP.

Let us mention a few simple generalizations of this matrix completion problem

that have not been considered in the literature but are readily solved as LMIPs, and

might be useful in applications.

Suppose that for each entry of the matrix we specify an upper and lower bound,

in other words, we consider an “interval matrix”. For entries that are completely

speciﬁed, the lower and upper bound coincide; for entries that are completely free

the lower bound is −∞ and the upper bound is +∞. The problem of completing

the matrix, subject to the bounds on each entry, to make it positive-deﬁnite (or a

contraction, i.e., have norm less than one) is readily formulated as an LMIP.

As another example, suppose we are given a set of matrices that have some given

sparsity pattern. We want to determine a matrix with the complementary sparsity

pattern that, when added to each of the given matrices, results in a positive-deﬁnite

(or contractive, etc.) matrix. We might call this the “simultaneous matrix completion

problem” since we seek a completion that serves for multiple original matrices. This

problem is also readily cast as an LMIP.

As a more speciﬁc example, suppose we are given upper and lower bounds for

some of the correlation coeﬃcients of a stationary time-series, and must determine

some complementary correlation coeﬃcients that are consistent with them. In other

words, we have a Toeplitz matrix with some entries within given intervals, while the

others are completely free (unspeciﬁed). We must determine these free entries so that

the Toeplitz matrix is positive-deﬁnite, for any choice of the speciﬁed elements in the

intervals. This is an LMIP.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

3.6 Quadratic Approximation of a Polytopic Norm 41

3.6 Quadratic Approximation of a Polytopic Norm

Consider a piecewise linear norm ·

pl

on R

n

deﬁned by

z

pl

∆

= max

i=1,...,p

¸

¸

a

T

i

z

¸

¸

,

where a

i

∈ R

n

for i = 1, . . . , p.

For any P > 0, the quadratic norm deﬁned by z

P

∆

=

√

z

T

Pz = P

1/2

z satisﬁes

1/

√

αz

P

≤ z

pl

≤

√

αz

P

for all z,

for some constant α ≥ 1. Thus, the quadratic norm

√

z

T

Pz approximates z

pl

within

a factor of α. We will show that the problem of ﬁnding P that minimizes α, i.e., the

problem of determining the optimal quadratic norm approximation of ·

pl

, is an

EVP.

Let v

1

, . . . , v

L

be the vertices of the unit ball B

pl

of ·

pl

, so that

B

pl

∆

= {z | z

pl

≤ 1} = Co{v

1

, . . . , v

L

}

and let B

P

denote the unit ball of ·

P

. Then

1/

√

αz

P

≤ z

pl

≤

√

αz

P

(3.7)

is equivalent to

1/

√

αB

P

⊆ B

pl

⊆

√

αB

P

.

The ﬁrst inclusion, 1/

√

αB

P

⊆ B

pl

, is equivalent to

a

T

i

P

−1

a

i

≤ α, i = 1, . . . , p,

and the second, B

pl

⊆

√

αB

P

, is equivalent to

v

T

i

Pv

i

≤ α, i = 1, . . . , L.

Therefore we can minimize α such that (3.7) holds for some P > 0 by solving the EVP

minimize α

subject to v

T

i

Pv

i

≤ α, i = 1, . . . , L,

_

P a

i

a

T

i

α

_

≥ 0, i = 1, . . . , p

Remark: The number of vertices L of the unit ball of | |

pl

can grow exponen-

tially in p and n. Therefore the result is uninteresting from the point of view of

computational complexity and of limited practical use except for problems with

low dimensions.

The optimal factor α in this problem never exceeds

√

n. Indeed, by computing

the maximum volume ellipsoid that lies inside the unit ball of | |

pl

, (which does

not require us to ﬁnd the vertices vi, i = 1, . . . , L), we can ﬁnd (in polynomial-

time) a norm for which α is less than

√

n. See the Notes and References for more

details.

This electronic version is for personal use and may not be duplicated or distributed.

42 Chapter 3 Some Matrix Problems

3.7 Ellipsoidal Approximation

The problem of approximating some subset of R

n

with an ellipsoid arises in many

ﬁelds and has a long history; see the Notes and References. To pose such a problem

precisely we need to know how the subset is described, whether we seek an inner or

outer approximation, and how the approximation will be judged (volume, major or

minor semi-axis, etc.).

In some cases the problem can be cast as a CP or an EVP, and hence solved

exactly. As an example consider the problem of ﬁnding the ellipsoid centered around

the origin of smallest volume that contains a polytope described by its vertices. We

saw in §2.2.4 that this problem can be cast as a CP.

In other cases, we can compute an approximation of the optimal ellipsoid by

solving a CP or an EVP. In some of these cases, the problem is known to be NP-hard,

so it is unlikely that the problem can be reduced (polynomially) to an LMI problem.

It also suggests that the approximations obtained by solving a CP or EVP will not be

good for all instances of problem data, although the approximations may be good on

“typical” problems, and hence of some use in practice.

As an example consider the problem of ﬁnding the ellipsoid of smallest volume

that contains a polytope described by a set of linear inequalities, i.e., { x | a

T

i

x ≤

b

i

, i = 1, . . . , p }. (This is the same problem as described above, but with a diﬀerent

description of the polytope.) This problem is NP-hard. Indeed, consider the problem

of simply verifying that a given, ﬁxed ellipsoid contains a polytope described by a

set of linear inequalities. This problem is equivalent to the general concave quadratic

programming problem, which is NP-complete.

In this section we consider subsets formed from ellipsoids E

1

, . . . , E

p

in various

ways: union, intersection, and addition. We approximate these sets by an ellipsoid E

0

.

We describe an ellipsoid E in two diﬀerent ways. The ﬁrst description uses convex

quadratic functions:

E = {x | T(x) ≤ 0, } , T(x) = x

T

Ax + 2x

T

b +c, (3.8)

where A = A

T

> 0 and b

T

A

−1

b −c > 0 (which ensures that E is nonempty and does

not reduce to a single point). Note that this description is homogeneous, i.e., we can

scale A, b and c by any positive factor without aﬀecting E.

The volume of E is given by

vol(E)

2

= β det

_

(b

T

A

−1

b −c)A

−1

_

where β is a constant that depends only on the dimension of x, i.e., n. The diameter

of E (i.e., twice the maximum semi-axis length) is

2

_

(b

T

A

−1

b −c)λ

max

(A

−1

).

We will also describe an ellipsoid as the image of the unit ball under an aﬃne

mapping with symmetric positive-deﬁnite matrix:

E =

_

x

¸

¸

(x −x

c

)

T

P

−2

(x −x

c

) ≤ 1

_

= { Pz +x

c

| z ≤ 1 } , (3.9)

where P = P

T

> 0. (This representation is unique, i.e., P and x

c

are uniquely

determined by the ellipsoid E.) The volume of E is then proportional to det P, and

its diameter is 2λ

max

(P).

Each of these descriptions of E is readily converted into the other. Given a repre-

sentation (3.8) (i.e., A, b, and c) we form the representation (3.9) with

P =

_

b

T

A

−1

b −c A

−1/2

, x

c

= −A

−1

b.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

3.7 Ellipsoidal Approximation 43

Given the representation (3.9), we can form a representation (3.8) of E with

A = P

−2

, b = −P

−2

x

c

, c = x

T

c

P

−2

x

c

−1.

(This forms one representation of E; we can form every representation of E by scaling

these A, b, c by positive factors.)

3.7.1 Outer approximation of union of ellipsoids

We seek a small ellipsoid E

0

that covers the union of ellipsoids E

1

, . . . , E

p

(or equiva-

lently, the convex hull of the union, i.e., Co

p

i=1

E

i

). We will describe these ellipsoids

via the associated quadratic functions T

i

(x) = x

T

A

i

x + 2x

T

b

i

+c

i

. We have

E

0

⊇

p

_

i=1

E

i

(3.10)

if and only if E

i

⊆ E

0

for i = 1, . . . , p. This is true if and only if, for each i, every x

such that T

i

(x) ≤ 0 satisﬁes T

0

(x) ≤ 0. By the S-procedure, this is true if and only if

there exist nonnegative scalars τ

1

, . . . , τ

p

such that

for every x, T

0

(x) −τ

i

T

i

(x) ≤ 0, i = 1, . . . , p,

or, equivalently, such that

_

A

0

b

0

b

T

0

c

0

_

−τ

i

_

A

i

b

i

b

T

i

c

i

_

≤ 0, i = 1, . . . , p.

Since our representation of E

0

is homogeneous, we will now normalize A

0

, b

0

and c

0

in a convenient way: such that b

T

0

A

−1

0

b

0

−c

0

= 1. In other words we set

c

0

= b

T

0

A

−1

0

b

0

−1 (3.11)

and parametrize E

0

by A

0

and b

0

alone. Thus our condition becomes:

_

A

0

b

0

b

T

0

b

T

0

A

−1

0

b

0

−1

_

−τ

i

_

A

i

b

i

b

T

i

c

i

_

≤ 0, i = 1, . . . , p.

Using Schur complements, we obtain the equivalent LMI

_

¸

_

A

0

b

0

0

b

T

0

−1 b

T

0

0 b

0

−A

0

_

¸

_ −τ

i

_

¸

_

A

i

b

i

0

b

T

i

c

i

0

0 0 0

_

¸

_ ≤ 0, i = 1, . . . , p, (3.12)

with variables A

0

, b

0

, and τ

1

, . . . , τ

p

. To summarize, the condition (3.10) is equivalent

to the existence of nonnegative τ

1

, . . . , τ

p

such that (3.12) holds. (Since A

0

> 0, it

follows that the τ

i

must, in fact, be positive.)

With the normalization (3.11), the volume of E

0

is proportional to

_

det A

−1

0

.

Thus we can ﬁnd the smallest volume ellipsoid containing the union of ellipsoids E

1

,

. . . , E

p

by solving the CP (with variables A

0

, b

0

, and τ

1

, . . . , τ

p

)

minimize log det A

−1

0

subject to A

0

> 0, τ

1

≥ 0, . . . , τ

p

≥ 0, (3.12)

This electronic version is for personal use and may not be duplicated or distributed.

44 Chapter 3 Some Matrix Problems

This ellipsoid is called the L¨ owner–John ellipsoid for

p

i=1

E

i

, and satisﬁes several

nice properties (e.g., aﬃne invariance, certain bounds on how well it approximates

Co

E

i

); see the Notes and References.

We can ﬁnd the ellipsoid (or equivalently, sphere) of smallest diameter that con-

tains

E

i

by minimizing λ

max

(A

−1

0

) subject to the constraints of the previous CP,

which is equivalent to solving the EVP

maximize λ

subject to A

0

> λI, τ

1

≥ 0, . . . , τ

p

≥ 0, (3.12)

The optimal diameter is 2/

√

λ

opt

, where λ

opt

is an optimal value of the EVP above.

3.7.2 Outer approximation of the intersection of ellipsoids

Simply verifying that

E

0

⊇

p

i=1

E

i

(3.13)

holds, given E

0

, E

1

, . . . , E

p

, is NP-complete, so we do not expect to recast it as one

of our LMI problems. A fortiori, we do not expect to cast the problem of ﬁnding the

smallest volume ellipsoid covering the intersection of some ellipsoids as a CP.

We will describe three ways that we can compute suboptimal ellipsoids for this

problem by solving CPs. The ﬁrst method uses the S-procedure to derive an LMI

that is a suﬃcient condition for (3.13) to hold.

Once again we normalize our representation of E

0

so that (3.11) holds. From the

S-procedure we obtain the condition: there exist positive scalars τ

1

, . . . , τ

p

such that

_

A

0

b

0

b

T

0

b

T

0

A

−1

0

b

0

−1

_

−

p

i=1

τ

i

_

A

i

b

i

b

T

i

c

i

_

≤ 0,

which can be written as the LMI (in variables A

0

, b

0

, and τ

1

, . . . , τ

p

):

_

¸

_

A

0

b

0

0

b

T

0

−1 b

T

0

0 b

0

−A

0

_

¸

_ −

p

i=1

τ

i

_

¸

_

A

i

b

i

0

b

T

i

c

i

0

0 0 0

_

¸

_ ≤ 0. (3.14)

This LMI is suﬃcient but not necessary for (3.13) to hold, i.e., it characterizes some

but not all of the ellipsoids that cover the intersection of E

1

, . . . , E

p

. We can ﬁnd the

best such outer approximation (i.e., the one of smallest volume) by solving the CP

(with variables A

0

, b

0

, and τ

1

, . . . , τ

p

)

minimize log det A

−1

0

subject to A

0

> 0, τ

1

≥ 0, . . . , τ

p

≥ 0, (3.14)

(3.15)

Remark: The intersection

p

i=1

ci is empty or a single point if and only if there

exist nonnegative τ1, . . . , τp, not all zero, such that

p

i=1

τi

_

Ai bi

b

T

i

ci

_

≥ 0.

In this case (3.15) is unbounded below.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

3.7 Ellipsoidal Approximation 45

Another method that can be used to ﬁnd a suboptimal solution for the problem of

determining the minimum volume ellipsoid that contains the intersection of E

1

, . . . , E

p

is to ﬁrst compute the maximum volume ellipsoid that is contained in the intersection,

which can be cast as a CP (see the next section). If this ellipsoid is scaled by a factor

of n about it center then it is guaranteed to contain the intersection. See the Notes

and References for more discussion.

Yet another method for producing a suboptimal solution is based on the idea of

the analytic center (see §2.4.1). Suppose that E

1

, . . . , E

p

are given in the form (3.9),

with x

1

, . . . , x

p

and P

1

, . . . , P

p

. The LMI

F(x) = diag

__

I P

−1

1

(x −x

1

)

(x −x

1

)

T

P

−1

1

1

_

, . . .

. . . ,

_

I P

−1

p

(x −x

p

)

(x −x

p

)

T

P

−1

p

1

__

> 0

in the variable x has as feasible set the interior of

p

i=1

E

i

. We assume now that the

intersection

p

i=1

E

i

has nonempty interior (i.e., is nonempty and not a single point).

Let x

**denote its analytic center, i.e.,
**

x

= arg min

F(x) > 0

log det F(x)

−1

and let H denote the Hessian of log det F(x)

−1

at x

**. Then it can be shown that
**

E

0

∆

=

_

x

¸

¸

(x −x

)

T

H(x −x

) ≤ 1

_

⊆

p

i=1

E

i

.

In fact we can give a bound on how poorly E

0

approximates the intersection. It can

be shown that

_

x

¸

¸

(x −x

)

T

H(x −x

) ≤ (3p + 1)

2

_

⊇

p

i=1

E

i

.

Thus, when E

0

is enlarged by a factor of 3p + 1, it covers

p

i=1

E

i

. See the Notes and

References for more discussion of this.

3.7.3 Inner approximation of the intersection of ellipsoids

The ellipsoid E

0

is contained in the intersection of the ellipsoids E

1

, . . . , E

p

if and only

if for every x satisfying

(x −x

0

)

T

P

−2

0

(x −x

0

) ≤ 1,

we have

(x −x

i

)

T

P

−2

i

(x −x

i

) ≤ 1, i = 1, . . . , p.

Using the S-procedure, this is equivalent to the existence of nonnegative λ

1

, . . . , λ

p

satisfying

for every x, (x −x

i

)

T

P

−2

i

(x −x

i

)

−λ

i

(x −x

0

)

T

P

−2

0

(x −x

0

) ≤ 1 −λ

i

, i = 1, . . . , p.

(3.16)

This electronic version is for personal use and may not be duplicated or distributed.

46 Chapter 3 Some Matrix Problems

We prove in the Notes and References section that (3.16) is equivalent to

_

¸

_

−P

2

i

x

i

−x

0

P

0

(x

i

−x

0

)

T

λ

i

−1 0

P

0

0 −λ

i

I

_

¸

_ ≤ 0, i = 1, . . . , p. (3.17)

Therefore, we obtain the largest volume ellipsoid contained in the intersection of

the ellipsoids E

1

, . . . , E

p

by solving the CP (with variables P

0

, x

0

, λ

1

, . . . , λ

p

)

minimize log det P

−1

0

subject to P

0

> 0, λ

1

≥ 0, . . . , λ

p

≥ 0, (3.17)

Among the ellipsoids contained in the intersection of E

1

, . . . , E

p

, we can ﬁnd one

that maximizes the smallest semi-axis by solving

minimize λ

subject to λI > P

0

> 0, λ

1

≥ 0, . . . , λ

p

≥ 0, (3.17)

Equivalently, this yields the sphere of largest diameter contained in the intersection.

3.7.4 Outer approximation of the sum of ellipsoids

The sum of the ellipsoids E

1

, . . . , E

p

is deﬁned as

A

∆

= {x

1

+· · · +x

p

| x

1

∈ E

1

, . . . , x

p

∈ E

p

.}

We seek a small ellipsoid E

0

containing A.

The ellipsoid E

0

contains A if and only if for every x

1

, . . . , x

p

such that T

i

(x

i

) ≤ 0,

i = 1, . . . , p, we have

T

0

_

p

i=1

x

i

_

≤ 0.

By application of the S-procedure, this condition is true if

there exist τ

i

≥ 0, i = 1, . . . , p such that

for every x

i

, i = 1, . . . , p, T

0

_

p

i=1

x

i

_

−

p

i=1

τ

i

T

i

(x

i

) ≤ 0.

(3.18)

Once again we normalize our representation of E

0

so that (3.11) holds. Let x

denote the vector made by stacking x

1

, . . . , x

p

, i.e.,

x =

_

¸

¸

_

x

1

.

.

.

x

p

_

¸

¸

_

.

Then both T

0

(

p

i=1

x

i

) and

p

i=1

τ

i

T

i

(x

i

) are quadratic functions of x. Indeed, let

E

i

denote the matrix such that x

i

= E

i

x, and deﬁne

E

0

=

p

i=1

E

i

,

˜

A

i

= E

T

i

A

i

E

i

,

˜

b

i

= E

T

i

b

i

, i = 0, . . . , p.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 47

With this notation, we see that condition (3.18) is equivalent to

there exist τ

i

≥ 0, i = 1, . . . , p such that

_

˜

A

0

˜

b

0

˜

b

T

0

b

T

0

A

−1

0

b

0

−1

_

−

p

i=1

τ

i

_

˜

A

i

˜

b

i

˜

b

T

i

c

i

_

≤ 0.

(3.19)

This condition is readily written as an LMI in variables A

0

, b

0

, τ

1

, . . . , τ

p

using Schur

complements:

_

¸

_

˜

A

0

˜

b

0

0

˜

b

T

0

−1 b

T

0

0 b

0

−A

0

_

¸

_ −

p

i=1

τ

i

_

¸

_

˜

A

i

˜

b

i

0

˜

b

T

i

c

i

0

0 0 0

_

¸

_ ≤ 0. (3.20)

Therefore, we compute the minimum volume ellipsoid, proven to contain the

sum of E

1

, . . . , E

p

via the S-procedure, by minimizing log det A

−1

0

subject to A

0

> 0

and (3.20). This is a CP.

Notes and References

Matrix scaling problems

The problem of scaling a matrix to reduce its condition number is considered in Forsythe

and Straus [FS55] (who describe conditions for the scaling T to be optimal in terms of a

new variable R = T

−∗

T

−1

, just as we do). Other references on the problem of improving

the condition number of a matrix via diagonal or block-diagonal scaling are [Bau63, GV74,

Sha85, Wat91]. We should point out that the results of ¸3.1 are not practically useful in

numerical analysis, since the cost of computing the optimal scalings (via the GEVP) exceeds

the cost of solving the problem for which the scalings are required (e.g., solving a set of linear

equations).

A special case of the problem of minimizing the condition number of a positive-deﬁnite matrix

is considered in [EHM92, EHM93]. Finally, Khachiyan and Kalantari ([KK92]) consider the

problem of diagonally scaling a positive semideﬁnite matrix via interior-point methods and

give a complete complexity analysis for the problem.

The problem of similarity-scaling a matrix to minimize its norm appears often in control

applications (see for example [Doy82, Saf82]). A recent survey article on this topic is by

Packard and Doyle [PD93]. A related quantity is the one-sided multivariable stability margin,

described in [TSC92]; the problem of computing this quantity can be easily transformed

into a GEVP. The quantity (3.5), used in the analysis of linear systems with parameter

uncertainties, was introduced by Fan, Tits and Doyle in [FTD91] (see also [BFBE92]). Some

closely related quantities were considered in [SC93, SL93a, CS92b]; see also Chapter 8 in

this book, where a diﬀerent approach to the same problem and its extensions is considered.

We also mention some other related quantities found in the articles by Kouvaritakis and

Latchman [KL85a, KL85b, LN92], and Rotea and Prasanth [RP93, RP94], which can be

computed via GEVPs.

Diagonal quadratic Lyapunov functions

The problem of rescaling a matrix M with a diagonal matrix D > 0 such that its Her-

mitian part is positive-deﬁnite is considered in the articles by Barker et al. [BBP78] and

Khalil [Kha82]. The problem is the same as determining the existence of diagonal quadratic

Lyapunov function for a given linear system; see [Ger85, Hu87, BY89, Kra91]. In [KB93],

Kaszkurewicz and Bhaya give an excellent survey on the topic of diagonal Lyapunov func-

tions, including many references and examples from mathematical ecology, power systems,

This electronic version is for personal use and may not be duplicated or distributed.

48 Chapter 3 Some Matrix Problems

and digital ﬁlter stability. Diagonal Lyapunov functions arise in the analysis of positive

orthant stability; see ¸10.3.

Matrix completion problems

Matrix completions problems are addressed in the article of Dym and Gohberg [DG81], in

which the authors examine conditions for the existence of completions of band matrices such

that the inverse is also a band matrix. In [GJSW84], their results are extended to a more

general class of partially speciﬁed matrices. In both papers the “sparsity pattern” result

mentioned in ¸3.5 appears. In [BJL89], Barrett et al. provide an algorithm for ﬁnding the

completion maximizing the determinant of the completed matrix (i.e., the analytic center A

**mentioned in ¸3.5). In [HW93], Helton and Woerdeman consider the problem of minimum
**

norm extension of Hankel matrices. In [GFS93], Grigoriadis, Frazho and Skelton consider

the problem of approximating a given symmetric matrix by a Toeplitz matrix and propose a

solution to this problem via convex optimization. See also [OSA93].

Dancis ([Dan92, Dan93] and Johnson and Lunquist [JL92]) study completion problems in

which the rank or inertia of the completed matrix is speciﬁed. As far as we know, such

problems cannot be cast as LMIPs.

A classic paper on the contractive completion problem (i.e., completing a matrix so its norm

is less than one) is Davis, Kahan, and Weinberger in [DKW82], in which an analytic solution

is given for contractive completion problems with a special block form. The result in this

paper is closely related to (and also more general than) the matrix elimination procedure

given in ¸2.6.2. See [Woe90, NW92, BW92] for more references on contractive completion

problems.

Maximum volume ellipsoid contained in a symmetric polytope

In this section, we establish some properties of the maximum volume ellipsoid contained in

a symmetric polytope described as:

1 =

_

z

¸

¸

¸

¸

a

T

i

z

¸

¸

≤ 1, i = 1, . . . , p

_

.

Consider an ellipsoid described by c = |x

¸

¸

x

T

Q

−1

x ≤ 1¦ where Q = Q

T

> 0. Its volume is

proportional to

√

det Q. The condition c ⊆ 1 can be expressed a

T

i

Qai ≤ 1 for i = 1, . . . , p.

Hence we obtain the maximum volume ellipsoid contained in 1 by solving the CP

minimize log det Q

−1

subject to Q > 0, a

T

i

Qai ≤ 1, i = 1, . . . , p

Suppose now that Q is optimal. From the standard optimality conditions, there are nonneg-

ative λ1, . . . , λp such that

Q

−1

=

p

i=1

λi

a

T

i

Qai

aia

T

i

,

and λi = 0 if a

T

i

Qai = 1. Multiplying by Q

1/2

on the left and right and taking the trace

yields

p

i=1

λi = n.

For any x ∈ R

n

, we have

x

T

Q

−1

x =

p

i=1

λi

(x

T

ai)

2

a

T

i

Qai

.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 49

For each i, either λi = 0 or a

T

i

Qai = 1, so we conclude

x

T

Q

−1

x =

p

i=1

λi(x

T

ai)

2

.

Suppose now that x belongs to 1. Then

x

T

Q

−1

x ≤

p

i=1

λi = n.

Equivalently, 1 ⊆

√

nc, so we have:

c ⊆ 1 ⊆

√

nc,

i.e., the maximum volume ellipsoid contained in 1 approximates 1 within the scale factor

√

n.

Ellipsoidal approximation

The problem of computing the minimum volume ellipsoid in R

n

that contains a given convex

set was considered by John and L¨ owner, who showed that the optimal ellipsoid, when shrunk

about its center by a factor of n, is contained in the given set (see [Joh85, GLS88]). Such

ellipsoids are called L¨ owner–John ellipsoids. In the case of symmetric sets, the factor n can

be improved to

√

n (using an argument similar to the one in the Notes and References section

above). A closely related problem is ﬁnding the maximum volume ellipsoid contained in a

convex set. Similar results hold for these ellipsoids.

Special techniques have been developed to compute smallest spheres and ellipsoids containing

a given set of points in spaces of low dimension; see e.g., Preparata and Shamos [PS85, p.255]

or Post [Pos84].

Minimum volume ellipsoids containing a given set arise often in control theory. Schweppe

and Schlaepfer [Sch68, SS72] use ellipsoids containing the sum and the intersection of two

given ellipsoids in the problem of state estimation with norm-bounded disturbances. See also

the articles by Bertsekas and Rhodes [BR71], Chernousko [Che80a, Che80b, Che80c] and

Kahan [Kah68]. In [RB92], ellipsoidal approximations of robot linkages and workspace are

used to rapidly detect or avoid collisions.

Minimum volume ellipsoids have been widely used in system identiﬁcation; see the ar-

ticles by Fogel [Fog79], Fogel and Huang [FH82], Deller [Del89], Belforte, Bona and

Cerone [BBC90], Lau, Kosut and Boyd [LKB90, KLB90, KLB92], Cheung, Yurkovich

and Passino [CYP93] and the book by Norton [Nor86].

In the ellipsoid algorithm we encountered the minimum volume ellipsoid that contains a given

half-ellipsoid; the references on the ellipsoid algorithm describe various extensions.

Eﬃcient interior-point methods for computing minimum volume outer and maximum volume

inner ellipsoids for various sets are described in Nesterov and Nemirovskii [NN94]. See

also [KT93].

Ellipsoidal approximation via analytic centers

Ellipsoidal approximations via barrier functions arise in the theory of interior-point algo-

rithms, and can be traced back to Dikin [Dik67] and Karmarkar [Kar84] for polytopes

described by a set of linear inequalities. For the case of more general LMIs, see Nesterov and

Nemirovskii [NN94], Boyd and El Ghaoui [BE93], and Jarre [Jar93c].

These approximations have the property that they are easy to compute, and come with

provable bounds on how suboptimal they are. These bounds have the following form: if

This electronic version is for personal use and may not be duplicated or distributed.

50 Chapter 3 Some Matrix Problems

the inner ellipsoid is stretched about its center by some factor α (which depends only on

the dimension of the space, the type and size of the constraints) then the resulting ellipsoid

covers the feasible set. The papers cited above give diﬀerent values of α; one may be better

than another, depending on the number and type of constraints.

In contrast, the maximum volume inner ellipsoid, i.e., the L¨ owner–John ellipsoid, is more dif-

ﬁcult to compute (still polynomial, however) but the factor α depends only on the dimension

of the space and not on the type or number of constraints, and is always smaller than the

ones obtained via barrier functions. So an ellipsoidal approximation from a barrier function

can serve as a “cheap substitute” for the L¨ owner–John ellipsoid.

Proof of lemma in §3.7.3

We prove that the statement

for every x, (x −x1)

T

P

−2

1

(x −x1) −λ(x −x0)

T

P

−2

0

(x −x0) ≤ 1 −λ (3.21)

where P0, P1 are positive matrices, λ ≥ 0, is equivalent to the matrix inequality

_

¸

_

−P

2

1

x1 −x0 P0

(x1 −x0)

T

λ −1 0

P0 0 −λI

_

¸

_

≤ 0. (3.22)

The maximum over x of

(x −x1)

T

P

−2

1

(x −x1) −λ(x −x0)

T

P

−2

0

(x −x0) (3.23)

is ﬁnite if and only if P

2

0

−λP

2

1

≤ 0 (which implies that λ > 0) and there exists x

∗

satisfying

P

−2

1

(x

∗

−x1) = λP

−2

0

(x

∗

−x0), (3.24)

or, equivalently, satisfying

(P

2

0

/λ −P

2

1

)P

−2

1

(x

∗

−x1) = x1 −x0. (3.25)

In this case, x

∗

is a maximizer of (3.23). Using (3.24), condition (3.21) implies

(x

∗

−x1)

T

P

−2

1

(P

2

1

−P

2

0

/λ)P

−2

1

(x

∗

−x1) ≤ 1 −λ.

Using Schur complements, this last condition is equivalent to

_

λ −1 (x

∗

−x1)

T

P

−2

1

(P

2

0

/λ −P

2

1

)

(P

2

0

/λ −P

2

1

)P

−2

1

(x

∗

−x1) P

2

0

/λ −P

2

1

_

≤ 0.

Using (3.25), this implies

_

λ −1 (x1 −x0)

T

x1 −x0 P

2

0

/λ −P

2

1

_

≤ 0,

which is equivalent to (3.22). Conversely, it is easy to show that (3.22) implies (3.21).

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Chapter 4

Linear Diﬀerential Inclusions

4.1 Diﬀerential Inclusions

A diﬀerential inclusion (DI) is described by:

˙ x ∈ F(x(t), t), x(0) = x

0

, (4.1)

where F is a set-valued function on R

n

×R

+

. Any x : R

+

→R

n

that satisﬁes (4.1) is

called a solution or trajectory of the DI (4.1). Of course, there can be many solutions

of the DI (4.1). Our goal is to establish that various properties are satisﬁed by all

solutions of a given DI. For example, we might show that every trajectory of a given

DI converges to zero as t →∞.

By a standard result called the Relaxation Theorem, we may as well assume

F(x, t) is a convex set for every x and t. The DI given by

˙ x ∈ CoF(x(t), t), x(0) = x

0

(4.2)

is called the relaxed version of the DI (4.1). Since CoF(x(t), t) ⊇ F(x(t), t), every

trajectory of the DI (4.1) is also a trajectory of relaxed DI (4.2). Very roughly speak-

ing, the Relaxation Theorem states that for many purposes the converse is true. (See

the References for precise statements.) As a speciﬁc and simple example, it can be

shown that for every DI we encounter in this book, the reachable or attainable sets of

the DI and its relaxed version coincide, i.e., for every T ≥ 0

{x(T) | x satisﬁes (4.1)} = {x(T) | x satisﬁes (4.2)} .

In fact we will not need the Relaxation Theorem, or rather, we will always get it

“for free”—every result we establish in the next two chapters extends immediately to

the relaxed version of the problem. The reason is that when a quadratic Lyapunov

function is used to establish some property for the DI (4.1), then the same Lyapunov

function establishes the property for the relaxed DI (4.2).

4.1.1 Linear diﬀerential inclusions

A linear diﬀerential inclusion (LDI) is given by

˙ x ∈ Ωx, x(0) = x

0

, (4.3)

where Ω is a subset of R

n×n

. We can interpret the LDI (4.3) as describing a family

of linear time-varying systems. Every trajectory of the LDI satisﬁes

˙ x = A(t)x, x(0) = x

0

, (4.4)

51

52 Chapter 4 Linear Differential Inclusions

for some A : R

+

→ Ω. Conversely, for any A : R

+

→ Ω, the solution of (4.4) is a

trajectory of the LDI (4.3). In the language of control theory, the LDI (4.3) might be

described as an “uncertain time-varying linear system,” with the set Ω describing the

“uncertainty” in the matrix A(t).

4.1.2 A generalization to systems

We will encounter a generalization of the LDI described above to linear systems with

inputs and outputs. We will consider a system described by

˙ x = A(t)x +B

u

(t)u +B

w

(t)w, x(0) = x

0

,

z = C

z

(t)x +D

zu

(t)u +D

zw

(t)w

(4.5)

where x : R

+

→ R

n

, u : R

+

→ R

nu

, w : R

+

→ R

nw

, z : R

+

→ R

nz

. x is referred

to as the state, u is the control input, w is the exogenous input and z is the output.

The matrices in (4.5) satisfy

_

A(t) B

u

(t) B

w

(t)

C

z

(t) D

zu

(t) D

zw

(t)

_

∈ Ω. (4.6)

for all t ≥ 0, where Ω ⊆ R

(n+nz)×(n+nu+nw)

. We will be more speciﬁc about the form

of Ω shortly.

In some applications we can have one or more of the integers n

u

, n

w

, and n

z

equal

to zero, which means that the corresponding variable is not used. For example, the

LDI ˙ x ∈ Ωx results when n

u

= n

w

= n

z

= 0. In order not to introduce another term

to describe the set of all solutions of (4.5) and (4.6), we will call (4.5) and (4.6) a

system described by LDIs or simply, an LDI.

4.2 Some Speciﬁc LDIs

We now describe some speciﬁc families of LDIs that we will encounter in the next two

chapters.

4.2.1 Linear time-invariant systems

When Ω is a singleton, the LDI reduces to the linear time-invariant (LTI) system

˙ x = Ax +B

u

u +B

w

w, x(0) = x

0

,

z = C

z

x +D

zu

u +D

zw

w,

where

Ω =

__

A B

u

B

w

C

z

D

zu

D

zw

__

. (4.7)

Although most of the results of Chapters 5–7 are well-known for LTI systems, some

are new; we will discuss these in detail when we encounter them.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

4.2 Some Specific LDIs 53

4.2.2 Polytopic LDIs

When Ω is a polytope, we will call the LDI a polytopic LDI or PLDI. Most of our

results require that Ω be described by a list of its vertices, i.e., in the form

Co

__

A

1

B

u,1

B

w,1

C

z,1

D

zu,1

D

zw,1

_

, . . . ,

_

A

L

B

u,L

B

w,L

C

z,L

D

zu,L

D

zw,L

__

. (4.8)

where the matrices (4.8) are given.

If instead Ω is described by a set of l linear inequalities, then the number of

vertices, i.e., L, will generally increase very rapidly (exponentially) with l. Therefore

results for PLDIs that require the description (4.8) are of limited interest for problems

in which Ω is described in terms of linear inequalities.

4.2.3 Norm-bound LDIs

Another special class of LDIs is the class of norm-bound LDIs (NLDIs), described by

˙ x = Ax +B

p

p +B

u

u +B

w

w, x(0) = x

0

,

q = C

q

x +D

qp

p +D

qu

u +D

qw

w,

z = C

z

x +D

zp

p +D

zu

u +D

zw

w,

p = ∆(t)q, ∆(t) ≤ 1.

(4.9)

where ∆ : R

+

→ R

np×nq

, with ∆(t) ≤ 1 for all t. We will often rewrite the

condition p = ∆(t)q, ∆ ≤ 1 in the equivalent form p

T

p ≤ q

T

q.

For the NLDI (4.9) the set Ω has the form

Ω =

_

˜

A+

˜

B∆(I −D

qp

∆)

−1

˜

C

¸

¸

¸ ∆ ≤ 1

_

,

where

˜

A =

_

A B

u

B

w

C

z

D

zu

D

zw

_

,

˜

B =

_

B

p

D

zp

_

,

˜

C =

_

C

q

D

qu

D

qw

_

.(4.10)

Thus, Ω is the image of the (matrix norm) unit ball under a (matrix) linear-fractional

mapping. Note that Ω is convex. The set Ω is well deﬁned if and only if D

T

qp

D

qp

< I,

in which case we say that the NLDI is well-posed. In the sequel we will consider only

well-posed NLDIs.

The equations (4.9) can be interpreted as an LTI system with inputs u, w, output

z, and a time-varying feedback matrix ∆(t) connected between q and p.

Remark: In the sequel, we often drop some terms in (4.9): Dqw and Dzp will

always be zero, and very often Dqp, Dzw, Dzu and Dqu will be assumed zero.

These assumptions greatly simplify the formulas and treatment. It is not hard to

work out the corresponding formulas and results for the more general case.

This electronic version is for personal use and may not be duplicated or distributed.

54 Chapter 4 Linear Differential Inclusions

4.2.4 Diagonal norm-bound LDIs

A useful variation on the NLDI is the diagonal norm-bound LDI (DNLDI), in which

we further restrict the matrix ∆ to be diagonal. It is given by

˙ x = Ax +B

p

p +B

u

u +B

w

w, x(0) = x

0

,

q = C

q

x +D

qp

p +D

qu

u +D

qw

w,

z = C

z

x +D

zp

p +D

zu

u +D

zw

w,

p

i

= δ

i

(t)q

i

, |δ

i

(t)| ≤ 1, i = 1, . . . , n

q

.

(4.11)

We will often express the condition p

i

= δ

i

(t)q

i

, |δ

i

(t)| ≤ 1 in the simpler form

|p

i

| ≤ |q

i

|. Here we have the componentwise inequalities |p

i

| ≤ |q

i

| instead of the single

inequality p ≤ q appearing in the NLDI. In the language of control theory, we

describe a DNLDI as a linear system with n

q

scalar, uncertain, time-varying feedback

gains, each of which is bounded by one. The DNLDI is described by the set

Ω =

_

˜

A+

˜

B∆(I −D

qp

∆)

−1

˜

C

¸

¸

¸ ∆ ≤ 1, ∆ diagonal

_

, (4.12)

where

˜

A,

˜

B and

˜

C are given by (4.10).

The condition of well-posedness for a DNLDI is that I −D

qp

∆ should be invertible

for all diagonal ∆ with ∆ ≤ 1. This is equivalent to det(I − D

qp

∆) > 0 when ∆

takes on its 2

nq

extreme values, i.e., |∆

ii

| = 1. Although this condition is explicit, the

number of inequalities grows exponentially with n

q

. In fact, determining whether a

DNLDI is well-posed is NP-hard; see the Notes and References.

It turns out that for a DNLDI, CoΩ is a polytope. In fact,

CoΩ = Co

_

˜

A+

˜

B∆(I −D

qp

∆)

−1

˜

C

¸

¸

¸ ∆ diagonal, |∆

ii

| = 1

_

,

i.e., the vertices of CoΩ are among 2

nq

images of the extreme points of ∆ under the

matrix linear-fractional mapping. Therefore a DNLDI can be described as a PLDI, at

least in principle. But there is a very important diﬀerence between the two descriptions

of the same LDI: the number of vertices required to give the PLDI representation of

an LDI increases exponentially with n

q

(see the Notes and References of Chapter 5

for more discussion).

Remark: More elaborate variations on the NLDI are sometimes useful. Doyle

introduced the idea of “structured feedback,” in which the matrix ∆ is restricted

to have a given block-diagonal form, and in addition, there can be equality con-

straints among some of the blocks. When we have a single block, we have an

NLDI; when it is diagonal (with no equality constraints), we have a DNLDI.

Most of the results we present for NLDIs and DNLDIs can be extended to these

more general types of LDIs.

4.3 Nonlinear System Analysis via LDIs

Much of our motivation for studying LDIs comes from the fact that we can use them

to establish various properties of nonlinear, time-varying systems using a technique

known as “global linearization”. The idea is implicit in the work on absolute stability

originating in the Soviet Union in the 1940s.

Consider the system

˙ x = f(x, u, w, t), z = g(x, u, w, t). (4.13)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

4.3 Nonlinear System Analysis via LDIs 55

Suppose that for each x, u, w, and t there is a matrix G(x, u, w, t) ∈ Ω such that

_

f(x, u, w, t)

g(x, u, w, t)

_

= G(x, u, w, t)

_

¸

_

x

u

w

_

¸

_ (4.14)

where Ω ⊆ R

(n+nz)×(n+nu+nw)

. Then of course every trajectory of the nonlinear

system (4.13) is also a trajectory of the LDI deﬁned by Ω. If we can prove that every

trajectory of the LDI deﬁned by Ω has some property (e.g., converges to zero), then

a fortiori we have proved that every trajectory of the nonlinear system (4.13) has this

property.

4.3.1 A derivative condition

Conditions that guarantee the existence of such a G are f(0, 0, 0, 0) = 0, g(0, 0, 0, 0) =

0, and

_

¸

¸

_

∂f

∂x

∂f

∂u

∂f

∂w

∂g

∂x

∂g

∂u

∂g

∂w

_

¸

¸

_

∈ Ω for all x, u, w, t. (4.15)

In fact we can make a stronger statement that links the diﬀerence between a pair

of trajectories of the nonlinear system and the trajectories of the LDI given by Ω.

Suppose we have (4.15) but not necessarily f(0, 0, 0, 0) = 0, g(0, 0, 0, 0) = 0. Then for

any pair of trajectories (x, w, z) and (˜ x, ˜ w, ˜ z) we have

_

˙ x −

˙

˜ x

z − ˜ z

_

∈ CoΩ

_

¸

_

x − ˜ x

u − ˜ u

w − ˜ w

_

¸

_,

i.e., (x − ˜ x, w − ˜ w, z − ˜ z) is a trajectory of the LDI given by CoΩ.

These results follow from a simple extension of the mean-value theorem: if φ :

R

n

→R

n

satisﬁes

∂φ

∂x

∈ Ω

throughout R

n

, then for any x and ˜ x we have

φ(x) −φ(˜ x) ∈ CoΩ(x − ˜ x). (4.16)

To see this, let c ∈ R

n

. By the mean-value theorem we have

c

T

(φ(x) −φ(˜ x)) = c

T

∂φ

∂x

(ζ)(x − ˜ x)

for some ζ that lies on the line segment between x and ˜ x. Since by assumption

∂φ

∂x

(ζ) ∈ Ω,

we conclude that

c

T

(φ(x) −φ(˜ x)) ≤ sup

A ∈ CoΩ

c

T

A(x − ˜ x).

Since c was arbitrary, we see that φ(x) −φ(˜ x) is in every half-space that contains the

convex set CoΩ(x − ˜ x), which proves (4.16).

This electronic version is for personal use and may not be duplicated or distributed.

56 Chapter 4 Linear Differential Inclusions

4.3.2 Sector conditions

Sometimes the nonlinear system (4.13) can be expressed in the form

˙ x = Ax +B

p

p +B

u

u +B

w

w,

q = C

q

x +D

qp

p +D

qu

u +D

qw

w,

z = C

z

x +D

zp

p +D

zu

u +D

zw

w,

p

i

= φ

i

(q

i

, t), i = 1, . . . , n

q

,

(4.17)

where φ

i

: R×R

+

→R satisfy the sector conditions

α

i

q

2

≤ qφ

i

(q, t) ≤ β

i

q

2

for all q and t ≥ 0, where α

i

and β

i

are given. In words, the system consists of a linear

part together with n

q

time-varying sector-bounded scalar nonlinearities.

The variables p

i

and q

i

can be eliminated from (4.17) provided the matrix I−D

qp

∆

is nonsingular for all diagonal ∆ with α

i

≤ ∆

ii

≤ β

i

. In this case, we say the system

is well-posed.

Assume now that the system (4.17) is well-posed, and let (4.13) be the equations

obtained by eliminating the variables p and q. Deﬁne

Ω =

_

˜

A+

˜

B∆(I −D

qp

∆)

−1

˜

C

¸

¸

¸ ∆ diagonal, ∆

ii

∈ [α

i

, β

i

]

_

,

where

˜

A,

˜

B and

˜

C are given by (4.10). Then, the condition (4.14) holds.

Notes and References

Diﬀerential inclusions

The books by Aubin and Cellina [AC84], Kisielewicz [Kis91], and the classic text by Filip-

pov [Fil88] cover the theory of diﬀerential inclusions. See also the article by Roxin [Rox65]

and references therein. Background material, e.g., set-valued analysis, can be found in the

book by Aubin and Frankowska [AF90].

The term linear diﬀerential inclusion is a bit misleading, since LDIs do not enjoy any particu-

lar linearity or superposition properties. We use the term only to point out the interpretation

as an uncertain time-varying linear system. Many authors refer to LDIs using some subset

of the phrase “uncertain linear time-varying system”. A more accurate term, suggested to

us and used by A. Filippov and E. Pyatnitskii, is selector-linear diﬀerential inclusion.

PLDIs come up in several articles on control theory, e.g., [MP89, KP87a, KP87b, BY89].

Integral quadratic constraints

The pointwise quadratic constraint p(t)

T

p(t) ≤ q(t)

T

q(t) that occur in NLDIs can be gener-

alized to the integral quadratic constraint

_

t

0

p

T

p dτ ≤

_

t

0

q

T

q dτ; see ¸8.2 and ¸8.3.

Global linearization

The idea of replacing a nonlinear system by a time-varying linear system can be found in Liu

et al. [Liu68, LSL69]. They call this approach global linearization. Of course, approximating

the set of trajectories of a nonlinear system via LDIs can be very conservative, i.e., there are

many trajectories of the LDI that are not trajectories of the nonlinear system. We will see

in Chapter 8 how this conservatism can be reduced in some special cases.

The idea of global linearization is implicit in the early Soviet literature on the absolute

stability problem, e.g., Lur’e and Postnikov [LP44, Lur57] and Popov [Pop73].

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 57

Modeling systems as PLDIs

One of the key issues in robust control theory is how to model or measure plant “uncertainty”

or “variation”. We propose a simple method that should work well in some cases. At the

least, the method is extremely simple and natural.

Suppose we have a real system that is fairly well modeled as a linear system. We collect

many sets of input/output measurements, obtained at diﬀerent times, under diﬀerent op-

erating conditions, or perhaps from diﬀerent instances of the system (e.g., diﬀerent units

from a manufacturing run). It is important that we have data sets from enough plants or

plant conditions to characterize or at least give some idea of the plant variation that can be

expected.

For each data set we develop a linear system model of the plant. To simplify the problem

we will assume that the state in this model is accessible, so the diﬀerent models refer to the

same state vector. These models should be fairly close, but of course not exactly the same.

We might ﬁnd for example that the transfer matrices of these models at s = 0 diﬀer by about

10%, but at high frequencies they diﬀer considerably more. More importantly, this collection

of models contains information about the “structure” of the plant variation.

We propose to model the system as a PLDI with the vertices given by the measured or

estimated linear system models. In other words, we model the plant as a time-varying linear

system, with system matrices that can jump around among any of the models we estimated.

It seems reasonable to conjecture that a controller that works well with this PLDI is likely

to work well when connected to the real system. Such controllers can be designed by the

methods described in Chapter 7.

Well-posedness of DNLDIs

See [BY89] for a proof that a DNLDI is well-posed if and only if det(I −Dqp∆) > 0 for every

∆ with [∆ii[ = 1; the idea behind this proof can be traced back to Zadeh and Desoer [ZD63,

¸9.17]; see also [Sae86]. An equivalent condition is in terms of 10 matrices (a matrix is 10

if every principal minor is nonnegative; see [FP62, FP66, FP67]): A DNLDI is well-posed if

and only if I +Dqp is invertible, and (I +Dqp)

−1

(I −Dqp) is a 10 matrix [Gha90]. Another

equivalent condition is that (I + Dqp, I − Dqp) is a J0 pair (see [Wil70]). The problem

of determining whether a general matrix is 10 is thought to be very diﬃcult, i.e., of high

complexity [CPS92, p149].

Standard branch-and-bound techniques can be applied to the problem of determining well-

posedness of DNLDIs; see for example [Sd86, GS88, SP89, BBB91, BB92]. Of course these

methods have no theoretical advantage over simply checking that the 2

nq

determinants are

positive. But in practice they may be more eﬃcient, especially when used with computation-

ally cheap suﬃcient conditions (see below).

There are several suﬃcient conditions for well-posedness of a DNLDI that can be checked in

polynomial-time. Fan, Tits, and Doyle [FTD91] show that if the LMI

D

T

qp

PDqp < P, P diagonal, P > 0, (4.18)

is feasible, then the DNLDI is well-posed. Condition (4.18) is equivalent to the existence of a

diagonal (scaling) matrix R such that |RDqpR

−1

| < 1. In [FTD91], the authors also give a

suﬃcient condition for well-posedness of more general “structured” LDIs, e.g., when equality

constraints are imposed on the matrix ∆ in (4.11). These conditions can also be obtained

using the S-procedure; see [Fer93].

Representing DNLDIs as PLDIs

For a well-posed DNLDI, CoΩ is a polytope, which means that the DNLDI (4.12) can also

be represented as a PLDI. More speciﬁcally, we have

CoΩ = Co

_

˜

A+

˜

B∆(I −Dqp∆)

−1

˜

C

¸

¸

∆ diagonal, [∆ii[ = 1

_

,

This electronic version is for personal use and may not be duplicated or distributed.

58 Chapter 4 Linear Differential Inclusions

where

˜

A,

˜

B and

˜

C are given by (4.10). Note that the set on the right-hand side is a polytope

with at most 2

nq

vertices. We now prove this result.

Deﬁne

l =

_

˜

A+

˜

B∆(I −Dqp∆)

−1

˜

C

¸

¸

∆ diagonal, [∆ii[ = 1

_

,

which is a set containing at most 2

nq

points. By deﬁnition

CoΩ = Co

_

˜

A+

˜

B∆(I −Dqp∆)

−1

˜

C

¸

¸

∆ diagonal, [∆ii[ ≤ 1

_

.

Clearly we have CoΩ ⊇ Col. We will now show that every extreme point of CoΩ is

contained in Col, which will imply that Col ⊇ CoΩ, completing the proof.

Let Wext be an extreme point of CoΩ. Then there exists a linear function ψ such that Wext

is the unique maximizer of ψ over CoΩ. In fact, Wext ∈ Ω, since otherwise, the half-space

|W [ ψ(W) < ψ(Wext) −¦, where > 0 is suﬃciently small, contains Ω but not Wext, which

contradicts Wext ∈ CoΩ.

Now, for W ∈ Ω, ψ(W) is a bounded linear-fractional function of the ﬁrst diagonal entry

∆11 of ∆, for ﬁxed ∆22, . . . , ∆nqnq

, i.e., it equals (a +b∆11)/(c +d∆11) for some a, b, c and

d. Moreover, the denominator c + d∆11 is nonzero for ∆11 ∈ [−1, 1], since the DNLDI is

well-posed. Therefore, for every ﬁxed ∆22, . . . , ∆nqnq

, the function ψ achieves a maximum at

an extreme value of ∆11, i.e., ∆11 = ±1. Extending this argument to ∆22, . . . , ∆nqnq

leads

us to the following conclusion: Wext =

˜

A +

˜

BS (I −DqpS)

−1

˜

C, where S satisﬁes Sii = ±1.

In other words, Wext ∈ Col, which concludes our proof.

We have already noted that the description of a DNLDI as a PLDI will be much larger than

its description as a DNLDI; see the Notes and References of Chapter 5, and also [BY89] and

[PS82, KP87a, Kam83].

Approximating PLDIs by NLDIs

It is often useful to conservatively approximate a PLDI as an NLDI. One reason, among

many, is the potentially much smaller size of the description as an NLDI compared to the

description of the PLDI. We consider the PLDI

˙ x = A(t)x, A(t) ∈ ΩPLDI

∆

= Co|A1, . . . , AL¦ , (4.19)

with x(t) ∈ R

n

, and the NLDI

˙ x = (A+Bp∆(t)Cq)x, |∆(t)| ≤ 1, (4.20)

with associated set ΩNLDI

∆

= |A+Bp∆Cq [ |∆| ≤ 1¦. Our goal is to ﬁnd A, Bp, and Cq

such that ΩPLDI ⊆ ΩNLDI with the set ΩNLDI as small as possible in some appropriate sense.

This will give an eﬃcient outer approximation of the PLDI (4.19) by the NLDI (4.20). Since

ΩPLDI ⊆ ΩNLDI implies that every trajectory of the PLDI is a trajectory of the NLDI, every

result we establish for all trajectories of the NLDI holds for all trajectories of the PLDI. For

L very large, it is much easier to work with the NLDI than the PLDI.

For simplicity, we will only consider the case with Bp ∈ R

n×n

and invertible. Evidently there

is substantial redundancy in our representation of the NLDI (4.20): We can replace Bp by

αBp and Cq by Cq/α where α ,= 0. (We can also replace Bp and Cq with BpU and V Cq

where U and V are any orthogonal matrices, without aﬀecting the set ΩNLDI, but we will

not use this fact.)

We have ΩPLDI ⊆ ΩNLDI if for every ξ and k = 1, . . . , L, there exists π such that

Bpπ = (A

k

−A)ξ, π

T

π ≤ ξ

T

C

T

q

Cqξ.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 59

This yields the equivalent condition

(A

k

−A)

T

B

−T

p

B

−1

p

(A

k

−A) ≤ C

T

q

Cq, k = 1, . . . , L,

which, in turn, is equivalent to

BpB

T

p

> 0,

_

C

T

q

Cq (A

k

−A)

T

A

k

−A BpB

T

p

_

≥ 0, k = 1, . . . , L.

Introducing new variables V = C

T

q

Cq and W = BpB

T

p

, we get the equivalent LMI in A, V

and W

W > 0,

_

V (A

k

−A)

T

A

k

−A W

_

≥ 0, k = 1, . . . , L. (4.21)

Thus we have ΩPLDI ⊆ ΩNLDI if condition (4.21) holds.

There are several ways to minimize the size of ΩNLDI ⊇ ΩPLDI. The most obvious is to

minimize Tr V + Tr W subject to (4.21), which is an EVP in A, V and W. This objective

is clearly related to several measures of the size of ΩNLDI, but we do not know of any simple

interpretation.

We can formulate the problem of minimizing the diameter of ΩNLDI as an EVP. We deﬁne the

diameter of a set Ω ⊆ R

n×n

as max ||F −G| [ F, G ∈ Ω¦. The diameter of ΩNLDI is equal

to 2

_

λmax(V )λmax(W). We can exploit the scaling redundancy in Cq to assume without loss

of generality that λmax(V ) ≤ 1, and then minimize λmax(W) subject to (4.21) and V ≤ I,

which is an EVP. It can be shown that the resulting optimal V

opt

satisﬁes λmax(V

opt

) = 1,

so that we have in fact minimized the diameter of ΩNLDI subject to ΩPLDI ⊆ ΩNLDI.

This electronic version is for personal use and may not be duplicated or distributed.

Chapter 5

Analysis of LDIs:

State Properties

In this chapter we consider properties of the state x of the LDI

˙ x = A(t)x, A(t) ∈ Ω, (5.1)

where Ω ⊆ R

n×n

has one of four forms:

• LTI systems: LTI systems are described by ˙ x = Ax.

• Polytopic LDIs: PLDIs are described by ˙ x = A(t)x, A(t) ∈ Co{A

1

, . . . , A

L

}.

• Norm-bound LDIs: NLDIs are described by

˙ x = Ax +B

p

p, q = C

q

x +D

qp

p, p = ∆(t)q, ∆(t) ≤ 1,

which we will rewrite as

˙ x = Ax +B

p

p, p

T

p ≤ (C

q

x +D

qp

p)

T

(C

q

x +D

qp

p). (5.2)

We assume well-posedness, i.e., D

qp

< 1.

• Diagonal Norm-bound LDIs: DNLDIs are described by

˙ x = Ax +B

p

p, q = C

q

x +D

qp

p,

p

i

= δ

i

(t)q

i

, |δ

i

(t)| ≤ 1, i = 1, . . . , n

q

.

(5.3)

which can be rewritten as

˙ x = Ax +B

p

p, q = C

q

x +D

qp

p, |p

i

| ≤ |q

i

|, i = 1, . . . , n

q

.

Again, we assume well-posedness.

5.1 Quadratic Stability

We ﬁrst study stability of the LDI (5.1), that is, we ask whether all trajectories of

system (5.1) converge to zero as t →∞. A suﬃcient condition for this is the existence

of a quadratic function V (ξ) = ξ

T

Pξ, P > 0 that decreases along every nonzero

trajectory of (5.1). If there exists such a P, we say the LDI (5.1) is quadratically

stable and we call V a quadratic Lyapunov function.

61

62 Chapter 5 Analysis of LDIs: State Properties

Since

d

dt

V (x) = x

T

_

A(t)

T

P +PA(t)

_

x,

a necessary and suﬃcient condition for quadratic stability of system (5.1) is

P > 0, A

T

P +PA < 0 for all A ∈ Ω. (5.4)

Multiplying the second inequality in (5.4) on the left and right by P

−1

, and deﬁning

a new variable Q = P

−1

, we may rewrite (5.4) as

Q > 0, QA

T

+AQ < 0 for all A ∈ Ω. (5.5)

This dual inequality is an equivalent condition for quadratic stability. We now show

that conditions for quadratic stability for LTI systems, PLDIs, and NLDIs can be

expressed in terms of LMIs.

• LTI systems: Condition (5.4) becomes

P > 0, A

T

P +PA < 0. (5.6)

Therefore, checking quadratic stability for an LTI system is an LMIP in the variable

P. This is precisely the (necessary and suﬃcient) Lyapunov stability criterion for LTI

systems. (In other words, a linear system is stable if and only if it is quadratically sta-

ble.) Alternatively, using (5.5), stability of LTI systems is equivalent to the existence

of Q satisfying the LMI

Q > 0, AQ+QA

T

< 0. (5.7)

Of course, each of these (very special) LMIPs can be solved analytically by solving a

Lyapunov equation (see §1.2 and §2.7).

• Polytopic LDIs: Condition (5.4) is equivalent to

P > 0, A

T

i

P +PA

i

< 0, i = 1, . . . , L. (5.8)

Thus, determining quadratic stability for PLDIs is an LMIP in the variable P. The

dual condition (5.5) is equivalent to the LMI in the variable Q

Q > 0, QA

T

i

+A

i

Q < 0, i = 1, . . . , L. (5.9)

• Norm-bound LDIs: Condition (5.4) is equivalent to P > 0 and

_

ξ

π

_

T

_

A

T

P +PA PB

p

B

T

p

P 0

__

ξ

π

_

< 0 (5.10)

for all nonzero ξ satisfying

_

ξ

π

_

T

_

_

−C

T

q

C

q

−C

T

q

D

qp

−D

T

qp

C

q

I −D

T

qp

D

qp

_

_

_

ξ

π

_

≤ 0. (5.11)

In order to apply the S-procedure we show that the set

A = {(ξ, π) | ξ = 0, (5.11) }

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

5.1 Quadratic Stability 63

equals the set

B = {(ξ, π) | (ξ, π) = 0, (5.11) } .

It suﬃces to show that {(ξ, π) | ξ = 0, π = 0, (5.11) } = ∅. But this is immediate:

If π = 0, then condition (5.11) cannot hold without having ξ = 0, since I−D

T

qp

D

qp

> 0.

Therefore, the condition that dV (x)/dt < 0 for all nonzero trajectories is equivalent

to (5.10) being satisﬁed for any nonzero (ξ, π) satisfying (5.11). (This argument recurs

throughout this chapter and will not be repeated.) Using the S-procedure, we ﬁnd

that quadratic stability of (5.2) is equivalent to the existence of P and λ satisfying

P > 0, λ ≥ 0,

_

_

A

T

P +PA+λC

T

q

C

q

PB

p

+λC

T

q

D

qp

(PB

p

+λC

T

q

D

qp

)

T

−λ(I −D

T

qp

D

qp

)

_

_

< 0.

(5.12)

Thus, determining quadratic stability of an NLDI is an LMIP. Since LMI (5.12) implies

that λ > 0, we may, by deﬁning

˜

P = P/λ, obtain an equivalent condition

˜

P > 0,

_

_

A

T

˜

P +

˜

PA+C

T

q

C

q

˜

PB

p

+C

T

q

D

qp

(

˜

PB

p

+C

T

q

D

qp

)

T

−(I −D

T

qp

D

qp

)

_

_

< 0, (5.13)

an LMI in the variable

˜

P. Thus, quadratic stability of the NLDI has the frequency-

domain interpretation that the H

∞

norm of the LTI system

˙ x = Ax +B

p

p, q = C

q

x +D

qp

p

is less than one.

With the new variables Q = P

−1

, µ = 1/λ, quadratic stability of NLDIs is also

equivalent to the existence of µ and Q satisfying the LMI

µ ≥ 0, Q > 0,

_

AQ+QA

T

+µB

p

B

T

p

µB

p

D

T

qp

+QC

T

q

(µB

p

D

T

qp

+QC

T

q

)

T

−µ(I −D

qp

D

T

qp

)

_

< 0.

(5.14)

Remark: Note that our assumption of well-posedness is in fact incorporated in

the LMI (5.12) since it implies I − DqpD

T

qp

> 0. Therefore the NLDI (5.2) is

well-posed and quadratically stable if and only if the LMIP (5.12) has a solution.

In the remainder of this chapter we will assume that D

qp

in (5.2) is zero; the

reader should bear in mind that all the following results hold for nonzero D

qp

as well.

• Diagonal Norm-bound LDIs: For the DNLDI (5.3), we can obtain only a

suﬃcient condition for quadratic stability. The condition dV (x)/dt < 0 for all nonzero

trajectories is equivalent to

ξ

T

(A

T

P +PA)ξ + 2ξ

T

PB

p

π < 0

for all nonzero (ξ, π) satisfying

π

2

i

≤ (C

q,i

ξ +D

qp,i

π)

T

(C

q,i

ξ +D

qp,i

π) , i = 1, . . . , L,

This electronic version is for personal use and may not be duplicated or distributed.

64 Chapter 5 Analysis of LDIs: State Properties

where we have used C

q,i

and D

qp,i

to denote the ith rows of C

q

and D

qp

respec-

tively. Using the S-procedure, we see that this condition is implied by the existence

of nonnegative λ

1

, . . . , λ

nq

such that

ξ

T

(A

T

P +PA)ξ + 2ξ

T

PB

p

π

+

L

i=1

λ

i

_

(C

q,i

ξ +D

qp,i

π)

T

(C

q,i

ξ +D

qp,i

π) −π

2

i

_

< 0

for all nonzero (ξ, π). With Λ = diag(λ, . . . , λ

nq

), this is equivalent to

_

A

T

P +PA+C

T

q

ΛC

q

PB

p

+C

T

q

ΛD

qp

B

T

p

P +D

T

qp

ΛC

q

D

T

qp

ΛD

qp

−Λ

_

< 0. (5.15)

Therefore, if there exist P > 0 and diagonal Λ ≥ 0 satisfying (5.15), then the

DNLDI (5.3) is quadratically stable. Checking this suﬃcient condition for quadratic

stability is an LMIP. Note that from (5.15), we must have Λ > 0.

Equivalently, quadratic stability is implied by the existence of Q > 0, M =

diag(µ

1

, . . . , µ

nq

) > 0 satisfying the LMI

_

AQ+QA

T

+B

p

MB

T

p

QC

T

q

+B

p

MD

T

qp

C

q

Q+D

qp

MB

T

p

−M +D

qp

MD

T

qp

_

< 0, (5.16)

another LMIP.

Remark: Quadratic stability of a DNLDI via the S-procedure has a simple

frequency-domain interpretation. Denoting by H the transfer matrix H(s) =

Dqp + Cq(sI − A)

−1

Bp, and assuming (A, Bp, Cq) is minimal, quadratic stabil-

ity is equivalent to the fact that, for some diagonal, positive-deﬁnite matrix Λ,

|Λ

1/2

HΛ

−1/2

|∞ < 1. Λ

−1/2

can then be interpreted as a scaling. (See the Notes

and References for more details on the connection between the S-procedure and

scaling.)

Remark: Here too our assumption of well-posedness is in fact incorporated

in the LMI (5.15) or (5.16). The bottom right block is precisely the suﬃcient

condition for well-posedness mentioned on page 57. Therefore the DNLDI (5.2)

is well-posed and quadratically stable if the LMI (5.15) or (5.16) is feasible.

Note the similarity between the corresponding LMIs for the NLDI and the DNLDI;

the only diﬀerence is that the diagonal matrix appearing in the LMI associated with

the DNLDI is ﬁxed as the scaled identity matrix in the LMI associated with the NLDI.

In general it is straightforward to derive the corresponding (suﬃcient) condition for

DNLDIs from the result for NLDIs. In the remainder of the book, we will often restrict

our attention to NLDIs, and pay only occasional attention to DNLDIs; we leave to

the reader the task of generalizing all the results for NLDIs to DNLDIs.

Remark: An LTI system is stable if and only if it is quadratically stable. For

more general LDIs, however, stability does not imply quadratic stability (see the

Notes and References at the end of this chapter). It is true that an LDI is

stable if and only if there exists a convex Lyapunov function that proves it; see

the Notes and References.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

5.1 Quadratic Stability 65

5.1.1 Coordinate transformations

We can give another interpretation of quadratic stability in terms of state-space coor-

dinate transformations. Consider the change of coordinates x = T ¯ x, where det T = 0.

In the new coordinate system, (5.1) is described by the matrix

¯

A(t) = T

−1

A(t)T.

We ask the question: Does there exist T such that in the new coordinate system, all

(nonzero) trajectories of the LDI are always decreasing in norm, i.e., d¯ x/dt < 0?

It is easy to show that this is true if and only if there exists a quadratic Lyapunov

function V (ξ) = ξ

T

Pξ for the LDI, in which case we can take any T with TT

T

= P

−1

,

for example, T = P

−1/2

(these T’s are all related by right orthogonal matrix multi-

plication). With this interpretation, it is natural to seek a coordinate transformation

matrix T that makes all nonzero trajectories decreasing in norm, and has the small-

est possible condition number κ(T). This turns out to be an EVP for LTI systems,

polytopic and norm-bound LDIs.

To see this, let P = T

−T

T

−1

, so that κ(T)

2

= κ(P). Minimizing the condition

number of T subject to the requirement that d¯ x/dt < 0 for all nonzero trajectories is

then equivalent to minimizing the condition number of P subject to d(x

T

Px)/dt < 0.

(Any T with TT

T

= P

−1

opt

where P

opt

is an optimal solution, is optimal for the original

problem.) For each of our systems (LTI systems, PLDIs and NLDIs), the change of

coordinates with smallest condition number is obtained by solving the following EVPs:

• LTI systems: The EVP in the variables P and η is

minimize η

subject to (5.6), I ≤ P ≤ ηI

(In this formulation, we have taken advantage of the homogeneity of the LMI (5.6) in

the variable P.)

• Polytopic LDIs: The EVP in the variables P and η is

minimize η

subject to (5.8), I ≤ P ≤ ηI

• Norm-bound LDIs: The EVP in the variables P, η and λ is

minimize η

subject to (5.12), I ≤ P ≤ ηI

5.1.2 Quadratic stability margins

Quadratic stability margins give a measure of how much the set Ω can be expanded

about some center with the LDI remaining quadratically stable.

• Polytopic LDIs: For the PLDI

˙ x = (A

0

+A(t))x, A(t) ∈ α Co{A

1

, . . . , A

L

} ,

we deﬁne the quadratic stability margin as the largest nonnegative α for which it is

quadratically stable. This quantity is computed by solving the following GEVP in P

This electronic version is for personal use and may not be duplicated or distributed.

66 Chapter 5 Analysis of LDIs: State Properties

and α:

maximize α

subject to P > 0, α ≥ 0,

A

T

0

P +PA

0

+α

_

A

T

i

P +PA

i

_

< 0, i = 1, . . . , L

• Norm-bound LDIs: The quadratic stability margin of the system

˙ x = Ax +B

p

p, p

T

p ≤ α

2

x

T

C

T

q

C

q

x,

is deﬁned as the largest α ≥ 0 for which the system is quadratically stable, and is

computed by solving the GEVP in P, λ and β = α

2

:

maximize β

subject to P > 0, β ≥ 0,

_

A

T

P +PA+βλC

T

q

C

q

PB

p

B

T

p

P −λI

_

< 0

Deﬁning

˜

P = P/λ, we get an equivalent EVP in

˜

P and β:

maximize β

subject to

˜

P > 0, β ≥ 0,

_

A

T

˜

P +

˜

PA+βC

T

q

C

q

˜

PB

p

B

T

p

˜

P −I

_

< 0

Remark: The quadratic stability margin obtained by solving this EVP is just

1/|Cq(sI −A)

−1

Bp|∞.

5.1.3 Decay rate

The decay rate (or largest Lyapunov exponent) of the LDI (5.1) is deﬁned to be the

largest α such that

lim

t→∞

e

αt

x(t) = 0

holds for all trajectories x. Equivalently, the decay rate is the supremum of

liminf

t→∞

−log x(t)

t

over all nonzero trajectories. (Stability corresponds to positive decay rate.)

We can use the quadratic Lyapunov function V (ξ) = ξ

T

Pξ to establish a lower

bound on the decay rate of the LDI (5.1). If dV (x)/dt ≤ −2αV (x) for all trajectories,

then V (x(t)) ≤ V (x(0))e

−2αt

, so that x(t) ≤ e

−αt

κ(P)

1/2

x(0) for all trajectories,

and therefore the decay rate of the LDI (5.1) is at least α.

• LTI systems: The condition that dV (x)/dt ≤ −2αV (x) for all trajectories is

equivalent to

A

T

P +PA+ 2αP ≤ 0. (5.17)

Therefore, the largest lower bound on the decay rate that we can ﬁnd using a quadratic

Lyapunov function can be found by solving the following GEVP in P and α:

maximize α

subject to P > 0, (5.17)

(5.18)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

5.1 Quadratic Stability 67

This lower bound is sharp, i.e., the optimal value of the GEVP (5.18) is the decay rate

of the LTI system (which is the stability degree of A, i.e., negative of the maximum

real part of the eigenvalues of A).

As an alternate condition, there exists a quadratic Lyapunov function proving

that the decay rate is at least α if and only if there exists Q satisfying the LMI

Q > 0, AQ+QA

T

+ 2αQ ≤ 0. (5.19)

• Polytopic LDIs: The condition that dV (x)/dt ≤ −2αV (x) for all trajectories is

equivalent to the LMI

A

T

i

P +PA

i

+ 2αP ≤ 0, i = 1, . . . , L. (5.20)

Therefore, the largest lower bound on the decay rate provable via quadratic Lyapunov

functions is obtained by maximizing α subject to (5.20) and P > 0. This is a GEVP

in P and α.

As an alternate condition, there exists a quadratic Lyapunov function proving

that the decay rate is at least α if and only if there exists Q satisfying the LMI

Q > 0, A

i

Q+QA

T

i

+ 2αQ ≤ 0, i = 1, . . . , L. (5.21)

• Norm-bound LDIs: Applying the S-procedure, the condition that dV (x)/dt ≤

−2αV (x) for all trajectories is equivalent to the existence of λ ≥ 0 such that

_

A

T

P +PA+λC

T

q

C

q

+ 2αP PB

p

B

T

p

P −λI

_

≤ 0. (5.22)

Therefore, we obtain the largest lower bound on the decay rate of (5.2) by maximizing

α over the variables α, P and λ, subject to P > 0, λ ≥ 0 and (5.22), a GEVP.

The lower bound has a simple frequency-domain interpretation: Deﬁne the α-

shifted H

∞

norm of the system by

H

∞,α

∆

= sup{ H(s) | Re s > −α} .

Then the optimal value of the GEVP is equal to the largest α such that H

∞,α

< 1.

This can be seen by rewriting (5.22) as

_

(A+αI)

T

P +P (A+αI) +λC

T

q

C

q

PB

p

B

T

p

P −λI

_

≤ 0,

and noting that the H

∞

norm of the system (A+αI, B

p

, C

q

) equals H

∞,α

.

An alternate necessary and suﬃcient condition for the existence of a quadratic

Lyapunov function proving that the decay rate of (5.2) is at least α is that there

exists Q and µ satisfying the LMI

Q > 0, µ ≥ 0,

_

AQ+QA

T

+µB

p

B

T

p

+ 2αQ QC

T

q

C

q

Q −µI

_

≤ 0.

(5.23)

This electronic version is for personal use and may not be duplicated or distributed.

68 Chapter 5 Analysis of LDIs: State Properties

5.2 Invariant Ellipsoids

Quadratic stability can also be interpreted in terms of invariant ellipsoids. For Q > 0,

let E denote the ellipsoid centered at the origin

E =

_

ξ ∈ R

n

¸

¸

ξ

T

Q

−1

ξ ≤ 1

_

.

The ellipsoid E is said to be invariant for the LDI (5.1) if for every trajectory x of the

LDI, x(t

0

) ∈ E implies x(t) ∈ E for all t ≥ t

0

. It is easily shown that this is the case

if and only if Q satisﬁes

QA

T

+AQ ≤ 0, for all A ∈ Ω,

or equivalently,

A

T

P +PA ≤ 0, for all A ∈ Ω, (5.24)

where P = Q

−1

. Thus, for LTI systems, PLDIs and NLDIs, invariance of E is charac-

terized by LMIs in Q or P, its inverse.

Remark: Condition (5.24) is just the nonstrict version of condition (5.4).

• LTI systems: The corresponding LMI in P is

P > 0, A

T

P +PA ≤ 0, (5.25)

and the LMI in Q is

Q > 0, AQ+PA

T

≤ 0, (5.26)

• Polytopic LDIs: The corresponding LMI in P is

P > 0, A

T

i

P +PA

i

≤ 0, i = 1, . . . , L, (5.27)

and the LMI in Q is

Q > 0, QA

T

i

+A

i

Q ≤ 0, i = 1, . . . , L. (5.28)

• Norm-bound LDIs: Applying the S-procedure, invariance of the ellipsoid E is

equivalent to the existence of λ such that

P > 0, λ ≥ 0,

_

_

A

T

P +PA+λC

T

q

C

q

PB

p

+λC

T

q

D

qp

(PB

p

+λC

T

q

D

qp

)

T

−λ(I −D

T

qp

D

qp

)

_

_

≤ 0.

(5.29)

Invariance of the ellipsoid E is equivalent to the existence of µ such that

µ ≥ 0, Q > 0, and

_

AQ+QA

T

+µB

p

B

T

p

µB

p

D

T

qp

+QC

T

q

(µB

p

D

T

qp

+QC

T

q

)

T

−µ(I −D

qp

D

T

qp

)

_

≤ 0.

(5.30)

In summary, the condition that E be an invariant ellipsoid can be expressed in

each case as an LMI in Q or its inverse P.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

5.2 Invariant Ellipsoids 69

5.2.1 Smallest invariant ellipsoid containing a polytope

Consider a polytope described by its vertices, P = Co{v

1

, . . . , v

p

}. The ellipsoid E

contains the polytope P if and only if

v

T

j

Q

−1

v

j

≤ 1, j = 1, . . . , p.

This condition may be expressed as an LMI in Q

_

1 v

T

j

v

j

Q

_

≥ 0, j = 1, . . . , p, (5.31)

or as an LMI in P = Q

−1

as

v

T

j

Pv

j

≤ 1, j = 1, . . . , p. (5.32)

Minimum volume

The volume of E is, up to a constant that depends on n,

√

det Q. We can minimize

this volume by solving appropriate CPs.

• LTI systems: The CP in P is

minimize log det P

−1

subject to (5.25), (5.32)

• Polytopic LDIs: The CP in P is

minimize log det P

−1

subject to (5.27), (5.32)

(5.33)

• Norm-bound LDIs: The CP in P and λ is

minimize log det P

−1

subject to (5.29), (5.32)

Minimum diameter

The diameter of the ellipsoid E is 2

_

λ

max

(Q). We can minimize this quantity by

solving appropriate EVPs.

• LTI systems: The EVP in the variables Q and λ is

minimize λ

subject to (5.26), (5.31), Q ≤ λI

• Polytopic LDIs: The EVP in Q and λ is

minimize λ

subject to (5.28), (5.31), Q ≤ λI

(5.34)

• Norm-bound LDIs: The EVP in Q, λ and µ is

minimize λ

subject to (5.30), (5.31), Q ≤ λI

This electronic version is for personal use and may not be duplicated or distributed.

70 Chapter 5 Analysis of LDIs: State Properties

Remark: These results have the following use. The polytope 1 represents our

knowledge of the state at time t0, i.e., x(t0) ∈ 1 (this knowledge may reﬂect

measurements or prior assumptions). An invariant ellipsoid c containing 1 then

gives a bound on the state for t ≥ t0, i.e., we can guarantee that x(t) ∈ c for all

t ≥ t0.

5.2.2 Largest invariant ellipsoid contained in a polytope

We now consider a polytope described by linear inequalities:

P =

_

ξ ∈ R

n

¸

¸

a

T

k

ξ ≤ 1, k = 1, . . . , q

_

.

The ellipsoid E is contained in the polytope P if and only if

max

_

a

T

k

ξ | ξ ∈ E

_

≤ 1, k = 1, . . . , q.

This is equivalent to

a

T

k

Qa

k

≤ 1, k = 1, . . . , q, (5.35)

which is a set of linear inequalities in Q. The maximum volume of invariant ellipsoids

contained in P can be found by solving CPs:

• LTI systems: For LTI systems, the CP in the variable Q is

minimize log det Q

−1

subject to (5.26), (5.35)

(Since the volume of E is, up to a constant,

√

det Q, minimizing log det Q

−1

will

maximize the volume of E).

• Polytopic LDIs: For PLDIs, the CP in the variable Q is

minimize log det Q

−1

subject to (5.28), (5.35)

• Norm-bound LDIs: For NLDIs, the CP in the variables Q and µ is

minimize log det Q

−1

subject to (5.30), (5.35)

We also can ﬁnd the maximum minor diameter (that is, the length of the minor

axis) of invariant ellipsoids contained in P by solving EVPs:

• LTI systems: For LTI systems, the EVP in the variables Q and λ is

maximize λ

subject to (5.26), (5.35), λI ≤ Q

• Polytopic LDIs: For PLDIs, the EVP in the variables Q and λ is

maximize λ

subject to (5.28), 5.35), λI ≤ Q

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

5.2 Invariant Ellipsoids 71

• Norm-bound LDIs: For NLDIs, the EVP in the variables Q, µ and λ is

maximize λ

subject to (5.30), (5.35), λI ≤ Q

Remark: These results can be used as follows. The polytope 1 represents the

allowable (or safe) operating region for the system. The ellipsoids found above

can be interpreted as regions of safe initial conditions, i.e., initial conditions for

which we can guarantee that the state always remains in the safe operating region.

5.2.3 Bound on return time

The return time of a stable LDI for the polytope P is deﬁned as the smallest T such

that if x(0) ∈ P, then x(t) ∈ P for all t ≥ T. Upper bounds on the return time can

be found by solving EVPs.

• LTI systems: Let us consider a positive decay rate α. If Q > 0 satisﬁes

QA

T

+AQ+ 2αQ ≤ 0,

then E is an invariant ellipsoid, and moreover x(0) ∈ E implies that x(t) ∈ e

−αt

E for

all t ≥ 0. Therefore if T is such that

e

−αT

E ⊆ P ⊆ E,

we can conclude that if x(0) ∈ P, then x(t) ∈ P for all t ≥ T, so that T is an upper

bound on the return time. If we use both representations of the polytope,

P =

_

x ∈ R

n

¸

¸

a

T

k

x ≤ 1, k = 1, . . . , q

_

= Co{v

1

, . . . , v

p

},

the constraint e

−αT

E ⊆ P is equivalent to the LMI in Q and γ

a

T

i

Qa

i

≤ γ, k = 1, . . . , q, (5.36)

where γ = e

2αT

, and the problem of ﬁnding the smallest such γ and therefore the

smallest T (for a ﬁxed α) is an EVP in the variables γ and Q:

minimize γ

subject to (5.19), (5.31), (5.36)

• Polytopic LDIs: For PLDIs, the smallest bound on the return time provable via

invariant ellipsoids, with a given decay rate α, is obtained by solving the EVP in the

variables Q and γ

minimize γ

subject to (5.21), (5.31), (5.36)

• Norm-bound LDIs: For NLDIs, the smallest bound on the return time provable

via invariant ellipsoids, with a given decay rate α, is obtained by solving the EVP in

the variables Q, γ and µ

minimize γ

subject to (5.23), (5.31), (5.36)

This electronic version is for personal use and may not be duplicated or distributed.

72 Chapter 5 Analysis of LDIs: State Properties

Notes and References

Quadratic stability for NLDIs

Lur’e and Postnikov [LP44] gave one of the earliest stability analyses for NLDIs: They

considered the stability of the system

˙ x = Ax +bpp, q = cqx, p = φ(q, t)q, (5.37)

where p and q are scalar. This problem came to be popularly known as the problem of

“absolute stability in automatic control”. In the original setting of Lur’e and Postnikov, φ was

assumed to be a time-invariant nonlinearity. Subsequently, various additional assumptions

were made about φ, generating diﬀerent special cases and solutions. (The family of systems of

the form (5.37), where φ(q, t) can be any sector [−1, 1] nonlinearity, is an NLDI.) Pyatnitskii,

in [Pya68], points out that by 1968, over 200 papers had been written about the system (5.37).

Among these, the ones most relevant to this book are undoubtedly from Yakubovich. As

far as we know, Yakubovich was the ﬁrst to make systematic use of LMIs along with the

S-procedure to prove stability of nonlinear control systems (see references [Yak62, Yak64,

Yak67, Yak66, Yak77, Yak82]). The main idea of Yakubovich was to express the re-

sulting LMIs as frequency-domain criteria for the transfer function G(s) = cq(sI − A)

−1

bp.

(Yakubovich calls this method the “method of matrix inequalities”.) These criteria are most

useful when dealing with experimental data arising from frequency response measurements;

they are described in detail in the book by Narendra and Taylor [NT73]. See also [Bar70b].

Popov [Pop73] and Willems [Wil71b, Wil74a] outline the relationship between the problem

of absolute stability of automatic control and quadratic optimal control.

The case when p and q are vector-valued signals has been considered much more recently;

see for instance, [ZK88, ZK87, KR88, BH89]. In [KPZ90, PD90], the authors remark

that quadratic stability of NLDIs is equivalent to the H∞ condition (5.13). However, as

stated in [PD90], there is no fundamental diﬀerence between these and the older results

of Yakubovich and Popov; the newer results can be regarded as extensions of the works

of Yakubovich and Popov to feedback synthesis and to the case of “structured perturba-

tions” [RCDP93, PZP

+

92].

While all the results presented in this chapter are based on Lyapunov stability techniques,

they have very strong connections with the input-output approach to study the stability

of nonlinear, uncertain systems. This approach was sparked by Popov [Pop62], who used

it to study the Lur’e system. Other major contributors include Zames [Zam66a, Zam66b,

ZF67], Sandberg [San64, San65a, San65b] and Brockett [Bro65, Bro66, BW65]. See

also the papers and the book by Willems [Wil69a, Wil69b, Wil71a]. The advantage of

the input-output approach is that the systems are not required to have ﬁnite-dimensional

state [Des65, DV75, Log90]; the corresponding stability criteria are usually most easily

expressed in the frequency domain, and result in inﬁnite-dimensional convex optimization

problems. Modern approaches from this viewpoint include Doyle’s µ-analysis [Doy82] and

Safonov’s Km-analysis [Saf82, SD83, SD84], implemented in the robustness analysis soft-

ware packages µ-tools and Km-tools [BDG

+

91, CS92a], which approximately solve these

inﬁnite-dimensional problems. See also [AS79, SS81].

Quadratic stability for PLDIs

This topic is much more recent than NLDIs, the likely reason being that the LMI expressing

quadratic stability of a general PLDI (which is just a number of simultaneous Lyapunov

inequalities (5.8)) cannot be converted to a Riccati inequality or a frequency-domain criterion.

One of the ﬁrst occurrences of PLDIs is in [HB76], where Horisberger and Belanger note

that the problem of quadratic stability for a PLDI is convex (the authors write down the

corresponding LMIs). For other results and computational procedures for PLDIs see [PS82,

PS86, BY89, KP87a, Kam83, EZKN90, KB93, AGG94]; the article [KP87a], where the

authors develop a subgradient method for proving quadratic stability of a PLDI, which is

further reﬁned in [Mei91]; the paper [PS86], where the discrete-time counterpart of this

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 73

problem is considered. See also the articles by Gu et al. [GZL90, GL93b, Gu94], Garofalo

et al. [GCG93], and the survey article by Barmish et al. [BK93].

Necessary and suﬃcient conditions for LDI stability

Much attention has also focused on necessary and suﬃcient conditions for stability of LDIs

(as opposed to quadratic stability); see for example [PR91b] or [MP86]. Earlier references on

this topic include [Pya70b] which connects the problem of stability of an LDI to an optimal

control problem, and [Pya70a, BT79, BT80], where the discrete-time counterpart of this

problem is considered.

Vector Lyapunov functions

The technique of vector Lyapunov functions also yields search problems that can be expressed

as LMIPs. One example is given in [CFA90]. Here we already have quadratic Lyapunov

functions for a number of subsystems; the problem is to ﬁnd an appropriate positive linear

combination that proves, e.g., stability of an interconnected system.

Finally, let us point out that quadratic Lyapunov functions have been used to determine

estimates of regions of stability for general nonlinear systems. See for example [Gha94,

BD85].

Stable LDIs that are not quadratically stable

An LTI system is stable if and only if it is quadratically stable; this is just Lyapunov’s

stability theorem for linear systems (see e.g., [Lya47, p277]). It is possible, however, for an

LDI to be stable without being quadratically stable. Here is a PLDI example:

˙ x = A(t)x, A(t) ∈ Co|A1, A2¦ ,

A1 =

_

−100 0

0 −1

_

, A2 =

_

8 −9

120 −18

_

.

From (2.8), this PLDI is not quadratically stable if there exist Q0 ≥ 0, Q1 ≥ 0 and Q2 ≥ 0,

not all zero, such that

Q0 = A1Q1 +Q1A

T

1

+A2Q2 +Q2A

T

2

.

It can be veriﬁed that the matrices

Q0 =

_

5.2 2

2 24

_

, Q1 =

_

0.1 3

3 90

_

, Q2 =

_

2.7 1

1 1

_

satisfy these duality conditions.

However, the piecewise quadratic Lyapunov function

V (x) = max

_

x

T

P1x, x

T

P2x

_

,

P1 =

_

14 −1

−1 1

_

, P2 =

_

0 0

0 1

_

,

(5.38)

proves that the PLDI is stable. To show this, we use the S-procedure. A necessary and

suﬃcient condition for the Lyapunov function V deﬁned in (5.38) to prove the stability of

This electronic version is for personal use and may not be duplicated or distributed.

74 Chapter 5 Analysis of LDIs: State Properties

the PLDI is the existence of four nonnegative numbers λ1, λ2, λ3, λ4 such that

A

T

1

P1 +P1A1 −λ1(P2 −P1) < 0,

A

T

2

P1 +P1A2 −λ2(P2 −P1) < 0,

A

T

1

P2 +P2A1 +λ3(P2 −P1) < 0, A

T

2

P2 +P2A2 +λ4(P2 −P1) < 0.

(5.39)

It can be veriﬁed that λ1 = 50, λ2 = 0, λ3 = 1, λ4 = 100 are such numbers.

For other examples of stable LDIs that are not quadratically stable, see Brockett [Bro65,

Bro77, Bro70] and Pyatnitskii [Pya71]. In [Bro77], Brockett uses the S-procedure to

prove that a certain piecewise quadratic form is a Lyapunov function and ﬁnds LMIs similar

to (5.39); see also [Vid78].

Finally, an LDI is stable if and only if there is a convex Lyapunov function that proves it.

One such Lyapunov function is

V (ξ) = sup ||x(t)| [ x satisﬁes (5.1), x(0) = ξ, t ≥ 0¦ .

See Brayton and Tong [BT79, BT80], and also [Zub55, MV63, DK71, VV85, PR91b,

PR91a, Rap90, Rap93, MP89, Mol87]. In general, computing such Lyapunov functions is

computationally intensive, if not intractable.

Nonlinear systems and fading memory

Consider the nonlinear system

˙ x = f(x, w, t), (5.40)

with

∂f

∂x

∈ Ω for all x, w, t

where Ω is convex. Suppose the LDI ˙ x ∈ Ωx is stable, i.e., all trajectories converge to zero as

t →∞. This implies the system has fading memory, that is, for ﬁxed input w, the diﬀerence

between any two trajectories of (5.40) converges to zero. In other words, the system “forgets”

its initial condition. (See [BC85].)

Fix an input w and let x and ˜ x denote any two solutions of (5.40). Using the mean-value

theorem from ¸4.3.1, we have

d

dt

(x − ˜ x) = f(x, w, t) −f(˜ x, w, t) = A(t)(x − ˜ x)

for some A(t) ∈ Ω. Since the LDI is stable, x(t) − ˜ x(t) converges to zero as t →∞.

Relation between S-procedure for DNLDIs and scaling

We discuss the interpretation of the diagonal matrix Λ in (5.15) as a scaling matrix.

Every trajectory of the DNLDI (5.3) is also a trajectory (and vice versa) of the DNLDI

˙ x = Ax +BpT

−1

p,

q = TCqx +TDqpT

−1

p,

pi = δi(t)qi, [δi(t)[ ≤ 1, i = 1, . . . , nq

(5.41)

for any diagonal nonsingular T. Therefore, if DNLDI (5.41) is stable for some diagonal

nonsingular T, then so is DNLDI (5.3). T is referred to as a scaling matrix.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 75

Treating DNLDI (5.41) as an NLDI, and applying the condition (5.13) for quadratic stability,

we require, for some diagonal nonsingular T,

P > 0,

_

A

T

P +PA+C

T

q

T

2

Cq PBpT

−1

+C

T

q

T

2

DqpT

−1

(PBpT

−1

+C

T

q

T

2

DqpT

−1

)

T

−(I −T

−1

D

T

qp

T

2

DqpT

−1

)

_

< 0,

An obvious congruence, followed by the substitution T

2

= Λ, yields the LMI condition in

P > 0 and diagonal Λ > 0:

_

A

T

P +PA+C

T

q

ΛCq PBp +CqΛDqp

B

T

P

P +D

T

qp

ΛCq D

T

qp

ΛDq −Λ

_

< 0.

This is precisely LMI (5.15), the condition for stability of the DNLDI (5.3) obtained using

the S-procedure. This relation between scaling and the diagonal matrices arising from the

S-procedure is described in Boyd and Yang [BY89].

Pyatnitskii and Skorodinskii [PS82, PS83], and Kamenetskii [Kam83] reduce the problem of

numerical search for appropriate scalings and the associated Lyapunov functions to a convex

optimization problem. Saeki and Araki [SA88] also conclude convexity of the scaling problem

and use the properties of M-matrices to obtain a solution.

Representing a DNLDI as a PLDI: Implications for quadratic stability

We saw in Chapter 4 that a DNLDI can be represented as a PLDI, where the number L of

vertices can be as large as 2

nq

. From the results of ¸5.1, checking quadratic stability for this

PLDI requires the solution of L simultaneous Lyapunov inequalities in P > 0 (LMI (5.8)). If

the system is represented as a DNLDI, a suﬃcient condition for quadratic stability is given

by the LMI (5.15), in the variables P > 0 and Λ > 0.

Comparing the two conditions for quadratic stability, we observe that while the quadratic

stability condition (5.8) for LDI (5.3), obtained by representing it as a PLDI, is both necessary

and suﬃcient (i.e., not conservative), the size of the LMI grows exponentially with nq, the

size of ∆. On the other hand, the size of the LMI (5.15) grows polynomially with nq, but

this LMI yields only a suﬃcient condition for quadratic stability of LDI (5.3). Discussion of

this issue can be found in Kamenetskii [Kam83] and Pyatnitskii and Skorodinskii [PS82].

This electronic version is for personal use and may not be duplicated or distributed.

Chapter 6

Analysis of LDIs:

Input/Output Properties

6.1 Input-to-State Properties

We ﬁrst consider input-to-state properties of the LDI

˙ x = A(t)x +B

w

(t)w, [A(t) B

w

(t)] ∈ Ω, (6.1)

where w is an exogenous input signal.

• LTI systems: The system has the form ˙ x = Ax +B

w

w.

• Polytopic LDIs: Here the description is

˙ x = A(t)x +B

w

(t)w, [A(t) B

w

(t)] ∈ Co{[A

1

B

w,1

] , . . . , [A

L

B

w,L

]} .

• Norm-bound LDIs: NLDIs have the form

˙ x = Ax +B

p

p +B

w

w, q = C

q

x, p = ∆(t)q, ∆(t) ≤ 1,

or equivalently,

˙ x = Ax +B

p

p +B

w

w, q = C

q

x, p

T

p ≤ q

T

q.

• Diagonal Norm-bound LDIs: DNLDIs have the form

˙ x = Ax +B

p

p +B

w

w, q = C

q

x,

p

i

= δ

i

(t)q

i

, |δ

i

(t)| ≤ 1, i = 1, . . . , n

q

,

or equivalently,

˙ x = Ax +B

p

p +B

w

w, q = C

q

x, |p

i

| ≤ |q

i

|, i = 1, . . . , n

q

.

6.1.1 Reachable sets with unit-energy inputs

Let R

ue

denote the set of reachable states with unit-energy inputs for the LDI (6.1),

i.e.,

R

ue

∆

=

_

¸

¸

_

¸

¸

_

x(T)

¸

¸

¸

¸

¸

¸

¸

¸

x, w satisfy (6.1), x(0) = 0

_

T

0

w

T

w dt ≤ 1, T ≥ 0

_

¸

¸

_

¸

¸

_

.

77

78 Chapter 6 Analysis of LDIs: Input/Output Properties

We will bound R

ue

by ellipsoids of the form

E =

_

ξ | ξ

T

Pξ ≤ 1

_

, (6.2)

where P > 0.

Suppose that the function V (ξ) = ξ

T

Pξ, with P > 0, satisﬁes

dV (x)/dt ≤ w

T

w for all x, w satisfying (6.1). (6.3)

Integrating both sides from 0 to T, we get

V (x(T)) −V (x(0)) ≤

_

T

0

w

T

w dt.

Noting that V (x(0)) = 0 since x(0) = 0, we get

V (x(T)) ≤

_

T

0

w

T

w dt ≤ 1,

for every T ≥ 0, and every input w such that

_

T

0

w

T

w dt ≤ 1. In other words, the

ellipsoid E contains the reachable set R

ue

.

• LTI systems: Condition (6.3) is equivalent to the LMI in P

P > 0,

_

A

T

P +PA PB

w

B

T

w

P −I

_

≤ 0, (6.4)

which can also be expressed as an LMI in Q = P

−1

Q > 0, AQ+QA

T

+B

w

B

T

w

≤ 0. (6.5)

Remark: For a controllable LTI system, the reachable set is the ellipsoid

|ξ [ ξ

T

W

−1

c

ξ ≤ 1¦, where Wc is the controllability Gramian, deﬁned by

Wc

∆

=

_

∞

0

e

At

BwB

T

w

e

A

T

t

dt.

Since Wc satisﬁes the Lyapunov equation

AWc +WcA

T

+BwB

T

w

= 0, (6.6)

we see that Q = Wc satisﬁes (6.5). Thus the ellipsoidal bound is sharp for LTI

systems.

• Polytopic LDIs: For PLDIs, condition (6.3) holds if and only if

P > 0 and

_

A(t)

T

P +PA(t) PB

w

(t)

B

w

(t)

T

P −I

_

≤ 0 for all t ≥ 0. (6.7)

Inequality (6.7) holds if and only if

P > 0,

_

A

T

i

P +PA

i

PB

w,i

B

T

w,i

P −I

_

≤ 0, i = 1, . . . , L. (6.8)

Thus the ellipsoid E contains R

ue

if the LMI (6.8) holds.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

6.1 Input-to-State Properties 79

Alternatively, we can write (6.8) as

Q > 0, QA

T

i

+A

i

Q+B

w,i

B

T

w,i

≤ 0, i = 1, . . . , L, (6.9)

where Q = P

−1

. Note that this condition implies that E is invariant.

• Norm-bound LDIs: Condition (6.3) holds if and only if P > 0 and

ξ

T

(A

T

P +PA)ξ + 2ξ

T

P (B

p

π +B

w

ω) −ω

T

ω ≤ 0

hold for every ω and for every ξ and π satisfying

π

T

π −ξ

T

C

T

q

C

q

ξ ≤ 0.

Using the S-procedure, this is equivalent to the existence of P and λ satisfying

P > 0, λ ≥ 0,

_

¸

_

A

T

P +PA+λC

T

q

C

q

PB

p

PB

w

B

T

p

P −λI 0

B

T

w

P 0 −I

_

¸

_ ≤ 0. (6.10)

Equivalently, deﬁning Q = P

−1

, condition (6.3) holds if and only if there exist Q and

µ satisfying

Q > 0, µ ≥ 0,

_

AQ+QA

T

+B

w

B

T

w

+µB

p

B

T

p

QC

T

q

C

q

Q −µI

_

≤ 0. (6.11)

• Diagonal Norm-bound LDIs: Condition (6.3) holds if and only if P > 0 and

for all ξ and π that satisfy

π

2

i

≤ (C

q,i

ξ)

T

(C

q,i

ξ) , i = 1, . . . , n

q

,

we have for all ω,

ξ

T

(A

T

P +PA)ξ + 2ξ

T

P (B

p

π +B

w

ω) −ω

T

ω ≤ 0.

It follows from the S-procedure that this condition holds if there exist P and Λ =

diag(λ, . . . , λ

nq

) satisfying

P > 0, Λ ≥ 0,

_

¸

_

A

T

P +PA+C

T

q

ΛC

q

PB

p

PB

w

B

T

p

P −Λ 0

B

T

w

P 0 −I

_

¸

_ ≤ 0.

Equivalently, deﬁning Q = P

−1

, condition (6.3) holds if there exist Q and a diagonal

matrix M such that

Q > 0, M ≥ 0,

_

AQ+QA

T

+B

w

B

T

w

+B

p

MB

T

p

QC

T

q

C

q

Q −M

_

≤ 0. (6.12)

For the remainder of this section we leave the extension to DNLDIs to the reader.

The results described above give us a set of ellipsoidal outer approximations of

R

ue

. We can optimize over this set in several ways.

This electronic version is for personal use and may not be duplicated or distributed.

80 Chapter 6 Analysis of LDIs: Input/Output Properties

Smallest outer approximations

• LTI systems: For LTI systems, we can minimize the volume over ellipsoids E with

P satisfying (6.3) by minimizing log det P

−1

over the variable P subject to (6.4). This

is a CP.

If (A, B

w

) is uncontrollable, the volume of the reachable set is zero, and this CP

will be unbounded below. If (A, B

w

) is controllable, then the CP ﬁnds the exact

reachable set, which corresponds to P = W

−1

c

.

• Polytopic LDIs: For PLDIs, the minimum volume ellipsoid of the form (6.2)

with P satisfying (6.3), is found by minimizing log det P

−1

subject to (6.8). This is a

CP.

