Documente Academic
Documente Profesional
Documente Cultură
CONTROL
DESIGN:
A POLYNOMIAL
APPROACH
ROBUST
CONTROL
DESIGN:
A POLYNOMIAL
APPROACH
Theodore E. Djaferis
Electrical and Computer Engineering
University0'Massachusetts
Amherst, Massachusetts
...
"
SPRINGER SCIENCE+BUSINESS MEDIA, LLC
ISBN 978-0-7923-9617-8 ISBN 978-1-4615-2293-5 (eBook)
DOI 10.1007/978-1-4615-2293-5
LIST OF FIGURES xi
LIST OF TABLES xv
PREFACE xvii
1 INTRODUCTION 1
1.1 The Control Problem 1
1.2 Book Outline 7
2 SYSTEM DYNAMICS 15
2.1 System Representations 15
2.2 Feedback Configurations 18
2.3 Stability 21
2.4 Stability of Interconnected Systems 24
2.5 'V-Stability 26
2.6 Performance 28
3 STABILITY TESTS 35
3.1 Polynomial Stability 35
3.2 The Routh-Hurwitz Stability Criterion 36
3.3 The Nyquist Stability Theorem 38
3.4 The Finite Nyquist Theorem 40
REFERENCES 253
INDEX 259
LIST OF FIGURES
Chapter 1
1.1 n-link Robotic Manipulator 2
Chapter 2
2.1 Single Revolute Link Robotic Manipulator 16
2.2 Unity Feedback Configuration 19
2.3 Feedback Configuration with Sensor 19
2.4 General Feedback Configuration 20
2.5 Poles of Second Order System 27
2.6 Desired Pole Locations for Second Order System 27
2.7 Unity Feedback Configuration 29
2.8 Basic Feedback Configuration 32
Chapter 3
3.1 (a) poles and zeros of f, (b) image of r under f 39
3.2 A Closed Jordan Curve with Convex Interior 40
3.3 A Closed Jordan Curve with Convex Interior 42
3.4 A Large Semi-circular Contour (R ~ 1) 45
Chapter 4
4.1 a) multiplicative, b) additive, c) divisive 54
4.2 Unity Feedback Configuration, Parametric Uncertainty 55
4.3 Multiplicative Perturbation Unity Feedback 56
4.4 Performance as Robust Stability 59
4.5 Robust Performance, Parametric Uncertainty 59
4.6 Feedback System Affected by Measurement Noise 62
4.7 Robust Performance as Robust Stability 63
xu ROBUST CONTROL DESIGN: A Polynomial Approach
Chapter 5
5.1 Parameter Overbound 102
5.2 Rectangular Overbound of a 6 Convex Parpolygon 103
Chapter 6
6.1 Two Value Sets 113
6.2 Two Value Sets 114
6.3 IP(jw,a)1 vs. w for a = (al,a2) E {4,6} X {3,5} 117
6.4 Unity Feedback Configuration 118
6.5 A Plot of 4>( il. 75, EztO ll ) 118
Chapter T
7.1 Unity Feedback Configuration, Parametric Uncertainty 125
7.2 A Mass-Spring-Mass System 128
7.3 Plot of the Value Set 4>(-0.7471 + j3.3, 011) 134
7.4 The Region 1) 135
7.5 Closed Loop Poles For Nominal Design 136
7.6 Closed Loop Poles of Parameter Space Method Controller 140
7.7 Closed Loop Poles of FIT Synthesis Controller 140
7.8 Bode plots of second order weighting functions 143
7.9 Unity Feedback Configuration, Parametric Uncertainty 147
7.10 Value set and Convex Hull 150
7.11 The Value Set 4>(jl.9281,0:IT) 152
7.12 The Value Set 4>(j2, O~SPIT) 154
Chapter 8
8.1 Robust Performance, Parametric Uncertainty 161
8.2 Unit Disc, Square and Octagon Overbounds 164
8.3 Bode plot of the cost function TII1"1 172
8.4 Bode plot of the function IW1SI + IW2 TI 172
List of Figures X111
Chapter 9
9.1 Feedback System with Tracking and Noise Requirements 176
9.2 Robust Tracking as Robust Stability 176
9.3 Noise Attenuation ai Robust Stability 177
9.4 Performance as Robust Stability 178
9.5 Seeker Stabilization Loop 185
9.6 Step Response up to .02 seconds 192
9.7 Step Response up to .5 seconds 192
9.8 Magnitude of Transfer Function from Tel to 8 193
9.9 Magnitude of the Transfer Function We to WA 193
Chapter 10
10.1 Unity Feedback Configuration, Parametric Uncertainty 196
10.2 Feedback System for Finite Plant Family 199
10.3 Unity Feedback Norm Bounded Multiplicative Uncertainty 200
10.4 Unity Feedback Norm Bounded Divisive Uncertainty 203
10.5 Unity Feedback Norm Bounded Additive/Divisive Uncertainty 208
10.6 Feedback Loop in Block Diagonal Bounded Perturbation Form 208
10.7 Feedback System with Tracking and Noise Requirements 218
10.8 Square Overbound of a Disc 219
Chapter 11
11.1 Two Link Planar Manipulator with Prismatic Joints 228
11.2 Unity Feedback Configuration 230
11.3 General Feedback Configuration 230
11.4 Feedback System with Tracking and Noise Requirements 236
11.5 RLC Circuit 240
LIST OF TABLES
Chapter 3
3.1 Frequencies and Sectors Proving Stability 47
Chapter 6
