Sbook PDF

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/242014364
Control systems with actuator saturation: analysis and design
Book · January 2001

DOI: 10.1007/978-1-4612-0205-9
CITATIONS READS
984 3,841
2 authors, including:
Tingshu Hu
University of Massachusetts Lowell
147 PUBLICATIONS 5,907 CITATIONS
SEE PROFILE
All content following this page was uploaded by Tingshu Hu on 09 May 2016.
The user has requested enhancement of the downloaded file.

Tingshu Hu and Zongli Lin
Department of Electrical Engineering
University of Virginia
Control Systems
with Actuator Saturation:
Analysis and Design
To be published by Birkhäuser, Boston
ISBN: 0-8176-4219-6
Charlottesville, Virginia
November, 2000
To Jianping and Sylvia
T.H.
To Jian, Tony, and Vivian

Z.L.
Contents
Preface xi
1 Introduction 1
1.1 Linear Systems with Actuator Saturation . . . . . . . . . . 1
1.2 Notation, Acronyms, and Terminology . . . . . . . . . . . . 3
2 Null Controllability – Continuous-Time Systems 11

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Preliminaries and Definitions . . . . . . . . . . . . . . . . . 12
2.3 General Description of Null Controllable Region . . . . . . 15
2.4 Systems with Only Real Eigenvalues . . . . . . . . . . . . . 21
2.5 Systems with Complex Eigenvalues . . . . . . . . . . . . . . 27
2.6 Some Remarks on the Description of C(T ) . . . . . . . . . . 33
2.7 Asymptotically Null Controllable Region . . . . . . . . . . . 34
2.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3 Null Controllability – Discrete-Time Systems 37

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Preliminaries and Definitions . . . . . . . . . . . . . . . . . 38
3.3 General Description of Null Controllable Region . . . . . . 41
3.4 Systems with Only Real Eigenvalues . . . . . . . . . . . . . 44
3.5 Systems with Complex Eigenvalues . . . . . . . . . . . . . . 48
3.6 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.7 Asymptotically Null Controllable Region . . . . . . . . . . . 51
3.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4 Stabilization on Null Controllable Region –

Continuous-Time Systems 55
vii
viii CONTENTS
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2 Domain of Attraction – Planar System under
Saturated Linear Feedback . . . . . . . . . . . . . . . . . . 57
4.3 Semi-Global Stabilization – Planar Systems . . . . . . . . . 67
4.4 Semi-Global Stabilization – Higher Order Systems . . . . . . 74
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5 Stabilization on Null Controllable Region –

Discrete-Time Systems 85
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.2 Global Stabilization at Set of Equilibria –
Planar Systems . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3 Global Stabilization – Planar Systems . . . . . . . . . . . . . 99
5.4 Semi-Global Stabilization – Planar Systems . . . . . . . . . 105
5.5 Semi-Global Stabilization – Higher Order Systems . . . . . . 108
5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6 Practical Stabilization on Null Controllable Region 113

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.2 Problem Statement and Main Results . . . . . . . . . . . . 114
6.2.1 Problem Statement . . . . . . . . . . . . . . . . . . . 114
6.2.2 Main Results: Semi-Global Practical Stabilization . 114
6.3 Proof of Main Results . . . . . . . . . . . . . . . . . . . . . 115
6.3.1 Properties of the Trajectories of Second Order
Linear Systems . . . . . . . . . . . . . . . . . . . . . 115
6.3.2 Properties of the Domain of Attraction . . . . . . . 119
6.3.3 Proof of Theorem 6.2.1: Second Order Systems . . . 127
6.3.4 Proof of Theorem 6.2.1: Higher Order Systems . . . 141
6.4 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.A Proof of Lemma 6.3.1 . . . . . . . . . . . . . . . . . . . . . 149
6.B Proof of Lemma 6.3.2 . . . . . . . . . . . . . . . . . . . . . 153
7 Estimation of the Domain of Attraction under

Saturated Linear Feedback 157
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.2 A Measure of Set Size . . . . . . . . . . . . . . . . . . . . . 159
7.3 Some Facts about Convex Hulls . . . . . . . . . . . . . . . . 160
CONTENTS ix
7.4 Continuous-Time Systems under State Feedback . . . . . . 163

7.4.1 A Set Invariance Condition Based on
Circle Criterion . . . . . . . . . . . . . . . . . . . . . 164
7.4.2 An Improved Condition for Set Invariance . . . . . . 165
7.4.3 The Necessary and Sufficient Condition –
Single Input Systems . . . . . . . . . . . . . . . . . . 167
7.4.4 Estimation of the Domain of Attraction . . . . . . . 169
7.5 Discrete-Time Systems under State Feedback . . . . . . . . 173
7.5.1 Condition for Set Invariance . . . . . . . . . . . . . . 173
7.5.2 The Necessary and Sufficient Condition –
Single Input Systems . . . . . . . . . . . . . . . . . . 177
7.5.3 Estimation of the Domain of Attraction . . . . . . . 179
7.6 Extension to Output Feedback . . . . . . . . . . . . . . . . 180
7.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
8 On Enlarging the Domain of Attraction 183

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
8.2 Continuous-Time Systems . . . . . . . . . . . . . . . . . . . 183
8.3 Discrete-Time Systems . . . . . . . . . . . . . . . . . . . . . 185
8.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
9 Semi-Global Stabilization with Guaranteed

Regional Performance 195
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
9.2 Expansion of the Domain of Attraction . . . . . . . . . . . 197
9.3 Semi-Globalization – Discrete-Time Systems . . . . . . . . . 199
9.4 Semi-Globalization – Continuous-Time Systems . . . . . . . 205
9.5 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
9.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
10 Disturbance Rejection with Stability 211

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
10.2.3 Disturbance Rejection with Guaranteed Domain
of Attraction . . . . . . . . . . . . . . . . . . . . . . 216
10.2.4 An Example . . . . . . . . . . . . . . . . . . . . . . 219
x CONTENTS

10.3.3 Disturbance Rejection with Guaranteed Domain
of Attraction . . . . . . . . . . . . . . . . . . . . . . 225
10.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
11 On Maximizing the Convergence Rate 229

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
11.2.1 Maximal Convergence Control and Maximal
Invariant Ellipsoid . . . . . . . . . . . . . . . . . . . 233
11.2.2 Saturated High Gain Feedback . . . . . . . . . . . . 242
11.2.3 Overall Convergence Rate . . . . . . . . . . . . . . . 247
11.2.4 Maximal Convergence Control in the Presence
of Disturbances . . . . . . . . . . . . . . . . . . . . . 255
11.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
12 Output Regulation – Continuous-Time Systems 265

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
12.2 Preliminaries and Problem Statement . . . . . . . . . . . . 267
12.2.1 Review of Linear Output Regulation Theory . . . . 267
12.2.2 Output Regulation in the Presence of
Actuator Saturation . . . . . . . . . . . . . . . . . . 270
12.3 The Regulatable Region . . . . . . . . . . . . . . . . . . . . 271
12.4 State Feedback Controllers . . . . . . . . . . . . . . . . . . 279
12.5 Error Feedback Controllers . . . . . . . . . . . . . . . . . . 290
12.6 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
12.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
13 Output Regulation – Discrete-Time Systems 305

13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
13.2 Preliminaries and Problem Statement . . . . . . . . . . . . 306
13.2.1 Review of Linear Output Regulation Theory . . . . 306
13.2.2 Output Regulation in the Presence of
Actuator Saturation . . . . . . . . . . . . . . . . . . 307
13.3 The Regulatable Region . . . . . . . . . . . . . . . . . . . . 309
CONTENTS xi
13.4 State Feedback Controllers . . . . . . . . . . . . . . . . . . 315

13.5 Error Feedback Controllers . . . . . . . . . . . . . . . . . . 324
13.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
14 Linear Systems with Non-Actuator Saturation 327

14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
14.2 Planar Linear Systems under State Saturation –
Continuous-Time Systems . . . . . . . . . . . . . . . . . . . 328
14.2.1 System Description and Problem Statement . . . . . 328
14.2.2 Main Results on Global Asymptotic Stability . . . . 328
14.2.3 Outline of the Proof . . . . . . . . . . . . . . . . . . 330
14.3 Planar Linear Systems under State Saturation –
Discrete-Time Systems . . . . . . . . . . . . . . . . . . . . . 344
14.3.1 System Description and Problem Statement . . . . . 344
14.3.2 Main Results on Global Asymptotic Stability . . . . 344
14.3.3 Outline of the Proof . . . . . . . . . . . . . . . . . . 347
14.4 Semi-Global Stabilization of Linear Systems
Subject to Sensor Saturation . . . . . . . . . . . . . . . . . 362
14.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . 362
14.4.2 Main Results . . . . . . . . . . . . . . . . . . . . . . 363
14.4.3 An Example . . . . . . . . . . . . . . . . . . . . . . 370
14.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
Bibliography 371
Index 386
xii CONTENTS
Preface
Saturation nonlinearities are ubiquitous in engineering systems. In control

systems, every physical actuator or sensor is subject to saturation owing to
its maximum and minimum limits. A digital filter is subject to saturation if
it is implemented in a finite word length format. Saturation nonlinearities
are also purposely introduced into engineering systems such as control sys-
tems and neural network systems. Regardless of how saturation arises, the
analysis and design of a system that contains saturation nonlinearities is
an important problem. Not only is this problem theoretically challenging,
but it is also practically imperative. This book intends to study control
systems with actuator saturation in a systematic way. It will also present
some related results on systems with state saturation or sensor saturation.
Roughly speaking, there are two strategies for dealing with actuator sat-
uration. The first strategy is to neglect the saturation in the first stage of
the control design process, and then to add some problem-specific schemes
to deal with the adverse effects caused by saturation. These schemes, known
as anti-windup schemes, are typically introduced using ad hoc modifications
and extensive simulations. The basic idea behind these schemes is to intro-
duce additional feedbacks in such a way that the actuator stays properly
within its limits. Most of these schemes lead to improved performance but
poorly understood stability properties.
The second strategy is more systematic. It takes into account the sat-
uration nonlinearities at the outset of the control design. Or, in the case
that a control law is designed a priori to meet either the performance or
stability requirement, it analyzes the closed-loop system under actuator
saturation systematically and redesigns the controller in such a way that
xiii
xiv Preface
the performance is retained while stability is improved or the other way

around. This is the approach we will take in this book. Such an approach
to dealing with actuator saturation entails the characterization of the null
controllable region, the set of all states that can be driven to the origin by
the saturating actuators, and the design of feedback laws that are valid on
the entire null controllable region or a large portion of it. More specifically,
the results that are to be presented in this book are outlined as follows.
In Chapter 1, after a short introduction to linear systems with saturation
nonlinearities, in particular, actuator saturation, we list some notation and
acronyms that are used throughout the book. Some technical terms will
also be defined here.
Chapters 2 and 3 give explicit descriptions of the null controllable region
of a linear system with the bounded controls delivered by the saturating
actuators. Chapter 2 deals with continuous-time systems. Chapter 3 deals
with discrete-time systems.
Chapters 4 and 5 study the stabilizability at the origin of linear systems
with saturating actuators. The main objective is to obtain a domain of
attraction that is arbitrarily close to the null controllable region. We refer
to such a stabilization problem as semi-global stabilization on the null con-
trollable region. Chapter 4 deals with continuous-time systems. Chapter 5
deals with discrete-time systems.
Chapter 6 considers continuous-time linear systems that are subject to
both actuator saturation and input-additive bounded disturbance. The
objective is to construct feedback laws that will cause all trajectories start-
ing from within any a priori specified (arbitrarily large) compact subset of
the null controllable region to converge to another a priori specified (arbi-
trarily small) neighborhood of the origin. We refer to such a problem as
semi-global practical stabilization on the null controllable region.
Chapter 7 looks at the problem of controlling a linear system with satu-
rating actuators from a different angle. An LMI-based method is developed
for estimating the domain of attraction of a linear system under an a pri-
ori designed saturated linear feedback law. This analysis method is then
utilized in Chapter 8 to arrive at a method for designing linear state feed-
back laws that would result in the largest estimated domain of attraction.
Each of these two chapters treats both continuous-time and discrete-time
systems.
Preface xv
Chapter 9 develops a design method for arriving at simple nonlinear

feedback laws that achieve semi-global stabilization on the null control-
lable region and, in the mean time, guarantee regional performance. Both
continuous-time and discrete-time systems are considered.
Chapter 10 addresses the problem of controlling linear systems subject
to both actuator saturation and disturbance. Unlike in Chapter 6, here
the disturbance is not input additive and can enter the system from any-
where. Design problems that capture both large domains of attraction and
strong disturbance rejection capability are formulated and solved. Both
continuous-time and discrete-time systems are considered.
Chapter 11 examines the problem of maximizing the convergence rate
inside a given ellipsoid for both continuous-time and discrete-time systems
subject to actuator saturation. Simple methods are also proposed for de-
termining the largest ellipsoid of a given shape that can be made invariant
with a saturated control. For continuous-time systems, the maximal conver-
gence rate is achieved by a bang-bang type control with a simple switching
scheme. A sub-optimal convergence rate can be achieved with saturated
high-gain linear feedback. For discrete-time systems, the maximal conver-
gence rate is achieved by a coupled saturated linear feedback.
Chapters 12 and 13 formulate and solve the classical problem of output
regulation for linear systems with saturating actuators. The problem is to
design stabilizing feedback laws that, in the presence of disturbances, cause
the plant output to track reference signals asymptotically. Both the refer-
ence signals and the disturbances are modeled by a reference system, called
the exosystem. The asymptotically regulatable region, the set of all initial
conditions of the plant and the exosystem for which the output regulation
is possible, is characterized. Feedback laws that achieve output regula-
tion on the asymptotically regulatable region are constructed. Chapter 12
deals with continuous-time systems. Chapter 13 deals with discrete-time
systems.
Finally, Chapter 14 collects some results on the analysis and design of
linear systems subject to saturation other than actuator saturation. This
includes sensor saturation and state saturation.
The intended audience of this monograph includes practicing engineers
and researchers in areas related to control engineering. An appropriate
background for this monograph would be some first year graduate courses
xvi Preface
in linear systems and multivariable control. Some background in nonlinear

control systems would greatly facilitate the reading of the book.
In such an active current research area as actuator saturation, it is
impossible to account for all the available results. Although we have tried
our best to relate our work to the available research, we are still frustrated
with our inability to do a better job in this regard and will strive to improve
in future work.
We would like to thank some of our colleagues who, through collabora-
tion on the topics of this book, motivated and helped us in many ways. They
are Professor Ben M. Chen of National University of Singapore, Profes-
sor Daniel Miller of University of Waterloo, Professor Li Qiu of Hong Kong
University of Science and Technology, and Professor Yacov Shamash of
State University of New York at Stony Brook. We would also like to thank
our colleague Dr. Yong-Yan Cao for a careful reading of the manuscript.
We are indebted to Professor William S. Levine, the series editor, for
his enthusiasm and encouragement of our efforts in completing this book.
We are also thankful to the staff at Birkhäuser, in particular, Ms. Louise
Farkas, Ms. Shoshanna Grossman, and Ms. Lauren Schultz, for their excel-
lent editorial assistance.
We are grateful to the United States Office of Naval Research’s Young
Investigator Program for supporting our research that leads to most of the
results presented in this book. We are also grateful to the University of
Virginia for an environment that allowed us to write this book.
Our special thanks go to our families, {Jianping, Sylvia, (T. H.)} and
{Jian, Tony, Vivian, (Z. L.)}. Without their sacrifice, encouragement, and
support, this book would not have been completed.
This monograph was typeset by the authors using LATEX. All simula-
tions and numerical computations were carried out in MATLAB.
Charlottesville, Virginia Tingshu Hu

November 2000 Zongli Lin
Chapter 1
Introduction
1.1. Linear Systems with Actuator Saturation
Every physical actuator is subject to saturation. For this reason, the orig-
inal formulations of many fundamental control problems, including con-
trollability and time optimal control, all reflect the constraints imposed by
actuator saturation. Control problems that involve hard nonlinearities such
as actuator saturation, however, turned out to be difficult to deal with. As
a result, even though there have been continual efforts in addressing actua-
tor saturation (see [4] for a chronological bibliography on this subject), its
effect has been ignored in most of the modern control literature.
On the other hand, it has been well known that, when the actuator
saturates, the performance of the closed-loop system designed without con-
sidering actuator saturation may seriously deteriorate. In the extreme case,
even the stability may be lost. A well-known example of performance degra-
dation (e.g., large overshoot and large settling time) occurs when a linear
compensator with integrators, say a PID compensator, is used in a closed-
loop system. During the time when the actuator saturates, the error is
continuously integrated even though the control is not what it should be,
and hence, the states of the compensator attain values that lead to larger
controls than the actuator limit. This phenomenon is called integrator
windup [23]. In the absence of integrators, a given reference setpoint might
result in a different steady state response, causing the need to reset the
reference to a value different from the desired setpoint. With integral con-
trol, the controllers automatically bring the output to the given reference
1
2 Chapter 1. Introduction
setpoint and hence the integrator does the reset. For this reason, integrator
windup is sometimes referred to as reset windup.
A practical approach to compensating this performance degradation due
to actuator saturation is to add some problem-specific anti-windup schemes
to deal with the adverse effects caused by saturation. These schemes are
typically introduced using ad hoc modifications and extensive simulations.
The basic idea behind these schemes is to introduce additional feedbacks in
such a way that the actuator stays properly within its limit. For example,
several schemes have been proposed to solve the reset windup problem when
integrators are present in the forward loop. Most of these schemes lead to
improved performance but poorly understood stability properties. More
recently, some researchers have attempted to provide more systematic and
more general schemes to deal with the problem (see, e.g., [11,12,30,53,60]).
In this book, we will take the approach of considering actuator satura-
tion at the outset of control design. As seen in the recent literature, there
has been a resurge of interest in this approach, possibly owing to its system-
atic nature. A fundamental issue is the identification of the class of linear
systems that are (globally) asymptotically null controllable by bounded
controls. A system is said to be globally asymptotically null controllable
by bounded controls if, for a given bound on the controls, every state in its
state space can be driven to the origin either in a finite time or asymptot-
ically by a bounded control. In particular, it was established in [76,89,90]
that a linear stabilizable system having all its poles in the closed left-half
plane is globally asymptotically null controllable. For this reason, a linear
stabilizable system with all its poles in the closed left-half plane is com-
monly said to be asymptotically null controllable with bounded controls,
or ANCBC, and most of the recent work has been focused on ANCBC sys-
tems. For such systems, various types of feedback laws have been proposed
that work globally (on the entire state space) or semi-globally (on any a
priori given arbitrarily large bounded set in the state space). We refer the
reader to [65] and the references therein for some accounts of these results.
It is clear that a linear system having poles in the open right-half plane
is not globally asymptotically null controllable with bounded controls. Any
feedback laws designed for such a system would not work globally. Two
natural questions to ask are:
• for a general, not necessarily ANCBC, linear system with saturating

actuators, what is the asymptotically null controllable region, the
1.2. Notation, Acronyms, and Terminology 3
set of all states that can be driven to the origin in a finite time or
asymptotically by a bounded control?
• how to design feedback laws that work on the entire asymptotically

null controllable region or a large portion of it?
This book is intended to answer these two questions, for both continuous-
time and discrete-time systems. We will start with explicit analytical de-
scriptions of the asymptotically null controllable region. Once we have
identified this region, we will design feedback laws that achieve various
closed-loop performance specifications on the entire asymptotically null
controllable region or a large portion of it. These performances range from
the basic control problem of stabilization to those beyond large domain of
attraction such as transient properties, disturbance rejection and output
regulation. A full list of the topics to be addressed in this book can be
found in the preface to this book, or in the table of contents.
Throughout this book, we will assume that the saturation function is

symmetric with respect to the origin and has a unity saturation level. Such
a saturation function is commonly known as the standard saturation func-
tion. We will use sat : R → R to denote the standard saturation function,
i.e.,
sat(s) = sign(s) min 1, |s| . (1.1.1)
With a slight abuse of notation and for simplicity, for a vector u ∈ Rm , we

will also use the same sat(u) to denote the vector saturation function, i.e.,
T
sat(u) = sat(u1 ) sat(u2 ) · · · sat(um ) . (1.1.2)
While assuming unity of saturation level is without loss of generality, the

asymmetry in the saturation function needs further consideration. Some
efforts in the treatment of asymmetry of saturation function have been
made in [46].
1.2. Notation, Acronyms, and Terminology

We shall adopt the following notation and acronyms:
R := the set of real numbers,

Rn := the set of n-dimensional real vectors,
Rn×n := the set of n × n real matrices,

N := the set of natural numbers,
C := the entire complex plane,
C− := the open left-half complex plane,
C+ := the open right-half complex plane,
C0 := the imaginary axis in the complex plane,
C := the set of complex numbers inside the unit circle,
C⊗ := the set of complex numbers outside the unit circle,
C◦ := the unit circle in the complex plane,
[a, b] := the closed real interval,
[K1 , K2 ] := the set of integers from K1 to K2 ,
[p1 , p2 ] := {λp1 + (1 − λ)p2 : 0 ≤ λ ≤ 1} for two vectors p1 and p2 ,
√
j := −1,
C := the null controllable region,

C a := the asymptotically null controllable region,
C(T ) := the null controllable region at time T (continuous-time),
C(K) := the null controllable region at step K (discrete-time),
R := the reachable region,
R(T ) := the reachable region at time T (continuous-time),
R(K) := the reachable region at step K (discrete-time),
Rg := the regulatable region,
Rag := the asymptotically regulatable region,
Rg (T ) := the regulatable region at time T (continuous-time),
Rg (K) := the regulatable region at step K (discrete-time),
|x| := the Euclidean norm, or 2-norm, of x ∈ Rn ,

|x|∞ := max |xi | for x ∈ Rn ,
i
|X| := the 2-norm of a matrix X,
x∞ := the L∞ -norm of a continuous-time signal x(t)
or the l∞ -norm of a discrete-time signal x(k),

I := an identity matrix,
Ik := an identity matrix of dimension k × k,
X T := the transpose of a matrix X,
det(X) := the determinant of a square matrix X,
λ(X) := the set of eigenvalues of a square matrix X,
λmax (X) := the maximum eigenvalues of X when λ(X) ⊂ R,
λmin (X) := the minimum eigenvalues of X when λ(X) ⊂ R,
D := the set of diagonal matrices with the diagonals being 1
or 0,
Di := an element of D,
Di− := I − Di ,
αX := {αx : x ∈ X } for a positive number α and a set X ,

co(X ) := the convex hull of a set X ,
int(X ) := the interior of a set X ,
X := the closure of a set X ,
∂X := the boundary of a set X ,
Ext(X ) := the set of extremal points of a convex set X ,
A(X ) := the area of a set X ⊂ R2 ,
dist(x, X ) := the distance from a point x to a set X ,
dist(X 1 , X 2 ) := the Hausdorff distance between two sets X 1 and X 2 ,
E(P, ρ) := the ellipsoid {x ∈ Rn : xT P x ≤ ρ},

E(P ) := E(P, 1),
L(F ) := the linear region of the saturated feedback sat(F x),
E c := the set of extremal controls,

Em
c := a minial representative of E c ,
E c (K) := the set of extremal controls on [0, K] (discrete-time),

Ua := the set of admissible controls,
sign(x) := the sign function of x ∈ R,

sat(x) := the scalar or vector saturation function of x,
ANCBC := asymptotically null controllable with bounded controls,

ARE := algebraic Riccati equation,
DARE := discrete-time algebraic Riccati equation,
LMI := linear matrix inequality,
LQR := linear quadratic regulator,
PLC := piecewise linear control,
2 := the end of a proof.
We will also use the following terminology.
• For a symmetric matrix P ∈ Rn×n , we use P > 0 and P ≥ 0 to denote

that P is positive definite and positive semi-definite, respectively.
Similarly, we use P < 0 and P ≤ 0 to denote that P is negative
definite and negative semi-definite, respectively.
• For two symmetric matrices P1 , P2 ∈ Rn×n , P1 > P2 means that

P1 − P2 > 0 and P1 ≥ P2 means that P1 − P2 ≥ 0. Similarly, P1 < P2
means that P1 − P2 < 0 and P1 ≤ P2 means that P1 − P2 ≤ 0.
• A continuous-time control signal u is said to be admissible if

u ∈ Ua = u : u is measurable and |u(t)|∞ ≤ 1, ∀ t ∈ R .
Similarly, a discrete-time control signal u is said to be admissible if
|u(k)|∞ ≤ 1, ∀ k ∈ N.
• For x ∈ Rn and y ∈ Rm , we sometimes, by an abuse of notation,

write (x, y) instead of [xT , y T ]T .
• For a positive definite matrix P ∈ Rn×n and a positive real number

ρ, an ellipsoid E(P, ρ) is defined as

E(P, ρ) = x ∈ Rn : xT P x ≤ ρ .
In the case that ρ = 1, we will simply use E(P ) to denote E(P, 1).
Also, for a continuous function V : Rn → R+ , a level set LV (c) is
defined as
LV (c) := x ∈ Rn : V (x) ≤ c .
• A continuous-time linear system,
ẋ = Ax + Bu,
is said to be
– asymptotically stable (or simply stable) if λ(A) ⊂ C− ,

– neutrally stable if λ(A) ⊂ C− ∪ C0 and every eigenvalue of A on
C0 has Jordan blocks of size one,
– semi-stable if λ(A) ⊂ C− ∪ C0 ,
– exponentially unstable if λ(A)
⊂ C− ∪ C0 , and
– anti-stable if λ(A) ⊂ C+ .
In each of these situations, we will also say that the matrix A is sta-
ble (or Hurwitz), neutrally stable, semi-stable, exponentially unstable
and anti-stable, respectively.
• A discrete-time linear system,
x(k + 1) = Ax(k) + Bu(k),
is said to be
– asymptotically stable (or simply stable) if λ(A) ⊂ C ,

– neutrally stable if λ(A) ⊂ C− ∪ C◦ and every eigenvalue of A on
C◦ has Jordan blocks of size one,
– semi-stable if λ(A) ⊂ C ∪ C◦ ,
– exponentially unstable if λ(A)
⊂ C ∪ C◦ , and
– anti-stable if λ(A) ⊂ C⊗ .
In each of these situations, we will also say that the matrix A is stable
(or Schur), neutrally stable, semi-stable, exponentially unstable and
anti-stable, respectively.
• By stabilizability, unless otherwise specified, we mean asymptotic sta-

bilizability at the origin, i.e., the existence of a feedback law such that
the closed-loop system is asymptotically stable at the origin. Simi-
larly, by global (semi-global, local) stabilization we mean it in the
asymptotic sense.
• Let X 1 and X 2 be two bounded subsets of Rn . Their Hausdorff

distance is defined as

1 , X 2 ), d(X
dist(X 1 , X 2 ) := max d(X 2, X 1) ,
where
1 , X 2 ) = sup
d(X inf |x1 − x2 |.
x 1 ∈X 1 x 2 ∈X 2
Here the vector norm used is arbitrary.
• A set X is convex if, for any two points x1 , x2 ∈ X ,
λx1 + (1 − λ)x2 ∈ X ,
for all 0 ≤ λ ≤ 1.
• For a convex set X ⊂ Rn , a point x0 ∈ X is said to be an extremal

point (or simply, an extreme) of X if there exists a vector c ∈ Rn
such that
cT x0 > cT x, ∀ x ∈ X \ {x0 }.
Intuitively, the hyperplane cT x = cT x0 divides the space into two

parts and the set X \ {x0 } lies completely in one part. We use Ext(X )
to denote the set of all the extremal points of X .
• A convex set X is said to be strictly convex if all of its boundary

points are extremal points, i.e., Ext(X ) = ∂X . It is strictly convex if
and only if for any two points x1 , x2 ∈ ∂X ,
λx1 + (1 − λ)x2 ∈
/ ∂X
for all λ ∈ (0, 1). For example, an ellipsoid is strictly convex while a
cube is not.
• The convex hull of a set X is the minimal convex set that contains X .
For a group of points x1 , x2 , · · · , xI ∈ Rn , the convex hull of these
points is
I
I

co x1 , x2 , · · · , xI = αi xi : αi = 1, αi ≥ 0 .
i=1 i=1
• For a matrix F ∈ Rm×n , denote the ith row of F as fi and define

L(F ) := x ∈ Rn : |fi x| ≤ 1, i = 1, 2, · · · , m .
If F is the feedback matrix, then L(F ) is the region where the feedback
control u = sat(F x) is linear in x. We call L(F ) the linear region of
the saturated feedback sat(F x), or simply, the linear region of the
saturation function.
Chapter 2
Null Controllability –
Continuous-Time
Systems
2.1. Introduction
This chapter studies null controllability of continuous-time linear systems

with bounded controls. Null controllability of a system refers to the possi-
bility of steering its state to the origin in a finite time by an appropriate
choice of the admissible control input. If the system is linear and is control-
lable, then any state can be steered to any other location, for example, the
origin, in the state space in a finite time by a control input. This implies
that any controllable linear system is null controllable from any point in
the state space. The control inputs that are used to drive certain states
to the origin, however, might have to be large in magnitude. As a result,
when the control input is limited by the actuator saturation, a linear con-
trollable system might not be globally null controllable. In this situation,
it is important to identify the set of all the states that can be steered to
the origin with the bounded controls delivered by the actuators. This set
is referred to as the null controllable region and is denoted as C.
Let C(T ) denote the set of states that can be steered to the origin in a
finite time T > 0 with bounded controls. Then, the null controllable region
C is given by the union of C(T ) for all T ∈ (0, ∞).
11
12 Chapter 2. Null Controllability – Continuous-Time Systems
In the earlier literature, the null controllable region, also called the
controllable set, was closely related to the time optimal control (see, e.g.,
[18,52,63,76]). For a given initial state x0 , the time optimal control problem
has a solution if and only if x0 ∈ C. If x0 is on the boundary of C(T ), then
the minimal time to steer x0 to the origin is T . The corresponding time
optimal control is a bang-bang control. In recent literature on control of
linear systems with saturating actuators, the characterization of the null
controllable region forms the guideline for searching feedback laws (see, e.g.,
[65,95]).
Because of the fundamental role the null controllable region plays in con-
trol theory for linear systems with bounded controls, there have been contin-
ual efforts towards its characterization (see, e.g., [3,4,18,21,52,54,62,89,90]
and the references therein). In this chapter, we will present simple and ex-
plicit descriptions of the null controllable region. Our presentation is based
on both the earlier understanding of the null controllable region and our
recent work on the topic [41].
Section 2.2 contains some preliminaries from which a complete descrip-
tion of the null controllable region can be made. Section 2.3 presents a
general description of the null controllable region. Although explicit and
complete, this general description can be drastically simplified by utiliz-
ing the eigenstructure of the given system. Sections 2.4 and 2.5 show how
such simplification is made. In particular, Section 2.4 deals with the case
that all eigenvalues are real, and Section 2.5 deals with the case of com-
plex eigenvalues. Section 2.6 explains how C(T ) can be described. Section
2.7 addresses the situation where the linear system contains some uncon-
trollable, but stable, poles. In this case, the notion of asymptotically null
controllability is introduced. Finally, a brief concluding remark is made in
Section 2.8.
2.2. Preliminaries and Definitions
Consider a linear system

ẋ = Ax + Bu, (2.2.1)
where x ∈ Rn is the state and u ∈ Rm is the control. Let

Ua = u : u is measurable and |u(t)|∞ ≤ 1, ∀ t ∈ R , (2.2.2)
2.2. Preliminaries and Definitions 13
where |u(t)|∞ = maxi |ui (t)|. A control signal u is said to be admissible

if u ∈ Ua . We are interested in the control of the system (2.2.1) by using
admissible controls. Our first concern is the set of states that can be steered
to the origin by admissible controls.
Definition 2.2.1.
1) A state x0 is said to be null controllable in time T > 0 if there exists

an admissible control u such that the state trajectory x(t) of the
system satisfies x(0) = x0 and x(T ) = 0. The set of all states that
are null controllable in time T , denoted by C(T ), is called the null
controllable region at time T .
2) A state x0 is said to be null controllable if x0 ∈ C(T ) for some T ∈

[0, ∞). The set of all null controllable states, denoted by C, is called
the null controllable region of the system.
With the above definition, we see that x0 ∈ C(T ) if and only if there
exists an admissible control u such that
T
AT
0 = e x0 + eA(T −τ ) Bu(τ )dτ
0

T
= eAT x0 + e−Aτ Bu(τ )dτ .
0
It follows that

T
−Aτ
C(T ) = x=− e Bu(τ )dτ : u ∈ Ua . (2.2.3)
0
The minus sign “−” before the integration can be removed since Ua is
symmetric. Also, we have

C= C(T ). (2.2.4)
T ∈[0,∞)
In what follows we recall from the literature some existing results on

the characterization of the null controllable region.
Proposition 2.2.1. Assume that (A, B) is controllable.
a) If A is semi-stable, then C = Rn ;
b) If A is anti-stable, then C is a bounded convex open set containing

the origin;
c) If
A1 0
A= ,
0 A2
with A1 ∈ Rn1 ×n1 anti-stable and A2 ∈ Rn2 ×n2 semi-stable, and B
is partitioned accordingly as

B1
B= ,
B2
then,
C = C 1 × Rn2 ,
where C 1 is the null controllable region of the anti-stable system
ẋ1 = A1 x1 + B1 u.
Statement a) can be found in [76,89,90]. Statements b) and c) are proven

in [29]. Because of this proposition, we can concentrate on the study of null
controllable regions of anti-stable systems. For an anti-stable system,
∞
C= x= e−Aτ Bu(τ )dτ : u ∈ Ua , (2.2.5)
0
where C denotes the closure of C. We will also use “∂” to denote the
boundary of a set. In this chapter, we will derive a method for explicitly
describing ∂C.
Let
B = b1 b2 · · · bm ,
and, for each i = 1 to m, let the null controllable region of the system
ẋ = Ax + bi ui
be C i , then it is clear that

m
C= Ci = x1 + x2 + · · · + xm : xi ∈ C i , i = 1, 2, · · · , m . (2.2.6)
i=1
In view of (2.2.6) and Proposition 2.2.1, in the study of null controllable re-
gions we will assume, without loss of generality, that (A, B) is controllable,
A is anti-stable, and m = 1. For clarity, we rename B as b.
2.3. General Description of Null Controllable Region 15
As will become apparent shortly, the concept of time-reversed system

will facilitate our characterization of the null controllable region. For a
general dynamical system
ẋ = f (x, u), (2.2.7)
its time-reversed system is
ż = −f (z, v). (2.2.8)
It is easy to see that x(t) solves (2.2.7) with x(0) = x0 , x(t1 ) = x1 and
certain u if and only if z(t) = x(t1 − t) solves (2.2.8) with z(0) = x1 ,
z(t1 ) = x0 and v(t) = u(t1 − t). The two systems have the same curves as
trajectories, but traverse in opposite directions.
Consider the time-reversed system of (2.2.1),
ż = −Az − bv. (2.2.9)
Definition 2.2.2.
1) A state zf is said to be reachable in time T if there exists an admissible

control v such that the state trajectory z(t) of the system (2.2.9)
satisfies z(0) = 0 and z(T ) = zf . The set of all states that are
reachable in time T , denoted by R(T ), is called the reachable region
at time T .
2) A state zf is said to be reachable if zf ∈ R(T ) for some T ∈ [0, ∞).

The set of all reachable states, denoted by R, is called the reachable
region of the system (2.2.9).
It is known that C(T ) and C of (2.2.1) are the same as R(T ) and R
of (2.2.9) (see, e.g., [76]). To avoid confusion, we will continue to use the
notation x, u, C(T ) and C for the original system (2.2.1), and z, v, R(T )
and R for the time-reversed system (2.2.9).
2.3. General Description of Null Controllable Region
In this section, we will show that the boundary of the null controllable
region of a general anti-stable linear system with saturating actuators is
composed of a set of extremal trajectories of its time-reversed system. The
description of this set will be further simplified for systems with only real
poles and for systems with complex poles in Sections 2.4 and 2.5, respec-
tively.
We will characterize the null controllable region C of the system (2.2.1)
through studying the reachable region R of its time-reversed system (2.2.9).
Since A is anti-stable, we have
∞
R= z= e−Aτ bv(τ )dτ : v ∈ Ua
0
0
= z= eAτ bv(τ )dτ : v ∈ Ua .
−∞
Noticing that eAτ = e−A(0−τ ) , we see that a point z in R is a state of

the time-reversed system (2.2.9) at t = 0 by applying an admissible control
v from −∞ to 0.
Theorem 2.3.1.
0

∂R = z = eAτ b sign cT eAτ b dτ : c
= 0 . (2.3.1)
−∞
R is strictly convex. Moreover, for each z ∗ ∈ ∂R, there exists a unique

admissible control v ∗ such that
0
∗
z = eAτ bv ∗ (τ )dτ. (2.3.2)
−∞
Proof. First, the convexity of R can easily be verified by definition. Let

z ∗ ∈ ∂R. Then, there exists a nonzero vector c ∈ Rn such that
0
cT z ∗ = max cT z = max cT eAτ bv(τ )dτ. (2.3.3)
z∈R v∈Ua −∞
Since c
= 0 and (A, b) is controllable, cT eAτ b
≡ 0. Since cT eAt b has a finite
number of zeros in any finite interval,

µ t : cT eAt b = 0 = 0, (2.3.4)
where µ(·) denotes the measure of a set.

It is easy to see that
v ∗ (t) = sign(cT eAt b)

maximizes the right hand side of (2.3.3). We maintain that v ∗ is the unique
optimal solution of (2.3.3). To verify this, we need to show that for any
v ∈ Ua , v
= v ∗ ,
0 0
cT eAτ bv ∗ (τ )dτ > cT eAτ bv(τ )dτ. (2.3.5)
−∞ −∞
Since v
= v ∗ , there are a set E1 ⊂ [−∞, 0] with nonzero measure, i.e.,
µ(E1 ) = δ1 > 0, and a number ε1 > 0 such that
|v(t) − v ∗ (t)| ≥ ε1 , ∀ t ∈ E1 .
By (2.3.4), there exist a set E ⊂ E1 , with µ(E) = δ > 0, and a positive

number ε > 0 such that
T At
c e b ≥ ε, ∀ t ∈ E.
Noting that v ∈ Ua , we have
cT eAt b (v ∗ (t) − v(t)) ≥ 0, ∀ t ∈ [−∞, 0].
It then follows that

0
cT eAτ b (v ∗ (τ ) − v(τ )) dτ ≥ cT eAτ b (v ∗ (τ ) − v(t)) dτ
−∞
E
= |cT eAτ b| |v ∗ (τ ) − v(τ )|dτ
E
≥ δεε1 > 0.
This shows that v ∗ (t) is the unique optimal solution of (2.3.3) and hence
the unique admissible control satisfying
0
∗
z = eAτ bv ∗ (τ )dτ. (2.3.6)
−∞
On the other hand, if

0
∗

z = eAτ b sign cT eAτ b dτ
−∞
for some nonzero c, then obviously
cT z ∗ = max cT z.
z∈R
This shows that z ∗ ∈ ∂R and we have (2.3.1).

Since for each c
= 0, the optimal solution v ∗ (t) and z ∗ of (2.3.3) is
unique, we see that R is strictly convex. 2
Theorem 2.3.1 says that for z ∗ ∈ ∂R, there is a unique admissible
control v ∗ satisfying (2.3.2). From (2.3.1), this implies that

v ∗ (t) = sign cT eAt b
for some c
= 0 (such c, |c| = 1, may be non-unique). So, if v is an admissible
control and there is no c such that v(t) = sign(cT eAt b) for all t ≤ 0, then
0
eAτ bv(τ )dτ ∈
/ ∂R
−∞
must be in the interior of R.

Since

sign kcT eAτ b = sign cT eAτ b , ∀ k > 0,
equation (2.3.1) shows that ∂R can be determined from the surface of a
unit ball in Rn . That is,
0
T Aτ
∂R = z = e b sign c e b dτ : |c| = 1 .
Aτ
(2.3.7)
−∞
In what follows, we will simplify (2.3.7) and describe ∂R in terms of a

set of trajectories of the time-reversed system (2.2.9).
Denote

E c := v(t) = sign cT eAt b , t ∈ R : c
= 0 , (2.3.8)
and for an admissible control v, denote

t
Φ(t, v) := − e−A(t−τ ) bv(τ )dτ. (2.3.9)
−∞
Since A is anti-stable, the integral in (2.3.9) exists for all t ∈ R. Hence

Φ(t, v) is well defined. We see that Φ(t, v) is the trajectory of the time-
reversed system (2.2.9) under the control v.

If v(t) = sign cT eAt b , then
t
Φ(t, v) = − e−A(t−τ ) bv(τ )dτ
−∞
0
=− eAτ b sign cT eAt eAτ b dτ
−∞
∈ ∂R, (2.3.10)
for all t ∈ R, i.e., Φ(t, v) lies entirely on ∂R. An admissible control v such
that Φ(t, v) lies entirely on ∂R is said to be an extremal control and such
Φ(t, v) an extremal trajectory. On the other hand, given an admissible
control v(t), if there exists no c such that v(t) = sign(cT eAt b) for all t ≤ 0,
then by Theorem 2.3.1, Φ(0, v) ∈ / ∂R must be in the interior of R. By
the time invariance property of the system, if there exists no c such that
v(t) = sign(cT eAt b) for all t ≤ t0 , Φ(t, v) must be in the interior of R for
all t ≥ t0 . Consequently, E c is the set of extremal controls. The following
lemma shows that ∂R is covered by the set of extremal trajectories.
Lemma 2.3.1.

∂R = Φ(t, v) : t ∈ R, v ∈ E c . (2.3.11)
Proof. For any fixed t ∈ R, it follows from (2.3.1) that

t

∂R = − e−A(t−τ ) b sign cT e−At eAτ b dτ : c
= 0
−∞

t
−A(t−τ )
T Aτ
= − e b sign c e b dτ : c
= 0 ,
−∞
i.e.,
∂R = Φ(t, v) : v ∈ E c ,
for any fixed t ∈ R. Hence ∂R can be viewed as the set of extremal
trajectories at any frozen time. Now let t vary, then each point on ∂R
moves along a trajectory but the whole set is invariant. So we can also
write
∂R = Φ(t, v) : v ∈ E c , t ∈ R ,
which is equivalent to (2.3.11). 2

Unlike (2.3.1), equation (2.3.11) shows that ∂R is covered by extremal
trajectories. It, however, introduces redundancy by repeating the same set
{Φ(t, v) : v ∈ E c } for all t ∈ R. This redundancy can be removed by a
careful examination of the set E c . Indeed, the set {Φ(t, v) : t ∈ R} can be
identical for a class of v ∈ E c .
Definition 2.3.1.
1) Two extremal controls v1 , v2 ∈ E c are said to be equivalent, denoted
by v1 ∼ v2 , if there exists an h ∈ R such that
v1 (t) = v2 (t − h), ∀ t ∈ R;
2) Two vectors c1 , c2 ∈ Rn are said to be equivalent, denoted by c1 ∼ c2 ,

if there exist a k > 0 and an h ∈ R such that
T
c1 = keA h
c2 .
Noting that a shift in time of the control corresponds to the same shift
of the state trajectory, we see that, if v1 ∼ v2 , then

Φ(t, v1 ) : t ∈ R = Φ(t, v2 ) : t ∈ R ;
and if c1 ∼ c2 , then
sign(cT1 eAt b) ∼ sign(cT2 eAt b).
Definition 2.3.2.
1) A set E m c ⊂ E c is called a minimal representative of E c if for any

v ∈ E c , there exists a unique v1 ∈ E m
c such that v ∼ v1 .
2) A set M ⊂ Rn is called a minimal representative of Rn if for any

c ∈ Rn , there exists a unique c1 ∈ M such that c ∼ c1 .
With this definition, there will be no pair of distinct elements in E m

c or in
M that are equivalent. It should be noted that the minimal representative
of E c or Rn is unique up to equivalence and E m c and M always exist. An
immediate consequence of these definitions and Lemma 2.3.1 is the following
theorem.
Theorem 2.3.2. If E m
c is a minimal representative of E c , then

∂R = Φ(t, v) : t ∈ R, v ∈ E m c .
If M is a minimal representative of Rn , then

∂R = Φ t, sign cT eAt b : t ∈ R, c ∈ M \ {0} .
It turns out that for some classes of systems, E m

c can be easily described.
For second order systems, E m
c contains only one or two elements, so ∂R can
be covered by no more than two trajectories. For third order systems, E m c
can be determined from some real intervals. We will see later that for
systems of different eigenvalue structures, the descriptions of E m c can be
quite different.
2.4. Systems with Only Real Eigenvalues 21
2.4. Systems with Only Real Eigenvalues

It follows from, for example, [76, p. 77], that if A has only real eigenvalues
and c
= 0, then cT eAt b has at most n − 1 zeros. This implies that an
extremal control can have at most n − 1 switches. We will show that the
converse is also true.
Theorem 2.4.1. For the system (2.2.9), assume that A has only real eigen-
values, then,
a) An extremal control has at most n − 1 switches;
b) Any bang-bang control with n − 1 or less switches is an extremal

control.
To prove this theorem, we first need to present a lemma. Let us use

P k to denote the set of real polynomials with degrees less than the integer
k. The number 0 is considered a polynomial with arbitrary degree or with
degree −1.
Lemma 2.4.1. Given N positive integers, k1 , k2 , · · · , kN , define a set of

functions

N
GN := g(t) = e fi (t) : fi ∈ P ki , ai ∈ R, g(t)
≡ 0 .
ai t
i=1
N
Then g(t) ∈ GN has at most i=1 ki − 1 zeros.
Proof. We prove this lemma by induction. It is easy to see that the

statement is true when N = 1. Now assume that it is true when N is
replaced by N − 1. Let g(t) ∈ GN . Suppose on the contrary that g has
N −aN t
N
i=1 ki or more zeros. Then g̃(t) = g(t)e also has i=1 ki or more
zeros. Therefore, the kN th derivative of g̃,
N −1 (kN )

(kN ) (ai −aN )t
[g̃(t)] = e fi (t) + fN (t)
i=1
N −1 (kN )

(ai −aN )t
= e fi (t)
i=1
∈ GN −1 ,
N −1
has at least i=1 ki zeros, which is a contradiction. 2
Proof of Theorem 2.4.1. The proof of a) was sketched in [76]. To show
b), assume that A has N distinct real eigenvalues λi , i = 1, 2, · · · , N , each
N
with a multiplicity of ki ( i=1 ki = n). It is well-known that

N
cT eAt b = eλi t fi (t),
i=1
for some fi ∈ P ki . If c
= 0, then cT eAt b
≡ 0 by the controllability of
(A, b). (Thus a) follows from Lemma 2.4.1). To complete the proof of
b), we first show that any bang-bang control v with n − 1 switches is an
extremal control.
Let t1 , t2 , · · · , tn−1 ∈ R be the distinct switching times of v. From the
following n − 1 linear equations
cT eAti b = 0, i = 1, 2, · · · , n − 1,
at least one nonzero vector c ∈ Rn can be solved. With any such a c, a)

implies that g(t) = cT eAt b
≡ 0 has no other zeros than the n − 1 zeros at
ti , i = 1, 2, · · · , n − 1.
Now the question is whether g(t) indeed changes the sign at each ti . If
it does, then v(t) = sign(cT eAt b) (or sign(−cT eAt b)) and v is an extremal
control.
We now show that g does change the sign at each ti . If g does not
change the sign at a certain ti , then g(t) must have a local extremum at
ti , so ġ(ti ) = 0. We argue that there is at most one ti such that ġ(ti ) = 0,
otherwise ġ would have at least n zeros, counting the at least n − 2 ones
lying within the intervals (ti , ti+1 )’s, which is impossible by Lemma 2.4.1,
since ġ has the same structure as g.
We further conclude that g, however, cannot have a local extremum at
any of these ti ’s.
Let
N
g(t) = eλi t fi (t).
i=1
Assume, without loss of generality, that fN (t)
≡ 0. Suppose on the contrary
that g has a local minimum (or maximum) at t1 , then g̃(t) = g(t)e−λN t also
has a local minimum (or maximum) at t1 , furthermore,
g̃(ti ) = 0, ˙ i )
= 0,
g̃(t i = 2, 3, · · · , n − 1.
Hence, there exists an > 0 (or < 0) such that

N −1
g̃(t) − = e(λi −λN )t fi (t) + fN (t) −
i=1
has n zeros, which contradicts Lemma 2.4.1. Therefore, g changes signs

at all ti . This shows that v(t) = sign(cT eAt b) (or sign(−cT eAt b) ) is an
extremal control.
Now consider the case that v has less than n − 1 switches, say n − 1 − j
switches, at ti , i = 1, 2, · · · , n − 1 − j. For simplicity and without loss
of generality, assume that A is in the Jordan canonical form (the state
transformation matrix can be absorbed in cT and b). Partition A and b as

A= , b= ,
0 A1 b1
where A1 is of size n − j. It is easy to see that A1 is also of the Jordan
canonical form and (A1 , b1 ) is controllable. Furthermore,

eAt = .
0 e A1 t
Accordingly, consider c of the form,

0
c= ,
c1
then
cT eAt b = cT1 eA1 t b1 .
By the forgoing proof for the full dimensional case, we see that there exists
c1 such that v(t) = sign(cT1 eA1 t b1 ) is a bang-bang control with switching
times exactly at ti , i = 1, 2, · · · , n − 1 − j.
Therefore, we conclude that any bang-bang control with less than n − 1
switches is also extremal. 2
By Theorem 2.4.1, the set of extremal controls can be described as
follows,
 
  

  1, −∞ ≤ t < t1 , 

E c = ± v : v(t) = (−1) ,
i
ti ≤ t < ti+1 , t1 < t2 ≤ · · · ≤ tn−1

  

 (−1)n−1 , tn−1 ≤ t < ∞, 

v(t) ≡ ±1 ,
where ti , i = 1, · · · , n − 1, are the switching times. If v(t) has a switch,

then the first switch occurs at t = t1 . Here we allow ti = ti+1 (i
= 1) and
tn−1 = ∞, so the above description of E c consists of all bang-bang controls
with n − 1 or less switches.
To obtain a minimal representative of E c , we can simply set t1 = 0, that
is,

 

  1, −∞ ≤ t < t1 ,
Emc = ± v : v(t) = (−1)i , ti ≤ t < ti+1 , 0 = t 1 < t2 ≤

 
 (−1)n−1 , tn−1 ≤ t < ∞,




· · · ≤ tn−1 v(t) ≡ ±1 .



For each v ∈ E mc , we have v(t) = 1 (or −1) for all t < 0. Hence, for
t ≤ 0, t
Φ(t, v) = − e−A(t−τ ) bdτ = −A−1 b (or A−1 b).
−∞
Afterwards, v(t) is a bang-bang control with n − 2 or less switches. Denote
ze+ = −A−1 b, ze− = A−1 b,
then from Theorem 2.3.2 we have,
Observation 2.4.1. ∂R = ∂C is covered by two bunches of trajectories.

The first bunch consists of trajectories of (2.2.9) when the initial state is ze+
and the input is a bang-bang control that starts at t = 0 with v = −1 and
has n − 2 or less switches. The second bunch consists of the trajectories
of (2.2.9) when the initial state is ze− and the input is a bang-bang control
that starts at t = 0 with v = +1 and has n − 2 or less switches.
Furthermore, ∂R can be simply described in terms of the open-loop

transition matrix. Note that for a fixed t ≥ 0,


 ti+1
n−1
−At + −A(t−τ )
Φ(t, v) : v ∈ E c = ± e
m
ze − e i
b(−1) dτ :

 i=1 ti



0 = t1 < t2 ≤ · · · ≤ tn−1 ≤ tn = t ∪ ± ze+



 n−1

i −A(t−ti )
= ± 2(−1) e + (−1) I A−1 b :
n

 i=1



0 = t1 < t2 ≤ · · · ≤ tn−1 ≤ t ∪ ± ze+ .


Hence,

∂R = Φ(t, v) : t ∈ R, v ∈ E m
c

 n−1

i −A(t−ti )
= ± 2(−1) e + (−1) I A−1 b :
n

 i=1



0 = t1 ≤ t2 · · · ≤ tn−1 ≤t≤∞ . (2.4.1)


Here, we allow t1 = t2 to include ±ze+ . For second order systems,

 

 t 

−At − −A(t−τ )
∂R = ± e ze − e bdτ : t ∈ [0, ∞]

 0 


= ± −2e−At + I A−1 b : t ∈ [0, ∞] . (2.4.2)
If n = 3, then one half of ∂R = ∂C is formed by the trajectories of

(2.2.9) starting from ze+ with the control initially being v = −1 and then
switching at any time to v = +1. So the trajectories go toward ze− at first
then turn back toward ze+ . The other half is just symmetric to the first
half. That is,


 t2 t
∂R = ± e−At ze+ + e−A(t−τ ) bdτ − e−A(t−τ ) bdτ :

 0 t2



0 ≤ t2 ≤ t ≤ ∞ . (2.4.3)


0.8
v=+1
0.6
0.4
0.2
−
0 z
e
+
ze
−0.2
−0.4
−0.6
v=−1
−0.8
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Figure 2.4.1: ∂R of a second order system.
Plotted in Figs. 2.4.1 and 2.4.2 are respectively the ∂R of a second order
system with
0 −0.5 0
A= , B= ,
1 1.5 −1
and that of a third order system with
   
0.2 1 0 1
A =  0 0.2 0 , B =  1 .
0 0 0.4 1
Since the trajectories of the original system and those of the time-reversed
system are the same but traverse in opposite directions, we can also say
that ∂R = ∂C is covered by a set of trajectories of the original system.
While all the trajectories of the time-reversed system start at ze+ or ze−
and are very easy to generate by simulation, it is impossible to get the
same trajectories from the original system. For example, when n = 2, one
half of ∂R is formed by the trajectory of the time-reversed system that
starts at ze− under a constant control v = +1. The trajectory goes from ze−
toward ze+ asymptotically but never reaches ze+ at a finite time. It seems
that if we apply u = +1 at ze+ to the original system, the trajectory will
2.5. Systems with Complex Eigenvalues 27
2 −
z
e
1
−1
−2
−3 z+
e
10
5 30
20
0 10
0
−5 −10
−20
−10 −30
Figure 2.4.2: ∂R of a third order system.
go from ze+ to ze− along the same trajectory of the time-reversed system.
However, this is not the case. The trajectory of the original system will
stay at ze+ under the constant control u = +1. The boundary ∂R can only
be partially generated from the original system if we know one point on it
other than ±ze+ . But this point is not easy to get from the original system
by simulation.
2.5. Systems with Complex Eigenvalues

For a system with complex eigenvalues, the minimal representative set E m
c
is harder to determine. In what follows, we consider two important cases.
Case 1. A ∈ R2×2 has a pair of complex eigenvalues α ± jβ, α, β > 0
Let V be the nonsingular matrix such that

α −β
A=V V −1 ,
β α
and let
c1 b1
= V c,T
= V −1 b,
c2 b2
then

At
cos(βt) − sin(βt) b1
T
c e b= c1 c2 eαt
sin(βt) cos(βt) b2

b1 b2 c1
= cos(βt) sin(βt) eαt .
−b2 b1 c2

b1 b2
Since is nonsingular, it follows that
−b2 b1

b1 b2 c1 sin(θ)
: c
= 0 = r : r
= 0, θ ∈ [0, 2π) .
−b2 b1 c2 cos(θ)
Hence

sign cT eAt b : c
= 0 = sign(sin(βt + θ)) : θ ∈ [0, 2π) ,
and the set of extremal controls is

E c = v(t) = sign(sin(βt + θ)), t ∈ R : θ ∈ [0, 2π) .
It is easy to see that all the elements of E c are equivalent. By setting

θ = 0, we obtain

Emc = v(t) = sign(sin(βt)), t ∈ R ,
which contains only one element. Denote

π
Tp = ,
β
then
e−ATp = −e−αTp I.
Let
−1 1 + e−αTp −
zs− = I + e−ATp I − e−ATp A−1 b = z . (2.5.1)
1 − e−αTp e
It can be verified that the extremal trajectory corresponding to
v(t) = sign(sin(βt))
is periodic with period 2Tp and,

t
∂R = ± e−At zs− − e−A(t−τ ) bdτ : t ∈ [0, Tp )
0

−At −
= ± e zs − I − e−At A−1 b : t ∈ [0, Tp ) . (2.5.2)
3
v=−1
z−e
1 z−
s
0
+
zs
−1 +
ze
−2
−3
v=1
−4
−5
−6 −4 −2 0 2 4 6
Figure 2.5.1: ∂R of a second order system with complex poles.
Plotted in Fig. 2.5.1 is the ∂R of a second order system with

0.6 −0.8 2
A= , B= .
0.8 −0.6 4
The trajectory goes from zs− to zs+ under the control v = +1 and then from
zs+ to zs− under the control v = −1.
Unlike second order systems with real eigenvalues, where ze+ and ze− are
on the boundary of R, for second order systems with complex eigenvalues,
zs+ and zs− are on the boundary, while ze+ and ze− are in the interior of R
(see Fig. 2.5.1, where ze+ and ze− are marked with “2” and zs+ and zs− are
marked with “∗”).
Case 2. A ∈ R3×3 has eigenvalues α ± jβ and α1 , with α, β, α1 > 0
We will consider the case of α = α1 and that of α
= α1 separately.
a) α = α1 . Similar to Case 1, we can obtain

E c = v(t) = sign(k + sin(βt + θ)), t ∈ R : k ∈ R, θ ∈ [0, 2π) .
Since sign(k + sin(βt + θ)) is the same for all k ≥ 1 (or k ≤ −1), we have

Emc = v(t) = sign(k + sin(βt)), t ∈ R : k ∈ [−1, 1] .
0.5
−0.5
−1
−1.5
1
0.5 1.5
1
0 0.5
0
−0.5 −0.5
−1
−1 −1.5
Figure 2.5.2: Extremal trajectories on ∂R, α = α1 .
Each v ∈ E mc is periodic with period 2Tp , but the length of v(t) being 1 and
that of v(t) being −1 vary with k. Φ(t, v) can be easily determined from
simulation or direct computation. Plotted in Fig. 2.5.2 are some extremal
trajectories on ∂R of the time-reversed system (2.2.9) with
   
0.8 0 0 1
A =  0 0.8 −2  , B =  1  .
0 2 0.8 1
b) α
= α1 . In this case,

E c = v(t) = sign k1 e(α1 −α)t + k2 sin(βt + θ) , t ∈ R :

(k1 , k2 )
= (0, 0), θ ∈ [0, 2π) .
This set E c can be decomposed as
E c = E c1 ∪ E c2 ∪ E c3 ,
where
E c1 = v(t) ≡ ±1 , (k2 = 0),

E c 2 = v(t) = ±sign(sin(βt + θ)) : θ ∈ [0, 2π) , (k1 = 0),
and,

E c3 = v(t) = ±sign ke(α1 −α)t + sin(βt + θ) : k > 0, θ ∈ [0, 2π) .
We will show that a minimal representative of E c3 is

(α1 −α)t
Em
c3 = v(t) = ±sign e + sin(βt + θ) : θ ∈ [0, 2π) . (2.5.3)
Let
v(t) = sign ke(α1 −α)t + sin(βt + θ) ∈ E c 3 .
Since k > 0, there is a number h ∈ R such that e(α1 −α)h = k. So

v(t) = sign e(α1 −α)(t+h) + sin(β(t + h) − βh + θ) = v1 (t + h)
for some v1 (t) ∈ E m

c3 , i.e., v ∼ v1 . On the other hand, given v1 , v2 ∈ E c3 ,
m
suppose that

v1 (t) = sign e(α1 −α)t + sin(βt + θ1 ) ,

v2 (t) = sign e(α1 −α)t + sin(βt + θ2 ) ,
and v1 ∼ v2 , i.e., v1 (t) = v2 (t − h), for some h, then

sign e(α1 −α)t + sin(βt + θ1 ) = sign e(α1 −α)(t−h) + sin(β(t − h) + θ2 ) .
When α1 < α (or α1 > α), both e(α1 −α)t and e(α1 −α)(t−h) go to zero as t
goes to ∞ (or −∞). For v1 (t) and v2 (t − h) to change signs at the same
time, we must have βt + θ1 = β(t − h) + θ2 + lπ, for some integer l. Since
at any switching time of v1 (t) and v2 (t),
sin(βt + θ1 ) < 0, sin(β(t − h) + θ2 ) < 0,
we conclude that
sin(βt + θ1 ) = sin(β(t − h) + θ2 )
and hence
e(α1 −α)t = e(α1 −α)(t−h) .
0.5
−0.5
−1
−1.5
−2
1
0.5 1.5
1
0 0.5
0
−0.5 −0.5
−1
−1 −1.5
Figure 2.5.3: Extremal trajectories on ∂R, α1 < α.
Therefore, we must have h = 0 and θ1 = θ2 , i.e., v1 = v2 . These show that

Emc3 is a minimal representative of E c 3 .
The minimal representative of E c 2 is the same as E m c in Case 1. It
follows that

Em
c = v(t) ≡ ±1 ∪ v(t) = sign(sin(βt)) ∪ E c3 .
m
When α1 < α, for each v ∈ E m c3 , v(t) = 1 (or −1) for all t ≤ 0, so the
corresponding extremal trajectory stays at ze+ = −A−1 b or ze− before t = 0.
And after some time, it goes toward a periodic trajectory since as t goes to
infinity, v(t) becomes periodic. When α1 > α, for each v ∈ E m c3 , v(t) = 1(or
−1) for all t ≥ 0, and the corresponding extremal trajectory starts from
near periodic and goes toward ze+ or ze− .
Plotted in Fig. 2.5.3 are some extremal trajectories on ∂R of the time-
reversed system (2.2.9) with
  

0.5 0 0 1
A= 0 0.8 −2  , B =  1 .
0 2 0.8 1
2.6. Some Remarks on the Description of C(T ) 33
For v(t) ≡ ±1 or v(t) = sign(sin(βt)), a closed form expression of Φ(t, v) can

be obtained. However, for v ∈ E m c3 , there exists no closed form expression
of Φ(t, v).
For higher order systems, the relative locations of the eigenvalues are
more diversified and the analysis will be technically much more involved. It
can, however, be expected that in general case, the number of parameters
used to describe E m
c is n − 2.
2.6. Some Remarks on the Description of C(T )

The problem of describing the null controllable region at a finite time T ,
C(T ), was studied in the earlier literature (see, e.g., [18,63,76]). A descrip-
tion of ∂R(T ) similar to Theorem 2.3.1 can be summarized from [63] and
[76], as was done in [47]. Also, a boundary formula for the case where A
has only real eigenvalues was derived in [47]. However, the result that ∂R
is composed of a set of extremal trajectories does not apply to ∂R(T ). In
this section, we will give a brief summary of the earlier results on R(T ) and
explain why there is no corresponding boundary description for R(T ).
Theorem 2.6.1 ([47,76]).

T
−A(T −τ )
T Aτ
∂R(T ) = z = e b sign c e b dτ : c
= 0 . (2.6.1)
0
R(T ) is strictly convex. Moreover, for each z ∗ ∈ ∂R(T ), there exists a

unique admissible control v ∗ such that
T
∗
z = e−A(T −τ ) bv ∗ (τ )dτ. (2.6.2)
0
The equation (2.6.1) was presented in [47] and the other properties
in Theorem 2.6.1 can be obtained similarly as those in Theorem 2.3.1.
From Theorem 2.6.1, we see that each point on ∂R(T ) can be obtained by
applying a control

v(t) = −sign cT eAt b
to the time-reversed system under zero initial condition. In the case that
A has only real eigenvalues, it was shown in [47] that the set of controls

v(t) = sign cT eAt b , t ∈ [0, T ] : c
= 0
is also the set of bang-bang controls with n − 1 or less switches. That is,

T
∂R(T ) = z= e−A(T −τ ) b v(τ )dτ : v ∈ E c (T ) , (2.6.3)
0
where E c (T ) is the set of bang-bang controls on [0, T ] that have n − 1 or

less switches. Based on (2.6.3), a formula similar to (2.4.1) was obtained
to describe ∂R(T ) in [47].
However, we don’t have a trajectory description similar to Theorem
2.3.2 for ∂R(T ). This is because the major step towards Theorem 2.3.2,
equation (2.3.10), does not go through if −∞ is replaced by a finite number
−T .
2.7. Asymptotically Null Controllable Region
Much as the notion of stabilizability relaxes the stronger notion of con-

trollability, asymptotically null controllability is a natural extension of the
notion of null controllability. This relaxation is introduced to deal with the
situation where the linear system contains some uncontrollable, but stable,
poles.
Definition 2.7.1. Consider the system (2.2.1). A state x0 is said to be

asymptotically null controllable if there exists an admissible control u such
that the state trajectory x(t) of the system satisfies x(0) = x0 and
lim x(t) = 0.
t→∞
The set of all states that are asymptotically null controllable, denoted by
C a , is called the asymptotically null controllable region.
The characterization of the asymptotically null controllable region is

trivial and can be carried out in terms of the null controllable region of the
anti-stable subsystem.
Theorem 2.7.1. Consider the system (2.2.1). Assume that (A, B) is sta-
bilizable and is given in the following form,

A1 0 B1
A= , B= ,
0 A2 B2
2.8. Conclusions 35
where A1 ∈ Rn1 ×n1 is anti-stable and A2 ∈ Rn1 ×n2 is semi-stable. Then,

the asymptotically null controllable region is given by
C a = C 1 × Rn2 ,
where C 1 is the null controllable region of the anti-stable subsystem
ẋ1 = A1 x1 + B1 u.
Proof. Without loss of generality, let us further assume that

A20 0 B20
A2 = , B2 = ,
0 A2− B2−
where A20 ∈ Rn20 ×n20 has all its eigenvalues on the imaginary axis and
A2− ∈ Rn2− ×n2− is Hurwitz. Let the state be partitioned accordingly as
x = [xT1 xT20 xT2− ]T , with x20 ∈ Rn20 , x2− ∈ Rn2− . Then,
% &
A1 0 B1
,
0 A20 B20
is controllable and the null controllable region corresponding to the state

[xT1 xT20 ]T is C 1 × Rn20 by Proposition 2.2.1. If x0 = [xT1 xT20 xT2− ]T ∈
C 1 × Rn2 , then there exist a finite time T and an admissible control such
that x1 (T ) = 0, x20 (T ) = 0. For t ≥ T , we can simply set u = 0 and the
state x2− will approach the origin asymptotically. 2
It is also clear that, if the system is controllable, then its asymptotically
null controllable region and null controllable region are identical.
2.8. Conclusions
In this chapter, we have obtained explicit descriptions of the boundary of
the null controllable region of a continuous-time linear system with bounded
controls. In the next chapter, the discrete-time counterparts of the results
of this chapter will be presented. Throughout the book we will be concerned
with feedback laws that are valid on the entire null controllable region or a
large portion of it.
Chapter 3
Null Controllability –
Discrete-Time
Systems
3.1. Introduction
This chapter studies null controllability of discrete-time linear systems with

bounded controls. As in the continuous-time case, when the control input
is limited by the actuator saturation, a discrete-time linear controllable
system might not be globally null controllable. We are thus led to the char-
acterization of the null controllable region C, the set of all states that can
be steered to the origin by the bounded controls delivered by the actuators.
Let C(K) denote the set of states that can be steered to the origin with
bounded controls in K steps. Then, the null controllable region C is given
by the union of C(K) for all K ∈ [1, ∞).
As seen in the literature, there has been continual interest in the study of
the null controllable region for a discrete-time linear system with bounded
controls (see, e.g., [10,18–21,26,28,52,54,61,62,64,80,89,90,103,106]). In this
study, C is typically approximated by C(K) with K sufficiently large. For
a fixed K, C(K) is characterized in terms of its boundary hyperplanes or
vertices which are usually computed via linear programming. As K in-
creases, the computational burden is more intensive and it is more difficult
to implement the control based on on-line computation. An exception is
[21], where an explicit formula for computing the vertices of C was provided
37
38 Chapter 3. Null Controllability – Discrete-Time Systems
for second order systems with complex eigenvalues. An interesting inter-

pretation of this set of vertices is that they form a special trajectory of the
system under a periodic bang-bang control.
It is notable that there is another trend in constrained control where
state constraints are also considered (see, e.g., [20,26,79]). The main idea
of the design proposed there is to determine a contractive positive invariant
set, with corresponding controls, to keep the states in a smaller and smaller
subsets. The tool employed is also linear programming. The complexity
of the controller depends on the conservativeness of the invariant set. As
the invariant set approaches the largest one, which, in the absence of state
constraints, is C, the on-line computational burden also increases and the
controller will be more difficult to implement.
In this chapter, we will develop an explicit description of the null con-
trollable region. This description will be instrumental in arriving at easily
implementable stabilizing feedback laws in Chapter 5. Our presentation in
this chapter is based on both the earlier understanding of the null control-
lable region and our recent work [44].
Section 3.2 contains some preliminaries from which a complete descrip-
tion of the null controllable region can be made. Section 3.3 presents a
general description of the null controllable region. Although explicit and
complete, this general description can be drastically simplified by utiliz-
ing the eigenstructure of the given system. Sections 3.4 and 3.5 show how
such simplification is made. In particular, Section 3.4 deals with the case
that all eigenvalues are real, and Section 3.5 deals with the case of complex
eigenvalues. Section 3.6 uses an example to illustrate how the boundary
of the null controllable region can be obtained. Section 3.7 addresses the
situation where the linear system contains some uncontrollable, but stable,
poles. Finally, a brief concluding remark is made in Section 3.8.
3.2. Preliminaries and Definitions
Consider a discrete-time system
x(k + 1) = Ax(k) + Bu(k), (3.2.1)
where x(k) ∈ Rn is the state and u(k) ∈ Rm is the control. A control

signal u is said to be admissible if |u(k)|∞ ≤ 1 for all integer k ≥ 0. We
are interested in the control of system (3.2.1) by using admissible controls.
3.2. Preliminaries and Definitions 39
Our first concern is the set of states that can be steered to the origin by
admissible controls.
Definition 3.2.1.
1) A state x0 is said to be null controllable in K steps (K > 0) if there

exists an admissible control u such that the time response x of the
system satisfies x(0) = x0 and x(K) = 0.
2) The set of all states that are null controllable in K steps is called the
null controllable region of the system at step K and is denoted by
C(K).
Definition 3.2.2.
1) A state x0 is said to be null controllable if x0 ∈ C(K) for some K < ∞.
2) The set of all null controllable states is called the null controllable
region of the system and is denoted by C.
Similar to the continuous-time case, we have the following separation

result.
Proposition 3.2.1. Assume that (A, B) is controllable.
a) If A is semi-stable, then C = Rn .
b) If A is anti-stable, then C is a bounded convex open set containing

the origin.
c) If
A1 0
A= ,
0 A2
with A1 ∈ Rn1 ×n1 anti-stable and A2 ∈ Rn2 ×n2 semi-stable, and B
is partitioned accordingly as

B1
B= ,
B2
then
C = C 1 × Rn2 ,
where C 1 is the null controllable region of the anti-stable sub-system
x1 (k + 1) = A1 x1 (k) + B1 u(k).
Because of this proposition, we can concentrate on the study of null

controllable regions for anti-stable systems. For anti-stable systems, C can
be approximated by C(K) for sufficiently large K.
If

B = b1 b2 · · · bm
and the null controllable region of the system
x(k + 1) = Ax(k) + bi ui (k), i = 1, 2, · · · , m,
is C i , then

m
C= Ci = x1 + x2 + · · · + xm : xi ∈ C i , i = 1, 2, · · · , m .
i=1
Hence we can begin our study of null controllable regions with systems
having only one input.
In summary, we will assume in the study of null controllable regions
that (A, B) is controllable, A is anti-stable, and m = 1.
In many situations, it may be more convenient to study the null con-
trollability of a system through the reachability of its time-reversed system.
The time-reversed system of (3.2.1) is
z(k + 1) = A−1 z(k) − A−1 Bv(k). (3.2.2)
Note that we have assumed that A is anti-stable, so A is invertible. It is

easy to see that x(k) satisfies (3.2.1) with x(0) = x0 , x(k1 ) = x1 , and a
given u if and only if z(k) := x(k1 − k) satisfies (3.2.2) with z(0) = x1 ,
z(k1 ) = x0 , and v(k) = u(k1 − k − 1). The two systems have the same set
of points as trajectories, but traverse in opposite directions.
Definition 3.2.3. Consider the system (3.2.2).
1) A state zf is said to be reachable in K steps if there exists an ad-

missible control v such that the time response z of the system (3.2.2)
satisfies z(0) = 0 and z(K) = zf .
2) The set of all states that are reachable in K steps is called the reach-
able region of the system (3.2.2) at step K and is denoted by R(K).
Definition 3.2.4. Consider the system (3.2.2).

1) A state zf is said to be reachable if zf ∈ R(K) for some K < ∞.
2) The set of all reachable states is called the reachable region of the
system (3.2.2) and is denoted by R.
It is easy to verify that C(K) and C of (3.2.1) are the same as R(K) and
R of (3.2.2). To avoid confusion, we will reserve the notation x, u, C(K),
and C for the original system (3.2.1), and use z, v, R(K), and R for the
time-reversed system (3.2.2).
To proceed, we need more notation. With K1 , K2 integers, for con-
venience, we will use [K1 , K2 ] to denote the set of integers {K1 , K1 +
1, · · · , K2 }. We will also use [a1 , a2 ] to denote the usual closed interval
of real numbers. The situation will be clear from the context.
3.3. General Description of Null Controllable Region
We have assumed in the last section that A is anti-stable, (A, B) is control-

lable, and m = 1. Since B is now a column vector, we rename it as b for
convenience. From Definitions 3.2.2 and 3.2.4, C(K), R(K), C and R can
be written as,
K−1

−(K−)
C(K) = R(K) = − A bv() : |v()| ≤ 1, ∀ ∈ [0, K − 1] ,
=0
and
∞

−(K−)
C=R= − A bv() : |v()| ≤ 1, ∀ ≥ 0 .
=0
It is easy to see that C(K), R(K), C and R are all convex and that C(K)
and R(K) are polytopes. An extremal point of a polytope is usually called
a vertex. In some special cases, R (or C) could also be a polytope of finite
many extremal points. But in general, R has infinitely many extremal
points. In any case, R is the convex hull of Ext(R), the set of all the
extremal points of R. In view of this, it suffices to characterize Ext(R).
Also, it can be shown with standard analysis that C(K), R(K), C and
R depend on A and b continuously in the Housdorff metric, even if (A, b)
is not controllable in the usual linear sense. For technical reasons, we first
consider the reachable region R(K).
Definition 3.3.1. An admissible control v is said to be an extremal control

on [0, K] if the response z(k) of the system (3.2.2), with z(0) = 0, is in
Ext(R(k)) for all k ∈ [0, K].
Lemma 3.3.1. If zf ∈ Ext(R(K)), and v is an admissible control that

steers the state from the origin to zf at step K, then v is an extremal
control on [0, K].
This lemma is obvious. Note that if z(k1 ) ∈

/ Ext(R(k1 )) for some k1 > 0,
then z(k) will not be in Ext(R(k)) for any k > k1 under any admissible
control. Denote the set of extremal controls on [0, K] as E c (K). It then
follows that
K−1

Ext(R(K)) = − A−(K−) bv() : v ∈ E c (K) . (3.3.1)
=0
Lemma 3.3.2 ([82]). An admissible control v ∗ is an extremal control on

[0, K] for the system (3.2.2) if and only if there is a vector c ∈ Rn such that
cT Ak b
= 0, ∀ k ∈ [0, K − 1],
and
v ∗ (k) = sign(cT Ak b), ∀ k ∈ [0, K − 1].
Proof. Let v ∗ be an admissible control and define

K−1
z ∗ := − A−(K−) bv ∗ ().
=0
From Definition 3.3.1 and Lemma 3.3.1, v ∗ is an extremal control iff z ∗ ∈

Ext(R(K)). This is the case if and only if there exists a nonzero vector
h ∈ Rn such that

hT z ∗ = min hT z : z ∈ R(K)
K−1

= min − hT A−(K−) bv() : |v()| ≤ 1, ∈ [0, K − 1]
=0
(3.3.2)
and
hT z ∗ < hT z, ∀ z ∈ R(K) \ {z ∗ }. (3.3.3)
Observing that if hT A−(K−i) b = 0 for some i ∈ [0, K − 1], we have
hT z ∗ = hT z, ∀ z = z ∗ + A−(K−i) bv(i), |v(i)| ≤ 1.
This means that (3.3.2) and (3.3.3) hold if and only if
hT A−(K−) b
= 0, ∀ ∈ [0, K − 1]
and
v ∗ () = sign hT A−(K−) b , ∈ [0, K − 1].
The result follows by replacing hT A−K with cT and with k. 2

This lemma says that an extremal control is a bang-bang control, i.e.,
a control only takes value 1 or −1. Because of this lemma, we can write
E c (K) as

E c (K) = v(k) = sign cT Ak b : cT Ak b
= 0, ∀ k ∈ [0, K − 1] . (3.3.4)
Consequently, it follows from (3.3.1) that

K−1

−(K−)
Ext(R(K)) = − A b sign(c A b) : c A b
= 0, ∀ ∈ [0, K − 1] .
T T
=0
(3.3.5)
Writing cT A as cT AK A−(K−) and replacing cT AK with cT and K −
with , we have
K

− − −
Ext(R(K)) = − A b sign(c A b) : c A b
= 0, ∀ ∈ [1, K] .
T T
=1
Letting K go to infinity, we arrive at the following result.