• Norm-bound LDIs: For NLDIs, the minimum volume ellipsoid of the form (6.2),

satisfying (6.3), is obtained by minimizing log det P

−1

over the variables P and λ

subject to (6.10). This is again a CP.

We can also minimize the diameter of ellipsoids of the form (6.2), satisfying (6.3),

by replacing the objective function log det P

−1

in the CPs by the objective function

λ

max

(P

−1

). This yields EVPs.

Testing if a point is outside the reachable set

The point x

0

lies outside the reachable set if there exists an ellipsoidal outer approx-

imation of R

ue

that does not contain it. For LTI systems, x

0

does not belong to the

reachable set if there exists P satisfying x

T

0

Px

0

> 1 and (6.4). This is an LMIP. Of

course, if (A, B

w

) is controllable, then x

0

does not belong to the reachable set if and

only if x

T

0

W

−1

c

x

0

> 1, where W

c

is deﬁned by the equation (6.6).

For PLDIs, a suﬃcient condition for a point x

0

to lie outside the reachable set

R

ue

can be checked via an LMIP in the variable P: x

T

0

Px

0

> 1 and (6.8). For NLDIs,

x

0

lies outside the reachable set R

ue

if there exist P and λ satisfying x

T

0

Px

0

> 1

and (6.10), another LMIP.

Testing if the reachable set lies outside a half-space

The reachable set R

ue

lies outside the half-space H =

_

ξ ∈ R

n

¸

¸

a

T

ξ > 1

_

if and only

if an ellipsoid E containing R

ue

lies outside H, that is, the minimum value of ξ

T

Q

−1

ξ

over ξ satisfying a

T

ξ > 1 exceeds one. We easily see that

min

_

ξ

T

Q

−1

ξ

¸

¸

a

T

ξ > 1

_

= 1/(a

T

Qa).

Therefore, for LTI systems, H does not intersect the reachable set if there exists

Q satisfying a

T

Qa < 1 and (6.5). This is an LMIP. If (A, B

w

) is controllable, then H

does not intersect the reachable set if and only if a

T

W

c

a < 1, where W

c

is deﬁned by

the equation (6.6).

For PLDIs, a suﬃcient condition for H not to intersect the reachable set R

ue

can be checked via an LMIP in the variable Q: a

T

Qa < 1 and (6.9). For NLDIs, H

does not intersect the reachable set R

ue

if there exists Q and µ satisfying a

T

Qa < 1

and (6.11). This is an LMIP.

Testing if the reachable set is contained in a polytope

Since a polytope can be expressed as an intersection of half-spaces (these determine

its “faces”), we can immediately use the results of the previous subsection in order to

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

6.1 Input-to-State Properties 81

check if the reachable set is contained in a polytope. (Note we only obtain suﬃcient

conditions.) We remark that we can use diﬀerent ellipsoidal outer approximations to

check diﬀerent faces.

6.1.2 Reachable sets with componentwise unit-energy inputs

We now turn to a variation of the previous problem; we consider the set of reachable

states with inputs whose components have unit-energy, that is, we consider the set

R

uce

∆

=

_

¸

_

¸

_

x(T)

¸

¸

¸

¸

¸

¸

¸

x, w satisfy (6.1), x(0) = 0,

_

T

0

w

2

i

dt ≤ 1, i = 1, . . . , n

w

, T ≥ 0

_

¸

_

¸

_

.

Suppose there is a positive-deﬁnite quadratic function V (ξ) = ξ

T

Pξ, and R =

diag(r

1

, . . . , r

nw

) such that

P > 0, R ≥ 0, Tr R = 1,

d

dt

V (x) ≤ w

T

Rw for all x and w satisfying (6.1).

(6.13)

Then the ellipsoid E given by (6.2) contains the reachable set. To prove this, we

integrate both sides of the last inequality from 0 to T, to get

V (x(T)) −V (x(0)) ≤

_

T

0

w

T

Rw dt.

Noting that V (x(0)) = 0 since x(0) = 0, we get

V (x(T)) ≤

_

T

0

w

T

Rw dt ≤ 1.

Let us now consider condition (6.13) for the various LDIs:

• LTI systems: For LTI systems, condition (6.13) is equivalent to the LMI in P

and R:

P > 0, R ≥ 0 and diagonal, Tr R = 1,

_

A

T

P +PA PB

w

B

T

w

P −R

_

≤ 0.

(6.14)

Therefore, if there exist P and R such that (6.14) is satisﬁed, then the ellipsoid E

contains the reachable set R

uce

.

Alternatively, using the variable Q = P

−1

, we can write an equivalent LMI in Q

and R:

Q > 0, R ≥ 0 and diagonal, Tr R = 1,

_

QA

T

+AQ B

w

B

T

w

−R

_

≤ 0.

(6.15)

Remark: It turns out that with LTI systems, the reachable set is equal to the

intersection of all such ellipsoids. (For more details on this, see the Notes and

References.)

This electronic version is for personal use and may not be duplicated or distributed.

82 Chapter 6 Analysis of LDIs: Input/Output Properties

• Polytopic LDIs: For PLDIs, condition (6.13) holds if and only if

P > 0, R ≥ 0 and diagonal, Tr R = 1,

_

A

T

i

P +PA

i

PB

w,i

B

T

w,i

P −R

_

≤ 0, i = 1, . . . , L.

(6.16)

Therefore, if there exist P and R such that (6.16) holds, then the ellipsoid E =

_

ξ | ξ

T

Pξ ≤ 1

_

contains R

uce

.

With Q = P

−1

, condition (6.13) is equivalent to

Q > 0, R ≥ 0 and diagonal, Tr R = 1,

_

QA

T

i

+A

i

Q B

w,i

B

T

w,i

−R

_

≤ 0, i = 1, . . . , L.

(6.17)

• Norm-bound LDIs: For NLDIs, applying the S-procedure, condition (6.13) holds

if and only if there exist P, λ and R such that

P > 0, R ≥ 0 and diagonal, Tr R = 1, λ ≥ 0,

_

¸

_

A

T

P +PA+λC

T

q

C

q

PB

p

PB

w

B

T

p

P −λI 0

B

T

w

P 0 −R

_

¸

_ ≤ 0.

(6.18)

Therefore, if there exists P, λ and R such that (6.18) is satisﬁed, then the ellipsoid E

contains the reachable set R

uce

.

With Q = P

−1

, condition (6.13) is also equivalent to the existence of Q, R and µ

such that

Q > 0, R ≥ 0 and diagonal, Tr R = 1, µ ≥ 0,

_

¸

_

AQ+QA

T

+µB

p

B

T

p

QC

T

q

B

w

C

q

Q −µI 0

B

T

w

0 −R

_

¸

_ ≤ 0.

(6.19)

As in the previous section, we can ﬁnd the minimum volume and minimum di-

ameter ellipsoids among all ellipsoids of the form (6.2), satisfying condition (6.13), by

solving CPs and EVPs, respectively. We can also check that a point does not belong

to the reachable set or that a half-space does not intersect the reachable set, by solving

appropriate LMIPs.

6.1.3 Reachable sets with unit-peak inputs

We consider reachable sets with inputs w that satisfy w

T

w ≤ 1. Thus, we are inter-

ested in the set

R

up

∆

=

_

x(T)

¸

¸

¸

¸

¸

x, w satisfy (6.1), x(0) = 0,

w

T

w ≤ 1, T ≥ 0

_

.

Suppose that there exists a quadratic function V (ξ) = ξ

T

Pξ with

P > 0, and dV (x)/dt ≤ 0 for all x, w

satisfying (6.1), w

T

w ≤ 1 and V (x) ≥ 1.

(6.20)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

6.1 Input-to-State Properties 83

Then, the ellipsoid E given by (6.2) contains the reachable set R

up

. Let us consider

now condition (6.20) for the various LDIs:

• LTI systems: For LTI systems, condition (6.20) is equivalent to P > 0 and

ξ

T

(A

T

P +PA)ξ +ξ

T

PB

w

w +w

T

B

T

w

Pξ ≤ 0

for any ξ and w satisfying

w

T

w ≤ 1 and ξ

T

Pξ ≥ 1.

Using the S-procedure, we conclude that condition (6.20) holds if there exist α ≥ 0

and β ≥ 0 such that for all x and w,

_

x

w

_

T

_

A

T

P +PA+αP PB

w

B

T

w

P −βI

__

x

w

_

+β −α ≤ 0, (6.21)

or equivalently

_

¸

_

A

T

P +PA+αP PB

w

0

B

T

w

P −βI 0

0 0 β −α

_

¸

_ ≤ 0. (6.22)

Clearly we must have α ≥ β. Next, if (6.22) holds for some (α

0

, β

0

), then it holds for

all α

0

≥ β ≥ β

0

. Therefore, we can assume without loss of generality that β = α, and

rewrite (6.21) as

_

A

T

P +PA+αP PB

w

B

T

w

P −αI

_

≤ 0. (6.23)

Therefore, if there exists P and α satisfying

P > 0, α ≥ 0, and (6.23), (6.24)

then the ellipsoid E contains the reachable set R

up

. Note that inequality (6.23) is not

an LMI in α and P; however, for ﬁxed α it is an LMI in P.

Alternatively, condition (6.23) is equivalent to the following inequality in Q = P

−1

and α:

_

QA

T

+AQ+αQ B

w

B

T

w

−αI

_

≤ 0. (6.25)

• Polytopic LDIs: Condition (6.20) holds if P > 0 and there exists α ≥ 0 satisfying

_

A

T

i

P +PA

i

+αP PB

w,i

B

T

w,i

P −αI

_

≤ 0, i = 1, . . . , L. (6.26)

Therefore, if there exists P and α satisfying

P > 0, α ≥ 0, and (6.26), (6.27)

then the ellipsoid E contains the reachable set R

up

. Once again, we note that condi-

tion (6.26) is not an LMI in α and P, but is an LMI in P for ﬁxed α. We can also

rewrite condition (6.26) in terms of Q = P

−1

as the equivalent inequality

_

QA

T

i

+A

i

Q+αQ B

w,i

B

T

w,i

−αI

_

≤ 0, i = 1, . . . , L. (6.28)

This electronic version is for personal use and may not be duplicated or distributed.

84 Chapter 6 Analysis of LDIs: Input/Output Properties

• Norm-bound LDIs: For NLDIs, condition (6.20) holds if and only if P > 0 and

ξ

T

(A

T

P +PA)ξ + 2ξ

T

P(B

w

ω +B

p

π) ≤ 0

for every ω, ξ and π satisfying

ω

T

ω ≤ 1, ξ

T

Pξ ≥ 1, π

T

π −ξ

T

C

T

q

C

q

ξ ≤ 0.

Therefore, using the S-procedure, a suﬃcient condition for (6.20) to hold is that there

exist α ≥ 0 and λ ≥ 0 such that

_

¸

_

A

T

P +PA+αP +λC

T

q

C

q

PB

p

PB

w

B

T

p

P −λI 0

B

T

w

P 0 −αI

_

¸

_ ≤ 0. (6.29)

If there exists P, α and λ satisfying

P > 0, α ≥ 0, λ ≥ 0, and (6.29), (6.30)

the ellipsoid E

P

−1 contains the reachable set R

up

. We note again that inequality (6.29)

is not an LMI in P and α, but is an LMI in P for ﬁxed α.

Deﬁning Q = P

−1

, condition (6.20) is also implied by the existence of µ ≥ 0 and

α ≥ 0 such that

_

¸

_

QA

T

+AQ+αQ+µB

T

p

B

p

QC

T

q

B

w

C

q

Q −µI 0

B

T

w

0 −αI

_

¸

_ ≤ 0. (6.31)

(In fact, the existence of λ ≥ 0 and α ≥ 0 satisfying (6.29) is equivalent to the existence

of µ ≥ 0 and α ≥ 0 satisfying (6.31).)

For ﬁxed α, we can minimize the volume or diameter among all ellipsoids E satis-

fying condition (6.24) for LTI systems, condition (6.27) for PLDIs and condition (6.30)

for NLDIs by forming the appropriate CPs, or EVPs, respectively.

Remark: Inequality (6.23) can also be derived by combining the results of ¸6.1.1

with an exponential time-weighting of the input w. See the Notes and References

for details.

As in ¸6.1.2, it is possible to determine outer ellipsoidal approximations of the

reachable set of LTI systems, PLDIs and NLDIs, subject to componentwise unit-

peak inputs.

6.2 State-to-Output Properties

We now consider state-to-output properties for the LDI

˙ x = A(t)x, z = C

z

(t)x (6.32)

where

_

A(t)

C

z

(t)

_

∈ Ω. (6.33)

• LTI systems: LTI systems have the form ˙ x = Ax, z = C

z

x.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

6.2 State-to-Output Properties 85

• Polytopic LDIs: PLDIs have the form ˙ x = A(t)x, z = C

z

(t)x, where

_

A(t)

C

z

(t)

_

∈ Co

__

A

1

C

1

_

, . . . ,

_

A

L

C

L

__

.

• Norm-bound LDIs: NLDIs have the form

˙ x = Ax +B

p

p, q = C

q

x, z = C

z

x p = ∆(t)q, ∆(t) ≤ 1,

or equivalently

˙ x = Ax +B

p

p, q = C

q

x, z = C

z

x, p

T

p ≤ q

T

q.

• Diagonal Norm-bound LDIs: DNLDIs have the form

˙ x = Ax +B

p

p, q = C

q

x,

z = C

z

x, p

i

= δ

i

(t)q

i

, |δ

i

(t)| ≤ 1, i = 1, . . . , n

q

,

or equivalently

˙ x = Ax +B

p

p, q = C

q

x,

z = C

z

x, |p

i

| ≤ |q

i

|, i = 1, . . . , n

q

.

6.2.1 Bounds on output energy

We seek the maximum output energy given a certain initial state,

max

__

∞

0

z

T

z dt

¸

¸

¸

¸

˙ x = A(t)x, z = C

z

(t)x

_

, (6.34)

where x(0) is given, and the maximum is taken over A(t), C

z

(t) such that (6.33) holds.

Suppose there exists a quadratic function V (ξ) = ξ

T

Pξ such that

P > 0 and

d

dt

V (x) ≤ −z

T

z, for every x and z satisfying (6.32). (6.35)

Then, integrating both sides of the second inequality in (6.35) from 0 to T, we get

V (x(T)) −V (x(0)) ≤ −

_

T

0

z

T

z dt.

for every T ≥ 0. Since V (x(T)) ≥ 0, we conclude that V (x(0)) = x(0)

T

Px(0) is an

upper bound on the maximum energy of the output z given the initial condition x(0).

We now derive LMIs that provide upper bounds on the output energy for the

various LDIs.

• LTI systems: In the case of LTI systems, condition (6.35) is equivalent to

P > 0, A

T

P +PA+C

T

z

C

z

≤ 0. (6.36)

Therefore, we obtain the best upper bound on the output energy provable via quadratic

functions by solving the EVP

minimize x(0)

T

Px(0)

subject to P > 0, (6.36)

(6.37)

This electronic version is for personal use and may not be duplicated or distributed.

86 Chapter 6 Analysis of LDIs: Input/Output Properties

In this case the solution can be found analytically, and it is exactly equal to the output

energy, which is x(0)

T

W

o

x(0), where W

o

is the observability Gramian of the system,

deﬁned by

W

o

∆

=

_

∞

0

e

A

T

t

C

T

z

C

z

e

At

dt.

Since W

o

satisﬁes the Lyapunov equation

A

T

W

o

+W

o

A+C

T

z

C

z

= 0, (6.38)

it satisﬁes (6.35).

With Q = P

−1

, condition (6.35) is equivalent to

Q > 0,

_

AQ+QA

T

QC

T

z

C

z

Q −I

_

≤ 0. (6.39)

If Q satisﬁes (6.39), then an upper bound on the output energy (6.34) is x(0)

T

Q

−1

x(0).

• Polytopic LDIs: For PLDIs, condition (6.35) is equivalent to

P > 0, A(t)

T

P +PA(t) +C

z

(t)

T

C

z

(t) ≤ 0 for all t ≥ 0. (6.40)

Inequalities (6.40) hold if and only if the following LMI in the variable P holds:

P > 0, A

T

i

P +PA

i

+C

T

z,i

C

z,i

≤ 0, i = 1, . . . , L. (6.41)

Therefore the EVP corresponding to ﬁnding the best upper bound on the output

energy provable via quadratic functions is

minimize x(0)

T

Px(0)

subject to P > 0, (6.41)

(6.42)

Deﬁning Q = P

−1

, condition (6.35) is also equivalent to the LMI in Q

Q > 0,

_

A

i

Q+QA

T

i

QC

T

z,i

C

z,i

Q −I

_

≤ 0.

• Norm-bound LDIs: For NLDIs, condition (6.35) is equivalent to P > 0 and

ξ

T

(A

T

P +PA+C

T

z

C

z

)ξ + 2ξ

T

PB

p

π ≤ 0

for every ξ and π that satisfy π

T

π ≤ ξ

T

C

T

q

C

q

ξ. Using the S-procedure, an equivalent

condition is the existence of P and λ such that

P > 0, λ ≥ 0,

_

A

T

P +PA+C

T

z

C

z

+λC

T

q

C

q

PB

p

B

T

p

P −λI

_

≤ 0.

(6.43)

Therefore we obtain the smallest upper bound on the output energy (6.34) provable

via quadratic functions satisfying condition (6.35) by solving the following EVP in the

variables P and λ

minimize x(0)

T

Px(0)

subject to P > 0, λ ≥ 0, (6.43)

(6.44)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

6.2 State-to-Output Properties 87

With Q = P

−1

, condition (6.35) is also equivalent to the existence of µ ≥ 0 satisfying

Q > 0,

_

¸

_

AQ+QA

T

+µB

p

B

T

p

QC

T

z

QC

T

q

C

z

Q −I 0

C

q

Q 0 −µI

_

¸

_ ≤ 0.

• Diagonal Norm-bound LDIs: For DNLDIs, condition (6.35) is equivalent to

ξ

T

(A

T

P +PA+C

T

z

C

z

)ξ + 2ξ

T

PB

p

π ≤ 0,

for every ξ and π that satisfy

π

2

i

≤ (C

q,i

ξ)

T

(C

q,i

ξ) , i = 1, . . . , n

q

.

Using the S-procedure, a suﬃcient condition is that there exists a diagonal Λ ≥ 0

such that

_

A

T

P +PA+C

T

z

C

z

+C

T

q

ΛC

q

PB

p

B

T

p

P −Λ

_

≤ 0. (6.45)

Therefore we obtain the smallest upper bound on the output energy (6.34) provable

via quadratic functions satisfying condition (6.35) and the S-procedure by solving the

following EVP in the variables P and Λ:

minimize x(0)

T

Px(0)

subject to P > 0, Λ ≥ 0 and diagonal, (6.45)

With Q = P

−1

, condition (6.35) is also equivalent to the existence of a diagonal matrix

M ≥ 0 satisfying

_

¸

_

AQ+QA

T

+B

p

MB

T

p

QC

T

z

QC

T

q

C

z

Q −I 0

C

q

Q 0 −M

_

¸

_ ≤ 0.

Once again, we leave it to the reader to extend the results in the remainder of the

section to DNLDIs.

Maximum output energy extractable from a set

As an extension, we seek bounds on the maximum extractable energy from x(0), which

is known only to lie in a polytope P = Co{v

1

, . . . , v

p

}.

We ﬁrst note that since P > 0, x(0)

T

Px(0) takes its maximum value on one of

the vertices v

1

, . . . , v

p

. Therefore, the smallest upper bound on the extractable energy

from P provable via quadratic Lyapunov functions for any of our LDIs is computed

by solving (6.37), (6.42) or (6.44) p times, successively setting x(0) = v

i

, i = 1, . . . , p.

The maximum of the p resulting optimal values is the smallest upper bound sought.

As a variation, suppose that x(0) is only known to belong to the ellipsoid E =

{ξ | ξ

T

Xξ ≤ 1}. Then, the smallest upper bound on the maximum output energy

extractable from E provable via quadratic functions satisfying (6.35) is obtained by

replacing the objective function x(0)

T

Px(0) in the EVPs (6.37), (6.42), and (6.44) by

the objective function λ

max

(X

−1/2

PX

−1/2

), and solving the corresponding EVPs. For

LTI systems, the exact value of the maximum extractable output energy is obtained.

This electronic version is for personal use and may not be duplicated or distributed.

88 Chapter 6 Analysis of LDIs: Input/Output Properties

Finally, if x

0

is a random variable with Ex

0

x

T

0

= X

0

, then we obtain the small-

est upper bound on the expected output energy provable via quadratic functions by

replacing the objective function x(0)

T

Px(0) in the EVPs (6.37), (6.42), and (6.44) by

the objective function Tr X

0

P and solving the corresponding EVPs.

6.2.2 Bounds on output peak

It is possible to derive bounds on z(t) using invariant ellipsoids.

Assume ﬁrst that the initial condition x(0) is known. Suppose E = {ξ | ξ

T

Pξ ≤ 1}

is an invariant ellipsoid containing x(0) for the LDI (6.32). Then

z(t)

T

z(t) ≤ max

ξ∈E

ξ

T

C

z

(t)

T

C

z

(t)ξ

for all t ≥ 0. We can express max

ξ∈E

ξ

T

C

z

(t)

T

C

z

(t)ξ as the square root of the

minimum of δ subject to

_

P C

T

z

C

z

δI

_

≥ 0. (6.46)

• LTI systems: The smallest bound on the output peak that can be obtained via

invariant ellipsoids is the square root of the optimal value of the EVP in the variables

P and δ

minimize δ

subject to P > 0, x(0)

T

Px(0) ≤ 1,

(6.46), A

T

P +PA ≤ 0

(6.47)

• Polytopic LDIs: For PLDIs, we obtain the smallest upper bound on z provable

via invariant ellipsoids by taking the square root of the optimal value of the following

EVP in the variables P and δ

minimize δ

subject to P > 0, x(0)

T

Px(0) ≤ 1, (6.46),

A

T

i

P +PA

i

≤ 0, i = 1, . . . , L

• Norm-bound LDIs: For NLDIs, we obtain the smallest upper bound on z

provable via invariant ellipsoids by taking the square root of the optimal value of the

following EVP in the variables P, λ and δ

minimize δ

subject to P > 0, x(0)

T

Px(0) ≤ 1, λ ≥ 0, (6.46),

_

_

A

T

P +PA+λC

T

q

C

q

PB

p

B

T

p

P −λI

_

_

≤ 0

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

6.3 Input-to-Output Properties 89

Remark: As variation of this problem, we can impose a decay rate constraint on

the output, that is, given α > 0, we can compute an upper bound on the smallest

γ such that the output z satisﬁes

|z(t)| ≤ γe

−αt

,

for all t ≥ 0. This also reduces to an EVP.

6.3 Input-to-Output Properties

We ﬁnally consider the input-output behavior of the LDI

˙ x = A(t)x +B

w

(t)w, x(0) = x

0

,

z = C

z

(t)x +D

zw

(t)w,

(6.48)

where

_

A(t) B

w

(t)

C

z

(t) D

zw

(t)

_

∈ Ω. (6.49)

6.3.1 Hankel norm bounds

In this section, we assume that D

zw

(t) = 0 for all t, and consider the quantity

φ

∆

= max

_

_

_

_

∞

T

z

T

z dt

¸

¸

¸

¸

¸

¸

_

T

0

w

T

w dt ≤ 1, x(0) = 0,

w(t) = 0, t > T ≥ 0

_

_

_

,

and the maximum is taken over A(t), B(t) and C(t) satisfying (6.49). For an LTI

system,

√

φ equals the Hankel norm, a name that we extend for convenience to LDIs.

An upper bound for φ can be computed by combining the ellipsoidal bounds on

reachable sets from §6.1.1 and the bounds on the output energy from §6.2.1. Suppose

that P > 0 and Q > 0 satisfy

d

dt

_

x

T

Px

_

≤ w

T

w,

d

dt

_

x

T

Qx

_

≤ −z

T

z (6.50)

for all x, w and z satisfying (6.48). Then, from the arguments in §6.1.1 and §6.2.1, we

conclude that

φ ≤ sup

__

∞

T

z

T

z dt

¸

¸

x(T)

T

Px(T) ≤ 1

_

≤ sup

_

x(T)

T

Qx(T)

¸

¸

x(T)

T

Px(T) ≤ 1

_

= λ

max

(P

−1/2

QP

−1/2

).

• LTI systems: We compute the smallest upper bound on the Hankel norm provable

via quadratic functions satisfying (6.50) by computing the square root of the minimum

γ such that there exist P > 0 and Q > 0 satisfying

A

T

P +PA+PB

w

B

T

w

P ≤ 0,

A

T

Q+QA+C

T

z

C

z

≤ 0,

γP −Q ≥ 0.

This electronic version is for personal use and may not be duplicated or distributed.

90 Chapter 6 Analysis of LDIs: Input/Output Properties

This is a GEVP in the variables γ, P, and Q. It is possible to transform it into an EVP

by introducing the new variable

˜

Q = Q/γ. The corresponding EVP in the variables

γ, P, and

˜

Q is then to minimize γ subject to

A

T

P +PA+PB

w

B

T

w

P ≤ 0,

A

T

˜

Q+

˜

QA+C

T

z

C

z

/γ ≤ 0,

P −

˜

Q ≥ 0

In this case, the optimal value is exactly the Hankel norm. It can be found analytically

as λ

max

(W

c

W

o

) where W

c

and W

o

are the controllability and observability Gramians,

i.e., the solutions of the Lyapunov equations (6.6) and (6.38), respectively.

• Polytopic LDIs: We compute the smallest upper bound on the Hankel norm

provable via quadratic functions satisfying (6.50) by computing the square root of the

minimum γ such that there exist P > 0 and Q > 0 satisfying

A

T

i

P +PA

i

+PB

w,i

B

T

w,i

P ≤ 0,

A

T

i

Q+QA

i

+C

T

z,i

C

z,i

≤ 0, i = 1, . . . , L

γP −Q ≥ 0.

This is a GEVP in the variables γ, P, and Q. Deﬁning

˜

Q = Q/γ, an equivalent EVP

in the variables γ, P, and

˜

Q is to minimize γ subject to

A

T

i

P +PA

i

+PB

w,i

B

T

w,i

P ≤ 0,

A

T

i

˜

Q+

˜

QA

i

+C

T

z,i

C

z,i

/γ ≤ 0, i = 1, . . . , L

P −

˜

Q ≥ 0

• Norm-bound LDIs: Similarly, the smallest upper bound on the Hankel norm

provable via quadratic functions satisfying (6.50) is computed by taking the square

root of the minimum γ such that there exist P > 0, Q > 0, λ ≥ 0, and µ ≥ 0 satisfying

_

¸

_

A

T

P +PA+λC

T

q

C

q

PB

p

PB

w

B

T

p

P −λI 0

B

T

w

P 0 −I

_

¸

_ ≤ 0,

_

A

T

Q+QA+C

T

z

C

z

+µC

T

q

C

q

QB

p

B

T

p

Q −µI

_

≤ 0,

γP −Q ≥ 0.

This is a GEVP in the variables γ, P, Q, λ, and µ. If we introduce the new variables

˜

Q = Q/γ, ν = µ/γ and equivalent EVP in the variables γ, P,

˜

Q, λ, and ν is to

minimize γ such that

_

¸

_

A

T

P +PA+λC

T

q

C

q

PB

p

PB

w

B

T

p

P −λI 0

B

T

w

P 0 −I

_

¸

_ ≤ 0,

_

A

T

˜

Q+

˜

QA+C

T

z

C

z

/γ +νC

T

q

C

q

˜

QB

p

B

T

p

˜

Q −νI

_

≤ 0,

P −

˜

Q ≥ 0.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

6.3 Input-to-Output Properties 91

6.3.2 L

2

and RMS gains

We assume D

zw

(t) = 0 for simplicity of exposition. We deﬁne the L

2

gain of the

LDI (6.48) as the quantity

sup

w2=0

z

2

w

2

where the L

2

norm of u is u

2

2

=

_

∞

0

u

T

u dt, and the supremum is taken over

all nonzero trajectories of the LDI, starting from x(0) = 0. The LDI is said to be

nonexpansive if its L

2

gain is less than one.

Now, suppose there exists a quadratic function V (ξ) = ξ

T

Pξ, P > 0, and γ ≥ 0

such that for all t,

d

dt

V (x) +z

T

z −γ

2

w

T

w ≤ 0 for all x and w satisfying (6.48). (6.51)

Then the L

2

gain of the LDI is less than γ. To show this, we integrate (6.51) from 0

to T, with the initial condition x(0) = 0, to get

V (x(T)) +

_

T

0

_

z

T

z −γ

2

w

T

w

_

dt ≤ 0.

since V (x(T)) ≥ 0, this implies

z

2

w

2

≤ γ.

Remark: It can be easily checked that if (6.51) holds for V (ξ) = ξ

T

Pξ, P > 0,

then γ is also an upper bound on the RMS gain of the LDI, where the root-mean-

square (RMS) value of ξ is deﬁned as

RMS(ξ)

∆

=

_

limsup

T→∞

1

T

_

T

0

ξ

T

ξ dt

_

1/2

,

and the RMS gain is deﬁned as

sup

RMS(w)=0

RMS(z)

RMS(w)

.

Now reconsider condition (6.51) for LDIs.

• LTI systems: Condition (6.51) is equivalent to

_

A

T

P +PA+C

T

z

C

z

PB

w

B

T

w

P −γ

2

I

_

≤ 0. (6.52)

Therefore, we compute the smallest upper bound on the L

2

gain of the LTI system

provable via quadratic functions by minimizing γ over the variables P and γ satisfying

conditions P > 0 and (6.52). This is an EVP.

Remark: Assuming that (A, Bw, Cz) is minimal, this EVP gives the exact value

of the L2 gain of the LTI system, which also equals the H∞ norm of its transfer

matrix, |Cz(sI −A)

−1

Bw|∞.

This electronic version is for personal use and may not be duplicated or distributed.

92 Chapter 6 Analysis of LDIs: Input/Output Properties

Assume B

w

= 0 and C

z

= 0; then the existence of P > 0 satisfying (6.52) is

equivalent to the existence of Q > 0 satisfying

_

AQ+QA

T

+B

w

B

T

w

/γ

2

QC

T

z

C

z

Q −I

_

≤ 0.

• Polytopic LDIs: Condition (6.51) is equivalent to

_

A

T

i

P +PA

i

+C

T

z,i

C

z,i

PB

w,i

B

T

w,i

P −γ

2

I

_

≤ 0, i = 1, . . . , L. (6.53)

Assume there exists i

0

for which B

w,i0

= 0, and j

0

for which C

z,j0

= 0. Then there

exists P > 0 satisfying (6.53) if and only if there exists Q > 0 satisfying

_

QA

T

i

+A

i

Q+B

w,i

B

T

w,i

QC

T

z,i

C

z,i

Q −γ

2

I

_

≤ 0, i = 1, . . . , L.

We get the smallest upper bound on the L

2

gain provable via quadratic functions by

minimizing γ (over γ and P) subject to (6.53) and P > 0, which is an EVP.

• Norm-bound LDIs: For NLDIs, condition (6.51) is equivalent to

ξ

T

(A

T

P +PA+C

T

z

C

z

)ξ + 2ξ

T

P (B

p

π +B

w

w) −γ

2

w

T

w ≤ 0

for all ξ and π satisfying

π

T

π ≤ ξ

T

C

T

q

C

q

ξ.

This is true if and only if there exists λ ≥ 0 such that

_

¸

_

A

T

P +PA+C

T

z

C

z

+λC

T

q

C

q

PB

p

PB

w

B

T

p

P −λI 0

B

T

w

P 0 −γ

2

I

_

¸

_ ≤ 0. (6.54)

Therefore, we obtain the best upper bound on the L

2

gain provable via quadratic

functions by minimizing γ over the variables γ, P and λ, subject to P > 0, λ ≥ 0

and (6.54).

If B

w

= 0 and C

z

= 0, then the existence of P > 0 and λ ≥ 0 satisfying (6.54) is

equivalent to the existence of Q > 0 and µ ≥ 0 satisfying

_

¸

_

QA

T

+AQ+B

w

B

T

w

+µB

p

B

T

p

QC

T

q

QC

T

z

C

q

Q −µI 0

C

z

Q 0 −γ

2

I

_

¸

_ ≤ 0.

• Diagonal Norm-bound LDIs: For DNLDIs, condition (6.51) is equivalent to

ξ

T

(A

T

P +PA+C

T

z

C

z

)ξ + 2ξ

T

P (B

p

π +B

w

w) −γ

2

w

T

w ≤ 0

for all ξ and π satisfying

π

T

i

π

i

≤ (C

q,i

ξ

i

)

T

(C

q,i

ξ

i

), i = 1, . . . , n

q

.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

6.3 Input-to-Output Properties 93

Using the S-procedure, a suﬃcient condition is that there exist a diagonal Λ ≥ 0 such

that

_

¸

_

A

T

P +PA+C

T

z

C

z

+C

T

q

ΛC

q

PB

p

PB

w

B

T

p

P −Λ 0

B

T

w

P 0 −γ

2

I

_

¸

_ ≤ 0. (6.55)

We obtain the best such upper bound by minimizing γ over the variables γ, P and λ,

subject to P > 0, λ ≥ 0 and (6.55).

If B

w

= 0 and C

z

= 0, then the existence of P > 0 and λ ≥ 0 satisfying (6.54) is

equivalent to the existence of Q > 0 and M = diag(µ

1

, . . . , µ

nq

) ≥ 0 satisfying

_

¸

_

QA

T

+AQ+B

w

B

T

w

+B

p

MB

T

p

QC

T

q

QC

T

z

C

q

Q −M 0

C

z

Q 0 −γ

2

I

_

¸

_ ≤ 0.

6.3.3 Dissipativity

The LDI (6.48) is said to be passive if every solution x with x(0) = 0 satisﬁes

_

T

0

w

T

z dt ≥ 0

for all T ≥ 0. It is said to have dissipation η if

_

T

0

_

w

T

z −ηw

T

w

_

dt ≥ 0 (6.56)

holds for all trajectories with x(0) = 0 and all T ≥ 0. Thus passivity corresponds to

nonnegative dissipation. The largest dissipation of the system, i.e., the largest number

η such that (6.56) holds, will be called its dissipativity.

Suppose that there is a quadratic function V (ξ) = ξ

T

Pξ, P > 0, such that

for all x and w satisfying (6.48),

d

dt

V (x) −2w

T

z + 2ηw

T

w ≤ 0. (6.57)

Then, integrating (6.57) from 0 to T with initial condition x(0) = 0 yields

V (x(T)) −

_

T

0

_

2w

T

z −2ηw

T

w

_

dt ≤ 0.

Since V (x(T)) ≥ 0, we conclude

_

T

0

_

w

T

z −ηw

T

w

_

dt ≥ 0,

which implies that the dissipativity of the LDI is at least η. Now reconsider condi-

tion (6.57) for the various LDIs.

• LTI systems: For LTI systems, condition (6.57) is equivalent to

_

A

T

P +PA PB

w

−C

T

z

B

T

w

P −C

z

2ηI −(D

T

zw

+D

zw

)

_

≤ 0. (6.58)

We ﬁnd the largest dissipation that can be guaranteed using quadratic functions by

maximizing η over the variables P and η, subject to P > 0 and (6.58), an EVP.

This electronic version is for personal use and may not be duplicated or distributed.

94 Chapter 6 Analysis of LDIs: Input/Output Properties

Assuming that the system (A, B

w

, C

z

) is minimal, the optimal value of this EVP

is exactly equal to the dissipativity of the system, which can be expressed in terms of

the transfer matrix H(s) = C

z

(sI −A)

−1

B

w

+D

zw

as

inf

Res>0

λ

min

(H(s) +H(s)

∗

)

2

. (6.59)

This follows from the PR lemma.

Deﬁning Q = P

−1

, condition (6.58) is equivalent to

_

QA

T

+AQ B

w

−QC

T

z

B

T

w

−C

z

Q 2ηI −(D

T

zw

+D

zw

)

_

≤ 0. (6.60)

• Polytopic LDIs: For PLDIs, condition (6.57) is equivalent to

_

A

T

i

P +PA

i

PB

w,i

−C

T

z,i

B

T

w,i

P −C

z,i

2ηI −(D

T

zw,i

+D

zw,i

)

_

≤ 0, i = 1, . . . , L. (6.61)

We ﬁnd the largest dissipation that can be guaranteed via quadratic functions by

maximizing η over the variables P and η satisfying P > 0 and (6.61). This is an EVP;

its optimal value is a lower bound on the dissipativity of the PLDI.

If we deﬁne Q = P

−1

, condition (6.57) is equivalent to

_

QA

T

i

+A

i

Q B

w,i

−QC

T

z,i

B

T

w,i

−C

z,i

Q 2ηI −(D

T

zw,i

+D

zw,i

)

_

≤ 0, i = 1, . . . , L. (6.62)

• Norm-bound LDIs: For NLDIs, using the S-procedure, the condition (6.57) is

true if and only if there exists a nonnegative scalar λ such that

_

¸

_

A

T

P +PA+λC

T

q

C

q

PB

w

−C

T

z

PB

p

B

T

w

P −C

z

2ηI −(D

T

zw

+D

zw

) 0

B

T

p

P 0 −λI

_

¸

_ ≤ 0. (6.63)

Therefore we can ﬁnd the largest dissipation provable with a quadratic Lyapunov

function by maximizing η (over η and P), subject to P > 0 and (6.63). This is an

EVP. Deﬁning Q = P

−1

, condition (6.57) is equivalent to the existence of µ ≥ 0

satisfying

_

¸

_

QA

T

+AQ+µB

p

B

T

p

B

w

−QC

T

z

QC

T

q

B

T

w

−C

z

Q 2ηI −(D

T

zw

+D

zw

) 0

C

q

Q 0 −µI

_

¸

_ ≤ 0. (6.64)

6.3.4 Diagonal scaling for componentwise results

We assume for simplicity that D

zw

(t) = 0. Assuming that the system (6.48) has as

many inputs as outputs, we consider the system

˙ x = A(t)x +B

w

(t)T

−1

˜ w, x(0) = 0,

˜ z = TC

z

(t)x.

(6.65)

where T is a positive-deﬁnite diagonal matrix, which has the interpretation of a scal-

ing. We will see that scaling enables us to conclude “componentwise” results for the

system (6.65).

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

6.3 Input-to-Output Properties 95

Consider for example, the “scaled L

2

gain”

α = inf

T diagonal,

T > 0

sup

˜ w2=0

˜ z

2

˜ w

2

.

The scaled L

2

gain has the following interpretation:

max

i=1,...,nz

sup

wi2=0

z

i

w

i

≤ α.

We show this as follows: For every ﬁxed scaling T = diag(t

1

, . . . , t

nz

),

sup

˜ w2=0

˜ z

2

2

˜ w

2

2

= sup

w2=0

nz

i

t

2

i

z

i

2

2

nz

i

t

2

i

w

i

2

2

≥ sup

wi2=0

z

i

2

2

w

i

2

2

.

Since the inequality is true for all diagonal nonsingular T, the desired inequality

follows. We now show how we can obtain bounds on the scaled L

2

gain using LMIs

for LTI systems and LDIs.

• LTI systems: The L

2

gain for the LTI system scaled by T is guaranteed to be

less than γ if there exists P > 0 such that the LMI

_

A

T

P +PA+C

T

z

SC

z

PB

w

B

T

w

P −γ

2

S

_

< 0

holds with S = T

T

T. The smallest scaled L

2

gain of the system can therefore be

computed as a GEVP.

• Polytopic LDIs: For PLDIs, the scaled L

2

gain is guaranteed to be lower than

γ if there exists S > 0, with S diagonal, and P > 0 which satisfy

_

A

T

i

P +PA

i

+C

T

z,i

SC

z,i

PB

w,i

B

T

w,i

P −γ

2

S

_

< 0, i = 1, . . . , L. (6.66)

The optimal scaled L

2

gain γ is therefore obtained by minimizing γ over (γ, P and

S) subject to (6.66), P > 0, S > 0 and S diagonal. This is a GEVP.

• Norm-bound LDIs: For NLDIs, the scaled L

2

gain (with scaling T) is less than γ

if there exist λ ≥ 0, P > 0, and S > 0, S diagonal, such that the following LMI holds:

_

¸

_

A

T

P +PA+C

T

z

SC

z

+λC

T

q

C

q

PB

p

PB

w

B

T

p

P −λI 0

B

T

w

P 0 −γ

2

S

_

¸

_ ≤ 0. (6.67)

Therefore, minimizing γ

2

subject to the constraint (6.67) is an EVP.

This electronic version is for personal use and may not be duplicated or distributed.

96 Chapter 6 Analysis of LDIs: Input/Output Properties

Notes and References

Integral quadratic constraints

In ¸8.2 (and ¸10.9) we consider a generalization of the NLDI in which the pointwise condition

p

T

p ≤ q

T

q is replaced by the integral constraint

_

∞

0

p

T

p dt ≤

_

∞

0

q

T

q dt.

Many of the results derived for NLDIs and DNLDIs then become necessary and suﬃcient.

Integral quadratic constraints were introduced by Yakubovich. For a general study of these,

we refer the reader to the articles of Yakubovich (see [Yak92] and references therein), and

also Megretsky [Meg93, Meg92a, Meg92b].

Reachable sets for componentwise unit-energy inputs

We study in greater detail the set of states reachable with componentwise unit-energy inputs

for LTI systems (see ¸6.1.2). We begin by observing that a point x0 belongs to the reachable

set if and only if the optimal value of the problem

minimize max

i

_

∞

0

w

2

i

dt

subject to ˙ x = −Ax +Bww, x(0) = x0, lim

t→∞

x(t) = 0

is less than one. This is a multi-criterion convex quadratic problem, considered in ¸10.8. In

this case, the problem reduces to checking whether the LMI

x

T

0

Px0 < 1,

_

A

T

P +PA PB

B

T

P −R

_

≤ 0,

P > 0, R > 0, diagonal, Tr R = 1

is feasible. This shows that the reachable set is the intersection of ellipsoids

_

ξ [ ξ

T

Pξ ≤ 1

_

satisfying equation (6.14).

A variation of this problem is considered in [SZ92], where the authors consider inputs w

satisfying

_

∞

0

ww

T

dt ≤ Q, or equivalently

_

∞

0

w

T

Ww dt ≤ Tr WQ for every W ≥ 0.

They get necessary conditions characterizing the corresponding reachable set. We now derive

LMI conditions that are both necessary and suﬃcient.

As before, x0 belongs to the reachable set if and only the optimal value of the problem

minimize

_

∞

0

w

T

Ww dt

subject to ˙ x = −Ax +Bww, x(0) = x0, lim

t→∞

x(t) = 0

is less than Tr WQ for every W ≥ 0. This is equivalent to infeasibility of the LMI (in the

variables P and W):

x

T

0

Px0 −Tr WQ ≥ 0, P > 0,

_

A

T

P +PA PB

B

T

P −W

_

≤ 0.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 97

These are precisely the conditions obtained in [SZ92].

Reachable sets with unit-peak inputs

Schweppe [Sch73, ¸4.3.3] considers the more general problem of time-varying ellipsoidal ap-

proximations of reachable sets with unit-peak inputs for time-varying systems; the positive-

deﬁnite matrix describing the ellipsoidal approximation satisﬁes a matrix diﬀerential equa-

tion, terms of which closely resemble those in the matrix inequalities described in this book.

Sabin and Summers [SS90] study the approximation of the reachable set via quadratic func-

tions. A survey of the available techniques can be found in the article by Gayek [Gay91].

See also [FG88].

The technique used to derive LMI (6.23) can be interpreted as an exponential time-weighting

procedure. Fix α > 0. Then for every T > 0, rewrite LDI (6.1), with new exponentially

time-weighted variables xT (t) = e

α(t−T)/2

x(t) and v(t) = e

α(t−T)/2

w(t) as

˙ xT =

_

A(t) +

α

2

I

_

xT +Bw(t)v, xT (0) = 0.

Since w(t)

T

w(t) ≤ 1, we have

_

T

0

v

T

v dτ ≤ 1/α.

Now suppose that P > 0 satisﬁes condition (6.23), which we rewrite for convenience as

_

(A+αI/2)

T

P +P(A+αI/2) PBw

B

T

w

P −αI

_

≤ 0. (6.68)

The results of ¸6.1.1 imply that xT (T) satisﬁes xT (T)

T

PxT (T) ≤ 1; therefore, x(T)

T

Px(T) ≤

1. Since LMI (6.68) is independent of T, the ellipsoid

_

x [ x

T

Px ≤ 1

_

contains the reachable

set with unit-peak inputs.

Bounds on overshoot

The results of ¸6.1.3 can be used to ﬁnd a bound on the step response peak for LDIs. Consider

a single-input single-output LDI, subject to a unit-step input:

˙ x = A(t)x +bw(t), x(0) = 0,

s = cz(t)x,

_

A(t) bw(t)

cz(t) 0

_

∈ Ω.

Since a unit-step is also a unit-peak input, an upper bound on max

t≥0

[s(t)[ (i.e., the over-

shoot) can be found by combining the results on reachable sets with unit-peak inputs (¸6.1.3)

and bounds on output peak (¸6.2.2).

Suppose there exist Q > 0 and α > 0 satisfying (6.25),

A(t)Q+QA(t)

T

+αQ+

1

α

bw(t)bw(t)

T

< 0.

Then, from ¸6.2.2, we conclude that max

t≥0

[s(t)[ does not exceed M, where M

2

= max

t≥0

cz(t)

T

Qcz(t). This can be used to ﬁnd LMI-based bounds on overshoot for LTI systems,

PLDIs and NLDIs. We note that the resulting bounds can be quite conservative. These

LMIs can be used to determine state-feedback matrices as well (see the Notes and References

of Chapter 7).

This electronic version is for personal use and may not be duplicated or distributed.

98 Chapter 6 Analysis of LDIs: Input/Output Properties

Bounds on impulse response peak for LTI systems

The impulse response h(t) = ce

At

b of the single-input single-output LTI system

˙ x = Ax +bu, x(0) = 0, y = cx

is just the output y with initial condition x(0) = b and zero input. Therefore, we can com-

pute a bound on the impulse response peak by combining the results on invariant ellipsoids

from ¸5.2 with those on the peak value of the output from ¸6.2.2. The bound obtained this

way can be quite conservative. For instance, consider the LTI system with transfer function

H(s) =

1

s +s1

n

i=2

s −si

s +si

,

where si+1 ¸ si > 0, i.e., the dynamics of this system are widely spaced. The bound on

the peak value of the impulse response of the system computed via the EVP (6.47) turns

out to be 2n − 1 times the actual maximum value of the impulse response, where n is the

dimension of a minimal realization of the system [Fer93]. We conjecture that the bound

can be no more conservative, that is, for every LTI system, the bound on the peak of the

impulse response computed via the EVP (6.47) can be no more than 2n−1 times the actual

maximum value of the impulse response.

Examples for which the bound obtained via the EVP (6.47) is sharp include the case of passive

LTI systems. In this case, we know there exists a positive-deﬁnite P such that A

T

P+PA ≤ 0,

Pb = c

T

, so that the maximum value of the impulse response is cb and is attained for t = 0.

This is an illustration of the fact that passive systems have nice “peaking” properties, which

is used in nonlinear control [Kok92].

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Chapter 7

State-Feedback Synthesis for LDIs

7.1 Static State-Feedback Controllers

We consider the LDI

˙ x = A(t)x +B

w

(t)w +B

u

(t)u, z = C

z

(t)x +D

zw

(t)w +D

zu

(t)u,

_

A(t) B

w

(t) B

u

(t)

C

z

(t) D

zw

(t) D

zu

(t)

_

∈ Ω,

(7.1)

where Ω has one of our special forms (i.e., singleton, polytope, image of a unit ball

under a matrix linear-fractional mapping). Here u : R

+

→ R

nu

is the control input

and w : R

+

→R

nw

is the exogenous input signal .

Let K ∈ R

nu×n

, and suppose that u = Kx. Since the control input is a linear

function of the state, this is called (linear, constant) state-feedback, and the matrix K

is called the state-feedback gain. This yields the closed-loop LDI

˙ x = (A(t) +B

u

(t)K) x +B

w

(t)w,

z = (C

z

(t) +D

zu

(t)K) x +D

zw

(t)w.

(7.2)

In this chapter we consider the state-feedback synthesis problem, i.e., the problem of

ﬁnding a matrix K so that the closed-loop LDI (7.2) satisﬁes certain properties or

speciﬁcations, e.g., stability.

Remark: Using the idea of global linearization described in ¸4.3, the methods of

this chapter can be used to synthesize a linear state-feedback for some nonlinear,

time-varying systems. As an example, we can synthesize a state-feedback for the

system

˙ x = f(x, w, u, t), z = g(x, w, u, t),

provided we have

_

¸

_

∂f

∂x

∂f

∂w

∂f

∂u

∂g

∂x

∂g

∂w

∂g

∂u

_

¸

_

(x, w, u, t) ∈ Ω,

and f(0, 0, 0, t) = 0, g(0, 0, 0, t) = 0, for all x, t, w and u.

99

100 Chapter 7 State-Feedback Synthesis for LDIs

7.2 State Properties

We consider the LDI

˙ x = A(t)x +B

u

(t)u, [A(t) B

u

(t)] ∈ Ω. (7.3)

• LTI systems: For LTI systems, (7.3) becomes

˙ x = Ax +B

u

u. (7.4)

• Polytopic LDIs: PLDIs are given by

˙ x = A(t)x +B

u

(t)u, [A(t) B

u

(t)] ∈ Co{[A

1

B

u,1

] , . . . . [A

L

B

u,L

]} . (7.5)

• Norm-bound LDIs: For NLDIs, (7.3) becomes

˙ x = Ax +B

u

u +B

p

p, q = C

q

x +D

qu

u +D

qp

p

p = ∆(t)q, ∆(t) ≤ 1.

(7.6)

Equivalently, we have

˙ x = Ax +B

u

u +B

p

p, q = C

q

x +D

qp

p +D

qu

u, p

T

p ≤ q

T

q.

• Diagonal Norm-bound LDIs: For DNLDIs, equation (7.3) becomes

˙ x = Ax +B

u

u +B

p

p, q = C

q

x +D

qp

p +D

qu

u,

p

i

= δ

i

(t)q

i

, |δ

i

(t)| ≤ 1, i = 1, . . . , L.

(7.7)

Equivalently, we have

˙ x = Ax +B

u

u +B

p

p, q = C

q

x +D

qp

p +D

qu

u,

|p

i

| ≤ |q

i

|, i = 1, . . . , n

q

.

7.2.1 Quadratic stabilizability

The system (7.1) is said to be quadratically stabilizable (via linear state-feedback) if

there exists a state-feedback gain K such that the closed-loop system (7.2) is quadrat-

ically stable (hence, stable). Quadratic stabilizability can be expressed as an LMIP.

• LTI systems: Let us ﬁx the matrix K. The LTI system (7.4) is (quadratically)

stable if and only if there exists P > 0 such that

(A+B

u

K)

T

P +P(A+B

u

K) < 0,

or equivalently, there exists Q > 0 such that

Q(A+B

u

K)

T

+ (A+B

u

K)Q < 0. (7.8)

Neither of these conditions is jointly convex in K and P or Q, but by a simple change

of variables we can obtain an equivalent condition that is an LMI.

Deﬁne Y = KQ, so that for Q > 0 we have K = Y Q

−1

. Substituting into (7.8)

yields

AQ+QA

T

+B

u

Y +Y

T

B

T

u

< 0, (7.9)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

7.2 State Properties 101

which is an LMI in Q and Y . Thus, the system (7.4) is quadratically stabilizable if

and only if there exist Q > 0 and Y such that the LMI (7.9) holds. If this LMI is

feasible, then the quadratic function V (ξ) = ξ

T

Q

−1

ξ proves (quadratic) stability of

system (7.4) with state-feedback u = Y Q

−1

x.

Remark: We will use the simple change of variables Y = KQ many times in

the sequel. It allows us to recast the problem of ﬁnding a state-feedback gain

as an LMIP with variables that include Q and Y ; the state-feedback gain K is

recovered as K = Y Q

−1

.

An alternate equivalent condition for (quadratic) stabilizability, involving fewer

variables, can be derived using the elimination of matrix variables described in §2.6.2:

There exist Q > 0 and a scalar σ such that

AQ+QA

T

−σB

u

B

T

u

< 0. (7.10)

Since we can always assume σ > 0 in LMI (7.10), and since the LMI is homogeneous

in Q and σ, we can without loss of generality take σ = 1, thus reducing the number of

variables by one. If Q > 0 satisﬁes the LMI (7.10), a stabilizing state-feedback gain is

given by K = −(σ/2)B

T

u

Q

−1

.

From the elimination procedure of §2.6.2, another equivalent condition is

˜

B

T

u

_

AQ+QA

T

_

˜

B

u

< 0, (7.11)

where

˜

B

u

is an orthogonal complement of B

u

. For any Q > 0 satisfying (7.11), a

stabilizing state-feedback gain is K = −(σ/2)B

T

u

Q

−1

, where σ is any scalar such

that (7.10) holds (condition (7.11) implies that such a scalar exists).

For LTI systems, these LMI conditions are necessary and suﬃcient for stabiliz-

ability. In terms of linear system theory, stabilizability is equivalent to the condition

that every unstable mode be controllable; it is not hard to show that this is equivalent

to feasibility of the LMIs (7.9), (7.11), or (7.10).

• Polytopic LDIs: For the PLDI (7.5), the same argument applies. Quadratic

stabilizability is equivalent to the existence of Q > 0 and Y with

QA

T

i

+A

i

Q+B

u,i

Y +Y

T

B

T

u,i

< 0, i = 1, . . . , L. (7.12)

• Norm-bound LDIs: With u = Kx, the NLDI (7.6) is quadratically stable (see

LMI (5.14)) if there exist Q > 0 and µ > 0 such that

_

¸

_

_

AQ+QA

T

+B

u

KQ

+QK

T

B

T

u

+µB

p

B

T

p

_

µB

p

D

T

qp

+Q(C

q

+D

qu

K)

T

µD

qp

B

T

p

+ (C

q

+D

qu

K)Q −µ(I −D

qp

D

T

qp

)

_

¸

_ < 0.

This condition has a simple frequency-domain interpretation: The H

∞

norm of the

transfer matrix from p to q for the LTI system ˙ x = (A + B

u

K)x + B

p

p, q = (C

q

+

D

qu

K)x +D

qp

p is less than one.

With Y = KQ, we conclude that the NLDI is quadratically stabilizable if there

exist Q > 0, µ > 0 and Y such that

_

¸

_

_

AQ+QA

T

+µB

p

B

T

p

+B

u

Y +Y

T

B

T

u

_

µB

p

D

T

qp

+QC

T

q

+Y

T

D

T

qu

µD

qp

B

T

p

+C

q

Q+D

qu

Y −µ(I −D

qp

D

T

qp

)

_

¸

_ < 0. (7.13)

This electronic version is for personal use and may not be duplicated or distributed.

102 Chapter 7 State-Feedback Synthesis for LDIs

Using the elimination procedure of §2.6.2, we obtain the equivalent conditions:

_

¸

¸

_

_

AQ+QA

T

−σB

u

B

T

u

+µB

p

B

T

p

_

QC

T

q

+µB

p

D

T

qp

−σB

u

D

T

qu

C

q

Q+µD

qp

B

T

p

−σD

qu

B

T

u

−µ(I −D

qp

D

T

qp

) −σD

qu

D

T

qu

_

¸

¸

_

< 0,

−µ(I −D

qp

D

T

qp

) < 0.

(7.14)

The latter condition is always satisﬁed since the NLDI is well-posed. By homogeneity,

we can set µ = 1. If Q > 0 and σ > 0 satisfy (7.14) with µ = 1, a stabilizing

state-feedback gain is K = (−σ/2)B

T

u

Q

−1

.

An alternate, equivalent condition for quadratic stabilizability is expressed in

terms of the orthogonal complement of [B

T

u

D

T

qu

]

T

, which we denote by

˜

G:

˜

G

T

_

AQ+QA

T

+µB

p

B

T

p

QC

T

q

+µB

p

D

T

qp

C

q

Q+µD

qp

B

T

p

−µ(I −D

qp

D

T

qp

)

_

˜

G < 0. (7.15)

(Again, we can freely set µ = 1 in this LMI.)

In the remainder of this chapter we will assume that D

qp

in (7.6) is zero in order

to simplify the discussion; all the following results can be extended to the case in

which D

qp

is nonzero.

• Diagonal Norm-bound LDIs: With u = Kx, the DNLDI (7.7) is quadratically

stable (see LMI (5.16)) if there exist Q > 0, M = diag(µ

1

, . . . , µ

nq

) > 0 satisfying

_

¸

_

_

AQ+QA

T

+B

u

KQ

+QK

T

B

T

u

+B

p

MB

T

p

_

B

p

MD

T

qp

+Q(C

q

+D

qu

K)

T

D

qp

MB

T

p

+ (C

q

+D

qu

K)Q −(M −D

qp

MD

T

qp

)

_

¸

_ < 0.

This condition leads to a simple frequency-domain interpretation for quadratic stabi-

lizability of DNLDIs: Let H denote the transfer matrix from p to q for the LTI system

˙ x = (A+B

u

K)x+B

p

p, q = (C

q

+D

qu

K)x+D

qp

p. Then the DNLDI is quadratically

stabilizable if there exists K such that for some diagonal positive-deﬁnite matrix M,

M

−1/2

HM

1/2

∞

≤ 1. There is no analytical method for checking this condition.

With Y = KQ, we conclude that the NLDI is quadratically stabilizable if there

exist Q > 0, M > 0 and diagonal, and Y such that

_

¸

_

_

AQ+QA

T

+B

p

MB

T

p

+B

u

Y +Y

T

B

T

u

_

B

p

MD

T

qp

+QC

T

q

+Y

T

D

T

qu

D

qp

MB

T

p

+C

q

Q+D

qu

Y −(M −D

qp

MD

T

qp

)

_

¸

_ < 0.

As before, we can eliminate the variable Y using the elimination procedure of §2.6.2,

and obtain equivalent LMIs in fewer variables.

7.2.2 Holdable ellipsoids

Just as quadratic stability can also be interpreted in terms of invariant ellipsoids, we

can interpret quadratic stabilizability in terms of holdable ellipsoids. We say that the

ellipsoid

E =

_

ξ ∈ R

n

¸

¸

ξ

T

Q

−1

ξ ≤ 1

_

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

7.2 State Properties 103

is holdable for the system (7.3) if there exists a state-feedback gain K such that E

is invariant for the system (7.3) with u = Kx. Therefore the LMIs described in the

previous section also characterize holdable ellipsoids for the system (7.3).

With this parametrization of quadratic stabilizability and holdable ellipsoids as

LMIs in Q > 0 and Y , we can solve various optimization problems, such as ﬁnding

the coordinate transformation that minimizes the condition number of Q, imposing

norm constraints on the input u = Kx, etc.

Remark: In Chapter 5, we saw that quadratic stability and invariant ellipsoids

were characterized by LMIs in a variable P > 0 and also its inverse Q = P

−1

.

In contrast, quadratic stabilizability and ellipsoid holdability can be expressed

as LMIs only in variables Q and Y ; it is not possible to rewrite these LMIs as

LMIs with Q

−1

as a variable. This restricts the extension of some of the results

from Chapter 5 (and Chapter 6) to state-feedback synthesis. As a rule of thumb,

results in the analysis of LDIs that are expressed as LMIs in the variable Q can

be extended to state-feedback synthesis, with a few exceptions; results expressed

as LMIs in P are not. For example, we will not be able to compute the minimum

volume holdable ellipsoid that contains a given polytope 1 (see problem (5.33))

as an optimization problem over LMIs; however, we will be able to compute the

minimum diameter holdable ellipsoid (see problem (5.34)) containing 1.

7.2.3 Constraints on the control input

When the initial condition is known, we can ﬁnd an upper bound on the norm of

the control input u(t) = Kx(t) as follows. Pick Q > 0 and Y which satisfy the

quadratic stabilizability condition (either (7.9), (7.12) or (7.13)), and in addition

x(0)

T

Q

−1

x(0) ≤ 1. This implies that x(t) belongs to E for all t ≥ 0, and conse-

quently,

max

t≥0

u(t) = max

t≥0

Y Q

−1

x(t)

≤ max

x∈E

Y Q

−1

x

= λ

max

(Q

−1/2

Y

T

Y Q

−1/2

).

Therefore, the constraint u(t) ≤ µ is enforced at all times t ≥ 0 if the LMIs

_

1 x(0)

T

x(0) Q

_

≥ 0,

_

Q Y

T

Y µ

2

I

_

≥ 0 (7.16)

hold, where Q > 0 and Y satisfy the stabilizability conditions (either (7.9), (7.12)

or (7.13)).

We can extend this to the case where x(0) lies in an ellipsoid or polytope. For

example suppose we require (7.16) to hold for all x(0 ≤ 1. This is easily shown to

be equivalent to

Q ≥ I,

_

Q Y

T

Y µ

2

I

_

≥ 0.

As another extension we can handle constraints on

u(t)

max

∆

= max

i

|u

i

(t)|

This electronic version is for personal use and may not be duplicated or distributed.

104 Chapter 7 State-Feedback Synthesis for LDIs

in a similar manner:

max

t≥0

u(t)

max

= max

t≥0

Y Q

−1

x(t)

max

≤ max

x∈E

Y Q

−1

x

max

= max

i

_

Y Q

−1

Y

T

_

ii

.

Therefore, the constraint u(t)

max

≤ µ for t ≥ 0 is implied by the LMI

_

1 x(0)

T

x(0) Q

_

≥ 0,

_

X Y

Y

T

Q

_

≥ 0, X

ii

≤ µ

2

,

where once again, Q > 0 and Y satisfy the stabilizability conditions (either (7.9),

(7.12) or (7.13)).

7.3 Input-to-State Properties

We next consider the LDI

˙ x = A(t)x +B

w

(t)w +B

u

(t)u. (7.17)

• LTI systems: For LTI systems, equation (7.17) becomes ˙ x = Ax +B

w

w +B

u

u.

• Polytopic LDIs: PLDIs are given by ˙ x = A(t)x + B

w

(t)w + B

u

(t)u, where

[A(t) B

w

(t) B

u

(t)] ∈ Co{[A

1

B

w,1

B

u,1

] , . . . , [A

L

B

w,L

B

u,L

]}.

• Norm-bound LDIs: For NLDIs, equation (7.17) becomes

˙ x = Ax +B

u

u +B

p

p +B

w

w, q = C

q

x +D

qu

u

p = ∆(t)q, ∆(t) ≤ 1

which we can also express as

˙ x = Ax +B

u

u +B

p

p +B

w

w, q = C

q

x +D

qu

u, p

T

p ≤ q

T

q.

• Diagonal Norm-bound LDIs: For DNLDIs, equation (7.17) becomes

˙ x = Ax +B

u

u +B

p

p +B

w

w, q = C

q

x +D

qu

u

p

i

= δ

i

(t)q

i

, |δ

i

(t)| ≤ 1, i = 1, . . . , L.

Equivalently, we have

˙ x = Ax +B

u

u +B

p

p +B

w

w, q = C

q

x +D

qu

u,

|p

i

| ≤ |q

i

|, i = 1, . . . , n

q

.

7.3.1 Reachable sets with unit-energy inputs

For the system (7.17) with u = Kx, the set of states reachable with unit energy is

deﬁned as

R

ue

∆

=

_

¸

_

¸

_

x(T)

¸

¸

¸

¸

¸

¸

¸

x, w, u satisfy (7.17), u = Kx, x(0) = 0,

_

T

0

w

T

w dt ≤ 1, T ≥ 0

_

¸

_

¸

_

.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

7.3 Input-to-State Properties 105

We will now derive conditions under which there exists a state-feedback gain K guar-

anteeing that a given ellipsoid E = {ξ ∈ R

n

¸

¸

ξ

T

Q

−1

ξ ≤ 1} contains R

ue

.

• LTI systems: From §6.1.1, the ellipsoid E contains R

ue

for the system (7.4) for

some state-feedback gain K if K satisﬁes

AQ+QA

T

+B

u

KQ+QK

T

B

T

u

+B

w

B

T

w

≤ 0.

Setting KQ = Y , we conclude that E ⊇ R

ue

for some K if there exist Y such that

QA

T

+AQ+B

u

Y +Y

T

B

T

u

+B

w

B

T

w

≤ 0.

For any Q > 0 and Y satisfying this LMI, the state-feedback gain K = Y Q

−1

is such

that the ellipsoid E contains the reachable set for the closed-loop system.

Using the elimination procedure of §2.6.2, we eliminate the variable Y to obtain

the LMI in Q > 0 and the variable σ:

QA

T

+AQ−σB

u

B

T

u

+B

w

B

T

w

≤ 0. (7.18)

For any Q > 0 and σ satisfying this LMI, a corresponding state-feedback gain is given

by K = −(σ/2)B

T

u

Q

−1

. Another equivalent condition for E to contain R

ue

is

˜

B

T

u

_

AQ+QA

T

+B

w

B

T

w

_

˜

B

u

≤ 0, (7.19)

where

˜

B

u

is an orthogonal complement of B

u

.

• Polytopic LDIs: For PLDIs, E contains R

ue

for some state-feedback gain K if

there exist Q > 0 and Y such that the following LMI holds (see LMI (6.9)):

QA

T

i

+A

i

Q+B

u,i

Y +Y

T

B

T

u,i

+B

w,i

B

T

w,i

< 0, i = 1, . . . , L.

• Norm-bound LDIs: For NLDIs, E contains R

ue

for some state-feedback gain K

if there exist µ > 0 and Y such that the following LMI holds (see LMI (6.11)):

_

QA

T

+AQ+B

u

Y +Y

T

B

T

u

+B

w

B

T

w

+µB

p

B

T

p

(C

q

Q+D

qu

Y )

T

C

q

Q+D

qu

Y −µI

_

≤ 0.

We can eliminate Y to obtain the LMI in Q > 0 and σ that guarantees that E contains

the reachable set R

ue

for some state-feedback gain K:

_

¸

_

_

AQ+QA

T

−σB

u

B

T

u

+µB

p

B

T

p

+B

w

B

T

w

_

QC

T

q

+µB

p

D

T

qp

−σB

u

D

T

qu

C

q

Q+µD

qp

B

T

p

−σD

qu

B

T

u

−µ(I −D

qp

D

T

qp

) −σD

qu

D

T

qu

_

¸

_ < 0. (7.20)

The corresponding state-feedback gain is K = (−σ/2)B

T

u

Q

−1

.

An equivalent condition is expressed in terms of

˜

G, the orthogonal complement

of [B

T

u

D

T

qu

]

T

:

˜

G

T

_

¸

_

_

AQ+QA

T

+µB

p

B

T

p

+B

w

B

T

w

_

QC

T

q

+µB

p

D

T

qp

C

q

Q+µD

qp

B

T

p

−µ(I −D

qp

D

T

qp

)

_

¸

_

˜

G < 0.

This electronic version is for personal use and may not be duplicated or distributed.

106 Chapter 7 State-Feedback Synthesis for LDIs

• Diagonal Norm-bound LDIs: For DNLDIs, E contains R

ue

for some state-

feedback gain K if there exist Q > 0, M > 0 and diagonal, and Y such that the

following LMI holds (see LMI (6.12)):

_

QA

T

+AQ+B

u

Y +Y

T

B

T

u

+B

w

B

T

w

+B

p

MB

T

p

(C

q

Q+D

qu

Y )

T

C

q

Q+D

qu

Y −M

_

≤ 0.

Using the elimination procedure of §2.6.2, we can eliminate the variable Y to obtain

an equivalent LMI in fewer variables.

Using these LMIs (see §6.1.1 for details):

• We can ﬁnd a state-feedback gain K such that a given point x

0

lies outside the

set of reachable states for the system (7.3).

• We can ﬁnd a state-feedback gain K such that the set of reachable states for the

system (7.3) lies in a given half-space. This result can be extended to check if

the reachable set is contained in a polytope. In this case, in contrast with the

results in Chapter 6, we must use the same outer ellipsoidal approximation to

check diﬀerent faces. (This is due to the coupling induced by the new variable

Y = KQ.)

7.3.2 Reachable sets with componentwise unit-energy inputs

With u = Kx, the set of reachable states for inputs with componentwise unit-energy

for the system (7.17) is deﬁned as

R

uce

∆

=

_

¸

_

¸

_

x(T)

¸

¸

¸

¸

¸

¸

¸

x, w, u satisfy (7.17), u = Kx, x(0) = 0,

_

T

0

w

T

i

w

i

dt ≤ 1, T ≥ 0

_

¸

_

¸

_

.

We now consider the existence of the state-feedback gain K such that the ellipsoid

E = {ξ ∈ R

n

¸

¸

ξ

T

Q

−1

ξ ≤ 1} contains R

uce

for LTI systems and LDIs.

• LTI systems: From LMI (6.15), E ⊇ R

uce

for the LTI system (7.4) if there exists

diagonal R > 0 with unit trace such that the LMI

_

QA

T

+AQ+B

u

Y +Y

T

B

T

u

B

w

B

T

w

−R

_

≤ 0.

• Polytopic LDIs: For PLDIs, E ⊇ R

uce

for some state-feedback gain K if there

exist Q > 0, Y (= KQ), and a diagonal R > 0 with unit trace (see LMI (6.17)), such

that

_

QA

T

i

+A

i

Q+B

u,i

Y +Y

T

B

T

u,i

B

w,i

B

T

w,i

−R

_

< 0, i = 1, . . . , L.

• Norm-bound LDIs: For NLDIs, E ⊇ R

uce

for some state-feedback gain K if there

exist Q > 0, µ > 0, Y and a diagonal R > 0 with unit trace (see LMI (6.19)), such

that

_

¸

_

QA

T

+Y

T

B

T

u

+AQ+B

u

Y +µB

p

B

T

p

QC

T

q

+Y

T

D

T

qu

B

w

C

q

Q+D

qu

Y −µI 0

B

T

w

0 −R

_

¸

_ ≤ 0.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

7.4 State-to-Output Properties 107

Remark: For LTI systems and NLDIs, it is possible to eliminate the variable Y .

7.3.3 Reachable sets with unit-peak inputs

For a ﬁxed state-feedback gain K, the set of states reachable with inputs with unit-

peak is deﬁned as

R

up

∆

=

_

x(T)

¸

¸

¸

¸

¸

x, w, u satisfy (7.17), u = Kx, x(0) = 0,

w(t)

T

w(t) ≤ 1, T ≥ 0

_

.

• LTI systems: From condition (6.25), E contains R

uce

for the LTI system (7.4) if

there exists α > 0 such that

AQ+QA

T

+B

u

Y +Y

T

B

T

u

+B

w

B

T

w

/α +αQ ≤ 0,

where Y = KQ.

• Polytopic LDIs: For PLDIs, E ⊇ R

uce

if there exist Q > 0 and Y (see condi-

tion (6.28)) such that

A

i

Q+QA

T

i

+B

u,i

Y +Y

T

B

T

u,i

+αQ+B

w,i

B

T

w,i

/α ≤ 0

for i = 1, . . . , L.

• Norm-bound LDIs: From Chapter 6, E contains R

uce

if there exist Q > 0, Y ,

α > 0 and µ > 0 such that

_

¸

_

_

AQ+QA

T

+αQ

+B

u

Y +Y

T

B

u

+µB

p

B

T

p

+B

w

B

T

w

/α

_

(C

q

Q+D

qu

Y )

T

C

q

Q+D

qu

Y −µI

_

¸

_ ≤ 0.

This condition is an LMI for ﬁxed α.

Remark: Again, for LTI systems and NLDIs, it is possible to eliminate the

variable Y .

7.4 State-to-Output Properties

We consider state-feedback design to achieve certain desirable state-to-output prop-

erties for the LDI (4.5). We set the exogenous input w to zero in (4.5), and consider

˙ x = A(t)x +B

u

(t)u, z = C

z

(t)x.

• LTI systems: For LTI systems, the state equations are

˙ x = Ax +B

u

u, z = C

z

x +D

zu

u. (7.21)

• Polytopic LDIs: For PLDIs we have

_

A(t) B

u

(t)

C

z

(t) D

zu

(t)

_

∈ Co

__

A

1

B

u,1

C

z,1

D

zu,1

_

, . . . ,

_

A

L

B

u,L

C

z,L

D

zu,L

__

.

This electronic version is for personal use and may not be duplicated or distributed.

108 Chapter 7 State-Feedback Synthesis for LDIs

• Norm-bound LDIs: NLDIs are given by

˙ x = Ax +B

u

u +B

p

p, z = C

z

x +D

zu

u

q = C

q

x +D

qu

u p = ∆(t)q, ∆(t) ≤ 1,

which can be rewritten as

˙ x = Ax +B

u

u +B

p

p, z = C

z

x +D

zu

u

q = C

q

x +D

qu

u, p

T

p ≤ q

T

q.

• Diagonal Norm-bound LDIs: Finally, DNLDIs are given by

˙ x = Ax +B

u

u +B

p

p, z = C

z

x +D

zu

u, q = C

q

x +D

qu

u,

p

i

= δ

i

(t)q

i

, |δ

i

(t)| ≤ 1, i = 1, . . . , L.

which can be rewritten as

˙ x = Ax +B

u

u +B

p

p, z = C

z

x +D

zu

u, q = C

q

x +D

qu

u,

|p

i

| ≤ |q

i

|, i = 1, . . . , n

q

.

7.4.1 Bounds on output energy

We ﬁrst show how to ﬁnd a state-feedback gain K such that the output energy (as

deﬁned in §6.2.1) of the closed-loop system is less than some speciﬁed value. We

assume ﬁrst that the initial condition x

0

is given.

• LTI systems: We conclude from LMI (6.39) that for a given state-feedback gain

K, the output energy of system (7.21) does not exceed x

T

0

Q

−1

x

0

, where Q > 0 and

Y = KQ satisfy

_

AQ+QA

T

+B

u

Y +Y

T

B

T

u

(C

z

Q+D

zu

Y )

T

C

z

Q+D

zu

Y −I

_

≤ 0, (7.22)

Regarding Y as a variable, we can then ﬁnd a state-feedback gain that guarantees an

output energy less than γ by solving the LMIP x

T

0

Q

−1

x

0

≤ γ and (7.22).

Of course, inequality (7.22) is closely related to the classical Linear-Quadratic

Regulator (LQR) problem; see the Notes and References.

• Polytopic LDIs: In this case, the output energy is bounded above by x

T

0

Q

−1

x

0

,

where Q satisﬁes the LMI

_

¸

¸

_

_

A

i

Q+QA

T

i

+B

u,i

Y +Y

T

B

T

u,i

_

(C

z,i

Q+D

zu,i

Y )

T

C

z,i

Q+D

zu,i

Y −I

_

¸

¸

_

≤ 0, i = 1, . . . , L

for some Y .

• Norm-bound LDIs: In the case of NLDIs, the output energy is bounded above

by x

T

0

Q

−1

x

0

, for any Q > 0, Y and µ ≥ 0 such that

_

¸

¸

¸

¸

_

_

AQ+QA

T

+B

u

Y

+Y

T

B

T

u

+µB

p

B

T

p

_

(C

z

Q+D

zu

Y )

T

(C

q

Q+D

qu

Y )

T

C

z

Q+D

zu

Y −I 0

C

q

Q+D

qu

Y 0 −µI

_

¸

¸

¸

¸

_

≤ 0. (7.23)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

7.5 Input-to-Output Properties 109

Remark: As before, we can eliminate the variable Y from this LMI.

• Diagonal Norm-bound LDIs: Finally, in the case of DNLDIs, the output energy

is bounded above by x

T

0

Q

−1

x

0

, for any Q > 0, Y and M ≥ 0 and diagonal such that

_

¸

¸

¸

¸

_

_

AQ+QA

T

+B

u

Y

+Y

T

B

T

u

+B

p

MB

T

p

_

(C

z

Q+D

zu

Y )

T

(C

q

Q+D

qu

Y )

T

C

z

Q+D

zu

Y −I 0

C

q

Q+D

qu

Y 0 −M

_

¸

¸

¸

¸

_

≤ 0.

Given an initial condition, ﬁnding a state-feedback gain K so as to minimize the

upper bound on the extractable energy for the various LDIs is therefore an EVP. We

can extend these results to the case when x

0

is speciﬁed to lie in a polytope or an

ellipsoid (see §6.2.1). If x

0

is a random variable with Ex

0

x

T

0

= X

0

, EVPs that yield

state-feedback gains that minimize the expected value of the output energy can be

derived.

7.5 Input-to-Output Properties

We ﬁnally consider the problem of ﬁnding a state-feedback gain K so as to achieve de-

sired properties between the exogenous input w and the output z for the system (7.1).

As mentioned in Chapter 6, the list of problems considered here is far from exhaustive.

7.5.1 L

2

and RMS gains

We seek a state-feedback gain K such that the L

2

gain

sup

w2=1

z

2

= sup

w2=0

z

2

w

2

of the closed-loop system is less than a speciﬁed number γ.

• LTI systems: As seen in 6.3.2, the L

2

gain for LTI systems is equal to the H

∞

norm of the corresponding transfer matrix. From §6.3.2, there exists a state-feedback

gain K such that the L

2

gain of an LTI system is less than γ, if there exist K and

Q > 0 such that,

_

¸

_

_

(A+B

u

K)Q+Q(A+B

u

K)

T

+B

w

B

T

w

_

Q(C

z

+D

zu

K)

T

(C

z

+D

zu

K)Q −γ

2

I

_

¸

_ ≤ 0. (7.24)

Introducing Y = KQ, this can be rewritten as

_

AQ+QA

T

+B

u

Y +Y

T

B

T

u

+B

w

B

T

w

(C

z

Q+D

zu

Y )

T

C

z

Q+D

zu

Y −γ

2

I

_

≤ 0. (7.25)

In the case of LTI systems, this condition is necessary and suﬃcient. Assuming

C

T

z

D

zu

= 0 and D

T

zu

D

zu

invertible, it is possible to simplify the inequality (7.25)

and get the equivalent Riccati inequality

AQ+QA

T

−B

u

(D

T

zu

D

zu

)

−1

B

T

u

+B

w

B

T

w

+QC

T

z

C

z

Q/γ

2

≤ 0. (7.26)

This electronic version is for personal use and may not be duplicated or distributed.

110 Chapter 7 State-Feedback Synthesis for LDIs

The corresponding Riccati equation is readily solved via Hamiltonian matrices. The

resulting controller, with state-feedback gain K = −(D

T

zu

D

zu

)

−1

B

T

u

, yields a closed-

loop system with H

∞

-norm less than γ.

• Polytopic LDIs: From §6.3.2, there exists a state-feedback gain such that the

L

2

gain of a PLDI is less than γ if there exist Y and Q > 0 such that

_

¸

_

_

A

i

Q+QA

T

i

+B

u,i

Y

+Y

T

B

T

u,i

+B

w,i

B

T

w,i

_

(C

z,i

Q+D

zu,i

Y )

T

C

z,i

Q+D

zu,i

Y −γ

2

I

_

¸

_ ≤ 0. (7.27)

• Norm-bound LDIs: For NLDIs, the LMI that guarantees that the L

2

gain is less

than γ for some state-feedback gain K is

_

¸

¸

¸

¸

¸

¸

¸

¸

_

_

_

_

AQ+QA

T

+B

u

Y +Y

T

B

T

u

+B

w

B

T

w

+µB

p

B

T

p

_

_

_ (C

z

Q+D

zu

Y )

T

(C

q

Q+D

qu

Y )

T

C

z

Q+D

zu

Y −γ

2

I 0

C

q

Q+D

qu

Y 0 −µI

_

¸

¸

¸

¸

¸

¸

¸

¸

_

≤ 0. (7.28)

We can therefore ﬁnd state-feedback gains that minimize the upper bound on the

L

2

gain, provable with quadratic Lyapunov functions, for the various LDIs by solving

EVPs.

7.5.2 Dissipativity

We next seek a state-feedback gain K such that the closed-loop system (7.2) is passive;

more generally, assuming D

zw

(t) is nonzero and square, we wish to maximize the

dissipativity, i.e., η satisfying

_

T

0

_

w

T

z −ηw

T

w

_

dt ≥ 0

for all T ≥ 0.

• LTI systems: Substituting Y = KQ, and using LMI (6.60), we conclude that the

dissipation of the LTI system is at least η if the LMI in η, Q > 0 and Y holds:

_

AQ+QA

T

+B

u

Y +Y

T

B

T

u

B

w

−QC

T

z

−Y

T

D

T

zu

B

T

w

−C

z

Q−D

zu

Y 2ηI −(D

zw

+D

T

zw

)

_

≤ 0.

• Polytopic LDIs: From (6.62), there exists a state-feedback gain such that the

dissipation exceeds η if there exist Q > 0 and Y such that

_

¸

¸

_

_

A

i

Q+QA

T

i

+B

u,i

Y +Y

T

B

T

u,i

_

B

w,i

−QC

T

z,i

−Y

T

D

zu,i

B

T

w,i

−C

z,i

Q−D

zu,i

Y 2ηI −(D

zw,i

+D

T

zw,i

)

_

¸

¸

_

≤ 0.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

7.6 Observer-Based Controllers for Nonlinear Systems 111

• Norm-bound LDIs: From (6.64), there exists a state-feedback gain such that the

dissipation exceeds η if there exist Q > 0 and Y such that

_

¸

¸

¸

¸

¸

¸

¸

_

_

AQ+QA

T

+B

u

Y

+Y

T

B

T

u

+µB

q

B

T

q

_ _

B

w

−QC

T

z

−Y

T

D

T

zu

_

(C

q

Q+D

qu

Y )

T

_

B

T

w

−C

z

Q

−D

zu

Y

_

−(D

zw

+D

T

zw

−2ηI) 0

C

q

Q+D

qu

Y 0 −µI

_

¸

¸

¸

¸

¸

¸

¸

_

≤ 0

We can therefore ﬁnd state-feedback gains that maximize the lower bound on the

dissipativity, provable with quadratic Lyapunov functions, for the various LDIs by

solving EVPs.

Remark: As with ¸6.3.4, we can incorporate scaling techniques into many of the

results above to derive componentwise results. Since the new LMIs thus obtained

are straightforward to derive, we will omit them here.

7.5.3 Dynamic versus static state-feedback controllers

A dynamic state-feedback controller has the form

˙

¯ x =

¯

A¯ x +

¯

B

y

x, u =

¯

C

u

¯ x +

¯

D

uy

x, (7.29)

where

¯

A ∈ R

r×r

,

¯

B

y

∈ R

r×n

,

¯

C

u

∈ R

nu×r

,

¯

D

uy

∈ R

nu×n

. The number r is called

the order of the controller. Note that by taking r = 0, this dynamic state-feedback

controller reduces to the state-feedback we have considered so far (which is called

static state-feedback in this context).

It might seem that the dynamic state-feedback controller (7.29) allows us to meet

more speciﬁcations than can be met using a static state-feedback. For speciﬁcations

based on quadratic Lyapunov functions, however, this is not the case. For exam-

ple, suppose there does not exist a static state-feedback gain that yields closed-loop

quadratic stability. In this case we might turn to the more general dynamic state-

feedback. We could work out the more complicated LMIs that characterize quadratic

stabilizability with dynamic state-feedback. We would ﬁnd, however, that these more

general LMIs are also infeasible. See the Notes and References.

7.6 Observer-Based Controllers for Nonlinear Systems

We consider the nonlinear system

˙ x = f(x) +B

u

u, y = C

y

x, (7.30)

where x : R

+

→ R

n

is the state variable, u : R

+

→ R

p

is the control variable,

and y : R

+

→ R

q

is the measured or sensed variable. We assume the function

f : R

n

→R

n

satisﬁes f(0) = 0 and

∂f

∂x

∈ Co{A

1

, . . . , A

M

}

where A

1

, . . . , A

L

are given, which is the same as (4.15).

We look for a stabilizing observer-based controller of the form

˙

¯ x = f(¯ x) +B

u

u +L(C

y

¯ x −y), u = K¯ x, (7.31)

This electronic version is for personal use and may not be duplicated or distributed.

112 Chapter 7 State-Feedback Synthesis for LDIs

i.e., we search for the matrices K (the estimated-state feedback gain) and L (the

observer gain) such that the closed-loop system

_

˙ x

˙

¯ x

_

=

_

f(x) +B

u

K¯ x

−LC

y

x +f(¯ x) + (B

u

K +LC

y

)¯ x

_

(7.32)

is stable. The closed-loop system (7.32) is stable if it is quadratically stable, which is

true if there exists a positive-deﬁnite matrix

˜

P ∈ R

2n×2n

such that for any nonzero

trajectory x, ¯ x, we have:

d

dt

_

x

¯ x

_

T

˜

P

_

x

¯ x

_

< 0.

In the Notes and References, we prove that this is true if there exist P, Q, Y , and W

such that the LMIs

Q > 0, A

i

Q+QA

T

i

+B

u

Y +Y

T

B

T

u

< 0, i = 1, . . . , M (7.33)

and

P > 0, A

T

i

P +PA

i

+WC

y

+C

T

y

W

T

< 0, i = 1, . . . , M (7.34)

hold. To every P, Q, Y , and W satisfying these LMIs, there corresponds a stabilizing

observer-based controller of the form (7.31), obtained by setting K = Y Q

−1

and

L = P

−1

W.

Using the elimination procedure of §2.6.2, we can obtain equivalent conditions in

which the variables Y and W do not appear. These conditions are that some P > 0

and Q > 0 satisfy

A

i

Q+QA

T

i

< σB

u

B

T

u

, i = 1, . . . , M (7.35)

and

A

T

i

P +PA

i

< µC

T

y

C

y

, i = 1, . . . , M (7.36)

for some σ and µ. By homogeneity we can freely set σ = µ = 1. For any P > 0, Q > 0

satisfying these LMIs with σ = µ = 1, we can obtain a stabilizing observer-based

controller of the form (7.31), by setting K = −(1/2)B

T

u

Q

−1

and L = −(1/2)P

−1

C

y

.

Another equivalent condition is that

˜

B

T

u

_

A

i

Q+QA

T

i

_

˜

B

u

< 0, i = 1, . . . , M

and

˜

C

y

_

A

T

i

P +PA

i

_

˜

C

T

y

< 0, i = 1, . . . , M

hold for some P > 0, Q > 0, where

˜

B

u

and

˜

C

T

y

are orthogonal complements of B

u

and

C

T

y

, respectively. If P and Q satisfy these LMIs, then they satisfy the LMIs (7.36)

and (7.35) for some σ and µ. The observer-based controller with K = −(σ/2)B

T

u

Q

−1

and L = −(µ/2)P

−1

C

y

stabilizes the nonlinear system (7.30).

Notes and References

Lyapunov functions and state-feedback

The extension of Lyapunov’s methods to the state-feedback synthesis problem has a long

history; see e.g., [Bar70a, Lef65]. It is clearly related to the theory of optimal control, in

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 113

which the natural Lyapunov function is the Bellman–Pontryagin “min-cost-to-go” or value

function (see [Pon61, Pon62]). We suspect that the methods described in this chapter can

be interpreted as a search for suboptimal controls in which the candidate value functions

are restricted to a speciﬁc class, i.e., quadratic. In the general optimal control problem,

the (exact) value function satisﬁes a partial diﬀerential equation and is hard to compute;

by restricting our search to quadratic approximate value functions we end up with a low

complexity task, i.e., a convex problem involving LMIs.

Several approaches use Lyapunov-like functions to synthesize nonlinear state-feedback control

laws. One example is the “Lyapunov Min-Max Controller” described in [Gut79, GP82,

Cor85, CL90, Wei94]. See also [Zin90, SC91, Rya88, GR88].

Quadratic stabilizability

The term “quadratic stabilizability” seems to have been coined by Hollot and Barmish

in [HB80], where the authors give necessary and suﬃcient conditions for it; see also [Lei79,

Gut79, BCL83, PB84, Cor85]. Petersen [Pet85], shows that there exist LDIs that are

quadratically stabilizable, though not via linear state-feedback; see also [SP94e]. Hol-

lot [Hol87] describes conditions under which a quadratically stabilizable LDI has inﬁnite

stabilizability margin; see also [SRC93]. Wei [Wei90] derives necessary and suﬃcient “sign-

pattern” conditions on the perturbations for the existence of a linear feedback for quadratic

stability for a class of single-input uncertain linear dynamical systems. Petersen [Pet87b]

derives a Riccati-based approach for stabilizing NLDIs, which is in fact the state-feedback

H∞ equation (7.28) (this is shown in [Pet87a]).

The change of variables Q = P

−1

and Y = KP

−1

, which enables the extension of many

of the results of Chapters 5 and 6 to state-feedback synthesis, is due to Bernussou, Peres

and Geromel [BPG89a]. See also [GPB91, BPG89b, BPG89a, PBG89, PGB93]. This

change of variables, though not explicitly described, is used in the paper by Thorp and

Barmish [TB81]; see also [HB80, Bar83, Bar85]. For a recent and broad review about

quadratic stabilization of uncertain systems, see [Cor94].

In [Son83], Sontag formulates a general theorem giving necessary and suﬃcient Lyapunov

type conditions for stabilizability. For LTI systems his criteria reduce to the LMIs in this

chapter.

Quadratic Lyapunov functions for proving performance

The addition of performance considerations in the analysis of LDIs can be traced back to

1955 and even earlier, when Letov [Let61] studied the performance of unknown, nonlinear

and possibly time-varying control systems and calls it “the problem of control quality”. At

that time, the performance criteria were decay rate and output peak deviations for systems

subject to bounded peak inputs (see [Let61, p.234] for details).

In the case of LTI systems, adding performance indices allows one to recover many classical

results of automatic control. The problem of maximizing decay rate is discussed in great

detail in [Yan92]; other references on decay rates and quadratic stability margins are [Gu92b,

Gu92a, GPB91, Gu92b, Sko91b, PZP

+

92, EBFB92, Cor90]. Arzelier et al. [ABG93]

consider the problem of robust stabilization with pole assignment for PLDIs via cutting-plane

techniques.

The problem of approximating reachable sets for LTI systems under various assumptions on

the energy of the exogenous input has been considered for example by Skelton [SZ92] and

references therein; see also the Notes and References of the previous chapter. See also [OC87].

The idea that the output variance for white noise inputs (and related LQR-like performance

measures) can be minimized using LMIs can be found in [BH88a, BH89, RK91, KR91,

KKR93, BB91, FBBE92, RK93, SI93]. Peres, Souza and Geromel [PSG92, PG93] con-

sider the extension to PLDIs. The counterpart for NLDIs is in Petersen, McFarlane and

Rotea [PM92, PMR93], as well as in Stoorvogel [Sto91].

The problem of ﬁnding a state-feedback gain to minimize the scaled L2 gain for LTI sys-

tems is discussed in [EBFB92]. The L2 gain of PLDIs has been studied by Obradovic and

This electronic version is for personal use and may not be duplicated or distributed.

114 Chapter 7 State-Feedback Synthesis for LDIs

Valavani [OV92], by Peres, Geromel and Souza [PGS91] and also by Ohara, Masubuchi and

Suda [OMS93]. NLDIs have been studied in this context by Petersen [Pet89], and DeSouza

et al. [XFdS92, XdS92, WXdS92, dSFX93], who point out explicitly that quadratic sta-

bility of NLDIs with an L2 gain bound is equivalent to the scaled H∞ condition (6.54); see

also [ZKSN92, Gu93]. For LTI systems, it is interesting to compare the inequality (7.25)

which provides a necessary and suﬃcient condition for a system to have L2 gain less than

γ with the corresponding matrix inequality found in the article by Stoorvogel and Trentel-

man [ST90]. In particular, the quadratic matrix inequality found there (expressed in our

notation) is

_

A

T

P +PA+γ

2

PBwB

T

w

P +C

T

z

Cz PBu +C

T

z

Dzu

B

T

u

P +D

T

zu

Cz D

T

zu

Dzu

_

≥ 0,

which does not possess any obvious convexity property. We note that the Riccati inequal-

ity (7.26) is also encountered in full-state feedback H∞ control [Pet87a]. Its occurrence in

control theory is much older; it appears verbatim in some articles on the theory of nonzero-

sum diﬀerential games; see for example Starr and Ho [SH68], Yakubovich [Yak70, Yak71]

and Mageirou [Mag76].

Kokotovic [Kok92] considers stabilization of nonlinear systems, and provides a motivation for

rendering an LTI system passive via state-feedback (see ¸6.3.3). In [PP92], the authors give

analytic conditions for making an LTI system passive via state-feedback; see also [SKS93].

Finally, we mention an instance of a Lyapunov function with a built-in performance criterion.

In the book by Aubin and Cellina [AC84, ¸6], we ﬁnd:

We shall investigate whether diﬀerential inclusions ˙ x ∈ F(x(t)), x(0) = x0

have trajectories satisfying the property

∀t > s, V (x(t)) −V (x(s)) +

_

t

s

W(x(τ), ˙ x(τ)) dτ ≤ 0. (7.37)

We shall then say that a function V [ ] satisfying this condition is a Lyapunov

function for F with respect to W.

Similar ideas can be found in [CP72]. We shall encounter such Lyapunov functions in ¸8.2.

Inequality (7.37) is often called a “dissipation inequality” [Wil71b].

The problem of minimizing the control eﬀort given a performance bound on the closed-loop

system can be found in the articles of Sch¨ omig, Sznaier and Ly [SSL93], and Grigoriadis,

Carpenter, Zhu and Skelton [GCZS93].

LMI formulation of LQR problem

For the LTI system ˙ x = Ax +Buu, z = Czx +Dzuu, the Linear-Quadratic Regulator (LQR)

problem is: Given an initial condition x(0), ﬁnd the control input u that minimizes the output

energy

_

∞

0

z

T

z dt. We assume for simplicity that (A, B, C) is minimal, D

T

zu

Dzu is invertible

and D

T

zu

Cz = 0.

It turns out that the optimal input u can be expressed as a constant state-feedback u = Kx,

where K = −(D

T

zu

Dzu)

−1

B

T

u

Pare and Pare is the unique positive-deﬁnite matrix that satisﬁes

the algebraic Riccati equation

A

T

Pare +PareA−PareBu(D

T

zu

Dzu)

−1

B

T

u

Pare +C

T

z

Cz = 0. (7.38)

The optimal output energy for an initial condition x(0) is then given by x(0)

T

Parex(0). With

Qare = P

−1

are

, we can rewrite (7.38) as

AQare +QareA

T

−Bu(D

T

zu

Dzu)

−1

B

T

u

+QareC

T

z

CzQare = 0, (7.39)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 115

and the optimal output energy as x(0)

T

Q

−1

are

x(0).

In section ¸7.4.1, we considered the closely related problem of ﬁnding a state-feedback gain

K that minimizes an upper bound on the energy of the output, given an initial condition.

For LTI systems, the upper bound equals the output energy, and we showed that in this case,

the minimum output energy was given by minimizing x(0)

T

Q

−1

x(0) subject to Q > 0 and

_

AQ+QA

T

+BuY +Y

T

B

T

u

(CzQ+DzuY )

T

CzQ+DzuY −I

_

≤ 0. (7.40)

Of course, the optimal value of this EVP must equal the optimal output energy given via

the solution to the Riccati solution (7.39); the Riccati equation can be thus be interpreted as

yielding an analytic solution to the EVP. We can derive this analytic solution for the EVP via

the following steps. First, it can be shown, using a simple completion-of-squares argument

that LMI (7.40) holds for some Q > 0 and Y if and only if it holds for Y = −(D

T

zu

Dzu)

−1

B

T

u

;

in this case, Q > 0 must satisfy

QA

T

+AQ+QC

T

z

CzQ−Bu

_

D

T

zu

Dzu

_

−1

B

T

u

≤ 0. (7.41)

Next, it can be shown by standard manipulations that if Q > 0 satisﬁes (7.41), then Q ≤ Qare.

Therefore for every initial condition x(0), we have x(0)

T

Q

−1

are

x(0) ≤ x(0)

T

Q

−1

x(0), and

therefore the optimal value of the EVP is just x(0)

T

Q

−1

are

x(0), and the optimal state-feedback

gain is Kopt = YoptQ

−1

are

= −(D

T

zu

Dzu)

−1

B

T

u

Pare.

It is interesting to note that with P = Q

−1

, the EVP is equivalent to minimizing x(0)

T

Px(0)

subject to

A

T

P +PA+C

T

z

Cz −PBu

_

D

T

zu

Dzu

_

−1

B

T

u

P ≤ 0, (7.42)

which is not a convex constraint in P. However, the problem of maximizing x(0)

T

Px(0)

subject P > 0 and the constraint

A

T

P +PA−PBu(D

T

zu

Dzu)

−1

B

T

u

P +CzC

T

z

≥ 0,

which is nothing other than (7.42), but with the inequality reversed, is an EVP; it is well-

known (see for example [Wil71b]) that this EVP is another formulation of the LQR problem.

Static versus dynamic state-feedback

An LTI system can be stabilized using dynamic state-feedback if and only if it can be stabi-

lized using static state-feedback (see for example, [Kai80]). In fact, similar statements can

be made for all the properties and for all the LDIs considered in this chapter, provided the

properties are speciﬁed using quadratic Lyapunov functions. We will demonstrate this fact

on one simple problem considered in this chapter.

Consider the LTI system

˙ x = Ax +Buu +Bww, x(0) = 0, z = Czx. (7.43)

From ¸7.5.1, the exists a static state-feedback u = Kx such that the L2 gain from w to z

does not exceed γ if the LMI in variables Q > 0 and Y is feasible:

_

AQ+QA

T

+BuY +Y

T

B

T

u

+BwB

T

w

QC

T

z

CzQ −γ

2

I

_

≤ 0.

This electronic version is for personal use and may not be duplicated or distributed.

116 Chapter 7 State-Feedback Synthesis for LDIs

(In fact this condition is also necessary, but this is irrelevant to the current discussion.)

Eliminating Y , we get the equivalent LMI in Q > 0 and scalar σ:

AQ+QA

T

+BwB

T

w

+

QC

T

z

CzQ

γ

2

≤ σBuB

T

u

. (7.44)

We next consider a dynamic state-feedback for the system:

˙ xc = Acxc +Bcx, xc(0) = 0, u = Ccxc +Dcx. (7.45)

We will show that there exist matrices Ac, Bc, Cc and Dc such that the L2 gain of the

system (7.43,7.45) does not exceed γ if and only if the LMI (7.44) holds. This will establish

that the smallest upper bound on the L2 gain, provable via quadratic Lyapunov functions,

is the same irrespective of whether a static state-feedback or a dynamic state-feedback is

employed.

