This is page i

Printer: Opaque this

A Course in Robust
Control Theory
a convex approach

Geir E. Dullerud Fernando G. Paganini

University of Illinois University of California
Urbana-Champaign Los Angeles
 This is page i
Printer: Opaque this

Contents

0 Introduction 1
0.1 System representations . . . . . . . . . . . . . . . . . . . 2
0.1.1 Block diagrams . . . . . . . . . . . . . . . . . . . 2
0.1.2 Nonlinear equations and linear decompositions . . 4
0.2 Robust control problems and uncertainty . . . . . . . . . 9
0.2.1 Stabilization . . . . . . . . . . . . . . . . . . . . . 9
0.2.2 Disturbances and commands . . . . . . . . . . . . 12
0.2.3 Unmodeled dynamics . . . . . . . . . . . . . . . . 15
1 Preliminaries in Finite Dimensional Space 18
1.1 Linear spaces and mappings . . . . . . . . . . . . . . . . 18
1.1.1 Vector spaces . . . . . . . . . . . . . . . . . . . . 19
1.1.2 Subspaces . . . . . . . . . . . . . . . . . . . . . . 21
1.1.3 Bases, spans, and linear independence . . . . . . 22
1.1.4 Mappings and matrix representations . . . . . . 24
1.1.5 Change of basis and invariance . . . . . . . . . . 28
1.2 Subsets and Convexity . . . . . . . . . . . . . . . . . . . 30
1.2.1 Some basic topology . . . . . . . . . . . . . . . . 31
1.2.2 Convex sets . . . . . . . . . . . . . . . . . . . . . 32
1.3 Matrix Theory . . . . . . . . . . . . . . . . . . . . . . . . 38
1.3.1 Eigenvalues and Jordan form . . . . . . . . . . . 39
1.3.2 Self-adjoint, unitary and positive denite matrices 41
1.3.3 Singular value decomposition . . . . . . . . . . . 45
1.4 Linear Matrix Inequalities . . . . . . . . . . . . . . . . . 47
ii Contents
1.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2 State Space System Theory 57
2.1 The autonomous system . . . . . . . . . . . . . . . . . . 58
2.2 Controllability . . . . . . . . . . . . . . . . . . . . . . . . 61
2.2.1 Reachability . . . . . . . . . . . . . . . . . . . . . 61
2.2.2 Properties of controllability . . . . . . . . . . . . 66
2.2.3 Stabilizability and the PBH test . . . . . . . . . . 69
2.2.4 Controllability from a single input . . . . . . . . . 72
2.3 Eigenvalue assignment . . . . . . . . . . . . . . . . . . . 74
2.3.1 Single input case . . . . . . . . . . . . . . . . . . 74
2.3.2 Multi input case . . . . . . . . . . . . . . . . . . . 75
2.4 Observability . . . . . . . . . . . . . . . . . . . . . . . . 77
2.4.1 The unobservable subspace . . . . . . . . . . . . . 78
2.4.2 Observers . . . . . . . . . . . . . . . . . . . . . . 81
2.4.3 Observer-Based Controllers . . . . . . . . . . . . 83
2.5 Minimal realizations . . . . . . . . . . . . . . . . . . . . 84
2.6 Transfer functions and state space . . . . . . . . . . . . . 87
2.6.1 Real-rational matrices and state space realizations 89
2.6.2 Minimality . . . . . . . . . . . . . . . . . . . . . . 92
2.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3 Linear Analysis 97
3.1 Normed and inner product spaces . . . . . . . . . . . . . 98
3.1.1 Complete spaces . . . . . . . . . . . . . . . . . . 101
3.2 Operators . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.2.1 Banach algebras . . . . . . . . . . . . . . . . . . . 107
3.2.2 Some elements of spectral theory . . . . . . . . . 110
3.3 Frequency domain spaces: signals . . . . . . . . . . . . . 113
3.3.1 The space L^ 2 and the Fourier transform . . . . . 113
3.3.2 The spaces H2 and H2? and the Laplace transform 115
3.3.3 Summarizing the big picture . . . . . . . . . . . . 119
3.4 Frequency domain spaces: operators . . . . . . . . . . . . 120
3.4.1 Time invariance and multiplication operators . . 121
3.4.2 Causality with time invariance . . . . . . . . . . . 122
3.4.3 Causality and H1 . . . . . . . . . . . . . . . . . 124
3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4 Model realizations and reduction 131
4.1 Lyapunov equations and inequalities . . . . . . . . . . . 131
4.2 Observability operator and gramian . . . . . . . . . . . . 134
4.3 Controllability operator and gramian . . . . . . . . . . . 137
4.4 Balanced realizations . . . . . . . . . . . . . . . . . . . . 140
4.5 Hankel operators . . . . . . . . . . . . . . . . . . . . . . 143
4.6 Model reduction . . . . . . . . . . . . . . . . . . . . . . . 147
 Contents iii
4.6.1 Limitations . . . . . . . . . . . . . . . . . . . . . 148
4.6.2 Balanced truncation . . . . . . . . . . . . . . . . 151
4.6.3 Inner transfer functions . . . . . . . . . . . . . . . 154
4.6.4 Bound for the balanced truncation error . . . . . 155
4.7 Generalized gramians and truncations . . . . . . . . . . . 160
4.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
5 Stabilizing Controllers 167
5.1 System Stability . . . . . . . . . . . . . . . . . . . . . . . 169
5.2 Stabilization . . . . . . . . . . . . . . . . . . . . . . . . . 172
5.2.1 Static state feedback stabilization via LMIs . . . 173
5.2.2 An LMI characterization of the stabilization prob-
lem . . . . . . . . . . . . . . . . . . . . . . . . . . 174
5.3 Parametrization of stabilizing controllers . . . . . . . . . 175
5.3.1 Coprime factorization . . . . . . . . . . . . . . . . 176
5.3.2 Controller Parametrization . . . . . . . . . . . . . 179
5.3.3 Closed-loop maps for the general system . . . . . 183
5.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
6 H2 Optimal Control 188
6.1 Motivation for H2 control . . . . . . . . . . . . . . . . . 190
6.2 Riccati equation and Hamiltonian matrix . . . . . . . . . 192
6.3 Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
6.4 State feedback H2 synthesis via LMIs . . . . . . . . . . . 202
6.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
7 H1 Synthesis 208
7.1 Two important matrix inequalities . . . . . . . . . . . . 209
7.1.1 The KYP Lemma . . . . . . . . . . . . . . . . . . 212
7.2 Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
7.3 Controller reconstruction . . . . . . . . . . . . . . . . . . 222
7.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
8 Uncertain Systems 227
8.1 Uncertainty modeling and well-connectedness . . . . . . 229
8.2 Arbitrary block-structured uncertainty . . . . . . . . . . 234
8.2.1 A scaled small-gain test and its suciency . . . . 236
8.2.2 Necessity of the scaled small-gain test . . . . . . . 239
8.3 The Structured Singular Value . . . . . . . . . . . . . . . 245
8.4 Time invariant uncertainty . . . . . . . . . . . . . . . . . 248
8.4.1 Analysis of time invariant uncertainty . . . . . . . 249
8.4.2 The matrix structured singular value and its upper
bound . . . . . . . . . . . . . . . . . . . . . . . . 257
8.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
iv Contents
9 Feedback Control of Uncertain Systems 270
9.1 Stability of feedback loops . . . . . . . . . . . . . . . . . 273
9.1.1 L2 -extended and stability guarantees . . . . . . . 274
9.1.2 Causality and maps on L2 -extended . . . . . . . . 277
9.2 Robust stability and performance . . . . . . . . . . . . . 280
9.2.1 Robust stability under arbitrary structured uncer-
tainty . . . . . . . . . . . . . . . . . . . . . . . . . 281
9.2.2 Robust stability under LTI uncertainty . . . . . . 281
9.2.3 Robust Performance Analysis . . . . . . . . . . . 282
9.3 Robust Controller Synthesis . . . . . . . . . . . . . . . . 284
9.3.1 Robust synthesis against
a;c .......... 285
9.3.2 Robust synthesis against
. . . . . . . . . . .
TI 289
9.3.3 D-K iteration: a synthesis heuristic . . . . . . . . 293
9.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
10 Further Topics: Analysis 298
10.1 Analysis via Integral Quadratic Constraints . . . . . . . 298
10.1.1 Analysis results . . . . . . . . . . . . . . . . . . . 303
10.1.2 The search for an appropriate IQC . . . . . . . . 308
10.2 Robust H2 Performance Analysis . . . . . . . . . . . . . 310
10.2.1 Frequency domain methods and their interpretation 311
10.2.2 State-Space Bounds Involving Causality . . . . . 316
10.2.3 Comparisons . . . . . . . . . . . . . . . . . . . . 320
10.2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . 321
11 Further Topics: Synthesis 323
11.1 Linear parameter varying and multidimensional systems 324
11.1.1 LPV synthesis . . . . . . . . . . . . . . . . . . . . 327
11.1.2 Realization theory for multidimensional systems . 333
11.2 A Framework for Time Varying Systems: Synthesis and
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
11.2.1 Block-diagonal operators . . . . . . . . . . . . . 338
11.2.2 The system function . . . . . . . . . . . . . . . . 340
11.2.3 Evaluating the `2 induced norm . . . . . . . . . . 344
11.2.4 LTV synthesis . . . . . . . . . . . . . . . . . . . . 347
11.2.5 Periodic systems and nite dimensional conditions 349
A Some Basic Measure Theory 352
A.1 Sets of zero measure . . . . . . . . . . . . . . . . . . . . 352
A.2 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 355
A.3 Comments on norms and Lp spaces . . . . . . . . . . . . 357
B Proofs of Strict Separation 359
C -Simple Structures 365
 Contents v
C.1 The case of
1; 1 . . . . . . . . . . . . . . . . . . . . . . 366
C.2 The case of
0; 3 . . . . . . . . . . . . . . . . . . . . . . 370
References 375
 This is page vi
Printer: Opaque this

Preface

Research in robust control theory has been one of the most active areas of
mainstream systems theory since the late 70s. This research activity has
been at the con uence of dynamical systems theory, functional analysis,
matrix analysis, numerical methods, complexity theory, and engineering ap-
plications. The discipline has involved interactions between diverse research
groups including pure mathematicians, applied mathematicians, computer
scientists and engineers, and during its development there has been a sur-
prisingly close connection between pure theory and tangible engineering
application. By now this research eort has produced a rather extensive
set of approaches using a wide variety of mathematical techniques, and
applications of robust control theory are spreading to areas as diverse as
control of uids, power networks, and the investigation of feedback mech-
anisms in biology. During the 90s the theory has seen major advances and
achieved a new maturity, centered around the notion of convexity. This em-
phasis is two-fold. On one hand, the methods of convex programming have
been introduced to the eld and released a wave of computational meth-
ods which, interestingly, have impact beyond the study of control theory.
Simultaneously a new understanding has developed on the computational
complexity implications of uncertainty modeling; in particular it has be-
come clear that one must go beyond the time invariant structure to describe
uncertainty in terms amenable to convex robustness analysis.
Our broad goal in this book is to give a graduate-level course on ro-
bust control theory that emphasizes these new developments, but at the
same time conveys the main principles and ubiquitous tools at the heart
of the subject. This course is intended as an introduction to robust control
 Contents vii
theory, and begins at the level of basic systems theory, but ends having
introduced the issues and machinery of current active research. Thus the
pedagogical objectives of the book are (1) to introduce a coherent and uni-
ed framework for studying robust control theory; (2) to provide students
with the control-theoretic background required to read and contribute to
the research literature; (3) the presentation of the main ideas and demon-
strations of the major results of robust control theory. We therefore hope
the book will be of value to mathematical researchers and computer sci-
entists wishing to learn about robust control theory, graduate students
planning to do research in the area, and engineering practitioners requiring
advanced control techniques. The book is meant to feature convex methods
and the viewpoint gained from a general operator theory setting, however
rather than be purist we have endeavored to give a balanced course which
ts these themes in with the established landscape of robust control the-
ory. The eect of this intention on the book is that as it progresses these
themes are increasingly emphasized, whereas more conventional techniques
appear less frequently. The current research literature in robust control
theory is vast and so we have not attempted to cover all topics, but have
instead selected those that we believe are central and most eectively form
a launching point for further study of the eld.
The text is written to comprise a two-quarter or two-semester gradu-
ate course in applied mathematics or engineering. The material presented
has been successfully taught in this capacity during the past few years by
the authors at Caltech, University of Waterloo, University of Illinois, and
UCLA. For students with background in state space methods a serious
approach at a subset of the material can be achieved in one semester. Stu-
dents are assumed to have familiarity with linear algebra, and otherwise
only advanced calculus and basic complex analysis are strictly required.
After an introduction and a preliminary technical chapter, the course
begins with a thorough introduction to state space systems theory. It then
moves on to cover open-loop systems issues using the newly introduced
concept of a norm. Following this the simplest closed-loop synthesis issue
is addressed, that of stabilization. Then there are two chapters on synthesis
which cover the H2 and H1 formulations. Next open-loop uncertain system
models are introduced; this chapter gives a comprehensive treatment of
structured uncertainty using perturbations that are either time invariant
or arbitrary. The results on open-loop uncertain systems are then applied to
feedback control in the following chapter where both closed-loop analysis
and synthesis are addressed. The nal two chapters are devoted to the
presentation of four advanced topics in a more descriptive manner. In the
preliminary chapter of the book some basic ideas from convex analysis are
presented as is the important concept of a linear matrix inequality (LMI).
Linear matrix inequalities are perhaps the major analytical tool used in
this text, and combined with the operator theory framework presented later
viii Contents
provide a powerful perspective. A more detailed summary of the chapters
is given below.

Chapter 1 Preliminaries in Finite Dimensional Space

Elementary linear algebra is rst reviewed, and a short summary of basic
concepts from convex analysis are provided. A selection of matrix theory
topics is presented including Jordan form and singular value decomposition.
The chapter ends with a section on linear matrix inequalities.

Chapter 2 State Space Systems Theory

This chapter introduces the basic state space model. Controllability, reach-
ability and stabilizability are then covered in conjunction with various tests
for these properties. Eigenvalue assignment is discussed in full for multi-
input systems. The concept of observability is then presented together with
an introduction to observers and observer based controllers. Minimal real-
izations are treated, and the connection between state space realizations
and transfer functions is made.

Chapter 3 Linear Analysis

The major objective of this chapter is to introduce the operator theory
needed in the sequel. It begins by introducing normed and inner product
spaces, and then the operators on Hilbert space. The focus of the text is
systems with L2 inputs, and so a number of related function spaces are
presented, including the H2 and H1 spaces. The various connections of
these spaces with time invariance and causality are introduced.

Chapter 4 Model Realization and Reduction

The chapter initiates the quantitative study of systems using norms as
the system measure. The open-loop characteristics of systems are exam-
ined using the controllabilty and observability gramians, and a geometric
motivation is given for balanced realizations. Hankel operators and singu-
lar values are then discussed, followed by model reduction using balanced
truncation.

Chapter 5 Stabilizing Controllers

The concepts of closed-loop well-posedness and stability are dened. An
LMI solution to nding a stabilizing controller is stated. Following this
the question of parametrizing all stabilizing controllers is pursued, and the
important idea of a coprime factorization is brought in. With this tool a
complete controller parametrization is given.
 Contents ix
Chapter 6 H2 Optimal Control
Various interpretations of the system H2 norm are given, and the optimal
H2 control problem is formulated. Subsequently the Riccati equation is
dened and some of its basic properties are elucidated. A solution to the
the optimal H2 problem is then derived. Finally an LMI solution is given
to the H2 state feedback synthesis problem.
Chapter 7 H1 Synthesis
The objective of this chapter is to pose the H1 synthesis problem and give
a complete LMI derivation. At the beginning of the chapter an LMI version
of the KYP lemma is proved making connections with the Riccati theory
of Chapter 6. Then the synthesis problem is solved in full, yielding general
constructive conditions.
Chapter 8 Uncertain Systems
In this chapter a framework for modeling systems with uncertainty is put
forth, based on structured perturbations in a feedback interconnection. The
notion of well-connectedness of these representations is presented. Then
necessary and sucient conditions are derived for well-connectedness un-
der arbitrary structured perturbations, which take the form of a generalized
small gain conditions. This leads to the extension of the small-gain approach
to more general uncertainty structures by introducing the structured sin-
gular value. Then the method is applied to the study of time-invariant
structured uncertainty, and yields conditions in terms of the matrix struc-
tured singular value. Finally, the latter and its upper bound are studied in
geometric terms by means of quadratic forms in Euclidean space.
Chapter 9 Feedback Control of Uncertain Systems
The concepts of Chapter 8 are applied to the problem of robustness
analysis for feedback control. In particular causality is introduced into
the well-connectedness problem and the questions of robust stability and
performance are analyized. Then the D-K iteration heuristic for robust
synthesis is presented in detail; this involves a combination of the synthesis
procedures from earlier chapters and the uncertain analysis conditions just
derived in Chapter 8.
Chapter 10 Further Topics: Analysis
Two advanced topics are presented. The rst is an introduction to integral
quadratic constraints (IQCs), which form and important and more general
way to dene uncertainty in systems; as with much of the earlier work
the associated analysis conditions reduce to LMIs. The second topic is
the robust H2 problem. This is an important control topic which aims to
x Contents
reconcile the superior performance characteristics of the H2 norm with the
robustness advantages of using perturbations measured in the L2 -induced
norm.
Chapter 11 Further Topics: Synthesis
This chapter considers two new topics as well. It starts with a generaliza-
tion of the one-dimensional state space framework, studied thus far in the
text, to multiple-dimensions. This approach has applications to uncertain
system realization, gain scheduling and distributed systems. Finally time
varying systems are studied using an operator theoretic framework, and
we show how these systems can be treated formally as though they were
time invariant. Thus analogs of many of results derived in the book are
immediately available in this more general setting, by simply converting
LMIs to structured operator inequalities.

Geir E. Dullerud Fernando G. Paganini

Urbana Los Angeles
 This is page xi
Printer: Opaque this

List of Figures

1 Basic block diagram . . . . . . . . . . . . . . . . . . . . . 3

2 System decomposition . . . . . . . . . . . . . . . . . . . . 5
3 Double pendulum with torque input . . . . . . . . . . . . 11
4 Controlled system . . . . . . . . . . . . . . . . . . . . . . 12
5 Double pendulum with attached uid-vessel. . . . . . . . 16
1.1 Convex and nonconvex sets . . . . . . . . . . . . . . . . . 33
1.2 Convex hull of nite number of points . . . . . . . . . . . 33
1.3 Convex and nonconvex sets . . . . . . . . . . . . . . . . . 34
1.4 A hyperplane . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.5 Separating hyperplane . . . . . . . . . . . . . . . . . . . . 35
1.6 Convex and nonconvex cones . . . . . . . . . . . . . . . . 38
1.7 Illustration of semidenite programming . . . . . . . . . . 51
4.1 Observability ellipsoid . . . . . . . . . . . . . . . . . . . . 136
4.2 Unbalanced system ellipsoids . . . . . . . . . . . . . . . . 141
4.3 Results of G and ;G operating on a given u 2 L2 (;1; 0]. 144
4.4 Log-log plots for the example. Top: jG^ (j!)j. Bottom: jG^ (j!)j
(dotted), jG^ 4 (j!)j (full) and jG^ 2 (j!)j (dashed). . . . . . . 159
5.1 General feedback arrangement . . . . . . . . . . . . . . . . 167
5.2 Input-output stability . . . . . . . . . . . . . . . . . . . . 171
5.3 Coprime factorization . . . . . . . . . . . . . . . . . . . . 180
xii List of Figures
6.1 Synthesis arrangement . . . . . . . . . . . . . . . . . . . . 189
8.1 Uncertain system model . . . . . . . . . . . . . . . . . . . 230
8.2 Cascade of uncertain systems . . . . . . . . . . . . . . . . 233
8.3 (a) D(; r) > 0; (b) , r separated by a hyperplane. . . 242
8.4 Example system. . . . . . . . . . . . . . . . . . . . . . . . 250
9.1 Feedback system with uncertainty . . . . . . . . . . . . . 270
9.2 Robustness analysis setup . . . . . . . . . . . . . . . . . . 272
9.3 Uncertain plant . . . . . . . . . . . . . . . . . . . . . . . . 275
9.4 Scaled synthesis problem . . . . . . . . . . . . . . . . . . . 286
10.1 Setup for robustness analysis . . . . . . . . . . . . . . . . 303
10.2 Setup for Robust H2 analysis. . . . . . . . . . . . . . . . . 310
10.3 Constraints on the accumulated spectrum . . . . . . . . . 314
11.1 NMD system . . . . . . . . . . . . . . . . . . . . . . . . . 324
11.2 NMD synthesis conguration . . . . . . . . . . . . . . . . 328
11.3 Closed-loop system . . . . . . . . . . . . . . . . . . . . . . 348
B.1 Signals q and p (dashed); q~ and p~ (solid). . . . . . . . . . 362
C.1 Illustration of the proof. . . . . . . . . . . . . . . . . . . . 373
 This is page 1
Printer: Opaque this

0
Introduction

In this course we will explore and study a mathematical approach aimed

directly at dealing with complex physical systems that are coupled in feed-
back. The general methodology we study has analytical applications to
both human-engineered systems and systems that arise in nature, and the
context of our course will be its use for feedback control.
The direction we will take is based on two related observations about
models for complex physical systems. The rst is that analytical or com-
putational models which closely describe physical systems are dicult or
impossible to precisely characterize and simulate. The second is that a
model, no matter how detailed, is never a completely accurate represen-
tation of a real physical system. The rst observation means that we are
forced to use simplied system models for reasons of tractability; the lat-
ter simply states that models are innately inaccurate. In this course both
aspects will be termed system uncertainty, and our main objective is to
develop systematic techniques and tools for the design and analysis of sys-
tems which are uncertain. The predominant idea that is used to contend
with such uncertainty or unpredictability is feedback compensation.
There are several ways in which systems can be uncertain, and in this
course we will target the main three:
 The initial conditions of a system may not be accurately specied or
completely known.
 Systems experience disturbances from their environment, and system
commands are typically not known a priori.
2 0. Introduction
 Uncertainty in the accuracy of a system model itself is a central
source. Any dynamical model of a system will neglect some physi-
cal phenomena, and this means that any analytical control approach
based solely on this model will neglect some regimes of operation.
In short: the major objective of feedback control is to minimize the eects
of unknown initial conditions and external in uences on system behavior,
subject to the constraint of not having a complete representation of the sys-
tem. This is a formidable challenge in that predictable behavior is expected
from a controlled system, and yet the strategies used to achieve this must
do so using an inexact system model. The term robust in the title of this
course refers to the fact that the methods we pursue will be expected to
operate in an uncertain environment with respect to the system dynamics.
The mathematical tools and models we use will be primarily linear, moti-
vated mainly by the requirement of computability of our methods; however
the theory we develop is directly aimed at the control of complex nonlinear
systems. In this introductory chapter we will devote some space to discuss,
at an informal level, the interplay between linear and nonlinear aspects in
this approach.
The purpose of this chapter is to provide some context and motivation
for the mathematical work and problems we will encounter in the course.
For this reason we do not provide many technical details here, however it
might be informative to refer back to this chapter periodically during the
course.

0.1 System representations

We will now introduce the diagrams and models used in this course.

0.1.1 Block diagrams

We will often view physical or mathematical systems a mappings. From
this perspective a system maps an input to an output; for dynamical sys-
tems these are regarded as functions of time. This is not the only or most
primitive way to view systems, although we will nd this viewpoint to be
very attractive both mathematically and for guiding and building intuition.
In this section we introduce the notion of a block diagram for representing
systems, and most importantly for specifying their interconnections.
We use the symbol P to denote a system that maps an input function
u(t) to an output function y(t). This relationship is denoted by
y = P (u):
Figure 1 illustrates this relationship. The direction of the arrows indicate
whether a function is an input or an output of the system P . The details
 0.1. System representations 3

y P u

Figure 1. Basic block diagram

of how P constructs y from the input u is not depicted in the diagram,

instead the benet of using such block diagrams is that interconnections of
systems can be readily visualized.
Consider the so-called cascade interconnection of the two subsystems.
This interconnection represents the equations

y P2 v P1 u

v = P1 (u)
y = P2 (v):
We see that this interconnection takes the two subsystems P1 and P2 to
form a system P dened by P (u) = P2 (P1 (u) ). Thus this diagram simply
depicts a composition of maps. Notice that the input to P2 is the output
of P1 .
z w
P
u

y
Q

Another type of interconnection involves feedback. In the gure above

we have such an arrangement. Here P has inputs given by the ordered pair
(w; u) and the outputs (z; y). The system Q has input y and output u.
This block diagram therefore pictorially represents the equations
(z; y) = P (w; u)
y = Q(y):
Since part of the output of P is an input to Q, and conversely the output
of Q is an input to P , these systems are coupled in feedback.
4 0. Introduction
We will now move on to discussing the basic modeling concept of this
course and in doing so will immediately make use of block diagrams.

0.1.2 Nonlinear equations and linear decompositions

We have just introduced the idea of representing a system as an input-
output mapping, and did not concern ourselves with how such a mapping
might be dened. We will now outline the main idea behind the modeling
framework used in this course, which is to represent a complex system as
a combination of a perturbation and a simpler system. We will illustrate
this by studying two important cases.
Isolating nonlinearities
The rst case considered is the decomposition of a system into a linear part
and a static nonlinearity. The motivation for this is so that later we can
replace the nonlinearity using objects more amenable to analysis.
To start consider the nonlinear system described by the equations
x_ = f (x; u) (1)
y = h(x; u);
with the initial condition x(0). Here x(t), y(t) and u(t) are vector valued
functions, and f and h are smooth vector valued functions. The rst of
these equations is a dierential equation and the second is purely algebraic.
Given an initial condition and some additional technical assumptions, these
equations dene a mapping from u to y. Our goal is now to decompose this
system into a linear part and a nonlinear part around a specied point; to
reduce clutter in the notation we assume this point is zero.
Dene the following equivalent system
x_ = Ax + Bu + g(x; u) (2)
y = Cx + Du + r(x; u);
where A, B , C and D provide a linear approximation to the dynamics, and
g(x; u) = f (x; u) ; Ax ; Bu
r(x; u) = h(x; u) ; h(0; 0) ; Cx ; Du:
For instance one could take the Jacobian linearization
A = d1 f (0; 0); B = d2 f (0; 0);
C = d1 h(0; 0); and D = d2 h(0; 0);
where d1 and d2 denote vector dierentiation by the rst and second vector
variables respectively. The following discussion, however, does not require
this assumption. The system in (2) consists of linear functions and the
possibly nonlinear functions g and r. It is clear that the solutions to this
 0.1. System representations 5
system have a one-to-one correspondence with the solutions of (1), since we
have simply rewritten the functions. Further let us write these equations
in the equivalent form
x_ = Ax + Bu + w1 (3)
y = Cx + Du + w2 (4)
(w1 ; w2 ) = (g(x; u); r(x; u)): (5)
Now let G be the mapping described by (3) and (4) which satises
G : (w1 ; w2 ; u) 7! (x; u; y);
given an initial condition x(0). Further let Q be the mapping which takes
(x; u) 7! (w1 ; w2 ) as described by (5). Thus the system of equations dened
by (3 - 5) has the block diagram below. The system G is totally described

G
y u
Figure 2. System decomposition
by linear dierential equations, and Q is a static nonlinear mapping. By
static we mean that the output of Q at any point in time depends only on
the input at that particular time, or equivalently that Q has no memory.
Thus all of the nonlinear behavior of the initial system (1) is captured in
Q and the feedback interconnection.
We will almost exclusively work with the case where the point (0; 0),
around which this decomposition is taken, is an equilibrium point of (1).
Namely
f (0; 0) = 0:
In this case the functions g and r satisfy g(0; 0) = 0 and r(0; 0) = 0, and
therefore Q(0; 0) = 0. Also the linear system described by
x_ = Ax + Bu
y = Cx + Du
is the linearization of (1) around the equilibrium point. The linear system
G is thus an augmented version of the linearization.
6 0. Introduction
Higher order dynamics
In the construction just considered we were able to isolate nonlinear system
aspects in the mapping Q, and our motivation for this was so that later
we will be able to replace Q with an alternative description which is more
easily analyzed. For the same reason we will sometimes wish to do this
not only with the nonlinear part of a system, but also with some of its
dynamics. Let us now move on to consider this more complex scenario. We
have the equations
   
x_ 1 = f1 (x1 ; x2 ; u) (6)
x_ 2 f2 (x1 ; x2 ; u)
y = h(x1 ; x2 ; u);
Following a similar procedure to the one we just carried out on the system
in (1), we can decompose the system described in (6) to arrive at the
equivalent set of equations:
x_ 1 = A1 x1 + B1 u + g1 (x1 ; x2 ; u) (7)
x_ 2 = f2 (x1 ; x2 ; u)
y = C1 x1 + Du + r(x1 ; x2 ; u):
This is done by focusing on the equations x_ 1 = f1 (x1 ; x2 ; u) and y =
h(x1 ; x2 ; u), and performing the same steps as before treating both x2 and
u as the inputs. The equations in (7) are equivalent to the linear equations
x_ 1 = A1 x1 + B1 u + w1 (8)
y = C1 x1 + Du + w2 ;
coupled with the nonlinear equations
x_ 2 = f2 (x1 ; x2 ; u) (9)
(w1 ; w2 ) = (g1 (x1 ; x2 ; u); r(x1 ; x2 ; u) ):
Now similar to before we set G to be the linear system
G : (w1 ; w2 ; u) 7! (x1 ; u; y)
which satises the equations in (8). Also dene Q to be the system described
by (9) where
Q : (x1 ; u) 7! (w1 ; w2 ):
With these new denitions of P and Q we see that Figure 2 depicts the
system described in (6). Furthermore part of the system dynamics and all
of the system nonlinearity is isolated in the mapping Q. Notice that the
decomposition we performed on (1) is a special case of the current one.
 0.1. System representations 7
Modeling Q
In each of the two decompositions just considered we split the initial sys-
tems, given by (1) or (6), into a G-part and a Q-part. The system G was
described by linear dierential equations, whereas nonlinearities were con-
ned to Q. This decomposing of systems into a linear low dimensional part,
and a potentially nonlinear and high dimensional part, is at the heart of
this course. The main approach adopted here to deal with the Q-part will
be to replace it with a set
of linear maps which capture its behavior.
The motivation for doing this is that the resulting analysis can frequently
be made tractable.
Formally stated we require the set
to have the following property: if
q = Q(p), for some input p, then there should exist a mapping
in the set

such that
q =
(p): (10)
The key idea here is that the elements of the set
can be much simpler
dynamically than Q. However when combined in a set they are actually
able to generate all of the possible input-output pairs (p; q) which satisfy
q = Q(p). Therein lies the power of introducing
: one complex object can
be replaced by a set of simpler ones.
We now discuss how this idea can be used for analysis of the system
depicted in Figure 2. Let
S(G; Q) denote the mapping u 7! y in Figure 2.
Now replace this map with the set of maps
S(G;
)
generated by choosing
from the set
. Then we see that if the input-
output behaviors associated with all the mappings S(G;
) satisfy a given
property, then so must any input-output behavior of S(G; Q). Thus any
property which holds over the set
is guaranteed to hold for the system
S(G; Q). However the converse is not true and so analysis using
can in
general be conservative. Let us consider this issue.
If a set
has the property described in (10), then providing that it has
more than one element, it will necessarily generate more input-output pairs
than Q. Specically
f(p; q) : q = Q(p)g  f(p; q) : there exists
2
, such that q =
qg:
Clearly the set on the left denes a function, whereas the input-output
pairs generated by
is in general only a relation. The degree of closeness
of these sets determines the level of conservatism introduced by using

in place of Q.
We now illustrate how the behavior of Q can be captured by
with two
simple examples.
8 0. Introduction
Examples:
We begin with the decomposition for (1). For simplicity assume that x, y
and u are all scalar valued functions. Now suppose that the functions r and
g, which dene Q, are known to satisfy the sector or Lipschitz bounds
jw1 (t)j  k11 jx(t)j + k12 ju(t)j (11)
jw2 (t)j  k21 jx(t)j + k22 ju(t)j;
for some positive constants kij . It follows that if for particular sig-
nals (w1 ; w2 ) = Q(x; u), then there exist scalar functions of time
11 (t),12 (t),21 (t) and 22 (t), each satisfying ij (t) 2 [;kij ; kij ], such that
w1 (t) = 11 (t)x(t) + 12 (t)u(t) (12)
w2 (t) = 21 (t)x(t) + 22 (t)u(t);
Dene the set
to consist of all 2  2 matrix functions
which satisfy
 

= 11 ((tt)) 12 ((tt)) ; where jij (t)j  kij for each time t  0.
21 22
From the above discussion it is clear the set
has the property that given
any inputs and outputs satisfying (w1 ; w2 ) = Q(x; u), there exists
2

satisfying (12).
Let us turn to an analogous construction associated with the decomposi-
tion of the system governed by (6), recalling that Q is now dynamic. Assume
x1 and u are scalar, and suppose it is known that if (w1 ; w2 ) = Q(x1 ; u)
then the following energy inequalities hold:
Z 1 Z 1 Z 1 
jw1 (t)j2 dt  k1 jx1 (t)j2 dt + ju(t)j2 dt
0 0 0
Z 1 Z 1 Z 1 
jw2 (t)j2 dt  k2 jx1 (t)j2 dt + ju(t)j2 dt ;
0 0 0
when the right hand side integrals are nite. Dene
to consist of all linear
mappings
: (x1 ; u) 7! (w1 ; w2 ) which satisfy the above inequalities for
all functions x1 and u from a suitably dened class. It is possible to show
that if (w1 ; w2 ) = Q(x1 ; u), for some bounded energy functions x1 and u,
then there exists a mapping
in
such that
(x1 ; u) = (w1 ; w2 ). In this
sense
can generate any behavior of Q. As a remark, inequalities such
as the above assume implicitly that initial conditions in the state x2 can
be neglected; in the language of the next section, there is a requirement of
stability in high-order dynamics that can be isolated in this way. 
Using a set
instead of the mapping Q has another purpose. As already
pointed out physical systems will never be exactly represented by models
of the form (1) or (6). Thus the introduction of the set
aords a way to
account for potential system behaviors without explicitly modeling them.
For example the inequalities in (11) may be all that is known about some
 0.2. Robust control problems and uncertainty 9
higher order dynamics of a system; note that given these bounds we would
not even need to know the order of these dynamics to account for them
using
. Therefore this provides a way to explicitly incorporate knowledge
about the unpredictability of a physical system into a formal model. Thus
the introduction of the
serves two related but distinct purposes:
 provides a technique for simplifying a given model;
 can be used to model and account for uncertain dynamics.
In the course we will study analysis and synthesis using these types of
models, particularly when systems are formed by the interconnection of
many such subsystems. We call these types of models uncertain systems.

0.2 Robust control problems and uncertainty

In this section we outline three of the basic control scenarios pursued in this
course. Our discussion is informal and is intended to provide motivation
for the mathematical analysis in the sequel.

0.2.1 Stabilization
One of the most basic goals of a feedback control system is stabilization.
This means nullifying the eects of the uncertainty surrounding the initial
conditions of a system. Before explaining this in more detail we review
some basic concepts.
Consider the autonomous system
x_ = f (x); with some initial condition x(0). (13)
We will be concerned with equilibrium points xe of this system, namely
points where f (xe ) = 0 is satised. Without loss of generality in this dis-
cussion we shall assume that xe = 0, since this can always be arranged by
redening f appropriately. We say that the equilibrium point zero is stable
if for any initial condition x(0) suciently near to zero, the time trajec-
tory x(t) remains near to zero. The equilibrium is exponentially stable if
it is stable and furthermore the function x(t) tends to zero at an exponen-
tial rate when x(0) is chosen suciently small. Stability is an important
property because it is unlikely that a physical system is ever exactly at an
equilibrium point. It says that if the initial state of the system is slightly
perturbed away from the equilibrium, the resulting state trajectory will
not diverge. Exponential stability goes further to say that if such initial
deviations are small then the system trajectory will tend quickly back to
the equilibrium point. Thus stable systems are insensitive to uncertainty
about their initial conditions.
10 0. Introduction
We now review a test for exponential stability. Suppose
x_ = Ax
is the linearization of (13) at the point zero. It is possible to show: the
zero point of (13) is exponentially stable if and only if the linearization
is exponentially stable at zero.1 The linearization is exponentially stable
exactly when all the eigenvalues of the matrix A have negative real part.
Thus exponential stability can be checked directly by calculating A, the
Jacobian matrix of f at zero. Further it can be shown that if any of the
eigenvalues have positive real part, then the equilibrium point zero is not
even stable.
We now move to the issue of stabilization, which is using a control law
to turn an unstable equilibrium point, into an exponentially stable one.
Below is a controlled nonlinear system
x_ = f (x; u); (14)
where the input is the function u. Suppose that (0; 0) is an equilibrium
point of this system. Our stability denitions are extended to such con-
trolled systems in the following way: the equilibrium point (0; 0) is dened
to be (exponentially) stable if zero is an (exponentially) stable equilibrium
point of the autonomous system x_ = f (x; 0).
Our rst task is to investigate conditions under which it is possible to
stabilize such an equilibrium point using a special type of control strategy
called a state feedback. In this scenario we seek a control feedback law of
the form
u(t) = p(x(t));
where p is a smooth function, such that the closed loop system
x_ = f (x; p(x(t)) )
is exponentially stable around zero. That is we want to nd a function p
which maps the state of the system x to a control action u. Let us assume
that such a p exists and examine some of its properties. First notice that in
order for zero to be an equilibrium point of the closed loop we may assume
p(0) = 0:
Given this the linearization of the closed loop is
x_ = (A + BF )x;
where A = d1 f (0; 0), B = d2 f (0; 0) and F = dp(0). Thus we see that the
closed loop system is exponentially stable if and only if all the eigenvalues
of A + BF are strictly in the left half of the complex plane. Conversely
1
This is under the assumption that f is suciently smooth.
 0.2. Robust control problems and uncertainty 11
notice that if a matrix F exists such that A + BF has the desired stability
property, then the state feedback law p(x) = Fx will stabilize the closed
loop.
In the scenario just discussed the state x was directly available to use
in the feedback law. A more general control situation occurs when only
an observation y = h(x; u) is available for feedback, or when a dynamic
control law is employed. For these the analysis is more complicated, we
defer the study to subsequent chapters. We now illustrate these concepts
with an example.
Stabilization of a double pendulum
Shown below in Figure 3 is a double pendulum. In the gure two rigid
links are connected by a hinge joint, so that they can rotate with respect
to each other. The rst link is also constrained to rotate about a point
which is xed in space. The control input to this rigid body system is a

 = torque

g = gravity

1

2

Figure 3. Double pendulum with torque input

torque  which is applied to the rst link. The conguration of this system
is completely specied by 1 and 2 , which are the respective angles that
the rst and second links make with the vertical. Since we have assumed
that the links are ideal rigid bodies, this system can be described by a
dierential equation of the form (14), where x = (1 ; 2 ; _1 ; _2 ) and u =  .
It is routine to show that for each xed angle 1e of the rst link, there
exists a torque e , such that
((1e ; 2e ; 0; 0); e ) is an equilibrium point;
where 2e is equal to either zero or . That is for any value of 1e we
have two equilibrium points of the system; both occur when the second
12 0. Introduction
link is vertical. When 2e = 0 the equilibrium point is stable. The 2e = 
equilibrium point is unstable.
We may wish to stabilize the pendulum about its upright position at such
an equilibrium point. To apply a state feedback control law as discussed
above we require the ability to measure x, namely we base our control law
on the link angles and their velocities. In the more general scheme, also
described above, we would only have access to some function of these four
measurements; a typical situation is that the observation is (1 ; 2 ) the two
angles but not their velocities.

0.2.2 Disturbances and commands

Dynamic stability is usually a basic requirement of a controlled system,
however typically much more is demanded, and in fact often feedback is
introduced in systems which are already stable, with the objective of im-
proving dierent aspects of the dynamic behavior. An important issue is
the eect of unknown environmental in uences. For instance consider the
ideal double pendulum just discussed above; more realistically such an ap-
paratus is also in uenced by ground vibrations felt at the point of xture,
or it may experience forces on its links as a result of air currents or \gusts".
Since such external eects are rarely expected to help achieve desired sys-
tem behavior (e.g. balance the pendulum) they are commonly referred as
disturbances. One of the main objectives of feedback is to render a system
insensitive to such disturbances, and a signicant portion of our course will
be devoted to the study of systems from this point of view.
A pictorial representation of a controlled system being in uenced by
unknown inputs, such as disturbances, is shown below. Here we have a dy-

system Dynamical unknown

output input
System

measurement action

Control law

Figure 4. Controlled system

namical system in feedback with a control law, which takes some actions
based on measurement information. However there are other environmen-
tal in uences acting on the systems, which are unknown at the time of
 0.2. Robust control problems and uncertainty 13
control design, and could be disturbances, or also external commands. The
behavior of the system is characterized by the outputs. To bring sharper
focus to our discussion we consider a second example.
Position control of an electric motor

d
+ i

v 


The gure depicts a schematic diagram of an electric motor controlled
by excitation: a voltage applied to the motor windings results in a torque
applied to the motor shaft. While physical details are not central to our
discussion, we will nd it useful to write down an elementary dynamical
model. The key variables are
 v applied voltage.
 i current in the eld windings.
  motor torque and d opposing torque from the environment.
  angle of the shaft and ! = _ angular velocity.
The objective of the control system is for the angular position  to follow
a reference command r , despite the eect of an unknown resisting torque
d . This so-called servomechanism problem is common in many applica-
tions, for instance we could think of moving a robot arm in an uncertain
environment.
We begin by writing down a dierential equation model for the motor:
di
v = Ri + L dt (15)
 = i (16)
d!
J dt =  ; d ; B! (17)
Here (15) models the electrical circuit in terms of its resistance R and
inductance L; (16) says the torque is a linear function of the current; and
14 0. Introduction
(17) is the rotational dynamics of the shaft, where J is the moment of
inertia and B is mechanical damping. Since the electrical transients of (15)
are typically much faster than the mechanical dynamics, it seems reasonable
to neglect the former, setting L = 0. Also in what follows we normalize, for
simplicity, the remaining constants (R, , J , B ) to unity in an appropriate
system of units.
We now address the problem of controlling our motor to achieve the
desired objective. We do this by a control system that measures the output
 and acts on the voltage v, following the law
v(t) = K (r ; );
where K > 0 is a proportionality constant to be designed. Intuitively,
the system applies torque in the adequate direction to counteract the error
between the command r and the actual output angle . It is an instructive
exercise, left for the reader, to express this system in terms of Figure 4.
Here the driving signals are the command r and the disturbance torque
d , which are unknown at the time we design K .
Given this control law, we can nd the equations of the resulting closed
loop dynamics, and with the above conventions obtain the following.
       
d  = 0 1  + 0 0 r
dt ! ;K ;1 ! K ;1 d
Thus our resulting dynamics have the form
x_ = Ax + Bw
encountered in the previous section.
We rst discuss system stability. It is straightforward to verify that the
eigenvalues of the above A matrix have negative real part whenever K >
0, therefore in the absence of external inputs the system is exponentially
stable. In particular, initial conditions will have asymptotically no eect.
Now suppose r and d are constant over time, then by solving the dif-
ferential equation it follows that the states (; !) converge asymptotically
to
d and !(1) = 0:
(1) = r ; K
Thus the motor achieves an asymptotic position which has an error of d =K
with respect to the command. We make the following remarks:
 Clearly if we make the constant K very large we will have accurate
tracking of r despite the eect of d . This highlights the central
role of feedback in achieving system reliability in the presence the
uncertainty about the environment. We will revisit this issue in the
following section.
 0.2. Robust control problems and uncertainty 15
 The success in this case depends strongly on the fact that r and d are
known to be constant; in other words while their value was unknown,
we had some a priori information about their characteristics.
The last observation is in fact general to any control design question.
That is some information about the unknown inputs is required for us to
be able to assess system performance. In the above example the signals were
specied except for a parameter. More generally the information available
is not so strongly specied, for example one may know something about the
energy or spectral properties of commands and disturbances. The available
information is typically expressed in one of two forms:
(i) We may specify a set D and impose that the disturbance should lie
in D;
(ii) We may give a statistical description of the disturbance signals.
The rst alternative typically leads to questions about the worst possible
system behavior caused by any element of D. In the second case, one usually
is concerned with statistically typical behavior. Which is more appropriate
is application dependent, and we will provide methods for both alternatives
in this course.
We are now ready to discuss a more complicated instance of uncertainty
in the next section.

0.2.3 Unmodeled dynamics

When modeling a physical system we always approximate some aspects of
the physical phenomena, due to an incomplete theory of physics. Also we
frequently further simplify our models by choice so as to make analysis
more tractable, when we believe such a simplication is innocuous. Now
we immediately wonder about the eect of such approximations when we
apply feedback to a system.
Take for instance our previous example of the electric motor, where we
deliberately neglected the eects of inductance in the electric circuit which
was deemed irrelevant to a study at the time scale of mechanical motion.
And the conclusion of our study was that we should make the constant K
as large as possible in order to achieve good tracking performance.
However if we keep the inductance L in our model, and repeat the anal-
ysis, it is not dicult to see that the resulting third order system becomes
unstable for a suciently high value of K . Thus we see that our seemingly
benign modeling error can be amplied by feedback to the point of making
the system unusable. Thus feedback is a double-edged sword: it can render
a system insensitive to uncertainty (e.g. the torque disturbance d ), but it
can also increase sensitivity to it, as was just seen. Thus feedback design
always involves a judicious balance of this fundamental tradeo.
16 0. Introduction
uid

 = torque

g = gravity

1

2

Figure 5. Double pendulum with attached uid-vessel.

At this point the reader may be thinking that this diculty was due
exclusively to careless modeling, we should have worked with the full, third
order model. Note however that there are many other dynamical aspects
which have been neglected. For instance the bending dynamics of the motor
shaft could also be described by additional state equations, and so on. We
could go to the level of spatially distributed, innite dimensional dynamics,
and there would still be neglected eects. No matter where one stops in
modeling, the reliability of the conclusions depends strongly on the fact that
whatever has been neglected will not become crucial later. In the presence of
feedback, this assessment is particularly challenging and is really a central
design question.
To emphasize this point in another example, consider the modied double
pendulum shown in Figure 5. In this new setup a vessel containing uid has
been rigidly attached to the end of the second link. Suppose following our
discussion of the previous two sections, that we wish to stabilize this system
about one of its equilibria. The addition of the uid-vessel to this system
signicantly complicates the modeling of this system, and transforms our
two rigid body system into an innite dimensional system which is highly
intractable, to the point where it is even beyond the scope of accurate
computer simulation.
 0.2. Robust control problems and uncertainty 17
However to balance this system an innite dimensional model is probably
not required, and perhaps a low dimensional one will suce. An extreme
model in this latter category would be one that modeled the uid-vessel
system as a point mass; it may well be that a feedback design which renders
our system insensitive to the value of this mass would perform well in the
real system. But possibly the oscillations of the uid inside the vessel may
compromise performance. In this case a more rened, but still tractable
model could consist of an oscillatory mass-spring type dynamical model.
These modeling issues become even more central if we interconnect many
uncertain or complex systems, to form a \system of systems". A very simple
example is the coupled system formed when the electric motor above is used
to generate the control torque for the uid-pendulum system.
The main conclusion we make is that there is no such thing as the \cor-
rect" model for control. A useful model is one in which the remaining
uncertainty or unpredictability of the system can be adequately compen-
sated by feedback. Thus we have set the stage for this course. The key
players are feedback, stability, performance, uncertainty and interconnec-
tion of systems. The mathematical theory to follow is motivated by the
challenging interplay between these aspects of designed dynamical systems.

Notes and References

For a precise denition of stability and theorems on linearization see any
standard text on dynamical systems theory; for instance [52]. For specic re-
sults on exponential stability and additional stability results in the context
of control theory see [69]. The double pendulum control example of x0.2.1
originates in [124], where it is named the pendubot. The primary focus of
this book is control theory; see [123] for more information on applications
and practical aspects of design.
 This is page 18
Printer: Opaque this

1
Preliminaries in Finite Dimensional
Space

This chapter is centered around nite dimension vector spaces, mappings

on them, and the convexity property.
Much of the material is standard linear algebra, with which the reader
is assumed to have familiarity; correspondingly, our emphasis here is to
provide a survey of the key ideas and tools, setting a common notation and
presenting some results for future reference. We provide few proofs, but
the reader can gain practice with the results and machinery presented by
completing some of the exercises at the end of the chapter.
We also cover some of the basic ideas and results from convex analysis
in nite dimensional space, which play a key role in this course. Having
completed these fundamentals we introduce a new object, linear matrix
inequalities or LMIs, which we will use throughout the course as a major
theoretical and computational tool.

1.1 Linear spaces and mappings

In this section we will introduce some of the basic ideas in linear algebra.
Our treatment is primarily intended as a review for the reader's conve-
nience, with some additional focus on the geometric aspects of the subject.
References are given at the end of the chapter for more details at both
introductory and advanced levels.
 1.1. Linear spaces and mappings 19
1.1.1 Vector spaces
The structure introduced now will pervade our course, that of a vector
space, also called a linear space. This is a set that has a natural addition
operation dened on it, together with scalar multiplication. Because this
is such an important concept, and arises in a number of dierent ways,
it is worth dening it precisely below. In the denition, the eld F can
be taken here to be the real numbers R, or the complex numbers C . The
terminology real vector space, or complex vector space is used to specify
these alternatives.
Denition 1.1. Suppose V is a nonempty set and F is a eld, and that
operations of vector addition and scalar multiplication are dened in the
following way.
(a) For every pair u, v 2 V a unique element u + v 2 V is assigned called
their sum;
(b) For each  2 F and v 2 V , there is a unique element v 2 V called
their product.
Then V is a vector space if the following properties hold for all u, v, w 2 V ,
and all ,  2 F:
(i) There exists a zero element in V , denoted by 0, such that v + 0 = v;
(ii) There exists a vector ;v in V , such that v + (;v) = 0;
(iii) The association u + (v + w) = (u + v) + w is satised;
(iv) The commutativity relationship u + v = v + u holds;
(v) Scalar distributivity (u + v) = u + v holds;
(vi) Vector distributivity ( +  )v = v + v is satised;
(vii) The associative rule ( )v = (v) for scalar multiplication holds;
(viii) For the unit scalar 1 2 F the equality 1v = v holds.
Formally, a vector space is an additive group together with a scalar mul-
tiplication operation dened over a eld F, which must satisfy the usual
rules (v){(viii) of distributivity and associativity. Notice that both V and
F contain the zero element, which we will denote by \0" regardless of the
instance.
Given two vector spaces V1 and V2 , with the same associated scalar eld,
we use V1 V2 to denote the vector space formed by their Cartesian product.
Thus every element of V1  V2 is of the form
(v1 ; v2 ) where v1 2 V1 and v2 2 V2 :
Having dened a vector space we now consider a number of examples.
20 1. Preliminaries in Finite Dimensional Space
Examples:
Both R and C can be considered as real vector spaces, although C is more
commonly regarded as a complex vector space. The most common example
of a real vector space is Rn =R 


 R; namely n copies of R. We represent
elements of Rn in a column vector notation
2 3
x1
x = 64 ... 75 2 Rn ; where each xk 2 R:
xn
Addition and scalar multiplication in Rn are dened componentwise:
2 3 2 3
x1 + y1 x1
6 x2 + y2 7 6 x2 7
x + y = 664 .. 775 ; x = 664 .. 775 ; for  2 R; x; y 2 Rn :
. .
xn + yn xn
Identical denitions apply to the complex space C n . As a further step,
consider the space C mn of complex m  n matrices of the form
2 3
a11


a1n
A = 64 ... . . . ... 75 :
am1


amn
Using once again componentwise addition and scalar multiplication, C mn
is a (real or complex) vector space.
We now dene two vector spaces of matrices which will be central in our
course. First, we dene the Hermitian conjugate or adjoint of the above
matrix A 2 C mn by
2 3
a11


am1
A = 64 ... . . . ... 75 2 C nm ;
a1n


amn
where we use a to denote the complex conjugate of a number a 2 C .
So A is the matrix formed by transposing the indices of A and taking the
complex conjugate of each element. A square matrix A 2 C nn is Hermitian
or self-adjoint if
A = A :
The space of Hermitian matrices is denoted H n , and is a real vector space.
If a Hermitian matrix A is in Rnn it is more specically referred to as
symmetric. The set of symmetric matrices is also a real vector space and
will be written Sn.
The set F (Rm ; Rn ) of functions mapping m real variables to Rn is a
vector space. Addition between two functions f1 and f2 is dened by
(f1 + f2 )(x1 ; : : : ; xm ) = f1 (x1 ; : : : ; xm ) + f2 (x1 ; : : : ; xm )
 1.1. Linear spaces and mappings 21
for any variables x1 ; : : : ; xm ; this is called pointwise addition. Scalar
multiplication by a real number  is dened by
(f )(x1 ; : : : ; xm ) = f (x1 ; : : : ; xm ):
An example of a less standard vector space is given by the set comprised
of multinomials in m variables, that have homogeneous order n. We denote
this set by Pm[n]. To illustrate the elements of this set consider
p1 (x1 ; x2 ; x3 ) = x21 x2 x3 ; p2 (x1 ; x2 ; x3 ) = x31 x2 ; p3 (x1 ; x2 ; x3 ) = x1 x2 x3 :
Each of these is a multinomial in 3 variables, however p1 and p2 have order
four, whereas the order of p3 is three. Thus only p1 and p2 are in P3[4] .
Similarly of
p4 (x1 ; x2 ; x3 ) = x41 + x2 x33 and p5 (x1 ; x2 ; x3 ) = x21 x2 x3 + x1
only p4 is in P3[4], whereas p5 is not in any P3[n] space since its terms are not
homogeneous. Some thought will convince you that Pm[n] is a vector space
under pointwise addition.


1.1.2 Subspaces
A subspace of a vector space V is a subset of V which is also a vector space
with respect to the same eld and operations; equivalently, it is a subset
which is closed under the operations on V .
Examples:
A vector space can have many subspaces, and the simplest of these is the
zero subspace, denoted by f0g. This is a subspace of any vector space and
contains only the zero element. Excepting the zero subspace and the entire
space, the simplest type of subspace in V is of the form
Sv = fs 2 V : s = v; for some  2 Rg;
given v in V . That is each element in V generates a subspace by multiplying
it by all possible scalars. In R2 or R3 , such subspaces correspond to lines
going through the origin.
Going back to our earlier examples of vector spaces we see that the
multinomials Pm[n] are subspaces of F (Rm ; R), for any n.
Now Rn has many subspaces and an important set is those associated
with the natural insertion of Rm into Rn , when m < n. Elements of these
subspaces are of the form
 
x = x0
22 1. Preliminaries in Finite Dimensional Space
where x 2 Rm and 0 2 Rn;m . 
Given two subspaces S1 and S2 we can dene the addition
S1 + S2 = fs 2 V : s = s1 + s2 for some s1 2 S1 and s2 2 S2 g
which is easily veried to be a subspace.

1.1.3 Bases, spans, and linear independence

We now dene some key vector space concepts. Given elements v1 ; : : : ; vm
in a vector space we denote their span by
spanfv1 ; : : : ; vm g;
which is the set of all vectors v that can be written as
v = 1 v1 +


+ m vm
for some scalars k 2 F; the above expression is called a linear combination
of the vectors v1 ; : : : ; vm . It is straightforward to verify that the span always
denes a subspace. If for some vectors we have
spanfv1 ; : : : ; vm g = V ;
we say that the vector space V is nite dimensional. If no such nite set of
vectors exists we say the vector space is innite dimensional. Our focus for
the remainder of the chapter is exclusively nite dimensional vector spaces.
We will pursue the study of some innite dimensional spaces in Chapter 3.
If a vector space V is nite dimensional we dene its dimension, denoted
dim(V ), to be the smallest number n such that there exist vectors v1 ; : : : ; vn
satisfying
spanfv1 ; : : : ; vn g = V :
In that case we say that the set
fv1 ; : : : ; vn g is a basis for V :
Notice that a basis will automatically satisfy the linear independence
property, which means that the only solution to the equation
1 v1 +


+ n vn = 0
is 1 =


= n = 0. Otherwise, one of the vi 's could be expressed as a
linear combination of the others and V would be spanned by fewer than n
vectors. Given this observation, it follows easily that for a given v 2 V , the
scalars (1 ; : : : ; n ) satisfying
1 v1 +


+ n vn = v
are unique; they are termed the coordinates of v in the basis fv1 ; : : : ; vn g.
Linear independence is dened analogously for any set of vectors
fv1 ; : : : ; vm g; it is equivalent to saying the vectors are a basis for their
 1.1. Linear spaces and mappings 23
span. The maximal number of linearly independent vectors is n, the di-
mension of the space; in fact any linearly independent set can be extended
with additional vectors to form a basis.
Examples:
From our examples so far Rn ; C mn and Pm[n] are all nite dimensional
vector spaces; however F (Rm ; Rn ) is innite dimensional. As the reader
may already be aware, the real vector space Rn and complex vector space
C mn are m and mn dimensional, respectively. The dimension of Pm is
[n]
more challenging to compute and its determination is an exercise at the
end of the chapter.
An important computational concept in vector space analysis is associ-
ating a general k dimensional vector space V with the vector space Fk . This
is done by taking a basis fv1 ; : : : ; vk g for V , and associating each vector v
in V with the vector of coordinates in the given basis,
2 3
1
6 .. 75 2 Fk :
4 .
k
Equivalently, each vector vi in the basis is associated with the vector
2 3
0
6 .. 7
6.7
6 7
607
6 7
ei = 66177 2 Fk :
607
6 7
6.7
4 .. 5
0
That is ei is the vector with zeros everywhere excepts its ith entry which
is one. Thus we are identifying the basis fv1 ; : : : ; vk g in V with the set
fe1; : : : ; ek g which is in fact a basis of Fk , called the canonical basis.
To see how this type of identication is made, suppose we are dealing
with Rnm , which has dimension k = nm. Then a basis for this vector
space is
2 3
0


0
Eij = 64 ... . . . 1 ... 75 ;
0


0
which are the matrices that are zero everywhere but their i; j th-entry which
is one. Then we identify each of these with the vector en(j;1)+i 2 Rk . Thus
addition or scalar multiplication on Rnm can be translated to equivalent
operations on Rk . 
24 1. Preliminaries in Finite Dimensional Space
1.1.4 Mappings and matrix representations
We are now ready to introduce the important concept of a linear mapping
between vector spaces. The mapping A : V ! W is linear if
A(v1 + v2 ) = Av1 + Av2
for all v1 ; v2 in V , and all scalars 1 and 2 . Here V and W are vector
spaces with the same associated eld F. The space V is called the domain
of the mapping, and W its codomain.
Given bases fv1 ; : : : ; vn g and fw1 ; : : : ; wm g for V and W respectively,
we associate scalars akj with the mapping A, dening them such that they
satisfy
Avk = a1k w1 + a2k w2 +


+ amk wm ;
for each 1  k  n. Namely given any basis vector vk , the coecients ajk
are the coordinates of Avk in the chosen basis for W . It turns out that these
mn numbers ajk completely specify the linear mapping A. To see this is true
consider any vector v 2 V , and let w = Av. We can express both vectors in
their respective bases as v = 1 v1 +


+ n vn and w = 1 w1 +


+ mwm .
Now we have

w = Av = A(1 v1 +


+ n vn )

= 1 Av1 +


+ n Avn
!
n X
X m m
X n
X
= k ajk wj = k ajk wj ;
k=1 j =1 j =1 k=1
and therefore by uniqueness of the coordinates we must have
m
X
j = k ajk ; j = 1; : : : ; m:
j =1
To express this relationship in a more convenient form, we can write the
set of numbers ajk as the m  n matrix
2 3
a11


a1n
[A] = 64 ... ... ..
.
7
5 :
am1


amn
Then via the standard matrix product we have
2 3 2 32 3
1 a11


a1n 1
6 .. 75 = 64 .. ... .. 76 .. 75 :
4 . . . 54 .
n am1


anm n
 1.1. Linear spaces and mappings 25
In summary any linear mapping A between vector spaces can be regarded
as a matrix [A] mapping Fn to Fm via matrix multiplication.
Notice that the numbers akj depend intimately on the bases fv1 ; : : : ; vn g
and fw1 ; : : : ; wm g. Frequently we use only one basis for V and one for W
and thus there is no need to distinguish between the map A and the basis
dependent matrix [A]. We therefore simply write A to denote either the
map or the matrix, making which is meant context dependent.
We now give two examples to more clearly illustrate the above discussion.

Examples:
Given matrices B 2 C kk and D 2 C ll we dene the map : C kl !
C kl by

(X ) = BX ; XD;

where the right hand-side is in terms of matrix addition and multiplication.

Clearly is a linear mapping since

(X1 + X2 ) = B (X1 + X2 ) ; (X1 + X2 )D

= (BX1 ; X1 D) +  (BX2 ; X2 D)

=  (X1 ) +  (X2 ):

If we now consider the identication between the matrix space C kl and
the product space C kl , then can be thought of as a map from C kl to C kl ,
and can accordingly be represented by a complex matrix which is kl  kl.
We now do an explicit 2  2 example for illustration. Suppose k = l = 2
and that
   
B = 13 24 and D = 50 00 :

We would like to nd a matrix representation for . Since the domain and

codomain of are equal, we will use the standard basis for C 22 for each.
26 1. Preliminaries in Finite Dimensional Space
This basis is given by the matrices Eij dened earlier. We have
 
(E11 ) = ;34 00 = ;4E11 + 3E21 ;
 
(E12 ) = 00 13 = E12 + 3E22 ;
 
(E21 ) = ;21 00 = 2E11 ; E21 ;
 
(E22 ) = 00 24 = 2E12 + 4E22 :
Now we identify the basis fE11 ; E12 ; E21 ; E22 g with the standard basis for
C 4 given by fe1 ; e2 ; e3 ; e4 g. Therefore we get that
2 3
;4 0 2 0
[ ] = 64 03 10 ;01 2075
6 7

0 3 0 4
in this basis.
Another linear operator involves the multinomial function Pm[n] dened
earlier in this section. Given an element a 2 Pm[k] we can dene the mapping
 : Pm[n] ! Pm[n+k] by function multiplication
(p)(x1 ; x2 ; : : : ; xm ) := a(x1 ; x2 ; : : : ; xm )p(x1 ; x2 ; : : : ; xm ):
Again  can be regarded as a matrix, which maps Rd1 ! Rd2 , where d1
and d2 are the dimensions of Pm[n] and Pm[n+k] respectively. 
Associated with any linear map A : V ! W is its image space, which is
dened by
ImA = fw 2 W : there exists v 2 V satisfying Av = wg:
This set contains all the elements of W which are the image of some point
in V . Clearly if fv1 ; : : : ; vn g is a basis for V then
ImA = spanfAv1 ; : : : ; Avn g
and is thus a subspace. The map A is called surjective when ImA = W .
The dimension of the image space is called the rank of the linear mapping
A, and the concept is applied as well to the associated matrix [A]. Namely
rank[A] = dim(ImA):
If S is a subspace of V , then the image of S under the mapping A is denoted
AS . That is
AS = fw 2 W : there exists s 2 S satisfying As = wg:
 1.1. Linear spaces and mappings 27
In particular, this means that AV = ImA.
Another important set related to A is its kernel, or null space, dened
by
ker A = fv 2 V : Av = 0g:
In words, ker A is the set of vectors in V which get mapped by A to the
zero element in W , and is easily veried to be a subspace of V .
If we consider the equation Av = w, suppose va and vb are both solutions;
then
A(va ; vb ) = 0:
Namely the dierence between any two solutions is in the kernel of A. Thus
given any solution va to the equation, all solutions are parametrized by
va + v0 ;
where v0 is any element in ker A.
In particular, when ker A is the zero subspace, there is at most a unique
solution to the equation Av = w. This means Ava = Avb only when va = vb ;
a mapping with this property is called injective.
In summary, a solution to the equation Av = w will exist if and only if
w 2 ImA; it will be unique only when ker A is the zero subspace.
The dimensions of the image and kernel of A are linked by the
relationship
dim(V ) = dim(ImA) + dim(ker A);
proved in the exercises at the end of the chapter.
A mapping is called bijective when it is both injective and surjective, i.e.
for every w 2 W there exists a unique v satisfying Av = w. In this case
there is a well dened inverse mapping A;1 : W ! V , such that
A;1 A = IV ; AA;1 = IW :
In the above, I denotes the identity mapping in each space, i.e. the map
that leaves elements unchanged. For instance, IV : v 7! v for every v 2 V .
From the above property on dimensions we see that if there exists a
bijective linear mapping between two spaces V and W , then the spaces
must have the same dimension. Also, if a mapping A is from V back to
itself, namely A : V ! V , then one of the two properties (injectivity or
surjectivity) suces to guarantee the other.
We will also use the terms nonsingular or invertible to describe bijective
mappings, and apply these terms as well to their associated matrices. Notice
that invertibility of the mapping A is equivalent to invertibility of [A] in
terms of the standard matrix product; this holds true regardless of the
chosen bases.
28 1. Preliminaries in Finite Dimensional Space
Examples:
To illustrate these notions let us return to the mappings and  dened
above. For the 2  2 numerical example given, maps C 22 back to itself.
It is easily checked that it is invertible by showing either
Im = C 22 ; or equivalently ker = 0:
In contrast  is not a map on the same space, instead taking Pm[n] to the
larger space Pm[n+k] . And we see that the dimension of the image of  is at
most n, and the dimension of its kernel at least k. Thus assuming k > 0
there are at least some elements w 2 Pm[n+k] for which
v = w
cannot be solved. These are exactly the values of w that are not in Im.


1.1.5 Change of basis and invariance

We have already discussed the idea of choosing a basis fv1 ; : : : ; vn g for the
vector space V , and then associating every vector x in V with its coordinates
2 3
1
xv = 64 ... 75 2 Fn ;
n
which are the unique scalars satisfying x = 1 v1 +


+ n vn . This raises
the question, suppose we choose another basis u1; : : : ; un for V , how can we
eectively move between these basis representations? That is given x 2 V ,
how are the coordinate vectors xv ; xu 2 Fn related?
The answer is as follows. Suppose that each basis vector uk is expressed
by
uk = t1k v1 +


+ tnk vn ;
in the basis fv1 ; : : : ; vn g. Then the coecients tik dene the matrix
2 3
t11


t1n
T = 64 ... . . . ... 75 :
tn1


tnn
Notice that such a matrix is nonsingular, since it represents the iden-
tity mapping IV in the bases fv1 ; : : : ; vn g and fu1; : : : ; un g. Then the
relationship between the two coordinate vectors is
Txu = xv :
 1.1. Linear spaces and mappings 29
Now suppose A : V ! V and that Av : Fn ! Fn is the representation
of A on the basis v1 ; : : : ; vn , and Au is the representation of A using the
basis u1; : : : ; un. How is Au related to Av ?
To study this, take any x 2 V and let xv , xu be its coordinates in the
respective bases, and zv , zu be the coordinates of Ax. Then we have
zu = T ;1 zv = T ;1 Av xv = T ;1Av Txu :
Since the above identity and
zu = Au xu
both hold for every xu , we conclude that
Au = T ;1Av T:
The above relationship is called a similarity transformation. This discussion
can be summarized in the following commutative diagram. Let E : V ! Fn
be the map that takes elements of V to their representation in Fn with
respect to the basis fv1 ; : : : ; vn g. Then
E T; 1

V Fn Fn

A Av Au = T ;1Av T

E T;
V
1

Fn Fn

Next we examine mappings when viewed with respect to a subspace. Sup-

pose that S  V is a k dimensional subspace of V , and that v1 ; : : : ; vn is a
basis for V with
spanfv1 ; : : : ; vk g = S :
That is the rst k vectors of this basis forms a basis for S . If E : V ! Fn
is the associated map which maps the basis vectors in V to the standard
basis on Fn , then
E S = Fk  f0g  Fn :
Thus in Fn we can view S as the elements of the form
 
x where x 2 Fk :
0
When a basis has this property we say it is a canonical basis for the subspace
S . From the point of view of a linear mapping A : V ! V this partitioning
of Fn gives a useful decomposition of the corresponding matrix [A]. Namely
30 1. Preliminaries in Finite Dimensional Space
we can regard [A] as
 
[A] = A 1 A2
A A ;
3 4
where A1 : Fk ! Fk ; A2 : Fn;k ! Fk ; A3 : Fk ! Fn;k , and A4 : Fn;k !
Fn;k . We have that
 
EAS = Im A 1
A3 :
Finally to end this section we have the notion of invariance of a subspace
to a mapping. We say that a subspace S  V is A-invariant if A : V ! V
and
AS  S :
Clearly every map has at least two invariant subspaces, the zero subspace
and entire domain V . For subspaces S of intermediate dimension, the in-
variance property is expressed most clearly in a canonical basis for the
subspace. When S is A-invariant, the partitioning of [A] as above yields a
matrix of the form
 
[A] = A01 A2
A3 :
Similarly if a matrix has this form the subspace Fk  f0g is [A]-invariant.
We will revisit the question of nding non-trivial invariant sub-
spaces later in the chapter, when studying eigenvectors and the Jordan
decomposition.
In the next section we will pursue some of the geometric properties of
sets in linear spaces.

1.2 Subsets and Convexity

Up to now in this chapter, the only sets we have encountered are those
that are also vector spaces. In this section we will consider more general
subsets of a vector space V , with an emphasis of introducing and examining
convexity. Convexity is one of the main geometric ideas underlying much
of global optimization theory, and in particular is the foundation for a
major analytical tool used in this course. The results presented here are
for work in nite dimensional vector spaces, but in many cases they have
innite dimensional versions as well. Throughout this section we have the
standing assumption that V is a nite dimensional real vector space. Before
we address convexity it will be useful to introduce some topological aspects
of vector spaces.
 1.2. Subsets and Convexity 31
1.2.1 Some basic topology
To start we seek to dene the notion of a neighborhood of a point in the
vector space V . We do this by rst dening the unit ball with respect to
a basis. Suppose that fu1; : : : ; ung is a basis for the vector space V . Then
the open unit ball in this basis is dened by
B(u1; : : : ; un ) = f1 u1 +


+ n un 2 V : i 2 R; 21 +


+ 2n < 1g:
This set contains all the points that can be expressed, in the basis, with
the vector of coecients  inside the unit sphere of Fn . Clearly this set is
basis-dependent. In particular, since ftu1; : : : ; tun g is also a basis for every
t 6= 0, then given any nonzero element v in V , there always exists a basis
such that v is in the corresponding unit ball, and another basis such that v
is not in the associated unit ball. The zero vector is the only element that
belongs to every unit ball.
We now dene the notion a neighborhood of a point, which intuitively
means any set that totally surrounds the given point in the vector space.
Denition 1.2. A subset N (0) of the vector space V is a neighborhood of
the zero element if there exists a basis u1 ; : : : ; un for V such that
B(u1; : : : ; un )  N (0):
Further, a subset N (w)  V is a neighborhood of the point w 2 V if the set
N = fv 2 V : v + w 2 N (w)g
is a neighborhood of the zero element.
This says that a set is a neighborhood of zero provided that one of its
subsets is the unit ball in some basis. A neighborhood of a general point w 2
V is any set that is equal to a neighborhood of zero that has been translated
by w. Notice that while a basis was used in the denition, since its choice
is allowed to be arbitrary the resulting denition of a neighborhood is in
fact basis-independent.
If Q is a subset of V , we say that it is open if
for every v 2 Q, there exists a neighborhood of v which is a subset of Q.
We denote the complement of the set Q by Qc, which we recall is dened
by
Qc = fv 2 V : v 62 Qg:
A set Q is a closed set if its complement is open.
Next we dene the closure of a set. The closure of Q, denoted Q , is
Q = fv 2 V : for every neighborhood N (v) of v, N (v) \ Q is nonemptyg:
So the closure contains all points that are arbitrarily near to Q. Clearly,
Q  Q. The closure is always a closed set, and is the smallest closed set
that contains Q; these facts are left as exercises.
32 1. Preliminaries in Finite Dimensional Space
One of the major objectives of this section is to develop tests for de-
termining when two subsets of V do not intersect. In addition to being
disjoint, we would like to have a notion of when disjoint sets are not ar-
bitrarily close to each other. We call this property strict separation of the
sets. We say that two subsets Q1 , Q2  V are strictly separated if there
exists a neighborhood N (0) of zero, such that for all v1 2 Q1 and v2 2 Q2
the condition
v1 ; v2 62 N (0) holds.
Clearly N (0) can be chosen such that v2 ; v1 62 N (0) simultaneously holds,
so this denition is symmetric with respect to the sets. Therefore sets are
strictly separated if the dierence between any two elements chosen from
each set is uniformly outside a neighborhood of the origin.
The condition of strict separation is not always easy to check, and so
we present an alternative condition for a special case. First we say that a
subset Q 2 V is bounded if there exists a basis u1 ; : : : ; un for V such that
Q  B(u1; : : : ; un ):
A set that is both closed and bounded is called compact.
We can now state the following property.
Proposition 1.3. Suppose Q1 and Q2 are subsets of V , and that Q1
is bounded. Then Q1 and Q2 are strictly separated if and only if the
intersection of their closures Q 1 \ Q 2 is empty.
The result states that provided that one set is bounded, two sets are strictly
separated if there closures are disjoint. Thus if two sets are closed, with one
of them compact, they are strictly separated exactly when they are disjoint.
Having introduced some topology for vector spaces, we are ready to move
on to discussing convexity.

1.2.2 Convex sets

Let us begin by dening the line segment that joins two points in V . Suppose
that v1 and v2 are in V , then we dene the line segment L(v1 ; v2 ) between
them as the set of points
L(v1 ; v2 ) = fv 2 V : v = v1 + (1 ; )v2 for some  2 [0; 1]g:
Clearly the end points of the line segment are v1 and v2 , which occur in
the parametrization when  = 1 and  = 0, respectively.
We can now turn to the idea of convexity. Suppose that Q is a nonempty
subset of the vector space V . Then Q is dened to be a convex set if
for any v1 ; v2 2 Q; the line segment L(v1 ; v2 ) is a subset of Q:
That is Q is convex if it contains all the line segments between its points.
Geometrically we have the intuitive picture shown in Figure 1.1.
 1.2. Subsets and Convexity 33

v1 v1 v2
v2

Figure 1.1. Convex and nonconvex sets

Clearly any vector space is convex, as is any subset fvg of a vector space
containing only a single element.
We can think of the expression v1 + (1 ; )v2 for a point on the line
L(v1 ; v2 ) as a weighted average. To see this instead write equivalently
v = 1 v1 + 2 v2 ;
where 1 ; 2 2 [0; 1] and satisfy 1 +2 = 1. Then we see that if both 1 and
2 are both equal to one half we have our usual notion of the average. And
if they take other values the weighted average \favors" one of the points.
Thus the clear generalization of such an average to n points v1 ; : : : ; vn is
v = 1 v1 +


+ n vn ;
where 1 +


+ n = 1 and 1 ; : : : ; n 2 [0; 1]. A line segment gave us
geometrically a point on the line between the two endpoints. The general-
ization of this to an average of n points, yields a point inside the perimeter
dened by the points v1 ; : : : ; vn . This is illustrated in Figure 1.2.
v co(fv1 ; : : : ; v6 g)
v1 2

v
v5 v6 v3

v4
Figure 1.2. Convex hull of nite number of points
Given v1 ; : : : ; vn we dene the convex hull of these points by
n
X n
X
co(fv1 ; : : : ; vn g) = fv 2 V : v = k vk ; with k 2 [0; 1]; k = 1g:
k=1 k=1
Thus this set is all the points inside the perimeter in Figure 1.2. In words
the convex hull of the points v1 ; : : : ; vn is simply the set comprised of all
weighted averages of these points. In particular we have that for two points
L(v1 ; v2 ) = co(fv1 ; v2 g). It is a straightforward exercise to show that if
34 1. Preliminaries in Finite Dimensional Space
Q is convex, then it necessarily contains any convex hull formed from a
collection of its points.
So far we have only dened the convex hull in terms of a nite number
of points. We now generalize this to an arbitrary set. Given a set Q, we
dene its convex hull co(Q) by
co(Q) =fv 2 V : there exists n and v1 ; : : : ; vn 2 Q
such that v 2 co(fv1 ; : : : ; vn g)g:
So the convex hull of Q is the collection of all possible weighted averages
of points in Q. It is straightforward to demonstrate that for any set Q:
 the subset condition Q  co(Q) is satised;
 the convex hull co(Q) is convex;
 the relationship co(Q) = co (co(Q)) holds.
We also have the following results which relates convexity of a set to its
convex hull.
Proposition 1.4. A set Q is convex if and only if co(Q) = Q is satised.
Notice that, by denition, the intersection of convex sets is always convex;
therefore, given a set Q, there exists a smallest convex set that contains Q;
it follows easily that this is precisely co(Q); in other words, if Y is convex
and Q  Y , then necessarily co(Q) is a subset of Y . Pictorially we have
Figure 1.3 to visualize Q and its convex hull.
Q co(Q)

Figure 1.3. Convex and nonconvex sets

A linear mapping F : V ! R is called a (linear) functional. If it is not
identical to zero (a standing assumption from now on), then it is always
surjective; namely for xed a 2 R the equation
F (v) = a
always has a solution in the variable v 2 V . Also if v1 satises F (v1 ) = a,
all solutions to the equation are given by
v = v1 + v0 ; where v0 2 ker F:
Thus we can view this set of solutions as the kernel subspace shifted away
from the origin. Also, the dimension of ker F is dim(V ) ; 1. A set in an
 1.2. Subsets and Convexity 35
n-dimensional space, obtained by shifting an n ; 1 dimensional subspace
away from the origin, is called a hyperplane. It is not dicult to show
that every hyperplane can be generated by a linear functional F and a real
number a as described above. A hyperplane in R3 is depicted in Figure 1.4.

v1
Figure 1.4. A hyperplane
An important property of a hyperplane, which is clear in the above ge-
ometric case, is that it always breaks up the space into two half-spaces:
these have the form fv : F (v)  ag and fv : F (v)  ag.
This leads to the notion of separating two sets with a hyperplane. Given
two sets Q1 and Q2 in V , we say that the hyperplane dened by (F; a)
separates the sets if
(a) F (v1 )  a, for all v1 2 Q1 ;
(b) F (v2 )  a, for all v2 2 Q2 .
Geometrically we have the illustration in Figure 1.5 below.

Q1

Q2

Figure 1.5. Separating hyperplane

Further we say that they are strictly separated if (b) is changed to
F (v2 )  a + ; for all v2 2 Q2 ;
36 1. Preliminaries in Finite Dimensional Space
for some xed  > 0. Such a separating hyperplane may not always exist:
in the gure if we move the sets suciently close together it will not be
possible to nd any hyperplane which separates the sets. However we have
the following major result, which says that if two sets are convex and
disjoint, then there always exists a separating hyperplane between them.
Theorem 1.5. Suppose Q1 and Q2 are two non-empty convex subsets of
the vector space V .
(a) If the intersection Q1 \ Q2 is empty, then there exists a hyperplane
which separates Q1 from Q2 ;
(b) There exists a strictly separating hyperplane between the sets if and
only if the sets are strictly separated.
This is a very powerful theorem which we will make critical use of later in
the course. While we defer the proof to the references, readers should gain
condence of its validity by drawing some geometrical pictures.
At this point we have developed the notions of convex sets and separation
in an abstract setting. We next consider two explicit examples which will
be key to us later.
Examples:
Here we will consider representing hyperplanes for two explicit vector
spaces, the spaces Rn and the symmetric matrices Sn. Let us rst consider
Rn .
Given a linear functional F : Rn ! R we see that it is completely dened
by the numbers
F (e1 ); : : : ; F (en )
where fe1 ; : : : ; en g is the canonical basis for Rn . Thus given any vector
x 2 Rn we have
F (x) = x1 F (e1 ) +


+ xn F (en )
Conversely given any y 2 Rn a linear functional F is dened on Rn by
F (x) = y1 x1 + : : : + yn xn , which we express more concisely using matrix
multiplication by
F (x) = y x:
Thus we see that any hyperplane in Rn is characterized by the equation
y x = a;
for some y 2 Rn and a 2 R.
We now wish to generalize this to the real square matrices Rnn . To do
this we rst introduce the trace of a matrix X 2 Rnn . The trace operation
is dened by
TrX = x11 +


+ xnn ;
 1.2. Subsets and Convexity 37
the sum of the diagonal entries. Therefore Tr : Rnn ! R. Also for X ,
Z 2 Rnn and ,  2 R we have
n
X n
X
Tr(X + Z ) = xkk + zkk = TrX +  TrZ;
k=1 k=1
and so the trace operation denes a particular linear functional on Rnn .
Now given an element Y 2 Rnn it is also routine to show that the mapping
F : Rnn ! R dened by
n
X
F (X ) = Tr(Y  X ) = xkj ykj (1.1)
k;j =1
is a linear functional. The last identity is a consequence of the denitions
of trace and matrix product.
The question we now ask is whether every linear functional on Rnn is
dened in this way? The answer is yes and can be seen by expanding X in
the standard basis of Rnn , as
n
X
X= xkj Ekj :
k;j =1
Given a linear functional F , we have
n
X
F (X ) = xkj F (Ekj ) = Tr(Y  X );
k;j =1
where we have dened the matrix Y by ykj = F (Ekj ). Similarly it is
straightforward to show that all linear functionals F on the symmetric
matrices Sn are of the form in (1.1) where Y 2 Sn, and an analogous
situation occurs over the space H n .
Later in the course we will specically require separating hyperplanes
for sets constructed from Cartesian products of the above spaces. We
summarize what we will require in the following result.
Proposition 1.6. Suppose that the vector space V is given by the
Cartesian product
V = Rn1 


 Rns  H ns+1 


 H ns+f :
Then F is a linear functional on V if and only if there exists
Y = (y1 ; : : : ; yns ; Yns+1 ; : : : ; Yns+f ) 2 V such that
F (V ) = y1 v1 +


+ yn s vns + Tr(Yns+1 Vns+1 ) +


+ Tr(Yns+f Vns+f );
for all V = (v1 ; : : : ; vns ; Vns+1 ; : : : ; Vns+f ) 2 V.

As a nal point in this section, we introduce the notion of cones in vector
space.
38 1. Preliminaries in Finite Dimensional Space
A set Q  V is called a cone if it is closed under positive scalar
multiplication, i.e. if
v 2 Q implies tv 2 Q for every t > 0:
Clearly subspaces are cones, but the latter denition is broader since it
includes, for example, the half-line Cv = fv :  > 0g for a xed vector v.
Of particular interest are cones which have the convexity property. In
fact convex cones are precisely those which are closed under addition (i.e.
v1 ; v2 2 C implies v1 + v2 2 C ). A canonical example of a convex cone is a
half-space
F (v)  0
that goes through the origin, or an intersection of half-spaces of this form.
More illustrations are given in Figure 1.6.

Figure 1.6. Convex and nonconvex cones

Up to this point we have introduced a number of concepts related to
vector spaces and have for the most part developed our results in a basis
independent way, emphasizing the structure of linear spaces. This provides
a powerful way to visualize linear algebra. However for most parts of the
course it is much more ecient to work in a particular basis, especially
since we are interested in developing results with computation in mind. In
this spirit we now turn to results about matrices.

1.3 Matrix Theory

The material of this section is aimed directly at both analysis and com-
putation. Our goals will be to review some basic facts about matrices,
and present some additional results for later reference, including two ma-
trix decompositions which have tremendous application, the Jordan form
and singular value decomposition. Both are extremely useful for analyti-
cal purposes, and the singular value decomposition is also very important
 1.3. Matrix Theory 39
in computations. We will also present some results about self-adjoint and
positive denite matrices.

1.3.1 Eigenvalues and Jordan form

In this section we are concerned exclusively with complex square matrices.
We begin with a denition: if A 2 C nn , we say that  2 C is an eigenvalue
of A if
Ax = x (1.2)
can be satised for some nonzero vector x in C n . Such a vector x is called
an eigenvector. Equivalently this means that ker(A ; I ) 6= 0 or A ; I
is singular. A matrix is singular exactly when its determinant is zero, and
therefore we have that  is an eigenvalue if and only if
det(A ; I ) = 0;
where det(
) denotes determinant. Regarding  as a variable we call the
polynomial
det(A ; I ) = (;1)n n + an;1 n;1 +


+ a0
the characteristic polynomial of A. If A is a real matrix then the coecients
ak will be real as well. The characteristic polynomial can be factored as
det(A ; I ) = (1 ; )


(n ; ):
The n complex roots k , which need not be distinct, are the eigenvalues of
A. Furthermore if A is a real matrix, then any non real eigenvalues must
appear in conjugate pairs. Also, a matrix has the eigenvalue zero if and
only if it is singular.
Associated with every eigenvalue k is the subspace
Ek = ker(A ; k I );
every nonzero element in Ek is an eigenvector corresponding to the
eigenvalue k . Now suppose that a set of eigenvectors satises
spanfx1 ; : : : ; xn g = C n :
 
Then we can dene the invertible matrix X = x1


xn , and from
the matrix product we nd
   
AX = Ax1


Axn = 1 x1


n xn = X 
where  is the diagonal matrix
2 3
1 0
 = 64 ... 7
5:
0 n
40 1. Preliminaries in Finite Dimensional Space
Thus in this case we have a similarity transformation X such that
X ;1AX =  is diagonal, and we say that the matrix A is diagonalizable.
Summarizing we have the following result.
Proposition 1.7. A matrix A is diagonalizable if and only if
E1 + E2 +


+ En = C n holds.
The following example shows that not all matrices can be diagonalized.
Consider the 2  2 matrix
 
0 1 :
0 0
It has a repeated eigenvalue at zero, but only one linearly independent
eigenvector. Thus it cannot be diagonalized. Matrices of this form have a
special role in the decomposition we are about to introduce: dene the n  n
matrix N by
2 3
0 1 0
6 . . . 77
N = 664 7;
15
0 0
where N = 0 if the dimension n = 1. Such matrices are called nilpotent
because N n = 0. Using these we dene a matrix to be a Jordan block if it
is of the form
2 3
 1 0
6 . . . 77
J = I + N = 664 7:
15
0 
Notice all scalars are 1  1 Jordan blocks. A Jordan block has one eigenvalue
 of multiplicity n. However it has only one linearly independent eigenvec-
tor. A key feature of a Jordan block is that it has precisely n J -invariant
subspaces. They are given by
C k  f0g;
for 1  k  n. When k = 1 this corresponds exactly to the subspace
associated with its eigenvector. We can now state the Jordan decomposition
theorem.
Theorem 1.8. Suppose A 2 C nn . Then there exists a nonsingular matrix
T 2 C nn , and an integer 1  p  n, such that
2 3
J1 0
6 J2 7
TAT ;1 = 664 ...
7
7
5
;
0 Jp
 1.3. Matrix Theory 41
where the matrices Jk are Jordan blocks.
This theorem states that a matrix can be transformed to one that is block-
diagonal, where each of the diagonal matrices is a Jordan block. Clearly
if a matrix is diagonalizable each Jordan block Jk will simply be a scalar
equal to an eigenvalue of A. In general each block Jk has a single eigenvalue
of A in all its diagonal entries; however a given eigenvalue of A may occur
in several blocks. If the dimensions of Jk is nk  nk , then from our earlier
discussion there are exactly nk invariant subspaces of A associated with Jk .
All invariant subspaces can be constructed from this collection associated
with the Jordan blocks.
We will not explicitly require a constructive method for transforming a
matrix to Jordan form, and will use this result solely for analysis.