6.1 Frequencies and Sectors Proving Robust Stability 113
6.2 Frequencies and Sectors Proving Robust Stability 114
Chapter 7
7.1 Controller Synthesis Data 130
7.2 Controller Synthesis Data 137
7.3 si
Values of 5 ) 139
Chapter 9
9.1 Seeker System Parameter Values 184
PREFACE
To a large extent, our lives on this earth depend on systems that operate auto-
matically. Many such systems can be found in nature and others are man made.
These systems can be biological, electrical, mechanical, chemical, or ecological,
to name just a few categories. Our human body is full of systems whose conti-
nued automatic operation is vital for our existence. On a daily basis we come
in contact with man made systems whose automatic operation ensures increa-
sed productivity, promotes economic development and improves the quality of
life. A primary component that is responsible for the automatic operation of
a system is a device or mechanism called the controller. In man made systems
one must first design and then implement such a controller either as a piece of
hardware or as software code in a computer. The safe and efficient automatic
operation of such systems is testimony to the success of control theorists and
practitioners over the years. This book presents new methods {or controller
design.
Frequently, linear time invariant system models are generated that involve para-
meter uncertainty. During the 1980's there was an explosion of robust analysis
results for systems with real parameter uncertainty. The hope was that these
XVlli ROBUST CONTROL DESIGN: A Polynomial Approach
would lead to methods for robust controller synthesis as was the case with
~/ Hoc and systems with norm bounded uncertainty.
My emphasis in this book is robust control design for linear time invariant sy-
stem models with parameter uncertainty. Control system design is a very broad
and iterative process involving many steps but my focus here is the design (or
synthesis) ofthe controller. Throughout the book I use the words "design" and
"synthesis" interchangeably. My intention is not to summarize and compare
all the results that have appeared on the subject, but to present a polynomial
approach that I believe does lead to powerful new methods for robust control
design. My emphasis will be on single input single output systems but in the
last chapter I lay the foundation for multivariable design. The most promising
techniques are those derived from the Finite Nyquist Theorem (FNT) and its
corollary the Finite Inclusions Theorem (FIT). These results reduce the robust
control synthesis problem to an iterative process where at each iteration one
solves a finite number of linear inequalitiu in the controller parameters. The
formulation facilitates the development of methods for robust V-stabilization
and robust multiobjective performance synthesis and can lead to less conserva-
tive designs. Furthermore, it allows greater flexibility in the choice of controller
order. Insightful and computationally efficient robust control algorithms can
be developed based on these results and then implemented in software, thus au-
tomating the design process. Some of the proposed synthesis algorithms have
been programmed in the MATLAB environment and used in the solution of
some of the examples. However, no resources were available in order to pro-
duce a "professional" version of this software package and none is included with
the book. contrary to my desires.
Most of the results presented in this book have appeared in journal articles
and conference proceedings over the last few years. I felt that collecting this
material and presenting it in book form gives a clearer and more compelling
picture. My hope is to convince the reader that these techniques are powerful
tools for robust control synthesis and to stimulate further development.
Acknowledgements
There are several individuals without whose help the book would not be a
reality. My sincere gratitude first goes to my wife, Mary Djaferis, for her love,
devotion and continuous encouragement. She is a real blessing in my life!
Preface XIX
The research would not have been possible without the support of my students.