We will show only the part that is not obvious. With state vector x

cl

= [x

T

x

T

c

]

T

, the

closed-loop system is

˙ x

cl

= A

cl

x

cl

+B

cl

u, x

cl

(0) = 0, z = C

cl

x

cl

,

where A

cl

= A

big

+B

big

K

big

with

A

big

=

_

A 0

0 0

_

, B

big

=

_

0 Bu

I 0

_

, K

big

=

_

Bc Ac

Dc Cc

_

,

B

cl

=

_

Bw

0

_

, C

cl

= [Cz 0] .

From ¸6.3.2, the L2 gain of the closed-loop system does not exceed γ if the LMI in P

big

> 0

is feasible:

_

A

T

big

P +P

big

A+PB

big

K

big

+K

T

big

B

T

big

P +C

T

cl

C

cl

PB

cl

B

T

cl

P −γ

2

I

_

≤ 0.

Eliminating K

big

from this LMI yields two equivalent LMIs in P

big

> 0 and a scalar τ.

_

A

T

big

P +P

big

A−τI +C

T

cl

C

cl

PB

cl

B

T

cl

P −γ

2

I

_

≤ 0,

_

A

T

big

P +P

big

A−τPB

big

B

T

big

P +C

T

cl

C

cl

PB

cl

B

T

cl

P −γ

2

I

_

≤ 0.

It is easy to verify that the ﬁrst LMI holds for large enough τ, while the second reduces to

τ > 0 and

AQ11 +Q11A

T

+

BwB

T

w

γ

2

+Q11CzC

T

z

Q11 ≤ τBuB

T

u

,

where Q11 is the leading nn block of P

−1

big

. The last LMI is precisely (7.44) with the change

of variables Q = γ

2

Q11 and σ = γ

2

τ.

A similar result is established [PZP

+

92]; other references that discuss the question of when

dynamic or nonlinear feedback does better than static feedback include Petersen, Khar-

gonekar and Rotea [Pet85, KPR88, RK88, KR88, Pet88, RK89].

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 117

Dynamic output-feedback for LDIs

A natural extension to the results of this chapter is the synthesis of output-feedback to meet

various performance speciﬁcations. We suspect, however, that output-feedback synthesis

problems have high complexity and are therefore unlikely to be recast as LMI problems.

Several researchers have considered such problems: Steinberg and Corless [SC85], Galimidi

and Barmish [GB86], and Geromel, Peres, de Souza and Skelton [PGS93]; see also [HCM93].

One variation of this problem is the “reduced-order controller problem”, that is, ﬁnding dy-

namic output-feedback controllers with the smallest possible number of states. It is shown

in [PZPB91, EG93, Pac94] that this problem can be reduced to the problem of minimizing

the rank of a matrix P ≥ 0 subject to certain LMIs. The general problem of minimizing the

rank of a matrix subject to an LMI constraint has high complexity; in fact, a special rank

minimization problem can be shown to be equivalent to the NP-hard zero-one linear pro-

gramming problem (see e.g. [Dav94]). Nevertheless several researchers have tried heuristic,

local algorithms for such problems and report practical success. Others studying this topic

include Iwasaki, Skelton, Geromel and de Souza [IS93b, ISG93, GdSS93, SIG93].

Observer-based controllers for nonlinear systems

We prove that the LMIs (7.33) and (7.34) (in the variables P > 0, Q > 0, W and Y ) ensure

the existence of an observer-based controller of the form (7.31) which makes the system (7.30)

quadratically stable.

We start by rewriting the system (7.32) in the coordinates ¯ x, ¯ x −x:

_

˙

¯ x

˙

¯ x − ˙ x

_

=

_

LCy(¯ x −x) +f(¯ x) +BuK¯ x

f(¯ x) −f(x) +LCy(¯ x −x)

_

. (7.46)

Using the results of ¸4.3, we can write

_

˙

¯ x

˙

¯ x − ˙ x

_

∈ Co

_

˜

Aij

¸

¸

1 ≤ i ≤ M, 1 ≤ j ≤ M

_

_

¯ x

¯ x −x

_

with

˜

Aij =

_

Ai +BuK LCy

0 Aj +LCy

_

, i, j = 1, . . . , M.

Simple state-space manipulations show the system (7.32) is quadratically stable if and only if

the system (7.46) is quadratically stable. Therefore, the system (7.32) is quadratically stable

if there exists

˜

P > 0 such that

˜

A

T

ij

˜

P +

˜

P

˜

Aij < 0, i, j = 1, . . . , M. (7.47)

We will now show that the LMI (7.47) has a solution if and only if the LMIs (7.33) and (7.34)

do.

First suppose that (7.47) has a solution

˜

P. With

˜

P partitioned as n n blocks,

˜

P =

_

P11 P12

P

T

12

P22

_

,

we easily check that the inequality (7.47) implies

(Ai +BuK)

T

P11 +P11(Ai +BuK) < 0, i = 1, . . . , M. (7.48)

This electronic version is for personal use and may not be duplicated or distributed.

118 Chapter 7 State-Feedback Synthesis for LDIs

With the change of variables Q = P

−1

11

and Y = BuQ, the inequality (7.48) is equivalent

to (7.33).

Deﬁne

˜

Q =

˜

P

−1

. The inequality (7.47) becomes

˜

Aij

˜

Q+

˜

Q

˜

A

T

ij

< 0, i, j = 1, . . . , M.

Denoting by Q22 the lower-right n n block of

˜

Q, we see that this inequality implies

(Ai +LCy)Q22 +Q22(Ai +LCy)

T

< 0, i = 1, . . . , M. (7.49)

With P = Q

−1

22

and W = PL, the inequality (7.49) is equivalent to (7.34).

Conversely, suppose that P, Q, Y and W satisfy the LMIs (7.34) and (7.33) and deﬁne

L = P

−1

W, K = Y Q

−1

. We now prove there exists a positive λ such that

˜

P =

_

λQ

−1

0

0 P

_

satisﬁes (7.47) and therefore proves the closed-loop system (7.32) is quadratically stable. We

compute

˜

A

T

ij

˜

P +

˜

P

˜

Aij =

_

¸

¸

¸

_

λ

_

(Ai +BuK)

T

Q

−1

+Q

−1

(Ai +BuK)

_

λQ

−1

LCy

λ(LCy)

T

Q

−1

(Aj +LCy)

T

P

+P(Aj +LCy)

_

¸

¸

¸

_

,

for i, j = 1, . . . , M. Therefore, using Schur complements, the closed-loop system (7.32) is

quadratically stable if λ > 0 satisﬁes

λ(Q

−1

LCy((Aj +LCy)

T

P +P(Aj +LCy))

−1

(LCy)

T

Q

−1

)

−(A+BuK)

T

Q

−1

−Q

−1

(A+BuK) > 0

for i, j = 1, . . . , M. This condition is satisﬁed for any λ > 0 such that

λ min

1≤j≤M

µj > max

1≤i≤M

νi,

where

µj = λmin(Q

−1

LCy((Aj +LCy)

T

P +P(Aj +LCy))

−1

(LCy)

T

Q

−1

), j = 1, . . . , M

and

νi = λmax((Ai +BuK)

T

Q

−1

+Q

−1

(Ai +BuK)), i = 1, . . . , M.

Since (7.34) and (7.33) are satisﬁed, such a λ always exists.

Problems of observer-based controller design along with quadratic Lyapunov functions that

prove stability have been considered by Bernstein and Haddad [BH93] and Yaz [Yaz93].

Gain-scheduled or parameter-dependent controllers

Several researchers have extended the results on state-feedback synthesis to handle the case

in which the controller parameters can depend on system parameters. Such controllers are

called gain-scheduled or parameter-dependent. We refer the reader to the articles cited for

precise explanations of what these control laws are, and which design problems can be recast

as LMI problems. Lu, Zhou and Doyle in [LZD91] and Becker and Packard [Bec93, BP91]

consider the problem in the context of NLDIs. Other relevant references include [PB92,

Pac94, PBPB93, BPPB93, GA94, IS93a, IS93b, AGB94].

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Chapter 8

Lur’e and Multiplier Methods

8.1 Analysis of Lur’e Systems

We consider the Lur’e system

˙ x = Ax +B

p

p +B

w

w, q = C

q

x,

z = C

z

x, p

i

(t) = φ

i

(q

i

(t)), i = 1, . . . , n

p

,

(8.1)

where p(t) ∈ R

np

, and the functions φ

i

satisfy the [0, 1] sector conditions

0 ≤ σφ

i

(σ) ≤ σ

2

for all σ ∈ R, (8.2)

or, equivalently,

φ

i

(σ)(φ

i

(σ) −σ) ≤ 0 for all σ ∈ R.

The data in this problem are the matrices A, B

p

, B

w

, C

q

and C

z

. The results that

follow will hold for any nonlinearities φ

i

satisfying the sector conditions. In cases

where the φ

i

are known, however, the results can be sharpened.

It is possible to handle the more general sector conditions

α

i

σ

2

≤ σφ

i

(σ) ≤ β

i

σ

2

for all σ ∈ R,

where α

i

and β

i

are given. Such systems are readily transformed to the form given

in (8.1) and (8.2) by a loop transformation (see the Notes and References).

An important special case of (8.1), (8.2) occurs when the functions φ

i

are linear,

i.e., φ

i

(σ) = δ

i

σ, where δ

i

∈ [0, 1]. In the terminology of control theory, this is referred

to as a system with unknown-but-constant parameters. It is important to distinguish

this case from the PLDI obtained with (8.1), φ

i

(σ) = δ

i

(t)σ and δ

i

(t) ∈ [0, 1], which

is referred to as a system with unknown, time-varying parameters.

Our analysis will be based on Lyapunov functions of the form

V (ξ) = ξ

T

Pξ + 2

np

i=1

λ

i

_

Ci,qξ

0

φ

i

(σ) dσ, (8.3)

where C

i,q

denotes the ith row of C

q

. Thus the data describing the Lyapunov function

are the matrix P and the scalars λ

i

, i = 1, . . . , n

p

. We require P > 0 and λ

i

≥ 0,

which implies that V (ξ) ≥ ξ

T

Pξ > 0 for nonzero ξ.

Note that the Lyapunov function (8.3) depends on the (not fully speciﬁed) non-

linearities φ

i

. Thus the data P and λ

i

should be thought of as providing a recipe for

119

120 Chapter 8 Lur’e and Multiplier Methods

constructing a speciﬁc Lyapunov function given the nonlinearities φ

i

. As an exam-

ple, consider the special case of a system with unknown-but-constant parameters, i.e.,

φ

i

(σ) = δ

i

σ. The Lyapunov function will then have the form

V (ξ) = ξ

T

_

P +C

T

q

∆ΛC

q

_

ξ

where ∆ = diag(δ

1

, . . . , δ

np

) and Λ = diag(λ

1

, . . . , λ

np

). In other words, we are really

synthesizing a parameter-dependent quadratic Lyapunov function for our parameter-

dependent system.

8.1.1 Stability

We set the exogenous input w to zero, and seek P and λ

i

such that

dV (x)

dt

< 0 for all nonzero x satisfying (8.1) and (8.2). (8.4)

Since

dV (x)

dt

= 2

_

x

T

P +

np

i=1

λ

i

p

i

C

i,q

_

(Ax +B

p

p) ,

condition (8.4) holds if and only if

_

ξ

T

P +

np

i=1

λ

i

π

i

C

i,q

_

(Aξ +B

p

π) < 0

for all nonzero ξ satisfying

π

i

(π

i

−c

i,q

ξ) ≤ 0, i = 1, . . . , n

p

. (8.5)

It is easily shown that

{(ξ, π) | ξ = 0, (8.5) } = {(ξ, π) | ξ = 0 or π = 0, (8.5) } .

The S-procedure then yields the following suﬃcient condition for (8.4): The LMI in

P > 0, Λ = diag(λ

1

, . . . , λ

np

) ≥ 0 and T = diag(τ

1

, . . . , τ

np

) ≥ 0

_

A

T

P +PA PB

p

+A

T

C

T

q

Λ +C

T

q

T

B

T

p

P + ΛC

q

A+TC

q

ΛC

q

B

p

+B

T

p

C

T

q

Λ −2T

_

< 0 (8.6)

holds.

Remark: When we set Λ = 0 we obtain the LMI

_

A

T

P +PA PBp +C

T

q

T

B

T

p

P +TCq −2T

_

< 0,

which can be interpreted as a condition for the existence of a quadratic Lyapunov

function for the Lur’e system, as follows. Perform a loop transformation on the

Lur’e system to bring the nonlinearities to sector [−1, 1]. Then the LMI above

is the same (to within a scaling of the variable T) as one that yields quadratic

stability of the associated DNLDI; see ¸5.1.

Remark: Condition (8.6) is only a suﬃcient condition for the existence of a Lur’e

Lyapunov function that proves stability of system (8.1). It is also necessary when

there is only one nonlinearity, i.e., when np = 1.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

8.1 Analysis of Lur’e Systems 121

8.1.2 Reachable sets with unit-energy inputs

We consider the set reachable with unit-energy inputs,

R

ue

∆

=

_

¸

¸

_

¸

¸

_

x(T)

¸

¸

¸

¸

¸

¸

¸

¸

x, w satisfy (8.3), x(0) = 0

_

T

0

w

T

w dt ≤ 1, T ≥ 0

_

¸

¸

_

¸

¸

_

.

The set

F =

_

ξ

¸

¸

¸

¸

¸

ξ

T

Pξ + 2

np

i=1

λ

i

_

Ci,qξ

0

φ

i

(σ) dσ ≤ 1

_

contains R

ue

if

d

dt

V (x) ≤ w

T

w for all x and w satisfying (8.1) and (8.2). (8.7)

From the S-procedure, the condition (8.7) holds if there exists T = diag(τ

1

, . . . , τ

np

) ≥

0 such that

_

¸

_

A

T

P +PA PB

p

+A

T

C

T

q

Λ +C

T

q

T PB

w

B

T

p

P + ΛC

q

A+TC

q

ΛC

q

B

p

+B

T

p

C

T

q

Λ −2T ΛC

q

B

w

B

T

w

P B

T

w

C

T

q

Λ −I

_

¸

_ ≤ 0 (8.8)

holds, where Λ = diag(λ

1

, . . . , λ

np

) ≥ 0.

Of course, the ellipsoid E = {ξ ∈ R

n

¸

¸

ξ

T

Pξ ≤ 1} contains F, so that E ⊇

F ⊇ R

ue

. Therefore, E gives an outer approximation of R

ue

if the nonlinearities

are not known. If the nonlinearities φ

i

are known, F gives a possibly better outer

approximation. We can use these results to prove that a point x

0

does not belong to

the reachable set by solving appropriate LMIPs.

8.1.3 Output energy bounds

We consider the system (8.1) with initial condition x(0), and compute upper bounds

on the output energy

J =

_

∞

0

z

T

z dt

using Lyapunov functions V of the form (8.3).

If

d

dt

V (x) +z

T

z ≤ 0 for all x satisfying (8.1), (8.9)

then J ≤ V (x(0)). The condition (8.9) is equivalent to

2

_

ξ

T

P +

np

i=1

λ

i

π

i

C

i,q

_

(Aξ +B

p

π) +ξ

T

C

T

z

C

z

ξ ≤ 0,

for every ξ satisfying

π

i

(π

i

−c

i,q

ξ) ≤ 0, i = 1, . . . , n

p

.

This electronic version is for personal use and may not be duplicated or distributed.

122 Chapter 8 Lur’e and Multiplier Methods

Using the S-procedure, we conclude that the condition (8.9) holds if there exists

T = diag(τ

1

, . . . , τ

np

) ≥ 0 such that

_

A

T

P +PA+C

T

z

C

z

PB

p

+A

T

C

T

q

Λ +C

T

q

T

B

T

p

P + ΛC

q

A+TC

q

ΛC

q

B

p

+B

T

p

C

T

q

Λ −2T

_

≤ 0. (8.10)

Since V (x(0)) ≤ x(0)

T

(P + C

T

q

ΛC

q

)x(0), an upper bound on J is obtained

by solving the following EVP in the variables P, Λ = diag(λ

1

, . . . , λ

np

) and T =

diag(τ

1

, . . . , τ

np

):

minimize x(0)

T

_

P +C

T

q

ΛC

q

_

x(0)

subject to (8.10), T ≥ 0, Λ ≥ 0, P > 0

If the nonlinearities φ

i

are known, we can obtain possibly better bounds on J by

modifying the objective in the EVP appropriately.

8.1.4 L

2

gain

We assume that D

zw

= 0 for simplicity. If there exists a Lyapunov function of the

form (8.3) and γ ≥ 0 such that

d

dt

V (x) ≤ γ

2

w

T

w −z

T

z for all x and w satisfying (8.1), (8.11)

then the L

2

gain of the system (8.1) does not exceed γ. The condition (8.11) is

equivalent to

2

_

ξ

T

P +

np

i=1

λ

i

π

i

C

i,q

_

(Aξ +B

p

π) ≤ γ

2

w

T

w −ξ

T

C

T

z

C

z

ξ

for any ξ satisfying π

i

(π

i

−c

i,q

ξ) ≤ 0, i = 1, . . . , n

p

. Using the S-procedure, this is

satisﬁed if there exists T = diag(τ

1

, . . . , τ

np

) ≥ 0 such that

_

¸

_

A

T

P +PA+C

T

z

C

z

PB

p

+A

T

C

T

q

Λ +C

T

q

T PB

w

B

T

p

P + ΛC

q

A+TC

q

ΛC

q

B

p

+B

T

p

C

T

q

Λ −2T ΛC

q

B

w

B

T

w

P B

T

w

C

T

q

Λ −γ

2

I

_

¸

_ ≤ 0. (8.12)

The smallest upper bound on the L

2

gain, provable using Lur’e Lyapunov functions,

is therefore obtained by minimizing γ over γ, P, Λ and T subject to P > 0, Λ =

diag(λ

1

, . . . , λ

np

) ≥ 0, T = diag(τ

1

, . . . , τ

np

) ≥ 0 and (8.12). This is an EVP.

8.2 Integral Quadratic Constraints

In this section we consider an important variation on the NLDI, in which the point-

wise constraint p(t)

T

p(t) ≤ q(t)

T

q(t) is replaced with a constraint on the integrals of

p(t)

T

p(t) and q(t)

T

q(t):

˙ x = Ax +B

p

p +B

u

u +B

w

w,

q = C

q

x +D

qp

p +D

qu

u +D

qw

w,

z = C

z

x +D

zp

p +D

zu

u +D

zw

w

_

t

0

p(τ)

T

p(τ) dτ ≤

_

t

0

q(τ)

T

q(τ) dτ.

(8.13)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

8.2 Integral Quadratic Constraints 123

Such a system is closely related to the NLDI with the same data matrices; indeed,

every trajectory of the associated NLDI is a trajectory of (8.13). This system is

not, however, a diﬀerential inclusion. In control theory terms, the system (8.13) is

described as a linear system with (dynamic) nonexpansive feedback. An important

example is when p and q are related by a linear system, i.e.,

˙ x

f

= A

f

x

f

+B

f

q, p = C

f

x

f

+D

f

q, x

f

(0) = 0,

where D

f

+C

f

(sI −A

f

)

−1

B

f

∞

≤ 1.

We can also consider a generalization of the DNLDI, in which we have compo-

nentwise integral quadratic constraints. Moreover, we can consider integral quadratic

constraints with diﬀerent sector bounds; as an example we will encounter constraints

of the form

_

t

0

p(τ)

T

q(τ) dτ ≥ 0 in §8.3.

We will now show that many of the results on NLDIs (and DNLDIs) from Chap-

ters 5–7 generalize to systems with integral quadratic constraints. We will demon-

strate this generalization for stability analysis and L

2

gain bounds, leaving others to

the reader.

As in Chapters 5–7, our analysis will be based on quadratic functions of the state

V (ξ) = ξ

T

Pξ, and roughly speaking, the very same LMIs will arise. However, the

interpretation of V is diﬀerent here. For example, in stability analysis, the V of

Chapters 5–7 decreases monotonically to zero. Here, the very same V decreases to

zero, but not necessarily monotonically. Thus, V is not a Lyapunov function in the

conventional sense.

8.2.1 Stability

Consider system (8.13) without w and z. Suppose P > 0 and λ ≥ 0 are such that

d

dt

x

T

Px < λ

_

p

T

p −q

T

q

_

, (8.14)

or

d

dt

_

x

T

Px +λ

_

t

0

_

q(τ)

T

q(τ) −p(τ)

T

p(τ)

_

dτ

_

< 0.

Note that the second term is always nonnegative. Using standard arguments from

Lyapunov theory, it can be shown that lim

t→∞

x(t) = 0, or the system is stable.

Now, let us examine the condition (8.14). It is exactly the same as the condition

obtained by applying the S-procedure to the condition

d

dt

x

T

Px < 0, whenever p

T

p ≤ q

T

q,

which in turn, leads to the LMI condition for quadratic stability of NLDI (5.3)

P > 0, λ ≥ 0,

_

_

A

T

P +PA+λC

T

q

C

q

PB

p

+λC

T

q

D

qp

(PB

p

+λC

T

q

D

qp

)

T

−λ(I −D

T

qp

D

qp

)

_

_

< 0.

8.2.2 L

2

gain

We assume that x(0) = 0 and, for simplicity, that D

zw

= 0 and D

qp

= 0.

This electronic version is for personal use and may not be duplicated or distributed.

124 Chapter 8 Lur’e and Multiplier Methods

Suppose P > 0 and λ ≥ 0 are such that

d

dt

x

T

Px < λ

_

p

T

p −q

T

q

_

+γ

2

w

T

w −z

T

z. (8.15)

Integrating both sides from 0 to T,

x(T)

T

Px(T) +λ

_

T

0

_

q(τ)

T

q(τ) −p(τ)

T

p(τ)

_

dτ <

_

T

0

_

γ

2

w

T

w −z

T

z

_

dt.

This implies that the L

2

gain from w to z for system (8.13) does not exceed γ.

Condition (8.15) leads to the same LMI (6.55) which guarantees that the L

2

gain

of the NLDI (4.9) from w to z does not exceed γ.

Remark: Almost all the results described in Chapters 5, 6 and 7 extend imme-

diately to systems with integral quadratic constraints. The only exceptions are

the results on coordinate transformations (¸5.1.1, also ¸7.2.2), and on reachable

sets with unit-peak inputs (¸6.1.3 and ¸7.3.3).

8.3 Multipliers for Systems with Unknown Parameters

We consider the system

˙ x = Ax +B

p

p +B

w

w, q = C

q

x +D

qp

p,

p

i

= δ

i

q

i

, i = 1, . . . , n

p

, z = C

z

x +D

zw

w,

(8.16)

where p(t) ∈ R

np

, and δ

i

, i = 1, . . . , n

p

are any nonnegative numbers. We can consider

the more general case when α

i

≤ δ

i

≤ β

i

using a loop transformation (see the Notes

and References).

For reasons that will become clear shortly, we begin by deﬁning n

p

LTI systems

with input q

i

and output q

m,i

, where q

m,i

(t) ∈ R:

˙ x

m,i

= A

m,i

x

m,i

+B

m,i

q

i

, x

m,i

(0) = 0,

q

m,i

= C

m,i

x

m,i

+D

m,i

q

i

,

(8.17)

where we assume that A

m,i

is stable and (A

m,i

, B

m,i

) is controllable. We further

assume that

_

t

0

q

i

(τ)q

m,i

(τ) dτ ≥ 0 for all t ≥ 0.

This last condition is equivalent to passivity of the LTI systems (8.17), which in turn

is equivalent to the existence of P

i

≥ 0, i = 1, . . . , n

p

, satisfying

_

A

T

m,i

P

i

+P

i

A

m,i

P

i

B

m,i

−C

T

m,i

B

T

m,i

P

i

−C

m,i

−

_

D

m,i

+D

T

m,i

_

_

≤ 0, i = 1, . . . , n

p

. (8.18)

(See §6.3.3.)

Thus, q and q

m

∆

= [q

m,1

, . . . , q

m,np

]

T

are the input and output respectively of the

diagonal system with realization

A

m

= diag(A

m,i

), B

m

= diag(B

m,i

), C

m

= diag(C

m,i

), D

m

= diag(D

m,i

),

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

8.3 Multipliers for Systems with Unknown Parameters 125

Deﬁning

˜

A =

_

A 0

B

m

C

q

A

m

_

,

˜

B

p

=

_

B

p

B

m

D

qp

_

,

˜

B

w

=

_

B

w

0

_

,

˜

C

q

= [D

m

C

q

C

m

],

˜

D

qp

= D

m

D

qp

,

˜

C

z

= [C

z

0].

consider the following system with integral quadratic constraints:

d˜ x

dt

=

˜

A˜ x +

˜

B

p

p +

˜

B

w

w,

q

m

=

˜

C

q

˜ x +

˜

D

qp

p,

z =

˜

C

z

˜ x +

˜

D

qp

p +D

zw

w,

_

t

0

p

i

(τ)q

m,i

(τ) dτ ≥ 0, i = 1, . . . , n

p

.

(8.19)

It is easy to show that if x is a trajectory of (8.16), then [x

T

x

T

m

]

T

is a trajectory

of (8.19) for some appropriate x

m

. Therefore, conclusions about (8.16) have implica-

tions for (8.19). For instance, if system (8.19) is stable, so is system (8.16).

8.3.1 Stability

Consider system (8.19) without w and z. Suppose P > 0 is such that

d

dt

˜ x

T

P ˜ x + 2

np

i=1

p

i

q

m,i

< 0, (8.20)

or

d

dt

_

˜ x

T

P ˜ x + 2

np

i=1

_

t

0

p

i

(τ)q

m,i

(τ) dτ

_

< 0.

Then, it can be shown using standard arguments that lim

t→∞

˜ x(t) = 0, i.e., the

system (8.19) is stable.

Condition (8.20) is just the LMI in P > 0, C

m

and D

m

:

_

˜

A

T

P +P

˜

A P

˜

B

p

+

˜

C

T

q

˜

B

T

p

P +

˜

C

q

˜

D

qp

+

˜

D

T

qp

_

< 0. (8.21)

Remark: LMI (8.21) can also be interpreted as a positive dissipation con-

dition for the LTI system with transfer matrix G(s) = −W(s)H(s), where

W(s) = Cm(sI −Am)

−1

Bm+Dm and H(s) = Cq(sI −A)

−1

Bp +Dqp. Therefore,

we conclude that system (8.16) is stable if there exists a passive W such that

−W(s)H(s) has positive dissipation. Hence W is often called a “multiplier”. See

the Notes and References for details.

8.3.2 Reachable sets with unit-energy inputs

The set of reachable states with unit-energy inputs for the system

˙ x = Ax +B

p

p +B

w

w, q = C

q

x +D

qp

p,

p = δq, δ ≥ 0,

(8.22)

This electronic version is for personal use and may not be duplicated or distributed.

126 Chapter 8 Lur’e and Multiplier Methods

is

R

ue

∆

=

_

¸

¸

_

¸

¸

_

x(T)

¸

¸

¸

¸

¸

¸

¸

¸

x, w satisfy (8.22), x(0) = 0

_

T

0

w

T

w dt ≤ 1, T ≥ 0

_

¸

¸

_

¸

¸

_

.

We can obtain a suﬃcient condition for a point x

0

to lie outside the reachable set R

ue

,

using quadratic functions for the augmented system. Suppose P > 0 is such that

d

dt

˜ x

T

P ˜ x + 2

np

i=1

p

i

q

m,i

< w

T

w, (8.23)

or

˜ x

T

P ˜ x + 2

np

i=1

_

t

0

p

i

(τ)q

m,i

(τ) dτ <

_

t

0

w(τ)

T

w(τ) dτ.

then the set E =

_

˜ x

¸

¸

˜ x

T

P ˜ x ≤ 1

_

contains the reachable set for the system (8.19).

Condition (8.23) is equivalent to the LMI in P > 0, C

m

and D

m

:

_

¸

_

˜

A

T

P +P

˜

A P

˜

B

w

P

˜

B

p

+

˜

C

T

q

˜

B

T

w

P −I 0

˜

B

T

p

P +

˜

C

q

0

˜

D

qp

+

˜

D

T

qp

_

¸

_ < 0. (8.24)

Now, the point x

0

lies outside the set of reachable states for the system (8.22) if

there exists P satisfying (8.23) and there exists no z such that

_

x

0

z

_

T

P

_

x

0

z

_

≤ 1.

Partitioning P conformally with the sizes of x

0

and z as

P =

_

P

11

P

12

P

T

12

P

22

_

,

we note that

min

z

_

x

0

z

_

T

_

P

11

P

12

P

T

12

P

22

__

x

0

z

_

= x

T

0

(P

11

−P

12

P

−1

22

P

T

12

)x

0

.

Therefore, x

0

does not belong to the reachable set for the system (8.22) if there exist

P > 0, P

i

≥ 0, λ ≥ 0, C

m

, and D

M

satisfying (8.18), (8.24) and

_

x

T

0

P

11

x

0

−1 x

T

0

P

12

P

T

12

x

0

P

22

_

> 0.

This is an LMIP.

Notes and References

Lur’e systems

The Notes and References of Chapters 5 and 7 are relevant to this chapter as well. Lur’e

and Postnikov were the ﬁrst to propose Lyapunov functions consisting of a quadratic form

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 127

plus the integral of the nonlinearity [LP44]. The book [Lur57] followed this article, in which

the author showed how to solve the problem of ﬁnding such Lyapunov functions analyti-

cally for systems of order 2 or 3. Another early book that uses such Lyapunov functions is

Letov [Let61]. Work on this topic was continued by Popov [Pop62], Yakubovich [Yak64].

Tsypkin [Tsy64d, Tsy64c, Tsy64a, Tsy64b], Szego [Sze63] and Meyer [Mey66]. Jury

and Lee [JL65] consider the case of Lur’e systems with multiple nonlinearities and derive a

corresponding stability criterion. See also [IR64].

The ellipsoid method is used to construct Lyapunov functions for the Lur’e stability problem

in the papers by Pyatnitskii and Skorodinskii [PS82, PS83]. See also [Sko91a, Sko91b] for

general numerical methods to prove stability of nonlinear systems via the Popov or circle

criterion. A precursor is the paper by Karmarkar and Siljak [KS75], who used a locally con-

vergent algorithm to determine margins for a Lur’e-Postnikov system with one nonlinearity.

Indeed, in the early books by Lur’e, Postnikov, and Letov the problem of constructing Lya-

punov functions reduces to the satisfaction of some inequalities. In some special cases they

even point out geometrical properties of the regions deﬁned by the inequalities, i.e., that they

form a parallelogram. But as far as we know, convexity of the regions is not noted in the

early work. In any case, it was not known in the 1940’s and 1950’s that solution of convex

inequalities is practically and theoretically tractable.

Construction of Lyapunov functions for systems with integral quadratic constraints

We saw in ¸8.2.1 that the quadratic positive function V (ξ) = ξ

T

Pξ is not necessarily a

Lyapunov function in the conventional sense for the system (8.13) with integral quadratic

constraints; although V (x(t)) →0 as t →∞, it does not do so monotonically. We now show

how we can explicitly construct a Lyapunov function given P, in cases when p and q are

related by

˙ x

f

= f(x

f

, q, t), p = g(x

f

, q, t), (8.25)

where x

f

(t) ∈ R

n

f

.

Since

_

t

0

p

T

pdτ ≤

_

t

0

q

T

q dτ, it can be shown (see for example [Wil72]) that V

f

: R

n

f

→R+

given by

V

f

(ξ

f

)

∆

= sup

_

−

_

T

0

_

q

T

q −p

T

p

_

dt

¸

¸

¸

¸

x

f

(0) = ξ

f

, x

f

satisﬁes (8.25), T ≥ 0

_

is a Lyapunov function that proves stability of system (8.25). Moreover, the Lyapunov

function V : R

n

R

n

f

→R+ given by

V (ξ, ξ

f

)

∆

= ξ

T

Pξ +V

f

(ξ

f

)

proves stability of the interconnection

˙ x = Ax +Bpp, q = Cqx +Dqpp,

˙ x

f

= f(x

f

, q, t), p = g(x

f

, q, t).

Note that this construction also shows that ξ

T

Pξ proves the stability of the NLDI (5.2).

Integral quadratic constraints have recently received much attention, especially in the former

Soviet Union. See for example [Yak88, Meg92a, Meg92b, Yak92, Sav91, SP94b, SP94h,

SP94d].

Systems with constant, unknown parameters

Problems involving system (8.16) are widely studied in the ﬁeld of robust control, sometimes

under the name “real-µ problems”; see Siljak [Sil89] for a survey. The results of ¸8.3 were

This electronic version is for personal use and may not be duplicated or distributed.

128 Chapter 8 Lur’e and Multiplier Methods

derived by Fan, Tits and Doyle in [FTD91] using an alternate approach, involving the ap-

plication of an oﬀ-axis circle criterion (see also ¸3.3). We also mention [TV91], where Tesi

and Vicino study a hybrid system, consisting of a Lur’e system with unknown parameters.

There are elegant analytic solutions for some very special problems involving parameter-

dependent linear systems, e.g., Kharitonov’s theorem [Kha78] and its extensions [BHL89].

A nice survey of these results can be found in Barmish [Bar93].

Complexity of stabilization problems

The stability (and performance) analysis problems considered in ¸8.3 have high complexity.

For example, checking if (8.16) is stable is NP-hard (see Coxson and Demarco [CD91], Braatz

and Young [BYDM93], Poljak and Rohn [PR94], and Nemirovskii [Nem94]). We conjecture

that checking whether a general DNLDI is stable is also NP-hard.

Likewise, many stabilization problems for uncertain systems are NP-hard. For instance, the

problem of checking whether system (8.16) is stabilizable by a constant state-feedback can

be shown to be NP-hard using the method of [Nem94].

A related result, due to Blondel and Gevers [BG94], states that checking whether there

exists a common stabilizing LTI controller for three LTI systems is undecidable. In contrast,

checking whether there exists a static, output-feedback control law for a single LTI system

is rationally decidable [ABJ75].

Multiplier methods

See [Wil69a, NT73, DV75, Sil69, Wil70, Wil76] for discussion of and bibliography on

multiplier theory and its connections to Lyapunov stability theory. In general multiplier

theory we consider the system

˙ x = Ax +Bpp +Bww, q = Cqx +Dqpp, p(t) = ∆(q, t).

The method involves checking that Hm = W

_

Cq(sI −A)

−1

Bp +Dqp

_

satisﬁes

Hm(jω) +Hm(jω)

∗

≥ 0 for every ω ∈ R,

for some choice of W from a set that is determined by the properties of ∆. When ∆ is

a nonlinear time-invariant memoryless operator satisfying a sector condition (i.e., when we

have a Lur’e system), the multipliers W are of the form (1 + qs) for some q ≥ 0. This is

the famous Popov criterion [Pop62, Pop64]; we must also cite Yakubovich, who showed,

using the S-procedure [Yak77], that feasibility of LMI (8.6) is equivalent to the existence of

q nonnegative such that (1 +qs)(cq(sI −A)

−1

bp) + 1/k has positive dissipation.

In the case when ∆(q, t) = δq(t) for some unknown real number δ, the multiplier W can be

any passive transfer matrix. This observation was made by Brockett and Willems in [BW65];

given a transfer function Hqp they show that the transfer function Hqp/(1+kHqp) is stable for

all values of k in (0, ∞) if and only if there exists a passive multiplier W such that WHqp is

also passive. This case has also been considered in [CS92b, SC93, SL93a, SLC94] where the

authors devise appropriate multipliers for constant real uncertainties. The paper [BHPD94]

by Balakrishnan et al., shows how various stability tests for uncertain systems can be red-

erived in the context of multiplier theory, and how these tests can be reduced to LMIPs. See

also [LSC94].

Safonov and Wyetzner use a convex parametrization of multipliers found by Zames and

Falb [ZF68, ZF67] to prove stability of systems subject to “monotonic” or “odd-monotonic”

nonlinearities (see [SW87] for details; see also [GG94]). Hall and How, along with Bern-

stein and Haddad, generalize the use of quadratic Lyapunov functions along with multipliers

for various classes of nonlinearities, and apply them to Aerospace problems; see for ex-

ample [HB91a, How93, HH93b, HHHB92, HH93a, HHH93a, HCB93, HB93b, HB93a,

HHH93b]. See also [CHD93].

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 129

Multiple nonlinearities for Lur’e systems

Our analysis method for Lur’e systems with multiple nonlinearities in ¸8.1 can be conser-

vative, since the S-procedure can be conservative in this case. Rapoport [Rap86, Rap87,

Rap88] devises a way of determining if there exists Lyapunov functions of the form (8.3)

nonconservatively by generating appropriate LMIs. We note, however, that the number of

these LMIs grows exponentially with the number of nonlinearities. Kamenetskii [Kam89]

and Rapoport [Rap89] derive corresponding frequency-domain stability criteria. Kamenet-

skii calls his results “convolution method for solving matrix inequalities”.

Loop transformations

Consider the Lur’e system with general sector conditions, i.e.,

˙ x = Ax +Bp, q = Cx +Dp,

pi(t) = φi(qi(t)), αiσ

2

≤ σφi(σ) ≤ βiσ

2

.

(8.26)

Here we have dropped the variables w and z and the subscripts on B, C and D to simplify

the presentation. We assume this system is well-posed, i.e., det(I −D∆) ,= 0 for all diagonal

∆ with αi ≤ ∆ii ≤ βi.

Deﬁne

¯ pi

∆

=

1

βi −αi

(φi(qi) −αiqi) =

¯

φi(qi).

It is readily shown that 0 ≤ σ

¯

φi(σ) ≤ σ

2

for all σ. Let Λ and Γ denote the diagonal matrices

Λ = diag(α1, . . . , αnq

), Γ = diag(β1 −α1, . . . , βnq

−αnq

),

so that ¯ p = Γ

−1

(p − Λq). We now substitute p = Γ¯ p + Λq into (8.26) and, using our well-

posedness assumption, solve for ˙ x and q in terms of x and ¯ p. This results in

˙ x =

_

A+BΛ(I −DΛ)

−1

C

_

x +B(I −ΛD)

−1

Γ¯ p,

q = (I −DΛ)

−1

Cx + (I −DΛ)

−1

DΓ¯ p.

We can therefore express (8.26) as

˙ x =

¯

Ax +

¯

B¯ p, q =

¯

Cx +

¯

D¯ p,

¯ pi(t) =

¯

φi(qi(t)), 0 ≤ σ

¯

φi(σ) ≤ σ

2

.

(8.27)

where

¯

A = A+BΛ(I −DΛ)

−1

C,

¯

B = B(I −ΛD)

−1

Γ,

¯

C = (I −DΛ)

−1

C,

¯

D = (I −DΛ)

−1

DΓ.

Note that (8.27) is in the standard Lur’e system form. So to analyze the more general Lur’e

system (8.26), we simply apply the methods of ¸8.1 to the loop-transformed system (8.27).

In practice, i.e., in a numerical implementation, it is probably better to derive the LMIs

associated with the more general Lur’e system than to loop transform to the standard Lur’e

system.

The construction above is a loop transformation that maps nonlinearities in sector [αi, βi]

into the standard sector, i.e., [0, 1]. Similar transformations can be used to map any set of

sectors into any other, including so-called inﬁnite sectors, in which, say, βi = +∞.

The term loop transformation comes from a simple block-diagram interpretation of the equa-

tions given above. For detailed descriptions of loop transformations, see the book by Desoer

and Vidyasagar [DV75, p50].

This electronic version is for personal use and may not be duplicated or distributed.

Chapter 9

Systems with Multiplicative Noise

9.1 Analysis of Systems with Multiplicative Noise

9.1.1 Mean-square stability

We ﬁrst consider the discrete-time stochastic system

x(k + 1) =

_

A

0

+

L

i=1

A

i

p

i

(k)

_

x(k), (9.1)

where p(0), p(1), . . ., are independent, identically distributed random variables with

Ep(k) = 0, Ep(k)p(k)

T

= Σ = diag(σ

1

, . . . , σ

L

). (9.2)

We assume that x(0) is independent of the process p.

Deﬁne M(k), the state correlation matrix, as

M(k)

∆

= Ex(k)x(k)

T

.

Of course, M satisﬁes the linear recursion

M(k + 1) = AM(k)A

T

+

L

i=1

σ

2

i

A

i

M(k)A

T

i

, M(0) = Ex(0)x(0)

T

. (9.3)

If this linear recursion is stable, i.e., regardless of x(0), lim

k→∞

M(k) = 0, we say

the system is mean-square stable. Mean-square stability implies, for example, that

x(k) →0 almost surely.

It can be shown (see the Notes and References) that mean-square stability is

equivalent to the existence of a matrix P > 0 satisfying the LMI

A

T

PA−P +

L

i=1

σ

2

i

A

T

i

PA

i

< 0. (9.4)

An alternative, equivalent condition is that the dual inequality

AQA

T

−Q+

L

i=1

σ

2

i

A

i

QA

T

i

< 0 (9.5)

holds for some Q > 0.

131

132 Chapter 9 Systems with Multiplicative Noise

Remark: For P > 0 satisfying (9.4), we can interpret the function V (ξ) = ξ

T

Pξ

as a stochastic Lyapunov function: EV (x) decreases along trajectories of (9.1).

Alternatively, the function V (M) = Tr MP is a (linear) Lyapunov function for

the deterministic system (9.3) (see the Notes and References).

Remark: Mean-square stability can be veriﬁed directly by solving the Lyapunov

equation in P

A

T

PA−P +

L

i=1

σ

2

i

A

T

i

PAi +I = 0

and checking whether P > 0. Thus the LMIPs (9.4) and (9.5) oﬀer no computa-

tional (or theoretical) advantages for checking mean-square stability.

We can consider variations on the mean-square stability problem, for example,

determining the mean-square stability margin. Here, we are given A

0

, . . . , A

L

, and

asked to ﬁnd the largest γ such that with Σ < γ

2

I, the system (9.1) is mean-square

stable. From the LMIP characterizations of mean-square stability, we can derive

GEVP characterizations that yield the exact mean-square stability margin. Again,

the GEVP oﬀers no computational or theoretical advantages, since the mean-square

stability margin can be obtained by computing the eigenvalue of a (large) matrix.

9.1.2 State mean and covariance bounds with unit-energy inputs

We now add an exogenous input w to our stochastic system:

x(k + 1) = Ax(k) +B

w

w(k) +

L

i=1

(A

i

x(k) +B

w,i

w(k)) p

i

(k), x(0) = 0. (9.6)

We assume that the exogenous input w is deterministic, with energy not exceeding

one, i.e.,

∞

k=0

w(k)

T

w(k) ≤ 1.

Let ¯ x(k) denote the mean of x(k), which satisﬁes ¯ x(k +1) = A¯ x(k) +B

w

w(k), and let

X(k) denote the covariance of x(k), i.e., X(k) = E(x(k) − ¯ x(k)) (x(k) − ¯ x(k))

T

. We

will develop joint bounds on ¯ x(k) and X(k).

Suppose V (ξ) = ξ

T

Pξ where P > 0 satisﬁes

EV (x(k + 1)) ≤ EV (x(k)) +w(k)

T

w(k) (9.7)

for all trajectories. Then,

EV (x(k)) ≤

k

j=0

w(j)

T

w(j), k ≥ 0,

which implies that, for all k ≥ 0,

Ex(k)

T

Px(k) = ¯ x(k)

T

P ¯ x(k) +E(x(k) − ¯ x(k))

T

P (x(k) − ¯ x(k))

= ¯ x(k)

T

P ¯ x(k) +Tr X(k)P

≤ 1.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

9.1 Analysis of Systems with Multiplicative Noise 133

Let us now examine condition (9.7). We note that

EV (x(k + 1)) =

_

¯ x(k)

w(k)

_

T

M

_

¯ x(k)

w(k)

_

+Tr

_

A

T

PA+

L

i=1

σ

2

i

A

T

i

PA

i

_

X(k),

where

M =

_

A

T

B

T

w

_

P

_

A B

w

_

+

L

i=1

σ

2

i

_

A

T

i

B

T

w,i

_

P

_

A

i

B

w,i

_

.

Next, the LMI conditions

_

A

T

B

T

w

_

P

_

A B

w

_

+

L

i=1

σ

2

i

_

A

T

i

B

T

w,i

_

P

_

A

i

B

w,i

_

≤

_

P 0

0 I

_

(9.8)

and

A

T

PA+

L

i=1

σ

2

i

A

T

i

PA

i

< P (9.9)

imply

EV (x(k + 1)) ≤ ¯ x

T

(k)P ¯ x(k) +w(k)

T

w(k)

+Tr

_

A

T

PA+

L

i=1

σ

2

i

A

T

i

PA

i

_

X(k)

≤ ¯ x

T

(k)P ¯ x(k) +w(k)

T

w(k)

+E(x(k) − ¯ x(k))

T

P (x(k) − ¯ x(k))

= EV (x(k)) +w(k)

T

w(k), k ≥ 0.

Of course, LMI (9.8) implies LMI (9.9); thus, we conclude that LMI (9.8) im-

plies (9.7). Therefore, we conclude that ¯ x(k) and X(k) must belong to the set

_

(

¯

x(k), X(k))

¸

¸

¯ x(k)

T

P ¯ x(k) +Tr X(k)P ≤ 1

_

for any P > 0 satisfying the LMI (9.8). For example, we have the bound Tr M(k)P ≤ 1

on the state covariance matrix.

We can derive further bounds on M(k). For example, since we have Tr M(k)P ≤

λ

min

(P) Tr M(k) and therefore Tr M(k) ≤ 1/λ

min

(P), we can derive bounds on Tr M

by maximizing λ

min

(P) subject to (9.8) and P > 0. This is an EVP. As another

example, we can derive an upper bound on λ

max

(M) by maximizing Tr P subject

to (9.8) and P > 0, which is another EVP.

9.1.3 Bound on L

2

gain

We now consider the system

x(k + 1) = Ax(k) +B

w

w(k) +

L

i=1

(A

i

x(k) +B

w,i

w(k)) p

i

(k), x(0) = 0,

z(k) = C

z

x(k) +D

zw

w(k) +

L

i=1

(C

z,i

x(k) +D

zw,i

w(k)) p

i

(k).

(9.10)

This electronic version is for personal use and may not be duplicated or distributed.

134 Chapter 9 Systems with Multiplicative Noise

with the same assumptions on p. We assume that w is deterministic.

We deﬁne the L

2

gain η of this system as

η

2

∆

= sup

_

E

∞

k=0

z(k)

T

z(k)

¸

¸

¸

¸

¸

∞

k=0

w(k)

T

w(k) ≤ 1

_

. (9.11)

Suppose that V (ξ) = ξ

T

Pξ, with P > 0, satisﬁes

EV (x(k + 1)) −EV (x(k)) ≤ γ

2

w(k)

T

w(k) −Ez(k)

T

z(k). (9.12)

Then γ ≥ η. The condition (9.12) can be shown to be equivalent to the LMI

_

A B

w

C

z

D

zw

_

T

_

P 0

0 I

__

A B

w

C

z

D

zw

_

−

_

P 0

0 γ

2

I

_

+

L

i=1

σ

2

i

_

A

i

B

w,i

C

z,i

D

zw,i

_

T

_

P 0

0 I

__

A

i

B

w,i

C

z,i

D

zw,i

_

≤ 0.

(9.13)

Minimizing γ subject to LMI (9.13), which is an EVP, yields an upper bound on η.

Remark: It is straightforward to apply the scaling method developed in ¸6.3.4

to obtain componentwise results.

9.2 State-Feedback Synthesis

We now add a control input u to our system:

x(k + 1) = Ax(k) +B

u

u(k) +B

w

w(k)+

L

i=1

(A

i

x(k) +B

u,i

u(k) +B

w,i

w(k)) p

i

(k),

z(k) = C

z

x(k) +D

zu

u(k) +D

zw

w(k)+

L

i=1

(C

z,i

x(k) +D

zu,i

u(k) +D

zw,i

w(k)) p

i

(k).

(9.14)

We seek a state-feedback u(k) = Kx(k) such that the closed-loop system

x(k + 1) = (A+B

u

K)x(k) +B

w

w(k)+

L

i=1

((A

i

+B

u,i

K)x(k) +B

w,i

w(k)) p

i

(k),

z(k) = (C

z

+D

zu

K)x(k) +D

zw

w(k)+

L

i=1

((C

z,i

+D

zu,i

K)x(k) +D

zw,i

w(k)) p

i

(k),

(9.15)

satisﬁes various properties.

9.2.1 Stabilizability

We seek K and Q > 0 such that

(A+B

u

K)Q(A+B

u

K)

T

−Q

+

L

i=1

σ

2

i

(A

i

+B

u,i

K)Q(A

i

+B

u,i

K)

T

< 0.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

9.2 State-Feedback Synthesis 135

With the change of variable Y = KQ, we obtain the equivalent condition

(AQ+B

u

Y )Q

−1

(AQ+B

u

Y )

T

−Q

+

L

i=1

σ

2

i

(A

i

Q+B

u,i

Y )Q

−1

(A

i

Q+B

u,i

Y )

T

< 0,

(9.16)

which is easily written as an LMI in Q and Y .

Thus, system (9.14) is mean-square stabilizable (with constant state-feedback) if

and only if the LMI Q > 0, (9.16) is feasible. In this case, a stabilizing state-feedback

gain is K = Y Q

−1

.

As an extension, we can maximize the closed-loop mean-square stability margin.

System (9.15) is mean-square stable for Σ ≤ γ

2

I if and only if

(AQ+B

u

Y )Q

−1

(AQ+B

u

Y )

T

−Q

+γ

2

L

i=1

(A

i

Q+B

u,i

Y )Q

−1

(A

i

Q+B

u,i

Y )

T

< 0

(9.17)

holds for some Y and Q > 0. Therefore, ﬁnding K that maximizes γ reduces to the

following GEVP:

maximize γ

subject to Q > 0, (9.17)

9.2.2 Minimizing the bound on L

2

gain

We now seek a state-feedback gain K which minimizes the bound on L

2

gain (as

deﬁned by (9.11)) for system (9.15). η ≤ γ for some K if there exist P > 0 and K

such that

L

i=1

σ

2

i

_

A

i

+B

u,i

K B

w,i

C

z,i

+D

zu,i

K D

zw,i

_

T

˜

P

_

A

i

+B

u,i

K B

w,i

C

z,i

+D

zu,i

K D

zw,i

_

+

_

A+B

u

K B

w

C

z

+D

zu

K D

zw

_

T

˜

P

_

A+B

u

K B

w

C

z

+D

zu

K D

zw

_

≤

_

P 0

0 γ

2

I

_

where

˜

P =

_

P 0

0 I

_

.

We make the change of variables Q = P

−1

, Y = KQ. Applying a congruence

with

˜

Q = diag(Q, I), we get

L

i=1

σ

2

i

_

A

i

Q+B

u,i

Y B

w,i

C

z,i

Q+D

zu,i

Y D

zw,i

_

T

˜

Q

−1

_

A

i

Q+B

u,i

Y B

w,i

C

z,i

Q+D

zu,i

Y D

zw,i

_

+

_

AQ+B

u

Y B

w

C

z

Q+D

zu

Y D

zw

_

T

˜

Q

−1

_

AQ+B

u

Y B

w

C

z

Q+D

zu

Y D

zw

_

≤

_

Q 0

0 γ

2

I

_

.

This inequality is readily transformed to an LMI.

This electronic version is for personal use and may not be duplicated or distributed.

136 Chapter 9 Systems with Multiplicative Noise

Notes and References

Systems with multiplicative noise

Systems with multiplicative noise are a special case of stochastic systems. For an introduction

to this topic see Arnold [Arn74] or

˚

Astr¨ om [˚Ast70].

A rigorous treatment of continuous-time stochastic systems requires much technical machin-

ery (see [EG94]). In fact there are many ways to formulate the continuous-time version of a

system such as (9.10); see Itˆ o [Ito51] and Stratonovitch [Str66].

Stability of system with multiplicative noise

For a survey of diﬀerent concepts of stability for stochastic systems see [Koz69] and the

references therein. The deﬁnition of mean-square stability can be found in [Sam59]. Mean-

square stability is a strong form of stability. In particular, it implies stability of the mean

Ex(k) and that all trajectories converge to zero with probability one [Kus67, Wil73]. For the

systems considered here, it is also equivalent to L2-stability, which is the property that (see

for example, [EP92]) E

T

t=0

|x(t)|

2

has a (ﬁnite) limit as T →∞ for all initial conditions.

The equation for the state correlation matrix (9.3) can be found in [McL71, Wil73]. An alge-

braic criterion for mean-square stability of a system with multiplicative white noise is given

in [NS72]. It is reminiscent of the classical (deterministic) Hurwitz criterion. Frequency-

domain criteria are given in [WB71]. The related Lyapunov equation was given by Klein-

mann [Kle69] for the case L = 1 (i.e., a single uncertainty). The criterion then reduces

to an H2 norm condition on a certain transfer matrix. (This is in parallel with the results

of [EP92], see below.)

Kats and Krasovskii introduced in [KK60] the idea of a stochastic Lyapunov function and

proved a related Lyapunov theorem. A thorough stability analysis of (nonlinear) stochastic

systems was made by Kushner in [Kus67] using this idea of a stochastic Lyapunov func-

tion. Kushner used this idea to derive bounds for return time, expected output energy, etc.

However, this approach does not provide a way to construct the Lyapunov functions.

Proof of stability criterion

We now prove that condition (9.4) is a necessary and suﬃcient condition for mean-square

stability. We ﬁrst prove that condition (9.4) is suﬃcient. Let P > 0 satisfy (9.4). Introduce

the (linear Lyapunov) function V (M) = Tr MP, which is positive on the cone of nonnegative

matrices. For nonzero M(k) satisfying (9.3),

V (M(k + 1)) = Tr

_

AM(k)A

T

+

L

i=1

σ

2

i

AiM(k)A

T

i

_

P

= Tr M(k)

_

A

T

PA+

L

i=1

σ

2

i

A

T

i

PAi

_

< Tr M(k)P = V (M(k)).

A further standard argument from Lyapunov theory proves that M(k) converges to zero as

k →∞.

To prove that condition (9.4) is also necessary, we assume that the system is mean-square

stable. This means that for every positive initial condition M(0) ≥ 0, the solution to the

linear recursion (9.3) goes to zero as k → 0. Linearity of system (9.3) implies that, for an

arbitrary choice of M(0), the corresponding solution M(k) converges to 0, i.e., (9.3) is stable.

We now write (9.3) as m(k +1) = /m(k), where m(k) is a vector containing the n

2

elements

of M(k), and / is a real matrix. (This matrix can be expressed via Kronecker products

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 137

involving A0, . . . , AL.) Stability of / is equivalent to that of /

T

, and the corresponding

matrix diﬀerence equation,

N(k + 1) = A

T

N(k)A+

L

i=1

σ

2

i

A

T

i

N(k)Ai, (9.18)

is also stable.

Now let N(0) > 0, and let N(k) be the corresponding solution of (9.18). Since N(k) satisﬁes

a stable ﬁrst-order linear diﬀerence equation with constant coeﬃcients, the function

P(k) =

k

j=0

N(j)

has a (ﬁnite) limit for k →∞, which we denote by P. The fact that N(0) > 0 implies P > 0.

Now (9.18) implies

P(k + 1) = N(0) +A

T

P(k)A+

L

i=1

σ

2

i

A

T

i

P(k)Ai

Taking the limit k →∞ shows that P satisﬁes (9.4).

Mean-square stability margin

System (9.10) can be viewed as a linear system subject to random parameter variations. For

systems with deterministic parameters only known to lie in ranges, the computation of an

exact stability margin is an NP-hard problem [CD92]. The fact that a stochastic analog

of the deterministic stability margin is easier to compute should not be surprising; we are

dealing with a very special form of stability, which only considers the average behavior of the

system.

An exact stability margin can also be computed for continuous-time systems with multiplica-

tive noise provided the Itˆ o framework is considered; see [EG94]. In contrast, the computation

of an exact stability margin seems diﬃcult for Stratonovitch systems.

In [EP92], El Bouhtouri and Pritchard provide a complete robustness analysis in a slightly

diﬀerent framework than ours, namely for so-called “block-diagonal” perturbations. Roughly

speaking, they consider a continuous-time Itˆ o system of the form

dx(t) = Ax(t)dt +

L

i=1

Bi∆i(Cx(t))dpi(t),

where the operators ∆i are Lipschitz continuous, with ∆i(0) = 0; the processes pi, i =

1, . . . , L, are standard Brownian motions (which, roughly speaking, are integrals of white

noise).

We note that our framework can be recovered by assuming that the operators ∆i are restricted

to a diagonal structure of the form ∆i(z) = σiz.

Measuring the “size” of the perturbation term by

_

L

i=1

|∆i|

2

_

1/2

, where | | denotes the

Lipschitz norm, leads the authors of [EP92] to the robustness margin

minimize γ

subject to P > 0, A

T

P +PA+C

T

C < 0, B

T

i

PBi ≤ γ, i = 1, . . . , L

This electronic version is for personal use and may not be duplicated or distributed.

138 Chapter 9 Systems with Multiplicative Noise

For L = 1, and B1 and C are vectors, that is, when a single scalar uncertain parameter

perturbs an LTI system, both this result and ours coincide with the H2 norm of (A, B1, C).

This is consistent with the earlier results of Kleinmann [Kle69] and Willems [WB71] (see

also [BB91, p114] for a related result). For L > 1 however, applying their results to our

framework would be conservative, since we only consider “diagonal” perturbations with “re-

peated elements”, ∆i = σiI. Note that the LMI approach can also solve the state-feedback

synthesis problem in the framework of [EP92]. El Bouhtouri and Pritchard provide a solution

of this problem in [EP94].

Stability conditions for arbitrary noise intensities are given in geometric terms in [WW83].

In our framework, this happens if and only if the optimal value of the corresponding GEVP

is zero.

Bounds on state covariance with noise input

We consider system (9.6) in which w is a unit white noise process, i.e., w(k) are independent,

identically distributed with Ew(k) = 0, Ew(k)w(k)

T

= W, and w is independent of p.

In this case, of course, the state mean ¯ x is zero. We derive an upper bound on the state

correlation (or covariance) of the system.

The state correlation M(k) satisﬁes the diﬀerence equation

M(k + 1) = AM(k)A

T

+BwWB

T

w

+

L

i=1

σ

2

i

_

AiM(k)A

T

i

+Bw,iWB

T

w,i

_

,

with M(0) = 0.

Since the system is mean-square stable, the state correlation M(k) has a (ﬁnite) limit M∞

which satisﬁes

M∞ = AM∞A

T

+BwWB

T

w

+

L

i=1

σ

2

i

_

AiM∞A

T

i

+Bw,iWB

T

w,i

_

.

In fact, the matrix M∞ can be computed directly as the solution of a linear matrix equation.

However, the above LMI formulation extends immediately to more complicated situations (for

instance, when the white noise input w has a covariance matrix with unknown oﬀ-diagonal

elements), while the “direct” method does not. See the Notes and References for Chapter 6

for details on how these more complicated situations can be handled.

Extension to other stochastic systems

The Lur’e stability problem considered in ¸8.1 can be extended to the stochastic framework;

see Wonham [Won66]. This extension can be cast in terms of LMIs.

It is also possible to extend the results in this chapter to “uncertain stochastic systems”, of

the form

x(k + 1) =

_

A(k) +

L

i=1

σiAipi(k)

_

x(k),

where the time-varying matrix A(k) is only known to belong to a given polytope.

State-feedback synthesis

Conditions for stabilizability for arbitrary noise intensities are given in geometric terms

in [WW83]. Apart from this work, the bulk of earlier research on this topic concentrated

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 139

on Linear-Quadratic Regulator theory for continuous-time Itˆ o systems, which addresses the

problem of ﬁnding an input u that minimize a performance index of the form

J(K) = E

_

∞

0

_

x(t)

T

Qx(t) +u(t)

T

Ru(t)

_

dt

where Q ≥ 0 and R > 0 are given. The solution of this problem, as found in [McL71]

or [Ben92], can be expressed as a (linear, constant) state-feedback u(t) = Kx(t) where the

gain K is given in terms of the solution of a non-standard Riccati equation. (See [Kle69,

Won68, FR75, BH88b] for additional details.) The existing methods (see e.g. [Phi89,

RSH90]) use homotopy algorithms to solve these equations, with no guarantee of global con-

vergence. We can solve these non-standard Riccati equations reliably (i.e., with guaranteed

convergence), by recognizing that the solution is an extremal point of a certain LMI feasibility

set.

This electronic version is for personal use and may not be duplicated or distributed.

Chapter 10

Miscellaneous Problems

10.1 Optimization over an Aﬃne Family of Linear Systems

We consider a family of linear systems,

˙ x = Ax +B

w

w, z = C

z

(θ)x +D

zw

(θ)w, (10.1)

where C

z

and D

zw

depend aﬃnely on a parameter θ ∈ R

p

. We assume that A is

stable and (A, B

w

) is controllable. The transfer function,

H

θ

(s)

∆

= C

z

(θ)(sI −A)

−1

B

w

+D

zw

(θ),

depends aﬃnely on θ.

Several problems arising in system and control theory have the form

minimize ϕ

0

(H

θ

)

subject to ϕ

i

(H

θ

) < α

i

, i = 1, . . . , p

(10.2)

where ϕ

i

are various convex functionals. These problems can often be recast as LMI

problems. To do this, we will represent ϕ

i

(H

θ

) < α

i

as an LMI in θ, α

i

, and possibly

some auxiliary variables ζ

i

(for each i):

F

i

(θ, α

i

, ζ

i

) > 0.

The general optimization problem (10.2) can then be expressed as the EVP

minimize α

0

subject to F

i

(θ, α

i

, ζ

i

) > 0, i = 0, 1, . . . , p

10.1.1 H

2

norm

The H

2

norm of the system (10.1), i.e.,

H

θ

2

2

=

1

2π

Tr

_

∞

0

H

θ

(jω)

∗

H

θ

(jω) dω,

is ﬁnite if and only if D

zw

(θ) = 0. In this case, it equals Tr C

z

W

c

C

T

z

, where W

c

> 0 is

the controllability Gramian of the system (10.1) which satisﬁes (6.6). Therefore, the

H

2

norm of the system (10.1) is less than or equal to γ if and only if the following

conditions on θ and γ

2

are satisﬁed:

D

zw

(θ) = 0, Tr C

z

(θ)W

c

C

z

(θ)

T

≤ γ

2

.

The quadratic constraint on θ is readily cast as an LMI.

141

142 Chapter 10 Miscellaneous Problems

10.1.2 H

∞

norm

From the bounded-real lemma, we have H

θ

∞

< γ if and only if there exists P ≥ 0

such that

_

A

T

P +PA PB

w

B

T

w

P −γ

2

I

_

+

_

C

z

(θ)

T

D

zw

(θ)

T

_

_

C

z

(θ) D

zw

(θ)

_

≤ 0.

Remark: Here we have converted the so-called “semi-inﬁnite” convex constraint

|H

θ

(iω)| < γ for all ω ∈ R into a ﬁnite-dimensional convex (linear matrix)

inequality.

10.1.3 Entropy

The γ-entropy of the system (10.1) is deﬁned as

I

γ

(H

θ

)

∆

=

_

_

_

−γ

2

2π

_

∞

−∞

log det(I −γ

2

H

θ

(iω)H

θ

(iω)

∗

) dω, if H

θ

∞

< γ,

∞, otherwise.

When it is ﬁnite, I

γ

(H

θ

) is given by Tr B

T

w

PB

w

, where P is a symmetric matrix

with the smallest possible maximum singular value among all solutions of the Riccati

equation

A

T

P +PA+C

z

(θ)

T

C

z

(θ) +

1

γ

2

PB

w

B

T

w

P = 0.

For the system (10.1), the γ-entropy constraint I

γ

≤ λ is therefore equivalent to an

LMI in θ, P = P

T

, γ

2

, and λ:

D

zw

(θ) = 0,

_

¸

_

A

T

P +PA PB

w

C

z

(θ)

T

B

T

w

P −γ

2

I 0

C

z

(θ) 0 −I

_

¸

_ ≤ 0, Tr B

T

w

PB

w

≤ λ.

10.1.4 Dissipativity

Suppose that w and z are the same size. The dissipativity of H

θ

(see (6.59)) exceeds

γ if and only if the LMI in the variables P = P

T

and θ holds:

_

A

T

P +PA PB

w

−C

z

(θ)

T

B

T

w

P −C

z

(θ) 2γI −D

zw

(θ) −D

zw

(θ)

T

_

≤ 0.

We remind the reader that passivity corresponds to zero dissipativity.

10.1.5 Hankel norm

The Hankel norm of H

θ

is less than γ if and only if the following LMI in Q, θ and γ

2

holds:

D

zw

(θ) = 0, A

T

Q+QA+C

z

(θ)

T

C

z

(θ) ≤ 0,

γ

2

I −W

1/2

c

QW

1/2

c

≥ 0, Q ≥ 0.

(W

c

is the controllability Gramian deﬁned by (6.6).)

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

10.2 Analysis of Systems with LTI Perturbations 143

10.2 Analysis of Systems with LTI Perturbations

In this section, we address the problem of determining the stability of the system

˙ x = Ax +B

p

p, q = C

q

x +D

qp

p, p

i

= δ

i

∗ q

i

, i = 1, . . . , n

p

, (10.3)

where p = [p

1

· · · p

np

]

T

, q = [q

1

· · · q

np

]

T

, and δ

i

, i = 1, . . . , n

p

, are impulse

responses of single-output single-output LTI systems whose H

∞

norm is strictly less

than one, and ∗ denotes convolution. This problem has many variations: Some of the

δ

i

may be impulse response matrices, or they may satisfy equality constraints such as

δ

1

= δ

2

. Our analysis can be readily extended to these cases.

Let H

qp

(s) = C

q

(sI −A)

−1

B

p

+D

qp

. Suppose that there exists a diagonal transfer

matrix W(s) = diag(W

1

(s), . . . , W

np

(s)), where W

1

(s), . . . , W

np

(s) are single-input

single-output transfer functions with no poles or zeros on the imaginary axis such

that

sup

ω∈R

_

_

W(iω)H

qp

(iω)W(iω)

−1

_

_

< 1 (10.4)

Then from a small-gain argument, the system (10.3) is stable. Such a W is referred

to as a frequency-dependent scaling.

Condition (10.4) is equivalent to the existence of γ > 0 such that

H

qp

(iω)

∗

W(iω)

∗

W(iω)H

qp

(iω) −W(iω)

∗

W(iω) +γI ≤ 0,

for all ω ∈ R. From spectral factorization theory, it can be shown that such a W

exists if and only if there exists a stable, diagonal transfer matrix V and positive µ

such that

H

qp

(iω)

∗

(V (iω) +V (iω)

∗

)H

qp

(iω) −(V (iω) +V (iω)

∗

) +γI ≤ 0,

V (iω) +V (iω)

∗

≥ µI,

(10.5)

for all ω ∈ R.

In order to reduce (10.5) to an LMI condition, we restrict V to belong to an aﬃne,

ﬁnite-dimensional set Θ of diagonal transfer matrices, i.e., of the form

V

θ

(s) = D

V

(θ) +C

V

(θ)(sI −A)

−1

B,

where D

V

and C

V

are aﬃne functions are θ.

Then, H

qp

(−s)

T

V

θ

(s)H

qp

(s) − V

θ

(s) has a state-space realization (A

aug

, B

aug

,

C

aug

(θ), D

aug

(θ)) where C

aug

and D

aug

are aﬃne in θ, and (A

aug

, B

aug

) is controllable.

From §10.1.4, the condition that V

θ

(iω) +V

θ

(iω)

∗

≥ µI for all ω ∈ R is equivalent

to the existence of P

1

≥ 0, θ and γ > 0 such that the LMI

_

A

T

V

P

1

+P

1

A

V

P

1

B

V

−C

V

(θ)

T

(P

1

B

V

−C

V

(θ)

T

)

T

−(D

V

(θ) +D

V

(θ)

T

) +µI

_

≤ 0

holds.

The condition H

qp

(iω)

∗

(V

θ

(iω) +V

θ

(iω)

∗

)H

qp

(iω) −(V

θ

(iω) +V

θ

(iω)

∗

) +γI ≤ 0

is equivalent to the existence of P

2

= P

T

2

and θ satisfying the LMI

_

A

T

aug

P

2

+P

2

A

aug

P

2

B

aug

+C

aug

(θ)

T

(P

2

B

aug

+C

aug

(θ)

T

)

T

D

aug

(θ) +D

aug

(θ)

T

+γI

_

≤ 0. (10.6)

(See the Notes and References for details.) Thus proving stability of the system (10.3)

using a ﬁnite-dimensional set of frequency-dependent scalings is an LMIP.

This electronic version is for personal use and may not be duplicated or distributed.

144 Chapter 10 Miscellaneous Problems

10.3 Positive Orthant Stabilizability

The LTI system ˙ x = Ax is said to be positive orthant stable if x(0) ∈ R

n

+

implies that

x(t) ∈ R

n

+

for all t ≥ 0, and x(t) → 0 as t → ∞. It can be shown that the system

˙ x = Ax is positive orthant stable if and only if A

ij

≥ 0 for i = j, and there exists

a diagonal P > 0 such that PA

T

+ AP < 0. Therefore checking positive orthant

stability is an LMIP.

We next consider the problem of ﬁnding K such that the system ˙ x = (A+BK)x

is positive orthant stable. Equivalently, we seek a diagonal Q > 0 and K such that

Q(A + BK)

T

+ (A + BK)Q < 0, with the oﬀ-diagonal entries of A + BK being

nonnegative. Since Q > 0 is diagonal, the last condition holds if and only if all

the oﬀ-diagonal entries of (A+BK)Q are nonnegative. Therefore, with the change of

variables Y = KQ, proving positive orthant stabilizability for the system ˙ x = Ax+Bu

is equivalent to ﬁnding a solution to the LMIP with variables Q and Y :

Q > 0, (AQ+BY )

ij

≥ 0, i = j, QA

T

+Y

T

B

T

+AQ+BY < 0. (10.7)

Remark: The method can be extended to invariance of more general (polyhedral)

sets. Positive orthant stability and stabilizability of LDIs can be handled in a

similar way. See the Notes and References.

10.4 Linear Systems with Delays

Consider the system described by the delay-diﬀerential equation

d

dt

x(t) = Ax(t) +

L

i=1

A

i

x(t −τ

i

), (10.8)

where x(t) ∈ R

n

, and τ

i

> 0. If the functional

V (x, t) = x(t)

T

Px(t) +

L

i=1

_

τi

0

x(t −s)

T

P

i

x(t −s) ds, (10.9)

where P > 0, P

1

> 0, . . . , P

L

> 0, satisﬁes dV (x, t)/dt < 0 for every x satisfying (10.8),

then the system (10.8) is stable, i.e., x(t) →0 as t →∞.

It can be veriﬁed that dV (x, t)/dt = y(t)

T

Wy(t), where

W =

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

A

T

P +PA+

L

i=1

P

i

PA

1

· · · PA

L

A

T

1

P −P

1

· · · 0

.

.

.

.

.

.

.

.

.

.

.

.

A

T

L

P 0 · · · −P

L

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

, y(t) =

_

¸

¸

¸

¸

_

x(t)

x(t −τ

1

)

.

.

.

x(t −τ

L

)

_

¸

¸

¸

¸

_

.

Therefore, we can prove stability of system (10.8) using Lyapunov functionals of the

form (10.9) by solving the LMIP W < 0, P > 0, P

1

> 0, . . . , P

L

> 0.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

10.5 Interpolation Problems 145

Remark: The techniques for proving stability of norm-bound LDIs, discussed

in Chapter 5, can also be used on the system (10.8); it can be veriﬁed that the

LMI (5.15) that we get for quadratic stability of a DNLDI, in the case when the

state x is a scalar, is the same as the LMI W < 0 above. We also note that we

can regard the delay elements of system (10.8) as convolution operators with unit

L2 gain, and use the techniques of ¸10.2 to check its stability.

Next, we add an input u to the system (10.8) and consider

d

dt

x(t) = Ax(t) +Bu(t) +

L

i=1

A

i

x(t −τ

i

). (10.10)

We seek a state-feedback u(t) = Kx(t) such that the system (10.10) is stable.

From the discussion above, there exists a state-feedback gain K such that a Lya-

punov functional of the form (10.9) proves the stability of the system (10.10), if

W =

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

(A+BK)

T

P +P(A+BK) +

L

i=1

P

i

PA

1

· · · PA

L

A

T

1

P −P

1

· · · 0

.

.

.

.

.

.

.

.

.

.

.

.

A

T

L

P 0 · · · −P

L

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

< 0

for some P, P

1

, . . . , P

L

> 0.

Multiplying every block entry of W on the left and on the right by P

−1

and setting

Q = P

−1

, Q

i

= P

−1

P

i

P

−1

and Y = KP

−1

, we obtain the condition

X =

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

AQ+QA

T

+BY +Y

T

B

T

+

L

i=1

Q

i

A

1

Q · · · A

L

Q

QA

T

1

−Q

1

· · · 0

.

.

.

.

.

.

.

.

.

.

.

.

QA

T

L

0 · · · −Q

L

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

< 0.

Thus, checking stabilizability of the system (10.8) using Lyapunov functionals of the

form (10.9) is an LMIP in the variables Q, Y , Q

1

, . . . , Q

L

.

Remark: It is possible to use the elimination procedure of ¸2.6.2 to eliminate

the matrix variable Y and obtain an equivalent LMIP with fewer variables.

10.5 Interpolation Problems

10.5.1 Tangential Nevanlinna-Pick problem

Given λ

1

, . . . , λ

m

with λ

i

∈ C

+

∆

= {s | Re s > 0} and distinct, u

1

, . . . , u

m

, with

u

i

∈ C

q

and v

1

, . . . , v

m

, with v

i

∈ C

p

, i = 1, . . . , m, the tangential Nevanlinna-Pick

This electronic version is for personal use and may not be duplicated or distributed.

146 Chapter 10 Miscellaneous Problems

problem is to ﬁnd, if possible, a function H : C −→ C

p×q

which is analytic in C

+

,

and satisﬁes

H(λ

i

)u

i

= v

i

, i = 1, . . . , m, with H

∞

≤ 1. (10.11)

Problem (10.11) arises in multi-input multi-output H

∞

control theory.

From Nevanlinna-Pick theory, problem (10.11) has a solution and only if the Pick

matrix N deﬁned by

N

ij

=

u

∗

i

u

j

−v

∗

i

v

j

λ

∗

i

+λ

j

)

is positive semideﬁnite. N can also be obtained as the solution of the Lyapunov

equation

A

∗

N +NA−(U

∗

U −V

∗

V ) = 0.

where A = diag(λ

1

, . . . , λ

m

), U = [u

1

· · · u

m

], V = [v

1

· · · v

m

]. For future reference,

we note that N = G

in

−G

out

, where

A

∗

G

in

+G

in

A−U

∗

U = 0, A

∗

G

out

+G

out

A−V

∗

V = 0.

Solving the tangential Nevanlinna-Pick problem simply requires checking whether N ≥

0.

10.5.2 Nevanlinna-Pick interpolation with scaling

We now consider a simple variation of problem (10.11): Given λ

1

, . . . , λ

m

with λ

i

∈

C

+

, u

1

, . . . , u

m

, with u

i

∈ C

p

and v

1

, . . . , v

m

, with v

i

∈ C

p

, i = 1, . . . , m, the problem

is to ﬁnd

γ

opt

= inf

_

DHD

−1

∞

¸

¸

¸

¸

¸

H is analytic in C

+

, D = D

∗

> 0

D ∈ D, H(λ

i

)u

i

= v

i

, i = 1, . . . , m

_

, (10.12)

where D is the set of m×m block-diagonal matrices with some speciﬁed block struc-

ture. Problem (10.12) corresponds to ﬁnding the smallest scaled H

∞

norm of all

interpolants. This problem arises in multi-input multi-output H

∞

control synthesis

for systems with structured perturbations.

With a change of variables P = D

∗

D and the discussion in the previous section,

it follows that γ

opt

is the smallest positive γ such that there exists P > 0, P ∈ D such

that the following equations and inequality hold:

A

∗

G

in

+G

in

A−U

∗

PU = 0,

A

∗

G

out

+G

out

A−V

∗

PV = 0,

γ

2

G

in

−G

out

≥ 0.

This is a GEVP.

10.5.3 Frequency response identiﬁcation

We consider the problem of identifying the transfer function H of a linear system

from noisy measurements of its frequency response at a set of frequencies. We seek H

satisfying two constraints:

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

10.6 The Inverse Problem of Optimal Control 147

• Consistency with measurements. For some n

i

with |n

i

| ≤ , we have f

i

=

H(jω

i

)+n

i

. Here, f

i

is the measurement of the frequency response at frequency

ω

i

, n

i

is the (unknown) measurement error, and is the measurement precision.

• Prior assumption. α-shifted H

∞

norm of H does not exceed M.

From Nevanlinna-Pick theory, there exist H and n with H(ω

i

) = f

i

+ n

i

satisfying

these conditions if and only if there exist n

i

with |n

i

| ≤ , G

in

> 0 and G

out

> 0 such

that

M

2

G

in

−G

out

≥ 0,

(A+αI)

∗

G

in

+G

in

(A+αI) −e

∗

e = 0,

(A+αI)

∗

G

out

+G

out

(A+αI) −(f +n)

∗

(f +n) = 0,

where A = diag(jω

1

, . . . , jω

m

), e = [1 · · · 1], f = [f

1

· · · f

m

], and n =

[n

1

· · · n

m

]. It can be shown that these conditions are equivalent to

M

2

G

in

−G

out

≥ 0,

(A+αI)

∗

G

in

+G

in

(A+αI) −e

∗

e = 0,

(A+αI)

∗

G

out

+G

out

(A+αI) −(f +n)

∗

(f +n) ≥ 0,

with |n

i

| ≤ .

With this observation, we can answer a number of interesting questions in fre-

quency response identiﬁcation by solving EVPs and GEVPs.

• For ﬁxed α and , minimize M. Solving this GEVP answers the question “Given

α and a bound on the noise values, what is the smallest possible α-shifted

H

∞

norm of the system consistent with the measurements of the frequency

response?”

• For ﬁxed α and M, minimize . Solving this EVP answers the question “Given

α and a bound on α-shifted H

∞

norm of the system, what is the “smallest”

possible noise such that the measurements are consistent with the given values

of α and M.

10.6 The Inverse Problem of Optimal Control

Given a system

˙ x = Ax +Bu, x(0) = x

0

, z =

_

Q

1/2

0

0 R

1/2

__

x

u

_

,

with (A, B) is stabilizable, (Q, A) is detectable and R > 0, the LQR problem is to

determine u that minimizes

_

∞

0

z

T

z dt.

The solution of this problem can be expressed as a state-feedback u = Kx with

K = −R

−1

B

T

P, where P is the unique nonnegative solution of

A

T

P +PA−PBR

−1

B

T

P +Q = 0.

The inverse problem of optimal control is the following. Given a matrix K, determine

if there exist Q ≥ 0 and R > 0, such that (Q, A) is detectable and u = Kx is the

This electronic version is for personal use and may not be duplicated or distributed.

148 Chapter 10 Miscellaneous Problems

optimal control for the corresponding LQR problem. Equivalently, we seek R > 0 and

Q ≥ 0 such that there exists P nonnegative and P

1

positive-deﬁnite satisfying

(A+BK)

T

P +P(A+BK) +K

T

RK +Q = 0, B

T

P +RK = 0

and A

T

P

1

+P

1

A < Q. This is an LMIP in P, P

1

, R and Q. (The condition involving

P

1

is equivalent to (Q, A) being detectable.)

10.7 System Realization Problems

Consider the discrete-time minimal LTI system

x(k + 1) = Ax(k) +Bu(k), y(k) = Cx(k), (10.13)

where x : Z

+

→R

n

, u : Z

+

→R

nu

y : Z

+

→R

ny

. The system realization problem is

to ﬁnd a change of state coordinates x = T ¯ x, such that the new realization

¯ x(k + 1) = T

−1

AT ¯ x(k) +T

−1

Bu(k) y(k) = CT ¯ x(k), (10.14)

satisﬁes two competing constraints: First, the input-to-state transfer matrix, T

−1

(zI−

A)

−1

B should be “small”, in order to avoid state overﬂow in numerical implementa-

tion; and second, the state-to-output transfer matrix, C(zI−A)

−1

T should be “small”,

in order to minimize eﬀects of state quantization at the output.

Suppose we assume that the RMS value of the input is bounded by α and require

the RMS value of the state to be less than one. This yields the bound

_

_

T

−1

(zI −A)

−1

B

_

_

∞

≤ 1/α, (10.15)

where the H

∞

norm of the transfer matrix H(z) is deﬁned as

H

∞

= sup{H(z) | |z| > 1} .

Next, suppose that the state quantization noise is modeled as a white noise sequence

w(k) with Ew(k)

T

w(k) = η, injected directly into the state, and its eﬀect on the

output is measured by its RMS value, which is just η times the H

2

norm of the

state-to-output transfer matrix:

p

noise

= η

_

_

C(zI −A)

−1

T

_

_

2

. (10.16)

Our problem is then to compute T to minimize the noise power (10.16) subject to the

overﬂow avoidance constraint (10.15).

The constraint (10.15) is equivalent to the existence of P > 0 such that

_

A

T

PA−P +T

−T

T

−1

A

T

PB

B

T

PA B

T

PB −I/α

2

_

≤ 0.

The output noise power can be expressed as

p

2

noise

= η

2

Tr T

T

W

o

T,

where W

o

is the observability Gramian of the original system (A, B, C), given by

W

o

=

∞

k=0

(A

T

)

k

C

T

CA

k

.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

10.7 System Realization Problems 149

W

o

is the solution of the Lyapunov equation

A

T

W

o

A−W

o

+C

T

C = 0.

With X = T

−T

T

−1

, the realization problem is: Minimize η

2

Tr W

o

X

−1

subject

to X > 0 and the LMI (in P > 0 and X)

_

A

T

PA−P +X A

T

PB

B

T

PA B

T

PB −I/α

2

_

≤ 0. (10.17)

This is a convex problem in X and P, and can be transformed into the EVP

minimize η

2

Tr Y W

c

subject to (10.17), P > 0,

_

Y I

I X

_

≥ 0

Similar methods can be used to handle several variations and extensions.

Input u has RMS value bounded componentwise

Suppose the RMS value of each component u

i

is less than α (instead of the RMS value

of u being less than α) and that the RMS value of the state is still required to be less

than one. Then, using the methods of §10.9, an equivalent condition is the LMI (with

variables X, P and R):

_

A

T

PA−P +X A

T

PB

B

T

PA B

T

PB −R/α

2

_

≤ 0, P > 0, (10.18)

where R > 0 is diagonal, with unit trace, so that the EVP (over X, Y , P and R) is

minimize η

2

Tr Y W

c

subject to (10.18), P > 0, X > 0

R > 0, R is diagonal, Tr R = 1

_

Y I

I X

_

≥ 0

Minimizing state-to-output H

∞

norm

Alternatively, we can measure the eﬀect of the quantization noise on the output with

the H

∞

norm, that is, we can choose T to minimize C(zI −A)

−1

T

∞

subject to the

constraint (10.15). The constraint C(zI − A)

−1

T

∞

≤ γ is equivalent to the LMI

(over Y > 0 and X = T

−T

T

−1

> 0):

_

A

T

Y A−Y +C

T

C A

T

Y

Y A Y −X/α

2

_

≤ 0. (10.19)

Therefore choosing T to minimize C(zI −A)

−1

T

∞

subject to (10.15) is the GEVP

minimize γ

subject to P > 0, Y > 0, X > 0, (10.17), (10.19)

An optimal state coordinate transformation T is any matrix that satisﬁes X

opt

=

(TT

T

)

−1

, where X

opt

is an optimal value of X in the GEVP.

This electronic version is for personal use and may not be duplicated or distributed.

150 Chapter 10 Miscellaneous Problems

10.8 Multi-Criterion LQG

Consider the LTI system given by

˙ x = Ax +Bu +w, y = Cx +v, z =

_

Q

1/2

0

0 R

1/2

__

x

u

_

,

where u is the control input; y is the measured output; z is the exogenous output;

w and v are independent white noise signals with constant, positive-deﬁnite spectral

density matrices W and V respectively; we further assume that Q ≥ 0 and R > 0,

that (A, B) is controllable, and that (C, A) and (Q, A) are observable.

The standard Linear-Quadratic Gaussian (LQG) problem is to minimize

J

lqg

= lim

t→∞

Ez(t)

T

z(t)

over u, subject to the condition that u(t) is measurable on y(τ) for τ ≤ t. The optimal

cost is given by

J

∗

lqg

= Tr (X

lqg

U +QY

lqg

) ,

where X

lqg

and Y

lqg

are the unique positive-deﬁnite solutions of the Riccati equations

A

T

X

lqg

+X

lqg

A−X

lqg

BR

−1

B

T

X

lqg

+Q = 0,

AY

lqg

+Y

lqg

A

T

−Y

lqg

C

T

V

−1

CY

lqg

+W = 0,

(10.20)

and U = Y

lqg

C

T

V

−1

CY

lqg

.

In the multi-criterion LQG problem, we have several exogenous outputs of interest

given by

z

i

=

_

Q

1/2

i

0

0 R

1/2

i

__

x

u

_

, Q

i

≥ 0, R

i

> 0, i = 0, . . . , p.

We assume that (Q

0

, A) is observable. For each z

i

, we associate a cost function

J

i

lqg

= lim

t→∞

Ez

i

(t)

T

z

i

(t), i = 0, . . . , p. (10.21)

The multi-criterion LQG problem is to minimize J

0

lqg

over u subject to the measurabil-

ity condition and the constraints J

i

lqg

< γ

i

, i = 1, . . . , p. This is a convex optimization

problem, whose solution is given by maximizing

Tr (X

lqg

U +QY

lqg

) −

p

i=1

γ

i

τ

i

,

over nonnegative τ

1

, . . . , τ

p

, where X

lqg

and Y

lqg

are the solutions of (10.20) with

Q = Q

0

+

p

i=1

τ

i

Q

i

and R = R

0

+

p

i=1

τ

i

R

i

.

Noting that X

lqg

≥ X for every X > 0 that satisﬁes

A

T

X +XA−XBR

−1

B

T

X +Q ≥ 0,

we conclude that the optimal cost is the maximum of

Tr

_

XU + (Q

0

+

p

i=1

τ

i

Q

i

)Y

lqg

_

−

p

i=1

γ

i

τ

i

,

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

10.9 Nonconvex Multi-Criterion Quadratic Problems 151

over X, τ

1

, . . . , τ

p

, subject to X > 0, τ

1

≥ 0, . . . , τ

p

≥ 0 and

A

T

X +XA−XB

_

R

0

+

p

i=1

τ

i

R

i

_

−1

B

T

X +Q

0

+

p

i=1

τ

i

Q

i

≥ 0.

Computing this is an EVP.

10.9 Nonconvex Multi-Criterion Quadratic Problems

In this section, we consider the LTI system ˙ x = Ax +Bu, x(0) = x

0

, where (A, B) is

controllable. For any u, we deﬁne a set of p + 1 cost indices J

0

, . . . , J

p

by

J

i

(u) =

_

∞

0

[x

T

u

T

]

_

Q

i

C

i

C

T

i

R

i

__

x

u

_

dt, i = 0, . . . , p.

Here the symmetric matrices

_

Q

i

C

i

C

T

i

R

i

_

, i = 0, . . . , p,

are not necessarily positive-deﬁnite. The constrained optimal control problem is:

minimize J

0

,

subject to J

i

≤ γ

i

, i = 1, . . . , p, x(t) →0 as t →∞

(10.22)

The solution to this problem proceeds as follows: We ﬁrst deﬁne

Q = Q

0

+

p

i=1

τ

i

Q

i

, R = R

0

+

p

i=1

τ

i

R

i

, C = C

0

+

p

i=1

τ

i

C

i

,

where τ

1

≥ 0, . . . , τ

p

≥ 0. Next, with τ = [τ

1

· · · τ

p

], we deﬁne

S(τ, u)

∆

= J

0

+

p

i=1

τ

i

J

i

−

p

i=1

τ

i

γ

i

.

Then, the solution of problem (10.22), when it is not −∞, is given by:

sup

τ

inf

u

{S(τ, u) | x(t) →0 as t →∞}

(See the Notes and References.)

For any ﬁxed τ, the inﬁmum over u of S(τ, u), when it is not −∞, is computed

by solving the EVP in P = P

T

maximize x(0)

T

Px(0) −

p

i=1

τ

i

γ

i

subject to

_

A

T

P +PA+Q PB +C

T

B

T

P +C R

_

≥ 0

(10.23)

Therefore, the optimal control problem is solved by the EVP in P = P

T

and τ:

maximize x(0)

T

Px(0) −

p

i=1

τ

i

γ

i

subject to

_

A

T

P +PA+Q PB +C

T

B

T

P +C R

_

≥ 0, τ

1

≥ 0, . . . , τ

p

≥ 0.

This electronic version is for personal use and may not be duplicated or distributed.

152 Chapter 10 Miscellaneous Problems

Notes and References

Optimization over an aﬃne family of linear systems

Optimization over an aﬃne family of transfer matrices arises in linear controller design via

Youla’s parametrization of closed-loop transfer matrices; see Boyd and Barratt [BB91] and

the references therein. Other examples include ﬁnite-impulse-response (FIR) ﬁlter design (see

for example [OKUI88]) and antenna array weight design [KRB89].

In [Kav94] Kavrano˘ glu considers the problem of approximating in the H∞ norm a given

transfer matrix by a constant matrix, which is a problem of the form considered in ¸10.1. He

casts the problem as an EVP.

Stability of systems with LTI perturbations

For references about this problem, we refer the reader to the Notes and References in Chap-

ter 3. LMI (10.6) is derived from Theorems 3 and 4 in Willems [Wil71b].

Positive orthant stabilizability

The problem of positive orthant stability and holdability was extensively studied by Berman,

Neumann and Stern [BNS89, ¸7.4]. They solve the positive orthant holdability problem

when the control input is scalar; our results extend theirs to the multi-input case. The study

of positive orthant holdability draws heavily from the theory of positive matrices, references

for which are [BNS89] and the book by Berman and Plemmons [BP79, ch6].

Diagonal solutions to the Lyapunov equation play a central role in the positive orthant

stabilizability and holdability problems. A related paper is by Geromel [Ger85], who gives

a computational procedure to ﬁnd diagonal solutions of the Lyapunov equation. Diagonal

quadratic Lyapunov functions also arise in the study of large-scale systems [MH78]; see

also ¸3.4. A survey of results and applications of diagonal Lyapunov stability is given in

Kaszkurewicz and Bhaya [KB93].

A generalization of the concept of positive orthant stability is that of invariance of polyhedral

sets. The problem of checking whether a given polyhedron is invariant as well as the associated

state-feedback synthesis problem have been considered by many authors, see e.g., [Bit88,

BBT90, HB91b] and references therein. The LMI method of ¸10.3 can be extended to the

polyhedral case, using the approach of [CH93]. Finally, we note that the results of ¸10.3 are

easily generalized to the LDIs considered in Chapter 4.

Stabilization of systems with time delays

The introduction of Lyapunov functionals of the form (10.9) is due to Krasovskii [Kra56].

Skorodinskii [Sko90] observed that the problem of proving stability of a system with delays

via Lyapunov–Krasovskii functionals is a convex problem. See also [PF92, WBL92, SL93b,

FBB92].

It is interesting to note that the “structured stability problem” or “µ-analysis problem”

can be cast in terms of linear systems with delays. Consider a stable LTI system with

transfer matrix H and L inputs and outputs, connected in feedback with a diagonal transfer

matrix ∆ with H∞ norm less than one. This system is stable for all such ∆ if and only if

sup

ω∈R

µ(H(jω)) < 1, where µ denotes the structured singular value (see [Doy82]). It can

be shown (see [Boy86]) that if the feedback system is stable for all ∆ which are diagonal,

with ∆ii(s) = e

−sT

i

, Ti > 0 (i.e., delays) then the system is stable for all ∆ with |∆|∞ ≤ 1.

Thus, verifying sup

ω∈R

µ(H(jω)) < 1 can be cast as checking stability of a linear system

with L arbitrary, positive delays. In particular we see that Krasovskii’s method from 1956

can be interpreted as a Lyapunov-based µ-analysis method.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 153

Interpolation problems

The classical Nevanlinna-Pick problem dates back at least to 1916 [Pic16, Nev19]. There

are two matrix versions of this problem: the matrix Nevanlinna-Pick problem [DGK79]

and the tangential Nevanlinna-Pick problem [Fed75]. A comprehensive description of the

Nevanlinna-Pick problem, its extensions and variations can be found in the book by Ball,

Gohberg and Rodman [BGR90].

A number of researchers have studied the application of Nevanlinna-Pick theory to problems

in system and control: Delsarte, Genin and Kamp discuss the role of the matrix Nevanlinna-

Pick problem in circuit and systems theory [DGK81]. Zames and Francis, in [ZF83], study

the implications of interpolation conditions in controller design; also see [OF85]. Chang

and Pearson [CP84] solve a class of H∞-optimal controller design problems using matrix

Nevanlinna-Pick theory. Safonov [Saf86] reduces the design of controllers for systems with

structured perturbations to the scaled matrix Nevanlinna-Pick problem. In [Kim87], Kimura

reduces a class of H∞-optimal controller design problems to the tangential Nevanlinna-Pick

problem; see also [Bal94].

System identiﬁcation

The system identiﬁcation problem considered in ¸10.5.3 can be found in [CNF92]. Other

system identiﬁcation problems can be cast in terms of LMI problems; see e.g., [PKT

+

94].

The inverse problem of optimal control

This problem was ﬁrst considered by Kalman [Kal64], who solved it when the control u is

scalar; see also [And66c]. It is discussed in great detail by Anderson and Moore [AM90,

¸5.6], Fujii and Narazaki [FN84] and the references therein; they solve the problem when

the control weighting matrix R is known, by checking that the return diﬀerence inequality

_

I −B

T

(−iωI −A

T

)

−1

K

_

R

_

I −K

T

(iωI −A)

−1

B

_

≥ R

holds for all ω ∈ R. Our result handles the case of unknown R, i.e., the most general form

of the inverse optimal control problem.

System realization problems

For a discussion of the problem of system realizations see [MR76, Thi84], which give an-

alytical solutions via balanced realizations for special realization problems. The book by

Gevers and Li [GL93a] describes a number of digital ﬁlter realization problems. Liu, Skelton

and Grigoriadis discuss the problem of optimal ﬁnite word-length controller implementa-

tion [LSG92]. See also Rotea and Williamson [RW94b, RW94a].

We conjecture that the LMI approach to system realization can also incorporate the constraint

of stability, i.e., that the (nonlinear) system not exhibit limit cycles due to quantization or

overﬂow; see, e.g., [Wil74b].

We also note that an H∞ norm constraint on the input-to-state transfer matrix yields,

indirectly, a bound on the peak gain from input to state; Boyd and Doyle [BD87]. This can

be used to guarantee no overﬂow given a maximum peak input level.

Multi-criterion LQG

The multi-criterion LQG problem is discussed in [BB91]; see also the references cited there.

Rotea [Rot93] calls this problem “generalized H2 control”. He shows that one can restrict

attention to observer-based controllers such as

d

dt

ˆ x = (A+BK)ˆ x +Y C

T

V

−1

(y −Cˆ x), u = Kˆ x,

This electronic version is for personal use and may not be duplicated or distributed.

154 Chapter 10 Miscellaneous Problems

where Y is given by (10.20). The covariance of the residual state ˆ x−x is Y and the covariance

P of the state ˆ x is given by the solution of the Lyapunov equation:

(A+BK)P +P(A+BK)

T

+Y C

T

V

−1

CY = 0. (10.24)

The value of each cost index J

i

lqg

deﬁned by (10.21) is then

J

i

lqg

= Tr RiKPK

T

+Tr Qi (Y +P) .

Now, taking Z = KP as a new variable, we replace (10.24) by the linear equality in W and

P

AP +PA

T

+BZ +Z

T

B

T

+Y C

T

V

−1

CY = 0

and each cost index becomes

J

i

lqg

= Tr RiZP

−1

Z

T

+Tr Qi (Y +P) .

The multi-criterion problem is then solved by the EVP in P, Z and X

minimize Tr R0X +Tr Q0(P +Y )

subject to

_

X Z

Z

T

P

_

≥ 0,

AP +PA

T

+BZ +Z

T

B

T

+Y C

T

V

−1

CY = 0,

Tr RiX +Tr Qi(P +Y ) ≤ γi, i = 1, . . . , L

This formulation of the multi-criterion LQG problem is the dual of our formulation; the two

formulations can be used together in an eﬃcient primal-dual method (see, e.g., [VB93b]). We

also mention an algorithm devised by Zhu, Rotea and Skelton [ZRS93] to solve a similar class

of problems. For a variation on the multi-criterion LQG problem discussed here, see [Toi84];

see also [TM89, Mak91].

Nonconvex multi-criterion quadratic problems

This section is based on Megretsky [Meg92a, Meg92b, Meg93] and Yakubovich [Yak92].

In our discussion we omitted important technical details, such as constraint regularity, which

is covered in these articles. Also, Yakubovich gives conditions under which the inﬁmum

is actually a minimum (that is, there exists an optimal control law that achieves the best

performance). The EVP (10.23) is derived from Theorem 3 in Willems [Wil71b].

Many of the suﬃcient conditions for NLDIs in Chapters 5–7 can be shown to be necessary

and suﬃcient conditions when we consider integral constraints on p and q, by analyzing them

as nonconvex multi-criterion quadratic problems. This is illustrated in the following section.

Mixed H

2

–H

∞

problem

We consider the analog of an NLDI with integral quadratic constraint:

˙ x = Ax +Bpp +Buu, q = Cqx, z =

_

Q

1/2

0

0 R

1/2

__

x

u

_

_

∞

0

p

T

p dt ≤

_

∞

0

q

T

q dt

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Notes and References 155

Deﬁne the cost function J =

_

∞

0

z

T

z dt. Our goal is to determine minK maxp J, over all

possible linear state-feedback strategies u = Kx. (Note that this is the same as the prob-

lem considered in ¸7.4.1 with integral constraints on p and q instead of pointwise in time

constraints.)

For ﬁxed K, ﬁnding the maximum of J over admissible p’s is done using the results of ¸10.9.