Theorem 3.3.1.
∞

− − −
Ext(C) = Ext(R) = − A T
b sign(c A T
b) : c A b
= 0, ∀ ≥ 1 .
=1
Theorem 3.3.1 can also be proven with a similar procedure as the proof
of Theorem 2.3.1. It should be noted that for the continuous-time case,
every point in ∂R is an extremal point. This is implied by the fact that R
is strictly convex for a continuous time system.
We note that in the above theorem, the infinite summation always exists
since A is anti-stable.
Since

sign cT A− b = sign γcT A− b
for any positive number γ, this formula shows that Ext(R) can be deter-
mined from the surface of a unit ball. It should be noted that each extreme
corresponds to a region in the surface of the unit ball rather than just
one point. This formula provides a straightforward method for computing
the extremal points of the null controllable region and no optimization is
involved. In the following, we will give a more attractive formula for com-
puting the extremal points of the null controllable region by exploring the
eigenstructure of the matrix A.
3.4. Systems with Only Real Eigenvalues

In comparison with the continuous-time systems, a little more technical
consideration is needed here. This difference can be illustrated with a simple
example. If A = −2, then cT Ak b changes the sign at each k. Hence, if A has
some negative real eigenvalues, an extremal control can have infinitely many
switches. This complexity can be avoided through a technical manipulation.
Suppose that A has only real eigenvalues including some negative ones.
Consider

y(k + 1) = A2 y(k) + Ab b w(k), (3.4.1)
where y(k) = x(2k) and

u(2k)
w(k) = .
u(2k + 1)
Then the null controllable region of (3.2.1) is the same as that of (3.4.1),
which is the sum of the null controllable regions of the following two sub-
systems:
y(k + 1) = A2 y(k) + Ab w1 (k),
and
y(k + 1) = A2 y(k) + b w2 (k),
both of which have positive real eigenvalues. Therefore, without loss of
generality, in this section we further assume that A has only positive real
eigenvalues. Under this assumption, it is known that any extremal control
can have at most n − 1 switches [82]. Here we will show that the converse
is also true. That is, any bang-bang control with n − 1 or less switches is
an extremal control.
Lemma 3.4.1. For the system (3.2.2), suppose that A has only positive
real eigenvalues. Then,
a) An extremal control has at most n − 1 switches;
b) Any bang-bang control with n − 1 or less switches is an extremal

control.
Proof. Since A has only positive real eigenvalues, systems (3.2.1) and
(3.2.2) can be considered as the discretized systems resulting from
ẋ(t) = Ac x(t) + bc u(t) (3.4.2)
and
ż(t) = −Ac z(t) − bc v(t) (3.4.3)
with sampling period h, where Ac has only positive real eigenvalues. Thus,
Ac h
A = e Ac h , b = A−1
c e − I bc
and
Ac h
cT Ak b = cT A−1
c e − I eAc hk bc .
By Lemma 2.4.1, the continuous function in t,

cT A−1
c I − e Ac h e Ac t b c ,
changes sign at most n − 1 times, it follows that sign(cT Ak b) has at most

n − 1 switches.
To prove b), let Rc (T ) be the reachable region of the continuous-time
system (3.4.3) at time T . Suppose that v ∗ is a discrete-time bang-bang con-
trol with n − 1 or less switches, and that the state of the system (3.2.2) at
step K under the control v ∗ is z ∗ ; equivalently, a corresponding continuous-
time bang-bang control will drive the state of the system (3.4.3) from the
origin to z ∗ at time Kh. It follows from (2.6.3) that z ∗ belongs to ∂Rc (Kh)
of the continuous-time system (3.4.3). Recall from Theorem 2.6.1 that
Rc (Kh) is strictly convex, i.e., every boundary point of Rc (Kh) is an ex-
tremal point, it follows that z ∗ is an extremal point of Rc (Kh). Since
z ∗ ∈ R(K) ⊂ Rc (Kh), we must have z ∗ ∈ Ext(R(K)). Therefore, v ∗ is an
extremal control. 2
It follows from the above lemma that the set of extremal controls on
[0, K] can be described as follows,

 

  1, 0 ≤ k < k1 ,
E c (K) = ± v : v(k) = (−1)i , ki ≤ k < ki+1 ,

 
 (−1)n−1 , kn−1 ≤ k ≤ K − 1,




0 ≤ k1 ≤ · · · ≤ kn−1 ≤ K − 1 .



Notice that we allow ki = ki+1 in the above expression to include all the
bang-bang controls with n − 1 or less switches.
For a square matrix X, it can be easily verified that
k −1

2
X k
(I − X) = X k1 − X k2 .
k=k1
Hence, if I − X is nonsingular, then

2 −1
k
X k = (X k1 − X k2 )(I − X)−1 .
k=k1
By applying this equality we have that, if v ∈ E c (K), then



K−1 1 −1
k 2 −1
k 3 −1
k
A−(K−) bv() = A−K  A − A + A − · · ·
=0 =0 =k1 =k2


K−1
+ (−1)n−1 A  b
=kn−1

= A−K I − Ak1 − Ak1 + Ak2 + Ak2 − · · ·


+ (−1)n−1 Akn−1 − (−1)n−1 AK  (I − A)−1 b

−K
=A I − 2Ak1 + 2Ak2 − · · · + (−1)n−1 2Akn−1

+ (−1) A (I − A)−1 b.
n K
So from (3.3.1), we have

K−1

−(K−)
Ext(R(K)) = − A bv() : v ∈ E c (K)
=0

n−1
−K −(K−ki )
= ± A +2 i
(−1) A + (−1) I (I − A)−1 b :
n
i=1


0 ≤ k1 ≤ · · · kn−1 ≤ K − 1


n−1
= ± A−K + 2 (−1)i A−i + (−1)n I (I − A)−1 b :
i=1


K ≥ 1 ≥ · · · ≥ n−1 ≥ 1 .

By letting K go to infinity, we arrive at the following theorem.
Theorem 3.4.1. If A has only real positive eigenvalues, then
Ext(C) = Ext(R)


n−1
i −i
= ± 2 (−1) A + (−1) I (I − A)−1 b :
n

i=1


∞ ≥ 1 ≥ · · · ≥ n−1 ≥ 1 .

In particular, for second order systems, we have,

Ext(R(K)) = ± A−K − 2A− + I (I − A)−1 b : 1 ≤ ≤ K .
Hence, there are exactly 2K extremal points (vertices), versus the upper
bound of 2K extremal points which emerges from a superficial analysis of
R(K). Furthermore, notice that

Ext(R) = ± 2A− − I (I − A)−1 b : 1 ≤ ≤ ∞ . (3.4.4)
Similarly, for third order systems,

Ext(R(K)) = ± A−K − 2A−1 + 2A−2 − I (I − A)−1 b :

1 ≤ 2 ≤ 1 ≤ K ,
which has K(K −1) extremal points, versus the upper bound of 2K extremal
points which emerges from a superficial analysis of R(K). As expected,

Ext(R) = ± 2A−1 − 2A−2 + I (I − A)−1 b : 1 ≤ 2 ≤ 1 ≤ ∞ .
A more interesting interpretation of Ext(R) can be obtained after some

manipulation. Let
x+
e := (I − A)
−1
b
be the equilibrium point of the system (3.2.2) under the constant control
v(k) = 1. Then for a second order system, it can be verified that

k−1
Ext(R) = ± A−k x+ e + A−(k−−1) (−A−1 b)(−1) : 1 ≤ k ≤ ∞ ,
=0
(3.4.5)
which is exactly the set of points formed by the trajectories of (3.2.2) start-
ing from x+ +
e or −xe under the constant control of v = −1 or +1.
For third order systems,

1 −1
k
Ext(R) = ± A−k2 x+ e + A−(k2 −−1) (−A−1 b)(−1)
=0

2 −1
k
+ A−(k2 −−1) (−A−1 b)(+1) : 1 ≤ k1 ≤ k2 ≤ ∞ .
=k1
We see that one half of Ext(R) is formed by the trajectories of (3.2.2)

starting from x+ e , first under the control of v = −1, and then switching to
v = +1 at any step k1 . The other half is symmetric to the first half.
Similarly, for higher-order systems with only positive real eigenvalues,
Ext(R) = Ext(C) can be interpreted as the set of points formed by the
trajectories of (3.2.2) starting from x+ +
e or −xe under any bang-bang control
with n − 2 or less switches.
3.5. Systems with Complex Eigenvalues

We now consider the situation where A ∈ R2×2 has a pair of complex
eigenvalues, r(cos β ± j sin β), with r > 1, 0 < β < π.
Theorem 3.5.1 ([21]). Suppose

p
β= π,
q
where p and q are coprime positive integers and p < q. Then,
Ext(R(2q))
2q−1 % &
π i
= − A−(2q−l) b sign sin β+ + π : i ∈ [0, 2q−1] , (3.5.1)
2q q
=0
and
r2q
Ext(R) = Ext(R(2q)). (3.5.2)
r2q − 1
In view of this theorem, we can compute the extremal points of R(2q)
using (3.5.1), and then scale them by
r2q
r2q − 1
to obtain the extremal points of R.
The set of extremal points Ext(R) also coincides with the steady state
trajectory of the time-reversed system (3.2.2) to a particular periodic bang-
bang control. Let the control be
% &
∗ π
v (k) = sign sin βk +
2q
and z ∗ (k) be the zero initial state response,

k−1 % &
∗ −(k−) π
z (k) = − A b sign sin β + .
2q
=0
Denote Γ(K) as

Γ(K) := ± z ∗ (K + k) : k ∈ [0, 2q − 1] , (3.5.3)
then the limit limK→∞ Γ(K) exists, and this limit is the union of the steady
state trajectories of (3.2.2) under v ∗ (k) and −v ∗ (k). Let
Γ = lim Γ(K).
K→∞
Proposition 3.5.1 ([21]). Suppose that

p
β= π, p < q,
q
and (p, q) are coprime positive integers. Then
Ext(C) = Ext(R) = Γ.
If
1
β= π,
N
where N is a positive integer, then v ∗ (k) has a period of 2N and
AN = −rN I.
If we denote
rN + 1 rN + 1 +
x+
s = (I − A)−1
b = x ,
rN − 1 rN − 1 e
and
x− +
s = −xs ,
then we can obtain an explicit expression for the steady state trajectory Γ,

i−1
−i − −(i−−1) −1
Γ = ± A xs + A (−A b)(1) : 1 ≤ i ≤ N . (3.5.4)
=0
This is composed of trajectories of (3.2.2) from ±x− +

s to ±xs under v(k) =
+ −
±1, or the trajectories of (3.2.1) from ±xs to ±xs under u(k) = ±1.
If πβ is irrational, then R and C have infinite many extremal points.
Since R and C depend continuously on the state matrix A, specifically on
β, they can be arbitrarily approximated with those having rational πβ . On
the other hand, we can also use a trajectory to approximate the extremal
points of R and C. Let
v ∗ (k) = sign(sin(βk))
and z ∗ (k) be the zero initial state response to the time-reversed system.
Although the control and the time response are not exactly periodic, there
exist a set of limit points of the state trajectories, which also form the
extremal points of R and C.
3.6. An Example
Consider the continuous-time system

0 1 0
ẋ = Ac x + bc u = x+ u, (3.6.1)
−0.5 1.5 1
and its time-reversed system
ż = −Ac z − bc v. (3.6.2)
3.7. Asymptotically Null Controllable Region 51
From Chapter 2, the boundary of the null controllable region of (3.6.1) is

∂C c = ± I − 2e−Ac t A−1 c b c : 0 ≤ t ≤ ∞
t
= ± e−Ac t x+e + e −Ac (t−τ )
(−b c )(−1)dτ : 0 ≤ t ≤ ∞ ,
0
where x+ −1
e = −Ac bc is the equilibrium point of (3.6.2) under the constant
control u = 1. The second equality shows that ∂C c is formed by the trajec-
tories of (3.6.2) starting from ±x+
e under the control v = ∓1.
With h the sampling period and
Ac h
A = e Ac h , b = A−1
c e − I bc ,
we consider the discrete-time system
x(k + 1) = Ax(k) + bu(k) (3.6.3)
under different h. From Theorem 3.4.1 we have

Ext(C) = ± I − 2A−k (I − A)−1 b : k = 0, 1, 2, · · ·

hk
= ± e−Ac hk x+
e + e −Ac (hk−τ )
(−bc )(−1)dτ : k = 0, 1, 2, · · · .
0
We note here that (I − A)−1 b = −A−1 c bc . The above equation shows

that the extremal points of C are the points on ∂C c , the trajectories of
(3.6.2) starting from ±x+ e under the control v = ∓1 at times hk, k =
0, 1, 2, · · · . The boundaries of C corresponding to different sampling periods
h = 0.1, 0.2, 1, 2, 4, 8 are plotted in Fig. 3.6.1. When h = 0.1, C is very
close to C c of the continuous-time system; when h = 8, C is diminished to
a narrow strip.
3.7. Asymptotically Null Controllable Region
As discussed in Section 2.7 in the context of continuous-time systems,

asymptotically null controllability is a natural extension of the notion of
null controllability. This relaxation is introduced to deal with the situation
where the linear system contains some uncontrollable, but stable, poles.
0.8
h=1
0.6 h=2 h=0.1,0.2
0.4
h=4
0.2
h=8
0
-0.2
-0.4
-0.6
-0.8
-1
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
Figure 3.6.1: Controllable regions under different sampling periods.
Definition 3.7.1. Consider the system (3.2.1). A state x0 is said to be

asymptotically null controllable if there exists an admissible control u such
that the state trajectory x(k) of the system satisfies x(0) = x0 and
lim x(k) = 0.
k→∞
The set of all states that are asymptotically null controllable, denoted by
C a , is called the asymptotically null controllable region.
The characterization of the asymptotically null controllable region is

trivial and can be carried out in terms of the null controllable region of
the anti-stable subsystem. Similar to Theorem 2.7.1 in the continuous-time
case, we have the following result.
Theorem 3.7.1. Consider the system (3.2.1). Assume that (A, B) is sta-
bilizable and is given in the following form,

A1 0 B1
A= , B= ,
0 A2 B2
3.8. Conclusions 53
where A1 ∈ Rn1 ×n1 is anti-stable and A2 ∈ Rn1 ×n2 is semi-stable. Then,

the asymptotically null controllable region is given by
C a = C 1 × Rn2 ,
where C 1 is the null controllable region of the unstable subsystem
x1 (k + 1) = A1 x1 (k) + B1 u(k).
It is clear that, if the system is controllable, then its asymptotically null

controllable region and null controllable region are identical.
3.8. Conclusions
In this chapter, we have presented an explicit description of the extremal
points of the null controllable region of a discrete-time linear system with
bounded controls. The continuous-time counterparts of the results pre-
sented here was given in the previous chapter. Throughout the book we
will be concerned with feedback laws that are valid on the entire null con-
trollable region or a large portion of it.
Chapter 4
Stabilization on Null
Controllable Region –
Continuous-Time Systems
4.1. Introduction
In this chapter, we will study the problem of stablizing a linear system with
saturating actuators. The key issue involved here is the size of the result-
ing domain of attraction of the equilibrium. Indeed, local stabilization, for
which the size of the domain of attraction is not a design specification, is
trivial. It is straightforward to see that any linear feedback that stabilizes
the system in the absence of actuator saturation would also locally sta-
bilize the system in the presence of actuator saturation. In fact, with the
given stabilizing linear feedback law, actuator saturation can be completely
avoided by restricting the initial states to a small neighborhood of the equi-
librium. Our focus in this chapter is on the construction of feedback laws
that would lead to a domain of attraction that contains any a priori given
bounded subset of the asymptotically null controllable region in its interior.
We refer to such a problem as semi-global stabilization on the asymptoti-
cally null controllable region, or simply, semi-global stabilization. We recall
from Chapter 2 that the null controllable region of a linear system subject
to actuator saturation is the set of all the states that can be steered to
the origin in a finite time by an admissible control, and the asymptotically
55
56 Chapter 4. Stabilization – Continuous-Time Systems
null controllable region is the set of all the states that can be driven to the
origin asymptotically by an admissible control.
Recent literature has witnessed a surge of interest in the stabilization of
linear systems subject to actuator saturation. Most of the existing results,
however, pertain to systems that are asymptotically null controllable (AN-
CBC). We recall from Section 2.7 that the asymptotically null controllable
region of an ANCBC system is the entire state space. For these systems,
it was shown in [93] that global stabilization is possible. A nested feed-
back design technique for constructing nonlinear globally asymptotically
stabilizing feedback laws was proposed in [98] for a chain of integrators and
was fully generalized in [95]. Alternative solutions to the global stabiliza-
tion problem consisting of scheduling a parameter in an algebraic Riccati
equation according to the size of the state vector were later proposed in
[94,99]. The question of whether or not a general linear ANCBC system
subject to actuator saturation can be globally asymptotically stabilized by
a linear feedback was answered in [24,96], where it was shown that a chain
of integrators of length greater than two cannot be globally asymptotically
stabilized by any saturated linear feedback. In a search for simple control
strategies, we have earlier constructed linear feedback laws ([67,65]) that
achieve semi-global stabilization of these systems.
The objective of this chapter is to construct feedback laws that would
achieve semi-global stabilization of linear systems that have one or two
exponentially unstable poles and are subject to actuator saturation. Semi-
global stabilization for general systems will be treated in Chapter 9. Our
motivation for considering separately systems with one or two exponen-
tially unstable poles here is twofold. First, for such systems, we are able to
construct simple feedback laws. For a planar system with both poles expo-
nentially unstable or systems having no exponentially unstable poles, linear
feedback is sufficient. For systems having both polynomially unstable poles
and one or two exponentially unstable poles, two linear feedback laws, one
for inside a certain invariant set and the other for the outside, would be
sufficient. Second, in establishing these stabilization results, we will arrive
at some results that are interesting in their own right. For example, we
show that, for a planar anti-stable linear system under any saturated linear
stabilizing feedback law, the boundary of the domain of attraction is formed
by the unique limit cycle of the closed-loop system. Our presentation in
this chapter is based on our recent work [42].
4.2. Domain of Attraction under Saturated Linear Feedback 57
In Section 4.2, we establish the just mentioned fact that, for a planar
anti-stable linear system under any stabilizing saturated linear feedback,
the boundary of the domain of attraction is formed by the unique limit
cycle of the closed-loop system. In Section 4.3, we will show how to appro-
priately choose the linear feedback gain such that this unique limit cycle
will approach the boundary of the null controllable region of the system.
We will thus establish semi-global stabilization of such systems. In Section
4.4 we will establish semi-global stabilization for higher order systems with
one or two exponentially unstable poles. In Section 4.5, we will draw a brief
conclusion to the chapter.
4.2. Domain of Attraction – Planar System under

Saturated Linear Feedback
Consider the open loop system
ẋ = Ax + bu, (4.2.1)
where x ∈ Rn is the state and u ∈ R is the control. Because of actuator

saturation, the actual control that is delivered to the system belongs to the
following set

Ua = u : u is measurable and |u(t)| ≤ 1, ∀ t ∈ R . (4.2.2)
We refer to any control u ∈ Ua as an admissible control. A saturated linear

state feedback is given by
u = sat(f x), (4.2.3)
where f ∈ R1×n is the feedback gain and sat : R → R is the standard

saturation function as defined in Chapter 1. The feedback law u = sat(f x)
is said to be stabilizing if A + bf is asymptotically stable. With a saturated
linear state feedback applied, the closed loop system is given by
ẋ = Ax + b sat(f x). (4.2.4)
Denote the state transition map of (4.2.4) by φ : (t, x0 ) → x(t). The domain
of attraction S of the equilibrium x = 0 of (4.2.4) is defined by

S := x0 ∈ Rn : lim φ(t, x0 ) = 0 .
t→∞
The objective of this section is to determine the domain of attraction S

for anti-stable planar systems under any saturated linear stabilizing feed-
back law of the form (4.2.3). In [2], it was shown that the boundary of S,
denoted by ∂S, is a closed trajectory, but no method for finding this closed
trajectory is provided. Generally, only a subset of S lying between f x = 1
and f x = −1 is detected as a level set of some Lyapunov function (see, e.g.,
[27]). Let P be a positive definite matrix such that
(A + bf )T P + P (A + bf ) < 0.
Since {x ∈ R2 : −1 < f x < 1} is an open neighborhood of the origin, it

must contain
E(P, ρ0 ) := x ∈ R2 : xT P x ≤ ρ0 (4.2.5)
for some ρ0 > 0. Clearly, E(P, ρ0 ) ⊂ S is an invariant set and serves as

an estimate of the domain of attraction. This estimation of the domain of
attraction can be very conservative (see, e.g., Fig. 4.2.1).
Lemma 4.2.1. The origin is the unique equilibrium point of the system
(4.2.4).
Proof. This is a result from [2]. A simpler proof goes as follows. The other
two candidate equilibrium points are x+ e = −A
−1
b and x− + +
e = −xe . For xe
to be an equilibrium, we must have f x+ +
e ≥ 1 so that u = sat(f xe ) = 1.
Since A is anti-stable and (A, b) is controllable, we can assume, without
loss of generality, that

0 1
A= , a1 , a2 > 0,
−a1 a2
and
0
b= .
1

This implies that if f = f1 f2 is stabilizing, then f1 < a1 . It can be
verified that
f1
f x+
e = −f A
−1
b= < 1.
a1
This rules out x+ −
e . Similarly, we can rule out xe . 2
Let us introduce the time-reversed system of (4.2.4),
ż = −Az − b sat(f z). (4.2.6)

Clearly (4.2.6) also has only one equilibrium point, an unstable one, at the
origin. Denote the state transition map of (4.2.6) as ψ : (t, z0 ) → z(t).
Theorem 4.2.1. ∂S is the unique limit cycle of the planar systems (4.2.4)
and (4.2.6). Furthermore, ∂S is the positive limit set of ψ(·, z0 ) for all
z0
= 0.
This theorem says that ∂S is the unique limit cycle of (4.2.4) and (4.2.6).
This limit cycle is a stable one for (4.2.6) (in a global sense) but an unstable
one for (4.2.4). Therefore, it is easy to determine ∂S by simulating the
time-reversed system (4.2.6). Shown in Fig. 4.2.1 is a typical result, where
two trajectories, one starting from outside, the solid curve, and the other
starting from inside, the dashed curve, both converge to the unique limit
cycle. The straight lines in Fig. 4.2.1 are f z = 1 and f z = −1.
-2
-4
-6
-8
-10 -8 -6 -4 -2 0 2 4 6 8 10
Figure 4.2.1: Determination of ∂S from the limit cycle.
To prove Theorem 4.2.1, we need some properties of the trajectories of

second order linear systems. The following two lemmas are direct conse-
quences of Lemmas 6.3.1 and 6.3.2 in Chapter 6.
Lemma 4.2.2. Suppose that A ∈ R2×2 is anti-stable and (f, A) is observ-

able. Given c > 0, let x1 , x2 , y1 and y2 (x1
= x2 ) be four points on the line
f x = c, satisfying
y1 = eAT1 x1 , y2 = eAT2 x2
for some T1 , T2 > 0 and
f eAt1 x1 > c, f eAt2 x2 > c, ∀ t1 ∈ (0, T1 ), t2 ∈ (0, T2 ),
then, |y1 − y2 | > |x1 − x2 |.
For an illustration of Lemma 4.2.2, see Fig. 4.2.2. In Fig. 4.2.2, the
curve from xi to yi is x(t) = eAt xi , t ∈ [0, Ti ], a segment of a trajectory
of the autonomous system ẋ = Ax. Lemma 4.2.2 indicates that if any two
different trajectories leave certain straight line on the same side, they will
be further apart when they return to it.
2.5
1.5
y y x x
2 1 1 2
1
0.5
−0.5
−1
−10 −5 0 5
Figure 4.2.2: Illustration of Lemma 4.2.2.
Lemma 4.2.3. Suppose that A ∈ R2×2 is asymptotically stable and (f, A)

is observable. Given c > 0, let x1 and x2 be two points on the line f x = c
and y1 , y2 be two points on f x = −c such that
y1 = eAT1 x1 , y2 = eAT2 x2
for some T1 , T2 > 0, and

At
f e 1 x1 < c, f eAt2 x2 < c, ∀ t1 ∈ (0, T1 ), t2 ∈ (0, T2 ),
then, |y1 − y2 | > |x1 − x2 |.
For an illustration of Lemma 4.2.3, see Fig. 4.2.3. It says that if two
different trajectories of the autonomous system ẋ = Ax enter the region
between f x = c and f x = −c, they will be further apart when they leave the
region. Notice that in Lemma 4.2.2, A is anti-stable, and in Lemma 4.2.3,
A is asymptotically stable.
1.5
x x
2 1
1
0.5
−0.5
y y
2 1
−1
−1.5
−2
−10 −5 0 5
It should be noted that in Lemmas 6.3.1 and 6.3.2, special forms of A and
f are assumed. Since (f, A) is observable, they can always be transformed
into the special form. Lemmas 6.3.1 and 6.3.2 are applicable to the general
form of (f, A) since linear transformation does not change the ratio between
the lengths of the two aligned vectors, y1 − y2 and x1 − x2 .
Proof of Theorem 4.2.1. We first prove that, for the system (4.2.6), every
trajectory ψ(t, z0 ), z0
= 0, converges to a periodic orbit as t → ∞. Recall
that E(P, ρ0 ) (defined in (4.2.5)) lies within the domain of attraction of the
equilibrium x = 0 of (4.2.4) and is an invariant set. Since the trajectories of

(4.2.4) and those of (4.2.6) are the same but traverse in opposite directions,
it follows that, for every state z0
= 0 of (4.2.6), there is some t0 ≥ 0 such
that ψ(t, z0 ) lies outside E(P, ρ0 ) for all t ≥ t0 . The state transition map
of the system (4.2.6) is,
t
−At
ψ(t, z0 ) = e z0 − e−A(t−τ )b sat(f z(τ ))dτ. (4.2.7)
0
Since −A is stable, the first term converges to the origin. Since
|sat(f z(τ ))| ≤ 1,
the second term belongs to C, the null controllable region of (4.2.1), for all
t. It follows that there exists a ρ1 > ρ0 such that
ψ T (t, z0 )P ψ(t, z0 ) ≤ ρ1 < ∞, ∀ t ≥ t0 .
Let
Q = z ∈ R 2 : ρ0 ≤ z T P z ≤ ρ1 .
Then, ψ(t, z0 ), t ≥ t0 , lies entirely in Q. It follows from Poincaré-Bendixon

Theorem (see e.g., [55]) that ψ(t, z0 ) converges to a periodic trajectory.
The preceding paragraph shows that systems (4.2.4) and (4.2.6) have
periodic trajectories. We claim that the two systems each has only one
periodic trajectory. For direct use of Lemma 4.2.2 and Lemma 4.2.3, we
prove this claim through the original system (4.2.4).
First notice that a periodic trajectory must enclose the unique equilib-
rium point x = 0 by the index theory (see e.g., [55]), and must be symmetric
to the origin (since the vector field is symmetric, −Γ is a periodic trajectory
if Γ is, hence if a periodic trajectory Γ is not symmetric, there will be two
intersecting trajectories). Also, it cannot be completely contained in the
linear region between f x = 1 and f x = −1 (otherwise the asymptotically
stable linear system
ẋ = (A + bf )x
would have a closed trajectory in this region. This is impossible). Hence it
has to intersect each of the lines f x = ±1 at least twice. Assume, without
loss of generality, that (f, A, b) is in the observer canonical form, i.e.,

0 −a1 b1
f = 0 1 , A= , b= ,
1 a2 b2
with a1 , a2 > 0, and denote

ξ1
x= .
ξ2
In this case, f x = ±1 are horizontal lines. The stability of A + bf requires

that −a1 + b1 < 0 and a2 + b2 < 0. Observe that on the line f x = 1, we
have ξ2 = 1 and
ξ˙2 = ξ1 + a2 + b2 .
Hence, if ξ1 > −a2 − b2 , then,
ξ̇2 > 0,
i.e., the trajectories go upwards, and if ξ1 < −a2 − b2 , then,
ξ̇2 < 0,
i.e., the trajectories go downwards. This implies that every periodic trajec-
tory crosses f x = 1 exactly twice. Similarly, every periodic trajectory also
crosses f x = −1 exactly twice. It also implies that a periodic trajectory
goes counterclockwise.
Now suppose on the contrary that (4.2.4) has two different periodic tra-
jectories Γ1 and Γ2 , with Γ1 enclosed by Γ2 , as illustrated in Fig. 4.2.4.
Since any periodic trajectory must enclose the origin and any two trajec-
tories cannot intersect, all the periodic trajectories must be ordered by
enclosement. Let x1 and y1 be the two intersections of Γ1 with f x = 1,
and x2 , y2 be the two intersections of Γ2 with f x = 1. Then along Γ1 , the
trajectory goes from x1 to y1 , −x1 , −y1 and returns to x1 ; and along Γ2 ,
the trajectory goes from x2 to y2 , −x2 , −y2 and returns to x2 .
Let
x+
e = −A
−1
b.
Since x1 → y1 along Γ1 and x2 → y2 along Γ2 are on the trajectories of

% &
d(x − x+
e ) +
ẋ = Ax + b, or = A(x − xe ) ,
dt
we have
y1 − x+
e =e
AT1
(x1 − x+
e ), y2 − x+
e = e
AT2
(x2 − x+
e )
2.5
2 Γ2
Γ1
1.5
y y x1 x
2 1 2
1
0.5 +
xe
−0.5
−x −x −y −y
2 1 1 2
−1
−1.5
−2
−2.5
−10 −8 −6 −4 −2 0 2 4 6 8 10
Figure 4.2.4: Illustration for the proof of Theorem 4.2.1.
for some T1 , T2 > 0. Furthermore, since

b1
f x+
e = < 1,
a1
we have
f (x1 − x+ + + + +
e ) = f (x2 − xe ) = f (y1 − xe ) = f (y2 − xe ) = 1 − f xe > 0,
and for all x on the two pieces of trajectories,
f (x − x+ +
e ) ≥ 1 − f xe .
It follows from Lemma 4.2.2 that
|y2 − y1 | > |x2 − x1 |. (4.2.8)
On the other hand, y1 → −x1 along Γ1 and y2 → −x2 along Γ2 are on

trajectories of
ẋ = (A + bf )x,
and satisfy
−x1 = e(A+bf )T3 y1 , −x2 = e(A+bf )T4 y2 ,
for some T3 , T4 > 0. It follows from Lemma 4.2.3 that
|x2 − x1 | > |y2 − y1 |,
which is a contradiction to (4.2.8). Therefore, Γ1 and Γ2 must be the same

periodic trajectory. This shows that the systems have only one periodic
trajectory and hence it is a limit cycle.
We have so far proven that both systems (4.2.4) and (4.2.6) have a
unique limit cycle and every trajectory ψ(t, z0 ), z0
= 0, of (4.2.6) converges
to this limit cycle. This implies that a trajectory φ(t, x0 ) of (4.2.4) converges
to the origin if and only if x0 is inside the limit cycle. Therefore, the limit
cycle is ∂S. 2
Remark 4.2.1.
1) In the above proof, we also showed that ∂S is symmetric and has two
intersections with f x = 1 and two with f x = −1;
2) This proof relies heavily on the symmetry of the saturation function.

An alternative proof, which does not require symmetry of the satu-
ration function can be found in [46].
Another nice feature of S is that it is convex.
Proposition 4.2.1. S is convex.
Proof. We start the proof by showing some general properties of a second

order linear autonomous system
ẋ = Ax, det (A)

= 0. (4.2.9)
Denote θ = x, r = |x|, then,

cos θ
x=r ,
sin θ
and it can be verified that

0 1 cos θ
θ̇ = cos θ sin θ A . (4.2.10)
−1 0 sin θ
Since
0 1 cos θ
cos θ sin θ = 0,
−1 0 sin θ
equation (4.2.10) has at most four equilibria on [0, 2π), which correspond to
the directions of the real eigenvectors of A. If θ(0) is an equilibrium, then
θ(t) is a constant, and x(t) is a straight line. If θ(0) is not an equilibrium,
then θ(t) will never reach an equilibrium at finite time, otherwise the tra-
jectory x(t) would intersect a straight line trajectory, which is impossible.
Hence, if x0 is not an eigenvector of A, θ̇(t) will never be equal to zero, thus
θ̇(t) will be sign definite. This shows that x(t) is strictly monotonically
increasing (or decreasing).
Let us now consider the direction angle of the trajectory, ẋ =: γ. Since
ẍ(t) = Aẋ(t), by the same argument, γ(t) also increases (or decreases)
monotonically. We claim that if A is asymptotically stable or anti-stable
(det (A) > 0), then θ̇(t) and γ̇(t) have the same sign, i.e., the trajectories
bend toward the origin; and if the signs of the two eigenvalues of A are
different (det (A) < 0), then θ̇(t) and γ̇(t) have opposite signs. This can be
simply shown as follows. Rewrite (4.2.10) as,

1 T 0 1
θ̇ = x Ax. (4.2.11)
|x|2 −1 0
Similarly,

1 T 0 1 1 T T 0 1
γ̇ = ẋ Aẋ = x A AAx.
|ẋ|2 −1 0 |ẋ|2 −1 0
It is trivial to verify that, for a 2 × 2 matrix A,

T 0 1 0 1
A A = det (A) .
−1 0 −1 0
Hence
det (A) T 0 1 |x|2 det (A)
γ̇ = x Ax = θ̇. (4.2.12)
|ẋ|2 −1 0 |ẋ|2
This shows that the claim is true.
Now we can apply the above claim to the limit cycle of the system
(4.2.4) (refer to Fig. 4.2.4). From x1 to y1 , [x(t) − x+
e ] increases, so ẋ(t)
increases. From y1 to −x1 , x(t) and ẋ(t) also increase. Similarly, ẋ(t)
increases from −x1 to −y1 , and from −y1 to x1 . It is straightforward to
verify that ẋ(t) is continuous at x1 , y1 , −x1 and −y1 . So ẋ(t), the direction
angle, is monotonically increasing along the limit cycle. This implies that
the region enclosed by the limit cycle, S, is convex. 2
4.3. Semi-Global Stabilization – Planar Systems 67
4.3. Semi-Global Stabilization – Planar Systems
In this section, we will establish semi-global stabilizability on the null con-

trollable region for second order anti-stable linear systems subject to actu-
ator saturation. In particular, we will show that the domain of attraction
of a second order anti-stable linear system under a saturated linear state
feedback law can be made to enclose any a priori given bounded subset
of the null controllable region in its interior by judiciously choosing the
feedback gain. In the next section, we will utilize this design method to
construct simple nonlinear feedback laws that semi-globally stabilize higher
order systems with two exponentially unstable poles.
Consider the system (4.2.1), with A ∈ R2×2 anti-stable and (A, b) con-
trollable. Let P be the unique positive definite solution to the following
algebraic Riccati equation (ARE),
AT P + P A − P bbT P = 0. (4.3.1)
Note that this equation is associated with the minimum energy regulation,
i.e., an LQR problem with cost
∞
J= uT (t)u(t)dt.
0
The corresponding minimum energy state feedback gain is given by
f0 = −bT P.
By the infinite gain margin and 50% gain reduction margin property of
LQR regulators, the origin is a stable equilibrium of the system
ẋ = Ax + b sat(kf0 x) (4.3.2)
for all k > 0.5. Let S(k) be the domain of attraction of the equilibrium x =
0 of (4.3.2). Then, the following result establishes semi-global stabilizability
by linear feedback.
Theorem 4.3.1.
lim dist(S(k), C) = 0.
k→∞
Proof. For simplicity and without loss of generality, we assume that

0 −a1
A= , a1 , a2 > 0,
1 a2
and
0
b= .
−1
Since A is anti-stable and (A, b) is controllable, A and b can always be
transformed into this form. Suppose that A has already taken this form
and
b1
b= .
b2
Let

V = −A−1 b −b ,
then, V is nonsingular and it can be verified that

0
V −1 AV = A, V −1 b = .
−1
With this special form of A and b, we have

a2
a1 0 0 −a1
P =2 , f0 = 0 2a2 , A + kbf0 = ,
0 a2 1 a2 (1 − 2k)
and
1 −1
ze+ = −A−1
b= , ze− = −ze+ = .
0 0
We also have f0 A−1 b = 0.
For a given k > 0.5, by Theorem 4.2.1, the system (4.3.2) has a unique
limit cycle which is the boundary of S(k). To visualize the proof, ∂C and
∂S(k) for some k are plotted in Fig. 4.3.1, where the inner closed curve is
∂S(k), and the outer one is ∂C.
We recall that when the eigenvalues of A are real (see (2.4.2)),
t
−At − −A(t−τ )
∂C = ± e ze − e bdτ : t ∈ [0, ∞] , (4.3.3)
0
and when the eigenvalues of A are complex (see (2.5.2)),

t
−At − −A(t−τ )
∂C = ± e zs − e bdτ : t ∈ [0, Tp ] . (4.3.4)
0
0.6
0.4
0.2
x1 −y1
−
ze
0
y1 −x1
−0.2
−0.4
−0.6
−0.8
−1.5 −1 −0.5 0 0.5 1 1.5
Figure 4.3.1: The domain of attraction and the null controllable region.
On the other hand, ∂S(k) is the limit cycle of the time-reversed system of
(4.3.2),
ż = −Az − b sat(kf0 z). (4.3.5)
Here the limit cycle as a trajectory goes clockwise. From Remark 4.2.1,
we know that the limit cycle is symmetric and has two intersections with
kf0 z = 1 and two with kf0 z = −1 (see Fig. 4.3.1). Let T be the time
required for the limit cycle trajectory to go from y1 to x1 , and T2 the time
from x1 to −y1 , then

∂S(k) = ±e−(A+kbf0 )t y1 : t ∈ [0, T ]
t
∪ ± e−At x1 − e−A(t−τ ) bdτ : t ∈ [0, T2 ] . (4.3.6)
0
Here and in the sequel, the dependence of x1 , y1 , T and T2 on k is omitted

for simplicity.
As k → ∞, the distance between the lines kf0 z = 1 and kf0 z = −1
approaches zero. By comparing (4.3.3), (4.3.4) and (4.3.6), we see that to
prove the theorem, it suffices to show that
lim T = 0,
k→∞
lim x1 = lim y1 = ze− (or zs− ),

k→∞ k→∞
and
lim T2 = ∞ (or Tp ).
k→∞
If these are true, the lengths of the parts of the limit cycle between the
lines kf0 z = 1 and kf0 z = −1 will tend to zero. We will first show that
lim T = 0.
k→∞
Let
x11 y11
x1 = 1 , y1 = 1 ,
2ka2 − 2ka 2
then, kf0 x1 = 1, kf0 y1 = −1,

x11 y11
1 = e−(A+kbf0 )T 1 , (4.3.7)
2ka2 − 2ka 2
and

−(A+kbf0 )t
y11
kf0 e 1 ≤ 1, ∀ t ∈ [0, T ].
− 2ka 2

We also note that the upward movement of the trajectory at x1 and y1
implies that
2k − 1 1 − 2k
x11 < , y11 < .
2k 2k
As k → ∞, the matrix

0 −a1
A + kbf0 =
1 a2 (1 − 2k)
has two distinct real eigenvalues −λ1 and −λ2 (Their dependence on k is
also omitted). Assume λ2 > λ1 . Since λ1 λ2 = a1 and λ1 + λ2 = a2 (2k − 1),
we have
lim λ1 = 0, lim λ2 = +∞.
k→∞ k→∞
Let
λ2 λ1
V = .
1 1
It can be verified that

−λ1 0
A + kbf0 = V V −1 .
0 −λ2
It follows that
−1
(A+kbf0 )T λ2 λ1 e−λ1 T 0 λ2 λ1
e = ,
1 1 0 e−λ2 T 1 1
and from (4.3.7), we obtain
1 λ2 − λ1 + λ2 e−λ2 T − λ1 e−λ1 T
x11 = ,
2ka2 e−λ2 T − e−λ1 T
and,
1 λ2 − λ1 + λ2 eλ2 T − λ1 eλ1 T
y11 = .
2ka2 eλ1 T − eλ2 T
Since
1 − 2k λ1 + λ2
y11 < =−
2k 2ka2
and
eλ1 T − eλ2 T < 0,
we have
λ1 eλ2 T < λ2 − λ1 + λ2 eλ1 T < 2λ2 eλ1 T
and
ln 2λ 2
1 2λ2
T < λ1
= ln 2 ,
λ2 − λ1 λ2 − λ1 a1
where we note that
a1
λ1 = .
λ2
Since
lim λ2 = ∞, lim λ1 = 0,
k→∞ k→∞
we obtain
lim T = 0.
k→∞
It follows that
y11 λ2 − λ1 + λ2 eλ2 T − λ1 eλ1 T
lim = lim
k→∞ x11 k→∞ (λ2 − λ1 )e(λ1 +λ2 )T + λ2 eλ1 T − λ1 eλ2 T
λ2 − λ1 1+e
λ1 T
1+eλ2 T
= lim λ2 T (1+eλ1 T ) = 1,
k→∞ λ2 eλ1 T − λ1 e 1+eλ2 T
where we have used the fact that
lim λ1 = 0.
k→∞
Since x1 and y1 are bounded by the null controllable region, x11 and y11
are finite numbers. Hence,
lim (y11 − x11 ) = 0. (4.3.8)

k→∞
On the limit cycle of the time-reversed system (4.3.5), we also have

T2
−y1 = e−AT2 x1 − e−A(T2 −τ ) bdτ,
0
i.e.,
y11 −AT2
x11
1 = −e 1 + I − e−AT2 A−1 b,
− 2ka2 2ka2

y 11 y 11 − x11 0
I + e−AT2 = I − e−AT2 A−1 b+e−AT2 1 + 1 .
0 − 2ka 2 2ka2
It follows from (4.3.8) that

y11
−AT2 −1 −AT2
−1
lim − I +e I −e A b = 0.
k→∞ 0
Hence,
−1
lim 0 1 I + e−AT2 I − e−AT2 A−1 b = 0.
k→∞
For different cases, it can be shown from the above equality that
1. If the eigenvalues of A are real, then

y11
lim T2 = ∞, lim y1 = lim x1 = lim = ze− .
k→∞ k→∞ k→∞ k→∞ 0
2. If the eigenvalues of A are complex, then

y11
lim T2 = Tp , lim y1 = lim x1 = lim = zs− .
k→∞ k→∞ k→∞ k→∞ 0
This completes the proof. 2

We would like to note that the use of high gain feedback is crucial in
establishing the results of the theorem. The minimum energy feedback f0
itself does not give a domain of attraction close to C. This is quite different
from the related result in [65,67] for semi-stable open loop systems. In the
later case, it was shown that if (A, B) is ANCBC, then low gain feedback
can lead to arbitrarily large domain of attraction.
In what follows, we will use two examples to demonstrate the results
presented in this section.
Example 4.3.1. Consider the system (4.2.1) with

0 −0.5 0
A= , b= .
1 1.5 −1
The matrix A has two real eigenvalues at 0.5 and 1. Following the design
procedure, we obtain

f0 = 0 3 .
In Fig. 4.3.2, the boundaries of the domains of attraction corresponding to

different f = kf0 , k = 0.50005, 0.65, 1, 3, 10, are plotted from the inner to
the outer. The domains of attraction do become bigger for greater k. The
outermost dashed boundary is ∂C. When k = 10, it can be seen that ∂S(k)
is very close to ∂C.
0.8
0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Figure 4.3.2: Domains of attraction under different feedbacks: the real

eigenvalue case.

0.6 −0.8 2
A= , b= .
0.8 0.6 4
The matrix A has a pair of complex eigenvalues λ = 0.6 ± j0.8. Following

the design procedure, we obtain

f0 = 0.12 −0.66 .
In Fig. 4.3.3, the boundaries of the domains of attraction corresponding to

different f = kf0 , k = 0.50001, 0.6, 0.8, 1, 2, are plotted from the inner to
the outer. The outermost dashed curve is ∂C. We see that when k = 2, the
domain of attraction is almost the same as the null controllable region.
−1
−2
−3
−4
−5
−6 −4 −2 0 2 4 6
Figure 4.3.3: Domains of attraction under different feedbacks: the complex

eigenvalue case.
4.4. Semi-Global Stabilization – Higher Order Systems
In this section, we will establish semi-global stabilizability for higher order

systems with one or two exponentially unstable poles. In particular, we
will focus on systems with two exponentially unstable poles. Following the
development for the two exponentially unstable pole systems, we can treat
the single exponentially unstable pole case in a straightforward way.
4.4. Semi-Global Stabilization – Higher Order Systems 75
Consider the following open-loop system,

A1 0 b1
ẋ = Ax + bu = x+ u, (4.4.1)
0 A2 b2
where x = (xa , xs ), xa ∈ R2 , xs ∈ Rn , A1 ∈ R2×2 is anti-stable and

A2 ∈ Rn is semi-stable. Assume that (A, b) is stabilizable. Denote the null
controllable region of the subsystem
ẋa = A1 xa + b1 u
as C 1 , then the asymptotically null controllable region of (4.4.1) is
C a = C 1 × Rn .
Given γ1 , γ2 > 0, let

Ω1 (γ1 ) := γ1 xa ∈ R2 : xa ∈ C 1 , (4.4.2)
and
Ω2 (γ2 ) := xs ∈ Rn : |xs | ≤ γ2 . (4.4.3)
When γ1 = 1, Ω1 (γ1 ) = C 1 and when γ1 < 1, Ω1 (γ1 ) lies in the interior

of C 1 . Our objective here is to show that, given any γ1 < 1 and γ2 > 0,
an admissible state feedback can be designed such that Ω1 (γ1 ) × Ω2 (γ2 )
is contained in the domain of attraction of the equilibrium x = 0 of the
closed-loop system. This would establish the semi-global stabilizability of
the system (4.4.1).
For an ε > 0, let

P1 (ε) P2 (ε)
P (ε) = ∈ R(2+n)×(2+n)
P2T (ε) P3 (ε)
be the unique positive definite solution to the algebraic Riccati equation

(ARE),
AT P + P A − P bbT P + εI = 0. (4.4.4)
Clearly, as ε ↓ 0, P (ε) decreases. Hence limε→0 P (ε) exists.

Let P 1 be the unique positive definite solution to the ARE
AT1 P 1 + P 1 A1 − P 1 b1 bT1 P 1 = 0.
Then by the continuity property of the solution of the Riccati equation

[108],
P1 0
lim P (ε) = .
ε→0 0 0
Let
f (ε) := −bT P (ε).
First, consider the domain of attraction of the equilibrium x = 0 of the
following closed-loop system
ẋ = Ax + b sat(f (ε)x). (4.4.5)
It is easy to see that

% &
1
D(ε) := E P (ε), 1
|bT P 2 (ε)|2

1
= x ∈ R2+n : xT P (ε)x ≤ 1 ⊂ L(f (ε))
|bT P 2 (ε)|2
is contained in the domain of attraction of the equilibrium x = 0 of (4.4.5)
and is an invariant set. Note that if x0 ∈ D(ε), then x(t) ∈ D(ε) and
|f (ε)x(t)| ≤ 1 for all t > 0. That is, x(t) will stay in the linear region of
the closed-loop system, and in D(ε).
Lemma 4.4.1. Denote

1
r1 (ε) := 1 ,
2 T 12
2 P1 (ε) b P (ε)
'
−|P2 (ε)| + |P2 (ε)|2 + 3|P1 (ε)| |P3 (ε)|
r2 (ε) := r1 (ε).
|P3 (ε)|
Then,

D1 (ε) := x ∈ R2+n : |xa | ≤ r1 (ε), |xs | ≤ r2 (ε) ⊂ D(ε).
Moreover,
lim r2 (ε) = ∞,
ε→0
and r1 (ε) increases with an upper bound as ε tends to zero.
Proof. It can be verified that

1
|P1 (ε)|r12 (ε) + 2|P2 (ε)|r1 (ε)r2 (ε) + |P3 (ε)|r22 (ε) = . (4.4.6)
T 12 2
b P (ε)
So for all x ∈ D1 (ε),

1
xT P (ε)x ≤ ,
T 12 2
b P (ε)
i.e., D1 (ε) ⊂ D(ε). By the definition of r1 (ε) and r2 (ε), we have
3|P1 (ε)| 1
r2 (ε) = ' · 1 .
|P2 (ε)| + |P2 (ε)| + 3|P1 (ε)| |P3 (ε)| 2 P (ε) bT P 12 (ε)
2 2
1
Since as ε goes to zero, P2 (ε), P3 (ε) → 0 and P1 (ε) → P 1 , r1 (ε) is bounded

whereas r2 (ε) → ∞. It follows from the monotonicity of P (ε) that r1 is a
monotonically decreasing function of ε. 2
Theorem 4.4.1. Let f0 = −bT1 P1 . For any γ1 < 1 and γ2 > 0, there exist
a k > 0.5 and an ε > 0 such that Ω1 (γ1 )×Ω2 (γ2 ) is contained in the domain
of attraction of the equilibrium x = 0 of the closed-loop system

sat(kf0 xa ), x ∈/ D(ε),
ẋ = Ax + bu, u = (4.4.7)
sat(f (ε)x), x ∈ D(ε).
Proof. Since γ1 < 1, by Theorem 4.3.1, there exists a k > 0.5 such that
Ω1 (γ1 ) lies in the interior of the domain of attraction of the equilibrium
xa = 0 of
ẋa = A1 xa + b1 sat(kf0 xa ). (4.4.8)
Let ε0 > 0 be given. For an initial state xa0 ∈ Ω1 (γ1 ), denote the trajectory
of (4.4.8) as ψ(t, xa0 ). Define

T (xa0 ) := min t ≥ 0 : |ψ(t, xa0 )| ≤ r1 (ε0 ) ,
then T (xa0 ) is the time when ψ(t, xa0 ) first enters the ball

B1 = xa ∈ R2 : |xa | ≤ r1 (ε0 ) .
Let
TM = max T (xa0 ) : xa0 ∈ ∂Ω1 (γ1 ) (4.4.9)
and

TM
A2 (TM −τ )
γ = max eA2 t γ2 + e b2 dτ, (4.4.10)
t∈[0,TM ] 0
then, by Lemma 4.4.1, there exists an ε < ε0 such that r1 (ε) ≥ r1 (ε0 ),
r2 (ε) ≥ γ and

D1 (ε) = x ∈ R2+n : |xa | ≤ r1 (ε), |xs | ≤ r2 (ε) ⊂ D(ε)
lies in the domain of attraction of the equilibrium x = 0 of (4.4.5).

Now consider an initial state of (4.4.7), x0 ∈ Ω1 (γ1 ) × Ω2 (γ2 ). If x0 ∈
D(ε), then x(t) will go to the origin since D(ε) is an invariant set and is
contained in the domain of attraction. If x0
∈ D(ε), we conclude that x(t)
will enter D(ε) at some T ≤ TM under the control u = sat(kf0 xa ). Observe
that under this control, xa (t) goes along a trajectory of (4.4.8). If there is
no switch, xa (t) will hit the ball B1 at time T (xa0 ). Clearly, T (xa0 ) ≤ TM
and at T (xa0 ),
|xs (T (xa0 ))| ≤ γ ≤ r2 (ε),
which implies that
x(T (xa0 )) ∈ D1 (ε).
Thus we see that if there is no switch, x(t) will be in D1 (ε) at T (xa0 ).
Since D1 (ε) ⊂ D(ε), x(t) must have entered D(ε) at some earlier time
T ≤ T (xa0 ) ≤ TM . So we have the desired conclusion. With the switching
control applied, once x(t) enters the invariant set D(ε), it will remain in it
and go to the origin asymptotically. 2
Remark 4.4.1.
1) For the case that the open-loop system has a single exponentially
unstable pole, Theorem 4.4.1 is also true. Let x, A, b and P (ε) be
partitioned similarly to the two exponentially unstable pole case, with
xa , A1 , b1 , P1 (ε) ∈ R. Then, it is easy to verify that
% &
|b1 | |b1 |
C1 = − , ,
A1 A1
and
2A1 2A1
P1 = , f0 = − .
b21 b1
Also, for all k > 0.5, the origin of the system
ẋa = A1 xa + b1 sat(kf0 xa )
has a domain of attraction equal to C 1 . Hence, the proof of Theo-

rem 4.4.1 can be readily adapted to this case.
2) For the case that the open-loop system has no exponentially unstable
poles (ANCBC), we have C a = Rn , and
lim P (ε) = 0.
ε→∞
Therefore, the set D(ε), which is in the domain of attraction for the
system under the control u = sat(f (ε)x), can be made arbitrarily
large to include any compact subset of Rn . This means that the
control u = sat(f (ε)x) can achieve semi-global stabilization and no
switching is necessary.
The proof of Theorem 4.4.1 provides a constructive method for com-

puting k and f (ε). In what follows, we illustrate this procedure with an
example.
Example 4.4.1. Consider an open-loop system described by (4.4.1) with

   
0.6 −0.8 0 0 0 2
 0.8 0.6 0 0 0   4 
   
A=  0 0 0 1 0 , b =  0 ,
  
 0 0 −1 0 0   1 
0 0 0 0 −1 −1
and
x = (xa , xs ), xa = (x1 , x2 ), xs = (x3 , x4 , x5 ).
The desired domain of attraction is Ω1 (γ1 )× Ω2 (γ2 ) with γ1 = 0.9 and γ2 =
10. Take ε0 = 0.000001, then r1 (ε0 ) = 0.5884. Here is the computational
result.
Step 1. We obtain

kf0 = 0.1360 −0.748 .
This is searched by comparing the boundary of Ω1 (γ1 ) with the limit
cycle of (4.4.8), which can always be made large enough to enclose
Ω1 (γ1 ).
Step 2. TM = 7.08 and γ = 17.2916 are computed from (4.4.9) and

(4.4.10) and each T (xa0 ) is obtained from simulation.
Step 3. Search ε such that r1 (ε) > r1 (ε0 ) and r2 (ε) > γ. One such ε is
given by ε = 0.45ε0, with r1 (ε) = 0.5887 > r1 (ε0 ), r2 (ε) = 17.7138 >
γ, and

f (ε) = 0.120097 −0.660525 0 0.000949 0 .
−2
−4
−6
0 100 200 300 400 500 600 700
Figure 4.4.1: Time response of x1 .
Applying the feedback matrices kf0 and f (ε) to the system (4.4.7), we
get the desired domain of attraction. Shown in Figs. 4.4.1 and 4.4.2 are the
time responses x1 (t) and x3 (t) of (4.4.7) with an initial state
 
4.7005
 0.70001 
 
x0 = 
 10 ,

 0 
0
which is on the boundary of Ω1 (γ1 ) × Ω2 (γ2 ).

The figures show that the convergence rate is very slow. This is because
A + bf (ε) has a pair of eigenvalues that are very close to the imaginary
axis, −0.0005 ± j1.
The convergence rate can be accelerated after the state enters D(ε) by
applying the piecewise-linear control law of [107]. The idea is as follows:
select a chain of εi , εN > εN −1 > · · · > ε1 > ε, compute P (εi ), f (εi ), then

1
D(εi ) = x ∈ R2+n : xT P (εi )x ≤ 1
|bT P 2 (εi )|2
15
10
−5
−10
−15
0 100 200 300 400 500 600 700
Figure 4.4.2: Time response of x3 .
is a sequence of nested invariant sets corresponding to each feedback control

u(t) = sat(f (εi )x(t)), i.e.,
D(εN ) ⊂ D(εN −1 ) ⊂ · · · ⊂ D(ε1 ) ⊂ D(ε).
With the following multiple switching control law



 sat(f (εN )x(t)), if x(t) ∈ D(εN ),




 sat(f (εN −1 )x(t)), if x(t) ∈ D(εN −1 ) \ D(εN ),

u= .. ..
 . .



 sat(f (ε)x(t)), if x(t) ∈ D(ε) \ D(ε1 ),



sat(kf0 xa (t)), if x(t)
∈ D(ε),
the convergence rate is increased. Applying the above control law, with
N = 20, to the system with the same initial state, the time response of x3
is plotted in Fig. 4.4.3. Fig. 4.4.4 is the plot of |x(t)|. In the figures, we see
that the convergence rate is indeed increased.
15
10
−5
−10
−15
0 100 200 300 400 500 600 700
Figure 4.4.3: Time-response of x3 under multiple switching control.
15
10
0
0 100 200 300 400 500 600 700
Figure 4.4.4: Time-response of |x| under multiple switching control.

4.5. Conclusions 83
4.5. Conclusions
In this chapter, we provided a simple semi-global stabilization strategy
for exponentially unstable linear systems with saturating actuators. For a
planar anti-stable system, the resulting feedback laws are saturated linear
state feedbacks and for higher order systems with one or two anti-stable
modes, they are piecewise linear state feedbacks with only one switch. As
a corollary, we also recovered the known result that an ANCBC system
can be semi-globally stabilized with a saturated linear feedback. The semi-
global stabilizability for general systems will be established in Chapter 9,
where more complicated feedback laws will be used.
Chapter 5
Stabilization on Null
Controllable Region –
Discrete-Time Systems
5.1. Introduction
In the previous chapter, we established semi-global stabilizability on the

null controllable region of continuous-time linear systems which have no
more than two exponentially unstable poles and are subject to actuator
saturation. In this chapter, we will establish similar results for discrete-
time systems.
Contrary to the usual expectation, the approach taken to establish these
results is completely different from that in the continuous-time case. We
recall that, in the continuous-time case, high gain feedback is crucial for
semi-global stabilization of the anti-stable subsystem. High gain feedback
is, however, infeasible for discrete-time systems. In fact, high gain feedback
in discrete-time systems is destabilizing. In this chapter, a feature unique
to a discrete-time system is captured such that a nonlinear controller can
be designed to globally stabilize a second order anti-stable subsystem on
its null controllable region. Moreover, this nonlinear controller produces a
deadbeat response. Based on this nonlinear controller, we will construct
another controller that achieves semi-global stabilization on the null con-
trollable region. Our motivation for sacrificing globality for semi-globality
85
86 Chapter 5. Stabilization – Discrete-Time Systems
is the simpler controller structure and that the controller is less sensitive
to system model uncertainties and disturbances.
Finally, as in the continuous-time case, we will untilize the above design
technique to construct feedback laws that achieve semi-global stabilization
on the null controllable region for higher order systems with no more than
two exponentially unstable poles. Our presentation in this chapter draws
on materials from [45].
In Sections 5.2 and 5.3, we will establish global stabilization on the
null controllable region for second order anti-stable linear systems subject
to actuator saturation. In particular, in Section 5.2, a saturated linear
feedback law is constructed that causes all the states in the null controllable
region to reach a set of equilibrium points in a finite number of steps. In
Section 5.3, based on this saturated linear feedback law, an overall nonlinear
feedback law is constructed that first causes all states in the null controllable
region to reach the set of equilibrium points in a finite number of steps and
then forces them from there to the origin, also in a finite number of steps.
Section 5.4 presents a semi-global stabilization strategy for second order
anti-stable linear systems. Section 5.5 deals with higher order systems with
one or two anti-stable poles. Finally, a brief conclusion to the chapter is
drawn in Section 5.6.
5.2. Global Stabilization at Set of Equilibria –

Planar Systems
Consider a second order anti-stable system,
x(k + 1) = Ax(k) + bu(k), (5.2.1)
where x(k) ∈ R2 , u(k) ∈ R, A is anti-stable and u∞ ≤ 1. The time-

reversed system of (5.2.1) is
z(k + 1) = A−1 z(k) − A−1 bv(k). (5.2.2)
Denote the equilibrium point of (5.2.1) under the constant control u(k) = 1
as x+
e := (I − A)
−1
b. This is also the equilibrium point of (5.2.2) under the
constant control v(k) = 1. Define x− +
e := −xe .
The study of stabilization of the above system depends on the eigen-
structure of the system matrix A. For simplicity, we will restrict our in-
vestigation on two cases, although Example 5.3.1 shows that the results
5.2. Global Stabilization at Set of Equilibria 87
developed for these two cases are also valid for the general case. These two
cases are:
Case 1. A has positive real eigenvalues, and,
Case 2. A has a pair of complex eigenvalues, eig(A) = r(cos(β)±j sin(β)),

where r > 1 and β = N1 π for some positive integer N > 2.
In Case 2, if (5.2.1) is a sampled-data system, β = N1 π means that the

sampling rate is 2N times that of the natural frequency of the continuous-
time system. In this case, AN = −rN I. Because N > 2, we have cos β ≥ 12 .
Denote
rN + 1 +
x+
s = N x , x− s = −xs ,
+
r −1 e
then x+ − +
s and xs are two extremal points of C (see (3.5.4)). Since r > 1, xe
and x− +
e are in the interior of C. This is different from Case 1 where xe and
−
xe are two extremal points of C.
To simplify the problem, we first perform a state space transformation.
Let

G = (I − A)−1 b (−2A−1 + I)b .
Since (A, b) is controllable, the matrix

b A−1 b
is nonsingular and hence G is also nonsingular. Let

−1 1
T = G−1 , Ā = T AT −1 , b̄ = T b,
0 1
with G defined by

G = (I − Ā)−1 b̄ (−2Ā−1 + I)b̄ ,
we have
−1 1
G = TG = .
0 1
Hence, for simplicity, we make the following assumption in this chapter.
Assumption 5.2.1. For the second order system (5.2.1), assume that A
and b satisfy

−1
−1
−1 1
(I − A) b (−2A + I)b = . (5.2.3)
0 1
If not so, a similarity transformation always exists to make (5.2.3) true.

Under this assumption, we have

−1 1
x+
e = (I − A)−1
b = , x −
e = .
0 0

Lemma 5.2.1. With f = −1 2 , we have

1 0
A + bf = ,
0 0
f x+
e = 1, f (A−1 x−
e −A
−1
b) = 1, (5.2.4)
and
f A−1 x+
e = 0, f A−1 b = −1. (5.2.5)
Proof. Multiplying both sides of (5.2.3) from the left with f yields
f (I − A)−1 b = 1, f (−2A−1 + I)(I − A)−1 b = 1. (5.2.6)
By Assumption 5.2.1,

−1
x+
e = (I − A)
−1
b= ,
0
so we have f x+
e = 1 and
1 = f (−2A−1 + I)(I − A)−1 b

= −f A−1 (I − A)−1 b + f (I − A−1 )(I − A)−1 b
= f A−1 x−
e − fA
−1
b.
Hence (5.2.4) holds. If we subtract the two equations in (5.2.6), we get
−2f A−1 (I − A)−1 b = 0 =⇒ f A−1 x+

e = fA
−1 −
xe = 0.
If we combine this with (5.2.4), we get f A−1 b = −1. Hence, (5.2.5) holds.
Now we have

−1 + + + −1
(A + bf ) = (A + bf )xe = Axe + b = xe = .
0 0
T
Multiplying both sides of (5.2.3) from right with 1 1 , we obtain

−1 0 0 0
−2A b = =⇒ 2b + A = . (5.2.7)
1 1 0
Therefore,
0 0 0
(A + bf ) =A + 2b = .
1 1 0

−1 0 −1 0 1 0
(A + bf ) = =⇒ (A + bf ) = .
0 1 0 0 0 0
2

Using this f = −1 2 , we construct a saturated linear feedback
u(k) = sat(f x(k)). (5.2.8)
Consider the closed-loop system
x(k + 1) = Ax(k) + b sat(f x(k)). (5.2.9)
We have the following main result of this section.
Theorem 5.2.1. Define

E := λx+ −
e + (1 − λ)xe : λ ∈ (0, 1) .
For any initial state x(0) ∈ C, the trajectory of (5.2.9) will reach the line
*
set E = E {x+ −
e , xe } in a finite number of steps and then remain in E.
Furthermore, E is the set of stable equilibrium points of (5.2.9).
This theorem is illustrated in Fig. 5.2.1 for Case 2, where ±x+ e are in
the interior of C. The two parallel lines are f x = ±1 and E is the horizontal
line segment between the two parallel lines.
Proof of Theorem 5.2.1. We will prove this theorem for Case 1 and
Case 2 separately.
Case 1. A has two real positive eigenvalues
In this case, recall (3.4.4) and (3.4.5) as follows,

Ext(C) = ± (2A−i − I)(I − A)−1 b : 1 ≤ i ≤ ∞

i−1
−i − −(i−−1) −1
= ± A xe + A (−A b)(1) : 1 ≤ i ≤ ∞ .
=0
1.5
0.5
−0.5
−1
−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
Figure 5.2.1: Illustration for Theorem 5.2.1.
Here we use the integer “i” instead of “” in (3.4.4) and “k” in (3.4.5). For
convenience, we have also replaced x+ −
e in (3.4.5) with xe and −1 in the
sum with 1 accordingly. We consider half of the set Ext(C),

i−1
A−i x−
e + A−(i−−1) (−A−1 b) : 1 ≤ i ≤ ∞ ,
=0
and for each integer i, we define

i−1

pi := A−i x−
e + A−(i−−1) −A−1 b .
=0
Let p0 = x− e . Then the extremal points of C, p0 , p1 , p2 , · · · , pi , · · ·, form

a trajectory of the time reversed system (5.2.2) under the control v(k) ≡
1 and with initial state z(0) = p0 = x− e . Conversely, for the original
system (5.2.1), if pi is the initial state, then under the control u(k) ≡ 1, the
trajectory will be pi−1 , pi−2 , · · · and reach p0 at the ith step. This implies
that
Api+1 + b = pi , i ≥ 0. (5.2.10)
Note that p1 = A−1 x−

e −A
−1
b. From (5.2.4), we have
f x+
e = 1, f p1 = 1. (5.2.11)
This shows that f x = 1 is a line passing through x+
e and p1 , as illustrated
in Fig. 5.2.2.
pi p4
p3
p2
p1 fx=1
x+ fx=-1
e
o xe- ( po )
-p
1
-p 2
-p
3
Figure 5.2.2: Illustration for Case 1.
The null controllable region C is cut into two symmetric parts by the
straight line passing through x+ −
e and xe . Denote the upper part with
+
pi , i ≥ 1, at its extremal points as C . This part is composed of the
triangles pi x+
e pi+1 , denoted by ∆i , i ≥ 0:

∆i := γ1 x+ e + γ2 pi + γ3 pi+1 : γ1 + γ2 + γ3 = 1, γ1 ∈ (0, 1), γ2 , γ3 ≥ 0 ,
where γ1 ∈ (0, 1) implies that ∆i does not include x+

e , nor the line segment
between pi and pi+1 .
The line f x = 1 cuts C + into two parts. One part is composed of
∆i , i ≥ 1, while the other part has only ∆0 . If x(k) ∈ ∆i , i ≥ 1, then
f x(k) ≥ 1, and if x(k) ∈ ∆0 , then |f x(k)| ≤ 1.
We claim that a trajectory of (5.2.9) starting from the ith sector ∆i
will reach ∆0 at step i and reach E at step i + 1. By symmetry, the same
argument applies to −∆i .
We first show that any state in ∆0 will reach E at the next step. For
x(k) ∈ ∆0 , there exist γ1 ∈ (0, 1), γ2 , γ3 ≥ 0, γ1 + γ2 + γ3 = 1, such that
x(k) = γ1 x+ + −
e + γ2 p0 + γ3 p1 = γ1 xe + γ2 xe + γ3 p1 .
We also have |f x(k)| ≤ 1, so sat(f x(k)) = f x(k) and
x(k + 1) = Ax(k) + b sat(f x(k)) = (A + bf )x(k).
Recalling that

−1 1 0
x+
e = , x− +
e = −xe , A + bf = ,
0 0 0
we have
(A + bf )x+ +
e = xe , (A + bf )x− −
e = xe .
In view of (5.2.11) and the fact that p1 = A−1 x−

e −A
−1
b, we have
(A + bf )p1 = Ap1 + b = x−
e .
It follows that
x(k + 1) = (A + bf )x(k) = γ1 x+ −
e + (γ2 + γ3 )xe ∈ E.
We next show that a trajectory starting from ∆i , i ≥ 1, will reach ∆0

at the ith step. Recall that before the trajectory reaches ∆0 , we have
f x(k) ≥ 1, so u(k) = sat(f x(k)) = 1, which means that
x(k + 1) = Ax(k) + b.
If x(k) ∈ ∆i , then there exist γ1 ∈ (0, 1), γ2 , γ3 ≥ 0, γ1 + γ2 + γ3 = 1,

such that
x(k) = γ1 x+
e + γ2 pi + γ3 pi+1 .
Hence,
x(k + 1) = Ax(k) + b
= A(γ1 x+
e + γ2 pi + γ3 pi+1 ) + (γ1 + γ2 + γ3 )b
= γ1 (Ax+
e + b) + γ2 (Api + b) + γ3 (Api+1 + b)
= γ1 x+
e + γ2 pi−1 + γ3 pi
∈ ∆i−1 ,
where the last equality follows from (5.2.10).

From the above relation, we see that x(k) and x(k + 1) have the same
coefficients γ1 , γ2 and γ3 . So there is a one to one correspondence between
∆i and ∆i−1 . This relation can be rewritten as
A∆i + b = ∆i−1 , i ≥ 1.
Therefore, if x(0) ∈ ∆i , then x(1) ∈ ∆i−1 , · · · , x(i) ∈ ∆0 . By the previous

argument, x(i + 1) ∈ E.
Finally, because
1 0
A + bf =
0 0
and E is in the linear region {x : |f x| ≤ 1}, we see that E is the set of
stable equilibrium points.
Case 2. eig(A) = r(cos(β) ± j sin(β)), β = π

N, r > 1, N > 2
Recall from (3.5.4) and Proposition 3.5.1 that

i−1
Ext(C) = ± A−i x− s + A−(i−−1) (−A−1 b)(1) : 1 ≤ i ≤ N
=0
is composed of the trajectories of (5.2.2) from ±x− +

s to ±xs under v(k) =
+ −
±1, or the trajectories of (5.2.1) from ±xs to ±xs under u(k) = ±1, where
rN + 1 −
x−
s = x .
rN − 1 e
Let p0 = x−
s and

i−1
pi = A−i p0 + A−(i−−1) (−A−1 b)(1), i ∈ [1, N ].
=0
Then,
Api + b = pi−1 , i ∈ [1, N ], (5.2.12)
and pN = x+
s . Since f A
−1 −
xe = 0 and f A−1 b = −1 (by (5.2.5)), we have

f p1 = f A−1 p0 − A−1 b = f A−1 x−
s − fA
−1
b = 1.
Since f x+e = 1 (by (5.2.4)), it follows that f x = 1 is a line passing through

x+
e and p 1.
p3
p2
p22
pi p21
p1 fx=1
1 p12
p 1
fx=-1
x+ ) pN pN
2 1
x -e ( po1) 11111
00000
s( pN11111111
00000000 o 00000p 2 x -s ( po )
11111
00000000
11111111
2 p N+1
1 x+e o
pN+1 pN+1
-p 1
-p
2
Figure 5.2.3: Illustration for Case 2.
The line between x+ −

s and xs cuts C into two symmetric parts. Denote
the part with pi , i ≥ 1, at its extremal points as C + . Then C + is composed
of N triangles, ∆i , i = 0, 1, · · · , N − 1, where

∆i = γ1 x+ e + γ2 pi + γ3 pi+1 : γ1 + γ2 + γ3 = 1, γ1 ∈ (0, 1), γ2 , γ3 ≥ 0 .
There are N −1 triangles above the line f x = 1, which are, ∆i , i ∈ [1, N −1].
In addition, there is a small triangle beside ∆N −1 , lying within −∆0 , which
is also above the line f x = 1 (the left gridded triangle in Fig. 5.2.3). We
need to deal with this triangle separately. The line f x = 1 intersects with
the boundary of C at two points − − one is p1 , and we label the other as
pN +1 . It can be shown that pN +1 lies on the line between x+ s (pN ) and −p1 .
So there exists a γ ∈ (0, 1) such that
pN +1 = γx+
s − (1 − γ)p1 , f pN +1 = 1. (5.2.13)
Since f x+
e = 1, we have
rN + 1
f x+
s = .
rN − 1
It follows that
rN − 1
γ= .
rN
From (5.2.13),
rN − 1 + 1
ApN +1 = Axs − N Ap1 . (5.2.14)
rN r
And from (5.2.12),
Ap1 = p0 − b = −x+
s − b. (5.2.15)
Noting that Ax+ +
e + b = xe , we have
rN + 1 + rN + 1 +
Ax+
s = Axe = N (x − b),
r −1
N r −1 e
and
rN + 1
Ax+ +
s = xs − b. (5.2.16)
rN − 1
Putting (5.2.15) and (5.2.16) into (5.2.14), we obtain
% &
rN − 1 + rN + 1 1
ApN +1 = x − b − N (−x+ +
s − b) = xs − b.
r N s
r −1
N r
This shows that
ApN +1 + b = x+
s = pN ,
which extends (5.2.12) to the case of i = N + 1. So we can treat this point

pN +1 along with pi , i = N, N − 1, · · · , 1, because they are all above or on
the line f x = 1. With pN +1 , we get an additional triangle,

∆N = γ1 x+ e + γ2 p N + γ3 p N +1 : γ1 + γ 2 + γ3 = 1, γ1 ∈ (0, 1), γ2 , γ3 ≥ 0 .
Because ∆i , i ∈ [1, N ], are above the line f x = 1 and the extremal points
pi , i ∈ [1, N + 1], satisfy (5.2.12), similar to Case 1, the relation between
∆i s can be simply put as follows,
A∆i + b = ∆i−1 , i ∈ [1, N ].
Hence any trajectory of (5.2.9) starting from C can be traced back to ∆N

or −∆N . By symmetry, we only need to study trajectories starting from
∆N .
Let p10 = x−
e and

i−1
p1i = A−i p10 + A−(i−−1) (−A−1 b)(1), i ∈ [0, N + 1].
=0
Then, similar to pi s, we have
Ap1i + b = p1i−1 , i ∈ [1, N + 1].

Hence,
Api + b = pi−1 , Ap1i + b = p1i−1 .
It follows from the equality Ax+ +
e + b = xe that
A(pi − x+ +
e ) = pi−1 − xe ,
A(p1i − x+ 1 +
e ) = pi−1 − xe ,
A(pi − p1i ) = pi−1 − p1i−1 .
By induction, we have
Ai (pi − x+ +
e ) = p0 − xe ,
Ai (p1i − x+ 1 +
e ) = p0 − xe ,
Ai (pi − p1i ) = p0 − p10 .
Recalling that
rN + 1 −
p10 = x−
e , p0 = x−
s = x ,
rN − 1 e
we can verify that
rN − 1
p10 − x+
e = p 0 − x+
e = 2x−
e , (5.2.17)
rN
and
1 2
p0 − p10 = p0 − x+
e = N x− . (5.2.18)
r N r −1 e
We thus have, by linearity,
rN − 1
p1i − x+
e = N
pi − x+
e
r
and
1
pi − p1i =
N
pi − x+ e .
r
From the above equalities, we see that pi , p1i and x+ e are on the same line.
Inductively, let pj+1
0 be a point between p j
0 and p 0 such that
1 1 +

p0 − pj+1
0 = N
p 0 − p j
0 = N (j+1) p0 − xe , (5.2.19)
r r
and

i−1

pj+1
i = A−i pj+1
0 + A−(i−−1) −A−1 b (1), i ∈ [0, N + 1]. (5.2.20)
=0
Then, again by linearity, we have,

1 1
pi − pj+1
i = N pi − pji = N (j+1) pi − x+
e (5.2.21)
r r
and that pi , p1i , · · · , pji , · · · and x+
e are on the same line (see Fig. 5.2.4 for
an illustration). Clearly,
lim pji = pi .
j→∞
From (5.2.20), we see that pj0 , pj1 , · · · , pjN +1 is a trajectory of (5.2.2) under
the constant control v(k) ≡ 1. Since pji , i ≥ 1, are all above the line f x = 1,
it follows that
sat f pji = 1.
This shows that the closed-loop system (5.2.9) has a reversed trajectory,
pjN +1 , pjN , · · · , pj0 . We will show that pj0 is closer to the origin than pjN .
Note that p10 = x− e . It follows from (5.2.17), (5.2.18) and (5.2.19) that
1 1
p0 − pj+1
0 = (p0 − x+ −
e ) = N j (p0 − xe ).
rN (j+1) r
Recalling that p0 = −pN , we have,
1
−pN − pj+1
0 = (−pN − x−
e ),
rN j
i.e.,
1
pN + pj+1
0 = (pN − x+
e ). (5.2.22)
rN j
From (5.2.21), we have,
1
pN − pjN = (pN − x+
e ).
rN j
Subtracting the above equation from (5.2.22), we get
pj+1
0 = −pjN . (5.2.23)
So a trajectory of the closed-loop system starting at pj+1N will reach pj+1

0 =
j
−pN at the N th step, which is closer to the origin. This is crucial to our
proof. Fig. 5.2.3 shows a trajectory starting at p2N . It reaches p20 = −p1N at
the N th step and reaches −p10 = x+ e at the 2N th step.
Denote the triangle pji+1 pji x+
e as ∆ji (see Fig. 5.2.4). Then by (5.2.20),
A∆ji+1 + b = ∆ji , i ∈ [0, N ], j ≥ 1.

p
p 3i+1 i+1
2
pi+1
p 1i+1
x+
e p1 p 2 p 3 pi
i i i
Figure 5.2.4: Illustration for the partition of ∆i .
So the triangle ∆jN is transfered to ∆j0 at the N th step. If x(0) ∈ ∆N , then

x(N ) ∈ ∆0 and if, in addition, |f x(N )| ≤ 1, then x(N + 1) will be on the
line between x+ −
e and xs by linearity (similar to Case 1). Because ∆0 (the
1
triangle xe p1 xe ) is in the linear region |f x| ≤ 1, all the states in ∆10 will

+ 1 −
be compressed to the line set E in the next step and will stay there.
We now need to know how the other trajectories travel. If x(0) ∈
∆N \ ∆1N (the quadrilateral p1N p2N p2N +1 p1N +1 ), then at the N th step, the
2
trajectory will reach ∆20 \ ∆10 (the quadrilateral p10 p20 p21 p11 ). Since p20 = −p1N ,
∆20 \ ∆10 is the union of a parallelogram and −∆1N (the right grided small
triangle in Fig. 5.2.3). The states in the parallelogram will reach the side
of −∆1N on the line between x− −
e and xs at the next step because the
parallelogram is in the linear region |f x| ≤ 1. No matter at the N th or the
(N + 1)th step, the set ∆2N \ ∆1N will be transfered to −∆1N . Similarly, the
set ∆jN \ ∆j−1
N will be transfered to −∆j−1 N \ (−∆j−2 N ) at the N th or the
(N + 1)th step (refer to (5.2.23)).
Now suppose x(0) ∈ ∆N , then there always exists a j ≥ 1 such that
x(0) ∈ ∆jN \ ∆j−1N . By the foregoing argument, the trajectory will reach
−∆j−1
N \ (−∆ j−2
N ) at the N th or the (N + 1)th step. By continuing this
5.3. Global Stabilization – Planar Systems 99
process, the trajectory will first reach ∆1N (or −∆1N ), then ∆10 (or −∆10 ) and
finally E and remain in E.
Since any state in C can be traced back to ∆N , we know that all the
trajectories starting from C will reach E in a finite number of steps. 2
5.3. Global Stabilization – Planar Systems
With the development in the previous section, we are ready to establish

global stabilizability on the null controllable region for second order anti-
stable linear systems subject to actuator saturation. In the previous section,
a saturated linear feedback
u(k) = sat(f x(k))
was constructed that drives all the states in the null controllable region C
to the set E. In what follows, we will construct a feedback law that first
drives all states in C to E in a finite number of steps, and then from there
to the origin, also in a finite number of steps.
With
x1 (k)
x(k) = ∈ E,
0
we might be tempted to look for a control that keeps the state in E, while
forcing it closer to the origin, i.e., to find a control |u(k)| ≤ 1 such that

x1 (k) x1 (k)
A + bu(k) = ρ,
0 0
with |ρ| < 1. Unfortunately, this is impossible because for each x(k) ∈ E,
there is only one control to keep it in E and this control can only keep the
state stationary. One solution is to use two controls u(k) and u(k + 1) in a
sequence. We use u(k) to drive the state away from E and use u(k + 1) to
drive it back to E but at a point closer to the origin.
A point in E can be expressed as

x1 (k)
x(k) = , x1 (k) ∈ [−1, 1].
0
By stacking the two controls together, we have

2
u(k)
x(k + 2) = A x(k) + Ab b . (5.3.1)
u(k + 1)
It is desired to find u(k) and u(k + 1) such that

x1 (k) u(k) x1 (k)
x(k + 2) = A2 + Ab b = ρ
0 u(k + 1) 0
with |ρ| < 1. A smaller ρ indicates that x(k + 2) is closer to the origin and
implies a faster convergence rate. Therefore, we would like to minimize |ρ|.
The problem can be formulated as:
inf |ρ| (5.3.2)

x1 (k) u(k) x1 (k)
s.t. A2 + Ab b = ρ,
0 u(k + 1) 0
|u(k)| ≤ 1, |u(k + 1)| ≤ 1.
This simple problem has a closed form solution.
Lemma 5.3.1. For |x1 (k)| ≤ 1, the optimization problem (5.3.2) has a
solution: with
1
ρ1 (x1 (k)) = − (det A − trA + 1) + trA − det A,
|x1 (k)|
1 det A − trA + 1 det A
ρ2 (x1 (k)) = − + ,
|x1 (k)| trA − 1 trA − 1
we have an optimal ρ given by

0, x1 (k) = 0,
∗
ρ (x1 (k)) = (5.3.3)
max{0, ρ1 (x1 (k)), ρ2 (x1 (k))}, otherwise.
And the optimal control signal is given by

u∗ (k) −1 ∗ 2
x1 (k)
= Ab b ρ (x1 (k))I − A .
u∗ (k + 1) 0
Furthermore, when |x1 (k)| = 1, we have ρ∗ (x1 (k)) = 1 and when |x1 (k)| <
1, we have ρ∗ (x1 (k)) < 1. Also, ρ∗ is an increasing function of |x1 (k)|.
Proof. It is easy to see that if x1 (k) = 0, then inf |ρ| = 0. We assume that
x1 (k) > 0. The case where x1 (k) < 0 can be dealt with in a similar way.
Recall from (5.2.3) that

1
b = (A − I) ,
0
which implies that

1 1 a11 1
Ab b = (A − I) A = (A − I) .
0 0 a21 0
So the equation for u in (5.3.2) becomes

−1
u(k) a11 1 −1 2 x1 (k)
= (A − I) (ρI − A ) . (5.3.4)
u(k + 1) a21 0 0
Direct computation shows that

u(k) 1
=
u(k + 1) det A − trA + 1
% &
−1 trA − det A
× ρ+ x1 (k). (5.3.5)
trA − 1 −det A
Because A is anti-stable, it can be shown that in both Case 1 and Case

2, we have det A − trA + 1 > 0 and trA − 1 > 0 (note that in Case 2,
N > 2). Since x1 (k) > 0, the constraint that |u(k)| ≤ 1 and |u(k + 1)| ≤ 1
is equivalent to
ρ 1 ≤ ρ ≤ ρ3 , ρ2 ≤ ρ ≤ ρ4 ,
where
1
ρ1 = − (det A − trA + 1) + trA − det A,
x1 (k)
1 (det A − trA + 1) det A
ρ2 = − + ,
x1 (k) trA − 1 trA − 1
1
ρ3 = (det A − trA + 1) + trA − det A,
x1 (k)
1 (det A − trA + 1) det A
ρ4 = + .
x1 (k) trA − 1 trA − 1
When x1 (k) = 1,
ρ2 = ρ3 = 1, ρ1 < 1, ρ4 > 1,
so there is only one feasible solution, namely that of ρ∗ = 1.

