ESSENTIALS OF

ROBUST CONTROL

Kemin Zhou

May 25, 1999

Preface

Robustness of control systems to disturbances and uncertainties has always been the
central issue in feedback control. Feedback would not be needed for most control systems
if there were no disturbances and uncertainties. Developing multivariable robust control
methods has been the focal point in the last two decades in the control community. The
state-of-the-art H∞ robust control theory is the result of this effort.
This book introduces some essentials of robust and H∞ control theory. It grew from
another book by this author, John C. Doyle, and Keith Glover, entitled Robust and
Optimal Control, which has been extensively class-tested in many universities around
the world. Unlike that book, which is intended primarily as a comprehensive reference of
robust and H∞ control theory, this book is intended to be a text for a graduate course
in multivariable control. It is also intended to be a reference for practicing control
engineers who are interested in applying the state-of-the-art robust control techniques
in their applications. With this objective in mind, I have streamlined the presentation,
added more than 50 illustrative examples, included many related Matlab R commands1
and more than 150 exercise problems, and added some recent developments in the area
of robust control such as gap metric, ν-gap metric, model validation, and mixed µ
problem. In addition, many proofs are completely rewritten and some advanced topics
are either deleted completely or do not get an in-depth treatment.
The prerequisite for reading this book is some basic knowledge of classical control
theory and state-space theory. The text contains more material than could be covered in
detail in a one-semester or a one-quarter course. Chapter 1 gives a chapter-by-chapter
summary of the main results presented in the book, which could be used as a guide for
the selection of topics for a specific course. Chapters 2 and 3 can be used as a refresher
for some linear algebra facts and some standard linear system theory. A course focusing
on H∞ control should cover at least most parts of Chapters 4–6, 8, 9, 11–13, and Sections
14.1 and 14.2. An advanced H∞ control course should also include the rest of Chapter
14, Chapter 16, and possibly Chapters 10, 7, and 15. A course focusing on robustness
and model uncertainty should cover at least Chapters 4, 5, and 8–10. Chapters 17 and
18 can be added to any advanced robust and H∞ control course if time permits.
I have tried hard to eliminate obvious mistakes. It is, however, impossible for me
to make the book perfect. Readers are encouraged to send corrections, comments, and
1 Matlab is a registered trademark of The MathWorks, Inc.

vii
viii PREFACE

suggestions to me, preferably by electronic mail, at

[email protected]
I am also planning to put any corrections, modifications, and extensions on the Internet
so that they can be obtained either from the following anonymous ftp:
ftp gate.ee.lsu.edu cd pub/kemin/books/essentials/
or from the author’s home page:
http://kilo.ee.lsu.edu/kemin/books/essentials/
This book would not be possible without the work done jointly for the previous
book with Professor John C. Doyle and Professor Keith Glover. I thank them for their
influence on my research and on this book. Their serious attitudes toward scientific
research have been reference models for me. I am especially grateful to John for having
me as a research fellow in Caltech, where I had two very enjoyable years and had
opportunities to catch a glimpse of his “BIG PICTURE” of control.
I want to thank my editor from Prentice Hall, Tom Robbins, who originally proposed
the idea for this book and has been a constant source of support for me while writing it.
Without his support and encouragement, this project would have been a difficult one.
It has been my great pleasure to work with him.
I would like to express my sincere gratitude to Professor Bruce A. Francis for giving
me many helpful comments and suggestions on this book. Professor Francis has also
kindly provided many exercises in the book. I am also grateful to Professor Kang-Zhi Liu
and Professor Zheng-Hua Luo, who have made many useful comments and suggestions.
I want to thank Professor Glen Vinnicombe for his generous help in the preparation of
Chapters 16 and 17. Special thanks go to Professor Jianqing Mao for providing me the
opportunity to present much of this material in a series of lectures at Beijing University
of Aeronautics and Astronautics in the summer of 1996.
In addition, I would like to thank all those who have helped in many ways in making
this book possible, especially Professor Pramod P. Khargonekar, Professor André Tits,
Professor Andrew Packard, Professor Jie Chen, Professor Jakob Stoustrup, Professor
Hans Henrik Niemann, Professor Malcolm Smith, Professor Tryphon Georgiou, Profes-
sor Tongwen Chen, Professor Hitay Özbay, Professor Gary Balas, Professor Carolyn
Beck, Professor Dennis S. Bernstein, Professor Mohamed Darouach, Dr. Bobby Boden-
heimer, Professor Guoxiang Gu, Dr. Weimin Lu, Dr. John Morris, Dr. Matt Newlin,
Professor Li Qiu, Professor Hector P. Rotstein, Professor Andrew Teel, Professor Ja-
gannathan Ramanujam, Dr. Linda G. Bushnell, Xiang Chen, Greg Salomon, Pablo A.
Parrilo, and many other people.
I would also like to thank the following agencies for supporting my research: National
Science Foundation, Army Research Office (ARO), Air Force of Scientific Research, and
the Board of Regents in the State of Louisiana.
Finally, I would like to thank my wife, Jing, and my son, Eric, for their generous
support, understanding, and patience during the writing of this book.

Kemin Zhou
PREFACE ix

Here is how H∞ is pronounced in Chinese:

It means “The joy of love is endless.”

Contents

Preface vii

Notation and Symbols xv

List of Acronyms xvii

1 Introduction 1
1.1 What Is This Book About? . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Highlights of This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Linear Algebra 11
2.1 Linear Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Matrix Inversion Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Invariant Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Vector Norms and Matrix Norms . . . . . . . . . . . . . . . . . . . . . . 16
2.6 Singular Value Decomposition . . . . . . . . . . . . . . . . . . . . . . . . 19
2.7 Semidefinite Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.8 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.9 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3 Linear Systems 27
3.1 Descriptions of Linear Dynamical Systems . . . . . . . . . . . . . . . . . 27
3.2 Controllability and Observability . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Observers and Observer-Based Controllers . . . . . . . . . . . . . . . . . 31
3.4 Operations on Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.5 State-Space Realizations for Transfer Matrices . . . . . . . . . . . . . . 35
3.6 Multivariable System Poles and Zeros . . . . . . . . . . . . . . . . . . . 38
3.7 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

xi
xii CONTENTS

4 H2 and H∞ Spaces 45
4.1 Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 H2 and H∞ Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3 Computing L2 and H2 Norms . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4 Computing L∞ and H∞ Norms . . . . . . . . . . . . . . . . . . . . . . . 55
4.5 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

5 Internal Stability 65
5.1 Feedback Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 Well-Posedness of Feedback Loop . . . . . . . . . . . . . . . . . . . . . . 66
5.3 Internal Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.4 Coprime Factorization over RH∞ . . . . . . . . . . . . . . . . . . . . . . 71
5.5 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

6 Performance Specifications and Limitations 81

6.1 Feedback Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2 Weighted H2 and H∞ Performance . . . . . . . . . . . . . . . . . . . . . 85
6.3 Selection of Weighting Functions . . . . . . . . . . . . . . . . . . . . . . 89
6.4 Bode’s Gain and Phase Relation . . . . . . . . . . . . . . . . . . . . . . 94
6.5 Bode’s Sensitivity Integral . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.6 Analyticity Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.7 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

7 Balanced Model Reduction 105

7.1 Lyapunov Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.2 Balanced Realizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.3 Model Reduction by Balanced Truncation . . . . . . . . . . . . . . . . . 117
7.4 Frequency-Weighted Balanced Model Reduction . . . . . . . . . . . . . . 124
7.5 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

8 Uncertainty and Robustness 129

8.1 Model Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
8.2 Small Gain Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.3 Stability under Unstructured Uncertainties . . . . . . . . . . . . . . . . 141
8.4 Robust Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
8.5 Skewed Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
8.6 Classical Control for MIMO Systems . . . . . . . . . . . . . . . . . . . . 154
8.7 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
8.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
CONTENTS xiii

9 Linear Fractional Transformation 165

9.1 Linear Fractional Transformations . . . . . . . . . . . . . . . . . . . . . 165
9.2 Basic Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
9.3 Redheffer Star Products . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
9.4 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
9.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

10 µ and µ Synthesis 183

10.1 General Framework for System Robustness . . . . . . . . . . . . . . . . 184
10.2 Structured Singular Value . . . . . . . . . . . . . . . . . . . . . . . . . . 187
10.3 Structured Robust Stability and Performance . . . . . . . . . . . . . . . 200
10.4 Overview of µ Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
10.5 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
10.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217

11 Controller Parameterization 221

11.1 Existence of Stabilizing Controllers . . . . . . . . . . . . . . . . . . . . . 222
11.2 Parameterization of All Stabilizing Controllers . . . . . . . . . . . . . . 224
11.3 Coprime Factorization Approach . . . . . . . . . . . . . . . . . . . . . . 228
11.4 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
11.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231

12 Algebraic Riccati Equations 233

12.1 Stabilizing Solution and Riccati Operator . . . . . . . . . . . . . . . . . 234
12.2 Inner Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
12.3 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
12.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246

13 H2 Optimal Control 253

13.1 Introduction to Regulator Problem . . . . . . . . . . . . . . . . . . . . . 253
13.2 Standard LQR Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
13.3 Extended LQR Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
13.4 Guaranteed Stability Margins of LQR . . . . . . . . . . . . . . . . . . . 259
13.5 Standard H2 Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
13.6 Stability Margins of H2 Controllers . . . . . . . . . . . . . . . . . . . . . 265
13.7 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
13.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267

14 H∞ Control 269
14.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
14.2 A Simplified H∞ Control Problem . . . . . . . . . . . . . . . . . . . . . 270
14.3 Optimality and Limiting Behavior . . . . . . . . . . . . . . . . . . . . . 282
14.4 Minimum Entropy Controller . . . . . . . . . . . . . . . . . . . . . . . . 286
14.5 An Optimal Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
xiv CONTENTS

14.6 General H∞ Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288

14.7 Relaxing Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
14.8 H2 and H∞ Integral Control . . . . . . . . . . . . . . . . . . . . . . . . 294
14.9 H∞ Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
14.10Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
14.11Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300

15 Controller Reduction 305

15.1 H∞ Controller Reductions . . . . . . . . . . . . . . . . . . . . . . . . . . 306
15.2 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
15.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313

16 H∞ Loop Shaping 315

16.1 Robust Stabilization of Coprime Factors . . . . . . . . . . . . . . . . . . 315
16.2 Loop-Shaping Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
16.3 Justification for H∞ Loop Shaping . . . . . . . . . . . . . . . . . . . . . 328
16.4 Further Guidelines for Loop Shaping . . . . . . . . . . . . . . . . . . . . 334
16.5 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
16.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342

17 Gap Metric and ν-Gap Metric 349

17.1 Gap Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
17.2 ν-Gap Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
17.3 Geometric Interpretation of ν-Gap Metric . . . . . . . . . . . . . . . . . 370
17.4 Extended Loop-Shaping Design . . . . . . . . . . . . . . . . . . . . . . . 373
17.5 Controller Order Reduction . . . . . . . . . . . . . . . . . . . . . . . . . 375
17.6 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
17.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375

18 Miscellaneous Topics 377

18.1 Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
18.2 Mixed µ Analysis and Synthesis . . . . . . . . . . . . . . . . . . . . . . . 381
18.3 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
18.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390

Bibliography 391

Index 407
Notation and Symbols

R and C fields of real and complex numbers

F field, either R or C
C− and C− open and closed left-half plane
C+ and C+ open and closed right-half plane
jR imaginary axis

∈ belong to
⊂ subset
∪ union
∩ intersection

2 end of proof
3 end of remark

:= defined as
' and / asymptotically greater and less than
 and  much greater and less than

α complex conjugate of α ∈ C
|α| absolute value of α ∈ C
Re(α) real part of α ∈ C

In n × n identity matrix
[aij ] a matrix with aij as its ith row and jth column element
diag(a1 , . . . , an ) an n × n diagonal matrix with ai as its ith diagonal element
AT and A∗ transpose and complex conjugate transpose of A
A−1 and A+ inverse and pseudoinverse of A
A−∗ shorthand for (A−1 )∗
det(A) determinant of A
trace(A) trace of A

xv
xvi NOTATION AND SYMBOLS

λ(A) eigenvalue of A
ρ(A) spectral radius of A
ρR (A) real spectrum radius of A
σ(A) and σ(A) the largest and the smallest singular values of A
σi (A) ith singular value of A
κ(A) condition number of A
kAk spectral norm of A: kAk = σ(A)
Im(A), R(A) image (or range) space of A
Ker(A), N(A) kernel (or null) space of A
X− (A) stable invariant subspace of A

Ric(H) the stabilizing solution of an ARE

g∗f convolution of g and f
∠ angle
h, i inner product
x⊥y orthogonal, hx, yi = 0
D⊥ orthogonal complement of D
S⊥ orthogonal complement of subspace S, e.g., H2⊥

L2 (−∞, ∞) time domain square integrable functions

L2+ := L2 [0, ∞) subspace of L2 (−∞, ∞) with functions zero for t < 0
L2− := L2 (−∞, 0] subspace of L2 (−∞, ∞) with functions zero for t > 0
L2 (jR) square integrable functions on C0 including at ∞
H2 subspace of L2 (jR) with functions analytic in Re(s) > 0
H2⊥ subspace of L2 (jR) with functions analytic in Re(s) < 0
L∞ (jR) functions bounded on Re(s) = 0 including at ∞
H∞ the set of L∞ (jR) functions analytic in Re(s) > 0
−
H∞ the set of L∞ (jR) functions analytic in Re(s) < 0

prefix B and Bo closed and open unit ball, e.g. B∆ and Bo ∆

prefix R real rational, e.g., RH∞ and RH2 , etc.

Rp (s) rational proper transfer matrices

∼
G
 (s)  shorthand for GT (−s)
A B
shorthand for state space realization C(sI − A)−1 B + D
C D
η(G(s)) number of right-half plane poles
η0 (G(s)) number of imaginary axis poles
wno(G) winding number

F` (M, Q) lower LFT

Fu (M, Q) upper LFT
M ?N star product
List of Acronyms

ARE algebraic Riccati equation

FDLTI finite dimensional linear time invariant
iff if and only if
lcf left coprime factorization
LFT linear fractional transformation
lhp or LHP left-half plane Re(s) < 0
LQG linear quadratic Gaussian
LTI linear time invariant
MIMO multi-input multioutput
nlcf normalized left coprime factorization
NP nominal performance
nrcf normalized right coprime factorization
NS nominal stability
rcf right coprime factorization
rhp or RHP right-half plane Re(s) > 0
RP robust performance
RS robust stability
SISO single-input single-output
SSV structured singular value (µ)
SVD singular value decomposition

xvii
xviii LIST OF ACRONYMS
Chapter 1

Introduction

This chapter gives a brief description of the problems considered in this book and the
key results presented in each chapter.

1.1 What Is This Book About?

This book is about basic robust and H∞ control theory. We consider a control system
with possibly multiple sources of uncertainties, noises, and disturbances as shown in
Figure 1.1.

uncertainty
disturbance
other controlled signals

uncertainty System Interconnection uncertainty

tracking errors noise

controller
reference signals

Figure 1.1: General system interconnection

1
2 INTRODUCTION

We consider mainly two types of problems:

• Analysis problems: Given a controller, determine if the controlled signals (in-

cluding tracking errors, control signals, etc.) satisfy the desired properties for all
admissible noises, disturbances, and model uncertainties.

• Synthesis problems: Design a controller so that the controlled signals satisfy the
desired properties for all admissible noises, disturbances, and model uncertainties.

Most of our analysis and synthesis will be done on a unified linear fractional transforma-
tion (LFT) framework. To that end, we shall show that the system shown in Figure 1.1
can be put in the general diagram in Figure 1.2, where P is the interconnection matrix,
K is the controller, ∆ is the set of all possible uncertainty, w is a vector signal including
noises, disturbances, and reference signals, z is a vector signal including all controlled
signals and tracking errors, u is the control signal, and y is the measurement.

- ∆
v η

z  w
P


y - K u

Figure 1.2: General LFT framework

The block diagram in Figure 1.2 represents the following equations:

   
v η
 z  = P w 
y u
η = ∆v
u = Ky.

Let the transfer matrix from w to z be denoted by Tzw and assume that the ad-
missible uncertainty ∆ satisfies k∆k∞ < 1/γu for some γu > 0. Then our analy-
sis problem is to answer if the closed-loop system is stable for all admissible ∆ and
kTzw k∞ ≤ γp for some prespecified γp > 0, where kTzw k∞ is the H∞ norm defined as
kTzw k∞ = supω σ̄ (Tzw (jω)). The synthesis problem is to design a controller K so that
the aforementioned robust stability and performance conditions are satisfied.
In the simplest form, we have either ∆ = 0 or w = 0. The former becomes the well-
known H∞ control problem and the later becomes the robust stability problem. The two
1.2. Highlights of This Book 3

problems are equivalent when ∆ is a single-block unstructured uncertainty through the

application of the small gain theorem (see Chapter 8). This robust stability consequence
was probably the main motivation for the development of H∞ methods.
The analysis and synthesis for systems with multiple-block ∆ can be reduced in most
cases to an equivalent H∞ problem with suitable scalings. Thus a solution to the H∞
control problem is the key to all robustness problems considered in this book. In the
next section, we shall give a chapter-by-chapter summary of the main results presented
in this book.
We refer readers to the book Robust and Optimal Control by K. Zhou, J. C. Doyle,
and K. Glover [1996] for a brief historical review of H∞ and robust control and for
some detailed treatment of some advanced topics.

1.2 Highlights of This Book

The key results in each chapter are highlighted in this section. Readers should consult
the corresponding chapters for the exact statements and conditions.

Chapter 2 reviews some basic linear algebra facts.

Chapter 3 reviews system theoretical concepts: controllability, observability, sta-

bilizability, detectability, pole placement, observer theory, system poles and zeros, and
state-space realizations.

Chapter 4 introduces the H2 spaces and the H∞ spaces. State-space methods of

computing real rational H2 and H∞ transfer matrix norms are presented. For example,
let  
A B
G(s) = ∈ RH∞ .
C 0
Then
kGk22 = trace(B ∗ QB) = trace(CP C ∗ )
and
kGk∞ = max{γ : H has an eigenvalue on the imaginary axis},
where P and Q are the controllability and observability Gramians and
 
A BB ∗ /γ 2
H= .
−C ∗ C −A∗
4 INTRODUCTION

Chapter 5 introduces the feedback structure and discusses its stability.

w1 e1
-e - P
+ 6
+

e2 + w2
?
+
K̂  e

We define that the above closed-loop system is internally stable if and only if
 −1  
I −K̂ (I − K̂P )−1 K̂(I − P K̂)−1
= ∈ RH∞ .
−P I P (I − K̂P )−1 (I − P K̂)−1

Alternative characterizations of internal stability using coprime factorizations are also

presented.

Chapter 6 considers the feedback system properties and design limitations. The
formulations of optimal H2 and H∞ control problems and the selection of weighting
functions are also considered in this chapter.

Chapter 7 considers the problem of reducing the order of a linear multivariable

dynamical system using the balanced truncation method. Suppose
 
A11 A12 B1
G(s) =  A21 A22 B2  ∈ RH∞
C1 C2 D

is a balanced realization with controllability and observability Gramians P = Q = Σ =

diag(Σ1 , Σ2 )

Σ1= diag(σ1 Is1 , σ2 Is2 , . . . , σr Isr )

Σ2= diag(σr+1 Isr+1 , σr+2 Isr+2 , . . . , σN IsN ).
 
A11 B1
Then the truncated system Gr (s) = is stable and satisfies an additive
C1 D
error bound:
XN
kG(s) − Gr (s)k∞ ≤ 2 σi .
i=r+1

Frequency-weighted balanced truncation method is also discussed.

Chapter 8 derives robust stability tests for systems under various modeling assump-
tions through the use of the small gain theorem. In particular, we show that a system,
shown at the top of the following page, with an unstructured uncertainty ∆ ∈ RH∞
1.2. Highlights of This Book 5

with k∆k∞ < 1 is robustly stable if and only if kTzw k∞ ≤ 1, where Tzw is the matrix
transfer function from w to z.

∆
z w

nominal system

Chapter 9 introduces the LFT in detail. We show that many control problems
can be formulated and treated in the LFT framework. In particular, we show that
every analysis problem can be put in an LFT form with some structured ∆(s) and some
interconnection matrix M (s) and every synthesis problem can be put in an LFT form
with a generalized plant G(s) and a controller K(s) to be designed.

z w
- ∆  
G 
 y u
z w
 M  -K

Chapter 10 considers robust stability and performance for systems with multiple
sources of uncertainties. We show that an uncertain system is robustly stable and
satisfies some H∞ performance criterion for all ∆i ∈ RH∞ with k∆i k∞ < 1 if and only
if the structured singular value (µ) of the corresponding interconnection model is no
greater than 1.

∆1

∆4 nominal system ∆2

∆3
6 INTRODUCTION

Chapter 11 characterizes in state-space all controllers that stabilize a given dy-

namical system G(s). For a given generalized plant
 
  A B1 B2
G11 (s) G12 (s)
G(s) = =  C1 D11 D12 
G21 (s) G22 (s)
C2 D21 D22
we show that all stabilizing controllers can be parameterized as the transfer matrix from
y to u below where F and L are such that A + LC2 and A + B2 F are stable and where
Q is any stable proper transfer matrix.

 z  w
G 
y u

D22 
R
c?
 c?
 C2   c c B2  c
− 66 6
- A

- F
- −L

u1 y1
- Q

Chapter 12 studies the stabilizing solution to an algebraic Riccati equation (ARE).

A solution to the following ARE

A∗ X + XA + XRX + Q = 0

is said to be a stabilizing solution if A + RX is stable. Now let

 
A R
H :=
−Q −A∗

and let X− (H) be the stable H invariant subspace and

 
X1
X− (H) = Im ,
X2

where X1 , X2 ∈ Cn×n . If X1 is nonsingular, then X := X2 X1−1 is uniquely determined

by H, denoted by X = Ric(H). A key result of this chapter is the so-called bounded
1.2. Highlights of This Book 7

real lemma, which states that a stable transfer matrix G(s) satisfies kG(s)k∞ < γ if
and only if there exists an X such that A + BB ∗ X/γ 2 is stable and

XA + A∗ X + XBB ∗ X/γ 2 + C ∗ C = 0.

The H∞ control theory in Chapter 14 will be derived based on this lemma.

Chapter 13 treats the optimal control of linear time-invariant systems with quadratic
performance criteria (i.e., H2 problems). We consider a dynamical system described by
an LFT with  
A B1 B2
G(s) =  C1 0 D12  .
C2 D21 0

z w
G

y u
- K

Define
∗ ∗
R1 = D12 D12 > 0, R2 = D21 D21 >0
 
A − B2 R1−1 D12
∗
C1 −B2 R1−1 B2∗
H2 := −1 ∗
∗
−C1 (I − D12 R1 D12 )C1 −(A − B2 R1−1 D12
∗
C1 )∗
 
∗
(A − B1 D21 R2−1 C2 )∗ −C2∗ R2−1 C2
J2 :=
−B1 (I − D21 R2−1 D21 )B1∗
∗ ∗
−(A − B1 D21 R2−1 C2 )
X2 := Ric(H2 ) ≥ 0, Y2 := Ric(J2 ) ≥ 0
F2 := −R1−1 (B2∗ X2 + ∗
D12 C1 ), L2 := −(Y2 C2∗ + B1 D21
∗
)R2−1 .
Then the H2 optimal controller (i.e., the controller that minimizes kTzw k2 ) is given by
 
A + B2 F2 + L2 C2 −L2
Kopt (s) := .
F2 0

Chapter 14 first considers an H∞ control problem with the generalized plant G(s)
as given in Chapter 13 but with some additional simplifications: R1 = I, R2 = I,
∗ ∗
D12 C1 = 0, and B1 D21 = 0. We show that there exists an admissible controller such
that kTzw k∞ < γ if and only if the following three conditions hold:
(i) H∞ ∈ dom(Ric) and X∞ := Ric(H∞ ) > 0, where
 
A γ −2 B1 B1∗ − B2 B2∗
H∞ := ;
−C1∗ C1 −A∗
8 INTRODUCTION

(ii) J∞ ∈ dom(Ric) and Y∞ := Ric(J∞ ) > 0, where

 
A∗ γ −2 C1∗ C1 − C2∗ C2
J∞ := ;
−B1 B1∗ −A

(iii) ρ(X∞ Y∞ ) < γ 2 .

Moreover, an admissible controller such that kTzw k∞ < γ is given by

 
Â∞ −Z∞ L∞
Ksub =
F∞ 0

where
Â∞ := A + γ −2 B1 B1∗ X∞ + B2 F∞ + Z∞ L∞ C2
F∞ := −B2∗ X∞ , L∞ := −Y∞ C2∗ , Z∞ := (I − γ −2 Y∞ X∞ )−1 .
We then consider further the general H∞ control problem. We indicate how various
assumptions can be relaxed to accommodate other more complicated problems, such
as singular control problems. We also consider the integral control in the H2 and H∞
theory and show how the general H∞ solution can be used to solve the H∞ filtering
problem.

Chapter 15 considers the design of reduced-order controllers by means of controller

reduction. Special attention is paid to the controller reduction methods that preserve
the closed-loop stability and performance. Methods are presented that give sufficient
conditions in terms of frequency-weighted model reduction.

Chapter 16 first solves a special H∞ minimization problem. Let P = M̃ −1 Ñ be a

normalized left coprime factorization. Then we show that
 
K  
inf (I + P K)−1
I P
K stabilizing I
∞
  q −1
K   2
= inf −1 −1
(I + P K) M̃ = 1 − Ñ M̃ H .
K stabilizing I ∞

This implies that there is a robustly stabilizing controller for

˜ M )−1 (Ñ + ∆
P∆ = (M̃ + ∆ ˜ N)

with  
∆˜N ˜M
∆ <
∞
if and only if q   2
≤ 1 − Ñ M̃ .
H
1.3. Notes and References 9

Using this stabilization result, a loop-shaping design technique is proposed. The pro-
posed technique uses only the basic concept of loop-shaping methods, and then a robust
stabilization controller for the normalized coprime factor perturbed system is used to
construct the final controller.

Chapter 17 introduces the gap metric and the ν-gap metric. The frequency domain
interpretation and applications of the ν-gap metric are discussed. The controller order
reduction in the gap or ν-gap metric framework is also considered.

Chapter 18 considers briefly the problems of model validation and the mixed real
and complex µ analysis and synthesis.

Most computations and examples in this book are done using Matlab. Since we
shall use Matlab as a major computational tool, it is assumed that readers have some
basic working knowledge of the Matlab operations (for example, how to input vec-
tors and matrices). We have also included in this book some brief explanations of
Matlab, Simulink R , Control System Toolbox, and µ Analysis and Synthesis Toolbox1
commands. In particular, this book is written consistently with the µ Analysis and
Synthesis Toolbox. (Robust Control Toolbox, LMI Control Toolbox, and other soft-
ware packages may equally be used with this book.) Thus it is helpful for readers to
have access to this toolbox. It is suggested at this point to try the following demo
programs from this toolbox.

 msdemo1

 msdemo2

We shall introduce many more Matlab commands in the subsequent chapters.

1.3 Notes and References

The original formulation of the H∞ control problem can be found in Zames [1981].
Relations between H∞ have now been established with many other topics in control:
for example, risk-sensitive control of Whittle [1990]; differential games (see Başar and
Bernhard [1991], Limebeer, Anderson, Khargonekar, and Green [1992]; Green and Lime-
beer [1995]); chain-scattering representation, and J-lossless factorization (Green [1992]
and Kimura [1997]). See also Zhou, Doyle, and Glover [1996] for additional discussions
and references. The state-space theory of H∞ has also been carried much further, by
generalizing time invariant to time varying, infinite horizon to finite horizon, and finite
dimensional to infinite dimensional, and even to some nonlinear settings.
1 Simulink is a registered trademark of The MathWorks, Inc.; µ-Analysis and Synthesis is a trade-

mark of The MathWorks, Inc. and MUSYN Inc.; Control System Toolbox, Robust Control Toolbox,
and LMI Control Toolbox are trademarks of The MathWorks, Inc.
10 INTRODUCTION

1.4 Problems
Problem 1.1 We shall solve an easy problem first. When you read a paper or a book,
you often come across a statement like this “It is easy ...”. What the author really
meant was one of the following: (a) it is really easy; (b) it seems to be easy; (c) it is
easy for an expert; (d) the author does not know how to show it but he or she thinks it
is correct. Now prove that when I say “It is easy” in this book, I mean it is really easy.
(Hint: If you can prove it after you read the whole book, ask your boss for a promotion.
If you cannot prove it after you read the whole book, trash the book and write a book
yourself. Remember use something like “it is easy ...” if you are not sure what you are
talking about.)
Chapter 2

Linear Algebra

Some basic linear algebra facts will be reviewed in this chapter. The detailed treatment
of this topic can be found in the references listed at the end of the chapter. Hence we shall
omit most proofs and provide proofs only for those results that either cannot be easily
found in the standard linear algebra textbooks or are insightful to the understanding of
some related problems.

2.1 Linear Subspaces

Let R denote the real scalar field and C the complex scalar field. For the interest of this
chapter, let F be either R or C and let Fn be the vector space over F (i.e., Fn is either Rn
or Cn ). Now let x1 , x2 , . . . , xk ∈ Fn . Then an element of the form α1 x1 +. . .+αk xk with
αi ∈ F is a linear combination over F of x1 , . . . , xk . The set of all linear combinations
of x1 , x2 , . . . , xk ∈ Fn is a subspace called the span of x1 , x2 , . . . , xk , denoted by

span{x1 , x2 , . . . , xk } := {x = α1 x1 + . . . + αk xk : αi ∈ F}.

A set of vectors x1 , x2 , . . . , xk ∈ Fn is said to be linearly dependent over F if there

exists α1 , . . . , αk ∈ F not all zero such that α1 x2 + . . . + αk xk = 0; otherwise the vectors
are said to be linearly independent.
Let S be a subspace of Fn , then a set of vectors {x1 , x2 , . . . , xk } ∈ S is called a basis
for S if x1 , x2 , . . . , xk are linearly independent and S = span{x1 , x2 , . . . , xk }. However,
such a basis for a subspace S is not unique but all bases for S have the same number
of elements. This number is called the dimension of S, denoted by dim(S).
A set of vectors {x1 , x2 , . . . , xk } in Fn is mutually orthogonal if x∗i xj = 0 for all
i 6= j and orthonormal if x∗i xj = δij , where the superscript ∗ denotes complex conjugate
transpose and δij is the Kronecker delta function with δij = 1 for i = j and δij = 0
for i 6= j. More generally, a collection of subspaces S1 , S2 , . . . , Sk of Fn is mutually
orthogonal if x∗ y = 0 whenever x ∈ Si and y ∈ Sj for i 6= j.

11
12 LINEAR ALGEBRA

The orthogonal complement of a subspace S ⊂ Fn is defined by

S ⊥ := {y ∈ Fn : y ∗ x = 0 for all x ∈ S}.

We call a set of vectors {u1 , u2 , . . . , uk } an orthonormal basis for a subspace S ∈ Fn if

the vectors form a basis of S and are orthonormal. It is always possible to extend such
a basis to a full orthonormal basis {u1 , u2 , . . . , un } for Fn . Note that in this case

S ⊥ = span{uk+1 , . . . , un },

and {uk+1 , . . . , un } is called an orthonormal completion of {u1 , u2 , . . . , uk }.

Let A ∈ Fm×n be a linear transformation from Fn to Fm ; that is,

A : Fn 7−→ Fm .

Then the kernel or null space of the linear transformation A is defined by

KerA = N (A) := {x ∈ Fn : Ax = 0},

and the image or range of A is

ImA = R(A) := {y ∈ Fm : y = Ax, x ∈ Fn }.

Let ai , i = 1, 2, . . . , n denote the columns of a matrix A ∈ Fm×n ; then

ImA = span{a1 , a2 , . . . , an }.

A square matrix U ∈ F n×n whose columns form an orthonormal basis for Fn is called
a unitary matrix (or orthogonal matrix if F = R), and it satisfies U ∗ U = I = U U ∗ .
Now let A = [aij ] ∈ Cn×n ; then the trace of A is defined as
X
n
trace(A) := aii .
i=1

Illustrative MATLAB Commands:

 basis of KerA = null(A); basis of ImA = orth(A); rank of A = rank(A);

2.2 Eigenvalues and Eigenvectors

Let A ∈ Cn×n ; then the eigenvalues of A are the n roots of its characteristic polynomial
p(λ) = det(λI − A). The maximal modulus of the eigenvalues is called the spectral
radius, denoted by
ρ(A) := max |λi |
1≤i≤n

if λi is a root of p(λ), where, as usual, | · | denotes the magnitude. The real spectral
radius of a matrix A, denoted by ρR (A), is the maximum modulus of the real eigenvalues
2.3. Matrix Inversion Formulas 13

of A; that is, ρR (A) := max |λi | and ρR (A) := 0 if A has no real eigenvalues. A nonzero
λi ∈R
vector x ∈ Cn that satisfies
Ax = λx
is referred to as a right eigenvector of A. Dually, a nonzero vector y is called a left
eigenvector of A if
y ∗ A = λy ∗ .
In general, eigenvalues need not be real, and neither do their corresponding eigenvectors.
However, if A is real and λ is a real eigenvalue of A, then there is a real eigenvector
corresponding to λ. In the case that all eigenvalues of a matrix A are real, we will
denote λmax (A) for the largest eigenvalue of A and λmin (A) for the smallest eigenvalue.
In particular, if A is a Hermitian matrix (i.e., A = A∗ ), then there exist a unitary matrix
U and a real diagonal matrix Λ such that A = U ΛU ∗ , where the diagonal elements of
Λ are the eigenvalues of A and the columns of U are the eigenvectors of A.

Lemma 2.1 Consider the Sylvester equation

AX + XB = C, (2.1)

where A ∈ Fn×n , B ∈ Fm×m , and C ∈ Fn×m are given matrices. There exists a
unique solution X ∈ Fn×m if and only if λi (A) + λj (B) 6= 0, ∀i = 1, 2, . . . , n, and
j = 1, 2, . . . , m.
In particular, if B = A∗ , equation (2.1) is called the Lyapunov equation; and the
necessary and sufficient condition for the existence of a unique solution is that
λi (A) + λ̄j (A) 6= 0, ∀i, j = 1, 2, . . . , n.

Illustrative MATLAB Commands:

 [V, D] = eig(A) % AV = V D

 X=lyap(A,B,-C) % solving Sylvester equation.

2.3 Matrix Inversion Formulas

Let A be a square matrix partitioned as follows:
 
A11 A12
A := ,
A21 A22

where A11 and A22 are also square matrices. Now suppose A11 is nonsingular; then A
has the following decomposition:
     
A11 A12 I 0 A11 0 I A−111 A12
=
A21 A22 A21 A−1
11 I 0 ∆ 0 I
14 LINEAR ALGEBRA

with ∆ := A22 − A21 A−1

11 A12 , and A is nonsingular iff ∆ is nonsingular. Dually, if A22
is nonsingular, then
     
A11 A12 I A12 A−122
ˆ
∆ 0 I 0
=
A21 A22 0 I 0 A22 A−1
22 A21 I

with ∆ˆ := A11 − A12 A−1 A21 , and A is nonsingular iff ∆

ˆ is nonsingular. The matrix ∆
22
ˆ
(∆) is called the Schur complement of A11 (A22 ) in A.
Moreover, if A is nonsingular, then
 −1  −1 
A11 A12 A11 + A−1
11 A12 ∆
−1
A21 A−1
11 −A−1
11 A12 ∆
−1
= −1 −1 −1
A21 A22 −∆ A21 A11 ∆
and  −1  
A11 A12 ∆ˆ −1 −∆ˆ −1 A12 A−1
= 22 .
−1 ˆ −1 −1 −1 ˆ −1 A12 A−1
A21 A22 −A22 A21 ∆ A22 + A22 A21 ∆ 22

The preceding matrix inversion formulas are particularly simple if A is block trian-
gular:
 −1  
A11 0 A−1
11 0
=
A21 A22 −A−1 −1
22 A21 A11 A−1
22
 −1  −1 −1 −1

A11 A12 A11 −A11 A12 A22
= .
0 A22 0 A−1
22

The following identity is also very useful. Suppose A11 and A22 are both nonsingular
matrices; then
(A11 − A12 A−1
22 A21 )
−1
= A−1 −1 −1
11 + A11 A12 (A22 − A21 A11 A12 )
−1
A21 A−1
11 .

As a consequence of the matrix decomposition formulas mentioned previously, we

can calculate the determinant of a matrix by using its submatrices. Suppose A11 is
nonsingular; then
det A = det A11 det(A22 − A21 A−1
11 A12 ).
On the other hand, if A22 is nonsingular, then
det A = det A22 det(A11 − A12 A−1
22 A21 ).

In particular, for any B ∈ Cm×n and C ∈ Cn×m , we have

 
Im B
det = det(In + CB) = det(Im + BC)
−C In
and for x, y ∈ Cn
det(In + xy ∗ ) = 1 + y ∗ x.

Related MATLAB Commands: inv, det

2.4. Invariant Subspaces 15

2.4 Invariant Subspaces

Let A : Cn 7−→ Cn be a linear transformation, λ be an eigenvalue of A, and x be a
corresponding eigenvector, respectively. Then Ax = λx and A(αx) = λ(αx) for any
α ∈ C. Clearly, the eigenvector x defines a one-dimensional subspace that is invariant
with respect to premultiplication by A since Ak x = λk x, ∀k. In general, a subspace
S ⊂ Cn is called invariant for the transformation A, or A-invariant, if Ax ∈ S for every
x ∈ S. In other words, that S is invariant for A means that the image of S under A
is contained in S: AS ⊂ S. For example, {0}, Cn , KerA, and ImA are all A-invariant
subspaces.
As a generalization of the one-dimensional invariant subspace induced by an eigen-
vector, let λ1 , . . . , λk be eigenvalues of A (not necessarily distinct), and let xi be the cor-
responding eigenvectors and the generalized eigenvectors. Then S = span{x1 , . . . , xk }
is an A-invariant subspace provided that all the lower-rank generalized eigenvectors
are included. More specifically, let λ1 = λ2 = · · · = λl be eigenvalues of A, and
let x1 , x2 , . . . , xl be the corresponding eigenvector and the generalized eigenvectors ob-
tained through the following equations:

(A − λ1 I)x1 = 0
(A − λ1 I)x2 = x1
..
.
(A − λ1 I)xl = xl−1 .

Then a subspace S with xt ∈ S for some t ≤ l is an A-invariant subspace only if all lower-
rank eigenvectors and generalized eigenvectors of xt are in S (i.e., xi ∈ S, ∀1 ≤ i ≤ t).
This will be further illustrated in Example 2.1.
On the other hand, if S is a nontrivial subspace1 and is A-invariant, then there is
x ∈ S and λ such that Ax = λx.
An A-invariant subspace S ⊂ Cn is called a stable invariant subspace if all the
eigenvalues of A constrained to S have negative real parts. Stable invariant subspaces
will play an important role in computing the stabilizing solutions to the algebraic Riccati
equations in Chapter 12.

Example 2.1 Suppose a matrix A has the following Jordan canonical form:
 
λ1 1
    λ1 
A x1 x2 x3 x4 = x1 x2 x3 x4  


λ3
λ4
1 We will say subspace S is trivial if S = {0}.
16 LINEAR ALGEBRA

with Reλ1 < 0, λ3 < 0, and λ4 > 0. Then it is easy to verify that
S1 = span{x1 } S12 = span{x1 , x2 } S123 = span{x1 , x2 , x3 }
S3 = span{x3 } S13 = span{x1 , x3 } S124 = span{x1 , x2 , x4 }
S4 = span{x4 } S14 = span{x1 , x4 } S34 = span{x3 , x4 }
are all A-invariant subspaces. Moreover, S1 , S3 , S12 , S13 , and S123 are stable A-invariant
subspaces. The subspaces S2 = span{x2 }, S23 = span{x2 , x3 }, S24 = span{x2 , x4 }, and
S234 = span{x2 , x3 , x4 } are, however, not A-invariant subspaces since the lower-rank
eigenvector x1 is not in these subspaces. To illustrate, consider the subspace S23 . Then
by definition, Ax2 ∈ S23 if it is an A-invariant subspace. Since
Ax2 = λx2 + x1 ,
Ax2 ∈ S23 would require that x1 be a linear combination of x2 and x3 , but this is
impossible since x1 is independent of x2 and x3 .

2.5 Vector Norms and Matrix Norms

In this section, we shall define vector and matrix norms. Let X be a vector space.
A real-valued function k·k defined on X is said to be a norm on X if it satisfies the
following properties:
(i) kxk ≥ 0 (positivity);
(ii) kxk = 0 if and only if x = 0 (positive definiteness);
(iii) kαxk = |α| kxk, for any scalar α (homogeneity);
(iv) kx + yk ≤ kxk + kyk (triangle inequality)
for any x ∈ X and y ∈ X.
Let x ∈ Cn . Then we define the vector p-norm of x as
!1/p
Xn
kxkp := |xi |p , for 1 ≤ p < ∞.
i=1

In particular, when p = 1, 2, ∞ we have

X
n
kxk1 := |xi |;
i=1
v
u n
uX
kxk2 := t |xi |2 ;
i=1
2.5. Vector Norms and Matrix Norms 17

kxk∞ := max |xi |.

1≤i≤n

Clearly, norm is an abstraction and extension of our usual concept of length in three-
dimensional Euclidean space. So a norm of a vector is a measure of the vector “length”
(for example, kxk2 is the Euclidean distance of the vector x from the origin). Similarly,
we can introduce some kind of measure for a matrix.
Let A = [aij ] ∈ Cm×n ; then the matrix norm induced by a vector p-norm is defined
as
kAxkp
kAkp := sup .
x6=0 kxkp

The matrix norms induced by vector p-norms are sometimes called induced p-norms.
This is because kAkp is defined by or induced from a vector p-norm. In fact, A can
be viewed as a mapping from a vector space Cn equipped with a vector norm k·kp to
another vector space Cm equipped with a vector norm k·kp . So from a system theoretical
point of view, the induced norms have the interpretation of input/output amplification
gains.
In particular, the induced matrix 2-norm can be computed as
p
kAk2 = λmax (A∗ A).

We shall adopt the following convention throughout this book for the vector and
matrix norms unless specified otherwise: Let x ∈ Cn and A ∈ Cm×n ; then we shall
denote the Euclidean 2-norm of x simply by

kxk := kxk2

and the induced 2-norm of A by

kAk := kAk2 .

The Euclidean 2-norm has some very nice properties:

Lemma 2.2 Let x ∈ Fn and y ∈ Fm .

1. Suppose n ≥ m. Then kxk = kyk iff there is a matrix U ∈ Fn×m such that x = U y
and U ∗ U = I.

2. Suppose n = m. Then |x∗ y| ≤ kxk kyk. Moreover, the equality holds iff x = αy
for some α ∈ F or y = 0.

3. kxk ≤ kyk iff there is a matrix ∆ ∈ Fn×m with k∆k ≤ 1 such that x = ∆y.
Furthermore, kxk < kyk iff k∆k < 1.

4. kU xk = kxk for any appropriately dimensioned unitary matrices U .

18 LINEAR ALGEBRA

Another often used matrix norm is the so called Frobenius norm. It is defined as
v
uX
p um X n
kAkF := trace(A∗ A) = t |aij |2 .
i=1 j=1

However, the Frobenius norm is not an induced norm.

The following properties of matrix norms are easy to show:

Lemma 2.3 Let A and B be any matrices with appropriate dimensions. Then

1. ρ(A) ≤ kAk (this is also true for the F -norm and any induced matrix norm).

2. kABk ≤ kAk kBk. In particular, this gives A−1 ≥ kAk
−1
if A is invertible.
(This is also true for any induced matrix norm.)

3. kU AV k = kAk, and kU AV kF = kAkF , for any appropriately dimensioned unitary

matrices U and V .

4. kABkF ≤ kAk kBkF and kABkF ≤ kBk kAkF .

Note that although premultiplication or postmultiplication of a unitary matrix on a

matrix does not change its induced 2-norm and F -norm, it does change its eigenvalues.
For example, let
 
1 0
A= .
1 0

Then λ1 (A) = 1, λ2 (A) = 0. Now let

" #
√1 √1
U= 2 2 ;
− √12 √1
2

then U is a unitary matrix and

 √ 
2 0
UA =
0 0
√
with λ1 (U A) = 2, λ2 (U A) = 0. This property is useful in some matrix perturbation
problems, particularly in the computation of bounds for structured singular values,
which will be studied in Chapter 9.

Related MATLAB Commands: norm, normest

2.6. Singular Value Decomposition 19

2.6 Singular Value Decomposition

A very useful tool in matrix analysis is singular value decomposition (SVD). It will be
seen that singular values of a matrix are good measures of the “size” of the matrix and
that the corresponding singular vectors are good indications of strong/weak input or
output directions.
Theorem 2.4 Let A ∈ Fm×n . There exist unitary matrices
U = [u1 , u2 , . . . , um ] ∈ Fm×m
V = [v1 , v2 , . . . , vn ] ∈ Fn×n
such that  
∗ Σ1 0
A = U ΣV , Σ= ,
0 0
where  
σ1 0 ··· 0
 0 σ2 ··· 0 
 
Σ1 =  .. .. .. .. 
 . . . . 
0 0 ··· σp
and
σ1 ≥ σ2 ≥ · · · ≥ σp ≥ 0, p = min{m, n}.

Proof. Let σ = kAk and without loss of generality assume m ≥ n. Then, from the
definition of kAk, there exists a z ∈ Fn such that
kAzk = σ kzk .
By Lemma 2.2, there is a matrix Ũ ∈ F m×n such that Ũ ∗ Ũ = I and
Az = σŨ z.
Now let
z Ũ z
x= ∈ Fn , y=
∈F .
m
kzk Ũ z
We have Ax = σy. Let  
V = x V1 ∈ Fn×n
and  
U= y U1 ∈ Fm×m
be unitary.2 Consequently, U ∗ AV has the following structure:
 ∗     
∗ y Ax y ∗ AV1 σy ∗ y y ∗ AV1 σ w∗
A1 := U AV = = = ,
U1∗ Ax U1∗ AV1 σU1∗ y U1∗ AV1 0 B
2 Recall that it is always possible to extend an orthonormal set of vectors to an orthonormal basis

for the whole space.

20 LINEAR ALGEBRA

where w := V1∗ A∗ y ∈ Fn−1 and B := U1∗ AV1 ∈ F(m−1)×(n−1) .

Since   2
1

 0 
∗  
A1  ..  = (σ2 + w∗ w),
 . 

0 2
2 ∗
it follows that kA1 k ≥ σ + w w. But since σ = kAk = kA1 k, we must have w = 0.
2

An obvious induction argument gives

U ∗ AV = Σ.

This completes the proof. 2

The σi is the ith singular value of A, and the vectors ui and vj are, respectively,
the ith left singular vector and the jth right singular vector. It is easy to verify that

Avi = σi ui
A∗ ui = σi vi .

The preceding equations can also be written as

A∗ Avi = σi2 vi
AA∗ ui = σi2 ui .

Hence σi2 is an eigenvalue of AA∗ or A∗ A, ui is an eigenvector of AA∗ , and vi is an

eigenvector of A∗ A.
The following notations for singular values are often adopted:

σ(A) = σmax (A) = σ1 = the largest singular value of A;

and
σ(A) = σmin (A) = σp = the smallest singular value of A.
Geometrically, the singular values of a matrix A are precisely the lengths of the semi-
axes of the hyperellipsoid E defined by

E = {y : y = Ax, x ∈ Cn , kxk = 1}.

Thus v1 is the direction in which kyk is largest for all kxk = 1; while vn is the direction
in which kyk is smallest for all kxk = 1. From the input/output point of view, v1 (vn )
is the highest (lowest) gain input (or control) direction, while u1 (um ) is the highest
(lowest) gain output (or observing) direction. This can be illustrated by the following
2 × 2 matrix:
   
cos θ1 − sin θ1 σ1 cos θ2 − sin θ2
A= .
sin θ1 cos θ1 σ2 sin θ2 cos θ2
2.6. Singular Value Decomposition 21

It is easy to see that A maps a unit circle to an ellipsoid with semiaxes of σ1 and σ2 .

Hence it is often convenient to introduce the following alternative definitions for the
largest singular value σ:
σ(A) := max kAxk
kxk=1

and for the smallest singular value σ of a tall matrix:

σ(A) := min kAxk .
kxk=1

Lemma 2.5 Suppose A and ∆ are square matrices. Then

(i) |σ(A + ∆) − σ(A)| ≤ σ(∆);
(ii) σ(A∆) ≥ σ(A)σ(∆);
1
(iii) σ(A−1 ) = if A is invertible.
σ(A)

Proof.
(i) By definition
σ(A + ∆) := min k(A + ∆)xk ≥ min {kAxk − k∆xk}
kxk=1 kxk=1

≥ min kAxk − max k∆xk

kxk=1 kxk=1

= σ(A) − σ(∆).
Hence −σ(∆) ≤ σ(A + ∆) − σ(A). The other inequality σ(A + ∆) − σ(A) ≤ σ(∆)
follows by replacing A by A + ∆ and ∆ by −∆ in the preceding proof.
(ii) This follows by noting that
σ(A∆) := min kA∆xk
kxk=1
r
= min x∗ ∆∗ A∗ A∆x
kxk=1

≥ σ(A) min k∆xk = σ(A)σ(∆).

kxk=1

(iii) Let the singular value decomposition of A be A = U ΣV ∗ ; then A−1 = V Σ−1 U ∗ .

Hence σ(A−1 ) = σ(Σ−1 ) = 1/σ(Σ) = 1/σ(A).
2

Note that (ii) may not be true if A and ∆ are not square matrices. For example,
1   √
consider A = and ∆ = 3 4 ; then σ(A∆) = 0 but σ(A) = 5 and σ(∆) = 5.
2
22 LINEAR ALGEBRA

Some useful properties of SVD are collected in the following lemma.

Lemma 2.6 Let A ∈ Fm×n and
σ1 ≥ σ2 ≥ · · · ≥ σr > σr+1 = · · · = 0, r ≤ min{m, n}.
Then
1. rank(A) = r;
2. KerA = span{vr+1 , . . . , vn } and (KerA)⊥ = span{v1 , . . . , vr };
3. ImA = span{u1 , . . . , ur } and (ImA)⊥ = span{ur+1 , . . . , um };
4. A ∈ Fm×n has a dyadic expansion:
X
r
A= σi ui vi∗ = Ur Σr Vr∗ ,
i=1

where Ur = [u1 , . . . , ur ], Vr = [v1 , . . . , vr ], and Σr = diag (σ1 , . . . , σr );

5. kAk2F = σ12 + σ22 + · · · + σr2 ;

6. kAk = σ1 ;
7. σi (U0 AV0 ) = σi (A), i = 1, . . . , p for any appropriately dimensioned unitary ma-
trices U0 and V0 ;
Pk
8. Let k < r = rank(A) and Ak := i=1 σi ui vi∗ ; then
min kA − Bk = kA − Ak k = σk+1 .
rank(B)≤k

Proof. We shall only give a proof for part 8. It is easy to see that rank(Ak ) ≤ k and
kA − Ak k = σk+1 . Hence, we only need show that min kA − Bk ≥ σk+1 . Let B
rank(B)≤k
be any matrix such that rank(B) ≤ k. Then
kA − Bk = kU ΣV ∗ − Bk = kΣ − U ∗ BV k
 
  Ik+1

≥ Ik+1 0 (Σ − U BV ) ∗ =
Σk+1 − B̂ ,
0
 
  ∗ Ik+1
where B̂ = Ik+1 0 U BV ∈ F(k+1)×(k+1) and rank(B̂) ≤ k. Let x ∈ Fk+1
0
be such that B̂x = 0 and kxk = 1. Then


kA − Bk ≥ Σk+1 − B̂ ≥ (Σk+1 − B̂)x = kΣk+1 xk ≥ σk+1 .

Since B is arbitrary, the conclusion follows. 2

2.7. Semidefinite Matrices 23

Illustrative MATLAB Commands:

 [U, Σ, V] = svd(A) % A = U ΣV ∗

Related MATLAB Commands: cond, condest

2.7 Semidefinite Matrices

A square Hermitian matrix A = A∗ is said to be positive definite (semidefinite), denoted
by A > 0 (≥ 0), if x∗ Ax > 0 (≥ 0) for all x 6= 0. Suppose A ∈ Fn×n and A = A∗ ≥ 0;
then there exists a B ∈ Fn×r with r ≥ rank(A) such that A = BB ∗ .

Lemma 2.7 Let B ∈ Fm×n and C ∈ Fk×n . Suppose m ≥ k and B ∗ B = C ∗ C. Then

there exists a matrix U ∈ Fm×k such that U ∗ U = I and B = U C.

Proof. Let V1 and V2 be unitary matrices such that

   
B1 C1
B = V1 , C = V2 ,
0 0

where B1 and C1 are full-row rank. Then B1 and C1 have the same number of rows
and V3 := B1 C1∗ (C1 C1∗ )−1 satisfies V3∗ V3 = I since B ∗ B = C ∗ C. Hence V3 is a unitary
matrix and V3∗ B1 = C1 . Finally, let
 
V3 0
U = V1 V2∗
0 V4

for any suitably dimensioned V4 such that V4∗ V4 = I. 2

We can define square root for a positive semidefinite matrix A, A1/2 = (A1/2 )∗ ≥ 0,
by
A = A1/2 A1/2 .
Clearly, A1/2 can be computed by using spectral decomposition or SVD: Let A = U ΛU ∗ ;
then
A1/2 = U Λ1/2 U ∗ ,
where p p
Λ = diag{λ1 , . . . , λn }, Λ1/2 = diag{ λ1 , . . . , λn }.
24 LINEAR ALGEBRA

Lemma 2.8 Suppose A = A∗ > 0 and B = B ∗ ≥ 0. Then A > B iff ρ(BA−1 ) < 1.

Proof. Since A > 0, we have A > B iff

0 < I − A−1/2 BA−1/2

i.e., iff ρ(A−1/2 BA−1/2 ) < 1. However, A−1/2 BA−1/2 and BA−1 are similar, hence
ρ(BA−1 ) = ρ(A−1/2 BA−1/2 ) and the claim follows. 2

2.8 Notes and References

A very extensive treatment of most topics in this chapter can be found in Brogan [1991],
Horn and Johnson [1990, 1991] and Lancaster and Tismenetsky [1985]. Golub and Van
Loan’s book [1983] contains many numerical algorithms for solving most of the problems
in this chapter.

2.9 Problems
Problem 2.1 Let  
1 1 0
1 0 1
 
A = 2 1 1.
 
1 0 1
2 0 2
Determine the row and column rank of A and find bases for Im(A), Im(A∗ ), and Ker(A).
 
1 4
Problem 2.2 Let D0 =  2 5 . Find a D such that D∗ D = I and ImD = ImD0 .
3 6
 
Furthermore, find a D⊥ such that D D⊥ is a unitary matrix.

Problem 2.3 Let A be a nonsingular matrix and x, y ∈ Cn . Show

Axy ∗ A
(A−1 + xy ∗ )−1 = A −
1 + y ∗ Ax
and
det A
det(A−1 + xy ∗ )−1 = .
1 + y ∗ Ax
Problem 2.4 Let A and B be compatible matrices. Show

B(I + AB)−1 = (I + BA)−1 B, (I + A)−1 = I − A(I + A)−1 .

2.9. Problems 25

Problem
 2.5 Find a basis for the maximum dimensional stable invariant subspace of
A R
H= with
−Q −A∗
   
−1 2 −1 −1
1. A = , R= , and Q = 0
3 0 −1 −1
     
0 1 0 0 1 2
2. A = , R= , and Q =
0 2 0 −1 2 4
 
1 2
3. A = 0, R = , and Q = I2 .
2 5

Problem 2.6 Let A = [aij ]. Show that α(A) := maxi,j |aij | defines a matrix norm.
Give examples so that α(A) < ρ(A) and α(AB) > α(A)α(B).
   
1 2 3 0
Problem 2.7 Let A = and B = . (a) Find all x such that Ax = B.
4 1 −1 1
(b) Find the minimal norm solution x: min {kxk : Ax = B}.
   
1 2 3
Problem 2.8 Let A =  −2 −5  and B =  4 . Find an x such that kAx − Bk
0 1 5
is minimized.

Problem 2.9 Let kAk < 1. Show

1. (I − A)−1 = I + A + A2 + · · · .

2. (I − A)−1 ≤ 1 + kAk + kAk + · · · =
2 1
1−kAk .

3. (I − A)−1 ≥ 1+kAk
1
.

Problem 2.10 Let A ∈ Cm×n . Show that

1 √
√ kAk2 ≤ kAk∞ ≤ n kAk2 ;
m
1 √
√ kAk2 ≤ kAk1 ≤ m kAk2 ;
n
1
kAk∞ ≤ kAk1 ≤ m kAk∞ .
n
Problem 2.11 Let A = xy ∗ and x, y ∈ Cn . Show that kAk2 = kAkF = kxk kyk.
 
1 1+j
. Find A 2 and a B ∈ C2 such that A = BB ∗ .
1
Problem 2.12 Let A =
1−j 2
26 LINEAR ALGEBRA

 
P11 P12
Problem 2.13 Let P = P ∗ = ∗ ≥ 0 with P11 ∈ Ck×k . Show λi (P ) ≥
P12 P22
λi (P11 ), ∀ 1 ≤ i ≤ k.
 
∗ X11 X12
Problem 2.14 Let X = X ≥ 0 be partitioned as X = ∗ . (a) Show
X12 X22
∗
KerX22 ⊂ KerX12 ; (b) let X22 = U2 diag(Λ1 , 0)U2 be such that Λ1 is nonsingular and
+
define X22 := U2 diag(Λ−1 ∗
1 , 0)U2 (the pseudoinverse of X22 ); then show that Y = X12 X22
+

solves Y X22 = X12 ; and (c) show that

    ∗
 
X11 X12 I X12 X22 +
X11 − X12 X22
+
X12 0 I 0
∗ = + ∗ .
X12 X22 0 I 0 X22 X22 X12 I
Chapter 3

Linear Systems

This chapter reviews some basic system theoretical concepts. The notions of controlla-
bility, observability, stabilizability, and detectability are defined and various algebraic
and geometric characterizations of these notions are summarized. Observer theory is
then introduced. System interconnections and realizations are studied. Finally, the
concepts of system poles and zeros are introduced.

3.1 Descriptions of Linear Dynamical Systems

Let a finite dimensional linear time invariant (FDLTI) dynamical system be described
by the following linear constant coefficient differential equations:
ẋ = Ax + Bu, x(t0 ) = x0 (3.1)
y = Cx + Du, (3.2)
where x(t) ∈ Rn is called the system state, x(t0 ) is called the initial condition of the
system, u(t) ∈ Rm is called the system input, and y(t) ∈ Rp is the system output. The
A, B, C, and D are appropriately dimensioned real constant matrices. A dynamical
system with single-input (m = 1) and single-output (p = 1) is called a SISO (single-
input and single-output) system; otherwise it is called a MIMO (multiple-input and
multiple-output) system. The corresponding transfer matrix from u to y is defined as
Y (s) = G(s)U (s),
where U (s) and Y (s) are the Laplace transforms of u(t) and y(t) with zero initial
condition (x(0) = 0). Hence, we have
G(s) = C(sI − A)−1 B + D.
Note that the system equations (3.1) and (3.2) can be written in a more compact matrix
form:     
ẋ A B x
= .
y C D u

27
28 LINEAR SYSTEMS

To expedite calculations involving transfer matrices, we shall use the following notation:
 
A B
:= C(sI − A)−1 B + D.
C D

In Matlab the system can also be written in the packed form using the command
 G=pck(A, B, C, D) % pack the realization in partitioned form
 seesys(G) % display G in partitioned format
 [A, B, C, D]=unpck(G) % unpack the system matrix
Note that  
A B
C D
is a real block matrix, not a transfer function.

Illustrative MATLAB Commands:

 G=pck([], [], [], 10) % create a constant system matrix

 [y, x, t]=step(A, B, C, D, Iu) % Iu=i (step response of the ith channel)
 [y, x, t]=initial(A, B, C, D, x0 ) % initial response with initial condition x0
 [y, x, t]=impulse(A, B, C, D, Iu) % impulse response of the Iuth channel
 [y,x]=lsim(A,B,C,D,U,T) % U is a length(T ) × column(B) matrix input; T is
the sampling points.

Related MATLAB Commands: minfo, trsp, cos tr, sin tr, siggen

3.2 Controllability and Observability

We now turn to some very important concepts in linear system theory.
Definition 3.1 The dynamical system described by equation (3.1) or the pair (A, B)
is said to be controllable if, for any initial state x(0) = x0 , t1 > 0 and final state x1 ,
there exists a (piecewise continuous) input u(·) such that the solution of equation (3.1)
satisfies x(t1 ) = x1 . Otherwise, the system or the pair (A, B) is said to be uncontrollable.
The controllability (and the observability introduced next) of a system can be verified
through some algebraic or geometric criteria.
3.2. Controllability and Observability 29

Theorem 3.1 The following are equivalent:

(i) (A, B) is controllable.
(ii) The matrix
Z t
∗
Wc (t) := eAτ BB ∗ eA τ dτ
0
is positive definite for any t > 0.
(iii) The controllability matrix
 
C= B AB A2 B
. . . An−1 B
Pn
has full-row rank or, in other words, hA |ImBi := i=1 Im(Ai−1 B) = Rn .
(iv) The matrix [A − λI, B] has full-row rank for all λ in C.
(v) Let λ and x be any eigenvalue and any corresponding left eigenvector of A (i.e.,
x∗ A = x∗ λ); then x∗ B 6= 0.

(vi) The eigenvalues of A+BF can be freely assigned (with the restriction that complex
eigenvalues are in conjugate pairs) by a suitable choice of F .

       
2 0 1 1 0
Example 3.1 Let A = and B = . Then x1 = and x2 =
0 2 1 0 1
are independent eigenvectors of A and x∗i B 6= 0, i = 1, 2. However, this should not lead
one to conclude that (A, B) is controllable. In fact, x = x1 − x2 is also an eigenvector
of A and x∗ B = 0, which implies that (A, B) is not controllable. Hence one must check
for all possible eigenvectors in using criterion (v).

Definition 3.2 An unforced dynamical system ẋ = Ax is said to be stable if all the

eigenvalues of A are in the open left half plane; that is, Reλ(A) < 0. A matrix A with
such a property is said to be stable or Hurwitz.

Definition 3.3 The dynamical system of equation (3.1), or the pair (A, B), is said to
be stabilizable if there exists a state feedback u = F x such that the system is stable
(i.e., A + BF is stable).

It is more appropriate to call this stabilizability the state feedback stabilizability to

differentiate it from the output feedback stabilizability defined later.
The following theorem is a consequence of Theorem 3.1.
30 LINEAR SYSTEMS

Theorem 3.2 The following are equivalent:

(i) (A, B) is stabilizable.
(ii) The matrix [A − λI, B] has full-row rank for all Reλ ≥ 0.
(iii) For all λ and x such that x∗ A = x∗ λ and Reλ ≥ 0, x∗ B 6= 0.
(iv) There exists a matrix F such that A + BF is Hurwitz.

We now consider the dual notions: observability and detectability of the system
described by equations (3.1) and (3.2).

Definition 3.4 The dynamical system described by equations (3.1) and (3.2) or by the
pair (C, A) is said to be observable if, for any t1 > 0, the initial state x(0) = x0 can be
determined from the time history of the input u(t) and the output y(t) in the interval
of [0, t1 ]. Otherwise, the system, or (C, A), is said to be unobservable.

Theorem 3.3 The following are equivalent:

(i) (C, A) is observable.
(ii) The matrix
Z t
∗
Wo (t) := eA τ C ∗ CeAτ dτ
0
is positive definite for any t > 0.
(iii) The observability matrix  
C
 CA 
 
 CA2 
O= 
 .. 
 . 
CAn−1
T
has full-column rank or ni=1 Ker(CAi−1 ) = 0.
 
A − λI
(iv) The matrix has full-column rank for all λ in C.
C

(v) Let λ and y be any eigenvalue and any corresponding right eigenvector of A (i.e.,
Ay = λy); then Cy 6= 0.
(vi) The eigenvalues of A+LC can be freely assigned (with the restriction that complex
eigenvalues are in conjugate pairs) by a suitable choice of L.
(vii) (A∗ , C ∗ ) is controllable.
3.3. Observers and Observer-Based Controllers 31

Definition 3.5 The system, or the pair (C, A), is detectable if A + LC is stable for
some L.

Theorem 3.4 The following are equivalent:

(i) (C, A) is detectable.
 
A − λI
(ii) The matrix has full-column rank for all Reλ ≥ 0.
C
(iii) For all λ and x such that Ax = λx and Reλ ≥ 0, Cx 6= 0.
(iv) There exists a matrix L such that A + LC is Hurwitz.
(v) (A∗ , C ∗ ) is stabilizable.

The conditions (iv) and (v) of Theorems 3.1 and 3.3 and the conditions (ii) and
(iii) of Theorems 3.2 and 3.4 are often called Popov-Belevitch-Hautus (PBH) tests. In
particular, the following definitions of modal controllability and observability are often
useful.

Definition 3.6 Let λ be an eigenvalue of A or, equivalently, a mode of the system.

Then the mode λ is said to be controllable (observable) if x∗ B 6= 0 (Cx 6= 0) for all
left (right) eigenvectors of A associated with λ; that is, x∗ A = λx∗ (Ax = λx) and
0 6= x ∈ Cn . Otherwise, the mode is said to be uncontrollable (unobservable).

It follows that a system is controllable (observable) if and only if every mode is control-
lable (observable). Similarly, a system is stabilizable (detectable) if and only if every
unstable mode is controllable (observable).

Illustrative MATLAB Commands:

 C= ctrb(A, B); O= obsv(A, C);

 Wc (∞)=gram(A, B); % if A is stable.
 F=-place(A, B, P) % P is a vector of desired eigenvalues.

Related MATLAB Commands: ctrbf, obsvf, canon, strans, acker

3.3 Observers and Observer-Based Controllers

It is clear that if a system is controllable and the system states are available for feed-
back, then the system closed-loop poles can be assigned arbitrarily through a constant
feedback. However, in most practical applications, the system states are not completely
accessible and all the designer knows are the output y and input u. Hence, the esti-
mation of the system states from the given output information y and input u is often
32 LINEAR SYSTEMS

necessary to realize some specific design objectives. In this section, we consider such an
estimation problem and the application of this state estimation in feedback control.
Consider a plant modeled by equations (3.1) and (3.2). An observer is a dynamical
system with input (u, y) and output (say, x̂), that asymptotically estimates the state x,
that is, x̂(t) − x(t) → 0 as t → ∞ for all initial states and for every input.
Theorem 3.5 An observer exists iff (C, A) is detectable. Further, if (C, A) is de-
tectable, then a full-order Luenberger observer is given by
q̇ = Aq + Bu + L(Cq + Du − y) (3.3)
x̂ = q, (3.4)
where L is any matrix such that A + LC is stable.
Recall that, for a dynamical system described by the equations (3.1) and (3.2), if (A, B)
is controllable and state x is available for feedback, then there is a state feedback u = F x
such that the closed-loop poles of the system can be arbitrarily assigned. Similarly, if
(C, A) is observable, then the system observer poles can be arbitrarily placed so that the
state estimator x̂ can be made to approach x arbitrarily fast. Now let us consider what
will happen if the system states are not available for feedback so that the estimated
state has to be used. Hence, the controller has the following dynamics:
x̂˙ = (A + LC)x̂ + Bu + LDu − Ly
u = F x̂.
Then the total system state equations are given by
    
ẋ A BF x
= .
x̂˙ −LC A + BF + LC x̂
Let e := x − x̂; then the system equation becomes
    
ė A + LC 0 e
=
x̂˙ −LC A + BF x̂
and the closed-loop poles consist of two parts: the poles resulting from state feedback
λi (A + BF ) and the poles resulting from the state estimation λj (A + LC). Now if
(A, B) is controllable and (C, A) is observable, then there exist F and L such that the
eigenvalues of A + BF and A + LC can be arbitrarily assigned. In particular, they can
be made to be stable. Note that a slightly weaker result can also result even if (A, B)
and (C, A) are only stabilizable and detectable.
The controller given above is called an observer-based controller and is denoted as
u = K(s)y
and  
A + BF + LC + LDF −L
K(s) = .
F 0
3.3. Observers and Observer-Based Controllers 33

Now denote the open-loop plant by

 
A B
G= ;
C D

then the closed-loop feedback system is as shown below:

y u
 G  K 

In general, if a system is stabilizable through feeding back the output y, then it is

said to be output feedback stabilizable. It is clear from the above construction that a
system is output feedback stabilizable if and only if (A, B) is stabilizable and (C, A) is
detectable.

   
1 2 1  
Example 3.2 Let A = ,B = , and C = 1 0 . We shall design a
1 0 0
state feedback u = F x such that the closed-loop poles are at {−2, −3}. This can be
done by choosing F = −6 −8 using

 F = −place(A, B, [−2, −3]).

Now suppose the states are not available for feedback and we want to construct
 an
−21
observer so that the observer poles are at {−10, −10}. Then L = can be
−51
obtained by using
 L = −acker(A0 , C 0 , [−10, −10])0
and the observer-based controller is given by

−534(s + 0.6966)
K(s) = .
(s + 34.6564)(s − 8.6564)

Note that the stabilizing controller itself is unstable. Of course, this may not be desirable
in practice.
34 LINEAR SYSTEMS

3.4 Operations on Systems

In this section, we present some facts about system interconnection. Since these proofs
are straightforward, we will leave the details to the reader.
Suppose that G1 and G2 are two subsystems with state-space representations:
   
A1 B1 A2 B2
G1 = , G2 = .
C1 D1 C2 D2
Then the series or cascade connection of these two subsystems is a system with the
output of the second subsystem as the input of the first subsystem, as shown in the
following diagram:
 G1  G2 

This operation in terms of the transfer matrices of the two subsystems is essentially the
product of two transfer matrices. Hence, a representation for the cascaded system can
be obtained as
  
A1 B1 A2 B2
G1 G2 =
C1 D1 C2 D2
   
A1 B1 C2 B1 D2 A2 0 B2
=  0 A2 B2  =  B1 C2 A1 B1 D2  .
C1 D1 C2 D1 D2 D1 C2 C1 D1 D2
Similarly, the parallel connection or the addition of G1 and G2 can be obtained as
 
    A1 0 B1
A1 B1 A2 B2
G1 + G2 = + =  0 A2 B2 .
C1 D1 C2 D2
C1 C2 D1 + D2
For future reference, we shall also introduce the following definitions.
Definition 3.7 The transpose of a transfer matrix G(s) or the dual system is defined
as
G 7−→ GT (s) = B ∗ (sI − A∗ )−1 C ∗ + D∗
or, equivalently,    
A B A∗ C∗
7−→ .
C D B∗ D∗
Definition 3.8 The conjugate system of G(s) is defined as
G 7−→ G∼ (s) := GT (−s) = B ∗ (−sI − A∗ )−1 C ∗ + D∗
or, equivalently,    
A B −A∗ −C ∗
7−→ .
C D B∗ D∗
3.5. State-Space Realizations for Transfer Matrices 35

∗
In particular, we have G∗ (jω) := [G(jω)] = G∼ (jω).
A real rational matrix Ĝ(s) is called an inverse of a transfer matrix G(s) if G(s)Ĝ(s) =
Ĝ(s)G(s) = I. Suppose G(s) is square and D is invertible. Then
 
A − BD−1 C −BD−1
G−1 = .
D−1 C D−1

The corresponding Matlab commands are

 
G1 G2 ⇐⇒ mmult(G1 , G2 ), G1 G2 ⇐⇒ sbs(G1 , G2 )

G1 + G2 ⇐⇒ madd(G1 , G2 ), G1 − G2 ⇐⇒ msub(G1 , G2 )
   
G1 G1
⇐⇒ abv(G1 , G2 ), ⇐⇒ daug(G1 , G2 ),
G2 G2
GT (s) ⇐⇒ transp(G), G∼ (s) ⇐⇒ cjt(G), G−1 (s) ⇐⇒ minv(G)
α G(s) ⇐⇒ mscl(G, α), α is a scalar.

Related MATLAB Commands: append, parallel, feedback, series, cloop,

sclin, sclout, sel

3.5 State-Space Realizations for Transfer Matrices

In some cases, the natural or convenient description for a dynamical system is in terms
of matrix transfer function. This occurs, for example, in some highly complex sys-
tems for which the analytic differential equations are too hard or too complex to write
down. Hence certain engineering approximation or identification has to be carried out;
for example, input and output frequency responses are obtained from experiments so
that some transfer matrix approximating the system dynamics can be obtained. Since
the state-space computation is most convenient to implement on the computer, some
appropriate state-space representation for the resulting transfer matrix is necessary.
In general, assume that G(s) is a real rational transfer matrix that is proper. Then
we call a state-space model (A, B, C, D) such that
 
A B
G(s) =
C D

a realization of G(s).

Definition 3.9 A state-space realization (A, B, C, D) of G(s) is said to be a minimal

realization of G(s) if A has the smallest possible dimension.

Theorem 3.6 A state-space realization (A, B, C, D) of G(s) is minimal if and only if

(A, B) is controllable and (C, A) is observable.
36 LINEAR SYSTEMS

We now describe several ways to obtain a state-space realization for a given multiple-
input and multiple-output transfer matrix G(s). We shall first consider SIMO (single-
input and multiple-output) and MISO (multiple-input and single-output) systems.
Let G(s) be a column vector of transfer function with p outputs:

β1 sn−1 + β2 sn−2 + · · · + βn−1 s + βn

G(s) = + d, βi ∈ Rp , d ∈ Rp .
sn + a1 sn−1 + · · · + an−1 s + an
 
A b
Then G(s) = with
C d
   
−a1 −a2 · · · −an−1 −an 1
 1 0 ··· 0 0   0 
   
 0 1 · · · 0 0   0 
A :=   b :=  
 .. .. .. ..   .. 
 . . . .   . 
0 0 ··· 1 0 0
 
C= β1 β2 · · · βn−1 βn
is a so-called controllable canonical form or controller canonical form.
Dually, consider a multiple-input and single-output system
η1 sn−1 + η2 sn−2 + · · · + ηn−1 s + ηn
G(s) = + d, ηi∗ ∈ Rm , d∗ ∈ Rm .
sn + a1 sn−1 + · · · + an−1 s + an
Then  
−a1 1 0 ··· 0 η1
 −a2 0 1 ··· 0 η2 
   
 .. .. .. .. .. 
A B  . . . . . 
G(s) = = 
c d  −an−1 0 0 · · · 1 ηn−1 
 
 −an 0 0 ··· 0 ηn 
1 0 0 ··· 0 d
is an observable canonical form or observer canonical form.
For a MIMO system, the simplest and most straightforward way to obtain a real-
ization is by realizing each element of the matrix G(s) and then combining all these
individual realizations to form a realization for G(s). To illustrate, let us consider a
2 × 2 (block) transfer matrix such as
 
G1 (s) G2 (s)
G(s) =
G3 (s) G4 (s)

and assume that Gi (s) has a state-space realization of the form

 
Ai Bi
Gi (s) = , i = 1, . . . , 4.
Ci Di
3.5. State-Space Realizations for Transfer Matrices 37

Then a realization for G(s) can be obtained as (G = abv(sbs(G1 , G2 ), sbs(G3 , G4 ))):

 
A1 0 0 0 B1 0
 0 A2 0 0 0 B2 
 
 0 0 A3 0 B3 0 
G(s) = 

.

 0 0 0 A4 0 B4 
 C1 C2 0 0 D1 D2 
0 0 C3 C4 D3 D4

Alternatively, if the transfer matrix G(s) can be factored into the product and/or
the sum of several simply realized transfer matrices, then a realization for G can be
obtained by using the cascade or addition formulas given in the preceding section.

A problem inherited with these kinds of realization procedures is that a realization

thus obtained will generally not be minimal. To obtain a minimal realization, a Kalman
controllability and observability decomposition has to be performed to eliminate the un-
controllable and/or unobservable states. (An alternative numerically reliable method to
eliminate uncontrollable and/or unobservable states is the balanced realization method,
which will be discussed later.)

We shall now describe a procedure that does result in a minimal realization by

using partial fractional expansion (the resulting realization is sometimes called Gilbert’s
realization due to E. G. Gilbert).
Let G(s) be a p × m transfer matrix and write it in the following form:

N (s)
G(s) =
d(s)

with d(s) a scalar polynomial. For simplicity, we shall assume that d(s) has only real
and distinct roots λi 6= λj if i 6= j and

d(s) = (s − λ1 )(s − λ2 ) · · · (s − λr ).

Then G(s) has the following partial fractional expansion:

Xr
Wi
G(s) = D + .
i=1
s − λi

Suppose
rank Wi = ki

and let Bi ∈ Rki ×m and Ci ∈ Rp×ki be two constant matrices such that

Wi = Ci Bi .
38 LINEAR SYSTEMS

Then a realization for G(s) is given by

 
λ1 Ik1 B1
 .. .. 
 . . 
G(s) =  .
 λr Ikr Br 
C1 ··· Cr D

It follows immediately from PBH tests that this realization is controllable and observ-
able, and thus it is minimal.
An immediate consequence of this minimal realization is that a transfer matrix with
an rth order polynomial denominator does not necessarily have an rth order state-space
realization unless Wi for each i is a rank one matrix.
This approach can, in fact, be generalized to more complicated cases where d(s) may
have complex and/or repeated roots. Readers may convince themselves by trying some
simple examples.

Illustrative MATLAB Commands:

 G=nd2sys(num, den, gain); G=zp2sys(zeros, poles, gain);
Related MATLAB Commands: ss2tf, ss2zp, zp2ss, tf2ss, residue, minreal

3.6 Multivariable System Poles and Zeros

A matrix is called a polynomial matrix of a variable if every element of the matrix is a
polynomial of the variable.

Definition 3.10 Let Q(s) be a (p × m) polynomial matrix (or a transfer matrix) of s.

Then the normal rank of Q(s), denoted normalrank (Q(s)), is the maximally possible
rank of Q(s) for at least one s ∈ C.

To show the difference between the normal rank of a polynomial matrix and the rank
of the polynomial matrix evaluated at a certain point, consider
 
s 1
Q(s) =  s2 1  .
s 1

Then Q(s) has normal rank 2 since rank Q(3) = 2. However, Q(0) has rank 1.
The poles and zeros of a transfer matrix can be characterized in terms of its state-
space realizations. Let  
A B
C D
be a state-space realization of G(s).
3.6. Multivariable System Poles and Zeros 39

Definition 3.11 The eigenvalues of A are called the poles of the realization of G(s).

To define zeros, let us consider the following system matrix:

 
A − sI B
Q(s) = .
C D

Definition 3.12 A complex number z0 ∈ C is called an invariant zero of the system

realization if it satisfies
   
A − z0 I B A − sI B
rank < normalrank .
C D C D

The invariant zeros are not changed by constant state feedback since
    
A + BF − z0 I B A − z0 I B I 0
rank = rank
C + DF D C D F I
 
A − z0 I B
= rank .
C D

It is also clear that invariant zeros are not changed under similarity transformation.
The following lemma is obvious.
 
A − sI B
Lemma 3.7 Suppose has full-column normal rank. Then z0 ∈ C is an
C D
invariant zero of a realization (A, B, C, D) if and only if there exist 0 6= x ∈ Cn and
u ∈ Cm such that   
A − z0 I B x
= 0.
C D u
Moreover, if u = 0, then z0 is also a nonobservable mode.
 
x
Proof. By definition, z0 is an invariant zero if there is a vector 6= 0 such that
u
  
A − z0 I B x
=0
C D u
 
A − sI B
since has full-column normal rank.
C D
 
x
On the other hand, suppose z0 is an invariant zero; then there is a vector 6= 0
u
such that   
A − z0 I B x
= 0.
C D u
40 LINEAR SYSTEMS

   
B A − sI B
We claim that x 6= 0. Otherwise, u = 0 or u = 0 since has
D C D
 
x
full-column normal rank (i.e., = 0), which is a contradiction.
u
Finally, note that if u = 0, then
 
A − z0 I
x=0
C
and z0 is a nonobservable mode by PBH test. 2

When the system is square, the invariant zeros can be computed by solving a gen-
eralized eigenvalue problem:
     
A B x I 0 x
= z0
C D u 0 0 u
| {z } | {z }
M N

using a Matlab command: eig(M, N).

 
A − sI B
Lemma 3.8 Suppose has full-row normal rank. Then z0 ∈ C is an
C D
invariant zero of a realization (A, B, C, D) if and only if there exist 0 6= y ∈ Cn and
v ∈ Cp such that  
 ∗  A − z0 I B
y v∗ = 0.
C D
Moreover, if v = 0, then z0 is also a noncontrollable mode.
 
A − sI B
Lemma 3.9 G(s) has full-column (row) normal rank if and only if has
C D
full-column (row) normal rank.

Proof. This follows by noting that

    
A − sI B I 0 A − sI B
=
C D C(A − sI)−1 I 0 G(s)
and  
A − sI B
normalrank = n + normalrank(G(s)).
C D
2

Lemma 3.10 Let G(s) ∈ Rp (s) be a p × m transfer matrix and let (A, B, C, D) be a
minimal realization. If the system input is of the form u(t) = u0 eλt , where λ ∈ C is not
a pole of G(s) and u0 ∈ Cm is an arbitrary constant vector, then the output due to the
input u(t) and the initial state x(0) = (λI − A)−1 Bu0 is y(t) = G(λ)u0 eλt , ∀t ≥ 0. In
particular, if λ is a zero of G(s), then y(t) = 0.
3.7. Notes and References 41

Proof. The system response with respect to the input u(t) = u0 eλt and the initial
condition x(0) = (λI − A)−1 Bu0 is (in terms of the Laplace transform)
Y (s) = C(sI − A)−1 x(0) + C(sI − A)−1 BU (s) + DU (s)
= C(sI − A)−1 x(0) + C(sI − A)−1 Bu0 (s − λ)−1 + Du0 (s − λ)−1
 
= C(sI − A)−1 x(0) + C (sI − A)−1 − (λI − A)−1 Bu0 (s − λ)−1
+C(λI − A)−1 Bu0 (s − λ)−1 + Du0 (s − λ)−1
= C(sI − A)−1 (x(0) − (λI − A)−1 Bu0 ) + G(λ)u0 (s − λ)−1
= G(λ)u0 (s − λ)−1 .
Hence y(t) = G(λ)u0 eλt . 2

Example 3.3 Let

 
−1 −2 1 1 2 3
   0 2 −1 3 2 1 
A B  
G(s) = =
 −4 −3 −2 1 1 1 
.
C D  1 1 1 0 0 0 
2 3 4 0 0 0
Then the invariant zeros of the system can be found using the Matlab command
 G=pck(A, B, C, D), z0 = szeros(G), % or
 z0 = tzero(A, B, C, D)
which gives z0 = 0.2. Since G(s) is full-row rank, we can find y and v such that
 
 ∗  A − z0 I B
y v∗ = 0,
C D
which can again be computed using a Matlab command:
 
0.0466
   0.0466 
y  
 null([A − z0 ∗ eye(3), B; C, D]0 ) =⇒ =
 −0.1866
.

v  −0.9702 
0.1399

Related MATLAB Commands: spoles, rifd

3.7 Notes and References

Readers are referred to Brogan [1991], Chen [1984], Kailath [1980], and Wonham [1985]
for extensive treatment of standard linear system theory.
42 LINEAR SYSTEMS

3.8 Problems
Problem 3.1 Let A ∈ Cn×n and B ∈ Cm×m . Show that X(t) = eAt X(0)eBt is the
solution to
Ẋ = AX + XB.

Problem 3.2 Given the impulse response, h1 (t), of a linear time-invariant system with
model
ÿ + a1 ẏ + a2 y = u
find the impulse response, h(t), of the system

ÿ + a1 ẏ + a2 y = b0 ü + b1 u̇ + b2 u.

Justify your answer and generalize your result to nth order systems.

Problem 3.3  Suppose

 a second order
 system is given
 by ẋ = Ax. Suppose
  it is known
4 1 5 1
that x(1) = for x(0) = and x(1) = for x(0) = . Find x(n)
 −2
 1 −2 2
1
for x(0) = . Can you determine A?
0

Problem 3.4 Assume (A, B) is controllable. Show that (F, G) with

   
A 0 B
F = , G=
C 0 0

is controllable if and only if  

A B
C 0
is a full-row rank matrix.

Problem 3.5 Let λi , i = 1, 2, . . . , n be n distinct eigenvalues of a matrix A ∈ Cn×n and

let xi and yi be the corresponding right- and left-unit eigenvectors such that yi∗ xi = 1.
Show that
Xn
yi∗ xj = δij , A = λi xi yi∗
i=1

and
X
n
(Cxi )(y ∗ B)
−1
C(sI − A) B= i
.
i=1
s − λi

Furthermore, show that the mode λi is controllable iff yi∗ B 6= 0 and the mode λi is
observable iff Cxi 6= 0.
3.8. Problems 43

Problem 3.6 Let (A, b, c) be a realization with A ∈ Rn×n , b ∈ Rn , and c∗ ∈ Rn .

Assume that λi (A) + λj (A) 6= 0 for all i, j. This assumption ensures the existence of
X ∈ Rn×n such that
AX + XA + bc = 0
Show that X is nonsingular if and only if (A, b) is controllable and (c, A) is observable.

Problem 3.7 Compute the system zeros and the corresponding zero directions of the
following transfer functions
   
1 2 2 2 −1 −2 1 2
 1 0 3 4   0 1 2 1 
G1 (s) =   
 0 1 1 2  , G2 (s) =  1
,

1 0 0
1 0 2 0 1 1 0 0
 1 s+3 
 
2(s + 1)(s + 2) s+2  s+1 (s + 1)(s − 2) 
G3 (s) = , G4 (s) = 
 10


s(s + 3)(s + 4) (s + 1)(s + 3) 5
s−2 s+3
Also find the vectors x and u whenever appropriate so that either
    
A − zI B x  ∗ ∗
 A − zI B
= 0 or x u = 0.
C D u C D
44 LINEAR SYSTEMS
Chapter 4

H2 and H∞ Spaces

The most important objective of a control system is to achieve certain performance

specifications in addition to providing internal stability. One way to describe the per-
formance specifications of a control system is in terms of the size of certain signals of
interest. For example, the performance of a tracking system could be measured by the
size of the tracking error signal. In this chapter, we look at several ways of defining a
signal’s size (i.e., at several norms for signals). Of course, which norm is most appro-
priate depends on the situation at hand. For that purpose, we shall first introduce the
Hardy spaces H2 and H∞ . Some state-space methods of computing real rational H2
and H∞ transfer matrix norms are also presented.

4.1 Hilbert Spaces

Recall the inner product of vectors defined on a Euclidean space Cn :
   
x1 y1
Xn
 ..   .. 
∗
hx, yi := x y = x̄i yi ∀x =  .  , y =  . ∈C .
n

i=1 xn yn
Note that many important metric notions and geometrical properties, such as length,
distance, angle, and the energy of physical systems, can be deduced from this inner
product. For instance, the length of a vector x ∈ Cn is defined as
p
kxk := hx, xi

and the angle between two vectors x, y ∈ Cn can be computed from

hx, yi
cos ∠(x, y) = , ∠(x, y) ∈ [0, π].
kxk kyk
The two vectors are said to be orthogonal if ∠(x, y) = π2 .

45
46 H2 AND H∞ SPACES

We now consider a natural generalization of the inner product on Cn to more general

(possibly infinite dimensional) vector spaces.
Definition 4.1 Let V be a vector space over C. An inner product1 on V is a complex-
valued function,
h·, ·i : V × V 7−→ C
such that for any x, y, z ∈ V and α, β ∈ C
(i) hx, αy + βzi = αhx, yi + βhx, zi
(ii) hx, yi = hy, xi
(iii) hx, xi > 0 if x 6= 0.
A vector space V with an inner product is called an inner product space.
p
It is clear that the inner product defined above induces a norm kxk := hx, xi, so
that the norm conditions in Chapter 2 are satisfied. In particular, the distance between
vectors x and y is d(x, y) = kx − yk.
Two vectors x and y in an inner product space V are said to be orthogonal if
hx, yi = 0, denoted x ⊥ y. More generally, a vector x is said to be orthogonal to a set
S ⊂ V , denoted by x ⊥ S, if x ⊥ y for all y ∈ S.
The inner product and the inner product induced norm have the following familiar
properties.
Theorem 4.1 Let V be an inner product space and let x, y ∈ V . Then
(i) |hx, yi| ≤ kxk kyk (Cauchy-Schwarz inequality). Moreover, the equality holds if
and only if x = αy for some constant α or y = 0.
(ii) kx + yk2 + kx − yk2 = 2 kxk2 + 2 kyk2 (Parallelogram law).
2 2 2
(iii) kx + yk = kxk + kyk if x ⊥ y.
A Hilbert space is a complete inner product space with the norm induced by its inner
product. For example, Cn with the usual inner product is a (finite dimensional) Hilbert
space. More generally, it is straightforward to verify that Cn×m with the inner product
defined as
Xn X m
hA, Bi := trace A∗ B = āij bij ∀A, B ∈ Cn×m
i=1 j=1

is also a (finite dimensional) Hilbert space.

A well-know infinite dimensional Hilbert space is L2 [a, b], which consists of all square
integrable and Lebesgue measurable functions defined on an interval [a, b] with the inner
product defined as
Z b
hf, gi := f (t)∗ g(t)dt
a
1 The
property (i) in the following list is the other way around to the usual mathematical convention
since we want to have hx, yi = x∗ y rather than y ∗ x for x, y ∈ Cn .
4.2. H2 and H∞ Spaces 47

for f, g ∈ L2 [a, b]. Similarly, if the functions are vector or matrix-valued, the inner
product is defined correspondingly as
Z b
hf, gi := trace [f (t)∗ g(t)] dt.
a

Some spaces used often in this book are L2 [0, ∞), L2 (−∞, 0], L2 (−∞, ∞). More pre-
cisely, they are defined as
L2 = L2 (−∞, ∞): Hilbert space of matrix-valued functions on R, with inner product
Z ∞
hf, gi := trace [f (t)∗ g(t)] dt.
−∞

L2+ = L2 [0, ∞): subspace of L2 (−∞, ∞) with functions zero for t < 0.
L2− = L2 (−∞, 0]: subspace of L2 (−∞, ∞) with functions zero for t > 0.

4.2 H2 and H∞ Spaces

Let S ⊂ C be an open set, and let f (s) be a complex-valued function defined on S:

f (s) : S 7−→ C.

Then f (s) is said to be analytic at a point z0 in S if it is differentiable at z0 and also

at each point in some neighborhood of z0 . It is a fact that if f (s) is analytic at z0 then
f has continuous derivatives of all orders at z0 . Hence, a function analytic at z0 has
a power series representation at z0 . The converse is also true (i.e., if a function has
a power series at z0 , then it is analytic at z0 ). A function f (s) is said to be analytic
in S if it has a derivative or is analytic at each point of S. A matrix-valued function
is analytic in S if every element of the matrix is analytic in S. For example, all real
rational stable transfer matrices are analytic in the right-half plane and e−s is analytic
everywhere.
A well-know property of the analytic functions is the so-called maximum modulus
theorem.
Theorem 4.2 If f (s) is defined and continuous on a closed-bounded set S and analytic
on the interior of S, then |f (s)| cannot attain the maximum in the interior of S unless
f (s) is a constant.