This maximum is obtained by solving the EVP

minimize x(0)

T

Px(0)

subject to

_

¸

_

(A+BuK)

T

P +P(A+BuK)+

Q+K

T

RK +τC

T

q

Cq

PBp

B

T

p

P −τI

_

¸

_

≤ 0, τ ≥ 0

The minimization over K is now simply done by introducing the new variables W = P

−1

,

Y = KW, µ = 1/τ. Indeed, in this particular case, we know we can add the constraint

P > 0. The corresponding EVP is then

minimize x(0)

T

W

−1

x(0)

subject to

_

¸

_

_

WA

T

+Y B

T

u

+AW +BuY +

WQW +Y

T

RY +WC

T

q

CqW/µ

_

µBp

µB

T

p

−µI

_

¸

_

≤ 0, µ ≥ 0

which is the same as the EVP (7.23).

This electronic version is for personal use and may not be duplicated or distributed.

Notation

R, R

k

, R

m×n

The real numbers, real k-vectors, real m×n matrices.

R

+

The nonnegative real numbers.

C The complex numbers.

Re(a) The real part of a ∈ C, i.e., (a +a

∗

)/2.

S The closure of a set S ⊆ R

n

.

I

k

The k ×k identity matrix. The subscript is omitted when k is not

relevant or can be determined from context.

M

T

Transpose of a matrix M: (M

T

)

ij

= M

ji

.

M

∗

Complex-conjugate transpose of a matrix M: (M

∗

)

ij

= M

∗

ji

, where

α

∗

denotes the complex-conjugate of α ∈ C.

Tr M Trace of M ∈ R

n×n

, i.e.,

n

i=1

M

ii

.

M ≥ 0 M is symmetric and positive semideﬁnite, i.e., M = M

T

and

z

T

Mz ≥ 0 for all z ∈ R

n

.

M > 0 M is symmetric and positive-deﬁnite, i.e., z

T

Mz > 0 for all nonzero

z ∈ R

n

.

M > N M and N are symmetric and M −N > 0.

M

1/2

For M > 0, M

1/2

is the unique Z = Z

T

such that Z > 0, Z

2

= M.

λ

max

(M) The maximum eigenvalue of the matrix M = M

T

.

λ

max

(P, Q) For P = P

T

, Q = Q

T

> 0, λ

max

(P, Q) denotes the maximum eigen-

value of the symmetric pencil (P, Q), i.e., λ

max

(Q

−1/2

PQ

−1/2

).

λ

min

(M) The minimum eigenvalue of M = M

T

.

M The spectral norm of a matrix or vector M, i.e.,

_

λ

max

(M

T

M).

Reduces to the Euclidean norm, i.e., x =

√

x

T

x, for a vector x.

diag(· · ·) Block-diagonal matrix formed from the arguments, i.e.,

diag (M

1

, . . . , M

m

)

∆

=

_

¸

¸

_

M

1

.

.

.

M

m

_

¸

¸

_

.

157

158 Notation

CoS Convex hull of the set S ⊆ R

n

, given by

CoS

∆

=

_

p

i=1

λ

i

x

i

¸

¸

¸

¸

¸

x

i

∈ S, p ≥ 0

_

.

(Without loss of generality, we can take p = n + 1 here.)

Ex Expected value of (the random variable) x.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Acronyms

Acronym Meaning Page

ARE Algebraic Riccati Equation 3

CP Convex Minimization Problem 11

DI Diﬀerential Inclusion 51

DNLDI Diagonal Norm-bound Linear Diﬀerential Inclusion 54

EVP Eigenvalue Minimization Problem 10

GEVP Generalized Eigenvalue Minimization Problem 10

LDI Linear Diﬀerential Inclusion 51

LMI Linear Matrix Inequality 7

LMIP Linear Matrix Inequality Problem 9

LQG Linear-Quadratic Gaussian 150

LQR Linear-Quadratic Regulator 114

LTI Linear Time-invariant 52

NLDI Norm-bound Linear Diﬀerential Inclusion 53

PLDI Polytopic Linear Diﬀerential Inclusion 53

PR Positive-Real 25

RMS Root-Mean-Square 91

159

Bibliography

[ABG93] D. Arzelier, J. Bernussou, and G. Garcia. Pole assignment of linear uncertain

systems in a sector via a Lyapunov-type approach. IEEE Trans. Aut. Control,

AC-38(7):1128–1131, July 1993.

[ABJ75] B. D. Anderson, N. K. Bose, and E. I. Jury. Output feedback stabilization and

related problems—Solution via decision methods. IEEE Trans. Aut. Control,

AC-20:53–66, 1975.

[AC84] J. P. Aubin and A. Cellina. Diﬀerential Inclusions, volume 264 of Comprehensive

Studies in Mathematics. Springer-Verlag, 1984.

[AF90] J. P. Aubin and H. Frankowska. Set-valued analysis, volume 2 of Systems &

Control. Birkhauser, Boston, 1990.

[AG64] M. A. Aizerman and F. R. Gantmacher. Absolute stability of regulator systems.

Information Systems. Holden-Day, San Francisco, 1964.

[AGB94] P. Apkarian, P. Gahinet, and G. Becker. Self-scheduled H∞ control of linear

parameter-varying systems. Submitted., 1994.

[AGG94] F. Amato, F. Garofalo, and L. Glielmo. Polytopic coverings and robust stability

analysis via Lyapunov quadratic forms. In A. Zinober, editor, Variable structure

and Lyapunov control, volume 193 of Lecture notes in control and information

sciences, chapter 13, pages 269–288. Springer-Verlag, 1994.

[Akg84] M. Akg¨ ul. Topics in Relaxation and Ellipsoidal Methods, volume 97 of Research

Notes in Mathematics. Pitman, 1984.

[AL84] W. F. Arnold and A. J. Laub. Generalized eigenproblem algorithms and software

for algebraic Riccati equations. Proc. IEEE, 72(12):1746–1754, December 1984.

[Ali91] F. Alizadeh. Combinatorial optimization with interior point methods and semi-

deﬁnite matrices. PhD thesis, Univ. of Minnesota, October 1991.

[Ali92a] F. Alizadeh. Combinatorial optimization with semideﬁnite matrices. In Pro-

ceedings of second annual Integer Programming and Combinatorial Optimization

conference. Carnegie-Mellon University, 1992.

[Ali92b] F. Alizadeh. Optimization over the positive-deﬁnite cone: interior point meth-

ods and combinatorial applications. In Panos Pardalos, editor, Advances in

Optimization and Parallel Computing. North-Holland, 1992.

[All89] J. Allwright. On maximizing the minimum eigenvalue of a linear combination of

symmetric matrices. SIAM J. on Matrix Analysis and Applications, 10:347–382,

1989.

[AM74] B. Anderson and P. J. Moylan. Spectral factorization of a ﬁnite-dimensional

nonstationary matrix covariance. IEEE Trans. Aut. Control, AC-19(6):680–692,

1974.

[AM90] B. Anderson and J. B. Moore. Optimal Control: Linear Quadratic Methods.

Prentice-Hall, 1990.

[And66a] B. Anderson. Algebraic description of bounded real matrices. Electron. Lett.,

2(12):464–465, December 1966.

161

162 Bibliography

[And66b] B. Anderson. Stability of control systems with multiple nonlinearities. J.

Franklin Inst., 282:155–160, 1966.

[And66c] B. Anderson. The testing for optimality of linear systems. Int. J. Control,

4(1):29–40, 1966.

[And67] B. Anderson. A system theory criterion for positive real matrices. SIAM J.

Control, 5:171–182, 1967.

[And73] B. Anderson. Algebraic properties of minimal degree spectral factors. Automat-

ica, 9:491–500, 1973. Correction in 11:321–322, 1975.

[Arn74] L. Arnold. Stochastic diﬀerential equations: theory and applications. Wiley, New

York, 1974.

[AS79] M. Araki and M. Saeki. Multivariable circle criteria and M-matrix condition. In

Proc. Joint American Control Conf., pages 11–17, 1979.

[˚Ast70] K. J.

˚

Astr¨ om. Introduction to stochastic control theory, volume 70 of Math. in

Science and Engr. Academic Press, 1970.

[AV73] B. Anderson and S. Vongpanitlerd. Network analysis and synthesis: a modern

systems theory approach. Prentice-Hall, 1973.

[Bad82] F. A. Badawi. On the quadratic matrix inequality and the corresponding alge-

braic Riccati equation. Int. J. Control, 36(2):313–322, August 1982.

[Bal94] V. Balakrishnan. H∞ controller synthesis with time-domain constraints using

interpolation theory and convex optimization. IEEE Trans. Aut. Control, 1994.

Submitted.

[Bar70a] E. A. Barbashin. Introduction to the theory of stability. Walters-Noordhoﬀ, 1970.

[Bar70b] A. I. Barkin. Suﬃcient conditions for the absence of auto-oscillations in pulse

systems. Aut. Remote Control, 31:942–946, 1970.

[Bar83] B. R. Barmish. Stabilization of uncertain systems via linear control. IEEE

Trans. Aut. Control, AC-28(8):848–850, August 1983.

[Bar85] B. R. Barmish. Necessary and suﬃcient conditions for quadratic stabilizabil-

ity of an uncertain system. Journal of Optimization Theory and Applications,

46(4):399–408, August 1985.

[Bar93] B. R. Barmish. New Tools for Robustness of Linear Systems. MacMillan, 1993.

[Bau63] F. L. Bauer. Optimally scaled matrices. Numerische Mathematik, 5:73–87, 1963.

[BB91] S. Boyd and C. Barratt. Linear Controller Design: Limits of Performance.

Prentice-Hall, 1991.

[BB92] V. Balakrishnan and S. Boyd. Global optimization in control system analysis

and design. In C. T. Leondes, editor, Control and Dynamic Systems: Advances

in Theory and Applications, volume 53. Academic Press, New York, New York,

1992.

[BBB91] V. Balakrishnan, S. Boyd, and S. Balemi. Branch and bound algorithm for

computing the minimum stability degree of parameter-dependent linear systems.

Int. J. of Robust and Nonlinear Control, 1(4):295–317, October–December 1991.

[BBC90] G. Belforte, B. Bona, and V. Cerone. Parameter estimation algorithms for a

set-membership description of uncertainty. Automatica, 26:887–898, 1990.

[BBFE93] S. Boyd, V. Balakrishnan, E. Feron, and L. El Ghaoui. Control system analysis

and synthesis via linear matrix inequalities. In Proc. American Control Conf.,

volume 2, pages 2147–2154, San Francisco, California, June 1993.

[BBK89] S. Boyd, V. Balakrishnan, and P. Kabamba. A bisection method for computing

the H∞ norm of a transfer matrix and related problems. Mathematics of Control,

Signals, and Systems, 2(3):207–219, 1989.

[BBP78] G. P. Barker, A. Berman, and R. J. Plemmons. Positive diagonal solutions to

the Lyapunov equations. Linear and Multilinear Algebra, 5:249–256, 1978.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 163

[BBT90] C. Burgat, A. Benzaouia, and S. Tarbouriech. Positively invariant sets of

discrete-time systems with constrained inputs. Int. J. Systems Science, 21:1249–

1271, 1990.

[BC85] S. Boyd and L. O. Chua. Fading memory and the problem of approximating

nonlinear operators with Volterra series. IEEE Trans. Circuits Syst., CAS-

32(11):1150–1171, November 1985.

[BCL83] B. R. Barmish, M. Corless, and G. Leitmann. A new class of stabilizing con-

trollers for uncertain dynamical systems. SIAM Journal on Control and Opti-

mization, 21(2):246–255, March 1983.

[BD85] V. G. Borisov and S. N. Diligenskii. Numerical method of stability analysis of

nonlinear systems. Automation and Remote Control, 46:1373–1380, 1985.

[BD87] S. Boyd and J. Doyle. Comparison of peak and RMS gains for discrete-time

systems. Syst. Control Letters, 9:1–6, June 1987.

[BDG

+

91] G. Balas, J. C. Doyle, K. Glover, A. Packard, and Roy Smith. µ-analysis and

synthesis. MUSYN, inc., and The Mathworks, Inc., 1991.

[BE93] S. Boyd and L. El Ghaoui. Method of centers for minimizing generalized eigenval-

ues. Linear Algebra and Applications, special issue on Numerical Linear Algebra

Methods in Control, Signals and Systems, 188:63–111, July 1993.

[Bec91] C. Beck. Computational issues in solving LMIs. In Proc. IEEE Conf. on Decision

and Control, volume 2, pages 1259–1260, Brighton, December 1991.

[Bec93] G. Becker. Quadratic Stability and Performance of Linear Parameter Dependent

Systems. PhD thesis, University of California, Berkeley, 1993.

[BEFB93] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear matrix inequalities

in system and control theory. In Proc. Annual Allerton Conf. on Communication,

Control and Computing, Allerton House, Monticello, Illinois, October 1993.

[Ben92] A. Bensoussan. Stochastic Control of Partially Observable Systems. Cambridge

University Press, Cambridge, Great Britain, 1992.

[BF63] R. Bellman and K. Fan. On systems of linear inequalities in Hermitian matrix

variables. In V. L. Klee, editor, Convexity, volume 7 of Proceedings of Symposia

in Pure Mathematics, pages 1–11. American Mathematical Society, 1963.

[BFBE92] V. Balakrishnan, E. Feron, S. Boyd, and L. El Ghaoui. Computing bounds for

the structured singular value via an interior point algorithm. In Proc. American

Control Conf., volume 4, pages 2195–2196, Chicago, June 1992.

[BG94] V. Blondel and M. Gevers. Simultaneous stabilizability question of three linear

systems is rationally undecidable. Mathematics of Control, Signals, and Systems,

6(1):135–145, 1994.

[BGR90] J. A. Ball, I. Gohberg, and L. Rodman. Interpolation of Rational Matrix Func-

tions, volume 45 of Operator Theory: Advances and Applications. Birkh¨ auser,

Basel, 1990.

[BGT81] R. G. Bland, D. Goldfarb, and M. J. Todd. The ellipsoid method: A survey.

Operations Research, 29(6):1039–1091, 1981.

[BH88a] D. Bernstein and W. Haddad. Robust stability and performance analysis for

state space systems via quadratic Lyapunov bounds. In Proc. IEEE Conf. on

Decision and Control, pages 2182–2187, 1988.

[BH88b] D. S. Bernstein and D. C. Hyland. Optimal projection equations for reduced-

order modeling, estimation, and control of linear systems with multiplicative

noise. J. Optimiz. Theory Appl., 58:387–409, 1988.

[BH89] D. Bernstein and W. Haddad. Robust stability and performance analysis for

linear dynamic systems. IEEE Trans. Aut. Control, AC-34(7):751–758, July

1989.

[BH93] D. Bernstein and W. Haddad. Nonlinear controllers for positive-real systems

with arbitrary input nonlinearities. In Proc. American Control Conf., volume 1,

pages 832–836, San Francisco, California, June 1993.

This electronic version is for personal use and may not be duplicated or distributed.

164 Bibliography

[BHL89] A. C. Bartlett, C. V. Hollot, and H. Lin. Root locations of an entire polytope

of polynomials: it suﬃces to check the edges. Mathematics of Control, Signals,

and Systems, 1(1):61–71, 1989.

[BHPD94] V. Balakrishnan, Y. Huang, A. Packard, and J. Doyle. Linear matrix inequalities

in analysis with multipliers. In Proc. American Control Conf., 1994. Invited

session on Recent Advances in the Control of Uncertain Systems.

[Bit88] G. Bitsoris. Positively invariant polyhedral sets of discrete-time linear systems.

Int. J. Control, 47(6):1713–1726, June 1988.

[BJL89] W. W. Barrett, C. R. Johnson, and M. Lundquist. Determinantal formulation for

matrix completions associated with chordal graphs. Linear Algebra and Appl.,

121:265–289, 1989.

[BK93] B. R. Barmish and H. I. Kang. A survey of extreme point results for robustness

of control systems. Automatica, 29(1):13–35, 1993.

[BNS89] A. Berman, M. Neumann, and R. Stern. Nonnegative Matrices in Dynamic

Systems. Wiley-Interscience, 1989.

[Boy86] S. Boyd. A note on parametric and nonparametric uncertainties in control sys-

tems. In Proc. American Control Conf., pages 1847–1849, Seattle, Washington,

June 1986.

[Boy94] S. Boyd. Robust control tools: Graphical user-interfaces and LMI algorithms.

Systems, Control and Information, 38(3):111–117, March 1994. Special issue on

Numerical Approaches in Control Theory.

[BP79] A. Berman and R. J. Plemmons. Nonnegative matrices in the mathematical

sciences. Computer Science and Applied Mathematics. Academic Press, 1979.

[BP91] G. Becker and A. Packard. Gain-scheduled state-feedback with quadratic stabil-

ity for uncertain systems. In Proc. American Control Conf., pages 2702–2703,

June 1991.

[BPG89a] J. Bernussou, P. L. D. Peres, and J. C. Geromel. A linear programming oriented

procedure for quadratic stabilization of uncertain systems. Syst. Control Letters,

13:65–72, 1989.

[BPG89b] J. Bernussou, P. L. D. Peres, and J. C. Geromel. Robust decentralized regulation:

A linear programming approach. IFAC/IFORS/IMACS Symposium Large Scale

Systems: Theory and Applications, Berlin-GDR, 1:135–138, 1989.

[BPPB93] G. Becker, A. Packard, D. Philbrick, and G. Balas. Control of parametrically

dependent linear systems: a single quadratic Lyapunov approach. In 1993 Amer-

ican Control Conference, volume 3, pages 2795–2799, June 1993.

[BR71] D. Bertsekas and I. Rhodes. Recursive state estimation for a set-membership

description of uncertainty. IEEE Trans. Aut. Control, AC-16(2):117–128, 1971.

[Bro65] R. Brockett. Optimization theory and the converse of the circle criterion. In

National Electronics Conference, volume 21, pages 697–701, San Francisco, Cal-

ifornia, 1965.

[Bro66] R. W. Brockett. The status of stability theory for deterministic systems. IEEE

Trans. Aut. Control, AC-11:596–606, July 1966.

[Bro70] R. W. Brockett. Finite Dimensional Linear Systems. John Wiley & Sons, New

York, 1970.

[Bro77] R. Brockett. On improving the circle criterion. In Proc. IEEE Conf. on Decision

and Control, pages 255–257, 1977.

[BT79] R. K. Brayton and C. H. Tong. Stability of dynamical systems: A constructive

approach. IEEE Trans. Circuits Syst., CAS-26(4):224–234, 1979.

[BT80] R. K. Brayton and C. H. Tong. Constructive stability and asymptotic stability

of dynamical systems. IEEE Trans. Circuits Syst., CAS-27(11):1121–1130, 1980.

[BW65] R. W. Brockett and J. L. Willems. Frequency domain stability criteria. IEEE

Trans. Aut. Control, AC-10:255–261, 401–413, 1965.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 165

[BW92] M. Bakonyi and H. J. Woerdeman. The central method for positive semi-deﬁnite,

contractive and strong Parrott type completion problems. Operator Theory:

Advances and Applications, 59:78–95, 1992.

[BY89] S. Boyd and Q. Yang. Structured and simultaneous Lyapunov functions for

system stability problems. Int. J. Control, 49(6):2215–2240, 1989.

[BYDM93] R. D. Braatz, P. M. Young, J. C. Doyle, and M. Morari. Computational com-

plexity of µ calculation. In Proc. American Control Conf., pages 1682–1683, San

Francisco, CA, June 1993.

[CD91] G. E. Coxson and C. L. DeMarco. Testing robust stability of general matrix

polytopes is an NP-hard computation. In Proc. Annual Allerton Conf. on Com-

munication, Control and Computing, pages 105–106, Allerton House, Monticello,

Illinois, 1991.

[CD92] G. E. Coxson and C. L. DeMarco. Computing the real structured singular value

is NP-hard. Technical Report ECE-92-4, Dept. of Elec. and Comp. Eng., Univ.

of Wisconsin-Madison, June 1992.

[CDW75] J. Cullum, W. Donath, and P. Wolfe. The minimization of certain nondiﬀer-

entiable sums of eigenvalues of symmetric matrices. Math. Programming Study,

3:35–55, 1975.

[CFA90] H.-D. Chiang and L. Fekih-Ahmed. A constructive methodology for estimating

the stability regions of interconnected nonlinear systems. IEEE Trans. Circuits

Syst., 37(5):577–588, 1990.

[CH93] E. B. Castelan and J. C. Hennet. On invariant polyhedra of continuous-time

linear systems. IEEE Trans. Aut. Control, 38(11):1680–1685, December 1993.

[CHD93] E. G. Collins, W. M. Haddad, and L. D. Davis. Riccati equation approaches for

robust stability and performance. In Proc. American Control Conf., volume 2,

pages 1079–1083, San Francisco, California, June 1993.

[Che80a] F. L. Chernousko. Optimal guaranteed estimates of indeterminancies with the

aid of ellipsoids I. Engineering Cybernetics, 18(3):1–9, 1980.

[Che80b] F. L. Chernousko. Optimal guaranteed estimates of indeterminancies with the

aid of ellipsoids II. Engineering Cybernetics, 18(4):1–9, 1980.

[Che80c] F. L. Chernousko. Optimal guaranteed estimates of indeterminancies with the

aid of ellipsoids III. Engineering Cybernetics, 18(5):1–7, 1980.

[CL90] M. Corless and G. Leitmann. Deterministic control of uncertain systems: a Lya-

punov theory approach. In A. Zinober, editor, Deterministic control of uncertain

systems. Peregrinus Ltd., 1990.

[CM81] B. Craven and B. Mond. Linear programming with matrix variables. Linear

Algebra and Appl., 38:73–80, 1981.

[CNF92] J. Chen, C. N. Nett, and M. K. H. Fan. Worst-case identiﬁcation in H∞: Vali-

dation of a priori information, essentially optimal algorithms, and error bounds.

In Proc. American Control Conf., pages 251–257, Chicago, June 1992.

[Cor85] M. Corless. Stabilization of uncertain discrete-time systems. Proceedings of the

IFAC workshop on model error concepts and compensation, Boston, USA, pages

125–128, 1985.

[Cor90] M. Corless. Guaranteed rates of exponential convergence for uncertain dynamical

systems. Journal of Optimization Theory and Applications, 64:481–494, 1990.

[Cor94] M. Corless. Robust stability analysis and controller design with quadratic Lya-

punov functions. In A. Zinober, editor, Variable structure and Lyapunov control,

volume 193 of Lecture notes in control and information sciences, chapter 9, pages

181–203. Springer-Verlag, 1994.

[CP72] S. Chang and T. Peng. Adaptive guaranteed cost control of systems with uncer-

tain parameters. IEEE Trans. Aut. Control, AC-17:474–483, August 1972.

[CP84] B. C. Chang and J. B. Pearson, Jr. Optimal disturbance rejection in linear

multivariable systems. IEEE Trans. Aut. Control, AC-29(10):880–887, October

1984.

This electronic version is for personal use and may not be duplicated or distributed.

166 Bibliography

[CPS92] R. Cottle, J. S. Pang, and R. E. Stone. The Linear Complementarity Problem.

Academic Press, 1992.

[CS92a] R. Chiang and M. G. Safonov. Robust Control Toolbox. The Mathworks, Inc.,

1992.

[CS92b] R. Y. Chiang and M. G. Safonov. Real Km-synthesis via generalized Popov mul-

tipliers. In Proc. American Control Conf., volume 3, pages 2417–2418, Chicago,

June 1992.

[CYP93] M. Cheung, S. Yurkovich, and K. M. Passino. An optimal volume ellipsoid

algorithm for parameter estimation. IEEE Trans. Aut. Control, AC-38(8):1292–

1296, 1993.

[Dan92] J. Dancis. Positive semideﬁnite completions of partial hermitian matrices. Linear

Algebra and Appl., 175:97–114, 1992.

[Dan93] J. Dancis. Choosing the inertias for completions of certain partially speciﬁed

matrices. SIAM Jour. Matrix Anal. Appl., 14(3):813–829, July 1993.

[Dav94] J. David. Algorithms for analysis and design of robust controllers. PhD thesis,

Kat. Univ. Leuven, ESAT, 3001 Leuven, Belgium, 1994.

[Del89] J. Deller. Set membership identiﬁcation in digital signal processing. IEEE Trans.

Acoust., Speech, Signal Processing, 6(4):4–20, 1989.

[Des65] C. A. Desoer. A generalization of the Popov criterion. IEEE Trans. Aut. Control,

AC-10:182–185, 1965.

[DG81] H. Dym and I. Gohberg. Extensions of band matrices with band inverses. Linear

Algebra and Appl., 36:1–24, 1981.

[DGK79] P. Delsarte, Y. Genin, and Y. Kamp. The Nevanlinna-Pick problem for matrix-

valued functions. SIAM J. Appl. Math., 36(1):47–61, February 1979.

[DGK81] P. Delsarte, Y. Genin, and Y. Kamp. On the role of the Nevanlinna-Pick problem

in circuit and systems theory. Circuit Theory and Appl., 9:177–187, 1981.

[Dik67] I. Dikin. Iterative solution of problems of linear and quadratic programming.

Soviet Math. Dokl., 8(3):674–675, 1967.

[DK71] E. J. Davison and E. M. Kurak. A computational method for determining

quadratic Lyapunov functions for non-linear systems. Automatica, 7(5):627–636,

September 1971.

[DKW82] C. Davis, W. M. Kahan, and H. F. Weinberger. Norm-preserving dilations and

their applications to optimal error bounds. SIAM J. Numerical Anal., 19(3):445–

469, June 1982.

[Doo81] P. Van Dooren. A generalized eigenvalue approach for solving Riccati equations.

SIAM J. Sci. Stat. Comp., 2:121–135, June 1981.

[Doy82] J. Doyle. Analysis of feedback systems with structured uncertainties. IEE Proc.,

129-D(6):242–250, November 1982.

[DRT92] D. Den Hertog, C. Roos, and T. Terlaky. A large-step analytic center method

for a class of smooth convex programming problems. SIAM J. Optimization,

2(1):55–70, February 1992.

[dSFX93] C. E. de Souza, M. Fu, and L. Xie. H∞ analysis and synthesis of discrete-time

systems with time-varying uncertainty. IEEE Trans. Aut. Control, AC-38(3),

March 1993.

[DV75] C. A. Desoer and M. Vidyasagar. Feedback Systems: Input-Output Properties.

Academic Press, New York, 1975.

[EBFB92] L. El Ghaoui, V. Balakrishnan, E. Feron, and S. Boyd. On maximizing a robust-

ness measure for structured nonlinear perturbations. In Proc. American Control

Conf., volume 4, pages 2923–2924, Chicago, June 1992.

[EG93] L. El Ghaoui and P. Gahinet. Rank minimization under LMI constraints: a

framework for output feedback problems. In Proc. of the European Contr. Conf.,

pages 1176–1179, July 1993.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 167

[EG94] L. El Ghaoui. Analysis and state-feedback synthesis for systems with multiplica-

tive noise. Syst. Control Letters, 1994. Submitted.

[EHM92] L. Elsner, C. He, and V. Mehrmann. Minimization of the norm, the norm of the

inverse and the condition number of a matrix by completion. Technical Report

92-028, Universit¨ at Bielefeld, POB 8640, D-4800 Bielefeld 1, Germany, 1992.

[EHM93] L. Elsner, C. He, and V. Mehrmann. Minimization of the condition number of a

positive-deﬁnite matrix by completion. Technical report, TU-Chemnitz-Zwikau,

Chemnitz, Germany, September 1993.

[EP92] A. El Bouhtouri and A. J. Pritchard. Stability radii of linear systems with respect

to stochastic perturbations. Syst. Control Letters, 19(1):29–33, July 1992.

[EP94] A. El Bouthouri and A. Pritchard. A Riccati equation approach to maximizing

the stability radius of a linear system under structured stochastic Lipschitzian

perturbations. Syst. Control Letters, January 1994.

[EZKN90] S. V. Emelyanov, P. V. Zhivoglyadov, S. K. Korovin, and S. V. Nikitin. Group of

isometries and some stabilization and stabilizability aspects of uncertain systems.

Problems of Control and Information Theory, 19(5-6):451–477, 1990.

[Fan93] M. K. H. Fan. A quadratically convergent algorithm on minimizing the largest

eigenvalue of a symmetric matrix. Linear Algebra and Appl., 188-189:231–253,

1993.

[Fau67] P. Faurre. Representation of Stochastic Processes. PhD thesis, Stanford Univer-

sity, 1967.

[Fau73] P. Faurre. R´ealisations Markoviennes de processus stationnaires. Technical

Report 13, INRIA, Domaine de Voluceau, Rocquencourt, 78150, Le Chesnay,

France, 1973.

[FBB92] E. Feron, V. Balakrishnan, and S. Boyd. Design of stabilizing state feedback for

delay systems via convex optimization. In Proc. IEEE Conf. on Decision and

Control, volume 4, pages 2195–2196, Tucson, December 1992.

[FBBE92] E. Feron, V. Balakrishnan, S. Boyd, and L. El Ghaoui. Numerical methods

for H2 related problems. In Proc. American Control Conf., volume 4, pages

2921–2922, Chicago, June 1992.

[FD77] P. Faurre and M. Depeyrot. Elements of System Theory. North-Holland, New

York, 1977.

[Fed75] P. I. Fedˇcina. Descriptions of solutions of the tangential Nevanlinna-Pick prob-

lem. Doklady Akad. Nauk. Arm., 60:37–42, 1975.

[Fer93] E. Feron. Linear Matrix Inequalities for the Problem of Absolute Stability of

Control Systems. PhD thesis, Stanford University, Stanford, CA 94305, October

1993.

[FG88] M. E. Fisher and J. E. Gayek. Estimating reachable sets for two-dimensional lin-

ear discrete systems. Journal of Optimization Theory and Applications, 56(1):67–

88, January 1988.

[FH82] E. Fogel and Y. Huang. On the value of information in system identiﬁcation-

bounded noise case. Automatica, 18(2):229–238, 1982.

[Fil88] A. F. Filippov. Diﬀerential Equations with Discontinuous Righthand Sides.

Kluwer Academic Publishers, Dordrect, The Netherlands, 1988. Translated from

1960 Russian version.

[Fin37] P. Finsler.

¨

Uber das Vorkommen deﬁniter und semi-deﬁniter Formen in Scharen

quadratischer Formen. Comentarii Mathematici Helvetici, 9:199–192, 1937.

[Fle85] R. Fletcher. Semideﬁnite matrix constraints in optimization. SIAM J. Control

and Opt., 23:493–513, 1985.

[FM68] A. Fiacco and G. McCormick. Nonlinear programming: sequential unconstrained

minimization techniques. Wiley, 1968. Reprinted 1990 in the SIAM Classics in

Applied Mathematics series.

This electronic version is for personal use and may not be duplicated or distributed.

168 Bibliography

[FN84] T. Fujii and M. Narazaki. A complete optimality condition in the inverse problem

of optimal control. SIAM J. on Control and Optimization, 22(2):327–341, March

1984.

[FN92] M. K. H. Fan and B. Nekooie. On minimizing the largest eigenvalue of a sym-

metric matrix. In Proc. IEEE Conf. on Decision and Control, pages 134–139,

December 1992.

[FNO87] S. Friedland, J. Nocedal, and M. Overton. The formulation and analysis of

numerical methods for inverse eigenvalue problems. SIAM J. Numer. Anal.,

24(3):634–667, June 1987.

[Fog79] E. Fogel. System identiﬁcation via membership set constraints with energy con-

strained noise. IEEE Trans. Aut. Control, AC-24(5):752–757, 1979.

[FP62] M. Fiedler and V. Ptak. On matrices with non-positive oﬀ-diagonal elements

and positive principal minors. Czechoslovak Mathematical Journal, 87:382–400,

1962.

[FP66] M. Fiedler and V. Ptak. Some generalizations of positive deﬁniteness and mono-

tonicity. Numerische Mathematik, 9:163–172, 1966.

[FP67] M. Fiedler and V. Ptak. Diagonally dominant matrices. Czechoslovak Mathe-

matical Journal, 92:420–433, 1967.

[FR75] W. H. Fleming and R. W. Rishel. Deterministic and stochastic optimal control.

Springer-Verlag, 1975.

[Fra87] B. A. Francis. A course in H∞ Control Theory, volume 88 of Lecture Notes in

Control and Information Sciences. Springer-Verlag, 1987.

[FS55] G. E. Forsythe and E. G. Straus. On best conditioned matrices. Proc. Amer.

Math. Soc., 6:340–345, 1955.

[FT86] M. K. H. Fan and A. L. Tits. Characterization and eﬃcient computation of the

structured singular value. IEEE Trans. Aut. Control, AC-31(8):734–743, August

1986.

[FT88] M. K. H. Fan and A. L. Tits. m-form numerical range and the computation

of the structured singular value. IEEE Trans. Aut. Control, AC-33(3):284–289,

March 1988.

[FTD91] M. K. H. Fan, A. L. Tits, and J. C. Doyle. Robustness in the presence of mixed

parametric uncertainty and unmodeled dynamics. IEEE Trans. Aut. Control,

AC-36(1):25–38, January 1991.

[FY79] A. L. Fradkov and V. A. Yakubovich. The S-procedure and duality relations in

nonconvex problems of quadratic programming. Vestnik Leningrad Univ. Math.,

6(1):101–109, 1979. In Russian, 1973.

[GA93] P. Gahinet and P. Apkarian. A LMI-based parametrization of all H∞ controllers

with applications. In Proc. IEEE Conf. on Decision and Control, pages 656–661,

San Antonio, Texas, December 1993.

[GA94] P. Gahinet and P. Apkarian. A convex characterization of parameter dependent

H∞ controllers. IEEE Trans. Aut. Control, 1994. To appear.

[Gah92] P. Gahinet. A convex parametrization of H∞ suboptimal controllers. In Proc.

IEEE Conf. on Decision and Control, pages 937–942, 1992.

[Gay91] J. E. Gayek. A survey of techniques for approximating reachable and controllable

sets. In Proc. IEEE Conf. on Decision and Control, pages 1724–1729, Brighton,

December 1991.

[GB86] A. R. Galimidi and B. R. Barmish. The constrained Lyapunov problem and its

application to robust output feedback stabilization. IEEE Trans. Aut. Control,

AC-31(5):410–419, May 1986.

[GCG93] F. Garofalo, G. Celentano, and L. Glielmo. Stability robustness of interval

matrices via Lyapunov quadratic forms. IEEE Trans. Aut. Control, 38(2):281–

284, February 1993.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 169

[GCZS93] K. M. Grigoriadis, M. J. Carpenter, G. Zhu, and R. E. Skelton. Optimal redesign

of linear systems. In Proc. American Control Conf., volume 3, pages 2680–2684,

San Francisco, California, June 1993.

[GdSS93] J. C. Geromel, C. C. de Souza, and R. E. Skelton. Static output feedback

controllers: Robust stability and convexity. Submitted to IEEE Trans. Aut.

Cont., 1993.

[Ger85] J. C. Geromel. On the determination of a diagonal solution of the Lyapunov

equation. IEEE Trans. Aut. Control, AC-30:404–406, April 1985.

[GFS93] K. M. Grigoriadis, A. E. Frazho, and R. E. Skelton. Application of alternating

convex projections methods for computation of positive Toeplitz matrices. IEEE

Trans. Signal Proc., September 1993. To appear.

[GG94] P. B. Gapski and J. C. Geromel. A convex approach to the absolute stability

problem. IEEE Trans. Aut. Control, November 1994.

[Gha90] L. El Ghaoui. Robustness of Linear Systems against Parameter Variations. PhD

thesis, Stanford University, Stanford CA 94305, March 1990.

[Gha94] L. El Ghaoui. Quadratic Lyapunov functions for rational systems: An LMI

approach. Submitted to IFAC Conf. on Robust Control., 1994.

[Gil66] E. Gilbert. An iterative procedure for computing the minimum of a quadratic

form on a convex set. SIAM J. on Control, 4(1):61–80, 1966.

[GJSW84] R. Grone, C. R. Johnson, E. M S´ a, and H. Wolkowicz. Positive deﬁnite com-

pletions of partial Hermitian matrices. Linear Algebra and Appl., 58:109–124,

1984.

[GL81] P. G´ acs and L. Lov´ asz. Khachiyan’s algorithm for linear programming. Math.

Program. Studies, 14:61–68, 1981.

[GL89] G. Golub and C. Van Loan. Matrix Computations. Johns Hopkins Univ. Press,

Baltimore, second edition, 1989.

[GL93a] M. Gevers and G. Li. Parametrizations in Control, Estimation and Filter-

ing Problems: Accuracy Aspects. Communications and Control Engineering.

Springer-Verlag, London, January 1993.

[GL93b] K. Gu and N. K. Loh. Direct computation of stability bounds for systems with

polytopic uncertainties. IEEE Trans. Aut. Control, 38(2):363–366, February

1993.

[GLS88] M. Gr¨ otschel, L. Lov´ asz, and A. Schrijver. Geometric Algorithms and Combina-

torial Optimization, volume 2 of Algorithms and Combinatorics. Springer-Verlag,

1988.

[GN93] P. Gahinet and A. Nemirovskii. LMI Lab: A Package for Manipulating and

Solving LMIs. INRIA, 1993.

[GP82] S. Gutman and Z. Palmor. Properties of min-max controllers in uncertain dy-

namical systems. SIAM Journal on Control and Optimization, 20(6):850–861,

November 1982.

[GPB91] J. C. Geromel, P. L. D. Peres, and J. Bernussou. On a convex parameter space

method for linear control design of uncertain systems. SIAM J. Control and

Optimization, 29(2):381–402, March 1991.

[GR88] D. P. Goodall and E. P. Ryan. Feedback controlled diﬀerential inclusions and

stabilization of uncertain dynamical systems. SIAM Journal on Control and

Optimization, 26(6):1431–1441, November 1988.

[GS88] R. R. E. De Gaston and M. G. Safonov. Exact calculation of the multiloop

stability margin. IEEE Trans. Aut. Control, AC-33(2):156–171, 1988.

[GS92] K. M. Grigoriadis and R. E. Skelton. Alternating convex projection methods for

covariance control design. In Proc. Annual Allerton Conf. on Communication,

Control and Computing, Allerton House, Monticello, Illinois, 1992.

[GT88] C. Goh and D. Teo. On minimax eigenvalue problems via constrained optimiza-

tion. Journal of Optimization Theory and Applications, 57(1):59–68, 1988.

This electronic version is for personal use and may not be duplicated or distributed.

170 Bibliography

[GTS

+

94] K. C. Goh, L. Turan, M. G. Safonov, G. P. Papavassilopoulos, and J. H. Ly. Bi-

aﬃne matrix inequality properties and computational methods. In Proc. Amer-

ican Control Conf., 1994.

[Gu92a] K. Gu. Designing stabilizing control of uncertain systems by quasiconvex opti-

mization. IEEE Trans. Aut. Control, AC-38(9), 1992.

[Gu92b] K. Gu. Linear feedback stabilization of arbitrary time-varying uncertain systems.

In Proc. American Control Conf., Chicago, June 1992.

[Gu93] K. Gu. The trade-oﬀ between H∞ attenuation and uncertainty. In Proc. Amer-

ican Control Conf., pages 1666–1670, San Fransico, June 1993.

[Gu94] K. Gu. Designing stabilizing control of uncertain systems by quasiconvex opti-

mization. IEEE Trans. Aut. Control, 39(1):127–131, January 1994.

[Gut79] S. Gutman. Uncertain dynamical systems — a Lyapunov min-max approach.

IEEE Trans. Aut. Control, AC-24(3):437–443, June 1979.

[GV74] G. Golub and J. Varah. On a characterization of the best 2 scaling of a matrix.

SIAM J. on Numerical Analysis, 11(3):472–479, June 1974.

[GZL90] K. Gu, M. A. Zohdy, and N. K. Loh. Necessary and suﬃcient conditions of

quadratic stability of uncertain linear systems. IEEE Trans. Aut. Control, AC-

35(5):601–604, May 1990.

[HB76] H. P. Horisberger and P. R. B´elanger. Regulators for linear, time invariant plants

with uncertain parameters. IEEE Trans. Aut. Control, AC-21:705–708, 1976.

[HB80] C. V. Hollot and B. R. Barmish. Optimal quadratic stabilizability of uncer-

tain linear systems. Proc. 18th Allerton Conf. on Communication, Control and

Computing, University of Illinois, Monticello, pages 697–706, 1980.

[HB91a] W. Haddad and D. Bernstein. Parameter-dependent Lyapunov functions, con-

stant real parameter uncertainty, and the Popov criterion in robust analysis

and synthesis. In Proc. IEEE Conf. on Decision and Control, pages 2274–2279,

2632–2633, Brighton, U.K., December 1991.

[HB91b] J. Hennet and J. Beziat. A class of invariant regulators for the discrete-time

linear constrained regulator problem. Automatica, 27:549–554, 1991.

[HB93a] W. Haddad and D. Bernstein. Explicit construction of quadratic Lyapunov

functions for the small gain, positivity, circle, and Popov theorems and their

application to robust stability. part I: Continuous-time theory. Int. J. Robust

and Nonlinear Control, 3:313–339, 1993.

[HB93b] W. Haddad and D. Bernstein. Oﬀ-axis absolute stability criteria and µ-bounds

involving non-positive-real plant-dependent multipliers for robust stability and

performance with locally slope-restricted monotonic nonlinearities. In Proc.

American Control Conf., volume 3, pages 2790–2794, San Francisco, Califor-

nia, June 1993.

[HCB93] W. Haddad, E. Collins, and D. Bernstein. Robust stability analysis using the

small gain, circle, positivity, and Popov theorems: A comparative study. IEEE

Trans. Control Sys. Tech., 1(4):290–293, December 1993.

[HCM93] W. M. Haddad, E. G. Collins, and R. Moser. Fixed structure computation of the

structured singular value. In Proc. American Control Conf., pages 1673–1674,

San Fransico, June 1993.

[Hes81] M. R. Hestenes. Optimization Theory, the Finite Dimensional Case. Krieger,

Huntington, New York, 1981.

[HH93a] S. R. Hall and J. P. How. Mixed H2/µ performance bounds using dissipation

theory. In Proc. IEEE Conf. on Decision and Control, San Antonio, Texas,

December 1993.

[HH93b] J. P. How and S. R. Hall. Connections between the Popov stability criterion

and bounds for real parametric uncertainty. In Proc. American Control Conf.,

volume 2, pages 1084–1089, June 1993.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 171

[HHH93a] J. P. How, W. M. Haddad, and S. R. Hall. Robust control synthesis examples

with real parameter uncertainty using the Popov criterion. In Proc. American

Control Conf., volume 2, pages 1090–1095, San Francisco, California, June 1993.

[HHH93b] J. P. How, S. R. Hall, and W. M. Haddad. Robust controllers for the Middeck

Active Control Experiment using Popov controller synthesis. In AIAA Guid.,

Nav., and Control Conf., pages 1611–1621, August 1993. Submitted to IEEE

Transactions on Control Systems Technology.

[HHHB92] W. Haddad, J. How, S. Hall, and D. Bernstein. Extensions of mixed-µ bounds to

monotonic and odd monotonic nonlinearities using absolute stability theory. In

Proc. IEEE Conf. on Decision and Control, Tucson, Arizona, December 1992.

[HJ91] R. Horn and C. Johnson. Topics in Matrix Analysis. Cambridge University

Press, Cambridge, 1991.

[Hol87] C. V. Hollot. Bound invariant Lyapunov functions: A means for enlarging the

class of stabilizable uncertain systems. Int. J. Control, 46(1):161–184, 1987.

[How93] J. P. How. Robust Control Design with Real Parameter Uncertainty using Abso-

lute Stability Theory. PhD thesis, Massachusetts Institute of Technology, Febru-

ary 1993.

[Hu87] H. Hu. An algorithm for rescaling a matrix positive deﬁnite. Linear Algebra and

its Applications, 96:131–147, 1987.

[Hua67] P. Huard. Resolution of mathematical programming with nonlinear constraints

by the method of centers. North Holland, Amsterdam, 1967.

[HUL93] J.-B. Hiriart-Urruty and C. Lemar´echal. Convex Analysis and Minimization

Algorithms I, volume 305 of Grundlehren der mathematischen Wissenschaften.

Springer-Verlag, New York, 1993.

[HUY92] J.-B. Hiriart-Urruty and D. Ye. Sensitivity analysis of all eigenvalues of a sym-

metric matrix. Technical report, Univ. Paul Sabatier, Toulouse, 1992.

[HW93] J. W. Helton and H. J. Woerdeman. Symmetric Hankel operators: Minimal

norm extensions and eigenstructures. Linear Algebra and Appl., 185:1–19, 1993.

[IR64] E. S. Ibrahim and Z. V. Rekasius. Construction of Liapunov functions for non-

linear feedback systems with several nonlinear elements. In Proc. Joint American

Control Conf., pages 256–260, June 1964.

[IS93a] T. Iwasaki and R. E. Skelton. A complete solution to the general H∞ control

problem: LMI existence conditions and state space formulas. In Proc. American

Control Conf., pages 605–609, San Francisco, June 1993.

[IS93b] T. Iwasaki and R. E. Skelton. Parametrization of all stabilizing controllers via

quadratic Lyapunov functions. Submitted to J. Optim. Theory Appl., 1993.

[ISG93] T. Iwasaki, R. E. Skelton, and J. C. Geromel. Linear quadratic suboptimal

control via static output feedback. Submitted to Syst. and Cont. Letters, 1993.

[Ito51] K. Ito. On a formula concerning stochastic diﬀerentials. Nagoya Math. J., 3(55),

1951.

[Iwa93] T. Iwasaki. A Uniﬁed Matrix Inequality Approach to Linear Control Design.

PhD thesis, Purdue University, West Lafayette, IN 47907, December 1993.

[Jar91] F. Jarre. Interior-point methods for convex programming. To appear, Appl.

Math. Opt., 1991.

[Jar93a] F. Jarre. An interior-point method for convex fractional programming. Submitted

to Math. Progr., March 1993.

[Jar93b] F. Jarre. An interior-point method for minimizing the maximum eigenvalue of

a linear combination of matrices. SIAM J. Control and Opt., 31(5):1360–1377,

September 1993.

[Jar93c] F. Jarre. Optimal ellipsoidal approximations around the analytic center. Tech-

nical report, Intit¨ ut f¨ ur Angewandte Mathematik, University of W¨ urzburg, 8700

W¨ urzburg, Germany, 1993.

This electronic version is for personal use and may not be duplicated or distributed.

172 Bibliography

[JL65] E. I. Jury and B. W. Lee. The absolute stability of systems with many nonlin-

earities. Automation and Remote Control, 26:943–961, 1965.

[JL92] C. R. Johnson and M. Lundquist. An inertia formula for Hermitian matrices

with sparse inverses. Linear Algebra and Appl., 162:541–556, 1992.

[Joh85] F. John. Extremum problems with inequalities as subsidiary conditions. In

J. Moser, editor, Fritz John, Collected Papers, pages 543–560. Birkhauser,

Boston, Massachussetts, 1985.

[Kah68] W. Kahan. Circumscribing an ellipsoid about the intersection of two ellipsoids.

Canad. Math. Bull., 11(3):437–441, 1968.

[Kai80] T. Kailath. Linear Systems. Prentice-Hall, New Jersey, 1980.

[Kal63a] R. E. Kalman. Lyapunov functions for the problem of Lur’e in automatic control.

Proc. Nat. Acad. Sci., USA, 49:201–205, 1963.

[Kal63b] R. E. Kalman. On a new characterization of linear passive systems. In Proc.

First Annual Allerton Conf. on Communication, Control and Computing, pages

456–470, 1963.

[Kal64] R. E. Kalman. When is a linear control system optimal? Trans. ASME, J.

Basic Eng., Ser. D., 86:51–60, 1964.

[Kam83] V. A. Kamenetskii. Absolute stability and absolute instability of control systems

with several nonlinear nonstationary elements. Automation and Remote Control,

44(12):1543–1552, 1983.

[Kam89] V. A. Kamenetskii. Convolution method for matrix inequalities. Automation

and Remote Control, 50(5):598–607, 1989.

[Kar84] N. Karmarkar. A new polynomial-time algorithm for linear programming. Com-

binatorica, 4(4):373–395, 1984.

[Kav94] D. Kavrano˘ glu. Computation of the optimal solution for the zeroth order L∞

norm approximation problem. Syst. Control Letters, 1994. Submitted.

[KB93] E. Kaszkurewicz and A. Bhaya. Robust stability and diagonal Lyapunov func-

tions. SIAM J. on Matrix Analysis and Applications, 14(2):508–520, April 1993.

[Kha78] V. L. Kharitonov. Asymptotic stability of an equilibrium position of a family of

systems of linear diﬀerential equations. Diﬀerential’nye Uraveniya, 14(11):1483–

1485, 1978.

[Kha79] L. Khachiyan. A polynomial algorithm in linear programming. Soviet Math.

Dokl., 20:191–194, 1979.

[Kha82] H. K. Khalil. On the existence of positive diagonal P such that PA+A

T

P < 0.

IEEE Trans. Aut. Control, AC-27:181–184, 1982.

[Kim87] H. Kimura. Directional interpolation approach to H

∞

-optimization and robust

stabilization. IEEE Trans. Aut. Control, 32(12):1085–1093, December 1987.

[Kis91] M. Kisielewicz. Diﬀerential inclusions and optimal control, volume 44 of Mathe-

matics and its applications; East European series. Kluwer Academic Publishers,

Boston, 1991.

[KK60] I. I. Kats and N. N. Krasovskii. On the stability of systems with random pa-

rameters. Prikl. Mat. Mek., 809, 1960.

[KK92] L. Khachiyan and B. Kalantari. Diagonal matrix scaling and linear program-

ming. SIAM J. on Optimization, 2(4):668–672, November 1992.

[KKR93] I. Kaminer, P. P. Khargonekar, and M. A. Rotea. Mixed H2/H∞ control for

discrete-time systems via convex optimization. Automatica, 29(1):57–70, 1993.

[KL85a] B. Kouvaritakis and D. Latchman. Necessary and suﬃcient stability criterion for

systems with structured uncertainties: The major principal direction alignment

principle. Int. J. Control, 42(3):575–598, 1985.

[KL85b] B. Kouvaritakis and D. Latchman. Singular value and eigenvalue techniques

in the analysis of systems with structured perturbations. Int. J. Control,

41(6):1381–1412, 1985.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 173

[KLB90] R. Kosut, M. Lau, and S. Boyd. Identiﬁcation of systems with parametric and

nonparametric uncertainty. In Proc. American Control Conf., pages 2412–2417,

San Diego, May 1990.

[KLB92] R. Kosut, M. Lau, and S. Boyd. Set-membership identiﬁcation of systems with

parametric and nonparametric uncertainty. IEEE Trans. Aut. Control, AC-

37(7):929–941, July 1992.

[Kle69] D. L. Kleinman. On the stability of linear stochastic systems. IEEE Trans. Aut.

Control, August 1969.

[Kok92] P. V. Kokotovic. The joy of feedback: Nonlinear and adaptive. IEEE Control

Systems, June 1992.

[Koz69] F. Kozin. A survey of stability of stochastic systems. Automatica, 5:95–112,

1969.

[KP87a] V. A. Kamenetskii and E. S. Pyatnitskii. Gradient method of constructing

Lyapunov functions in problems of absolute stability. Automation and Remote

Control, pages 1–9, 1987.

[KP87b] V. A. Kamenetskii and Y. S. Pyatnitskii. An iterative method of Lyapunov

function construction for diﬀerential inclusions. Syst. Control Letters, 8:445–

451, 1987.

[KPR88] P. P. Khargonekar, I. R. Petersen, and M. A. Rotea. H∞-optimal control with

state-feedback. IEEE Trans. Aut. Control, AC-33(8):786–788, August 1988.

[KPY91] K. Kortanek, F. Potra, and Y. Ye. On some eﬃcient interior point methods for

nonlinear convex programming. Linear Algebra and Appl., 152:169–189, 1991.

[KPZ90] P. P. Khargonekar, I. R. Petersen, and K. Zhou. Robust stabilization of uncertain

linear systems: quadratic stabilizability and H∞ control theory. IEEE Trans.

Aut. Control, AC-35(3):356–361, March 1990.

[KR88] P. P. Khargonekar and M. A. Rotea. Stabilization of uncertain systems with

norm bounded uncertainty using control Lyapunov functions. Proc. IEEE Conf.

on Decision and Control, 1:503–507, 1988.

[KR91] P. P. Khargonekar and M. A. Rotea. Mixed H2/H∞ control: A convex opti-

mization approach. IEEE Trans. Aut. Control, AC-36(7):824–837, July 1991.

[Kra56] N. N. Krasovskii. Application of Lyapunov’s second method for equations with

time delay. Prikl. Mat. Mek., 20(3):315–327, 1956.

[Kra91] J. Kraaijevanger. A characterization of Lyapunov diagonal stability using

Hadamard products. Linear Algebra and Appl., 151:245–254, 1991.

[KRB89] N. Khraishi, R. Roy, and S. Boyd. A speciﬁcation language approach to solving

convex antenna weight design problems. Technical Report 8901, Stanford Uni-

versity, Information Systems Laboratory, Electrical Engineering Department,

January 1989.

[KS75] V. J. S. Karmarkar and D. D. Siljak. Maximization of absolute stability regions

by mathematical programming methods. Regelungstechnik, 2:59–61, 1975.

[KT93] L. G. Khachiyan and M. J. Todd. On the complexity of approximating the

maximal inscribed ellipsoid for a polytope. Mathematical Programming, 61:137–

159, 1993.

[Kus67] H. J. Kushner. Stochastic stability and control. Academic Press, New York,

1967.

[Lau79] A. J. Laub. A Schur method for solving algebraic Riccati equations. IEEE

Trans. Aut. Control, AC-24(6):913–921, December 1979.

[Lef65] S. Lefschetz. Stability of Nonlinear Control Systems. Mathematics in Science

and Engineering. Academic Press, New York, 1965.

[Lei79] G. Leitmann. Guaranteed asymptotic stability for some linear systems with

bounded uncertainties. Trans. of the ASME: Journal of Dynamic Systems, Mea-

surement, and Control, 101:212–216, September 1979.

This electronic version is for personal use and may not be duplicated or distributed.

174 Bibliography

[Let61] A. M. Letov. Stability in Nonlinear Control Systems. Princeton University Press,

Princeton, 1961.

[Lev59] J. J. Levin. On the matrix Riccati equation. Proc. Amer. Math. Soc., 10:519–524,

1959.

[LH66] B. Lieu and P. Huard. La m´ethode des centres dans un espace topologique.

Numerische Mathematik, 8:56–67, 1966.

[Liu68] R. W. Liu. Convergent systems. IEEE Trans. Aut. Control, AC-13(4):384–391,

August 1968.

[LKB90] M. Lau, R. Kosut, and S. Boyd. Parameter set estimation of systems with

uncertain nonparametric dynamics and disturbances. In Proc. IEEE Conf. on

Decision and Control, pages 3162–3167, Honolulu, Hawaii, December 1990.

[LLJ90] R. Lozano-Leal and S. Joshi. Strictly positive real transfer functions revisited.

IEEE Trans. Aut. Control, AC-35:1243–1245, 1990.

[LN92] H. A. Latchman and R. J. Norris. On the equivalence between similarity and

nonsimilarity scalings in robustness analysis. IEEE Trans. Aut. Control, AC-

37(11):1849–1853, 1992.

[Log90] H. Logemann. Circle criteria, small-gain conditions and internal stability for

inﬁnite-dimensional systems. Technical Report 222, Institut f¨ ur Dynamische

Systeme, Universit¨ at Bremen, April 1990.

[LP44] A. I. Lur’e and V. N. Postnikov. On the theory of stability of control systems.

Applied mathematics and mechanics, 8(3), 1944. In Russian.

[LR80] P. Lancaster and L. Rodman. Existence and uniqueness theorems for the alge-

braic Riccati equation. Int. J. Control, 32(2):285–309, 1980.

[LSA94] J. H. Ly, M. G. Safonov, and F. Ahmad. Positive real Parrott theorem with

application to LMI controller synthesis. In Proc. American Control Conf., 1994.

[LSC94] J. H. Ly, M. G. Safonov, and R. Y Chiang. Real/complex multivariable stability

margin computation via generalized Popov multiplier — LMI approach. In Proc.

American Control Conf., 1994.

[LSG92] K. Liu, R. E. Skelton, and K. Grigoriadis. Optimal controllers for ﬁnite

wordlength implementation. IEEE Trans. Aut. Control, 372, 1992.

[LSL69] R. Liu, R. Saeks, and R. J. Leake. On global linearization. SIAM-AMS proceed-

ings, pages 93–102, 1969.

[Lur57] A. I. Lur’e. Some Nonlinear Problems in the Theory of Automatic Control. H.

M. Stationery Oﬀ., London, 1957. In Russian, 1951.

[Lya47] A. M. Lyapunov. Probl`eme g´en´eral de la stabilit´e du mouvement, volume 17 of

Annals of Mathematics Studies. Princeton University Press, Princeton, 1947.

[LZD91] W. M. Lu, K. Zhou, and J. C. Doyle. Stabilization of LFT systems. In Proc.

IEEE Conf. on Decision and Control, volume 2, pages 1239–1244, Brighton,

December 1991.

[Mak91] P. M¨ akil¨ a. Multiple models, multiplicative noise and linear quadratic control-

algorithm aspects. Int. J. Control, 54(4):921–941, October 1991.

[Mag76] E. F. Mageirou. Values and strategies in inﬁnite duration linear quadratic games.

IEEE Trans. Aut. Control, AC-21:547–550, 1976.

[McL71] P. J. McLane. Optimal stochastic control of linear systems with state- and

control-dependent disturbances. IEEE Trans. Aut. Control, 16(6):292–299, De-

cember 1971.

[Meg92a] A. Megretsky. S–procedure in optimal non-stochastic ﬁltering. Technical Report

TRITA/MAT-92-0015, Department of Mathematics, Royal Institute of Technol-

ogy, S-100 44 Stockholm, Sweden, March 1992.

[Meg92b] A. Megretsky. Power distribution approach in robust control. Technical Report

TRITA/MAT-92-0027, Department of Mathematics, Royal Institute of Technol-

ogy, S-100 44 Stockholm, Sweden, 1992.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 175

[Meg93] A. Megretsky. Necessary and suﬃcient conditions of stability: A multiloop

generalization of the circle criterion. IEEE Trans. Aut. Control, AC-38(5):753–

756, May 1993.

[Meh91] S. Mehrotra. On the implementation of a primal-dual interior point method.

SIAM Journal on Optimization, 2(4):575–601, November 1991.

[Mei91] A. M. Meilakhs. Construction of a Lyapunov function for parametrically per-

turbed linear systems. Automation and Remote Control, 52(10):1479–1481, Oc-

tober 1991.

[Mey66] K. R. Meyer. On the existence of Lyapunov functions for the problem of Lur’e.

J. SIAM Control, 3(3):373–383, 1966.

[MH78] P. J. Moylan and D. J. Hill. Stability criteria for large-scale systems. IEEE

Trans. Aut. Control, AC-23:143–149, 1978.

[Mol87] A. P. Molchanov. Lyapunov functions for nonlinear discrete-time control sys-

tems. Automation and Remote Control, pages 728–736, 1987.

[MP86] A. P. Molchanov and E. S. Pyatnitskii. Lyapunov functions that specify nec-

essary and suﬃcient conditions of absolute stability of nonlinear nonstationary

control system Pt. I–III. Automation and Remote Control, 46(3,4,5):(344–354),

(443–451), (620–630), 1986.

[MP89] A. P. Molchanov and E. S. Pyatnitskii. Criteria of asymptotic stability of dif-

ferential and diﬀerence inclusions encountered in control theory. Syst. Control

Letters, 13:59–64, 1989.

[MR76] C. T. Mullis and R. A. Roberts. Synthesis of minimum roundoﬀ noise ﬁxed-point

digital ﬁlters. IEEE Trans. Circuits Syst., 23:551–562, September 1976.

[MV63] S. G. Margolis and W. G. Vogt. Control engineering applications of V. I. Zubov’s

construction procedure for Lyapunov functions. IEEE Trans. Aut. Control, AC-

8:104–113, April 1963.

[Nem94] A. Nemirovskii. Several NP-Hard problems arising in robust stability analysis.

Mathematics of Control, Signals, and Systems, 6(1):99–105, 1994.

[Nev19] R. Nevanlinna.

¨

Uber beschr¨ ankte funktionen die in gegebenen punkten

vorgeschreibene funktionswerte bewirkt werden. Anal. Acad. Sci. Fenn., Ser.

A, 13:1–71, 1919.

[NF92] B. Nekooie and M. K. H. Fan. A quadratically convergent local algorithm on

minimizing sums of the largest eigenvalues of a symmetric matrix. Submitted to

Computational Optimization and Applications., October 1992.

[NN88] Yu. Nesterov and A. Nemirovsky. A general approach to polynomial-time algo-

rithms design for convex programming. Technical report, Centr. Econ. & Math.

Inst., USSR Acad. Sci., Moscow, USSR, 1988.

[NN90a] Yu. Nesterov and A. Nemirovsky. Optimization over positive semideﬁnite ma-

trices: Mathematical background and user’s manual. USSR Acad. Sci. Centr.

Econ. & Math. Inst., 32 Krasikova St., Moscow 117418 USSR, 1990.

[NN90b] Yu. Nesterov and A. Nemirovsky. Self-concordant functions and polynomial time

methods in convex programming. Technical report, Centr. Econ. & Math. Inst.,

USSR Acad. Sci., Moscow, USSR, April 1990.

[NN91a] Yu. Nesterov and A. Nemirovsky. Conic formulation of a convex programming

problem and duality. Technical report, Centr. Econ. & Math. Inst., USSR

Academy of Sciences, Moscow USSR, 1991.

[NN91b] Yu. Nesterov and A. Nemirovsky. An interior point method for generalized linear-

fractional programming. Submitted to Math. Programming, Series B., 1991.

[NN93] Yu. E. Nesterov and A. S. Nemirovskii. Acceleration of the path-following

method for optimization over the cone of positive semideﬁnite matrices.

Manuscript., 1993.

[NN94] Yu. Nesterov and A. Nemirovsky. Interior-point polynomial methods in convex

programming, volume 13 of Studies in Applied Mathematics. SIAM, Philadelphia,

PA, 1994.

This electronic version is for personal use and may not be duplicated or distributed.

176 Bibliography

[Nor86] J. P. Norton. An Introduction to Identiﬁcation. Academic Press, 1986.

[NS72] T. Nakamizo and Y. Sawaragi. Analytical study on n-th order linear system with

stochastic coeﬃcients. In R. F. Curtain, editor, Lecture notes on Mathematics:

Stability of stochastic dynamical systems, pages 173–185. Springer verlag, 1972.

[NT73] K. S. Narendra and J. H. Taylor. Frequency domain criteria for absolute stability.

Electrical science. Academic Press, New York, 1973.

[NW92] G. Naevdal and H. J. Woerdeman. Partial matrix contractions and intersections

of matrix balls. Linear Algebra and Appl., 175:225–238, 1992.

[NY83] A. Nemirovskii and D. Yudin. Problem Complexity and Method Eﬃciency in

Optimization. John Wiley & Sons, 1983.

[OC87] A. I. Ovseevich and F. L. Chernousko. On optimal ellipsoids approximating

reachable sets. Prob. of Control and Info. Th., 16(2):125–134, 1987.

[OF85] S. D. O’Young and B. A. Francis. Sensitivity tradeoﬀs for multivariable plants.

IEEE Trans. Aut. Control, AC-30(7):625–632, July 1985.

[OKUI88] Y. Ohashi, H. Kando, H. Ukai, and T. Iwazumi. Optimal design of FIR linear

phase ﬁlters via convex quadratic programming method. Int. J. Systems Sci.,

19(8):1469–1482, 1988.

[Ola94] A. Olas. Construction of optimal Lyapunov functions for systems with structured

uncertainties. IEEE Trans. Aut. Control, 39(1):167–171, January 1994.

[OMS93] E. Ohara, I. Masubuchi, and N. Suda. Quadratic stabilization and robust dis-

turbance attenuation using parametrization of stabilizing state feedback gains

– convex optimization approach. In Proc. American Control Conf., volume 3,

pages 2816–2820, San Francisco, California, June 1993.

[OS93] A. Olas and S. So. Application of the simplex method to the multi-variable sta-

bility problem. In Proc. American Control Conf., pages 1680–1681, San Fransico,

June 1993.

[OS94] A. Ohara and T. Sugie. Control systems synthesis via convex optimization.

Systems, Control and Information, 38(3):139–146, March 1994. Special issue on

Numerical Approaches in Control Theory. In Japanese.

[OSA93] A. Ohara, N. Suda, and A. Amari. Dualistic diﬀerential geometry of positive

deﬁnite matrices and its application to optimizations and matrix completion

problems. In MTNS-93, 1993.

[OV92] D. Obradovic and L. Valavani. Stability and performance in the presence of mag-

nitude bounded real uncertainty: Riccati equation based state space approaches.

Automatica, July 1992.

[Ove88] M. Overton. On minimizing the maximum eigenvalue of a symmetric matrix.

SIAM J. on Matrix Analysis and Applications, 9(2):256–268, 1988.

[Ove92] M. Overton. Large-scale optimization of eigenvalues. SIAM J. Optimization,

pages 88–120, 1992.

[OW92] M. Overton and R. Womersley. On the sum of the largest eigenvalues of a

symmetric matrix. SIAM J. on Matrix Analysis and Applications, pages 41–45,

1992.

[OW93] M. Overton and R. Womersley. Optimality conditions and duality theory for

minimizing sums of the largest eigenvalues of symmetric matrices. Mathematical

Programming, 62:321–357, 1993.

[Pac94] A. Packard. Gain scheduling via linear fractional transformations. Syst. Control

Letters, 22:79–92, 1994.

[Pan89] E. Panier. On the need for special purpose algorithms for minimax eigenvalue

problems. Journal Opt. Theory Appl., 62(2):279–287, August 1989.

[Par78] S. Parrott. On a quotient norm and the Sz.-Nagy-Foias lifting theorem. Journal

of Functional Analysis, 30:311–328, 1978.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 177

[PB84] I. Petersen and B. Barmish. The stabilization of single-input uncertain linear

systems via control. In Proc. 6th International Conference on Analysis and

Optimization of Systems, pages 69–83. springer-Verlag, 1984. Lecture notes in

control and information sciences, Vol. 62.

[PB92] A. Packard and G. Becker. Quadratic stabilization of parametrically-dependent

linear systems using parametrically-dependent linear, dynamic feedback. ASME

J. of Dynamic Systems, Measurement and Control, July 1992. Submitted.

[PBG89] P. L. D. Peres, J. Bernussou, and J. C. Geromel. A cutting plane technique

applied to robust control synthesis. In IFAC workshop on Control Applications

of Nonlinear Programming and Optimization, Paris, 1989.

[PBPB93] A. Packard, G. Becker, D. Philbrick, and G. Balas. Control of parameter de-

pendent systems: applications to H∞ gain scheduling. In 1993 IEEE Regional

Conference on Aerospace Control Systems, pages 329–333, May 1993.

[PD90] A. Packard and J. C. Doyle. Quadratic stability with real and complex pertur-

bations. IEEE Trans. Aut. Control, AC-35(2):198–201, February 1990.

[PD93] A. Packard and J. Doyle. The complex structured singular value. Automatica,

29(1):71–109, 1993.

[Pet85] I. R. Petersen. Quadratic stabilizability of uncertain linear systems: existence of

a nonlinear stabilizing control does not imply existence of a stabilizing control.

IEEE Trans. Aut. Control, AC-30(3):291–293, 1985.

[Pet87a] I. R. Petersen. Disturbance attenuation and H∞ optimization: a design method

based on the algebraic Riccati equation. IEEE Trans. Aut. Control, AC-

32(5):427–429, May 1987.

[Pet87b] I. R. Petersen. A stabilization algorithm for a class of uncertain systems. Syst.

Control Letters, 8(4):351–357, 1987.

[Pet88] I. R. Petersen. Quadratic stabilizability of uncertain linear systems containing

both constant and time-varying uncertain parameters. Journal of Optimization

Theory and Applications, 57(3):439–461, June 1988.

[Pet89] I. R. Petersen. Complete results for a class of state-feedback disturbance atten-

uation problems. IEEE Trans. Aut. Control, AC-34(11):1196–1199, November

1989.

[PF92] S. Phoojaruenchanachai and K. Furuta. Memoryless stabilization of uncertain

linear systems including time-varying state delays. IEEE Trans. Aut. Control,

37(7), July 1992.

[PG93] P. L. D. Peres and J. C. Geromel. H2 control for discrete-time systems, opti-

mality and robustness. Automatica, 29(1):225–228, 1993.

[PGB93] P. L. D. Peres, J. C. Geromel, and J. Bernussou. Quadratic stabilizability of

linear uncertain systems in convex-bounded domains. Automatica, 29(2):491–

493, 1993.

[PGS91] P. L. D. Peres, J. C. Geromel, and S. R. Souza. Convex analysis of discrete-time

uncertain H∞ control problems. In Proc. IEEE Conf. on Decision and Control,

Brighton, UK, December 1991.

[PGS93] P. L. D. Peres, J. C. Geromel, and S. R. Souza. H∞ robust control by static

output feedback. In Proc. American Control Conf., volume 1, pages 620–621,

San Francisco, California, June 1993.

[PH86] I. R. Petersen and C. V. Hollot. A Riccati equation approach to the stabilization

of uncertain linear systems. Automatica, 22(4):397–411, 1986.

[Phi89] Y. A. Phillis. Algorithms for systems with multiplicative noise. J. Contr. Dyn.

Syst., 30:65–81, 1989.

[Pic16] G. Pick.

¨

Uber die beschr¨ ankungen analytischer funktionen, welche durch

vorgegebene funktionswerte bewirkt werden. Math. Ann., 77:7–23, 1916.

[PKT

+

94] K. Poolla, P. Khargonekar, A. Tikku, J. Krause, and K. Nagpal. A time-domain

approach to model validation. IEEE Trans. Aut. Control, May 1994.

This electronic version is for personal use and may not be duplicated or distributed.

178 Bibliography

[PM92] I. R. Petersen and D. C. McFarlane. Optimal guaranteed cost control of uncer-

tain linear systems. In Proc. American Control Conf., Chicago, June 1992.

[PMR93] I. R. Petersen, D. C. McFarlane, and M. A. Rotea. Optimal guaranteed cost

control of discrete-time uncertain systems. In 12th IFAC World Congress, pages

407–410, Sydney, Australia, July 1993.

[Pon61] L. Pontryagin. Optimal Regulation Processes, volume 18. American Mathemat-

ical Society Translations, 1961.

[Pon62] L. Pontryagin. The Mathematical Theory of Optimal Processes. John Wiley &

Sons, 1962.

[Pop62] V. M. Popov. Absolute stability of nonlinear systems of automatic control.

Automation and Remote Control, 22:857–875, 1962.

[Pop64] V. M. Popov. One problem in the theory of absolute stability of controlled

systems. Automation and Remote Control, 25(9):1129–1134, 1964.

[Pop73] V. M. Popov. Hyperstability of Control Systems. Springer-Verlag, New York,

1973.

[Pos84] M. Post. Minimum spanning ellipsoids. In Proc. ACM Symp. Theory Comp.,

pages 108–116, 1984.

[Pot66] J. E. Potter. Matrix quadratic solutions. J. SIAM Appl. Math., 14(3):496–501,

May 1966.

[PP92] G. Picci and S. Pinzoni. On feedback-dissipative systems. Jour. of Math. Sys-

tems, Estimation and Control, 2(1):1–30, 1992.

[PR91a] E. S. Pyatnitskii and L. B. Rapoport. Existence of periodic motions and test for

absolute stability of nonlinear nonstationary systems in the three-dimensional

case. Automation and Remote Control, 52(5):648–658, 1991.

[PR91b] E. S. Pyatnitskii and L. B. Rapoport. Periodic motion and tests for absolute

stability of non-linear nonstationary systems. Automation and Remote Control,

52(10):1379–1387, October 1991.

[PR94] S. Poljak and J. Rohn. Checking robust nonsingularity is NP-complete. Mathe-

matics of Control, Signals, and Systems, 6(1):1–9, 1994.

[PS82] E. S. Pyatnitskii and V. I. Skorodinskii. Numerical methods of Lyapunov func-

tion construction and their application to the absolute stability problem. Syst.

Control Letters, 2(2):130–135, August 1982.

[PS83] E. S. Pyatnitskii and V. I. Skorodinskii. Numerical method of construction

of Lyapunov functions and absolute stability criteria in the form of numerical

procedures. Automation and Remote Control, 44(11):1427–1437, 1983.

[PS85] F. P. Preparata and M. I. Shamos. Computational Geometry, an Introduction.

Texts and monographs in Computer Science. Springer-Verlag, New York, 1985.

[PS86] E. S. Pyatnitskii and V. I. Skorodinskii. A criterion of absolute stability of

non-linear sampled-data control systems in the form of numerical procedures.

Automation and Remote Control, 47(9):1190–1198, August 1986.

[PSG92] P. L. D. Peres, S. R. Souza, and J. C. Geromel. Optimal H2 control for uncertain

systems. In Proc. American Control Conf., Chicago, June 1992.

[PW82] E. Polak and Y. Wardi. A nondiﬀerentiable optimization algorithm for the design

of control systems subject to singular value inequalities over a frequency range.

Automatica, 18(3):267–283, 1982.

[Pya68] E. S. Pyatnitskii. New research on the absolute stability of automatic control

(review). Automation and Remote Control, 29(6):855–881, June 1968.

[Pya70a] E. S. Pyatnitskii. Absolute stability of nonlinear pulse systems with nonsta-

tionary nonlinearity. Automation and Remote Control, 31(8):1242–1249, August

1970.

[Pya70b] E. S. Pyatnitskii. Absolute stability of nonstationary nonlinear systems. Au-

tomation and Remote Control, 31(1):1–11, January 1970.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 179

[Pya71] E. S. Pyatnitskii. Criterion for the absolute stability of second-order non-linear

controlled systems with one nonlinear nonstationary element. Automation and

Remote Control, 32(1):1–11, January 1971.

[PZP

+

92] A. Packard, K. Zhou, P. Pandey, J. Leonhardson, and G. Balas. Optimal, con-

stant I/O similarity scaling for full-information and state-feedback control prob-

lems. Syst. Control Letters, 19:271–280, October 1992.

[PZPB91] A. Packard, K. Zhou, P. Pandey, and G. Becker. A collection of robust control

problems leading to LMI’s. In Proc. IEEE Conf. on Decision and Control,

volume 2, pages 1245–1250, Brighton, December 1991.

[Rap86] L. B. Rapoport. Lyapunov stability and positive deﬁniteness of the quadratic

form on the cone. Prikladnaya matematika i mekhanika, 50(4):674–679, 1986.

In Russian; English translation in Applied Mathematics and Mechanics.

[Rap87] L. B. Rapoport. Absolute stability of control systems with several nonlinear

stationary elements. Automation and Remote Control, pages 623–630, 1987.

[Rap88] L. B. Rapoport. Sign-deﬁniteness of a quadratic form with quadratic constraints

and absolute stability of nonlinear control systems. Sov. Phys. Dokl., 33(2):96–

98, 1988.

[Rap89] L. B. Rapoport. Frequency criterion of absolute stability for control systems

with several nonlinear stationary elements. Automation and Remote Control,

50(6):743–750, 1989.

[Rap90] L. B. Rapoport. Boundary of absolute stability of nonlinear nonstationary sys-

tems and its connection with the construction of invariant functions. Automation

and Remote Control, 51(10):1368–1375, 1990.

[Rap93] L. B. Rapoport. On existence of non-smooth invariant functions on the boundary

of the absolute stability domain. Automation and Remote Control, 54(3):448–

453, 1993.

[RB92] E. Rimon and S. Boyd. Eﬃcient distance computation using best ellipsoid ﬁt.

Proceedings of the 1992 IEEE International Symposium on Intelligent Control,

pages 360–365, August 1992.

[RCDP93] M. A. Rotea, M. Corless, D. Da, and I. R. Petersen. Systems with structured

uncertainty: Relations between quadratic and robust stability. IEEE Trans.

Aut. Control, AC-38(5):799–803, 1993.

[Rei46] W. T. Reid. A matrix diﬀerential equation of the Riccati type. Amer. J. Math.,

68:237–246, 1946.

[Rin91] U. T. Ringertz. Optimal design of nonlinear shell structures. Technical Report

FFA TN 1991-18, The Aeronautical Research Institute of Sweden, 1991.

[RK88] M. A. Rotea and P. P. Khargonekar. Stabilizability of linear time-varying and

uncertain systems. IEEE Trans. Aut. Control, 33(9), September 1988.

[RK89] M. A. Rotea and P. P. Khargonekar. Stabilization of uncertain systems with

norm bounded uncertainty - a control Lyapunov function approach. SIAM J.

on Optimization, 27(6):1402–1476, November 1989.

[RK91] M. A. Rotea and P. P. Khargonekar. Generalized H2/H∞ control via convex

optimization. In Proc. IEEE Conf. on Decision and Control, volume 3, pages

2719–2720, Brighton, England, June 1991. Also in Proc. IMA Workshop on

Robust Control, University of Minnesota, 1993.

[RK93] M. A. Rotea and I. Kaminer. Generalized H2/H∞ control of discrete-time sys-

tems. In 12th IFAC World Congress, volume 2, pages 239–242, Sydney, Aus-

tralia, July 1993.

[Rob89] G. Robel. On computing the inﬁnity norm. IEEE Trans. Aut. Control, AC-

34(8):882–884, August 1989.

[Roc70] R. T. Rockafellar. Convex Analysis. Princeton Univ. Press, Princeton, second

edition, 1970.

[Roc82] R. T. Rockafellar. Convex Analysis and Optimization. Pitman, Boston, 1982.

This electronic version is for personal use and may not be duplicated or distributed.

180 Bibliography

[Roc93] R. T. Rockafellar. Lagrange multipliers and optimality. SIAM Review, 35:183–

283, 1993.

[Rot93] M. A. Rotea. The generalized H2 control problem. Automatica, 29(2):373–383,

1993.

[Rox65] E. Roxin. On generalized dynamical systems deﬁned by contingent equations.

Journal of Diﬀerential Equations, 1:188–205, 1965.

[RP93] M. A. Rotea and R. Prasanth. The ρ performance measure: A new tool for

controller design with multiple frequency domain speciﬁcations. Submitted to

Trans. Aut. Control, 1993.

[RP94] M. A. Rotea and R. Prasanth. The ρ performance measure: A new tool for con-

troller design with multiple frequency domain speciﬁcations. In Proc. American

Control Conf., Baltimore, 1994.

[RSH90] S. Richter and A. Scotteward Hodel. Homotopy methods for the solution of

general modiﬁed algebraic Riccati equations. In Proc. IEEE Conf. on Decision

and Control, pages 971–976, 1990.

[RVW93] F. Rendl, R. Vanderbei, and H. Wolkowocz. A primal-dual interior-point method

for the max-min eigenvalue problem. Technical report, University of Waterloo,

Dept. of Combinatorics and Optimization, Waterloo, Ontario, Canada, 1993.

[RW94a] M. A. Rotea and D. Williamson. Optimal realizations of ﬁnite wordlength dig-

ital controllers via aﬃne matrix inequalities. In Proc. American Control Conf.,

Baltimore, 1994.

[RW94b] M. A. Rotea and D. Williamson. Optimal synthesis of ﬁnite wordlength ﬁlters

via aﬃne matrix inequalities. IEEE Trans. Circuits Syst., 1994. To appear.

Abridged version in Proc. 31st Annual Allerton Conference on Communication,

Control, and Computing, Monticello, IL, pp. 505-514, 1993.

[Rya88] E. P. Ryan. Adaptive stabilization of a class of uncertain nonlinear systems: A

diﬀerential inclusion approach. Systems and Control Letters, 10:95–101, 1988.

[SA88] M. Saeki and M. Araki. Search method for the weighting matrix of the multi-

variable circle criterion. Int. J. Control, 47(6):1811–1821, 1988.

[Sae86] M. Saeki. A method of robust stability analysis with highly structured uncer-

tainties. IEEE Trans. Aut. Control, AC-31(10):935–940, October 1986.

[Saf80] M. G. Safonov. Stability and Robustness of Multivariable Feedback Systems. MIT

Press, Cambridge, 1980.

[Saf82] M. G. Safonov. Stability margins of diagonally perturbed multivariable feedback

systems. IEE Proc., 129-D:251–256, 1982.

[Saf86] M. G. Safonov. Optimal diagonal scaling for inﬁnity-norm optimization. Syst.

Control Letters, 7:257–260, 1986.

[Sam59] J. C. Samuels. On the mean square stability of random linear systems. Trans.

IRE, PGIT-5(Special supplement):248–, 1959.

[San64] I. W. Sandberg. A frequency-domain condition for the stability of feedback

systems containing a single time-varying nonlinear element. Bell Syst. Tech. J.,

43(3):1601–1608, July 1964.

[San65a] I. W. Sandberg. On the boundedness of solutions of non-linear integral equations.

Bell Syst. Tech. J., 44:439–453, 1965.

[San65b] I. W. Sandberg. Some results in the theory of physical systems governed by

nonlinear functional equations. Bell Syst. Tech. J., 44:871–898, 1965.

[Sav91] A. V. Savkin. Absolute stability of nonlinear control systems with nonstationary

linear part. Automation and Remote Control, 41(3):362–367, 1991.

[SC85] A. Steinberg and M. Corless. Output feedback stabilization of uncertain dynam-

ical systems. IEEE Trans. Aut. Control, AC-30(10):1025–1027, October 1985.

[SC91] A. G. Soldatos and M. Corless. Stabilizing uncertain systems with bounded

control. Dynamics and Control, 1:227–238, 1991.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 181

[SC93] M. G. Safonov and R. Y. Chiang. Real/complex Km-synthesis without curve

ﬁtting. In C. T. Leondes, editor, Control and Dynamic Systems, volume 56,

pages 303–324. Academic Press, New York, 1993.

[Sch68] F. C. Schweppe. Recursive state estimation: Unknown but bounded errors and

system inputs. IEEE Trans. Aut. Control, AC-13(1):22–28, 1968.

[Sch73] F. C. Schweppe. Uncertain dynamic systems. Prentice-Hall, Englewood Cliﬀs,

N.J., 1973.

[SD83] M. G. Safonov and J. Doyle. Optimal scaling for multivariable stability margin

singular value computation. In Proceedings of MECO/EES 1983 Symposium,

1983.

[SD84] M. G. Safonov and J. Doyle. Minimizing conservativeness of robust singular

values. In S. G. Tzafestas, editor, Multivariable Control, pages 197–207. D.

Reidel, 1984.

[Sd86] A. Sideris and R. deGaston. Multivariable stability margin calculation with

uncertain correlated parameters. In Proceedings of 25th Conference on Decision

and Control, pages 766–771. IEEE, December 1986.

[SGL94] M. G. Safonov, K. C. Goh, and J. H. Ly. Control system synthesis via bilinear

matrix inequalities. In Proc. American Control Conf., 1994.

[SH68] A. W. Starr and Y. C. Ho. Nonzero-sum diﬀerential games. Technical Report

564, Harvard University - Division of Engineering and Applied Physics, Cam-

bridge, Massachusetts, 1968.

[Sha85] A. Shapiro. Extremal problems on the set of nonnegative deﬁnite matrices.

Linear Algebra and Appl., 67:7–18, 1985.

[Sho85] N. Z. Shor. Minimization Methods for Non-diﬀerentiable Functions. Springer

Series in Computational Mathematics. Springer-Verlag, Berlin, 1985.

[SI93] R. E. Skelton and T. Iwasaki. Liapunov and covariance controllers. Int. J.

Control, 57(3):519–536, March 1993.

[SIG93] R. E. Skelton, T. Iwasaki, and K. M. Grigoriadis. A uniﬁed approach to linear

control design, 1993. Manuscript.

[Sil69] D. D. Siljak. Nonlinear Systems: Parameter Analysis and Design. John Wiley

& Sons, New York, 1969.

[Sil71] D. D. Siljak. New algebraic criteria for positive realness. J. Franklin Inst.,

291(2):109–118, February 1971.

[Sil73] D. D. Siljak. Algebraic criteria for positive realness relative to the unit circle.

J. Franklin Inst., 296(2):115–122, August 1973.

[Sil89] D. D. Siljak. Parameter space methods for robust control design: A guided tour.

IEEE Trans. Aut. Control, 34(7):674–688, July 1989.

[Sko90] V. I. Skorodinskii. Iterational method of construction of Lyapunov-Krasovskii

functionals for linear systems with delay. Automation and Remote Control,

51(9):1205–1212, 1990.

[Sko91a] V. I. Skorodinskii. Iterational methods of testing frequency domain criteria of

absolute stability for continuous control systems. I. Automation and Remote

Control, 52(7):941–945, 1991.

[Sko91b] V. I. Skorodinskii. Iterational methods of veriﬁcation for frequency domain

criteria of absolute stability for continuous control systems. II. Automation and

Remote Control, 52(8):1082–1088, 1991.

[SKS93] W. Sun, P. Khargonekar, and D. Shim. Controller synthesis to render a closed

loop transfer function positive real. In Proc. American Control Conf., volume 2,

pages 1074–1078, San Francisco, California, June 1993.

[SL93a] M. G. Safonov and P. H. Lee. A multiplier method for computing real mul-

tivariable stability margin. In IFAC World Congress, Sydney, Australia, July

1993.

This electronic version is for personal use and may not be duplicated or distributed.

182 Bibliography

[SL93b] T. J. Su and P. L. Liu. Robust stability for linear uncertain time-delay systems

with delay-dependence. Int. J. Systems Sci., 24(6):1067–1080, 1993.

[SLC94] M. G. Safonov, J. Ly, and R. Y. Chiang. On computation of multivariable

stability margin using generalized Popov multiplier—LMI approach. In Proc.

American Control Conf., Baltimore, June 1994.

[Son83] E. Sontag. A Lyapunov-like characterization of asymptotic controllability. SIAM

J. Contr. Opt., 21:462–471, May 1983.

[Son88] G. Sonnevend. New algorithms in convex programming based on a notion of

‘centre’ (for systems of analytic inequalities) and on rational extrapolation. In-

ternational Series of Numerical Mathematics, 84:311–326, 1988.

[Son90] E. D. Sontag. Mathematical Control Theory, volume 6 of Texts in Applied Math-

ematics. Springer-Verlag, 1990.

[SP89] A. Sideris and R. S. S. Pe˜ na. Fast computation of the multivariable stability

margin for real interrelated uncertain parameters. IEEE Trans. Aut. Control,

AC-34(12):1272–1276, December 1989.

[SP94a] M. G. Safonov and G. P. Papavassilopoulos. The diameter of an intersection of

ellipsoids and BMI robust synthesis. In Symposium on Robust Control Design.

IFAC, 1994.

[SP94b] A. V. Savkin and I. Petersen. A connection between H

∞

control and the absolute

stabilizability of uncertain systems. Syst. Control Letters, 1994. To appear.

[SP94c] A. V. Savkin and I. Petersen. A method for robust stabilization related to the

Popov stability criterion. Technical report, Dept. of Elec. Engineering, Aus-

tralian Defence Force Academy, Campbell 2600, Australia, 1994.

[SP94d] A. V. Savkin and I. Petersen. Minimax optimal control of uncertain systems

with structured uncertainty. Int. J. Robust and Nonlinear Control, 1994. To

appear.

[SP94e] A. V. Savkin and I. Petersen. Nonlinear versus linear control in the absolute

stabilizability of uncertain systems with structured uncertainty. IEEE Trans.

Aut. Control, 1994. Submitted.

[SP94f] A. V. Savkin and I. Petersen. Output feedback ultimate boundedness control

for uncertain dynamic systems with structured uncertainty. Technical report,

Dept. of Elec. Engineering, Australian Defence Force Academy, Campbell 2600,

Australia, 1994.

[SP94g] A. V. Savkin and I. Petersen. Time-varying problems of H

∞

control and absolute

stabilizability of uncertain systems. Technical report, Dept. of Elec. Engineering,

Australian Defence Force Academy, Campbell 2600, Australia, 1994. Submitted

to 1994 ACC.

[SP94h] A. V. Savkin and I. Petersen. An uncertainty averaging approach to optimal

guaranteed cost control of uncertain systems with structured uncertainty. Tech-

nical report, Dept. of Elec. Engineering, Australian Defence Force Academy,

Campbell 2600, Australia, 1994.

[SRC93] S. M. Swei, M. A. Rotea, and M. Corless. System order reduction in robust

stabilization problems. Int. J. Control, 1993. To appear; also in 12th IFAC

World Congress, Australia 1993, (6)27-30, July, 1993.

[SS72] F. M. Schlaepfer and F. C. Schweppe. Continuous-time state estimation under

disturbances bounded by convex sets. IEEE Trans. Aut. Control, AC-17(2):197–

205, 1972.

[SS81] T. N. Sharma and V. Singh. On the absolute stability of multivariable discrete-

time nonlinear systems. IEEE Trans. Aut. Control, AC-26(2):585–586, 1981.

[SS90] G. C. Sabin and D. Summers. Optimal technique for estimating the reachable set

of a controlled n-dimensional linear system. Int. J. Systems Sci., 21(4):675–692,

1990.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 183

[SSHJ65] D. G. Schultz, F. T. Smith, H. C. Hsieh, and C. D. Johnson. The generation

of Liapunov functions. In C. T. Leondes, editor, Advances in Control Systems,

volume 2, pages 1–64. Academic Press, New York, New York, 1965.

[SSL93] E. Sch¨ omig, M. Sznaier, and U. L. Ly. Minimum control eﬀort state-feedback

H∞-control. In Proc. American Control Conf., volume 1, pages 595–599, San

Francisco, California, June 1993.

[ST90] A. Stoorvogel and H. Trentelman. The quadratic matrix inequality in singular

H∞ control with state feedback. SIAM J. Control and Optimization, 28(5):1190–

1208, September 1990.

[Sto91] A. A. Stoorvogel. The robust H2 problem: A worst case design. In Conference

on Decision and Control, pages 194–199, Brighton, England, December 1991.

[Str66] R. L. Stratonovitch. A new representation of stochastic integrals and equations.

SIAM J. Control and Opt., pages 362–371, 1966.

[SW87] M. G. Safonov and G. Wyetzner. Computer-aided stability criterion renders

Popov criterion obsolete. IEEE Trans. Aut. Control, AC-32(12):1128–1131, De-

cember 1987.

[SZ92] R. E. Skelton and G. Zhu. Optimal L∞ bounds for disturbance robustness.

IEEE Trans. Aut. Control, AC-37(10):1568–1572, October 1992.

[Sze63] G. P. Szeg¨ o. On the absolute stability of sampled-data control systems. Proc.

Nat. Acad. Sci., 50:558–560, 1963.

[TB81] J. S. Thorp and B. R. Barmish. On guaranteed stability of uncertain linear

systems via linear control. Journal of Optimization Theory and Applications,

35(4):559–579, December 1981.

[Thi84] L. Thiele. State space discrete systems. Int. J. of Circuit Theory and Applica-

tions, 12, 1984.

[TM89] H. T. Toivonen and P. M. M¨ akil¨ a. Computer-aided design procedure for multi-

objective LQG control problems. Int. J. Control, 49(2):655–666, February 1989.

[Toi84] H. T. Toivonen. A multiobjective linear quadratic Gaussian control problem.

IEEE Trans. Aut. Control, AC-29(3):279–280, March 1984.

[TSC92] J. A. Tekawy, M. G. Safonov, and R. Y. Chiang. Convexity property of the one-

sided multivariable stability margin. IEEE Trans. Aut. Control, AC-37(4):496–

498, April 1992.

[TSL92] J. A. Tekawy, M. G. Safonov, and C. T. Leondes. Robustness measures for

one-sided parameter uncertainties. Syst. Control Letters, 19(2):131–137, August

1992.

[Tsy64a] Ya. Tsypkin. Absolute stability of a class of nonlinear automatic sampled data

systems. Aut. Remote Control, 25:918–923, 1964.

[Tsy64b] Ya. Tsypkin. A criterion for absolute stability of automatic pulse systems with

monotonic characteristics of the nonlinear element. Sov. Phys. Doklady, 9:263–

266, 1964.

[Tsy64c] Ya. Tsypkin. Frequency criteria for the absolute stability of nonlinear sampled

data systems. Aut. Remote Control, 25:261–267, 1964.

[Tsy64d] Ya. Tsypkin. Fundamentals of the theory of non-linear pulse control systems.

In Proc. 2nd IFAC, pages 172–180, London, 1964. Butterworth & Co.

[TV91] A. Tesi and A. Vicino. Robust absolute stability of Lur’e control systems in

parameter space. Automatica, 27(1):147–151, January 1991.

[Uhl79] F. Uhlig. A recurring theorem about pairs of quadratic forms and extensions: A

survey. Linear Algebra and Appl., 25:219–237, 1979.

[Vav91] S. A. Vavasis. Nonlinear Optimization: Complexity Issues. Oxford University

Press, New York, 1991.

[VB93a] L. Vandenberghe and S. Boyd. A polynomial-time algorithm for determining

quadratic Lyapunov functions for nonlinear systems. In Eur. Conf. Circuit Th.

and Design, pages 1065–1068, 1993.

This electronic version is for personal use and may not be duplicated or distributed.

184 Bibliography

[VB93b] L. Vandenberghe and S. Boyd. Primal-dual potential reduction method for prob-

lems involving matrix inequalities. To be published in Math. Programming, 1993.

[Vid78] M. Vidyasagar. On matrix measures and convex Liapunov functions. J. Math.

Anal. and Appl., 62:90–103, 1978.

[Vid92] M. Vidyasagar. Nonlinear Systems Analysis. Prentice-Hall, Englewood Cliﬀs,

second edition, 1992.

[VV85] A. Vannelli and M. Vidyasagar. Maximal Lyapunov functions and domains of

attraction for autonomous nonlinear systems. Automatica, 21(1):69–80, 1985.

[Wat91] G. A. Watson. An algorithm for optimal 2 scaling of matrices. IMA Journal of

Numerical Analysis, 11:481–492, 1991.

[WB71] J. C. Willems and G. L. Blankenship. Frequency domain stability criteria for

stochastic systems. IEEE Trans. Aut. Control, AC-16(4):292–299, 1971.

[WBL92] M. Wang, D. Boley, and E. B. Lee. Robust stability and stabilization of time de-

lay systems in real parameter space. In Proc. American Control Conf., volume 1,

pages 85–86, Chicago, June 1992.

[Wei90] K. Wei. Quadratic stabilizability of linear systems with structural independent

time-varying uncertainties. IEEE Trans. Aut. Control, AC-35(3):268–277, March

1990.

[Wei94] A. Weinmann. Uncertain models and robust control. Springer-Verlag, 1994.

[Wen88] J. T. Wen. Time domain and frequency domain conditions for strict positive

realness. IEEE Trans. Aut. Control, 33:988–992, 1988.

[Wil69a] J. C. Willems. The Analysis of Feedback Systems, volume 62 of Research Mono-

graphs. MIT Press, 1969.

[Wil69b] J. C. Willems. Stability, instability, invertibility, and causality. SIAM J. Control,

7(4):645–671, November 1969.

[Wil70] J. L. Willems. Stability Theory of Dynamical Systems. John Wiley & Sons, New

York, 1970.

[Wil71a] J. C. Willems. The generation of Lyapunov functions for input-output stable

systems. SIAM J. Control, 9(1):105–134, February 1971.

[Wil71b] J. C. Willems. Least squares stationary optimal control and the algebraic Riccati

equation. IEEE Trans. Aut. Control, AC-16(6):621–634, December 1971.

[Wil72] J. C. Willems. Dissipative dynamical systems I: General theory. II: Linear sys-

tems with quadratic supply rates. Archive for Rational Mechanics and Analysis,

45:321–343, 1972.

[Wil73] J. L. Willems. The circle criterion and quadratic Lyapunov functions for stability

analysis. IEEE Trans. Aut. Control, 18:184–186, 1973.

[Wil74a] J. C. Willems. On the existence of a nonpositive solution to the Riccati equation.

IEEE Trans. Aut. Control, AC-19(5), October 1974.

[Wil74b] A. N. Willson Jr. A stability criterion for nonautonomous diﬀerence equations

with application to the design of a digital FSK oscillator. IEEE Trans. Circuits

Syst., 21:124–130, 1974.

[Wil76] J. C. Willems. Mechanisms for the stability and instability in feedback systems.

Proc. IEEE, 64:24–35, 1976.

[Wil79] A. S. Willsky. Digital Signal Processing and Control and Estimation Theory:

Points of Tangency, Areas of Intersection, and Parallel Directions. MIT Press,

1979.

[Wil70] A. L. Willson. New theorems on the equations of nonlinear DC transistor net-

works. Bell System technical Journal, 49:1713–1738, October 70.

[Woe90] H. J. Woerdeman. Strictly contractive and positive completions for block ma-

trices. Linear Algebra and Appl., 105:63–105, 1990.

[Won66] W. M. Wonham. Lyapunov criteria for weak stochastic stability. J. of Diﬀ.

Equations, 2:195–207, 1966.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Bibliography 185

[Won68] W. M. Wonham. On a matrix Riccati equation of stochastic control. SIAM J.

Control and Opt., 6:681–697, 1968.

[Wri92] M. Wright. Interior methods for constrained optimization. Acta Numerica, 1,

1992.

[WW83] J. L. Willems and J. C. Willems. Robust stabilization of uncertain systems.

SIAM J. Control and Opt., 21(3):352–373, May 1983.

[WXdS92] Y. Wang, L. Xie, and C. E. de Souza. Robust control of a class of uncertain

non-linear systems. Syst. Control Letters, 19:139–149, 1992.

[XdS92] L. Xie and C. E. de Souza. Robust H∞ control for linear systems with norm-

bounded time-varying uncertainty. IEEE Trans. Aut. Control, AC-37(8), August

1992.

[XFdS92] L. Xie, M. Fu, and C. E. de Souza. H∞ control and quadratic stabilization

of systems with parameter uncertainty via output feedback. IEEE Trans. Aut.

Control, 37(8):1253–1256, August 1992.

[Yak62] V. A. Yakubovich. The solution of certain matrix inequalities in automatic

control theory. Soviet Math. Dokl., 3:620–623, 1962. In Russian, 1961.

[Yak64] V. A. Yakubovich. Solution of certain matrix inequalities encountered in non-

linear control theory. Soviet Math. Dokl., 5:652–656, 1964.

[Yak66] V. A. Yakubovich. Frequency conditions for the existence of absolutely stable

periodic and almost periodic limiting regimes of control systems with many

nonstationary elements. In IFAC World Congress, London, 1966.