1.3.2 Self-adjoint, unitary and positive denite matrices

We have already introduced the adjoint A of a complex matrix A; in this
section we study in more detail the structure given to the space of matrices
by this operation. A rst observation, which will be used extensively below
is that
(AB ) = B  A
for matrices A, B of compatible dimensions; this follows directly by
denition.
Another basic concept closely related to the adjoint is the Euclidean
length of a vector x 2 C n , dened by
p
jxj = x x
This extends the usual denition of magnitude of a complex number, so
our notation will not cause any ambiguity. In particular,
n
X
jxj2 = x x = jxi j2 :
i=1
Clearly jxj is never negative, and is zero only when the vector x = 0. Later
in the course we will discuss generalizations of this concept in more general
vector spaces.
We have already encountered the notion of a Hermitian matrix, charac-
terized by the self-adjoint property Q = Q. Recall the notation H n for the
(real) vector space of complex Hermitian matrices. We now collect some
properties and introduce some new denitions, for later use. Everything we
will state will apply as well to the set Sn of real, symmetric matrices.
Our rst result about self-adjoint matrices is that their eigenvalues are
always real. Suppose Ax = x for nonzero x. Then we have
x x = x Ax = (Ax) x =  x x:
42 1. Preliminaries in Finite Dimensional Space
Since x x > 0 we conclude that  =  .
We say that two vectors x,y 2 C n are orthogonal if
y x = 0:
Given a set of vectors fv1 ; : : : ; vk g in C n we say the vectors are orthonormal
if

vi vj = 10;; ifif ii = j;
6= j .
The vectors are orthonormal if each has unit length and is orthogonal
to all the others. It easy to show that orthonormal vectors are linearly
independent, so such a set can have at most n members. If k < n, then
it is always possible to nd a vector vk+1 such that fv1 ; : : : ; vk+1 g is an
orthonormal set. To see this, form the k  n matrix
2 3
v1
Vk = 64 ... 75 :
vk
The kernel of Vk has the nonzero dimension n ; k, and therefore any
element of the kernel is orthogonal to the vectors fv1 ; : : : ; vk g. We conclude
that any element of unit length in ker Vk is a suitable candidate for vk+1 .
Applying this procedure repeatedly we can generate an orthonormal basis
fv1 ; : : : ; vn g for C n .
A square matrix U 2 C nn is called unitary if it satises
U U = I:
If U 2 Rnn and satises the above identity it is more specically called
an orthogonal matrix. From this denition we see that the columns of any
unitary matrix forms an orthonormal basis for C n . Further, since U is
square it must be that U  = U ;1 and therefore UU  = I . So the columns
of U  also form an orthonormal basis. A key property of unitary matrices
is that if y = Ux, for some x 2 C n , then the length of y is equal to that of
x:
p
jyj = py y = (Ux) (Ux) = x U Ux = jxj:
p

Unitary matrices are the only matrices that leave the length of every vector
unchanged. We are now ready to state the spectral theorem for Hermitian
matrices.
Theorem 1.9. Suppose H is a matrix in H n . Then there exist a unitary
matrix U and a real diagonal matrix  such that
H = U U  :
Furthermore, if H is in Sn then U can be chosen to be orthogonal.
 1.3. Matrix Theory 43
Notice that since U  = U ;1 for a unitary U , the above expression is a
similarity transformation. Therefore the theorem says that a self-adjoint
matrix can be diagonalized by a unitary similarity transformation. Thus
the columns of U are all eigenvectors of H . Since the proof of this result
assembles a number of concepts from this chapter we provide it below.
Proof . We will use an induction argument. Clearly the result is true if H
is simply a scalar, and it is therefore sucient to show that if the result
holds for matrices in H n;1 then it holds for H 2 H n . We proceed with
the assumption that the decomposition result holds for (n ; 1)  (n ; 1)
Hermitian matrices.
The matrix H has at least one eigenvalue 1 , and 1 is real since H is
Hermitian. Let x1 be an eigenvector associated with this eigenvalue, and
without loss of generality we assume it to have length one. Dene X to be
any unitary matrix with x1 as its rst column, namely
X = [x1


xn ]:
Now consider the product X AX . Its rst column is given by X  Ax1 =
1 X  x1 = 1 e1 , where e1 is the rst element of the canonical basis. Its
rst row is described by x1 AX which is equal to 1 x1 X = 1 e1 , since
x1 A = 1 x1 because A is self-adjoint. Thus we have
 

X AX = 0 A ;  1 0
2
where A2 a Hermitian matrix in H n;1 . By the inductive hypothesis there
exists a unitary matrix X2 in C (n;1)(n;1) such that A2 = X2 2 X2, where
2 is both diagonal and real. We conclude that
       
A= X 0 X I 0  1 0 I 0 
2 0 2 0 X2 X :
The right-hand side gives the desired decomposition.
If H is a real matrix, that is in H n , then all the matrices in the
construction above are also real, proving the latter part of the theorem. 
We remark in addition that the (real) eigenvalues of H can be arranged
in decreasing order in the diagonal of . This follows directly from the
above induction argument: just take 1 to be the largest eigenvalue.
We now focus on the case where these eigenvalues have a denite sign.
Given Q 2 H n , we say it is positive denite, denoted Q > 0, if
x Qx > 0;
for all nonzero x 2 C n . Similarly Q is positive semidenite, denoted Q  0,
if the inequality is nonstrict; and negative denite and negative semidenite
are similarly dened. If a matrix is not positive or negative semidenite,
then it is indenite.
44 1. Preliminaries in Finite Dimensional Space
The following properties of positive matrices follow directly from the
denition, and are left as exercises:
 If Q > 0 and A 2 C nn , then A QA  0. If A is invertible, then
A QA > 0.
 If Q1 > 0, Q2 > 0, then 1 Q1 + 2 Q2 > 0 whenever 1 > 0, 2  0.
In particular, the set of positive denite matrices is a convex cone in
H n , as dened in the previous section.
At this point we may well ask, how can we check whether a matrix is
positive denite? The following answer is derived from Theorem 1.9:
If Q 2 H n ; then Q > 0 if and only if the eigenvalues of Q are all positive.
Notice in particular that a positive denite matrix is always invertible, and
its inverse is also positive denite. Also a matrix is positive semi-denite
exactly when none of its eigenvalues are negative; in that case the number
of strictly positive eigenvalues is equal to the rank of the matrix.
An additional useful property for positive matrices is the existence of a
square root. Let Q = U U   0, in other1 words the diagonal elements of
 are non-negative.
12
Then we can dene  2 to be the matrix with diagonal
elements k , and
Q 21 := U  12 U  :
Then
12 21
Q 21  0 (also Q 21 > 0 when Q > 0) and it is easily veried that
Q Q = Q.
Having dened a notion of positivity, our next aim is to generalize the
idea of ordering to matrices, namely what does it mean for a matrix to be
larger than another matrix? We write
Q>S
for matrices Q, S 2 H n to denote that Q ; S > 0. We refer to such
expressions generally as matrix inequalities. Note that for matrices that
it may be that neither Q  S nor Q  S holds, i.e. not all matrices are
comparable.
We conclude our discussion by establishing a very useful result, known
as the Schur complement formula.
Theorem 1.10. Suppose that Q; M , and R are matrices and that M and
Q are self-adjoint. Then the following are equivalent
(a) The matrix inequalities Q > 0 and
M ; RQ;1 R > 0 both hold;
(b) The matrix inequality
 
M R
R Q > 0 is satised:
 1.3. Matrix Theory 45
Proof . The two inequalities listed in (a) are equivalent to the single block
inequality.
 
M ; RQ;1 R 0 > 0 :
0 Q
Now left and right multiply this inequality by the nonsingular matrix
 
I RQ;1
0 I
and its adjoint, respectively, to get
      
M R I RQ;1 M ; RQ;1R 0 I 0
R Q = 0 I
0 Q Q;1 R I > 0:
Therefore inequality (b) holds if and only if (a) holds. 
We remark that an identical result holds in the negative denite case,
replacing all \<" by \>".
Having assembled some facts about self-adjoint matrices, we move on to
our nal matrix theory topic.

1.3.3 Singular value decomposition

Here we introduce the singular value decomposition of a rectangular matrix,
which will have many applications in our analysis, and is of very signicant
computational value. The term singular value decomposition or SVD refers
to the product U V  in the statement of the theorem below.
Theorem 1.11. Suppose A 2 C mn and that p = minfm; ng. Then there
exist unitary matrices U 2 C mm and V 2 C nn such that
A = U V  ;
where  2 Rmn and its scalar entries satisfy
(a) the condition ij = 0, for i 6= j ;
(b) the ordering 11  22 


 pp  0.
Additionally, if A 2 Rmn then U and V can be chosen to be orthogonal
matrices.
Proof . Since the result holds for A if and only if it holds for A, we assume
without loss of generality that n  m. To start let r be the rank of A A,
and therefore by Theorem 1.9 we have
2 3
 2
  1 0
A A = V 0 00 V  ; where  = 64 . . . 75 > 0 and V is unitary.
0 r
46 1. Preliminaries in Finite Dimensional Space
We also assume that the nonstrict ordering 1 


 r holds. Now dene

 0
J= 0 I
and we have  
;1   ; 1 ; 1 
J V A AV J = (AV J ) (AV J ) = 0 0 ; ; 1 I r 0

where Ir denotes the r  r identity matrix. From the right-hand side we

see that the rst r columns of AV J ;1 form an orthonormal set, and the
remaining columns must be zero. Thus
 
AV J ;1 = U1 0 ;
where U1 2 C mr . This leads to

    0   
   0 
A= 1 U 0 U U
0 I V = 1 2 0 0 V ;
m(n;r) . So choose U2
 side is valid for any U2 2 C
where the right-hand
such that U1 U2 is unitary.
To prove the nal part of the theorem, simply note that if A is a real
matrix then all of the constructions above can be made with real matrices.

When n = m the matrix  in the SVD is diagonal. When these dimensions
are not equal  has the form of either
2 3
2
11 0
3  11 0
6 . 7
6
6 ... 7
7
4 . . 5 when n > m, or 6 7 when n < m.
0 mm 0
4 nn 5
0 0
The rst p non negative scalars kk are called the singular values of the
matrix A, and are denoted by the ordered set 1 ; : : : p , where k = kk . As
we already saw in the proof, the decomposition of the theorem immediately
gives us that
A A = V ( )V  and AA = U ( )U  ;
which are singular value decompositions of A A and AA . But since V  =
V ;1 and U  = U ;1 it follows that these are also the diagonalizations of
the matrices. We see
12  22 


 p2  0
are exactly the p largest eigenvalues of A A and AA ; the remaining eigen-
values of either matrix are all necessarily equal to zero. This observation
provides a straightforward method to obtain the singular value decomposi-
tion of any matrix A, by simply diagonalizing the Hermitian matrices A A
and AA .
 1.4. Linear Matrix Inequalities 47
The SVD of a matrix has many useful properties. We use  (A) to de-
note the largest singular value 1 , which from the SVD has the following
property.
 (A) = maxfjAvj : v 2 C n and jvj = 1g:
Namely it gives the maximum magnication of length a vector v can
experience when acted upon by A.   
Finally, partition U = u1


um and V = v1


vn and
suppose that A has r nonzero singular values. Then
   
ImA = Im u1


ur and ker A = Im vr+1


vn :
That is the SVD provides an orthonormal basis for both the image and
kernel of A. Furthermore notice that the rank of A is equal to r, precisely
the number of nonzero singular values.

1.4 Linear Matrix Inequalities

This section is devoted to introducing the central concept of a linear ma-
trix inequality which we use throughout the course. The reasons for its
importance are:
 From an analytical point of view, many important problems can be
reduced to this form. Such reductions will be presented as we go along
during the rest of the course.
 The resulting computational problem can be eciently and com-
pletely solved by recently established numerical algorithms. While
this second issue is beyond the scope of this course, we will devote
some space in this section to explain some of the properties behind
this computational tractability.
A linear matrix inequality, abbreviated LMI, in the variable X is an
inequality of the form
F (X ) < Q;
where
 the variable X takes values in a real vector space X ;
 the mapping F : X ! H n is linear;
 the matrix Q is in the set of Hermitian matrices H n :
The above is a strict inequality and F (X )  Q is a nonstrict linear matrix
inequality. Thus to determine whether an inequality is an LMI, we simply
see whether the above conditions are satised. Let us consider some explicit
examples.
48 1. Preliminaries in Finite Dimensional Space
Examples:
To start we note that every LMI can be written in a vector form. Suppose
that V1 ; : : : ; Vm is a basis for the vector space X . Then for any X in X we
have that there exist scalars x1 ; : : : ; xm such that
X = x1 V1 +


+ xm Vm :
If we substitute into the LMI F (X ) < Q, and use the linearity of F , we
have
x1 F (V1 ) +


+ xm (Vm ) < Q:
Thus the variables are the scalars xk , and the F (Vk ) are xed Hermitian
matrices. In other words every LMI can be converted to the form
x1 F1 +


+ xm Fm < Q;
where x 2 R is the variable. While this coordinate form could also be
m
taken as a denition of an LMI, it is not how we typically encounter LMIs
in our course, and is often cumbersome for analysis.
Consider the inequality
A XA ; X < Q; (1.3)
where A 2 Rnn ; Q 2 H n and the variable X is in Sn. If we dene
F (X ) = A XA ; X;
then clearly this is a linear mapping Sn ! Sn. Therefore we see that (1.3)
is an LMI.
Now look at the matrix inequality
A XA + BY + Y  B  + T < 0;
where A 2 C nn ; B 2 C nm ; T 2 H n , and the variables X and Y are in Sn
and Rmn respectively. This too is an LMI in the variables X and Y . To
see this explicitly let
Z := (X; Y ) 2 Sn  Rmn
and dene F (Z ) = A XA + BY + Y  B  . Then F : Sn  Rmn ! H n is a
linear map and the LMI can be written compactly as
F (Z ) < ;T:
With these examples and denition in hand, we will easily be able to
recognize an LMI. 
Here we have formulated LMIs in terms of the Hermitian matrices, which
is the most general situation for our later analysis. In some problems LMIs
are written over the space of symmetric matrices Sn, and this is the usual
form employed for computation. This change is inconsequential in regard
 1.4. Linear Matrix Inequalities 49
to the following discussion, and furthermore in the exercises we will see
that the Hermitian form can always be converted to the symmetric form.
An important property of LMIs is that a list of them can always be
converted to one LMI. To see this suppose we have the two LMIs
F0 (X0 ) < Q0 and F1 (X1 ) < Q1
in the variables X0 and X1 in vector spaces X0 and X1 respectively. Then
this is equivalent to the single LMI
F (X ) < Q;
where
 
X = (X0 ; X1 ) 2 X0  X1 and Q = Q00 Q0
1
and we dene the function F : X0  X1 ! H n0 +n1 by
 
F (X ) = 0 F0 ( X0 ) 0
F1 (X1 ) :
Notice that if X0 and X1 are equal, or have common components then X
should be dened in a corresponding appropriate manner.
An important point, which will appear repeatedly in this course, is that
conditions that do not look like LMIs at rst glance, can sometimes be
converted to them. The Schur complement given in Theorem 1.10 will be
very useful in this regard. We illustrate this by a simple example.
Example:
Let A 2 Rnn , b; c 2 Rn , and d 2 R. The inequality
(Ax + b) (Ax + b) ; c x ; d < 0
is not an LMI, since the expression is quadratic in the variable x. However
the Schur complement formula implies it is equivalent to
  
c x + d (Ax + b) > 0:
Ax + b I
Since the left hand side matrix now depends anely on x, the latter is an
LMI and can easily be expressed in the form F (x) < Q. 
As our course progresses we will nd that many control problems can be
formulated in terms of nding solutions to LMIs. More generally we will
have optimization problems with LMI constraints on the solution space.
This class of optimization is known as semidenite programming. The
simplest semidenite program is a decision problem, and is of the form:
Does there exist X 2 X , satisfying F (X ) < Q?
50 1. Preliminaries in Finite Dimensional Space
This is known as a feasibility problem and asks only whether there is any
element which satises the LMI. Computational aspects will be discussed
later, but we can already note an important property of the solution set.
Proposition 1.12. The set C := fX 2 X such that F (X ) < Qg is convex
in X .
Proof . Suppose X1, X2 2 C , which means they satisfy F (X1 ) < Q and
F (X2 ) < Q. Consider any point X3 in L(X1 ; X2 ), namely X3 = X1 +
(1 ; )X2 , for some value  2 [0; 1]. Using linearity of the function F we
have
F (X3 ) = F (X1 ) + (1 ; )F (X2 ) < Q + (1 ; )Q = Q:
The inequality follows from the fact that positive denite matrices are a
convex cone. Therefore X3 2 C . 
We remark that the above proposition does not say anything about the
LMI being feasible; indeed C could be empty.
We now turn to a more general semidenite optimization problem.
minimize: c(X );
subject to: F (X )  Q and X 2 X ,
where c(X ) is a linear functional on X . It is referred to as the linear
objective problem.
In this formulation we are being informal about the meaning of \mini-
mize"; in rigor we want to compute the inmum, which may or may not
be achieved at a given X .
Example:
Let X = R2 ; nd the inmum of x1 subject to
   
;x1 0  0 1 :
0 ;x2 1 0
It is easy to show that the answer is zero, but this value is not achieved by
any matrix satisfying the constraint; notice this happens even in the case
of non-strict LMI constraints. 
Clearly the linear objective problem makes sense only if the LMI is feasi-
ble, which seems to imply that such problems are of higher diculty than
feasibility. In fact, both are of very similar nature:
 In the exercises you will show how the linear objective problem can
be tackled by a family of feasibility questions.
 Conversely, the feasibility problem can be recast as the problem
J := inf t
subject to F (X ) ; tI  Q:
 1.4. Linear Matrix Inequalities 51
The latter is a linear objective problem in the variables t 2 R and X 2
X , and the corresponding LMI constraint is automatically (strictly)
feasible. It is an easy exercise to see that J < 0 if and only if the LMI
F (X ) < Q is feasible.
Given these relationships, we focus for the rest of the section on the linear
objective problem, with the assumption that the strict LMI F (X ) < Q is
feasible. A pictorial view of the problem is given in Figure 1.7, for the case
 
X = xx1 2 R2 :
2
The convex set depicted in the gure represents the feasibility set C =
fX : F (X )  Qg for the linear objective problem; while we have drawn a
bounded set, we remark that this is not necessarily the case.

C Xn+1
Xn  
 Xmin

Figure 1.7. Illustration of semidenite programming

Since C (X ) is a linear functional on R2 , it has the form

C (X ) = Y  X = y1 x1 + y2 x2
for a xed vector Y 2 R2 ; therefore the point Xmin that solves the problem
is the element of C with the most negative projection in the direction of Y ,
as depicted in Figure 1.7. Also the picture suggests that there are no other
local minima for the function in the set, namely for every other point there
is a \descent" direction. This property, fundamental to convex optimization
problems, is now stated precisely.
Proposition 1.13. Suppose X0 is a local minimum of the linear objective
problem, that is c(X0 )  c(X ) for every X 2 N (X0 ) \ C . Then X0 is the
global minimum of the problem over C .
52 1. Preliminaries in Finite Dimensional Space
Proof . Let X1 be another point in C . Since C is convex, it contains the
line segment
L(X1; X0 ) = fX1 + (1 ; )X0 ;  2 [0; 1]g:
Also by denition the neighborhood N (X0 ) will contain the points X1 +
(1 ; )X0 for suciently small , say  2 [0; ). Now the function f () =
c(X1 + (1 ; )X0 ) is linear in  2 [0; 1], and f (0)  f () for  2 [0; ) by
hypothesis. Then f () is non-decreasing and f (0)  f (1), or equivalently
c(X0 )  c(X1 ), which concludes our proof. 
The above property generalizes (see the exercises) to any convex opti-
mization problem, and plays a strong role in ensuring that these problems
can be solved globally, and not just locally, by numerical algorithms.
At this point the reader may be wondering about the nature of semidef-
inite programming algorithms; we will purposely stay away from any
detailed discussion of these methods, which would take us far aeld, and
defer to the authoritative references provided at the end of the chapter.
However we will provide a few remarks aimed mainly at reinforcing the
idea that these problems are fundamentally tractable.
A rst observation is that clearly the minimum, if it exists, must lie on
the boundary of the feasible set; thus one could think one restricting the
search to the boundary. This is in fact a popular alternative in the case of
linear programming, which corresponds to the case where the feasible set
is a polytope, the intersection of a nite number of half-spaces. However
in the semidenite programming problem the boundary is in general quite
complicated, and thus methods involving interior points are favored.
Going back to Figure 1.7, suppose we have a point Xn in our feasible
set; an immediate consequence is that we need only keep the set
fX 2 C : c(X )  c(Xn )g
for our remaining search for the global minimum. This amounts to inter-
secting C with a half-space; thus we can progressively shrink the feasibility
region to zero, provided we are able to successively generate a \good" new
feasible point Xn+1 . Many optimization algorithms are based on this prin-
ciple; one of the simplest is the so-called ellipsoid algorithm that alternates
between \cutting" and overbounding the resulting set by an ellipsoid; Xn+1
would then be the center of such ellipsoid; for details see the references.
More ecient methods for semidenite programming are based on barrier
functions to impose the feasibility constraint. The idea is to minimize the
function
c(X ) + (X ) (1.4)
where  > 0, and the barrier function (X ) is convex and approaches
innity on the boundary of the feasible set. For a denition of a convex
function, and some basic properties see the exercises at the end of the
 1.5. Exercises 53
chapter. Also you will show there that the function
(X ) = ; log(det[Q ; F (X )])
serves as a barrier function for the set C .
Provided we start from a feasible point, the minimization in (1.4) can
be globally solved by unconstrained optimization methods (e.g. Newton's
algorithm). By successively reducing the weight of the barrier function,
an iteration is produced which can be shown to converge to the global
minimum. The computational complexity of these algorithms has moderate
(polynomial) growth with problem size, the latter being characterized by
the dimensionality of the variable space X and of the constraint set H n . For
extensive details on this active area of optimization, as well as many other
alternative algorithms, the reader is encouraged to consult the references.
We have now completed our preliminary preparation for the course, and
are ready to investigate some control theory problems.

1.5 Exercises
1. Find a basis for C mn as a real vector space. What is the dimension
of this real vector space?
2. The spaces H n and Sn are both real vector spaces, nd bases for each.
How are they related?
3. Determine the dimension of the set of homogeneous multinomials
P3[4] . What is the general formula for the dimension of Pm[n] .
4. We consider the mapping  dened in x1.1.4. Let a 2 P3[1] be
a(x1 ; x2 ; x3 ) = x2 , and consider  : P3[1] ! P3[2] , which is dened
by
(p)(x1 ; x2 ; x3 ) = a(x1 ; x2 ; x3 )p(x1 ; x2 ; x3 ):
Choose bases for P3[1] and P3[2] , and represent  as the corresponding
matrix [].
5. Suppose A : V ! W . Let fAv1 ; : : : ; Avr g be a basis for ImA and
fu1 ; : : : ; uk g be a basis for ker A. Show that fv1 ; : : : ; vr ; u1; : : : ; uk g
is a basis for V and deduce that dim(V ) = dim(ker A) + dim(ImA).
6. Given a mapping A : V ! V show that both ker A and ImA are
A-invariant.
7. By direct calculation, show that given A 2 C mn and B 2 C nm ,
the identity TrAB = TrBA holds. Use this to prove that the trace of
any square matrix is the sum of its eigenvalues.
54 1. Preliminaries in Finite Dimensional Space
8. Prove that every linear functional F on Sn can be expressed as
F (X ) = Tr(Y X ) for some xed Y 2 Sn.
9. Suppose A is a Hermitian matrix, and that 0 and 1 are two eigenval-
ues with corresponding eigenvectors x0 and x1 . Prove that if 0 6= 1 ,
then x0 and x1 are orthogonal.
10. Show that if P  0 and Q  0, then Tr(PQ)  0.
11. Suppose A 2 C mn has the singular value decomposition U V  .
(a) Show that if x 2 spanfvk+1 ; : : : ; vn g then jAxj  k+1 jxj.
(b) If A is n  n, let 1  2 


n be the diagonal elements of
. Denote (A) = n . Show that A is invertible if and only if
(A) > 0, and in that case
 (A;1 ) =  (1A) :
(c) If A 2 H n , then ; (A) I  A   (A) I .
12. Suppose that > 0 and that X 2 Rnm .
(a) Show that  (X )  if and only if X X  2 ;
(b) Convert the constraint  (X )  to an equivalent LMI
condition.
13. The spectral radius of a matrix M 2 C nn is dened as
(M ) := maxfjj such that  is an eigenvalue of M g:
(a) Show that (M )   (M ) and nd both numbers for
2 3
0 1 0
6 . . 7
M = 664 . . . . 775
1
0 0
(b) Show that
(M )  inf (DMD;1 )
D invertible
(c) Prove that there is equality in (b). Hint: use the Jordan form.
(d) Deduce that (M ) < 1 if and only if the set of LMIs
X > 0; M  XM ; X < 0
is feasible over X 2 H n .
14. Consider a real LMI given by
F (X ) < Q;

where the linear map F : X ! Sn, Q 2 Sn, and X is a real vector
space. Show that any LMI of the standard form given in x1.4 with
 1.5. Exercises 55
respect to the Hermitian matrices H n , can be converted to a real LMI
with n = 2n. Hint: rst nd an condition for a matrix A 2 H n to be
positive denite, in the form of an LMI on Re(A) and Im(A).
15. In x1.4 two types of LMI problems were introduced, the feasibility
problem and the linear objective problem. Show how the feasibility
problem can be used with iteration to solve the latter.
16. Another type of LMI optimization problem is the so-called generalized
eigenvalue minimization problem, which is
minimize: ;
subject to: F0 (X ) + F1 (X ) < Q0 + Q1 and X 2 X ,
where F1 and F2 are linear mapping from X to Sn, and Q0 , Q1 2 Sn.
Show that the linear objective problem can be reformulated in this
format. Further show that this problem can be solved by iteration
using the feasibility problem.
17. Let C be a convex set in a real vector space X . A function  : C ! R
is said to be convex if it satises
(x1 + (1 ; )x2 )  (x1 ) + (1 ; )(x2 )
for every x1 , x2 in C and every  2 [0; 1]. The minimization of such
a function is called a convex optimization problem. As an important
example, the function (x) = ; log(x) is convex in (0; 1). Clearly,
any linear function is convex.
(a) Prove that for a convex function, every local minimum is a global
minimum.
(b) Show a function  is convex if and only if for any x1 , x2 in C ,
the function f () := (x1 + (1 ; )x2 ) is convex in  2 [0; 1].
(c) Prove that (X ) = ; log(det(X )) is convex in the set of positive
matrices fX > 0g. Hint: use the identity
1 1 1 1
X1 + (1 ; )X2 = X22 I + X2; 2 (X1 ; X2 )X2; 2 X22
and express det(X1 + (1 ; )X12 ) in terms of the
12
eigenvalues of
;
the Hermitian matrix H := X2 (X1 ; X2 )X2 .
2 ;
(d) Deduce that if F : X ! H n is linear, ; log(det[Q ; F (X )]) is a
barrier function for the set C = fX 2 X : F (X ) < Qg.

Notes and references

Given its ubiquitous presence in analytical subjects, there are many ex-
cellent books on linear algebra at the introductory level; one choice is for
56 1. Preliminaries in Finite Dimensional Space
example [128]. For an advanced treatment from a geometric perspective
the reader is referred to [51].
For many more details on convexity see the standard reference [109].
Two excellent sources for matrix theory are [58] and the companion work
[59]. For information and algorithms for computing with matrices see [49].
LMIs have a long history as an analytical tool, but it is only with re-
cent advances in semidenite programming that they have acquired a more
central role in computation. These advances have been largely motivated
by the pioneering work [85] on interior point methods for general convex
optimization problems. Semidenite programming is still undergoing rapid
development; a recent survey is [133]. Also see [13] for an introduction to
LMI optimization with many examples from control theory.
Reference [79] is a more general work on optimization primarily aimed
at innite dimensional vector spaces.
 This is page 57
Printer: Opaque this

2
State Space System Theory

We will now begin our study of system theory. This chapter is devoted to
examining one of the building blocks used in the foundation of this course,
the continuous time, state space system. Our goal is to cover the fundamen-
tals of state space systems, and we will consider and answer questions about
their basic structure, controlling and observing them, and representations
of them.
The following two equations dene a state space system.
x_ (t) = Ax(t) + Bu(t); with x(0) = x0 (2.1)
y(t) = Cx(t) + Du(t);
where u(t); x(t) and y(t) are vector valued functions, and A; B; C and D
are matrices. We recall that the derivative x_ (t) is simply the vector formed
from the derivatives of each scalar entry in x(t). The rst of the above
equations is called the state equation and the other the output equation.
The variable t  0 is time and the function u(t) is referred to as the system
input. The functions x(t) and y(t) are called the state and output of the
system respectively, and depend on the input.
For later reference we dene the dimensions of the vectors by
u(t) 2 Rm ; x(t) 2 Rn and y(t) 2 Rp :
Thus A is an nn matrix; the matrix B is nm; matrices C and D are pn
and p  m respectively. We will restrict ourselves to real matrices during
the chapter, however all the results we prove hold for complex matrices as
well. Notice that the system given above is a rst order linear dierential
equation with an initial condition, and therefore has a unique solution.
58 2. State Space System Theory
The state space formulation above is very general because many systems
of higher order linear dierential equations can be reformulated in this
way. This generality motivates our study of state space systems. Before
considering this system as a whole we will examine simpler versions of it
to successively build our understanding.

2.1 The autonomous system

In this short section we will develop some intuition about the state equation
by considering one of its simplications. This will lead us to introducing
and studying the generalization of the exponential function to matrices, a
function we will nd repeatedly useful in the chapter.
Let us focus on the autonomous system
x_ (t) = Ax(t); with x(0) = x0 : (2.2)
This system is a special case of (2.1) in which there are no inputs and no
outputs. The state function x(t), for t  0, is completely specied by the
initial condition x(0). The immediate question that comes to mind is, can
we explicitly state the solution to this dierential equation?
Consider rst the case where x(t) is a scalar, namely n = 1. Then A is
equal to some scalar a. Thus we have that
x(t) = eat x(0)
is the unique solution for x(t). There are a number of equivalent ways to
dene the meaning of the scalar function eat , and one of these is using its
power or Taylor series expansion.
We follow this lead for the multivariable case when n > 1 and make the
denition: the matrix exponential eM of a square matrix M is dened to
be the matrix sum of the power series
eM = I + M + 2!1 M 2 + 3!1 M 3 + 4!1 M 4 +


:
It is not dicult to show that eM is well-dened; namely the above series
always converges. Notice that if M is a scalar then this denition agrees
with the familiar power series expansion of the scalar exponential function.
Our main focus will be the time dependent function
eAt = I + At + 2!1 A2 t2 + 3!1 A3 t3 +



which is dened for every t. The following are some basic properties of the
matrix exponential.
(a) e0 = I , where 0 denotes the zero element in Rnn ;
(b) eM  = (eM ) ;
(c) dtd eAt = AeAt = eAt A;
 2.1. The autonomous system 59
(d) If the square matrices M and N commute, that is MN = NM , then
the relationship eM +N = eM eN holds.
The rst three properties follow immediately from the denition, for the
last one see the exercises at the end of the chapter. We will now illustrate
the matrix types which result from this denition by looking at two special
cases; by using the Jordan form, these two examples can be used to specify
the exponential of any matrix.
Examples:
Our rst example pertains to the case where A is a diagonalizable matrix,
namely a similarity transformation T exists such that
2 3
1 0
T ;1AT =  = 64 . . . 75 ; (2.3)
0 n
holds where k are the eigenvalues of A. This is a convenient transformation
because, for k  0, we have Ak = T k T ;1. Therefore
2  t 3
e1 0
; 1
2!


eAt = T I + t + 2 t2 +


T ;1 = T 64 ... 7 ;1
5T :
0 
en t
Thus we see that the matrix exponential consists of linear combinations of
exponentials whose exponents are the eigenvalues of the matrix.
The second case we consider is where A is exactly a Jordan block, and
so has the form
2 3
 1 0
6 . . 7
A = 664 . . . . 775 = I + N; (2.4)
1
0 
where N is the matrix with ones above the diagonal and zeros in its re-
maining entries. An interesting and useful feature of N is that its powers
are easily computed. In particular its square is
2 3
0 0 1 0
6
6 . . . . . . . . . 77
2
N = 666 7
0 0 177
4 0 05
0 0
and the successive powers of N have analogous structure, with the diagonal
of ones shifting upwards, until eventually we nd N n = 0. That is N is
nilpotent of order n.
60 2. State Space System Theory
This feature of N can be exploited to compute the exponential eAt. Since
the matrices I and N commute we can invoke property (d) above to nd
that
 
eAt = eIt eNt = et I + Nt + 2!1 N 2 t2 +


(n;1 1)! N n;1 tn;1
This gives us the nal result
2
1 t t2


tn;1 3
6 2 (n;1)! 7
6 ... ... ... .. 7
t 6 .2 7
eAt = e 6
6 1 t t 7
7 :
6 2 7
40 1 t 5
0 1
So we have shown that the matrix exponential of a single Jordan block
consists of functions of the form tkk! et . 
Having gained some familiarity with the matrix exponential we now re-
turn to the dierential equation (2.2). Using the above properties of the
matrix exponential, it follows immediately that the unique solution to this
autonomous equation is given by
x(t) = eAt x(0); for t  0.
One of the main properties of interest for such a system is that of stability.
We say that the system described by (2.2) is internally stable if, for every
initial condition x(0) 2 Rn , the limit
!1 0 holds:
x(t) t;!
This limit simply says that each scalar entry of x(t) tends to zero with
time. In the exercises you will show that the system is internally stable if
and only if each eigenvalue  of A satises
Re < 0;
namely they are all in the left half of the complex plane. A matrix that
has this property is called a Hurwitz matrix. Thus we have the following
summarizing result.
Proposition 2.1. The autonomous system in (2.2) is internally stable if
and only if A is Hurwitz.
At this point we have covered all the key properties of the autonomous
state equation, and saw that these follow directly from those of the matrix
exponential function. We are now ready to consider systems systems with
inputs, controlled systems.
 2.2. Controllability 61
2.2 Controllability
This section is the most important of the chapter as it contains machinery
and concepts that are crucial to the rest of the course. The problem we
intend to pose and answer is a fundamental question about the control of
state space systems. We will study the equation
x_ (t) = Ax(t) + Bu(t); with x(0) = 0; (2.5)
which is the state equation from (2.1). Notice that for now we set the initial
condition of the state to zero. Thus it is easy to verify that the solution to
this equation is
Z t
x(t) = eA(t; )Bu( )d; for t  0.
0
The integrand in the equation is a vector, so each scalar entry of x(t) is just
equal to the integral of of the corresponding scalar entry of eA(t; )Bu( ).
Our rst objective will be to determine what states can be reached by
manipulating the input u(t). In other words how much control do we have
over the values of x(t) through choice of the function u? We will nd that
this question has a surprising answer.

2.2.1 Reachability
We begin our study by asking, given a xed time t, what are the possible
values of the state vector x(t)? Or asked another way: given a vector in Rn
is it possible to steer x(t) to this value by choosing an appropriate input
function u(t)? We will answer this question completely, and will nd that
it has a surprisingly simple answer. To do this we require three related
concepts.
Set of reachable states
For a xed time t > 0, let Rt denote the states that are reachable at time
t by some input function u. Namely Rt is the set
Rt := f 2 Rn : there exists u such that x(t) =  g :
This denition is made respecting the state equation given in (2.5), where
the initial condition is zero. It turns out that Rt is a subspace of Rn . This is
a simple consequence of the linearity of (2.5), which we leave as an exercise.
Controllable subspace
Next we dene a subspace associated with any state equation. Given the
state equation x_ (t) = Ax(t) + Bu(t) we call the matrix
 
B AB A2 B


An;1 B
62 2. State Space System Theory
the associated controllability matrix. Equivalently we say it is the con-
trollability matrix of the matrix pair (A; B ). Recall that A is an n  n
matrix, and notice that the dimension n plays an important role in the
above denition.
Associated with the controllability matrix is the controllability subspace,
denoted CAB , which is dened to be the image of the controllability matrix.
 
CAB := Im B AB


An;1 B :
Thus we see that CAB , like Rt above is a subspace of Rn . When dimension
of CAB is n, or equivalently the controllability matrix has the full rank of n,
we say that (A; B ) is a controllable pair. In the same vein we refer to a state
space system as being controllable when the associated matrix pair (A; B )
is controllable. We shall soon see the motivation for this terminology.
Controllability gramian
Here we dene yet another object associated with the state equation,
a matrix which depends on time. For each t > 0, the time dependent
controllability gramian is dened to be the n  n matrix
Z t
Wt := eA BB  eA d :
0
Having dened the set of reachable states, the controllability subspace
and the controllability gramian we can now state the main result of this
section. It will take a number of steps to prove.
Theorem 2.2. For each time t > 0 the set equality
Rt = CAB = Im Wt holds:
This theorem says that the set of reachable states is always equal to the
controllability subspace, and is also equal to the image of the controllability
gramian. Since the controllability subspace CAB is independent of time so
is the set of reachable states: if a state can be reached at a particular time,
then it can be reached at any t > 0, no matter how small. According to
the theorem, if (A; B ) is controllable then Rt is equal to the entire state
space Rn ; that is all the states are reachable by appropriate choice of the
input function u.
Let us now move on to proving Theorem 2.2. We will accomplish this by
proving three lemmas, showing sequentially that
 Rt is a subset of CAB ;
 CAB is a subset of ImWt ;
 ImWt is a subset of Rt .
These facts will be proved in Lemmas 2.6, 2.8 and 2.9 respectively.
 2.2. Controllability 63
Before our rst result we require a fact about matrices known as the
Cayley-Hamilton theorem. Denote the characteristic polynomial of matrix
A by charA (s). That is
charA (s) := det(sI ; A) =: sn + an;1 sn;1 +


a0 ;
where det(
) denotes the determinant of the argument matrix. Recall that
the eigenvalues of A are the same as the roots of charA (s). We now state
the Cayley-Hamilton theorem.
Theorem 2.3. Given a square matrix A the following matrix equation is
satised
An + an;1 An;1 + an;2 An;2 +


+ a0 I = 0;
where ak denote the scalar coecients of the characteristic polynomial of
A.
That is the Cayley-Hamilton theorem says a matrix satises its own
characteristic equation. In shorthand notation we write
charA (A) = 0:
We will not prove this result here but instead illustrate the idea behind the
proof, using the example matrices considered above.
Examples:
First consider the case of a diagonalizable matrix A as in (2.3). Then it
follows that
2 3
charA (1 ) 0
charA (A) = T charA ()T ;1 = T 64 ... 7 ;1
5T :
0 charA (n )
Now by denition each of the eigenvalues is a root of charA (s), and so we
see in this case charA (A) = 0.
Next we turn to the case of a Jordan block (2.4). Clearly in this case A
has n identical eigenvalues, and has characteristic polynomial
charA (s) = (s ; )n :
Now it is easy to see that charA (A) = 0 since
charA (A) = (A ; I )n = N n = 0;
where N is the nilpotent matrix dened in the earlier examples.
The general case of the Cayley-Hamilton theorem can be proved using
the ideas from these examples and the Jordan decomposition, and you are
asked to do this in the exercises. 
The signicance of the Cayley-Hamilton theorem for our purposes is
that it says the matrix An is a linear combination of the matrix set
64 2. State Space System Theory
fAn;1 ; An;2 ; : : : ; A; I g. Namely
An 2 spanfAn;1 ; An;2 ; : : : ; A; I g:
More generally we have the next proposition.
Proposition 2.4. Suppose k  n. Then there exist scalar constants
0 ; : : : ; n;1 satisfying
Ak = 0 I + 1 A +


+ n;1 An;1
We now return for a moment to the matrix exponential and have the
following result.
Lemma 2.5. There exist scalar functions 0 (t); : : : ; n;1(t) such that
eAt = 0 (t)I + 1 (t)A +


+ n;1 (t)An;1 ;
for every t  0.
The result says that the time dependent matrix exponential eAt can be
written as a nite sum, where the time dependence is isolated in the scalar
functions k (t). The result is easily proved by observing that, for each t  0,
the matrix exponential has the expansion
2 3
eAt = I + At + (At2!) + (At3!) +


:
The result then follows by expanding Ak , for k  n, using Proposition 2.4.
We are ready for the rst step in the proof of our main result, which
is to prove that the reachable states are a subset of the image of the
controllability matrix.
Lemma 2.6. The set of reachable states Rt is a subset of the controllability
subspace CAB .
Proof . Fix t > 0 and choose any reachable state  2 Rt. It is sucient to
show that  2 CAB . Since  2 Rt there exists an input function u such that
Z t
= eA(t; )Bu( )d :
0
Now substitute the expansion for the matrix exponential from Lemma 2.5
to get
Z t Z t
= 0 (t ;  )Bu( )d +


+ An;1 n;1 (t ;  )Bu( )d :
0 0
Writing this as a product we get
2 Rt 3
0 0 (t ;  )u( )d

 = B AB


An;1 B
6 .. 7
:
4
Rt
. 5

0 n;1 (t ;  )u( )d

 2.2. Controllability 65
The right most factor is a vector in Rnm , and so we see that  is in the
image of the controllability matrix. 
Next in our plan is to show that CAB is contained in the image of Wt .
Again we require some preliminaries. Given a subspace X we dene its
orthogonal complement X ? by
X ? = fx 2 Rn : x y = 0; for all y 2 Xg:
It is routine to show that the dimension of X ? is n minus the dimension
of X . We have the following elementary result which can be proved as an
exercise.
Proposition 2.7. Given any matrix W the subspace equality
(ImW )? = ker W  holds:
We now prove the next step in the demonstration of Theorem 2.2.
Lemma 2.8. The controllability subspace satises
CAB  ImWt :
Proof . We show equivalently that
?  (ImWt )? :
CAB
From Proposition 2.7 we have that
(ImWt )? = ker Wt = ker Wt ;
where the latter equation is true since the controllability gramian is
? .
symmetric. So we need to show that if  2 ker Wt then  2 CAB
Let  2 ker Wt and thus we have
  Wt  = 0 :
Applying this to the denition of Wt we have
Z t  Z t
0 =  eA BB  eA d = 
(  eA B )(B  eA   )d :
0 0
Let y( ) = B  eA   and the last equation says
Z t
y ( )y( )d = 0 :
0
Thus it follows that
y ( ) =   eA  B = 0
for each 0    t. Hence we have that the right-sided derivatives at zero
satisfy

dk y  = 0
d k  =0
66 2. State Space System Theory
for all k  0. Now

dk y  =   Ak B ;
d k  =0
and so it follows that
 
  B AB


An;1 B = 0;
where we have written n vector equations as a single matrix equation. The
latter equation says that the vector  is orthogonal to all vectors in the
? as required. 
image of the controllability matrix. This means  is in CAB
To nish our proof of Theorem 2.2 we complete the planned chain of
containments proving that any element in the image of the gramian is
necessarily a reachable state.
Lemma 2.9. The image ImWt  Rt , where Rt is the set of reachable
states.
Proof . Select any time t > 0 and  2 ImWt . Then by denition there
exists  in Rm so that
 = Wt  :
Now dene
u( ) = B  eA(t; ); for 0    t :
Then the solution to x_ = Ax + Bu; x(0) = 0 at time t is
Z t Z t
x(t) = eA(t; )Bu( )d = eA(t; )BB  eA(t; ) d (2.6)
0 0
Z t
= eA BB  eA d  = Wt  =  : (2.7)
0
By denition this means that  is indeed in the set of reachable states. 
Thus we have successfully proved the theorem of this section, which says
that Rt = CAB = ImWt . Since CAB is the image of the controllability
matrix this gives us a simple way to compute the reachable states of the
system. Also notice that in the proof of Lemma 2.9 we constructed an
explicit input u for reaching a given state in the set of reachable states.
Summarizing, we have shown precisely which states are reachable and how
to reach them.

2.2.2 Properties of controllability

Recall that we said the pair (A; B ) is controllable if CAB has maximum
dimension n, namely all states are reachable. This is an important situation
 2.2. Controllability 67
and we now investigate this special case further. Let us now return to the
full state space system of (2.1) which is
x_ (t) = Ax(t) + Bu(t); x(0) = x0
y(t) = Cx(t) + Du(t):
This set of equations species a relationship between the input u and the
output y. It is natural to ask whether this is the unique set of equations of
this form which relate u to y. The answer is on the contrary, there are many
state-space systems which provide the same relationship. We elaborate on
this point later, but for the moment we will exhibit a family of state-space
systems obtained from the given one by a change of coordinates in the
state. By dening
x~(t) = Tx(t);
where T 2 Rnn is a similarity transformation, we have
x~_ (t) = T x_ (t) = T (Ax(t) + Bu(t))
= TAT ;1x~(t) + TBu(t) ;
which has the initial condition x~(0) = Tx(0). Similarly we have that
y(t) = Cx(t) + Du(t) = CT ;1x~(t) + Du(t) :
Thus we see that a similarity transformation denes a new state and thus
a new realization for the same input-output relationship. Summarizing we
have that any similarity transformation denes a mapping
(A; B; C; D) 7! (A;~ B;
~ C;~ D~ );
from one system realization to another by A~ = TAT ;1, B~ = TB , C~ =
CT ;1 and D~ = D. Notice that D is unaected by a state transformation.
We have the rst property of controllability now stated, which says that
systems remain controllable under state transformations.
Proposition 2.10. The pair (A; B) is controllable, if and only if, (TAT ;1; TB)
is controllable.
Proof . The controllability matrix of (TAT ;1; TB) is
   
TB (TAT ;1)TB


(TAT ;1)n;1 TB = T B AB


An;1 B
Since T is a similarity transformation the rank of this matrix is clearly
equal to that of
 
B AB


An;1 B ;
which of course is the controllability matrix associated with the matrix
pair (A; B ). Thus either of these controllability matrices has rank n exactly
when the other does. 
68 2. State Space System Theory
In the language of subspaces the above result says
CA~B~ = T CAB ;
where A~ = TAT ;1 and B~ = TB . This should be readily clear from the
proof and the denition of the controllable subspace.
Recall the denition of a subspace being invariant to a matrix: suppose W
is a subspace of Rn and A is an n  n matrix, then we say W is A-invariant
if
AW  W :
Also recall that if W is r dimensional and A-invariant, then there exists a
similarity transformation T such that

;1 A~11 A~12 
TAT = 0 A~
21
where A~11 2 R and
r  r
 
T W = Im I0r :
We use this denition and the latter fact in the next two very important
results.
Proposition 2.11. Suppose A 2 Rnn and B 2 Rnm . Then the
controllable subspace CAB is A-invariant.
Proof . We need to show that ACAB  CAB . Using the denition of CAB
we have
 
ACAB = A Im B AB


An B
 
= Im AB A2 B


An B :
Now by the Cayley-Hamilton theorem we see that
   
Im AB A2 B


An B  Im B AB


An;1 B
and therefore the result follows. 
An immediate consequence of this proposition and the preceding discussion
is the following.
Theorem 2.12. Given a matrix pair (A; B) with the dimension dim(CAB ) =
r. Then there exists a similarity transformation T such that
(a) The transformed pair has the form

~11 A~12 
A
 
~
; 1
TAT = 0 A~ and TB = B01 ;
22
where A~11 2 Rrr and B~1 2 Rrm ;
 2.2. Controllability 69
(b) The transformed controllable subspace is
 
T CAB = Im I0r  Rn ;
(c) The pair (A~11 ; B~1 ) is controllable.
The rst two parts of the theorem are essentially complete since we already
know that CAB is invariant to A, and that ImB  CAB holds by denition.
Part (c) follows by evaluating the rank of the controllability matrix for the
transformed system. We leave this as an exercise. When a matrix pair is
in the form given in (a) of the theorem, with (c) satised we say it is in
controllability form.
We now summarize the procedure for transforming a pair (A; B ) into
this form.

General procedure for controllability form

(a) Find a basis fv1 ; : : : ; vr g in Rn for CAB ;
(b) Augment this basis to get fv1 ; : : : ; vn g a complete basis for Rn ;
(c) Dene T ;1 = [v1 ; : : : ; vn ] and set
A~ = TAT ;1 and B~ = TB ;
(d) Partition these matrices as

~ ~12   
B~1 :
A~ = A011 A A~22 and ~
B = 0
It is important to understand exactly why the above procedure works, as it
collects a number of key ideas. We now move on to examine the implications
of the controllability form.

2.2.3 Stabilizability and the PBH test

When a system is put in controllability form, as the name suggests, the
controllable and uncontrollable parts of the system are isolated. Let see
this explicitly. Consider the state equation
x_ (t) = Ax(t) + Bu(t); x(0) = x0 :
Dene a new state x~(t) = Tx(t), where T is a similarity transformation that
transforms the system to controllability form. Thus the new state equations
are

_  ~ ~12  x~1 (t) B~1 
x~_ (t) = xx~~_ 1 ((tt)) = A011 A
A~22 x~2 (t) + 0 u(t);
2
70 2. State Space System Theory
with the initial condition given by
 
x
~1 (0)
x~(0) = x~ (0) = Tx(0):
2
Notice that x~2 (t) evolves according to the equation
x~_ 2 (t) = A~22 x~2 (t) ;
which only depends on the initial condition x~2 (0), and is entirely inde-
pendent of both x~1 (0) and the input u. Therefore x~2 is said to be the
uncontrollable part of the system. Now suppose that the matrix A~22 is not
Hurwitz, then there exists an initial condition x~2 (0) such that
x~2 (t) does not tend to zero as t ! 1 :
Namely there exists an initial condition x(0) so that no matter what u
is, the state x(t) does not tend to zero as the time t tends to innity.
Conversely, it turns out that if A~22 above is Hurwitz then for every initial
condition x(0), there exists an input u such that x(t) tends asymptotically
to zero. This latter fact will become clear to us soon, and motivates the
following denition.
Denition 2.13. Suppose (A; B) is a matrix pair. If A~22 in a controlla-
bility form is Hurwitz, then we say the pair is stabilizable.
Thus a matrix pair is said to be stabilizable if for every initial condition it
is possible to nd an input that asymptotically steers the state x(t) to the
origin; clearly this is the weakest form of stability for a controlled system.
If the matrix pair is not stabilizable there exists an initial condition so that
irrespective of u(t), the state function x(t) does not tend to the origin.
So far the only way we have to determine stabilizability of a pair (A; B )
is to convert the system to controllability form. This is both inconvenient
and unnecessary, and our goal is now to develop a more sophisticated test.
To do this we rst develop an alternate method for checking controllability,
the Popov-Belevitch-Hautus or PBH test. The test is stated below.
Theorem 2.14. The pair (A; B) is controllable if and only if, for each
 2 C , the rank condition
 
rank A ; I B = n holds :
Proof . First notice that this condition need only be satised at the eigen-
values of A, since the rank of A ; I is n otherwise.
We rst prove \only if", by a contrapositive argument. Suppose  is an
eigenvalue of A and
 
rank A ; I B < n :
Then there exists a nonzero vector x such that
 
x A ; I B = 0 :
 2.2. Controllability 71
Thus we have that both
x A = x and x B = 0 hold.
From the former it follows that
x A2 = 2 x
or more generally x Ak = k x for each k  1. Hence
 
x B AB


An;1 B = 0 ;
and we see (A; B ) is not controllable.
Now we prove \if", again using the contrapositive. Suppose (A; B ) is not
controllable, then by Theorem 2.12 there exists a similarity transformation
such that

A~11 A~21   
~
A = TAT = 0 A~ and B01 ;
~ ; 1
22
where the dimensions of A~22 are nonzero. Let  be an eigenvalue of A~22
and we see that
 
rank A~ ; I B~ < n :
 
Therefore A ; I B has rank less than n since
 ;1 
A ; I B = T A~ ; I B T 0 I0 :
   

This completes the contrapositive argument.


In view of this theorem we say that an eigenvalue  of matrix A is
controllable if
 
rank A ; I B = n ;
and so every eigenvalue of A is controllable, if and only if, (A; B ) is a con-
trollable pair. Thinking back on matrix A~ in the above proof, we see that
A~11 corresponds to the controllable states, whereas all the eigenvalues of
A~22 are uncontrollable. Note that the eigenvalues of A~11 and A~22 need not
be distinct. From this discussion and the proof of Theorem 2.14 the next
result follows readily.

Corollary 2.15. The matrix pair (A; B) is stabilizable if and only if the
condition
 
rank A ; I B = n holds for all  2 C+ :
This corollary states that one need only check the above rank condition at
the unstable eigenvalues of A, that is those in the closed right half-plane
C + . This is a much simpler task than determining stabilizability from the
denition.
72 2. State Space System Theory
2.2.4 Controllability from a single input
This section examines the special case of a pair (A; B ) where B 2 Rn1 .
That is our state space system has only a single scalar input. This investi-
gation will furnish us with a new system realization which has important
applications.
A matrix of the form
2 3
0 1 0
6
6 ... ... 7
7
6 7
4 0 0 1 5
;a0 ;a1 ;an;1
is called a companion matrix, and has the useful property that its
characteristic polynomial is
sn + an;1 sn;1 +


+ a0 :
The latter fact is easily veried. We now prove an important theorem which
relates single input systems to companion matrices.
Theorem 2.16. Suppose (A; B) is a controllable pair, and B is a column
vector. Then there exists a similarity transformation T such that
2 3 2 3
0 1 0 0
6 ... ... 7 6.7
TAT ;1 = 66 7
7 and TB = 66 .. 77 :
4 0 0 1 5 405
;a0 ;a1 ;an;1 1
The theorem states that if a single input system is controllable, then it can
be transformed to the special realization above, where A is in companion
form and B is the rst standard basis vector. This is called the controllable
canonical realization, and features the A-matrix in companion form with
the B -matrix the n-th standard basis vector of Rn . To prove this result we
will again use the Cayley-Hamilton theorem.

Proof . Referring to the characteristic polynomial of A and invoking the

Cayley-Hamilton theorem we get
An + an;1 An;1 +


+ a0 I = 0 :
Post multiplying this by B gives
An B = ;a0 B ;


; an;1 An;1 B :
 2.2. Controllability 73
From this equation, keeping in mind that B is a column vector, it is easy
to verify that
2 3
0 0 ;a0
6 ... 7
 ;  
A B


A B = B AB


A B 6 .
n 1 n ;1  61
6 ; a 1 7
7
| {z } 4 .. 0 .. 75
P .
0 1 ;an;1
| {z }
M
Note that P is both square and nonsingular. From the last equation we
have
P ;1 AP = M ;
and also that
2 3
1
  07
6
B = B AB


An;1 B 664 .. 775 :
.
0
Therefore we see
2 3
1
607
P ;1 B = 664 .. 775 :
.
0
Let us hold these forms in abeyance for the moment, and introduce two
new matrices.
2 3 2 3
0 1 0 0
6 . . 7 6.7
A~ := 664 . . . . 7 ~ = 66 .. 77 :
7 and B
0 0 1 5 405
;a0 ;a1 ;an;1 1
These are the matrices to which we want to transform. Dene
 
P~ := B~ A~B~


A~n;1 B~ :
Now (A;~ B~ ) is controllable; to see this simply write out the controllability
matrix. Therefore since A~ has the same characteristic polynomial as A, we
can use the same argument already followed to show that
2 3
1
607
P~ ;1 A~P~ = M and P~ ;1 B~ = 664 .. 775 :
.
0
74 2. State Space System Theory
Returning to our expressions earlier we have that
P ;1 AP = P~ ;1 A~P~ and P ;1 B = P~ ;1 B:
~
~ ;1 and we arrive at the desired result of
Set T = PP
TAT ;1 = A~ and TB = B~ :

We are now ready to leave this section and move on to one of the most
common types of closed-loop control. The decomposition given here will
have a direct role in the next section. Realize that the proof above provides
us with an explicit procedure for putting a matrix pair in the controllable
canonical form.

2.3 Eigenvalue assignment

In this section we consider a state space problem associated with a special
type of closed-loop control called state feedback. Take the state equation
x_ (t) = Ax(t) + Bu(t); x(0) = x0
and suppose we use the feedback law u(t) = Fx(t) where F is a xed matrix
in Rmn . Then the state equation becomes
x_ (t) = (A + BF )x(t); x(0) = x0 :
This is an autonomous system whose behavior depends only on A + BF .
Thus to control the system we might wish to stabilize it, meaning that
A + BF is Hurwitz. More generally we could try to specify the eigenvalues
of A + BF exactly by selecting F , thus achieving certain dynamical charac-
teristics. The question we answer in this section is, when are the eigenvalues
of A + BF freely assignable through choice of F . We answer this by rst
treating the single input case, and then tackling the more challenging multi
input case second.

2.3.1 Single input case

Here we consider the case where B 2 Rn1 . Suppose (A; B ) is controllable,
then by Theorem 2.16 we know there exists a similarity transformation
such that
2 3
0
  6.7
A~ = ;0a ;a


I ;an;  and B~ = 664 .. 775 :
0 1 1
0
1
 2.3. Eigenvalue assignment 75
Let E = f0 ; : : : ; n;1 g be the desired set of eigenvalues for A + BF . Since
the matrix A + BF is real we necessarily restrict E to be a set which is
symmetric with respect to the real line;  namely if  isin E , then so is the
complex conjugate of . Next let F~ = F~0


F~n;1 where each F~k is a
real scalar to be chosen. Then
 
A~ + B~ F~ = ;a 0+ F~ ;a + F~


I ;a + F~  :
0 0 1 1 1 n; n;
1

Now this matrix is in companion form and so its characteristic polynomial

is
sn + (an;1 ; F~n;1 )sn;1 +


+ (a0 ; F~0 ) :
Clearly by appropriate choice of the scalars F~k we can arrange for the
characteristic polynomial to be equal to
(s ; 0 )(s ; 1 )


(s ; n;1 );
which is a polynomial with real coecient since E is symmetric. Note that
the choice of these scalars F~k is unique. Thus we have assigned the eigen-
values of A~ + B~ F~ to be the set E . Recalling that A~ is related to A by a
similarity transformation we have
T (A~ + B~ F~ )T ;1 = A + B (FT ~ ;1 ) =: A + BF
where F is dened to by FT ~ ;1 . Therefore the eigenvalues of A + BF have
been successfully assigned. We have the following general result.
Theorem 2.17. Suppose B 2 Rn1 . The eigenvalues of A + BF can be
arbitrarily assigned if and only if (A; B ) is controllable.
That controllability ensures the assignability of eigenvalues can be seen
from the preceding discussion. We leave the \only if" part of the proof as
an exercise. Hint: transform the system to controllability form. We are now
ready to deal with the harder proof of the multi input case.

2.3.2 Multi input case

In this section we treat the case of eigenvalue assignment when B 2 Rnm
and m  1; so far we have only looked at this problem with m = 1. As
we showed for the single input case, it turns out that (A; B ) controllable
means we can assign A + BF freely. Also if (A; B ) is not controllable, then
some eigenvalues of A + BF remain xed no matter how we choose F .
The rst step to proving the general multi input result is the following
key lemma.
Lemma 2.18. Suppose (A; B) is controllable. If B1 2 Rn1 is any nonzero
column of B , then there exists an F1 such that the matrix pair (A+BF1 ; B1 )
is controllable.
76 2. State Space System Theory
The proof of this lemma is rather involved, and so we show its application
in the following major theorem rst. This theorem is called the eigenvalue
or pole placement theorem.
Theorem 2.19. The eigenvalues of A + BF are freely assignable, by
choosing F 2 Rmn , if and only if the pair (A; B ) is controllable.
Proof . To show \if" start by choosing any nonzero column of B. Now by
Lemma 2.18 there exists a matrix F1 such that (A+BF1 ; B1 ) is controllable.
Having chosen such an F1 we can invoke Theorem 2.17 to ensure the
existence of a matrix F2 such that
(A + BF1 ) + B1 F2
has any desired eigenvalue assignment. Finally using the denition of B1 ,
choose a matrix F such that A + BF equals (A + BF1 ) + B1 F2 above.
As with the single input case we leave the \only if" part of the proof as
an exercise.

We remark here that unlike the single input case, in the multi input case
the state feedback matrix F may not be unique. Having established that
Lemma 2.18 is indeed worth proving, we are ready to proceed to its proof.
Proof of Lemma 2.18. The proof revolves around the articial discrete
time system
xk+1 = Axk + Buk ; with the initial condition x0 = B1 ; (2.8)
which we dene for the sole purpose of simplifying our notation in
this proof. Suppose we can prove that there exists an input sequence
fu0; : : : ; un;2 g such that the resulting state trajectory
fx0 ; : : : ; xn;1 g spans Rn :
Then let the matrix F1 be a solution to
   
F1 x0 x1


xn;1 = u0 u1


un;1 ; (2.9)
where un;1 is any arbitrary vector. It is then easy to see that (A + BF1 ; B1 )
is controllable: observe from (2.8) and (2.9) that
xk+1 = (A + BF1 )xk x0 = B1 ;
and therefore
   
x0


xn;1 = B1 (A + BF1 )B1


(A + BF1 )n;1 B1 :
Since the left hand side has full rank so must the right hand side, which
establishes the claim.
Thus it is sucient to show that there exists input sequence fu0; : : : ; un;2 g
such that fx0 ; : : : ; xn;1 g are linearly independent. We use induction. Sup-
pose that k < n and fu0; : : : ; uk;1 g results in the linearly independent
 2.4. Observability 77
sequence
fx0 ; : : : ; xk g:
 
Dene X = Im x0


xk  Rn , and therefore we need to show that
there exists a new input vector uk such that xk+1 is not in the subspace
X . That is there exists uk such that
Axk + Buk 62 X : (2.10)
We will prove this by contradiction.
Suppose on the contrary that for any choice of the k-th input uk 2 Rn
we have
Axk + Buk 2 X :
Then in particular Axk is in X for uk = 0, and since X is a subspace we
have that
Buk 2 X
for all uk 2 Rm . This latter condition means ImB  X .
We now show that X must be A-invariant, or equivalently
Axj 2 X for 0  j  k :
This holds for j = k from our assumption above. For j < k we note that
Axj = xj+1 ; Buj :
Observe that this means the right hand side is in X , since xj+1 2 X and
Buj 2 X by assumption. Therefore X is A-invariant. Now this means
ImAB  AX  X and more generally that
ImAj B  X for 0  j  n ; 1 :
This implies
 
Im B AB


An;1 B  X :
Since X is of dimension less than n we have a contradiction since (A; B ) is
by hypothesis controllable. Therefore (2.10) must be true.

This proof tells us how to explicitly construct F , however much better
numerical methods exist; a complete discussion of numerically reliable
techniques would take us too far aeld. This concludes our study of
controllability.

2.4 Observability
In the last section we studied a special case of the system equations pre-
sented in (2.1), which had an input but no output. We will now in contrast
78 2. State Space System Theory
consider a system with an output but no input, the system is
x_ (t) = Ax(t); with x(0) = x0
y(t) = Cx(t):
This system has no input, and its solution depends entirely on the initial
condition x(0) = x0 . The solution of this equation is clearly
y(t) = CeAt x0 ; for t  0:
We regard the function y(t) as an output and will now focus on the question
of whether we can determine the value of x0 by observing the variable y(t)
over a time interval.

2.4.1 The unobservable subspace

In this section we take the perspective of an onlooker who is only able to
measure or observe the output function y(t) over a nite interval [0; T ],
but wants to nd the value of x0 . We can regard the expression for y as
a map from the state space Rn to the space C ([0; T ]; Rp ) of continuous
functions on [0; T ] that are Rp -valued; that is : Rn ! C ([0; T ]; Rp ) and
is dened by

x0 7! CeAt x0 :
So we have that
y = x0
and our task is, if possible, to determine x0 given the function y. Now this
is possible exactly when there is only one solution to the above equation.
When y is a given solution we can be sure that it is the only solution, if
and only if the kernel of is zero. Namely
ker = 0 :
Thus if this condition is met we will be able to determine x0 , and if it is
violated then the initial conditions cannot be unambiguously discerned by
observing y. For this reason we say the matrix pair (C; A) is observable
if ker = 0. The following result gives us an explicit test for checking
observability.
Theorem 2.20. The kernel of is given by
2 3
C
6 CA 7
ker = ker C \


\ ker CAn;1 = ker 664 .. 7
7 :
. 5
CAn;1
 2.4. Observability 79
The subspace ker is usually denoted NCA and is called the unobservable
subspace of the pair (C; A) for reasons we will soon see. Thus a system is
observable when NCA = 0. The matrix on the right in the theorem is named
the observability matrix; from the theorem, observability is equivalent to
this matrix having full column rank.
Proof . We rst show that ker  ker C \


\ CAn;1 . Let x0 2 ker
and therefore by denition we know that
CeAt x0 = 0 for t  0:
It is straightforward to verify that for each k  0 the following holds

dk CeAt x  = CAk x :
dtk 0 0
t=0
But by the equation that precedes the above we see that the left hand-side
must be zero, and so x0 must be in the kernel of any matrix CAk where k
is non negative.
To complete our proof we must show that ker C \


\ker CAn;1  ker .
By Lemma 2.5 we know there exist scalar functions k (t) such that
eAt = 0 (t)I +


+ n;1 (t)An;1
for t  0. From this it is clear that if x0 is in ker C \


\ ker CAn;1 then
CeAt x0 =0. 
Although observability and controllability have been motivated in very
dierent ways they are intimately related algebraically.
Proposition 2.21. The following are equivalent
(a) (C; A) is observable;
(b) (A ; C  ) is controllable.
Proof . Condition (b) holds when the controllability matrix
 
P := C  A C 


(A )n;1 C 
has the full rank of n. This is true, if and only if, the transpose matrix
P  has rank n. It is routine to verify that P  is exactly the observability
matrix associated with (C A). 
Having established this algebraic link between the concepts of controlla-
bility and observability, we can now easily nd analogous versions of all our
results on controllability for observability. These analogs are listed in the
following proposition, and their proofs are left as an exercise and review.
Proposition 2.22.
80 2. State Space System Theory
(a) The pair (C; A) is observable, if and only if
 
rank A ;C I = n for all  2 C ;
(b) The eigenvalues of A + LC are freely assignable by choosing L, if and
only if, (C; A) is observable;
(c) The unobservable subspace NCA is A-invariant;
(d) There exists a similarity transformation T such that

~11 0 
A  
TAT = A~ A~ ; CT ;1 = C~1 0
; 1
21 22
where (C;~ A~11 ) is observable;
(e) The matrix A~22 in part (d) is Hurwitz, if and only if, the condition
 
A ; I
rank C = n for all  2 C+ holds :
All these properties can be obtained from our work on controllability. Notice
that in part (d) the following holds.
 
T N = Im 0 :
CA I
We will call the decomposition of part (d) an observability form, because
it explicitly isolates the invariant subspace NCA . Writing out the state
equations in this form gives us particular insight:
x~_ 1 (t) = A~11 x~1 (t);
x~_ 2 (t) = A~21 x~1 (t) + A~22 x~2 (t);
 
y(t) = C~1 x~1 (t); with the initial condition xx~~1 (0) = Tx:
2 (0)
From here it is clear that x~1 only depends on the initial condition x~1 (0),
and is completely unaected by x~2 ; therefore y is entirely independent of
x~2 . For this reason we say
 the vector x~1 contains the observable state;
 the vector x~2 contains the unobservable state.
Now suppose x~1 (0) = 0. Then x~1 (t) is zero for all time,
x~_ 2 (t) = A~22 x~2 (t);
and y = 0. If A~22 is not Hurwitz we have for some initial state x~2 (0), that
y is zero for all time yet x~2 (t) does not tend to zero. In contrast if A~22 is
Hurwitz we know that x~2 (t) will tend to zero.
 2.4. Observability 81
Because this situation is important, it is worthwhile to dene a term:
when A~22 is Hurwitz we say that (C; A) is a detectable matrix pair. Part
(e) of the proposition provides a way to check detectability via a PBH type
test.
Here we have introduced the notion of observability, motivated by the
fact that we wish to determine x(0) given the function y. We now turn to
issues related to actually constructing a method for determining the state
given the system output.

2.4.2 Observers
Frequently in control problems it is of interest to obtain an estimate of the
state x(t) based on past values of the output y. From our investigation in
the last section we learned under what conditions x(0) can be determined
exactly by measuring y. If they are met there are a number of ways to do
this; we will investigate an important one in the exercises. Right now we
focus on determining asymptotic estimates for x(t).
We will pursue this goal using our full state space system
x_ (t) = Ax(t) + Bu(t); x(0) = x0 ;
y(t) = Cx(t) + Du(t):
Our objective is to nd an asymptotic approximation of x(t) given the
input u and the output y, but without knowledge of the initial condition
x(0).
There are a number of ways to address this problem, and the one we
pursue here is using a dynamical system called an observer. The system
equations for an observer are
w_ (t) = Mw(t) + Ny(t) + Pu(t); w(0) = w0
x^(t) = Qw(t) + Ry(t) + Su(t): (2.11)
So the inputs to the observer are u and y, and the output is x^. The key
property that the observer must have is that the error
!1 0
x^(t) ; x(t) t;!
for all initial conditions x(0) and w(0), and all system inputs u.
Theorem 2.23. An observer exists, if and only if, (C; A) is detectable.
Furthermore, in that case one such observer is given by
x^_ (t) = (A + LC )^x(t) ; Ly(t) + (B + LD)u(t) (2.12)
where the matrix L is chosen such that A + LC is Hurwitz.
Notice in the observer equation (2.12) that we have x^ = w, so we have
removed w for simplicity. An observer with this structure is called a full-
order Luenberger observer.
82 2. State Space System Theory
Proof . We rst show the necessity of the detectability condition, by con-
trapositive. We will show that if (C; A) is not detectable, there always exist
initial conditions x(0) and w(0), and an input u(t), where the observer er-
ror does not decay to zero. In particular, we will select u = 0, w(0) = 0,
and x(0) to excite only the unobservable states, as is now explained.
Without loss of generality we will assume the system is already in observ-
ability form. For the general case, a state-transformation must be added to
the following argument; this is left as an exercise. With this assumption,
and u = 0, we have the equations
x_ 1 (t) = A11 x1 (t)
x_ 2 (t) = A21 x1 (t) + A22 x2 (t)
y(t) = C1 x1 (t):
Notice that if (C; A) is not detectable, A22 is not a Hurwitz matrix.
Therefore x2 (0) can be chosen so that the solution to
x_ 2 (t) = A22 x2 (t)
does not tend to zero as t ! 1. Now choosing x1 (0) = 0, it is clear that
x1 (t), and also y(t), will be identically zero for all time, while x2 (t) evolves
according to the autonomous equation and does not tend to zero.
Now turning to a generic observer equation of the form (2.11), we see
that if w(0) = 0, then w(t) and therefore x^(t) are identically zero. This
contradicts the fact that the error x^(t) ; x(t) tends to zero, so we have
shown necessity.
The proof of suciency is constructive. We know that if (C; A) is de-
tectable, L can be chosen to make A + LC a Hurwitz matrix. With this
choice, we construct the observer in (2.12). Substituting the expression for
y into (2.12) we have
x^_ (t) = (A + LC )^x(t) ; L(Cx(t) + Du(t) ) + (B + LD)u(t)
= (A + LC )^x(t) ; LCx(t) + Bu(t):
Now if we subtract the state equation from this, we obtain
x^_ (t) ; x_ (t) = (A + LC )^x(t) ; LCx(t) + Bu(t) ; Ax(t) ; Bu(t)
= (A + LC )(^x(t) ; x(t) )
This means that the error e = x^ ; x satises the autonomous dierential
equation
e_(t) = (A + LC )e(t)
Since A + LC is a Hurwitz matrix, e(t) tends to zero as required. 
 2.4. Observability 83
2.4.3 Observer-Based Controllers
We will now brie y exhibit a rst example of feedback control, combin-
ing the ideas of the previous sections. Feedback consists of connecting a
controller of the form
x_ K (t) = AK xK (t) + BK y(t)
u(t) = CK xK (t) + DK y(t)
to the open-loop system (2.1), so that the combined system achieves some
desired properties. A fundamental requirement is the stability of the re-
sulting autonomous system. We have already seen that in the case where
the state is available for feedback (i.e. when y = x), then a static control
law of the form
u(t) = Fx(t)
can be used to achieve desired closed-loop eigenvalues, provided that (A; B )
is controllable. Or if it is at least stabilizable, the closed loop can be made
internally stable.
The question arises as to whether we can achieve similar properties for
the general case, where the output contains only partial information on the
state. The answer, not surprisingly, will combine controllability properties
with observability of the state from the given output. We state the following
result.
Proposition 2.24. Given open-loop system (2.1), the controller
x^_ (t) = (A + LC + BF + LDF )^x(t) ; Ly(t)
u(t) = F x^(t)
is such that the closed loop system has eigenvalues exactly at the eigenvalues
of the matrices A + BF and A + LC .
Notice that given the choice of u, the rst equation can be rewritten as
x^_ (t) = (A + LC )^x(t) ; Ly(t) + (B + LD)u(t)
and has exactly the structure of a Luenberger observer (2.12) for the state
x. Given this observation, the expression for u can be interpreted as be-
ing analogous to the state feedback law considered above, except that the
estimate x^ is used in place of the unavailable state.
Proof . Combining the open loop system with the controller, and elimi-
nating the variables u, y, leads to the combined equations
    
x_ (t) = A BF x(t)
x^_ (t) ;LC A + LC + BF x^(t)
Now the similarity transformation
 
T = I0 ;II
84 2. State Space System Theory
(which amounts to replacing x by x ; x^ as a state variable) leads to
   
A BF ; 1
T ;LC A + LC + BF T = ;LC A + BF A + LC 0

which exhibits the eigenvalue structure. 

As a consequence of this result, we conclude that controllability and
observability of (A; B; C; D) suces to place the closed-loop eigenvalues
at any desired locations, whereas stabilizability and detectability suce to
obtain an internally stabilizing feedback law. In addition, notice that the
state-feedback and observer components appear in a decoupled fashion, so
they can be designed independently; this property is commonly called the
separation principle.

2.5 Minimal realizations

In the rest of the chapter we look at the state-space system
x_ (t) = Ax(t) + Bu(t); x(0) = x0
y(t) = Cx(t) + Du(t)
from the point of view of the relationship it establishes between inputs and
outputs. We focus on the case where the system is initially at rest, with
x(0) = 0, and see that in this case the system establishes a map between
inputs and outputs, given by the formula
Z t
y(t) = CeA(t; )Bu( )d + Du(t) :
0
In this section take the standpoint that this input-output map is the pri-
mary object of study, and the state-space system is merely a realization of
this map, i.e. a way to dene such a map in terms of rst-order dierential
equations. From this point of view, a natural question is to study when
realizations are equivalent. In other words when does
Z t Z t
CeA(t; )Bu( )dt + Du(t) = C1 eA1 (t; )B1 u( )d + D1 u( )
0 0
hold for all input functions u and all times t  0. In particular we will be
interested in nding, among all equivalent realizations, the one with the
minimal dynamic order, which we dene as the dimension n of the square
matrix A.
We rst characterize equivalence in the following result.
Lemma 2.25. Two system realizations (A; B; C; D) and (A1 ; B1; C1 ; D1)
are equivalent, if and only if, D = D1 and
CeAt B = C1 eA1 t B1 holds for all t  0.
 2.5. Minimal realizations 85
Proof . It is clear from the input-output formula that the said conditions
are sucient for equivalence. We now show necessity.
First, it is clear that D = D1 must hold since in particular Du(0) =
D1 u(0) for all u(0). With this in place we can now rewrite the equivalence
condition as
Z t
fCeA(t; )B ; C1 eA1 (t; )B1 gu( )d = 0; (2.13)
0
for all functions u. It thus remains to show that the factor multiplying u(t)
in (2.13) must be identically zero. For simplicity we show this in the case
where CeAt B and C1 eA1 t B1 are scalar functions; the multivariable case
follows using a very similar argument.
Suppose, by contrapositive, that for some t0  0 the functions are not
equal. Then dene
u(t) = jCeA(t0 +1;t) B ; C1 eA1 (t0 +1;t) B1 j
and notice that u(1) > 0. Thus applying this input to the condition in
(2.13), with t = t0 + 1, we see that
Z t0 +1 Z t0 +1
fCeA(t0 +1; )B ; C1 eA1 (t0 +1; )B1 gu( )d = ju( )j2 d > 0;
0 0
where the last inequality follows from the fact that u is a continuous
function and u(1) > 0. Thus we have a contradiction and necessity
follows. 
This lemma reduces the equivalence of realizations to checking that two
functions of time are equal. We now provide an alternative characterization,
in terms of the matrices CAk B , which are called the Markov parameters
of the system realization (A; B; C; D).
Lemma 2.26. The two realizations (A; B; C; D) and (A1 ; B1; C1; D1) are
equivalent, if and only if D = D1 and
CAk B = C1 Ak1 B1 for all k  0: (2.14)
Proof . By Lemma 2.25 we must show the equivalence of (2.14) with
CeAt B = C1 eA1 t B1 :
for all t  0. The suciency of (2.14) follows clearly by denition of the
matrix exponential, since
2
CeAt B = CB + CABt + CA2 B t2 +



For the converse, assume the time functions are equal and notice that the
k-th (right) derivative of CeAt B is given by
dk CeAt B = CAk eAt B;
dtk
86 2. State Space System Theory
with a similar expression for the derivatives of C1 eA1 t B1 . These derivatives
all must be equal, so we have
 
dk CeAt B  = CAk B = C Ak B = dk C eA1 t B  as required.
dtk  1 1 1 dtk 1 1
t=0 t=0

Now we turn to the question of minimality: given a realization
(A; B; C; D) of order n, does there exist an equivalent realization (A1 ; B1 ; C1 ; D)
of order n1 < n? The next theorem tells us that this question is related to
observability and controllability.
Proposition 2.27. Suppose (A; B; C; D) is a system realization. If either
(C; A) is not observable or (A; B ) is not controllable, then there exists a
lower order realization (A1 ; B1 ; C1 ; D) for the system.
Proof . There are two cases to consider: (A; B) not controllable and (C; A)
unobservable. We will assume that (A; B ) is not controllable since the unob-
servable case is nearly identical in proof. With this assumption we convert
to controllability form. That is choose a similarity transformation T such
that

A~11 A~12   
~
TAT = 0 A~ and TB = B01 ;
;1
22
where the dimensions of A~22 are nonzero. Let
 
CT ;1 = C~1 C~2
and we see that
  
 A~11 t ~
CeAt B = C~1 C~2 e 0 ?? B01


= C~1 eA~11 t B~1 :

Thus setting A1 = A~11 ; B1 = B~1 and C1 = C~1 and invoking Lemma 2.25
we have found a lower order realization. 
The result above says that if a realization is not both controllable and
observable, then there exists a lower order equivalent realization. Immedi-
ately we wonder whether controllability and observability of a realization
guarantees that it is of the lowest order; such a lowest order realization is
called a minimal realization. The answer is armative, as we show in the
following result.
Theorem 2.28. If (C; A) is observable and (A; B) is controllable, then
(A; B; C; D) is a minimal realization.
Proof . We will show that if (A1 ; B1; C1; D1) is a realization which is equiv-
alent to (A; B; C; D), then the dimension of A1 is no smaller than that of
A.
 2.6. Transfer functions and state space 87
From Lemma 2.26 we know that
CAk B = C1 Ak1 B1 for k  0 :
Therefore in particular we have that
2 3
C
6 CA 7 
6
6 .. 7
7 B AB


An;1 B
4 . 5
CAn;1
is equal to
2 3
C1
6 C1 A1 7 
6
6 .. 7
7 B1 A1 B1


An1 ;1 B1 :
4 . 5
C1 An1 ;1
Now both the matrices on the left side of the equation have the full rank
of n; thus their product has rank n. This means that the right hand-side
has rank n, and simply by the dimensions we see that A1 must be at least
n  n in size. 
We have completed our study of minimal realizations for input-output
systems with only time domain considerations. We now turn to an alterna-
tive characterization of systems which is of great importance and will be
used throughout the course.

2.6 Transfer functions and state space

An alternative method for the study of systems from an input-output per-
spective, is to employ transforms of the relevant signals. In particular, the
Laplace transform of a time function f (t) is dened as
Z 1
f^(s) := f (t)e;st dt
0
where s varies in a region of the complex plane C where the above integral
converges. Thus we dene a linear transformation between functions of
t and functions of s. We will not pursue here the question of describing
function spaces in which this transform is well-dened; an important case
will be discussed in detail in the next chapter. For the moment assume that
all time functions of interest increase no faster than exponentially. More
precisely, given a function f assume that there exist real constants  and
 satisfying
jf (t)j  et for all t  0:
88 2. State Space System Theory
In this case the above transform is well dened for Re(s) > .
A basic property is that if the time derivative f_(t) has a well dened
Laplace transform, the latter is given by
sf^(s) ; f (0):
This property can be used to derive a Laplace transform version of our
basic state-space system (2.1) of the form
sx^(s) ; x(0) = Ax^(s) + B u^(s);
y^(s) = C x^(s) + Du^(s):
In the special case of x(0) = 0, the input output relationship
y^(s) = G^ (s)^u(s) (2.15)
follows, where
G^ (s) = C (Is ; A);1 B + D
is called the transfer function of the system. This function is well dened
whenever the matrix (Is ; A) has an inverse, and in particular the identity
(2.15) will hold in some right-half plane Re(s) > , where this inverse, as
well as the transforms y^(s) and u^(s), are well-dened.
We can also relate the relationship (2.15) to the time domain description
Z t
y(t) = CeA(t; )Bu( )d + Du(t) (2.16)
0
considered in the previous section. To do this, rst notice that the transform
of the term Du(t) is Du^(s). Next we can show that (Is;A);1 is the Laplace
transform of the matrix eAt . For this purpose consider
Re(s) > max Re( eig(A) );
and write
Z 1 Z 1 Z 1 d h i
eAt e;st dt = e;(sI ;A)t dt = ;(sI ; A);1 dt e;(sI ;A)t dt
0 0   0
;
= (sI ; A) I ; t!lim
1 e ; (sI ; A )t = (sI ; A);1
+1
where the last identity follows from the fact that since the eigenvalues of
;(Is ; A) have all negative real parts, all the entries in the matrix will
have exponentially decaying terms. We refer to the study of the formula
for the matrix exponential, leaving details as an exercise.
It follows immediately from linearity that C (Is ; A);1 B is the transform
of CeAt B . Then the rst term in (2.16) can be interpreted as a convolution
integral between CeAt B and u(t), whose transform will be the product
C (Is ; A);1 B u^(s)
 2.6. Transfer functions and state space 89
completing the relationship (2.15). We will not expand here on the issue
of convolutions and their relationship with transforms, since we return to
this topic in the next chapter.
The transfer function provides yet another version of the equivalence
between two state-space realizations.
Lemma 2.29. The two realizations (A; B; C; D) and (A1 ; B1; C1; D1) are
equivalent, if and only if
G^ (s) = C (Is ; A);1 B + D = C1 (Is ; A1 );1 B1 + D1 = G^ 1 (s) ; (2.17)
for all s where the inverses are well dened
Proof . One can argue that this must be true based on the input-output
relationship (2.15). Alternatively, using Lemma 2.25 and the transform of
CeAt B computed above, it suces to show that G^ (s) = G^ 1 (s) if and only
if
D = D1 ; C (Is ; A);1 B = C1 (Is ; A1 );1 B1 :
Here the \if" direction is obvious, and \only if" follows by noticing that
D = slim ^ ^
!1 G(s) = slim
!1 G1 (s) = D1 ;
which in turn uses the fact that
lim ( Is ; A ) ;1 = lim 1 (I ; A=s);1 = 0 ;
s!1 s!1 s
and similarly with A1 . 

2.6.1 Real-rational matrices and state space realizations

We now discuss the structure of the transfer function G^ (s), constituted by
real rational functions of the variable s.
A rational function is the quotient of two real polynomials, of the general
form
m;1
g^(s) = sbnm+s a +sn
;

1++b
1
s
++ ba0
n+1 0
for appropriate scalars ak and bk , where m and n are the order of numerator
and denominator polynomials. The function is called real rational if the ak ,
bk are real numbers. This function is dened everywhere in the complex
plane, except at the roots of the denominator polynomial. Recall that a
function has a pole in the complex plane if it tends to innity as that point
is approached; for the rational function g^(s) poles can only occur at the
roots of its denominator.
We say that a rational function is proper if n  m, and strictly proper
if n > m. Notice these are equivalent to the statements
lim g^(s) exists in C and slim
s!1 !1 g^(s) = 0 respectively.
90 2. State Space System Theory
Our interest here is primarily in rational, proper functions. We extend our
notion of such scalar functions to the matrix case: suppose
2 3
g^11 (s)


g^m1 (s)
G^ (s) = 64 ... ... .. 75
.
g^1p (s)


g^mp (s)
where each g^ij (s) is a scalar function. Then we dene G^ (s) to be real
rational and proper if each scalar function g^ij (s) has these properties. We
say that G^ (s) has a pole at a point if one of its scalar entries does.
Now returning to our transfer function, we see that it is indeed rational
since (Is ; A);1 can be written using Cramer's rule as
(Is ; A);1 = det(Is1 ; A) adj(Is ; A);
where adj (
) denotes the classical adjoint of a matrix and is formed of
cofactors of elements of the matrix. Therefore adj(Is ; A) is a matrix of
polynomial entries, each of which has order less than n. The determinant
det(Is ; A) is also a polynomial, indeed the characteristic polynomial of
A, and has order n. Therefore (Is ; A);1 is a strictly proper, real rational
matrix, and it follows that C (Is ; A);1 B + D is a proper, real rational
matrix.
We now turn to the converse question: given a real rational transfer
function, is it always the transfer function of some state-space system?
The next proposition establishes the converse in the strictly proper scalar
case.
Lemma 2.30. Suppose that
n;1
g^1 (s) = sn c+n;a1 s sn;+1
+


+

c+0 a ;
n;1 0
then there exists a state space realization (A; B; C; 0) such that
g^1 (s) = C (Is ; A);1 B; where A is n  n.
Proof . Let
2 3 2 3
0 1 0 0
6 ... ... 7 6.7
6 .. 7
C = [c0


cn;1 ]; A = 6
6
7
7 ; and B = 6 7 :
4 0 0 1 5 405
;a0 ;a1 ;an;1 1
Then it is routine to verify that this realization satises the claim. 
From this lemma we can prove the more general scalar result.
 2.6. Transfer functions and state space 91
Proposition 2.31. If g^(s) is a proper rational function given by
n;1
g^(s) = csnns ++a cn;s1ns;1 ++





++ac0 ;
n
n;1 0
then it has a state space realization (A; B; C; D) where A is n  n.
Proof . First since g^(s) is proper it is possible to write it as
g^(s) = g^1 (s) + D;
where g^1 is strictly proper and D is a real constant. Now invoke
Lemma 2.30. 
Having treated the scalar case, we now turn to the matrix case, starting
with the following fact. Consider two transfer functions G^ 1 (s) = C1 (Is ;
A1 );1 B1 + D1 and G^ 2 (s) = C2 (Is ; A2 );1 B2 + D2 . If we stack them to
form a larger transfer function the following can be veried algebraically.
   