The theorem implies that |f (s)| can only achieve its maximum on the boundary of S;
that is,
max |f (s)| = max |f (s)|
s∈S s∈∂S

where ∂S denotes the boundary of S. Next we consider some frequently used complex
(matrix) function spaces.
48 H2 AND H∞ SPACES

L2 (jR) Space
L2 (jR) or simply L2 is a Hilbert space of matrix-valued (or scalar-valued) func-
tions on jR and consists of all complex matrix functions F such that the following
integral is bounded:
Z ∞
trace [F ∗ (jω)F (jω)] dω < ∞.
−∞

The inner product for this Hilbert space is defined as

Z ∞
1
hF, Gi := trace [F ∗ (jω)G(jω)] dω
2π −∞
for F, G ∈ L2 , and the inner product induced norm is given by
p
kF k2 := hF, F i.

For example, all real rational strictly proper transfer matrices with no poles on the
imaginary axis form a subspace (not closed) of L2 (jR) that is denoted by RL2 (jR) or
simply RL2 .
H2 Space2
H2 is a (closed) subspace of L2 (jR) with matrix functions F (s) analytic in
Re(s) > 0 (open right-half plane). The corresponding norm is defined as
 Z ∞ 
1
trace [F ∗ (σ + jω)F (σ + jω)] dω .
2
kF k2 := sup
σ>0 2π −∞

It can be shown3 that

Z ∞
1
kF k22 = trace [F ∗ (jω)F (jω)] dω.
2π −∞

Hence, we can compute the norm for H2 just as we do for L2 . The real rational
subspace of H2 , which consists of all strictly proper and real rational stable transfer
matrices, is denoted by RH2 .
H2⊥ Space
H2⊥ is the orthogonal complement of H2 in L2 ; that is, the (closed) subspace of
functions in L2 that are analytic in the open left-half plane. The real rational
subspace of H2⊥ , which consists of all strictly proper rational transfer matrices
with all poles in the open right-half plane, will be denoted by RH⊥ 2 . It is easy to
see that if G is a strictly proper, stable, and real rational transfer matrix, then
G ∈ H2 and G∼ ∈ H2⊥ . Most of our study in this book will be focused on the real
rational case.
2 The H space and H
2 ∞ space defined in this subsection together with the Hp spaces, p ≥ 1,
which will not be introduced in this book, are usually called Hardy spaces and are named after the
mathematician G. H. Hardy (hence the notation of H).
3 See Francis [1987].
4.2. H2 and H∞ Spaces 49

The L2 spaces defined previously in the frequency domain can be related to the L2
spaces defined in the time domain. Recall the fact that a function in L2 space in the
time domain admits a bilateral Laplace (or Fourier) transform. In fact, it can be shown
that this bilateral Laplace transform yields an isometric isomorphism between the L2
spaces in the time domain and the L2 spaces in the frequency domain (this is what is
called Parseval’s relations):
L2 (−∞, ∞) ∼ = L2 (jR)
L2 [0, ∞) ∼
= H2
L2 (−∞, 0] ∼
= H2⊥ .
As a result, if g(t) ∈ L2 (−∞, ∞) and if its bilateral Laplace transform is G(s) ∈ L2 (jR),
then
kGk2 = kgk2 .
Hence, whenever there is no confusion, the notation for functions in the time domain
and in the frequency domain will be used interchangeably.
Laplace Transform
 -
L2 [0, ∞) H2
Inverse Transform

6 6
P+ P+

Laplace Transform
 -
L2 (−∞, ∞) L2 (jR)
Inverse Transform

P− P−
? ?
Laplace Transform
 -
L2 (−∞, 0] H2⊥
Inverse Transform

Figure 4.1: Relationships among function spaces

Define an orthogonal projection

P+ : L2 (−∞, ∞) 7−→ L2 [0, ∞)
such that, for any function f (t) ∈ L2 (−∞, ∞), we have g(t) = P+ f (t) with

f (t), for t ≥ 0;
g(t) :=
0, for t < 0.
50 H2 AND H∞ SPACES

In this book, P+ will also be used to denote the projection from L2 (jR) onto H2 .
Similarly, define P− as another orthogonal projection from L2 (−∞, ∞) onto L2 (−∞, 0]
(or L2 (jR) onto H2⊥ ). Then the relationships between L2 spaces and H2 spaces can be
shown as in Figure 4.1.
Other classes of important complex matrix functions used in this book are those
bounded on the imaginary axis.

L∞ (jR) Space
L∞ (jR) or simply L∞ is a Banach space of matrix-valued (or scalar-valued) func-
tions that are (essentially) bounded on jR, with norm

kF k∞ := ess sup σ [F (jω)] .

ω∈R

The rational subspace of L∞ , denoted by RL∞ (jR) or simply RL∞ , consists of

all proper and real rational transfer matrices with no poles on the imaginary axis.

H∞ Space
H∞ is a (closed) subspace of L∞ with functions that are analytic and bounded in
the open right-half plane. The H∞ norm is defined as

kF k∞ := sup σ [F (s)] = sup σ [F (jω)] .

Re(s)>0 ω∈R

The second equality can be regarded as a generalization of the maximum modulus

theorem for matrix functions. (See Boyd and Desoer [1985] for a proof.) The real
rational subspace of H∞ is denoted by RH∞ , which consists of all proper and real
rational stable transfer matrices.
−
H∞ Space
−
H∞ is a (closed) subspace of L∞ with functions that are analytic and bounded in
−
the open left-half plane. The H∞ norm is defined as

kF k∞ := sup σ [F (s)] = sup σ [F (jω)] .

Re(s)<0 ω∈R

The real rational subspace of H∞ −

is denoted by RH− ∞ , which consists of all proper,
real rational, antistable transfer matrices (i.e., functions with all poles in the open
right-half plane).

Let G(s) ∈ L∞ be a p × q transfer matrix. Then a multiplication operator is defined as

MG : L2 7−→ L2

MG f := Gf.
4.2. H2 and H∞ Spaces 51

In writing the preceding mapping, we have assumed that f has a compatible dimension.
A more accurate description of the foregoing operator should be

MG : Lq2 7−→ Lp2 .

That is, f is a q-dimensional vector function with each component in L2 . However, we

shall suppress all dimensions in this book and assume that all objects have compatible
dimensions.
A useful fact about the multiplication operator is that the norm of a matrix G in
L∞ equals the norm of the corresponding multiplication operator.
Theorem 4.3 Let G ∈ L∞ be a p × q transfer matrix. Then kMG k = kGk∞ .

Remark 4.1 It is also true that this operator norm equals the norm of the operator
restricted to H2 (or H2⊥ ); that is,

kMG k = kMG |H2 k := sup {kGf k2 : f ∈ H2 , kf k2 ≤ 1} .

This will be clear in the proof where an f ∈ H2 is constructed. 3

Proof. By definition, we have

kMG k = sup {kGf k2 : f ∈ L2 , kf k2 ≤ 1} .

First we see that kGk∞ is an upper bound for the operator norm:
Z ∞
1
f ∗ (jω)G∗ (jω)G(jω)f (jω) dω
2
kGf k2 =
2π −∞
Z ∞
1
≤ kGk2∞ kf (jω)k2 dω
2π −∞
= kGk2∞ kf k22 .

To show that kGk∞ is the least upper bound, first choose a frequency ω0 where σ [G(jω)]
is maximum; that is,
σ [G(jω0 )] = kGk∞ ,
and denote the singular value decomposition of G(jω0 ) by
X
r
G(jω0 ) = σu1 (jω0 )v1∗ (jω0 ) + σi ui (jω0 )vi∗ (jω0 )
i=2

where r is the rank of G(jω0 ) and ui , vi have unit length.

Next we assume that G(s) has real coefficients and we shall construct a function
f (s) ∈ H2 with real coefficients so that the norm is approximately achieved. [It will
be clear in the following that the proof is much simpler if f is allowed to have complex
coefficients, which is necessary when G(s) has complex coefficients.]
52 H2 AND H∞ SPACES

If ω0 < ∞, write v1 (jω0 ) as

 
α1 ejθ1
 α2 ejθ2 
 
v1 (jω0 ) =  .. 
 . 
αq ejθq

where αi ∈ R is such that θi ∈ (−π, 0] and q is the column dimension of G. Now let
0 ≤ βi ≤ ∞ be such that  
βi − jω0
θi = ∠
βi + jω0
(with βi → ∞ if θi = 0) and let f be given by
 
α1 ββ11 −s
+s
 α2 ββ22 −s 
 
f (s) =   f(s)
+s
ˆ
 .. 
 . 
βq −s
αq βq +s

βi −s
(with 1 replacing βi +s if θi = 0), where a scalar function fˆ is chosen so that

ˆ c if |ω − ω0 | <  or |ω + ω0 | < 
|f(jω)| =
0 otherwise

where ˆ
p  is a small positive number and c is chosen so that f has unit 2-norm (i.e.,
c = π/2). This, in turn, implies that f has unit 2-norm. Then

1 h 2 2
i
kGf k22 ≈ σ [G(−jω0 )] π + σ [G(jω0 )] π
2π
= σ [G(jω0 )]2 = kGk2∞ .

Similarly, if ω0 = ∞, the conclusion follows by letting ω0 → ∞ in the foregoing. 2

Illustrative MATLAB Commands:

 [sv, w]=sigma(A, B, C, D); % frequency response of the singular values; or

 w=logspace(l, h, n); sv=sigma(A, B, C, D, w); % n points between 10l and

10h .

Related MATLAB Commands: semilogx, semilogy, bode, freqs, nichols, frsp,

vsvd, vplot, pkvnorm
4.3. Computing L2 and H2 Norms 53

4.3 Computing L2 and H2 Norms

Let G(s) ∈ L2 and recall that the L2 norm of G is defined as
s Z ∞
1
kGk2 := trace{G∗ (jω)G(jω)} dω
2π −∞
= kgk
sZ2
∞
= trace{g ∗ (t)g(t)} dt
−∞

where g(t) denotes the convolution kernel of G.

It is easy to see that the L2 norm defined previously is finite iff the transfer matrix G
is strictly proper; that is, G(∞) = 0. Hence, we will generally assume that the transfer
matrix is strictly proper whenever we refer to the L2 norm of G (of course, this also
applies to H2 norms). One straightforward way of computing the L2 norm is to use
contour integral. Suppose G is strictly proper; then we have
Z ∞
1
kGk2 =
2
trace{G∗ (jω)G(jω)} dω
2π −∞
I
1
= trace{G∼ (s)G(s)} ds.
2πj
The last integral is a contour integral along the imaginary axis and around an infinite
semicircle in the left-half plane; the contribution to the integral from this semicircle
equals zero because G is strictly proper. By the residue theorem, kGk22 equals the sum
of the residues of trace{G∼ (s)G(s)} at its poles in the left-half plane.
Although kGk2 can, in principle, be computed from its definition or from the method
just suggested, it is useful in many applications to have alternative characterizations and
to take advantage of the state-space representations of G. The computation of a RH2
transfer matrix norm is particularly simple.

Lemma 4.4 Consider a transfer matrix

 
A B
G(s) =
C 0

with A stable. Then we have

kGk22 = trace(B ∗ QB) = trace(CP C ∗ ) (4.1)

where Q and P are observability and controllability Gramians that can be obtained from
the following Lyapunov equations:

AP + P A∗ + BB ∗ = 0 A∗ Q + QA + C ∗ C = 0.
54 H2 AND H∞ SPACES

Proof. Since G is stable, we have


−1 CeAt B, t≥0
g(t) = L (G) =
0, t<0
and
Z ∞ Z ∞
kGk22 = trace{g ∗ (t)g(t)} dt =
trace{g(t)g(t)∗ } dt
Z0 ∞ 0
Z ∞
∗ A∗ t ∗ ∗
= At
trace{B e C Ce B} dt = trace{CeAt BB ∗ eA t C ∗ } dt.
0 0

The lemma follows from the fact that the controllability Gramian of (A, B) and the
observability Gramian of (C, A) can be represented as
Z ∞ Z ∞
∗ ∗
Q= eA t C ∗ CeAt dt, P = eAt BB ∗ eA t dt,
0 0

which can also be obtained from

AP + P A∗ + BB ∗ = 0 A∗ Q + QA + C ∗ C = 0.
2

To compute the L2 norm of a rational transfer function, G(s) ∈ RL2 , using the
state-space approach, let G(s) = [G(s)]+ + [G(s)]− with G+ ∈ RH2 and G− ∈ RH⊥ 2;
then
2 2 2
kGk2 = k[G(s)]+ k2 + k[G(s)]− k2
∼
where k[G(s)]+ k2 and k[G(s)]− k2 = k[G(−s)]+ k2 = k([G(s)]− ) k2 can be computed
using the preceding lemma.
Still another useful characterization of the H2 norm of G is in terms of hypothetical
input-output experiments. Let ei denote the ith standard basis vector of Rm , where m
is the input dimension of the system. Apply the impulsive input δ(t)ei [δ(t) is the unit
impulse] and denote the output by zi (t)(= g(t)ei ). Assume D = 0; then zi ∈ L2+ and
X
m
kGk22 = kzi k22 .
i=1

Note that this characterization of the H2 norm can be appropriately generalized for
nonlinear time-varying systems; see Chen and Francis [1992] for an application of this
norm in sampled-data control.

Example 4.1 Consider a transfer matrix

 
3(s + 3) 2
 (s − 1)(s + 2) s−1 =G +G
G= s+1 1  s u

(s + 2)(s + 3) s−4
4.4. Computing L∞ and H∞ Norms 55

with    
−2 0 −1 0 1 0 4 2
 0 −3 2 0   0 4 0 1 
Gs = 
 1
, Gu =  .
0 0 0   1 0 0 0 
1 1 0 0 0 1 0 0
Then the command h2norm(G qs ) gives kGs k2 = 0.6055 and h2norm(cjt(Gu )) gives
2 2
kGu k2 = 3.182. Hence kGk2 = kGs k2 + kGu k2 = 3.2393.

Illustrative MATLAB Commands:

 P = gram(A, B); Q = gram(A0 , C0 ); or P = lyap(A, B ∗ B0 );

 [Gs , Gu ] = sdecomp(G); % decompose into stable and antistable parts.

4.4 Computing L∞ and H∞ Norms

We shall first consider, as in the L2 case, how to compute the ∞ norm of an RL∞
transfer matrix. Let G(s) ∈ RL∞ and recall that the L∞ norm of a matrix rational
transfer function G is defined as

kGk∞ := sup σ{G(jω)}.

ω

The computation of the L∞ norm of G is complicated and requires a search. A control

engineering interpretation of the infinity norm of a scalar transfer function G is the
distance in the complex plane from the origin to the farthest point on the Nyquist plot
of G, and it also appears as the peak value on the Bode magnitude plot of |G(jω)|.
Hence the ∞ norm of a transfer function can, in principle, be obtained graphically.
To get an estimate, set up a fine grid of frequency points:

{ω1 , · · · , ωN }.

Then an estimate for kGk∞ is

max σ{G(jωk )}.

1≤k≤N

This value is usually read directly from a Bode singular value plot. The RL∞ norm can
also be computed in state-space.

Lemma 4.5 Let γ > 0 and

 
A B
G(s) = ∈ RL∞ . (4.2)
C D
56 H2 AND H∞ SPACES

Then kGk∞ < γ if and only if σ(D) < γ and the Hamiltonian matrix H has no eigen-
values on the imaginary axis where
 
A + BR−1 D∗ C BR−1 B ∗
H := (4.3)
−C ∗ (I + DR−1 D∗ )C −(A + BR−1 D∗ C)∗

and R = γ 2 I − D∗ D.

Proof. Let Φ(s) = γ 2 I − G∼ (s)G(s). Then it is clear that kGk∞ < γ if and only if
Φ(jω) > 0 for all ω ∈ R. Since Φ(∞) = R > 0 and since Φ(jω) is a continuous function
of ω, Φ(jω) > 0 for all ω ∈ R if and only if Φ(jω) is nonsingular for all ω ∈ R ∪ {∞};
that is, Φ(s) has no imaginary axis zero. Equivalently, Φ−1 (s) has no imaginary axis
pole. It is easy to compute by some simple algebra that
   
BR−1
 H
−C ∗ DR−1 
Φ−1 (s) =  .
 −1 ∗ 
R D C R−1 B ∗ R−1

Thus the conclusion follows if the above realization has neither uncontrollable modes
nor unobservable modes on the imaginary axis. Assume that jω0 is an eigenvalue
of H but not a pole of Φ−1 (s). Then jω0 must be either anunobservable mode of
  BR−1
( R−1 D∗ C R−1 B ∗ , H) or an uncontrollable mode of (H, ). Now
−C ∗ DR−1
 −1 ∗ −1 ∗

 0 is an unobservable mode of ( R D C R B
suppose jω , H). Then there exists
x1
an x0 = 6= 0 such that
x2
 
Hx0 = jω0 x0 , R−1 D∗ C R−1 B ∗ x0 = 0.

These equations can be simplified to

(jω0 I − A)x1 = 0
(jω0 I + A∗ )x2 = −C ∗ Cx1
D∗ Cx1 + B ∗ x2 = 0.

Since A has no imaginary axis eigenvalues, we have x1 = 0 and x2 = 0. This contradicts

our assumption, and hence the realization has no unobservable modes on the imaginary
axis.
Similarly, a contradiction
 will also
 be arrived at if jω0 is assumed to be an uncon-
BR−1
trollable mode of (H, ). 2
−C ∗ DR−1
4.4. Computing L∞ and H∞ Norms 57

Bisection Algorithm
Lemma 4.5 suggests the following bisection algorithm to compute RL∞ norm:
(a) Select an upper bound γu and a lower bound γl such that γl ≤ kGk∞ ≤ γu ;
(b) If (γu − γl )/γl ≤specified level, stop; kGk ≈ (γu + γl )/2. Otherwise go to the next
step;
(c) Set γ = (γl + γu )/2;
(d) Test if kGk∞ < γ by calculating the eigenvalues of H for the given γ;
(e) If H has an eigenvalue on jR, set γl = γ; otherwise set γu = γ; go back to step
(b).
Of course, the above algorithm applies to H∞ norm computation as well. Thus L∞
norm computation requires a search, over either γ or ω, in contrast to L2 (H2 ) norm
computation, which does not. A somewhat analogous situation occurs for constant
matrices with the norms kM k22 = trace(M ∗ M ) and kM k∞ = σ[M ]. In principle, kM k22
can be computed exactly with a finite number of operations, as can the test for whether
σ(M ) < γ (e.g., γ 2 I − M ∗ M > 0), but the value of σ(M ) cannot. To compute σ(M ),
we must use some type of iterative algorithm.

Remark 4.2 It is clear that kGk∞ < γ iff γ −1 G ∞ < 1. Hence, there is no loss of
generality in assuming γ = 1. This assumption will often be made in the remainder of
this book. It is also noted that there are other fast algorithms to carry out the preceding
norm computation; nevertheless, this bisection algorithm is the simplest. 3
Additional interpretations can be given for the H∞ norm of a stable matrix transfer
function. When G(s) is a single-input and single-output system, the H∞ norm of the
G(s) can be regarded as the largest possible amplification factor of the system’s steady-
state response to sinusoidal excitations. For example, the steady-state response of the
system with respect to a sinusoidal input u(t) = U sin(ω0 t + φ) is
y(t) = U |G(jω0 )| sin (ω0 t + φ + ∠G(jω0 ))
and thus the maximum possible amplification factor is sup |G(jω0 )|, which is precisely
ω0
the H∞ norm of the transfer function.
In the multiple-input and multiple-output case, the H∞ norm of a transfer matrix
G ∈ RH∞ can also be regarded as the largest possible amplification factor of the
system’s steady-state response to sinusoidal excitations in the following sense: Let the
sinusoidal inputs be
   
u1 sin(ω0 t + φ1 ) u1
 u2 sin(ω0 t + φ2 )   u2 
   
u(t) =  ..  , û =  ..  .
 .   . 
uq sin(ω0 t + φq ) uq
58 H2 AND H∞ SPACES

Then the steady-state response of the system can be written as

   
y1 sin(ω0 t + θ1 ) y1
 y2 sin(ω0 t + θ2 )   y2 
   
y(t) =  .. , ŷ =  .. 
 .   . 
yp sin(ω0 t + θp ) yp

for some yi , θi , i = 1, 2, . . . , p, and furthermore,

kŷk
kGk∞ = sup
φi ,ωo ,û kûk

where k·k is the Euclidean norm. The details are left as an exercise.

Example 4.2 Consider a mass/spring/damper system as shown in Figure 4.2.

F1
x1

m1

k1 b1
F2 x2

m2

k2 b2

Figure 4.2: A two-mass/spring/damper system

The dynamical system can be described by the following differential equations:

   
ẋ1 x1  
 ẋ2   x2 
    + B F1
 ẋ3  = A  x3  F2
ẋ4 x4
4.4. Computing L∞ and H∞ Norms 59

2
10

The largest singular value

1
10

0
10

−1
10 The smallest singular value

−2
10 −1 0 1
10 10 10
frequency (rad/sec)

Figure 4.3: kGk∞ is the peak of the largest singular value of G(jω)

with
   
0 0 1 0 0 0
 0 0 0 1   0 0 
   
 k1 k1 b1 b1   1 
A= − − , B =  0 .
 m1 m1 m1 m1   m1 
 k1 k1 + k2 b1 b1 + b2   1 
− − 0
m2 m2 m2 m2 m2

Suppose that G(s) is the transfer matrix from (F1 , F2 ) to (x1 , x2 ); that is,
 
1 0 0 0
C= , D = 0,
0 1 0 0

and suppose k1 = 1, k2 = 4, b1 = 0.2, b2 = 0.1, m1 = 1, and m2 = 2 with appropriate

units. The following Matlab commands generate the singular value Bode plot of the
above system as shown in Figure 4.3.

 G=pck(A,B,C,D);

 hinfnorm(G,0.0001) or linfnorm(G,0.0001) % relative error ≤ 0.0001

 w=logspace(-1,1,200); % 200 points between 1 = 10−1 and 10 = 101;

 Gf=frsp(G,w); % computing frequency response;

 [u,s,v]=vsvd(Gf ); % SVD at each frequency;

60 H2 AND H∞ SPACES

 vplot(0 liv, lm0 , s), grid % plot both singular values and grid.

Then the H∞ norm of this transfer matrix is kG(s)k∞ = 11.47, which is shown as
the peak of the largest singular value Bode plot in Figure 4.3. Since the peak is achieved
at ωmax = 0.8483, exciting the system using the following sinusoidal input
   
F1 0.9614 sin(0.8483t)
=
F2 0.2753 sin(0.8483t − 0.12)

gives the steady-state response of the system as

   
x1 11.47 × 0.9614 sin(0.8483t − 1.5483)
= .
x2 11.47 × 0.2753 sin(0.8483t − 1.4283)

This shows that the system response will be amplified 11.47 times for an input signal
at the frequency ωmax , which could be undesirable if F1 and F2 are disturbance force
and x1 and x2 are the positions to be kept steady.

Example 4.3 Consider a two-by-two transfer matrix

 
10(s + 1) 1
 s2 + 0.2s + 100 s + 1 
G(s) = 
.

s+2 5(s + 1)
s2 + 0.1s + 10 (s + 2)(s + 3)

A state-space realization of G can be obtained using the following Matlab commands:

 G11=nd2sys([10,10],[1,0.2,100]);
 G12=nd2sys(1,[1,1]);
 G21=nd2sys([1,2],[1,0.1,10]);
 G22=nd2sys([5,5],[1,5,6]);
 G=sbs(abv(G11,G21),abv(G12,G22));
Next, we set up a frequency grid to compute the frequency response of G and the
singular values of G(jω) over a suitable range of frequency.
 w=logspace(0,2,200); % 200 points between 1 = 100 and 100 = 102 ;
 Gf=frsp(G,w); % computing frequency response;
 [u,s,v]=vsvd(Gf ); % SVD at each frequency;
4.5. Notes and References 61

 vplot(0 liv, lm0 , s), grid % plot both singular values and grid;

 pkvnorm(s) % find the norm from the frequency response of the singular values.

The singular values of G(jω) are plotted in Figure 4.4, which gives an estimate of
kGk∞ ≈ 32.861. The state-space bisection algorithm described previously leads to
kGk∞ = 50.25 ± 0.01 and the corresponding Matlab command is

 hinfnorm(G,0.0001) or linfnorm(G,0.0001) % relative error ≤ 0.0001.

2
10

1
10

0
10

−1
10

−2
10 0 1 2
10 10 10

Figure 4.4: The largest and the smallest singular values of G(jω)

The preceding computational results show clearly that the graphical method can lead
to a wrong answer for a lightly damped system if the frequency grid is not sufficiently
dense. Indeed, we would get kGk∞ ≈ 43.525, 48.286 and 49.737 from the graphical
method if 400, 800, and 1600 frequency points are used, respectively.

Related MATLAB Commands: linfnorm, vnorm, getiv, scliv, var2con, xtract,

xtracti

4.5 Notes and References

The basic concept of function spaces presented in this chapter can be found in any
standard functional analysis textbook, for instance, Naylor and Sell [1982] and Gohberg
and Goldberg [1981]. The system theoretical interpretations of the norms and function
62 H2 AND H∞ SPACES

spaces can be found in Desoer and Vidyasagar [1975]. The bisection L∞ norm compu-
tational algorithm was first developed in Boyd, Balakrishnan, and Kabamba [1989]. A
more efficient L∞ norm computational algorithm is presented in Bruinsma and Stein-
buch [1990].

4.6 Problems
Problem 4.1 Let G(s) be a matrix in RH∞ . Prove that
 
G 2

I = kGk∞ + 1.
2

Problem 4.2 (Parseval relation) Let f (t), g(t) ∈ L2 , F (jω) = F{f (t)}, and G(jω) =
F{g(t)}. Show that
Z ∞ Z ∞
1
f (t)g(t)dt = F (jω)G∗ (jω)dω
−∞ 2π −∞
and Z ∞ Z ∞
1
|f (t)|2 dt = |F (jω)|2 dω.
−∞ 2π −∞
Note that
Z ∞ Z ∞
1
F (jω) = f (t)e−jωt dt, f (t) = F −1 (F (jω)) = F (jω)ejωt dω.
−∞ 2π −∞
−1
where F denotes the inverse Fourier transform.
Problem 4.3 Suppose A is stable. Show
Z ∞
(jωI − A)−1 dω = πI.
−∞
 
A B
Suppose G(s) = ∈ RH∞ and let Q = Q∗ be the observability Gramian. Use
C 0
the above formula to show that
Z ∞
1
G∼ (jω)G(jω)dω = B ∗ QB.
2π −∞
[Hint: Use the fact that G∼ (s)G(s) = F ∼ (s) + F (s) and F (s) = B ∗ Q(sI − A)−1 B.]
Problem 4.4 Compute the 2-norm and ∞-norm of the following systems:
 1 
s+3  
 s + 1 (s + 1)(s − 2)  1 0 1
G1 (s) = 
 10
 , G2 (s) =  2
 3 1 
5 1 2 0
s−2 s+3
4.6. Problems 63

 
  −1 −2 −3 1 2
−1 −2 1  1 0 0 0 1 
 
G3 (s) =  1 0 0 , G4 (s) = 
 0 1 0 2 0 
.
2 3 0  1 0 0 1 0 
0 1 1 0 2
Problem 4.5 Let r(t) = sin ωt be the input signal to a plant
ωn2
G(s) =
s2 + 2ξωn s + ωn2
√
with 0 < ξ < 1/ 2. Find the steady-state response of the system y(t). Also find the
frequency ω that gives the largest magnitude steady-state response of y(t).
Problem 4.6 Let G(s) ∈ RH∞ be a p × q transfer matrix and y = G(s)u. Suppose
   
u1 sin(ω0 t + φ1 ) u1
 u2 sin(ω0 t + φ2 )   u2 
   
u(t) =  ..  , û =  ..  .
 .   . 
uq sin(ω0 t + φq ) uq
Show that the steady-state response of the system is given by
   
y1 sin(ω0 t + θ1 ) y1
 y2 sin(ω0 t + θ2 )   y2 
   
y(t) =  ..  , ŷ =  .. 
 .   . 
yp sin(ω0 t + θp ) yp
for some yi and θi , i = 1, 2, . . . , p. Show that sup kŷk2 = kGk∞ .
φi ,ωo ,kûk2 ≤1

Problem 4.7 Write a Matlab program to plot, versus γ, the distance from the imag-
inary axis to the nearest eigenvalue of the Hamiltonian matrix for a given state-space
model with stable A. Try it on
 s+1 s 
 (s + 2)(s + 3) s+1 
 
 .
 s2 − 2 s+4 
(s + 3)(s + 4) (s + 1)(s + 2)
Read off the value of the H∞ -norm. Compare with the Matlab function hinfnorm.
1
Problem 4.8 Let G(s) = . Compute kG(s)k∞ using the Bode
(s2 + 2ξs + 1)(s + 1)
plot and state-space algorithm, respectively for ξ = 1, 0.1, 0.01, 0.001 and compare the
results.
64 H2 AND H∞ SPACES
Chapter 5

Internal Stability

This chapter introduces the feedback structure and discusses its stability and various
stability tests. The arrangement of this chapter is as follows: Section 5.1 discusses the
necessity for introducing feedback structure and describes the general feedback con-
figuration. Section 5.2 defines the well-posedness of the feedback loop. Section 5.3
introduces the notion of internal stability and various stability tests. Section 5.4 intro-
duces the stable coprime factorizations of rational matrices. The stability conditions in
terms of various coprime factorizations are also considered in this section.

5.1 Feedback Structure

In designing control systems, there are several fundamental issues that transcend the
boundaries of specific applications. Although they may differ for each application and
may have different levels of importance, these issues are generic in their relationship to
control design objectives and procedures. Central to these issues is the requirement to
provide satisfactory performance in the face of modeling errors, system variations, and
uncertainty. Indeed, this requirement was the original motivation for the development
of feedback systems. Feedback is only required when system performance cannot be
achieved because of uncertainty in system characteristics. A more detailed treatment
of model uncertainties and their representations will be discussed in Chapter 8.
For the moment, assuming we are given a model including a representation of un-
certainty that we believe adequately captures the essential features of the plant, the
next step in the controller design process is to determine what structure is necessary
to achieve the desired performance. Prefiltering input signals (or open-loop control)
can change the dynamic response of the model set but cannot reduce the effect of un-
certainty. If the uncertainty is too great to achieve the desired accuracy of response,
then a feedback structure is required. The mere assumption of a feedback structure,
however, does not guarantee a reduction of uncertainty, and there are many obstacles
to achieving the uncertainty-reducing benefits of feedback. In particular, since for any

65
66 INTERNAL STABILITY

reasonable model set representing a physical system uncertainty becomes large and the
phase is completely unknown at sufficiently high frequencies, the loop gain must be
small at those frequencies to avoid destabilizing the high-frequency system dynamics.
Even worse is that the feedback system actually increases uncertainty and sensitivity in
the frequency ranges where uncertainty is significantly large. In other words, because
of the type of sets required to model physical systems reasonably and because of the
restriction that our controllers be causal, we cannot use feedback (or any other control
structure) to cause our closed-loop model set to be a proper subset of the open-loop
model set. Often, what can be achieved with intelligent use of feedback is a signifi-
cant reduction of uncertainty for certain signals of importance with a small increase
spread over other signals. Thus, the feedback design problem centers around the trade-
off involved in reducing the overall impact of uncertainty. This tradeoff also occurs, for
example, when using feedback to reduce command/disturbance error while minimizing
response degradation due to measurement noise. To be of practical value, a design
technique must provide means for performing these tradeoffs. We shall discuss these
tradeoffs in more detail in the next chapter.
To focus our discussion, we shall consider the standard feedback configuration shown
in Figure 5.1. It consists of the interconnected plant P and controller K forced by
command r, sensor noise n, plant input disturbance di , and plant output disturbance
d. In general, all signals are assumed to be multivariable, and all transfer matrices are
assumed to have appropriate dimensions.

di d
r-e - K u ? u
- e p- P ?
- e y-
− 6

?
e
n

Figure 5.1: Standard feedback configuration

5.2 Well-Posedness of Feedback Loop

Assume that the plant P and the controller K in Figure 5.1 are fixed real rational
proper transfer matrices. Then the first question one would ask is whether the feedback
interconnection makes sense or is physically realizable. To be more specific, consider a
simple example where
s−1
P =− , K=1
s+2
5.2. Well-Posedness of Feedback Loop 67

are both proper transfer functions. However,

(s + 2) s−1
u= (r − n − d) + di .
3 3
That is, the transfer functions from the external signals r − n − d and di to u are not
proper. Hence, the feedback system is not physically realizable.

Definition 5.1 A feedback system is said to be well-posed if all closed-loop transfer

matrices are well-defined and proper.

Now suppose that all the external signals r, n, d, and di are specified and that the
closed-loop transfer matrices from them to u are, respectively, well-defined and proper.
Then y and all other signals are also well-defined and the related transfer matrices are
proper. Furthermore, since the transfer matrices from d and n to u are the same and
differ from the transfer matrix from r to u by only a sign, the system is well-posed if
di
and only if the transfer matrix from to u exists and is proper.
d
To be consistent with the notation used in the rest of this book, we shall denote

K̂ := −K (5.1)

and regroup the external input signals into the feedback loop as w1 and w2 and regroup
the input signals of the plant and the controller as e1 and e2 . Then the feedback loop
with the plant and the controller can be simply represented as inFigure 5.2 and the
w1
system is well-posed if and only if the transfer matrix from to e1 exists and is
w2
proper.
w1 e1
-e - P
+ 6
+

e2 +
? + w2
K̂  e

Figure 5.2: Internal stability analysis diagram

Lemma 5.1 The feedback system in Figure 5.2 is well-posed if and only if

I − K̂(∞)P (∞) (5.2)

is invertible.
68 INTERNAL STABILITY

Proof. The system in Figure 5.2 can be represented in equation form as

e1 = w1 + K̂e2
e2 = w2 + P e1 .
Then an expression for e1 can be obtained as
(I − K̂P )e1 = w1 + K̂w2 .
Thus well-posedness is equivalent to the condition that (I − K̂P )−1 exists and is proper.
But this is equivalent to the condition that the constant term of the transfer function
I − K̂P is invertible. 2

It is straightforward to show that equation (5.2) is equivalent to either one of the

following two conditions:
 
I −K̂(∞)
is invertible; (5.3)
−P (∞) I

I − P (∞)K̂(∞) is invertible.
The well-posedness condition is simple to state in terms of state-space realizations.
Introduce realizations of P and K̂:
  " #
A B Â B̂
P = , K̂ = .
C D Ĉ D̂

Then P (∞) = D, K̂(∞) = D̂ and  the well-posedness

 condition in equation (5.3) is
I −D̂
equivalent to the invertibility of . Fortunately, in most practical cases
−D I
we shall have D = 0, and hence well-posedness for most practical control systems is
guaranteed.

5.3 Internal Stability

Consider a system described by the standard block diagram in Figure 5.2 and assume
that the system is well-posed.
Definition 5.2 The system of Figure 5.2 is said to be internally stable if the transfer
matrix
 −1  
I −K̂ (I − K̂P )−1 K̂(I − P K̂)−1
= (5.4)
−P I P (I − K̂P )−1 (I − P K̂)−1
 
I + K̂(I − P K̂)−1 P K̂(I − P K̂)−1
=
(I − P K̂)−1 P (I − P K̂)−1
from (w1 , w2 ) to (e1 , e2 ) belongs to RH∞ .
5.3. Internal Stability 69

Note that to check internal stability, it is necessary (and sufficient) to test whether each
of the four transfer matrices in equation (5.4) is in H∞ . Stability cannot be concluded
even if three of the four transfer matrices in equation (5.4) are in H∞ . For example, let
an interconnected system transfer function be given by
s−1 1
P = , K̂ = − .
s+1 s−1
Then it is easy to compute
 s+1 s+1 
  −  
e1  s+2 (s − 1)(s + 2)  w1
=


 w2 ,
e2 s−1 s+1
s+2 s+2
which shows that the system is not internally stable although three of the four transfer
functions are stable.

Remark 5.1 Internal stability is a basic requirement for a practical feedback system.
This is because all interconnected systems may be unavoidably subject to some nonzero
initial conditions and some (possibly small) errors, and it cannot be tolerated in practice
that such errors at some locations will lead to unbounded signals at some other locations
in the closed-loop system. Internal stability guarantees that all signals in a system are
bounded provided that the injected signals (at any locations) are bounded. 3

However, there are some special cases under which determining system stability is
simple.
Corollary 5.2 Suppose K̂ ∈ RH∞ . Then the system in Figure 5.2 is internally stable
if and only if it is well-posed and P (I − K̂P )−1 ∈ RH∞ .

Proof. The necessity is obvious. To prove the sufficiency, it is sufficient to show that
(I − P K̂)−1 ∈ RH∞ . But this follows from

(I − P K̂)−1 = I + (I − P K̂)−1 P K̂

and (I − P K̂)−1 P, K̂ ∈ RH∞ . 2

Also, we have the following:

Corollary 5.3 Suppose P ∈ RH∞ . Then the system in Figure 5.2 is internally stable
if and only if it is well-posed and K̂(I − P K̂)−1 ∈ RH∞ .

Corollary 5.4 Suppose P ∈ RH∞ and K̂ ∈ RH∞ . Then the system in Figure 5.2 is
internally stable if and only if (I − P K̂)−1 ∈ RH∞ , or, equivalently, det(I − P (s)K̂(s))
has no zeros in the closed right-half plane.
70 INTERNAL STABILITY

Note that all the previous discussions and conclusions apply equally to infinite di-
mensional plants and controllers. To study the more general case, we shall limit our
discussions to finite dimensional systems and define

nk := number of open right-half plane (rhp) poles of K̂(s)

np := number of open right-half plane (rhp) poles of P (s).

Theorem 5.5 The system is internally stable if and only if it is well-posed and
(i) the number of open rhp poles of P (s)K̂(s) = nk + np ;

(ii) (I − P (s)K̂(s))−1 is stable.

Proof. It is easy to show that P K̂ and (I − P K̂)−1 have the following realizations:
 
A B Ĉ B D̂ " #
 0  −1
B̂  , (I − P K̂) =
Ā B̄
P K̂ =  Â
C̄ D̄
C DĈ DD̂
where
   
A B Ĉ B D̂  
Ā = + (I − DD̂)−1 C DĈ
0 Â B̂
 
B D̂
B̄ = (I − DD̂)−1
B̂
 
C̄ = (I − DD̂)−1 C DĈ
D̄ = (I − DD̂)−1 .

Hence, the system is internally stable iff Ā is stable. (see Problem 5.2.)
Now suppose that the system is internally stable; then (I − P K̂)−1 ∈ RH∞ . So we
only need to show that given condition (ii), condition (i) is necessary and sufficient for
the internal stability. This follows by noting that (Ā, B̄) is stabilizable iff
   
A B Ĉ B D̂
, (5.5)
0 Â B̂

is stabilizable; and (C̄, Ā) is detectable iff

  
  A B Ĉ
C DĈ , (5.6)
0 Â

is detectable. But conditions (5.5) and (5.6) are equivalent to condition (i). 2

Condition (i) in the preceding theorem implies that there is no unstable pole/zero
cancellation in forming the product P K̂.
5.4. Coprime Factorization over RH∞ 71

The preceding theorem is, in fact, the basis for the classical control theory, where
the stability is checked only for one closed-loop transfer function with the implicit
assumption that the controller itself is stable (and most probably also minimum phase;
or at least marginally stable and minimum phase with the condition that any imaginary
axis pole of the controller is not in the same location as any zero of the plant).

Example 5.1 Let P and K̂ be two-by-two transfer matrices


1   1−s 
0 −1
 s−1 
P = , K̂ =  s + 1 .
1
0 0 −1
s+1
Then
   
−1 −1 s+1 (s + 1)2
−
 s+1 s−1   s+2 (s + 2)2 (s − 1) 
P K̂ = 

, (I − P K̂)−1 = .
−1  
s+1

0 0
s+1 s+2
So the closed-loop system is not stable even though

(s + 2)2
det(I − P K̂) =
(s + 1)2

has no zero in the closed right-half plane and the number of unstable poles of P K̂ =
nk + np = 1. Hence, in general, det(I − P K̂) having no zeros in the closed right-half
plane does not necessarily imply (I − P K̂)−1 ∈ RH∞ .

5.4 Coprime Factorization over RH∞

Recall that two polynomials m(s) and n(s), with, for example, real coefficients, are said
to be coprime if their greatest common divisor is 1 (equivalent, they have no common
zeros). It follows from Euclid’s algorithm1 that two polynomials m and n are coprime
iff there exist polynomials x(s) and y(s) such that xm + yn = 1; such an equation is
called a Bezout identity. Similarly, two transfer functions m(s) and n(s) in RH∞ are
said to be coprime over RH∞ if there exists x, y ∈ RH∞ such that

xm + yn = 1.
1 See, for example, Kailath [1980], pages 140–141.
72 INTERNAL STABILITY

The more primitive, but equivalent, definition is that m and n are coprime if every
common divisor of m and n is invertible in RH∞ ; that is,

h, mh−1 , nh−1 ∈ RH∞ =⇒ h−1 ∈ RH∞ .

More generally, we have the following:

Definition 5.3 Two matrices M and N in RH∞ are right coprime over RH∞ if they
have the same number of columns and if there exist matrices Xr and Yr in RH∞ such
that  
  M
Xr Yr = Xr M + Yr N = I.
N
Similarly, two matrices M̃ and Ñ in RH∞ are left coprime over RH∞ if they have the
same number of rows and if there exist matrices Xl and Yl in RH∞ such that
 
  Xl
M̃ Ñ = M̃ Xl + ÑYl = I.
Yl
 
M
Note that these definitions are equivalent to saying that the matrix is left in-
N
 
vertible in RH∞ and the matrix M̃ Ñ is right invertible in RH∞ . These two
equations are often called Bezout identities.
Now let P be a proper real rational matrix. A right coprime factorization (rcf )
of P is a factorization P = N M −1 , where N and M are right coprime over RH∞ .
Similarly, a left coprime factorization (lcf ) has the form P = M̃ −1 Ñ , where Ñ and M̃
are left-coprime over RH∞ . A matrix P (s) ∈ Rp (s) is said to have double coprime
factorization if there exist a right coprime factorization P = N M −1 , a left coprime
factorization P = M̃ −1 Ñ , and Xr , Yr , Xl , Yl ∈ RH∞ such that
  
Xr Yr M −Yl
= I. (5.7)
−Ñ M̃ N Xl

Of course, implicit in these definitions is the requirement that both M and M̃ be square
and nonsingular.

Theorem 5.6 Suppose P (s) is a proper real rational matrix and

 
A B
P =
C D

is a stabilizable and detectable realization. Let F and L be such that A+BF and A+LC
are both stable, and define
 
  A + BF B −L
M −Yl
= F I 0  (5.8)
N Xl
C + DF D I
5.4. Coprime Factorization over RH∞ 73

 
  A + LC −(B + LD) L
Xr Yr
= F I 0 . (5.9)
−Ñ M̃
C −D I

Then P = N M −1 = M̃ −1 Ñ are rcf and lcf, respectively, and, furthermore, equation

(5.7) is satisfied.

Proof. The theorem follows by verifying equation (5.7). 2

Remark 5.2 Note that if P is stable, then we can take Xr = Xl = I, Yr = Yl = 0,

N = Ñ = P , M = M̃ = I. 3

Remark 5.3 The coprime factorization of a transfer matrix can be given a feedback-
control interpretation. For example, right coprime factorization comes out naturally
from changing the control variable by a state feedback. Consider the state-space equa-
tions for a plant P :

ẋ = Ax + Bu
y = Cx + Du.

Next, introduce a state feedback and change the variable

v := u − F x

where F is such that A + BF is stable. Then we get

ẋ = (A + BF )x + Bv
u = Fx + v
y = (C + DF )x + Dv.

Evidently, from these equations, the transfer matrix from v to u is

 
A + BF B
M (s) =
F I

and that from v to y is

 
A + BF B
N (s) =
C + DF D
Therefore,
u = M v, y = N v
so that y = N M −1 u; that is, P = N M −1 . 3
74 INTERNAL STABILITY

We shall now see how coprime factorizations can be used to obtain alternative charac-
terizations of internal stability conditions. Consider again the standard stability analysis
diagram in Figure 5.2. We begin with any rcf’s and lcf’s of P and K̂:

P = N M −1 = M̃ −1 Ñ (5.10)

K̂ = U V −1 = Ṽ −1 Ũ . (5.11)

Lemma 5.7 Consider the system in Figure 5.2. The following conditions are equiva-
lent:
1. The feedback system is internally stable.
 
M U
2. is invertible in RH∞ .
N V
 
Ṽ −Ũ
3. is invertible in RH∞ .
−Ñ M̃

4. M̃V − ÑU is invertible in RH∞ .

5. Ṽ M − ŨN is invertible in RH∞ .

Proof. Note that the system is internally stable if

 −1
I −K̂
∈ RH∞
−P I

or, equivalently,
 −1
I K̂
∈ RH∞ (5.12)
P I
Now       
I K̂ I U V −1 M U M −1 0
= =
P I N M −1 I N V 0 V −1
so that  −1   −1
I K̂ M 0 M U
=
P I 0 V N V
Since the matrices    
M 0 M U
,
0 V N V
are right coprime (this fact is left as an exercise for the reader), equation (5.12) holds
iff  −1
M U
∈ RH∞
N V
5.4. Coprime Factorization over RH∞ 75

This proves the equivalence of conditions 1 and 2. The equivalence of conditions 1 and
3 is proved similarly.
Conditions 4 and 5 are implied by conditions 2 and 3 from the following equation:
    
Ṽ −Ũ M U Ṽ M − Ũ N 0
=
−Ñ M̃ N V 0 M̃ V − Ñ U

Since the left-hand side of the above equation is invertible in RH∞ , so is the right-hand
side. Hence, conditions 4 and 5 are satisfied. We only need to show that either condition
4 or condition 5 implies condition 1. Let us show that condition 5 implies condition 1;
this is obvious since
 −1  −1
I K̂ I Ṽ −1 Ũ
=
P I N M −1 I
  −1  
M 0 Ṽ M Ũ Ṽ 0
= ∈ RH∞
0 I N I 0 I
 −1
Ṽ M Ũ
if ∈ RH∞ or if condition 5 is satisfied. 2
N I

Combining Lemma 5.7 and Theorem 5.6, we have the following corollary.

Corollary 5.8 Let P be a proper real rational matrix and P = N M −1 = M̃ −1 Ñ be the

corresponding rcf and lcf over RH∞ . Then there exists a controller

K̂0 = U0 V0−1 = Ṽ0−1 Ũ0

with U0 , V0 , Ũ0 , and Ṽ0 in RH∞ such that

    
Ṽ0 −Ũ0 M U0 I 0
= (5.13)
−Ñ M̃ N V0 0 I

Furthermore, let F and L be such that A+BF and A+LC are stable. Then a particular
set of state-space realizations for these matrices can be given by
 
  A + BF B −L
M U0
= F I 0  (5.14)
N V0
C + DF D I
 
  A + LC −(B + LD) L
Ṽ0 −Ũ0
= F I 0  (5.15)
−Ñ M̃
C −D I
76 INTERNAL STABILITY

Proof. The idea behind the choice of these matrices is as follows. Using the observer
theory, find a controller K̂0 achieving internal stability; for example
 
A + BF + LC + LDF −L
K̂0 := (5.16)
F 0

Perform factorizations
K̂0 = U0 V0−1 = Ṽ0−1 Ũ0 ,
which are analogous to the ones performed on P . Then Lemma 5.7 implies that each
of the two left-hand side block matrices of equation (5.13) must be invertible in RH∞ .
In fact, equation (5.13) is satisfied by comparing it with equation (5.7). 2

Finding a coprime factorization for a scalar transfer function is fairly easy. Let
P (s) = num(s)/den(s) where num(s) and den(s) are the numerator and the denomi-
nator polynomials of P (s), and let α(s) be a stable polynomial of the same order as
den(s). Then P (s) = n(s)/m(s) with n(s) = num(s)/α(s) and m(s) = den(s)/α(s) is
a coprime factorization. However, finding an x(s) ∈ H∞ and a y(s) ∈ H∞ such that
x(s)n(s) + y(s)m(s) = 1 needs much more work.

s−2
Example 5.2 Let P (s) = and α = (s + 1)(s + 3). Then P (s) = n(s)/m(s)
s(s + 3)
s−2 s
with n(s) = and m(s) = forms a coprime factorization. To find an
(s + 1)(s + 3) s+1
x(s) ∈ H∞ and a y(s) ∈ H∞ such that x(s)n(s) + y(s)m(s) = 1, consider a stabilizing
s−1
controller for P : K̂ = − . Then K̂ = u/v with u = K̂ and v = 1 is a coprime
s + 10
factorization and
(s + 11.7085)(s + 2.214)(s + 0.077)
m(s)v(s) − n(s)u(s) = =: β(s)
(s + 1)(s + 3)(s + 10)

Then we can take

(s − 1)(s + 1)(s + 3)
x(s) = −u(s)/β(s) =
(s + 11.7085)(s + 2.214)(s + 0.077)

(s + 1)(s + 3)(s + 10)

y(s) = v(s)/β(s) =
(s + 11.7085)(s + 2.214)(s + 0.077)

Matlab programs can be used to find the appropriate F and L matrices in state-
space so that the desired coprime factorization can be obtained. Let A ∈ Rn×n , B ∈
Rn×m and C ∈ Rp×n . Then an F and an L can be obtained from
 F=-lqr(A, B, eye(n), eye(m)); % or
5.5. Notes and References 77

 F=-place(A, B, Pf ); % Pf= poles of A+BF

 L = −lqr(A0 , C0 , eye(n), eye(p))0 ; % or
 L = −place(A0 , C0 , Pl)0 ; % Pl=poles of A+LC.

5.5 Notes and References

The presentation of this chapter is based primarily on Doyle [1984]. The discussion of in-
ternal stability and coprime factorization can also be found in Francis [1987], Vidyasagar
[1985], and Nett, Jacobson, and Balas [1984].

5.6 Problems
Problem 5.1 Recall that a feedback system is said to be internally stable if all closed-
loop transfer functions are stable. Describe the conditions for internal stability of the
following feedback system:
d
r ?
- d - G1 - d - G2
−6

? n
H  d

How can the stability conditions be simplified if H(s) and G1 (s) are both stable?
 −1
I −K̂
Problem 5.2 Show that ∈ RH∞ if and only if
−P I
   
A B Ĉ B D̂  
Ā := + (I − DD̂)−1 C DĈ
0 Â B̂
is stable.
Problem 5.3 Suppose N, M, U, V ∈ RH∞ and N M −1 and U V −1 are right coprime
factorizations, respectively. Show that
  −1
M 0 M U
0 V N V
is also a right coprime factorization.
s−1
Problem 5.4 Let G(s) = . Find a stable coprime factorization G =
(s + 2)(s − 3)
n(s)/m(s) and x, y ∈ RH∞ such that xn + ym = 1.
78 INTERNAL STABILITY

(s − 1)(s + α) (s − 3)(s + α)
Problem 5.5 Let N (s) = and M (s) = . Show that
(s + 2)(s + 3)(s + β) (s + 3)(s + β)
(N, M ) is also a coprime factorization of the G in Problem 5.4 for any α > 0 and β > 0.

Problem 5.6 Let G = N M −1 be a right coprime factorization over RH∞ . It is called

a normalized coprime factorization if N ∼ N + M ∼ M = I. Now consider scalar transfer
function G. Then the following procedure can be used to find a normalized coprime
factorization: (a) Let G = n/m be any coprime factorization over RH∞ . (b) Find
a stable and minimum phase spectral factor w such that w∼ w = n∼ n + m∼ m. Let
N = n/w and M = m/w; then G = N/M is a normalized coprime factorization. Find
a normalized coprime factorization for Problem 5.4.

Problem 5.7 The following procedure constructs a normalized right coprime factor-
ization when G is strictly proper:

1. Get a stabilizable, detectable realization A, B, C.

2. Do the Matlab command F = −lqr(A, B, C 0 C, I).

3. Set  
  A + BF B
N
(s) =  C 0 
M
F I

Verify that the procedure produces factors that satisfy G = N M −1 . Now try the
procedure on
 
1 1
 s−1 s−2 
G(s) = 
 2

1 
s s+2
Verify numerically that

N (jω)∗ N (jω) + M (jω)∗ M (jω) = I, ∀ω. (5.17)

Problem 5.8 Use the procedure in Problem 5.7 to find the normalized right coprime
factorization for  1 
s+3
 s + 1 (s + 1)(s − 2) 
G1 (s) = 
 10


5
s−2 s+3
 
2(s + 1)(s + 2) s+2
G2 (s) =
s(s + 3)(s + 4) (s + 1)(s + 3)
5.6. Problems 79

 
−1 −2 1 1 2 3
 0 2 −1 3 2 1 
 
G3 (s) =  −4 −3 −2 1 1 1 

 1 1 1 0 0 0 
2 3 4 0 0 0
 
−1 −2 1 2
 0 1 2 1 
G4 (s) = 
 1

1 0 0 
1 1 0 0

Problem 5.9 Define the normalized left coprime factorization and describe a procedure
to find such factorizations for strictly proper transfer matrices.
80 INTERNAL STABILITY
Chapter 6

Performance Specifications
and Limitations

In this chapter, we consider further the feedback system properties and discuss how to
achieve desired performance using feedback control. We also consider the mathematical
formulations of optimal H2 and H∞ control problems. A key step in the optimal
control design is the selection of weighting functions. We shall give some guidelines to
such selection process using some SISO examples. We shall also discuss in some detail
the design limitations imposed by bandwidth constraints, the open-loop right-half plane
zeros, and the open-loop right-half plane poles using Bode’s gain and phase relation,
Bode’s sensitivity integral relation, and the Poisson integral formula.

6.1 Feedback Properties

In this section, we discuss the properties of a feedback system. In particular, we consider
the benefit of the feedback structure and the concept of design tradeoffs for conflicting
objectives — namely, how to achieve the benefits of feedback in the face of uncertainties.

di d
r-e - K u ? u
- e p- P ?
- e y-
− 6

?
e
n

Figure 6.1: Standard feedback configuration

81
82 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

Consider again the feedback system shown in Figure 5.1. For convenience, the system
diagram is shown again in Figure 6.1. For further discussion, it is convenient to define
the input loop transfer matrix, Li , and output loop transfer matrix, Lo , as

Li = KP, Lo = P K,

respectively, where Li is obtained from breaking the loop at the input (u) of the plant
while Lo is obtained from breaking the loop at the output (y) of the plant. The input
sensitivity matrix is defined as the transfer matrix from di to up :

Si = (I + Li )−1 , up = Si di .

The output sensitivity matrix is defined as the transfer matrix from d to y:

So = (I + Lo )−1 , y = So d.

The input and output complementary sensitivity matrices are defined as

Ti = I − Si = Li (I + Li )−1

To = I − So = Lo (I + Lo )−1 ,
respectively. (The word complementary is used to signify the fact that T is the comple-
ment of S, T = I − S.) The matrix I + Li is called the input return difference matrix
and I + Lo is called the output return difference matrix.
It is easy to see that the closed-loop system, if it is internally stable, satisfies the
following equations:

y = To (r − n) + So P di + So d (6.1)
r − y = So (r − d) + To n − So P di (6.2)
u = KSo (r − n) − KSo d − Ti di (6.3)
up = KSo (r − n) − KSo d + Si di . (6.4)

These four equations show the fundamental benefits and design objectives inherent in
feedback loops. For example, equation (6.1) shows that the effects of disturbance d on
the plant output can be made “small” by making the output sensitivity function So
small. Similarly, equation (6.4) shows that the effects of disturbance di on the plant
input can be made small by making the input sensitivity function Si small. The notion
of smallness for a transfer matrix in a certain range of frequencies can be made explicit
using frequency-dependent singular values, for example, σ(So ) < 1 over a frequency
range would mean that the effects of disturbance d at the plant output are effectively
desensitized over that frequency range.
Hence, good disturbance rejection at the plant output (y) would require that

1
σ(So ) = σ (I + P K)−1 = (for disturbance at plant output, d),
σ(I + P K)


σ(So P ) = σ (I + P K)−1 P = σ(P Si ) (for disturbance at plant input, di )
6.1. Feedback Properties 83

be made small and good disturbance rejection at the plant input (up ) would require
that

1
σ(Si ) = σ (I + KP )−1 = (for disturbance at plant input, di ),
σ(I + KP )


σ(Si K) = σ K(I + P K)−1 = σ(KSo ) (for disturbance at plant output, d)

be made small, particularly in the low-frequency range where d and di are usually
significant.
Note that

σ(P K) − 1 ≤ σ(I + P K) ≤ σ(P K) + 1

σ(KP ) − 1 ≤ σ(I + KP ) ≤ σ(KP ) + 1

then
1 1
≤ σ(So ) ≤ , if σ(P K) > 1
σ(P K) + 1 σ(P K) − 1
1 1
≤ σ(Si ) ≤ , if σ(KP ) > 1
σ(KP ) + 1 σ(KP ) − 1
These equations imply that

σ(So )  1 ⇐⇒ σ(P K)  1
σ(Si )  1 ⇐⇒ σ(KP )  1.

Now suppose P and K are invertible; then

1
σ(P K)  1 or σ(KP )  1 ⇐⇒ σ(So P ) = σ (I + P K)−1 P ≈ σ(K −1 ) =
σ(K)

1
σ(P K)  1 or σ(KP )  1 ⇐⇒ σ(KSo ) = σ K(I + P K)−1 ≈ σ(P −1 ) =
σ(P )
Hence good performance at plant output (y) requires, in general, large output loop gain
σ(Lo ) = σ(P K)  1 in the frequency range where d is significant for desensitizing d and
large enough controller gain σ(K)  1 in the frequency range where di is significant for
desensitizing di . Similarly, good performance at plant input (up ) requires, in general,
large input loop gain σ(Li ) = σ(KP )  1 in the frequency range where di is significant
for desensitizing di and large enough plant gain σ(P )  1 in the frequency range where
d is significant, which cannot be changed by controller design, for desensitizing d. [In
general, So 6= Si unless K and P are square and diagonal, which is true if P is a scalar
system. Hence, small σ(So ) does not necessarily imply small σ(Si ); in other words,
good disturbance rejection at the output does not necessarily mean good disturbance
rejection at the plant input.]
Hence, good multivariable feedback loop design boils down to achieving high loop (and
possibly controller) gains in the necessary frequency range.
84 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

Despite the simplicity of this statement, feedback design is by no means trivial.

This is true because loop gains cannot be made arbitrarily high over arbitrarily large
frequency ranges. Rather, they must satisfy certain performance tradeoff and design
limitations. A major performance tradeoff, for example, concerns commands and dis-
turbance error reduction versus stability under the model uncertainty. Assume that the
plant model is perturbed to (I + ∆)P with ∆ stable, and assume that the system is
nominally stable (i.e., the closed-loop system with ∆ = 0 is stable). Now the perturbed
closed-loop system is stable if
det (I + (I + ∆)P K) = det(I + P K) det(I + ∆To )
has no right-half plane zero. This would, in general, amount to requiring that k∆To k
be small or that σ̄(To ) be small at those frequencies where ∆ is significant, typically at
high-frequency range, which, in turn, implies that the loop gain, σ(Lo ), should be small
at those frequencies.
Still another tradeoff is with the sensor noise error reduction. The conflict between
the disturbance rejection and the sensor noise reduction is evident in equation (6.1).
Large σ(Lo (jω)) values over a large frequency range make errors due to d small. How-
ever, they also make errors due to n large because this noise is “passed through” over
the same frequency range, that is,
y = To (r − n) + So P di + So d ≈ (r − n)
Note that n is typically significant in the high-frequency range. Worst still, large loop
gains outside of the bandwidth of P — that is, σ(Lo (jω))  1 or σ(Li (jω))  1 while
σ(P (jω))  1 — can make the control activity (u) quite unacceptable, which may cause
the saturation of actuators. This follows from
u = KSo (r − n − d) − Ti di = Si K(r − n − d) − Ti di ≈ P −1 (r − n − d) − di
Here, we have assumed P to be square and invertible for convenience. The resulting
equation shows that disturbances and sensor noise are actually amplified at u whenever
the frequency range significantly exceeds the bandwidth of P , since for ω such that
σ(P (jω))  1 we have
1
σ[P −1 (jω)] = 1
σ[P (jω)]
Similarly, the controller gain, σ(K), should also be kept not too large in the frequency
range where the loop gain is small in order not to saturate the actuators. This is because
for small loop gain σ(Lo (jω))  1 or σ(Li (jω))  1
u = KSo (r − n − d) − Ti di ≈ K(r − n − d)
Therefore, it is desirable to keep σ(K) not too large when the loop gain is small.
To summarize the above discussion, we note that good performance requires in some
frequency range, typically some low-frequency range (0, ωl ),
σ(P K)  1, σ(KP )  1, σ(K)  1
6.2. Weighted H2 and H∞ Performance 85

and good robustness and good sensor noise rejection require in some frequency range,
typically some high-frequency range (ωh , ∞),

σ(P K)  1, σ(KP )  1, σ(K) ≤ M

where M is not too large. These design requirements are shown graphically in Figure 6.2.
The specific frequencies ωl and ωh depend on the specific applications and the knowledge
one has of the disturbance characteristics, the modeling uncertainties, and the sensor
noise levels.

HH
@ HH6
@ HH
@ @
XX @ σ(L)
XX @ @
Z @ @
Z @
Z @ @
Z S @
Z @
Z S
S Z
H Z ωh
HH Z -
ωl HH ZZ log ω

S @HH
σ(L) S @ HH
S @ H
HH
S @ HH
S @
S @

Figure 6.2: Desired loop gain

6.2 Weighted H2 and H∞ Performance

In this section, we consider how to formulate some performance objectives into math-
ematically tractable problems. As shown in Section 6.1, the performance objectives of
a feedback system can usually be specified in terms of requirements on the sensitivity
functions and/or complementary sensitivity functions or in terms of some other closed-
loop transfer functions. For instance, the performance criteria for a scalar system may
be specified as requiring 
|S(jω)| ≤ ε, ∀ω ≤ ω0 ,
|S(jω)| ≤ M, ∀ω > ω0
where S(jω) = 1/(1 + P (jω)K(jω)). However, it is much more convenient to reflect
the system performance objectives by choosing appropriate weighting functions. For
86 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

example, the preceding performance objective can be written as

|We (jω)S(jω)| ≤ 1, ∀ω
with 
1/ε, ∀ω ≤ ω0
|We (jω)| =
1/M, ∀ω > ω0
To use We in control design, a rational transfer function We (s) is usually used to ap-
proximate the foregoing frequency response.
The advantage of using weighted performance specifications is obvious in multi-
variable system design. First, some components of a vector signal are usually more
important than others. Second, each component of the signal may not be measured in
the same units; for example, some components of the output error signal may be mea-
sured in terms of length, and others may be measured in terms of voltage. Therefore,
weighting functions are essential to make these components comparable. Also, we might
be primarily interested in rejecting errors in a certain frequency range (for example, low
frequencies); hence some frequency-dependent weights must be chosen.

6
ũ d˜i d˜
? ?
Wu Wi Wd
6 di d
- Wr r-e - K u -?
e - P -?
e y- We e-

− 6

?n
e Wn 
ñ

Figure 6.3: Standard feedback configuration with weights

In general, we shall modify the standard feedback diagram in Figure 6.1 into Fig-
ure 6.3. The weighting functions in Figure 6.3 are chosen to reflect the design objectives
and knowledge of the disturbances and sensor noise. For example, Wd and Wi may be
chosen to reflect the frequency contents of the disturbances d and di or they may be used
to model the disturbance power spectrum depending on the nature of signals involved
in the practical systems. The weighting matrix Wn is used to model the frequency
contents of the sensor noise while We may be used to reflect the requirements on the
shape of certain closed-loop transfer functions (for example, the shape of the output
sensitivity function). Similarly, Wu may be used to reflect some restrictions on the con-
trol or actuator signals, and the dashed precompensator Wr is an optional element used
to achieve deliberate command shaping or to represent a nonunity feedback system in
equivalent unity feedback form.
6.2. Weighted H2 and H∞ Performance 87

It is, in fact, essential that some appropriate weighting matrices be used in order
to utilize the optimal control theory discussed in this book (i.e., H2 and H∞ theory).
So a very important step in the controller design process is to choose the appropriate
weights, We , Wd , Wu , and possibly Wn , Wi , Wr . The appropriate choice of weights for a
particular practical problem is not trivial. In many occasions, as in the scalar case, the
weights are chosen purely as a design parameter without any physical bases, so these
weights may be treated as tuning parameters that are chosen by the designer to achieve
the best compromise between the conflicting objectives. The selection of the weighting
matrices should be guided by the expected system inputs and the relative importance
of the outputs.
Hence, control design may be regarded as a process of choosing a controller K such
that certain weighted signals are made small in some sense. There are many different
ways to define the smallness of a signal or transfer matrix, as we have discussed in
the last chapter. Different definitions lead to different control synthesis methods, and
some are much harder than others. A control engineer should make a judgment of the
mathematical complexity versus engineering requirements.
Next, we introduce two classes of performance formulations: H2 and H∞ criteria.
For the simplicity of presentation, we shall assume that di = 0 and n = 0.

H2 Performance
Assume, for example, that the disturbance d˜ can be approximately modeled as an
impulse with random input direction; that is,
˜ = ηδ(t)
d(t)

and
E(ηη ∗ ) = I
where E denotes the expectation. We may choose to minimize the expected energy of
˜
the error e due to the disturbance d:
n o Z ∞ 
2 2 2
E kek2 = E kek dt = kWe So Wd k2
0

In general, a controller minimizing only the above criterion can lead to a very large
control signal u that could cause saturation of the actuators as well as many other
undesirable problems. Hence, for a realistic controller design, it is necessary to include
the control signal u in the cost function. Thus, our design criterion would usually be
something like this:
n o

We So Wd
 2

E kek2 + ρ2 kũk2 =
2 2
ρWu KSo Wd
2

with some appropriate choice of weighting matrix Wu and scalar ρ. The parameter ρ
clearly defines the tradeoff we discussed earlier between good disturbance rejection at
88 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

the output and control effort (or disturbance and sensor noise rejection at the actuators).
Note that ρ can be set to ρ = 1 by an appropriate choice of Wu . This problem can
be viewed as minimizing the energy consumed by the system in order to reject the
disturbance d.
This type of problem was the dominant paradigm in the 1960s and 1970s and is
usually referred to as linear quadratic Gaussian control, or simply as LQG. (Such prob-
lems will also be referred to as H2 mixed-sensitivity problems for consistency with the
H∞ problems discussed next.) The development of this paradigm stimulated extensive
research efforts and is responsible for important technological innovation, particularly
in the area of estimation. The theoretical contributions include a deeper understanding
of linear systems and improved computational methods for complex systems through
state-space techniques. The major limitation of this theory is the lack of formal treat-
ment of uncertainty in the plant itself. By allowing only additive noise for uncertainty,
the stochastic theory ignored this important practical issue. Plant uncertainty is par-
ticularly critical in feedback systems. (See Paganini [1995,1996] for some recent results
on robust H2 control theory.)

H∞ Performance
Although the H2 norm (or L2 norm) may be a meaningful performance measure and
although LQG theory can give efficient design compromises under certain disturbance
and plant assumptions, the H2 norm suffers a major deficiency. This deficiency is due
to the fact that the tradeoff between disturbance error reduction and sensor noise error
reduction is not the only constraint on feedback design. The problem is that these
performance tradeoffs are often overshadowed by a second limitation on high loop gains
— namely, the requirement for tolerance to uncertainties. Though a controller may
be designed using FDLTI models, the design must be implemented and operated with
a real physical plant. The properties of physical systems (in particular, the ways in
which they deviate from finite-dimensional linear models) put strict limitations on the
frequency range over which the loop gains may be large.
A solution to this problem would be to put explicit constraints on the loop gain in
the cost function. For instance, one may chose to minimize

sup kek2 = kWe So Wd k∞

kd˜k2 ≤1

subject to some restrictions on the control energy or control bandwidth:

sup kũk2 = kWu KSo Wd k∞

kd̃k2 ≤1

Or, more frequently, one may introduce a parameter ρ and a mixed criterion
n o
  2

We So Wd
sup kek22 + ρ2 kũk22 = ρWu KSo Wd

kd̃k2 ≤1 ∞
6.3. Selection of Weighting Functions 89

Alternatively, if the system robust stability margin is the major concern, the weighted
complementary sensitivity has to be limited. Thus the whole cost function may be
 
We So Wd

ρW1 To W2
∞

where W1 and W2 are the frequency-dependent uncertainty scaling matrices. These

design problems are usually called H∞ mixed-sensitivity problems. For a scalar system,
an H∞ norm minimization problem can also be viewed as minimizing the maximum
magnitude of the system’s steady-state response with respect to the worst-case sinusoidal
inputs.

6.3 Selection of Weighting Functions

The selection of weighting functions for a specific design problem often involves ad hoc
fixing, many iterations, and fine tuning. It is very hard to give a general formula for
the weighting functions that will work in every case. Nevertheless, we shall try to give
some guidelines in this section by looking at a typical SISO problem.
Consider an SISO feedback system shown in Figure 6.1. Then the tracking error is
e = r − y = S(r − d) + T n − SP di . So, as we have discussed earlier, we must keep |S|
small over a range of frequencies, typically low frequencies where r and d are significant.
To motivate the choice of our performance weighting function We , let L = P K be a
standard second-order system
ωn2
L=
s(s + 2ξωn )
It is well-known from the classical control theory that the quality of the (step) time
response can be quantified by rise time tr , settling time ts , and percent overshoot
100Mp%. Furthermore, these performance indices can be approximately calculated as

0.6 + 2.16ξ 4 − √ πξ
tr ≈ , 0.3 ≤ ξ ≤ 0.8; ts ≈ ; Mp = e 1−ξ2 , 0 < ξ < 1
ωn ξωn

The key points to note are that (1) the speed of the system response is proportional to
ωn and (2) the overshoot of the system response is determined only by the damping ratio
ξ. It is well known that the frequency ωn and the damping ratio ξ can be essentially
captured in the frequency domain by the open-loop crossover frequency and the phase
margin or the bandwidth and the resonant peak of the closed-loop complementary
sensitivity function T .
Since our performance objectives are closely related to the sensitivity function, we
shall consider in some detail how these time domain indices or, equivalently, ωn and ξ
are related to the frequency response of the sensitivity function

1 s(s + 2ξωn )
S= = 2
1+L s + 2ξωn s + ωn2
90 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

1
10

0
10
sensitivity function

−1
10

−2
10 −1 0 1
10 10 10
normalized frequency

Figure 6.4: Sensitivity function S for ξ = 0.05, 0.1, 0.2, 0.5, 0.8, and 1 with normalized
frequency (ω/ωn )

The frequency
√ response of the sensitivity function S is shown in Figure √ 6.4. Note that
|S(jωn / 2)| = 1. We can regard the closed-loop bandwidth ωb ≈ ωn / 2, since beyond
this frequency the closed-loop system will not be able to track the reference and the
disturbance will actually be amplified.
Next, note that
p
α α2 + 4ξ 2
Ms := kSk∞ = |S(jωmax )| = p
(1 − α2 )2 + 4ξ 2 α2
q p
where α = 0.5 + 0.5 1 + 8ξ 2 and ωmax = αωn . For example, Ms = 5.123 when
ξ = 0.1. The relationship between ξ and Ms is shown in Figure 6.5. It is clear that the
overshoot can be excessive if Ms is large. Hence a good control design should not have
a very large Ms .
Now suppose we are given the time domain performance specifications then we can
determine the corresponding requirements in frequency domain in terms of the band-
width ωb and the peak sensitivity Ms . Hence a good control design should result in
a sensitivity function S satisfying both the bandwidth ωb and the peak sensitivity Ms
requirements, as shown in Figure 6.6. These requirements can be approximately repre-
sented as  
 s 
|S(s)| ≤   , s = jω, ∀ ω
s/Ms + ωb 
6.3. Selection of Weighting Functions 91

4.5

3.5
peak sensitivity

2.5

1.5

1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
damping ratio

Figure 6.5: Peak sensitivity Ms versus damping ratio ξ

1/|We|

Ms

ωb |S(jω)|

Figure 6.6: Performance weight We and desired S

92 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

Or, equivalently, |We S| ≤ 1 with

s/Ms + ωb
We = (6.5)
s
The preceding discussion applies in principle to most control design and hence the
preceding weighting function can, in principle, be used as a candidate weighting function
in an initial design. Since the steady-state error with respect to a step input is given by
|S(0)|, it is clear that |S(0)| = 0 if the closed-loop system is stable and kWe Sk∞ < ∞.
Unfortunately, the optimal control techniques described in this book cannot be used
directly for problems with such weighting functions since these techniques assume that all
unstable poles of the system (including plant and all performance and control weighting
functions) are stabilizable by the control and detectable from the measurement outputs,
which is clearly not satisfied if We has an imaginary axis pole since We is not detectable
from the measurement. We shall discuss in Chapter 14 how such problems can be
reformulated so that the techniques described in this book can be applied. A theory
dealing directly with such problems is available but is much more complicated both
theoretically and computationally and does not seem to offer much advantage.

1/|We|

Ms

ωb |S(j ω)|

Figure 6.7: Practical performance weight We and desired S

Now instead of perfect tracking for step input, suppose we only need the steady-
state error with respect to a step input to be no greater than  (i.e., |S(0)| ≤ );
then it is sufficient to choose a weighting function We satisfying |We (0)| ≥ 1/ε so that
kWe Sk∞ ≤ 1 can be achieved. A possible choice of We can be obtained by modifying
the weighting function in equation (6.5):

s/Ms + ωb
We = (6.6)
s + ωb ε
6.3. Selection of Weighting Functions 93

Hence, for practical purpose, one can usually choose a suitable ε, as shown in Figure 6.7,
to satisfy the performance specifications. If a steeper transition between low-frequency
and high-frequency is desired, the weight We can be modified as follows:
 √ k
s/ k Ms + ωb
We = √ (6.7)
s + ωb k ε

for some integer k ≥ 1.

The selection of control weighting function Wu follows similarly from the preceding
discussion by considering the control signal equation

u = KS(r − n − d) − T di

The magnitude of |KS| in the low-frequency range is essentially limited by the allowable
cost of control effort and saturation limit of the actuators; hence, in general, the max-
imum gain Mu of KS can be fairly large, while the high-frequency gain is essentially
limited by the controller bandwidth (ωbc ) and the (sensor) noise frequencies. Ideally,
one would like to roll off as fast as possible beyond the desired control bandwidth so
that the high-frequency noises are attenuated as much as possible. Hence a candidate
weight Wu would be
s + ωbc /Mu
Wu = (6.8)
ωbc

1/|Wu |
Mu

|KS(jω)|

ω bc

ε1

Figure 6.8: Control weight Wu and desired KS

However, again the optimal control design techniques developed in this book cannot
be applied directly to a problem with an improper control weighting function. Hence
94 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

we shall introduce a far away pole to make Wu proper:

s + ωbc /Mu
Wu = (6.9)
ε1 s + ωbc

for a small ε1 > 0, as shown in Figure 6.8. Similarly, if a faster rolloff is desired, one
may choose
 √ k
s + ωbc / k Mu
Wu = √
k ε s + ω
(6.10)
1 bc

for some integer k ≥ 1.

The weights for MIMO problems can be initially chosen as diagonal matrices with
each diagonal term chosen in the foregoing form.

6.4 Bode’s Gain and Phase Relation

One important problem that arises frequently is concerned with the level of perfor-
mance that can be achieved in feedback design. It has been shown in Section 6.1 that
the feedback design goals are inherently conflicting, and a tradeoff must be performed
among different design objectives. It is also known that the fundamental requirements,
such as stability and robustness, impose inherent limitations on the feedback properties
irrespective of design methods, and the design limitations become more severe in the
presence of right-half plane zeros and poles in the open-loop transfer function.
In the classical feedback theory, Bode’s gain-phase integral relation (see Bode [1945])
has been used as an important tool to express design constraints in scalar systems. This
integral relation says that the phase of a stable and minimum phase transfer function
is determined uniquely by the magnitude of the transfer function. More precisely, let
L(s) be a stable and minimum phase transfer function: then
Z ∞
1 d ln |L| |ν|
∠L(jω0 ) = ln coth dν (6.11)
π −∞ dν 2

|ν| e|ν|/2 + e−|ν|/2

where ν := ln(ω/ω0 ). The function ln coth = ln |ν|/2 is plotted in Fig-
2 e − e−|ν|/2
ure 6.9.
|ν|
Note that ln coth decreases rapidly as ω deviates from ω0 and hence the integral
2
d ln |L(jω)|
depends mostly on the behavior of near the frequency ω0 . This is clear
dν
from the following integration:
 
Z  1.1406 (rad), α = ln 3  65.3o , α = ln 3
1 α |ν|
ln coth dν = 1.3146 (rad), α = ln 5 = 75.3o , α = ln 5
π −α 2  
1.443 (rad), α = ln 10 82.7o , α = ln 10.
6.4. Bode’s Gain and Phase Relation 95

4.5

3.5

3
ln coth |ν |/ 2

2.5

1.5

0.5

0
−3 −2 −1 0 1 2 3
ν

|ν|
Figure 6.9: The function ln coth vs ν
2

d ln |L(jω)|
Note that is the slope of the Bode plot, which is generally negative for
dν
almost all frequencies. It follows that ∠L(jω0 ) will be large if the gain L attenuates
slowly near ω0 and small if it attenuates rapidly near ω0 . For example, suppose the
d ln |L(jω)|
slope = −`; that is, (−20` dB per decade), in the neighborhood of ω0 ; then
dν
it is reasonable to expect

 −` × 65.3 , if the slope of L = −` for 3 ≤ ω0 ≤ 3
o 1 ω

∠L(jω0 ) < −` × 75.3 , if the slope of L = −` for 5 ≤ ωω0 ≤ 5

o 1

−` × 82.7o , if the slope of L = −` for 101
≤ ωω0 ≤ 10.

The behavior of ∠L(jω) is particularly important near the crossover frequency ωc , where
|L(jωc )| = 1 since π + ∠L(jωc ) is the phase margin of the feedback system. Further,
the return difference is given by
 
 π + ∠L(jωc ) 
|1 + L(jωc )| = |1 + L−1 (jωc )| = 2 sin ,

2

which must not be too small for good stability robustness. If π + ∠L(jωc ) is forced to
be very small by rapid gain attenuation, the feedback system will amplify disturbances
and exhibit little uncertainty tolerance at and near ωc . Since it is generally required
that the loop transfer function L roll off as fast as possible in the high-frequency range,
it is reasonable to expect that ∠L(jωc ) is at most −` × 90o if the slope of L(jω) is −`
near ωc . Thus it is important to keep the slope of L near ωc not much smaller than
−1 for a reasonably wide range of frequencies in order to guarantee some reasonable
96 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

performance. The conflict between attenuation rate and loop quality near crossover is
thus clearly evident.
Bode’s gain and phase relation can be extended to stable and nonminimum phase
transfer functions easily. Let z1 , z2 , . . . , zk be the right-half plane zeros of L(s), then L
can be factorized as
−s + z1 −s + z2 −s + zk
L(s) = ··· Lmp (s)
s + z1 s + z2 s + zk
where Lmp is stable and minimum phase and |L(jω)| = |Lmp (jω)|. Hence

Y
k
−jω0 + zi
∠L(jω0 ) = ∠Lmp (jω0 ) + ∠
i=1
jω0 + zi
Z ∞ Xk
1 d ln |Lmp | |ν| −jω0 + zi
= ln coth dν + ∠ ,
π −∞ dν 2 i=1
jω0 + zi
which gives
Z ∞ Xk
1 d ln |L| |ν| −jω0 + zi
∠L(jω0 ) = ln coth dν + ∠ . (6.12)
π −∞ dν 2 i=1
jω0 + zi

−jω0 + zi
Since ∠ ≤ 0 for each i, a nonminimum phase zero contributes an additional
jω0 + zi
phase lag and imposes limitations on the rolloff rate of the open-loop gain. For example,
suppose L has a zero at z > 0; then

−jω0 + z 
φ1 (ω0 /z) := ∠ = −90o , −53.13o, −28o,
jω0 + z ω0 =z,z/2,z/4

as shown in Figure 6.10. Since the slope of |L| near the crossover frequency is, in
general, no greater than −1, which means that the phase due to the minimum phase
part, Lmp , of L will, in general, be no greater than −90o , the crossover frequency (or
the closed-loop bandwidth) must satisfy
ωc < z/2 (6.13)
in order to guarantee the closed-loop stability and some reasonable closed-loop perfor-
mance.
Next suppose L has a pair of complex right-half zeros at z = x ± jy with x > 0; then

−jω0 + z −jω0 + z̄ 
φ2 (ω0 /|z|) := ∠
jω0 + z jω0 + z̄ ω0 =|z|,|z|/2,|z|/3,|z|/4

 −180o , −106.26o, −73.7o , −56o , Re(z)  =(z)
≈ −180 ,o
−86.7 , −55.9 , −41.3o, Re(z) ≈ =(z)
o o

−360 ,o
0o , 0o , 0o , Re(z)  =(z)
6.4. Bode’s Gain and Phase Relation 97

−10

−20

phase φ1(ω0 /z) (in degree) −30

−40

−50

−60

−70

−80

−90
ω0 / |z|
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 6.10: Phase φ1 (ω0 /z) due to a real zero z > 0

−20
y/x=10 y/x=100
−40
phase φ2 ( ω0 / |z| ) (in degree)

−60

−80 y/x=1 y/x=3

−100

−120

−140 y/x=0.01

−160

−180

−200
ω0 / |z|
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 6.11: Phase φ2 (ω0 /|z|) due to a pair of complex zeros: z = x ± jy and x > 0
98 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

as shown in Figure 6.11. In this case we conclude that the crossover frequency must
satisfy 
 |z|/4, Re(z)  =(z)
ωc < |z|/3, Re(z) ≈ =(z) (6.14)

|z|, Re(z)  =(z)
in order to guarantee the closed-loop stability and some reasonable closed-loop perfor-
mance.

6.5 Bode’s Sensitivity Integral

In this section, we consider the design limitations imposed by the bandwidth constraints
and the right-half plane poles and zeros using Bode’s sensitivity integral and Poisson
integral. Let L be the open-loop transfer function with at least two more poles than
zeros and let p1 , p2 , . . . , pm be the open right-half plane poles of L. Then the following
Bode’s sensitivity integral holds:
Z ∞ X
m
ln |S(jω)|dω = π Re(pi ) (6.15)
0 i=1

In the case where L is stable, the integral simplifies to

Z ∞
ln |S(jω)|dω = 0 (6.16)
0

These integrals show that there will exist a frequency range over which the magnitude
of the sensitivity function exceeds one if it is to be kept below one at other frequencies,
as illustrated in Figure 6.12. This is the so-called water bed effect.

|S(j ω)|

+
1
ω
−

Figure 6.12: Water bed effect of sensitivity function

6.5. Bode’s Sensitivity Integral 99

Suppose that the feedback system is designed such that the level of sensitivity re-
duction is given by
|S(jω)| ≤  < 1, ∀ω ∈ [0, ωl ]
where  > 0 is a given constant.
Bandwidth constraints in feedback design typically require that the open-loop trans-
fer function be small above a specified frequency, and that it roll off at a rate of more
than one pole-zero excess above that frequency. These constraints are commonly needed
to ensure stability robustness despite the presence of modeling uncertainty in the plant
model, particularly at high frequencies. One way of quantifying such bandwidth con-
straints is by requiring the open-loop transfer function to satisfy

Mh
|L(jω)| ≤ ≤ ˜ < 1, ∀ω ∈ [ωh , ∞)
ω 1+β
where ωh > ωl , and Mh > 0, β > 0 are some given constants.
Note that for ω ≥ ωh ,
1 1
|S(jω)| ≤ ≤
1 − |L(jω)| 1 − ωMh
1+β

and
Z ∞   ∞ Z
X ∞  i
Mh 1 Mh
− ln 1 − 1+β dω = dω
ωh ω i=1 ωh i ω 1+β
∞
!i
X 1 ωh Mh
=
i=1
i i(1 + β) − 1 ωh1+β
∞
!i !
ωh X 1 Mh ωh Mh
≤ =− ln 1 − 1+β
β i=1 i ωh1+β β ωh
ωh
≤ − ln(1 − ˜).
β

Then Z
X
m ∞
π Re(pi ) = ln |S(jω)|dω
i=1 0

Z ωl Z ωh Z ∞
= ln |S(jω)|dω + ln |S(jω)|dω + ln |S(jω)|dω
0 ωl ωh
Z ∞  
Mh
≤ ωl ln  + (ωh − ωl ) max ln |S(jω)| − ln 1 − 1+β dω
ω∈[ωl ,ωh ] ωh ω
ωh
≤ ωl ln  + (ωh − ωl ) max ln |S(jω)| − ln(1 − ˜),
ω∈[ωl ,ωh ] β
100 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

which gives
  ω ω−ω
l
1 h l ωh
max |S(jω)| ≥ e α
(1 − ˜) β(ωh −ωl )
ω∈[ωl ,ωh ] 
where Pm
π i=1 Re(pi )
α= .
ωh − ωl
The above lower bound shows that the sensitivity can be very significant in the transition
band.
Next, using the Poisson integral relation, we investigate the design constraints on
sensitivity properties imposed by open-loop nonminimum phase zeros. Suppose L has
at least one more poles than zeros and suppose z = x0 + jy0 with x0 > 0 is a right-half
plane zero of L. Then
Z Ym  
∞
x0  z + pi 
ln |S(jω)| 2 dω = π ln   (6.17)
x0 + (ω − y0 )2  z − pi 
−∞ i=1

This integral implies that the sensitivity reduction ability of the system may be severely
limited by the open-loop unstable poles and nonminimum phase zeros, especially when
these poles and zeros are close to each other.
Define Z ωl
x0
θ(z) := 2 + (ω − y )2 dω
−ωl 0x 0

Then
m 
Y
 Z
 z + pi  ∞
x0
π ln   ln |S(jω)|
 z − pi  = x20 + (ω − y0 )2
dω
i=1 −∞

≤ (π − θ(z)) ln kS(jω)k∞ + θ(z) ln(),

which gives
  π−θ(z)
θ(z)
m  ! π
1 Y  z + pi  π−θ(z)
kS(s)k∞ ≥  
  z − pi 
i=1

This lower bound on the maximum sensitivity shows that for a nonminimum phase
system, its sensitivity must increase significantly beyond one at certain frequencies if
the sensitivity reduction is to be achieved at other frequencies.

6.6 Analyticity Constraints

Let p1 , p2 , . . . , pm and z1 , z2 , . . . , zk be the open right-half plane poles and zeros of L,
respectively. Suppose that the closed-loop system is stable. Then

S(pi ) = 0, T (pi ) = 1, i = 1, 2, . . . , m
6.6. Analyticity Constraints 101

and
S(zj ) = 1, T (zj ) = 0, j = 1, 2, . . . , k
The internal stability of the feedback system is guaranteed by satisfying these analyticity
(or interpolation) conditions. On the other hand, these conditions also impose severe
limitations on the achievable performance of the feedback system.
Suppose S = (I + L)−1 and T = L(I + L)−1 are stable. Then p1 , p2 , . . . , pm are the
right-half plane zeros of S and z1 , z2 , . . . , zk are the right-half plane zeros of T . Let
Y
m
s − pi Yk
s − zj
Bp (s) = , Bz (s) =
i=1
s + pi j=1
s + zj

Then |Bp (jω)| = 1 and |Bz (jω)| = 1 for all frequencies and, moreover,
Bp−1 (s)S(s) ∈ H∞ , Bz−1 (s)T (s) ∈ H∞ .
Hence, by the maximum modulus theorem, we have

kS(s)k = Bp−1 (s)S(s) ≥ |Bp−1 (z)S(z)|
∞ ∞
for any z with Re(z) > 0. Let z be a right-half plane zero of L; then
Ym  
 z + pi 
kS(s)k∞ ≥ |Bp−1 (z)| =  
 z − pi 
i=1

Similarly, one can obtain

k 
Y

 p + zj 
kT (s)k∞ ≥ |Bz−1 (p)| =  
 p − zj 
j=1

where p is a right-half plane pole of L.

The weighted problem can be considered in the same fashion. Let We be a weight
such that We S is stable. Then
Ym  
 z + pi 
kWe (s)S(s)k∞ ≥ |We (z)|  
 z − pi 
i=1

s/Ms + ωb
Now suppose We (s) = , kWe Sk∞ ≤ 1, and z is a real right-half plane zero.
s + ωb 
Then m  
z/Ms + ωb Y  z − pi 
≤  
z + ωb   z + pi  =: α,
i=1
which gives
z 1 1
ωb ≤ (α − ) ≈ z(α − )
1 − α Ms Ms
where α = 1 if L has no right-half plane poles. This shows that the bandwidth of the
closed-loop must be much smaller than the right-half plane zero. Similar conclusions
can be arrived at for complex right-half plane zeros.
102 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

6.7 Notes and References

The loop-shaping design is well-known for SISO systems in the classical control theory.
The idea was extended to MIMO systems by Doyle and Stein [1981] using the LQG
design technique. The limitations of the loop-shaping design are discussed in detail in
Stein and Doyle [1991]. Chapter 16 presents another loop-shaping method using H∞
control theory, which has the potential to overcome the limitations of the LQG/LTR
method. Some additional discussions on the choice of weighting functions can be found
in Skogestad and Postlethwaite [1996]. The design tradeoffs and limitations for SISO
systems are discussed in detail in Bode [1945], Horowitz [1963], and Doyle, Francis,
and Tannenbaum [1992]. The monograph by Freudenberg and Looze [1988] contains
many multivariable generalizations. The multivariable generalization of Bode’s integral
relation can be found in Chen [1995], on which Section 6.5 is based. Some related results
can be found in Boyd and Desoer [1985]. Additional related results can be found in a
recent book by Seron, Braslavsky, and Goodwin [1997].

6.8 Problems
Problem 6.1 Let P be an open-loop plant. It is desired to design a controller so that
the overshoot ≤ 10% and settling time ≤ 10 sec. Estimate the allowable peak sensitivity
Ms and the closed-loop bandwidth.
1
Problem 6.2 Let L1 = be an open-loop transfer function of a unity feedback
s(s + 1)2
system. Find the phase margin, overshoot, settling time, and the corresponding Ms .

Problem 6.3 Repeat Problem 6.2 with

100(s + 10)
L2 = .
(s + 1)(s + 2)(s + 20)
10(1 − s)
Problem 6.4 Let P = . Use classical loop-shaping method to design a first-
s(s + 10)
order lead or lag controller so that the system has at least 30o phase margin and as
large a crossover frequency as possible.

Problem 6.5 Use the root locus method to show that a nonminimum phase system
cannot be stabilized by a very high-gain controller.
5
Problem 6.6 Let P = . Design a lead or lag controller so that the system
(1 − s)(s + 2)
has at least 30 phase margin with loop gain ≥ 2 for any frequency ω ≤ 0.1 and the
o

smallest possible bandwidth (or crossover frequency).

Problem 6.7 Use the root locus method to show that an unstable system cannot be
stabilized by a very low gain controller.
6.8. Problems 103

Problem 6.8 Consider the unity-feedback loop with proper controller K(s) and strictly
proper plant P (s), both assumed square. Assume internal stability.
1. Let w(s) be a scalar weighting function, assumed in RH∞ . Define

 = kw(I + P K)−1 k∞ , δ = kK(I + P K)−1 k∞

so  measures, say, disturbance attenuation and δ measures, say, control effort.