As x1 (k) < 1 decreases from 1, we have ρ1 ↓, ρ2 ↓, ρ3 ↑, ρ4 ↑, so there
will be feasible solutions and the optimal ρ is
ρ∗ = max(0, ρ1 , ρ2 ) < 1.
We also see that ρ∗ decreases as x1 (k) decreases. With this ρ∗ , we have an

optimal control of

u∗ (k) −1 ∗ 2 x1 (k)
= Ab b (ρ I − A ) .
u∗ (k + 1) 0
2
We next establish two lemmas that would directly lead to our main
results on global asymptotic stabilization on the null controllable region.
Lemma 5.3.2. Consider the closed-loop system

2
u(k)
x(k + 2) = A x(k) + Ab b , x(0) ∈ E, (5.3.6)
u(k + 1)
with

u(k) −1 ∗ x1 (k)
= Ab b ρ (x1 (k))I − A2 (5.3.7)
u(k + 1) 0
and ρ∗ (x1 (k)) given by (5.3.3). Then the solution of (5.3.6) lies in E for
even k, the sequence {|x1 (2)|, |x1 (4)|, · · ·} is monotonically decreasing and
there exists a k < ∞ such that x(k) = 0.
Proof. If x1 (0) = 0, then u = 0, so x1 (k) = 0 for all k ≥ 0. Now we assume

that x1 (0) > 0. From Lemma 5.3.1, we know that x(2), x(4), · · · ∈ E and
x1 (k + 2) = ρ∗ (x1 (k))x1 (k).
Since 0 < x1 (0) < 1, we have ρ∗ (x1 (0)) < 1 and x1 (2) < x1 (0). Because ρ∗
is an increasing function of |x1 (k)|, we have that
x1 (k + 2) ≤ ρ∗ (x1 (0))x1 (k), k ≥ 0, k even, (5.3.8)
hence x1 (2), x1 (4), · · · is an exponentially decreasing sequence.

From the definition of ρ1 and ρ2 in Lemma 5.3.1, since
det (A) − tr(A) + 1 > 0,
we see that there exists some xm > 0 such that ρ1 , ρ2 < 0 for all |x1 (k)| ≤
xm . From (5.3.8), there exists a k < ∞ such that |x1 (k)| ≤ xm , so
ρ∗ (x1 (k)) = 0 and we get x1 (k + 2) = 0 and x(k + 2) = 0. 2
Now we know that under the saturated linear feedback (5.2.8), all the
states in C will be steered to the line segment E (or E in Case 2) in a finite
number of steps, and under the two step controller (5.3.7), any state in E
will be driven to the origin in finite number of steps. What remains is to
combine these two controllers in a simple way.
The question is, given a state x(k) ∈ C, which controller should we use?
If x(k) ∈ E, we have to use the two step controller, which generates two
controls u(k) and u(k + 1) at a time. After u(k) is applied, the state reaches
x(k + 1) which may not be in E. This same x(k + 1), however, might arise
from using the saturated linear feedback (5.2.8). One approach would be
to add some memory in the control law which keeps track of where the
previous state was, namely in E or not in E. There is a simpler approach,
as is shown below.
Lemma 5.3.3. Let x(k) ∈ E and u(k), u(k + 1) be the two controls gen-
erated by the controller (5.3.7). Then,
a) u(k + 1) = f x(k + 1).
b) x(k + 1) = Ax(k) + bu(k) lies between the lines f x = ±1, i.e.,
|f x(k + 1)| ≤ 1.
Proof. After applying the two controls u(k) and u(k + 1), we have

x1 (k + 2)
x(k + 2) = Ax(k + 1) + bu(k + 1) = , x1 (k + 2) ∈ (−1, 1).
0
It follows that

−1 x1 (k + 2)
x(k + 1) = A − A−1 bu(k + 1).
0
From (5.2.5), we have

1
f A−1 = −f A−1 x+
e = 0, f A−1 b = −1.
0
Therefore,
f x(k + 1) = u(k + 1).
This proves a). Item b) follows from the above equation and the fact that
|u(k + 1)| ≤ 1. 2
This result says that the second control u(k + 1) coincides with the
saturated linear feedback control u = sat(f x). So we don’t need to know
whether a state is a result of the controller (5.2.8) or of the controller (5.3.7).

For a state not in E, we can simply apply the saturated linear feedback
u = sat(f x), and if x(k) ∈ E, then we just apply u(k) in (5.3.7). In the
case where A has complex eigenvalues, we need to pay special attention to
the two end points of E, i.e., x+ −
e and xe . They are in the interior of C and
are unstable equilibria under both (5.2.8) and (5.3.7). If x(k) = x+ e , then
u(k) = 1 under both (5.2.8) and (5.3.7), hence the state will stay at x+ e .
We can make the state leave these two points by applying u = ±(1 − η)
with a small number η > 0. Overall, we have a combined controller given
by 

 sat(f x), x∈ / {E ∪ ±x+ e },

u = h(x) = g(x), x ∈ E, (5.3.9)


 +
±(1 − η), x = ±xe ,
where

−1 ∗ 2x1
f = [−1 2], g(x) = 1 0 Ab b (ρ (x1 )I − A ) ,
0
and

∗
0, x1 = 0,
ρ (x1 ) =
max {0, ρ1 (x1 ), ρ2 (x1 )} , otherwise,
1
ρ1 (x1 ) = − (det A − trA + 1) + trA − det A,
|x1 |
1 det A − trA + 1 det A
ρ2 (x1 ) = − + .
|x1 | trA − 1 trA − 1
Note that g(x) is the first control u(k) in (5.3.7). Under the control
u = h(x), the closed-loop system is
x(k + 1) = Ax(k) + bh(x(k)). (5.3.10)
Combining Theorem 5.2.1, Lemma 5.3.2 and Lemma 5.3.3, we immedi-

ately arrive at the main result of this section.
Theorem 5.3.1. Consider the system (5.3.10). For every x(0) ∈ C, there
exists a k < ∞ such that x(k) = 0.
Example 5.3.1. We illustrate the results of this section with an example

where β/π is irrational. Consider the system (5.3.10) with

0.6570 0.6861 −0.3430
A= , b= ,
−0.9164 1.8328 −0.9164
and
√
2
eig(A) = r(cos β ± sin β), r = 1.3538, β = π.
11
Two trajectories of the system are plotted in Fig. 5.3.1. The initial points
are marked with “∗”. They are very close to the boundary of C. The
extremal points of C are marked with “◦”. We see that the two trajectories
reach the origin in finite number of steps.
1.5
0.5
−0.5
−1
−1.5
−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Figure 5.3.1: Trajectories of the closed-loop system.
5.4. Semi-Global Stabilization – Planar Systems
In this section, we will establish semi-global stabilizability on the null con-

trollable region for second order anti-stable linear systems subject to actu-
ator saturation of the form (5.2.1). For such systems, it was established in
the previous section that global stabilization can be achieved by controllers
of the form (5.3.9). Our motivation for considering semi-global stabiliza-
tion here is simpler controllers which are also less sensitive to system model
uncertainties and disturbances.
Recall from Lemma 5.2.1 that

1 0
A + bf = .
0 0
Since (A + bf, b) is controllable, b1 and b2 are both nonzero. If we choose

f¯ = f − δ 0
such that 0 < b1 δ < 1, then

1 − b1 δ 0
A + bf¯ =
−b2 δ 0
is asymptotically stable. Now consider the closed-loop system

x(k + 1) = Ax(k) + b sat f¯x(k) . (5.4.1)
We need to determine an invariant set in the linear region |f¯x| ≤ 1 for this
system. Let
1 0 b1 δ(2 − b1 δ)
P = , p2 = .
0 p2 2(b2 δ)2
Then,
(A + bf¯)T P (A + bf¯) − P < 0.
Define a level set

E P, c2 := x : xT P x ≤ c2
and
c0 = max c > 0 : |f¯x| ≤ 1, ∀ x ∈ E(P, c2 ) .
Then, E(P, c20 ) is the largest level set that lies entirely in the linear region.
We see that both P and E(P, c20 ) depend on δ. When δ is very small, p2
will be very big and E(P, c20 ) will be a flat ellipsoid that encloses almost
all points in E (Recall that E is a horizontal line between f x = 1 and
f x = −1). As δ → 0, p2 → ∞ and E(P, c20 ) → E (see Fig. 5.4.1, where the
inner ellipsoid is E(P, c20 )).
Now consider the switching control law,

sat(f x), x ∈ / E(P, c20 ),
u(x) = (5.4.2)
sat(f¯x), x ∈ E(P, c20 ).
Note that E(P, c20 ) is a contractive invariant set: once the trajectory enters
E(P, c20 ), it will remain in E(P, c20 ) and then converge to the origin under
the control law u = sat(f¯x) = f¯x. Because any closed interval in E can be
enclosed in E(P, c20 ) by decreasing δ, the domain of attraction of the origin
of the closed-loop system can be made arbitrarily close to C for the case
that A has two real eigenvalues.
Example 5.4.1. The open-loop system has

0.8911 0.2178 −0.1089
A= , b= .
−0.9111 1.8221 −0.9111
We choose δ = 0.01, so,

f¯ = −0.99 2
and c0 = 0.9747. The simulation result is shown in Fig. 5.4.1, where “∗”
is the initial state of the trajectory. The trajectory reaches the set E(P, c20 )
in 5 steps, then remains in E(P, c20 ) and converges to the origin.
1.5
0.5
−0.5
−1
−1.5
−2
−1.5 −1 −0.5 0 0.5 1 1.5
Figure 5.4.1: Illustration for Example 5.4.1.

5.5. Semi-Global Stabilization – Higher Order Systems
Consider the following open loop system,

A1 0 b1
x(k + 1) = Ax(k) + bu(k) = x(k) + u(k), (5.5.1)
0 A2 b2
where x = (xa , xs ), xa ∈ R2 , xs ∈ Rn , A1 ∈ R2×2 is anti-stable and

A2 ∈ Rn is semi-stable. Assume that (A, b) is stabilizable. Denote the null
controllable region of the subsystem
xa (k + 1) = A1 xa (k) + b1 u(k)
as C 1 . Then the asymptotically null controllable region of (5.5.1) is C 1 ×Rn .

Given γ1 , γ2 > 0, let

Ω1 (γ1 ) := γ1 xa ∈ R2 : xa ∈ C 1 , (5.5.2)
and
Ω2 (γ2 ) := xs ∈ Rn : |xs | ≤ γ2 . (5.5.3)
When γ1 = 1, Ω1 (γ1 ) = C 1 and when γ1 < 1, Ω1 (γ1 ) lies in the interior of

C 1 . In this section, we will show that given any γ1 < 1 and γ2 > 0, a state
feedback can be designed such that Ω1 (γ1 ) × Ω2 (γ2 ) is contained in the
domain of attraction of the closed-loop system. The basic idea in Chapter
4 for continuous-time systems can be easily adapted for our purpose except
for some technical consideration on the discrete algebraic Riccati equation
(DARE).
For an ε > 0, let

P1 (ε) P2 (ε)
P (ε) = ∈ R(2+n)×(2+n)
P2T (ε) P3 (ε)
be the unique positive definite solution to the DARE
P = AT P A − AT P b(bT P b + 1)−1 bT P A + εI. (5.5.4)
As ε ↓ 0, P (ε) decreases.
Let P 1 be the unique positive definite solution to the DARE
P 1 = AT1 P 1 A1 − AT1 P 1 b1 (bT1 P 1 b1 + 1)−1 bT1 P 1 A1 .

Then by the continuity property of the solution of the DARE (see [86,
Theorem 4.2.7], we have

P1 0
lim P (ε) = .
ε→0 0 0
Let
f1 (ε) := −(bT P (ε)b + 1)−1 bT P (ε)A.
First, consider the domain of attraction of the origin of the following closed-
loop system
x(k + 1) = Ax(k) + b sat(f1 (ε)x(k)). (5.5.5)
It can be shown by a similar method of [65, Lemma 2.3.6] that there exist
constants c > 0 and ε∗ > 0 such that

1 1
f1 (ε)P − 2 (ε) = (bT P (ε)b + 1)−1 bT P (ε)AP − 2 (ε) < c, ∀ ε ∈ (0, ε∗ ].
It follows that

1
D(ε) := x ∈ R2+n : xT P (ε)x ≤
c2
is contained in the domain of attraction of (5.5.5) and is an invariant set for

all ε ∈ (0, ε∗ ]. Note that if x0 ∈ D(ε), then x(k) ∈ D(ε) and |f1 (ε)x(k)| ≤ 1
for all k > 0. That is, x(k) will stay in the linear region of the closed-loop
system, and in D(ε).
Lemma 5.5.1. Denote

1
r1 (ε) := 1 ,
2
2 P1 (ε) c
'
−|P2 (ε)| + |P2 (ε)|2 + 3|P1 (ε)| |P3 (ε)|
r2 (ε) := r1 (ε).
|P3 (ε)|
Then,

D1 (ε) := x ∈ R2+n : |xa | ≤ r1 (ε), |xs | ≤ r2 (ε) ⊂ D(ε).
Moreover,
lim r2 (ε) = ∞,
ε→0

Proof. It can be verified that

1
|P1 (ε)|r12 (ε) + 2|P2 (ε)|r1 (ε)r2 (ε) + |P3 (ε)|r22 (ε) = . (5.5.6)
c2
Hence
1
xT P (ε)x ≤ , ∀ x ∈ D1 (ε),
c2
i.e., D1 (ε) ⊂ D(ε). By the definition of r1 (ε) and r2 (ε), we have
3|P1 (ε)| 1
r2 (ε) = ' · 1 .
|P2 (ε)| + |P2 (ε)| + 3|P1 (ε)| |P3 (ε)| 2 P 2 (ε) c
2
1
Since as ε goes to zero, P2 (ε), P3 (ε) → 0 and P1 (ε) → P 1 , r1 (ε) is bounded

whereas r2 (ε) → ∞. It follows from the monotonicity of P (ε) that r1 is a
monotonically decreasing function of ε. 2
Theorem 5.5.1. Let h(·) be defined in (5.3.9). For any γ1 < 1 and γ2 > 0,
there exists an ε > 0 such that Ω1 (γ1 )×Ω2 (γ2 ) is in the domain of attraction
of the origin for the following closed-loop system

h(xa ), x∈
/ D(ε),
x(k + 1) = Ax(k) + bu(k), u(k) = (5.5.7)
sat(f1 (ε)x), x ∈ D(ε).
Proof. The proof is analogous to that of Theorem 4.4.1. 2
Remark 5.5.1.
1) For the case that the open loop system has a single exponentially
unstable pole, there exists a simpler nonlinear function h(·) such that
Theorem 5.5.1 is true. Let x, A and b be partitioned similarly to the
two exponentially unstable pole case, with xa , A1 , b1 ∈ R. Then, it is
easy to verify that
% &
|b1 | |b1 |
C1 = − , .
|A1 | − 1 |A1 | − 1
Also, the origin of the system
% &
A1
xa (k + 1) = A1 xa (k) − b1 sat xa (k)
b1
has a domain of attraction equal to C 1 . Hence, if we let
% &
A1
h(xa ) = −sat xa ,
b1
then Theorem 5.5.1 is also true for this case.
5.6. Conclusions 111
2) For the case that the open loop system has no exponentially unstable
poles (ANCBC), we have C = Rn , and
lim P (ε) = 0.
ε→∞
Therefore, the set D(ε), which is in the domain of attraction for the
system under the control u = sat(f1 (ε)x), can be made arbitrarily
large to include any compact subset of Rn . This means that the
control u = sat(f1 (ε)x) can achieve semi-global stabilization and no
switching is necessary.
5.6. Conclusions
In this chapter, we established global and semi-global stabilizability for
exponentially unstable discrete-time linear systems with saturating actu-
ators. For a planar anti-stable system, controllers are constructed that
achieve global and semi-global stabilization, respectively. For higher order
systems with one or two anti-stable modes, semi-global stabilizability was
established also by explicit construction of feedback laws. The semi-global
stabilizability for general systems will be established in Chapter 9.
Chapter 6
Practical Stabilization on
Null Controllable Region
6.1. Introduction
In the previous chapters, we have characterized the asymptotically null

controllable region of a linear systems subject to actuator saturation. We
have also constructed stabilizing feedback laws that would result in a do-
main of attraction that is either the entire asymptotically null controllable
region or a large portion of it. In this chapter, we start to address closed-
loop performances beyond large domains of attraction. In particular, we
will design feedback laws that not only achieve semi-global stabilization on
asymptotically null controllable region but also have the ability to reject
input-additive bounded disturbances to an arbitrary level of accuracy. We
refer to such a design problem as semi-global practical stabilization on the
null controllable region.
More precisely, for any set X 0 in the interior of the asymptotically null
controllable region, any (arbitrarily small) set X ∞ containing the origin in
its interior, and any (arbitrarily large) bound on the disturbance, we would
like to design a feedback law such that any trajectory of the closed-loop
system enters and remains in the set X ∞ in a finite time as long as it starts
from the set X 0 .
We will again restrict our attention to systems with no more than two
exponentially unstable poles. A solution of the problem for systems with no
exponentially unstable poles can be found in [65,85]. General linear systems
113
114 Chapter 6. Practical Stabilization
subject to actuator saturation and non-input-additive disturbances will be

investigated in Chapter 10.
Section 6.2 formulates the problem and summarizes the main results.
Section 6.3 contains the proof of the main results. For clarity in the presen-
tation, we have delegated the proof of some technical lemmas to Appendices
6.A and 6.B. Section 6.4 uses an aircraft model to demonstrate the results
presented in this chapter. Section 6.5 contains a brief conclusion to the
chapter.
6.2. Problem Statement and Main Results

6.2.1. Problem Statement
Consider an open-loop system subject to both actuator saturation and dis-

turbance,
ẋ = Ax + b sat(u + w), (6.2.1)
where x ∈ Rn is the state, u ∈ R is the control input and w ∈ R is the

disturbance. Assume that (A, b) is stabilizable. We consider the following
set of disturbances:

W := w : [0, ∞) → R, w is measurable and w∞ ≤ D ,
where D is a known constant.

Let C and C a be the null controllable region and the asymptotically null
controllable region, respectively, of the system (6.2.1) in the absence of the
disturbance w. Recall that, if (A, b) is controllable, then C a = C.
Our objective in this chapter is to achieve semi-global practical stabi-
lization on the null controllable region for the system (6.2.1), i.e., to design
a family of feedback laws such that, given any (arbitrarily large) set X 0 in
the interior of C a and any (arbitrarily small) set X ∞ containing the origin
in its interior, there is a feedback law from this family such that any tra-
jectory of the closed-loop system that starts from within X 0 will enter X ∞
in a finite time and remain there.
6.2.2. Main Results: Semi-Global Practical Stabilization
Given any (arbitrarily small) set that contains the origin in its interior, we
will show that its domain of attraction can be made to include any compact
subset of C a in the presence of disturbances bounded by an (arbitrarily
6.3. Proof of Main Results 115
large) given number. More specifically, we will establish the following result
on semi-global practical stabilization on the null controllable region for the
system (6.2.1).
Theorem 6.2.1. Consider the system (6.2.1) with A having two exponen-
tially unstable eigenvalues. Given any set X 0 ⊂ int(C a ), any set X ∞ such
that 0 ∈ int(X ∞ ), and any positive number D, there is a feedback law
u = F (x) such that any trajectory of the closed-loop system enters the set
X ∞ in a finite time an remains there as long as it starts from within the
set X 0 .
6.3. Proof of Main Results

We outline the proof of the main result, Theorem 6.2.1, as follows. In Sec-
tion 6.3.1, we will establish some properties of second order linear systems,
in the absence of actuator saturation. In Section 6.3.2, we use the results in
Section 6.3.1 to establish the continuity and monotonicity of the domain of
attraction for a class of second order systems. The main part of the proof
is given in Sections 6.3.3 and 6.3.4. Section 6.3.3 deals with second order
anti-stable systems, while Section 6.3.4 deals with higher order systems
with one or two exponentially unstable poles.
6.3.1. Properties of the Trajectories of Second Order

Linear Systems
We first consider the second order anti-stable system

0 −a1
ẋ = Ax = x, a1 , a2 > 0. (6.3.1)
1 a2
We will examine its trajectories with respect to a horizontal line kf x = 1
where

f = 0 1 , k > 0.
On this line, x2 = k1 . If x1 > − ak2 , then ẋ2 > 0, i.e., the vector ẋ points
upward. If x1 < − ak2 , then ẋ2 < 0, i.e., the vector ẋ points downward.
Above the line, ẋ1 < 0, hence the trajectories all go leftward. Denote
λ
− k1 , if A has real eigenvalues λ1 and λ2 , λ1 ≥ λ2 > 0,
am =
∞, if A has a pair of complex eigenvalues.
Then, we have
Lemma 6.3.1. Let x11 ≥ − ak2 and

x11
p= 1
k
be a point on the line kf x = 1. The trajectory x(t) = eAt p, t ≥ 0, will

return to this line if and only if x11 < am . Let T be the first time when it
returns and
y11
p = 1
k
be the corresponding intersection, i.e., p = eAT p. This defines two func-

tions: x11 → y11 and x11 → T . Then for all x11 ∈ (− ak2 , am ),
dy11 d2 y11 dT
< −1, < 0, > 0. (6.3.2)
dx11 dx211 dx11
Proof. See Appendix 6.A. 2
2.5
1.5
1 , , ,
p3 p2 p1 p1 p2 p3
0.5
0
−14 −12 −10 −8 −6 −4 −2 0 2 4 6
It may be easier to interpret Lemma 6.3.1 by writing (6.3.2) as

d(−y11 ) d2 (−y11 )
> 1, > 0.
dx11 dx211
An illustration of Lemma 6.3.1 is given in Fig. 6.3.1, where p1 , p2 and p3

are three points on kf x = 1, i.e.,

xi11 a
2
pi = 1 , x i
11 ∈ − , a m , i = 1, 2, 3,
k
k
and p1 , p2 and p3 are the first intersections of the trajectories that start
from p1 , p2 and p3 . Then,
|p3 − p2 | |p − p1 |

> 2 > 1. (6.3.3)
|p3 − p2 | |p2 − p1 |
It follows that
|p2 −p1 |
|p2 − p1 | |p2 − p1 | 1+ |p3 −p2 |
< =⇒ < 1.
|p3 − p2 | |p3 − p2 | 1+ |p2 −p1 |
|p3 −p2 |
Hence,
|p2 −p1 |
|p3 − p1 | |p − p2 | + |p2 − p1 | |p − p2 | 1 + |p3 −p2 | |p3 − p2 |
= 3 = 3 < .
|p3 − p1 | |p3 − p2 | + |p2 − p1 | |p3 − p2 | 1 + |p2 −p1 | |p3 − p2 |
|p3 −p2 |
(6.3.4)
Also from (6.3.3), we have
|p3 −p2 |
|p3 − p2 | |p3 − p2 | 1+ |p2 −p1 |
> =⇒ > 1.
|p2 − p1 | |p2 − p1 | 1+ |p3 −p2 |
|p2 −p1 |
Hence,
|p3 −p2 |
|p3 − p1 | |p − p2 | + |p2 − p1 | |p − p1 | 1 + |p2 −p1 | |p2 − p1 |
= 3 = 2 > .
|p3 − p1 | |p3 − p2 | + |p2 − p1 | |p2 − p1 | 1 + |p3 −p2 | |p2 − p1 |
|p2 −p1 |
(6.3.5)
Combining (6.3.4) and (6.3.5), we obtain
|p3 − p2 | |p − p1 | |p − p1 |

> 3 > 2 > 1. (6.3.6)
|p3 − p2 | |p3 − p1 | |p2 − p1 |
We next consider a second order stable linear system,

0 −a1
ẋ = Ax = x, a1 , a2 > 0. (6.3.7)
1 −a2
We will study the trajectories of (6.3.7) with respect to two horizontal lines
kf x = 1 and kf x = −1 where

f= 0 1
and k > 0. For some points on the line kf x = 1, the trajectories of (6.3.7)
starting from these points will enter the region

x ∈ R2 : |kf x| < 1
and then intersect the line kf x = −1. We will define some functions to
relate these points on kf x = 1 to the first intersections of the trajectories
with kf x = −1. The functions will be defined in terms of the point

− ak2
p0 =
− k1
on kf x = −1. It is easy to see that, if a point x is on kf x = −1 and is to
the left of p0 , then the vector ẋ points downward; if x is to the right of p0 ,
then the vector ẋ points upward (see Fig. 6.3.2).
Definition 6.3.1. Let p0 be the unique point on kf x = 1 and Td > 0 be

the unique number satisfying

eATd p0 = p0 , kf eAt p0 ≤ 1, ∀ t ∈ [0, Td ].
Denote the first coordinate of p0 as xm , i.e.,

xm
p0 = 1 .
k
For x11 ∈ (−∞, xm ], let

x11
p = 1 ,
k
be a point on kf x = 1, then there is a unique

y11
p=
− k1
on kf x = −1, where y11 ∈ (−∞, − ak2 ] and a unique T ∈ (0, Td ] such that

p = eAT p , kf eAt p ≤ 1, ∀ t ∈ [0, T ]. (6.3.8)
This defines two functions x11 → y11 , and x11 → T with x11 ∈ (−∞, xm ],
y11 ∈ (−∞, − ak2 ] and T ∈ (0, Td ].
Lemma 6.3.2. For all x11 ∈ (−∞, xm ), we have x11 < y11 and
dy11 d2 y11 dT
> 1, > 0, > 0.
dx11 dx211 dx11
1.5
, , , ,
p p p p
3 2 1 0
1
0.5
−0.5
p p p p
3 2 1 0
−1
−1.5
−2
−2.5
−12 −10 −8 −6 −4 −2 0 2 4 6
Proof. See Appendix 6.B. 2

This lemma is illustrated with Fig. 6.3.2, where and are three p1 , p2 p3
points on kf x = 1, and p1 , p2 and p3 are the three first intersections of
kf x = −1 with the three trajectories starting from p1 , p2 and p3 , respec-
tively. Then,
|p1 − p2 | |p1 − p3 | |p2 − p3 |
> > > 1.
|p1 − p2 | |p1 − p3 | |p2 − p3 |
6.3.2. Properties of the Domain of Attraction
ẋ = Ax + b sat(kf x), x ∈ R2 , (6.3.9)
where
0 −a1 −b1
A= , b= ,
1 a2 −b2
a1 , a2 , b2 > 0, b1 ≥ 0, and

f= 0 1 .
If k > a2 /b2 , then A + kbf is Hurwitz and the origin is the unique equi-
librium point of (6.3.9) and it is stable. Denote the domain of attraction
of the origin as S(k), then by Theorem 4.2.1, ∂S(k) is the unique limit
cycle of (6.3.9). We will further show that the domain of attraction S(k)
increases as k increases.
Consider k0 > a2 /b2 . Denote the increment of k as δk . Remark 4.2.1
says that ∂S(k0 ) is symmetric with respect to the origin and has two inter-
sections with each of the lines k0 f x = 1 and k0 f x = −1. In Fig. 6.3.3, the
closed curve is ∂S(k0 ) and p1 , p2 , p3 (= −p1 ) and p4 (= −p2 ) are the four
intersections.
In the sequel, for a vector p ∈ R2 , we use (p)i to denote the ith coordi-
nate of p.
Since at p2 , the trajectory goes downward, i.e., ẋ2 < 0, so,
a2 − k0 b2
(p2 )1 < < 0.
k0
From Lemma 6.3.2, we have (p1 )1 < (p2 )1 < 0 (x11 < y11 in the lemma).
Hence both p1 and p2 are on the left half plane. Define
(p2 )1 a2
∆(k0 ) = − k0 + − k0 .
b2 b2
Then ∆(k0 ) > 0 due to the fact that the trajectory goes downward at p2 .
Proposition 6.3.1. Suppose k0 > a2 /b2 . Then, for all δk ∈ (0, ∆(k0 )),
S(k0 ) ⊂ S(k0 + δk ).
Proof. Since δk > 0, the two lines (k0 + δk )f x = ±1 are between the lines
k0 f x = 1 and k0 f x = −1. It follows that the vector field above k0 f x = 1
and that below k0 f x = −1 are the same for
ẋ = Ax + b sat(k0 f x) (6.3.10)
and
ẋ = Ax + b sat((k0 + δk )f x). (6.3.11)
Hence, if a trajectory of (6.3.11) starts at p4 (or p2 ), it will go along ∂S(k0 )
to p1 (or p3 ).
1.5
p p
1 1 4
s s
0.5 1 4
0
s
0
−0.5 ,, ,
s s s s
2 2 2 3
−1 p p
2 3
−1.5
−2
−8 −6 −4 −2 0 2 4 6 8
Figure 6.3.3: Illustration for the proof of Proposition 6.3.1.
Claim: If a trajectory of (6.3.11) starts at a point on ∂S(k0 ) between p1

and p2 and intersects the line k0 f x = −1, then the intersection must be
inside S(k0 ).
It follows from the claim that any trajectory of (6.3.11) that starts from
∂S(k0 ) will stay inside of S(k0 ) when it returns to the lines k0 f x = ±1. So
it is bounded and hence belongs to S(k0 + δk ). Note that any trajectory
outside of S(k0 + δk ) will diverge because the system has a unique unstable
limit cycle as the boundary of the domain of attraction. Since the two sets
are convex and open, we will have
S(k0 ) ⊂ S(k0 + δk ).
It remains to prove the claim.
Since S(k0 ) is convex, (A + bk0 f )x from p1 to p2 along ∂S(k0 ) is in-

creasing. Let s0 be the intersection of ∂S(k0 ) with the abscissa. Then
at s0 , (A + bk0 f )x = − π2 ; from p1 to s0 , (A + bk0 f )x ∈ (−π, − π2 ); and
from s0 to p2 , (A + bk0 f )x ∈ (− π2 , 0). Now consider a point x along S(k0 )
between p1 and p2 ,
1) If x is between p1 and s0 , then,
k0 f x ≤ sat((k0 + δk )f x).
If (A + bk0 f )x < b, then, ẋ of (6.3.11) directs inward of ∂S(k0 ) and

if (A + bk0 f )x) > b, then ẋ of (6.3.11) directs outward of ∂S(k0 ).
Since (A + bk0 f )x is increasing, the vector ẋ may direct outward of
∂S(k0 ) for the whole segment or for a lower part of the segment.
2) If x is between s0 and p2 , then
k0 f x ≥ sat((k0 + δk )f x).
Since b ∈ (−π, − π2 ), we have
(A + bk0 f )x) ≤ (Ax + b sat((k0 + δk )f x)),
i.e., the vector ẋ of (6.3.11) directs inward of ∂S(k0 ).
Let
x11
s1 = , h>0
h
be a point on ∂S(k0 ) between p1 and s0 such that ẋ of (6.3.11) at s1 directs
outward of ∂S(k0 ). Let

y11
s2 =
−h
be the intersection of ∂S(k0 ) with x2 = −h. Then by 1) the trajectory of
(6.3.11) starting at s1 will remain outside of ∂S(k0 ) above the abscissa. We
will show that when the trajectory reaches the line x2 = −h at s2 , it must
be inside ∂S(k0 ).
Let
0 0
s3 = , s4 =
−h h
(see Fig. 6.3.3). Denote the region enclosed by s1 s2 s3 s4 s1 as G0 , where the
part s1 s2 is on ∂S(k0 ) and the other parts are straight lines. Since this
region lies between k0 f x = ±1, the vector field of (6.3.10) on this region is,
ẋ1 = −(a1 + k0 b1 )x2 =: f1 (x)

ẋ2 = x1 + (a2 − k0 b2 )x2 =: f2 (x).
Applying Green’s Theorem to the system (6.3.10) on G0 , we get

+ % &
∂f1 ∂f2
f2 dx1 − f1 dx2 = − + dx1 dx2 . (6.3.12)
∂G0 G0 ∂x1 ∂x2
Note that the left-hand side integral from s1 to s2 and that from s3 to s4
are zero. Denote the area of G0 as Q0 , then, from (6.3.12), we have,
1 2 1 2
x11 +(a2 −k0 b2 )hx11 − y11 +(a2 −k0 b2 )hy11 = −(a2 −k0 b2 )Q0 . (6.3.13)
2 2
Clearly,
Q0 > −h(x11 + y11 ) (6.3.14)
by the convexity of S(k0 ) and the region G0 .
On the other hand, we consider a trajectory of (6.3.11) starting at s1
but crossing the line x2 = −h at

y11 + δy11
s2 = .
−h
First, we assume that s1 lies between (k0 + δk )f x = ±1. Apply Green’s

Theorem to (6.3.11) on the region enclosed by s1 s2 s3 s4 s1 , where the part
s1 s2 is on a trajectory of (6.3.11). Denote the area of the region as Q0 + δQ .
Similarly,
1 2 1
x11 + (a0 − k0 b2 − δk b2 )hx11 − (y11 + δy11 )2
2 2
+ (a2 − k0 b2 − δk b2 )h(y11 + δy11 )
= −(a2 − k0 b2 − δk b2 )(Q0 + δQ ). (6.3.15)
Subtracting (6.3.13) from (6.3.15), we obtain
−[y11 − (a2 − k0 b2 − δk b2 )h]δy11 = (k0 b2 − a2 )δQ + δk b2 (Q0 + hx11 + hy11 )

1
+ δy211 + δk b2 δQ . (6.3.16)
2
Q0 + hx11 + hy11 > 0.
Also recall that k0 b2 − a2 > 0.

From the definition of ∆(k0 ), we have
1
(p2 )1 − (a2 − k0 b2 − δk b2 ) < 0,
k0
1
for all δk ∈ [0, ∆(k0 )). Since y11 < (p2 )1 , h < k0 and −(a2 −k0 b2 −δk b2 ) > 0,
it follows that
y11 − (a2 − k0 b2 − δk b2 )h < 0, ∀ δk ∈ [0, ∆(k0 )).
Suppose that δk ∈ [0, ∆(k0 )). If δy11 < 0, then s2 is outside of ∂S(k0 )
and we must have δQ > 0. In this case, the left hand side of (6.3.16) is
negative and the righthand side is positive. This is a contradiction. Hence
δy11 must be positive and s2 must be inside ∂S(k0 ). By 2), the vector ẋ of
the system (6.3.11) directs inward of ∂S(k0 ) from s2 to p2 , therefore, when
the trajectory reaches k0 f x = −1, it must be to the right of p2 , i.e., still
inside ∂S(k0 ).
Now suppose s1 lies between (k0 + δk )f x = 1 and k0 f x = 1. Then, by
applying Green’s Theorem, we get exactly the same equation as (6.3.16),
although we need to partition the region enclosed by s1 s2 s3 s4 s1 into three
parts. And similar arguments apply. Thus we conclude that for all δk ∈
[0, ∆(k0 )), S(k0 ) ⊂ S(k0 + δk ). 2
Proposition 6.3.2. Consider
ẋ = Ax + b sat(f x), x ∈ R2 , (6.3.17)
where A ∈ R2×2 and b ∈ R2×1 are constant matrices, A is anti-stable and

f ∈ R1×2 is a variable. Denote the domain of attraction of the origin for
(6.3.17) as S(f ). Then, at any f such that A + bf is Hurwitz and has
distinct eigenvalues, S(f ) is continuous.
Proof. We only need to show that ∂S(f ) is continuous. Recall from

Theorem 4.2.1 that ∂S(f ) is a closed trajectory and has four intersections
with f x = ±1. Since the vector ẋ = Ax + b sat(f x) is continuous in f at
each x, it suffices to show that one of the intersections is continuous in f .
Actually we can show that the intersections are also differentiable in f . For
simplicity and for direct use of Lemmas 6.3.1 and 6.3.2, we apply a state
space transformation, x̂ = V (f )x, to the system (6.3.17), such that,

−1 0 −a1
V (f )AV (f ) = =: Â, (6.3.18)
1 a2

b1 (f )
V (f )b = =: b̂(f ), (6.3.19)
b2 (f )
and

f V −1 (f ) = 0 1 =: fˆ. (6.3.20)
Such a transformation always exists. To see this, assume that A is already
in this form. Since A is anti-stable and A + bf is stable, (f, A) must be
observable. So,
f A − a2 f
V (f ) =
f
is nonsingular and it can be verified that this V (f ) is the desired transfor-
mation matrix. Moreover, V (f ), V −1 (f ), b1 (f ) and b2 (f ) are all analytic
in f . Now consider the transformed system
x̂˙ = Âx̂ + b̂(f )sat(fˆx̂). (6.3.21)
Note that Â and fˆ are both independent of f . Under the state transforma-
tion, S(f ) is transformed into
Ŝ(f ) = {V (f )x : x ∈ S(f )},
the domain of attraction for (6.3.21) and ∂ Ŝ(f ) is its unique limit cycle.
Let
x̂11
p1 =
1
be a point on the line fˆx̂ = 1 such that a trajectory starting at p1 will go
above the line and return to the line (for the first time) at

ŷ11
p1 = .
1
Let T1 be the time for the trajectory to go from p1 to p1 , then

eÂT1 p1 + Â−1 b̂(f ) = p1 + Â−1 b̂(f ) ,
or equivalently,

x̂11 +(Â−1 b̂(f ))1 ŷ11 +(Â−1 b̂(f ))1
ÂT1 1+(Â−1 b̂(f ))2 1+(Â−1 b̂(f ))2
e = ,
1 1
where (·)i , i = 1, 2, denotes the ith coordinate of a vector. It can be verified

from the stability of Â + b̂(f )fˆ that 1 + (Â−1 b̂(f ))2 > 0. So Lemma 6.3.1
applies here with a changing of variables. Let us express ŷ11 as a function
of x̂11 and f , i.e., ŷ11 = ŷ11 (f, x̂11 ). By Lemma 6.3.1, ŷ11 is continuously
differentiable in x̂11 . It is easy to see that ŷ11 is also continuously differen-

tiable in f .
Suppose that the trajectory continues from p1 and intersects the line
fˆx̂ = −1 at a nonzero angle. Let

ẑ11
p1 =
−1
be the first intersection of the trajectory with fˆx̂ = −1. Note that between
fˆx̂ = 1 and fˆx̂ = −1, the vector field of (6.3.21) is

˙x̂ = (Â + b̂(f )fˆ)x̂ = 0 −a1 + b1 (f ) x̂
1 a2 + b2 (f )
and that Â+b̂(f )fˆ is Hurwitz. Hence Lemma 6.3.2 applies and we know that
ẑ11 is continuously differentiable in ŷ11 . To see that ẑ11 is also continuously
differentiable in f , recall that we have assumed that A + bf has distinct
eigenvalues, so the eigenvalues are analytic in f . From (6.B.1) in the proof
of Lemma 6.3.2, we see that T is continuously differentiable in λ1 and λ2
and hence in f for T < Td . Thus, ẑ11 is also continuously differentiable
in f . Note here that ẑ11 corresponds to y11 in (6.B.2) and ŷ11 to x11 in
(6.B.1). In summary, we can write
ẑ11 = ẑ11 (f, x̂11 ),
where ẑ11 is continuously differentiable in f and x̂11 . Now, suppose that

x̂11
p1 =
1
is a point on the limit cycle ∂ Ŝ(f ), then, we must have ẑ11 = −x̂11 , i.e.,
ẑ11 (f, x̂11 ) + x̂11 = 0, (6.3.22)
due to the symmetry of ∂ Ŝ(f ). We write
g(f, x̂11 ) = ẑ11 (f, x̂11 ) + x̂11 = 0.
By the uniqueness of the limit cycle, x̂11 is uniquely determined by f .

By Lemma 6.3.1 and Lemma 6.3.2, we know
∂ ẑ11 ∂ ẑ11 ∂ ŷ11
= < −1,
∂ x̂11 ∂ ŷ11 ∂ x̂11
hence,
∂g

= 0,
∂ x̂11
and by the implicit function theorem, x̂11 is differentiable in f . Recall that

x̂11
p1 =
1
is a point in the vector field of (6.3.21). The corresponding intersection in
the original system (6.3.17) is

−1 x̂11
V (f ) ,
1
clearly, it is also differentiable in f . 2
Combining Propositions 6.3.1 and 6.3.2, we have
Corollary 6.3.1. Consider the system (6.3.9) with A, b and f in the spec-
ified form. Given k1 and k2 , k2 > k1 > ab22 , suppose that A + kbf has
distinct eigenvalues for all k ∈ [k1 , k2 ], then
S(k) ⊂ S(k + δk ),
for all k ∈ [k1 , k2 ], δk ∈ [0, k2 − k].
Proof. By Proposition 6.3.2, ∂S(k) is continuous in k for all k ∈ [k1 , k2 ].

So (p2 )1 and hence the function ∆(k) are also continuous in k. It follows
that
min ∆(k) : k ∈ [k1 , k2 ] > 0.
By applying Proposition 6.3.1, we have the corollary. 2
It can be seen that there exists a k0 > 0 such that A + kbf has distinct
eigenvalues for all k > k0 . Thus, by Corollary 6.3.1, S(k) will be continuous
and monotonically increasing for all k > k0 .
6.3.3. Proof of Theorem 6.2.1: Second Order Systems
We will prove the theorem by explicit construction of a family of feedback

laws that solve the problem. To this end, let us first establish some prelim-
inary results for a general system (6.2.1), not necessarily of second order
or anti-stable. Let P (ε) be the positive definite solution of the algebraic
Riccati equation,
AT P + P A − P bbT P + εI = 0. (6.3.23)
It is known that P (ε) is continuous in ε for ε ≥ 0. Let
f (ε) = −bT P (ε).
With u = kf (ε)x, we have the closed-loop system
ẋ = Ax + b sat (kf (ε)x + w) . (6.3.24)
Clearly, A + kbf (ε) is Hurwitz for all k ≥ 0.5. It can also be shown that
there exists some k0 > 0 such that A + kbf (ε) has distinct eigenvalues for
all k > k0 and ε ∈ [0, 1].
For x(0) = x0 and w ∈ W, denote the state trajectory of (6.3.24) as
ψ(t, x0 , w).
Lemma 6.3.3. Consider the system (6.3.24). Let ε > 0 be given. Denote
σmax (P (ε))D2 4
c∞ = , c0 = .
ε(2k − 1) bT P (ε)b
Suppose that k is sufficiently large such that c∞ < c0 . Denote

S p (ε) := E(P (ε), c0 ) = x : xT P (ε)x ≤ c0 ,
and
S ∞ (ε, k) := E(P (ε), c∞ ) = x : xT P (ε)x ≤ c∞ .
Then, S p (ε) and S ∞ (ε, k) are invariant sets, and, for any w ∈ W, x0 ∈
S p (ε), ψ(t, x0 , w) will enter S ∞ (ε, k) in a finite time and remain in it.
Proof. Let V (x) = xT P (ε)x. It suffices to show that V̇ < 0 for all
x ∈ S p (ε) \ S ∞ (ε, k) and for all |w| ≤ D. In what follows, we simply write
P (ε) as P and f (ε) as f , since in this lemma, ε is fixed. Note that
V̇ = xT (AT P + P A)x + 2xT P b sat(kf x + w).
We will consider the case where xT P b ≥ 0. The case where xT P b ≤ 0 can

be dealt with in a similar way.
If kf x + w ≤ −1, then,
V̇ = xT (AT P + P A)x − 2xT P b

= xT P bbT P x − 2xT P b − εxT x
= xT P b(xT P b − 2) − εxT x.
Since
4
xT P x ≤ c0 = ,
bT P b
we have 1
1
bT P x ≤ bT P 2 P 2 x ≤ 2,
and hence V̇ < 0.

If kf x + w > −1, then
sat(kf x + w) ≤ kf x + w,
and
V̇ ≤ xT (AT P + P A)x + 2xT P b(kf x + w)

= −(2k − 1)xT P bbT P x − εxT x + 2xT P bw
% &2
√ w w2
=− 2k − 1xT P b − √ + − εxT x.
2k − 1 2k − 1
Since
σmax (P )D2
xT P x > c∞ = ,
ε(2k − 1)
we have
D2
xT x > .
ε(2k − 1)
It follows that V̇ < 0. 2
It is clear from Lemma 6.3.3 that, as k goes to infinity, S ∞ (ε, k) con-
verges to the origin. In particular, there exists a k such that S ∞ (ε, k) ⊂
X ∞.
For any ANCBC system, as ε → 0, P (ε) → 0, and c0 → ∞. Thus S p (ε)
can be made arbitrarily large, and with a fixed ε, we can increase k to
make c∞ arbitrarily small. So the proof of Theorem 6.2.1 would have been
completed here. However, for exponentially unstable systems, S p (ε) can be
a quite small subset of C a as ε → 0 (see Chapter 4) and hence considerable
work needs to be carried out before completing the proof.
Define the domain of attraction of the origin in the absence of distur-
bance as
S(ε, k) := x0 : lim ψ(t, x0 , 0) = 0 ,
t→∞
and in the presence of disturbance, define the domain of attraction of the
set S ∞ (ε, k) as

S D (ε, k) := x0 : lim dist (ψ(t, x0 , w), S ∞ (ε, k)) = 0, ∀ w ∈ W ,
t→∞
where dist (ψ(t, x0 , w), S ∞ (ε, k)) is the distance from the point
ψ(t, x0 , w) to the set S ∞ (ε, k). Our objective is to choose ε and k such
that
X 0 ⊂ S D (ε, k), S ∞ (ε, k) ⊂ X ∞ .
Let’s now assume that the condition c∞ < c0 in Lemma 6.3.3 is satisfied.
Then we have S p (ε) ⊂ S D (ε, k) ⊂ S(ε, k). By using the Lyapunov function
V (x) = xT P (ε)x, we can only determine a subset S p (ε) of S D (ε, k). As
ε decreases, P (ε) decreases. It was shown in [107] that if ε1 < ε2 , then
S p (ε2 ) ⊂ S p (ε1 ). Hence by decreasing ε, we can enlarge S p (ε). However,
since limε→0 S p (ε) can be much smaller than C a , we are unable to prove
that S D (ε, k) is close to C a by simply enlarging S p (ε) as was done in [85].
For this reason, we will resort to the detailed investigation on the vector
field of (6.3.24) in the presence of the disturbance.
We now continue with the proof of the theorem and focus on the second
order systems. Also assume that A is anti-stable. In this case, C a = C.
We will prove the theorem by showing that, given any X 0 ⊂ int(C), any
(arbitrarily small) X ∞ such that 0 ∈ int(X ∞ ), and any D > 0, there exist
an ε > 0 and a k ≥ 0.5 such that X 0 ⊂ S D (ε, k) and S ∞ (ε, k) ⊂ X ∞ .
Theorem 4.3.1 applies to the case where ε = 0. It means that
lim dist(S(0, k), C) = 0.

k→∞
However, when ε = 0, it is impossible to achieve disturbance rejection by

increasing the value of k even if there is no saturation. What we can do is
to first set ε = 0, choose k0 sufficiently large so that A+k0 bf (ε) has distinct
eigenvalues for all k > k0 and ε ∈ [0, 1] and X 0 ⊂ int(S(0, k0 )). Then by
the continuity of the domain of attraction stated in Proposition 6.3.2 and
the continuity of the solution of the Riccati equation, we can fix this k0 and
choose ε sufficiently small so that X 0 ⊂ int(S(ε, k0 )). By Corollary 6.3.1,
we know that S(ε, k) is nondecreasing, so X 0 ⊂ int(S(ε, k)) for all k ≥ k0 .
What remains to be shown is that for any given positive number D and a
fixed ε, we can choose k sufficiently large so that dist(S D (ε, k), S(ε, k)) is
arbitrarily small. Then we will have X 0 ⊂ S D (ε, k) for some k.
Let’s fix an ε such that
X 0 ⊂ int(S(ε, k)), ∀ k ≥ k0 .
Since ε is fixed, we can assume that a state transformation x̂ = V x, as in

(6.3.18)-(6.3.20) is performed so that

fˆ = −bT P (ε)V −1 = 0 1 , (6.3.25)

0 −a1 −b1
Â = V AV −1 = , b̂ = V b = , a1 , a2 , b1 , b2 > 0,
1 a2 −b2
(6.3.26)
where a1 , a2 > 0 is from the anti-stability of A and b1 , b2 > 0 follows from
the fact that an LQR controller has infinite gain margin and ε
= 0. (Note
that b1 = 0 iff ε = 0). Under this state transformation, the sets S p (ε),
S D (ε, k), S(ε, k), S ∞ (ε, k), C, X 0 and X ∞ are transformed respectively
into Ŝ p (ε), Ŝ D (ε, k), Ŝ(ε, k), Ŝ ∞ (ε, k), Ĉ, X̂ 0 and X̂ ∞ , all defined in an
obvious way. For example,

Ĉ = V x : x ∈ C .
Let
P̂ (ε) = (V −1 )T P (ε)V −1 .
Since ε is now fixed, we denote P̂ (ε), Ŝ p (ε), Ŝ D (ε, k), Ŝ(ε, k) and Ŝ ∞ (ε, k)
as P̂ , Ŝ p , Ŝ D (k), Ŝ(k) and Ŝ ∞ (k), respectively.
Now we consider
x̂˙ = Âx̂ + b̂ sat(k fˆx̂ + w). (6.3.27)
This standard form fits just right into Corollary 6.3.1 , so we can be sure
that Ŝ(k) increases as k increases. It follows that
Ŝ(k0 ) ⊂ Ŝ(k), ∀ k > k0 .
To satisfy the design requirement, it is necessary that no point in X̂ 0 \

X̂ ∞ can be made stationary with any |w| ≤ D. Let us first exclude this
possibility by appropriate choice of k.
For a constant w, there are three candidate equilibrium points, x̂+ e =
−1 − −1 −1
−Â b̂, x̂e = Â b̂ and x̂e = −(Â+k b̂f ) b̂w, corresponding to sat(k fˆx̂+
w ˆ
w) = 1, sat(k fˆx̂ + w) = −1 and sat(k fˆx̂ + w) = k fˆx̂ + w, respectively. For
each of them to be an actual equilibrium point, we must have,
k fˆx̂+
e + w ≥ 1, k fˆx̂−
e + w ≤ −1, or |k fˆx̂w
e + w| ≤ 1,
respectively.
Here we have

1 a2 b 1 + a1 b 2
x̂+
e = , x̂− +
e = −x̂e ,
a1 −b1
and
1 a2 b 1 + a1 b 2
x̂w = w.
e
a1 + b 1 k −b1
If Â has no complex eigenvalues, then x̂+ −
e , x̂e ∈ ∂ Ĉ (see Chapter 2), so
x̂+ −
e , x̂e ∈
/ X̂ 0 for any X̂ 0 ⊂ int(Ĉ). But if Â has a pair of complex eigen-
values, x̂+ −
e , x̂e ∈ int(Ĉ) and will be in X̂ 0 if X̂ 0 is close enough to Ĉ. So,
it is desirable that x̂+ −
e and x̂e cannot be made stationary by any |w| ≤ D.
This requires
k fˆx̂+
e + w < 1, k fˆx̂−
e + w > −1, ∀ |w| ≤ D,
which is equivalent to
b1
k + w > −1, ∀ |w| ≤ D.
a1
If D < 1, this is satisfied for all k. If D > 1, we need to choose k such that
a1
k> (D − 1).
b1
Note that this will be impossible if b1 =0, which corresponds to the case
where ε = 0. This is one reason that ε should be nonzero.
Finally, as k → ∞, x̂w
e → 0 for all w, |w| ≤ D. Hence k can be chosen
e
∈ X̂ 0 \ X̂ ∞ .
large enough such that x̂w
In summary, from the above analysis, we will restrict ourselves to k such
that

a1 D a2 b 1 + a1 b 2
k > (D − 1), ∈ X ∞. (6.3.28)
b1 a1 + b 1 k −b1
To study the vector field of (6.3.27), we rewrite it as
x̂˙ 1 = −a1 x̂2 − b1 sat(k fˆx̂ + w),

x̂˙ 2 = x̂1 + a2 x̂2 − b2 sat(k fˆx̂ + w).
The vector field is much complicated by the presence of the disturbance.

However, it still exhibits some properties which we will make use in our
construction of the desired controller.
• Above the line k fˆx̂ = D + 1, k fˆx̂ + w ≥ 1 for all |w| ≤ D, so

sat(k fˆx̂ + w) = 1, i.e., the vector x̂˙ is independent of w and is affine
in x̂. Similarly, below k fˆx̂ = −(D + 1), sat(k fˆx̂ + w) = −1.
• In the ellipsoid Ŝ p , we have shown in Lemma 6.3.3 that all the tra-
jectories will converge to Ŝ ∞ (k), which can be made arbitrarily small
by increasing the value of k.
Suppose that k is sufficiently large such that the boundary of Ŝ p inter-

sects the lines k fˆx̂ = ±(D + 1). Denote the region between k fˆx̂ = (D + 1)
and k fˆx̂ = −(D + 1), and to the left of Ŝ p as Q(k) (see the shaded region
in Fig. 6.3.4). Let

x̂m (k) = − max x̂1 : x̂ ∈ Q(k) .
If k is sufficiently large, then Q(k) lies entirely in the left-half plane, so

x̂m (k) > 0. Choose K such that
D+1 −xm (K) + a2 D+1 b2

−xm (K) + a2 < 0, K
> . (6.3.29)
K −a1 K
D+1 b1
Note that xm (k) increases as k increases Then, the vector field in Q(k) has
the following property:
Lemma 6.3.4. Suppose k > K. Then, for all x̂ ∈ Q(k), |w| ≤ D,

% & % & % &
−1 b2
−xm (k) −1 b2
tan −π < Â D+1
˙
+ b ≤ x̂ < tan . (6.3.30)
b1 k b1
This implies that for any straight line E with slope bb21 , if x̂ ∈ E ∩ Q(k),
then the vector x̂˙ points to the right of E for all w, |w| ≤ D.
Proof. Between the lines k fˆx̂ = D + 1 and k fˆx̂ = −(D + 1), sat(k fˆx̂ + w)
takes value in [−1, 1] and hence,

˙x̂ ∈ −a1 x̂2 b1
+λ : λ ∈ [−1, 1] . (6.3.31)
x̂1 + a2 x̂2 b2
For x̂ ∈ Q(k), if
% & % &
−1 b2 −1 b2
tan − π < Âx̂ < tan
,
b1 b1
0.8
0.6
E
0.4
0.2
^ ^^
k f x=D+1
Sp
Q(k)
0
−0.2
^^
k f x=−(D+1)
−0.4
−0.6
−0.8
−1
−10 −8 −6 −4 −2 0 2 4 6 8 10
Figure 6.3.4: Partition of the vector field of the system (6.3.27).
then,
% & % &
−1 b2 −1 b2
tan − π < (Âx̂ + λb̂) < tan
, ∀ λ ∈ [−1, 1]. (6.3.32)
b1 b1
Since xm (k) is increasing with k, we see from (6.3.29) that for all k > K,
D+1 −xm (k) + a2 D+1 b2

−xm (k) + a2 < 0, k
> .
k −a1 kD+1 b1
It follows that
% & % &
−xm (k)
−π <
b2
tan−1 Â D+1
b1 k
% &
−a D+1
=
1 k π −1 b2
< − < tan .
−xm (k) + a2 D+1
k 2 b 1
For all x̂ ∈ Q(k), we have

D+1
x̂1 ≤ −xm (k), |x̂2 | ≤ .
k
So, % &

−xm (k)
Â D+1 ≤ Âx̂ ≤ 0 =⇒
k
% & % &
b2 b2
tan−1 − π < Âx̂ < tan−1 .
b1 b1
Hence, by (6.3.32),
% & % &
−1 b2 b2
tan − π < x̂˙ < tan−1 ,
b1 b1
for all x̂ ∈ Q(k) and |w| ≤ D. With similar arguments, it can be further
verified that
% &

−xm (k)
min Âx̂ + λb̂ : x̂ ∈ Q(k), λ ∈ [−1, 1] ≥
Â D+1 +b
% &k
b2
> tan−1 − π.
b1
Thus, (6.3.30) follows. 2

This lemma means that any trajectory of (6.3.27) starting from inside
of Q(k) and to the right of E will remain to the right of E before it leaves
Q(k).
Based on Lemma 6.3.4, we can construct an invariant set Ŝ I (k) ⊂ Ŝ(k)
and show that it is also a subset of Ŝ D (k). Moreover, it can be made
arbitrarily close to Ŝ(k).
Lemma 6.3.5. Let K > 0 be chosen such that (6.3.29) is satisfied.
a) If k > K satisfies (6.3.28) and

% &
a2 (D + 1) b1 (D + 1)
b2 − > , (6.3.33)
k kb2
then, there exist unique p1 , p2 ∈ Ŝ(k) on the line k fˆx̂ = D + 1

such that a trajectory of (6.3.27) starting at p1 goes upward, returns
to the line at p2 and the line from p2 to −p1 has a slope of b2 /b1
(see Fig. 6.3.5, where the outer closed curve is ∂ Ŝ(k) and the upper
straight line is k fˆx̂ = D + 1).
b) Denote the region enclosed by the trajectories from ±p1 to ±p2 , and
the straight lines from ±p2 to ∓p1 as Ŝ I (k) (in Fig. 6.3.5, the region
enclosed by the inner closed curve). Then,

lim dist Ŝ I (k), Ŝ(k) = 0.
k→∞
2.5
1.5
0.5
s2 p2 p0 p1 s1
−s −p (p ) −p −s
1 1 3 2 2
−0.5
−1
−1.5
−2
−2.5
−10 −8 −6 −4 −2 0 2 4 6 8 10
Figure 6.3.5: Illustration for Lemma 6.3.5.
c) Ŝ I (k) is an invariant set and Ŝ I (k) ⊂ Ŝ D (k), i.e., it is inside the

domain of attraction of Ŝ ∞ (k).
Proof. a) Recall that ∂ Ŝ(k) is a closed trajectory of (6.3.27) with w ≡ 0.

Denote the intersections of ∂ Ŝ(k) with k fˆx̂ = D + 1 as s1 and s2 (see
Fig. 6.3.5). Let
b2 − a2 (D+1)
k
p0 = D+1
,
k
then x̂˙ 2 = 0 at p0 and to the left (right) of p0 , x̂˙ 2 < 0 (> 0). Let p1 be a
point on k fˆx̂ = D + 1 between p0 and s1 , then a trajectory starting at p1
goes upward and will return to k fˆx̂ = D + 1 at some p2 between p0 and s2 .
The point p2 is uniquely determined by p1 . We then draw a straight line
from p2 with slope bb21 . Let the intersection of the line with k fˆx̂ = −(D + 1)
be p3 . Clearly, p2 and p3 depend on p1 continuously. And the quantity
(p3 − (−s1 ))1
r(p1 ) :=
(s1 − p1 )1
also depends on p1 continuously. If p1 = s1 , then p2 = s2 . Note that the
trajectories above the line k fˆx̂ = D + 1 are independent of w and hence
are the same as those with w = 0. Since s2 and −s1 are on a trajectory of
(6.3.27) with w = 0, so by Lemma 6.3.4, −s1 must be to the right of the
straight line with slope bb21 that passes s2 . This shows that −s1 is to the
right of p3 (with p1 = s1 ) and hence
lim r(p1 ) = −∞. (6.3.34)

p1 →s1
If p1 = p0 , then p2 = p0 and
a2 (D+1)

b2 − k − 2(D+1)b
kb2
1
p3 = .
− D+1
k
So,
(s1 )1 + b2 − a2 (D+1) − 2(D+1)b1
r(p1 = p0 ) = k
kb2
.
(s1 )1 − b2 − a2 (D+1)
k
And, by condition (6.3.33), r(p1 = p0 ) > 1. In view of (6.3.34) and by the

continuity of r(p1 ), there exists a p1 between s1 and p0 such that r(p1 ) = 1,
i.e., p3 = −p1 and hence the line from p2 to −p1 has a slope of bb21 . This
shows the existence of (p1 , p2 ). Suppose on the contrary that such pair
(p1 , p2 ) is not unique and there exists (p1 , p2 ) with the same property, say, p1
to the left of p1 and p2 to the right of p2 , by Lemma 6.3.1, |p2 −p2 | > |p1 −p1 |.
But |(−p1 ) − (−p1 )| = |p2 − p2 | since the line from p2 to −p1 and that from
p2 to −p1 have the same slope. This is a contradiction.
b) We see that x̂˙ 2 = 0 at p0 . By applying Lemma 6.3.1 with a shifting
of the origin, we have
|p2 − s2 | |p0 − s2 |
> > 1.
|p1 − s1 | |p0 − s1 |
(refer to (6.3.6)). As k → ∞,

b2
s2 + s1 → 0, p0 → .
0
Since s1 and s2 are restricted to the null controllable region Ĉ, there exist
some K1 > 0 and γ > 0 such that, for all k > K1 ,
|p0 − s2 | |p2 − s2 |
≥ 1 + γ =⇒ > 1 + γ. (6.3.35)
|p0 − s1 | |p1 − s1 |
From Fig. 6.3.5 , we see that
|p2 − s2 | = |p1 − s1 | + (−s1 − s2 )1 + (p2 − (−p1 ))1 .

2(D+1)
As k → ∞, k → 0, so (p2 − (−p1 ))1 → 0. Since s1 + s2 → 0, we have
|p2 − s2 | − |p1 − s1 | → 0.
From (6.3.35),
|p2 − s2 | − |p1 − s1 | > γ|p1 − s1 |.
So we must have |p1 − s1 | → 0 and hence |p2 − s2 | → 0. Therefore,

lim dist Ŝ I (k), Ŝ(k) = 0.
k→∞
c) First we show that Ŝ I (k) is an invariant set. Note that the part of
∂ Ŝ I (k) from p1 to p2 and that from −p1 to −p2 are trajectories of (6.3.27)
under any |w| ≤ D. At any point on the line from p2 to −p1 , Lemma 6.3.4
says that x̂˙ directs to the right side of the line, i.e., no trajectory can cross
the line between p2 and −p1 leftward. By symmetry, no trajectory can cross
the line between −p2 and p1 rightward. These show that no trajectory can
cross ∂ Ŝ I (k) outward, thus Ŝ I (k) is an invariant set. Since Ŝ p is also an
invariant set and any trajectory that starts from inside of it will converge
to Ŝ ∞ (k), it suffices to show that any trajectory that starts from inside of
Ŝ I (k) will enter Ŝ p . We will do this by contradiction.
Suppose that there exist an x̂0 ∈ Ŝ I (k) \ Ŝ p and a w ∈ W such that
ψ(t, x̂0 , w) ∈ Ŝ I (k) \ Ŝ p for all t > 0, then there must be a point x̂∗ ∈
Ŝ I (k) \ Ŝ p such that either
i) limt→∞ ψ(t, x̂0 , w) = x̂∗ ; or
ii) there exits a sequence t1 , t2 , · · · , ti , · · · such that
lim ψ(ti , x̂0 , w) = x̂∗ ,

i→∞
and there is an ε > 0 such that for any T > 0, there exists a t > T
satisfying |ψ(t, x̂0 , w) − x̂∗ | > ε.
Item i) implies that x̂∗ can be made stationary by some w ∈ W. This

is impossible as we have shown that k has been chosen such that all the
stationary points are inside Ŝ ∞ (k). Item ii) implies that there is a closed
trajectory with length greater than 2ε that passes through x̂∗ . There are
two possibilities here: the closed trajectory encloses Ŝ p or it does not enclose
Ŝ p . We will show that none of the cases is possible.
2.5
1.5
0.5 ^
s2 p2 q2 Sp q1 p1 s
1
−s1 −p1 q3 q4 −p2 −s2

−0.5
−1
−1.5
−2
−2.5
−10 −8 −6 −4 −2 0 2 4 6 8 10
Figure 6.3.6: Illustration of the proof.
Suppose that there is a closed trajectory that encloses Ŝ p . Let q1 , q2 , q3

and q4 be the four intersections of the closed trajectory with k fˆx̂ = ±(D+1)
as shown in Fig. 6.3.6.
By Lemma 6.3.1,
|p2 − q2 | > |p1 − q1 |, |q4 − (−p2 )| > |q3 − (−p1 )|,
and by Lemma 6.3.4,
|q3 − (−p1 )| ≥ |p2 − q2 |, |p1 − q1 | ≥ |q4 − (−p2 )|.
So we have
|p2 − q2 | > |p1 − q1 | ≥ |q4 − (−p2 )| > |q3 − (−p1 )| ≥ |p2 − q2 |.
This is a contradiction. Therefore, there exists no closed trajectory that

encloses Ŝ p . We next exclude the other possibility.
Clearly, there can be no closed trajectory that is completely above
k f x̂ = D + 1 or below k fˆx̂ = −(D + 1). So if there is a closed trajec-
ˆ
tory, it must intersect k fˆx̂ = D + 1 or k fˆx̂ = −(D + 1) to the left (or
to the right) of Ŝ p at least twice, or lie completely within Q(k). We as-

sume that the intersections are to the left of Ŝ p . Since k > K satisfies
(6.3.33), so xm (k) > 0 and b2 − a2 (D+1) k > 0. Hence, for all points on the
line k fˆx̂ = D + 1 to the left of Ŝ p , x̂˙ 2 < 0, so no closed trajectory lying
between Ŝ I (k) and Ŝ p will cross this piece of straight line twice. Below the
line k fˆx̂ = −(D + 1), x̂˙ 1 > 0. By Lemma 6.3.4, no trajectory in Q(k) will
cross a line that has slope b2 /b1 leftward. Hence there will be no closed
trajectory crossing the line k fˆx̂ = −(D + 1) to the left of Ŝ p twice. In view
of Lemma 6.3.4, there exists no closed trajectory completely inside Q(k).
These show that there exists no closed trajectory that does not enclose Ŝ p
either.
In conclusion, for every x̂0 ∈ S I (k) and w ∈ W, there must be a T < ∞
such that ψ(T, x̂0 , w) ∈ Ŝ p . And since Ŝ p is in the domain of attraction of
Ŝ ∞ (k), it follows that x̂0 ∈ Ŝ D (k) and hence S I (k) ⊂ Ŝ D (k). 2
The proof of Theorem 6.2.1 can be completed by invoking Lemmas 6.3.3
and 6.3.5. For clarity, we organize it as follows. Our proof will also include
a constructive method to choose the parameters ε and k.
Proof of Theorem 6.2.1. Given an X 0 ⊂ int(C), an X ∞ such that 0 ∈
int(X ∞ ) and a D > 0, we need to choose ε and k such that X 0 ⊂ S D (ε, k)
and S ∞ (ε, k) ⊂ X ∞ .
Step 1. Let ε = 0 and find k0 such that X 0 ⊂ int(S(0, k0 )). This is

guaranteed by Theorem 4.3.1. Increase k0 , if necessary, such that
A + kbf (ε) has distinct eigenvalues for all k > k0 and ε ∈ [0, 1].
Step 2. Find an ε ∈ (0, 1] such that X 0 ⊂ int(S(ε, k0 )). This is possible

because, by Proposition 6.3.2, S(ε, k0 ) is continuous in f (ε), and f (ε)
is continuous in ε.
Step 3. Fix ε and perform state transformation x̂ = V x such that

(fˆ, Â, b̂) is in the form of (6.3.25) and (6.3.26). Also perform this
transformation to the sets X 0 and X ∞ to get X̂ 0 and X̂ ∞ . We
don’t need to transform S(ε, k0 ) to Ŝ(k0 ) but should remember that
X̂ 0 ⊂ int(Ŝ(k0 )).
Step 4. Find a k > K satisfying (6.3.28) and (6.3.33) such that X̂ 0 ⊂

Ŝ I (k). Since X̂ 0 ⊂ int(Ŝ(k0 )), we have X̂ 0 ⊂ int(Ŝ(k)) for all k > k0 .
Hence by Lemma 6.3.5, X̂ 0 ⊂ Ŝ I (k) ⊂ Ŝ D (k) for some k > 0. It
follows that X 0 ⊂ S D (ε, k).
Step 5. Increase k, if necessary, so that S ∞ (ε, k) ⊂ X ∞ . This can be

satisfied due to Lemma 6.3.3. 2
6.3.4. Proof of Theorem 6.2.1: Higher Order Systems
As with the stabilization problem considered in Chapter 4, where the dis-

turbance is absent, the main idea in this proof is first to bring those expo-
nentially unstable states to a “safe set” by using partial state feedback, and
then to switch to a full state feedback that steers all the states to a neigh-
borhood of the origin. The first step control is justified in Section 6.3.3 and
the second step control is guaranteed by Lemma 6.3.3 and the property of
the solution of the Riccati equation, which allows the states that are not
exponentially unstable to grow freely.
For easy reference, we recall some assumptions and notation used in
Section 4.4. Assume that the matrix pair (A, b) in the system (6.2.1) is in
the form of
A1 0 b1
A= , b= ,
0 A2 b2
where x = (xa , xs ), A1 ∈ R2×2 is anti-stable and A2 ∈ Rn is semi-stable.
Assume that (A, b) is stabilizable. Denote the null controllable region of
the subsystem
ẋa = A1 xa + b1 u
as C 1 . Then, by Theorem 2.7.1, the asymptotically null controllable region
of (6.2.1) is C a = C 1 × Rn . Given any γ1 ∈ (0, 1), and γ2 > 0, denote

Ω1 (γ1 ) := γ1 xa ∈ R2 : xa ∈ C 1 ,
and
Ω2 (γ2 ) := xs ∈ Rn : |xs | ≤ γ2 .
For any compact subset X 0 of C a = C 1 × Rn , there exist a γ1 ∈ (0, 1) and
a γ2 > 0 such that X 0 ⊂ Ω1 (γ1 ) × Ω2 (γ2 ). For this reason, we assume,
without loss of generality, that X 0 = Ω1 (γ1 ) × Ω2 (γ2 ).
For an ε > 0, let

P1 (ε) P2 (ε)
P (ε) = ∈ R(2+n)×(2+n)
P2T (ε) P3 (ε)
be the unique positive definite solution to the ARE
AT P + P A − P bbT P + εI = 0. (6.3.36)
Clearly, as ε ↓ 0, P (ε) decreases. Hence limε→0 P (ε) exists.

Let P 1 be the unique positive definite solution to the ARE
AT1 P 1 + P 1 A1 − P 1 b1 bT1 P 1 = 0.
Then,
P1 0
lim P (ε) = .
ε→0 0 0
Let
f (ε) := −bT P (ε).
We first study the following closed-loop system
ẋ = Ax + b sat(kf (ε)x + w). (6.3.37)
Recall from Lemma 6.3.3 that the invariant set S p (ε) is in the domain
of attraction of the set S ∞ (ε, k) for sufficiently large k’s.
Lemma 6.3.6. Denote

1
r1 (ε) := 1 ,
2 T 12
P1 (ε) b P (ε)
'
−|P2 (ε)| + |P2 (ε)|2 + 3|P1 (ε)| |P3 (ε)|
r2 (ε) := r1 (ε).
|P3 (ε)|
Then,

D1 (ε) := x ∈ R2+n : |xa | ≤ r1 (ε), |xs | ≤ r2 (ε) ⊂ S p (ε).
Moreover,
lim r2 (ε) = ∞,
ε→0
Proof. Similar to the proof of Lemma 4.4.1 in Chapter 4. 2

Proof of Theorem 6.2.1. Denote X a0 = Ω(γ1 ), then X a0 ⊂ int(C 1 ).
Given an ε0 > 0, let

X a∞ = xa ∈ R2 : |xa | ≤ r1 (ε0 ) .
By the result on second order systems, there exists a controller u = f1 xa

such that any trajectory of
ẋa = A1 xa + b1 sat(f1 xa + w) (6.3.38)

that starts from within X a0 will converge to X a∞ in a finite time and stay
there. Denote the trajectory of (6.3.38) that starts at xa0 as ψ1 (t, xa0 , w)
and define

TM := max min t > 0 : ψ1 (t, xa0 , w) ∈ X a∞ .
xa0 ∈∂ X a0 ,w∈W
It can be shown that TM < ∞. Let

TM
A2 (TM −τ )
γ = max eA2 t γ2 + e b2 dτ, (6.3.39)
t∈[0,TM ] 0
then, by Lemma 6.3.6, there exists an ε < ε0 such that r1 (ε) ≥ r1 (ε0 ),
r2 (ε) ≥ γ and

D1 (ε) = x ∈ R2+n : |xa | ≤ r1 (ε), |xs | ≤ r2 (ε) ⊂ S p (ε)
lies in the domain of attraction of S ∞ (ε, k).