G^ 1 (s) = C1 (Is ; A1 );1 B1 + D1
G^ 2 (s) C2 (Is ; A2 );1 B2 + D2
   ;1    
C1 0
= 0 C Is ; 0 A A 1 0 B1 + D1 :
2 2 B2 D2
That is a block column with two transfer function entries is exactly a trans-
fer function. Similarly for appropriately dimensioned transfer functions we
have
  ;1  
G^ 1 (s) G^ 2 (s) = C1 C2 Is ; A01 A02 B1 0 + D D  :
   
0 B2 1 2

Consequently if we are given a matrix G^ (s) of proper, real rational

functions, we can begin by nding realizations
g^ij (s) = Cij (Is ; Aij );1 Bij + Dij :
for each entry using Proposition 2.31, and then use the previous rules
successively to form a state-space realization for G^ (s). We thus have the
following result, which summarizes the previous discussion.
Proposition 2.32. The set of proper real rational functions is equal to the
set of state space transfer functions.
The above construction provides a concrete algorithm for constructing a
state space realization for a rational proper function. However the proce-
dure may not be economical from the point of view of the dimension of the
state space system. These issues are discussed next.
92 2. State Space System Theory
2.6.2 Minimality
In the above algorithm we assign separate states to each entry of the matrix,
which may lead to an unnecessarily high state dimension. This fact is now
illustrated.
Example:
Consider the transfer function
 1 2 
G^ (s) = s+1 s+1 :
The previous construction would lead to
   
A = ;01 ;01 ; B = 10 01 ; C = 1 2 ;
 

but clearly a lower order realization is given by

 
A = ;1; B = 1; C = 1 2 :

This raises once more the question of minimality of a state-space realiza-
tion, in this case as a representation of a rational transfer function. As
before a realization of a rational function G^ (s) is minimal if it has the
lowest possible dimension n for the A-matrix.
How does lack of minimality re ect itself in the realization? In principle
the fact that (A; B; C; D) has more states than are necessary to describe
G^ (s) would imply that pole-zero cancellations will occur if we calculate
 
1
C char (s) adj (Is ; A) B + D ;
A
since this expression must be the same as the one obtained with a lower
order characteristic polynomial in the denominator.
In the case of a scalar G^ (s), it can be shown that minimality is equivalent
to there being no pole-zero cancellations occurring when the state space
transfer function is computed. For the matrix case however the issue is
more complicated.
Examples:
For example it is not immediately obvious that
"
1 0#
G^ 1 (s) = 1s 1
s2 s
can be realized with n = 2, whereas
" #
0 1
G^ 2 (s) = 1 1s
s2 s
 2.7. Exercises 93
requires that n be at least three, when in both cases the algorithm described
above would lead to n = 4. 
Therefore, a necessary complement to the above constructive procedure
is a test for the minimality of a given state-space realization of a transfer
function, and a procedure to remove unnecessary states. Fortunately we
have already provided the tools to address this problem.
Theorem 2.33. A realization (A; B; C; D) of a transfer function G^(s) is
minimal, if and only if, (A; B ) is controllable and (C; A) is observable.
Proof . We have already established that equality of transfer functions is
the same as equivalence of the corresponding input-output maps. Therefore
the proof reduces to Proposition 2.27 and Theorem 2.28. 
In addition to being able to test for minimality, our work in the earlier
sections provides a method to reduce the order of a non-minimal realiza-
tion, namely eliminating uncontrollable or unobservable states. This can be
easily done by transforming the realization to the controllability form or
to the observability form (these steps can be done successively). As an ex-
ercise, the reader is invited to construct minimal realizations for the above
transfer functions G^ 1 (s) and G^ 2 (s).
This brings us to the end of our work on basic state space system theory.
We have provided algebraic methods to describe a general class of linear
systems, both in the time domain and in the transfer function method.
In the next chapter we will move beyond the algebraic and attempt to
establish quantitative descriptions of signals and systems, from an input-
output perspective.

2.7 Exercises
1. Consider the following set of dierential equations
u1 + c u2 = q1 + b q_2 + q1 ;
u2 = q2 + d q1 + e q_2 + f q1 ;
where the dependent functions are q1 and q2 . Convert this to a 2  2
state space system.
2. Suppose A is Hermitian and is Hurwitz. Show that if x(t) = eAt x(0)
then jx(t)j < jx(0)j, for each t > 0. Here j
j denotes the Euclidean
length of the vector.
3. Prove that e(N +M )t = eNteMt if N and M commute. Use the
following steps:
(a) Show that eNt and M commute;
94 2. State Space System Theory
(b) By taking derivatives directly show that the time derivative of
Q(t) = eNt eMt satises Q_ (t) = (N + M )eNteMt ;
(c) Since Q(t) satises Q_ (t) = (N + M )Q(t), with Q(0) = I , show
that Q(t) = e(N +M )t .
Notice in particular this demonstrates eN +M = eN eM .
4. Using Jordan decomposition prove Proposition 2.4.
5. In the proof of Lemma 2.8 we use that fact that given two subspaces
V1 and V2 , then V1  V2 if and only if V1?  V2? . Prove this fact.
Hint: rst show that for any subspace V = (V ? )? .
6. Prove Proposition 2.7.
7. Using a change of basis transform the pair
2 3 2 3
0 1 0 1 1 1
A = 64 00 02 ;11 00 75 B = 64 11 00 75
6 7 6 7

0 ;1 1 1 0 0
   
~ ~12 B~1 where (A~11 ; B~1 ) is
to the form A~ = A011 A A~22 and B~ = 0
controllable.
8. Fill in the details of the proof of Theorem 2.12 on controllability form.
9. (a) Consider the discrete time state equation
xk+1 = Axk + Buk ; with x0 = 0:
A state  2 Rn is said to be reachable if there exists a sequence
u0 ; : : : ; uN ;1 such that xN =  .
(i) Show that   reachable if and only
 if it is in the image of the
matrix B AB


An;1 B ;
(ii) If  is reachable the length N of the sequence uk can be
chosen to be at most n.
(b) Given the uncontrolled discrete time state space system
xk+1 = Axk ; with initial condition x0
yk = Cxk ;
we say it is observable if each x0 gives rise to a unique output
sequence yk .
(i) Show that the above system is observable if and only if
(A ; C  ) is controllable;
(ii) Prove that if the system is observable, the initial condition
x0 can be determined from the nite sequence y0 ; : : : ; yn;1 .
10. Provide the details of a proof for Corollary 2.15.
 2.7. Exercises 95
11. Using our proofs on controllability and stabilizability as a guide, prove
all parts of Proposition 2.22.
12. Given a pair (C; A), devise a test to determine whether it is possible
to nd a matrix L so that all the eigenvalues of A + LC are equal to
;1; you should not have to explicitly construct L.
13. In this question we derive the famous Kalman decomposition, which
is a generlization of the controllability and oberservability forms we
have studied in this chapter. Suppose we are given a matrix triple
(C; A; B ).
(a) Prove that the subspace intersection CAB \ NCA is A-invariant;
(b) Show that there exists a similarity transformation T so that
2 3 2 3
A1 0 A6 0 B1
T ;1AT = 64 02 03 A4 05 75 ; T ;1B = 64B02 75 ;
A A A A
6 7 6 7
7
0 0 A8 A9 0
 
and CT = C1 0 C2 0 . Furthermore the following proper-
ties are satised.
 the pair (C1 ; A1 ) is observable;
 the pair 
(A1 ; B1 ) iscontrollable;
 
 the pair A A ; BB1 is controllable;
A 1
2 3
0
2
  
 the pair C1 ; C2 ; A01 AA67 is observable.
 

To do this use the A-invariance properties of the controllable

subspace CAB and the unobservable subspace NCA , and use T
to decompose the state space into the four independent sub-
spaces that are: controllable and observable; controllable and
unobservable; observable and uncontrollable; unobservable and
uncontrollable.
14. Given the system matrices
2 3 2 3
2 2 1 2 2
A = 4 1 1 1 5; B = 4 2 1 5
4 1 0 3 2
C = [2 1 0];
determine whether it is possible to construct an observer for the
system
x_ = Ax + Bu x(0) = x0
y = Cx:
96 2. State Space System Theory
15. (i) Given the system
x_ (t) = A x(t) + B u(t); x(0) = x0 :
Provide an explicit solution for the state x(1) in terms of u(t).
(ii) Suppose in the system of (i) that
2
;1 2 0 3 2
1
3 2 ;3 3
e
A = 4 ;4 5 0 5 B = 4 1 5 x(0) = 4 2e;3 5
0 0 4 0 e;4
Does there exist a function u(t) so that
2 3
11
x(1) = 4 12 5?
1
Give full justication for your answer, but note that you do not
necessarily have to nd u(t) explicitly.
16. Recall the double pendulum system of the introductory chapter.
Discuss whether there any equilibrium points around which the sys-
tem linearization loses controllability. Is it stabilizable around these
points?

Notes and references

The material covered in this chapter is standard, although the presenta-
tion here is largely based on [40]. For this reason we primarily highlight the
origin of the ideas and concepts. The fundamental ideas of controllability
and observability introduced in this chapter are due to Kalman [64]. The
PBH test, as its name suggests was developed independently by the three
researchers in [102, 8, 55]. The well-known result on controllable canoni-
cal form presented in conjunction with single input eigenvalue assignment
originates from [101]. The rst proof of the multivariable eigenvalue assign-
ment result appears in [143]. The controllability and observabilty forms of
this chapter are special cases of the so-called Kalman decomposition which
was introduced in [65, 45]. This also led to the notion of system minimality.
The key idea involved in constructing the Luenberger observer presented
here was rst reported in [78].
For more detailed historical references on time domain state space theory
than provided above, and an in depth look at state space systems from
a geometric perspective, see [144]. In the nal section of this chapter we
introduced the basics of rational function theory and connections with state
space realizations; a more complete treatment in a systems context can be
found in [62].
 This is page 97
Printer: Opaque this

3
Linear Analysis

One of the prevailing viewpoints for the study of systems and signals is
that in which a dynamical system is viewed as a mapping between input
and output functions. This concept underlies most of the basic treatments
of signal processing, communications, and control. Although a functional
analytic perspective is implicitly used in these areas, the associated ma-
chinery is not typically used directly for the study of dynamical systems.
However incorporating more machinery from analysis (function spaces, op-
erators) into this conceptual picture, leads to methods of key importance
for the study of systems. In particular, operator norms provide a natural
way to quantify the \size" of a system, a fundamental requirement for a
quantitative theory of system uncertainty and model approximation.
In this chapter we introduce some of the basic concepts from analysis
that are required for the development of robust control theory. This involves
assembling some denitions, providing examples of the central objects and
presenting their important properties. This is the most mathematically
abstract chapter of the course, and is intended to provide us with a solid
and rigorous framework. At the same time, we will not attempt a completely
self contained mathematical treatment, which would constitute a course in
itself. To build intuition, the reader is encouraged to supply examples and
proofs in a number of places. For detailed proofs of the most technical
results presented consult the references listed at the end of the chapter.
During the chapter we will encounter the terminology \for almost ev-
ery" and \essential supremum", which refer to advanced mathematical
concepts that are not part of the core of our course. These clauses are
used to make our statements precise, but can be replaced by \for every"
98 3. Linear Analysis
and \supremum", respectively, without compromising understanding of the
material. If these former terms are unfamiliar, precise denitions and a
short introduction to them are given in Appendix A.

3.1 Normed and inner product spaces

The most important mathematical concept in this course is that of a
normed space, which we will use continually to measure both signals and
systems. We will start by dening exactly what we mean by a norm, and
then move on to some examples.
Denition 3.1. A norm k
kV on a vector space V is a function mapping
V ! [0; 1) which, for each v 2 V , satises
(a) kvkV = 0 if and only if v = 0;
(b) jj
kvkV = kvkV , for all scalars ;
(c) ku + vkV  kukV + kvkV , for all u 2 V .
Dening a norm on a vector space is done to make the notion of the \size"
of an element precise, that is the size of v is kvkV . With this in mind
condition (a) says that the zero element is the only member of V which
has size zero. The second norm requirement (b) means that the size of v
must scale linearly according to the sizes of  and v. The third condition
is called the triangle inequality, inspired by the case of Euclidean space: if
u and v are vectors that specify two sides of a triangle, then (c) states that
the third side u + v must have length no larger than the sum of the lengths
of the other two.
A vector space together with a norm is called a normed space and is
denoted (V ; k
kV ). Frequently when the norm and space is clear we simply
write k
k, suppressing the subscript on the norm symbol. Understanding
normed spaces will provide us with a powerful point of view, and we now
begin to explore this new concept.
We say a sequence vk 2 V converges in a normed space V if there exists
v 2 V such that
!1 0:
kv ; vk k k;!
We use limk!1 vk = v to denote the above relationship; thus this preserves
our usual notation for limits with numbers, although our interpretation is
slightly more involved.
Examples:
In Chapter 1 we have already encountered the Euclidean norm on the nite
dimensional space C n . A variety of other norms can be put on this space;
 3.1. Normed and inner product spaces 99
a standard family is formed by the so-called p-norms. Given a p  1 we
dene the p-norm
jvjp := (jv1 jp +


+ jvn jp ) p1 ; where v 2 C n :
Notice that we use the notation j
jp instead of k
kp ; the reason for this
convention is to avoid confusion with function norms, dened below. When
the sub-index is suppressed, we will assume by default that we are referring
to p = 2, that is the Euclidean norm jvj = jvj2 .
This family of norms is extended to the case where p = 1 by dening
the 1-norm to be
jvj1 := 1maxk n k
jv j:
We can also dene norms for matrices, in particular the Frobenius norm
is dened as
jM jF := (TrM  M ) 12 for M 2 C mn ,
and the maximum singular value from Chapter 1
1
 (M ) := (maximum eigenvalue of M  M ) 2 :
is also a norm. Once again we use the notations j
jF and  (
) instead of
the norm sign k
k.
Our primary interest is norms dened on spaces of functions. For p  1,
let Lnp (;1; 1) denote the vector space of functions mapping R to C n that
satisfy
Z 1
ju(t)jpp dt < 1;
;1
where j
jp denotes the vector norm on C n dened above. Then
Z 1 1
p
kukp := p
ju(t)jp dt
;1
is a norm on this space, and is the one tacitly associated with it. The
superscript n is called the spatial dimension of the space Lnp(;1; 1). In
most cases we will not need to keep track of this dimension and so we
simply write Lp (;1; 1).
We complete the family of Lp spaces with the normed space L1 (;1; 1)
consisting of functions such that
kuk1 := ess sup ju(t)j1
t2R
is nite. Once again juj1 is the previously dened vector norm.
We also dene the space Lp [0; 1) to be the set of functions in
Lp (;1; 1) with support in [0; 1), more precisely
Lp [0; 1) = fu(t) 2 Lp (;1; 1) : u(t) = 0 for t < 0g:
100 3. Linear Analysis
This convention is non-standard, more commonly Lp [0; 1) is the space of
functions with restricted domain. However it is clear that given a function
on [0; 1), one can trivially identify it with a function on the above set by
dening it to be zero for t < 0.
By adopting the above rule, all our functions will be dened for all t
and we are allowed to write Lp [0; 1)  Lp (;1; 1), in words Lp [0; 1)
is a normed subspace of Lp (;1; 1). Analogously we dene the subspace
Lp (;1; 0] of functions that are zero for t > 0. 
Another very important notion that we will frequently encounter is an
inner product, and it is closely related to the idea of a norm.
Denition 3.2. An inner product h
;
iV on a vector space V is a function
mapping V  V ! C so that
(a) The inner product hv; viV is non negative, for all v 2 V ;
(b) hv; viV = 0 if and only if v = 0;
(c) If v 2 V , then hv; 1 u1 + 2 u2i = 1 hv; u1 i + 2 hv; u2 i, for all ui 2 V
and scalars i ; i.e the mapping u 7! hv; ui is linear on V .
(d) hu; viV is the complex conjugate of hv; uiV , for all v; u 2 V .
Geometrically the inner product captures the idea of angle between two
elements of V in the same way that the so-called \dot" product does in
Euclidean space.
A vector space V together with an inner product h
;
iV is called an inner
product space. It can be veried that
p
kvk = hv; vi
satises the properties of a norm. The rst two properties are straight-
forward; the triangle inequality can be established by rst proving the
Cauchy-Schwarz inequality
jhu; vij  kuk
kvk;
both are proved in the exercises. Thus the properties of an inner product
are more restrictive than those of a norm.
We say that two elements u; v in V are orthogonal if
hu; vi = 0:
The notation u?v is sometimes used to indicate this relationship. It is easy
to show that if u?v, then
ku + vk2 = kuk2 + kvk2:
This is a generalization of Pythagoras' theorem.
 3.1. Normed and inner product spaces 101
Examples:
The standard example of an inner product space, from which much intuition
about inner product spaces is gained, is Euclidean space. That is Rn or C n
with the inner product
hx; yi := x y := x1 y1 +


+ xn yn ;
where again  denotes complex conjugate transpose.
The matrix generalization of this inner product is dened on C mn and
is given by
hA; B i := TrA B;
for two elements A and B of C mn . Notice that this inner product induces
the Frobenius norm jAjF . Also if n = 1 then this is exactly the inner
product space C m from above.
We now introduce what for us will be the most important example of
an inner product space. Starting with the space of functions L2 (;1; 1)
dened before, we introduce
Z 1
hx; yi := x (t)y(t) dt:
;1
It can be shown that this integral is well dened, and is an inner product;
also the corresponding norm coincides with the norm k
k2 dened above.
Thus L2 (;1; 1) is an inner product space. Also the subspaces L2(;1; 0]
and L2[0; 1) of functions with support in the \past" and \future" are
themselves inner product spaces. 
We are now ready to consider normed and inner product spaces which
have an additional convenient property.

3.1.1 Complete spaces

We have already talked about convergence of sequences in normed and
inner product spaces. We said that the limit limk!1 vk exists, if there
exists an element v in the inner product space, such that the numerical
limit limk!1 kv ; vk k = 0. This is a precise denition, but means that
if we are to determine whether a sequence vk converges it is in general
necessary to nd such an element v. Finding v may be cumbersome or
impossible, which motivates the topic of this short section. We start with
the following concept.
Denition 3.3. Suppose V is a normed space. A sequence vk is a Cauchy
sequence if, for each  > 0, there exists M  0 such that
kvk ; vl k < ; for all k; l  M:
102 3. Linear Analysis
This denition simply says that a sequence is Cauchy if it satises
l!1
kvk ; vl k k;;! 0:
A Cauchy sequence is a sequence that \appears" to be converging. It is
not hard to show that every convergent sequence is necessarily a Cauchy
sequence, but the converse may not be true.
A normed space is complete if every Cauchy sequence in it converges,
and such a space is referred to as a Banach space. A complete inner product
space is called a Hilbert space. Thus if we know that a space is complete,
we can denitively check the convergence of any sequence vk by simply
determining whether or not it is a Cauchy sequence. More importantly
complete spaces behave very much like the complex numbers with regard
to convergence and are thus very convenient to work with.
Examples:
There is a plethora of Banach and Hilbert space examples. Our standard
spaces C n and Rn are complete, with the norms and inner products we
have dened thus far. All the Lp spaces we have dened are complete; see
Appendix A for some additional technical detail on this. In particular L2
is an example of a Hilbert space.
An example of an inner product space that is not complete is
W = fw 2 L2 [0; 1) : there exists a T > 0 such that w(t) = 0; for all t  T g;

equipped with the L2 norm. To see this consider the sequence of functions
given by

wk (t) := e;t ; for 0  t  k,
0; otherwise:
This sequence is clearly in W , and is a Cauchy sequence since kwk ; wl k 
e; min(k;l) . The sequence wk is in L2 [0; 1) and it is easy to verify that it
converges in that space to the function

w(t) := e;t ; for 0  t,
0; otherwise:
But w is not in W , and thus W is not complete.

We have introduced the concept of completeness of a normed or inner
product space, and throughout the course we will work with spaces of this
type.
 3.2. Operators 103
3.2 Operators
As we have mentioned in the previous examples, normed spaces can be used
to characterize time domain functions, which we informally call signals.
We now examine mappings from one normed space V to another normed
space Z . These are central to the course as they will eventually be used to
represent systems. The focus of interest will be linear, bounded mappings.
Denition 3.4. Suppose V and Z are Banach spaces. A mapping from V
to Z is called a linear, bounded operator if
(a) (Linearity) F (1 v1 + 2 v2 ) = 1 F (v1 ) + 2 F (v2 ) for all v1 ; v2 2 V
and scalars 1 , 2 .
(b) (Boundedness) There exists a scalar   0, such that
kFvkZ  
kvkV for all v 2 V : (3.1)
The space of all linear, bounded operators mapping V to Z is denoted by
L( V ; Z ) ;
and we usually refer to linear, bounded operators as simply operators. We
dene the induced norm on this space by
kF kV!Z = sup kkFv v k
kZ ;
v2V ; v6=0 V
and it can be veried that it satises the properties of a norm. Notice that
kF kV!Z is the smallest number  that satises (3.1). When the spaces in-
volved are obvious we write simply kF k. It is possible to show that L(V ; Z )
is a complete space, and we take this fact for granted. If V = Z we use the
abbreviation L(V ) for L(V ; V ).
We also have a natural notion of composition of operators. If F 2 L(V ; Z )
and G 2 L(Z ; Y ) and the composition GF is dened by
(GF )v := G(Fv); for each v in V .
Clearly, GF is a linear mapping, and it is not dicult to show that
kGF kV!Y  kGkZ!Y kF kV!Z ; (3.2)
which implies GF 2 L(V ; Y ). The above submultiplicative property of in-
duced norms has great signicance in robust control theory; we shall have
more to say about it below.
Examples:
As always our simplest example comes from vectors and matrices. Given
a matrix M 2 C mn , it denes a linear operator C n ! C m by matrix
multiplication in the familiar way
z = Mv where v 2 C n :
104 3. Linear Analysis
As in Chapter 1, we will not distinguish between the matrix M and the
linear operator it denes. From the singular value decomposition in Chapter
1, if we put the 2-norm on C m and C n then
kM kCm !C n =  (M ):
If we put norms dierent from the 2-norm on C m and C n , then it should be
clear that the induced norm of M may be given by something other than
 (M ).
Another example that is crucial in studying systems is the case of con-
volution operators. Suppose f is in the space L11 [0; 1) of scalar functions,
then it denes an operator F : L1 [0; 1) ! L1 [0; 1) by
Z t
(Fu)(t) := f (t ;  )u( )d; t  0;
0
for u in L1 [0; 1). To see that F is a bounded mapping on L1[0; 1) notice
that if u 2 L1 [0; 1) and we set y = Fu then the following inequalities are
satised for any t  0.
Z t
jy(t)j = j f (t ;  )u( )d j
0
Z t
 jf (t ;  )j ju( )jd
Z0 1
 jf ( )jd kuk1 = kf k1 kuk1:
0
Therefore we see that kyk1  kf k1 kuk1 since t was arbitrary, and more
generally that
kF kL1!L1  kf k1:
It follows that F is a bounded operator. The next proposition says that
the induced norm is exactly given by the 1-norm of f , as one might guess
from our argument above.
Proposition 3.5. With F dened on L1[0; 1) as above we have
kF kL1!L1 = kf k1:
Proof . We have already shown above that kF kL1!L1  kf k1 and so
it remains to demonstrate kF kL1!L1  kf k1. We accomplish this by
showing that, for every  > 0, there exists u, with kuk1 = 1 such that
kFuk1 +   kf k1:
Choose  > 0 and t such that
Z t
jf (t ;  )jd +  > kf k1:
0
 3.2. Operators 105
Now set u( ) = sgn( f (t ;  ) ), that is the sign of f (t ;  ), and notice that
kuk1  1. Then we get
Z t
(Fu)(t) = f (t ;  )sgn(f (t ;  )) d
0
Z t
= jf (t ;  )j d:
0
The function (Fu)(t) is continuous and therefore we get kFuk1 +  
kf k1. 
Thus this example shows us how to calculate the L1 induced norm of
a convolution operator exactly in terms of the 1-norm of its convolution
kernel.

We now introduce the adjoint of an operator in Hilbert space.
Denition 3.6. Suppose V and Z are Hilbert spaces, and F 2 L(V , Z ).
The operator F  2 L(Z ; V ) is the adjoint of F if
hz; FviZ = hF  z; viV
for all v 2 V and z 2 Z .
We now look at some examples.
Examples:
The simplest example of this is where V = C n and Z = C m , which are
equipped with the usual inner product, and F 2 C mn . Then the adjoint
of the operator F is exactly equal to its conjugate transpose of the matrix
F , which motivates the notation F  .
Another example is given by convolution operators: suppose that f is a
scalar function in L1 [0; 1) and the operator Q on L2 [0; 1) is again dened
by
Z t
(Qu)(t) := f (t ;  )u( )d; (3.3)
0
for u in L2 [0; 1). Then we leave it to the reader to show that the adjoint
operator Q is given by
Z 1
(Q z )(t) = f  ( ; t)z ( ) d; for z in L2 [0; 1),
t
where f  denotes the complex conjugate of the function f . 
Some basic properties of the adjoint are given in the following statement;
while the existence proof is beyond the scope of this course, the remaining
facts are covered in the exercises at the end of the chapter.
106 3. Linear Analysis
Proposition 3.7. For any F 2 L(V ; Z ), the adjoint exists, is unique and
satises
kF k = kF  k = kF  F k 21 : (3.4)
An operator F is called self-adjoint if F  = F , which generalizes the
notion of a Hermitian matrix. It is easy to show that when F is self-adjoint,
the quadratic form
(v) := hFv; vi
takes only real values. Therefore we have  : (V ; V ) ;! R. Such quadratic
forms play an important role later in the course.
As in the matrix case, it is natural to inquire about the sign of such
forms. We will say that a self-adjoint operator is
 Positive semidenite (denoted F  0) if hFv; vi  0 for all v 2 V .
 Positive denite (denoted F > 0) if, there exists  > 0, such that
hFv; vi  kvk2, for all v 2 V .
We remark that an operator F satisfying hFv; vi  0, for all nonzero
v 2 V , is not guaranteed to be positive as it is in the matrix case; the
exercises provide examples of this. An important property is that given an
operator in one of these classes, there always exists a square root operator
F 21 of the same class, such that (F 21 )2 = F .
We now introduce another important denition involving the adjoint. An
operator U 2 L(V ; Z ) is called isometric if it satises
U U = I;
The reason for this terminology is that these operators satisfy
hUv1 ; Uv2 i = hU  Uv1 ; v2 i = hv1 ; v2 i;
for any v1 ; v2 2 V , i.e. the operator preserves inner products. In particular,
isometries preserve norms and therefore distances: they are \rigid" trans-
formations. A consequence of this is that they satisfy kU k = 1, but the
isometric property is clearly more restrictive.
An isometric operator is called unitary if U  = U ;1 , in other words
U  U = I and UU  = I:
Unitary operators are bijective mappings that preserve the all the structure
of a Hilbert space; if U 2 L(V ; Z ) exists, the spaces V ; Z are isomorphic,
they can be identied from an abstract point of view.
Example:
A matrix U 2 C mn whose columns u1 ; : : : ; un are orthonormal vectors in
C m is an isometry; if in addition m = n, U is unitary.


 3.2. Operators 107
Before leaving this section we again emphasize that we use the same
notation \ * " to denote complex conjugate of a scalar, complex conjugate
transpose of a matrix, and adjoint of an operator; a moment's thought will
convince you that the former two are just special cases of the latter and
thus this helps keep our notation economical.

3.2.1 Banach algebras

We have already seen that the set of operators L(V ) on a Banach space V
is itself a Banach space. However, the composition of operators endows it
with an additional algebraic structure, which is of crucial importance. In
this section we will isolate this structure and prove an important associated
result.
Denition 3.8. A Banach algebra B is a Banach space with a multiplica-
tion operation dened for elements of B, mapping B  B ! B, satisfying
the following properties:
(a) Algebraic properties:
(i) There exists an element I 2 B, such that F
I = I
F = F , for
all F 2 B;
(ii) F (GH ) = (FG)H , for all F; G; H in B;
(iii) F (G + H ) = FG + FH , for all F; G; H in B;
(iv) For all F; G in B, and each scalar , we have F (G) =
(F )G = FG.
(b) Property involving the norm: for all elements F; G in B, we have
kFGk  kF k
kGk:
This denition says that a Banach algebra has a multiplication opera-
tion dened between its elements, one element is the identity element, and
satises the standard properties of being associative, distributive, and com-
mutative with scalar multiplication. The key property of a Banach algebra
is that its norm satises the submultiplicative inequality, listed in (b).
While the above denition is abstract, we should keep in mind that the
main motivating example for the study of Banach algebras is the space
B = L(V ) of operators on a Banach space V , equipped with the induced
norm. For this example the algebraic properties are immediate, and we
have already seen in (3.2) that the submultiplicative property holds.
Examples:
Of course C itself is a Banach algebra, but perhaps the simplest nontrivial
example is the space C nn of n  n complex matrices equipped with the
maximum singular value norm  (
). Clearly this is a special case of L(V )
with V = C n .
108 3. Linear Analysis
An example of Banach algebra which does not correspond to a space of
operators is C nn endowed with the Frobenius norm j
jF . It is shown in
the exercises that the submultiplicative inequality holds, but this is not an
induced norm. 
We now turn to the concept of an inverse of an element J in a Banach
algebra. We use J ;1 to denote an element of B which satises
J J ;1 = J ;1 J = I:
If such an element exists it is not hard to see that it is necessarily unique.
The following result is very important in robust control theory, where it
is commonly referred to as the small gain theorem.
Theorem 3.9. Suppose Q is a member of the Banach algebra B. If kQk <
1, then (I ; Q);1 exists. Furthermore
1
X
(I ; Q);1 = Qk :
k=0
This theorem says that if the operator Q has \gain" kQk < 1, then (I ;
Q);1 is well-dened, and a power series expression for this inverse is given.
Although this norm condition is sucient for the existence of the inverse
of I ; Q it is not necessary as we see by example.
Examples:
If we take the Banach algebra (Rnn ;  (
) ) from the last example we see
that
 1
with Q = 01 02 the maximum singular value (Q) = 21 :
2
Therefore by the theorem we know that the matrix I ; Q has an inverse;
obviously it is easy forus tosimply compute this inverse and show directly
that (I ; Q);1 = 23 21 12 , which also agrees with the formula in the
theorem.
This theorem gives sucient conditions for the existence of (I ; Q);1
but the following simple example shows it is not necessary: let
   
~ 0 10 ~
Q = 0 0 and then (I ; Q) = 0 1 ; ; 1 1 ; 10

but  (Q~ ) = 10 which does not satisfy the hypothesis of the theorem. 
Having considered some simple examples | soon we will consider innite
dimensional ones | let usPturn to the proof of the theorem. We start by
explaining the meaning of 1k=0 Q . If this sum is a member of B , this sum
k
 3.2. Operators 109
represents the unique element L that satises
n
X
lim kL ;
n!1
Qk k = 0;
k=0
Pn k
that is L = limn!1 k=0 Q . We also have the following technical lemma;
we leave the proof to the reader as it is a good exercise using the basic
properties discussed so far.
Lemma 3.10. Suppose Q and P1k=0 Qk are both in the Banach algebra B.
Then
1
X 1
X
Q Qk = Qk :
k=0 k=1
We are now ready to prove the theorem. The proof of the theorem relies
on the completeness property of the space B, and the submultiplicative
property.
Proof of Theorem 3.9. Our rst task is to demonstrate that P1k=0 Qk
is an element of B. Since by assumption
Pn B is complete it is sucient to
show that the sequence Tn := k=0 Qk is a Cauchy sequence. By the
submultiplicative inequality we have that
kQk k  kQkk :
For m > n we see that
m
X m
X m
X
kTm ; Tnk = k Qk k  kQk k  kQkk ;
k=n+1 k=n+1 k=n+1
where the left hand inequalityP follows by the triangle inequality.
m;n
It is
m
straightforward to show that k=n+1 kQk = kQk
k n +1 1 ;kQ k
1;kQk , since
the former is a geometric series. We conclude that
kTm ; Tn k  1k;QkkQk ;
n+1

so Tn is a Cauchy sequencePas we required.

Having established that 1 k=0 Q is an element of B we now show that
k
it is the inverse of I ; Q. We look at the product
1
X 1
X 1
X 1
X 1
X
(I ; Q) Qk = Qk ; Q Qk = I + Qk ; Q Qk = I;
k=0 k=0 k=0 k=1 k=0
where the rightPhand side follows by invoking Lemma 3.10. Similarly we
can show that 1 k=0 Q k
(I ; Q) = I , and therefore P1 Qk is the inverse
k=0
of I ; Q. 
110 3. Linear Analysis
Thus we conclude our brief incursion into the theory of Banach algebras.
We now return to the special case of operators, to discuss the spectrum;
here we will nd a rst application for the above small gain result.

3.2.2 Some elements of spectral theory

In this section we introduce the spectrum of an operator and discuss a
few elementary properties. The objective is to cover only some basic facts
which will be used later in the course. For an extensive treatment of this
important topic in analysis, see the references at the end of the chapter.
While the spectrum can be studied for operators on Banach space, or
more generally for Banach algebras, we conne our discussion to operators
on a Hilbert space.
Denition 3.11. Let V be a Hilbert space, and M 2 L(V ). The spectrum
of M is dened by
spec(M ) := f 2 C : (I ; M ) is not invertible in L(V )g;
and the spectral radius of M by
(M ) := supfjj :  2 spec(M )g:
A non-obvious fact which is implicit in the denition of (M ) is that the
spectrum is non-empty. We will accept this fact without proof.
Examples:
As usual we begin by considering the nite dimensional case, with M 2
C nn . It is clear that in this case the spectrum consists of the set of eigen-
values of M , and the spectral radius coincides with the largest absolute
value of any eigenvalue of M .
The next example shows that there may be more to the spectrum than
eigenvalues. Let V = L2 [0; 1) and dene the operator M as
M : u(t) 7! e;t u(t):
It is easy to show that the operator is bounded since e;t is a bounded
function on [0; 1). Now we claim that spec(M ) is the real interval [0; 1].
To see this notice that
(I ; M ) : u(t) 7! ( ; e;t )u(t):
If  62 [0; 1], then we can dene an inverse function
'(t) = 1
 ; e;t for t  0;
that is bounded on [0; 1), dening a bounded inverse operator on L2 [0; 1),
(I ; M );1 v(t) 7! 1 v(t):
 ; e;t
 3.2. Operators 111
So such  is not in the spectrum, and spec(M )  [0; 1]. If 2[0; 1], the
inverse function '(t) will go to innity at a t = ; log() (or t = 1 when
 = 0), and it can be deduced that there is no bounded inverse to the
operator (I ; M ). Therefore spec(M ) = [0; 1].
However notice that
( ; e;t )u(t) = 0 implies u(t) = 0; for almost every t,
so the operator M has no eigenvalues. 
The next proposition relates the spectral radius to the norm, and in
particular implies the spectral radius is always nite, or equivalently that
the spectrum is a bounded set. The proof is based on the small gain result
of Theorem 3.9.
Proposition 3.12. Given an operator M 2 L(V ), the inequality (M ) 
kM k holds.
Proof . Suppose jj > kM k. Then setting Q = M= we have kQk < 1 and
we invoke Theorem 3.9 to see that I ; Q is invertible, with inverse
 ;1 1 Mk :
I ; M
X
= k
k=0 
Now it follows immediately that
1
X M k = ;1 I ; M ;1 = (I ; M );1 ;
 
k+1 (3.5)
k=0  
so  62 spec(M ). This means that spec(M ) is included in the closed ball of
radius kM k, which completes the proof. 
The previous result generalizes the well-known relationship between spec-
tral radius and maximum singular value of a matrix, and in general there
is a gap between these two quantities. The spectral radius can, however,
be characterized exactly in terms of the norms of the powers of M .
Proposition 3.13. For M 2 L(V ), the spectral radius satises (M ) =
limk!1 kM k k k1 .
This statement implies, in particular, that the limit on the right hand side
exists for every M . We will not prove this result here, but remark that
this can be done by studying the radius of convergence of the power series
in (3.5), that has similar properties as scalar power series. For the case of
matrices, the reader is invited to give a more direct proof based on the
Jordan form.
A special case in which the spectral radius coincides with the norm, is
when the operator is self-adjoint.
Proposition 3.14. Let M 2 L(V ), and suppose that M = M . Then
(M ) = kM k.
112 3. Linear Analysis
The proof is a simple exercise based on Propositions 3.7 and 3.13, and is left
for the reader. The following is an immediate corollary, that generalizes the
familiar matrix property that the square of  (A) is equal to the maximum
eigenvalue of A A.
Corollary 3.15. Let M 2 L(U ; V ), where U ; V are Hilbert spaces. Then
kM k2 = (M  M ).
We remark, without proof, that the spectrum of a self-adjoint opera-
tor is contained in the real line, and a positive semi-denite operator has
spectrum in the non-negative half-line.
To conclude the section we will show a result regarding the invariance of
the spectral radius to changes in the order of multiplication of operators.
The key observation is the following.
Lemma 3.16. Consider the operators M 2 L(U ; V ) and Q 2 L(V ; U ),
so that MQ 2 L(V ) and QM 2 L(U ) are well dened. If IU , IV are the
identity operators on each space, then
IV ; MQ is invertible if and only if IU ; QM is invertible.
Proof . Assuming IU ; MQ is invertible, we can explicitly construct the
inverse
(IU ; QM );1 = IU + Q(IV ; MQ);1 M:
In other words, the operator on the right hand side is well-dened, and rou-
tine operations show it is the inverse of IU ; QM . The converse implication
is analogous. 
Having shown this, we can state the following result regarding the
spectrum.
Proposition 3.17. Consider the operators M 2 L(U ; V ) and Q 2
L(V ; U ), so that MQ 2 L(V ) and QM 2 L(U ) are well dened. Then the
sets spec(MQ) and spec(QM ) coincide except possibly for the value 0 2 C .
In particular,
(MQ) = (QM ):
Proof . Consider 0 6=  2 C . Then
 62 spec(MQ) if and only if I ; MQ invertible,
if and only if I ; M
 Q invertible,
if and only if I ; Q M invertible,
if and only if I ; QM invertible,
if and only if  62 spec(QM );
 3.3. Frequency domain spaces: signals 113
where the only non-trivial step follows from Lemma 3.16. Therefore the
spectra can dier at most in the element  = 0; this potential dierence
cannot aect the spectral radius. 
Example:
Consider two matrices M 2 C mn and Q 2 C nm , where e.g. m < n. The
previous result implies that the nonzero eigenvalues of MQ 2 C mm and
QM 2 C nn coincide. Clearly QM will have additional eigenvalues at 0,
since its rank is smaller than its dimension n.

So far in this chapter we have introduced a number of general mathe-
matical concepts and tools. We will now turn to some more customized
material.

3.3 Frequency domain spaces: signals

In the previous sections we have introduced, by way of examples, time
domain spaces of functions which will play the role of signals in our control
problems. Frequently in the sequel it will be advantageous to reformulate
problems in the frequency domain where they can be simpler to solve or
conceptualize. In particular we will see that L2 spaces are particularly
suited for a frequency domain treatment. In this section we will introduce
frequency domain spaces, and describe the various relationships they have
with the time domain spaces.

3.3.1 The space L^ 2 and the Fourier transform

We now dene the complex inner product space L^ 2(j R) which consists of
functions mapping j R to C n with the inner product
Z 1
hu^; v^i2 := 21 u^ (j!)^v(j!)d! :
;1
Thus a function u^ : j R ! C is in L^ 2 (j R) if hu^; u^i2 = ku^k22 < 1. Here
n
we use the same notation for the norm and inner product of L^ 2 as we did
for L2 ; which is intended to always be clear from the context as frequency
domain objects will be denoted by \ b ".
The Fourier transform of a function u : R ! C n , is dened to be
Z 1
u^(j!) = u(t)e;j!t dt:
;1
As usual, writing such an integral raises the issue of its convergence. It is
easy to see that the integral converges absolutely for u(t) 2 L1 (;1; 1),
114 3. Linear Analysis
but we are interested in dening Fourier transforms on L2 (;1; 1), where
this strong convergence may not hold. One approach to avoid this diculty
is to consider the \principal value" limit
Z T
u^(j!) = Tlim
!1
u(t)e;j!t dt:
;T
We will adopt the following convention: u^(j!) is dened as the above limit
at values of ! where it exists, and u^(j!) = 0 otherwise. In this way the
transform is dened for any function which is integrable on nite intervals.
For convenience we denote this operation by the map , and write simply
u^ = u
to indicate the above relationship. It can be shown (see the references) that
when u(t) 2 L2 (;1; 1) this limit exists for almost all !.
Given an function u^ : j R ! C n we dene the inverse Fourier transform
of u^ by
Z 1
u(t) = 2 1 u^(j!)ej!t d!;
;1
with an analogous convention regarding the convergence of the integral,
and use u = ;1 u^ to indicate this transformation.
As the notation would indicate, for certain classes of functions these
maps are inverses of each other. For our purposes, it suces to note that
for u 2 L2 (;1; 1) and v^ 2 L^ 2 (j R), we have
u(t) = ;1 (u)(t); for almost every t,
and
v^(j!) = (;1 v^)(j!); for almost every !.
Furthermore, we have the following key result known as the Plancherel
theorem.
Theorem 3.18. With the Fourier transform and its inverse dened as
above:
(i) The map  : L2(;1; 1) ! L^ 2(j R), and given any u; v 2
L2 (;1; 1) the equality
hu; vi2 = hu; vi2 holds:
(ii) The map ;1 : L^ (j R) ! L2 (;1; 1), and if u^; v^ 2 L^ 2 (j R), then
hu^; v^i2 = h;1 u^; ;1 v^i2 is satised.
This theorem says that the Fourier transform is an invertible map between
the spaces L2 (;1; 1) and L^ 2 (j R), and more importantly, this map pre-
serves the inner product; thus the Fourier transform is a unitary operator
 3.3. Frequency domain spaces: signals 115
which makes the two spaces isomorphic. In particular the theorem says for
any v^ 2 L^ 2 (j R) and u 2 L2(;1; 1) we have
kuk2 = kuk2 and k;1v^k2 = kv^k2 ;
namely norms are preserved under this mapping.

3.3.2 The spaces H2 and H2? and the Laplace transform

In this section we introduce two subspaces of L^ 2 (j R) that will be important
in our studies of causal systems.
Recall the denition of L2 [0; 1), the Hilbert subspace of functions in
L2 (;1; 1) which are zero in (;1; 0). Given u 2 L2 [0; 1), we dene its
Laplace transform by the integral
Z T Z 1
u^(s) := Tlim
!1
e;st u(t)dt = e;st u(t)dt; (3.6)
0 0
when this limit exists, and set u^(s) = 0 at the divergent values of s. We will
soon see, however, that the integral converges absolutely when Re(s) > 0.
We use the notation
u^ = u
to indicate this transformation. If we evaluate the transform at s = j!, for
some real number !, we see that since u(t) 2 L2 [0; 1),
Z 1 Z 1
u^(j!) = e;j!t u(t)dt = e;j!t u(t)dt
0 ;1
is the Fourier transform of u(t); therefore there is no source of ambiguity
in the notation \^u"; depending on its argument it will denote the Laplace
transform or the Fourier restriction. Notice that by the Plancherel theorem,
u^(j!) 2 L^ 2 (j R), and ku^k2 = kuk2.
Next notice that for s 2 C , with Re(s) > 0, we have that the scalar
function dened by
(
e ;s t for t  0,
gs (t) =
0 otherwise;
is in L2 [0; 1). Thus we see that if u(t) is scalar, the Laplace transform u^(s)
is given by the following inner product
u^(s) = hgs ; ui2 ;
and therefore the integral (3.6) converges for all s satisfying Re(s) > 0.
This is also true if u(t) is vector-valued, applying the previous argument
for each component.
To state some additional properties of u^(s) we introduce a new function
space H2 in the following denition.
116 3. Linear Analysis
Denition 3.19. A function u^ : C+ ! C n is in H2 if
(a) u^(s) is analytic in the open right half plane C + ;
(b) For almost every real number !,
lim+ u^( + j!) = u^(j!) ;
!0
Z 1
(c) sup0 ju^( + j!)j22 d! < 1.
;1
We remark that the supremum in (c) is always achieved at  = 0 when u^
is in H2 . Also that any function u^ which is only dened on the open half
plane C + can be extended to C+ so that it is in H2 if (a) holds, and (c) is
satised when the supremum taken over C + .
The key result we now state is that the Laplace transform of a function
L2 [0; 1) satises these properties, and furthermore every function in H2
can be obtained in this form.
Theorem 3.20.
(a) If u 2 L2 [0; 1) then u 2 H2 .
(b) If u^ 2 H2 , then there exists u 2 L2[0; 1) satisfying u = u^.
We will not prove this theorem, since it would take us too far aeld, but
only make a few remarks. Both parts are non-trivial. Part (a) says that all
the requirements in Denition 3.19 are met by a Laplace transform u^(s) of
a function in L2 [0; 1).
Part (b) is called the Paley-Wiener theorem, and in essence says that
the mapping  has an inverse dened on H2 . This mapping is the inverse
Laplace transform and can be shown to be given explicitly by
Z 1
u(t) = (;1 u^)(t) = 21 et
ej!t u^( + j!)dw;
;1
where  is any positive real number. In particular it can be shown that
for u^ 2 H2 , the above integral is independent of  and is equal to zero for
t < 0.
A useful consequence of Theorem 3.20 is the following:
Corollary 3.21. Let u^ and v^ be functions in H2 , and u^(j!) = v^(j!) for
every ! 2 R. Then u^(s) = v^(s) for every s 2 C+ .
Proof . From Theorem 3.20 we nd u(t) and v(t) in L2[0; 1) such that
u^ = u, v^ = v. Then u^ ; v^ = (u ; v).
Since u^(j!) ; v^(j!) = 0, we conclude that the Fourier transform (u ;
v) = 0. Since  is an isomorphism we must have u(t) ; v(t) = 0 for almost
all t. Now the Laplace transform gives u^(s) = v^(s), for every s 2 C+ . 
From the previous corollary we nd that there is a one to one corre-
spondence between functions in H2 and their restrictions to the imaginary
 3.3. Frequency domain spaces: signals 117
axis. Thus we are justied in identifying these two functions and writing
H2  L^ 2(j R), which involves some abuse of notation but no ambiguity. We
say a function u^(j!) 2 L^ 2(j R) is in H2 if it has an analytic continuation
u^(s) to the right half plane which satises the conditions of Denition 3.19.
Given this identication, we can endow H2 with the inner product
inherited from L^ 2 (j R), dening for u^ and v^ in H2
Z 1
1
hu^; v^i2 = 2 u^(j!)^v (j!) d! :
;1
In this way H2 is a Hilbert subspace of L^ 2 (j R).
Example:
To gain insight into Denition 3.19, consider the function
s
u^(s) = s e+ 1 :
It is easily shown that u^(j!) 2 L^ 2 (j R), and also u^(s) satises (a) and (b)
in the denition. However (c) is not satised, so u^(s) 62 H2 .
In the time domain, it is not dicult to show that
 ;(t+1)
;1 u^ = u(t) = e0; ; for t  ;1;
otherwise,
that is not a function in L2 [0; 1). 
Having identied H2 as a subspace of L^ 2 (j R), we immediately won-
der about the dierence between these two spaces, how do we need to
complement H2 to obtain all of L^ 2 (j R)?
To work towards an answer it is useful to return to the time domain;
here it is very easy to characterize the dierence between L2 (;1; 1) and
L2 [0; 1). In particular, the space L2 (;1; 0] is the orthogonal complement
of L2[0; 1): functions in both spaces are orthogonal to each other, and any
function u(t) 2 L2 (;1; 1) can be trivially written as a sum
u(t) = u+ (t) + u;(t); u+ (t) 2 L2[0; 1); u; (t) 2 L2 (;1; 0]:
We denote this relationship by
L2 (;1; 0] = L2 [0; 1)?:
Some properties of orthogonal complements in general Hilbert spaces are
included in the exercises at the end of the chapter.
Returning to the frequency domain, we are led to ask the question: what
is the Fourier image of L2 (;1; 0]? Not surprisingly, the answer involves
the use of Laplace transforms. For u(t) 2 L2 (;1; 0], dene the left sided
Laplace transform of u by
Z 0
(; u)(s) := e;st u(t)dt;
;1
118 3. Linear Analysis
with the usual convention regarding convergence. Now it is easy to see that
this transform converges on the left hand plane s 2 C ; . Furthermore, for
s = j! it coincides with the Fourier transform.
This symmetric situation leads to introduce the space H2? of functions
mapping C; to C n which satisfy: y^ is in H2? if
y^(;s) is a member of H2 :
In particular, functions in H2? are analytic on C ; and have limit almost
everywhere when approaching the imaginary axis horizontally from the left.
As expected we have the following result.
Proposition 3.22.
(a) If u 2 L2 (;1; 0] then ; u 2 H2?;
(b) If u^ 2 H2? , then there exists u 2 L2 (;1; 0] satisfying ; u = u^.
Proof . We reduce the statement to Theorem 3.20 by a symmetry in the
time axis. We have
u(t) 2 L2 (;1; 0] if and only if u(;t) 2 L2 [0; 1);
if and only if [u(;t)] 2 H2 ;
and it is easy to show that
Z 1 Z 0
[u(;t)](s) = u(;t)e;stdt = u(t)est dt = ;[u(t)](;s):
0 ;1
So [u(;t)] 2 H2 if and only if ;u 2 H2? . 
Analogously to Corollary 3.21 it follows that functions in H2? are char-
acterized by their values on the imaginary axis, so once again we can write
H2?  L^ 2 (j R), identifying H2? with a Hilbert subspace of L^ 2(j R), con-
sisting of functions u^(j!) which have analytic continuation to the left half
plane with the corresponding boundedness property.
To complete our picture we inquire as to how do the subspaces H2 and
H2? relate to each other? As the notation suggests, we have the following
answer.
Proposition 3.23. The space H2? is the orthogonal complement of H2 in
L^ 2 (j R). Namely, if u^ 2 H2 and v^ 2 H2? then hu^; v^i2 = 0, and also every
function u^(j!) 2 L^ 2(j R) can be written uniquely as a sum
u^(j!) = u^+ (j!) + u^;(j!); u^+ 2 H2 ; u^; 2 H2?:
The proof is a direct application of the Plancherel theorem, since we have
established that as subspaces of L^ 2(j R), H2 and H2? are the Fourier images
of L2 [0; 1) and L2 (;1; 0]. We leave details to the reader.
 3.3. Frequency domain spaces: signals 119
3.3.3 Summarizing the big picture
It is useful now to summarize what we have learned about frequency do-
main signal spaces. The main conclusions are represented in the following
diagram which includes all the spaces and relevant transforms.

L2 (;1; 0]  L2 (;1; 1)  L2[0; 1)

? x ? x ? x
? ? ? ? ? ?
; ??y ?
?
?
;;1 ??y ?
?
?
;1 ??y ?
?
?
;1
H2?  L^ 2 (j R)  H2
The time domain space L2(;1; 1) is isomorphic to the frequency domain
space L^ 2 (j R) by means of the Fourier transform, which is invertible and
preserves inner products. In the time domain we can identify the natural
subspaces L2 (;1; 0] and L2 [0; 1), which are mutually orthogonal. In the
frequency domain, these are mapped by the Fourier transform to the sub-
spaces H2? and H2 of L^ 2 (j R); functions in these spaces are dened over the
imaginary axis, but also admit analytic continuations respectively to each
half-plane, with bounded norms over vertical lines. The Laplace transforms
; and  extend the Fourier transform to each analytic domain.
Examples:
To gain more insight into the frequency domain spaces that have been
introduced, it is useful to discuss which rational functions belong to each
of them. Recall from Chapter 2 that a scalar rational function is a ratio of
polynomials
u^(s) = pq((ss)) ;
and the concept extends to vectors or matrices with components of this
form. We will denote by RL^ 2, RH2 , and RH2? the sets of rational functions
which belong, respectively, to L^ 2 (j R), H2 , and H2? . For example u^(s) is in
RL^ 2 if u^(j!) 2 L^ 2 (j R).
It is not dicult to show that a vector rational function is:
(i) in RL^ 2 if it is strictly proper and has no poles on the imaginary axis;
(ii) in RH2 if it is strictly proper and has no poles on the closed right
half plane C+ ;
(iii) in RH2? if it is strictly proper and has no poles on the closed left half
plane C; .
The subspaces RL^ 2, RH2 , and RH2? are themselves inner product spaces,
but are not complete. In particular it can be shown that they are dense
in their respective Hilbert spaces L^ 2 (j R), H2 and H2? ; this means that for
120 3. Linear Analysis
every element in H2 is the limit of a sequence of functions in RH2 , and
analogously for the other spaces.
It follows directly from the partial fraction expansion that every function
in RL^ 2 can be written as a sum of functions RH2 , and RH2?; what is not so
immediate is that these subspaces are perpendicular, i.e. given a rational
function u^(s) with no poles in C+ , and another one v^(s) with no poles in
C ; ,
1 Z 1 u^ (j!)^v (j!) d! = 0:
2 ;1
This latter fact follows from the preceding theory. 
Our next goal is to move from alternate representations for signals to
equivalent system representations.

3.4 Frequency domain spaces: operators

Having studied the structure of the L2 and L^ 2 Hilbert spaces, we now
turn to the study of operators. Since the above spaces are isomorphic it is
clear that so must L(L2 ) and L(L^ 2 ). Namely with each M 2 L(L2 ) we can
associate M~ = M ;1 2 L(L^ 2 ), where  is the Fourier transform. Thus
we can choose to study operators in whichever domain is most convenient.
In particular, we will see in this section that the frequency domain provides
a very elegant, exact characterization of an important class of operators on
L2 : those which are linear and time-invariant.
Our study begins by introducing the class of multiplication operators on
L^ 2 (j R). First dene the space L^ 1 (j R) of matrix-valued functions j R !
C mn such that

kG^ k1 = ess sup  (G^ (j!)) < 1:

! 2R
The subscript 1 in this norm corresponds to the essential supremum over
frequency, analogously to the time domain L1 space introduced before.
Notice however that we are using matrix functions and taking a maximum
singular value norm at each frequency. More importantly, we should bear in
mind that, even in the scalar case, this space is not related to L1 (;1; 1)
via the Fourier transform.
The next result shows that L^ 1(j R) is a representation for a set of linear
bounded operators on L^ 2(j R).
Proposition 3.24. Every function G^ 2 L^ 1(j R) denes a bounded linear
operator
MG^ : L^ 2 (j R) ! L^ 2(j R) ;
 3.4. Frequency domain spaces: operators 121
via the relationship (MG^ u^)(j!) = G^ (j!)^u(j!). Also
kMG^ kL^ 2!L^ 2 = kG^ k1 :
The operator MG^ is called, for obvious reasons, a multiplication operator;
we see that the induced norm of this operator is exactly kG^ k1 , which is
left as an exercise. We also have that G^ denes an operator on L2 (;1; 1),
denoted by G, and dened by
G = ;1 MG^  :
This operator has some very special properties. We call G^ (j!) the fre-
quency response of G, when G satises the last equation. The diagram
below summarizes the relationships between operators and spaces dened
so far.

L2 (;1?; 1)
?
;!
G
L2 (;1
? ; 1)
?
??y ?
?
y

L^ 2(j R) ;!
M^G
L^ 2(j R)

3.4.1 Time invariance and multiplication operators

In this section we investigate some of the temporal properties of multipli-
cation operators. Our discussion is brought into focus by dening the shift
operator S , for   0, on time domain vector valued functions. If u is a
function R ! C n the action of the shift operator is given by
(S u)(t) = u(t ;  ) :
The inverse S;1 is also dened on these functions, and satises
(S;1 u)(t) = u(t +  ):
Suppose u 2 L2 (;1; 1), and let y = S u. Then taking Fourier transforms
we have
y^(j!) = e;j!
u^(j!) ;
so S is represented in the frequency domain by the multiplication by e;j!
in L^ 1(j R).
This brings us to the denition of time invariance: an operator Q mapping
L2 (;1; 1) ! L2(;1; 1) is time invariant if
S Q = QS ; for all   0 :
From this denition it is immediate to see that every multiplication opera-
tor MG^ denes a time invariant G = ;1 MG^ . In fact, S G is represented
122 3. Linear Analysis
by
e;jw G^ (j!) = G^ (j!)e;jw ;
where the right hand expression is a representation for GS . Thus G is
necessarily time invariant.
A more surprising fact is the converse, namely all time invariant operators
on L2 (;1; 1) can be represented in this way.
Theorem 3.25. An operator G : L2(;1; 1) ! L2(;1; 1) is time in-
variant if and only if it can be represented in L^ 2 (j R) by a multiplication by
a function
G^ 2 L^ 1 (j R) :
Thus this theorem says that the set of time invariant operators is in ex-
act correspondence with the set of functions L^ 1 (j R). In mathematical
terminology these spaces are isomorphic.

3.4.2 Causality with time invariance

We now formalize the notion of causality, which informally means that
the present is only determined by the past and present. It turns out that
the space of functions L^ 1 (j R) gives us representations for multiplication
operators that are not always causal. One of our goals is to classify those
operators that are both time invariant and causal.
To begin dene the truncation operator P , for  2 R, on vector valued
functions by

(P u)(t) = u0;(t); tt  ;
> ;
where u : R ! C n . The function (P u)(t) has support on the interval
(;1;  ], which means it is zero outside this interval.
We dene an operator G : L2 (;1; 1) ! L2 (;1; 1) to be causal if
P GP = P G; (3.7)
for all  2 R. If we apply this identity to some element u 2 L2 (;1; 1) we
get
P GP u = P Gu : (3.8)
Notice that (P u)(t) is zero for t >  . Thus the above equation says that
(Gu)(t), for t   , only depends on u(t) for t   . That is the past only
depends on the past.
The above denition of causality depends on the real parameter  , and
we now attempt to simplify this condition in the special case of G being a
 3.4. Frequency domain spaces: operators 123
time-invariant operator. Suppose that G is time invariant and satises the
above causality condition for the special case where  = 0, namely
P0 G(I ; P0 ) = 0 holds. (3.9)
We now show that this in fact guarantees that the causality condition holds
for all  , and therefore that G is causal. To see this notice that the following
relationship holds, for any   0, immediately from the denitions of S
and P :
P S = S P0 : (3.10)
That is, if we truncate a function over positive time and shift it by  , this
is the same as shifting it by  and then truncating it after time  .
Turning back to (3.9) we x a positive  and see it holds if and only if
S P0 G(I ; P0 ) = 0
is satised, since S is non singular. Now applying (3.10) we obtain the
equivalent conditions
0 = S P0 (I ; P ) = P0 S G(I ; P )
= P G(I ; P )S ;
where the latter two equalities follow from (3.10) and the fact that G is
time invariant. Finally, since the shift operator is invertible we see this
latter equality holds exactly when P G(I ; P ) = 0. Therefore condition
(3.8) holds for every positive  whenever equation (3.9) does. For  < 0
we can use the same argument starting with P S;;1 = S;;1 P0 instead of
(3.10), to show that (3.9) implies (3.8). Clearly (3.9) is a special case of
(3.8), and so we have the following result.
Proposition 3.26. A time invariant operator G on L2(;1; 1) is causal
if and only if
P0 G(I ; P0 ) = 0 is satised: (3.11)
We can also write the following equivalent characterization, that follows
by noticing that P0 and I ; P0 are the natural projections from L2 (;1; 1)
to L2 (;1; 0] and L2 [0; 1). We leave details to the reader.
Corollary 3.27. A time invariant operator G : L2(;1; 1) ! L2(;1; 1)
is causal if and only if it maps L2 [0; 1) to L2 [0; 1).
In plain language this corollary says that a time invariant operator on
L2 (;1; 1) is causal exactly when it maps every function that is zero for
negative time to a function which also is zero on the negative time axis.
124 3. Linear Analysis
3.4.3 Causality and H1
Having characterized causality in the time domain, we now return to fre-
quency domain spaces. The key question is: which multiplication operators
MG^ on L^ 2 (j R) correspond to causal time invariant operators in the time
domain?
A partial answer is given by translating Corollary 3.27 to the frequency
domain. From our discussion of signal spaces, it follows that
G = ;1 MG^  is causal if and only if MG^ maps H2 to H2 :
Recall that H2 is made of functions u^(j!) which have analytic continua-
tion to the right half plane, with a certain boundedness condition. Which
property of G^ (j!) is required for the product G^ (j!)^u(j!) to have a similar
continuation? Clearly if G^ itself can be analytically extended to G^ (s), ana-
lytic and bounded on C + , then G^ (s)^u(s) would be in H2 whenever u^(s) is.
Now, is this a necessary requirement for G^ ? The following example suggests
that this is indeed the case.
Example:
Suppose G is a causal, time invariant operator between scalar L2 spaces.
Choose
(
;t
u(t) = e for t  0,
0 otherwise:
1 and we have
Then u^(s) = s+1
z^(j!) = G^ (j!) j! 1+ 1 ;
for a scalar function G^ . Thus we see that G^ (j!) = (j! + 1)^z(j!). Since
z^ 2 H2 , we can extend the domain of G^ , setting
G^ (s) := (s + 1)^z(s)
so that G^ (s) is dened, and analytic, for all s 2 C+ . It is also possible to
show that
kG^ k1 = sup+ jG^ (s)j :
s2C
So we see from this generic scalar example that causality seems to imply
that G^ (s) is naturally dened on C+ rather than just j R. 
This discussion leads us to dene the following important class of
functions
Denition 3.28. A function G^ : C+ ! C nm is in H1 if
(a) G^ (s) is analytic in C + ;
 3.4. Frequency domain spaces: operators 125
(b) for almost every real number !,
lim+ G^ ( + j!) = G^ (j!);
 !0
(c) sup  (G^ (s)) < 1.
s2C+
As with H2 , it can be shown that H1 functions are determined by their
values on the imaginary axis. Thus we can regard H1 as a subspace of
L^ 1 (j R), with the same norm
kG^ k1 = ess sup (G^ (j!)) :
!2R
We remark that the above quantity also coincides with the supremum in
Denition 3.28(c). This maximum principle is discussed in the exercises.
We also point out that if a function G^ is only dened on C + , but satises
(a) and the supremum condition in (c) restricted to C + , then one can show
that its domain of denition can be extended to include the imaginary axis
so that it is in H1 .
We are now ready to characterize causal, time invariant operators in the
frequency domain.
Theorem 3.29.
(a) Every G^ 2 H1 denes a causal, time invariant operator G on
L2 (;1; 1), where z = Gu is dened by z^(j!) = G^ (j!)^u(j!);
(b) For each causal, time invariant operator G on L2 (;1; 1) there exists
a function G^ 2 H1 such that z = Gu satises z^(j!) = G^ (j!)^u(j!)
for all u in L2 (;1; 1).
The following diagram complements the one given before, focusing exclu-
sively on causal, time invariant operators, or equivalently for multiplication
functions G^ 2 H1 .

L2[0?; 1)
?
;!
G
L2 [0? ; 1)
?
??y ?
?
y

H2 ;!
M^G
H2
Examples:
As a nal illustrative example, we turn once more to the case of rational
functions. We dene the sets RL^ 1 and RH1 to consist of rational, matrix
functions that belong respectively to L^ 1 and H1 . It is not hard to see
that a matrix rational function is:
126 3. Linear Analysis
(i) in RL^ 1 if it is proper and has no poles on the imaginary axis.
(ii) in RH1 if it is proper and has no poles on the closed right half plane
C+ :

It follows from Chapter 2 that every function in RL^ 1 can be expressed

in the form
G^ (j!) = C (j!I ; A);1 B + D;
where A has no eigenvalues on the imaginary axis, and every function in
RH1 can be expressed in the form
G^ (s) = C (sI ; A);1 B + D;
where A is Hurwitz. In the latter case, the time domain operator G =
;1 MG^ , restricted to L2 [0; 1), has the convolution form
Z t
u(t) 7! CeA(t; )Bu( )d + Du(t);
0
that was studied in Chapter 2 when we discussed minimal realizations.
A remark is in order regarding the RL^ 1 case. When A has a right half
plane eigenvalue, rather than the previous convolution which would give a
causal, but unbounded map, the multiplication MG^ corresponds to a non-
causal, yet bounded operator. More insight into this distinction is provided
in the exercises. 
Finally we have the following theorem to end this section and indeed the
chapter.
Theorem 3.30.

(a) Every G^ 2 H1 denes a causal, time invariant operator G on

L2 [0; 1), where z = Gu is dened by z^(j!) = G^ (j!)^u(j!);
(b) If the operator G 2 L(L2 [0; 1) ) is time invariant, then there exists
a function G^ 2 H1 such that z = Gu satises z^(j!) = G^ (j!)^u(j!),
for all u in L2 [0; 1).
This theorem states that all LTI operators on L2 [0; 1) are represented by
functions in H1 . Notice this means that an LTI operator on L2 [0; 1) is
necessarily causal. This is in contrast to the LTI operators on L2 (;1; 1)
which we saw need not be causal.
We have thus concluded our survey of function spaces and basic oper-
ator theory, and are now ready to put these tools to use for the study of
controlled systems.
 3.5. Exercises 127
3.5 Exercises
1. Let h ; i be the inner product on a Hilbert space. Use the fact that
hu + v; u + v i  0 for every scalar  to derive the Cauchy-Schwarz
inequality. Deduce from it the triangle inequality.
2. Prove the submultiplicative inequality (3.2) for operators.
3. Prove Proposition 3.7.
4. Let H be a Hilbert space, and S a linear subspace. We dene the
orthogonal complement of S by
S ? = fv 2 H : hv; si = 0 for all s 2 Sg:
A subspace S is closed if every convergent sequence with elements in
S has limit in S . The Projection Theorem states that if S is closed,
every v 2 H can be written in a unique way as
v = v1 + v2 ; v1 2 S ; v2 2 S ? ;
v1 is called the projection of v in S , and PS 2 L(H) dened by
PS : v 7! v1 is the orthogonal projection operator.
a) If S is closed, show that (S ? )? = S . What happens if S is not
closed?
b) Show that PS = PS2 = PS .
c) For any operator M 2 L(H), show that Im(M )? = ker(M  ).
d) Consider the multiplication operator MG^ on H2 , dened by
G^ (s) = s+1
s : Describe the subspaces in part c) for this case.
Is Im(M ) closed?
5. Consider the space C nn of square matrices. Show that the Frobenius
norm jM jF makes this space a Banach algebra, but that this norm
cannot be induced by any vector norm in C n .
6. Consider the scalar spaces Lp (;1; 1); for which value(s) of p does
the pointwise multiplication fg(t) = f (t)g(t) give the structure of a
Banach algebra?
7. Use the Jordan1 form to show that for any M 2 C nn , (M ) =
limk!1  (M k ) k .
8. Prove Proposition 3.14.
9. Given an operator M on a Hilbert space V , show that
(M ) = inf kDMD;1k;
D2L(V ); invertible
by proving the equivalence between:
(i) (M ) < 1;
128 3. Linear Analysis
(ii) There exists D 2 L(V ), invertible such that kDMD;1k < 1;
(iii) There exists P 2 L(V ), P > 0 such that
M  PM ; P < 0:
Hint: Consider the series 1
P k k
k=0 (M ) M :
10. Prove Proposition 3.24, in particular show that the induced norm is
exactly kG^ k1 . You can assume that G^ (j!) is a continuous function.
Hint: Choose an input that concentrates its energy around the peak
frequency response.
11. Consider the scalar transfer function G^ (s) = C (Is ; A);1 B , with
A Hurwitz, and the corresponding time domain mapping G =
;1 MG^ , ( is the Fourier transform).
a) Recalling the denitions of the time domain spaces L2[0; 1) and
L1 [0; 1), prove that
kGkL2!L1 = kG^ kH2 :
b) Prove that kG^ k2H2 = CXC  , where X satises the Lyapunov
equation
AX + XA + BB  = 0:
12. Consider the multiplication operator MG^ , where G^ (j!) = C (j!I ;
A);1 B + D 2 L^ 1 (j R).
a) Characterize the adjoint MG^ as a multiplication operator, and
give a state-space realization for the corresponding function.
b) Give conditions in terms of G^ (j!) that make:
(i) MG^ self-adjoint; (ii) MG^  0; (ii) MG^ > 0.
c) Assuming MG^ is self-adjoint, determine when hu^; MG^ u^i > 0
holds, for all nonzero u^ 2 L^ (j R). Are these conditions equivalent
to MG^ > 0?
d) What is the spectrum of MG^ ?
13. Maximum modulus principle. It follows from analytic function theory
that given a scalar function f (s) in H1 (analytic and bounded in
Re(s) > 0), its maximum modulus is achieved at the boundary, i.e.
sup jf (s)j = ess sup jf (j!)j = kf k1:
Re(s)>0 !
Using this fact, extend it to matrix-valued functions with the H1
norm.
s
14. Decompose the function u^(s) = se+1 as a sum of functions in H2 and
H2? .
 3.5. Exercises 129
15. a) Given two functions u, v in L2(;1; 1) with Fourier transforms
satisfying
ju^(j!)j  jv^(j!)j for almost all ! 2 R: (3.12)
Show that there exists a linear time invariant operator G, with
kGk  1, satisfying Gu = v. Can G always be chosen to be
causal?
b) Let Gu = v, kGk  1, but G not necessarily time invariant. Is
(3.12) satised?
16. Let G^ (j!) = j!1;1 . Express G = ;1 MG^ , as a convolution in
L2 (;1; 1), giving the corresponding convolution function g(t). Is
G causal?
17. Prove part (b) of Theorem 3.30. To do this show that the domain of
G can be extended to L2 (;1; 1) by setting
;1
!1 S GS (I ; P )u;
Gu := lim
and that G on this extended domain is LTI. Now use Proposition 3.26
and Theorem 3.29.
18. Consider the operator Q on L2 [0; 1) dened in (3.3). Is it both time
invariant and causal? Derive the explicit formula for its adjoint Q .
Is Q time invariant?

Notes and references

For additional background material on linear functional analysis consult
[12] for an excellent introduction; also see [149] for more basic information
on Hilbert space. More advanced work on Hilbert space operator theory,
and a main source of references on this topic, is the book [54].
Proofs of all the results in this chapter (excepting Theorems 3.25 and
3.29) on the function spaces Lp , H2 and H1 can be found in [111]. These
proofs are for the scalar valued case, but easily generalize to the vector or
matrix valued case. In particular Lebesgue theory is developed there, and
proofs of the Plancherel and Paley-Wiener theorems are given. We point out
that some aspects of our presentation are non-standard, and dierent from
[111]. In particular the denition of the Fourier transform, pointwise as a
principal value limit is proved in [68]. Also our denitions for H2 and H1
are made on the closed, rather than open, half-plane. These modications
to the usual denitions have been done for clarity of exposition, and do not
aect any results.
Theorem 3.25 is proved in [11, Thms71-72] for the scalar case; the ex-
tension to the vector case given here is routine. Theorem 3.29 follows from
130 3. Linear Analysis
Theorem 3.25 and Proposition 3.26, and standard analytic continuation
theory; see [111]. See also [139] for more information on time invariant op-
erators and functions. The book [37] has many results on the discrete time
operator valued case. This book also contains much additional information
on the connections between function theory and operator theory, for both
the matrix and operator valued cases, and is a particularly good reference
for such results as related to control theory.
Classic references for work on Hp and related spaces are [32, 57]; an-
other important reference is [43]. Additional advanced information on these
spaces in the operator valued case can be located in [129, 110].
 This is page 131
Printer: Opaque this

4
Model realizations and reduction

In this chapter we start our investigation of quantitative input and output

properties of systems. To do this we will require the state space systems the-
ory of Chapter 2 combined with the new viewpoint and framework gained
in the preceding chapter. We rst consider issues related to the relative con-
trollability and observability of system states, and their relationships with
the overall input-output characteristics of a system. We then turn to the
important question of systematically nding reduced order approximations
to systems. We will develop a powerful technique to accomplish this, and
the operator perspective of the previous chapter will play a central role.

4.1 Lyapunov equations and inequalities

In this section we describe a basic tool for the study of stability, control-
lability and observability in linear systems. This topic is closely related to
the material in Chapter 2, but we have chosen to present it in this chapter,
where it will be extensively applied for the rst time.
Lyapunov equations come in the two dual forms
A X + XA + Q = 0
and
AX + XA + Q = 0:
In both cases A, Q are given square matrices, and X is the unknown. For
concreteness we will focus on the rst case, but by the trivial substitution
132 4. Model realizations and reduction
of A by A , all the following properties have their dual counterpart. Notice
in particular that the Hurwitz property is preserved by the \" operation.
We now see how Lyapunov equations can be solved explicitly when A is
Hurwitz.
Theorem 4.1. Suppose A and Q are square matrices and that A is
Hurwitz. Then
Z 1
X= eA QeA d (4.1)
0
is the unique solution to the Lyapunov equation A X + XA + Q = 0.
Proof . First, we verify that the given X is a solution, observing that
Z 1 d
A X + XA = A A A  A 1 
0 d fe Qe gd = e Qe 0 = ;Q:
It remains to show that it is the unique solution. For this purpose, dene
the linear mapping ; : C nn ! C nn by
;(X ) := A X + XA ;
where X 2 C nn . Now for any matrix Q 2 C nn we have seen that the
equation
;(X ) = Q
has a solution. This means that the dimension of the image of ; is n2 ; but
the dimension of the domain of ; is also n2 and therefore
ker ; = 0 :
This means ;(X ) = Q has a unique solution for each Q. 
In terms of computation, we emphasize that the Lyapunov equation is
linear in the matrix variable X , and therefore it is in essence no more than
a system of n2 linear equations in the n2 entries of X , that can be solved by
standard techniques. In particular when A is Hurwitz we are guaranteed to
have a unique solution; direct computation of the integral in (4.1) is never
required.
Another useful property of the Lyapunov solution is that
       
I 0 A 0 I 0 = A 0 = A 0 :
X I Q ;A ;X I A X + XA + Q ;A 0 ; A
Thus X can also be found by performing an eigenvalue decomposition.
From now on we focus on the case where Q = Q , for which one al-
ways seeks solutions X = X  to the Lyapunov equations. The following
result shows that the Hurwitz nature of A is directly related to the sign
deniteness of the solutions.
Proposition 4.2. Suppose Q > 0. Then A is Hurwitz if and only if there
exists a solution X > 0 to the Lyapunov equation A X + XA + Q = 0.
 4.1. Lyapunov equations and inequalities 133
Proof . We rst establish the \only if" direction. Since A is Hurwitz we
already know that the unique solution is
Z 1
X= eA QeA d:
0
Now since Q > 0 it is straightforward to verify that X > 0. We leave details
as an exercise.
For the converse, suppose X > 0 solves the equation. If  is an eigenvalue
of A, with eigenvector v 6= 0, we have
0 = v (A X + XA + Q)v =  v Xv + v Xv + v Qv:
Since v Xv > 0 we have

2Re() = ; vv Xv
Qv < 0;
so A is Hurwitz. 
We make the following remarks:
 The preceding proof is purely algebraic; an alternative argument
which helps explain our terminology, is to consider the Lyapunov
function V (x) = x Xx as a tool to show that x_ = Ax is asymptot-
ically stable. This approach can be found in standard references on
dierential equations.
 The preceding result remains true if the hypothesis Q > 0 is weakened
to Q = C  C  0, with (C; A) observable. The verication of this
stronger theorem is covered in the exercises at the end of the chapter.
In that case, the Lyapunov solution is denoted by Yo , satisfying
A Yo + Yo A + C  C = 0; (4.2)
and is called the observability gramian of (C; A); this is the topic of
the next section.
 It is sometimes convenient to rewrite Proposition 4.2 as a Linear
Matrix Inequality:
Corollary 4.3. The matrix A is Hurwitz if and only if there exists X > 0
satisfying
A X + XA < 0:
We nish the section elaborating some more on such Lyapunov
inequalities. In general, these have the form
A X + XA + Q  0
or the strict version
A X + XA + Q < 0:
134 4. Model realizations and reduction
Focusing on the case where A is Hurwitz, it is clear that these two LMIs
always have solutions. To see this notice that the left hand side of the
inequality in the above corollary can be made less than any matrix simply
by scaling any solution X . Now we ask: how do the solutions to these
inequalities relate to the corresponding Lyapunov equation solution? We
have the following result, which is left as an exercise.
Proposition 4.4. Suppose that A is Hurwitz, and X0 satises AX0 +
X0 A + Q = 0, where Q is a symmetric matrix. If X satises A X + XA +
Q  0, then
X  X0 :
This concludes our general study of Lyapunov equations. We now focus
on the problem of quantifying the degree of controllability and observability
of state-space realizations.

4.2 Observability operator and gramian

We begin our study by focusing on the autonomous system given by
x_ (t) = Ax(t); x(0) = x0 2 C n
y(t) = Cx(t);
where A is a Hurwitz matrix. This is our usual state space system with the
input set to zero, and a non-zero initial condition on the state. The solution
to this system is CeAt x0 , for t  0. Dene the observability operator o :
C n ! L2 [0; 1) by
(
x0 7! Ce x0 for t  0;
At
0 otherwise.
This operator is analogous to the map encountered in Chapter 2; here,
however, we exploit the fact that A is Hurwitz. In particular, we can easily
show there exist positive constants  and  satisfying
jCeAt x0 j  
e;t jx0 j ;
for all initial conditions x0 and times t. If we set y = ox0 this gives
kyk2  p jx0 j ;
2
so we see that y 2 L2 [0; 1) and o is a bounded operator. Let us now take
a more careful look at the energy of y = o x0 , for x0 2 C n . It is given by
kyk22 = h o x0 ; o x0 i = hx0 ; o ox0 i ;
 4.2. Observability operator and gramian 135
where o : L2 [0; 1) ! C n and is the adjoint of o. An exact expression
for the adjoint operator can be found from its denition and the following
equations. For each z 2 L2 [0; 1) and x0 2 C n
Z 1
h o z; x0 iC n = hz; ox0 i2 = z ( )CeA x0 d
0 
Z 1
= eA C  z ( )d x0
0
Z 1 
= eA C  z ( )d; x0 :
0 Cn

Thus we see that o is given by

Z 1
 z = eA C  z ( )d ;
0
for z 2 L2 [0; 1). Now the 2-norm of y = o x0 is given by hx0 ; o o x0 i
and we have
Z 1
o ( ox0 ) = eA  C  CeA x0 d
0
Z 1 
= eA  C  CeA d x0 = ( o o )x0 :
0
Therefore the operator o o is given by the matrix
Z 1
Yo :=  o =
o eA C  CeA d ;
0
which is the observability gramian of (C; A), already encountered in the
previous section. In fact, using Theorem 4.1 it is clear that Yo satises the
Lyapunov equation (4.2).
This gramian is positive semi-denite
Yo  0;
and as mentioned before is positive denite when (C; A) is observable. How-
ever the previous calculation shows that the gramian carries quantitative
information in addition to the a yes or no answer regarding observability.
We have shown that the energy of the output y = o x0 , starting from an
initial condition x0 2 C n , is given by
kyk22 = x0 Yo x0 : (4.3)
In particular, if we only consider states that satisfy jx0 j = 1, then clearly
some of these states will yield higher output norms kyk2 than others.
Therefore the gramian measures \how observable" a given initial condition
is.
136 4. Model realizations and reduction

Figure 4.1. Observability ellipsoid

This idea can be described geometrically in a very intuitive way. Consider

the vector
1
Yo2 x0 ;
where x0 has unit length. This vector has length equal to the 2-norm of
y, from (4.3), and as we will see contains directional information as well.
Dene the set
1
E := fYo2 x0 : x0 2 C n and jx0 j = 1g ;
which is the collection of all such vectors. Since Yo is positive semi-denite
this set is an ellipsoid, and is depicted two dimensionally in Figure 4.2.
This shape is called the observability ellipsoid, and tells us the output
norm associated with a particular direction in state space. Let
1  2 


 n  0 ;
1
be the eigenvalues of Yo2 and
v1 ; : : : ; vn ;
their respective unit-norm eigenvectors. Then the vk give the directions of
the principal axes of the ellipsoid, and the k the length of each axis.
If k = 0 for some k, the corresponding vk produces no output energy
and is therefore unobservable. In fact the span of all vk 's with k = 0 is
precisely the unobservable subspace NCA from Chapter 2. Geometrically,
the ellipsoid degenerates into a subspace of lower dimension, orthogonal to
this space.
If k , l are both nonzero but, for example, k >> l , then the output
energy resulting from initial state vl is much smaller than that observed
when the initial condition is vk . Thus both states are observable, but intu-
itively state vk is \more observable" than state vl . We conclude that the
observability gramian provides us with a way to assess the relative observ-
ability of various directions in the state space. When a state corresponds
to a large eigenvalue k we say it is strongly observable; if instead k > 0
 4.3. Controllability operator and gramian 137
is small compared with the other eigenvalues we say the state is weakly
observable.
As an aside let us see the signicance of this gramian for determining
the observability of a pair (C; A) where A is Hurwitz. In Chapter 2 we saw
how this could be done by checking whether the matrix
 
Q = C  A C 


(A )n;1 C 
has rank n. Another alternative, also based on rank, is the PBH test. This
raises the question of numerically determining the rank of a matrix, in par-
ticular the question of numerical tolerance for this determination; perhaps
one would check the eigenvalues of Q Q, but we have no interpretation
assigned to the size of these eigenvalues. In particular, you will show in
the exercises that these are unrelated to the eigenvalues of Yo . In short,
the above rank test can only provide a yes or no answer. In contrast, the
eigenvalues of Yo have a clear association with the degree of observability
and thus provide a much more sound numerical method for observability
assessment.
In line with this reasoning, it seems natural that these eigenvalues could
be used as a basis of a model reduction scheme, based on the elimination
of weakly observable states. We will return to this idea later, but rst we
study the dual situation of controllability.

4.3 Controllability operator and gramian

Here our focus is the dual idea of that pursued in the preceding sec-
tion. Given a matrix pair (A; B ) dene the controllability operator c :
L2 (;1; 0] ! C n by
Z 0
u 7! e;A Bu( )d :
;1
Notice this can be thought of as the response of the system described by
x_ (t) = Ax(t) + Bu(t); x(;1) = 0 (4.4)
to an input function u 2 L2 (;1; 0] where the output is the vector x(0).
To see the signicance of this operator, we ask the following question:
Given x(0) 2 C n with unit norm, jx0 j = 1, what u 2 L2 (;1; 0]
solves
cu = x0
with the smallest norm kuk2?
In system terms this question is, what is the input with smallest energy
kuk2 which drives the state in (4.4) to satisfy x(0) = x0 at time zero?
138 4. Model realizations and reduction
Clearly if the matrix pair (A; B ) is not controllable it may not be possible
even to satisfy cu = x0 , so for the next result we focus on the case where
the pair is controllable.
Proposition 4.5. Suppose that (A; B) is controllable. Then
(i) the matrix c c =: Xc is nonsingular;
(ii) For any x0 2 C n , the input uopt := c Xc;1x0 is the element of
minimum norm in the set
fu 2 L2 (;1; 0]; cu = x0 g: (4.5)
Proof . First, we nd an expression for Xc. Proceeding analogously to the
observability case, the denition of the adjoint implies that c : C n !
L2 (;1; 0] is given by
(
 ;A t for t  0;
 :  7! B e 
c 0 otherwise,
and therefore
Z 0 Z 1
Xc = c  =c e;At BB  e;At dt = eAtBB  eAt dt;
;1 0
called the controllability gramian of the matrix pair (A; B ). Notice that
this is an innite horizon version of the denition given in Chapter 2, and
also that Xc satises the Lyapunov equation
AXc + XcA + BB  = 0;
dual to the observability case. It follows from here that (A; B ) controllable
if and only if Xc > 0, which establishes (i).
Next, we verify that given x0 2 C n ,
cuopt = c c Xc;1 x0 = x0 ;
and thus u = uopt belongs to the set of allowable inputs described in (4.5).
So it remains to show, for a general u in this set, that kuk2  kuoptk. Dene
the operator
P = c Xc;1 c
on L2 (;1; 0], and observe that P 2 = P = P  . An operator satisfying these
identities is called an orthogonal projection. In particular, it satises
hPu; (I ; P )ui = hu; P (I ; P )ui = 0;
and consequently
kuk22 = kPuk22 + k(I ; P )uk22  kPuk22;
for any u 2 L2(;1; 0]. This is called the Bessel inequality.
 4.3. Controllability operator and gramian 139
Now let u 2 L2(;1; 0] be any function satisfying cu = x0 . Applying
c Xc;1 on both sides we get
Pu = c Xc;1 cu = c Xc;1 x0 = uopt ;
so the Bessel inequality gives
kuk22  kuopt k22:

The proposition says that if we want to reach a state x0 then
uopt = c Xc;1x0
is the most economical input in terms of energy. This input energy is given
by
kuoptk22 = h c Xc;1 x0 ; c Xc;1 x0 i
= hXc;1 x0 ; c c Xc;1x0 i
= x0 Xc;1 x0 :
We now provide a geometric interpretation of the controllability gramian,
similar to that for Yo . The key question is: what are all the nal states
x0 = c u
that can result from an input u 2 L2(;1; 0] of unit norm? The answer
is given in the following proposition, which holds even if the pair (A; B )
is not controllable. We will, however, concentrate here on the controllable
case; the general case is covered in the exercises at the end of the chapter.
Proposition 4.6. The following sets are equal:
(a) f cu : u 2 L2 (;1; 0] and kuk2  1g;
1
(b) fXc2 xc : xc 2 C n and jxc j  1g.
Proof . We rst show that set (a) is contained in set (b): choose u 2
L2 (;1; 0] with kuk2  1. Set
1
xc = Xc; 2 cu;
which has norm
1 1
jxc j2 = hXc; 2 cu; Xc; 2 c ui = hu; c Xc;1 cui = hu; Pui = kPuk22  kuk22  1:
1
This means that c u = Xc2 xc is in the set (b).
We now demonstrate that set (b) is contained in set (a). Let xc be any
vector that has length less than or equal to one. We choose the input of
140 4. Model realizations and reduction
1
minimum norm which gives cu = Xc2 xc . From the previous results this
input exists and its norm is
1 1
kuopt k22 = (Xc2 xc ) Xc;1(Xc2 xc ) = jxc j22  1:
1
Therefore Xc2 xc is in the set (a).

This result says that all the states reachable with u satisfying kuk2  1
are given by
1
Xc2 xc ;
where jxc j  1. Notice the norm squared of any such state is given by
xc Xc xc . We dene the controllability ellipsoid by
1
Ec = fXc2 xc : xc 2 C n and jxc j = 1g :
This gives us the boundary of the set given in the proposition. Let
1  2 


 n  0 ;
1
be the eigenvalues of Xc2 and
v1 ; : : : ; vn
their corresponding orthonormal eigenvectors. Then clearly the principal
axes of the ellipsoid Ec are given by
k vk :
The unit vectors v1 ; : : : ; vn give a basis for the state space and the values
1 ; : : : ; n tell us that given an input kuk2 = 1, the largest we can make
a state in the direction vk is k . Thus we conclude that if k = 0, then vk
is an unreachable state. In the same vein, if k >> l then state direction
vk is \more" controllable than direction vl . As with observability this gives
rise to the terms strong and weak controllability.

4.4 Balanced realizations

The preceding two sections have given us geometric ways, in terms of
the gramians introduced, to assess which directions in the state space are
strongly or weakly controllable and observable. We now look at our usual
system given by
x_ (t) = Ax(t) + Bu(t); x(0) = x0
y(t) = Cx(t);
 4.4. Balanced realizations 141
for t  0. Here the direct feedthrough term given by the D-matrix is not
included, since it makes no dierence to our investigation. We are inter-
ested in nding a natural basis for this state space that gives some idea
as to which states dominate the system behavior. In our study so far we
concluded that states corresponding to small eigenvalues of the observabil-
ity gramian are not very observable. Does this mean such states do not
contribute much to the input-output behavior of the above system? The
answer is not necessary since such states may be very controllable. This
phenomenon is easily appreciated by looking at Figure 4.2, which shows
both the observability and controllability ellipsoids on the same plot.