Derive the following inequality, which shows that  and δ cannot both be small
simultaneously in general. For every Re s0 ≥ 0

|w(s0 )| ≤  + |w(s0 )|σmin [P (s0 )]δ.

2. If we want very good disturbance attenuation at a particular frequency, you might

guess that we need high controller gain at that frequency. Fix ω with jω not a
pole of P (s), and suppose

 := σmax [(I + P K)−1 (jω)] < 1.

Derive a lower bound for σmin [K(jω)]. This lower bound should blow up as  → 0.

Problem 6.9 Suppose that P is proper and has one right half plane zero at s = z > 0.
w
Suppose that y = 1+P K , where w is a unit step at time t = 0, and that our performance
specification is 
α, if 0 ≤ t ≤ T ;
|y(t)| ≤
β, if T < t
for some α > 1 > β > 0. Show that for a proper, internally stabilizing, LTI controller
K to exist that meets the specification, we must have that
 
α−β
ln ≤ zT.
α−1

What tradeoffs does this imply?

Problem 6.10 Let K be a stabilizing controller for the plant

s−α
P =
(s − β)(s + γ)

α > 0, β > 0, γ ≥ 0. Suppose |S(jω)| ≤ δ < 1, ∀ω ∈ [−ω0 , ω0 ] where

1
S(s) = .
1 + PK
Find a lower bound for kSk∞ and calculate the lower bound for α = 1, β = 2, γ = 10,
δ = 0.2, and ω0 = 1.
104 PERFORMANCE SPECIFICATIONS AND LIMITATIONS
Chapter 7

Balanced Model Reduction

Simple linear models/controllers are normally preferred over complex ones in control
system design for some obvious reasons: They are much easier to do analysis and
synthesis with. Furthermore, simple controllers are easier to implement and are more
reliable because there are fewer things to go wrong in the hardware or bugs to fix in
the software. In the case when the system is infinite dimensional, the model/controller
approximation becomes essential. In this chapter we consider the problem of reducing
the order of a linear multivariable dynamical system. There are many ways to reduce the
order of a dynamical system. However, we shall study only one of them: the balanced
truncation method. The main advantage of this method is that it is simple and performs
fairly well.
A model order-reduction problem can, in general, be stated as follows: Given a full-
order model G(s), find a lower-order model (say, an rth order model Gr ), such that G
and Gr are close in some sense. Of course, there are many ways to define the closeness
of an approximation. For example, one may desire that the reduced model be such that

G = Gr + ∆a

and ∆a is small in some norm. This model reduction is usually called an additive model
reduction problem. We shall be only interested in L∞ norm approximation in this book.
Once the norm is chosen, the additive model reduction problem can be formulated as

inf kG − Gr k∞ .
deg(Gr )≤r

In general, a practical model reduction problem is inherently frequency-weighted (i.e.,

the requirement on the approximation accuracy at one frequency range can be drastically
different from the requirement at another frequency range). These problems can, in
general, be formulated as frequency-weighted model reduction problems:

inf kWo (G − Gr )Wi k∞

deg(Gr )≤r

105
106 BALANCED MODEL REDUCTION

with an appropriate choice of Wi and Wo . We shall see in this chapter how the bal-
anced realization can give an effective approach to the aforementioned model reduction
problems.

7.1 Lyapunov Equations

Testing stability, controllability, and observability of a system is very important in linear
system analysis and synthesis. However, these tests often have to be done indirectly. In
that respect, the Lyapunov theory is sometimes useful. Consider the following Lyapunov
equation:
A∗ Q + QA + H = 0 (7.1)
with given real matrices A and H. It is well known that this equation has a unique
solution iff λi (A) + λ̄j (A) 6= 0, ∀i, j. In this section, we shall study the relationships
between the solution Q and the stability of A. The following results are standard.

Lemma 7.1 Assume that A is stable, then the following statements hold:
R∞ ∗
(i) Q = 0 eA t HeAt dt.

(ii) Q > 0 if H > 0 and Q ≥ 0 if H ≥ 0.

(iii) If H ≥ 0, then (H, A) is observable iff Q > 0.

An immediate consequence of part (iii) is that, given a stable matrix A, a pair (C, A)
is observable if and only if the solution to the following Lyapunov equation

A∗ Q + QA + C ∗ C = 0

is positive definite, where Q is the observability Gramian. Similarly, a pair (A, B) is

controllable if and only if the solution to

AP + P A∗ + BB ∗ = 0

is positive definite, where P is the controllability Gramian.

In many applications, we are given the solution of the Lyapunov equation and need
to conclude the stability of the matrix A.

Lemma 7.2 Suppose Q is the solution of the Lyapunov equation (7.1), then

(i) Reλi (A) ≤ 0 if Q > 0 and H ≥ 0.

(ii) A is stable if Q > 0 and H > 0.

(iii) A is stable if Q ≥ 0, H ≥ 0, and (H, A) is detectable.

7.2. Balanced Realizations 107

Proof. Let λ be an eigenvalue of A and v 6= 0 be a corresponding eigenvector, then

Av = λv. Premultiply equation (7.1) by v ∗ and postmultiply equation (7.1) by v to get

2Re λ(v ∗ Qv) + v ∗ Hv = 0.

Now if Q > 0, then v ∗ Qv > 0, and it is clear that Reλ ≤ 0 if H ≥ 0 and Reλ < 0 if
H > 0. Hence (i) and (ii) hold. To see (iii), we assume Reλ ≥ 0. Then we must have
v ∗ Hv = 0 (i.e., Hv = 0). This implies that λ is an unstable and unobservable mode,
which contradicts the assumption that (H, A) is detectable. 2

7.2 Balanced Realizations

Although there are infinitely many different state-space realizations for a given transfer
matrix, some particular realizations have proven to be very useful in control engineering
and signal processing. Here we will only introduce one class of realizations for stable
transfer matrices that are most useful in control applications. To motivate the class of
realizations, we first consider some simple facts.
 
A B
Lemma 7.3 Let be a state-space realization of a (not necessarily stable)
C D
transfer matrix G(s). Suppose that there exists a symmetric matrix
 
P1 0
P = P∗ =
0 0

with P1 nonsingular such that

AP + P A∗ + BB ∗ = 0.

Now partition the realization (A, B, C, D) compatibly with P as

 
A11 A12 B1
 A21 A22 B2  .
C1 C2 D
 
A11 B1
Then is also a realization of G. Moreover, (A11 , B1 ) is controllable if A11
C1 D
is stable.

Proof. Use the partitioned P and (A, B, C) to get

 
∗ ∗ A11 P1 + P1 A∗11 + B1 B1∗ P1 A∗21 + B1 B2∗
0 = AP + P A + BB = ,
A21 P1 + B2 B1∗ B2 B2∗
108 BALANCED MODEL REDUCTION

which gives B2 = 0 and A21 = 0 since P1 is nonsingular. Hence, part of the realization
is not controllable:
   
A11 A12 B1 A11 A12 B1  
 A21 A22 B2  =  A11 B1
0 A22 0 = .
C1 D
C1 C2 D C1 C2 D

Finally, it follows from Lemma 7.1 that (A11 , B1 ) is controllable if A11 is stable. 2

We also have the following:

 
A B
Lemma 7.4 Let be a state-space realization of a (not necessarily stable)
C D
transfer matrix G(s). Suppose that there exists a symmetric matrix
 
∗ Q1 0
Q=Q =
0 0

with Q1 nonsingular such that

QA + A∗ Q + C ∗ C = 0.

Now partition the realization (A, B, C, D) compatibly with Q as

 
A11 A12 B1
 A21 A22 B2  .
C1 C2 D
 
A11 B1
Then is also a realization of G. Moreover, (C1 , A11 ) is observable if A11
C1 D
is stable.

The preceding two lemmas suggest that to obtain a minimal realization from a stable
nonminimal realization, one only needs to eliminate all states corresponding to the zero
block diagonal term of the controllability Gramian P and the observability Gramian
Q. In the case where P is not block diagonal, the following procedure can be used to
eliminate noncontrollable subsystems:
 
A B
1. Let G(s) = be a stable realization.
C D
2. Compute the controllability Gramian P ≥ 0 from

AP + P A∗ + BB ∗ = 0.
 
  Λ1 0  ∗
3. Diagonalize P to get P = U1 U2 U1 U2 with Λ1 > 0 and
  0 0
U1 U2 unitary.
7.2. Balanced Realizations 109

 
U1∗ AU1 U1∗ B
4. Then G(s) = is a controllable realization.
CU1 D

A dual procedure can also be applied to eliminate nonobservable subsystems.

Now assume that Λ1 > 0 is diagonal and is partitioned as Λ1 = diag(Λ11 , Λ12 ) such
that λmax (Λ12 )  λmin (Λ11 ); then it is tempting to conclude that one can also discard
those states corresponding to Λ12 without causing much error. However, this is not
necessarily true, as shown in the following example.

Example 7.1 Consider a stable transfer function

3s + 18
G(s) = .
s2 + 3s + 18
Then G(s) has a state-space realization given by
 
−1 −4/α 1
G(s) =  4α −2 2α 
−1 2/α 0

where α is any nonzero number. It is easy to check that the controllability Gramian of
the realization is given by  
0.5
P = .
α2
Since the last diagonal term of P can be made arbitrarily small by making α small,
the controllability of the corresponding state can be made arbitrarily weak. If the state
corresponding to the last diagonal term of P is removed, we get a transfer function
 
−1 1 −1
Ĝ = = ,
−1 0 s+1

which is not close to the original transfer function in any sense. The problem may be
easily detected if one checks the observability Gramian Q, which is
 
0.5
Q= .
1/α2

Since 1/α2 is very large if α is small, this shows that the state corresponding to the last
diagonal term is strongly observable.

This example shows that the controllability (or observability) Gramian alone cannot
give an accurate indication of the dominance of the system states in the input/output
110 BALANCED MODEL REDUCTION

behavior. This motivates the introduction of a balanced realization that gives balanced
Gramians for controllability
 and observability.
A B
Suppose G = is stable (i.e., A is stable). Let P and Q denote the
C D
controllability Gramian and observability Gramian, respectively. Then by Lemma 7.1,
P and Q satisfy the following Lyapunov equations:

AP + P A∗ + BB ∗ = 0 (7.2)

A∗ Q + QA + C ∗ C = 0, (7.3)

and P ≥ 0, Q ≥ 0. Furthermore, the pair (A, B) is controllable iff P > 0, and (C, A) is
observable iff Q > 0.
Suppose the state is transformed by a nonsingular T to x̂ = T x to yield the realiza-
tion " #  
 B̂ T AT −1 T B
G= = .
Ĉ D̂ CT −1 D

Then the Gramians are transformed to P̂ = T P T ∗ and Q̂ = (T −1 )∗ QT −1 . Note that

P̂ Q̂ = T P QT −1, and therefore the eigenvalues of the product of the Gramians are
invariant under state transformation.
Consider the similarity transformation T , which gives the eigenvector decomposition

P Q = T −1 ΛT, Λ = diag(λ1 Is1 , . . . , λN IsN ).

Then the columns of T −1 are eigenvectors of P Q corresponding to the eigenvalues {λi }.

Later, it will be shown that P Q has a real diagonal Jordan form and that Λ ≥ 0, which
are consequences of P ≥ 0 and Q ≥ 0.
Although the eigenvectors are not unique, in the case of a minimal realization they
can always be chosen such that

P̂ = T P T ∗ = Σ,

Q̂ = (T −1 )∗ QT −1 = Σ,

where Σ = diag(σ1 Is1 , σ2 Is2 , . . . , σN IsN ) and Σ2 = Λ. This new realization with con-
trollability and observability Gramians P̂ = Q̂ = Σ will be referred to as a balanced
realization (also called internally balanced realization). The decreasingly ordered num-
bers, σ1 > σ2 > . . . > σN ≥ 0, are called the Hankel singular values of the system.
More generally, if a realization of a stable system is not minimal, then there is a trans-
formation such that the controllability and observability Gramians for the transformed
realization are diagonal and the controllable and observable subsystem is balanced. This
is a consequence of the following matrix fact.
7.2. Balanced Realizations 111

Theorem 7.5 Let P and Q be two positive semidefinite matrices. Then there exists a
nonsingular matrix T such that
   
Σ1 Σ1
 Σ2   0 
TPT∗ =  
,
 (T −1 )∗ QT −1 = 



0 Σ3
0 0

respectively, with Σ1 , Σ2 , Σ3 diagonal and positive definite.

Proof. Since P is a positive semidefinite matrix, there exists a transformation T1 such

that  
I 0
T1 P T1∗ =
0 0
Now let  
Q11 Q12
(T1∗ )−1 QT1−1 =
Q∗12 Q22
and there exists a unitary matrix U1 such that
 2 
Σ1 0
U1 Q11 U1∗ = , Σ1 > 0
0 0

Let  
U1 0
(T2∗ )−1 =
0 I
and then  
Σ21 0 Q̂121
(T2∗ )−1 (T1∗ )−1 QT1−1 (T2 )−1 = 0 0 Q̂122 
Q̂∗121 Q̂∗122 Q22
But Q ≥ 0 implies Q̂122 = 0. So now let
 
I 0 0
(T3∗ )−1 =  0 I 0 
−Q̂∗121 Σ−2
1 0 I

giving
 
Σ21 0 0
(T3∗ )−1 (T2∗ )−1 (T1∗ )−1 QT1−1 (T2 )−1 (T3 )−1 =  0 0 0 
∗ −2
0 0 Q22 − Q̂121 Σ1 Q̂121

Next find a unitary matrix U2 such that

 
Σ3 0
U2 (Q22 − Q̂∗121 Σ−2 ∗
1 Q̂121 )U2 = , Σ3 > 0
0 0
112 BALANCED MODEL REDUCTION

Define  −1/2

Σ1 0 0
(T4∗ )−1 = 0 I 0 
0 0 U2
and let
T = T4 T3 T2 T1
Then
   
Σ1 Σ1
 Σ2   0 
TPT∗ = 

,
 (T ∗ )−1 QT −1 = 



0 Σ3
0 0

with Σ2 = I. 2

Corollary 7.6 The product of two positive semidefinite matrices is similar to a positive
semidefinite matrix.

Proof. Let P and Q be any positive semidefinite matrices. Then it is easy to see that
with the transformation given previously
 2 
−1 Σ1 0
T P QT =
0 0
2

 
A B
Corollary 7.7 For any stable system G = , there exists a nonsingular trans-
C D
" #
T AT −1 T B
formation T such that G = has controllability Gramian P and ob-
CT −1 D
servability Gramian Q given by
   
Σ1 Σ1
 Σ2   0 
P = 
,
 Q=



0 Σ3
0 0

respectively, with Σ1 , Σ2 , Σ3 diagonal and positive definite.

7.2. Balanced Realizations 113

 
A B
In the special case where is a minimal realization, a balanced realization
C D
can be obtained through the following simplified procedure:
1. Compute the controllability and observability Gramians P > 0, Q > 0.
2. Find a matrix R such that P = R∗ R.
3. Diagonalize RQR∗ to get RQR∗ = U Σ2 U ∗ .
" #
−1 ∗ −1/2 ∗ ∗ −1 −1
T AT −1 TB
4. Let T = R UΣ . Then T P T = (T ) QT = Σ and
CT −1 D
is balanced.
Assume that the Hankel singular values of the system are decreasingly ordered so
that Σ = diag(σ1 Is1 , σ2 Is2 , . . . , σN IsN ) with σ1 > σ2 > . . . > σN and suppose σr 
σr+1 for some r. Then the balanced realization implies that those states corresponding
to the singular values of σr+1 , . . . , σN are less controllable and less observable than those
states corresponding to σ1 , . . . , σr . Therefore, truncating those less controllable and less
observable states will not lose much information about the system.
Two other closely related realizations are called input normal realization with P = I
and Q = Σ2 , and output normal realization with P = Σ2 and Q = I. Both realizations
can be obtained easily from the balanced realization by a suitable scaling on the states.
Next we shall derive some simple and useful bounds for the H∞ norm and the L1
norm of a stable system.

Theorem 7.8 Suppose  

A B
G(s) = ∈ RH∞
C 0
is a balanced realization; that is, there exists

Σ = diag(σ1 Is1 , σ2 Is2 , . . . , σN IsN ) ≥ 0

with σ1 > σ2 > . . . > σN ≥ 0, such that

AΣ + ΣA∗ + BB ∗ = 0 A∗ Σ + ΣA + C ∗ C = 0

Then
Z ∞ X
N
σ1 ≤ kGk∞ ≤ kg(t)k dt ≤ 2 σi
0 i=1

where g(t) = CeAt B.

Remark 7.1 It should be clear that the inequalities stated in the theorem do not
depend on a particular state-space realization of G(s). However, use of the balanced
realization does make the proof simple. 3
114 BALANCED MODEL REDUCTION

Proof. Let G(s) have the following state-space realization:

ẋ = Ax + Bw
z = Cx. (7.4)

Assume without loss of generality that (A, B) is controllable and (C, A) is observable.
Then Σ is nonsingular. Next, differentiate x(t)∗ Σ−1 x(t) along the solution of equation
(7.4) for any given input w as follows:

d ∗ −1
(x Σ x) = ẋ∗ Σ−1 x + x∗ Σ−1 ẋ = x∗ (A∗ Σ−1 + Σ−1 A)x + 2hw, B ∗ Σ−1 xi
dt
Using the equation involving controllability Gramian to substitute for A∗ Σ−1 + Σ−1 A
and completion of the squares gives
d ∗ −1
(x Σ x) = kwk2 − kw − B ∗ Σ−1 xk2
dt
Integration from t = −∞ to t = 0 with x(−∞) = 0 and x(0) = x0 gives

x∗0 Σ−1 x0 = kwk22 − kw − B ∗ Σ−1 xk22 ≤ kwk22

Let w = B ∗ Σ−1 x; then ẋ = (A + BB ∗ Σ−1 )x = −ΣA∗ Σ−1 x =⇒ x ∈ L2 (−∞, 0]

=⇒ w ∈ L2 (−∞, 0] and
 
inf kwk22  x(0) = x0 = x∗0 Σ−1 x0 .
w∈L2 (−∞,0]

Given x(0) = x0 and w = 0 for t ≥ 0, the norm of z(t) = CeAt x0 can be found from
Z ∞ Z ∞
∗
x∗0 eA t C ∗ CeAt x0 dt = x∗0 Σx0
2
kz(t)k dt =
0 0

To show σ1 ≤ kGk∞ , note that

qR
∞
kg ∗ wk2 −∞
kz(t)k2 dt
kGk∞ = sup = sup qR∞
w∈L2 (−∞,∞) kwk2 w∈L2 (−∞,∞) kw(t)k2 dt −∞

qR s
∞ 2
0
kz(t)k dt x∗0 Σx0
≥ sup qR = sup = σ1
w∈L2 (−∞,0] 0 2
kw(t)k dt x0 6=0 x∗0 Σ−1 x0
−∞

We shall now show the other inequalities. Since

Z ∞
G(s) := g(t)e−st dt, Re(s) > 0,
0
7.2. Balanced Realizations 115

by the definition of H∞ norm, we have

Z ∞

kGk∞ = sup g(t)e −st
dt
Re(s)>0 0
Z ∞

≤ sup g(t)e−st dt
Re(s)>0 0
Z ∞
≤ kg(t)k dt.
0

To prove the last inequality, let ei be the ith unit vector and define
   
E1 = e1 · · · es1 , E2 = es1 +1 · · · es1 +s2 , ...,
 
EN = es1 +···+sN −1 +1 · · · es1 +···+sN .
X
N
Then Ei Ei∗ = I and
i=1

Z Z

At/2 X
∞ ∞ N
∗ At/2
kg(t)k dt = Ce Ei Ei e B dt
0 0
i=1
N Z ∞
X
At/2
≤ Ce Ei Ei∗ eAt/2 B dt
i=1 0
N Z
X ∞
At/2 ∗ At/2
≤ Ce Ei Ei e B dt
i=1 0
sZ sZ
X
N ∞ 2 ∞ X
N
≤ CeAt/2 Ei dt E ∗ eAt/2 B 2 dt ≤ 2 σi
i
i=1 0 0 i=1

where we have used Cauchy-Schwarz inequality and the following relations:

Z ∞ Z ∞  
At/2 2 ∗
Ce Ei dt = λmax Ei∗ eA t/2 C ∗ CeAt/2 Ei dt = 2λmax (Ei∗ ΣEi ) = 2σi
0 0
Z ∞ Z ∞  
∗ At/2 2 ∗
Ei e B dt = λmax Ei∗ eAt/2 BB ∗ eA t/2 Ei dt = 2λmax (Ei∗ ΣEi ) = 2σi
0 0
2

Example 7.2 Consider a system

 
−1 −2 1
G(s) =  1 0 0 
2 3 0
116 BALANCED MODEL REDUCTION

It is easy to show that the Hankel singular values of G are σ1 = 1.6061 and σ2 = 0.8561.
The H∞ norm of G is kGk∞ = 2.972 and the L1 norm of g(t) can be computed as
Z ∞
|g(t)|dt = h1 + h2 + h3 + h4 + . . .
0

where hi , i =R 1, 2, . . . are the variations of the step response of G shown in Figure 7.1,
∞
which gives 0 |g(t)|dt ≈ 3.5. (See Problem 7.2.)

2.5

2
h2

h4
1.5
step response

h3

h1
1

0.5

0
0 1 2 3 4 5 6 7 8 9 10
time

Figure 7.1: Estimating the L1 norm of g(t)

So we have
Z ∞
1.6061 = σ1 ≤ kGk∞ = 2.972 ≤ |g(t)|dt = 3.5 ≤ 2(σ1 + σ2 ) = 4.9244.
0

Illustrative MATLAB Commands:

 [Ab, Bb, Cb, sig, Tinv]=balreal(A, B, C); % sig is a vector of Hankel

singular values and Tinv = T −1 ;
 [Gb , sig] = sysbal(G);

Related MATLAB Commands: ssdelete, ssselect, modred, strunc

7.3. Model Reduction by Balanced Truncation 117

7.3 Model Reduction by Balanced Truncation

 
A B
Consider a stable system G ∈ RH∞ and suppose G = is a balanced realiza-
C D
tion (i.e., its controllability and observability Gramians are equal and diagonal). Denote
the balanced Gramians by Σ; then

AΣ + ΣA∗ + BB ∗ = 0 (7.5)

A∗ Σ + ΣA + C ∗ C = 0. (7.6)
 
Σ1 0
Now partition the balanced Gramian as Σ = and partition the system
0 Σ2
accordingly as  
A11 A12 B1
G =  A21 A22 B2  .
C1 C2 D
The following theorem characterizes the properties of these subsystems.

Theorem 7.9 Assume that Σ1 and Σ2 have no diagonal entries in common. Then both
subsystems (Aii , Bi , Ci ), i = 1, 2 are asymptotically stable.

Proof. It is clearly sufficient to show that A11 is asymptotically stable. The proof for
the stability of A22 is similar. Note that equations (7.5) and (7.6) can be written in
terms of their partitioned matrices as

A11 Σ1 + Σ1 A∗11 + B1 B1∗ = 0 (7.7)

Σ1 A11 + A∗11 Σ1 + C1∗ C1 = 0 (7.8)
A21 Σ1 + Σ2 A∗12 + B2 B1∗ = 0 (7.9)
Σ2 A21 + A∗12 Σ1 + C2∗ C1 = 0 (7.10)
A22 Σ2 + Σ2 A∗22 + B2 B2∗ = 0 (7.11)
Σ2 A22 + A∗22 Σ2 + C2∗ C2 = 0. (7.12)

By Lemma 7.3 or Lemma 7.4, Σ1 can be assumed to be positive definite without loss of
generality. Then it is obvious that Reλi (A11 ) ≤ 0 by Lemma 7.2. Assume that A11 is
not asymptotically stable; then there exists an eigenvalue at jω for some ω. Let V be
a basis matrix for Ker(A11 − jωI). Then we have

(A11 − jωI)V = 0, (7.13)

which gives
V ∗ (A∗11 + jωI) = 0.
118 BALANCED MODEL REDUCTION

Equations (7.7) and (7.8) can be rewritten as

(A11 − jωI)Σ1 + Σ1 (A∗11 + jωI) + B1 B1∗ = 0 (7.14)
Σ1 (A11 − jωI) + (A∗11 + jωI)Σ1 + C1∗ C1 = 0. (7.15)
Multiplication of equation (7.15) from the right by V and from the left by V ∗ gives
V ∗ C1∗ C1 V = 0, which is equivalent to
C1 V = 0.
Multiplication of equation (7.15) from the right by V now gives
(A∗11 + jωI)Σ1 V = 0.
Analogously, first multiply equation (7.14) from the right by Σ1 V and from the left by
V ∗ Σ1 to obtain
B1∗ Σ1 V = 0.
Then multiply equation (7.14) from the right by Σ1 V to get
(A11 − jωI)Σ21 V = 0.
It follows that the columns of Σ21 V are in Ker(A11 − jωI). Therefore, there exists a
matrix Σ̄1 such that
Σ21 V = V Σ̄21 .
Since Σ̄21 is the restriction of Σ21 to the space spanned by V , it follows that it is possible
to choose V such that Σ̄21 is diagonal. It is then also possible to choose Σ̄1 diagonal and
such that the diagonal entries of Σ̄1 are a subset of the diagonal entries of Σ1 .
Multiply equation (7.9) from the right by Σ1 V and equation (7.10) by V to get
A21 Σ21 V + Σ2 A∗12 Σ1 V = 0
Σ2 A21 V + A∗12 Σ1 V = 0,
which gives
(A21 V )Σ̄21 = Σ22 (A21 V ).
This is a Sylvester equation in (A21 V ). Because Σ̄21 and Σ22 have no diagonal entries in
common, it follows that
A21 V = 0 (7.16)
is the unique solution. Now equations (7.16) and (7.13) imply that
    
A11 A12 V V
= jω ,
A21 A22 0 0
which means that the A-matrix of the original system has an eigenvalue at jω. This
contradicts the fact that the original system is asymptotically stable. Therefore, A11
must be asymptotically stable. 2
7.3. Model Reduction by Balanced Truncation 119

Corollary 7.10 If Σ has distinct singular values, then every subsystem is asymptoti-
cally stable.

The stability condition in Theorem 7.9 is only sufficient as shown in the following
example.

Example 7.3 Note that

 
−2 −2.8284 −2
(s − 1)(s − 2) 
= 0 −1 −1.4142 
(s + 1)(s + 2)
2 1.4142 1

is a balanced realization with Σ = I, and every subsystem of the realization is stable.

On the other hand,
 
−1 1.4142 1.4142
s2 − s + 2  
= −1.4142 0 0
s2 + s + 2
−1.4142 0 1

is also a balanced realization with Σ = I, but one of the subsystems is not stable.

Theorem 7.11 Suppose G(s) ∈ RH∞ and

 
A11 A12 B1
G(s) =  A21 A22 B2 
C1 C2 D

is a balanced realization with Gramian Σ = diag(Σ1 , Σ2 )

Σ1 = diag(σ1 Is1 , σ2 Is2 , . . . , σr Isr )

Σ2 = diag(σr+1 Isr+1 , σr+2 Isr+2 , . . . , σN IsN )

and
σ1 > σ2 > · · · > σr > σr+1 > σr+2 > · · · > σN
where σi has multiplicity si , i = 1, 2, . . . , N and s1 + s2 + · · · + sN = n. Then the
truncated system  
A11 B1
Gr (s) =
C1 D
is balanced and asymptotically stable. Furthermore,

kG(s) − Gr (s)k∞ ≤ 2(σr+1 + σr+2 + · · · + σN )

120 BALANCED MODEL REDUCTION

Proof. The stability of Gr follows from Theorem 7.9. We shall first show the one step
model reduction. Hence we shall assume Σ2 = σN IsN . Define the approximation error
 
A11 A12 B1  
  A11 B1
E11 := A21 A22 B2 −
C1 D
C1 C2 D
 
A11 0 0 B1
 0 A A B1 
=   0
11 12 
A21 A22 B2 
−C1 C1 C2 0
Apply a similarity transformation T to the preceding state-space realization with
   
I/2 I/2 0 I I 0
T =  I/2 −I/2 0  , T −1 =  I −I 0 
0 0 I 0 0 I
to get
 
A11 0 A12 /2 B1
 0 A11 −A12 /2 0 
E11 = 
 A21

−A21 A22 B2 
0 −2C1 C2 0
Consider a dilation of E11 (s):
 
E11 (s) E12 (s)
E(s) =
E21 (s) E22 (s)
 
A11 0 A12 /2 B1 0
 0 A −A12 /2 0 σN Σ−1
1 C1
∗ 
 11 
=   A 21 −A 21 A22 B2 −C2 ∗ 

 0 −2C1 C2 0 2σN I 
−2σN B1∗ Σ−1
1 0 −B2∗ 2σN I 0
" #
à B̃
=:
C̃ D̃
Then it is easy to verify that
 
Σ1 0
P̃ =  0 2 −1
σN Σ1 0 
0 0 2σN IsN
satisfies
ÃP̃ + P̃ Ã∗ + B̃ B̃ ∗ = 0
P̃ C̃ ∗ + B̃ D̃∗ = 0
7.3. Model Reduction by Balanced Truncation 121

Using these two equations, we have

 
à −B̃ B̃ ∗ B̃ D̃∗
∼  0 −Ã∗ C̃ ∗ 
E(s)E (s) =  
C̃ −D̃B̃ ∗ D̃D̃∗
 
à −ÃP̃ − P̃ Ã∗ − B̃ B̃ ∗ P̃ C̃ ∗ + B̃ D̃∗
 −Ã∗ C̃ ∗ 
=  0 
C̃ −C̃ P̃ − D̃ B̃ ∗ D̃D̃ ∗
 
à 0 0
 0 −Ã∗ C̃ ∗ 
=  
C̃ 0 D̃D̃∗
= D̃D̃∗ = 4σN
2
I

where the second equality is obtained by applying a similarity transformation

 
I P̃
T =
0 I

Hence kE11 k∞ ≤ kEk∞ = 2σN , which is the desired result.

The remainder of the proof isachieved byusing the order reduction by one-step re-
A11 B1
sults and by noting that Gk (s) = obtained by the “kth” order partitioning
C1 D
is internally balanced with balanced Gramian given by

Σ1 = diag(σ1 Is1 , σ2 Is2 , . . . , σk Isk )

Let Ek (s) = Gk+1 (s) − Gk (s) for k = 1, 2, . . . , N − 1 and let GN (s) = G(s). Then

σ [Ek (jω)] ≤ 2σk+1

since Gk (s) is a reduced-order model obtained from the internally balanced realization
of Gk+1 (s) and the bound for one-step order reduction holds.
Noting that
X
N −1
G(s) − Gr (s) = Ek (s)
k=r

by the definition of Ek (s), we have

X
N −1 X
N −1
σ [G(jω) − Gr (jω)] ≤ σ [Ek (jω)] ≤ 2 σk+1
k=r k=r

This is the desired upper bound. 2

A useful consequence of the preceding theorem is the following corollary.

122 BALANCED MODEL REDUCTION

Corollary 7.12 Let σi , i = 1, . . . , N be the Hankel singular values of G(s) ∈ RH∞ .

Then
kG(s) − G(∞)k∞ ≤ 2(σ1 + . . . + σN )

The above bound can be tight for some systems.

Example 7.4 Consider an nth-order transfer function

X
n
bi
G(s) = ,
i=1
s + ai
P
with ai > 0 and bi > 0. Then kG(s)k∞ = G(0) = ni=1 bi /ai and G(s) has the following
state-space realization:
 √ 
−a1 √b1
 −a2 b2 
 
 .. .. 
G= . 
√. 
 −a 
 n bn 
√ √ √
b1 b2 · · · bn 0

and the controllability and observability Gramians of the realization are given by
"p #
bi bj
P =Q=
ai + aj

It is easy to see that σi = λi (P ) = λi (Q) and

X
n X
n Xn
bi 1 1
σi = λi (P ) = trace(P ) = = G(0) = kGk∞
i=1 i=1 i=1
2a i 2 2

In particular, let ai = bi = α2i ; then P = Q → 12 In (i.e., σj → 12 as α → ∞). This

example also shows that even when the Hankel singular values are extremely close, they
may not be regarded as repeated singular values.

The model reduction bound can also be loose for systems with Hankel singular values
close to each other.
7.3. Model Reduction by Balanced Truncation 123

Example 7.5 Consider the balanced realization of a fourth-order system:

(s − 0.99)(s − 2)(s − 3)(s − 4)

G(s) =
(s + 1)(s + 2)(s + 3)(s + 4)
 
−9.2e + 00 −5.7e + 00 −2.7e + 00 1.3e + 00 −4.3e + 00
 5.7e + 00 −8.1e − 07 −6.4e − 01 1.5e − 06 1.3e − 03 
 
=
 −2.7e + 00 6.4e − 01 −7.9e − 01 7.1e − 01 −1.3e + 00 
 −1.3e + 00 1.5e − 06 −7.1e − 01 −2.7e − 06 −2.3e − 03 
4.3e + 00 1.3e − 03 1.3e + 00 −2.3e − 03 1.0e + 00
with Hankel singular values given by

σ1 = 0.9998, σ2 = 0.9988, σ3 = 0.9963, σ4 = 0.9923.

The approximation errors and the estimated bounds are listed in the following table.
The table shows that the actual error for an rth-order approximation is almost the same
as 2σr+1 , which would be the estimated bound if we regard σr+1 = σr+2 = · · · = σ4 . In
general, it is not hard to construct an nth-order system so that the rth-order balanced
model reduction error is approximately 2σr+1 but the error bound is arbitrarily close
to 2(n − r)σr+1 . One method to construct such a system is as follows: Let G(s) be a
stable all-pass function, that is, G∼ (s)G(s) = I. Then there is a balanced realization
for G so that the controllability and observability Gramians are P = Q = I. Next,
make a very small perturbation to the balanced realization, then the perturbed system
has a balanced realization with distinct singular values and P = Q ≈ I. This perturbed
system will have the desired properties.

r 0 1 2 3
kG − Gr k∞ 1.9997 1.9983 1.9933 1.9845
P
Bounds: 2 4i=r+1 σi 7.9744 5.9748 3.9772 1.9845
2σr+1 1.9996 1.9976 1.9926 1.9845

The balanced realization and truncation can be done using the following Matlab
commands:

 [Gb , sig] = sysbal(G); % find a balanced realization Gb and the Hankel singular
values sig.

 Gr = strunc(Gb , 2); % truncate to the second-order.

Related MATLAB Commands: reordsys, resid, Hankmr

124 BALANCED MODEL REDUCTION

7.4 Frequency-Weighted Balanced Model Reduction

This section considers the extension of the balanced truncation method to the frequency-
weighted case. Given the original full-order model G ∈ RH∞ , the input weighting
matrix Wi ∈ RH∞ , and the output weighting matrix Wo ∈ RH∞ , our objective is to
find a lower-order model Gr such that

kWo (G − Gr )Wi k∞

is made as small as possible. Assume that G, Wi , and Wo have the following state-space
realizations:
     
A B Ai Bi Ao Bo
G= , Wi = , Wo =
C 0 Ci Di Co Do

with A ∈ Rn×n . Note that there is no loss of generality in assuming D = G(∞) = 0

since otherwise it can be eliminated by replacing Gr with D + Gr .
Now the state-space realization for the weighted transfer matrix is given by
 
A 0 BCi BDi " #
 Bo C Ao 0 0  Ā B̄
Wo GWi =    =:
0 0 Ai Bi  C̄ 0
Do C Co 0 0

Let P̄ and Q̄ be the solutions to the following Lyapunov equations:

ĀP̄ + P̄ Ā∗ + B̄ B̄ ∗ = 0 (7.17)

Q̄Ā + Ā∗ Q̄ + C̄ ∗ C̄ = 0 (7.18)

Then the input weighted Gramian P and the output weighted Gramian Q are defined
by    
  In   In
P := In 0 P̄ , Q := In 0 Q̄
0 0
It can be shown easily that P and Q satisfy the following lower-order equations:
     ∗   ∗
A BCi P P12 P P12 A BCi BDi BDi
∗ + ∗ + =0
0 Ai P12 P22 P12 P22 0 Ai Bi Bi
(7.19)

    ∗     ∗
Q Q12 A 0 A 0 Q Q12 C ∗ Do∗ C ∗ Do∗
+ + =0
Q∗12 Q22 Bo C Ao Bo C Ao Q∗12 Q22 Co∗ Co∗
(7.20)

The computation can be further reduced if Wi = I or Wo = I. In the case of Wi = I,

P can be obtained from
P A∗ + AP + BB ∗ = 0 (7.21)
7.4. Frequency-Weighted Balanced Model Reduction 125

while in the case of Wo = I, Q can be obtained from

QA + A∗ Q + C ∗ C = 0 (7.22)

Now let T be a nonsingular matrix such that

 
∗ −1 ∗ −1 Σ1
T P T = (T ) QT =
Σ2

(i.e., balanced) with Σ1 = diag(σ1 Is1 , . . . , σr Isr ) and Σ2 = diag(σr+1 Isr+1 , . . . , σN IsN )
and partition the system accordingly as
" #  A A12 B1

T AT −1 T B 11
=  A21 A22 B2 
CT −1 0 C1 C2 0

Then a reduced-order model Gr is obtained as

 
A11 B1
Gr =
C1 0

Unfortunately, there is generally no known a priori error bound for the approximation
error and the reduced-order model Gr is not guaranteed to be stable either.
A very special frequency-weighted model reduction problem is the relative error
model reduction problem where the objective is to find a reduced-order model Gr so
that
Gr = G(I + ∆rel )
and k∆rel k∞ is made as small as possible. ∆rel is usually called the relative error. In
the case where G is square and invertible, this problem can be simply formulated as

min G−1 (G − Gr ) ∞ .
degGr ≤r

Of course, the dual approximation problem

Gr = (I + ∆rel )G

can be obtained by taking the transpose of G. It turns out that the approximation Gr
obtained below also serves as a multiplicative approximation:

G = Gr (I + ∆mul )

where ∆mul is usually called the multiplicative error.

Error bounds can be derived if the frequency-weighted balanced truncation method
is applied to the relative and multiplicative approximations.
126 BALANCED MODEL REDUCTION

Theorem 7.13 Let G, G−1 ∈ RH∞ be an nth-order square transfer matrix with a
state-space realization  
A B
G(s) =
C D
Let P and Q be the solutions to
P A∗ + AP + BB ∗ = 0 (7.23)
−1 −1 ∗ ∗ −1 ∗ −1
Q(A − BD C) + (A − BD C) Q + C (D ) D C=0 (7.24)
Suppose
P = Q = diag(σ1 Is1 , . . . , σr Isr , σr+1 Isr+1 , . . . , σN IsN ) = diag(Σ1 , Σ2 )
with σ1 > σ2 > . . . > σN ≥ 0, and let the realization of G be partitioned compatibly with
Σ1 and Σ2 as  
A11 A12 B1
G(s) =  A21 A22 B2 
C1 C2 D
Then  
A11 B1
Gr (s) =
C1 D
is stable and minimum phase. Furthermore,
N 
Y q 
k∆rel k∞ ≤ 1 + 2σi ( 1 + σi + σi ) − 1
2

i=r+1
N 
Y q 
k∆mul k∞ ≤ 1 + 2σi ( 1 + σi + σi ) − 1
2

i=r+1

Related MATLAB Commands: srelbal, sfrwtbal

7.5 Notes and References

Balanced realization was first introduced by Mullis and Roberts [1976] to study roundoff
noise in digital filters. Moore [1981] proposed the balanced truncation method for model-
reduction. The stability properties of the reduced-order model were shown by Pernebo
and Silverman [1982]. The error bound for the balanced model reduction was shown
by Enns [1984a, 1984b], and Glover [1984] subsequently gave an independent proof.
The frequency-weighted balanced model-reduction method was also introduced by Enns
[1984a, 1984b]. The error bounds for the relative error are derived in Zhou [1995]. Other
related results are shown in Green [1988]. Other weighted model-reduction methods
can be found in Al-Saggaf and Franklin [1988], Glover [1986b], Glover, Limebeer and
Hung [1992], Green [1988], Hung and Glover [1986], Zhou [1995], and references therein.
Discrete-time balance model-reduction results can be found in Al-Saggaf and Franklin
[1987], Hinrichsen and Pritchard [1990], and references therein.
7.6. Problems 127

7.6 Problems
Problem 7.1 Use the following relation
d  At A∗ t  ∗ ∗
e Qe = AeAt QeA t + eAt QeA t A
dt
R∞ ∗
to show that P = 0 eAt QeA t dt solves

AP + P A∗ + Q = 0

if A is stable.

Problem 7.2 Let G(s) ∈ H∞ and let g(t) be the inverse Laplace transform of G(s).
Let hi , i = 1, 2, . . . be the variations of the step response of G. Show that
Z ∞
|g(t)|dt = h1 + h2 + h3 + h4 + . . .
0

Problem 7.3 Let Q ≥ 0 be the solution to

QA + A∗ Q + C ∗ C = 0

Suppose Q has m zero eigenvalues. Show that there is a nonsingular matrix T such that
" #  A 0 B1

T AT −1 T B 11
=  A21 A22 B2  , A22 ∈ Rm×m .
CT −1 D C1 0 D

Apply the above result to the following state-space model:

   
−4 −7 −2 1 2  
0 2 1
A= 1 0 0  , B =  0 −1  , C = , D=0
1 1 0
−1 1 0 0 2

Problem 7.4 Let

X
5
α2i
G(s) =
i=1
s + α2i
Find a balanced realization for each of the following α:

α = 2, 4, 20, 100.

Discuss the behavior of the Hankel singular values as α → ∞.

Problem 7.5 Find a transformation so that T P T ∗ = Σ2 , (T ∗ )−1 QT −1 = I. (This

realization is called output normalized realization.)
128 BALANCED MODEL REDUCTION

Problem 7.6 Consider the model reduction error:

 
A11 0 A12 /2 B1  
 0 A −A 12 /2 0  Ae Be
E11 =   A21 −A21
11  =:
 .
A22 B2 Ce 0
0 −2C1 C2 0

Show that  
Σ1 0
P̃ =  0 2 −1
σN Σ1 0 
0 0 2σN IsN
satisfies
1
Ae P̃ + P̃ A∗e + Be Be∗ + ∗
2 P̃ Ce Ce P̃ = 0.
4σN

Problem 7.7 Suppose P and Q are the controllability and observability Gramians of
G(s) = C(sI − A)−1 B ∈ RH∞ . Let Gd (z) = G(s)|s= z+1 = Cd (zI − Ad )−1 Bd + Dd .
z−1
Compute the controllability and observability Gramians Pd and Qd and compare P Q
and Pd Qd .

Problem 7.8 Note that a delay can be approximated as

 τ n
1 − 2n s
e−τ s ≈ τ
1 + 2n s

e−s
for a sufficiently large n. Let a process model be approximated by
1 + Ts
 10
1 − 0.05s 1
G(s) =
1 + 0.05s 1 + sT

For each T = 0, 0.01, 0.1, 1, 10, find a reduced-order model, if possible, using balanced
truncation such that the approximation error is no greater than 0.1.
Chapter 8

Uncertainty and Robustness

In this chapter we briefly describe various types of uncertainties that can arise in phys-
ical systems, and we single out “unstructured uncertainties” as generic errors that are
associated with all design models. We obtain robust stability tests for systems under
various model uncertainty assumptions through the use of the small gain theorem. We
also obtain some sufficient conditions for robust performance under unstructured un-
certainties. The difficulty associated with MIMO robust performance design and the
role of plant condition numbers for systems with skewed performance and uncertainty
specifications are revealed. A simple example is also used to indicate the fundamental
difference between the robustness of an SISO system and that of a MIMO system. In
particular, we show that applying the SISO analysis/design method to a MIMO system
may lead to erroneous results.

8.1 Model Uncertainty

Most control designs are based on the use of a design model. The relationship between
models and the reality they represent is subtle and complex. A mathematical model
provides a map from inputs to responses. The quality of a model depends on how closely
its responses match those of the true plant. Since no single fixed model can respond
exactly like the true plant, we need, at the very least, a set of maps. However, the
modeling problem is much deeper — the universe of mathematical models from which a
model set is chosen is distinct from the universe of physical systems. Therefore, a model
set that includes the true physical plant can never be constructed. It is necessary for
the engineer to make a leap of faith regarding the applicability of a particular design
based on a mathematical model. To be practical, a design technique must help make
this leap small by accounting for the inevitable inadequacy of models. A good model
should be simple enough to facilitate design, yet complex enough to give the engineer
confidence that designs based on the model will work on the true plant.
The term uncertainty refers to the differences or errors between models and reality,

129
130 UNCERTAINTY AND ROBUSTNESS

and whatever mechanism is used to express these errors will be called a representation
of uncertainty. Representations of uncertainty vary primarily in terms of the amount
of structure they contain. This reflects both our knowledge of the physical mechanisms
that cause differences between the model and the plant and our ability to represent these
mechanisms in a way that facilitates convenient manipulation. For example, consider
the problem of bounding the magnitude of the effect of some uncertainty on the output
of a nominally fixed linear system. A useful measure of uncertainty in this context is
to provide a bound on the power spectrum of the output’s deviation from its nominal
response. In the simplest case, this power spectrum is assumed to be independent
of the input. This is equivalent to assuming that the uncertainty is generated by an
additive noise signal with a bounded power spectrum; the uncertainty is represented as
additive noise. Of course, no physical system is linear with additive noise, but some
aspects of physical behavior are approximated quite well using this model. This type
of uncertainty received a great deal of attention in the literature during the 1960s and
1970s, and elegant solutions are obtained for many interesting problems (e.g., white
noise propagation in linear systems, Wiener and Kalman filtering, and LQG optimal
control). Unfortunately, LQG optimal control did not address uncertainty adequately
and hence had less practical impact than might have been hoped.
Generally, the deviation’s power spectrum of the true output from the nominal will
depend significantly on the input. For example, an additive noise model is entirely in-
appropriate for capturing uncertainty arising from variations in the material properties
of physical plants. The actual construction of model sets for more general uncertainty
can be quite difficult. For example, a set membership statement for the parameters of
an otherwise known FDLTI model is a highly structured representation of uncertainty.
It typically arises from the use of linear incremental models at various operating points
(e.g., aerodynamic coefficients in flight control vary with flight environment and air-
craft configurations, and equation coefficients in power plant control vary with aging,
slag buildup, coal composition, etc.). In each case, the amounts of variation and any
known relationships between parameters can be expressed by confining the parameters
to appropriately defined subsets of parameter space. However, for certain classes of sig-
nals (e.g., high-frequency), the parameterized FDLTI model fails to describe the plant
because the plant will always have dynamics that are not represented in the fixed order
model.
In general, we are forced to use not just a single parameterized model but model sets
that allow for plant dynamics that are not explicitly represented in the model struc-
ture. A simple example of this involves using frequency domain bounds on transfer
functions to describe a model set. To use such sets to describe physical systems, the
bounds must roughly grow with frequency. In particular, at sufficiently high frequencies,
phase is completely unknown (i.e., ±180o uncertainties). This is a consequence of dy-
namic properties that inevitably occur in physical systems. This gives a less structured
representation of uncertainty.
8.1. Model Uncertainty 131

Examples of less structured representations of uncertainty are direct set membership

statements for the transfer function matrix of the model. For instance, the statement

P∆ (s) = P (s) + W1 (s)∆(s)W2 (s), σ[∆(jω)] < 1, ∀ω ≥ 0, (8.1)

where W1 and W2 are stable transfer matrices that characterize the spatial and frequency
structure of the uncertainty, confines the matrix P∆ to a neighborhood of the nominal
model P . In particular, if W1 = I and W2 = w(s)I, where w(s) is a scalar function,
then P∆ describes a disk centered at P with radius w(jω) at each frequency, as shown in
Figure 8.1. The statement does not imply a mechanism or structure that gives rise to ∆.
The uncertainty may be caused by parameter changes, as mentioned previously or by
neglected dynamics, or by a host of other unspecified effects. An alternative statement
to equation (8.1) is the so-called multiplicative form:

P∆ (s) = (I + W1 (s)∆(s)W2 (s))P (s). (8.2)

This statement confines P∆ to a normalized neighborhood of the nominal model P . An

advantage of equation (8.2) over (8.1) is that in equation (8.2) compensated transfer
functions have the same uncertainty representation as the raw model (i.e., the weighting
functions apply to P K as well as P ). Some other alternative set membership statements
will be discussed later.

P(j ω )

w(jω )

Figure 8.1: Nyquist diagram of an uncertain model

The best choice of uncertainty representation for a specific FDLTI model depends,
of course, on the errors the model makes. In practice, it is generally possible to repre-
sent some of these errors in a highly structured parameterized form. These are usually
the low-frequency error components. There are always remaining higher-frequency er-
rors, however, which cannot be covered this way. These are caused by such effects as
infinite-dimensional electromechanical resonance, time delays, diffusion processes, etc.
Fortunately, the less structured representations, such as equations (8.1) and (8.2), are
well suited to represent this latter class of errors. Consequently, equations (8.1) and
132 UNCERTAINTY AND ROBUSTNESS

(8.2) have become widely used “generic” uncertainty representations for FDLTI models.
An important point is that the construction of the weighting matrices W1 and W2 for
multivariable systems is not trivial.
Motivated from these observations, we will focus for the moment on the multiplica-
tive description of uncertainty. We will assume that P∆ in equation (8.2) remains a
strictly proper FDLTI system for all ∆. More general perturbations (e.g., time varying,
infinite dimensional, nonlinear) can also be covered by this set provided they are given
appropriate “conic sector” interpretations via Parseval’s theorem. This connection is
developed in [Safonov, 1980] and [Zames, 1966] and will not be pursued here.
When used to represent the various high-frequency mechanisms mentioned previ-
ously, the weighting functions in equation (8.2) commonly have the properties illustrated
in Figure 8.2. They are small ( 1) at low frequencies and increase to unity and above
at higher frequencies. The growth with frequency inevitably occurs because phase un-
certainties eventually exceed ±180 degrees and magnitude deviations eventually exceed
the nominal transfer function magnitudes. Readers who are skeptical about this reality
are encouraged to try a few experiments with physical devices.

nominal model

log ω

actual model

Figure 8.2: Typical behavior of multiplicative uncertainty: pδ (s) = [1 + w(s)δ(s)]p(s)

Also note that the representation of uncertainty in equation (8.2) can be used to
include perturbation effects that are in fact certain. A nonlinear element may be quite
accurately modeled, but because our design techniques cannot effectively deal with the
nonlinearity, it is treated as a conic sector nonlinearity.1 As another example, we may
deliberately choose to ignore various known dynamic characteristics in order to achieve
a simple nominal design model. One such instance is the model reduction process
discussed in the last chapter.
1 See, for example, Safonov [1980] and Zames [1966].
8.1. Model Uncertainty 133

Example 8.1 Let a dynamical system be described by

10 (2 + 0.2α)s2 + (2 + 0.3α + 0.4β)s + (1 + 0.2β)
P (s, α, β) = , α, β ∈ [−1, 1]
(s2 + 0.5s + 1)(s2 + 2s + 3)(s2 + 3s + 6)
Then for each frequency, all possible frequency responses with varying parameters α and
β are in a box, as shown in Figure 8.3. We can also obtain an unstructured uncertainty
bound for this system. In fact, we have

P (s, α, β) ∈ {P0 + W ∆ | k∆k ≤ 1}

with P0 := P (s, 0, 0) and

10 0.2s2 + 0.7s + 0.2
W (s) = P (s, 1, 1) − P (s, 0, 0) = 2
(s + 0.5s + 1)(s2 + 2s + 3)(s2 + 3s + 6)
The frequency response P0 + W ∆ is shown in Figure 8.3 as circles.

0.5

−0.5

−1

−1.5

−2

−2.5

−3

−3.5
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5

Figure 8.3: Nyquist diagram of uncertain system and disk covering

Another way to bound the frequency response is to treat α and β as norm bounded
uncertainties; that is,

P (s, α, β) ∈ {P0 + W1 ∆1 + W2 ∆2 | k∆i k∞ ≤ 1}

with P0 = P (s, 0, 0) and

10(0.2s2 + 0.3s)
W1 = ,
(s2 + 0.5s + 1)(s2 + 2s + 3)(s2 + 3s + 6)
134 UNCERTAINTY AND ROBUSTNESS

10(0.4s + 0.2)
W2 =
(s2 + 0.5s + 1)(s2 + 2s + 3)(s2 + 3s + 6)
It is in fact easy to show that
{P0 + W1 ∆1 + W2 ∆2 | k∆i k∞ ≤ 1} = {P0 + W ∆ | k∆k∞ ≤ 1}
with |W | = |W1 | + |W2 |. The frequency response P0 + W ∆ is shown in Figure 8.4. This
bounding is clearly more conservative.
1

0.5

−0.5

−1

−1.5

−2

−2.5

−3

−3.5
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5

Figure 8.4: A conservative covering

Example 8.2 Consider a process control model

ke−τ s
G(s) = , 4 ≤ k ≤ 9, 2 ≤ T ≤ 3, 1 ≤ τ ≤ 2.
Ts + 1
Take the nominal model as
6.5
G0 (s) =
(2.5s + 1)(1.5s + 1)
Then for each frequency, all possible frequency responses are in a box, as shown in
Figure 8.5. To obtain an unstructured uncertainty bound for this system, plot the error
∆a (jω) = G(jω) − G0 (jω)
for a set of parameters, as shown in Figure 8.6, and then use the Matlab command
ginput to pick a set of upper-bound frequency responses and use fitmag to fit a stable
and minimum phase transfer function to the upper-bound frequency responses.
8.1. Model Uncertainty 135

ω =2
0
ω =1 ω =3 ω = 0.01

−2
ω = 0.05
Imaginary

ω = 0.1

−4
ω = 0.8
ω = 0.2
ω = 0.5
−6 ω = 0.3

−8
−5 0 5 10
Real

Figure 8.5: Uncertain delay system and G0

 mf= ginput(50) % pick 50 points: the first column of mf is the frequency points
and the second column of mf is the corresponding magnitude responses.
 magg=vpck(mf(:,2),mf(:,1)); % pack them as a varying matrix.
 Wa =fitmag(magg); % choose the order of Wa online. A third-order Wa is
sufficient for this example.
 [A,B,C,D]=unpck(Wa) % converting into state-space.
 [Z, P, K]=ss2zp(A,B,C,D) % converting into zero/pole/gain form.

We get
0.0376(s + 116.4808)(s + 7.4514)(s + 0.2674)
Wa (s) =
(s + 1.2436)(s + 0.5575)(s + 4.9508)
and the frequency response of Wa is also plotted in Figure 8.6. Similarly, define the
multiplicative uncertainty
G(s) − G0 (s)
∆m (s) :=
G0 (s)
and a Wm can be found such that |∆m (jω)| ≤ |Wm (jω)|, as shown in Figure 8.7. A
Wm is given by
2.8169(s + 0.212)(s2 + 2.6128s + 1.732)
Wm =
s2 + 2.2425s + 2.6319
136 UNCERTAINTY AND ROBUSTNESS

1
10

0
10

−1
10

−2
10 −2 −1 0 1 2
10 10 10 10 10

Figure 8.6: ∆a (dashed line) and a bound Wa (solid line)

3
10

2
10

1
10

0
10

−1
10

−2
10

−3
10

−4
10

−5
10 −2 −1 0 1 2
10 10 10 10 10

Figure 8.7: ∆m (dashed line) and a bound Wm (solid line)

8.2. Small Gain Theorem 137

Note that this Wm is not proper since G0 and G do not have the same relative degrees.
To get a proper Wm , we need to choose a nominal model G0 having the same the relative
order as that of G.

The following terminologies are used in this book:

Definition 8.1 Given the description of an uncertainty model set Π and a set of per-
formance objectives, suppose P ∈ Π is the nominal design model and K is the resulting
controller. Then the closed-loop feedback system is said to have

Nominal Stability (NS): if K internally stabilizes the nominal model P .

Robust Stability (RS): if K internally stabilizes every plant belonging to Π.

Nominal Performance (NP): if the performance objectives are satisfied for the
nominal plant P .

Robust Performance (RP): if the performance objectives are satisfied for every
plant belonging to Π.

The nominal stability and performance can be easily checked using various standard
techniques. The conditions for which the robust stability and robust performance are
satisfied under various assumptions on the uncertainty set Π will be considered in the
following sections.

Related MATLAB Commands: magfit, drawmag, fitsys, genphase, vunpck,

vabs, vinv, vimag, vreal, vcjt, vebe

8.2 Small Gain Theorem

This section and the next section consider the stability test of a nominally stable system
under unstructured perturbations. The basis for the robust stability criteria derived in
the sequel is the so-called small gain theorem.
Consider the interconnected system shown in Figure 8.8 with M (s) a stable p × q
transfer matrix.

Theorem 8.1 (Small Gain Theorem) Suppose M ∈ RH∞ and let γ > 0. Then the
interconnected system shown in Figure 8.8 is well-posed and internally stable for all
∆(s) ∈ RH∞ with

(a) k∆k∞ ≤ 1/γ if and only if kM (s)k∞ < γ

(b) k∆k∞ < 1/γ if and only if kM (s)k∞ ≤ γ

138 UNCERTAINTY AND ROBUSTNESS

w1 e1
-e - ∆
+ 6
+

e2 + w2
?
e+
M 

Figure 8.8: M − ∆ loop for stability analysis

Proof. We shall only prove part (a). The proof for part (b) is similar. Without loss
of generality, assume γ = 1.
(Sufficiency) It is clear that M (s)∆(s) is stable since both M (s) and ∆(s) are stable.
Thus by Theorem 5.5 (or Corollary 5.4) the closed-loop system is stable if det(I − M ∆)
has no zero in the closed right-half plane for all ∆ ∈ RH∞ and k∆k∞ ≤ 1. Equivalently,
the closed-loop system is stable if

inf σ (I − M (s)∆(s)) 6= 0
s∈C+

for all ∆ ∈ RH∞ and k∆k∞ ≤ 1. But this follows from

inf σ (I − M (s)∆(s)) ≥ 1− sup σ̄ (M (s)∆(s)) = 1−kM (s)∆(s)k∞ ≥ 1−kM (s)k∞ > 0.

s∈C+ s∈C+

(Necessity) This will be shown by contradiction. Suppose kM k∞ ≥ 1. We will show

that there exists a ∆ ∈ RH∞ with k∆k∞ ≤ 1 such that det(I − M (s)∆(s)) has a
zero on the imaginary axis, so the system is unstable. Suppose ω0 ∈ R+ ∪ {∞} is
such that σ̄(M (jω0 )) ≥ 1. Let M (jω0 ) = U (jω0 )Σ(jω0 )V ∗ (jω0 ) be a singular value
decomposition with  
U (jω0 ) = u1 u2 · · · up
 
V (jω0 ) = v1 v2 · · · vq
 
σ1
 σ2 
Σ(jω0 ) =  
..
.
To obtain a contradiction, it now suffices to construct a ∆ ∈ RH∞ such that ∆(jω0 ) =
∗
σ1 v1 u1 and k∆k∞ ≤ 1. Indeed, for such ∆(s),
1

det(I − M (jω0 )∆(jω0 )) = det(I − U ΣV ∗ v1 u∗1 /σ1 ) = 1 − u∗1 U ΣV ∗ v1 /σ1 = 0

and thus the closed-loop system is either not well-posed (if ω0 = ∞) or unstable (if
ω ∈ R). There are two different cases:
8.2. Small Gain Theorem 139

(1) ω0 = 0 or ∞: then U and V are real matrices. In this case, ∆(s) can be chosen as

1
∆= v1 u∗1 ∈ Rq×p
σ1

(2) 0 < ω0 < ∞: write u1 and v1 in the following form:

 
v11 ejφ1
   v12 ejφ2 
 
u∗1 = u11 ejθ1 u12 ejθ2 · · · u1p ejθp , v1 =  .. 
 . 
v1q ejφq

where u1i ∈ R and v1j ∈ R are chosen so that θi , φj ∈ [−π, 0) for all i, j.

Choose βi ≥ 0 and αj ≥ 0 so that

   
βi − jω0 αj − jω0
∠ = θi , ∠ = φj
βi + jω0 αj + jω0

for i = 1, 2, . . . , p and j = 1, 2, . . . , q. Let

 1 −s

v11 α
1 
α1 +s h i
..  β −s β −s
∆(s) =  .  u11 β11 +s · · · u1p βpp +s ∈ RH∞
σ1 αq −s
v1q αq +s

∗
Then k∆k∞ = 1/σ1 ≤ 1 and ∆(jω0 ) = 1
σ1 v1 u1 .

The theorem still holds even if ∆ and M are infinite dimensional. This is summarized
as the following corollary.

Corollary 8.2 The following statements are equivalent:

(i) The system is well-posed and internally stable for all ∆ ∈ H∞ with k∆k∞ < 1/γ;

(ii) The system is well-posed and internally stable for all ∆ ∈ RH∞ with k∆k∞ < 1/γ;

(iii) The system is well-posed and internally stable for all ∆ ∈ Cq×p with k∆k < 1/γ;

(iv) kM k∞ ≤ γ.

Remark 8.1 It can be shown that the small gain condition is sufficient to guarantee
internal stability even if ∆ is a nonlinear and time-varying “stable” operator with an
appropriately defined stability notion, see Desoer and Vidyasagar [1975]. 3
140 UNCERTAINTY AND ROBUSTNESS

The following lemma shows that if kM k∞ > γ, there exists a destabilizing ∆ with
k∆k∞ < 1/γ such that the closed-loop system has poles in the open right-half plane.
(This is stronger than what is given in the proof of Theorem 8.1.)
Lemma 8.3 Suppose M ∈ RH∞ and kM k∞ > γ. Then there exists a σ0 > 0 such
that for any given σ ∈ [0, σ0 ] there exists a ∆ ∈ RH∞ with k∆k∞ < 1/γ such that
det(I − M (s)∆(s)) has a zero on the axis Re(s) = σ.

Proof. Without loss of generality, assume γ = 1. Since M ∈ RH∞ and

kM k∞ > 1, there exists a 0 < ω0 < ∞ such that kM (jω0 )k > 1. Given any γ such that
1 < γ < kM (jω0 )k, there is a sufficiently small σ0 > 0 such that
min kM (σ + jω0 )k ≥ γ
σ∈[0,σ0 ]

and s s
ω02 + (σ0 + α)2 ω02 + (σ0 + β)2
<γ
ω02 + (σ0 − α)2 ω02 + (σ0 − β)2
for any α ≥ 0 and β ≥ 0.
Now let σ ∈ [0, σ0 ] and let M (σ + jω0 ) = U ΣV ∗ be a singular value decomposition
with  
U = u1 u2 · · · up
 
V = v1 v2 · · · vq
 
σ1
 σ2 
Σ= .
..
.
Write u1 and v1 in the following form:
 
v11 ejφ1
   v12 ejφ2 
 
u∗1 = u11 ejθ1 u12 ejθ2 · · · u1p ejθp , v1 =  .. 
 . 
v1q ejφq
where u1i ∈ R and v1j ∈ R are chosen so that θi , φj ∈ [−π, 0) for all i, j.
Choose βi ≥ 0 and αj ≥ 0 so that
   
βi − σ − jω0 αj − σ − jω0
∠ = θi , ∠ = φj
βi + σ + jω0 αj + σ + jω0
for i = 1, 2, . . . , p and j = 1, 2, . . . , q. Let
 α1 −s 
α̃1 v11 α
1 
1 +s h i
..  β −s β −s
∆(s) =  .  β̃1 u11 β11 +s · · · β̃p u1p βpp +s ∈ RH∞
σ1 αq −s
α̃q v1q αq +s
8.3. Stability under Unstructured Uncertainties 141

where s s
ω02 + (σ + βi )2 ω02 + (σ + αj )2
β̃i := , α̃j :=
ω02 + (σ − βi )2 ω02 + (σ − αj )2
Then n o n o
maxj {α̃j } maxi β̃i maxj {α̃j } maxi β̃i
k∆k∞ ≤ ≤ <1
σ1 γ
and
1
∆(σ + jω0 ) = v1 u∗1
σ1
det (I − M (σ + jω0 )∆(σ + jω0 )) = 0
Hence s = σ + jω0 is a zero for the transfer function det (I − M (s)∆(s)). 2

The preceding lemma plays a key role in the necessity proofs of many robust stability
tests in the sequel.

8.3 Stability under Unstructured Uncertainties

The small gain theorem in the last section will be used here to derive robust stability
tests under various assumptions of model uncertainties. The modeling error ∆ will again
be assumed to be stable. (Most of the robust stability tests discussed in the sequel can
be generalized easily to the unstable ∆ case with some mild assumptions on the number
of unstable poles of the uncertain model; we encourage readers to fill in the details.)
In addition, we assume that the modeling error ∆ is suitably scaled with weighting
functions W1 and W2 (i.e., the uncertainty can be represented as W1 ∆W2 ).

-f - K - Π -
6
−

Figure 8.9: Unstructured robust stability analysis

We shall consider the standard setup shown in Figure 8.9, where Π is the set of un-
certain plants with P ∈ Π as the nominal plant and with K as the internally stabilizing
controller for P . The sensitivity and complementary sensitivity matrix functions are
defined, as usual, as
So = (I + P K)−1 , To = I − So
and
Si = (I + KP )−1 , Ti = I − Si .
142 UNCERTAINTY AND ROBUSTNESS

Recall that the closed-loop system is well-posed and internally stable if and only if
 −1  
I K (I + KΠ)−1 −K(I + ΠK)−1
= ∈ RH∞
−Π I (I + ΠK)−1 Π (I + ΠK)−1

for all Π ∈ Π.

8.3.1 Additive Uncertainty

We assume that the model uncertainty can be represented by an additive perturbation:

Π = P + W1 ∆W2 .

Theorem 8.4 Let Π = {P + W1 ∆W2 : ∆ ∈ RH∞ } and let K be a stabilizing con-

troller for the nominal plant P . Then the closed-loop system is well-posed and internally
stable for all k∆k∞ < 1 if and only if kW2 KSo W1 k∞ ≤ 1.

Proof. Let Π = P + W1 ∆W2 ∈ Π. Then

 −1
I K
−Π I
 
(I + KSo W1 ∆W2 )−1 Si −KSo(I + W1 ∆W2 KSo )−1
=
(I + So W1 ∆W2 K)−1 So (P + W1 ∆W2 ) So (I + W1 ∆W2 KSo )−1

is well-posed and internally stable if (I + ∆W2 KSo W1 )−1 ∈ RH∞ since

det(I + KSo W1 ∆W2 ) = det(I + W1 ∆W2 KSo ) = det(I + So W1 ∆W2 K)

= det(I + ∆W2 KSo W1 ).

But (I + ∆W2 KSo W1 )−1 ∈ RH∞ is guaranteed if k∆W2 KSo W1 k∞ < 1 (small gain
theorem). Hence kW2 KSo W1 k∞ ≤ 1 is sufficient for robust stability.
To show the necessity, note that robust stability implies that

K(I + ΠK)−1 = KSo (I + W1 ∆W2 KSo )−1 ∈ RH∞

for all admissible ∆. This, in turn, implies that

∆W2 K(I + ΠK)−1 W1 = I − (I + ∆W2 KSo W1 )−1 ∈ RH∞

for all admissible ∆. By the small gain theorem, this is true for all ∆ ∈ RH∞ with
k∆k∞ < 1 only if kW2 KSo W1 k∞ ≤ 1. 2
8.3. Stability under Unstructured Uncertainties 143

dm d˜
z w ? ?
- W2 - ∆ - e
- W1 Wd

d
r - e- - P ? ?
- e - e y- We e-
K
− 6

Figure 8.10: Output multiplicative perturbed systems

8.3.2 Multiplicative Uncertainty

In this section, we assume that the system model is described by the following set of
multiplicative perturbations:

Π = (I + W1 ∆W2 )P

with W1 , W2 , ∆ ∈ RH∞ . Consider the feedback system shown in Figure 8.10.

Theorem 8.5 Let Π = {(I + W1 ∆W2 )P : ∆ ∈ RH∞ } and let K be a stabilizing con-
troller for the nominal plant P . Then the closed-loop system is well-posed and internally
stable for all ∆ ∈ RH∞ with k∆k∞ < 1 if and only if kW2 To W1 k∞ ≤ 1.

Proof. We shall first prove that the condition is necessary for robust stability. Suppose
kW2 To W1 k∞ > 1. Then by Lemma 8.3, for any given sufficiently small σ > 0, there
is a ∆ ∈ RH∞ with k∆k∞ < 1 such that (I + ∆W2 To W1 )−1 has poles on the axis
Re(s) = σ. This implies that

(I + ΠK)−1 = So (I + W1 ∆W2 To )−1

has poles on the axis Re(s) = σ since σ can always be chosen so that the unstable poles
are not cancelled by the zeros of So . Hence kW2 To W1 k∞ ≤ 1 is necessary for robust
stability. The sufficiency follows from the small gain theorem. 2
144 UNCERTAINTY AND ROBUSTNESS

8.3.3 Coprime Factor Uncertainty

As another example, consider a left coprime factor perturbed plant described in Fig-
ure 8.11.
z1 − z2
-∆
˜N - e ˜M 
∆
w
r ? y
-e - K - Ñ -e - M̃ −1 -
−6

Figure 8.11: Left coprime factor perturbed systems

Theorem 8.6 Let

˜ M )−1 (Ñ + ∆
Π = (M̃ + ∆ ˜ N)
˜M, ∆
with M̃ , Ñ, ∆ ˜ N ∈ RH∞ . The transfer matrices (M̃ , Ñ ) are assumed to be a stable
left coprime factorization of P (i.e.,
 P = M̃−1 Ñ ), and K internally stabilizes the
nominal system P . Define ∆ := ∆N ∆ ˜ ˜ M . Then the closed-loop system is well-
posed and internally stable for all k∆k∞ < 1 if and only if
 
K
−1 −1
I (I + P K) M̃ ≤ 1
∞

Proof. Let K = U V −1 be a right coprime factorization over RH∞ . By Lemma 5.7,

the closed-loop system is internally stable if and only if
 −1
˜ N )U + (M̃ + ∆
(Ñ + ∆ ˜ M )V ∈ RH∞ (8.3)

Since K stabilizes P , (Ñ U + M̃ V )−1 ∈ RH∞ . Hence equation (8.3) holds if and only if
 −1
I + (∆ ˜ M V )(Ñ U + M̃V )−1
˜ NU + ∆ ∈ RH∞

By the small gain theorem, the above is true for all k∆k∞ < 1 if and only if
   
U K
−1 −1 −1
V (Ñ U + M̃V ) = I (I + P K) M̃ ≤1
∞ ∞

2
8.3. Stability under Unstructured Uncertainties 145

8.3.4 Unstructured Robust Stability Tests

Table 8.1 summaries robust stability tests on the plant uncertainties under various
assumptions. All of the tests pertain to the standard setup shown in Figure 8.9, where
Π is the set of uncertain plants with P ∈ Π as the nominal plant and with K as the
internally stabilizing controller of P .

W1 ∈ RH∞ W2 ∈ RH∞ ∆ ∈ RH∞ k∆k∞ < 1

Perturbed Model Sets Representative Types of Robust Stability Tests

Π Uncertainty Characterized
output (sensor) errors
(I + W1 ∆W2 )P neglected HF dynamics kW2 To W1 k∞ ≤ 1
uncertain rhp zeros
input (actuators) errors
P (I + W1 ∆W2 ) neglected HF dynamics kW2 Ti W1 k∞ ≤ 1
uncertain rhp zeros
LF parameter errors
(I + W1 ∆W2 )−1 P uncertain rhp poles kW2 So W1 k∞ ≤ 1

LF parameter errors
P (I + W1 ∆W2 )−1 uncertain rhp poles kW2 Si W1 k∞ ≤ 1

additive plant errors

P + W1 ∆W2 neglected HF dynamics kW2 KSo W1 k∞ ≤ 1
uncertain rhp zeros
LF parameter errors
P (I + W1 ∆W2 P )−1 uncertain rhp poles kW2 So P W1 k∞ ≤ 1

˜ M )−1 (Ñ + ∆
(M̃ + ∆ ˜ N) LF parameter errors  
K
−1
P = M̃ Ñ neglected HF dynamics −1
≤1
I So M̃
  ∞
∆= ∆ ˜N ∆˜M uncertain rhp poles & zeros
−1
(N + ∆N )(M + ∆M ) LF parameter errors −1
P =N −1 M Si [K I] ≤ 1
M  neglected HF dynamics ∞
∆N
∆= uncertain rhp poles & zeros
∆M

Table 8.1: Unstructured robust stability tests (HF: high frequency, LF: low frequency)
146 UNCERTAINTY AND ROBUSTNESS

Table 8.1 should be interpreted as follows:

UNSTRUCTURED ANALYSIS THEOREM

Given NS & Perturbed Model Sets

Then Closed-Loop Robust Stability
if and only if Robust Stability Tests

The table also indicates representative types of physical uncertainties that can be
usefully represented by cone-bounded perturbations inserted at appropriate locations.
For example, the representation P∆ = (I + W1 ∆W2 )P in the first row is useful for out-
put errors at high frequencies (HF), covering such things as unmodeled high-frequency
dynamics of sensors or plants, including diffusion processes, transport lags, electrome-
chanical resonances, etc. The representation P∆ = P (I + W1 ∆W2 ) in the second row
covers similar types of errors at the inputs. Both cases should be contrasted with
the third and the fourth rows, which treat P (I + W1 ∆W2 )−1 and (I + W1 ∆W2 )−1 P .
These representations are more useful for variations in modeled dynamics, such as low-
frequency (LF) errors produced by parameter variations with operating conditions, with
aging, or across production copies of the same plant.
Note from the table that the stability requirements on ∆ do not limit our ability
to represent variations in either the number or locations of rhp singularities, as can be
seen from some simple examples.

Example 8.3 Suppose an uncertain system with changing numbers of right-half plane
poles is described by  
1
P∆ = : δ ∈ R, |δ| ≤ 1 .
s−δ
1 1
Then P1 = ∈ P∆ has one right-half plane pole and P2 = ∈ P∆ has no
s−1 s+1
right-half plane pole. Nevertheless, the set of P∆ can be covered by a set of feedback
uncertain plants:

P∆ ⊂ Π := P (1 + δP )−1 : δ ∈ RH∞ , kδk∞ ≤ 1

1
with P = .
s
8.4. Robust Performance 147

Example 8.4 As another example, consider the following set of plants:

s+1+α
P∆ = , |α| ≤ 2.
(s + 1)(s + 2)
This set of plants has changing numbers of right-half plane zeros since the plant has no
right-half plane zero when α = 0 and has one right-half plane zero when α = −2. The
uncertain plant can be covered by a set of multiplicative perturbed plants:
 
1 2δ
P∆ ⊂ Π := (1 + ), δ ∈ RH∞ , kδk∞ ≤ 1 .
s+2 s+1

It should be noted that this covering can be quite conservative.

8.4 Robust Performance

Consider the perturbed system shown in Figure 8.12 with the set of perturbed models
described by a set Π. Suppose the weighting matrices Wd , We ∈ RH∞ and the perfor-

d˜
?
Wd

d
r-e - K - P∆ ∈ Π ?
- e y- We e-

−6

Figure 8.12: Diagram for robust performance analysis

mance criterion is to keep the error e as small as possible in some sense for all possible
models belonging to the set Π. In general, the set Π can be either a parameterized set
or an unstructured set such as those described in Table 8.1. The performance specifica-
tions are usually specified in terms of the magnitude of each component e in the time
domain with respect to bounded disturbances, or, alternatively and more conveniently,
some requirements on the closed-loop frequency response of the transfer matrix between
d˜ and e (say, integral of square error or the magnitude of the steady-state error with
respect to sinusoidal disturbances). The former design criterion leads to the so-called
148 UNCERTAINTY AND ROBUSTNESS

L1 -optimal control framework and the latter leads to H2 and H∞ design frameworks,
respectively. In this section, we will focus primarily on the H∞ performance objectives
with unstructured model uncertainty descriptions. The performance under structured
uncertainty will be considered in Chapter 10.
Suppose the performance criterion is to keep the worst-case energy of the error e as
small as possible over all d˜ of unit energy, for example,

sup kek2 ≤ 
kd˜k2 ≤1

for some small . By scaling the error e (i.e., by properly selecting We ) we can assume
without loss of generality that  = 1.
Let Ted˜ denote the transfer matrix between d˜ and e, then

Ted̃ = We (I + P∆ K)−1 Wd , P∆ ∈ Π. (8.4)

Then the robust performance criterion in this case can be described as requiring that
the closed-loop system be robustly stable and that

T ≤ 1, ∀P∆ ∈ Π. (8.5)
ed̃ ∞

More specifically, an output multiplicatively perturbed system will be analyzed first.

The analysis for other classes of models can be done analogously. The perturbed model
can be described as

Π := {(I + W1 ∆W2 )P : ∆ ∈ RH∞ , k∆k∞ < 1} (8.6)

with W1 , W2 ∈ RH∞ . The explicit system diagram is as shown in Figure 8.10. For this
class of models, we have

Ted˜ = We So (I + W1 ∆W2 To )−1 Wd ,

and the robust performance is satisfied iff

kW2 To W1 k∞ ≤ 1

and
T ≤ 1, ∀∆ ∈ RH∞ , k∆k < 1.
ed̃ ∞ ∞

The exact analysis for this robust performance problem is not trivial and will be given
in Chapter 10. However, some sufficient conditions are relatively easy to obtain by
bounding these two inequalities, and they may shed some light on the nature of these
problems. It will be assumed throughout that the controller K internally stabilizes the
nominal plant P .

Theorem 8.7 Suppose P∆ ∈ {(I + W1 ∆W2 )P : ∆ ∈ RH∞ , k∆k∞ < 1} and K in-
ternally stabilizes P . Then the system robust performance is guaranteed if either one of
the following conditions is satisfied:
8.4. Robust Performance 149

(i) for each frequency ω

σ(Wd )σ(We So ) + σ(W1 )σ(W2 To ) ≤ 1; (8.7)

(ii) for each frequency ω

κ(W1−1 Wd )σ(We So Wd ) + σ(W2 To W1 ) ≤ 1 (8.8)
where W1 and Wd are assumed to be invertible and κ(W1−1 Wd ) is the condition
number.

Proof. It is obvious that both condition (8.7) and condition (8.8) guarantee that
kW2 To W1 k∞ ≤ 1. So it is sufficient to show that Ted˜ ∞ ≤ 1, ∀∆ ∈ RH∞ , k∆k∞ < 1.
Now for any frequency ω, it is easy to see that
σ(Ted˜) ≤ σ(We So )σ[(I + W1 ∆W2 To )−1 ]σ(Wd )
σ(We So )σ(Wd ) σ(We So )σ(Wd )
= ≤
σ(I + W1 ∆W2 To ) 1 − σ(W1 ∆W2 To )
σ(We So )σ(Wd )
≤ .
1 − σ(W1 )σ(W2 To )σ̄(∆)
Hence condition (8.7) guarantees σ(Ted̃ ) ≤ 1 for all ∆ ∈ RH∞ with k∆k∞ < 1 at all
frequencies.
Similarly, suppose W1 and Wd are invertible; write
Ted̃ = We So Wd (W1−1 Wd )−1 (I + ∆W2 To W1 )−1 (W1−1 Wd ),
and then
σ(We So Wd )κ(W1−1 Wd )
σ(Ted˜) ≤ .
1 − σ(W2 To W1 )σ̄(∆)
Hence by condition (8.8), σ(Ted̃ ) ≤ 1 is guaranteed for all ∆ ∈ RH∞ with k∆k∞ < 1
at all frequencies. 2

Remark 8.2 It is not hard to show that either one of the conditions in the theorem is
also necessary for scalar valued systems.
Remark 8.3 Suppose κ(W1−1 Wd ) ≈ 1 (weighting matrices satisfying this condition
are usually called round weights). This is particularly the case if W1 = w1 (s)I and
Wd = wd (s)I. Recall that σ(We So Wd ) ≤ 1 is the necessary and sufficient condition for
nominal performance and that σ(W2 To W1 ) ≤ 1 is the necessary and sufficient condition
for robust stability. Hence the condition (ii) in Theorem 8.7 is almost guaranteed by
NP + RS (i.e., RP is almost guaranteed by NP + RS). Since RP implies NP + RS, we
have NP + RS ≈ RP. (In contrast, such a conclusion cannot be drawn in the skewed
case, which will be considered in the next section.) Since condition (ii) implies NP+RS,
we can also conclude that condition (ii) is almost equivalent to RP (i.e., beside being
sufficient, it is almost necessary). 3
150 UNCERTAINTY AND ROBUSTNESS

Remark 8.4 Note that in light of the equivalence relation between the robust stabil-
ity and nominal performance, it is reasonable to conjecture that the preceding robust
performance problem is equivalent to the robust stability problem in Figure 8.9 with
the uncertainty model set given by
Π := (I + Wd ∆e We )−1 (I + W1 ∆W2 )P
and k∆e k∞ < 1, k∆k∞ < 1, as shown in Figure 8.13. This conjecture is indeed true;
however, the equivalent model uncertainty is structured, and the exact stability analysis
for such systems is not trivial and will be studied in Chapter 10. 3

d˜
∆e 
z w ?
- W2 - ∆ - W1 Wd

e- K - P -?
e -?
e - We e

− 6

Figure 8.13: Robust performance with unstructured uncertainty vs. robust stability
with structured uncertainty

Remark 8.5 Note that if W1 and Wd are invertible, then Ted˜ can also be written as
 −1
Ted̃ = We So Wd I + (W1−1 Wd )−1 ∆W2 To W1 (W1−1 Wd ) .
So another alternative sufficient condition for robust performance can be obtained as
σ(We So Wd ) + κ(W1−1 Wd )σ(W2 To W1 ) ≤ 1.

A similar situation also occurs in the skewed case below. We will not repeat all these
variations. 3

8.5 Skewed Specifications

We now consider the system with skewed specifications (i.e., the uncertainty and per-
formance are not measured at the same location). For instance, the system performance
is still measured in terms of output sensitivity, but the uncertainty model is in input
multiplicative form:
Π := {P (I + W1 ∆W2 ) : ∆ ∈ RH∞ , k∆k∞ < 1} .
8.5. Skewed Specifications 151

z
- ∆
w d˜
? ?
W2 W1 Wd
e
e - K 6 -?
e - P -?
e - We -
− 6

Figure 8.14: Skewed problems

The system block diagram is shown in Figure 8.14.

For systems described by this class of models, the robust stability condition becomes

kW2 Ti W1 k∞ ≤ 1,

and the nominal performance condition becomes

kWe So Wd k∞ ≤ 1.

To consider the robust performance, let T̃ed˜ denote the transfer matrix from d˜ to e.
Then

T̃ed̃ = We So (I + P W1 ∆W2 KSo )−1 Wd

 −1
= We So Wd I + (Wd−1 P W1 )∆(W2 Ti W1 )(Wd−1 P W1 )−1 .

The last equality follows if W1 , Wd , and P are invertible and, if W2 is invertible, can
also be written as
 −1
T̃ed̃ = We So Wd (W1−1 Wd )−1 I + (W1−1 P W1 )∆(W2 P −1 W2−1 )(W2 To W1 ) (W1−1 Wd ).

Then the following results follow easily.

Theorem 8.8 Suppose P∆ ∈ Π = {P (I + W1 ∆W2 ) : ∆ ∈ RH∞ , k∆k∞ < 1} and K

internally stabilizes P . Assume that P, W1 , W2 , and Wd are square and invertible. Then
the system robust performance is guaranteed if either one of the following conditions is
satisfied:
(i) for each frequency ω

σ(We So Wd ) + κ(Wd−1 P W1 )σ(W2 Ti W1 ) ≤ 1; (8.9)

(ii) for each frequency ω

κ(W1−1 Wd )σ(We So Wd ) + σ(W1−1 P W1 )σ(W2 P −1 W2−1 )σ(W2 To W1 ) ≤ 1. (8.10)

152 UNCERTAINTY AND ROBUSTNESS

Remark 8.6 If the appropriate invertibility conditions are not satisfied, then an alter-
native sufficient condition for robust performance can be given by

σ(Wd )σ(We So ) + σ(P W1 )σ(W2 KSo ) ≤ 1.

Similar to the previous case, there are many different variations of sufficient conditions
although equation (8.10) may be the most useful one. 3

Remark 8.7 It is important to note that in this case, the robust stability condition is
given in terms of Li = KP while the nominal performance condition is given in terms
of Lo = P K. These classes of problems are called skewed problems or problems with
skewed specifications.2 Since, in general, P K 6= KP , the robust stability margin or
tolerances for uncertainties at the plant input and output are generally not the same.
3

Remark 8.8 It is also noted that the robust performance condition is related to the
condition number of the weighted nominal model. So, in general, if the weighted nominal
model is ill-conditioned at the range of critical frequencies, then the robust performance
condition may be far more restrictive than the robust stability condition and the nominal
performance condition together. For simplicity, assume W1 = I, Wd = I and W2 = wt I,
where wt ∈ RH∞ is a scalar function. Further, P is assumed to be invertible. Then
the robust performance condition (8.10) can be written as

σ(We So ) + κ(P )σ(wt To ) ≤ 1, ∀ω.

Comparing these conditions with those obtained for nonskewed problems shows that
the condition related to robust stability is scaled by the condition number of the plant.3
Since κ(P ) ≥ 1, it is clear that the skewed specifications are much harder to satisfy
if the plant is not well conditioned. This problem will be discussed in more detail in
Section 10.3.3 of Chapter 10. 3

Remark 8.9 Suppose K is invertible, then T̃ed̃ can be written as

T̃ed˜ = We K −1 (I + Ti W1 ∆W2 )−1 Si KWd .

Assume further that We = I, Wd = ws I, W2 = I, where ws ∈ RH∞ is a scalar function.

Then a sufficient condition for robust performance is given by

κ(K)σ(Si ws ) + σ(Ti W1 ) ≤ 1, ∀ω,

with κ(K) := σ(K)σ(K −1 ). This is equivalent to treating the input multiplicative plant
uncertainty as the output multiplicative controller uncertainty. 3
2 See Stein and Doyle [1991].
3 Alternativecondition can be derived so that the condition related to nominal performance is scaled
by the condition number.
8.5. Skewed Specifications 153

The fact that the condition number appeared in the robust performance test for skewed
problems can be given another interpretation by considering two sets of plants Π1 and
Π2 , as shown in Figure 8.15 and below.
Π1 := {P (I + wt ∆) : ∆ ∈ RH∞ , k∆k∞ < 1}
Π2 := {(I + w̃t ∆)P : ∆ ∈ RH∞ , k∆k∞ < 1} .

- wt - ∆ - w̃t - ∆

-?
e - P - - P -?
e -

Figure 8.15: Converting input uncertainty to output uncertainty

Assume that P is invertible; then

Π2 ⊇ Π1 if |w̃t | ≥ |wt |κ(P ) ∀ω
since P (I + wt ∆) = (I + wt P ∆P −1 )P .
The condition number of a transfer matrix can be very high at high frequency, which
may significantly limit the achievable performance. The example below, taken from the
textbook by Franklin, Powell, and Workman [1990, page 788], shows that the condition
number shown in Figure 8.16 may increase with the frequency:
 
−0.2 0.1 1 0 1
 −0.05 0 0 0 0.7   
 
P (s) =  0 0 −1 1 0 = 1 s (s + 1)(s + 0.07)
  a(s) −0.05 0.7(s + 1)(s + 0.13)
 1 0 0 0 0 
0 1 0 0 0
where a(s) = (s + 1)(s + 0.1707)(s + 0.02929).
It is appropriate to point out that the skewed problem setup, although more com-
plicated than that of the nonskewed problem, is particularly suitable for control system
design. To be more specific, consider the transfer function from w and d˜ to z and e:
   
z w
= G(s) ˜
e d
where
 
−W2 Ti W1 −W2 KSo Wd
G(s) :=
We So P W1 We So Wd
    
−W2 0 K   W1 0
= (I + P K)−1 P I
0 We I 0 Wd
154 UNCERTAINTY AND ROBUSTNESS

3
10

2
10
condition number

1
10

0
10 -3 -2 -1 0 1 2
10 10 10 10 10 10
frequency

Figure 8.16: Condition number κ(ω) = σ̄(P (jω))/σ(P (jω))

Then a suitable performance criterion is to make kG(s)k∞ small. Indeed, small kG(s)k∞
implies that Ti , KSo , So P and So are small in some suitable frequency ranges, which
are the desired design specifications discussed in Section 6.1 of Chapter 6. It will be
clear in Chapter 16 and Chapter 17 that the kGk∞ is related to the robust stability
margin in the gap metric, ν-gap metric, and normalized coprime factor perturbations.
Therefore, making kGk∞ small is a suitable design approach.

8.6 Classical Control for MIMO Systems

In this section, we show through an example that the classical control theory may not
be reliable when it is applied to MIMO system design.
Consider a symmetric spinning body with torque inputs, T1 and T2 , along two or-
thogonal transverse axes, x and y, as shown in Figure 8.17. Assume that the angular
velocity of the spinning body with respect to the z axis is constant, Ω. Assume fur-
ther that the inertias of the spinning body with respect to the x, y, and z axes are I1 ,
I2 = I1 , and I3 , respectively. Denote by ω1 and ω2 the angular velocities of the body
with respect to the x and y axes, respectively. Then the Euler’s equation of the spinning
8.6. Classical Control for MIMO Systems 155

body is given by

I1 ω̇1 − ω2 Ω(I1 − I3 ) = T1
I1 ω̇2 − ω1 Ω(I3 − I1 ) = T2

Figure 8.17: Spinning body

Define    
u1 T1 /I1
:= , a := (1 − I3 /I1 )Ω.
u2 T2 /I1
Then the system dynamical equations can be written as
      
ω̇1 0 a ω1 u1
= +
ω̇2 −a 0 ω2 u2
Now suppose that the angular rates ω1 and ω2 are measured in scaled and rotated
coordinates:
       
y1 1 cos θ sin θ ω1 1 a ω1
= =
y2 cos θ − sin θ cos θ ω2 −a 1 ω2

where tan θ := a. (There is no specific physical meaning for the measurements of y1 and
y2 but they are assumed here only for the convenience of discussion.) Then the transfer
matrix for the spinning body can be computed as

Y (s) = P (s)U (s)

156 UNCERTAINTY AND ROBUSTNESS

with  
1 s − a2 a(s + 1)
P (s) = 2
s + a2 −a(s + 1) s − a2
Suppose the control law is chosen to be a unit feedback u = −y. Then the sensitivity
function and the complementary sensitivity function are given by
   
1 s −a 1 1 a
S = (I + P )−1 = , T = P (I + P )−1 =
s+1 a s s + 1 −a 1
1
Note that each single loop has the open-loop transfer function as , so each loop has
s
90o phase margin and ∞ gain margin.
Suppose one loop transfer function is perturbed, as shown in Figure 8.18.

w 6
z
δ 
y1 u1
 c?
  c
6
c −
P u2
 
y2 6
−

Figure 8.18: One-loop-at-a-time analysis

Denote
z(s) 1
= −T11 = −
w(s) s+1
Then the maximum allowable perturbation is given by
1
kδk∞ < = 1,
kT11 k∞
which is independent of a. Similarly the maximum allowable perturbation on the other
loop is also 1 by symmetry. However, if both loops are perturbed at the same time,
then the maximum allowable perturbation is much smaller, as shown next.
Consider a multivariable perturbation, as shown in Figure 8.19; that is, P∆ = (I +
∆)P , with  
δ11 δ12
∆= ∈ RH∞
δ21 δ22
a 2 × 2 transfer matrix such that k∆k∞ < γ. Then by the small gain theorem, the
system is robustly stable for every such ∆ iff
1 1
γ≤ =√ ( 1 if a  1).
kT k∞ 1 + a2
8.7. Notes and References 157

−
y1 ? r̃
 d d g11  d 1
6o δ11 
S 6
g12 
δ12 

δ21  g21 

y2 δ22 
 ?
d
/
 ?
d g22  dr̃2
6
−

Figure 8.19: Simultaneous perturbations

In particular, consider  
δ11
∆ = ∆d = ∈ R2×2 .
δ22
Then the closed-loop system is stable for every such ∆ iff
1

det(I + T ∆d ) = 2
s2 + (2 + δ11 + δ22 )s + 1 + δ11 + δ22 + (1 + a2 )δ11 δ22
(s + 1)
has no zero in the closed right-half plane. Hence the stability region is given by
2 + δ11 + δ22 > 0
1 + δ11 + δ22 + (1 + a2 )δ11 δ22 > 0.
It is easy to see that the system is unstable with
1
δ11 = −δ22 = √ .
1 + a2
The stability region for a = 5 is drawn in Figure 8.20, which shows how checking the
axis misses nearby regions of instability, and that for a >> 5, things just get that much
worse. The hyperbola portion of the picture gets arbitrarily close to (0,0). This clearly
shows that the analysis of a MIMO system using SISO methods can be misleading and
can even give erroneous results. Hence an MIMO method has to be used.