Choose a k such that S ∞ (ε, k) ⊂ X ∞ , and let the combined controller
be
f1 xa , if x ∈
/ S p (ε),
u= (6.3.40)
kf (ε)x, if x ∈ S p (ε).
Consider an initial state of the closed-loop system of (6.2.1) with (6.3.40),
x0 ∈ Ω1 (γ1 )× Ω2 (γ2 ). If x0 ∈ S p (ε), then x(t) will enter S ∞ (ε, k) ⊂ X ∞ . If
x0
∈ S p (ε), we conclude that x(t) will enter S p (ε) at some T1 ≤ TM under
the control u = f1 xa . Observe that, under this control, xa (t) goes along
a trajectory of (6.3.38). If there is no switch, xa (t) will hit the ball X a∞
at T1 ≤ TM and at this instant |xs (T1 )| ≤ γ ≤ r2 (ε), so x(T1 ) ∈ D1 (ε).
Thus, we see that, if there is no switch, x(t) will be in D1 (ε) at T1 . Since
D1 (ε) ⊂ S p (ε), x(t) must have entered S p (ε) at some earlier time T ≤ T1 ≤
TM . So we have the conclusion. With the switching control applied, once
x(t) enters the invariant set S p (ε), it will converge to S ∞ (ε, k) and remain
there. 2
Remark 6.3.1.
1) Similar to Remark 4.4.1, Theorem 6.2.1 is also true for the case that
the open loop system has a single exponentially unstable pole. Let
x, A, b and P (ε) be partitioned similarly to the two exponentially
unstable pole case, with xa , A1 , b1 , P1 (ε) ∈ R. For the anti-stable
sub-system, the controller has the form
2A1 1
u = f1 xa = −k xa , k> .
b1 2
It can be easily shown that given any X a0 in the interior of

% &
|b1 | |b1 |
C1 = − ,
A1 A1
1
and any X a∞ containing the origin in its interior, there exists a k > 2
such that any trajectory of
% &
2A1
ẋa = A1 xa + b1 sat − kxa + w
b1
that starts from within X a0 will converge to X a∞ in a finite time

and remain there. Hence, the proof of Theorem 6.2.1 can be readily
adapted to this case.
Unlike Remark 4.4.1, where any k > 12 is eligible, here in the presence
of disturbance, as X a0 becomes larger or X a∞ becomes smaller, k has
to be increased to meet the requirements.
2) For the case that the open-loop system has no exponentially unstable
poles (ANCBC), the controller has the form of u = kf (ε)x and no
switching is necessary. But unlike Remark 4.4.1, we have to increase
k if X 0 is increased or X ∞ is decreased.
6.4. An Example
In this section, we will use an aircraft model to demonstrate the results ob-
tained in this chapter. Consider the longitudinal dynamics of the TRANS3
aircraft under certain flight condition [51],
    
ż1 0 14.3877 0 −31.5311 z1
 ż2   −0.0012 −0.4217 1.0000 −0.0284   
     z2 
 ż3  =  0.0002 −0.3816 −0.4658 0   z3 
ż4 0 0 1.0000 0 z4
 
4.526
 −0.0337 
+ 
 −1.4566  v.
0
The states z1 , z2 , z3 and z4 are the velocity, the angle of attack, the pitch
rate and the Euler angle rotation of aircraft about the inertial y-axis, re-
spectively. The control v is the elevator input, which is bounded by 10
6.4. An Example 145
degree, or 0.1745 rad. With a state transformation of the form x = T z and

the input normalization such that the control is bounded by 1, we obtain

ẋa A1 0 xa b1
= + sat(u + w),
ẋs 0 A2 xs b2
where

0.0212 0.1670 −0.4650 0.6247
A1 = , A2 = ,
−0.1670 0.0212 −0.6247 −0.4650
and
8.2856 0.7584
b1 = , b2 = .
−2.4303 −1.8562
The system has two stable modes −0.4650 ± j0.6247 and two anti-stable
ones, 0.0212 ± j0.1670. Suppose that w is bounded by |w| ≤ D = 2.
For the anti-stable xa -subsystem, we take γ1 = 0.9. With the technique
used in the proof of Theorem 6.2.1 for second order systems, we obtain a
feedback u = f1 xa , where

f1 = −0.4335 0.2952 ,
such that Ω1 (γ1 ) is inside some invariant set S I . Moreover, for all initial
xa0 ∈ S I , under the control u = f1 xa , xa (t) will enter the ball

X a = xa ∈ R2 : |xa | ≤ 29.8501 .
In Fig. 6.4.1, the outermost closed curve is the boundary of the null con-
trollable region ∂C 1 , the inner dash-dotted closed curve is ∂S I , the dashed
closed curve is ∂Ω1 (γ1 ), and the innermost solid closed curve is ∂X a .
The xs -subsystem is exponentially stable. Under the saturated control,
it can be shown that for any initial value xs0 ∈ R2 , there exists a T > 0
such that xs (t) will enter a bounded ball at time T and remain there. The
bounded ball is computed as

X s = xs ∈ R2 : |xs | ≤ 4 .
We see that, for any (xa0 , xs0 ) ∈ S I × R2 , under the partial feedback
control u = f1 xa , the state (xa , xs ) will enter the set X a × X s in a finite
time and remain there. The next step is to design a full state feedback
to make the set X a × X s inside the domain of attraction of an arbitrarily
small set.
300
200
100
−100
−200
−300
−300 −200 −100 0 100 200 300
Figure 6.4.1: Design partial feedback u = f1 xa such that Ω1 (γ1 ) ⊂ S I .
Choose ε = 0.03, we get

 
0.9671 0.0005 −0.0686 0.0375
 0.0005 0.9664 0.0345 −0.0410 
P (ε) = 0.001 × 
 −0.0686
,
0.0345 4.1915 −0.7462 
0.0375 −0.0410 −0.7462 11.3408

f (ε) = 0.001 × −0.0729 1.408 −36.4271 33.6402 ,
and

S p (ε) = x ∈ R4 : xT P (ε)x ≤ 10.3561 .
It can be verified that X a × X s ⊂ S p (ε). This implies that, under the

control u = f1 xa , every trajectory starting from within X a × X s will enter
S p (ε) in a finite time. If k is sufficiently large, then, under the control
u = kf (ε)x, S p (ε) will be an invariant set. In this case, the switching
controller (6.3.40) is well-defined.
The final step is to choose k sufficiently large such that the state will
converge to an arbitrarily small neighborhood of the origin. We illustrate
this point by simulation results for different values of k. In the simulation,
6.. Conclusions 147
300
x12
k=2.5
200
100
−100
−200
x
11
−300
−300 −200 −100 0 100 200 300
Figure 6.4.2: A trajectory of xa with ε = 0.03, k = 2.5.
we choose w(t) = 2 sin(0.1t) and xa0 to be a point very close to the bound-
ary of S I (see the point marked with “◦” in Fig. 6.4.2 and Fig. 6.4.4). We
also set

1000
xs0 = ,
1000
which is very far away from the origin. When k = 2.5, the disturbance is
not satisfactorily rejected (see Fig. 6.4.2 for a trajectory of xa and Fig. 6.4.3
for the time response of |x(t)|). When k = 30, the disturbance is rejected
to a much higher level of accuracy (see Fig. 6.4.4 and Fig. 6.4.5).
6.5. Conclusions
In this chapter, we solved the problem of semi-global practical stabilization

for linear exponentially unstable systems subject to actuator saturation and
input additive disturbance. We have assumed that the open loop system
has no more than two anti-stable modes. Our development relies heavily
on limit cycle theory and vector field analysis of the exponentially unstable
subsystem. It is, however, not expected that these results can be extended
500
450
k=2.5
400
350
300
2
||x(t)||
250
200
150
100
50
0
0 50 100 150 200 250 300 350
time t
Figure 6.4.3: Time response of |x(t)| with ε = 0.03, k = 2.5.
300
x12
k=30
200
100
−100
−200
x
11
−300
−300 −200 −100 0 100 200 300
Figure 6.4.4: A trajectory of xa with ε = 0.03, k = 30.

6.A. Proof of Lemma 6.3.1 149
500
450
k=30
400
350
300
2
||x(t)||
250
200
150
100
50
0
0 50 100 150 200 250 300 350
time t
Figure 6.4.5: Time response of |x(t)| with ε = 0.03, k = 30.
in a direct way to systems with more than two exponentially unstable open
loop poles.
6.A. Proof of Lemma 6.3.1
Since at the intersection p , the trajectory goes downward, it follows that

y11 < − ak2 . Using the fact that
1
f p = f p =
k
and
p = eAT p,
we have,

x11
0 k eAT 1 = 1, (6.A.1)
k

−AT y11
0 k e 1 = 1. (6.A.2)
k
From (6.A.1) and (6.A.2), we can also express x11 and y11 as functions of T .
In other words, x11 and y11 are related to each other through the parameter
T . Since the valid domain of x11 can be finite or infinite depending on
the location of the eigenvalues of A, it is necessary to break the proof for
different cases. We will see soon that the relation among x11 , y11 and T is
quite different for different cases.
Case 1. The matrix

0 −λ1 λ2
A=
1 λ1 + λ2
has two different real eigenvalues λ1 , λ2 > 0. Assume that λ1 > λ2 .
Let
−λ2 −λ1
V = ,
1 1
then,
eλ1 T 0
eAT = V V −1 .
0 eλ2 T
From (6.A.1) and (6.A.2) we have,
1 λ1 − λ2 + λ2 eλ2 T − λ1 eλ1 T
x11 (T ) = · , (6.A.3)
k eλ1 T − eλ2 T
1 λ1 − λ2 + λ2 e−λ2 T − λ1 e−λ1 T
y11 (T ) = · . (6.A.4)
k e−λ1 T − e−λ2 T
Due to the uniqueness of the trajectory, T is also uniquely determined by
x11 . So, x11 ↔ T , x11 ↔ y11 , y11 ↔ T are all one to one maps. From
the above two equations, we know that x11 (T ) and y11 (T ) are analytic on
(0, ∞). It can be verified from (6.A.3) that
λ1 + λ2 a2 λ1
lim x11 = − =− , lim x11 = − = am .
T →0 k k T →∞ k
So we know the valid domain of x11 is (− ak2 , am ). It can also be verified
that
dx11
> 0,
dT
or
dT
> 0.
dx11
Denote
dy11
g(T ) := − ,
dx11
6.A. Proof of Lemma 6.3.1 151
then,
λ1 − λ2 + λ2 eλ1 T − λ1 eλ2 T
g(T ) = .
λ1 − λ2 + λ2 e−λ1 T − λ1 e−λ2 T
lim g(T ) = 1,
T →0
and
dg λ1 λ2 (eλ1 T − eλ2 T )
= h(T ),
dT (λ1 − λ2 + λ2 e−λ1 T − λ1 e−λ2 T )2
where

h(T ) = (λ1 − λ2 ) 1 − e−(λ1 +λ2 )T + (λ1 + λ2 ) e−λ1 T − e−λ2 T .
Since h(0) = 0 and
dh
= (λ1 + λ2 )e−(λ1 +λ2 ) λ1 − λ2 + λ2 eλ1 T − λ1 eλ2 T > 0, ∀ T > 0,
dT
we have h(T ) > 0, hence
dg
> 0.
dT
This shows that g(T ) > 1 for all T > 0, i.e.,
dy11
< −1.
dx11
Since
dg dg(T ) dx11 d2 y11 dx11
= · =− 2 · ,
dT dx11 dT dx11 dT
and
dg dx11
> 0, > 0,
dT dT
it follows that
d2 y11
< 0.
dx211
Case 2. The matrix

0 −λ2
A=
1 2λ
has two identical real eigenvalues λ > 0.
Let
−λ 1
V = ,
1 0
then,
1 T
eAT = V V −1 eλT .
0 1
In this case,
1
x11 (T ) = − 1 + λT − e−λT ,
kT
1
y11 (T ) = 1 − λT − eλT .
kT
Similar to Case 1, it can be shown that
2λ a2 λ
lim x11 = − =− , lim x11 = − = am .
T →0 k k T →∞ k
So the valid domain of x11 is (− ak2 , am ). It can also be verified that
dx11
> 0.
dT
Denote
dy11
g(T ) := − ,
dx11
then,
1 − eλT + λT eλT
g(T ) = , lim g(T ) = 1,
1 − e−λT − λT e−λT T →0
and
dg λ2 T
= 2 h(T ),
dT (1 − e−λT − λT e−λT )
where
h(T ) = eλT − e−λT − 2λT.
It can be shown that h(T ) > 0, hence
dg
> 0.
dT
The remaining part is similar to Case 1.
Case 3. The matrix
0 −(α2 + β 2 )
A=
1 2α
has two complex eigenvalues α ± jβ, α, β > 0.
Let
β −α
V = ,
0 1
6.B. Proof of Lemma 6.3.2 153
then,
cos βT − sin βT
eAT = V V −1 eαT .
sin βT cos βT
From (6.A.1) and (6.A.2) we have,
1
x11 (T ) = −β cos βT − α sin βT + βe−αT ,
k sin βT
1
y11 (T ) = β cos βT − α sin βT − βeαT .
k sin βT
The valid domain of T is (0, πβ ), this can be obtained directly from the
vector field and also from the above equations. Notice that
2α a2
lim x11 (T ) = − =− , lim x11 (T ) = ∞.
T →0 k k T→π
β
So we have am = ∞ in this case.

Define g(T ) similarly as in Case 1, we have
β + (α sin βT − β cos βT )eαT
g(T ) = , lim g(T ) = 1,
β − (α sin βT + β cos βT )e−αT T →0
and
dg (α2 + β 2 ) sin βT
= 2 h(T ),
dT (β − (α sin βT + β cos βT )e−αT )
where
h(T ) = βeαT − βe−αT − 2α sin βT.
It can be verified that h(0) = 0 and
dh
> 0,
dT
thus, % &
dg π
> 0, ∀ T ∈ 0, .
dT β
The remaining part is similar to that in Case 1. 2
6.B. Proof of Lemma 6.3.2

Similar to the proof of Lemma 6.3.1, from (6.3.8), we can express x11 and y11
as functions of T , x11 (T ) and y11 (T ). Clearly these functions are analytic.
Denote
dy11 (T )
dT
g(T ) := dx11 (T )
.
dT
It suffices to show that

dx11 dg
> 0, g(T ) > 1, > 0.
dT dT
We need to break the proof into three different cases.
Case 1. The matrix
0 −λ1 λ2
A=
1 −(λ1 + λ2 )
has two different real eigenvalues −λ1 , −λ2 > 0. Assume that λ1 > λ2 .
Let
λ2 λ1
V = ,
1 1
then,
e−λ1 T 0
eAT = V V −1 .
0 e−λ2 T
From (6.3.8) and the fact that kf p = 1, kf p = −1, we have, for T ∈ (0,Td ),
1 λ2 − λ1 + λ2 e−λ2 T − λ1 e−λ1 T
x11 (T ) = , (6.B.1)
k e−λ2 T − e−λ1 T
1 λ2 − λ1 + λ2 eλ2 T − λ1 eλ1 T
y11 (T ) = , (6.B.2)
k eλ1 T − eλ2 T
and,
λ2 − λ1 + λ2 e−λ1 T − λ1 e−λ2 T
g(T ) = .
λ2 − λ1 + λ2 eλ1 T − λ1 eλ2 T
By the definition of Td ,
a2 λ1 + λ2
y11 (Td ) = =− .
k k
It can be shown that as T → Td , g(T ) → ∞. Since g(0) = 1 and
dg 2λ1 λ2
= (λ1 + λ2 )[ch(λ1 T − λ2 T ) − 1]
dT (λ2 − λ1 + λ2 eλ1 T − λ1 eλ2 T )2

+ (λ2 − λ1 )[ch(λ2 T ) − ch(λ1 T )] > 0,
where
ea + e−a
ch(a) = ≥1
2
is monotonously increasing, we have that
g(T ) > 1, ∀ T ∈ (0, Td ).

6.B. Proof of Lemma 6.3.2 155
It can also be verified that

dx11
> 0.
dT
The remaining proof is similar to the proof of Lemma 6.3.1.
Case 2. The matrix
0 −λ2
A=
1 −2λ
has two identical real eigenvalues.
Let
λ 1
V = ,
1 0
then,
1 T
e AT
=V V −1 e−λT .
0 1
In this case, for all T ∈ (0, Td ), we have,
1
x11 (T ) = − 1 − λT + eλT ,
kT
1
y11 (T ) = − 1 + λT + e−λT ,
kT
and
1 + λT e−λT + e−λT
g(T ) = .
1 − λT eλT + eλT
Since g(0) = 1 and

dg λ2 T 2λT + eλT − e−λT
= 2 > 0,
dT (1 − λT eλT + eλT )
we have g(T ) > 1 for all T ∈ (0, Td ). It can be verified that

dx11
> 0.
dT
Case 3. The matrix

0 −(α2 + β 2 )
A=
1 −2α
has two complex eigenvalues −α ± jβ, α, β > 0.
Let
β α
V = ,
0 1
then,
cos βT − sin βT
eAT = V V −1 e−αT .
sin βT cos βT
π
In this case, Td < β,
1
x11 (T ) = − β cos βT − α sin βT + βeαT ,
k sin βT
1
y11 (T ) = − β cos βT + α sin βT + βe−αT ,
k sin βT
and
β + (β cos βT + α sin βT )e−αT
g(T ) = , T ∈ (0, Td ).
β + (β cos βT − α sin βT )eαT
Since g(0) = 1 and
dg (α2 + β 2 ) sin βT
= 2
dT (β + (β cos βT − α sin βT )eαT )

× 2α sin βT + β(eαT − e−αT ) > 0
we have g(T ) > 1 for all T ∈ (0, Td ). It can also be verified that
dx11
> 0.
dT
For all the above three cases, since g(T ) > 1, i.e.,
dy11 dx11
> , ∀ T ∈ (0, Td )
dT dT
and
x11 (T )
lim = 1.
T →0 y11 (T )
We thus have y11 > x11 . 2

Chapter 7
Estimation of the Domain

of Attraction under
Saturated Linear
Feedback
7.1. Introduction
The problem of estimating the domain of attraction of an equilibrium

of a nonlinear dynamical system has been extensively studied for several
decades (see, e.g., [16,17,25,32,40,50,55,58,75,81,84,104,105] and the refer-
ences therein). In Section 4.2, we presented a simple method for deter-
mining the domain of attraction for a second order linear system under a
saturated linear feedback. For general higher order systems, exact descrip-
tion of the domain of attraction seems impossible. Our objective in this
chapter is to obtain an estimate of the domain of attraction, with the least
conservatism, for general linear systems under saturated linear feedback.
Our presentation draws on materials from our recent work [40].
Consider the following linear state feedback systems subject to actuator
saturation,
ẋ = Ax + Bsat(F x),
and
x(k + 1) = Ax(k) + Bsat(F x(k)),
157
158 Chapter 7. Estimation of Domain of Attraction
where x ∈ Rn , u ∈ Rm and F is a pre-designed state feedback matrix.

The problem of estimating the domains of attraction for such systems has
been a focus of study in recent years. Various results have been developed
toward less conservative and computable estimations. In particular, simple
and general methods have been derived by applying the absolute stability
analysis tools, such as the circle and Popov criteria, where the saturation is
treated as a locally sector bounded nonlinearity and the domain of attrac-
tion is estimated by use of quadratic and Lur’e type Lyapunov functions.
The multivariable circle criterion in [55] is translated into (nonlinear) ma-
trix inequalities in [32,81]. The matrix inequalities contain the controller
parameters and other auxiliary parameters, such as the positive definite
matrix P in the Lyapunov function V (x) = xT P x and the saturation lev-
els. By fixing some of the parameters, these matrix inequalities simplify to
linear matrix inequalities(LMIs) and can be treated with the LMI software.
A nice feature of these analysis tools is that they can easily be adapted for
controller synthesis by simply considering the feedback gain matrix as an
additional optimization parameter.
Since the circle criterion is applicable to general memoryless sector
bounded nonlinearities, we can expect the conservatism in estimating the
domain of attraction when it is applied to the saturation nonlinearity. In
this chapter, less conservative conditions for an ellipsoid to be in the domain
of attraction are derived by exploring the special property of the saturation
function. These conditions are given directly in terms of LMIs. Hence they
are very easy to handle in both analysis and design. The main results are
based on the idea of placing the saturated control sat(F x) in the convex
hull of a group of linear controls. By further exploiting this idea, we will
reveal a surprising fact for single input systems: suppose that an ellipsoid
is made invariant with a linear feedback u = F x, then it is invariant under
the saturated linear feedback u = sat(F x) if and only if it can be made
invariant with any saturated, possibly nonlinear, feedback. This means
that the set invariance property under saturated linear feedback is in some
sense independent of a particular feedback as long as the corresponding
linear feedback makes the ellipsoid invariant.
This chapter is organized as follows. Section 7.2 introduces a measure
of the size of a set. Section 7.3 presents some simple facts about the convex
hulls which will be used for placing sat(F x) into the convex hull of a group
of linear controls. Sections 7.4 and 7.5 consider the continuous-time and
7.2. A Measure of Set Size 159
discrete-time systems under state feedback, respectively. More specifically,

Section 7.4.1 gives a brief review of the circle criterion. Section 7.4.2 derives
a new condition for set invariance and shows that this new condition is less
conservative than the existing ones. Section 7.4.3 provides a necessary and
sufficient condition for set invariance. Based on the results in Section 7.4.2,
an LMI-based approach to enlarging the estimate of the domain of attrac-
tion is developed in Section 7.4.4. The development of Section 7.5 parallels
that of Section 7.4. Section 7.6 extends the results to systems under output
feedback. Section 7.7 draws conclusions to the chapter.
7.2. A Measure of Set Size

The objective of this chapter is to obtain estimates of the domain of attrac-
tion and to reduce the conservatism in the estimation. At this point, we
need a suitable measure of the size of a set so that the problem of optimizing
the estimate can be exactly and meaningfully formulated, and moreover,
can be treated easily. A traditional measure of the size of a set is its vol-
ume. For an ellipsoid E(P, ρ), its volume is proportional to det(P/ρ)−1 . In
this chapter, we will introduce a new measure which takes the shape of a
set into consideration. The idea is to introduce a shape reference set.
Let X R ⊂ Rn be a bounded convex set of some desired shape. We call
it a shape reference set. Suppose that 0 ∈ X R . For a positive real number
α, denote
αX R = αx : x ∈ X R .
Our desired measure of the largeness of sets should have the following prop-
erty: A set X 1 is larger than another set X 2 if the largest α1 X R contained
in X 1 is larger than the largest α2 X R contained in X 2 . The following
definition of size will possess this property.
For a set S ⊂ Rn , define the size of S with respect to X R as

αR (S) := sup α > 0 : αX R ⊂ S .
If αR (S) ≥ 1, then X R ⊂ S. Two typical types of X R are the ellipsoids

X R = x ∈ Rn : xT Rx ≤ 1 ,
with R > 0, and the polyhedrons

X R = co x1 , x2 , · · · , xl ,
where co{·} denotes the convex hull of a group of vectors.

For example, suppose that X R is an ellipsoid and R is a diagonal ma-
trix, then the state corresponding to a smaller diagonal element is desired
to have a larger range of variation in the domain of attraction. The rela-
tive weightings at the diagonals of R can be determined from the physical
operating range of the states. On the other hand, if we know the possible
initial conditions of the system, we can take X R as a polyhedron with all
the possible initial conditions on its vertices. At the extreme case, we can
take X R = co{0, x0 }. By using the methods to be developed in this chapter,
we can verify if x0 is in the domain of attraction.
It turns out that this measure of size is very flexible to use if we have
some information about the initial conditions. For example, we know that
the initial condition of the dynamic output feedback controller is 0. So
we can put a large weighting on the diagonals of R corresponding to these
states, as will be seen in Section 7.6. Also, we will see later that this measure
of size is very easy to handle in the optimization problems resulting from
estimation of the domain of attraction.
7.3. Some Facts about Convex Hulls

The goal of this section is to place sat(F x) into the convex hull of a group
of linear feedbacks. To this end, we need to establish some simple facts
about the convex hull. Recall that for a group of points, u1 , u2 , · · · , uI , the
convex hull of these points is defined as,
I
I
co u : i ∈ [1, I] :=
i i
αi u : αi = 1, αi ≥ 0 .
i=1 i=1
Lemma 7.3.1. Let u, u1 , u2 , · · · , uI ∈ Rm1 , v, v 1 , v 2 , · · · , v J ∈ Rm2 . If

u ∈ co{ui : i ∈ [1, I]} and v ∈ co{v j : j ∈ [1, J ]}, then
i
u u
∈ co : i ∈ [1, I], j ∈ [1, J ] . (7.3.1)
v vj
Proof. Since u ∈ co{ui : i ∈ [1, I]} and v ∈ co{v j : j ∈ [1, J ]}, there exist
αi , βj ≥ 0, i = 1, 2, · · · , I, j = 1, 2, · · · , J , such that
I
J

αi = βj = 1,
i=1 j=1
7.3. Some Facts about Convex Hulls 161
and
I
J

u= αi ui , v= βj v j .
i=1 j=1
Therefore,
 
I I i J
αi u i α
i=1 i u β
j=1 j
u i=1  
= J = 
v βj v j J j I
j=1 j=1 βj v i=1 αi
 J 
I
αi βj ui I
J
i=1 j=1 ui
= J = αi βj .
I
αi βj v j vj
i=1 j=1 i=1 j=1
Noting that
I
J I
J

αi βj = αi βj = 1,
i=1 j=1 i=1 j=1
we obtain (7.3.1). 2
Let D be the set of m × m diagonal matrices whose diagonal elements

are either 1 or 0. For example, if m = 2, then

0 0 0 0 1 0 1 0
D= , , , .
0 0 0 1 0 0 0 1
There are 2m elements in D. Suppose that each element of D is labeled as

Di , i = 1, 2, · · · , 2m . Then,

D = Di : i ∈ [1, 2m ] .
Denote Di− = I − Di . Clearly, Di− is also an element of D if Di ∈ D. Given

two vectors, u, v ∈ Rm ,

Di u + Di− v : i ∈ [1, 2m ]
is the set of vectors formed by choosing some elements from u and the rest
from v. Given two matrices F, H ∈ Rm×n ,

Di F + Di− H : i ∈ [1, 2m ]
is the set of matrices formed by choosing some rows from F and the rest
from H.
Lemma 7.3.2. Let u, v ∈ Rm ,

   
u1 v1
 u2   v2 
   
u= . , v= .. .
 ..   . 
um vm
Suppose that |vi | ≤ 1 for all i ∈ [1, m], then

sat(u) ∈ co Di u + Di− v : i ∈ [1, 2m ] .
Proof. Since |vi | ≤ 1, we have

sat(ui ) ∈ co ui , vi , ∀ i ∈ [1, m].
By applying Lemma 7.3.1 inductively, we have

sat(u1 ) ∈ co u1 , v1 ;
% &
u1 u1 u1 v1 v1
sat ∈ co , , , ;
u2 u2 v2 u2 v2
         
u1  u1 u1 u1 u1
sat  u2  ∈ co  u2 ,  u2 ,  v2 ,  v2 ,

u3 u3 v3 u3 v3
       
v1 v1 v1 v1 
 u2 ,  u2 ,  v2 ,  v2  ;

u3 v3 u3 v3
..
.
and finally,
sat(u) ∈ co Di u + Di− v : i ∈ [1, 2m ] .
2
Lemma 7.3.2 is illustrated in Fig. 7.3.1 for the case where m = 2.

Given two feedback matrices F, H ∈ Rm×n , suppose that |Hx|∞ ≤ 1,
then by Lemma 7.3.2, we have

sat(F x) ∈ co Di F x + Di− Hx : i ∈ [1, 2m ] .
In this way, we have placed sat(F x) into the convex hull of a group of linear
feedbacks.
7.4. Continuous-Time Systems under State Feedback 163
2
v u
[ u1] [ u1]
2 2
sat(u)
1
v u
[ v1] [ v 1]
0 2 2
−1
−2
−3
−3 −2 −1 0 1 2 3
7.4. Continuous-Time Systems under State Feedback

Consider an open-loop system
ẋ = Ax + Bu.
Under the saturated linear state feedback u = sat(F x), the closed-loop
system is
ẋ = Ax + Bsat(F x). (7.4.1)
For a matrix F ∈ Rm×n , denote the ith row of F as fi and define

L(F ) := x ∈ Rn : |fi x| ≤ 1, i = 1, 2, · · · , m .
If F is the feedback matrix, then L(F ) is the region where the feedback
control u = sat(F x) is linear in x. We call L(F ) the linear region of the
saturated feedback sat(F x), or simply, the linear region of saturation.
For x(0) = x0 ∈ Rn , denote the state trajectory of the system (7.4.1)
as ψ(t, x0 ). The domain of attraction of the origin is then given by

S := x0 ∈ Rn : lim ψ(t, x0 ) = 0 .
t→∞
A set is said to be invariant if all the trajectories starting from within it

will remain in it. Clearly, S is an invariant set.
Let P ∈ Rn×n be a positive-definite matrix. Denote

E(P, ρ) = x ∈ Rn : xT P x ≤ ρ .
Let V (x) = xT P x. The ellipsoid E(P, ρ) is said to be contractive invariant

if
V̇ (x) = 2xT P (Ax + Bsat(F x)) < 0
for all x ∈ E(P, ρ) \ {0}. Clearly, if E(P, ρ) is contractive invariant, then it

is inside the domain of attraction.
We will develop conditions under which E(P, ρ) is contractive invariant
and hence obtain an estimate of the domain of attraction.
7.4.1. A Set Invariance Condition Based on

Circle Criterion
A multivariable circle criterion is presented in [55, Theorem 10.1]. This

circle criterion is used in [32,81] to arrive at an estimate of the domain of
attraction for the system (7.4.1), with a given feedback gain matrix F .
Proposition 7.4.1 ([55,81]). Assume that (F, A, B) is controllable and

observable. Given an ellipsoid E(P, ρ), suppose that there exist positive
diagonal matrices K1 , K2 ∈ Rn×n with K1 < I, K2 ≥ I such that
(A + BK1 F )T P + P (A + BK1 F )
1
+ [F T (K2 − K1 ) + P B)][(K2 − K1 )F + B T P ] < 0, (7.4.2)
2
and E(P, ρ) ⊂ L(K1 F ). Then E(P, ρ) is a contractive invariant set and
hence inside the domain of attraction.
Remark 7.4.1. For a given sector bounded nonlinearity defined with K1

and K2 , the stability condition and the positive real condition in [55, The-
orem 10.1] is equivalent to the existence of a P > 0 such that (7.4.2) is
satisfied.
Here we simplified and restated the result in [55,81] to give conditions

for a given E(P, ρ) to be contractive invariant. If A + BF is Hurwitz, then
there are infinitely many E(P, ρ) satisfying the condition in Proposition
7.4.1. To estimate the domain of attraction, we may choose from all these
E(P, ρ) the “largest” one (the one with the largest volume or contains the
largest set with a fixed shape, etc.). A similar condition based on circle
criterion is given in [32]. These conditions are then used for stability and
performance analysis with LMI software in [32,81]. Since the inequality
(7.4.2) is not jointly convex in K1 and P , the two parameters need to be
optimized separately and iteratively. However, it is not guaranteed that
the global optimum can be obtained.
7.4.2. An Improved Condition for Set Invariance
For a given state feedback law u = sat(F x), we will develop a less con-
servative set invariance condition by exploring the special property of the
saturation nonlinearity. It is based on direct Lyapunov function analysis
in terms of an auxiliary feedback matrix H ∈ Rm×n . This condition turns
out to be equivalent to some LMIs. Denote the ith row of H as hi .
Theorem 7.4.1. Given an ellipsoid E(P, ρ), if there exists an H ∈ Rm×n

such that
(A + B(Di F + Di− H))T P + P (A + B(Di F + Di− H)) < 0,

∀ i ∈ [1, 2m ], (7.4.3)
and E(P, ρ) ⊂ L(H), i.e.,
|hi x| ≤ 1, ∀ x ∈ E(P, ρ), i = 1, 2, · · · , m,
then E(P, ρ) is a contractive invariant set and hence inside the domain of
attraction.
Proof. Let V (x) = xT P x, we need to show that
V̇ (x) = 2xT P (Ax + Bsat(F x)) < 0, ∀ x ∈ E(P, ρ) \ {0}. (7.4.4)
Since |hi x| ≤ 1 for all x ∈ E(P, ρ), i = 1, 2, · · · , m, by Lemma 7.3.2, for

every x ∈ E(P, ρ),

It follows that

Ax + Bsat(F x) ∈ co Ax + B(Di F + Di− H)x : i ∈ [1, 2m ] .
Therefore,
2xT P (Ax + Bsat(F x)) ≤ maxm 2xT P (A + B(Di F + Di− H))x,

i∈[1,2 ]
for every x ∈ E(P, ρ).

Since the condition (7.4.3) is satisfied, we have
max 2xT P (A + B(Di F + Di− H))x < 0

i∈[1,2m ]
for all x
= 0. Therefore, for every x ∈ E(P, ρ) \ {0},
V̇ (x) = 2xT P (Ax + Bsat(F x)) < 0.
This verifies (7.4.4). 2
If we restrict H to be K1 F , where K1 is a diagonal matrix and 0 <

K1 < I, then we have
Corollary 7.4.1. Given an ellipsoid E(P, ρ), if there exists a positive di-
agonal matrix K1 ∈ Rn×n , 0 < K1 < I, such that
(A + B(Di F + Di− K1 F ))T P + P (A + B(Di F + Di− K1 F )) < 0,

∀ i ∈ [1, 2m ], (7.4.5)
and E(P, ρ) ⊂ L(K1 F ), then E(P, ρ) is a contractive invariant set.
This is equivalent to Theorem 10.4 in [55] when applied to saturation

nonlinearity. Obviously, the condition in Corollary 7.4.1 is more conserva-
tive than that in Theorem 7.4.1 because the latter allows more freedom in
choosing the H matrix. However, it is implied in [55] that the condition
in Proposition 7.4.1 is even more conservative than that in Corollary 7.4.1.
Hence, the condition in Theorem 7.4.1 is the least conservative. Another
important advantage of Theorem 7.4.1 is that, when the estimation of do-
main of attraction is formulated into a constrained optimization problem,
the constraints are in the form of linear matrix inequalities while from
Proposition 7.4.1 and Corollary 7.4.1, we can only obtain bilinear matrix
inequalities. The formulation and solution of such an LMI constrained
optimization problem will be given in Section 7.4.4.
7.4.3. The Necessary and Sufficient Condition –

Single Input Systems
For single input systems (m = 1), Di = 0 (or 1). So the condition in

Theorem 7.4.1 for E(P, ρ) to be contractive invariant simplifies to: there
exists an H ∈ R1×n such that
(A + BF )T P + P (A + BF ) < 0,
(A + BH)T P + P (A + BH) < 0,

and E(P, ρ) ∈ L(H). This means that, if there exists another saturated
linear feedback u = sat(Hx), |Hx| ≤ 1 for all x ∈ E(P, ρ), that makes
E(P, ρ) contractive invariant, then E(P, ρ) is also contractive invariant under
u = sat(F x). In fact, we can go one step further to extend sat(Hx) to a
general nonlinear function h(x).
Theorem 7.4.2. Assume m = 1. Given an ellipsoid E(P, ρ), suppose that
(A + BF )T P + P (A + BF ) < 0. (7.4.6)
Then E(P, ρ) is contractive invariant under u = sat(F x) if and only if there

exists a function h(x) : Rn → R, |h(x)| ≤ 1 for all x ∈ E(P, ρ), such that
E(P, ρ) is contractive invariant under the control u = h(x), i.e.,
xT P (Ax + Bh(x)) < 0, ∀ x ∈ E(P, ρ) \ {0}. (7.4.7)
Proof. The “only if” part is obvious. Now we show the “if” part. Here we
have |h(x)| ≤ 1 for all x ∈ E(P, ρ). It follows from Lemma 7.3.2 that for
sat(F x) ∈ co F x, h(x) .
Hence,

xT P (Ax + Bsat(F x)) ≤ max xT P (Ax + Bh(x)), xT P (Ax + BF x) .
By (7.4.6) and (7.4.7), we obtain
xT P (Ax + Bsat(F x)) < 0, ∀ x ∈ E(P, ρ) \ {0}.
This shows that E(P, ρ) is contractive invariant under u = sat(F x). 2

Here we note that the condition (7.4.6) is necessary for E(P, ρ) to be
contractive invariant.
Theorem 7.4.2 implies that, for a single input system, the invariance of
an ellipsoid E(P, ρ) under a saturated linear control u = sat(F x) is in some
sense independent of F as long as the condition
(A + BF )T P + P (A + BF ) < 0
is satisfied. In other words, suppose that both F1 and F2 satisfy the condi-
tion
(A + BFi )T P + P (A + BFi ) < 0, i = 1, 2,
then, the maximal invariant ellipsoid E(P, ρ) (with ρ maximized) is the
same under either u = sat(F1 x) or u = sat(F2 x). This means that the
invariance of an ellipsoid depends on its shape rather than a particular
feedback. For a given P > 0, Theorem 11.2.1 of Chapter 11 will give a
way of determining the largest ρ such that E(P, ρ) can be made contractive
invariant with some control u = h(x), |h(x)| ≤ 1.
Example 7.4.1. Consider the closed-loop system (7.4.1) with

0.6 −0.8 2
A= , B= ,
0.8 0.6 4
and

F = 1.2231 −2.2486 .
The P matrix is given as

2.4628 −1.5372
P = .
−1.5372 1.3307
Combining Theorem 7.4.2 and Theorem 11.2.1, the largest contractive in-
variant ellipsoid is determined as E(P, ρ∗ ) with ρ∗ = 1.4696. Here we have
three other feedback matrices,

F1 = 1 −2 , F2 = 0.5 −3 , F3 = 0.3 −2.8 .
They all satisfy

(A + BFi )T P + P (A + BFi ) < 0.
Hence the largest contractive invariant ellipsoids under the feedbacks u =
sat(F x) and sat(Fi x), i = 1, 2, 3, are the same. Fig. 7.4.1 illustrates this
result. In the figure, the curves show V̇ (x) under different feedbacks along
the boundary of E(P, ρ∗ ), which is indicated by the angle of x ∈ ∂E(P, ρ∗ ).
As a comparison, we also plotted V̇ (x) under the bang-bang control u =
−sign(B T P x), which minimizes V̇ (x) (see Chapter 11).
F F2
3
−1 F1
−2
T
u=−sgn(B Px)
dV/dt
−3
−4
−5
−6
1.5 2 2.5 3 3.5 4 4.5 5
θ
Figure 7.4.1: V̇ (x) along the boundary of E(P, ρ∗ ).
7.4.4. Estimation of the Domain of Attraction
With all the ellipsoids satisfying the set invariance condition in Theo-
rem 7.4.1, we would like to choose from among them the “largest” one as
the least conservative estimate of the domain of attraction, i.e., we would
like to choose from all the E(P, ρ)’s that satisfy the set invariance condition
such that the quantity αR (E(P, ρ)) is maximized, where αR (E(P, ρ)) is the
size of E(P, ρ) with respect to some shape reference set X R . This problem
can be formulated as:
sup α (7.4.8)
P >0,ρ,H
s.t. a) αX R ⊂ E(P, ρ),
b) (A + B(Di F + Di− H))T P + P (A + B(Di F + Di− H)) < 0,
i ∈ [1, 2m ],
c) E(P, ρ) ⊂ L(H).
We define the supremum of α as α∗ .

Here we note that the above formulation can be easily extended to
the problem of maximizing the volume of E(P, ρ). If we replace α with
log det(P/ρ)−1 and remove constraint a), then we obtain the problem of
maximizing the volume of E(P, ρ). The reason for using log det(P/ρ)−1
instead of det(P/ρ)−1 is that the former is a convex function of P . Similar
modification can be made to other optimization problems to be formulated
in this chapter and Chapters 9 and 10. Moreover, the following proce-
dure to transform (7.4.8) into a convex optimization problem with LMI
constraints can be adapted to the corresponding volume maximization (or
minimization) problems.
To solve the optimization problem (7.4.8), we need to transform the set
inclusion constraints a) and c) into inequalities. If X R is a polyhedron,

X R = co x1 , x2 , · · · , xl ,
then constraint a) is equivalent to
α2 xTi P xi ≤ ρ, i = 1, 2, · · · , l. (7.4.9)
If X R is an ellipsoid

X R = x ∈ Rn : xT Rx ≤ 1 ,
then constraint a) is equivalent to,
α2 P ≤ ρR. (7.4.10)
On the other hand, constraint c) is equivalent to

min xT P x : hi x = 1 ≥ ρ, i = 1, 2, · · · , m. (7.4.11)
To see this, note that E(P, ρ) ⊂ L(H) if and only if all the hyperplanes
hi x = ±1, i = 1, 2, · · · , m, lie completely outside of the ellipsoid

E(P, ρ) = x ∈ Rn : xT P x ≤ ρ ,
i.e., at each point x on the hyperplanes hi x = ±1, we have xT P x ≥ ρ.

The left-hand side of (7.4.11) is a convex optimization problem and has
a unique minimum. By using the Lagrange multiplier method, we obtain

min xT P x : hi x = 1 = (hi P −1 hTi )−1 .
Consequently, constraint c) is equivalent to
ρhi P −1 hTi ≤ 1, i = 1, 2, · · · , m. (7.4.12)

Thus, if X R is a polyhedron, then (7.4.8) can be rewritten as follows,
sup α (7.4.13)
P >0,ρ,H
s.t. a1) α2 xTi P xi ≤ ρ, i = 1, 2, · · · , l,

i ∈ [1, 2m ],
c) ρhi P −1 hTi ≤ 1, i = 1, 2, · · · , m.
If X R is an ellipsoid, we just need to replace a1) with (7.4.10).

Constraints a), b) and c) are nonlinear and convex. The standard tool
to transform such constraints into LMI is Schur complements: Suppose
Q > 0, then
R S
≥0
ST Q
if and only if
R − SQ−1 S T ≥ 0.
If X R is a polyhedron, then by Schur complement, constraint a) is equivalent

to
% & 1
P 1 xTi
xiT
xi ≤ 2 ⇐⇒
α2 −1 ≥ 0, i = 1, 2, · · · , l. (7.4.14)
P
ρ α xi ρ
If X R is an ellipsoid, then constraint a) is equivalent to

1
α2 R I
−1
P ≥ 0. (7.4.15)
I ρ
Constraint b) is equivalent to
% &−1 % &−1
P T P
A + B(Di F + Di− H) + A + B(Di F + Di− H) < 0,
ρ ρ
i ∈ [1, 2m ]. (7.4.16)
And also by Schur complement, constraint c) is equivalent to

 −1 
P
 1 h i ρ 
 −1 −1  ≥ 0, i = 1, 2, · · · , m. (7.4.17)
P T P
ρ hi ρ
Let % &−1 % &−1

1 P P
γ= , Q= , Z=H .
α2 ρ ρ
Also let the ith row of Z be zi , i.e.,
% &−1
P
z i = hi .
ρ
Note that
(Di F + Di− H)Q = Di F Q + Di− HQ = Di F Q + Di− Z.
If X R is a polyhedron, then from (7.4.14), (7.4.16) and (7.4.17) the opti-

mization problem (7.4.8) can be rewritten as
inf γ (7.4.18)
Q>0,Z

γ xTi
s.t. a1) ≥ 0, i = 1, 2, · · · , l,
xi Q
b) QAT + AQ + (Di F Q + Di− Z)T B T + B(Di F Q + Di− Z) < 0,
i ∈ [1, 2m ],

1 zi
c) ≥ 0, i = 1, 2, · · · , m,
ziT Q
where all the constraints are given in LMIs.

If X R is an ellipsoid, we just need to replace a1) with

γR I
a2) ≥ 0.
I Q
Let the optimum of the optimization problem be γ ∗ , with the solution

Q and Z ∗ . Then, the optimal value of (7.4.8) is given by
∗
1
α∗ = (γ ∗ )− 2 ,
with H ∗ = Z ∗ (Q∗ )−1 and the resulting invariant set is E((Q∗ )−1 , 1). (Here,
we have, without loss of generality, let ρ = 1.)
We note that if E(Pi , ρi ), i ∈ [1, N ], are all contractive invariant, then
the union of these ellipsoids is also in the domain of attraction. Hence we
can obtain a better estimate by choosing different shape reference sets and
obtaining the union of the resulting invariant ellipsoids, as will be illustrated
in the following example.
7.5. Discrete-Time Systems under State Feedback 173
Example 7.4.2. We use an example of [81] to illustrate our results. The

system is described by (7.4.1) with

0 1 0
A= , B=
1 0 5
and the feedback gain is given by

F = −2 −1 .
Let
X R = co 0, x1

−1
with x1 = . We solve (7.4.18) and get α∗ = 4.3711. The maximal
0.8
ellipsoid is E(P ∗ , 1) with

0.1170 0.0627
P∗ = .
0.0627 0.0558
This ellipsoid is plotted in Fig. 7.4.2 in solid curve. The inner dashed
ellipsoid is an invariant set obtained by the circle criterion method in [81]
and the region bounded by the dash-dotted curve is obtained by the Popov
method, also in [81]. We see that both the regions obtained by the circle
criterion and by the Popov method can be actually enclosed in a single
invariant ellipsoid.
To arrive at an even better estimate, we vary x1 over a unit circle, and
solve (7.4.18) for each x1 . Let the optimal α be α∗ (x1 ). The union of all the
resulting ellipsoids gives a better estimate of the domain of attraction. The
outermost dotted boundary in Fig. 7.4.2 is formed by the points α∗ (x1 )x1
as x1 varies along the unit circle.
7.5. Discrete-Time Systems under State Feedback

7.5.1. Condition for Set Invariance
Consider an open loop system
x(k + 1) = Ax(k) + Bu(k)
under a saturated linear state feedback u = sat(F x). The closed-loop

system is
x(k + 1) = Ax(k) + Bsat(F x(k)). (7.5.1)
−2
−4
−6
−8
−5 −4 −3 −2 −1 0 1 2 3 4 5
Figure 7.4.2: The invariant sets obtained with different methods.
Denote the state trajectory starting from x(0) = x0 ∈ Rn as ψ(k, x0 ). The

domain of attraction of the origin is

S := x0 ∈ R : lim ψ(k, x0 ) = 0 .
n
k→∞
A set is said to be invariant if all the trajectories starting from within it

will remain in it. Also, it is easy to see that S is an invariant set.
Let P ∈ Rn×n be a positive definite matrix. Let V (x) = xT P x. The
ellipsoid E(P, ρ) is said to be contractive invariant if
∆V (x) := (Ax + Bsat(F x))T P (Ax + B sat(F x)) − xT P x < 0,
for all x ∈ E(P, ρ) \ {0}. Clearly, if E(P, ρ) is contractive invariant, then it

is inside the domain of attraction.
We are interested in knowing the conditions under which E(P, ρ) is con-
tractive invariant. The following proposition is a discrete-time counterpart
of Theorem 10.4 in [55] (when applied to saturation nonlinearities) and of
Corollary 7.4.1 in this chapter. It is also a special case of Theorem 7.5.1 to
be presented next.
Proposition 7.5.1. Given an ellipsoid E(P, ρ), if there exists a positive

diagonal matrix K ∈ Rm×m , K < I, such that,
T
A + B(Di F + Di− KF ) P A + B(Di F + Di− KF ) − P < 0,
∀ i ∈ [1, 2m ], (7.5.2)
and E(P, ρ) ⊂ L(KF ), then E(P, ρ) is a contractive invariant set.
To interpret Proposition 7.5.1, let us write
ui (k) = sat(fi x(k)) = γi (k)fi x(k),
where
sat(fi x(k))
γi (k) =
fi x(k)
denotes the saturation level. Here, we view γi (k) as the varying gain of each
control channel that takes value between K(i, i) and 1. Then, the quadratic
stability (within E(P, ρ)) of the system is guaranteed by the quadratic sta-
bility of the linear systems corresponding to the 2m vertices of the box of
varying gains, which are
γi = K(i, i) or 1, i = 1, 2, · · · , m.
And each vertex in turn specifies a linear feedback matrix in the set

Di F + Di− KF : i ∈ [1, 2m ] .
Using this idea, Proposition 7.5.1 can be easily shown with some standard
techniques in robustness analysis.
Since K < I, we see that L(F ) is in the interior of L(KF ). Hence, the
condition E(P, ρ) ⊂ L(KF ) allows E(P, ρ) to go beyond the linear region
of the saturation function sat(F x).
Similar to the continuous-time case, we have the following less conser-
vative criterion for an ellipsoid to be contractive invariant.
Theorem 7.5.1. Given an ellipsoid E(P, ρ), if there exists an H ∈ Rm×n

such that
(A + B(Di F + Di− H))T P (A + B(Di F + Di− H)) − P < 0,

∀ i ∈ [1, 2m ], (7.5.3)
and E(P, ρ) ⊂ L(H), i.e.,
|hi x| ≤ 1, ∀ x ∈ E(P, ρ), i = 1, 2, · · · , m,
then, E(P, ρ) is a contractive invariant set.
Proof. Let V (x) = xT P x, we need to show that
∆V (x) = (Ax + Bsat(F x))T P (Ax + Bsat(F x)) − xT P x < 0,

∀ x ∈ E(P, ρ) \ {0}. (7.5.4)
Since |hi x| ≤ 1 for all x ∈ E(P, ρ), i = 1, 2, · · · , m, by Lemma 7.3.2, for


It follows that

Ax + Bsat(F x) ∈ co Ax + B(Di F + Di− H)x : i ∈ [1, 2m ] .
By the convexity of the function V (x) = xT P x, we have
(Ax + Bsat(F x))T P (Ax + Bsat(F x))

≤ maxm xT (A + B(Di F + Di− H))T P (A + B(Di F + Di− H))x,
i∈[1,2 ]
for every x ∈ E(P, ρ).

Since the condition (7.5.3) is satisfied, we have
max xT (A + B(Di F + Di− H))T P (A + B(Di F + Di− H))x < xT P x,

i∈[1,2m ]
for all x
= 0. Therefore, for every x ∈ E(P, ρ) \ {0},
(Ax + Bsat(F x))T P (Ax + Bsat(F x)) < xT P x.
This verifies (7.5.4). 2
We see that Proposition 7.5.1 is a special case of Theorem 7.5.1 by

setting H = KF . So the condition in Theorem 7.5.1 is less conservative
than that in Proposition 7.5.1. This will be illustrated in Example 7.5.1.
7.5.2. The Necessary and Sufficient Condition –

Single Input Systems
For single input systems (m = 1), Di = 0 (or 1). So the condition in

Theorem 7.5.1 for E(P, ρ) to be contractive invariant simplifies to: there
exists an H ∈ R1×n such that
(A + BF )T P (A + BF ) − P < 0,
(A + BH)T P (A + BH) − P < 0,
and E(P, ρ) ∈ L(H). The following theorem is a discrete-time counter-

part to Theorem 7.4.2 and can be proved in a similar way by utilizing the
convexity of the function V (x) = xT P x.
Theorem 7.5.2. Assume m = 1. Given an ellipsoid E(P, ρ), suppose that
(A + BF )T P (A + BF ) − P < 0. (7.5.5)
Then, E(P, ρ) is contractive invariant under u = sat(F x) if and only if there

exists a function h(x) : Rn → R, |h(x)| ≤ 1 for all x ∈ E(P, ρ), such that
E(P, ρ) is contractive invariant under the control u = h(x), i.e.,
(Ax + Bh(x))T P (Ax + Bh(x)) − xT P x < 0, ∀ x ∈ E(P, ρ) \ {0}. (7.5.6)
The condition in Theorem 7.5.2 can be checked with Theorem 11.3.1

of Chapter 11, which also provides a method for determining the largest ρ
such that E(P, ρ) is contractive invariant.
Example 7.5.1. Consider the closed-loop system (7.5.1) with

0.8876 −0.5555 −0.1124
A= , B= ,
0.5555 1.5542 0.5555
and

F = −0.7651 −2.0299 .
Given
5.0127 −0.6475
P = ,
−0.6475 4.2135
we would like to find the maximal ρ such that E(P, ρ) is contractive in-
variant. By applying Theorem 7.5.2 and Theorem 11.3.1 with a bisection
search, the maximal E(P, ρ) is E(P, ρ∗ ) with ρ∗ = 2.349.
Let us compare the largest invariant ellipsoid, E(P, ρ∗ ), with estimates
obtained by other methods.
• The maximal ρ such that E(P, ρ) ⊂ L(F ) is ρ1 = 0.8237;
• The maximal ρ satisfying the condition in Proposition 7.5.1 is ρ2 =

1.0710;
• The maximal ρ satisfying the condition in Theorem 7.5.1 is ρ3 = ρ∗ =

2.3490, with

H = −0.1389 −1.3018 .
Shown in Fig. 7.5.1 is a comparison of the estimates of the invariant

ellipsoids obtained with different methods. It is interesting to note that
ρ3 = ρ∗ . In this case, the largest estimate of the invariant ellipsoid obtained
by Theorem 7.5.1 is not conservative at all.
Hx=−1
0.8
Fx=−1
0.6
0.4
0.2
0
x’Px=ρ
Fx=1 1
−0.2
−0.4 x’Px=ρ x’Px=ρ =ρ*

2 3
−0.6
Hx=1
−0.8
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Figure 7.5.1: Different estimates of the domain of attraction.

7.5.3. Estimation of the Domain of Attraction
Given a shape reference set X R , the problem of maximizing the contractive

invariant ellipsoid with respect to X R can be formulated as:
sup α (7.5.7)
P >0,ρ,H
s.t. a) αX R ⊂ E(P, ρ),
b) (A + B(Di F + Di− H))T P (A + B(Di F + Di− H)) − P ≤ 0,
i ∈ [1, 2m ],
c) E(P, ρ) ⊂ L(H).
Like in the continuous-time case, the above constraints can be trans-

formed into LMIs. Since constraints a) and c) are the same as those in
(7.4.8), we only deal with constraint b). By Schur complement, constraint
b) is equivalent to

P (A + B(Di F + Di− H))T
≥ 0, i ∈ [1, 2m ].
A + B(Di F + Di− H) P −1
Multiplying both sides from left and right with

−1
P 0
,
0 I
we have
 −1 −1 
P P − T
 ρ ρ (A+B(Di F +Di H)) 
 −1 −1  ≥ 0,
(A+B(Di F +Di− H)) P
ρ
P
ρ
i ∈ [1, 2m ]. (7.5.8)
Let % &−1 % &−1

1 P P
γ = 2, Q= , Z=H .
α ρ ρ
Also let the ith row of Z be zi , i.e.,
% &−1
P
z i = hi .
ρ
If X R is a polyhedron, then the optimization problem (7.5.7) can be rewrit-

ten as
inf γ (7.5.9)
Q>0,Z

γ xTi
s.t. a1) ≥ 0, i = 1, 2, · · · , l,
xi Q

Q (AQ + B(Di F Q + Di− Z))T
b) ≥ 0,
AQ + B(Di F Q + Di− Z) Q
i ∈ [1, 2m ],

1 zi
c) ≥ 0, i = 1, 2, · · · , m.
ziT Q
where all the constraints are given in LMIs.

If X R is an ellipsoid, we just need to replace a1) with

γR I
a2) ≥ 0.
I Q
7.6. Extension to Output Feedback

Consider the continuous-time system
ẋ = Ax + B sat(u),
y = Cx,
under the output feedback control,
ż = Ac z + Bc y,
u = Cc z + Dc y.
The closed-loop system has the following state equation

% &
ẋ x x
=A + B sat F , (7.6.1)
ż z z
where

A 0 B
A= , B= , F = Dc C Cc .
Bc C Ac 0
We see that the system (7.6.1) has the form of (7.4.1). Hence all the
methods developed in the previous sections can be utilized to estimate
the domain of attraction of (7.6.1). Very often, the initial condition z(0)
of the output feedback controller is set to 0. This information can be
utilized in choosing the shape reference set to reduce the conservatism of
the estimation. For example, if X R is a polyhedron, its vertices can be
chosen of the form (xi , 0). In this way, X R is in a subspace corresponding
to the x state. If X R is an ellipsoid, then we can choose

R1 0
R= ,
0 R2
with the elements of R2 much larger than those of R1 so that the ellipsoid
is very thin along the direction of the z state.
When an invariant ellipsoid E(P, ρ) is obtained as an estimation of the
domain of attraction, we may restrict out attention to only a subset of
E(P, ρ), namely, its intersection with the subspace z = 0. Suppose that

P1 P2
P = ,
P2T P3
with P1 ∈ Rn×n , then the intersection of E(P, ρ) with z = 0 is

x
: xT P1 x ≤ ρ .
0
The above discussion also applies to discrete-time systems.
7.7. Conclusions
In this chapter, we have considered the problem of estimating the domain of
attraction for a linear system under a pre-designed saturated linear feedback
law. We used ellipsoids as our estimates of the domain of attraction. A
simple condition was derived in terms of an auxiliary feedback matrix for
determining if a given ellipsoid is contractive invariant. This condition was
shown to be less conservative than the existing conditions which are based
on the circle criterion or the vertex analysis. An important feature of this
new condition is that it can be expressed as LMIs in terms of all the varying
parameters. This makes the problem of maximizing the invariant ellipsoid
with respect to some shape reference set or the volume of the invariant
ellipsoid very easy. With a little modification, this analysis problem can
be turned into a controller design problem, as will be seen in the next
chapter. Moreover, the results in this chapter will be further extended
in Chapter 10 to determine the invariant sets for systems with persistent

disturbances and to design controllers that achieve disturbance rejection
with guaranteed stability requirements.
Chapter 8
On Enlarging the Domain

of Attraction
8.1. Introduction
In this chapter, we will present a method for designing feedback gains that
result in large domains of attraction. Our approach is to formulate the
problem into a constrained optimization problem. Since the precise do-
main of attraction under a feedback law is hard to identify, we will first
obtain invariant ellipsoids as estimates of the domain of attraction and
then maximize the estimate over stabilizing feedback laws.
In solving the optimization problem, we will also reveal a surprising
aspect of the design for large domain of attraction. If our purpose is solely
to enlarge the domain of attraction, we might as well restrict the invariant
ellipsoid (an estimate of the domain of attraction) in the linear region of
the saturation function, although allowing saturation will provide us more
freedom in choosing controllers. Another interesting aspect is that, for a
discrete-time system, the domain of attraction can be further enlarged if
the design is performed on its lifted system.
8.2. Continuous-Time Systems
Consider the open-loop system
ẋ = Ax + Bu, x ∈ Rn , u ∈ Rm . (8.2.1)
183
184 Chapter 8. On Enlarging the Domain of Attraction
system is given by
ẋ = Ax + Bsat(F x). (8.2.2)
For a fixed F , Chapter 7 provides a method for estimating the domain of
attraction of the origin for the system (8.2.2) by searching for the largest
invariant ellipsoid. In this section, we will design a feedback matrix F
such that this estimate is maximized with respect to a given reference set
X R . By applying Theorem 7.4.1, this optimization problem can be readily
formulated as follows,
sup α (8.2.3)
P >0,ρ,F,H
s.t. a) αX R ⊂ E(P, ρ),
i ∈ [1, 2m ],
c) E(P, ρ) ⊂ L(H).
The only difference of the optimization problem (8.2.3) from (7.4.8) is
that (8.2.3) has an extra optimization parameter F . Denote the supremum
of α as α∗1 .
Let us consider a simpler optimization problem
sup α (8.2.4)
P >0,ρ,F
s.t. a) αX R ⊂ E(P, ρ),
b) (A + BF )T P + P (A + BF ) < 0,
c) E(P, ρ) ⊂ L(F ).
Denote the supremum of α in (8.2.4) as α∗2 . Recalling that Di + Di− = I,
we can view (8.2.4) as a result from forcing F = H in (8.2.3). It follows
that
α∗2 ≤ α∗1 .
What is somewhat surprising is that,
α∗1 = α∗2 .
This can be seen as follows. Suppose that (α1 , ρ1 , P1 , F1 , H1 ) satisfies the
constraints of (8.2.3). By choosing Di = 0, then we obtain from b) of
(8.2.3) the following inequality,
(A + BH1 )T P + P (A + BH1 ) < 0.
8.3. Discrete-Time Systems 185
By setting (α, ρ, P, F ) = (α1 , ρ1 , P1 , H1 ), then clearly (α, ρ, P, F ) satisfies

the constraints of (8.2.4). This shows that α∗1 ≤ α∗2 . Hence we conclude
that α∗1 = α∗2 .
From the above analysis, we see that, if our only purpose is to enlarge
the domain of attraction, we might as well solve the simpler optimization
problem (8.2.4). The resulting invariant ellipsoid E(P, ρ) will lie completely
inside the linear region of the saturation function sat(F x). If other per-
formance requirements are involved, we may set H in (8.2.3) b) equal to
the optimal value of (8.2.4), say F2∗ , to guarantee a known domain of at-
traction, and choose from all the F satisfying (8.2.3) b) to optimize other
performances, such as convergence rate and disturbance rejection.
Both optimization problems (8.2.3) and (8.2.4) can be solved with LMI
method as (7.4.8) in Chapter 7. For example, (8.2.3) can be transformed
into an LMI problem similar to (7.4.18) with F Q replaced by a free param-
eter Y . It should be noted that as the domain of attraction is enlarged by
solving (8.2.4), some eigenvalues of A + BF will get very close to the imag-
inary axis, indicating a low convergence rate. The problem of maximizing
the convergence rate inside a given ellipsoid will be discussed in Chapter 11.
8.3. Discrete-Time Systems

x(k + 1) = Ax(k) + Bu(k), x ∈ Rn , u ∈ Rm . (8.3.1)
system is given by
x(k + 1) = Ax(k) + Bsat(F x(k)). (8.3.2)
Similar to the continuous-time case, we can solve the following optimization

problem to enlarge the domain of attraction,
sup α (8.3.3)
P >0,ρ,F
s.t. a) αX R ⊂ E(P, ρ),
b) (A + BF )T P (A + BF ) − P < 0,
c) E(P, ρ) ⊂ L(F ).
In what follows, we will present a method for further enlargement of

the domain of attraction by using the lifting technique. The motivation of
using the lifting technique is as follows: by solving (8.3.3), we obtain an

invariant set E(P, ρ) where the trajectory of the system is kept inside at
every step. It would be less restrictive if we only require the state to be
inside E(P, ρ) at every L steps, say, 0, L, 2L, · · ·. This may result in a larger
E(P, ρ). Let L be a positive integer. Denoting
AL = AL , BL = [ AL−1 B AL−2 B ···B]
and  
u(kL)
 u(kL + 1) 
 
xL (k) = x(kL), uL (k) =  .. ,
 . 
u(kL + L − 1)
we obtain the lifted L-step system
xL (k + 1) = AL xL (k) + BL uL (k), xL ∈ Rn , uL ∈ RmL . (8.3.4)
Let
uL (k) = sat(FL xL (k)), FL ∈ RmL×n ,
be a stabilizing feedback. Under this feedback law, the closed-loop system
is given by
xL (k + 1) = AL xL (k) + BL sat(FL xL (k)). (8.3.5)
Note that the control
uL (k) = sat(FL xL (k))
is periodic for the original system (8.3.1) with period L. Under this control,
if E(P, ρ) is an invariant set for the original unlifted system, then it is also
invariant for the lifted system. But an invariant set E(P, ρ) for the lifted
system need not be invariant for the state of the original system. However,
the domain of attraction is the same for both the lifted system and the
original system. This can be seen as follows. Clearly, x(k) → 0 implies
xL (k) → 0. On the other hand, if xL (k) → 0, then uL (k) → 0 and x(kL+j),
j = 1, 2, · · · , L, will be arbitrarily close to x(kL) = xL (k). Hence, we also
have x(k) → 0. Because of this, we can enlarge the domain of attraction
by enlarging the invariant ellipsoid for the lifted closed-loop system.
Similar to the one-step case, the problem of maximizing the invariant
ellipsoid can be described as
sup α (8.3.6)
P >0,FL
s.t. a) αX R ⊂ E(P, ρ),

b) (AL + BL FL )T P (AL + BL FL ) − P < 0,
c) E(P, ρ) ⊂ L(FL ),
which can also be solved by the LMI approach proposed in Chapter 7.

Denoting the supremum of α corresponding to an L step lifting as α∗ (L),
we have the following theorem that justifies the use of lifting technique.
Theorem 8.3.1. For any integers p, L ≥ 1, α∗ (p) ≤ α∗ (pL).
Proof.
Case 1. p = 1
Denote the set of feasible (α, P ) satisfying constraints a), b) and c) as

Φ(L) = (α, P ) : there exists an FL s.t. a), b) and c) are satisfied .
It suffices to show that Φ(1) ⊂ Φ(L).

Suppose that (α, P ) ∈ Φ(1), then there exists an F ∈ Rm×n , with its
ith row labeled as fi , such that
fi P −1 fiT ≤ ρ−1 , i = 1, 2, · · · , m, (8.3.7)
and
(A + BF )T P (A + BF ) − P < 0, (8.3.8)
which is equivalent to

P (A + BF )T
> 0,
A + BF P −1
and to
(A + BF )P −1 (A + BF )T − P −1 < 0. (8.3.9)
Let  
F
 F (A + BF ) 
 
FL =  .. ,
 . 
F (A + BF )L−1
then,
AL + BL FL = AL + AL−1 BF + AL−2 BF (A + BF ) + · · ·
+BF (A + BF )L−1
= (A + BF )L .
It then follows from (8.3.8) that
(AL + BL FL )T P (AL + BL FL ) = ((A + BF )T )L P (A + BF )L

< ((A + BF )T )L−1 P (A + BF )L−1
< ···
< P,
which shows that P and FL satisfy constraint b).

Denote the jth row of FL as fL j . Since fL j = fi (A + BF )q for some
i ≤ m, q ≤ L − 1, we have
fL j P −1 fL Tj = fi (A + BF )q P −1 ((A + BF )T )q fiT .
It follows from (8.3.7) and (8.3.9) that
fL j P −1 fL Tj ≤ fi (A + BF )q−1 P −1 ((A + BF )T )q−1 fiT

≤ ···
≤ fi P −1 fiT
≤ ρ−1 ,
which shows that P and FL also satisfy constraint c). Hence, (α, P ) ∈ Φ(L).
Case 2. p > 1
Let
Ap = Ap , Bp = [ Ap−1 B Ap−2 B ···B]
and
Ap L = ApL , Bp L = [ ApL−1 B ApL−2 B · · · B ],
then,
Ap L = (Ap )L , Bp L = [ AL−1
p Bp AL−2
p Bp · · · Bp ].
Suppose we first lift the system (8.2.1) with step p to get xp (k) = x(kp)
and
xp (k + 1) = Ap xp (k) + Bp sat(up (k)),
and then lift the above system with step L to get xp L (k) = xp (kL) = x(kLp)
and
xp L (k + 1) = Ap L xp L (k) + Bp L sat(up L (k)).
Applying the result in Case 1, we immediately have α∗ (p) ≤ α∗ (pL). 2
Remark 8.3.1. The equality α∗ (p) = α∗ (pL) with L > 1 can occur in
some special cases. For example, let A = a > 1, B = 1, and X R = [−1, 1].
1
α∗ (L) = , ∀ L ≥ 1.
a−1
From the above theorem, we see that
α∗ (1) ≤ α∗ (2) ≤ α∗ (4) ≤ α∗ (8) · · · ,
and
α∗ (1) ≤ α∗ (3) ≤ α∗ (6) ≤ α∗ (12) · · · .
But α∗ (L1 ) ≤ α∗ (L2 ) does not necessarily hold for all L1 < L2 .
Example 8.3.1. Consider a second order system in the form of (8.3.1)

with
0.9510 0.5408 0.0980
A= , B= .
−0.2704 1.7622 0.5408
The matrix A has two unstable eigenvalues {1.2214, 1.4918}. The shape
reference set
X R = x ∈ R2 : xT Rx ≤ 1 ,
with
1.2862 −1.0310
R= ,
−1.0310 4.7138
is chosen according to the shape of the null controllable region. Table 8.3.1
shows the computational result for α∗ (L), L = 1, 2, 4, 8, 16, 32.
L 1 2 4 8 16 32
α∗ (L) 1.0650 1.0930 1.1896 1.4017 1.5164 1.5426
Table 8.3.1: The increase of α∗ (L).
Fig. 8.3.1 demonstrates the effectiveness of the lifting technique. The

innermost curve is the boundary of α∗ (1)X R . For L = 2, 4, 8, 16, 32, the
set α∗ (L)X R grows bigger as L is increased. The outermost curve is the
boundary of the null controllable region obtained by the method presented
in Chapter 3. We see that the increase from α∗ (16) to α∗ (32) is small so
we take L = 16 as the lifting step. The optimal Po corresponding to α∗ (16)
is
0.5593 −0.4483 1
Po = = 0.4348R = ∗ R.
−0.4483 2.0497 (α (16))2
0.8
0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
−1
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Figure 8.3.1: The sets α∗ (L)X R .
The optimal feedback gain matrix is,

0.3504 0.4636 0.6129 0.7324 0.7279
F̄oT =
−1.4294 −1.3917 −1.2360 −0.8490 −0.2872
0.6374 0.5467 0.4777 0.4279 0.3918 0.3653

0.1679 0.4481 0.6167 0.7225 0.7924 0.8406

0.3454 0.3302 0.3185 0.3094 0.3021
.
0.8750 0.9003 0.9193 0.9337 0.9447
The eigenvalues of (Ā + B̄ F̄o )T Po (Ā + B̄ F̄o ) − Po are −0.0006 and −0.4293.
The eigenvalues of Ā + B̄ F̄o are 0.2758 ± j0.8814. These show that the
convergence rate is slow.
Example 8.3.2. A third order system is described by (8.3.1) with

   
1.1972 1.0775 0 1.4431
A= 0 1.1972 0  , B =  0.9861  .
0 0 1.4333 1.0833
All of the eigenvalues of A are unstable. For the purpose of comparison,
we choose 18 points on the boundary of the null controllable region as the
vertices of X R (see Fig. 8.3.2, where the vertices of X R are marked with
“∗” and the vertices of the null controllable region are marked with “·”).
5 3
x2
x3
0 0
−1
−2
−5 −3
−40 −20 0 20 40 −5 0 5
x x
1 2
2 5
1
0
x3
x3
−1
−5
5
−2 50
0
0
−3
−40 −20 0 20 40 x −5 −50 x
2 1
x1
Figure 8.3.2: The vertices of X R .
Table 8.3.2 shows the computational result for α∗ (L), L = 1, 2, 4, 8,

16, 32.
L 1 2 4 8 16 32
α∗ (L) 0.4274 0.4382 0.4593 0.4868 0.5564 0.6041
Table 8.3.2: The increase of α∗ (L).
We also note that α∗ (L) increases significantly as L increases. This

can be seen in Figs. 8.3.3 and 8.3.4, where the vertices of α∗ (1)X R and
α∗ (32)X R are shown in comparison with the null controllable region.
8.4. Conclusions
We presented some design methods for enlarging the domain of attraction

with saturated linear feedback. Examples were worked out in detail to illus-
5 3
x2
x3
0 0
−1
−2
−5 −3
−40 −20 0 20 40 −5 0 5
x x
1 2
2 5
1
0
x3
x3
−1
−5
5
−2 50
0
0
−3
−40 −20 0 20 40 x2 −5 −50 x1
x1
Figure 8.3.3: α∗ (1)X R and the null controllable region.
5 3
1
x2
x3
0 0
−1
−2
−5 −3
−40 −20 0 20 40 −5 0 5
x1 x2
2 5
1
0
x3
x3
−1
−5
5
−2 50
0
0
−3
−40 −20 0 20 40 x2 −5 −50 x1
x1
Figure 8.3.4: α∗ (32)X R and the null controllable region.

trate the effectiveness of the design method. Also shown in the examples is,
however, the slow convergence rate of the resulting feedback system. This
issue will be discussed in detail in Chapter 11.
Chapter 9
Semi-Global Stabilization
with Guaranteed
Regional Performance
9.1. Introduction
We revisit the problem of semi-globally stabilizing a linear system on its

null controllable region with saturating actuators. This problem has been
solved for single input systems with one or two anti-stable poles in both
the continuous-time and discrete-time settings in Chapters 4 and 5, respec-
tively. In this chapter, we will consider more general systems, possibly
multiple input and with more than two anti-stable poles.
We first consider a linear discrete-time system subject to actuator sat-
uration,
x(k + 1) = Ax(k) + Bsat(u(k)), x ∈ Rn , u ∈ Rm . (9.1.1)
Note that the saturation function sat : Rm → Rm represents the con-

straints imposed by the saturating actuators. Assume that a feedback law
u = F0 (x) has been designed such that the resulting closed-loop system in
the absence of the actuator saturation
x(k + 1) = Ax(k) + BF0 (x(k)) (9.1.2)
195
196 Chapter 9. Semi-Global Stabilization with Regional Performance
has the desired performance. We would like to study the stability and
performance of the system in the presence of actuator saturation,
x(k + 1) = Ax(k) + Bsat(F0 (x(k))). (9.1.3)
Let D0 be an invariant set of the closed-loop system and be inside the linear
region of the saturation function, i.e.,

D 0 ⊂ x ∈ Rn : |F0 (x)|∞ ≤ 1 .
For example, a linear state feedback law u = F0 x could be constructed that

places the closed-loop poles at certain desired locations and D0 could be
an ellipsoid of the form

E(P0 ) = x ∈ Rn : xT P0 x ≤ 1 ,
where P0 > 0 satisfies
(A + BF0 )T P0 (A + BF0 ) − P0 < 0. (9.1.4)
Suppose that D 0 is in the linear region, then it is an invariant set and within
D0 , the saturation function does not have an effect and hence the desired
closed-loop performance is preserved.
The objective of this chapter is to construct feedback laws that semi-
globally stabilize the system (9.1.1) on its null controllable region and in
the mean time preserve the desired closed-loop performance in the region
D0 . We will design our controllers by combining a sequence of feedback
laws
u = Fi (x), i = 0, 1, · · · , M,
in such a way that the union of the invariant sets corresponding to each
of the feedback laws is also an invariant set, which is shown to be in the
domain of attraction. By appropriately selecting this sequence of feedback
laws, the union of these invariant sets can then be made large enough to
include any subset in the interior of the null controllable region. This idea
will be made feasible by the use of lifting technique.
This chapter is organized as follows. In Section 9.2, we propose a method
for expanding the domain of attraction by switching among a finite sequence
of feedback laws. This switching design is then used in Section 9.3 to show
that the domain of attraction can be enlarged to include any subset in the
interior of the null controllable region. Section 9.4 extends the results of
9.2. Expansion of the Domain of Attraction 197
Section 9.3 to continuous-time systems. An example is given in Section 9.5

to illustrate our design method. Finally, a brief concluding remark is made
in Section 9.6.
9.2. Expansion of the Domain of Attraction

Let
u = Fi (x), i = 0, 1, · · · M,
be a finite sequence of stabilizing feedback laws. Among these feedback
laws, u = F0 (x) can be viewed as the one that was originally designed to
guarantee certain desired closed-loop performance in a given region and the
remaining feedback laws have been introduced for the purpose of enlarging
the domain of attraction while preserving the regional performance of the
original feedback law u = F0 (x).
For each i = 0, 1, · · · , M , let D i be an invariant set inside the domain
of attraction of the equilibrium x = 0 of the closed-loop system under the
feedback law u = Fi (x),
x(k + 1) = Ax(k) + B sat(Fi (x)). (9.2.1)
Denote

i
Ωi = Dj , i = 0, 1, · · · M.
j=0
Then, Ω0 ⊂ Ω1 ⊂ · · · ⊂ ΩM .
Theorem 9.2.1. For each i = 0, 1, · · · , M , Ωi is an invariant set inside the

domain of attraction of x = 0 of the closed-loop system
x(k + 1) = Ax(k) + Bsat(Gi (x(k)), (9.2.2)
where 

 F0 (x), if x ∈ Ω0 ,


 F (x), if x ∈ Ω1 \ Ω0 ,
1
Gi (x) := (9.2.3)


.. ..

 . .

Fi (x), if x ∈ Ωi \ Ωi−1 .
Here we note that, for each i = 1, 2, · · · , M ,

i−1
Ωi \ Ωi−1 = Di \ Dj .
j=0
Proof of Theorem 9.2.1. We prove the theorem by induction. The

statement is trivially true for i = 0. Now suppose it is true for i ≥ 0, we
need to show that it is also true for i + 1.
Now assume that Ωi is an invariant set inside the domain of attraction
under the control u = Gi (x). We can write Gi+1 (x) as

Gi (x), if x ∈ Ωi ,
Gi+1 (x) = (9.2.4)
Fi+1 (x), if x ∈ Ωi+1 \ Ωi .
If x(0) ∈ Ωi , then under the feedback u = Gi (x), x(k) ∈ Ωi for all k and
lim x(k) = 0,
k→∞
because Ωi is an invariant set inside the domain of attraction under u =

Gi (x).
On the other hand, suppose x(0) ∈ Ωi+1 \ Ωi = D i+1 \ Ωi . Since
Di+1 is an invariant set inside the domain of attraction under the feedback
u = Fi+1 (x) and Ωi is a neighborhood of the origin, x(k) will enter Ωi at
some k1 < ∞. After that, the control is switched to u = Gi (x) and by the
foregoing argument, we also have
lim x(k) = 0.
k→∞
This shows that Ωi+1 is inside the domain of attraction.