[Yak67] V. A. Yakubovich. The method of matrix inequalities in the stability theory

of nonlinear control systems, I, II, III. Automation and Remote Control, 25-

26(4):905–917, 577–592, 753–763, April 1967.

[Yak70] V. A. Yakubovich. On synthesis of optimal controls in a linear diﬀerential game

with quadratic payoﬀ. Soviet Math. Dokl., 11(6):1478–1481, 1970.

[Yak71] V. A. Yakubovich. On synthesis of optimal controls in a linear diﬀerential game

on a ﬁnite time interval with quadratic payoﬀ. Soviet Math. Dokl., 12(5):1454–

1458, 1971.

[Yak73] V. A. Yakubovich. Minimization of quadratic functionals under the quadratic

constraints and the necessity of a frequency condition in the quadratic criterion

for absolute stability of nonlinear control systems. Soviet Math. Doklady, 14:593–

597, 1973.

[Yak77] V. A. Yakubovich. The S-procedure in non-linear control theory. Vestnik

Leningrad Univ. Math., 4:73–93, 1977. In Russian, 1971.

[Yak82] V. A. Yakubovich. Frequency methods of qualitative investigation into nonlin-

ear systems. In A. Yu. Ishlinsky and F. L. Chernousko, editors, Advances in

Theoretical and Applied Mechanics, pages 102–124. MIR, Moscow, 1982.

[Yak88] V. A. Yakubovich. Dichotomy and absolute stability of nonlinear systems with

periodically nonstationary linear part. Syst. Control Letters, 1988.

[Yak92] V. A. Yakubovich. Nonconvex optimization problem: The inﬁnite-horizon linear-

quadratic control problem with quadratic constraints. Syst. Control Letters, July

1992.

[Yan92] Q. Yang. Minimum Decay Rate of a Family of Dynamical Systems. PhD thesis,

Stanford University, Stanford, CA 94305, January 1992.

[Yaz93] E. Yaz. Stabilizing compensator design for uncertain nonlinear systems. Syst.

Control Letters, 21(1):11–17, 1993.

[Yos94] A. Yoshise. An optimization method for convex programs—interior-point

method and analytical center. Systems, Control and Information, 38(3):155–

160, March 1994. Special issue on Numerical Approaches in Control Theory. In

Japanese.

[Zam66a] G. Zames. On the input-output stability of time-varying nonlinear feedback

systems—Part I: Conditions derived using concepts of loop gain, conicity, and

positivity. IEEE Trans. Aut. Control, AC-11:228–238, April 1966.

This electronic version is for personal use and may not be duplicated or distributed.

186 Bibliography

[Zam66b] G. Zames. On the input-output stability of time-varying nonlinear feedback

systems—Part II: Conditions involving circles in the frequency plane and sector

nonlinearities. IEEE Trans. Aut. Control, AC-11:465–476, July 1966.

[ZD63] L. A. Zadeh and C. A. Desoer. Linear System Theory: The State Space Approach.

McGraw-Hill, New York, 1963. Reprinted by Krieger, 1979.

[ZF67] G. Zames and P. L. Falb. On the stability of systems with monotone and odd

monotone nonlinearities. IEEE Trans. Aut. Control, 12(2):221–223, April 1967.

[ZF68] G. Zames and P. L. Falb. Stability conditions for systems with monotone and

slope-restricted nonlinearities. SIAM J. Control, 6(1):89–108, July 1968.

[ZF83] G. Zames and B. A. Francis. Feedback, minimax sensitivity and optimal robust-

ness. IEEE Trans. Aut. Control, AC-28(5):585–601, May 1983.

[Zin90] A. S. I. Zinober, editor. Deterministic control of uncertain systems. Peter Pere-

grinus Ltd., London, 1990.

[ZK87] K. Zhou and P. P. Khargonekar. Stability robustness bounds for linear state-

space models with structured uncertainty. IEEE Trans. Aut. Control, AC-

32(7):621–623, July 1987.

[ZK88] K. Zhou and P. P. Khargonekar. Robust stabilization of linear systems with

norm-bounded time-varying uncertainty. Syst. Control Letters, 10:17–20, 1988.

[ZKSN92] K. Zhou, P. P. Khargonekar, J. Stoustrup, and H. H. Niemann. Robust stability

and performance of uncertain systems in state space. In Proc. IEEE Conf. on

Decision and Control, pages 662–667, Tucson, December 1992.

[ZRS93] G. Zhu, M. A. Rotea, and R. E. Skelton. A convergent feasible algorithm for the

output covariance constraint problem. In Proc. American Control Conf., pages

1675–1679, San Fransico, June 1993.

[Zub55] V. I. Zubov. Problems in the theory of the second method of Lyapunov, construc-

tion of the general solution in the domain of asymptotic stability. Prikladnaya

Matematika i Mekhanika, 19:179–210, 1955.

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Index

A

Absolute stability, 54, 56, 72

Acronyms

apology for too many, 5

list, 159

Aﬃne

family of transfer matrices, 141

matrix inequality, 27

Algebraic Riccati equation, 3, 26, 110, 115

Algorithm

conjugate-gradients, 18

ellipsoid, 49

for Lyapunov inequalities, 18

interior-point, 30

Karmarkar’s, 4

method of centers, 16, 30

Nesterov and Nemirovskii’s, 4

Newton, 15

projective, 30

Almost sure convergence, 131

system with multiplicative noise, 136

α-shifted H∞ norm, 67

Analytic center, 15

ellipsoidal approximation, 45

ellipsoidal bound, 49

Analytic solution, 2, 62, 115

LMI, 24

Approximation

ellipsoidal, 49

of PLDI, 58

ARE, 3, 26, 110, 115

Augmented system, 124

B

Barrier function, 15

Bellman function, 112

Bounded-real lemma, 26

Brownian motion, 137

C

Central path, 16

Cheap ellipsoid approximation, 50

Circle criterion, 2

Closed-loop system

dynamic state-feedback, 116

static state-feedback, 99

Youla parametrization, 152

Co, 12, 51

Complement

orthogonal, 22

Schur, 7

Completion problem, 40, 48

Complexity

convex optimization, 18, 29

S-procedure, 24

stabilization problem, 128

Componentwise results

using scaling, 111, 134

via scaling, 94

Componentwise unit-energy input

LDI, 81, 96

state-feedback, 106

Concave quadratic programming, 42

Condition number, 37

coordinate transformation, 65, 103

Conjugate-gradients, 18

Constrained optimal control problem, 151

Constraint

equality, 9, 19

integral quadratic vs. pointwise, 96,

122

on control input, 103

qualiﬁcation, 19

Contractive completion problem, 48

Control input, 52

norm constraint, 103

Controllability Gramian, 78

Controller

dynamic feedback, 111

gain-scheduled, 118

order, 111

reduced-order, 117

state-feedback, 99

Convex

Lyapunov function, 74

Convex function

LMI representation, 29

Lyapunov, 64

Convex hull, 12, 51, 158

Convex optimization

complexity, 18, 29

duality, 5

ellipsoid algorithm, 12

interior-point method, 14

problem structure, 18

software, 31

Convex problem, 11

187

188 Index

Convolution, 143

Coordinate transformation

condition number, 65

digital ﬁlter realization, 148

state-feedback, 103

CP, 11

Cutting plane, 12

D

Decay rate, 66

Delay, 144

DI, 51

Diagonal

Lyapunov function, 40, 47, 144

norm-bound LDI, 54

Diameter, 69

invariant ellipsoid of LDI, 80

NLDI, 59

Diﬀerential inclusion, 51

selector-linear, 56

Dissipativity

LDI, 93

state-feedback, 110

DNLDI, 54

well-posedness, 57

Duality, 5, 9, 29, 49

Dynamic state-feedback, 111

E

c, 11

Eigenvalue problem, 10

Elimination

in LMI, 22

of variable, 9

procedure, 48

Ellipsoid

algorithm, 12, 29, 49

approximating intersection, 44

approximating polytope, 42

approximating reachable set, 78, 104,

106, 121, 125, 132

approximating sum, 46

approximating union, 43

approximation, 49

diameter, 69, 80

extractable energy, 87, 109

holdable, 102

invariant, 68

L¨ owner–John, 44, 49

minimum volume, 11, 44

poor man’s, 50

via analytic center, 45, 49

Entropy, 142

Equality constraint, 19

Euclidean norm, 8

EVP, 10

Exogenous input, 52

E, 5

Expected output energy

LDI, 87

state-feedback, 109

Exponential time-weighting, 84, 97

Extractable energy

LDI, 87

state-feedback, 109

F

Fading memory, 74

Family of transfer matrices

examples, 152

parametrization, 143

Feedback

diagonal norm-bound, 54

nonexpansive, 123

structured, 54

time-varying, 53

Finsler’s lemma, 22

Frequency measurements, 146

Function

quasiconvex, 29

value, 112

G

Gain

L2 or RMS, 91

state-feedback, 99

Gain-scheduled controller, 118

γ-entropy, 142

Generalized eigenvalue problem, 10

GEVP, 10

Global linearization, 54

state-feedback, 99

Gramian

controllability, 78

observability, 85

H

Half-space

cutting plane, 12

Hamiltonian matrix, 26

Hankel norm

LDI, 89

transfer matrix, 142

H∞ norm, 27, 91, 142

Holdability, 102

Holdable ellipsoid

approximating reachable set, 106

LDI, 102

output energy, 108

H2 norm, 141

H2–H∞ problem, 154

I

Identiﬁcation, 146

Iγ, 142

Implicit equality constraint, 19

Impulse response, 98

Inequality

return-diﬀerence, 153

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Index 189

Infeasible LMI, 29

Initial condition

extractable energy, 85, 108

safe, 71

Input

control or exogenous, 52

unit-energy, 77, 104, 121, 125, 132

unit-peak, 82, 107

Input-to-output properties

LDI, 89

Lur’e system, 122

state-feedback, 109

system with multiplicative noise, 133

Input-to-state properties

ﬁlter realization, 148

LDI, 77

Lur’e system, 121

state-feedback, 104

system with multiplicative noise, 132

system with unknown parameters, 125

Integral quadratic constraint, 96, 122

Interior-point method, 4, 14, 30

primal-dual, 18

Interval matrix, 40

Invariance

of positive orthant, 152

Invariant ellipsoid

approximating reachable set, 78, 104,

106

extractable energy, 87

for LDI, 68

output energy, 108

Inverse problem of optimal control, 147

Itˆ o stochastic system, 136

K

κ, 37

Kalman-Yakubovich-Popov lemma, 2

Karmarkar’s algorithm, 4

L

L¨ owner–John ellipsoid, 49

LDI

better name, 56

componentwise unit-energy input, 81

decay rate, 66

deﬁnition, 51

diagonal norm-bound, 54

dissipativity, 93, 110

expected output energy, 87

Hankel norm, 89

invariant ellipsoid, 68

L2 gain, 91, 109

norm-bound, 53

output energy, 85, 108

output peak, 88

polytope, 53

positive orthant stabilizability, 144

quadratically stabilizable, 100

quadratic stability, 61

reachable set, 77, 104

return time, 71

scaling, 94, 111

stability, 61

stability margin, 65

stable, not quadratically stable, 64,

73

state-feedback, 99

unit-energy input, 77

unit-peak input, 82

Lemma

bounded-real, 26

Finsler, 22

PR, 25, 93

Linear

inequality, 7

matrix inequality, 7

program, 10

Linear diﬀerential inclusion, 51

Linear-fractional

mapping, 53

programming, 11

Linearization

global, 54

Linear-Quadratic Regulator, 108

Line-search, 28

LMI

analytic center, 15

analytic solution, 24

convex function representation, 29

deﬁnition, 7

equality constraint, 9

father, 4

feasibility problem, 9

grandfather, 4

graphical condition, 2

history, 2

infeasible, 29

interior-point method, 4, 14

multiple, 7

nonstrict, 7, 18

optimization, 30

reduction, 19, 31

Riccati equation solution, 3

semideﬁnite terms, 22

slack variable, 8

software, 31

standard problems, 9

strict, 18

LMI-Lab, 31

LMIP, 9

LMI-tool, 31

Loop transformation, 119

sector condition, 129

L¨ owner–John ellipsoid, 44

LP, 10

LQG

multi-criterion, 150

system with multiplicative noise, 138

LQR, 108, 114

This electronic version is for personal use and may not be duplicated or distributed.

190 Index

LTI system

deﬁnition, 52

input-to-state properties, 78

overshoot, 97

realization, 148

stability, 62

transfer matrix, 91

with delay, 144

with nonparametric uncertainty, 143

L2 gain

LDI, 91

LTI system, 142

Lur’e system, 122

scaled, 94

state-feedback, 109

system with multiplicative noise, 133

Lur’e resolving equation, 3

Lur’e system

absolute stability, 56

deﬁnition, 119

history, 126

Lur’e term

in Lyapunov function, 119

Lyapunov

equation, 2, 25, 62, 78, 85

exponent, 66

functional, 144

inequality, 2, 8

Lyapunov function

convex, 64, 74

diagonal, 40, 47, 144

integral quadratic constraints, 123, 127

Lur’e, 119

parameter-dependent, 119

quadratic, 61

stochastic, 132, 136

Lyapunov inequality

reduction to a strict one, 20

M

Margin

quadratic stability, 65

quadratic stabilizability, 113

system with multiplicative noise, 132

matlab, 31

Matrix

completion problem, 40

elimination of variable, 22, 48

Hamiltonian, 26

inequality, 7

interval, 40

linear-fractional mapping, 53

M, 75

Moore–Penrose inverse, 28

norm, 8

10, 57

pencil, 28

Pick, 146

problems, 37

quadratic inequality, 8

transfer, 25

Maximum singular value, 8

Mean-square stability

deﬁnition, 132

state-feedback, 135

Method of centers, 16, 30

Minimum size ellipsoid, 69

Minimum volume ellipsoid, 11

Mixed H2–H∞ problem, 154

M-matrix, 75

Moore–Penrose inverse, 28

µ-analysis, 152

Multi-criterion LQG problem, 96, 150, 153

Multiple LMI, 7

Multiplicative noise, 131

Multiplier

nonparametric uncertainty, 143

system with unknown parameters, 124

theory, 128

N

Nesterov and Nemirovskii’s algorithm, 4

Nevanlinna-Pick interpolation, 145

Nevanlinna-Pick theory, 146

Newton’s method, 15

NLDI

approximating PLDI, 58

deﬁnition, 53

input-to-state properties, 79

quadratic stability, 62

state-to-output properties, 84

well-posedness, 53, 63

Noise

in interpolation problems, 146

multiplicative, 131

Nonconvex quadratic problem

mixed LQR–L2 problem, 154

multi-criterion LQG, 151

Nonexpansive, 26

feedback, 123

LDI, 91

Nonlinear system

fading memory, 74

state-feedback, 99

Nonparametric uncertainty

in LTI system, 143

Nonstrict LMI, 7, 18

Norm

α-shifted H∞, 67

Euclidean, 8

Hankel, 89

H∞, 27, 91

H2, 141

matrix, 8

maximum singular value, 8

of LTI system, 148

quadratic approximation of piecewise

linear, 41

scaled singular value, 38

Norm-bound LDI, 53

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Index 191

Notation

diﬀerential equations, 5

table, 157

NP-complete, 42

NP-hard, 128

well-posedness of DNLDI, 54, 57

O

Observability Gramian, 85

Ω, 51

Optimal control, 112, 151

Optimization

ellipsoid algorithm, 12

interior-point method, 14

linear-fractional, 11

linear program, 10

LMI, 30

quasiconvex, 11

optin, 31

Order of controller, 111

Orthant

invariant, 152

stability, 144

Orthogonal complement, 22

Outer approximation of reachable set

state-feedback, 106

Output, 52

Output energy

LDI, 85

Lur’e system, 121

state-feedback, 108

Output peak

LDI, 88

Output variance for white noise input, 113,

138

Overﬂow, 148

Overshoot

LTI system, 97

P

10 matrix, 57

Parameter-dependent

controller, 118

Lyapunov function, 119

Parametric uncertainty, 124

scaling matrix problem, 39

Parametrization

family of transfer matrices, 143

Youla, 152

Passive, 25

Passivity, 3

dissipativity, 93

state-feedback, 110

Path of centers, 16

Pencil, 28

Penrose inverse, 28

Performance

LQG cost, 150

of LDI, 113

system with multiplicative noise, 133

Perturbation

LTI, 143

norm-bound, 53

stochastic, 137

structured, 39, 54

Pick matrix, 146

Piecewise linear norm, 41

PLDI, 53

approximation by NLDI, 58

input-to-state properties, 78

quadratic stability, 62

state-to-output properties, 84

Pointwise constraint, 96, 122

Polyhedron

invariant, 152

Polynomial-time, 12

Polytope, 12

approximated by invariant ellipsoid,

69

containing reachable set, 80

ellipsoid approximation, 42

extractable energy, 87, 109

LDI, 53

return time, 71

symmetric, 48

Polytopic LDI, 53

Polytopic norm, 41

Popov criterion, 2, 119

Positive-deﬁnite

completion, 40

matrix, 7

representable, 29

Positive orthant stabilizability, 144, 152

Positive-real

lemma, 25

transfer matrix, 25

Primal-dual interior-point method, 18

Procedure

elimination of variable, 22

S, 23

Projective algorithm, 30

PR lemma, 2, 25, 93

Q

Quadratic

Lyapunov function, 61

matrix inequality, 8, 114

Quadratic approximation

of polytopic norm, 41

of reachable set, 105

Quadratic programming

indeﬁnite, 42

Quadratic stability

invariant ellipsoid, 68

LDI, 61

margin, 65

nonlinear system, 74

vs. stability, 64

Quadratic stabilizability, 100

This electronic version is for personal use and may not be duplicated or distributed.

192 Index

Qualiﬁcation

constraint, 19

Quantization, 148

Quasiconvex

function, 29

problem, 11

R

Reachable set

LDI, 77

Lur’e system, 121

relaxed DI, 51

state-feedback, 104

system with multiplicative noise, 132

system with unknown parameters, 125

1uce, 81

1ue, 77

1up, 82

Realization of system, 148

Reduced-order controller, 117

Reduction

Lyapunov inequality, 20

strictly feasible LMI, 19

Regularity conditions, 4

Relaxed version of a DI, 51

Response

step or impulse, 97

Return-diﬀerence inequality, 153

Return time

LDI, 71

Riccati equation, 3, 26, 115

Riccati inequality, 110, 115

RMS gain

LDI, 91

state-feedback, 109

RMS norm, 148

Root-Mean-Square, 91

Routh–Hurwitz algorithm, 35

S

Safe initial condition, 71

Scaled L2 gain

LDI, 94

state-feedback, 111

system with multiplicative noise, 134

Scaled singular value, 38

Scaling

and S-procedure, 64

componentwise results, 94

for componentwise results, 111, 134

in matrix problem, 37

interpolation, 145

Schur complement

nonstrict, 28

strict, 7

scilab, 31

Sector condition, 56

loop transformation, 129

Selector-linear

diﬀerential inclusion, 56

Self-concordant barrier, 15

Simultaneous matrix completion, 40

Singular value plot, 27

Slack variable in LMIs, 8

Small-gain theorem, 3, 143

Software for LMI problems, 31

Sparsity pattern, 40

Sphere

smallest or largest, 44, 46

S-procedure

and scaling, 64

description, 23

Stability

absolute, 56, 72

degree, 67

LDI, 61

LTI system, 62

Lur’e system, 120

positive orthant, 144

system with multiplicative noise, 132

system with unknown parameters, 125

Stability margin

LDI, 65

mean-square, 132, 137

Stabilizability

LDI, 100

positive orthant, 144

system with multiplicative noise, 134

Stabilization problem

complexity, 128

Standard LMI problems, 9

∗, 143

State, 52

State-feedback

controller, 99

global linearization, 99

in linear system with delay, 145

LDI, 99

nonlinear system, 99

positive orthant stability, 144

scaling, 111

synthesis, 99

system with multiplicative noise, 134

State properties

LDI, 61

Lur’e system, 120

state-feedback, 100

system with multiplicative noise, 131

State-to-output properties

ﬁlter realization, 148

LDI, 84

Lur’e system, 121

state-feedback, 107

Static state-feedback, 99

Step

length, 28

response, 97

Stochastic stability margin, 132

Stochastic system

continuous-time, 136

Copyright c _ 1994 by the Society for Industrial and Applied Mathematics.

Index 193

discrete-time, 131

Stopping criterion

method of centers, 17

Stratonovitch stochastic system, 136

Strict feasibility

reduction to, 19

Strict LMI, 18

Structured

feedback, 54, 152

singular value, 152

Structured uncertainty

interpolation, 146

matrix problem, 39

Sturm method, 26, 35

Symmetric polytope, 48

Synthesis

output-feedback, 117, 150

state-feedback, 99, 134

System

realization, 148

stochastic, 136

uncertain, 56

with multiplicative noise, 131

System theory duality, 5

System with unknown parameters

matrix problem, 39

multipliers for, 124

T

Tangential Nevanlinna-Pick interpolation,

145

Time-varying

feedback matrix, 53

system, 51

Tr, 5

Trace, 5

Transfer matrix, 25

LTI system, 64, 91

optimization, 141

positive-real, 25

Transformation

loop, 119

Tsypkin criterion, 2

U

Uncertainty

LTI, 143

parametric, 124

random, 131

time-varying, 51

Unit-energy input

LDI, 77

Lur’e system, 121

state-feedback, 104

system with multiplicative noise, 132

system with unknown parameters, 125

Unit-intensity white noise, 138

Unit-peak input, 97

LDI, 82

state-feedback, 107

Unknown parameters

system with, 124

V

Value function, 112

Variable

elimination, 9

matrix, 8

slack, 8

Vertex

of polytope, 41

vol, 13

Volume

ellipsoid algorithm, 13

invariant ellipsoid, 69

of ellipsoid, 41

W

Well-posedness

DNLDI, 54, 57

NLDI, 53, 63

White noise

Brownian motion, 137

uncertainty, 131

Y

Youla parametrization, 152

This electronic version is for personal use and may not be duplicated or distributed.

**SIAM Studies in Applied Mathematics
**

This series of monographs focuses on mathematics and its applications to problems of current concern to industry, government, and society. These monographs will be of interest to applied mathematicians, numerical analysts, statisticians, engineers, and scientists who have an active need to learn useful methodology.

Series List

Vol. 1 Vol. 2 Vol. 3 Lie-B¨cklund Transformations in Applications a Robert L. Anderson and Nail H. Ibragimov Methods and Applications of Interval Analysis Ramon E. Moore Ill-Posed Problems for Integrodiﬀerential Equations in Mechanics and Electromagnetic Theory Frederick Bloom Solitons and the Inverse Scattering Transform Mark J. Ablowitz and Harvey Segur Fourier Analysis of Numerical Approximations of Hyperbolic Equations Robert Vichnevetsky and John B. Bowles Numerical Solution of Elliptic Problems Garrett Birkhoﬀ and Robert E. Lynch Analytical and Numerical Methods for Volterra Equations Peter Linz Contact Problems in Elasticity: A Study of Variational Inequalities and Finite Element Methods N. Kikuchi and J. T. Oden Augmented Lagrangian and Operator-Splitting Methods in Nonlinear Mechanics Roland Glowinski and P. Le Tallec Boundary Stabilization of Thin Plate Splines John E. Lagnese Electro-Diﬀusion of Ions Isaak Rubinstein Mathematical Problems in Linear Viscoelasticity Mauro Fabrizio and Angelo Morro Interior-Point Polynomial Algorithms in Convex Programming Yurii Nesterov and Arkadii Nemirovskii The Boundary Function Method for Singular Perturbation Problems Adelaida B. Vasil’eva, Valentin F. Butuzov, and Leonid V. Kalachev Linear Matrix Inequalities in System and Control Theory Stephen Boyd, Laurent El Ghaoui, Eric Feron, and Venkataramanan Balakrishnan

Vol. 4 Vol. 5 Vol. 6 Vol. 7 Vol. 8

Vol. 9

Vol. 10 Vol. 11 Vol. 12 Vol. 13 Vol. 14 Vol. 15

Stephen Boyd, Laurent El Ghaoui, Eric Feron, and Venkataramanan Balakrishnan

Linear Matrix Inequalities in System and Control Theory

**Society for Industrial and Applied Mathematics
**

®

Philadelphia

Copyright c 1994 by the Society for Industrial and Applied Mathematics. All rights reserved. No part of this book may be reproduced, stored, or transmitted in any manner without the written permission of the Publisher. For information, write to the Society for Industrial and Applied Mathematics, 3600 University City Science Center, Philadelphia, Pennsylvania 19104-2688.

The royalties from the sales of this book are being placed in a fund to help students attend SIAM meetings and other SIAM related activities. This fund is administered by SIAM and qualiﬁed individuals are encouraged to write directly to SIAM for guidelines.

Library of Congress Cataloging-in-Publication Data Linear matrix inequalities in system and control theory / Stephen Boyd . . . [et al.]. p. cm. -- (SIAM studies in applied mathematics ; vol. 15) Includes bibliographical references and index. ISBN 0-89871-334-X 1. Control theory. 2. Matrix inequalities. 3. Mathematical optimization. I. Boyd, Stephen P. II. Series: SIAM studies in applied mathematics : 15. QA402.3.L489 1994 515’.64--dc20 94-10477

2. . . . . .2 Some Standard Problems . . . . . . . . 3. . . . . . . . . . 3. . . . . . . . . . . . . . .Contents Preface Acknowledgments 1 Introduction 1. . . . . . . . . . . . . . . . . . . . . . . . . . .7 Ellipsoidal Approximation . . . . . . . . . . . . . 2. . . . . . . . . . . . . . . . Linear Diﬀerential Inclusions 4.2 Invariant Ellipsoids .2 Minimizing Condition Number of a Positive-Deﬁnite Matrix 3. . . . . . . . . . . . .6 Quadratic Approximation of a Polytopic Norm . . . . . 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . . . . .1 Quadratic Stability .2 A Brief History of LMIs in Control Theory 1. . . . . 4 5 Analysis of LDIs: State Properties 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Diﬀerential Inclusions . . . . . . . . . . . . . . . . . . . .7 Some LMI Problems with Analytic Solutions Notes and References . . . . . . . . 1. . . . . . . . . . . . . . .5 Matrix Completion Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Interior-Point Methods . . . . . . . . . . . .2 Some Speciﬁc LDIs . . . . . . .6 Miscellaneous Results on Matrix Inequalities 2. . 1. . 3. . . . 2 Some Standard Problems Involving LMIs 2. . . . . . . . . . . . . . . . .3 Ellipsoid Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Notes on the Style of the Book . . . . . . . . . . . vii ix 1 1 2 4 5 7 7 9 12 14 18 22 24 27 37 37 38 38 39 40 41 42 47 51 51 52 54 56 61 61 68 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v . . . .1 Linear Matrix Inequalities . . . . . . . . . . Notes and References . . . . . . . . . . . . . . 4. . . . 3 Some Matrix Problems 3. . . . . . . . . 4. . . . . . . . . . . . . . . . . . . . . . . . 3. . . . . . . . . . . . . . . . . . . . . . 5. . . . . . . . . . . .3 Nonlinear System Analysis via LDIs Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Origin of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Minimizing Condition Number by Scaling .5 Strict and Nonstrict LMIs . . . . . 2. .1 Overview . . . . . . .3 Minimizing Norm by Scaling .4 Rescaling a Matrix Positive-Deﬁnite . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Interpolation Problems . . . . . . . . . . . . . . 136 . . . . . . . . . . . . . . . . .2 State-to-Output Properties . . . . . . . . . . . . . . . Notes and References . . . . 7 State-Feedback Synthesis for LDIs 7. . . . . . . . . . 134 Notes and References . . . . . . . . . . . . . . 6. . 8. 10. . . . . . . . . . 8. . . . . . . . . . . . . . . . . . 8 9 Systems with Multiplicative Noise 131 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7. 6. . . . . . . . . . . .8 Multi-Criterion LQG . . . . 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10. . 141 141 143 144 144 145 147 148 150 151 152 157 159 161 187 10 Miscellaneous Problems 10. .3 Multipliers for Systems with Unknown Parameters Notes and References . .1 Static State-Feedback Controllers . . . . . . . . . .6 Observer-Based Controllers for Nonlinear Systems Notes and References . . . . . . . .6 The Inverse Problem of Optimal Control . . . . . 131 9. . . . . . . . . . . . . . . 10. . . . . . . . . . . . . . . . .2 State Properties . . . . . . . . . . . . . . . . . . . . . . . .2 Integral Quadratic Constraints . 10. . . . . . . . . . . . . . . . 7. . . . . . . .5 Input-to-Output Properties . . . . . . . 10. . . . . . . . . . . . . .1 Analysis of Lur’e Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Positive Orthant Stabilizability . . . . . . . . . . . . . . 7. . 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 State-Feedback Synthesis . . Notes and References . . . . . . . . . . . . . . .7 System Realization Problems . . . . . . . . . . . . . . . . .1 Input-to-State Properties . . . . . . . . . . . . 10. . . .4 Linear Systems with Delays . . . . . . . . . . . . . . . . . . .3 Input-to-State Properties . . . . . . . . . . . Notation List of Acronyms Bibliography Index Copyright c 1994 by the Society for Industrial and Applied Mathematics.vi Contents Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 State-to-Output Properties . . . . . .3 Input-to-Output Properties . . . . . . . . . . . . . . . . .1 Analysis of Systems with Multiplicative Noise . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Analysis of Systems with LTI Perturbations . . . . . . . . .9 Nonconvex Multi-Criterion Quadratic Problems . . . . . . . . . 72 77 77 84 89 96 99 99 100 104 107 109 111 112 119 119 122 124 126 6 Analysis of LDIs: Input/Output Properties 6. . . . . . . . . . . 10. . . . . . . . . . . . . . Lur’e and Multiplier Methods 8. . . . . . . . . . . . . . . . . . . . . . . . . .1 Optimization over an Aﬃne Family of Linear Systems 10. . . . . .

Therefore. the ellipsoid algorithm of Shor. The reader will soon see that their ideas and methods play a critical role in the basic idea presented in this book. and Convex Analysis and Minimization Algorithms I [HUL93] by Hiriart– Urruty and Lemar´chal. but our main point is that they can be solved numerically in all cases. Nonlinear Systems Analysis [Vid92] by Vidyasagar. For a few special cases there are “analytic solutions” to these problems. Optimal Control: Linear Quadratic Methods [AM90] by Anderson and Moore. These standard problems can be solved in polynomialtime (by.. and so are tractable. we hope that this book will later be considered as the ﬁrst book on the topic. e. at least in a theoretical sense. Sontag’s book Mathematical Control Theory [Son90] is an excellent survey. Although we believe that the methods described in this book have great practical value. and we make only brief comments about the implications of the results for practical control engineering. Recently developed interior-point methods for these standard problems have been found to be extremely eﬃcient in practice. vii . This is a research monograph: We present no speciﬁc examples or numerical results. To put it in a more positive light.g. The background required of the reader is knowledge of basic system and control theory and an exposure to optimization. but can also serve as a source of application problems for researchers in convex optimization. and Yudin). Nemirovskii.Preface The basic topic of this book is solving problems from system and control theory using convex optimization. Further background material is covered in the texts Linear Systems [Kai80] by Kailath. This book is primarily intended for the researcher in system and control theory. not the most readable or accessible. we consider the original problems from system and control theory as solved. we should warn the reader whose primary interest is applied control engineering. We show that a wide variety of problems arising in system and control theory can be reduced to a handful of standard convex and quasiconvex optimization problems that involve matrix inequalities. e We also highly recommend the book Interior-point Polynomial Algorithms in Convex Programming [NN94] by Nesterov and Nemirovskii as a companion to this book.

.

V. M. Nemirovskii for encouraging us to write this book. Vandenberghe. Petersen. J. This book was greatly improved by the suggestions of J. Dankoski. S. El Ghaoui and E. Savastiouk. Massachusetts College Park. Balakrishnan was supported in part by the Control and Dynamical Systems Group at Caltech and by NSF (under NSFD CDR-8803012). Abedor. Kosut. C. Feron was supported in part by the Charles Stark Draper Career Development Chair at MIT. The research reported in this book was supported in part by AFOSR (under F49620-92-J-0013). T. California Paris. and J. and many macros originally written by Craig Barratt. Kailath. A This book was typeset by the authors using L TEX. realization theory. M. France Cambridge. It is a special pleasure to thank V.Acknowledgments We thank A. I. Packard for introducing us to multiplier methods. We are grateful to S. G. respectively. Feron were supported in part by the D´l´gation G´n´rale pour ee e e l’Armement. L. R. Pyatnitskii. NSF (under ECS-9222391). Yakubovich for his comments and suggestions. M. Stephen Boyd Laurent El Ghaoui Eric Feron Venkataramanan Balakrishnan Stanford. and ARPA (under F49620-93-1-0085). Gevers. and state-feedback synthesis methods. Franklin. L. Safonov. Maryland ix . A. Rotea. Hall. Willems. E. and A. E. Tits. Doyle. P.

.

Linear Matrix Inequalities in System and Control Theory .

.

By scanning the list above or the table of contents. it seems that the results are new.g. minimizing condition number by diagonal scaling • construction of quadratic Lyapunov functions for stability and performance analysis of linear diﬀerential inclusions • joint synthesis of state-feedback and quadratic Lyapunov functions for linear diﬀerential inclusions • synthesis of state-feedback and quadratic Lyapunov functions for stochastic and delay systems • synthesis of Lur’e-type Lyapunov functions for nonlinear systems • optimization over an aﬃne family of transfer matrices. in others.. the reader will see that Lyapunov’s methods will be our main focus. In many cases. which are traditionally 1 . The types of problems we consider include: • matrix scaling problems. e. published results. This is to show that Lyapunov’s methods. however. Here we have a secondary goal. beyond showing that many problems from Lyapunov theory can be cast as convex or quasiconvex problems. our reduction constitutes a solution to the original problem. but also in several other senses as well. including synthesis of multipliers for analysis of linear systems with unknown parameters • positive orthant stability and state-feedback synthesis • optimal system realization • interpolation problems. Since these resulting optimization problems can be solved numerically very eﬃciently using recently developed interior-point methods. certainly in a practical sense. the more conventional approach is to seek an analytic or frequency-domain solution to the matrix inequalities. In comparison. including scaling • multicriterion LQG/LQR • inverse problem of optimal control In some cases.Chapter 1 Introduction 1. we are extending known results.1 Overview The aim of this book is to show that we can reduce a very wide variety of problems arising in system and control theory to a few standard convex or quasiconvex optimization problems involving linear matrix inequalities (LMIs). we are describing known.

e. which is a special form of an LMI. Lyapunov also showed that this ﬁrst LMI could be explicitly solved. In summary. Although they did not explicitly form matrix inequalities. He showed that the diﬀerential equation d x(t) = Ax(t) (1.2) The requirement P > 0. which is guaranteed to be positive-deﬁnite if the system (1. the contribution was to show how to solve a certain family of LMIs by graphical methods. by hand. These criteria could be applied to higher order systems. when Lyapunov published his seminal work introducing what we now call Lyapunov theory. Popov. provided we do not insist on an “analytic solution”. The LMIs that resulted were solved analytically. but did not gracefully or usefully extend to systems containing more than one nonlinearity.. the problem of stability of a control system with a nonlinearity in the actuator. small systems). T (1. Copyright c 1994 by the Society for Industrial and Applied Mathematics. The next major breakthrough came in the early 1960’s. circle criterion. can just as well be used to ﬁnd bounds on system performance. all trajectories converge to zero) if and only if there exists a positivedeﬁnite matrix P such that AT P + P A < 0.2). especially. the ﬁrst LMI used to analyze stability of a dynamical system was the Lyapunov inequality (1. Lur’e. Kalman. 1. Tsypkin criterion. From the point of view of our story (LMIs in control theory). These inequalities were reduced to polynomial inequalities which were then checked “by hand” (for. using what we now call the positive-real (PR) lemma (see §2. Lur’e and others were the ﬁrst to apply Lyapunov’s methods to practical control engineering problems. A P + P A < 0 is what we now call a Lyapunov inequality on P . This resulted in the celebrated Popov criterion.2). and other researchers succeeded in reducing the solution of the LMIs that arose in the problem of Lur’e to simple graphical criteria. we can pick any Q = QT > 0 and then solve the linear equation AT P +P A = −Q for the matrix P . third order) systems.7.2 A Brief History of LMIs in Control Theory The history of LMIs in the analysis of dynamical systems goes back more than 100 years.2 Chapter 1 Introduction applied to the analysis of system stability. Nevertheless they were justiﬁably excited by the idea that Lyapunov’s theory could be applied to important (and diﬃcult) practical problems in control engineering. are now about to become a real medium for the examination of the urgent problems of contemporary engineering. and others in the Soviet Union applied Lyapunov’s methods to some speciﬁc practical problems in control engineering. Indeed. their stability criteria have the form of LMIs. and many variations. From the introduction of Lur’e’s 1951 book [Lur57] we ﬁnd: This book represents the ﬁrst attempt to demonstrate that the ideas expressed 60 years ago by Lyapunov. . Of course this limited their application to small (second. The story begins in about 1890. which even comparatively recently appeared to be remote from practical application. needless to say. which can be solved analytically (by solving a set of linear equations). The next major milestone occurs in the 1940’s. when Yakubovich. Postnikov.1) dt is stable (i. In summary.1) is stable.

The next major advance (in our view) was the simple observation that: The LMIs that arise in system and control theory can be formulated as convex optimization problems that are amenable to computer solution. and quadratic optimal control.3) and points out that it can be solved by studying the symmetric solutions of the ARE AT P + P A − (P B + C T )R−1 (B T P + C) + Q = 0. which in turn can be found by an eigendecomposition of a related Hamiltonian matrix. From our point of view. clearly and completely. it was known that the LMI appearing in the PR lemma could be solved not only by graphical means. Yak64. By 1970. and The method of matrix inequalities in the stability theory of nonlinear control systems (1965. It would be interesting to see whether or not it can be exploited in computational algorithms. which depends almost entirely on the problem size and not the particular problem data. Willems is led to the LMI AT P + P A + Q B P +C T P B + CT R ≥ 0. and by solving Lyapunov or Riccati equations. J. Pyatnitskii and Skorodinskii [PS82] were perhaps the ﬁrst researchers to make this point. especially by Yakubovich [Yak62. where the ARE was called the Lur’e resolving equation (see [Yak88]). In a 1971 paper [Wil71b] on quadratic optimal control. Yak67]. researchers knew several methods for solving special types of LMIs: direct (for small systems). and not the more general form that we consider in this book. Of course it cannot be solved exactly in a ﬁnite number of arithmetic steps for systems of ﬁfth and higher order.2 for details. and were found to be related to the ideas of passivity. these methods are all “closed-form” or “analytic” solutions that can be used to solve special forms of LMIs.7. the most important of which is that we can reliably solve many LMIs for which no “analytic solution” has been found (or is likely to be found). The solution of certain matrix inequalities in automatic control theory (1962). Although this is a simple observation.) In Willems’ 1971 paper we ﬁnd the following striking quote: The basic importance of the LMI seems to be largely unappreciated. The PR lemma and extensions were intensively studied in the latter half of the 1960s. So by 1971.2 A Brief History of LMIs in Control Theory 3 The important role of LMIs in control theory was already recognized in the early 1960’s. which they This electronic version is for personal use and may not be duplicated or distributed. (1. They reduced the original problem of Lur’e (extended to the case of multiple nonlinearities) to a convex optimization problem involving LMIs. (See §2. the small-gain criteria introduced by Zames and Sandberg. Here Willems refers to the speciﬁc LMI (1. . (Most control researchers and engineers consider the Riccati equation to have an “analytic” solution.3). since the standard algorithms that solve it are very predictable in terms of the eﬀort required. This observation was made explicitly by several researchers. C.1. for example. English translation 1967). Still. it has some important consequences. This is clear simply from the titles of some of his papers from 1962–5..) This connection had been observed earlier in the Soviet Union.g. e. graphical methods. Willems’ suggestion that LMIs might have some advantages in computational algorithms (as compared to the corresponding Riccati equations) foreshadows the next chapter in the story. but also by solving a certain algebraic Riccati equation (ARE).

but in contrast to the ellipsoid method. we often fail to mention regularity or other technical conditions. And of course. to formulate the search for a Lyapunov function as a convex optimization problem. in a 1965 paper by Schultz et al. • Early 1980’s: Recognition that many LMIs can be solved by computer via convex programming. analytic solution of the Lyapunov LMI via Lyapunov equation. In a 1976 paper.. Essentially all of this research activity concentrated on algorithms for linear and (convex) quadratic programs. and in particular. Karmarkar’s work spurred an enormous amount of work in the area of interior-point methods for linear programming (including the rediscovery of eﬃcient methods that were developed in the 1960s but ignored). several interior-point algorithms for LMI problems have been implemented and tested on speciﬁc families of LMIs that arise in control theory. and found to be extremely eﬃcient. Although there remains much to be done in this area.) Pyatnitskii and Skorodinskii were the ﬁrst. for example. • Late 1960’s: Observation that the same family of LMIs can be solved by solving an ARE. . • 1940’s: Application of Lyapunov’s methods to real control engineering problems. N. A summary of key events in the history of LMIs in control theory is then: • 1890: First LMI appears. The new development is the ability to directly solve (general) LMIs. Every statement is to be interpreted as being true modulo appropriate technical conditions (that in most cases are trivial to ﬁgure out).g. is also very eﬃcient in practice. to the problems we encounter in this book. [SSHJ65]. The ﬁnal chapter in our story is quite recent and of great practical importance: the development of powerful and eﬃcient interior-point methods to solve the LMIs that arise in system and control theory. It is fair to say that Yakubovich is the father of the ﬁeld. and Lyapunov the grandfather of the ﬁeld. We should also mention several precursors. • Late 1980’s: Development of interior-point algorithms for LMIs. (This problem had been studied before. e. We are very informal. Small LMIs solved “by hand”. We sometimes skip “details” that would be important if the optimization problem were to be solved numerically.3 Notes on the Style of the Book We use a very informal mathematical style. 1. In 1984. • Early 1960’s: PR lemma gives graphical techniques for solving another family of LMIs. Horisberger and Belanger [HB76] had remarked that the existence of a quadratic Lyapunov function that simultaneously proves stability of a collection of linear systems is a convex problem involving LMIs. it may be Copyright c 1994 by the Society for Industrial and Applied Mathematics. perhaps even cavalier. in our reduction of a problem to an optimization problem. as far as we know. Then in 1988.4 Chapter 1 Introduction then solved using the ellipsoid algorithm. Nesterov and Nemirovskii developed interior-point methods that apply directly to convex problems involving LMIs. and then apply an algorithm guaranteed to solve the optimization problem. like the ellipsoid method. As an example. Karmarkar introduced a new linear programming algorithm that solves linear programs in polynomial-time. the idea of having a computer search for a Lyapunov function was not new—it appears. but the “solutions” involved an arbitrary scaling matrix.

we adopt 0 the convention that the operators Tr (trace of a matrix) and E (expected value) have lower precedence than multiplication. and soon it was clear that we were writing a book. even though good ones may be available. and so on. The list of problems that we consider is meant only to be representative. by Boyd and El Ghaoui [BE93]. we use the standard convention of dropping the time argument from the variables in diﬀerential equations. despite its size (over 500 entries). Thus. and later Balakrishnan. The appendix contains a list of notation and a list of acronyms used in the book. Once we reduce a problem arising in control theory to a convex program. for each reduced problem we could state. Presumably these dual problems and lower bounds can be given interesting system-theoretic interpretations. when we encounter timevarying coeﬃcients. T T T u y dt for 0 u(t)T y(t)dt. as in x = A(t)x. To lighten the notation. in which we hide proofs.g. We do not discuss initial guesses for the optimization problems. we do not pursue any theoretical aspects of reducing a problem to a convex problem involving matrix inequalities. started adding material. but grew too large to be a section. Thus. and certainly not exhaustive. The order of the authors’ names reﬂects this history. ˙ Similarly. In a similar way. To reduce the number of parentheses required. 1. we drop the dummy variable from deﬁnite integrals. The completeness of the bibliography should not be overestimated. Then Feron. In the ﬁrst chapter on analysis of linear diﬀerential inclusions. we describe fewer and fewer variations. probably simplify. and assume that the reader can translate the results from the continuous-time case to the discrete-time case. This electronic version is for personal use and may not be duplicated or distributed. assuming that the reader could work out the extensions. We switch to discrete-time systems when we consider system realization problems (which almost always arise in this form) and also when we consider stochastic systems (to avoid the technical details of stochastic diﬀerential equations). Therefore. we can consider various dual optimization problems. For a few months it was a manuscript (that presumably would have been submitted for publication as a paper) entitled Generalized Eigenvalue Problems Arising in Control Theory. and then interpret in system or control theoretic terms the optimality conditions for the resulting convex problem. For example. we describe many variations on problems (e. We mostly consider continuous-time systems. not a paper. lower bounds for the problem. x = Ax is our short ˙ form for dx/dt = Ax(t). in later chapters..4 Origin of the Book This book started out as a section of the paper Method of Centers for Minimizing Generalized Eigenvalues. and bibliography and historical notes. We apologize to the reader for the seven new acronyms we introduce. Each chapter concludes with a section entitled Notes and References. writing for example. Tr AT B means Tr AT B . Another fascinating topic that could be explored is the relation between system and control theory duality and convex programming duality. etc.1. transpose. precise statements.4 Origin of the Book 5 necessary to add constraints to the optimization problem for normalization or to ensure boundedness. the reader who wishes to implement an algorithm that solves a problem considered in this book should be prepared to make a few modiﬁcations or additions to our description of the “solution”. . our treatment of problems becomes more terse as the book progresses. To avoid excessive repetition. computing bounds on margins and decay rates). elaborations. we will explicitly show the time dependence. Here A is a constant matrix.

.

Of course.. such as Lyapunov and convex quadratic matrix inequalities. . 7 (2.1) may seem to have a specialized form. i.1) and the nonstrict LMI (2. F (p) (x) > 0 can be expressed as the single LMI diag(F (1) (x).1) where x ∈ Rm is the variable and the symmetric matrices Fi = FiT ∈ Rn×n . . linear inequalities.5. matrix norm inequalities. . i = 0. . When the matrices Fi are diagonal. can all be cast in the form of an LMI.1 Linear Matrix Inequalities ∆ m A linear matrix inequality (LMI) has the form F (x) = F0 + i=1 xi Fi > 0. (2.1) is equivalent to a set of n polynomial inequalities in x. Therefore we will make no distinction between a set of LMIs and a single LMI. the set {x | F (x) > 0} is convex. . (convex) quadratic inequalities. are given.2) . . the LMI F (x) > 0 is just a set of linear inequalities.3) F (x) ≥ 0. Although the LMI (2. m. (2. i. . . .. but a precise statement of the relation is a bit involved. Nonlinear (convex) inequalities are converted to LMI form using Schur complements.e. In particular. and constraints that arise in control theory. .. We will also encounter nonstrict LMIs. The basic idea is as follows: the LMI Q(x) S(x)T S(x) R(x) > 0. . The inequality symbol in (2. it can represent a wide variety of convex constraints on x. uT F (x)u > 0 for all nonzero u ∈ Rn . F (p) (x)) > 0. which have the form The strict LMI (2.2) are closely related. the leading principal minors of F (x) must be positive. .. the LMI (2. The LMI (2. “the LMI F (1) (x) > 0.e. . . In the next few sections we consider strict LMIs.Chapter 2 Some Standard Problems Involving LMIs 2. F (p) (x) > 0” will mean “the LMI diag(F (1) (x).1) is a convex constraint on x.1) means that F (x) is positivedeﬁnite. . so we defer it to §2. . i. . .e. . F (p) (x)) > 0”.e. Multiple LMIs F (1) (x) > 0. . i.

In this case we will not write out the LMI explicitly in the form F (x) > 0. and the LMI (in x and X): Tr X < 1. (2. As an example. R(x) = R(x)T .. . the Lyapunov inequality AT P + P A < 0. is expressed as the LMI P (x) c(x)T c(x) 1 > 0.4) In other words.3). the (maximum singular value) matrix norm constraint Z(x) < 1. where Z(x) ∈ Rp×q and depends aﬃnely on x. P (x) > 0. .4. P (x) > 0. Copyright c 1994 by the Society for Industrial and Applied Mathematics.4.5) is readily put in the form (2. the set of nonlinear inequalities (2. and S(x) depend aﬃnely on x. see §2. X S(x) S(x)T P (x) > 0. where c(x) ∈ Rn and P (x) = P (x)T ∈ Rn×n depend aﬃnely on x. As another related example.5).1 Matrices as variables We will often encounter problems in which the variables are matrices. More generally.g. the Lyapunov inequality (2. (Of course. is equivalent to R(x) > 0. Let P1 . . Many other convex constraints on x can be expressed in the form of an LMI. in addition to saving notation. Pm be a basis for symmetric n × n matrices (m = n(n + 1)/2).4) can be represented as the LMI (2. where P (x) = P (x)T ∈ Rn×n and S(x) ∈ Rn×p depend aﬃnely on x.) Leaving LMIs in a condensed form such as (2. consider the quadratic matrix inequality AT P + P A + P BR−1 B T P + Q < 0. is represented as the LMI I Z(x)T Z(x) I >0 (since Z < 1 is equivalent to I − ZZ T > 0).1). see the Notes and References. 2. the constraint Tr S(x)T P (x)−1 S(x) < 1.6) . .8 Chapter 2 Some Standard Problems Involving LMIs where Q(x) = Q(x)T . as follows.1. The constraint c(x)T P (x)−1 c(x) < 1.5) where A ∈ Rn×n is given and P = P T is the variable. e. (2. may lead to more eﬃcient computation. Note that the case q = 1 reduces to a general convex quadratic inequality on x. Then take F0 = 0 and Fi = −AT Pi − Pi A. but instead make clear which matrices are the variables. is handled by introducing a new (slack) matrix variable X = X T ∈ Rp×p . (2. The phrase “the LMI AT P + P A < 0 in P ” means that the matrix P is a variable. Q(x) − S(x)R(x)−1 S(x)T > 0.

Then take F0 = diag(P0 . . AT P + P A < 0. . L.2 Some Standard Problems 9 where A. this means: Find a nonzero G ≥ 0 such that Tr GFi = 0 for i = 1. m. 2. . . . . We will say “solving the LMI F (x) > 0” to mean solving the corresponding LMIP. . leaving any required elimination of equality constraints to the reader. QL . . It can be expressed as the linear matrix inequality −AT P − P A − Q BT P PB R > 0. and P = P T is the variable. . . −AT Pi − Pi A) for i = 1. . Note that this is a quadratic matrix inequality in the variable P . .7) in the form F (x) > 0. Pm be a basis for symmetric k×k matrices with trace zero (m = (k(k + 1)/2) − 1) and let P0 be a symmetric k × k matrix with Tr P0 = 1. . and need to ﬁnd P satisfying the LMI P > 0. . . . AT P + P Ai < 0.7) as LMIs. . Q0 = i=1 Qi A T + A i Qi . (2. . the corresponding LMI Problem (LMIP) is to ﬁnd xfeas such that F (xfeas ) > 0 or determine that the LMI is infeasible. . This representation also clearly shows that the quadratic matrix inequality (2. i = 1. −AT P0 − P0 A) and Fi = diag(Pi . . or determine that no such P exists. . Determining that no such P exists is equivalent to ﬁnding Q0 .2. We will refer to constraints such as (2.2. . which is not obvious. . QL ≥ 0. 2.6) is convex in P . . . . Tr P = 1. Q = QT . consider the “simultaneous Lyapunov stability problem” (which we will see in §5. m and Tr GF0 ≤ 0. such that L Q0 ≥ 0. . As an example of an LMIP. 2.1.2 Some Standard Problems Here we list some common convex and quasiconvex problems that we will encounter in the sequel. not all zero. this is a convex feasibility problem. see the Notes and References. e. Of course we can eliminate the equality constraint to write (2. This electronic version is for personal use and may not be duplicated or distributed.1 LMI problems Given an LMI F (x) > 0. . P > 0.8) which is another (nonstrict) LMIP. (By duality.) Of course. i i = 1.g. B. R = RT > 0 are given matrices of appropriate sizes. Let P1 .7) where P ∈ Rk×k is the variable.2 Linear equality constraints In some problems we will encounter linear equality constraints on the variables. .1): We are given Ai ∈ Rn×n . L. i (2.

any can be transformed into any other. this EVP can also be expressed in terms of the associated quadratic matrix inequality: minimize subject to γ AT P + P A + C T C + γ −1 P BB T P < 0. and C ∈ Rm×n are given..10 Chapter 2 Some Standard Problems Involving LMIs 2. In the special case when the matrices Fi are all diagonal.e. B(x) > 0. i. . This is a convex optimization problem. minimize subject to λ λI − A(x) > 0. From our remarks above. and P and γ are the optimization variables. this problem reduces to the general linear programming problem: minimizing the linear function cT x subject to a set of linear inequalities on x. The general form of a GEVP is: minimize subject to λ λB(x) − A(x) > 0.e.2): minimize subject to γ −AT P − P A − C T C BT P PB γI > 0. B(x)) B(x) > 0. i.3. We leave it to the reader to verify that these forms are equivalent. EVPs can appear in the equivalent form of minimizing a linear function subject to an LMI.e. subject to an LMI constraint (or determine that the constraint is infeasible).. subject to an LMI constraint. We can express this as minimize subject to λmax (A(x). P >0 2.3 Generalized eigenvalue problems The generalized eigenvalue problem (GEVP) is to minimize the maximum generalized eigenvalue of a pair of matrices that depend aﬃnely on a variable. C(x) > 0 Copyright c 1994 by the Society for Industrial and Applied Mathematics.2. consider the problem (which appears in §6. Another equivalent form for the EVP is: minimize subject to λ A(x. i. λ).2 Eigenvalue problems The eigenvalue problem (EVP) is to minimize the maximum eigenvalue of a matrix that depends aﬃnely on a variable. B(x) > 0 where A and B are symmetric matrices that depend aﬃnely on the optimization variable x.2. λ) > 0 where A is aﬃne in (x.. B ∈ Rn×p . P >0 where the matrices A ∈ Rn×n . As an example of an EVP. B and C are symmetric matrices that are aﬃne functions of x.9) with F an aﬃne function of x. minimize subject to cT x F (x) > 0 (2. C(x) > 0 where A.

1.) 2. B(θx + (1 − θ)˜)) ≤ x x max {λmax (A(x)..4. µ). . This GEVP is a quasiconvex optimization problem since the constraint is convex and the objective. log det A−1 is a convex function of A. B(x)). (This problem arises in §5. we leave these problems in the more natural form. λ) ≥ A(x. .2. An equivalent alternate form for a GEVP is minimize subject to λ A(x.11) i = 1. see the Notes and References. which we will abbreviate CP: minimize subject to log det A(x)−1 A(x) > 0. see [NN94. In addition. B(x)). x x Note that when the matrices are all diagonal and A(x) and B(x) are scalar. λmax (A(x).2 Some Standard Problems 11 where λmax (X.) Remark: Problem CP can be transformed into an EVP. minimizing a linear-fractional function subject to a set of linear inequalities. we will also encounter the following convex problem.4 A convex problem Although we will be concerned mostly with LMIPs. B.. this problem reduces to the general linear-fractional programming problem. i. . the largest eigenvalue of the matrix Y −1/2 XY −1/2 . (2. many nonlinear quasiconvex functions can be represented in the form of a GEVP with appropriate A. consider the problem maximize subject to α −AT P − P A − 2αP > 0. vi P vi ≤ 1. and satisﬁes the monotonicity condition λ > µ =⇒ A(x. Let E denote the ellipsoid centered at the origin determined by P . B(x) > 0 (2. and GEVPs. B(˜))} . since det A(x) > λ can be represented as an LMI in x and λ. This means that for feasible x. EVPs.e.10) where A and B are symmetric matrices that depend aﬃnely on x.e. . (Note that when A > 0. This electronic version is for personal use and may not be duplicated or distributed. Y ) denotes the largest generalized eigenvalue of the pencil λY − X with Y > 0. and C. As an example of a GEVP. P >0 where the matrix A is given. We will encounter several variations of this problem. λ) is aﬃne in x for ﬁxed λ and aﬃne in λ for ﬁxed x. is quasiconvex.2. consider the problem: minimize subject to log det P −1 T P > 0. x and 0 ≤ θ ≤ 1. As an example of CP. §6. λmax (A(˜). L here vi ∈ Rn are given and P = P T ∈ Rn×n is the variable.3]. i. . which has the following interpretation. ˜ λmax (A(θx + (1 − θ)˜). λ) > 0 where A(x. As we do with variables which are matrices. and the optimization variables are the symmetric matrix P and the scalar α.

e. 2. EVPs. . and CPs) are tractable. and if it is. . minimizing log det P −1 is the same as minimizing the volume of E. . or equivalently. .) The basic idea of the algorithm is as follows. .) We then know that the sliced half-ellipsoid E (0) ∩ z g (0)T (z − x(0) ) ≤ 0 contains an optimal point.. we will assume that the problem we are solving has at least one optimal point.11). E (1) is then guaranteed to contain an optimal point. the polytope Co{v1 . from both theoretical and practical viewpoints: • They can be solved in polynomial-time (indeed with a variety of interpretations for the term “polynomial-time”). So by solving (2. we consider any feasible point as being optimal. ˜ ˜ ˜ Copyright c 1994 by the Society for Industrial and Applied Mathematics. roughly speaking. We then compute a cutting plane for our problem that passes through the center x (0) of E (0) . The minimum volume ellipsoid that contains the half-ellipsoid (z − a)T A−1 (z − a) ≤ 1. Since the volume of E is proportional to (det P )−1/2 . we ﬁnd the minimum volume ellipsoid. In practice. By “solve the problem” we mean: Determine whether or not the problem is feasible. . An ellipsoid E can be described as E= z z (z − a)T A−1 (z − a) ≤ 1 where A = AT > 0. the constraints are feasible. vL .12 ∆ Chapter 2 Some Standard Problems Involving LMIs E = z z T P z ≤ 1 . 2. depending on the situation). . Now we compute the ellipsoid E (1) of minimum volume that contains this sliced half-ellipsoid. We describe it here because it is very simple and. The constraints are simply vi ∈ E. . The process is then repeated.5 Solving these problems The standard problems (LMIPs. . that contains the points v1 . where Co denotes the convex hull. We now describe the algorithm more explicitly. Although more sophisticated versions of the ellipsoid algorithm can detect infeasible constraints. the interior-point algorithms described in the next section are much more eﬃcient. from a theoretical point of view. • They can be solved in practice very eﬃciently. compute a feasible point with an objective value that exceeds the global minimum by less than some prespeciﬁed accuracy. is guaranteed to solve the standard problems. eﬃcient (polynomial-time). We start with an ellipsoid E (0) that is guaranteed to contain an optimal point. This means that we ﬁnd a nonzero vector g (0) such that an optimal point lies in the half-space z g (0)T (z − x(0) ) ≤ 0 (or in the half-space z g (0)T (z − x(0) ) < 0 . (In the feasibility problem. g T (z − a) ≤ 0 is given by ˜ E= z (z − a)T A−1 (z − a) ≤ 1 . vL }. i. (We will explain how to do this for each of our problems later. however.3 Ellipsoid Algorithm We ﬁrst describe a simple ellipsoid algorithm that.2. centered at the origin. GEVPs.

the minimum length interval containing a half-interval is the half-interval itself. (We refer the reader to the papers cited in the Notes and References for precise statements of what we mean by “polynomial-time” and indeed. (We note that these formulas hold only for m ≥ 2. m.) The ellipsoid algorithm is initialized with x(0) and A(0) such that the corresponding ellipsoid contains an optimal point. If x does not satisfy this LMI. 2.” as well as proofs. The algorithm then proceeds as follows: for k = 1.3 Ellipsoid Algorithm 13 where a=a− ˜ A˜ g . It turns out that the volume of these ellipsoids decreases geometrically. In the case of ˜ one variable. compute a g (k) that deﬁnes a cutting plane at x(k) g := g (k)T A(k) g (k) ˜ x(k+1) := x(k) − A(k+1) := m2 m2 −1 −1/2 g (k) 1 (k) g ˜ m+1 A A(k) − 2 (k) g g T A(k) ˜˜ m+1 A This recursion generates a sequence of ellipsoids that are guaranteed to contain an optimal point. the ellipsoid algorithm. . “convergence.2. . Deﬁne g by gi = −uT Fi u. i = 1. . in this case. there exists a nonzero u such that m uT F (x)u = uT F0 + i=1 x i Fi u ≤ 0. reduces to the familiar bisection algorithm.. and this fact can be used to prove polynomial-time convergence of the algorithm. m+1 ˜ A= m2 m2 − 1 A− 2 A˜g T A . . . this g deﬁnes a cutting plane for the LMIP at the point x. We have vol(E (k) ) ≤ e− 2m vol(E (0) ). i. g˜ m+1 and g = g/ g T Ag.) We now show how to compute cutting planes for each of our standard problems. EVPs: Consider the EVP minimize subject to cT x F (x) > 0 This electronic version is for personal use and may not be duplicated or distributed. . It follows that every feasible point lies in the half-space {z | g T (z − x) < 0}. . Then for any z satisfying g T (z − x) ≥ 0 we have m uT F (z)u = uT F0 + i=1 zi Fi u = uT F (x)u − g T (z − x) ≤ 0. . LMIPs: Consider the LMI m k F (x) = F0 + i=1 xi Fi > 0.e.

14 Chapter 2 Some Standard Problems Involving LMIs Suppose ﬁrst that the given point x is infeasible.4 Interior-Point Methods Since 1988.e. and hence z cannot be optimal. Now suppose we are given a point x. C(x) > 0 m m m Here. So we assume that x is feasible.. Suppose now that x is feasible. a cutting plane is given by the gradient of the objective m uT (λmax (A(x). Then we can construct a cutting plane for this problem using the method described above for LMIPs. g T (z − x) > 0 implies log det A(z)−1 > log det A(x)−1 . we use the method described for LMIPs to generate a cutting plane at x. . B(x)) B(x) > 0. To see this. Deﬁne g by gi = −uT (λmax (A(x). When the given point x is infeasible we already know how to generate a cutting plane. B(x))Bi − Ai )u. Here. Hence. In this case. GEVPs: We consider the formulation minimize subject to λmax (A(x). In this case. If the constraints are violated.10). B(x))B(x) − A(x)) u = 0. B(x)) − log det A(x) = − log det A0 + xi Ai i=1 at x. i. . and hence cannot be optimal. log det A(z)−1 ≥ log det A(x)−1 + g T (z − x). we are discarding the half-space {z | g T (z − x) > 0} because all such points (whether feasible or not) have an objective value larger than x. CPs: We now consider our standard CP (2.e. We claim g deﬁnes a cutting plane for this GEVP at the point x.. F (x) > 0. F (x) > 0. Now assume that the given point x is feasible. A(x) = A0 + i=1 xi Ai . In particular. we note that for any z. we are discarding the half-space {z | g T (z − x) ≥ 0} because all such points are infeasible. i. m. B(z)) ≥ λmax (A(x). Since the objective function is convex we have for all z. i = 1. The references describe more sophisticated interior-point methods for the other problems (including the problem of computing a feasible starting point for an EVP). . given a feasible starting point.. the vector g = c deﬁnes a cutting plane for the EVP at the point x. λmax (A(z). In this case. 2. .e. B(x))B(z) − A(z)) u = −g T (z − x). if g T (z − x) ≥ 0 we ﬁnd that which establishes our claim. interior-point methods have been developed for the standard problems. In this section we describe a simple interior-point method for solving an EVP. . i. Copyright c 1994 by the Society for Industrial and Applied Mathematics. Pick any u = 0 such that (λmax (A(x). B(x) = B0 + i=1 xi Bi and C(x) = C0 + i=1 xi Ci . gi = − Tr Ai A(x)−1 .

can be used to eﬃciently compute x . i. (2. The analytic center is.4 Interior-Point Methods 15 2. . so it has a unique minimizer. . . (2. We now turn to the problem of computing the analytic center of an LMI.2]. In [NN94. and in particular.16) This electronic version is for personal use and may not be duplicated or distributed. with appropriate step length selection.) We suppose now that the feasible set is nonempty and bounded.14) ∆ (2. We consider the algorithm: x(k+1) := x(k) − α(k) H(x(k) )−1 g(x(k) ). (This is a special form of our problem CP. which we denote x : x = arg min φ(x). respectively. x We refer to x as the analytic center of the LMI F (x) > 0. starting from a feasible initial point. F (x ) has maximum determinant among all positive-deﬁnite matrices of the form F (x). Nesterov and Nemirovskii give a simple step length rule appropriate for a general class of barrier functions (called self-concordant). Their damping factor is: α(k) := 1 1/(1 + δ(x(k) )) if δ(x(k) ) ≤ 1/4. F (x) > 0 (2.15) where α(k) is the damping factor of the kth iteration. Remark: The analytic center depends on the LMI (i. .4. the data F0 . invariant under congruence of the LMI: F (x) > 0 and T T F (x)T > 0 have the same analytic center provided T is nonsingular. Consider the LMI m F (x) = F0 + i=1 xi Fi > 0.. Fm ) and not its feasible set: We can have two LMIs with diﬀerent analytic centers but identical feasible sets. (We have already encountered this function in our standard problem CP. . (2. The function φ(x) = ∆ log det F (x)−1 ∞ F (x) > 0 otherwise. and is important in its own right. . the function φ. which implies that the matrices F1 . It can be shown that φ is strictly convex on the feasible set. Equivalently. at x(k) .e.13) that is. §2.) Newton’s method. Fm are linearly independent (otherwise the feasible set will contain a line). .12) is ﬁnite if and only if F (x) > 0. .1 Analytic center of an LMI The notion of the analytic center of an LMI plays an important role in interior-point methods. however. and g(x(k) ) and H(x(k) ) denote the gradient and Hessian of φ. where Fi = FiT ∈ Rn×n . x = arg max det F (x). and becomes inﬁnite as x approaches the boundary of the feasible set {x|F (x) > 0}. it is a barrier function for the feasible set..2. otherwise. .e.

4. given a feasible starting point. i.18) has a bounded feasible set. so for each λ > λopt the LMI F (x) > 0.3 Method of centers The method of centers is a simple interior-point algorithm that solves an EVP. Nesterov and Nemirovskii show that this step length always results in x(k+1) feasible.16). the matrices F0 .e. The curve given by x (λ) for λ > λopt is called the path of centers for the EVP. Fm . (2. and the numbers m and n. and c3 are three absolute constants. It can be shown that it is analytic and has a limit as λ → λopt . The point xopt is optimal (or more precisely. x (λ) is feasible and satisﬁes cT x (λ) < λ. The algorithm is initialized with λ(0) and x(0) . (2.16 where δ(x) = ∆ Chapter 2 Some Standard Problems Involving LMIs g(x)T H(x)−1 g(x) is called the Newton decrement of φ at x. The ﬁrst and second terms on the right-hand side do not depend on any problem data.. i. 2. i. T x (λ) λ−c 1 λ − cT x (2.17) where c1 . .19) i = 1. . 2. c2 . Indeed. with F (x(0) ) > 0 and cT x(0) < λ(0) . . . The point x (λ) is characterized by ∂ ∂xi log det F (x)−1 + log x (λ) = − Tr F (x (λ))−1 Fi + ci = 0. F (x(k+1) ) > 0.4. We will also assume that the LMI (2. the limit of a minimizing sequence) since for λ > λopt .. m. The last term on the right-hand side of (2. . and proceeds as follows: λ(k+1) x (k+1) := := (1 − θ)cT x(k) + θλ(k) x (λ(k+1) ) Copyright c 1994 by the Society for Industrial and Applied Mathematics. They prove that φ(x(k) ) − φ(x ) ≤ whenever k ≥ c1 + c2 log log(1/ ) + c3 φ(x(0) ) − φ(x ) . .17) depends on how “centered” the initial point is. .2 The path of centers Now consider the standard EVP: minimize subject to cT x F (x) > 0 cT x < λ Let λopt denote its optimal value..e. and convergence of x(k) to x . and therefore has an analytic center which we denote x (λ): x (λ) = arg min x ∆ log det F (x)−1 + log 1 λ − cT x . . they give sharp bounds on the number of steps required to compute x to a given accuracy using Newton’s method with the step length (2. . The second term grows so slowly with required accuracy that for all practical purposes it can be lumped together with the ﬁrst and considered an absolute constant.e. speciﬁc numbers. which we denote xopt .18) is feasible.

Multiplying (2. however. and therefore can be used as the initial point for computing the next iterate x (λ(k+1) ) via Newton’s method. cT x < λ. In other words we use x (λ) = arg min x ∆ log det F (x)−1 + q log 1 λ − cT x . This electronic version is for personal use and may not be duplicated or distributed.e. . • Instead of deﬁning x (λ) to be the analytic center of the LMI F (x) > 0. We make a few comments here. The modiﬁcations are: • Instead of the Nesterov–Nemirovskii step length. With θ > 0. First. the current iterate x(k) is feasible for the inequality cT x < λ(k+1) . but more sophisticated versions are. see the Notes and References.19) by (xi − xopt ) i and summing over i yields: 1 cT (x(k) − xopt ). n ≥ (k) λ − cT x(k) where θ is a parameter with 0 < θ < 1. Note that we can also express the inequality above in the form which shows that the stopping criterion until λ(k) − cT x(k) < /n cT x(k) − λopt ≤ n(λ(k) − cT x(k) ). an exact line-search) will yield faster convergence to the analytic center. Second. Using q > 1.. and refer the reader to the Notes and References for further elaboration. guarantees that on exit. we conclude that 1 (cT x(k) − λopt ). The most eﬃcient algorithms developed so far appear to be primal-dual algorithms (and variations) and the projective method of Nemirovskii. however. where q > 1 is an integer. the optimal value has been found within . Tr F (x(k) )−1 F (x(k) ) − F (xopt ) = (k) λ − cT x(k) which proves that λ(k) approaches λopt with at least geometric convergence. this variation on the method of centers is not polynomialtime. The classic method of centers is obtained with θ = 0. n+1 Since Tr F (x(k) )−1 F (xopt ) ≥ 0. F (x) > 0. x (k) does not (quite) satisfy the new inequality cT x < λ(k+1) . we deﬁne it to be the analytic center of the LMI F (x) > 0 along with q copies of cT x < λ. the method of centers is not the most eﬃcient. yields much faster reduction of λ per iteration. (k) We now give a simple proof of convergence. Since computing an analytic center is itself a special type of CP. we can view the method of centers as a way of solving an EVP by solving a sequence of CPs (which is done using Newton’s method). In this case. say q ≈ n where F (x) ∈ Rn×n . a step length chosen to minimize φ along the Newton search direction (i. Among interior-point methods for the standard problems. we note that two simple modiﬁcations of the method of centers as described above yield an algorithm that is fairly eﬃcient in practice.4 Interior-Point Methods 17 Replacing cT x(k) by (λ(k+1) − θλ(k) )/(1 − θ) yields (λ(k+1) − λopt ) ≤ n + θ (k) (λ − λopt ). and perhaps more important.2.

E ∈ n×n . We will also encounter these problems with nonstrict LMIs. In comparison. Second.5 Strict and Nonstrict LMIs We have so far assumed that the optimization problems LMIP. symmetric matrices D1 .2 ). Copyright c 1994 by the Society for Industrial and Applied Mathematics. with a mixture of strict and nonstrict LMIs. the cost of solving L Lyapunov equations of the same size is O(m1.20) to a given accuracy in O(m 3. the computational eﬀort required to solve the leastsquares problems is much smaller than by standard “direct” methods such as QR or Cholesky factorization. It turns out that this Newton direction can be expressed as the solution of a weighted least-squares problem of the same size as the original problem. . . by exploiting problem structure in the conjugate-gradient iterations. . with α ≈ 2. or more generally. . See the Notes and References for more discussion. DL . In interior-point methods most of the computational eﬀort is devoted to computing the Newton direction of a barrier or similar function. the ellipsoid method solves (2.5 L). . AL ∈ Rn×n . L (2. GEVP. . We are given matrices A1 . exploiting the problem structure. First.4. the relative cost of solving L coupled Lyapunov inequalities.75 L1. and CP involve strict LMIs. This complexity estimate is remarkably small. . and still obtain an approximation of the Newton direction suitable for interior-point methods. The idea is very roughly as follows. For example.1. The problem is an EVP that we will encounter in §6. i i = 1. we ﬁnd that F (x) ∈ RN ×N with N = Ln. 2.4 Interior-point methods and problem structure An important feature of interior-point methods is that problem structure can be exploited to increase eﬃciency. When the Lyapunov inequalities are combined into one large LMI F (x) > 0. compared to the cost of solving L independent Lyapunov inequalities is O(m0. We consider the EVP R minimize subject to Tr EP AT P + P Ai + Di < 0. the constant hidden in the O(·) notation is much larger for the ellipsoid algorithm). Numerical experiments on families of problems with randomly generated data and families of problems arising in system and control theory show that the actual performance of the interior-point method is much better than the worst-case estimate: O(mα Lβ ) arithmetic operations. it is possible to terminate the conjugate-gradient iterations before convergence. EVP. .1 and β ≈ 1. This LMI has much structure: It is block-diagonal with each block a Lyapunov inequality.5 ) arithmetic operations to solve the problem to a given accuracy.2. . This example illustrates one of our points: The computational cost of solving one of the standard problems (which has no “analytic solution”) can be comparable to the computational cost of evaluating the solution of a similar problem that has an “analytic solution”. They prove the worst-case estimate of O(m2.18 Chapter 2 Some Standard Problems Involving LMIs 2.20).20) In this problem the variable is the matrix P . Using conjugate-gradient and other related methods to solve these least-squares systems gives two advantages. . An example will demonstrate the eﬃciencies that can be obtained using the techniques sketched above.2. .5 L) arithmetic operations (moreover. . so the dimension of the optimization variable is m = n(n + 1)/2. Vandenberghe and Boyd have developed a (primal-dual) interior-point method that solves (2.6 L0. . Therefore.

e. We will say that the LMI F (x) ≥ 0 is strictly feasible if its strict version is feasible. Fp ∈ Rq×q with q ≤ n such that: F (x) ≥ 0 ⇐⇒ x = Az + b for some z ∈ Rp with p ∆ ˜ ˜ F (z) = F0 + i=1 ˜ zi Fi ≥ 0 This electronic version is for personal use and may not be duplicated or distributed. then we have {x ∈ Rm | F (x) ≥ 0 } = {x ∈ Rm | F (x) > 0 }. .5.. is strictly feasible. When an LMI is feasible but not strictly feasible. say. . We have just seen that when an LMI is strictly feasible. This is true for the problems GEVP and CP as well.. Then there is a matrix A ∈ Rm×p with p ≤ m. Fm ∈ Rn×n be symmetric. Once again.2). . −x) with x ∈ R. and symmetric matrices ˜ ˜ F0 . however. ∞) so the right-hand side is −∞. The problem here is that F (x) is always singular. 2.1. Of course. i.2. we can replace nonstrict inequality with strict inequality in the problems EVP.. Note that this reduced LMI satisﬁes the constraint qualiﬁcation. minimize subject to cT x F (x) ≥ 0 19 Intuition suggests that we could simply solve the strict EVP (by. the feasible set of the nonstrict LMI is the closure of the feasible set of the strict LMI. The left-hand side. This example shows one of the two pathologies that can occur: The nonstrict inequality contains an implicit equality constraint (in contrast with an explicit equality constraint as in §2. the nonstrict LMI F (x) ≥ 0 is equivalent (in the sense of deﬁning equal feasible sets) to ˜ the “reduced” LMI F (x) = x ≥ 0. by eliminating implicit equality constraints and then reducing the resulting LMI by removing any constant nullspace.22) is +∞. a vector b ∈ Rm .e. If the strict LMI F (x) > 0 is feasible. It follows that inf cT x | F (x) ≥ 0 = inf cT x | F (x) > 0 . . and CP in order to solve them. i. (2. 0) with x ∈ R and c = −1. we can solve the strict EVP to obtain a solution of the nonstrict EVP. As a simple example. This is correct in most but not all cases.22) So in this case. the requirement of strict feasibility is a (very strong) constraint qualiﬁcation.21) i. . since the LMI F (x) ≥ 0 has the single feasible point x = 0. GEVP. .9). . . The other pathology is demonstrated by the example F (x) = diag(x.1 Reduction to a strictly feasible LMI It turns out that any feasible nonstrict LMI can be reduced to an equivalent LMI that is strictly feasible. .e. and the EVPs with the strict and nonstrict LMIs can be very diﬀerent. if there is some x0 ∈ Rm such that F (x0 ) > 0.21) need not hold. i.22) is +∞ since the strict LMI F (x) > 0 is infeasible. (2. The feasible set for the nonstrict LMI is the interval [0. The right-hand side of (2. the strict LMI is infeasible so the right-hand side of (2. consider F (x) = diag(x. is always 0. In the language of optimization theory. (2. The precise statement is: Let F0 .5 Strict and Nonstrict LMIs As an example consider the nonstrict version of the EVP (2.e. an interior-point method) to obtain the solution of the nonstrict EVP.

Consider the interesting case where the LMI (2. Astab where 0 < ω1 < · · · < ωr .. the part corresponding to each imaginary axis eigenvalue.20 Chapter 2 Some Standard Problems Involving LMIs ˜ where the LMI F (z) ≥ 0 is either infeasible or strictly feasible. and all the eigenvalues of the matrix Astab ∈ Rs×s have negative real part. so we can just take A = I. there are no implicit equality constraints or nontrivial ˜ common nullspace for F . 2. In most of the problems encountered in this book. or equivalently. the LMI F (z) ≥ 0 can be interpreted as the original LMI with its constant nullspace removed (see the Notes and References).23) is strictly feasible if and only if all trajectories of x = Ax converge to zero. and the part associated with eigenvalue zero.. We know from system theory that the LMI (2. if the eigenvalues of A have ˙ nonpositive real part. Roughly speaking.23) is feasible but not strictly feasible. See the Notes and References for a proof. we have separated out the stable part of A. For example we have inf cT x F (x) ≥ 0 = = inf cT (Az + b) inf cT (Az + b) ˜ F (z) ≥ 0 ˜ F (z) > 0 ˜ since the LMI F (z) ≥ 0 is either infeasible or strictly feasible. The matrix A and vector b describe the implicit equality constraints for the LMI ˜ F (x) ≥ 0.5.23) where A ∈ Rk×k is given. P > 0. correspond to Jordan blocks of size one. i. ˙ all the eigenvalues of A have negative real part. and the symmetric matrix P is the variable.e. Using this reduction we can.. Note that this LMI contains a strict inequality as well as a nonstrict inequality.23) is feasible if and only if all trajectories of x = Ax are bounded. 0kr+1 denotes the zero matrix in Rkr+1 ×kr+1 . Copyright c 1994 by the Society for Industrial and Applied Mathematics. (2. Similarly. the Lyapunov inequality: AT P + P A ≤ 0. . at least in principle. 0kr+1 . or equivalently. we see that by a change of coordinates we can put A in the form ˜ ∆ A = = T −1 AT diag 0 −ω1 Ik1 ω 1 I k1 0 .. and F = F . and those with zero real part are nondefective. 0 −ωr Ikr ω r I kr 0 .. From the remarks above. b = 0.2 Example: Lyapunov inequality To illustrate the previous ideas we consider the simplest LMI arising in control theory. always deal with strictly feasible LMIs. We also know from system theory that the LMI (2.

. stab In summary. AT Pstab + Pstab Astab T stab where the zero matrix has size 2k1 +· · ·+2kr +kr+1 . . We can take the symmetric matrices P1 . . r + 1. . P > 0 = ˜ P = diag P1 QT 1 Q1 P1 . . Pr+1 ..e. . . Pr QT r Qr Pr . Pstab and the skewsymmetric matrices Q1 . . The reduced LMI (2. . .25) has as variable the symmetric matrices P1 . . Thus. 2 2 The original LMI (2. . since we can take P1 . Qr as the “free” variable z.. The reduced LMI (2.2. r i Pr+1 > 0. i. . the dimension of the original variable x is m = k(k + 1)/2). .. Pr+1 . .e. ..23) involves a matrix inequality of size n = 2k (i. .. i = 1.. T Astab Pstab (2.23) is Pi QT i Qi Pi > 0.. AT Pstab + Pstab Astab ≤ 0. QT = −Qi . . This reduced LMI is strictly feasible. . . r. + Pstab Astab ≤ 0. and Pstab as the solution of the Lyapunov equation AT Pstab + Pstab Astab + I = 0. Pr+1 .. r. Pr QT r Qr Pr . . . Pr+1 > 0. . Now we substitute P in the form (2. . . . i = 1. the aﬃne mapping from z into x simply maps these matrices into P = T −T diag P1 QT 1 Q1 P1 . . the reduced stab LMI corresponding to (2.. .. .23) are that P must have this special structure. . . Pstab and the skew-symmetric matrices Q1 . .. . . so the dimension of the variable z in the reduced LMI is r p= i=1 2 ki + kr+1 (kr+1 + 1) s(s + 1) + < m. Q1 . We ﬁnd that AT P + P A = T T diag 0. . .24) Put another way. . .25) Pstab > 0. Pstab T −1 . .5 Strict and Nonstrict LMIs Using standard methods of Lyapunov theory it can be shown that P AT P + P A ≤ 0. We remove the constant nullspace to obtain the reduced version. i = 1. Pr+1 . (2. Pr+1 as identity matrices.23). . the original LMI (2.e. i = 1. This electronic version is for personal use and may not be duplicated or distributed. Pi QT i Qi Pi > 0. two inequalities of size k). T Astab Pstab + Pstab Astab ≤ 0 From this characterization we can ﬁnd a reduced LMI that is strictly feasible. Qr as zero matrices. Qr . Pstab > 0 . Pstab 21 ˜ T −T P T −1 Pi ∈ Rki ×ki .23) has one symmetric matrix of size k as variable (i. .24) back into the original LMI (2.25) involves a matrix inequality of size k + s < n. the equality constraints implicit in the LMI (2.

and xm is large enough that F (x) > 0 if and only if U ˜ x ˜ ˜ ˜ x ˜ ˜ x U T F (˜)U + xm (U T U )2 > U T F (˜)U U T F (˜)U Therefore. we can derive an equivalent inequality without those variables.22 Chapter 2 Some Standard Problems Involving LMIs 2. xm−1 such that U T F (x1 . Then (2.28) hold with z = z0 . we have eliminated the matrix variable X from (2.. . Copyright c 1994 by the Society for Industrial and Applied Mathematics. by taking xm extremely large. let Fm = U U T with U full-rank. . we see that ˜ ˜˜ x ˜ T F U (˜) > 0. e.6. . feasibility of the matrix inequality (2. (2.26) is equivalent to ˜ ˜ ˜ ﬁnd x1 . . Suppose for example that Fm ≥ 0 and has rank r < n. we have F (x) > 0 Since ˜ U U T ⇐⇒ ˜ U U T ˜ F (x) U U > 0. U (z) and V (z) do not depend on X. U U = 0 and [U U ˜ just means that [U U ] is nonsingular).28) disappears. Matrix inequalities of the form (2. Of course. ˜ x ∆ where x = [x1 · · · xm−1 ]T and F (˜) = F0 + x1 F1 + · · · + xm−1 Fm−1 . ˜ ˜ Suppose that for every z.27) arise in the controller synthesis problems described in Chapter 7. . Consider G(z) + U (z)XV (z)T + V (z)X T U (z)T > 0. xm−1 )U > 0. U (z) and V (z) are orthogonal complements of U (z) and V (z) respectively. .27) to form (2.27) where the vector z and the matrix X are (independent) variables.27) with variables X and z is equivalent to the feasibility of (2. ˜ F (x) U U = ˜ ˜ x ˜ U T F (˜)U ˜ x ˜ U T F (˜)U ˜ ˜ x U T F (˜)U ˜ x U T F (˜)U + xm (U T U )2 . Roughly speaking.2 Elimination of matrix variables When a matrix inequality has some variables that appear in a certain form. 2. In other words.27) holds for some X and z = z0 if and only if the inequalities ˜ ˜ U (z)T G(z)U (z) > 0. −1 ˜ ˜ x U T F (˜)U.1 Elimination of semideﬁnite terms In the LMIP ﬁnd x such that F (x) > 0. (2. we can ensure that F is positive-deﬁnite on the range of Fm . Note that if U (z) or V (z) has rank n for all z. the LMIP (2.28) with variable z.6 Miscellaneous Results on Matrix Inequalities 2. We prove this lemma in the Notes and References. ˜ More precisely. ˜ ˜ V (z)T G(z)V (z) > 0 (2.g. and G(z) ∈ Rn×n . so the LMIP reduces to making F positive-deﬁnite on the nullspace of Fm . and let U be an orthogonal T ˜ ˜ ] is of maximum rank (which in this case complement of U . .28). then the ﬁrst or second inequality in (2. .26) we can eliminate any variables for which the corresponding coeﬃcient in F is semidefinite. .6.

29) and reduced the size of the matrices in the LMI.30) are equivalent. in other cases. Thus we have eliminated the variable Y from (2. . ˜ ˜ with variable Q.30) F0 (ζ) ≥ 0 for all ζ such that Fi (ζ) ≥ 0. Fp be quadratic functions of the variable ζ ∈ Rn : Fi (ζ) = ζ T Ti ζ + 2uT ζ + vi . . . In some cases. (2. . AQ + QAT + BY + Y T B T < 0.31) can be written as T0 uT 0 u0 v0 p − τi i=1 Ti uT i ui vi ≥ 0. . . Note that (2. . (2.29) where Q and Y are the variables. . G(z) − σV (z)V (z)T > 0 for some σ ∈ R. The S-procedure for quadratic functions and nonstrict inequalities Let F0 .6. the converse holds. It is also equivalent to the LMIP Q > 0. (2.31) and (2. then (2. We consider the following condition on F0 . ˜ ˜ B T AQ + QAT B < 0.31) then (2. . . It is a nontrivial fact that when p = 1. . This electronic version is for personal use and may not be duplicated or distributed. p. . Obviously if there exist τ1 ≥ 0. . τp ≥ 0 such that p for all ζ.6 Miscellaneous Results on Matrix Inequalities 23 We can express (2. . . .2. where B is any matrix of maximum rank such that B T B = 0. where the variables are Q and σ ∈ R.30) holds. F0 (ζ) − i=1 τi Fi (ζ) ≥ 0. . As an example.3 The S-procedure i = 0. . p. Remark: If the functions Fi are aﬃne. we will encounter in §7. .28) in another form using Finsler’s lemma (see the Notes and References): G(z) − σU (z)U (z)T > 0. this is the Farkas lemma. AQ + QAT < σBB T . . ∆ 2.1 the LMIP with LMI Q > 0. this constraint can be expressed as an LMI in the data deﬁning the quadratic functions or forms. . we can form an LMI that is a conservative but often useful approximation of the constraint. This LMIP is equivalent to the LMIP with LMI Q > 0. i where Ti = TiT . provided that there is some ζ0 such that F1 (ζ0 ) > 0.2. We will often encounter the constraint that some quadratic function (or quadratic form) be negative whenever some other quadratic functions (or quadratic forms) are all negative. Fp : i = 1.

. At the same time we introduce and deﬁne several important terms from system and control theory.. . Then the condition (2. .34) holds can be expressed as an LMIP (in P and the scalar variable τ ). We brieﬂy describe some of these results in this section. . τp ≥ 0 such that T0 − τi Ti > 0. Let T0 . .e. (2. ξ π T AT P + P A P B BT P 0 ξ π < 0. i=1 (2. the converse holds. p. . . .34) Applying the second version of the S-procedure. i. . .32) holds.24 Chapter 2 Some Standard Problems Involving LMIs The S-procedure for quadratic forms and strict inequalities We will use another variation of the S-procedure. . τp . BT P −τ I Thus the problem of ﬁnding P > 0 such that (2. τp .30) is convex in ν. See the Notes and References for further discussion. . Tp ∈ Rn×n be symmetric matrices.30) holds has low complexity. provided T that there is some ζ0 such that ζ0 T1 ζ0 > 0. this problem has low complexity. however. . which involves quadratic forms and strict inequalities. . Remark: The ﬁrst version of the S-procedure deals with nonstrict inequalities and quadratic functions that may include constant and linear terms. (2. The second version deals with strict inequalities and quadratic forms only. Therefore.32) there exists τ1 ≥ 0. checking whether (2. quadratic functions without constant or linear terms. . Tp : ζ T T0 ζ > 0 for all ζ = 0 such that ζ T Ti ζ ≥ 0. On the other hand. . . We consider the following condition on T0 . This does not.7 Some LMI Problems with Analytic Solutions There are analytic solutions to several LMI problems of special form. .33) is an LMI in the variables T0 and τ1 . . . It is obvious that if p i = 1. Copyright c 1994 by the Society for Industrial and Applied Mathematics. (2. . . often with important system and control theoretic interpretations. u0 and v0 depend aﬃnely on some parameter ν.31) holds for some ν is an LMIP in the variables ν and τ1 .33) then (2. Remark: Suppose that T0 . . . S-procedure example In Chapter 5 we will encounter the following constraint on the variable P : for all ξ = 0 and π satisfying π T π ≤ ξ T C T Cξ. It is a nontrivial fact that when p = 1. mean that the problem of checking whether there exists ν such that (2. . Note that (2.34) is equivalent to the existence of τ ≥ 0 such that AT P + P A + τ C T C P B < 0. 2.

a numerical solution procedure via Riccati equations as well. This set of linear equations will be solvable and result in P > 0 if and only if the LMI is feasible. We give a simpliﬁed discussion here. In fact this procedure not only ﬁnds a solution when the LMI is feasible.37). and under some additional assumptions.e.7. The link with system and control theory is given by the following result. all trajectories of x = Ax converge to zero as ˙ t → ∞. ..3. it parametrizes all solutions as Q varies over the positive-deﬁnite cone.35) is equivalent to the quadratic matrix inequality AT P + P A + (P B − C T )(D + D T )−1 (P B − C T )T ≤ 0. The LMI considered is: P > 0.37) u(t)T y(t) dt ≥ 0 for all solutions of (2. and refer the reader to the References for more complete statements. (We will encounter a variation on this LMI in §6. B ∈ Rn×p . ∆ This electronic version is for personal use and may not be duplicated or distributed.1 Lyapunov’s inequality We have already mentioned the LMIP associated with Lyapunov’s inequality. P > 0. Passivity is equivalent to the transfer matrix H being positive-real.2 The positive-real lemma Another important example is given by the positive-real (PR) lemma. the LMI (2. B. Lyapunov showed that this LMI is feasible if and only the matrix A is stable. C) is minimal. which is nothing but a set of n(n + 1)/2 linear equations for the n(n + 1)/2 scalar variables in P . AT P + P A < 0 where P is variable and A ∈ Rn×n is given. (2.37) with x(0) = 0.36) We assume for simplicity that A is stable and the system (A. or equivalently.35) is feasible if and only if the linear system x = Ax + Bu. ˙ is passive. AT P + P A BT P − C −D T − D P B − CT ≤ 0. The LMI (2. all eigenvalues of A must have negative real part. i. which means that H(s) + H(s)∗ ≥ 0 for all Re s > 0.e. T 0 y = Cx + Du (2. which yields a “frequency-domain” interpretation for a certain LMIP. deﬁned as H(s) = C(sI − A)−1 B + D for s ∈ C.7. Passivity can also be expressed in terms of the transfer matrix of the linear system (2.3. 2. and the matrix P = P T ∈ Rn×n is the variable. i.) Note that if D + D T > 0. i. C ∈ Rp×n . we pick any Q > 0 and solve the Lyapunov equation AT P + P A = −Q.2..7 Some LMI Problems with Analytic Solutions 25 2. and D ∈ Rp×p are given. (2.35) where A ∈ Rn×n .e. To solve this LMIP.

Thus we have a graphical.38): if P = P T satisﬁes (2. or equivalently. e. frequency-domain condition for feasibility of the LMI (2. Then the system (2.g. The solution Pare thus obtained is the minimal element among the set of solutions of (2. With these assumptions the LMI (2. and D ∈ Rp×p are given. plotting the curve given by the real and imaginary parts of H(iω) for ω ∈ R (called the Nyquist plot of the linear system).37) is nonexpansive. e.. and D.. This approach was used in much of the work in the 1960s and 1970s described in §1. For simplicity we assume that A is stable and (A. Much more discussion of this method. With a few further technical assumptions. C ∈ Rp×n . B.35) can be solved by a method based on Riccati equations and Hamiltonian matrices. Pick V ∈ R2n×n so that its range is a basis for the stable eigenspace of M . V = [v1 · · · vn ] where v1 .38) which is just the quadratic matrix inequality (2.36) with equality substituted for inequality. then set Pare = V2 V1−1 . vn are a set of independent eigenvectors of M associated with its n eigenvalues with negative real part.38) we ﬁrst form the associated Hamiltonian matrix M= A − B(D + D T )−1 C −C (D + D ) T T −1 B(D + D T )−1 B T −A + C T (D + D T )−1 B T T C .39) where A ∈ Rn×n . (2. and the matrix P = P T ∈ Rn×n is the variable.37) is passive. Partition V as V = V1 V2 .35) is feasible if and only if there exists a real matrix P = P T satisfying the ARE AT P + P A + (P B − C T )(D + D T )−1 (P B − C T )T = 0. B.g. then P ≥ Pare . . To solve the ARE (2.2.3 The bounded-real lemma The same results appear in another important form.35). . T 0 T y(t)T y(t) dt ≤ u(t)T u(t) dt 0 Copyright c 1994 by the Society for Industrial and Applied Mathematics. including D + D T > 0. (2. if and only if M has no pure imaginary eigenvalues. When M has no pure imaginary eigenvalues we can construct a solution Pare as follows.7..e. Note that P > 0. 2. C.26 Chapter 2 Some Standard Problems Involving LMIs When p = 1 this condition can be checked by various graphical means. . where V1 and V2 are square. C) is minimal. can be found in the References. the LMI (2. . i. . This fact can be used to form a Routh– Hurwitz (Sturm) type test for passivity as a set of polynomial inequalities in the data A. the LMI (2. the bounded-real lemma. B ∈ Rn×p .38). This LMI is feasible if and only the linear system (2. AT P + P A + C T C B T P + DT C P B + CT D DT D − I ≤ 0.35) is feasible. including the precise technical conditions. Here we consider the LMI P > 0.

7. We can solve this equation by forming the Hamiltonian matrix M= A + B(I − D T D)−1 DT C −C T (I − DD T )−1 C −AT − C T D(I − D T D)−1 B T B(I − D T D)−1 B T .39) is feasible. For example.40) as described in §2. Notes and References LMIs The term “Linear Matrix Inequality” was coined by J. of constructing an appropriate P . the LMI (2. This condition is easily checked graphically. by plotting H(iω) versus ω ∈ R (called a singular value plot for p > 1 and a Bode magnitude plot for p = 1).7. B) is stabilizable.g.. AP + P AT < BB T where A ∈ Rn×n and B ∈ Rn×p are given and P is the variable. or equivalently.e.g. 2.37) with x(0) = 0. ∞ This is sometimes expressed as H ≤ 1 where = sup { H(s) | Re s > 0} is called the H∞ norm of the transfer matrix H. As a simple example consider the LMI P > 0. if and only if M has no imaginary eigenvalues.39). the term was used in several papers from the 1970s to refer to the This electronic version is for personal use and may not be duplicated or distributed. the LMI (2. i. the diagonal matrix has maximum determinant. With some appropriate technical conditions (see the Notes and References) including D T D < I.2. there exists a matrix K ∈ R p×n such that A + BK is stable.. i. . it is well-known that among all symmetric positive-deﬁnite matrices with given diagonal elements. e.5).39) to an ARE.Notes and References 27 for all solutions of (2. Several standard results from linear algebra can also be interpreted as analytic solutions to certain LMI problems of special form. Willems and is widely used now.. This LMI is feasible if and only if the pair (A.. As we mentioned in §1. Then the system (2.2. In this case we can ﬁnd the minimal solution to (2. or estimator gains for observers.4 Others Several other LMI problems have analytic solutions. We will describe this and other results in Chapter 7. Nonexpansivity is equivalent to the transfer matrix H satisfying the bounded-real condition. Once again this is the quadratic matrix inequality associated with the LMI (2.37) is nonexpansive. There are simple methods for determining whether this is the case. and when it is. This condition can also be expressed in terms of the transfer matrix H. e. Once again we can relate the LMI (2. C. with equality instead of the inequality.40) has a real solution P = P . This gives an analytic solution to a special (and fairly trivial) CP (see §3. the ones that arise in the synthesis of state-feedback for linear systems.e.39) is feasible if and only the ARE AT P + P A + C T C + (P B + C T D)(I − D T D)−1 (P B + C T D)T = 0 T (2. H(s)∗ H(s) ≤ I H ∆ ∞ for all Re s > 0.

41) holds if and only if I 0 0 U T Q S T S R I 0 0 U where [S1 S2 ] = SU . since the matrix is an aﬃne and not linear function of the variable. but it does arise in the “line-search (sub)problem”. Inequality (2. and Q T S1 T = S1 Q T S2 S1 Σ 0 S2 0 ≥ 0. In this context the variable represents the “step length” to be taken in an iteration of an algorithm. For the case of m = 1 all of our standard problems are readily solved. Golub and Van Loan [GL89] and the references therein). i. where R† denotes the Moore–Penrose inverse of R. Q − SR† S T ≥ 0. . We must then have S2 = 0. so that U T RU = Σ 0 0 0 . We think of the term “Linear Matrix Inequality” as analogous to the term “Linear Inequalities” used to describe aT x ≤ bi . .g. The term is also consistent with the title of an early paper by Bellman and Ky Fan: On Systems of Linear Inequalities in Hermitian Matrix Variables [BF63] (see below). Schur complements for nonstrict inequalities The Schur complement result of section §2. which holds if and only if Q − SR† S T ≥ 0. and it useful to know that such problems can be solved extremely eﬃciently. p. S(I − RR† ) = 0.g. e. i = 1. . when we restrict one of our standard problems to a speciﬁc line. . The condition Q ST S R ≥0 (2. by simultaneous diagonalization of F0 and F1 by an appropriate congruence.1 can be generalized to nonstrict inequalities as follows.3). A more accurate term might be “Aﬃne Matrix Inequality”.. The theory of matrix pencils is very old and quite complete. e. 0 S1 Σ ≥ 0. where Σ > 0 and diagonal. and there is an extensive literature (see. let U be an orthogonal matrix that diagonalizes R. Suppose Q and R are symmetric.. i When m = 1. which holds if and only if S(I − RR† ) = 0. The case m = 1 does not arise on its own in any interesting problem from system or control theory.e. Formulating convex problems in terms of LMIs The idea that LMIs can be used to represent a wide variety of convex constraints can be found in Nesterov and Nemirovskii’s report [NN90b] (which is really a preliminary version Copyright c 1994 by the Society for Industrial and Applied Mathematics.28 Chapter 2 Some Standard Problems Involving LMIs speciﬁc LMI (1. with appropriate partitioning. To see this. an LMI is nothing but a matrix pencil. .41) is equivalent to R ≥ 0..