8.7 Notes and References

The small gain theorem was first presented by Zames [1966]. The book by Desoer and
Vidyasagar [1975] contains an extensive treatment and applications of this theorem in
158 UNCERTAINTY AND ROBUSTNESS

1.5

0.5

−0.5

−1

−1.5

−2

−2 −1.5 −1 −0.5 0 0.5 1 1.5

Figure 8.20: Stability region for a = 5

various forms. Robust stability conditions under various uncertainty assumptions are
discussed in Doyle, Wall, and Stein [1982].

8.8 Problems
Problem 8.1 This problem shows that the stability margin is critically dependent on
the type of perturbation. The setup is a unity-feedback loop with controller K(s) = 1
and plant Pnom (s) + ∆(s), where
10
Pnom (s) = .
s2 + 0.2s + 1
1. Assume ∆(s) ∈ RH∞ . Compute the largest β such that the feedback system is
internally stable for all k∆k∞ < β.
2. Repeat but with ∆ ∈ R.
Problem 8.2 Let M ∈ Cp×q be a given complex matrix. Then it is shown in Qiu et
al [1995] that I − ∆M is invertible for all ∆ ∈ Rq×p such that σ(∆) ≤ γ if and only if
µ∆ (M ) < 1/γ, where
 
ReM −α=M
µ∆ (M ) = inf σ2 .
α∈(0,1] α−1 =M ReM
8.8. Problems 159

It follows that (I − ∆M (s))−1 ∈ RH∞ for a given M (s) ∈ RH∞ and all ∆ ∈ Rq×p with
σ(∆) ≤ γ if and only if supω µ∆ (M (jω)) < 1/γ. Write a Matlab program to compute
µ∆ (M ) and apply it to the preceding problem.

Ke−τ s
Problem 8.3 Let G(s) = and K ∈ [10, 12], τ ∈ [0, 0.5], T = 1. Find a nominal
Ts + 1
model Go (s) ∈ RH∞ and a weighting function W (s) ∈ RH∞ such that

G(s) ∈ {Go (s) (1 + W (s)∆(s)) : ∆ ∈ H∞ , k∆k ≤ 1} .

Problem 8.4 Think of an example of a physical system with the property that the
number of unstable poles changes when the system undergoes a small change, for ex-
ample, when a mass is perturbed slightly or the geometry is deformed slightly.

Problem 8.5 Let X be a space of scalar-valued transfer functions. A function f (s) in

X is a unit if 1/f (s) is in X .
1. Prove that the set of units in RH∞ is an open set, that is, if f is a unit, then

(∃ > 0) (∀g ∈ RH∞ ) kgk∞ <  =⇒ f + g is a unit. (8.11)

2. Here is an application of the preceding fact. Consider the unity feedback system
with controller k(s) and plant p(s), both SISO, with k(s) proper and p(s) strictly
proper. Do coprime factorizations over RH∞ :
np nk
p= , k= .
mp mk

Then the feedback system is internally stable iff np nk + mp mk is a unit in RH∞ .

Assume it is a unit. Perturb p(s) to

np + ∆n
p= , ∆n , ∆m ∈ RH∞ .
mp + ∆m

Show that internal stability is preserved if k∆n k∞ and k∆m k∞ are small enough.
The conclusion is that internal stability is preserved if the perturbations are small
enough in the H∞ norm.
3. Give an example of a unit f (s) in RH∞ such that equation (8.11) fails for the H2
norm, that is, such that

(∀ > 0) (∃g ∈ RH2 ) kgk2 <  and f + g is not a unit.

What is the significance of this fact concerning robust stability?

Problem 8.6 Let ∆ and M be square constant matrices. Prove that the following
three conditions are equivalent:
160 UNCERTAINTY AND ROBUSTNESS

 
I −∆
1. is invertible;
−M I

2. I − M ∆ is invertible;

3. I − ∆M is invertible.

Problem 8.7 Consider the unity feedback system

- j - K - G -
−6

For  1 1 
 s s 
G(s) = 



1 1
s s
design a proper controller K(s) to stabilize the feedback system internally. Now perturb
G(s) to
 1+ 1 
 s s 
  ,  ∈ R.
 
1 1
s s
Is the feedback system internally stable for all sufficiently small ?

1
Problem 8.8 Consider the unity feedback system with K(s) = 3, G(s) = . Com-
s−2
pute by hand (i.e., without Matlab) a normalized coprime factorization of G(s). Con-
sidering perturbations ∆N and ∆M of the factors
 of G(s), compute
 by hand the stability
radius , that is, the least upper bound on ∆N ∆M ∞ such that feedback sta-
bility is preserved.

1
Problem 8.9 Let a unit feedback system with a controller K(s) = and a nominal
s
s+1
plant model Po (s) = 2 . Construct a smallest destabilizing ∆ ∈ RH∞ in the
s + 0.2s + 5
sense of k∆k∞ for each of the following cases:

(a) P = Po + ∆;

0.2(s + 10)
(b) P = Po (1 + W ∆) with W (s) = ;
s + 50
8.8. Problems 161

N + ∆n s+1 s2 + 0.2s + 5  
(c) P = ,N= 2
,M= 2
, and ∆ = ∆n ∆m .
M + ∆m (s + 2) (s + 2)

Problem 8.10 This problem concerns the unity feedback system with controller K(s)
and plant  
1 1 2
G(s) = .
s+1 3 4
1. Take K(s) = kI2 (k a real scalar) and find the range of k for internal stability.

2. Take  
k1 0
K(s) =
0 k2
(k1 , k2 real scalars) and find the region of (k1 , k2 ) in R2 for internal stability.

Problem 8.11 (Kharitonov’s Theorem) Let a(s) be an interval polynomial

a(s) = [a− − + − + 2
0 , a0 ] + [a1 , a1 ]s + [a2 , a2 ]s + · · · .
+

Kharitonov’s theorem shows that a(s) is stable if and only if the following four Kharitonov
polynomials are stable:

K1 (s) = a− − − 4 − 5
0 + a1 s + a2 s + a3 s + a4 s + a5 s + a6 s + · · ·
+ 2 + 3 + 6

K2 (s) = a− + + 2 − 3 − 4
0 + a1 s + a2 s + a3 s + a4 s + a5 s + a6 s + · · ·
+ 5 + 6

− 2 − 3 − 6
0 + a1 s + a2 s + a3 s + a4 s + a5 s + a6 s + · · ·
K3 (s) = a+ + + 4 + 5

− − 2 − 5 − 6
0 + a1 s + a2 s + a3 s + a4 s + a5 s + a6 s + · · ·
K4 (s) = a+ + 3 + 4

Let ai := (a− +
i + ai )/2 and let

anom (s) = a0 + a1 s + a2 s2 + · · · .

Find a least conservative W (s) such that

a(s)
∈ {1 + W (s)∆(s) | k∆k∞ ≤ 1} .
anom (s)

Problem 8.12 One of the main tools in this chapter was the small-gain theorem. One
way to state it is as follows: Define a transfer matrix F (s) in RH∞ to be contractive
if kF k∞ ≤ 1 and strictly contractive if kF k∞ < 1. Then for the unity feedback system
the small gain theorem is this: If K is contractive and G is strictly contractive, then
the feedback system is stable.
This problem concerns passivity and the passivity theorem. This is an important
tool in the study of the stability of feedback systems, especially robotics, that is com-
plementary to the small gain theorem.
162 UNCERTAINTY AND ROBUSTNESS

Consider a system with a square transfer matrix F (s) in RH∞ . This is said to be
passive if
F (jω) + F (jω)∗ ≥ 0, ∀ω.
Here, the symbol ≥ 0 means that the matrix is positive semidefinite. If the system is
SISO, the condition is equivalent to

Re F (jω) ≥ 0, ∀ω;

that is, the Nyquist plot of F lies in the right-half plane. The system is strictly passive
if F − I is passive for some  > 0.

1. Consider a mechanical system with input vector u(t) (forces and torques) and
output vector y(t) (velocities) modeled by the equation

M ẏ + Ky = u

where M and K are symmetric, positive definite matrices. Show that this system
is passive.

2. If F is passive, then (I + F )−1 ∈ RH∞ and (I + F )−1 (I − F ) is contractive; if

F is strictly passive, then (I + F )−1 (I − F ) is strictly contractive. Prove these
statements for the case that F is SISO.

3. Using the results so far, show (in the MIMO case) that the unity feedback system
is stable if K is passive and G is strictly passive.

Problem 8.13 Consider a SISO feedback system shown below with P = Po + W2 ∆2 .

d
?
W3
z
e - K - P -?
e -
− 6

Assume that P0 and P have the same number of right-half plane poles, W2 is stable,
and
|Re{∆2 }| ≤ α, |={∆2 }| ≤ β.
Derive the necessary and sufficient conditions for the feedback system to be robustly
stable.
8.8. Problems 163

 
P11 P12
Problem 8.14 Let P = ∈ RH∞ be a two-by-two transfer matrix. Find
P21 P22
sufficient (and necessary, if possible) conditions in each case so that Fu (P, ∆) is stable
for all possible stable ∆ that satisfies the following conditions, respectively:
1. at each frequency
Re∆(jω) ≥ 0, |∆(jω)| < α

2. at each frequency
Re∆(jω)e±jθ ≥ 0, |∆(jω)| < α
where θ ≥ 0.

3. at each frequency

Re∆(jω) ≥ 0, =∆(jω) ≥ 0, Re∆(jω) + =∆(jω) < α

Problem 8.15 Let P = (I + ∆W )P0 such that P and P0 have the same number of
unstable poles for all admissible ∆, k∆k∞ < γ. Show that K robustly stabilizes P if
and only if K stabilizes P0 and

W P0 K(I + P0 K)−1 ≤ 1.
∞

Problem 8.16 Give appropriate generalizations of the preceding problem to other

types of uncertainties.

Problem 8.17 Let K = I and

 1 2 
 s+1 s+3 
P0 =  .
1 1
s+1 s+1
1. Let P = P0 + ∆ with k∆k∞ ≤ γ. Determine the largest γ for robust stability.
 
k1
2. Let ∆ = ∈ R2×2 . Determine the stability region.
k2

Problem 8.18 Repeat the preceding problem with

 
s−1 5s + 1
 (s + 1)2 (s + 1)2 
P0 = 

.

−1 s−1
(s + 1)2 (s + 1)2
164 UNCERTAINTY AND ROBUSTNESS
Chapter 9

Linear Fractional
Transformation

This chapter introduces a new matrix function: linear fractional transformation (LFT).
We show that many interesting control problems can be formulated in an LFT frame-
work and thus can be treated using the same technique.

9.1 Linear Fractional Transformations

This section introduces the matrix linear fractional transformations. It is well known
from the one-complex-variable function theory that a mapping F : C 7→ C of the form
a + bs
F (s) =
c + ds
with a, b, c, and d ∈ C is called a linear fractional transformation. In particular, if c 6= 0
then F (s) can also be written as

F (s) = α + βs(1 − γs)−1

for some α, β and γ ∈ C. The linear fractional transformation described above for
scalars can be generalized to the matrix case.

Definition 9.1 Let M be a complex matrix partitioned as

 
M11 M12
M= ∈ C(p1 +p2 )×(q1 +q2 ) ,
M21 M22

and let ∆` ∈ Cq2 ×p2 and ∆u ∈ Cq1 ×p1 be two other complex matrices. Then we can
formally define a lower LFT with respect to ∆` as the map

F` (M, •) : Cq2 ×p2 7→ Cp1 ×q1

165
166 LINEAR FRACTIONAL TRANSFORMATION

with
F` (M, ∆` ) := M11 + M12 ∆` (I − M22 ∆` )−1 M21
provided that the inverse (I − M22 ∆` )−1 exists. We can also define an upper LFT with
respect to ∆u as
Fu (M, •) : Cq1 ×p1 7→ Cp2 ×q2
with
Fu (M, ∆u ) = M22 + M21 ∆u (I − M11 ∆u )−1 M12
provided that the inverse (I − M11 ∆u )−1 exists.

The matrix M in the preceding LFTs is called the coefficient matrix. The motivation for
the terminologies of lower and upper LFTs should be clear from the following diagram
representations of F` (M, ∆` ) and Fu (M, ∆u ):

z1 w1
  - ∆u
M  y2 u2
y1 u1 z2  w2
- ∆`  M 

The diagram on the left represents the following set of equations:

      
z1 w1 M11 M12 w1
= M = ,
y1 u1 M21 M22 u1
u1 = ∆` y1

while the diagram on the right represents

      
y2 u2 M11 M12 u2
= M = ,
z2 w2 M21 M22 w2
u2 = ∆u y2 .

It is easy to verify that the mapping defined on the left diagram is equal to F` (M, ∆` )
and the mapping defined on the right diagram is equal to Fu (M, ∆u ). So from the above
diagrams, F` (M, ∆` ) is a transformation obtained from closing the lower loop on the left
diagram; similarly, Fu (M, ∆u ) is a transformation obtained from closing the upper loop
on the right diagram. In most cases, we shall use the general term LFT in referring to
both upper and lower LFTs and assume that the context will distinguish the situations
since one can use either of these notations
 to express a given object. Indeed, it is clear
M22 M21
that Fu (N, ∆) = F` (M, ∆) with N = . It is usually not crucial which
M12 M11
expression is used; however, it is often the case that one expression is more convenient
than the other for a given problem. It should also be clear to the reader that in writing
F` (M, ∆) [or Fu (M, ∆)] it is implied that ∆ has compatible dimensions.
9.1. Linear Fractional Transformations 167

A useful interpretation of an LFT [e.g., F` (M, ∆)] is that F` (M, ∆) has a nominal
mapping, M11 , and is perturbed by ∆, while M12 , M21 , and M22 reflect a prior knowl-
edge as to how the perturbation affects the nominal map, M11 . A similar interpretation
can be applied to Fu (M, ∆). This is why LFT is particularly useful in the study of
perturbations, which is the focus of the next chapter.
The physical meaning of an LFT in control science is obvious if we take M as a proper
transfer matrix. In that case, the LFTs defined previously are simply the closed-loop
transfer matrices from w1 7→ z1 and w2 7→ z2 , respectively; that is,

Tzw1 = F` (M, ∆` ), Tzw2 = Fu (M, ∆u )

where M may be the controlled plant and ∆ may be either the system model uncer-
tainties or the controllers.

Definition 9.2 An LFT, F` (M, ∆), is said to be well-defined (or well-posed) if

(I − M22 ∆) is invertible.

Note that this definition is consistent with the well-posedness definition of the feed-
back system, which requires that the corresponding transfer matrix be invertible in
Rp (s). It is clear that the study of an LFT that is not well-defined is meaningless;
hence throughout this book, whenever an LFT is invoked, it will be assumed implicitly
that it is well-defined. It is also clear from the definition that, for any M , F` (M, 0)
is well-defined; hence any function that is not well-defined at the origin cannot be ex-
pressed as an LFT in terms of its variables. For example, f (δ) = 1/δ is not an LFT of
δ.
In some literature, LFT is used to refer to the following matrix functions:

(A + BQ)(C + DQ)−1 or (C + QD)−1 (A + QB)

where C is usually assumed to be invertible due to practical consideration. The following

results follow from some simple algebra.

Lemma 9.1 Suppose C is invertible. Then

(A + BQ)(C + DQ)−1 = F` (M, Q)

(C + QD)−1 (A + QB) = F` (N, Q)

with    
AC −1 B − AC −1 D C −1 A C −1
M= , N= .
C −1 −C −1 D B − DC A −DC −1
−1

The converse also holds if M satisfies certain conditions.

168 LINEAR FRACTIONAL TRANSFORMATION

 
M11 M12
Lemma 9.2 Let F` (M, Q) be a given LFT with M = .
M21 M22

(a) If M12 is invertible, then

F` (M, Q) = (C + QD)−1 (A + QB)

−1 −1 −1 −1
with A = M12 M11 , B = M21 − M22 M12 M11 , C = M12 , and D = −M22 M12 ;
that is,
  
  0 0 −I
A C
= F`  M21 0 M22  , −M12−1 
B D
M11 I 0
  
0 0 −I
= F`  M21 0 M22  , E −1 
M11 I M12 + E

for any nonsingular matrix E.

(b) If M21 is invertible, then

F` (M, Q) = (A + BQ)(C + DQ)−1

−1 −1 −1 −1
with A = M11 M21 , B = M12 − M11 M21 M22 , C = M21 , and D = −M21 M22 ;
that is,
  
  0 M12 M11
A B
= F`  0 0 I  , −M21 −1 
C D
−I M22 0
  
0 M12 M11
= F`  0 0 I  , E −1 
−I M22 M21 + E

for any nonsingular matrix E.

However, for an arbitrary LFT F` (M, Q), neither M21 nor M12 is necessarily square
and invertible; therefore, the alternative fractional formula is more restrictive.
It should be pointed out that some seemingly simple functions do not have simple
LFT representations. For example,

(A + QB)(I + QD)−1

cannot always be written in the form of F` (M, Q) for some M ; however, it can be
written as
(A + QB)(I + QD)−1 = F` (N, ∆)
9.1. Linear Fractional Transformations 169

with  
A I A  
Q
N =  −B 0 −B  , ∆= .
Q
D 0 D
Note that the dimension of ∆ is twice of Q.
The following lemma shows that the inverse of an LF T is still an LFT.
 
M11 M12
Lemma 9.3 Let M = and M22 is nonsingular. Then
M21 M22

(Fu (M, ∆))−1 = Fu (N, ∆)

with N , is given by
 −1 −1

M11 − M12 M22 M21 −M12 M22
N= −1 −1 .
M22 M21 M22
LFT is a very convenient tool to formulate many mathematical objects. We shall
illustrate this by the following two examples.

Simple Block Diagrams

A feedback system with the following block diagram

 W1 
uf
d
?
i y - K - i - P - W2 v-
−6 u

?
F  i n

can be rearranged as an LFT:

z w
G 
y u
- K

with  
    W2 P 0 W2 P
d v
w= , z= , G= 0 0 W1  .
n uf
−F P −F −F P
170 LINEAR FRACTIONAL TRANSFORMATION

A state-space realization for the generalized plant G can be obtained by directly

realizing the transfer matrix G using any standard multivariable realization techniques
(e.g., Gilbert realization). However, the direct realization approach is usually com-
plicated. Here we shall show another way to obtain the realization for G based on
the realizations of each component. To simplify the expression, we shall assume that
the plant P is strictly proper and P , F , W1 , and W2 have, respectively, the following
state-space realizations:
       
Ap Bp Af Bf Au Bu Av Bv
P = , F = , W1 = , W2 = .
Cp 0 Cf Df Cu Du Cv Dv

That is,
ẋp = Ap xp + Bp (d + u), yp = Cp xp ,
ẋf = Af xf + Bf (yp + n), −y = Cf xf + Df (yp + n),
ẋu = Au xu + Bu u, uf = Cu xu + Du u,
ẋv = Av xv + Bv yp , v = Cv xv + Dv yp .
Now define a new state vector  
xp
 xf 
x= 
 xu 
xv
and eliminate the variable yp to get a realization of G as

ẋ = Ax + B1 w + B2 u
z = C1 x + D11 w + D12 u
y = C2 x + D21 w + D22 u

with
     
Ap 0 0 0 Bp 0 Bp
 Bf Cp Af 0 0     
A=  , B1 =  0 Bf  , B2 =  0 
0 0 Au 0   0 0   Bu 
Bv Cp 0 0 Av 0 0 0
   
Dv Cp 0 0 Cv 0
C1 = , D11 = 0, D12 =
0 0 Cu 0 Du
   
C2 = −Df Cp −Cf 0 0 , D21 = 0 −Df , D22 = 0.
9.1. Linear Fractional Transformations 171

Parametric Uncertainty: A Mass/Spring/Damper System

One natural type of uncertainty is unknown coefficients in a state-space model. To

motivate this type of uncertainty description, we shall begin with a familiar mechanical
system, shown in Figure 9.1.

6
F

X
X

k 
XXX c

X
X

Figure 9.1: A mass/spring/damper system

The dynamical equation of the system motion can be described by

c k F
ẍ + ẋ + x = .
m m m

Suppose that the three physical parameters m, c, and k are not known exactly, but are
believed to lie in known intervals. In particular, the actual mass m is within 10% of
a nominal mass, m̄, the actual damping value c is within 20% of a nominal value of
c̄, and the spring stiffness is within 30% of its nominal value of k̄. Now introducing
perturbations δm , δc , and δk , which are assumed to be unknown but lie in the interval
[−1, 1], the block diagram for the dynamical system is as shown in Figure 9.2.
1
It is easy to check that m can be represented as an LFT in δm :

1 1 1 0.1
= = − δm (1 + 0.1δm )−1 = F` (M1 , δm )
m m̄(1 + 0.1δm ) m̄ m̄

" #
1
m̄ − 0.1
m̄
with M1 = . Suppose that the input signals of the dynamical system
1 −0.1
are selected as x1 = x, x2 = ẋ, F , and the output signals are selected as ẋ1 and ẋ2 . To
represent the system model as an LFT of the natural uncertainty parameters δm , δc , and
δk , we shall first isolate the uncertainty parameters and denote the inputs and outputs
of δk , δc , and δm as yk , yc , ym and uk , uc , um , respectively, as shown in Figure 9.3.
172 LINEAR FRACTIONAL TRANSFORMATION

x F
1  ẋ 1 ẍ 1  d
s s m̄(1+0.1δm )
6−

- c̄(1 + 0.2δc ) -d
6+

- k̄(1 + 0.3δk )

Figure 9.2: Block diagram of mass/spring/damper equation

x1 x2
1  1   e F
s s M1  6−
ym um
- δm

- c̄ -e -e
6 6
- 0.2 - δc
yc uc

- k̄ -e
6
- 0.3 - δ
yk k uk

Figure 9.3: A block diagram for the mass/spring/damper system with uncertain pa-
rameters
9.2. Basic Principle 173

Then
 
    x1
ẋ1 0 1 0 0 0 0  
    x2     
 ẋ2   − m̄ k̄
− m̄
c̄ 1
− m̄
1
− m̄
1
− 0.1   uk yk
   m̄ m̄  
 yk  =  0.3k̄  F     
   0 0 0 0 0   ,  u c  = ∆  yc  .
    uk 
 yc   0 0.2c̄ 0 0 0 0 


 um ym
 uc 
ym −k̄ −c̄ 1 −1 −1 −0.1
um

That is,


" # x1
ẋ1  
= F` (M, ∆)  x2 
ẋ2
F
where
 
0 1 0 0 0 0
   
 − m̄
k̄
− m̄
c̄ 1
− m̄
1
− m̄
1
− 0.1  δk 0 0
 m̄ m̄ 
M = 0.3k̄ 0 0 0 0 0 , ∆ = 
 0 δc

0 .
 
 
 0 0.2c̄ 0 0 0 0  0 0 δm
−k̄ −c̄ 1 −1 −1 −0.1

9.2 Basic Principle

We have studied two simple examples of the use of LFTs and, in particular, their role in
modeling uncertainty. The basic principle at work here in writing a matrix LFT is often
referred to as “pulling out the ∆’s”. We will try to illustrate this with another picture.
Consider a structure with four substructures interconnected in some known way, as
shown in Figure 9.4. This diagram can be redrawn as a standard one via “pulling out
the ∆’s” in Figure 9.5.
Now the matrix M of the LFT can be obtained by computing the corresponding
transfer matrix in the shadowed box.
We shall illustrate the preceding principle with an example. Consider an input/output
relation
a + bδ2 + cδ1 δ22
z= w =: Gw
1 + dδ1 δ2 + eδ12
where a, b, c, d, and e are given constants or transfer functions. We would like to write
G as an LFT in terms of δ1 and δ2 . We shall do this in three steps:

1. Draw a block diagram for the input/output relation with each δ separated as
shown in Figure 9.6.
174 LINEAR FRACTIONAL TRANSFORMATION

.. .. . . . . . .. .. .. .. .. .. .. .. .. . . .. .. . .
..... . . . . . ............................. . . ....... .. ..
. .... ∆1  ......................... .. ... .... .. .. .. .
.
.. . . .
.. ............... ... . .. .
.
. . . . ..  . . ............... ........ ∆2  ... .. .
. . . . ...... .... . . ............... ...... .... . . . ..
. .... .. ... . . ............... ... .. ....... . . . ......
 .. ... .. .. . . ....... ... ............ .... ... ... .. ........
...... . . . . . . . . . . .. . .
.. . . .  ................... .... . . . . ...... ....
∆3 .. .. .. .. .. .. .. . . . ..
....  . . . . .. . . . . .. .. K  ......
. .... . . .. . ... .................. ..  . .. ..
.. ......... ........................ ....... . . . . .......
.. .. .. .. .. .. .. .. .. .. .. .. .. . . . . . .. ..

Figure 9.4: Multiple source of uncertain structure

- ∆1
- ∆2
- ∆3

.. .. . . . . . .. .. .. .. .. .. .. .. .. . . .. .. . .
 ..... . . . . . ............................. . . ....... .. ..
 . ....
.  ......................... .. ... .... .. .. .. .
 .. . . . . . . . .
.. ............... ... . . ...
. . . . .  . . ............... ........ ? .
. . . . ...... .... . . ............... ...... .... .... . ..
. . . . . . . . . . . . .
. .. . .. . .. .. .. .. .. .. . .. .. . . . . .. ... .
 .. ... .. .. . . ....... ... ............ .... ... ... .. ........
...... .. .. ... .. .. .. .. .. .. .. ........ ....
.. . . . ........ ......... ..... .. .
.
....  . .. ............ .. ..
. . .  .......
 . .... . . .. . ... .................. ..  .......
.. ......... ........................ ....... . . . . .......
.. .. .. .. .. .. .. .. .. .. .. .. .. . . . . . .. ..

- K

Figure 9.5: Pulling out the ∆’s

9.2. Basic Principle 175

a 

z ? u y4
 e 4 δ2  e b 
6

c
6
u3 y3 u1 y1 w
δ2  δ1  e
6
?
y2 u2
−d - δ1 - −e -e
6

Figure 9.6: Block diagram for G

2. Mark the inputs and outputs of the δ’s as y’s and u’s, respectively. (This is
essentially pulling out the ∆’s.)

3. Write z and y’s in terms of w and u’s with all δ’s taken out. (This step is equivalent
to computing the transformation in the shadowed box in Figure 9.5.)
   
y1 u1
   
 y2   u2 
   
 y3 =M u3 
   
   
 y4   u4 
z w

where  
0 −e −d 0 1
 
 1 0 0 0 0 
 
M =
 1 0 0 0 0 .

 
 0 −be −bd + c 0 b 
0 −ae −ad 1 a
Then " #
δ1 I2 0
z = Fu (M, ∆)w, ∆ = .
0 δ2 I2

All LFT examples in Section 9.1 can be obtained following the preceding steps.
176 LINEAR FRACTIONAL TRANSFORMATION

For Simulink users, it is much easier to do all the computations using Simulink
block diagrams, as shown in the following example.

Example 9.1 Consider the HIMAT (highly maneuverable aircraft) control problem
from the µ Analysis and Synthesis Toolbox (Balas et al. [1994]). The system diagram
is shown in Figure 9.7 where
   
50(s + 100) 0.5(s + 3)
 0   0 
Wdel =  s + 10000 , Wp =  s + 0.03
50(s + 100)  0.5(s + 3) 
,
0 0
s + 10000 s + 0.03
 
2(s + 1.28)
 0 
Wn =  s + 320 2(s + 1.28)  ,
0
s + 320
 
−0.0226 −36.6 −18.9 −32.1 0 0
 
 0 −1.9 0.983 0 −0.414 0 
 
 0.0123 −11.7 −2.63 −77.8 22.4 
 0 
P0 =  
 0 0 1 0 0 0 
 
 0 
 0 57.3 0 0 0 
0 0 0 57.3 0 0

" #
z1
" #
z2 d1
" #
- ∆ p1 d2
p2
" #
Wdel e1
6 -?
f - P0 ?
-f - Wp - e2

" #
n1
u y ?
K  f Wn  n2

Figure 9.7: HIMAT closed-loop interconnection

9.2. Basic Principle 177

The open-loop interconnection is

 
  p1
 
z1  p2 
   
   
 z2   d1 
   
 e1   d2 
  = Ĝ(s) 

.

 e2   n1 
   
   
 y1   n2 
 
y2  u1 
u2

The Simulink block diagram of this open-loop interconnection is shown in Figure 9.8.

Figure 9.8: Simulink block diagram for HIMAT (aircraft.m)

" #
A B
The Ĝ(s) = can be computed by
C D

 [A, B, C, D] = linmod(0 aircraft0 )

178 LINEAR FRACTIONAL TRANSFORMATION

which gives
 
−10000I2 0 0 0 0 0 0 0
 −0.0226 −36.6 −18.9 −32.1 
 0 0 0 0 
 
 0 0 −1.9 0.983 0 0 0 0 
 
 0 0.0123 −11.7 −2.63 0 0 0 0 
A=



 0 0 0 1 0 0 0 0 
 
 0 0 −54.087 0 0 −0.018 0 0 
 
 0 0 0 0 −54.087 0 −0.018 0 
0 0 0 0 0 0 0 −320I2
 
0 0 0 0 −703.5624 0
 −703.5624 
 0 0 0 0 0 
 
 0 0 0 0 0 0 
 
 −0.4140 0 0 0 −0.4140 0 
B=
 −77.8


 22.4 0 0 −77.8 22.4 
 
 0 0 0 0 0 0 
 
 0 0 −0.9439I2 0 0 0 
0 0 0 −25.2476I2 0 0
 
703.5624I2 0 0 0 0 0 0 0 0
 −0.9439 
 0 0 28.65 0 0 0 0 0 
 
C= 0 0 0 0 28.65 0 −0.9439 0 0 
 
 0 0 57.3 0 0 0 0 25.2476 0 
0 0 0 0 57.3 0 0 0 25.2476
 
0 0 0 0 0 0 50 0
 50 
 0 0 0 0 0 0 0 
 
 0 0 0.5 0 0 0 0 0 
D= .
 0 0 0 0.5 0 0 0 0 
 
 0 0 1 0 2 0 0 0 
0 0 0 1 0 2 0 0

9.3 Redheffer Star Products

The most important property of LFTs is that any interconnection of LFTs is again an
LFT. This property is by far the most often used and is the heart of LFT machinery.
Indeed, it is not hard to see that most of the interconnection structures discussed earlier
(e.g., feedback and cascade) can be viewed as special cases of the so-called star product.
9.3. Redheffer Star Products 179

Suppose that P and K are compatibly partitioned matrices

" # " #
P11 P12 K11 K12
P = , K=
P21 P22 K21 K22
such that the matrix product P22 K11 is well-defined and square, and assume further
that I − P22 K11 is invertible. Then the star product of P and K with respect to this
partition is defined as
" #
−1
Fl (P, K11 ) P12 (I − K11 P22 ) K12
P ? K := −1 . (9.1)
K21 (I − P22 K11 ) P21 Fu (K, P22 )
Note that this definition is dependent on the partitioning of the matrices P and K. In
fact, this star product may be well-defined for one partition and not well-defined for
another; however, we will not explicitly show this dependence because it is always clear
from the context. In a block diagram, this dependence appears, as shown in Figure 9.9.

z
  w z
 w
P  ẑ

P ?K
 ŵ
ya
aa!!!
u
! a
!! aa

ẑ K ŵ
 

Figure 9.9: Interconnection of LFTs

Now suppose that P and K are transfer matrices with state-space representations:
   
A B1 B2 AK BK1 BK2
   
P =  C1 D11 D12  K =  CK1 DK11 DK12  .
C2 D21 D22 CK2 DK21 DK22
Then the transfer matrix " # " #
w z
P ?K : 7→
ŵ ẑ
has a representation
 
Ā B̄1 B̄2 " #
  Ā B̄
P ?K =
 C̄1 D̄11 D̄12 =

C̄ D̄
C̄2 D̄21 D̄22
180 LINEAR FRACTIONAL TRANSFORMATION

where
" #
A + B2 R̃−1 DK11 C2 B2 R̃−1 CK1
Ā =
BK1 R−1 C2 AK + BK1 R−1 D22 CK1
" #
B1 + B2 R̃−1 DK11 D21 B2 R̃−1 DK12
B̄ =
BK1 R−1 D21 BK2 + BK1 R−1 D22 DK12
" #
C1 + D12 DK11 R−1 C2 D12 R̃−1 CK1
C̄ =
DK21 R−1 C2 CK2 + DK21 R−1 D22 CK1
" #
D11 + D12 DK11 R−1 D21 D12 R̃−1 DK12
D̄ =
DK21 R−1 D21 DK22 + DK21 R−1 D22 DK12

R = I − D22 DK11 , R̃ = I − DK11 D22 .

In fact, it is easy to show that

" # " #
A B2 DK11 CK1
Ā = ? ,
C2 D22 BK1 AK
" # " #
B1 B2 DK11 DK12
B̄ = ? ,
D21 D22 BK1 BK2
" # " #
C1 D12 DK11 CK1
C̄ = ? ,
C2 D22 DK21 CK2
" # " #
D11 D12 DK11 DK12
D̄ = ? .
D21 D22 DK21 DK22

The Matlab command starp can be used to compute the star product:

 P ? K = starp(P, K, dimy, dimu)

where dimy and dimu are the dimensions of y and u, respectively. In the particular case
when dim(ẑ) = 0 and dim(ŵ) = 0, we have

 F` (P, K) = starp(P, K)

9.4 Notes and References

This chapter is based on the lecture notes by Packard [1991] and the paper by Doyle,
Packard, and Zhou [1991].
9.5. Problems 181

9.5 Problems
˜ = M ∆N , where ∆ is block diago-
Problem 9.1 Find M and N matrices such that ∆
nal.
h i
˜ = ∆1 ∆2
1. ∆
 
˜ = ∆1 0 0
2. ∆
0 0 ∆2
 h i
∆1 0 0
" # 
˜ =
3. ∆ 
 ∆2 
∆3
0
 
∆1 0 0
˜ =  ∆2
4. ∆ ∆3 0 
0 0 ∆4
 
∆1 ∆3 0
˜ =  ∆2
5. ∆ ∆4 ∆6 
0 ∆5 0

−1
Problem 9.2 Let G = (I − P (s)∆) . Find a matrix M (s) such that G = Fu (M, ∆).

Problem 9.3 Consider the unity feedback system with G(s) of size 2 × 2. Suppose
G(s) has an uncertainty model of the form
" #
[1 + ∆11 (s)]g11 (s) [1 + ∆12 (s)]g12 (s)
G(s) = .
[1 + ∆21 (s)]g21 (s) [1 + ∆22 (s)]g22 (s)

Suppose also that we wish to study robust stability of the feedback system. Pull out the
∆’s and draw the appropriate block diagram in the form of a structured perturbation
of a nominal system.

Problem 9.4 Let

 
A B1 B2 " #
  Â B̂
P (s) =  C1 D11 D12 , K(s) = .
Ĉ D̂
C2 D21 D22

Find state-space realizations for F` (P, K) and F` (P, D̂).

182 LINEAR FRACTIONAL TRANSFORMATION

Problem 9.5 Suppose D21 is nonsingular and

 
A B1 B2
 
M (s) =  C1 D11 D12  .
C2 D21 D22

Find a state-space realization for

 " # " # 
0 M12 M11
 
M̂(s) = 
 h 0 0 i I 

−I M22 M21 + E

where E is a constant matrix. Find the state-space realization for F` (M̂ , E −1 ) when
E = I.
Chapter 10

µ and µ Synthesis

It is noted that the robust stability and robust performance criteria derived in Chap-
ter 8 vary with the assumptions about the uncertainty descriptions and performance
requirements. We shall show in this chapter that they can all be treated in a unified
framework using the LFT machinery introduced in the last chapter and the structured
singular value to be introduced in this chapter. This, of course, does not mean that
those special problems and their corresponding results are not important; on the con-
trary, they are sometimes very enlightening to our understanding of complex problems,
such as those in which complex problems are formed from simple problems. On the
other hand, a unified approach may relieve the mathematical burden of dealing with
specific problems repeatedly. Furthermore, the unified framework introduced here will
enable us to treat exactly the robust stability and robust performance problems for
systems with multiple sources of uncertainties, which is a formidable problem from the
standpoint of Chapter 8, in the same fashion as single unstructured uncertainty. In-
deed, if a system is subject to multiple sources of uncertainties, in order to use the
results in Chapter 8 for unstructured cases, it is necessary to reflect all sources of un-
certainties from their known point of occurrence to a single reference location in the
loop. Such reflected uncertainties invariably have a great deal of structure, which must
then be “covered up” with a large, arbitrarily more conservative perturbation in order
to maintain a simple cone-bounded representation at the reference location. Readers
might have some idea about the conservativeness in such reflection based on the skewed
specification problem, where an input multiplicative uncertainty of the plant is reflected
at the output and the size of the reflected uncertainty is proportional to the condition
number of the plant. In general, the reflected uncertainty may be proportional to the
condition number of the transfer matrix between its original location and the reflected
location. Thus it is highly desirable to treat the uncertainties as they are and where
they are. The structured singular value is defined exactly for that purpose.

183
184 µ AND µ SYNTHESIS

10.1 General Framework for System Robustness

As we illustrated in Chapter 9, any interconnected system may be rearranged to fit the
general framework in Figure 10.1. Although the interconnection structure can become
quite complicated for complex systems, many software packages, such as Simulink and
µ Analysis and Synthesis Toolbox, are available that could be used to generate the
interconnection structure from system components. Various modeling assumptions will
be considered, and the impact of these assumptions on analysis and synthesis methods
will be explored in this general framework.

- ∆


z  w
P


- K

Figure 10.1: General framework

Note that uncertainty may be modeled in two ways, either as external inputs or
as perturbations to the nominal model. The performance of a system is measured in
terms of the behavior of the outputs or errors. The assumptions that characterize
the uncertainty, performance, and nominal models determine the analysis techniques
that must be used. The models are assumed to be FDLTI systems. The uncertain
inputs are assumed to be either filtered white noise or weighted power or weighted
Lp signals. Performance is measured as weighted output variances, or as power, or as
weighted output Lp norms. The perturbations are assumed to be themselves FDLTI
systems that are norm-bounded as input-output operators. Various combinations of
these assumptions form the basis for all the standard linear system analysis tools.
Given that the nominal model is an FDLTI system, the interconnection system has
the form  
P11 (s) P12 (s) P13 (s)
 
P (s) =  P21 (s) P22 (s) P23 (s) 
P31 (s) P32 (s) P33 (s)
and the closed-loop system is an LFT on the perturbation and the controller given by
z = Fu (F` (P, K), ∆) w
= F` (Fu (P, ∆), K) w.
We shall focus our discussion in this section on analysis methods; therefore, the
controller may be viewed as just another system component and absorbed into the
10.1. General Framework for System Robustness 185

interconnection structure. Denote

" #
M11 (s) M12 (s)
M (s) = F` (P (s), K(s)) = .
M21 (s) M22 (s)

Then the general framework reduces to Figure 10.2, where

 
z = Fu (M, ∆)w = M22 + M21 ∆(I − M11 ∆)−1 M12 w.

- ∆

z  w
 M 

Figure 10.2: Analysis framework

Suppose K(s) is a stabilizing controller for the nominal plant P . Then M (s) ∈ RH∞ .
In general, the stability of Fu (M, ∆) does not necessarily imply the internal stability of
the closed-loop feedback system. However, they can be made equivalent with suitably
chosen w and z. For example, consider again the multiplicatively perturbed system
shown in Figure 10.3.
e3 d3
d2 - W2 - ∆ - W1
d1 e1 e2
?-
-e - K -e -?
e
P
−6

Figure 10.3: Multiplicatively perturbed systems

Now let " # " #

d1 e1
w := , z := .
d2 e2
Then the system is robustly stable for all ∆(s) ∈ RH∞ with k∆k∞ < 1 if and only if
Fu (M, ∆) ∈ RH∞ for all admissible ∆ with M11 = −W2 P K(I + P K)−1 W1 , which is
guaranteed by kM11 k∞ ≤ 1.
The analysis results presented in the previous chapters together with the associ-
ated synthesis tools are summarized in Table 10.1 with various uncertainty modeling
assumptions.
However, the analysis is not so simple for systems with multiple sources of model
uncertainties, including the robust performance problem for systems with unstructured
186 µ AND µ SYNTHESIS

Input Performance Perturbation Analysis Synthesis

Assumptions Specifications Assumptions Tests Methods

E(w(t)w(τ )∗ ) E(z(t)∗ z(t)) ≤ 1 LQG

= δ(t − τ )I

∆=0 kM22 k2 ≤ 1 Wiener-Hopf

w = U0 δ(t) 2
E(kzk2 ) ≤1
E(U0 U0∗ ) = I H2

kwk2 ≤ 1 kzk2 ≤ 1 ∆=0 kM22 k∞ ≤ 1 Singular Value

Loop Shaping

kwk2 ≤ 1 Internal Stability k∆k∞ < 1 kM11 k∞ ≤ 1 H∞

Table 10.1: General analysis for single source of uncertainty

uncertainty. As we shown in Chapter 9, if a system is built from components that are

themselves uncertain, then, in general, the uncertainty in the system level is structured,
involving typically a large number of real parameters. The stability analysis involving
real parameters is much more difficult and will be discussed in Chapter 18. Here we shall
simply cover the real parametric uncertainty with norm-bounded dynamical uncertainty.
Moreover, the interconnection model M can always be chosen so that ∆(s) is block
diagonal, and, by absorbing any weights, k∆k∞ < 1. Thus we shall assume that ∆(s)
takes the form of

∆(s) = {diag [δ1 Ir1 , . . . , δs IrS , ∆1 , . . . , ∆F ] : δi (s) ∈ RH∞ , ∆j ∈ RH∞ }

with kδi k∞ < 1 and k∆j k∞ < 1. Then the system is robustly stable iff the intercon-
nected system in Figure 10.4 is stable.
The results of Table 10.1 can be applied to analysis of the system’s robust stability
in two ways:

(1) kM11 k∞ ≤ 1 implies stability, but not conversely, because this test ignores the
known block diagonal structure of the uncertainties and is equivalent to regarding
∆ as unstructured. This can be arbitrarily conservative in that stable systems can
have arbitrarily large kM11 k∞ .
10.2. Structured Singular Value 187

- ∆F (s)

..
.

- δ1 (s)I


.. M11 (s) ..
. .


Figure 10.4: Robust stability analysis framework

(2) Test for each δi (∆j ) individually (assuming no uncertainty in other channels).
This test can be arbitrarily optimistic because it ignores interaction between the
δi (∆j ). This optimism is also clearly shown in the spinning body example in
Section 8.6.
The difference between the stability margins (or bounds on ∆) obtained in (1) and (2)
can be arbitrarily far apart. Only when the margins are close can conclusions be made
about the general case with structured uncertainty.
The exact stability and performance analysis for systems with structured uncertainty
requires a new matrix function called the structured singular value (SSV), which is
denoted by µ.

10.2 Structured Singular Value

10.2.1 Definitions of µ
We shall motivate the definition of the structured singular value by asking the following
question: Given a matrix M ∈ Cp×q , what is the smallest perturbation matrix ∆ ∈ Cq×p
in the sense of σ(∆) such that

det(I − M ∆) = 0?

That is, we are interested in finding


αmin := inf σ(∆) : det(I − M ∆) = 0, ∆ ∈ Cq×p .
188 µ AND µ SYNTHESIS

It is easy to see that

 1
αmin = inf α : det(I − αM ∆) = 0, σ(∆) ≤ 1, ∆ ∈ Cq×p =
max ρ(M ∆)
σ(∆)≤1

and
max ρ(M ∆) = σ(M ).
σ(∆)≤1

Hence the smallest norm of a “destabilizing” perturbation matrix is 1/σ(M ) with a

smallest “destabilizing” ∆:
1
∆des = v1 u∗1 , det(I − M ∆des ) = 0
σ(M )

where M = σ(M )u1 v1∗ + σ2 u2 v2∗ + · · · is a singular value decomposition.

So the reciprocal of the largest singular value of a matrix is a measure of the smallest
“destabilizing” perturbation matrix. Hence it is instructive to introduce the following
alternative definition for the largest singular value:
1
σ(M ) := .
inf{σ(∆) : det(I − M ∆) = 0, ∆ ∈ Cq×p }
Next we consider a similar problem but with ∆ structurally restricted. In particular,
we consider the block diagonal matrix ∆. We shall consider two types of blocks: repeated
scalar and full blocks. Let S and F represent the number of repeated scalar blocks and
the number of full blocks, respectively. To bookkeep their dimensions, we introduce
positive integers r1 , . . . , rS ; m1 , . . . , mF . The ith repeated scalar block is ri × ri , while
the jth full block is mj × mj . With those integers given, we define ∆ ⊂ Cn×n as

∆ = diag [δ1 Ir1 , . . . , δs IrS , ∆1 , . . . , ∆F ] : δi ∈ C, ∆j ∈ Cmj ×mj . (10.1)

For consistency among all the dimensions, we must have

X
S X
F
ri + mj = n.
i=1 j=1

Often, we will need norm-bounded subsets of ∆, and we introduce the following nota-
tion:
B∆ = {∆ ∈ ∆ : σ(∆) ≤ 1} (10.2)
Bo ∆ = {∆ ∈ ∆ : σ(∆) < 1} (10.3)
where the superscript “o” symbolizes the open ball. To keep the notation as simple as
possible in equation (10.1), we place all of the repeated scalar blocks first; in actuality,
they can come in any order. Also, the full blocks do not have to be square, but restricting
them as such saves a great deal in terms of notation.
10.2. Structured Singular Value 189

Now we ask a similar question: Given a matrix M ∈ Cp×q , what is the smallest
perturbation matrix ∆ ∈ ∆ in the sense of σ(∆) such that

det(I − M ∆) = 0?

That is, we are interested in finding

αmin := inf {σ(∆) : det(I − M ∆) = 0, ∆ ∈ ∆} .

Again we have
1
αmin = inf {α : det(I − αM ∆) = 0, ∆ ∈ B∆} = .
max ρ(M ∆)
∆∈B∆

Similar to the unstructured case, we shall call 1/αmin the structured singular value and
denote it by µ∆ (M ).
Definition 10.1 For M ∈ Cn×n , µ∆ (M ) is defined as
1
µ∆ (M ) := (10.4)
min {σ(∆) : ∆ ∈ ∆, det (I − M ∆) = 0}
unless no ∆ ∈ ∆ makes I − M ∆ singular, in which case µ∆ (M ) := 0.
Remark 10.1 Without a loss in generality, the full blocks in the minimal norm ∆ can
each be chosen to be dyads (rank = 1). To see this, assume S = 0 (i.e., all blocks are
full blocks). Suppose that I − M ∆ is singular for some ∆ ∈ ∆. Then there is an x ∈ Cn
such that M ∆x = x. Now partition x conformably with ∆:
 
x1
 
 x2 
x= 
 ..  , xi ∈ C , i = 1, . . . , F
mi
 . 
xF

and let 
 ∆i xi x∗i

 kx k2 , xi 6= 0;

i
˜i =
∆ for i = 1, 2, . . . , F.



 0, xi = 0
Define
˜ = diag{∆
∆ ˜ 1, ∆
˜ 2, . . . , ∆
˜ F }.
Then σ(∆) ˜ ≤ σ(∆), ∆x˜ = ∆x, and thus (I − M ∆)x ˜ = (I − M ∆)x = 0 (i.e., I − M ∆ ˜
is also singular). Hence we have replaced a general perturbation ∆ that satisfies the
singularity condition with a perturbation ∆ ˜ that is no larger [in the σ(·) sense] and has
rank 1 for each block but still satisfies the singularity condition. 3
190 µ AND µ SYNTHESIS

Lemma 10.1 µ∆ (M ) = max ρ (M ∆)

∆∈B∆

In view of this lemma, continuity of the function µ : Cn×n → R is apparent. In general,

though, the function µ : Cn×n → R is not a norm, since it does not satisfy the triangle
inequality; however, for any α ∈ C, µ (αM ) = |α|µ (M ), so in some sense, it is related
to how “big” the matrix is.
We can relate µ∆ (M ) to familiar linear algebra quantities when ∆ is one of two
extreme sets.
• If ∆ = {δI : δ ∈ C} (S = 1, F = 0, r1 = n), then µ∆ (M ) = ρ (M ), the spectral
radius of M .

Proof. The only ∆’s in ∆ that satisfy the det (I − M ∆) = 0 constraint are
reciprocals of nonzero eigenvalues of M . The smallest one of these is associated
with the largest (magnitude) eigenvalue, so, µ∆ (M ) = ρ (M ). 2

• If ∆ = Cn×n (S = 0, F = 1, m1 = n), then µ∆ (M ) = σ(M ).

Obviously, for a general ∆, as in equation (10.1), we must have
{δIn : δ ∈ C} ⊂ ∆ ⊂ Cn×n . (10.5)
Hence directly from the definition of µ and from the two preceding special cases, we
conclude that
ρ (M ) ≤ µ∆ (M ) ≤ σ (M ) . (10.6)
These bounds alone are not sufficient for our purposes because the gap between ρ and
σ can be arbitrarily large.

Example 10.1 Suppose " #

δ1 0
∆=
0 δ2
and consider
" #
0 β
(1) M = for any β > 0. Then ρ(M ) = 0 and σ(M ) = β. But µ(M ) = 0
0 0
since det(I − M ∆) = 1 for all admissible ∆.
" #
−1/2 1/2
(2) M = . Then ρ(M ) = 0 and σ(M ) = 1. Since
−1/2 1/2
δ1 − δ2
det(I − M ∆) = 1 + ,
2

it is easy to see that min maxi |δi | : 1 + δ1 −δ
2
2
= 0 = 1, so µ(M ) = 1.
10.2. Structured Singular Value 191

Thus neither ρ nor σ provide useful bounds even in simple cases. The only time they
do provide reliable bounds is when ρ ≈ σ.

However, the bounds can be refined by considering transformations on M that do

not affect µ∆ (M ), but do affect ρ and σ. To do this, define the following two subsets
of Cn×n :

U = {U ∈ ∆ : U U ∗ = In } (10.7)
(   )
diag D1 , . . . , DS , d1 Im1 , . . . , dF −1 ImF −1 , ImF :
D = . (10.8)
Di ∈ Cri ×ri , Di = Di∗ > 0, dj ∈ R, dj > 0
Note that for any ∆ ∈ ∆, U ∈ U, and D ∈ D,

U∗ ∈ U U∆ ∈ ∆ ∆U ∈ ∆ σ (U ∆) = σ (∆U ) = σ (∆) (10.9)

D∆ = ∆D. (10.10)
Consequently, we have the following:

Theorem 10.2 For all U ∈ U and D ∈ D

µ∆ (M U ) = µ∆ (U M ) = µ∆ (M ) = µ∆ DM D−1 . (10.11)

Proof. For all D ∈ D and ∆ ∈ ∆,

det (I − M ∆) = det I − M D−1 ∆D = det I − DM D−1 ∆


since D commutes with ∆. Therefore µ∆ (M ) = µ∆ DM D−1 . Also, for each
U ∈ U, det (I − M ∆) = 0 if and only if det (I − M U U ∗ ∆) = 0. Since U ∗ ∆ ∈ ∆
and σ (U ∗ ∆) = σ (∆), we get µ∆ (M U ) = µ∆ (M ) as desired. The argument for U M is
the same. 2

Therefore, the bounds in equation (10.6) can be tightened to

max ρ(U M ) ≤ max ρ (∆M ) = µ∆ (M ) ≤ inf σ DM D−1 (10.12)
U∈U ∆∈B∆ D∈D

where the equality comes from Lemma 10.1. Note that the last element in the D matrix
−1
is normalized to 1 since for any nonzero scalar γ, DM D−1 = (γD) M (γD) .

Remark 10.2 Note that the scaling set D in Theorem 10.2 and in inequality (10.12)
is not necessarily restricted to being Hermitian. In fact, it can be replaced by any set of
nonsingular matrices that satisfy equation (10.10). However, enlarging the set of scaling
matrices does not improve the upper-bound in inequality (10.12). This can be shown
192 µ AND µ SYNTHESIS

as follows: Let D be any nonsingular matrix such that D∆ = ∆D. Then there exist a
Hermitian matrix 0 < R = R∗ ∈ D and a unitary matrix U such that D = U R and




inf σ DM D−1 = inf σ U RM R−1 U ∗ = inf σ RM R−1 .
D D R∈D

Therefore, there is no loss of generality in assuming D to be Hermitian. 3

10.2.2 Bounds
In this section we will concentrate on the bounds


max ρ (U M ) ≤ µ∆ (M ) ≤ inf σ DM D−1 .
U∈U D∈D

The lower bound is always an equality (Doyle [1982]).

Theorem 10.3 max ρ(M U ) = µ∆ (M )

U∈U

Unfortunately, the quantity ρ (U M ) can have multiple local maxima that are not global.
Thus local search cannot be guaranteed to obtain µ, but can only yield a lower bound.
For computation purposes one can derive a slightly different formulation of the lower
bound as a power algorithm that is reminiscent of power algorithms for eigenvalues and
singular values (Packard and Doyle [1988a, 1988b]). While there are open questions
about convergence, the algorithm usually works quite well and has proven to be an
effective method to compute µ.
The upper-bound can be reformulated as a convex optimization problem, so the
global minimum can, in principle, be found. Unfortunately, the upper-bound is not
always equal to µ. For block structures ∆ satisfying 2S + F ≤ 3, the upper-bound is
always equal to µ∆ (M ), and for block structures with 2S + F > 3, there exist matrices
for which µ is less than the infimum. This can be summarized in the following diagram,
which shows for which cases the upper-bound is guaranteed to be equal to µ. See
Packard and Doyle [1993] for details.

Theorem 10.4 µ∆ (M ) = inf σ(DM D−1 ) if 2S + F ≤ 3

D∈D

F= 0 1 2 3 4
S=
0 yes yes yes no
1 yes yes no no no
2 no no no no no

Several of the boxes have connections with standard results.

• S = 0, F = 1 : µ∆ (M ) = σ (M ).
10.2. Structured Singular Value 193

• S = 1, F = 0 : µ∆ (M ) = ρ (M ) = inf σ DM D−1 . This is a standard result in
D∈D
linear algebra. In fact, without a loss in generality, the matrix M can be assumed
in Jordan canonical form. Now let
   
λ 1 1
   
 λ 1   k 
   
 . .   . 
J1 =  .. ..  , D1 =  ..  ∈ Cn1 ×n1 .
   
   
 λ 1   k n1 −2 
λ k n1 −1

Then inf σ(D1 J1 D1−1 ) = lim σ(D1 J1 D1−1 ) = |λ|. (Note that by Re-
D1 ∈Cn1 ×n1 k→∞
mark 10.2, the scaling matrix does not need to be Hermitian.) The conclusion
follows by applying this result to each Jordan block.
That µ equals to the preceding upper-bound in this case is also equivalent to the
fact that Lyapunov asymptotic stability and exponential stability are equivalent
for discrete time systems. This is because ρ (M ) < 1 (exponential stability of a
discrete time system matrix M ) implies for some nonsingular D ∈ Cn×n

σ(DM D−1 ) < 1 or (D−1 )∗ M ∗ D∗ DM D−1 − I < 0,

which, in turn, is equivalent to the existence of a P = D∗ D > 0 such that

M ∗P M − P < 0

(Lyapunov asymptotic stability).

• S = 0, F = 2 : This case was studied by Redheffer [1959].
• S = 1, F = 1 : This is equivalent to a state-space characterization of the H∞
norm of a discrete time transfer function.
• S = 2, F = 0 : This is equivalent to the fact that for multidimensional systems
(two dimensional, in fact), exponential stability is not equivalent to Lyapunov
stability.
• S = 0, F ≥ 4 : For this case, the upper-bound is not always equal to µ. This
is important, as these are the cases that arise most frequently in applications.
Fortunately, the bound seems to be close to µ. The worst known example has a
ratio of µ over the bound of about .85, and most systems are close to 1.
The preceding bounds are much more than just computational schemes. They are
also theoretically rich and can unify a number of apparently different results in linear
systems theory. There are several connections with Lyapunov asymptotic stability,
two of which were hinted at previously, but there are further connections between the
194 µ AND µ SYNTHESIS

upper-bound scalings and solutions to Lyapunov and Riccati equations. Indeed, many
major theorems in linear systems theory follow from the upper-bounds and from some
results of linear fractional transformations. The lower bound can be viewed as a natural
generalization of the maximum modulus theorem.
Of course, one of the most important uses of the upper-bound is as a computational
scheme when combined with the lower bound. For reliable use of the µ theory, it is
essential to have upper and lower bounds. Another important feature of the upper-
bound is that it can be combined with H∞ controller synthesis methods to yield an ad
hoc µ-synthesis method. Note that the upper-bound when applied to transfer functions
is simply a scaled H∞ norm. This is exploited in the D − K iteration procedure to
perform approximate µ synthesis (Doyle [1982]), which will be briefly introduced in
Section 10.4.
The upper and lower bounds of the structured singular value and the scaling matrix
D can be computed using the MATLAB command
 [bounds,rowd] = mu(M,blk)
where the structure of the ∆ is specified by a two-column matrix blk. for example, a
 
δ1 I2 0 0 0 0 0
 
 0 δ2 0 0 0 0 
 
 0 0 
 0 ∆3 0 0 
∆= 
 0 0 0 ∆4 0 0 
 
 0 0 δ5 I3 0 
 0 0 
0 0 0 0 0 ∆6

δ1 , δ2 , δ5 , ∈ C, ∆3 ∈ C2×3 , ∆4 ∈ C3×3 , ∆6 ∈ C2×1

can be specified by  
2 0
 
 1 1 
 
 
 2 3 
blk =  .
 3 3 
 
 
 3 0 
2 1
Note that ∆j is not required to be square. The outputs of the program include a 2 × 1
vector bounds containing the upper and lower bounds of µ∆ (M ) and the row vector
rowd containing the scaling D. The D matrix can be recovered by
 [D` , Dr ] = unwrapd(rowd, blk)
where D` and Dr denote the
left and right scaling matrices used in computing the
upper-bound inf σ D` M Dr−1 when some full blocks are not necessarily square and
they are equal if all full blocks are square.
10.2. Structured Singular Value 195

Example 10.2 Let

 
j 2 2j 0 −1 −1 + 3j 2 + 3j
 
 3+j 2−j −1 + j 2+j −1 + j 1 −1 + j 
 
 3+j j 2 + 2j −1 + 2j 3−j 3j −1 + j 
 
 
M =  −1 + j −1 − j j 0 1−j 2−j 2 + 2j 
 
 3 j 1+j 3j 1+j 3j −j 
 
 
 1 3 + 2j 2 + 2j 3j 1 + 2j 2+j −1 + 2j 
2+j −1 − j −1 3 + 3j 2 + 3j 2j 1−j
and
   

 δ1 I2 


   

  δ2  
∆= ∆=

 : δ1 , δ2 ∈ C, ∆3 ∈ C2×3 , ∆4 ∈ C2×1 .


  ∆3  


 

 ∆4 
 
2 0
  h i
 1 1 
Then blk = 

 and the Matlab program gives bounds = 10.5955 10.5518

 2 3 
2 1
and  
D1
 
 0.7638 
D` = 



 0.8809I3 
1.0293
 
D1
 
 0.7638 
Dr = 



 0.8809I2 
1.0293I2
where " #
1.0260 − 0.0657j 0.2174 − 0.3471j
D1 = .
−0.0701 + 0.3871j −0.4487 − 0.6953j
In fact, D` and Dr can be replaced by Hermitian matrices without changing the upper-
bound by replacing D1 with
" #
1.0992 0.0041 − 0.0591j
D̂1 =
0.0041 + 0.0591j 0.9215
196 µ AND µ SYNTHESIS

since D1 = U1 D̂1 and

" #
0.9155 − 0.0713j 0.2365 − 0.3177j
U1 =
−0.1029 + 0.3824j −0.5111 − 0.7629j

is a unitary matrix.

10.2.3 Well-Posedness and Performance for Constant LFTs

Let M be a complex matrix partitioned as
" #
M11 M12
M= (10.13)
M21 M22

and suppose there are two defined block structures, ∆1 and ∆2 , which are compatible
in size with M11 and M22 , respectively. Define a third structure ∆ as
(" # )
∆1 0
∆= : ∆1 ∈ ∆1 , ∆2 ∈ ∆2 . (10.14)
0 ∆2

Now we may compute µ with respect to three structures. The notations we use to keep
track of these computations are as follows: µ1 (·) is with respect to ∆1 , µ2 (·) is with
respect to ∆2 , and µ∆ (·) is with respect to ∆. In view of these notations, µ1 (M11 ),
µ2 (M22 ), and µ∆ (M ) all make sense, though, for instance, µ1 (M ) does not.
This section is interested in following constant matrix problems:
• Determine whether the LFT F` (M, ∆2 ) is well-defined for all ∆2 ∈ ∆2 with
σ(∆2 ) ≤ β (< β).
• If so, determine how “large” F` (M, ∆2 ) can get for this norm-bounded set of
perturbations.
Let ∆2 ∈ ∆2 . Recall that F` (M, ∆2 ) is well-defined if I − M22 ∆2 is invertible. The
first theorem is nothing more than a restatement of the definition of µ.

Theorem 10.5 The linear fractional transformation F` (M, ∆2 ) is well-defined

(a) for all ∆2 ∈ B∆2 if and only if µ2 (M22 ) < 1.
(b) for all ∆2 ∈ Bo ∆2 if and only if µ2 (M22 ) ≤ 1.

As the “perturbation” ∆2 deviates from zero, the matrix F` (M, ∆2 ) deviates from
M11 . The range of values that µ1 (F` (M, ∆2 )) takes on is intimately related to µ∆ (M ),
as shown in the following theorem:
10.2. Structured Singular Value 197

Theorem 10.6 (main loop theorem) The following are equivalent:




 µ2 (M22 ) < 1, and
µ∆ (M ) < 1 ⇐⇒


 max µ1 (F` (M, ∆2 )) < 1
∆2 ∈B∆2


 µ2 (M22 ) ≤ 1, and

µ∆ (M ) ≤ 1 ⇐⇒


 sup µ1 (F` (M, ∆2 )) ≤ 1.
∆2 ∈Bo ∆2

Proof. We shall only prove the first part of the equivalence. The proof for the second
part is similar.
⇐ Let ∆i ∈ ∆i be given, with σ (∆i ) ≤ 1, and define ∆ = diag [∆1 , ∆2 ]. Obviously
∆ ∈ ∆. Now " #
I − M11 ∆1 −M12 ∆2
det (I − M ∆) = det . (10.15)
−M21 ∆1 I − M22 ∆2
By hypothesis I − M22 ∆2 is invertible, and hence det (I − M ∆) becomes
 
−1
det (I − M22 ∆2 ) det I − M11 ∆1 − M12 ∆2 (I − M22 ∆2 ) M21 ∆1 .

Collecting the ∆1 terms leaves

det (I − M ∆) = det (I − M22 ∆2 ) det (I − F` (M, ∆2 ) ∆1 ) . (10.16)

But µ1 (F` (M, ∆2 )) < 1 and ∆1 ∈ B∆1 , so I − F` (M, ∆2 ) ∆1 must be nonsingular.

Therefore, I − M ∆ is nonsingular and, by definition, µ∆ (M ) < 1.
⇒ Basically, the argument above is reversed. Again let ∆1 ∈ B∆1 and
∆2 ∈ B∆2 be given, and define ∆ = diag [∆1 , ∆2 ]. Then ∆ ∈ B∆ and, by hypothesis,
det (I − M ∆) 6= 0. It is easy to verify from the definition of µ that (always)

µ (M ) ≥ max {µ1 (M11 ) , µ2 (M22 )} .

We can see that µ2 (M22 ) < 1, which gives that I−M22 ∆2 is also nonsingular. Therefore,
the expression in equation (10.16) is valid, giving

det (I − M22 ∆2 ) det (I − F` (M, ∆2 ) ∆1 ) = det (I − M ∆) 6= 0.

Obviously, I − F` (M, ∆2 ) ∆1 is nonsingular for all ∆i ∈ B∆i , which indicates that

the claim is true. 2

Remark 10.3 This theorem forms the basis for all uses of µ in linear system robustness
analysis, whether from a state-space, frequency domain, or Lyapunov approach. 3
198 µ AND µ SYNTHESIS

The role of the block structure ∆2 in the main loop theorem is clear — it is the
structure that the perturbations come from; however, the role of the perturbation struc-
ture ∆1 is often misunderstood. Note that µ1 (·) appears on the right-hand side of
the theorem, so that the set ∆1 defines what particular property of F` (M, ∆2 ) is
considered. As an example, consider the theorem applied with the two simple block
structures considered right after Lemma 10.1. Define ∆1 := {δ1 In : δ1 ∈ C}. Hence,
for A ∈ Cn×n , µ1 (A) = ρ (A). Likewise, define ∆2 = Cm×m ; then for D ∈ Cm×m ,
µ2 (D) = σ(D). Now, let ∆ be the diagonal augmentation of these two sets, namely
(" # )
δ1 In 0n×m
∆ := : δ1 ∈ C, ∆2 ∈ Cm×m
⊂ C(n+m)×(n+m) .
0m×n ∆2

Let A ∈ Cn×n , B ∈ Cn×m , C ∈ Cm×n , and D ∈ Cm×m be given, and interpret them as
the state-space model of a discrete time system

xk+1 = Axk + Buk

yk = Cxk + Duk .

Let M ∈ C(n+m)×(n+m) be the block state-space matrix of the system

" #
A B
M= .
C D

Applying the theorem with these data gives that the following are equivalent:
• The spectral radius of A satisfies ρ (A) < 1, and
 
max σ D + Cδ1 (I − Aδ1 )−1 B < 1. (10.17)
δ1 ∈C
|δ1 |≤1

• The maximum singular value of D satisfies σ(D) < 1, and

 
−1
max ρ A + B∆2 (I − D∆2 ) C < 1. (10.18)
∆2 ∈Cm×m
σ(∆2 )≤1

• The structured singular value of M satisfies

µ∆ (M ) < 1. (10.19)

The first condition is recognized by two things: The system is stable, and the || · ||∞
norm on the transfer function from u to y is less than 1 (by replacing δ1 with 1z ):
   
kGk∞ := max σ D + C (zI − A)−1 B = max σ D + Cδ1 (I − Aδ1 )−1 B .
z∈C δ1 ∈C
|z|≥1 |δ1 |≤1
10.2. Structured Singular Value 199

−1
The second condition implies that (I − D∆2 ) is well defined for all σ(∆2 ) ≤ 1 and
that a robust stability result holds for the uncertain difference equation
 
xk+1 = A + B∆2 (I − D∆2 )−1 C xk

where ∆2 is any element in Cm×m with σ(∆2 ) ≤ 1, but otherwise unknown.

This equivalence between the small gain condition, kGk∞ < 1, and the stability
robustness of the uncertain difference equation is well-known. This is the small gain
theorem, in its necessary and sufficient form for linear, time invariant systems with one
of the components norm bounded, but otherwise unknown. What is important to note
is that both of these conditions are equivalent to a condition involving the structured
singular value of the state-space matrix. Already we have seen that special cases of µ are
the spectral radius and the maximum singular value. Here we see that other important
linear system properties — namely, robust stability and input-output gain — are also
related to a particular case of the structured singular value.

Example 10.3 Let M , ∆1 , and ∆2 be defined as in the beginning of this section. Now
suppose µ2 (M22 ) < 1. Find

max µ1 (F` (M, ∆2 )) .

∆2 ∈B∆2

This can be done iteratively as follows:

max µ1 (F` (M, ∆2 )) = α

∆2 ∈B∆2
" # !!
M11 /α M12 /α
⇐⇒ max µ1 F` , ∆2 =1
∆2 ∈B∆2 M21 M22
" #!
M11 /α M12 /α
⇐⇒ µ∆ = 1.
M21 M22
Hence
( " #! )
M11 /α M12 /α
max µ1 (F` (M, ∆2 )) = α: µ∆ =1 .
∆2 ∈B∆2 M21 M22

For example, let ∆1 = δI2 , ∆2 ∈ C2×2 :

" # " # " # " #
0.1 0.2 1 0 1 2 0.5 0
A= , B= , C= , D= .
1 0 1 1 1 3 0 0.8

Find
αmax = sup ρ(A + B∆2 (I − D∆2 )−1 C).
σ(∆2 )≤1
200 µ AND µ SYNTHESIS

" #
δI2
Define ∆ = . Then a bisection search can be done to find
∆2
( " #! )
A/α B/α
αmax = α: µ∆ =1 = 21.77.
C D

Related MATLAB Commands: unwrapp, muunwrap, dypert, sisorat

10.3 Structured Robust Stability and Performance

10.3.1 Robust Stability
The most well-known use of µ as a robustness analysis tool is in the frequency domain.
Suppose G(s) is a stable, real rational, multi-input, multioutput transfer function of
a linear system. For clarity, assume G has q1 inputs and p1 outputs. Let ∆ be a
block structure, as in equation (10.1), and assume that the dimensions are such that
∆ ⊂ Cq1 ×p1 . We want to consider feedback perturbations to G that are themselves
dynamical systems with the block diagonal structure of the set ∆.
Let M (∆) denote the set of all block diagonal and stable rational transfer functions
that have block structures such as ∆.

M (∆) := ∆(·) ∈ RH∞ : ∆(so ) ∈ ∆ for all so ∈ C+

Theorem 10.7 Let β > 0. The loop shown below is well-posed and internally stable
for all ∆(·) ∈ M (∆) with k∆k∞ < β1 if and only if

sup µ∆ (G(jω)) ≤ β.
ω∈R

w1 e1
-e - ∆
+ 6
+

e2 + w2
?
+
G(s)  e

Proof. (⇐=) Suppose sups∈C+ µ∆ (G(s)) ≤ β. Then det(I − G(s)∆(s)) 6= 0 for all
s ∈ C+ ∪ {∞} whenever k∆k∞ < 1/β (i.e., the system is robustly stable). Now it is
sufficient to show that

sup µ∆ (G(s)) = sup µ∆ (G(jω)).

s∈C+ ω∈R
10.3. Structured Robust Stability and Performance 201

It is clear that
sup µ∆ (G(s)) = sup µ∆ (G(s)) ≥ sup µ∆ (G(jω)).
s∈C+ s∈C+ ω

Now suppose sups∈C+ µ∆ (G(s)) > β; then by the definition of µ, there is an so ∈

C+ ∪ {∞} and a complex structured ∆ such that σ(∆) < 1/β and det(I − G(so )∆) = 0.
This implies that there is a 0 ≤ ω̂ ≤ ∞ and 0 < α ≤ 1 such that det(I − G(j ω̂)α∆) =
0. This, in turn, implies that µ∆ (G(j ω̂)) > β since σ(α∆) < 1/β. In other words,
sups∈C+ µ∆ (G(s)) ≤ supω µ∆ (G(jω)). The proof is complete.
(=⇒) Suppose supω∈R µ∆ (G(jω)) > β. Then there is a 0 < ωo < ∞ such that
µ∆ (G(jωo )) > β. By Remark 10.1, there is a complex ∆c ∈ ∆ that each full block
has rank 1 and σ(∆c ) < 1/β such that I − G(jωo )∆c is singular. Next, using the same
construction used in the proof of the small gain theorem (Theorem 8.1), one can find a
rational ∆(s) such that k∆(s)k∞ = σ(∆c ) < 1/β, ∆(jωo ) = ∆c , and ∆(s) destabilizes
the system. 2

Hence, the peak value on the µ plot of the frequency response determines the size
of perturbations that the loop is robustly stable against.
Remark 10.4 The internal stability with a closed ball of uncertainties is more compli-
cated. The following example is shown in Tits and Fan [1995]. Consider
" #
1 0 −1
G(s) =
s+1 1 0

and ∆ = δ(s)I2 . Then

1
sup µ∆ (G(jω)) = sup = µ∆ (G(j0)) = 1.
ω∈R ω∈R |jω + 1|

On the other hand, µ∆ (G(s)) < 1 for all s 6= 0, s ∈ C+ , and the only matrices in the
form of Γ = γI2 with |γ| ≤ 1 for which
det(I − G(0)Γ) = 0
are the complex matrices ±jI2 . Thus, clearly, (I − G(s)∆(s))−1 ∈ RH∞ for all real
rational ∆(s) = δ(s)I2 with kδk∞ ≤ 1 since ∆(0) must be real. This shows that
supω∈R µ∆ (G(jω)) < 1 is not necessary for (I − G(s)∆(s))−1 ∈ RH∞ with the closed
ball of structured uncertainty k∆k∞ ≤ 1. Similar examples with no repeated blocks are
1
generated by setting G(s) = s+1 M , where M is any real matrix with µ∆ (M ) = 1 for
which there is no real ∆ ∈ ∆ with σ(∆) = 1 such that det(I − M ∆) = 0. For example,
let
    
0 β " #  δ1 
 
  −β α α  
M = γ α  , ∆=  δ2  , δi ∈ C
0 −γ γ 
 

γ −α δ3
202 µ AND µ SYNTHESIS

with γ 2 = 12 and β 2 + 2α2 = 1. Then it is shown in Packard and Doyle [1993] that
µ∆ (M ) = 1 and all ∆ ∈ ∆ with σ(∆) = 1 that satisfy det(I − M ∆) = 0 must be
complex. 3

Remark 10.5 Let ∆ ∈ RH∞ be a structured uncertainty and

" #
G11 (s) G12 (s)
G(s) = ∈ RH∞ .
G21 (s) G22 (s)

Then Fu (G, ∆) ∈ RH∞ does not necessarily imply (I − G11 ∆)−1 ∈ RH∞ whether ∆
is in an open ball or is in a closed ball. For example, consider
 1 
s+1 0 1
 
G(s) =  0 10
s+1 0 
1 0 0
" #
δ1 1
and ∆ = with k∆k∞ < 1. Then Fu (G, ∆) = 1 ∈ RH∞ for all
δ2 1 − δ1 s+1
admissible ∆ (k∆k∞ < 1) but (I − G11 ∆)−1 ∈ RH∞ is true only for k∆k∞ < 0.1. 3

10.3.2 Robust Performance

Often, stability is not the only property of a closed-loop system that must be robust to
perturbations. Typically, there are exogenous disturbances acting on the system (wind
gusts, sensor noise) that result in tracking and regulation errors. Under perturbation,
the effect that these disturbances have on error signals can greatly increase. In most
cases, long before the onset of instability, the closed-loop performance will degrade to
the point of unacceptability (hence the need for a “robust performance” test). Such
a test will indicate the worst-case level of performance degradation associated with a
given level of perturbations.
Assume Gp is a stable, real-rational, proper transfer function with q1 + q2 inputs
and p1 + p2 outputs. Partition Gp in the obvious manner
" #
G11 G12
Gp (s) =
G21 G22

so that G11 has q1 inputs and p1 outputs, and so on. Let ∆ ⊂ Cq1 ×p1 be a block
structure, as in equation (10.1). Define an augmented block structure:
(" # )
∆ 0 q2 ×p2
∆P := : ∆ ∈ ∆, ∆f ∈ C .
0 ∆f

The setup is to address theoretically the robust performance questions about the
following loop:
10.3. Structured Robust Stability and Performance 203

- ∆(s)

z  w
 Gp (s) 

The transfer function from w to z is denoted by Fu (Gp , ∆).

Theorem 10.8 Let β > 0. For all ∆(s) ∈ M (∆) with k∆k∞ < β1 , the loop shown
above is well-posed, internally stable, and kFu (Gp , ∆)k∞ ≤ β if and only if

sup µ∆P (Gp (jω)) ≤ β.

ω∈R

Note that by internal stability, supω∈R µ∆ (G11 (jω)) ≤ β, then the proof of this
theorem is exactly along the lines of the earlier proof for Theorem 10.7, but also appeals
to Theorem 10.6. This is a remarkably useful theorem. It says that a robust performance
problem is equivalent to a robust stability problem with augmented uncertainty ∆, as
shown in Figure 10.5.

- ∆f

- ∆


Gp (s)


Figure 10.5: Robust performance vs robust stability

Example 10.4 We shall consider again the HIMAT problem from Example 9.1. Use
the Simulink block diagram in Example 9.1 and run the following commands to get
an interconnection model Ĝ, an H∞ stabilizing controller K and a closed-loop transfer
matrix Gp (s) = F` (Ĝ, K). (Do not bother to figure out how hinfsyn works; it will be
considered in detail in Chapter 14.)
204 µ AND µ SYNTHESIS

 [A, B, C, D] = linmod(0 aircraft0 )

 Ĝ = pck(A, B, C, D);

 [K, Gp , γ] = hinfsyn(Ĝ, 2, 2, 0, 10, 0.001, 2);

which gives γ = 1.8612 = kGp k∞ , a stabilizing controller K, and a closed loop transfer
matrix Gp :
 
  p 1
 
z1  p2 
    " #
 z2   d 
   1  G p11 G p12
 e  = Gp (s) 
 d2 
 , Gp (s) = .
 1    Gp21 Gp22
 n 
e2  1 
n2

maximum singular value

1.5

0.5

0 −3 −2 −1 0 1 2 3
10 10 10 10 10 10 10
frequency (rad/sec)

Figure 10.6: Singular values of Gp (jω)

Now generate the singular value frequency responses of Gp :

 w=logspace(-3,3,300);

 Gpf = frsp(Gp , w); % Gpf is the frequency response of Gp ;

 [u, s, v] = vsvd(Gpf );
10.3. Structured Robust Stability and Performance 205

 vplot(0 liv, m0 , s)
The singular value frequency responses of Gp are shown in Figure 10.6. To test the
robust stability, we need to compute kGp11 k∞ :
 Gp11 = sel(Gp , 1 : 2, 1 : 2);
 norm of Gp11 = hinfnorm(Gp11 , 0.001);
which gives kGP 11 k∞ = 0.933 < 1. So the system is robustly stable. To check the
robust performance, we shall compute the µ∆P (Gp (jω)) for each frequency with
" #
∆
∆P = , ∆ ∈ C2×2 , ∆f ∈ C4×2 .
∆f

Maximum Singular Value and mu

2

maximum singular value

1.5

mu bounds

0.5 −3 −2 −1 0 1 2 3
10 10 10 10 10 10 10
frequency (rad/sec)

Figure 10.7: µ∆P (Gp (jω)) and σ(Gp (jω))

 blk=[2,2;4,2];
 [bnds,dvec,sens,pvec]=mu(Gpf,blk);
 vplot(0 liv, m0 , vnorm(Gpf ), bnds)
 title(0 Maximum Singular Value and mu0 )
 xlabel(0 frequency(rad/sec)0 )
206 µ AND µ SYNTHESIS

 text(0.01, 1.7,0 maximum singular value0 )

 text(0.5, 0.8,0 mu bounds0 )

The structured singular value µ∆P (Gp (jω)) and σ(Gp (jω)) are shown in Figure 10.7.
It is clear that the robust performance is not satisfied. Note that
" #!
Gp11 Gp12
max kFu (Gp , ∆)k∞ ≤ γ ⇐⇒ sup µ∆P ≤ 1.
k∆k∞ ≤1 ω Gp21 /γ Gp22 /γ

Using a bisection algorithm, we can also find the worst performance:

max kFu (Gp , ∆)k∞ = 12.7824.

k∆k∞ ≤1

10.3.3 Two-Block µ: Robust Performance Revisited

Suppose that the uncertainty block is given by
" #
∆1
∆= ∈ RH∞
∆2

with k∆k∞ < 1 and that the interconnection model G is given by

" #
G11 (s) G12 (s)
G(s) = ∈ RH∞ .
G21 (s) G22 (s)

Then the closed-loop system is well-posed and internally stable iff supω µ∆ (G(jω)) ≤ 1.
Let " #
dω I
Dω = , dω ∈ R+ .
I
Then " #
G11 (jω) dω G12 (jω)
Dω G(jω)Dω−1 = 1
.
G
dω 21 (jω) G22 (jω)
Hence, by Theorem 10.4, at each frequency ω
" #!
G11 (jω) dω G12 (jω)
µ∆ (G(jω)) = inf σ 1
. (10.20)
dω ∈R+
dω G21 (jω) G22 (jω)

Since the minimization is convex in log dω (see, Doyle [1982]), the optimal dω can
be found by a search; however, two approximations to dω can be obtained easily by
approximating the right-hand side of equation (10.20):
10.3. Structured Robust Stability and Performance 207

(1) Note that

" #!
kG11 (jω)k dω kG12 (jω)k
µ∆ (G(jω)) ≤ inf σ
dω kG21 (jω)k kG22 (jω)k
dω ∈R+ 1

s  
2 2 1 2 2
≤ kG11 (jω)k + d2ω kG12 (jω)k + 2 kG21 (jω)k + kG22 (jω)k
inf
dω ∈R+ dω
q
2 2
= kG11 (jω)k + kG22 (jω)k + 2 kG12 (jω)k kG21 (jω)k

with the minimizing dω given by

 q

 kG21 (jω)k
if G12 6= 0 & G21 6= 0,
 kG12 (jω)k
dˆω = 0 if G21 = 0, (10.21)


 ∞ if G12 = 0.

(2) Alternative approximation can be obtained by using the Frobenius norm:

" #
dω G12 (jω)
G11 (jω)
µ∆ (G(jω)) ≤ inf 1
dω ∈R+
dω G21 (jω) G22 (jω)
F

s  
1
= inf kG11 (jω)k2F + d2ω kG12 (jω)k2F + kG21 (jω)k2F + kG22 (jω)k2F
dω ∈R+ d2ω
q
2 2
= kG11 (jω)kF + kG22 (jω)kF + 2 kG12 (jω)kF kG21 (jω)kF

with the minimizing dω given by

 q


kG21 (jω)kF
if G12 6= 0 & G21 6= 0,
 kG12 (jω)kF
˜
dω = 0 if G21 = 0, (10.22)


 ∞ if G12 = 0.

It can be shown that the approximations for the scalar dω obtained previously are exact
for a 2 × 2 matrix G. For higher dimensional G, the approximations for dω are still
reasonably good. Hence an approximation of µ can be obtained as
" #!
G11 (jω) dˆω G12 (jω)
µ∆ (G(jω)) ≤ σ 1 (10.23)
d̂
G21 (jω) G22 (jω)
ω
208 µ AND µ SYNTHESIS

or, alternatively, as
" #!
G11 (jω) d˜ω G12 (jω)
µ∆ (G(jω)) ≤ σ 1 . (10.24)
d˜
G21 (jω) G22 (jω)
ω

We can now see how these approximated µ tests are compared with the sufficient
conditions obtained in Chapter 8.

Example 10.5 Consider again the robust performance problem of a system with out-
put multiplicative uncertainty in Chapter 8 (see Figure 8.10):

P∆ = (I + W1 ∆W2 )P, k∆k∞ < 1.

Then it is easy to show that the problem can be put in the general framework by
selecting " #
−W2 To W1 −W2 To Wd
G(s) =
We So W1 We So Wd
and that the robust performance condition is satisfied if and only if

kW2 To W1 k∞ ≤ 1 (10.25)

and
kFu (G, ∆)k∞ ≤ 1 (10.26)
for all ∆ ∈ RH∞ with k∆k∞ < 1. But equations (10.25) and (10.26) are satisfied iff
for each frequency ω
" #!
−W2 To W1 −dω W2 To Wd
µ∆ (G(jω)) = inf σ 1
≤ 1.
dω ∈R+
dω We So W1 We So Wd

Note that, in contrast to the sufficient condition obtained in Chapter 8, this condition is
an exact test for robust performance. To compare the µ test with the criteria obtained
in Chapter 8, some upper-bounds for µ can be derived. Let
s
kWe So W1 k
dω = .
kW2 To Wd k

Then, using the first approximation for µ, we get

q
2 2
µ∆ (G(jω)) ≤ kW2 To W1 k + kWe So Wd k + 2 kW2 To Wd k kWe So W1 k
q
kW2 To W1 k + kWe So Wd k + 2κ(W1−1 Wd ) kW2 To W1 k kWe So Wd k
2 2
≤
≤ kW2 To W1 k + κ(W1−1 Wd ) kWe So Wd k
10.3. Structured Robust Stability and Performance 209

where W1 is assumed to be invertible in the last two inequalities. The last term is
exactly the sufficient robust performance criteria obtained in Chapter 8. It is clear that
any term preceding the last forms a tighter test since κ(W1−1 Wd ) ≥ 1. Yet another
alternative sufficient test can be obtained from the preceding sequence of inequalities:
q
µ∆ (G(jω)) ≤ κ(W1−1 Wd )(kW2 To W1 k + kWe So Wd k).
Note that this sufficient condition is not easy to get from the approach taken in Chapter 8
and is potentially less conservative than the bounds derived there.

Next we consider the skewed specification problem, but first the following lemma is
needed in the sequel.

Lemma 10.9 Suppose σ = σ1 ≥ σ2 ≥ . . . ≥ σm = σ > 0, then

 
2 1 σ σ
inf max (dσi ) + = + .
d∈R+ i (dσi )2 σ σ

Proof. Consider a function y = x + 1/x; then y is a convex function and the maxi-
mization over a closed interval is achieved at the boundary of the interval. Hence for
any fixed d
   
1 1 1
max (dσi )2 + = max (dσ)2
+ , (dσ)2
+ .
i (dσi )2 (dσ)2 (dσ)2

Then the minimization over d is obtained iff

1 1
(dσ)2 + 2
= (dσ)2 + ,
(dσ) (dσ)2

which gives d2 = 1
σ σ. The result then follows from substituting d. 2

Example 10.6 As another example, consider again the skewed specification problem
from Chapter 8. Then the corresponding G matrix is given by
" #
−W2 Ti W1 −W2 KSo Wd
G= .
We So P W1 We So Wd

So the robust performance specification is satisfied iff

" #!
−W2 Ti W1 −dω W2 KSo Wd
µ∆ (G(jω)) = inf σ 1
≤1
dω ∈R+
dω We So P W1 We So Wd
210 µ AND µ SYNTHESIS

for all ω ≥ 0. As in the last example, an upper-bound can be obtained by taking

s
kWe So P W1 k
dω = .
kW2 KSo Wd k

Then q
µ∆ (G(jω)) ≤ κ(Wd−1 P W1 )(kW2 Ti W1 k + kWe So Wd k).
In particular, this suggests that the robust performance margin is inversely proportional
to the square root of the plant condition number if Wd = I and W1 = I. This can be
further illustrated by considering a plant-inverting control system.
To simplify the exposition, we shall make the following assumptions:
We = ws I, Wd = I, W1 = I, W2 = wt I,
and P is stable and has a stable inverse (i.e., minimum phase) (P can be strictly proper).
Furthermore, we shall assume that the controller has the form
K(s) = P −1 (s)l(s)
where l(s) is a scalar loop transfer function that makes K(s) proper and stabilizes
the closed loop. This compensator produces diagonal sensitivity and complementary
sensitivity functions with identical diagonal elements; namely,
1 l(s)
So = Si = I, To = Ti = I.
1 + l(s) 1 + l(s)
Denote
1 l(s)
ε(s) = , τ (s) =
1 + l(s) 1 + l(s)
and substitute these expressions into G; we get
" #
−wt τ I −wt τ P −1
G= .
ws εP ws εI
The structured singular value for G at frequency ω can be computed by
" #!
−wt τ I −wt τ (dP )−1
µ∆ (G(jω)) = inf σ .
d∈R+ ws εdP ws εI

Let the singular value decomposition of P (jω) at frequency ω be

P (jω) = U ΣV ∗ , Σ = diag(σ1 , σ2 , . . . , σm )
with σ1 = σ and σm = σ, where m is the dimension of P . Then
" #!
−wt τ I −wt τ (dΣ)−1
µ∆ (G(jω)) = inf σ
d∈R+ ws εdΣ ws εI
10.3. Structured Robust Stability and Performance 211

since unitary operations do not change the singular values of a matrix. Note that
" #
−wt τ I −wt τ (dΣ)−1
= P1 diag(M1 , M2 , . . . , Mm )P2
ws εdΣ ws εI

where P1 and P2 are permutation matrices and where

" #
−wt τ −wt τ (dσi )−1
Mi = .
ws εdσi ws ε

Hence
" #!
−wt τ−wt τ (dσi )−1
µ∆ (G(jω)) = inf max σ
d∈R+ i ws εdσi ws ε
" # !
−wt τ h i
= inf max σ −1
1 (dσi )
d∈R+ i ws εdσi
p
= inf max (1 + |dσi |−2 )(|ws εdσi |2 + |wt τ |2 )
d∈R+ i
s  
 wt τ 2
= inf max |ws ε| + |wt τ | + |ws εdσi | + 
2 2 2   .
d∈R+ i dσi 

Using Lemma 10.9, it is easy to show that the maximum is achieved at either σ or σ
and that optimal d is given by
|wt τ |
d2 = ,
|ws ε|σσ
so the structured singular value is
s
1
µ∆ (G(jω)) = |ws ε|2 + |wt τ |2 + |ws ε||wt τ |[κ(P ) + ]. (10.27)
κ(P )

Note that if |ws ε| and |wt τ | are not too large, which is guaranteed if the nominal
performance and robust stability conditions are satisfied, then the structured singular
value is proportional to the square root of the plant condition number:
p
µ∆ (G(jω)) ≈ |ws ε||wt τ |κ(P ) . (10.28)

This confirms our intuition that an ill-conditioned plant with skewed specifications is
hard to control.
212 µ AND µ SYNTHESIS

10.3.4 Approximation of Multiple Full Block µ

The approximations given in the last subsection can be generalized to the multiple-block
µ problem by assuming that M is partitioned consistently with the structure of

∆ = diag(∆1 , ∆2 , . . . , ∆F )

so that  
M11 M12 ··· M1F
 
 M21 M22 ··· M2F 
M =
 .. .. .. 

 . . . 
MF 1 MF 2 · · · MF F
and
D = diag(d1 I, . . . , dF −1 I, I).
Now  
−1 di
DM D = Mij , dF := 1.
dj
Hence
 
−1 di
µ∆ (M ) ≤ inf σ(DM D ) = inf σ Mij
D∈D D∈D dj
v
  u F F
di uX X 2 d
2
≤ inf σ kMij k ≤ inf t kMij k i2
D∈D dj D∈D
i=1 j=1
dj
v
u F F
uX X 2 d
2
≤ inf t kMij kF i2 .
D∈D
i=1 j=1
dj

An approximate D can be found by solving the following minimization problem:

X
F X
F
2 d2i
inf kMij k
D∈D
i=1 j=1
d2j

or, more conveniently, by minimizing

X
F X
F
2 d2i
inf kMij kF
D∈D
i=1 j=1
d2j

with dF = 1. The optimal di minimizing the preceding two problems satisfies, respec-
tively,
P 2 2
i6=k kMik k di
4
dk = P 2 , k = 1, 2, . . . , F − 1 (10.29)
j6=k kMkj k /dj
2
10.4. Overview of µ Synthesis 213

and P 2
i6=k kMik kF d2i
d4k =P 2 , k = 1, 2, . . . , F − 1. (10.30)
j6=k kMkj kF /d2j
Using these relations, dk can be obtained by iterations.