I gratefully acknowledge the efforts of Richard D. Kaminsky, who spearhea-
ded the development of the Finite Nyquis.t Theorem and the Finite Inclusions
Theorem. My professional association with him has been very productive and
enjoyable. Many thanks go to Wojciech Krawiec and Daniel P. Boudreau for
their valued efforts in working out examples which demonstrate that the theory
works. They were the ones who generated the first "experimental" version of
software code in MATLAB that implements some of the synthesis algorithms.
I also would like to take this opportunity to thank some other colleagues who
have had an impact on my work. I first would like to thank Christopher V.
Hollot who over the last decade provided support and encouragement. Ii was
a pleasure to cooperate with him on a number of projects including some that
are reflected in this book. Our discussions on technical matters as well as other
issues of general interest have been very rewarding. I wish to thank Bob Bar-
mish for his sincere help throughout the years. I have benefited a great deal
from our professional association and look forward to our future interactions. I
also wish to thank Doug Loose for his influence in broadening my perspective
on control system design.
Theodore E. Djaferis
1
INTRODUCTION
actuator n-link
reference_ +
link errors commands y
Robotic
positions + Controller
- Manipulator
..
actual link positions
The controller, plays a key role in the efficient automatic operation of such a
feedback control system. At the conceptual level one can think of this controller
as a function that takes the error signals (which frequently are "voltages"), and
maps them to appropriate actuator command signals (also "voltages"). Alter-
natively one can also think of such a controller as a dynamic system where the
"inputs" are the error signals and the "outputs" are the actuator commands. A
crucial question arises: How does one construct a controller that guarantees an
efficient task execution? This is perhaps the most fundamental question that
Control Theory seeks to answer. Another related question might be, given a
controller how does it affect the operation of the system? The first question is
concerned with 8ynthe8U or de8ign, whereas the second with analy8i8.
In general, these two are very challenging questions and there are ample reasons
why this is so. To begin with, the dynamic behavior of most physical systems
is nonlinear. It is almost impossible to accurately express by mathematical
relationships the system dynamics. Even if one could, these expressions would
Introduction 3
be extremely complicated. This is certainly true for the 6-link robotic mani-
pulator. In many cases there are system components that may not be known
precisely. Specifically, the description may involve parameters whose values are
only known to lie in some interval. Furthermore, the controller will rely on
measurements provided by the sensors, which will inevitably be corrupted by
"noise".
For "special classes" of systems one can provide very concrete answers to the
questions of controller analysis and design. One such important class, are
systems whose dynamic behavior can be described by a set of linear differen-
tial equations with constant, coefficients (i.e. Linear Time Invariant (LTI),
Systems). For this class of systems many very elegant solutions have been pro-
vided throughout the years. Chronologically, Cla88ical Control methods were
developed first and they were followed by Multivariable Control techniques.
New contributions to both methodologies continue to be made.
In Cla88ical Control the focus is on systems which are single input, single output
(SISO). System descriptions are either in the "time domain" in terms of "state
equations", or the "frequency domain" in terms of input-output "transfer func-
tions". The concept of .ttability is defined which captures important qualitative
dynamic features. Specifically, a very natural definition of stability requires
that if a system is excited by a "bounded" input then its output must also be
"bounded". Systems that do not exhibit this desirable dynamic behavior are
labeled un.ttable. It is true that many physical systems are inherently unstable
and require some form of compensation in order to be "stabilized". The use of
"feedback" has been shown to be a very successful compensation technique. Not
only is one interested in stability but in measures of relative .ttability. In other
words, one would like to know how far off is a stable system from being made
unstable. For a given system model parameters such as the "gain margin" and
"phase margin" provide partial information for this purpose. A system with
"large" gain and phase margin possesses some measure of rob'lJ.8tneu in that
the system would remain stable even if the model was perturbed. Fundamental
results such as the Nyquist Theorem are used to test for stability and the Ny-
quist plot answers questions about relative stability. The concept of stability
captures only some of the important dynamic characteristics. Others relate to
quic/cneu and accuracy of the response to excitation, the ability to reject di.ttur-
bance8, track reference command.t and attenuate noi.te effect.t. It is known that
the "sensitivity" and "complementary sensitivity" transfer functions of feedb-
ack systems have a strong influence on performance and that specific types of
robust stability ensure certain performance characteristics. The primary Classi-
4 CHAPTER 1
We are fortunate that even though the vast majority of physical systems have
distinct nonlinear dynamic characteristics, they nonetheless can be very fre-
quently modelled (at specific operating points) by LTI dynamic equations. This
allows for the development of insight about the operation of the physical sy-
stems and provides the basis for controller design, using the formidable LTI
tools generated over the last three decades. Very frequently, these LTI designs
suitably modified after simulation and testing on the nonlinear system itself,
provide very effective means of control. This approach is a very successful me-
thod of controller analysis and design for a great variety of dynamic systems,
including n--link robotic manipulators.