Notes and References

29

of their book [NN94]) and software manual [NN90a]. They formalize the idea of a “positivedeﬁnite representable” function; see §5.3, §5.4, and §6.4 of their book [NN94]. The idea is also discussed in the article [Ali92b] and thesis [Ali91] of Alizadeh. In [BE93], Boyd and El Ghaoui give a list of quasiconvex functions that can be represented as generalized eigenvalues of matrices that depend aﬃnely on the variable.

**Infeasibility criterion for LMIs
**

The LMI F (x) > 0 is infeasible means the aﬃne set {F (x) | x ∈ Rm } does not intersect the positive-deﬁnite cone. From convex analysis, this is equivalent to the existence of a linear functional ψ that is positive on the positive-deﬁnite cone and nonpositive on the aﬃne set of matrices. The linear functionals that are positive on the positive-deﬁnite cone are of the form ψ(F ) = Tr GF , where G ≥ 0 and G = 0. From the fact that ψ is nonpositive on the aﬃne set {F (x)|x ∈ Rm }, we can conclude that Tr GFi = 0, i = 1, . . . , m and Tr GF0 ≤ 0. These are precisely the conditions for infeasibility of an LMI that we mentioned in §2.2.1. For the special case of multiple Lyapunov inequalities, these conditions are given Bellman and Ky Fan [BF63] and Kamenetskii and Pyatnitskii [KP87a, KP87b].

It is straightforward to derive optimality criteria for the other problems, using convex analysis. Some general references for convex analysis are the books [Roc70, Roc82] and survey article [Roc93] by Rockafellar. The recent text [HUL93] gives a good overview of convex analysis; LMIs are used as examples in several places.

**Complexity of convex optimization
**

The important role of convexity in optimization is fairly widely known, but perhaps not well enough appreciated, at least outside the former Soviet Union. In a standard (Western) treatment of optimization, our standard problems LMIP, EVP, GEVP, and CP would be considered very diﬃcult since they are nondiﬀerentiable and nonlinear. Their convexity properties, however, make them tractable, both in theory and in practice. In [Roc93, p194], Rockafellar makes the important point: One distinguishing idea which dominates many issues in optimization theory is convexity . . . An important reason is the fact that when a convex function is minimized over a convex set every locally optimal solution is global. Also, ﬁrstorder necessary conditions turn out to be suﬃcient. A variety of other properties conducive to computation and interpretation of solutions ride on convexity as well. In fact the great watershed in optimization isn’t between linearity and nonlinearity, but convexity and nonconvexity. Detailed discussion of the (low) complexity of convex optimization problems can be found in the books by Gr¨tschel, Lov´sz, Schrijver [GLS88], Nemirovskii and Yudin [NY83], and o a Vavasis [Vav91]. Complete and detailed worst-case complexity analyses of several algorithms for our standard problems can be found in Chapters 3, 4, and 6 of Nesterov and Nemirovskii [NN94].

Ellipsoid algorithm

The ellipsoid algorithm was developed by Shor, Nemirovskii, and Yudin in the 1970s [Sho85, NY83]. It was used by Khachiyan in 1979 to prove that linear programs can be solved in polynomial-time [Kha79, GL81, GLS88]. Discussion of the ellipsoid method, as well as several extensions and variations, can be found in [BGT81] and [BB91, ch14]. A detailed history of its development, including English and Russian references, appears in chapter 3 of Akg¨l [Akg84]. u This electronic version is for personal use and may not be duplicated or distributed.

30

Chapter 2

Some Standard Problems Involving LMIs

**Optimization problems involving LMIs
**

One of the earliest papers on LMIs is also, in our opinion, one of the best: On systems of linear inequalities in Hermitian matrix variables, by Bellman and Ky Fan [BF63]. In this paper we ﬁnd many results, e.g., a duality theory for the multiple Lyapunov inequality EVP and a theorem of the alternative for the multiple Lyapunov inequality LMIP. The paper concentrates on the associated mathematics: there is no discussion of how to solve such problems, or any potential applications, although they do comment that such inequality systems arise in Lyapunov theory. Given Bellman’s legendary knowledge of system and control theory, and especially Lyapunov theory, it is tempting to conjecture that Bellman was aware of the potential applications. But so far we have found no evidence for this. Relevant work includes Cullum et al. [CDW75], Craven and Mond [CM81], Polak and Wardi [PW82], Fletcher [Fle85], Shapiro [Sha85], Friedland et al. [FNO87], Goh and Teo [GT88], Panier [Pan89], Allwright [All89], Overton and Womersley [Ove88, Ove92, OW93, OW92], Ringertz [Rin91], Fan and Nekooie [FN92, Fan93] and Hiriart–Urruty and Ye [HUY92]. LMIPs are solved using various algorithms for convex optimization in Boyd and Yang [BY89], Pyatnitskii and Skorodinskii [PS83], Kamenetskii and Pyatnitskii [KP87a, KP87b]. A survey of methods for solving problems involving LMIs used by researchers in control theory can be found in the paper by Beck [Bec91]. The software packages [BDG+ 91] and [CS92a] use convex programming to solve many robust control problems. Boyd [Boy94] outlines how interior-point convex optimization algorithms can be used to build robust control software tools. Convex optimization problems involving LMIs have been used in control theory since about 1985: Gilbert’s method [Gil66] was used by Doyle [Doy82] to solve a diagonal scaling problem; another example is the musol program of Fan and Tits [FT86]. Among other methods used in control to solve problems involving multiple LMIs is the “method of alternating convex projections” described in the article by Grigoriadis and Skelton [GS92]. This method is essentially a relaxation method, and requires an analytic expression for the projection onto the feasible set of each LMI. It is not a polynomial-time algorithm; its complexity depends on the problem data, unlike the ellipsoid method or the interior-point methods described in [NN94], whose complexities depend on the problem size only. However, the alternating projections method is reported to work well in many cases.

**Interior-point methods for LMI problems
**

Interior-point methods for various LMI problems have recently been developed by several researchers. The ﬁrst were Nesterov and Nemirovskii [NN88, NN90b, NN90a, NN91a, NN94, NN93]; others include Alizadeh [Ali92b, Ali91, Ali92a], Jarre [Jar93c], Vandenberghe and Boyd [VB93b], Rendl, Vanderbei, and Wolkowocz [RVW93], and Yoshise [Yos94]. Of course, general interior-point methods (and the method of centers in particular) have a long history. Early work includes the book by Fiacco and McCormick [FM68], the method of centers described by Huard et al. [LH66, Hua67], and Dikin’s interior-point method for linear programming [Dik67]. Interest in interior-point methods surged in 1984 when Karmarkar [Kar84] gave his interior-point method for solving linear programs, which appears to have very good practical performance as well as a good worst-case complexity bound. Since the publication of Karmarkar’s paper, many researchers have studied interior-point methods for linear and quadratic programming. These methods are often described in such a way that extensions to more general (convex) constraints and objectives are not clear. However, Nesterov and Nemirovskii have developed a theory of interior-point methods that applies to more general convex programming problems, and in particular, every problem that arises in this book; see [NN94]. In particular, they derive complexity bounds for many diﬀerent interior-point algorithms, including the method of centers, Nemirovskii’s projective algorithm and primal-dual methods. The most eﬃcient algorithms seem to be Nemirovskii’s projective algorithm and primal-dual methods (for the case of linear programs, see [Meh91]). Other recent articles that consider interior-point methods for more general convex programming include Sonnevend [Son88], Jarre [Jar91, Jar93b], Kortanek et al. [KPY91], Den Hertog, Roos, and Terlaky [DRT92], and the survey by Wright [Wri92]. Copyright c 1994 by the Society for Industrial and Applied Mathematics.

Notes and References

31

Interior-point methods for GEVPs are described in Boyd and El Ghaoui [BE93] (variation on the method of centers), and Nesterov and Nemirovskii [NN91b] and [NN94, §4.4] (a variation on Nemirovskii’s projective algorithm). Since GEVPs are not convex problems, devising a reliable stopping criterion is more challenging than for the convex problems LMIP, EVP, and CP. A detailed complexity analysis (in particular, a statement and proof of the polynomial-time complexity of GEVPs) is given in [NN91b, NN94]. See also [Jar93a]. Several researchers have recently studied the possibility of switching from an interior-point method to a quadratically convergent local method in order to improve on the ﬁnal convergence; see [Ove92, OW93, OW92, FN92, NF92]. Interior-point methods for CPs and other related extremal ellipsoid volume problems can be found in [NN94, §6.5].

**Interior-point methods and problem structure
**

Many researchers have developed algorithms that take advantage of the special structure of the least-squares problems arising in interior-point methods for linear programming. As far as we know, the ﬁrst (and so far, only) interior-point algorithm that takes advantage of the special (Lyapunov) structure of an LMI problem arising in control is described in Vandenberghe and Boyd [VB93b, VB93a]. Nemirovskii’s projective method can also take advantage of such structure; these two algorithms appear to be the most eﬃcient algorithms developed so far for solving the LMIs that arise in control theory.

**Software for solving LMI problems
**

Gahinet and Nemirovskii have recently developed a software package called LMI-Lab [GN93] based on an earlier FORTRAN code [NN90a], which allows the user to describe an LMI problem in a high-level symbolic form (not unlike the formulas that appear throughout this book!). LMI-Lab then solves the problem using Nemirovskii’s projective algorithm, taking advantage of some of the problem structure (e.g., block structure, diagonal structure of some of the matrix variables). Recently, El Ghaoui has developed another software package for solving LMI problems. This noncommercial package, called LMI-tool, can be used with matlab. It is available via anonymous ftp (for more information, send mail to elghaoui@ensta.fr). Another version of LMItool, developed by Nikoukhah and Delebecque, is available for use with the matlab-like freeware package scilab; in this version, LMI-tool has been interfaced with Nemirovskii’s projective code. scilab can be obtained via anonymous ftp (for more information, send mail to Scilab@inria.fr). A commercial software package that solves a few specialized control system analysis and design problems via LMI formulation, called optin, was recently developed by Olas and Associates; see [OS93, Ola94].

**Reduction to a strictly feasible LMI
**

In this section we prove the following statement. Let F0 , . . . , Fm ∈ Rn×n be symmetric matrices. Then there is a matrix A ∈ Rm×p with p ≤ m, a vector b ∈ Rm , and symmetric ˜ ˜ matrices F0 , . . . , Fp ∈ Rq×q with q ≤ n such that:

p ∆ ˜ ˜ F (x) ≥ 0 if and only if x = Az + b for some z ∈ Rp and F (z) = F0 +

i=1

˜ zi Fi ≥ 0.

˜ In addition, if the LMI F (x) ≥ 0 is feasible, then the LMI F (z) ≥ 0 is strictly feasible. This electronic version is for personal use and may not be duplicated or distributed.

32

Consider the LMI

m

Chapter 2

Some Standard Problems Involving LMIs

F (x) = F0 +

i=1

xi Fi ≥ 0,

(2.42)

where Fi ∈ Rn×n , i = 0, . . . , m. Let X denote the feasible set {x | F (x) ≥ 0}. ˜ ˜ If X is empty, there is nothing to prove; we can take p = m, A = I, b = 0, F0 = F0 , . . . , Fp = Fp . Henceforth we assume that X is nonempty. If X is a singleton, say X = {x0 }, then with ˜ ˜ p = 1, A = 0, b = x0 , F0 = 1, and F1 = 0, the statement follows. Now consider the case when X is neither empty nor a singleton. Then, there is an aﬃne subspace A of minimal dimension p ≥ 1 that contains X. Let a1 , . . . ap a basis for the linear part of A. Then every x ∈ A can be written as x = Az + b where A = [a1 · · · ap ] is full-rank p and b ∈ Rm . Deﬁning G0 = F (b), Gi = F (ai )−F0 , i = 1, . . . , p, and G(z) = G0 + i=1 zi Gi , p we see that F (x) ≥ 0 if and only if there exists z ∈ R satisfying x = Az + b and G(z) ≥ 0. Let Z = {z ∈ Rp | G(z) ≥ 0}. By construction, Z has nonempty interior. Let z0 a point in the interior of Z and let v lie in the nullspace of G(z0 ). Now, v T G(z)v is a nonnegative aﬃne function of z ∈ Z and is zero at an interior point of Z. Therefore, it is identically zero over Z, and hence over Rp . Therefore, v belongs to the intersection B of the nullspaces of the Gi , i = 0, . . . , p. Conversely, any v belonging to B will satisfy v T G(z)v = 0 for any z ∈ Rp . B may therefore be interpreted as the “constant” nullspace of the LMI (2.42). Let q be the dimension of B. If q = n (i.e., G0 = · · · = Gp = 0), then obviously F (x) ≥ 0 if and only if x = Az + b. In this ˜ ˜ case, the statement is satisﬁed with F0 = I ∈ Rp×p , Fi = 0, i = 1, . . . , p.

∆

If q < n, let v1 , . . . , vq a basis for B. Complete the basis v1 , . . . , vq by vq+1 , . . . , vn to ˜ ˜ obtain a basis of Rn . Deﬁne U = [vq+1 · · · vn ], Fi = U T Gi U , i = 1, . . . , p and F (z) = p ˜i . Then, the LMI G(z) ≥ 0 is equivalent to the LMI F (z) ≥ 0, and by ˜ ˜0 + F zF i=1 i ˜ construction, there exists z0 such that F (z0 ) > 0. This concludes the proof.

**Elimination procedure for matrix variables
**

We now prove the matrix elimination result stated in §2.6.2: Given G, U , V , there exists X such that G + U XV T + V X T U T > 0 if and only if ˜ ˜ U T GU > 0, ˜ ˜ V T GV > 0 (2.44) (2.43)

˜ ˜ holds, where U and V are orthogonal complements of U and V respectively. It is obvious that if (2.43) holds for some X, so do inequalities (2.44). Let us now prove the converse. Suppose that inequalities (2.44) are feasible. Suppose now that inequality (2.43) is not feasible for any X. In other words, suppose that G + U XV T + V X T U T > 0, for every X. By duality, this is equivalent to the condition there exists Z = 0 with Z ≥ 0, V T ZU = 0, and Tr GZ ≤ 0. Copyright c 1994 by the Society for Industrial and Applied Mathematics. (2.45)

and we have ˜ ˜ ˜ ˜ Z = RRT = (RT )(RT )T = AAT + BB T = V HH T V T + U KK T U T . Pac94. See also the book by Hestenes [Hes81. Let R be a Cholesky factor of Z. For a complete discussion and references. ˜ ˜ matrices A.46) for some matrices H. This means that there exists a unitary matrix T . . The name “S-procedure” was introduced much later by Aizerman and Gantmacher in their 1964 This electronic version is for personal use and may not be duplicated or distributed. K. i. due to Parrott [Par78]. B such that V T A = 0. when Lur’e and Postnikov used it to prove the stability of some particular nonlinear systems.Notes and References Now let us show that Z ≥ 0 and V T ZU = 0 imply that ˜ ˜ ˜ ˜ Z = V HH T V T + U KK T U T 33 (2. The equivalence between the conditions ˜ ˜ U (z)T G(z)U (z) > 0 and G(z) − σU (z)U (z)T > 0 for some real scalar σ is through Finsler’s lemma [Fin37] (see also §3.44) implies that Tr GZ > 0 for Z of the form (2. The elimination lemma is related to a matrix dilation problem considered in [DKW82]. Gah92. and matrices M and N such that V T RT = [0 M ] and U T RT = [N 0] . KR88. The condition V T ZU = 0 is equivalent to V T RRT U = (V T R)(U T R)T = 0. which is the desired result. p78-86] for proofs of various S-procedure results.e. This will ﬁnish our proof. we refer the reader to the survey article by Uhlig [Uhl79]. Finsler’s lemma has also been directly used to eliminate variables in certain matrix inequalities (see for example [PH86. it is closely related to the S-procedure. which was used in control problems in [PZPB91. Iwa93] (see also the Notes and References of Chapter 7). In other words.6 of [Sch73]). which states that if xT Qx > 0 for all nonzero x such that xT Ax = 0. where the number of columns of M and N add up to that of R. then there exists a real scalar such that Q − σA > 0. Z = RRT . since (2. The S-procedure The problem of determining if a quadratic form is nonnegative when other quadratic forms are nonnegative has been studied by mathematicians for at least seventy years. IS93a. We also note another result on elimination of matrix variables.46).45). PZP+ 92. since otherwise Z = 0). B can be written A = V H. BPG89b]). RT can be written as RT = [A B] ˜ ˜ for some matrices A. real matrices. The application of the S-procedure to control problems dates back to 1944. K (at least one of which is nonzero. p354-360] and Horn and Johnson [HJ91. which contradicts Tr GZ ≤ 0 in (2. where Q and A are symmetric. From the deﬁnition of V and U . U T B = 0. B = U K for some matrices H.2.

i. and v0 depend aﬃnely on some variable ν. Yakubovich [Yak62. described in several books on control and systems theory. is as hard as solving a general indeﬁnite quadratic program. We could determine whether there exists ν such that (2. B) is controllable and D + D T > 0. the PR Lemma states that the transfer function c(sI − A)−1 b of the single-input single-output minimal system (A. ˙ is passive. We might therefore imagine that the problem of checking whether there exists ν such that (2. FY79]. Copyright c 1994 by the Society for Industrial and Applied Mathematics. satisﬁes T 0 y = Cx + Du. Complexity of S-procedure condition Suppose that T0 . B. Indeed. the problem is that ﬁnding a cutting plane (which is required for each step of the ellipsoid algorithm) is itself an NP-complete problem. In its original form. merely verifying that the condition (2. Yak77. Yak73].34 Chapter 2 Some Standard Problems Involving LMIs book [AG64]. Kal63b] established the connection between the Popov criterion and the existence of a positivedeﬁnite matrix satisfying certain matrix inequalities. The following statements are equivalent: 1. i.47) if and only if there exists P > 0 such that AT P + P A ≤ 0 and P b = cT . Recent and important developments about the S-procedure can be found in [Yak92] and references therein. The system x = Ax + Bu.30) is convex in ν. the condition (2.47) can be checked graphically. Faurre and Depeyrot [FD77].e. and Willems [Wil70]. Popov gave the famous Popov frequency-domain stability criterion for the absolute stability problem [Pop62]. since then Yakubovich has given more general formulations and corrected errors in some earlier proofs [Yak77. certain quadratic optimal control problems and the existence of symmetric solutions to the ARE. And73] extended the PR lemma to multi-input multi-output systems. which is NP-complete. and D such that all eigenvalues of A have negative real part.30) holds has low complexity since it can be cast as a convex feasibility problem. by verifying that the Nyquist plot of the “linear part” of the nonlinear system was conﬁned to a speciﬁc region in the complex plane. (A. Wil74a] described connections between the PR lemma. e.e. Yak64] and Kalman [Kal63a. Anderson [And67. Positive-real and bounded-real lemma In 1961. In fact. ui . and derived similar results for nonexpansive systems [And66a]. Willems [Wil71b. The proof of the two S-procedure results described in this chapter can be found in the articles by Yakubovich [FY79. LMIs and AREs can be summarized as follows.30) holds with polynomially many steps of the ellipsoid algorithm. for ﬁxed ζ. Vidyasagar [Vid92. Re c(sI − A)−1 b ≥ 0 for all Re s > 0 (2. The PR lemma is now standard material.30) holds for ﬁxed data Ti . Brockett [Bro70]. Narendra and Taylor [NT73]. The PR lemma is also known by various names such as the Yakubovich–Kalman–Popov–Anderson Lemma or the Kalman–Yakubovich Lemma. x(0) = 0 u(t)T y(t)dt ≥ 0 for all u and T ≥ 0. the constraint F0 (ζ) ≥ 0 is a linear constraint on ν and the constraint (2. ch5–7]. and vi . u0 . . The connections between passivity. it is in [Wil71b] that we ﬁnd Willems’ quote on the role of LMIs in quadratic optimal control (see page 3). The condition (2. Popov’s criterion could be checked via graphical means. pp474–478]. c) is positive-real..g. And66b. C. This is wrong. b.. We consider A.30) is simply an inﬁnite number of these linear constraints. Anderson and Vongpanitlerd [AV73.

it is explicitly stated in Potter [Pot66]. The transfer matrix H(s) = C(sI − A)−1 B + D is positive-real. The LMI −(D + D T ) ≤0 (2. 3. see the articles by Wen [Wen88]. AV73]. this method is an obvious variation of Laub’s algorithm [Lau79. An excellent discussion of this result and its connections with spectral factorization theory can be found in [AV73]. Origins of this result can be traced back to Reid [Rei46] and Levin [Lev59]. It is also possible to check passivity (or nonexpansivity) of a system symbolically using Sturm methods. Lozano–Leal and Joshi [LLJ90]. There exists P = P T satisfying the ARE AT P + P A + (P B − C T )(D + D T )−1 (P B − C T )T = 0. interconnected systems [MH78] and stochastic processes and stochastic realization theory [Fau67. Pot66.e. 2. and 4 can be found in Theorem 4 of [Wil71b] and also [Wil74a]. . AL84]. AM74]. Sil73] and Boyd.2. see Siljak [Sil71. A less stable numerical method can be found in [Rei46. i. or Van Dooren’s [Doo81]. (2. 4. AT P + P A BT P − C P B − CT 35 3. is in [BBK89].g. which yields Routh–Hurwitz like algorithms. see also [Fra87. H(s) + H(s)∗ ≥ 0 for all s with Re s ≥ 0.. Fau73. A − B(D + D T )−1 C −C (D + D ) T T −1 B(D + D T )−1 B T −A + C T (D + DT )−1 B T T C This electronic version is for personal use and may not be duplicated or distributed. The PR lemma is used in many areas.Notes and References 2. The method for constructing Pare from the stable eigenspace of M .48) and the solutions of the ARE (2. The equivalence of 5 is found in. described in §2. The sizes of the Jordan blocks corresponding to the pure imaginary eigenvalues of the Hamiltonian matrix M= are all even. Lev59. Connections between the extremal points of the set of solutions of the LMI (2.48) in the variable P = P T is feasible. e. For other discussions of strict passivity. Rob89].7. The equivalence of 1.. Balakrishnan. BBK89.49) 5.49) are explored in [Bad82]. Theorem 1 of [LR80]. for example. and Kabamba [BBK89].

.

Let us ﬁx γ > 1. i. which is the same as (RRT )−1 ≤ M T (LT L)M ≤ γ 2 (RRT )−1 .. diagonal L ∈ Rp×p and R ∈ Rq×q and µ > 0 such that µI ≤ (LM R)T (LM R) ≤ µγ 2 I.1 Minimizing Condition Number by Scaling The condition number of a matrix M ∈ Rp×q . with p ≥ q.1) where L and R are the optimization variables. Q ∈ Rq×q with P > 0. For square invertible matrices this reduces to κ(M ) = M M −1 . 37 (3. diagonal and nonsingular R ∈ Rq×q .2) This is equivalent to the existence of diagonal P ∈ Rp×p . We will show that this problem can be transformed into a GEVP.3) . We assume without loss of generality that p ≥ q and M is full-rank. There exist nonsingular. diagonal L ∈ R and R ∈ Rq×q such that I ≤ (LM R)T (LM R) ≤ γ 2 I.Chapter 3 Some Matrix Problems 3. √ Since we can absorb the factor 1/ µ into L. (3. and the matrix M ∈ Rp×q is given. diagonal and nonsingular (3. and κ(M ) = ∞ otherwise. diagonal L ∈ Rp×p and R ∈ Rq×q such that κ(LM R) ≤ γ if and only if there are nonsingular. Q > 0 and Q ≤ M T P M ≤ γ 2 Q. is the ratio of its largest and smallest singular values. We consider the problem: minimize subject to κ (LM R) L ∈ Rp×p .e. κ(M ) = ∆ λmax (M T M ) λmin (M T M ) 1/2 for M full-rank. this is equivalent to the existence of p×p nonsingular.

This problem can be reformulated as the GEVP minimize subject to γ F (x) > 0. Complex matrices are readily handled by substituting M ∗ (complex-conjugate transpose) for M T .38 Chapter 3 Some Matrix Problems To see this.3) holds with P = LT L and Q = (RRT )−1 . the constraint that one or more blocks are equal. ˜ I < νM0 + i=1 xi Mi < γI ˜ 3. and more generally. and γ).e. µI < M (x) < γµI (the variables here are x. ﬁrst suppose that L ∈ Rp×p and R ∈ Rq×q are nonsingular and diagonal.4) µ > 0. Hence we can solve (3. and (3. We will show that ν(M ) can be computed by solving a GEVP. Then. Q ≤ M T P M ≤ γ2Q P > 0.2 Minimizing Condition Number of a Positive-Deﬁnite Matrix A related problem is minimizing the condition number of a positive-deﬁnite matrix M that depends aﬃnely on the variable x. Conversely. x = x/µ. Suppose m m M (x) = M0 + i=1 x i Mi . µ. largest singular value) of the matrix M . Q>0 Remark: This result is readily extended to handle scaling matrices that have a given block-diagonal structure. 3. (3.1) by solving the GEVP: minimize subject to γ2 P ∈ Rp×p and diagonal.3 Minimizing Norm by Scaling ∆ The optimal diagonally scaled norm of a matrix M ∈ Cn×n is deﬁned as ν(M ) = inf DM D−1 D ∈ Cn×n is diagonal and nonsingular .. F (x) = F0 + i=1 x i Fi . Q > 0. .4) as the EVP with ˜ variables ν = 1/µ.3) holds for diagonal P ∈ Rp×p and Q ∈ Rq×q with P > 0. Q ∈ Rq×q and diagonal. we can express (3. suppose that (3.2) holds for L = P 1/2 and R = Q−1/2 . subject to the LMI constraint F (x) > 0. (3. Copyright c 1994 by the Society for Industrial and Applied Mathematics. We can reformulate this GEVP as an EVP as follows. where M is the norm (i. x = x/µ and γ: ˜ minimize subject to γ m m νF0 + i=1 xi Fi > 0.2) holds. Deﬁning the new variables ν = 1/µ. Then (3.

. . For example. and nonsingular we simply compute the optimal value of the GEVP minimize subject to γ P > 0 and diagonal. nonsingular D M ∗ P M ≤ γ 2 P for some diagonal P = P T > 0 . This is true if and only if the LMI M ∗ D + DM > 0. the more general case of block-diagonal similarity-scaling. . i = 1. which plays an important role in the stability analysis of linear systems with uncertain parameters. As another example. .6) is feasible. ∗ 39 inf γ inf γ Therefore ν(M ) is the optimal value of the GEVP minimize subject to γ P > 0 and diagonal. . Mi∗ P Mi < γ 2 P. M ∗ D + DM > 0. . we can solve the problem of simultaneous optimal diagonal scaling of several matrices M1 . M ∗P M < γ2P Remark: This result can be extended in many ways. is readily handled. . is inf γ P > 0. we can ﬁnd a feasible scaling D with the smallest condition number by solving the EVP minimize subject to µ D diagonal. and ask whether there is a diagonal matrix D > 0 such that the Hermitian part of DM is positive-deﬁnite. . . nonsingular D M ∗ D∗ DM ≤ γ 2 D∗ D for some diagonal. When this LMI is feasible. p DMi D−1 D ∈ Cn×n is diagonal . . . with equality constraints among the blocks.5) P and G diagonal and real which can also be computed by solving a GEVP. So determining whether a matrix can be rescaled positive-deﬁnite is an LMIP. I < D < µ2 I This electronic version is for personal use and may not be duplicated or distributed. . (3. p Another closely related quantity.4 Rescaling a Matrix Positive-Definite Note that ν(M ) = = = inf γ DM D−1 DM D−1 < γ 2 I for some diagonal.3. . . 3. M ∗ P M + i(M ∗ G − GM ) < γ 2 P. Mp : To compute inf max i = 1. D>0 (3.4 Rescaling a Matrix Positive-Deﬁnite We are given a matrix M ∈ Cn×n .

g. Evidently this is an LMIP. We can assume that the diagonal entries are speciﬁed (otherwise we simply make them very large. As another example. and we are to complete the matrix to make it positive-deﬁnite. Thus the matrix A −1 has a zero entry in every location corresponding to an unspeciﬁed entry in the original matrix.e. We can easily recover some known results about this problem. in other words.1). Many of these problems can be cast as LMIPs. i. i. For entries that are completely speciﬁed.3. Then the set of all positive-deﬁnite completions. we consider an “interval matrix”. etc.. the lower and upper bound coincide. if nonempty. Copyright c 1994 by the Society for Industrial and Applied Mathematics. for entries that are completely free the lower bound is −∞ and the upper bound is +∞. suppose we are given a set of matrices that have some given sparsity pattern. As a more speciﬁc example..4.2).. Hence the associated LMI has an analytic center A (see §2. suppose we are given upper and lower bounds for some of the correlation coeﬃcients of a stationary time-series. see §2. see the Notes and References and §10. This problem is also readily cast as an LMIP. subject to the bounds on each entry. But of course an arbitrary positive-deﬁnite completion problem is readily solved as an LMIP. The matrix A is characterized by Tr A −1 B = 0 for every B = B T that has zero entries where A has speciﬁed entries (see §2.e.6. This is an LMIP. Suppose that for each entry of the matrix we specify an upper and lower bound. . results in a positive-deﬁnite (or contractive. We want to determine a matrix with the complementary sparsity pattern that. and might be useful in applications.6) can also be interpreted as the condition for the existence of a diagonal quadratic Lyapunov function that proves stability of −M . to make it positive-deﬁnite (or a contraction. Depending on the pattern of speciﬁed entries.e. e. have norm less than one) is readily formulated as an LMIP. We must determine these free entries so that the Toeplitz matrix is positive-deﬁnite. we have a Toeplitz matrix with some entries within given intervals.) matrix. while the others are completely free (unspeciﬁed). it has the same sparsity structure as the original matrix. 3. i. when added to each of the given matrices.4. We might call this the “simultaneous matrix completion problem” since we seek a completion that serves for multiple original matrices. The problem of completing the matrix. eliminate them. Let us mention a few simple generalizations of this matrix completion problem that have not been considered in the literature but are readily solved as LMIPs. this condition can lead to an “analytic solution” of the problem. In other words.40 Chapter 3 Some Matrix Problems Remark: The LMI (3. is bounded.1). The simplest example is positive-deﬁnite completion: Some entries of a symmetric matrix are given. is Toeplitz and positive-deﬁnite. for any choice of the speciﬁed elements in the intervals. and must determine some complementary correlation coeﬃcients that are consistent with them..5 Matrix Completion Problems In a matrix completion problem some entries of a matrix are ﬁxed and the others are to be chosen so the resulting (“completed”) matrix has some property.

.7) holds for some P > 0 by solving the EVP minimize subject to α T vi P vi ≤ α. ∆ √ For any P > 0. . .. . vL be the vertices of the unit ball Bpl of · pl ..6 Quadratic Approximation of a Polytopic Norm 41 3. L. . i = 1. . . . Indeed. . . vL } P. . Therefore we can minimize α such that (3. .. p.√ = 1. We will show that the problem of ﬁnding P that minimizes α. . . is an EVP. by computing the maximum volume ellipsoid that lies inside the unit ball of · pl . . . the quadratic norm z T P z approximates z pl within a factor of α. . P aT i ai α ≥ 0. i and the second. . Therefore the result is uninteresting from the point of view of computational complexity and of limited practical use except for problems with low dimensions. . Thus.p aT z .7) is equivalent to √ √ 1/ αBP ⊆ Bpl ⊆ αBP . . 1/ αBP ⊆ Bpl . i = 1. Bpl T vi P vi ≤ α. we can ﬁnd (in polynomiali time) a norm for which α is less than n.. i. .6 Quadratic Approximation of a Polytopic Norm · pl Consider a piecewise linear norm z pl on Rn deﬁned by = max ∆ i=1.e. . Let v1 . . . . . . √ ⊆ αBP .3. . √ for some constant α ≥ 1.. . L. p. and let BP denote the unit ball of · √ √ 1/ α z P ≤ z pl ≤ α z P Then (3. . (which does not require us to ﬁnd the vertices vi . . √ The optimal factor α in this problem never exceeds n. the problem of determining the optimal quadratic norm approximation of · pl . the quadratic norm deﬁned by z P = z T P z = P 1/2 z satisﬁes √ √ 1/ α z P ≤ z pl ≤ α z P for all z. is equivalent to i = 1. so that Bpl = {z | z ∆ pl ≤ 1 } = Co {v1 . L). See the Notes and References for more details. . i where ai ∈ Rn for i = 1. This electronic version is for personal use and may not be duplicated or distributed. √ The ﬁrst inclusion. is equivalent to aT P −1 ai ≤ α. i = 1. . p Remark: The number of vertices L of the unit ball of · pl can grow exponentially in p and n.

Given a representation (3. We describe an ellipsoid E in two diﬀerent ways.). Indeed. (This is the same problem as described above. twice the maximum semi-axis length) is 2 (bT A−1 b − c)λmax (A−1 ).. and its diameter is 2λmax (P ). so it is unlikely that the problem can be reduced (polynomially) to an LMI problem. . In some of these cases.9) Copyright c 1994 by the Society for Industrial and Applied Mathematics. b.9) with P = bT A−1 b − c A−1/2 . T T T (x) = xT Ax + 2xT b + c. Note that this description is homogeneous. We saw in §2. but with a diﬀerent description of the polytope. i. p }. As an example consider the problem of ﬁnding the ellipsoid centered around the origin of smallest volume that contains a polytope described by its vertices. although the approximations may be good on “typical” problems. and hence solved exactly. We will also describe an ellipsoid as the image of the unit ball under an aﬃne mapping with symmetric positive-deﬁnite matrix: where P = P T > 0. n. Each of these descriptions of E is readily converted into the other. It also suggests that the approximations obtained by solving a CP or EVP will not be good for all instances of problem data. . E= x (x − xc )T P −2 (x − xc ) ≤ 1 = { P z + xc | z ≤ 1 } . and hence of some use in practice. which is NP-complete. i. and addition..) This problem is NP-hard. xc = −A−1 b. see the Notes and References. (This representation is unique. The diameter of E (i. . etc. consider the problem of simply verifying that a given. we can scale A. .. This problem is equivalent to the general concave quadratic programming problem. In some cases the problem can be cast as a CP or an EVP. . intersection.4 that this problem can be cast as a CP. and c) we form the representation (3.e. we can compute an approximation of the optimal ellipsoid by solving a CP or an EVP. The ﬁrst description uses convex quadratic functions: E = {x | T (x) ≤ 0. In this section we consider subsets formed from ellipsoids E1 . Ep in various ways: union. To pose such a problem precisely we need to know how the subset is described.7 Ellipsoidal Approximation The problem of approximating some subset of Rn with an ellipsoid arises in many ﬁelds and has a long history. whether we seek an inner or outer approximation. b and c by any positive factor without aﬀecting E. and how the approximation will be judged (volume. The volume of E is given by vol(E)2 = β det (bT A−1 b − c)A−1 where β is a constant that depends only on the dimension of x. i.e.42 Chapter 3 Some Matrix Problems 3. .e.e. P and xc are uniquely determined by the ellipsoid E.. ﬁxed ellipsoid contains a polytope described by a set of linear inequalities.. i = 1. the problem is known to be NP-hard. . In other cases. −1 (3. As an example consider the problem of ﬁnding the ellipsoid of smallest volume that contains a polytope described by a set of linear inequalities. { x | a T x ≤ i bi . i. (3.e. major or minor semi-axis.e. We approximate these sets by an ellipsoid E 0 . .2.8) (i..8) where A = A > 0 and b A b − c > 0 (which ensures that E is nonempty and does not reduce to a single point). } . .) The volume of E is then proportional to det P . A.

) complements. .11) with variables A0 . c = xT P −2 xc − 1.) 3. Co i=1 Ei ). . . b = −P −2 xc . . Since our representation of E0 is homogeneous. b0 and c0 in a convenient way: such that bT A−1 b0 − c0 = 1. . . . i = 1. .8) of E with A = P −2 . . in fact. i = 1. we can form every representation of E by scaling these A. i. be positive.e. p. τp such that (3. . .3. T0 (x) − τi Ti (x) ≤ 0. τp such that for every x. . . the condition (3. . 0 0 (3. . . every x such that Ti (x) ≤ 0 satisﬁes T0 (x) ≤ 0. 0 Thus we can ﬁnd the smallest volume ellipsoid containing the union of ellipsoids E 1 . (Since A0 > 0. this is true if and only if there exist nonnegative scalars τ1 .12) holds.10) is equivalent to the existence of nonnegative τ1 . . . . This is true if and only if. . . In other words we set 0 0 c0 = bT A−1 b0 − 1 0 0 and parametrize E0 by A0 and b0 alone. (3. p. . . We will describe these ellipsoids via the associated quadratic functions Ti (x) = xT Ai x + 2xT bi + ci . . . p. and τ1 .12) This electronic version is for personal use and may not be duplicated or distributed. we can form a representation (3. b. . . τ1 ≥ 0. b0 .9). . τp ≥ 0. we obtain Ai b0 0 −1 bT − τi bT 0 i 0 b0 −A0 the equivalent LMI bi 0 ci 0 ≤ 0. . Thus our condition becomes: A0 bT 0 Using Schur A0 T b0 0 b0 T −1 b0 A 0 b0 −1 − τi Ai bT i bi ci ≤ 0. . . i = 1.. Ep by solving the CP (with variables A0 . equivalently. b0 . (3. . c by positive factors. . i = 1. . . . or. the convex hull of the union. To summarize. . it follows that the τi must. . . the volume of E0 is proportional to det A−1 . . Ep (or equivap lently. . . . By the S-procedure. we will now normalize A0 . for each i. τp . p. τp ) minimize subject to log det A−1 0 A0 > 0.10) if and only if Ei ⊆ E0 for i = 1. .12) With the normalization (3.7 Ellipsoidal Approximation 43 Given the representation (3. . and τ1 . . p. . We have p E0 ⊇ i=1 Ei (3. . .11).7. c (This forms one representation of E. such that A0 bT 0 b0 c0 − τi Ai bT i bi ci ≤ 0.1 Outer approximation of union of ellipsoids We seek a small ellipsoid E0 that covers the union of ellipsoids E1 .

g. . .15) Remark: The intersection i=1 Ei is empty or a single point if and only if there exist nonnegative τ1 . . Ep . not all zero. . . given E0 .11) holds..14) This LMI is suﬃcient but not necessary for (3. 0 which is equivalent to solving the EVP λ A0 > λI. .. .13) holds.44 Chapter 3 Some Matrix Problems p This ellipsoid is called the L¨wner–John ellipsoid for i=1 Ei . maximize subject to 3. τ1 ≥ 0. . b0 −1 bT − 0 i i=1 0 b0 −A0 0 0 0 (3. i. τp . . Once again we normalize our representation of E0 so that (3. such that p τi i=1 Ai bT i bi ci ≥ 0. . . . so we do not expect to recast it as one of our LMI problems. we do not expect to cast the problem of ﬁnding the smallest volume ellipsoid covering the intersection of some ellipsoids as a CP. . . which can be written as the LMI (in variables A0 .e. A fortiori.13) to hold. . .2 Outer approximation of the intersection of ellipsoids Simply verifying that p E0 ⊇ i=1 Ei (3. τp ≥ 0. E1 . and τ1 . . aﬃne invariance. From the S-procedure we obtain the condition: there exist positive scalars τ1 . τp ): A 0 b0 0 A i bi 0 p T τi bT ci 0 ≤ 0. Ep .e.7. τp ) minimize subject to log det A−1 0 A0 > 0. b0 . We can ﬁnd the ellipsoid (or equivalently.14) (3. (3. . .15) is unbounded below. τp such that A0 bT 0 b0 −1 bT A 0 b0 − 1 0 p − τi i=1 Ai bT i bi ci ≤ 0. . . .13) to hold. . sphere) of smallest diameter that contains Ei by minimizing λmax (A−1 ) subject to the constraints of the previous CP.12) √ The optimal diameter is 2/ λopt . it characterizes some but not all of the ellipsoids that cover the intersection of E1 . the one of smallest volume) by solving the CP (with variables A0 .. τp ≥ 0. and τ1 . is NP-complete. . . In this case (3. . and satisﬁes several o nice properties (e. certain bounds on how well it approximates Co Ei ). Copyright c 1994 by the Society for Industrial and Applied Mathematics. . . p (3. τ1 ≥ 0. see the Notes and References. We can ﬁnd the best such outer approximation (i. where λopt is an optimal value of the EVP above. . . . b0 . We will describe three ways that we can compute suboptimal ellipsoids for this problem by solving CPs. The ﬁrst method uses the S-procedure to derive an LMI that is a suﬃcient condition for (3. . .

9). Suppose that E1 .1). .. . . (x − xi )T Pi−2 (x − xi ) −2 −λi (x − x0 )T P0 (x − x0 ) ≤ 1 − λi . . Ei . . The LMI F (x) = diag I −1 (x − x1 )T P1 −1 P1 (x − x1 ) 1 . . . .. Using the S-procedure. . xp and P1 . Ep are given in the form (3. . . See the Notes and References for more discussion. this is equivalent to the existence of nonnegative λ 1 . Let x denote its analytic center. i. .3 Inner approximation of the intersection of ellipsoids The ellipsoid E0 is contained in the intersection of the ellipsoids E1 . .3. λp satisfying for every x. when E0 is enlarged by a factor of 3p + 1.. it covers References for more discussion of this. .. .. This electronic version is for personal use and may not be duplicated or distributed. Then it can be shown that p E0 = x ∆ (x − x )T H(x − x ) ≤ 1 ⊆ i=1 Ei . .4. ..e. . . p i=1 Thus. . Yet another method for producing a suboptimal solution is based on the idea of the analytic center (see §2. .7. is nonempty and not a single point). we have (x − xi )T Pi−2 (x − xi ) ≤ 1. . (3. . . . with x1 . Pp . In fact we can give a bound on how poorly E0 approximates the intersection. .e. which can be cast as a CP (see the next section).. Ep if and only if for every x satisfying −2 (x − x0 )T P0 (x − x0 ) ≤ 1.7 Ellipsoidal Approximation 45 Another method that can be used to ﬁnd a suboptimal solution for the problem of determining the minimum volume ellipsoid that contains the intersection of E 1 .. . We assume now that the p intersection i=1 Ei has nonempty interior (i. x = arg min log det F (x)−1 F (x) > 0 and let H denote the Hessian of log det F (x)−1 at x . I −1 (x − xp )T Pp p 1 >0 in the variable x has as feasible set the interior of i=1 Ei . . See the Notes and 3. . −1 Pp (x − xp ) . It can be shown that p x (x − x )T H(x − x ) ≤ (3p + 1)2 ⊇ i=1 Ei . . . p. . i = 1. Ep is to ﬁrst compute the maximum volume ellipsoid that is contained in the intersection.16) i = 1. If this ellipsoid is scaled by a factor of n about it center then it is guaranteed to contain the intersection. . p.

b i i = 0. and deﬁne p Once again we normalize our representation of E0 so that (3. .7. . x = . (3. xp such that Ti (xi ) ≤ 0. . p. xp p p E0 = i=1 Ei .46 Chapter 3 Some Matrix Problems We prove in the Notes and References section that (3. i = 1. The ellipsoid E0 contains A if and only if for every x1 . i = 1.18) Then both T0 ( i=1 xi ) and i=1 τi Ti (xi ) are quadratic functions of x. . (xi − x0 )T λi − 1 P0 0 −λi I (3. . 3. λ1 ≥ 0. (3. By application of the S-procedure. i = 1. . . this yields the sphere of largest diameter contained in the intersection. λp ) minimize subject to −1 log det P0 P0 > 0. . ˜i = E T bi . .17) Therefore.} We seek a small ellipsoid E0 containing A. . Let x denote the vector made by stacking x1 . xp . we obtain the largest volume ellipsoid contained in the intersection of the ellipsoids E1 .e. . Ep is deﬁned as A = {x1 + · · · + xp | x1 ∈ E1 . T ˜ A i = E i A i Ei . x1 . . . . . λ1 . . we have p ∆ T0 i=1 xi ≤ 0. . let Ei denote the matrix such that xi = Ei x. . . . T0 i=1 xi − i=1 τi Ti (xi ) ≤ 0. . . . p. . . . xp ∈ Ep . . . i = 1. . . . Ep . . p. . (3. . . .4 Outer approximation of the sum of ellipsoids The sum of the ellipsoids E1 . λ1 ≥ 0. λp ≥ 0. . x0 .. . .16) is equivalent to −Pi2 xi − x 0 P0 0 ≤ 0. . . p such that p p for every xi . . Indeed.17) Among the ellipsoids contained in the intersection of E1 . . Copyright c 1994 by the Society for Industrial and Applied Mathematics. . this condition is true if there exist τi ≥ 0. . p. . . . λp ≥ 0.17) Equivalently.11) holds. . . . i. . Ep by solving the CP (with variables P0 . we can ﬁnd one that maximizes the smallest semi-axis by solving minimize subject to λ λI > P0 > 0. . . . . . . .

The quantity (3. A related quantity is the one-sided multivariable stability margin. which can be computed via GEVPs. GV74. including many references and examples from mathematical ecology. Other references on the problem of improving the condition number of a matrix via diagonal or block-diagonal scaling are [Bau63. . In [KB93]. .g.20). . τp using Schur complements: ˜ ˜ b A0 ˜0 b 0 Ai ˜i 0 p ˜T T τi ˜T ci 0 ≤ 0. Wat91]. used in the analysis of linear systems with parameter uncertainties. power systems. . We should point out that the results of §3. just as we do). by minimizing log det A0 subject to A0 > 0 and (3. Diagonal quadratic Lyapunov functions The problem of rescaling a matrix M with a diagonal matrix D > 0 such that its Hermitian part is positive-deﬁnite is considered in the articles by Barker et al. . SL93a. . LN92]. i = 1. Notes and References Matrix scaling problems The problem of scaling a matrix to reduce its condition number is considered in Forsythe and Straus [FS55] (who describe conditions for the scaling T to be optimal in terms of a new variable R = T −∗ T −1 . KL85b. and Rotea and Prasanth [RP93. Sha85. . Saf82]). This condition is readily written as an LMI in variables A0 . The problem is the same as determining the existence of diagonal quadratic Lyapunov function for a given linear system. Kaszkurewicz and Bhaya give an excellent survey on the topic of diagonal Lyapunov functions. b0 . We also mention some other related quantities found in the articles by Kouvaritakis and Latchman [KL85a. . since the cost of computing the optimal scalings (via the GEVP) exceeds the cost of solving the problem for which the scalings are required (e. . Kra91]. where a diﬀerent approach to the same problem and its extensions is considered. This electronic version is for personal use and may not be duplicated or distributed. Tits and Doyle in [FTD91] (see also [BFBE92]). Some closely related quantities were considered in [SC93. A recent survey article on this topic is by Packard and Doyle [PD93]. RP94]. solving a set of linear equations). p such that ˜ A0 ˜T b 0 47 ˜0 b bT A−1 b0 − 1 0 0 p − τi i=1 ˜ Ai ˜T b i ˜i b ci (3. Hu87. A special case of the problem of minimizing the condition number of a positive-deﬁnite matrix is considered in [EHM92.18) is equivalent to there exist τi ≥ 0.Notes and References With this notation. we see that condition (3. Khachiyan and Kalantari ([KK92]) consider the problem of diagonally scaling a positive semideﬁnite matrix via interior-point methods and give a complete complexity analysis for the problem.20) bi b0 −1 b0 − i=1 0 b0 −A0 0 0 0 Therefore.19) ≤ 0.. Ep via the S-procedure. . τ1 .5). see [Ger85. . see also Chapter 8 in this book. . This is a CP. The problem of similarity-scaling a matrix to minimize its norm appears often in control applications (see for example [Doy82. we compute the minimum volume ellipsoid. BY89. Finally. CS92b]. (3. EHM93]. proven to contain the −1 sum of E1 . .1 are not practically useful in numerical analysis. described in [TSC92]. the problem of computing this quantity can be easily transformed into a GEVP. [BBP78] and Khalil [Kha82]. was introduced by Fan.

Matrix completion problems Matrix completions problems are addressed in the article of Dym and Gohberg [DG81].e. In [HW93]. i aT Qai i and λi = 0 if aT Qai = 1. their results are extended to a more general class of partially speciﬁed matrices. i i = 1. In [GJSW84]. Grigoriadis. Helton and Woerdeman consider the problem of minimum norm extension of Hankel matrices. aT Qai ≤ 1. BW92] for more references on contractive completion problems. in which an analytic solution is given for contractive completion problems with a special block form. . aT Qai i Copyright c 1994 by the Society for Industrial and Applied Mathematics. . . See [Woe90. . Its volume is √ proportional to det Q. The result in this paper is closely related to (and also more general than) the matrix elimination procedure given in §2. p Suppose now that Q is optimal. In [GFS93]. Consider an ellipsoid described by E = {x xT Q−1 x ≤ 1 } where Q = QT > 0. .6. Multiplying by Q1/2 on the left and right and taking the trace i p yields λ = n. . Maximum volume ellipsoid contained in a symmetric polytope In this section.48 Chapter 3 Some Matrix Problems and digital ﬁlter stability. p. . see §10. such problems cannot be cast as LMIPs. . Kahan. we have p xT Q−1 x = i=1 λi (xT ai )2 . . p . . in which the authors examine conditions for the existence of completions of band matrices such that the inverse is also a band matrix. . The condition E ⊆ P can be expressed aT Qai ≤ 1 for i = 1. . NW92. Barrett et al. . . the analytic center A mentioned in §3. A classic paper on the contractive completion problem (i.e. Dancis ([Dan92.. As far as we know.3. See also [OSA93]. . i i = 1. . λp such that p Q−1 = i=1 λi ai aT . In [BJL89].5 appears. completing a matrix so its norm is less than one) is Davis. we establish some properties of the maximum volume ellipsoid contained in a symmetric polytope described as: P= z aT z ≤ 1. i Hence we obtain the maximum volume ellipsoid contained in P by solving the CP minimize subject to log det Q−1 Q > 0. provide an algorithm for ﬁnding the completion maximizing the determinant of the completed matrix (i. In both papers the “sparsity pattern” result mentioned in §3. Frazho and Skelton consider the problem of approximating a given symmetric matrix by a Toeplitz matrix and propose a solution to this problem via convex optimization..2. Diagonal Lyapunov functions arise in the analysis of positive orthant stability. Dan93] and Johnson and Lunquist [JL92]) study completion problems in which the rank or inertia of the completed matrix is speciﬁed. there are nonnegative λ1 .5). . From the standard optimality conditions. and Weinberger in [DKW82]. i=1 i For any x ∈ Rn .

GLS88]). and can be traced back to Dikin [Dik67] and Karmarkar [Kar84] for polytopes described by a set of linear inequalities. Preparata and Shamos [PS85. i=1 Equivalently. √ the maximum volume ellipsoid contained in P approximates P within the scale factor n. Chernousko [Che80a. Such ellipsoids are called L¨wner–John ellipsoids. is contained in the given set (see [Joh85. Yurkovich and Passino [CYP93] and the book by Norton [Nor86]. p. Suppose now that x belongs to P. Che80c] and Kahan [Kah68]. Bona and Cerone [BBC90].. see Nesterov and Nemirovskii [NN94]. KLB90. Ellipsoidal approximation via analytic centers Ellipsoidal approximations via barrier functions arise in the theory of interior-point algorithms. the references on the ellipsoid algorithm describe various extensions. either λi = 0 or aT Qai = 1.g. Che80b. Fogel and Huang [FH82]. Minimum volume ellipsoids containing a given set arise often in control theory.255] or Post [Pos84].. ellipsoidal approximations of robot linkages and workspace are used to rapidly detect or avoid collisions. P ⊆ E ⊆P⊆ √ √ nE. Cheung. In the case of symmetric sets. Then p xT Q−1 x ≤ λi = n. Ellipsoidal approximation The problem of computing the minimum volume ellipsoid in Rn that contains a given convex set was considered by John and L¨wner. Deller [Del89]. In the ellipsoid algorithm we encountered the minimum volume ellipsoid that contains a given half-ellipsoid. SS72] use ellipsoids containing the sum and the intersection of two given ellipsoids in the problem of state estimation with norm-bounded disturbances. Eﬃcient interior-point methods for computing minimum volume outer and maximum volume inner ellipsoids for various sets are described in Nesterov and Nemirovskii [NN94]. Kosut and Boyd [LKB90. Schweppe and Schlaepfer [Sch68. Belforte. These approximations have the property that they are easy to compute. when shrunk o about its center by a factor of n. i. so we have: nE. Minimum volume ellipsoids have been widely used in system identiﬁcation. so we conclude i p 49 xT Q−1 x = i=1 λi (xT ai )2 . See also [KT93]. . Boyd and El Ghaoui [BE93]. For the case of more general LMIs. see e. Special techniques have been developed to compute smallest spheres and ellipsoids containing a given set of points in spaces of low dimension. In [RB92].e. and Jarre [Jar93c]. KLB92]. A closely related problem is ﬁnding the maximum volume ellipsoid contained in a convex set. Lau.Notes and References For each i. Similar results hold for these ellipsoids. who showed that the optimal ellipsoid. and come with provable bounds on how suboptimal they are. the factor n can o √ be improved to n (using an argument similar to the one in the Notes and References section above). see the articles by Fogel [Fog79]. See also the articles by Bertsekas and Rhodes [BR71]. These bounds have the following form: if This electronic version is for personal use and may not be duplicated or distributed.

3 We prove that the statement for every x.22) The maximum over x of −2 −2 (x − x1 )T P1 (x − x1 ) − λ(x − x0 )T P0 (x − x0 ) (3. depending on the number and type of constraints. this implies λ−1 (x1 − x0 )T x1 − x0 2 2 P0 /λ − P1 ≤ 0. (3. So an ellipsoidal approximation from a barrier function can serve as a “cheap substitute” for the L¨wner–John ellipsoid. it is easy to show that (3.24) or. o Proof of lemma in §3.22) implies (3. is equivalent to the matrix inequality (x1 − x0 )T P0 2 −P1 x1 − x0 λ−1 0 P0 0 −λI ≤ 0. x∗ is a maximizer of (3.25) In this case. the type and size of the constraints) then the resulting ellipsoid covers the feasible set.25).. Using (3.7. (3. one may be better than another.22).21) implies −2 2 2 −2 (x∗ − x1 )T P1 (P1 − P0 /λ)P1 (x∗ − x1 ) ≤ 1 − λ.23) 2 2 is ﬁnite if and only if P0 − λP1 ≤ 0 (which implies that λ > 0) and there exists x∗ satisfying −2 −2 P1 (x∗ − x1 ) = λP0 (x∗ − x0 ). condition (3.e.21) where P0 . Using Schur complements. this last condition is equivalent to λ−1 −2 2 2 (x∗ − x1 )T P1 (P0 /λ − P1 ) 2 2 P0 /λ − P1 2 (P0 /λ − −2 2 P1 )P1 (x∗ − x1 ) ≤ 0. (3.23). and is always smaller than the ones obtained via barrier functions.50 Chapter 3 Some Matrix Problems the inner ellipsoid is stretched about its center by some factor α (which depends only on the dimension of the space. the maximum volume inner ellipsoid. In contrast.24). . satisfying 2 2 −2 (P0 /λ − P1 )P1 (x∗ − x1 ) = x1 − x0 . i. equivalently. λ ≥ 0. the L¨wner–John ellipsoid. which is equivalent to (3. Using (3.21). P1 are positive matrices. Copyright c 1994 by the Society for Industrial and Applied Mathematics. −2 −2 (x − x1 )T P1 (x − x1 ) − λ(x − x0 )T P0 (x − x0 ) ≤ 1 − λ (3. however) but the factor α depends only on the dimension of the space and not on the type or number of constraints. Conversely. is more difo ﬁcult to compute (still polynomial. The papers cited above give diﬀerent values of α.

Our goal is to establish that various properties are satisﬁed by all solutions of a given DI.1) is called a solution or trajectory of the DI (4. Every trajectory of the LDI satisﬁes x = A(t)x.2).1). Very roughly speaking.2). ˙ x(0) = x0 . x ∈ Co F (x(t). t) is a convex set for every x and t. i.) As a speciﬁc and simple example. We can interpret the LDI (4. ˙ x(0) = x0 .4) x ∈ Ωx. t) ⊇ F (x(t). 51 (4.2)} . for every T ≥ 0 In fact we will not need the Relaxation Theorem. t). there can be many solutions of the DI (4. Any x : R+ → R that satisﬁes (4. By a standard result called the Relaxation Theorem. The DI given by is called the relaxed version of the DI (4. it can be shown that for every DI we encounter in this book.1).1) where F is a set-valued function on R × R+ . we might show that every trajectory of a given DI converges to zero as t → ∞.1.2) 4.1 Linear diﬀerential inclusions A linear diﬀerential inclusion (LDI) is given by where Ω is a subset of R . Of course. Since Co F (x(t).1). then the same Lyapunov function establishes the property for the relaxed DI (4.1). t).1) is also a trajectory of relaxed DI (4. t).1)} = {x(T ) | x satisﬁes (4. ˙ x(0) = x0 . we will always get it “for free”—every result we establish in the next two chapters extends immediately to the relaxed version of the problem. n n A diﬀerential inclusion (DI) is described by: (4.3) .e. n×n (4. ˙ x(0) = x0 (4. (See the References for precise statements. The reason is that when a quadratic Lyapunov function is used to establish some property for the DI (4. the Relaxation Theorem states that for many purposes the converse is true. every trajectory of the DI (4. the reachable or attainable sets of the DI and its relaxed version coincide. we may as well assume F (x. or rather.Chapter 4 Linear Diﬀerential Inclusions 4.1 Diﬀerential Inclusions x ∈ F (x(t). For example. {x(T ) | x satisﬁes (4..3) as describing a family of linear time-varying systems.

For example. The matrices in (4. In order not to introduce another term ˙ to describe the set of all solutions of (4. we will discuss these in detail when we encounter them. (4. where Ω ⊆ R(n+nz )×(n+nu +nw ) . some are new. We will be more speciﬁc about the form of Ω shortly.1 Linear time-invariant systems When Ω is a singleton. which means that the corresponding variable is not used. ˙ z = Cz (t)x + Dzu (t)u + Dzw (t)w x(0) = x0 .5) and (4. w : R+ → Rnw . and nz equal to zero. (4. 4. 4.6) a system described by LDIs or simply. the LDI x ∈ Ωx results when nu = nw = nz = 0. Conversely. u : R+ → Rnu . In the language of control theory. u is the control input.7) Although most of the results of Chapters 5–7 are well-known for LTI systems. where Ω= A Cz Bu Dzu Bw Dzw .3). the solution of (4. Copyright c 1994 by the Society for Industrial and Applied Mathematics.6).5) where x : R+ → Rn . nw .2 A generalization to systems We will encounter a generalization of the LDI described above to linear systems with inputs and outputs.5) and (4.6) for all t ≥ 0.4) is a trajectory of the LDI (4. the LDI (4. an LDI.3) might be described as an “uncertain time-varying linear system. (4.2.1. z : R+ → Rnz . for any A : R+ → Ω. w is the exogenous input and z is the output.” with the set Ω describing the “uncertainty” in the matrix A(t). we will call (4.5) satisfy A(t) Cz (t) Bu (t) Dzu (t) Bw (t) Dzw (t) ∈ Ω. .52 Chapter 4 Linear Differential Inclusions for some A : R+ → Ω. x is referred to as the state.2 Some Speciﬁc LDIs We now describe some speciﬁc families of LDIs that we will encounter in the next two chapters. the LDI reduces to the linear time-invariant (LTI) system x = Ax + Bu u + Bw w. 4. ˙ x(0) = x0 . We will consider a system described by x = A(t)x + Bu (t)u + Bw (t)w. z = Cz x + Dzu u + Dzw w. In some applications we can have one or more of the integers nu .

L Bw. If instead Ω is described by a set of l linear inequalities. z = Cz x + Dzp p + Dzu u + Dzw w. The set Ω is well deﬁned if and only if Dqp Dqp < I.9): Dqw and Dzp will always be zero. we often drop some terms in (4.1 Dzu.1 Bu.. Dzw . For the NLDI (4. In the sequel we will consider only well-posed NLDIs.(4. in which case we say that the NLDI is well-posed . Most of our results require that Ω be described by a list of its vertices.. Thus. i. ˜ C= Cq Dqu Dqw . i. output z.L Dzu. Ω is the image of the (matrix norm) unit ball under a (matrix) linear-fractional T mapping..2. ∆(t) ≤ 1. These assumptions greatly simplify the formulas and treatment.1 .2 Some Specific LDIs 53 4. . we will call the LDI a polytopic LDI or PLDI.1 Bw..8) are given..2.2 Polytopic LDIs When Ω is a polytope.L . Remark: In the sequel.9) where ∆ : R+ → Rnp ×nq . x(0) = x0 . then the number of vertices.. with ∆(t) ≤ 1 for all t.e.e. Dzu and Dqu will be assumed zero. 4.4. ∆ ≤ 1 in the equivalent form pT p ≤ q T q. Note that Ω is convex.10) −1 ˜ ˜ ˜ A + B∆ (I − Dqp ∆) C ∆ ≤1 .L Bu. in the form Co A1 Cz. described by x = Ax + Bp p + Bu u + Bw w. It is not hard to work out the corresponding formulas and results for the more general case. Therefore results for PLDIs that require the description (4.1 Dzw.8) where the matrices (4.9) can be interpreted as an LTI system with inputs u. AL Cz. ˙ q = Cq x + Dqp p + Dqu u + Dqw w. (4. The equations (4. (4. L. and very often Dqp . w.L Dzw.3 Norm-bound LDIs Another special class of LDIs is the class of norm-bound LDIs (NLDIs). will generally increase very rapidly (exponentially) with l. and a time-varying feedback matrix ∆(t) connected between q and p. p = ∆(t)q.9) the set Ω has the form Ω= where ˜ A= A Cz Bu Dzu Bw Dzw . We will often rewrite the condition p = ∆(t)q. ˜ B= Bp Dzp . This electronic version is for personal use and may not be duplicated or distributed.8) are of limited interest for problems in which Ω is described in terms of linear inequalities.

The DNLDI is described by the set Ω= −1 ˜ ˜ ˜ A + B∆ (I − Dqp ∆) C ∆ ≤ 1.4 Diagonal norm-bound LDIs A useful variation on the NLDI is the diagonal norm-bound LDI (DNLDI).13) Copyright c 1994 by the Society for Industrial and Applied Mathematics. . in which we further restrict the matrix ∆ to be diagonal. It is given by x = Ax + Bp p + Bu u + Bw w. i.e. z = Cz x + Dzp p + Dzu u + Dzw w. −1 ˜ ˜ ˜ Co Ω = Co A + B∆ (I − Dqp ∆) C ∆ diagonal. In the language of control theory. Remark: More elaborate variations on the NLDI are sometimes useful. we have an NLDI. Although this condition is explicit. ∆ diagonal . we have a DNLDI. time-varying feedback gains. (4. t). u. u. The condition of well-posedness for a DNLDI is that I −Dqp ∆ should be invertible for all diagonal ∆ with ∆ ≤ 1. When we have a single block. when it is diagonal (with no equality constraints). (4. determining whether a DNLDI is well-posed is NP-hard. and in addition. there can be equality constraints among some of the blocks. B and C are given by (4. w. . see the Notes and References. (4. ˙ z = g(x. |∆ii | = 1 . . In fact. Here we have the componentwise inequalities |pi | ≤ |qi | instead of the single inequality p ≤ q appearing in the NLDI. we describe a DNLDI as a linear system with nq scalar. 4. i. the vertices of Co Ω are among 2nq images of the extreme points of ∆ under the matrix linear-fractional mapping. Therefore a DNLDI can be described as a PLDI.” in which the matrix ∆ is restricted to have a given block-diagonal form.2. w. We will often express the condition pi = δi (t)qi . pi = δi (t)qi . t). at least in principle. |∆ii | = 1. the number of inequalities grows exponentially with nq . time-varying systems using a technique known as “global linearization”. . It turns out that for a DNLDI.e. In fact. Consider the system x = f (x. nq . The idea is implicit in the work on absolute stability originating in the Soviet Union in the 1940s. But there is a very important diﬀerence between the two descriptions of the same LDI: the number of vertices required to give the PLDI representation of an LDI increases exponentially with nq (see the Notes and References of Chapter 5 for more discussion). Doyle introduced the idea of “structured feedback. |δi (t)| ≤ 1 in the simpler form |pi | ≤ |qi |. Co Ω is a polytope. . Most of the results we present for NLDIs and DNLDIs can be extended to these more general types of LDIs. This is equivalent to det(I − Dqp ∆) > 0 when ∆ takes on its 2nq extreme values. each of which is bounded by one. |δi (t)| ≤ 1. uncertain.10).12) ˜ ˜ ˜ where A. i = 1. ˙ x(0) = x0 ...54 Chapter 4 Linear Differential Inclusions 4.3 Nonlinear System Analysis via LDIs Much of our motivation for studying LDIs comes from the fact that we can use them to establish various properties of nonlinear.11) q = Cq x + Dqp p + Dqu u + Dqw w.

w.. 0. ˜ z−z ˜ w−w ˜ i. 0. 0. w. 0) = 0. w − w. 0. and t there is a matrix G(x. t) ∈ Ω such that x f (x. w. Since by assumption ˜ ∂φ (ζ) ∈ Ω.13) has this property. ∂x ∂u ∂w In fact we can make a stronger statement that links the diﬀerence between a pair of trajectories of the nonlinear system and the trajectories of the LDI given by Ω. t) (4.. u. ˜ This electronic version is for personal use and may not be duplicated or distributed. and ∂f ∂f ∂f ∂x ∂u ∂w (4. 0. By the mean-value theorem we have cT (φ(x) − φ(˜)) = cT x φ(x) − φ(˜) ∈ Co Ω(x − x). converges to zero). u. 0.15) but not necessarily f (0. w. 0) = 0. g(0. z) and (˜.3 Nonlinear System Analysis via LDIs 55 where Ω ⊆ R(n+nz )×(n+nu +nw ) . 0) = 0. w. u. Suppose we have (4. we see that φ(x) − φ(˜) is in every half-space that contains the x convex set Co Ω(x − x).g. w. ∂x we conclude that cT (φ(x) − φ(˜)) ≤ x sup cT A(x − x).1 A derivative condition Conditions that guarantee the existence of such a G are f (0. t.16). Suppose that for each x. t) w 4. w.14) = G(x. which proves (4. x ˜ n (4. Then of course every trajectory of the nonlinear system (4. then for any x and x we have ˜ To see this. z ) we have x ˜ ˜ x−x ˜ ˙ x−x ˙ ˜ ∈ Co Ω u − u . ˜ ˜ ˜ These results follow from a simple extension of the mean-value theorem: if φ : Rn → Rn satisﬁes ∂φ ∈Ω ∂x throughout Rn . z − z ) is a trajectory of the LDI given by Co Ω. u. u. let c ∈ R . . then a fortiori we have proved that every trajectory of the nonlinear system (4. g(0. 0.16) ∂φ (ζ)(x − x) ˜ ∂x for some ζ that lies on the line segment between x and x.15) ∂g ∂g ∂g ∈ Ω for all x. u.4. (x − x. w. 0.e. ˜ A ∈ Co Ω Since c was arbitrary. t) u g(x. Then for any pair of trajectories (x.13) is also a trajectory of the LDI deﬁned by Ω. 0) = 0.3. If we can prove that every trajectory of the LDI deﬁned by Ω has some property (e.

[Liu68. See also the article by Roxin [Rox65] and references therein. PLDIs come up in several articles on control theory.14) holds. there are many trajectories of the LDI that are not trajectories of the nonlinear system.2 Sector conditions Sometimes the nonlinear system (4. We use the term only to point out the interpretation as an uncertain time-varying linear system. is selector-linear diﬀerential inclusion. we say the system is well-posed. [MP89. KP87b. Background material. suggested to us and used by A.13) be the equations obtained by eliminating the variables p and q. βi ] . Of course. Many authors refer to LDIs using some subset of the phrase “uncertain linear time-varying system”. where αi and βi are given. the condition (4. Global linearization The idea of replacing a nonlinear system by a time-varying linear system can be found in Liu et al. Pyatnitskii. since LDIs do not enjoy any particular linearity or superposition properties.g. and let (4.g. Integral quadratic constraints The pointwise quadratic constraint p(t)T p(t) ≤ q(t)T q(t) that occur in NLDIs can be genert t alized to the integral quadratic constraint 0 pT p dτ ≤ 0 q T q dτ . . approximating the set of trajectories of a nonlinear system via LDIs can be very conservative. i = 1.17) αi q 2 ≤ qφi (q. In this case.56 Chapter 4 Linear Differential Inclusions 4. t) ≤ βi q 2 ∆ diagonal. KP87a. can be found in the book by Aubin and Frankowska [AF90]. BY89]. nq .17) provided the matrix I−Dqp ∆ is nonsingular for all diagonal ∆ with αi ≤ ∆ii ≤ βi .17) is well-posed. B and C are given by (4. e.. A more accurate term.. Assume now that the system (4. We will see in Chapter 8 how this conservatism can be reduced in some special cases. . i.10). e. see §8. ˙ q = Cq x + Dqp p + Dqu u + Dqw w. the system consists of a linear part together with nq time-varying sector-bounded scalar nonlinearities. Copyright c 1994 by the Society for Industrial and Applied Mathematics. . and the classic text by Filippov [Fil88] cover the theory of diﬀerential inclusions. They call this approach global linearization. The idea of global linearization is implicit in the early Soviet literature on the absolute stability problem..2 and §8. where φi : R × R+ → R satisfy the sector conditions for all q and t ≥ 0.3.3. Then.e.13) can be expressed in the form x = Ax + Bp p + Bu u + Bw w. ∆ii ∈ [αi .g.. Filippov and E. Lur57] and Popov [Pop73]. Notes and References Diﬀerential inclusions The books by Aubin and Cellina [AC84]. pi = φi (qi . . t). set-valued analysis. ˜ ˜ ˜ where A. Kisielewicz [Kis91]. e. z = Cz x + Dzp p + Dzu u + Dzw w. The term linear diﬀerential inclusion is a bit misleading. The variables pi and qi can be eliminated from (4. LSL69]. Lur’e and Postnikov [LP44. In words. . Deﬁne Ω= −1 ˜ ˜ ˜ A + B∆ (I − Dqp ∆) C (4.

An equivalent condition is in terms of P0 matrices (a matrix is P0 if every principal minor is nonnegative. SP89. so the diﬀerent models refer to the same state vector. but at high frequencies they diﬀer considerably more. the authors also give a suﬃcient condition for well-posedness of more general “structured” LDIs.g.Notes and References Modeling systems as PLDIs 57 One of the key issues in robust control theory is how to model or measure plant “uncertainty” or “variation”. Standard branch-and-bound techniques can be applied to the problem of determining wellposedness of DNLDIs.. For each data set we develop a linear system model of the plant. we have ˜ ˜ ˜ Co Ω = Co A + B∆ (I − Dqp ∆)−1 C ∆ diagonal. which means that the DNLDI (4. FP67]): A DNLDI is well-posed if and only if I + Dqp is invertible. Another equivalent condition is that (I + Dqp . see [FP62. see for example [Sd86. (4. .12) can also be represented as a PLDI. and Doyle [FTD91] show that if the LMI T Dqp P Dqp < P. under diﬀerent operating conditions. then the DNLDI is well-posed. |∆ii | = 1 . we model the plant as a time-varying linear system. Suppose we have a real system that is fairly well modeled as a linear system. Well-posedness of DNLDIs See [BY89] for a proof that a DNLDI is well-posed if and only if det(I − Dqp ∆) > 0 for every ∆ with |∆ii | = 1. We propose to model the system as a PLDI with the vertices given by the measured or estimated linear system models.18) is equivalent to the existence of a diagonal (scaling) matrix R such that RDqp R−1 < 1. These models should be fairly close. of high complexity [CPS92. with system matrices that can jump around among any of the models we estimated. I − Dqp ) is a W0 pair (see [Wil70]). GS88.. At the least. We collect many sets of input/output measurements. Tits. In [FTD91]. see also [Sae86]. Condition (4.e. the method is extremely simple and natural. diﬀerent units from a manufacturing run). P > 0. More speciﬁcally. this collection of models contains information about the “structure” of the plant variation. But in practice they may be more eﬃcient. when equality constraints are imposed on the matrix ∆ in (4. It is important that we have data sets from enough plants or plant conditions to characterize or at least give some idea of the plant variation that can be expected. Fan. see [Fer93]. or perhaps from diﬀerent instances of the system (e.18) is feasible. BB92]. the idea behind this proof can be traced back to Zadeh and Desoer [ZD63. BBB91. More importantly. Representing DNLDIs as PLDIs For a well-posed DNLDI. To simplify the problem we will assume that the state in this model is accessible. There are several suﬃcient conditions for well-posedness of a DNLDI that can be checked in polynomial-time. The problem of determining whether a general matrix is P0 is thought to be very diﬃcult. These conditions can also be obtained using the S-procedure. but of course not exactly the same. In other words. We might ﬁnd for example that the transfer matrices of these models at s = 0 diﬀer by about 10%. §9. It seems reasonable to conjecture that a controller that works well with this PLDI is likely to work well when connected to the real system. especially when used with computationally cheap suﬃcient conditions (see below). FP66. i. Of course these methods have no theoretical advantage over simply checking that the 2nq determinants are positive. Such controllers can be designed by the methods described in Chapter 7.17]. Co Ω is a polytope. p149].11). and (I + Dqp )−1 (I − Dqp ) is a P0 matrix [Gha90].. P diagonal. This electronic version is for personal use and may not be duplicated or distributed. We propose a simple method that should work well in some cases.g. e. obtained at diﬀerent times.

Wext ∈ Ω. which concludes our proof.e. Wext ∈ Co K. where > 0 is suﬃciently small. (We can also replace Bp and Cq with Bp U and V Cq where U and V are any orthogonal matrices. Deﬁne K= ˜ ˜ ˜ A + B∆ (I − Dqp ∆)−1 C ∆ diagonal. . the denominator c + d∆11 is nonzero for ∆11 ∈ [−1. . . we will only consider the case with Bp ∈ Rn×n and invertible. |∆ii | = 1 . Since ΩPLDI ⊆ ΩNLDI implies that every trajectory of the PLDI is a trajectory of the NLDI. ∆ (4. Approximating PLDIs by NLDIs It is often useful to conservatively approximate a PLDI as an NLDI. completing the proof. . We will now show that every extreme point of Co Ω is contained in Co K. among many. . . ψ(W ) is a bounded linear-fractional function of the ﬁrst diagonal entry ∆11 of ∆.e. For simplicity. This will give an eﬃcient outer approximation of the PLDI (4. ∆nq nq leads ˜ ˜ ˜ us to the following conclusion: Wext = A + BS (I − Dqp S)−1 C. and the NLDI x = (A + Bp ∆(t)Cq )x. which is a set containing at most 2nq points. . see the Notes and References of Chapter 5. . Let Wext be an extreme point of Co Ω. Now. every result we establish for all trajectories of the NLDI holds for all trajectories of the PLDI. . and Cq such that ΩPLDI ⊆ ΩNLDI with the set ΩNLDI as small as possible in some appropriate sense. ∆11 = ±1. We have already noted that the description of a DNLDI as a PLDI will be much larger than its description as a DNLDI. . 1]. For L very large. ˙ ∆ ∆(t) ≤ 1. We now prove this result. c and d. Kam83]. ∆nq nq . contains Ω but not Wext . One reason. . . By deﬁnition ˜ ˜ ˜ Co Ω = Co A + B∆ (I − Dqp ∆)−1 C ∆ diagonal. there exists π such that Bp π = (Ak − A)ξ. where S satisﬁes Sii = ±1. Moreover. . without aﬀecting the set ΩNLDI .20). We consider the PLDI x = A(t)x. Evidently there is substantial redundancy in our representation of the NLDI (4. since otherwise. Then there exists a linear function ψ such that Wext is the unique maximizer of ψ over Co Ω. . Clearly we have Co Ω ⊇ Co K. ˙ A(t) ∈ ΩPLDI = Co {A1 . for every ﬁxed ∆22 . KP87a. . Our goal is to ﬁnd A. . . which will imply that Co K ⊇ Co Ω. which contradicts Wext ∈ Co Ω. Copyright c 1994 by the Society for Industrial and Applied Mathematics. Note that the set on the right-hand side is a polytope with at most 2nq vertices. the half-space {W | ψ(W ) < ψ(Wext ) − }. .19) by the NLDI (4.. (4. for W ∈ Ω. since the DNLDI is well-posed. Therefore. . Bp . |∆ii | ≤ 1 . b. B and C are given by (4. it is much easier to work with the NLDI than the PLDI. but we will not use this fact. In other words. AL } . In fact. the function ψ achieves a maximum at an extreme value of ∆11 . i.20): We can replace B p by αBp and Cq by Cq /α where α = 0.) We have ΩPLDI ⊆ ΩNLDI if for every ξ and k = 1. ∆nq nq . Extending this argument to ∆22 . i. for ﬁxed ∆22 . it equals (a + b∆11 )/(c + d∆11 ) for some a.. T π T π ≤ ξ T Cq Cq ξ.10). .20) with associated set ΩNLDI = {A + Bp ∆Cq | ∆ ≤ 1}. . L. is the potentially much smaller size of the description as an NLDI compared to the description of the PLDI.58 Chapter 4 Linear Differential Inclusions ˜ ˜ ˜ where A.19) with x(t) ∈ Rn . and also [BY89] and [PS82.

**Notes and References
**

This yields the equivalent condition

−T −1 T (Ak − A)T Bp Bp (Ak − A) ≤ Cq Cq ,

59

k = 1, . . . , L,

**which, in turn, is equivalent to
**

T Bp Bp > 0, T Cq Cq

Ak − A

(Ak − A)T

T B p Bp

≥ 0,

k = 1, . . . , L.

T T Introducing new variables V = Cq Cq and W = Bp Bp , we get the equivalent LMI in A, V and W

W > 0,

V Ak − A

(Ak − A)T W

≥ 0,

k = 1, . . . , L.

(4.21)

There are several ways to minimize the size of ΩNLDI ⊇ ΩPLDI . The most obvious is to minimize Tr V + Tr W subject to (4.21), which is an EVP in A, V and W . This objective is clearly related to several measures of the size of ΩNLDI , but we do not know of any simple interpretation. We can formulate the problem of minimizing the diameter of ΩNLDI as an EVP. We deﬁne the diameter of a set Ω ⊆ Rn×n as max { F − G | F, G ∈ Ω}. The diameter of ΩNLDI is equal to 2 λmax (V )λmax (W ). We can exploit the scaling redundancy in Cq to assume without loss of generality that λmax (V ) ≤ 1, and then minimize λmax (W ) subject to (4.21) and V ≤ I, which is an EVP. It can be shown that the resulting optimal V opt satisﬁes λmax (V opt ) = 1, so that we have in fact minimized the diameter of ΩNLDI subject to ΩPLDI ⊆ ΩNLDI .

Thus we have ΩPLDI ⊆ ΩNLDI if condition (4.21) holds.

This electronic version is for personal use and may not be duplicated or distributed.

Chapter 5

**Analysis of LDIs: State Properties
**

In this chapter we consider properties of the state x of the LDI x = A(t)x, ˙ where Ω ⊆ R

n×n

A(t) ∈ Ω, has one of four forms:

(5.1)

• LTI systems: LTI systems are described by x = Ax. ˙ • Polytopic LDIs: PLDIs are described by x = A(t)x, A(t) ∈ Co {A1 , . . . , AL }. ˙ • Norm-bound LDIs: NLDIs are described by ˙ x = Ax + Bp p, which we will rewrite as x = Ax + Bp p, ˙ pT p ≤ (Cq x + Dqp p)T (Cq x + Dqp p). (5.2) q = Cq x + Dqp p, p = ∆(t)q, ∆(t) ≤ 1,

We assume well-posedness, i.e., Dqp < 1. • Diagonal Norm-bound LDIs: DNLDIs are described by x = Ax + Bp p, ˙ pi = δi (t)qi , which can be rewritten as x = Ax + Bp p, ˙ q = Cq x + Dqp p, |pi | ≤ |qi |, i = 1, . . . , nq . q = Cq x + Dqp p, |δi (t)| ≤ 1, i = 1, . . . , nq . (5.3)

Again, we assume well-posedness.

5.1

Quadratic Stability

We ﬁrst study stability of the LDI (5.1), that is, we ask whether all trajectories of system (5.1) converge to zero as t → ∞. A suﬃcient condition for this is the existence of a quadratic function V (ξ) = ξ T P ξ, P > 0 that decreases along every nonzero trajectory of (5.1). If there exists such a P , we say the LDI (5.1) is quadratically stable and we call V a quadratic Lyapunov function. 61

62 Since

Chapter 5

Analysis of LDIs: State Properties

d V (x) = xT A(t)T P + P A(t) x, dt a necessary and suﬃcient condition for quadratic stability of system (5.1) is P > 0, AT P + P A < 0 for all A ∈ Ω. (5.4)

Multiplying the second inequality in (5.4) on the left and right by P −1 , and deﬁning a new variable Q = P −1 , we may rewrite (5.4) as Q > 0, QAT + AQ < 0 for all A ∈ Ω. (5.5)

This dual inequality is an equivalent condition for quadratic stability. We now show that conditions for quadratic stability for LTI systems, PLDIs, and NLDIs can be expressed in terms of LMIs. • LTI systems: Condition (5.4) becomes P > 0, AT P + P A < 0. (5.6) Therefore, checking quadratic stability for an LTI system is an LMIP in the variable P . This is precisely the (necessary and suﬃcient) Lyapunov stability criterion for LTI systems. (In other words, a linear system is stable if and only if it is quadratically stable.) Alternatively, using (5.5), stability of LTI systems is equivalent to the existence of Q satisfying the LMI Q > 0, AQ + QAT < 0. (5.7)

Of course, each of these (very special) LMIPs can be solved analytically by solving a Lyapunov equation (see §1.2 and §2.7). • Polytopic LDIs: Condition (5.4) is equivalent to P > 0, AT P + P Ai < 0, i i = 1, . . . , L. (5.8) Thus, determining quadratic stability for PLDIs is an LMIP in the variable P . The dual condition (5.5) is equivalent to the LMI in the variable Q Q > 0, QAT + Ai Q < 0, i i = 1, . . . , L. (5.9)

**• Norm-bound LDIs: Condition (5.4) is equivalent to P > 0 and ξ π
**

T

AT P + P A T Bp P

P Bp 0

ξ π

<0

(5.10)

for all nonzero ξ satisfying T T T −Cq Cq −Cq Dqp ξ T T π −Dqp Cq I − Dqp Dqp A = {(ξ, π) | ξ = 0, (5.11) }

ξ π

≤ 0.

(5.11)

**In order to apply the S-procedure we show that the set
**

Copyright c 1994 by the Society for Industrial and Applied Mathematics.

. the condition that dV (x)/dt < 0 for all nonzero trajectories is equivalent to (5. π) | (ξ. ˙ q = Cq x + Dqp p < 0. T i = 1. The condition dV (x)/dt < 0 for all nonzero trajectories is equivalent to ξ T (AT P + P A)ξ + 2ξ T P Bp π < 0 for all nonzero (ξ.) Using the S-procedure. (5. π) | ξ = 0. In the remainder of this chapter we will assume that Dqp in (5.13) P > 0. (This argument recurs throughout this chapter and will not be repeated.14) < 0. (5. obtain an equivalent condition T T ˜ ˜ ˜ AT P + P A + Cq Cq P Bp + Cq Dqp ˜ < 0. is less than one.12) Thus.10) being satisﬁed for any nonzero (ξ. This electronic version is for personal use and may not be duplicated or distributed. Remark: Note that our assumption of well-posedness is in fact incorporated in T the LMI (5. .11) } .1 Quadratic Stability 63 equals the set B = {(ξ. λ ≥ 0.12) has a solution. π) satisfying (5.3).2) is zero. we may. we can obtain only a suﬃcient condition for quadratic stability. With the new variables Q = P −1 . T T ˜ (P Bp + Cq Dqp )T −(I − Dqp Dqp ) ˜ an LMI in the variable P . π = 0. . It suﬃces to show that {(ξ. Therefore the NLDI (5.i ξ + Dqp. • Diagonal Norm-bound LDIs: For the DNLDI (5.11) cannot hold without having ξ = 0.i ξ + Dqp. . quadratic stability of the NLDI has the frequencydomain interpretation that the H∞ norm of the LTI system x = Ax + Bp p. . then condition (5.i π) (Cq. we ﬁnd that quadratic stability of (5. But this is immediate: T If π = 0. Since LMI (5. Thus. the reader should bear in mind that all the following results hold for nonzero D qp as well.2) is equivalent to the existence of P and λ satisfying P > 0.i π) . Q > 0. determining quadratic stability of an NLDI is an LMIP. π) satisfying 2 πi ≤ (Cq.11) } = ∅. Therefore.2) is well-posed and quadratically stable if and only if the LMIP (5.12) since it implies I − Dqp Dqp > 0. (5. π) = 0. T AT P + P A + λCq Cq T (P Bp + λCq Dqp )T T P Bp + λCq Dqp T −λ(I − Dqp Dqp ) (5.11). T T µBp Dqp + QCq T −µ(I − Dqp Dqp ) T AQ + QAT + µBp Bp T T (µBp Dqp + QCq )T (5. since I −D qp Dqp > 0. by deﬁning P = P/λ. µ = 1/λ.12) implies ˜ that λ > 0.5. L. quadratic stability of NLDIs is also equivalent to the existence of µ and Q satisfying the LMI µ ≥ 0.

. this is equivalent to T AT P + P A + Cq ΛCq T Bp P T P Bp + Cq ΛDqp T Dqp ΛDqp − Λ + T Dqp ΛCq < 0. . (See the Notes and References for more details on the connection between the S-procedure and scaling. Remark: Quadratic stability of a DNLDI via the S-procedure has a simple frequency-domain interpretation. if there exist P > 0 and diagonal Λ ≥ 0 satisfying (5.15) or (5.i π) − πi < 0 T for all nonzero (ξ.i ξ + Dqp. For more general LDIs.) Remark: Here too our assumption of well-posedness is in fact incorporated in the LMI (5.i ξ + Dqp. Note the similarity between the corresponding LMIs for the NLDI and the DNLDI. positive-deﬁnite matrix Λ. the only diﬀerence is that the diagonal matrix appearing in the LMI associated with the DNLDI is ﬁxed as the scaled identity matrix in the LMI associated with the NLDI.2) is well-posed and quadratically stable if the LMI (5. . we leave to the reader the task of generalizing all the results for NLDIs to DNLDIs. stability does not imply quadratic stability (see the Notes and References at the end of this chapter).15) or (5. however. Λ−1/2 can then be interpreted as a scaling. Bp . . Equivalently. we will often restrict our attention to NLDIs. for some diagonal.16) another LMIP.15). then the DNLDI (5. Copyright c 1994 by the Society for Industrial and Applied Mathematics. Using the S-procedure. and assuming (A. Remark: An LTI system is stable if and only if it is quadratically stable.i π) (Cq. . and pay only occasional attention to DNLDIs. Note that from (5.i to denote the ith rows of Cq and Dqp respectively.i and Dqp. The bottom right block is precisely the suﬃcient condition for well-posedness mentioned on page 57. quadratic stability is implied by the existence of Q > 0. It is true that an LDI is stable if and only if there exists a convex Lyapunov function that proves it. λnq such that ξ T (AT P + P A)ξ + 2ξ T P Bp π L + i=1 2 λi (Cq. µnq ) > 0 satisfying the LMI T AQ + QAT + Bp M Bp T Cq Q + Dqp M Bp T T QCq + Bp M Dqp T −M + Dqp M Dqp < 0. . . . see the Notes and References.64 Chapter 5 Analysis of LDIs: State Properties where we have used Cq. In the remainder of the book. . .15) Therefore. Therefore the DNLDI (5. π).3) is quadratically stable. (5. Denoting by H the transfer matrix H(s) = Dqp + Cq (sI − A)−1 Bp . Λ1/2 HΛ−1/2 ∞ < 1. Checking this suﬃcient condition for quadratic stability is an LMIP. we see that this condition is implied by the existence of nonnegative λ1 . (5. With Λ = diag(λ. . In general it is straightforward to derive the corresponding (suﬃcient) condition for DNLDIs from the result for NLDIs.16).16) is feasible. λnq ). . Cq ) is minimal. . M = diag(µ1 . quadratic stability is equivalent to the fact that. we must have Λ > 0.15).

.) • Polytopic LDIs: The EVP in the variables P and η is minimize subject to η (5. We ask the question: Does there exist T such that in the new coordinate system. . we have taken advantage of the homogeneity of the LMI (5. T = P −1/2 (these T ’s are all related by right orthogonal matrix multiplication).6) in the variable P . η and λ is minimize subject to η (5.. This turns out to be an EVP for LTI systems. Consider the change of coordinates x = T x. ¯ ¯ In the new coordinate system. so that κ(T )2 = κ(P ). where det T = 0.8).1 Coordinate transformations We can give another interpretation of quadratic stability in terms of state-space coordinate transformations. let P = T −T T −1 .1) is described by the matrix A(t) = T −1 A(t)T .12). the change of coordinates with smallest condition number is obtained by solving the following EVPs: • LTI systems: The EVP in the variables P and η is minimize subject to η (5. . To see this.1 Quadratic Stability 65 5.1. With this interpretation. −1 (Any T with T T T = Popt where Popt is an optimal solution. . it is natural to seek a coordinate transformation matrix T that makes all nonzero trajectories decreasing in norm.6). I ≤ P ≤ ηI (In this formulation. PLDIs and NLDIs). .1.2 Quadratic stability margins Quadratic stability margins give a measure of how much the set Ω can be expanded about some center with the LDI remaining quadratically stable. I ≤ P ≤ ηI • Norm-bound LDIs: The EVP in the variables P .) For each of our systems (LTI systems. I ≤ P ≤ ηI 5. all (nonzero) trajectories of the LDI are always decreasing in norm. polytopic and norm-bound LDIs. in which case we can take any T with T T T = P −1 . This quantity is computed by solving the following GEVP in P This electronic version is for personal use and may not be duplicated or distributed. is optimal for the original problem. i. • Polytopic LDIs: For the PLDI x = (A0 + A(t))x.5.e. d x /dt < 0? ¯ It is easy to show that this is true if and only if there exists a quadratic Lyapunov function V (ξ) = ξ T P ξ for the LDI. and has the smallest possible condition number κ(T ). ˙ A(t) ∈ α Co {A1 . Minimizing the condition ¯ number of T subject to the requirement that d x /dt < 0 for all nonzero trajectories is then equivalent to minimizing the condition number of P subject to d(xT P x)/dt < 0. we deﬁne the quadratic stability margin as the largest nonnegative α for which it is quadratically stable. (5. AL } . for example.

the largest lower bound on the decay rate that we can ﬁnd using a quadratic Lyapunov function can be found by solving the following GEVP in P and α: maximize subject to α P > 0. and is computed by solving the GEVP in P .1). . (5. T AT P + P A + βλCq Cq T Bp P P Bp −λI <0 ˜ ˜ Deﬁning P = P/λ. ˙ T pT p ≤ α2 xT Cq Cq x.) We can use the quadratic Lyapunov function V (ξ) = ξ T P ξ to establish a lower bound on the decay rate of the LDI (5. lim inf t→∞ • LTI systems: equivalent to The condition that dV (x)/dt ≤ −2αV (x) for all trajectories is (5. β ≥ 0. Chapter 5 Analysis of LDIs: State Properties α ≥ 0. is deﬁned as the largest α ≥ 0 for which the system is quadratically stable. the decay rate is the supremum of − log x(t) t over all nonzero trajectories. . 0 i • Norm-bound LDIs: The quadratic stability margin of the system x = Ax + Bp p. Therefore.1) is at least α.18) Copyright c 1994 by the Society for Industrial and Applied Mathematics. 5. so that x(t) ≤ e−αt κ(P )1/2 x(0) for all trajectories.1. . . (Stability corresponds to positive decay rate. L AT P + P A0 + α AT P + P Ai < 0.66 and α: maximize subject to α P > 0.3 Decay rate The decay rate (or largest Lyapunov exponent) of the LDI (5. β ≥ 0. If dV (x)/dt ≤ −2αV (x) for all trajectories.1) is deﬁned to be the largest α such that t→∞ lim eαt x(t) = 0 holds for all trajectories x. Equivalently. then V (x(t)) ≤ V (x(0))e−2αt .17) (5. i = 1. . λ and β = α2 : maximize subject to β P > 0. we get an equivalent EVP in P and β: maximize subject to β ˜ P > 0.17) AT P + P A + 2αP ≤ 0. T ˜ ˜ AT P + P A + βCq Cq T ˜ Bp P ˜ P Bp −I <0 Remark: The quadratic stability margin obtained by solving this EVP is just 1/ Cq (sI − A)−1 Bp ∞ . and therefore the decay rate of the LDI (5.

(5. As an alternate condition. we obtain the largest lower bound on the decay rate of (5.21) • Norm-bound LDIs: Applying the S-procedure. there exists a quadratic Lyapunov function proving that the decay rate is at least α if and only if there exists Q satisfying the LMI Q > 0. L. . i. T QCq −µI T AQ + QAT + µBp Bp + 2αQ Cq Q (5.2) by maximizing α over the variables α. Ai Q + QAT + 2αQ ≤ 0.19) • Polytopic LDIs: The condition that dV (x)/dt ≤ −2αV (x) for all trajectories is equivalent to the LMI AT P + P Ai + 2αP ≤ 0. Cq ) equals H ∞. the largest lower bound on the decay rate provable via quadratic Lyapunov functions is obtained by maximizing α subject to (5. .. a GEVP. there exists a quadratic Lyapunov function proving that the decay rate is at least α if and only if there exists Q satisfying the LMI Q > 0. This electronic version is for personal use and may not be duplicated or distributed. . AQ + QAT + 2αQ ≤ 0. As an alternate condition.20) Therefore.α ∆ Then the optimal value of the GEVP is equal to the largest α such that H This can be seen by rewriting (5.α = sup { H(s) | Re s > −α} . . the condition that dV (x)/dt ≤ −2αV (x) for all trajectories is equivalent to the existence of λ ≥ 0 such that T AT P + P A + λCq Cq + 2αP T Bp P P Bp −λI ≤ 0.2) is at least α is that there exists Q and µ satisfying the LMI Q > 0. i i = 1. .22). Bp .e. (5. .23) ≤ 0. (5. . This is a GEVP in P and α. subject to P > 0. ..20) and P > 0. An alternate necessary and suﬃcient condition for the existence of a quadratic Lyapunov function proving that the decay rate of (5. L. negative of the maximum real part of the eigenvalues of A).22) Therefore. . P and λ. µ ≥ 0. λ ≥ 0 and (5.α .1 Quadratic Stability 67 This lower bound is sharp. the optimal value of the GEVP (5. (5. ∞.18) is the decay rate of the LTI system (which is the stability degree of A.22) as T (A + αI) P + P (A + αI) + λCq Cq T Bp P T < 1.e. P Bp −λI ≤ 0. i i = 1.5. i. The lower bound has a simple frequency-domain interpretation: Deﬁne the αshifted H∞ norm of the system by H ∞. and noting that the H∞ norm of the system (A + αI.

AT P + P Ai ≤ 0.2 Invariant Ellipsoids Quadratic stability can also be interpreted in terms of invariant ellipsoids. invariance of E is characterized by LMIs in Q or P . (5.29) Invariance of the ellipsoid E is equivalent to the existence of µ such that and T T µBp Dqp + QCq T −µ(I − Dqp Dqp ) T AQ + QAT + µBp Bp T T (µBp Dqp + QCq )T ≤ 0.4). i = 1. (5. • LTI systems: The corresponding LMI in P is P > 0. (5. Remark: Condition (5. . L.26) • Polytopic LDIs: The corresponding LMI in P is P > 0. For Q > 0. Q > 0. the condition that E be an invariant ellipsoid can be expressed in each case as an LMI in Q or its inverse P .24) is just the nonstrict version of condition (5. PLDIs and NLDIs. λ ≥ 0. . . or equivalently. its inverse.25) and the LMI in Q is Q > 0. . for LTI systems. (5. In summary. (5. i QAT + Ai Q ≤ 0. for all A ∈ Ω. (5. AT P + P A ≤ 0. invariance of the ellipsoid E is equivalent to the existence of λ such that P > 0.24) for all A ∈ Ω. .1) if for every trajectory x of the LDI. T AT P + P A + λCq Cq T (P Bp + λCq Dqp )T µ ≥ 0.28) • Norm-bound LDIs: Applying the S-procedure. i i = 1. . Thus. AT P + P A ≤ 0.30) ≤ 0. T P Bp + λCq Dqp T −λ(I − Dqp Dqp ) (5. It is easily shown that this is the case if and only if Q satisﬁes QAT + AQ ≤ 0.27) and the LMI in Q is Q > 0. . where P = Q−1 . x(t0 ) ∈ E implies x(t) ∈ E for all t ≥ t0 . AQ + P AT ≤ 0. Copyright c 1994 by the Society for Industrial and Applied Mathematics. let E denote the ellipsoid centered at the origin E = ξ ∈ Rn ξ T Q−1 ξ ≤ 1 . L.68 Chapter 5 Analysis of LDIs: State Properties 5. . . The ellipsoid E is said to be invariant for the LDI (5.