Figure 4.2. Unbalanced system ellipsoids

As this picture shows, these two ellipsoids need not be aligned, and there-
fore we can have a situation as drawn here, where the major and minor
axes of the ellipsoids are nearly opposite. Therefore we reason that the
most intuition would be gained about the system if the controllability and
observability ellipsoids were exactly aligned, and is certainly the natural
setting for this discussion. This raises the question of whether or not it is
possible to arrange such an alignment by means of a state transformation;
answering this is our next goal.
A change of basis to the state space of the above system yields the
familiar transformed realization
A~ = TAT ;1; B~ = TB; and C~ = CT ;1 ;
where T is a similarity transformation. The controllability gramian
associated with this new realization is
1 ~
Z
X~ c = eA B~ B~  eA~ d
Z0 1
= TeA T ;1TBB  T (T  );1 eA T d
0
= TXcT 
142 4. Model realizations and reduction
Similarly the observability gramian in the transformed basis is
Y~o = (T  );1 Yo T ;1 :
The above are called congruence transformations, which typically arise
from quadratic forms under a change of basis. The following result concerns
simultaneous diagonalization of positive denite matrices by this kind of
transformation.
Proposition 4.7. Given positive denite matrices X and Y , there exists
a nonsingular matrix T such that
TXT  = (T );1 Y T ;1 = 
where  is a diagonal, positive denite matrix.
Proof . Perform a singular value decomposition of the matrix X 21 Y X 21 to
get
X 12 Y X 21 = U 2 U  ;
where U is unitary and  is diagonal, positive denite. Therefore we get
; 12 U  X 12 Y X 12 U ; 12 = :
Now set T ;1 = X 21 U ; 21 and the above states that (T ;1 ) Y T ;1 = .
Also
TXT  = ( 12 U  X ; 21 ) X (X ; 12 U  21 ) = :

Applying this proposition to the gramians, the following conclusion can
be drawn:
Corollary 4.8. Suppose (A; B; C ) is a controllable and observable realiza-
tion. Then there exists a state transformation T such that the equivalent
~ B;
realization (A; ~ C~ ) = (TAT ;1; TB; CT ;1) satises
X~c = Y~o = 
with  > 0 diagonal.
A state space realization such that the controllability and observability
gramians are equal and diagonal, is called a balanced realization. The pre-
vious corollary implies that there always exists a balanced realization for
a transfer function in RH1 ; in fact, starting with any minimal realization
(which will necessarily have A Hurwitz), a balanced one can be obtained
from the above choice of state transformation.
Clearly the controllability and observability ellipsoids for the system are
exactly aligned when a system is balanced. Thus the states which are the
least controllable are also the least observable. Balanced realizations play
a key role in the model reduction studies of the rest of this chapter.
 4.5. Hankel operators 143
Before leaving this section, we state a generalization of Proposition
4.7 which covers the general case, where the original realization is not
necessarily controllable or observable.
Proposition 4.9. Given positive semi-denite matrices X and Y there
exists a nonsingular matrix T such that
2 3
1
(a) TXT  = 64
6 2 7
0 5,
7

0
2 3
1
6
(b) (T  );1 Y T ;1 = 64 0 7
3 5,
7

0
where the matrices k are diagonal and positive denite.
When applied to the gramians Xc and Yo , we nd that if the system is
either uncontrollable or unobservable, then each of the k blocks of Y~o and
X~c have the following interpretation:
 1 captures controllable and observable states
 2 captures controllable and unobservable states
 3 captures observable and uncontrollable states
 4 captures unobservable and uncontrollable states.
Under such a transformation the state matrix A~ has the form
2~
A1 0 A~6 0 3
6~ ~ ~4 A~5 77
A~ = 64A02 A03 AA~7 0 5 ;
0 0 A~8 A~9
which is the so-called Kalman decomposition; this can be veried from
the various invariance properties of the controllability and observability
subspaces.

4.5 Hankel operators

In the previous sections we have introduced an input-to-state operator c
and a state-to-output operator o that where naturally related to the de-
gree of controllability and observability of a state-space system. It seems
natural, then, that these pieces should come together to study input-output
properties of a system. These ideas are now explored.
144 4. Model realizations and reduction
We begin with a causal, bounded input-output operator G : L2 (;1; 1) !
L2 (;1; 1) described by the state space convolution
Z t
(Gu)(t) := CeA(t; ) Bu( )d;
;1
where A is a Hurwitz matrix. In the language of Chapter 3, this operator
is G = ;1 MG^ , where G^ (s) = C (sI ; A);1 B is a transfer function in
RH1 . We focus on the strictly proper case; the presence of a D term would
make no dierence to the discussion in this section.
Now we dene the Hankel operator ;G : L2(;1; 0] ! L2 [0; 1) of G by

;G := P+ GL2 (;1;0] ;

where GL2 (;1;0] denotes the restriction of G to the subspace L2(;1; 0],
and P+ is the operator projects a signal in L2(;1; 1) into L2 [0; 1) by
truncation. In other words, ;G takes an input supported in the \past", and
maps it to the future output, as illustrated in Figure 4.3 for scalar inputs
and outputs.

t
Gu

t
;G u

t
Figure 4.3. Results of G and ;G operating on a given u 2 L (;1; 0].
2

The Hankel operator is closely related to the controllability and observ-

ability operators, due to the causality of G; in fact, if u(t) is supported in
the past, the future output P+ y depends on u only through the initial state
x(0); thus we can think of composing two stages:
 c which maps u 2 L2 (;1; 0] to x(0).
 o which maps x(0) to y(t); t  0, with no input applied for t  0.
 4.5. Hankel operators 145
It is not dicult to formalize this idea and obtain
;G = o c;
using the denitions of these operators; we leave details to the reader. Thus
the following commutative diagram is satised.
;G
L2 (;1; 0] L2 [0; 1)

c o
Cn

Notice that ;G has at most rank n for a state space system of order n;
namely the dimension of Im(;G ) is at most n.
Our rst result on the Hankel operator tells us how to compute its
L2 (;1; 0] ! L2[0; 1) induced norm. Below max denotes the largest
eigenvalue of the argument matrix.
Proposition 4.10. The norm of ;G is given by
k;G kL2!L2 = (max (Yo Xc ) ) 21 ;
where Yo and Xc are the gramians of the realization (A; B; C ).
Proof . We begin with the following identity established in Chapter 3:
k;Gk2 = (;G ;G ) = ( c o o c ) ;
where (
) denotes spectral radius. We recall as well that the spectral radius
is invariant under commutation, and obtain
k;Gk2 = ( o o c c ) = (Yo Xc ) :
Now, Yo Xc is a matrix, and has only non-negative eigenvalues; for this
latter point,1 observe
12
that the nonzero eigenvalues of Yo Xc coincide with
those of Xc Yo Xc  0.
2
Therefore (Yo Xc ) = max (Yo Xc), which completes the proof. 
We remark that it can also be shown that ;G ;G , a nite rank operator,
has only eigenvalues in its spectrum, namely the eigenvalues of Yo Xc in
addition to the eigenvalue zero. For this reason, in analogy to the matrix
case, the square roots of the eigenvalues of Yo Xc are called the singular
values of ;G or the Hankel singular values of the system G. We order these
and denote them by
1  2 


 n :
146 4. Model realizations and reduction
Clearly 1 = k;Gk.
Since the operator ;G depends only on the input-output properties of G,
it is clear that the Hankel singular values must be the same for two equiv-
alent state-space realizations. This can be veried directly for realizations
related by a state transformation T . Recall that the gramians transform
via
Y~0 = (T ;1 ) Yo T ;1;
X~c = TXc T  ;
and therefore
Y~0 X~c = (T ;1) Yo XcT :
Thus we have Y~0 X~c and Yo Xc are related by similarity and their eigenvalues
coincide.
If we have realizations of dierent order (one of them non-minimal) of a
given G^ , their Hankel singular values can only dier on the value 0, which
appears due to lack of observability or controllability; in fact, the minimal
order can be identied with the number of non-zero Hankel singular values
of any realization.
In the special case of a balanced realization Xc = Y0 = , we see that
the diagonal of  is formed precisely by the Hankel singular values.
We now turn to the question of relating the Hankel singular values to the
norm of the original operator G. While these are dierent objects we will
see that important bounds can be established. The rst one is the following.
Proposition 4.11. With the above denitions,
1 = k;G k  kGk = kG^ k1:
Proof . Since the projection P+ has induced norm 1, then
 
k;Gk = kP+ GL2 (;1;0] k  kGL2 (;1;0] k  kGk:
For the last step, notice that the norm of an operator cannot increase when
restricting it to a subspace. In fact, it is not dicult to show that this step
is an equality. 
Thus the largest Hankel singular value provides us with a lower bound
for the H1 norm of the transfer function G^ . The next result provides an
upper bound in terms of the Hankel singular values.
Proposition 4.12. Suppose G^(s) = C (Is;A);1 B and that A is a Hurwitz
matrix. Then
kG^ k1  2(1 +


+ n ) ;
where the k are the Hankel singular values associated with (A; B; C ).
 4.6. Model reduction 147
A proof with elementary tools is given in the exercises at the end of the
chapter. In the next section we will see a slight renement of this bound
using more advanced methods.
The Hankel operator plays an important part in systems theory, as well
as pure operator theory, and is intimately related to the question of ap-
proximation in the k
k1 norm. The latter has many connections to robust
control theory; in this chapter we will use it in connection to model re-
duction. For further discussion consult the references at the end of the
chapter.

4.6 Model reduction

We are now ready to consider approximation of systems. Frequently in
modeling or control synthesis it is of value to use a simple model that ap-
proximates a more complex one. In our context, complexity is measured by
the dimensionality of state-space realizations. Given a system, we would like
to reduce the order of a state-space realization, while keeping the system
input output properties \approximately" the same.
An extreme case of model reduction we have already studied, is the
elimination of uncontrollable or unobservable states. As shown in Chapter
2, this can be done without aecting the input output transfer function, i.e.
with \zero error". We are now ready to take the next step, that is to allow
some error, and study the following problem: given a transfer function G^ (s)
of order n, nd a transfer function G^ r (s) of order r < n, such that G^ and
G^ r are \close".
Before this problem has a precise meaning, a notion of distance between
transfer functions must be chosen; indeed there are many metrics that can
be used, all based on the input-output behavior of the systems. We will
focus on the H1 model reduction problem where
^ G^ r ) = kG^ ; G^ r k1 = kG ; Gr kL2!L2 :
dist(G;
This choice deserves some comment.
In the rst place, with this choice we are restricting our model reduction
to stable systems G, i.e. systems which correspond to bounded operators
on L2 . Later on in the course we will remark on how these techniques can
be extended to study distance between unstable systems.
Secondly, even under the stability restriction, there are other system
norms which could be candidates for model approximation. The reason
for our preference towards an induced norm can be illustrated as follows.
Suppose our system G is to be connected in cascade with another system
H ; if we are going to use the approximation Gr instead of G, we would
like to be able to easily quantify the error between the composed systems
HG and HGr . For any norm satisfying the submultiplicative inequality we
148 4. Model realizations and reduction
have
kHG ; HGr k = kH k
kG^ ; G^ r k;
which tells us how to propagate our error to the cascaded system. Thus we
see the convenience of operator norms, in particular the H1 norm.
We recapitulate our discussion by stating more explicitly the H1 model
reduction problem:
 
Given: transfer function G^ (s) = C A B = C (sI ; A);1 B +
D
D, where A 2 Rnn is Hurwitz.  
Find: lower order function G^ r (s) = A r Br where A 2
Cr Dr r
R r  r such that
kG^ ; G^ r k1 is minimized:
An immediate observation is that we can assume D = 0, since
   
A B ^ A B ^
C D ; Gr = C 0 ; (Gr ; D)
and feed through terms do not aect the order of the realization.
What is not so clear is whether Dr can be set to zero; in other words, is
the best approximation to a strictly proper system, itself strictly proper?
Indeed there may be advantages in using Dr 6= D, as is shown in the
exercises at the end of the chapter; in this section, however, we concentrate
on the case where both G^ and G^ r are strictly proper.
As stated, the H1 model reduction problem is open, in the sense that
there is no known computationally tractable method to obtain the optimal
approximation of a given order. In the absence of this optimal solution, we
will study model reduction methods in terms of bounds on the approxi-
mation error. This includes both lower bounds on the error achievable by
any model reduction method, and upper bounds on the error achieved by
a specic scheme, in our case the balanced truncation method.

4.6.1 Limitations
Faced with the approximation problem above it is natural for us to ask
whether there are some fundamental limits to how well we can approximate
a given system with a lower order one. To work towards the answer, as
well as to build intuition, we begin by studying a matrix approximation
problem that is strongly related to the system approximation problem:
given a square matrix N 2 Rnn , how well can it be approximated, in
the maximum singular value norm, by a matrix of rank r? We have the
following result:
 4.6. Model reduction 149
Lemma 4.13. Suppose that N 2 Rnn , with singular values 1  2 



 n . Then for any R 2 Rnn , rank(R)  r,
 (N ; R)  r+1 :
Proof . We start by taking the singular value decomposition of N :
2 3
1
N = U V  where  = 64 ... 7
5 ;
n
and matrices U; V are both unitary.
Now dene the column vectors vk by
[v1


vn ] = V ;
and consider the r+1-dimensional subspace described by spanfv1 ; : : : ; vr+1 g.
Since the subspace ker(R) has dimension at least n ; r, these two sub-
spaces must intersect non-trivially. Therefore spanfv1 ; : : : ; vr+1 g contains
an element x, with jxj = 1, such that
Rx = 0 :
The vectors vk are orthonormal since V is unitary and so x can be expressed
by
r+1
X r+1
X
x= k vk with 2k = 1;
k=1 k=1
for appropriately chosen scalars k .
We now let the matrix N ; R act on this vector x to get
r+1
X r+1
X
(N ; R)x = Nx = k Nvk = k k uk
k=1 k=1
where we dene uk by [u1


un ] = U . The uk are orthonormal and so we
have
r+1
X
j(N ; R)xj2 = k2 2k :
k=1
To complete the proof notice that since the singular values are ordered we
have
r+1
X r+1
X
k2 2k  r2+1 2k = r2+1 ;
k=1 k=1
and so
 (N ; R)  j(N ; R)xj  r+1 :

150 4. Model realizations and reduction
The preceding Lemma shows us that there are fundamental limits to the
matrix approximation problem in the  (
) norm, and that singular values
play a key role in characterizing these limits. What does this have to do with
the system model reduction problem? The key connection is given by the
Hankel operator, since as we saw the dimension of a state-space realization
is the same as the rank of the Hankel operator. Not surprisingly, then, we
can express the fundamental limitations to model reduction in terms of the
Hankel singular values:
Theorem 4.14. Let 1  2 


 r  r+1


 n > 0 be the
Hankel singular values associated with the realization (A; B; C ) of G^ . Then
for any G^ r of order r,
kG^ ; G^ r k1  r+1 :
This result says that if we are seeking a reduced order model of state
dimension r, then we cannot make the approximation error smaller than
the (r + 1)-th Hankel singular value of the original system.
Proof . To begin we have that
kG^ ; G^ r k1  k;G;Gr k = k;G ; ;Gr k :
It therefore suces to demonstrate that
k;G ; ;Gr k  r+1 :
As we have noted before these Hankel operators satisfy
rank (;G )  n and rank (;Gr )  r ;
which is a fact we will now exploit.
We recall the denitions of the observability and controllability operators
o and c associated with (A; B; C ), and the identity
;G = o c :
Now we dene the maps Po : L2[0; 1) ! Rn and Pc : Rn ! L2 (;1; 0] by
1 1
Po = Y0; 2 o and Pc = c Xc; 2 ;
and verify that kPo k = kPck = 1. We therefore have
k;G ; ;Gr k = kPo k
k;G ; ;Gr k
kPc k   (Po ;G Pc ; P0 ;Gr Pc ) ;
(4.6)
where the submultiplicative inequality is used. We further have that
1 1 1 1
Po ;G Pc = Yo2 o o c c Xo2 = Yo2 Xc2 ;
which has rank equal to n since its singular values are 1 ; : : : ; n which are
all positive. But matrix Po ;Gr Pc has rank at most equal to r since rank
(;Gr )  r. Invoke Lemma 4.13 to see that
 (Po ;GPc ; Po ;Gr Pc )  r+1 :
 4.6. Model reduction 151
Now apply (4.6) to arrive at the conclusion we seek.


4.6.2 Balanced truncation

We now move on to one of the major methods for nding a reduced order
approximation. Recall we showed that a balanced realization for an input-
output system can always be found. Formally this means a realization such
that the gramians satisfy
2 3
1
6 2 7
Xc = Yo = 664 ...
7
7
5
:
n
Also we saw that if a given k was small compared with the other eigen-
values, the corresponding state had the interpretation of being weakly
controllable and observable. The technique for model reduction we now
consider, essentially amounts to discarding or truncating such states.
Here we necessarily assume that the realization (A; B; C ) is a balanced
realization and order the Hankel singular values
1  2 


 n :
Let us suppose we want to nd a realization of order r < n. We assume
that the strict inequality
r+1 < r is satised;
i.e. we only attempt to truncate the system at an order where the Hankel
singular values have a clear separation. Then we compatibly partition our
realization as
   
A= A 11 A12 B1
A21 A22 ; B = B2 ; and C = [C1 C2 ] ;
where A11 2 Rrr . The reduced order model is chosen to be
 
G^ r (s) = C1 (Is ; A11 );1 B1 = AC11 B01 :
1
The minimum requirement on G^ r is that A11 be Hurwitz, and this is indeed
the case.
Proposition 4.15.
(a) The matrix A11 , as dened above, is Hurwitz;
(b) The realization (A11 ; B1 ; C1 ) is balanced with Hankel singular values
1 ; : : : ; r .
152 4. Model realizations and reduction
Proof . We begin by writing down the Lyapunov equations of the balanced
realization (A; B; C ), highlighting the relevant partition:
        
A11 A21 1 0 + 1 0 A11 A12 + C1 [C C ] = 0;
A12 A22 0 2 0 2 A21 A22 C2 1 2
(4.7)
       
A11 A12 1 0 + 1 0 A11 A21 + B1 [B  B  ] = 0:
A21 A22 0 2 0 2 A12 A22 B2 1 2
(4.8)
Here 1 and 2 are diagonal matrices with, respectively 1 ; : : : ; r , and
r+1 ; : : : ; n in the diagonal. Notice that, by assumption, they have no
eigenvalues in common.
From here we extract the top left blocks and obtain
A11 1 + 1 A11 + C1 C1 = 0; (4.9)
 
A11 1 + 1 A11 + B1 B1 = 0: (4.10)
We see then that if A11 is Hurwitz, we immediately have that 1 is a bal-
anced gramian for the truncated system, proving (b). Thus we concentrate
on part (a).
Let  be an eigenvalue of A11 , and V be a full column rank matrix that
generates the eigenvector space, i.e.
ker(A11 ; I ) = Im(V ):
Multiplying (4.9) on the left by V  and on the right by V leads to
( + )V  1 V + V  C1 C1 V = 0:
Since V  1 V > 0 we see that Re()  0. Thus it only remains to rule out
the possibility of a purely imaginary eigenvalue; we make this assumption
from now on.
Setting  = j!, the preceding equation implies that the corresponding
V must satisfy
C1 V = 0:
Now we multiply (4.9) only on the right by V ; we conclude that
A11 (1 V ) = ;j!1V:
We have now an invariant subspace of A11 , so we turn to an analogous
study of the dual equation (4.10). First we multiply it on the left by V  1
and on the right by 1 V to show
B1 1 V = 0;
then we return to (4.10) and perform only the right multiplication by 1 V
to yield
A11 (21 V ) = j!21 V:
 4.6. Model reduction 153
So we are back with an eigenspace of A11 ; in particular, we have
Im(21 V )  ker(A11 ; j!I ) = Im(V );
namely Im(V ) is invariant under 21 . We can now nish this exercise in
invariant subspaces by concluding that 21 must have an eigenvector v in
Im(V ), satisfying
21 v = 2 v:
We have thus found v 6= 0 that is simultaneously an eigenvector of A11
with eigenvalue j!, and of 21 , with eigenvalue 2 . In addition, from the
previous development we have
C1 v = 0 and B1 1 v = 0:
We claim that
    
A11 A12 v = j! v : (4.11)
A21 A22 0 0
Equivalently, we must show that A21 v = 0. To do this we return to the
original Lyapunov equations (4.7) and (4.8) and multiply them respectively
by
   
v and 1 v :
0 0
Looking only at the second row of equations and using the above properties
of v, we get
A12 1 v + 2 A21 v = 0;
A21 2 v + 2 A12 1 v = 0;
which lead to the key relation
22 A21 v = 2 A21 v:
At this point (only!) we bring in the assumption that 1 and 2 have no
common eigenvalues; thus 2 cannot be an eigenvalue of 22 , so A21 v must
be zero, which proves (4.11).
Now (4.11) contradicts the fact that A is Hurwitz, and thus the
assumption that  = j! is an eigenvalue of A11 is ruled out.

In the exercises you will show, by counterexample, that the requirement
of distinct eigenvalues for 1 and 2 (r > r+1 ) is indeed essential to
ensure the stability of the truncation.
Having discussed stability, we turn to the key model reduction question:
what is the H1 norm error achieved by this method? This type of question
may appear to be premature at this point in the course, since we have
not yet discussed any general purpose method for H1 norm computation.
154 4. Model realizations and reduction
However the model reduction error can be quantied by a \special purpose"
argument based on so-called \allpass dilations"; to develop it requires a
short diversion into the topic of inner transfer functions, after which we
will return to the balanced truncation error.

4.6.3 Inner transfer functions

We begin by recalling the notion of an isometric operator between Hilbert
spaces, characterized by the identity
U U = I;
where U  is the usual adjoint. Such operators preserve norms and hence
have induced norm kU k = 1.
Focusing now on linear-time invariant operators characterized by a
transfer function U^ (s) 2 RH1 , we inquire which of these are isometric.
Converting to frequency domain we see that U is an isometry if and only
if
hU^ x^; U^ x^i; holds for all x 2 L2
Therefore we conclude that a system denes an isometric operator if and
only if its transfer function satises
U^ (j!) U^ (j!) = I; for all !:
Rational transfer functions satisfying the above identity are called inner
functions. The term allpass function is also used sometimes, since in the
scalar case these transfer functions have unit gain at every frequency.
The next step is to obtain a state-space test for inner functions. For this
purpose, notice that if
 
U^ (s) = A B C D
then the realization
   
U^  (s) = ;BA ;DC
satises U^  (j!) = U^ (j!) . The rational function U^  (s) is called the para-
Hermitian conjugate of U^ (s). If U^ (s) 2 RH1 , U^  (s) will be analytic on
the closed left half-plane.
Composing U^  and U^ , we obtain the joint realization
2 3
;A ;C  C ;C  D
U^  (s)U^ (s) = 4 0 A B 5 : (4.12)
B D C D D
Clearly, the only way the above rational function can be constant across
frequency, as required for U^ to be inner, is that all states must be uncon-
 4.6. Model reduction 155
trollable or unobservable. An important case where this can be guaranteed
is given in the next lemma.
 
Lemma 4.16. Consider the transfer function U^ (s) = CA DB where A
is Hurwitz. If the observability gramian Yo of (A; C ) satises the additional
property C  D + Yo B = 0, then
U^  (s)U^ (s) = D D:
Proof . Starting with the realization (4.12), we introduce the similarity
transformation
 
T = I0 ;IYo
which leads to the equivalent realization
2
;A ;(A Yo + Yo A + C  C ) ;(C  D + Yo B ) 3
U^  (s)U^ (s) = 4 0 A B 5:
B D C + B  Yo D D
Now the Lyapunov equation for Yo , and the additional property in the
hypothesis give
2
;A 0 0 3
U^  (s)U^ (s) = 4 0 A B 5 = D D;
B  0 D D
since all states in this latter realization are uncontrollable or unobservable.

We are now ready to return to the model reduction problem.

4.6.4 Bound for the balanced truncation error

In this section we develop a bound for the H1 norm error associated with
balanced truncation. This bound will show that if the r+1 ; : : : ; n are
small as compared with the remaining Hankel singular values, then Gr is
a near approximation to G.
We rst consider the case where the n ; r smallest Hankel singular values
are equal.
Lemma 4.17. Suppose the Hankel singular values of G satisfy
1  2 


 r > r+1 = r+2


= n :
If G^ r is obtained by r-th order balanced truncation, then kG^ ; G^ r k1 
2r+1 .
156 4. Model realizations and reduction
Proof . We assume that r+1 6= 0 otherwise we are immediately done.
Start by dening the error system
E^11 (s) = G^ (s) ; G^ r (s) ;
which has state space realization
2 3
A11 0 0 B1
E^11 (s) = 64 00 A 11 A12 B1 7
6
A21 A22 B2 5 :
7

;C1 C1 C2 0
Notice that by Proposition 4.15, this realization has a Hurwitz \A"-matrix.
Our technique of proof will be to construct an \allpass dilation" of E11 (s),
by nding
 
E 11 (s
E (s) = E (s) E (s)) E 12 (s )
21 22
that contains E11 as a sub-block and is, up to a constant, inner. If this can
be done, the error norm can be bounded by the norm of E (s) which is easy
to compute.
We rst use the similarity transformation
2 3
I I 0
T = 4I ;I 05 ;
0 0 I
to get
2 3
A11 0 A12 =2 B1
E^11 (s) = 64 A021 ;AA1121 ;AA1222=2 B02 75 :
6 7

0 ;2C1 C2 0
Next, we dene the augmentation as
2 3
A11 0 A12 =2 B1 0
6 0 A11 ;A12 =2 0 ;
r+1 1 C1 77
1
6
^
E (s) = 6
6 A 21 ; A 21 A 22 B 2 ;C2 77
4 0 ;2C1 C2 0 2r+1 I 5
;2r+1 B1 1 ; 1 0 
;B2 2r+1 I 0
 

A B :
=: C  D
Here we are using the notation
 
 = 01  0 I
r+1
for the balanced gramian of the original system G.
 4.6. Model reduction 157
While this construction is as yet unmotivated, the underlying aim is to be
able to apply Lemma 4.16. Indeed, it can be veried by direct substitution
that the observability gramian of the above realization is
2 3
4r2+1 ;1 1 0 0
Yo = 4 0 41 0 5;
0 0 2r+1 I
and that it satises the additional restriction
C  D + Yo B = 0:
While somewhat tedious, this verication is based only the Lyapunov
equations satised by the gramian .
Therefore we can apply Lemma 4.16 to conclude that
E (j!) E (j!) = D  D = 4r2+1 I for all !;
and therefore we have
kE^11 k1  kE^ k1 = 2r+1 :

We now turn to the general case, where the truncated eigenvalues
r+1 ; : : : ; n are not necessarily all equal. Still, there may be repetitions
among them, so we introduce the notation 1t ; : : : kt to denote the distinct
values of this tail. More precisely, we assume that
1t > 2t >


> kt
and fr+1 ; : : : ; n g = f1t ; : : : kt g. Equivalently, the balanced gramian of
the full order system is given by
2 3
1
6
6
... 7
7
2
1
3
6
6 r
7
7 6 1t I 7
= 6
6 r+1
7
7 = 664 ...
7
7
5
6 7
6
4 ... 7
5 kt I
n
where the block dimensions of the last expression correspond to the number
of repetitions of each it .
We are now ready to state the main result:
Theorem 4.18. With the above notation for the Hankel singular values of
G^ , let G^ r be obtained by r-th order balanced truncation. Then the following
inequality is satised:
kG^ ; G^ r k1  2(1t +


kt ):
158 4. Model realizations and reduction
Proof . The idea is to successively apply the previous Lemma, truncat-
ing one block of repeated Hankel singular values at every step. By virtue
or Proposition 4.15, in this process the realizations we obtain remain bal-
anced and stable, and the Hankel singular values are successively removed.
Therefore at each step we can apply the previous Lemma and obtain an
error bound of twice the corresponding it . Finally, the triangle inequality
implies that the overall error bound is the sum of these terms.
More explicitly, consider the transfer functions G^ (0) ; : : : ; G^ (k) , where
G k) = G^ , and for each i 2 f1; : : : ; kg, G^ (i;1) is obtained by balanced
^ (
truncation of the repeated Hankel singular value it of G^ (i) . By induction,
each G^ (i) has a stable, balanced realization with gramian
2 3
1
6
6 1t I 7
7
6
4 ... 7
5
:
it I
Therefore Lemma 4.17 applies at each step and gives
kG^ (i) ; G^ (i;1) k1  2it :
Also G^ (0) = G^ r , so we have
k 
X 
kG^ ; G^ r k = k G^ (i) ; G^ (i;1) k1
i=0
k
X
 kG^ (i) ; G^ (i;1) k1  2(1t +


kt ):
i=0

As a comment, notice that by specializing the previous result to the
case r = 0 we can bound kG^ k1 by twice the sum of its Hankel singular
values; since repeated singular values are only counted once, this is actually
a tighter version of Proposition 4.12.
We have therefore developed a method by which to predictably reduce
the dynamic size of a state space model, with the H1 norm as the quality
measure on the error.
Example:
We now give a numerical example of the uses of this method. Consider the
7-th order transfer function in RH1 ,
(s + 10)(s ; 5)(s2 + 2s + 5)(s2 ; 0:5s + 5)
G^ (s) = (s + 4)( s2 + 4s + 8)(s2 + 0:2s + 100)(s2 + 5s + 2000) ;
with magnitude Bode plot depicted in the top portion of Figure 4.4.
 4.6. Model reduction 159
0
10

−5
10
−1 0 1 2 3 4
10 10 10 10 10 10

0
10

−2
10

−4
10

−6
10
−1 0 1 2 3 4
10 10 10 10 10 10

Figure 4.4. Log-log plots for the example. Top: jG^ (j!)j. Bottom: jG^ (j!)j (dotted),
jG^ (j!)j (full) and jG^ (j!)j (dashed).
4 2

The Hankel singular values of G^ can be computed to be

2 3 2 3
1 0:17926
62 7 60:178787
6 7 6 7
63 7 60:107687
6 7 6 7
64 7 = 60:107567 :
6 7 6 7
65 7 60:000767
6 7 6 7
46 5 40:000085
7 0:00003
Given these values, it is apparent that a natural model reduction would be
to r = 4 states, since the remaining states are clearly very weakly observable
and controllable. We may also wish to study a more drastic reduction to
r = 2 states. The magnitude bode plots of G^ 4 and G^ 2 , obtained by balanced
truncation, are shown, together with G^ , in the bottom portion of Figure
4.4. We see that G^ 4 \picks-up" with great accuracy the high-frequency
behavior of the system (where it has its greatest gain), but commits errors
at low frequency. G^ 2 focuses its approximation on the largest resonant peak,
around ! = 10.
160 4. Model realizations and reduction
We can compute the model reduction bounds from Theorem 4.18, and
also (e.g. by frequency gridding, or with methods to be covered later in the
course) the actual H1 errors. We nd
0:0014 = kG^ ; G^ 4 k1  2(5 + 6 + 7 ) = 0:0017;
0:2153 = kG^ ; G^ 2 k1  2(3 + 4 + 5 + 6 + 7 ) = 0:4322:


We conclude the section with two remarks:
First, we emphasize again that we are not claiming that the balanced
truncation method is optimal in any sense; the problem of minimizing kG^ ;
G^ r k1 remains computationally dicult.
Second, given the presence of the Hankel operator and its singular values
in the above theory, we may inquire: does balanced truncation minimize the
Hankel norm k;G^ ; ;G^ r k? Once again, the answer is negative. This other
problem, however, can indeed be solved by methods of a similar nature.
For details consult the references at the end of the chapter.

4.7 Generalized gramians and truncations

The goal of this nal section is to generalize some of the above methods to
LMI tools, essentially by replacing Lyapunov equations with inequalities.
We will see that this extra exibility provides some valuable generalizations
and additional insight into the model reduction problem.
Let us generalize our notion of system gramians. A symmetric matrix X
is called a generalized controllability gramian if it satises the Lyapunov
inequality
AX + XA + BB   0;
and similarly a symmetric matrix Y is a generalized observability gramian
if A Y + Y A + C  C  0 holds.
The system gramians Xc and Yo are both generalized gramians, but are
not unique in this regard; in general there is a set of generalized gramians,
even when (A; B; C ) is a minimal realization. What special property to
the actual gramians enjoy? From Proposition 4.4 we know, in fact, they
are the minimal elements among the generalized gramians, i.e.
X  Xc ; Y  Yo (4.13)
for any generalized gramians X , Y .
 4.7. Generalized gramians and truncations 161
As with Xc and Yo we say that two generalized gramians X and Y are
balanced when they are diagonal and equal:
2 3
1
X = Y = 64 ... 7
5 :
n
Generalized gramians can be balanced in the exact same way as before;
when (A; B; C ) is minimal, X and Y are positive denite and we can simply
apply the simultaneous diagonalization result of Proposition 4.7.
The entries k are called the generalized Hankel singular values. It can
be shown, based on the inequalities (4.13), that with the order 1  2 



 n , we have as expected
i  i ; i = 1 : : : n:
Generalized gramians can also be used for model reduction by truncation;
we have the following generalization to Proposition 4.15 and Theorem 4.18:
 
Proposition 4.19. If G^ = CA B0 is minimal, with A Hurwitz, and
 
G^ r = Ar Br is formed by truncating according to the balanced
Cr 0
generalized gramians, then Ar is Hurwitz and
kG^ ; G^ r k1  2( 1t +


kt ) ; (4.14)
where it denote the distinct generalized Hankel singular values which satisfy
it < r .
The proof of this result follows readily from our earlier balanced truncation
result, and is included in the exercises.
At rst sight, the previous generalization may not appear to have a
particular application, because we have seen that the generalized Han-
kel singular values are always larger than the i . Therefore it seems that
Theorem 4.18 would always give a smaller error bound.
Notice, however, that there is the issue of repetition in the singular val-
ues. Suppose, for example, that a third order system has Hankel singular
values (1 ; 2 ; 3 ) = (1; 0:4; 0:3). If by some means we managed to ob-
tain generalized Hankel singular values ( 1 ; 2 ; 3 ) = (1; 0:4; 0:4), then in a
truncation to r = 1 states the generalized method will have a smaller error
bound.
Can such repetition always be arranged? The interesting fact is that
the answer is armative, and can be used to characterize exactly the H1
model reduction problem. This is stated in the following result.
 
Theorem 4.20. Given G^ = CA DB , with A 2 Rnn Hurwitz, the
following are equivalent:
162 4. Model realizations and reduction
 
(i) There exists G^r = A r Br
C D of order r such that kG^ ; G^r k1 < .
r r
(ii) There exist X > 0, Y > 0 satisfying
AX + XA + BB  < 0;
A Y + Y A + C  C < 0;
min(XY ) = 2 ; with multiplicity at least n ; r:
We will not prove this theorem here, since it will follow as a corollary of a
general result on H1 control, to be proved in Chapter 7. We will also see
there how G^ r can be constructed if we have the appropriate X , Y .
This theorem says that if there exists an H1 model reduction with error
less than , then we can always nd (strict) generalized gramians such
that the smallest generalized Hankel singular value is , repeated n ; r
times. In particular, there would be only one term in the error bound from
truncation (4.14), although truncation is not the method we will use to
obtain the above result. Notice also that the factor of 2 does not appear in
the H1 norm bound, which is related to the use of a term Dr 6= D (see
the exercises).
At this point, it might appear that we can solve the H1 model reduction
problem, since we have an exact nite dimensional characterization of when
a reduction of order r with a certain error can be achieved. The diculty
is that the above conditions, while easy to state, are not computationally
tractable. In particular the condition on the minimum eigenvalue is not
easy to enforce in a convex way. This means that the above theorem cannot
truly be considered a solution. Nevertheless, it is a valuable addition to our
insight on this problem, since it provides a very compact description of
where the computational diculty lies.
There are other important reasons to work with generalized gramians,
perhaps the most compelling one is that these generalize to multidimen-
sional and uncertain systems, as discussed in the references. In addition, we
remark that for discrete time systems, generalized gramians appear natu-
rally since the actual gramians are not preserved by truncation; for more
details see the exercises and the references.

4.8 Exercises
1. Suppose A, X and C satisfy A X + XA + C  C = 0. Show that any
two of the following implies the third:
(i) A Hurwitz.
(ii) (C; A) observable.
(iii) X > 0.
2. Prove Proposition 4.4.
 4.8. Exercises 163
3. Use the controllability canonical form to prove Proposition 4.6 in the
general case of uncontrollable (A; B ).
4. Controllability gramian vs. controllability matrix. We have seen that
the singular values of the controllability gramian Xc can be used to
determine \how controllable" the states are. In this problem you will
show that the controllability matrix
Mc = [B AB A2 B


An;1 B ]
cannot be used for the same purpose, since its singular values are
unrelated to those of Xc . In particular, construct examples (A 2
C 22 ; B 2 C 21 suces) such that:
a) Xc = I , but (Mc ) is arbitrarily small.
b) Mc = I , but (Xc ) is arbitrarily small.
5. Proof of Proposition 4.12.
a) For an integrable matrix function M (t), show that
!
Z b Z b
 M (t) dt   (M (t)) dt:
a a
You can use
 the fact
 that the property holds in the scalar case.
A B
b) Let G^ = C 0 . Using a) and the fact that G^ (j!) is the
Fourier transform of CeAt B , derive the inequality
Z 1
kG^ k1   (Ce At2 )  (e At2 B )dt:
0
c) If Xc , Yo are the gramians of (A; B; C ), show that
Z 1
 (Ce At2 )2 dt  2Tr(Yo );
0
Z 1
 (e At2 B )2 dt  2Tr(Xc ):
0
d) Combine b) and c) for a balanced realization (A; B; C ), to show
that
kG^ k1  2(1 +


+ n ) :
6. In Proposition 4.15 the strict separation r > r+1 of Hankel singular
values was used to ensure the stability of the truncation (i.e. that
A11 is Hurwitz). Show that indeed this is a necessary requirement, by
constructing an example (e.g. with n = 2, r = 1) where the truncated
matrix is not Hurwitz.
7. Consider the transfer function G^ (s) = s+1
1 .
164 4. Model realizations and reduction
a) Find the (only) Hankel singular value of G^ . Compare the error
bound and actual error when truncating to 0 states (clearly the
truncation is G^ 0 = 0).
b) Show that by allowing a nonzero dr term, the previous error can
be reduced by a factor of 2. Do this by solving explicitly
min
d2R
kG^ ; dk1 :
8. a) Let X and Y be two positive denite matrices in C nn . Show
that the following are equivalent:
 
(i) X  Y ;1 ; (ii) min (XY )  1; (iii) XI YI  0:
b) Under the conditions of part a), show that: min(XY ) = 1, with
multiplicity k, if and only if XI YI  0 and has rank 2n ; k.
c) Let (A; B; C ) be a state-space realization of G^ (s) with A
Hurwitz. Assume there exist X0 , Y0 in C rr such that
   
A X00 I 0 + X00 I 0 A + BB  < 0
n;r n;r
   
A Y00 I 0 + Y00 I 0 A + C  C < 0
n;r n;r
 
X0 Ir  0
Ir Y0
Using Theorem 4.20 prove that there exists a r-th order re-
alization G^ r (s) with kG^ ; G^ r k1 < . Is the above problem
convex?
9. Discrete time gramians and truncation. This problem concerns the
discrete time system
xk+1 = Axk + Bwk ;
zk = Cxk :
(a) The discrete Lyapunov equations are
A Lo A ; Lo + C  C = 0;
AJc A ; Jc + BB  = 0:
Assume that each eigenvalue in the set eig(A) has absolute value
less than one (stable in the discrete sense). Explicitly express the
solutions Jc and Lo in terms of A, B and C ; prove these solutions
are unique.
A realization in said to be balanced if Jc and Lo are diagonal
and equal. Is it always possible to nd a balanced realization?
 4.8. Exercises 165
(b) In this part we assume, in contrast to (a), that J > 0 and
L > 0 are generalized gramians that are solutions to the strict
Lyapunov inequalities
A LA ; L + C  C < 0;
AJA ; J + BB  < 0:
Show that Lo < L and Jc < J . Can generalized gramians be
balanced?  
(c) Suppose we are given generalized gramians J = L = 01 I 0 ,
where 1 > 0 is diagonal. Partition the state space accordingly:
   
A= A 11 A12
A21 A22 ; B= B 1
B2 ;
 
C = C1 C2 ;
and dene
 
G^ 11 = AC11 B01 :
1
Show that (A11 ; B1 ; C1 ) is a balanced realization in the sense of
generalized gramians, and that all the eigenvalues of A11 have
absolute value less than one.
10. Prove Proposition 4.19. Use the following steps:
a) Use the balanced Lyapunov inequalities to show that there exits
an augmented system
2 3
A B Ba
G^ a = 4 C D 0 5
Ca 0 0
whose gramians are X and Y .
b) Apply Theorem 4.18 directly to G^ a to give a reduced system
G^ ar . Show that G^ r of Proposition 4.19 is embedded in G^ ar and
must therefore satisfy the proposition.

Notes and references

Balanced realizations were rst introduced into the control theory litera-
ture in [84], motivated by principal component analysis. In this paper weak
and strong controllability and observability were introduced, and balanced
truncation model reduction was suggested. A proof that balanced trun-
cation yields a stable system was subsequently given in [97]. The initial
proof of the error bound in Theorem 4.18 appeared in [33], and an inde-
pendent but subsequent proof in [46]. The concise proof given here of the
166 4. Model realizations and reduction
key Lemma 4.17 is due to [46]. For computational methods on obtaining
balanced realizations see [76] and [116].
The books [96] and [103] can be consulted for in depth treatments of
Hankel operators.
Another major approach to model reduction, also involving Hankel sin-
gular values is that of so-called optimal Hankel norm approximation. The
main results in this area were presented in [46], based on scalar results
appearing in [1]. A more compact treatment of this approach is given in
[47]. An interesting comment is that this approach, in conjunction with a
nonzero Dr term, can be used to improve the error bound in Theorem 4.18
by a factor of two.
The discrete time version of balanced truncation model reduction was
initiated in [2] using discrete gramians. The use of gramians in discrete
time only yields an analog of Lemma 4.17, but surprisingly Theorem 4.18
does not hold in general: the induction argument breaks down since bal-
ancing is not preserved by truncation. In [56] it was shown that if strict
generalized gramians are used in lieu of gramians then a discrete version of
Proposition 4.19 does hold.
The use of generalized gramians to characterize exactly the H1 model
reduction problem, is essentially from [67], although our version of Theo-
rem 4.20 emphasizes the LMI methods we will use later for H1 control
synthesis. A discrete time version, based on LMI methods, was obtained
in [7]. In this latter work, notions of balanced realizations and model re-
duction are extended to multi-dimensional and uncertain systems. We will
give an outline of these generalizations in Chapter 11.
 This is page 167
Printer: Opaque this

5
Stabilizing Controllers

We begin here our study of feedback design, which will occupy our atten-
tion in the following three chapters. We will consider systematic design
methods in the sense that objectives are rst specied, without restricting
to a particular technique for achieving them.
The only a priori structure will be a very general feedback arrangement
which is described below. Once introduced, we will focus in this chapter
on a rst necessary specication for any feedback system: that it be stable
in some appropriate sense. In particular we will precisely dene feedback
stability and then proceed to parametrize all controllers that stabilize the
feedback system.

z w
G
u

y
K

Figure 5.1. General feedback arrangement

168 5. Stabilizing Controllers
The general feedback setup we are concerned with is shown above. As
depicted the so-called closed-loop system has one external input and one
output, given by w and z respectively. The signal or function w captures
the eects of the environment on the feedback system; for instance noise,
disturbances and commands. The signal z contains all characteristics of the
feedback system that are to be controlled. The maps G and K represent
linear subsystems where G is a given system, frequently called the plant,
which is xed, and K is the controller or control law whose aim is to ensure
that the mapping from w to z has the desired characteristics. To accomplish
this task the control law utilizes signal y, and chooses an action u which
directly aects the behavior of G.
Here G and K are state space systems, with G evolving according to
 
x_ (t) = A x(t) + B1 B2 wu((tt)) ;
 

      
z (t) = C1 x(t) + D11 D12 w(t) ;
y(t) C2 D21 D22 u(t)
and K being described by
x_ K (t) = AK xK (t) + BK y(t);
u(t) = CK xK (t) + DK y(t):
Throughout this chapter we have the standing assumption that the matrix
triples (C; A; B ) and (CK ; AK ; BK ) are both stabilizable and detectable.
As shown in the gure, G is naturally partitioned with respect to its two
inputs and two outputs. We therefore partition the transfer function of G
as
2 3
A B1 B2 
^ ^ 
G^ (s) = 4 C1 D11 D12 5 = G^ 11 (s) G^ 12 (s) ;
C2 D21 D22 G21 (s) G22 (s)
so that we can later refer to these constituent transfer functions.
At rst we must determine under what conditions this interconnection of
components makes sense. That is, we need to know when these equations
have a solution for an arbitrary input w. The system of Figure 5.1 is well-
posed if unique solutions exist for x(t), xK (t), y(t) and u(t), for all initial
conditions x(0), xK (0) and all (suciently regular) input functions w(t).
Proposition 5.1. The connection of G and K in Figure 5.1 is well-posed,
if and only if, I ; D22 DK is nonsingular.
Proof . Writing out the state equations of the overall system, we have
x_ (t) = Ax(t) + B1 w(t) + B2 u(t) (5.1)
x_ K (t) = AK xK (t) + BK y(t);
 5.1. System Stability 169
and
       
I ;DK u(t) 0 CK x(t) 0
;D22 I y(t) = C2 0 xK (t) + D21 w(t): (5.2)
Now it is easily seen that the left hand side matrix is invertible if and only
if I ; D22 DK is nonsingular. If this holds, clearly one can substitute u,
y into (5.1) and nd a unique solution to the state equations. Conversely
if this does not hold, from (5.2) we can nd a linear combination of x(t),
xK (t), and w(t) which must be zero, which means that x(0), xK (0), w(0)
cannot be chosen arbitrarily. 
Notice in particular that if either D22 = 0 or DK = 0 (strictly proper
G^ 22 or K^ ), then the interconnection in Figure 5.1 is well-posed.
We are now ready to talk about stability.

5.1 System Stability

In this section we discuss the notion of internal stability, and discuss its
relation to the boundedness of input-output maps. Internal stability was al-
ready introduced in Chapter 2 for an autonomous system; we now specialize
the denition to our feedback arrangement.
Denition 5.2. The system in Figure 5.1 is internally stable if it is well-
posed, and for every initial condition x(0) of G, and xK (0) of K , the limits
!1 0 hold ;
x(t); xk (t) t;!
when w = 0.
The following is an immediate test for internal stability.
Proposition 5.3. The system of Figure 5.1 is internally stable if and only
if I ; D22 DK is invertible and
    ;1  
A = A0 A0 + B02 B0 ;DI ;DI K 0 CK
C2 0 (5.3)
K K 22
is Hurwitz.
Proof . Given the well-posedness condition as in Proposition 5.1, it is easy
to solve from (5.1) and (5.2) to show that A is the A-matrix of the closed-
loop; therefore the result follows from our work in Chapter 2. 
As dened, internal stability refers to the autonomous system dynamics
in the absence of an input w; in this regard it coincides with the stan-
dard notion of asymptotic stability of dynamical systems. However it has
immediate implications on the input-output properties of the system. In
particular, the transfer function from w to z , denoted T^(s), is proper and
170 5. Stabilizing Controllers
will have as poles a subset of the eigenvalues of A; for example, when
DK = 0 we have
 
T^(s) = C1 D12 CK (Is ; A);1 B BD1 + D11 :
 
K 21
If A is Hurwitz, this function has all its poles in the left half plane: in the
language of Chapter 3, it is an element of RH1 . Equivalently, the closed
loop map w 7! z is a bounded, causal operator on L2 [0; 1); this is termed
input-output stability.
The question immediately arises as to whether the two notions are in-
terchangeable, that is whether the boundedness of w 7! z implies internal
stability. Clearly, the answer is negative: an extreme example would be to
have C1 , D11 , D12 be all zero which gives T^(s) = 0 but clearly says nothing
about A. In other words, the internal dynamics need not be re ected in the
external map: they could be unobservable or uncontrollable..
We are still interested, however, in seeking an external characterization
of internal stability; this will allow us, in a later section, to describe all
possible stabilizing controllers. Such characterization should only impose
non-restrictive requirements on the internal dynamics. From this point of
view, the variables w and z are not relevant to the discussion, since the
conditions of Proposition 5.3 depend exclusively on the components in u
and y of the state space system G.
These considerations lead us to discuss a second interconnection diagram,
represented in Figure 5.2. This is precisely the same as the bottom loop in
Figure 5.1, except that we have injected interconnection noise. In particular,
the controller K has the same description as before, and the system G22 is
the lower block of G, described by the state space equations
x_ 22 (t) = Ax22 (t) + B2 v1 (t)
v2 (t) = C2 x22 (t) + D22 v1 (t);
where (C2 ; A; B2 ; D22 ) are the same matrices as in the state space
description of G.
The only addition to the diagram are the external inputs d1 and d2 ,
introduced at the interconnection between G22 and K .
As with our more general system, we say that this new system is well-
posed if there exist unique solutions for x22 (t), xK (t), v1 and v2 for all
inputs d1 (t) and d2 (t) and initial conditions x22 (0), xK (0). We say it is
internally stable if it is well posed and for di = 0 we have
!1 0 holds
x22 (t); xK (t) t;!
for every initial condition.
It is an easy exercise to see that the system is well-posed, if and only
if, I ; D22 DK is nonsingular; this is the same well-posedness condition we
have for Figure 5.1. Also noticing that all the states in the description of G
 5.1. System Stability 171

v2 v1 d1
G22

d2 K
Figure 5.2. Input-output stability

are included in the equations for G22 , it follows immediately that internal
stability of one is equivalent to internal stability of the other.
Lemma 5.4. Given a controller K , Figure 5.1 is internally stable, if and
only if, Figure 5.2 is internally stable.
The next result shows that with this new set of inputs, internal stability
can be characterized by the boundedness of an input-output map.
Lemma 5.5. Suppose that (C2 ; A; B2) is stabilizable and detectable. Then 
Figure 5.2 is internally stable if and only if the transfer function of dd1 7!
  2
v1 is in RH .
v2 1
Proof . We begin by nding an expression for the transfer function. For
convenience denote
 
D = I ;DK
;D22 I ;
then routine calculations lead to the following relationship:
   
v^1 (s) = M^ (s) d^(s) ;
v^2 (s) d^2 (s)
where
     
^  ; 0 C K  ;
M (s) := D C 0 (Is ; A) 0 B D + D + 0 ;I ;
1 1 B 2 0  ; 1  ; 1 0 0
2 K
and A is the closed loop matrix from (5.3). Therefore the \only if" direction
follows immediately, since the poles of this transfer function are a subset
of the eigenvalues of A, which is by assumption Hurwitz; see Proposition
5.3 and Lemma 5.4.
172 5. Stabilizing Controllers
To prove \if": assume that the transfer function has no poles in C+ ,
therefore the same is true of
   
0 CK (Is ; A);1 B2 0
C2 0 0 BK :
| {z } | {z }
C B
We need to show that A is Hurwitz; it is therefore sucient to show that
 A;
(C;  B ) is a stabilizable and detectable realization. Let
   
F = F0 F0 ; D ;1 C0 C0K ;
K 2
where F and FK are chosen so that A + B2 F and AK + BK FK are both
Hurwitz. It is routine to show that
 
A + B F = A +0B2 F A + 0B F ;
K K K
and thus (A; B ) is stabilizable.
A formally similar argument shows that (C;  A) is detectable. 
The previous characterization will be used later on to develop a
parametrization of all stabilizing controllers. As for the stabilizability
and detectability assumption on (C2 ; A; B2 ), we will soon see that it is
non-restrictive, i.e. it is necessary for the existence of any stabilizing
controller.

5.2 Stabilization
In the previous section we have discussed the analysis of stability of a
given feedback conguration; we now turn to the question of design of
a stabilizing controller. The following result explains when this can be
achieved.
Proposition 5.6. A necessary and sucient condition for the existence of
an internally stabilizing K for Figure 5.1, is that (C2 ; A; B2 ) is stabilizable
and detectable. In that case, one such controller is given by
 
K^ (s) = A + B2 F + LC 2 + LD22 F ;L ;
F 0
where F and L are matrices such that A + B2 F and A + LC2 are Hurwitz.
Proof . If the stabilizability or detectability of (C2 ; A; B2) is violated, we
can choose an initial condition which excites the unstable hidden mode. It
is not dicult to show that the state will not converge to zero, regardless
of the controller. Details are left as an exercise. Consequently no internally
stabilizing K exists, which proves necessity.
 5.2. Stabilization 173
For the suciency side, it is enough to verify that the given controller
is indeed internally stabilizing. Notice that this is precisely the observer-
based controller encountered in Chapter 2. In particular, well-posedness is
ensured since DK = 0, and it was shown in Chapter 2 that the closed loop
eigenvalues are exactly the eigenvalues of A + B2 F and A + LC2 . 
The previous result is already a solution to the stabilization problem,
since it provides a constructive procedure for nding a stabilizer, in this
case with the structure of a state feedback combined with an asymptotic
observer. However there are other aspects to the stabilization question, in
particular:
 Can we nd a stabilizer with lower order? The above construction
provides a controller of the order of the plant.
 What are all the stabilizing controllers? This is an important ques-
tion since one is normally interested in other performance properties
beyond internal stability.
We will address the rst question using purely LMI techniques, starting
with the state feedback problem, and later considering the general case.
The second problem will then be addressed in Section 5.3.

5.2.1 Static state feedback stabilization via LMIs

We begin our search for stabilizing controllers in terms of LMIs by consid-
ering the simplest type of control, static state feedback. A state feedback
controller has direct access to the state of G, namely
y = x:
This means that C2 = I , D22 = 0. The feedback is static if it has no
dynamics, i.e. it is given simply by
u(t) = DK y(t):
As a beginning for controller parametrization we would like to nd all such
DK that are stabilizing. That is, referring to Proposition 5.3, all matrices
DK such
AL := A + B2 DK
is Hurwitz. We already know that such a DK exists exactly when (A; B2 ) is
stabilizable, and we now seek a constructive way to nd such DK 's in terms
of an LMI. The natural path to follow is to impose that AL is Hurwitz by
means of a Lyapunov inequality of the type discussed in Chapter 4. Here
we will impose that the inequality
AL X + XAL < 0
174 5. Stabilizing Controllers
has a solution X > 0 (dual to the test in Corollary 3, Chapter 4).
Substituting for AL we get
(A + B2 DK )X + X (A + B2 DK ) < 0;
that can be expanded out as
AX + B2 (DK X ) + XA + (XDK )B2 < 0:
The above is not an LMI in DK and X since their product appears. However
it is natural to introduce the new variable Y = DK X , which makes the
above an LMI in X and Y ; once this LMI is solved, the desired DK can be
easily reconstructed.
Thus we nd that all stabilizing feedback laws can be found from a
convex feasibility problem. We summarize the conclusions in the following
statement.
Theorem 5.7. A static state feedback DK stabilizes Figure 5.1, if and only
if, there exist matrices X > 0 and Y such that
DK = Y X ;1
and
   

A B2 X

+

X Y   A < 0 is satised:
Y B2
We are now ready to move on to a more general LMI characterization of
xed order controllers.

5.2.2 An LMI characterization of the stabilization problem

Given a xed controller order, a direct approach to the stabilization prob-
lem is to write down a generic controller realization (AK ; BK ; CK ; DK ),
and impose that the closed loop matrix A in (5.3) is Hurwitz by means of
the Lyapunov inequality
XL > 0; A XL + XLA < 0:
The question is whether the above condition, with unknowns XL, AK , BK ,
CK and DK can be computed tractably. The following is a result in this
direction, stated for simplicity in the case of a strictly proper plant.
Theorem 5.8. Consider the system of Figure 5.1, with D22 = 0. There
exists a controller of order nK , which internally stabilizes the system, if
 5.3. Parametrization of stabilizing controllers 175
and only if, there exist n  n matrices X > 0, Y > 0, such that
NX (A X + XA)NX < 0; (5.4)
NY (AY + Y A )NY < 0;
 
(5.5)
X I
I Y  0; and (5.6)
 
rank XI I  n+n ;
Y K (5.7)
where NX and NY are full column rank matrices such that
ImNX = ker C2 ;
ImNY = ker B2 :
We make the following remarks: the LMIs (5.4) and (5.5) are equiva-
lent, respectively, to the detectability and stabilizability of the realization
(C2 ; A; B2 ). Since they are homogeneous in the unknowns, the solutions
can always be scaled up to satisfy (5.6). This means that if there is no
constraint in the order nk , the existence of a stabilizer is equivalent to sta-
bilizability and detectability of (C2 ; A; B2 ), consistently with the previous
analysis.
Furthermore, clearly the rank constraint (5.7) is not felt when nK  n,
which implies that we never need an order greater than n to stabilize,
once again consistently with Proposition 5.6. In this case, the set of LMIs
(5.4-5.6) is a convex problem that can be solved for X , Y . For nK <
n, stabilizability and detectability may not suce to obtain a stabilizer;
furthermore, the rank constraint is not convex and makes the computation
of X , Y more dicult; still, the above proposition provides useful insight
into the role of controller order.
Once X , Y are obtained, algebraic operations can be used to compute
XL, and from it an additional LMI leads to the controller. We will not
extend ourselves on this construction here, or on the proof of Theorem 5.8,
since these topics will all be covered in Chapter 7, where we extend the LMI
method for the solution of the H1 control problem. After going through
Chapter 7, the reader is invited to prove Theorem 5.8 as an exercise.

5.3 Parametrization of stabilizing controllers

In this section we pursue the objective of fully characterizing the family of
all linear controllers which stabilize a given plant.
This is a very important tool since in many problems one wishes
to optimize a performance objective with stability as a constraint. A
parametrization of the stabilizing controllers reduces these problems to
unconstrained optimization over the parameter.
176 5. Stabilizing Controllers
The parametrization will be strongly based on transfer function meth-
ods, for which we will exploit the input-output characterization of internal
stability, summarized in the following corollary of Lemmas 5.4 and 5.5 and
Proposition 5.6.
Corollary 5.9. A necessary and sucient condition for the existence of an
internally stabilizing controller for Figure 5.1, is that (C2 ; A; B2 ) is stabi-
lizable and detectable. In that case, the internally stabilizing controllers are
exactly those for which the conguration of Figure 5.2 makes the transfer
function of
   
d1 7! v1
d2 v2
an element of RH1 .
It is important to note that given any two appropriately dimensioned func-
tions Q^ 1 and Q^ 2 in RH1 , their product Q^ 1 Q^ 2 will be in RH1 , as will be
their sum Q^ 1 + Q^ 2 ; namely RH1 has the structure of an algebra. The fol-
lowing subsection introduces tools to analyze the structure of this algebra,
which are key to the controller parametrization.

5.3.1 Coprime factorization

Throughout this section we work mostly with transfer functions. So far we
have used the \hat" convention to denote a transfer function, for instance
T^(s). We will now also use the \wave" notation to denote functions, as in
Q~ (s).
Given two matrix valued functions M^ and N^ in RH1 , they are right
~ Y~ 2 RH1 such that
coprime if there exist X;
X~ M^ + Y~ N^ = I :
Similarly, M~ and N~ are left coprime if there exist X^ and Y^ in RH1 , such
that
M~ X^ + N~ Y^ = I :
These equations are called Bezout identities.
A right coprime factorization of a proper rational function P^ is a
factorization
P^ = N^ M^ ;1 ;
where
^ M^ 2 RH1 ; M^ ;1 is proper,
N;
and N^ and M^ are right coprime. A left coprime factorization for P^ would
be
P^ = M~ ;1 N~ ;
 5.3. Parametrization of stabilizing controllers 177
with similar restrictions imposed on N~ and M~ as in a right factorization.
Example:
To gain insight into the denitions, consider the case of a scalar p^(s). In
this case a (right or left) coprime factorization is
n^ (s)
p^(s) = m^ (s)
where n^ (s), m^ (s) are in RH1 , and
x^(s)m^ (s) + y^(s)^n(s) = 1 for some x^(s); y^(s) 2 RH1 :
It is not dicult to show that coprimeness over RH1 is equivalent to
the property that n^ (s) and m^ (s) have no common zeros on the closed right
half plane C + , or at s = 1.
Take, for instance, p^(s) = ss;;21 ; then
n^ (s) = ss +; 1 ; m^ (s) = s ; 2
a s+a
denes a coprime factorization over RH1 for any a > 0, but
n^(s) = (ss+;a1)2 ; m^ (s) = (ss+;a2)2
does not. 
Do these factorizations exist in the general matrix case? The following
result shows constructively that this is the case for any proper rational
matrix.
Proposition 5.10. Given a proper rational function P^ (s), there exist both
right and left coprime factorizations
P^ = N^ M^ ;1 = M~ ;1 N~ ;
satisfying
  
X~ ;Y~ M^ Y^ = I (5.8)
;N~ M~ N^ X^
for appropriate functions X~ , Y~ , X^ and Y^ in RH1 .
Notice that the Bezout identities in (5.8) have a sign change as compared
to the ones above, which is of course inconsequential. Also, here X~ , Y~ X^
and Y^ satisfy an additional identity, and (5.8) is therefore termed a doubly
coprime factorization of P^ .
Proof . Our proof is by direct construction. Let
 
P^ (s) = CA D
B
178 5. Stabilizing Controllers
where (C; A; B ) is a stabilizable and detectable realization. Introduce the
state space system.
x_ (t) = Ax(t) + Bu(t) x(0) = 0
y(t) = Cx(t) + Du(t) :
Now choose a matrix F such that AF := A + BF is Hurwitz. Further dene
v(t) = u(t) ; Fx(t)
CF = C + DF :
Then we dene the system
x_ (t) = AF x(t) + Bv(t) x(0) = 0
u(t) = Fx(t) + v(t)
y(t) = CF x(t) + Dv(t) ;
whose solutions have a one-to-one correspondence with the solutions to the
rst system. The transfer function from v to u is
 
M^ (s) = AFF BI ;
and from v to y has transfer function
 
N^ (s) = AF B CF D :
Thus we see that
P^ (s) = N^ (s)M^ ;1 (s) ;
a fact that can also be checked via direct state space manipulations.
Similarly a factorization
P^ (s) = M~ ;1 (s)N~ (s) ;
can be found, with
   
M~ (s) = ACH HI ; N~ (s) = ACH BDH ;
where H is chosen such that AH = A + HC is Hurwitz, and BH = B + HD.
A quick way to derive these is to transpose the problem to
 
AT C T
P^ T (s) = B = N~ T (s)(M~ T (s));1
T DT
and reproduce the steps of the right coprime case.
The other four transfer functions are given by
   
X^ (s) = A F ;H Y^ (s) = AFF ;0H
CF I
 5.3. Parametrization of stabilizing controllers 179
and
   
X~ (s) = AH ;BH
F I Y~ (s) = AFH ;0H :
That these satisfy the claim can be veried by routine state space
manipulations.


5.3.2 Controller Parametrization

We are now ready to parametrize all stabilizing controllers for Figure 5.1.
From Corollary 5.9, we may focus on the input-output properties of the
arrangement of Figure 5.2. To motivate what follows, we start by looking
again at the case of scalar functions.
Example:
Let G^ 22 be a scalar transfer function with coprime factorization G^ 22 = mn^^ ((ss)) .
Let x^(s); y^(s) 2 RH1 satisfy
x^(s)m^ (s) ; y^(s)^n(s) = 1: (5.9)
Now we claim that K^ (s) = xy^^((ss)) is input-output stabilizing for Figure 5.2.
   
To see this, it is routine to check that the transfer function of dd1 !
7 v1
v2
2
is given by
     
1 1 K^ = 1 x^m^ y^m^ = x^m^ y^m^ :
1 ; K^ G^ 22 G^ 22 K^ G^ 22 x^m^ ; y^n^ x^n^ y^n^ x^n^ y^n^
Notice that the relationship (5.9) played a key role in making the common
denominator disappear. The last transfer function is in RH1 , which proves
stability.
In the same vein, noticing that
(^x ; n^ q^)m^ ; (^y ; m ^ q^)^n = 1
for any q^ in RH1 , it follows analogously that
K^ (s) = y^x^((ss)) ;; mn^^ ((ss)^
)^q (s)
q(s)
is a stabilizing controller for any q^ in RH1 that makes the above transfer
function proper. Thus we see that a coprime factorization generates a family
of stabilizing controllers over the parameter q^(s).
The striking fact, which is proved below for the general matrix case, is
that indeed all stabilizing controllers can be expressed in the latter form.

180 5. Stabilizing Controllers

v2 q v1
N M ;1 d1

p
d2 V ;1 U
Figure 5.3. Coprime factorization

We now proceed to extend the idea in the example to the matrix setting,
and also build up the theory necessary to show the ensuing parametrization
is complete. Let
  
X~ ;Y~ M^ Y^ = I
;N~ M~ N^ X^
be a doubly coprime factorization of G^ 22 , and
K^ = U^ V^ ;1 = V~ ;1 U~
be coprime factorizations for K^ .
The following gives us a new condition for stability.
Lemma 5.11. Given the above denitions the following are equivalent
(a) The controller K input-output stabilizes G22 in Figure 5.2;

^ ^
(b) M^ U^ is invertible in RH1 ;
N V
 
~ ~
(c) ;VN~ ;M~U is invertible in RH1 .
Proof . First we demonstrate that condition (b) implies condition (a). This
proof centers around Figure 5.3, that is exactly Figure 5.2 redrawn with
the factorizations for G22 and K . Clearly
    
M^ ;U^ q^ = d^1 : (5.10)
;N^ V^ p^ d^2
Thus we see
( )
I 0 + 0 U^ M^ ;U^ ;1
      
v^1 = d^1 :
v^2 0 0 N^ 0 ;N^ V^ d^2
 5.3. Parametrization of stabilizing controllers 181
The above transfer function inverse is in RH1 by assumption (b), therefore
we see that (a) holds.
Next we must show that (a) implies (b). To do this we use the Bezout
identity.
X~ M^ ; Y~ N^ = I :
Referring to the gure we obtain
M^ q^ = v^1
N^ q^ = v^2 :
Multiplying the Bezout identity by q^ and substituting we get
q^ = X~ v^1 ; Y~ v^2 :
Now by assumption the transfer functions from d^1 and d^2 to v^1 and v^2 are
in RH1 . Thus by the last equation the transfer function from the inputs
d^1 and d^2 to q^ must be in RH1 .
Similarly we can show that the transfer functions from the inputs to p^
are in RH1 , by instead starting with a Bezout identity X~K V^ + Y~K U^ = I
for the controller. Recalling the relationship in (5.10) we see that (b) must
be satised.
To show that (a) and (c) are equivalent we simply use the left coprime
factorizations for G^ 22 and K^ , and follow the argument above.

We can now prove the main synthesis result of the chapter.
Theorem 5.12. A controller K input-output stabilizes G22 in Figure 5.2,
if and only if, there exists Q^ 2 RH1 such that
K^ = (Y^ ; M^ Q^ )(X^ ; N^ Q^ );1 = (X~ ; Q^ N~ );1 (Y~ ; Q^ M~ ) (5.11)
and the latter two inverses exist as proper rational functions.
Proof . We begin by showing that the latter equality holds for any Q^ 2
RH1 such that the inverses exist. Given such a Q^ we have by the doubly
coprime factorization formula that
    
I Q^ X~ ;Y~ M^ Y^ I ;Q^ = I ;
0 I ;N~ M~ N^ X^ 0 I
which yields
  
X~ ; Q^ N~ ;(Y~ ; Q^ M~ ) M^ Y^ ; M^ Q^ = I (5.12)
;N~ M~ N^ X^ ; N^ Q^
182 5. Stabilizing Controllers
Taking this product we get nally
 
? (X~ ; Q~ N~ )(Y^ ; M^ Q^ ) ; (Y~ ; Q^ M~ )(X^ ; N^ Q^ ) = I :
? ?
Here \?" denotes irrelevant entries, and from the top right entry we see
that the two quotients of the theorem must be equal if the appropriate
inverses exist.
We now turn to showing that this parametrization is indeed stabilizing. So
choose a
Q^ 2 RH1 , where the above inverses exist, and dene
U^ = Y^ ; M^ Q^ V^ = X^ ; N^ Q^
U~ = Y~ ; Q^ M~ V~ = X~ ; Q^ N~ :
From (5.12) we see
  
V~ ;U~ M^ U^ = I ;
;N~ M~ N^ V^
which implies that U; ^ V^ and U; ~ V~ are right and left coprime factorizations
^
of K respectively. Also it clearly says
M^ U^ ;1 2 RH :
 

N^ V^ 1

Therefore invoking Lemma 5.11 we see that K^ is stabilizing.

Finally, we show that every stabilizing controller is given by the parametriza-
tion in (5.11). Fix a controller K^ and let
K^ =: U^ V^ ;1
be a right coprime factorization. Dene
^ := M~ V^ ; N~ U^ ;
and observe
    
X~ ;Y~ M^ U^ = I X~ U^ ; Y~ V^
;N~ M~ N^ V^ 0 ^
from the doubly coprime factorization of G^ 22 . Now both the matrix func-
tions on the left hand side have inverses in RH1 : the rst by the doubly
coprime factorization; the second by Lemma 5.11. Immediately we see that
^ ;1 2 RH1 :
Set Q^ := ;(X~ U^ ; Y~ V^ )^ ;1 , which is in RH1 , and left-multiply the above
matrix equation by
 
M^ Y^
N^ X^
 5.3. Parametrization of stabilizing controllers 183
to get
      
M^ U^ = M^ Y^ I ;Q^ ^ = M^ (Y^ ; M^ Q^ )^ :
N^ V^ N^ X^ 0 ^ N^ (X^ ; N^ Q^ )^
Just looking at the second block-column we see that
(X^ ; N^ Q^ );1 = ^ V^ ;1
exists and is proper, and also that
K^ = (Y^ ; M^ Q^ )(X^ ; N^ Q^ );1 :

In the above statement, we have described all stabilizing controllers for
Figure 5.2 in terms of parameter Q^ (s). The only constraints on this Q^ are
that it is RH1 , and that the inverse
(X^ ; N^ Q^ );1
exists and is proper. This latter condition can be expressed as the require-
ment that X^ (1) ; N^ (1)Q^ (1) should be invertible, and ensures our family
of controllers remains proper.

5.3.3 Closed-loop maps for the general system

In this section we summarize and assemble the results we have developed,
by returning to stabilization of the general setup of Figure 5.1. First recall
our partition of the system G:
2 3
A B1 B2 
^ ^ 
G^ = 4 C1 D11 D12 5 = G^ 11 G^ 12 :
C2 D21 D22 G21 G22
Next take a doubly coprime factorization of G^ 22
  
X~ ;Y~ M^ Y^ = I ;
;N~ M~ N^ X^
where G^ 22 = N^ M^ ;1 = M~ ;1 N~ . With these denitions we have the
following general result.
Theorem 5.13. Suppose (C2 ; A; B2) is a detectable and stabilizable real-
ization. Then all controllers that internally stabilize the system of Figure
5.1 are given by
K^ = (Y^ ; M^ Q^ )(X^ ; N^ Q^ );1 = (X~ ; Q^ N~ );1 (Y~ ; Q^ M~ ) ;
where Q^ (s) 2 RH1 and X^ (1) ; N^ (1)Q^ (1) is invertible.
Proof . This follows immediately from Corollary 5.9 and Theorem 5.12.

184 5. Stabilizing Controllers
Having parametrized all stabilizing controllers in terms of the RH1 pa-
rameter Q^ , we now turn to how this parametrizes the closed-loop map from
w to z .
Dene the three transfer functions
T^1 = G^ 11 + G^ 12 Y^ M~ G^ 21 ;
T^2 = G^ 12 M;
^
T^3 = M~ G^ 21 :
With these denitions in hand we have
Theorem 5.14.
(a) T^k 2 RH1 , for 1  k  3;
(b) With K as in Theorem 5.13, the transfer function from w to z in
Figure 5.1 is given by
T^1 ; T^2 Q^ T^3 :
Proof . This follows by routine algebra, and is left as an exercise.

The expression in (b) above shows that the closed-loop transfer function
from w to z is an ane function of the control parameter Q^ . This is a
property that can be exploited in many problems, and in particular is very
important for optimization.
In this chapter we have dened stability for the general feedback cong-
uration of Figure 5.1. We found conditions under which stabilization was
possible, and gave a complete characterization of all stabilizing controllers.
The next topic of the course is optimal controller synthesis, where our
performance demands are more than just stability.

5.4 Exercises
 
   
1. Consider the region R = s 2 C : (ss+ ;s s) (ss+;ss ) < 0 in
the complex plane. Here s denotes conjugation, and  > 0.
a) Sketch the region R.
b) For A, X in C nn we consider the 2n  2n LMI
 
AX + XA (AX ; XA ) < 0
(XA ; AX ) AX + XA
that is obtained by formally replacing s by AX , s by XA in
the denition of R. Show that if there exists X > 0 satisfying
this LMI, then the eigenvalues of A are in the region R. It can
be shown that the converse is also true.
 5.4. Exercises 185
c) Now we have a system x_ = Ax + Bu, and we wish to design a
state feedback u = Fx such that the closed loop poles fall in R.
Derive an LMI problem to do this.
2. Complete the proof of Proposition 5.10 by showing that the given
transfer functions satisfy equation (5.8). To do this, rst write
realizations of order n for each of
   
X~ ;Y~ and M^ Y^ ;
;N~ M~ N^ X^
then compose them and eliminate unobservable/uncontrollable states.
Yes, it's a little tedious, but signicantly less than working with
matrices of rational functions.
3. In this problem we consider the system in the gure, where P is a
stable plant, i.e. P^ (s) 2 RH1 .
d
r- - K -? - P z
6;

a) Show that all stabilizing controllers are given by K^ = Q^ (I ;

P^ Q^ );1 , with Q^ (s) 2 RH1 and such that the inverse is proper.
b) Now consider the case of single input-output P and K .
Parametrize all controllers that stabilize the plant and give per-
fect asymptotic tracking of a step applied in r, for d = 0. Is this
always possible?
c) Repeat part b) for asymptotically rejecting a step applied in d.
4. Prove Theorem 5.16.
5. In the standard feedback picture of the gure we assume that H ,

are both in RH1 .
v2
v1 d1

d2 H
a) Show that the system is input-output stable (i.e. the map from
d to v is in RH1 ) if and only if I ;
H is invertible in RH1 .
b) If k
k1 kH k1 < 1, show the system is input-output stable.
186 5. Stabilizing Controllers
6. Let P^ (s) be a proper rational matrix.
A right

coprime factorization
^
N ( s )
P^ = N^ M^ ;1 is called normalized if M^ (s) is inner (see Chapter 4).
It is shown in the exercises of the next chapter that every P^ admits
a normalized right coprime factorization (n.r.c.f.).
a) Given a n.r.c.f. P^ = N^ M^ ;1 , we now consider the perturbed
plant
P^
= (N^ +
^ N )(M^ +
^ M );1 ;

^ ^
^ N
where
N ,
M are in RH1 and ^  :


M 1
Show that P^
corresponds to the system inside the dashed box
in the following Figure.
P

N ;
M
p1 p2

N q M ;1
d1 d2

Further show that given a controller K^ , the above conguration

can be redrawn into the equivalent gure below, where
   
p = pp1 ; d = dd1 :
 
H^ = M^ ;1 (I ; K^ P^ );1 K^ I ;
2 2
d
p
N q
H
;
M

b) Using Problem 5, show that if K stabilizes P and

  ;1
P^ ;  
^ ^
 < I (I ; K P ) K^ I
1
1
 5.4. Exercises 187
then the controller will stabilize the perturbed plant.
c) Based on these results, propose a method of model reduction
for unstable plants. The method should be well behaved from
the point of view of the stabilization question (explain why).
Describe the algorithm in terms of concrete steps which can be
calculated in state space.

Notes and references

The standard closed loop formulation used here originates in [28], and owes
much to the detailed presentation in [41]. These sources concentrate on
pure input-output stability of a form equivalent to the internal stability
considered here.
The solution to the general stabilization problem in terms of LMIs orig-
inates in [77], where multi-dimensional systems were treated; the present
form with a constraint on controller order, is a special case of the optimal
synthesis results in [42, 89], to be covered later in the book.
The transfer function parametrization of all stabilizing controllers was
introduced in [148] and [73] in terms of matrix polynomial fractions. The
parametrization given here in terms of RH1 is from [20]. This parametriza-
tion is not crucial in future chapters of this course. However it has played
an important part in the development of robust control theory, and contin-
ues to be the launching point for many synthesis approaches; for instance
in L1 optimal control [17].
Coprime factorizations over RH1 were introduced in [134], to obtain
a factorization theory suited to the study of stability. A more extensive
presentation of this approach is given in [136]. To deal with the issue of
computation, state space realizations for coprime factorizations were de-
veloped in [72], and the state space doubly coprime factorization formulae
used here are from [86]. Coprime factorizations of this sort have numerous
applications, and are used in many branches of robust control theory; see
for example [44] and [48].
 This is page 188
Printer: Opaque this

6
H2 Optimal Control

In this chapter we begin our study of optimal synthesis and in particular

will derive controllers that optimize the H2 performance criterion.
The performance criterion of the chapter is dened by the space RH2
of matrix valued transfer functions. This space consists of rational, strictly
proper transfer functions, which have all their poles in C ; . Thus it is the
natural extension to matrix functions of the space RH2 of vector valued
functions introduced in Chapter 3.
The inner product on this space is dened as
Z 1
^ ^
hF ; Gi2 = 2 1 TrfF^  (j!)G^ (j!)gd!
;1
for two elements F^ and G^ in RH2 .
The synthesis problem we wish to address is concerned with the familiar
feedback arrangement shown in Figure 6.1, where both G and K are state
space systems. It is convenient to introduce the following notation for the
transfer function from w to z in the diagram.
^ K^ ) := G^ 11 + G^ 12 K^ (I ; G^ 22 K^ );1 G^ 21 = transfer function w 7! z :
S (G;
This is sometimes called the star-product between transfer functions G^ and
K^ , or equivalently their linear fractional transformation. S (G; ^ K^ ) is well-
dened and proper provided the interconnection is well-posed, which will
be a basic requirement.
The optimal H2 synthesis problem is
Given: nominal state space system G;
 6. H Optimal Control
2 189

z w
G
u

y
K

Figure 6.1. Synthesis arrangement

Find: a state space controller K , which internally stabilizes G and

^ K^ )k2 :
minimizes kS (G;
The major goal of the chapter is to construct, or synthesize, the solution
K to this optimization problem. A motivation for this optimization will be
given in the next section.
Let us rst consider the pure analysis question of computing the H2
norm of an element P^ of RH2 . Take a state space realization
 
P^ = A B
C 0 ;
where A is Hurwitz. Then we have
Z 1
kP^ k22 = 21 TrfP^  (j!)P^ (j!)gd!
;1
Z 1 
= TrfB  eA t C  CeAt B gdt ;
0
where the last equality follows by the Plancherel Theorem. Using linearity
of the trace gives us
Z 1
kP^ k22 = TrfB  eAt C  CeAt dtB g = TrB  Yo B
0
with Yo denoting the observability gramian. A similar argument shows that
kP^ k22 = TrCXc C  where Xc is the controllability gramian of this realiza-
tion. At the end of the chapter we will see that LMIs can also be used for
this calculation.
190 6. H Optimal Control
2

6.1 Motivation for H2 control

In this section we brie y discuss the motivation, from a control system
perspective, for the design of a feedback control from the point of view of
optimal H2 performance.
First, a general comment on the desire to minimize a norm of the map
from w to z ; this comment also applies to H1 control, to be considered in
the next chapter. Clearly, for this kind of objective to be meaningful the
arrangement of Figure 6.1 must be such that z represents variables that
must be \kept small". For example, in a tracking problem the conguration
should be set up so that z contains the tracking error, rather than the
tracking output. Also z usually contains some penalty on the control eort
u, to make the problem meaningful. Although we will not discuss these
modeling issues here, this philosophy underlies the objective of \making
maps small" which we are adopting.
Generally speaking, the H2 norm is an appropriate measure of perfor-
mance whenever there is information about the spectral content of the
driving variable w. The most common example of this is when w is station-
ary noise, which we can model for instance as a stationary random process.
Frequently when given this interpretation H2 optimal control is referred to
as linear quadratic gaussian (LQG) control. This motivation is now brie y
described.
For a stationary stochastic process, the autocorrelation matrix is dened
as
Rw ( ) = E w(t +  )w(t)
where E denotes expected value, and the spectral density Sw (j!) is the
Fourier transform of Rw ( ). While the specic signal trajectory w(t) is a
priori unknown, it is common that one has a good model of the spectral
density. The signal variance or power is related to it by
Z 1
Ejw(t)j2 = TrRw (0) = 21 TrSw (j!)d!
;1
If z and w are related by z = Pw for a stable, linear time invariant system
P , it can be shown that
Sz (j!) = P^ (j!)Sw (j!)P^ (j!)
and therefore
Z 1
Ejz (t)j2 = 21 TrP^ (j!)Sw (j!)P^ (j!) d!
;1
Z 1
= 21 TrP^ (j!) P^ (j!)Sw (j!)d!
;1
The right hand side of the previous equation looks exactly like kP k22, except
for the weighting matrix Sw (j!). In fact the identity is exact for the case of
 6.1. Motivation for H control
2 191
white noise1 Sw (j!) = I and therefore the H2 norm can be interpreted as
the output variance for white noise applied at the input. More generally, by
writing Sw (j!) = W^  (j!)W^ (j!) and incorporating the \weight" W (j!)
into the setup of Figure 6.1, the H2 norm can be used to study the response
to processes of known spectral density.
A second example where we have information about the input spectrum
is when the input signal w(t) is known a priori. This arises, for example in
the case of tracking of xed reference signals. Let us consider the special
case where w(t) is scalar valued, and in particular
w(t) = (t) ;
the Dirac delta function. Then directly from the denition we have that
Z 1
kz k22 = z (t)z (t)dt
;1Z
1
= 21 z^ (j!)^z(j!)d!
Z 1
;1
1
= 2 P^  (j!)P^ (j!)d!
;1
= kP^ k22 :
Thus we see that when the system input is scalar valued, the H2 norm
of P^ gives us a direct measure of the system output energy. If instead of
the Dirac delta function we wished to study the response to a dierent
(transient) input, this signal can be generated as the impulse response of
a known lter W^ (s), which once again can be incorporated into our H2
optimization setup.
To extend this interpretation to vector-valued inputs: suppose that the
input to P has m spatial channels; that is w(t) takes values in Rm . Then
dene
wk (t) = (t)
ek ;
where ek is the kth standard basis vector in Rm . If zk is the response
zk = Pwk , then it follows that
m
X
kzk k22 = kP^ k22
k=1
Details are left as an exercise. Thus kP^ k22 gives us the sum of the output
energies, which can also be thought of as a measure of the average output
energy over the impulsive inputs.

1
Strictly speaking, this idealization escapes the theory of stationary random
processes; a rigorous treatment is not in the scope of our course.
192 6. H Optimal Control
2

Finally, we present a third motivation for the H2 norm which is of a

dierent nature; here the input is an arbitrary L2 signal and has no known
characteristics, but we wish to measure the output signal in the L1 norm
(over time). That is, our criterion is the induced norm from L2 to L1 ,
kP kL2!L1 = sup kz k1
kwk2 =1
This interpretation applies to the case of a scalar output z (t). We can write
 Z 1 
 1 j!t 
jz (t)j =  2
 z^(j!)e d!
Z 1
;1
 21 jP^ (j!)w^(j!)jd!
;1
Z 1
 21 jP^ (j!)jjw^(j!)jd!
;1
 kP^ k2kw^k2
Hence kP kL2!L1  kP^ k2 , and equality is not dicult to show.
Here we have discussed three dierent views of the H2 norm of a system
transfer function. There is also a fourth which we leave until Chapter 10.
Having motivated this norm we proceed to develop an important tool for
tackling the optimal synthesis problem.

6.2 Riccati equation and Hamiltonian matrix

In this section we develop some matrix theory which will be required during
our development of the H2 optimal synthesis procedure.
A Hamiltonian matrix is a real matrix of the form
 
H = ;AQ ;RA
where Q and R are n  n symmetric matrices, and A is also in Rnn . The
Riccati equation associated with H is
A X + XA + XRX + Q = 0 :
The connection between the two is seen in the following similarity
transformation of H which uses a solution to the Riccati equation
     
I 0 A R I 0 = A + RX R
;X I ;Q ;A X I 0 ;A ; XR
Since the transformed representation of H is a zero in its lower left cor-
ner, we see that X provides us with the natural
  coordinate system for an
invariant subspace of H . Specically Im XI is an invariant subspace of
 6.2. Riccati equation and Hamiltonian matrix 193
H . This observation, and the fact that A + RX is a representation for H
on that subspace is the key to this section. Hamiltonian matrices have the
following elementary property.
Proposition 6.1. H and ;H are similar matrices (i.e. there exists T such
that H = ;T ;1HT ).
Proof . Let
 
0
J= I 0 ; ; I

and observe that J ;1 HJ = ;H . So H is similar to (;H ) , the transpose

of the real matrix ;H . Now any matrix is similar to its transpose, therefore
H is similar of ;H . 
This result implies that if  is an eigenvalue of H , ; is an eigenvalue of
the same multiplicity; in particular, their corresponding invariant subspaces
are of the same dimension. Note that since H is real its eigenvalues are also
mirrored across the real axis in C ; so eigenvalues will be mirrored over both
axis in the complex plane.
If H has no purely imaginary eigenvalues, it must have an equal number
in the left and right half planes of C . Therefore the invariant subspaces
corresponding to stable and unstable eigenvalues have both dimension n in
this case. This leads to the following denition:
Denition 6.2. Suppose that H is a Hamiltonian matrix. H is said to be
in the domain of the Riccati operator if there exist square, n  n matrices
H; and X such that
   
I I
X H; = H X ; (6.1)
and H; is Hurwitz. In this case we dene the function Ric(H ) = X .
The identity (6.1) implies that
 
Im XI
is an invariant subspace of H , of dimension n. Since H; is Hurwitz this
subspace corresponds to all eigenvalues of H in C ; . From the previous
discussion this implies that H can have no purely imaginary eigenvalues;
however (6.1) is slightly stronger in that it sets a particular structure on the
subspace (the rst n coordinates are independent). Also it should be clear
that X is necessarily unique, and thus the notation Ric(H ) is unambiguous.
This latter fact is shown more explicitly in Proposition 6.4.
For Hamiltonian matrices of the above type we prove the following crucial
properties.
194 6. H Optimal Control
2

Theorem 6.3. Suppose H is in the domain of the Riccati operator and

X = Ric(H ). Then
(a) X is a symmetric matrix;
(b) X satises the Riccati equation;
(c) A + RX is Hurwitz.
Proof . We prove each of the properties in turn, starting with (a): again
let
 
J = I0 ;0I ;
and left multiply property (6.1) by J to get
   
I J I H = I  JH I :
   

X X ; X X
Now JH is symmetric and therefore the left hand side above must be
symmetric meaning
 
I J I H = ;H  I  J I :
     

X X ; ; X X
From this we have that
(X  ; X )H; + H; (X  ; X ) = 0 :
Recall that H; is Hurwitz, and therefore this latter Lyapunov equation has
a unique solution; that is X  ; X = 0.  
We now show property (b): left multiply (6.1) by ;X I to arrive at
   
;X I H XI = [;X I ] XI H; = 0 :
 

Now expand the left hand side, using the denition of H and get the Riccati
equation.
Finally we demonstrate that property (c)  holds.
 Again we use the
relationship (6.1); this time left multiply by I 0 and see that
   
I 0 H XI = I 0 XI H; = H; :
   

The left hand side is equal to A + RX .