In particular, consider the case when the family of plants is given by the transfer
function:
(1.1)
where P(s) is some "nominal" system transfer function, W(s) is some given
minimum phase stable transfer function and t::.. is any arbitrary stable and pro-
per transfer function with IA(jw)1 $ 1, for all w E [0,00). This is "norm
bounded" uncertainty and is referred to as a multiplicative perturbation. This
type of uncertainty can be used to describe a system model which is obtained
from input-output experiments on the actual system or to take into account
"unmodelled dynamics". Many analysis and synthesis results exist for feedback
systems with such plant descriptions. Not only is the theory highly developed,
but software routines in MATLAB are readily available for implementing the
required procedures.
A transfer function or state space description can include entries that are "pa-
rameters". It could be that a number of physical parameters (SUCh as masses,
inertias, friction coefficients etc.) are not known exactly, but rather have values
that lie in known intervals. This form of uncertainty is "parametric". A family
of transfer functions can then be given as:
P(s, a) (1.2)
where np(s, a), d,(s, a) are the numerator and denominator polynomials which
have coefficients that depend on the real parameter vector a, which takes va-
lues in some hypercube Oil' Parameter Space Methods have been developed for
such systems in the 1960's and 1970's which focus on the characteristic poly-
nomial. However, these methods do not provide complete answers for analysis
and synthesis problems and lack the mathematical rigor, theordical develop-
ment and power of H oo • The 1980's brought a flurry of research activity in
the area of analysis of systems with parameter uncertainty and existing lite-
rature contains many important results. Unfortunately, the effort was not as
successful in providing insight into the synthesis problem and the development
of robust design methods. It is true, that norm bounded perturbations can be
used to overbound plants with parameter uncertainty. Once this is done then
H oo techniques can be employed. However, the resulting solutions will invaria-
6 CHAPTER 1
We believe that some recent results have made such an approach possible. The
primary purpose of this book is to provide a comprehensive exposition of this
new approach to robust analysis and synthesis of systems with parameter un-
certainty. We present the underlying theory, and show how this formulation
can be used to solve a number of important control problems. We demonstrate
the resulting analysis and design methods on several examples. The approach
exploits the properties of the closed loop characteristic polynomial family and
its "value set" in the complex plane. The primary vehicle used for the inve-
stigation of robust stability is the Finite Incltuion6 Theorem (FIT), which is
derived from the Nyquist Theorem. In some sense, the strength of any ana-
lysis tool can be measured by the amount of insight it provides for controller
synthesis. A prime example of this is the Nyquist Theorem, which has been
the springboard for a multitude of design methods, including some of the ones
presented here. We will demonstrate that not only is FIT useful in robustness
analysis, but that it also provides an appropriate setting for the development
of powerful robust design tools for systems with parameter uncertainty.
and redundancy. As important as these issues are, in this book we will just be
interested in controller design.
The importance of the characteristic polynomial and its roots in the dynamic
behavior of a feedback system are very well known. There is extensive Control
literature on the subject that goes back at least half a century. The Parameter
Space Methods developed in the 1960's and 1970's are prime examples of work
in this direction. The study of root locations of polynomials, has a tradition
in the Mathematics literature of almost two centuries. It is a fact, that most
of the initial results were concerned with only a single polynomial rather than
a family of polynomials. In addition, early results dealt mainly with analysis,
as the issue of "design" is only of major concern to Control theorists and en-
gineers. During the 1980's, great strides were made in the area of analysis of
root locations of families of polynomials whose coefficients depended on para-
meters. However, most of these results did not provide the appropriate setting
for the development of design techniques.
There is no question that the problem of controller synthesis for systems with
parameter uncertainty is very challenging. The problem has not been comple-
tely solved and there are sincere doubts as to whether an elegant, computatio-
nally tractable solution such as the one for norm bounded uncertainty exists.
We believe that the methods presented in this book provide the control system
designer with powerful new tools. In this book we present a new characteristic
polynomial approach to the robust analysis and synthesis of systems with pa-
rameter uncertainty. Most of the book is devoted to single input, single output
systems. This is not due to an inherent limitation in the theory, but reflects
8 CHAPTER 1
the current state of development. Chapter 11, presents initial results on robust
V-stabilization and robust performance of multi input, multi output (MIMO)
systems.