In the first case above, x(k) ∈ Ωi ⊂ Ωi+1 for all k ≥ 0, and in the second
*
case, x(k) ∈ D i+1 Ωi = Ωi+1 for all k ≥ 0. Therefore, Ωi+1 is also an
invariant set under u = Gi+1 (x). 2
From (9.2.3), we see that if x ∈ Ω0 = D0 , then u = F0 (x) is in effect

and hence the pre-designed performance is guaranteed on D0 .
For later use in Section 9.4, it can be verified in a similar way that
Theorem 9.2.1 is also true for a continuous-time system
ẋ = f (x, u),
in particular,
ẋ = Ax + Bsat(u), (9.2.5)
with a sequence of stabilizing feedback laws
u = Fi (x), i = 0, 1, · · · , M.
9.3. Semi-Globalization – Discrete-Time Systems 199
In the context of continuous-time systems, the existence and uniqueness of

the solution of the closed-loop system is guaranteed by the fact that Ωi ’s
are invariant sets and nested to each other. In other words, a trajectory
starting from a set Ωi will remain in it. Once it enters a smaller set Ωj ,
j < i, it will again remain in it.
9.3. Semi-Globalization – Discrete-Time Systems

In this section, we will utilize the lifting technique to design a sequence of
ellipsoids that cover any prescribed compact subset of the null controllable
region. Each ellipsoid is invariant and in the domain of attraction for the
lifted closed-loop system under an appropriately chosen linear feedback.
This, by Theorem 9.2.1, would achieve semi-global stabilization for the
lifted system, and hence for the original system. We thus refer to such a
design process as semi-globalization.
Recall from Chapter 3 that the null controllable region of (9.1.1) at step
K, denoted as C(K), is the set of states that can be steered to the origin
in K steps. An initial state x0 ∈ C(K) if and only if there exists a control
u satisfying |u(k)|∞ ≤ 1, k = 0, 1, · · · , K − 1, such that

K−1
AK x0 + AK−i−1 Bu(i) = 0. (9.3.1)
i=0
The null controllable region, denoted as C, is the set of states that can be
steered to the origin in a finite number of steps. We also have

C= C(K).
K≥0
Also, it can be shown by standard analysis that any compact subset of C is

a subset of C(K) for some K. For simplicity, we assume that the pair (A, B)
is controllable and A is nonsingular (otherwise the order of the system can
be reduced to make it so). Then there is an integer n0 ≤ n such that the
matrix
AK−1 B AK−2 B · · · , AB B
has full row rank for all K ≥ n0 . It follows from (9.3.1) that C(K) contains
the origin in its interior and is bounded for all K ≥ n0 .
For a positive integer L, the lifted system of (9.1.1) with a lifting step
L is given by
xL (k + 1) = AL xL (k) + BL sat(uL (k)), (9.3.2)
where  
u(kL)
xL (k) = x(kL), uL (k) =  u(kL + 1)  ,
u(kL + L − 1)
and
AL = AL , BL = [AL−1 B AL−2 B · · · AB B]. (9.3.3)
We have more flexibility in the design of a system by using the lifting
technique because it allows us to see further the effect of a control law and
to consider the combined effect of the control actions at several steps.
For a feedback matrix F ∈ RmL×n , denote the unsaturated region (lin-
ear region) of the closed-loop system
xL (k + 1) = AL xL (k) + BL sat(F xL (k)) (9.3.4)
as
L(F ) := x ∈ Rn : |fj x| ≤ 1, j = 1, 2, · · · , mL ,
where fj is the jth row of F . If L ≥ n0 , then there exists an F such that
AL + BL F = 0.
For such an F , there is a corresponding L(F ) and for all xL 0 = x0 ∈ L(F ),
AL x0 + BL sat(F x0 ) = (AL + BL F )x0 = 0.
Hence L(F ) is an invariant set of the lifted system (9.3.4) and is inside the
domain of attraction.
For a positive definite matrix P ∈ Rn×n , denote

E(P ) = x ∈ Rn : xT P x ≤ 1 .
Suppose that E(P ) ⊂ L(F ), then under the feedback law uL = F xL , E(P ) is
also an invariant set inside the domain of attraction. Here we are interested
in the ellipsoids because they can be generalized to the Lyapunov level sets
for the case AL + BL F
= 0. We will show that any compact subset of the
null controllable region can be covered by the union of a finite sequence of
such ellipsoids.
Lemma 9.3.1. Given an integer L ≥ n0 and a positive number β < 1,

there exists a family of Fi ∈ RmL×n , i = 1, 2, · · · , M , with corresponding
positive definite matrices Pi ’s, such that AL + BL Fi = 0,
E(Pi ) ⊂ L(Fi ), i = 1, 2, · · · , M,
and

M
βC(L) ⊂ E(Pi ),
i=1
where
βC(L) = βx : x ∈ C(L) .
Proof. Let ∂(βC(L)) be the boundary of βC(L). First, we show that, there
exists an ε > 0 such that, for any x1 ∈ ∂(βC(L)), there exist an F ∈ RmL×n
and a P > 0 that satisfy
AL + BL F = 0,
and
B(x1 , ε) ⊂ E(P ) ⊂ L(F ),
where
B(x1 , ε) = x ∈ Rn : |x − x1 | ≤ ε .
Let e be the unit vector in Rn whose th element is 1 and other elements
are zeros. For simplicity, assume x1 = γe1 , otherwise we can use a unitary
transformation x → V x, V T V = I, to satisfy this. Note that a unitary
transformation is equivalent to rotating the state space and does not change
the shapes of B(x1 , ε), E(P ) and C(L).
Since x1 = γe1 ∈ βC(L), it follows from (9.3.1) and (9.3.3) that there
exists a uL 1 ∈ RmL , |uL 1 |∞ ≤ β, such that
AL γe1 + BL uL 1 = 0. (9.3.5)
Define
max{|x| : x ∈ ∂C(L)}
µ= .
min{|x| : x ∈ ∂C(L)}
Since L ≥ n0 , C(L) includes the origin in its interior and hence µ < ∞. It
follows that γe ∈ µβC(L) for all ≥ 2. Therefore, for each ≥ 2, there
exists a uL ∈ RmL , |uL |∞ ≤ µβ, such that
AL γe + BL uL = 0. (9.3.6)
Let F = {fj } be chosen as
1
F = uL 1 uL 2 · · · uL n ,
γ
then,
β µβ
|fj1 | ≤ , |fj | ≤ , ∀ = 2, · · · , n, j = 1, 2, · · · , mL.
γ γ
From (9.3.5) and (9.3.6), we have
1
(AL + BL F )e = AL e + BL uL = 0, = 1, 2, · · · , n.
γ
This shows that AL + BL F = 0.
Let
p1 0
P = ,
0 p2 In−1
where
% &2 % &2 % &−1
1 2β βµ (β + 1)2
p1 = , p2 = (n − 1) 1− .
γ2 β+1 γ 4
Let
γmin = min |x| : x ∈ ∂(βC(L))
and
% & '
−1
2β 2β 2 n − 1)βµ
ε= 1− γmin max ,' .
β+1 β+1 4 − (β + 1)2
Then,
1
2 2β
P ε ≤ 1 − .
β+1
Note that ε is independent of γ and a particular x1 .
We also have
% &2 % &2
1 2 1 1 β n − 1 βµ
fj P −1 fjT = fj1 + Σn=2 fj2
≤ + = 1, (9.3.7)
p1 p2 p1 γ p2 γ
which implies that E(P ) ⊂ L(F ). To see this, we verify that, for any
x ∈ E(P ),
1
1 1 1
|fj x| = fj P − 2 P 2 x ≤ fj P −1 fjT 2 (xT P x) 2 ≤ 1.
For x ∈ B(x1 , ε), we have

1 1 1 1
2 2 2 2β
P x ≤ P x1 + P (x − x1 ) ≤ + P 2 ε ≤ 1.
β+1
This shows that xT P x ≤ 1 and hence B(x1 , ε) ⊂ E(P ) ⊂ L(F ).

Since ∂(βC(L)) is a compact set, there exists a finite set of xi ∈ ∂(βC(L)),
i = 1, 2, · · · , M , such that

M
∂(βC(L)) ⊂ B(xi , ε).
i=1
By the foregoing proof, we know that for each xi ∈ ∂(βC(L)), there exist
an Fi and a Pi such that AL + BL Fi = 0 and
B(xi , ε) ⊂ E(Pi ) ⊂ L(Fi ).
Hence,

M
∂(βC(L)) ⊂ E(Pi ).
i=1

M
βC(L) ⊂ E(Pi ).
i=1
To see this, for any x ∈ βC(L), let y be the intersection of ∂(βC(L)) with
the straight line passing through the origin and x. Then, y ∈ E(Pi0 ) for
some i0 . Since E(Pi0 ) is convex and contains the origin, x ∈ E(Pi0 ). 2
Remark 9.3.1. We would like to point out that, the family of Fi ’s may
contain repeated members with different Pi ’s. This is the case, for example,
when the system has a single input (m = 1) and the lifting step L is the
same as n, the dimension of the state space. In this case, we have only a
−1
unique Fi = −BL AL with C(L) ⊂ L(Fi ).
Lemma 9.3.1 shows that βC(L) can be covered by a finite number of

ellipsoids and within each ellipsoid there is a corresponding linear feedback
law such that the state of (9.3.2) will be steered to the origin the next
step, or equivalently, the state of (9.1.1) will be steered to the origin in
L steps. Since β can be made arbitrarily close to 1 and L can be made
arbitrarily large, any compact subset of C can be covered by a family of
such ellipsoids. It should be noted that as β gets closer to 1, ε will decrease
and we need more ellipsoids to cover βC(L), although the determination
of these ellipsoids could be technically involved for higher order systems.
Also, in the above development, we need to lift the system by L steps to
cover βC(L). Actually, the lifting step can be reduced if we replace the
dead-beat condition AL + BL F = 0 with a less restrictive one:
(AL + BL F )T P (AL + BL F ) − cP ≤ 0,
where c ∈ (0, 1) specifies the requirement of the convergence rate. A direct

consequence of Lemma 9.3.1 is
Theorem 9.3.1. Given any compact subset X 0 of C and a number c ∈

(0, 1), there exist an L ≥ 1 and a family of Fi ∈ RmL×n , i = 1, 2, · · · , M ,
with corresponding positive definite matrices Pi ’s, such that
(AL + BL Fi )T Pi (AL + BL Fi ) − cPi ≤ 0, (9.3.8)
E(Pi ) ⊂ L(Fi ), i = 1, 2, · · · , M, (9.3.9)

and

M
X0 ⊂ E(Pi ). (9.3.10)
i=1
Because of (9.3.8) and (9.3.9), E(Pi ) is an invariant set inside the domain
of attraction for the closed-loop system
xL (k + 1) = AL xL (k) + BL sat(Fi xL (k)).

*
By Theorem 9.2.1, we can use a switching controller to make M i=1 E(Pi )
inside the domain of attraction. Once the state enters the region E(P0 ),
the controller switches to the feedback law
 
F0 (xL (k))
 F0 (x(kL + 1)) 
 
uL (k) = F̄0 (xL (k)) =  .. , (9.3.11)
 . 
F0 (x(kL + L − 1))
where the variables x(kL + i), i = 1, 2, · · · , L − 1, can be recursively com-

puted from the state xL (k) as follows,
x(kL + 1) = AxL (k) + BF0 (xL (k))

x(kL + 2) = Ax(kL + 1) + BF0 (x(kL + 1)
= A(AxL (k) + BF0 (xL (k))) + BF0 (AxL (k) + BF0 (xL (k)))
..
.
9.4. Semi-Globalization – Continuous-Time Systems 205
x(kL + i + 1) = Ax(kL + i) + BF0 (x(kL + i)).
Since the feedback law (9.3.11) is equivalent to u = F0 (x) in the original

time index, under which E(P0 ) is an invariant set, E(P0 ) is also an invariant
set under the feedback law (9.3.11) in the lifted time index and the desired
performance in this region is preserved.
We also observe that, due to the switching and lifting that are involved
in the construction of feedback laws, our final semi-globally stabilizing feed-
back laws, when implemented in the original system (9.1.1), are nonlinear
and periodic in time.
9.4. Semi-Globalization – Continuous-Time Systems
In this section, we consider the continuous-time counterpart of the system

(9.1.1)
ẋ = Ax + Bsat(u), x ∈ Rn , u ∈ Rm . (9.4.1)
The null controllable region at time T , denoted as C(T ), is the set of states
that can be steered to the origin in time T by an admissible control u ∈ Ua .
The null controllable region, denoted as C, is given by

C(T ).
T ≥0
Let h > 0 be the lifting period. We are now interested in controlling

the state of (9.4.1) at times kh, k = 1, 2, · · ·. Denote xh (k) = x(kh) and
uh (k, τ ) = u(kh + τ ). Let Ah = eAh , then the lifted system is
h
xh (k + 1) = Ah xh (k) + eA(h−τ ) Bsat(uh (k, τ ))dτ. (9.4.2)
0
Denote the set of m × n dimensional measurable functions defined on

[0, h) as F m×n . With a matrix function F ∈ F m×n , let the feedback control
be uh (k, τ ) = F (τ )xh (k). Then the closed-loop system is
h
xh (k + 1) = Ah xh (k) + eA(h−τ ) Bsat(F (τ )xh (k))dτ. (9.4.3)
0
The unsaturated region of the feedback law is then given by,

L(F ) := x ∈ Rn : |fj (τ )x| ≤ 1, j = 1, 2, · · · , m, τ ∈ [0, h) ,
where fj ∈ F 1×n is the jth row of F . If xh (k) ∈ L(F ), then
sat(F (τ )xh (k)) = F (τ )xh (k)
and
h
xh (k + 1) = Ah + eA(h−τ ) BF (τ )dτ xh (k). (9.4.4)
0
The feedback uh (k, τ ) = F (τ )xh (k) is stabilizing if there exists a P > 0

such that
h
T h
A(h−τ ) A(h−τ )
Ah + e BF (τ )dτ P Ah + e BF (τ )dτ − P < 0.
0 0
Note that P can be scaled such that E(P ) ⊂ L(F ). In this case, E(P ) is an
invariant set inside the domain of attraction for the system (9.4.3). Since
for all xh (k) ∈ E(P ), the control is linear in xh (k), so, when xh (k) tends
to the origin, the control uh (k, τ ) = F (τ )xh (k) will gets smaller and hence
the state of the original system (9.4.1) between t = kh and t = (k + 1)h will
stay close to xh (k). Similar to the discrete-time case, we have the following
lemma.
Lemma 9.4.1. Given h > 0 and a positive number β < 1, there exists a
family of Fi ∈ F m×n , i = 1, 2, · · · , M , with corresponding positive definite
matrices Pi ’s, such that
h
Ah + eA(h−τ )BFi (τ )dτ = 0,
0
E(Pi ) ⊂ L(Fi ), i = 1, 2, · · · , M,
and

M
βC(h) ⊂ E(Pi ).
i=1
Proof. The idea of the proof is the same as that of Lemma 9.3.1. Here we
just show how to construct ε, F and P for a given x1 ∈ ∂(βC(h)). We also
assume that x1 = γe1 . Since γe1 ∈ ∂(βC(h)), there exists a u1 ∈ F m×1 ,
|u1 (τ )|∞ ≤ β for all τ ∈ [0, h), such that
h
Ah γe1 + eA(h−τ )Bu1 (τ )dτ = 0,
0
9.5. An Example 207
and for ≥ 2, there exists a u ∈ F m×1 , |u (τ )|∞ ≤ βµ for all τ ∈ [0, h),
such that h
Ah γe + eA(h−τ )Bu (τ )dτ = 0.
0
Let
1
u1 u2 · · · un ,
F =
γ
and P and ε be the same as those in the proof of Lemma 9.3.1, the remaining
part of the proof will be the same as that of Lemma 9.3.1 except that (9.3.7)
is replaced by
fj (τ )P −1 fjT (τ ) ≤ 1, ∀ τ ∈ [0, h), j = 1, 2, · · · , m.
2
The following is the counterpart of Theorem 9.3.1 for the discrete-time
system (9.1.1).
Theorem 9.4.1. Given any compact subset X 0 of C and a number c ∈

(0, 1), there exist an h > 0 and a family of Fi ∈ F m×n , i = 1, 2, · · · , M ,
with corresponding positive definite matrices Pi ’s, such that

T
h h
Ah + eA(h−τ ) BFi (τ )dτ Pi Ah + eA(h−τ ) BFi (τ )dτ − cPi ≤ 0,
0 0
E(Pi ) ⊂ L(Fi ), i = 1, 2, · · · , M,
and

M
X0 ⊂ E(Pi ).
i=1
Again, by Theorem 9.2.1, we can use a switching controller to make

*M
i=0 E(Pi ) inside the domain of attraction and hence semi-global stabiliza-
tion can be achieved. Moreover, once the state enters the region E(P0 ), the
controller switches to the feedback law u = F0 (x) and hence the desired
performance in this region is preserved.
9.5. An Example
Consider the system (9.1.1) with

0.8876 −0.5555 −0.1124
A= , B= .
0.5555 1.5542 0.5555
The matrix A is exponentially unstable with a pair of eigenvalues 1.2209 ±

j0.4444. The LQR controller corresponding to the cost function
∞

J= x(k)T Qx(k) + u(k)T Ru(k) ,
k=0
with Q = I and R = 1 is

u = F0 (x) = −0.2630 −2.1501 x.
Let D0 be obtained as E(P0 ) with

2.1367 −0.2761
P0 =
−0.2761 1.7968
(see the ellipsoid enclosed by the solid curve in Fig. 9.5.1).
To enlarge the domain of attraction, we take a lifting step of 8 and
obtain 16 invariant ellipsoids with corresponding feedback controllers (see
the ellipsoids enclosed by the dotted curves in Fig. 9.5.1). Each invariant
ellipsoid is optimal with respect to certain xi in the sense that it contains
αxi with |α| maximized (see the points marked with “∗”). This is computed
by using the LMI method of Chapter 8. The outermost curve in Fig. 9.5.1
is the boundary of the null controllable region C. We see that the union of
the ellipsoids covers a large portion of C.
Figs. 9.5.2-9.5.4 show some simulation results of the closed-loop system
under the multiple switching controls. The initial state is very close to the
boundary of C. In Fig. 9.5.2, the dashed trajectory is that of the unlifted
system (9.1.1) under the switching control, and the trajectory of the lifted
system is marked with “∗”. Figs. 9.5.3 and 9.5.4 are the states and control
of the original unlifted system.
9.6. Conclusions
In this chapter, we proposed a control design method for semi-global sta-
bilization on the null controllable region with guaranteed regional perfor-
mance. This design method applies to general (possibly exponentially un-
stable) systems in either continuous-time or discrete-time. The resulting
feedback laws expand the domain of attraction achieved by an a priori de-
signed feedback law to include any bounded set in the interior of the null
controllable region, while preserving the desired performance of the original
feedback law in a fixed region.
1.5
0.5
−0.5
−1
−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
Figure 9.5.1: The union of the invariant ellipsoids.
1.5
0.5
−0.5
−1
−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
Figure 9.5.2: A trajectory under the multiple switching control.

0.5
−0.5
−1
−1.5
0 10 20 30 40 50 60 70 80 90
Figure 9.5.3: The states: x1 , ‘-’; x2 , ‘- -’.
0.8
0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
−1
0 10 20 30 40 50 60 70 80
Figure 9.5.4: The control.

Chapter 10
Disturbance Rejection
with Stability
10.1. Introduction
In this chapter, we will study the following linear systems subject to actu-
ator saturation and persistent disturbances,
ẋ = Ax + B sat(u) + Ew, (10.1.1)
and
x(k + 1) = Ax(k) + Bsat(u(k)) + Ew(k), (10.1.2)
where x ∈ Rn is the state, u ∈ Rm is the control and w ∈ Rq is the dis-

turbance. Also, sat : Rm → Rm is the standard saturation function that
represents the constraints imposed by the actuators. Since the terms Ew
and Ew(k) are outside of the saturation function, a trajectory might go
unbounded no matter where it starts and whatever control we apply. Our
primary concern is the boundedness of the trajectories in the presence of
disturbances. We are interested in knowing if there exists a bounded set
such that all the trajectories starting from inside of it can be kept within
it. If there is such a bounded set, we would further like to synthesize feed-
back laws that have the ability to reject the disturbance. Here disturbance
rejection is in the sense that, there is a small (as small as possible) neigh-
borhood of the origin, such that all the trajectories starting from inside of
it (in particular, the origin) will remain in it. This performance is analyzed,
211
212 Chapter 10. Disturbance Rejection with Stability
for example, for the class of disturbances with finite energy in [32]. In this
chapter, we will deal with persistent disturbances.
In addition, we are also interested in the problem of asymptotic distur-
bance rejection with nonzero initial states. This is the problem of designing
a feedback law that, in the presence of disturbances, causes closed-loop sys-
tem trajectories starting from a set to converge to a smaller set. A related
problem was addressed in [85] and Chapter 6, where the disturbances are
input additive and enter the system before the saturating actuators, i.e.,
the system has the following state equation,
ẋ = Ax + Bsat(u + w).
Reference [85] deals with the class of systems with A having all its eigen-
values in the closed left-half plane, and Chapter 6 deals with exponentially
unstable systems with A having two or fewer eigenvalues in the open right-
half plane. It is shown in [85] and Chapter 6 that there exist a family
of linear feedbacks that achieve asymptotic disturbance decoupling. This
is in the sense that, given any positive number D, any compact subset
X 0 of the asymptotically null controllable region and any arbitrarily small
neighborhood X ∞ of the origin, there is a feedback control from this fam-
ily such that any trajectory starting from within X 0 will enter X ∞ in a
finite time for all the disturbance w, w∞ ≤ D. We, however, could not
expect to have this nice result for the systems (10.1.1) and (10.1.2), where
the disturbance enters the system after the saturating actuator. If w or E
is sufficiently large, it may even be impossible to keep the state bounded.
What we expect is to have a set X 0 (as large as we can get) and a set X ∞
(as small as we can get) such that all the trajectories starting from within
X 0 will enter X ∞ in a finite time and remain in it thereafter.
This chapter is organized as follows. Section 10.2 deals with continuous-
time systems. Several problems related to set invariance in the presence
of disturbances and disturbance rejection are formulated in Section 10.2.1.
The solutions to these problems are given in Section 10.2.2 and 10.2.3. In
particular, Section 10.2.2 gives a condition for set invariance in the presence
of disturbances. Section 10.2.3 solves the problem of disturbance rejection
with guaranteed domain of attraction. An example is presented in Sec-
tion 10.2.4 to illustrate the effectiveness of the design methods. Section 10.3
deals with discrete-time systems and parallels Section 10.2. Section 10.4
draws a brief conclusion to the chapter.
10.2. Continuous-Time Systems 213

ẋ = Ax + Bsat(u) + Ew, x ∈ Rn , u ∈ Rm , w ∈ Rq , (10.2.1)
where, for simplicity and without loss of generality, we assume that the
bounded disturbance w belongs to the set

W := w : w(t)T w(t) ≤ 1, ∀ t ≥ 0 .
Let the state feedback be u = F x. Then the closed-loop system is given by
ẋ = Ax + Bsat(F x) + Ew. (10.2.2)
For an initial state x(0) = x0 , denote the state trajectory of the closed-loop
system in the presence of w as ψ(t, x0 , w). A set in Rn is said to be invariant
if all the trajectories starting from within it will remain in it regardless of
w ∈ W. Let P ∈ Rn×n be a positive definite matrix and let V (x) = xT P x.
An ellipsoid E(P, ρ) is said to be strictly invariant if
V̇ (x, w) = 2xT P (Ax + Bsat(F x) + Ew) < 0
for all x ∈ ∂E(P, ρ) and all w, wT w ≤ 1. The notion of invariant set plays
an important role in studying the stability and other performances of a
system (see [7,9] and the references therein). If an invariant set is bounded,
then all the trajectory starting from within it will be bounded.
Our primary concern is the boundedness of the trajectories for some (as
large as possible) set of initial states. This requires a large invariant set.
On the other hand, for the purpose of disturbance rejection, we would also
like to have a small invariant set containing the origin in its interior so that
a trajectory starting from the origin will stay close to the origin.
To formally state the objectives of this section, we need to extend the
notion of the domain of attraction as follows.
Definition 10.2.1. Let X be a bounded invariant set of (10.2.2). The

domain of attraction of X is

S(X ) := x0 ∈ Rn : lim dist(ψ(t, x0 , w), X ) = 0, ∀ w ∈ W ,
t→∞
where
dist(ψ(t, x0 , w), X ) = inf |ψ(t, x0 , w) − x|
x∈X
is the distance from ψ(t, x0 , w) to X .
In the above definition, | · | can be any vector norm. The problems we
are to address in this section can be formulated as follows.
Problem 10.2.1 (Set invariance analysis). Let F be known. Determine if
a given ellipsoid E(P, ρ) is (strictly) invariant.
Problem 10.2.2 (Invariant set enlargement). Given a shape reference set
X 0 ⊂ Rn , design an F such that the closed-loop system has a bounded
invariant set E(P, ρ) which contains α2 X 0 with α2 maximized.
Problem 10.2.3 (Disturbance rejection). Given a shape reference set
X ∞ ⊂ Rn , design an F such that the closed-loop system has an invari-
ant set E(P, ρ) ⊂ α3 X ∞ with α3 minimized. Here we can also take X ∞ to
be the (possibly unbounded) polyhedron

x ∈ Rn : |ci x| ≤ 1, i = 1, 2, · · · , p .
Then, the minimization of α3 leads to the minimization of the L∞ -norm of

the output y = Cx ∈ Rp .
Problem 10.2.4 (Disturbance rejection with guaranteed domain of at-
traction). Given two shape reference sets, X ∞ and X 0 . Design an F
such that the closed-loop system has an invariant set E(P, 1) that contains
X 0 , and for all x0 ∈ E(P, 1), ψ(t, x0 , w) will enter a smaller invariant set
E(P, ρ1 ) ⊂ α4 X ∞ with α4 minimized.
We consider the closed-loop system (10.2.2) with a given F . Following

Chapter 7, we use D = {Di : i ∈ [1, 2m ]} to denote the set of m × m
diagonal matrices whose diagonal elements are eithe 1 or 0. We also denote
Di− = I − Di .
Theorem 10.2.1. For a given ellipsoid E(P, ρ), if there exist an H ∈
Rm×n and a positive number η such that
(A + B(Di F + Di− H))T P + P (A + B(Di F + Di− H))
1 η
+ P EE T P + P ≤ 0 (< 0), ∀ i ∈ [1, 2m ], (10.2.3)
η ρ
and E(P, ρ) ⊂ L(H), then E(P, ρ) is a (strictly) invariant set for the system
(10.2.2).
Proof. We will prove the strict invariance. That is, for V (x) = xT P x, we
will show that
V̇ (x, w) = 2xT P (Ax + Bsat(F x) + Ew) < 0, ∀ x ∈ ∂E(P, ρ), wT w ≤ 1.
Following the procedure of the proof of Theorem 7.4.1, we can show that
for each x ∈ E(P, ρ),
2xT P (Ax + Bsat(F x)) ≤ maxm 2xT P (A + B(Di F + Di− H))x.

i∈[1,2 ]
Recall that for any positive number η,

1 T
2aT b ≤ a a + ηbT b, ∀ a, b ∈ Rn .
η
It follows that
1 T 1
2xT P Ew ≤ x P EE T P x + ηwT w ≤ xT P EE T P x + η.
η η
Hence,
1
V̇ (x, w) ≤ maxm 2xT P (A + B(Di F + Di− H))x + xT P EE T P x + η.
i∈[1,2 ] η
It follows from (10.2.3) that for all x ∈ E(P, ρ), wT w ≤ 1,
η
V̇ (x, w) < − xT P x + η.
ρ
Observing that on the boundary of E(P, ρ), xT P x = ρ, hence V̇ (x, w) < 0.
This shows that E(P, ρ) is a strictly invariant set. 2
Theorem 10.2.1 deals with Problem 10.2.1 and can be easily used for
controller design in Problems 10.2.2 and 10.2.3. For Problem 10.2.2, we
can solve the following optimization problem:
sup α2 (10.2.4)
P >0,ρ,η>0,F,H
s.t. a) α2 X 0 ⊂ E(P, ρ),
b) (A + B(Di F + Di− H))T P + P (A + B(Di F + Di− H))
1 η
+ P EE T P + P < 0, i ∈ [1, 2m ],
η ρ
c) |hi x| ≤ 1, ∀ x ∈ E(P, ρ), i ∈ [1, m].
Let % &−1
P
Q= , Y = F Q, Z = HQ,
ρ
then, constraint b) is equivalent to
QAT + AQ + (Di Y + Di− Z)T B T + B(Di Y + Di− Z)

ρ η
+ EE T + Q < 0, i ∈ [1, 2m ].
η ρ
If we fix ηρ , then the original optimization constraints can be transformed
into LMIs as with (7.4.8). The global maximum of α2 will be obtained by
running ηρ from 0 to ∞. Since ρ can be absorbed into other parameters, we
simply set ρ = 1.
For Problem 10.2.3, we have
inf α3 (10.2.5)
P >0,ρ,η>0,F,H
s.t. a) E(P, ρ) ⊂ α3 X ∞ ,
b) (A + B(Di F + Di− H))T P + P (A + B(Di F + Di− H))
1 η
+ P EE T P + P < 0, i ∈ [1, 2m ],
η ρ
c) |hi x| ≤ 1, ∀ x ∈ E(P, ρ), i ∈ [1, m],
which can be solved similarly as Problem 10.2.2.
10.2.3. Disturbance Rejection with Guaranteed Domain

of Attraction
Given X 0 ⊂ Rn , if the optimal solution of Problem 10.2.2 is α∗2 > 1, then

there are infinitely many choices of the feedback gain matrix F such that X 0
is contained in some invariant ellipsoid. We will use this extra freedom for
disturbance rejection. That is, to construct another invariant set E(P, ρ1 )
which is as small as possible with respect to some shape reference set X ∞ .
Moreover, X 0 is inside the domain of attraction of E(P, ρ1 ). In this way,
all the trajectories starting from X 0 will enter E(P, ρ1 ) ⊂ α4 X ∞ for some
α4 > 0. Here the number α4 is a measure of the degree of disturbance
rejection.
Before addressing Problem 10.2.4, we need to answer the following ques-
tion: Suppose that for a given F and P , both E(P, ρ1 ) and E(P, ρ2 ), ρ1 < ρ2 ,
are strictly invariant sets, then under what conditions will the other ellip-
soids E(P, ρ), ρ ∈ (ρ1 , ρ2 ) also be strictly invariant? If they are, then all the
trajectories starting from within E(P, ρ2 ) will enter E(P, ρ1 ) and remain in-
side it. This can be seen as follows. If all the ellipsoids E(P, ρ), ρ ∈ [ρ1 , ρ2 ],
are strictly invariant, then we have
V̇ (x, w) < 0, ∀ xT P x ∈ [ρ1 , ρ2 ], wT w ≤ 1.
Since V̇ (x, w) is a continuous function in x and w, there exists a positive

number ε > 0 such that
V̇ (x, w) < −ε, ∀ xT P x ∈ [ρ1 , ρ2 ], wT w ≤ 1.
It follows that all the trajectories starting from within E(P, ρ2 ) will enter
E(P, ρ1 ) and remain inside it.
Theorem 10.2.2. Given two ellipsoids, E(P, ρ1 ) and E(P, ρ2 ), ρ2 > ρ1 >
0, if there exist H1 , H2 ∈ Rm×n and a positive η such that
(A + B(Di F + Di− H1 ))T P + P (A + B(Di F + Di− H1 ))

1 η
+ P EE T P + P < 0, ∀ i ∈ [1, 2m ], (10.2.6)
η ρ1
(A + B(Di F + Di− H2 ))T P + P (A + B(Di F + Di− H2 ))
1 η
+ P EE T P + P < 0, ∀ i ∈ [1, 2m ], (10.2.7)
η ρ2
and E(P, ρ1 ) ⊂ L(H1 ), E(P, ρ2 ) ⊂ L(H2 ), then, for every ρ ∈ [ρ1 , ρ2 ], there
exists an H ∈ Rm×n such that
(A + B(Di F + Di− H))T P + P (A + B(Di F + Di− H))

1 η
+ P EE T P + P < 0, ∀ i ∈ [1, 2m ], (10.2.8)
η ρ
and E(P, ρ) ∈ L(H). This implies that E(P, ρ) is also strictly invariant.
Proof. Let h1,i and h2,i be the ith row of H1 and H2 respectively. The
conditions E(P, ρ1 ) ⊂ L(H1 ) and E(P, ρ2 ) ⊂ L(H2 ) are equivalent to
1 1
ρ1 h1,i ρ2 h2,i
≥ 0, ≥ 0, i = 1, 2, · · · , m.
hT1,i P hT2,i P
Since ρ ∈ [ρ1 , ρ2 ], there exists a λ ∈ [0, 1] such that

1 1 1
= λ + (1 − λ) .
ρ ρ1 ρ2
Let
H = λH1 + (1 − λ)H2 .
Clearly
1
ρ hi
≥ 0.
hTi P
From (10.2.6) and (10.2.7), and by convexity, we have (10.2.8). 2
In view of Theorem 10.2.2, to solve Problem 10.2.4, we only need to
construct two invariant ellipsoids E(P, ρ1 ) and E(P, ρ2 ) satisfying the con-
dition of Theorem 10.2.2 such that X 0 ⊂ E(P, ρ2 ) and E(P, ρ1 ) ⊂ α4 X ∞
with α4 minimized. Since ρ2 can be absorbed into other parameters, we
assume for simplicity that ρ2 = 1 and ρ1 < 1. Problem 10.2.4 can then be
formulated as
inf α4 (10.2.9)
P >0,0<ρ1 <1,η>0,F,H1 ,H2 ,
s.t. a) X 0 ⊂ E(P, 1), E(P, ρ1 ) ⊂ α4 X ∞ ,
b) (A + B(Di F + Di− H1 ))T P + P (A + B(Di F + Di− H1 ))
1 η
+ P EE T P + P < 0, i ∈ [1, 2m ],
η ρ1
c) (A + B(Di F + Di− H2 ))T P + P (A + B(Di F + Di− H2 ))
1
+ P EE T P + ηP < 0, i ∈ [1, 2m ],
η
d) |h1,i x| ≤ 1, ∀ x ∈ E(P, ρ1 ), i ∈ [1, m],
e) |h2,i x| ≤ 1, ∀ x ∈ E(P, 1), i ∈ [1, m].
Suppose that X ∞ and X 0 are ellipsoids, X ∞ = E(R1 , 1), and X 0 =

E(R2 , 1). Let
γ = α24 , Q = P −1 , Y = F Q, Z1 = H1 Q, Z2 = H2 Q,
and denote the ith row of Z1 and Z2 as z1,i , z2,i , respectively. Then, the
above problem is equivalent to
inf γ (10.2.10)
Q>0,0<ρ1 <1,η>0,Y,Z1 ,Z2 ,
s.t. a) Q ≥ R2−1 , ρ1 Q ≤ γR1−1 ,

b) QAT + AQ + (Di Y + Di− Z1 )T B T + B(Di Y + Di− Z1 )
1 η
+ EE T + Q < 0, i ∈ [1, 2m ],
η ρ1
c) QAT + AQ + (Di Y + Di− Z2 )T B T + B(Di Y + Di− Z2 )

1
+ EE T + ηQ < 0, i ∈ [1, 2m ],
η
1
ρ1 z 1,i
d) ≥ 0, i ∈ [1, m],
T
z1,i Q

1 z2,i
e) T
≥ 0, i ∈ [1, m].
z2,i Q
If we fix ρ1 and η, then the constraints of the optimization problem

(10.2.10) become LMIs. To obtain the global infimum, we may vary ρ1
from 0 to 1 and η from 0 to ∞.
10.2.4. An Example
The open loop system is described by (10.2.1) with

0.6 −0.8 2 0.1
A= , B= , E= .
0.8 0.6 4 0.1
The system has a pair of unstable complex poles. We first ignore the
disturbance and use the method of Chapter 8 (by solving (8.2.3)) to find
a feedback with the objective of maximizing the domain of attraction with
respect to the unit ball, X R = E(I, 1). The result is,
α∗1 = 2.4417,

0.0752 −0.0566
P1∗ = ,
−0.0566 0.1331
and

F1∗ = −0.0025 −0.2987 .
The invariant ellipsoid E(P1∗ , 1) is plotted in Fig. 10.2.1 as the larger ellip-
soid. As a comparison, we also plotted the boundary of the null controllable
region of the open loop system (see the dashed outer curve).
We next deal with Problem 10.2.2. By solving (10.2.4) with X 0 being
a unit ball, we obtain α∗2 = 2.3195, with η2∗ = 0.019. The corresponding
feedback matrix is

F2∗ = −0.0009 −0.3110 .
The resulting invariant ellipsoid is E(P2∗ , 1) with

0.0835 −0.0639
P2∗ = .
−0.0639 0.1460
It is the inner dash-dotted ellipsoid in Fig. 10.2.1.
−1
−2
−3
−4
−5
−6 −4 −2 0 2 4 6
Figure 10.2.1: The invariant ellipsoids and the null controllable region.
To deal with Problem 10.2.3, we solve (10.2.5), with X ∞ also being a

unit ball, we obtain α∗3 = 0.0606, which shows that the disturbance can be
rejected to a very small level. The corresponding feedback gain is

F3∗ = 1955.5 −1161.1 ,
which is very large as compared with F2∗ , the one that achieves large domain
of attraction.
Now we turn to Problem 10.2.4. The optimization result from solving
Problem 10.2.2 gives us some guide in choosing X 0 . If X 0 is a disk, we
know that X 0 must be smaller than the maximal disk contained in E(P2∗ , 1),

which is, E I, (α∗2 )2 = E(I, 2.31952). Here we choose X 0 = E(I, 22 ) and
X ∞ = E(I, 1). To obtain the optimal solution, we vary η from 0 to ∞ and
ρ1 from 0 to 1 and for fixed ρ1 and η, we solve (10.2.10). The solution is

α∗4 = 0.9725, η ∗ = 0.006, ρ∗1 = 0.0489,

F4∗ = 0.2844 −1.4430 ,
and
0.1145 −0.0922
P4∗ = .
−0.0922 0.1872
In Fig. 10.2.2, the larger ellipsoid is E(P4∗ , 1), the smaller ellipsoid is
E(P4∗ , ρ1 ) and the outermost dashed closed-curve is the boundary of the
null controllable region. A trajectory is plotted from x0 ∈ ∂E(P4∗ , 1) with
w =sign(sin(0.3t)).
−1
−2
−3
−4
−5
−6 −4 −2 0 2 4 6
Figure 10.2.2: The invariant ellipsoids and a trajectory.

x(k + 1) = Ax(k) + Bsat(u(k)) + Ew(k),

x ∈ Rn , u ∈ Rm , w ∈ Rq , (10.3.1)
where, without loss of generality, we assume that the bounded disturbance

w belongs to the set

W := w : w(k)T w(k) ≤ 1, ∀ k ≥ 0 .
Let the state feedback be u = F x. Then, the closed-loop system is given

by
x(k + 1) = Ax(k) + Bsat(F x(k)) + Ew(k). (10.3.2)
For an initial state x(0) = x0 , denote the state trajectory of the closed-loop
system in the presence of w as ψ(k, x0 , w).
Our primary concern is the boundedness of the trajectories. A set in
n
R is said to be invariant if all the trajectories starting from within it will
remain in it regardless of w ∈ W. An ellipsoid E(P, ρ) is said to be strictly
invariant if
(Ax + Bsat(F x) + Ew)T P (Ax + B sat(F x) + Ew) < ρ
for all x ∈ E(P, ρ) and all w, wT w ≤ 1.

Similar to the continuous-time case, we need to extend the notion of
the domain of attraction of an equilibrium to that of an invariant set as
follows.
Definition 10.3.1. Let X be a bounded invariant set of (10.3.2). The

domain of attraction of X is

S(X ) := x0 ∈ R : lim dist(ψ(k, x0 , w), X ) = 0, ∀ w ∈ W ,
n
k→∞
where
dist(ψ(k, x0 , w), X ) = inf |ψ(k, x0 , w) − x|
x∈X
is the distance from ψ(k, x0 , w) to X .
In parallel to the continuous-time case, the problems we are to address

in this section can be formulated as follows.
Problem 10.3.1 (Set invariance analysis). Let F be known. Determine if

a given ellipsoid E(P, ρ) is (strictly) invariant.
Problem 10.3.2 (Invariant set enlargement). Given a shape reference set

X 0 ⊂ Rn , design an F such that the closed-loop system has a bounded
invariant set E(P, ρ) ⊃ α2 X 0 with α2 maximized.
Problem 10.3.3 (Disturbance rejection). Given a shape reference set

X ∞ ⊂ Rn , design an F such that the closed-loop system has an invari-
ant set E(P, ρ) ⊂ α3 X ∞ with α3 minimized. Here we can also take X ∞ to
be the (possibly unbounded) polyhedron

x ∈ Rn : |ci x| ≤ 1, i = 1, 2, · · · , p .
In this case, the minimization of α3 leads to the minimization of the L∞ -

norm of the output y = Cx ∈ Rp .
Problem 10.3.4 (Disturbance rejection with guaranteed domain of at-

traction). Given two shape reference sets, X ∞ and X 0 . Design an F such
that the closed-loop system has an invariant set E(P, 1) ⊃ X 0 , and for all
x0 ∈ E(P, 1), ψ(k, x0 , w) will enter a smaller invariant set E(P, ρ1 ) ⊂ α4 X ∞
with α4 minimized.
We consider the closed-loop system (10.3.2) with a given F .
Theorem 10.3.1. For a given ellipsoid E(P, ρ), if there exist an H ∈

Rm×n and a positive number η such that
(1 + η)(A + B(Di F + Di− H))T P (A + B(Di F + Di− H))

% &
1+η
+ λmax (E T P E) − 1 P ≤ (<)0, ∀ i ∈ [1, 2m ], (10.3.3)
ρη
and E(P, ρ) ⊂ L(H), then E(P, ρ) is a (strictly) invariant set for the system
(10.3.2).
Proof. We prove the strict invariance. That is, we will show that
(Ax + Bsat(F x) + Ew)T P (Ax + Bsat(F x) + Ew) < ρ,
for all x ∈ E(P, ρ) and all w, wT w ≤ 1. Following the procedure of the

proof of Theorem 7.5.1, we can show that for every x ∈ E(P, ρ), and every
w, wT w ≤ 1,
(Ax + Bsat(F x) + Ew)T P (Ax + Bsat(F x) + Ew)

≤ maxm (Ax+B(Di F +Di− H)x+Ew)T P (Ax+B(Di F +Di− H)x+Ew).
i∈[1,2 ]
Using the fact that, for any η > 0,

% &
1 T
(a + b)T (a + b) ≤ (1 + η)aT a + 1 + b b, ∀ a, b ∈ Rn ,
η
we have,
(Ax + Bsat(F x) + Ew)T P (Ax + Bsat(F x) + Ew)

≤ maxm (1 + η)xT (A + B(Di F + Di− H))T P (A + B(Di F + Di− H))x
i∈[1,2 ]
% &
1
+ 1+ wT E T P Ew
η
≤ maxm (1 + η)xT (A + B(Di F + Di− H))T P (A + B(Di F + Di− H))x
i∈[1,2 ]
% &
1
+ 1+ λmax (E T P E).
η
To prove the strict invariance, it suffices to show that for all x ∈ ∂E(P, ρ)
and for all i ∈ [1, 2m ],
(1 + η)xT (A + B(Di F + Di− H))T P (A + B(Di F + Di− H))x

% &
1
+ 1+ λmax (E T P E) < ρ. (10.3.4)
η
Noticing that % &
P
x T
x = 1, ∀x ∈ ∂E(P, ρ),
ρ
we see that (10.3.4) is guaranteed by (10.3.3). 2
Theorem 10.3.1 deals with Problem 10.3.1 and can be easily used for
controller design in Problems 10.3.2 and 10.3.3. For Problem 10.3.2, we
can solve the following optimization problem,
sup α2 (10.3.5)
P >0,ρ,η>0,F,H
s.t. a) α2 X 0 ⊂ E(P, ρ),
b) (1 + η)(A + B(Di F + Di− H))T P (A + B(Di F + Di− H))
% % T & &
1+η E PE
+ λmax − 1 P ≤ 0, i ∈ [1, 2m ],
η ρ
c) |hi x| ≤ 1, ∀ x ∈ E(P, ρ), i ∈ [1, m].
Let % &−1
P
Q= , Y = F Q, Z = HQ,
ρ
then with a similar procedure as in dealing with the optimization problem

(7.5.7), it can be shown that constraint b) is equivalent to the existence of
% &
η
λ ∈ 0,
1+η
such that

1+η
1− η λ Q (AQ + BDi Y + BDi− Z)T
≥ 0,
AQ + BDi Y + BDi− Z 1
1+η Q
and
λ ET
≥ 0.
E Q
If we fix η and λ, then the original optimization constraints can be
transformed into LMIs as with (7.4.8). The global maximum of α2 will be
obtained by running η from 0 to ∞ and λ from 0 to 1+ηη
. Since ρ can be
absorbed into other parameters, we simply set ρ = 1.
For Problem 10.3.3, we have
inf α3 (10.3.6)
P >0,ρ,η>0,F,H
s.t. a) E(P, ρ) ⊂ α3 X ∞ ,
b) (1 + η)(A + B(Di F + Di− H))T P (A + B(Di F + Di− H))
% % T & &
1+η E PE
+ λmax − 1 P ≤ 0, i ∈ [1, 2m ],
η ρ
c) |hi x| ≤ 1, ∀ x ∈ E(P, ρ), i ∈ [1, m],
which can be solved similarly as Problem 10.3.2.
10.3.3. Disturbance Rejection with Guaranteed Domain

of Attraction
Given X 0 ⊂ Rn , if the optimal solution of Problem 10.3.2 is α∗2 > 1, then

there are infinitely many choices of the feedback matrix F such that X 0 is
contained in some invariant ellipsoid. We will use this extra freedom for
disturbance rejection. That is, to construct another invariant set E(P, ρ1 )
which is as small as possible with respect to some X ∞ . Moreover, X 0 is
inside the domain of attraction of E(P, ρ1 ). In this way, all the trajectories
starting from within X 0 will enter E(P, ρ1 ) ⊂ α4 X ∞ for some α4 > 0. Here
the number α4 is a measure of the degree of disturbance rejection.
Similar to the continuous-time case, before addressing Problem 10.3.4,

we need to answer the following question: Suppose that for given F and
P , both E(P, ρ1 ) and E(P, ρ2 ), ρ1 < ρ2 , are strictly invariant sets, then
under what conditions will the other ellipsoids E(P, ρ), ρ ∈ (ρ1 , ρ2 ), also be
strictly invariant? If they are, then all the trajectories starting from within
E(P, ρ2 ) will enter E(P, ρ1 ) and remain inside it.
Theorem 10.3.2. Given two ellipsoids, E(P, ρ1 ) and E(P, ρ2 ), ρ2 > ρ1 > 0,
if there exist H1 , H2 ∈ Rm×n and a positive η such that
(1 + η)(A + B(Di F + Di− H1 ))T P (A + B(Di F + Di− H1 ))

% &
1+η
+ λmax (E T P E) − 1 P < 0, (10.3.7)
ρ1 η
(1 + η)(A + B(Di F + Di− H2 ))T P (A + B(Di F + Di− H2 ))
% &
1+η
+ λmax (E T P E) − 1 P < 0, (10.3.8)
ρ2 η
for all i ∈ [1, 2m ], and E(P, ρ1 ) ⊂ L(H1 ), E(P, ρ2 ) ⊂ L(H2 ), then for every
ρ ∈ [ρ1 , ρ2 ], there exists an H ∈ Rm×n such that
(1 + η)(A + B(Di F + Di− H))T P (A + B(Di F + Di− H)) +

% &
1+η
λmax (E P E) − 1 P < 0.
T
(10.3.9)
ρη
and E(P, ρ) ∈ L(H). This implies that E(P, ρ) is also strictly invariant.
Proof. Let h1,i and h2,i be the ith rows of H1 and H2 respectively. The
conditions E(P, ρ1 ) ⊂ L(H1 ) and E(P, ρ2 ) ⊂ L(H2 ) are equivalent to
1 1
ρ1 h1,i ρ2 h2,i
≥ 0, ≥ 0, i = 1, 2, · · · , m.
hT1,i P hT2,i P
Since ρ ∈ [ρ1 , ρ2 ], there exists an α ∈ [0, 1] such that

1 1 1
= α + (1 − α) .
ρ ρ1 ρ2
Let
H = αH1 + (1 − α)H2 .
Clearly,
1
ρ hi
≥ 0.
hTi P
Since (10.3.7) and (10.3.8) are equivalent to

1 − 1+η T
ρ1 η λmax (E P E) P (A + B(Di F + Di− H1 ))T
>0
A + B(Di F + Di− H1 ) P −1
and

1+η
1− T
ρ2 η λmax (E P E) P (A + B(Di F + Di− H2 ))T
> 0,
A + B(Di F + Di− H2 ) P −1
by convexity, we have

1 − 1+η T
ρη λmax (E P E) P (A + B(Di F + Di− H))T
> 0,
A + B(Di F + Di− H) P −1
which is equivalent to (10.3.9). 2
In view of Theorem 10.3.2, to solve Problem 10.3.4, we can construct

two invariant ellipsoids E(P, ρ1 ) and E(P, ρ2 ) satisfying the condition of
Theorem 10.3.2 such that X 0 ⊂ E(P, ρ2 ) and E(P, ρ1 ) ⊂ α4 X ∞ with α4
minimized. Since ρ2 can be absorbed into other parameters, we assume for
simplicity that ρ2 = 1 and ρ1 < 1. Problem 10.3.4 can then be formulated
as
inf α4 (10.3.10)
P >0,0<ρ1 <1,η>0,F,H1 ,H2 ,
s.t. a) X 0 ⊂ E(P, 1), E(P, ρ1 ) ⊂ α4 X ∞ ,

1 − 1+η
ρ η λmax (E T
P E) P (A + B(D i F + D −
i H 1 ))T
b) 1 > 0,
A + B(Di F + Di− H1 ) P −1
i ∈ [1, 2m ],

1 − 1+η
η λmax (E T
P E) P (A + B(Di F + Di− H2 ))T
c) > 0,
A + B(Di F + Di− H2 ) P −1
i ∈ [1, 2m ],
d) |h1,i x| ≤ 1, ∀ x ∈ E(P, ρ1 ), i ∈ [1, m],
e) |h2,i x| ≤ 1, ∀ x ∈ E(P, 1), i ∈ [1, m].
Suppose that X ∞ and X 0 are ellipsoids, X ∞ = E(R1 , 1), and X 0 =

E(R2 , 1). Let
γ = α24 , Q = P −1 , Y = F Q, Z1 = H1 Q, Z2 = H2 Q,
then the above problem is equivalent to
inf γ (10.3.11)
Q>0,0<ρ1 <1,η>0,λ,Y,Z1 ,Z2 ,
s.t. a) Q ≥ R2−1 , ρ1 Q ≤ γR1−1 ,

1 − 1+η
ρ1 η λ Q (A + B(D i Y + D −
i Z 1 ))T
b) > 0,
A + B(Di Y + Di− Z1 ) Q
i ∈ [1, 2m ],

1+η
1− η λ (A + B(Di Y + Di− Z2 ))T
Q
c) > 0,
A + B(Di Y + Di− Z2 ) Q
i ∈ [1, 2m ],
1
z1,i
d) ρ1
T
≥ 0, i ∈ [1, m],
z1,i Q

1 z2,i
e) T ≥ 0, i ∈ [1, m],
z2,i Q

λ ET
f) ≥ 0.
E Q
If we fix ρ1 , λ and η, then the constraints of the optimization problem

(10.3.11) become LMIs. To obtain the global infimum, we may vary ρ1
from 0 to 1, η from 0 to ∞ and λ from 0 to 1+ηρ1 η
.
10.4. Conclusions
In this chapter, we have considered linear systems subject to actuator sat-
uration and disturbances. Conditions for determining if a given ellipsoid
is strictly invariant were derived. With the aid of these conditions, we
developed analysis and design methods, both for enlarging the invariant el-
lipsoid and for disturbance rejection. Examples were used to demonstrate
the effectiveness of these methods.
Chapter 11
On Maximizing the
Convergence Rate
11.1. Introduction
Fast response is always a desired property for control systems. The time
optimal control problem was formulated for this purpose. Although it is
well known that the time optimal control is a bang-bang control, this control
strategy is rarely implemented in real systems. The main reason is that it
is generally impossible to characterize the switching surface. For discrete-
time systems, online computation has been proposed in the literature, but
the computation burden is very heavy since linear programming has to be
solved recursively with increasing time-horizon. Also, as the time-horizon
is extended, numerical problems become more severe. Another reason is
that even if the optimal control can be obtained exactly and efficiently, it
results in open-loop control.
Another notion related to fast response is the convergence rate. For
linear systems, the overall convergence rate is measured by the maximal
real part of the closed-loop eigenvalues. Consider the linear system
ẋ = Ax, x ∈ Rn . (11.1.1)
Assume that the system is asymptotically stable. Let

α = − max Re(λi (A)) : i = 1, 2, · · · , n ,
229
230 Chapter 11. On Maximizing the Convergence Rate
where Re(λi (A)) is the real part of the ith eigenvalue of A, then α > 0.
For simplicity, assume that A + αI is neutrally stable, then there exists a
positive definite matrix P such that
AT P + P A ≤ −2αP.
(If A + αI is semi-stable but not neutrally stable, we can replace α in the

above equation with a number α1 < α arbitrarily close to α.) Also, there
exists a nonzero x ∈ Rn such that
xT (AT P + P A)x = −2αxT P x. (11.1.2)
Let
V (x) = xT P x.
Then,
V̇ (x) = xT (AT P + P A)x ≤ −2αxT P x = −2αV (x).
Hence,
V̇ (x)
≤ −2α, ∀ x ∈ Rn \ {0}. (11.1.3)
V (x)
Furthermore, because of (11.1.2), we have

1 V̇ (x)
α = min − : x ∈ R \ {0} .
n
(11.1.4)
2 V (x)
From (11.1.3), we obtain
V (x(t)) ≤ e−2αt V (x0 ).
Therefore, we call α the overall convergence rate of the system (11.1.1). For
a general nonlinear system, the convergence rate can be defined similarly
as in (11.1.4). Since local stability is a more general property than global
stability, we would like to define the term convergence rate on some subset
of the domain of attraction. Consider a nonlinear system
ẋ = f (x).
Assume that the system is asymptotically stable. Given a Lyapunov func-

tion V (x), let LV (ρ) be a level set

LV (ρ) = x ∈ Rn : V (x) ≤ ρ .
11.1. Introduction 231
Suppose that V̇ (x) < 0 for all x ∈ LV (ρ) \ {0}. Then, the overall conver-
gence rate of the system on LV (ρ) can be defined as

1 V̇ (x)
α := inf − : x ∈ LV (ρ) \ {0} . (11.1.5)
2 V (x)
For a discrete-time system
x(k + 1) = f (x(k)),
we can define the overall convergence rate on LV (ρ) as

1 ∆V (x)
α := inf − : x ∈ LV (ρ) \ {0} ,
2 V (x)
where ∆V (x) is the increment of V (x) along the trajectory of the system.
We can also define the convergence rate at each point x in the state space
as % &
V̇ (x) ∆V (x)
− or − .
V (x) V (x)
For linear systems subject to actuator saturation, the problem of max-
imizing the convergence rate at each x ∈ Rn is well defined. The objective
of this chapter is to find a control law constrained by the actuator satura-
tion such that −V̇ (x) is maximized at each x. It turns out that the optimal
control law is a bang-bang type control with a simple switching scheme.
Since the discontinuity might be undesirable, for example, causing chatter-
ing around the switching surface, we will also derive a continuous control
law with some loss of optimality in the convergence rate. The proposed
continuous control law is a saturated high gain linear feedback. As the gain
goes to infinity, the saturated high gain feedback approaches the optimal
bang-bang control law.
For a discrete-time system, the control law that maximizes the conver-
gence rate is a coupled saturated linear feedback. If the system has one or
two inputs, the control law can be put into a simple formula. If the system
has more than two inputs, the controller is more complicated. It is linear
inside some polyhedron. Outside of this polyhedron, we need to solve a
simple convex optimization problem.
We should note that under the maximal convergence control, the de-
crease of V (x) over a fixed time interval needn’t be maximal. This is be-
cause we don’t take the path of a trajectory into consideration. However, a
very important consequence of the maximal convergence control is that it

produces the maximal invariant ellipsoid of a given shape (in the absence
or in the presence of disturbances). It is easy to see that an ellipsoid can
be made invariant if and only if the maximal V̇ (x) (or minimal −V̇ (x))
on the boundary of the ellipsoid under the maximal convergence control is
less than 0 (or greater than 0). As we have discussed in Chapters 7 and
10, invariant ellipsoids are used to estimate the domain of attraction and
to study disturbance rejection capability of the closed-loop system. In this
chapter, we will present a simple method for checking if an ellipsoid can be
made invariant and for determining the largest ellipsoid that can be made
invariant.
In Chapter 8, we presented a method for enlarging the invariant ellipsoid
with respect to some shape reference set. As a result, the closed-loop system
behaves linearly in the ellipsoid. In other words, the feedback control does
not saturate in the ellipsoid. In this chapter, the control that achieves
maximal convergence rate and maximal invariant ellipsoid saturates almost
everywhere in the ellipsoid (for continuous-time systems). This seemingly
contradiction can be explained with the result in Chapter 7, where it was
shown that for a single input system, the largest invariant ellipsoid is in
some sense independent of a particular stabilizing controller. Although
both methods in Chapter 8 and in this chapter produce large invariant
ellipsoids, the focuses of the two chapters are different. Chapter 8 puts
the optimization problems into LMI framework and makes it very easy to
choose the shape of the ellipsoid. This chapter assumes that the shape of
the ellipsoid is given, say, produced by the method of Chapter 8, and tries
to maximize the convergence rate and to find the maximal invariant set of
the given shape.
This chapter is organized as follows. Section 11.2 and Section 11.3
study the maximal convergence control problems for continuous-time and
discrete-time systems, respectively. In particular, Section 11.2.1 shows that
the maximal convergence control is a bang-bang type control with a simple
switching scheme and that it produces the maximal invariant ellipsoid of
a given shape. A method for determining the largest ellipsoid that can
be made invariant with a bounded control is also given in this section.
Section 11.2.2 presents a saturated high gain feedback strategy to avoid
the discontinuity of the bang-bang control. An example is included in
Section 11.2.2 to illustrate the effectiveness of high gain feedback control.
Section 11.2.3 reveals some properties and limitations about the overall
convergence rate and provides methods to deal with these limitations. Sec-
tion 11.2.4 shows that the maximal convergence control also achieves both
the maximal and the minimal invariant ellipsoids in the presence of distur-
bances. A brief concluding remark is made in Section 11.4.

11.2.1. Maximal Convergence Control and Maximal
Invariant Ellipsoid
Consider a linear system subject to actuator saturation,
ẋ = Ax + Bu, x ∈ Rn , u ∈ Rm , |u|∞ ≤ 1. (11.2.1)
Assume that the system is stabilizable and that B has full column rank.
Let

B = b1 b2 · · · bm .
In Chapter 8, we developed a design method for enlarging the domain
of attraction under a state feedback. The approach was to maximize the
invariant ellipsoid with repect to some shape reference set. The optimized
controller, however, results in very slow convergence rate of the closed-
loop system. The objective of this chapter is to design a controller that
maximizes the convergence rate inside a given ellipsoid. Given a positive
definite matrix P , let
V (x) = xT P x.
For a positive number ρ, the level set associated with V (x) is the ellipsoid,

E(P, ρ) = x ∈ Rn : xT P x ≤ ρ .
Along the trajectory of the system (11.2.1),
V̇ (x, u) = 2xT P (Ax + Bu)

m
= xT (AT P + P A)x + 2 xT P bi ui .
i=1
Under the constraint that |u|∞ ≤ 1, the control that maximizes the con-
vergence rate, or minimizes V̇ (x, u), is simply
ui = −sign(bTi P x), i = 1, 2, · · · , m, (11.2.2)

where sign : R → R is the sign function. Under this bang-bang control, we

have

m
V̇ (x) = xT (AT P + P A)x − 2 xT P bi sign(bTi P x).
i=1
Now consider the closed-loop system

m
ẋ = Ax − bi sign(bTi P x). (11.2.3)
i=1
Because of the discontinuity of the sign function, equation (11.2.3) may have
no solution for some x(0) or have solution only in a finite time interval. For
example, for the single input case m = 1, equation (11.2.3) will have no
solution if B T P x(0) = 0 and ẋ at each side of the switching plane B T P x = 0
points to the other side. We will use a continuous feedback law,
ui = −sat(kbTi P x),
where sat : R → R is the standard saturation function, to approximate the

bang-bang control
ui = −sign(bTi P x)
in the next section. In what follows, we use the bang-bang control law to
investigate the possibility that an ellipsoid can be made invariant with a
bounded control |u|∞ ≤ 1.
Recall that an ellipsoid E(P, ρ) is invariant for a system ẋ = f (x) if
all the trajectories starting from it will stay inside of it. It is contractive
invariant if
V̇ (x) = 2xT P f (x) < 0, ∀ x ∈ E(P, ρ) \ {0}.
Since the bang-bang control
ui = −sign(bT P x)
minimizes V̇ (x, u) at each x, we have the following obvious fact.
Fact 11.2.1. The following two statements are equivalent:
a) The ellipsoid E(P, ρ) can be made contractive invariant for (11.2.1)

with a bounded control |u|∞ ≤ 1;
b) The ellipsoid E(P, ρ) is contractive invariant for (11.2.3), i.e., the fol-
lowing condition is satisfied,

m
V̇ (x) = xT (AT P + P A)x − 2 xT P bi sign(bTi P x) < 0,
i=1
∀ x ∈ E(P, ρ) \ {0}. (11.2.4)
It is clear from Fact 11.2.1 that the maximal convergence control pro-
duces the maximal invariant ellipsoid of a given shape. We will see in
the next section that if (11.2.4) is satisfied, there also exists a continuous
feedback law such that E(P, ρ) is contractive invariant. In this case, all the
trajectories starting from E(P, ρ) will converge to the origin asymptotically.
For an arbitrary positive definite matrix P , there may exists no ρ such
that E(P, ρ) can be made invariant. In what follows we give condition for P
such that E(P, ρ) can be made invariant for some ρ and provide a method
for finding the largest ρ.
Proposition 11.2.1. For a given positive definite matrix P , the following

three statements are equivalent:
a) There exists a ρ > 0 such that (11.2.4) is satisfied;
b) There exists an F ∈ Rm×n such that
(A + BF )T P + P (A + BF ) < 0; (11.2.5)
c) There exists a k > 0 such that
(A − kBB T P )T P + P (A − kBB T P ) < 0. (11.2.6)
Proof. b)→ a). If (11.2.5) is satisfied, then there exists a ρ > 0 such that

E(P, ρ) ⊂ x ∈ Rn : |F x|∞ ≤ 1 .
If x0 ∈ E(P, ρ), then under the control u = F x, x(t) will stay in E(P, ρ) and
we also have |u|∞ ≤ 1 for all t ≥ 0. This means that E(P, ρ) can be made
contractive invariant with a bounded control. Hence by the equivalence of
the statements a) and b) in Fact 11.2.1, we have (11.2.4).
c) → b). It is obvious.
a) → c). Let us assume that

0
PB = ,
R
where R is an m × m nonsingular matrix. If not so, we can use a state

transformation, x̄ = T x, with T nonsingular such that
P → P̄ = (T −1 )T P T −1 ,
B → B̄ = T B
and
0
P B → P̄ B̄ = (T −1 )T P B = .
R
Recall that we have assumed that B has full column rank. Also, let us
accordingly partition x as
x1
,
x2
and AT P + P A and P as

Q1 Q12 P1 P12
AT P + P A = , P = .
QT12 Q2 T
P12 P2
For all
x1
∈ ∂E(P, ρ),
0
we have xT P B = 0. So, if a) is true, then (11.2.4) holds for some ρ > 0,
which implies that
xT1 Q1 x1 < 0,
for all x1 such that xT1 P1 x1 = ρ. It follows that Q1 < 0. Hence there exists
a k > 0 such that

Q1 Q12
(A − kBB T P )T P + P (A − kBB T P ) = < 0.
QT12 Q2 − kRRT
This shows that c) is true. 2

Suppose that we are given a shape of ellipsoid, characterized by P0 > 0,
and that the maximal convergence rate is desired with respect to V (x) =
xT P0 x, but this shape cannot be made invariant for any size, then we
can take E(P0 , 1) as a shape reference set and find another invariant set
E(P, ρ) such that αE(P0 , 1) ⊂ E(P, ρ) with α maximized by the method in
Chapter 8. The shape of the resulting invariant ellipsoid E(P, ρ) will be the
closest to that of E(P0 , 1).
Now assume that we have a P > 0 such that the conditions in Proposi-
tion 11.2.1 are satisfied. Given ρ > 0, we would like to determine if E(P, ρ)
is contractive invariant for the closed-loop system (11.2.3). Let’s start with
the single input case. In this case, the condition (11.2.4) simplifies to
V̇ (x) = xT (AT P + P A)x − 2xT P Bsign(B T P x) < 0,

∀ x ∈ E(P, ρ) \ {0}. (11.2.7)
We claim that (11.2.7) is equivalent to
xT (AT P + P A)x − 2xT P Bsign(B T P x) < 0, ∀ x ∈ ∂E(P, ρ). (11.2.8)
To see this, we consider kx for k ∈ (0, 1] and x ∈ ∂E(P, ρ). Suppose

that
xT (AT P + P A)x − 2xT P Bsign(B T P x) < 0.
Since
−2xT P Bsign(B T P x) ≤ 0,
we have
2xT P B
xT (AT P + P A)x − sign(B T P x) < 0, ∀ k ∈ (0, 1].
k
Therefore,
(kx)T (AT P + P A)(kx) − 2(kx)T P Bsign(B T P kx)

% &
2xT P B
= k 2 xT (AT P + P A)x − sign(B T P x)
k
< 0,
for all k ∈ (0, 1]. This shows that the condition (11.2.7) is equivalent
to (11.2.8). Based on this equivalence property, we have the following
necessary and sufficient condition for the contractive invariance of a given
ellipsoid.
Theorem 11.2.1. Assume that m = 1. Suppose that E(P, ρ) can be made

contractive invariant for some ρ > 0. Let λ1 , λ2 , · · · , λJ > 0 be real num-
bers such that

λj P − AT P − P A P
det =0 (11.2.9)
ρ−1 P BB T P λj P − AT P − P A
and
B T P (AT P + P A − λj P )−1 P B > 0. (11.2.10)
Then, E(P, ρ) is contractive invariant for the system (11.2.3) if and only if
λj ρ − B T P (AT P + P A − λj P )−1 P B < 0, ∀ j = 1, 2, · · · , J .
If there exists no λj > 0 satisfying (11.2.9) and (11.2.10), then E(P, ρ) is

contractive invariant.
In the proof of Theorem

11.2.1, we will use the following algebraic fact.
X1 X2
Suppose that X1 , X4 and are square matrices. If X1 is non-
X3 X4
singular, then

X1 X2
det = det (X1 )det (X4 − X3 X1−1 X2 ), (11.2.11)
X3 X4
and if X4 is nonsingular, then

X1 X2
det = det (X4 )det (X1 − X2 X1−1 X3 ). (11.2.12)
X3 X4
Proof of Theorem 11.2.1. Denote
g(x) = xT (AT P + P A)x − 2xT P B.
By the equivalence of (11.2.7) and (11.2.8), the contractive invariance of

E(P, ρ) is equivalent to

max g(x) : B T P x ≥ 0, xT P x = ρ < 0. (11.2.13)
Since E(P, ρ) can be made contractive invariant for some ρ > 0, we must
have g(x) < 0 for all B T P x = 0. In this case, the contractive invariance of
E(P, ρ) is equivalent to that all the extrema of g(x) in the surface xT P x = ρ,
B T P x > 0, if any, are less than zero.
By the Lagrange multiplier method, an extremum of g(x) in the surface
x P x = ρ, B T P x > 0, must satisfy
T
(AT P + P A − λP )x = P B, xT P x = ρ, xT P B > 0, (11.2.14)
for some real number λ. And at the extremum, we have
g(x) = λρ − xT P B.
If λ ≤ 0, then g(x) < 0 since xT P B > 0. So we only need to consider λ > 0.

Now suppose that λ > 0. From
(AT P + P A − λP )x = P B,
we conclude that
det (AT P + P A − λP )
= 0.
To show this, we assume, without loss of generality, that

Q1 Q12 P1 P12 0
AT P + P A = , P = , P B = ,
QT12 q2 T
P12 p2 r
as in the proof of Proposition 11.2.1, it follows then that Q1 < 0. Since

λ > 0, Q1 < 0 and P1 > 0, Q1 − λP1 is nonsingular. Let

x1
x= , x2 ∈ R
x2
and suppose that x

= 0 satisfies
(AT P + P A − λP )x = P B,
then,
x1 = −(Q1 − λP1 )−1 (Q12 − λP12 )x2 ,
and

− (QT12 − λP12
T
)(Q1 − λP1 )−1 (Q12 − λP12 ) + q2 − λp2 x2 = r.
Multiplying both sides with det (Q1 − λP1 ) and applying (11.2.11), we ob-
tain
det (AT P + P A − λP )x2 = det (Q1 − λP1 )r.
Since r
= 0 and det (Q1 − λP1 )
= 0, we must have
det (AT P + P A − λP )
= 0.
So for all λ > 0 and x satisfying (11.2.14), we have
x = (AT P + P A − λP )−1 P B,
and hence from xT P x = ρ,
B T P (AT P + P A − λP )−1 P (AT P + P A − λP )−1 P B = ρ. (11.2.15)

Denote
Φ = λP − AT P − P A,
then the equation (11.2.15) can be written as,
B T P Φ−1 P Φ−1 P B = ρ.
By invoking (11.2.11) and (11.2.12), we obtain

ρ −B T P Φ−1
det =0
−Φ−1 P B P −1

% &
ρ 0 T
B P 0 Φ−1 0 0 I
det − =0
0 P −1 0 I 0 Φ−1 PB 0

% &
Φ 0 0 I ρ−1 0 BTP 0
det − =0
0 Φ PB 0 0 P 0 I

λP − AT P − P A P
det = 0.
ρ−1 P BB T P λP − AT P − P A
This last equation is (11.2.9).
Also, at the extremum, we have xT P B > 0. This is equivalent to
(11.2.10),
B T P (AT P + P A − λP )−1 P B > 0.
Finally, at the extremum
g(x) = xT (AT P + P A)x − 2xT P B

= λρ − B T P (AT P + P A − λP )−1 P B.
Hence the result of the theorem follows. 2

Here we note that all the λj ’s satisfying (11.2.9) are the eigenvalues of
the matrix
−1 T 1 1 1
P 2 A P 2 + P 2 AP − 2 −I
1 1 1 1 1 1 .
−ρ−1 P 2 BB T P 2 P − 2 AT P 2 + P 2 AP − 2
Hence the condition of Theorem 11.2.1 can be easily checked.

Recall that the condition (11.2.7) is equivalent to (11.2.8). This implies
that there is a ρ∗ > 0 such that E(P, ρ) is contractive invariant if and only
if ρ < ρ∗ . Therefore, the maximum value ρ∗ can be obtained by checking

the condition of Theorem 11.2.1 bisectionally.
For systems with multiple inputs, we may divide the state space into
cones: xT P bi < 0 (> 0, or = 0), and check the maximum value of

m
V̇ (x) = x (A P + P A)x − 2
T T
xT P bi sign(bTi P x)
i=1
within each cone. For example, consider m = 2, the surface of the ellipsoid
E(P, ρ) can be divided into the following subsets:

S1 = x ∈ Rn : bT1 P x = 0, bT2 P x ≥ 0, xT P x = ρ , −S1 ,

S2 = x ∈ Rn : bT1 P x ≥ 0, bT2 P x = 0, xT P x = ρ , −S2 ,

S3 = x ∈ Rn : bT1 P x > 0, bT2 P x > 0, xT P x = ρ , −S3 ,

S4 = x ∈ Rn : bT1 P x > 0, bT2 P x < 0, xT P x = ρ , −S4 .
With this partition, E(P, ρ) is contractive invariant if and only if
max V̇ (x) < 0, max V̇ (x) < 0, (11.2.16)

x∈S1 x∈S2
and all the local extrema of V̇ (x) in S3 and S4 are negative.

In S1 ,
V̇ (x) = xT (AT P + P A)x − 2xT P b2 .
Let N ∈ Rn×(n−1) be a matrix of rank n − 1 such that bT1 P N = 0, i.e.,
{N y : y ∈ Rn−1 } is the kernel of bT1 P . The constraint bT1 P x = 0 can be
replaced by x = N y, y ∈ Rn−1 . Thus,

max V̇ (x) = max y T N T (AT P + P A)N y − 2y T N T P b2 :
x∈S1

bT2 P N y ≥ 0, y T N T P N y = ρ .
This is similar to the optimization problem (11.2.13) in the proof of Theo-

rem 11.2.1 except with a reduced order. The second optimization problem
in (11.2.16) can be handled in the same way.
In S3 ,
V̇ (x) = xT (AT P + P A)x − 2xT P (b1 + b2 ).
All the local extrema of V̇ (x) in S3 (and in S4 ) can be obtained like those
of g(x) in the proof of Theorem 11.2.1.
11.2.2. Saturated High Gain Feedback
With the bang-bang type control law (11.2.2), V (x) of the closed-loop sys-
tem will decrease with a maximal convergence rate. As we have noted in the
last section, the discontinuity of the bang-bang control law may cause the
state equation to have no solution. When the control law is implemented
on a sampled-data system, it may cause frequent switching (chattering)
around the switching surface. In this section, we will replace the bang-
bang control law with a saturated high gain linear feedback at the cost of
a slight reduction in convergence rate. We also start with the single input
case.
Theorem 11.2.2. Assume that m = 1. Suppose that E(P, ρ) can be made

contractive invariant with a bounded control. If
(A − k0 BB T P )T P + P (A − k0 BB T P ) < 0, (11.2.17)
then E(P, ρ) is contractive invariant under the saturated linear control
u = −sat(kB T P x),
for all k ≥ k0 .
This theorem says that the saturated linear control u = −sat(kB T P x)

can produce the same invariant set as the bang-bang control law as long
as (11.2.17) is satisfied and k > k0 . As has been pointed out in Propo-
sition 11.2.1, there always exists a k > 0 satisfying (11.2.17) under the
condition that E(P, ρ) can be made contractive invariant. So the invariance
of E(P, ρ) does not require high gain feedback. But high gain is necessary
for achieving fast convergence rate.
Proof of Theorem 11.2.2. Suppose that E(P, ρ) can be made contractive
invariant, then by the equivalence of statements a) and b) in Fact 11.2.1,
xT (AT P +P A)x−2xT P Bsign(B T P x) < 0, ∀ x ∈ E(P, ρ)\{0}. (11.2.18)
If |B T P x| ≥ k1 , then
sat(kB T P x) = sign(B T P x).
xT (AT P + P A)x − 2xT P Bsat(kB T P x) < 0 (11.2.19)

for x ∈ E(P, ρ) \ {0} such that |B T P x| ≥ k1 .

If |B T P x| < k1 , then
sat(kB T P x) = kB T P x.
Since k ≥ k0 , it follows from (11.2.17) that
(A − kBB T P )T P + P (A − kBB T P ) < 0.
Hence we also obtain (11.2.19). This shows that E(P, ρ) is contractive

invariant under the control
u = −sat(kB T P x).
Theorem 11.2.2 can also be proved by applying Theorem 7.4.2.

Let us now compare the convergence rates under the two controls
u = −sat(kB T P x)
and
u = −sign(B T P x).
The difference between the two V̇ (x)’s under these two controls is
 1

 = 0, if |B T P x| > ,
T k
2x P B sign(B P x) − sat(kB P x)
T T

 ≤ 2 , if |B T P x| ≤ 1
.
k k
Note that |sign(B T P x) − sat(kB T P x)| ≤ 1. By letting k → ∞, the differ-
ence between the two V̇ (x)’s will go to zero uniformly on E(P, ρ). Thus we
can say that high gain saturated linear feedback will produce sub-optimal
convergence rate.
We now consider the multiple input case. For v ∈ Rm , denote
T
sat(v) = sat(v1 ) sat(v2 ) · · · sat(vm ) .
Here, as usual, we have slightly abused the notation by using sat to denote
both the scalar and the vector saturation functions. We have the following
result.
Theorem 11.2.3. For a multiple input system, suppose that E(P, ρ) can
be made contractive invariant with a bounded control, then there exists a
k > 0 such that E(P, ρ) is contractive invariant under the control
Proof. The condition that E(P, ρ) can be made contractive invariant im-
plies that

m
xT (AT P + P A)x − 2 xT P bi sign(bTi P x) < 0,
i=1
∀ x ∈ E(P, ρ) \ {0}. (11.2.20)
We need to show that there exists a k > 0 such that

m
xT (AT P + P A)x − 2 xT P bi sat(kbTi P x) < 0,
i=1
∀ x ∈ E(P, ρ) \ {0}. (11.2.21)
First, we will show that (11.2.21) is equivalent to

m
xT (AT P + P A)x − 2 xT P bi sat(kbTi P x) < 0,
i=1
∀ x ∈ ∂E(P, ρ). (11.2.22)
If this is true, then we only need to consider x ∈ ∂E(P, ρ). Obviously,

(11.2.22) is implied by (11.2.21). We need to prove the converse.
Suppose that x ∈ ∂E(P, ρ) satisfies (11.2.22). For γ > 0, define

m
sat(kbTi P γx)
g(γ) = xT (AT P + P A)x − 2 xT P bi ,
i=1
γ
then g(1) < 0. Since

xT P bi sat(kbTi P x) ≥ 0
and
sat(kbTi P γx)

γ
decreases as γ increases, it follows that g(γ) increases as γ increases. Hence,
g(γ) ≤ g(1) < 0, ∀ γ ∈ (0, 1].

Therefore, for all γ ∈ (0, 1],

m
(γx)T (AT P + P A)(γx) − 2 (γx)T P bi sat(kbTi P γx) = γ 2 g(γ) < 0.
i=1
This shows that (11.2.21) is implied by (11.2.22).