Example 10.7 Consider a 3 × 3 complex matrix

 
1+j 10 − 2j −20j
 
M =  5j 3+j −1 + 3j 
−2 j 4−j

with structured ∆ = diag(δ1 , δ2 , δ3 ). The largest singular value of M is σ(M ) = 22.9094

and the structured singular value of M computed using the µ Analysis and Synthesis
Toolbox is equal to its upper-bound:

µ∆ (M ) = inf σ(DM D−1 ) = 11.9636

D∈D

with the optimal scaling Dopt = diag(0.3955, 0.6847, 1). The optimal D minimizing

X
F X
F
2 d2i
inf kMij k
D∈D
i=1 j=1
d2j

is Dsubopt = diag(0.3212, 0.4643, 1), which is solved from equation (10.29). Using this
Dsubopt , we obtain another upper-bound for the structured singular value:
−1
µ∆ (M ) ≤ σ(Dsubopt M Dsubopt ) = 12.2538.

One may also use this Dsubopt as an initial guess for the exact optimization.

10.4 Overview of µ Synthesis

This section briefly outlines various synthesis methods. The details are somewhat com-
plicated and are treated in the other parts of this book. At this point, we simply want
to point out how the analysis theory discussed in the previous sections leads naturally
to synthesis questions.
From the analysis results, we see that each case eventually leads to the evaluation
of
kM kα α = 2, ∞, or µ (10.31)
214 µ AND µ SYNTHESIS

for some transfer matrix M . Thus when the controller is put back into the problem, it
involves only a simple linear fractional transformation, as shown in Figure 10.8, with

M = F` (G, K) = G11 + G12 K(I − G22 K)−1 G21

" #
G11 G12
where G = is chosen, respectively, as
G21 G22
" #
P22 P23
• nominal performance only (∆ = 0): G =
P32 P33
" #
P11 P13
• robust stability only: G =
P31 P33
 
P11 P12 P13
 
• robust performance: G = P =  P21 P22 P23 .
P31 P32 P33

z w
 
G 

- K

Figure 10.8: Synthesis framework

Each case then leads to the synthesis problem

min kF` (G, K)kα for α = 2, ∞, or µ, (10.32)

K

which is subject to the internal stability of the nominal.

The solutions of these problems for α = 2 and ∞ are the focus of the rest of this
book. The α = 2 case was already known in the 1960s, and the result presented in this
book is simply a new interpretation. The two Riccati solutions for the α = ∞ case were
new products of the late 1980s.
The synthesis for the α = µ case is not yet fully solved. Recall that µ may be
obtained by scaling and applying k·k∞ (for F ≤ 3 and S = 0); a reasonable approach
is to “solve”
min inf DF` (G, K)D−1 (10.33)
K D,D −1
∈H∞ ∞

by iteratively solving for K and D. This is the so-called D-K iteration. The stable and
minimum phase scaling matrix D(s) is chosen such that D(s)∆(s) = ∆(s)D(s). [Note
10.4. Overview of µ Synthesis 215

that D(s) is not necessarily belonging to D since D(s) is not necessarily Hermitian,

see Remark 10.2.] For a fixed scaling transfer matrix D, minK DF` (G, K)D−1 ∞ is
a standard H∞ optimization problem that will be solved later in this book. For a
given stabilizing controller K, inf D,D−1 ∈H∞ DF` (G, K)D−1 ∞ is a standard convex
optimization problem and it can be solved pointwise in the frequency domain:
 
sup inf σ Dω F` (G, K)(jω)Dω−1 .
ω Dω ∈D

Indeed,
 
inf DF` (G, K)D−1 = sup inf σ Dω F` (G, K)(jω)Dω−1 .
∞
D,D−1 ∈H∞ ω Dω ∈D

This follows intuitively from the following arguments: The left-hand side is always no
smaller than the right-hand side, and, on the other hand, given Dω ∈ D, there is always a
real-rational function D(s), stable with stable inverse, such that the Hermitian positive
definite factor in the polar decomposition of D(jω) uniformly approximates Dω over
ω in R. In particular, in the case of scalar blocks, the magnitude |D(jω)| uniformly
approximates Dω over R.
Note that when S = 0 (no scalar blocks),

1 I, . . . , dF −1 I, I) ∈ D,
Dω = diag(dω ω

which is a block diagonal scaling matrix applied pointwise across frequency to the fre-
quency response F` (G, K)(jω).

 D   D−1 
G


- K

Figure 10.9: µ synthesis via scaling

D-K iterations proceed by performing this two-parameter minimization in sequential

fashion: first minimizing over K with D fixed, then minimizing pointwise over D with K
fixed, then again over K, and again over D, etc. Details of this process are summarized
in the following steps:
(i) Fix an initial estimate of the scaling matrix Dω ∈ D pointwise across frequency.
(ii) Find scalar transfer functions di (s), d−1
i (s) ∈ RH∞ for i = 1, 2, . . . , (F − 1) such
that |di (jω)| ≈ dω
i . This step can be done using the interpolation theory (Youla
and Saito [1967]); however, this will usually result in very high-order transfer
functions, which explains why this process is currently done mostly by graphical
matching using lower-order transfer functions.
216 µ AND µ SYNTHESIS

(iii) Let
D(s) = diag(d1 (s)I, . . . , dF −1 (s)I, I).
Construct a state-space model for system
" # " #
D(s) D−1 (s)
Ĝ(s) = G(s) ,
I I

as shown in Figure 10.9.

(iv) Solve an H∞ -optimization problem to minimize

F` (Ĝ, K)
∞

over all stabilizing K’s. Note that this optimization problem uses the scaled
version of G. Let its minimizing controller be denoted by K̂.

(v) Minimize σ[Dω F` (G, K̂)Dω−1 ] over Dω , pointwise across frequency.1 Note that
this evaluation uses the minimizing K̂ from the last step, but that G is not scaled.
The minimization itself produces a new scaling function. Let this new function
be denoted by D̂ω .

(vi) Compare D̂ω with the previous estimate Dω . Stop if they are close, but otherwise
replace Dω with D̂ω and return to step (ii).

With either K or D fixed, the global optimum in the other variable may be found
using the µ and H∞ solutions. Although the joint optimization of D and K is not
convex and the global convergence is not guaranteed, many designs have shown that
this approach works very well (see, e.g., Balas [1990]). In fact, this is probably the
most effective design methodology available today for dealing with such complicated
problems. Detailed treatment of µ analysis is given in Packard and Doyle [1993]. The
rest of this book will focus on the H∞ optimization, which is a fundamental tool for µ
synthesis.

Related MATLAB Commands: musynfit, musynflp, muftbtch, dkit

10.5 Notes and References

This chapter is partially based on the lecture notes given by Doyle [1984] and partially
based on the lecture notes by Packard [1991] and the paper by Doyle, Packard, and
Zhou [1991]. Parts of Section 10.3.3 are based on the paper by Stein and Doyle [1991].
The small µ theorem for systems with nonrational plants and uncertainties is proven
in Tits [1995]. Connections were established in Poolla and Tikku [1995] between the
1 The approximate solutions given in the preceding section may be used.
10.6. Problems 217

frequency-dependent D-scaled upper bounds of the structured singular value and the
robust performance of a system with arbitrarily slowly varying structured linear per-
turbations. Robust performance of systems with structured time-varying perturbations
was also considered in Shamma [1994] using the constant D-scaled upper bounds of the
structured singular value. Other results on µ can be found in Fan and Tits [1986], Fan,
Tits, and Doyle [1991], Packard and Doyle [1993], Packard and Pandey [1993], Young
[1993], and references therein.

10.6 Problems
Problem 10.1 Let M and N be suitably dimensioned matrices and let ∆ be a struc-
tured uncertainty. Prove or disprove

(a) µ∆ (M ) = 0 =⇒ M = 0;

(b) µ∆ (M1 + M2 ) ≤ µ∆ (M1 ) + µ∆ (M2 ).

(c) µ∆ (αM ) = |α|µ∆ (M ).

(d) µ∆ (I) = 1.

(e) µ∆ (M N ) ≤ σ(M )µ∆ (N ).

(f) µ∆ (M N ) ≤ σ(N )µ∆ (M ).

" #
∆1 0
Problem 10.2 Let ∆ = , where ∆i are structured uncertainties. Show
0 ∆2
" #!
M11 M12
that µ∆ = max{µ∆1 (M11 ), µ∆2 (M22 )}.
0 M22

Problem 10.3 Matlab exercise. Let M be a 7 × 7 random real matrix. Take the
perturbation structure to be
  

 δ1 I3 0 0 

 
∆=  0 ∆2 0  : δ1 , δ3 ∈ C, ∆2 ∈ C2×2
.

 

0 0 δ3 I2

Compute µ(M ) and a singularizing perturbation.

" # " #
0 M12 ∆1
Problem 10.4 Let M = be a complex matrix and let ∆ = .
M21 0 ∆2
Show that p
µ∆ (M ) = σ(M12 )σ(M21 ).
218 µ AND µ SYNTHESIS

" # " #
M11 M12 ∆1
Problem 10.5 Let M = be a complex matrix and let ∆ = .
M21 M22 ∆2
Show that p
σ(M12 )σ(M21 ) − max{σ(M11 ), σ(M22 )} ≤ µ∆ (M )
p
≤ σ(M12 )σ(M21 ) + max{σ(M11 ), σ(M22 )}.

Problem 10.6 Let ∆ be all diagonal full blocks and M be partitioned as M = [Mij ],
where Mij are matrices with suitable dimensions. Show that

µ∆ (M ) ≤ π ([kMij k]) (= ρ ([kMij k]))

where π(·) denotes the Perron eigenvalue.

Problem 10.7 Show

inf DM D−1 = ρ(M )
D∈Cn×n p

where k·kp is the induced p-norm, 1 ≤ p ≤ ∞.

Problem 10.8 Let D be a nonsingular diagonal matrix D = diag(d1 , d2 , . . . , dn ). Show

inf DM D−1 p = π(M )
D

if either M = [|mi j|] and 1 ≤ p ≤ ∞ or p = 1 or ∞. Moreover, the optimal D is given

by
1/p 1/q 1/p 1/q
D = diag(y1 /x1 , y2 /x2 , . . . , yn1/p /x1/q
n )

where 1/p + 1/q = 1 and

[|mij |]x = π(M )x, y T [|mij |] = π(M )y T .

Problem 10.9 Let ∆ be a structured uncertainty defined in the book. Suppose M =

xy ∗ with x, y ∈ Cn . Derive an exact expression for µ∆ (M ) in terms of the compo-
nents of x and y. [Note that ∆ = diag(δ1 I, δ2 I, . . . , δm I, ∆1 , . . . , ∆F ) and µ∆ (M ) =
maxU∈U ρ(M U ) = max∆∈B∆ ρ(M ∆).]

Problem 10.10 Let {xk }∞ ∞ ∞

k=0 , {zk }k=0 , and {dk }k=0 be sequences satisfying

2 2 2 2
kxk+1 k + kzk k ≤ β 2 (kxk k + kdk k )

for some β < 1 and all k ≥ 0. If d ∈ `2 , show that both x ∈ `2 and z ∈ `2 and the
norms are bounded by
2 2 2 2
kzk2 + (1 − β 2 ) kxk2 ≤ β 2 kdk2 + kx0 k

Give a system interpretation of this result.

10.6. Problems 219

Problem 10.11 Consider a SISO feedback system shown here with

P = P0 (1 + W1 ∆1 ) + W2 ∆2 , ∆i ∈ RH∞ , k∆i k∞ < 1, i = 1, 2.

Suppose W1 and W2 are stable, and P and P0 have the same number of poles in
Re{s} > 0.

d
?
W3
z
e - K - P -?
e -
− 6

(a) Show that the feedback system is robustly stable if and only if K stabilizes P0
and
k |W1 T | + |W2 KS| k∞ ≤ 1
where
1 P0 K
S= , T = .
1 + P0 K 1 + P0 K

(b) Show that the feedback system has robust performance; that is, kTzd k∞ ≤ 1, if
and only if K stabilizes P0 and

k |W3 S| + |W1 T | + |W2 KS| k∞ ≤ 1.

Problem 10.12 In Problem 10.11, find a matrix

" #
M11 M12
M=
M21 M22

such that " #

∆1
z = Fu (M, ∆)d, ∆s = .
∆2
Assume that K stabilizes P0 . Show that at each frequency

µ∆s (M11 ) = inf σ(Ds M11 Ds−1 ) = |W1 T | + |W2 KS|

Ds ∈Ds
" #
d
with Ds = .
1
220 µ AND µ SYNTHESIS

 
" # d1
∆s  
Next let ∆ = and D =  d2 . Show that
∆p
1

µ∆ (M ) = inf σ(DM D−1 ) = |W3 S| + |W1 T | + |W2 KS|

D∈D

Problem 10.13 Let ∆ = diag(∆1 , . . . , ∆F ) be a structured uncertainty and suppose

M = xy ∗ with x, y ∈ Cn . Let x and y be partitioned compatibly with the ∆:
   
x1 y1
 x2   y2 
x=  
 ...  , y =  ..  .

.
xm+F ym+F
Show that
µ∆ (M ) = inf σ(DM D−1 )
D∈D

and the minimizing di are given by

kyi k kxF k
d2i = .
kxi k kyF k
Problem 10.14 Consider M ∈ C2n×2n to be given. Let ∆2 be a n × n block structure,
and suppose that µ2 (M22 ) < 1. Suppose also that Fl (M, ∆2 ) is invertible for all ∆2 ∈
B∆2 . For each α > 1, find a matrix Wα and a block structure ∆ ˆ such that

max κ(Fl (M, ∆2 )) < α

∆2 ∈B∆2

if and only if
µ∆
ˆ (Wα ) < 1

where κ is the condition number.

Problem 10.15 Let G(s) ∈ RH∞ be an m × m symmetric transfer matrix, i.e.,

GT (s) = G(s), and let ∆(s) ∈ RH∞ be a diagonal perturbation, i.e.,
 δ (s) 
1
 δ2 (s) 
∆(s) = 
 .. .

.
δm (s)
Show (I − G∆)−1 ∈ RH∞ for all k∆k∞ ≤ γ if and only if kGk∞ < 1/γ. [Hint: Note
that for a complex symmetric matrix M = M T ∈ Cm×m , there is a unitary matrix U
and a diagonal matrix Σ = diag(σ1 , σ2 , . . . , σm ) ≥ 0 such that M = U ΣU T , see Horn
and Johnson [1990, page 204]. For detailed discussion of this problem, see Qiu [1995].]
Chapter 11

Controller Parameterization

The basic configuration of the feedback systems considered in this chapter is an LFT ,
as shown in Figure 11.1, where G is the generalized plant with two sets of inputs: the
exogenous inputs w, which include disturbances and commands, and control inputs u.
The plant G also has two sets of outputs: the measured (or sensor) outputs y and the
regulated outputs z. K is the controller to be designed. A control problem in this setup
is either to analyze some specific properties (e.g., stability or performance) of the closed
loop or to design the feedback control K such that the closed-loop system is stable
in some appropriate sense and the error signal z is specified (i.e., some performance
specifications are satisfied). In this chapter we are only concerned with the basic internal
stabilization problems. We will see again that this setup is very convenient for other
general control synthesis problems in the coming chapters.

z w
G

y u
- K

Figure 11.1: General system interconnection

Suppose that a given feedback system is feedback stabilizable. In this chapter, the
problem we are mostly interested in is parameterizing all controllers that stabilize the
system. The parameterization of all internally stabilizing controllers was first introduced
by Youla et al. [1976a, 1976b] using the coprime factorization technique. We shall,
however, focus on the state-space approach in this chapter.

221
222 CONTROLLER PARAMETERIZATION

11.1 Existence of Stabilizing Controllers

Consider a system described by the standard block diagram in Figure 11.1. Assume
that G(s) has a stabilizable and detectable realization of the form
 
" # A B1 B2
G11 (s) G12 (s)  
G(s) = =  C1 D11 D12  . (11.1)
G21 (s) G22 (s)
C2 D21 D22

The stabilization problem is to find feedback mapping K such that the closed-loop
system is internally stable; the well-posedness is required for this interconnection.

Definition 11.1 A proper system G is said to be stabilizable through output feedback

if there exists a proper controller K internally stabilizing G in Figure 11.1. Moreover,
a proper controller K(s) is said to be admissible if it internally stabilizes G.

The following result is standard and follows from Chapter 3.

Lemma 11.1 There exists a proper K achieving internal stability iff (A, B2 ) is stabiliz-
able and (C2 , A) is detectable. Further, let F and L be such that A + B2 F and A + LC2
are stable; then an observer-based stabilizing controller is given by
" #
A + B2 F + LC2 + LD22 F −L
K(s) = .
F 0

Proof. (⇐) By the stabilizability and detectability assumptions, there exist F and
L such that A + B2 F and A + LC2 are stable. Now let K(s) be the observer-based
controller given in the lemma, then the closed-loop A matrix is given by
" #
A B2 F
à = .
−LC2 A + B2 F + LC2

It is easy to check that this matrix is similar to the matrix

" #
A + LC2 0
.
−LC2 A + B2 F

Thus the spectrum of à equals the union of the spectra of A + LC2 and A + B2 F . In
particular, Ã is stable.
(⇒) If (A, B2 ) is not stabilizable or if (C2 , A) is not detectable, then there are some
eigenvalues of à that are fixed in the right-half plane, no matter what the compensator
is. The details are left as an exercise. 2
11.1. Existence of Stabilizing Controllers 223

The stabilizability and detectability conditions of (A, B2 , C2 ) are assumed through-

out the remainder of this chapter.1 It follows that the realization for G22 is stabilizable
and detectable, and these assumptions are enough to yield the following result.

G22 

y u

- K

Figure 11.2: Equivalent stabilization diagram

" #
A B2
Lemma 11.2 Suppose the inherited realization for G22 is stabilizable and
C2 D22
detectable. Then the system in Figure 11.1 is internally stable iff the one in Figure 11.2
is internally stable.
In other words, K(s) internally stabilizes G(s) if and only if it internally stabilizes
G22 [provided that (A, B2 , C2 ) is stabilizable and detectable].

Proof. The necessity follows from the definition. To show the sufficiency, it is sufficient
to show that the system in Figure 11.1 and that in Figure 11.2 share the same A matrix,
which is obvious. 2

From Lemma 11.2, we see that the stabilizing controller for G depends only on G22 .
Hence all stabilizing controllers for G can be obtained by using only G22 .
Remark 11.1 There should be no confusion between a given realization for a transfer
matrix G22 and the inherited realization from G, where G22 is a submatrix. A given
realization for G22 may be stabilizable and detectable while the inherited realization
may not be. For instance,
" #
1 −1 1
G22 = =
s+1 1 0
is a minimal realization but the inherited realization of G22 from
 
−1 0 0 1
" #  
G11 G12  0 1 1 0 
= 

G21 G22  0 1 0 0  
1 0 0 0
1 It should be clear that the stabilizability and detectability of a realization for G do not guarantee

the stabilizability and/or detectability of the corresponding realization for G22 .

224 CONTROLLER PARAMETERIZATION

is  
−1 0 1  
  1
G22 = 0 1 0  = ,
s+1
1 0 0
which is neither stabilizable nor detectable. 3

11.2 Parameterization of All Stabilizing Controllers

Consider again the standard system block diagram in Figure 11.1 with
 
A B1 B2 " #
  G11 (s) G12 (s)
G(s) =  C1 D11 D12 = .
G21 (s) G22 (s)
C2 D21 D22

Suppose (A, B2 ) is stabilizable and (C2 , A) is detectable. In this section we discuss the
following problem:

Given a plant G, parameterize all controllers K that internally stabilize G.

This parameterization for all stabilizing controllers is usually called Youla parameteri-
zation. The parameterization of all stabilizing controllers is easy when the plant itself
is stable.

Theorem 11.3 Suppose G ∈ RH∞ ; then the set of all stabilizing controllers can be
described as
K = Q(I + G22 Q)−1 (11.2)
for any Q ∈ RH∞ and I + D22 Q(∞) nonsingular.

Remark 11.2 This result is very natural considering Corollary 5.3, which says that a
controller K stabilizes a stable plant G22 iff K(I − G22 K)−1 is stable. Now suppose
Q = K(I −G22 K)−1 is a stable transfer matrix, then K can be solved from this equation
which gives exactly the controller parameterization in the preceding theorem. 3

Proof. Note that G22 (s) is stable by the assumptions on G. Then it is straightforward
to verify that the controllers given previously stabilize G22 . On the other hand, suppose
K0 is a stabilizing controller; then Q0 := K0 (I − G22 K0 )−1 ∈ RH∞ , so K0 can be
expressed as K0 = Q0 (I + G22 Q0 )−1 . Note that the invertibility in the last equation is
guaranteed by the well-posedness condition of the interconnected system with controller
K0 since I + D22 Q0 (∞) = (I − D22 K0 (∞))−1 . 2

However, if G is not stable, the parameterization is much more complicated. The

results can be more conveniently stated using state-space representations.
11.2. Parameterization of All Stabilizing Controllers 225

Theorem 11.4 Let F and L be such that A + LC2 and A + B2 F are stable; then all
controllers that internally stabilize G can be parameterized as the transfer matrix from
y to u:

u y  
J A + B2 F + LC2 + LD22 F −L B2 + LD22
  
J = F 0 I 
- Q −(C2 + D22 F ) I −D22

with any Q ∈ RH∞ and I + D22 Q(∞) nonsingular. Furthermore, the set of all closed-
loop transfer matrices from w to z achievable by an internally stabilizing proper con-
troller is equal to

F` (T, Q) = {T11 + T12 QT21 : Q ∈ RH∞ , I + D22 Q(∞) invertible}

where T is given by
 
A + B2 F −B2 F B1 B2
" #  
T11 T12  0 A + LC2 B1 + LD21 0 
T = =
 C +D F
.
T21 T22  1 12 −D12 F D11 D12 

0 C2 D21 0

Proof. Let K = F` (J, Q). Then it is straightforward to verify, by using the state-
space star product formula and some tedious algebra, that F` (G, K) = T11 + T12 QT21
with the T given in the theorem. Hence the controller K = F` (J, Q) for any given
Q ∈ RH∞ does internally stabilize G. Now let K be any stabilizing controller for G;
then F` (J,
ˆ K) ∈ RH∞ , where
 
A −L B2
 
Jˆ =  −F 0 I .
C2 I D22

(Jˆ is stabilized by K since it has the same G22 matrix as G.)

Let Q0 := F` (J, ˆ K) ∈ RH∞ ; then F` (J, Q0 ) = F` (J, F` (J,
ˆ K)) =: F` (Jtmp , K),
where Jtmp can be obtained by using the state-space star product formula given in
Chapter 9:
 
A + LC2 + B2 F + LD22 F −(B2 + LD22 )F −L B2 + LD22
 
 L(C2 + D22 F ) A − LD22 F −L B2 + LD22 
Jtmp =   
−F 
 F 0 I 
−(C2 + D22 F ) C2 + D22 F I 0
226 CONTROLLER PARAMETERIZATION

 
A + LC2 −(B2 + LD22 )F −L B2 + LD22
 
 0 A + B2 F 0 0 
= 



 0 −F 0 I 
0 C2 I 0
" #
0 I
= .
I 0

Hence F` (J, Q0 ) = F` (Jtmp , K) = K. This shows that any stabilizing controller can be
expressed in the form of F` (J, Q0 ) for some Q0 ∈ RH∞ . 2

An important point to note is that the closed-loop transfer matrix is simply an

affine function of the controller parameter matrix Q. The proper K’s achieving internal
stability are precisely those represented in Figure 11.3.

 z  w
G 
y u

D22 

? R
c ?
c C2   c c B2  c
− 66 6
- A

- F
- −L

y1 u1
- Q

Figure 11.3: Structure of stabilizing controllers

It is interesting to note that the system in the dashed box is an observer-based

stabilizing controller for G (or G22 ). Furthermore, it is easy to show that the transfer
function between (y, u1 ) and (u, y1 ) is J; that is,
" # " #
u y
=J .
y1 u1

It is also easy to show that the transfer matrix from u1 to y1 is T22 = 0.

11.2. Parameterization of All Stabilizing Controllers 227

This diagram of the parameterization of all stabilizing controllers also suggests an

interesting interpretation: Every internal stabilization amounts to adding stable dy-
namics to the plant and then stabilizing the extended plant by means of an observer.
The precise statement is as follows: For simplicity of the formulas, only the cases of
strictly proper G22 and K are treated.

Theorem 11.5 Assume that G22 and K are strictly proper and the system in Fig-
ure 11.1 is internally stable. Then G22 can be embedded in a system
" #
Ae Be
Ce 0

where " # " #

A 0 B2 h i
Ae = , Be = , Ce = C2 0 (11.3)
0 Aa 0
and where Aa is stable, such that K has the form
" #
Ae + Be Fe + Le Ce −Le
K= (11.4)
Fe 0

where Ae + Be Fe and Ae + Le Ce are stable.

Proof. K is representable as in Figure 11.3 for some Q in RH∞ . For K to be strictly

proper, Q must be strictly proper. Take a minimal realization of Q:
" #
Aa Ba
Q= .
Ca 0

Since Q ∈ RH∞ , Aa is stable. Let x and xa denote state vectors for J and Q, respec-
tively, and write the equations for the system in Figure 11.3:

ẋ = (A + B2 F + LC2 )x − Ly + B2 u1
u = F x + u1
y1 = −C2 x + y
ẋa = Aa xa + Ba y1
u1 = Ca xa

These equations yield

ẋe = (Ae + Be Fe + Le Ce )xe − Le y

u = Fe xe
228 CONTROLLER PARAMETERIZATION

where " # " #

x h i L
xe := , Fe := F Ca , Le :=
xa −Ba

and where Ae , Be , Ce are as in equation (11.3). 2

1
Example 11.1 Consider a standard feedback system shown in Figure 5.1 with P = .
s−1
We shall find all stabilizing controllers for P such that the steady-state errors with re-
spect to the step input and sin 2t are both zero. It is easy to see that the controller
must provide poles at 0 and ±2j. Now let the set of stabilizing controllers for a mod-
(s + 1)3
ified plant be Km . Then the desired set of controllers is given by
(s − 1)s(s2 + 22 )
(s + 1)3
K= Km .
s(s2 + 22 )

11.3 Coprime Factorization Approach

In this section, all stabilizing controller parameterization will be derived using the con-
ventional coprime factorization approach. Readers should be familiar with the results
presented in Section 5.4 of Chapter 5 before preceding further.

Theorem 11.6 Let G22 = N M −1 = M̃ −1 Ñ be the rcf and lcf of G22 over RH∞ , re-
spectively. Then the set of all proper controllers achieving internal stability is parame-
terized either by

K = (U0 + M Qr )(V0 + N Qr )−1 , det(I + V0−1 N Qr )(∞) 6= 0 (11.5)

for Qr ∈ RH∞ or by

K = (Ṽ0 + Ql Ñ )−1 (Ũ0 + Ql M̃ ), det(I + Ql Ñ Ṽ0−1 )(∞) 6= 0 (11.6)

for Ql ∈ RH∞ , where U0 , V0 , Ũ0 , Ṽ0 ∈ RH∞ satisfy the Bezout identities:

Ṽ0 M − Ũ0 N = I, M̃ V0 − Ñ U0 = I.

Moreover, if U0 , V0 , Ũ0 , and Ṽ0 are chosen such that U0 V0−1 = Ṽ0−1 Ũ0 ; that is,
" #" # " #
Ṽ0 −Ũ0 M U0 I 0
=
−Ñ M̃ N V0 0 I
11.3. Coprime Factorization Approach 229

then

K = (U0 + M Qy )(V0 + N Qy )−1

= (Ṽ0 + Qy Ñ )−1 (Ũ0 + Qy M̃ )
= F` (Jy , Qy ) (11.7)

where " #
U0 V0−1 Ṽ0−1
Jy := (11.8)
V0−1 −V0−1 N
and where Qy ranges over RH∞ such that (I + V0−1 N Qy )(∞) is invertible.

Proof. We shall prove the parameterization given in equation (11.5) first. Assume
that K has the form indicated, and define

U := U0 + M Qr , V := V0 + N Qr .

Then

M̃ V − Ñ U = M̃ (V0 + N Qr ) − Ñ (U0 + M Qr ) = M̃V0 − ÑU0 + (M̃ N − Ñ M )Qr = I.

Thus K achieves internal stability by Lemma 5.7.

Conversely, suppose K is proper and achieves internal stability. Introduce an rcf of
K over RH∞ as K = U V −1 . Then by Lemma 5.7, Z := M̃ V − Ñ U is invertible in
RH∞ . Define Qr by the equation

U0 + M Qr = U Z −1 , (11.9)

so
Qr = M −1 (U Z −1 − U0 ).
Then, using the Bezout identity, we have

V0 + N Qr = V0 + N M −1 (U Z −1 − U0 )
= V0 + M̃ −1 Ñ(U Z −1 − U0 )
= M̃ −1 (M̃ V0 − Ñ U0 + ÑU Z −1 )
= M̃ −1 (I + Ñ U Z −1 )
= M̃ −1 (Z + Ñ U )Z −1
= M̃ −1 M̃ V Z −1
= V Z −1 . (11.10)

Thus,

K = U V −1
= (U0 + M Qr )(V0 + N Qr )−1 .
230 CONTROLLER PARAMETERIZATION

To see that Qr belongs to RH∞ , observe first from equation (11.9) and then from
equation (11.10) that both M Qr and N Qr belong to RH∞ . Then
Qr = (Ṽ0 M − Ũ0 N )Qr = Ṽ0 (M Qr ) − Ũ0 (N Qr ) ∈ RH∞ .
Finally, since V and Z evaluated at s = ∞ are both invertible, so is V0 + N Qr from
equation (11.10), and hence so is I + V0−1 N Qr .
Similarly, the parameterization given in equation (11.6) can be obtained.
To show that the controller can be written in the form of equation (11.7), note that
(U0 + M Qy )(V0 + N Qy )−1 = U0 V0−1 + (M − U0 V0−1 N )Qy (I + V0−1 N Qy )−1 V0−1
and that U0 V0−1 = Ṽ0−1 Ũ0 . We have
(M − U0 V0−1 N ) = (M − Ṽ0−1 Ũ0 N ) = Ṽ0−1 (Ṽ0 M − Ũ0 N ) = Ṽ0−1
and
K = U0 V0−1 + Ṽ0−1 Qy (I + V0−1 N Qy )−1 V0−1 .
2

Corollary 11.7 Given an admissible controller K with coprime factorizations K =

U V −1 = Ṽ −1 Ũ , the free parameter Qy ∈ RH∞ in Youla parameterization is given by
Qy = M −1 (U Z −1 − U0 )
where Z := M̃V − Ñ U .
Next, we shall establish the precise relationship between the preceding all stabilizing
controller parameterization and the state-space parameterization in the last section.
The following theorem follows from some algebraic manipulation.
Theorem 11.8 Let the doubly coprime factorizations of G22 be chosen as
 
" # A + B2 F B2 −L
M U0  
= F I 0 
N V0
C2 + D22 F D22 I
 
" # A + LC2 −(B2 + LD22 ) L
Ṽ0 −Ũ0  
= F I 0 
−Ñ M̃
C2 −D22 I
where F and L are chosen such that A + B2 F and A + LC2 are both stable. Then Jy
can be computed as
 
A + B2 F + LC2 + LD22 F −L B2 + LD22
 
Jy =  F 0 I .
−(C2 + D22 F ) I −D22
11.4. Notes and References 231

Remark 11.3 Note that Jy is exactly the same as the J in Theorem 11.4 and that
K0 := U0 V0−1 is an observer-based stabilizing controller with
" #
A + B2 F + LC2 + LD22 F −L
K0 := .
F 0

11.4 Notes and References

The conventional Youla parameterization can be found in Youla et al. [1976a, 1976b],
Desoer et al. [1980], Doyle [1984], Vidyasagar [1985], and Francis [1987]. The state-
space derivation of all stabilizing controllers was reported in Lu, Zhou, and Doyle [1996].
The paper by Moore et al. [1990] contains some other related interesting results. The
parameterization of all two-degree-of-freedom stabilizing controllers is given in Youla
and Bongiorno [1985] and Vidyasagar [1985].

11.5 Problems
1
Problem 11.1 Let P = . Find the set of all stabilizing controllers K = F` (J, Q).
s−1
Now verify that K0 = −4 is a stabilizing controller and find a Q0 ∈ RH∞ such that
K0 = F` (J, Q0 ).

Problem 11.2 Suppose that {Pi : i = 1, . . . , n} is a set of MIMO plants and that
there is a single controller K that internally stabilizes each Pi in the set. Show that
there exists a single transfer function P such that the set

P = {Fu (P, ∆) | ∆ ∈ H∞ , k∆k∞ ≤ 1 }

is also robustly stabilized by K and that {Pi } ⊂ P.

Problem 11.3 Internal Model Control (IMC): Suppose a plant P is stable. Then
it is known that all stabilizing controllers can be parameterized as K(s) = Q(I − P Q)−1
for all stable Q. In practice, the exact plant model is not known, only a nominal model
P0 is available. Hence the controller can be implemented as in the following diagram:

r -e -e - y
Q - P -
−6 6

P0 
232 CONTROLLER PARAMETERIZATION

The control diagram can be redrawn as follows:

r -e y
- Q - P -
−6

- P0 -e?
−

This control implementation is known as internal model control (IMC). Note that no
signal is fed back if the model is exact. Discuss the advantage of this implementation
and possible generalizations.

Problem 11.4 Use the Youla parameterization (the coprime factor form) to show that
a SISO plant cannot be stabilized by a stable controller if the plant does not satisfy
the parity interlacing properties. [A SISO plant is said to satisfy the parity interlacing
property if the number of unstable real poles between any two unstable real zeros is
even; +∞ counts as a unstable zero if the plant is strictly proper. See Youla, Jabr, and
Lu [1974] and Vidyasagar [1985].]
Chapter 12

Algebraic Riccati Equations

We studied the Lyapunov equation in Chapter 7 and saw the roles it played in some
applications. A more general equation than the Lyapunov equation in control theory is
the so-called algebraic Riccati equation or ARE for short. Roughly speaking, Lyapunov
equations are most useful in system analysis while AREs are most useful in control
system synthesis; particularly, they play central roles in H2 and H∞ optimal control.
Let A, Q, and R be real n × n matrices with Q and R symmetric. Then an algebraic
Riccati equation is the following matrix equation:

A∗ X + XA + XRX + Q = 0. (12.1)

Associated with this Riccati equation is a 2n × 2n matrix:

" #
A R
H := . (12.2)
−Q −A∗

A matrix of this form is called a Hamiltonian matrix. The matrix H in equation (12.2)
will be used to obtain the solutions to the equation (12.1). It is useful to note that
the spectrum of H is symmetric about the imaginary axis. To see that, introduce the
2n × 2n matrix: " #
0 −I
J :=
I 0

having the property J 2 = −I. Then

J −1 HJ = −JHJ = −H ∗

so H and −H ∗ are similar. Thus λ is an eigenvalue iff −λ̄ is.

This chapter is devoted to the study of this algebraic Riccati equation.

233
234 ALGEBRAIC RICCATI EQUATIONS

12.1 Stabilizing Solution and Riccati Operator

Assume that H has no eigenvalues on the imaginary axis. Then it must have n eigen-
values in Re s < 0 and n in Re s > 0. Consider the n-dimensional invariant spectral
subspace, X− (H), corresponding to eigenvalues of H in Re s < 0. By finding a basis for
X− (H), stacking the basis vectors up to form a matrix, and partitioning the matrix, we
get " #
X1
X− (H) = Im
X2
where X1 , X2 ∈ Cn×n . (X1 and X2 can be chosen to be real matrices.) If X1 is
nonsingular or, equivalently, if the two subspaces
" #
0
X− (H), Im (12.3)
I

are complementary, we can set X := X2 X1−1 . Then X is uniquely determined by H

(i.e., H 7−→ X is a function, which will be denoted Ric). We will take the domain of
Ric, denoted dom(Ric), to consist of Hamiltonian matrices H with two properties: H
has no eigenvalues on the imaginary axis and the two subspaces in equation(12.3) are
complementary. For ease of reference, these will be called the stability property and
the complementarity property, respectively. This solution will be called the stabilizing
solution. Thus, X = Ric(H) and

Ric : dom(Ric) ⊂ R2n×2n 7−→ Rn×n .

Remark 12.1 It is now clear that to obtain the stabilizing solution to the Riccati
equation, it is necessary to construct bases for the stable invariant subspace of H.
One way of constructing this invariant subspace is to use eigenvectors and generalized
eigenvectors of H. Suppose λi is an eigenvalue of H with multiplicity k (then λi+j = λi
for all j = 1, . . . , k − 1), and let vi be a corresponding eigenvector and vi+1 , . . . , vi+k−1
be the corresponding generalized eigenvectors associated with vi and λi . Then vj are
related by
(H − λi I)vi = 0
(H − λi I)vi+1 = vi
..
.
(H − λi I)vi+k−1 = vi+k−2 ,
and the span{vj , j = i, . . . , i + k − 1} is an invariant subspace of H. The sum of all
invariant subspaces corresponding to stable eigenvalues is the stable invariant subspace
X− (H). 3
12.1. Stabilizing Solution and Riccati Operator 235

Example 12.1 Let

" # " # " #
−3 2 0 0 0 0
A= R= , Q= .
−2 1 0 −1 0 0

The eigenvalues of H are 1, 1, −1, −1, and the corresponding eigenvectors and general-
ized eigenvectors are
       
1 −1 1 1
       
 2   −3/2   1   3/2 

v1 =        .
 , v2 =  1 
v3 =   , v4 =  
 2     0   0 
−2 0 0 0

It is easy to check that {v3 , v4 } form a basis for the stable invariant subspace X− (H),
{v1 , v2 } form a basis for the antistable invariant subspace, and {v1 , v3 } form a basis for
another invariant subspace corresponding to eigenvalues {1, −1} so
" # " #
−10 6 −2 2
X = 0, X̃ = , X̂ =
6 −4 2 −2

are all solutions of the ARE with the property

λ(A + RX) = {−1, −1}, λ(A + RX̃) = {1, 1}, λ(A + RX̂) = {1, −1}.

Thus only X is the stabilizing solution. The stabilizing solution can be found using the
following Matlab command:
 [X1 , X2 ] = ric schr(H), X = X2 /X1

The following well-known results give some properties of X as well as verifiable

conditions under which H belongs to dom(Ric).

Theorem 12.1 Suppose H ∈ dom(Ric) and X = Ric(H). Then

(i) X is real symmetric;
(ii) X satisfies the algebraic Riccati equation

A∗ X + XA + XRX + Q = 0;

(iii) A + RX is stable.
236 ALGEBRAIC RICCATI EQUATIONS

Proof. (i) Let X1 , X2 be as before. It is claimed that

X1∗ X2 is Hermitian. (12.4)

To prove this, note that there exists a stable matrix H− in Rn×n such that
" # " #
X1 X1
H = H− .
X2 X2

(H− is a matrix representation of H|X− (H) .) Premultiply this equation by

" #∗
X1
J
X2

to get " #∗ " # " #∗ " #

X1 X1 X1 X1
JH = J H− . (12.5)
X2 X2 X2 X2
Since JH is symmetric, so is the left-hand side of equation (12.5) and so is the right-hand
side:
(−X1∗ X2 + X2∗ X1 )H− = H− ∗
(−X1∗ X2 + X2∗ X1 )∗
∗
= −H− (−X1∗ X2 + X2∗ X1 ).
This is a Lyapunov equation. Since H− is stable, the unique solution is

−X1∗ X2 + X2∗ X1 = 0.

This proves equation (12.4). Hence X := X2 X1−1 = (X1−1 )∗ (X1∗ X2 )X1−1 is Hermitian.
Since X1 and X2 can always be chosen to be real and X is unique, X is real symmetric.
(ii) Start with the equation
" # " #
X1 X1
H = H−
X2 X2

and postmultiply by X1−1 to get

" # " #
I I
H = X1 H− X1−1 . (12.6)
X X

Now pre-multiply by [X − I]:

" #
I
[X − I]H = 0.
X
12.1. Stabilizing Solution and Riccati Operator 237

This is precisely the Riccati equation.

(iii) Premultiply equation (12.6) by [I 0] to get

A + RX = X1 H− X1−1 .

Thus A + RX is stable because H− is. 2

Now we are going to state one of the main theorems of this section; it gives the
necessary and sufficient conditions for the existence of a unique stabilizing solution of
equation (12.1) under certain restrictions on the matrix R.

Theorem 12.2 Suppose H has no imaginary eigenvalues and R is either positive semi-
definite or negative semidefinite. Then H ∈ dom(Ric) if and only if (A, R) is stabilizable.

Proof. (⇐) To prove that H ∈ dom(Ric), we must show that

" #
0
X− (H), Im
I

are complementary. This requires a preliminary step. As in the proof of Theorem 12.1
define X1 , X2 , H− so that " #
X1
X− (H) = Im
X2
" # " #
X1 X1
H = H− . (12.7)
X2 X2
We want to show that X1 is nonsingular (i.e., Ker X1 = 0). First, it is claimed that
Ker X1 is H− invariant. To prove this, let x ∈ Ker X1 . Premultiply equation (12.7) by
[I 0] to get
AX1 + RX2 = X1 H− . (12.8)
Premultiply by x∗ X2∗ , postmultiply by x, and use the fact that X2∗ X1 is symmetric [see
equation (12.4)] to get
x∗ X2∗ RX2 x = 0.
Since R is semidefinite, this implies that RX2 x = 0. Now postmultiply equation (12.8)
by x to get X1 H− x = 0 (i.e., H− x ∈ Ker X1 ). This proves the claim.
Now to prove that X1 is nonsingular, suppose, on the contrary, that Ker X1 6= 0.
Then H− |Ker X1 has an eigenvalue, λ, and a corresponding eigenvector, x:

H− x = λx (12.9)

Re λ < 0, 0 6= x ∈ Ker X1 .
238 ALGEBRAIC RICCATI EQUATIONS

Premultiply equation (12.7) by [0 I]:

−QX1 − A∗ X2 = X2 H− . (12.10)

Postmultiply the above equation by x and use equation (12.9):

(A∗ + λI)X2 x = 0.

Recall that RX2 x = 0; we have

x∗ X2∗ [A + λI R] = 0.

Then the stabilizability

" #of (A, R) implies X2 x = 0. But if both X1 x = 0 and X2 x = 0,
X1
then x = 0 since has full column rank, which is a contradiction.
X2
(⇒) This is obvious since H ∈ dom(Ric) implies that X is a stabilizing solution and
that A + RX is asymptotically stable. It also implies that (A, R) must be stabilizable.
2

The following result is the so-called bounded real lemma, which follows immediately
from the preceding theorem.
" #
A B
Corollary 12.3 Let γ > 0, G(s) = ∈ RH∞ , and
C D
" #
A + BR−1 D∗ C BR−1 B ∗
H :=
−C ∗ (I + DR−1 D∗ )C −(A + BR−1 D∗ C)∗

where R = γ 2 I − D∗ D. Then the following conditions are equivalent:

(i) kGk∞ < γ.
(ii) σ̄(D) < γ and H has no eigenvalues on the imaginary axis.
(iii) σ̄(D) < γ and H ∈ dom(Ric).
(iv) σ̄(D) < γ and H ∈ dom(Ric) and Ric(H) ≥ 0 (Ric(H) > 0 if (C, A) is observable).
(v) σ̄(D) < γ and there exists an X = X ∗ ≥ 0 such that

X(A+BR−1 D∗ C)+(A+BR−1 D∗ C)∗ X +XBR−1 B ∗ X +C ∗ (I +DR−1 D∗ )C = 0

and A + BR−1 D∗ C + BR−1 B ∗ X has no eigenvalues on the imaginary axis.

(vi) σ̄(D) < γ and there exists an X = X ∗ > 0 such that

X(A+BR−1 D∗ C)+(A+BR−1 D∗ C)∗ X +XBR−1 B ∗ X +C ∗ (I +DR−1 D∗ )C < 0.

12.1. Stabilizing Solution and Riccati Operator 239

(vii) There exists an X = X ∗ > 0 such that

 
XA + A∗ X XB C∗
 
 B∗X −γI D∗  < 0.
C D −γI

Proof. The equivalence between (i) and (ii) has been shown in Chapter 4. The equiv-
alence between (iii) and (iv) is obvious by noting the fact that A + BR−1 D∗ C is stable
if kGk∞ < γ (See Problem 12.15). The equivalence between (ii) and (iii) follows from
the preceding theorem. It is also obvious that (iv) implies (v). We shall now show that
(v) implies (i). Thus suppose that there is an X ≥ 0 such that
X(A + BR−1 D∗ C) + (A + BR−1 D∗ C)∗ X + XBR−1 B ∗ X + C ∗ (I + DR−1 D∗ )C = 0
and A + BR−1 (B ∗ X + D∗ C) has no eigenvalues on the imaginary axis. Then
" #
A −B
W (s) :=
B ∗ X + D∗ C R

has no zeros on the imaginary axis since

" #
−1 A + BR−1 (B ∗ X + D∗ C) BR−1
W (s) =
R−1 (B ∗ X + D∗ C) R−1

has no poles on the imaginary axis. Next, note that

−X(jωI − A) − (jωI − A)∗ X + XBR−1 D∗ C + C ∗ DR−1 B ∗ X
+XBR−1 B ∗ X + C ∗ (I + DR−1 D∗ )C = 0.
Multiplying B ∗ {(jωI−A)∗ }−1 on the left and (jωI−A)−1 B on the right of the preceding
equation and completing square, we have
G∗ (jω)G(jω) = γ 2 I − W ∗ (jω)R−1 W (jω).
Since W (s) has no zeros on the imaginary axis, we conclude that kGk∞ < γ.
The equivalence between (vi) and (vii) follows from Schur complement. It is also
easy to show that (vi) implies (i) by following the similar procedure as above. To show
that (i) implies (vi), let
 
A B
 
Ĝ =  C D  .
I 0


Then there exists an  > 0 such that Ĝ < γ. Now (vi) follows by applying part (v)
∞
to Ĝ. 2
240 ALGEBRAIC RICCATI EQUATIONS

Theorem 12.4 Suppose H has the form

" #
A −BB ∗
H= .
−C ∗ C −A∗

Then H ∈ dom(Ric) iff (A, B) is stabilizable and (C, A) has no unobservable modes on
the imaginary axis. Furthermore, X = Ric(H) ≥ 0 if H ∈ dom(Ric), and Ker(X) = {0}
if and only if (C, A) has no stable unobservable modes.

Note that Ker(X) ⊂ Ker(C), so that the equation XM = C ∗ always has a solution
for M , and a minimum F -norm solution is given by X + C ∗ .

Proof. It is clear from Theorem 12.2 that the stabilizability of (A, B) is necessary,
and it is also sufficient if H has no eigenvalues on the imaginary axis. So we only need
to show that, assuming (A, B) is stabilizable, H has no imaginary eigenvalues iff (C, A)
has no unobservable
" # modes on the imaginary axis. Suppose that jω is an eigenvalue
x
and 0 6= is a corresponding eigenvector. Then
z

Ax − BB ∗ z = jωx

−C ∗ Cx − A∗ z = jωz.
Rearrange:
(A − jωI)x = BB ∗ z (12.11)
−(A − jωI)∗ z = C ∗ Cx. (12.12)
Thus
hz, (A − jωI)xi = hz, BB ∗ zi = kB ∗ zk2
−hx, (A − jωI)∗ zi = hx, C ∗ Cxi = kCxk2
so hx, (A − jωI)∗ zi is real and

−kCxk2 = h(A − jωI)x, zi = hz, (A − jωI)xi = kB ∗ zk2 .

Therefore, B ∗ z = 0 and Cx = 0. So from equations (12.11) and (12.12)

(A − jωI)x = 0

(A − jωI)∗ z = 0.
Combine the last four equations to get

z ∗ [A − jωI B] = 0
12.1. Stabilizing Solution and Riccati Operator 241

" #
A − jωI
x = 0.
C
The stabilizability of (A, B) gives z = 0. Now it is clear that jω is an eigenvalue of H
iff jω is an unobservable mode of (C, A).
Next, set X := Ric(H). We will show that X ≥ 0. The Riccati equation is

A∗ X + XA − XBB ∗ X + C ∗ C = 0

or, equivalently,

(A − BB ∗ X)∗ X + X(A − BB ∗ X) + XBB ∗ X + C ∗ C = 0. (12.13)

Noting that A − BB ∗ X is stable (Theorem 12.1), we have

Z ∞
∗ ∗ ∗
X= e(A−BB X) t (XBB ∗ X + C ∗ C)e(A−BB X)t dt. (12.14)
0

Since XBB ∗ X + C ∗ C is positive semidefinite, so is X.

Finally, we will show that KerX is nontrivial if and only if (C, A) has stable unob-
servable modes. Let x ∈ KerX, then Xx = 0. Premultiply equation (12.13) by x∗ and
postmultiply by x to get
Cx = 0.
Now postmultiply equation (12.13) again by x to get

XAx = 0.

We conclude that Ker(X) is an A-invariant subspace. Thus if Ker(X) 6= {0}, then

there is a 0 6= x ∈ Ker(X) and a λ such that λx = Ax = (A − BB ∗ X)x and Cx =
0. Since (A − BB ∗ X) is stable, Reλ < 0; thus λ is a stable unobservable mode.
Conversely, suppose (C, A) has an unobservable stable mode λ (i.e., there is an x such
that Ax = λx, Cx = 0). By premultiplying the Riccati equation by x∗ and post-
multiplying by x, we get

2Reλx∗ Xx − x∗ XBB ∗ Xx = 0.

Hence x∗ Xx = 0 (i.e., X is singular) since Reλ < 0. 2

Example 12.2 This example shows that the observability of (C, A) is not necessary
for the existence of a positive definite stabilizing solution. Let
" # " #
1 0 1 h i
A= , B= , C= 0 0 .
0 2 1
242 ALGEBRAIC RICCATI EQUATIONS

Then (A, B) is stabilizable, but (C, A) is not detectable. However,

" #
18 −24
X= >0
−24 36
is the stabilizing solution.

Corollary 12.5 Suppose that (A, B) is stabilizable and (C, A) is detectable. Then the
Riccati equation
A∗ X + XA − XBB ∗ X + C ∗ C = 0
has a unique positive semidefinite solution. Moreover, the solution is stabilizing.

Proof. It is obvious from the preceding theorem that the Riccati equation has a unique
stabilizing solution and that the solution is positive semidefinite. Hence we only need
to show that any positive semidefinite solution X ≥ 0 must also be stabilizing. Then
by the uniqueness of the stabilizing solution, we can conclude that there is only one
positive semidefinite solution. To achieve that goal, let us assume that X ≥ 0 satisfies
the Riccati equation but that it is not stabilizing. First rewrite the Riccati equation as
(A − BB ∗ X)∗ X + X(A − BB ∗ X) + XBB ∗ X + C ∗ C = 0 (12.15)
and let λ and x be an unstable eigenvalue and the corresponding eigenvector of
A − BB ∗ X, respectively; that is,
(A − BB ∗ X)x = λx.
Now premultiply and postmultiply equation (12.15) by x∗ and x, respectively, and we
have
(λ̄ + λ)x∗ Xx + x∗ (XBB ∗ X + C ∗ C)x = 0.
This implies
B ∗ Xx = 0, Cx = 0
since Re(λ) ≥ 0 and X ≥ 0. Finally, we arrive at
Ax = λx, Cx = 0.
That is, (C, A) is not detectable, which is a contradiction. Hence Re(λ) < 0 (i.e., X ≥ 0
is the stabilizing solution). 2

Lemma 12.6 Suppose D has full column rank and let R = D∗ D > 0; then the following
statements are equivalent:
" #
A − jωI B
(i) has full column rank for all ω.
C D


(ii) (I − DR−1 D∗ )C, A − BR−1 D∗ C has no unobservable modes on the jω axis.
12.1. Stabilizing Solution and Riccati Operator 243

Proof. Suppose jω is an unobservable mode of (I − DR−1 D∗ )C, A − BR−1 D∗ C ;
then there is an x 6= 0 such that

(A − BR−1 D∗ C)x = jωx, (I − DR−1 D∗ )Cx = 0;

that is, " #" #" #

A − jωI B I 0 x
= 0.
C D −R−1 D∗ C I 0
But this implies that " #
A − jωI B
(12.16)
C D
does not have full-column rank. Conversely, suppose
" # equation (12.16) does not have
u
full-column rank for some ω; then there exists 6= 0 such that
v
" #" #
A − jωI B u
= 0.
C D v

Now let " # " #" #

u I 0 x
= .
v −R−1 D∗ C I y
Then " # " #" #
x I 0 u
= −1 ∗
6= 0
y R D C I v
and
(A − BR−1 D∗ C − jωI)x + By = 0 (12.17)
−1 ∗
(I − DR D )Cx + Dy = 0. (12.18)
Premultiply equation (12.18) by D∗ to get y = 0. Then we have

(A − BR−1 D∗ C)x = jωx, (I − DR−1 D∗ )Cx = 0;

that is, jω is an unobservable mode of (I − DR−1 D∗ )C, A − BR−1 D∗ C . 2

h i
Remark 12.2 If D is not square, then there is a D⊥ such that D⊥ DR−1/2 is
∗ −1 ∗
unitary and that D⊥ D⊥ = I−DR D . Hence, in some cases we will write the condition
∗
(ii) in the preceding lemma as (D⊥ C, A−BR−1 D∗ C) having no imaginary unobservable
modes. Of course, if D is square, the condition is simplified to A − BR−1 D∗ C having no
imaginary eigenvalues. Note also that if D∗ C = 0, condition (ii) becomes (C, A) having
no imaginary unobservable modes. 3
244 ALGEBRAIC RICCATI EQUATIONS

Corollary 12.7 Suppose D has full column rank and denote R = D∗ D > 0. Let H
have the form
" # " #
A 0 B h i
−1 ∗ ∗
H = − R D C B
−C ∗ C −A∗ −C ∗ D
" #
A − BR−1 D∗ C −BR−1 B ∗
= .
−C ∗ (I − DR−1 D∗ )C −(A − BR−1 D∗ C)∗
" #
A − jωI B
Then H ∈ dom(Ric) iff (A, B) is stabilizable and has full-column rank
C D
for all ω. Furthermore, X = Ric(H) ≥ 0 if H ∈ dom(Ric), and Ker(X) = {0} if and
∗
only if (D⊥ C, A − BR−1 D∗ C) has no stable unobservable modes.

Proof. This is the consequence of Lemma 12.6 and Theorem 12.4. 2

Remark 12.3 It is easy to see that the detectability (observability) of

∗
(D⊥ C, A − BR−1 D∗ C) implies the detectability (observability) of (C, A); however, the
converse is, in general, not true. Hence the existence of a stabilizing solution to the
Riccati equation in the preceding corollary is not guaranteed by the stabilizability of
(A, B) and detectability of (C, A). Furthermore, even if a stabilizing solution exists,
the positive definiteness of the solution is not guaranteed by the observability of (C, A)
unless D∗ C = 0. As an example, consider
" # " # " # " #
0 1 0 1 0 1
A= , B= , C= , D= .
0 0 −1 0 0 0
Then (C, A) is observable, (A, B) is controllable, and
" #
0 1 h i
∗ ∗
A − BD C = , D⊥ C= 0 0 .
1 0
A Riccati equation with the preceding data has a nonnegative definite stabilizing so-
∗
lution since (D⊥ C, A − BR−1 D∗ C) has no unobservable modes on the imaginary axis.
∗
However, the solution is not positive definite since (D⊥ C, A − BR−1 D∗ C) has a stable
unobservable mode. On the other hand, if the B matrix is changed to
" #
0
B= ,
1
then the corresponding Riccati equation has no stabilizing solution since, in this case,
(A − BD∗ C) has eigenvalues on the imaginary axis although (A, B) is controllable and
(C, A) is observable. 3

Related MATLAB Commands: ric eig, are

12.2. Inner Functions 245

12.2 Inner Functions

A transfer function N is called inner if N ∈ RH∞ and N ∼ N = I and co-inner if
N ∈ RH∞ and N N ∼ = I. Note that N need not be square. Inner and co-inner are
dual notions (i.e., N is an inner iff N T is a co-inner). A matrix function N ∈ RL∞ is
called all-pass if N is square and N ∼ N = I; clearly a square inner function is all-pass.
We will focus on the characterizations of inner functions here, and the properties of
co-inner functions follow by duality.
Note that N inner implies that N has at least as many rows as columns. For
N inner and any q ∈ Cm , v ∈ L2 , kN (jω)qk = kqk, ∀ω and kN vk2 = kvk2 since
N (jω)∗ N (jω) = I for all ω. Because of these norm preserving properties, inner matrices
will play an important role in the control synthesis theory in this book. In this section,
we present a state-space characterization of inner transfer functions.
" #
A B
Lemma 12.8 Suppose N = ∈ RH∞ and X = X ∗ ≥ 0 satisfies
C D

A∗ X + XA + C ∗ C = 0. (12.19)

Then
(a) D∗ C + B ∗ X = 0 implies N ∼ N = D∗ D.
(b) (A, B) is controllable, and N ∼ N = D∗ D implies that D∗ C + B ∗ X = 0.

Proof. Conjugating the states of

 
A 0 B
 

N N =  −C ∗ C
∼
−A∗ −C ∗ D 

D∗ C B∗ D∗ D
" # " #−1 " #
I 0 I 0 I 0
by on the left and = on the right yields
−X I −X I X I
 
A 0 B
 
N ∼N =  ∗ ∗ ∗
 −(A X + XA + C C) −A −(XB + C D) 
∗ 
B ∗ X + D∗ C B∗ D∗ D
 
A 0 B
 
= 
 0 −A∗ −(XB + C ∗ D) 
.
B ∗ X + D∗ C B∗ D∗ D

Then (a) and (b) follow easily. 2

246 ALGEBRAIC RICCATI EQUATIONS

This lemma immediately leads to one characterization of inner matrices in terms

of their state-space representations. Simply add the condition that D∗ D = I to
Lemma 12.8 to get N ∼ N = I.
" #
A B
Corollary 12.9 Suppose N = is stable and minimal, and X is the observ-
C D
ability Gramian. Then N is an inner if and only if

(a) D∗ C + B ∗ X = 0

(b) D∗ D = I.

A transfer matrix N⊥ is called a complementary inner factor (CIF) of N if [N N⊥ ]

is square and is an inner. The dual notion of the complementary co-inner factor is
defined in the obvious way. Given an inner N , the following lemma gives a construction
of its CIF. The proof of this lemma follows from straightforward calculation and from
the fact that CX + X = C since Im(I − X + X) ⊂ Ker(X) ⊂ Ker(C).
" #
A B
Lemma 12.10 Let N = be an inner and X be the observability Gramian.
C D
Then a CIF N⊥ is given by
" #
A −X + C ∗ D⊥
N⊥ =
C D⊥

where D⊥ is an orthogonal complement of D such that [D D⊥ ] is square and orthogonal.

12.3 Notes and References

The general solutions of a Riccati equation are given by Martensson [1971]. The paper by
Willems [1971] contains a comprehensive treatment of ARE and the related optimization
problems. Some matrix factorization results are given in Doyle [1984]. Numerical
methods for solving ARE can be found in Arnold and Laub [1984], Van Dooren [1981],
and references therein. See Zhou, Doyle, and Glover [1996] and Lancaster and Rodman
[1995] for a more extensive treatment of this subject.

12.4 Problems
" #
A B
Problem 12.1 Assume that G(s) := ∈ RL∞ is a stabilizable and de-
C D
tectable realization and γ > kG(s)k∞ . Show that there exists a transfer matrix M ∈
12.4. Problems 247

RL∞ such that M ∼ M = γ 2 I − G∼ G and M −1 ∈ RH∞ . A particular realization of M

is " #
A B
M (s) =
−R1/2 F R1/2
where

R = γ 2 I − D∗ D
F = R−1 (B ∗ X + D∗ C)
" #
A + BR−1 D∗ C BR−1 B ∗
X = Ric
−C ∗ (I + DR−1 D∗ )C −(A + BR−1 D∗ C)∗

and X ≥ 0 if A is stable.
" #
A B
Problem 12.2 Let G(s) = be a stabilizable and detectable realization.
C D
" #
∼ A − jω B
Suppose G (jω)G(jω) > 0 for all ω or has full-column rank for all ω.
C D
Let " #
A − BR−1 D∗ C −BR−1 B ∗
X = Ric
−C ∗ (I − DR−1 D∗ )C −(A − BR−1 D∗ C)∗
with R := D∗ D > 0. Show
W ∼ W = G∼ G
where W −1 ∈ RH∞ and
" #
A B
W = .
R−1/2 (D∗ C + B ∗ X) R1/2

Problem 12.3 A square (m × m) matrix function G(s) ∈ RH∞ is said to be positive

real (PR) if G(jω) + G∗ (jω) ≥ 0 for all finite ω; and G(s) "is said to#be strictly positive
A B
real (SPR) if G(jω) + G∗ (jω) > 0 for all ω ∈ R. Let be a state-space
C D
realization of G(s) with A stable (not necessarily a minimal realization). Suppose there
exist X ≥ 0, Q, and W such that

XA + A∗ X = −Q∗ Q (12.20)
B∗X + W ∗Q = C (12.21)
D + D∗ = W ∗ W, (12.22)

Show that G(s) is positive real and

G(s) + G∼ (s) = M ∼ (s)M (s)

248 ALGEBRAIC RICCATI EQUATIONS

" #
A B
with M (s) = . Furthermore, if M (jω) has full-column rank for all ω ∈ R,
Q W
then G(s) is strictly positive real.

Problem 12.4 Suppose (A, B, C, D) is a minimal realization of G(s) with A stable and
G(s) positive real. Show that there exist X ≥ 0, Q, and W such that

XA + A∗ X = −Q∗ Q
B∗X + W ∗Q = C
D + D∗ = W ∗ W

and
G(s) + G∼ (s) = M ∼ (s)M (s)
" #
A B
with M (s) = . Furthermore, if G(s) is strictly positive real, then M (jω) has
Q W
full-column rank for all ω ∈ R.
" #
A B
Problem 12.5 Let be a state-space realization of G(s) ∈ RH∞ with A
C D
stable and R := D + D∗ > 0. Show that G(s) is strictly positive real if and only if there
exists a stabilizing solution to the following Riccati equation:

X(A − BR−1 C) + (A − BR−1 C)∗ X + XBR−1 B ∗ X + C ∗ R−1 C = 0.

" #
A B
Moreover, M (s) = is minimal phase and
R− 2 (C − B ∗ X) R 2
1 1

G(s) + G∼ (s) = M ∼ (s)M (s).

Problem 12.6 Assume p ≥ m. Show that there exists an rcf G = N M −1 such that N
is an inner if and only if G∼ G > 0 on the jω axis, including at ∞. This factorization is
unique" up to a#constant unitary multiple." Furthermore, assume
# that the realization of
A B A − jωI B
G= is stabilizable and that has full column rank for all
C D C D
ω ∈ R. Then a particular realization of the desired coprime factorization is
 
" # A + BF BR−1/2
M  
:=  F R−1/2  ∈ RH∞
N
C + DF DR−1/2

where
R = D∗ D > 0
12.4. Problems 249

F = −R−1 (B ∗ X + D∗ C)
and " #
A − BR−1 D∗ C −BR−1 B ∗
X = Ric ≥ 0.
−C ∗ (I − DR−1 D∗ )C −(A − BR−1 D∗ C)∗
Moreover, a complementary inner factor can be obtained as
" #
A + BF −X † C ∗ D⊥
N⊥ =
C + DF D⊥

if p > m.
" #
A B
Problem 12.7 Assume that G = ∈ Rp (s) and (A, B) is stabilizable. Show
C D
that there exists a right coprime factorization G = N M −1 such that M ∈ RH∞ is an
inner if and only if G has no poles on the jω axis. A particular realization is
 
" # A + BF B
M  
:=  F I  ∈ RH∞
N
C + DF D

where
F = −B ∗ X
" #
A −BB ∗
X = Ric ≥ 0.
0 −A∗

Problem 12.8 A right coprime factorization of G = N M −1 with N, M ∈ RH " ∞ is #

∼ ∼ M
called a normalized right coprime factorization if M M + N N = I; that is, if
N
is an inner. Similarly, an lcf G = M̃ −1 Ñ is called a normalized left "coprime factorization
#
h i A B
if M̃ Ñ is a co-inner. Let a realization of G be given by G = and define
C D
R = I + D∗ D > 0 and R̃ = I + DD∗ > 0.

(a) Suppose (A, B) is stabilizable and (C, A) has no unobservable modes on the imagi-
nary axis. Show that there is a normalized right coprime factorization G = N M −1
 
" # A + BF BR−1/2
M  
:=  F R−1/2  ∈ RH∞
N
C + DF DR−1/2
250 ALGEBRAIC RICCATI EQUATIONS

where
F = −R−1 (B ∗ X + D∗ C)
and " #
A − BR−1 D∗ C −BR−1 B ∗
X = Ric ≥ 0.
−C ∗ R̃−1 C −(A − BR−1 D∗ C)∗

(b) Suppose (C, A) is detectable and (A, B) has no uncontrollable modes on the imag-
inary axis. Show that there is a normalized left coprime factorization G = M̃ −1 Ñ
" #
h i A + LC L B + LD
M̃ Ñ :=
R̃−1/2 C R̃−1/2 R̃−1/2 D

where
L = −(BD∗ + Y C ∗ )R̃−1
and " #
(A − BD∗ R̃−1 C)∗ −C ∗ R̃−1 C
Y = Ric ≥ 0.
−BR−1 B ∗ −(A − BD∗ R̃−1 C)

(c) Show
" #that the controllability Gramian P and the observability Gramian Q of
M
are given by
N
P = (I + Y X)−1 Y, Q=X
h i
while the controllability Gramian P̃ and observability Gramian Q̃ of M̃ Ñ
are given by
P̃ = Y, Q̃ = (I + XY )−1 X.
" #
A B
Problem 12.9 Let G(s) = . Find M1 and M2 such that M1−1 , M2−1 ∈ RH∞
C D
and
M1 M1∼ = I + GG∼ , M2∼ M2 = I + G∼ G.

Problem 12.10 Let A ∈ Rm×m , B ∈ Rn×n , C ∈ Rm×n , and consider the Sylvester
equation
AX + XB = C
for an unknown matrix X ∈ Rm×n . Let
" # " #
B 0 B 0
M= , N= .
C −A 0 −A
12.4. Problems 251

" #
U
1. Let the columns of ∈ Cn+m×n be the eigenvectors of M associated with
V
the eigenvalues of B and suppose U is nonsingular. Show that

X = V U −1

solves the Sylvester equation. Moreover, every solution of the Sylvester equation
can be written in the above form.

2. Show that the Sylvester equation has a solution if and only if M and N are similar.
(See Lancaster and Tismenetsky [1985, page 423].)

Problem 12.11 Let A ∈ Rn×n . Show that

Z t
∗
P (t) = eA τ QeAτ dτ
0

satisfies
Ṗ (t) = A∗ P (t) + P (t)A + Q, P (0) = 0.

Problem 12.12 A more general case of the above problem is when the given matrices
are time varying and the initial condition is not zero. Let A(t), Q(t), P0 ∈ Rn×n . Show
that Z t
P (t) = ΦT (t, t0 )P0 Φ(t, t0 ) + ΦT (t, τ )Q(τ )Φ(t, τ )dτ
t0

satisfies
Ṗ (t) = A∗ P (t) + P (t)A + Q(t), P (t0 ) = P0
where Φ(t, τ ) is the state transition matrix for the system ẋ = A(t)x.

Problem 12.13 Let A ∈ Rn×n , R = R∗ , Q = Q∗ . Define

" #
A R
H= .
−Q −A∗

Let " #
Θ11 (t) Θ12 (t)
Θ(t) = = eHt .
Θ21 (t) Θ22 (t)
Show that
P (t) = (Θ21 (t) + Θ22 P0 )(Θ11 (t) + Θ12 (t)P0 )−1
is the solution to the following differential Riccati equation:

−Ṗ (t) = A∗ P (t) + P (t)A + P RP + Q, P (0) = P0 .

252 ALGEBRAIC RICCATI EQUATIONS

Problem 12.14 Let A ∈ Rn×n , R = R∗ , Q = Q∗ . Define

" #
A R
H= .
−Q −A∗

Let " #
Θ11 (t) Θ12 (t)
Θ(t) = = eH(t−T ) .
Θ21 (t) Θ22 (t)
Show that
P (t) = Θ21 (t)Θ−1
11 (t)

is the solution to the following differential Riccati equation:

−Ṗ (t) = A∗ P (t) + P (t)A + P RP + Q, P (T ) = 0.

" #
A B
Problem 12.15 Suppose G(s) = ∈ RH∞ and kGk∞ < γ. Show that
C D
A + BR−1 D∗ C with R = γ 2 I − D∗ D is stable. (Hint: Show A + B(I − ∆D/γ)−1 ∆C/γ
is stable for all ∆ with k∆k ≤ 1.)
Chapter 13

H2 Optimal Control

In this chapter we treat the optimal control of linear time-invariant systems with a
quadratic performance criterion.

13.1 Introduction to Regulator Problem

Consider the following dynamical system:
ẋ = Ax + B2 u, x(t0 ) = x0 (13.1)
where x0 is given but arbitrary. Our objective is to find a control function u(t) defined
on [t0 , T ], which can be a function of the state x(t), such that the state x(t) is driven to
a (small) neighborhood of origin at time T . This is the so-called regulator problem. One
might suggest that this regulator problem can be trivially solved for any T > t0 if the
system is controllable. This is indeed the case if the controller can provide arbitrarily
large amount of energy since, by the definition of controllability, one can immediately
construct a control function that will drive the state to zero in an arbitrarily short
time. However, this is not practical since any physical system has energy limitation
(i.e., the actuator will eventually saturate). Furthermore, large control action can easily
drive the system out of the region, where the given linear model is valid. Hence certain
limitations have to be imposed on the control in practical engineering implementation.
The constraints on control u may be measured in many different ways; for example,
Z T Z T
2
kuk dt, kuk dt, sup kuk ;
t0 t0 t∈[t0 ,T ]

That is, in terms of L1 norm, L2 norm, and L∞ norm or, more generally, weighted L1
norm, L2 norm, and L∞ norm
Z T Z T
kWu uk dt, kWu uk2 dt, sup kWu uk
t0 t0 t∈[t0 ,T ]

253
254 H2 OPTIMAL CONTROL

for some constant weighting matrix Wu .

Similarly, one might also want to impose some constraints on the transient response
x(t) in a similar fashion:
Z T Z T
2
kWx xk dt, kWx xk dt, sup kWx xk
t0 t0 t∈[t0 ,T ]

for some weighting matrix Wx . Hence the regulator problem can be posed as an optimal
control problem with certain combined performance index on u and x. In this chap-
ter, we shall be concerned exclusively with the L2 performance problem or quadratic
performance problem. Moreover, we shall focus on the infinite time regulator problem
(i.e., T → ∞) and, without loss of generality, we shall assume t0 = 0. In this case, our
problem is as follows: Find a control u(t) defined on [0, ∞) such that the state x(t) is
driven to the origin as t → ∞ and the following performance index is minimized:
Z ∞" #∗ " #" #
x(t) Q S x(t)
min dt (13.2)
u 0 u(t) S∗ R u(t)

for some Q = Q∗ , S, and R = R∗ > 0. This problem is traditionally called a linear

quadratic regulator problem or, simply, an LQR problem. Here we have assumed R > 0
to emphasize that the control energy has to be finite (i.e., u(t) ∈ L2 [0, ∞)). So L2 [0, ∞)
is the space over which the integral is minimized. Moreover, it is also generally assumed
that " #
Q S
≥ 0. (13.3)
S∗ R

Since R is positive definite, it has a square root, R1/2 , which is also positive-definite.
By the substitution
u ← R1/2 u,
we may as well assume at the start that R = I. In fact, we can even assume S = 0
by using a pre-state feedback u = −S ∗ x + v provided some care is exercised; however,
this will not be assumed in the sequel. Since the matrix in equation (13.3) is positive
semi-definite with R = I, it can be factored as
" # " #
Q S C1∗ h i
= C1 D12 .
S∗ I ∗
D12

Then equation (13.2) can be rewritten as

2
min kC1 x + D12 uk2 .
u∈L2 [0,∞)
13.2. Standard LQR Problem 255

In fact, the LQR problem is posed traditionally as the minimization problem:

min kC1 x + D12 uk22 (13.4)

u∈L2 [0,∞)

subject to: ẋ = Ax + Bu, x(0) = x0 (13.5)

without explicitly mentioning the condition that the control should drive the state to the
origin. Instead some assumptions are imposed on Q, S, and R (or, equivalently, on C1
and D12 ) to ensure that the optimal control law u has this property. To see what assump-
tion one needs to make to ensure that the minimization problem formulated in equations
(13.4) and (13.5) has a sensible solution, let us consider a simple example with A = 1,
B = 1, Q = 0, S = 0, and R = 1:
Z ∞
min u2 dt, ẋ = x + u, x(0) = x0 .
u∈L2 [0,∞) 0

It is clear that u = 0 is the optimal solution. However, the system with u = 0 is

unstable and x(t) diverges exponentially to infinity since x(t) = et x0 . The problem
with this example is that this performance index does not “see” the unstable state x.
Thus, to ensure that the minimization problem in equations (13.4) and (13.5) is sensible,
we must assume that all unstable states can be “seen” from the performance index; that
is, (C1 , A) must be detectable. An LQR problem with such an assumption will be called
a standard LQR problem.
On the other hand, if the closed-loop stabilityR is imposed on the preceding mini-
∞
mization, then it can be shown that minu∈L2 [0,∞) 0 u2 dt = 2x20 and u(t) = −2x(t) is
the optimal control. This can also be generalized to a more general case where (C1 , A)
is not necessarily detectable. Such a LQR problem will be referred to as an Extended
LQR problem.

13.2 Standard LQR Problem

In this section, we shall consider the LQR problem as traditionally formulated.

Standard LQR Problem

Let a dynamical system be described by

ẋ = Ax + B2 u, x(0) = x0 given but arbitrary (13.6)

z = C1 x + D12 u (13.7)

and suppose that the system parameter matrices satisfy the following as-
sumptions:

(A1) (A, B2 ) is stabilizable;

256 H2 OPTIMAL CONTROL

h i
(A2) D12 has full column rank with D12 D⊥ unitary;
(A3) (C1 , A) is detectable;
" #
A − jωI B2
(A4) has full column rank for all ω.
C1 D12

Find an optimal control law u ∈ L2 [0, ∞) such that the performance criterion
2
kzk2 is minimized.

Remark 13.1 Assumption (A1) is clearly necessary for the existence of a stabilizing
control function u. The assumption (A2) is made for simplicity of notation and is
∗
actually a restatement that R = D12 D12 = I. Note also that D⊥ drops out when D12
is square. It is interesting to point out that (A3) is not needed in the Extended LQR
problem. The assumption (A3) enforces that the unconditional optimization problem
will result in a stabilizing control law. In fact, the assumption (A3) together with (A1)
guarantees that the input/output stability implies the internal stability; that is, u ∈ L2
and z ∈ L2 imply x ∈ L2 , which will be shown in Lemma 13.1. Finally note that (A4)
∗ ∗
is equivalent to the condition that (D⊥ C1 , A − B2 D12 C1 ) has no unobservable modes
on the imaginary axis and is weaker than the popular assumption of detectability of
∗ ∗
(D⊥ C1 , A − B2 D12 C1 ). (A4), together with the stabilizability of (A, B2 ), guarantees
by Corollary 12.7 that the following Hamiltonian matrix belongs to dom(Ric) and that
X = Ric(H) ≥ 0:

" # " #
A 0 B2 h i
∗
H = − D12 C1 B2∗
−C1∗ C1 −A∗ −C1∗ D12
" #
∗
A − B2 D12 C1 −B2 B2∗
= ∗ ∗ ∗
. (13.8)
−C1 D⊥ D⊥ C1 −(A − B2 D12 C1 )∗
∗
Note also that if D12 C1 = 0, then (A4) is implied by the detectability of (C1 , A). 3

Note that the Riccati equation corresponding to equation (13.8) is

∗
(A − B2 D12 C1 )∗ X + X(A − B2 D12
∗
C1 ) − XB2 B2∗ X + C1∗ D⊥ D⊥
∗
C1 = 0. (13.9)

Now let X be the corresponding stabilizing solution and define

F := −(B2∗ X + D12
∗
C1 ). (13.10)

Then A + B2 F is stable. Denote

AF := A + B2 F, CF := C1 + D12 F
13.2. Standard LQR Problem 257

and rearrange equation (13.9) to get

A∗F X + XAF + CF∗ CF = 0. (13.11)
Thus X is the observability Gramian of (CF , AF ).
Consider applying the control law u = F x to the system equations (13.6) and (13.7).
The controlled system becomes
ẋ = AF x, x(0) = x0
z = CF x
or, equivalently,
ẋ = AF x + x0 δ(t), x(0− ) = 0
z = CF x.
The associated transfer matrix is
" #
AF I
Gc (s) =
CF 0
and
kGc x0 k22 = x∗0 Xx0 .

The proof of the following theorem requires a preliminary result about internal
stability given input-output stability.
Lemma 13.1 If u, z ∈ L2 [0, ∞) and (C1 , A) is detectable in the system described by
equations (13.6) and (13.7), then x ∈ L2 [0, ∞). Furthermore, x(t) → 0 as t → ∞.

Proof. Since (C1 , A) is detectable, there exists L such that A + LC1 is stable. Let x̂
be the state estimate of x given by
x̂˙ = (A + LC1 )x̂ + (LD12 + B2 )u − Lz.
Then x̂ ∈ L2 [0, ∞) and x̂ → 0 (see Problem 13.1) since z and u are in L2 [0, ∞). Now
let e = x − x̂; then
ė = (A + LC1 )e
and e ∈ L2 [0, ∞). Therefore, x = e + x̂ ∈ L2 [0, ∞). It is easy to see that e(t) → 0 as
t → ∞ for any initial condition e(0). Finally, x(t) → 0 since x̂ → 0. 2

Theorem 13.2 There exists a unique optimal control for the LQR problem, namely
u = F x. Moreover,
min kzk2 = kGc x0 k2 .
u∈L2 [0,∞)

Note that the optimal control strategy is a constant gain state feedback, and this
gain is independent of the initial condition x0 .
258 H2 OPTIMAL CONTROL

Proof. With the change of variable v = u − F x, the system can be written as

" # " #" #
ẋ AF B2 x
= , x(0) = x0 . (13.12)
z CF D12 v

Now if v ∈ L2 [0, ∞), then x, z ∈ L2 [0, ∞) and x(∞) = 0 since AF is stable.

Hence u = F x + v ∈ L2 [0, ∞). Conversely, if u, z ∈ L2 [0, ∞), then from Lemma 13.1
x ∈ L2 [0, ∞). So v ∈ L2 [0, ∞). Thus the mapping v = u − F x between v ∈ L2 [0, ∞)
and those u ∈ L2 [0, ∞) that make z ∈ L2 [0, ∞) is one-to-one and onto. Therefore,

min kzk2 = min kzk2 .

u∈L2 [0,∞) v∈L2 [0,∞)

By differentiating x(t)∗ Xx(t) with respect to t along a solution of the differential equa-
tion (13.12) and by using equation (13.9) and the fact that CF∗ D12 = −XB2 , we see
that
d ∗
x Xx = ẋ∗ Xx + x∗ X ẋ = x∗ (A∗F X + XAF )x + 2x∗ XB2 v
dt
= −x∗ CF∗ CF x + 2x∗ XB2 v
= −(CF x + D12 v)∗ (CF x + D12 v) + 2x∗ CF∗ D12 v + v ∗ v + 2x∗ XB2 v
= − kzk2 + kvk2 . (13.13)

Now integrate equation (13.13) from 0 to ∞ to get

kzk2 = x∗0 Xx0 + kvk2 .

2 2

Clearly, the unique optimal control is v = 0, i.e., u = F x. 2

13.3 Extended LQR Problem

This section considers the extended LQR problem where no detectability assumption is
made for (C1 , A).

Extended LQR Problem

Let a dynamical system be given by

ẋ = Ax + B2 u, x(0) = x0 given but arbitrary

z = C1 x + D12 u

with the following assumptions:

(A1) (A, B2 ) is stabilizable;
13.4. Guaranteed Stability Margins of LQR 259

h i
(A2) D12 has full column rank with D12 D⊥ unitary;
" #
A − jωI B2
(A3) has full column rank for all ω.
C1 D12
Find an optimal control law u ∈ L2 [0, ∞) such that the system is internally
stable (i.e., x ∈ L2 [0, ∞)) and the performance criterion kzk22 is minimized.

Assume the same notation as in the last section, we have:

Theorem 13.3 There exists a unique optimal control for the extended LQR problem,
namely u = F x. Moreover,
min kzk2 = kGc x0 k2 .
u∈L2 [0,∞)

Proof. The proof of this theorem is very similar to the proof of the standard LQR
problem except that, in this case, the input/output stability may not necessarily imply
the internal stability. Instead, the internal stability is guaranteed by the way of choosing
control law.
Suppose that u ∈ L2 [0, ∞) is such a control law that the system is stable, i.e.,
x ∈ L2 [0, ∞). Then v = u − F x ∈ L2 [0, ∞). On the other hand, let v ∈ L2 [0, ∞) and
consider " # " #" #
ẋ AF B2 x
= , x(0) = x0 .
z CF D12 v
Then x, z ∈ L2 [0, ∞) and x(∞) = 0 since AF is stable. Hence u = F x + v ∈ L2 [0, ∞).
Again the mapping v = u − F x between v ∈ L2 [0, ∞) and those u ∈ L2 [0, ∞) that make
z ∈ L2 [0, ∞) and x ∈ L2 [0, ∞) is one to one and onto. Therefore,
min kzk2 = min kzk2 .
u∈L2 [0,∞) v∈L2 [0,∞)

Using the same technique as in the proof of the standard LQR problem, we have
kzk22 = x∗0 Xx0 + kvk22 .
Thus, the unique optimal control is v = 0, i.e., u = F x. 2

13.4 Guaranteed Stability Margins of LQR

Now we will consider the system described by equation (13.6) with the LQR control law
u = F x. The closed-loop block diagram is as shown in Figure 13.1.
The following result is the key to stability margins of an LQR control law.
Lemma 13.4 Let F = −(B2∗ X + D12 ∗
C1 ) and define G12 = D12 + C1 (sI − A)−1 B2 .
Then


I − B2∗ (−sI − A∗ )−1 F ∗ I − F (sI − A)−1 B2 = G∼ 12 (s)G12 (s).
260 H2 OPTIMAL CONTROL

u x
- F - ẋ = Ax + B2 u -

Figure 13.1: LQR closed-loop system

Proof. Note that the Riccati equation (13.9) can be written as

XA + A∗ X − F ∗ F + C1∗ C1 = 0.

Add and subtract sX to the above equation to get

−X(sI − A) − (−sI − A∗ )X − F ∗ F + C1∗ C1 = 0.

Now multiply the above equation from the left by B2∗ (−sI − A∗ )−1 and from the right
by (sI − A)−1 B2 to get

−B2∗ (−sI − A∗ )−1 XB2 − B2∗ X(sI − A)−1 B2 − B2∗ (−sI − A∗ )−1 F ∗ F (sI − A)−1 B2

+B2∗ (−sI − A∗ )−1 C1∗ C1 (sI − A)−1 B2 = 0.

Using −B2∗ X = F + D12
∗
C1 in the above equation, we have

B2∗ (−sI − A∗ )−1 F ∗ + F (sI − A)−1 B2 − B2∗ (−sI − A∗ )−1 F ∗ F (sI − A)−1 B2

+B2∗ (−sI − A∗ )−1 C1∗ D12 + D12

∗
C1 (sI − A)−1 B2

+B2∗ (−sI − A∗ )−1 C1∗ C1 (sI − A)−1 B2 = 0.

∗
Then the result follows from completing the square and from the fact that D12 D12 = I.
2

∗
Corollary 13.5 Suppose D12 C1 = 0. Then



I − B2∗ (−sI − A∗ )−1 F ∗ I − F (sI − A)−1 B2 = I+B2∗ (−sI−A∗ )−1 C1∗ C1 (sI−A)−1 B2 .

In particular,



I − B2∗ (−jωI − A∗ )−1 F ∗ I − F (jωI − A)−1 B2 ≥ I (13.14)

and



I + B2∗ (−jωI − A∗ − F ∗ B2∗ )−1 F ∗ I + F (jωI − A − B2 F )−1 B2 ≤ I. (13.15)
13.5. Standard H2 Problem 261

Note that the inequality (13.15) follows from taking the inverse of inequality (13.14).
Define G(s) = −F (sI − A)−1 B2 and assume for the moment that the system is
single-input. Then the inequality (13.14) shows that the open-loop Nyquist diagram
of the system G(s) in Figure 13.1 never enters the unit disk centered at (−1, 0) of the
complex plane. Hence the system has at least a 6 dB (= 20 log 2) gain margin and a
60o phase margin in both directions. A similar interpretation may be generalized to
multiple-input systems.
Next, it is noted that the inequality (13.15) can also be given some robustness
interpretation. In fact, it implies that the closed-loop system in Figure 13.1 is stable
even if the open-loop system G(s) is perturbed additively by a ∆ ∈ RH∞ as long as
k∆k∞ < 1. This can be seen from the following block diagram and the small gain
theorem, where the transfer matrix from w to z is exactly I + F (jωI − A − B2 F )−1 B2 .

z w
- ∆

- F - ẋ = Ax + B2 u -?
e -

13.5 Standard H2 Problem

The system considered in this section is described by the following standard block
diagram:
z w
G

y u
- K

The realization of the transfer matrix G is taken to be of the form

 
A B1 B2
 
G(s) =  C1 0 D12  .
C2 D21 0

Notice the special off-diagonal structure of D: D22 is assumed to be zero so that G22
is strictly proper;1 also, D11 is assumed to be zero in order to guarantee that the H2
1 This assumption is made without loss of generality since a substitution of K = K(I + D K)−1
D 22
would give the controller for D22 6= 0.
262 H2 OPTIMAL CONTROL

problem is properly posed.2

The following additional assumptions are made for the output feedback H2 problem
in this chapter:
(i) (A, B2 ) is stabilizable and (C2 , A) is detectable;
∗ ∗
(ii) R1 = D12 D12 > 0 and R2 = D21 D21 > 0;
" #
A − jωI B2
(iii) has full column rank for all ω;
C1 D12
" #
A − jωI B1
(iv) has full row rank for all ω.
C2 D21
The first assumption is for the stabilizability of G by output feedback, and the third and
the fourth assumptions together with the first guarantee that the two Hamiltonian ma-
trices associated with the following H2 problem belong to dom(Ric). The assumptions
in (ii) guarantee that the H2 optimal control problem is nonsingular.

H2 Problem The H2 control problem is to find a proper, real rational

controller K that stabilizes G internally and minimizes the H2 norm of the
transfer matrix Tzw from w to z.
In the following discussions we shall assume that we have state models of G and K.
Recall that a controller is said to be admissible if it is internally stabilizing and proper.
By Corollary 12.7 the two Hamiltonian matrices
" #
A − B2 R1−1 D12 ∗
C1 −B2 R1−1 B2∗
H2 :=
−C1∗ (I − D12 R1−1 D12 ∗
)C1 −(A − B2 R1−1 D12 ∗
C1 )∗
" #
(A − B1 D21 ∗
R2−1 C2 )∗ −C2∗ R2−1 C2
J2 :=
−B1 (I − D21 ∗
R2−1 D21 )B1∗ −(A − B1 D21 ∗
R2−1 C2 )
belong to dom(Ric), and, moreover, X2 := Ric(H2 ) ≥ 0 and Y2 := Ric(J2 ) ≥ 0. Define
F2 := −R1−1 (B2∗ X2 + D12
∗
C1 ), L2 := −(Y2 C2∗ + B1 D21
∗
)R2−1
and
AF2 := A + B2 F2 , C1F2 := C1 + D12 F2
AL2 := A + L2 C2 , B1L2 := B1 + L2 D21
Â2 := A + B2 F2 + L2 C2
" # " #
AF2 I AL2 B1L2
Gc (s) := , Gf (s) := .
C1F2 0 I 0
Before stating the main theorem, we note the following fact:
2 Recall that a rational proper stable transfer function is an RH2 function iff it is strictly proper.
13.5. Standard H2 Problem 263

Lemma 13.6 Let U, V ∈ RH∞ be defined as

" −1/2
# " #
AF2 B2 R1 AL2 B1L2
U := −1/2 , V := −1/2 −1/2 .
C1F2 D12 R1 R2 C2 R2 D21

Then U is an inner and V is a co-inner, U ∼ Gc ∈ RH⊥

2 , and Gf V
∼
∈ RH⊥
2.

Proof. The proof uses standard manipulations of state-space realizations. From U we

get " #
∼ −A∗F2 −C1F∗
U (s) = −1/2 −1/2 ∗
2
.
R1 B2∗ R1 D12
Then it is easy to compute
 −1/2 
−A∗F2 ∗
−C1F C1F2 ∗
−C1F D12 R1
 
2 2
−1/2
U ∼U =  0 AF2 B2 R1 
−1/2 −1/2
R1 B2∗ R1 ∗
D12 C1F2 I
 
−A∗F2 ∗
−C1F C1F2 0
 
2

U ∼ Gc =  0 AF2 I .
−1/2 −1/2
R1 B2∗ R1 ∗
D12 C1F2 0
Now do the similarity transformation
" #
I −X2
0 I

on the states of the preceding transfer matrices and note that

A∗F2 X2 + X2 AF2 + C1F

∗
2
C1F2 = 0.

We get  
−A∗F2 0 0
 −1/2 
U ∼U =  0 AF2 B2 R1 =I
−1/2
R1 B2∗ 0 I
 
−A∗F2 0 −X2 " #
  −A∗F2 −X2
∼
U Gc =  0 AF2 I = −1/2 ∈ RH⊥
2.
−1/2 R1 B2∗ 0
R1 B2∗ 0 0

It follows by duality that Gf V ∼ ∈ RH⊥

2 and V is a co-inner. 2
264 H2 OPTIMAL CONTROL

Theorem 13.7 There exists a unique optimal controller

" #
Â2 −L2
Kopt (s) := .
F2 0

Moreover,

min kTzw k22 = kGc B1 k22 + kR1 F2 Gf k22 = trace (B1∗ X2 B1 ) + trace (R1 F2 Y2 F2∗ ) .
1/2

Proof. Consider the all-stabilizing controller parameterization K(s) = F` (M2 , Q),

Q ∈ RH∞ with
 
Â2 −L2 B2
 
M2 (s) =  F2 0 I 
−C2 I 0
and consider the following system diagram:
z  w
G 
y u
XXX 

X
 XXX

M2 
y1 u1

- Q

Then Tzw = F` (N, Q) with

 
AF2 −B2 F2 B1 B2
 
 0 AL2 B1L2 0 
N =
 C

 1F2 −D12 F2 0 D12 

0 C2 D21 0

and
1/2 1/2 1/2
Tzw = Gc B1 − U R1 F2 Gf + U R1 QR2 V.
It follows from Lemma 13.6 that Gc B1 and U are orthogonal. Thus
2
2 2 1/2 1/2 1/2
kTzw k2 = kGc B1 k2 + U R1 F2 Gf − U R1 QR2 V
2
2
2 1/2 1/2 1/2
= kGc B1 k2 + R1 F2 Gf − R1 QR2 V .
2
13.6. Stability Margins of H2 Controllers 265

Since Gf and V are also orthogonal by Lemma 13.6, we have

2
1/2 1/2 1/2
kTzw k22 = kGc B1 k22 + R1 F2 Gf − R1 QR2 V
2
2 2
2 1/2 1/2 1/2
= kGc B1 k2 + R1 F2 Gf + R1 QR2 .
2 2

This shows clearly that Q = 0 gives the unique optimal control, so K = F` (M2 , 0) is
the unique optimal controller. 2

The optimal H2 controller, Kopt , and the closed-loop transfer matrix, Tzw , can be
obtained by the following Matlab program:
 [K, Tzw ] = h2syn(G, ny , nu )
where ny and nu are the dimensions of y and u, respectively.