Q ≤ λI This electronic version is for personal use and may not be duplicated or distributed. p. .31). λmax (Q). p. . Q ≤ λI • Polytopic LDIs: The EVP in Q and λ is minimize subject to λ (5.34) • Norm-bound LDIs: The EVP in Q. (5. (5.26). .32) Minimum diameter The diameter of the ellipsoid E is 2 solving appropriate EVPs. . p. (5. j = 1. Q ≤ λI (5. λ and µ is minimize subject to λ (5. . (5.29). . We can minimize this quantity by • LTI systems: The EVP in the variables Q and λ is minimize subject to λ (5.32) (5. . j = 1. j = 1.27). (5. ≥ 0. . . vp }.28).2 Invariant Ellipsoids 69 5. . . The ellipsoid E contains the polytope P if and only if This condition may be expressed as an LMI in Q 1 vj T vj Q T vj Q−1 vj ≤ 1. up to a constant that depends on n.2. .33) • Norm-bound LDIs: The CP in P and λ is minimize subject to log det P −1 (5. We can minimize • Polytopic LDIs: The CP in P is minimize subject to log det P −1 (5. (5.32) √ det Q.30). • LTI systems: The CP in P is minimize log det P −1 subject to (5.5. this volume by solving appropriate CPs. . . (5. . P = Co{v1 .31). . (5. .25).1 Smallest invariant ellipsoid containing a polytope Consider a polytope described by its vertices.31).32) Minimum volume The volume of E is.31) or as an LMI in P = Q−1 as T vj P vj ≤ 1.

. x(t0 ) ∈ P (this knowledge may reﬂect measurements or prior assumptions). (5. the CP in the variables Q and µ is minimize subject to log det Q−1 (5. . An invariant ellipsoid E containing P then gives a bound on the state for t ≥ t0 . i. k k = 1. q . 5. .e. . (5.2..30).. k = 1. the CP in the variable Q is • Norm-bound LDIs: For NLDIs.35) • Polytopic LDIs: For PLDIs. . . the length of the minor axis) of invariant ellipsoids contained in P by solving EVPs: • LTI systems: For LTI systems.2 Largest invariant ellipsoid contained in a polytope We now consider a polytope described by linear inequalities: P = ξ ∈ Rn aT ξ ≤ 1. . (5. which is a set of linear inequalities in Q. q.35) √ det Q. we can guarantee that x(t) ∈ E for all t ≥ t0 . k ≤ 1. the CP in the variable Q is minimize subject to log det Q−1 (5.e. q. the EVP in the variables Q and λ is maximize subject to λ (5. λI ≤ Q Copyright c 1994 by the Society for Industrial and Applied Mathematics. (5. . i.35) k = 1.28). 5.70 Chapter 5 Analysis of LDIs: State Properties Remark: These results have the following use. the EVP in the variables Q and λ is maximize subject to λ (5. minimizing log det Q−1 will (Since the volume of E is. The polytope P represents our knowledge of the state at time t0 .26). up to a constant. λI ≤ Q • Polytopic LDIs: For PLDIs. minimize subject to log det Q−1 (5. The maximum volume of invariant ellipsoids contained in P can be found by solving CPs: • LTI systems: For LTI systems. (5.35) We also can ﬁnd the maximum minor diameter (that is. maximize the volume of E).35). . .26).28). The ellipsoid E is contained in the polytope P if and only if max aT ξ | ξ ∈ E k This is equivalent to aT Qak ≤ 1. . . .35).

and the problem of ﬁnding the smallest such γ and therefore the smallest T (for a ﬁxed α) is an EVP in the variables γ and Q: minimize subject to γ (5. . . . vp }. . . . i aT x ≤ 1. .5. (5. (5.36) This electronic version is for personal use and may not be duplicated or distributed.31). k = 1. . we can conclude that if x(0) ∈ P. the smallest bound on the return time provable via invariant ellipsoids.23).3 Bound on return time The return time of a stable LDI for the polytope P is deﬁned as the smallest T such that if x(0) ∈ P.. The ellipsoids found above can be interpreted as regions of safe initial conditions.35). (5. with a given decay rate α.21). • LTI systems: Let us consider a positive decay rate α. .36) • Norm-bound LDIs: For NLDIs.30).2 Invariant Ellipsoids 71 • Norm-bound LDIs: For NLDIs. (5.e. (5. λI ≤ Q Remark: These results can be used as follows. q = Co{v1 .36) • Polytopic LDIs: For PLDIs. with a given decay rate α. then x(t) ∈ P for all t ≥ T . k k = 1. the EVP in the variables Q. P = x ∈ Rn aT Qai ≤ γ.2. then E is an invariant ellipsoid.36) the constraint e−αT E ⊆ P is equivalent to the LMI in Q and γ where γ = e2αT . (5. i. . If we use both representations of the polytope. If Q > 0 satisﬁes QAT + AQ + 2αQ ≤ 0. initial conditions for which we can guarantee that the state always remains in the safe operating region. 5.31). . so that T is an upper bound on the return time. (5. is obtained by solving the EVP in the variables Q. The polytope P represents the allowable (or safe) operating region for the system. and moreover x(0) ∈ E implies that x(t) ∈ e −αt E for all t ≥ 0. µ and λ is maximize subject to λ (5. Therefore if T is such that e−αT E ⊆ P ⊆ E.31). q. Upper bounds on the return time can be found by solving EVPs. then x(t) ∈ P for all t ≥ T .19). γ and µ minimize subject to γ (5. is obtained by solving the EVP in the variables Q and γ minimize subject to γ (5. . the smallest bound on the return time provable via invariant ellipsoids. . (5.

see for instance. over 200 papers had been written about the system (5. Yakubovich was the ﬁrst to make systematic use of LMIs along with the S-procedure to prove stability of nonlinear control systems (see references [Yak62. KB93. BY89. [ZK88. Yak66. as stated in [PD90]. there is no fundamental diﬀerence between these and the older results of Yakubovich and Popov.8)) cannot be converted to a Riccati inequality or a frequency-domain criterion. the corresponding stability criteria are usually most easily expressed in the frequency domain. Wil71a]. is an NLDI. Popov [Pop73] and Willems [Wil71b. San65a. For other results and computational procedures for PLDIs see [PS82. The case when p and q are vector-valued signals has been considered much more recently. SS81]. PD90]. who used it to study the Lur’e system. KP87a. and result in inﬁnite-dimensional convex optimization problems.) These criteria are most useful when dealing with experimental data arising from frequency response measurements. ZK87. See also [AS79. which is further reﬁned in [Mei91]. where the authors develop a subgradient method for proving quadratic stability of a PLDI. SD83. where Horisberger and Belanger note that the problem of quadratic stability for a PLDI is convex (the authors write down the corresponding LMIs).37). which approximately solve these inﬁnite-dimensional problems.37). This approach was sparked by Popov [Pop62]. Kam83. BH89]. One of the ﬁrst occurrences of PLDIs is in [HB76]. BW65]. (Yakubovich calls this method the “method of matrix inequalities”. (5. The advantage of the input-output approach is that the systems are not required to have ﬁnite-dimensional state [Des65. See also [Bar70b]. SD84].37) where p and q are scalar. PS86. This problem came to be popularly known as the problem of “absolute stability in automatic control”. Zam66b. (The family of systems of the form (5. in [Pya68]. they have very strong connections with the input-output approach to study the stability of nonlinear. Other major contributors include Zames [Zam66a. Yak82]). generating diﬀerent special cases and solutions. points out that by 1968. As far as we know. where φ(q.) Pyatnitskii. AGG94]. Subsequently. ZF67]. various additional assumptions were made about φ. the paper [PS86]. In the original setting of Lur’e and Postnikov. Yak77. where the discrete-time counterpart of this Copyright c 1994 by the Society for Industrial and Applied Mathematics. The main idea of Yakubovich was to express the resulting LMIs as frequency-domain criteria for the transfer function G(s) = cq (sI − A)−1 bp .72 Chapter 5 Analysis of LDIs: State Properties Notes and References Quadratic stability for NLDIs Lur’e and Postnikov [LP44] gave one of the earliest stability analyses for NLDIs: They considered the stability of the system x = Ax + bp p. the authors remark that quadratic stability of NLDIs is equivalent to the H∞ condition (5. the ones most relevant to this book are undoubtedly from Yakubovich. Wil74a] outline the relationship between the problem of absolute stability of automatic control and quadratic optimal control. implemented in the robustness analysis software packages µ-tools and Km -tools [BDG+ 91. Log90]. CS92a]. the likely reason being that the LMI expressing quadratic stability of a general PLDI (which is just a number of simultaneous Lyapunov inequalities (5. Wil69b. Quadratic stability for PLDIs This topic is much more recent than NLDIs. However. Among these. PZP+ 92]. See also the papers and the book by Willems [Wil69a. Modern approaches from this viewpoint include Doyle’s µ-analysis [Doy82] and Safonov’s Km -analysis [Saf82.13). uncertain systems. ˙ q = cq x. t) can be any sector [−1. 1] nonlinearity. Bro66. DV75. φ was assumed to be a time-invariant nonlinearity. KR88. p = φ(q. While all the results presented in this chapter are based on Lyapunov stability techniques. San65b] and Brockett [Bro65. In [KPZ90. Sandberg [San64. t)q. . the article [KP87a]. Yak67. the newer results can be regarded as extensions of the works of Yakubovich and Popov to feedback synthesis and to the case of “structured perturbations” [RCDP93. EZKN90. Yak64. they are described in detail in the book by Narendra and Taylor [NT73].

see for example [PR91b] or [MP86]. Finally. this PLDI is not quadratically stable if there exist Q0 ≥ 0. . See also the articles by Gu et al.8). A2 } . Gu94]. where the discrete-time counterpart of this problem is considered. stability of an interconnected system. A necessary and suﬃcient condition for the Lyapunov function V deﬁned in (5. However. xT P2 x . [GZL90. A2 = 8 120 −9 . BT79. Here is a PLDI example: x = A(t)x. for an LDI to be stable without being quadratically stable. GL93b.g. Q1 ≥ 0 and Q2 ≥ 0. It is possible. −18 From (2. Stable LDIs that are not quadratically stable An LTI system is stable if and only if it is quadratically stable.7 1 1 1 satisfy these duality conditions.1 3 3 90 . e. 2 It can be veriﬁed that the matrices Q0 = 5. [BK93].. ˙ A1 = −100 0 A(t) ∈ Co {A1 . this is just Lyapunov’s stability theorem for linear systems (see e. One example is given in [CFA90]. Garofalo et al. proves that the PLDI is stable. such that T Q0 = A 1 Q1 + Q 1 A 1 + A 2 Q2 + Q 2 A T .Notes and References 73 problem is considered.2 2 2 24 . Vector Lyapunov functions The technique of vector Lyapunov functions also yields search problems that can be expressed as LMIPs.. p277]). not all zero. let us point out that quadratic Lyapunov functions have been used to determine estimates of regions of stability for general nonlinear systems. Earlier references on this topic include [Pya70b] which connects the problem of stability of an LDI to an optimal control problem. To show this.g. Here we already have quadratic Lyapunov functions for a number of subsystems. Q2 = 2. we use the S-procedure. [GCG93]. Necessary and suﬃcient conditions for LDI stability Much attention has also focused on necessary and suﬃcient conditions for stability of LDIs (as opposed to quadratic stability). See for example [Gha94. BT80]. Q1 = 0. P2 = 0 0 0 1 (5. and [Pya70a.38) to prove the stability of This electronic version is for personal use and may not be duplicated or distributed. 0 −1 . [Lya47. the problem is to ﬁnd an appropriate positive linear combination that proves.38) . however. the piecewise quadratic Lyapunov function V (x) = max xT P1 x. P1 = 14 −1 −1 1 . BD85]. and the survey article by Barmish et al.

t).) Fix an input w and let x and x denote any two solutions of (5. w. ˜ (5.3. Brockett uses the S-procedure to prove that a certain piecewise quadratic form is a Lyapunov function and ﬁnds LMIs similar to (5. In general. In other words. Suppose the LDI x ∈ Ωx is stable. see also [Vid78]. for ﬁxed input w. the diﬀerence between any two trajectories of (5.3).74 Chapter 5 Analysis of LDIs: State Properties the PLDI is the existence of four nonnegative numbers λ1 . w. . we have d (x − x) = f (x. |δi (t)| ≤ 1.39). w. i = 1. w.41) is stable for some diagonal nonsingular T . nq (5. DK71. λ3 = 1.40) converges to zero. computing such Lyapunov functions is computationally intensive. . For other examples of stable LDIs that are not quadratically stable. if DNLDI (5. λ4 such that AT P1 + P1 A1 − λ1 (P2 − P1 ) < 0. x(t) − x(t) converges to zero as t → ∞. t ∂x where Ω is convex. then so is DNLDI (5. Bro77. x(0) = ξ. (See [BC85]. This implies the system has fading memory.3) is also a trajectory (and vice versa) of the DNLDI x = Ax + Bp T −1 p. AT P2 + P2 A2 + λ4 (P2 − P1 ) < 0. and also [Zub55. Using the mean-value ˜ theorem from §4. t) = A(t)(x − x) ˜ x ˜ dt for some A(t) ∈ Ω. MP89. i. 1 AT P1 + P1 A2 − λ2 (P2 − P1 ) < 0. all trajectories converge to zero as ˙ t → ∞. Rap90. if not intractable.40). MV63. PR91b. Rap93.e.40) Relation between S-procedure for DNLDIs and scaling We discuss the interpretation of the diagonal matrix Λ in (5. Therefore. 2 (5. Finally. λ3 . . . Every trajectory of the DNLDI (5. Copyright c 1994 by the Society for Industrial and Applied Mathematics. .. In [Bro77]. 1 2 It can be veriﬁed that λ1 = 50. ˙ q = T Cq x + T Dqp T −1 p. t ≥ 0 } . Mol87]. See Brayton and Tong [BT79. VV85. PR91a. λ4 = 100 are such numbers. Nonlinear systems and fading memory Consider the nonlinear system x = f (x. an LDI is stable if and only if there is a convex Lyapunov function that proves it. Bro70] and Pyatnitskii [Pya71]. t) − f (˜. see Brockett [Bro65. One such Lyapunov function is V (ξ) = sup { x(t) | x satisﬁes (5. that is.15) as a scaling matrix . T is referred to as a scaling matrix. pi = δi (t)qi . BT80]. λ2 = 0.1. ˙ with ∂f ∈ Ω for all x.1). λ2 .41) for any diagonal nonsingular T . Since the LDI is stable. the system “forgets” its initial condition.39) AT P2 + P2 A1 + λ3 (P2 − P1 ) < 0.

is both necessary and suﬃcient (i.3) obtained using the S-procedure. This is precisely LMI (5. An obvious congruence. the condition for stability of the DNLDI (5. On the other hand. but this LMI yields only a suﬃcient condition for quadratic stability of LDI (5. not conservative). < 0. the size of the LMI grows exponentially with n q .15).15). yields the LMI condition in P > 0 and diagonal Λ > 0: T AT P + P A + Cq ΛCq T T BP P + Dqp ΛCq P Bp + Cq ΛDqp T Dqp ΛDq − Λ < 0.3). Representing a DNLDI as a PLDI: Implications for quadratic stability We saw in Chapter 4 that a DNLDI can be represented as a PLDI. T AT P + P A + C q T 2 Cq T (P Bp T −1 + Cq T 2 Dqp T −1 )T T P Bp T −1 + Cq T 2 Dqp T −1 T −(I − T −1 Dqp T 2 Dqp T −1 ) P > 0.8) for LDI (5.15) grows polynomially with n q . PS83]..41) as an NLDI. we require. we observe that while the quadratic stability condition (5. obtained by representing it as a PLDI. and Kamenetskii [Kam83] reduce the problem of numerical search for appropriate scalings and the associated Lyapunov functions to a convex optimization problem.Notes and References 75 Treating DNLDI (5. Discussion of this issue can be found in Kamenetskii [Kam83] and Pyatnitskii and Skorodinskii [PS82]. .8)). This relation between scaling and the diagonal matrices arising from the S-procedure is described in Boyd and Yang [BY89].13) for quadratic stability. checking quadratic stability for this PLDI requires the solution of L simultaneous Lyapunov inequalities in P > 0 (LMI (5. where the number L of vertices can be as large as 2nq . in the variables P > 0 and Λ > 0. for some diagonal nonsingular T . Saeki and Araki [SA88] also conclude convexity of the scaling problem and use the properties of M-matrices to obtain a solution. followed by the substitution T 2 = Λ. the size of ∆. the size of the LMI (5. and applying the condition (5. Comparing the two conditions for quadratic stability. Pyatnitskii and Skorodinskii [PS82.e. This electronic version is for personal use and may not be duplicated or distributed.1. If the system is represented as a DNLDI. From the results of §5.3). a suﬃcient condition for quadratic stability is given by the LMI (5.

.

or equivalently. ˙ q = Cq x. ˙ q = Cq x. . • Diagonal Norm-bound LDIs: DNLDIs have the form x = Ax + Bp p + Bw w. x(0) = 0 ∆ Rue = x(T ) .Chapter 6 Analysis of LDIs: Input/Output Properties 6. • Norm-bound LDIs: NLDIs have the form x = Ax + Bp p + Bw w.1 ] .1) We ﬁrst consider input-to-state properties of the LDI where w is an exogenous input signal. . [AL Bw. x = Ax + Bp p + Bw w. . . pT p ≤ q T q. i. . ˙ q = Cq x. . . pi = δi (t)qi . ˙ • Polytopic LDIs: Here the description is x = A(t)x + Bw (t)w. ∆(t) ≤ 1.1). Bw. • LTI systems: The system has the form x = Ax + Bw w.. i = 1. x = Ax + Bp p + Bw w.1. x.1 Input-to-State Properties x = A(t)x + Bw (t)w. .1 Reachable sets with unit-energy inputs Let Rue denote the set of reachable states with unit-energy inputs for the LDI (6.e.L ]} . nq . w satisfy (6. . |δi (t)| ≤ 1. T ≥0 0 77 . ˙ q = Cq x. or equivalently. i = 1. (6. T wT w dt ≤ 1. . 6. . ˙ [A(t) Bw (t)] ∈ Co {[A1 p = ∆(t)q. ˙ [A(t) Bw (t)] ∈ Ω. |pi | ≤ |qi |. . nq .1).

8) holds.4) which can also be expressed as an LMI in Q = P −1 Q > 0. (6.7) Inequality (6. we get T V (x(T )) − V (x(0)) ≤ wT w dt. Suppose that the function V (ξ) = ξ T P ξ. Copyright c 1994 by the Society for Industrial and Applied Mathematics.3) holds if and only if P > 0 and A(t)T P + P A(t) Bw (t) P T P Bw (t) −I ≤ 0 for all t ≥ 0. the reachable set is the ellipsoid −1 {ξ | ξ T Wc ξ ≤ 1}. with P > 0. the • LTI systems: Condition (6. . • Polytopic LDIs: For PLDIs.78 Chapter 6 Analysis of LDIs: Input/Output Properties We will bound Rue by ellipsoids of the form E = ξ | ξT P ξ ≤ 1 . T AQ + QAT + Bw Bw ≤ 0.6) we see that Q = Wc satisﬁes (6.i −I ≤ 0. −I ≤ 0. . AT P + P A P B w T Bw P wT w dt ≤ 1.1). (6. . where P > 0. (6. where Wc is the controllability Gramian. (6. T Ai P + P A i T Bw.5). satisﬁes dV (x)/dt ≤ w T w for all x. L. w satisfying (6. 0 Noting that V (x(0)) = 0 since x(0) = 0.5) Remark: For a controllable LTI system. condition (6. . Since Wc satisﬁes the Lyapunov equation T AWc + Wc AT + Bw Bw = 0. .3) is equivalent to the LMI in P P > 0. (6.3) (6. T 0 for every T ≥ 0. and every input w such that ellipsoid E contains the reachable set Rue . In other words. i = 1.7) holds if and only if P > 0. Thus the ellipsoidal bound is sharp for LTI systems. we get T V (x(T )) ≤ 0 wT w dt ≤ 1.8) Thus the ellipsoid E contains Rue if the LMI (6.i P P Bw.2) Integrating both sides from 0 to T . deﬁned by Wc = 0 ∆ ∞ T eAt Bw Bw eA T t dt. (6.

L. this is equivalent to the existence of P and T T A P + P A + λCq Cq P Bp P Bw T P > 0.9) where Q = P −1 . M ≥ 0. . . T Condition (6. we have for all ω. deﬁning Q = P −1 . (6. . Note that this condition implies that E is invariant.10) ≤ 0. .6.3) holds if and only if P > 0 and i = 1.12) For the remainder of this section we leave the extension to DNLDIs to the reader. we can write (6.3) holds if there exist Q and a diagonal matrix M such that Q > 0. T T AQ + QAT + Bw Bw + µBp Bp Cq Q T QCq −µI Using the S-procedure. i 79 i = 1. This electronic version is for personal use and may not be duplicated or distributed. P Bp −Λ 0 Equivalently.i ξ) . T T AQ + QAT + Bw Bw + Bp M Bp Cq Q T QCq −M P Bw 0 ≤ 0.1 Input-to-State Properties Alternatively. . . λnq ) satisfying T AT P + P A + Cq ΛCq T Bp P T Bw P P > 0.i ≤ 0. Bp P −λI 0 T 0 −I Bw P λ satisfying ≤ 0. We can optimize over this set in several ways. condition (6.i Bw.3) holds if and only if there exist Q and µ satisfying Q > 0.i ξ) (Cq. Equivalently. • Norm-bound LDIs: Condition (6. The results described above give us a set of ellipsoidal outer approximations of Rue . µ ≥ 0. . (6. λ ≥ 0. condition (6.8) as Q > 0. . . . (6. −I ≤ 0. T QAT + Ai Q + Bw. deﬁning Q = P −1 .11) • Diagonal Norm-bound LDIs: for all ξ and π that satisfy 2 πi ≤ (Cq. (6. It follows from the S-procedure that this condition holds if there exist P and Λ = diag(λ. Λ ≥ 0. .3) holds if and only if P > 0 and ξ T (AT P + P A)ξ + 2ξ T P (Bp π + Bw ω) − ω T ω ≤ 0 hold for every ω and for every ξ and π satisfying T π T π − ξ T Cq Cq ξ ≤ 0. ξ T (AT P + P A)ξ + 2ξ T P (Bp π + Bw ω) − ω T ω ≤ 0. . nq . .

then H does not intersect the reachable set if and only if aT Wc a < 1.8).4). Testing if the reachable set lies outside a half-space The reachable set Rue lies outside the half-space H = ξ ∈ Rn aT ξ > 1 if and only if an ellipsoid E containing Rue lies outside H.6). is found by minimizing log det P −1 subject to (6.2). • Norm-bound LDIs: For NLDIs. This is a CP. satisfying (6. H does not intersect the reachable set Rue if there exists Q and µ satisfying aT Qa < 1 and (6. 0 x0 lies outside the reachable set Rue if there exist P and λ satisfying xT P x0 > 1 0 and (6. Bw ) is controllable. that is. If (A. For NLDIs. Testing if the reachable set is contained in a polytope Since a polytope can be expressed as an intersection of half-spaces (these determine its “faces”). the volume of the reachable set is zero. Bw ) is uncontrollable. a suﬃcient condition for H not to intersect the reachable set Rue can be checked via an LMIP in the variable Q: aT Qa < 1 and (6. by replacing the objective function log det P −1 in the CPs by the objective function λmax (P −1 ).6). if (A.9). x0 does not belong to the reachable set if there exists P satisfying xT P x0 > 1 and (6.11). the minimum volume ellipsoid of the form (6. is obtained by minimizing log det P −1 over the variables P and λ subject to (6. Of 0 course. Bw ) is controllable. which corresponds to P = Wc . Therefore.5).8). then x0 does not belong to the reachable set if and −1 only if xT Wc x0 > 1. For NLDIs. and this CP will be unbounded below.3). If (A. the minimum volume ellipsoid of the form (6. This is an LMIP. then the CP ﬁnds the exact −1 reachable set. where Wc is deﬁned by the equation (6.3) by minimizing log det P −1 over the variable P subject to (6. another LMIP. the minimum value of ξ T Q−1 ξ over ξ satisfying aT ξ > 1 exceeds one.4).10).3). • Polytopic LDIs: For PLDIs. satisfying (6. Testing if a point is outside the reachable set The point x0 lies outside the reachable set if there exists an ellipsoidal outer approximation of Rue that does not contain it. 0 For PLDIs. a suﬃcient condition for a point x0 to lie outside the reachable set Rue can be checked via an LMIP in the variable P : xT P x0 > 1 and (6.80 Chapter 6 Analysis of LDIs: Input/Output Properties Smallest outer approximations • LTI systems: For LTI systems. We can also minimize the diameter of ellipsoids of the form (6.10). This is again a CP. For PLDIs.3). we can immediately use the results of the previous subsection in order to Copyright c 1994 by the Society for Industrial and Applied Mathematics.2). We easily see that min ξ T Q−1 ξ aT ξ > 1 = 1/(aT Qa). for LTI systems. This yields EVPs.2) with P satisfying (6. This is an LMIP. we can minimize the volume over ellipsoids E with P satisfying (6. where Wc is deﬁned by the equation (6. For LTI systems. . If (A. This is an LMIP. This is a CP. H does not intersect the reachable set if there exists Q satisfying aT Qa < 1 and (6. Bw ) is controllable.

then the ellipsoid E contains the reachable set Ruce . (Note we only obtain suﬃcient conditions. −R Therefore. . 2 T ≥0 wi dt ≤ 1. and R = diag(r1 . using the variable Q = P −1 . if there exist P and R such that (6. Remark: It turns out that with LTI systems. R ≥ 0.13) Then the ellipsoid E given by (6. i = 1. . condition (6. d V (x) ≤ w T Rw for all x and w satisfying (6. w satisfy (6.15) ≤ 0. we consider the set of reachable states with inputs whose components have unit-energy. to get T V (x(T )) − V (x(0)) ≤ T wT Rw dt. see the Notes and References. x(0) = 0. To prove this. Tr R = 1. Tr R = 1.1 Input-to-State Properties 81 check if the reachable set is contained in a polytope. ∆ T Ruce = x(T ) .14) is satisﬁed.1.6.1). . the reachable set is equal to the intersection of all such ellipsoids. we consider the set x. dt (6. we can write an equivalent LMI in Q and R: Q > 0. 6. . 0 Let us now consider condition (6. we get V (x(T )) ≤ wT Rw dt ≤ 1. . we integrate both sides of the last inequality from 0 to T .2 Reachable sets with componentwise unit-energy inputs We now turn to a variation of the previous problem.) We remark that we can use diﬀerent ellipsoidal outer approximations to check diﬀerent faces.13) for the various LDIs: • LTI systems: and R: For LTI systems. . . 0 Suppose there is a positive-deﬁnite quadratic function V (ξ) = ξ T P ξ. R ≥ 0 and diagonal. 0 Noting that V (x(0)) = 0 since x(0) = 0. . nw .14) ≤ 0. AT P + P A P B w T Bw P (6. Alternatively. QAT + AQ T Bw Bw −R (6. that is. .) This electronic version is for personal use and may not be duplicated or distributed.2) contains the reachable set.1). Tr R = 1. R ≥ 0 and diagonal. rnw ) such that P > 0.13) is equivalent to the LMI in P P > 0. (For more details on this.

(6. T ≥0 . . (6. T T A P + P A + λCq Cq P Bp P Bw T Bp P −λI 0 T 0 −R Bw P λ ≥ 0.82 Chapter 6 Analysis of LDIs: Input/Output Properties • Polytopic LDIs: For PLDIs.i P P Bw. by solving appropriate LMIPs. by solving CPs and EVPs. if there exists P .13) holds if and only if there exist P . µ ≥ 0. R ≥ 0 and diagonal.1. λ and R such that P > 0. Tr R = 1. we are interested in the set Rup = ∆ x(T ) x. Suppose that there exists a quadratic function V (ξ) = ξ T P ξ with P > 0. 6.13) holds if and only if P > 0.13). condition (6. w satisfying (6.i −R (6.18) Therefore. Tr R = 1.1).18) is satisﬁed.19) 0 −R As in the previous section. . condition (6. Thus. (6. x(0) = 0.13) is equivalent to Q > 0. Copyright c 1994 by the Society for Industrial and Applied Mathematics. With Q = P −1 . T T AQ + QAT + µBp Bp QCq Bw Cq Q −µI 0 ≤ 0. Tr R = 1. Therefore. and dV (x)/dt ≤ 0 for all x.3 Reachable sets with unit-peak inputs We consider reachable sets with inputs w that satisfy w T w ≤ 1.16) holds. if there exist P and R such that (6. condition (6. then the ellipsoid E contains the reachable set Ruce .20) . satisfying condition (6. L. R ≥ 0 and diagonal.2). respectively.13) is also equivalent to the existence of Q. . .16) ≤ 0. w T w ≤ 1 and V (x) ≥ 1.i Bw. • Norm-bound LDIs: For NLDIs. QAT + Ai Q i T Bw. .17) ≤ 0. applying the S-procedure. . T Bw ≤ 0. w satisfy (6. We can also check that a point does not belong to the reachable set or that a half-space does not intersect the reachable set. R ≥ 0 and diagonal. AT P + P A i i T Bw. i = 1. . R ≥ 0 and diagonal. Tr R = 1. L. we can ﬁnd the minimum volume and minimum diameter ellipsoids among all ellipsoids of the form (6. With Q = P −1 .1). condition (6. . T w w ≤ 1. then the ellipsoid E = ξ | ξ T P ξ ≤ 1 contains Ruce . λ and R such that (6.i −R (6. i = 1. R and µ such that Q > 0.

Once again.24) then the ellipsoid E contains the reachable set Rup . x w T AT P + P A + αP T Bw P P Bw −βI x w + β − α ≤ 0. L. (6. (6. and (6.26) is not an LMI in α and P .23) is not an LMI in α and P .20) holds if P > 0 and there exists α ≥ 0 satisfying T Ai P + P Ai + αP T Bw. β−α (6. .i −αI ≤ 0.23) is equivalent to the following inequality in Q = P −1 and α: QAT + AQ + αQ T Bw Bw −αI ≤ 0. . if (6.20) holds if there exist α ≥ 0 and β ≥ 0 such that for all x and w. if there exists P and α satisfying P > 0.26). α ≥ 0. and rewrite (6.i P P Bw. β0 ).20) is equivalent to P > 0 and T ξ T (AT P + P A)ξ + ξ T P Bw w + wT Bw P ξ ≤ 0 for any ξ and w satisfying wT w ≤ 1 and ξ T P ξ ≥ 1. and (6. .1 Input-to-State Properties 83 Then.25) • Polytopic LDIs: Condition (6.23) Therefore. . Note that inequality (6.21) as AT P + P A + αP T Bw P 0 ≤ 0. .22) holds for some (α0 . however. L. We can also rewrite condition (6.27) then the ellipsoid E contains the reachable set Rup . Using the S-procedure. Alternatively. Therefore. Let us consider now condition (6. condition (6.i Bw. (6.21) or equivalently T A P + P A + αP T Bw P 0 P Bw −βI 0 0 Clearly we must have α ≥ β. (6. (6. i = 1. then it holds for all α0 ≥ β ≥ β0 . the ellipsoid E given by (6. we conclude that condition (6. we note that condition (6.26) Therefore.6. if there exists P and α satisfying P > 0.28) This electronic version is for personal use and may not be duplicated or distributed. . for ﬁxed α it is an LMI in P .22) P Bw −αI ≤ 0. (6. we can assume without loss of generality that β = α. but is an LMI in P for ﬁxed α.2) contains the reachable set Rup .i −αI ≤ 0.20) for the various LDIs: • LTI systems: For LTI systems. . i = 1. condition (6. Next. (6.26) in terms of Q = P −1 as the equivalent inequality T QAi + Ai Q + αQ T Bw.23). α ≥ 0. . .

but is an LMI in P for ﬁxed α. (6. If there exists P . α and λ satisfying P > 0. Bp P T Bw P 0 −αI and (6.29) −λI 0 ≤ 0. We note again that inequality (6.84 Chapter 6 Analysis of LDIs: Input/Output Properties • Norm-bound LDIs: For NLDIs. Remark: Inequality (6. Therefore. . T π T π − ξ T Cq Cq ξ ≤ 0. α ≥ 0. 6.20) to hold is that there exist α ≥ 0 and λ ≥ 0 such that T T A P + P A + αP + λCq Cq P Bp P Bw T (6. z = Cz x.1 with an exponential time-weighting of the input w.2. respectively.29) is equivalent to the existence of µ ≥ 0 and α ≥ 0 satisfying (6.1. the existence of λ ≥ 0 and α ≥ 0 satisfying (6.1.24) for LTI systems.27) for PLDIs and condition (6.) For ﬁxed α.32) We now consider state-to-output properties for the LDI where A(t) Cz (t) ∈ Ω. condition (6. ˙ Copyright c 1994 by the Society for Industrial and Applied Mathematics.2 State-to-Output Properties x = A(t)x.29).23) can also be derived by combining the results of §6. it is possible to determine outer ellipsoidal approximations of the reachable set of LTI systems. a suﬃcient condition for (6.29) is not an LMI in P and α.33) • LTI systems: LTI systems have the form x = Ax.30) for NLDIs by forming the appropriate CPs.31).30) the ellipsoid EP −1 contains the reachable set Rup . we can minimize the volume or diameter among all ellipsoids E satisfying condition (6. See the Notes and References for details. λ ≥ 0. ˙ z = Cz (t)x (6. subject to componentwise unitpeak inputs. condition (6. ξ and π satisfying ω T ω ≤ 1. ξ T P ξ ≥ 1. As in §6.20) is also implied by the existence of µ ≥ 0 and α ≥ 0 such that T T QAT + AQ + αQ + µBp Bp QCq Bw (6.31) Cq Q −µI 0 ≤ 0. Deﬁning Q = P −1 . condition (6. using the S-procedure. or EVPs. (6. PLDIs and NLDIs.20) holds if and only if P > 0 and ξ T (AT P + P A)ξ + 2ξ T P (Bw ω + Bp π) ≤ 0 for every ω. T Bw 0 −αI (In fact.

.1 Bounds on output energy We seek the maximum output energy given a certain initial state. or equivalently x = Ax + Bp p. .32). we get V (x(T )) − V (x(0)) ≤ − z T z dt. T AT P + P A + Cz Cz ≤ 0. Suppose there exists a quadratic function V (ξ) = ξ T P ξ such that P > 0 and d V (x) ≤ −z T z.36) Therefore. ˙ q = Cq x.35) is equivalent to P > 0. ∞ max 0 z T z dt x = A(t)x. q = Cq x...33) holds. and the maximum is taken over A(t). . z = Cz x p = ∆(t)q.37) This electronic version is for personal use and may not be duplicated or distributed. ∆(t) ≤ 1. integrating both sides of the second inequality in (6. i = 1. pi = δi (t)qi . ˙ or equivalently x = Ax + Bp p. |pi | ≤ |qi |. nq . 6. . 0 for every T ≥ 0. (6. q = Cq x. nq .2.6. • LTI systems: In the case of LTI systems. |δi (t)| ≤ 1. . Since V (x(T )) ≥ 0. Cz (t) such that (6. z = Cz x. ˙ q = Cq x. we conclude that V (x(0)) = x(0)T P x(0) is an upper bound on the maximum energy of the output z given the initial condition x(0). • Norm-bound LDIs: NLDIs have the form x = Ax + Bp p.. pT p ≤ q T q.35) Then. We now derive LMIs that provide upper bounds on the output energy for the various LDIs. . AL CL . we obtain the best upper bound on the output energy provable via quadratic functions by solving the EVP minimize subject to x(0)T P x(0) P > 0. z = Cz x. (6. . z = Cz (t)x. where ˙ A(t) Cz (t) ∈ Co A1 C1 . i = 1. • Diagonal Norm-bound LDIs: DNLDIs have the form x = Ax + Bp p. (6.2 State-to-Output Properties 85 • Polytopic LDIs: PLDIs have the form x = A(t)x. condition (6.36) (6. (6. ˙ z = Cz x. . dt T for every x and z satisfying (6. ˙ z = Cz (t)x .35) from 0 to T . . .34) where x(0) is given.

.43) (6.i −I ≤ 0.35) is equivalent to P > 0. (6.44) Copyright c 1994 by the Society for Industrial and Applied Mathematics.i Cz. condition (6. With Q = P −1 . Using the S-procedure. . • Polytopic LDIs: For PLDIs. deﬁned by Wo = 0 ∆ ∞ T eA t Cz Cz eAt dt. . λ ≥ 0. • Norm-bound LDIs: For NLDIs. (6. T T AT P + P A + Cz Cz + λCq Cq T Bp P Therefore we obtain the smallest upper bound on the output energy (6.i Q T QCz.42) Deﬁning Q = P −1 . then an upper bound on the output energy (6.41) (6.39). and it is exactly equal to the output energy. condition (6. . (6. condition (6.35) is equivalent to Q > 0. (6.35) is equivalent to P > 0 and T ξ T (AT P + P A + Cz Cz )ξ + 2ξ T P Bp π ≤ 0 T for every ξ and π that satisfy π T π ≤ ξ T Cq Cq ξ. condition (6. an equivalent condition is the existence of P and λ such that P > 0. i for all t ≥ 0. (6.35) by solving the following EVP in the variables P and λ minimize subject to x(0)T P x(0) P > 0. Ai Q + QAT i Cz.43) ≤ 0. T Since Wo satisﬁes the Lyapunov equation T AT Wo + Wo A + Cz Cz = 0.34) is x(0) T Q−1 x(0). λ ≥ 0. A(t)T P + P A(t) + Cz (t)T Cz (t) ≤ 0 T AT P + P Ai + Cz. where Wo is the observability Gramian of the system.35).i ≤ 0. which is x(0)T Wo x(0).34) provable via quadratic functions satisfying condition (6.86 Chapter 6 Analysis of LDIs: Input/Output Properties In this case the solution can be found analytically. .40) Inequalities (6.41) Therefore the EVP corresponding to ﬁnding the best upper bound on the output energy provable via quadratic functions is minimize subject to x(0)T P x(0) P > 0. P Bp −λI (6. i = 1.38) it satisﬁes (6.40) hold if and only if the following LMI in the variable P holds: P > 0. AQ + QAT Cz Q T QCz −I ≤ 0. (6.35) is also equivalent to the LMI in Q Q > 0.39) If Q satisﬁes (6. L.

(6. vp . (6. . . With Q = P −1 . the smallest upper bound on the maximum output energy extractable from E provable via quadratic functions satisfying (6.44) p times. . This electronic version is for personal use and may not be duplicated or distributed. Therefore. T i = 1.35) is equivalent to for every ξ and π that satisfy 2 πi ≤ (Cq. and (6.35) is obtained by replacing the objective function x(0)T P x(0) in the EVPs (6. As a variation. The maximum of the p resulting optimal values is the smallest upper bound sought. Cq Q 0 −M Once again. Λ ≥ 0 and diagonal. .2 State-to-Output Properties 87 existence of µ ≥ 0 satisfying ≤ 0.35) is also equivalent to the existence of a diagonal matrix M ≥ 0 satisfying T T T AQ + QAT + Bp M Bp QCz QCq Cz Q −I 0 ≤ 0. . nq . .35) and the S-procedure by solving the following EVP in the variables P and Λ: minimize subject to x(0)T P x(0) P > 0. we leave it to the reader to extend the results in the remainder of the section to DNLDIs. Using the S-procedure. condition (6. which is known only to lie in a polytope P = Co{v1 . successively setting x(0) = v i .37). condition (6. a suﬃcient condition is that there exists a diagonal Λ ≥ 0 such that T T AT P + P A + Cz Cz + Cq ΛCq T Bp P P Bp −Λ ≤ 0. the exact value of the maximum extractable output energy is obtained.45) With Q = P −1 . . vp }. x(0)T P x(0) takes its maximum value on one of the vertices v1 . condition (6. Then.6. Cz Q −I 0 Cq Q 0 −µI T ξ T (AT P + P A + Cz Cz )ξ + 2ξ T P Bp π ≤ 0. .44) by the objective function λmax (X −1/2 P X −1/2 ). • Diagonal Norm-bound LDIs: For DNLDIs. . (6. .45) Therefore we obtain the smallest upper bound on the output energy (6. suppose that x(0) is only known to belong to the ellipsoid E = {ξ | ξ T Xξ ≤ 1}.42) or (6. .i ξ) (Cq. We ﬁrst note that since P > 0. . . For LTI systems. and solving the corresponding EVPs. . . p. Maximum output energy extractable from a set As an extension. we seek bounds on the maximum extractable energy from x(0). . i = 1.35) is also equivalent to the T T T AQ + QAT + µBp Bp QCz QCq Q > 0.34) provable via quadratic functions satisfying condition (6. the smallest upper bound on the extractable energy from P provable via quadratic Lyapunov functions for any of our LDIs is computed by solving (6.37).42). .i ξ) . (6.

Then z(t)T z(t) ≤ max ξ T Cz (t)T Cz (t)ξ ξ∈E for all t ≥ 0.47) • Polytopic LDIs: For PLDIs.44) by the objective function Tr X0 P and solving the corresponding EVPs. i = 1. Assume ﬁrst that the initial condition x(0) is known. (6. T AT P + P A + λCq Cq P Bp ≤0 T Bp P −λI (6.42).88 Chapter 6 Analysis of LDIs: Input/Output Properties Finally. Copyright c 1994 by the Society for Industrial and Applied Mathematics. We can express maxξ∈E ξ T Cz (t)T Cz (t)ξ as the square root of the minimum of δ subject to P Cz T Cz δI ≥ 0. if x0 is a random variable with E x0 xT = X0 . L (6.46). . Suppose E = {ξ | ξ T P ξ ≤ 1} is an invariant ellipsoid containing x(0) for the LDI (6. . then we obtain the small0 est upper bound on the expected output energy provable via quadratic functions by replacing the objective function x(0)T P x(0) in the EVPs (6. i • Norm-bound LDIs: For NLDIs.2. x(0)T P x(0) ≤ 1.46). (6. λ and δ minimize subject to δ P > 0. x(0)T P x(0) ≤ 1.46).32). λ ≥ 0. . and (6. x(0)T P x(0) ≤ 1. AT P + P Ai ≤ 0. .2 Bounds on output peak It is possible to derive bounds on z(t) using invariant ellipsoids.46) • LTI systems: The smallest bound on the output peak that can be obtained via invariant ellipsoids is the square root of the optimal value of the EVP in the variables P and δ minimize subject to δ P > 0. we obtain the smallest upper bound on z provable via invariant ellipsoids by taking the square root of the optimal value of the following EVP in the variables P . (6.37). . AT P + P A ≤ 0 (6. 6. we obtain the smallest upper bound on z provable via invariant ellipsoids by taking the square root of the optimal value of the following EVP in the variables P and δ minimize subject to δ P > 0.

(6. we conclude that φ ≤ sup T z T z dt x(T )T P x(T ) ≤ 1 ≤ sup x(T )T Qx(T ) x(T )T P x(T ) ≤ 1 = λmax (P −1/2 QP −1/2 ). that is.2. . This also reduces to an EVP. for all t ≥ 0. T w(t) = 0. w and z satisfying (6. x(0) = x0 . we assume that Dzw (t) = 0 for all t. we can impose a decay rate constraint on the output.6.1. An upper bound for φ can be computed by combining the ellipsoidal bounds on reachable sets from §6.1 Hankel norm bounds In this section. and consider the quantity T ∞ wT w dt ≤ 1. Suppose that P > 0 and Q > 0 satisfy d T x P x ≤ wT w.3 Input-to-Output Properties Remark: As variation of this problem. φ equals the Hankel norm.3. x(0) = 0.49) 6. We ﬁnally consider the input-output behavior of the LDI (6.50) for all x.1.50) by computing the square root of the minimum γ such that there exist P > 0 and Q > 0 satisfying T AT P + P A + P Bw Bw P ≤ 0. dt d T x Qx ≤ −z T z dt (6.48) where A(t) Cz (t) Bw (t) Dzw (t) ∈ Ω. we can compute an upper bound on the smallest γ such that the output z satisﬁes z(t) ≤ γe−αt .1. This electronic version is for personal use and may not be duplicated or distributed. Then.3 Input-to-Output Properties x = A(t)x + Bw (t)w.1 and the bounds on the output energy from §6. t > T ≥ 0 and the √ maximum is taken over A(t). For an LTI system. T AT Q + QA + Cz Cz ≤ 0. B(t) and C(t) satisfying (6. γP − Q ≥ 0. 0 ∆ φ = max z T z dt .1 and §6. 89 6.49). ˙ z = Cz (t)x + Dzw (t)w.48).1. given α > 0. ∞ • LTI systems: We compute the smallest upper bound on the Hankel norm provable via quadratic functions satisfying (6.2. from the arguments in §6. a name that we extend for convenience to LDIs.

. P . respectively. . • Polytopic LDIs: We compute the smallest upper bound on the Hankel norm provable via quadratic functions satisfying (6. the smallest upper bound on the Hankel norm provable via quadratic functions satisfying (6. P . i = 1. γP − Q ≥ 0. ˜ BT Q −νI p ˜ P − Q ≥ 0. If we introduce the new variables ˜ ˜ Q = Q/γ. and µ ≥ 0 satisfying T T A P + P A + λCq Cq P Bp P Bw T Bp P −λI 0 ≤ 0.e. .38).. T Bw P 0 A Q + QA T T + C z Cz T Bp Q + T µCq Cq −I γP − Q ≥ 0.i P ≤ 0.50) by computing the square root of the minimum γ such that there exist P > 0 and Q > 0 satisfying T T Ai P + P Ai + P Bw. Deﬁning Q = Q/γ. and µ.i Bw.i Cz.i ≤ 0. . and ν is to minimize γ such that T T A P + P A + λCq Cq P Bp P Bw T Bp P −λI 0 ≤ 0.90 Chapter 6 Analysis of LDIs: Input/Output Properties This is a GEVP in the variables γ. The corresponding EVP in the variables ˜ γ. L • Norm-bound LDIs: Similarly. ν = µ/γ and equivalent EVP in the variables γ. P . T Bw P 0 −I T ˜ T T ˜ + Cz Cz /γ + νCq Cq QBp ˜ A Q + QA ≤ 0. λ ≥ 0. P . Q. Q. T T Ai Q + QAi + Cz. ˜ P −Q≥0 i = 1. . .6) and (6. . and Q is then to minimize γ subject to T AT P + P A + P Bw Bw P ≤ 0. L ˜ This is a GEVP in the variables γ. i. P . It is possible to transform it into an EVP ˜ by introducing the new variable Q = Q/γ. T ˜ T ˜ Ai Q + QAi + Cz.i Bw.i /γ ≤ 0.i Cz. and Q. QBp −µI ≤ 0. . It can be found analytically as λmax (Wc Wo ) where Wc and Wo are the controllability and observability Gramians. and Q.i P ≤ 0.50) is computed by taking the square root of the minimum γ such that there exist P > 0. . and Q is to minimize γ subject to T T Ai P + P Ai + P Bw. an equivalent EVP ˜ in the variables γ. λ. This is a GEVP in the variables γ. Copyright c 1994 by the Society for Industrial and Applied Mathematics. λ. Q > 0. the solutions of the Lyapunov equations (6. the optimal value is exactly the Hankel norm. T ˜ T ˜ A Q + QA + Cz Cz /γ ≤ 0. ˜≥0 P −Q In this case. P .

and the supremum is taken over 2 all nonzero trajectories of the LDI.51) Then the L2 gain of the LDI is less than γ. with the initial condition x(0) = 0.2 L2 and RMS gains We assume Dzw (t) = 0 for simplicity of exposition. Remark: It can be easily checked that if (6.51) from 0 to T . and the RMS gain is deﬁned as sup RMS(w)=0 RMS(z) . P > 0. which also equals the H∞ norm of its transfer matrix. Remark: Assuming that (A. The LDI is said to be nonexpansive if its L2 gain is less than one. We deﬁne the L2 gain of the LDI (6.3. d V (x) + z T z − γ 2 wT w ≤ 0 dt for all x and w satisfying (6. since V (x(T )) ≥ 0. . (6. Cz ) is minimal. P > 0. suppose there exists a quadratic function V (ξ) = ξ T P ξ. This electronic version is for personal use and may not be duplicated or distributed. and γ ≥ 0 such that for all t. this EVP gives the exact value of the L2 gain of the LTI system. then γ is also an upper bound on the RMS gain of the LDI.48). RMS(w) Now reconsider condition (6. • LTI systems: Condition (6. This is an EVP.48) as the quantity sup w 2 =0 z w 2 2 ∞ where the L2 norm of u is u 2 = 0 uT u dt. starting from x(0) = 0.51) is equivalent to T A T P + P A + C z Cz T Bw P P Bw −γ 2 I ≤ 0. this implies z w 2 2 ≤ γ.51) for LDIs.52) Therefore.51) holds for V (ξ) = ξ T P ξ. to get T V (x(T )) + 0 z T z − γ 2 wT w dt ≤ 0. where the root-meansquare (RMS) value of ξ is deﬁned as RMS(ξ) = ∆ lim sup T →∞ 1 T T 1/2 ξ T ξ dt 0 .52). Bw . Now. we integrate (6.3 Input-to-Output Properties 91 6. Cz (sI − A)−1 Bw ∞ . (6. we compute the smallest upper bound on the L2 gain of the LTI system provable via quadratic functions by minimizing γ over the variables P and γ satisfying conditions P > 0 and (6.6. To show this.

then the existence of P > 0 and λ ≥ 0 satisfying (6. subject to P > 0. . i = 1. .i Cz. then the existence of P > 0 satisfying (6. • Norm-bound LDIs: For NLDIs. . .i T Bw. condition (6.i Bw. .i ξi )T (Cq.51) is equivalent to T ξ T (AT P + P A + Cz Cz )ξ + 2ξ T P (Bp π + Bw w) − γ 2 wT w ≤ 0 This is true if and only if there exists λ ≥ 0 such that T T T A P + P A + Cz Cz + λCq Cq P Bp P Bw T Bp P −λI 0 ≤ 0. condition (6.54) 0 −γ I 2 for all ξ and π satisfying T πi πi ≤ (Cq.i i T QCz. i = 1. .92 Chapter 6 Analysis of LDIs: Input/Output Properties Assume Bw = 0 and Cz = 0. λ ≥ 0 and (6.i P P Bw.52) is equivalent to the existence of Q > 0 satisfying T AQ + QAT + Bw Bw /γ 2 Cz Q T QCz −I ≤ 0. .i0 = 0.53) if and only if there exists Q > 0 satisfying T QAT + Ai Q + Bw.54).51) is equivalent to T ξ T (AT P + P A + Cz Cz )ξ + 2ξ T P (Bp π + Bw w) − γ 2 wT w ≤ 0 for all ξ and π satisfying T π T π ≤ ξ T Cq Cq ξ. and j0 for which Cz.i −γ 2 I ≤ 0. If Bw = 0 and Cz = 0. . Copyright c 1994 by the Society for Industrial and Applied Mathematics. . i = 1.51) is equivalent to T T Ai P + P Ai + Cz.j0 = 0. We get the smallest upper bound on the L2 gain provable via quadratic functions by minimizing γ (over γ and P ) subject to (6. L. nq .i ξi ). . . which is an EVP. Then there exists P > 0 satisfying (6.i Q −γ 2 I ≤ 0. T Bw P (6. Therefore. P and λ.54) is equivalent to the existence of Q > 0 and µ ≥ 0 satisfying T T T T QAT + AQ + Bw Bw + µBp Bp QCq QCz Cq Q −µI 0 ≤ 0. (6. . . • Polytopic LDIs: Condition (6.53) and P > 0. Cz Q 0 −γ 2 I • Diagonal Norm-bound LDIs: For DNLDIs. we obtain the best upper bound on the L2 gain provable via quadratic functions by minimizing γ over the variables γ.i Cz.53) Assume there exists i0 for which Bw. L.

56) holds for all trajectories with x(0) = 0 and all T ≥ 0. integrating (6.e. This electronic version is for personal use and may not be duplicated or distributed. we conclude wT z − ηwT w dt ≥ 0.. i. subject to P > 0. Suppose that there is a quadratic function V (ξ) = ξ T P ξ. P > 0. . subject to P > 0 and (6.48). such that for all x and w satisfying (6. . If Bw = 0 and Cz = 0. µnq ) ≥ 0 satisfying T T T T QAT + AQ + Bw Bw + Bp M Bp QCq QCz Cq Q −M 0 ≤ 0.3 Dissipativity The LDI (6. condition (6. • LTI systems: For LTI systems. (6. It is said to have dissipation η if 0 wT z − ηwT w dt ≥ 0 (6.55) Bp P −Λ 0 ≤ 0.55).6.54) is equivalent to the existence of Q > 0 and M = diag(µ1 .57) for the various LDIs. T Bw P 0 −γ 2 I We obtain the best such upper bound by minimizing γ over the variables γ. a suﬃcient condition is that there exist a diagonal Λ ≥ 0 such that T T T A P + P A + Cz Cz + Cq ΛCq P Bp P Bw T (6. . Thus passivity corresponds to nonnegative dissipation.48) is said to be passive if every solution x with x(0) = 0 satisﬁes T 0 wT z dt ≥ 0 T for all T ≥ 0.3 Input-to-Output Properties 93 Using the S-procedure. the largest number η such that (6.57) is equivalent to AT P + P A T Bw P − C z T P B w − Cz T 2ηI − (Dzw + Dzw ) ≤ 0. dt (6. will be called its dissipativity. λ ≥ 0 and (6. Now reconsider condition (6. an EVP.57) from 0 to T with initial condition x(0) = 0 yields T V (x(T )) − T 0 0 2wT z − 2ηw T w dt ≤ 0. P and λ. d V (x) − 2w T z + 2ηw T w ≤ 0.3. which implies that the dissipativity of the LDI is at least η. Cz Q 0 −γ 2 I 6. The largest dissipation of the system. Since V (x(T )) ≥ 0. then the existence of P > 0 and λ ≥ 0 satisfying (6.58).58) We ﬁnd the largest dissipation that can be guaranteed using quadratic functions by maximizing η over the variables P and η.57) Then. . .56) holds.

i T 2ηI − (Dzw. . Assuming that the system (6. L. using the S-procedure.63) Bw P − C z 2ηI − (Dzw + Dzw ) 0 ≤ 0.i − Cz. Deﬁning Q = P −1 .61) We ﬁnd the largest dissipation that can be guaranteed via quadratic functions by maximizing η over the variables P and η satisfying P > 0 and (6.i + Dzw. condition (6. This is an EVP. (6. Cz ) is minimal. condition (6. ˙ ˜ z = T Cz (t)x. L. . .i − Cz.64) Bw − C z Q 2ηI − (Dzw + Dzw ) 0 ≤ 0. . . T Bp P 0 −λI Therefore we can ﬁnd the largest dissipation provable with a quadratic Lyapunov function by maximizing η (over η and P ).i P − Cz.58) is equivalent to inf QAT + AQ T Bw − C z Q T Bw − QCz T 2ηI − (Dzw + Dzw ) (6. the optimal value of this EVP is exactly equal to the dissipativity of the system.57) is true if and only if there exists a nonnegative scalar λ such that T T T A P + P A + λCq Cq P Bw − Cz P Bp T T (6.65). the condition (6.i ≤ 0.i ) T P Bw.i ≤ 0. . (6. We will see that scaling enables us to conclude “componentwise” results for the system (6. Deﬁning Q = P −1 . Copyright c 1994 by the Society for Industrial and Applied Mathematics.62) • Norm-bound LDIs: For NLDIs. . If we deﬁne Q = P −1 .48) has as many inputs as outputs. (6. Cq Q 0 −µI 6. (6.94 Chapter 6 Analysis of LDIs: Input/Output Properties Assuming that the system (A.59) ≤ 0. which has the interpretation of a scaling. which can be expressed in terms of the transfer matrix H(s) = Cz (sI − A)−1 Bw + Dzw as λmin (H(s) + H(s)∗ ) . subject to P > 0 and (6. Bw . Res>0 2 This follows from the PR lemma. we consider the system x = A(t)x + Bw (t)T −1 w.3.i ) T Bw.60) • Polytopic LDIs: For PLDIs.i + Dzw. .63). i = 1. its optimal value is a lower bound on the dissipativity of the PLDI.i − QCz.61).57) is equivalent to QAT + Ai Q i T Bw.4 Diagonal scaling for componentwise results We assume for simplicity that Dzw (t) = 0. condition (6.57) is equivalent to AT P + P A i i T Bw. condition (6.57) is equivalent to the existence of µ ≥ 0 satisfying T T T QCq QAT + AQ + µBp Bp Bw − QCz T T (6. This is an EVP.i Q T 2ηI − (Dzw. i = 1. .65) where T is a positive-deﬁnite diagonal matrix. ˜ x(0) = 0.

. This electronic version is for personal use and may not be duplicated or distributed. the scaled L2 gain (with scaling T ) is less than γ if there exist λ ≥ 0. L. S diagonal. . .nz w i zi ≤ α. . sup w ˜ 2 =0 z ˜ w ˜ 2 2 2 2 = ≥ sup sup w 2 =0 wi 2 =0 nz 2 i ti nz 2 i ti zi 2 2 . P > 0.6. .. such that the following LMI holds: T T AT P + P A + Cz SCz + λCq Cq T Bp P T Bw P P Bp −λI 0 P Bw 0 ≤ 0. tnz ). the desired inequality follows.66) The optimal scaled L2 gain γ is therefore obtained by minimizing γ over (γ. P and S) subject to (6.i i T Bw. minimizing γ 2 subject to the constraint (6. wi We show this as follows: For every ﬁxed scaling T = diag(t1 . the scaled L2 gain is guaranteed to be lower than γ if there exists S > 0..67) Therefore. • Norm-bound LDIs: For NLDIs.66). −γ 2 S (6. i = 1.. 95 sup w ˜ 2 =0 z ˜ w ˜ 2 2 . wi 2 2 zi wi 2 2 2 2 Since the inequality is true for all diagonal nonsingular T . and S > 0. .i SCz.i P P Bw.i −γ 2 S < 0. • LTI systems: The L2 gain for the LTI system scaled by T is guaranteed to be less than γ if there exists P > 0 such that the LMI T AT P + P A + Cz SCz T Bw P P Bw −γ 2 S <0 holds with S = T T T . The smallest scaled L2 gain of the system can therefore be computed as a GEVP. and P > 0 which satisfy T AT P + P Ai + Cz. . (6. • Polytopic LDIs: For PLDIs.. with S diagonal. This is a GEVP. P > 0. . S > 0 and S diagonal. We now show how we can obtain bounds on the scaled L2 gain using LMIs for LTI systems and LDIs. . the “scaled L2 gain” α= inf T diagonal.3 Input-to-Output Properties Consider for example.67) is an EVP. T >0 The scaled L2 gain has the following interpretation: max sup 2 =0 i=1.

R > 0. Integral quadratic constraints were introduced by Yakubovich. considered in §10. As before. AT P + P A BT P PB −W ≤ 0. A variation of this problem is considered in [SZ92]. t→∞ lim x(t) = 0 is less than one. We begin by observing that a point x0 belongs to the reachable set if and only if the optimal value of the problem ∞ minimize subject to max i 0 2 wi dt x = −Ax + Bw w. we refer the reader to the articles of Yakubovich (see [Yak92] and references therein). We now derive LMI conditions that are both necessary and suﬃcient. For a general study of these. AT P + P A B P T PB −R ≤ 0. . P > 0. This is a multi-criterion convex quadratic problem.8.96 Chapter 6 Analysis of LDIs: Input/Output Properties Notes and References Integral quadratic constraints In §8. x0 belongs to the reachable set if and only the optimal value of the problem ∞ minimize 0 wT W w dt x = −Ax + Bw w. Meg92b].2 (and §10. P > 0. ˙ x(0) = x0 . the problem reduces to checking whether the LMI T x0 P x0 < 1. In this case. 0 Many of the results derived for NLDIs and DNLDIs then become necessary and suﬃcient. diagonal.9) we consider a generalization of the NLDI in which the pointwise condition pT p ≤ q T q is replaced by the integral constraint ∞ 0 ∞ pT p dt ≤ q T q dt. This is equivalent to infeasibility of the LMI (in the variables P and W ): T x0 P x0 − Tr W Q ≥ 0. and also Megretsky [Meg93. Copyright c 1994 by the Society for Industrial and Applied Mathematics. or equivalently ∞ 0 wT W w dt ≤ Tr W Q for every W ≥ 0. They get necessary conditions characterizing the corresponding reachable set. This shows that the reachable set is the intersection of ellipsoids ξ | ξ T P ξ ≤ 1 satisfying equation (6. where the authors consider inputs w ∞ satisfying 0 wwT dt ≤ Q.1. Meg92a.2). Tr R = 1 is feasible. ˙ x(0) = x0 . t→∞ subject to lim x(t) = 0 is less than Tr W Q for every W ≥ 0.14). Reachable sets for componentwise unit-energy inputs We study in greater detail the set of states reachable with componentwise unit-energy inputs for LTI systems (see §6.

§4.25).2.23) can be interpreted as an exponential time-weighting procedure. Sabin and Summers [SS90] study the approximation of the reachable set via quadratic functions. where M 2 = maxt≥0 cz (t)T Qcz (t). We note that the resulting bounds can be quite conservative. Now suppose that P > 0 satisﬁes condition (6.. .3) and bounds on output peak (§6. A(t)Q + QA(t)T + αQ + 1 bw (t)bw (t)T < 0. Fix α > 0.2. with new exponentially time-weighted variables xT (t) = eα(t−T )/2 x(t) and v(t) = eα(t−T )/2 w(t) as xT = A(t) + ˙ α I xT + Bw (t)v.23). Since LMI (6. the overshoot) can be found by combining the results on reachable sets with unit-peak inputs (§6. the positivedeﬁnite matrix describing the ellipsoidal approximation satisﬁes a matrix diﬀerential equation. we conclude that maxt≥0 |s(t)| does not exceed M . x(0) = 0. See also [FG88]. we have v T v dτ ≤ 1/α.68) The results of §6.Notes and References These are precisely the conditions obtained in [SZ92]. The technique used to derive LMI (6.68) is independent of T . Then for every T > 0. This can be used to ﬁnd LMI-based bounds on overshoot for LTI systems. terms of which closely resemble those in the matrix inequalities described in this book.2.3.1. (6. the ellipsoid x | xT P x ≤ 1 contains the reachable set with unit-peak inputs. an upper bound on maxt≥0 |s(t)| (i. x(T )T P x(T ) ≤ 1.1 imply that xT (T ) satisﬁes xT (T )T P xT (T ) ≤ 1.3] considers the more general problem of time-varying ellipsoidal approximations of reachable sets with unit-peak inputs for time-varying systems. Suppose there exist Q > 0 and α > 0 satisfying (6.1. 97 Reachable sets with unit-peak inputs Schweppe [Sch73. This electronic version is for personal use and may not be duplicated or distributed. ˙ s = cz (t)x. subject to a unit-step input: x = A(t)x + bw (t).e. 2 T 0 xT (0) = 0. A survey of the available techniques can be found in the article by Gayek [Gay91]. α Then. rewrite LDI (6. PLDIs and NLDIs. from §6. Since a unit-step is also a unit-peak input.3 can be used to ﬁnd a bound on the step response peak for LDIs. which we rewrite for convenience as (A + αI/2)T P + P (A + αI/2) T Bw P P Bw −αI ≤ 0. A(t) cz (t) bw (t) 0 ∈ Ω. Since w(t)T w(t) ≤ 1.2). Consider a single-input single-output LDI.1). These LMIs can be used to determine state-feedback matrices as well (see the Notes and References of Chapter 7).1. therefore. Bounds on overshoot The results of §6.

2.47) turns out to be 2n − 1 times the actual maximum value of the impulse response.2. In this case. which is used in nonlinear control [Kok92].e. i. where n is the dimension of a minimal realization of the system [Fer93]. the dynamics of this system are widely spaced. for every LTI system. Copyright c 1994 by the Society for Industrial and Applied Mathematics. consider the LTI system with transfer function H(s) = 1 s + s1 n i=2 s − si . that is. We conjecture that the bound can be no more conservative.2 with those on the peak value of the output from §6.. The bound on the peak value of the impulse response of the system computed via the EVP (6.98 Chapter 6 Analysis of LDIs: Input/Output Properties Bounds on impulse response peak for LTI systems The impulse response h(t) = ceAt b of the single-input single-output LTI system x = Ax + bu. The bound obtained this way can be quite conservative. P b = cT . For instance. ˙ x(0) = 0. so that the maximum value of the impulse response is cb and is attained for t = 0. . the bound on the peak of the impulse response computed via the EVP (6. This is an illustration of the fact that passive systems have nice “peaking” properties.47) can be no more than 2n − 1 times the actual maximum value of the impulse response. Examples for which the bound obtained via the EVP (6. we can compute a bound on the impulse response peak by combining the results on invariant ellipsoids from §5. Therefore. s + si where si+1 si > 0.47) is sharp include the case of passive LTI systems. y = cx is just the output y with initial condition x(0) = b and zero input. we know there exists a positive-deﬁnite P such that AT P +P A ≤ 0.

t. e. 0.2) satisﬁes certain properties or speciﬁcations. u. (7. As an example. and the matrix K is called the state-feedback gain. t) = 0. w and u. z = Cz (t)x + Dzw (t)w + Dzu (t)u. This yields the closed-loop LDI x = (A(t) + Bu (t)K) x + Bw (t)w.3. (7. the methods of this chapter can be used to synthesize a linear state-feedback for some nonlinear. u. ˙ z = (Cz (t) + Dzu (t)K) x + Dzw (t)w. g(0. polytope. ˙ A(t) Cz (t) Bw (t) Dzw (t) Bu (t) Dzu (t) ∈ Ω. this is called (linear. and suppose that u = Kx. 0.g. u. ∂f ∂w ∂g ∂w and f (0. t). Here u : R+ → Rnu is the control input and w : R+ → Rnw is the exogenous input signal . ∂g ∂u 99 . for all x. w. w.2) In this chapter we consider the state-feedback synthesis problem. singleton.. ∂f ∂u (x. w. t) ∈ Ω. 0. i. ˙ provided we have ∂f ∂x ∂g ∂x z = g(x. Let K ∈ Rnu ×n .1 Static State-Feedback Controllers x = A(t)x + Bw (t)w + Bu (t)u.e..1) We consider the LDI where Ω has one of our special forms (i. Since the control input is a linear function of the state. constant) state-feedback . 0.e.Chapter 7 State-Feedback Synthesis for LDIs 7. the problem of ﬁnding a matrix K so that the closed-loop LDI (7. time-varying systems. stability.. image of a unit ball under a matrix linear-fractional mapping). t). we can synthesize a state-feedback for the system x = f (x. t) = 0. Remark: Using the idea of global linearization described in §4.

we have x = Ax + Bu u + Bp p. ˙ • Polytopic LDIs: PLDIs are given by x = A(t)x + Bu (t)u. or equivalently.1 Quadratic stabilizability The system (7. (7. i = 1. (7. ˙ [A(t) Bu (t)] ∈ Co {[A1 Bu. . (7.8) Neither of these conditions is jointly convex in K and P or Q. q = Cq x + Dqp p + Dqu u.3) We consider the LDI • LTI systems: For LTI systems. ˙ q = Cq x + Dqp p + Dqu u. [AL Bu. .3) becomes x = Ax + Bu u + Bp p. ˙ p = ∆(t)q.4) is (quadratically) stable if and only if there exists P > 0 such that (A + Bu K)T P + P (A + Bu K) < 0. Equivalently. .2. Substituting into (7. q = Cq x + Dqu u + Dqp p ∆(t) ≤ 1. i = 1. The LTI system (7. but by a simple change of variables we can obtain an equivalent condition that is an LMI. (7. . nq . .1 ] . . there exists Q > 0 such that Q(A + Bu K)T + (A + Bu K)Q < 0. |δi (t)| ≤ 1.2) is quadratically stable (hence. .5) (7. Deﬁne Y = KQ. equation (7. . Equivalently.1) is said to be quadratically stabilizable (via linear state-feedback) if there exists a state-feedback gain K such that the closed-loop system (7. so that for Q > 0 we have K = Y Q−1 . .6) • Diagonal Norm-bound LDIs: For DNLDIs. .7) 7.100 Chapter 7 State-Feedback Synthesis for LDIs 7.9) Copyright c 1994 by the Society for Industrial and Applied Mathematics. we have x = Ax + Bu u + Bp p. pT p ≤ q T q.4) • Norm-bound LDIs: For NLDIs. (7.3) becomes x = Ax + Bu u. (7. Quadratic stabilizability can be expressed as an LMIP. . . (7. ˙ pi = δi (t)qi . • LTI systems: Let us ﬁx the matrix K. ˙ [A(t) Bu (t)] ∈ Ω. ˙ |pi | ≤ |qi |. (7.3) becomes x = Ax + Bu u + Bp p. q = Cq x + Dqp p + Dqu u. L. .8) yields T AQ + QAT + Bu Y + Y T Bu < 0.2 State Properties x = A(t)x + Bu (t)u. stable).L ]} .

• Polytopic LDIs: For the PLDI (7.i Y + Y T Bu.7. i i = 1. (7. Thus. .4) with state-feedback u = Y Q−1 x.i < 0.11).2 State Properties 101 which is an LMI in Q and Y . If Q > 0 satisﬁes the LMI (7. . L. For any Q > 0 satisfying (7.2.10). another equivalent condition is ˜ ˜T Bu AQ + QAT Bu < 0.11) implies that such a scalar exists).11). thus reducing the number of variables by one. these LMI conditions are necessary and suﬃcient for stabilizability. T µDqp Bp + (Cq + Dqu K)Q T −µ(I − Dqp Dqp ) This condition has a simple frequency-domain interpretation: The H∞ norm of the transfer matrix from p to q for the LTI system x = (A + Bu K)x + Bp p.10) holds (condition (7. the NLDI (7. the same argument applies.6.9) holds.2: There exist Q > 0 and a scalar σ such that T AQ + QAT − σBu Bu < 0. For LTI systems.4) is quadratically stabilizable if and only if there exist Q > 0 and Y such that the LMI (7. we conclude that the NLDI is quadratically stabilizable if there exist Q > 0. µ > 0 and Y such that T AQ + QAT + µBp Bp T T T µBp Dqp + QCq + Y T Dqu T (7. a stabilizing state-feedback gain is T given by K = −(σ/2)Bu Q−1 . involving fewer variables.14)) if there exist Q > 0 and µ > 0 such that AQ + QAT + Bu KQ T T µBp Dqp + Q(Cq + Dqu K) T T + QK T Bu + µBp Bp < 0. . then the quadratic function V (ξ) = ξ T Q−1 ξ proves (quadratic) stability of system (7. a T stabilizing state-feedback gain is K = −(σ/2)Bu Q−1 . (7. . In terms of linear system theory. It allows us to recast the problem of ﬁnding a state-feedback gain as an LMIP with variables that include Q and Y . q = (Cq + ˙ Dqu K)x + Dqp p is less than one. or (7. (7. where σ is any scalar such that (7. we can without loss of generality take σ = 1. stabilizability is equivalent to the condition that every unstable mode be controllable.12) • Norm-bound LDIs: With u = Kx. If this LMI is feasible.11) ˜ where Bu is an orthogonal complement of Bu .13) +Bu Y + Y T Bu < 0. it is not hard to show that this is equivalent to feasibility of the LMIs (7.10).9).10) Since we can always assume σ > 0 in LMI (7. An alternate equivalent condition for (quadratic) stabilizability.5). the system (7. From the elimination procedure of §2. T µDqp Bp + Cq Q + Dqu Y T − µ(I − Dqp Dqp ) This electronic version is for personal use and may not be duplicated or distributed. Quadratic stabilizability is equivalent to the existence of Q > 0 and Y with T QAT + Ai Q + Bu. Remark: We will use the simple change of variables Y = KQ many times in the sequel. With Y = KQ. and since the LMI is homogeneous in Q and σ. the state-feedback gain K is recovered as K = Y Q−1 . can be derived using the elimination of matrix variables described in §2.6) is quadratically stable (see LMI (5.10). . (7.6.

a stabilizing T state-feedback gain is K = (−σ/2)Bu Q−1 . the DNLDI (7. .6) is zero in order to simplify the discussion. T Dqp M Bp + (Cq + Dqu K)Q T −(M − Dqp M Dqp ) This condition leads to a simple frequency-domain interpretation for quadratic stabilizability of DNLDIs: Let H denote the transfer matrix from p to q for the LTI system x = (A + Bu K)x + Bp p.) In the remainder of this chapter we will assume that Dqp in (7. we can set µ = 1.2. The latter condition is always satisﬁed since the NLDI is well-posed. M −1/2 HM 1/2 ∞ ≤ 1. (7. − σBu Bu + µBp Bp (7.7) is quadratically stable (see LMI (5. We say that the ellipsoid E = ξ ∈ Rn ξ T Q−1 ξ ≤ 1 Copyright c 1994 by the Society for Industrial and Applied Mathematics. There is no analytical method for checking this condition.14) T T T T Cq Q + µDqp Bp − σDqu Bu −µ(I − Dqp Dqp ) − σDqu Dqu T −µ(I − Dqp Dqp ) < 0.6. equivalent condition for quadratic stabilizability is expressed in ˜ terms of the orthogonal complement of [B T DT ]T .15) (Again.16)) if there exist Q > 0. which we denote by G: u qu ˜ GT T AQ + QAT + µBp Bp T Cq Q + µDqp Bp T T QCq + µBp Dqp T −µ(I − Dqp Dqp ) ˜ G < 0. By homogeneity.2 Holdable ellipsoids Just as quadratic stability can also be interpreted in terms of invariant ellipsoids. q = (Cq + Dqu K)x + Dqp p. T Dqp M Bp + Cq Q + Dqu Y T − (M − Dqp M Dqp ) As before.14) with µ = 1. we can interpret quadratic stabilizability in terms of holdable ellipsoids. . With Y = KQ. and obtain equivalent LMIs in fewer variables. 7. . we conclude that the NLDI is quadratically stabilizable if there exist Q > 0. • Diagonal Norm-bound LDIs: With u = Kx. we can freely set µ = 1 in this LMI. M = diag(µ1 . M > 0 and diagonal. If Q > 0 and σ > 0 satisfy (7. Then the DNLDI is quadratically ˙ stabilizable if there exists K such that for some diagonal positive-deﬁnite matrix M . .102 Chapter 7 State-Feedback Synthesis for LDIs Using the elimination procedure of §2. An alternate. all the following results can be extended to the case in which Dqp is nonzero.6. and Y such that T AQ + QAT + Bp M Bp T T T T Bp M Dqp + QCq + Y Dqu T +Bu Y + Y T Bu < 0. we obtain the equivalent conditions: AQ + QAT T T T QCq + µBp Dqp − σBu Dqu T T < 0.2. . we can eliminate the variable Y using the elimination procedure of §2. µnq ) > 0 satisfying AQ + QAT + Bu KQ T T Bp M Dqp + Q(Cq + Dqu K) T T + QK T Bu + Bp M Bp < 0.2.

In contrast.3) with u = Kx. Therefore the LMIs described in the previous section also characterize holdable ellipsoids for the system (7. This is easily shown to be equivalent to Q ≥ I. max u(t) t≥0 = ≤ = maxt≥0 Y Q−1 x(t) maxx∈E Y Q−1 x λmax (Q−1/2 Y T Y Q−1/2 ). For example suppose we require (7. Remark: In Chapter 5. results expressed as LMIs in P are not. the constraint u(t) ≤ µ is enforced at all times t ≥ 0 if the LMIs 1 x(0) x(0)T Q ≥ 0.33)) as an optimization problem over LMIs. and in addition x(0)T Q−1 x(0) ≤ 1.9). it is not possible to rewrite these LMIs as LMIs with Q−1 as a variable. however.34)) containing P.13)). This implies that x(t) belongs to E for all t ≥ 0. imposing norm constraints on the input u = Kx. we can ﬁnd an upper bound on the norm of the control input u(t) = Kx(t) as follows.3 Constraints on the control input When the initial condition is known. (7. and consequently. This restricts the extension of some of the results from Chapter 5 (and Chapter 6) to state-feedback synthesis. As another extension we can handle constraints on u(t) max = max |ui (t)| i ∆ This electronic version is for personal use and may not be duplicated or distributed.13)).16) hold. we will not be able to compute the minimum volume holdable ellipsoid that contains a given polytope P (see problem (5. For example. As a rule of thumb. with a few exceptions. such as ﬁnding the coordinate transformation that minimizes the condition number of Q. Q Y YT µ2 I ≥0 (7. where Q > 0 and Y satisfy the stabilizability conditions (either (7.9).12) or (7.12) or (7. we will be able to compute the minimum diameter holdable ellipsoid (see problem (5. etc. We can extend this to the case where x(0) lies in an ellipsoid or polytope.3) if there exists a state-feedback gain K such that E is invariant for the system (7. quadratic stabilizability and ellipsoid holdability can be expressed as LMIs only in variables Q and Y . we saw that quadratic stability and invariant ellipsoids were characterized by LMIs in a variable P > 0 and also its inverse Q = P −1 . (7. Pick Q > 0 and Y which satisfy the quadratic stabilizability condition (either (7. we can solve various optimization problems.7. With this parametrization of quadratic stabilizability and holdable ellipsoids as LMIs in Q > 0 and Y . .2 State Properties 103 is holdable for the system (7.16) to hold for all x(0 ≤ 1. Q Y YT µ2 I ≥ 0. 7. Therefore.2.3). results in the analysis of LDIs that are expressed as LMIs in the variable Q can be extended to state-feedback synthesis.

we have x = Ax + Bu u + Bp p + Bw w. 7.1 Bu. u satisfy (7. the set of states reachable with unit energy is deﬁned as x. . L.17) becomes x = Ax + Bu u + Bp p + Bw w.L ]}.104 in a similar manner: max u(t) t≥0 max Chapter 7 State-Feedback Synthesis for LDIs = = ≤ maxt≥0 Y Q−1 x(t) maxx∈E Y Q−1 x maxi Y Q max −1 max max ii Y T .1 ] .17) becomes x = Ax + Bu u + Bp p + Bw w. . i = 1. w. ˙ (7. . u = Kx.17) with u = Kx. 7.13)). T ≥ 0 0 Copyright c 1994 by the Society for Industrial and Applied Mathematics. Therefore.L Bu. nq . ˙ q = Cq x + Dqu u pi = δi (t)qi . [AL Bw. wT w dt ≤ 1. ˙ q = Cq x + Dqu u • Diagonal Norm-bound LDIs: For DNLDIs. .9).1 Reachable sets with unit-energy inputs For the system (7. ˙ |pi | ≤ |qi |. equation (7.17) We next consider the LDI • LTI systems: For LTI systems.17). . |δi (t)| ≤ 1. pT p ≤ q T q. equation (7. Xii ≤ µ2 . ∆ T Rue = x(T ) .3. equation (7. . • Norm-bound LDIs: For NLDIs. .12) or (7. Q > 0 and Y satisfy the stabilizability conditions (either (7. which we can also express as x = Ax + Bu u + Bp p + Bw w. . ˙ p = ∆(t)q.17) becomes x = Ax + Bw w + Bu u. ˙ • Polytopic LDIs: PLDIs are given by x = A(t)x + Bw (t)w + Bu (t)u.3 Input-to-State Properties x = A(t)x + Bw (t)w + Bu (t)u. where once again. . Equivalently. q = Cq x + Dqu u. i = 1. ≤ µ for t ≥ 0 is implied by the LMI X Y Q T Y ≥ 0. . . ∆(t) ≤ 1 q = Cq x + Dqu u. . where ˙ [A(t) Bw (t) Bu (t)] ∈ Co {[A1 Bw. the constraint u(t) 1 x(0) x(0)T Q ≥ 0. x(0) = 0. . (7.

a corresponding state-feedback gain is given T by K = −(σ/2)Bu Q−1 . ˜ GT AQ + QAT T T +µBp Bp + Bw Bw T Cq Q + µDqp Bp T T QCq + µBp Dqp ˜ G < 0. T The corresponding state-feedback gain is K = (−σ/2)Bu Q−1 .4) for some state-feedback gain K if K satisﬁes T T AQ + QAT + Bu KQ + QK T Bu + Bw Bw ≤ 0. For any Q > 0 and Y satisfying this LMI. .i + Bw. T T Cq Q + µDqp Bp − σDqu Bu T T −µ(I − Dqp Dqp ) − σDqu Dqu This electronic version is for personal use and may not be duplicated or distributed. . Another equivalent condition for E to contain Rue is T ˜ ˜T Bu AQ + QAT + Bw Bw Bu ≤ 0. (7.3 Input-to-State Properties 105 We will now derive conditions under which there exists a state-feedback gain K guaranteeing that a given ellipsoid E = {ξ ∈ Rn ξ T Q−1 ξ ≤ 1 } contains Rue . • Polytopic LDIs: For PLDIs.7.18) For any Q > 0 and σ satisfying this LMI.1.i < 0. ˜ An equivalent condition is expressed in terms of G. L. the ellipsoid E contains Rue for the system (7. Using the elimination procedure of §2. the state-feedback gain K = Y Q−1 is such that the ellipsoid E contains the reachable set for the closed-loop system.2.i Y + Y T Bu. Setting KQ = Y . (7. E contains Rue for some state-feedback gain K if there exist µ > 0 and Y such that the following LMI holds (see LMI (6.6.19) ˜ where Bu is an orthogonal complement of Bu . • Norm-bound LDIs: For NLDIs. . we conclude that E ⊇ Rue for some K if there exist Y such that T T QAT + AQ + Bu Y + Y T Bu + Bw Bw ≤ 0. we eliminate the variable Y to obtain the LMI in Q > 0 and the variable σ: T T QAT + AQ − σBu Bu + Bw Bw ≤ 0. .9)): T T T QAi + Ai Q + Bu.i Bw. E contains Rue for some state-feedback gain K if there exist Q > 0 and Y such that the following LMI holds (see LMI (6. i = 1.1. • LTI systems: From §6. the orthogonal complement T T of [Bu Dqu ]T : We can eliminate Y to obtain the LMI in Q > 0 and σ that guarantees that E contains the reachable set Rue for some state-feedback gain K: T AQ + QAT − σBu Bu T T T QCq + µBp Dqp − σBu Dqu T T (7. T −µ(I − Dqp Dqp ) .11)): T T T QAT + AQ + Bu Y + Y T Bu + Bw Bw + µBp Bp (Cq Q + Dqu Y )T −µI Cq Q + Dqu Y ≤ 0.20) +µBp Bp + Bw Bw < 0.

2 Reachable sets with componentwise unit-energy inputs With u = Kx.i i T Bw. we can eliminate the variable Y to obtain an equivalent LMI in fewer variables. . µ > 0. T Bw 0 −R Copyright c 1994 by the Society for Industrial and Applied Mathematics. • We can ﬁnd a state-feedback gain K such that the set of reachable states for the system (7.3.17). Ruce = x(T ) T wi wi dt ≤ 1. x(0) = 0. E ⊇ Ruce for some state-feedback gain K if there exist Q > 0. E contains Rue for some statefeedback gain K if there exist Q > 0. . and Y such that the following LMI holds (see LMI (6. w. i = 1. L.106 Chapter 7 State-Feedback Synthesis for LDIs • Diagonal Norm-bound LDIs: For DNLDIs. • LTI systems: From LMI (6. In this case. the set of reachable states for inputs with componentwise unit-energy for the system (7. and a diagonal R > 0 with unit trace (see LMI (6. E ⊇ Ruce for the LTI system (7.i Bw.6. Y (= KQ). T ≥0 0 We now consider the existence of the state-feedback gain K such that the ellipsoid E = {ξ ∈ Rn ξ T Q−1 ξ ≤ 1 } contains Ruce for LTI systems and LDIs.i Y + Y T Bu. .) 7. ∆ T . • Polytopic LDIs: For PLDIs.15).19)). (This is due to the coupling induced by the new variable Y = KQ.1.3). Using these LMIs (see §6. such that T T T T QAT + Y T Bu + AQ + Bu Y + µBp Bp QCq + Y T Dqu Bw Cq Q + Dqu Y −µI 0 ≤ 0. such that T QAT + Ai Q + Bu. M > 0 and diagonal.17)).3) lies in a given half-space. E ⊇ Ruce for some state-feedback gain K if there exist Q > 0.17) is deﬁned as x. . u satisfy (7.12)): T T T QAT + AQ + Bu Y + Y T Bu + Bw Bw + Bp M Bp (Cq Q + Dqu Y )T −M Cq Q + Dqu Y ≤ 0. Y and a diagonal R > 0 with unit trace (see LMI (6.4) if there exists diagonal R > 0 with unit trace such that the LMI T QAT + AQ + Bu Y + Y T Bu T Bw Bw −R ≤ 0. This result can be extended to check if the reachable set is contained in a polytope. . we must use the same outer ellipsoidal approximation to check diﬀerent faces. u = Kx. in contrast with the results in Chapter 6.i −R < 0. • Norm-bound LDIs: For NLDIs.1 for details): • We can ﬁnd a state-feedback gain K such that a given point x0 lies outside the set of reachable states for the system (7.2. Using the elimination procedure of §2.

This electronic version is for personal use and may not be duplicated or distributed.. • LTI systems: For LTI systems. Y . it is possible to eliminate the variable Y . w.i Y + Y T Bu..25).5).4 State-to-Output Properties Remark: For LTI systems and NLDIs.4) if there exists α > 0 such that T T AQ + QAT + Bu Y + Y T Bu + Bw Bw /α + αQ ≤ 0.L Dzu.i Bw. .1 Bu. and consider x = A(t)x + Bu (t)u.i + αQ + Bw. .1 Dzu. it is possible to eliminate the variable Y . . the set of states reachable with inputs with unitpeak is deﬁned as Rup = ∆ x(T ) x. E ⊇ Ruce if there exist Q > 0 and Y (see condition (6. • Norm-bound LDIs: From Chapter 6.L . 107 7. α > 0 and µ > 0 such that AQ + QAT + αQ T (Cq Q + Dqu Y ) T T +Bu Y + Y T Bu + µBp Bp + Bw Bw /α ≤ 0. the state equations are x = Ax + Bu u.. AL Cz. • Polytopic LDIs: For PLDIs.28)) such that T T T Ai Q + QAi + Bu.21) • Polytopic LDIs: For PLDIs we have A(t) Cz (t) Bu (t) Dzu (t) ∈ Co A1 Cz. ˙ z = Cz x + Dzu u. T ≥0 x(0) = 0.L Bu.5). E contains Ruce for the LTI system (7.3.7.3 Reachable sets with unit-peak inputs For a ﬁxed state-feedback gain K. ˙ z = Cz (t)x.1 . Cq Q + Dqu Y −µI This condition is an LMI for ﬁxed α. u satisfy (7. for LTI systems and NLDIs. where Y = KQ.4 State-to-Output Properties We consider state-feedback design to achieve certain desirable state-to-output properties for the LDI (4. (7. E contains Ruce if there exist Q > 0. Remark: Again. .i /α ≤ 0 for i = 1. T u = Kx. We set the exogenous input w to zero in (4.17).. L. . 7. • LTI systems: From condition (6. w(t) w(t) ≤ 1. .

q = Cq x + Dqu u.39) that for a given state-feedback gain K. which can be rewritten as x = Ax + Bu u + Bp p.1) of the closed-loop system is less than some speciﬁed value.4.i Cz. the output energy is by xT Q−1 x0 .i Y −I for some Y . for any Q > 0.1 Bounds on output energy We ﬁrst show how to ﬁnd a state-feedback gain K such that the output energy (as deﬁned in §6. . L. 0 (Cz Q + Dzu Y )T −I ≤ 0. 7. pi = δi (t)qi . ˙ q = Cq x + Dqu u. i = 1. • Diagonal Norm-bound LDIs: Finally.i Y )T T T +Bu. |δi (t)| ≤ 1. p = ∆(t)q. the output energy of system (7. nq . L • Norm-bound LDIs: In the case of NLDIs.108 Chapter 7 State-Feedback Synthesis for LDIs • Norm-bound LDIs: NLDIs are given by x = Ax + Bu u + Bp p. where Q > 0 and 0 Y = KQ satisfy T AQ + QAT + Bu Y + Y T Bu Cz Q + Dzu Y Regarding Y as a variable. (7.i Q + Dzu. . . . |pi | ≤ |qi |. q = Cq x + Dqu u. • Polytopic LDIs: In this case. ˙ z = Cz x + Dzu u.i Y + Y Bu.23) Copyright c 1994 by the Society for Industrial and Applied Mathematics. • LTI systems: We conclude from LMI (6. . i = 1. . We assume ﬁrst that the initial condition x0 is given. ˙ z = Cz x + Dzu u. the output energy where Q satisﬁes the LMI Ai Q + QAT i (Cz.22). . (7. ˙ q = Cq x + Dqu u which can be rewritten as x = Ax + Bu u + Bp p. is bounded above by x T Q−1 x0 .22) ≤ 0. DNLDIs are given by x = Ax + Bu u + Bp p. . we can then ﬁnd a state-feedback gain that guarantees an output energy less than γ by solving the LMIP xT Q−1 x0 ≤ γ and (7. Y and µ ≥ 0 such that 0 AQ + QAT + Bu Y (Cz Q + Dzu Y )T (Cq Q + Dqu Y )T T T +Y T Bu + µBp Bp Cz Q + Dzu Y −I 0 Cq Q + Dqu Y 0 −µI bounded above ≤ 0.21) does not exceed xT Q−1 x0 .i Q + Dzu.2. . see the Notes and References.22) is closely related to the classical Linear-Quadratic Regulator (LQR) problem. i = 1. . . inequality (7. . z = Cz x + Dzu u p p ≤ q T q. . 0 Of course. T z = Cz x + Dzu u ∆(t) ≤ 1.

Y and M ≥ 0 and diagonal such that 0 AQ + QAT + Bu Y T T (Cz Q + Dzu Y ) (Cq Q + Dqu Y ) T T +Y T Bu + Bp M Bp ≤ 0.7. in the case of DNLDIs. it is possible to simplify the inequality (7. We can extend these results to the case when x0 is speciﬁed to lie in a polytope or an ellipsoid (see §6.5 Input-to-Output Properties Remark: As before.1).1).25) In the case of LTI systems.26) This electronic version is for personal use and may not be duplicated or distributed.3. the list of problems considered here is far from exhaustive. the L2 gain for LTI systems is equal to the H∞ norm of the corresponding transfer matrix. EVPs that yield 0 state-feedback gains that minimize the expected value of the output energy can be derived. Cz Q + Dzu Y −I 0 0 −M Cq Q + Dqu Y 7.1 L2 and RMS gains We seek a state-feedback gain K such that the L2 gain sup w 2 =1 z 2 = sup w 2 =0 z w 2 2 of the closed-loop system is less than a speciﬁed number γ.2.24) +Bw Bw ≤ 0. (Cz + Dzu K)Q −γ 2 I Introducing Y = KQ.25) and get the equivalent Riccati inequality T T T T AQ + QAT − Bu (Dzu Dzu )−1 Bu + Bw Bw + QCz Cz Q/γ 2 ≤ 0. the output energy is bounded above by xT Q−1 x0 . we can eliminate the variable Y from this LMI. for any Q > 0. (7. If x0 is a random variable with E x0 xT = X0 . • LTI systems: As seen in 6.5 Input-to-Output Properties We ﬁnally consider the problem of ﬁnding a state-feedback gain K so as to achieve desired properties between the exogenous input w and the output z for the system (7.2. ﬁnding a state-feedback gain K so as to minimize the upper bound on the extractable energy for the various LDIs is therefore an EVP. this condition is necessary and suﬃcient. (A + Bu K)Q + Q(A + Bu K)T T Q(Cz + Dzu K) T (7. . As mentioned in Chapter 6.2. if there exist K and Q > 0 such that.5. Assuming T T Cz Dzu = 0 and Dzu Dzu invertible. • Diagonal Norm-bound LDIs: Finally. 7. this can be rewritten as T T AQ + QAT + Bu Y + Y T Bu + Bw Bw (Cz Q + Dzu Y )T −γ 2 I Cz Q + Dzu Y ≤ 0. From §6.3. there exists a state-feedback gain K such that the L2 gain of an LTI system is less than γ. 109 Given an initial condition. (7.

• LTI systems: Substituting Y = KQ. +Bu. yields a closedloop system with H∞ -norm less than γ. there exists a state-feedback gain such that the L2 gain of a PLDI is less than γ if there exist Y and Q > 0 such that Ai Q + QAT + Bu.i − Cz. and using LMI (6. (7.i ) Copyright c 1994 by the Society for Industrial and Applied Mathematics. • Polytopic LDIs: From (6. Q > 0 and Y holds: T AQ + QAT + Bu Y + Y T Bu T Bw T T Bw − QCz − Y T Dzu T 2ηI − (Dzw + Dzw ) − Cz Q − Dzu Y ≤ 0. The T T resulting controller. provable with quadratic Lyapunov functions.5.i + Dzw.2) is passive.i Q + Dzu.60).i Bw.28) 2 Cz Q + Dzu Y −γ I 0 Cq Q + Dqu Y 0 −µI We can therefore ﬁnd state-feedback gains that minimize the upper bound on the L2 gain.62).3.i Q − Dzu.i + Bw.i − QCz. we conclude that the dissipation of the LTI system is at least η if the LMI in η. for the various LDIs by solving EVPs.e.110 Chapter 7 State-Feedback Synthesis for LDIs The corresponding Riccati equation is readily solved via Hamiltonian matrices.. • Polytopic LDIs: From §6.i Y + Y T Bu. the LMI that guarantees that the L2 gain is less than γ for some state-feedback gain K is AQ + QAT T +Bu Y + Y T Bu (Cz Q + Dzu Y )T (Cq Q + Dqu Y )T T T +Bw Bw + µBp Bp ≤ 0.i T ≤ 0.i Y i T (Cz.i ≤ 0.i T Bw. assuming Dzw (t) is nonzero and square. η satisfying T 0 wT z − ηwT w dt ≥ 0 for all T ≥ 0.i Q + Dzu.i Y ) T T (7.i Y −γ 2 I • Norm-bound LDIs: For NLDIs. . with state-feedback gain K = −(Dzu Dzu )−1 Bu . Cz. more generally. 7.2 Dissipativity We next seek a state-feedback gain K such that the closed-loop system (7. we wish to maximize the dissipativity. there exists a state-feedback gain such that the dissipation exceeds η if there exist Q > 0 and Y such that Ai Q + QAT i T Bw. i.27) +Y T Bu.2.i − Y T Dzu.i Y T 2ηI − (Dzw.

¯ x ¯ u = K x. AM } ∂x where A1 . . however. ˙ n We consider the nonlinear system y = Cy x. .29) allows us to meet more speciﬁcations than can be met using a static state-feedback.6 Observer-Based Controllers for Nonlinear Systems 111 We can therefore ﬁnd state-feedback gains that maximize the lower bound on the dissipativity. We could work out the more complicated LMIs that characterize quadratic stabilizability with dynamic state-feedback. which is the same as (4. u : R+ → R is the control variable. p (7. It might seem that the dynamic state-feedback controller (7. By ∈ Rr×n . AL are given. this is not the case.6 Observer-Based Controllers for Nonlinear Systems x = f (x) + Bu u.3. Since the new LMIs thus obtained are straightforward to derive. . we can incorporate scaling techniques into many of the results above to derive componentwise results.7. .31) This electronic version is for personal use and may not be duplicated or distributed. . . Cu ∈ Rnu ×r . for the various LDIs by solving EVPs.30) where x : R+ → R is the state variable. 7. We would ﬁnd. For speciﬁcations based on quadratic Lyapunov functions. The number r is called the order of the controller.3 Dynamic versus static state-feedback controllers A dynamic state-feedback controller has the form ¯x ¯ ˙ x = A¯ + By x.5. ¯ ¯ ¯ ¯ u = Cu x + Duy x.4.29) ¯ ¯ ¯ ¯ where A ∈ Rr×r . (7. that these more general LMIs are also infeasible. We look for a stabilizing observer-based controller of the form ˙ x = f (¯) + Bu u + L(Cy x − y). this dynamic state-feedback controller reduces to the state-feedback we have considered so far (which is called static state-feedback in this context). . we will omit them here. suppose there does not exist a static state-feedback gain that yields closed-loop quadratic stability. and y : R+ → Rq is the measured or sensed variable. For example.15). We assume the function f : Rn → Rn satisﬁes f (0) = 0 and ∂f ∈ Co {A1 . . Duy ∈ Rnu ×n . See the Notes and References. . • Norm-bound LDIs: From (6. Note that by taking r = 0. Remark: As with §6. however. In this case we might turn to the more general dynamic statefeedback. provable with quadratic Lyapunov functions. there exists a state-feedback gain such that the dissipation exceeds η if there exist Q > 0 and Y such that T AQ + QAT + Bu Y Bw − QCz T (Cq Q + Dqu Y ) T T T +Y T Bu + µBq Bq −Y T Dzu T ≤0 Bw − C z Q T −(Dzw + Dzw − 2ηI) 0 −Dzu Y Cq Q + Dqu Y 0 −µI 7. ¯ (7.64).

30). [Bar70a. It is clearly related to the theory of optimal control. M (7. M (7. These conditions are that some P > 0 and Q > 0 satisfy T T Ai Q + QAi < σBu Bu . . where Bu and Cy are orthogonal complements of Bu and T Cy .g. i and T ˜T ˜ Cy Ai P + P Ai Cy < 0. which is ˜ true if there exists a positive-deﬁnite matrix P ∈ R2n×2n such that for any nonzero trajectory x. in Copyright c 1994 by the Society for Industrial and Applied Mathematics. .32) is stable. then they satisfy the LMIs (7.31). Y .34) hold. M i = 1. M (7. see e.112 Chapter 7 State-Feedback Synthesis for LDIs i. we search for the matrices K (the estimated-state feedback gain) and L (the observer gain) such that the closed-loop system x ˙ ˙ x ¯ = f (x) + Bu K x ¯ −LCy x + f (¯) + (Bu K + LCy )¯ x x (7.. To every P . . . .36) T and (7. . . there corresponds a stabilizing observer-based controller of the form (7. we can obtain a stabilizing observer-based T controller of the form (7. . i i = 1. The observer-based controller with K = −(σ/2)B u Q−1 −1 and L = −(µ/2)P Cy stabilizes the nonlinear system (7. by setting K = −(1/2)Bu Q−1 and L = −(1/2)P −1 Cy . and W such that the LMIs Q > 0. .35) and T AT P + P Ai < µCy Cy . . M (7. If P and Q satisfy these LMIs. Y . i = 1.33) i = 1. . . . .32) is stable if it is quadratically stable. . obtained by setting K = Y Q−1 and L = P −1 W . and P > 0. The closed-loop system (7. Q. respectively. Q > 0. . i T Ai Q + QAT + Bu Y + Y T Bu < 0. i i = 1. we can obtain equivalent conditions in which the variables Y and W do not appear. By homogeneity we can freely set σ = µ = 1. we prove that this is true if there exist P . Notes and References Lyapunov functions and state-feedback The extension of Lyapunov’s methods to the state-feedback synthesis problem has a long history. .. Another equivalent condition is that ˜T ˜ Bu Ai Q + QAT Bu < 0. Q. . .35) for some σ and µ. x. T AT P + P Ai + W Cy + Cy W T < 0. .2. i = 1. . . . Q > 0 satisfying these LMIs with σ = µ = 1. In the Notes and References. and W satisfying these LMIs. Using the elimination procedure of §2. Lef65].36) for some σ and µ. For any P > 0.31). .e. M ˜ ˜T hold for some P > 0. . we have: ¯ d dt x x ¯ T ˜ P x x ¯ < 0.6.