Let

m
ε = − max x (A P + P A)x − 2
T T T T
x P bi sign(bi P x) .
x∈∂ E (P,ρ)
i=1
Then, from (11.2.20), we have ε > 0. For all x ∈ ∂E(P, ρ),

m
xT (AT P + P A)x − 2 xT P bi sat(kbTi P x)
i=1

m
= xT (AT P + P A)x − 2 xT P bi sign(bTi P x)
i=1

m
+2 xT P bi sign(bTi P x) − sat(kbTi P x)
i=1

m
≤ −ε + 2 xT P bi sign(bTi P x) − sat(kbTi P x) . (11.2.23)
i=1
Since
 1

 = 0, if |bTi P x| > ,
T k
x P bi sign(bi P x) − sat(kbi P x)
T T

 ≤ 1 , if |bT P x| ≤ 1
i ,
k k
we have m
2m

2 x P bi sign(bi P x) − sat(kbi P x) ≤
T T T
.
k
i=1
2m
If we choose k > ε , then from (11.2.23),

m
xT (AT P + P A)x − 2 xT P bi sat(kbTi P x) < 0, ∀ x ∈ ∂E(P, ρ),
i=1
which is (11.2.22). It follows from the equivalence of (11.2.21) and (11.2.22)

that E(P, ρ) is contractive invariant under the control
2

0.6 −0.8 2
A= , B= .
0.8 0.6 4
Let
0.0836 −0.0639
P = .
−0.0639 0.1460
By checking the condition of Theorem 11.2.1 bisectionally, the largest el-
lipsoid that can be made contractive invariant with a bounded control is
E(P, ρ∗ ) with ρ∗ = 1.059. By using the design method in Chapter 8, a
feedback

u = sat(F0 x), F0 = 0.0036 −0.3057
is found such that
(A + BF0 )T P + P (A + BF0 ) < 0,
with E(P, ρ), ρ = 1.058 inside the linear region of the saturation

L(F0 ) = x ∈ R2 : |F0 x|∞ ≤ 1
(see Fig. 11.2.1). The eigenvalues of A + BF0 are −0.0078 ± j0.8782. It

can be expected that the convergence rate under this feedback is very slow.
The convergence rate is accelerated by using a saturated high gain feedback
u = −sat(5B T P x).
This is illustrated in Figs. 11.2.1 and 11.2.2. In Fig. 11.2.1, “∗” represents
the initial state, the solid trajectory is under the control of u = sat(F0 x) and
the dash-dotted one is under the control of u = −sat(5B T P x). Fig. 11.2.2
shows V (x) = xT P x as a function of time. Also, the solid one is under the
control of u = sat(F0 x) and the dash-dotted one is under the control of
From Fig. 11.2.2, we see that the decrease of xT P x becomes slower and
slower under the control
This will be the case even if we increase the gain k in
F0x=−1
0
T
5B Px=1
−1
T
x Px=1.058
−2
−3
−4
−8 −6 −4 −2 0 2 4 6 8
Figure 11.2.1: Comparison of the trajectory convergence rates.
to infinity. Actually, the overall convergence rate under the bang-bang

control
ui = −sign(bTi P x), i = 1, 2, · · · , m,
is limited by the shape of the ellipsoid or the matrix P . This problem will
be discussed in the next section.
11.2.3. Overall Convergence Rate
We also consider the system (11.2.3) under the maximal convergence con-
trol,
m
ẋ = Ax − bi sign(bTi P x). (11.2.24)
i=1
Assume that E(P, ρ) is contractive invariant for (11.2.24). We would like

to know the overall convergence rate in E(P, ρ). We will see later that as
ρ decreases (note that a trajectory goes into smaller E(P, ρ) as time goes
by), the overall convergence rate increases but is limited by the shape of
E(P, ρ). This limit can be raised by choosing P properly.
1.4
1.2
1
u=sat(F0x)
xTPx 0.8
0.6
0.4
u=−sat(5BTPx)
0.2
0
0 2 4 6 8 10 12 14 16 18 20
t
Figure 11.2.2: Comparison of the convergence rates of xT P x.
The overall convergence rate, denoted by α, is defined by (11.1.5) in the

introduction. Here we would like to examine its dependence on ρ, so we
write,
1 V̇ (x)
α(ρ) := inf − : x ∈ E(P, ρ) \ {0} .
2 V (x)
The main results of this subsection are contained in the following theorem.
Theorem 11.2.4.
a)
1 V̇ (x)
α(ρ) = min − T
: x Px = ρ ; (11.2.25)
2 ρ
b) α(ρ) increases as ρ decreases;
c) Let

β0 = min − xT (AT P + P A)x : xT P x = 1, xT P B = 0 ,
then,
β0
lim α(ρ) = . (11.2.26)
ρ→0 2
Proof.
a) Consider x ∈ ∂E(P, ρ) and k ∈ (0, 1],

V̇ (kx) k 2 xT (AT P + P A)x − 2k m T T
i=1 x P bi sign(bi P kx)
− =−
V (kx) k 2 xT P x
2
m
−x (A P + P A)x + k i=1 xT P bi sign(bTi P x)
T T
= . (11.2.27)
xT P x
Since

m
xT P bi sign(bTi P x) ≥ 0,
i=1
(kx) (x)
− VV̇ (kx) increases as k decreases. It follows that the minimal value of − VV̇ (x)
is obtained on the boundary of E(P, ρ), which implies (11.2.25).
b) This follows from the proof of a).
c) From a), we see that

T T m
−x (A P +P A)x+2 i=1 xT P bi sign(bTi P x) T
2α(ρ) = min : x P x=ρ
ρ

2 T
m
= min − xT (AT P + P A)x + √ x P bi sign(bTi P x) :
ρ i=1

xT P x = 1

2 T
m
≤ min −xT (AT P + P A)x + √ x P bi sign(bTi P x) :
ρ i=1

T T
x P x = 1, x P B = 0

= min − xT (AT P + P A)x : xT P x = 1, xT P B = 0
= β0 .
It follows then 2α(ρ) ≤ β0 for all ρ > 0. To prove (11.2.26), it suffices to

show that given any ε > 0, there exists a ρ > 0 such that
2α(ρ) ≥ β0 − ε.
Denote
X 0 = x ∈ Rn : xT P x = 1, xT P B = 0 ,
and
X (δ) = x ∈ Rn : xT P x = 1, |xT P B|∞ ≤ δ .
It is clear that
lim dist(X (δ), X 0 ) = 0,
δ→0
where dist(·, ·) is the Hausdorff distance. By the uniform continuity of

xT (AT P + P A)x on the surface {x ∈ Rn : xT P x = 1}, we have that, given
any ε, there exists a δ > 0 such that

min − xT (AT P + P A)x : xT P x = 1, |xT P B|∞ ≤ δ ≥ β0 − ε.
(11.2.28)
Since

m
xT P bi sign(bTi P x) ≥ 0,
i=1
we have,

2 T
m
min −xT (AT P + P A)x + √ x P bi sign(bTi P x) :
ρ i=1

xT P x = 1, |xT P B|∞ ≤ δ ≥ β0 − ε, (11.2.29)
for all ρ > 0.

Let

β1 = min − xT (AT P + P A)x : xT P x = 1, |xT P B|∞ ≥ δ .
If β1 ≥ β0 − ε, then for all ρ > 0,

2 T
m
ρ i=1

xT P x = 1, |xT P B|∞ ≥ δ ≥ β1 ≥ β0 − ε.
Combining the above with (11.2.29), we have

2 T
m
2α(ρ) = min −xT (AT P + P A)x + √ x P bi sign(bTi P x) :
ρ i=1

xT P x = 1 ≥ β0 − ε. (11.2.30)
for all ρ > 0. This shows that
2α(ρ) ≥ β0 − ε, ∀ρ > 0.
If β1 < β0 − ε, then for

% &2
2δ
ρ< ,
−β1 + β0 − ε
we have

2 T
m
ρ i=1

xT P x = 1, |xT P B|∞ ≥ δ
2δ
≥ β1 + √
ρ
> β0 − ε.
Combining the above with (11.2.29), we also obtain (11.2.30) and
2α(ρ) ≥ β0 − ε
for % &2
2δ
ρ< .
−β1 + β0 − ε
This completes the proof. 2
Theorem 11.2.4 says that α(ρ) can be obtained by computing the max-
imum of V̇ (x) over ∂E(P, ρ). For the single input case, Theorem 11.2.1
provides a method for determining if this maximum is negative. The exact
value of α(ρ) can be computed with a procedure similar to the proof of
Theorem 11.2.1. To avoid too much technical detail, we assume for sim-
plicity that, for any real eigenvalue λj of A + P −1 AT P , there exists no x
satisfying
(P A + AT P − λj P )x = P B.
This is the case if (A + P −1 AT P, B) is controllable, thus the assumption is
generally true. Let λ1 , λ2 , · · · , λJ be real numbers satisfying (11.2.9) and
(11.2.10). Denote
1
βj = −λj + B T P (AT P + P A − λj P )−1 P B,
ρ
then,
1
α(ρ) = min βj : j = 0, 1, · · · , J ,
2
where β0 is as defined in Theorem 11.2.4.
Since the overall convergence rate is limited by β0 /2 and it approaches
this limit as ρ goes to 0, we would like β0 not to be too small. For a fixed
P , a formula for computing β0 can be derived directly from the definition.
Let the kernel of B T P be {N y : y ∈ Rn−m }. Then

β0 = −λmax (N T P N )−1 N T (AT P + P A)N . (11.2.31)
Since N depends on P , the above formula gives us no hint for choosing P

to obtain a satisfactory β0 . The following proposition will lead to an LMI
approach to choosing P .
Proposition 11.2.2. Let P be given. Then,
β0 =sup λ (11.2.32)
F
s.t. (A + BF )T P + P (A + BF ) ≤ −λP.
Proof. Notice that, for any F , we have
xT ((A + BF )T P + P (A + BF ))x = xT (AT P + P A)x, ∀ xT P B = 0.
It follows that

β0 = min − xT ((A + BF )T P + P (A + BF ))x : xT P x = 1, xT P B = 0

≥ min − xT ((A + BF )T P + P (A + BF ))x : xT P x = 1 ,
and hence,

β0 ≥ sup min − xT ((A + BF )T P + P (A + BF ))x : xT P x = 1 .
F
(11.2.33)
In what follows, we will prove that

β0 = sup min − xT ((A + BF )T P + P (A + BF ))x : xT P x = 1 .
F
(11.2.34)
In view of (11.2.33), it suffices to show that for any ε > 0, there exists an
F = −kB T P , with k > 0, such that

min − xT ((A + BF )T P + P (A + BF ))x : xT P x = 1 ≥ β0 − ε.
(11.2.35)
From the definition of β0 , we see that there exists a δ > 0 such that

min − xT (AT P + P A)x : xT P x = 1, |xT P B|2 ≤ δ ≥ β0 − ε
(refer to (11.2.28). Since xT P BB T P x ≥ 0,

min − xT ((A − kBB T P )T P + P (A − kBB T P ))x :

xT P x = 1, |xT P B| ≤ δ
≥ β0 − ε, (11.2.36)
for all k > 0.

For every x ∈ Rn ,
−xT ((A − kBB T P )T P + (A − kBB T P ))x
is an increasing function of k. Similar to the proof of Theorem 11.2.4, it

can be shown that there exists a k > 0, such that

min − xT ((A − kBB T P )T P + P (A − kBB T P ))x :

xT P x = 1, |xT P B| ≥ δ
≥ β0 − ε. (11.2.37)
Combining (11.2.36) and (11.2.37), we have

min − xT ((A − kBB T P )T P + P (A − kBB T P ))x : xT P x = 1 ≥ β0 − ε.
This proves (11.2.35) and hence (11.2.34).

Denote

β(F ) = min − xT ((A + BF )T P + P (A + BF ))x : xT P x = 1 .

β0 = sup β(F ).
F
It can be shown that

β(F ) = max λ : (A + BF )T P + P (A + BF ) ≤ −λP .
This brings us to (11.2.32). 2

For a fixed P , β0 is a finite value given by (11.2.31). Assume that

(A, B) is controllable, then the eigenvalues of (A + BF ) can be arbitrarily
assigned. If we also take P as a variable, then −β0 /2 can be made equal to
the largest real part of the eigenvalues of A+BF (see the definition, as given
in Section 11.1, of the overall convergence rate for a linear system). This
means that β0 can be made arbitrarily large. But generally, as β0 becomes
very large, the matrix P will be badly conditioned, i.e., the ellipsoid E(P, ρ)
will become very thin in certain direction, and hence very “small”, with
respect to a fixed shape reference set. On the other hand, as mentioned in
Chapter 8, if our only objective is to enlarge the domain of attraction with
respect to a reference set, some eigenvalues of A + BF will be very close
to the imaginary axis, resulting in very small β0 . These two conflicting
objectives can be balanced, for example, by pre-specifying a lower bound
on β0 and then maximizing the invariant ellipsoid. This mixed problem can
be described as follows (by adding a new constraint d) to the optimization
problem (8.2.4)):
sup α (11.2.38)
P >0,ρ,F,H
s.t. a) αX R ⊂ E(P, ρ),
b) (A + BF )T P + P (A + BF ) < 0,
c) E(P, ρ) ⊂ L(F ),
d) (A + BH)T P + P (A + BH) ≤ −βP.
The last constraint d) is imposed to guarantee a lower bound β on

the convergence rate. By solving (11.2.38), we obtain the optimal ellip-
soid E(P, ρ) along with two feedback matrices F and H. We may actually
discard both F and H but instead use the bang-bang control law
ui = −sign(bTi P x), i = 1, 2, · · · , m, (11.2.39)
or the high gain controller
ui = −sat(kbTi P x), i = 1, 2, · · · , m. (11.2.40)
The final outcome is that under the control of (11.2.39) or (11.2.40), the
closed-loop system will have a contractive invariant set E(P, ρ) and a guar-
anteed limit of the convergence rate
β0 β
≥ .
2 2
11.2.4. Maximal Convergence Control in the Presence

of Disturbances
ẋ = Ax + Bu + Ew, x ∈ Rn , u ∈ Rm , w ∈ Rq , (11.2.41)
where u is the control bounded by |u|∞ ≤ 1 and w is the disturbance

bounded by |w|∞ ≤ 1. Note the difference between the bound on w and
that in Chapter 10. This is assumed for simplicity. We also consider the
quadratic Lyapunov function V (x) = xT P x.
Along the trajectory of the system (11.2.41),
V̇ (x, u, w) = 2xT P (Ax + Bu + Ew)

m
T T
= x (A P + P A)x + 2 xT P bi ui + 2xT P Ew.
i=1
No matter what w is, the control that maximizes the convergence rate, or
minimizes V̇ (x, u, w), is also
ui = −sign(bTi P x), i = 1, 2, · · · , m. (11.2.42)
Under this control, we have

m
V̇ (x, w) = xT (AT P + P A)x − 2 xT P bi sign(bTi P x) + 2xT P Ew. (11.2.43)
i=1
Now consider the closed-loop system

m
ẋ = Ax − bi sign(bTi P x) + Ew. (11.2.44)
i=1
An ellipsoid E(P, ρ) is invariant for the system (11.2.44) if V̇ (x, w) ≤ 0 for

all possible w bounded by |w|∞ ≤ 1. Unlike the system in the absence
of disturbance, this system does not possess the property of contractive
invariance. We may define strict invariance as in Chapter 10. Since the
invariance property can be easily extended to the strict invariance property
(e.g., by changing “≤” to “<”), we will only investigate the former.
Denote the ith column of E as ei . From (11.2.43), we see that, for
x ∈ Rn , the worst w that maximizes V̇ (x, w) is
wi = sign(eTi P x), i = 1, 2, · · · , q.
Thus, we have the following obvious fact.

Fact 11.2.2. The following three statements are equivalent:
a) The ellipsoid E(P, ρ) can be made invariant for (11.2.41) with a

bounded control |u|∞ ≤ 1;
b) The ellipsoid E(P, ρ) is invariant for (11.2.44);
c) The following condition is satisfied,

m
x (A P + P A)x − 2
T T
i=1

q
+ xT P ei sign(eTi P x) ≤ 0, ∀ x ∈ ∂E(P, ρ). (11.2.45)
i=1
Denote

m
V̇ (x) = x (A P + P A)x − 2
T T
i=1

q
+2 xT P ei sign(eTi P x).
i=1
Fact 11.2.2 says that E(P, ρ) is invariant for (11.2.44) if and only if V̇ (x) ≤ 0
for all x ∈ ∂E(P, ρ). In the following, when we say that E(P, ρ) is invariant,
we mean it is invariant for the system (11.2.44), or for the system (11.2.41)
under the control
ui = −sign(bTi P x).
Proposition 11.2.3. Suppose that E(P, ρ1 ) and E(P, ρ2 ) are both

(strictly) invariant and ρ1 < ρ2 . Then E(P, ρ) is (strictly) invariant for
all ρ ∈ [ρ1 , ρ2 ].
Proof. For a fixed x ∈ Rn , consider V̇ (kx), k ≥ 0, as a function of k.

Clearly, this is a quadratic function and we also have V̇ (0x) = 0. Hence,
if both V̇ (k1 x) and V̇ (k2 x) are negative, we must have V̇ (kx) negative for
all k between k1 and k2 . The result of the proposition follows readily from
this argument. 2
From Proposition 11.2.3, we see that all the ρ such that E(P, ρ) is in-
variant form an interval. By Fact 11.2.2, this is the largest interval we
can achieve with any feedback control constrained by the input saturation.
In other words, the maximal convergence control produces both the max-
imal invariant ellipsoid and the minimal invariant ellipsoid. As we have
mentioned in Chapter 10, a large invariant ellipsoid is desired for a large
domain of attraction, and a small invariant ellipsoid indicates a good ca-
pability of disturbance rejection. In summary, the maximal convergence
control is ideal for dealing with disturbances.
Similar to the system in the absence of disturbance, it can be shown
that the saturated high gain feedback
can be used to approximate the bang-bang control
ui = −sign(bTi P x), i = 1, 2, · · · , m,
and to produce a sub-optimal disturbance rejection performance.

The condition in Fact 11.2.2 c) can be verified by solving a set of opti-
mization problems. Consider the single input and single disturbance case,
m = q = 1. In this case, we have
V̇ (x) = xT (AT P + P A)x − 2xT P Bsign(B T P x) + 2xT P Esign(E T P x).
Similar to the two-input case in Section 11.2.1, we can divide the surface
∂E(P, ρ) into 8 subsets:

S1 = x ∈ R n : B T P x = 0, E T P x ≥ 0, xT P x = ρ , −S1 ,

S2 = x ∈ R n : B T P x ≥ 0, E T P x = 0, xT P x = ρ , −S2 ,

S3 = x ∈ R n : B T P x > 0, E T P x > 0, xT P x = ρ , −S3 ,

S4 = x ∈ R n : B T P x > 0, E T P x < 0, xT P x = ρ , −S4 .
With this partition, E(P, ρ) is invariant if and only if
max V̇ (x) ≤ 0, max V̇ (x) ≤ 0, (11.2.46)

x∈S1 x∈S2
and all the local extrema of V̇ (x) in S3 and S4 are non-positive. These
optimization problems can be handled similarly as those in Section 11.2.1.

Consider the discrete-time system
x(k + 1) = Ax(k) + Bu(k), x ∈ Rn , u ∈ Rm , |u|∞ ≤ 1. (11.3.1)
Assume that the system is stabilizable and that B has full column rank.
Given V (x) = xT P x, we would like to maximize the convergence rate within
E(P, ρ).
Define
∆V (x, u) = (Ax + Bu)T P (Ax + Bu) − xT P x.
Our objective is to minimize ∆V (x, u) under the constraint |u|∞ ≤ 1 for

every x ∈ E(P, ρ).
Since B has full column rank, B T P B is nonsingular and we have
∆V (x, u) = uT B T P Bu + 2uT B T P Ax + xT AT P Ax − xT P x
T
= u + (B T P B)−1 B T P Ax B T P B u + (B T P B)−1 B T P Ax
−xT AT P B(B T P B)−1 B T P Ax + xT AT P Ax − xT P x.
Let
F0 = −(B T P B)−1 B T P A.
It is clear that the convergence maximization problem is equivalent to
min (u − F0 x)T B T P B(u − F0 x). (11.3.2)

|u|∞ ≤1
Let us first consider the single input case. In this case, the control that
maximizes the convergence rate is simply,
u = sat(F0 x). (11.3.3)
This control is continuous in x, so it is better behaved than its continuous-

time counterpart.
x(k + 1) = Ax(k) + Bsat(F0 x). (11.3.4)
The system is asymptotically stable at the origin if
(A + BF0 )T P (A + BF0 ) − P < 0,

i.e.,
T
A − B(B T P B)−1 B T P A P A − B(B T P B)−1 B T P A − P < 0.
(11.3.5)
As in the continuous-time case, we also have
Fact 11.3.1. For the single input case, the following two statements are
equivalent:
a) The ellipsoid E(P, ρ) can be made contractive invariant for the system
(11.3.1) with a bounded control;
b) The ellipsoid E(P, ρ) is contractive invariant for (11.3.4), i.e., the fol-
lowing condition is satisfied,
(Ax + Bsat(F0 x))T P (Ax + Bsat(F0 x)) − xT P x < 0,

∀ x ∈ E(P, ρ) \ {0}. (11.3.6)
Proposition 11.3.1. For the single input case, the ellipsoid E(P, ρ) can be
made contractive invariant for some ρ > 0 if and only if (11.3.5) is satisfied.
Proof. There exists a ρ0 > 0 such that

E(P, ρ0 ) ⊂ x ∈ Rn : |F0 x|∞ ≤ 1 .
Suppose that (11.3.5) is satisfied. Let u = sat(F0 x), then u = F0 x for

x ∈ E(P, ρ0 ). Hence E(P, ρ0 ) is contractive invariant under u = sat(F0 x).
On the other hand, suppose that E(P, ρ) can be made contractive in-
variant. By Fact 11.3.1, we have (11.3.6). Let ρ1 = min{ρ0 , ρ}, then
sat(F0 x) = F0 x for x ∈ E(P, ρ1 ). It follows from (11.3.6) that (11.3.5) is
true. 2
Suppose that (11.3.5) is satisfied, we would like to obtain the condition
for E(P, ρ) to be invariant for (11.3.4) and to find the largest ρ such that
E(P, ρ) is invariant.
Because of (11.3.5), we have
(Ax + Bsat(F0 x))T P (Ax + Bsat(F0 x)) − xT P x < 0
for all

x ∈ x ∈ Rn : |F0 x| ≤ 1, x
= 0

= x ∈ Rn : |B T P Ax| ≤ B T P B, x
= 0 .
Hence, E(P, ρ) is invariant if and only if
(Ax − B)T P (Ax − B) − xT P x < 0,

∀ x ∈ E(P, ρ) such that B T P Ax ≥ B T P B. (11.3.7)
Like the equivalence of (11.2.7) and (11.2.8), it can be shown that (11.3.7)
is equivalent to
(Ax − B)T P (Ax − B) − xT P x < 0,

∀ x ∈ ∂E(P, ρ) such that B T P Ax ≥ B T P B. (11.3.8)
And we have
Theorem 11.3.1. For the single input case, assume that (11.3.5) is satis-
fied. Let λ1 , λ2 , · · · , λJ > 1 be real numbers such that

λj P − AT P A P
det =0 (11.3.9)
ρ−1 AT P BB T P A λj P − AT P A
and
B T P A(AT P A − λj P )−1 AT P B ≥ B T P B. (11.3.10)
Then, E(P, ρ) is contractive invariant for the system (11.3.4) if and only if
(λj − 1)ρ − B T P A(AT P A − λj P )−1 AT P B + B T P B < 0,

∀ j = 1, 2, · · · J . (11.3.11)
If there is no λj > 1 satisfying (11.3.9) and (11.3.10), then E(P, ρ) is con-

tractive invariant.
Proof. Denote
g(x) = (Ax − B)T P (Ax − B) − xT P x.
On the plane B T P Ax = B T P B, we have
Ax − B(B T P B)−1 B T P Ax = Ax − B.
Since P satisfies (11.3.5), we have g(x) < 0 for all x on the plane B T P Ax =
B T P B. So the invariance of E(P, ρ) is equivalent to that all the extrema in
the surface xT P x = ρ, B T P Ax > B T P B, if any, are less than zero.
Suppose that x is an extremum in the surface, then by Lagrange multi-

plier method, there exists a real number λ such that
(AT P A − λP )x = AT P B, xT P x = ρ, B T P Ax > B T P B. (11.3.12)
At the extremum, we have
g(x) = (λ − 1)ρ − B T P Ax + B T P B.
Since B T P Ax > B T P B, we have g(x) < 0 for λ ≤ 1. So we only need to

consider λ > 1. Assume that

0
AT P B =
r
and partition AT P A and P accordingly as

Q1 Q12 P1 P12
AT P A = , P = .
QT12 q2 T
P12 p2
It follows from (11.3.5) that Q1 − P1 < 0 and Q1 − λP1 < 0 for all λ > 1.
It can be shown as in the proof of Theorem 11.2.1 that (AT P A − λP )x =
AT P B and λ > 1 imply that AT P A − λP is nonsingular. Hence
x = (AT P A − λP )−1 AT P B,
and from xT P x = ρ,
B T P A(AT P A − λP )−1 P (AT P A − λP )−1 AT P B = ρ.
The rest of the proof is similar to that of Theorem 11.2.1. 2
The supremum of ρ such that E(P, ρ) is contractive invariant can be

obtained by checking the condition in Theorem 11.3.1 bisectionally.
Let us now consider the multiple input case and examine the optimiza-
tion problem (11.3.2). If B T P B is diagonal, then we also have
u = sat(F0 x),
as the optimal control. But for the general case, the optimal u cannot be
put into this simple form. Actually, (11.3.2) is a minimal distance problem.
Let
v = F0 x = −(B T P B)−1 B T P Ax,
then the optimization problem is to find a point in the box |u|∞ ≤ 1 that
is closest to v, with the distance induced by the weighted 2-norm, i.e.,
1
|u − v|B T P B = ((u − v)T B T P B(u − v)) 2 .
This is illustrated in Fig. 11.3.1. In Fig. 11.3.1, v is marked with “◦” and
the optimal u is marked with “∗”. Suppose that v ∈ R2 is outside of the
3
R4
2 R
3
R
2
v v
v
1
u u u
v
u
0 −R R0
1
R1
−1
−2 −R2 −R3 −R4
−3
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Figure 11.3.1: Illustration for the relation between v and the optimal u.
unit square, then there exists a smallest ellipsoid centered at v,

E 0 = u ∈ R2 : (u − v)T B T P B(u − v) ≤ r2 ,
which includes a point of the unit square. Since the unit square is convex
and E 0 is strictly convex, they have a unique intersection, which is the
optimal u.
For the case m = 2, the solution can be put into an explicit formula.
For simplicity, let
T a −c
B PB =
−c b
and assume that c > 0. We partition the plane into 9 regions (see
Fig. 11.3.1, where the regions are divided by the dash-dotted lines):

R0 = v ∈ R2 : |v|∞ ≤ 1 ,
c c
R1 = v ∈ R2 : v1 > 1, −1 + (v1 − 1) < v2 ≤ 1 + (v1 − 1) ,
b b
2 c a
R2 = v ∈ R : 1 + (v1 − 1) < v2 ≤ 1 + (v1 − 1) ,
b c
2 c c
R3 = v ∈ R : v2 > 1, −1 + (v2 − 1) ≤ v1 < 1 + (v2 − 1) ,
a a
2 c c
R4 = v ∈ R : v1 < −1 + (v2 − 1), v2 ≥ 1 + (v1 + 1) ,
a b
−R1 , −R2 , −R3 , and −R4 . With this partition, the maximal convergence
control that solves (11.3.2) can be written as,

 v, if v ∈ R0 ,





 1

 , if v ∈ R1 ,

 v − c(v1 − 1)/b

 2




 1
, if v ∈ R2 ,
u= 1 (11.3.13)





 v1 − c(v2 − 1)/a

 , if v ∈ R3 ,

 1





 −1

 , if v ∈ R4 .
1
and for v ∈ −R1 , −R2 , −R3 or −R4 , the optimal control u is symmetric.
Since ui does not solely depend on vi , we can view u as a coupled saturation
function of v. If m > 2, it is hard to put the optimal solution into an explicit
formula. However, the optimization problem (11.3.2) can be transformed
into a very simple LMI problem:
min γ (11.3.14)
u,γ

γ (u − v)T
s.t. ≥ 0, |u|∞ ≤ 1.
u−v (B T P B)−1
This optimization problem can be efficiently solved so that the control can
be computed and implemented on line. It can be shown that the optimal
control u resulting from (11.3.14) is continuous in v and hence continuous
in x.
We have some remarks for the case that B has full row rank. It is known
that there exists a linear dead-beat control for this case. Let B + be the
right inverse of B, then
AT P B(B T P B)−1 B T P A = AT P B(B T P B)−1 B T P BB + A = AT P A.

It follows that
T
V (Ax + Bu) = u + (B T P B)−1 B T P Ax B T P B

× u + (B T P B)−1 B T P Ax .
It is easy to see that the optimal convergence control is also a dead-beat

control for x in the region,

x ∈ Rn : |(B T P B)−1 B T P Ax|∞ ≤ 1 .
Outside of this region, the optimal control can be obtained by solving a

simple LMI problem.
Similar to the continuous-time case, it can be shown that the overall
convergence rate increases as ρ decreases. And if E(P, ρ) is in the linear
region, the overall convergence rate is a constant which equals to that of a
linear system.
The problem of disturbance rejection for a discrete-time system is much
more complicated than its continuous-time counterpart. This is because
the expression of ∆V (x, u, w) contains some cross terms between u and w
and the maximal convergence control will be dependent on w.
11.4. Conclusions
We have shown in this chapter that, for a continuous-time system subject
to input saturation and persistent disturbance, the maximal convergence
control is a bang-bang type control with a simple switching strategy. A
saturated high gain linear feedback is developed to avoid the discontinuity
of the bang-bang control. For a discrete-time system, the maximal con-
vergence control is a coupled saturated linear feedback. We also provided
methods for determining the largest ellipsoid that can be made invariant
with a bounded control for both continuous-time and discrete-time systems.
Chapter 12
Output Regulation –
Continuous-Time
Systems
12.1. Introduction
Earlier in the book, we presented ways of constructing stabilizing feedback

laws that result in large domains of attraction. In this chapter, we will
study the classical problem of output regulation for linear systems subject
to actuator saturation.
The problem of output regulation is one of regulating the outputs to
the reference signals asymptotically in the presence of disturbances. A key
feature of this problem formulation is that both the reference signals and
the disturbances are generated by an external system, commonly known as
the exogenous system or exosystem. This problem is well studied for both
linear systems (see, e.g., [22,109]) and nonlinear systems (see, e.g., [48]).
In comparison with the problem of stabilization, the problem of output
regulation for linear systems subject to actuator saturation, however, has
received relatively less attention. The few works that have motivated our
current research are [88], [71] and [101]. In [71,101], the problem of output
regulation was studied for ANCBC systems subject to actuator saturation.
Necessary and sufficient conditions on the plant/exosystem and their initial
conditions were derived under which output regulation can be achieved.
Under these conditions, feedback laws that achieve output regulation were
265
266 Chapter 12. Output Regulation – Continuous-Time Systems
constructed based on the semi-global stabilizing feedback laws of [65,67].

The recent work [88] made an attempt to address the problem of output
regulation for exponentially unstable linear systems subject to actuator
saturation. The attempt was to enlarge the set of initial conditions of the
plant and the exosystem under which output regulation can be achieved.
In particular, for plants with only one positive pole and exosystems that
contain only one frequency component, feedback laws were constructed that
achieve output regulation on what we will characterize in this chapter as
the asymptotically regulatable region.
The objective of this chapter is to systematically study the problem of
output regulation for general linear systems subject to actuator saturation.
The presentation of this chapter is based on our recent work [37]. In partic-
ular, we will first characterize the regulatable region, the set of plant and
exosystem initial conditions for which output regulation is possible with
the saturating actuators. It turns out that the asymptotically regulatable
region can be characterized in terms of the null controllable region of the
anti-stable subsystem of the plant. We then construct feedback laws that
achieve regulation on the regulatable region. These feedback laws are con-
structed from the stabilizing feedback laws in such a way that a stabilizing
feedback law that achieves a larger domain of attraction leads to a feedback
law that achieves output regulation on a larger subset of the regulatable
region and, a stabilizing feedback law on the entire null controllable region
leads to a feedback law that achieves output regulation on the entire regu-
latable region. In the situation that output regulation is required to occur
on the entire regulatable region, the problem is referred to as global output
regulation. If, on the other hand, output regulation is only required to
occur on any (arbitrarily large) compact subset of the regulatable region,
the problem is then referred to as semi-global output regulation.
This chapter is organized as follows. In Section 12.2, we precisely for-
mulate the problem of output regulation for linear systems with saturating
actuators. As a motivation to our problem formulation and our solution
to the problem, we will also briefly recall the classical problem of output
regulation for linear systems as originally formulated and solved in [22].
Section 12.3 characterizes the regulatable region. Sections 12.4 and 12.5
respectively construct state feedback and error feedback laws that achieve
output regulation on the regulatable region. In Section 12.6, we illustrate
12.2. Preliminaries and Problem Statement 267
our results on an aircraft model. Finally, Section 12.7 draws a brief con-
clusion to this chapter.
12.2. Preliminaries and Problem Statement

In this section, we first recall from [22,48] the classical formulation of and
results on the problem of output regulation for linear systems in Section
12.2.1 and then formulate the problem of output regulation for linear sys-
tems subject to actuator saturation in Section 12.2.2. This brief review will
motivate our problem formulation as well as the solution to the formulated
problems.
12.2.1. Review of Linear Output Regulation Theory
Consider a linear system

ẋ = Ax + Bu + P w,
ẇ = Sw, (12.2.1)
e = Cx + Qw.
The first equation of this system describes a plant, with state x ∈ Rn
and input u ∈ Rm , subject to the effect of a disturbance represented by
P w. The third equation defines the error e ∈ Rq between the actual plant
output Cx and a reference signal −Qw that the plant output is required to
track. The second equation describes an autonomous system, often called
the exogenous system or the exosystem, with state w ∈ Rr . The exosystem
models the class of disturbances and references taken into consideration.
The control action to the plant, u, can be provided either by state
feedback or by error feedback. A state feedback controller has the form of
u = F x + Gw, (12.2.2)
while an error feedback controller has the form of

ξ˙ = Ac ξ + Bc e, ξ ∈ Rl ,
(12.2.3)
u = Cc ξ + Dc e.
The objective is to achieve internal stability and output regulation. Internal
stability means that if we disconnect the exosystem and set w equal to zero
then the closed-loop system is asymptotically stable. Output regulation
means that for any initial conditions of the closed-loop system, we have
that e(t) → 0 as t → ∞. Formally, all of these can be summarized in the
following two synthesis problems.
Definition 12.2.1. The problem of output regulation by state feedback

for linear system (12.2.1) is to find, if possible, a state feedback law of the
form (12.2.2) such that
1) the system
ẋ = (A + BF )x
is asymptotically stable;
2) for all (x0 , w0 ) ∈ Rn+r , the interconnection of (12.2.1) and the feed-
back law (12.2.2) results in
lim e(t) = 0.
t→∞
Definition 12.2.2. The problem of output regulation by error feedback

for linear system (12.2.1) is to find, if possible, an error feedback law of the
form (12.2.3) such that
1) the system
ẋ = Ax + BCc ξ + BDc Cx,
ξ˙ = Ac ξ + Bc Cx
is asymptotically stable;
2) for all (x0 , ξ0 , w0 ) ∈ Rn+l+r , the interconnection of (12.2.1) and

(12.2.3) results in
lim e(t) = 0.
t→∞
The solution to these two problems is based on the following assump-

tions.
Assumption 12.2.1.
A1. The eigenvalues of S have nonnegative real parts.
A2. The pair (A, B) is stabilizable.
A3. The pair % &

A P
C Q ,
0 S
is detectable.
The first of these assumptions does not involve a loss of generality,

because asymptotically stable modes in the exosystem do not affect the
regulation of the output. The second is indeed necessary for asymptotic
stabilization via either state or error feedback. The third assumption is
stronger than the detectability assumption on the pair (C, A), which is
necessary for asymptotic stabilization via error feedback. However, it does
not involve a loss of generality, as discussed in detail in [22]. In fact, if
the pair (C, A) is detectable and Assumption A3 does not hold, it is al-
ways possible to reduce the state space of the exosystem to the part which
actually affects the error, and the reduced system thus obtained satisfies
Assumption A3.
The following results, due to Francis [22], describe necessary and suffi-
cient conditions for the existence of solutions to the above two problems.
Proposition 12.2.1. Suppose Assumptions A1 and A2 hold. Then, the

problem of output regulation by state feedback is solvable if and only if
there exist matrices Π and Γ that solve the linear matrix equations
ΠS = AΠ + BΓ + P,
(12.2.4)
CΠ + Q = 0.
Moreover, if in addition Assumption A3 also holds, the solvability of the
above linear matrix equations is also a necessary and sufficient condition
for the solvability of the problem of output regulation by error feedback.
In summary, if Assumptions A1-A3 hold, then the problem of output

regulation by error feedback is solvable if and only if the problem of output
regulation by state feedback is solvable, and the conditions for the existence
of solutions can be expressed in terms of the solvability of certain linear
matrix equations. In [31], it was shown that the possibility of solving these
matrix equations can be characterized in terms of a comparison between the
transmission polynomials of the system (12.2.1), in which u is considered
as the input and e as the output, and those of the system
ẋ = Ax + Bu,
ẇ = Sw, (12.2.5)
e = Cx.
The latter can be interpreted as the system obtained from (12.2.1) by cut-
ting the connections between the exosystem and the plant. More specifi-
cally, the following results were proven in [31].
Proposition 12.2.2. The linear matrix equations (12.2.4) are solvable if

and only if the systems (12.2.1) and (12.2.5) have the same transmission
polynomials.
12.2.2. Output Regulation in the Presence of

Actuator Saturation
Motivated by the classical formulation of output regulation for linear sys-

tems, we consider the following linear system subject to actuator saturation
and the exosystem,
ẋ = Ax + Bu + P w, x ∈ Rn , u ∈ Rm ,
ẇ = Sw, w ∈ Rr , (12.2.6)
e = Cx + Qw, e ∈ Rq .
Here the input u represents the output of the saturating actuators. Without
loss of generality, we assume that the input signal u is measurable and
satisfy the bound u∞ ≤ 1. The set of u’s that satisfy these assumptions
is referred to as admissible controls in Chapter 2 and is denoted by Ua (see
(2.2.2)). Because of the bound on the control input, both the plant and
the exosystem cannot operate in the entire state space. For this reason, we
assume that
(x0 , w0 ) ∈ Y0 ,
for some Y0 ∈ Rn × Rr . Let

X 0 = x0 ∈ Rn : (x0 , 0) ∈ Y0 .
The problems to be addressed in this chapter are the following:
Problem 12.2.1. The problem of output regulation by state feedback for

the system (12.2.6) is to find, if possible, a state feedback law u = φ(x, w),
with |φ(x, w)|∞ ≤ 1 and φ(0, 0) = 0, such that
1. the equilibrium x = 0 of the system
ẋ = Ax + Bφ(x, 0)
is asymptotically stable with X 0 contained in its domain of attraction;
2. for all (x0 , w0 ) ∈ Y0 , the interconnection of (12.2.6) and the feedback

law u = φ(x, w) results in bounded state trajectory x(t) and
lim e(t) = 0.
t→∞
12.3. The Regulatable Region 271
Problem 12.2.2. The problem of output regulation by error feedback for

the system (12.2.6) is to find, if possible, an error feedback law of the form
ξ̇ = ψ(ξ, e), ξ0 ∈ Y0 ,
u = φ(ξ),
with ψ(0, 0) = 0, φ(0) = 0 and |φ(ξ)|∞ ≤ 1, such that
1. the equilibrium (x, ξ) = (0, 0) of the system
ẋ = Ax + Bφ(ξ),
(12.2.7)
ξ˙ = ψ(ξ, Cx)
is asymptotically stable with X 0 × Y0 contained in its domain of

attraction;
2. for all (x0 , w0 , ξ0 ) ∈ Y0 × Y0 , the interconnection of (12.2.6) and

(12.2.7) results in bounded trajectory (x(t), ξ(t)) and
lim e(t) = 0.
t→∞
Our objectives are to characterize the maximal set of initial conditions

(x0 , w0 ), the largest possible Y0 , on which the above two problems are
solvable and to explicitly construct feedback laws that result in Y0 as large
as possible.
We will assume that (A, B) is stabilizable. We will also assume that S
is neutrally stable and all its eigenvalues are on the imaginary axis. The
stabilizability of (A, B) is clearly necessary for the stabilization of the plant.
The assumption on S is without loss of generality. This is because asymp-
totically stable modes in the exosystem do not affect the regulation of the
output and output regulation cannot be expected with bounded controls if
the exosystem has unstable modes.
12.3. The Regulatable Region

In this section, we will characterize the set of all initial states of the plant
and the exosystem on which the problems of output regulation are solvable
under the restriction that u∞ ≤ 1. We will refer to this set as the
asymptotically regulatable region.
To begin with, we observe from the classical output regulation theory
(see Section 12.2.1) that for these two problems to be solvable, there must
exist matrices Π ∈ Rn×r and Γ ∈ Rm×r that solve the following linear
matrix equations,
(12.3.1)
0 = CΠ + Q.
Given the matrices Π and Γ, we define a new state z = x − Πw and
rewrite the system equations as
ż = Az + Bu − BΓw,
ẇ = Sw, (12.3.2)
e = Cz.
From these new equations, it is clear that e(t) goes to zero asymptotically
if z(t) goes to zero asymptotically. This is possible only if (see [71])

sup ΓeSt w0 ∞ < 1. (12.3.3)
t≥0
For this reason, we will restrict our attention to exosystem initial conditions
in the following compact set

W 0 = w0 ∈ Rr : |Γw(t)|∞ = ΓeSt w0 ∞ ≤ ρ, ∀ t ≥ 0 , (12.3.4)
for some ρ ∈ [0, 1). For later use, we also denote δ = 1 − ρ. We note that
the compactness of W 0 can be guaranteed by the observability of (Γ, S).
Indeed, if (Γ, S) is not observable, the exosystem can be reduced to make
it so.
We can now precisely define the notion of asymptotically regulatable
region as follows.
Definition 12.3.1.
1) Given T > 0, a pair (z0 , w0 ) ∈ Rn × W 0 is regulatable in time T if
there exists an admissible control u, such that the response of (12.3.2)
satisfies z(T ) = 0. The set of all (z0 , w0 ) regulatable in time T is
denoted as Rg (T ).
2) A pair (z0 , w0 ) is regulatable if (z0 , w0 ) ∈ Rg (T ) for some T ∈ [0, ∞).

The set of all regulatable (z0 , w0 ) is referred to as the regulatable
region and is denoted as Rg .
3) The set of all (z0 , w0 ) for which there exist admissible controls such
that the response of (12.3.2) satisfies
lim z(t) = 0
t→∞
is referred to as the asymptotically regulatable region and is denoted

as Rag .
Remark 12.3.1. Note that we have defined the asymptotically regulatable

region in terms of
lim z(t) = 0
t→∞
rather than
lim e(t) = 0.
t→∞
Requiring the former instead of the latter will also guarantee the closed-loop
stability in the absence of w. As will be explained in detail in Remark 12.3.2,
this will result in essentially the same description of the asymptotically
regulatable region.
We will describe Rg (T ), Rg and Rag in terms of the asymptotically null

controllable region of the plant
v̇ = Av + Bu, u ∈ Ua . (12.3.5)
The asymptotically null controllable region C a and the related C and C(T ),
were defined in Chapter 2. Since our description of Rg (T ), Rg and Rag
relies heavily on them, we briefly recall these definitions as follows.
Definition 12.3.2. The null controllable region of the system (12.3.5) at

time T , denoted as C(T ), is the set of v0 ∈ Rn that can be driven to the
origin in time T and the null controllable region, denoted as C, is the set
of v0 ∈ Rn that can be driven to the origin in finite time by admissible
controls. The asymptotically null controllable region, denoted as C a , is the
set of all v0 that can be driven to the origin asymptotically by admissible
controls.
Clearly,
C= C(T ), (12.3.6)
T ∈[0,∞)
and
T
−Aτ
C(T ) = e Bu(τ )dτ : u ∈ Ua . (12.3.7)
0
It is also clear that the null controllable region and the asymptotically null
controllable region are identical if the pair (A, B) is controllable. Some
simple methods to describe C and C a were developed in Chapter 2. To

simplify the characterization of Rg and Rag , and without loss of generality,
let us assume that

z1
z= , z1 ∈ Rn1 , z2 ∈ Rn2 ,
z2
and
A1 0 B1
A= , B= , (12.3.8)
0 A2 B2
where A1 ∈ Rn1 ×n1 is semi-stable (i.e., all its eigenvalues are in the closed
left-half plane) and A2 ∈ Rn2 ×n2 is anti-stable (i.e., all its eigenvalues are
in the open right-half plane). The anti-stable subsystem,
ż2 = A2 z2 + B2 u − B2 Γw,
(12.3.9)
ẇ = Sw,
is of crucial importance. Denote its regulatable regions as Rg 2 (T ) and Rg 2 ,

and the null controllable regions of the system
v̇2 = A2 v2 + B2 u
as C 2 (T ) and C 2 . Then, the asymptotically null controllable region of the

system
v̇ = Av + Bu
is given by C a = Rn1 × C 2 (see Chapter 2), where C 2 is a bounded convex

open set. Denote the closure of C 2 as C 2 ; then
∞
C2 = e−A2 τ B2 u(τ )dτ : u ∈ Ua .
0
Also, if a set D is in the interior of C 2 , then, there is a finite T > 0 such

that D is in the interior of C 2 (T ).
Theorem 12.3.1. Let V2 ∈ Rn2 ×r be the unique solution to the linear

matrix equation
−A2 V2 + V2 S = −B2 Γ (12.3.10)
and let
V (T ) = V2 − e−A2 T V2 eST .
Then,

a) Rg 2 (T ) = (z2 , w) ∈ Rn2 × W 0 : z2 − V (T )w ∈ C 2 (T ) ; (12.3.11)

b) Rg 2 = (z2 , w) ∈ Rn2 × W 0 : z2 − V2 w ∈ C 2 ; (12.3.12)
c) Rag = Rn1 × Rg 2 . (12.3.13)
Proof.
a) Given (z20 , w0 ) ∈ Rn2 × W 0 and an admissible control u, the solution
of (12.3.9) at t = T is,
T
z2 (T ) = eA2 T z20 + e−A2 τ B2 u(τ )dτ
0

T
−A2 τ
− e B2 Γe Sτ
w0 dτ . (12.3.14)
0
Applying (12.3.10), we have

T T
− e−A2 τ B2 ΓeSτ dτ = de−A2 τ V2 eSτ
0 0
−A2 T
=e V2 eST − V2
= −V (T ). (12.3.15)
Thus,
T
−A2 T
e z2 (T ) = z20 − V (T )w0 + e−A2 τ B2 u(τ )dτ.
0
By setting z2 (T ) = 0, we immediately obtain (12.3.11) from the definition

of Rg (T ) and (12.3.7).
b) Since A2 is anti-stable and S is neutrally stable, we have that
lim V (T ) = V2 .
T →∞

∞
V2 = e−A2 τ B2 ΓeSτ dτ.
0
First we show that
z20 − V2 w0 ∈ C 2 =⇒ (z20 , w0 ) ∈ Rg 2 .
Since C 2 is open, if z20 − V2 w0 ∈ C 2 , then, there exists an ε > 0 such that

z20 − V2 w0 + z2 : |z2 |∞ ≤ ε ⊂ C 2 .
Also, there exists a T1 > 0 such that

z20 − V2 w0 + z2 : |z2 |∞ ≤ ε ⊂ C 2 (T1 ).
Since
lim V (T ) = V2 ,
T →∞
there is a T2 > T1 such that
|(V2 − V (T2 ))w0 |∞ < ε.
Hence,
z20 − V (T2 )w0 = z20 − V2 w0 + (V2 − V (T2 )w0 ∈ C 2 (T1 ) ⊂ C 2 (T2 ).
It follows from a) that (z20 , w0 ) ∈ Rg 2 (T2 ) ⊂ Rg 2 .

Next we show that
(z20 , w0 ) ∈ Rg 2 =⇒ z20 − V2 w0 ∈ C 2 .
If (z20 , w0 ) ∈ Rg 2 , then (z20 , w0 ) ∈ Rg 2 (T1 ) for some T1 > 0. It follows

from the definition of Rg (T ) that there exists an admissible control u1 such
that
T1 T1
z20 − e−A2 τ B2 ΓeSτ w0 dτ + e−A2 τ B2 u1 (τ )dτ = 0.
0 0
Denote
Z2 = δe−A2 T1 v2 : v2 ∈ C 2 .
For each z2 ∈ Z2 , there is an admissible control u2 such that
∞ ∞
z2 = δe−A2 T1 e−A2 τ B2 u2 (τ )dτ = e−A2 τ B2 δu2 (τ − T1 )dτ.
0 T1
Hence,
∞
z20 − V2 w0 + z2 = z20 − e−A2 τ B2 ΓeSτ w0 dτ
0
∞
+ e−A2 τ B2 δu2 (τ − T1 )dτ
T1
T1
= z20 − e−A2 τ B2 ΓeSτ w0 dτ
∞0

+ e−A2 τ B2 δu2 (τ − T1 ) − ΓeSτ w0 dτ
T1
T1
=− e−A2 τ B2 u1 (τ )dτ
0
∞
+ e−A2 τ B2 δu2 (τ − T1 ) − ΓeSτ w0 dτ
T1
∈ C2,
where the last step follows from the fact that

Sτ
Γe w0 ) ≤ ρ = 1 − δ, ∀ τ > 0.
∞
This implies that

z20 − V2 w0 + z2 : z2 ∈ Z2 ⊂ C 2 .
Since e−A2 T1 is nonsingular, the set Z2 contains the origin in its interior.
It follows that
z20 − V2 w0 ∈ C 2 .
c) It is easy to see that Rag ⊂ Rn1 × Rg 2 . We need to show that

Rn1 × Rg 2 ⊂ Rag .
Suppose that (z0 , w0 ) = (z10 , z20 , w0 ) ∈ Rn1 × Rg 2 . There exist a T ≥ 0
and an admissible u such that z2 (T ) = 0. For t ≥ T , let u = Γw + uδ , with
uδ ∞ ≤ δ = 1 − ρ. Since w0 ∈ W 0 , we have Γw∞ ≤ ρ and hence,
u∞ ≤ 1 and
ż = Az + Bu − BΓw = Az + B uδ .
Since
z1 (T )
0
is inside the asymptotically null controllable region of the above system
under the constraint uδ ∞ ≤ δ (see Chapter 2), there exists a uδ such
that
lim z(t) = 0.
t→∞
Hence (z0 , w0 ) ∈ Rag . This establishes that Rn1 × Rg 2 ⊂ Rag and hence
Rn1 × Rg 2 = Rag . 2
Remark 12.3.2. We can now continue with the argument in Remark

12.3.1. From the above theorem, we see that (z0 , w0 ) is regulatable if and
only if z20 − V2 w0 ∈ C 2 . This is essentially the necessary condition to keep
z2 (t) bounded. We explain this as follows: Suppose that z20 − V2 w0 ∈ / C 2,
then, there exists an ε > 0 such that
|z20 − V2 w0 − z2 | > ε, ∀ z2 ∈ C 2 . (12.3.16)
Since
lim V (T ) = V2 ,
T →∞
there exists a T0 > 0 such that

ε
|(V2 − V (T ))w0 | <
2
for all T > T0 . Since
T
− e−A2 τ B2 u(τ )dτ ∈ C 2 , ∀ T > 0,
0
it follows from (12.3.16) and the inequality
|a + b| > |a| − |b|, ∀ a, b ∈ Rn2 ,
that
T T

z20 − e−A2 τ B2 ΓeSτ w0 dτ + e−A2 τ B2 u(τ )dτ
0 0
T

= z20 − V (T )w0 + e−A2 τ B2 u(τ )dτ
0
ε ε
>ε− = ,
2 2
for all T > T0 and u ∈ Ua . Since the smallest singular value of eA2 T
increases exponentially with T , it follows from (12.3.14) that z2 (T ) will grow
unbounded. Recall from the definition of Problems 12.2.1 and 12.2.2 that
the state x(t) and hence z(t) has to be bounded, so even if the requirement
lim z(t) = 0 (12.3.17)

t→∞
is replaced with
lim e(t) = 0, (12.3.18)
t→∞
12.4. State Feedback Controllers 279
we still require z20 − V2 w0 ∈ C 2 to achieve output regulation. The gap

only arises from the boundary of C 2 . It is unclear whether it is possible to
achieve (12.3.18) with z(t) bounded for z20 − V2 w0 ∈ ∂C 2 . Since this prob-
lem involves too much technical detail and is of little practical importance
(one will not take the risk to allow z20 − V2 w0 ∈ ∂C 2 , otherwise a small
perturbation will cause the state to grow unbounded), we will not dwell on
this subtle technical point here.
Remark 12.3.3. Given (z0 , w0 ), there exists an admissible control u such

that
lim z(t) = 0
t→∞
if and only if (z0 , w0 ) ∈

Rag . Recalling that z = x − Πw, we observe that,
for a given pair of initial states in the original coordinates, (x0 , w0 ), there
is an admissible control u such that
lim (x(t) − Πw(t)) = 0

t→∞
if and only if
x20 − (Π2 + V2 )w0 ∈ C 2 ,
where

Π2 = 0 In2 Π.
12.4. State Feedback Controllers
In this section, we will construct a feedback law that solves the problem of
output regulation by state feedback for linear systems subject to actuator
saturation. We will assume that a stabilizing state feedback law
u = f (v), |f (v)|∞ ≤ 1, v ∈ Rn ,
has been designed and the equilibrium v = 0 of the closed-loop system
v̇ = Av + Bf (v) (12.4.1)
has a domain of attraction S ⊂ C a . We will construct our feedback law

u = g(z, w) from this stabilizing feedback law.
Given a state feedback
u = g(z, w),
with
|g(z, w)|∞ ≤ 1, ∀ (z, w) ∈ Rn × W 0 ,
the closed-loop system is given by,
ż = Az + B g(z, w) − BΓw, (12.4.2)

ẇ = Sw.
Denote the time response of z(t) to the initial state (z0 , w0 ) as z(t, z0 , w0 )
and define

S zw := (z0 , w0 ) ∈ Rn × W 0 : lim z(t, z0 , w0 ) = 0 .
t→∞
Since Rag is the set of all (z0 , w0 ) for which z(t) can be driven to the origin
asymptotically, we must have S zw ⊂ Rag . Our objective is to design a
control law u = g(z, w) such that S zw is as large as possible, or as close to
Rag as possible.
First we need a mild assumption which will be removed later. Assume
that there exists a matrix V ∈ Rn×r such that
−AV + V S = −BΓ. (12.4.3)
This will be the case if A and S have no common eigenvalues. With the
decomposition in (12.3.8), if we partition V accordingly as

V1
V = ,
V2
then V2 satisfies
−A2 V2 + V2 S = −B2 Γ.
Denote
Dzw := (z, w) ∈ Rn × W 0 : z − V w ∈ S , (12.4.4)
on which the following observation can be made.
Observation 12.4.1.
a) The set Dzw increases as S increases, and Dzw = Rag if S = C a ;
b) In the absence of w,
x0 ∈ S =⇒ (z0 , 0) ∈ Dzw .
Proof. The fact that Dzw increases as S increases is easy to see. To see
the rest of a), we note that, for a general plant, C a = Rn1 × C 2 . If S = C a ,
then S = Rn1 × C 2 , and

Dzw = (z, w) ∈ Rn × W 0 : z − V w ∈ Rn1 × C 2

= (z1 , z2 , w) ∈ Rn1 × Rn2 × W 0 : z1 − V1 w ∈ Rn1 , z2 − V2 w ∈ C 2
= Rn1 × Rg 2 = Rag .
Part b) is also clear if we note that z0 = x0 − Πw0 = x0 for w0 = 0. 2

With this observation, we see that our objective of making S zw as large
as possible will be achieved by designing a feedback law such that Dzw ⊂
S zw . We will reach this objective through a series of technical lemmas.
Lemma 12.4.1. Let u = f (z − V w). Consider the closed-loop system
ż = Az + Bf (z − V w) − BΓw,
(12.4.5)
ẇ = Sw.
For this system, Dzw is an invariant set and for all (z0 , w0 ) ∈ Dzw ,
lim (z(t) − V w(t)) = 0.

t→∞
Proof. Substituting (12.4.3) into system (12.4.5), we obtain
ż = Az + Bf (z − V w) − AV w + V S w
= A(z − V w) + Bf (z − V w) + V ẇ.
Define the new state v := z − V w; we have
v̇ = Av + Bf (v),
which has a domain of attraction S. This also implies that S is an invariant

set for the v-system.
If (z0 , w0 ) ∈ Dzw , then v0 = z0 − V w0 ∈ S. It follows that
v(t) = z(t) − V w(t) ∈ S, ∀ t ≥ 0,
and
lim (z(t) − V w(t)) = lim v(t) = 0.
t→∞ t→∞
2
Lemma 12.4.1 says that, in the presence of w, the simple feedback u =

f (z − V w) will cause z(t) − V w(t) to approach zero and z(t) to approach
V w(t), which is bounded. Our next step is to construct a finite sequence
of controllers
u = fk (z, w, α), k = 0, 1, 2, · · · , N,
all parameterized in α ∈ (0, 1). By judiciously switching among these
controllers, we can cause z(t) to approach αk V w(t) for any k. By choosing
N large enough, z(t) will become arbitrarily small in a finite time. Once
z(t) becomes small enough, we will use the controller
% &
Fz
u = Γw + δ sat ,
δ
with F to be specified later, to cause z(t) to converge to the origin.

Let F ∈ Rm×n be such that
v̇ = Av + B sat(F v), (12.4.6)
is asymptotically stable at the origin. Let X > 0 be such that
(A + BF )T X + X(A + BF ) < 0
and that the ellipsoid

E(X) := v ∈ Rn : v T Xv ≤ 1
is in the linear region of the feedback control u = sat(F v), i.e.,
|F v|∞ ≤ 1, ∀ v ∈ E(X).
Then E(X) is an invariant set and is in the domain of attraction of the

origin for the closed-loop system (12.4.6).
Lemma 12.4.2. Suppose that D ⊂ Rn is an invariant set in the domain

of attraction for the system
v̇ = Av + Bf (v). (12.4.7)
Then, for any α > 0, αD is an invariant set in the domain of attraction for
the system v
v̇ = Av + αBf . (12.4.8)
α
Proof. Write (12.4.8) as
v̇ v v
= A + Bf ,
α α α
v v0
and replace with v, then we get (12.4.7). If v0 ∈ αD, i.e., ∈ D, then
α α
v(t)
∈ D, ∀t > 0
α
and
lim v(t) = 0.
t→∞
For any α ∈ (0, 1), since W 0 is a compact set, there exists a positive
integer N such that
1

αN X 2 V w < δ, ∀ w ∈ W 0 . (12.4.9)
That is,
αN V w ∈ δE(X), ∀ w ∈ W 0.
Define a sequence of subsets in Rn × W 0 as,

k
Dzw = (z, w) ∈ Rn × W 0 : z − αk V w ∈ αk E(X) , k = 0, 1, · · · , N,

N +1
Dzw = (z, w) ∈ Rn × W 0 : z ∈ δE(X) ,
and, on each of these sets, define a state feedback law as follows,

% &
F (z − αk V w)
fk (z, w, α) = (1 − αk )Γw + αk sat , k = 0, 1, · · · , N,
αk
% &
Fz
fN +1 (z, w) = Γw + δ sat .
δ
It can be verified that, for any k = 0 to N + 1,
|fk (z, w, α)|∞ ≤ 1, ∀ (z, w) ∈ Rn × W 0 .
Lemma 12.4.3. Let u = fk (z, w, α). Consider the closed-loop system
ż = Az + Bfk (z, w, α) − BΓw,

(12.4.10)
ẇ = Sw.
k
For this system, Dzw is an invariant set. Moreover, if k = 0, 1, · · · , N , then,

lim z(t) − αk V w(t) = 0, ∀ (z0 , w0 ) ∈ Dzw
k
,
t→∞
and if k = N + 1, then,
N +1
lim z(t) = 0, ∀ (z0 , w0 ) ∈ Dzw .
t→∞
Proof. With
u = fk (z, w, α), k = 0, 1, · · · , N,
we have
% &
F (z − αk V w)
ż = Az + 1 − αk BΓw + αk Bsat − BΓw
αk
% &
F (z − αk V w)
= Az + αk Bsat − αk BΓw. (12.4.11)
αk
Let vk = z − αk V w, then by (12.4.3),

% &
F vk
v̇k = Avk + αk Bsat . (12.4.12)
αk
It follows from Lemma 12.4.2 that αk E(X) is an invariant set in the domain
k
of attraction for the vk -system. Hence Dzw is invariant for the system
(12.4.10) and if (z0 , w0 ) ∈ Dzw
k
, i.e.,
vk0 = z0 − αk V w0 ∈ αk E(X),
then,

lim z(t) − αk V w(t) = lim vk (t) = 0.
t→∞ t→∞
For % &
Fz
u = fN +1 (z, w) = Γw + δsat ,
δ
we have % &
Fz
ż = Az + δBsat
δ
and the same argument applies. 2
Based on the technical lemmas established above, we construct our final
state feedback law as follows,
u = g(z, w, α, N )


 fN +1 (z, w), if (z, w) ∈ ΩN +1 := Dzw
N +1
,


 f (z, w, α), if (z, w) ∈ Ωk := Dzw
k
\ ∪N +1 j
k j=k+1 Dzw ,
=

 k = 0, 1, · · · , N,


 +1 j
f (z − V w), if (z, w) ∈ Ω := Rn × W 0 \ ∪N
j=0 Dzw .
Since Ω, Ω0 , · · · , ΩN +1 are disjoint and their union is Rn ×W 0 , the controller

is well defined on Rn ×W 0 . What remains to be shown is that this controller
will accomplish our objective if the parameter α is properly chosen.
Let
12
X V w
α0 = max 1
.
w∈W 0
X 2 V w + 1
It is obvious that α0 ∈ (0, 1).
Theorem 12.4.1. Choose any α ∈ (α0 , 1) and let N be specified as in

(12.4.9). Then for all (z0 , w0 ) ∈ Dzw , the solution of the closed-loop system,
ż = Az + B g(z, w, α, N ) − BΓw, (12.4.13)

ẇ = Sw,
satisfies
lim z(t) = 0,
t→∞
i.e., Dzw ⊂ S zw .
Proof. The control u = g(z, w, α, N ) is executed by choosing one feedback

law from fk (z, w, α), k = 0, 1, · · · , N + 1, and f (z − V w). The crucial point
is to guarantee that no chattering will occur and that (z, w) will move
successively from Ω, to Ω0 , Ω1 , · · ·, finally entering ΩN +1 , in which z(t) will
converge to the origin.
Without loss of generality, we assume that (z0 , w0 ) ∈ Ω ∩ Dzw , so the
control u = f (z − V w) is in effect at the beginning. By Lemma 12.4.1,
lim (z(t) − V w(t)) = 0.

t→∞
Hence there is a finite time t0 ≥ 0 such that z(t0 ) − V w(t0 ) ∈ E(X), i.e.,
0
(z(t0 ), w(t0 )) ∈ Dzw . The condition that (z(t), w(t)) ∈ Dzw
k
, k > 0, might
be satisfied at an earlier time t1 ≤ t0 . In any case, there is a finite time
t1 ≥ 0 such that
+1
(z(t1 ), w(t1 )) ∈ Ωk = Dzw
k
\ ∪N j
j=k+1 Dzw ,
for some k = 0, 1, · · · N + 1. After that, the control u = fk (z, w, α) will be

in effect.
We claim that, for any (z(t1 ), w(t1 )) ∈ Ωk , under the control u =
fk (z, w, α), there is a finite time t2 > t1 such that (z(t2 ), w(t2 )) ∈ Dzw
k+1
.
Since Ω ⊂ Dzw , by Lemma 12.4.3, we have that, under the control
k k
u = fk (z, w, α),

lim z(t) − αk V w(t) = 0.
t→∞
Since α ∈ (α0 , 1), we have
1

(1 − α) X 2 V w < α, ∀ w ∈ W 0.
Therefore, for k < N ,

1 1 1
2
X (z − αk+1 V w) ≤ X 2 (z − αk V w) + αk (1 − α) X 2 V w
1

< X 2 (z − αk V w) + αk+1 . (12.4.14)
Since the first term goes to zero asymptotically, there exists a finite time
t2 > t1 such that
1
2
X (z(t2 ) − αk+1 V w(t2 )) ≤ αk+1 .
This implies that
z(t2 ) − αk+1 V w(t2 ) ∈ αk+1 E(X),
i.e.,
(z(t2 ), w(t2 )) ∈ Dzw
k+1
.
If k = N , then, by (12.4.9),
1 1 1
2 2
X z ≤ X (z − αN V w) + αN X 2 V w
1

< X 2 (z − αN V w) + δ.
Also, the first term goes to zero asymptotically, so there exists a finite time
t2 > t1 such that 1
2
X z(t2 ) ≤ δ,
N +1
i.e., (z(t2 ), w(t2 )) ∈ Dzw .
k+l
Just as before, (z, w) might have entered Dzw , l > 1, before it enters
k+1
Dzw . In any case, there is a finite time t such that
+1
(z(t), w(t)) ∈ Ωk+l = Dzw
k+l
\ ∪N j
j=k+l+1 Dzw ,
for some l ≥ 1. After that, the controller will be switched to fk+l (z, w, α).
k
It is also important to note that, by Lemma 12.4.3, Dzw is invariant
under the control u = fk (z, w, α). Once (z, w) ∈ Ω ⊂ Dzw , it will never go
k k
back to Ωl , l < k (or Ω) since Ωl , l < k and Ω have no intersection with Dzw k
l k
(But Ω , l > k, might have intersection with Dzw ). In summary, for any
(z0 , w0 ) ∈ Dzw , suppose (z0 , w0 ) ∈ Ωk , the control will first be fk (z, w, α)
and then switch successively to fk1 , fk2 , · · ·, with k1 , k2 , · · ·, strictly increas-
N +1
ing until (z, w) enters Dzw and remains there. Hence, by Lemma 12.4.3,
we have
lim z(t) = 0.
t→∞
From the proof of Theorem 12.4.1, we see that for all (z0 , w0 ) ∈ Dzw ,
the number of switches is at most N + 2.
For a better understanding, we illustrate the proof of Theorem 12.4.1
k
with Fig. 12.4.1. The sets Dzw for the simplest case where both z and w are
one dimensional are plotted. Here we have X = V = 1, α = 0.6, δ = 0.2
and N = 3. The parallelogram bounded by the straight lines Lk (along
with the two vertical lines w = ±1) is

k
Dzw = (z, w) : z − αk w ≤ αk , |w| ≤ 1 , k = 0, 1, 3, 4.
The dotted line passing through the origin is in parallel to the line L0 .
Suppose that (z0 , w0 ) ∈ Ω0 ⊂ Dzw0
, then under the control u = f0 (z, w, α),
(z, w) will converge to the dotted line. Since α is chosen such that this
1 1
dotted line is inside Dzw (see (12.4.14)), (z, w) will enter Dzw in a finite
time. After that, the controller will be switched to f1 (z, w, α) and so on.
4
Finally (z, w) will enter Dzw and z(t) will go to zero. If z is two dimensional,
k
we will have cylinders as Dzw instead of parallelograms.
In what follows, we will deal with the case that there is no V satisfying
−AV + V S = −BΓ.
This will occur if A and S have some same eigenvalues on the imaginary
axis. Another case is that some eigenvalues of A and S on the imaginary
axis are very close. This will result in large elements of V , so α could be
very close to 1 and N could be very large. The following method is derived
to deal with these two cases.
2.5
z
2 L0
1.5
L
1
1
0.5 L
3
L
4
0
L
4
−0.5 L3
−1
L1
−1.5
−2 L
0
w
−2.5
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Figure 12.4.1: Illustration for the proof of Theorem 12.4.1.
Suppose that the z-system (12.3.2) has the following form,

ż1 A1 0 z1 B1 B1 Γ
= + u− w, (12.4.15)
ż2 0 A2 z2 B2 B2 Γ
with A1 ∈ Rn1 ×n1 semi-stable and A2 ∈ Rn2 ×n2 anti-stable. Also suppose
that there is a known function f (v2 ), |f (v2 )|∞ ≤ 1 for all v2 ∈ Rn2 , such
that the origin of the following system
v̇2 = A2 v2 + B2 f (v2 )
has a domain of attraction S 2 , which is a bounded set. Then, by Lemma

12.4.2, the system v
2
v̇2 = A2 v2 + δB2 f
δ
has a domain of attraction δS 2 . Also, by [102], there exists a control u =
δ sat(h(v)) such that the origin of
v̇ = Av + δBsat(h(v))
has a domain of attraction S δ = Rn1 × δS 2 .

Since there is a V2 satisfying
−A2 V2 + V2 S = −B2 Γ,
by the foregoing development of Theorem 12.4.1, there exists a controller

u = g(z2 , w, α, N ) such that for any (z0 , w0 ) satisfying
z20 − V2 w0 ∈ S 2 , w0 ∈ W 0 ,
the time response of the closed-loop system satisfies
lim z2 (t) = 0.
t→∞
Hence there is a finite time t1 > 0 such that z(t1 ) ∈ Rn1 × δS 2 . After that
if we switch to the control u = Γw + δ sat(h(z), we will have
ż = Az + δBsat(h(z))
and z will stay in S δ = Rn1 ×δS 2 . To avoid the possibility that the state z2
might be trapped on the boundary of δS 2 , we need to modify the controller
further such that it will not switch before z2 has entered γδS 2 for some
γ ∈ (0, 1). For this purpose, we introduce a switching state variable s(t):


 0, if z2 ∈ Rn2 \ δS 2 ,

s(t) = 1, if z2 ∈ γδS 2 ,


 −
s(t ), if z2 ∈ δS 2 \ γδS 2 .
We may simply assume that s(0) = 0. Now the modified controller is,

g(z2 , w, α, N ), if z2 ∈ Rn2 \ γδS 2 and s = 0,
u=
Γw + δ sat(h(z)), if z2 ∈ δS 2 and s = 1.
We conclude that under this control, the following set

Dzw = (z, w) ∈ Rn × W 0 : z2 − V2 w ∈ S 2
will be a subset of S zw . The argument goes as follows. Given (z0 , w0 ) ∈

Dzw , without loss of generality assume that z20 ∈ Rn2 \ δS 2 , then s(0) = 0
and u = g(z2 , w, α, N ) will be in effect and by Theorem 12.4.1 we have
lim z2 (t) = 0.
t→∞
So there is a finite time t1 > 0 such that z2 (t1 ) ∈ δS 2 . At t1 we still

have s(t) = 0 so the controller will not switch until z2 (t2 ) ∈ γδS 2 for
some t2 > t1 . After t2 , we have s(t) = 1 and the controller will switch to
Γw + δ sat(h(z)). Under this control Rn1 × δS 2 is an invariant set and a
domain of attraction. So z2 (t) will stay in δS 2 , s(t) will stay at 1 and we
have
lim z(t) = 0.
t→∞
12.5. Error Feedback Controllers

Consider again the open loop system (12.3.2). Here in this section, we
assume that only the error e = Cz is available for feedback. Also, without
loss of generality, assume that the pair
% &
A −BΓ
(C, A) = [C 0],
0 S
is observable. If it is detectable, but not observable, then the unobservable
modes must be the asymptotically stable eigenvalues of A, which do not
affect the output regulation (see (12.3.2)) and hence can be left out.
We use the following observer to reconstruct the states z and w,
z̄˙ = Az̄ + Bu − BΓw̄ − L1 (e − C z̄), (12.5.1)

w̄˙ = S w̄ − L2 (e − C z̄).
Letting
z̃ = z − z̄, w̃ = w − w̄,
we can write the composite system as
ż = Az + Bu − BΓw, (12.5.2)
ẇ = Sw,

z̃˙ A + L1 C −BΓ z̃
= .
w̃˙ L2 C S w̃
Now we have to use (z̄, w̄) instead of (z, w) to construct a feedback
controller. Since (C, A) is observable, we can choose

L1
L=
L2
appropriately such that the estimation error (z̃, w̃) decays arbitrarily fast.
Moreover, the following fact is easy to establish,
12.5. Error Feedback Controllers 291
Lemma 12.5.1. Denote

A + L1 C −BΓ
Ã = .
L2 C S
Given any (arbitrarily small) positive numbers T and ε, there exists an L

such that

max eÃt , |L| · eÃt ≤ ε, ∀ t ≥ T,
where | · | can be any matrix norm.
Because of this lemma, it is desirable that the controller based on the

observer can achieve almost the same performance as the state feedback
controller. Nevertheless, we need more consideration in the construction
of the switched controller. We also need an additional assumption on the
existing stabilizing controller f (v) so that it can tolerate some class of
disturbances.
Consider the system
v̇ = Av + Bf (v + η),
where η stands for the disturbance arising from, for example, the observer
error. Assume that |f (v)|∞ ≤ 1 for all v ∈ Rn and that there exist a
set D0 ⊂ Rn and positive numbers γ and d0 such that the solution of the
system satisfies
v∞ ≤ γ max (|v0 |∞ , η∞ ) , va ≤ γηa , ∀ v0 ∈ D0 , η∞ ≤ d0 ,
where
va = lim sup |v(t)|∞ .
t→∞
This system is said to satisfy an asymptotic bound from D0 with gain γ

and restriction d0 [102]. For an explicit construction of feedback laws that
achieve such a closed-loop property, see, e.g., Chapter 6. In Chapter 6,
a saturated linear feedback u = f (v) = sat(F0 v) with such property is
constructed for second order anti-stable systems. Moreover, the set D0 can
be made arbitrarily close to the null controllable region.
Let D ∈ Rn be in the interior of D0 . Given a positive number M ,
denote

z̃
DM = (z, w, z̃, w̃) ∈ R × W 0 × R
n n+r
: z − V w ∈ D, ≤M .
w̃
Consider the system (12.5.2). For simplicity, we also assume that

(12.4.3) is satisfied. Letting v = z − V w, we obtain
v̇ = Av + Bu.
Suppose that (z0 , w0 , z̃ 0 , w̃0 ) ∈ DM , then v0 = z0 − V w0 ∈ D. Since D is

in the interior of D0 , there exists a T0 > 0 such that, with every admissible
control u ∈ Ua , we have
v(t) = z(t) − V w(t) ∈ D0 , ∀ t ≤ T0 . (12.5.3)
What we are going to do is to choose L such that the observation error is

sufficiently small after T0 , and to design switching feedback laws to make
z̄(t) → αk V w̄(k) with increasing k and finally drive z̄(t) and z(t) to the
origin.
Lemma 12.5.2. There exists an L ∈ R(n+r)×q such that, under the con-
trol
u1 (t), t < T0 ,
u=
f (z̄ − V w̄), t ≥ T0 ,
where u1 is any admissible control, the solution of the system (12.5.2)
satisfies
lim (z(t) − V w(t)) = 0, ∀ (z0 , w0 , z̃ 0 , w̃ 0 ) ∈ DM .
t→∞
Proof. Let v = z −V w. Since u1 is an admissible control, we have (12.5.3).

In particular,
v(T0 ) = z(T0 ) − V w(T0 ) ∈ D0 . (12.5.4)
Let ṽ = z̃ − V w̃, by Lemma 12.5.1, there exists an L ∈ R(n+r)×q such
that for all (z0 , w0 , z̃ 0 , w̃ 0 ) ∈ DM ,
|ṽ(t)|∞ = |z̃(t) − V w̃(t)|∞ ≤ d0 , ∀ t ≥ T0 . (12.5.5)
We now consider the system after T0 . For t ≥ T0 , we have u = f (z̄−V w̄)

and the closed-loop system is
v̇ = Av + Bf (z̄ − V w̄) = Av + Bf (v − ṽ), v(T0 ) ∈ D0 .
By assumption, this system satisfies an asymptotic bound from D0 with a

finite gain and restriction d0 . It follows from (12.5.4), (12.5.5) and
lim ṽ(t) = 0,
t→∞
that
lim (z(t) − V w(t)) = lim v(t) = 0.
t→∞ t→∞
2
Lemma 12.5.2 means that we can keep z(t) bounded if (z0 , w0 , z̃ 0 , w̃ 0 ) ∈
DM . Just as the state feedback case, we want to move z(t) to the origin
gradually by making z(t) − αk V w(t) small with increased k. Due to the
switching nature of the final controller and that the feedback has to be based
on (z̄, w̄), we need to construct a sequence of sets which are invariant with
respect to (z̄, w̄) rather than (z, w) under the corresponding controllers.
Using linear system theory, it is easy to design an F ∈ Rm×n , along
with a matrix X > 0 such that A + BF is Hurwitz and the set

E(X) = v ∈ Rn : v T Xv ≤ 1
is inside the linear region of the saturation function sat(F v) and that for
some positive number d1 , E(X) is invariant for the system
v̇ = Av + Bsat(F v) − η, |η|∞ ≤ d1 . (12.5.6)
Let α and N be determined from X in the same way as with the state
feedback controller. With F ∈ Rm×n , we form a sequence of controllers,
% &
F (z̄ − αk V w̄)
u = fk (z̄, w̄, α) = (1 − α )Γw̄ + α sat
k k
, k = 0, 1, · · · , N,
αk
and % &
F z̄
u = fN +1 (z̄, w̄) = Γw̄ + δ sat .
δ
Under the control u = fk (z̄, w̄, α), consider vk = z̄ − αk V w̄, then we get
% &
F vk
v̇k = Avk + αk B sat − (L1 − αk V L2 )C z̃. (12.5.7)
αk
Note the difference between this equation and the corresponding (12.4.12)
in the state feedback case. Here we need to take into account the extra
term (L1 − αk V L2 )C z̃. For clarity, we split the discussion into three cases.
Case 1. k = 0
Let v = z̄ − V w̄. Then the system
v̇ = Av + Bsat(F v) − (L1 − V L2 )C z̃,

has an invariant set E(X) for all z̃ satisfying
|(L1 − V L2 )C z̃|∞ ≤ d1 . (12.5.8)
Since
lim z̃(t) = 0,
t→∞
we also have
lim v(t) = 0, ∀ v0 ∈ E(X).
t→∞
Case 2. 0 < k ≤ N
Similar to Lemma 12.4.3, we have an invariant set αk E(X) for the system
(12.5.7) under the restriction

(L1 − αk V L2 )C z̃ ≤ αk d1 . (12.5.9)
∞
Also, because
lim z̃(t) = 0,
t→∞
we have
lim vk (t) = 0, ∀ v0 ∈ αk E(X).
t→∞
Case 3. k = N + 1
We have an invariant set δE(X) for the system

% &
F z̄
z̄˙ = Az̄ + δB sat − L1 C z̃
δ
under the restriction

|L1 C z̃|∞ ≤ δd1 . (12.5.10)
Now we define a sequence of sets in Rn × Rr :

Dz̄k w̄ = (z̄, w̄) ∈ Rn × Rr : z̄ − αk V w̄ ∈ αk E(X) , k = 0, 1, 2, · · · , N,
and
Dz̄Nw̄+1 = (z̄, w̄) ∈ Rn × Rr : z̄ ∈ δE(X) .
A counterpart of Lemma 12.4.3 is,

Lemma 12.5.3. Suppose that F is chosen such that E(X) is invariant for
the system (12.5.6) and that z̃ satisfies all the conditions (12.5.8), (12.5.9)
and (12.5.10). Then, under the control u = fk (z̄, w̄, α), the set Dz̄k w̄ is
invariant for the system (12.5.1) and

lim z̄(t) − αk V w̄(t) = 0.
t→∞
With the preliminaries we developed above, we can now construct an

error feedback law as follows:
u = sat(g(z̄, w̄, α, N )), (12.5.11)
where


 fN +1 (z̄, w̄), if (z̄, w̄) ∈ ΩN +1 := Dz̄Nw̄+1 ,


 +1 j
fk (z̄, w̄, α), if (z̄, w̄) ∈ Ωk := Dz̄k w̄ \ ∪N j=k+1 Dz̄ w̄ ,
g(z̄, w̄, α, N ) =

 k = 0, 1, . . . , N,


 +1 j
f (z̄ − V w̄), if (z̄, w̄) ∈ Ω := Rn × Rr \ ∪N j=0 Dz̄ w̄ .
Suppose that the controller (12.5.11) is connected to the system (12.5.2).