Related MATLAB Commands: lqg, lqr, lqr2, lqry, reg, lqe

13.6 Stability Margins of H2 Controllers

It is well-known that a system with LQR controller has at least 60o phase margin
and 6 dB gain margin. However, it is not clear whether these stability margins will
be preserved if the states are not available and the output feedback H2 (or LQG)
controller has to be used. The answer is provided here through a counterexample from
Doyle [1978]: There are no guaranteed stability margins for a H2 controller.
Consider a single-input and single-output two-state generalized dynamical system:
 " # " √ # " # 
1 1 σ 0 0
 √ 
 0 1 σ 0 1 
 
 " # " # 
 √ √ 
G(s) =  q q 0 .
 0 
 0 0 1 
 
 h i h i 
1 0 0 1 0

It can be shown analytically that

" # " #
2α α 2β β
X2 = , Y2 =
α α β β

and " #
h i 1
F2 = −α 1 1 , L2 = −β
1
266 H2 OPTIMAL CONTROL

where p √
α=2+ 4+q , β =2+ 4 + σ.
Then the optimal output H2 controller is given by
 
1−β 1 β
 
Kopt =  −(α + β) 1 − α β  .
−α −α 0

Suppose that the resulting closed-loop controller (or plant G22 ) has a scalar gain k with
a nominal value k = 1. Then the controller implemented in the system is actually

K = kKopt ,

and the closed-loop system A matrix becomes

 
1 1 0 0
 
 0 1 −kα −kα 
à = 
 β
.

 0 1−β 1 
β 0 −α − β 1−α

It can be shown that the characteristic polynomial has the form

det(sI − Ã) = s4 + a3 s3 + a2 s2 + a1 s + a0

with
a1 = α + β − 4 + 2(k − 1)αβ, a0 = 1 + (1 − k)αβ.
Note that for closed-loop stability it is necessary to have a0 > 0 and a1 > 0. Note also
that a0 ≈ (1 − k)αβ and a1 ≈ 2(k − 1)αβ for sufficiently large α and β if k 6= 1. It is
easy to see that for sufficiently large α and β (or q and σ), the system is unstable for
arbitrarily small perturbations in k in either direction. Thus, by choice of q and σ, the
gain margins may be made arbitrarily small.
It is interesting to note that the margins deteriorate as control weight (1/q) gets
small (large q) and/or system driving noise gets large (large σ). In modern control
folklore, these have often been considered ad hoc means of improving sensitivity.
It is also important to recognize that vanishing margins are not only associated with
open-loop unstable systems. It is easy to construct minimum phase, open-loop stable
counterexamples for which the margins are arbitrarily small.
The point of this example is that H2 (LQG) solutions, unlike LQR solutions, pro-
vide no global system-independent guaranteed robustness properties. Like their more
classical colleagues, modern LQG designers are obliged to test their margins for each
specific design.
It may, however, be possible to improve the robustness of a given design by relaxing
the optimality of the filter with respect to error properties. A successful approach in
13.7. Notes and References 267

this direction is the so called LQG loop transfer recovery (LQG/LTR) design technique.
The idea is to design a filtering gain, L2 , in such way so that the LQG (or H2 ) control
law will approximate the loop properties of the regular LQR control. This will not be
explored further here; interested readers may consult related references.

13.7 Notes and References

Detailed treatment of H2 related theory, LQ optimal control, Kalman filtering, etc.,
can be found in Anderson and Moore [1989] or Kwakernaak and Sivan [1972]. The
LQG/LTR control design was first introduced by Doyle and Stein [1981], and much
work has been reported in this area since then. Additional results on the LQR stability
margins can be found in Zhang and Fu [1996].

13.8 Problems
Problem 13.1 Let v(t) ∈ L2 [0, ∞). Let y(t) be the output of the system G(s) = 1
s+1
with input v. Prove that limt→∞ y(t) = 0.

Problem 13.2 Parameterize all stabilizing controllers satisfying kTzw k2 ≤ γ for a given
γ > 0.

Problem 13.3 Consider the feedback system in Figure 6.3 and suppose

s − 10 1 s+2
P = , We = , Wu = .
(s + 1)(s + 10) s + 0.001 s + 10

Design a controller that minimizes

" #

We So
.
Wu KSo
2

Simulate the time response of the system when r is a step.

Problem 13.4 Repeat Problem 13.3 when We = 1/s. (Note that the solution given in
this chapter cannot be applied directly.)

Problem 13.5 Consider the model matching (or reference) control problem shown
here:
268 H2 OPTIMAL CONTROL

uw
- W -

-e - K - P
−
6 u
r ?−
e
e
-
6
- M

Let M (s) ∈ H∞ be a strictly proper transfer matrix and W (s), W −1 (s) ∈ RH∞ . For-
mulate an H2 control problem that minimizes uw and the error e through minimizing
the H2 norm of the transfer matrix from r to (e, uw ). Apply your formula to

4 10(s + 2) 0.1(s + 1)
M (s) = , P (s) = , W (s) = .
s2 + 2s + 4 (s + 1)3 s + 10

Problem 13.6 Repeat Problem 13.5 with W =  for  = 0.01 and 0.0001. Study the
behavior of the controller when  → 0.

Problem 13.7 Repeat Problem 13.5 and Problem 13.6 with

10(2 − s)
P = .
(s + 1)3
Chapter 14

H∞ Control

In this chapter we consider H∞ control theory. Specifically, we formulate the optimal

and suboptimal H∞ control problems in Section 14.1. However, we will focus on the
suboptimal case in this book and discuss why we do so. In Section 14.2 a suboptimal
controller is characterized together with an algebraic proof for a class of simplified
problems while leaving the more general problems to a later section. The behavior of
the H∞ controller as a function of performance level γ is considered in Section 14.3. The
optimal controllers are also briefly considered in this section. Some other interpretations
of the H∞ controllers are given in Section 14.4. Section 14.5 presents the formulas for an
optimal H∞ controller. Section 14.6 considers again the standard H∞ control problem
but with some assumptions in the previous sections relaxed. Since the proof techniques
in Section 14.2 can, in principle, be applied to this general case except with some more
involved algebra, the detailed proof for the general case will not be given; only the
formulas are presented. We shall indicate how the assumptions in the general case can
be relaxed further to accommodate other more complicated problems in Section 14.7.
Section 14.8 considers the integral control in the H2 and H∞ theory and Section 14.9
considers how the general H∞ solution can be used to solve the H∞ filtering problem.

14.1 Problem Formulation

Consider the system described by the block diagram
z w
G

y u
- K

where the plant G and controller K are assumed to be real rational and proper. It will
be assumed that state-space models of G and K are available and that their realizations

269
270 H∞ CONTROL

are assumed to be stabilizable and detectable. Recall again that a controller is said to
be admissible if it internally stabilizes the system. Clearly, stability is the most basic
requirement for a practical system to work. Hence any sensible controller has to be
admissible.
Optimal H∞ Control: Find all admissible controllers K(s) such that
kTzw k∞ is minimized.
It should be noted that the optimal H∞ controllers as just defined are generally not
unique for MIMO systems. Furthermore, finding an optimal H∞ controller is often both
numerically and theoretically complicated, as shown in Glover and Doyle [1989]. This
is certainly in contrast with the standard H2 theory, in which the optimal controller
is unique and can be obtained by solving two Riccati equations without iterations.
Knowing the achievable optimal (minimum) H∞ norm may be useful theoretically since
it sets a limit on what we can achieve. However, in practice it is often not necessary
and sometimes even undesirable to design an optimal controller, and it is usually much
cheaper to obtain controllers that are very close in the norm sense to the optimal ones,
which will be called suboptimal controllers. A suboptimal controller may also have other
nice properties (e.g., lower bandwidth) over the optimal ones.

Suboptimal H∞ Control: Given γ > 0, find all admissible controllers

K(s), if there are any, such that kTzw k∞ < γ.

For the reasons mentioned above, we focus our attention in this book on suboptimal
control. When appropriate, we briefly discuss what will happen when γ approaches the
optimal value.

14.2 A Simplified H∞ Control Problem

The realization of the transfer matrix G is taken to be of the form
 
A B1 B2
 
G(s) =  C1 0 D12 .
C2 D21 0

The following assumptions are made:

(i) (A, B1 ) is controllable and (C1 , A) is observable;
(ii) (A, B2 ) is stabilizable and (C2 , A) is detectable;
h i h i
∗
(iii) D12 C1 D12 = 0 I ;
" # " #
B1 ∗ 0
(iv) D21 = .
D21 I
14.2. A Simplified H∞ Control Problem 271

Two additional assumptions that are implicit in the assumed realization for G(s) are
that D11 = 0 and D22 = 0. As we mentioned in the last chapter, D22 6= 0 does not pose
any problem since it is easy to form an equivalent problem with D22 = 0 by a linear
fractional transformation on the controller K(s). However, relaxing the assumption
D11 = 0 complicates the formulas substantially.
The H∞ solution involves the following two Hamiltonian matrices:
" # " #
A γ −2 B1 B1∗ − B2 B2∗ A∗ γ −2 C1∗ C1 − C2∗ C2
H∞ := , J∞ := .
−C1∗ C1 −A∗ −B1 B1∗ −A
The important difference here from the H2 problem is that the (1,2)-blocks are not sign
definite, so we cannot use Theorem 12.4 in Chapter 12 to guarantee that H∞ ∈ dom(Ric)
or Ric(H∞ ) ≥ 0. Indeed, these conditions are intimately related to the existence of
H∞ suboptimal controllers. Note that the (1,2)-blocks are a suggestive combination of
expressions from the H∞ norm characterization in Chapter 4 (or bounded real lemma
in Chapter 12) and from the H2 synthesis of Chapter 13. It is also clear that if γ
approaches infinity, then these two Hamiltonian matrices become the corresponding H2
control Hamiltonian matrices. The reasons for the form of these expressions should
become clear through the discussions and proofs for the following theorem.
Theorem 14.1 There exists an admissible controller such that kTzw k∞ < γ iff the
following three conditions hold:
(i) H∞ ∈ dom(Ric) and X∞ := Ric(H∞ ) > 0;
(ii) J∞ ∈ dom(Ric) and Y∞ := Ric(J∞ ) > 0;
(iii) ρ(X∞ Y∞ ) < γ 2 .
Moreover, when these conditions hold, one such controller is
" #
Â∞ −Z∞ L∞
Ksub (s) :=
F∞ 0
where
Â∞ := A + γ −2 B1 B1∗ X∞ + B2 F∞ + Z∞ L∞ C2
F∞ := −B2∗ X∞ , L∞ := −Y∞ C2∗ , Z∞ := (I − γ −2 Y∞ X∞ )−1 .
Furthermore, the set of all admissible controllers such that kTzw k∞ < γ equals the set
of all transfer matrices from y to u in
u y  
M∞ Â∞ −Z∞ L∞ Z∞ B2
  
M∞ (s) =  F∞ 0 I 
- Q −C2 I 0

where Q ∈ RH∞ , kQk∞ < γ.

272 H∞ CONTROL

We shall only give a proof of the first part of the theorem; the proof for the all-
controller parameterization needs much more work and is omitted (see Zhou, Doyle,
and Glover [1996] for a comprehensive treatment of the related topics). We shall first
show some preliminary results.

Lemma 14.2 Suppose that X ∈ Rn×n , Y ∈ Rn×n , with X = X ∗ > 0, and Y = Y ∗ > 0.
Let r be a positive integer. Then there exist matrices X12 ∈ Rn×r , X2 ∈ Rr×r such that
X2 = X2∗
" # " #−1 " #
X X12 X X12 Y ?
∗
>0 and ∗
=
X12 X2 X12 X2 ? ?

if and only if
" # " #
X In X In
≥0 and rank ≤ n + r.
In Y In Y

Proof. (⇐) By the assumption, there is a matrix X12 ∈ Rn×r such that X − Y −1 =
∗
X12 X12 . Defining X2 := Ir completes the construction.
(⇒) Using Schur complements,

Y = X −1 + X −1 X12 (X2 − X12

∗
X −1 X12 )−1 X12
∗
X −1 .

Inverting, using the matrix inversion lemma, gives

Y −1 = X − X12 X2−1 X12

∗
.

Hence, X −Y −1 = X12 X2−1 X12

∗
≥ 0, and, indeed, rank(X −Y −1 ) =rank(X12 X2−1 X12
∗
)≤
r. 2

Lemma 14.3 There exists an rth-order admissible controller such that kTzw k∞ < γ
only if the following three conditions hold:
(i) There exists a Y1 > 0 such that

AY1 + Y1 A∗ + Y1 C1∗ C1 Y1 /γ 2 + B1 B1∗ − γ 2 B2 B2∗ < 0. (14.1)

(ii) There exists an X1 > 0 such that

X1 A + A∗ X1 + X1 B1 B1∗ X1 /γ 2 + C1∗ C1 − γ 2 C2∗ C2 < 0. (14.2)

" # " #
X1 /γ In X1 /γ In
(iii) ≥0 rank ≤ n + r.
In Y1 /γ In Y1 /γ
14.2. A Simplified H∞ Control Problem 273

Proof. Suppose that there exists an rth-order controller K(s) such that kTzw k∞ < γ.
Let K(s) have a state-space realization
" #
 B̂
K(s) = .
Ĉ D̂

Then
 
A + B2 D̂C2 B2 Ĉ B1 + B2 D̂D21 " #
  Ac Bc
Tzw 
= F` (G, K) =  B̂C2 Â B̂D21 
 =: C Dc
.
c
C1 + D12 D̂C2 D12 Ĉ D12 D̂D21

Denote
R = γ 2 I − Dc∗ Dc , R̃ = γ 2 I − Dc Dc∗ .
" #
X1 X12
By Corollary 12.3, there exists an X̃ = ∗
> 0 such that
X12 X2

X̃(Ac + Bc R−1 Dc∗ Cc ) + (Ac + Bc R−1 Dc∗ Cc )∗ X̃ + X̃Bc R−1 Bc∗ X̃ + Cc∗ R̃−1 Cc < 0. (14.3)

This gives, after much algebraic manipulation,

X1 A + A∗ X1 + X1 B1 B1∗ X1 /γ 2 + C1∗ C1 − γ 2 C2∗ C2

+(X1 B1 D̂ + X12 B̂ + γ 2 C2∗ )(γ 2 I − D̂∗ D̂)−1 (X1 B1 D̂ + X12 B̂ + γ 2 C2∗ )∗ < 0,
which implies that

X1 A + A∗ X1 + X1 B1 B1∗ X1 /γ 2 + C1∗ C1 − γ 2 C2∗ C2 < 0.

On the other hand, let

Ỹ = γ 2 X̃ −1
" #
Y1 Y12
and partition Ỹ as Ỹ = ∗
> 0. Then
Y12 Y2

(Ac + Bc R−1 Dc∗ Cc )Ỹ + Ỹ (Ac + Bc R−1 Dc∗ Cc )∗ + Ỹ Cc∗ R̃−1 Cc Ỹ + Bc R−1 Bc∗ < 0. (14.4)

This gives
AY1 + Y1 A∗ + B1 B1∗ − γ 2 B2 B2∗ + Y1 C1∗ C1 Y1 /γ 2
+(Y1 C1∗ D̂∗ + Y12 Ĉ ∗ + γ 2 B2 )(γ 2 I − D̂D̂∗ )−1 (Y1 C1∗ D̂∗ + Y12 Ĉ ∗ + γ 2 B2 )∗ < 0,
which implies that

AY1 + Y1 A∗ + B1 B1∗ − γ 2 B2 B2∗ + Y1 C1∗ C1 Y1 /γ 2 < 0.

274 H∞ CONTROL

By Lemma 14.2, given X1 > 0 and Y1 > 0, there exists X12 and X2 such that Ỹ = γ 2 X̃ −1
or Ỹ /γ = (X̃/γ)−1 :
" #−1 " #
X1 /γ X12 /γ Y1 /γ ?
∗
=
X12 /γ X2 /γ ? ?
if and only if
" # " #
X1 /γ In X1 /γ In
≥0 rank ≤ n + r.
In Y1 /γ In Y1 /γ
2

To show that the inequalities in the preceding lemma imply the existence of the
stabilizing solutions to the Riccati equations of X∞ and Y∞ , we need the following
theorem.

Theorem 14.4 Let R ≥ 0 and suppose (A, R) is controllable and there is an X = X ∗

such that
Q(X) := XA + A∗ X + XRX + Q < 0. (14.5)
Then there exists a solution X+ > X to the Riccati equation

X+ A + A∗ X+ + X+ RX+ + Q = 0 (14.6)

such that A + RX+ is antistable.

Proof. Let R = BB ∗ for some B. Note the fact that (A, R) is controllable iff (A, B)
is. Let X be such that Q(X) < 0. Since (A, B) is controllable, there is an F0 such that

A0 := A − BF0

is antistable. Now let X0 = X0∗ be the unique solution to the Lyapunov equation

X0 A0 + A∗0 X0 − F0∗ F0 + Q = 0.

Define
F̂0 := F0 + B ∗ X,
and we have the following equation:

(X0 − X)A0 + A∗0 (X0 − X) = F̂0∗ F̂0 − Q(X) > 0.

The antistability of A0 implies that

X0 > X.

Starting with X0 , we shall define a nonincreasing sequence of Hermitian matrices {Xi }.

Associated with {Xi }, we shall also define a sequence of antistable matrices {Ai } and a
14.2. A Simplified H∞ Control Problem 275

sequence of matrices {Fi }. Assume inductively that we have already defined matrices
{Xi }, {Ai }, and {Fi } for i up to n − 1 such that Xi is Hermitian and
X0 ≥ X1 ≥ · · · ≥ Xn−1 > X,
Ai = A − BFi is antistable, i = 0, . . . , n − 1;
Fi = −B ∗ Xi−1 , i = 1, . . . , n − 1;
Xi Ai + A∗i Xi = Fi∗ Fi − Q, i = 0, 1, . . . , n − 1. (14.7)
Next, introduce
Fn = −B ∗ Xn−1 ,
An = A − BFn .
First we show that An is antistable. Using equation (14.7), with i = n − 1, we get
Xn−1 An + A∗n Xn−1 + Q − Fn∗ Fn − (Fn − Fn−1 )∗ (Fn − Fn−1 ) = 0. (14.8)
Let
F̂n := Fn + B ∗ X;
then
(Xn−1 −X)An +A∗n (Xn−1 −X) = −Q(X)+F̂n∗ F̂n +(Fn −Fn−1 )∗ (Fn −Fn−1 ) > 0, (14.9)
which implies that An is antistable by Lyapunov theorem since Xn−1 − X > 0.
Now we introduce Xn as the unique solution of the Lyapunov equation:
Xn An + A∗n Xn = Fn∗ Fn − Q. (14.10)
Then Xn is Hermitian. Next, we have
(Xn − X)An + A∗n (Xn − X) = −Q(X) + F̂n∗ F̂n > 0,
and, by using equation (14.8),
(Xn−1 − Xn )An + A∗n (Xn−1 − Xn ) = (Fn − Fn−1 )∗ (Fn − Fn−1 ) ≥ 0.
Since An is antistable, we have
Xn−1 ≥ Xn > X.
We have a nonincreasing sequence {Xi }, and the sequence is bounded below by Xi > X.
Hence the limit
X+ := lim Xn
n→∞
exists and is Hermitian, and we have X+ ≥ X. Passing the limit n → ∞ in equation
(14.10), we get Q(X+ ) = 0. So X+ is a solution of equation (14.6).
Note that X+ − X ≥ 0 and
(X+ − X)A+ + A∗+ (X+ − X) = −Q(X) + (X+ − X)R(X+ − X) > 0. (14.11)
Hence, X+ − X > 0 and A+ = A + RX+ is antistable. 2
276 H∞ CONTROL

Lemma 14.5 There exists an admissible controller such that kTzw k∞ < γ only if the
following three conditions hold:
(i) There exists a stabilizing solution X∞ > 0 to

X∞ A + A∗ X∞ + X∞ (B1 B1∗ /γ 2 − B2 B2∗ )X∞ + C1∗ C1 = 0. (14.12)

(ii) There exists a stabilizing solution Y∞ > 0 to

AY∞ + Y∞ A∗ + Y∞ (C1∗ C1 /γ 2 − C2∗ C2 )Y∞ + B1 B1∗ = 0. (14.13)

" #
−1
γY∞ In
(iii) −1
>0 or ρ(X∞ Y∞ ) < γ 2 .
In γX∞

Proof. Applying Theorem 14.4 to part (i) of Lemma 14.3, we conclude that there
exists a Y > Y1 > 0 such that

AY + Y A∗ + Y C1∗ C1 Y /γ 2 + B1 B1∗ − γ 2 B2 B2∗ = 0

and A + C1∗ C1 Y /γ 2 is antistable. Let X∞ := γ 2 Y −1 ; we have

X∞ A + A∗ X∞ + X∞ (B1 B1∗ /γ 2 − B2 B2∗ )X∞ + C1∗ C1 = 0 (14.14)

and

A + (B1 B1∗ /γ 2 − B2 B2∗ )X∞ = −X∞

−1
(A + C1∗ C1 X∞
−1 −1
)X∞ = −X∞ (A + C1∗ C1 Y /γ 2 )X∞

is stable.
Similarly, applying Theorem 14.4 to part (ii) of Lemma 14.3, we conclude that there
exists an X > X1 > 0 such that

XA + A∗ X + XB1 B1∗ X/γ 2 + C1∗ C1 − γ 2 C2∗ C2 = 0

and A + B1 B1∗ X/γ 2 is antistable. Let Y∞ := γ 2 X −1 , we have

AY∞ + Y∞ A∗ + Y∞ (C1∗ C1 /γ 2 − C2∗ C2 )Y∞ + B1 B1∗ = 0 (14.15)

and A + (C1∗ C1 /γ 2 − C2∗ C2 )Y∞ is stable.

Finally, note that the rank condition in part (iii) of Lemma 14.3 is automatically
satisfied by r ≥ n, and
" # " # " #
−1
γY∞ In X/γ In X1 /γ In
−1
= > ≥0
In γX∞ In Y /γ In Y1 /γ

or ρ(X∞ Y∞ ) < γ 2 . 2
14.2. A Simplified H∞ Control Problem 277

Proof of Theorem 14.1: To complete the proof, we only need to show that the
controller Ksub given in Theorem 14.1 renders kTzw k∞ < γ. Note that the closed-loop
transfer function with Ksub is given by
 
A B2 F∞ B1 " #
  A B
Tzw =  −Z∞ L∞ D21 
c c
 −Z∞ L∞ C2 Â∞  =: C D .
c c
C1 D12 F∞ 0

Define " #
−1 −1 −1
γ 2 Y∞ −γ 2 Y∞ Z∞
P = ∗ −1 −1 2 −1 −1
.
−γ 2 (Z∞ ) Y∞ γ Y∞ Z∞
Then it is easy to show that P > 0 and

P Ac + A∗c P + P Bc Bc∗ P/γ 2 + Cc∗ Cc = 0.

Moreover,
" #
A + B1 B1∗ Y∞
−1
B2 F∞ − B1 B1∗ Y∞−1 −1
Z∞
Ac + Bc Bc∗ P/γ 2 = ∗ 2
0 A + B1 B1 X∞ /γ + B2 F∞

has no eigenvalues on the imaginary axis since A + B1 B1∗ X∞ /γ 2 + B2 F∞ is stable and

A + B1 B1∗ Y∞
−1
is antistable. Thus, by Corollary 12.3, kTzw k∞ < γ. 2

Remark 14.1 It is appropriate to point out that the conditions stated in Lemma 14.3
are, in fact, necessary and sufficient; see Gahinet and Apkarian [1994] and Gahinet
[1996] for a linear matrix inequality (LMI) approach to the H∞ problem. But the
necessity should be suitably interpreted. For example, if one finds an X1 > 0 and a
Y1 > 0 satisfying conditions (i) and (ii) but not condition (iii), this does not imply that
there is no admissible H∞ controller since there might be other X1 > 0 and Y1 > 0 that
satisfy all three conditions. For example, consider γ = 1 and
 h i 
−1 1 0 1
 " # " # 
 
 1 0 
G(s) =  0 .
 0 1 
 h i 
1 0 1 0

It is easy to check that X1 = Y1 = 0.5 satisfy (i) and (ii) but not (iii). Nevertheless,
we shall show in the next section that γopt = 0.7321 and thus a suboptimal controller
exists for γ = 1. In fact, we can check that 1 < X1 < 2, 1 < Y1 < 2 also satisfy (i), (ii)
and (iii). 3
278 H∞ CONTROL

Example 14.1 Consider the feedback system shown in Figure 6.3 with
50(s + 1.4) 2 s+1
P = , We = , Wu = .
(s + 1)(s + 2) s + 0.2 s + 10
" # " #
d e
We shall design a controller so that the H∞ norm from w = to z = is
di ũ
minimized. Note that
" # " #" # " #
e We (I + P K)−1 We (I + P K)−1 P d d
= =: Tzw .
ũ −Wu K(I + P K)−1 −Wu K(I + P K)−1 P di di

Then the problem can be set up in an LFT framework with

 
−0.2 2 2 0 2 0 0
 
 0 −1 0 0 0 20 −20 
   
We We P −We P  0 0 −2 0 0 30 −30 
 
   
G(s) =  0 0 −Wu  =  0 0 0 −10 0 0 −3 .
 
I P −P  0 
 1 0 0 0 0 0 
 
 0 0 0 −3 0 0 −1 
0 1 1 0 1 0 0

A suboptimal H∞ controller can be computed by using the following command:

 [K, Tzw , γsubopt ] = hinfsyn(G, ny , nu , γmin , γmax , tol)
where ny and nu are the dimensions of y and u; γmin and γmax are, respectively a lower
bound and an upper bound for γopt ; and tol is a tolerance to the optimal value. Set
ny = 1, nu = 1, γmin = 0, γmax = 10, tol = 0.0001; we get γsubopt = 0.7849 and a
suboptimal controller
12.82(s/10 + 1)(s/7.27 + 1)(s/1.4 + 1)
K= .
(s/32449447.67 + 1)(s/22.19 + 1)(s/1.4 + 1)(s/0.2 + 1)
If we set tol = 0.01, we would get γsubopt = 0.7875 and a suboptimal controller
12.78(s/10 + 1)(s/7.27 + 1)(s/1.4 + 1)
K̃ = .
(s/2335.59 + 1)(s/21.97 + 1)(s/1.4 + 1)(s/0.2 + 1)

The only significant difference between K and K̃ is the exact location of the far-away
stable controller pole. Figure 14.1 shows the closed-loop frequency response of σ (Tzw )
and Figure 14.2 shows the frequency responses of S, T, KS, and SP .
14.2. A Simplified H∞ Control Problem 279

0
10

−1
10 −2 −1 0 1 2 3 4
10 10 10 10 10 10 10
frequency (rad/sec)

Figure 14.1: The closed-loop frequency responses of σ(Tzw ) with

K (solid line) and K̃ (dashed line)

1
10

T
0
10

KS

−1
10 SP

S
−2
10

−3
10 −2 −1 0 1 2 3 4
10 10 10 10 10 10 10
frequency (rad/sec)

Figure 14.2: The frequency responses of S, T, KS, and SP with K

280 H∞ CONTROL

Example 14.2 Consider again the two-mass/spring/damper system shown in Figure 4.2.
Assume that F1 is the control force, F2 is the disturbance force, and the measurements
of x1 and x2 are corrupted by measurement noise:
 
" # " # " # 0.01(s + 10)
y1 x1 n1  0 
Wn =  s + 100
y=
y2
=
x2
+ Wn
n2
, 0.01(s + 10)  .
0
s + 100
Our objective is to design a control law so that the effect of the disturbance force F2 on
the positions of the two masses, x1 and x2 , are reduced in a frequency" range 0 ≤#ω ≤ 2.
W1 0
The problem can be set up as shown in Figure 14.3, where We = is the
0 W2
performance weight and Wu is the control weight. In order to limit the control force,
we shall choose
s+5
Wu = .
s + 50

6z2

w1 = F2
Wu " #
x1
6 ? x2
y z1
- K - Plant - We -
u = F1 " #
n1
w2 =
n2
?
e Wn 

Figure 14.3: Rejecting the disturbance force F2 by a feedback control

 
F2
 
Now let u = F1 , w =  n1 ; then the problem can be formulated in an LFT form
n2
with  " # " # 
We P1 0 We P2
 
G(s) = 
 h 0 0i Wu 

P1 Wn P2
14.2. A Simplified H∞ Control Problem 281

" #
x1
where P1 and P2 denote the transfer matrices from F1 and F2 to , respectively.
x2
Let
5
W1 = , W2 = 0.
s/2 + 1
That is, we only want to reject the effect of the disturbance force F2 on the position x1 .
Then the optimal H2 performance is kF` (G, K2 )k2 = 2.6584 and the H∞ performance
with the optimal H2 controller is kF` (G, K2 )k∞ = 2.6079 while the optimal H∞ perfor-
mance with an H∞ controller is kF` (G, K∞ )k∞ = 1.6101. This means that the effect
of the disturbance force F2 in the desired frequency rang 0 ≤ ω ≤ 2 will be effectively
reduced with the H∞ controller K∞ by 5/1.6101 = 3.1054 times at x1 . On the other
hand, let
5
W1 = 0, W2 = .
s/2 + 1
That is, we only want to reject the effect of the disturbance force F2 on the position x2 .
Then the optimal H2 performance is kF` (G, K2 )k2 = 0.1659 and the H∞ performance
with the optimal H2 controller is kF` (G, K2 )k∞ = 0.5202 while the optimal H∞ perfor-
mance with an H∞ controller is kF` (G, K∞ )k∞ = 0.5189. This means that the effect
of the disturbance force F2 in the desired frequency rang 0 ≤ ω ≤ 2 will be effectively
reduced with the H∞ controller K∞ by 5/0.5189 = 9.6358 times at x2 .
1
10

H∞ Control
The largest singular value

H2 Control
0
10

−1
10 −1 0 1
10 10 10
frequency (rad/sec)

Figure 14.4: The largest singular value plot of the closed-loop system Tzw with an H2
controller and an H∞ controller

Finally, set
5
W1 = W2 = .
s/2 + 1
282 H∞ CONTROL

That is, we want to reject the effect of the disturbance force F2 on both x1 and x2 . Then
the optimal H2 performance is kF` (G, K2 )k2 = 4.087 and the H∞ performance with
the optimal H2 controller is kF` (G, K2 )k∞ = 6.0921 while the optimal H∞ performance
with an H∞ controller is kF` (G, K∞ )k∞ = 4.3611. This means that the effect of the
disturbance force F2 in the desired frequency rang 0 ≤ ω ≤ 2 will only be effectively
reduced with the H∞ controller K∞ by 5/4.3611 = 1.1465 times at both x1 and x2 .
This result shows clearly that it is very hard to reject the disturbance effect on both
positions at the same time. The largest singular value Bode plots of the closed-loop
system are shown in Figure 14.4. We note that the H∞ controller typically gives a
relatively flat frequency response since it tries to minimize the peak of the frequency
response. On the other hand, the H2 controller would typically produce a frequency
response that rolls off fast in the high-frequency range but with a large peak in the
low-frequency range.

14.3 Optimality and Limiting Behavior

In this section, we will discuss, without proof, the behavior of the H∞ suboptimal
solution as γ varies, especially as γ approaches the infima achievable norm, denoted by
γopt . Since Theorem 14.1 gives necessary and sufficient conditions for the existence of
an admissible controller such that kTzw k∞ < γ, γopt is the infimum over all γ such that
conditions (i)–(iii) are satisfied. Theorem 14.1 does not give an explicit formula for γopt ,
but, just as for the H∞ norm calculation, it can be computed as closely as desired by a
search technique.
Although we have not focused on the problem of H∞ optimal controllers, the as-
sumptions in this book make them relatively easy to obtain in most cases. In addition to
describing the qualitative behavior of suboptimal solutions as γ varies, we will indicate
why the descriptor version of the controller formulas below can usually provide formulas
for the optimal controller when γ = γopt .
As γ → ∞, H∞ → H2 , X∞ → X2 , etc., and Ksub → K2 . This fact is the result of
the particular choice of the suboptimal controller. While it could be argued that Ksub
is a natural choice, this connection with H2 actually hints at deeper interpretations. In
fact, Ksub is the minimum entropy solution (see Section 14.4) as well as the minimax
controller for kzk22 − γ 2 kwk22 .
If γ2 ≥ γ1 > γopt , then X∞ (γ1 ) ≥ X∞ (γ2 ) and Y∞ (γ1 ) ≥ Y∞ (γ2 ). Thus X∞ and
Y∞ are decreasing functions of γ, as is ρ(X∞ Y∞ ). At γ = γopt , any one of the three
conditions in Theorem 14.1 can fail. If only condition (iii) fails, then it is relatively
straightforward to show that the descriptor formulas below for γ = γopt are optimal;
that is, the optimal controller is given by
−2
(I − γopt Y∞ X∞ )x̂˙ = As x̂ − L∞ y (14.16)
u = F∞ x̂ (14.17)
14.3. Optimality and Limiting Behavior 283

−2 −2 −2
where As := A + B2 F∞ + L∞ C2 + γopt Y∞ A∗ X∞ + γopt B1 B1∗ X∞ + γopt Y∞ C1∗ C1 . (See
Example 14.3.)
The formulas in Theorem 14.1 are not well-defined in the optimal case because the
−2
term (I − γopt X∞ Y∞ ) is not invertible. It is possible but far less likely that conditions
(i) or (ii) would fail before (iii). To see this, consider (i) and let γ1 be the largest γ
for which H∞ fails to be in dom(Ric) because the H∞ matrix fails to have either the
stability property or the complementarity property. The same remarks will apply to (ii)
by duality.
If complementarity fails at γ = γ1 , then ρ(X∞ ) → ∞ as γ → γ1 . For γ < γ1 ,
H∞ may again be in dom(Ric), but X∞ will be indefinite. For such γ, the controller
u = −B2∗ X∞ x would make kTzw k∞ < γ but would not be stabilizing. (See part 1
of Example 14.3.) If the stability property fails at γ = γ1 , then H∞ 6∈ dom(Ric)
but Ric can be extended to obtain X∞ so that a controller can be obtained to make
kTzw k∞ = γ1 . The stability property will also not hold for any γ ≤ γ1 , and no controller
whatsoever exists that makes kTzw k∞ < γ1 . In other words, if stability breaks down
first, then the infimum over stabilizing controllers equals the infimum over all controllers,
stabilizing or otherwise. (See part 2 of Example 14.3.) In view of this, we would typically
expect that complementarity would fail first.
Complementarity failing at γ = γ1 means ρ(X∞ ) → ∞ as γ → γ1 , so condition (iii)
would fail at even larger values of γ, unless the eigenvectors associated with ρ(X∞ ) as
γ → γ1 are in the null space of Y∞ . Thus condition (iii) is the most likely of all to fail
first. If condition (i) or (ii) fails first because the stability property fails, the formulas
in Theorem 14.1 as well as their descriptor versions are optimal at γ = γopt . This is
illustrated in Example 14.3. If the complementarity condition fails first, [but (iii) does
not fail], then obtaining formulas for the optimal controllers is a more subtle problem.

Example 14.3 Let an interconnected dynamical system realization be given by

 h i 
a 1 0 1
 " # " # 
 
 1 0 
G(s) =  0 .
 0 1 
 h i 
1 0 1 0

Then all assumptions for output feedback problem are satisfied and
" 2
# " #
1−γ 2
a 1−γ γ 2 a γ 2
H∞ = , J∞ = .
−1 −a −1 −a

The eigenvalues of H∞ and J∞ are given by

( p )
(a2 + 1)γ 2 − 1
± .
γ
284 H∞ CONTROL

1
If γ > √ , then X− (H∞ ) and X− (J∞ ) exist and
2
a +1
" √ 2 #
(a +1)γ −1−aγ
2

X− (H∞ ) = Im γ
1
" √ #
(a2 +1)γ 2 −1−aγ
X− (J∞ ) = Im γ .
1
We shall consider two cases:

1) a > 0: In this case, the complementary property of dom(Ric) will fail before the
stability property fails since
p
(a2 + 1)γ 2 − 1 − aγ = 0

when γ = 1.
1
Nevertheless, if γ > √ and γ 6= 1, then H∞ ∈ dom(Ric) and
2
a +1
(
γ > 0; if γ > 1
X∞ = p =
(a + 1)γ − 1 − aγ
2 2 < 0; if √a12 +1 < γ < 1.

Note that if γ > 1, then H∞ ∈ dom(Ric), J∞ ∈ dom(Ric), and

γ
X∞ = p >0
(a2 + 1)γ 2 − 1 − aγ
γ
Y∞ = p > 0.
(a + 1)γ 2 − 1 − aγ
2

Hence conditions (i) and (ii) in Theorem 14.1 are satisfied, and we need to check
condition (iii). Since

γ2
ρ(X∞ Y∞ ) = p ,
( (a2 + 1)γ 2 − 1 − aγ)2

it is clear that ρ(X∞ Y∞ ) → ∞ when γ → 1. So condition (iii) will fail before

condition (i) or (ii) fails.
2) a < 0: In this case, the complementary property is always satisfied, and, further-
more, H∞ ∈ dom(Ric), J∞ ∈ dom(Ric), and
γ
X∞ = p >0
(a2 + 1)γ 2 − 1 − aγ
14.3. Optimality and Limiting Behavior 285

γ
Y∞ = p >0
(a2 + 1)γ 2 − 1 − aγ
1
for γ > √ .
2
a +1
1
However, for γ ≤ √ , H∞ 6∈ dom(Ric) since stability property fails. Nev-
a2 + 1
1
ertheless, in this case, if γ0 = √ , we can extend the dom(Ric) to include
a2 + 1
those matrices H∞ with imaginary axis eigenvalues as
" #
−a
X− (H∞ ) = Im
1
1
such that X∞ = − is a solution to the Riccati equation
a
A∗ X∞ + X∞ A + C1∗ C1 + γ0−2 X∞ B1 B1∗ X∞ − X∞ B2 B2∗ X∞ = 0
and A + γ0−2 B1 B1∗ X∞ − B2 B2∗ X∞ = 0. It can be shown that
γ2
ρ(X∞ Y∞ ) = p < γ2
( (a2 + 1)γ 2 − 1 − aγ)2
is satisfied if and only if
p  
1
γ> a2 + 2 + a > √ .
a2 + 1
So condition (iii) of Theorem 14.1 will fail before either (i) or (ii) fails.
In both a > 0 and a < 0 cases, the optimal γ for the output feedback is given by
p
γopt = a2 + 2 + a
and the optimal controller given by the descriptor formula in equations (14.16) and (14.17)
is a constant. In fact,
γopt
uopt = − q y.
2 2 − 1 − aγ
(a + 1)γopt opt
√
For instance, let a = −1 then γopt = 3 − 1 = 0.7321 and uopt = −0.7321 y. Further,
 
−1.7321 1 −0.7321
 
Tzw =  1 0 0 .
−0.7321 0 −0.7321
It is easy to check that kTzw k∞ = 0.7321.
286 H∞ CONTROL

14.4 Minimum Entropy Controller

Let T be a transfer matrix with kT k∞ < γ. Then the entropy of T (s) is defined by
Z
γ2 ∞ 

I(T, γ) = − ln det I − γ −2 T ∗ (jω)T (jω)  dω.
2π −∞

It is easy to see that

Z X
γ2 ∞  
I(T, γ) = − ln 1 − γ −2 σi2 (T (jω)) dω
2π −∞ i

and I(T, γ) ≥ 0, where σi (T (jω)) is the ith singular value of T (jω). It is also easy to
show that Z ∞X
1 2
lim I(T, γ) = σ2 (T (jω)) dω = kT k2 .
γ→∞ 2π −∞ i i

Thus the entropy I(T, γ) is, in fact , a performance index measuring the tradeoff between
the H∞ optimality (γ → kT k∞ ) and the H2 optimality (γ → ∞).
It has been shown in Glover and Mustafa [1989] that the suboptimal controller given
in Theorem 14.1 is actually the controller that satisfies the norm condition kTzw k∞ < γ
and minimizes the following entropy:
Z
γ 2 ∞ 

− ln det I − γ −2 Tzw
∗
(jω)Tzw (jω)  dω.
2π −∞

Therefore, the given suboptimal controller is also called the minimum entropy controller
˜ γ) = −I(T, γ)].
[maximum entropy controller if the entropy is defined as I(T,

Related MATLAB Commands: hinfsyne, hinffi

14.5 An Optimal Controller

To offer a general idea about the appearance of an optimal controller, we shall give in
the following (without proof) the conditions under which an optimal controller exists
and an explicit formula for an optimal controller.

Theorem 14.6 There exists an admissible controller such that kTzw k∞ ≤ γ iff the
following three conditions hold:

(i) There exists a full column rank matrix

" #
X∞1
∈ R2n×n
X∞2
14.5. An Optimal Controller 287

such that " # " #

X∞1 X∞1
H∞ = TX , Re λi (TX ) ≤ 0 ∀i
X∞2 X∞2

and
∗ ∗
X∞1 X∞2 = X∞2 X∞1 ;

(ii) There exists a full column rank matrix

" #
Y∞1
∈ R2n×n
Y∞2

such that " # " #

Y∞1 Y∞1
J∞ = TY , Re λi (TY ) ≤ 0 ∀i
Y∞2 Y∞2

and
∗ ∗
Y∞1 Y∞2 = Y∞2 Y∞1 ;

(iii)
" #
∗
X∞2 X∞1 γ −1 X∞2
∗
Y∞2
−1 ∗ ∗
≥ 0.
γ Y∞2 X∞2 Y∞2 Y∞1

Moreover, when these conditions hold, one such controller is

Kopt (s) := CK (sEK − AK )+ BK

where
∗
EK := Y∞1 X∞1 − γ −2 Y∞2
∗
X∞2
∗ ∗
BK := Y∞2 C2
CK := −B2∗ X∞2
AK := EK TX − BK C2 X∞1 = TY∗ EK + Y∞1
∗
B2 CK .

Remark 14.2 It is simple to show that if X∞1 and Y∞1 are nonsingular and if
−1 −1
X∞ = X∞2 X∞1 and Y∞ = Y∞2 Y∞1 , then condition (iii) in the preceding theorem
is equivalent to X∞ ≥ 0, Y∞ ≥ 0, and ρ(Y∞ X∞ ) ≤ γ 2 . So, in this case, the conditions
for the existence of an optimal controller can be obtained from “taking the limit” of
the corresponding conditions in Theorem 14.1. Moreover, the controller given above is
reduced to the descriptor form given in equations (14.16) and (14.17). 3
288 H∞ CONTROL

14.6 General H∞ Solutions

Consider the system described by the block diagram
z w
G

y u
- K

where, as usual, G and K are assumed to be real rational and proper with K constrained
to provide internal stability. The controller is said to be admissible if it is real rational,
proper, and stabilizing. Although we are taking everything to be real, the results
presented here are still true for the complex case with some obvious modifications. We
will again only be interested in characterizing all suboptimal H∞ controllers.
The realization of the transfer matrix G is taken to be of the form
 
A B1 B2 " #
  A B
G(s) =  C1 D11 D12  = ,
C D
C2 D21 0

which is compatible with the dimensions of z(t) ∈ Rp1 , y(t) ∈ Rp2 , w(t) ∈ Rm1 ,
u(t) ∈ Rm2 , and the state x(t) ∈ Rn . The following assumptions are made:

(A1) (A, B2 ) is stabilizable and (C2 , A) is detectable;

" #
0 h i
(A2) D12 = and D21 = 0 I ;
I
" #
A − jωI B2
(A3) has full column rank for all ω;
C1 D12
" #
A − jωI B1
(A4) has full row rank for all ω.
C2 D21

Assumption (A1) is necessary for the existence of stabilizing controllers. The as-
sumptions in (A2) mean that the penalty on z = C1 x + D12 u includes a nonsingular,
normalized penalty on the control u, that the exogenous signal w includes both plant
disturbance and sensor noise, and that the sensor noise weighting is normalized and
nonsingular. Relaxation of (A2) leads to singular control problems; see Stroorvogel
[1992]. For those problems that have D12 full column rank and D21 full row rank but
do not satisfy assumption (A2), a normalizing procedure is given in the next section so
that an equivalent new system will satisfy this assumption.
14.6. General H∞ Solutions 289

Assumptions (A3) and (A4) are made for a technical reason: Together with (A1)
they guarantee that the two Hamiltonian matrices in the corresponding H2 problem
belong to dom(Ric), as we have seen in Chapter 13. Dropping (A3) and (A4) would
make the solution very complicated. A further discussion of the assumptions and their
possible relaxation will be provided in Section 14.7.
The main result is now stated in terms of the solutions of the X∞ and Y∞ Riccati
equations together with the “state feedback” and “output injection” matrices F and L.

" #
∗ γ 2 Im1 0
R := D1• D1• − , where D1• := [D11 D12 ]
0 0
" # " #
∗ γ 2 Ip1 0 D11
R̃ := D•1 D•1 − , where D•1 :=
0 0 D21
" # " #
A 0 B h i
H∞ := − R−1 ∗
D1• C1 B∗
−C1∗ C1 −A∗ −C1∗ D1•
" # " #
A∗ 0 C∗ h i
J∞ := − R̃−1 D•1 B1∗ C
−B1 B1∗ −A ∗
−B1 D•1

X∞ := Ric(H∞ ) Y∞ := Ric(J∞ )

" #
F1∞
F := := −R−1 [D1•
∗
C1 + B ∗ X∞ ]
F2∞
h i
∗
L := L1∞ L2∞ := −[B1 D•1 + Y∞ C ∗ ]R̃−1

Partition D, F1∞ , and L1∞ are as follows:

 
∗ ∗ ∗
F11∞ F12∞ F2∞
" #  
F0  L∗11∞ D1111 D1112 0 
=  .
L0 D  L∗ 
 12∞ D1121 D1122 I 
L∗2∞ 0 I 0

Remark 14.3 In the above matrix partitioning, some matrices may not exist depend-
ing on whether D12 or D21 is square. This issue will be discussed further later. For the
time being, we shall assume that all matrices in the partition exist. 3

Theorem 14.7 Suppose G satisfies the assumptions (A1)–(A4).

(a) There exists an admissible controller K(s) such that ||F` (G, K)||∞ < γ (i.e.,
kTzw k∞ < γ) if and only if
290 H∞ CONTROL

∗ ∗
(i) γ > max(σ[D1111 , D1112 , ], σ[D1111 , D1121 ]);
(ii) H∞ ∈ dom(Ric) with X∞ = Ric(H∞ ) ≥ 0;
(iii) J∞ ∈ dom(Ric) with Y∞ = Ric(J∞ ) ≥ 0;
(iv) ρ(X∞ Y∞ ) < γ 2 .

(b) Given that the conditions of part (a) are satisfied, then all rational internally
stabilizing controllers K(s) satisfying ||F` (G, K)||∞ < γ are given by

K = F` (M∞ , Q) for arbitrary Q ∈ RH∞ such that kQk∞ < γ

where  
 B̂1 B̂2
 
M∞ = 
 Ĉ1 D̂11 D̂12 

Ĉ2 D̂21 0
∗ ∗
D̂11 = −D1121 D1111 (γ 2 I − D1111 D1111 )−1 D1112 − D1122
D̂12 ∈ Rm2 ×m2 and D̂21 ∈ Rp2 ×p2 are any matrices (e.g., Cholesky factors) satisfying
∗ ∗
D̂12 D̂12 = I − D1121 (γ 2 I − D1111 D1111 )−1 D1121
∗
,
∗ ∗ ∗ −1
D̂21 D̂21 = I − D1112 (γ I − D1111 D1111 ) D1112 ,
2

and

B̂2 = Z∞ (B2 + L12∞ )D̂12 ,

Ĉ2 = −D̂21 (C2 + F12∞ ),
−1
B̂1 = −Z∞ L2∞ + B̂2 D̂12 D̂11 ,
−1
Ĉ1 = F2∞ + D̂11 D̂21 Ĉ2 ,
−1
 = A + BF + B̂1 D̂21 Ĉ2

where
Z∞ = (I − γ −2 Y∞ X∞ )−1 .

(Note that if D11 = 0 then the formulas are considerably simplified.)

Some Special Cases:

Case 1: D12 = I
In this case

1. In part (a), (i) becomes γ > σ(D1121 ).

14.7. Relaxing Assumptions 291

2. In part (b)

D̂11 = −D1122
∗
D̂12 D̂12 = I − γ −2 D1121 D1121
∗

∗
D̂21 D̂21 = I.

Case 2: D21 = I
In this case

1. In part (a), (i) becomes γ > σ(D1112 ).

2. In part (b)

D̂11 = −D1122
∗
D̂12 D̂12 = I
∗
D̂21 D̂21 = I − γ −2 D1112
∗
D1112 .

Case 3: D12 = I & D21 = I

In this case

1. In part (a), (i) drops out.

2. In part (b)

D̂11 = −D1122
∗
D̂12 D̂12 = I
∗
D̂21 D̂21 = I.

14.7 Relaxing Assumptions

In this section, we indicate how the results of Section 14.6 can be extended to more
general cases. Let a given problem have the following diagram, where zp (t) ∈ Rp1 ,
yp (t) ∈ Rp2 , wp (t) ∈ Rm1 , and up (t) ∈ Rm2 :

zp wp
P

yp up
- Kp
292 H∞ CONTROL

The plant P has the following state-space realization with Dp12 full column rank and
Dp21 full row rank:
 
Ap Bp1 Bp2
 
P (s) =  Cp1 Dp11 Dp12  .
Cp2 Dp21 Dp22

The objective is to find all rational proper controllers Kp (s) that stabilize P and
||F` (P, Kp )||∞ < γ. To solve this problem, we first transform it to the standard one
treated in the last section. Note that the following procedure can also be applied to the
H2 problem (except the procedure for the case D11 6= 0).

Normalize D12 and D21

Perform singular value decompositions to obtain the matrix factorizations
" #
0 h i
Dp12 = Up Rp , Dp21 = R̃p 0 I Ũp
I

such that Up and Ũp are square and unitary and Rp and R̃p are square and invertible.
Now let
zp = Up z, wp = Ũp∗ w, yp = R̃p y, up = Rp−1 u
and
K(s) = Rp Kp (s)R̃p

" # " #
Up∗ 0 Ũp∗ 0
G(s) = P (s)
0 R̃p−1 0 Rp−1
 
Ap Bp1 Ũp∗ Bp2 Rp−1
 
=  U ∗C Up∗ Dp11 Ũp∗ Up∗ Dp12 Rp−1 
 p p1 
−1 −1 ∗ −1 −1
R̃p Cp2 R̃p Dp21 Ũp R̃p Dp22 Rp
 
A B1 B2 " #
  A B
=:  C1 D11 D12  = .
C D
C2 D21 D22

Then " #
0 h i
D12 = D21 = 0 I ,
I

and the new system is as follows:

14.7. Relaxing Assumptions 293

z w
G

y u
- K

Furthermore, ||F` (P, Kp )||α = ||Up F` (G, K)Ũp ||α = kF` (G, K)kα for α = 2 or ∞
since Up and Ũp are unitary. Moreover, Assumptions (A1), (A3), and (A4) remain
unaffected.

Remove the Assumption D22 = 0

Suppose K(s) is a controller for G with D22 set to zero. Then the controller for D22 6= 0
is K(I + D22 K)−1 . Hence there is no loss of generality in assuming D22 = 0.

Relax (A3) and (A4)

Suppose that
 
0 0 1
 
G= 0 0 1 
1 1 0

which violates both (A3) and (A4) and corresponds to the robust stabilization of an
integrator. If the controller u = −x, where  > 0 is used, then

−s
Tzw = , with kTzw k∞ = .
s+

Hence the norm can be made arbitrarily small as  → 0, but  = 0 is not admissible
since it is not stabilizing. This may be thought of as a case where the H∞ optimum is
not achieved on the set of admissible controllers. Of course, for this system, H∞ optimal
control is a silly problem, although the suboptimal case is not obviously so.

Relax (A1)
If assumption (A1) is violated, then it is obvious that no admissible controllers exist.
Suppose (A1) is relaxed to allow unstabilizable and/or undetectable modes on the jω
axis and internal stability is also relaxed to also allow closed-loop jω axis poles, but
(A2)–(A4) is still satisfied. It can be easily shown that under these conditions the
closed-loop H∞ norm cannot be made finite and, in particular, that the unstabilizable
and/or undetectable modes on the jω axis must show up as poles in the closed-loop
system.
294 H∞ CONTROL

Violate (A1) and either or both (A3) and (A4)

Sensible control problems can be posed that violate (A1) and either or both (A3) and
(A4). For example, cases when A has modes at s = 0 that are unstabilizable through
B2 and/or undetectable through C2 arise when an integrator is included in a weight
on a disturbance input or an error term. In these cases, either (A3) or (A4) are also
violated, or the closed-loop H∞ norm cannot be made finite. In many applications, such
problems can be reformulated so that the integrator occurs inside the loop (essentially
using the internal model principle) and is hence detectable and stabilizable. We will
show this process in the next section.
An alternative is to introduce  perturbations so that (A1), (A3), and (A4) are
satisfied. Roughly speaking, this would produce sensible answers for sensible problems,
but the behavior as  → 0 could be problematic.

Relax (A2)
In the cases that either D12 is not full column rank or that D21 is not full row rank,
improper controllers can give a bounded H∞ norm for Tzw , although the controllers
will not be admissible by our definition. Such singular filtering and control problems
have been well-studied in H2 theory and many of the same techniques go over to the
H∞ case (e.g., Willems [1981] and Willems, Kitapci, and Silverman [1986]). A complete
solution to the singular problem can be found in Stroorvogel [1992].

14.8 H2 and H∞ Integral Control

It is interesting to note that the H2 and H∞ design frameworks do not, in general,
produce integral control. In this section we show how to introduce integral control into
the H2 and H∞ design framework through a simple disturbance rejection problem. We
consider a feedback system shown in Figure 14.5. We shall assume that the frequency
contents of the disturbance w are effectively modeled by the weighting Wd ∈ RH∞ and
the constraints on control signal are limited by an appropriate choice of Wu ∈ RH∞ .
In order to let the output y track the reference signal r, we require K to contain an
integrator [i.e., K(s) has a pole at s = 0]. (In general, K is required to have poles on
the imaginary axis.)
There are several ways to achieve the integral design. One approach is to introduce
an integral in the performance weight We . Then the transfer function between w and
z1 is given by
z1 = We (I + P K)−1 Wd w.
Now if the resulting controller K stabilizes the plant P and makes the norm (2-norm
or ∞-norm) between w and z1 finite, then K must have a pole at s = 0 that is the zero
of the sensitivity function (assuming Wd has no zeros at s = 0). (This follows from the
well-known internal model principle.) The problem with this approach is that the H∞
(or H2 ) control theory presented in this chapter and in the previous chapters cannot be
14.8. H2 and H∞ Integral Control 295

6
z2 w
?
Wu Wd
6
- ?
r - g y z
- K - P g - We -1
−6 u

Figure 14.5: A simple disturbance rejection problem

applied to this problem formulation directly because the pole s = 0 of We becomes an

uncontrollable pole of the feedback system and the very first assumption for the H∞
(or H2 ) theory is violated.
However, the obstacles can be overcome by appropriately reformulating the problem.
Suppose We can be factorized as follows:

We = W̃e (s)M (s)

where M (s) is proper, containing all the imaginary axis poles of We , and
M −1 (s) ∈ RH∞ , W̃e (s) is stable and minimum phase. Now suppose there exists a
controller K(s) that contains the same imaginary axis poles that achieves the perfor-
mance specifications. Then, without loss of generality, K can be factorized as

K(s) = −K̂(s)M (s)

such that there is no unstable pole/zero cancellation in forming the product K̂(s)M (s).
Now the problem can be reformulated as in Figure 14.6. Figure 14.6 can be put in the
general LFT framework as in Figure 14.7 with
 " # " # 
W̃e M Wd W̃e M P
 
G(s) =  0 Wu .
M Wd MP

We shall illustrate this design through a simple numerical example. Let

 
0 1 0
s−2  
P = = 3 2 1  , Wd = 1,
(s + 1)(s − 3)
−2 1 0
" #
s + 10 −100 −90 1
Wu = = , We = .
s + 100 1 1 s
296 H∞ CONTROL

6z2 w
?
Wu Wd
6
- ?
y1 z
- K̂ - P g - M - W̃ -1
e
u

Figure 14.6: Disturbance rejection with imaginary axis poles

w
z1  W̃e  Wd 

 
z2  Wu  

?
M  g P 
y1 u

- K̂

Figure 14.7: LFT framework for the disturbance rejection problem

Then we can choose without loss of generality that

s+α 1
M= , W̃e = , α > 0.
s s+α
This gives the following generalized system:
 
−α 0 1 −2 1 1 0
 
 0 −100 0 0 0 0 −90 
 
 0 0 −2α α 
 0 α 0 
 
 0 0 0 0 1 0 0 
G(s) = 
 0
.

 0 0 3 2 0 1 
 
 1 0 0 0 0 0 0 
 
 0 
 1 0 0 0 0 1 
0 0 1 −2 1 1 0
14.9. H∞ Filtering 297

The suboptimal H∞ controller K̂∞ for each α can be computed easily as

−2060381.4(s + 1)(s + α)(s + 100)(s − 0.1557)
K̂∞ = ,
(s + α)2 (s + 32.17)(s + 262343)(s − 19.89)

which gives the closed-loop H∞ norm 7.854. Hence the controller K∞ = −K̂∞ (s)M (s)
is given by
2060381.4(s + 1)(s + 100)(s − 0.1557) 7.85(s + 1)(s + 100)(s − 0.1557)
K∞ (s) = ≈ ,
s(s + 32.17)(s + 262343)(s − 19.89) s(s + 32.17)(s − 19.89)
which is independent of α as expected. Similarly, we can solve an optimal H2 controller
−43.487(s + 1)(s + α)(s + 100)(s − 0.069)
K̂2 =
(s + α)2 (s2 + 30.94s + 411.81)(s − 7.964)
and
43.487(s + 1)(s + 100)(s − 0.069)
K2 (s) = −K̂2 (s)M (s) = .
s(s2 + 30.94s + 411.81)(s − 7.964)

An approximate integral control can also be achieved without going through the
preceding process by letting
1
We = W̃e = , M (s) = 1
s+
for a sufficiently small  > 0. For example, a controller for  = 0.001 is given by
316880(s + 1)(s + 100)(s − 0.1545) 7.85(s + 1)(s + 100)(s − 0.1545)
K∞ = ≈ ,
(s + 0.001)(s + 32)(s + 40370)(s − 20) s(s + 32)(s − 20)
which gives the closed-loop H∞ norm of 7.85. Similarly, an approximate H2 integral
controller is obtained as
43.47(s + 1)(s + 100)(s − 0.0679) 43.47(s + 1)(s + 100)(s − 0.0679)
K2 = ≈ .
(s + 0.001)(s + 30.93s + 411.7)(s − 7.972)
2 s(s2 + 30.93s + 411.7)(s − 7.972)

14.9 H∞ Filtering
In this section we show how the filtering problem can be solved using the H∞ theory
developed earlier. Suppose a dynamic system is described by the following equations:

ẋ = Ax + B1 w(t), x(0) = 0 (14.18)

y = C2 x + D21 w(t) (14.19)
z = C1 x + D11 w(t) (14.20)

The filtering problem is to find an estimate ẑ of z in some sense using the measurement
of y. The restriction on the filtering problem is that the filter has to be causal so
298 H∞ CONTROL

that it can be realized (i.e., ẑ has to be generated by a causal system acting on the
measurements). We will further restrict our filter to be unbiased; that is, given T > 0
the estimate ẑ(t) = 0 ∀t ∈ [0, T ] if y(t) = 0, ∀t ∈ [0, T ]. Now we can state our H∞
filtering problem.
H∞ Filtering: Given a γ > 0, find a causal filter F (s) ∈ RH∞ if it exists
such that
2
kz − ẑk2
J := sup 2 < γ2
w∈L2 [0,∞) kwk2

with ẑ = F (s)y.
A diagram for the filtering problem is shown in Figure 14.8.

z   w
A B1
 
 C1 D11 
z∆ ? ẑ y
 j F (s)  C2 D21
−

Figure 14.8: Filtering problem formulation

The preceding filtering problem can also be formulated in an LFT framework: Given
a system shown below
z∆ w  
G(s) A B1 0
  
G(s) =  C1 D11 −I 
y ẑ
- F (s) C2 D21 0

find a filter F (s) ∈ RH∞ such that

2
kz∆ k2
sup 2 < γ2. (14.21)
w∈L2 kwk2
Hence the filtering problem can be regarded as a special H∞ problem. However, com-
pared with control problems, there is no internal stability requirement in the filtering
problem. Hence the solution to the above filtering problem can be obtained from the H∞
solution in the previous sections by setting B2 = 0 and dropping the internal stability
requirement.
Theorem 14.8 Suppose (C2 , A) is detectable and
" #
A − jωI B1
C2 D21
14.10. Notes and References 299

has full row rank for all ω. Let D21 be normalized and D11 partitioned conformably as
" # " #
D11 D111 D112
= .
D21 0 I

Then there exists a causal F (s) ∈ RH∞ such that J < γ 2 if and only if σ(D111 ) < γ
and J∞ ∈ dom(Ric) with Y∞ = Ric(J∞ ) ≥ 0, where
" #" #∗ " #
D11 D11 γ2I 0
R̃ := −
D21 D21 0 0
" # " # " #
A∗ 0 C1∗ C2∗ −1 D11 B1∗ C1
J∞ := − R̃ .
−B1 B1∗ −A ∗
−B1 D11 ∗
−B1 D21 D21 B1∗ C2

Moreover, if the above conditions are satisfied, then a rational causal filter F (s)
satisfying J < γ 2 is given by
" #
A + L2∞ C2 + L1∞ D112 C2 −L2∞ − L1∞ D112
ẑ = F (s)y = y
C1 − D112 C2 D112
where h i h i
L1∞ L2∞ := − ∗
B1 D11 + Y∞ C1∗ ∗
B1 D21 + Y∞ C2∗ R̃−1 .
∗
In the case where D11 = 0 and B1 D21 = 0 the filter becomes much simpler:
" #
A − Y∞ C2∗ C2 Y∞ C2∗
ẑ = y
C1 0
where Y∞ is the stabilizing solution to
Y∞ A∗ + AY∞ + Y∞ (γ −2 C1∗ C1 − C2∗ C2 )Y∞ + B1 B1∗ = 0.

14.10 Notes and References

The first part of this chapter is based on Sampei, Mita, and Nakamichi [1990], Packard
[1994], Doyle, Glover, Khargonekar, and Francis [1989], and Chen and Zhou [1996].
The proof of Theorem 14.4 comes from Ran and Vreugdenhil [1988]. The detailed
derivation of the H∞ solution for the general case is treated in Glover and Doyle [1988,
1989]. A fairly complete solution to the singular H∞ problem is obtained in Stoorvogel
[1992]. The H∞ filtering and smoothing problems are considered in detail in Nagpal
and Khargonekar [1991]. There is a rich body of literature on the LMI approach to H∞
control and related problems. In particular, readers are referred to the monograph by
Boyd, Ghaoui, Feron, and Balakrishnan [1994] for a comprehensive treatment of LMIs
and their applications in control. See also the paper by Chilali and Gahinet [1996] for
an application of LMIs in H∞ control with closed-loop pole constraints.
300 H∞ CONTROL

14.11 Problems
Problem 14.1 Figure 14.9 shows a single-loop analog feedback system. The plant is

6 6
z1 w2 z2
?
W 2 1

6 6
w1 e ?y u
- j - F - j - K - P -
−6

Figure 14.9: Analog feedback system

P and the controller K; F is an antialiasing filter for future digital implementation of

the controller (it is a good idea to include F at the start of the analog design so that
there are no surprises later due to additional phase lag). The basic control specification
is to get good tracking over a certain frequency range, say [0, ω1 ]; that is, to make the
magnitude of the transfer function from w1 to e small over this frequency range. The
weighted tracking error is z1 in the figure, where the weight W is selected to be a low-
pass filter with bandwidth ω1 . We could attempt to minimize the H∞ norm from w1
to z1 , but this problem is not regular. To regularize it, another input, w2 , is added and
another signal, z2 , is penalized. The two weights 1 and 2 are small positive scalars.
The design problem is to minimize the H∞ norm
" # " #
w1 z1
from w = to z = .
w2 z2

Figure 14.9 can then be converted to the standard setup by stacking the states of P , F ,
and W to form the state of G.
The plant transfer function is taken to be
20 − s
P (s) = .
(s + 0.01)(s + 20)
With a view toward subsequent digital control with h = 0.5, the filter F is taken to
have bandwidth π/0.5, and Nyquist frequency ωN :
1
F (s) = .
(0.5/π)s + 1
14.11. Problems 301

The weight W is then taken to have bandwidth one-fifth the Nyquist frequency:
 2
1
W (s) = .
(2.5/π)s + 1
Finally, 1 and 2 are both set to 0.01.
Run hinfsyn and show your plots of the closed-loop frequency responses.
Problem 14.2 Make the same assumptions as in Chapter 13 for H2 control and derive
the H∞ controller parameterization by using the normalization procedure and Theo-
rem 14.7.
Problem 14.3 Consider the feedback system in Figure 6.3 and suppose
s − 10 1 s+2
P = , We = , Wu = .
(s + 1)(s + 10) s + 0.001 s + 10
Design a controller that minimizes
" #

We So
.
Wu KSo
∞

Simulate the time response of the system when r is a step.

Problem 14.4 Repeat Problem 14.3 when We = 1/s.
Problem 14.5 Consider again Problem 13.5 and design a controller that minimizes
the H∞ norm of the transfer matrix from r to (e, uw ).
Problem 14.6 Repeat Problem 14.5 with W =  for  = 0.01 and 0.0001. Study the
behavior of the controller when  → 0.
Problem 14.7 Repeat Problem 14.5 and Problem 14.6 with
10(2 − s)
P = .
(s + 1)3
Problem 14.8 Let N ∈ RH−
∞ . The Nehari problem is the following approximation
problem:
inf kN − Qk∞ .
Q∈RH∞

Formulate the Nehari problem as a standard H∞ control problem.

Problem 14.9 Consider a generalized plant
 
−4 25 0.8 −1
 
 −10 29 0.9 −1 
P = 
−25 
 10 0 1 
13 25 1 0
302 H∞ CONTROL

with an SISO controller K. Find the optimal H∞ performance γopt . Calculate the
central controller for each γ ∈ (γopt , 2) and the corresponding H∞ performance, γ∞ .
Plot γ∞ versus γ. Is γ∞ monotonic with respect to γ? (See Ushida and Kimura [1996]
for a detailed discussion.)
" #
A B
Problem 14.10 Let a satellite model be given by Po (s) = , where
C 0
   
0 1 0 0 0
   
 0 0   −5 
A=
0 0  , B =  1.7319 × 10 ,
 0 0   
 0 1   0 
0 0 −1.5392 −2 × 0.003 × 1.539 3.7859 × 10−4
h i
C = 1 0 1 0 , D = 0.
Suppose the true system is described by
P = (N + ∆N )(M + ∆M )−1
where Po = N M −1 is a normalized coprime factorization. Design a controller so that
the controller stabilizes the largest
" #
∆N
∆= .
∆M
Problem 14.11 Consider the feedback system shown below and let
0.5(1 − s) s/1.245 + 1 s/0.502 + 1
P = , W1 = 50 , W2 = 0.1256 .
(s + 2)(s + 0.5) s/0.007 + 1 s/2 + 1

z1 w1 w2
- W2 - ∆
C
C
z2
e - K - P -CW ?
e - W1 -
− 6

Figure 14.10: System with additive uncertainty

Then " # " #" # " #

z1 −W2 KS −W2 KS w1 w1
= =M
z2 W1 S W1 S w2 w2
where S = (I + P K)−1 .
14.11. Problems 303

(a) Design a controller K such that

inf kM k∞ .
K stabilizing

(b) Design a controller K so that

" #
∆1
inf sup µ∆ (M ), ∆= .
K stabilizing ω ∆2

Note that µ∆ (M ) = |W1 S| + |W2 KS|.

Problem 14.12 Design a µ-synthesis controller for the HIMAT control problem in
Example 9.1.

Problem 14.13 Let G(s) ∈ H∞ be a square transfer matrix and α > 0. Show that G
is (extended)
strictly positive real (i.e., G∗ (jω) + G(jω) > 0, ∀ ω ∈ R ∪ {∞}) if and
only if (αI − G)(αI + G)−1 ∞ < 1.

Problem 14.14 Consider a generalized system

 
A B1 B2
 
G(s) =  C1 D11 D12 
C2 D21 D22

and suppose we want to find a controller K such that F` (G, K) is (extended) strictly
postive real. Show that the problem can be converted to a standard H∞ control problem
by using the transformation in the last problem.

Problem 14.15 (Synthesis Using Popov Criterion) A stability criterion by Popov in-
volves finding a controller and a multiplier matrix N such that

Q + (I + sN )F` (G, K)

is strictly positive real where Q is a constant matrix. Assume D11 = 0 and D12 = 0 in
the realization of G. Find the G̃ so that

Q + (I + sN )F` (G, K) = F` (G̃, K)

Hence the results in the last problem can be applied.

304 H∞ CONTROL
Chapter 15

Controller Reduction

We have shown in the previous chapters that the H∞ control theory and µ synthesis
can be used to design robust performance controllers for highly complex uncertain sys-
tems. However, since a great many physical plants are modeled as high-order dynamical
systems, the controllers designed with these methodologies typically have orders com-
parable to those of the plants. Simple linear controllers are normally preferred over
complex linear controllers in control system designs for some obvious reasons: They
are easier to understand and computationally less demanding; they are also easier to
implement and have higher reliability since there are fewer things to go wrong in the
hardware or bugs to fix in the software. Therefore, a lower-order controller should be
sought whenever the resulting performance degradation is kept within an acceptable
level. There are usually three ways to arrive at a lower-order controller. A seemingly
obvious approach is to design lower-order controllers directly based on the high-order
models. However, this is still largely an open research problem. Another approach is
first to reduce the order of a high-order plant and, second, based on the reduced plant
model, design a lower-order controller. A potential problem associated with this ap-
proach is that such a lower-order controller may not even stabilize the full-order plant
since the error information between the full-order model and the reduced-order model
is not considered in the design of the controller. On the other hand, one may seek to
design first a high-order, high-performance controller and subsequently proceed with a
reduction of the designed controller. This approach is usually referred to as controller
reduction. A crucial consideration in controller order reduction is to take into account
the closed loop so that closed-loop stability is guaranteed and the performance degrada-
tion is minimized with the reduced-order controllers. The purpose of this chapter is to
introduce several controller reduction methods that can guarantee closed-loop stability
and possibly closed-loop performance as well.

305
306 CONTROLLER REDUCTION

15.1 H∞ Controller Reductions

In this section, we consider an H∞ performance-preserving controller order reduction
problem. We consider the feedback system shown in Figure 15.1 with a generalized
plant realization given by
 
A B1 B2
 
G(s) =  C1 D11 D12  .
C2 D21 D22

z w
 
G(s)

y u
- K(s)

Figure 15.1: Closed-loop system diagram

The following assumptions are made:

(A1) (A, B2 ) is stabilizable and (C2 , A) is detectable;
(A2) D12 has full column rank and D21 has full row rank;
" #
A − jωI B2
(A3) has full column rank for all ω;
C1 D12
" #
A − jωI B1
(A4) has full row rank for all ω.
C2 D21

As stated in Chapter 14, all stabilizing controllers satisfying kTzw k∞ < γ can be pa-
rameterized as
K = F` (M∞ , Q), Q ∈ RH∞ , kQk∞ < γ (15.1)
where M∞ is of the form
 
" # Â B̂1 B̂2
M11 (s) M12 (s)  
M∞ = =  Ĉ1 D̂11 D̂12 
M21 (s) M22 (s)
Ĉ2 D̂21 D̂22

−1 −1
such that D̂12 and D̂21 are invertible and  − B̂2 D̂12 Ĉ1 and  − B̂1 D̂21 Ĉ2 are both
−1 −1
stable (i.e., M12 and M21 are both stable).
15.1. H∞ Controller Reductions 307

The problem to be considered here is to find a controller K̂ with a minimal possible

order such that the H∞ performance requirement F` (G, K̂) < γ is satisfied. This is
∞
clearly equivalent to finding a Q so that it satisfies the preceding constraint and the order
of K̂ is minimized. Instead of choosing Q directly, we shall approach this problem from
a different perspective. The following lemma is useful in the subsequent development
and can be regarded as a special case of Theorem 10.6 (main loop theorem).

Lemma 15.1 Consider a feedback system shown below

z w
N

y u
- Q

where N is a suitably partitioned transfer matrix

" #
N11 N12
N (s) = .
N21 N22

Then the closed-loop transfer matrix from w to z is given by

Tzw = F` (N, Q) = N11 + N12 Q(I − N22 Q)−1 N21 .

Assume that the feedback loop is well-posed [i.e., det(I − N22 (∞)Q(∞)) 6= 0] and either
N21 (jω) has full row rank for all ω ∈ R ∪ ∞ or N12 (jω) has full column rank for all
ω ∈ R ∪ ∞ and kN k∞ ≤ 1; then kF` (N, Q)k∞ < 1 if kQk∞ < 1.

Proof. We shall assume that N21 has full row rank. The case when N12 has full
column rank can be shown in the same way.
To show that kTzw k∞ < 1, consider the closed-loop system at any frequency s = jω
with the signals fixed as complex constant vectors. Let kQk∞ =:  < 1 and note that
Twy = N21+
(I − N22 Q), where N21
+
is a right inverse of N21 . Also let κ := kTwy k∞ . Then
kwk2 ≤ κkyk2, and kN k∞ ≤ 1 implies that kzk22 + kyk22 ≤ kwk22 + kuk22 . Therefore,

kzk22 ≤ kwk22 + (2 − 1)kyk22 ≤ [1 − (1 − 2 )κ−2 ]kwk22 ,

which implies kTzw k∞ < 1 . 2

308 CONTROLLER REDUCTION

15.1.1 Additive Reduction

Consider the class of (reduced-order) controllers that can be represented in the form
K̂ = K0 + W2 ∆W1 ,
where K0 may be interpreted as a nominal, higher-order controller, and ∆ is a stable
perturbation with stable, minimum phase and invertible weighting functions W1 and
W2 . Suppose that kF` (G, K0 )k∞ < γ. A natural question is whether it is possible to
obtain a reduced-order controller K̂ in this class such that the H∞ performance bound
remains valid when K̂ is in place of K0 . Note that this is somewhat a special case of the
preceding general problem: The specific form of K̂ means that K̂ and K0 must possess
the same right-half plane poles, thus to a certain degree limiting the set of attainable
reduced-order controllers.
Suppose K̂ is a suboptimal H∞ controller; that is, there is a Q ∈ RH∞ with
kQk∞ < γ such that K̂ = F` (M∞ , Q). It follows from simple algebra that

Q = F` (K̄a−1 , K̂)
where " # " #
0 I 0 I
K̄a−1 := −1
M∞ .
I 0 I 0
Furthermore, it follows from straightforward manipulations that


kQk∞ < γ ⇐⇒ F` (K̄a−1 , K̂) < γ
∞

⇐⇒ F` (K̄a−1 , K0 + W2 ∆W1 ) ∞ < γ


⇐⇒ F` (R̃, ∆) < 1
∞

where " #" #" #

γ −1/2 I 0 R11 R12 γ −1/2 I 0
R̃ =
0 W1 R21 R22 0 W2
and R is given by the star product
" # " #
R11 R12 Ko I
= S(K̄a−1 , ).
R21 R22 I 0

It is easy to see that R̃12 and R̃21 are both minimum phase and invertible and hence
have full column and full row rank, respectively, for all ω ∈ R ∪ ∞. Consequently,
by
invoking Lemma 15.1, we conclude that if R̃ is a contraction and k∆k∞ < 1, then

F` (R̃, ∆) < 1. This guarantees the existence of a Q such that kQk∞ < γ or,
∞

equivalently, the existence of a K̂ such that F` (G, K̂) < γ. This observation leads
∞
to the following theorem.
15.1. H∞ Controller Reductions 309

Theorem 15.2 Suppose W1 and W2 are stable, minimum phase and invertible transfer
matrices such that R̃ is a contraction. Let K0 be a stabilizing controller
such
that

kF` (G, K0 )k∞ < γ. Then K̂ is also a stabilizing controller such that F` (G, K̂) < γ
∞
if

k∆k∞ = W2−1 (K̂ − K0 )W1−1 < 1.
∞

Since R̃ can always be made contractive for sufficiently small W1 and W2 , there are
infinite many
W1 and W2 that satisfy the conditions in the theorem. It is obvious that

to make W2−1 (K̂ − K0 )W1−1 < 1 for some K̂, one would like to select the “largest”
∞
W1 and W2 such that R̃ is a contraction.

Lemma 15.3 Assume that kR22 k∞ < γ and define

 
" # 0 −R11 0 R12
 
L1 L2  −R11∼
0 ∼
R21 0  −1
L= = F (  , γ I).
L∼
`  −R22 
2 L3  0 R21 0 
∼ ∼
R12 0 −R22 0

Then R̃ is a contraction if W1 and W2 satisfy

" # " #
(W1∼ W1 )−1 0 L1 L2
∼ −1
≥ .
0 (W2 W2 ) L∼
2 L3

Proof. See Goddard and Glover [1993]. 2

An algorithm that maximizes det(W1∼ W1 ) det(W2 W2∼ ) has been developed by God-
dard and Glover [1993]. The procedure below, devised directly from the preceding
theorem, can be used to generate a required
reduced-order
controller that will preserve

the closed-loop H∞ performance bound F` (G, K̂) < γ.
∞

1. Let K0 be a full-order controller such that kF` (G, K0 )k∞ < γ;

2. Compute W1 and W2 so that R̃ is a contraction;

3. Use the weighted model reduction method
in Chapter 7 or any other methods to
−1 −1
find a K̂ so that W2 (K̂ − K0 )W1 < 1.
∞

Note that all controller reduction methods introduced in this book are only sufficient;
that is, there may be desired reduced-order controllers that cannot be found from the
proposed procedures.
310 CONTROLLER REDUCTION

15.1.2 Coprime Factor Reduction

The H∞ controller reduction problem can also be considered in the coprime factor
framework. For that purpose, we need the following alternative representation of all
admissible H∞ controllers:

Lemma 15.4 The family of all admissible controllers such that kTzw k∞ < γ can also
be written as

K(s) = F` (M∞ , Q) = (Θ11 Q + Θ12 )(Θ21 Q + Θ22 )−1 := U V −1

= (QΘ̃12 + Θ̃22 )−1 (QΘ̃11 + Θ̃21 ) := Ṽ −1 Ũ

where Q ∈ RH∞ , kQk∞ < γ, and U V −1 and Ṽ −1 Ũ are, respectively, right and left
coprime factorizations over RH∞ , and
 
" # −1
 − B̂1 D̂21 Ĉ2 −1
B̂2 − B̂1 D̂21 D̂22 −1
B̂1 D̂21
Θ11 Θ12  −1 −1 −1 
Θ= =  Ĉ1 − D̂11 D̂21 Ĉ2 D̂12 − D̂11 D̂21 D̂22 D̂11 D̂21 
Θ21 Θ22 −1 −1 −1
−D̂21 Ĉ2 −D̂21 D̂22 D̂21
 
" # −1
 − B̂2 D̂12 Ĉ1 −1
B̂1 − B̂2 D̂12 D̂11 −1
−B̂2 D̂12
Θ̃11 Θ̃12  −1 −1 −1 
Θ̃ = =  Ĉ2 − D̂22 D̂12 Ĉ1 D̂21 − D̂22 D̂12 D̂11 −D̂22 D̂12 
Θ̃21 Θ̃22 −1 −1 −1
D̂12 Ĉ1 D̂12 D̂11 D̂12
 −1 −1 −1

 − B̂2 D̂12 Ĉ1 B̂2 D̂12 B̂1 − B̂2 D̂12 D̂11
 
Θ−1 = −1
−D̂12 Ĉ1 −1
D̂12 −1
−D̂12 D̂11 
−1 −1 −1
Ĉ2 − D̂22 D̂12 Ĉ1 D̂22 D̂12 D̂21 − D̂22 D̂12 D̂11
 −1 −1 −1

 − B̂1 D̂21 Ĉ2 −B̂1 D̂21 B̂2 − B̂1 D̂21 D̂22
 
Θ̃−1 =  −1
D̂21 Ĉ2 −1
D̂21 −1
D̂21 D̂22 .
−1 −1 −1
Ĉ1 − D̂11 D̂21 Ĉ2 −D̂11 D̂21 D̂12 − D̂11 D̂21 D̂22

Proof. The results follow immediately from Lemma 9.2. 2

Theorem 15.5 Let K0 = Θ12 Θ−1 22 be the central H∞ controller such that
kF` (G, K0 )k∞ < γ and let Û , V̂ ∈ RH∞ with det V̂ (∞) 6= 0 be such that
" # " # " #!
γ −1 I √
0 −1 Θ12 Û
Θ − < 1/ 2. (15.2)
0 I Θ22 V̂
∞

Then K̂ = Û V̂ −1 is also a stabilizing controller such that kF` (G, K̂)k∞ < γ.
15.1. H∞ Controller Reductions 311

Proof. Note that by Lemma 15.4, K is an admissible controller such that kTzw k∞ < γ
if and only if there exists a Q ∈ RH∞ with kQk∞ < γ such that
" # " # " #
U Θ11 Q + Θ12 Q
:= =Θ (15.3)
V Θ21 Q + Θ22 I

and
K = U V −1 .
Hence, to show that K̂ = Û V̂ −1 with Û and V̂ satisfying equation (15.2) is also a
stabilizing controller such that kF` (G, K̂)k∞ < γ, we need to show that there is another
coprime factorization for K̂ = U V −1 and a Q ∈ RH∞ with kQk∞ < γ such that
equation (15.3) is satisfied.
Define " # " # " #!
γ −1 I 0 −1 Θ12 Û
∆ := Θ −
0 I Θ22 V̂
and partition ∆ as " #
∆U
∆ := .
∆V
Then " # " # " # " #
Û Θ12 γI 0 −γ∆U
= −Θ ∆=Θ
V̂ Θ22 0 I I − ∆V
and " # " #
Û(I − ∆V )−1 −γ∆U (I − ∆V )−1
=Θ .
V̂ (I − ∆V )−1 I

Define U := Û (I −∆V )−1 , V := V̂ (I −∆V )−1 , and Q := −γ∆U (I −∆V )−1 . Then U V −1
is another coprime factorization for K̂. To show that K̂ = U V −1
= Û V̂
−1
is a stabilizing

controller such that kF` (G, K̂)k∞ < γ, we need to show that γ∆U (I − ∆V )−1 ∞ < γ

or, equivalently, ∆U (I − ∆V )−1 ∞ < 1. Now
h  h i i −1
∆U (I − ∆V )−1 = ∆ I− 0 I ∆
I 0
 h i  
0 I 0 √
= F`  √ h √ i  , 2∆
I/ 2 0 I/ 2

and, by Lemma 15.1, ∆U (I − ∆V )−1 ∞ < 1 since
 h i 
0 I 0
 √ h √ i 
I/ 2 0 I/ 2
312 CONTROLLER REDUCTION

√
is a contraction and 2∆ ∞ < 1. 2

Similarly, we have the following theorem:

Theorem 15.6 Let K0 = Θ̃−1 22 Θ̃21 be the central H∞ controller such that
kF` (G, K0 )k∞ < γ and let Ũ , Ṽ ∈ RH∞ with det Ṽˆ (∞) 6= 0 be such that
ˆ ˆ
" #
h i h i −1 √
ˆ Ṽˆ −1 γ I 0
Θ̃21 Θ̃22 − Ũ Θ̃ < 1/ 2.
0 I ∞
−1
Then K̂ = Ṽˆ ˆ is also a stabilizing controller such that kF (G, K̂)k < γ.
Ũ ` ∞

The preceding two theorems show that the sufficient conditions for H∞ controller
reduction problem are equivalent to frequency-weighted H∞ model reduction problems.

H∞ Controller Reduction Procedures

(i) Let K0 = Θ12 Θ−1 −1
22 (= Θ̃22 Θ̃21 ) be a suboptimal H∞ central controller (Q = 0) such
that kTzw k∞ < γ.
−1
(ii) Find a reduced-order controller K̂ = Û V̂ −1 (or Ṽˆ Ũ ˆ ) such that
" # " # " #!
γ −1 I 0 √
Θ Û
Θ−1
12
− < 1/ 2
0 I Θ22 V̂
∞
or " #
h i h i γ −1 I √
ˆ −1 0
Θ̃21 Θ̃22 − Ũ Ṽˆ Θ̃ < 1/ 2.
0 I
∞
Then the closed-loop system with the reduced-order controller K̂ is stable and the
performance is maintained with the reduced-order controller; that is,


kTzw k∞ = F` (G, K̂) < γ.
∞

15.2 Notes and References

The main results presented in this chapter are based on the work of Goddard and
Glover [1993, 1994]. Other controller reduction methods include the stability-oriented
controller reduction criterion proposed by Enns [1984a, 1984b]; the weighted and un-
weighted coprime factor controller reduction methods studied by Liu and Anderson
[1986, 1990]; Liu, Anderson, and Ly [1990]; Anderson and Liu [1989]; and Anderson
[1993]; the normalized H∞ controller reduction studied by Mustafa and Glover [1991];
the normalized coprime factor method studied by McFarlane and Glover [1990] in the
H∞ loop-shaping setup; and the controller reduction in the ν-gap metric setup studied
by Vinnicombe [1993b]. Lenz, Khargonekar, and Doyle [1987] have also proposed an-
other H∞ controller reduction method with guaranteed performance for a class of H∞
problems.
15.3. Problems 313

15.3 Problems
Problem 15.1 Find a lower-order controller for the system in Example 10.4 when
γ = 2.

Problem 15.2 Find a lower-order controller for Problem 14.3 when γ = 1.1γopt where
γopt is the optimal norm.

Problem 15.3 Find a lower-order controller for the HIMAT control problem in Prob-
lem 14.12 when γ = 1.1γopt where γopt is the optimal norm. Compare the controller
reduction methods presented in this chapter with other available methods.

Problem 15.4 Let G be a generalized plant and K be a stabilizing controller. Let

∆ = diag(∆p , ∆k )"be a #suitably dimensioned
" # perturbation and let Tẑŵ be the transfer
w z
matrix from ŵ = to ẑ = in the following diagram:
w1 z1

z w
 
G 

zw z1 - f w1
W −1  - K

- (K̂ − K)W

Let W, W −1 ∈ H∞ be a given transfer matrix. Show that the following statements are
equivalent:
" # !
I 0
1. µ∆ Tẑŵ < 1;
0 W −1

2. kF` (G, K)k∞ < 1 and W −1 Fu (Tẑŵ , ∆p ) ∞ < 1 for all σ(∆p ) ≤ 1;
" # !
−1
I 0
3. W Tz1 w1 ∞ < 1 and F` T ẑ ŵ , ∆k < 1 for all σ(∆k ) ≤ 1.
0 W −1
∞

Problem 15.5 In the part "3 of Problem# 15.4, if we! let ∆k = (K̂ − K)W , then Tz1 w1 =
I 0
G22 (I − KG22 )−1 and F` Tẑŵ , ∆k = F` (G, K̂). Thus K̂ stabilizes the
0 W −1


system and satisfies F` (G, K̂) < 1 if k∆k k∞ = (K̂ − K)W ≤ 1 and part 2 of
∞ ∞
314 CONTROLLER REDUCTION

Problem 15.4 is satisfied by a controller K. Hence to reduce the order of the controller
K, it is sufficient to solve a frequency-weighted model reduction problem if W can be
calculated. In the single-input and single-output case, a “smallest” weighting function
W (s) can be calculated using part 2 of Problem 15.4 as follows:

|W (jω)| ≥ sup |Fu (Tẑŵ (jω), ∆p )|.

σ(∆p )≤1

Repeat Problems 15.1 and 15.2 using the foregoing method. (Hint: W can be computed
frequency by frequency using µ software and then fitted by a stable and minimum phase
transfer function.)

Problem 15.6 One way to generalize the method in Problem 15.5 to the MIMO case
is to take a diagonal W
W = diag(W1 , W2 , . . . , Wm )
and let Ŵi be computed from
T
|Ŵi (jω)| ≥ sup e Fu (Tẑŵ (jω), ∆p )
i
σ(∆p )≤1

where ei is the ith unit vector. Next let α(s) be computed from


|α(jω)| ≥ sup Ŵ −1 Fu (Tẑŵ (jω), ∆p )
σ(∆p )≤1

where Ŵ = diag(Ŵ1 , Ŵ2 , . . . , Ŵm ). Then a suitable W can be taken as

W = αŴ .

Apply this method to Problem 15.3.

Problem 15.7 Generalize the procedures in Problems 15.5 and 15.6 to problems with
additional structured uncertainty cases. (A more general case can be found in Yang and
Packard [1995].)
Chapter 16

H∞ Loop Shaping

This chapter introduces a design technique that incorporates loop-shaping methods to

obtain performance/robust stability tradeoffs, and a particular H∞ optimization prob-
lem to guarantee closed-loop stability and a level of robust stability at all frequencies.
The proposed technique uses only the basic concept of loop-shaping methods, and then
a robust stabilization controller for the normalized coprime factor perturbed system is
used to construct the final controller. This chapter is arranged as follows: The H∞
theory is applied to solve the stabilization problem of a normalized coprime factor
perturbed system in Section 16.1. The loop-shaping design procedure is described in
Section 16.2. The theoretical justification for the loop-shaping design procedure is given
in Section 16.3. Some further loop-shaping guidelines are given in Section 16.4.

16.1 Robust Stabilization of Coprime Factors

In this section, we use the H∞ control theory developed in previous chapters to solve
the robust stabilization of a left coprime factor perturbed plant given by
˜ M )−1 (Ñ + ∆
P∆ = (M̃ + ∆ ˜ N)
h i
˜M, ∆
˜N ˜ ˜M
with M̃ , Ñ, ∆ ∈ RH∞ and ∆ N ∆ <  (see Figure 16.1). The transfer
∞
matrices (M̃ , Ñ ) are assumed to be a left coprime factorization of P (i.e., P = M̃ −1 Ñ ),
and K internally stabilizes the nominal system.
It has been shown in Chapter 8 that the system is robustly stable iff
" #
K
−1 −1
(I + P K) M̃ ≤ 1/.
I
∞

Finding a controller such that the above norm condition holds is an H∞ norm mini-
mization problem that can be solved using H∞ theory developed in previous chapters.