The last part of the preceding theorem, and the remark made ear-
lier about the Riccati operator being well dened, immediately imply the
following:
Proposition 6.4. There is at most one solution to the Riccati equation
such that A + RX is Hurwitz.
 6.2. Riccati equation and Hamiltonian matrix 195
To complete our work on Hamiltonian matrices and the Riccati equation
we consider a result which ties it directly to systems theory.
Theorem 6.5. Suppose that H is a Hamiltonian matrix and
(a) H has no purely imaginary axis eigenvalue;
(b) R is either positive or negative semidenite;
(c) (A; R) is a stabilizable matrix pair.
Then H is in the domain of the Riccati operator.
Proof . Since H has no imaginary axis eigenvalue we see that by Propo-
sition 6.1 it has n eigenvalues in each of the open half planes, where
A 2 Rnn . Thus there exist matrices X1 ; X2 and a Hurwitz matrix H;
in Rnn such that
   
H X 1 X1
X2 = X2 H; (6.2)
It is sucient to show that X1 is nonsingular, then multiplication by X1;1
turns (6.2) to the form (6.1) with X = X2 X1;1. Equivalently, we must show
that ker X1 = 0. We accomplish this in two steps.
Step A: Show X1 X2 is symmetric.
Left multiply the last equation by
X1  J where J = 0 ;I
   

X2 I 0
to arrive at
X1  JH X1 = X1  J X1 H :
       

X2 X2 X2 X2 ;
As noted earlier JH is symmetric and so from the right hand side above
we get
(;X1X2 + X2X1 )H; = ;H; (;X1X2 + X2X1 ) :
Moving everything to one side of the equality we get a Lyapunov equation;
since H; is Hurwitz it has the unique solution of zero. That is
;X1X2 + X2 X1 = 0 :
Step B: Show X1 is nonsingular.
We will show that ker X1 = 0. First wedemonstrate
 that it is an invariant
subspace of H; . Left multiply (6.2) by I 0 to get this is
AX1 + RX2 = X1 H; : (6.3)
Now for any x 2 ker X1 we have
x X2 (AX1 + RX2)x = x X2X1 H; x = x X1X2 H; x = 0 ;
196 6. H Optimal Control
2

where we have used the result of Step A. We have apparently

x X2RX2 x = 0 ;
and since R is semi-denite that RX2 x = 0. Finally multiply (6.3) by x to
get
RX2 x = X1 H; x = 0 ;
which says that ker X1 is invariant to H; .
We use this latter fact to show that ker X1 = 0. We know that if this
kernel is nonzero, then there exists a nonzero element x 2 ker X1 such that
H; x = x ;
where the real part of  is negative. Thus it is enough to show that any x
satisfying this equation must be zero, which we now do.
Left multiply (6.2) by [0 I ] to get
;QX1 ; A X2 = X2 H; ;
which therefore implies
(A + I )X2 x = 0 :
Also from above RX2x = 0 and so
x X2 [A +  I R] = 0 :
Now x X2 = 0 since (A; R) is stabilizable, and so
 
X1 x = 0 :
X2
Since the left hand matrix has full rank we see that x is necessarily zero.

To end we have the following exercise.
Corollary 6.6. Suppose (C; A; B) is a stabilizable and detectable matrix
triple, and
  
H = ;CA C ;;BB
A :

Then H is in the domain of the Riccati operator.
We have developed all Riccati results that are required to pursue the
optimal H2 synthesis.

6.3 Synthesis
Our solution to the optimal synthesis problem will be strongly based on
the concept of an inner function, introduced in Chapter 3.5. In fact the
 6.3. Synthesis 197
connection between the H2 problem and Riccati equation techniques will
be based on a key Lemma about inner functions.
We recall that a transfer function U^ (s) 2 H1 is called inner when it
satises
U^  (j!)U^ (j!) = I for all ! 2 R:
Multiplication by an inner function denes an isometry on H2 . This was
indicated before; however the property also extends to the matrix space
RH2 which we are considering in the present chapter. To see this, suppose
U^ is inner and that G^ 1 and G^ 2 are in RH2 , then
Z 1
hU^ G^ 1 ; U^ G^ 2 i2 = Tr(U^ (j!)G^ 1 (j!)) U^ (j!)G^ 2 (j!)dw
;1
Z 1
= TrG^ 1 (j!)G^ 2 (j!)d!
;1
=hG^ 1 ; G^ 2 i2
Namely inner products are unchanged when such a function is inserted.
We now relate inner functions with Riccati equations.
Lemma 6.7. Consider the state-space matrices A; B; C; D, where (C; A; B)
is stabilizable and detectable and D [C D] = [0 I ]. For
  
H = ;CA C ;;BB A ;

dene X = Ric(H ) and F = ;B X . Then
 
U^ (s) = CA +
+ BF B
DF D (6.4)
is inner.
Proof . Since X is a stabilizing solution, we have that A;BBX = A+BF
is Hurwitz, therefore U^ (s) 2 RH1 . Now from the Riccati equation
A X + XA + C  C ; XBB  X = 0
and the hypothesis it is veried routinely that
(A + BF ) X + X (A + BF ) + (C + DF ) (C + DF ) = 0:
Therefore X is the observability Gramian corresponding to the realization
(6.4). Also, the Gramian X satises
(C + DF ) D + XB = F  + XB = 0:
Therefore we are in a position to apply Lemma 4.16, and conclude that
U^ (s) is inner. 
As a nal preliminary, we have the following matrix version of our earlier
vector valued result; this extension is a straightforward exercise.
198 6. H Optimal Control
2

Proposition 6.8. If G^1 and G^2 are matrix valued function in RH2 and
RH2? , respectively, then
hG^ 1 ; G^ 2 iL2 = 0 :
Observe that if Q^ 2 RH2 then the function Q^  2 RH2? .
We are now in a position to state the main result, which explicitly gives
the optimal solution to the H2 synthesis problem. We spend the remainder
of the section proving the result using a sequence of lemmas. The technique
used exploits the properties of the Riccati equation, and in particular its
stabilizing solution.
Theorem 6.9. Suppose G is a state space system with realization
2 3
A B1 B2
G^ (s) = 4 C1 0 D12 5
C2 D21 0
where
(a) (C1 ; A; B1 ) is a stabilizable and detectable;
(b) (C2 ; A; B2 ) is a stabilizable and detectable;
 C1 D12  = 0 I ;
(c) D12
   = 0 I .
(d) D21 B1 D21
Then the optimal stabilizing controller to the H2 synthesis problem is given
by
 
K^ 2 (s) = AF2 ;0L ;
with
L = ;Y C2 ; F = ;B2 X and A2 = A + B2 F + LC2
where X = Ric(H ); Y = Ric(J ) and the Hamiltonian matrices are
     
H = ;CA C ;;BA2B 2 ; J = ;BA B  ;C;2AC2 :
1 1 1 1
Furthermore the performance achieved by K2 is
kG^ c B1 k22 + kF G^ f k22
where
   
G^ c = A F I ^
CF 0 and Gf = I 0
AL BL

with AF = A + B2F; CF = C1 + D12 F; AL = A + LC2 and BL = B1 + LD21.

 6.3. Synthesis 199
There are a number of conditions imposed on the state space matrices of
G^ (s) which we now comment on. The rst two are that D11 and D22 both
be zero. That D11 = 0 is a reasonable assumption, which ensures that the
closed loop transfer is strictly proper provided K^ is strictly proper. That
D22 = 0 is done purely out of convenience and can be removed.
Conditions (a) and (b) guarantee that both H and J are in the domain
of the Riccati operator; (a) is a technical condition that can be relaxed at
the expense of a more complicated proof. Clearly however the condition in
(b) is necessary for a stabilizing controller to exist.
The nal two conditions are also made out of convenience. Condition
(c) can be interpreted as saying that the plant output C1 x and the weight
D12 u on the control eort are orthogonal, and all the control channels af-
fect the system performances directly. Condition (d) states that the system
disturbances and measurement noise are orthogonal; furthermore that the
noise channels aect all the system measurements. The relaxing of these
assumptions results in more complicated controller expressions and deriva-
tions, but the main method now pursued remains largely unchanged in such
a general scenario.
We can now embark on our quest to prove Theorem 6.9. The next result
is absolutely central to our technique of proof. Note that the hypothesis
are a subset of those in Theorem 6.9.
Lemma 6.10. Suppose G is a system with state space realization
2 3
A B1 B2
G^ (s) = 4 C1 0 D12 5
C2 D21 0
satisfying
(a) (C1 ; A; B1 ) is stabilizable and detectable;
(b) (C2 ; A; B2 ) is stabilizable and detectable;
(c) D12 C1 D12  = 0 I .
If a controller K internally stabilizes G, then the closed loop performance
satises
^ K^ )k22 = kG^ c B1 k22 + kS (G^ tmp ; K^ )k22 ;
kS (G;
where
2 3
  A B1 B2
G^ c = ACF I0 and G^ tmp = 4 ;F 0 I 5
F C2 D21 0
with AF = A + B2 F , CF = C1 + D12 F , F = ;B2 X ,
 
H = ;CA C ;;BA2 B 2 and X = Ric(H ) :
1 1
200 6. H Optimal Control
2

Proof . First we establish that all the objects in the theorem statement are
dened. Use hypothesis (a) and (b), and invoke Corollary 6.6 to see that
H is in the domain of the Riccati operator. From the same result we see
that AF is Hurwitz, and so G^ c 2 RH2 . Also notice that K must stabilize
Gtmp since the closed loop A-matrix is the same as that with G; hence
S (G^ tmp ; K^ ) 2 RH2 .
The closed loop state equations are
x_ (t) = Ax(t) + B1 w(t) + B2 u(t)
z (t) = C1 x(t) + D12 u(t)
y(t) = C2 x(t) + D21 w(t) ;
with u = Ky. Now make the substitution v(t) = u(t) ; Fx(t) to get
x_ (t) = (A + B2 F )x(t) + B1 w(t) + B2 v(t)
z (t) = (C1 + D12 F )x(t) + D12 v(t) :
Recalling the denition of G^ c (s) and converting to transfer functions we
get
z^(s) = G^ c (s)B1 w^(s) + U^ (s)^v (s) ;
where U^ has realization (AF ; B2 ; CF ; D12 ). It is routine to show that the
transfer function from w to v is S (G^ tmp ; K^ ) and so
S (G;^ K^ ) = G^ c B1 + US ^ (G^ tmp ; K ) :
Next let us take the norm of S (G; ^ K^ ) making use of this new expression
^ K^ )k22 = kG^ cB1 k22 + kUS
kS (G; ^ (G^ tmp ; K^ )k22 + 2RefhUS ^ (G^ tmp ; K ); G^ c B^1 ig :
The key to the proof is that:
(i) U^ is inner; this was proved in Lemma 6.7;
(ii) U^  G^ c B1 is in RH2? ; this fact follows by similar state-space
manipulations and is left as an exercise.
The conclusion now follows since kS (G^ tmp ; K^ )k2 = kUS ^ (G^ tmp ; K^ )k2 ,
and
^ (G^ tmp ; K^ ); G^ c B1 i2 = hS (G^ tmp ; K^ ); U^  G^ c B1 )i2 = 0 ;
hUS
where the orthogonality is clear because S (G^ tmp ; K^ ) 2 RH2 as noted above.

The major point of this lemma is that it puts a lower bound on the achiev-
able performance, which is kG^ c B1 k22 . Now K stabilizes G of the corollary
if and only if Gtmp is stabilized, and thus minimizing the closed loop
performance is equivalent to minimizing kS (G^ tmp ; K^ )k2 by choosing K .
 6.3. Synthesis 201
An additional remark is that in the special case of state feedback, i.e.
when C2 = I , D21 = 0, this second term can be made zero by the static
control law K = F , as follows from the direct substitution into S (G^ tmp ; K^ ).
Alternatively, revisiting the proof we note that u = Fx gives v = 0, and the
second term of the cost would not appear. Therefore Lemma 6.10 provides
a solution to the H2 problem in the state feedback case.
In the general case, the auxiliary variable v with its associated cost re-
ects the price paid by not having the state available for measurement.
This additional cost can now be optimized as well. Before addressing this,
we state a so-called duality result; its proof is an exercise involving only
the basic properties of matrix transpose and trace
Lemma 6.11. Suppose G has a realization (A; B; C; D); K has a realiza-
tion (AK ; BK ; CK ; DK ), and > 0. Then the following are equivalent.
(i) K internally stabilizes G and kS (G; ^ K^ )k2 = ;
(ii) K 0 internally stabilizes G0 and kS (G^ 0 ; K^ 0 )k2 = .
Here G0 has realization (A ; C  ; B  ; D ) and K 0 has realization (AK ; CK ; BK ; DK )
The previous three results combine to give a further breakdown of the
closed loop transfer function of G and K .
Lemma 6.12. Let Gtmp be as dened in Lemma 6.10. If K internally
stabilizes Gtmp then
kS (G^ tmp ; K^ )k22 = kF G^ f k22 + kS (E;
^ K^ )k22
where
2 3
A ;L B2
E^ (s) = 4 ;F 0 I 5
C2 I 0
Proof . Start by applying Lemma 6.11 with G0 = Gtmp and K 0 = K to
get
kS (G^ tmp ; K^ )k2 = kS (G^ 0tmp ; K^ 0 )k2 :
Now apply Lemma 6.10 with G = G0tmp and K = K 0 to get
kS (G^ 0tmp ; K^ 0 )k22 = kG^ 0f F  k22 + kS (E^ 0 ; K^ 0)k22
Finally applying Lemma 6.11 to the right hand side we get
kS (G^ 0tmp ; K^ 0 )k22 = kF G^ f k22 + kS (E;
^ K^ )k22 :

We can now completely prove Theorem 6.9.
202 6. H Optimal Control
2

Proof . Observe that K2 is an observer based controller of the standard

form; also by our assumption, Theorem 6.3 and Corollary 6.6 we see AF
and AL are Hurwitz. Therefore K2 internally stabilizes G.
Invoke Lemma 6.10 and Lemma 6.12 to get
^ K2 )k22 = kG^ c B1 k22 + kF G^ f k22 + kS (E;
kS (G; ^ K^ 2 )k22 :
Since the rst two terms on the right do not depend on K^ 2 at all, it is
sucient to show that kS (E; ^ K^ 2)k2 = 0. This is easily shown by writing
out the state space equations for the closed loop system S (E; ^ K^ 2 ).

At this point we have completely proved our optimal synthesis result.
Notice from the last proof that regardless of the controller K we have
^ K^ )k22 = kG^ c B1 k22 + kF G^ f k22 + kS (E;
kS (G; ^ K^ )k22 ;
providing that K internally stabilizes G, or equivalently E . An alternative
approach to the H2 synthesis problem we just solved involves using the
controller parameterization of Chapter 5 in terms of the RH1 parameter
Q^ . Then our synthesis problem becomes
minimize kT^1 ; T^2Q^ T^3 k2 with Q^ 2 RH1 ,
a so-called model matching problem. If we use the stabilizing feedback
gains F and L of this section, in the state space coprime factorization of
Chapter 5, we can show that the optimal solution to this model matching
problem occurs when Q^ = 0. We leave this as a possible exercise.

6.4 State feedback H2 synthesis via LMIs

We have just described a method of constructing the optimal solution to
the H2 synthesis problem, which uses the Riccati equation as a central
tool. In this section we investigate obtaining a solution via LMI machinery
instead, and focus our attention on the simplied case of state feedback.
Our goal is to catch a quick glimpse of how LMI techniques can be used to
solve this type of synthesis problem.
Our rst result characterizes the RH2 norm and internal stability in
terms of linear matrix inequalities.
Proposition 6.13. Suppose P is a state space system with realization
(A; B; C ). Then
A is Hurwitz and kP^ k2 < 1;
if and only if, there exists a symmetric matrix X > 0 such that TrCXC  <
1 and
AX + XA + BB  < 0:
 6.4. State feedback H synthesis via LMIs
2 203
Proof . To show the \only if" part: by hypothesis the system controllability
gramian Xc satises
TrCXc C  < 1
and A is Hurwitz. Now the expression
Z 1  
eAt B I BI eAt dt;
 
X=
0
is continuous in , and equal to Xc when  = 0. Therefore, for some  > 0, we
have TrCXC  < 1. This matrix X clearly satises the Lyapunov equation
AX + XA + BB  + 2 I = 0;
which means that the inequality of the claim is met.
The \if" direction follows the same idea. By assumption there exists X
which is a positive denite solution to
AX + XA + BB  < 0;
and satises TrCXC  < 1. Thus A is necessarily Hurwitz, and X is a
generalized controllability gramian; so we know from Chapter 3.5 that X 
Xc, where Xc is the controllability gramian. This implies that TrCXC  
TrCXc C  and thus kP^ k2 < 1. 
We are now ready to formulate the state feedback synthesis problem.
The plant G has the following form
2 3
A B1 B2
^
G(s) := C1 0 D12 5 ;
4
I 0 0
where A 2 R ; B1 2 R ; B2 2 Rnm ; C1 2 Rqn . The controller is
n n n p
taken to be simply a static feedback gain K^ (s) = DK . While more generally
we could use a dynamic controller, we have already seen in the previous
section that the static restriction entails no loss of performance in the state
feedback case. Our rst synthesis result is based on the same change of
matrix variables which was encountered in the study of stabilization.
Proposition 6.14. There exists a feedback gain K^ (s) = DK that
internally stabilizes G and satises
^ K^ )k2 < 1;
kS (G;
if and only if, there exist a rectangular matrix Y 2 Rmn such that
DK = Y X ;1;
where X > 0 and satises the inequalities
   

A B2 X

+

X Y   A + B1 B1 < 0; (6.5)
Y B2
Tr(C1 X + D12 Y )X (C1 X + D12 Y ) < 1:
; 1 (6.6)
204 6. H Optimal Control
2

Proof . Simply write out the state space equations corresponding to the
interconnection shown in Figure 6.1.
x_ (t) = (A + B2 DK )x(t) + B1 w(t)
z (t) = (C1 + D12 DK )x(t):
Now invoking Proposition 6.13 we see that our problem can be solved if
and only if there exist DK and X > 0 satisfying
(A + B2 DK )X + X (A + B2 DK ) + B1 B1  < 0
Tr(C1 + D12 DK )X (C1 + D12 DK ) < 1
Now introducing the change of variables DK X = Y these conditions are
equivalent to (6.5) and (6.6). 
While we have transformed the H2 synthesis problem to a set of matrix
inequalities, this is not yet an LMI problem since (6.6) is not convex in X ,
Y . Now we recall our discussion from Chapter 1, where it was mentioned
that often problems which do not appear to be convex can be transformed
into LMI problems by a Schur complement operation. This is indeed the
case here, although it will also be necessary to introduce an additional vari-
able, a so-called slack variable. We explain these techniques while obtaining
our main result.
Theorem 6.15. There exists a feedback gain K^ (s) = DK that internally
stabilizes G and satises
^ K^ )k2 < 1;
kS (G;
if and only if, there exist square matrices X 2 Rnn Z 2 Rqq and a
rectangular matrix Y 2 Rmn such that
DK = Y X ;1;
and the inequalities
   
A B2 X
 
+

X Y   A + B1 B1 < 0; (6.7)
Y B2
 
X (C1 X + D12 Y ) > 0; (6.8)
(C1 X + D12 Y ) Z
Tr(Z ) < 1 (6.9)
are satised.
Proof . It suces to show that conditions (6.8) and (6.9) are equivalent to
(6.6) and X > 0.
Suppose (6.6) holds; since the trace is monotonic under matrix in-
equalities, then we can always nd a matrix Z satisfying Tr(Z ) < 1
and
(C1 X + D12 Y )X ;1 (C1 X + D12 Y ) < Z
 6.5. Exercises 205
indeed it suces to perturb the matrix on the left by a small positive
matrix. Now from the Schur complement formula of Theorem 1.10, the
above inequality and X > 0 are equivalent to (6.8).
The previous steps can be reversed to obtain the converse implication.

Thus we have reduced the static state feedback H2 synthesis problem to
a set of convex conditions in the three variables X , Y and Z .
The above derivations exhibit a common feature of tackling problems via
LMIs: there is always an element of \art" involved in nding the appro-
priate transformation or change of variables that would render a problem
convex, and success is never guaranteed. In the next chapter we will present
additional tools to aid us in this process. The references contain a more ex-
tensive set of such tools, in particular how to tackle the general (dynamic
output feedback) H2 synthesis problem via LMI techniques.
We are now ready to turn our attention to a new performance criterion;
it will be studied purely from an LMI point of view.

6.5 Exercises
1. Verify that for a multi-input sytsem P , kP^ k22 is the sum of out-
put energies corresponding to impulses (t)ek applied in each input
channel.
2. Prove Corollary 6.6.
3. Normalized coprime factorizations. This exercise complements Ex-
ercise 6, Chapter 5. We recall that a right coprime factorization
 
P^ = N^ M^ ;1 of a proper P^ (s) is called normalized when M N^ (s)
^ (s)
is inner.  
(a) Consider P^ = CA B0 , and assume the realization is minimal.
Let F = ;B  X , where
  
X = Ric ;CA C ;;BB
A :

Show that
2 3
N^ (s) = 4 A +CBF B0
 
5
M^ (s) F I
denes a normalized right coprime factorization for P^ .
(b) Extend the result for any P^ (not necessarily strictly proper).
206 6. H Optimal Control
2

4. In the proof of Lemma 6.10, verify the claim (ii) that U^  G^ c B1 is in

RH2? .
5. Prove the dualization result of Lemma 6.11.
6. Let Xs be denote the stabilizing solution of the Riccati equation
A X + XA + C  C ; XBB X = 0; (6.10)
where we assume that (A; B; C ) is controllable and observable.
(a) For any other symmetric solution X to (6.10) show that
(A ; BB  X )(Xs ; X ) + (Xs ; X )(A ; BB  X ) +
(Xs ; X )BB  (Xs ; X ) = 0
Use this to prove X  Xs , i.e. Xs is the maximizing solution to
(6.10).
(b) Now introduce the LMI
  
A X + XA + C  C XB  0:
BX I
Show that if X satises the LMI, then also X  Xs .
(c) Show that there exists an anti-stabilizing solution Xa to the Ric-
cati equation (6.10), i.e. a solution such that all the eigenvalues
of the matrix A ; BB  Xa are on the right half plane Re(s) > 0.
(d) Prove that Xa  X  Xs for any solution to (6.10), or the LMI
in (b).
(e) Show that the inequality Xa < Xs is strict, and that these
are the only two solutions to (6.10) with this property. Hint:
Relate Xs ; Xa to the controllability gramian of the pair (A ;
BB  Xs ; B ).
7. Here we investigate some uniqueness properties of H2 optimal control.
(a) Show by counter example that the optimal controller con-
structed in this chapter need not be the only controller that
achieves optimum performance;
(b) As already noted, the H2 optimal control synthesis can be posed
as the model matching problem
minimize kT^1 ; T^2 Q^ T^3 k2; with Q^ 2 RH1 .
This is simply done by invoking the Youla parametrization. Show
that the optimal closed loop transfer function, if one exists, is
unique. Hint: recall that RH2 is an inner product space, and
realize that the above model matching problem can be viewed
as nding the nearest point in a subspace of RH2 to T^1.
 6.5. Exercises 207
Notes and references
Optimal H2 control has a long history, and is more commonly known as
linear quadratic gaussian (LQG) control, because of its stochastic interpre-
tation outlined at the beginning of this chapter. It gets its roots from the
papers [63, 66] on Kalman lters. A more extensive historical treatment of
this subject, including detailed references, can be found in the books [74]
and [3]. The solution given here is based on the famous paper [24], where
in addition to the optimal controller derived here, all possible controllers
satisfying a particular H2 performance level are parametrized. The state
feedback solution given here in terms of LMIs rst appears in [36].
The main analytical tool used in this chapter is the Riccati equation,
which has a substantial research literature associated with it. We have just
provided a glimpse for our purposes; however see the recent monograph [75],
and also the classic survey article [141] for additional details and references
on these equations. For work on computational algorithms see for instance
[21].
LMI solutions to many control synthesis problems have appeared in
recent years. For a general synthesis method applicable in particular to
output-feedback H2 synthesis, see [120].
 This is page 208
Printer: Opaque this

7
H1 Synthesis

In this chapter we consider optimal synthesis with respect to the H1 norm

introduced in Chapter 3. Again we are concerned with the feedback ar-
rangement of Figure 6.1 where we have two state space systems G and K ,
each having their familiar role.
We will pursue the answer to the following question: does there exist a
state space controller K such that
 The closed loop system is internally stable;
 The closed loop performance satises
^ K^ )k1 < 1 :
kS (G;
Thus we only plan to consider the problem of making the closed loop con-
tractive in the sense of H1 . It is clear, however, that determining whether
^ K^ )k1 < , for some con-
there exists a stabilizing controller so that kS (G;
stant , can be achieved by rescaling the dependent problem to arrive
at the contractive version given above. Furthermore, by searching over ,
our approach will allow us to get as close to the minimal H1 norm as we
desire, but in contrast to our work on H2 optimal control, we will not seek
a controller that exactly optimizes the H1 norm.
There are many approaches for solving the H1 control problem. Proba-
bly the most celebrated solution is in terms of Riccati equations of a similar
style to the H2 solution of Chapter 6. Here we will present a solution based
entirely on linear matrix inequalities, which has the main advantage that it
can be obtained with relatively straightforward matrix tools, and without
any restrictions on the problem data. In fact Riccati equations and LMIs
 7.1. Two important matrix inequalities 209
are intimately related, an issue we will explain when proving the Kalman-
Yakubovich-Popov lemma concerning the analysis of the H1 norm of a
system, which will be key to the subsequent synthesis solution.
Before getting into the details of the problem, we make a few comments
about the motivation for this optimization.
As discussed in Chapter 3, the H1 norm is the L2 -induced norm of a
causal, stable, linear-time invariant system. More precisely, given a causal
linear time-invariant operator G : L2 (;1; 1) ! L2(;1; 1), the cor-
responding operator in the isomorphic space L^ 2 (j R) is a multiplication
operator MG^ for a certain G^ (s) 2 H1 , and
kGkL2!L2 = kMG^ kL^ 2!L^ 2 = kG^ k1
What is the motivation for minimizing such an induced norm? If we refer
back to the philosophy of \making error signals small" discussed in Chapter
6, we are minimizing the maximum \gain" of the system in the energy or
L2 sense. Equivalently, the excitation w is considered to be an arbitrary L2
signal and we wish to minimize its worst-case eect on the energy of z . This
may be an appropriate criterion if, as opposed to the situation of Chapter
6, we know little about the spectral characteristics of w. We will discuss,
in more detail, alternatives and tradeos for noise modeling in Chapter 9.
There is however a more important reason than noise rejection that mo-
tivates an induced norm criterion; as seen in Section 3.1.2, a contractive
operator Q has the property that the invertibility of I ; Q is ensured;
this so-called small-gain property will be key to ensuring stability of cer-
tain feedback systems, in particular when some of the components are not
precisely specied. This reason has made H1 control a central subject in
control theory; further discussion of this application is given later in the
course.

7.1 Two important matrix inequalities

The entire synthesis approach of the chapter revolves around the two tech-
nical results presented here. The rst of these is a result purely about
matrices; the second is an important systems theory result and is frequently
called the Kalman-Yacubovich-Popov lemma, or KYP lemma for short.
We begin by stating the following which the reader can prove as an
exercise.
Lemma 7.1. Suppose P and Q are matrices satisfying ker P = 0 and
ker Q = 0. Then for every matrix Y there exists a solution J to
P  J Q = Y:
The above lemma is used to prove the next one which is one of the two
major technical results of this section.
210 7. H1 Synthesis
Lemma 7.2. Suppose
(a) P; Q and H are matrices and that H is symmetric;
(b) The matrices WP and WQ are full rank matrices satisfying ImWP =
ker P and ImWQ = ker Q.
Then there exists a matrix J such that
H + P  J  Q + Q JP < 0 ; (7.1)
if and only if, the inequalities
WP HWP < 0 and WQ HWQ < 0
both hold.
Observe that when the kernels of P and Q are not both nonzero the result
does not apply as stated. However it is readily seen from Lemma 7.1, that
if both of the kernels are zero then there is always a solution J . If for
example only ker P = 0 then WQ HWQ < 0 is a necessary and sucient
condition for a solution to (7.1) to exist, as follows by simplied version of
the following proof.
Proof . We will show the equivalence of the conditions directly by
construction. To begin dene V1 to be a matrix such that
ImV1 = ker P \ ker Q;
and V2 and V3 such that
   
Im V1 V2 = ker P and Im V1 V3 = ker Q:
Without loss of generality we assume that V1 , V2 and V3 have full column
rank and dene V4 so that
 
V = V1 V2 V3 V4
is square and nonsingular. Therefore the LMI (7.1) above holds, if and only
if
V  HV + V  P  J  QV + V  Q JPV < 0 does. (7.2)
Now PV and QV are simply the matrices P and Q on the domain basis
dened by V ; therefore they have the form
   
PV = 0 0 P1 P2 and QV = 0 Q1 0 Q2 ;
we also dene the block components
2 3
H11 H12 H13 H14
 6H  H22 H23 H24 77 :
V HV =: 64H12 H23 H33 H34 5
13
H14 H24 H34 H44
 7.1. Two important matrix inequalities 211
Further dene the variable Y by
  
 
Y = YY11 YY12 = PP1 J  Q1 Q2 :
 
21 22 2
   
From their denitions ker P1 P2 = 0 and ker Q1 Q2 = 0, and so by
Lemma 7.1 we see that Y is freely assignable by choosing an appropriate
matrix J .
Writing out inequality (7.2) using the above denitions we get
2 3
H11 H12 H13 H14
6H  H 22 H 23 + Y  H  7
24 + Y21
6 12 11
H34 + Y12 5 < 0:
7
 
4H13 H23 + Y11 H33
H14 H24 + Y21 H34 + Y12 H44 + Y22 + Y22
Apply the Schur complement formula to the upper 3  3 block, and we see
the above holds, if and only if, the two following inequalities are met.
2 3
H11 H12 H13
H := 4H12 H22 H23 + Y11 5 < 0
H13 H23 + Y11 H33
H14 
2 3 2 3
H14
and H44 + Y22 + Y22 ; 4H24 + Y21 5 H ;1 4H24 + Y21 5 < 0
H34 + Y12 H34 + Y12
As already noted above Y is freely assignable and so we see that provided
the rst inequality can be achieved by choosing Y11 , the second can always
be met by appropriate choice of Y12 , Y21 and Y22 . That is the above two
inequalities can be achieved, if and only if, H < 0 holds for some Y11 . Now
applying a Schur complement on H with respect to H11 , we obtain
2 3
H11 0 0
4 0 H22 ; H12 H11;1H12 Y11 + X  5 < 0;
0 Y11 + X H33 ; H13 H11;1 H13
where X = H23  H ;1 H12 . Now since Y11 is freely assignable we see
 ; H13
11
readily that the last condition can be satised, if and only if, the diagonal
entries of the left hand matrix are all negative denite. Using the Schur
complement result twice these three conditions can be converted to the
equivalent conditions
   
H11 H12 < 0 and H11 H13 < 0:
H12 H22 H13 H33
By the choice of our basis we see that these hold, if and only if, WP HWP <
0 and WQ HWQ < 0 are both met. 
Having proved this matrix result we move on to our second result, the
KYP lemma.
212 7. H1 Synthesis
7.1.1 The KYP Lemma
There are many versions of this result, which establishes the equivalence
between a frequency domain inequality and a state-space condition in terms
of either a Riccati equation or an LMI. The version given below turns an
H1 norm condition into an LMI. Being able to do this is very helpful for
attaining our goal of controller synthesis, however it is equally important
simply as a nite dimensional analysis test for transfer functions.
Lemma 7.3. Suppose M^ (s) = C (Is ; A);1B + D. Then the following are
equivalent conditions.
(i) The matrix A is Hurwitz and
kM^ k1 < 1 ;
(ii) There exists a matrix X > 0 such that
   
C C D + A X + XA XB < 0 : (7.3)
D BX ;I
The condition in (ii) is clearly an LMI and gives us a very convenient way
to evaluate the H1 norm of a transfer function. In the proof below we see
proving that condition (ii) implies that (i) holds is reasonably straightfor-
ward, and involves showing the direct connection between the above LMI
and the state space equations that describe M . Proving the converse is con-
siderably harder; fortunately we will be able to exploit the Riccati equation
techniques which were introduced in Chapter 6. An alternative proof, which
employs only matrix arguments, will be given later in the course.
Proof . We begin by showing (ii) implies (i). The top left block in (7.3)
states that A X + XA + C  C < 0. Since X > 0 we see that A must be
Hurwitz.
It remains to show contractiveness which we do by employing a system-
theoretic argument based on the state equations for M . Using the strict
inequality (7.3) choose 1 >  > 0 such that
   
C C D + A X + XA XB < 0 (7.4)
D BX ;(1 ; )I
holds. Let w 2 L2[0; 1) and realize that in order to show that M is
contractive, it is sucient to show that kz k2  (1;)kwk2 , where z := Mw.
The state space equations relating w and z are
x_ (t) = Ax(t) + Bw(t); x(0) = 0;
z (t) = Cx(t) + Dw(t):
 
Now multiplying inequality (7.4) on the left by x (t) w (t) and on the
right by the adjoint we have
jz (t)j22 + x (t)X (Ax(t) + Bw(t)) + (Ax(t) + Bw(t)) Xx(t) ; (1 ; )jw(t)j22  0
 7.1. Two important matrix inequalities 213
By introducing the storage function V : Rn ! R, dened by V (x(t)) =
x (t)Xx(t), we arrive at the so-called dissipation inequality
V_ + jz (t)j22  (1 ; )jw(t)j22 :
Integrating on an interval [0; T ], recalling that x(0) = 0, gives
Z T Z T
x(T ) Xx(T ) + jz (t)j22 dt  (1 ; ) jw(t)j22 dt:
0 0
Now let T ! 1; since w 2 L2 and A is Hurwitz, then x(T ) converges to
zero and therefore we nd
kz k22  (1 ; )kwk22 ;
which completes this direction of the proof.
We now tackle the direction (i) implies (ii). To simplify the expressions
we will write the derivation in the special case D = 0, but an analogous
argument applies to the general case (see the exercises). Starting from
 
M^ (s) = CA B0 ;
and recalling the denition of M^  (s) from Chapter 6, we derive the state-
space representation
2 3
A 0 ;B
I ; M^  (s)M^ (s) = 4 ;C  C ;A 0 5 :
0 B I
It is easy to verify that
2 3
A BB  B
 ^ ;
[I ; M (s)M (s)] = ;C  C ;A 0 5 :
^ 1 4 (7.5)
0 B I
Since kM^ k1 < 1 by hypothesis, we conclude that [I ; M^  (s)M^ (s)];1 has
no poles on the imaginary axis. Furthermore we now show, using the PBH
test, that the realization (7.5) has no unobservable eigenvalues that are
purely imaginary. Suppose that
2 3
j!0 I ; A ;BB  x 
4 C C j!0 I + A 5 x1 = 0;
0 B 2
for some vectors x1 and x2 . Then we have the following chain of
implications:
B  x2 = 0 implies (j!0 I ; A)x1 = 0;
therefore x1 = 0 since A is Hurwitz;
this means (j!0 I + A )x2 = 0;
which implies x2 = 0 again because A is Hurwitz.
214 7. H1 Synthesis
We conclude that (7.5) has no unobservable eigenvalues on the imagi-
nary axis; an analogous argument shows the absence of uncontrollable
eigenvalues. This means that the matrix
  
H = ;CA C BB
;A 
has no purely imaginary eigenvalues. Referring to Theorem 5 in Chapter 6,
notice that BB   0 and (A; BB  ) is stabilizable since A is Hurwitz. Hence
H is in the domain of the Riccati operator, and we can dene X0 = Ric(H )
satisfying
A X0 + X0 A + C  C + X0 BB  X0 = 0 (7.6)
 
and A + BB X0 Hurwitz. Also note that (7.6) implies A X0 + X0 A  0,
therefore from our work on Lyapunov equations we see that
X0  0
since A is Hurwitz. To obtain the LMI characterization of (ii) we must
slightly strengthen the previous relationships. For this purpose dene X to
be the solution of the Lyapunov equation
(A + BB  X0 ) X + X (A + BB  X0 ) = ;I: (7.7)
 
Since (A + BB X0 ) is Hurwitz we have X > 0. Now let X = X0 + X > 0, 
which is positive denite for all  > 0. Using (7.6) and (7.7) we have
A X + XA + C  C + XBB  X = ;I + 2 XBB   X:
Choose  > 0 suciently small so that this equation is negative denite.
Hence we have found X > 0 satisfying the strict Riccati Inequality
A X + XA + C  C + XBB  X < 0:
Now applying a Schur complement operation, this inequality is equivalent
to
  
A X + XA + C  C XB < 0 ;
BX ;I
which is (7.3) for the special case D = 0. 
The preceding proof illustrates some of the deepest relationships of linear
systems theory. We have seen that frequency domain inequalities are asso-
ciated with dissipativity of storage functions in the time domain, and also
the connection between LMIs (linked to dissipativity) and Riccati equations
(which arise in quadratic optimization).
In fact this latter connection extends as well to problems of H1 syn-
thesis, where both Riccati equations and LMIs can be used to solve the
suboptimal control problem. In this course we will pursue the LMI solu-
tion. Surprisingly the two results of this section are all we require, together
with basic matrix algebra, to solve our control problem.
 7.2. Synthesis 215
7.2 Synthesis
We start with the state space realizations that describe the systems G and
K:
2 3
A B1 B2 
A B

^ 4 5 ^
G(s) = C1 D11 D12 ; K (s) = C D : K K
C2 D21 0 K K
Notice that we have assumed D22 = 0. Removing this assumption leads
to more complicated formulae, but the technique is identical. We make no
other assumptions about the state space systems. The state dimensions of
the nominal system and controller will be important: A 2 Rnn , AK 2
RnK nK .
Our rst step is to combine these two state space realizations into one
which describes the map from w to z . We obtain
2 3

A B
 A + B2 DK C2 B2 CK B1 + B2 DK D21
^ K^ ) = L L = 4
S (G; BK C2 AK BK D21 5:
CL DL C1 + D12 DK C2 D12 CK D11 + D12 DK D21
Now dene the matrix
 
J= A K BK
CK DK ;
which collects the representation for K into one matrix. We can parametrize
the closed-loop relation in terms of the controller realization as follows. First
make the following denitions.
   

A= 0 0 A 0 B = B01
 (7.8)
 
C = C0 I0
 
C = C1 0
2
 
0
B = I 02 B 
D12 = 0 D12


 
D = 0
21 D21
which are entirely in terms of the state space matrices for G. Then we have
AL = A + BJC BL = B + BJD21
 (7.9)
CL = C + D12 JC DL = D11 + D12 JD21
The crucial point here is that the parametrization of the closed loop state
space matrices is ane in the controller matrix J .
Now we are looking for a controller K such that the closed loop is con-
tractive and internally stable. The following form of the KYP lemma will
help us.
216 7. H1 Synthesis
Corollary 7.4. Suppose M^ L(s) = CL(Is ; AL);1BL + DL. Then the
following are equivalent conditions.
(a) The matrix AL is Hurwitz and kM^Lk1 < 1;
(b) There exists a symmetric positive denite matrix XL such that
2  3
AL XL + XLAL XL BL CL
4 BL XL ;I DL 5 < 0 :
CL DL ;I
This result is readily proved from Lemma 7.3 by applying the Schur comple-
ment formula. Notice that the matrix inequality in (b) is ane in XL and
J individually, but it is not jointly ane in both variables. The main task
now is to obtain a characterization where we do have a convex problem.
Now dene the matrices
 
PXL = B  XL 0 D12
 
Q = C D21 0
and further
2  3
A XL + XLA XLB C 
HXL = 4 B  XL ;I D11 5 :
C D11 ;I
It follows that the inequality in (b) above is exactly
HXL + Q J  PXL + PX L JQ < 0 :
Lemma 7.5. Given the above denitions there exists a controller synthesis
K if and only if there exists a symmetric matrix XL > 0 such that
WPXL HXL WPXL < 0 and WQ HXL WQ < 0 ;
where WPXL and WQ are as dened in Lemma 7.2.
Proof . From the discussion above we see that a controller K exists if and
only if there exists XL > 0 satisfying
HXL + Q J  PXL + PX L JQ < 0 :
Now invoke Lemma 7.2.

This lemma says that a controller exists if and only if the two matrix
inequalities can be satised. Each of the inequalities is given in terms of the
state space matrices of G and the variable XL . However we must realize
that since XL appears in both HXL and WPXL , that these are not LMI
conditions. Converting to an LMI formulation is our next goal, and will
 7.2. Synthesis 217
require a number of steps. Given a matrix XL > 0 dene the related matrix
2
 L;1 + XL;1 A B XL;1 C  3
AX
TXL = 4 B  ;I D11 5 ; (7.10)

CXL ; 1 D11 ;I
and the matrix
 
P = B  0 D12 (7.11)
which only depends on the state space realization of G. The next lemma
converts one of the two matrix inequalities of the lemma, involving HXL ,
to one in terms of TXL .
Lemma 7.6. Suppose XL > 0. Then
WPXL HXL WPXL < 0; if and only if, WP TXL WP < 0:
Proof . Start by observing that
PXL = PS ;
where
2 3
XL 0 0
S = 4 0 I 05 :
0 0 I
Therefore we have
ker PXL = S ;1 ker P :
Then using the denitions of WPXL and WP we can set
WPXL = S ;1 WP :
Finally we have that WPXL HXL WPXL < 0 if and only if
WP (S ;1 ) HXL S ;1 WP < 0
and it is routine to verify (S ;1 ) HXL S ;1 = TXL .

Combining the last two lemmas we see that there exists a controller of
state dimension nK if and only if there exists a symmetric matrix XL > 0
such that
WP TXL WP < 0 and WQ HXL WQ < 0 : (7.12)
The rst of these inequalities is an LMI in the matrix variable XL;1 , where
as the second is an LMI in terms of XL. However the system of both
inequalities is not an LMI. Our intent is to convert these seemingly non-
convex conditions into an LMI condition.
218 7. H1 Synthesis
Recall that XL is a real and symmetric (n + nK )  (n + nK ) matrix; here
n and nK are state dimensions of G and K . Let us now dene the matrices
X and Y which are n  n submatrices of XL and XL;1 , by
   
X X 2 ;
XL =: X  X and XL =: Y  Y :1 Y Y 2 (7.13)
2 3 2 3
We now show that the two inequality conditions listed in (7.12), only
constrain the submatrices X and Y .
Lemma 7.7. Suppose XL is a positive denite (n + nK )  (n + nK ) matrix
and X and Y are n  n matrices dened as in (7.13). Then
WP TXL WP < 0 and WQ HXL WQ < 0 ;
if and only if, the following two matrix inequalities are satised
(a)
2  3

NX 0  4A XB +XXA XB
 1 C1 N 0


;I D11 0 I < 0 ;
5 X
0 I 1
C 1 D
11 ;I
(b)
2   3
NY 0 4AY C+YY A Y;CI1 DB1 5 NY 0 < 0 ;
   

0 I 1 11 0 I
B1  ;I
D11
where NX and NY are full-rank matrices whose images satisfy
 
ImNX = ker C2 D21
  :
ImNY = ker B2 D12
Proof . The proof amounts to writing out the denitions and removing
redundant constraints. Let us show that WP TXL WP < 0 is equivalent to
the LMI in (b).
From the denitions of TXL in (7.10), and A, B and C in (7.8) we get
2 3
AY + Y A AY2 B1 Y C1
  0 Y2 C1 77
TXL = 64 Y2BA 0
6
1 0 ;I D 5 11
C1 Y C1 Y2 D11 ;I
Also recalling the denition of P in (7.11), and substituting for B and D12
from (7.8) yields
 
P = B0 I0 00 D0 :
2 12
 7.2. Synthesis 219
Thus the kernel of P is the image of
2 3
V1 0
WP = 64 00 I075
6 7

V2 0
where
 
V1 = N
V2 Y
 
spans the kernel of B2 D12  as dened above. Notice that the second
block row of WP is exactly zero, and therefore the second block-row and
block-column of TXL , as explained above, do not enter into the constraint
WP TXL WP < 0. Namely this inequality is
V1 0  AY + Y A B1 Y C1 V1 0
2 3 2 32 3
4 0 I5 4 B1 ;I D11 5 4 0 I 5 < 0 :
V2 0 C1 Y D11 ;I V2 0
By applying the permutation
2 3 2 3
V1 0 I 0 0 N 0
4 0 I 5 = 40 0 I 5 Y
V 0 0 I 0 0 I
2
we arrive at (b).
Using a nearly identical argument, we can readily show that WQ HXL WQ <
0 is equivalent to LMI (a) in the theorem statement.

What we have shown is that a controller synthesis exists if and only if
there exists an (n + nK )  (n + nK ) matrix XL that satises conditions (a)
and (b) of the last lemma. These latter two conditions only involve X and
Y , which are submatrices of XL and XL;1 respectively. Our next result tell
us under what conditions, given arbitrary matrices X and Y , it is possible
to nd a positive denite matrix XL that satises (7.13).
Lemma 7.8. Suppose X and Y are symmetric, positive denite matrices
in Rnn ; and nK is a positive integer. Then there exist matrices X2 ; Y2 2
RnnK and symmetric matrices X3 ; Y3 2 RnK nK , satisfying

X X2 > 0 and X X2 ;1 = Y Y2
     

X2 X3 X2 X3 Y2 Y3

if and only if
   
X I  0 and rank X I  n + n : (7.14)
I Y I Y K
220 7. H1 Synthesis
Proof . First we prove that the rst two conditions imply the second two.
From
  
X X2 Y Y2
X2 X3 Y2 Y3 = I (7.15)
it is routine to verify that
     
I 0 X X 2
0  Y Y X X 0 Y  = I Y : I Y X I
2 2 3 2
Also the Schur complement relationship
     
X I = I Y ;1 X ; Y ;1 0 I 0 (7.16)
I Y 0 I 0 Y Y ;1 I
implies that
 
rank XI YI = n + rank(X ; Y ;1 ) = n + rank(XY ; I )  n + nk ;
where the last inequality follows from (7.15): I ; XY = X2 Y2 and X2 2
RnnK .
To prove \if" we start with the assumption that (7.14) holds; therefore
(7.16) gives
X ; Y ;1  0 and rank (X ; Y ;1 )  nK :
These conditions ensure that there exists a matrix X2 2 RnnK so that
X ; Y ;1 = X2 X2  0 :
From this and the Schur complement argument we see that
 
X X2
X2 I > 0 :
Also
X X2 ;1 = Y
   
;Y X2
X2 I ;X2Y X2 Y X2 + I
and so we set X3 = I .

The lemma states that a matrix XL in R(n+nK )(n+nK ) , satisfying (7.13),
can be constructed from X and Y exactly when the LMI and rank condi-
tions in (7.14) are satised. The rank condition is not in general an LMI,
but notice that
 
rank XI YI  2n :
Therefore we see that if nK  n in the lemma, the rank condition becomes
vacuous and we are left with only the LMI condition. We can now prove
the synthesis theorem.
 7.2. Synthesis 221
Theorem 7.9. A synthesis exists for the H1 problem, if and only if there
exist symmetric matrices X > 0 and Y > 0 such that
(a)
2  3

NX 0  4A XB +XXA XB
 1 C1 N 0

 X
;I D11 0 I < 0 ;
5
0 I 1
C 1 D ;I
11
(b)
2   3
NY 0 4AY C+YY A Y;CI1 DB1 5 NY 0 < 0 ;
   

0 I 1 11 0 I
B1  ;I
D11
(c)
 
X I  0;
I Y
where NX and NY are full-rank matrices whose images satisfy
 
ImNX = ker C2 D21
  :
ImNY = ker B2 D12
Proof . Suppose a controller exists, then by Lemma 7.7 a controller exists
if and only if the inequalities
WP TXL WP < 0 and WQ HXL WQ < 0
hold for some symmetric, positive denite matrix XL in R(n+nK )(n+nK ) .
By Lemma 7.7 these LMIs being satised imply that (a) and (b) are met.
Also invoking Lemma 7.8 we see that (c) is satised.
Showing that (a{c) imply the existence of a synthesis is essentially the
reverse process. We choose nK  n, in this way the rank condition in
Lemma 7.8 is automatically satised, and thus there exists an XL in
R(n+nK )(n+nK ) which satises (7.13). The proof is now completed by using
XL and (a{b) together with Lemma 7.7.

This theorem gives us exact conditions under which a solution exists
to our H1 synthesis problem. Notice that in the suciency direction we
required that nk  n, but clearly it suces to choose nk = n. In other words
a synthesis exists if and only if one exists with state dimension nK = n.
What if we want controllers of order nK less than n? Then clearly from
the above proof we have the following characterization.
Corollary 7.10. A synthesis of order nk exists for the H1 problem, if and
only if there exist symmetric matrices X > 0 and Y > 0 satisfying (a), (b),
222 7. H1 Synthesis
and (c) in Theorem 7.9 plus the additional constraint
 
rank XI YI  n + nK :
Unfortunately this constraint is not convex when nK < n, so this
says that in general the reduced order H1 problem is computationally
much harder than the full order problem. Nevertheless, the above explicit
condition can be exploited in certain situations.

7.3 Controller reconstruction

The results of the last section provide us with an explicit way to determine
whether a synthesis exists which solves the H1 problem. Implicit in our
development is a method to construct controllers when the conditions of
Theorem 7.9 are met. We now outline this procedure, which simply retraces
our steps so far.
Suppose X and Y have been found satisfying Theorem 7.9 then by
Lemma 7.8 there exists a matrix XL 2 RnnK satisfying
   
XL = X? ?? and XL;1 = Y? ?? :
From the proof of the lemma we can construct XL by nding a matrix
X2 2 RnnK such that X ; Y ;1 = X2 X2. Then
 
XL = X IX X 2
2
has the properties desired above. As seen before, the order nK need be no
larger than n, and in general can be chosen to be the rank of X ; Y ;1 .
Next by Lemma 7.2 we know that there exists a solution to
HXL + Q J  PXL + PX L JQ < 0 ;
and that any such solution J provides the state space realization for a
feasible controller K . The solution of this LMI can be accomplished using
standard techniques, and there is clearly an open set of solutions J .

7.4 Exercises
1. Prove Lemma 7.1.
 7.4. Exercises 223
2. Generalization
 of the KYP Lemma. Let A be a Hurwitz matrix, and
Q S
let = S  R be a symmetric matrix with R > 0. We dene
 ;1   ;1 
^(j!) = (j!I ; A) B (j!I ; A) B
I I
Show that the following are equivalent:
(i) ^(j!)   > 0 for all ! 2 R.
(ii) The Hamiltonian matrix
 ;1 S  
H = ;(AQ;;BR ;BR;1B 
BR;1 S  ) ;(A ; BR;1 S  )
is in the domain of the Riccati operator.
(iii) The LMI
  
A X + XA XB + > 0
BX 0
admits a symmetric solution X .
(iv) There exists a quadratic storage function V (x) = x Px such the
dissipation inequality
    
_V  x(t) Q S x(t)
u(t) S R ; I u(t)
is satised over any solutions to the equation x_ = Ax + Bu.
Hint: The method of proof from x7.1.1 can be replicated here.
3. Spectral Factorization. This exercise is a continuation of the previous
one on the KYP Lemma. We take the same denitions for , H , etc.,
and assume the above equivalent conditions are satised. Now set
 
M^ (s) = A B
R; 12 (S  + B  X ) R 21 ;
where X = Ric(H ). Show that M^ (s) 2 RH1 , M^ (s);1 2 RH1 , and
the factorization
^(j!) = M^ (j!) M^ (j!)
holds for every ! 2 R.
4. Mixed H2 =H1 control.
(a) We are given a stable system with the inputs partitioned in two
channels w1 , w2 and a common output z :
 
P^ = P^1 P^2 = CA BD1 B02 :
 
224 7. H1 Synthesis
Suppose there exists X > 0 satisfying
2 3
A X + XA XB1 C 
4 B1 X ;I D 5 < 0; (7.17)
C D ;I
Tr(B2 XB2) < 2 : (7.18)
Show that P satises the specications kP^1 kH1 < 1; kP^2 kH2 <
. Is the converse true?
(b) We now wish to use part (a) for state-feedback synthesis. In
other words, given an open loop system
x_ = A0 x + B1 w1 + B2 w2 + B0 u;
z = C0 x + Dw1 + D0 u;
we want to nd a state feedback u = Fx such that the closed
loop satises (7.17)-(7.18). Substitute the closed loop matrices
into (7.17); does this give an LMI problem for synthesis?
(c) Now modify (7.17) to an LMI in the variable X ;1, and show
how to replace (7.18) by two convex conditions in X ;1 and an
appropriately chosen slack variable Z .
(d) Use part (c) to obtain a convex method for mixed H1 /H2 state
feedback synthesis.
5. As a special case of reduced order H1 synthesis, prove Theorem 4.20
on model reduction.
6. Prove Theorem 5.8 involving stabilization. Hint: Reproduce the steps
of the H1 synthesis proof, using a Lyapunov inequality in place of
the KYP Lemma.
7. Connections to Riccati solutions for the H1 problem. Let
2 3
A B1 B2
^
G(s) = C1 0 D12
4 5
C2 D21 0
satisfy the normalization conditions

 C1 D12 = 0 I      
D12 and D21 B1 D21
 = 0 I :

Notice that these (and D11 = 0) are part of the conditions we imposed
in our solution to the H2 -optimal control in the previous chapter.
 7.4. Exercises 225
(a) Show that the H1 synthesis is equivalent to the feasibility of
the LMIs X > 0, Y > 0 and
  
A X + XA + C1 C1 ; C2 C2 XB1 < 0;
B1 X ;I
 
AY + Y A + B1 B1 ; B2 B2 Y C1 < 0;
  
C1 Y

;I 
X I  0:
I Y
(b) Now denote Q = Y ;1 , P = X ;1. Convert the above conditions
to the following:
A P + PA + C1 C1 + P (B1 B1 ; B2 B2 )P < 0;
AQ + QA + B1 B1 + Q(C1 C1 ; C2 C2 )Q < 0;
(PQ)  1
These are two Riccati inequalities plus a spectral radius cou-
pling condition. Formally analogous conditions involving the
corresponding Riccati equations can be obtained when the
plant satises some additional technical assumptions. For details
consult the references.

Notes and references

The H1 control problem was formulated in [152], and was motivated by
the necessity for a control framework that could systematically incorporate
errors in the plant model. At the time this had been a goal of control
research for a number of years, and H2 control seemed poorly suited for
this task [22]. The main observation of [152] was that these requirements
could be met by working in a Banach algebra such as H1 , but not H2 which
lacks this structure. We will revisit this question in subsequent chapters.
The formulation of the H1 problem precipitated an enormous research
eort into its solution. The initial activity was based on the parametriza-
tion of stabilizing controllers discussed in Chapter 5, which reduced the
problem to approximation in analytic function space. This problem was
solved in the multivariable case by a combination of function theory and
state space methods, notably Riccati equations. For an extensive account
of this approach to the H1 problem, see the book [41]. Recent extensions
to innite dimensional systems can be found in [38].
Ultimately, these eorts led to a solution to the H1 problem in terms
of two Riccati equations and based entirely on state-space methods [24];
see the books [50, 155] for an extensive presentation of this approach and
historical references. This solution has close ties to the theory of dier-
ential games (see [6]). For extensions of the Riccati equation method to
distributed parameter systems, see [132].
226 7. H1 Synthesis
One of the drawbacks of the Riccati equation theory was that it required
unnecessary rank conditions on the plant; the ensuing research in removing
such restrictions [16, 126] led to the use of Riccati inequalities [117, 119]
which pointed in the direction of LMIs. Complete LMI solutions to the
problem with no unnecessary system requirements appeared in [42, 89];
these papers form the basis for this chapter, particularly the presentation
in [42].
The LMI solution has, however, other advantages beyond this regular-
ity question. In the rst place, a family of controllers is parameterized,
in contrast to the Riccati solution which over-emphasizes the so-called
\central" solution. This increased exibility can be exploited to impose
other desirable requirements on the closed loop; for a recent survey of these
multi-objective problems see [120].
Also, the LMI solution has led to powerful generalizations: in [18] a
more general version is solved where spatial constraints can be specied;
[31] solves the time varying and periodic problems; nally, the extension
of this solution to multi-dimensional systems forms the basis of Linear
Parameter-Varying control, a powerful method for gain-scheduling design
[89].
 This is page 227
Printer: Opaque this

8
Uncertain Systems

In the last three chapters we have developed synthesis techniques for feed-
back systems where the plant model was completely specied, in the sense
that given any input there is a uniquely determined output. Also our plant
models were linear, time invariant and nite dimensional. We now return
our focus to analysis, but move beyond our previous restriction of having
complete system knowledge to the consideration of uncertain systems.
In a narrow sense, uncertainty arises when some aspect of the system
model is not completely known at the time of analysis and design. The
typical example here is the value of a parameter which may vary according
to operating conditions. As discussed in the introduction to this course, we
will use the term uncertainty in a broader sense to include also the result
of deliberate under-modeling, when this occurs to avoid very detailed and
complex models.
To illustrate this latter point, let us brie y and informally examine some
of the issues involved with modeling a complex system. In the gure we
have a conceptual illustration of a complex input-output system.
As depicted it has an input w and an output z. A complete description of
such a system is given by the set of input-output pairs
P = f(w;
 z) : z is the output given input wg:
Here the complex system can be nonlinear or innite dimensional, but is
assumed to be deterministic.
Now suppose we attempt to model this complex system with a mapping
G. Then similar to the complex system, we can describe this model by an
228 8. Uncertain Systems

z Complex system w

input-output set
RG = f(w; z ) : z = Gwg :
Our intuitive notion of a good model is, given an input-output pair (w;
 z)
in the set of system behaviors P and the corresponding model pair (w;
 z) 2
RG that
z is a good approximation to z :
If however G is much simpler than the system it is intended to model, this
is clearly an unreasonable expectation for all possible inputs w.

z G w

Now consider instead a set R of input-output pairs generated by a family

G which is a specied set of input-output maps.
R = f(w; z ) : there exists G 2 G such that z = Gwg: (8.1)
In words this set is just the union of the sets RG when G ranges over
the set G . Such a set R is a relation with respect to input-output pairs,
but may no longer specify a function; namely there could be many pairs
(w; z0 ); (w; z1 ); : : : ; (w; zn ), with distinct zk , associated with a given input
w. Notice this can never be the case in set RG . When would R provide a
good model for the system given by P ? Descriptively, the basic requirement
would be that
most elements of P are close to some element of R.
In this way R would approximately contain or cover the complex model
(strict covering P  R is also desirable, but may be more dicult to
ascertain). Now the key observation is that this requirement can indeed be
satised even when the description of G is much simpler than the complex
 8.1. Uncertainty modeling and well-connectedness 229
system; for example each G can be a linear mapping, yet the system to be
modeled could be signicantly nonlinear.
Example:
Consider the static input-output system specied by the set
P = f(w; z ) : z (t) = f (w(t)), for each tg ;
where f is a highly complex nonlinear function, but is known to satisfy the
so-called sector bound condition
jf (w(t) )j  jw(t)j; for each t:
Suppose we focus on signals in L2 . Then clearly the relation
R = f(w; z ) : kz k2  kwk2 g
will satisfy R  P and thus cover the complex behavior. Now it is easily
shown (see Lemma 8.4 below) that R has the form (8.1) where G is the
unit ball of linear operators on L2. 
What is the price paid for this containment procedure? Mainly, that the
relation R often describes a much larger behavior than P ; in particular it
is usually not true that \most elements of R are close to some element of
P ". In the above example, pairs (w; z ) 2 R may for instance have the two
signals supported in disjoint time intervals, a situation which is not present
in P .
In other words, this modeling technique decreases complexity by intro-
ducing uncertainty into the system description. This kind of tradeo is
implicitly present in most of engineering modeling. What is special about
the robust control theory that follows is that the uncertainty is made ex-
plicit in terms of a set of input-output models G . For the remainder of
the chapter we study an important way to describe such sets, and pro-
vide methods to analyze such descriptions. In the next chapter we will
explicitly relate our current work to robust stability and robust perfor-
mance of complex feedback systems, and will discuss synthesis under these
conditions.

8.1 Uncertainty modeling and well-connectedness

In this section we introduce precisely the type of uncertain systems we will
be studying. The idea is to parametrize the set G from the above discussion
using perturbations. Then by varying the nature of these perturbations we
can alter the characteristics of the set G and ultimately the set of input-
output pairs R. We will seldom need to directly refer to the sets G and
R, however it is important to keep the above motivation for our system
230 8. Uncertain Systems
models in mind as they will help us to chart our course. The basic setup
we now use is shown in Figure 8.1.

p q

M
z w
Figure 8.1. Uncertain system model
Here M and
are bounded operators on L2[0; 1), and w the system input
is in L2[0; 1). The picture represents the map w 7! z which is formally
dened by the equations
q =
p
    
p = M11 M12 q ;
z M21 M22 w
where M is compatibly partitioned with respect to the inputs and outputs.
Our goal will be to understand the possible maps w 7! z when all that is
known about
is that it resides in a pre-specied subset
of the bounded
linear operators on L2 . We will consider two fundamental types of sets in
this chapter which are specied by spatial structure, norm bounds and their
dynamical characteristics.
Before embarking on this specic investigation let us quickly look at the
algebraic form of maps that can be obtained using the above arrangement.
Observe that if (I ; M11
) is nonsingular then
w 7! z = M22 + M21
(I ; M11
);1 M12 =: S(M;
);
where this expression denes the notation S(M;
), the upper star product.
Recall that we dened the lower star product S (
;
) earlier when studying
synthesis.
Examples:
We now present two examples of the most common types of uncertain
models. Suppose that
 
M= 0 I : I G
 8.1. Uncertainty modeling and well-connectedness 231
Then we see
S(M;
) = G +
:
That is the sets of operators generated by S(M;
), when
2
is simply
that of an additive perturbation to an operator G.
Similarly let
 
M = I0 G
G
and get that
S(M;
) = G +
G = (I +
)G;
which is a multiplicative perturbation to the operator G.

Two examples have been presented above, which show that the setup
of Figure 8.1 can be used to capture two simple sets. It turns out that
this arrangement has a surprisingly rich number of possibilities, which are
prescribed by choosing the form of the operator M and the uncertainty or
perturbation set
. One reason for this is that this modeling technique is
ideally suited to handle system interconnection (e.g. cascade or feedback),
as discussed in more detail below. The exercises at the end of the chapter
will also highlight some of these possibilities. First, though, we lay the basis
for a rigorous treatment of these representations by means of the following
denition.
Denition 8.1. Given an operator M and a set
, the uncertain system
in Figure 8.1 is robustly well-connected if
I ; M11
is nonsingular for all
2
:
We will also use the terminology (M11 ;
) is robustly well-connected to
refer to this property.
When the uncertain system is robust well-connected, holds we are assured
that the map w 7! z of Figure 8.1 is bounded and given by S(M;
), for
every
2
. At rst sight, we may get the impression that robust well-
connectedness is just a technicality. It turns out however, that it is the
central question for robustness analysis, and many fundamental system
properties reduce to robust well-connectedness; we will, in fact, dedicate
this chapter to answering this question in a variety of cases. Subsequently,
in Chapter 9 these techniques will be applied to some important problems
of robust feedback design.
A simple rst case of evaluation of robust well-connectedness is when the
set
is the unit ball in operator space. In this case of so-called unstructured
uncertainty, the analysis can be reduced to a small-gain property. The term
232 8. Uncertain Systems
unstructured uncertainty is used because we are not imposing any structure
on the perturbations considered except that they are contractive.
Theorem 8.2. Let Q be an operator and
= f
2 L(L2) : k
k  1g.
Then I ; Q
is nonsingular for all
2
if and only if
kQk < 1:
Proof . The \if" direction, observe that for any
2
, we have the norm
inequalities
kQ
k  kQkk
k < 1:
Then it follows from the small-gain theorem in Chapter 3 that I ; Q
is
invertible.
For the \only if" direction, we must show that if kQk  1 then there
exists
2
with I ; M11
singular. From our work in Chapter 3, we
obtain the spectral radius condition
(QQ ) = kQk2 = kQk2  1;
where QQ has only nonnegative spectrum, so I ; QQ is singular with
 = kQk2.
Dividing by  we see that I ; QQ;1 is singular, and thus we set

= ;1 Q , which is contractive since k
k = kQk;1  1. 
The preceding result reduces the analysis of whether the system of Figure
8.1 is robust well-connected under unstructured uncertainty, to evaluating
the norm of M11. If for instance M11 is a nite dimensional, causal LTI
operator, a computational evaluation follows from the tools of Chapter 7.
We remark that there is nothing special about perturbations of size one.
If the uncertainty set were specied as f
2 L(L2 ) : k
k   g, the cor-
responding test would be kM11 k < 1= ; now, since a normalizing constant
can always be included in the description of M , we will assume from now
on that this normalization has already been performed and our uncertainty
balls are of unit size.
Therefore the analysis of well-connectedness is simple for unstructured
uncertainty. There are, however, usually important reasons to impose ad-
ditional structure on the perturbation set
, in addition to a norm bound.
We explain two such reasons here, the rst is illustrated by the following
example.
Example:
The gure depicts the cascade interconnection of two uncertain systems
of our standard form; i.e., a composition of the uncertain operators. It is
a routine exercise, left for the reader, to show that the interconnection is
equivalent to a system of the form of Figure 8.1, for an appropriately chosen
 8.1. Uncertainty modeling and well-connectedness 233

-
1 -
2
p1 q1 p2 q2
 
 N  H 
Figure 8.2. Cascade of uncertain systems

M and
     
p = pp1 ; q = qq1 ;
=
01
0 :
2 2 2
Therefore we nd that the uncertainty set
for the composite system
will have, by construction, a block diagonal structure. 
Thus we see that system interconnection generates structure; while com-
ponents can be modeled by unstructured balls as described in previous
examples, a spatially structured uncertainty set can be used to re ect the
complexity of their interconnection. We remark that many other forms of
interconnection (e.g. feedback) can also be accommodated in this setting.
We now move to a second source of uncertainty structure; this arises
when in addition to robust well-connectedness, one wishes to study the
norm of the resulting operator set S (M;
). That is assuming (M11 ;
) is
well-connected, does
kS(M;
)k < 1
hold for all
2
. We have the following result which states that the latter
contractiveness problem can be recast as a well-connectedness question.
Proposition 8.3. Dene the perturbation set
   

=
=
0u
0p :
u 2
; and
p 2 L(L2 ) with k
pk  1 :
p

Then (M11 ;
) is robustly well-connected and

kS(M;
u )k < 1; for all
u 2
,
if and only if I ; M
is invertible for all
2
. p

Proof . We rst demonstrate \only if". For any

 

=
0u
0 2
;
p p
234 8. Uncertain Systems
we have  
I ; M
= I ;;MM11

u I ;;MM12

p :
21 u 22 p
Since by denition of
we know
u 2
, it follows by assumption that
p
(I ; M11
u );1 exists. It is easily veried that
   
I ; M11
u ;M12
p = I 0
(8.2)
;M21
u I ; M22
p ;M21
u (I ; M11
u );1 I
 
I ; M11
u ;M12
p
0 I ; S(M;
u )
p
Since the rst operator on the right of the equality is nonsingular, (I ;
M
);1 exists if I ; S(M;
u )
p is nonsingular. This follows from Theorem
8.2, setting Q = S(M;
u ) which by assumption is contractive.
The \if" direction: suppose
u 2
then clearly
 

:=
0u 00 2
p

and  
I ; M
= I ;;MM11

u I0
21 u
is nonsingular by assumption. Therefore (I ; M11
u );1 exists for all
u 2

.
Now looking at (8.2) for any xed
u 2
, and all
p satisfying k
p k 
1, we have that
I ; S(M;
u )
p is nonsingular :
Once again, Theorem 8.2 implies that we must have kS(M;
u )k < 1 for
all
u 2
. 
Thus we have identied two reasons for the introduction of spatial,
block-diagonal structure to our perturbation set
. The rst is that if
a system is formed by the interconnection of subsystems, then this struc-
ture arises immediately due to perturbations which may be present in the
subsystems. Our second motivation is because this will enable us to con-
sider the performance associated with the closed-loop. The next section
is devoted to studying the implications of this structure in the analysis
of well-connectedness. Later on in this chapter we will explore additional
structural constraints which can be imposed on our uncertainty set.

8.2 Arbitrary block-structured uncertainty

This section is devoted to developing an analysis test for robust well-
connectedness when the only restriction imposed on the uncertainty is that
 8.2. Arbitrary block-structured uncertainty 235
it have a block-diagonal spatial structure. In particular, the main question
is whether one can generalize the small gain test which was valid under
unstructured perturbations.
We start by dening the block-diagonal uncertainty set
, which is a
a

spatially structured subset of the unit ball in L(Lm 2 ).

m

= fdiag(
1 ; : : : ;
d) :
k 2 L(L2 k ) and k
k k  1 holds.g;
a

where the spatial dimensions mk are xed and satisfy m = m1 +


+ md .

The above notation means that every perturbation
in this set
is of a
the form
2 3

1 0


0
6 . . . . . . .. 77

= 66 0. . . . 7;
6
7
4 .. . .
. . 05
0


0
d
where each
k can be any contractive linear operator on the space Lm k
2 .
This set constrains the spatial structure of its members, but allows any
other dynamical characteristics.
In line with our discussion at the beginning of the chapter, it is useful to
study the relation formed by the input-output pairs which are parametrized
by
. Referring to Figure 8.1, consider the set
a

Ra = f(p; q) 2 L2 : kpk k  kqk k; where k = 1; : : : ; dg

Here pk , qk denote the parts of the vectors p, q which correspond to the
spatial structure
. We claim that
a

Ra is equal to the set f(p; q) 2 L2 : q =
p; for some
2
g; (8.3)

a

this follows from the following lemma applied to each block.

Lemma 8.4. Suppose p and q are elements of L2[0; 1). The following are
equivalent.
(i) kpk  kqk.
(ii) There exists an operator
2 L(L2 ) with k
k  1, such that
p = q.
Proof . It is clear that (ii) implies (i), so we focus on the other direction.
Assume that p is nonzero, otherwise the result is immediate. Dene
by

u := q hkp;pku2i :
2
Now
p = q as required, and the norm bound on
follows by an
application of the Cauchy-Schwartz inequality. 
This Lemma completes the example in the introduction of this chapter,
where a relation in terms of a single norm constraint was used to cover
236 8. Uncertain Systems
a complex nonlinearity. Here the relation Ra allows us to impose a nite
number of such constraints.
A word on terminology: because the operators in
are linear and in
a

general not time invariant, this structure is sometimes called linear time-
varying (LTV) uncertainty. We have used the more general term arbitrary
block-structured uncertainty since
(i) From a modeling perspective, this uncertainty set is often not mo-
tivated by time variation. As explained above, we could well be
modeling a nonlinearity;
(ii) From a mathematical perspective, the relation Ra is indeed time
invariant; that is if (p; q) 2 Ra , then (S p; S q) 2 Ra , where S is
the shift on L2 [0; 1). While a linear parametrization of Ra involves
time-varying operators, we will see below that the time invariance of
Ra is indeed the central property required for our analysis.
We now proceed to the study of robust well-connectedness over the set

. Throughout we will assume that the nominal system M is an LTI
a
operator. Our objective is to study under what conditions (M11 ;
) is a
robustly well-connected. Recall that by this we mean
I ; M11
is invertible, for all
2
:a

For convenience, we will drop the subindex from M11 in the following
discussion and simply write M .

8.2.1 A scaled small-gain test and its suciency

We begin with the simple observation that since
is subset of the unit
a

ball of L(L2 ), then clearly the small gain condition

kM k < 1
for unstructured uncertainty must be sucient for robust well-connectedness
over
. However the above norm condition is in general conservative since
a
the block-diagonal restriction did not come into play. In other words the
condition kM k  1 does not necessarily imply that I ; M
is singular for
some perturbation in
. a
A basic tool to reduce the conservatism of the contractiveness condition
above is to introduce operators which commute with the perturbations. We
dene the commutant of the uncertainty set
by a

; = f; 2 L(L2 ) : ; is invertible and ;
=
; for all
2
g: a

Notice that the inverse ;;1 of any ; 2 ; automatically has the same
commuting property, and we can write the identity
;(I ; M
);;1 = I ; ;M ;;1
; for all
2
: a
 8.2. Arbitrary block-structured uncertainty 237
Thus I ; M
is singular if and only if I ; ;M ;;1
is singular. This means
that if we can nd an element ; of the commutant satisfying
k;M ;;1k < 1;
then we can guarantee that (M;
) is robustly well-connected simply
a
by invoking the small gain condition. This motivates us to describe these
commuting operators.
You will show in the exercises at the end of the chapter, that ; 2 ; if
and only if ; is of the form
2 3
1 I 0


0
6 ... ... 7
; = 66 0. . .
6 7
7;
4 .. . . . . 0 75
0


0 d I
for some constant nonzero scalars k in C . Clearly since such an operator
; is memoryless it can also be thought of as a matrix.
With this description we recapitulate our previous discussion as follows.
Proposition 8.5. Suppose M is a bounded operator. If there exists ; in
; such that
k;M ;;1k < 1;
then (M;
) is robustly well-connected.
a

It will also be convenient to consider the set of positive matrices of the

same spatial structure; we dene the set
P ; = f; 2 Rmm : ; = diag( 1 Im1 ; : : : ; d Imd ); where the scalars k > 0g:
(8.4)
Observe that P ; is a convex cone of matrices, as dened in Chapter 1.
Also notice that:
 If ; 2 ;, then ; ; 2 P ;.
 ; is in P ; if and only if its square root ; 21 is in P ;.
These properties allow us to claim that in the above proposition it suces
to restrict attention to the positive elements of the commutant. We now
state this equivalence more explicitly, together with two other restatement
of the above test.
Proposition 8.6. Suppose M is a time invariant bounded operator, with
transfer function M^ (s) = C (Is ; A);1 B + D, and A is Hurwitz of order
n. Then the following are equivalent:
(i) There exists ; in ; such that k;M ;;1k < 1;
(ii) There exists ; in P ; such that k; 21 M ;; 21 k < 1;
238 8. Uncertain Systems
(iii) There exists ; in P ; such that
M ;M ; ; < 0; (8.5)
(iv) There exists ; in P ; and a symmetric n  n matrix X > 0 such that
   
C ; C D + A X + XA XB < 0 : (8.6)
D BX ;;
Here we have added the additional assumption that M^ is in RH1 for
condition (iv), however the rst three conditions are equivalent for any
operator M . The reformulation given in (iii), expresses our norm condition
in terms of a convex operator inequality; we recall that (8.5) means the
operator is negative denite, that is there exists  > 0 such that
hv; (M  ;M ; ;)vi  ;kvk2 for all v 2 L2 :
The second reformulation says that (i) can be reduced to an equivalent
LMI feasibility condition.
Proof . We rst show that (i) implies (iii). For this purpose we write
k;M ;;1vk2  (1 ; )kvk2
for any v 2 L2 and some  > 0; equivalently, setting u 2 ;;1 v we nd that
k;Muk2  (1 ; )k;uk2
for any u 2 L2. Now by rewriting the above inequality we obtain
hv; (M  ; ;M ; ; ;)v i  ;k;uk2  ; k;;1k2 kuk2;
which implies (iii) since ; ; 2 P ;.
The above argument can be essentially 12
reversed to show that (iii) implies
(ii); also, since ; 2 P ; implies ; 2 P ;  ;, it is clear that (ii) implies
(i).
This shows the equivalence of the rst three conditions, with no reference
to nite dimensionality or time invariance of the operator M . We now
impose this restriction and show that (ii) implies (iv). Since M is LTI and
nite dimensional, and ; is constant, the norm condition in (i) amounts to
the H1 norm condition
k; 21 M^ ;; 12 k1 = k; 21 C (Is ; A);1 B ;; 12 + ; 21 D;; 21 k1 < 1:
Since A is Hurwitz, we can apply the KYP lemma and conclude that there
exists a symmetric matrix X > 0 such that
C1  ; C D;; 12  + A X + XA XB ;; 12 < 0:
   
1
;; 2 D ;; 2 B  X ;I
 8.2. Arbitrary block-structured uncertainty 239
Now left 1 and right multiply this inequality by the symmetric matrix
diag(I; ; 2 ) to get
   
C ; C D + A X + XA XB < 0:
D BX ;;
This is precisely the LMI in (8.6). Clearly the above steps can be reversed
to establish that (iv) implies (ii). 
The preceding results provide a tractable means to establish robust well-
connectedness over
. Indeed the existence of an appropriate scaling
a

matrix ; 2 P ; can be determined from (8.6) by LMI computation. Since

the identity matrix is a member of P ; there always exists an ; in P ; such
that k; 21 M ;; 12 k  kM k. Therefore this scaled small-gain test provides a
less conservative test for robust well-connectedness than the previous condi-
tion kM k < 1. Our next task is to determine whether this scaling procedure
has eliminated all conservatism from the small-gain test. The main result
of this section is that this is indeed the case: the scaled small-gain condition
is necessary as well as sucient.

8.2.2 Necessity of the scaled small-gain test

In this section we will concentrate our eorts on showing that the scaled
small-gain condition must hold if (M;
) is a well-connected uncertain
a

system. In order to achieve this goal we will employ some of the ideas and
results from convex analysis introduced in x1.2.2. Our strategy will be to
translate our problem to one in terms of the separation of two specic sets
in Rd . We will be able to show that well-connectedness implies that these
sets are strictly separated, and that the scaled small-gain condition implies
there exists a hyperplane that strictly separates these sets. We will then
obtain our necessity result by showing that the strict separation of the
sets implies the stronger condition that there exists a strictly separating
hyperplane between them.
It will be advantageous to again look at the relation Ra introduced in
the preliminary discussion of this section and dened in (8.3), rather than
its parametrization by
. Also we will wish to describe our relation Ra
a
in terms of quadratic inequalities of the form kpk k2 ; kqk k2  0. Out of
convenience we rewrite these as
kEk pk2 ; kEk qk2  0; k = 1 : : : d; (8.7)
where the mk  m matrix
 
Ek = 0


0 I 0


0
selects the k-th block of the corresponding vector. The spatial dimensions
of Ek are such that any element ; = diag( 1 I; : : : dI ) in the set P ; can
240 8. Uncertain Systems
be written as
; = 1 E1 E1 +


d Ed Ed : (8.8)
Inequalities such as the one in (8.7) are called integral quadratic constraints
(IQCs); the name derives from the fact that they are quadratic in p and
q, and involve the energy integral. We will have more to say about general
IQCs in Chapter 10.
We wish to study the interconnection of Ra to the nominal system M ;
for this purpose, we impose the equation p = Mq on the expression (8.7)
and dene the quadratic form k : L2 ! R by
k (q) := kEk Mqk2 ; kEk qk2 = hq; (M  Ek Ek M ; Ek Ek )q i (8.9)
With this new notation, we make the following observation.
Proposition 8.7. If (M;
) is robustly well-connected, then there cannot
a
exist a nonzero q 2 L2 such that the following inequalities hold:
k (q)  0; for each k = 1; : : : ; d: (8.10)
This proposition follows from the reasoning: if such a q existed, we would be
assured that (Mq; q) 2 Ra . Thus by the earlier discussion, see Lemma 8.4,
there would exist a
2
such that
a

Mq = q;
implying that the operator (I ;
M ) would be singular. And therefore,
invoking Lemma 3.16, the operator I ; M
would also be singular. Thus
we have related robust well-connectedness to the infeasibility of the set of
inequalities in (8.10).
We now wish to relate the quadratic forms k to the scaled small-gain
condition. In fact this connection follows readily from version (ii) in Propo-
sition 8.6 by expressing the elements of P ; as in (8.8), which leads to the
identity
d
X d
X
hq; (M  ;M ; ;)qi = k hq; (M  Ek Ek M ; Ek Ek )q i = k k (q):
k=1 k=1
The following result follows immediately by considering condition (ii) of
Proposition 8.6; the reader can furnish a proof.
Proposition 8.8. The equivalent conditions in Proposition 8.6 hold if and
only if there exist scalars k > 0 and  > 0 such that
1 1 (q) +


+ d d (q)  ;kqk2; for all q 2 L2: (8.11)
Furthermore, if such a solution exists for (8.11) then there does not exist
a nonzero q in L2 satisfying (8.10).
We have thus related both robust well-connectedness and the scaled
small-gain test, to two dierent conditions involving the quadratic forms k .
 8.2. Arbitrary block-structured uncertainty 241
The former condition states the infeasibility of the set of constraints (8.10);
the latter condition is expressed as a single quadratic condition in (8.11)
with multipliers k , and implies the former. In the eld of optimization a
step of this sort, using multipliers to handle a set of quadratic constraints,
is sometimes called the S-procedure. The S-procedure is termed lossless
when there is no conservatism involved, namely when the conditions are
equivalent. In what follows we will show this is the case here for us.
To demonstrate this we rst express our ndings in geometric terms so
that we can bring convex analysis to bear on the problem. To this end
introduce the following subsets of Rd :
 = f(r1 ; : : : ; rd ) 2 Rd : rk  0; for each k = 1; : : : ; dg;
r = f(1 (q); : : : ; d (q)) 2 Rd : q 2 L2 ; with kqk2 = 1g:
Here  is the positive orthant in Rd , and r describes the range of the
quadratic forms k as q varies in the unit sphere of L2. For brevity we will
write (q) := (1 (q); : : : ; d (q)) and = ( 1 ; : : : d ). Therefore with the
standard inner product in Rd , we have
h ; (q)i = 1 1 (q) +


+ dd (q):
Now we are ready to interpret our conditions geometrically. From the
discussion and Proposition 8.7 we see that robust well-connectedness im-
plies that the closed set  and the set r are disjoint, that is  \r = ;. Now
r is a bounded set but it is not necessarily closed. Therefore we cannot
invoke Proposition 1.3 to conclude these sets are strictly separated when
they are disjoint. Nonetheless this turns out to be true.
Proposition 8.9. Suppose that (M;
) is robustly well-connected over

. Then the sets  and r are strictly separated, namely
a

D(; r) := r2inf;y2r jr ; yj > 0:

This result states that not only are the sets  and r disjoint, but they are
separated by a positive distance. Its proof is technically quite involved and
adds little of conceptual signicance to our discussion. For this reason the
proof is relegated to Appendix B.
Turning now to the condition in (8.11), we see that it species the
existence of a vector with positive entries and an  > 0, such that
h ; yi  ;; for all y 2 r:
In other words, the set r is constrained to lie strictly inside the negative
half-space dened by the hyperplane h ; yi = 0. Notice also that since the
entries k of are positive, the set  will automatically lie in the opposite
half-space. We have the following result.
Proposition 8.10. The conditions of Proposition 8.6 are satised if and
only if there is a hyperplane strictly separating the sets r and .
242 8. Uncertain Systems
Proof . From the denition of a strictly separating hyperplane, we need to
show that the conditions in Proposition 8.6 are equivalent to the existence
of a vector 2 Rd , and scalars  and  , such that
h ; yi   <   h ; ri holds, for all y 2 r; r 2 : (8.12)
The \only if" direction follows from the above comments and Proposi-
tion 8.8 by choosing  = ; and  = 0.
To establish \if", suppose that (8.12) holds for some ,  and  . First
note, from the structure of  that we must have k  0, for each k; other-
wise we could make the corresponding rk arbitrarily large and violate the
bound   h ; ri. Hence we see that 0  h ; ri, for all r in ; therefore
without loss of generality we set  = 0. Observe this implies that  < 0.
Finally we show there exists an  > 0 such that each of the entries k
can be perturbed to be strictly positive, satisfying a bound of the form
h ; yi  ; < 0, for all y 2 r. To see that this is is always possible simply
recall that each k is non negative,  < 0 and r is bounded. This establishes
inequality (8.11) and completes our proof by invoking Proposition 8.8. 
Figure 8.3 contains an illustration of the preceding conditions for the case
where d = 2. In part (a) of the gure we see the pair of sets r and  where
they have a positive distance between them, as implied by robust well-
connectedness according to Proposition 8.9. However the depiction in (b)
of Figure 8.3 imposes a strict hyperplane separation between the two sets,
which is equivalent to the scaled small-gain condition by the proposition
just proved.

r r
(a) (b)
Figure 8.3. (a) D(; r) > 0; (b) , r separated by a hyperplane.
Consequently, if we wish to show that the conditions of Proposition 8.6
are necessary for robust well-connectedness, we need to prove that strict
separation of the sets r and  automatically implies the existence of a
strictly separating hyperplane. This leads us to think of convex separation
theory, reviewed in x1.2.2. The closed positive orthant  is clearly a convex
set. As for the set r, we have the following result. Its proof will rely strongly
 8.2. Arbitrary block-structured uncertainty 243
on the time invariance of the operator M since it implies the forms k also
enjoy such a property.
Lemma 8.11. The closure r is convex.
Proof . We introduce the notation
Tk := M  Ek Ek M ; Ek Ek :
For each k = 1; : : : ; d, the operator Tk is self adjoint and time invariant;
also from (8.9) we know that k (q) = hTk q; qi.
Choose two elements y = (q) and y~ = (~q ) from the set r. By denition,
kqk = kq~k = 1. We wish to demonstrate any point on the line segment
that joins them is an element of r . That is for any  2 [0; 1] we have
y + (1 ; )~y 2 r . Given such an  let
p p
q :=  q + 1 ;  S q~;
where S is the usual time shift operator.
Our rst goal is to examine the behavior of (q ) as  tends to innity.
It is convenient to do this by considering each component k (q ). So for
any given k we compute
k (q ) = hTk q ; q i
p
= hTk q; qi + (1 ; )hTk S q~; S q~i + 2 (1 ; )RehTk q; S q~i:
(8.13)
We rst observe, from the time invariance of Tk , that the second term on
the right hand side satises
hTk S q~; S q~i = hS Tk q~; S q~i = hTk q~; q~i = k (~q ):
We this observation we can rewrite (8.13) as
p
k (q ) = k (q) + (1 ; )k (~q) + 2 (1 ; )RehTk q; S q~i:
Now we let  ! 1. Clearly the inner product on the right converges to zero
since the rst element is xed, and the second is being shifted to innity.
Thus we have
!1 k (q ) = k (q) + (1 ; )k (~q ):
lim
Collect all components, for k = 1; : : : ; d, to conclude that
lim (q ) = y + (1 ; )~y:
 !1 
Thus we have succeeded in obtaining the convex combination y + (1 ;
)~y as a limit of vectors in the range of ; however we have not established
that the q have unit norm, as required in the denition of r; to address
this, note that
lim kq k2 = kqk2 + (1 ; )kq~k2 = 1:
 !1 
244 8. Uncertain Systems
This follows by the same argument as the above, replacing Tk by the iden-
tity operator. Thus the q have asymptotically unit norm, which implies
that
 
lim  kqq k = y + (1 ; )~y:
 !1 
Now by denition the elements on the left are in r, for every  , so we
conclude that
y + (1 ; )~y 2 r :
Finally, by continuity the same will hold if we choose the original y and y~
from r rather than r; this establishes convexity. 
We can now assemble all the results of this section to show that the
scaled small-gain condition is indeed necessary for well-connectedness of
(M;
). This result is a direct consequence of the following.
a

 Suppose the uncertain system (M;
) is robustly well-connected,

a

then by Proposition 8.9 we conclude that sets r and  are strictly

separated.
 Since r and  are strictly separated, so must be their closures r and
; now the positive orthant  is convex, an so is r by Lemma 8.11;
therefore the conditions of Theorem 1.5 are satised, and there exists
a strictly separating hyperplane, i.e. there exists 2 Rd such that
(8.12) is satised.
 Now by Proposition 8.10, any and all of the equivalent scaled small-
gain conditions of Proposition 8.6 must hold.
This argument clearly shows that well-connectedness implies the scaled
small-gain condition, and we already know the converse statement is true
from Proposition 8.5. Thus we have obtained a necessary and sucient
condition for robustness analysis over
, which is summarized now as the
a
main result of this section.
Theorem 8.12. Suppose M is a time invariant bounded operator on
L2 [0; 1). The uncertain system (M;
) is robustly well-connected if and
a
only if
inf k;M ;;1k < 1: (8.14)
;2;
The theorem states that if the operator M is time-invariant, then the in-
mum condition is an exact test for well-connectedness. The importance of
this result is immediately apparent from Proposition 8.6. Part (iii) shows
that the above test is equivalent to a convex condition over the positive
scaling set P ;. Also if M is a state space system, then well-connectedness
 8.3. The Structured Singular Value 245
of (M;
) reduces to an LMI feasibility problem by part (iv) of the propo-
a
sition. It is the fact that this condition can be checked with surety which
makes the above theorem very attractive.
Having provided computable conditions for well-connectedness with
respect to the uncertainty set
of spatially structured, but otherwise ar-
a
bitrary perturbations, we are now ready to move on to consider uncertainty
sets which are further constrained in their dynamical characteristics.