At t = T0 , we have z(T0 )−V w(T0 ) ∈ D0 . To apply Lemma 12.5.3 for t ≥ T0 ,
we have to ensure that all the conditions (12.5.8), (12.5.9) and (12.5.10) are
satisfied. Moreover, we also require that
|fk (z̄, w̄, α)|∞ ≤ 1, |fN +1 (z̄, w̄)|∞ ≤ 1. (12.5.12)
Note that
% &
F (z̄ − αk V w̄)
|fk (z̄, w̄, α)|∞
= (1 − α )Γw̄ + α sat
k k
≤1
αk
∞
can be satisfied if
|Γw̄|∞ ≤ |Γw|∞ + |Γw̃|∞ ≤ 1. (12.5.13)
Let d0 and d1 be given, then by Lemma 12.5.1, all the conditions
(12.5.5), (12.5.8), (12.5.9), (12.5.10) and (12.5.13) can be satisfied for t ≥ T0
by suitably choosing L. Note that, in (12.5.13), |Γw|∞ ≤ ρ < 1.
Theorem 12.5.1. Consider the feedback system (12.5.2) with (12.5.11).

Suppose that F is chosen such that E(X) is invariant for the system (12.5.6).
Also suppose that L is chosen such that the conditions (12.5.5), (12.5.8),
(12.5.9), (12.5.10) and (12.5.13) are satisfied for t ≥ T0 . Then for all
(z0 , w0 , z̃ 0 , w̃ 0 ) ∈ DM ,
lim z(t) = 0.
t→∞
Proof. At t = T0 , (z̄, w̄) belongs to one of Ω, Ωk , k = 0, 1, · · · , N + 1.
Case 1. (z̄(T0 ), w̄(T0 )) ∈ Ω
Then u = f (z̄ − V w̄) will be in effect. This is just the situation described
in Lemma 12.5.2. So we have
lim (z̄(t) − V w̄(t)) = 0

t→∞
and similar to the proof of Theorem 12.4.1, it can be shown that (z̄, w̄) will
enter Ω0 , or some other Ωk at a finite time.
Case 2. (z̄(T0 ), w̄(T0 )) ∈ Ωk
Since z̃(t), t ≥ T0 satisfies the conditions (12.5.8), (12.5.9) and (12.5.10),

each Dz̄k w̄ is invariant under the control u = fk (z̄, w̄, α). Since
lim z̃(t) = 0,
t→∞
we have (see 12.5.7)

lim z̄(t) − αk V w̄(t) = lim vk (t) = 0.
t→∞ t→∞
Similar to the proof of Theorem 12.4.1, it can be shown that (z̄, w̄) will
enter Ωk1 for some k1 > k.
By repeating this procedure, we will have (z̄, w̄) ∈ ΩN +1 at some finite
time, and hence
lim z̄(t) = 0.
t→∞
Also, since
lim z̃(t) = 0,
t→∞
we have
lim z(t) = 0.
t→∞
Remark 12.5.1. It is difficult to tell which set will (z̄, w̄) belong for the
time interval [0, T0 ] because of the observation error. It might switch among
the sets Ω, Ωk , k = 0, · · · , N + 1, frequently. To avoid frequent switching of
the controller, we may simply set u = 0 for t < T0 .
12.6. An Example 297
12.6. An Example
In this section, we will apply the results developed above to the control of
an aircraft. Consider the longitudinal dynamics of the TRANS3 aircraft
under certain flight condition [51],
ẋ = Ax + Bu,
(12.6.1)
y = Cx,
with  
0 14.3877 0 −31.5311
 −0.0012 −0.4217 1.0000 −0.0284 
A=
 0.0002 −0.3816 −0.4658
,
0 
0 0 1.0000 0
 
4.526
 −0.0337 

B = 0.1745 ×  ,
−1.4566 
0
and

C= 1 0 0 0 ,
where the state consists of the velocity x1 , the angle of attack x2 , the pitch
rate x3 and the Euler angle rotation of aircraft about the inertial y-axis x4 ,
the control u is the elevator input, whose value is scaled between +1 and −1
(corresponding to ±10o). The design objective is to reject the disturbance
P w, where
 
−0.6526 −0.3350 0.4637 0.9185
 0.0049 0.0025 −0.0035 −0.0068 
P =  0.2100
,
0.1078 −0.1492 −0.2956 
0 0 0 0
and w contains the frequencies of 0.1 rad/s and 0.3 rad/s. Clearly, this
problem can be cast into an output regulation problem for the system
(12.2.6) with  
0 −0.1 0 0
 0.1 0 0 0 
S=  0
,
0 0 −0.3 
0 0 0.3 0
and Q = 0. A solution to the linear matrix equations (12.3.1) is

Π = 0, Γ = 0.8263 0.4242 −0.5871 −1.1630 .
Assume that the disturbance is bounded by Γw∞ ≤ ρ = 0.9. Thus,

δ = 0.1. Since Π = 0, we have z = x − Πw = x.
The matrix A has two stable eigenvalues −0.4650 ± j0.6247 and two
anti-stable ones, 0.0212 ± j0.1670. With state transformation, we get the
matrices for the anti-stable subsystem:

0.0212 0.1670 8.2856
A2 = , B2 = ,
−0.1670 0.0212 −2.4303
and V2 is solved as

17.8341 46.6790 −46.5057 25.0374
V2 = .
66.6926 4.4420 21.7554 22.6661
We don’t need to worry about the exponentially stable z1 -subsystem since
its state is bounded under any bounded input u − Γw and will converge
to the origin as the combined input goes to zero. We now investigate two
types of controllers.
Case 1. The state feedback
With the design method in Chapter 4, we obtain a semi-global
stabilizer

u = f (z2 ) = sat(βF0 z2 ), F0 = −0.0093 0.0041 ,
for the system in the absence of disturbance. With β = 1, the achieved do-
main of attraction is very close to the null controllable region (see Fig. 12.6.1,
where the outermost solid closed curve is the boundary of C 2 and the dotted
closed curve is the boundary of the domain of attraction S 2 ).
With the method proposed in Section 12.4, we take F = F0 and the
switching parameters be α = 0.4 and N = 3. So there are at most 5
switches in the controller u = g(z2 , w, α, 3). The closed-loop system is
simulated with z20 − V2 w0 very close to the boundary of S 2 (see the point
marked with “◦” in Fig. 12.6.1). The trajectory shown in the figure is that
of z2 (t) with the initial state marked with “∗”. We note here that the initial
state z20 is not necessarily inside S 2 , not even inside the null controllable
region. But z20 −V2 w0 has to be in the null controllable region, i.e., (z20 , w0 )
has to be in the regulatable region Rg 2 .
The output e = Cz is plotted in Fig. 12.6.2. The control (solid curve)
and the switching history (dash-dotted curve) are plotted in Fig. 12.6.3,
where for the dash-dotted curve, the number −0.2 indicates that the control
u = f (z2 − V2 w) = sat(F (z2 − V2 w))
12.6. An Example 299
300
200
100
−100
−200
−300
−300 −200 −100 0 100 200 300
Figure 12.6.1: A trajectory of z2 – state feedback.
is applied and the number 0.2k, indicates that the controller
u = fk (z2 , w, α), k = 0, 1, · · · , 4,
is applied, respectively. We see that the controller u = f3 (z2 , w, α) is

skipped.
Case 2. The error feedback
Here we assume that only the error signal is available. So we need to build
an observer to reconstruct the plant and exosystem states. For this system,
% &
A −BΓ
C 0 ,
0 S
is observable.
With the design method proposed in Chapter 6, we obtain a controller

u = f (v2 ) = sat(F0 v2 ), F0 = −0.0175 0.0103 ,
such that the system
v̇2 = A2 v2 + B2 f (v2 + η)
1500
1000
500
error e
−500
−1000
0 50 100 150 200 250 300 350
time t
Figure 12.6.2: The time response of the error – state feedback.
1.5
1
−: control u; −.: the switching history
0.5
−0.5
−1
−1.5
0 50 100 150 200 250 300 350
time t
Figure 12.6.3: The control and the switching history – state feedback.
satisfies an asymptotic bound from D0 with a finite gain and nonzero re-
striction, where D0 is also very close to the null controllable region (see
Fig. 12.6.4, where the outermost solid closed curve is the boundary of C 2
and the dash-dotted closed curve is the boundary of D0 ).
With the error feedback design method proposed in this chapter, we
obtain for the anti-stable subsystem

F = 0.0378 0.0357
and the switching parameters α = 0.5 and N = 3. Also, to achieve fast

recovery of the state, we choose

L= 24.11 −263.61 68.59 −115.96 161070.21
T
−231551.67 65778.10 −3347.56 ,
for the observer.

The closed-loop system is simulated with z20 − V w0 very close to the
boundary of D0 (see the point marked with “◦” in Fig. 12.6.4). We simply
set the initial state of the observer 0. The trajectory shown in the figure is
that of z2 (t), with the initial point marked with “∗”.
The error output e = Cz is plotted in Fig. 12.6.5 and the observer error
is plotted in Fig. 12.6.6. The control and the switching history is plotted
in Fig. 12.6.7. The simulation results verify the effectiveness of our design
method.
12.7. Conclusions
In this chapter, we have systematically studied the problem of output regu-
lation for linear systems subject to actuator saturation. The plants consid-
ered here are general and can be exponentially unstable. We first charac-
terized the asymptotically regulatable region, the set of initial conditions of
the plant and the exosystem for which output regulation can be achieved.
We then constructed feedback laws, of both state feedback and error feed-
back type, that achieve output regulation on the asymptotically regulatable
region.
300
200
100
−100
−200
−300
−300 −200 −100 0 100 200 300
Figure 12.6.4: The trajectory of z2 – error feedback.
1500
1000
500
error e
−500
−1000
−1500
0 50 100 150 200 250 300 350
time t
Figure 12.6.5: The error output e = Cz – error feedback.

6
x 10
5
The observer error

1
−1
−2
−3
−4
−5
0 5 10 15 20 25 30
time t
Figure 12.6.6: The observer error z̃ and w̃ – error feedback.
1.5
1
−: control u; −.: the switching history
0.5
−0.5
−1
−1.5
0 50 100 150 200 250 300 350
time t
Figure 12.6.7: The control and the switching history – error feedback.
Chapter 13
Output Regulation –
Discrete-Time
Systems
13.1. Introduction
In Chapter 12, we systematically studied the problem of output regula-

tion for continuous-time linear systems subject to actuator saturation. In
particular, we characterized the regulatable region, the set of plant and
exosystem initial conditions for which output regulation is possible with
the saturating actuators. It turned out that the asymptotically regulatable
region can be characterized in terms of the null controllable region of the
anti-stable subsystem of the plant. We then constructed feedback laws that
achieve regulation on the regulatable region. These feedback laws were con-
structed from the stabilizing feedback laws in such a way that a stabilizing
feedback law that achieves a larger domain of attraction leads to a feedback
law that achieves output regulation on a larger subset of the regulatable
region and, a stabilizing feedback law on the entire asymptotically null con-
trollable region leads to a feedback law that achieves output regulation on
the entire asymptotically regulatable region.
The objective of this chapter is to extend the above results to discrete-
time systems. In Section 13.2, we formulate the problem of output reg-
ulation for discrete-time linear systems with saturating actuators. As a
motivation to our problem formulation and our solution to the problem,
305
306 Chapter 13. Output Regulation – Discrete-Time Systems
we will also briefly state the discrete-time version of the classical results on
the problem of output regulation for continuous-time linear systems of [22].
Section 13.3 characterizes the regulatable region. Sections 13.4 and 13.5
respectively construct state feedback and error feedback laws that achieve
output regulation on the regulatable region. Finally, Section 13.6 draws a
brief conclusion to the chapter.
13.2. Preliminaries and Problem Statement
In this section, we state the discrete-time version of the classical results on

the problem of output regulation for continuous-time linear systems [22,48].
These results will motivate our formulation of as well as the solution to the
problem of output regulation for discrete-time linear systems subject to
actuator saturation.
13.2.1. Review of Linear Output Regulation Theory
Consider a discrete-time linear system
x(k + 1) = Ax(k) + Bu(k) + P w(k),

w(k + 1) = Sw(k), (13.2.1)
e(k) = Cx(k) + Qw(k).
As discussed in Section 12.2.1, the first equation of this system describes

a plant, with state x ∈ Rn and input u ∈ Rm , subject to the effect of
a disturbance represented by P w. The third equation defines the error
e ∈ Rq between the actual plant output Cx and a reference signal −Qw
that the plant output is required to track. The second equation describes an
autonomous system, often called the exogenous system or the exosystem,
with state w ∈ Rr . The exosystem models the class of disturbances and
references taken into consideration.
The control action to the plant, u, can be provided either by state
feedback or by error feedback. A state feedback controller has the form of
u(k) = F x(k) + Gw(k), (13.2.2)
while an error feedback controller has the form of
ξ(k + 1) = Ac ξ(k) + Bc e(k), ξ ∈ Rl ,

(13.2.3)
u(k) = Cc ξ(k) + Dc e(k).
The objective is to achieve internal stability and output regulation. Internal

stability means that, if we disconnect the exosystem and set w equal to zero,
then the closed-loop system is asymptotically stable. Output regulation
means that for any initial conditions of the closed-loop system, we have
that e(k) → 0 as k → ∞. Precise problem statements are similar to their
continuous-time counterparts, Definitions 12.2.1 and 12.2.2.
The solution to these problems is based on the following three assump-
tions.
Assumption 13.2.1.
A1. The eigenvalues of S are on or outside of the unit circle.
A2. The pair (A, B) is stabilizable.
A3. The pair % &

A P
C Q ,
0 S
is detectable.
For continuous-time systems, complete solutions to the output regula-

tion problems were established in [22] by Francis. These solutions can be
adapted for discrete-time systems as follows.
Proposition 13.2.1. Suppose Assumptions A1 and A2 hold. Then, the

problem of output regulation by state feedback is solvable if and only if
there exist matrices Π and Γ that solve the linear matrix equations
(13.2.4)
CΠ + Q = 0.
Moreover, if in addition Assumption A3 also holds, the solvability of the
above linear matrix equations is also a necessary and sufficient condition
for the solvability of the problem output regulation by error feedback.
13.2.2. Output Regulation in the Presence of

Actuator Saturation
Motivated by the classical formulation of output regulation for linear sys-

tems, we consider the following plant and the exosystem,
x(k + 1) = Ax(k) + Bu(k) + P w(k),
w(k + 1) = Sw(k), (13.2.5)
e(k) = Cx(k) + Qw(k),
where u is the output of saturating actuators and is constrained by |u|∞ ≤

1. A control u that satisfies this constraint is referred to as an admissible
control. Because of the bound on the control input, both the plant and
the exosystem cannot operate in the entire state space. For this reason, we
assume that
(x0 , w0 ) ∈ Y 0
for some Y 0 ⊂ Rn × Rr . Let

X 0 = x0 ∈ Rn : (x0 , 0) ∈ Y 0 .
The problem to be addressed in this chapter is the following.
Problem 13.2.1. The problem of output regulation by state feedback for

the system (13.2.5) is to find, if possible, a state feedback law u = φ(x, w),
with |φ(x, w)|∞ ≤ 1 and φ(0, 0) = 0, such that
1. the equilibrium x = 0 of the system
x(k + 1) = Ax(k) + Bφ(x(k), 0)
is asymptotically stable with X 0 contained in its domain of attraction;
2. for all (x0 , w0 ) ∈ Y 0 , the interconnection of (13.2.5) and the feedback

law u = φ(x, w) results in bounded state trajectories x(k) and
lim e(k) = 0.
k→∞
If only the error e is available, the state (x, w) can be reconstructed

after a finite number of steps. But the initial condition (x0 , w0 ) might have
to be constrained in a subset of Y 0 .
Our objective is to characterize the maximal set of initial conditions
(x0 , w0 ), the largest possible Y 0 , on which the above problem is solvable
and to explicitly construct feedback law that actually solves the problem
for Y 0 as large as possible.
We will assume that (A, B) is stabilizable. We will also assume that
S is neutrally stable and all its eigenvalues are on the unit circle. The
stabilizability of (A, B) is clearly necessary for the stabilization of the plant.
The assumption on S is without loss of generality because asymptotically
stable modes in the exosystem do not affect the regulation of the output
and output regulation cannot be expected with bounded controls if the
exosystem has unstable modes.
13.3. The Regulatable Region

In this section, we will characterize the set of all initial states of the plant
and the exosystem on which the problem of output regulation is solvable
under the restriction that u∞ ≤ 1. We will refer to this set as the
asymptotically regulatable region.
To begin with, we observe from the linear output regulation theory
(see Section 13.2.1) that for this problem to be solvable, there must exist
matrices Π ∈ Rn×r and Γ ∈ Rm×r that solve the following linear matrix
equations,
(13.3.1)
0 = CΠ + Q.
Given the matrices Π and Γ, we define a new state z = x − Πw and
rewrite the system equations as
z(k + 1) = Az(k) + Bu(k) − BΓw(k),
w(k + 1) = Sw(k), (13.3.2)
e(k) = Cz(k).
From these new equations, it is clear that e(k) goes to zero asymptotically
if z(k) goes to zero asymptotically. This is possible only if

sup ΓS k w0 ∞ < 1. (13.3.3)
k≥0
For this reason, we will restrict our attention to exosystem initial conditions
in the following compact set

W 0 = w0 ∈ Rr : |Γw(k)|∞ = ΓS k w0 ∞ ≤ ρ, ∀ k ≥ 0 , (13.3.4)
for some ρ ∈ [0, 1). For later use, we also denote δ = 1 − ρ. We note that
the compactness of W 0 can be guaranteed by the observability of (Γ, S).
Indeed, if (Γ, S) is not observable, then the exosystem can be reduced to
make it so.
We can now precisely define the notion of asymptotically regulatable
region as follows.
Definition 13.3.1.
1) Given K > 0, a pair (z0 , w0 ) ∈ Rn × W 0 is regulatable in K steps if
there exists an admissible control u, such that the response of (13.3.2)
satisfies z(K) = 0. The set of all (z0 , w0 ) regulatable in K steps is
denoted as Rg (K).
2) A pair (z0 , w0 ) is regulatable if (z0 , w0 ) ∈ Rg (K) for some K < ∞.

The set of all regulatable (z0 , w0 ) is referred to as the regulatable
region and is denoted as Rg .
3) The set of all (z0 , w0 ) for which there exist admissible controls such
that the response of (13.3.2) satisfies
lim z(k) = 0
k→∞
is referred to as the asymptotically regulatable region and is denoted

as Rag .
Remark 13.3.1. We would like to note that the regulatable region is de-
fined in terms of
lim z(k) = 0
k→∞
rather than
lim e(k) = 0.
k→∞
Requiring the former instead of the latter will also guarantee the closed-
loop stability in the absence of w. As in the continuous-time case (Chapter
12), this will result in essentially the same description of the regulatable
region.
We will describe Rg (K), Rg and Rag in terms of the asymptotically null

controllable region of the plant
v(k + 1) = Av(k) + Bu(k), u∞ ≤ 1.
Definition 13.3.2. The null controllable region at step K, denoted as

C(K), is the set of v0 ∈ Rn that can be driven to the origin in K steps and
the null controllable region, denoted as C, is the set of v0 ∈ Rn that can be
driven to the origin in finite number of steps by admissible controls. The
asymptotically null controllable region, denoted as C a , is the set of all v0
that can be driven to the origin asymptotically by admissible controls.
Clearly,
C= C(K)
K∈[0,∞)
and K−1

−i−1
C(K) = A Bu(i) : u∞ ≤ 1 . (13.3.5)
i=0
It is also clear that the null controllable region and the asymptotically null
controllable region are identical if the pair (A, B) is controllable. Some
simple methods to describe C and C a were developed in Chapter 3.
To simplify the characterization of Rg and Rag and without loss of gen-
erality, let us assume that

z1
z= , z1 ∈ Rn1 , z2 ∈ Rn2 ,
z2
and
A1 0 B1
A= , B= , (13.3.6)
0 A2 B2
where A1 ∈ Rn1 ×n1 is semi-stable (i.e., all its eigenvalues are on or inside
the unit circle) and A2 ∈ Rn2 ×n2 is anti-stable (i.e., all its eigenvalues are
outside of the unit circle). The anti-stable subsystem
z2 (k + 1) = A2 z2 (k) + B2 u(k) − B2 Γw(k),

(13.3.7)
w(k + 1) = Sw(k),
is of crucial importance. Denote its regulatable regions as Rg 2 (K) and

Rg 2 , and the null controllable regions of the system
v2 (k + 1) = A2 v2 (k) + B2 u(k)
as C 2 (K) and C 2 . Then, the asymptotically null controllable region of the

system
v(k + 1) = Av(k) + Bu(k)
is given by C a = Rn1 × C 2 (see Proposition 3.2.1), where C 2 is a bounded
convex open set. Denote the closure of C 2 as C 2 , then
∞

−i−1
C2 = A2 B2 u(i) : u∞ ≤ 1 .
i=0
Also, if a set D is in the interior of C 2 , then there is a finite integer K > 0

such that D is in the interior of C 2 (K).
Theorem 13.3.1. Let V2 ∈ Rn2 ×r be the unique solution to the linear

matrix equation
−A2 V2 + V2 S = −B2 Γ (13.3.8)
and let
V (K) = V2 − A−K V2 S K .
Then,

a) Rg 2 (K) = (z2 , w) ∈ Rn2 × W 0 : z2 − V (K)w ∈ C 2 (K) ; (13.3.9)

b) Rg 2 = (z2 , w) ∈ Rn2 × W 0 : z2 − V2 w ∈ C 2 ; (13.3.10)
c) Rag = Rn1 × Rg 2 . (13.3.11)
Proof.
a) Given (z20 , w0 ) ∈ Rn2 × W 0 and an admissible control u, the solution
of (13.3.7) at k = K is,

K−1
K−1
z2 (K) = AK z20 + A−i−1
2 B2 u(i) − A−i−1
2 B2 ΓS i w0 . (13.3.12)
i=0 i=0
Applying (13.3.8), we have

K−1
K−1
− A−i−1
2 B2 ΓS i = A−i−1
2 (−A2 V2 + V2 S)S i
i=0 i=0

K−1

= −A−i i −i−1
2 V2 S + A2 V2 S i+1
i=0
= −V2 + A−K
2 V2 S
K
= −V (K), (13.3.13)
where the third “=” is simply obtained by expanding the terms in the
summation and cancelling all the middle terms. Thus,

K−1
A−K z2 (K) = z20 − V (K)w0 + A−i−1 B2 u(i).
i=0
By setting z2 (K) = 0, we immediately obtain a) from the definition of

Rg (K) and (13.3.5).
b) Since A2 is anti-stable and S is neutrally stable, we have that
lim V (K) = V2 .
K→∞

∞

V2 = A−i−1 B2 ΓS i . (13.3.14)
i=0
First we show that
z20 − V2 w0 ∈ C 2 =⇒ (z20 , w0 ) ∈ Rg 2 .
Since C 2 is open, if z20 − V2 w0 ∈ C 2 , then there exists an ε > 0 such that

z20 − V2 w0 + z2 : |z2 |∞ ≤ ε ⊂ C 2 .
Also, there exists a K1 > 0 such that

z20 − V2 w0 + z2 : |z2 |∞ ≤ ε ⊂ C 2 (K1 ).
Since
lim V (K) = V2 ,
K→∞
there is a K2 > K1 such that z20 − V (K2 )w0 ∈ C 2 (K1 ) ⊂ C 2 (K2 ). It follows
from a) that (z20 , w0 ) ∈ Rg 2 (K2 ) ⊂ Rg 2 .
Next we show that
(z20 , w0 ) ∈ Rg 2 =⇒ z20 − V2 w0 ∈ C 2 .
If (z20 , w0 ) ∈ Rg 2 , then (z20 , w0 ) ∈ Rg 2 (K1 ) for some K1 > 0. It follows

from the definition of Rg (K) that there is an admissible control u1 such
that

K 1 −1 1 −1
K
z20 − A−i−1 B2 ΓS i w0 + A−i−1 B2 u1 (i) = 0. (13.3.15)
i=0 i=0
Denote
Z2 = δA−K
2
1
v2 : v2 ∈ C 2 .
For each z2 ∈ Z2 , there is an admissible control u2 such that
∞
∞

z2 = δA−K
2
1
A−i−1 B2 u2 (i) = A−i−1 B2 δu2 (i − K1 ). (13.3.16)
i=0 i=K1
It follows from (13.3.14), (13.3.15) and (13.3.16) that

∞
∞

z20 − V2 w0 + z2 = z20 − A−i−1 B2 ΓS i w0 + A−i−1 B2 δu2 (i − K1 )
i=0 i=K1
1 −1
K
= z20 − A−i−1 B2 ΓS i w0
i=0
∞

+ A−i−1 B2 δu2 (i − K1 ) − ΓS i w0
i=K1

K 1 −1
=− A−i−1 B2 u1 (i)
i=0
∞

+ A−i−1 B2 δu2 (i − K1 ) − ΓS i w0
i=K1
∈ C2.
The last step follows from the fact that |ΓS i w0 |∞ ≤ ρ = 1 − δ for all i ≥ 0
and thus |δu2 (i − K1 ) − ΓS i w0 |∞ ≤ 1. This implies that

z20 − V2 w0 + z2 : z2 ∈ Z2 ⊂ C 2 .
Since A−K
2
1
is nonsingular, the set Z2 contains the origin in its interior. It
follows that z20 − V2 w0 ∈ C 2 .
c) It is easy to see that Rag ⊂ Rn1 × Rg 2 . We need to show that
R × Rg 2 ⊂ Rag .
n1
Suppose that (z0 , w0 ) = (z10 , z20 , w0 ) ∈ Rn1 ×Rg 2 . There exist a K ≥ 0

and an admissible u1 such that z2 (K) = 0. For k ≥ K, let u = Γw + uδ ,
with uδ ∞ ≤ δ = 1 − ρ. Since w0 ∈ W 0 , we have Γw∞ ≤ ρ and hence,
u∞ ≤ 1. With u = Γw + uδ , we have,
z(k + 1) = Az(k) + Bu(k) − BΓw(k) = Az(k) + B uδ (k).
Since
z1 (K)
0
is inside the asymptotically null controllable region of the above system
under the constraint uδ ∞ ≤ δ (see Chapter 3), there exists a uδ such
that
lim z(k) = 0.
k→∞
Hence (z0 , w0 ) ∈ Rag . This establishes that Rn1 × Rg 2 ⊂ Rag and hence
Rn1 × Rg 2 = Rag . 2
Remark 13.3.2. Given (z0 , w0 ), there exists an admissible control u such

that
lim z(k) = 0
k→∞
if and only if (z0 , w0 ) ∈ Rag . Recalling that z = x − Πw, we observe that,

for a given pair of initial states in the original coordinates, (x0 , w0 ), there
is an admissible control u such that
lim (x(k) − Πw(k)) = 0

k→∞
if and only if
x20 − (Π2 + V2 )w0 ∈ C 2 ,
where

Π2 = 0 In2 Π.
13.4. State Feedback Controllers
In this section, we will construct a feedback law that solves the problem of
output regulation by state feedback for linear systems subject to actuator
saturation. We assume that a stabilizing state feedback law u = f (v),
|f (v)|∞ ≤ 1 for all v ∈ Rn , has been designed and the equilibrium v = 0 of
the closed-loop system
v(k + 1) = Av(k) + Bf (v(k)) (13.4.1)
has a domain of attraction S ⊂ C a . We will construct our feedback law

from this f (v).
Now consider the system (13.3.2). Given a state feedback
u = g(z, w), |g(z, w)|∞ ≤ 1, ∀ (z, w) ∈ Rn × W 0 ,
we have the closed-loop system
z(k + 1) = Az(k) + B g(z(k), w(k)) − BΓw(k), (13.4.2)

w(k + 1) = Sw(k).
Denote the time response of z(k) to the initial state (z0 , w0 ) as z(k, z0 , w0 )
and define

S zw := (z0 , w0 ) ∈ R × W 0 : lim z(k, z0 , w0 ) = 0 .
n
k→∞
Since Rag is the set of all (z0 , w0 ) for which z(k) can be driven to the origin
asymptotically, we must have S zw ⊂ Rag . Our objective is to design a
control law u = g(z, w) such that S zw is as large as possible, or as close to

Rag as possible.
First we need a mild assumption which will be removed later. Assume
that there exists a matrix V ∈ Rn×r such that
−AV + V S = −BΓ. (13.4.3)
This will be the case if A and S have no common eigenvalues. With the
decomposition in (13.3.6), if we partition V accordingly as

V1
V = ,
V2
then V2 satisfies
−A2 V2 + V2 S = −B2 Γ.
Denote
Dzw := (z, w) ∈ Rn × W 0 : z − V w ∈ S . (13.4.4)
Similar to the continuous-time case, we have the following observation.
Observation 13.4.1.
a) The set Dzw increases as S increases, and if S = C a , then Dzw = Rag ;
b) In the absence of w,
x0 ∈ S =⇒ (z0 , 0) ∈ Dzw .
With this observation, we see that our objective of enlarging S zw is

simply to design a feedback law such that Dzw ⊂ S zw . We will reach this
objective through a series of technical lemmas.
Lemma 13.4.1. Let u = f (z − V w). Consider the closed-loop system
z(k + 1) = Az(k) + Bf (z(k) − V w(k)) − BΓw(k),

(13.4.5)
w(k + 1) = Sw(k).
For this system, Dzw is an invariant set and for all (z0 , w0 ) ∈ Dzw ,
lim (z(k) − V w(k)) = 0.

k→∞
Proof. Substitute (13.4.3) into system (13.4.5), we obtain
z(k + 1) = Az(k) + Bf (z(k) − V w(k)) − AV w(k) + V S w(k)

= A(z(k) − V w(k)) + Bf (z(k) − V w(k)) + V w(k + 1).
Define the new state v := z − V w, we have
v(k + 1) = Av(k) + Bf (v(k)),
which has a domain of attraction S. This also implies that S is an invariant

set for the v-system.
If (z0 , w0 ) ∈ Dzw , then v0 = z0 − V w0 ∈ S. It follows that
v(k) = z(k) − V w(k) ∈ S, ∀ k ≥ 0,
and
lim (z(k) − V w(k)) = lim v(k) = 0.
k→∞ k→∞
Lemma 13.4.1 says that, in the presence of w, the simple feedback u =

f (z − V w) will cause z(k) − V w(k) to approach zero and z(k) to approach
V w(k), which is bounded. Our next step is to construct a finite sequence
of controllers
u = f (z, w, α), = 0, 1, 2, · · · , N,
all parameterized in α ∈ (0, 1). By judiciously switching between these

controllers, we can cause z(k) to approach α V w(k) for any . By choosing
N large enough, z(k) will become arbitrarily small in a finite number of
steps. Once z(k) becomes small enough, we will use the controller
% &
Fz
u = Γw + δ sat ,
δ
with F to be specified later, to make z(k) converge to the origin.

Let F ∈ Rm×n be such that
v(k + 1) = Av(k) + B sat(F v(k)), (13.4.6)
is asymptotically stable at the origin. Let X > 0 be such that
(A + BF )T X(A + BF ) − X < 0
and the ellipsoid

E(X) := v ∈ Rn : v T Xv ≤ 1
is in the linear region of the saturation function sat(F v), i.e.,
|F v|∞ ≤ 1, ∀ v ∈ E(X).
Then E(X) is an invariant set and is in the domain of attraction for the
closed-loop system (13.4.6).
Lemma 13.4.2. Suppose that D ⊂ Rn is an invariant set in the domain

of attraction for the system
v(k + 1) = Av(k) + Bf (v(k)), (13.4.7)
then for any α > 0, αD is an invariant set in the domain of attraction for
the system % &
v(k)
v(k + 1) = Av(k) + αBf . (13.4.8)
α
Proof. Write (13.4.8) as
% &
v(k + 1) v(k) v(k)
=A + Bf ,
α α α
v(k)
and replace α with v(k), then we get (13.4.7). If v0 ∈ αD, i.e.,
v0
∈ D,
α
then,
v(k)
∈ D, ∀k > 0
α
and
lim v(k) = 0.
k→∞
2
For any α ∈ (0, 1), there exists a positive integer N such that
1

αN X 2 V w < δ, ∀ w ∈ W 0 , (13.4.9)
i.e., αN V w ∈ δE(X), for all w ∈ W 0 . Define a sequence of subsets in

Rn × W 0 as,

Dzw = (z, w) ∈ Rn × W 0 : z − α V w ∈ α E(X) , = 0, 1, · · · , N,

N +1
Dzw = (z, w) ∈ Rn × W 0 : z ∈ δE(X) ,
and, on each of these sets, define a state feedback law as follows,

% &
F (z − α V w)
f (z, w, α) = (1 − α )Γw + α sat

, = 0, 1, · · · , N,
α
% &
Fz
fN +1 (z, w) = Γw + δ sat .
δ
It can be verified that, for each = 0 to N + 1,
|f (z, w, α)|∞ ≤ 1, ∀ (z, w) ∈ Rn × W 0 .
Lemma 13.4.3. Let u = f (z, w, α). Consider the closed-loop system
z(k + 1) = Az(k) + Bf (z(k), w(k), α) − BΓw(k),

(13.4.10)
w(k + 1) = Sw(k).

For this system, Dzw is an invariant set. Moreover, if = 0, 1, · · · , N , then,

lim z(k) − α V w(k) = 0, ∀ (z0 , w0 ) ∈ Dzw

,
k→∞
and if = N + 1, then,
N +1
lim z(k) = 0, ∀ (z0 , w0 ) ∈ Dzw .
k→∞
Proof. With u = f (z, w, α), = 0, 1, · · · , N , we have
z(k + 1) = Az(k) + (1 − α )BΓw(k)

% &
F (z(k) − α V w(k))

+α Bsat − BΓw(k)
α
% &
F (z(k) − α V w(k))
= Az(k) + α Bsat
α
−α BΓw(k). (13.4.11)
Let v = z − α V w. Then by (13.4.3),

% &
F v (k)
v (k + 1) = Av (k) + α Bsat . (13.4.12)
α
It follows from Lemma 13.4.2 that α E(X) is an invariant set in the domain

of attraction for the v -system. Hence Dzw is invariant for the system
(13.4.10) and if (z0 , w0 ) ∈ Dzw , i.e.,

v0 = z0 − α V w0 ∈ α E(X),
then,
lim (z(k) − α V w(k)) = lim v (k) = 0.
k→∞ k→∞
For % &
Fz
u = fN +1 (z, w) = Γw + δsat ,
δ
we have % &
F z(k)
z(k + 1) = Az(k) + δBsat
δ
and the same argument applies. 2
Based on the technical lemmas established above, we construct our final

state feedback law as follows,
u = g(z, w, α, N )


 fN +1 (z, w), if (z, w) ∈ ΩN +1 := Dzw
N +1
,


 f (z, w, α), if (z, w) ∈ Ω := D \ ∪N +1 Dj ,
zw j=+1 zw
=

 = 0, 1, · · · , N,


 +1 j
f (z − V w), if (z, w) ∈ Ω := R × W 0 \ ∪N
n
j=0 Dzw .
Since Ω, Ω0 , · · · , ΩN +1 are disjoint and their union is Rn ×W 0 , the controller

is well defined on Rn ×W 0 . What remains to be shown is that this controller
will accomplish our objective if the parameter α is properly chosen.
Let
1
|X 2 V w|
α0 = max 1 .
w∈W 0 |X 2 V w| + 1
It is obvious that α0 ∈ (0, 1).
Theorem 13.4.1. Choose any α ∈ (α0 , 1) and let N be specified as in

(13.4.9). Then, for all (z0 , w0 ) ∈ Dzw , the solution of the closed-loop
system,
z(k + 1) = Az(k) + B g(z(k), w(k), α, N ) − BΓw(k), (13.4.13)

w(k + 1) = Sw(k),
satisfies
lim z(k) = 0,
k→∞
i.e., Dzw ⊂ S zw .
Proof. The control u = g(z, w, α, N ) is executed by choosing one from

f (z, w, α), = 0, 1, · · · , N + 1, and f (z − V w). The crucial point is to
guarantee that (z, w) will move successively from Ω, to Ω0 , Ω1 , · · · , finally
entering ΩN +1 , in which z(k) will converge to the origin.
Without loss of generality, we assume that (z0 , w0 ) ∈ Ω ∩ Dzw , so the
control u = f (z − V w) is in effect at the beginning. By Lemma 13.4.1,
lim (z(k) − V w(k)) = 0.

k→∞
Hence there is a finite step k0 ≥ 0 such that z(k0 ) − V w(k0 ) ∈ E(X), i.e.,
0
(z(k0 ), w(k0 )) ∈ Dzw . The condition (z(k), w(k)) ∈ Dzw
, > 0, might be
satisfied at a smaller step k1 ≤ k0 . In any case, there is a finite step k1 ≥ 0
such that
+1
(z(k1 ), w(k1 )) ∈ Ω = Dzw

\ ∪N j
j=+1 Dzw
for some = 0, 1, · · · N + 1. After that, the control u = f (z, w, α) will be

in effect.
We claim that, for any (z(k1 ), w(k1 )) ∈ Ω , under the control u =
f (z, w, α), there is a finite integer k2 > k1 such that (z(k2 ), w(k2 )) ∈ Dzw
+1
.
Since Ω ⊂ Dzw , by Lemma 13.4.3, we have that, under the control

u = f (z, w, α),
lim (z(k) − α V w(k)) = 0.
k→∞
Since α ∈ (α0 , 1), we have

1

(1 − α) X 2 V w < α, ∀ w ∈ W 0.
Therefore, for < N ,

1 1 1
2
X (z − α+1 V w) ≤ X 2 (z − α V w) + α (1 − α) X 2 V w
1

< X 2 (z − α V w) + α+1 . (13.4.14)
Since the first term on the right-hand side goes to zero asymptotically, there
exists a finite k2 > k1 such that
1
2
X (z(k2 ) − α+1 V w(k2 )) ≤ α+1 .
This implies that
z(k2 ) − α+1 V w(k2 ) ∈ α+1 E(X),

i.e.,
(z(k2 ), w(k2 )) ∈ Dzw
+1
.
If = N , then, by (13.4.9),
1 1 1
2 2
X z ≤ X (z − αN V w) + αN X 2 V w
1

< X 2 (z − αN V w) + δ.
Also, the first term goes to zero asymptotically, so there exists a finite
integer k2 such that 1
2
X z(k2 ) ≤ δ,
N +1
i.e., (z(k2 ), w(k2 )) ∈ Dzw .
+q
Just as before, (z, w) might have entered Dzw , q > 1, before it enters
+1
Dzw . In any case, there is a finite k such that
+1
(z(k), w(k)) ∈ Ω+q = Dzw
+q
\ ∪N j
j=+q+1 Dzw ,
for some q ≥ 1. After that, the controller will be switched to f+q (z, w, α).

It is also important to note that, by Lemma 13.4.3, Dzw is invariant
under the control u = f (z, w, α). Once (z, w) ∈ Ω ⊂ Dzw , it will never

go back to Ωq , q < (or Ω) since Ωq , q < and Ω have no intersection

with Dzw , (but Ωq , q > , might have intersection with Dzw
). In sum-
mary, for any (z0 , w0 ) ∈ Dzw , suppose (z0 , w0 ) ∈ Ω , the control will first
be f (z, w, α) and then switch successively to f1 , f2 , · · ·, with 1 , 2 , · · ·,
N +1
strictly increasing until (z, w) enters Dzw and remains there. Hence,
lim z(k) = 0.
k→∞
From the proof of Theorem 13.4.1, we see that for all (z0 , w0 ) ∈ Dzw ,
the number of switches is at most N + 2.
In what follows, we will deal with the case that there is no V that
satisfies
−AV + V S = −BΓ.
This will occur if A and S have some same eigenvalues on the unit circle.
Another case is that some eigenvalues of A and S on the unit circle are very
close. This will result in large elements of V , so α could be very close to 1
and N could be very large. The following method is derived to deal with
these two cases.
Suppose that the z-system (13.3.2) has the following form,

z1 (k + 1) A1 0 z1 (k) B1 B1 Γ
= + u(k) − w(k)
z2 (k + 1) 0 A2 z2 (k) B2 B2 Γ
(13.4.15)
n1 ×n1 n2 ×n2
with A1 ∈ R semi-stable and A2 ∈ R anti-stable. Also suppose
that there is a known function f (v2 ), |f (v2 )|∞ ≤ 1 for all v2 ∈ Rn2 , such
that the origin of the following system
v2 (k + 1) = A2 v2 (k) + B2 f (v2 (k))
has a domain of attraction S 2 , which is a bounded set. Then, by Lemma

13.4.2, the system
% &
v2 (k)
v2 (k + 1) = A2 v2 (k) + δB2 f
δ
has a domain of attraction δS 2 . With the same technique as used in [102],

which applies to the continuous-time systems, it can be shown that there
exists a control u = δ sat(h(v)) such that the origin of
v(k + 1) = Av(k) + δBsat(h(v(k)))
has a domain of attraction S δ = Rn1 × δS 2 .

Since there is a V2 satisfying
−A2 V2 + V2 S = −B2 Γ,
by the foregoing development of Theorem 13.4.1, there exists a controller

u = g(z2 , w, α, N ) such that for all (z0 , w0 ) satisfying
z20 − V2 w0 ∈ S 2 , ∀ w0 ∈ W 0 ,
the response of the closed-loop system satisfies
lim z2 (k) = 0.
k→∞
Hence, there is a finite integer k1 > 0 such that z(k1 ) ∈ Rn1 × δS 2 . After
that if we switch to the control u = Γw + δ sat(h(z), we have
z(k + 1) = Az(k) + δBsat(h(z(k)))

and z will stay in S δ = Rn1 ×δS 2 and converge to the origin asymptotically.
In summary, we have the controller,

g(z2 , w, α, N ), if z2 ∈ Rn2 \ δS 2 ,
u=
Γw + δ sat(h(z)), if z2 ∈ δS 2 ,
and under this control, the following set

(z, w) ∈ Rn × W 0 : z2 − V2 w ∈ S 2
will be a subset of S zw .
13.5. Error Feedback Controllers

In the continuous-time case of the previous chapter, the set of initial states
on which
lim z(k) = 0
k→∞
is achieved by an error feedback can be made arbitrarily close to that by
a state feedback. This is because the observer error can be made arbitrar-
ily small in an arbitrarily short time interval. However, for discrete-time
systems, this is impossible. If the pair
% &
A −BΓ
[C 0],
0 S
is observable, then there is a minimal number of steps for any observer to
reconstruct all the states. Let this minimal number of steps be n0 . We
also assume that a stabilizing state feedback law u = f (v), |f (v)|∞ ≤ 1
for all v ∈ Rn , has been designed such that the origin of the closed-loop
system (13.4.1) has a domain of attraction S ⊂ C a . By using the design of
Section 13.4, the set

Dzw := (z, w) ∈ Rn × W 0 : z − V w ∈ S ,
can be made a subset of S zw with a state feedback u = g(z, w, α, N ).

Now for the case where only the tracking error e is available for feed-
back, a simple strategy is to let the control u be zero before the states are
completely recovered, and after that we let u = g(z, w, α, N ) as in (13.4.13),
i.e.,
0, if k < n0 ,
u= (13.5.1)
g(z, w, α, N ), if k ≥ n0 .
The question is: what is the set of initial states on which
lim z(k) = 0
k→∞
under the control of (13.5.1)? The answer is very simple. With u = 0, we

have
z(k + 1) = Az(k) − BΓw(k),
w(k + 1) = Sw(k).
By applying (13.4.3), we have
(z − V w)(k + 1) = A(z − V w)(k).
Hence,
(z − V w)(n0 ) = An0 (z − V w)(0) = An0 (z0 − V w0 ).
For (z0 , w0 ) to be in S zw , it suffices to have
(z(n0 ), w(n0 )) ∈ Dzw ,
i.e.,
z(n0 ) − V w(n0 ) ∈ S.
This is in turn equivalent to

(z0 , w0 ) ∈ Dzw := (z, w) ∈ R × W 0 : An0 (z − V w) ∈ S .
In summary, we have Dzw ⊂ S zw under the control of (13.5.1).

The set Dzw is close to Dzw if An0 is close to the identity matrix.
13.6. Conclusions
In this chapter, we have systematically studied the problem of output reg-
ulation for linear discrete-time systems subject to actuator saturation. The
plants considered here are general and can be exponentially unstable. We
first characterized the regulatable region, the set of initial conditions of the
plant and the exosystem for which output regulation can be achieved. We
then constructed feedback laws, of both state feedback and error feedback
type, that achieve output regulation on the regulatable region.
Chapter 14
Linear Systems with

Non-Actuator Saturation
14.1. Introduction
Throughout the previous chapters, we have closely examined linear systems

with saturating actuators in a systematic manner. We will conclude this
book with some results on linear systems subject to non-actuator satura-
tion. We will establish necessary and sufficient conditions under which a
planar linear system under state saturation is globally asymptotically sta-
ble. Another result that we will present in this chapter is the design of a
family of linear saturated output feedback laws that would semi-globally
stabilize a class of linear systems. Here, the saturated output is a result
of sensor saturation. In general, the effects, and hence the treatment, of
sensor saturation are very different from those of actuator saturation and
are far from being systematically studied. The result presented here is only
a very special situation which happens to be dual to a special case of the
result presented in Chapter 4 on actuator saturation. Even in this special
case, the mechanism behind the feedback laws is completely different.
As mentioned in the preface of this book, saturation nonlinearity is
ubiquitous in engineering systems and presents exciting topics for research.
Other systems that involve saturation nonlinearity include recurrent neural
networks [91,92] and associate memories [74].
For clarity, we will organize these results in three separate sections. In
Sections 14.2 and 14.3, we give a complete stability analysis of a planar
327
328 Chapter 14. Non-Actuator Saturation
linear systems under state saturation. Section 14.2 deals with continuous-
time systems and Section 14.3 deals with discrete-time systems. Finally, in
Section 14.4, we present a semi-global stabilization result for linear systems
subject to sensor saturation.
14.2. Planar Linear Systems under State Saturation –

Continuous-Time Systems
14.2.1. System Description and Problem Statement
Consider the following system

ẋ = sat(Ax), x ∈ R2 . (14.2.1)
Systems of the form (14.2.1) and their discrete-time counterparts mainly
arise in neural networks [92] and digital filters [74]. As with any dynami-
cal system, stability of these systems is of primary concern and has been
heavily studied in the literature for a long period of time (see, e.g., [1,72–
74,83] and the references therein). As seen in the literature, the stability
analysis of such seemingly simple systems is highly nontrivial. Only suffi-
cient conditions for global asymptotic stability are available ([1,74,83]). In
this section, we will present a complete analysis of the system (14.2.1). In
particular, necessary and sufficient conditions for the system to be glob-
ally asymptotically stable or to have a limit circle are explicitly given in
terms of the entries of the matrix A. We will also describe a surprising,
but appealing, phenomenon that, even with an unstable matrix A, it is still
possible for the system to have a bounded global attractor.
14.2.2. Main Results on Global Asymptotic Stability
Given an initial state x0 ∈ R2 , denote the trajectory of the system (14.2.1)

that passes through x0 at t = 0 as φ(t, x0 ). The global asymptotic stability
of the system (14.2.1) is defined as follows,
Definition 14.2.1. The equilibrium x = 0 of the system (14.2.1) is said
to be stable if, for every ε > 0, there exists a δ > 0 such that
φ(t, x0 ) ≤ ε, ∀ t ≥ 0, ∀ |x0 | ≤ δ.
It is said to be globally asymptotically stable if it is a stable equilibrium
and
lim φ(t, x0 ) = 0, ∀ x0 ∈ R2 .
t→∞
14.2. State Saturation – Continuous-Time Systems 329
Also, it is said to be locally asymptotically stable if it is stable and
lim φ(t, x0 ) = 0, ∀ x0 ∈ U0 ,
t→∞
for some neighborhood U0 of x = 0.

For convenience, we will often simply say that the system (14.2.1) is
stable, locally asymptotically stable or globally asymptotically stable when
it is clear from the context that the equilibrium x = 0 is referred to.
Obviously, x = 0 is a locally asymptotically stable equilibrium if and

only if A is Hurwitz. Because of this, we assume throughout this section
that A is nonsingular, i.e., det (A)
= 0. In this case, the system has a
unique equilibrium point at the origin.
To simplify our presentation, we will use some equivalent transforma-
tions to transfer the system (14.2.1) into a standard form. First, we observe
that the dynamics of the system (14.2.1) is left unchanged under the state
transformation z = T x if

±1 0 0 ±1
T = , or .
0 ±1 ±1 0
Following the idea of [1], let z = Ax. The system (14.2.1) is then trans-
formed into the following form,
ż = Aẋ = A sat(Ax) = A sat(z).
We see that the dynamics of the system (14.2.1) and hence its stability
properties are equivalent to those of the system
ẋ = A sat(x). (14.2.2)
For this reason, we will focus on (14.2.2) in this section. This is also the rea-
son why we refer to system (14.2.1) as linear system under state saturation.
Since A is Hurwitz, at least one of its diagonal elements must be nega-

tive. Without loss of generality, we assume throughout the remaining part
of this section that

−a11 a12
A= , a11 > 0, a21 ≥ 0. (14.2.3)
−a21 a22
Otherwise, we can use
0 1
T =
1 0
as the state transformation matrix to make a11 > 0 or use

1 0
T =
0 −1
to make a21 ≥ 0.
Our main result in this section, presented in the following theorem, gives
a complete description of the stability properties of the system (14.2.2) with
A given in (14.2.3). As explained above, any system of the form (14.2.1)
with Hurwitz A can be transformed into the form of (14.2.3).
Theorem 14.2.1. The system (14.2.2) is globally asymptotically stable if

and only if A is Hurwitz and one of the following conditions is satisfied,
a) a22 < 0;
b) a22 ≥ 0, and a11 a21 ≥ a12 a22 .
On the other hand, if none of a) and b) is satisfied, the system will have
diverging trajectories and there will also be a closed trajectory.
Remark 14.2.1. We recall a recent sufficient condition for global asymp-

totic stability of the system (14.2.2) from [1]. The results of [1], tailored to
the special form of A in (14.2.3), are summarized as follows. The system
(14.2.2) is globally asymptotically stable if A is Hurwitz and one of the
following conditions is satisfied,
i) a22 < 0;
ii) a22 ≥ 0, and a11 > a12 .
The facts that a22 ≥ 0 and A is Hurwitz imply that a12 > 0.
The detailed proof of Theorem 14.2.1 is rather involved and can be

found in [35]. In what follows, we will outline the main idea of this proof.
14.2.3. Outline of the Proof
The proof can be divided into three parts. In the first part, the vector field
is studied in detail and some constants are captured to characterize the
vector field. In the second part, we show that it is these constants, rather
than the stability of the A matrix, that determine the global boundedness
of the trajectories. Clearly, the global boundedness is a necessary condi-

tion for global asymptotic stability. In the third part, we show that the
global boundedness condition plus the stability of the matrix A constitute
the necessary and sufficient condition for global asymptotic stability. The
crucial step in this part is to show the nonexistence of closed trajectories.
In view of Remark 14.2.1, we only need to consider the case where
a22 ≥ 0 and a11 ≤ a12 . In this case, the four parameters a11 , a12 , a21 and
a22 are all non-negative.
Part 1. The Vector Field
Consider the vector field of the system (14.2.2),
ẋ1 = −a11 sat(x1 ) + a12 sat(x2 ) =: f1 (x),

ẋ2 = −a21 sat(x1 ) + a22 sat(x2 ) =: f2 (x), (14.2.4)
where 0 < a11 ≤ a12 , a21 ≥ 0, a22 ≥ 0, and det (A)

= 0. Denote the slope
of the trajectory at x as
f2 (x)
η(x) := .
f1 (x)
For an initial state x0 ∈ R2 , denote the trajectory of (14.2.4) as ψ(t, x0 ).
The vector field of (14.2.4) is partitioned into 9 regions according to
the saturation function, by two vertical lines, x1 = ±1, and two horizontal
lines, x2 = ±1 (see Fig. 14.2.1).
In the central unit square,
ẋ = Ax.
In the region
U := x : |x1 | ≤ 1, x2 ≥ 1 ,
we have
ẋ1 = −a11 x1 + a12 ,

ẋ2 = −a21 x1 + a22 .
Since a11 ≤ a12 , ẋ1 ≥ 0 in this region and the trajectories go rightward.
Also note that ẋ is independent of x2 , so for all the points on a vertical
line x1 = c in this region, ẋ is the same. Because of this, if x0 ∈ U and
ψ(t, x0 ) ∈ U for all t ∈ [0, t1 ], then with ∆ > 0,
% &
0 0
ψ t, x0 + = ψ(t, x0 ) + , ∀ t ∈ [0, t1 ]. (14.2.5)
∆ ∆
U
3
−R V
2
0 −W W
−1
−2
−V R
−3
−U
−4
−4 −3 −2 −1 0 1 2 3 4
Figure 14.2.1: The partitioning of the vector field.
We call (14.2.5) the vertical shifting property in U . Specifically, let

−1
x0 = , x02 ≥ 1,
x02
be a point on the line x1 = −1 and suppose x(t) remains in the region U

before it intersects the line x1 = 1 at t = T with

1
x(T ) = ,
xT 2
then the increment of x2 from t = 0 to t = T is independent of x02 . We

denote this constant xT 2 − x02 as h2 . It can be verified that if a11 < a12 ,
then
2a21 det A a12 + a11
h2 = − 2 log . (14.2.6)
a11 a11 a12 − a11
As Proposition 14.2.1 will show, a11 = a12 automatically ensures global
asymptotic stability of the system (14.2.2) if A is Hurwitz.
In the region

V := x : x1 ≥ 1, x2 ≥ 1 ,
ẋ and the slope of the trajectories are constants. We denote the constant
slope η(x) as
−a21 + a22
α := . (14.2.7)
−a11 + a12
In the region
W := x : x1 ≥ 1, |x2 | ≤ 1 ,
we have
ẋ1 = −a11 + a12 x2 ,

ẋ2 = −a21 + a22 x2 .
In contrast with the region U , ẋ is independent of x1 . If x0 ∈ W and

ψ(t, x0 ) ∈ W for all t ∈ [0, t1 ], then, with ∆ > 0, we have
% &
∆ ∆
ψ t, x0 + = ψ(t, x0 ) + , ∀ t ∈ [0, t1 ]. (14.2.8)
0 0
We call (14.2.8) the horizontal shifting property in the region W . As Propo-

sition 14.2.1 will show, if a21 ≤ a22 , the system (14.2.2) will not be globally
asymptotically stable. Now for the case that a21 > a22 , ẋ2 < 0 and ẋ points
downward in this region. In this case, if a trajectory starts at a point

x01
x0 = , x01 ≥ 1,
1
on the line x2 = 1 and crosses the line x2 = −1 at a point

xT 1
x(T ) = , xT 1 ≥ 1,
−1
then xT 1 − x01 is a constant. We denote this constant as h1 . It can be
verified that

 2a12 det A a21 + a22

 − a + a2 log a − a , if a22 > 0,
22 22 21 22
h1 = (14.2.9)


 − 2a11 , if a22 = 0.
a21
In the region
R := x : x1 ≥ 1, x2 ≤ −1 ,
ẋ is a constant. We denote the constant slope η(x) in this region as
a21 + a22
β := . (14.2.10)
a11 + a12
The remaining four regions are symmetric to U , V , W and R. We denote

them as −U , −V , −W and −R, respectively.
We next digress to address two special cases, a11 = a12 and a21 ≤ a22 .
Proposition 14.2.1. Assume that A is Hurwitz.
a) If a11 = a12 , then the system (14.2.4) is globally asymptotically sta-

ble.
b) If a21 ≤ a22 (which implies that a11

= a12 ), then the system has
diverging trajectories and also has a closed trajectory.
This proposition can be easily proved with Green’s Theorem.

Now the two special cases are cleared, we only need to study the remain-
ing case where a11 < a12 and a21 > a22 . For this case, the four constants h1 ,
h2 , α and β are all well defined. Moreover, all the trajectories go clockwise.
Shown in Fig. 14.2.1 are some typical trajectories.
Part 2. Conditions for Global Boundedness
In this part, we consider the system

−a11 a12
ẋ = A sat(x) = sat(x), a12 > a11 > 0, a21 > a22 ≥ 0.
−a21 a22
(14.2.11)
We don’t assume that A is Hurwitz in this part since the critical case where
A has a pair of pure imaginary eigenvalues will be useful to our study. It
turns out that the system can have a bounded global attractor even if A is
unstable. The global boundedness depends on β/α, h1 and h2 rather than
the stability of A.
Proposition 14.2.2. The system (14.2.11) has a bounded global attractor

if and only if one of the following conditions is satisfied,
a) a11 a21 > a12 a22 ;
b) a11 a21 = a12 a22 and βh1 + h2 < 0.
If a11 a21 = a12 a22 and βh1 + h2 = 0, then outside certain region, all
the trajectories are closed. If a11 a21 < a12 a22 (or a11 a21 = a12 a22 and
βh1 + h2 > 0), there will be diverging trajectories and if in addition A is
Hurwitz, there exists a closed trajectory.
Proof. Since a12 > a11 > 0 and a21 > a22 ≥ 0, we have α < 0, β > 0,
|h1 |, |h2 | < ∞.
Denote % &
α
u∗ = max 0, (βh1 + h2 ) , αh1 .
β
Let
1
p1 = , uk ≥ u∗ ,
uk + 1
be a point on the line x1 = 1 (see the point labeled “1” in Fig. 14.2.2). Let
the trajectory starting from p1 be ψ(t, p1 ). We will show later that ψ(t, p1 )
will go through regions V , W , R, and −U consecutively (not fall into the
central square before leaving −U ). Let the intersections of ψ(t, p1 ) with
the lines x2 = 1, x2 = −1, x1 = 1 and x1 = −1 be, respectively,

1 + vk 1 + wk 1 −1
, , , ,
1 −1 −1 − rk −1 − uk+1
which correspond to the points “2, 3, 4, 5” in Fig. 14.2.2. Then,

1 1 β
vk = − uk , wk = − uk + h1 , rk = − uk + βh1 ,
α α α
and uk+1 = rk + h2 , i.e.,
β
uk+1 = − uk + βh1 + h2 . (14.2.12)
α
The requirement that the trajectory do not enter the central square is
equivalent to vk , wk , rk , uk+1 ≥ 0. This is guaranteed by uk ≥ u∗ . If
we also have uk+1 ≥ u∗ , then we can continue
with the
above process to
1
get an intersection with the line x1 = 1 at (see point “9” in
1 + uk+2
Fig. 14.2.2), where
β
uk+2 = − uk+1 + βh1 + h2 ,
α
and so on. Equation (14.2.12) defines a first order linear time invariant
discrete-time system.
Case 1. a11 a21 < a12 a22
This inequality is equivalent to − αβ > 1. So the discrete-time system
(14.2.12) is unstable. If uk is large enough, then uk+1 > uk , uk+2 >
uk+1 , · · · will be an exponentially increasing sequence and the trajectory
9
8
2 1
7 2
6 3
−2 4
5
−4
−6
−8 −6 −4 −2 0 2 4 6

1
starting from will be unbounded. It can also be shown with
1 + uk
Poincaré-Bendixon Theorem that there exists a closed trajectory within
some region.
Case 2. a11 a21 > a12 a22
In this case, 0 < − αβ < 1. So (14.2.12) is stable. A global attractor can be

constructed from the numbers α, β, h1 and h2 .
Case 3. a11 a21 = a12 a22
In this case, − α
β
= 1 and uk+2 = uk + 2(βh1 + h2 ). If βh1 + h2 < 0, the
sequence uk , uk+2 , · · ·, will decrease steadily before the trajectory touches
the central square. A global attractor can be constructed.
If βh1 + h2 > 0, the sequence uk , uk+2 , · · ·, will increase steadily and
the trajectory will go unbounded. And similar to Case 1, there exists a
closed trajectory.
If βh1 + h2 = 0, then for uk sufficiently large, the trajectory
% &
1
ψ t,
1 + uk
is closed. 2
Part 3. Proof of Theorem 14.2.1
In view of Remark 14.2.1 and Proposition 14.2.1, we only need to consider
the following system

−a11 a12
ẋ = A sat(x) = sat(x), a12 > a11 > 0, a21 > a22 ≥ 0.
−a21 a22
(14.2.13)
Proposition 14.2.3. Assume that A is Hurwitz, the system (14.2.13) is

globally asymptotically stable if and only if a11 a21 ≥ a12 a22 .
This proposition can be established as follows. If a11 a21 < a12 a22 , then
by Proposition 14.2.2, the system is not globally asymptotically stable.
So the necessity of the condition is obvious. What remains is to show the
sufficiency of the condition. If a11 a21 = a12 a22 and A is Hurwitz, then it can
be verified that βh1 + h2 < 0. So, by Proposition 14.2.2, if a11 a21 ≥ a12 a22 ,
then the system has a global attractor. Now that the global boundedness
of the trajectories is guaranteed, the only thing that needs to be shown is
that the system has no closed trajectory.
Since all the trajectories are kept unchanged when the vector field is
multiplied by a positive constant, we assume that a11 = 1 in the sequel for
simplicity. Now we have,

−1 a12
A= .
−a21 a22
We first deal with the case where a22 = 0.
Lemma 14.2.1. Assume that A is Hurwitz. If a22 = 0, then (14.2.13) is

globally asymptotically stable.
This lemma can also be proven with Green’s Theorem.

In what follows, we consider the case that a22 > 0. Let
a21
k= ,
a12 a22
then we can assume that

−1 a12
A= , k > 0, a22 > 0. (14.2.14)
−ka12 a22 a22
The restriction a12 > a11 > 0 and the assumption that A is Hurwitz trans-
late to a12 > 1, a22 < 1 and ka212 > 1. And the condition a11 a21 ≥ a12 a22
is equivalent to k ≥ 1. Therefore, we can establish Proposition 14.2.3 by
showing that the system

−1 a12
ẋ = A sat(x) = sat(x), a12 > 1, k ≥ 1, 0 < a22 < 1,
−ka12 a22 a22
(14.2.15)
is globally asymptotically stable. The proof will be carried out by evolving
A from the simplest form where a22 = 1, k = 1, to the case a22 = 1, k ≥ 1,
and finally to the general case 0 < a22 < 1, k ≥ 1. When a22 = 1, the system
is surely not globally asymptotically stable because A is not Hurwitz, but
the trajectories in this case will be used as a reference for showing the
convergence of the trajectories when a22 decreases.
Let’s start with a simple model (the primary model),

−1 a12
ẋ = sat(x), a12 > 1. (14.2.16)
−a12 1
Given an initial condition x0 , denote the trajectory of (14.2.16) as ψ1 (t, x0 ).

Then we have
Lemma 14.2.2. All the trajectories of (14.2.16) are closed. Each trajec-
tory is symmetric with respect to the line x1 = x2 and the line x1 = −x2 .
Now we consider the secondary model,

−1 a12
ẋ = sat(x), a12 > 1, k ≥ 1. (14.2.17)
−ka12 1
Denote the trajectory of (14.2.17) as ψ2 (t, x0 ). The following two lemmas

give a clear characterization of the trajectories of the system (14.2.17).
Lemma 14.2.3. Assume that k > 1. Let

1
x∗ = ka12
.
1
Then, ψ2 (t, x∗ ) is a closed curve that lies within the central square. Denote
the region enclosed by ψ2 (t, x∗ ) as S0 , then every point inside S0 is on a
closed trajectory. And outside S0 , every trajectory will converge to ψ2 (t, x∗ )
(see Fig. 14.2.3).
−1
−2
−3
−4
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
Figure 14.2.3: Illustration for Lemma 14.2.3.
Lemma 14.2.4. Assume that k > 1. Let

x01 1
x0 = , x01 < ,
1 ka12
be a point on the line x2 = 1, then ψ2 (t, x0 ) goes upward at first and will
return to the line x2 = 1 and then intersect with the line x2 = −1. Let

xc1
xc =
−1
be the first intersection with x2 = −1, then,
1
− < xc1 < −x01 .
ka12
Lemmas 14.2.3 and 14.2.4 give us a clear picture of the trajectories of
(14.2.17), where a22 = 1, k > 1. Lemma 14.2.4 says that if x0 is outside
of S0 , a trajectory ψ2 (t, x0 ) will move closer and closer to x∗ (or −x∗ ) as
it reaches the lines x2 = ±1. Next we will show that as a22 is decreased,
a trajectory ψ(t, x0 ) of (14.2.15) will move even closer to x∗ (or −x∗ ) as
compared with ψ2 (t, x0 ) or not intersect with the lines. This will lead to
the proof of Proposition 14.2.3. Rewrite (14.2.15) as follows,
ẋ1 = −sat(x1 ) + a12 sat(x2 ),
ẋ2 = a22 (−ka12 sat(x1 ) + sat(x2 )) , (14.2.18)
where a12 > 1, k ≥ 1 and 0 < a22 < 1. We will consider the perturba-
tion of the trajectories as a22 varies. Denote the trajectory of (14.2.18) as
ψ(t, x0 , a22 ). As compared with (14.2.17), ẋ1 is the same but ẋ2 is multi-
plied with a scalar a22 . Because of this, the trajectories of (14.2.18) exhibit
some interesting properties.
Fact 14.2.1.
a) Let

x01
x0 = , x02 ≥ 1,
x02
be a point above the line x2 = 1, then for all a22 > 0,

0 1 ψ(t, x0 , a22 ) − x02 = a22 0 1 ψ2 (t, x0 ) − x02 ,
and

1 0 ψ(t, x0 , a22 ) = 1 0 ψ2 (t, x0 ),
as long as ψ(t, x0 , a22 ) stays above the line x2 = 1.
b) Let

x01
x0 = , x01 ≥ 1,
x02
be a point to the right of the line x1 = 1, then for all a22 > 0,
1
1 0 ψ(t, x0 , a22 ) − x01 = 1 0 ψ2 (a22 t, x0 ) − x01 ,
a22
and

0 1 ψ(t, x0 , a22 ) = 0 1 ψ2 (a22 t, x0 ),
as long as ψ(t, x0 , a22 ) stays to the right of x1 = 1.
An illustration for Fact 14.2.1 is given in Fig. 14.2.4, where the ‘∗’s are
x0 , the solid curves are ψ2 (t, x0 ), and the dashed curves are ψ(t, x0 , a22 ),
a22 < 1.
With Fact 14.2.1, we are ready to present a final lemma that leads to
the proof of Proposition 14.2.3.
2.5
1.5
0.5
−0.5
−1
−1.5
−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Figure 14.2.4: Illustration for Fact 14.2.1.
Lemma 14.2.5. Let

x01 x02
x0 = , x02 ∈ (0, 1], x01 ≥ .
x02 ka12
Suppose ψ(t, x0 , a22 ) intersects the line x2 = −x02 at a point

xc1
xc =
−x02
and xc1 < x01 , then if δ ∈ (0, a22 ) is sufficiently small, ψ(t, x0 , a22 − δ) will
intersect with ψ(t, x0 , a22 ) at a point to the right of x0 . If it also intersects
the line x2 = −x02 , the intersection will be to the left of xc .
Proof of Proposition 14.2.3. The necessity of the condition simply fol-

lows from Proposition 14.2.2. With Lemma 14.2.1, it remains to show that
the system (14.2.18) (or (14.2.15)) is globally asymptotically stable. We
will first show that any point on the line x2 = 1 is not on a closed
trajec-
1
tory. We can restrict our attention to the points to the left of ka12 ,
1
since for the points to its right, they can be traced back to the left as the
trajectories go rightward above the line x2 = 1. Let

x01 1
x0 = , x01 < ,
1 ka12
then ψ2 (t, x0 ) of the system (14.2.17) (see the solid curve in Fig. 14.2.5)
will return to the line x2 = 1 at a point x0 . From Fact 14.2.1, ψ(t, x0 , a22 )
1.5
XO ’
XO
1
0.5
−0.5
−1
XC
−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
will also return to x0 for all a22 > 0 (see the dashed curve in Fig. 14.2.5).

1
ka12
We have shown in Lemma 14.2.4 that for any x0 to the left of ,
1
if ψ2 (t, x0 ) reaches the line x2 = −1 at some point

xc1
xc = ,
−1
then xc1 < −x01 , i.e., xc is to the left of −x0 . It can be shown that xc is
also to the left of x0 . From Lemma 14.2.5, we know that as a22 decreases
from 1, the intersection of ψ(t, x0 , a22 ) and x2 = −1 will move leftward,
hence remain to the left of −x0 and x0 . Note that x0 is on ψ(t, x0 , a22 ),
14.3. State Saturation – Discrete-Time Systems 343
so ψ(t, x0 , a22 ) overlaps with ψ(t, x0 , a22 ). Therefore, x0 is not on a closed
trajectory, since a closed trajectory must be symmetric.
The possibility of the existence of a closed trajectory that does not
intersect with x2 = 1 can be excluded similarly.
If k > 1, then a11 a21 > a12 a22 ; if k = 1, then a11 a21 = a12 a22 and it
can be verified that β = a22 and
βh1 + h2 = a22 (1 − a22 )h1 < 0,
(assuming a11 = 1). So by Proposition 14.2.2, all the trajectories of

(14.2.18) are bounded, and they must converge to the origin. 2
Proof of Theorem 14.2.1. Combining Remark 14.2.1, Propositions

14.2.1, 14.2.2 and 14.2.3, we can obtain the necessary and sufficient con-
dition for the system (14.2.2) to be globally asymptotically stable. Con-
dition b) in Theorem 14.2.1 is a simple combination of condition b) in
Remark 14.2.1 and Propositions 14.2.1, 14.2.2 and 14.2.3. This is justified
as follows.
Since A is Hurwitz and a22 ≥ 0, we must have a12 > 0. If a12 ≤ a11 ,
Remark 14.2.1 b) and Proposition 14.2.1 say that the system is globally
asymptotically stable. If a12 > a11 , because a11 a21 ≥ a12 a22 , we have
a21 > a22 , so the system is also globally asymptotically stable by Proposi-
tion 14.2.3.
Conversely, suppose a22 ≥ 0 but a11 a21 < a12 a22 , then we have
a12 a21
> .
a11 a22
Since A is Hurwitz,
a12 a22
> .
a11 a21
Therefore,
% &
a12 a21 a22
> max , ≥ 1,
a11 a22 a21
i.e., a12 > a11 . Hence by Propositions 14.2.1 and 14.2.2, the system is not
globally asymptotically stable whether a21 > a22 or not. And in both cases,
the system has unbounded trajectories and there is also a closed trajectory.
2
14.3. Planar Linear Systems under State Saturation –

Discrete-Time Systems
14.3.1. System Description and Problem Statement
Consider the following system,

x(k + 1) = sat(Ax(k)), x ∈ R2 . (14.3.1)
This system and its continuous-time counterpart (14.2.1) have a similar
description, but they behave quite differently. First of all, (14.2.1) operates
on the entire plane while (14.3.1) operates only on the unit square. Second,
the trajectories of (14.2.1) do not intersect with each other but the con-
nected trajectory of (14.3.1) (by connecting x(k) and x(k +1)) can intersect
with itself. Third, the limit trajectories of (14.2.1) must be periodic but a
limit trajectory of (14.3.1) need not be.
Because of its importance in neural networks and digital filters, the sta-
bility of the system (14.3.1) has been heavily studied in the literature for
a long period of time (see, e.g., [1,72–74,77,83] and the references therein).
Various sufficient conditions are available for global asymptotic stability
([74,83]). The necessary and sufficient condition, however, has remained
unknown until very recently. In [36], a complete stability analysis is per-
formed on the system (14.3.1). In particular, the necessary and sufficient
condition for the system to be globally asymptotically stable is identified
and in the process of establishing this condition, the behaviors of the tra-
jectories are examined in detail. This section will present the main result
obtained in [36] and give an outline of its proof.
14.3.2. Main Results on Global Asymptotic Stability
Given an initial state x(0) = x0 , denote the trajectory of the system (14.3.1)
that passes through x0 at k = 0 as ψ(k, x0 ). In this section, we only consider
the positive trajectories. Hence, throughout the section, k ≥ 0.
Definition 14.3.1. The equilibrium x = 0 of the system (14.3.1) is said
to be stable if, for any ε > 0, there exists a δ > 0 such that,
|ψ(k, x0 )| ≤ ε, ∀ k ≥ 0, ∀ |x0 | ≤ δ.
It is said to be globally asymptotically stable if it is a stable equilibrium
and
lim ψ(k, x0 ) = 0, ∀ x0 ∈ R2 .
k→∞
Also, it is said to be locally asymptotically stable if it is stable and
lim ψ(k, x0 ) = 0, ∀ x0 ∈ U0 ,
k→∞
for some neighborhood U0 of x = 0.

For convenience, we will often simply say that the system (14.2.1) is
stable, locally asymptotically stable or globally asymptotically stable when
it is clear from the context that the equilibrium x = 0 is referred to.
The system is globally asymptotically stable only if it is locally asymp-

totically stable (A has eigenvalues inside the unit circle). In this case, A is
said to be asymptotically stable, Schur stable, or simply, stable. We assume
that A is stable throughout this section. Denote the closed unit square as
S and its boundary as ∂S. It is easy to see that no matter where x(0) is, we
always have x(1) ∈ S. Hence, the global asymptotic stability is equivalent
to limk→∞ ψ(k, x0 ) = 0 for all x0 ∈ S. The main result of this section is
presented as follows,

and only if A is asymptotically stable and there exists no x0 ∈ ∂S and
N > 0 such that ψ(N, x0 ) = ±x0 and ψ(k, x0 ) = Ak x0 ∈ S for all k < N .
If ψ(k, x0 ) = Ak x0 ∈ S for all k < N , then ψ(N, x0 ) = sat(AN x0 ).

Hence, this theorem can be interpreted as follows. Assume that A is stable,
then the system (14.3.1) is globally asymptotically stable if and only if none
of the following statements is true,
a) There exist an N ≥ 1 and d1 , d2 ≥ 0 such that

1 1 + d1 1
A N
=± , A k
∈ S, ∀ k < N;
1 1 + d2 1
b) There exist an N ≥ 1 and d1 , d2 ≥ 0 such that

−1 −1 − d1 −1
A N
=± , A k
∈ S, ∀ k < N;
1 1 + d2 1
c) There exist an N ≥ 1, d2 > 0 and x1 ∈ (−1, 1) such that

x1 x1 x1
AN
=± , A k
∈ S, ∀ k < N ;
1 1 + d2 1
d) There exist an N ≥ 1, d1 > 0 and x2 ∈ (−1, 1) such that

1 1 + d1 1
A N
=± , A k
∈ S, ∀ k < N.
x2 x2 x2
Each of the above conditions implies that there is a simple periodic

trajectory that starts at some x0 with period N or 2N . The trajectory
stays inside S as that of the corresponding linear system for the first N − 1
steps, and when the linear trajectory goes out of S at step N , the saturation
function makes ψ(N, x0 ) = sat(AN x0 ) return exactly at x0 or −x0 . These
conditions can be verified. Since A is stable, there exists an integer N0 such
that Ak x0 ∈ S for all k > N0 and all x0 ∈ ∂S. Hence, it suffices to check
the four conditions only for N < N0 .
Conditions a) and b) are very easy to check. As to c) (or d)), for each
N , at most two x1 s can be solved from

x1 x1
A N
=± .
1 1 + d2
To see this, denote the elements of AN as (AN )ij , i, j = 1, 2. Then from

c), we have
(AN )11 x1 + (AN )12 = ±x1 . (14.3.2)
If (AN )11
= ±1, then there are two x1 ’s that satisfy (14.3.2). If (AN )11 =
±1, we must have (AN )12
= 0. Otherwise AN would have an eigenvalue
±1, which is impossible since A is stable. In this case, (14.3.2) has only
one solution. It remains to check if d2 > 0 and

x1
A k
∈ S, ∀ k < N.
1
A more efficient method to check the conditions is developed in [36].

1.5840 −1.3990
A= .
3.9702 −2.9038
The following results are presented with accuracy up to 4 decimal digits.