315
316 H∞ LOOP SHAPING

z1 − z2
- ∆
˜N - f ˜M 
∆
w
r y
-f - K - Ñ -?
f - M̃ −1 -
−6

Figure 16.1: Left coprime factor perturbed systems

Suppose P has a stabilizable and detectable state-space realization given by

" #
A B
P =
C D
and let L be a matrix such that A + LC is stable. Then a left coprime factorization of
P = M̃ −1 Ñ is given by
" #
h i A + LC B + LD L
Ñ M̃ =
ZC ZD Z
where Z can be any nonsingular matrix. In particular, we shall choose Z = (I +
DD∗ )−1/2 if P = M̃ −1 Ñ is chosen to be a normalized left coprime factorization. Denote
K̂ = −K.
Then the system diagram can be put in an LFT form, as in Figure 16.2, with the
generalized plant
 
 " # " #  A −LZ −1 B
0 I  " # " # " # 
   0 0 I 
G(s) =  M̃ −1
P  =   
Z −1 
−1  C D 
M̃ P
C Z −1 D
 
A B1 B2
 
=:  C1 D11 D12  .
C2 D21 D22

To apply the H∞ control formulas in Chapter 14, we need to normalize the D12 and
D21 first. Note that
" # " # " #
D∗ (I + DD∗ )− 2 (I + D∗ D)− 2
1 1
I 0 ∗ 1
=U (I + D D) , where U =
2
−(I + DD∗ )− 2 D(I + D∗ D)− 2
1 1
D I
16.1. Robust Stabilization of Coprime Factors 317

z1  w

z 
2

?
M̃ −1  i Ñ 

y u

- K̂

Figure 16.2: LFT diagram for coprime factor stabilization

and U is a unitary matrix. Let

K̂ = (I + D∗ D)− 2 K̃Z
1

" # " #
z1 ẑ1
= U .
z2 ẑ2

Then kTzw k∞ = kU ∗ Tzw k∞ = kTẑw k∞ and the problem becomes one of finding a
controller K̃ so that kTẑw k∞ < γ with the following generalized plant:
" # " #
U∗ 0 I 0
Ĝ = G
0 (I + D∗ D)− 2
1
0 Z
 
A −LZ −1 B
 " # " # " # 
 −(I + DD ∗ − 12
) C −(I + DD ∗ − 12 −1
) Z 0 
=

.

(I + D∗ D)− 2 D∗ C (I + D∗ D)− 2 D∗ Z −1
1 1
 I 
ZD(I + D∗ D)− 2
1
ZC I

Now the formulas in Chapter 14 can be applied to Ĝ to obtain a controller K̃ and then
the K can be obtained from K = −(I + D∗ D)− 2 K̃Z. We shall leave the detail to the
1

reader. In the sequel, we shall consider the case D = 0 and Z = I. In this case, we have
γ > 1 and

LC LC ∗ LL∗ γ2C ∗C
X∞ (A − ) + (A − 2 ) X∞ − X∞ (BB ∗ − 2 )X∞ + 2 =0
γ2 −1 γ −1 γ −1 γ −1

Y∞ (A + LC)∗ + (A + LC)Y∞ − Y∞ C ∗ CY∞ = 0.

318 H∞ LOOP SHAPING

It is clear that Y∞ = 0 is the stabilizing solution. Hence by the formulas in Chapter 14

we have h i h i
L1∞ L2∞ = 0 L
and p
γ2 − 1
Z∞ = I, D̂11 = 0, D̂12 = I, D̂21 = I.
γ
The results are summarized in the following theorem.
Theorem 16.1 Let D = 0 and let L be such that A + LC is stable. Then there exists
a controller K such that
" #
K
−1 −1
(I + P K) M̃ < γ
I
∞

iff γ > 1 and there exists a stabilizing solution X∞ ≥ 0 solving

LC LC ∗ LL∗ γ2C ∗C
X∞ (A − ) + (A − 2 ) X∞ − X∞ (BB ∗ − 2 )X∞ + 2 = 0.
γ2−1 γ −1 γ −1 γ −1
Moreover, if the above conditions hold a central controller is given by
" #
A − BB ∗ X∞ + LC L
K= .
− B ∗ X∞ 0
It is clear that the existence of a robust stabilizing controller depends on the choice of
the stabilizing matrix L (i.e., the choice of the coprime factorization). Now let Y ≥ 0
be the stabilizing solution to
AY + Y A∗ − Y C ∗ CY + BB ∗ = 0
and let L = −Y C ∗ . Then the left coprime factorization (M̃ , Ñ ) given by
" #
h i A − Y C ∗ C B −Y C ∗
Ñ M̃ =
C 0 I
is a normalized left coprime factorization (see Chapter 12). Let k·kH denote the Hankel
norm (i.e., the largest Hankel singular value). Then we have the following result.
Corollary 16.2 Let D = 0 and L = −Y C ∗ , where Y ≥ 0 is the stabilizing solution to
AY + Y A∗ − Y C ∗ CY + BB ∗ = 0.
Then P = M̃ −1 Ñ is a normalized left coprime factorization and
" #
K
−1 −1 1
inf (I + P K) M̃ = p
K stabilizing I 1 − λmax (Y Q)
∞
 h i 2 −1/2

= 1 − Ñ M̃ =: γmin
H
16.1. Robust Stabilization of Coprime Factors 319

where Q is the solution to the following Lyapunov equation:

Q(A − Y C ∗ C) + (A − Y C ∗ C)∗ Q + C ∗ C = 0.

Moreover, if the preceding conditions hold, then for any γ > γmin a controller achieving
" #
K
−1 −1
(I + P K) M̃ < γ
I
∞

is given by " #
A − BB ∗ X∞ − Y C ∗ C −Y C ∗
K(s) =
− B ∗ X∞ 0
where  −1
γ2 γ2
X∞ = Q I − Y Q .
γ2 − 1 γ2 − 1

Proof. Note that the Hamiltonian matrix associated with X∞ is given by

" #
A + γ 21−1 Y C ∗ C −BB ∗ + γ 21−1 Y C ∗ CY
H∞ = 2 .
− γ 2γ−1 C ∗ C −(A + γ 21−1 Y C ∗ C)∗

Straightforward calculation shows that

" 2 # " 2 #−1
I − γ 2γ−1 Y I − γ 2γ−1 Y
H∞ = γ2
Hq γ2
0 γ 2 −1 I 0 γ 2 −1 I

where " #
A − Y C ∗C 0
Hq = .
−C ∗ C −(A − Y C ∗ C)∗
It is clear that the stable invariant subspace of Hq is given by
" #
I
X− (Hq ) = Im
Q

and the stable invariant subspace of H∞ is given by

" 2 # " 2 #
I − γ 2γ−1 Y I − γ 2γ−1 Y Q
X− (H∞ ) = γ2
X− (Hq ) = Im γ2
.
0 γ 2 −1 I γ 2 −1 Q

Hence there is a nonnegative definite stabilizing solution to the algebraic Riccati equa-
tion of X∞ if and only if
γ2
I− 2 YQ>0
γ −1
320 H∞ LOOP SHAPING

or
1
γ>p
1 − λmax (Y Q)
and the solution, if it exists, is given by
 −1
γ2 γ2
X∞ = Q I − Y Q .
γ2 − 1 γ2 − 1

Note
h that iY and Q are the controllability Gramian and the observability Gramian
h ofi
Ñ M̃ respectively. Therefore, we also have that the Hankel norm of Ñ M̃
p
is λmax (Y Q). 2

Corollary 16.3 Let P = M̃ −1 Ñ be a normalized left coprime factorization and

˜ M )−1 (Ñ + ∆
P∆ = (M̃ + ∆ ˜ N)

with h i
˜
∆N ˜M
∆ < .
∞
Then there is a robustly stabilizing controller for P∆ if and only if
r h i 2
p
 ≤ 1 − λmax (Y Q) = 1 − Ñ M̃ .
H

The solutions to the normalized left coprime factorization stabilization problem are
also solutions to a related H∞ problem, which is shown in the following lemma.

Lemma 16.4 Let P = M̃ −1 Ñ be a normalized left coprime factorization. Then

" # " #
K K h i

(I + P K)−1 M̃ −1 = (I + P K)−1 I P .
I I
∞ ∞

Proof. Since (M̃ , Ñ) is a normalized left coprime factorization of P , we have

h ih i∼
M̃ Ñ M̃ Ñ =I

and h i h i∼

M̃ Ñ = M̃ Ñ = 1.
∞ ∞
Using these equations, we have
" #
K

(I + P K)−1 M̃ −1
I
∞
16.1. Robust Stabilization of Coprime Factors 321

" #
h ih i∼
K −1 −1
= (I + P K) M̃ M̃ Ñ M̃ Ñ
I
∞
" #
h i h i∼
K
≤ (I + P K)−1 M̃ −1 M̃ Ñ M̃ Ñ
I ∞
∞
" #
h i
K −1
= (I + P K) I P
I
∞
" #
h i
K
≤ (I + P K)−1 M̃ −1 M̃ Ñ
I ∞
" # ∞

K
= (I + P K)−1 M̃ −1 .
I
∞

This implies
" # " #
K K h i
−1 −1
(I + P K) M̃ = (I + P K)−1 I P .
I I
∞ ∞

Combining Corollary 16.3 and Lemma 16.4, we have the following result.

Corollary 16.5 A controller solves the normalized left coprime factor robust stabiliza-
tion problem if and only if it solves the following H∞ control problem:
" #
I h i

(I + P K)−1 I P < γ
K
∞

and
" #
I h i
1
inf (I + P K)−1 I P = p
K stabilizing K 1 − λmax (Y Q)
∞
 h i 2 −1/2

= 1 − Ñ M̃ .
H

The solution Q can also be obtained in other ways. Let X ≥ 0 be the stabilizing
solution to
XA + A∗ X − XBB ∗ X + C ∗ C = 0.
Then it is easy to verify that
Q = (I + XY )−1 X.
322 H∞ LOOP SHAPING

Hence
 h i 2 −1/2 p
1
γmin = p = 1 − Ñ M̃ = 1 + λmax (XY ).
1 − λmax (Y Q) H

Similar results can be obtained if one starts with a normalized right coprime factoriza-
tion. In fact, a rather strong relation between the normalized left and right coprime
factorization problems can be established using the following matrix fact.

Lemma 16.6 Let M be a square matrix such that M 2 = M . Then σi (M ) = σi (I − M )

for all i such that 0 < σi (M ) 6= 1.

Proof. We first show that the eigenvalues of M are either 0 or 1 and M is diagonaliz-
able. In fact, assume that λ is an eigenvalue of M and x is a corresponding eigenvector.
Then λx = M x = M M x = M (M x) = λM x = λ2 x; that is, λ(1 − λ)x = 0. This
implies that either λ = 0 or λ = 1. To show that M is diagonalizable, assume that
M = T JT −1, where J is a Jordan canonical form. It follows immediately that J must
be diagonal by the condition M = M 2 .
Next, assume that M is diagonalized by a nonsingular matrix T such that
" #
I 0
M =T T −1 .
0 0

Then " #
0 0
N := I − M = T T −1 .
0 I
Define " #
A B
:= T ∗ T
B∗ D
and assume 0 < λ 6= 1. Then A > 0 and

det(M ∗ M − λI) = 0
" # " #
I 0 I 0
⇔ det( T ∗T − λT ∗ T ) = 0
0 0 0 0
" #
(1 − λ)A −λB
⇔ det =0
−λB ∗ −λD
λ2
⇔ det(−λD − B ∗ A−1 B) = 0
1−λ
1−λ
⇔ det( D + B ∗ A−1 B) = 0
λ
16.1. Robust Stabilization of Coprime Factors 323

" #
−λA −λB
⇔ det =0
−λB ∗ (1 − λ)D
⇔ det(N ∗ N − λI) = 0.

This implies that all nonzero eigenvalues of M ∗ M and N ∗ N that are not equal to 1 are
equal; that is, σi (M ) = σi (I − M ) for all i such that 0 < σi (M ) 6= 1. 2

Using this matrix fact, we have the following corollary.

Corollary 16.7 Let K and P be any compatibly dimensioned complex matrices. Then
" # " #
I h i I h i

(I + P K)−1 I P = (I + KP )−1 I K .
K P

Proof. Define
" # " #
I h i −P h i
M= (I + P K)−1 I P , N= (I + KP )−1 −K I .
K I

Then it is easy to verify that M 2 = M and N = I − M . By Lemma 16.6, we have

kM k = kN k. The corollary follows by noting that
" # " # " #
I h i 0 I 0 −I
−1
(I + KP ) I K = N .
P −I 0 I 0
2

Corollary 16.8 Let P = M̃ −1 Ñ = N M −1 be, respectively, the normalized left and

right coprime factorizations. Then
" #
K h i
−1
(I + P K) M̃ = M −1 (I + KP )−1 I K .
−1
I ∞
∞

Proof. This follows from Corollary 16.7 and the fact that
" #
h i I h i
−1 −1
M (I + KP ) I K = (I + KP )−1 I K .
∞ P
∞

This corollary says that any H∞ controller for the normalized left coprime factor-
ization is also an H∞ controller for the normalized right coprime factorization. Hence
one can work with either factorization.
324 H∞ LOOP SHAPING

For future reference, we shall define

 " # !−1
 I h i



 (I + P K)−1 I P if K stabilizes P
K
bP,K := ∞





0 otherwise
and
bopt := sup bP,K .
K
Then bP,K = bK,P and
r h i 2
p
bopt = 1 − λmax (Y Q) = 1 − Ñ M̃ .
H

The number bP,K can be related to the classical gain and phase margins as shown in
Vinnicombe [1993b].
Theorem 16.9 Let P be a SISO plant and K be a stabilizing controller. Then
1 + bP,K
gain margin ≥
1 − bP,K
and
phase margin ≥ 2 arcsin(bP,K ).

Proof. Note that for SISO system

|1 + P (jω)K(jω)|
bP,K ≤ p p , ∀ω.
1 + |P (jω)|2 1 + |K(jω)|2
So, at frequencies where k := −P (jω)K(jω) ∈ R+ ,
 
|1 − k| |1 − k| 1 − k
bP,K ≤ s ≤s  
 2
 1 + k ,
=
k2 k
(1 + |P |2 )(1 + ) min (1 + |P |2 )(1 + )
|P |2 P |P |2

which implies that

1 − bP,K 1 + bP,K
k≤ , or k ≥
1 + bP,K 1 − bP,K
from which the gain margin result follows. Similarly, at frequencies where P (jω)K(jω) =
−ejθ ,
|1 − ejθ | |2 sin θ2 | |2 sin θ2 |
bP,K ≤ s ≤s  = ,
1 1 2
(1 + |P |2 )(1 + ) min (1 + |P | )(1 +
2
)
|P |2 P |P |2
16.2. Loop-Shaping Design 325

which implies θ ≥ 2 arcsin bP,K . 2

For example, bP,K = 1/2 guarantees a gain margin of 3 and a phase margin of 60o .

Illustrative MATLAB Commands:

 bp,k = emargin(P, K); % given P and K, compute bP,K .

 [Kopt , bp,k ] = ncfsyn(P, 1); % find the optimal controller Kopt .
 [Ksub , bp,k ] = ncfsyn(P, 2); % find a suboptimal controller Ksub .

16.2 Loop-Shaping Design

This section considers the H∞ loop-shaping design. The objective of this approach is
to incorporate the simple performance/robustness tradeoff obtained in the loop-shaping
with the guaranteed stability properties of H∞ design methods. Recall from Section 6.1
of Chapter 6 that good performance controller design requires that





σ (I + P K)−1 , σ (I + P K)−1 P , σ (I + KP )−1 , σ K(I + P K)−1 (16.1)

be made small, particularly in some low-frequency range. Good robustness requires that



σ P K(I + P K)−1 , σ KP (I + KP )−1 (16.2)

be made small, particularly in some high-frequency range. These requirements, in turn,

imply that good controller design boils down to achieving the desired loop (and con-
troller) gains in the appropriate frequency range:

σ(P K)  1, σ(KP )  1, σ(K)  1

in some low-frequency range and

σ(P K)  1, σ(KP )  1, σ(K) ≤ M

in some high-frequency range where M is not too large.

The H∞ loop-shaping design procedure is developed by McFarlane and Glover [1990,
1992] and is stated next.

Loop-Shaping Design Procedure

(1) Loop-Shaping: The singular values of the nominal plant, as shown in Figure 16.3,
are shaped, using a precompensator W1 and/or a postcompensator W2 , to give a
desired open-loop shape. The nominal plant P and the shaping functions W1 , W2
are combined to form the shaped plant, Ps , where Ps = W2 P W1 . We assume that
W1 and W2 are such that Ps contains no hidden modes.
326 H∞ LOOP SHAPING

di d
r-e - K -?
e u -
P -?
e y-
− 6

?n
e

Figure 16.3: Standard feedback configuration

(2) Robust Stabilization: a) Calculate max (i.e., bopt (Ps )), where
" # !−1
I

max = inf (I + Ps K)−1 M̃s−1
K stabilizing K
∞
r h i 2

= 1 − Ñs M̃s < 1
H

and M̃s , Ñs define the normalized coprime factors of Ps such that Ps = M̃s−1 Ñs
and
M̃s M̃s∼ + Ñs Ñs∼ = I.
If max  1 return to (1) and adjust W1 and W2 .
b) Select  ≤ max ; then synthesize a stabilizing controller K∞ that satisfies
" #

I −1
(I + Ps K∞ ) M̃s ≤ −1 .
−1
K∞
∞

(3) The final feedback controller K is then constructed by combining the H∞ con-
troller K∞ with the shaping functions W1 and W2 , as shown in Figure 16.4, such
that
K = W1 K∞ W2 .

A typical design works as follows: the designer inspects the open-loop singular values
of the nominal plant and shapes these by pre- and/or postcompensation until nominal
performance (and possibly robust stability) specifications are met. (Recall that the
open-loop shape is related to closed-loop objectives.) A feedback controller K∞ with
associated stability margin (for the shaped plant)  ≤ max , is then synthesized. If max
is small, then the specified loop shape is incompatible with robust stability requirements
and should be adjusted accordingly; then K∞ is reevaluated.
In the preceding design procedure we have specified the desired loop shape by
W2 P W1 . But after Stage (2) of the design procedure, the actual loop shape achieved
16.2. Loop-Shaping Design 327

- W1 - P - W2 -

e - K∞ - W1 - P - W2
−6

Ps

e - W2 - K∞ - W1 - P
−6

Figure 16.4: The loop-shaping design procedure

is, in fact, given by W1 K∞ W2 P at plant input and P W1 K∞ W2 at plant output. It is

therefore possible that the inclusion of K∞ in the open-loop transfer function will cause
deterioration in the open-loop shape specified by Ps . In the next section, we will show
that the degradation in the loop shape caused by the H∞ controller K∞ is limited at
frequencies where the specified loop shape is sufficiently large or sufficiently small. In
particular, we show in the next section that  can be interpreted as an indicator of the
success of the loop-shaping in addition to providing a robust stability guarantee for the
closed-loop systems. A small value of max (max  1) in Stage (2) always indicates
incompatibility between the specified loop shape, the nominal plant phase, and robust
closed-loop stability.

Remark 16.1 Note that, in contrast to the classical loop-shaping approach, the loop-
shaping here is done without explicit regard for the nominal plant phase information.
That is, closed-loop stability requirements are disregarded at this stage. Also, in con-
trast with conventional H∞ design, the robust stabilization is done without frequency
weighting. The design procedure described here is both simple and systematic and only
assumes knowledge of elementary loop-shaping principles on the part of the designer.
3

Remark 16.2 The preceding robust stabilization objective can also be interpreted
as the more standard H∞ problem formulation of minimizing the H∞ norm of the
frequency-weighted gain from disturbances on the plant input and output to the con-
328 H∞ LOOP SHAPING

troller input and output as follows:

" # " #
h i
I −1 −1 I −1
(I + Ps K∞ ) M̃s = (I + Ps K∞ ) I Ps
K∞ K∞
∞ ∞
" #
h i
W2
= (I + P K)−1 W2−1 P W1
−1
W1 K
∞
" #
h i
I
= (I + K∞ Ps )−1 I K∞
Ps
∞
" #
−1 h i
W1 −1 −1
= (I + KP ) W1 KW2
W2 P
∞

This shows how all the closed-loop objectives in equations (16.1) and (16.2) are incor-
porated. As an example, it is easy to see that the signal relationship in Figure 16.5 is
given by " # " # " #
z1 W2 h i w
−1 −1 1
= (I + P K) W2 P W1 .
z2 W1−1 K w2
3

z2 w2 w1
6
? ?
W1−1 W1 W2−1
?- z1
e - K 6 -e P -?
e - W2 -
− 6

Figure 16.5: An equivalent H∞ formulation

16.3 Justification for H∞ Loop Shaping

The objective of this section is to provide justification for the use of parameter  as a
design indicator. We will show that  is a measure of both closed-loop robust stability
and the success of the design in meeting the loop-shaping specifications. Readers are
encouraged to consult the original reference by McFarlane and Glover [1990] for further
details.
We first examine the possibility of loop shape deterioration at frequencies of high
loop gain (typically low-frequency). At low-frequency [in particular, ω ∈ (0, ωl )], the de-
terioration in loop shape at plant output can be obtained by comparing σ(P W1 K∞ W2 )
16.3. Justification for H∞ Loop Shaping 329

to σ(Ps ) = σ(W2 P W1 ). Note that

σ(P K) = σ(P W1 K∞ W2 ) = σ(W2−1 W2 P W1 K∞ W2 ) ≥ σ(W2 P W1 )σ(K∞ )/κ(W2 )

(16.3)
where κ(·) denotes condition number. Similarly, for loop shape deterioration at plant
input, we have

σ(KP ) = σ(W1 K∞ W2 P ) = σ(W1 K∞ W2 P W1 W1−1 ) ≥ σ(W2 P W1 )σ(K∞ )/κ(W1 ).

(16.4)
In each case, σ(K∞ ) is required to obtain a bound on the deterioration in the loop shape
at low-frequency. Note that the condition numbers κ(W1 ) and κ(W2 ) are selected by
the designer.
Next, recalling that Ps denotes the shaped plant and that K∞ robustly stabilizes
the normalized coprime factorization of Ps with stability margin , we have
" #

I
(I + Ps K∞ )−1 M̃s−1 ≤ −1 := γ (16.5)
K∞
∞

where (Ñs , M̃s ) is a normalized left coprime factorization of Ps , and the parameter γ
is defined to simplify the notation to follow. The following result shows that σ(K∞ )
is explicitly bounded by functions of  and σ(Ps ), the minimum singular value of the
shaped plant, and hence by equation (16.3) and (16.4) K∞ will only have a limited
effect on the specified loop shape at low-frequency.

Theorem 16.10 Any controller K∞ satisfying equation (16.5), where Ps is assumed

square, also satisfies
p
σ(Ps (jω)) − γ 2 − 1
σ(K∞ (jω)) ≥ p
γ 2 − 1σ(Ps (jω)) + 1
for all ω such that p
σ(Ps (jω)) > γ 2 − 1.
p p
Furthermore, if σ(Ps )  γ 2 − 1, then σ(K∞ (jω)) ' 1/ γ 2 − 1, where ' denotes
asymptotically greater than or equal to as σ(Ps ) → ∞.
p
Proof. First note that σ(Ps ) > γ 2 − 1 implies that

I + Ps Ps∗ > γ 2 I.

Further, since (Ñs , M̃s ) is a normalized left coprime factorization of Ps , we have

M̃s M̃s∗ = I − Ñs Ñs∗ = I − M̃s Ps Ps∗ M̃s∗ .

Then
M̃s∗ M̃s = (I + Ps Ps∗ )−1 < γ −2 I.
330 H∞ LOOP SHAPING

Now " #

I
(I + Ps K∞ )−1 M̃s−1 ≤γ
K∞
∞

can be rewritten as
∗ ∗
(I + K∞ K∞ ) ≤ γ 2 (I + K∞ Ps∗ )(M̃s∗ M̃s )(I + Ps K∞ ). (16.6)

We will next show that K∞ is invertible. Suppose that there exists an x such that
K∞ x = 0, then x∗ × equation (16.6) × x gives

γ −2 x∗ x ≤ x∗ M̃s∗ M̃s x,

which implies that x = 0 since M̃s∗ M̃s < γ −2 I, and hence K∞ is invertible. Equa-
tion (16.6) can now be written as
−∗ −1 −∗
(K∞ K∞ + I) ≤ γ 2 (K∞ + Ps∗ )M̃s∗ M̃s (K∞
−1
+ Ps ). (16.7)

Define W such that

(W W ∗ )−1 = I − γ 2 M̃s∗ M̃s = I − γ 2 (I + Ps Ps∗ )−1 .

−1
Completing the square in equation (16.7) with respect to K∞ yields
−∗
(K∞ + N ∗ )(W W ∗ )−1 (K∞
−1
+ N ) ≤ (γ 2 − 1)R∗ R

where

N= γ 2 Ps ((1 − γ 2 )I + Ps∗ Ps )−1

R R = (I + Ps∗ Ps )((1 − γ 2 )I + Ps∗ Ps )−1 .
∗

Hence we have

R−∗ (K∞
−∗
+ N ∗ )(W W ∗ )−1 (K∞
−1
+ N )R−1 ≤ (γ 2 − 1)I

and
p
σ W −1 (K∞
−1
+ N )R−1 ≤ γ 2 − 1.


Use σ W −1 (K∞
−1
+ N )R−1 ≥ σ(W −1 )σ(K∞−1
+ N )σ(R−1 ) to get
p
−1
σ(K∞ + N) ≤ γ 2 − 1σ(W )σ(R)
−1
and use σ(K∞ + N ) ≥ σ(K∞ ) − σ(N ) to get
n o−1
σ(K∞ ) ≥ (γ 2 − 1)1/2 σ(W )σ(R) + σ(N ) . (16.8)
16.3. Justification for H∞ Loop Shaping 331

Next, note that the eigenvalues of W W ∗ , N ∗ N , and R∗ R can be computed as follows:

1 + λ(Ps Ps∗ )
λ(W W ∗ ) =
1 − γ 2 + λ(Ps Ps∗ )

γ 4 λ(Ps Ps∗ )
λ(N ∗ N ) =
(1 − γ 2 + λ(Ps Ps∗ ))2
1 + λ(Ps Ps∗ )
λ(R∗ R) = .
1 − γ 2 + λ(Ps Ps∗ )
Therefore,
 1/2  1/2
p 1 + λmin (Ps Ps∗ ) 1 + σ2 (Ps )
σ(W ) = λmax (W W ∗ ) = =
1 − γ 2 + λmin (Ps Ps∗ ) 1 − γ 2 + σ 2 (Ps )
p
p γ 2
λmin (Ps Ps∗ ) γ 2 σ(Ps )
σ(N ) = λmax (N ∗ N ) = ∗
=
1 − γ + λmin (Ps Ps )
2 1 − γ 2 + σ2 (Ps )
 1/2  1/2
p 1 + λmin (Ps Ps∗ ) 1 + σ 2 (Ps )
∗
σ(R) = λmax (R R) = = .
1 − γ 2 + λmin (Ps Ps∗ ) 1 − γ 2 + σ 2 (Ps )
Substituting these formulas into equation (16.8), we have
 −1 p
(γ 2 − 1)1/2 (1 + σ 2 (Ps )) + γ 2 σ(Ps ) σ(Ps )) − γ 2 − 1
σ(K∞ ) ≥ =p .
σ2 (Ps ) − (γ 2 − 1) γ 2 − 1σ(Ps ) + 1
2

The main implication of Theorem 16.10 is that the bound on σ(K∞ ) depends only
on the selected loop shape and the stability margin of the shaped plant. The value
of γ(= −1 ) directly determines the frequency range over which this result is valid – a
small γ (large ) is desirable, as we would expect. Further, Ps has a sufficiently large
loop gain; then so also will Ps K∞ provided that γ(= −1 ) is sufficiently small.
In an analogous manner, we now examine the possibility of deterioration in the loop
shape at high-frequency due to the inclusion of K∞ . Note that at high frequency [in
particular, ω ∈ (ωh , ∞)] the deterioration in plant output loop shape can be obtained
by comparing σ(P W1 K∞ W2 ) to σ(Ps ) = σ(W2 P W1 ). Note that, analogous to equation
(16.3) and (16.4), we have

σ(P K) = σ(P W1 K∞ W2 ) ≤ σ(W2 P W1 )σ(K∞ )κ(W2 ).

Similarly, the corresponding deterioration in plant input loop shape is obtained by

comparing σ(W1 K∞ W2 P ) to σ(W2 P W1 ), where

σ(KP ) = σ(W1 K∞ W2 P ) ≤ σ(W2 P W1 )σ(K∞ )κ(W1 ).

332 H∞ LOOP SHAPING

Hence, in each case, σ(K∞ ) is required to obtain a bound on the deterioration in the
loop shape at high-frequency. In an identical manner to Theorem 16.10, we now show
that σ(K∞ ) is explicitly bounded by functions of γ and σ(Ps ), the maximum singular
value of the shaped plant.

Theorem 16.11 Any controller K∞ satisfying equation (16.5) also satisfies

p
γ 2 − 1 + σ(Ps (jω))
σ(K∞ (jω)) ≤ p
1 − γ 2 − 1 σ(Ps (jω))

for all ω such that

1
σ(Ps (jω)) < p .
γ −1
2

p p
Furthermore, if σ(Ps )  1/ γ 2 − 1, then σ(K∞ (jω)) / γ 2 − 1, where / denotes
asymptotically less than or equal to as σ(Ps ) → 0.

Proof. The proof of Theorem 16.11 is similar to that of Theorem 16.10 and is only
sketched here: As in the proof of Theorem 16.10, we have M̃s∗ M̃s = (I + Ps Ps∗ )−1 and
∗ ∗
(I + K∞ K∞ ) ≤ γ 2 (I + K∞ Ps∗ )(M̃s∗ M̃s )(I + Ps K∞ ). (16.9)

Since σ(Ps ) < √ 1

,
γ 2 −1
I − γ 2 Ps∗ (I + Ps Ps∗ )−1 Ps > 0
and there exists a spectral factorization

V ∗ V = I − γ 2 Ps∗ (I + Ps Ps∗ )−1 Ps .

Now, completing the square in equation (16.9) with respect to K∞ yields

∗
(K∞ + M ∗ )V ∗ V (K∞ + M ) ≤ (γ 2 − 1)Y ∗ Y

where

M = γ 2 Ps∗ (I + (1 − γ 2 )Ps Ps∗ )−1

∗
Y Y = (γ 2 − 1)(I + Ps Ps∗ )(I + (1 − γ 2 )Ps Ps∗ )−1 .

Hence we have

p
σ V (K∞ + M )Y −1 ≤ γ 2 − 1,
which implies p
γ2 − 1
σ(K∞ ) ≤ + σ(M ). (16.10)
σ(V )σ(Y −1 )
16.3. Justification for H∞ Loop Shaping 333

As in the proof of Theorem 16.10, it is easy to show that

 1/2
1 − (γ 2 − 1)σ2 (Ps )
σ(V ) = σ(Y −1 ) =
1 + σ2 (Ps )

γ 2 σ(Ps )
σ(M ) = .
1 − (γ 2 − 1)σ 2 (Ps )
Substituting these formulas into equation (16.10), we have
p
(γ 2 − 1)1/2 (1 + σ2 (Ps )) + γ 2 σ(Ps ) γ 2 − 1 + σ(Ps )
σ(K∞ ) ≤ = p .
1 − (γ 2 − 1)σ (Ps )
2
1 − γ 2 − 1σ(Ps )
2

The results in Theorems 16.10 and 16.11 confirm that γ (alternatively ) indicates the
compatibility between the specified loop shape and closed-loop stability requirements.

Theorem 16.12 Let P be the nominal plant and let K = W1 K∞ W2 be the associated
controller obtained from loop-shaping design procedure in the last section. Then if
" #

I
(I + Ps K∞ )−1 M̃s−1 ≤ γ
K∞
∞

we have


σ K(I + P K)−1 ≤ γσ(M̃s )σ(W1 )σ(W2 ) (16.11)

n o
σ (I + P K)−1 ≤ min γσ(M̃s )κ(W2 ), 1 + γσ(Ns )κ(W2 ) (16.12)

n o
σ K(I + P K)−1 P ≤ min γσ(Ñs )κ(W1 ), 1 + γσ(Ms )κ(W1 ) (16.13)

γσ(Ñs )
σ (I + P K)−1 P ≤ (16.14)
σ(W1 )σ(W2 )

n o
−1
σ (I + KP ) ≤ min 1 + γσ(Ñs )κ(W1 ), γσ(Ms )κ(W1 ) (16.15)

n o
σ G(I + KP )−1 K ≤ min 1 + γσ(M̃s )κ(W2 ), γσ(Ns )κ(W2 ) (16.16)

where  1/2
σ2 (W2 P W1 )
σ(Ñs ) = σ(Ns ) = (16.17)
1 + σ2 (W2 P W1 )
 1/2
1
σ(M̃s ) = σ(Ms ) = (16.18)
1 + σ2 (W2 P W1 )
and (Ñs , M̃s ), (Ns , Ms ) is a normalized left coprime factorization and right coprime
factorization, respectively, of Ps = W2 P W1 .
334 H∞ LOOP SHAPING

Proof. Note that

M̃s∗ M̃s = (I + Ps Ps∗ )−1
and
M̃s M̃s∗ = I − Ñs Ñs∗ .
Then
1 1
σ 2 (M̃s ) = λmax (M̃s∗ M̃s ) = =
1 + λmin (Ps Ps∗ ) 1 + σ2 (Ps )
σ 2 (Ps )
σ2 (Ñs ) = 1 − σ2 (M̃s ) = . 2
1 + σ (Ps )
The proof for the normalized right coprime factorization is similar. All other inequalities
follow from noting that
" #

I −1 −1
(I + Ps K∞ ) M̃s ≤ γ
K∞
∞

and
" # " #
h i
I −1 −1 W2 −1
(I + Ps K∞ ) M̃s = (I + P K) W2−1 P W1
K∞ W1−1 K
∞
" # ∞
W −1 h i

= 1
(I + KP )−1 W1 P W2−1
W2 P
∞
2

This theorem shows that all closed-loop objectives are guaranteed to have bounded
magnitude and the bounds depend only on γ, W1 , W2 , and P .

16.4 Further Guidelines for Loop Shaping

Let P = N M −1 be a normalized right coprime factorization. It was shown in Georgiou
and Smith [1990] that
" #!
M (s)
bopt (P ) ≤ λ(P ) := inf σ .
Res>0 N (s)

Hence a small λ(P ) would necessarily imply a small bopt (P ). We shall now discuss the
performance limitations implied by this relationship for a scalar system. The following
argument is based on Vinnicombe [1993b], to which the reader is referred for further
discussions. Let z1 , z2 , . . . , zm and p1 , p2 , . . . , pk be the open right-half plane zeros and
poles of the plant P . Define
z1 − s z2 − s zm − s p1 − s p2 − s pk − s
Nz (s) = ... , Np (s) = ... .
z1 + s z2 + s zm + s p1 + s p2 + s pk + s
16.4. Further Guidelines for Loop Shaping 335

Then P can be written as

P (s) = P0 (s)Nz (s)/Np (s)
where P0 (s) has no open right-half plane poles or zeros. Let N0 (s) and M0 (s) be stable
and minimum phase spectral factors:
 −1
1
N0 (s)N0∼ (s) = 1+ , M0 (s)M0∼ (s) = (1 + P (s)P ∼ (s))−1 .
P (s)P ∼ (s)

Then P0 = N0 /M0 is a normalized coprime factorization and (N0 Nz ) and (M0 Np ) form
a pair of normalized coprime factorizations of P . Thus
q
bopt (P ) ≤ |N0 (s)Nz (s)|2 + |M0 (s)Np (s)|2 , ∀Re(s) > 0. (16.19)

Since N0 and M0 are both stable and have no zeros in the open right-half plane,
ln(N0 (s)) and ln(M0 (s)) are both analytic in Re(s) > 0 and so can be determined
from their boundary values on Re(s) = 0 via Poisson integrals (see also Problem 16.15):

Z !
∞
1
ln |N0 (re )| =
jθ
ln p Kθ (ω/r) d(ln ω)
−∞ 1 + 1/|P (jω)|2
Z !
∞
1
ln |M0 (rejθ )| = ln p Kθ (ω/r) d(ln ω)
−∞ 1 + |P (jω)|2

where r > 0, −π/2 < θ < π/2, and

1 2(ω/r)[1 + (ω/r)2 ] cos θ

Kθ (ω/r) =
π [1 − (ω/r)2 ]2 + 4(ω/r)2 cos2 θ

The function Kθ (ω/r) is plotted in Figure 16.6 against logarithmic frequency for
various values of θ. Note that the function is symmetric to ω = r in log ω and it attends
the maximum at ω = r. The function converges to an impulse function at ω = r when
θ approaches 90o ; that is, when |N0 (s)| or |M0 (s)| is evaluated close to the imaginary
axis.
Since the kernel function Kθ (ω/r) has the greatest weighting near ω = r, the Poisson
integral is largely determined by the frequency response near that frequency. Thus it is
clear that |N0 (rejθ )| will be small if |P (jω)| is small near ω = r. Similarly, |M0 (rejθ )|
will be small if |P (jω)| is large near ω = r.
It is also important to note that a very large percentage of weighting is concentrated
in a very narrow frequency range for a large θ (i.e., when s = rejθ has a much larger
imaginary part than the real part). Hence |N0 (rejθ )| and |M0 (rejθ )| will essentially be
determined by |P (jω)| in a very narrow frequency range near ω = r when θ is large. On
the other hand, when θ is small, a larger range of frequency response |P (jω)| around
ω = r will have affect on the value |N0 (rejθ )| and |M0 (rejθ )|. (This, in fact, will imply
336 H∞ LOOP SHAPING

1.5 o
θ = 80

o
1 θ = 60
Kθ (ω/r)

o
θ = 30

0.5
o
θ =0

0 −2 −1 0 1 2
10 10 10 10 10
ω/r

Figure 16.6: Kθ (ω/r) vs. normalized frequency ω/r

that a right-plane zero (pole) with a much larger real part than the imaginary part will
have much worse effect on the performance than a right-plane zero (pole) with a much
larger imaginary part than the real part.)
Let s = rejθ . Consider again the bound of equation (16.19) and note that Nz (zi ) = 0
and Np (pj ) = 0, we see that there are several ways in which the bound may be small
(i.e., bopt (P ) is small).

 |Nz (s)| and |Np (s)| are both small for some s. That is, |Nz (s)| ≈ 0 (i.e., s is close
to a right-half plane zero of P ) and |Np (s)| ≈ 0 (i.e., s is close to a right-half
plane pole of P ). This is only possible if P (s) has a right-half plane zero near a
right-half plane pole. (See Example 16.1.)

 |Nz (s)| and |M0 (s)| are both small for some s. That is, |Nz (s)| ≈ 0 (i.e., s is
close to a right-half plane zero of P ) and |M0 (s)| ≈ 0 (i.e., |P (jω)| is large around
ω = |s| = r). This is only possible if |P (jω)| is large around ω = r, where r is the
modulus of a right-half plane zero of P . (See Example 16.2.)

 |Np (s)| and |N0 (s)| are both small for some s. That is, |Np (s)| ≈ 0 (i.e., s is
close to a right-half plane pole of P ) and |N0 (s)| ≈ 0 (i.e., |P (jω)| is small around
ω = |s| = r). This is only possible if |P (jω)| is small around ω = r, where r is the
modulus of a right-half plane pole of P . (See Example 16.3.)

 |N0 (s)| and |M0 (s)| are both small for some s. That is, |N0 (s)| ≈ 0 (i.e., |P (jω)|
is small around ω = |s| = r) and |M0 (s)| ≈ 0 (i.e., |P (jω)| is large around
ω = |s| = r). The only way in which |P (jω)| can be both small and large
16.4. Further Guidelines for Loop Shaping 337

at frequencies near ω = r is that |P (jω)| is approximately equal to 1 and the

absolute value of the slope of |P (jω)| is large. (See Example 16.4.)

Example 16.1 Consider an unstable and nonminimum phase system

K(s − r)
P1 (s) = .
(s + 1)(s − 1)

The frequency responses of P1 (s) with r = 0.9 and K = 0.1, 1, and 10 are shown
in Figure 16.7. The following table shows that bopt (P1 ) will be very small for all K
whenever r is close to 1 (i.e., whenever there is an unstable pole close to an unstable
zero).

r 0.5 0.7 0.9 1.1 1.3 1.5

K = 0.1 bopt (P1 ) 0.0125 0.0075 0.0025 0.0025 0.0074 0.0124
r 0.5 0.7 0.9 1.1 1.3 1.5
K=1 bopt (P1 ) 0.1036 0.0579 0.0179 0.0165 0.0457 0.0706
r 0.5 0.7 0.9 1.1 1.3 1.5
K = 10 bopt (P1 ) 0.0658 0.0312 0.0088 0.0077 0.0208 0.0318

1
10
K=10

0
10
K=1

−1
10
K=0.1

−2
10 −2 −1 0 1 2
10 10 10 10 10

Figure 16.7: Frequency responses of P1 for r = 0.9 and K = 0.1, 1, and 10

338 H∞ LOOP SHAPING

Example 16.2 Consider a nonminimum phase plant

K(s − 1)
P2 (s) = .
s(s + 1)

The frequency responses of P2 (s) with K = 0.1, 1, and 10 are shown in Figure 16.8.
The following table shows clearly that bopt (P2 ) will be small if |P2 (jω)| is large around
ω = 1, the modulus of the right-half plane zero.

K 0.01 0.1 1 10 100

bopt (P2 ) 0.7001 0.6451 0.3827 0.0841 0.0098

3
10

K=10
2
10

1
10
K=1

0
10

K=0.1
−1
10

−2
10

−3
10 −2 −1 0 1 2
10 10 10 10 10

Figure 16.8: Frequency responses of P2 for K = 0.1, 1, and 10

Note that bopt (L/s) = 0.707 for any L and bopt (P2 ) −→ 0.707 as K −→ 0. This
is because |P2 (jω)| around the frequency of the right-half plane zero is very small as
K −→ 0.
Next consider a plant with a pair of complex right-half plane zeros:

K[(s − cos θ)2 + sin2 θ]

P3 (s) = .
s[(s + cos θ)2 + sin2 θ]

The magnitude frequency response of P3 is the same as that of P2 for the same K. The
optimal bopt (P3 ) for various θ’s are listed in the following table:
16.4. Further Guidelines for Loop Shaping 339

θ (degree) 0 45 60 80 85
K = 0.1 bopt (P3 ) 0.5952 0.6230 0.6447 0.6835 0.6950
θ (degree) 0 45 60 80 85
K =1 bopt (P3 ) 0.2588 0.3078 0.3568 0.4881 0.5512
θ (degree) 0 45 60 80 85
K = 10 bopt (P3 ) 0.0391 0.0488 0.0584 0.0813 0.0897

It can also be concluded from the table that bopt (P3 ) will be small if |P3 (jω)| is
large around the frequency of ω = 1 (the modulus of the right-half plane zero). It can
be further concluded that, for zeros with the same modulus, bopt (P3 ) will be smaller
for a plant with relatively larger real part zeros than for a plant with relatively larger
imaginary part zeros (i.e., a pair of real right-half plane zeros has a much worse effect
on the performance than a pair of almost pure imaginary axis right-half plane zeros of
the same modulus).

Example 16.3 Consider an unstable plant

K(s + 1)
P4 (s) = .
s(s − 1)

The magnitude frequency response of P4 is again the same as that of P2 for the same
K. The following table shows that bopt (P4 ) will be small if |P4 (jω)| is small around
ω = 1 (the modulus of the right-half plane pole).

K 0.01 0.1 1 10 100

bopt (P4 ) 0.0098 0.0841 0.3827 0.6451 0.7001

Note that bopt (P4 ) −→ 0.707 as K −→ ∞. This is because |P4 (jω)| is very large
around the frequency of the modulus of the right-half plane pole as K −→ ∞.
Next consider a plant with complex right-half plane poles:

K[(s + cos θ)2 + sin2 θ]

P5 (s) = .
s[(s − cos θ)2 + sin2 θ]

The optimal bopt (P5 ) for various θ’s are listed in the following table:
340 H∞ LOOP SHAPING

θ (degree) 0 45 60 80 85
K = 0.1 bopt (P5 ) 0.0391 0.0488 0.0584 0.0813 0.0897
θ (degree) 0 45 60 80 85
K =1 bopt (P5 ) 0.2588 0.3078 0.3568 0.4881 0.5512
θ (degree) 0 45 60 80 85
K = 10 bopt (P5 ) 0.5952 0.6230 0.6447 0.6835 0.6950

It can also be concluded from the table that bopt (P5 ) will be small if |P5 (jω)| is
small around the frequency of the modulus of the right-half plane pole. It can be
further concluded that, for poles with the same modulus, bopt (P5 ) will be smaller for
a plant with relatively larger real part poles than for a plant with relatively larger
imaginary part poles (i.e., a pair of real right-half plane poles has a much worse effect
on the performance than a pair of almost pure imaginary axis right-half plane poles of
the same modulus).

Example 16.4 Let a stable and minimum phase transfer function be

K(0.2s + 1)4
P6 (s) = .
s(s + 1)4

15
10

10
K=10000
10

5
10 K=0.1

0
10
K=0.00001

−5
10

−10
10

−15
10 −3 −2 −1 0 1 2 3
10 10 10 10 10 10 10

Figure 16.9: Frequency response of P6 for K = 10−5 , 10−1 and 104

16.5. Notes and References 341

The frequency responses of P6 with K = 10−5 , 10−1 , and 104 are shown in Figure 16.9.
It is clear that the slope of the frequency response near the crossover frequency for
K = 10−5 is not too large, which implies a reasonably good loop shape. Thus we should
expect bopt (P6 ) to be not too small. A similar conclusion applies to K = 104 . On the
other hand, the slope of the frequency response near the crossover frequency for K = 0.1
is quite large which implies a bad loop shape. Thus we should expect bopt (P6 ) to be
quite small. This is clear from the following table:

K 10−5 10−3 0.1 1 10 102 104

bopt (P6 ) 0.3566 0.0938 0.0569 0.0597 0.0765 0.1226 0.4933

Based on the preceding discussion, we can give some guidelines for the loop-shaping
design.

♥ The loop transfer function should be shaped in such a way that it has low gain
around the frequency of the modulus of any right-half plane zero z. Typically, it
requires that the crossover frequency be much smaller than the modulus of the
right-half plane zero; say, ωc < |z|/2 for any real zero and ωc < |z| for any complex
zero with a much larger imaginary part than the real part (see Figure 16.6).

♥ The loop transfer function should have a large gain around the frequency of the
modulus of any right-half plane pole.

♥ The loop transfer function should not have a large slope near the crossover fre-
quencies.

These guidelines are consistent with the rules used in classical control theory (see
Bode [1945] and Horowitz [1963]).

16.5 Notes and References

The H∞ loop-shaping using normalized coprime factorization was developed by McFar-
lane and Glover [1990, 1992], on which most parts of this chapter is based. In the same
references, some design examples were also shown. The method has been applied to the
design of scheduled controllers for a VSTOL aircraft in Hyde and Glover [1993]. Some
limitations of this loop-shaping design are discussed in detail in Vinnicombe [1993b] (on
which Section 16.4 is based) and Christian and Freudenberg [1994]. The robust sta-
bilization of normalized coprime factors is closely related to the robustness in the gap
metric and ν-gap metric, which will be discussed in the next chapter, see El-Sakkary
[1985], Georgiou and Smith [1990], Glover and McFarlane [1989], McFarlane, Glover,
and Vidyasagar [1990], Qiu and Davison [1992a, 1992b], Vinnicombe [1993a, 1993b],
Vidyasagar [1984, 1985], Zhu [1989], and references therein.
342 H∞ LOOP SHAPING

16.6 Problems
1
Problem 16.1 Consider a feedback system with G(s) = . Compute by hand max ,
s−2
the maximum stability radius for a normalized coprime factor perturbation of G(s).

Problem 16.2 In Corollary 16.2, find the parameterization of all H∞ controllers.

Problem 16.3 Let P∆ = (N + ∆N )(M + ∆M )−1 be a right coprime factor perturbed

plant with a nominal plant P = N M −1 , where (N, M ) is a pair of normalized right
coprime factorization. Formulate the corresponding robust stabilization problem as
an H∞ control problem and find a stabilizing controller using the H∞ formulas in
Chapter 14. Are there any connections between this stabilizing controller and the
controller obtained in this chapter for left coprime stabilization?

Problem 16.4 Let P have coprime factorizations P = N M −1 = M̃ −1 Ñ . Then there

exist U, V, Ũ , Ṽ ∈ H∞ such that
" #" #
M U Ṽ −Ũ
= I.
N V Ñ M̃

Furthermore, all stabilizing controllers for P can be written as

K = (U + M Q)(V + N Q)−1 , Q ∈ H∞ .

Show that " # " #

K U + MQ
(I + P K)−1 M̃ −1 = .
I V + NQ

Suppose P = N M −1 = M̃ −1 Ñ are normalized coprime factorizations. Show that

" # " # " #
U R+Q
−1 M
bP,K = + Q =
V N I
∞ ∞

where R = M ∼ U + N ∼ V .

Problem 16.5 Let K be a controller that stabilizes the plant P . Show that
1. K stabilizes P̃ = P + ∆a such that ∆a ∈ H∞ and k∆a k < bP,K ;

2. K stabilizes P̃ = P (I + ∆m ) such that ∆m ∈ H∞ and k∆m k < bP,K ;

3. K stabilizes P̃ = (I + ∆m )P such that ∆m ∈ H∞ and k∆m k < bP,K ;

4. K stabilizes P̃ = P (I + ∆f )−1 such that ∆f ∈ H∞ and k∆f k < bP,K ;

5. K stabilizes P̃ = (I + ∆f )−1 P such that ∆f ∈ H∞ and k∆f k < bP,K .

16.6. Problems 343

Discuss the possible implications of the preceding results.

Problem 16.6 Let K be a controller that stabilizes the plant P . Show that
−1
# in the form of K̃ = (U + ∆U )(V + ∆V ) such that ∆U , ∆V ∈ H∞
1. any controller
"
∆
U
and < bP,K also stabilizes P ;
∆V
∞

2. any controller K̃ = K + ∆a such that ∆a ∈ H∞ and k∆a k < bP,K also stabilizes
P;
3. any controller K̃ = K(I + ∆m ) such that ∆m ∈ H∞ and k∆m k < bP,K also
stabilizes P ;
4. any controller K̃ = (I + ∆m )K such that ∆m ∈ H∞ and k∆m k < bP,K also
stabilizes P ;
5. any controller K̃ = K(I + ∆f )−1 such that ∆f ∈ H∞ and k∆f k < bP,K also
stabilizes P ;
6. any controller K̃ = (I + ∆f )−1 K such that ∆f ∈ H∞ and k∆f k < bP,K also
stabilizes P .
Discuss the possible implications of the preceding results.
Problem 16.7 Let an uncertain plant be given by
s+α
Pδ = , α ∈ [1, 3], ζ ∈ [0.2, 0.4]
s2 + 2ζs + 1
and let a nominal model be
s + α0
P = .
s2 + 2ζ0 s + 1
1. Let α0 = 2 and ζ0 = 0.3. Find the largest possible k∆add k∞ and k∆mul k∞ where
∆add = Pδ − P, ∆mul = (Pδ − P )/P.

2. Let α0 = 2 and ζ0 = 0.3. Show that P = N/M with

s+2 s2 + 0.6s + 1
N= , M= 2
s2 + 1.9576s + 2.2361 s + 1.9576s + 2.2361
is a normalized coprime factorization. Now let
s+α s2 + 2ζs + 1
Nδ = , M δ =
s2 + 1.9576s + 2.2361 s2 + 1.9576s + 2.2361
∆n = Nδ − N, ∆m = Mδ − M.
h i

Find the largest possible ∆n ∆m .
∞
344 H∞ LOOP SHAPING

3. In part 2, let (N hδ , Mδ ) be a i
normalized coprime factorization of Pδ . Find the

largest possible ∆n ∆m .
∞

4. Find the optimal 0 and ζ0 such that the largest possible k∆add k∞ ,
h nominal αi

k∆mul k∞ , and ∆n ∆m are minimized, respectively.
∞

Discuss the advantages of each uncertainty modeling method in terms of robust stabi-
lizations.
−10
Problem 16.8 Let P = . Design (a) a precompensator W of order no greater
s(s − 1)
than 2 such that the crossover frequency ωc ≤ 2 and bopt (W P ) is as large as possible;
(b) find the optimal loop-shaping controller K = K∞ W with the W obtained in part
(a).

100(1 − s)
Problem 16.9 Let P = . Design (a) a precompensator W of order no
s(s + 10)
greater than 2 such that the crossover frequency ωc ≥ 1 and bopt (W P ) is as large as
possible; (b) find the optimal loop-shaping controller K = K∞ W with the W obtained
in part (a).
" #
A B
Problem 16.10 Let G(s) = and G(s) = N M −1 with
C 0
 
" # A + BF B
N  
= C 0 
M
F I

where #F is chosen such that N M −1 is a normalized

" " right
# coprime factorization.
" # Let
N̂ N N̂
be an rth order balanced truncation of . Show that is also a
M̂ M M̂
normalized right coprime factorization.

Problem 16.11 (Reduced-Order Controllers by Controller Model" Reduction; # see Mc-

A B
Farlane and Glover [1990], Zhou and Chen [1995].) Let G(s) = = M̃ −1 Ñ
C 0
be a normalized left coprime factorization and let K(s) be a suboptimal controller given
in Corollary 18.2 (with performance γ). Let K = U V −1 be a right coprime factorization
 
" # A − BB ∗ X∞ −Y C ∗
U  
= −C I 
V ∗
−B X∞ 0
16.6. Problems 345

and Û , V̂ ∈ RH∞ be approximations of U and V . Define

" # " #
U Û

 := −
V V̂ ∞

and Kr = Û V̂ −1 . Show that Kr is a stabilizing controller for G if  < 1 and

" # " #
K K h i
r −1 −1 r −1 γ
(I + GKr ) M̃ = (I + GKr ) I G < .
I I 1−
∞ ∞

Problem 16.12 (Reduced Order Controllers by Plant Model Reduction; see McFarlane
and Glover [1990].) Let G = M̃ −1 Ñ be a normalized left coprime factorization and let
K be a stabilizing controller such that
" #
K
−1
(I + GK) M̃ ≤ δ −1 .
−1
I
∞

Let Gr := M̃r−1 Ñr be an approximation of G and

h i

 := M̃ − M̃r Ñ − Ñr < δ.
∞

(a) Show that K stabilizes Gr and

" #
K
−1 −1
(I + Gr K) M̃r ≤ (δ − )−1 .
I
∞

(b) Let W (s), W −1 (s) ∈ RH∞ be obtained from the following spectral factorization:

W −1 W −∗ = M̃r M̃r∗ + Ñr Ñr∗ .

1
Show that kW k∞ ≤ and W −1 ∞ ≤ 1 + .
1−
(c) Show that " #
K
−1 1 −1 −1
δrn:= inf (I + Gr K1 ) (W M̃r )
K1 I
∞
" #
K h i W −1
1+
(I + Gr K1 )−1 I Gr ≤ ∞
1
= inf ≤ .
K1 I δ −  δ−
∞
and " #
K

(I + Gr K1 )−1 M̃r−1 −1
1
≤ δrn kW k∞ .
I
∞
346 H∞ LOOP SHAPING

(d) With the controller K1 given in (c), show that

" # " #
K K h i
−1
(I + GK1 )−1 M̃ −1 = (I + GK1 )−1 I
1 1
G ≤ δred
I I
∞ ∞

where
δrn δ− 1−
δred := −≤ −≤ (δ − ) − .
kW k∞ kW −1 k∞ kW k∞ 1+

Note that ifpÑr and M̃r are the kth-order balanced

Pn truncation of Ñ and M̃ , then
δ = δrn = 1 − σ12 , δred = δ − , and  ≤ 2 i=k+1 σi , where σi are the Hankel
h i
singular values of M̃ Ñ .

(e) Show that

" #
K
−1
δ̃r−1 (I + Gr K2 ) M̃r ≤ (δ − )−1 .
−1
2
:= inf
K2 I
∞

(f) With the controller K2 given in (e), show that

" # " #
K K h i
−1 −1
(I + GK2 )−1 I
2 2
(I + GK2 ) M̃ = G
I I
∞ ∞

≤ (δ̃r − )−1 ≤ (δ − 2)−1 .

Again note that if Ñr and M̃r are the kth-order balanced truncation of Ñ and M̃ ,
then δ̃r = δ.
(Note that K1 and K2 are reduced-order controllers.)
" #
A B
Problem 16.13 Let G(s) = = M̃ −1 Ñ be a normalized left coprime factor-
C 0
ization and let K(s) be a suboptimal controller given in Corollary 18.2 (with perfor-
mance γ): " #
A − BB ∗ X∞ − Y C ∗ C −Y C ∗
K(s) =
−B ∗ X∞ 0
where  −1
γ2 γ2
X∞ = Q I − Y Q
γ2 − 1 γ2 − 1
and
AY + Y A∗ − Y C ∗ CY + BB ∗ = 0
16.6. Problems 347

Q(A − Y C ∗ C) + (A − Y C ∗ C)∗ Q + C ∗ C = 0.
Suppose Y and Q are balanced; that is,

Y = Q = diag(σ1 , . . . , σr , σr+1 , . . . , σn ) = diag(Σ1 , Σ2 )

and let G(s) be partitioned accordingly as

 
A11 A12 B1
 
G(s) =  A21 A22 B2  .
C1 C2 0
 −1
γ2 γ2
Denote Y1 = Σ1 and X1 = γ 2 −1 Σ1 I− 2
γ 2 −1 Σ1 . Show that
" #
A11 − B1 B1∗ X1 − Y1 C1∗ C1 −Y1 C1∗
Kr (s) =
−B1∗ X1 0

is exactly the reduced-order controller obtained from the last problem with balanced
model reduction procedure. (It is also interesting to note that

Q = X(I + Y X)−1

where X = X ∗ ≥ 0 is the stabilizing solution to

XA + A∗ X − XBB ∗ X + C ∗ C = 0.

Hence balancing Y and Q is equivalent to balancing X and Y . This is called Riccati

balancing; see Jonckheere and Silverman [1983].)

Problem 16.14 Apply the controller reduction

" methods
# in the last three problems,
A B
respectively, to a satellite system G(s) = where
C 0
   
0 1 0 0 0
   
 0 0 0 0   1.7319 × 10−5 
A=

,
 B=

,

 0 0 0 1   0 
0 0 −1.5392 −2 × 0.003 × 1.539 3.7859 × 10−4
h i
C= 1 0 1 0 , D = 0.

Compare the results (see McFarlane and Glover [1990] for further details).
348 H∞ LOOP SHAPING

Problem 16.15 Let f (s) be analytic in the closed right-half plane and suppose

|f (rejθ )
lim max = 0.
r→∞ θ∈[−π/2,π/2] r

Then the Poisson integral formula (see, for example, Freudenberg and Looze [1988],
page 37) says that f (s) at any point s = x + jy in the open right-half plane can be
recovered from f (jω) via the integral relation:
Z
1 ∞ x
f (s) = f (jω) 2 dω.
π −∞ x + (y − ω)2

Let s = rejθ (i.e., x = r cos θ and y = r sin θ) with r > 0 and −π/2 < θ < π/2. Suppose
f (jω) = f (−jω). Show that
Z ∞
f (rejθ ) = f (jω)Kθ (ω/r) d(ln ω)
−∞

where
1 2(ω/r)[1 + (ω/r)2 ] cos θ
Kθ (ω/r) =
π [1 − (ω/r)2 ]2 + 4(ω/r)2 cos2 θ
Chapter 17

Gap Metric and ν-Gap Metric

In the previous chapters, we have seen that all of robust control design techniques assume
that we have some description of the model uncertainties (i.e., we have a measure of
the distance from the nominal plant to the set of uncertainty systems). This measure
is usually chosen to be a metric or a norm. However, the operator norm can be a poor
measure of the distance between systems in respect to feedback control system design.
For example, consider
1 1
P1 (s) = , P2 (s) = .
s s + 0.1
The closed-loop complementary sensitivity functions corresponding to P1 and P2 with
unity feedback are relatively close and their difference is

P1 (I + P1 )−1 − P2 (I + P2 )−1 = 0.0909,
∞

but the difference between P1 and P2 is

kP1 − P2 k∞ = ∞.

This shows that the closed-loop behavior of two systems can be very close even though
the norm of the difference between the two open-loop systems can be arbitrarily large.
To deal with such problems, the gap metric and the ν-gap metric were introduced
into the control literature by Zames and El-Sakkary [1980] (see also El-Sakkary [1985]
and Vinnicombe [1993]) as being appropriate for the study of uncertainty in feedback
systems. An alternative metric, the graph metric, was also introduced by Vidyasagar
[1984] in terms of normalized coprime factorizations. All of these metrics are equivalent,
and thus induce the same topology. This topology is the weakest in which feedback
stability is a robust property. The metrics define notions of distance in the space of
(possible) unstable systems that do not assume that the plants have the same number
of poles in the right-half plane.
We shall briefly introduce the gap metric in Section 17.1 and study some of its
applications in robust control. Our focus in this chapter is Sections 17.2–17.4, which

349
350 GAP METRIC AND ν-GAP METRIC

study in some detail the ν-gap metric. In particular, we introduce the ν-gap metric
in Section 17.2 and explore its frequency domain interpretation and applications in
Section 17.3 and Section 17.4. Finally, we consider controller order reduction in the gap
or ν-gap metric framework in Section 17.5.

17.1 Gap Metric

In this section we briefly introduce the gap metric and discuss some of its applications
in controller design.
Let P (s) be a p × m rational transfer matrix and let P have the following normalized
right and left stable coprime factorizations:

P = N M −1 = M̃ −1 Ñ .

That is,
M ∼ M + N ∼ N = I, M̃ M̃ ∼ + Ñ Ñ ∼ = I.
The graph of the operator P is the subspace of H2 consisting of all pairs (u, y) such
that y = P u. This is given by
" #
M
H2
N

and is a closed subspace of H2 . The gap between two systems P1 and P2 is defined by




" # − Π" #
δg (P1 , P2 ) = Π
M1 M2
H2 H2
N1 N2

where ΠK denotes the orthogonal projection onto K and P1 = N1 M1−1 and P2 =

N2 M2−1 are normalized right coprime factorizations.
It is shown by Georgiou [1988] that the gap metric can be computed as follows:

Theorem 17.1 Let P1 = N1 M1−1 and P2 = N2 M2−1 be normalized right coprime fac-
torizations. Then n o
δg (P1 , P2 ) = max ~δ(P1 , P2 ), ~δ(P2 , P1 )

where ~δg (P1 , P2 ) is the directed gap and can be computed by

" # " #
M

~δg (P1 , P2 ) = inf 1 M 2
− Q .
Q∈H∞ N1 N2
∞
17.1. Gap Metric 351

The following procedures can be used in computing the directed gap ~δg (P1 , P2 ).
Computing ~
δg (P1 , P2 ): Let
" # " #
A1 B1 A2 B2
P1 = , P2 = .
C1 D1 C2 D2

1. Let Pi = Ni Mi−1 , i = 1, 2 be normalized right coprime factorizations. Then

 
" # Ai + Bi Fi Bi Ri
−1/2
Ri = I + Di∗ Di
Mi  −1/2 
= Fi Ri , R̃i = I + Di Di∗
Ni −1/2
Ci + Di Fi Di Ri Fi = −Ri−1 (Bi∗ Xi + Di∗ Ci )

where " #
Ai − Bi Ri−1 Di∗ Ci −Bi Ri−1 Bi∗
Xi = Ric .
−Ci∗ R̃i−1 Ci −(Ai − Bi Ri−1 Di∗ Ci )∗

2. Define a generalized system

 " # " # 
M1 M2
 
G(s) =  N1 N2 
−I 0
 −1/2 
A1 + B1 F1 0 B1 R1 0
 −1/2 
 0 A2 + B2 F2 0 B2 R2 
 

= F1 F2 R1
−1/2
R2
−1/2 .

 −1/2 −1/2 
 C1 + D1 F1 C2 + D2 F2 D1 R1 D2 R2 
0 0 −I 0

3. Apply standard H∞ algorithm to

" # " #
M

~δg (P1 , P2 ) = inf 1 M 2
− Q = inf kF` (G, Q)k∞ .
Q∈H∞ N1 N2 Q∈H∞
∞

Using the above procedure, it is easy to show that

 
1 1
δg , = 0.0995,
s s + 0.1

which confirms that the two systems given at the beginning of this chapter are indeed
close in the gap metric. This example shows an important feature about the gap metric
(similarly, the ν-gap metric defined in the next section): The distance between two
352 GAP METRIC AND ν-GAP METRIC

plants, as measured by the gap metric δg (or the ν-gap metric δν ), has very little to do
with any difference between their open-loop behavior (indeed, there is no reason why it
should). This point will be further illustrated by an example in the next section.
A lower bound for the gap metric can also be obtained easily without actually solving
the corresponding H∞ optimization. Let
" #
M2∼ N2∼
Φ= .
−Ñ2 M̃2

Then Φ∼ Φ = ΦΦ∼ = I and

" # (" # " # )
M∼ N∼
~δg (P1 , P2 ) = M 1 M 2
inf 2 2
− Q
Q∈H∞ −Ñ2 M̃2 N1 N2
∞
" #
M ∼M + N ∼N − Q
1 1
= inf 2 2
≥ kΨ(P1 , P2 )k∞
Q∈H∞ −Ñ2 M1 + M̃2 N1
∞

where
" #" #
h i 0 I M1
Ψ(P1 , P2 ) := −Ñ2 M1 + M̃2 N1 = M̃2 Ñ2 . (17.1)
−I 0 N1

It will be seen in the next section that kΨ(P1 , P2 )k∞ is related to the ν-gap metric. The
above lower bound may actually be achieved. Consider, for example,
k1 k2
P1 = , P2 = .
s+1 s+1
Then it is easy to verify that Pi = Ni /Mi , i = 1, 2, with
k s+1
Ni = pi , Mi = p ,
s + 1 + ki2 s + 1 + ki2
are normalized coprime factorizations and it can be further shown, as in Georgiou and
Smith [1990], that

 |k1 − k2 |

 , if |k1 k2 | > 1;

 1 | + |k2 |
|k
δg (P1 , P2 ) = kΨ(P1 , P2 )k∞ =



 |k1 − k2 |
 p , if |k1 k2 | ≤ 1.
(1 + k12 )(1 + k22 )
An immediate consequence of Theorem 17.1 is the connection between the uncer-
tainties in the gap metric and the uncertainties characterized by the normalized coprime
factors. The following corollary states that a ball of uncertainty in the directed gap is
equivalent to a ball of uncertainty in the normalized coprime factors.
17.1. Gap Metric 353

Corollary 17.2 Let P have a normalized coprime factorization P = N M −1 . Then for

all 0 < b ≤ 1,
n o
P1 : ~δg (P, P1 ) < b
( " # )
∆
−1 N
= P1 : P1 = (N + ∆N )(M + ∆M ) , ∆N , ∆M ∈ H∞ , <b .
∆M
∞

Proof. Suppose ~δg (P, P1 ) < b and let P1 = N1 M1−1 be a normalized right coprime
factorization. Then there exists a Q ∈ H∞ such that
" # " #
M
M1
− Q < b.
N N1
∞

Define " # " # " #

∆M M1 M
:= Q− ∈ H∞ .
∆N N1 N
" #
∆
M
Then < b and P1 = (N1 Q)(M1 Q)−1 = (N + ∆N )(M + ∆M )−1 .
∆N
∞
To show the converse, note that P1 = (N + ∆N )(M + ∆M )−1 and there exists a
n on o−1
Q̃−1 ∈ H∞ such that P1 = (N + ∆N )Q̃ (M + ∆M )Q̃ is a normalized right
coprime factorization. Hence by definition, ~δg (P, P1 ) can be computed as
" # " # " # " #
M M

~δg (P, P1 ) = inf M + ∆M M + ∆M
− Q̃Q ≤ − <b
Q N N + ∆N N N + ∆N ∞
∞

where the first inequality follows by taking Q = Q̃−1 ∈ H∞ . 2

The following is a list of useful properties of the gap metric shown by Georgiou and
Smith [1990].

• If δg (P1 , P2 ) < 1, then δg (P1 , P2 ) = ~δg (P1 , P2 ) = ~δg (P2 , P1 ).

" #!
M (s)
• If b ≤ λ(P ) := inf σ , then
Res>0 N (s)
n o
P1 : ~δ(P, P1 ) < b = {P1 : δ(P, P1 ) < b} .
354 GAP METRIC AND ν-GAP METRIC

Recall that
( " # )−1
I h i
−1
bobt (P ) := inf (I + P K) I P
K stabilizing K
∞
r h i 2
p
= 1 − λmax (Y Q) = 1 − Ñ M̃
H

and
" # " #
I h i −1 I h i −1

bP,K := (I + P K)−1 I P = (I + KP )−1 I K .
K P
∞ ∞

The following results were shown by Qiu and Davison [1992a].

Theorem 17.3 Suppose the feedback system with the pair (P0 , K0 ) is stable. Let P :=
{P : δg (P, P0 ) < r1 } and K := {K : δg (K, K0 ) < r2 }. Then
(a) The feedback system with the pair (P, K) is also stable for all P ∈ P and K ∈ K
if and only if
arcsin bP0 ,K0 ≥ arcsin r1 + arcsin r2 .

(b) The worst possible performance resulting from these sets of plants and controllers
is given by

inf arcsin bP,K = arcsin bP0 ,K0 − arcsin r1 − arcsin r2 .

P ∈P, K∈K

The sufficiency part of the theorem follows from Theorem 17.8 in the next section. Note
that the theorem is still true if one of the uncertainty balls is taken as closed ball. In
particular, one can take either r1 = 0 or r2 = 0.

Example 17.1 Consider

s−1 2s − 1
P1 = , P2 = .
s+1 s+1
Then P1 = N1 /M1 and P2 = N2 /M2 with
1 s−1 1 2s − 1 s+1
N1 = √ , M1 = √ , N2 = √ √ , M2 = √ √
2s+1 2 5s + 2 5s + 2
are normalized coprime factorizations. It is easy to show that
|ω| 1
δg (P1 , P2 ) = 1/3 > kΨ(P1 , P2 )k∞ = sup √ =√ ,
ω 2
10ω + 4 10
17.1. Gap Metric 355

which implies that any controller K that stabilizes P1 and achieves only bP1 ,K > 1/3
will actually stabilize P2 by Theorem 17.3. The following Matlab command can be
used to compute the gap:

 δg (P1 , P2 ) = gap(P1 , P2 , tol)

√
Next, note that bobt (P1 ) = 1/ 2 and the optimal controller√achieving bobt (P1 ) is Kobt =
0. There must be a plant P with δν (P1 , P ) = bobt (P1 ) = 1/ 2 that can not be stabilized
by√Kobt = 0; that is, there must be an unstable plant P such that δν (P1 , P ) = bobt (P1 ) =
1/ 2. A such P can be found using Corollary 17.2:

{P : δg (P1 , P ) ≤ bobt (P1 )}

( " # )
∆
N1 + ∆N N
= P : P = , ∆N , ∆M ∈ H∞ , ≤ bobt (P1 ) .
M1 + ∆M ∆M
∞
" #
∆
N
that is, there must be ∆N , ∆M ∈ H∞ , = bobt (P1 ) such that
∆M
∞

N1 + ∆N
P =
M1 + ∆M
is unstable. Let
1 s−1
∆N = 0, ∆M = √ .
2s+1
Then
N1 + ∆N s−1 √
P = = , δν (P1 , P ) = bobt (P1 ) = 1/ 2.
M1 + ∆M 2s

Example 17.2 We shall now consider the following question: Given an uncertain plant
k
P (s) = , k ∈ [k1 , k2 ],
s−1
k0
(a) Find the best nominal design model P0 = in the sense
s−1
inf sup δg (P, P0 ).
k0 ∈[k1 ,k2 ] k∈[k1 ,k2 ]

(b) Let k1 be fixed and k2 be variable. Find the k0 so that the largest family of the
plant P can be guaranteed to be stabilized a priori by any controller satisfying
bP0 ,K = bobt (P0 ).
356 GAP METRIC AND ν-GAP METRIC

|k0 −k|
For simplicity, suppose k1 ≥ 1. It can be shown that δg (P, P0 ) = k0 +k . Then the
optimal k0 for question (a) satisfies
k0 − k1 k2 − k0
= ;
k0 + k1 k2 + k0
√
that is, k0 = k1 k2 and
√ √
k2 − k1
inf sup δg (P, P0 ) = √ √ .
k0 ∈[k1 ,k2 ] k∈[k1 ,k2 ] k2 + k1
To answer question (b), we note that by Theorem 17.3, a family of plants satisfying
δg (P, P0 ) ≤ r with P0 = k0 /(s + 1) is stabilizable a priori by any controller satisfying
bP0 ,K = bobt (P0 ) if, and only if, r < bP0 ,K . Since P0 = N0 /M0 with
k0 s−1
N0 = p , M0 = p
s+ 1 + k02 s+ 1 + k02
is a normalized coprime factorization, it is easy to show that
" # q p
N k02 + (1 − 1 + k02 )2
0
= p
M0 2 1 + k02
H

and v !
u
u1 1
bobt (P0 ) = t 1+ p .
2 1 + k02
Hence we need to find a k0 such that
 
k0 − k1 k2 − k0
bobt (P0 ) ≥ max , ;
k0 + k1 k2 + k0
that is, v
u !  
u1 1 k0 − k1 k2 − k0
t 1+ p ≥ max ,
2 1 + k02 k0 + k1 k2 + k0
for a largest possible k2 . The optimal k0 is given by the solution of the equation:
v !
u
u1 1 k0 − k1
t 1+ p =
2 1 + k02 k0 + k1

and the largest k2 = k02 /k1 . For example, if k1 = 1, then k0 = 7.147 and k2 = 51.0793.
In general, given a family of plant P , it is not easy to see how to choose a best
nominal model P0 such that (a) or (b) is true. This is still a very important open
question.
17.2. ν-Gap Metric 357

17.2 ν-Gap Metric

The shortfall of the gap metric is that it is not easily related to the frequency response
of the system. On the other hand, the ν-gap metric to be introduced in this section
has a clear frequency domain interpolation and can, in general, be computed from
frequency response. The presentation given in this section, Sections 17.3, 17.4, and
17.5 are based on Vinnicombe [1993a, 1993b], to which readers are referred for further
detailed discussions.
To define the new metric, we shall first review a basic concept from the complex
analysis.

Definition 17.1 Let g(s) be a scalar transfer function and let Γ denote a Nyquist
contour indented around the right of any imaginary axis poles of g(s), as shown in
Figure 17.1. Then the winding number of g(s) with respect to this contour, denoted
by wno(g), is the number of counterclockwise encirclements around the origin by g(s)
evaluated on the Nyquist contour Γ. (A clockwise encirclement counts as a negative
encirclement.)

Figure 17.1: The Nyquist contour

The following argument principle is standard and can be found from any complex
analysis book.

Lemma 17.4 (The Argument Principle) Let Γ be a closed contour in the complex
plane. Let f (s) be a function analytic along the contour; that is, f (s) has no poles on Γ.
Assume f (s) has Z zeros and P poles inside Γ. Then f (s) evaluated along the contour
Γ once in an anticlockwise direction will make Z − P anticlockwise encirclements of the
origin.
358 GAP METRIC AND ν-GAP METRIC

Let G(s) be a matrix (or scalar) transfer matrix. We shall denote η(G) and η0 (G),
respectively, the number of open right-half plane and imaginary axis poles of G(s).
The winding number has the following properties:

Lemma 17.5 Let g and h be biproper rational scalar transfer functions and let F be a
square transfer matrix. Then

(a) wno(gh) = wno(g)+wno(h);

(b) wno(g) = η(g −1 ) − η(g);

(c) wno(g ∼ ) = −wno(g) − η0 (g −1 ) + η0 (g);

(d) wno(1 + g) = 0 if g ∈ RL∞ and kgk∞ < 1;

(e) wno det(I + F ) = 0 if F ∈ RL∞ and kF k∞ < 1.

Proof. Part (a) is obvious by the argument principle. To show part (b), note that by
the argument principle wno(g) equals the excess of the number of open right-half plane
zeros of g over the number of open right-half plane poles of g; that is, wno(g) = η(g −1 )−
η(g), since the number of right-half plane zeros of g is the number of right-half plane
−1
poles of
 g∼ −1
. Next,
 suppose the order of g in part (c) is n. Then η(g ∼ ) = n−η(g)−η
 0 (g)
and η (g ) = n − η(g −1 ) − η0 (g −1 ), which gives wno(g ∼ ) = η (g ∼ )−1 − η(g ∼ ) =
η(g) − η(g −1 ) − η0 (g −1 ) + η0 (g) = −wno(g) − η0 (g −1 ) + η0 (g). Part (d) follows from the
fact that 1 + Reg(jω)Qm > 0, ∀ω since kgk∞ < 1. Finally, part (e) follows from part (d)
and det(I + F ) = i=1 (1 + λi (F )) with |λi (F )| < 1. 2

Example 17.3 Let

1.2(s + 3) s−1 2(s − 1)(s − 2) (s − 1)(s + 3)

g1 = , g2 = , g3 = , g4 = .
s−5 s−2 (s + 3)(s + 4) (s − 2)(s − 4)

Figure 17.2 shows the functions, g1 , g2 , g3 , and g4 , evaluated on the Nyquist contour Γ.
Clearly, we have

wno(g1 ) = −1, wno(g2 ) = 0, wno(g3 ) = 2, wno(g4 ) = −1

and they are consistent with the results computed from using Lemma 17.5.

The ν-gap metric introduced in Vinnicombe [1993a, 1993b] is defined as follows:

17.2. ν-Gap Metric 359

1.5

1
g1
g2
0.5 g3
g4
imaginary

−0.5

−1

−1.5

−2
−1 −0.5 0 0.5 1 1.5 2
real

Figure 17.2: g1 , g2 , g3 , and g4 evaluated on Γ

Definition 17.2 The ν-gap metric is defined as



 kΨ(P1 , P2 )k∞ , if det Θ(jω) 6= 0 ∀ω


 and wno det Θ(s) = 0,
δν (P1 , P2 ) =




 1, otherwise

where Θ(s) := N2∼ N1 + M2∼ M1 and Ψ(P1 , P2 ) := −Ñ2 M1 + M̃2 N1 .

Note that it can be shown as in Vinnicombe [1993a] that

δν (P1 , P2 ) = δν (P2 , P1 ) = δν (P1T , P2T )

and δν is indeed a metric (a proof of this fact is quite complex).

360 GAP METRIC AND ν-GAP METRIC

Computing δν (P1 , P2 ): Let

" # " #
A1 B1 A2 B2
P1 = , P2 = .
C1 D1 C2 D2

1. Let Pi = Ni Mi−1 , i = 1, 2 be normalized right coprime factorizations. Then

 
" # Ai + Bi Fi Bi Ri
−1/2
Ri = I + Di∗ Di
Mi  −1/2 
= Fi Ri , R̃i = I + Di Di∗
Ni −1/2
Ci + Di Fi Di Ri Fi = −Ri−1 (Bi∗ Xi + Di∗ Ci )

where " #
Ai − Bi Ri−1 Di∗ Ci −Bi Ri−1 Bi∗
Xi = Ric .
−Ci∗ R̃i−1 Ci −(Ai − Bi Ri−1 Di∗ Ci )∗

2. Compute the zeros of

" #∼ " #
M2 M1
Θ(s) := N2∼ N1 + M2∼ M1 = .
N2 N1

Let n0 = number of imaginary axis zeros of Θ, n+ = number of open right-half

plane zeros of Θ, and n = number of open right-half plane poles of Θ. Then
wno det(N2∼ N1 + M2∼ M1 ) = n+ − n.
3. If either n0 6= 0 or n+ 6= n, δν (P1 , P2 ) = 1. Otherwise, δν (P1 , P2 ) = kΨ(P1 , P2 )k∞
with Ψ(P1 , P2 ) = −Ñ2 M1 + M̃2 N1 :
" # " #
M̃i Ai + Li Ci Li Bi + Li Di
= −1/2 −1/2 −1/2
Ñi R̃i Ci R̃i R̃i Di

Li = −(Bi Di∗ + Yi Ci∗ )R̃i−1

where " #
(Ai − Bi Di∗ R̃i−1 Ci )∗ −Ci∗ R̃i−1 Ci
Yi = Ric .
−Bi Ri−1 Bi∗ −(Ai − Bi Di∗ R̃i−1 Ci )

The Matlab command nugap can be used to carry out the preceding computation:

 δν (P1 , P2 ) = nugap(P1 , P2 , tol)

where tol is the computational tolerance.

17.2. ν-Gap Metric 361

1
Example 17.4 Consider, for example, P1 = 1 and P2 = . Then
s
1 s 1
M 1 = N1 = √ , M2 = , N2 = .
2 s+1 s+1
Hence
1 1−s 1 1 s−1
Θ(s) = √ =√ , Ψ(P1 , P2 ) = √ ,
2 1 − s 2 2s+1
and δν (P1 , P2 ) = √1 . (Note that Θ has no poles or zeros!)
2

The ν-gap metric can also be computed directly from the system transfer matrices
without first finding the normalized coprime factorizations.

Theorem 17.6 The ν-gap metric can be defined as



 kΨ(P1 , P2 )k∞ , if det(I + P2∼ P1 ) 6= 0 ∀ω and


 wno det(I + P2∼ P1 ) + η(P1 ) − η(P2 ) − η0 (P2 ) = 0,
δν (P1 , P2 ) =




 1, otherwise

where Ψ(P1 , P2 ) can be written as

Ψ(P1 , P2 ) = (I + P2 P2∼ )−1/2 (P1 − P2 )(I + P1∼ P1 )−1/2 .

Proof. Since the number of unstable zeros of M1 (M2 ) is equal to the number of
unstable poles of P1 (P2 ), and

N2∼ N1 + M2∼ M1 = M2∼ (I + P2∼ P1 )M1 ,

we have
wno det(N2∼ N1 + M2∼ M1 ) = wno det {M2∼ (I + P2∼ P1 )M1 }
= wno det M2∼ + wno det(I + P2∼ P1 ) + wno det M1 .
Note that wno det M1 = η(P1 ), wno det M2∼ = −wno det M2 − η0 (M2−1 ) = −η(P2 ) −
η0 (P2 ), and

wno det(N2∼ N1 + M2∼ M1 ) = −η(P2 ) − η0 (P2 ) + wno det(I + P2∼ P1 ) + η(P1 ).

Furthermore,

det(N2∼ N1 + M2∼ M1 ) 6= 0, ∀ω ⇐⇒ det(I + P2∼ P1 ) 6= 0, ∀ω.

362 GAP METRIC AND ν-GAP METRIC

The theorem follows by noting that

Ψ(P1 , P2 ) = (I + P2 P2∼ )−1/2 (P1 − P2 )(I + P1∼ P1 )−1/2
since Ψ(P1 , P2 ) = −Ñ2 M1 + M̃2 N1 = M̃2 (P1 − P2 )M1 and
M̃2∼ M̃2 = (I + P2 P2∼ )−1 , M1 M1∼ = (I + P1∼ P1 )−1 .
2

This alternative formula is useful when doing the hand calculation or when com-
puting from the frequency response of the plants since it does not need to compute the
normalized coprime factorizations.

Example 17.5 Consider two plants P1 = 1 and P2 = 1/s. Then wno det(1 + P2∼ P1 ) =
wno[(s − 1)/s] = 1, as shown in Figure 17.3(a), and wno det(1 + P2∼ P1 ) + η(P1 ) −
η(P2 ) − η0 (P2 ) = 0. On the other hand, wno det(1 + P1∼ P2 ) + η(P2 ) − η(P1 ) = wno
(s + 1)/s = 0, as shown in Figure 17.3(b).

(s+1)/s
10 10

5 (s−1)/s 5

0 0

−5 −5

−10 −10

−10 −5 0 0 5 10
(a) (b)

s−1 s+1
Figure 17.3: and evaluated on Γ
s s

Similar to the gap metric, it is shown by Vinnicombe [1993a, 1993b] that the ν-gap
metric can also be characterized as an optimization problem (however, we shall not use
it for computation).
17.2. ν-Gap Metric 363

Theorem 17.7 Let P1 = N1 M1−1 and P2 = N2 M2−1 be normalized right coprime fac-
torizations. Then
" # " #
M
1 M2
δν (P1 , P2 ) = inf − Q .
−1 N1 N
Q, Q ∈ L ∞
2
∞

wno det(Q) = 0

Moreover, δg (P1 , P2 )bobt (P1 ) ≤ δν (P1 , P2 ) ≤ δg (P1 , P2 ).

It is now easy to see that

{P : δν (P0 , P ) < r}
( " # " # )
∆
−1 ∆N N
⊃ P = (N0 + ∆N )(M0 + ∆M ) : ∈ H∞ , <r .
∆M ∆M
∞
Define " # !
1 I h i
−1
:= σ (I + P (jω)K(jω)) I P (jω)
bP,K (ω) K(jω)
and
ψ(P1 (jω), P2 (jω)) = σ (Ψ(P1 (jω), P2 (jω))) .
The following theorem states that robust stability can be checked using the frequency-
by-frequency test.

Theorem 17.8 Suppose (P0 , K) is stable and δν (P0 , P1 ) < 1. Then (P1 , K) is stable if

bP0 ,K (ω) > ψ(P0 (jω), P1 (jω)), ∀ω.

Moreover,

arcsin bP1 ,K (ω) ≥ arcsin bP0 ,K (ω) − arcsin ψ(P0 (jω), P1 (jω)), ∀ω

and
arcsin bP1 ,K ≥ arcsin bP0 ,K − arcsin δν (P0 , P1 ).

Proof. Let P1 = M̃1−1 Ñ1 , P0 = N0 M0−1 = M̃0−1 Ñ0 and K = U V −1 be normalized

coprime factorizations, respectively. Then
" # !
1 V h i  
−1
=σ (M̃1 V + Ñ1 U ) M̃1 Ñ1 = σ (M̃1 V + Ñ1 U )−1 .
bP1 ,K (ω) U

That is, " #!

h i V
bP1 ,K (ω) = σ(M̃1 V + Ñ1 U ) = σ M̃1 Ñ1 .
U
364 GAP METRIC AND ν-GAP METRIC

Similarly, " #!
h i V
bP0 ,K (ω) = σ(M̃0 V + Ñ0 U ) = σ M̃0 Ñ0 .
U
Note that " #!
h i N0
ψ(P0 (jω), P1 (jω)) = σ M̃1 Ñ1
−M0
" #∼ " #
N0 M̃0∼ N0 M̃0∼
= I.
−M0 Ñ0∼ −M0 Ñ0∼
To simplify the derivation, define
" # " #
N0 h i h i V
G0 = , G̃0 = M̃0 Ñ0 , G̃1 = M̃1 Ñ1 , F = .
−M0 U

Then
ψ(P0 , P1 ) = σ(G̃1 G0 ), bP0 ,K (ω) = σ(G̃0 F ), bP1 ,K (ω) = σ(G̃1 F )
and h i∼ h i h ih i∼
G0 G̃∼
0 G0 G̃∼
0 = I =⇒ G0 G̃∼
0 G0 G̃∼
0 = I.

That is,
G0 G∼ ∼
0 + G̃0 G̃0 = I.

Note that

I = G̃1 G̃∼ ∼ ∼ ∼ ∼ ∼ ∼ ∼
1 = G̃1 (G0 G0 + G̃0 G̃0 )G̃1 = (G̃1 G0 )(G̃1 G0 ) + (G̃1 G̃0 )(G̃1 G̃0 ) .

Hence
σ2 (G̃1 G̃∼
0 ) = 1 − σ (G̃1 G0 ).
2

Similarly,

I = F ∼ F = F ∼ (G0 G∼ ∼ ∼ ∼ ∼ ∼
0 + G̃0 G̃0 )F = (G0 F ) (G0 F ) + (G̃0 F ) (G̃0 F )

=⇒ σ2 (G∼
0 F ) = 1 − σ (G̃0 F ).
2

By the assumption, ψ(P0 , P1 ) < bP0 ,K (ω); that is,

σ(G̃1 G0 ) < σ(G̃0 F ), ∀ω

and q q
σ(G∼
0 F) = 1 − σ2 (G̃0 F ) < 1 − σ 2 (G̃1 G0 ) = σ(G̃1 G̃∼
0 ).

Hence
σ(G̃1 G0 )σ(G∼ ∼
0 F ) < σ(G̃1 G̃0 )σ(G̃0 F );
17.2. ν-Gap Metric 365

that is,
σ(G̃1 G0 G∼ ∼
0 F ) < σ(G̃1 G̃0 G̃0 F ), ∀ω


=⇒ (G̃1 G̃∼
0 G0 F )
−1
(G̃1 G0 G∼
0 F ) < 1.
∞
Now
G̃1 F = G̃1 (G̃∼ ∼ ∼ ∼
0 G̃0 + G0 G0 )F = (G̃1 G̃0 G̃0 F ) + (G̃1 G0 G0 F )
 
= (G̃1 G̃∼ ∼
0 G̃0 F ) I + (G̃1 G̃0 G̃0 F )
−1
(G̃1 G0 G∼
0 F) .

By Lemma 17.5,

wno det(G̃1 F ) = wno det(G̃1 G̃∼ ∼

0 G̃0 F ) = wno det(G̃1 G̃0 ) + wno det(G̃0 F ).

Since (P0 , K) is stable =⇒ (G̃0 F )−1 ∈ H∞ =⇒ η((G̃0 F )−1 ) = 0

=⇒ wno det(G̃0 F ) := η((G̃0 F )−1 ) − η(G̃0 F ) = 0.

Next, note that

P0T = (Ñ0T )(M̃0T )−1 , P1T = (Ñ1T )(M̃1T )−1
and δν (P0T , P1T ) = δν (P0 , P1 ) < 1; then, by definition of δν (P0T , P1T ),

wno det((Ñ0T )∼ (Ñ1T ) + (M̃0T )∼ (M̃1T )) = wno det(G̃1 G̃∼ T ∼

0 ) = wno det(G̃1 G̃0 ) = 0.

Hence wno det(G̃1 F ) = 0, but wno det(G̃1 F ) := η((G̃1 F )−1 ) − η(G̃1 F ) = η((G̃1 F )−1 )
since η(G̃1 F ) = 0, so η((G̃1 F )−1 ) = 0; that is, (P1 , K) is stable.
Finally, note that

G̃1 F = G̃1 (G̃∼ ∼ ∼ ∼

0 G̃0 + G0 G0 )F = (G̃1 G̃0 )(G̃0 F ) + (G̃1 G0 )(G0 F )

and
σ(G̃1 F ) ≥ σ(G̃1 G̃∼ ∼
0 )σ(G̃0 F ) − σ(G̃1 G0 )σ(G0 F )
q q
= 1 − σ 2 (G̃1 G0 )σ(G̃0 F ) − σ(G̃1 G0 ) 1 − σ2 (G̃0 F )

= sin(arcsin σ(G̃0 F ) − arcsin σ(G̃1 G0 ))

= sin(arcsin bP0 ,K (ω) − arcsin ψ(P0 (jω), P1 (jω)))
and, consequently,

arcsin bP1 ,K (ω) ≥ arcsin bP0 ,K (ω) − arcsin ψ(P0 (jω), P1 (jω))

and

inf arcsin bP1 ,K (ω) ≥ inf arcsin bP0 ,K (ω) − sup arcsin ψ(P0 (jω), P1 (jω)).
ω ω ω

That is, arcsin bP1 ,K ≥ arcsin bP0 ,K − arcsin δν (P0 , P1 ). 2

366 GAP METRIC AND ν-GAP METRIC

The significance of the preceding theorem can be illustrated using Figure 17.4. It is
clear from the figure that δν (P0 , P1 ) > bP0 ,K . Thus a frequency-independent stability
test cannot conclude that a stabilizing controller K for P0 will stabilize P1 . However,
the frequency-dependent test in the preceding theorem shows that K stabilizes both P0
and P1 since bP0 ,K (ω) > ψ(P0 (jω), P1 (jω)) for all ω. Furthermore,

bP1 ,K ≥ inf sin (arcsin bP0 ,K (ω) − arcsin ψ(P0 , P1 )) > 0.