8.3 The Structured Singular Value

After having studied some robustness problems at some level of detail in
the preceding section, we will now extend some of the lessons learned to ro-
bustness analysis of more general uncertainty models. In x8.4 these general
tools will be applied to another special case of robust well-connectedness,
structured LTI uncertainty.
So far our uncertain systems have been characterized by a nominal com-
ponent M , which we have taken to be LTI, and an uncertain component.
For the latter, we have alternated between a description in terms of a re-
lation R between signals, and a parameterization of the relation by a set
of operators
. While these are interchangeable notions for the analysis
we have performed, some aspects of the problem are best illuminated with
each version.
Specically, we rst studied the unstructured uncertainty set where

was just the unit ball of operators; in this case the analysis of well-
connectedness reduced to small-gain condition on M . We also studied
spatially structured uncertainty, where the analysis of well-connectedness
reduced to a scaled small-gain condition (8.14). In both cases there is the
common theme that robust well-connectedness depends only on a gain con-
dition involving the nominal system M . We now explore how this notion
of small-gain can be generalized to more complex uncertainty structures.
The structures we are interested in consist of imposing some additional
property P (
) to the unit ball of operators, that is

= f
2 L(L2 ) : k
k  1 and the property P (
) holds.g
Examples of constraints imposed by P (
) are block-diagonal structure as
in
, or dynamic restrictions on some operator blocks (e.g. specifying
a

that they are LTI, or static, or memoryless, etc.). We will assume that
P (
) imposes no norm restrictions, and in fact will further assume that if

satises P (
), then so does
for any > 0. Namely the set
C
:= f
2 L(L2 ) : the property P (
) is satised.g
is a cone (the cone generated by
). These assumptions are implicit in
what follows.
246 8. Uncertain Systems
Denition 8.13. The structured singular value of an operator M with
respect to a set
which satises the above assumptions, is
(M;
) := inf fk
k :
2 C
and 1
I ; M
is singularg (8.15)
when the inmum is dened. Otherwise (M;
) is dened to be zero.
The inmum above is undened only in situations were there are no per-
turbations that satisfy the singularity condition. This occurs for instance
if M = 0. If perturbations exist which make I ; M
singular, the inmum
will be strictly positive and therefore (M;
) will be nite. We remark
that the terminology structured singular value originates with the matrix
case, which was the rst to be considered historically, and is discussed in
the next section.
To start our investigation we have the following result which provides an
upper bound on the value of the structured singular value. This result is a
restatement of the small-gain result.
Proposition 8.14. (M;
)  kM k, with equality in the unstructured
case
= f
2 L(L2 ) : k
k  1g.
Proof . If k
k < kM k;1, we know that I ;
M is nonsingular by small-
gain. Therefore the inmum in (8.15) is no less than kM k;1, and thus the
rst statement follows by inversion.
In the unstructured uncertainty case, C
is the entire space L(L2 ); by
scaling in Theorem 8.2, we can always construct
2 L(L2 ), k
k = kM k;1,
and I ;
M singular. This
achieves the inmum and proves the desired
equality. 
So we see that the structured singular value reduces to the norm if
is
unstructured; more precisely (M; BL(L2) ) = kM k where BL(L2) denotes
the unit ball of L(L2 ). Furthermore if the uncertainty is arbitrary block-
structured we have the following version of Theorem 8.12.
Proposition 8.15. Let M be an LTI, bounded operator on L2. Then
(M;
) = inf k;M ;;1k
;2;
a

This proposition follows from Theorem 8.12. We leave details as an exercise.

Returning to the general case, it is clear by now that the structured
singular value is closely related to robust well-connectedness. In fact one
can write the following chain of statements:
(M;
) < 1 if and only if inf fk
k :
2 C
, I ; M
is singularg > 1;
only if I ; M
nonsingular, for all
2
;
if and only if (M;
) is robustly well-connected.
The rst and last equivalences are by denition. In the intermediate step,
we did not make an \if" statement because the inmum might be one, not
 8.3. The Structured Singular Value 247
achieved by any
. In other words, in general one can only say that robust
well-connectedness implies (M;
)  1. This diculty does not appear in
the two cases considered up to now where the inma are indeed achieved.
The issue is a little more delicate in the case of LTI uncertainty, to be
considered in the next section. Nevertheless it is customary to regard the
condition (M;
) < 1 as interchangeable with robust well-connectedness;
this is our desired generalization of a small-gain test.
The following example gives more insight into the broad scope of prob-
lems covered by the structured singular value, as well as its mathematical
structure.
Example:
Let the uncertainty set
= fI :  2 C ; jj  1g. Then we have the
following chain of equalities.
(M;
) = (inf fjj :  2 C ; I ; M is singularg);1
= supfj;1 j : ;1 2 C and I;1 ; M is singularg
= supfjj :  2 spec(M )g
= (M );
where (
) is the operator spectral radius. 
So we see that the spectral radius also ts into this very general notion
of system gain. Notice in particular that while it easy to show that the
structured singular value satises the scaling property
(M;
) = jj(M;
); for all  2 C ;
in general it does not dene a norm. For instance the spectral radius is not
a norm.
We have thus introduced a system property, the structured singular
value, which is directly linked to the analysis of robust well-connectedness
under very general uncertainty structures. At this level of generality, how-
ever, our denition has not yet accomplished much, since it may not be
easy to evaluate . What we would like is a characterization, like the one
obtained in Proposition 8.15 for  with respect to
, that directly lends
a

itself to computation. Such a strong result, however, will not always be

available, which should not be surprising given the eort which was re-
quired to obtain this special case. Still we will nd it valuable to again
adopt our previous the approach based on scaled norms for the general
situation.
Analogously to x8.2.1, we introduce a set of scalings which commute
with perturbations in our uncertainty structure. Dene the nonsingular
commutant
;
:= f; 2 L(L2) : ; is invertible and ;
=
; for all
2
g: (8.16)
248 8. Uncertain Systems
This set does not in general have the structure (8.4) of constant, block-
diagonal matrices, but is nevertheless a well-dened entity. The following
can then be easily established with the methods of x8.2.1 and Proposi-
tion 8.14. Note that this result is true for any bounded operator M ; that
is M need not be a state space system or even time invariant.
Proposition 8.16. Let ;
be the commutant of a general uncertainty set

, as dened in (8.16). Then
(M;
)  inf k;M ;;1k:
;2;

Thus we have obtained an upper bound for the structured singular
value based on commuting scalings; equivalently, the existence of ; in the
appropriate set ;
satisfying
k;M ;;1k < 1 (8.17)
is always sucient for robust well-connectedness over
. What is not true
in general is the necessity of the above condition. In fact only in very special
cases is the bound of Proposition 8.16 known to be an equality; one of
these is of course when M is time invariant and
=
. Thus the scaling
a
method provides in general a conservative test of well-connectedness.
It is also possible in this general setting to introduce a set of positive
scaling operators, by
P ;
= f; 2 L(L2 ) : ; = ;~  ;~ for some ;~ 2 ;
g:
With this denition, it follows that (8.17) is feasible for ; 2 ;
if and only
if either of the following two conditions holds.
 k; 21 M ;; 12 k < 1 is satised for some ; 2 P ;
.
 The operator inequality
M ;M ; ; < 0;
is feasible over ; 2 P ;
.
The above equivalence is analogous to that of conditions (i) through (iii)
in Proposition 8.6, and can be established with similar methods, with one
important exception: here the set P ;
is not in general a subset of ;
.
This makes one of the above steps more involved; we will not pursue this
here, however, since we will not rely on the above equivalence at this level
of generality. Later on we will remark on this issue for the case of time
invariant uncertainty, which we are now ready to discuss.

8.4 Time invariant uncertainty

Having laid out the general framework of the structured singular value, we
now return to a specic uncertainty model, where the perturbations are
 8.4. Time invariant uncertainty 249
assumed to be time invariant in addition to spatially structured. Dene
the perturbation set
by TI

= f
= diag(
1 : : : ;
d) 2 L(L2 ) :
time invariant and k
k  1g:

TI

Notice that this set is simply the intersection of the LTI operators on
L2 [0; 1) with the set
. Bringing in the Laplace transform from Chap-
a

ter 3 we see that each
2
can be identied with its transfer function

^ 2 H1 , which inherits the corresponding spatial structure.
TI

We now introduce the relation RTI that goes along with the operator
set
. It is possible to show given an element q in L2 [0; 1) that
TI

q =
p; for some
2
; if and only if (p; q) 2 RTI ;

TI

where
RTI =
f(p; q) 2 L2 :
^ k 2 H1 ; k
^ k k1  1; j
^ k (j!)^pk (j!)j = jq^k (j!)jg:
From this we see that given any (p; q) 2 RTI the relationship
jp^k (j!)j  jq^k (j!)j; for almost every !;
must be satised since  (
^ k (j!) )  1. As with Ra we have established a
constraint between the sizes of the various components of p and q. However
the description is now much tighter since it is imposed at (almost) every
frequency. In contrast Ra can be written as
 Z 1 Z 1 
Ra = (p; q) 2 L2 : jp^ (j!)j2 d! 
k jq^ (j!)j2 d!;
k k = 1; : : : ; d ;
0 0
which clearly shows that the relation Ra only imposes quadratic constraints
over frequency in an aggregate manner.
As a modeling tool, time invariant uncertainty is targeted at describing
dynamics which are fundamentally linear and time invariant, but which we
do not desire to model in detail in the nominal description M . Thus M
could be a low order approximation to the linearized system, leaving out
high dimensional eects which arise, for instance, from small scale spatially
distributed dynamics. Instead of simply neglecting these dynamics, time
invariant perturbations provide a formalism for including or containing
them in RTI using frequency domain bounds.

8.4.1 Analysis of time invariant uncertainty

We will now discuss questions of robust well-connectedness for uncertain
systems with linear time-invariant uncertainty. What we will nd in the
sequel is that robust well-connectedness in this particular setting can be
reduced to a matrix test at each point in the frequency domain. Before
setting out to show this and the related results we look at an illuminating
250 8. Uncertain Systems
example. The following example points out the dierence we might expect
from the case of arbitrary structured uncertainty.
Example:

p1  q2
G1
?

1
2
q1-
G2
p2 6
Figure 8.4. Example system.
The uncertain system of Figure 8.4 is comprised of two xed linear time
invariant systems G1 , G2 , and two uncertainty blocks
1 ,
2 . All blocks are
single-input and -output. We are interested in whether this conguration
is well-connected in the sense of I ;
1 G1
2 G2 being nonsingular. We can
routinely redraw this system in the standard conguration of Figure 8.1,
with
   
0 G 1
M = M11 = G 0 and
= 0
:
1 0
2 2
Clearly I ;
1 G1
2 G2 is singular if and only if I ; M
is singular. Suppose
k
1 k  1 and k
2 k  1. To investigate the well-connectedness of this
system we can invoke the small gain theorem and impose the sucient
condition for robust well-connectedness given by
k
1 G1
2 G2 k < 1:
If the uncertainty is time invariant, then
2 and G1 commute and we
can use the submultiplicative inequality to obtain the stronger sucient
condition
kG1 G2 k = kG^ 1 G^ 2 k1 = ess sup jG^1 (j!)G^2 (j!)j < 1 (8.18)
!2R
for robust well-connectedness. If instead
1 and
2 are arbitrary contrac-
tive operators, we are not allowed to commute the operators, so we can
only write the small-gain condition
kG1 kkG2k = kG^ 1 k1 kG^ 2 k1 = (ess sup jG^1 (j!)j)(ess sup jG^2 (j!)j) < 1:
!2R !2R
(8.19)
These conditions are dierent in general: frequently (8.19) can be more
restrictive than (8.18) since values of jG^ 1 (j!)j and jG^ 2 (j!)j at dierent
 8.4. Time invariant uncertainty 251
frequencies can give a larger product. For instance
G^1 (s) = s + 1 ; G^2 (s) = s +s 1
will satisfy (8.18) for  < 2, but (8.19) only for  < 1. An extreme case
would be if G^ 1 (j!) and G^ 2 (j!) had disjoint support.
To interpret these frequency domain conditions, notice that for
2
,
the transfer functions
^ 1 (j!) and
^ 2 (j!) are contractive at every !, thus
TI

the small gain analysis can be decoupled in frequency, which makes (8.18)
sucient for well-connectedness. When the perturbation class is
the a
operators
1 and
2 are still contractive, but they are allowed to \transfer
energy" between the frequencies where G^1 and G^2 achieve their maximum
gain, making well-connectedness harder to achieve.
It turns out that these conditions are also necessary for robust well-
connectedness in each respective case. While this can be shown directly in
the conguration of Figure 8.4, it is illustrative to write it in the standard
form and apply the robustness analysis techniques of this chapter. This is
left as an exercise. 
We now proceed with the analysis of uncertain systems over
. A TI
standing assumption throughout the rest of this section is that the nominal
operator M is nite dimensional LTI, i.e. it has a transfer function M^ (s) 2
RH1 .
Our main tool will be the structured singular value , dened as in
Denition 8.13 for the present uncertainty set
. Following the general
TI

approach outlined in the previous section, the rst task is to identify the
operators ; which commute with the structure
. Notice that such an
TI
operator must commute with the delay operator S , for every   0, since
the delay S is itself a member of
. Thus we see that such an ; is
TI
necessarily time invariant. If we also take into account the spatial structure,
then it is shown in the exercises that the nonsingular commuting set is
; =
TI

f; 2 L(L2 ) : ; nonsingular, LTI and ;^ (s) = diag(^ 1 (s)I; : : : ; ^d (s)I )g:

In other words, at every value of s in C+ the matrix ;^ (s) must have the
block-diagonal structure of (8.4), but it is no longer restricted to be con-
stant. Thus the commuting set has grown as we reduced the perturbation
set from
to
. Notice also that the inverse ;;1 is automatically LTI,
or equivalently ;^ ;1 2 H1 .
a TI

The following result provides a summary of our robustness conditions so

far.
Theorem 8.17.
(M;
)  inf k;^ M^ ;^ ;1 k1  inf k;M^ ;;1 k1 = (M;
)
;2; ;2;
TI a
TI
252 8. Uncertain Systems
Proof . The rst inequality is a direct application of Proposition 8.16 to
the set
. The second inequality is clear since we are taking inmum
TI
over a smaller set. The third equality is a restatement of Theorem 8.12 and
is exactly Proposition 8.15. 
The immediate question is whether the rst inequality above is also an
equality: is there a counterpart of Theorem 8.12 for the time invariant
uncertainty case? Unfortunately we will nd that the answer is negative,
except in some specic cases. That is the inequality is usually strict. The
rest of the section will focus on this problem and pursue the analysis of

in more detail.
TI
To provide more generality, at this point we will extend the spatial
structure of our uncertainty class; consider the cone of complex matrices
C
s;f =
fdiag(1 Im1 ; : : : ; s Ims ;
s+1 ; : : : ;
s+f ) : k 2 C and
k 2 C mk mk g;
and denote its unit ball by
s;f = f
2 C
s;f :  (
)  1g. Now dene
the time invariant uncertainty set as

= f
2 L(L2) :
is LTI and
(
TI
^ s) 2
s;f ; for every Re(s)  0g:
In addition to the full blocks
k considered before, we are introducing
repeated scalar blocks of the form k I in the uncertainty. This means that
the same uncertainty operator acts on a number of scalar channels. One
motivation for such repetition is to describe the rational dependence of a
transfer function on an uncertain parameter, which can be expressed in
this way (see the exercises); another motivation will be seen at the end of
this chapter.
The next denition denes the matrix structured singular value, and is
entirely analogous to Denition 8.13. This denition will be central in our
reduction of LTI well-connectedness to a test on matrices.
Denition 8.18. Given a matrix Q 2 C mm we dene the structured
singular value of Q with respect to
s;f by
(Q;
s;f ) = minf (
) :
2 C
1and I ; Q
is singularg
s;f
when this minimum is dened, and (Q;
s;f ) = 0 otherwise.
The following properties of  with respect to the matrix set
s;f are left
as an exercise:
 (Q;
s;f ) = jj(Q;
s;f );
 (Q;
s;f ) = maxf(Q
) :
2
s;f g.
The second property implies that (Q;
s;f ) is a continuous function of Q.
This follows from the fact that the spectral radius function (
) is continu-
ous on the space of matrices, and the above maximum is over a compact set.
 8.4. Time invariant uncertainty 253
While we have not explicitly reviewed these topological facts in this course,
it is safe to assume they are known by readers studying this material.
The next major objective we have is to show how robustness analysis
over
can be converted into a matrix structured singular value test
TI

at each frequency. Before we can accomplish this we rst need one more
property of the matrix structured singular value, in addition to the ones
above, namely that it satises a maximum principle over matrix functions
that are analytic in a complex domain.
To obtain this result we need a few preliminary facts regarding polyno-
mials with complex variables. The rst is a continuity property of the roots
of a complex polynomial as a function of its coecients. In plain language
this result says that if all the coecients of two polynomials are sucient
close to each other, then their respective roots must also be near each other.
Lemma 8.19. Let p() = ( ; 1 )( ; 2 )


( ; n ) be an n-th order
polynomial over C ( 6= 0). If p(k) ( ) is a sequence of n-th order polyno-
mials with coecients converging to those of p( ), then there exists a root
1(k) of p(k) ( ) such that 1(k) ! 1 .
Proof . At each k, we dene the factorization
p(k) ( ) = (k) ( ; 1(k) )( ; 2(k) )


( ; n(k) )
such that 1(k) is the root closest to 1 (i.e. j1 ; 1(k) j  j1 ; i(k) j for every
i = 1 : : : n).
Since the coecients of p(k) ( ) converge to those of p( ), we can evaluate
at  = 1 and see that p(k) (1 ) ! p(1 ) = 0 as k ! 1. Also, (k) !  6= 0.
Therefore
n
Y
 (k)
 p (1 ) 

(k ) n
j ;  j  j ;  j = (k )   ;! 0 as k ! 1:
1 1 1 i
i=1
 (k) 


The next lemma concerns the roots of a polynomial of two complex
variables. It states that a particular type of root must always exist for such
polynomials, one in which the moduli of the arguments are equal.
Lemma 8.20. Let p(;  ) be a non-constant polynomial p : C 2 ! C . Dene
 = minfmax(j j; j j) : p(;  ) = 0g (8.20)
Then there exist  ? ;  ? such that p( ? ;  ? ) = 0 and j ? j = j ? j =  .
Proof . We start with ?;  ? that achieve the minimum in (8.20). The only
case we need to consider is when  > 0 and one of the above numbers has
magnitude less than  . Take, for instance, j ? j =  , j ? j <  .
254 8. Uncertain Systems
By setting  =  ? and grouping terms in the powers of  , we write
N
X
p(;  ? ) = an ( ? ) n : (8.21)
n=0
Suppose that the above polynomial in  is not identically zero. If we replace
 ? by (1 ; ) ? , its coecients are perturbed continuously, so we know from
Lemma 8.19 that as  ! 0+, the perturbed polynomial will have a root 
converging to  ? . Thus, for small enough , we have
j j < ; j(1 ; ) ? j < ; p( ; (1 ; ) ? ) = 0:
This contradicts the denition of  ; therefore the only alternative is that
the polynomial (8.21) must be identically zero in  . But then the value of
 can be set arbitrarily, in particular to have magnitude  , yielding a root
with the desired property. 
We are now in a position to state the maximum principle for the matrix
structured singular value. Just like the maximum modulus theorem for
scalar analytic functions, this theorem asserts that structured singular value
of a function that is analytic in the right half-plane achieves its supremum
on the boundary of the half-plane; we will actually treat the case of a
rational function.
Theorem 8.21. Let M^ (s) be a function in RH1, and
s;f be a
perturbation structure in the set of complex matrices. Then
sup (M^ (s);
s;f ) = sup (M^ (j!)
s;f ):
Re(s)0 !2R
Proof . We rst convert the problem to a maximization over the unit disk,
by means of the change of variables
s = 11 + :
;
This linear fractional transformation maps fj j  1;  6= 1g to fRe(s)  0g,
and the point  = 1 to s = 1. Also the boundary j j = 1 of the disk maps
to the imaginary axis.
Noticing that M^ (s) 2 RH1 has no poles in Re(s)  0 and has a nite
limit at s = 1, we conclude that the rational function
 
^ ^
Q( ) = M 1 ; 1 + 

has no poles over the closed unit disk j j  1. With this change, our theorem
reduces to showing that
0 := sup (Q^ ( );
s;f ) = sup (Q^ ( );
s;f ):
jj1 jj=1
 8.4. Time invariant uncertainty 255
Notice that by continuity of the matrix structured singular value, the supre-
mum 0 on the left is a maximum, achieved at some 0 , j0 j  1; we must
show it occurs at the disk boundary.
It suces to consider the case 0 > 0. By denition, we know that
1=0 =  (
0 ), where

0 := arg minf (
) :
2
s;f and I ; Q^ (0 )
is singular g:
Now consider the equation
det[I ; Q^ ( )
0 ] = 0
in the two complex variables  ,  . Since Q^ ( ) is rational with no poles in
j j  1, we can eliminate the common denominator and obtain a polynomial
equation
p(;  ) = 0
equivalent to the above for j j  1. We claim that
1 = minfmax(j j; j j) : p(;  ) = 0g:
In fact, the choice  = 0 ,  = 1 gives p(0 ; 1) = 0 and max(j0 j; 1) = 1: If
we found p(;  ) = 0 for some j j < 1, j j < 1, we would have
I ; Q^ ( )
0 singular, 
0 2
s;f ;  (
0 ) < 1=0:
This would give
(Q^ ( )
s;f ) > 0
contradicting the denition of 0 .
Thus we have proved our claim, which puts us in a position to apply
Lemma 8.20 and conclude there exists a root ( ? ;  ? ) of p(;  ) with j ? j =
j ? j = 1. Consequently,
I ; Q^ ( ? ) ?
0
is singular, with  ( ?
0 ) = 1=0 . So we conclude that, for j ? j = 1,
(Q^ ( ? )
s;f ) = 0 :

At this point we have assembled all the results required to prove the
following theorem, which is a major step in our analysis of time invariant
uncertainty. This theorem converts the well-connectedness of (M;
) TI
to a pure matrix test at each point in the frequency domain. It there-
fore has clear computational implication for the analysis of time invariant
uncertainty.
Theorem 8.22. Assume M is a time invariant bounded operator, with its
transfer function in RH1 . The following are equivalent:
256 8. Uncertain Systems
(a) (M;
) is robustly well-connected;
TI

(b) (M;
) < 1;

TI

(c) sup!2R (M^ (j!);
s;f ) < 1.

Furthermore, the left hand side quantities in parts (b) and (c) are equal.
Proof . It will be convenient to extend M^ to the compactied right half-
plane C+ [f1g, dening M^ (1) = D. Thus M^ (s) is a continuous function
on a compact set, and so is (M^ (s);
s;f ).
The fact that (b) implies (a) was already established in the previous
section.
To show that (a) implies (c), suppose that (c) does not hold, so
(M^ (j!0 );
s;f ) = sup (M^ (j!);
s;f )  1
! 2R
for some !0 2 R [ f1g. Therefore we can nd
0 2 C
s;f such that
 (
0 )  1 and I ; M^ (j!0 )
0 is singular. Choosing the constant LTI
^ s) =
0 , clearly it belongs to
and I ; M
is singular,
perturbation
( TI

contradicting condition (a).

To show that (c) implies (b), set
 = sup+ (M^ (s);
s;f ):
s2C
From condition (c), invoking Theorem 8.21, we know that  < 1. Fix any
scalar  satisfying 1 <  < ;1 . By the denition of the structured singular
value we have that
 < minf (
0 ) :
0 2 C
s;f and I ; M (s)
0 is singularg;
for all s in C+ [ f1g. Therefore I ; M^ (s)
0 is an invertible matrix for
every s in C+ [ f1g, and every matrix
0 2 C
s;f with  (
)   .
Since both s and
0 are varying over compact sets, we can uniformly
bound the norm of the matrix inverse, leading to the conclusion that
 (fI ; M^ (s)
0 g;1 ) is uniformly bounded.
Now for any time invariant operator
2 C
, with k
k   we conclude
TI

that I ; M
is a nonsingular operator. Therefore from Denition 8.13 we

nd that (M;
)   ;1 < 1.
Finally note that both (M;
) and sup!2R (M^ (j!);
s;f ) are
TI

TI

quantities which scale linearly with M . Thus if they were dierent, we

could re-scale M to make one of them less than 1 and the other greater
than 1, contradicting their equivalence with condition (a). Therefore they
must be equal. 
As a rst conclusion from the above theorem, we see that the test
(M;
) < 1 is necessary and sucient for robust well-connectedness.
TI
Recalling our general discussion in Section 8.3, we have avoided here the
 8.4. Time invariant uncertainty 257
diculty that was mentioned in establishing necessity. The reader should
be warned, however, that for this to hold we have allowed in the step \(a)
implies (c)" the use of a constant, complex perturbation
0 . This is legit-
imate since we have been working with complex signals and systems, but
it would not be allowed if we wished to constrain time domain signals to
be real-valued, as is often done; in this case the necessity direction will not
hold, as shown in the references at the end of the chapter.
As a second, important conclusion, we see that the test (M;
) < 1 TI
reduces exactly to a matrix structured singular value condition over the
frequency axis. Henceforth we concentrate exclusively on the latter problem
since it provides a direct avenue for computing well-connectedness.

8.4.2 The matrix structured singular value and its upper

bound
In this subsection we provide tools to examine and compute the structured
singular value of a matrix M . What we will nd is that in general it is not
amenable to computation by convex methods except in a number of special
cases.
As usual, we begin by considering the set of matrices ; which commute
with the perturbations, that is
;
=
; for all
2
s;f :
Restricting ourselves to positive denite matrices, the commuting set has
the form
;s;f =
fdiag(;1 ; : : : ; ;s ; s+1 I; : : : ; s+f I ) : ;k 2 H mk ; ;k > 0 and k > 0g;
where H mk denotes the set of Hermitian, mk  mk matrices. These full
blocks ;k appear in correspondence with the repeated scalar blocks k I .
We have the following result which provides an upper bound.
Proposition 8.23.
(a) (M;
s;f )  inf ;2;s;f  (;M ;;1 );
(b) A matrix ; 2 ;s;f satises  (; 21 M ;; 12 ) < 1 if and only if
M ;M ; ; < 0:
Clearly the above proposition is just a specialization to the matrix case of
our work in x8.3. An interesting point is however that in this case both con-
ditions ; 2 ;s;f and M ;M ; ; < 0 of (b) are Linear Matrix Inequalities
in the blocks of ;; therefore they lead directly to an ecient computation.
258 8. Uncertain Systems
Also by combining part (a) with Theorem 8.22, we obtain
(M;
) = sup (M^ (j!);
s;f )
TI
! 2R  
 sup ;!inf  ;! M^ (j!);;! 1
!2R 2;s;f
 inf k;^ M^ ;^ ;1 k1 :
;2; TI

In the rst inequality above, notice that we allow the scaling matrix ;!
to be frequency dependent. If in particular we choose this dependence to
be of the form ;^ (j!) where ; 2 ; , then we obtain the second inequality,
TI
which in fact re-derives the bound of Theorem 8.17 for this generalized
spatial structure. Now the intermediate bound appears to be potentially
sharper; in fact it is not dicult to show both bounds are equal and we
return to this issue in the next chapter.
Another important comment is that verication of the test
 
sup ; inf  ;! M^ (j!);;! 1 < 1
! 2R ! 2;s;f
amounts to a decoupled LMI problem over frequency, which is particularly
attractive for computation.
We still have not addressed, however, the conservatism of this bound. The
rest of the section is devoted to this issue; our approach will be to revisit
the method from convex analysis used in x8.2.2, which was based on the
language of quadratic forms, and see how far the analysis can be extended.
We will nd that most of the methods indeed have a counterpart here, but
the extension will fail at one crucial point; thus the bound will in general be
conservative. However this upper bound is equal to the structured singular
value in a number of cases and the tools we present here can be used to
demonstrate this.
We begin by characterizing the quadratic constraints equivalent to the
relation q =
p,
2
s;f . The rst part of the following lemma is matrix
version of Lemma 8.4; the second extends the method to repeated scalar
perturbations.
Lemma 8.24. Let p and q be two complex vectors. Then
(a) There exists a matrix
,  (
)  1 such that
p = q if and only if
p p ; q q  0:
(b) There exists a matrix I , jj  1, such that Ip = q if and only if
pp ; qq  0:
Proof . Assume p =6 0, otherwise the result is trivial.
 8.4. Time invariant uncertainty 259
(a) Clearly if
p = q,  (
)  1 then jpj  jqj or p p ; q q  0.
Conversely, if jpj  jqj we can choose the contractive rank one matrix


= jqp
pj2 :
(b) If p = q, jj  1, then pp ; qq = pp (1 ; jj2 )  0. Conversely, let
pp ; qq  0; this implies ker p  ker q and therefore Im q  Im p.
So necessarily there exists a complex scalar  which solves
q = p:
Now 0  pp ; qq = pp(1 ; jj2 ) implies jj  1.

Having established a quadratic characterization of the uncertainty
blocks, we now apply it to the structured singular value question.
Suppose I ; M
is singular, with
2
s;f . It will be more convenient
to work with I ;
M which is also singular (see Lemma 3.16). Let q 2 C m
be nonzero satisfying (I ;
M )q = 0, so q =
(Mq). Given the structure
of
, we can use the previous lemma to write quadratic constraints for the
components of q and Mq. We proceed analogously to x8.2.2. First, we write
these components as Ek q and Ek Mq where
 
Ek = 0


0 I 0


0 :
Next we dene the quadratic functions
k (q) =Ek Mqq M  Ek ; Ek qq Ek ;
k (q) =q M  Ek Ek Mq ; q Ek Ek q :
Finally we bring in the sets
rs;f := f(1 (q); : : : ; s (q); s+1 (q); : : : ; s+f (q)) : q 2 C m ; jqj = 1g
s;f := f(R1 ; : : : ; Rs ; rs+1 ; : : : ; rs+f ); Rk = Rk  0; rk  0g:
The latter are subsets of V = H m1 


 H ms  R 


 R, that is a real
vector space with the inner product
s
X sX
+f
hY; Ri = Tr(Yk Rk ) + yk rk :
k=1 k=s+1
The following characterization is the counterpart of Proposition 8.9.
Proposition 8.25. The following are equivalent:
(a) (M;
s;f ) < 1;
(b) the sets rs;f and s;f are disjoint.
260 8. Uncertain Systems
Proof . The sets rs;f and s;f intersect if and only if there exists q, with
jqj = 1 satisfying
k (q)  0; k = 1 : : : s;
k (q)  0; k = s + 1; : : : s + f:
Using Lemma 8.24, this happens if and only if there exist contractive k ,
k = 1; : : : ; s, and
k , k = s + 1; : : : ; s + f satisfying
k Ek Mq =Ek q; k = 1; : : : ; s

k Ek Mq =Ek q; k = s + 1; : : : ; s + f:
Putting these blocks together, the latter is equivalent to the existence of

2
s;f and q such that  (
)  1, jqj = 1 and

Mq = q
which is equivalent to (I ; M
) being singular for some
2
s;f ,  (
) 
1. By denition, this is the negation of (a). 
Having characterized the structured singular value in terms of properties
of the set rs;f , we now do the same with the upper bound of Proposition
8.23. Once again, the parallel with x8.3 carries through and we have the
following counterpart of Proposition 8.10.
Proposition 8.26. The following are equivalent:
(a) There exists ; 2 ;s;f satisfying  (;M ;;1) < 1;
(b) The convex hull co(rs;f ) is disjoint with s;f ;
(c) There exists a hyperplane in V which strictly separates rs;f and s;f ;
that is, there exists ; 2 V and ;  2 R such that
h;; Y i   <   h;; Ri for all Y 2 rs;f ; R 2 s;f : (8.22)
Proof . Notice that s;f is convex, so (b) implies (c) by the hyperplane
separation theorem in nite dimensional space. Conversely, if a hyperplane
strictly separates two sets it strictly separates their convex hulls, so (c)
implies (b). It remains to show that these conditions are equivalent to (a).
Starting from (a), we consider the LMI equivalent M  ;M ; ; < 0, for
some ; 2 ;s;f . For every vector q of jqj = 1, we have
q (M  ;M ; ;)q < 0:
Now we write
s
X sX
+f
;= Ek ;k Ek + k Ek Ek
k=1 k=s+1
 8.4. Time invariant uncertainty 261
that leads to the inequality
s
X
(q M  Ek ;k Ek Mq ; q Ek ;k Ek q) +
k=1
sX
+f
( k q M  Ek Ek Mq ; k q Ek Ek q) < 0
k=s+1
for jqj = 1. Taking a trace, we rewrite the inequality as
s
X
Tr (;k (Ek Mqq M  Ek ; Ek qq Ek )) +
k=1
sX
+f
k (q M  Ek Ek Mq ; q Ek Ek q) < 0
k=s+1
which we recognize as h;; Y i < 0, with
Y = (1 (q); : : : ; s (q); s+1 (q); : : : ; s+f (q)) and
; = (;1 ; : : : ; ;s ; s+1 ; : : : ; s+f )
Also since ;k > 0, k > 0 we conclude that
h;; Ri  0 for all R 2 s;f
so we have shown (8.22). The converse implication follows in a similar way
and is left as an exercise. 
The last two results have characterized the structured singular value and
its upper bound in terms of the sets rs;f and s;f ; if in particular the set
rs;f were convex, we would conclude that the bound is exact. This was
exactly the route we followed in x8.2.2. However in the matrix case the
set rs;f is not convex, except in very special cases. Recalling the proof of
Lemma 8.11, the key feature was the ability to shift in time, which has
no counterpart in the current situation. In fact, only for structures with
a small number of blocks can the bound be guaranteed to be exact. Such
structures are called -simple. We have the following classication.
Theorem 8.27.
(M;
s;f ) = inf  (;M ;;1 )
;2;s;f
holds for all matrices M if and only if the block speciers s and f satisfy
2s + f  3 :
For the alternatives (s; f ) 2 f(0; 1); (0; 2); (0; 3); (1; 0); (1; 1)g that satisfy
2s + f  3, the proof of suciency can be obtained using the tools intro-
duced in this section. In particular, an intricate study of the set rs;f is
262 8. Uncertain Systems
required in each case; details are provided in Appendix C. Counterexam-
ples exist for all cases with 2s + f > 3. We remark that the equality may
hold if M has special structure.
As an interesting application of the theorem for the case (s; f ) = (1; 1),
we invite the reader to prove the following discrete time version of the KYP
lemma. As a remark, in this case the I block appears due to the frequency
variable z , not to uncertainty; this indicates another application of the
structured singular value methods, which also extends to the consideration
of multi-dimensional systems, as will be discussed in Chapter 11.
Proposition 8.28. Given state space matrices A; B; C and D, the follow-
ing are equivalent.
(a) The eigenvalues of A are in the open unit disc and
sup  (C (I ; zA);1 zB + D) < 1 ;
z2D
(b) There exists a symmetric matrix ; > 0 such that
A B  ; 0 A B ; ; 0 <0:
      

C D 0 I C D 0 I
This completes our investigation of time invariant uncertainty and in
fact our study of models for uncertain systems. We now move to the next
chapter where we use our results and framework to investigate stability and
performance of uncertain feedback systems.

8.5 Exercises
1. This example is a follow-up to Exercise 6, Chapter 5 (with a slight
change in notation). As described there, a natural way to perturb a
system is to use coprime factor uncertainty of the form
P^
= (N^ +
^ N )(D^ +
^ D );1 :
Here P^0 = N^ D^ ;1 is a nominal model, expressed as a normalized
coprime factorization, and

^ 

=
^ N

D
is a perturbation with a given norm bound.
a) Find M such that algebraically we have P
= S(M;
). Is your
M always a bounded operator?
b) Repeat the questions of part a) to describe the closed loop map-
ping from d1 , d2 to q in the diagram given in Exc. 6, Chapter
5.
 8.5. Exercises 263
2. Consider the partitioned operators
   
N= N 11 N12 H11 H12
N21 N22 ; H = H21 H22 :
Find the operator M such that Figure 8.1 represents
a) The cascade of Figure 8.2, i.e.
S(M;
) = S(N;
1 )S(H;
2 ) with
= diag(
1 ;
2 ):
b) The composition S(M;
) = S(H; S(N;
)), where we assume
that I ; N22 H11 has a bounded inverse. Draw a diagram to
represent this composition.
c) The inverse S(M;
) = S(N;
);1 , where we assume that N22
has a bounded inverse.
3. Commutant sets.
a) Let ; 2 C nn . Prove that if ;
=
; for all
2 C nn , then
; = In , 2 C .
b) Let ; 2 C nn . Show that if ;
=
; for all structured
=
diag(
1 ; : : : ;
d ), then ; = diag( 1 I; : : : ; d I ), where i 2 C
and the identity matrices have appropriate dimensions.
c) Let ; 2 C nn . Show that if ;
=
; for all
2
s;f , then
; 2 ;s;f .
d) Characterize the commutant of
in L(L2 ) (i.e. the set ;
TI TI
of operators that commute with all members of
). TI

e) Characterize the commutant of
in L(L2 ).

a

4. Derive Proposition 8.15 from Theorem 8.12.

5. Consider the example of x8.4.1 (note all systems are SISO). It is
shown there that
(i) If
1 ,
2 are arbitrary operators, then kG1 kkG2k < 1 is
sucient for robust well-connectedness.
(ii) If
1 ,
2 are LTI operators, then kG1 G2 k1 < 1 is sucient for
robust well-connectedness.
Show that both of the above conditions are necessary as well as suf-
cient, by reducing them to the standard (M;
) form and applying
the theory of this chapter. What happens when only one of the
i is
known to be LTI?
6. Show that
a) (Q;
s;f ) = jj
s;f (Q);
b) (Q;
s;f ) = maxf(Q
) :
2
s;f g.
7. Rational functions in linear fractional form.
264 8. Uncertain Systems
a) Given the polynomial p() = pn n +


+ p1  + po , nd M such
that S(M; I ) = p() for an identity matrix of the appropriate
size.
b) Using Problem 2 above, show that a similar representation is
available for a rational function r() = pq(()) as long as q(0) 6= 0.
c) Generalize the above to matrices of rational functions, and to
rational functions of many variables 1 ; : : : d.
d) Apply to the following modeling problem. Consider the second
order transfer function
H^ (s) = s2 + s1 + !2
n
Assume the natural frequency !n is only known up to some error,
i.e. !n = !0 + k,  2 [;1; 1]. Find M^ (s) and the corresponding
dimension of the identity matrix, such that H^ (s) = S(M^ (s); I ):
8. Analysis with real parametric uncertainty. From the preceding exer-
cise we see that rational dependence of a system on a parameter
can be represented in our setup of Figure 8.1 by an appropriate
block I ,  2 R. This leads us to consider the mixed parametric/LTI
uncertainty structure

=
fdiag(1 I; : : : ; s I;
s+1 ; : : : ;
s+f ) 2
: k 2 [;1; 1]; 1  k  rg :
TI

In other words, the rst r blocks (r  s) in the general LTI structure

correspond to real parameters. For this structure it is possible to
generalize Theorem 8.22 and reduce the analysis over frequency to
the structured singular value test
sup (M^ (j!);
rs;f ) < 1;
! 2R
where the structure
rs;f  C mm is dened as

rs;f =
fdiag(1 I; : : : ; s I;
s+1 ; : : : ;
s+f ) 2
s;f : k 2 [;1; 1]; 1  k  rg :
We focus on a xed frequency !, let Q = M^ (j!). Since
rs;f 
s;f
we know that the LMI condition
Q ;Q ; ; < 0; ; 2 ;s;f
is sucient for (Q;
rs;f ) < 1. We wish to rene this bound by
exploiting the real blocks of
. Consider the additional matrix scaling
set
G = fG = diag(G1 ; : : : ; Gr ; 0; : : : ; 0) : Gk = Gk g  C mm :
 8.5. Exercises 265
a) Show that if q 2 ker(I ;
Q), with
2
rs;f , then
q (GQ ; Q G)q = 0 for all G 2 G :
b) Deduce that the LMI test in ; and G,
Q;Q ; ; < 0 + j (GQ ; QG) < 0; ; 2 ;s;f ; G 2 G ;
is sucient for (Q;
rs;f ) < 1.
c) Discuss how to generalize this procedure to study struc-
tures which combine real-parametric I blocks with arbitrary
operators
s+k as in x8.2.
9. Extension of maximum modulus principle for .
a) Extend Lemma 8.20 to polynomials of several complex variables.
b) Given a matrix M and a complex uncertainty structure
s;f ,
show that if

0 = arg
min
2

f (
) : I ; M
is singularg;
s;f
then without loss of generality
0 can be taken to be of the form
U , with  > 0 and U unitary.
c) Deduce that

s;f (M ) =
max
2
s;f
(M
) = max (MU ):
U 2
s;f ;U unitary
10. In this problem we will generalize the robust well-connectedness re-
sults of x8.2 by expanding the set
to include \diagonal" elements
a

similar to those we saw when studying the matrix structured singular

value. Dene
s;f to be the set
a

s;f := fdiag(1Im1 ; : : : ; sIms ;
s+1 ; : : : ;
s+f ) :

a

k 2 L(L12 );
k 2 L(Lm2 k ); and k
k < 1g:

Here the notation k Imk is shorthand for the operator diag(k ; : : : ; k )
which acts on Lm 2 ; note that k itself acts on the space of scalar
k
1
functions L2 . Realize that if these new types of blocks are not present
we have the usual denition of
from our work in x8.2.
a

a) Find the commutant set of this expanded perturbation set
s;f , a

and therefore the corresponding scaled small-gain theorem.

b) Here we will derive a generalization of Proposition 8.7, which
will relate well-connectedness to an infeasibility condition on a
new set of quadratic forms.
i) Suppose x1 ; x2 ; y1 ; y2 2 Rn , and that x1 and x2 are linearly
independent. Prove: there exists a matrix A 2 Rnn with
 (A)  1 such that
y1 = Ax1 and y2 = Ax2
266 8. Uncertain Systems
if and only if the matrix
   
y1 [y y ] ; x1 [x x ]  0:
y2 1 2 x2 1 2
ii) Suppose p1 ; p2 ; q1 ; q2 are scalar-valued functions in L12, and
that p1 and p2 are linearly independent. Show that there
exists an operator  2 L(L12 ) with kk  1 such that
q1 = p1 and q2 = p2
if and only if the matrix
Z 1   
q1 (t) [q (t) q (t)] ; p1 (t) [p (t) p (t)] dt  0:
0 q2 (t) 1 2 p2 (t) 1 2

iii) Now prove the generalization: given two Rm -valued func-

tions p; q 2 Lm 2 , there exists an operator  2 L(L2 ) with
1
kk  1 such that
q = Im p
if and only if the matrix
Z 1
q(t)q (t) ; p(t)p (t) dt  0:
0
Here there are no independence conditions as there were
above.
iv) Dene the matrices
 
Ek = 0


0 I 0


0 ; for 1  l  s + f;
which are dened so that sl=1
P +f 
El El = I , and we have

Ek diag(1 Im1 ; : : : ; s Ims ;
s+1 ; : : : ;
s+f )Ek =

k ; 1  k  s

k ; s + 1  k  s + f
for every
2
. Use these to dene the quadratic forms
a
k : L2 ! H mk and k : L2 ! R by
Z 1
k (p) = fEk q(t)q (t)Ek ; Ek p(t)p (t)Ek g dt;
0
for 1  k  s where q := Mp, and
Z 1
k (p) = fq (t)Ek Ek q(t) ; p (t)Ek Ek p(t)g dt;
0
for s +1  k  s + f . Prove that if there exists a nonzero p in
L2 so that k (p)  0 for 1  k  s, and k (p)  0 for s+1 
k  s + f , then (M;
s;f ) is not robustly well-connected.
a
Observe this is a generalization of Proposition 8.7.
 8.5. Exercises 267
c) Use the quadratic forms in (b, iv) to generalize the work in x8.2
to the set (M;
s;f ), and thus provide a more general version
a
of Theorem 8.12. The work in x8.4.2 will be helpful.
11. Suppose Q and R are state space operators in L(L2 ), that Q Q =
QQ = I , and that Q12 is surjective. Also assume that the system in
the gure below is well-connected.

- R
p q

Q

z   w

a) Prove the norm kRk  1 if and only if kS (Q; R)k  1. Consulting

the gure may be helpful.
b) Suppose A, B , C , D is a realization for the state space system
P , where  (D) < 1 and A is Hurwitz. Show that kP^ k1  1, if
and only if, kP^Q k1  1. Here PQ is the state space system with
realization
AQ = A + B (I ; D D);1 D C
BQ = B (I ; D D); 21
CQ = (I ; DD ); 12 C
DQ = 0
Hint: exploit part (a) by setting R^ = P^ and
   12 
Q^ (s) = (I ; DDD ) 21 (I ;;DDD) :

Notice: this question converts a question about a system with

a D-term to one without a D-term. This procedure is called
loop-shifting.
c) Prove that AQ in (b) must be Hurwitz if kP^ k1  1 and A is
Hurwitz.
d) Generalize these results to strict inequalities; that is systems
that are strictly contractive.
268 8. Uncertain Systems
Notes and references
The contemporary methods of robustness analysis are rooted in the work
of the 1960s on feedback stability for nonlinear systems from an input-
output perspective. In this category fall the small gain theorem [118, 151]
and methods based on passivity, particularly emphasized by the Russian
school where they were termed \absolute stability theory" [100, 146]. In
this literature we already nd \multiplier" methods for stability analysis
of various component nonlinearities which are direct precursors of the ;-
scalings considered in this chapter.
The observation that these methods could be used to study uncertain sys-
tems can be traced back to [151], but the subject was largely not pursued
until the late 70s when robustness began to be emphasized (see [115]) and
the connection to H1 norms was pointed out [152]. In particular the formal-
ization of uncertainty in terms of
blocks appeared in the papers [112, 23],
and [23] introduced the matrix structured singular value for studying time
invariant uncertainty. Subsequently a great deal of research activity was
devoted to its study, and to its extension to mixed real parametric and LTI
uncertainty [34]. Much of this research is summarized in [88, 90, 150]. The
observation that the test (M;
) < 1 can be conservative when dealing
TI
with real-valued signals can be found in [130]. Computing the structured
singular value for an arbitrary number of blocks appeared to be a vexing
task, and in [14, 131] it was demonstrated that the problem is NP-hard.
Given these diculties, it was welcome news when it was discovered in
the early 90s that convex upper bounds had an interpretation in their own
right. The rst result came from the parallel theory of robust control using
L1 signal norms, which we have not covered in this course (see [17], and
also [5] for structured singular value methods in this context). It was shown
in [70] that scaled small gain conditions in the L1 -induced norm, analogous
to Theorem 8.12 where exact for the analysis of arbitrary time-varying
operators. These results led to the study of the analogous question for the
case of L2 norms, leading to the proof of Theorem 8.12 in [80, 122] (see also
the related work [9]). The proof we presented is based on [82], which brought
in and extended the S-procedure [147] that had been developed in parallel
in the absolute stability literature. An extension of this viewpoint, which
makes IQCs [81] the central object, will be discussed in a later chapter. In
a sense, robust control has come full-circle to reinterpret the tools of input-
output stability analysis: instead of a known nonlinearity we are dealing
with an unknown ball of operators, and thus we obtain necessary as well
as sucient conditions.
In [99] these results are extended to show that the test with frequency
varying scalings, presented in Theorem 8.17, has a similar interpretation as
an exact test for the analysis of structured slowly varying perturbations.
The monograph [30] studies these structured robustness problems for hy-
 8.5. Exercises 269
brid sampled-data systems, and shows that many of them have unique
properties in this setting.
At the beginning of this chapter sets of systems were dened in terms of
explicit perturbations to a nominal plant. Another way to obtain a set of
systems is to put a topology on the space of all plants, and consider open
sets around the nominal system. This approach was initiated in [153] and
[135] where the gap metric and graph topology were dened, respectively.
For details on this research direction see for instance [44, 138] and the
references therein; at present these approaches do not explicitly consider
subsystem and component uncertainty.
 This is page 270
Printer: Opaque this

9
Feedback Control of Uncertain
Systems

In this chapter we bring together the separate threads of synthesis of feed-

back controllers in the absence of uncertainty, and analysis of uncertain
systems, into a common problem involving both uncertainty and control.
This problem is represented by the diagram shown in Figure 9.1, where G
is the generalized plant as in earlier chapters, but now also describes depen-
dence on system uncertainty. The perturbation
belongs to a structured
ball, and K represents the controller.

p q

z G w

y u
K

Figure 9.1. Feedback system with uncertainty

We will assume throughout this chapter that G and K are standard state
space systems, and they are therefore LTI and causal; nonetheless many of
 9. Feedback Control of Uncertain Systems 271
our results can be extended to an innite dimensional setting provided time
invariance and causality are maintained. Before proceeding any further, we
should note that feedback diagrams have been used in the preceding chap-
ters in two dierent settings. In Chapters 5 through 7, they represented the
feedback interconnection of causal, LTI, nite dimensional systems, moti-
vated by the modeling of an actual physical interconnection. In contrast,
in Chapter 8 we explored the possibilities of models which are not LTI, or
even completely specied in the sense that they are only known up to a per-
turbation set. In this chapter our goal is to combine these viewpoints, and
investigate both analysis and synthesis for closed-loop feedback systems
using the tools we have already developed.
We assume our system consists of a number of dierent interconnected
subsystems, and we need to specify the minimal desired properties of this
closed-loop setup. What we will require of the feedback system depicted in
Figure 9.1 is:
 Given any initial states for the state space systems G and K , and any
w 2 L2 [0; 1), there exist unique solutions in L2 for the functions
z; p; q; u and y;
 The states of G and K tend asymptotically to zero, from any initial
conditions, if w = 0;
 The ve maps on L2 from the input w to the functions z; p; q; u and
y are all bounded, assuming zero initial conditions for G and K .
Note that
is a bounded operator on L2 and therefore has no initial condi-
tions associated with it. When the feedback system satises these conditions
we will say it is stable. These conditions ensure that (a) the equations which
describe the feedback system always have meaning; (b) deviations in the
initial states of G and K away from equilibrium are innocuous; and (c) the
eect of w on the system is bounded.
Our rst objective is to nd conditions under which we can guarantee
that the system in Figure 9.1 is stable. Unless otherwise noted, we have
the following standing assumptions during this chapter:
 G is internally stabilized by K , in the sense of Chapter 5;

belongs to the class

= f
2
;
is causal g  L(L2)
c

where
is one of the uncertainty classes considered in the previous

chapter.
These assumptions deserve some comment, particularly with regard to the
causality of
. The rst of the assumptions says that K internally stabi-
lizes G. This means that the feedback system in Figure 9.1 is stable if

is set to zero, and we therefore call this nominal stability. Since we will
272 9. Feedback Control of Uncertain Systems
eventually treat
as a perturbation lying in a set which always contains
zero, this assumption is not restrictive. The causality of
is not as strongly
justiable, and at this point we make it for purely technical reasons which
will become apparent in x9.1.
To end this section we provide a sketch of the path we will take in
the chapter, where we will consider robust stability and performance of
the above feedback system. First, because of our assumption of nominal
stability, we can dene the interconnections of G and K by the star product
M = S (G; K );
which is an LTI system with transfer function in RH1 . Thus we can think
formally about our uncertain system as the conguration of Figure 9.2
below.

p q

M11 M12
z M21 M22 w
Figure 9.2. Robustness analysis setup

Now this setup loops exactly like the arrangement from Chapter 8. Indeed
our strategy is to study the robust stability of the overall system Figure
9.1 using the approach and results of Chapter 8.
We say the system of Figure 9.1 is robustly stable when it is stable for
every
2
. When robust stability holds we will consider questions of
c

performance, as usual measured in terms of keeping the mapping from w

to z \small". If I ; M11
is nonsingular we have
w 7! z = S(M;
) = M22 + M21
(I ; M11
);1 M12 ; (9.1)
which is a bounded operator on L2 . In this chapter we will take its induced
norm as measure of performance; for alternatives see Chapter 10. The sys-
tem of Figure 9.1 is said to have robust performance if it is robustly stable
and
kS(M;
)k < 1;
for every
2
. Notice that robust performance implies in particular
c

both robust stability and the nominal performance condition kM22k < 1; it
 9.1. Stability of feedback loops 273
should be clear, however, that robust performance is a stricter requirement
than the combination of these two.
At this point it is apparent our work in Chapter 8 will pay large divi-
dends provided that we can rigorously connect the equations which satisfy
the interconnection in Figure 9.1 to the simplied setup in Figure 9.2. Al-
though the diagrams appear to make this obvious, we must be careful to
ensure that the diagrams say the same thing as the equations which actu-
ally describe the system. We have the following theorem which provides us
with a guarantee of this.
Theorem 9.1. Suppose that
2 L(L2) is causal, and the assumption
of nominal stability holds. If (I ; M11
);1 exists as a bounded, causal
mapping on L2 , then the feedback system in Figure 9.1 is stable, and the
furthermore the mapping w to z is given by (9.1).
This theorem says that analysis of Figure 9.2 provides us with guarantees
for both stability and performance of the system in Figure 9.1, assuming
that (I ; M11
);1 is causal in addition to being in L(L2 ). In the following
section we will prove this theorem and in doing so the motivation behind
the causality constraint will become clear.
To prove Theorem 9.1 rigorously will require some additional prepara-
tion, and this proof is the main topic of the next section. Subsequently,
once we have established Theorem 9.1, we will proceed in x9.2 to applying
the results of Chapter 8 directly to analysis of robust stability and per-
formance of feedback systems. Finally in x9.3 we discuss the synthesis of
feedback controllers for robust performance.

9.1 Stability of feedback loops

In this section we will prove Theorem 9.1 and in fact show stability under
more general circumstances than we have specied above. The results of
the section apply specically when G and K are nite dimensional state
space systems, however with appropriate modications they hold for more
general G and K .
Our rst observation is that the state space plant G does not necessarily
dene a bounded LTI operator on L2 , nor does the state space controller K .
That is their transfer functions G^ (s) and K^ (s) are not necessarily in RH1 .
Indeed one of the purposes of K is to internally stabilize G. To be precise,
given a state space system R with realization (A; B; C; D) we dene its
eect on an input u from L2 [0; 1) by
Z t
(Ru)(t) = C eA(t; )Bu( ) d + Du(t):
0
So if A is not Hurwitz, this does not dene a bounded mapping on L2
assuming that (A; B; C ) is a minimal realization. However it is routine to
274 9. Feedback Control of Uncertain Systems
verify that, for any T > 0, the inequality
Z T
j(Ru)(t)j22 dt < 1 is satised. (9.2)
0
So although Ru may not be an element of L2 , the energy of (Ru)(t) is
bounded on any nite interval.

9.1.1 L2 -extended and stability guarantees

This observation motivates the denition1 of the space of functions L2 -
extended denoted by L2e . Dene
L2e = f w : [0; 1) ! C n satisfying PT w 2 L2 [0; 1), for all T > 0g:
Recall that PT is the mapping which truncates the support of any function
to the interval [0; T ], thus this is just the space of functions whose energy
is nite on any nite interval. The space L2e is a vector space by pointwise
addition, where we equate two elements v and g if they satisfy kPT (v ;
g)k2 = 0, for all T  0. Thus it contains L2 as a subspace, however L2e
itself is not a normed space.
Now it is routine to verify that any state space system such as R dened
above satises
R : L2e ! L2e :
Namely any state space system denes a mapping on L2e . Therefore G and
K will always dene maps on L2e . Notice also that any such state space
system is causal since
PT R = PT RPT ; for all T  0.
We make some observations about linear mappings on L2e . Since L2e is a
vector space the linear mappings on it are familiar, and in particular we
can dene the inverse of a linear mapping on L2e in the usual way. We say
that a mapping on L2e is bounded if it denes a bounded linear mapping
when restricted to L2 .
We can now use these notions to directly address the stability of Fig-
ure 9.1. First we consider the uncertain system shown in Figure 9.3. It
shows G and
without the controller. Now this creates the model for the
uncertain system we aim to control. Notice however that if G is an unstable
state space system we cannot assume that p is an L2 signal. Thus we need
to assume for now that
is a well-dened linear mapping on L2e which is
bounded.
We are now ready to prove a version Theorem 9.1. The important in-
gredient we now have in our analysis is that we will be able to postulate
1
this space is also called L -local, but the term L -extended is more apt here.
2 2
 9.1. Stability of feedback loops 275

p q

G
z w
Figure 9.3. Uncertain plant
solutions to our closed-loop systems equations over L2e , which is a set con-
taining both the domain and codomain of G and K . If we instead attempted
to prove uniqueness of solutions to the equations governing Figure 9.1 over
just L2 , we could not rule out the existence of solutions in L2e . Namely the
fact that the solutions p, q, u, y and z in Figure 9.1 are in L2 , when w is
in L2, should be a conclusion of our analysis not an a priori assumption.
We have the following result, and remark that if a system satises (i){(iii)
we say it is L2e -stable.
Theorem 9.2. Pertaining to Figure 9.1, suppose
(a) That
is a linear mapping on L2e and is bounded on L2;
(b) The state space system G is internally stabilized by the state space
controller K ;
(c) The inverse (I ; M11
);1 exists as a mapping on L2e and is bounded
on the normed space L2.
Then
(i) Given any w in L2e and initial states of G and K , there exist unique
solutions in L2e for p, q, u, y and z . Furthermore if w is in L2 , so
are the other ve functions;
(ii) The states of G and K tend asymptotically to zero, from any initial
condition, when w = 0;
(iii) The maps w to p, q, u, y and z are all bounded on L2 , when the initial
states of G and K are zero. Furthermore, if the latter condition holds
the mapping w to z is given by (9.1).
This result diers from Theorem 9.1 in two main respects. One: the inverse
(I ; M11
);1 need not be causal, it simply needs to be dened on all of L2e
and bounded on L2 . Two: here we look for solutions to our loop equations
without assuming up front that they necessarily lie in L2 ; if w is chosen in
276 9. Feedback Control of Uncertain Systems
L2 then we see that the solutions to loop equations necessarily lie in L2 . In
x9.1.2 we will see that Theorem 9.1 is a special case of the current theorem.
Proof . We rst prove (i). Choose w in L2e and an initial state
(x(0); xK (0) ) for the state space systems G and K . Now the equations
governing this system are the state space equations for G and K , and the
operator equation q =
p. We initially suppose that solutions for p, q, u,
y and z exist in L2e , and we will show that they are unique. Now by rou-
tine algebraic manipulations we can write the equations that govern the
feedback loops of this system as
q =
p
x_ L = AL xL + BL1 q + BL2 w; where xL (0) = (x(0); xK (0) )
p = CL1 xL + DL11 q + DL12 w:
Here AL , BL , CL and DL is the state space representation for the inter-
connection of G and K , and thus we know from hypothesis (b) that AL is
Hurwitz. Clearly this is a state space realization for the operator M . From
this we see that
p = M11q + M12 w + CL0 A~L xL (0);
where A~L : C n ! L2 via the relationship A~L xL (0) = eAL t xL (0). Now
applying assumption (c) above we have that
p = (I ; M11
);1 M12 w + (I ; M11
);1 CL0 A~L xL (0): (9.3)
Thus we see that p must be uniquely determined since w and xL (0) are
xed, and then the function q is uniquely specied by p. Since the inputs
q, w, and the initial conditions of the state space systems G and K are all
specied, we know from (b) and the work in Chapter 5 that u, y and z are
unique and in L2e .
To complete our demonstration of (i) we must show the existence of
solutions. To see this simply start by dening p from (9.3), given any w 2
L2e and initial condition xL (0), and essentially reverse the above argument.
Also if w is in L2 then by the boundedness of M , (I ; M11
);1 , and A~L
we see that p must be in L2 . Immediately it follows that q must be in L2
since
is bounded, and again by the nominal stability in (b) we have now
that the other functions are also in L2.
We now prove (ii). From above we know that if w = 0 we have
p = (I ; M11
);1 CL0 A~L xL (0);
and so q =
p is in L2 since AL is Hurwitz and
is bounded. Continuing
under these conditions we have
x_ L = AL xL + BL0 q; with some initial condition xL (0).
It is not dicult to show that xL (t) ! 0 as t ! 1, since q is in L2.
 9.1. Stability of feedback loops 277
Finally we need to show that the maps from w to the other functions are
bounded on L2 . From above we already know that the maps from w to p
and q are bounded on L2 . This means that the map from w to the state
xL must be bounded on L2 since AL is Hurwitz. Finally the functions z , u
and y are all simply given by matrices times w, xL and q and so the maps
from w to them must also be bounded on L2.
To end, the mapping w to z is given by (9.1) follows routinely from the
above formula for p, and the state space denition of the operator M . 
The theorem just proved gives us a set of sucient conditions for L2e
stability where the mapping
is dened on L2e and is bounded. Note that
it is not necessarily causal. We would however like to dene our perturba-
tions from the operators on L2 in keeping with our work in Chapter 8, and
we now address this topic.

9.1.2 Causality and maps on L2 -extended

From our discussion so far it is clear that any bounded mapping on L2e
denes an element of L(L2 ) simply by restricting its domain. We would
now like to examine the converse problem. Suppose that Q 2 L(L2 ), then
Q is not dened on all of L2e and it is not necessarily clear that the domain
of Q can be extended to L2e while maintaining the linearity of Q. Consider
the following example.
Example:
Dene the convolution operator Q on L2 by
Z 1
(Qu)(t) = e(t; )u( )d;
t
which can be shown to have an L2-induced norm of one. However we see
that it is not possible to make this meaningful for some inputs in L2e ; for
instance set u( ) = e which yields a diverging integral, for every t  0.

The problem with the above example can be attributed to the non causal-
ity of Q. In contrast, we now see that causal operators on L2 can always
be extended to L2e . To this end observe that given any causal operator
Q 2 L(L2 ), we can dene Qu = v, for u 2 L2e , by
PT v = PT Qu := PT QPT u;
for each T . We leave it to the reader to verify that this provides a valid de-
nition, namely that if PT1 v and PT2 v are dened as above for T1 < T2 , then
PT1 PT2 v = PT1 v; this follows from the causality of Q in L2 . Summarizing,
we have:
278 9. Feedback Control of Uncertain Systems
Proposition 9.3. Suppose Q is a causal operator in L(L2). Then Q
denes a linear causal mapping on L2e , which is bounded on L2 .
We are now in a position to see why Theorem 9.1 is a special case of
Theorem 9.2. In Theorem 9.1 we assume that both
and (I ; M11
);1
are causal and in L(L2 ). Thus by the above proposition these assumptions
imply that both these operators dene bounded mapping on L2e . This is
exactly what is required in conditions (a) and (c) of Theorem 9.2. Both
theorems assume that G is stabilized by K , and therefore the hypothesis
of Theorem 9.2 is less restrictive. However the conclusions (i){(iii) of this
theorem are stronger than that of Theorem 9.1, since it guarantees L2e -
stability not just stability. Hence Theorem 9.1 is a special case.
To complete our discussion it is natural to inquire whether non-causality
always prevents us from extending the domain to of denition to L2e . In
other words is existence of an extension to L2e any more general than
causality? the following example shows the answer is armative.
Example:
Let us dene the mapping Q on L2 by
(Qu)(t) = u(2t); for t  0:
Clearly this mapping is linear. It is straightforward to verify that 2kQuk2 =
kuk2, for any u in L2 , and so Q is a bounded operator with an induced
norm of one-half. This operator is anti causal since the value of (Qu)(t) is
equal to the value of u at time 2t in the future. From the above formula
for Q we see that it can be immediately extended to a linear mapping on
L2e . 
Thus we see that our results of Theorem 9.2 are strictly more general
than those requiring causality; nevertheless, the latter condition is the most
common means of extending operators to L2e ; accordingly, in the rest of
this chapter we will restrict
to be causal.
We now brie y remark on the implications of the causality restriction
from the point of view of uncertainty modeling. The usual argument in
favor of causal models is physical realizability; in this regard, it is natural
that a component such as K , which directly models a physical controller,
be taken as causal. However the situation is less obvious in the case of

, which often does not directly model a physical system, but instead is
a part of a parametrization of a relation (e.g. Ra ) which is the actual
model of a physical component. In these cases the main reason to choose
a causal model is mathematical convenience: this allows us to isolate
as
though it were a component, and in this way to standardize a large class
of uncertainty models into a single prototypical form. While this might
restrict somewhat our modeling possibilities, experience seems to indicate
that the restriction is not severe. Furthermore in the next section we will
 9.1. Stability of feedback loops 279
see that the restriction to causal operators has no eect on the robustness
conditions we are able to obtain.
A nal question in regard to causality is whether it is possible for the
inverse (I ; M11
);1 to be non causal, even when M11 and
are both
causal. An example shows us that this is indeed possible2 .
Example:
First recall the anti causal operator Q dened in the last example. Its
inverse is
(Q;1 u)(t) = u(t=2); for t  0:
This mapping is linear and causal; it is straightforward to verify that
kQ;1uk2 = 2kuk2, for any u in L2 , and so Q;1 is a bounded operator.
We can now construct our counter example to (I ;M11
);1 being causal.
Set
M11 = I and
= I ; Q;1 ;
and we see (I ; M11
);1 = Q which is anti causal as discussed earlier. 
Despite this counterexample, we will now see that the causality of (I ;
M11
);1 can be assured if
is causal and a spectral radius condition
is satised. This result is signicant for our robustness analysis because it
shows that if (I ; M11
);1 is not causal, then there exists a complex scalar
jj  1, such that I ; M11
is singular. In words, if the above inverse is
not causal, then there exists a smaller perturbation of the same structure
which causes our well-connectedness condition to fail. Now for the result.
Proposition 9.4. Let M11 and
be causal operators in L(L2). If
rad(M11
) < 1, then (I ; M11
);1 exists in L(L2 ) and is causal.
Proof . Clearly I ; M11
is invertible since its spectrum is by assumption
inside the unit disc. Also by this assumption we know from Chapter 3 that
its inverse has a series expansion
1
X
(I ; M11
);1 = (M11
)k :
k=0
Now since M ,
, are causal, so is (M
)k for every k, and thus also the
series which gives (I ; M
);1 . 
Having clearly established the relationship between feedback systems in-
volving uncertainty and the tools of Chapter 8, we are now ready to analyze
stability and performance in the feedback arrangement of Figure 9.1.

2
Such examples are impossible in discrete time if starting at time zero.
280 9. Feedback Control of Uncertain Systems
9.2 Robust stability and performance
We are now ready to confront robustness for the general feedback control
setup of Figure 9.1. In this section we will concentrate on the analysis of
robust stability and performance for a given controller K .
We start with the following denition which can be interpreted in terms
of Figure 9.2. With respect to a causal set
of contractive perturbations
c

and a causal operator M , both residing in L(L2 ), we say:

 The uncertain system (M11 ;
) is robustly stable if (I ; M11
);1
c

exists in L(L2 ) and is causal, for each
in
;

c

 The uncertain system (M;
) has robust performance if (M11 ;
)

c c

is robustly stable and kS(M;
)k  1, for every
2
.c

Stability here is identical to our denition of well-connectedness in Chap-

ter 8, except that now we demand that the relevant inverse be causal. We
dene robust performance to be stability plus contractiveness of the closed-
loop, which again is identical to the performance criterion we imposed in
Chapter 8.
Having made this denition we have the following result which is a di-
rect consequence of our work in x9.1. It states that if (M;
) has robust
c

stability or performance, then the feedback system of Figure 9.1 will also
enjoy the respective property.
Theorem 9.5. Suppose K internally stabilizes G. If the uncertain system
(M11 ;
) has robust stability, then the feedback system described by Fig-
c

ure 9.1 is robustly stable. Furthermore, if (M;
) has robust performance,

c

then the system of Figure 9.1 has robust performance.

This result follows immediately from Theorem 9.1, and gives sucient con-
ditions for robustness properties to hold for Figure 9.1 in terms of the
robustness properties of (M;
). We note that by Theorem 9.2 robust
c

stability of (M11 ;
) actually implies the stronger condition of robust

c

L2e -stability of the feedback system.

Remark 9.6. We remark here that if our denition of stability for Fig-
ure 9.1 is modied so that the system be tolerant to disturbances injected
directly into the loop between G and
, then robust stability of (M11 ;
)
c

would be a both necessary and sucient condition for robust stability of

Figure 9.1. However in many cases such an added condition is not natural,
and in this sense our conditions may be conservative.
Having established Theorem 9.5 we see that by concentrating on the
robustness properties of (M;
) we will guarantee the corresponding prop-
c

erties in the feedback conguration of Figure 9.1. Henceforth in this section
we will focus our eort on the uncertain system (M;
). The remainder
c

of this section is devoted to methods of analyzing robust stability and per-

 9.2. Robust stability and performance 281
formance. As we will see, the results will be completely parallel to those on
robust well-connectedness in the previous chapter.

9.2.1 Robust stability under arbitrary structured uncertainty

We rst consider the uncertainty structure

= fdiag(
1 ; : : : ;
d) :
k 2 L(L2 );
k causal ; k
k k  1g;
a;c

which is obtained by imposing causality on our class
from Chapter

a

8. We will assume nominal stability, so M in Figure 9.2 will be a causal,

bounded LTI system.
The main result, stated below, tells us that robust stability can be studied
analogously to the corresponding robust well-connectedness problem. The
set ; is the commutant set corresponding to
, and is used here with the
a
new set
. Also P ; is the positive subset of ;.
a;c

Theorem 9.7. Suppose that M is a causal, bounded LTI operator. Then

the following are equivalent:
(i) The uncertain system (M11 ;
) is robustly stable;
a;c

(ii) The inequality inf k;M11;;1 k < 1 holds.

;2;
Proof . The suciency direction (ii) implies (i), is essentially identical to
our work in Chapter 8. From the small-gain condition and the fact that
; commutes with
we have 1 > k;M11;;1 k  rad(;M11 ;;1
) =
rad(M11
). Thus I ; M11
must be nonsingular, and (I ; M11
);1 is
causal by virtue of Proposition 9.4.
For the necessity direction (i) implies (ii), we can take advantage of
most of our work in x8.2.2. Referring back to this section, we see that the
following step, analogous to Proposition 8.9, is all that is required.
Proposition 9.8. Suppose that (M11;
) is robustly stable. Then the
a;c
sets  and r are strictly separated, i.e.
d(; r) := r2inf;y2r jr ; yj > 0:
In fact, given this proposition we can then follow the rest of x8.2.2, leading
to condition (ii). Proposition 9.8 stated above is proved in Appendix B. 
We now turn to time invariant uncertainty.

9.2.2 Robust stability under LTI uncertainty

For the case of LTI uncertainty we recall the denition of the structure

from Chapter 8.
TI

= f
= diag(
1 : : : ;
d) 2 L(L2 ) :
time invariant and k
k  1g:

TI
282 9. Feedback Control of Uncertain Systems
Now each
in
is LTI on the space of functions L2[0; 1). By Corol-
TI
lary 3.27 we see that this immediately means that
must be causal. That
is, our LTI uncertainty set is already comprised of operators that are causal.
Having made this observation we can state the following theorem.
Proposition 9.9. Let M be a causal, bounded LTI system with transfer
function M^ 2 RH1 . Then the following are equivalent:
(i) The uncertain system (M11 ;
) is robustly stable;
TI

(ii) The structured singular value condition sup (M^ 11 (j!);
s;f ) < 1 holds.
!2R
Proof . By Theorem 8.22 it is sucient to show that robust well-
connectedness of (M11 ;
) is equivalent to robust stability. Since by
TI
denition the latter condition implies the former we need only prove the
converse.
Thus we need to show that if I ; M11
is nonsingular, where
2
,
then (I ; M11
);1 is causal. This follows directly from frequency domain
TI

analysis. The nonsingularity condition implies that (I ; M^ (s)
(

^ s) );1 is in
H1 , and thus by Corollary 3.27 we know it represents a causal operator.

Thus we see that if M is causal, robust stability of (M11 ;
) is TI

equivalent to robust well-connectedness; they both amount to a complex

structured singular value condition parametrized over frequency. Let us
move forward to robust performance.

9.2.3 Robust Performance Analysis

Having characterized robust stability, we now turn to the robust perfor-
mance problem for the conguration of Figure 9.2, with the performance
measured by the induced norm of the map from w to z .
In Chapter 8 we explored a similar issue in the context of abstract L2 op-
erators, and saw how a restriction on the norm of S(M;
) could be turned
into a robust well-connectedness problem with an augmented uncertainty
structure. We now see that the procedure carries through in the context of
causal uncertainties for the classes treated in the preceding sections.
Proposition 9.10. Suppose M is a causal LTI operator on L2[0; 1).
Dene the perturbation set
  

u 0

= 0
p :
u 2
;
p 2 L(L2) causal and k
pk  1 :
c
a;p a;c

The following are equivalent:

(a) The uncertain system (M;
) satises robust performance: it is
robustly stable and kS(M;
u )k < 1, for every
u 2
.
a;c

a;c
 9.2. Robust stability and performance 283
(b) The uncertain system (M;
) is robustly stable;
c
a;p

(c) There exists ; 2 ; such that

   ;1

; 0 ; 0

0 I
M 0 I < 1: (9.4)
Proof . The fact that (a) implies (b) follows analogously to Proposition 8.3
in Chapter 8. Here we have by hypothesis that (I ; M11
u ) has a causal,
bounded inverse, and the identity (8.2) can be written:
 
I ; M11
u ;M12
p =
;M21
u I ; M22
p
  
I 0 I ; M11
u ;M12
p 
;M21
u (I ; M11
u );1 I 0 I ; S(M;
u )
p
By Proposition 9.4, the hypothesis implies that I ; S(M;
u )
p has a
causal, bounded inverse, so the same happens with I ; M
.
The step from (b) to (c) is a direct application of Theorem 9.7, once we
recognize that the scaling set corresponding to
is c
a;p
  
; 0
;p = 0 pI : ; 2 ;; p 6= 0
and we realize that p can be normalized to one without loss of generality.
Finally to prove (c) to (a), rst note that the top-left block of (9.4)
implies k;M11 ;;1 k < 1 and hence robust stability. Next, we rewrite (9.4)
in operator inequality form
   
 ; 0 ;
:= M 0 I M ; 0 (1 ; )I < 0;
0 0 0

where ;0 = ; ; and by continuity we may introduce a suitably small  > 0

in the bottom block.
Applying the signals q and w from Figure 9.2, we have
   
0 q q
w ; w
           
= pz ; ;00 I0 pz ; wq ; ;00 (1 ;0 )I wq
d
X
= k (kpk k2 ; kqk k2 ) + kz k2 ; (1 ; )kwk2
k=1
 kz k2 ; (1 ; )kwk2 :
In the last step, the fact that qk =
k pk and k
k k  1 was used. The
nal inequality says that the map from w to z has induced norm less than
one. 
We now give an analogous result for the time invariant case.
284 9. Feedback Control of Uncertain Systems
Proposition 9.11. Suppose the conguration of Figure 9.2 is nominally
stable, and the uncertainty class is
. Dene the perturbation set
TI
  

u 0

= 0
p :
u 2
;
p 2 L(L2) LTI, causal, k
pk  1 :
T I;p TI

^ s) takes values in
s;f +1 for any
2

Notice that
( c
(i.e., there
is one extra full block with respect to the uncertainty structure
s;f ). The
T I;p

following are equivalent:

(a) The uncertain system (M;
) satises robust performance: it is
robustly stable and kS(M;
u )k < 1, for every
u 2
.
TI

TI

(b) The uncertain system (M;
) is robustly stable;

T I;p

(c)
sup (M^ (j!);
s;f +1 ) < 1: (9.5)
!2R
The proof follows along similar lines as that of Proposition 9.10 and is left
as an exercise. Notice that here the so-called performance block
p can be
taken to be LTI.
This concludes our work on analysis of robust stability and performance
in causal feedback systems. We now turn our attention to the synthesis of
controllers.

9.3 Robust Controller Synthesis

In this section we tackle the problem of synthesizing feedback controllers
for robust performance in the presence of uncertainty. In other words,
going back to our original setup of Figure 9.1 our objective is to design
the controller K as in earlier chapters, however we are no longer satised
with nominal stability and performance, but wish these properties to be
maintained throughout the uncertainty class
. c

As in previous sections, we will concentrate here on the L2 -induced norm

from w to z as a measure of performance. The advantage of this choice is
that robust performance can be studied with completely analogous methods
as robust stability. We thus focus our discussion on the robust stabilization
problem: given an LTI state space system G, and an uncertainty class
, c

nd a controller K such that

 K internally stabilizes G;
 The uncertain system (M;
) is robustly stable, where M :=
c

S (G; K ).
We will only consider perturbation structures for which we have developed
the necessary analysis tools. In this way our problem turns into the search
 9.3. Robust Controller Synthesis 285
for a controller K that satises a precise mathematical condition. The
simplest case is when we have unstructured uncertainty, i.e. when

= f
2 L(L2);
causal ; k
k  1g :
In this case a controller K will be robustly stabilizing if and only if it
satises
kS (G; K )k1 < 1;
this small-gain property follows, for instance, from our more general result
of Theorem 9.7. So our problem reduces to H1 synthesis; in fact, this
observation is the main motivation behind H1 control as a design method,
completing the discussion which was postponed from Chapter 7.
What this method does not consider is uncertainty structure, which as
we have seen can arise in two ways:
 Uncertainty models derived from interconnection of more simple
component structures.
 Performance block, added to account for a performance specication
in addition to robust stability.
The remainder of this chapter is dedicated to the robust synthesis problem
under the structured uncertainty classes
and
which we have
a;c TI
studied in detail.

9.3.1 Robust synthesis against
a;c

We begin our discussion with the class
. Let K be the set of state
a;c
space controllers K which internally stabilize G. Then robust synthesis is
reduced, via Theorem 9.7, to the optimization problem
inf k; 21 S (G; K );; 12 k;
;2P ;;K 2K
which is H1 synthesis under constant scaling matrices; we have constrained
the latter to the positive set P ;, which is a slight change from Theorem
9.7 but clearly inconsequential.
Robust stabilization is achieved if and only if the above inmum is less
than one, since the set
contains only contractive operators. We there-
a;c
fore have a recipe to tackle the problem. The main question is whether the
above optimization can be reduced to a tractable computational method.
Note that we have already considered and answered two restrictions of this
problem:
 For xed ; we have H1 synthesis;
 For xed K we have robustness analysis over
. a;c
286 9. Feedback Control of Uncertain Systems
As seen in previous chapters, both these subproblems can be reduced to
LMI computation, and the new challenge is the joint search over the matrix
variable ; and the system variable K .
The rst result is encouraging. It says that we can restrict our search for
a controller to those which have the same dynamic order as the plant.
Proposition 9.12. Let G be an LTI system of order n. If
inf k; 21 S (G; K );; 12 k < 1; (9.6)
;2P ;;K 2K
then there exits a controller K 2 K of order at most n, such that
k; 21 S (G; K );; 12 k < 1:
The proposition states that if the the robust stabilization problem can be
solved, then it can be solved using a controller K of state dimension no
greater than the plant G. Figure 9.4 illustrates the constructions used in
the proof.
G;

; 21 ;; 12
G

K
Figure 9.4. Scaled synthesis problem

Proof . Consider ;, K satisfying (9.6). Dening

 1   
;1
G; := ;02 0I G ; 0 2 I0
(see Figure 9.4) we have
; 12 S (G; K );; 21 = S (G; ; K ):
Also G; has order n, so from Chapter 7 we know that the condition
kS (G; ; K )k < 1
can be achieved with K of order no greater than n. 
Therefore we see that for purposes of robust synthesis over
, we can
a;c
conne our attention to controllers of the order of the generalized plant
 9.3. Robust Controller Synthesis 287
G. In particular, our optimization problem is nite, in the sense that it
involves a nite set of variables, ; and the controller state-space matrices
AK , BK , CK , DK of a xed order. This is substantially simpler than the
problem we initially started with, namely arbitrary state space controllers
and uncertainties in an abstract operator space. One could search over this
nite number of variables with a variety of computational tools.
Unfortunately this still does not mean that the problem is compu-
tationally tractable by the standards we have employed in this course,
namely problems which can be globally optimized in a moderate amount
of computational time, such as LMI optimization problems. Can a further
simplication be achieved? To explore this issue it is natural to combine
our robust stability condition from x9.2.1, which resulted from robustness
analysis, with our LMI method for H1 synthesis described in Chapter 7.
This is done next.
We begin with a state-space realization for the generalized plant
2 3
A B1 B2
G^ (s) = 4 C1 D11 D12 5 ;
C2 D21 0
and incorporate the scalings ; 12 and ;; 12 to obtain
2
A B1 ;; 12 B2 3
1 1 1 1
G^ ; (s) = 4 ; 2 C1 ; 2 D11 ;; 2 ; 2 D12 5 :
C2 D21 ;; 12 0
Next, we apply Theorem 7.9 to obtain necessary and sucient conditions
for the existence of a controller satisfying kS (G; ; K )k1 < 1. These take
the form

XA XB1 ;; 21 C1 1 ; 12 1 N 0
2 3
NX 0  4A;;X21 +
 

0 I BX
12 1 12
;I 1 ;; 2 D11 ; 2 5 0X I < 0 ;
; C1 ; D11 ;; 2 ;I
(9.7)
Y C1 ; 21 B ;; 12 
2  3
NY 0  4AY; 21+CYYA
  
12 1 ; 12 5 NY 0
0 I 1 ;I 1 ; D11 ; 0 I < 0 (9.8)
;; 21 B1 ;; 12 D11
 ;2 ;I
 
X I  0 (9.9)
I Y
where NX and NY are full-rank matrices whose images satisfy
 
ImNX = ker C2 D21 ;; 21 ;

ImNY = ker B2 D12  ; 21  :
288 9. Feedback Control of Uncertain Systems
So we see that the robust stabilization problem is solvable if and only if
conditions (9.7), (9.8) and (9.9) in the variables X , Y and ; are satised.
This once again emphasizes the nite dimensionality of the problem.
To get these conditions into a more transparent form, it is useful to
redene the outer multiplication factors so that they are independent of ;.
Dene NX , NY to be full rank matrices whose images satisfy
 
ImNX = ker C2 D21 ;
 :
ImNY = ker B2 D12
Then NX , NY are constant, and clearly we can take
 
I 0
NX = 0 ; 12 NX ; (9.10)
 
NY = I0 ;;0 12 NY : (9.11)
Substituting with (9.10) into (9.7) gives
C1 ; 211 N 0
2  3
NX 0  4A XB +XXA XB
  1
0 I 121
;; D11 ; 2 5 0X I < 0:
12
; C1 ; D11 ;I
As a nal simplication, we can multiply the last row and column of the
preceding LMI by ; 21 . This gives the condition
2  3

NX 0  4A XB +XXA XB
 1 C1 ; N 0


;; D11 ; 0 I < 0:
5 X (9.12)
0 I 1
;C 1 ;D 11 ;;
An analogous procedure combining (9.8) and (9.11) leads to
2   ;1  3
NY 0  4AY C+YY A ;Y;C;11 DB1 ;;;1 5 NY 0 < 0:
  

0 I 1 11 0 I (9.13)
;;1 B1 ;;1 D11 ;;;1
To summarize, we have now reduced robust stabilization to conditions (9.9),
(9.12) and (9.13). Condition (9.12) is an LMI on the variables X , ;, but
unfortunately we nd that (9.13) is an LMI on Y and ;;1 , not on ;. This
means that the conditions, as written are not jointly convex on the variables
X , Y and ;. Thus we have not been able to reduce robust synthesis to LMI
computation. In particular, examples can be given where the allowable set
of ;-scalings is even disconnected.
Of course at this point we may wonder whether some further manip-
ulations and possibly additional changes of variables may not yield an
equivalent problem which is convex. No such conversion is known, but at
the time of writing no mathematical proof to the contrary is known ei-
 9.3. Robust Controller Synthesis 289
ther. However insights gained from computational complexity theory (see
remarks below) seem to indicate that such a conversion is impossible.
Since our approach does not provide a convex answer for the general
stabilization problem, the next question is whether there are special cases
where the conclusion is dierent. Clearly if somehow we could get rid of one
of the two conditions (9.12) and (9.13), an LMI would result. One such case
is the so-called full information control problem where the measurements
are
 
y = xq :
That is the controller has direct access to the states and all the outputs of
the uncertainty blocks. Here
   
I
C2 = 0 ; D21 = I0 ;
therefore the kernel of [C2 D21 ] is trivial, so the constraint (9.12) disappears
completely. Consequently the variable X can be eliminated and the robust
synthesis problem reduces to (9.13) and Y > 0, hence an LMI problem in
Y and ;;1 .
A dual situation which is also convex is the full control case, where
(9.13) disappears. A few other instances of simplication in synthesis are
mentioned in the references at the end of the chapter. However these cases
are very special and in general robust design under structured uncertainty
remains a dicult problem. An alternative viewpoint which reinforces
this conclusion, is work on bilinear matrix inequalities (BMIs): these are
feasibility problems of the form
f (X; Y ) < 0;
where f is matrix valued and bilinear in the variables X , Y . Our robust
synthesis problem can indeed be rewritten in this form, as is shown in the
references. This is unfortunately not very helpful, since BMIs do not share
the attractive computational features of LMIs; rather, the general BMI
problem falls in an intractable computational complexity class3.
Given this complexity we are led to consider heuristic methods for
optimization; these will be mentioned after discussing synthesis issues
associated with time invariant uncertainty.