Chapter 2, begins with the state space and transfer function representations
of LTI systems with no uncertainty. Some basic feedback configurations are
described as well as a general system interconnection structure. The notions
of stability (Hurwitz) and V-stability are defined, and their relationship to the
"poles" of the system is demonstrated. The "poles" of a system are merely the
roots of its characteristic polynomial. Very simply, a system is called stable
"Hurwitz" if all its poles lie in the open left half complex plane, it is called
"V-stable" if its poles lie inside a region V in the complex plane. Stability and
V-stability are then defined for interconnected systems, and the link is made
to the closed loop characteristic polynomial. Even though stability is of vital
interest in the operation of a dynamic system it is not the only interest. Other
performance specifications need to be imposed in order to achieve a desired
dynamic response. These can be requirements on the transient or steady state
response. For second order systems many of these requirements can be shown
to be expressible as constraints on the location of system poles. By employing
approximations these can also be used for higher order systems. Other perfor-
mance objectives such as asymptotic tracking and noise attenuation are shown
to be expressible as bounds on the magnitude of the sensitivity and comple-
mentary sensitivity transfer functions respectively.
Chapter 3, lists a number of "tests" for Hurwitz stability and V-stability. In-
deed, one of the most well known criteria for stability is the Routh-Hurwitz
criterion. Even though it has been heavily used in analysis it has not had a
substantial impact in synthesis. A more powerful result is the Nyquist Stabi-
lity Theorem, a result whose influence in control has been dramatic. One can
determine V-stability of a polynomial by examining its Nyquist plot which is
a curve in the complex plane. A variant of this result is introduced in this
Chapter, which provides an essential part of the foundation for this book. The
Finite Nyquist Theorem (FNT) is a test for V-stability of a polynomial when V
is a convez set. In contrast to the Nyquist Theorem, it only requires knowledge
of the approximate location of the value of the polynomial at a finite number
of points in the complex plane. This number could be as low as 2n + 1 where
n is the degree of the polynomial.
tainty can be norm bounded or parametric. One is now faced with having to
deal with not only individual systems but families of systems. The notions
of robust stability and robust V-stability need to be defined for interconnec-
tions of families of systems. As a result one is now concerned with the robust
stability of families of closed loop characteristic polynomials with parametric
coefficients. It is also shown that performance specifications can be expressed
as robust polynomial stability problems. It follows that this is also true for ro-
bust performance, i.e. performance specifications on families of systems. Some
multiobjective robust performance problems can be posed as simultaneous ro-
bust polynomial stability problems. This implies the need for the development
of robust analysis and synthesis methods. If one chooses a Nyquist Theorem
based characteristic polynomial approach, this would require the manipulation
of "value sets" of uncertain polynomials. In the case of polynomials with af-
fine parametric uncertainty in their coefficients, one can show that at a specific
frequency these value sets are parpolygon6. For more general types parameter
uncertainty (i.e. multiaffine, polynomic), the polynomial value sets have more
complicated shapes that present difficulties in analysis and synthesis. One way
of dealing with this problem is to introduce value set overbounds. It will later
be shown that such overbounds, even though they inevitably introduce con-
servatism, have a "twofold" benefit. They allow one to state analytical results
for robust stabilization and also lead to efficient algorithms for robust synthesis.
Chapter 5, presents some tests for robust stability. Since we are concerned
with polynomials with parameter uncertainty, our approach which is Nyquist
Theorem based requires the checking of polynomial value set locations in the
complex plane. If we know that a single polynomial in our family is stable (i.e.
its Nyquist plot has the appropriate number of encirclements of the origin),
then the entire family will be stable if for each point on the Nyquist path the
polynomial value set does not include the origin. This is referred to as the zero
ezcl'IUion condition. For interval polynomial families, Kharitonov's Theorem
showed that stability of four polynomials is a necessary and sufficient condition
for robust stability. For affine polynomial families the Edge Theorem guaran-
tees that checking the stability of the edge polynomials is a sufficient (and of
course necessary) condition for robust stability. One can show that for the af-
fine polynomial family one need not check for zero exclusion at every point on
the Nyquist path, but only at a finite number of "special" frequencies. In this
Chapter we present one such test which we refer to as the Finite Frequency Test
(FFTEST). In this test the "special" frequencies can be computed apriori from
given data. Another stability test that requires a finite number of frequency
checks is the Finite Matched Phase Theorem (FMPT). A degree n polynomial
q( 6) is stable if and only if there exists some stable degree n polynomial p( 6)
10 CHAPTER 1
such that their phases "match" mod 2?r at a specific number of frequencies. If
one is willing to employ overbounding and thus introduce some conservatism,
one can show that robust stability of the affine family can be guaranteed from
the stability of a finite number of polynomials. In Chapter 10, this rectangular
overbound of a polynomial value set will be shown to allow one to state analy-
tical results for robust stabilization and robust performance synthesis.