See also [GPB91. GP82. One example is the “Lyapunov Min-Max Controller” described in [Gut79. is due to Bernussou. Rya88. The change of variables Q = P −1 and Y = KP −1 .. KKR93. see [Cor94]. EBFB92. McFarlane and Rotea [PM92. RK91. PMR93]. shows that there exist LDIs that are quadratically stabilizable. adding performance indices allows one to recover many classical results of automatic control. RK93. Quadratic stabilizability The term “quadratic stabilizability” seems to have been coined by Hollot and Barmish in [HB80]. Gu92a. PG93] consider the extension to PLDIs. Sko91b. Arzelier et al. is used in the paper by Thorp and Barmish [TB81]. We suspect that the methods described in this chapter can be interpreted as a search for suboptimal controls in which the candidate value functions are restricted to a speciﬁc class. . PGB93]. For a recent and broad review about quadratic stabilization of uncertain systems. In the general optimal control problem. see also [HB80. as well as in Stoorvogel [Sto91]. a convex problem involving LMIs.Notes and References 113 which the natural Lyapunov function is the Bellman–Pontryagin “min-cost-to-go” or value function (see [Pon61. Wei94]. The L2 gain of PLDIs has been studied by Obradovic and This electronic version is for personal use and may not be duplicated or distributed. i. For LTI systems his criteria reduce to the LMIs in this chapter. Petersen [Pet87b] derives a Riccati-based approach for stabilizing NLDIs. though not via linear state-feedback. The problem of ﬁnding a state-feedback gain to minimize the scaled L2 gain for LTI systems is discussed in [EBFB92].e. Petersen [Pet85]. BPG89a. see also [Lei79. when Letov [Let61] studied the performance of unknown. The problem of approximating reachable sets for LTI systems under various assumptions on the energy of the exogenous input has been considered for example by Skelton [SZ92] and references therein. Cor85]. The counterpart for NLDIs is in Petersen. Bar83. PZP+ 92. Bar85]. [ABG93] consider the problem of robust stabilization with pole assignment for PLDIs via cutting-plane techniques.e. Hollot [Hol87] describes conditions under which a quadratically stabilizable LDI has inﬁnite stabilizability margin. The idea that the output variance for white noise inputs (and related LQR-like performance measures) can be minimized using LMIs can be found in [BH88a.234] for details). Peres. In [Son83]. BPG89b. GR88]. Wei [Wei90] derives necessary and suﬃcient “signpattern” conditions on the perturbations for the existence of a linear feedback for quadratic stability for a class of single-input uncertain linear dynamical systems. PBG89. See also [Zin90. Several approaches use Lyapunov-like functions to synthesize nonlinear state-feedback control laws. though not explicitly described. i. p.. Quadratic Lyapunov functions for proving performance The addition of performance considerations in the analysis of LDIs can be traced back to 1955 and even earlier. At that time. Sontag formulates a general theorem giving necessary and suﬃcient Lyapunov type conditions for stabilizability. nonlinear and possibly time-varying control systems and calls it “the problem of control quality”. KR91. GPB91. PB84. BH89. See also [OC87].28) (this is shown in [Pet87a]). the performance criteria were decay rate and output peak deviations for systems subject to bounded peak inputs (see [Let61. quadratic. FBBE92. see also [SP94e]. The problem of maximizing decay rate is discussed in great detail in [Yan92]. see also the Notes and References of the previous chapter. other references on decay rates and quadratic stability margins are [Gu92b. In the case of LTI systems. the (exact) value function satisﬁes a partial diﬀerential equation and is hard to compute. by restricting our search to quadratic approximate value functions we end up with a low complexity task. where the authors give necessary and suﬃcient conditions for it. Gut79. Gu92b. see also [SRC93]. CL90. BCL83. Cor90]. Souza and Geromel [PSG92. BB91. Peres and Geromel [BPG89a]. Cor85. which enables the extension of many of the results of Chapters 5 and 6 to state-feedback synthesis. SC91. which is in fact the state-feedback H∞ equation (7. Pon62]). SI93]. This change of variables.

For LTI systems. we ﬁnd: We shall investigate whether diﬀerential inclusions x ∈ F (x(t)). We note that the Riccati inequality (7. it appears verbatim in some articles on the theory of nonzerosum diﬀerential games. z = Cz x + Dzu u. and DeSouza et al. B. Zhu and Skelton [GCZS93]. who point out explicitly that quadratic stability of NLDIs with an L2 gain bound is equivalent to the scaled H∞ condition (6. In [PP92]. ˙ have trajectories satisfying the property t x(0) = x0 ∀t > s. the quadratic matrix inequality found there (expressed in our notation) is T T AT P + P A + γ 2 P B w Bw P + C z Cz T Bu P T P Bu + Cz Dzu T Dzu Dzu + T Dzu Cz ≥ 0. §6]. Sznaier and Ly [SSL93].26) is also encountered in full-state feedback H∞ control [Pet87a]. T T where K = −(Dzu Dzu )−1 Bu Pare and Pare is the unique positive-deﬁnite matrix that satisﬁes the algebraic Riccati equation T T T AT Pare + Pare A − Pare Bu (Dzu Dzu )−1 Bu Pare + Cz Cz = 0.3. see for example Starr and Ho [SH68]. Finally.37) is often called a “dissipation inequality” [Wil71b]. by Peres. C) is minimal.39) Copyright c 1994 by the Society for Industrial and Applied Mathematics. (7. WXdS92. LMI formulation of LQR problem For the LTI system x = Ax + Bu u. Dzu Dzu is invertible T and Dzu Cz = 0. [XFdS92. In particular.3). It turns out that the optimal input u can be expressed as a constant state-feedback u = Kx. Yak71] and Mageirou [Mag76]. . With −1 Qare = Pare . V (x(t)) − V (x(s)) + s W (x(τ ). XdS92.54). which does not possess any obvious convexity property. the authors give analytic conditions for making an LTI system passive via state-feedback. o Carpenter. Yakubovich [Yak70. In the book by Aubin and Cellina [AC84. Inequality (7.2. Kokotovic [Kok92] considers stabilization of nonlinear systems. ˙ (7.25) which provides a necessary and suﬃcient condition for a system to have L2 gain less than γ with the corresponding matrix inequality found in the article by Stoorvogel and Trentelman [ST90]. We assume for simplicity that (A.37) We shall then say that a function V [· · ·] satisfying this condition is a Lyapunov function for F with respect to W . Its occurrence in control theory is much older. we can rewrite (7. (7. it is interesting to compare the inequality (7. Geromel and Souza [PGS91] and also by Ohara.114 Chapter 7 State-Feedback Synthesis for LDIs Valavani [OV92]. Similar ideas can be found in [CP72]. we mention an instance of a Lyapunov function with a built-in performance criterion. and provides a motivation for rendering an LTI system passive via state-feedback (see §6. see also [SKS93]. NLDIs have been studied in this context by Petersen [Pet89]. dSFX93]. Masubuchi and Suda [OMS93]. x(τ )) dτ ≤ 0.38) as T T T AQare + Qare AT − Bu (Dzu Dzu )−1 Bu + Qare Cz Cz Qare = 0. The problem of minimizing the control eﬀort given a performance bound on the closed-loop system can be found in the articles of Sch¨mig.38) The optimal output energy for an initial condition x(0) is then given by x(0)T Pare x(0). We shall encounter such Lyapunov functions in §8. Gu93]. and Grigoriadis. the Linear-Quadratic Regulator (LQR) ˙ problem is: Given an initial condition x(0). see also [ZKSN92. ﬁnd the control input u that minimizes the output ∞ T energy 0 z T z dt.

(7. [Kai80]). we have x(0)T Q−1 x(0) ≤ x(0)T Q−1 x(0).40) Of course.40) holds for some Q > 0 and Y if and only if it holds for Y = −(Dzu Dzu )−1 Bu . (7.42). (7.Notes and References and the optimal output energy as x(0)T Q−1 x(0). We can derive this analytic solution for the EVP via the following steps. the minimum output energy was given by minimizing x(0)T Q−1 x(0) subject to Q > 0 and T AQ + QAT + Bu Y + Y T Bu (Cz Q + Dzu Y )T −I Cz Q + Dzu Y ≤ 0. in this case. the problem of maximizing x(0)T P x(0) subject P > 0 and the constraint T T T AT P + P A − P Bu (Dzu Dzu )−1 Bu P + Cz Cz ≥ 0. Consider the LTI system x = Ax + Bu u + Bw w. given an initial condition.4. the upper bound equals the output energy. using a simple completion-of-squares argument T T that LMI (7. This electronic version is for personal use and may not be duplicated or distributed. the Riccati equation can be thus be interpreted as yielding an analytic solution to the EVP. then Q ≤ Q are . and we showed that in this case. it can be shown. Therefore for every initial condition x(0). the optimal value of this EVP must equal the optimal output energy given via the solution to the Riccati solution (7. and the optimal state-feedback are T T gain is Kopt = Yopt Q−1 = −(Dzu Dzu )−1 Bu Pare . However.5. which is nothing other than (7. but with the inequality reversed. . the EVP is equivalent to minimizing x(0)T P x(0) subject to T T AT P + P A + Cz Cz − P Bu Dzu Dzu −1 T Bu P ≤ 0.43) From §7. it is wellknown (see for example [Wil71b]) that this EVP is another formulation of the LQR problem. are It is interesting to note that with P = Q−1 . it can be shown by standard manipulations that if Q > 0 satisﬁes (7. the exists a static state-feedback u = Kx such that the L2 gain from w to z does not exceed γ if the LMI in variables Q > 0 and Y is feasible: T T AQ + QAT + Bu Y + Y T Bu + Bw Bw T QCz Cz Q −γ 2 I ≤ 0. First.42) which is not a convex constraint in P . we considered the closely related problem of ﬁnding a state-feedback gain K that minimizes an upper bound on the energy of the output.41). and are therefore the optimal value of the EVP is just x(0)T Q−1 x(0).1. We will demonstrate this fact on one simple problem considered in this chapter. For LTI systems. In fact. are 115 In section §7. is an EVP.39). provided the properties are speciﬁed using quadratic Lyapunov functions.41) Next.1. similar statements can be made for all the properties and for all the LDIs considered in this chapter. z = Cz x. ˙ x(0) = 0. (7. Q > 0 must satisfy T T QAT + AQ + QCz Cz Q − Bu Dzu Dzu −1 T Bu ≤ 0. Static versus dynamic state-feedback An LTI system can be stabilized using dynamic state-feedback if and only if it can be stabilized using static state-feedback (see for example.

2.) Eliminating Y . With state vector xcl = [xT closed-loop system is xcl = Acl xcl + Bcl u. This will establish that the smallest upper bound on the L2 gain. we get the equivalent LMI in Q > 0 and scalar σ: T AQ + QAT + Bw Bw + T QCz Cz Q T ≤ σBu Bu . Cc and Dc such that the L2 gain of the system (7. Kbig = . KR88. Bc . Bbig = . A similar result is established [PZP+ 92].44) holds. T AT P + Pbig A − τ I + Ccl Ccl big T Bcl P P Bcl −γ 2 I ≤ 0. other references that discuss the question of when dynamic or nonlinear feedback does better than static feedback include Petersen. Bcl = Ccl = [Cz 0] .45) We will show that there exist matrices Ac . is the same irrespective of whether a static state-feedback or a dynamic state-feedback is employed. γ2 (7. P Bcl −γ 2 I ≤ 0. Copyright c 1994 by the Society for Industrial and Applied Mathematics. z = Ccl xcl .43. provable via quadratic Lyapunov functions. Eliminating Kbig from this LMI yields two equivalent LMIs in Pbig > 0 and a scalar τ . RK88. but this is irrelevant to the current discussion. From §6. u = Cc xc + Dc x. T T AT P + Pbig A − τ P Bbig Bbig P + Ccl Ccl big T Bcl P It is easy to verify that the ﬁrst LMI holds for large enough τ .44) We next consider a dynamic state-feedback for the system: xc = Ac xc + Bc x. the c where Acl = Abig + Bbig Kbig with A 0 Bw 0 0 0 . ˙ xc (0) = 0.3. ˙ xcl (0) = 0. We will show only the part that is not obvious. Pet88. The last LMI is precisely (7.116 Chapter 7 State-Feedback Synthesis for LDIs (In fact this condition is also necessary. 0 I Bu 0 Bc Dc Ac Cc Abig = . (7. the L2 gain of the closed-loop system does not exceed γ if the LMI in Pbig > 0 is feasible: T T T AT P + Pbig A + P Bbig Kbig + Kbig Bbig P + Ccl Ccl big T Bcl P P Bcl −γ 2 I ≤ 0.7. .44) with the change 2 2 of variables Q = γ Q11 and σ = γ τ . KPR88.45) does not exceed γ if and only if the LMI (7. Khargonekar and Rotea [Pet85. γ2 −1 where Q11 is the leading n × n block of Pbig . while the second reduces to τ > 0 and AQ11 + Q11 AT + T Bw Bw T T + Q11 Cz Cz Q11 ≤ τ Bu Bu . RK89]. xT ]T .

a special rank minimization problem can be shown to be equivalent to the NP-hard zero-one linear programming problem (see e. (7. however. j = 1.33) and (7. Geromel and de Souza [IS93b. we can write ˙ x ¯ ˙ x−x ¯ ˙ with ˜ Aij = Ai + B u K 0 LCy Aj + LCy . i.g. ˜ ∈ Co Aij 1 ≤ i ≤ M. Several researchers have considered such problems: Steinberg and Corless [SC85]. local algorithms for such problems and report practical success. . j = 1. [Dav94]). ˜ P = P11 T P12 P12 P22 . We start by rewriting the system (7. that is.46) is quadratically stable.3.47) has a solution P . and Geromel.32) in the coordinates x. (7. Skelton. Galimidi and Barmish [GB86]. SIG93].47) has a solution if and only if the LMIs (7. It is shown in [PZPB91.33) and (7.47) implies (Ai + Bu K)T P11 + P11 (Ai + Bu K) < 0.34) do. GdSS93.30) quadratically stable. Observer-based controllers for nonlinear systems We prove that the LMIs (7. in fact. i. ISG93.47) We will now show that the LMI (7. .34) (in the variables P > 0.32) is quadratically stable if and only if the system (7. . The general problem of minimizing the rank of a matrix subject to an LMI constraint has high complexity. EG93.46) Using the results of §4. .31) which makes the system (7. i = 1. M. With P partitioned as n × n blocks. x − x: ¯ ¯ ˙ x ¯ ˙ x−x ¯ ˙ = LCy (¯ − x) + f (¯) + Bu K x x x ¯ f (¯) − f (x) + LCy (¯ − x) x x . . Pac94] that this problem can be reduced to the problem of minimizing the rank of a matrix P ≥ 0 subject to certain LMIs. de Souza and Skelton [PGS93]. . Q > 0. . . Therefore. . see also [HCM93]. Peres. (7. that output-feedback synthesis problems have high complexity and are therefore unlikely to be recast as LMI problems. . M. W and Y ) ensure the existence of an observer-based controller of the form (7. ˜ ˜ First suppose that (7. We suspect. . Nevertheless several researchers have tried heuristic. . ﬁnding dynamic output-feedback controllers with the smallest possible number of states.48) This electronic version is for personal use and may not be duplicated or distributed. we easily check that the inequality (7. Others studying this topic include Iwasaki. the system (7.Notes and References Dynamic output-feedback for LDIs 117 A natural extension to the results of this chapter is the synthesis of output-feedback to meet various performance speciﬁcations. M.32) is quadratically stable ˜ if there exists P > 0 such that ˜ij ˜ ˜˜ AT P + P Aij < 0. 1≤j≤M x ¯ x−x ¯ Simple state-space manipulations show the system (7. One variation of this problem is the “reduced-order controller problem”. .

. i = 1. Gain-scheduled or parameter-dependent controllers Several researchers have extended the results on state-feedback synthesis to handle the case in which the controller parameters can depend on system parameters. . j = 1. . .48) is equivalent to (7.118 Chapter 7 State-Feedback Synthesis for LDIs −1 With the change of variables Q = P11 and Y = Bu Q. such a λ always exists. .47) and therefore proves the closed-loop system (7. . BPPB93. suppose that P . M . IS93a. . j = 1. ˜ Denoting by Q22 the lower-right n × n block of Q.49) is equivalent to (7. (7.33). we see that this inequality implies (Ai + LCy )Q22 + Q22 (Ai + LCy )T < 0. and νi = λmax ((Ai + Bu K)T Q−1 + Q−1 (Ai + Bu K)). Problems of observer-based controller design along with quadratic Lyapunov functions that prove stability have been considered by Bernstein and Haddad [BH93] and Yaz [Yaz93]. . i. 22 Conversely. M.34) and (7.34). M . . .33) and deﬁne L = P −1 W . BP91] consider the problem in the context of NLDIs. . Lu. for i.33) are satisﬁed. PBPB93. The inequality (7. IS93b. Therefore. . the inequality (7. j = 1. . Pac94. AGB94]. . . . i = 1. . j = 1. . . We now prove there exists a positive λ such that ˜ P = λQ−1 0 0 P satisﬁes (7. . We compute ˜˜ ˜T ˜ Aij P + P Aij = λ (Ai + Bu K)T Q−1 + Q−1 (Ai + Bu K) λ(LCy )T Q−1 λQ −1 LCy (Aj + LCy )T P + P (Aj + LCy ) for i. using Schur complements. This condition is satisﬁed for any λ > 0 such that λ min µj > max νi . the closed-loop system (7. . Q.32) is quadratically stable if λ > 0 satisﬁes λ(Q−1 LCy ((Aj + LCy )T P + P (Aj + LCy ))−1 (LCy )T Q−1 ) − (A + Bu K)T Q−1 − Q−1 (A + Bu K) > 0 . Copyright c 1994 by the Society for Industrial and Applied Mathematics. . the inequality (7. We refer the reader to the articles cited for precise explanations of what these control laws are. . K = Y Q−1 . . 1≤j≤M 1≤i≤M where µj = λmin (Q−1 LCy ((Aj + LCy )T P + P (Aj + LCy ))−1 (LCy )T Q−1 ).32) is quadratically stable. Other relevant references include [PB92. ˜ ˜ Deﬁne Q = P −1 . GA94. M. . and which design problems can be recast as LMI problems.49) With P = Q−1 and W = P L.47) becomes ˜ ˜ ˜ ˜ij Aij Q + QAT < 0. M Since (7. Y and W satisfy the LMIs (7.34) and (7. M. Such controllers are called gain-scheduled or parameter-dependent. Zhou and Doyle in [LZD91] and Becker and Packard [Bec93.

2) occurs when the functions φ i are linear.Chapter 8 Lur’e and Multiplier Methods 8. where αi and βi are given.e. which is referred to as a system with unknown. z = Cz x. The results that follow will hold for any nonlinearities φi satisfying the sector conditions. 1]. equivalently. . time-varying parameters.1). i. Such systems are readily transformed to the form given in (8. which implies that V (ξ) ≥ ξ T P ξ > 0 for nonzero ξ. . Bw . φi (σ)(φi (σ) − σ) ≤ 0 for all We consider the Lur’e system (8. .q ξ λi 0 φi (σ) dσ. 1]. .2) by a loop transformation (see the Notes and References). np .1). An important special case of (8. np .1) and (8. . this is referred to as a system with unknown-but-constant parameters.1 Analysis of Lur’e Systems ˙ x = Ax + Bp p + Bw w. In cases where the φi are known. where p(t) ∈ Rnp .3) where Ci. pi (t) = φi (qi (t)).. Bp . (8. Note that the Lyapunov function (8. Cq and Cz . where δi ∈ [0. the results can be sharpened. . q = Cq x. φi (σ) = δi σ. It is important to distinguish this case from the PLDI obtained with (8.1) The data in this problem are the matrices A.q denotes the ith row of Cq . φi (σ) = δi (t)σ and δi (t) ∈ [0. . . Our analysis will be based on Lyapunov functions of the form np V (ξ) = ξ P ξ + 2 i=1 T Ci. Thus the data describing the Lyapunov function are the matrix P and the scalars λi .2) or. i = 1.3) depends on the (not fully speciﬁed) nonlinearities φi . It is possible to handle the more general sector conditions αi σ 2 ≤ σφi (σ) ≤ βi σ 2 for all σ ∈ R. Thus the data P and λi should be thought of as providing a recipe for 119 . (8. and the functions φi satisfy the [0. We require P > 0 and λi ≥ 0. 1] sector conditions 0 ≤ σφi (σ) ≤ σ 2 for all σ ∈ R. In the terminology of control theory. i = 1. however. σ ∈ R. (8.

The Lyapunov function will then have the form T V (ξ) = ξ T P + Cq ∆ΛCq ξ where ∆ = diag(δ1 . λnp ). The S-procedure then yields the following suﬃcient condition for (8.6) is only a suﬃcient condition for the existence of a Lur’e Lyapunov function that proves stability of system (8. i.1) and (8. Perform a loop transformation on the Lur’e system to bring the nonlinearities to sector [−1. Then the LMI above is the same (to within a scaling of the variable T ) as one that yields quadratic stability of the associated DNLDI. (8. .120 Chapter 8 Lur’e and Multiplier Methods constructing a speciﬁc Lyapunov function given the nonlinearities φi . Copyright c 1994 by the Society for Industrial and Applied Mathematics.2). . . .e. As an example. .5) } = {(ξ. . (8. .1. . we are really synthesizing a parameter-dependent quadratic Lyapunov function for our parameterdependent system.q ξ) ≤ 0.q dt i=1 condition (8. . and seek P and λi such that dV (x) < 0 for all nonzero x satisfying (8. . which can be interpreted as a condition for the existence of a quadratic Lyapunov function for the Lur’e system. Remark: When we set Λ = 0 we obtain the LMI AT P + P A T Bp P + T C q T P B p + Cq T −2T < 0. It is easily shown that {(ξ. .q (Aξ + Bp π) < 0 for all nonzero ξ satisfying πi (πi − ci. .1). τnp ) ≥ 0 AT P + P A T Bp P + ΛCq A + T Cq T T P B p + A T Cq Λ + C q T T T ΛCq Bp + Bp Cq Λ − 2T i = 1. i. φi (σ) = δi σ. . dt Since dV (x) = 2 xT P + λi pi Ci.5) } . consider the special case of a system with unknown-but-constant parameters. π) | ξ = 0. In other words. π) | ξ = 0 or π = 0. .5) <0 (8..1 Stability We set the exogenous input w to zero. δnp ) and Λ = diag(λ1 .4) (Ax + Bp p) . 1]. ξ P+ i=1 T λi πi Ci. (8. λnp ) ≥ 0 and T = diag(τ1 .4) holds if and only if np np (8. .1. .4): The LMI in P > 0. . np .e.6) holds. It is also necessary when there is only one nonlinearity. when np = 1. as follows. Remark: Condition (8. . see §5. Λ = diag(λ1 . 8. . . ..

3). Of course.7) From the S-procedure.q T (Aξ + Bp π) + ξ T Cz Cz ξ ≤ 0.8) Bp P + ΛCq A + T Cq ΛCq Bp + Bp Cq Λ − 2T ΛCq Bw ≤ 0 T T T Bw P B w Cq Λ −I holds.3). If the nonlinearities φi are known.q ξ λi 0 φi (σ) dσ ≤ 1 contains Rue if d V (x) ≤ w T w dt for all x and w satisfying (8.2). for every ξ satisfying πi (πi − ci.1) with initial condition x(0). The condition (8. i = 1. so that E ⊇ F ⊇ Rue . . np . w satisfy (8.7) holds if there exists T = diag(τ1 . .9) is equivalent to np 2 ξT P + i=1 λi πi Ci. (8. We can use these results to prove that a point x0 does not belong to the reachable set by solving appropriate LMIPs. τnp ) ≥ 0 such that T T AT P + P A P B p + A T Cq Λ + C q T P Bw T T T (8. (8.1) and (8.2 Reachable sets with unit-energy inputs We consider the set reachable with unit-energy inputs. . T ≥0 0 .1). x.1. . If d V (x) + z T z ≤ 0 dt for all x satisfying (8. and compute upper bounds on the output energy ∞ J= 0 z T z dt using Lyapunov functions V of the form (8. the ellipsoid E = {ξ ∈ Rn ξ T P ξ ≤ 1 } contains F.q ξ) ≤ 0. λnp ) ≥ 0. Therefore. . . the condition (8. . The set np F= ξ ξ Pξ + 2 i=1 T Ci. E gives an outer approximation of Rue if the nonlinearities are not known. F gives a possibly better outer approximation. . where Λ = diag(λ1 . .1 Analysis of Lur’e Systems 121 8.3 Output energy bounds We consider the system (8. This electronic version is for personal use and may not be duplicated or distributed.1. .9) then J ≤ V (x(0)). .8. . x(0) = 0 ∆ Rue = x(T ) T wT w dt ≤ 1. 8. .

. . . .1) does not exceed γ. Λ and T subject to P > 0. . . . (8. . . . If there exists a Lyapunov function of the form (8. This is an EVP. . . τnp ) ≥ 0 and (8. this is −γ I 2 ≤ 0.3) and γ ≥ 0 such that d V (x) ≤ γ 2 wT w − z T z for all x and w satisfying (8.1). . τnp ) ≥ 0 such that T A T P + P A + C z Cz T Bp P T T P B p + A T Cq Λ + C q T T T ΛCq Bp + Bp Cq Λ − 2T + ΛCq A + T Cq ≤ 0. . z = Cz x + Dzp p + Dzu u + Dzw w t 0 t (8. . P > 0 If the nonlinearities φi are known.12) 8.4 L2 gain We assume that Dzw = 0 for simplicity. . . . in which the pointwise constraint p(t)T p(t) ≤ q(t)T q(t) is replaced with a constraint on the integrals of p(t)T p(t) and q(t)T q(t): x = Ax + Bp p + Bu u + Bw w. we can obtain possibly better bounds on J by modifying the objective in the EVP appropriately.10) T Since V (x(0)) ≤ x(0)T (P + Cq ΛCq )x(0).9) holds if there exists T = diag(τ1 .122 Chapter 8 Lur’e and Multiplier Methods Using the S-procedure. λnp ) and T = diag(τ1 . . T ≥ 0. Λ = diag(λ1 .10). T = diag(τ1 . we conclude that the condition (8.11) is equivalent to np 2 ξT P + i=1 λi πi Ci. P .q T (Aξ + Bp π) ≤ γ 2 wT w − ξ T Cz Cz ξ The smallest upper bound on the L2 gain. is therefore obtained by minimizing γ over γ. The condition (8. λnp ) ≥ 0.13) p(τ )T p(τ ) dτ ≤ q(τ )T q(τ ) dτ. for any ξ satisfying πi (πi − ci. τnp ): minimize subject to T x(0)T P + Cq ΛCq x(0) (8. Using the satisﬁed if there exists T = diag(τ1 .q ξ) ≤ 0. Λ = diag(λ1 . . (8. . .12). (8. .11) dt then the L2 gain of the system (8. an upper bound on J is obtained by solving the following EVP in the variables P . . . Λ ≥ 0. . provable using Lur’e Lyapunov functions. i = 1. . np .1.2 Integral Quadratic Constraints In this section we consider an important variation on the NLDI. 0 Copyright c 1994 by the Society for Industrial and Applied Mathematics. τnp ) ≥ 0 such that T T T T P Bw A P + P A + C z Cz P B p + A T Cq Λ + C q T T T T Bp P + ΛCq A + T Cq ΛCq Bp + Bp Cq Λ − 2T ΛCq Bw T Bw P T T B w Cq Λ S-procedure. . . ˙ q = Cq x + Dqp p + Dqu u + Dqw w. 8.

leaving others to the reader.14) xT P x + λ 0 q(τ )T q(τ ) − p(τ )T p(τ ) dτ < 0. it can be shown that limt→∞ x(t) = 0.2. Moreover. the system (8. or the system is stable. but not necessarily monotonically. Using standard arguments from Lyapunov theory.3) P > 0.13) is described as a linear system with (dynamic) nonexpansive feedback.e.2. An important example is when p and q are related by a linear system. As in Chapters 5–7. and roughly speaking. however. It is exactly the same as the condition obtained by applying the S-procedure to the condition d T x P x < 0. Thus. This system is not. In control theory terms. our analysis will be based on quadratic functions of the state V (ξ) = ξ T P ξ.2 L2 gain We assume that x(0) = 0 and.1 Stability Consider system (8. ˙ p = Cf xf + Df q. Suppose P > 0 and λ ≥ 0 are such that d T x P x < λ pT p − q T q .13) without w and z. Here. indeed. We can also consider a generalization of the DNLDI. xf (0) = 0. dt which in turn. the interpretation of V is diﬀerent here. a diﬀerential inclusion.. We will now show that many of the results on NLDIs (and DNLDIs) from Chapters 5–7 generalize to systems with integral quadratic constraints. For example. every trajectory of the associated NLDI is a trajectory of (8. for simplicity. where Df + Cf (sI − Af )−1 Bf ∞ ≤ 1. the V of Chapters 5–7 decreases monotonically to zero.8.3. the very same V decreases to zero. leads to the LMI condition for quadratic stability of NLDI (5. T AT P + P A + λCq Cq T (P Bp + λCq Dqp )T T P Bp + λCq Dqp T −λ(I − Dqp Dqp ) < 0. This electronic version is for personal use and may not be duplicated or distributed. that Dzw = 0 and Dqp = 0. we can consider integral quadratic constraints with diﬀerent sector bounds. We will demonstrate this generalization for stability analysis and L2 gain bounds. Now.14). dt or d dt t (8. xf = Af xf + Bf q. 8. in which we have componentwise integral quadratic constraints. Note that the second term is always nonnegative. the very same LMIs will arise. . in stability analysis. as an example we will encounter constraints t of the form 0 p(τ )T q(τ ) dτ ≥ 0 in §8. However. let us examine the condition (8. whenever pT p ≤ q T q.13).2 Integral Quadratic Constraints 123 Such a system is closely related to the NLDI with the same data matrices. V is not a Lyapunov function in the conventional sense. i. λ ≥ 0. 8.

3.3. Bm.i + Dm. np are any nonnegative numbers. i = 1.np ]T are the input and output respectively of the diagonal system with realization Am = diag(Am. z = Cz x + Dzw w.3). qm.i ).) ∆ Thus. (8.i + Bm.i . This implies that the L2 gain from w to z for system (8.i T Pi Bm.i − Cm.i ≤ 0. which in turn is equivalent to the existence of Pi ≥ 0. . 6 and 7 extend immediately to systems with integral quadratic constraints. The only exceptions are the results on coordinate transformations (§5. (8.13) does not exceed γ. and on reachable sets with unit-peak inputs (§6. we begin by deﬁning np LTI systems with input qi and output qm. np .15) Integrating both sides from 0 to T . .i xm.i (t) ∈ R: xm. Bm = diag(Bm.1 . .124 Chapter 8 Suppose P > 0 and λ ≥ 0 are such that d T x P x < λ pT p − q T q + γ 2 wT w − z T z. . . satisfying AT Pi + Pi Am.i T − Dm.i + Dm. We further assume that t 0 qi (τ )qm.2. . This last condition is equivalent to passivity of the LTI systems (8. We can consider the more general case when αi ≤ δi ≤ βi using a loop transformation (see the Notes and References). . . np .i is stable and (Am. Condition (8. Copyright c 1994 by the Society for Industrial and Applied Mathematics.17) where we assume that Am.i xm. Dm = diag(Dm.9) from w to z does not exceed γ.i qi . i = 1. . .1.3 Multipliers for Systems with Unknown Parameters x = Ax + Bp p + Bw w.2). . 8.i ˙ qm.17). Cm.i .i m.15) leads to the same LMI (6.i T Bm. .3. . i = 1. . . Cm = diag(Cm. xm. dt Lur’e and Multiplier Methods (8.i = = Am. We consider the system (8. T T x(T )T P x(T ) + λ 0 q(τ )T q(τ ) − p(τ )T p(τ ) dτ < 0 γ 2 wT w − z T z dt. where qm.3 and §7. ˙ pi = δ i q i .i Pi − Cm.i (τ ) dτ ≥ 0 for all t ≥ 0. q and qm = [qm.i ).55) which guarantees that the L 2 gain of the NLDI (4.16) where p(t) ∈ Rnp . .1. . . Remark: Almost all the results described in Chapters 5. np .18) (See §6. For reasons that will become clear shortly.i qi . i = 1. q = Cq x + Dqp p.i ).i ).i (0) = 0. also §7.i ) is controllable. . . and δi . .1.

consider the following system with integral quadratic constraints: d˜ x ˜x ˜ ˜ = A˜ + Bp p + Bw w. dt ˜ ˜ ˜ qm = Cq x + Dqp p. Suppose P > 0 is such that d T pi qm. it can be shown using standard arguments that limt→∞ x(t) = 0.2 Reachable sets with unit-energy inputs The set of reachable states with unit-energy inputs for the system x = Ax + Bp p + Bw w. . np . 8. Cm and Dm : ˜ ˜ ˜ ˜T A T P + P A P B p + Cq ˜T ˜T ˜ ˜ Bp P + Cq Dqp + Dqp < 0. ˜ Dqp = Dm Dqp .21) can also be interpreted as a positive dissipation condition for the LTI system with transfer matrix G(s) = −W (s)H(s).8. where W (s) = Cm (sI − Am )−1 Bm + Dm and H(s) = Cq (sI − A)−1 Bp + Dqp . then [xT xT ]T is a trajectory m of (8. See the Notes and References for details. 8.i (τ ) dτ ≥ 0. It is easy to show that if x is a trajectory of (8. ˙ p = δq.3 Multipliers for Systems with Unknown Parameters 125 Deﬁning ˜ A= A B m Cq 0 Am .19) is stable. (8. i.3.19) without w and z.i (τ ) dτ < 0. q = Cq x + Dqp p.21) Remark: LMI (8. x Px + 2 ˜ ˜ dt i=1 or d dt np t np (8. Therefore.e. if system (8. . we conclude that system (8. (8.22) This electronic version is for personal use and may not be duplicated or distributed.3. pi (τ )qm.19) i = 1.1 Stability Consider system (8.20) is just the LMI in P > 0. ..16) have implications for (8. Hence W is often called a “multiplier”.20) x Px + 2 ˜ ˜ i=1 0 T pi (τ )qm. Condition (8. conclusions about (8. ˜ Cz = [Cz 0].16) is stable if there exists a passive W such that −W (s)H(s) has positive dissipation. the ˜ system (8.19) is stable.19) for some appropriate xm . ˜ Bw = Bw 0 . ˜ Cq = [Dm Cq Cm ]. ˜ ˜ ˜ z = Cz x + Dqp p + Dzw w.i < 0. .16). Therefore.16). ˜ Bp = Bp Bm Dqp .19). δ ≥ 0. Then. t 0 (8. so is system (8. For instance. .

24) and xT P11 x0 − 1 0 T P12 x0 This is an LMIP. then the set E = x x P x ≤ 1 contains the reachable set for the system (8. (8. the point x0 lies outside the set of reachable states for the system (8. xT P12 0 P22 > 0.22) if there exist P > 0. w satisfy (8. Partitioning P conformally with the sizes of x0 and z as P = we note that minz x0 z T P11 T P12 P12 P22 .19).22). We can obtain a suﬃcient condition for a point x0 to lie outside the reachable set Rue .i < wT w. T ˜ ˜ qp + Dqp ˜T ˜ p P + Cq 0 B D Now. Cm . dt i=1 or np t t np (8. λ ≥ 0. and DM satisfying (8.23) is equivalent to the LMI in P > 0. 0 Therefore. . x0 does not belong to the reachable set for the system (8. Suppose P > 0 is such that d T x Px + 2 ˜ ˜ pi qm. Pi ≥ 0. ˜ ˜ ˜ Condition (8.126 is Rue = ∆ Chapter 8 Lur’e and Multiplier Methods x.18).23) xT P x + 2 ˜ ˜ i=1 0 pi (τ )qm. x(T ) 0 T wT w dt ≤ 1.i (τ ) dτ < 0 T w(τ )T w(τ ) dτ. Notes and References Lur’e systems The Notes and References of Chapters 5 and 7 are relevant to this chapter as well. Cm and Dm : T ˜ ˜ ˜ ˜ ˜T A P + P A P B w P B p + Cq ˜T (8. P11 T P12 P12 P22 x0 z = −1 T xT (P11 − P12 P22 P12 )x0 . using quadratic functions for the augmented system.22) if there exists P satisfying (8.24) −I 0 Bw P < 0.23) and there exists no z such that x0 z T P x0 z ≤ 1. x(0) = 0 T ≥0 . Lur’e and Postnikov were the ﬁrst to propose Lyapunov functions consisting of a quadratic form Copyright c 1994 by the Society for Industrial and Applied Mathematics.

Since 0 pT p dτ ≤ given by ∆ t p = g(xf . p = g(xf . Tsy64b]. Tsy64a. SP94d]. Another early book that uses such Lyapunov functions is Letov [Let61]. q. Yakubovich [Yak64]. it can be shown (see for example [Wil72]) that Vf : Rnf → R+ T Vf (ξf ) = sup − 0 q T q − pT p dt xf (0) = ξf . In any case. it was not known in the 1940’s and 1950’s that solution of convex inequalities is practically and theoretically tractable. t). Yak92. Sko91b] for general numerical methods to prove stability of nonlinear systems via the Popov or circle criterion. A precursor is the paper by Karmarkar and Siljak [KS75]. in the early books by Lur’e. see Siljak [Sil89] for a survey. sometimes under the name “real-µ problems”. t). Meg92b. ∆ Note that this construction also shows that ξ T P ξ proves the stability of the NLDI (5. especially in the former Soviet Union. that they form a parallelogram.2. T ≥ 0 is a Lyapunov function that proves stability of system (8. Moreover. Tsypkin [Tsy64d.13) with integral quadratic constraints.2). In some special cases they even point out geometrical properties of the regions deﬁned by the inequalities. We now show how we can explicitly construct a Lyapunov function given P . Systems with constant. and Letov the problem of constructing Lyapunov functions reduces to the satisfaction of some inequalities. ˙ where xf (t) ∈ Rnf .Notes and References 127 plus the integral of the nonlinearity [LP44].25) t 0 q T q dτ .. Integral quadratic constraints have recently received much attention. See for example [Yak88. SP94h. t). See also [IR64]. in cases when p and q are related by xf = f (xf .16) are widely studied in the ﬁeld of robust control. The results of §8. xf satisﬁes (8. Szego [Sze63] and Meyer [Mey66].3 were This electronic version is for personal use and may not be duplicated or distributed. Tsy64c. ξf ) = ξ T P ξ + Vf (ξf ) proves stability of the interconnection x = Ax + Bp p. Indeed. Sav91. The book [Lur57] followed this article. Meg92a. Construction of Lyapunov functions for systems with integral quadratic constraints We saw in §8. t). q. the Lyapunov function V : Rn × Rnf → R+ given by V (ξ. Work on this topic was continued by Popov [Pop62]. q. SP94b. PS83]. q. . in which the author showed how to solve the problem of ﬁnding such Lyapunov functions analytically for systems of order 2 or 3. ˙ xf = f (xf .25). convexity of the regions is not noted in the early work.1 that the quadratic positive function V (ξ) = ξ T P ξ is not necessarily a Lyapunov function in the conventional sense for the system (8. See also [Sko91a. ˙ q = Cq x + Dqp p.e. (8. unknown parameters Problems involving system (8. it does not do so monotonically.25). The ellipsoid method is used to construct Lyapunov functions for the Lur’e stability problem in the papers by Pyatnitskii and Skorodinskii [PS82. who used a locally convergent algorithm to determine margins for a Lur’e-Postnikov system with one nonlinearity. i. Postnikov. But as far as we know. although V (x(t)) → 0 as t → ∞. Jury and Lee [JL65] consider the case of Lur’e systems with multiple nonlinearities and derive a corresponding stability criterion.

This case has also been considered in [CS92b. we must also cite Yakubovich.6) is equivalent to the existence of q nonnegative such that (1 + qs)(cq (sI − A)−1 bp ) + 1/k has positive dissipation. the problem of checking whether system (8. involving the application of an oﬀ-axis circle criterion (see also §3.. the multipliers W are of the form (1 + qs) for some q ≥ 0. given a transfer function Hqp they show that the transfer function Hqp /(1+kHqp ) is stable for all values of k in (0.128 Chapter 8 Lur’e and Multiplier Methods derived by Fan.16) is stable is NP-hard (see Coxson and Demarco [CD91]. HB93b. This is the famous Popov criterion [Pop62. For example. For instance.e. along with Bernstein and Haddad. In contrast. Wil76] for discussion of and bibliography on multiplier theory and its connections to Lyapunov stability theory. HHHB92. In the case when ∆(q. when we have a Lur’e system). for some choice of W from a set that is determined by the properties of ∆. A related result. and Nemirovskii [Nem94]). and apply them to Aerospace problems. See also [CHD93]. HH93a. .3). SLC94] where the authors devise appropriate multipliers for constant real uncertainties. and how these tests can be reduced to LMIPs. Copyright c 1994 by the Society for Industrial and Applied Mathematics. Braatz and Young [BYDM93]. checking whether there exists a static.16) is stabilizable by a constant state-feedback can be shown to be NP-hard using the method of [Nem94]. We conjecture that checking whether a general DNLDI is stable is also NP-hard. t).g.. ZF67] to prove stability of systems subject to “monotonic” or “odd-monotonic” nonlinearities (see [SW87] for details. that feasibility of LMI (8. ˙ q = Cq x + Dqp p. Sil69. HB93a.. See also [LSC94]. HHH93b]. Likewise. The method involves checking that Hm = W Cq (sI − A)−1 Bp + Dqp satisﬁes Hm (jω) + Hm (jω)∗ ≥ 0 for every ω ∈ R. Tits and Doyle in [FTD91] using an alternate approach. the multiplier W can be any passive transfer matrix. Multiplier methods See [Wil69a. HH93b. NT73. A nice survey of these results can be found in Barmish [Bar93]. due to Blondel and Gevers [BG94]. Complexity of stabilization problems The stability (and performance) analysis problems considered in §8. Kharitonov’s theorem [Kha78] and its extensions [BHL89]. output-feedback control law for a single LTI system is rationally decidable [ABJ75]. Hall and How. This observation was made by Brockett and Willems in [BW65]. checking if (8. see for example [HB91a.3 have high complexity. t) = δq(t) for some unknown real number δ. many stabilization problems for uncertain systems are NP-hard. generalize the use of quadratic Lyapunov functions along with multipliers for various classes of nonlinearities. Pop64]. consisting of a Lur’e system with unknown parameters. p(t) = ∆(q. Poljak and Rohn [PR94]. e. using the S-procedure [Yak77]. states that checking whether there exists a common stabilizing LTI controller for three LTI systems is undecidable. SC93. HCB93. Safonov and Wyetzner use a convex parametrization of multipliers found by Zames and Falb [ZF68. Wil70. We also mention [TV91]. who showed. The paper [BHPD94] by Balakrishnan et al. where Tesi and Vicino study a hybrid system. see also [GG94]). SL93a. When ∆ is a nonlinear time-invariant memoryless operator satisfying a sector condition (i. DV75. There are elegant analytic solutions for some very special problems involving parameterdependent linear systems. In general multiplier theory we consider the system x = Ax + Bp p + Bw w. HHH93a. ∞) if and only if there exists a passive multiplier W such that W Hqp is also passive. shows how various stability tests for uncertain systems can be rederived in the context of multiplier theory. How93.

including so-called inﬁnite sectors. p We can therefore express (8.3) nonconservatively by generating appropriate LMIs. This results in ˙ ¯ x ˙ q = = (I − DΛ)−1 Cx + (I − DΛ)−1 DΓ¯.e. i. p50]. −1 ¯ ¯ C = (I − DΛ) C. it is probably better to derive the LMIs associated with the more general Lur’e system than to loop transform to the standard Lur’e system. 1].26) and. pi (t) = φi (qi (t)). Γ = diag(β1 − α1 .Notes and References Multiple nonlinearities for Lur’e systems 129 Our analysis method for Lur’e systems with multiple nonlinearities in §8. C and D to simplify the presentation. i. We assume this system is well-posed. solve for x and q in terms of x and p. We now substitute p = Γ¯ + Λq into (8. This electronic version is for personal use and may not be duplicated or distributed. . (8. ˙ q = Cx + Dp. The term loop transformation comes from a simple block-diagram interpretation of the equations given above. In practice. i. that the number of these LMIs grows exponentially with the number of nonlinearities. . B = B(I − ΛD)−1 Γ.e. αi σ 2 ≤ σφi (σ) ≤ βi σ 2 . however. D = (I − DΛ)−1 DΓ. .27) . .27)..1 to the loop-transformed system (8.26) Here we have dropped the variables w and z and the subscripts on B. αnq ). .. since the S-procedure can be conservative in this case. So to analyze the more general Lur’e system (8. βi ] into the standard sector.26) as ¯ ¯¯ ¯ ¯¯ x = Ax + B p..26).. . Kamenetskii [Kam89] and Rapoport [Rap89] derive corresponding frequency-domain stability criteria. using our well¯ p posedness assumption. Deﬁne pi = ¯ ∆ 1 ¯ (φi (qi ) − αi qi ) = φi (qi ). Note that (8. Similar transformations can be used to map any set of sectors into any other. we simply apply the methods of §8. so that p = Γ−1 (p − Λq). Kamenetskii calls his results “convolution method for solving matrix inequalities”. Rap87. βi = +∞. Let Λ and Γ denote the diagonal matrices Λ = diag(α1 . det(I − D∆) = 0 for all diagonal ∆ with αi ≤ ∆ii ≤ βi . . The construction above is a loop transformation that maps nonlinearities in sector [α i . βi − α i ¯ It is readily shown that 0 ≤ σ φi (σ) ≤ σ 2 for all σ. p A + BΛ(I − DΛ)−1 C x + B(I − ΛD)−1 Γ¯.27) is in the standard Lur’e system form. Rap88] devises a way of determining if there exists Lyapunov functions of the form (8. (8.1 can be conservative. Loop transformations Consider the Lur’e system with general sector conditions.e. in which.e. For detailed descriptions of loop transformations. q = Cx + Dp. ˙ x = Ax + Bp. ¯i (qi (t)). in a numerical implementation. Rapoport [Rap86. i. see the book by Desoer and Vidyasagar [DV75. [0. pi (t) = φ ¯ 0 ≤ σφ where ¯ ¯ A = A + BΛ(I − DΛ)−1 C. say. . We note. ¯i (σ) ≤ σ 2 . βnq − αnq ).

.

we say the system is mean-square stable. i M (0) = E x(0)x(0)T .4) An alternative. Of course.e. regardless of x(0). as M (k) = E x(k)x(k)T . 131 . . . that x(k) → 0 almost surely. the state correlation matrix. . p(1).. M satisﬁes the linear recursion L ∆ M (k + 1) = AM (k)AT + i=1 2 σi Ai M (k)AT . Mean-square stability implies. σL ). E p(k)p(k)T = Σ = diag(σ1 .3) If this linear recursion is stable. (9.Chapter 9 Systems with Multiplicative Noise 9. . Deﬁne M (k). ..1 Analysis of Systems with Multiplicative Noise 9.2) We assume that x(0) is independent of the process p.1) where p(0). (9. equivalent condition is that the dual inequality L AQAT − Q + 2 σi Ai QAT < 0 i i=1 (9.1.5) holds for some Q > 0. i=1 (9. . identically distributed random variables with E p(k) = 0.1 Mean-square stability We ﬁrst consider the discrete-time stochastic system L x(k + 1) = A0 + i=1 Ai pi (k) x(k). . It can be shown (see the Notes and References) that mean-square stability is equivalent to the existence of a matrix P > 0 satisfying the LMI L AT P A − P + 2 T σi Ai P Ai < 0. are independent. (9. i. for example. limk→∞ M (k) = 0.

Again. i. the GEVP oﬀers no computational or theoretical advantages. We can consider variations on the mean-square stability problem. . since the mean-square stability margin can be obtained by computing the eigenvalue of a (large) matrix.3) (see the Notes and References). Remark: Mean-square stability can be veriﬁed directly by solving the Lyapunov equation in P L AT P A − P + 2 σi A T P A i + I = 0 i i=1 and checking whether P > 0.e. .7) E V (x(k)) ≤ w(j)T w(j). Here. for example. (9. Let x(k) denote the mean of x(k). i. . Then. x(0) = 0. ¯ Suppose V (ξ) = ξ T P ξ where P > 0 satisﬁes E V (x(k + 1)) ≤ E V (x(k)) + w(k)T w(k) for all trajectories. with energy not exceeding one. which satisﬁes x(k + 1) = A¯(k) + Bw w(k). We ¯ ¯ will develop joint bounds on x(k) and X(k). and let ¯ ¯ x T X(k) denote the covariance of x(k). which implies that.5) oﬀer no computational (or theoretical) advantages for checking mean-square stability. we are given A0 . E x(k)T P x(k) = x(k)T P x(k) + E (x(k) − x(k)) P (x(k) − x(k)) ¯ ¯ ¯ ¯ = x(k)T P x(k) + Tr X(k)P ¯ ¯ ≤ 1..2 State mean and covariance bounds with unit-energy inputs We now add an exogenous input w to our stochastic system: L x(k + 1) = Ax(k) + Bw w(k) + i=1 (Ai x(k) + Bw.i w(k)) pi (k). determining the mean-square stability margin. ∞ k=0 w(k)T w(k) ≤ 1. 9.4) and (9. . and asked to ﬁnd the largest γ such that with Σ < γ 2 I. the system (9.1. T Copyright c 1994 by the Society for Industrial and Applied Mathematics.4).e. k (9.1). j=0 k ≥ 0. From the LMIP characterizations of mean-square stability. Thus the LMIPs (9.6) We assume that the exogenous input w is deterministic. for all k ≥ 0.. X(k) = E (x(k) − x(k)) (x(k) − x(k)) . we can derive GEVP characterizations that yield the exact mean-square stability margin. AL . Alternatively. .132 Chapter 9 Systems with Multiplicative Noise Remark: For P > 0 satisfying (9. we can interpret the function V (ξ) = ξ T P ξ as a stochastic Lyapunov function: E V (x) decreases along trajectories of (9. the function V (M ) = Tr M P is a (linear) Lyapunov function for the deterministic system (9.1) is mean-square stable.

we conclude that x(k) and X(k) must belong to the set ¯ ¯ X(k)) x(k)T P x(k) + Tr X(k)P ≤ 1 (x(k).i x(k) + Dzw.9). We note that E V (x(k + 1)) = x(k) ¯ w(k) T 133 M x(k) ¯ w(k) L 2 σi AT P A i i i=1 + Tr AT P A + where M= AT T Bw L X(k).9. Therefore.i w(k)) pi (k). As another example. we can derive an upper bound on λmax (M ) by maximizing Tr P subject to (9.8). thus.i .8) and P > 0. For example. we conclude that LMI (9. the LMI conditions AT T Bw and L L P A Bw + i=1 2 σi AT i T Bw.i ≤ P 0 0 I (9. x(0) = 0.i P Ai Bw.i P Ai Bw.1. (9. we can derive bounds on Tr M by maximizing λmin (P ) subject to (9. since we have Tr M (k)P ≤ λmin (P ) Tr M (k) and therefore Tr M (k) ≤ 1/λmin (P ).i w(k)) pi (k). which is another EVP.8) implies (9. z(k) = Cz x(k) + Dzw w(k) + i=1 This electronic version is for personal use and may not be duplicated or distributed.8) and P > 0. P A Bw + i=1 2 σi AT i T Bw.8) AT P A + i=1 2 σi AT P A i < P i (9. 9. LMI (9. This is an EVP.7).7). ¯ ¯ for any P > 0 satisfying the LMI (9.3 Bound on L2 gain We now consider the system L x(k + 1) = Ax(k) + Bw w(k) + i=1 L (Ai x(k) + Bw. We can derive further bounds on M (k).10) (Cz. Next.1 Analysis of Systems with Multiplicative Noise Let us now examine condition (9.8) implies LMI (9. For example. k ≥ 0. we have the bound Tr M (k)P ≤ 1 on the state covariance matrix. . Of course.9) imply E V (x(k + 1)) ≤ xT (k)P x(k) + w(k)T w(k) ¯ ¯ L + Tr AT P A + ≤ = i=1 T 2 σi AT P A i i X(k) xT (k)P x(k) + w(k) w(k) ¯ ¯ T + E (x(k) − x(k)) P (x(k) − x(k)) ¯ ¯ E V (x(k)) + w(k)T w(k).

We deﬁne the L2 gain η of this system as η = sup E k=0 2 ∆ ∞ ∞ z(k) z(k) k=0 T w(k)T w(k) ≤ 1 .i x(k) + Dzu.i u(k) + Bw. satisﬁes Then γ ≥ η.i w(k)) pi (k).i Bw. Bw Dzw 2 σi i=1 T (9. i=1 z(k) = (Cz + Dzu K)x(k) + Dzw w(k)+ L (9. The condition (9.i w(k)) pi (k).i P 0 0 γ2I Bw. 9. .i K)Q(Ai + Bu.4 to obtain componentwise results.i w(k)) pi (k).i K)x(k) + Bw.i + Dzu.13) + Ai Cz. i=1 We seek a state-feedback u(k) = Kx(k) such that the closed-loop system x(k + 1) = (A + Bu K)x(k) + Bw w(k)+ L ((Ai + Bu. 9.i Dzw.i K)T < 0.2. We assume that w is deterministic.14) (Cz. Copyright c 1994 by the Society for Industrial and Applied Mathematics. yields an upper bound on η. with P > 0.i Minimizing γ subject to LMI (9. i=1 satisﬁes various properties.2 State-Feedback Synthesis x(k + 1) = Ax(k) + Bu u(k) + Bw w(k)+ L We now add a control input u to our system: (Ai x(k) + Bu.i Dzw.i K)x(k) + Dzw. Remark: It is straightforward to apply the scaling method developed in §6.12) P 0 0 I T A Cz P 0 Bw Dzw 0 I − Ai Cz. (9.134 Chapter 9 Systems with Multiplicative Noise with the same assumptions on p.11) Suppose that V (ξ) = ξ T P ξ.15) ((Cz.i u(k) + Dzw. which is an EVP.13).1 Stabilizability We seek K and Q > 0 such that (A + Bu K)Q(A + Bu K)T − Q L + i=1 2 σi (Ai + Bu.i w(k)) pi (k).3. i=1 z(k) = Cz x(k) + Dzu u(k) + Dzw w(k)+ L (9.i ≤ 0. (9.12) can be shown to be equivalent to the LMI A Cz L E V (x(k + 1)) − E V (x(k)) ≤ γ 2 w(k)T w(k) − E z(k)T z(k).

ﬁnding K that maximizes γ reduces to the following GEVP: maximize subject to γ Q > 0.2.i K Cz. system (9. .i K Bw Dzw ≤ Bw. (9.i Y Cz.2 Minimizing the bound on L2 gain We now seek a state-feedback gain K which minimizes the bound on L2 gain (as deﬁned by (9.14) is mean-square stabilizable (with constant state-feedback) if and only if the LMI Q > 0. Therefore.9.17) holds for some Y and Q > 0.i T T ˜ Q−1 Ai Q + Bu.2 State-Feedback Synthesis 135 With the change of variable Y = KQ.11)) for system (9.i + Dzu.i Dzw.i Q + Dzu. Thus.i K Cz.i Q + Dzu.i Y )Q−1 (Ai Q + Bu.15) is mean-square stable for Σ ≤ γ 2 I if and only if (AQ + Bu Y )Q−1 (AQ + Bu Y )T − Q L +γ 2 i=1 (Ai Q + Bu. System (9.i Y )Q−1 (Ai Q + Bu. In this case.i T T ˜ P Ai + Bu.i Y )T < 0 (9.i Y Bw Dzw Bw.17) 9. Applying a congruence ˜ with Q = diag(Q.i Y )T < 0.i Dzw.i P 0 0 γ2I + where ˜ P = A + Bu K Cz + Dzu K ˜ P A + Bu K Cz + Dzu K P 0 0 I .i Dzw. This electronic version is for personal use and may not be duplicated or distributed. we obtain the equivalent condition (AQ + Bu Y )Q−1 (AQ + Bu Y )T − Q L + i=1 2 σi (Ai Q + Bu. Y = KQ. a stabilizing state-feedback gain is K = Y Q−1 . + AQ + Bu Y Cz Q + Dzu Y ˜ Q−1 AQ + Bu Y Cz Q + Dzu Y This inequality is readily transformed to an LMI. we get L 2 σi i=1 Ai Q + Bu.16) is feasible.i Y Cz.i K Bw Dzw Bw.i Dzw. (9.i Y Bw Dzw ≤ Bw.i Q 0 0 γ2I . I). We make the change of variables Q = P −1 . we can maximize the closed-loop mean-square stability margin. η ≤ γ for some K if there exist P > 0 and K such that L 2 σi i=1 Ai + Bu.i + Dzu. As an extension. (9.15).16) which is easily written as an LMI in Q and Y .

The related Lyapunov equation was given by Kleinmann [Kle69] for the case L = 1 (i. For the systems considered here. Linearity of system (9. for an arbitrary choice of M (0). To prove that condition (9. [EP92]) E t=0 x(t) 2 has a (ﬁnite) limit as T → ∞ for all initial conditions. and A is a real matrix. A o A A rigorous treatment of continuous-time stochastic systems requires much technical machinery (see [EG94]). Wil73]. see below. where m(k) is a vector containing the n 2 elements of M (k). Introduce the (linear Lyapunov) function V (M ) = Tr M P . (This matrix can be expressed via Kronecker products Copyright c 1994 by the Society for Industrial and Applied Mathematics. Kushner used this idea to derive bounds for return time. For nonzero M (k) satisfying (9. Meansquare stability is a strong form of stability. this approach does not provide a way to construct the Lyapunov functions.3) is stable. The equation for the state correlation matrix (9.4) is a necessary and suﬃcient condition for mean-square stability.10). In particular. the solution to the linear recursion (9. the corresponding solution M (k) converges to 0. L V (M (k + 1)) = = < Tr AM (k)AT + i=1 2 σi Ai M (k)AT i L P Tr M (k) AT P A + i=1 2 σi A T P A i i Tr M (k)P = V (M (k)). We ﬁrst prove that condition (9. The criterion then reduces to an H2 norm condition on a certain transfer matrix. It is reminiscent of the classical (deterministic) Hurwitz criterion. .. which is positive on the cone of nonnegative matrices.3) goes to zero as k → 0.3).4) is also necessary. Proof of stability criterion We now prove that condition (9. we assume that the system is mean-square stable. a single uncertainty). In fact there are many ways to formulate the continuous-time version of a system such as (9.4) is suﬃcient.e.) Kats and Krasovskii introduced in [KK60] the idea of a stochastic Lyapunov function and proved a related Lyapunov theorem. However. Frequencydomain criteria are given in [WB71].3) can be found in [McL71. A thorough stability analysis of (nonlinear) stochastic systems was made by Kushner in [Kus67] using this idea of a stochastic Lyapunov function. it implies stability of the mean E x(k) and that all trajectories converge to zero with probability one [Kus67. The deﬁnition of mean-square stability can be found in [Sam59]. (This is in parallel with the results of [EP92].3) as m(k + 1) = Am(k).3) implies that. For an introduction to this topic see Arnold [Arn74] or ˚str¨m [˚st70]. (9.4). This means that for every positive initial condition M (0) ≥ 0..136 Chapter 9 Systems with Multiplicative Noise Notes and References Systems with multiplicative noise Systems with multiplicative noise are a special case of stochastic systems. Let P > 0 satisfy (9. etc. which is the property that (see T for example. A further standard argument from Lyapunov theory proves that M (k) converges to zero as k → ∞. We now write (9. expected output energy. o Stability of system with multiplicative noise For a survey of diﬀerent concepts of stability for stochastic systems see [Koz69] and the references therein.e. An algebraic criterion for mean-square stability of a system with multiplicative white noise is given in [NS72]. it is also equivalent to L2 -stability. i. Wil73]. see Itˆ [Ito51] and Stratonovitch [Str66].

and let N (k) be the corresponding solution of (9.18) implies has a (ﬁnite) limit for k → ∞. .18) is also stable. T Bi P Bi ≤ γ.4). The fact that N (0) > 0 implies P > 0. namely for so-called “block-diagonal” perturbations. are integrals of white noise). Mean-square stability margin System (9. which only considers the average behavior of the system. the function k P (k) = j=0 N (j) Now (9. and the corresponding matrix diﬀerence equation. Since N (k) satisﬁes a stable ﬁrst-order linear diﬀerence equation with constant coeﬃcients. Now let N (0) > 0. . . the processes pi . where i=1 Lipschitz norm. L This electronic version is for personal use and may not be duplicated or distributed. In [EP92]. L P (k + 1) = N (0) + AT P (k)A + i=1 2 σi AT P (k)Ai i Taking the limit k → ∞ shows that P satisﬁes (9. For systems with deterministic parameters only known to lie in ranges. . are standard Brownian motions (which. L N (k + 1) = AT N (k)A + i=1 2 σi AT N (k)Ai . L. . .10) can be viewed as a linear system subject to random parameter variations.18). . leads the authors of [EP92] to the robustness margin minimize subject to γ P > 0. the computation of an exact stability margin is an NP-hard problem [CD92]. .) Stability of A is equivalent to that of AT . In contrast. AT P + P A + C T C < 0. .Notes and References 137 involving A0 . ∆i 2 Measuring the “size” of the perturbation term by . L 1/2 · denotes the i = 1. . with ∆i (0) = 0. Roughly speaking. where the operators ∆i are Lipschitz continuous. . i = 1. they consider a continuous-time Itˆ system of the form o L dx(t) = Ax(t)dt + i=1 Bi ∆i (Cx(t))dpi (t). roughly speaking. An exact stability margin can also be computed for continuous-time systems with multiplicative noise provided the Itˆ framework is considered. we are dealing with a very special form of stability. El Bouhtouri and Pritchard provide a complete robustness analysis in a slightly diﬀerent framework than ours. We note that our framework can be recovered by assuming that the operators ∆ i are restricted to a diagonal structure of the form ∆i (z) = σi z. the computation o of an exact stability margin seems diﬃcult for Stratonovitch systems. . see [EG94]. which we denote by P . i (9. AL . . The fact that a stochastic analog of the deterministic stability margin is easier to compute should not be surprising.

Since the system is mean-square stable. Note that the LMI approach can also solve the state-feedback synthesis problem in the framework of [EP92]. This is consistent with the earlier results of Kleinmann [Kle69] and Willems [WB71] (see also [BB91. State-feedback synthesis Conditions for stabilizability for arbitrary noise intensities are given in geometric terms in [WW83]. Stability conditions for arbitrary noise intensities are given in geometric terms in [WW83]. and B1 and C are vectors. In this case.. identically distributed with E w(k) = 0. where the time-varying matrix A(k) is only known to belong to a given polytope. See the Notes and References for Chapter 6 for details on how these more complicated situations can be handled. see Wonham [Won66]. B1 . This extension can be cast in terms of LMIs. In our framework. Bounds on state covariance with noise input We consider system (9. . of course.1 can be extended to the stochastic framework.138 Chapter 9 Systems with Multiplicative Noise For L = 1. The state correlation M (k) satisﬁes the diﬀerence equation L T M (k + 1) = AM (k)AT + Bw W Bw + i=1 2 T σi Ai M (k)AT + Bw. the above LMI formulation extends immediately to more complicated situations (for instance. Extension to other stochastic systems The Lur’e stability problem considered in §8. p114] for a related result). We derive an upper bound on the state ¯ correlation (or covariance) of the system. that is.i . However. the matrix M∞ can be computed directly as the solution of a linear matrix equation. It is also possible to extend the results in this chapter to “uncertain stochastic systems”. ∆i = σi I.i . w(k) are independent. For L > 1 however. i In fact. since we only consider “diagonal” perturbations with “repeated elements”. applying their results to our framework would be conservative. the state mean x is zero. of the form L x(k + 1) = A(k) + i=1 σi Ai pi (k) x(k). when a single scalar uncertain parameter perturbs an LTI system. this happens if and only if the optimal value of the corresponding GEVP is zero. i. C). i with M (0) = 0. while the “direct” method does not.6) in which w is a unit white noise process. Apart from this work. E w(k)w(k)T = W . and w is independent of p. El Bouhtouri and Pritchard provide a solution of this problem in [EP94].i W Bw. the state correlation M (k) has a (ﬁnite) limit M ∞ which satisﬁes L T M∞ = AM∞ AT + Bw W Bw + i=1 2 T σi Ai M∞ AT + Bw. the bulk of earlier research on this topic concentrated Copyright c 1994 by the Society for Industrial and Applied Mathematics. when the white noise input w has a covariance matrix with unknown oﬀ-diagonal elements). both this result and ours coincide with the H2 norm of (A.i W Bw.e.

The solution of this problem.g.Notes and References 139 on Linear-Quadratic Regulator theory for continuous-time Itˆ systems. with no guarantee of global convergence.e. can be expressed as a (linear. as found in [McL71] or [Ben92]. by recognizing that the solution is an extremal point of a certain LMI feasibility set. with guaranteed convergence).. . Won68. RSH90]) use homotopy algorithms to solve these equations. This electronic version is for personal use and may not be duplicated or distributed. constant) state-feedback u(t) = Kx(t) where the gain K is given in terms of the solution of a non-standard Riccati equation. (See [Kle69. FR75. We can solve these non-standard Riccati equations reliably (i. [Phi89. BH88b] for additional details.) The existing methods (see e. which addresses the o problem of ﬁnding an input u that minimize a performance index of the form ∞ J(K) = E 0 x(t)T Qx(t) + u(t)T Ru(t) dt where Q ≥ 0 and R > 0 are given.

.

Chapter 10

Miscellaneous Problems

10.1 Optimization over an Aﬃne Family of Linear Systems

We consider a family of linear systems, x = Ax + Bw w, ˙ z = Cz (θ)x + Dzw (θ)w,

p

(10.1)

where Cz and Dzw depend aﬃnely on a parameter θ ∈ R . We assume that A is stable and (A, Bw ) is controllable. The transfer function, depends aﬃnely on θ. Several problems arising in system and control theory have the form minimize subject to ϕ0 (Hθ ) ϕi (Hθ ) < αi , i = 1, . . . , p (10.2) Hθ (s) = Cz (θ)(sI − A)−1 Bw + Dzw (θ),

∆

where ϕi are various convex functionals. These problems can often be recast as LMI problems. To do this, we will represent ϕi (Hθ ) < αi as an LMI in θ, αi , and possibly some auxiliary variables ζi (for each i): Fi (θ, αi , ζi ) > 0. The general optimization problem (10.2) can then be expressed as the EVP minimize subject to α0 Fi (θ, αi , ζi ) > 0, i = 0, 1, . . . , p

10.1.1

Hθ

H2 norm

2 2

**The H2 norm of the system (10.1), i.e., = 1 Tr 2π
**

∞

Hθ (jω)∗ Hθ (jω) dω,

0

T is ﬁnite if and only if Dzw (θ) = 0. In this case, it equals Tr Cz Wc Cz , where Wc > 0 is the controllability Gramian of the system (10.1) which satisﬁes (6.6). Therefore, the H2 norm of the system (10.1) is less than or equal to γ if and only if the following conditions on θ and γ 2 are satisﬁed:

Dzw (θ) = 0,

The quadratic constraint on θ is readily cast as an LMI. 141

Tr Cz (θ)Wc Cz (θ)T ≤ γ 2 .

142

Chapter 10

Miscellaneous Problems

10.1.2

H∞ norm

∞

From the bounded-real lemma, we have Hθ such that AT P + P A T Bw P P Bw −γ 2 I + Cz (θ)T Dzw (θ)T

< γ if and only if there exists P ≥ 0 Cz (θ) Dzw (θ) ≤ 0.

Remark: Here we have converted the so-called “semi-inﬁnite” convex constraint Hθ (iω) < γ for all ω ∈ R into a ﬁnite-dimensional convex (linear matrix) inequality.

10.1.3

Entropy

The γ-entropy of the system (10.1) is deﬁned as ∞ 2 −γ log det(I − γ 2 Hθ (iω)Hθ (iω)∗ ) dω, ∆ 2π −∞ Iγ (Hθ ) = ∞,

if Hθ

∞

< γ,

otherwise.

T When it is ﬁnite, Iγ (Hθ ) is given by Tr Bw P Bw , where P is a symmetric matrix with the smallest possible maximum singular value among all solutions of the Riccati equation

AT P + P A + Cz (θ)T Cz (θ) +

1 T P Bw Bw P = 0. γ2 λ is therefore equivalent to an

For the system (10.1), the γ-entropy constraint Iγ ≤ LMI in θ, P = P T , γ 2 , and λ: T A P + P A P Bw Cz (θ)T T Dzw (θ) = 0, Bw P −γ 2 I 0 Cz (θ) 0 −I

≤ 0,

T Tr Bw P Bw ≤ λ.

10.1.4

Dissipativity

Suppose that w and z are the same size. The dissipativity of Hθ (see (6.59)) exceeds γ if and only if the LMI in the variables P = P T and θ holds: AT P + P A T Bw P − Cz (θ) P Bw − Cz (θ)T 2γI − Dzw (θ) − Dzw (θ)T ≤ 0.

We remind the reader that passivity corresponds to zero dissipativity.

10.1.5

Hankel norm

The Hankel norm of Hθ is less than γ if and only if the following LMI in Q, θ and γ 2 holds: AT Q + QA + Cz (θ)T Cz (θ) ≤ 0, Dzw (θ) = 0, 1/2 1/2 2 γ I − Wc QWc ≥ 0, Q ≥ 0. (Wc is the controllability Gramian deﬁned by (6.6).)

Copyright c 1994 by the Society for Industrial and Applied Mathematics.

10.2

Analysis of Systems with LTI Perturbations

143

**10.2 Analysis of Systems with LTI Perturbations
**

In this section, we address the problem of determining the stability of the system x = Ax + Bp p, ˙ q = Cq x + Dqp p, pi = δi ∗ qi , i = 1, . . . , np , (10.3)

where p = [p1 · · · pnp ]T , q = [q1 · · · qnp ]T , and δi , i = 1, . . . , np , are impulse responses of single-output single-output LTI systems whose H∞ norm is strictly less than one, and ∗ denotes convolution. This problem has many variations: Some of the δi may be impulse response matrices, or they may satisfy equality constraints such as δ1 = δ2 . Our analysis can be readily extended to these cases. Let Hqp (s) = Cq (sI −A)−1 Bp +Dqp . Suppose that there exists a diagonal transfer matrix W (s) = diag(W1 (s), . . . , Wnp (s)), where W1 (s), . . . , Wnp (s) are single-input single-output transfer functions with no poles or zeros on the imaginary axis such that sup W (iω)Hqp (iω)W (iω)−1 < 1

ω∈R

(10.4)

Then from a small-gain argument, the system (10.3) is stable. Such a W is referred to as a frequency-dependent scaling. Condition (10.4) is equivalent to the existence of γ > 0 such that Hqp (iω)∗ W (iω)∗ W (iω)Hqp (iω) − W (iω)∗ W (iω) + γI ≤ 0, for all ω ∈ R. From spectral factorization theory, it can be shown that such a W exists if and only if there exists a stable, diagonal transfer matrix V and positive µ such that Hqp (iω)∗ (V (iω) + V (iω)∗ )Hqp (iω) − (V (iω) + V (iω)∗ ) + γI ≤ 0, V (iω) + V (iω)∗ ≥ µI, (10.5)

for all ω ∈ R. In order to reduce (10.5) to an LMI condition, we restrict V to belong to an aﬃne, ﬁnite-dimensional set Θ of diagonal transfer matrices, i.e., of the form Vθ (s) = DV (θ) + CV (θ)(sI − A)−1 B, where DV and CV are aﬃne functions are θ. Then, Hqp (−s)T Vθ (s)Hqp (s) − Vθ (s) has a state-space realization (Aaug , Baug , Caug (θ), Daug (θ)) where Caug and Daug are aﬃne in θ, and (Aaug , Baug ) is controllable. From §10.1.4, the condition that Vθ (iω) + Vθ (iω)∗ ≥ µI for all ω ∈ R is equivalent to the existence of P1 ≥ 0, θ and γ > 0 such that the LMI A T P1 + P 1 A V V (P1 BV − CV (θ)T )T P1 BV − CV (θ)T −(DV (θ) + DV (θ)T ) + µI ≤0

holds. The condition Hqp (iω)∗ (Vθ (iω) + Vθ (iω)∗ )Hqp (iω) − (Vθ (iω) + Vθ (iω)∗ ) + γI ≤ 0 T is equivalent to the existence of P2 = P2 and θ satisfying the LMI AT P2 + P2 Aaug aug (P2 Baug + Caug (θ)T )T P2 Baug + Caug (θ)T Daug (θ) + Daug (θ)T + γI ≤ 0. (10.6)

(See the Notes and References for details.) Thus proving stability of the system (10.3) using a ﬁnite-dimensional set of frequency-dependent scalings is an LMIP.

This electronic version is for personal use and may not be duplicated or distributed.

144

Chapter 10

Miscellaneous Problems

**10.3 Positive Orthant Stabilizability
**

˙ The LTI system x = Ax is said to be positive orthant stable if x(0) ∈ R n implies that + x(t) ∈ Rn for all t ≥ 0, and x(t) → 0 as t → ∞. It can be shown that the system + x = Ax is positive orthant stable if and only if Aij ≥ 0 for i = j, and there exists ˙ a diagonal P > 0 such that P AT + AP < 0. Therefore checking positive orthant stability is an LMIP. We next consider the problem of ﬁnding K such that the system x = (A + BK)x ˙ is positive orthant stable. Equivalently, we seek a diagonal Q > 0 and K such that Q(A + BK)T + (A + BK)Q < 0, with the oﬀ-diagonal entries of A + BK being nonnegative. Since Q > 0 is diagonal, the last condition holds if and only if all the oﬀ-diagonal entries of (A + BK)Q are nonnegative. Therefore, with the change of variables Y = KQ, proving positive orthant stabilizability for the system x = Ax+Bu ˙ is equivalent to ﬁnding a solution to the LMIP with variables Q and Y : Q > 0, (AQ + BY )ij ≥ 0, i = j, QAT + Y T B T + AQ + BY < 0. (10.7)

Remark: The method can be extended to invariance of more general (polyhedral) sets. Positive orthant stability and stabilizability of LDIs can be handled in a similar way. See the Notes and References.

**10.4 Linear Systems with Delays
**

Consider the system described by the delay-diﬀerential equation d x(t) = Ax(t) + dt

L i=1

Ai x(t − τi ),

(10.8)

**where x(t) ∈ Rn , and τi > 0. If the functional
**

L

V (x, t) = x(t)T P x(t) +

i=1 0

τi

x(t − s)T Pi x(t − s) ds,

(10.9)

**where P > 0, P1 > 0, . . . , PL > 0, satisﬁes dV (x, t)/dt < 0 for every x satisfying (10.8), then the system (10.8) is stable, i.e., x(t) → 0 as t → ∞. It can be veriﬁed that dV (x, t)/dt = y(t)T W y(t), where
**

L

**Therefore, we can prove stability of system (10.8) using Lyapunov functionals of the form (10.9) by solving the LMIP W < 0, P > 0, P1 > 0, . . . , PL > 0.
**

Copyright c 1994 by the Society for Industrial and Applied Mathematics.

A P + PA + AT P 1 W = . . . AT P L

T

Pi

P A1 −P1 . . . 0

i=1

···

··· .. .

···

P AL 0 , . . . −PL

x(t − τ1 ) y(t) = . . .

x(t)

x(t − τL )

.

. PL > 0. . . .6.10) is stable. we obtain the condition L (A + BK) P + P (A + BK) + AT P 1 W = . we add an input u to the system (10. . . there exists a state-feedback gain K such that a Lyapunov functional of the form (10. . 0 ··· .2 to check its stability. . . We also note that we can regard the delay elements of system (10.15) that we get for quadratic stability of a DNLDI. .1 Tangential Nevanlinna-Pick problem ∆ Given λ1 . AT P L T Pi P A1 −P1 . .8) and consider d x(t) = Ax(t) + Bu(t) + dt L i=1 Ai x(t − τi ).10. . . . if L for some P . checking stabilizability of the system (10. .8). the tangential Nevanlinna-Pick This electronic version is for personal use and may not be duplicated or distributed.10). and use the techniques of §10. Multiplying every block entry of W on the left and on the right by P −1 and setting Q = P −1 . . . ··· AL Q 0 < 0. . 145 Next.8) as convolution operators with unit L2 gain. AQ + QAT + BY + Y T B T + Qi i=1 T QA1 X= . is the same as the LMI W < 0 above. . Y . .5 Interpolation Problems 10. vm . . discussed in Chapter 5. . um . . From the discussion above. u1 . QL . −QL 10.5 Interpolation Problems Remark: The techniques for proving stability of norm-bound LDIs.2 to eliminate the matrix variable Y and obtain an equivalent LMIP with fewer variables. . Qi = P −1 Pi P −1 and Y = KP −1 . . . Q1 .5. . . . ··· P AL 0 <0 . with ui ∈ Cq and v1 . (10.8) using Lyapunov functionals of the form (10.9) is an LMIP in the variables Q. Remark: It is possible to use the elimination procedure of §2. m.9) proves the stability of the system (10.10) We seek a state-feedback u(t) = Kx(t) such that the system (10. with vi ∈ Cp .. . . P1 . . . . .. 0 i=1 ··· ··· . . T QAL A1 Q ··· −Q1 . . i = 1. . can also be used on the system (10. it can be veriﬁed that the LMI (5. −PL Thus. . in the case when the state x is a scalar. λm with λi ∈ C+ = {s | Re s > 0} and distinct. .

A∗ Gout + Gout A − V ∗ V = 0. with vi ∈ Cp . um .5. m . i = 1. and satisﬁes H(λi )ui = vi .3 Frequency response identiﬁcation We consider the problem of identifying the transfer function H of a linear system from noisy measurements of its frequency response at a set of frequencies. λm with λi ∈ C+ . A∗ Gout + Gout A − V ∗ P V = 0. .11) Problem (10. For future reference. This problem arises in multi-input multi-output H∞ control synthesis for systems with structured perturbations. We seek H satisfying two constraints: Copyright c 1994 by the Society for Industrial and Applied Mathematics. ∞ D = D∗ > 0 D ∈ D. . 10. u1 . .11) has a solution and only if the Pick matrix N deﬁned by Nij = ∗ u ∗ u j − v i vj i λ∗ + λ j ) i is positive semideﬁnite. . λm ). . where A∗ Gin + Gin A − U ∗ U = 0. 10. . N can also be obtained as the solution of the Lyapunov equation A∗ N + N A − (U ∗ U − V ∗ V ) = 0. a function H : C −→ Cp×q which is analytic in C+ . P ∈ D such that the following equations and inequality hold: A∗ Gin + Gin A − U ∗ P U = 0. With a change of variables P = D ∗ D and the discussion in the previous section. From Nevanlinna-Pick theory. . This is a GEVP.5.2 Nevanlinna-Pick interpolation with scaling We now consider a simple variation of problem (10. it follows that γopt is the smallest positive γ such that there exists P > 0. U = [u1 · · · um ]. . the problem is to ﬁnd γopt = inf DHD−1 H is analytic in C+ . i = 1. problem (10. . m. vm .11): Given λ1 . (10. . . . . m. . . i = 1. . (10. if possible. . H(λi )ui = vi . . . . . . . . Solving the tangential Nevanlinna-Pick problem simply requires checking whether N ≥ 0. γ 2 Gin − Gout ≥ 0. . . .12) where D is the set of m × m block-diagonal matrices with some speciﬁed block structure. Problem (10. . with ui ∈ Cp and v1 . where A = diag(λ1 .11) arises in multi-input multi-output H∞ control theory.146 Chapter 10 Miscellaneous Problems problem is to ﬁnd. .12) corresponds to ﬁnding the smallest scaled H∞ norm of all interpolants. with H ∞ ≤ 1. we note that N = Gin − Gout . V = [v1 · · · vm ].

jωm ). and is the measurement precision. ∗ (A + αI) Gout + Gout (A + αI) − (f + n)∗ (f + n) ≥ 0. we have fi = H(jωi ) + ni . determine if there exist Q ≥ 0 and R > 0.10. M 2 Gin − Gout ≥ 0. Gin > 0 and Gout > 0 such that (A + αI) Gin + Gin (A + αI) − e∗ e = 0. B) is stabilizable. and n = [n1 · · · nm ]. ni is the (unknown) measurement error. The inverse problem of optimal control is the following. ∗ 10. such that (Q. fi is the measurement of the frequency response at frequency ωi . 0 The solution of this problem can be expressed as a state-feedback u = Kx with K = −R−1 B T P . there exist H and n with H(ωi ) = fi + ni satisfying these conditions if and only if there exist ni with |ni | ≤ . with (A. ∗ (A + αI) Gout + Gout (A + αI) − (f + n)∗ (f + n) = 0. minimize M . where P is the unique nonnegative solution of AT P + P A − P BR−1 B T P + Q = 0. Here. ˙ x(0) = x0 . the LQR problem is to determine u that minimizes ∞ z T z dt. ∗ M 2 Gin − Gout ≥ 0. • For ﬁxed α and .6 The Inverse Problem of Optimal Control Given a system x = Ax + Bu. where A = diag(jω1 . . (Q. what is the “smallest” possible noise such that the measurements are consistent with the given values of α and M . . we can answer a number of interesting questions in frequency response identiﬁcation by solving EVPs and GEVPs. Given a matrix K. It can be shown that these conditions are equivalent to (A + αI) Gin + Gin (A + αI) − e∗ e = 0. • Prior assumption. A) is detectable and u = Kx is the This electronic version is for personal use and may not be duplicated or distributed. with |ni | ≤ . For some ni with |ni | ≤ . what is the smallest possible α-shifted H∞ norm of the system consistent with the measurements of the frequency response?” • For ﬁxed α and M . A) is detectable and R > 0. f = [f1 · · · fm ]. . . minimize . With this observation. Solving this GEVP answers the question “Given α and a bound on the noise values. . e = [1 · · · 1].6 The Inverse Problem of Optimal Control 147 • Consistency with measurements. Solving this EVP answers the question “Given α and a bound on α-shifted H∞ norm of the system. z= Q1/2 0 0 R 1/2 x u . α-shifted H∞ norm of H does not exceed M . From Nevanlinna-Pick theory.

. (10. u : Z+ → Rnu y : Z+ → Rny . in order to avoid state overﬂow in numerical implementation. B. where Wo is the observability Gramian of the original system (A. y(k) = Cx(k). Suppose we assume that the RMS value of the input is bounded by α and require the RMS value of the state to be less than one.16) subject to the overﬂow avoidance constraint (10. we seek R > 0 and Q ≥ 0 such that there exists P nonnegative and P1 positive-deﬁnite satisfying (A + BK)T P + P (A + BK) + K T RK + Q = 0. suppose that the state quantization noise is modeled as a white noise sequence w(k) with E w(k)T w(k) = η. ¯ (10. given by ∞ Wo = k=0 (AT )k C T CAk . R and Q. and its eﬀect on the output is measured by its RMS value. T −1 (zI − A)−1 B should be “small”. Equivalently. The system realization problem is ¯ to ﬁnd a change of state coordinates x = T x.148 Chapter 10 Miscellaneous Problems optimal control for the corresponding LQR problem. The constraint (10. C(zI −A)−1 T should be “small”. This yields the bound T −1 (zI − A)−1 B H ∞ ≤ 1/α.14) satisﬁes two competing constraints: First. the state-to-output transfer matrix.15). in order to minimize eﬀects of state quantization at the output. and second. C). B T P + RK = 0 and AT P1 + P1 A < Q. Copyright c 1994 by the Society for Industrial and Applied Mathematics. the input-to-state transfer matrix.) 10. This is an LMIP in P . (The condition involving P1 is equivalent to (Q.16) Our problem is then to compute T to minimize the noise power (10. | |z| > 1} . injected directly into the state. (10.7 System Realization Problems Consider the discrete-time minimal LTI system x(k + 1) = Ax(k) + Bu(k). (10. P1 . which is just η times the H2 norm of the state-to-output transfer matrix: pnoise = η C(zI − A)−1 T 2 .13) where x : Z+ → Rn .15) is equivalent to the existence of P > 0 such that AT P A − P + T −T T −1 BT P A AT P B B T P B − I/α2 ≤ 0. A) being detectable.15) where the H∞ norm of the transfer matrix H(z) is deﬁned as ∞ = sup { H(z) Next. The output noise power can be expressed as 2 T p2 noise = η Tr T Wo T. such that the new realization ¯ x(k + 1) = T −1 AT x(k) + T −1 Bu(k) ¯ y(k) = CT x(k).

P > 0. Input u has RMS value bounded componentwise Suppose the RMS value of each component ui is less than α (instead of the RMS value of u being less than α) and that the RMS value of the state is still required to be less than one. where Xopt is an optimal value of X in the GEVP. P and R): AT P A − P + X BT P A AT P B B T P B − R/α2 ≤ 0. Y > 0. (10. . so that the EVP (over X. ∞ (10. Y . Then. X > 0. (10.15) is the GEVP Therefore choosing T to minimize C(zI − A)−1 T minimize subject to γ P > 0. This electronic version is for personal use and may not be duplicated or distributed.17) This is a convex problem in X and P . the realization problem is: Minimize η 2 Tr Wo X −1 subject to X > 0 and the LMI (in P > 0 and X) AT P A − P + X BT P A AT P B B T P B − I/α2 ≤ 0.18) where R > 0 is diagonal.9. R > 0. Y I I X ≥0 Similar methods can be used to handle several variations and extensions. (10. The constraint C(zI − A)−1 T ∞ ≤ γ is equivalent to the LMI (over Y > 0 and X = T −T T −1 > 0): AT Y A − Y + C T C YA AT Y Y − X/α2 ≤ 0. we can measure the eﬀect of the quantization noise on the output with the H∞ norm. ≥0 Minimizing state-to-output H∞ norm Alternatively.17). an equivalent condition is the LMI (with variables X. P > 0. we can choose T to minimize C(zI − A)−1 T ∞ subject to the constraint (10.19) An optimal state coordinate transformation T is any matrix that satisﬁes X opt = (T T T )−1 . and can be transformed into the EVP minimize subject to η 2 Tr Y Wc (10.7 System Realization Problems 149 Wo is the solution of the Lyapunov equation AT Wo A − Wo + C T C = 0. With X = T −T T −1 . using the methods of §10.15).18). that is. (10. Y I I X X>0 Tr R = 1 R is diagonal. P > 0. with unit trace. P and R) is minimize subject to η 2 Tr Y Wc (10.10.19) subject to (10.17).

8 Multi-Criterion LQG Consider the LTI system given by x = Ax + Bu + w. τp . A) and (Q. Ri > 0. . y is the measured output. . we further assume that Q ≥ 0 and R > 0. AYlqg + Ylqg AT − Ylqg C T V −1 CYlqg + W = 0. . and U = Ylqg C T V −1 CYlqg . . p. . (10. A) are observable. we associate a cost function i Jlqg = lim E zi (t)T zi (t). B) is controllable. . . . This is a convex optimization problem. that (A. . Qi ≥ 0. For each zi . i=1 Copyright c 1994 by the Society for Industrial and Applied Mathematics. In the multi-criterion LQG problem. z is the exogenous output. . . . where Xlqg and Ylqg are the solutions of (10. A) is observable. The standard Linear-Quadratic Gaussian (LQG) problem is to minimize Jlqg = lim E z(t)T z(t) t→∞ over u. where Xlqg and Ylqg are the unique positive-deﬁnite solutions of the Riccati equations AT Xlqg + Xlqg A − Xlqg BR−1 B T Xlqg + Q = 0. The optimal cost is given by ∗ Jlqg = Tr (Xlqg U + QYlqg ) . i = 0. w and v are independent white noise signals with constant. t→∞ i = 0. Noting that Xlqg ≥ X for every X > 0 that satisﬁes AT X + XA − XBR−1 B T X + Q ≥ 0. z= Q1/2 0 0 R1/2 x u . . .150 Chapter 10 Miscellaneous Problems 10.20) 0 1/2 Ri x u . whose solution is given by maximizing p Tr (Xlqg U + QYlqg ) − γ i τi . p we conclude that the optimal cost is the maximum of p Tr XU + (Q0 + i=1 τi Qi )Ylqg − γ i τi . subject to the condition that u(t) is measurable on y(τ ) for τ ≤ t.20) with p p Q = Q0 + i=1 τi Q i and R = R0 + i=1 τi R i .21) 0 The multi-criterion LQG problem is to minimize Jlqg over u subject to the measurabili ity condition and the constraints Jlqg < γi . . p. we have several exogenous outputs of interest given by zi = Qi 0 1/2 (10. i=1 over nonnegative τ1 . positive-deﬁnite spectral density matrices W and V respectively. ˙ y = Cx + v. . . i = 1. p. We assume that (Q0 . where u is the control input. and that (C.

. p x(t) → 0 as t → ∞ p (10. are not necessarily positive-deﬁnite. is given by: sup inf {S(τ. we deﬁne p p S(τ. . subject to X > 0. The constrained optimal control problem is: minimize subject to p J0 . . p. where τ1 ≥ 0. the solution of problem (10. when it is not −∞. i = 0. .22) The solution to this problem proceeds as follows: We ﬁrst deﬁne Q = Q0 + i=1 τi Q i . with τ = [τ1 · · · τp ]. . we deﬁne a set of p + 1 cost indices J0 .9 Nonconvex Multi-Criterion Quadratic Problems In this section. . For any u. . where (A.23) ≥0 Therefore. C = C0 + i=1 τi C i . Jp by ∞ Ji (u) = 0 [xT uT ] Qi T Ci Ci Ri x u dt. . .9 Nonconvex Multi-Criterion Quadratic Problems 151 over X. Next. . u) = J0 + i=1 ∆ τi J i − τi γ i . . J i ≤ γi . . . is computed by solving the EVP in P = P T p maximize subject to x(0)T P x(0) − τi γ i i=1 AT P + P A + Q BT P + C p P B + CT R (10. the optimal control problem is solved by the EVP in P = P T and τ : maximize subject to x(0) P x(0) − T τi γ i i=1 AT P + P A + Q BT P + C P B + CT R ≥ 0. τp ≥ 0. Here the symmetric matrices Qi T Ci Ci Ri . This electronic version is for personal use and may not be duplicated or distributed. u). .10. u) | x(t) → 0 as t → ∞ } τ u (See the Notes and References. . 10. . . i = 0. . τ1 . p. . the inﬁmum over u of S(τ. Computing this is an EVP. x(0) = x0 . when it is not −∞. . . . . . . . we consider the LTI system x = Ax + Bu. . τp ≥ 0 and p −1 p A X + XA − XB T R0 + i=1 τi R i B X + Q0 + i=1 T τi Qi ≥ 0. R = R0 + i=1 τi R i . τp ≥ 0.) For any ﬁxed τ . .22). . τ1 ≥ 0. . i=1 Then. . i = 1. τp . . . B) is ˙ controllable. τ1 ≥ 0. p.

where µ denotes the structured singular value (see [Doy82]). FBB92]. we refer the reader to the Notes and References in Chapter 3. Skorodinskii [Sko90] observed that the problem of proving stability of a system with delays via Lyapunov–Krasovskii functionals is a convex problem.6) is derived from Theorems 3 and 4 in Willems [Wil71b]. It can be shown (see [Boy86]) that if the feedback system is stable for all ∆ which are diagonal. with ∆ii (s) = e−sTi . Thus. SL93b. He casts the problem as an EVP. In [Kav94] Kavrano˘lu considers the problem of approximating in the H∞ norm a given g transfer matrix by a constant matrix. .4]. LMI (10.e. verifying supω∈R µ(H(jω)) < 1 can be cast as checking stability of a linear system with L arbitrary. Stabilization of systems with time delays The introduction of Lyapunov functionals of the form (10. see e. our results extend theirs to the multi-input case. [Bit88. The problem of checking whether a given polyhedron is invariant as well as the associated state-feedback synthesis problem have been considered by many authors.152 Chapter 10 Miscellaneous Problems Notes and References Optimization over an aﬃne family of linear systems Optimization over an aﬃne family of transfer matrices arises in linear controller design via Youla’s parametrization of closed-loop transfer matrices.. A related paper is by Geromel [Ger85]. This system is stable for all such ∆ if and only if supω∈R µ(H(jω)) < 1. Diagonal quadratic Lyapunov functions also arise in the study of large-scale systems [MH78]. positive delays.9) is due to Krasovskii [Kra56]. Copyright c 1994 by the Society for Industrial and Applied Mathematics. It is interesting to note that the “structured stability problem” or “µ-analysis problem” can be cast in terms of linear systems with delays. Ti > 0 (i.. The study of positive orthant holdability draws heavily from the theory of positive matrices.g. Positive orthant stabilizability The problem of positive orthant stability and holdability was extensively studied by Berman. see Boyd and Barratt [BB91] and the references therein. references for which are [BNS89] and the book by Berman and Plemmons [BP79. using the approach of [CH93]. Stability of systems with LTI perturbations For references about this problem. delays) then the system is stable for all ∆ with ∆ ∞ ≤ 1. Other examples include ﬁnite-impulse-response (FIR) ﬁlter design (see for example [OKUI88]) and antenna array weight design [KRB89]. which is a problem of the form considered in §10. Consider a stable LTI system with transfer matrix H and L inputs and outputs. who gives a computational procedure to ﬁnd diagonal solutions of the Lyapunov equation.3 are easily generalized to the LDIs considered in Chapter 4. ch6]. HB91b] and references therein. A survey of results and applications of diagonal Lyapunov stability is given in Kaszkurewicz and Bhaya [KB93]. Neumann and Stern [BNS89. see also §3. Finally. Diagonal solutions to the Lyapunov equation play a central role in the positive orthant stabilizability and holdability problems.4. The LMI method of §10. A generalization of the concept of positive orthant stability is that of invariance of polyhedral sets. WBL92. §7.1. BBT90. we note that the results of §10. In particular we see that Krasovskii’s method from 1956 can be interpreted as a Lyapunov-based µ-analysis method. See also [PF92.3 can be extended to the polyhedral case. They solve the positive orthant holdability problem when the control input is scalar. connected in feedback with a diagonal transfer matrix ∆ with H∞ norm less than one.

Genin and Kamp discuss the role of the matrix NevanlinnaPick problem in circuit and systems theory [DGK81]. There are two matrix versions of this problem: the matrix Nevanlinna-Pick problem [DGK79] and the tangential Nevanlinna-Pick problem [Fed75]. Kimura reduces a class of H∞ -optimal controller design problems to the tangential Nevanlinna-Pick problem. who solved it when the control u is scalar. A number of researchers have studied the application of Nevanlinna-Pick theory to problems in system and control: Delsarte. see also the references cited there. Thi84]. see e. The book by Gevers and Li [GL93a] describes a number of digital ﬁlter realization problems. A comprehensive description of the Nevanlinna-Pick problem. a bound on the peak gain from input to state. see. Gohberg and Rodman [BGR90]. This can be used to guarantee no overﬂow given a maximum peak input level. they solve the problem when the control weighting matrix R is known.. its extensions and variations can be found in the book by Ball. study the implications of interpolation conditions in controller design. Rotea [Rot93] calls this problem “generalized H2 control”.e. We also note that an H∞ norm constraint on the input-to-state transfer matrix yields. ˆ This electronic version is for personal use and may not be duplicated or distributed. We conjecture that the LMI approach to system realization can also incorporate the constraint of stability..Notes and References Interpolation problems 153 The classical Nevanlinna-Pick problem dates back at least to 1916 [Pic16. Boyd and Doyle [BD87]. In [Kim87]. indirectly. It is discussed in great detail by Anderson and Moore [AM90. that the (nonlinear) system not exhibit limit cycles due to quantization or overﬂow. ˆ x ˆ dt u = K x. i. Multi-criterion LQG The multi-criterion LQG problem is discussed in [BB91].e. He shows that one can restrict attention to observer-based controllers such as d x = (A + BK)ˆ + Y C T V −1 (y − C x). Liu. System realization problems For a discussion of the problem of system realizations see [MR76.. Skelton and Grigoriadis discuss the problem of optimal ﬁnite word-length controller implementation [LSG92]. see also [Bal94]. also see [OF85]. which give analytical solutions via balanced realizations for special realization problems. Fujii and Narazaki [FN84] and the references therein. see also [And66c]. Chang and Pearson [CP84] solve a class of H∞ -optimal controller design problems using matrix Nevanlinna-Pick theory. e. Safonov [Saf86] reduces the design of controllers for systems with structured perturbations to the scaled matrix Nevanlinna-Pick problem. The inverse problem of optimal control This problem was ﬁrst considered by Kalman [Kal64]. See also Rotea and Williamson [RW94b. Nev19]. System identiﬁcation The system identiﬁcation problem considered in §10. RW94a]. §5. the most general form of the inverse optimal control problem.3 can be found in [CNF92]. Zames and Francis. Our result handles the case of unknown R. . [Wil74b]. by checking that the return diﬀerence inequality I − B T (−iωI − AT )−1 K R I − K T (iωI − A)−1 B ≥ R holds for all ω ∈ R. [PKT + 94].6]. Other system identiﬁcation problems can be cast in terms of LMI problems.5.g..g. in [ZF83]. i.

Many of the suﬃcient conditions for NLDIs in Chapters 5–7 can be shown to be necessary and suﬃcient conditions when we consider integral constraints on p and q. Nonconvex multi-criterion quadratic problems This section is based on Megretsky [Meg92a. Mak91]. This is illustrated in the following section. Yakubovich gives conditions under which the inﬁmum is actually a minimum (that is. which is covered in these articles.g.24) Now. Z and X minimize subject to Tr R0 X + Tr Q0 (P + Y ) X ZT Z P ≥ 0. i = 1. Meg93] and Yakubovich [Yak92]. i The value of each cost index Jlqg deﬁned by (10. . We also mention an algorithm devised by Zhu. ˙ ∞ 0 ∞ q = Cq x. taking Z = KP as a new variable. (10. . The covariance of the residual state x −x is Y and the covariance ˆ P of the s