ω

b P , K (ω)
0

b P ,K
0

ψ(P0 (jω), P(
1
jω))
δ v(P0 , P1 )

Figure 17.4: K stabilizes both P0 and P1 since bP0 ,K (ω) > ψ(P0 , P1 ) for all ω

The following theorem is one of the main results on the ν-gap metric.

Theorem 17.9 Let P0 be a nominal plant and β ≤ α < bobt (P0 ).

(i) For a given controller K,

arcsin bP,K > arcsin α − arcsin β

for all P satisfying δν (P0 , P ) ≤ β if and only if bP0 ,K > α.

(ii) For a given plant P ,

arcsin bP,K > arcsin α − arcsin β

for all K satisfying bP0 ,K > α if and only if δν (P0 , P ) ≤ β.

Proof. The sufficiency follows essentially from Theorem 17.8. The necessity proof is
harder, see Vinnicombe [1993a, 1993b] for details. 2

The preceding theorem shows that any plant at a distance less than β from the
nominal will be stabilized by any controller stabilizing the nominal with a stability
17.2. ν-Gap Metric 367

margin of β. Furthermore, any plant at a distance greater than β from the nominal will
be destabilized by some controller that stabilizes the nominal with a stability margin of
at least β.
Similarly, one can consider the system robust performance with simultaneous per-
turbations on the plant and controller.

Theorem 17.10 Suppose the feedback system with the pair (P0 , K0 ) is stable. Then

arcsin bP,K ≥ arcsin bP0 ,K0 − arcsin δν (P0 , P ) − arcsin δν (K0 , K)

for any P and K.

Proof. Use the fact that bP,K = bK,P and apply Theorem 17.8 to get

arcsin bP,K ≥ arcsin bP0 ,K − arcsin δν (P0 , P ).

Dually, we have

arcsin bP0 ,K ≥ arcsin bP0 ,K0 − arcsin δν (K0 , K).

Hence the result follows. 2

Example 17.6 Consider again the following example, studied in Vinnicombe [1993b],
with
s−1 2s − 1
P1 = , P2 =
s+1 s+1
and note that
−2s − 1 s − 1 3s + 2
1 + P2∼ P1 = 1 + = .
−s + 1 s + 1 s+1
Then

1 + P2∼ (jω)P1 (jω) 6= 0, ∀ω, wno det(I + P2∼ P1 ) + η(P1 ) − η(P2 ) = 0

and
|P1 − P2 | |ω| 1
δν (P1 , P2 ) = kΨ(P1 , P2 )k∞ = sup p p = sup √ =√ .
ω 1 + |P1 |2 1 + |P2 |2 ω 10ω 2 + 4 10
√
This implies that any controller K that stabilizes P1 and achieves only bP1 ,K > 1/ 10
will actually stabilize P2 . This result is clearly less conservative than that√of using
the gap metric. Furthermore, there exists a controller such that bP1 ,K = 1/ 10 that
destabilizes P2 . Such a controller is K = −1/2, which results in a closed-loop system
with P2 illposed.
368 GAP METRIC AND ν-GAP METRIC

Example 17.7 Consider the following example taken from Vinnicombe [1993b]:
100 100 100
P1 = , P2 = , P3 = .
2s + 1 2s − 1 (s + 1)2
P1 and P2 have very different open-loop characteristics—one is stable, the other unsta-
ble. However, it is easy to show that

δν (P1 , P2 ) = δg (P1 , P2 ) = 0.02, δν (P1 , P3 ) = δg (P1 , P3 ) = 0.8988,

δν (P2 , P3 ) = δg (P2 , P3 ) = 0.8941,

which show that P1 and P2 are very close while P1 and P3 (or P2 and P3 ) are quite far
away. It is not surprising that any reasonable controller for P1 will do well for P2 but
not necessarily for P3 . The closed-loop step responses under unity feedback,

K1 = 1,

are shown in Figure 17.5.

1.8

1.6

1.4

1.2

0.8

0.6

0.4
P1
P2
0.2
P3

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 17.5: Closed-loop step responses with K1 = 1

The corresponding stability margins for the closed-loop systems with P1 and P2 are

bP1 ,K1 = 0.7071, and bP2 ,K1 = 0.7,

respectively, which are very close to their maximally possible margins,

bobt (P1 ) = 0.7106, and bobt (P2 ) = 0.7036

17.2. ν-Gap Metric 369

(in fact, the optimal controllers for P1 and P2 are K = 0.99 and K = 1.01, respectively).
While the stability margin for the closed-loop system with P3 is

bP3 ,K1 = 0.0995,

which is far away from its optimal value, bobt (P3 ) = 0.4307, and results in poor perfor-
mance of the closed loop. In fact, it is not hard to find a controller that will perform
well for both P1 and P2 but will destabilize P3 .
Of course, this does not necessarily mean that all controllers performing reasonably
well with P1 and P2 will do badly with P3 , merely that some do — the unit feedback
being an example. It may be harder to find a controller that will perform reasonably
well with all three plants; the maximally stabilizing controller of P3 ,

2.0954s + 10.8184
K3 = ,
s + 23.2649

is a such controller, which gives

bP1 ,K3 = 0.4307, bP2 ,K3 = 0.4126, and bP3 ,K3 = 0.4307.

The step responses under this control law are shown in Figure 17.6.

1.2

0.8

0.6

P1
P2
0.4 P3

0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

2.0954s + 10.8184
Figure 17.6: Closed-loop step responses with K3 =
s + 23.2649
370 GAP METRIC AND ν-GAP METRIC

17.3 Geometric Interpretation of ν-Gap Metric

The most salient feature of the ν-gap metric is that it can be computed pointwise in
frequency domain:
δν (P1 , P2 ) = sup ψ(P1 (jω), P2 (jω))
ω

provided the winding number condition is satisfied. (For a more extensive coverage of
material presented in this section and the next two sections, readers are encouraged to
consult the original references by Vinnicombe [1992a, 1992b, 1993a, 1993b].)
In particular, for a single-input single-output system,

|P1 (jω) − P2 (jω)|

ψ(P1 (jω), P2 (jω)) = p p . (17.2)
1 + |P1 (jω)|2 1 + |P2 (jω)|2

This function has the interpretation of being the chordal distance between P1 (jω) and
P2 (jω). To illustrate this, consider the Riemann sphere, which is a unit sphere tangent
at its “south pole” to the complex plant at its origin shown in Figure 17.7.

Figure 17.7: Projection onto the Riemann sphere

A point s1 (e.g., s1 = 1−j) in the complex plane is stereographically projected on the

Riemann sphere by connecting the “north pole” to s1 and determining the intersection
of this straight line with the Riemann sphere, resulting in the projection, q1 , of s1 . The
17.3. Geometric Interpretation of ν-Gap Metric 371

coordinates of q1 are
Res1 =s1 |s1 |2
x1 = , y1 = , z1 = .
1 + |s1 |2 1 + |s1 |2 1 + |s1 |2
Thus, the north pole represents the point at infinity and the unit circle is projected onto
the “equator.” The chordal distance between two points, s1 and s2 , is the Euclidean
distance between their stereographical projections, q1 and q2 :
p |s1 − s2 |
d(s1 , s2 ) = (x1 − x2 )2 + (y1 − y2 )2 + (z1 − z2 )2 = p p .
1 + |s1 |2 1 + |s2 |2

0.8

0.6

0.4

0.2

0
−0.5

0.5
0.5
1 0
−0.5
1.5
real axis imaginary axis

Figure 17.8: Projection of a disk on the Nyquist diagram onto the Riemann sphere

Now consider a circle of chordal radius r centered at P0 (jω0 ) on the Riemann sphere
for some frequency ω0 ; that is,
|P (jω0 ) − P0 (jω0 )|
p p = r.
1 + |P (jω0 )|2 1 + |P0 (jω0 )|2

Let P (jω0 ) = R + jI and P0 (jω0 ) = R0 + jI0 . Then it is easy to show that

 2  2
R0 I0 α(1 + |P0 |2 − α)
R− + I− = 6 1
, if α =
1−α 1−α (1 − α)2

where α = r2 (1 + |P0 |2 ). This means that a ball of uncertainty on the ν-gap metric
corresponds to a (large) ball of uncertainty on the Nyquist diagram. Figure 17.8 shows
372 GAP METRIC AND ν-GAP METRIC

a circle of chordal radius 0.2 centered at the stereographical projection of P0 (jω0 ) = 1

and the corresponding circle on the Nyquist diagram.
Figure 17.9 and Figure 17.10 illustrate the uncertainty on the Nyquist diagram
corresponding to the balls of uncertainty on the Riemann sphere centered at p0 with
chordal radius 0.2. For example, an uncertainty of 0.2 at |p0 (jω0 )| = 1 for some ω0 (i.e.,
δν (p0 , p) ≤ 0.2) implies that 0.661 ≤ |p(jω0 )| ≤ 1.513 and the phase difference between
p0 and p is no more than 23.0739o at ω0 .

0.8

0.6

0.4

0.2

0
−1

1 1
0.5
2 0
−0.5
3 −1
real axis imaginary axis

Figure 17.9: Uncertainty on the Riemann sphere and the corresponding uncertainty on
the Nyquist diagram

Note that kΨ(P1 , P2 )k∞ on its own without the winding number condition is useless
for the study of feedback systems. This is illustrated through the following example.

Example 17.8 Consider

s−1−
P1 = 1, P2 = .
s−1
It is clear that P2 becomes increasingly difficult to stabilize as  → 0 due to the near
unstable pole/zero cancellation. In fact, any stabilizing controller for P1 will destabilize
all P2 for  sufficiently small. This is confirmed by noting that bobt (P1 ) = 1, bobt (P2 ) ≈
/2, and
δg (P1 , P2 ) = δν (P1 , P2 ) = 1,  ≥ −2.
||
However, kΨ(P1 , P2 )k∞ = √
4+4+22
≈ 
2 in itself fails to indicate the difficulty of the
problem.
17.4. Extended Loop-Shaping Design 373

0.5
Imaginary Axis 23.0739 degree

p0=0.1
p0=0.5
−0.5 p0=1

p0=1.5
−1
0 0.5 1 1.5 2
Real Axis

Figure 17.10: Uncertainty on the Nyquist diagram corresponding to the balls of uncer-
tainty on the Riemann sphere centered at p0 with chordal radius 0.2

17.4 Extended Loop-Shaping Design

Let P be a family of parametric uncertainty systems and let P0 ∈ P be a nominal design
model. We are interested in finding a controller so that we have the largest possible
robust stability margin; that is,
sup inf bP,K .
K P ∈P

Note that by Theorem 17.8, for any P1 ∈ P, we have

arcsin bP1 ,K (ω) ≥ arcsin bP0 ,K (ω) − arcsin ψ(P0 (jω), P1 (jω)), ∀ω.

Now suppose we need inf P ∈P bP,K > α. Then it is sufficient to have

arcsin bP0 ,K (ω) − arcsin ψ(P0 (jω), P1 (jω)) > arcsin α, ∀ω, P1 ∈ P;

that is,

bP0 ,K (ω) > sin (arcsin ψ(P0 (jω), P1 (jω)) + arcsin α) , ∀ω, P1 ∈ P.

Let W (s) ∈ H∞ be such that

|W (jω)| ≥ sin (arcsin ψ(P0 (jω), P1 (jω)) + arcsin α) , ∀ω, P1 ∈ P.

374 GAP METRIC AND ν-GAP METRIC

Then it is sufficient to guarantee

|W (jω)|
< 1.
bP0 ,K (ω)

Let P0 = M̃0−1 Ñ0 be a normalized left coprime factorization and note that
" # !
1 I
:= σ (I + P0 (jω)K(jω))−1 M̃0−1 (jω) .
bP0 ,K (ω) K(jω)

Then it is sufficient to find a controller so that

" #
I
−1
(I + P0 K) M̃0−1 W < 1.
K
∞

The process can be iterated to find the largest possible α.

Combining the preceding robust stabilization idea and the H∞ loop-shaping in Chap-
ter 16, one can devise an extended loop-shaping design procedure as follows. (A more
advanced loop-shaping procedure can be found in Vinnicombe [1993b].)
Design Procedure:
Let P be a family of parametric uncertain systems and let P0 be a nominal model.

(a) Loop-Shaping: The singular values of the nominal plant are shaped, using a prec-
ompensator W1 and/or a postcompensator W2 , to give a desired open-loop shape.
The nominal plant P0 and the shaping functions W1 , W2 are combined to form
the shaped plant, Ps , where Ps = W2 P0 W1 . We assume that W1 and W2 are such
that Ps contains no hidden modes.
(b) Compute frequency-by-frequency:

f (ω) = sup ψ(Ps (jω), W2 (jω)P (jω)W1 (jω)).

P ∈P

Set α = 0.
(b) Fit a stable and minimum phase rational transfer function W (s) so that

|W (jω)| ≥ sin(arcsin f (ω) + arcsin α) ∀ω.

(c) Find a K∞ such that

" #

I −1 −1
β := inf (I + P0 K∞ ) M̃0 W .
K∞ K∞
∞

(d) If β ≈ 1, stop and the final controller is K = W1 K∞ W2 . If β  1, increase α and

go back to (b). If β  1, decrease α and go back to (b).
17.5. Controller Order Reduction 375

17.5 Controller Order Reduction

The controller order-reduction procedure described in Chapter 15 can, of course, be
applied to the loop-shaping controller design in Chapter 16 and the gap or ν-gap metric
optimization here. However, the controller order-reduction in the loop-shaping con-
troller design, or gap metric, or ν-gap metric optimization is especially simple. The
following theorem follows immediately from Theorems 17.7 and 17.10.
Theorem 17.11 Let P0 be a nominal plant and K0 be a stabilizing controller such
that bP0 ,K0 ≤ bobt (P0 ). Let K0 = U V −1 be a normalized coprime factorization and let
Û , V̂ ∈ RH∞ be such that
" # " #
U Û

− ≤ ε.
V V̂ ∞

Then K := Û V̂ −1 stabilizes P0 if ε < bP0 ,K0 . Furthermore,

arcsin bP,K ≥ arcsin bP0 ,K0 − arcsin ε − arcsin β
for all {P : δν (P, P0 ) ≤ β}.
Hence to reduce the controller order one only needs to approximate the normalized
coprime factors of the controller. An algorithm for finding the best approximation is
also presented in Vinnicombe [1993a, 1993b].

17.6 Notes and References

This chapter is based on Georgiou and Smith [1990] and Vinnicombe [1993a, 1993b].
Early studies of the gap metric can be found in Zames and El-Sakkary [1980] and El-
Sakkary [1985]. The pointwise gap metric was introduced by Qiu and Davision [1992b].

17.7 Problems
Problem 17.1 Calculate the gap δg (Pi , Pj ) with
1 1 s+2 s−2 1
P1 = , P2 = , P3 = , P4 = , P5 = .
s+1 s−1 (s + 1)2 (s + 1)2 (s + 1)2
Problem 17.2 Let P = 10
τ s+1 , τ ∈ [1, 3] and let P0 = 10
τ0 s+1 . Find the optimal τ0 ∈ [1, 3]
minimizing
min max δg (P, P0 ).
τo τ

Problem 17.3 Repeat Problems 17.1–17.2 for the ν-gap metric.

376 GAP METRIC AND ν-GAP METRIC
Chapter 18

Miscellaneous Topics

This chapter considers two somewhat different problems. The first section gives a brief
introduction into the problem of model validation and the second section considers the
mixed real and complex µ analysis and synthesis.

18.1 Model Validation

A key to the success of the robust control theory developed in this book is to have
appropriate descriptions of the uncertain system (whether it is an additive uncertainty
model or a general linear fractional model). Then an important question is how one can
decide if a model description is appropriate (i.e., how to validate a model).
For simplicity of presentation, we have chosen to present the discrete time model
validation in this section with the understanding that a continuous time problem can be
approximated by fast sampling. Suppose we have modeled a set of uncertain dynamical
systems by
∆ := {∆ : ∆ ∈ H∞ , k∆k∞ ≤ 1}

where the H∞ norm of a discrete time system ∆ ∈ H∞ is defined as k∆(z)k∞ =

sup|z|>1 σ (∆(z)). In order to verify whether this model assumption is correct, some
experimental data are collected. For example, let the input to the system be the sequence
u = (u0 , u1 , . . . , ul−1 ) and the output y = (y0 , y1 , . . . , yl−1 ). A natural question is
whether these data are consistent with our modeling assumption. In other words, does
there exist a model ∆ ∈ ∆ such that the output of the ∆ for the period of t =
0, 1, . . . , l − 1 is exactly y = (y0 , y1 , . . . , yl−1 ) with the input u = (u0 , u1 , . . . , ul−1 )? If
there does not exist a such ∆, then the model is invalidated. If there exists a such ∆,
however, it does not mean that the model is validated but it only means that the model
is not validated by this set of data and it may be invalidated by another set of data in
the future. Hence it is actually more accurate to say our validation procedure is model
invalidation.

377
378 MISCELLANEOUS TOPICS

Let ∆ be a stable, causal, linear, time-invariant system with transfer matrix

∆(z) = h0 + h1 z −1 + h2 z −2 + · · ·
where hi , i = 0, 1, . . . are the matrix Markov parameters. Suppose we have applied the
input sequence u = (u0 , u1 , . . . , ul−1 ) to the system and collected the output for the
period t = 0, 1, . . . , ` − 1, y = (y0 , y1 , . . . , yl−1 ). Then the input and output sequences
are related by a Toeplitz matrix:
    
y0 h0 0 ··· 0 u0
     
 y1  
..
. 0  u1 
 . = h1 h0
 . 

 .   . 
.
 .   
.. .. ..
 . . . 0  . 
yl−1 hl−1 hl−2 · · · h0 ul−1

This equation shows that, for u0 6= 0 and SISO ∆, the inputs and outputs uniquely
determine the first ` Markov parameters of the transfer function ∆(z). The model is
validated (or more accurately not invalidated) if the remaining Markov parameters can
be chosen so that ∆(z) ∈ ∆. The existence of such a choice is the classical tangential
Carathéodory-Fejér interpolation problem, for which a solution to the MIMO case can
be found in Foias and Frazho [1990, page 195]. We shall state this result in the following
theorem. But we shall define some notation first.
Let (v0 , v1 , . . . , v`−1 , v` , v`+1 , . . .) be a sequence and let π` denote the truncation
operator such that
π` (v0 , v1 , . . . , v`−1 , v` , v`+1 , . . .) = (v0 , v1 , . . . , v`−1 ).
Let v = (v0 , v1 , . . . , v`−1 ) be a sequence of vectors and denote
 
v0 0 ··· 0
 
 v ..
. 0 
 1 v0 
Tv :=  . .
 . .. .. 
 . . . 0 
vl−1 vl−2 · · · v0

Theorem 18.1 Given u = (u0 , u1 , . . . , ul−1 ) and y = (y0 , y1 , . . . , yl−1 ), there exists a
∆ ∈ H∞ , k∆k∞ ≤ 1 such that
y = π` ∆u
if and only if Ty∗ Ty ≤ Tu∗ Tu .
Note that the output of ∆ after time t = ` − 1 is irrelevant to the test. The condition
Ty∗ Ty ≤ Tu∗ Tu is equivalent to

X
i X
i
kyj k2 ≤ kuj k2 , i = 0, 1, . . . , ` − 1
j=1 j=1
18.1. Model Validation 379

or
kπi yk2 ≤ kπi uk2 , i = 0, 1, . . . , ` − 1,
which is obviously necessary. In fact, the last condition holds for stable, linear, time-
k∆uk
varying operator ∆ with supu6=0 kuk 2 ≤ 1; see Poolla et al. [1994]. Note that if
2
u0 6= 0, then  Tu is of full column rank and the condition can also be written as
σ Ty (Tu∗ Tu )− 2 ≤ 1.
1

Using the above theorem, we can derive solutions to some model validation problems
easily. For example, consider a set of additive models shown in Figure 18.1.

d
?
D(z) ∆(z)  W (z) 

 ?
f ?
f P (z) 
y
u

Figure 18.1: Model validation for additive uncertainty

In this case,
y = (P + ∆W )u + Dd, k∆k∞ ≤ 1
where P (z), W (z), D(z) and ∆(z) are causal, linear, time-invariant systems (but not
necessarily stable). The disturbance d is assumed to come from a convex set, d ∈ Dconvex ;
for example, Dconvex = {d : d ∈ `2 [0, ∞), kdk2 ≤ 1}. For simplicity, we shall also assume
that W (∞) is of full column rank. Let

D(z) = D0 + D1 z −1 + D2 z −2 + · · · .

Theorem 18.2 Given a set of input-output data uexpt = (u0 , u1 , . . . , u`−1 ) with u0 6= 0,
yexpt = (y0 , y1 , . . . , y`−1 ) for the additive perturbed uncertainty system with an additive
disturbance d ∈ Dconvex , where Dconvex is a convex set, let

û = (û0 , û1 , . . . , û`−1 ) = π` (W uexpt )

ŷ = (ŷ0 , ŷ1 , . . . , ŷ`−1 ) = yexpt − π` P uexpt .

Then there exists a ∆ ∈ H∞ , k∆k∞ ≤ 1 such that

yexpt = π` ((P + ∆W )uexpt + Dd)

for some d ∈ Dconvex if and only if there exists a d = (d0 , d1 , . . . , dl−1 ) ∈ π` Dconvex such
that h i
σ (Tŷ − TD Td )(Tû∗ Tû )−1/2 ≤ 1
380 MISCELLANEOUS TOPICS

where  
D0 0 ··· 0
 
 D ..
0 
 1 D0 . 
TD :=  . .
 . .. .. 
 . . . 0 
Dl−1 Dl−2 ··· D0

Proof. Note that the system input-output equation can be written as

(y − P u) − Dd = ∆(W u).

Since P, W, D, and ∆ are causal, linear, and time invariant, we have π` Dd = π` Dπ` d,
π` (y − P u) = yexpt − π` P π` u = yexpt − π` P uexpt and π` W u = π` W π` u = π` W uexpt .
Denote
dˆ = (dˆ0 , dˆ1 , . . . , dˆ`−1 ) = π` (Dd).
Then it is easy to show that
   
dˆ0 d0
   
 dˆ1   d1 
  = TD  
 ..   .. 
 .   . 
ˆ
d`−1 d`−1

and Td̂ = TD Td . Now note that

Tπ` (y−P u−Dd) = Tπ` (y−P u) − Tπ` (Dd) = Tŷ − TD Td , Tπ` W u = Tû

and π` ∆W u = π` ∆π` (W u) since ∆ is causal. Applying Theorem 18.1, there exists a

∆ ∈ H∞ , k∆k∞ ≤ 1 such that

π` [(y − P u) − Dd] = π` ∆(W u) = π` ∆π` (W u)

if and only if
(Tŷ − TD Td )∗ (Tŷ − TD Td ) ≤ Tû∗ Tû
which is equivalent to h i
σ (Tŷ − TD Td )(Tû∗ Tû )− 2 ≤ 1.
1

Note that Tû is of full column rank since W (∞) is of full column rank and u0 6= 0,
6 0.
which implies û0 = 2

Note that h i
σ (Tŷ − TD Td )(Tû∗ Tû )− 2 ≤ 1
1
inf
d∈Dconvex

is a convex problem and can be checked numerically.

18.2. Mixed µ Analysis and Synthesis 381

Many other classes of model validation problems can be solved analogously. For
example, consider a coprime factor model validation problem with

y = (M + ∆M WM )−1 (N + ∆N WN )u + d

whereh M, N, WM , iand WN are causal, linear, time-invariant systems, and ∆M , ∆N ∈

H∞ , ∆M ∆N ≤ 1. Then the problem can be solved by multiplying M +∆M WM
∞
from the left of the system equation and rewriting the system equation as
" #
h i W (d − y)
M
(M y − N u − M d) = ∆M ∆N .
WN u

The model validation of a general LFT uncertainty system is considered in Davis [1995]
and Chen and Wang [1996]. For continuous time model validation, see Rangan and
Poolla [1996] and Smith and Dullerud [1996].

18.2 Mixed µ Analysis and Synthesis

In Chapter 10, we considered analysis and synthesis of systems with complex uncertain-
ties. However, in practice, many systems involve parametric uncertainties that are real
(for example, the uncertainty about a spring constant in a mechanical system). In this
case, one has to cover this real parameter variation with a complex disk in order to
use the complex µ analysis and synthesis tools, which usually results in a conservative
solution. In this section, we shall consider briefly the analysis and synthesis problems
with possibly both real parametric and complex uncertainties.
The mixed real and complex µ involves three types of blocks: repeated real scalar,
repeated complex scalar, and full blocks. Three nonnegative integers, Sr , Sc , and F ,
represent the number of repeated real scalar blocks, the number of repeated complex
scalar blocks, and the number of full blocks, and they satisfy

X
Sr X
Sc X
F
ki + ri + mj = n.
i=1 i=1 j=1

The ith repeated real scalar block is ki × ki , the jth repeated complex scalar block is
rj × rj , and the `th full block is m` × m` . The admissible set of uncertainties ∆ ⊂ Cn×n
is defined as  
∆ = diag φ1 Ik1 , . . . , φsr Iksr , δ1 Ir1 , . . . , δsc Irsc ,

∆1 , . . . , ∆F ] : φi ∈ R, δj ∈ C, ∆` ∈ Cm` ×m` . (18.1)
The mixed µ is defined in the same way as for the complex µ: Let M ∈ Cn×n ; then

µ∆ (M ) := (min {σ(∆) : ∆ ∈ ∆, det (I − M ∆) = 0})−1 (18.2)

382 MISCELLANEOUS TOPICS

unless no ∆ ∈ ∆ makes I −M ∆ singular, in which case µ∆ (M ) := 0. Or, equivalently,

1
:= inf {α : det(I − αM ∆) = 0, σ(∆) ≤ 1, ∆ ∈ ∆} .
µ∆ (M )
Let ρR (M ) be the real spectral radius (i.e., the largest magnitude of the real eigen-
values of M ). For√example, if a 4 × 4 matrix M has eigenvalues 1 ± j3, −2, 1, then
ρ(M ) = |1 + j3| = 10 and ρR (M ) = | − 2| = 2. It is easy to see that

µ∆ (M ) = max ρR (M ∆)
∆∈B∆

where B∆ := {∆ : ∆ ∈ ∆, σ(∆) ≤ 1}. Note that max ρR (M ∆) = max ρ(M ∆) if

∆∈B∆ ∆∈B∆
sr = 0. [This should not be confused with the fact that, for a given matrix ∆ ∈ B∆
π
and M , ρR (M ∆) may not be equal to ρ(M ∆). For example, M = 2ej 4 and ∆ = 1;
then ρ(M ∆) = 2 but ρR (M ∆) = 0 since M has no real eigenvalues. However, one can
choose another ∆1 = e−j 4 such that ρR (M ∆1 ) = 2 = ρ(M ∆).]
π

Define
Q = {∆ ∈ ∆ : φi ∈ [−1, 1], |δi | = 1, ∆i ∆∗i = Imi }
( h i )
diag D̃1 , . . . , D̃sr , D1 , . . . , Dsc , d1 Im1 , . . . , dF −1 ImF −1 , ImF :
D = .
D̃i ∈ Cki ×ki , D̃i = D̃i∗ > 0, Di ∈ Cri ×ri , Di = Di∗ > 0, dj ∈ R, dj > 0

G = diag [G1 , . . . , Gsr , 0, . . . , 0] : Gi = G∗i ∈ Cki ×ki .
It was shown in Young [1993] that

µ∆ (M ) = max ρR (QM ).
Q∈Q

Note that the above maximization is not necessarily achieved on the vertices for the
real parameters; hence one must search over the entire interval for each real parameter.
Again this maximization problem can have many local maximums and a power algorithm
has been developed in Young [1993] to compute a lower bound.
It should also be noted that even though the complex µ (i.e., sr = 0) is a continuous
function of the data, the mixed µ (i.e., sr 6= 0) may only be upper semicontinuous; see
Packard and Pandey [1993]. It was also shown in Braatz et al. [1994] and Toker and
Özbay [1995] that the computation of µ is a NP hard problem, which means that it
may not be computable in a polynomial time. Of course, it should not be interpreted
as every µ problem will not be solvable in a polynomial time; merely some might not.
Obviously, the upper bound for the complex µ can be applied for the mixed µ when
the intervals of the real parameters are covered by complex disks. However, a better
bound can be obtained for the mixed µ by exploiting the phase information of the real
parameters. To motivate the improved bound for the mixed µ, we consider again the
upper bound for the complex µ problem. It is known that

µ∆ (M ) ≤ inf σ(DM D−1 ).

D∈D
18.2. Mixed µ Analysis and Synthesis 383

This bound can be reformulated using linear matrix inequalities by noting the following:

σ(DM D−1 ) ≤ β ⇐⇒ (DM D−1 )∗ DM D−1 ≤ β 2 I ⇐⇒ M ∗ D∗ DM − β 2 D∗ D ≤ 0.

Since D is nonsingular and D∗ D ∈ D, we have


µ∆ (M ) ≤ inf min β : M ∗ DM − β 2 D ≤ 0 .
D∈D β

The following upper bound for the mixed µ was derived by Fan, Tits, and Doyle [1991]
and reformulated in the current form by Young [1993].

Theorem 18.3 Let M ∈ Cn×n and ∆ ∈ ∆. Then


µ∆ (M ) ≤ inf min β : M ∗ DM + j(GM − M ∗ G) − β 2 D ≤ 0 .
D∈D,G∈G β

Proof. Suppose we have a Q ∈ Q such that QM has a real eigenvalue λ ∈ R. Then

there is a vector x ∈ Cn such that

QM x = λx.
1 1 1
Let D ∈ D. Then D ∈ D, D Q = QD 2 and
2 2

1 1 1
D 2 QM x = QD 2 M x = λD 2 x.

Since σ(Q) ≤ 1, it follows that

1 2 2 1 2
1
λ2 D 2 x = QD 2 M x ≤ D 2 M x .

Hence
x∗ (M ∗ DM − λ2 D)x ≥ 0.
Next, let G ∈ G and note that Q∗ G = QG = GQ; then
 ∗
1 1 1
x∗ GM x = QM x GM x = x∗ M ∗ Q∗ GM x = x∗ M ∗ QGM x
λ λ λ
1 ∗ ∗ 1
= x M GQM x = x∗ M ∗ G(QM x) = x∗ M ∗ Gx.
λ λ
That is,
x∗ (GM − M ∗ G)x = 0.
Note that j(GM − M ∗ G) is a Hermitian matrix, so it follows that for such x

x∗ (M ∗ DM + j(GM − M ∗ G) − λ2 D)x ≥ 0.

It is now easy to see that if we have D ∈ D, G ∈ G and 0 ≤ β ∈ R such that

M ∗ DM + j(GM − M ∗ G) − β 2 D ≤ 0
384 MISCELLANEOUS TOPICS

then |λ| ≤ β, and hence µ∆ (M ) ≤ β. 2

This upper bound has an interesting interpretation: covering the uncertainties on

the real axis using possibly off-axis disks. To illustrate, let M ∈ C be a scalar and
∆ ∈ [−1, 1]. We can cover this real interval using a disk as shown in Figure 18.2.
The off-axis disk can be expressed as
s  2  
G
j + 1+
G ˜
∆, ˜ ∈ C, ∆
∆ ˜  ≤ 1.
β β

Im

Im
j G
j
β

-1 1 Re -1 1 Re

-j

Centered Disk Off-Axis Disk

Figure 18.2: Covering real parameters with disks

Hence 1 − ∆ M
β 6= 0 for all ∆ ∈ [−1, 1] is guaranteed if
 s 
 2  
G
1 − j + 1+
G ˜  M 6= 0,
∆ ˜ ∈ C, ∆
∆ ˜  ≤ 1
β β β

r  2
1+ G M
 
˜ ∈ C, ∆
˜  ≤ 1
β β
⇐⇒ 1 − ˜ 6= 0,
∆ ∆
1− jG M
β β
r  2 ∗  r  2 
G M G M
 1+ β β   1+ β β 
⇐⇒ 
 1−jG M
 
  1−jG M
≤1

β β β β

M∗ M GM M∗ G
⇐⇒ + j( − )−1≤0
β β β β β β
⇐⇒ M ∗ M + j(GM − M ∗ G) − β 2 ≤ 0.
18.2. Mixed µ Analysis and Synthesis 385

The scaling G allows one to exploit the phase information about the real parameters so
that a better upper bound can be obtained. We shall demonstrate this further using a
simple example.

Example 18.1 Let

s2 + 2s + 1
G(s) = .
s3 + s2 + 2s + 1

0.5

−0.5

−1

Nyquist diagram 1/G

disk centered at (0,0)
−1.5
disk centered at (0,−0.2j)
disk centered at (0, −j)

−2
−1.5 −1 −0.5 0 0.5 1 1.5

Figure 18.3: Computing the real stability margin by covering with disks

We are interested in finding the largest k such that 1 + ∆G(s) has no zero in the
right-half plane for all ∆ ∈ [−k, k]. Of course, the largest k can be found very easily by
using well-known stability test, which gives
 −1  −1
kmax = sup µ∆ (G(jω)) = sup max ρR (φG(jω))
ω ω φ∈[−1,1]

 −1  
1
= sup {|G(jω)| : =G(jω) = 0} = inf : =G(jω) = 0 = 0.5.
ω ω |G(jω)|
386 MISCELLANEOUS TOPICS

Now we use the complex covering idea to find the best possible k. Note that we only
need to find the smallest |∆| so that 1 + ∆G(jω0 ) = 0 for some ω0 or, equivalently,
∆ + 1/G(jω0 ) = 0. The frequency response of 1/G and the disks covering an interval
[−k, k] are shown in Figure 18.3. It is clear that a centered disk would give k =
1/ kGk∞ = 0.2970 and an off-axis disk centered at (0, −0.2j) would give k = 0.3984
while an off-axis disk centered at (0, −j) would give the exactly value k = 0.5.

The following alternative characterization of the upper bound is useful in the mixed
µ synthesis.

Theorem 18.4 Given β > 0, there exist D ∈ D and G ∈ G such that

M ∗ DM + j(GM − M ∗ G) − β 2 D ≤ 0

if and only if there are D1 ∈ D and G1 ∈ G such that

  
D1 M D1−1
− jG1 (I + G21 )− 2 ≤ 1.
1
σ
β

Proof. Let D = D12 and G = βD1 G1 D1 . Then

M ∗ DM + j(GM − M ∗ G) − β 2 D ≤ 0

⇐⇒ M ∗ D12 M + j(βD1 G1 D1 M − βM ∗ D1 G1 D1 ) − β 2 D12 ≤ 0

⇐⇒ (D1 M D1−1 )∗ (D1 M D1−1 ) + j(βG1 D1 M D1−1 − β(D1 M D1−1 )∗ G1 ) − β 2 I ≤ 0

 ∗  
D1 M D1−1 D1 M D1−1
⇐⇒ − jG1 − jG1 − (I + G21 ) ≤ 0
β β
  
D1 M D1−1 2 − 12
⇐⇒ σ − jG1 (I + G1 ) ≤ 1.
β
2

Similarly, the following corollary can be shown.

Corollary 18.5 µ∆ (M ) ≤ rβ if there are D1 ∈ D and G1 ∈ G such that

  
D1 M D1−1 2 − 12
σ − jG1 (I + G1 ) ≤ r ≤ 1.
β
18.2. Mixed µ Analysis and Synthesis 387

Proof. This follows by noting that

  
D1 M D1−1
− jG1 (I + G21 )− 2 ≤ r ≤ 1
1
σ
β
 ∗    2
D1 M D1−1 G1 D1 M D1−1 G1 G1
=⇒ −j −j ≤I+ G21 ≤I+ .
rβ r rβ r r
G1
Let G2 = ∈ G. Then
r
 ∗  
D1 M D1−1 D1 M D1−1
− jG2 − jG2 ≤ I + G22
rβ rβ
  
D1 M D1−1 2 − 12
=⇒ σ − jG2 (I + G2 ) ≤1
rβ
=⇒ µ∆ (M ) ≤ rβ.
2

Note that this corollary is not necessarily true if r > 1. It is fairly easy to check
that the well-posedness condition, main loop theorem, robust stability, and robust per-
formance theorems for the mixed µ setup are exactly the same as the ones for complex
µ problems.
We are now in the position to consider the synthesis problem with mixed uncer-
tainties. Consider again the general system diagram in Figure 18.4. By the robust
performance condition, we need to find a stabilizing controller K so that

min sup µ∆ (F` (P, K)) ≤ β.

K ω

z w
 
P 

- K

Figure 18.4: Synthesis framework

By Theorems 18.3 and 18.4, µ∆ (F` (P (jω), K(jω))) ≤ β, ∀ω if there are frequency-
dependent scaling matrices Dω ∈ D and Gω ∈ G such that
  
Dω (F` (P (jω), K(jω))) Dω−1 2 − 12
sup σ − jGω (I + Gω ) ≤ 1, ∀ω.
ω β
388 MISCELLANEOUS TOPICS

Similar to the complex µ synthesis, we can now describe a mixed µ synthesis procedure
that involves D, G − K iterations.

D, G − K Iteration:

(1) Let K be a stabilizing controller. Find initial estimates of the scaling matrices
Dω ∈ D, Gω ∈ G and a scalar β1 > 0 such that
  
Dω (F` (P (jω), K(jω))) Dω−1 2 − 12
sup σ − jGω (I + Gω ) ≤ 1, ∀ω.
ω β1

Obviously, one may start with Dω = I, Gω = 0, and a large β1 > 0.

(2) Fit the frequency response matrices Dω and jGω with D(s) and G(s) so that

D(jω) ≈ Dω , G(jω) ≈ jGω , ∀ ω.

Then for s = jω
  
Dω (F` (P (jω), K(jω))) Dω−1
− jGω (I + G2ω )− 2
1
sup σ
ω β1
  
D(s) (F` (P (s), K(s))) D−1 (s) ∼ − 12
≈ sup σ − G(s) (I + G (s)G(s)) .
ω β1

(3) Let D(s) be factorized as

∼ −1
D(s) = Dap (s)Dmin (s), Dap (s)Dap (s) = I, Dmin (s), Dmin (s) ∈ H∞ .

That is, Dap is an all-pass and Dmin is a stable and minimum phase transfer
matrix. Find a normalized right coprime factorization
∼
Dap (s)G(s)Dap (s) = GN G−1
M , GN , GM ∈ H∞

such that
G∼ ∼
M GM + GN GN = I.

Then
G−1 ∼ ∼
M Dap (I + G G)
−1
Dap (G−1 ∼
M ) =I

and, for each frequency s = jω, we have

  
D(s) (F` (P (s), K(s))) D−1 (s) ∼ − 12
σ − G(s) (I + G (s)G(s))
β1
 −1  
Dmin (F` (P, K)) Dmin ∼ ∼ ∼ − 12
=σ − Dap GDap Dap (I + G G)
β1
18.3. Notes and References 389

 −1  
Dmin (F` (P, K)) Dmin
=σ − GN G−1
M D ∼
ap (I + G∼
G)− 12
β1
 −1  
Dmin (F` (P, K)) Dmin GM −1 ∼ ∼ − 12
=σ − GN GM Dap (I + G G)
β1
 −1 
Dmin (F` (P, K)) Dmin GM
=σ − GN .
β1

(4) Define
" # " # " #
−1
Dmin (s) Dmin (s)GM (s) GN
Pa = P (s) − β1
I I 0

and find a controller Knew minimizing kF` (Pa , K)k∞ .

(5) Compute a new β1 as

β1 = sup inf {β(ω) : Γ ≤ 1}

ω D̃ω ∈D,G̃ω ∈G

where " ! #
D̃ω F` (P, Knew )D̃ω−1
G̃2ω )− 2
1
Γ := σ − j G̃ω (I + .
β(ω)

(6) Find D̂ω and Ĝω such that

" ! #
D̂ω F` (P, Knew )D̂ω−1
Ĝ2ω )− 2
1
inf σ − j Ĝω (I + .
D̂ω ∈D,Ĝω ∈G β1

(7) Compare the new scaling matrices D̂ω and Ĝω with the previous estimates Dω
and Gω . Stop if they are close, else replace Dω , Gω and K with D̂ω , Ĝω and
Knew , respectively, and go back to step (2).

18.3 Notes and References

The model validation problems are discussed in Smith and Doyle [1992] in the frequency
domain; in Poolla, Khargonekar, Tikku, Krause, Nagpal [1994 in the discrete time
domain (on which Section 18.1 is based); and in Rangan and Poolla [1996] and Smith
and Dullerud [1996] in the continuous time domain. See also Davis [1995] and Chen
and Wang [1996]. The mixed µ problems are discussed in detail in Young [1993] (on
which Section 18.2 is based), Fan, Tits, and Doyle [1991], Packard and Pandey [1993],
and references therein.
390 MISCELLANEOUS TOPICS

18.4 Problems
Problem 18.1 Write a Matlab program for the additive model validation problem
and try it on a simple experiment in your laboratory.
Bibliography

[1] Al-Saggaf, U. M. and G. F. Franklin (1987). “An error bound for a discrete reduced
order model of a linear multivariable system,” IEEE Trans. Automat. Contr., Vol.
AC-32, pp. 815-819.

[2] Al-Saggaf, U. M. and G. F. Franklin (1988). “Model reduction via balanced realiza-
tions: an extension and frequency weighting techniques,” IEEE Trans. Automat.
Contr., Vol. AC-33, No. 7, pp. 687-692.
[3] Anderson, B. D. O. (1967). “An algebraic solution to the spectral factorization
problem,” IEEE Trans. Automat. Contr., Vol. AC-12, pp. 410-414.

[4] Anderson, B. D. O. (1993). “Bode Prize Lecture (Control design: moving from
theory to practice),” IEEE Control Systems, Vol. 13, No.4, pp. 16-25.

[5] Anderson, B. D. O. and Y. Liu (1989). “Controller reduction: concepts and ap-
proaches, ” IEEE Trans. Automat. Contr. , Vol. AC-34, No. 8, pp. 802-812.

[6] Anderson, B. D. O. and J. B. Moore (1989). Optimal Control: Linear Quadratic

Methods. Prentice Hall, Englewood Cliffs, New Jersey.
[7] Anderson, B. D. O. and S. Vongpanitlerd (1973). Network Analysis and Synthesis:
A Modern System Theory Approach. Prentice Hall, Englewood Cliffs, New Jersey.
[8] Arnold, W. F. and A. J. Laub (1984). “Generalized Eigenproblem algorithms and
software for algebraic Riccati equations,” Proceedings of the IEEE, Vol. 72, No.
12, pp. 1746-1754.
[9] Balas, G. (1990). Robust Control of Flexible Structures: Theory and Experiments,
PhD. Thesis, California Institute of Technology.
[10] Balas, G., J. C. Doyle, K. Glover, A. Packard, and R. Smith (1994). µ-Analysis
and Synthesis Toolbox, MUSYN Inc. and The MathWorks, Inc.
[11] Başar, T. and P. Bernhard (1991). H∞ -Optimal Control and Related Minimax
Design Problems: A Dynamic Game Approach. Systems and Control: Foundations
and Applications. Birkhäuser, Boston.

391
392 BIBLIOGRAPHY

[12] Bernstein, D. S. and W. M. Haddard (1989). “LQG control with an H∞ perfor-

mance bound: A Riccati equation approach,” IEEE Trans. Automat. Contr., Vol.
AC-34, pp. 293-305.

[13] Bode, H. W. (1945). Network Analysis and Feedback Amplifier Design, Van Nos-
trand, Princeton.

[14] Boyd, S., V. Balakrishnan, and P. Kabamba (1989). “A bisection method for com-
puting the H∞ norm of a transfer matrix and related problems,” Math. Control,
Signals, and Systems, Vol. 2, No. 3, pp. 207-220.

[15] Boyd, S. and C. Barratt (1991). Linear Controller Design -Limits of Performance,
Prentice Hall, Englewood Cliffs, New Jersey.

[16] Boyd, S. and C. A. Desoer (1985). “Subharmonic functions and performance

bounds in linear time-invariant feedback systems,” IMA J. Math. Contr. and
Info., Vol. 2, pp. 153-170.

[17] Boyd, S., L. El Ghaoui, E. Feron, and V. Balakrishnan (1994). Linear Matrix
Inequalities in Systems and Control Theory, SIAM, Philadelphia.

[18] Braatz, R. D., P. M. Young, J. C. Doyle, and M. Morari (1994). “Computational

complexity of µ calculation,” IEEE Trans. Automat. Contr., Vol. 39, No. 5, pp.
1000-1002.

[19] Brogan, W. L. (1991). Modern Control Theory, 3rd ed., Prentice Hall, Englewood
Cliffs, New Jersey.

[20] Bruinsma, N. A. and M. Steinbuch (1990). “A fast algorithm to compute the

H∞ -norm of a transfer function matrix”, Systems and Control Letters, Vol. 14,
pp. 287-293.

[21] Bryson Jr., A. E. and Y-C. Ho (1975). Applied Optimal Control, Hemisphere
Publishing Corporation.

[22] Chen, C. T. (1984). Linear System Theory and Design, Holt, Rinehart and Win-
ston.

[23] Chen, J. (1995). “Sensitivity integral relations and design trade-offs in linear mul-
tivariable feedback systems,” IEEE Trans. Auto. Contr., Vol. 40, No. 10, pp.
1700-1716.

[24] Chen, J. and S. Wang (1996). “Validation of linear fractional uncertain models:
solutions via matrix inequalities,” IEEE Trans. Auto. Contr., Vol. 41, No. 6, pp.
844-849.

[25] Chen, T. and B. A. Francis (1992). “H2 -optimal sampled-data control,” IEEE
Trans. Automat. Contr., Vol. 36, No. 4, pp. 387-397.
BIBLIOGRAPHY 393

[26] Chen, X. and K. Zhou (1996). “On the algebraic approach to H∞ control,” 1996
IFAC World Congress, San Francisco, California, pp. 505-510.

[27] Chilali, M. and P. Gahinet (1996). “H∞ design with pole placement constraints:
an LMI approach,” IEEE Trans. Automat. Contr., Vol. 41, No. 3, pp. 358-367.

[28] Christian, L. and J. Freudenberg (1994). “Limits on achievable robustness against

coprime factor uncertainty,” Automatica, Vol. 30, No. 11, pp. 1693-1702.

[29] Chu, C. C. (1985). H∞ optimization and robust multivariable control. Ph.D. thesis,
University of Minnesota.

[30] Dahleh, M. and I. J. Diaz-Bobillo (1995). Control of Uncertain Systems: A Linear

Programming Approach, Prentice Hall, Englewood Cliffs, New Jersey.

[31] Daniel, R. W., B. Kouvaritakis, and H. Latchman (1986). “Principal direction

alignment: a geometric framework for the complete solution to the µ problem,”
IEE Proceedings, Vol. 133, Part D, No. 2, pp. 45-56.

[32] Davis, R. A. (1995). Model Validation for Robust Control, Ph.D. dissertation,
University of Cambridge, UK.

[33] Desoer, C. A., R. W. Liu, J. Murray, and R. Saeks (1980). “Feedback system
design: the fractional representation approach to analysis and synthesis,” IEEE
Trans. Automat. Contr., Vol. AC-25, No. 6, pp. 399-412.

[34] Desoer, C. A. and M. Vidyasagar (1975). Feedback Systems: Input-Output Prop-

erties, Academic Press, New York.

[35] Doyle, J. C. (1978). “Guaranteed Margins for LQG Regulators,” IEEE Trans.
Automat. Contr., Vol. AC-23, No. 4, pp. 756-757.

[36] Doyle, J.C. (1982). “Analysis of feedback systems with structured uncertainties,”
IEE Proceedings, Part D, Vol.133, pp. 45-56.

[37] Doyle, J. C. (1984). “Lecture notes in advances in multivariable control,”

ONR/Honeywell Workshop, Minneapolis.

[38] Doyle, J. C. (1985). “Structured uncertainty in control system design”, Proc.

IEEE Conf. Dec. Contr., Ft. Lauderdale.

[39] Doyle, J. C., B. Francis, and A. Tannenbaum (1992). Feedback Control Theory,
Macmillan Publishing Company.

[40] Doyle, J. C., K. Glover, P.P . Khargonekar, B. A. Francis (1989). “State-space so-
lutions to standard H2 and H∞ control problems,” IEEE Trans. Automat. Contr.,
Vol. AC-34, No. 8, pp. 831-847. Also see 1988 American Control Conference, At-
lanta, June 1988.
394 BIBLIOGRAPHY

[41] Doyle, J., K. Lenz, and A. Packard (1986). “Design examples using µ synthesis:
space shuttle lateral axis FCS during reentry,” IEEE Conf. Dec. Contr., December
1986, pp. 2218-2223.

[42] Doyle, J. C., A. Packard and K. Zhou (1991). “Review of LFTs, LMIs and µ,”
Proc. IEEE Conf. Dec. Contr., England, pp. 1227-1232

[43] Doyle, J. C., R. S. Smith, and D. F. Enns (1987). “Control of plants with input
saturation nonlinearities,” American Control Conference, pp. 1034-1039.

[44] Doyle, J. C. and G. Stein (1981). “Multivariable feedback design: Concepts for
a classical/modern synthesis,” IEEE Trans. Automat. Contr., Vol. AC-26, pp.
4-16, Feb. 1981.

[45] Doyle, J. C., J. Wall and G. Stein (1982). “Performance and robustness analysis
for structured uncertainty,” in Proc. IEEE Conf. Dec. Contr., pp. 629-636.

[46] Doyle, J. C., K. Zhou, K. Glover, and B. Bodenheimer (1994). “Mixed H2 and
H∞ performance objectives II: optimal control,” IEEE Trans. Automat. Contr.,
Vol. 39, No. 8, pp. 1575-1587.

[47] El-Sakkary, A. (1985). “The gap metric: robustness of stabilization of feedback

systems,” IEEE Trans. Automat. Contr., Vol. 30, pp. 240-247.

[48] Enns, D. (1984a). Model Reduction for Control System Design, Ph.D. dissertation,
Department of Aeronautics and Astronautics, Stanford University, Stanford, Cal-
ifornia.

[49] Enns, D. (1984b). “Model reduction with balanced realizations: an error bound
and a frequency weighted generalization,” Proc. Conf. Dec. Contr., Las Vegas,
Nevada.

[50] Fan, M. K. H. and A. L. Tits (1986). “Characterization and efficient computation

of the structured singular value,” IEEE Trans. Automat. Contr., Vol. AC-31, No.
8, pp. 734-743.

[51] Fan, M. K. H., A. L. Tits, and J. C. Doyle (1991). “Robustness in the presence of
mixed parametric uncertainty and unmodeled dynamics,” IEEE Trans. Automat.
Contr., Vol. AC-36, No. 1, pp. 25-38.

[52] Foias, C. and A. E. Frazho (1990). The commutant Lifting Approach to Interpo-
lation Problems, Birkhäuser.

[53] Francis, B. A. (1987). A course in H∞ control theory, Lecture Notes in Control

and Information Sciences, Vol. 88, Springer-Verlag.

[54] Francis, B. A. and J. C. Doyle (1987). “Linear control theory with an H∞ opti-
mality criterion,” SIAM J. Contr. Optimiz., Vol. 25, pp. 815-844.
BIBLIOGRAPHY 395

[55] Franklin, G. F., J. D. Powell, and M. L. Workman (1990), Digital Control of

Dynamic Systems, 2nd Ed, Addison-Wesley Publishing company.

[56] Freudenberg, J. S. (1986). “The general structured singular problem with two
blocks of uncertainty.” Proc. 1986 Allerton Conf., Monticello, Illinois.

[57] Freudenberg, J. S. and D. P. Looze (1988). Frequency Domain Properties of Scalar

and Multivariable Feedback Systems, Lecture Notes in Contr. and Info. Science,
Vol. 104, Springer-Verlag, Berlin.

[58] Gahinet, P. (1996). “Explicit controller formulas for LMI-based H∞ synthesis,”

Automatica, Vol. 32, No. 7, pp. 1007-1014.

[59] Gahinet, P. and P. Apkarian (1994). “A linear matrix inequality approach to H∞

control,” Int. J. Robust and Nonlinear Control, Vol. 4, pp. 421-448.

[60] Garnett, J. B. (1981). Bounded Analytic Functions, Academic Press.

[61] Georgiou, T. T. (1988). “On the computation of the gap metric,” Systems and
Control Letters, Vol. 11, pp. 253-257.

[62] Georgiou, T. T. and M. C. Smith (1990). “Optimal robustness in the gap metric,”
IEEE Trans. Automat. Contr., Vol. AC-35, No. 6, pp. 673-686.

[63] Georgiou, T. T. and M. C. Smith (1992). “Robust stabilization in the gap metric:
controller design for distributed plants,” IEEE Trans. Automat. Contr., Vol. AC-
37, No. 8, pp. 1133-1143.

[64] Glover, K. (1984). “All optimal Hankel-norm approximations of linear multivari-

able systems and their L∞ -error bounds,” Int. J. Contr., Vol. 39, pp. 1115-1193,
1984.

[65] Glover, K. (1986a). “Robust stabilization of linear multivariable systems: Rela-

tions to approximation,” Int. J. Contr., Vol. 43, No. 3, pp. 741-766.

[66] Glover, K. (1986b). “Multiplicative approximation of linear multivariable systems

with L∞ error bounds, ” Proc. Amer. Contr. Conf., Seattle, WA, pp. 1705-1709.

[67] Glover, K. and J. Doyle (1988). “State-space formulae for all stabilizing controllers
that satisfy an H∞ norm bound and relations to risk sensitivity,” Systems and
Control Letters, Vol. 11, pp. 167-172.

[68] Glover, K. and J. C. Doyle (1989). “A state space approach to H∞ optimal con-
trol,” in Three Decades of Mathematical Systems Theory: A Collection of Surveys
at the Occasion of the 50th Birthday of Jan C. Willems, H. Nijmeijer and J. M.
Schumacher (Eds.), Springer-Verlag, Lecture Notes in Control and Information
Sciences, Vol. 135, 1989.
396 BIBLIOGRAPHY

[69] Glover, K., D. J. N. Limebeer, J. C. Doyle, E. M. Kasenally, and M. G. Safonov

(1991). “A characterization of all solutions to the four block general distance
problem,” SIAM J. Contr. Optimiz., Vol. 29, No. 2, pp. 283-324.
[70] Glover, K., D. J. N. Limebeer, and Y. S. Hung (1992). “A structured approxima-
tion problem with applications to frequency weighted model reduction, ” IEEE
Trans. Automat. Contr., Vol. AC-37, No. 4, pp. 447-465.
[71] Glover, K. and D. McFarlane (1989). “Robust stabilization of normalized coprime
factor plant descriptions with H∞ bounded uncertainty,” IEEE Trans. Automat.
Contr., Vol. 34, No. 8, pp. 821-830.
[72] Glover, K. and D. Mustafa (1989). “Derivation of the maximum entropy H∞
controller and a state space formula for its entropy,” Int. J. Contr., Vol. 50, pp.
899.
[73] Goddard, P. J. and K. Glover (1993). “Controller reduction: weights for stability
and performance preservation,” Proceedings of the 32nd Conference on Decision
and Control, San Antonio, Texas, December 1993, pp. 2903-2908.
[74] Goddard, P. J. and K. Glover (1994). “Performance preserving frequency weighted
controller approximation: a coprime factorization approach,” Proceedings of the
33nd Conference on Decision and Control, Orlando, Florida, December 1994.
[75] Gohberg, I., and S. Goldberg (1981). Basic Operator Theory. Birkhauser, Boston.
[76] Gohberg, I., P. Lancaster, and L. Rodman (1986). “On Hermitian solutions of the
symmetric algebraic Riccati equation,” SIAM J. Contr. Optimiz., Vol. 24, No. 6,
November 1986, pp. 1323-1334.
[77] Golub, G. H., and C. F. Van Loan (1983). Matrix Computations, Johns Hopkind
Univ. Press, Baltimore.
[78] Green, M. (1988). “A relative error bound for balanced stochastic truncation,”
IEEE Trans. Automat. Contr., Vol. AC-33, No. 10, pp. 961-965.
[79] Green, M. (1992). “H∞ controller synthesis by J-lossless coprime factorization,”
SIAM J. Contr. Optimiz., Vol. 30, No. 3, pp. 522-547.
[80] Green, M., K. Glover, D. J. N. Limebeer, and J. C. Doyle (1988), “A J-spectral
factorization approach to H∞ control,” SIAM J. Contr. Optimiz., Vol. 28, pp.
1350-1371.
[81] Green, M. and D. J. N. Limebeer (1995). Linear Robust Control, Prentice Hall,
Englewood Cliffs, New Jersey.
[82] Hinrichsen, D., and A. J. Pritchard (1990). “An improved error estimate for
reduced-order models of discrete time system,” IEEE Trans. Automat. Contr.,
AC-35, pp. 317-320.
BIBLIOGRAPHY 397

[83] Hoffman, K. (1962). Banach Spaces of Analytic Functions, Prentice Hall, Engle-
wood Cliffs, New Jersey.

[84] Horn, R. A. and C. R. Johnson (1990). Matrix Analysis, Cambridge University

Press, first published in 1985.

[85] Horn, R. A. and C. R. Johnson (1991). Topics in Matrix Analysis, Cambridge

University Press.

[86] Horowitz, I. M. (1963). Synthesis of Feedback Systems, Academic Press, London.

[87] Hung, Y. S. (1989).“H∞ -optimal control–Part I: model matching, – Part II: solu-
tion for controllers,” Int. J. Contr., Vol. 49, pp. 1291-1359.

[88] Hung, Y. S. and K. Glover (1986). “Optimal Hankel-norm approximation of stable

systems with first order stable weighting functions,” Systems and Control Letters,
Vol. 7, pp. 165-172.

[89] Hyde, R. A. and K. Glover (1993). “The application of scheduled H∞ controllers

to a VSTOL aircraft,” IEEE Trans. Automat. Contr., Vol. 38, No. 7, pp. 1021-
1039.

[90] Jonckheere, E. and L. M. Silverman (1983). “A new set of invariants for linear
systems - applications to reduced order compensator design,” IEEE Transactions
on Automatic Control, Vol 28, No. 10, pp. 953-964.

[91] Kailath, T. (1980). Linear Systems, Prentice Hall, Englewood Cliffs, New Jersey.

[92] Kalman, R. E. (1964). “When is a control system optimal?” ASME Trans. Series
D: J. Basic Engineering, Vol. 86, pp. 1-10.

[93] Kalman, R. E. and R. S. Bucy (1960). “New results in linear filtering and predic-
tion theory,” ASME Trans. Series D: J. Basic Engineering, Vol. 83, pp. 95-108.

[94] Khargonekar, P. P., I. R. Petersen, and M. A. Rotea (1988). “H∞ - optimal control
with state-feedback,” IEEE Trans. Automat. Contr., Vol. AC-33, pp. 786-788.

[95] Khargonekar, P. P., I. R. Petersen, and K. Zhou (1990). “Robust stabilization and
H∞ -optimal control,” IEEE Trans. Automat. Contr., Vol. 35, No. 3, pp. 356-361.

[96] Khargonekar, P. P. and E. Sontag (1982). “On the relation between stable matrix
fraction factorizations and regulable realizations of linear systems over rings,”
IEEE Trans. Automat. Contr., Vol. 27, pp. 627-638.

[97] Khargonekar, P. P. and A. Tannenbaum (1985). “Noneuclidean metrics and the ro-
bust stabilization of systems with parameter uncertainty,” IEEE Trans. Automat.
Contr., Vol. AC-30, pp. 1005-1013.
398 BIBLIOGRAPHY

[98] Kimura, H. (1997). Chain-Scattering Approach to H∞ -Control, Birkhäuser,

Boston.
[99] Kwakernaak, H. and R. Sivan (1972). Linear Optimal Control Systems, Wiley-
Interscience, New York.
[100] Kucera, V. (1972). “A contribution to matrix quadratic equations,” IEEE Trans.
Automat. Contr., AC-17, No. 3, 344-347.
[101] Lancaster, P. and L. Rodman (1995). Algebraic Riccati Equations. Oxford Uni-
versity Press, Oxford.
[102] Lancaster, P. and M. Tismenetsky (1985). The Theory of Matrices: with applica-
tions, 2nd ed., Academic Press.
[103] Latham, G. A. and B. D. O. Anderson (1986). “Frequency weighted optimal
Hankel-norm approximation of stable transfer function,” Systems and Control
Letters, Vol. 5, pp. 229-236.
[104] Laub, S. J. (1980). “Computation of ’balancing’ transformations,” Proc. 1980
JACC, Session FA8-E.
[105] Lenz, K. E., P. P. Khargonekar, and John C. Doyle (1987). “Controller order
reduction with guaranteed stability and performance, ” American Control Con-
ference, pp. 1697-1698.
[106] Limebeer, D. J. N., B. D. O. Anderson, P. P. Khargonekar, and M. Green (1992).
“A game theoretic approach to H∞ control for time varying systems,” SIAM J.
Contr. Optimiz., Vol. 30, No. 2, pp. 262-283.
[107] Liu, K. Z., T. Mita, and R. Kawtani (1990). “Parameterization of state feedback
H∞ controllers,” Int. J. Contr., Vol. 51, No. 3, pp. 535-551.
[108] Liu, Y. and B. D. O. Anderson (1986). “Controller reduction via stable factoriza-
tion and balancing,” Int. J. Contr., Vol. 44, pp. 507-531.
[109] Liu, Y. and B. D. O. Anderson (1989). “Singular perturbation approximation of
balanced systems,” Int. J. Contr., Vol. 50, No. 4, pp. 1379-1405.
[110] Liu, Y. and B. D. O. Anderson (1990). “Frequency weighted controller reduction
methods and loop transfer recovery,” Automatica, Vol. 26, No. 3, pp. 487-497.
[111] Liu, Y., B. D. O. Anderson, and U. Ly (1990). “Coprime factorization controller
reduction with Bezout identity induced frequency weighting,” Automatica, Vol.
26, No. 2, pp. 233-249.
[112] Lu, W. M., K. Zhou, and J. C. Doyle (1996). “Stabilization of uncertain linear sys-
tems and linear multidimensional systems,” IEEE Trans. on Automatic Control,
vol. 41, No. 1, pp. 50-65.
BIBLIOGRAPHY 399

[113] Luenberger, D. G. (1971). “An introduction to observers,” IEEE Trans. Automat.

Contr., Vol. 16, No. 6, pp. 596-602.

[114] Maciejowski, J. M. (1989). Multivariable Feedback Design, Addison-Wesley, Read-

ing, Massachusetts.

[115] Mageirou, E. F. and Y. C. Ho (1977). “Decentralized stabilization via game the-

oretic methods,” Automatica, Vol. 13, pp. 393-399.

[116] Martensson, K. (1971). “On the matrix Riccati equation,” Information Sciences,
Vol. 3, pp. 17-49.

[117] McFarlane, D. C., K. Glover, and M. Vidyasagar (1990). “Reduced-order con-

troller design using coprime factor model reduction,” IEEE Trans. Automat.
Contr., Vol. 35, No. 3, pp. 369-373.

[118] McFarlane, D. C. and K. Glover (1990). Robust Controller Design Using Normal-
ized Coprime Factor Plant Descriptions, Vol. 138, Lecture Notes in Control and
Information Sciences, Springer-Verlag.

[119] McFarlane, D. C. and K. Glover (1992). “A loop shaping design procedure using
H∞ synthesis,” IEEE Trans. Automat. Contr., Vol. 37, No. 6, pp. 759-769.

[120] Meyer, D. G. (1990). “Fractional balanced reduction: model reduction via frac-
tional representation,” IEEE Trans. Automat. Contr., Vol. 35, No. 12, pp. 1341-
1345.

[121] Meyer, D. G. and G. Franklin (1987). “A connection between normalized co-

prime factorizations and linear quadratic regulator theory,” IEEE Trans. Au-
tomat. Contr., Vol. 32, pp. 227-228.

[122] Molinari, B. P. (1975). “The stabilizing solution of the discrete algebraic Riccati
equation,” IEEE Trans. Automat. Contr., June 1975, pp. 396-399.

[123] Moore, B. C. (1981). “Principal component analysis in linear systems: control-

lability, observability, and model reduction,” IEEE Trans. Automat. Contr., Vol.
26, No. 2, pp. 17-32.

[124] Moore, J. B., K. Glover, and A. Telford (1990). “All stabilizing controllers as
frequency shaped state estimate feed back,” IEEE Trans. Auto. Contr., Vol. AC-
35, pp. 203-208.

[125] Morari, M. and E. Zafiriou (1989). Robust Process Control. Prentice Hall, Engle-
wood Cliffs, New Jersey.

[126] Mullis, C. T. and R. A. Roberts (1976). “Synthesis of minimum roundoff noise

fixed point digital filters,” IEEE Trans. Circuits Syst., Vol. 23, pp. 551-562.
400 BIBLIOGRAPHY

[127] Mustafa, D. (1989). “Relations between maximum entropy / H∞ control and

combined H∞ /LQG control,” Systems and Control Letters, Vol. 12, No. 3, p.
193.

[128] Mustafa, D. and K. Glover (1990). Minimum Entropy H∞ Control, Lecture Notes
in Control and Information Sciences, Springer-Verlag.

[129] Mustafa, D. and K. Glover (1991). “Controller reduction by H∞ -balanced trun-

cation,” IEEE Trans. Automat. Contr., Vol. AC-36, No. 6, pp. 669-682.

[130] Naylor, A. W., and G. R. Sell (1982). Linear Operator Theory in Engineering and
Science, Springer-Verlag.

[131] Nagpal, K. M. and P. P. Khargonekar (1991). “Filtering and smoothing in an H∞

setting,” IEEE Trans. Automat. Contr., Vol. AC-36, No. 2, pp. 152-166.

[132] Nehari, Z. (1957). “On bounded bilinear forms,” Annals of Mathematics, Vol. 15,
No. 1, pp. 153-162.

[133] Nett, C. N., C. A. Jacobson and N. J. Balas (1984). “A connection between state-
space and doubly coprime fractional representations,” IEEE Trans. Automat.
Contr., Vol. AC-29, pp. 831-832.

[134] Nett, C. N. and J. A. Uthgenannt (1988). “An explicit formula and an optimal
weight for the 2-block structured singular value interaction measure,” Automatica,
Vol. 24, No. 2, pp. 261-265.

[135] Packard, A. (1991). Notes on µ∆ . Unpublished lecture notes.

[136] Packard, A. (1994). “Gain scheduling via linear fractional transformations.” Sys-
tems and Control Letters, Vol. 22, pp. 79-92.

[137] Packard, A. and J. C. Doyle (1988a). “Structured singular value with repeated
scalar blocks,” Proc. American Control Conference, Atlanta.

[138] Packard, A. and J. C. Doyle (1988b). Robust Control of Multivariable and Large
Scale Systems, Final Technical Report for Air Force Office of Scientific Research.

[139] Packard, A. and J. C. Doyle (1993), “The complex structured singular value,”
Automatica, Vol. 29, pp. 71-109.

[140] Packard, A. and P. Pandey (1993). “Continuity properties of the real/complex

structured singular value,” IEEE Trans. Automat. Contr., Vol. 38, No. 3, pp.
415-428.

[141] Packard, A., K. Zhou, P. Pandey, J. Leonhardson, and G. Balas (1992). “Optimal,
constant I/O similarity scaling for full information and state feedback control
problems,” Systems and Control Letters, Vol. 19, No. 4, pp. 271-280.
BIBLIOGRAPHY 401

[142] Paganini, F. (1995). “Necessary and sufficient conditions for robust H2 perfor-
mance,” Proc. IEEE Conf. dec. Contr., New Orleans, Louisiana, pp. 1970-1975.

[143] Paganini, F. (1996). “A set-based approach for white noise modeling,” IEEE
Trans. Automat. Contr., Vol. AC-41, No. 10, pp. 1453-1465.

[144] Pernebo, L. and L. M. Silverman (1982). ”Model reduction via balanced state
space Representation,” IEEE Trans. Automat. Contr., Vol. AC-27, No. 2, pp.
382-387.

[145] Petersen, I. R. (1987). “Disturbance attenuation and H∞ optimization: a design

method based on the algebraic Riccati equation,” IEEE Trans. Automat. Contr.,
Vol. AC-32, pp. 427-429.

[146] Poolla, K., P. P. Khargonekar, A. Tikku, J. Krause and K. Nagpal (1994). “A

time domain approach to model validation,” IEEE Trans. Automat. Contr., Vol.
39, No. 5, pp. 951-959.

[147] Poolla, K. and A. Tikku (1995). “Robust performance against time-varying struc-
tured perturbations,” IEEE Trans. Automat. Contr., Vol. 40, No. 9, pp. 1589-
1602.

[148] Qiu, L. and E. J. Davison (1992a). “Feedback stability under simultaneous gap
metric uncertainties in plant and controller,” Systems and Control Letters, Vol.
18, pp. 9-22.

[149] Qiu, L. and E. J. Davison (1992b). “Pointwise gap metrics on transfer matrices,”
IEEE Trans. Automat. Contr., Vol. 37, No. 6, pp. 741-758.

[150] Qiu, L. (1995). “On the robustness of symmetric systems,” Proc. IEEE Conf. Dec.
Contr., New Orleans, Louisiana, pp. 2659-2660.

[151] Qiu, L., B. Bernhardsson, A. Rantzer, E. J. Davison, P. M. Young, and J. C. Doyle

(1995). “A formula for computation of the real stability Radius,” Automatica, Vol.
31, No. 6, pp. 879-890.

[152] Ran, A. C. M. and R. Vreugdenhil (1988). “Existence and comparison theorems

for algebraic Riccati equations for continuous- and discrete-time systems,” Linear
Algebra and its Applications, Vol. 99, pp. 63-83.

[153] Redheffer, R. M. (1959). “Inequalities for a matrix Riccati equation,” Journal of

Mathematics and Mechanics, Vol. 8, No. 3.

[154] Redheffer, R. M. (1960). “On a certain linear fractional transformation,” J. Math.

and Physics, Vol. 39, pp. 269-286.

[155] Rangan, S. and K. Poolla (1996). “Time domain validation for sampled-data un-
certainty models,” IEEE Trans. Automat. Contr., Vol. 41, No. 7, pp. 980-991.
402 BIBLIOGRAPHY

[156] Rosenbrock, H. H. (1974). Computer-Aided Control System Design. Academic

Press, London.
[157] Safonov, M. G. (1980). Stability and Robustness of Multivariable Feedback Systems,
MIT Press.
[158] Safonov, M. G. (1982). “Stability margins of diagonally perturbed multivariable
feedback systems,” Proc. IEE, Part D, Vol. 129, No. 6, pp. 251-256.
[159] Safonov, M. G. (1984). “Stability of interconnected systems having slope-bounded
nonlinearities,”6th Int. Conf. on Analysis and Optimization of Systems, Nice,
France.
[160] Safonov, M. G. and J. C. Doyle (1984). “Minimizing conservativeness of robustness
singular values,” in Multivariable Control, S.G. Tzafestas, Ed., Reidel, New York.
[161] Safonov, M. G. and D. J. N. Limebeer, and R. Y. Chiang (1990). “Simplifying
the H∞ Theory via Loop Shifting, matrix pencil and descriptor concepts,” Int. J.
Contr., Vol 50, pp. 2467-2488.
[162] Sampei, M., T. Mita, and M. Nakamichi (1990). “An algebraic approach to H∞
output feedback control problems,” Systems and Control Letters, Vol. 14, pp.
13-24.
[163] Scherer, C. (1992). “H∞ control for plants with zeros on the imaginary axis,”
SIAM J. Contr. Optimiz., Vol. 30, No. 1, pp. 123-142.
[164] Scherer, C. (1992). “H∞ optimization without assumptions on finite or infinite
zeros,” SIAM J. Contr. Optimiz., Vol. 30, No. 1, pp. 143-166.
[165] Seron, M. M., J. H. Braslavsky, and G. C. Goodwin (1997). Fundamental Limi-
tations in Filtering and Control, Springer, Berlin.
[166] Shamma, J. (1994). “Robust stability with time-varying structured uncertainty,”
IEEE Trans. Automat. Contr., Vol. 39, No. 4, pp. 714-724.
[167] Skogestad, S. and I. Postlethwaite (1996). Multivariable Feedback Control, John
Wiley & Sons, New York.
[168] Smith, R. S. and J. C. Doyle (1992). “Model validation: a connection between
robust control and identification,” IEEE Trans. Automat. Contr., Vol. 37, No. 7,
pp. 942-952.
[169] Smith, R. S. and G. Dullerud (1996). “Continuous-time control model validation
using finite experimental data,” IEEE Trans. Automat. Contr., Vol. 41, No. 8, pp.
1094-1105.
[170] Stein, G. and J. C. Doyle (1991). “Beyond singular values and loop shapes,” AIAA
Journal of Guidance and Control, Vol. 14, No. 1, pp. 5-16.
BIBLIOGRAPHY 403

[171] Stoorvogel, A. A. (1992). The H∞ Control Problem: A State Space Approach,

Prentice Hall, Englewood Cliffs, New Jersey.
[172] Tadmor, G. (1990). “Worst-case design in the time domain: the maximum princi-
ple and the standard H∞ problem,” Mathematics of Control, Systems and Signal
Processing, pp. 301-324.
[173] Tits, A. L. (1995). “On the small mu theorem,” Proc. American Control Confer-
ence, Seattle, Washington, pp. 2870-2872.
[174] Tits, A. L. and M. K. H. Fan (1995). “On the small µ theorem,” Automatica, Vol.
31, No. 8, pp. 1199-1201.
[175] Toker, O. and H. Özbay (1995). “On the NP-hardness of the purely complex µ
computation, analysis/synthesis, and some related problems in multidimensional
systems,” Proc. American Control Conference, Seattle, Washington, pp. 447-451.
[176] Ushida, S. and H. Kimura (1996). “A counterexample to Mustafa-Glovers’ mono-
tonicity conjecture,” Systems and Control Letters, Vol. 28, No. 3, pp. 129-137.
[177] Van Dooren, P. (1981). “A generalized eigenvalue approach for solving Riccati
equations,” SIAM J. Sci. Stat. Comput., Vol. 2, No. 2, pp. 121-135.
[178] Vidyasagar, M. (1984). “The graph metric for unstable plants and robustness
estimates for feedback stability,” IEEE Trans. Automat. Contr., Vol. 29, pp. 403-
417.
[179] Vidyasagar, M. (1985). Control System Synthesis: A Factorization Approach. MIT
Press, Cambridge, Massachusetts.
[180] Vidyasagar, M. (1988). “Normalized coprime factorizations for non-strictly proper
systems,” IEEE Trans. Automat. Contr., Vol. 33, No. 3, pp. 300-301.
[181] Vidyasagar, M. and H. Kimura (1986). “Robust controllers for uncertain linear
multivariable systems,” Automatica, Vol. 22, pp. 85-94.
[182] Vinnicombe, G. (1992a). “On the frequency response interpretation on an indexed
L2 -gap metric,” Proceedings of the American Control Conference, Chicago, Illi-
nois, pp. 1133-1137.
[183] Vinnicombe, G. (1992b). “Robust Design in the graph topology: a benchmark
example,” Proceedings of the American Control Conference, Chicago, Illinois, pp.
2063-2064.
[184] Vinnicombe, G. (1993a). “Frequency domain uncertainty and the graph topology,”
IEEE Trans. Automat. Contr., Vol. 38, No. 9, pp. 1371-1383.
[185] Vinnicombe, G. (1993b). Measuring the Robustness of Feedback Systems, Ph.D.
dissertation, Cambridge University.
404 BIBLIOGRAPHY

[186] Whittle, P. (1990). Risk-sensitive Optimal Control, John Wiley and Sons, New
York.
[187] Willems, J.C. (1971). “Least Squares Stationary Optimal Control and the Alge-
braic Riccati Equation”, IEEE Trans. Automat. Contr., Vol. AC-16, No. 6, pp.
621-634.
[188] Willems, J. C. (1981). “Almost invariant subspaces: an approach to high gain
feedback design – Part I: almost controlled invariant subspaces,” IEEE Trans.
Automat. Contr., Vol. AC-26, pp. 235-252.
[189] Willems, J. C., A. Kitapci and L. M. Silverman (1986). “Singular optimal control:
a geometric approach,” SIAM J. Contr. Optimiz., Vol. 24, pp. 323-337.
[190] Wimmer, H. K. (1985). “Monotonicity of Maximal solutions of algebraic Riccati
equations,” Systems and Control Letters, Vol. 5, April 1985, pp. 317-319.
[191] Wonham, W. M. (1985). Linear Multivariable Control: A Geometric Approach,
third edition, Springer-Verlag, New York.
[192] Yang, X. H. and A. Packard (1995). “A low order controller design method,” Proc.
IEEE Conf. Dec. Contr., New Orleans, Louisiana, pp. 3068-3073.
[193] Youla, D. C. and J. J. Bongiorno, Jr. (1985). “A feedback theory of two-degree-of-
freedom optimal Wiener-Hopf design,” IEEE Trans. Automat. Contr., Vol. AC-30,
No. 7, pp. 652-665.
[194] Youla, D. C., H. A. Jabr, and J. J. Bongiorno (1976a). “Modern Wiener-Hopf
design of optimal controllers: part I,” IEEE Trans. Automat. Contr., Vol. AC-21,
pp. 3-13.
[195] Youla, D. C., H. A. Jabr, and J. J. Bongiorno (1976b). “Modern Wiener-Hopf
design of optimal controllers: part II,” IEEE Trans. Automat. Contr., Vol. AC-
21, pp. 319-338.
[196] Youla, D. C., H. A. Jabr, and C. N. Lu (1974). “Single-loop feedback stabilization
of linear multivariable dynamical plants,” Automatica, Vol. 10, pp. 159-173.
[197] Youla, D. C. and M. Saito (1967). “Interpolation with positive-real functions,”
Journal of The Franklin Institute, Vol. 284, No. 2, pp. 77-108.
[198] Young, N. (1988). An Introduction to Hilbert Space, Cambridge University Press.
[199] Young, P. M. (1993). Robustness with Parametric and Dynamic Uncertainty,
Ph.D. thesis, California Institute of Technology.
[200] Zames, G. (1966). “On the input-output stability of nonlinear time-varying feed-
back systems, parts I and II,” IEEE Trans. Automat. Contr., Vol. AC-11, pp. 228
and 465.
BIBLIOGRAPHY 405

[201] Zames, G. (1981). “Feedback and optimal sensitivity: model reference transfor-
mations, multiplicative seminorms, and approximate inverses,” IEEE Trans. Au-
tomat. Contr., Vol. AC-26, pp. 301-320.
[202] Zames, G. and A. K. El-Sakkary (1980). “Unstable systems and feedback: The
gap metric,” Proc. Allerton Conf., pp. 380-385.
[203] Zhang, C. and M. Fu (1996). “A revisit to the gain and phase margins of linear
quadratic regulators,” IEEE Trans. Automat. Contr., Vol. 41, No. 10, pp. 1527-
1530.
[204] Zhou, K. (1992). “On the parameterization of H∞ controllers,” IEEE Trans. Au-
tomat. Contr., Vol. 37, No. 9, pp. 1442-1445.
[205] Zhou, K. (1995). “Frequency weighted L∞ norm and optimal Hankel norm model
reduction,” IEEE Trans. Automat. Contr., vol. 40, no. 10, pp. 1687-1699, October
1995.
[206] Zhou K. and J. Chen (1995). “Performance bounds for coprime factor controller
reductions,” Systems and Control Letters, Vol. 26, No. 2, pp. 119-127.
[207] Zhou, K., J. C. Doyle and K. Glover (1996). Robust and Optimal Control. Prentice
Hall, Upper Saddle River, New Jersey.
[208] Zhou, K. and P. P. Khargonekar (1988). “An algebraic Riccati equation approach
to H∞ optimization,” Systems and Control Letters, Vol. 11, pp. 85-92.
[209] Zhou, K., K. Glover, B. Bodenheimer, and J. Doyle (1994). “Mixed H2 and H∞
performance objectives I: robust performance analysis,” IEEE Trans. Automat.
Contr., Vol. 39, No. 8, pp. 1564-1574.
[210] Zhou, T. and H. Kimura (1992). “Minimal H∞ -norm of transfer functions con-
sistent with prescribed finite input-output data,” in Proc. SICE’ 92, Kumamotp,
Japan, pp. 1079-1082.
[211] Zhu, S. Q. (1989). “Graph topology and gap topology for unstable systems,” IEEE
Trans. Automat. Contr., Vol. 34, No. 8, pp. 848-855.
406 BIBLIOGRAPHY
Index

additive approximation, 105 bounded real lemma, 7, 238

additive uncertainty, 131, 142 Boyd, S., 50, 62, 102, 299
admissible controller, 222 Braatz, R. D., 382
algebraic Riccati equation, 6, 233 Braslavsky, J. H., 102
complementarity property, 234 Brogan, W., 24, 41
solutions, 15 Bruinsma, N. A., 62
stability property, 234
stabilizing solution, 234
all-pass dilation, 120 Cauchy-Schwarz inequality, 46
all-pass function, 245 Chen, C. T., 41
Al-Saggaf, U. M., 126 Chen, J., 102, 344, 381, 389
analytic function, 47 Chen. T., 54
Anderson, B. D. O., 312 Chen, X., 299
Argument Principle, 357 Chilali, M., 299
Arnold, W. F., 246 Christian, L., 341
co-inner function, 245
complementary inner function, 246
Balakrishnan, V., 62, 299 complementary sensitivity, 82
balanced model reduction, 99 conjugate system, 34
additive, 105 controllability, 3, 27
error bound, 119 Gramian, 3, 53, 106
frequency-weighted, 124 matrix, 29
multiplicative, 125 controllable canonical form, 36
relative, 125 controller parameterization, 221
stability, 117 controller reduction, 8, 305
balanced realization, 4, 37, 105, 107 coprime factor uncertainty, 144, 315
Balas, G., 216 coprime factorization, 4, 71, 228, 315
Balas, N. J, 77 normalized, 318
basis, 11
Bezout identity, 71
bisection algorithm, 57 Davis, R. A., 381, 389
Bode, H. W., 102 Davison, E. J., 341, 354, 375
Bode integral, 81 design limitation, 4, 81
Bode’s gain and phase relation, 94 design model, 129
Bongiorno, J. J., 231 design tradeoff, 81

407
408 INDEX

Desoer, C. A., 50, 62, 102, 139, 158, 231 Georgiou, T. T., 334, 341, 350, 375
detectability, 27, 31 Gilbert’s realization, 37
D-G-K iteration, 388 Glover, K., 3, 126, 246, 299, 309, 312,
D-K iteration, 214 325, 341, 344
directed gap, 350 Goddard, P. J., 309, 312
dom(Ric), 234 Gohberg, I., 62
double coprime factorization, 71 Goldberg, I., 62
Doyle, J. C., 3, 77, 102, 152, 180, 192, Golub, G. H., 24
194, 206, 216, 231, 246, 267, Goodwin, G. C., 102
299, 312, 383, 389 Gramian, 106, 110
dual system, 34 graph metric, 341
Dullerud, G., 381, 389 graph topology, 341
Green, M., 126
eigenvalue, 12
eigenvector, 12, 234 Hamiltonian matrix, 56, 233
El Ghaoui, L., 299 Hankel singular value, 110
El-Sakkary, A., 341, 349, 375 Hardy spaces, 48
Enns, D., 126, 312 Hermitian matrix, 13
entropy, 286 H∞ control, 2, 7, 269
error bound, 4 loop shaping, 315
singular problem, 294
F` , 165 H∞ filtering, 8, 297
Fu , 166 H∞ norm, 2, 55
Fan, M. K. H., 201, 217, 383, 389 H∞ optimal controller, 282, 286
Feron, E., 299 H∞ performance, 85, 88
feedback, 65 H∞ space, 45, 47, 50
−
filtering, 297 H∞ space, 50
Foias, C., 378 Hilbert space, 45
Fourier transform, 49 Hinrichsen, D., 126
Francis, B. A., 54, 77, 102, 231, 299 Horn, R. A., 24
Franklin, G. F., 126 Horowitz, I. M., 102, 341
Frazho, A. E., 378 H2 norm, 53
frequency weighting, 85 H2 optimal control, 253
frequency-weighted balanced reduction, H2 performance, 85, 87
8, 124 H2 space, 45, 47
Freudenberg, J. S., 102, 341, 348 H2 stability margin, 265
Frobenius norm, 18 H2⊥ space, 48
Hung, Y. S., 126
Hurwitz, 29
Gahinet, P., 299 Hyde, R. A., 341
gain, 94
gap metric, 9, 154, 341, 349
generalized eigenvector, 15, 234 image, 12
INDEX 409

induced norm, 17 LQG stability margin, 265

inner function, 245 LQG/LTR, 102
inner product, 46 LQR porblem, 255
inner-outer factorization, 248 LQR stability margin, 259
input sensitivity, 82 L2 norm, 53
integral control, 8, 294 L2 space, 48
internal model principle, 294 L2 (−∞, ∞) space, 47
internal stability, 4, 68 Lu, W. M., 231
invariant subspace, 15 Ly, U., 312
invariant zero, 39, 242 Lyapunov equation, 13, 53, 106
inverse of a transfer function, 35

Jacobson, V. A., 77 main loop theorem, 197

Johnson, C. R., 24 Martensson, K., 246
Jonckheere, E., 347 matrix
Hermitian, 13
inequality, 239
Kabamba, P., 62 inversion formulas, 13
Kailath, T., 41, 71 norm, 16
kernel, 12 square root of a, 23
Khargonekar, P. P., 299, 312, 379, 389 maximum modulus theorem, 47
Kimura, H., 302 McFarlane, D. C., 312, 325, 341, 344
Kitapci, A., 294 minimal realization, 35, 109
Krause, J., 379, 389 minimum entropy controller, 286
Kwakernaak, H., 267 Mita, T., 299
mixed µ, 9, 381
Lancaster, P., 24, 246 modal controllability, 31
Laub, A. J., 246 modal observability, 31
Lebesgue measure, 46 model invalidation, 377
left coprime factorization, 71 model reduction, 105
Lenz, K., 312 model uncertainty, 65, 129
L∞ norm, 55 model validation, 9, 377
L∞ space, 50 Moore, B. C., 126
Limebeer, D. J. N., 126 Moore, J. B., 231
linear combination, 11 µ, 5, 183
linear fractional transformation (LFT), lower bound, 192
2, 5, 163 synthesis, 213
linear matrix inequality (LMI), 239, 277 upper bound, 192
Liu, Y., 312 Mullis, C. T., 126
loop gain, 83 multiplication operator, 50
loop shaping, 9, 315, 325 multiplicative approximation, 125
loop transfer matrix, 82 multiplicative uncertainty, 131, 143
Looze, D. P., 102, 348 Mustafa, D., 312
410 INDEX

Nagpal, K. M., 299, 379, 389 Qiu, L., 220, 341, 354, 375
Nakamichi, M., 299 quadratic performance, 253
Naylor, A. W., 62
Nett, C. N., 77
nominal performance (NP), 137 Ran, A. C. M., 299
nominal stability (NS), 137 Rangan, S., 381, 389
nonminimum phase zero, 81 range, 12
norm, 16 real µ, 381
normal rank, 38 real spectral radius, 13
normalized coprime factorization, 8, 154 realization, 35
loop shaping, 325 balanced, 110
ν-gap metric, 9, 154, 349 input normal, 113
null space, 12 minimal, 35
output normal, 113
Redheffer star product, 178
observability, 3, 27 reduced-order controller, 8
Gramian, 3, 53, 106 regulator problem, 253
observable canonical form, 36 relative approximation, 125
observable mode, 31 return difference, 82
observer, 31 RH∞ space, 50
observer-based controller, 31 RH− ∞ space, 50
optimality of H∞ controller, 282 RH2 space, 48
orthogonal complement, 12 RH⊥ 2 space, 48
orthogonal matrix, 12 Riccati equation, 233
output sensitivity, 82 Riccati operator, 234
Özbay, H., 382 right coprime factorization, 71
Roberts, R. A., 126
robust performance, 137
Packard, A., 180, 192, 216, 299, 314, H2 performance, 147
382, 389 H∞ performance, 147, 197
Pandey, P., 217, 382, 389 structured, 202
Parseval’s relations, 49 robust stability (RS), 5, 137
PBH (Popov-Belevitch-Hautus) tests, 31 structured, 200
performance limitation, 81 robust stabilization, 315
Pernebo, L., 126 Rodman, L., 246
phase, 94
plant condition number, 150
Poisson integral, 81, 335, 348 Safonov, M. G., 132
pole, 38 Saito, M., 215
Poolla, K., 379, 381, 389 Sampei, M., 299
positive (semi-)definite matrix, 23 Schur complement, 14
positive real, 247 Sell, G. R., 62
Postlethwaite, I., 102 sensitivity function, 82
Pritchard, A. J., 126 Seron, M. M., 102
INDEX 411

Silverman, L. M., 126, 347 unstructured, 5, 129

singular value decomposition (SVD), 19, unitary matrix, 12
51 Ushida, S., 302
singular H∞ problem, 294
singular vector, 20
Sivan, R., 267 Van Dooren, P., 246
skewed performance specification, 150 Van Loan, C. F., 24
Skogestad, S., 102 Vidyasagar, M., 62, 77, 139, 158, 231,
small gain theorem, 3, 129, 137 341, 349
Smith, M. C., 334, 341, 375 Vinnicombe, G., 312, 341, 349, 366, 375
Smith, R. S., 381, 389 Vreugdenhil, R., 299
span, 11
spectral radius, 12 Wall, J., 158
stability, 27 Wang, S., 381, 389
internal, 68 weighted model reduction, 124
margin, 265 weighting function, 4, 85, 89
stabilizability, 27 well-posedness, 66, 167
stabilizable, 29 Willems, J. C., 246
stabilization, 221 winding number, 357
stabilizing controller, 6, 221 Wonham, W. M., 41
stable invariant subspace, 15, 234
star product, 178
Stein, G., 102, 152, 217, 267 Yang, X. H., 314
Steinbuch, M., 62 Youla, D. C., 215, 221, 231
strictly positive real, 247 Youla parameterization, 224, 228
Stoorvogel, A. A., 294, 299 Young, P. M., 217, 382, 389
structured singular value, 5, 183
lower bound, 192
Zames, G., 9, 132, 158, 349, 375
upper bound, 192
zero, 3, 38
structured uncertainty, 5, 183
Zhou, K., 3, 126, 180, 217, 246, 299, 344
Sylvester equation, 13
Zhu, S. Q., 341

Tannenbaum, A., 102

Tikku, A., 379, 389
Tismenetsky, M., 24
Tits, A. L., 201, 217, 383, 389
Toker, O., 382
trace, 12
tradeoff, 81

uncertainty, 1, 65, 129

state space, 171