There are two points on ∂S that satisfy condition c), one with
(A3 )12
x1 = = 0.7308,
1 − (A3 )11
and the other with

(A2 )12
x1 = = 0.9208.
−1 − (A2 )11
But there are four periodic trajectories as listed below,

0.7308 −0.2414 −0.3791 0.7308
1) , , , ;
1.0000 −0.0023 −0.9516 1.0000

0.9028 0.0310 −0.9028 −0.0310 0.9028
2) , , , , ;
1.0000 0.6804 −1.0000 −0.6804 1.0000

0.7424 −0.2230 −0.4145 0.7424
3) , , , ;
1.0000 0.0438 −1.0000 1.0000

1.0000 0.1850 −1.0000 −0.1850 1.0000
4) , , , , .
1.0000 1.0000 −1.0000 −1.0000 1.0000
As we can see from the example, there are other kinds of periodic trajecto-
ries than what are inferred by the conditions a)-d), e.g., trajectories 3) and
4). There may also be trajectories that neither are periodic nor converge
to the origin. We will show in the proof that if none of the conditions a)-d)
is true, then there exists no non-convergent trajectory of any kind.
14.3.3. Outline of the Proof
The proof of Theorem 14.3.1 consists of three parts. In the first part,
some general properties of the limit trajectories of the system (14.3.1) are
characterized. An outstanding feature is that a nontrivial limit trajectory
can only intersect with two pairs of opposite sides of the unit square and
it cannot have intersections with both of the neighboring sides. This result
turns our attention to a simpler system which has only one saturated state,

a11 x1 (k) + a12 x2 (k)
x(k + 1) = . (14.3.3)
sat(a21 x1 (k) + a22 x2 (k))
This system will be studied in detail in Part 2. In particular, we will
establish a relation between the present intersection of a trajectory with
the lines x2 = ±1 and the next intersection. This relation helps us to
identify the condition for the system (14.3.3) to be globally asymptotically
stable, which in turn leads to our final proof of the main result in Part 3.
Part 1. Limit Trajectories
To prove that the system (14.3.1) is globally asymptotically stable, we
need to show that the only limit point of any trajectory is the origin. It
is well-known [74] that the system may have stationary points other than
the origin, and it may have periodic trajectories and even trajectories that
neither are periodic, nor go to a stationary point. In this part, we are going
to characterize some general properties of the non-convergent trajectories.
These properties will facilitate us to exclude the existence of such non-
convergent trajectories under the condition of Theorem 14.3.1.
Since every trajectory is bounded by the unit square, there exists a set
of points such that the trajectory will go arbitrarily close to them infinitely
many times.
Definition 14.3.2. For a given x0 ∈ R2 , a point x∗ ∈ R2 is called a

(positive) limit point of the trajectory ψ(k, x0 ) if there exists a subsequence
of ψ(k, x0 ), ψ(ki , x0 ), i = 1, 2, · · ·, such that
lim ψ(ki , x0 ) = x∗ .
i→∞
The set of all such limit points is called the limit set of the trajectory. We
denote this limit set as Γ(x0 ).
Since the function sat(Ax) is continuous in x, if a trajectory ψ(k, x0 )

returns arbitrarily close to x ∈ Γ(x0 ), it will also return arbitrarily close to
sat(Ax). We collect this property in the following lemma.
Lemma 14.3.1. If y0 ∈ Γ(x0 ), then ψ(k, y0 ) ∈ Γ(x0 ) for all k ≥ 0. Given

any (arbitrarily small) ε > 0 and any (arbitrarily large) integer N ≥ 1,
there exists an integer K0 > 0 such that
|ψ(k + K0 , x0 ) − ψ(k, y0 )|∞ < ε, ∀ k ≤ N.
Because of Lemma 14.3.1, for y0 ∈ Γ(x0 ), ψ(k, y0 ) is called a limit

trajectory of ψ(k, x0 ). It is periodic if and only if Γ(x0 ) has finite number
of elements.
The following notation is defined for simplicity. Denote

x1
Lh = : x1 ∈ (−1, 1) ;
1

1
Lv = : x2 ∈ (−1, 1) .
x2
We see that Lh and −Lh are the two horizontal sides of S, and Lv and −Lv
are the two vertical sides of S. Notice that they do not include the four
vertices of the unit square. Also, denote

1 −1
v1 = , v2 =
1 1
as the two upper vertices of the square.
Let y0 be a limit point of some non-convergent trajectory and for sim-
plicity, let yk = ψ(k, y0 ). Denote

Y = ± yk : k ≥ 0
and
AY = ± Ayk : k ≥ 0 .
Clearly, Y must have an intersection with the boundary of the unit square.
If Y ∩ Lh is not empty, define

x1
γ1 = inf x1 : ∈ Y ∩ (Lh ∪ {v1 , v2 }) ,
1
and
x1
γ2 = sup x1 : ∈ Y ∩ (Lh ∪ {v1 , v2 }) .
1
If Y ∩ Lv is not empty, define

1
γ3 = sup x2 : ∈ Y ∩ (Lv ∪ {v1 , −v2 }) ,
x2
and
1
γ4 = inf x2 : ∈ Y ∩ (Lv ∪ {v1 , −v2 }) .
x2
Proposition 14.3.1. Let y0 be a limit point of some trajectory.
a) If y0 ∈ Lh , then ψ(k, y0 ) will not touch Lv or −Lv for all k ≥ 0.

Moreover, ψ(k, y0 ) will stay inside the strip

x1
: |x1 | ≤ max |γ1 |, |γ2 | ;
x2
b) If y0 ∈ Lv , then ψ(k, y0 ) will not touch Lh or −Lh and will stay inside
the strip
x1
: |x2 | ≤ max |γ3 |, |γ4 | ;
x2
c) The set Y cannot include both v1 and v2 .
Proof. The proof is built up on a simple geometric fact. Let X be a set

in R2 and let AX be the image of X under the linear map x → Ax. Then
the area of AX equals to the area of X times |det (A)|.
a) We first assume that Y contains a finite number of elements, i.e.,
ψ(N, y0 ) = y0 for some N . Suppose on the contrary that the trajectory
will touch Lv or −Lv at some step. The main idea of the proof is to show
that the area of the convex hull of AY is no less than that of Y , which
contradicts the fact that |det (A)| < 1.
Since Y contains points on both Lh and Lv , γi , i = 1, 2, 3, 4, are all
defined.
If yk is in the interior of the unit square, then yk = Ayk−1 ; if yk ∈ Lh ,
then yk = sat(Ayk−1 ) and

x1
Ayk−1 = ,
1+d
for some |x1 | < 1 and d ≥ 0 (note that y0 = yN = sat(AyN −1 )); if yk ∈ Lv ,

then
1+d
Ayk−1 = ,
x2
for some |x2 | < 1 and d ≥ 0; if yk = v1 (or v2 ), then yk = sat(Ayk−1 ) and
% &
1 + d1 −1 − d1
Ayk−1 = or ,
1 + d2 1 + d2
for some d1 , d2 ≥ 0. Hence, AY contains all the elements of Y that are in

the interior of S, and for those yk on the boundary of S, if yk ∈ Lh , there
is a point in AY that is just above yk (on the same vertical line) and if
yk ∈ Lv , then there is a point in AY that is just to the right of yk (on the
same horizontal line).
Denote the areas of the convex hulls of Y and AY as A(Y ) and A(AY ),
respectively. Also, let

γ1 γ2 1 1
u1 = , u2 = , u3 = , u4 = ,
1 1 γ3 γ4
as shown in Fig. 14.3.1. In the figure, the points marked with “∗” belong
to AY , the polygon with dash-dotted boundary is the convex hull of AY
and the polygon with vertices ±ui , i = 1, 2, 3, 4, and some points in the
1.5
1
u1 u2
−u
4
0.5
−u3
0
u3
−0.5
u4
−u2 −u1
−1
−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
interior of S is the convex hull of Y . Since there is at least one point in Y

that is to the left of u1 , one to the right of u2 , one above u3 and one below
u4 , the convex hull of Y is a subset of the convex hull of AY . (This may not
be true if u1 is the leftmost point in Y , or if u2 is the rightmost). It follows
that A(AY ) ≥ A(Y ). This is a contradiction since A(AY ) = |det (A)|A(Y )
and |det (A)| < 1.
If on the contrary that Y has a point outside of the strip

x1
: |x1 | < max |γ1 |, |γ2 | ,
x2
then there will be a point in Y that is to the left of u1 (or on the same
vertical line with u1 ), and a point to the right of u2 (or on the same hori-
zontal line with u2 ). In this case, we also have A(AY ) ≥ A(Y ), which is a
contradiction.
Now we extend the result to the case that Y has infinite many elements.
Also suppose on the contrary that the trajectory will touch Lv , −Lv or go
outside of the strip at some step. By Lemma 14.3.1, for any ε > 0 and any
integer N ≥ 1, there exists a K0 > 0 such that
|ψ(k + K0 , x0 ) − ψ(k, y0 )|∞ < ε,

for all k ≤ N and in particular,
|ψ(K0 , x0 ) − y0 |∞ < ε.
So the trajectory ψ(k + K0 , x0 ), k ≥ 0, will also touch (or almost touch) Lv ,

−Lv , or go outside of the strip. Since y0 is a limit point of ψ(k + K0 , x0 ),
there exists a K1 > 0 such that
|ψ(K1 + K0 , x0 ) − y0 |∞ < ε.
Define
Z(ε) = ψ(k + K0 , x0 ) : 0 ≤ k ≤ K1
and
AZ(ε) = Aψ(k + K0 , x0 ) : 0 ≤ k ≤ K1 .
Using similar arguments as in the finite element case, we can show that
A(AZ(ε))
|det (A)| = ≥ 1 − O(ε).
A(Z(ε))
Letting ε → 0, we obtain |det (A)| ≥ 1, which is a contradiction.
b) Similar to a).
c) If, on the contrary, Y contains both v1 and v2 , then the convex hull
of Y is S. Also, AY contains a point

x1
Ayj = , x1 ≤ −1, x2 ≥ 1,
x2
and a point
x1
Ayk = , x1 ≥ 1, x2 ≥ 1,
x2
hence the convex hull of AY contains S. This also leads to A(AY ) ≥ A(Y ),
a contradiction. 2
Part 2. Systems with One Saturated State
Now we are clear from Proposition 14.3.1 that if there is any limit trajec-
tory, it can intersect only one pair of the sides of the unit square, either
(Lh , −Lh ), or (Lv , −Lv ), not both of them. Hence, we only need to investi-
gate the possibility that a limit trajectory only intersects ±Lh . The other
possibility that it only intersects ±Lv is similar. For this reason, we restrict
our attention to the following system,

a11 x1 (k) + a12 x2 (k)
x(k + 1) = := sat2 (Ax(k)). (14.3.4)
sat(a21 x1 (k) + a22 x2 (k))
Assume that the matrix

a11 a12
A=
a21 a22
is asymptotically stable. If a21 = 0 or a12 = 0, it is easy to see that both

systems (14.3.1) and (14.3.4) are globally asymptotically stable and none
of the conditions a)-d) of Theorem 14.3.1 can be satisfied. So we assume
in the following that a21 , a12
= 0.
The terms globally asymptotically stable, limit point and limit trajec-
tory for the system (14.3.1) are extended to the system (14.3.4) in a natural
way.
For a given initial state x(0) = x0 , denote the trajectory of the system
(14.3.4) as ψ2 (k, x0 ). Denote the line x2 = 1 as Leh , the line x2 = −1 as
−Leh and the region between these two lines (including ±Leh ) as Se . In this
part, we study the global asymptotic stability of the system (14.3.4) and
will also determine a subset in Leh which is free of limit points.
Proposition 14.3.2. The system (14.3.4) is globally asymptotically stable

if and only if A is stable and the following statement is not true for any
x1 ∈ R: There exist an integer N ≥ 1 and a real number d > 0 such that

x1 x1 x1
AN =± , Ak ∈ Se , ∀ k < N. (14.3.5)
1 1+d 1
Let
αs = min |x1 | : x1 satisfies (14.3.5) ,
then no limit trajectory can exist completely within the strip

x1
: |x1 | < αs .
x2
To prove Proposition 14.3.2, we need to establish the relation between

the next intersection of a trajectory with ±Leh and the present one. For x0 ∈
Leh , suppose that ψ2 (k, x0 ) will intersect ±Leh at k = ki , i = 1, 2, · · ·, with
0 < k1 < k2 < · · ·. Since the trajectory can be switched to −ψ2 (k, x0 ) at
any k without changing its convergence property, we assume for simplicity

that all the intersections ψ2 (ki , x0 ) are in Leh (If not so, just multiply it
with −1). Denote
x10 = ψ2 (k1 , x0 ), x20 = ψ2 (k2 , x0 ), ···.
We call x0 , x10 and x20 the first, the second and the third intersections,
respectively. We also call x0 and x10 the present and the next intersections.
Clearly, x10 is uniquely determined by x0 . We also see that the relation
x0 → x10 is a map from Leh to itself. To study the global asymptotic stability
of the system (14.3.4), it suffices to characterize the relation between x0 and
x10 . Through this relation, we can show that if (14.3.5) is not true for any
x1 , then for every x0 ∈ Leh , the intersections x10 , x20 , · · · will move closer and
closer toward an interval, and all the trajectories starting from this interval
will not touch the lines ±Leh and will converge to the origin.
Let x0 ∈ Leh . The next intersection of ψ2 (k, x0 ) with Leh occurs at step
k1 if

0 1 Ak1 x0 ≥ 1, 0 1 Ak x0 < 1, ∀ k < k1 .
The next intersection is x10 = ψ2 (k1 , x0 ) = sat2 (Ak1 x0 ) (or −sat2 (Ak1 x0 )).
Since for different x0 ∈ Leh , the number of steps for the trajectories to
return to ±Leh , i.e., the number k1 as defined above, is different, we see
that the relation between x0 and x10 must be discontinuous.
We will first determine an interval on Leh from which a trajectory will
not intersect ±Leh again (no x10 ) and will converge to the origin.
Since A is asymptotically stable, there exists a positive definite matrix
P such that
AT P A − P < 0.
Define the Lyapunov function as
V (x) := xT P x,
then for every x ∈ R2 , V (Ak x) < V (x) for all k > 1.

Given a real number ρ > 0, denote the Lyapunov level set as

E(P, ρ) := x ∈ R2 : xT P x ≤ ρ .
Let ρc be such that E(P, ρc ) ⊂ Se and E(P, ρc ) just touches ±Leh . In this
case, E(P, ρc ) has only one intersection with Leh . Let this intersection be

αc
pc = .
1
If x0 = pc , then the linear trajectory Ak x0 will be inside E(P, ρc ) ⊂ Se .

Hence, ψ2 (k, x0 ) = sat2 (Ak x0 ) = Ak x0 for all k > 0 and will converge to
the origin. Since
AT P A − P < 0,
there exists an interval around pc in Leh , of nonzero length, such that for
every x0 in this interval, ψ2 (k, x0 ) = Ak x0 , k ≥ 1, will never touch ±Leh
and will converge to the origin.
Here we will use a simple way to denote a line segment. Given two
points p1 , p2 ∈ R2 , denote

[p1 , p2 ] := λp1 + (1 − λ)p2 : 0 ≤ λ ≤ 1 ,
and similarly,
(p1 , p2 ] = [p1 , p2 ] \ {p1 },
[p1 , p2 ) = [p1 , p2 ] \ {p2 },
(p1 , p2 ) = [p1 , p2 ] \ {p1 , p2 }.
Define

α
α0 := min α < αc : Ak ∈ Se , ∀ k ≥ 0 ,
1
and
β
β0 := max β > αc : Ak ∈ Se , ∀ k ≥ 0 .
1
Since a21
= 0, the line ALeh := {Ax : x ∈ Leh }, has intersections with
both Leh and −Leh . So there exist points on both sides of pc which will be
mapped out of Se under A. Hence, α0 and β0 are finite numbers. Now let

α0 β0
p0 = , q0 = .
1 1
Because of the extremal nature in the definition of α0 , we must have Ak p0 ∈

±Leh for some k, otherwise α0 would not be the minimum of the set. The
same argument applies to q0 . Also from the definition, we have

Ak x : x ∈ [p0 , q0 ] ⊂ Se , ∀ k ≥ 1.
It follows that for every x0 ∈ (p0 , q0 ), the linear trajectory Ak x0 will never
touch ±Leh for all k ≥ 1. This implies that p0 and q0 are the inner most
pair of points on Leh where the discontinuity occurs on the relation between
x0 and x10 . Based on p0 and q0 , all the other discontinuous points on Leh
can be characterized with an inductive procedure (see [36]). Suppose that
the number of discontinuous points to the left of p0 is I and the number to
the right of q0 is J . Let the points pi , qj , i = 1, 2, · · · , I, j = 1, 2, · · · , J ,
be labeled such that pi is to the left of pi−1 and that qj is to the right of
qj−1 . For simplicity, denote

−∞ ∞
p∞ = , q∞ = .
1 1
Denote the second intersection of the trajectory ψ2 (k, pi ) with Leh as p1i
and that of ψ2 (k, qj ) with Leh as qj1 . The following lemma collects some
simple facts about the relations between pi , p1i , qj and qj1 .
Lemma 14.3.2.
a) p10 , q01 ∈ [p0 , q0 ];
b) For i ≥ 1, p1i ∈ (q0 , q∞ ) and for j ≥ 1, qj1 ∈ (p∞ , p0 );
c) p1i ∈ (p1i−1 , q∞ ), and qj1 ∈ (p∞ , qj−1

1
);
d) For i, j ≥ 1, p1i and qj1 cannot be both in [pi , qj ], nor both outside
of [pi , qj ], i.e., there must be one of them inside [pi , qj ] and the other
one outside of the interval.
Item c) in Lemma 14.3.2 shows that
[pi−1 , p1i−1 ] ⊂ [pi , p1i ],
and
1
[qj−1 , qj−1 ] ⊂ [qj1 , qj ].
Item d) shows that we have either
[pi , p1i ] ⊂ [qj1 , qj ]
or
[qj1 , qj ] ⊂ [pi , p1i ].
Item b) shows that all these intervals must include [p0 , q0 ]. In summary, the
facts in Lemma 14.3.2 jointly show that the intervals [pi , p1i ] and [qj1 , qj ],
i = 0, 1, · · · , I, j = 0, 1, · · · J , are ordered by inclusion. More properties

of the set of points pi , qj are revealed in [36] and a method for computing
the set of points is also provided. Based on these properties, the following
lemma can be established.
Lemma 14.3.3. Assume that the condition (14.3.5) is not true for any
x1 ∈ R. Consider x0 ∈ Leh . If x0 ∈ (p∞ , p0 ], then x10 ∈ (x0 , q∞ ) and one of
the following must be true:
a) x10 ∈ (p0 , q0 ) and there is no third intersection x20 ;
b) x10 ∈ (x0 , p0 ];
c) x10 ∈ [q0 , q∞ ) and x20 ∈ (x0 , x10 ).
Similarly, if x0 ∈ [q0 , q∞ ), then x10 ∈ (p∞ , x0 ) and one of the following must
be true.
d) x10 ∈ (p0 , q0 ) and there is no third intersection x20 ;
e) x10 ∈ [q0 , x0 );
f) x10 ∈ (p∞ , p0 ] and x20 ∈ (x10 , x0 ).
Also, if x0 ∈ (pi , p1i ) (or x0 ∈ (qj1 , qj )), then x10 , x20 and the subsequent
intersections will all be on the interval (pi , p1i ) (or (qj1 , qj )). Furthermore,
for any x0 ∈ Leh , there is a finite k1 such that ψ2 (k1 , x0 ) ∈ (p0 , q0 ). After
that, ψ2 (k, x0 ) will have no more intersection with Leh and will converge to
the origin.
This lemma simply says that if there is no x1 ∈ R that satisfies (14.3.5),

the intersections of a trajectory with the line Leh will move closer and closer
to the interval (p0 , q0 ) and finally fall on it and then the trajectory will
converge to the origin. Next, we assume that there is some x1 ∈ R that
satisfies (14.3.5) and would like to determine an interval in Leh from which
the convergence of trajectories is guaranteed.
Recall that pc is defined to be the unique intersection of the Lyapunov
level set E(P, ρc ) with the line Leh . Also, αc is the first coordinate of pc ,
i.e.,

αc = 1 0 pc .
Lemma 14.3.4.
Case 1. αc ≤ 0. If x0 ∈ [q0 , q∞ ), then x10 is to the left of x0 ;
Case 2. αc ≥ 0. If x0 ∈ (p∞ , p0 ], then x10 is to the right of x0 .
Case 1 and Case 2 correspond to two orientations of the Lyapunov level

sets. Lemma 14.3.4 can be easily proven from the orientation and the fact
that Ak x0 , k > 0, belongs to a smaller level set than x0 . Lemma 14.3.4 also
implies that, in Case 1, if x10 = x0 , then x0 ∈/ [q0 , q∞ ) and must belong to
(p∞ , p0 ].
Lemma 14.3.5. Let

αs = min |x1 | : x1 satisfies (14.3.5) .
Case 1. αc ≤ 0. Let
−αs
ps = ,
1
then p1s = ps ∈ (p∞ , p0 ). Suppose that ps ∈ [pi+1 , pi ). Then, for every
x0 ∈ (ps , p1i ], the trajectory ψ2 (k, x0 ) will converge to the origin;
Case 2. αc > 0. Let

αs
ps = ,
1
then p1s = ps ∈ (q0 , q∞ ). Suppose that ps ∈ (qj , qj+1 ]. Then for every
x0 ∈ [qj1 , ps ], the trajectory ψ2 (k, x0 ) will converge to the origin.
In both cases, no limit trajectory can be formed completely inside the strip

x1
: |x1 | < αs .
x2
Proof of Proposition 14.3.2. It follows immediately from Lemmas

14.3.3 and 14.3.5. 2
A computational method is provided in [36] for determining the point

±αs
ps =
1
and all the x1 ’s that satisfy (14.3.5).

Part 3. Proof of the Main Result
Now, we turn back to the system (14.3.1)
x(k + 1) = sat(Ax(k)). (14.3.6)
For easy reference, we restate Theorem 14.3.1 as follows.

and only if A is stable and none of the following statements is true:
a) There exists an N ≥ 1 such that
sat(AN v1 )) = ±v1 , Ak v1 ∈ S, ∀ k < N;
b) There exists an N ≥ 1 such that
sat(AN v2 )) = ±v2 , Ak v2 ∈ S, ∀ k < N;
c) There exist an x1 ∈ (−1, 1) and an N ≥ 1 such that

% &
x1 x1 x1
sat A N
=± , Ak
∈ S, ∀ k < N;
1 1 1
d) There exist an x2 ∈ (−1, 1) and an N ≥ 1 such that

% &
1 1 1
sat A N
=± , A k
∈ S, ∀ k < N.
x2 x2 x2
Proof. We will exclude the possibility of the existence of limit trajectories

(except for the trivial one at the origin) under the condition that none of
statements a)-d) in the theorem is true. In the following, when we say a
limit trajectory, we mean a nontrivial one other than the origin. Clearly,
every limit trajectory must include at least one point on the boundary of
the unit square, i.e., a point in the set ±(Lh ∪ Lv ∪ {v1 , v2 }). By Propo-
sition 14.3.1, we know that a limit trajectory cannot have points in both
±Lh and ±Lv . So we have two possibilities here, limit trajectories including
points in ±(Lh ∪ {v1 , v2 }), and those including points in ±(Lv ∪ {v1 , v2 }).
Because of the similarity, we only exclude the first possibility under the
condition that none of a)-c) is true, the second possibility can be excluded
under the condition that none of a), b) and d) is true.
For a given initial state x0 , we denote the trajectory of the system

(14.3.6) as ψ(k, x0 ) and the trajectory of (14.3.4) as ψ2 (k, x0 )
Clearly, if x1 ∈ (−1, 1) satisfies c), then this x1 also satisfies (14.3.5).
On the other hand, suppose that there is some x1 that satisfies (14.3.5).
Let ps be as defined in Lemma 14.3.5 for the system (14.3.4) (if there is no
x1 that satisfies (14.3.5), then we can assume

±∞
ps =
1
and the following argument also goes through). Note that, if there is some
x1 ∈ R, |x1 | ≤ 1, that satisfies (14.3.5), i.e.,

x1 x1 k x1
AN
=±
, 0 1 A ≤ 1, ∀ k < N,
1 1+d 1
we must also have

1 0 Ak x1 ≤ |x1 |, ∀ k < N,
1
which indicates that x1 satisfies c). Otherwise, as in the proof of Proposi-

tion 14.3.1, the area of the convex hull of the set

± x0 , ±Ax0 , · · · , ±AN −1 x0
would be less than the area of the convex hull of the set

± Ax0 , ±A2 x0 , · · · , ±AN x0 .
This would be a contradiction to the fact that |det (A)| < 1.

Hence, if no x1 satisfies c), then ps must be outside of S. By Propo-
sition 14.3.2, no limit trajectory of (14.3.4) can lie completely inside the
strip
x1
: |x1 | < αs .
x2
It follows that no limit trajectory of (14.3.6) can lie completely between
−Lv and Lv . Therefore, no limit trajectory of (14.3.6) can include only
boundary points in ±Lh . On the other hand, if a limit trajectory include
only boundary points ±v1 (or ±v2 , note that, by Proposition 14.3.1, no
limit trajectory can include both ±v1 and ±v2 ), then a) or b) must be
true, which contradicts to our assumption. In short, if there is a limit
14.4. Sensor Saturation 361
trajectory that include points in ±(Lh ∪ {v1 , v2 }), it must include at least
one point on ±Lh and one on ±v1 (or ±v2 ). Here we assume that it includes
v2 .
Let’s consider the trajectories ψ(k, v2 ) and ψ2 (k, v2 ). Suppose that
ψ(k, v2 ) has an intersection with ±Lh but does not include v1 and any
point in ±Lv , we conclude that ψ(k, v2 ) = ψ2 (k, v2 ) will converge to the
origin. The argument goes as follows.
Let k0 be the smallest k such that ψ(k, v2 ) intersects ±Lh . Denote
1
v2 = ψ(k0 , v2 ). Since b) is not true, k0 must also be the smallest k such
that

0 1 Ak v2 ≥ 1.
So, we have ψ2 (k, v2 ) = ψ(k, v2 ) for all k ≤ k0 . Here we have two cases.
Case 1. αc ≤ 0
In this case, ps is to the left of v2 . Since v21 = sat(Ak0 v2 ) = sat2 (Ak0 v2 )

goes to the right of v2 , by Lemma 14.3.4, v2 must be to the left of p0 . It
follows that v2 ∈ (ps , p1i ], where (ps , p1i ] is the interval in Lemma 14.3.5
b). Hence, ψ2 (k, v2 ) will converge to the origin. Moreover, the subsequent
intersections of ψ2 (k, v2 ) with ±Lh are between v2 and v21 . Since ψ(k, v2 )
does not touch ±Lv , we must have ψ(k, v2 ) = ψ2 (k, v2 ) and hence ψ(k, v2 )
will also converge to the origin.
Case 2. αc > 0
In this case ps is to the right of v1 . By the assumption that ψ(k, v2 ) does not
include v1 , the intersections of ψ2 (k, v2 ) with ±Lh will stay to the left of v1
(or to the right of −v1 ). Since αc > 0, by Lemma 14.3.4, the intersections
will move rightward until falling on [qj1 , ps ), where [qj1 , ps ) is the interval in
Lemma 14.3.5 c). Similar to Case 1, we have that ψ2 (k, v2 ) converges to
the origin and ψ(k, v2 ) = ψ2 (k, v2 ).
So far, we have excluded the possibility that a limit trajectory includes
any point in the set ±(Lh ∪{v1 , v2 }). The possibility that a limit trajectory
includes any point in the set ±(Lv ∪ {v1 , v2 }) can be excluded in a similar
way. Thus, there exists no limit trajectory of any kind and the system
(14.3.6) must be globally asymptotically stable. 2
14.4. Semi-Global Stabilization of Linear Systems

Subject to Sensor Saturation
14.4.1. Introduction
While actuator saturation has been addressed in much detail, only a few
results are available that deal with sensor saturation. In particular, issues
related to the observability of a linear system subject to sensor saturation
were discussed in detail in [57]. A discontinuous dead beat controller was
recently constructed for single input single output linear systems in the
presence of sensor saturation [59] that drives every initial state to the origin
in a finite time.
In this section, we consider the problem of semi-globally stabilizing lin-
ear systems using linear feedback of the saturated output measurement.
Here, by semi-global stabilization we mean the construction of a stabiliz-
ing feedback law that yields a domain of attraction that contains any a
priori given (arbitrarily large) bounded set. This problem was motivated
by its counterpart for linear systems subject to actuator saturation [65,67].
More specifically, it was established in [65,67] (see also Remark 4.4.1 of
Chapter 4) that a linear system subject to actuator saturation can be semi-
globally stabilized using linear feedback if the system is stabilizable and
detectable in the usual linear sense and all its open loop poles are in the
closed left-half plane, no matter where the invariant zeros are. What we
will show in this section is that a single input single output linear system
subject to sensor saturation can be semi-globally stabilized by linear sat-
urated output feedback if the system is stabilizable and detectable in the
usual linear sense and all its invariant zeros are in the closed left-half plane,
no matter where the open loop poles are. This result thus complements the
results of [59] in the sense that, it requires an extra condition that the
invariant zeros of the system be on the closed left-half plane to conclude
semi-global stabilizability by linear feedback. It can also be viewed as dual
to its actuator saturation counterpart in [65,67]. We, however, note that in
the dual situation [67], the condition of all poles being in the closed left-half
plane is necessary even with nonlinear feedback [90], while in the current
situation, the condition of all invariant zero being in the closed left-half
plane is not necessary with nonlinear feedback (by the result of [59]). It is
not clear at this time if it would become necessary for linear feedback.
Although this result can be viewed as dual to its actuator saturation

counterpart in [65,67], the mechanism behind the stabilizing feedback laws
is completely different. In the case of actuator saturation, we construct
low gain feedback laws that avoid the saturation of the input signal for
all initial states inside the a priori given set and the closed-loop system
behaves linearly. Here in the case of sensor saturation, the output matrix is
fixed and the output signal is always saturated for large initial states. Once
the output is saturated, no information other than its sign is available for
feedback. Our linear feedback laws are designed in such a way that they
use the saturated output to cause the system output to oscillate into the
linear region of sensor saturation function and remain there in a finite time.
The same linear feedback laws then stabilize the system at the origin. This
is possible since all the invariant zeros are in the closed left-half plane and
the feedback gains can be designed such that the overshoot of the output
is arbitrarily small.
Our presentation of this section is based on our recent work [66].
14.4.2. Main Results
Consider the following single input single output linear system subject to
sensor saturation,
ẋ = Ax + Bu, x ∈ Rn , u ∈ R,
(14.4.1)
y = sat(Cx), y ∈ R,
where sat : R → R is the standard saturation function. Our main results
on semi-global stabilizability of the system (14.4.1) is given in the following
theorem.
Theorem 14.4.1. The system (14.4.1) is semi-globally asymptotically sta-
bilizable by linear feedback of the saturated output if
• The pair (A, B) is stabilizable;
• The pair (A, C) is detectable; and
• All invariant zeros of the triple (A, B, C) are in the closed left-half
plane.
More specifically, for any a priori given bounded set X 0 ⊂ R2n , there exists
a linear dynamic output feedback law of the form
ż = F z + Gy, z ∈ Rn ,
(14.4.2)
u = Hz + H0 y,
such that the equilibrium (x, z) = (0, 0) of the closed-loop system is asymp-
totically stable with X 0 contained in its domain of attraction.
Proof. We will establish this result in two steps. In the first step, we will
construct a family of feedback laws of the form (14.4.2), parameterized in
ε ∈ (0, 1]. In the second step, we will show that, for any a priori given
bounded set X 0 ⊂ R2n , there exists an ε∗ ∈ (0, 1] such that, for each
ε ∈ (0, ε∗ ], the equilibrium (x, z) = (0, 0) of the closed-loop system is
asymptotically stable with X 0 contained in its domain of attraction.
The construction of the feedback laws follows the following algorithm.
Step 1. Find a state transformation [87],

T T
x = T x̄, x̄ = xT0 xT1 , x1 = x11 x12 · · · x1r ,
such that the system can be written in the following form,
ẋ0 = A0 x0 + B0 x11 , x0 ∈ Rn0 ,

ẋ11 = x12 ,
ẋ12 = x13 ,
.. (14.4.3)
.
ẋ1r = C0 x0 + a1 x11 + a2 x12 + · · · + ar x1r + u,
y = sat(x11 ),
where (A0 , B0 ) is stabilizable and the eigenvalues of A0 are the in-

variant zeros of the triple (A, B, C) and hence are all in the closed
left-half plane.
We note that the multiple input multiple output counterpart of the
above canonical form will in general also require a transformation on
the input and the output. The latter cannot be performed due to the
presence of sensor saturation.
Step 2. Let F0 (ε) be such that

λ(A0 + B0 F0 (ε)) = − ε + λ0 (A0 ) ∪ λ− (A0 ), ε ∈ (0, 1],
where λ0 (A0 ) and λ− (A0 ) denote respectively the sets of eigenvalues

of A0 that are on the imaginary axis and that are in the open left-half
plane. It is clear that A0 + B0 F0 (ε) is Hurwitz for any ε ∈ (0, 1] and
|F0 (ε)| ≤ α0 ε, ∀ ε ∈ (0, 1], (14.4.4)

for some α0 independent of ε.

Such an F0 (ε) exists since (A0 , B0 ) is stabilizable. We summarize
some properties for the triple (A0 , B0 , F0 (ε)) from [65, Lemmas 2.2.3
and 2.2.4 and Theorem 3.3.1].
Lemma 14.4.1. For the given triple (A0 , B0 , F0 (ε)), there exists a
nonsingular matrix T0 (ε) ∈ Rn0 ×n0 such that
|T0 (ε)| ≤ τ0 , (14.4.5)

|F0 (ε)T0−1 (ε)| ≤ β0 ε, (14.4.6)
|F0 (ε)A0 T0−1 (ε)| ≤ β1 ε, (14.4.7)
T0 (ε)(A0 + B0 F0 (ε))T0−1 (ε) = J0 (ε), (14.4.8)
where τ0 , β0 and β1 are some constants independent of ε and J0 (ε) ∈

Rn0 ×n0 is a real matrix. Moreover, there exists a P0 > 0, independent
of ε, such that,
ε
J0T (ε)P0 + P0 J0 (ε) ≤ − I. (14.4.9)
2
Step 3. Let L be such that A + LC is Hurwitz. Such an L exists since

the pair (A, C) is detectable.
Step 4. Construct the family of output feedback laws as follows,
ż = Az + Bu + L(Cz − y),
r
α1 α2 αr
u = −C0 z0 − ai z1i − r (y − F0 (ε)z0 ) − r−1 z12 − · · · − z1r ,
i=1
ε ε ε
where z0 and z1i , i = 1, 2, · · · , r, are defined as follows,

T T
z = T z̄, z̄ = z0T z1T , z1 = z11 z12 · · · z1r ,
and αi ’s are chosen such that
sr + αr sr−1 + αr−1 sr−2 + · · · + α2 s + α1 = (s + 1)r ,
i.e.,
r!
αi = Cri−1 = , i = 1, 2, · · · , r.
(i − 1)!(r − i + 1)!
We now proceed with the second step of the proof: to show that, for any
a priori given bounded set X 0 ⊂ R2n , there exists an ε∗ ∈ (0, 1] such that,
for each ε ∈ (0, ε∗ ], the equilibrium (x, z) = (0, 0) of the closed-loop system
is asymptotically stable with X 0 contained in its domain of attraction.
Without loss of generality, let us assume that the system is already in the
form of (14.4.3), i.e., T = I. Letting e = x − z, we can write the closed-loop
system as follows,
ẋ0 = A0 x0 + B0 x11 ,
ẋ11 = x12 ,
ẋ12 = x13 ,
..
.
ẋ1r = C0 (x0 − z0 ) + a1 (x11 − z11 ) + a2 (x12 − z12 ) + · · ·
α1 α2 αr
+ar (x1r − z1r ) − r (y − F0 (ε)z0 ) − r−1 z12 − · · · − z1r ,
ε ε ε
ż = Az + L(Cz − y) + B −C0 z0 − a1 z11 − · · · − ar z1r
α1 α2 αr
− r (y − F0 (ε)z0 ) − r−1 z12 − · · · − z1r ,
ε ε ε
y = sat(x11 ).
We next define a new set of state variables as follows,
x̃0 = T0 (ε)x0 ,
x̃11 = x11 − F0 (ε)x0 ,
1 2
x̃1i = εi−1 x1i + Ci−1 εi−2 x1i−1 + Ci−1 εi−3 x1i−2 + · · · + Ci−1
i−2
εx12
i−1
+Ci−1 (x11 − F0 (ε)x0 ), i = 2, 3, · · · , r,
e0 = x0 − z0 ,
e1i = x1i − z1i , i = 1, 2, · · · , r,
(14.4.10)
and denote
T
e = eT0 e11 e12 · · · e1r .
With these new state variables, the closed-loop system can be written
as follows,
x̃˙ 0 = J0 (ε)x̃0 + T0 (ε)B0 x̃11 ,

1 1
x̃˙ 11 = − x̃11 + x̃12
ε ε
− F0 (ε)A0 T0−1 (ε) + F0 (ε)B0 F0 (ε)T0−1 (ε) x̃0 − F0 (ε)B0 x̃11 ,
1 1
x̃˙ 12 = − x̃12 + x̃13
ε ε
− F0 (ε)A0 T0−1 (ε) + F0 (ε)B0 F0 (ε)T0−1 (ε) x̃0 − F0 (ε)B0 x̃11 ,
..
.
1 1
x̃˙ 1r−1 = − x̃1r−1 + x̃1r
ε ε
−[F0 (ε)A0 T0−1 (ε) + F0 (ε)B0 F0 (ε)T0−1 (ε)]x̃0 − F0 (ε)B0 x̃11 ,
1 1 1
x̃˙ 1r = − x̃1r + [x11 − sat(x11 )] − F0 (ε)e0
ε ε ε
+εr−1 [C0 e0 + a1 e11 + a2 e12 + · · · + ar e1r ]
+α2 e12 + α3 εe13 + · · · + αr εr−2 e1r
−[F0 (ε)A0 T0−1 (ε) + F0 (ε)B0 F0 (ε)T0−1 (ε)]x̃0 − F0 (ε)B0 x̃11 ,
ė = (A + LC)e − L[x11 − sat(x11 )].
(14.4.11)
Choose a Lyapunov function candidate as follows,

r
√ T
V (x̃0 , x̃11 , · · · , x̃1r , e) = ν x̃T0 P0 x̃0 + x̃21i + εe P e, (14.4.12)
i=1
where ν ∈ (0, 1], independent of ε, is a constant whose value is to be

determined later, P0 is as defined in Lemma 14.4.1, and P > 0 is such that
(A + LC)T P + P (A + LC) = −I. (14.4.13)
Let c > 0, independent of ε, be such that,
c≥ sup V (x̃0 , x̃11 , · · · , x̃1r , e). (14.4.14)

(x,z)∈X 0 ,ε∈(0,1],ν∈(0,1]
Such a c exists due to the boundedness of X 0 and the definition of the state
variables as given by (14.4.10). With this choice of c, it is obvious that
(x, z) ∈ X 0 implies that

(x̃0 , x̃11 , · · · , x̃1r ) ∈ LV (c) := (x̃0 , x̃11 , · · · , x̃1r , e) ∈ R2n : V ≤ c .
Using Lemma 14.4.1, we can calculate the derivative of V inside the

level set LV (c) along the trajectories of the closed-loop system (14.4.11) as
follows,
r
2 2
V̇ = −ν x̃0 x̃0 + 2ν x̃0 P0 T0 (ε)B0 x̃11 +
T T
− x̃21i + x̃1i x̃1i+1
i=1
ε ε

−2x̃1i [F0 (ε)A0 T0−1 (ε) + F0 (ε)B0 F0 (ε)T0−1 (ε)]x̃0 − 2x̃1i F0 (ε)B0 x̃11

1 1
+2x1r [x11 − sat(x11 )] − F0 (ε)e0
ε ε
+εr−1 [C0 e0 + a1 e11 + · · · + ar e1r ]

√
+α2 e12 + α3 εe13 + · · · + αr εr−2 e1r − εeT e
√
−2 εeT P L[x11 − sat(x11 )]
≤ −ν x̃T0 x̃0 + 2δ01 ν|x̃0 ||x̃11 |
r
2 2 2
+ − x̃1i + x̃1i x̃1i+1 + 2δi0 ε|x̃1i ||x̃0 | + 2δi1 ε|x̃1i ||x̃11 |
i=1
ε ε
2 √
+ |x1r ||x11 − sat(x11 )| + 2η1 |x̃1r ||e| − εeT e
ε √
+2η2 ε|e||x11 − sat(x11 )|, (14.4.15)
where δij ’s and ηi ’s are some constants, independent of ε.

We will continue our evaluation of V̇ by considering two separated cases,
|x11 | ≤ 1 and |x11 | > 1.
Case 1. |x11 | ≤ 1
In this case, we have,
1
V̇ ≤ −ν x̃T0 x̃0 + 2δ01 ν|x̃0 ||x̃11 − x̃11 x̃12 · · · x̃1r−1 x̃1r
 ε  
2 −1 · · · 0 0 x̃11
 −1 2 ··· 0 0   
   x̃12 
 .. .
.. . . . . .   . 
× . .. ..   .. 
  
 0 0 ··· 2 −1   x̃1r−1 
0 0 · · · −1 2 x̃1r
r
√
+ [2δi0 ε|x̃1i ||x̃0 | + 2δi1 ε|x̃1i ||x̃11 |] + 2η1 |x̃1r ||e| − εeT e
i=1

r
2 2
≤− ν− δi0 ε − δ01 ν |x̃0 |2
i=1
r

δ1
− − δ01 − δ10 − 2δ11 ε − δi1 ε x̃211
ε i=2
r−1
√
δ1 2 δ1 2η12 2 ε 2
− − δi0 − δi1 ε |x̃1i | − − δr1 ε − √ x̃1r − |e| ,
i=2
ε ε ε 2
where we have used the fact that the matrix

 
2 −1 · · · 0 0
 −1 2 · · · 0 0 
 
 .. .. . . .. .. 
 . . . . . 
 
 0 0 ··· 2 −1 
0 0 · · · −1 2
is positive definite with its maximum eigenvalue denoted as δ1 > 0.
Let ν be such that
1
ν≤
2δ01
and let ε∗1 ∈ (0, 1] be such that the following hold for all ε ∈ (0, ε∗1 ],
r
ν
ν− δi0 ε2 − δ01 ν 2 ≥ ,
i=1
4
2
δ1 δ1
− δ01 − δ10 − 2δ11 ε − δi1 ε ≥
ε i=2
2ε
δ1 δ1
− δi0 − δi1 ε ≥ , i = 2, 3, · · · , r − 1,
ε 2ε
δ1 2η 2 δ1
− δr1 ε − √ 1 ≥ .
ε ε 2ε
With these choices of ν and ε∗1 , we conclude that, for any |x11 | ≤ 1,
√
δ1
r
ν 2 2 ε 2
V̇ ≤ − |x̃0 | − |x̃1i | − |e| , ε ∈ (0, ε∗1 ].
4 2ε i=1 2
Case 2. |x11 | > 1

In this case, we have,
1 2 √
V̇ ≤ −ν x̃T0 x̃0 + 2δ01 ν|x̃0 ||x̃11 | − x̃11 − (|x11 | − 1)2 − εeT e
ε

r

+ (δi0 + δi1 )εx̃21i + δi0 ε|x̃0 |2 + δi1 εx̃211 + η1 x̃21r
i=1
√ √
+ η1 + η2 ε |e|2 + η2 ε(|x11 | − 1)2 .
Now let ε∗2 ∈ (0, 1] be such that, for all ε ∈ (0, ε∗2 ],
(x̃0 , x̃11 , · · · , x̃1r , e) ∈ LV (c)
implies that,
1
|F0 (ε)x0 | ≤ ,
2
and

r

2δ01 ν|x̃0 ||x̃11 | + (δi0 + δi1 )εx̃21i + δi0 ε|x̃0 |2 + δi1 εx̃211
i=1
√ √ 1
+η1 x̃21r + η1 + η2 ε |e|2 + η2 ε(|x11 | − 1)2 ≤ .
8ε
The first inequality is due to (14.4.4) and implies that,
1
x̃211 − (|x11 | − 1)2 ≥ .
4
With this choice of ε∗2 , we have that, for any |x11 | > 1,
√ T 1
V̇ ≤ −ν x̃T0 x̃0 − εe e − , ε ∈ (0, ε∗2 ].
8ε
Combining Cases 1 and 2, we conclude that, for any ε ∈ (0, ε∗ ] with

ε = min{ε∗1 , ε∗2 },
∗
V̇ < 0, ∀ (x̃0 , x̃11 , x̃12 , · · · , x̃1e , e) ∈ LV (c) \ {0},
which, in turn, shows that the equilibrium (x, z) = (0, 0) of the closed-
loop system is asymptotically stable with X 0 contained in its domain of
attraction. 2
14.4.3. An Example
In this section, we will use a simple example to demonstrate the behavior

of the closed-loop system under the family of feedback laws that we just
constructed. Consider the system (14.4.1) with
     T
0 1 0 0 0 0
 0 0 1 0   0   1 
A=   
 0 0 0 1 , B =  0 , C =  0  .
 
1 0 0 0 1 0
It can be easily verified that this system is controllable and observable with
an invariant zero at s = 0. The open loop poles are located at {−1, ±j, 1}.
Following the design algorithm proposed above, we construct a family of
parameterized output feedback laws as follows,
ż1 = z2 − 2(z2 − y),

ż2 = z3 − 4(z2 − y),
ż3 = z4 − 6(z2 − y),
ż4 = z1 − 4(z2 − y) + u,
1 3 3
u = −z1 − 3 (y + εz1 ) − 2 z3 − z4 .
ε ε ε
Some simulation results are shown in Figs. 14.4.1 to 14.4.4. In the
simulation, initial conditions are taken randomly as
   
−4.5625 34.5569
 −18.8880   
x(0) =   , z(0) =  15.4136  .
 8.2065   −25.4760 
−4.0685 15.8338
In Figs. 14.4.1 and 14.4.2, ε is chosen to be ε = 0.1. It is clear that with this
choice of ε, the initial conditions are not inside the domain of attraction.
In Figs. 14.4.3 and 14.4.4, ε is chosen to be ε = 0.001. We see that, the
output is out of saturation after some time and the closed-loop system
become linear and all its states converge to zero. This demonstrates that
as ε decreases, the domain of attraction is enlarged.
14.5. Conclusions
In this chapter, we have presented a few results on linear systems subject
to state or sensor saturation. These results only serve as an indication
that saturation occurs in system components other than actuators. In fact,
saturation nonlinearity is ubiquitous in engineering systems and remains as
an exciting topic for research.
200
100
−100
states x and z
−200
−300
−400
−500
−600
−700
0 0.5 1 1.5 2 2.5 3 3.5 4
time t
Figure 14.4.1: ε = 0.1: the states.
0.8
0.6
0.4
0.2
output y
−0.2
−0.4
−0.6
−0.8
−1
0 0.5 1 1.5 2 2.5 3 3.5 4

time t
Figure 14.4.2: ε = 0.1: the saturated output.

300
250
200
states x and z 150
100
50
−50
−100
0 0.5 1 1.5 2 2.5 3 3.5 4
time t
Figure 14.4.3: ε = 0.001: the states.
0.5
0
output y
−0.5
−1
0 0.5 1 1.5 2 2.5 3 3.5 4

time t
Figure 14.4.4: ε = 0.001: the saturated output.

Bibliography
[1] F. Albertini and D. D’Alessandro, “Asymptotic stability of

continuous-time systems with saturation nonlinearities,” Systems &
Control Letters, Vol. 29, pp. 175-180, 1996.
[2] J. Alvarez, R. Suarez and J. Alvarez, “Planar linear systems with sin-
gle saturated feedback,” Systems & Control Letters, Vol. 20, pp. 319-
326, 1993.
[3] B. R. Barmish and W. E. Schmitendorf, “A necessary and sufficient

condition for local constrained controllability of a linear system,”
IEEE Transactions on Automatic Control, Vol. 25, pp. 97-100, 1980.
[4] D. S. Bernstein and A. N. Michel, “A chronological bibliography on

saturating actuators,” International Journal of Robust and Nonlin-
ear Control, Vol. 5, pp. 375-380, 1995.
[5] G. Bitsoris, “On the positive invariance of polyhedral sets for

discrete-time systems,” Systems & Control Letters, Vol. 11, pp. 243-
248, 1988.
[6] F. Blanchini, “Feedback control for LTI systems with state and con-
trol bounds in the presence of disturbances,” IEEE Transactions on
Automatic Control, Vol. 35, pp. 1231-1243, 1990.
[7] F. Blanchini, “Ultimate boundedness control for uncertain discrete-

time systems via set-induced Lyapunov function,” IEEE Transac-
tions on Automatic Control, Vol. 39, pp. 428-433, 1994.
[8] F. Blanchini, “Set invariance in control – a survey,” Automatica,

Vol. 35, pp. 1747-1767, 1999.
375
376 Bibliography
[9] S. Boyd, L. El Ghaoui, E. Feron and V. Balakrishnan, Linear Matrix

Inequalities in Systems and Control Theory, SIAM, Philadelphia,
1994.
[10] P. Bruck, “A new result in time optimal control of discrete sys-

tems,” IEEE Transactions on Automatic Control, Vol. 19, pp. 597-
598, 1974.
[11] P. J. Campo and M. Morari, “Robust control of processes subject

to saturation nonlinearities,” Computers and Chemical Engineering,
Vol. 14, pp. 343-358, 1990.
[12] P. J. Campo, M. Morari and C. N. Nett, “Multivariable anti-windup

and bumpless transfer: a general theory,” Proceedings of American
Control Conference, pp. 1706-1711, 1989.
[13] A. Casavola and E. Mosca, “Global regulation of null-controllable

input-saturated systems with arbitrary l2 state disturbance,”
preprint.
[14] H. Choi, “On the stabilization of linear discrete-time systems subject

to input saturation,” Systems & Control Letters, Vol. 36, pp. 241-
244, 1999.
[15] J. Collado, R. Lozano and A. Ailon, “Semi-global stabilization

of discrete-time systems with bounded inputs using periodic con-
troller,” Systems & Control Letters, Vol. 36, pp. 267-275, 1999.
[16] E. J. Davison and K. C. Cowan, “A computational method for de-

termining the stability region of second-order non-linear autonomous
system,” International Journal of Control, Vol. 9, pp. 349-357, 1969.
[17] E. J. Davison and E. M. Kurak, “A computational method for deter-

mining quadratic Lyapunov functions for non-linear systems,” Au-
tomatica, Vol. 7, pp. 627-636, 1971.
[18] C. A. Desoer and J. Wing, “An optimal strategy for a saturat-

ing sampled data system,” IRE Transactions on Automatic Control,
Vol. 6, pp. 5-15, 1961.
[19] M. E. Evans and D. N. P. Murthy, “Controllability of discrete time

systems with positive controls,” IEEE Transactions on Automatic
Control, Vol. 22, pp. 943-945, 1977.
Bibliography 377
[20] A. Feuer and M. Heymann, “Admissible sets in linear feedback

systems with bounded controls,” International Journal of Control,
Vol. 23, pp. 381-392, 1976.
[21] M. E. Fisher and J. E. Gayek, “Estimating reachable sets for two-

dimensional linear discrete systems,” Journal of Optimization The-
ory and Applications, Vol. 56, pp. 67-88, 1987.
[22] B. A. Francis, “The linear multivariable regulator problem,” SIAM

Journal on Control and Optimization, Vol. 15, pp. 486-505, 1975.
[23] G. F. Franklin, J. D. Powell and A. Emami-Naeini, Feedback Control

of Dynamic Systems, third edition, Addison Wesley, New York, 1994.
[24] A. T. Fuller, “In-the-large stability of relay and saturating control

systems with linear controller,” International Journal of Control,
Vol. 10, pp. 457-480, 1969.
[25] E. G. Gilbert and K. T. Tan, “Linear systems with state and con-
trol constraints: the theory and application of maximal output ad-
missible sets,” IEEE Transactions on Automatic Control, Vol. 36,
pp. 1008-1020, 1991.
[26] P.-O. Gutman and M. Cwikel, “Admissible sets and feedback con-
trol for discrete-time linear dynamical systems with bounded control
and dynamics,” IEEE Transactions on Automatic Control, Vol. 31,
pp. 373-376, 1986.
[27] P.-O. Gutman and P. Hagander, “A new design of constrained con-

trollers for linear systems,” IEEE Transactions on Automatic Con-
trol, Vol. 30, pp. 22-33, 1985.
[28] M. Hamza and M. E. Rasmy, “A simple method for determining

the reachable set for linear discrete systems,” IEEE Transactions on
[29] O. Hájek, Control Theory in the Plane, Springer-Verlag, Berlin, 1991.
[30] R. Hanus, M. Kinnaert and J. L. Henrotte, “Conditioning technique,

a general anti-windup and bumpless transfer method,” Automatica,
Vol. 23, pp. 729-739, 1987.
378 Bibliography
[31] M. Hautus, “Linear matrix equations with applications to the

regulator problem,” in Outils and Modèles Mathématique pour
l’Automatique, I. D. Landou, Ed., Paris: C.N.R.S., pp. 399-412,
1983.
[32] H. Hindi and S. Boyd, “Analysis of linear systems with saturating us-
ing convex optimization,” Proceedings of the 37th IEEE Conference
on Decision & Control, pp. 903-908, 1998.
[33] L. M. Hocking, Optimal Control, An Introduction to the Theory and

Applications, Oxford University Press, Oxford, 1991.
[34] L. Hou and A. N. Michel, “Asymptotic stability of systems with

saturation constraints,” IEEE Transactions on Automatic Control,
Vol. 43, pp. 1148-1154, 1998.
[35] T. Hu and Z. Lin, “A complete stability analysis of planar linear

systems under saturation,” IEEE Transactions on Circuits and Sys-
tems - I: Fundamental Theory and Applications, Vol. 47, pp. 498-512,
2000.
[36] T. Hu and Z. Lin, “A complete stability analysis of planar linear

discrete-time systems under saturation,” IEEE Transactions on Cir-
cuits and Systems I: Fundamental Theory and Applications, to ap-
pear.
[37] T. Hu and Z. Lin, “Output regulation for general linear systems with
saturating actuators,” submitted to Automatica, 1999.
[38] T. Hu and Z. Lin, “On enlarging the basin of attraction for linear sys-
tems under saturated linear feedback,” Systems & Control Letters,
Vol. 40, pp. 59-69, 2000.
[39] T. Hu and Z. Lin, “Robust stabilization of exponentially unstable

linear systems with saturating actuators,” International Journal of
Robust and Nonlinear Control, to appear. Also in Proceedings of
American Control Conference, pp. 3196-3200, 1999.
[40] T. Hu, Z. Lin and B. M. Chen, “An analysis and design method
for linear systems subject to actuator saturation and disturbance,”
submitted to Automatica, 1999. Also in Proceedings of American
Control Conference, pp. 725-729, 2000.
Bibliography 379
[41] T. Hu, Z. Lin and L. Qiu, “An explicit description of the null control-
lable regions of linear systems with saturating actuators,” submitted
for publication, 2000.
[42] T. Hu, Z. Lin and L. Qiu, “Stabilization of exponentially unstable

linear systems with saturating actuators,” IEEE Transaction on Au-
tomatic Control, to appear.
[43] T. Hu, Z. Lin and Y. Shamash, “Semi-global stabilization with guar-

anteed regional performance of linear systems subject to actuator
saturation,” Systems & Control Letters, to appear. Also in Proceed-
ings of American Control Conference, pp. 4388-4392, 2000.
[44] T. Hu, D. E. Miller and L. Qiu, “Controllable regions of LTI discrete-

time systems with input saturation,” Proceedings of the 37th IEEE
Confernece on Decision and Control, pp. 371-376, 1998.
[45] T. Hu, D. E. Miller and L. Qiu, “Null controllability and stabilizabil-

ity of LTI discrete-time systems with input saturation,” submitted
for publication, 1999.
[46] T. Hu, A. N. Pitsillides and Z. Lin, “Null controllability and sta-

bilization of linear systems subject to asymmetric actuator satura-
tion,” in Actuator Saturation Control, edited by V. Kapila and K.M.
Grigoriadis, Marcel Dekker, New York, to appear.
[47] T. Hu, L. Qiu and Z. Lin, “The controllability and stabilization of

unstable LTI systems with input saturation,” Proceedings of the 36th
IEEE Conference on Decision and Control, pp. 4498-4503, 1997.
[48] A. Isidori and C. I. Byrnes, “Output regulation for nonlinear sys-

tems,” IEEE Transactions on Automatic Control, Vol. 35, pp. 131-
140, 1990.
[49] L. Jin, P. N. Nikiforuk and M. M. Gupta, “Absolute stability con-

ditions for discrete-time recurrent neural networks,” IEEE Transac-
tions on Neural Networks, Vol. 5, pp. 954-964, 1994.
[50] S. M. Joshi, “Stability of multiloop LQ regulators with nonlineari-

ties - part I: region of attraction,” IEEE Transactions on Automatic
Control, Vol. 31, pp. 364-367, 1986.
380 Bibliography
[51] J. Junkins, J. Valasek and D. Ward, Report of ONR UCAV Modeling

Effort, Department of Aerospace Engineering, Texas A&M Univer-
sity, 1999.
[52] R. E. Kalman, “Optimal nonlinear control of saturating systems by

intermittent action,” IRE Wescon Convention Record, Pt. 4, pp. 130-
135, 1957.
[53] P. Kapasouris, M. Athans and G. Stein, “Design of feedback control

systems for stable plants with saturating actuators,” Proceedings of
the 27th IEEE Conference on Decision and Control, pp. 469–479,
1988.
[54] S. S. Keerthi and E. G. Gilbert, “Computation of minimum-time

feedback control laws for discrete-time systems with state-control
constraints,” IEEE Transactions on Automatic Control, Vol. 32,
pp. 432-435, 1987.
[55] H. K. Khalil, Nonlinear Systsems, second edition, Prentice Hall, Up-

per Saddle River, 1996.
[56] P. P. Khargonekar, K. Poolla and A. Tannenbaum, “Robust control

of linear time invariant plants using periodic compensation,” IEEE
Transactions on Automatic Control, Vol. 30, pp. 1088-1096, 1985.
[57] R. B. Koplon, M. L. J. Hautus and E. D. Sontag, “Observability of

linear systems with saturated outputs,” Linear Algebra and Appli-
cations, Vols. 205-206, pp. 909-936, 1994.
[58] R. L. Kosut, “Design of linear systems with saturating linear con-

trol and bounded states,” IEEE Transactions on Automatic Control,
Vol. 28, pp. 121-124.
[59] G. Kreisselmeier, “Stabilization of linear systems in the presence

of output measurement saturation,” Systems & Control Letters,
Vol. 29, pp. 27-30, 1996.
[60] N. J. Krikelis and S. K. Barkas, “Design of tracking systems subject

to actuator saturation and integrator windup,” International Journal
of Control, Vol. 39, pp. 667-682, 1984.
Bibliography 381
[61] J. B. Lasserre, “On reachable and controllable sets for two-

dimensional linear discrete-time systems,” Journal of Optimization
Theory and Applications, Vol. 70, pp. 583-595, 1991.
[62] J. B. Lasserre, “Reachable, controllable sets and stabilizing control

of constrained linear systems,” Automatica, Vol. 29, pp. 531-536,
1993.
[63] E. B. Lee and L. Markus, Foundations of Optimal Control, John

Wiley and Sons Inc., New York, 1967.
[64] J. N. Lin, “Determination of reachable set for a linear discrete sys-

tem,” IEEE Transactions on Automatatic Control, Vol. 15, pp. 339-
342, 1970.
[65] Z. Lin, Low Gain Feedback, Lecture Notes in Control and Information
Sciences, Vol. 240, Springer-Verlag, London, 1998.
[66] Z. Lin and T. Hu, “Semi-global stabilization of linear system subject

to output saturation,” Systems & Control Letters, to appear.
[67] Z. Lin and A. Saberi, “Semi-global exponential stabilization of linear

systems subject to ‘input saturation’ via linear feedbacks,” Systems
& Control Letters, Vol. 21, pp. 225-239, 1993.
[68] Z. Lin and A. Saberi, “Semi-global exponential stabilization of lin-

ear discrete-time systems subject to input saturation via linear feed-
backs,” Systems & Control Letters, Vol. 24, pp. 125-132, 1995.
[69] Z. Lin and A. Saberi, “A semi-global low-and-high design for linear

systems with input saturation - stabilization and disturbance rejec-
tion,” International Journal of Robust and Nonlinear Control, Vol. 5,
pp. 381-398, 1995.
[70] Z. Lin and A. Saberi, “Semi-global exponential stabilization of linear

discrete-time systems subject to ‘input saturation’ via linear feed-
backs,” Systems & Control Letters, Vol. 24, pp. 125-132, 1995.
[71] Z. Lin, A. A. Stoorvogel and A. Saberi, “Output regulation for linear

systems subject to input saturation,” Automatica, Vol. 32, pp. 29-47,
1996.
382 Bibliography
[72] D. Liu and A.N. Michel, “Asymptotic stability of systems operating

on a closed hypercube,” Systems & Control Letters, vol. 19, pp. 281-
285, 1992.
[73] D. Liu and A.N. Michel, “Sparsely interconnected neural networks

for associative memories with applications to cellular neural net-
works,” IEEE Transactions on Circuits and Systems - II: Analog
and Digital Signal Processing, Vol. 41, pp. 295-307, 1994.
[74] D. Liu and A.N. Michel, Dynamical Systems with Saturation Nonlin-
earities, Lecture Notes in Control and Information Sciences, Vol. 195,
Springer-Verlag, London, 1994.
[75] K. A. Loparo and G. L. Blankenship, “Estimating the domain of

attraction of of nonlinear feedback systems,” IEEE Transactions on
[76] J. Macki and M. Strauss, Introduction to Optimal Control, Springer-

Verlag, Berlin, 1982.
[77] R. Mantri, A. Saberi and V. Venkatasubramanian, “Stability analy-

sis of continuous time planar systems with state saturation nonlinear-
ity,” IEEE Transactions on Circuits and Systems - I: Fundamental
Theory and Applications, Vol. 45, pp. 989-993, 1998.
[78] A. Megretski, “L2 BIBO output feedback stabilization with satu-

rated control,” 13th IFAC World Congress, Vol. D, pp. 435-440,
1996.
[79] A. Nagata, S. Kodama and S. Kumagai, “Time optimal discrete

control system with bounded state variable,” IEEE Transactions on
[80] L. Pandolfi, “Linear control systems: controllability with constrained

controls,” Journal of Optimization Theory and Applications, Vol. 19,
pp. 577-585, 1976.
[81] C. Pittet, S. Tarbouriech and C. Burgat, “Stability regions for lin-

ear systems with saturaing controls via circle and Popov criteria,”
Proceedings of the 36th IEEE Conference on Decision and Control,
pp. 4518-4523, 1997.
Bibliography 383
[82] X. Qian and J. Song, Engineering Cybernetics, in Chinese, Science

Academics, Beijing, 1980.
[83] J. H. F. Ritzerfeld, “A condition for the overflow stability of second-

order digital filters that is satisfied by all scaled state-space struc-
tures using saturation,” IEEE Transactions on Circuits and Systems,
Vol. 36, pp. 1049-1057, 1989.
[84] B. G. Romanchuk, “Computing regions for attraction with poly-

topes: planar case,” Automatica, Vol. 32, pp. 1727-1732, 1996.
[85] A. Saberi, Z. Lin and A. R. Teel, “Control of linear systems with

saturating actuators,” IEEE Transactions on Automatic Control,
Vol. 41, pp. 368-378, 1996.
[86] A. Saberi, P. Sannuti and B. M. Chen, H2 -Optimal Control, Prentice

Hall, London, 1995.
[87] P. Sannuti and A. Saberi, “A special coordinate basis of multivariable

linear systems–Finite and infinite zero structure, squaring down and
decoupling,” International Journal of Control, Vol. 45, pp. 1655-
1704, 1987.
[88] R. De Santis and A. Isidori, “Output regulation for linear systems

with anti-stable eigenvalues in the presence of input saturation,”
Proceedings of the 38th IEEE Conference on Decision and Control,
pp. 2106-2111, 1999.
[89] W. E. Schmitendorf and B. R. Barmish, “Null controllability of linear

systems with constrained controls,” SIAM Journal on Control and
Optimization, Vol. 18, pp. 327-345, 1980.
[90] E. D. Sontag, “An algebraic approach to bounded controllability of

linear systems,” International Journal of Control, Vol. 39, pp. 181-
188, 1984.
[91] E. D. Sontag, Mathematical Control Theory, second edition,

Springer, New York, 1998.
[92] E. D. Sontag, and H. J. Sussmann, “Complete controllability of

continuous-time recurrent neural networks,” Systems & Control Let-
ters, Vol. 30, pp. 197-183, 1997.
384 Bibliography
[93] E. D. Sontag and H.J. Sussmann, “Nonlinear output feedback design

for linear systems with saturating controls,” Proceedings of the 29th
IEEE Conference Decision and Control, pp. 3414-3416, 1990.
[94] R. Suarez, J. Alvarez-Ramirez and J. Solis-Daun, “Linear sys-

tems with bounded inputs: global stabilization with eigenvalue
placement,” International Journal of Robust and Nonlinear Control,
Vol. 7, pp. 835-845, 1997.
[95] H. J. Sussmann, E. D. Sontag and Y. Yang, “A general result on

the stabilization of linear systems using bounded controls,” IEEE
Transactions on Automatic Control, Vol. 39, pp. 2411-2425, 1994.
[96] H. J. Sussmann and Y. Yang, “On the stabilizability of multiple

integrators by means of bounded feedback controls,” Proceedings of
the 30th IEEE Conference on Decision and Control, pp. 70-72, 1991.
[97] M. Sznaier, “A set induced norm approach to the robust control of

constrained systems,” SIAM Journal on Control and Optimization,
Vol. 31, pp. 733-746, 1993.
[98] A. R. Teel, “Global stabilization and restricted tracking for multi-

ple integrators with bounded controls,” Systems & Control Letters,
Vol. 18, pp. 165-171, 1992.
[99] A. R. Teel, “Linear systems with input nonlinearities: global stabi-

lization by scheduling a family of H∞ -type controllers,” International
Journal of Robust and Nonlinear Control, Vol. 5, pp. 399-441, 1995.
[100] A. R. Teel, “Semi-global stabilization of linear null controllable sys-

tems with input nonlinearities,” IEEE Transactions on Automatic
Control, Vol. 40, pp. 96-100, 1995.
[101] A. R. Teel, Feedback Stabilization: Nonlinear Solutions to Inher-

ently Nonlinear Problems, Ph.D dissertation, University of Califor-
nia, Berkeley, 1992.
[102] A. R. Teel, “A nonlinear small gain theorem for the analysis of con-
trol systems,” IEEE Transactions on Automatic Control, Vol. 42,
pp. 1256-1270, 1996
Bibliography 385
[103] R. P. V. Til and W. E. Schmitendorf, “Constrained controllability

of discrete-time systems,” International Journal of Control, Vol. 43,
pp. 941-956, 1986.
[104] A. Vanelli and M. Vidyasagar, “Maximal Lyapunov functions and

domain of attraction for autonomous nonlinear systems,” Automat-
ica, Vol. 21, pp. 69-80, 1985.
[105] S. Weissenberger, “Application of results from the absolute stability

to the computation of finite stability domains,” IEEE Transactions
on Automatic Control, Vol. 13, pp. 124-125, 1968.
[106] J. Wing and C. A. Desoer, “The multiple-input minimal time reg-

ulator problem (general theory),” IEEE Transactions on Automatic
Control, Vol. 8, pp. 125-136, 1963.
[107] G. F. Wredenhagen and P. R. Belanger, “Piecewise-linear LQ control

for systems with input constraints,” Automatica, Vol. 30, pp. 403-
416, 1994.
[108] J. C. Willems, “Least squares stationary optimal control and the al-
gebraic Riccati equations,” IEEE Transactions on Automatic Con-
trol, Vol. 16, pp. 621-634, 1971.
[109] W. M. Wonham, Linear Multivariable Control: A Geometric Ap-

proach, Springer-Verlag, New York, 1979.
[110] Y. Yang, E. D. Sontag and H. J. Sussmann, “Global stabilization

of linear discrete-time systems with bounded feedback,” Systems &
Control Letters, Vol. 30, pp. 273-281, 1997.
386 Bibliography
Index
actuator saturation, 1 definition, 164, 174

admissible control, 6, 13, 38, 57, 270, example, 168
308 under saturated high gain
algebraic Riccati equation, 6, 67, 75, feedback, 244
127 contractive invariant
ANCBC, 6, 56, 79, 111, 129 example, 177
anti-stable, 7 convergence rate, 229, 231
anti-windup, 2 maximal, 233, 258
ARE, 6, 67, 75, 127, 141 maximal, with disturbance,
associate memory, 327 255
asymptotically null controllable of a linear system, 230
definition, 34, 51 of a nonlinear system, 231
asymptotically null controllable overall, 247
region, 34, 51, 75 sub-optimal, 243
definition, 34, 52 convex hull, 9, 160
description, 34, 52 convex set, 8
asymptotically null controllable strictly, 8
with bounded controls
(see: ANCBC), 6 DARE, 6, 108
asymptotically stable, 7 detectability, 268, 307, 363
auxiliary feedback matrix, 165 digital filter, 328, 344
disturbance rejection, 211
bang-bang control, 21, 24, 34, 43–45, example, 219
232, 234, 242, 247, 254, 257 LMI formulation, 216, 225
bounded global attractor, 328 problem, 214, 223
with guaranteed domain of
chattering, 231, 242, 285 attraction, 214, 216, 223, 225
circle criterion, 164 with guaranteed domain of at-
closed trajectory, 331, 334, 338 traction, LMI formulation,
contractive invariance 218, 227
387
388 Index
domain of attraction, 57, 73, 74, 164 example, 104

continuity, 124 Green’s Theorem, 123, 334, 337
convexity, 65
definition, 57, 163, 174 Hausdorff distance, 8, 67, 250
enlargement, 183, 185 high gain feedback, 72, 85
enlargement, example, Hurwitz, 7
189, 190
enlargement, index theory, 62
LMI formulation, 184, integrator windup, 1
185, 187 internal stability, 267, 307
estimation, 157, 164, 169, 179 invariant ellipsoid, 234
estimation, example, 173, 177 contractive, 234
expansion, 197 contractive, condition, 164–166,
expansion, example, 207 175, 176, 259
expansion, theorem, 197 contractive, definition, 164,
monotonicity, 120, 127 174
of a set, definition, 129, 213, 222 contractive, example, 168,
177
ellipsoid, 6 contractive, necessary and
volume of, 159 sufficient condition, 167, 177,
equilibrium, 58 238, 260
equivalent transformation, 329 contractive, under saturated high
exogenous system (or exosystem), 265, gain feedback, 244
267, 306 maximal, 235
exponentially unstable, 7 with disturbance, 255, 256
extremal control, 44, 45 with disturbance, strictly, 255
definition, 19 invariant set, 196
description, 23, 28, 29, 32, 46 definition, 164, 174
equivalence, 19 with disturbance, 136
in finite steps, definition, 41
minimal representative, 20 Jordan block, 7
property, 21 Jordan canonical form, 23
extremal point, 8
extremal trajectory, 15 Lagrange multiplier method, 170, 238,
definition, 19 261
level set, 7
global stabilization lifted system, 186
at set of equilibria, theorem, 89 lifting technique, 186, 199
Index 389
limit cycle, 59 example, 26, 29, 30, 32

limit set, 348 formula, 25, 28
limit trajectory, 347, 348 general description, 16
properties, 348, 349 trajectory description, 20, 24, 29,
linear matrix equation, 274, 280, 311, 32
316 null controllable region, extremes, 43
regulator equations, 269, 272, 307, example, 50
309 formula, 47, 48
linear region of saturation, 9, 163 trajectory description, 48, 49
LMI, 6, 158, 166, 216, 219, 225, 228,
252 observability, 290, 324
LQR, 6, 67, 208 observer, 290, 324
estimation error, 290
maximal convergence control, 233, 258,
fast observer, 290
263
output regulation, 265, 305
example, 246
definition, 267, 307
with disturbance, 255
output regulation, linear theory, 267,
maximal invariant ellipsoid, 235
306
minimum energy regulation, 67
assumptions, 268, 307
negative definite, 6 error feedback, 267, 306
negative semi-definite, 6 error feedback, problem, 268
neural network, 328, 344 solutions, 269, 307
neutrally stable, 7, 230, 271, 308 state feedback, 267, 306
norm state feedback, problem, 268
of matrix, 4 output regulation, with saturating ac-
of signal, 4 tuator, 270, 307
of vector, 4 assumptions, 268, 271, 272, 307,
null controllability, 11, 13, 37, 39 308
null controllable error feedback design, 290,
in a finite time, 13, 39 324
null controllable region, 11, 13, 39, error feedback law, 295
219, 220 error feedback, problem, 271
asymptotically, 34, 51, 212 example, 297
at a finite time, 13, 33, 39 state feedback design, 279,
general description, 15, 41 315
separation result, 13, 39 state feedback law, 284, 320
null controllable region, boundary, 15 state feedback, problem, 270, 308
390 Index
performance degradation, 1 vector, 3

periodic trajectory, 62, 346 Schur complements, 171
persistent disturbance, 211 Schur stable, 7
piecewise-linear control(PLC), 80 semi-global practical stabilization
PLC , 6
example, 144
Poincaré-Bendixon Theorem, 62, 336
problem, 114
positive definite, 6
theorem, 115
positive limit set, 59
positive semi-definite, 6 semi-global stabilization, 67, 74, 195,
practical stabilization, 113, 114 199, 205
semi-global, 113, 114 example, 73, 79, 107, 207
semi-global, theorem, 115 theorem, 67, 77, 110, 204, 207
with regional performance,
quadratic Lyapunov function, 164 195
with regional performance,
reachable, 15, 41
example, 207
in a finite time, 15, 40
semi-globalization, 199, 205
reachable region, 15, 16, 41
at a finite time, 15, 40 semi-stable, 7
recurrent neural network, 327 sensor saturation, 327, 362
reference signal, 265, 267, 306 example, 370
regulatable region, 271, 309 semi-global stabilizability,
asymptotic, definition, 273, theorem, 363
310 semi-global stabilization,
definition, 272, 309 example, 370
description, theorem, 274, semi-global stabilization,
311 feedback laws, 365
reset windup, 2 set
shape reference, 159, 214,
saturated linear feedback, 57
223
high gain, 67, 72, 242
size of, 159
saturation
of actuator, 1, 57 set invariance, 234
of sensor, 327, 362 contractive, 234
of state, 327–329, 344 contractive, condition, 164–166,
saturation function, 3 175, 176, 259, 260
coupled, 263 contractive, definition, 164,
scalar, 3 174
Index 391
contractive, necessary and local, 55

sufficient condition, 167, 177, semi-global, 55, 105, 108,
238 199, 205
definition, 164, 174 semi-global, example, 207
set invariance, with disturbance, 213, semi-global, theorem, 67, 77, 110,
255 204, 207
analysis, 222 state saturation, 327–329, 344
analysis, problem, 214 bounded global attractor,
condition, 215, 223 334
definition, 222 global asymptotic stability,
enlargement, LMI formula- theorem, 330
tion, 215, 224 global boundedness,
enlargement, problem, 214, condition, 334
222 globally asymptotic stability, the-
problem, 222 orem, 345
property, 256 periodic trajectories,
strictly, 213, 255 example, 347
strictly, definition, 222 with one saturated state,
shape reference set, 159, 184, 214, 347, 352
223 state transition map, 57, 59
stability strictly convex, 8, 16, 33
anti-stable, 7 supremum, 184
asymptotically stable, 7 switching control law, 81, 110,
exponentially unstable, 7 143, 197
globally asymptotically stable, 328,switching surface, 231, 242
344
locally asymptotically stable, 329, time optimal control, 1, 12, 229
345 time-reversed system, 15, 40, 58, 86
neutrally stable, 7, 230, 271, 308 tracking error, 267, 306, 324
of a matrix, 7 TRANS3 aircraft, 144, 297
of a system, 7 longitudinal dynamics, 144,
semi-stable, 7 297
stable, 7 transmission polynomial, 270
stabilizability, 8, 268, 307, 363
stabilization, 55 vector field
global, at set of equilibria, 86 state saturation, horizontal
global, theorem, 89 shifting property, 333
392 Index
state saturation, vertical

shifting property, 332
with disturbance, 132
with disturbance, properties, 132,
133
with state saturation, 331
with state saturation, partition,
331
vertex, 41
View publication stats

Sbook PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sbook PDF

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Control systems with actuator saturation: analysis and design

Book · January 2001

The user has requested enhancement of the downloaded file.

To be published by Birkhäuser, Boston

To Jian, Tony, and Vivian

2 Null Controllability – Continuous-Time Systems 11

3 Null Controllability – Discrete-Time Systems 37

4 Stabilization on Null Controllable Region –

5 Stabilization on Null Controllable Region –

6 Practical Stabilization on Null Controllable Region 113

7 Estimation of the Domain of Attraction under

7.4 Continuous-Time Systems under State Feedback . . . . . . 163

8 On Enlarging the Domain of Attraction 183

9 Semi-Global Stabilization with Guaranteed

10 Disturbance Rejection with Stability 211

10.3 Discrete-Time Systems . . . . . . . . . . . . . . . . . . . . . 221

11 On Maximizing the Convergence Rate 229

12 Output Regulation – Continuous-Time Systems 265

13 Output Regulation – Discrete-Time Systems 305

13.4 State Feedback Controllers . . . . . . . . . . . . . . . . . . 315

14 Linear Systems with Non-Actuator Saturation 327

Saturation nonlinearities are ubiquitous in engineering systems. In control

the performance is retained while stability is improved or the other way

Chapter 9 develops a design method for arriving at simple nonlinear

in linear systems and multivariable control. Some background in nonlinear

Charlottesville, Virginia Tingshu Hu

1.1. Linear Systems with Actuator Saturation

• for a general, not necessarily ANCBC, linear system with saturating

• how to design feedback laws that work on the entire asymptotically

Throughout this book, we will assume that the saturation function is

With a slight abuse of notation and for simplicity, for a vector u ∈ Rm , we

While assuming unity of saturation level is without loss of generality, the

1.2. Notation, Acronyms, and Terminology

R := the set of real numbers,

Rn×n := the set of n × n real matrices,

C := the null controllable region,

|x| := the Euclidean norm, or 2-norm, of x ∈ Rn ,

or the l∞ -norm of a discrete-time signal x(k),

αX := {αx : x ∈ X } for a positive number α and a set X ,

E(P, ρ) := the ellipsoid {x ∈ Rn : xT P x ≤ ρ},

E c := the set of extremal controls,

E c (K) := the set of extremal controls on [0, K] (discrete-time),

sign(x) := the sign function of x ∈ R,

ANCBC := asymptotically null controllable with bounded controls,

2 := the end of a proof.

We will also use the following terminology.

• For a symmetric matrix P ∈ Rn×n , we use P > 0 and P ≥ 0 to denote

• For two symmetric matrices P1 , P2 ∈ Rn×n , P1 > P2 means that

• A continuous-time control signal u is said to be admissible if

Similarly, a discrete-time control signal u is said to be admissible if

• For x ∈ Rn and y ∈ Rm , we sometimes, by an abuse of notation,

• For a positive deﬁnite matrix P ∈ Rn×n and a positive real number

• A continuous-time linear system,

– asymptotically stable (or simply stable) if λ(A) ⊂ C− ,

• A discrete-time linear system,

x(k + 1) = Ax(k) + Bu(k),

– asymptotically stable (or simply stable) if λ(A) ⊂ C ,

• By stabilizability, unless otherwise speciﬁed, we mean asymptotic sta-

• Let X 1 and X 2 be two bounded subsets of Rn . Their Hausdorﬀ

Here the vector norm used is arbitrary.

• A set X is convex if, for any two points x1 , x2 ∈ X ,

• For a convex set X ⊂ Rn , a point x0 ∈ X is said to be an extremal