9.3.2 Robust synthesis against
TI

We now turn our attention to the uncertainty class
. From the pre-
TI
ceding theory the robust stabilization problem reduces in this case to the
3
To be precise it is provably NP-hard, in the terminology of complexity theory.
290 9. Feedback Control of Uncertain Systems
optimization
;

inf sup  S^(G; K )(j!);
s;f :
K 2K !2R
This synthesis problem is however even harder than the one studied just
now in x9.3.1, since the function  to be optimized is dicult to evaluate. In
other words we are starting with a weaker result from the analysis side, so
the addition of K can only make matters worse. For this reason the above
optimization is rarely attempted, and what usually goes by the name of -
synthesis is the minimization based on the upper bound for the structured
singular value. This is the optimization
 
inf sup inf  ;! S^(G; K )(j!);;! 1 :
K 2K !2R ;! 2;s;f
(9.14)
If the above inmum is made less than one, we have a robustly stabilizing
controller from the analysis theory. One should note, however, that except
for -simple structures, the converse is not true, that is the previous method
might fail even when robust stabilization is achievable.
At this point we take the opportunity to discuss the relationship between
the above problem and the one obtained by using scaling operators from
; , namely
TI

inf k;^ S^(G; K );^ ;1 k1: (9.15)

K 2K;;2;T I
This topic pertains in fact to the analysis problem for a xed K , but has
been postponed from Chapter 8, since it is for the synthesis step that this
issue acquires the most relevance.
We will now argue that (9.14) and (9.15) give the same optimal value.
Clearly, if we nd a controller K 2 K and ; 2 ; satisfying
TI

k;^ S^(G; K );^ ;1 k1 < 1; (9.16)

then the answer to (9.14) will also be less than one, because we can always
pick ;! = ;^ (j!) 2 ;s;f at each !.
What is less obvious is the converse implication. So suppose that we have
a xed stabilizing controller K and a family of matrix scalings ;! 2 ;s;f
satisfying
 
sup  ;! S^(G; K )(j!);;! 1 < 1;
!2R
or equivalently the LMI form
M^ (j!) ;! ;! M^ (j!) ; ;! ;!  ;I for all ! 2 R;
where  > 0 and we have set M = S (G; K ). Notice that ;! ;! is an element
of P ;s;f for each !, i.e. it is a positive matrix and has the block structure
of ;s;f .
 9.3. Robust Controller Synthesis 291
Now a topological argument, outlined in the exercises at the end of the
chapter, shows that since M^ (s) is in RH1 , then in the above ;! ;! can be
replaced by a rational function ^ (j!), satisfying:
 ^ (j!) 2 RL1.
 ^ (j!) = ^ (j!) 2 P ;s;f for all !.
 For a xed  > 0,
M^ (j!) ^ (j!)M^ (j!) ; ^ (j!)  ;I for all ! 2 R:
What remains to be done, in order to show the equivalence of the two
bounds, is a spectral factorization step where one writes
^ (j!) = ;^ (j!) ;^ (j!)
with ;^ (s) 2 H1 , and ;^ (s) 2 ;s;f for each s in the right half-plane. This
nal step follows by observing that each of the diagonal blocks of ^ (j!) is
lower bounded by I at every frequency, and invoking the following result.
Proposition 9.13. Given a rational function Q^ 2 RL1, such that
Q^ (j!) > I > 0 for all !, then there exists F^ 2 RH1 , with F^ ;1 2 RH1 ,
such that
F^ (j!) F^ (j!) = Q^ (j!) holds for all ! 2 R: (9.17)
This spectral factorization theorem is covered in the exercises of Chapter 7.
Invoking this result for each block, we construct the desired ;^ (s) 2 RH1 ,
with ;^ (s) 2 ;s;f . In other words ;^ represents an operator in ; . Now the
TI
inequality
M^ (j!) ;^ (j!) ;^ (j!)M^ (j!) ; ;^ (j!) ;^ (j!)  ;I for all ! 2 R:
implies that k;^ M^ ;^ ;1 k1 < 1 as required.
Remark 9.14. In the above process we have in fact learned about the
structure of the positive scaling set P ; which goes along with the
TI
commutant ; . We recall from x8.3 that such set is dened by
TI

P ; = f; 2 L(L2 ) : ; = ;~  ;~ ; ;~ 2 ; g;
TI TI

and we remarked there that such sets are not always contained in the cor-
responding commutant. Here, in fact, operators in P ; are self-adjoint
TI
over L2 [0; 1), so in general they are not LTI; in other words they are not
represented by a transfer function in H1 , except when they are memoryless.
Now the above factorization provides in eect a representation for mem-
bers of P ; in terms of transfer functions in L1 . To be more precise, the
TI

operator 0 = ; ; 2 L(L2 ) that corresponds to ; 2 ; would be given by

TI


0 = P + 
L2 [0;1)
292 9. Feedback Control of Uncertain Systems
where  is the LTI operator on L2 (;1; 1) associated with the L1 func-
tion ^ (j!) = ;^ (j!) ;^ (j!), and P+ is the projection from L2(;1; 1) to
L2 [0; 1).
We are now in a position to return to the controller synthesis problem,
which we have reformulated as (9.15); further note from the above argument
that we can restrict the attention to rational scaling functions; explicitly
our optimization problem is to inmize
k;^ S^(G; K );^ ;1 k1 (9.18)
over the stabilizing controllers K 2 K, and scaling functions ;^ 2 RH1
with ;^ ;1 2 RH1 and ;^ (s) 2 ;s;f for every s.
As expected from the discussion in the previous section, this robust syn-
thesis problem is dicult as well, and we will not be able to provide a
computationally tractable solution to the optimization4 . Thus we are led
to consider heuristic algorithms which search for a minimum, without global
guarantees; the most commonly used one, known as D-K Iteration, will be
introduced in the next section.
As a further complication, in the case of (9.18) we will not be able to
bound a priori the order of the optimizing controller. To discuss this, sup-
pose we knew the correct scaling function ;^ (s) and it had order nd . Then
;^ (s);1 can be realized with the same order and we see from an analogous
argument to that in Proposition 9.12 that the controller order need not be
greater than n +2nd. However there is no way to know the required nd: this
issue will become more transparent once we look at the typical methods
for searching for scaling functions.
Suppose we have a xed controller K and thus a xed M^ = S^(G; K ) 2
RH1 . There are two main ways of nding a scaling ;^ which satises (9.16):
(a) Frequency-by-frequency analysis plus curve tting: In this alterna-
tive, for a given M we minimize, at each !
;

 ;^ (j!)M (j!);^ (j!);1
over ;^ (j!). For example we can check for the feasibility of the LMI
M^ (j!) ^ (j!)M^ (j!) ; 2 ^ (j!) < 0; (9.19)
and search over . In practice this is done over a discrete grid of fre-
quencies. Subsequently, a curve tting step is done to nd a rational
;(s) that approximately satises (9.17). The tighter the approxima-
tion, the higher order of ;(s) which will be required, and the step
gives no guarantees for frequencies between grid points.
(b) Basis function approach: Here ;^ (j!) is constrained a priori to lie on
a certain nite dimensional space of transfer functions, for instance
4
Some results exist for special cases; see the chapter references.
 9.3. Robust Controller Synthesis 293
one can take
 ;1  (j!I ; A; );1 
^(j!) = ;^ (j!) ;^ (j!) = (j!I ; A; ) Q
B; B ;
where A; and B; are xed and Q is allowed to vary. Endowing A; ,
B; and Q with appropriate spatial structure, a family of transfer
functions ^ (j!) 2 P ;s;f is parametrized, as discussed in the exer-
cises. It is also shown there how the search over ^ in this family
satisfying (9.19) over frequency, can be reduced to state-space LMI
computation.
This method involves some conservatism over the unconstrained scal-
ings ;^ (j!); as a counterpart, the frequency search and curve-tting
is avoided.
More details on both approaches are given in the references at the end
of the chapter. Having reviewed these approaches, we see that in both the
order of the scaling (determined either by the curve-tting step or by the
order of the basis expansion) can be chosen to be arbitrarily high. Unless
some result could be given showing we need not go higher than a certain
bound, there will no bound on the resulting nd and thus on the order of
the optimizing controller. No such result exists, and in fact examples show
that approximating the inmum in our optimization may indeed require
controllers of arbitrary high order.
We are now ready to move on to a design heuristic.

9.3.3 D-K iteration: a synthesis heuristic

In the previous two sections we have seen that robust synthesis for our
standard setup in Figure 9.1 can be recast as the nding the inmum of
k;S (G; K );;1 k
where K ranges over the stabilizing class K, and ; ranges over:
 the set of constant matrices ; for the case of uncertainty in
. a;c

 the set of rational functions in ; for uncertainty in
.

TI T I;c

We have also mentioned that this is not in general solvable by convex,

nite dimensional methods. This reality leads us to consider heuristic al-
gorithms, and this section is devoted to introducing the most commonly
used approach which is termed D-K iteration. The idea is to split the prob-
lem of robust synthesis, which we cannot solve, into two simpler problems
which we can solve. The two simpler problems are (a) synthesizing an H1
controller; and (b) nding a scaling that inmizes a scaled gain. We have
the following algorithm, which diers for each of these problems only in
the set from which ; is chosen.
294 9. Feedback Control of Uncertain Systems
1. Set ;1 = I , 0 = 1 and the counter k=1;

2. Solve for Kk in the H1 synthesis inf Kk k;k S (G; Kk );;k 1 k; let k

denote the achieved norm.

3. Compare k with k;1 ; if these are approximately equal stop, and set
K = Kk as the nal controller. Otherwise continue to next step.
4. Solve for ;k+1 in the scaled gain problem

inf k;k+1 S (G; Kk );;k+1

;k+1
1 k;

increment k and go to step 2.

This is the D-K iteration algorithm, which is so named because the scaling
variable ; is often denoted by D, and thus the algorithm iterates between
nding solutions for D and K . Starting with the trivial scaling ; = I , the
algorithm begins by performing an H1 synthesis in step (2); later in step
(4) a new scaling ; is found, which can be chosen from either ; or ; TI
depending on which type of uncertainty is involved. Then these scalings
are included for a new H1 synthesis; the algorithm stops when there is no
signicant improvement in the scaled norm.
What are the properties of this algorithm? First the achieved perfor-
mance k at any step forms a nonincreasing sequence up to the tolerance
employed in the inmization steps (an exercise). Also if we are dealing with
the uncertainty set
the scaling ;k is a constant matrix, and therefore
a;c
the controller Kk is never required to have order larger than that of G. In
contrast, when dealing with the uncertainty set
we must nd ratio-
TI

nal scalings by any of the two methods discussed in x9.3.2; in general the
scaling ;k may need to be of arbitrarily high order, and thus so must Kk .
Now notice that the scaled problem always involves the original generalized
plant G; this means that the ;k are modied, not accumulated, in the pro-
cess. Therefore if we impose a restriction on the order of the scalings to be
t, this automatically forces our algorithm to search only over controllers
up to a certain order.
However there is no guarantee that this process converges to the global
minimum, in fact not even to a local minimum: the iteration can get \stuck"
in values which are minima with respect to each variable separately, but
not for both of them at a time. This kind of diculty is not surprising given
the lack of convexity of the problem. Consequently these methods should
be viewed as ways of searching the controller space to improve an initial
design, rather than global solutions to the robust synthesis problem.
 9.4. Exercises 295
9.4 Exercises
1. Complete the discussion in x9.1.2 by providing the details for the
proof of Proposition 9.3. Also show that if S and S ;1 are causal and
in L(L2 ), then their extensions to L2e preserve the inverse.
2. Small gain and nonlinearity. Consider the static function
(
f (x) = 2 x for jxj  1;
1
0 for jxj < 1:
Clearly, jf (x)j  12 x so f has small gain. If I is the identity, is the
function I ; f invertible?
3. Prove Proposition 9.11.
4. Consider the robust synthesis problem under a constant scaling ; and
static state feedback, i.e.
2 3
A B1 B2
G^ (s) = 4 C1 D11 D12 5
I 0 0
and K (s) = F , a static matrix. Show that the search for F and ;
reduces to a bilinear matrix inequality (BMI) condition in the vari-
ables F and (X; ;), where X is a square state space matrix as in the
KYP lemma.
5. Show that the sequence k in the D-K iteration of x9.3.3 is nonin-
creasing. Thus the performance of the controllers Kk generated by
the algorithm will improve, or remain unchanged, as the iteration
proceeds.
6. Basis function approach to uncertainty scalings.
a) Consider the scalar transfer function
^(j!) = 0 + 1 1 +1 j! + 1 1 ;1j! + 2 1 +1 !2 ;
which is real valued for every !. Find xed A0 , B0 , and Q ane
in (0 ; 1 ; 2 ), so that
 ;1  (j!I ; A0 ));1 
^(j!) = (j!I ; A0 )
B0 Q B0 : (9.20)
b) Discuss how to modify the previous construction to describe:
(i) A spatially structured matrix
^ (j!) = 0 + 1 1 +1j! + 1 1 ;1 j! + 2 1 +1 !2 ;
(ii) Terms of higher order in 1+1j! and its conjugate.
296 9. Feedback Control of Uncertain Systems
c) Given a scaling ^ (j!) of this form (9.20), and
 
M^ (s) = CA D
B ;

nd xed matrices A, B and a matrix , anely dependent on

Q, such that
M^ (j!) ^ (j!)M^ (j!) ; ^ (j!) =
(j!I ; A);1  (j!I ; A);1 :
   

B B
d) Explain how to reduce condition (9.19) to an LMI.
7. Comparison of frequency scaling approaches.
a) Suppose that a rational scaling from ; is required to minimize
the gain k;S (G; K );;1 k subject to the constraint of having
TI

a xed maximum order. Can both of the frequency scaling

methods of x9.3.2 guarantee that such a solution will be found?
b) Discuss the computational tradeos between the two methods.
8. Continuous frequency scalings.
Given M^ 2 RH1 we wish to show that if there exists a family of
matrices ! 2 P ;s;f such that
M^ (j!) ! M^ (j!) ; !  ;I for all ! 2 R;
then such a ! exists which is continuous on ! 2 R [ f1g. A sketch
of a proof is provided below, ll in the details.
a) At each frequency ! in R [ f1g, there exist a xed matrix
! 2 P ;s;f and an open interval I! containing ! , such that
M^ (j!) ! M^ (j!) ; ! for all ! 2 I! .
b) Using the matrices and intervals from a) show that there exist a
nite number of intervals Ik and matrices k 2 P ;s;f such that
M^  (j!)k M^ (j!) ; k < 0 for all ! 2 Ik , and
[Nk=1 Ik = R [ f1g holds.
c) From this show that the desired function ! , continuous in R [
f1g can be constructed.
Note: it is a fact that any bounded function ! which is continuous
on R [ f1g can be approximated as closely as desired by a function
in RL1 . This proves the claim made in x9.3.2 when arguing that
rational scalings ; 2 ; can always be constructed.
TI
 9.4. Exercises 297
Notes and references
The input-output perspective for the study of stability for possibly nonlin-
ear or time varying systems goes back to the 1960, in particular to [151] on
the small gain theorem, and the absolute stability theory of [100, 146]. Stan-
dard references on this early theory, in particular covering the nonlinear
case, are [15] and [142]. The case of structured perturbations was already
considered in [146], but mostly gained attention in the late 70s [113]. The
emphasis up to this point was on sucient conditions; for subsequent work
on necessity of small-gain conditions including nonlinear nominal plants,
see [121] and references therein. As mentioned before, necessity theorems
for the structured case were developed in [70, 80, 122].
The fact that robust performance problems could be reduced to robust
stability by addition of a ctitious block is from [27]. The D ; K iteration
for robust synthesis was proposed in [29], including curve-tting in the
D-step; the alternative basis function approach is from [114].
The observation that full-information and full-control problem gives
convex robust synthesis is due to [91]. Other special problems with this
property include the \rank one" case studied in [106], and problems where
uncertainty is represented by a certain type of L2 constraints [19]. Also the
synthesis for robust sensitivity under additive LTI uncertainty has been
reduced in [154, 87] to a convex, yet innite dimensional problem. A recent
survey of the BMI approach for the general (non-convex) case is [83].
 This is page 298
Printer: Opaque this

10
Further Topics: Analysis

10.1 Analysis via Integral Quadratic Constraints

In previous chapters we pursued the modeling of uncertain systems in terms
of spatially structured perturbation balls, and developed the corresponding
methods for robustness analysis. It should be clear by now that the above
modeling strategy is general enough to handle a wide variety of situations
by manipulating the structure of these operator parametrizations. The cor-
responding robustness analysis conditions take the form of generalizations
of the small gain theorem.
Other strategies are, however, available for this class of problems. A
popular alternative has been, for example, to make system passivity (see
a denition below) the central concept instead of contractiveness, which
leads to parallel robustness analysis tools. While it is not dicult to
convert from one framework to the other by suitable (linear fractional) re-
parametrizations, the question arises as to whether one can give a unied
viewpoint that would directly cover the available robustness analysis meth-
ods. The recently developed framework of analysis via integral quadratic
constraints not only provides this but has also led to useful generalizations
of the classical methodologies.
We have already incorporated some elements of this framework in our
presentation of the previous chapters, namely when we exploited quadratic
forms over L2 as a tool to recast some of the robustness analysis condi-
tions. To point the way to the generalizations to follow, we will rst revisit
these results giving a slightly dierent perspective. The next step will be
 10.1. Analysis via Integral Quadratic Constraints 299
to generalize the method to arbitrary quadratic characterizations of uncer-
tainty, and develop robustness analysis theorems that go along with these
characterizations. Finally, at the end of the section we address the question
of computation of the resulting analysis conditions.
We begin by recalling from Chapter 8 that the uncertainty set
was a
characterized by the relation
  
Ra = pq 2 L2  L2 : q =
p; for some
2
a

  
p
= q 2 L2  L2 : kEk pk  kEk qk; k = 1; : : : ; d ;
 
where Ek = 0


0 I 0


0 . Introducing the quadratic forms
 
p = kE pk2 ; kE qk2;
k q k k
the above can be expressed as
    
Ra = pq 2 L2  L2 : k pq  0; k = 1; : : : ; d :
Notice that the forms k map L2  L2 ! R, which is slightly dierent from
our k 's from Chapter 8. We can also rewrite these quadratic forms as
     
p = h p ; p i
k q q k q
with
  
k = Ek0Ek ;E0E :
k k
Now let us combine the constraints in the k by means of multipliers k > 0.
We dene
  d      
p :=
X
k k p = h pq ; pq i
q k=1
q
where the matrix
d  
k k = ;0 ;0; ;
X
= (10.1)
k=1
and we recall that
d
X
;= k Ek Ek = diag( 1 I; : : : ; d I )
k=1
is an element of P ;, the set of positive scaling matrices from earlier
chapters.
300 10. Further Topics: Analysis
From the above discussion, for any p q related by q =
p,
2
we a
have the inequality
 
p  0: (10.2)
q
Next we turn our attention to the robustness analysis results. It was
shown in the previous chapters that the feasibility of the operator inequality
M  ;M ; ; < 0
for ; 2 P ;, is necessary and sucient for robust well-connectedness of
(M;
), and also for robust stability of (M;
) when M is causal.
a a;c

The above condition is easily rewritten as

    
M ; 0 M < 0;
I 0 ;; I
which means that
      
q; I 0 ;; MI q  ;2kqk2 ;
M ; 0

or equivalently that
 
Mq  ;kqk2 for every q 2 L2 : (10.3)
q
Notice that the set
  
RM := Mq q : q 2 L2  L2  L2
is the relation or graph dened by the operator M .
Consequently we can interpret the above results in geometric terms inside
the space L2  L2. Condition (10.3) species that the relation RM lies in
the strict negative cone dened by the quadratic form , whereas (10.2)
states that the uncertainty relation Ra lies in the non-negative cone. Thus
the two relations are quadratically separated by the form .
The results of Chapters 8 and 9 imply that the robust well-connectedness
of (M;
) (respectively the robust stability of (M;
)) is equivalent to
a a;c
the existence of of the form (10.1) such that the resulting form provides
this quadratic separation.
What happens with the analysis over
? While the exact (and dicult
TI
to compute) structured singular value tests are of a dierent nature, its
convex upper bound has indeed the same form.
Let be an LTI, self-adjoint operator on L2 (;1; 1), characterized by
the frequency domain L^ 1 function
 
^
^ (j!) = ;(j!) ^ 0 ;
0 ;;(j!)
 10.1. Analysis via Integral Quadratic Constraints 301
where ;^ (j!) 2 P ;s;f , corresponding to the spatial structure
s;f of
.
TI
The quadratic form
Z 1
(v) = hv; vi = 21 v^(j!) ^ (j!)^v (j!)d!
;1
is thus dened on L2 (;1; 1), and can in particular be restricted to
L2 [0; 1). In particular we will have
 
p 0
q
whenever q =
p and
2
. Also, the L1 condition
TI

k;^ (j!) 12 M^ (j!);^ (j!); 12 k1 < 1

is once again equivalent to (10.3). So here we can also interpret the above
convex analysis test in terms of quadratic separation. Notice that in this
case the test is only sucient.
The above discussion suggests a generalization of our procedure: instead
of constraining ourselves to 's of the above special structure, what if we
are allowed complete freedom in our search for a quadratic form? In other
words we pose the question of nding a suitable that satises (10.2) over
our uncertainty set
, and (10.3) over our nominal system. If such a is
found, we might have a general means of establishing robustness properties
of the (M;
) system. We give the following denition.
Denition 10.1. Let be a self-adjoint operator on L2  L2. The uncer-
tainty set
is said to satisfy the integral quadratic constraint (IQC) dened
by if for every p, q in L2 related by q =
p,
2
, the inequality
     
p = p ; p  0:
q q q
is satised.
Remark 10.2. In the preceding denition is allowed to be any self-
adjoint operator. However the typical situation is that the form is time
invariant on L2 , i.e. (v) = (S v) for every positive shift S . Such forms
can be written as
Z 1
1
(v) = 2 v^(j!) ^ (j!)^v(j!)d!  0;
;1
for a certain L^ 1 function (
^ j!), which helps explain the \IQC" terminol-
ogy.
We now look at dierent examples of uncertainty properties that can be
encompassed by IQCs, in addition to the structured contractiveness which
was studied above.
302 10. Further Topics: Analysis
Examples:
Consider the matrix
 
= I0 I0 :
A component q =
p satisfying
 
p = 2Rehp; qi  0
q
for every q 2 L2 is called passive; this property arises naturally in physical
(electrical, mechanical) modeling.
Now suppose
is a scalar, time varying real parameter, q(t) = (t)p(t).
Then for any anti-Hermitian matrix  = ; we have
p(t)  0  p(t) = 1 (t)[p(t)  p(t) + p(t) p(t)]dt = 0;
Z 1     Z

0 q (t)  0 q(t) 0
so
satises the IQC dened by
 
= 0 0 :
If in addition we have the contractiveness condition j(t)j  1, it follows
easily that
satises the IQC dened by
 
= ;0 ;0;
for any matrix ; > 0. Now one can always superimpose two IQCs, so we
nd that a contractive, time-varying parameter gain always satises the
IQC dened by
 
= ; ;; :
Finally, assume the parameter is real, contractive, and also constant over
time (LTI). Then it follows analogously that the component q = p satises
the IQC dened by
 
^ j!) = ^;^ (j!) ^^(j!) :
( (j!) ;;(j!)
for any bounded ^ (j!) = ;( ^ j!) and ;^ (j!) > 0. 
As a nal remark on the modeling aspect, we have not imposed any
a priori restrictions on the allowable choices of IQC; to be interesting,
however, they must allow for a rich enough set of signals. For instance a
negative denite will only describe the trivial set (p; q) = 0. Furthermore
this example shows that arbitrary IQCs need not respect the restriction
that the variable p is free, which is implicit in any model of the operator
 10.1. Analysis via Integral Quadratic Constraints 303
form q =
p. While a departure from our traditional operator viewpoint
might indeed have interest, in what follows we will assume that our IQCs
are models for some set
of operators.

10.1.1 Analysis results

Having introduced IQC descriptions of uncertainty, we are now ready to
pursue the generalization of our robustness analysis work based on IQCs.
The following discussion refers to the setup of Figure 10.1, relevant to
robust well-connectedness or robust stability questions. We assume M and

are bounded operators on L2 [0; 1). While robust performance problems
can also be included in this setup, they will not be discussed here; interested
readers can consult the references at the end of the chapter.

p q

Mq q
M

Figure 10.1. Setup for robustness analysis

We will assume we have found an IQC satised on the uncertainty (i.e.

is non-negative on the relation dened by
), and such that is strictly
negative on the relation RM , as in (10.3), reproduced here for convenience:
 
Mq  ;kqk2 for every q 2 L2 : (10.4)
q
The rst obvious consequence is that both relations have trivial
intersection: if q =
p and p = Mq, then
   
0 p = Mq  ;kqk2
q q
and therefore q must be zero. So if d = 0 the only L2 solutions to the
diagram of Figure 10.1 are p = 0, q = 0. Equivalently, ker(I ; M
) = f0g;
however, something slightly stronger can be said.
304 10. Further Topics: Analysis
 
Proposition 10.3. Suppose
is described by the IQC p  0, and
q
that (10.4) holds. Then there exists a constant  > 0 such that
kpk  k(I ; M
)pk
holds for every p 2 L2 .
To prove this result we will rely on the following Lemma about quadratic
forms on L2 .
Lemma 10.4. Let (v) = hv; vi, where 2 L(L2 ). Then there exists a
constant , depending only on , such that any z , v in L2 satisfying
(z )  0; (v)  ;kvk2;
will satisfy the inequality
kz ; vk  kz k:
Proof . We rst write the identity
(z ) ; (v) = hv; (z ; v)i + h(z ; v); vi + h(z ; v); (z ; v)i;
that leads to the inequality
kvk2  (z ) ; (v)  2k k kvk kz ; vk + k k kz ; vk2 : (10.5)
Separately, we nd  large enough so that
0  12 kvk2 ; 2k k kvk kz ; vk + 2 kz ; vk2: (10.6)
p
In fact one can take  = 2k k (depending only on ) which makes the
above a perfect square. Now combine (10.5) and (10.6) and obtain
1 kvk2  ;2 + k k
kz ; vk2
2
which is a bound of the form kvk  C kz ; vk. Finally, from the triangle
inequality we have
kz k  kvk + kz ; vk  (C + 1)kz ; vk;
which is the desired bound with  := C + 1. 
We now return to the proof of Proposition 10.3.
Proof . To use the above lemma set
   
z = pq ; v = Mq q :
By hypothesis we know that (z )  0, and
 
(v) = Mq  ;kqk  ;  kvk2 :
q 1 + kM k2
 10.1. Analysis via Integral Quadratic Constraints 305
This means we are in the hypothesis of Lemma 10.4 if we normalize
by a constant factor (dependent on  and M ). Therefore we can nd  =
( ; ; M ) such that
     
kpk 

p


p ; Mq
= kp ; Mqk = k(I ; M
)pk:
q q q


The above result provides an important rst step for robustness analysis
with IQCs. It tells us that the mapping (I ; M
) is injective over L2, and
that the inverse mapping dened on Im(I ; M
) has a norm bound of
. However this still falls short of establishing that the inverse is dened
over all of L2 (i.e. that the mapping is surjective), which would be required
to state that (M;
) is robustly well-connected. The following example
illustrates this diculty.
Example:
Suppose M is the constant gain M = 2, and
is the LTI system with
transfer function
^ s) = s ; 1 :

( s+1
It is easy to see that
is isometric, so kpk = kqk whenever q =
p.
Therefore
satises, for example, the IQC
 
p = ;kpk2 + 2kqk2  0:
q
corresponding to
 
= ;0I 20I :
Also it is clear that 
Mq = ;4kqk2 + 2kqk2  ;2kqk2;
q
Applying Proposition 10.3 we see that
kpk  k(I ; M
)pk
for some constant  (it is easily veried here that  = 1 suces). However
Q := I ; M
has transfer function
;s
Q^ (s) = s3 + 1
and does not have a bounded inverse on L2 [0; 1). In fact it is easy to see
that the operator Q, while injective, is not surjective, since the Laplace
transform of any element in its image must belong to the set
fv^(s) 2 H2 : v^(3) = 0g:
306 10. Further Topics: Analysis
In fact the above subspace of H2 exactly characterizes the image of Q. 
The above discussion implies that in addition to separation of the graphs
of M and
by an IQC, some additional property is required to prove
the invertibility of I ; M
. For the specic IQCs considered in previous
chapters, this stronger suciency result was provided by the small-gain
theorem. We seek an extended argument to establish this for a richer class
of IQCs. The solution discussed below is to assume that the uncertainty set
is closed under linear homotopy to
= 0. Before making this precise, we
state a property about the image of operators such as those in the previous
proposition.
Lemma 10.5. Suppose Q 2 L(L2 ) satises
kQpk  kpk
for all p 2 L2 . Then ImQ is a closed subspace of L2 .
Proof . Take a sequence Qpn which converges to z 2 L2. Then Qpn is a
Cauchy sequence so kQpn ; Qpmk <  holds for suciently large n, m.
Now applying the hypothesis we nd
kpn ; pm k  
so pn is also a Cauchy sequence, hence pn ! p 2 L2 . By continuity of Q,
Qpn ! Qp and therefore z = Qp is in the image of Q. 
We can now state the main result.
Theorem 10.6. Suppose the set
is such that if
2
, then 
2

for every  2 [0; 1]. If
satises the IQC dened by , and (10.4) holds,
then (M;
) is robustly well-connected, i.e. I ; M
is invertible over
.
Proof . Fix
2
. Given Proposition 10.3, it suces to show that I ;M

is surjective, because in that case it follows that (I ; M
);1 exists and
has norm bounded by .
Let us suppose that Im(I ; M0
) = L2 for some 0 2 [0; 1]. The key
step is to show that the property is maintained when 0 is replaced by 
satisfying
j ; 0 j  kM 1k k
k : (10.7)
Before proving this step, we note that it suces to establish our result.
Indeed, in that case we can begin at  = 0, where Im(I ) = L2 trivially,
and successively increment the value of  by the above steps, which are
constant, independent of the current  . In a nite number of steps we will
cover the interval  2 [0; 1].
We thus focus on the perturbation argument for a given 0 , and 
satisfying (10.7). By contradiction, suppose
Im(I ; M
)
 10.1. Analysis via Integral Quadratic Constraints 307
is a strict subspace of L2. Since it is closed by Lemma 10.5, then by the
projection theorem it has a non-trivial orthogonal complement. Namely we
can nd a function v 2 L2 , kvk = 1, such that
h(I ; M
)p; vi = 0 for all p 2 L2:
Now observe that
(I ; M0
) = (I ; M
) + M ( ; 0 )

therefore
h(I ; M0
)p; vi = hM ( ; 0 )
p; vi for all p 2 L2 : (10.8)
Since I ; M0
is surjective we can nd p0 satisfying
(I ; M0
)p0 = v;
and furthermore such p0 has norm bounded by . Substitution into (10.8)
gives
1 = hv; vi = hM ( ; 0 )
p0 ; vi  kM k j ; 0 j k
k  < 1;
which is a contradiction. Therefore I ; M
must be surjective as required.

We have thus obtained a fairly general robust well-connectedness test
based on IQCs; the extra condition imposed on the uncertainty set
is
quite mild, since one usually wishes to consider the nominal system (
= 0)
as part of the family, and thus it is not too restrictive to impose that the
segment of operators 
,  2 [0; 1] also keep us within
. The reader
can verify that this is true with the IQC models presented in the above
examples.
An important comment here is that we have only shown the suciency
of this test for a given IQC. When studying the uncertainty set
in a

Chapter 8 we showed as well that the method is non-conservative, in the

sense that if the system is robustly well-connected
 this fact can always be
; 0
established by an IQC of the form = 0 ;; with ; 2 P ;. Notice
that such family of IQCs is obtained as the conic hull of a nite number of
time invariant IQCs ( k considered at the begin of the section), and the
uncertainty set is rich enough to be characterized exactly by such family.
Only in such quite rare situations can one expect necessity theorems to
hold; this subject will not be pursued further here.
Having discuss well-connectedness, we now turn our attention to the
question of robust stability of the conguration of Figure 10.1, understood
as establishing a relationship in the extended space L2e . The following is a
simple immediate corollary of Theorem 10.6.
Corollary 10.7. Suppose the hypothesis of Theorem 10.6 hold, and that
M and the uncertainty class
are well dened over L2e . Suppose further
308 10. Further Topics: Analysis
that I ; M
is invertible as a mapping in L2e for
2
. Then the system
(M;
) is robustly stable.
Proof . Since (I ; M
);1 is assumed to be well dened over L2e, it only
remains to establish boundedness over L2. But this follows directly from
Theorem 10.6. 
The most common situation in which the conditions of the above corol-
lary can be satised, is when M and
are taken to be causal, and I ; M

has a causal inverse over L2e ; the IQC analysis comes in as a way of ensuring
boundedness over L2 .
As a nal remark in this section, we note that we have developed the
theory exclusively for linear operators M and
, in tune with the style
of this course. In fact the analysis applies, with minor modications to
nonlinear components; in a sense, IQCs provide a more sophisticated tool
to \cover" nonlinearities than the sector bounds that we used to motivate
our uncertainty
. For nonlinear formulations of the analysis, as well as
a
application examples of IQCs for some common nonlinearities, we refer the
reader to the references at the end of the chapter.

10.1.2 The search for an appropriate IQC

A question we have not yet addressed explicitly is how one nds an IQC
to satisfy the assumptions of the above theory, namely to quadratically
separate the relation RM from the uncertain relation. The success of the
method as an analysis technique clearly hinges on this question.
There are two aspects to the above issue:
1. How to nd IQCs that describe a given uncertainty set
.
2. How to search, among the above class, one IQC that satises (10.4)
and hence quadratic separation.
The rst question cannot have a general purpose answer, since it depends
strongly on how the uncertainty set
is presented to us; in particular it
may be more or less easy to extract, from the problem formulation, a valid
family of IQCs, as was done for instance in the examples of the previous
section.
What is true in general, however, is that the set of  which describe a
given relation R (i.e. which are non-negative over pq 2 R), is always a
convex cone: if 1 and 2 are valid IQCs for R, then so is
1 1 + 2 2
for any 1  0, 2  0. Similarly, condition
M  M <0
   

I I (10.9)
 10.1. Analysis via Integral Quadratic Constraints 309
is convex in . Therefore the search for an IQC is a (possibly innite
dimensional) convex feasibility condition over .
In more practical terms, one usually considers a family of known IQCs
for the relevant uncertainty class
, described by some free parameters;
this is potentially a subset of all the valid IQCs. Then one searches over the
set of parameters to nd one that satises (10.9). If successful, the search
is over; if not, it is conceivable that a richer \catalog" of IQCs may be able
to solve the problem.
How is such a parametrization of IQCs obtained? Clearly, one can be
generated from any nite set 1 ; : : : ; d of valid IQCs, by means of the
relevant convex cone. Sometimes, however, as in the example of scalar pa-
rameters discussed above, a matrix parameter such as  = ; or ; > 0
is more convenient. In some cases, most notably for LTI uncertainty, one
disposes of an innite family of IQCs characterized in the frequency domain
by
^ (j!) 2 S
at every !. In this case we have the same options that were discussed in
Chapter 9: either keep the parametrization innite dimensional, or restrict
it to a nite dimensional space of transfer functions by writing
 ;1  (j!I ; A );1 
^ (j!) = (j!I ; A ) Q (10.10)
B B

where A and B are xed and impose the structure of the relevant set
S , and Q is a free matrix parameter. The latter form can fairly gener-
ally accommodate a nite parametrization of a family of self-adjoint, LTI
operators.
The two options are distinguished when we address the second basic
question, namely the search for an IQC satisfying (10.9). Assuming M is
LTI, the frequency domain alternative is to impose
M^ (j!)  (
   
^ j!) M^ (j!)  ; for all !;
I I
which in practice implies a one-dimensional gridding approach.
For nite parametrizations of the form (10.10), the search for the param-
eter Q reduces to a state-space LMI by application of the KYP Lemma, as
discussed in the exercises of Chapter 9.
So we see that given a family of IQCs that satisfy the uncertainty, the
search for an appropriate one to establish robust stability can be handled
with our standard tools. Thus we have an attractive general methodology
for a variety of robustness analysis problems.
310 10. Further Topics: Analysis
10.2 Robust H2 Performance Analysis
In Chapter 8 we introduced a language for the characterization of system
uncertainty, based on uncertainty balls in the L2-induced norm. This pro-
cedure led naturally to conditions for robust stability analysis based on H1
norms of the nominal system, which in fact constitute the main motivation
for the use of H1 norms in control theory. When treating the problem of
system performance, it was shown in Chapter 9 how a disturbance rejection
specication in the H1 norm could be imposed with the same methods as
robust stability analysis.
Other than this mathematical convenience, however, the motivation for
H1 norms in disturbance rejection problems is often not very strong. An
H1 criterion measures the eect of the worst case adversary in a game
where we know nothing about the spectral content of the disturbance.
As argued in Chapter 6, very commonly one has available information on
the disturbance spectrum, that leads more naturally to an H2 measure of
performance.
In short, H1 norms are largely motivated by robustness, and H2 norms
by disturbance rejection. Is there are suitable compromise for robust per-
formance problems? This has been a long-lasting problem, that goes back
to the question of robustness of LQG regulators studied since the late 1970s
(see the references). We will present two approaches for this \Robust H2 "
problem. While the picture is not as tight as for H1 performance, we will
see that similar tools can be brought to bear for this case.

-

p M11 M12  q
z  M21 M22  w

Figure 10.2. Setup for Robust H analysis.

2

Throughout the section we will consider the uncertain system of Figure

10.2. The nominal map M is taken to be a nite dimensional LTI system
with state-space realization
x_ = Ax + Bq q + Bw w
p = Cp x + Dpq q
z = Cz x + Dzq q
 10.2. Robust H Performance Analysis
2 311
with no feed-through terms for the input w. The uncertainty operator

will be of the block structured class

= fdiag(
1 ; : : : ;
d) :
k 2 L(L2); k
k k  1g
a

considered in Chapter 8, or the time-invariant subset
. We will also

TI
consider the causal classes
and
. The robust well-connectedness
a;c T I;c
(respectively robust stability for the causal case) of the interconnection will
be assumed.
For LTI uncertainty, the robust H2 performance analysis problem is to
evaluate
sup kS(M;
)k2 ;

2
TI
where S(M;
) is the closed loop map from w to z , and we recall the
denition
Z 1
kG^ k22 = 21 TrfG^ (j!) G^ (j!)gd!: (10.11)
;1
Unfortunately the above supremum is dicult to compute; this is not un-
expected since the analysis of robust stability alone is already a structured
singular value problem which we have seen is hard. Thus we are led to
seek convex upper bounds for this supremum, similarly to the procedure
we followed in Chapter 8, using the set of scaling matrices
P ; = f; 2 Rmm : ; = diag( 1 Im1 ; : : : ; d Imd ); k > 0g:
We will also study such bounds for the perturbation class
; here the
a
above denition of the H2 norm does not apply; we will discuss some
possible generalizations in the sequel.

10.2.1 Frequency domain methods and their interpretation

In this section we will study a frequency domain characterization of robust
H2 performance, that is closely related to our work in Chapter 8. Consider
the optimization
Problem 10.8. Find Jf ;TI := inf R;1 1 Tr(Y (!)) d! , subject to ;(!) 2 P ;
2
and
   
M^ (j!) ;(0!) I0 M^ (j!) ; ;(0!) Y (0!)  0 8 !: (10.12)

This is an innite dimensional convex optimization problem, consisting of

the minimization of a linear objective subject to a family of LMI constraints
over frequency.
312 10. Further Topics: Analysis
In fact the LMI (10.12) is very closely related to the upper bounds for
the structured singular value: if we replace Y (!) by 2 I in (10.12), we
would impose the robust H1 performance specication that the worst-
case gain of S(M;
), for LTI perturbations, across every frequency and
spatial direction is bounded by .
Here we have added a slack variable Y (!) that allows such gain to vary,
provided that the accumulated eect over frequency and spatial direction
is minimized, re ecting an H2 -type specication. For LTI uncertainty, this
argument is easily formalized.
Theorem 10.9. Suppose the system in Figure 10.2 is robustly well-
connected over the class
. ThenTI

sup kS(M;
)k22  Jf ;TI :

2
TI

Proof . Consider a xed frequency !; by introducing signals q^(j!) , w^(j!)

we obtain from (10.12) the inequality
q^(j!)  M^ (j!) ;(!) 0 M^ (j!) ; ;(!) 0
        
q^(j!)  0;
w^(j!) 0 I 0 Y (!) w^(j!)
that can be expanded into
d
X d
X
jz^(j!)j2 + k (!)jp^k (j!)j2  w^(j!) Y (!)w^(j!) + k (!)jq^k (j!)j2 ;
k=1 k=1
using the signal conventions of Figure 10.2. Since
is time-invariant, and
contractive we have jp^k (j!)j  jq^k (j!)j, so k (!) > 0 implies that
jz^(j!)j2  w^(j!) Y (!)w^(j!):
Now, z^(j!) = [S(M;
)(j!)]w^(j!), so
 
w^(j!) S(M;
)(j!) S(M;
)(j!) ; Y (!) w^(j!)  0
holds for any w^(j!). Therefore we conclude that the preceding matrix is
negative semidenite and in particular
TrfS(M;
)(j!) S(M;
)(j!)g  Tr(Y (!)):
Using denition 10.11, we have
Z 1
kS(M;
)k22  Tr(Y (!)) d!
2 :
;1
Taking supremum over
and inmum over Y^ , we complete the proof. 
The preceding proof is remarkably simple and close in spirit to the meth-
ods of the structured singular value theory. In an analogous manner, we
can study the conservatism of this bound. Not surprisingly, this issue is
 10.2. Robust H Performance Analysis
2 313
related to whether the associated uncertainty structure is -simple or not,
but there are other issues as well. We state the following result.
Proposition 10.10. Assume that the number of blocks d  2, and that
the signal w is scalar. Then
sup kS(M;
)k22 = Jf ;TI :

2
TI
The proof of this result is a standard application of the structured sin-
gular value theory. Since w^ is scalar, so is Y (!), and it can be identied as
the worst-case gain of S(M;
)(j!) at that particular frequency, using the
classication of -simple structures. We leave details as an exercise.
Having identied the non-conservative case, we will now comment on the
sources of conservatism in the general case:
 Non -simple structures. Clearly as the number of blocks increases,
we expect a similar conservatism as the one we have in the structured-
singular value.
 Multivariable noise. As we see in the preceding proof, we are impos-
ing the matrix bound S(M;
)(j!) S(M;
)(j!)  Y (!) as a way
of constraining their traces. Examples can be given showing this is
conservative.
 Causal uncertainty. As we have seen, for the robust stability question
it is natural to consider causal perturbations. This information, how-
ever, is not used in the preceding bound. We will discuss this issue in
more detail below.
In view of these limitations, it is natural to inquire what exactly is
the bound Jf ;TI computing in regard to this problem. It turns out that
frequency domain conditions can be interpreted as tests for white noise re-
jection in a worst-case perspective, as explained below. This approach will
apply as well to the case of arbitrary structured uncertainty
; in that
a

case the conditions will involve the use of constant uncertainty scalings
; 2 P ;, analogously to the situation of previous chapters. In particular,
we will consider the modied problem
Z 1
Jf ;a := inf Tr(Y (!)) d!
2 ; subject to ; 2 P ; and
;1   
^  ; 0 ^ ; 0
M (j!) 0 I M (j!) ; 0 Y (!)  0 8 !: (10.13)
We now outline the set-based approach to white noise rejection, that
is directly tailored to the robustness analysis problem. By treating both
noise and uncertainty from a worst-case perspective, exact characteriza-
tions can be obtained. At rst sight, this proposition may seem strange
to readers accustomed to a stochastic treatment of noise; notice, however,
that a stochastic process is merely a model for the generation of signals
314 10. Further Topics: Analysis
with the required statistical spectrum, and other models (e.g. determinis-
tic chaos) are possible. Here we take the standpoint that rather than model
the generating mechanism, we can directly characterize a set of signals of
white spectrum, dened by suitable constraints, and subsequently pursue
worst-case analysis over such set.
One way to do this for scalar L2 signals is to constrain the cumulative
spectrum by dening the set
  Z   )
 B
W;B := w 2 L2 : min  ; ;  ;   jw^(j!)j2 d!   + ;
; 2 
(10.14)

kwk2

0
B
Figure 10.3. Constraints on the accumulated spectrum
This approach is inspired in statistical tests for white noise that are com-
monly used in the time series analysis literature. The constraints, depicted
in Figure 10.3, impose that signals in W;B have approximately unit spec-
trum (controlled by the accuracy  > 0) up to bandwidth B , since the
integrated spectrum must exhibit approximately linear growth in this re-
gion. Notice that this integrated spectrum will have a nite limit as  ! 1,
for L2 signals, so we only impose an sublinear upper bound for frequencies
above B .
Having dened such approximate sets, the white noise rejection measure
will be based on the worst-case rejection of signals in W;B in the limit as
 ! 0, B ! 1:
kS(M;
)k2;wn := lim!0
sup kS(M;
)wk2 :
B !1 w2W;B
For an LTI system under some regularity assumptions, k
k2;wn can be shown
to coincide with the standard H2 norm; see the references for details. The
method can be extended to multivariable noise signals, where the compo-
 10.2. Robust H Performance Analysis
2 315
nents are restricted to have low cross-correlation. Notice that the preceding
denition can apply to any bounded operator, even nonlinear.
We are now ready to state a characterization of the frequency domain
test.

Theorem 10.11. Suppose the system in Figure 10.2 is robustly well-

connected over the class
. With the above denitions,
a

Jf ;a = lim sup kS(M;
)wk22 :

!0 w2W
B !1
2
;B
a

The proof of this result involves extending the \S-procedure" method

employed in Chapter 8, to accommodate the additional constraints on the
signal spectrum. Notice that the constraints (10.14) are indeed IQCs in
the signal w, although in contrast to the situation of Chapter 8, we have
an innite family of them. Thus the proof is based on a convex duality
argument in innite dimensional space; for details see the references. An
additional comment is that the characterization remains true if
is a

replaced by the causal subset
.a;c

We remark that a similar interpretation can be given to the cost function

Jf ;TI, as a test for robust performance over white noise sets in the sense
dened above, for uncertainty of arbitrarily slow time variation. This also
parallels results available for H1 -type performance; references are given at
the end of the section.
In summary, by using the worst-case interpretation of white noise re-
jection, which in itself entails some conservatism, we are able to obtain a
parallel theory to the one obtained in the previous chapters for the H1
performance measure.
We end the discussion on these frequency domain characterizations with
a few remarks on the computational aspect. Problem 10.8 is innite dimen-
sional as posed, however we note that the problem decouples over frequency,
into the minimization of Tr(Y (!)) subject to (10.12) at each xed fre-
quency. This decoupling can be exploited to build practical approximation
tools, that complement the design tools of -analysis.
If ; is constant as in (10.13) the problem is coupled across frequency.
However in this case the it can be reduced exactly to a nite dimensional
state-space LMI optimization.
316 10. Further Topics: Analysis
Proposition 10.12. Jf ;a = inf Tr(Z ), subject to
2   3
AX; + X; A +Bp ;Bp X;Cq X; Cz
Cq X; 5 < 0;
; ;0 0I
4
Cz X;
2   3
AX+ + X+ A +Bp ;Bp X+Cq X+ Cz
C q X+ 5 < 0;
; ;0 0I
4
Cz X+
 
Z Bv
Bv X+ ; X; > 0:
A proof is provided in the references at the end of the chapter. Here we
will turn our attention to an alternative state-space method.

10.2.2 State-Space Bounds Involving Causality

The above frequency domain methods have been interpreted as picking the
worst among possible \white" disturbances. Now we will present a dierent
bound, that is obtained by focusing on impulse response interpretation of
the H2 norm, but in addition applies to average case white noise rejection.
As we will see, in these interpretations the causality of
plays a larger
role.
We state a second convex optimization problem, this time based on the
state space description. To simplify the formulas we will assume that Dpq
and Dzq are zero, but the general case can be treated similarly.
Problem 10.13. Find Js;a := inf Tr(Bw XBw ), subject to X > 0, ; 2
P ;, and
 
A X + XA + Cp ;Cp + Cz Cz XBq  0:
Bq X ;;

To interpret this problem, we dene another generalization of the H2

norm for time-varying operators. Recall from Chapter 6 that, for LTI
systems, the H2 norm is equal to the sum of the energies of the output
responses corresponding to impulses applied at each input channel. This
leads to the following denition for arbitrary operators:
m
X
kS(M;
)k22;imp := kzk k22;
k=1
where zk = S(M;
)k , and k is the Dirac impulse applied at the k-th
input. We have the following result.
 10.2. Robust H Performance Analysis
2 317
Theorem 10.14. Suppose the system of Figure 10.2 is robustly stable
against the ball
of structured, contractive, causal operators. Then
a;c

sup kS(M;
)k22;imp  Js;a:

2
a;c

Proof . The rst observation is that the eect of the impulse at the i-th
input is to \load" an initial condition x0 = Bw ei in the system, which
subsequently responds autonomously. Here ei denotes the i-th coordinate
vector in Rm . For this reason we rst focus on the problem for xed initial
condition and no input,
J (x0 ) := sup kz k22:

2
a;c ;x(0)=x0
We now write the bound
J (x0 )  sup kz k22
p2L2 [0;1);kp^k k22 kq^k k22
x(0)=x0
d !
 inf sup kz k2 + X (kp^ k2 ; kq^ k2 ) : (10.15)
k >0 q2L2 [0;1) 2 k k 2 k 2
k=1
In the rst step the k-th uncertainty block is replaced by the Integral
Quadratic Constraint kp^k k22  kq^k k22 ; this constraint would characterize
the class of contractive (possibly non-causal) operators. However by re-
quiring p 2 L2[0; 1), we are imposing some causality in the problem by
not allowing p to anticipate the impulse. This does not, however, impose
full causality in the map from q to p, hence the inequality.
Secondly, we are bounding the cost by using the Lagrange multipliers
k > 0 to take care of the constraints. It is straightforward to show the
stated inequality. This step is closely related to the \S-procedure" method
explained in Chapter 8 when studying the structured well-connectedness
problem. In that case we showed the procedure was not conservative; in fact,
a slight extension of those results can be used to show there is equality in
the second step of (10.15).
To compute the right hand side of (10.15), observe that for xed k we
have
Z 1
 
sup x(t) (Cp ;Cp + Cz Cz )x(t) ; q(t) ;q(t) dt: (10.16)
q2L2 [0;1) 0
This is a linear-quadratic optimization problem, whose solution is closely
related to our work in H2 control and the KYP Lemma. We refer to the
exercises of Chapter 7 for the main ideas behind the following Proposition.
Proposition 10.15. If the H1 norm condition
 1   
;2 0
M ;; 12 < 1
0 I 0 1
318 10. Further Topics: Analysis
holds, then the optimal value of (10.16) is given by
x0 Xx0 ;
where X is the stabilizing solution of the Algebraic Riccati Equation
A X + XA + Cp ;Cp + Cz Cz + XBq ;;1 Bq X = 0:
Furthermore, this solution X is the minimizing solution of the LMI
  
A X + XA + Cp ;Cp + Cz Cz XBq  0: (10.17)
Bq X ;;
Notice that the above LMI is the same as the one considered in Problem
10.13. To apply this result, rst notice that the norm condition is
 1
; 2 M11 ;; 21

< 1:
;; 21 M
21 1
Given the robust stability assumption, we know by Theorem 8.12 in Chap-
ter 8 that the norm of the top block can be made less than one by
appropriate choice of ;. Now since ; can be scaled up, the norm of the bot-
tom block can be made as small as desired, yielding the required condition,
and thus the feasibility of (10.17).
Now we wish to combine the solution X with the minimization over ;.
Here is where the LMI (10.17) is most advantageous, since it is jointly ane
in ; and X . We have
J (x0 )  inf x0 Xx0 ;
X;;>0 satisfying (10.17)
that is a convex optimization problem.
The nal step is to return to the sum over the impulses applied at the
input channels:
m
X
kS(M;
)k22;imp = kS(M;
)i k22
i=1
m
X
 J (Bw ei ) (10.18)
i=1
Xm
 inf ei Bw XBw ei (10.19)
i=1 X;;>0 satisfying (10.17)
Xm
 inf ei Bw XBw ei (10.20)
X;;>0 satisfying (10.17) i=1
= inf Tr(Bw XBw )
X;;>0 satisfying (10.17)
= Js;a :
 10.2. Robust H Performance Analysis
2 319
In the above chain of inequalities, (10.18) comes from nding the worst-case

for each initial condition (they need not be the same across i), (10.19) is
the previous derivation, and (10.20) results from exchanging the sum with
the inmum. 
The previous result focuses on the H2 norm as a measure of transient
performance; we immediately wonder if the same bound applies to the other
notions that were used to motivate the H2 norm, in particular in regard to
the rejection of stationary white noise. It can, in fact, be shown that
sup kS(M;
)k22;aov  Js;a ;

2
a;c
where
Z T
kS(M;
)k22;aov := lim sup T1 Ejz (t)j2 dt
T !1 0
is the average output variance of the time-varying system when the input
is stochastic white noise. This bound can be derived from Theorem 10.14,
using the fact that kS(M;
)k2;aov equals the average of the impulse re-
sponse energy as the impulse is shifted over time; see the references for
this equivalence. Alternatively, a stochastic argument can be given using
Ito calculus to show directly this bound; the references contain the main
ideas of this method, that also applies to nonlinear uncertainty. Notice,
however, that in these interpretations we can only prove a bound, not an
exact characterization as was done with Jf ;a .
We end the discussion of this method by explaining how the bound can
be rened in the case of LTI uncertainty. Returning to the proof we would
consider in this case frequency depending scalings k (!), and write
J (x0 )  sup kz k22
q2L2 [0;1);jp^k (j!)j2 jq^k (j!)j2
x(0)=x0
!
d Z 1
 (inf sup kz k 2+X k (!)(jp^k (j!)j2 ; jq^k (j!)j2 )d! :
2
k !)>0 q2L2 [0;1) k=1 ;1
(10.21)
However at this level of generality the restriction q 2 L2 [0; 1) (related
to causality) is not easily handled. The only available methods to impose
this are based on state-space computation, for which the k (!) must be
constrained a priori to the span of a nite set of rational basis functions.
We will not pursue this here (see the references), but remark that in this
way one generates a family of optimization costs JN s;TI of state dimension
increasing with number of elements in the basis, in the limit approaching
the optimization (10.21); we will have more to say about this below.
320 10. Further Topics: Analysis
10.2.3 Comparisons
We have shown two alternative methods to approach the robust H2 perfor-
mance problem; we end the section with a few remarks on the comparison
between them. For simplicity, we focus on the case of scalar disturbances
w.
We rst discuss perturbations in
and state the following relation-
a
ships:
sup kS(M;
)k22;imp  sup kS(M;
)k22;imp

2
a;c
2
a
 lim
!0
sup kS(M;
)wk22 (10.22)
B !1 w
22W
;B
a
= lim
!0
sup kS(M;
)wk22 = Jf ;a :
B !1 w
22W
;B
a;c
The rst inequality is clear; (10.22) follows from the fact that the impulse
(or more exactly, an L2 approximation) is always an element of the \white"
set W;B . The equalities with W;B were stated before.
Notice that the previous inequality does not transparently relate Jf ;a and
Js;a, since we only know that Js;a is an upper bound for the rst quantity.
Nevertheless, we have the following:
Proposition 10.16. Js;a  Jf ;a.
Proof . The exercises provide a direct proof based on the state-space ver-
sion of Jf ;a . Here we will give a more insightful argument for the case of
scalar w. Notice that in the case of a scalar impulse, x0 = Bw and the right
hand side of (10.15) is directly Js;a. Rewriting (10.15) in the frequency
domain we have
Z 1 " d #
inf sup 1 X
jz (j!)j + k (jpk (j!)j ; jqk (j!)j ) d!:
2 2 2
; q2L2 [0;1) 2 ;1
k=1
Now, introducing the slack variable Y (!) to bound R 1the above integrand we
can rewrite this problem as the minimization of ;1 Y (!) d!
2 subject to
d
X
jz (j!)j2 + k (jpk (j!)j2 ; jqk (j!)j2 )  Y (!)
k=1
for all q^(j!) in the H2 , Fourier image of L2 [0; 1). Now, since w(t) = (t)
we have    
p^(j!) = M^ (j!) q^(j!) ;
z^(j!) 1
which translates the previous inequality to
q^(j!)  M^ (j!) ; 0 M^ (j!) ; ; 0
        
q^(j!)  0:
1 0 I 0 Y (!) 1
 10.2. Robust H Performance Analysis
2 321
Now we have an expression that closely resembles Problem 10.8; if q^(j!)
were allowed to be any frequency function, the inequality would reduce to
(10.12) and the two problems would be equivalent. However the constraint
q^(j!) 2 H2 that embeds some causality in Problem 10.13 will lead in
general to a smaller supremum. 
We remark that examples can be given where the inequality in
Proposition 10.16 is strict.
An interesting consequence of the above proof is that for scalar w, there
is equality in (10.22). In fact if we remove the causality constraint in the
previous derivation, the result is exactly Jf ;a . This means that the impulse
can be taken as the worst-case \white" disturbance when we allow for
non-causal uncertainty.
A few more comments are in order regarding the role of causality. As
remarked before, if we are considering the worst-case white noise problem,
the cost does not change by the causality restriction; what happens is
that the signal achieving this cost ceases to be the impulse. When dealing
directly with the impulse response norm, or with average case stochastic
noise, then causality of
does indeed aect the cost, and in this case Js;a
provides a tighter bound.
Finally, we discuss the comparison for the case of LTI uncertainty. Notice
that here we have an unambiguous H2 norm we are trying to compute, for
which both approaches provide bounds.
In this regard, once again we nd that removing the restriction q 2
L2 [0; 1) from (10.21) will lead to the result Jf ;TI , but that there is a gap
between the two. This would mean that if the uncertainty set is
, T I;c
the state-space approach could in principle give a tighter bound. Notice,
however, that we do not have a Js;TI bound, only a family JN s;TI obtained
by basis expansions of order N for the frequency varying scalings. This
means that while
inf JN  Jf ;TI;
N s;TI
we know nothing about the situation with a given, nite N . This is par-
ticularly relevant since the computational cost of state space LMIs grows
dramatically with the state dimension, in contrast with a more tractable
growth rate for computation based on Problem 10.8.

10.2.4 Conclusion
In summary, we have presented two approaches to the Robust H2 perfor-
mance problem, and discussed their interpretation. The two methods are
not equivalent; one oers a tighter characterization of causal uncertainty,
the other the benet of frequency domain interpretation and computation.
A rare feature of H1 theory has been the complete duality between two
ways of thinking, one based on state-space and linear-quadratic optimiza-
322 10. Further Topics: Analysis
tion, the other on operator theory and the frequency domain. After two
decades of research, the robust H2 problem has found these two faces but
not achieved the same unity: LQG does not quite adapt to the world of
robustness.

Notes and References

The method of Integral Quadratic Constraints has its roots in the absolute
stability theory of [100, 146], in particular on the use of the S-procedure
[147] for nonlinear control. However the extensive use of quadratic sepa-
ration for robustness analysis appeared only very recently in the work of
[80, 82, 108, 81]. In addition to the input-output viewpoint presented here,
connections to state-space theory such as Lypunov functions for nonlinear
analysis can be given [107].
The question of robustness in H2 control goes back to the late 70s, a
time when \modern control theory" was put to the test of robustness. In
particular H2 (LQG) regulators were studied from the point of view of
classical stability margins [115], but found to oer in general no guarantees
[22]. This motivated eorts to reconcile LQG with classical concepts [25,
26, 125]. However at the same time (early 80s) the H1 paradigm was
being put forth [152], that more easily blended with classical concepts and
quickly became the centerpiece of multivariable control, since it allowed a
transparent combination of performance and robustness [27].
Nevertheless, the impression has remained that the H2 metric is more
appropriate for system performance. A new opportunity for a better treat-
ment of its robustness came in the late 80s, with the discovery of close ties
between the state-space methods of H2 and H1 control [24]. In particu-
lar the multi-objective H2 /H1 control design problem has received wide
attention, a few of the many references are [10, 71, 156, 120].
Concentrating on robust H2 performance analysis, the state-space
method we presented for robust H2 performance originates in [127, 98, 156],
although the presentation in terms of LMIs is more recent [61]; the exten-
sion to LTI perturbations is due to [35]. The frequency domain method
and the worst-case white noise interpretation are developed in [93, 94]. In
particular [94] contains the state-space computation for Jf ;a . The results
on characterization of Jf ;TI as a test for slowly-varying uncertainty, as well
as the extensions to causal perturbations, are found in [92]. More extensive
details on these and other approaches, as well as many more references can
be found in the survey chapter [95].
 This is page 323
Printer: Opaque this

11
Further Topics: Synthesis

We have arrived at the nal chapter of this course. As with the preceding
chapter our main objective is to acquire some familiarity with two new
topics; again our treatment will be of a survey nature. The two areas we
will consider are linear parameter varying systems, and linear time varying
(LTV) systems. The previous chapter considered advanced analysis, our
aim in this chapter is synthesis.
Up to this point in the course we have worked entirely with systems and
signals of a real time variable, namely continuous time systems, however
in this chapter we will instead consider discrete time systems. One reason
for this is that the concepts of the chapter are more easily developed and
understood in discrete time. This change also gives us the opportunity to
re ect on how our results in earlier parts of the course translate to discrete
time.
The basic state space form of a discrete time system is given below.
xk+1 = Axk + Bwk x0 = 0 (11.1)
zk = Cxk + Dwk :
This system is described by a dierence equation with an initial state con-
dition 0 , and these replace the dierential equation and initial condition
we are accustomed to in continuous time. Thus every matrix realization
(A; B; C; D) species both a discrete time system, and a continuous time
system. We will only use discrete time systems in this chapter.
Before starting on the new topics we brie y digress, and dene the space
of sequences on which the systems discussed in this chapter will act. We use
324 11. Further Topics: Synthesis
`n2 (N ) to denote the space of square summable sequences mapping the non
negative integers N to Rn . This is a Hilbert space with the inner product
1
X
hx; yi`2 = xk yk :
k=0
The space `n2 (N ) is the discrete time analog of the continuous time space
Ln2 [0; 1). We will usually write just `2 when the spatial dimension and
argument are clear.

11.1 Linear parameter varying and

multidimensional systems
So far the controllers we have aimed to synthesize have been pure state
space systems. Here we will extend our state space framework to cover a
broader class of systems. This extension will allow us to incorporate both
linear parameter varying (LPV) systems and multidimensional systems
into a common setting, with performance guarantees during synthesis.

1 In1 0
...
0 d Ind

A B
z C D w

Figure 11.1. NMD system

The new set of state space systems that we introduce are shown in Fig-
ure 11.1. The picture shows the upper star product between two systems.
The upper system is spatially diagonal, and the lower system is memoryless.
 11.1. Linear parameter varying and multidimensional systems 325
Each of the blocks in the upper system is written i Ini which means
2 3
i 0
i Ini = 64 . . . 75 2 L(`n2 i );
0 i
for some operator i on `12(N ). In words the operator i Ini is just the spa-
tially diagonal operator formed from ni copies of the bounded operator i .
For reference we call the operators i scalar-operators, because they act on
scalar sequences. Here (A; B; C; D) is simply a set of state space matri-
ces; dene n such that A 2 Rnn and therefore n1 +


+ nd = n holds.
Referring to Figure 11.1, let us set
2 3
1 In1 0

= 64 ... 7
5 (11.2)
0 d Ind
for convenient reference. Therefore we have the formal equations
xk = (
Ax)k + (
Bw)k (11.3)
zk = Cxk + Dwk
describing the interconnection. We use (
Ax)k to denote the k-th element
of the sequence given by
Ax; note that since the state space matrix A
has no dynamics (Ax)k = Axk in this notation. Here w 2 `2 and we
dene this system to be well-posed if I ;
A is nonsingular. Thus there
is a unique x 2 `2 which satises the above equations when the system is
well-posed. Now the map w 7! z is given by C (I ;
A);1
B + D and
is rational in the operators 1 ; : : : ; d when they commute. However these
operators i do not commute in general, and we therefore call these types
of systems noncommuting multidimensional systems, or NMD systems for
short. NMD systems can be used to model numerous linear situations, and
we now consider some examples.
Examples:
First we show that our standard state space system can be described in
this setup. Let Z denote the shift operator, or delay, on `2 . That is, given
x = (x0 ; x1 ; x2 ; : : : ) 2 `2 we have
Zx = (0; x0 ; x1 ; : : : ): (11.4)
We will not distinguish between shifts that act on `2 of dierent spatial
dimension. Therefore, given a sequence x 2 `n2 we have
2 3 2 32 3 2 (1) 3
Zx(1) Z 0 x(1) x
6 . 7 6 . 76 . 7 6 . 7
Zx = 4 . 5 = 4
. . . 5 4 . 5 ; for x partitioned as 4 .. 5 2 `n2 :
.
Zx(n) 0 Z x(n) x(n)
326 11. Further Topics: Synthesis
Namely Z acts independently on every scalar sequence x(i) comprising the
vector sequence x; so it is spatially diagonal. Now by setting
= Z we see
that the above system is exactly of the form in (11.3)
Our next example involves varying parameters. Suppose we have a sys-
tem whose state space realization depends on the real scalar parameters
1k ; : : : ; rk which vary with respect to the discrete time parameter k; the
variation with k may not be known a priori. Let (A~(); B~ (); C~ (); D~ () )
be the realization, where  signies the dependence on the i . If the de-
pendence of each of the matrices is rational in the parameters i , then it is
frequently possible to convert this system to the form in (11.3) with 1 = Z
and 2 = 1 ; 3 = 2 ; : : : ; r+1 = r , for some state space realization of
constant matrices (A; B; C; D). See the Chapter 8 exercises.
For our third example let us consider a multidimensional system in two
independent variables k1 and k2 . The state equation for such a system
follows.
      
xk1 +1; k2 = A11 A12 xk1 ; k2 + B1 w :
xk1 ; k2 +1 A21 A22 xk1 ; k2 B2 k1 ; k2
Now suppose that w 2 `2 (N  N ). Then let Z1 and Z2 be the shift operators
on the variables k1 and k2 respectively. By setting
 

= Z01 Z0
2
this system can immediately be converted to the form in (11.3). Clearly this
construction can be extended to any multidimensional system with inputs
and states in `2 (N 


 N ). Furthermore with some additional technical
considerations, systems with inputs in the space `2 (N 


 N  Z 


 Z)
can also be treated; see the references for details. 
The examples above provide motivation for the use of this model, and
has a clear analogy to the full block uncertainty introduced in Chapters 8
and 9 We will say more about these ties later.
Dene the set of operators

= f
2 L(`2 ) :
is diagonal as in (11.2), and satises k
k`2!`2  1g:
This is the set of contractive operators that have the diagonal form in
(11.2). We dene X to be the set of positive symmetric matrices in the
commutant of
by
X = fX 2 Sn : X > 0 and X
=
X , for all
2
g:
Therefore every element of X has the block-diagonal form
2 3
X1 0
X = 64 ... 7
5; (11.5)
0 Xd
 11.1. Linear parameter varying and multidimensional systems 327
where each Xi 2 Sni and is positive. We are now ready to state a ma-
jor result pertaining to NMD systems. The following theorem mirrors the
continuous time, full block uncertainty case of Theorem 8.12.
Theorem 11.1. The operators I ;
A and C (I ;
A);1 B + D are non-
singular and contractive respectively, for all
2
, if and only if there
exists X 2 X such that
A B  X 0 A B ; X 0 < 0:
      

C D 0 I C D 0 I
This theorem provides a necessary and sucient condition for such a NMD
system to be both well-posed and contractive with respect to the xed set

. We remark that this condition remains exact when 1 is xed to be
the shift Z . Thus it is exactly the scalar-operator-times-identity version of
Theorem 8.12 in discrete time; see also the exercises in Chapter 8 for a
starting point for proving this theorem.
The pure LMI test given above is also reminiscent of the KYP lemma,
however X is now structured. The result in Theorem 11.1 aords the oppor-
tunity to develop synthesis methods directly for NMD systems. Synthesis
is the topic of the next subsection, after which we discuss realization theory
for models of this type.

11.1.1 LPV synthesis

We now look at a synthesis problem associated with NMD systems. This
is commonly known as LPV synthesis because of the example given above
using varying parameters. This problem has a direct connection with gain
scheduling, as well as synthesis for systems with both temporal and spatial
variables. After posing the synthesis problem we will show its application
to gain scheduling.
The arrangement for the synthesis problem is shown below in Fig-
ure 11.2, and corresponds to our standard synthesis problem, but now the
constituent systems are NMD systems. In the setup we have an NMD sys-
tem G(1 ; : : : ; d), which is given and dependent on the scalar-operators
1 ; : : : ; d . The aim is to synthesize an NMD controller K (1 ; : : : ; d) in
terms of the same scalar-operators. Using the variables shown in the gure,
the equations which describe the plant are
  
xk = (
Ax)k +
B1 B2 wu
 
(11.6)
       k
zk C1 D11 D12 wk
yk = C2 xk + D21 0 uk ;
328 11. Further Topics: Synthesis

z w
G(1 ; : : : ; d )

y
K (1 ; : : : ; d )

Figure 11.2. NMD synthesis conguration

for given state space matrices (A; B; C; D), where
2 3
1 In1 0

= 64 ... 7
5:
0 dInd
These are the analogous equations to those we began with in H1 synthesis.
Similar to this the controller equations are
xk = (
 Ax)k + (
 By
 )k (11.7)
 
uk = C xk + Dyk ;
with
2 3
1 In 1 0

 = 64 ... 7
5:
0 dIn d
We dene the state dimensions of the plant and controller respectively to
be n and n, so that A is n  n, the matrix A is n  n , and therefore the
partition dimensions satisfy
d
X d
X
ni = n and n i = n:
k=1 k=1
The synthesis goal is to nd suitable matrices (A;  B;
 C;
 D ), where the
dimension n can be chosen freely as can the partitioning n1 ; : : : ; n d re-
specting the condition n 1 +


+ n d = n . However the dimension n and
the partition n1 ; : : : ; nd belonging to the plant is xed. Our synthesis ob-
 11.1. Linear parameter varying and multidimensional systems 329
jective is that K ensure the closed-loop is both stable and contractive. We
will make this precise soon, and show that we can easily solve this problem
using the procedure developed in Chapter 7. Before doing this let us rst
look at the applicability of the NMD synthesis problem to gain scheduling.
Example:
We return to the parameter varying setup described as an example above.
We have a state space realization (A~(); B~ (); C~ (); D~ () ), which de-
pends on the time varying parameters 1k ; : : : ; rk . Now consider a gain
scheduling scenario where these scalar parameters jk would not be known
a priori, but are known as each time k is reached. Further suppose that
although they are not known ahead of time, these parameters reside in
normalized intervals, so that jk 2 [;1; 1] always holds. In this scenario
it would be desirable to design a controller that made use of the measured
parameters and their interval bounds. Such a controller would be of the
form K (1 ; : : : ; r+1 ), where 1 = Z and i = i;1 , for i  2. Thus if our
plant can also be converted to an NMD system G(1 ; : : : ; r+1 ), nding a
controller can be posed directly as an NMD synthesis problem. 
Having provided some additional motivation for the setup let us return
to equations governing the closed-loop in Figure 11.2. The equations in
(11.6) and (11.7) can be written more compactly as
xL k = (
L AL xL )k + (
L BL w)k
zk = CL xLk + DL wk ;
where the matrices (AL ; BL; CL ; DL) are appropriately dened, and
   

L =
0
0 ; with xL = xx :
Notice that (AL ; BL; CL ; DL) have the same denitions as in the H1
synthesis of x7.2. We say an NMD controller K is admissible if for all
1 ; : : : ; d 2 L(`12 ), with ki k`2 !`2  1, the following two conditions are
met:
 the operator I ;
L AL is nonsingular;
 the map w 7! z is contractive on `2 .
These conditions state that a controller must satisfy a stability or
well-posedness condition, and simultaneously attenuate the input-output
mapping w 7! z . They have clear analogies with the criteria we set for H1
synthesis.
As mentioned already this NMD synthesis problem can solved using es-
sentially the same methodology we employed in solving the H1 synthesis
problem in x7.2 of Chapter 7. We will now provide a sketch of the steps in
the proof. The NMD synthesis result will then be stated in Theorem 11.5.
330 11. Further Topics: Synthesis
To start we require a temporary change of basis via a permutation.
Consider the form of
L :
22 3 3
1 In1 0
66
64 ... 7
5 0 7
7
  6 7



L = 0
 = 60 6
6 0 d In d 2 3
7
7
6  1 I n
 1 0 7
7
6
4 0 6
4 . .. 77
55
0 d In d
Each scalar-operator i appears twice in the block structure, once in
and
 Let P be the permutation matrix that satises
once in
.
2 3
1 InL1 0
P 
L P = 64 ... 7
5; (11.8)
0 dInLd
where nLi = ni + n i . That is this permutation matrix simply rearranges
the blocks so that each i only appears in a single block. Dene the positive
symmetric commutant of such operators
XL = fXL 2 Sn+n : XL > 0 and XL(P 
L P ) = (P 
L P )XL ,
for all i 2 L(`2 ) with
L dened in (11.8)g:
That is, any element of XL has the form
2 3
XL1 0
XL = 64 ... 7
5;
0 XLd
where each XLk 2 Sni+ni and is positive denite. We can now state the
following result.
Proposition 11.2. Suppose K (1; : : : ; d) is a NMD controller. Then K
is admissible if and only if there exists XL 2 XL such that
PAP PB  XL 0 PAP PB ; XL 0 < 0:
      

CP D 0 I CP D 0 I (11.9)
Proof . We need to show that both well-posedness and contractiveness are
tantamount to the LMI condition of the proposition. That is I ;
L AL
must be nonsingular and CL (I ;
LAL );1
L BL + D contractive. Ob-
serve that P ;1 = P  since P is a permutation matrix. Thus the two
conditions are equivalent to I ; (P 
L P )(P  ALP ) is nonsingular and
CL P (I ; (P 
L P )(P  AL P ));1 (P 
L P )P  BL + D is contractive. Now
P 
L P is of the form in (11.8), and so by invoking Theorem 11.1 we see
these two latter conditions hold if and only if (11.9) has a solution. 
 11.1. Linear parameter varying and multidimensional systems 331
Checking the admissibility of any given controller is equivalent to an LMI
feasibility problem. Our ultimate goal is to obtain synthesis conditions, so
we will need to examine the block structure of XL in (11.9). Any XL in XL
is of the block diagonal form diag(XL1 ; : : : ; XLd) with blocks given by
 
XLi = XXi X X
2i ;
2i 3i
and the matrices Xi 2 Sni, X2i 2 Rni n i and X3i 2 Sn i. Now consider the
eect of the permutation matrix P in (11.8) on XL . Let XP = PXLP and
we have
 
XP = XX X 2 ; (11.10)
2 X3
where the constituent matrices are
X = diag(X1 ; : : : ; Xd ); X2 = diag(X21 ; : : : ; X2d); and
X3 = diag(X31 ; : : : ; X3d ). We can pre- and post-multiply the expression
in (11.9) by the matrices diag(P  ; I ) and diag(P; I ) respectively, to arrive
at
AL BL  XP 0 AL BL ; XP 0 < 0;
      

C L DL 0 I CL DL 0 I
Now from (11.10) the matrix XP has the same partitioned structure as
the XL-matrix that appears in the H1 synthesis of x7.2; however each
of the matrices X1 , X2 and X3 is block-diagonal. If we apply the Schur
complement formula twice to the above LMI we get the following equivalent
inequality.
2 3
;XP;1 AL BL 0
6 A 
L ;XP 0 CL 7
0 ;I DL 5 < 0: (11.11)
6 7
4 B
L
0 CL DL ;I
That is a controller is admissible exactly when a solution can be found to
this inequality. This is the critical form for solving the synthesis problem.
The left hand side of this inequality is ane in the state space matrices
for the controller. We can use Lemma 7.2 to show that the solution to this
controller dependent LMI, implies the existence of solutions to two matrix
inequalities which do not explicitly or implicitly involve the controller. An-
other application of the Schur complement converts these inequalities into
LMIs, and yields the following lemma.
Lemma 11.3. An admissible controller exists, with partition dimensions
n 1 ; : : : ; n d , if and only if there exist symmetric matrices X and Y such
that
332 11. Further Topics: Synthesis
(i)
2 3

NY 0 4AYCAY A; Y C AY
 C1 B1 N 0
Y C  ; I D Y
0 I 1 1 1 11 5 0 I < 0
B1 
D11 ;I
(ii)
 2 3

NX 0  4A BXA
 ; X A XB1 C1 NX 0
1 XA B1 XB1 ; I D11
5 <0
0 I C1 D11 ;I 0 I
(iii) The identities
   
PXL P = X? ?? and PXL;1P = Y? ??
hold, for some XL 2 XL .
where the operators NY , NX satisfy
 
Im NY = ker B2 D12 NY NY = I
 
Im NX = ker C2 D21 NX NX = I:
We have concluded that an admissible controller exists exactly when there
exist solutions X and Y to conditions (i){(iii) in Lemma 11.3. Now (i)
and (ii) are LMIs and are completely specied in terms of the given plant
G. However (iii) is not an LMI condition, but only depends on controller
dimension. Note that X and Y must necessarily be members of the set X
dened in (11.5) if they satisfy (iii).
So the next step is to convert (iii) to an LMI-type condition. We have
the following result.
Lemma 11.4. The block-diagonal matrices X , Y 2 X satisfy condition
(iii) in Lemma 11.3 if and only if, for each 1  k  d, the following
inequalities hold
   
Xi I  0 and rank Xi I  n + n : (11.12)
I Yi I Yi i i
Proof . By Lemma 7.8 we see that there exist positive matrices XLk 2
Sni+ni exist such that
   
XLk = X?i ?? ; and XLk
;1 = Yi ?
? ?
for 1  k  d, exactly when the inequalities in (11.12) are satised. Clearly
the block-diagonal matrix XL 2 XL , and PXLP satises the condition (iii)
in Lemma 11.3. 
The rst of these conditions is an LMI, however the rank inequality is not.
Note that the rank conditions are trivially met if n i  ni , and therefore
 11.1. Linear parameter varying and multidimensional systems 333
can be eliminated if the controller order is chosen to be the same as the
plant. This situation was observed in our earlier H1 synthesis, where we
had only one variable.
We can now state the NMD system synthesis result.
Theorem 11.5. An admissible NMD controller exists if and only if, there
exist matrices X , Y 2 X , which satisfy LMI conditions (i) and (ii) in
Lemma 11.3 , and the LMI conditions given in (11.12), for each k =
1; : : : ; d,.
This gives exact convex conditions for the existence of a solution to the
NMD synthesis problem. Notice that if the rank conditions in (11.12) are
also achieved then a controller synthesis exists with dimensions n 1 ; : : : ; n d .
Furthermore, an admissible controller exists if and only if one exists satis-
fying n 1 = n1 ; : : : ; n d = nd. Also if d = 1, namely there is only one i , this
result corresponds exactly to the discrete time H1 synthesis problem; this
is more clearly apparent by consulting Proposition 8.28. When solutions X
and Y are found an explicit controller can be computed by constructing a
scaling XL , and then nding a controller realization which solves the LMI
given above in (11.11); this is similar to the procedure in x7.3.

11.1.2 Realization theory for multidimensional systems

By using the system construct shown in Figure 11.1 and described in the
associated text above, a number of systems realization concepts can be
generalized to NMD systems by utilizing an LMI approach. Most notable
of these realization topics are minimality and the associated notions of
reducibility and model reduction. Additional realization results that may
be derived for NMD systems include reachability, controllability and ob-
servability conditions, and the construction of Kalman{like decomposition
structures.
We begin by noting that the LMI given in (11.9) has a Lyapunov in-
equality embedded in its top, left corner. As we know from Chapters 2
and 4, in standard state space systems theory Lyapunov equations play a
formidable role in the development of conditions for testing the stability,
minimality, controllability and observability of a given set of realization
matrices (A; B; C; D). Specically, the following are well-known results
for standard state space system realizations in discrete time. They are all
discrete time analogs of results from Chapters 2 and 4:
(i) A matrix A is discrete time stable if and only if, for each Q > 0,
there exists a matrix X > 0 satisfying
A XA ; X + Q = 0:
This equation is one example of a discrete time Lyapunov equation.
334 11. Further Topics: Synthesis
(ii) Given a stable realization, the pair (C; A) is observable if and only if
there exists a positive denite solution X to the following Lyapunov
equation:
A XA ; X + C  C = 0:
(iii) Given a stable realization, the pair (A; B ) is controllable if and only if
there exists a positive denite solution Y to the following Lyapunov
equation:
AY A ; Y + BB  = 0:
Recall that the terms observable and controllable carry much deeper mean-
ings than are obvious from the above simple statements given in terms of
Lyapunov equations. Observability of a realization implies that given an
output sequence y0; : : : ; yn;1 , the initial state x0 can be uniquely deter-
mined, and controllability of a realization implies that the initial state can
be transferred to any other xed state x 2 Rn via a nite input sequence.
However, in this section, our main purpose is to emphasize the connec-
tions between the solutions to the above Lyapunov equations, namely the
Gramians, and the notion of minimality for a standard state space re-
alization. Given a realization (A; B; C; D) for a system G, we say the
realization is minimal if there exists no other realization for G, for example
(Ar ; Br ; Cr ; Dr ), of lower order. That is a realization satisfying dim(Ar ) <
dim(A).
As we saw in Chapter 4 a stable realization (A; B; C; D) for a system
G is minimal if and only if it is both controllable and observable, that
is, if and only if the Gramians X and Y are strictly positive denite. If a
given realization is not controllable (or observable), the resulting associated
Gramian will be semi-denite.
Returning to NMD systems, a parallel result on minimality can been
shown to hold. First we need to state what we mean by equivalent re-
alizations for NMD systems. We say that two realizations (A; B; C; D)
and (Ar ; Br ; Cr ; Dr ) are equivalent if the NMD systems G(1 ; : : : ; d) and
Gr (1 ; : : : ; d) they represent satisfy
kG(1 ; : : : ; d) ; Gr (1 ; : : : ; d )k = 0; for all contractive k 2 L(L12 ).
More explicitly the systems G and Gr are dened by
G(1 ; : : : ; d ) = C (I ;
A);1
B + D;
Gr (1 ; : : : ; d ) = Cr (I ;
r Ar );1
r Br + Dr ;
where
2 3 2 3
1 In1 0 1 Inr1 0

= 64 ... 7
5 and
r = 4
6 ... 7
5:
0 dInd 0 d Inrd
 11.1. Linear parameter varying and multidimensional systems 335
Thus we see that two realizations are equivalent if the operators they dene
on `2 are the same, for each k . For multidimensional systems we dene
minimal as follows:
Denition 11.6. A realization (A; B; C; D) for an NMD system is
minimal if dim(A) is lowest among all equivalent realizations.
As it turns out it can be shown that any minimal NMD realization will
further satisfy the condition that ni is lowest, for each i = 1; : : : ; d, among
all equivalent realizations.
We are now in a position to state an NMD system minimality result that
is directly based on LMIs. It says a realization is minimal if and only if
there exist no singular solutions to either of a pair of associated Lyapunov
inequalities, where these solutions are restricted to be in the set X . The
result now given is more general in that it addresses so-called reducibility
and contains minimality as a special case. Note that it is assumed that the
D terms are the same in both the full and reduced systems.
Theorem 11.7. Given a system realization (A; B; C; D) , then there
exists a reduced system realization (Ar ; Br ; Cr ; Dr ) such that
sup kG ; Gr k  
kk k1
if and only if there exist X  0 and Y  0, both in the closure X, satisfying
(i) AXA ; X + BB   0
(ii) A Y A ; Y + C  C  0
Pd
(iii) min (XY ) = 2 ; with multiplicity i=1 (ni ; ri )
where   0.
Notice that the  = 0 case gives an exact reducibility condition, which leads
directly to a minimality result; that is a realization is minimal if and only if
all solutions to the system Lyapunov inequalities are nonsingular. If a rank
decient structured Gramian exists for either of the Lyapunov inequalities,
then an equivalent lower dimension realization exists, and vice versa. Via
the proofs for this case we can also show that for a given NMD system, all
minimal realizations may be found by similarity transformations in X and
truncations.
The  > 0 case leads to a model reduction result. A proof can be derived
directly from the NMD synthesis results above in Theorem 11.5 by viewing
the desired reduced dimension system in the role of the controller. As we
have already noted, such rank constraints are not easily incorporated into
computations. We therefore seek a form of model reduction along the lines
of our earlier work in Chapter 4.
336 11. Further Topics: Synthesis
Balanced Truncation
An alternative approach to model reduction of NMD systems is that of
balanced truncation. For this method, the same LMIs must be solved, that
is structured solutions must be found to the system Lyapunov inequalities.
Following this, the system is balanced exactly as in the standard case of
Chapter 4 where now all balancing transformations are required to be in X .
Truncating the balanced system repetitively for each i leads to an additive
error bound that is greater than that stated above by a factor of two. More
specically, we proceed as follows.
In order to derive the model reduction error bounds for balanced NMD
systems, we partition the system matrices A; B; C and the balanced struc-
tured Gramian Y = X =  so as to separate the subblocks which will be
truncated. That is, A, B and C are partitioned compatibly with the block
structure
as
2 3 2 3
A11


A1d B1
A = 64 ... . . . ... 75 ; B = 64 ..  
.
7
5; C = C 1


Cd :
Ad1


Add Bd
We further partition each block of  by i = diag[^ 1i ; 2i ], for i = 1; : : : ; d,
where the realization submatrices corresponding to 2i will be truncated.
Denote
^ 1i = diag[i1 Ini1 ; : : : ; iki Iniki ];
and
2i = diag[i(ki +1) Ini(ki +1) ; : : : ; iti Initi ] ki  ti :
We then truncate both 2i and the corresponding parameter matrices, for
example, we truncate
   
^ 1112 ; B = B^1
h i
A11 = AA11 A A 1 B and C 1 = ^1 C12
C
1121 1122 12
to A^11 , B^1 and C^1 . Partitioning and truncating each Aij , Bj and Ci , i; j =
1; : : : ; d similarly results in the truncated system realization
2
A^11


A^1d B^1 3
Ar Br = 666 ... . . . ... ... 777 ;
 

Cr D 4 A^d1


A^d B^d 5

C^1


C^d D
with an associatedPset structure of
r = fdiag[1 Ir1 ; : : : ; d Ird ] : i 2
L(l2 )g where ri = kj=1i n . Note that
is constructed with the same
ij r
uncertainty variables i as is
, thus the
2
and
r 2
r are not
independent.
 11.2. A Framework for Time Varying Systems: Synthesis and Analysis 337
As in the 1D case, truncating a balanced stable NMD system realization
results in a lower dimension realization which is balanced and stable, which
is easily seen by considering the system Lyapunov inequalities. We now
state the balanced truncation model reduction error bound theorem for
multidimensional systems.
Theorem 11.8. Suppose (Ar ; Br ; Cr ; D) is the reduced model obtained
from the balanced stable system (A; B; C; D). Then
ti
d X
X
sup kG(1 ; : : : ; d) ; Gr (1 ; : : : ; d)k  2 ij : (11.13)
kk k1 i=1 j =ki +1
This result states that the reduced order system is guaranteed to satisfy an
error bound very much similar to the one we saw for standard systems in
Chapter 4. Notice that this result requires the use of Lyapunov inequalities
rather than equations. Also, unlike the previous reduction result, it can be
computed directly using LMIs.
We now leave NMD systems and remove one of the constraints we have
insisted on throughout the course for state space systems, that of time
invariance.

11.2 A Framework for Time Varying Systems:

Synthesis and Analysis
We consider a fundamental class of linear time varying systems in discrete
time. The standard way of describing such a system G is using the state
space system
xk+1 = Ak xk + Bk wk
(11.14)
zk = Ck xk + Dk wk
for wk an input sequence, where Ak , Bk , Ck and Dk are bounded matrix
sequences which are given a priori. The initial condition of the system is
x0 = 0.
Such system models arise in a number of applications. One way they
arise is when a nonlinear system is linearized along a trajectory; they can
arise in multirate control or optimal lter design, where the system is usu-
ally periodic; also they occur when plants and subsystem components are
naturally time varying.
Our main objective is to present a framework for LTV systems that
is consistent with the techniques we have learned so far, and that will
allow us to pursue synthesis and analysis. Thus many of the standard state
space methods we have learned for LTI systems can be applied directly
to LTV systems using the framework we now introduce. We illustrate this
by focusing on controller synthesis for the `2-induced norm (i.e. an H1
338 11. Further Topics: Synthesis
synthesis for LTV systems). In short the technique we now pursue makes a
strong connection with state space and LMI methods so that the transition
from time invariant to time varying analysis appears straightforward.
In x11.2.1 and x11.2.2, to which we now turn, our focus will be intro-
ducing tools for working with LTV systems. Following this we illustrate
the use of this machinery by considering analysis of LTV systems. Having
established these results we state the LTV synthesis conditions in x11.2.4.
The results we derive will be operator inequalities when our initial systems
are general LTV systems. If the the system has the specic structure of be-
ing periodic or of nite time horizon, then the general operator conditions
become structured LMIs. In both cases the conditions we obtain will be
convex.

11.2.1 Block-diagonal operators

We now introduce our rst concept for LTV systems analysis.
Denition 11.9. A bounded operator Q mapping `m2 to `n2 is block-
diagonal if there exists a sequence of matrices Qk in Rmn such that, for
all w; z , if z = Qw then zk = Qk wk . Then Q has the representation
2 3
Q0 0
6
6 Q1 7
7
6
4 Q2 7
5
:
...
0
Further, if Pk 2 Rmn is a uniformly bounded sequence of matrices we say
P = diag(P0 ; P1 ; : : : ) is the block-diagonal operator for Pk , and conversely
given P a block-diagonal operator, the blocks are denote by Pk , for k  0.
Suppose F , G, R and S are block-diagonal operators, and let A be a
partitioned operator, each of whose elements is a block-diagonal operator,
such as
 
A = FR G
S
We now dene the following notation:
     
F G := diag( F0 G0 ; F1 G1 ; : : : );
R S R0 S0 R1 S1
 
which we call the diagonal realization of A. Implicit in the denition of A
is the underlying block structure of the partitioned operator
 A. Clearly for
any given operator A of this particular structure, A is simply A with the
rows and columns permuted appropriately so that
   
F G = Fk Gk :
R S k Rk Sk
 11.2. A Framework for Time Varying Systems: Synthesis and Analysis 339

 which we shall denote by Pl (A)

Hence there exist permutation operators,
and Pr (A), such that Pl (A)APr (A) = A or equivalently
       
Pl FR GS FR GS Pr FR G S = R S
F G

For any operator A whose elements are block-diagonal operators

Pl (A)Pl (A) = Pl (A) Pl (A) = I;
Pr (A)Pr (A) = Pr (A) Pr (A) = I
and if Ais self-adjoint,
 then Pl (A) = Pr (A) . For a concrete example,
consider F G . Then
 
Pl
;
F G


= I; Pr
;
F G

E
= Z  EZ
where
2 3
1 0 0
60 0 1 0 7
E = 6640 0 0 0 1 0 775 :
.. ...
.
The following is immediate.
Proposition 11.10. For any real number , and any partitioned operator
A consisting
  of elements which are block-diagonal, A < I holds if and only
if A < I . That is, positivity is preserved under permutation.
Two further useful facts for the above permutations are the following.
Proposition 11.11.
(i) Suppose that A and B are partitioned operators consisting of block-
diagonal elements, and that their structures are the same. Then
     
A+B = A + B ;
(ii) Suppose that A and C are partitioned operators, each of which consists
of elements which are block-diagonal. Further suppose that the block
structures are compatible, so that the product A^C^ is block-diagonal
for any operators A^ and C^ with the same block structures as A and
C . Then
    
AC = A C :
Proof . Part (i) is obvious.
Part (ii) is simple to see, since Pr (A), the right permutation of A, depends
only on the column dimensions of the blocks in A. Since A and C have
340 11. Further Topics: Synthesis
compatible block structure, Pl (C ) = Pr (A) , and hence
  
A C = Pl (A)APr (A)Pl (C )CPr (C )
= Pl (A)ACPr (C )
= P l (AC )ACPr (AC )
= AC
which is the required result. 
 
The diagonal realization of a partitioned operator, and the notation

will allow us to easily manipulate expressions involving diagonal operators.
We now introduce another new tool.

11.2.2 The system function

To start recall the denition of the unilateral shift operator Z on `2 from
(11.4).
Using the previously dened notation, clearly Ak , Bk , Ck and Dk in
(11.14) dene block-diagonal operators. Using the shift operator Z , we can
rewrite equation (11.14) as
x = ZAx + ZBw
z = Cx + Dw:
The question of whether this set of equations is well-dened, that is whether
or not there exists an x 2 `2 such that they are satised, is one of stability
of the system. If the equations are well-dened, then we can write
G = C (I ; ZA);1 ZB + D; (11.15)
and z = Gw. These equations are clearly well-dened if 1 62 spec(ZA). The
next result shows that this condition is equivalent to a standard notion of
stability of LTV systems, that is exponential stability.
Denition 11.12. The system G is exponentially stable if, when w = 0,
there exist constants c > 0 and 0 <  < 1 such that, for each k0  0 and
any initial condition xk0 2 Rn , the inequality kxk kRn  c(k;k0 ) kxk0 kRn
holds for all k  k0 .
Proposition 11.13. Suppose Ak is a bounded sequence in L(X ) where
X is a Hilbert space. Then the dierence equation xk+1 = Ak xk is
exponentially stable if and only if 1 62 spec(ZA).
This is the well-known result that exponential stability is equivalent to `2
stability of the system xk+1 = Ak xk + vk ; versions of this result can be
found in any standard reference on Lyapunov theory. Thus the system is
stable if and only if 1 62 spec(ZA); we will work with this latter condition.
Throughout the sequel we will refer to the block-diagonal operators A,
B , C and D, and the operator G they dene, without formal reference to
their denitions in (11.14) and (11.15).
 11.2. A Framework for Time Varying Systems: Synthesis and Analysis 341
We now consider the properties of operators of the form of equa-
tion (11.15). Formally, this equation looks very much like the frequency
domain description of a discrete-time time invariant system. It is well known
that for such systems, one can replace the shift operator Z with a complex
number, and then the induced norm of the system is given by the maximum
norm of this transfer function over the unit ball in the complex plane.
We will show that, for linear time varying systems, very similar state-
ments can be made. Indeed, the induced norm of a linear time varying
system can be analyzed by computing the maximum norm of an operator-
valued function over a complex ball. However in this context, we will use a
bounded sequence k 2 C of complex numbers as our notion of frequency.
Robust control techniques to date have been primarily developed for LTI
systems; the system function derived here provides an important and di-
rect link between LTI and LTV systems, making the techniques of robust
control available for LTV systems. In particular, this allows the construc-
tion of convex upper-bounds for structured uncertainty problems for LTV
systems.
Given such a sequence, we will make use of two associated block-diagonal
operators. These are
2 3 2 3
0 I 0 0 I 0
6 1 I 7 6 0 1 I 7
 = 664 2 I
7
7
5

= 664 0 1 2 I
7
7
5
... ...
0 0
(11.16)
on `2 . Observe that

Z = Z
; (11.17)
which is easily veried. Also note that if each element of the sequence k
is on the unit circle T then
is invertible in L(`2 ). Using the denition of
 we dene the system function of the operator G by
G^ () := C (I ; ZA);1 ZB + D;
when the inverse is dened. We can now state the main result of this section.
Theorem 11.14. Suppose 1 62 spec(ZA). Then
kC (I ; ZA);1 ZB + Dk = sup kG^ ()k;
k 2D
where  depends on k as in (11.16).
This theorem says that the induced `2 norm of the system G, which equals
kC (I ; ZA);1 ZB + Dk, is given by the maximum of the norm kG^ ()k,
when the k are chosen in the unit disk. This result looks similar to the
well-known result for transfer functions of time invariant systems, and it is
342 11. Further Topics: Synthesis
the key element in allowing time invariant techniques to be applied to time
varying systems.
In particular, we will see that we can use this result to derive a time
varying version of the KYP lemma, characterizing those systems which are
contractive. This allows the development of time varying analogs of well-
known results in structured singular value analysis. However rst we must
prove a preliminary result.
Lemma 11.15. Suppose 1 62 spec(ZA). Then given any sequence k in the
unit circle T, the operator I ; ZA is invertible and we have
kC (I ; ZA);1 ZB + Dk = kG^ ()k:
Proof . Fix a sequence k 2 T and dene the operator
as in (11.16).
Now notice that both
and
;1 are isometries and therefore
kC (I ; ZA);1 ZB + Dk = k
fC (I ; ZA);1 ZB + Dg
;1 k:
To complete the proof consider the operator on the right-hand side above

fC (I ; ZA);1 ZB + Dg
;1 = C
(I ; ZA);1 Z
;1 B + D
= C
(I ; ZA);1
;1 ZB + D
= C (I ; ZA);1 ZB + D = G^ ();
where we have used the fact that
commutes with A, B , C and D, and
the relationship described by equation (11.17). 
This lemma states that it is possible to scale the system matrices A and
B by any complex sequence on the unit circle without aecting the norm
of the system. Note that this can equivalently be thought of as scaling Z ,
the shift operator. The next lemma describes the eect of the operator 
on the spectrum of ZA.
Lemma 11.16. Suppose that k is a sequence in the closed unit disc D
and dene  as in (11.16):
(i) If  62 spec(ZA), then  62 spec(ZA).
(ii) If the sequence k is further restricted to be in T, then spec(ZA) =
spec(ZA).
Proof . First note that without loss of generality we may assume that
 = 1 in (i), and therefore will show that 1 62 spec(ZA) implies that
1 62 spec(ZA).
We begin proving (i) by invoking Proposition 11.13 to see that, since
1 62 spec(ZA), the dierence equation xk+1 = Ak xk is exponentially stable.
Each k satises jk j  1 and so
xk+1 = k+1 Ak xk
 11.2. A Framework for Time Varying Systems: Synthesis and Analysis 343
is also exponentially stable. Again use Proposition 11.13 to conclude that
1 62 spec(ZQ) where Q is the block-diagonal operator corresponding to
Qk = k+1 Ak . It is routine to verify that ZQ = ZA.
Part (ii) is immediate by applying (11.17) to see that
ZA
;1 = ZA.

Note that in particular (i) and (ii) imply, the (apparently) well-known
result, that the spectrum of ZA is an entire disc centered at zero1; to
see this, set  = I and let  be in D . We can now prove the main result
of this section.
Proof of Theorem 11.14. For convenience dene := kG^ (I )k which is
equal to kGk by denition. Suppose contrary to the theorem that there
exists a sequence k 2 D such that kG^ ()k > . Then there exist elements
x; y 2 `2 satisfying kxk2 = kyk2 = 1 and
jhy; G^ ()wi2 j > :
Without loss of generality we may assume that w and y have nite support,
which we denote by n.
Now it is routine to verify that G^ () is lower triangular and has the
representation
2 3
D0 0
6
6 1 T10 D1 7
7
G^ () =
6
6 2 1 T 20  2 T21 D2 7
7 (11.18)
6 . ...
3 2 1 T30 ..
7
6 7
4 5
..
.
where Tkl = Ck Ak;1


Al+1 Bl . Therefore, recalling that w and y have
nite support, the inner product
hy; G^ ()wi2 = p(1 ; : : : ; n );
where p(
; : : : ;
) is some multinomial. Multinomials satisfy a maximum
principle; specically p satises
max jp(1 ; : : : ; n )j = max jp( ; : : : ; n )j:
2T 1
k 2D k
Thus there exist numbers 01 ; : : : ; 0n on the unit circle T so that
jp(01 ; : : : ; 0n )j  jp(1 ; : : : ; n )j > : (11.19)
Let 0 be the operator, of form (11.16), that corresponds to the sequence
f1; 01 ; : : : ; 0n ; 1; : : : g. Observe that by Lemma 11.15 we have kG^ (I )k =
1
operators of the form ZA are commonly known as weighted shifts
344 11. Further Topics: Synthesis
kG^ (0 )k. Also note that G^ (0 ) has the same lower triangular form as G^ ()
in (11.18) and therefore
hy; G^ (0 )wi2 = p(01 ; : : : ; 0n ):
Thus by (11.19) the inequality jhy; G^ (0 )wi2 j > holds.
Now certainly kG^ (0 )k  jhy; G^ (0 )wi2 j and hence kG^ (0 )k > ; also
recall that kG^ (I )k = kG^ (0 )k. But this is a contradiction since by denition
= kG^ (I )k. 
In the sequel we primarily work with the system function when  = I ,
where  is a complex scalar. Observe by dening the notation
G^ () := C (I ; ZA);1 ZB + D
this specialized function G^ () looks and acts very much like the transfer
function of an LTI system, and therefore plays an instrumental role in our
viewpoint in the next section.

11.2.3 Evaluating the `2 induced norm

The previous section showed that the induced norm of a linear time varying
system was given by the maximum of an operator norm over a complex
ball. In this section, our primary goal is to show that this can be recast
into a convex condition on the system matrices. We will see that the results
derived appear very similar to those we have already seen for time invariant
systems, and indeed the methodology parallels it closely.
To start we state the following technical lemma.
Lemma 11.17. The following conditions are equivalent
(i) sup2D kC (I ; ZA);1 ZB + Dk < 1 and rad(ZA) < 1;
(ii) There exists X 2 L(`2 ), which is self-adjoint and X > 0, such that
ZA ZB  X 0 ZA ZB ; X 0 < 0:
      

C D 0 I C D 0 I (11.20)
This is an operator version the matrix KYP lemma. It does not depend on
the structure of A, B , C or D, or the presence of the operator Z .
For comparison, the corresponding standard result for linear time in-
variant discrete time systems is now stated from Proposition 8.28. Given
a system G with transfer function G^ (z ) := C0 (I ; zA0 );1 zB0 + D0 in a
minimal realization, the H1 norm of G is less than 1, if and only if, there
exists a matrix X0 > 0 such that
A0 B0  X0 0 A0 B0 ; X0 0 < 0:
      

C0 D0 0 I C0 D0 0 I
 11.2. A Framework for Time Varying Systems: Synthesis and Analysis 345
Thus the above LTV result looks very similar to the time invariant result
just stated.
In Lemma 11.17 the variable X does not necessarily have a block-diagonal
structure, it is simply self-adjoint and positive denite. Our next goal is
therefore to improve upon this and obtain a formulation in which the vari-
able is block-diagonal. To this end dene the set X to consist of positive
denite self-adjoint operators X of the form
2 3
X0 0
6 X1 7
X = 664 X2
7
7
5
> 0; (11.21)
...
0
where the block structure is the same as that of the operator A. With this
denition we can state the main result of this section.
Theorem 11.18. The following conditions are equivalent
(i) kC (I ; ZA);1 ZB + Dk < 1 and 1 62 spec(ZA);
(ii) There exists X 2 X such that
ZA ZB  X 0 ZA ZB ; X 0 < 0:
      

C D 0 I C D 0 I (11.22)
Formally, the result is the same as that for the linear time invariant case,
but the operators ZA and ZB replace the usual A-matrix and B -matrix,
and X is block-diagonal. We shall see in the sequel that this is a general
property of this formalism, and that this gives a simple way to construct
and to understand the relationship between time invariant and time varying
systems.

Proof . We start by invoking Theorem 11.14 and Lemma 11.16 with  :=

I : condition (i) above is equivalent to condition (i) in Lemma 11.17. There-
fore it suces to show that (ii) above is equivalent to (ii) in Lemma 11.17.
Also, a solution X 2 X to (11.22) immediately satises (ii) in Lemma 11.17
with X := X .
It only remains to show that a solution X to (11.20) implies that there
exists X 2 X satisfying (11.22), which we now demonstrate. Suppose X 2
L(`2 ) is self-adjoint, and satises both X > 0 and (11.22). Our goal is to
construct X 2 X from X and show that it has the desired properties.
Dene the operator Ek = [0|
{z

0} I 0


] , for k  0, mapping Rn ! `2
k zeros
which then satises
Ek A = [0


0 Ak 0


]:
346 11. Further Topics: Synthesis
Observe that Ek Ek = I . Using Ek dene X to be the block-diagonal
operator `2 ! `2 corresponding to the sequence dened by
Xk = Ek XE
 k ; for each k  0.
Thus, X is a block-diagonal operator, whose elements are the blocks on the
diagonal of X . Clearly X is self-adjoint and satises X > 0 because X has
these properties. This proves X 2 X .
To complete the proof we must now demonstrate that X satises (11.22).
Grouping Z in (11.22) with X we apply Proposition 11.10 to see that
(11.22) holds if and only if the permuted inequality
A B  Z  XZ 0 A B ; X 0 < 0
      

C D 0 I C D 0 I
holds. Now we can apply Proposition 11.11 to show that the above is
tantamount to
A B  Z  XZ 0 A B ; X 0 < 0:
      

C D 0 I C D 0 I (11.23)
We will now show that this inequality is satised.
Observe that, for each k  0, the following holds 2
Ek C = [0


0 Ck 0


]:
Now using the fact Ek Ek = I it is routine to verify the important property
      
A B Ek 0 = Ek 0 A B
C D 0 Ek 0 Ek C D k holds, (11.24)
for each k  0.
Since X by assumption satises (11.20) there exists a  > 0 such that
A B  Z  XZ
  
 0 A B  X 0
C D 0 I C D ; 0 I < ;I:
Pre and post multiply this by diag(Ek ; Ek ) and diag(Ek ; Ek ) respectively,
and use (11.24) to get that the matrix inequality
A B  Ek 0 Z  XZ
   
 0 Ek 0  A B 
C D k 0 Ek 0 I 0 Ek C D k
  
E 0  0 Ek 0 
X
; 0 E  0 I 0 Ek < ;I
k
k

2
here we do not distinguish between versions of Ek that dier only in the spatial
dimension of the identity block
 11.2. A Framework for Time Varying Systems: Synthesis and Analysis 347
holds, for every k  0. Finally use the denition of X to see that this last
inequality is exactly
A B  Z  XZ 0 A B ; X 0 < ;I;
       

C D k 0 I k C D k 0 I k (11.25)
for each k  0. This immediately implies that inequality (11.23) is satised.

The following corollary relates the innite dimensional linear matrix
inequality to the pointwise properties of the system matrices.
Corollary 11.19. The following conditions are equivalent
(i) kC (I ; ZA);1 ZB + Dk < 1 and 1 62 spec(ZA);
(ii) There exists a sequence of matrices Xk > 0, bounded above and below,
such that the inequality
Ak Bk  Xk+1 0
      
Ak Bk ; Xk 0 < 0;
C k Dk 0 I Ck Dk 0 I
holds uniformly.
Proof . The result follows immediately from equation (11.25) in the proof
of Theorem 11.18 using the fact that (Z  XZ )k = Xk+1 . 
In this section we have developed an analysis condition for evaluating
the induced norm of an LTV system. In this framework the condition looks
formally equivalent to the KYP lemma for LTI systems.

11.2.4 LTV synthesis

The previous two sections have developed a framework for dealing with
LTV systems and provided analysis results of the form we require to ap-
ply the approach of Chapter 7 to solve the LTV synthesis problem. The
synthesis problem here is, given a discrete linear time-varying system, we
would like to nd a controller such that the closed-loop is contractive. In
the results of the previous section we saw that, using the framework devel-
oped, it was possible to perform the analysis for the time-varying case by
following directly the methods for the time-invariant case.
Let the LTV system G be dened by the following state space equations
xk+1 = Ak xk + B1k wk + B2k uk x0 = 0
zk = C1k xk + D11k wk + D12k uk (11.26)
yk = C2k xk + D21k wk :
We make the physical and technical assumption that the matrices
A; B; C; D are uniformly bounded functions of time. The only restrictions
on this system are that the direct feedthrough term D22 = 0. This is a
348 11. Further Topics: Synthesis
simple condition which is easy to ensure during implementation of such a
system.
We suppose this system is being controlled by an LTV controller K
characterized by
xKk+1 = AKk xKk + BkK yk
(11.27)
uk = CkK xKk + DkK yk :
Here we use n to dentote the number of states of G and n is the number
of states of K . The connection of G and K is shown in Figure 11.3.

z w
G

Figure 11.3. Closed-loop system

We write the realization of the closed loop system as

xLk+1 = ALk xLk + BkL wk
(11.28)
zk = CkL xLk + DkL wk ;
where xk contains the combined states of G and K ,and ALk , BkL , CkL and
DkL are appropriately dened.
We are only interested in controllers K that both stabilize G and provide
acceptable performance as measured by the induced norm of the map w 7!
z . The following denition expresses our synthesis goal. A controller K is
an admissible synthesis for G in Figure 11.3, if
 the spectral condition 1 62 spec(ZAL ) holds;
 the closed-loop performance kw 7! z k`2!`2 < 1 is achieved.
Hence, recalling Proposition 11.13 we are requiring the closed-loop system
dened by equations (11.28) be exponentially stable, in addition to be-
ing strictly contractive. We have the following theorem, which is written
entirely in terms of block-diagonal operators and the shift.
Theorem 11.20. There exists an admissible synthesis K for G, with state
dimension m  n, if and only if there exist block-diagonal operators Y > 0
and X > 0 satisfying
 11.2. A Framework for Time Varying Systems: Synthesis and Analysis 349
2 3
  ZAY A Z  ; Y ZAY C1 B1 N 0
N Y
(i) 0 I 0 4 
C1 Y A Z  C1 Y C1 ; I D11 5 0Y I < 0
B1 
D11 ;I
2   3

N 0
 A Z XZA ; X A Z  XZB1 C1 N 0
(ii) 0X I 4 B1 Z  XZA B1 Z XZB1 ; I D11 5 X <
C 1 D ;I 0 I
11
0
 
(iii) YI XI  0
where the operators NY , NX satisfy
 
Im NY = ker B2 Z  D12
 NY NY = I
 
Im NX = ker C2 D21 NX NX = I:
This makes the correspondence with the time-invariant case, and the for-
mulae of Chapter 7 clear; formally one can simply replace the A-matrix by
ZA and the B -matrix by ZB in the latter to arrive at the former.
The above synthesis theorem can be derived routinely by using the ma-
chinery developed in x11.2.1 and x11.2.2, and the KYP lemma of x11.2.3.
This is done simply by following the methodology used in the time-invariant
case in Chapter 7. Further, this solution has the important property of be-
ing convex. This oers not only powerful computational properties, but
also gives insight into the structure of the solution.

11.2.5 Periodic systems and nite dimensional conditions

The synthesis condition stated in Theorem 11.20 is in general innite di-
mensional, as is the analysis condition of Theorem 11.18. However there
are two important cases in which these results reduce to nite dimensional
convex problems. The rst is when one is only interested in behavior on the
nite horizon. In this case the matrix sequences Ak ,Bk ,Ck and Dk would be
chosen to be zero for k  N the length of the horizon. Thus the associated
analysis and synthesis inequalities immediately reduce to nite dimensional
conditions. The second major case that reduces occurs when the system G
is periodic, and explaining this connection is the purpose of this section.
An operator P on `2 is said to be q-periodic if
Z q P = PZ q ;
namely it commutes with q shifts. Throughout the sequel we x q  1 to
be some integer.
Before stating the next result we require some additional notation. Sup-
pose Q is a q-periodic block-diagonal operator, then we dene Q~ to be the
350 11. Further Topics: Synthesis
rst period truncation of Q, namely
2 3
Q0 0
Q~ := 64 ... 7
5 ;
0 Qq;1
which is a matrix. Also dene the cyclic shift matrix Z~, for q  2, by
2 3
0


0 I 2
Q1 0
3
6 .. 7
...
Z~ = 66 I . . 0 77 ; such that Z~  Q~ Z~ = 666
6 7
7
. 7 7:
4 .. .. 5 4 Qq;1 5
I 0 0 Q 0

For q = 1 set Z~ = I . Also dene the truncation of the set X , dened in

(11.21), by
X~ := fX~ : X 2 Xg:
Using these new denitions we have the following theorem which is a
periodic version of Theorem 11.18.
Theorem 11.21. Suppose A, B, C and D are q-periodic operators. The
following conditions are equivalent
(i) kC (I ; ZA);1 ZB + Dk < 1 and 1 62 spec(ZA);
(ii) There exists a matrix X~ 2 X~ such that
Z~ A~ Z~B~  X~ 0 Z~ A~ Z~ B~ ; X~ 0 < 0:
      

C~ D~ 0 I C~ D~ 0 I (11.29)

Thus this result gives an LMI condition to determine the `2 induced norm
of a periodic system of the form in (11.14). Notice that the statement of
this theorem simply involves replacing all the objects in Theorem 11.18
with their \wave" equivalent; for instance A now appears as A~. We have
the following synthesis theorem which mirrors this pattern exactly.
Theorem 11.22. Suppose G has a q-periodic realization. Then an admis-
sible controller of order n  n exists if and only if there exist solutions to
the inequalities of Theorem 11.20, where A, B , C , D, X , Y and Z , are
replaced by the block-matrices dened by A~, B~ , C~ , D~ , X~ , Y~ and Z~
Given this correspondence one wonders whether a nite dimensional
system function can be dened. The answer is yes. Dene
G~ (~ ) := C~ (I ; ~ Z~A~);1 ~ Z~B~ + D:
~
 11.2. A Framework for Time Varying Systems: Synthesis and Analysis 351
Here the matrix ~ is dened by
2 3
0 0
~ = 64 ... 7
5 ;
q;1
with the k being complex scalars. Analogs of our recent results on the
system function exist for this periodic systems function.
This brings us to the end of our quick look at LTV systems. The frame-
work presented here reduces general time varying problems to solutions in
terms of structured operator inequalities; these inequalities follow from the
standard LTI problems we have studied in earlier chapters. We have explic-
itly illustrated this by applying the new tools to deriving a KYP lemma
for LTV systems, and then provided the corresponding synthesis results
without proof. These general results become bona de LMI problems when
periodic or nite time horizon systems are being considered. In summary
the main feature of the framework is that it makes direct ties to the stan-
dard LMI techniques for time invariant systems, and thus many results and
derivations become formally equivalent.

Notes and references

The work in x11.1 on NMD systems is based on related ideas and concepts
from [77, 89, 7]. The rst of these papers studies LPV control for stabiliza-
tion only, whereas the second includes performance. The paper [7] focuses
on model reduction of NMD systems.
The presentation and technical machinery for LTV systems developed in
x11.2 are based on [31]. For the Lyapunov theory required on exponential
stability in this section, see for example [137]. In Theorem 11.14 a max-
imum principle for multinomials is required, and can be found in [104];
this monograph also provides many useful details on complex functions of
several variables. A proof of the Lemma 11.17 can be found in [145].
Problems involving time varying systems have been extensively studied
and there is a large literature. For references in a Riccati equation context
see [50] and the references therein. In particular the papers [4, 60] use
techniques related to the one presented here.
 This is page 352
Printer: Opaque this

AppendixA
Some Basic Measure Theory

In this appendix we will make precise some of the terminology used in

Chapter 3. Specically we target the terms \for almost every" and \essen-
tial supremum". These terms are introduced because many mathematical
properties of functions (of a real variable) hold in nearly all instances, but
fail on very small subsets of the real line; for this reason it is worthwhile
to specify rigorously what we mean by small sets. What we present here
is a very brief treatment aimed only at providing basic understanding of
these terms and the required concepts from measure theory. The subject
of measure theory is vast, and we refer the reader to [140] for more details
at an introductory level, and to [53] for a more advanced treatment of the
subject.

A.1 Sets of zero measure

We will conne ourselves to considering subsets of the real line R, with the
specic goal of classifying the smallest of these. To do this our rst task is
to investigate some of the basic notions of size or measure 1 of sets. Suppose
we have two subsets S1 and S2 of R, and S2  S1 . Then clearly S2 is no
larger than S1 and our usual intuition about measuring size must preserve
this. The notion we aim to make precise is, when will we say that these

1
We will restrict ourselves to Lebesgue measure.
 A.1. Sets of zero measure 353
sets are the same size. Clearly this question distills down to:
When is the dierence S1 nS2 small?
Said yet another way, when do we say that the size of the dierence between
the sets is insignicant? Thus in order to answer this question we need to
precisely dene what we mean by insignicant in size. We will introduce
the idea of a set having zero size, or in mathematical language having zero
measure.
The rst thing we require is a denition for the size of an open interval
(a; b) in R. We dene the size of this set to be b ; a, which is exactly our
usual notion of length. Generalizing this, suppose we have a collection of n
disjoint intervals
(a1 ; b1 ); (a2 ; b2 ); : : : ; (an ; bn );
and dene the associated set
G = (a1 ; b1 ) [ (a2 ; b2 ) [ : : : [ (an ; bn )
to be their union. Since the intervals are disjoint, our usual intuition about
size would dictate that the size of G should be additively based on these
interval lengths. We therefore accordingly dene the size of this set to be
n
X
(bk ; ak ) = size of G :
k=1
With these denitions in place, consider a subset S which is contained
in such a union:
S  [nk=1 (ak ; bk ):
Then if S has some size associated with it we would want this size to satisfy
n
X
size of S  (bk ; ak ):
k=1
Notice that this bound will remain true regardless of whether or not these
intervals are disjoint, we therefore proceed assuming that they are not
necessarily disjoint. Let us generalize this idea to a countable number of
intervals. Suppose that
S  [1k=1 (ak ; bk ):
Then if we have a size associated with S , we would naturally conclude that
1
X
size of S  (bk ; ak ) must hold:
k=1
If the series on the right converges we have an upper bound on the possible
size, or measure, of the set S . Of course if the series diverges then the above
inequality gives us no information about the set S .
354 AppendixA. Some Basic Measure Theory
Having done a little exploration above about measuring sets, we are now
ready to dene a set of zero measure.
Denition A.1. A subset S  R has zero measure if, for every  > 0,
there exists a countable family of intervals (ak ; bk ) such that the following
conditions hold:
(a) The set S is a subset of the union [1 k=1 (ak ; bk );
P1
(b) The sum k=1 (bk ; ak ) < .
Given our discussion above this denition can be interpreted as follows:
a set S has zero measure if the upper bound on its size can be made as
small as desired. To see how this denition applies we consider two simple
examples.
Examples:
First we consider the simplest nonempty sets in R, those containing one
element; let S = ftg be such a set. For  > 0 this set is contained in the
interval (t ; 2 ; t + 2 ), whose length is . Thus directly from the denition
this set has zero measure. More intuitively this construction says that
size S  :
Therefore S has zero size, since the above inequality holds for any  > 0.
Using the same argument it is not dicult to show that any set comprised
of a nite number of points ft1 ; : : : ; tn g has zero measure.
Let us now turn to the set of natural numbers N = f1; 2; 3; : : : g. This
set contains a countably innite number of points, yet we will now see that
it too has zero measure. Set  to be any number satisfying 0 <  < 1, and
dene the intervals
 k
(1 ; 
(ak ; bk ) = k ; 2 ; k + 2)  k (1 ;  )  ;
for each k 2 N . Since k 2 (ak ; bk ), for each k > 0, we see that
N  [1 k=1 (ak ; bk ):
The length of each interval (ak ; bk ) is (1 ; )k , and so
1
X 1
X
(bk ; ak ) = (1 ; )k = ;
k=1 k=1
where we have used the geometric series formula. From our denition above
we conclude that N has zero measure; its size is clearly smaller than any
. We leave as an exercise the extension of this example to show that the
integers Z are also a subset of R that has zero measure. Similarly it is pos-
sible to show that any countable subset of the real line has measure zero.
In particular the set of rational numbers Q is of zero measure; this fact
 A.2. Terminology 355
is perhaps surprising at rst glance since the rationals are so densely dis-
tributed on the real line. The examples we have given here of zero measure
sets all have a nite or countable number of elements; not all sets of zero
measure are countable, but constructing them is more involved. 

A.2 Terminology
Having introduced the denition of a set of zero measure, we can explain
the meaning of the term \for almost every". Suppose that P (t) is a logical
condition which depends on the real variable t. Then recall that a statement
\For every t 2 R the condition P (t) holds"
means that, for any chosen value t0 2 R, the condition P (t0 ) is true. Then
we dene the following terminology.
Denition A.2. Given a logical condition P (t), which depends on the real
variable t, the expression \For almost every t 2 R the condition P (t) holds",
means that the set
S = ft0 2 R : P (t0 ) is falseg has zero measure.
This denition states that \for almost every" means that the condition P (t)
can fail for some values of t, provided that it only fails on a very small set
of points. Put more precisely, the set S of points where the condition P (t)
is false has zero measure. Notice that this means that \for every" implies
for \for almost every" but the converse is not true; namely the former is
the stronger condition. To further see the implications of this terminology
we consider some examples.
Examples:
Consider the function f (t) = sin2 t. This function does not satisfy f (t) >
0, for all t 2 R, since the positivity condition fails when t is an integer.
Since we know Z is a set of measure zero it follows that
f (t) > 0, for almost all t 2 R
For the purpose of another example consider the function

d(t) = 10;; for t 2 Q;
for t 62 Q :
Then we see that d(t) = 0, for almost all t. Further given any function g(t),
it follows from the properties of d that that (g d)(t) = 0, for almost all t.

So far we have assumed that P (t) is dened on R, however it not un-
common for logical conditions to depend on subsets of the real line, and
356 AppendixA. Some Basic Measure Theory
we therefore extend our above denition. If D is a subset of R then \For
almost every t 2 D the condition P (t) holds" is dened to mean that the
set S = ft0 2 D : P (t0 ) is falseg is of zero measure.
We can now turn to the denition of the essential supremum of a function.
Denition A.3. Suppose that D is a subset of R, and the function f :
R ! R. The essential supremum of the function over D is dened by
ess sup f (t) = inf f 2 R : the function f (t) <  , for almost every t 2 D.g
t2D
In other words a function is never greater than its essential supremum,
except on a set of measure zero, and the essential supremum is the smallest
number that has this property. Thus we immediately see that the essential
supremum of a function can never be greater than the supremum. The
basic property which makes the essential supremum useful is that it ignores
values of the function that are only approached on a set of zero measure.
Again we look at some concrete examples to make this denition clear.
Examples:
Dene the function h : [0; 1) ! R by

h(t) = e;t ; for t > 0;
2; for t = 0:
Then according the denitions of supremum and essential supremum we
have
sup h(t) = 2 and ess sup h(t) = 1:
t2[0; 1) t2[0; 1)
The distinguishing property here is that the supremum of 2 is only ap-
proached (in fact achieved) at one point, namely t = 0, whereas the function
is otherwise less than one, so the essential supremum can be no greater than
one. However for any value of  < 1, the set of points for which h(t)   is
never of zero measure for it always contains an interval. Thus the essential
supremum is indeed one.
Recall the function d(t) just dened above. It satises
sup d(t) = 1 and ess sup d(t) = 0:
t2R t2R
To see this simply realize that d(t) is only near the value one on a set of
zero measure, namely the rational numbers; otherwise it is always equal to
zero. In fact, given any function g(t) we have that ess sup(g d)(t) = 0.
Finally we leave as an exercise the verication of the fact that if f (t) is a
continuous function, then its supremum is equal to its essential supremum.

 A.3. Comments on norms and Lp spaces 357
A.3 Comments on norms and Lp spaces
To end this appendix we discuss how sets of measure zero play a role in
dening the elements in an Lp (;1; 1) space. We begin by focusing on
L1 and an example.
Example:
Let f (t) be the function that is zero at every time, and then clearly kf k1 =
0. Also dene the function g by

g(t) = 10;; tt 262 ZZ:
From the above discussion of the essential supremum we know that kgk1 =
0. Thus we have that f and g are functions in L1 which both have norm
zero. In fact it is clear that we can dene many dierent functions with
zero innity norm.

This example seems to indicate that k
k1 is not a norm, since it violates
the requirement that only one element can have zero norm. What is needed
to reconcile this dichotomy is a reinterpretation of what we mean by an
element of L1 :
Functions that dier only on a set of measure zero are
considered to represent the same element.
Thus in our example above f and g both represent the zero element in L1 .
Furthermore if h and w are L1 functions, and satisfy kh ; wk1 = 0, then
they represent the same element. Thus strictly speaking the elements of
L1 are not functions but instead sets of functions, where each set contains
functions that are equivalent.
We now generalize to Lp spaces, for 1  p < 1. Recall that the norm is
dened by
Z 1  p1
khkp = p
jh(t)jp dt :
;1
It is a fact that if h is zero everywhere except on a set of zero measure,
then
khkp = 0:
That is function values on a measure zero set do not contribute to the
integral.2 Thus we see that the two example functions f and g given above
have zero norm in every Lp space, and as a result we cannot rigorously
2
This is based on Lebesgue integration theory.
358 AppendixA. Some Basic Measure Theory
regard them as distinct. So just as in L1 we regard any functions that
dier only on a set of measure zero as representing the same element in Lp .
In doing this all the mappings k
kp indeed dene norms.
To conclude we emphasize that for our purposes in this course the dis-
tinction between functions and elements of an Lp space is not crucial, and
elements of Lp spaces can be viewed as functions without compromising
understanding.
 This is page 359
Printer: Opaque this

AppendixB
Proofs of Strict Separation

This appendix presents two technical proofs which were omitted in Chap-
ters 8 and 9, part of the argument to establish necessity of scaled small-gain
conditions for robustness. Specically we will prove two propositions which
concerned the strict separation of the sets r and  in Rd . These were
dened as
 = f(r1 ; : : : ; rd ) 2 Rd : rk  0; for each k = 1; : : : ; dg;
r = f(1 (q); : : : ; d (q)) 2 Rd : q 2 L2 satisfying kqk2 = 1g:
We recall that k (q) = kEk Mqk2 ; kEk qk2 , M is the nominal LTI
system under consideration, and the projection matrices Ek break up
signals in components conformably with the uncertainty structure
=
diag(
1 ; : : : ;
d ):
In what follows, L2 [a; b] denotes the subspace of functions in L2 [0; 1)
with support in the interval [a; b], and P[a;b] : L2[0; 1) ! L2 [a; b] is the
natural projection. We now state the rst pending result.
Proposition B.1 (Proposition 8.9, Chapter 8). Suppose that (M;
) a
is robustly well-connected. Then the sets  and r are strictly separated, i.e.
D(; r) := r2inf;y2r jr ; yj > 0:
We will give a proof by contrapositive, based on the following key lemma.
Lemma B.2. Suppose D(r; ) = 0. Given any  > 0 and any t0  0 the
following conditions can be satised:
360 AppendixB. Proofs of Strict Separation
1. There exists a closed interval [t0 ; t1 ], and two functions p; q 2
L2 [t0 ; t1 ], with kqk = 1, such that
kEk pk  kEk qk; for each k = 1; : : : ; d: (B.1)
 > k(I ; P[t0 ;t1 ] )Mqk
2 (B.2)
p
 d = kp ; P[t0 ;t1 ] Mqk (B.3)
2. With the above choice of [t0 ; t1 ] and q, there exists an operator
=
diag(
1 ; : : : ;
d ) in L(L2 [t0 ; t1 ]) \
, such that k
k  1 and
p
a

;

k I ;
P[t0 ;t1 ] M qk   d: (B.4)
Proof . Fix  > 0 and t0  0. By hypothesis, there exists q 2 L2, kqk = 1,
satisfying k (q) > ;2 for each k = 1; : : : ; d. This amounts to
2 + kEk Mqk2 > kEk qk2 ; for each k = 1; : : : ; d:
Now clearly if the support of q is truncated to a suciently long interval,
and q is rescaled to have unit norm, the above inequality will still be satis-
ed by continuity of the norm. Also since Mq 2 L2 , by possibly enlarging
this truncation interval we can obtain [t0 ; t1 ] satisfying (B.2), and also
2 + kEk P[t0 ;t1 ] Mqk2 > kEk qk2 ; for each k = 1; : : : ; d:
Next choose  2 L2[t0 ; t1 ] such that Ek  has norm  and is orthogonal to
Ek P[t0 ;t1 ] Mq, for each k = 1; : : : ; d. Then dene
p = P[t0 ;t1 ] Mq + :
p
Now kk =  d so (B.3) follows, and also
kEk pk2 = 2 + kEk P[t0 ;t1 ]Mqk2 > kEk qk2 ; for every k = 1; : : : ; d;
which proves (B.1) and completes Part 1.
For Part 2, we start from (B.1) and invoke Lemma 8.4, Chapter 8 (notice
that it holds in any L2 space), to construct a contractive, block diagonal

satisfying
p = q. Then
;
;

I ;
P[t0 ;t1 ]M q =
p ; P[t0 ;t1 ] Mq
so (B.4) follows from (B.3). 
Proof . (Proposition B.1) The argument is by contrapositive: we assume
that D(r; ) = 0, the objective is to construct a perturbation
2
a
such that I ;
M is singular.
Fix any positive sequence n ! 0 as n tends to 1. For each n, we
construct q(n) and
(n) as in Lemma B.2. Since their supports can be
shifted arbitrarily, we choose them to be of the form [tn ; tn+1 ], with t0 =
0, so that these intervals form a complete partition of [0; 1). Now we
can combine the
(n) 2 L(L2 [tn ; tn+1 ]) \
to construct a single
2
a
 AppendixB. Proofs of Strict Separation 361
L(L2 [0; 1)), dened by
1
X

=
(n) P[tn ;tn+1] : (B.5)
n=1
Descriptively, this operator breaks up a signal u into its components in the
time partition [tn ; tn+1 ], applies
(n) to each \piece" P[tn ;tn+1] u, and puts
the resulting pieces back together. It is easy to see that k
k  1, since all
the
(n) are contractive. Furthermore
inherits the block-diagonal spatial
structure so
2
. a
Now apply
to the signal Mq(n) for a xed n. We can write


Mq(n) =
P[tn ;tn+1] + (I ; P[tn ;tn+1 ] ) Mq(n)
=
(n) P[tn ;tn+1] Mq(n) +
(I ; P[tn ;tn+1 ])Mq(n) ;
Applying the triangle inequality this leads to
 
k (I ;
M ) q(n) k  k I ;
(n) P[tn ;tn+1] M q(n) k + k(I ; P[tn ;tn+1 ] )Mq(n) k
p
 n d + 2n ;
where we have used (B.4) and (B.2) respectively. Now we let n ! 1
to see that the right hand side tends to zero, and thus so does the left
hand side. Therefore I ;
M cannot have a bounded inverse since for
each n we know by denition that kq(n)k = 1. This contradicts robust
well-connectedness. 
We turn now to our second result which states that if we restrict ourselves
to the causal operators in
, our rst result still holds.
a

Proposition B.3 (Proposition 9.8, Chapter 9). Suppose (M;
) is a;c

robustly stable. Then D(; r) > 0.
As compared to Proposition B.1, the hypothesis has now changed to state
that I ;
M has a causal, bounded inverse, for every causal
2
. a

This means that we would already have a proof by contradiction if the

we constructed in the previous proposition were causal. Looking more
closely, we see that the issue is the causality of each term
(n) P[tn ;tn+1] ; un-
fortunately, the basic construction of
(n) mapping p(n) to q(n) in Lemma
B.2 cannot guarantee causality inside the interval [tn ; tn+1 ]. Obtaining the
desired causality requires a more rened argument.
Lemma B.4. Suppose D(r; ) = 0. Given  = p1n > 0 and t0  0 there
exist:
(i) an interval [t0 ; t~1 ];
(ii) a signal q~ 2 L2[t0 ; t~1 ], kq~k = 1;
(iii) a contractive operator
~ in L(L2 [t0 ; t~1 ]) \
, with
~ P[t0 ;t~1 ] causal,
a
362 AppendixB. Proofs of Strict Separation
satisfying
k(I ; P[t0 ;t~1 ] )M q~k  p1n (B.6)
k I ;
~ P[t0 ;t~1 ] M q~k  pn
 
(B.7)
for some constant .
Before embarking on the proof of this lemma, we observe that it suces
to prove Proposition B.3. In fact, we can repeat the construction of (B.5)
and obtain
which is now causal and makes I ;
M singular. The latter
fact is established by using (B.6) and (B.7) instead of (B.2) and (B.4) , for
n = 1=pn.
Therefore we concentrate our eorts in proving Lemma B.4.
Proof . We rst invoke Lemma B.2 to construct an interval [t0; t1], func-
tions p; q and an operator
with the stated properties. For simplicity, we
will take t0 = 0 from now on; it is clear that everything can be shifted
appropriately. Also we denote h = t1 .

q
q~
p
p~
0 h nh nh + h

Figure B.1. Signals q and p (dashed); q~ and p~ (solid).

An illustration of the functions q and p is given by the broken line in
Figure B.1. Notice that in this picture p appears to have greater norm than
q, but this \energy" appears later in time; this would preclude a causal,
contractive
from mapping p to q.
To get around this diculty, we introduce a periodic repetition (n = 1=2
times) of the signals p and q, dening
n ;1
nX
q~ = p1n Sih q; p~ = p1n
X
Sih p;
i=1 i=0
where S denotes time shift as usual. The signals are sketched in Figure
B.1. Notice that there is an extra delay for q~; this is done deliberately so
 AppendixB. Proofs of Strict Separation 363
that p~ anticipates its energy to q~. Also, the normalizing factor is added to
ensure kq~k = 1. Both signals are supported in [0; t~1 ] where t~1 = (n + 1)h.
Now we introduce the operator
;1
nX

~ = S(i+1)h
S;ih P[ih;(i+1)h]
i=0
Notice that each term of the above sum truncates a signal to [ih; (i + 1)h],
shifts back to the interval [0; h], applies
we had obtained from Lemma
B.2, and shifts it again forward to [(i + 1)h; (i + 2)h] (i.e. by one extra
interval of h). A little thought will convince the reader that

~ maps L2 [0; (n + 1)h] to itself;
 Since
is contractive, so is
;~
 Since
p = q, then
~ ~ p = q~.
We claim that
~ is causal. By denition, this means that PT
~ PT = PT
~
for all T , where PT denotes truncation to [0; T ]; the only non-trivial case
here is when T 2 [0; (n + 1)h]. In particular assume that
i0 h < T  (i0 + 1)h;
for some integer i0 between 0 and n. First observe that
0 ;1
iX
PT
~ = PT S(i+1)h
S;ih P[ih;(i+1)h] ; (B.8)
i=0
since the remaining terms in the sum
~ have their image supported in
[(i0 + 1)h; 1). For the terms in (B.8) we have (i + 1)h  i0 h < T so
P[ih;(i+1)h] PT = P[ih;(i+1)h] :
Therefore multiplying (B.8) on the right by PT is inconsequential, i.e.
PT
~ PT = PT
.~
It only remains to show that the given
~ and q~ satisfy (B.6) and (B.7).
We rst write
n
k(I ; P )M q~k  p1 k(I ; P
X
[0;t~1 ] )S Mqk
[0;(n+1)h] ih
n i=1
n
 p1n
X
k(I ; P[ih;(i+1)h] )Sih Mqk
i=1
=p 1 n
X
n i=1 kSih (I ; P[0;h] )Mqk
p
= n k(I ; P[0;h] )Mqk  p1n : (B.9)
364 AppendixB. Proofs of Strict Separation
The rst step relies on the time invariance of M , and the last bound follows
from (B.2), since 2 = 1=n. This proves (B.6).
To prove (B.7) it suces to show that for some constant ,
kp~ ; P[0;t~1 ] M q~k  pn ; (B.10)
in this case contractiveness of
~ gives (B.7) because
   

~ p~ ; P[0;t~1 ] M q~ = I ;
~ P[0;t~1 ]M q~:
We thus focus on (B.10); this bound is broken in two parts, the rst is


1 n
S P Mq  p1 ;
X

P M q~ ; p

[0;t~1 ] n i=1 ih [0;h] (B.11)

n
and its derivation is left as an exercise since it is almost identical to (B.9).
The second quantity to bound is


1 Xn

p~ ; p


S P Mq =
n i=1 ih [0;h]


;1
nX
p1n p ; Snh P[0;h]Mq + Sih (p ; P[0;h] Mq) : (B.12)
i=1
Notice that we have isolated two terms inside the norm sign on the right
hand side, since the sums we are comparing have slightly dierent index
ranges; these two terms have a bounded norm, however, since kqk = 1, M
is a bounded operator, and p is close to Mq because of (B.2 ). As for the
last sum, the terms have disjoint support so
2
n ;1 ;1
nX
X


Sih (p ; P[0;h] Mq ) = kp ; P[0;h]Mqk2  (n ; 1)2 d  d;
i=1 i=1
where we invoked (B.3). This means p that right hand side of (B.12) is
bounded by some constant times 1= n. Now combining this with (B.11)
we have (B.10), concluding the proof. 

Notes and references

The preceding proofs follow the ideas in [122], that in particular proposed
the periodic repetition method to construct a causal, destabilizing
.
 This is page 365
Printer: Opaque this

AppendixC
-Simple Structures

This appendix is devoted to the proof of Theorem 8.27 which characterizes

the uncertainty structures for which the structured singular value is equal
to its upper bound, i.e.
(M;
s;f ) = ;2inf
;
 (;M ;;1):
s;f
We will focus on showing that the condition 2s + f  3 is sucient for the
above; the references can be consulted for counterexamples in the remaining
cases.
Clearly due to the scalability of  and its upper bound it suces to show
that
(M;
s;f ) < 1 implies ;2inf;
 (;M ;;1 ) < 1:
s;f
This step is fairly technical and found in very few places in the literature;
our treatment here is based on the common language of quadratic forms,
developed in Chapter 8.
We recall the denition of the sets
rs;f := f(1 (q); : : : ; s (q); s+1 (q); : : : ; s+f (q)) : q 2 C m ; jqj = 1g
s;f := f(R1 ; : : : ; Rs ; rs+1 ; : : : ; rs+f ); Rk = Rk  0; rk  0g;
where the quadratic functions
k (q) =Ek Mqq M  Ek ; Ek qq Ek ;
k (q) =q M  Ek Ek Mq ; q Ek Ek q ;
366 AppendixC. -Simple Structures
were used to characterize the uncertainty structure. In particular, we
showed in Propositions 8.25 and 8.26 that:
 (M;
s;f ) < 1 if and only if rs;f and s;f are disjoint.
 inf ;2;s;f  (;M ;;1 ) < 1 if and only if co(rs;f ) and s;f are disjoint.
Thus our problem reduces to establishing that when 2s + f  3, if rs;f
and s;f are disjoint, then co(rs;f ) and s;f are also disjoint. Unfortu-
nately this cannot be established by invoking convexity of rs;f , which in
general does not hold; thus a specialized argument is required in each case.
We will concentrate our eorts in the \extreme" cases (s; f ) = (1; 1)
and (s; f ) = (0; 3). These suce to cover all cases since if the bound is
exact for a certain structure, it must also be exact with fewer uncertainty
blocks of each type. This can be shown by starting with a smaller problem,
e.g. (s; f ) = (0; 2), then dening an augmented problem with an extra
uncertainty block which is \inactive", i.e. the added blocks of the M matrix
are zero. Then the result for the larger structure can be invoked; we leave
details to the reader.
The two key cases are covered respectively in Sections C.1 and C.2.

C.1 The case of
1; 1

Let us start our proof by writing the partition
 
M= M11 M12
M21 M11
in correspondence with the two blocks in
1;1 . A rst observation is that
if (M11 )  1, then there exists a complex scalar  satisfying jj  1, such
that
    
I 0 ; M11 M12 I 0 is singular.
0 I M21 M11 0 0
Now the matrix on the right is a member of
1;1 and has a maximum
singular value of at most one, and therefore we see that (M;
1;1 )  1.
This means that if (M;
1;1 ) < 1, then necessarily (M11 ) < 1 is satised.
We will therefore assume as we proceed that the latter condition holds.
We now recast our problem in terms of the sets r1;1 and 1;1 . It will
also be convenient to introduce the subset of 1;1 given by
1;1 := f(0; r) : 0 2 Sm1; r 2 R; r  0g:
We can state the main result:
Theorem C.1. Suppose that (M11) < 1. The following are equivalent:
(a) co(r1;1 ) and 1;1 are disjoint;
 C.1. The case of
;
1 1 367
(b) r1;1 and 1;1 are disjoint;
(c) r1;1 and 1;1 are disjoint;
(d) co(r1;1 ) and 1;1 are disjoint.
Clearly what we are after is the equivalence of (a) and (b), which by Propo-
sitions 8.25 and 8.26 implies the equality of the structured singular value
and its upper bound. The other two steps will be convenient for the proof.
An important comment is that the result is true even though the set r1;1
is not in general convex.
Let us now examine these conditions. Condition (a) obviously implies
all the others; also (c) is immediately implied by all the other conditions.
Therefore to prove the theorem it is therefore sucient to show that (c)
implies (a). We do this in two steps. First we show that (d) implies (a)
in Lemma C.2 below; and then nally the most challenging part, that (c)
implies (d), is proved in Lemma C.6 of the sequel. Having made these
observations we are ready to begin proving the theorem, which starts with
the following lemma.
Lemma C.2. Suppose that (M11) < 1. If co(r1;1) and 1;1 are disjoint,
then co(r1;1 ) and 1;1 are disjoint.
Proof . Start by noting that co(r1;1 ) and 1;1 are disjoint convex sets in V,
with co(r1;1 ) compact and 1;1 closed. Hence they are strictly separated
by a hyperplane; namely there exists a symmetric matrix ; and a real
number such that
Tr(;1 (q)) + 2 (q) <   r for every q; jqj = 1; and every r  0:
It follows that  0, since   r for all positive numbers r; and there-
fore that we can choose  = 0 in the above separation. Now analogously
to the proof of Proposition 8.26 the rst inequality can be rewritten as
   
M  ;0 I0 M ; ;0 I0 < 0:
Using the partition for M , the top-left block of this matrix inequality is
M11 ;M11 + M21 M21 ; ; < 0
and therefore
M11 ;M11 ; ; < 0:
This is a discrete time Lyapunov inequality, so using the hypothesis
(M11 ) < 1 we conclude that ; > 0. Now this implies that
Tr(;R) + r  0
for every (R; r) 2 1;1 , and therefore the hyperplane strictly separates
co(r1;1 ) and 1;1 .

368 AppendixC. -Simple Structures
Thus we have now shown above that (d) does indeed imply (a) in The-
orem C.1. It only remains for us to demonstrate that (c) implies (d). This
is the key step in proving the theorem and will require a little prelimi-
nary work. The rst step is to obtain a more convenient characterization
of co(r1;1 ), which will allow us to bring some matrix theory to bear on our
problem.
By denition  2 r1;1 means there exists a vector q of unit norm such
that  
 =  1((qq)) :
2
Consider a convex combination of two points q and v in r1;1 .
 
  ( q ) + (1 ;
q + (1 ; )v =  (q) + (1 ; ) (v) :
1  )1 ( v )
2 2
The following is readily obtained using the transposition property of the
matrix trace:
1 (q) + (1 ; )1 (v) =E1 MWM E1 ; E1 WE1 ;
2 (q) + (1 ; )2 (v) =TrfW (M  E2 E2 M ; E2 E2 )g;
where W = qq + (1 ; )vv . Given a symmetric matrix V we dene the
extended notation
1 (V ) =E1 MV M  E1 ; E1 V E1
2 (V ) =TrfV (M  E2 E2 M ; E2 E2 )g :
Thus the above equations can be written compactly as
1 (q) + (1 ; )1 (v) = 1 (W )
2 (q) + (1 ; )2 (v) = 2 (W ) :
This leads to the following parametrization of the convex hull of r1;1 .
Proposition C.3. The convex hull co(r1;1 ) is equal to the set of points
f(1 (W ); 2 (W )) : for some r  1,
r
X r
X
W= i qi qi ; jqi j = 1; i  0 and i = 1g
i=1 i=1
We remark that in the above parametrization, W is always positive semi-
denite and nonzero. Next we prove two important technical lemmas.
Lemma C.4. Suppose P and Q are matrices of the same dimensions such
that
PP  = QQ :
Then there exists a matrix U satisfying
P = QU and UU  = I :
 C.1. The case of
;
1 1 369
Proof . Start by taking the singular value decomposition
PP  = QQ = V 2 V  :
Clearly this means the singular value decompositions of P and Q are
P = V U1 and Q = V U2 :
Now set U = U2 U1 .

We use the lemma just proved in demonstrating the next result, which is
the key to our nal step.
Lemma C.5. Suppose P and Q are both n  r matrices. If PWP  ;
QWQ = 0, for some symmetric matrix W  0, then
r
X
W= wi wi for some vectors wi ;
i=1
such that for each 1  i  r we have
0 = Pwi wi P  ; Qwi wi Q :
Proof . By Lemma C.4 there exists a unitary matrix U such that
PW 21 = QW 12 U :
Since U is unitary, there exists scalars ri on the unit circle, and orthonormal
vectors qi such that
r
X r
X
U= ri qi qi and qi qi = I :
i=1 i=1
Thus we have for each i that
PW 21 qi = QW 12 Uqi = ri QW 12 qi :
Thus we set wi = W 21 qi to obtain the desired result. 
We now complete our proof of Theorem C.1, and establish that (c)
implies (d) in the following lemma.
Lemma C.6. If the sets r1;1 and 1;1 are disjoint, then co(r1;1 ) and
1;1 are disjoint.
Proof . We prove the result using the contrapositive. Suppose that
co(r1;1 ) and 1;1 intersect. Then by Proposition C.3 there exists a nonzero
positive semi denite matrix W such that
1 (W ) = 0 and 2 (W )  0 :
By denition this means
E1 MWM E1 ; E1 WE1 = 0 and TrfW (M E2 E2 M ; E2 E2 )g  0
370 AppendixC. -Simple Structures
are both satised.
Focusing on the former
P equality we see that by Lemma C.5 there exist
vectors so that W = ri=1 wi wi and for each i the following holds.
E1 Mwi wi M  E1 ; E1 wi wi E1 = 0 : (C.1)
Now looking at the inequality
TrfW (M  E2 E2 M ; E2 E2 )g  0 ;
we substitute for W to get
r
X
wi fM E2 E2 M ; E2 E2 gwi  0 :
i=1
Thus we see that there must exist a nonzero wi0 such that
wi0 fM E2 E2 M ; E2 E2 gwi0  0 :
Also by (C.1) we see that
E1 Mwi0 wi0 M  E1 ; E1 wi0 wi0 E1 = 0 :
Set q = jwwii00 j and then we have that 1 (q) = 0 and 2 (q)  0 both hold,
where jqj = 1. This directly implies that r1;1 and 1;1 intersect, which
completes our contrapositive argument. 
With the above lemma proved we have completely proved Theorem C.1.

C.2 The case of
0; 3

We begin by reviewing the denition of the set r0;3 , namely
r0;3 = f(q H1 q; q H2 q; q H3 q) : q 2 C m ; jqj = 1g  R3 ;
where
Hk := Mk Ek Ek Mk ; Ek Ek 2 H m for k = 1; 2; 3:
In fact this structure for the Hk will be irrelevant to us from now on.
The proof hinges around the following key Lemma.
Lemma C.7. Given two distinct points x and y in r0;3, there exists an
ellipsoid E in R3 which contains both points and is a subset of r0;3 .
By an ellipsoid we mean here the image through an ane mapping of
the unit sphere (with no interior)
S = fx 2 R3 : x21 + x22 + x23 = 1g:
In other words, E = fv0 + Tx : x 2 Sg for some xed v0 2 R3 , T 2 R33 .
 C.2. The case of
;
0 3 371
Proof . Let
x = (qx H1 qx ; qx H2 qx ; qx H3 qx );
y = (qy H1 qy ; qy H2 qy ; qy H3 qy );
where qx ; qy 2 C m and jqx j = jqy j = 1. Since x 6= y, it follows that the
vectors qx and qy must be linearly independent, and thus the matrix
Q := qx qy  qx qy > 0:
   

Now consider the two by two matrices

H~ k := Q; 21 qx qy  Hk qx qy Q; 12 ; k = 1; 2; 3;
   

and dene the set

E := f( H~ 1 ;  H~ 2 ;  H~ 3 ) :  2 C 2 ; jj = 1g:
We have the following properties:
 E  r0;3 . In fact  H~ k  = q Hk q for q = qx qy Q; 21 , and if
 

jj = 1 it follows from the denition of Q that

 
2
qx qy Q; 12  =  Q; 12 qx qy  qx qy Q; 21  = 1:
     


 x; y 2 E . Taking
 
x = Q 21 1
0
we have qx qy Q; 21 x = qx , and also
 


   
x x = 10 Q 10 = qx qx = 1:
An analogous construction holds for y.
 E is an ellipsoid. To see this, we rst parametrize the generating 's
by
 
 = r re1j' ;
2
where r1  0, r2  0, r12 + r22 = 1, and ' 2 [0; 2). Notice that we
have made the rst component real and positive; this restriction does
not change the set E since a complex factor of unit magnitude applied
to  does not aect the value of the quadratic forms  H~ k .
We can also parametrize the valid r1 , r2 and write
 
 = sin(cos( 2 ) ;  2 [0; ]; ' 2 [0; 2):
 )ej'
2
372 AppendixC. -Simple Structures
Now setting
 
H~ k = ak bk
bk ck
we have
       
 H~ k  = ak cos2 2 + ck sin2 2 + 2 sin 2 cos 2 Re(bk ei' ):
Employing some trigonometric identities and some further manipu-
lations, the latter is rewritten as
2 3
a k + c k

a k ; c k
 cos()
 H~ k  = 2 + 2 Re(bk ) ; Im(bk ) 4sin() cos(')5 :
sin() sin(')
Collecting the components for k = 1; 2; 3 we arrive at the formula
2 3
cos()
v = v0 + T 4sin() cos(')5 ;
sin() sin(')
3 3
where v0 2 R and T 2 R are xed, and  and ' vary respec-
3
tively over [0; ] and [0; 2). Now we recognize that the above vector
varies precisely over the unit sphere in R3 (the parametrization corre-
sponds to the standard spherical coordinates). Thus E is an ellipsoid
as claimed.

The above Lemma does not imply that the set r0;3 is convex; indeed
such an ellipsoid can have \holes" in it. However it is geometrically clear
that if the segment between two points intersects the positive orthant 0;3 ,
the same happens with any ellipsoid going through these two points; this
is the direction we will follow to establish that co(r0;3 ) \ 0;3 nonempty
implies r0;3 \ 0;3 nonempty.
However the diculty is that not all points in co(r0;3 ) lie in a segment
between two points in r0;3 : convex combinations of more than two points
are in general required. The question of how many points are actually
required is answered by a classical result from convex analysis; see the
references for a proof.
Lemma C.8 (Caratheodory). Let K  V , where V is a d dimensional
real vector space. Every point in co(K) is a convex combination of at most
d + 1 points in K.
We will require the following minor renement of the above statement.
Corollary C.9. If K  V is compact, then every point in the boundary of
co(K) is a convex combination of at most d points in K.
 C.2. The case of
; 0 3 373
Proof . The Caratheodory result implies that for every V 2 co(K), there
exists a nite convex hull of the form
( )
Xd+1 d+1
X
cofv1 ; : : : ; vd+1 g = k vk : k  0; k = 1
k=1 k=1
with vertices vk 2 K, which contains v.
If the vk are in a lower dimensional hyperplane, then d points will suf-
ce to generate v by invoking the same result. Otherwise, every point in
cofv1 ; : : : ; vd+1 g which is generated by k > 0 for every k will be interior
to cofv1 ; : : : ; vd+1 g  co(K). Therefore for points v in the boundary of
co(K), one of the k 's must be 0 and a convex combination of d points will
suce. 
Equipped with these tools, we are now ready to tackle the main result.
Theorem C.10. If co(r0;3 ) \ 0;3 is nonempty, then r0;3 \ 0;3 is
nonempty.
Proof . By hypothesis there exists a point v 2 co(r0;3) \ 0;3; since
co(r0;3 ) is compact, such a point can be chosen from its boundary. Since
we are in the space R3 , Corollary C.9 implies that there exist three points
x, y, z in r0;3 such that
v = x + y + z 2 0;3 ;
with ,  , non-negative and  +  + = 1. Geometrically, the triangle
S (x; y; z ) intersects the positive orthant at some point v.
Claim: v lies in a segment between two points in r0;3.
This is obvious if x,y, z are aligned or if any of , , is 0. We thus focus
on the remaining case, where the triangle S (x; y; z ) is non-degenerate and
v is interior to it, as illustrated in Figure C.1.
x

y v
w z
u

Figure C.1. Illustration of the proof.

We rst write
v = x + y + z = x + ( + )  +1 (y + z ) = x + ( + )w;
374 AppendixC. -Simple Structures
where the constructed w lies in the segment L(y; z ). Now consider the ellip-
soid E  r0;3 through y and z , obtained from Lemma C.7. If it degenerates
to 1 or 2 dimensions, then w 2 E  r0;3 and the claim is proved. If not, w
must lie inside the ellipsoid E . The half line starting at x, through w must
\exit" the ellipsoid at a point u 2 E  r0;3 such that w is in the segment
L(x; u). Therefore v 2 L(x; w)  L(x; u). Since x; u 2 r0;3 , we have proved
the claim.
Now to nish the proof, we have found two points in r0;3 such that
the segment between them intersects 0;3 . The corresponding ellipsoid
E  r0;3 between these points must clearly also intersect 0;3 . Therefore
r0;3 \ 0;3 is non-empty. 
We remark that the above argument depends strongly on the dimen-
sionality of the space; an extension to e.g. four dimensions would require
a construction of four-dimensional ellipsoids going through three points,
extending Lemma C.7. In fact such extension is not possible and the result
is not true for block structures of the form
0;f with f  4.

Notes and references

The above results are basically from [23] and [90] in the structured sin-
gular value literature, although the presentation is dierent, in particular
in regard to the denitions of the r sets. Also parallel result have been
obtained in the Russian literature in terms of the \S-procedure" [147, 39].
Our proof for the case of
1; 1 follows that of [105] where the focus is
the KYP lemma, and our proof of the
0;3 case is from [92]. A reference
for the Caratheodory Theorem is [109].
 This is page 375
Printer: Opaque this

References

[1] V.M. Adamjan, D.Z. Arov, and M.G. Krein. Innite block hankel ma-
trices and related extension problems. American Mathematical Society
Translations, 111:133{156, 1978.
[2] U.M. Al-Saggaf and G.F. Franklin. An error bound for a discrete re-
duced order model of a linear multivariable system. IEEE Transactions
on Automatic Control, 32:815{819, 1987.
[3] B. D. O. Anderson and J. B. Moore. Optimal Control: Linear Quadratic
Methods. Prentice Hall, 1990.
[4] J.A. Ball, I. Gohberg, and M.A. Kaashoek. Nevanlinna-pick interpolation
for time-varying input-output maps: the discrete case. Operator Theory:
Advances and Applications, Birkauser, 56:1{51, 1992.
[5] B. Bamieh and M. Daleh. On robust stability with structured time-invariant
perturbation. Systems and Control Letters, 21:103{108, 1993.
[6] T. Basar and P. Bernhard. H1 -Optimal Control and Related Mini-Max
Design Problems: A Dynamic Game Approach. Birkhauser, 1991.
[7] C.L. Beck, J.C. Doyle, and K. Glover. Model reduction of multi-dimensional
and uncertain systems. IEEE Transactions on Automatic Control, 41:1466{
1477, 1996.
[8] V. Belevitch. Classical Network Theory. Holden-Day, 1968.
[9] H. Bercovici, C. Foias, and A. Tannenbaum. Structured interpolation
theory. Operator Theory Advances and Applications, 47:195{220, 1990.
[10] D.S. Bernstein and W. H. Haddad. LQG control with an H1 performance
bound: A Ricatti equation approach. IEEE Transactions on Automatic
Control, 34:293{305, 1989.
376 References
[11] S. Bochner and K. Chandrasekharan. Fourier Transforms. Princeton
University Press, 1949.
[12] B. Bollobas. Linear Analysis. Cambridge University Press, 1990.
[13] S.P. Boyd, El Ghoui, E. Feron, and V. Balakrishnan. Linear Matrix In-
equalities in System and Control Theory. Society for Industrial and Applied
Mathematics, 1994.
[14] R. Braatz, P. Young, J. C. Doyle, and M. Morari. Computational
complexity of  calculation. IEEE Transactions on Automatic Control,
39:1000{1002, 1994.
[15] C.A.Desoer and M. Vidyasagar. Feedback Systems: Input-Output Proper-
ties. Academic Press, 1975.
[16] B. R. Copeland and M. G. Safonov. A generalized eigenproblem solution
for singular H and H1 problems. In Control and Dynamic Systems; editor
2

C. T. Leondes. Academic Press, 1992.

[17] M. A. Dahleh and I. J. Diaz-Bobillo. Control of Uncertain Systems. Prentice
Hall, 1995.
[18] R. D'Andrea. H-innity optimization with spatial constraints. In Proc.
IEEE Conference on Decision and Control, pages 4327{4332, 1995.
[19] R. D'Andrea. Generalized l synthesis: A new framework for control design.
2

In Proc. IEEE Conference on Decision and Control, 1996. To appear in

IEEE Transactions on Automatic Control.
[20] C.A. Desoer, R.W. Liu, J. Murray, and R. Saeks. Feedback system design:
the fractional representation approach. IEEE Transactions on Automatic
Control, 25:399{412, 1980.
[21] P. Van Dooren. A generalized eigenvalue approach for solving Riccati equa-
tions. SIAM Journal of Scientic and Statistical Computing, 2:121{135,
1981.
[22] J. C. Doyle. Guaranteed margins for LQG regulators. IEEE Transactions
on Automatic Control, 23:756{757, 1978.
[23] J. C. Doyle. Analysis of feedback systems with structured uncertainty. In
IEE Proceedings, pages 242{250, 1982.
[24] J. C. Doyle, K. Glover, P. Khargonekar, and B. A. Francis. State-Space so-
lutions to standard H2 and H-innity control problems. IEEE Transactions
on Automatic Control, 34:831{847, 1989.
[25] J. C. Doyle and G. Stein. Robustness with observers. IEEE Transactions
on Automatic Control, 24:607{611, 1979.
[26] J. C. Doyle and G. Stein. Multivariable feedback design: Concepts for a clas-
sical/modern synthesis. IEEE Transactions on Automatic Control, 26:4{16,
1981.
[27] J. C. Doyle, J . E. Wall, and G. Stein. Performance and robustness analy-
sis for structured uncertainty. In Proc. IEEE Conference on Decision and
Control, 1982.
[28] J.C. Doyle. Lecture notes in advances in multivariable control. ONR/Honeywell
Workshop, 1984.
 References 377
[29] J.C. Doyle. Structured uncertainty in control system design. In Proc. IEEE
Conference on Decision and Control, 1985.
[30] G.E. Dullerud. Control of Uncertain Sampled-Data Systems. Birkhauser,
1996.
[31] G.E. Dullerud and S.G. Lall. A new approach to analysis and synthesis of
time-varying systems. In Proc. IEEE Conference on Decision and Control,
1997. To appear in IEEE Transactions on Automatic Control.
[32] P.L. Duren. Theory of Hp Spaces. Academic Press, 1970.
[33] D.F. Enns. Model reduction with balanced realizations: an error bound and
frequency weighted generalization. In Proc. IEEE Conference on Decision
and Control, 1984.
[34] M. Fan, A. Tits, and J. C. Doyle. Robustness in the presence of mixed
parametric uncertainty and unmodeled dynamics. IEEE Transactions on
Automatic Control, 36:25{38, 1991.
[35] E. Feron. Analysis of robust H performance using multiplier theory. SIAM
2

Journal of Control and Optimization, 35:160{177, 1997.

[36] E. Feron, V. Balakrishnan, S. Boyd, and L. El Ghaoui. Numerical methods
for H related problems. In Proc. American Control Conference, 1992.
2

[37] C. Foias and A.E. Frazho. The Commutant Lifting Approach to Interpola-
tion Problems. Birkhauser, 1990.
[38] C. Foias, H. Ozbay, and A. Tannenbaum. Robust Control of Innite
Dimensional Systems. Springer-Verlag, 1996.
[39] A. Fradkov and V. A. Yakubovich. The S-procedure and duality theo-
rems for nonconvex problems of quadratic programming. Vestnik Leningrad
University, 31:81{87, 1973. In Russian.
[40] B.A. Francis. Notes on introductory state space systems. Personal commu-
nication.
[41] B.A. Francis. A Course in H1 Control Theory. Springer-Verlag, 1987.
[42] P. Gahinet and P. Apkarian. A Linear Matrix Inequality approach to H1
control. International Journal of Robust and Nonlinear Control, 4:421{448,
1994.
[43] J.B. Garnett. Bounded Analytic Functions. Academic Press, 1981.
[44] T. Georgiou and M. Smith. Optimal robustness in the gap metric. IEEE
Transactions on Automatic Control, 35:673{686, 1990.
[45] E. Gilbert. Controllability and observability in multivariable control
systems. SIAM Journal of Control, 1:128{151, 1963.
[46] K. Glover. All optimal hankel-norm approximations of linear multivari-
able systems and their l1 error bounds. IEEE Transactions on Automatic
Control, 39:1115{1193, 1984.
[47] K. Glover. A tutorial on model reduction. In From Data to Model; editor
J.C. Willems. Springer-Verlag, 1989.
[48] K. Glover and D. McFarlane. Robust stabilization of normalized coprime
factor plant descriptions with H1 -bounded uncertainty. IEEE Transactions
on Automatic Control, 34:821{830, 1989.
378 References
[49] G.H. Golub and C.F. Van Loan. Matrix Computations. The Johns Hopkins
University Press, 1996.
[50] M. Green and David J.N. Limebeer. Linear Robust Control. Prentice Hall,
1995.
[51] W.H. Greub. Linear Algebra. Springer, 1981.
[52] J. Guckenheimer and P. Holmes. Nonlinear oscillations, dynamical systems
and bifurcations of vector elds. Springer, 1986.
[53] P.R. Halmos. Mesure Theory. Springer, 1974.
[54] P.R. Halmos. A Hilbert Space Problem Book. Springer-Verlag, 1982.
[55] M.L.J. Hautus. Controllability and observability conditions of linear
autonomous systems. In Proc. Kon. Ned. Akad. Wetensch. Ser. A., 1969.
[56] D. Hinrichsen and A.J. Pritchard. An improved error estimate for reduced
order models of discrete time systems. IEEE Transactions on Automatic
Control, 35:317{320, 1990.
[57] K. Homan. Banach Spaces of Analytic Functions. Prentice-Hall, 1962.
[58] R.A. Horn and C.R. Johnson. Matrix Analysis. Cambridge University Press,
1991.
[59] R.A. Horn and C.R. Johnson. Topics in Matrix Analysis. Cambridge
University Press, 1995.
[60] P.A. Iglesias. An entropy formula for time-varying discrete-time control
systems. SIAM Journal of Control and Optimization, 34:1691{1706, 1996.
[61] T. Iwasaki. Robust performance analysis for systems with norm-bounded
time-varying structured uncertainty. In Proc. American Control Confer-
ence, 1994.
[62] T. Kailath. Linear Systems. Prentice-Hall, 1980.
[63] R.E. Kalman. A new approach to linear ltering and prediction theory.
ASME Transactions, Series D: Journal of Basic Engineering, 82:35{45,
1960.
[64] R.E. Kalman. On the general theory of control systems. In Proc. IFAC
World Congress, 1960.
[65] R.E. Kalman. Mathematical descriptions of linear systems. SIAM Journal
of Control, 1:152{192, 1963.
[66] R.E. Kalman and R.S. Bucy. New results in linear ltering and predic-
tion theory. ASME Transactions, Series D: Journal of Basic Engineering,
83:95{108, 1960.
[67] D. Kavranoglu and M. Bettayeb. Characterization of the solution to the
optimal H1 model reduction problem. Systems and Control Letters, 20:99{
107, 1993.
[68] C. Kenig and P. Tomas. P. maximal operators dened by fourier multipliers.
Studia Math., 68:79{83, 1980.
[69] H.K. Khalil. Nonlinear Systems. Prentice-Hall, 1996.
[70] M. Khammash and J. B. Pearson. Performance robustness of discrete-
time systems with structured uncertainty. IEEE Transactions on Automatic
Control, 36:398{412, 1991.
 References 379
[71] P. Khargonekar and M. Rotea. Mixed H /H1 control: A convex opti-
2

mization approach. IEEE Transactions on Automatic Control, 36:824{837,

1991.
[72] P.P. Khargonekar and E. Sontag. On the relation between stable matrix
fraction factorizations and regulable realizations of linear systems over
rings. IEEE Transactions on Automatic Control, 27:627{638, 1982.
[73] V. Kucera. Discrete Linear Control: the Polynomial Approach. Wiley, 1979.
[74] H. Kwakernaak and R. Sivan. Linear Optimal Control Systems. Wiley-
Interscience, 1972.
[75] P. Lancaster and L. Rodman. Algebraic Riccati Equations. Clarendon Press,
1995.
[76] A.J. Laub, M.T. Heath, C.C. Page, and R.C. Ward. Computation of balanc-
ing transformations and other applications of simultaneous diagonalization
algorithms. IEEE Transactions on Automatic Control, 32:115{122, 1987.
[77] W. M. Lu, K. Zhou, and J. C. Doyle. Stabilization of LFT systems. In Proc.
IEEE Conference on Decision and Control, 1991.
[78] D.G. Luenberger. Observing the state of a linear system. IEEE Transac-
tions on Military Electronics, 8:74{80, 1964.
[79] D.G. Luenberger. Optimization by Vector Space Methods. Wiley, 1969.
[80] A. Megretski. Neccessary and sucient conditions of stability: A multi-
loop generalization of the circle criterion. IEEE Transactions on Automatic
Control, 38, 1993.
[81] A. Megretski and A. Rantzer. System-analysis via integral quadratic
constraints. IEEE Transactions on Automatic Control, 42:819{830, 1997.
[82] A. Megretski and S. Treil. Power distribution inequalities in optimization
and robustness of uncertain systems. JMSEC, 3, 1993.
[83] M. Mesbahi, M.G. Safonov, and G.P. Papavassilopoulos. Bilinearity and
complementarity in robust control. In Recent Advances on LMI Methods in
Control; editors L. El Ghaoui, S. Niculescu. SIAM, 1999.
[84] B.C. Moore. Principal component analysis in linear systems: controllabil-
ity, observablity, and model reduction. IEEE Transactions on Automatic
Control, 26:17{32, 1981.
[85] Y. Nesterov and A. Nemirovskii. Interior-point Polynomial Algorithms in
Convex Programming. Society for Industrial and Applied Mathematics,
1994.
[86] C.N. Nett, C.A. Jacobson, and M.J. Balas. A connection between state
space and doubly coprime fractional representations. IEEE Transactions
on Automatic Control, 29:831{832, 1984.
[87] J. G. Owen and G. Zames. Duality theory for robust disturbance
attenuation. Automatica, 29:695{705, 1993.
[88] A. Packard. What's new with : Structured Uncertainty in Multivariable
Control. PhD thesis, University of California, Berkeley, 1988.
[89] A. Packard. Gain scheduling via linear fractional transformations. Systems
and Control Letters, 22(2):79{92, 1994.
380 References
[90] A. Packard and J. C. Doyle. The complex structured singular value.
Automatica, 29(1):71{109, 1993.
[91] A. Packard, K. Zhou, P. Pandey, J. Leonhardson, and G. Balas. Opti-
mal, constant I/O similarity scaling for full-information and state-feedback
control problems. Systems and Control Letters, 19:271{280, 1992.
[92] F. Paganini. Sets and Constraints in the Analysis of Uncertain Systems.
PhD thesis, California Institute of Technology, 1996.
[93] F. Paganini. Frequency domain conditions for robust H performance.
2

IEEE Transactions on Automatic Control, 1998. To appear.

[94] F. Paganini. Convex methods for robust H analysis of continuous time
2

systems. IEEE Transactions on Automatic Control, 1999. To appear.

[95] F. Paganini and E. Feron. Lmi methods for robust h analysis: a survey
2

with comparisons. In Recent Advances on LMI Methods in Control; editors

L. El Ghaoui and S. Niculescu. SIAM, 1999.
[96] J.R. Partington. An Introduction to Hankel Operators. Cambridge Univer-
sity Press, 1988.
[97] L. Pernabo and L.M. Silverman. Model reduction by balanced state space
representations. International Journal of Control, 27:382{387, 1982.
[98] I. Petersen, D. McFarlane, and M. Rotea. Optimal guaranteed cost control
of discrete-time uncertain linear systems. In Proc. IFAC World Congress,
1993.
[99] K. Poolla and A. Tikku. Robust performance against time-varying struc-
tured perturbations. IEEE Transactions on Automatic Control, 40:1589{
1602, 1995.
[100] V. M. Popov. Absolute stability of nonlinear systems of automatic control.
Automation Remote Control, 22:857{875, 1961.
[101] V.M. Popov. The solution of a new stability problem for controlled system.
Automatic and Remote Control, 24:1{23, 1963.
[102] V.M. Popov. Hyperstability of Control Systems. Springer, 1966.
[103] S.C. Power. Hankel Operators on Hilbert Space. Pitman, 1981.
[104] R.M. Range. Holomorphic functions and integrable representations in
several complex variables. Springer, 1986.
[105] A. Rantzer. On the Kalman-Yakubovich-Popov lemma. Systems and
Control Letters, 28:7{10, 1996.
[106] A. Rantzer and A. Megretski. A convex parameterization of robustly stabi-
lizing controllers. IEEE Transactions on Automatic Control, 39:1802{1808,
1994.
[107] A. Rantzer and A. Megretski. Tutorial on integral quadratic constraints.
1999. Preprint.
[108] A. Rantzer and A. Megretskii. System analysis via integral quadratic
constraints. In Proc. IEEE Conference on Decision and Control, pages
3062{3067, 1994.
[109] T.R. Rockafellar. Convex Analysis. Princeton University Press, 1997.
 References 381
[110] M. Rosenblum and J. Rovnyak. Hardy Classes and Operator Theory. Oxford
University Press, 1985.
[111] W. Rudin. Real and Complex Analysis. McGraw-Hill, 1987.
[112] M. G. Safonov. Tight bounds on the response of multivariable systems with
component uncertainty. In Proc. Allerton, pages 451{460, 1978.
[113] M. G. Safonov and M. Athans. A multiloop generalization of the circle
criterion for stability margin analysis. IEEE Transactions on Automatic
Control, 26:415:422, 1981.
[114] M. G. Safonov and R.Y. Chiang. Real/complex Km synthesis without curve
tting. In Control and Dynamic Systems; editor C. T. Leondes. Academic
Press, 1993.
[115] M.G. Safonov. Stability and Robustness of Multivariable Feedback Systems.
MIT Press, 1980.
[116] M.G. Safonov and R.Y. Chiang. A schur method for balanced-truncation
model reduction. IEEE Transactions on Automatic Control, 45:729{733,
1989.
[117] M. Sampei, T. Mita, and M. Nakamichi. An algebraic approach to H1
output feedback control problems. Systems and Control Letters, 14:13{24,
1990.
[118] I.W. Sandberg. An observation concerning the application of the con-
traction mapping xed-point theorem and a result concerning the norm-
boundedness of solutions of nonlinear functional equations. Bell Systems
Technical Journal, 44:809{1812, 1965.
[119] C. Scherer. H-innity optimization without assumptions on nite or innite
zeros. SIAM Journal of Control and Optimization, 30:143{166, 1992.
[120] C. Scherer, P. Gahinet, and M. Chilali. Multiobjective output-feedback
control via lmi-optimization. IEEE Transactions on Automatic Control,
42:896{911, 1997.
[121] J. Shamma. The necessity of the small-gain theorem for time-varying and
nonlinear systems. IEEE Transactions on Automatic Control, 36:1138{
1147, 1991.
[122] J. Shamma. Robust stability with time varying structured uncertainty.
IEEE Transactions on Automatic Control, 39:714{724, 1994.
[123] S. Skogestad and I. Postlethwaite. Multivariable feedback control : analysis
and design. Wiley, 1996.
[124] M.W. Spong and D.J. Block . The pendubot: A mechatronic system for
control research and education. In Proc. IEEE Conference on Decision and
Control, pages 555{556, 1995.
[125] G. Stein and M. Athans. The LQG/LTR procedure for multivariable feed-
back control design. IEEE Transactions on Automatic Control, 32:105{114,
1987.
[126] A. A. Stoorvogel. The H1 Control Problem: A State Space Approach.
Prentice Hall, 1992.
[127] A. A. Stoorvogel. The robust H control problem: A worst-case design.
2

IEEE Transactions on Automatic Control, 38:1358{1370, 1993.

382 References
[128] G. Strang. Linear Algebra and its Applications. Academic Press, 1980.
[129] B. Sz.-Nagy and C. Foias. Harmonic Analysis of Operators on Hilbert Space.
North-Holland, 1970.
[130] A. L. Tits and M. K. H. Fan. On the small- theorem. Automatica, 31:1199{
1201, 1995.
[131] O. Toker and H. Ozbay. On the NP-hardness of the purely complex 
computation, analysis/synthesis, and some related problems in multidi-
mensional systems. In Proc. American Control Conference, pages 447{451,
1995.
[132] B. van Keulen. H1 Control for Distributed Parameter Systems: A State-
Space Approach. Birkhauser, 1993.
[133] L. Vandenberghe and S. Boyd. Semidenite programming. SIAM Review,
38:49{95, 1996.
[134] M. Vidyasagar. Input-output stability of a broad class of linear time invari-
ant multvariable feedback systems. SIAM Journal of Control, 10:203{209,
1972.
[135] M. Vidyasagar. The graph metric for unstable plants and robustness es-
timates for feedback stability. IEEE Transactions on Automatic Control,
29:403{417, 1984.
[136] M. Vidyasagar. Control system synthesis: A Factorization Approach. MIT
Press, 1985.
[137] M. Vidyasagar. Nonlinear Systems Analysis. Prentice-Hall, 1993.
[138] G. Vinnicombe. Frequency domain uncertainty and the graph topology.
IEEE Transactions on Automatic Control, 1993.
[139] G. Weiss. Representation of shift-invariant operators on l by h1 transfer
2

functions: an elementary proof, a generalization to lp , and a counterexample

for l1 . MCSS, 4:193{203, 1991.
[140] R.L. Wheeden and A. Zygmund. Measure and Integral. Marcel Dekker,
1977.
[141] J. C. Willems. Least squares stationary optimal control and the algebraic
Riccati equation. IEEE Transactions on Automatic Control, 16:621{634,
1971.
[142] J.C. Willems. The Analysis of Feedback Systems. MIT Press, 1990.
[143] W.M. Wonham. On pole assignment in muli-input controllable linear
systems. IEEE Transactions on Automatic Control, 12(6):660{665, 1967.
[144] W.M. Wonham. Linear Multivariable Control. Springer-Verlag, 1985.
[145] V.A. Yacubovich. A frequency theorem for the case in which the state and
control spaces are hilbert spaces with an application to some problems of
synthesis of optimal controls. Sibirskii Mat. Zh., 15:639{668, 1975. english
translation in Siberian Mathematics Journal.
[146] V. A. Yakubovich. Frequency conditions for the absolute stability of con-
trol systems with several nonlinear or linear nonstationary units. Automat.
Telemech., pages 5{30, 1967.
 References 383
[147] V. A. Yakubovich. S-procedure in nonlinear control theory. Vestnik
Leningrad Univ., pages 62{77, 1971. In Russian. English translation in
Vestnik Leningrad Univ. Math., (1977), 73-93.
[148] D.C. Youla, H.A. Jabr, and J.J. Bongiorno Jr. Modern Wiener-Hopf design
of optimal controllers: part 2. IEEE Transactions on Automatic Control,
21:319{338, 1976.
[149] N. Young. An Introduction to Hilbert Space. Cambridge University Press,
1988.
[150] P. M. Young. Robustness with Parametric and Dynamic Uncertainty. PhD
thesis, California Institute of Technology, 1993.
[151] G. Zames. On the input-output stability of time varying nonlinear feedback
systems. Part I: Conditions derived using concepts of loop gain, conicity,
and positivity. IEEE Transactions on Automatic Control, 11(2):228{238,
1966.
[152] G. Zames. Feedback and optimal sensitivity: Model reference trans-
formations, multiplicative seminorms, and approximate inverses. IEEE
Transactions on Automatic Control, 26:301{320, 1981.
[153] G. Zames and A.K. El-Sakkary. Unstable systems and feedback: the gap
metric. In Proc. Allerton, pages 380{385, 1980.
[154] G. Zames and J. G. Owen. Duality theory for mimo robust disturbance
rejection. IEEE Transactions on Automatic Control, 38:743{52., 1993.
[155] K. Zhou, J.C. Doyle, and K. Glover. Robust and Optimal Control. Prentice
Hall, 1996.
[156] K. Zhou, K. Glover, B. Bodenheimer, and J.C. Doyle. Mixed H and H1
2

performance objectives I: Robust performance analysis. IEEE Transactions

on Automatic Control, 39:1564{1574, 1994.