Chapter 6, suggests another test for robust 1>-stability, which is singled out
because of its importance. The Finite Inclusions Theorem (FIT) is actually a
corollary of the Finite Nyquist Theorem presented in Chapter 3. Assume that
the set 1> is the interior of some Jordan curve r, and that it is convex. This
new test guarantees robust 1>-stability for a general polynomial family, if the
polynomial value set at a finite number of points on r lies in an appropriately
chosen set of sectors in the complex plane. The number of such points is not
large and can be as small as 2n + I, where n is the degree of the polynomial fa-
mily. In the case of robust stability FIT requires a fewer number of points, and
this is further reduced because of symmetry in the case when the polynomials
have real coefficients. In the case of affine polynomial uncertainty FIT provides
a necessary and sufficient condition for robust 1>-stability. In view of the Map-
ping Theorem, FIT can be used to state a sufficient robust 1>-stability result for
multiaffine polynomial families. Furthermore, it is a test based only on "vertex
information". Examples are presented of how FIT is used as a robust stability
test and how it can be used in other robust analysis applications like the com-
putation of a weighted Boo norm. One can also show a relationship between
FIT and the robust stability test obtained through the rectangular overbound.
A certain equivalence is shown which involves the Hermite-Biehler Theorem in
the case when 90° sectors are used. It is important to mention that a major
difference between FIT and the Finite Frequency Test presented in Chapter 5,
is the fact in FFTEST the test frequencies can be computed apriori from the
given data.
The first six chapters complete the first part of the book and lay the foundation
for the synthesis methods presented in Chapters 7 through 11.
In Chapter 10, we show how the rectangular value set overbound allows one
to formulate "analytical" robust stabilisation results. Specifically, for an affine
plant family with k uncertain parameters, finding a controller that simulta-
neously stabilises 41: polynomials is sufficient to robustly stabilise the entire
family. For a family of interval plants, the number of polynomials that need
to be checked is reduced to 64. It is important to note that the overbounding
introduced makes these only sufficient robust stabilisation results. The issue
of how to construct controllers that simultaneously stabilize a finite number
of polynomials is also addressed. It is shown how these controllers can be ob-
tained as solutions of appropriately posed Hoc problems. Since overbounding
is introduced one is inevitably faced with the issue of conservatism. In some
cases one can compare the solutions obtained from Simultaneous Polynomial
Stabilization (SPS) with with those obtained from Hoc techniques and show
that they are not more conservative. In view of the connection between robust
stability and robust performance it is evident that the simultaneous polynomial
stabilization approach can provide solutions to robust performance synthesis.
Due to the introduction of overbounding one should expect that the solutions
obtained would involve conservatism. Hoc synthesis and the simultaneous po-
lynomial stabilization approach are carried out in an example and the results
are presented.
noise attenuation design problems which turns them into simultaneous polyno-
mial family stabiliJation problems. One can thus carry out multiple objective
robust performance synthesis using a modified SSFIT algorithm. This is again
an iterative process where at each iteration a linear program is solved.
The theory of Linear Time Invariant systems is very well developed and over
the last decade results have been produced that have successfully dealt with
many robustness issues. This book focuses on systems with parameter uncer-
tainty and suggests a different approach for robust analysis and synthesis. Our
methodology exploits the properties of polynomials and is based on fundamen-
tal results from mathematics and control. The methods can be applied to both
single input single output and multivariable systems. The results presented
complement existing theory and provide computationally efficient algorithms
for robust analysis and synthesis. The methods have been used for control-
ler design in several examples that employed software written in a MATLAB
environment. However, this software package does not meet "professional"
standards.
2
SYSTEM DYNAMICS
(2.2)
The transfer function P( 6) is just a rational function in 6, where np(6) and dp(6)
are polynomials is 6, with 8(d(6» = n, (8(·) denotes degree), and 8(np(6» < n,
(i.e. the transfer function is atrictl1l proper). We assume that np(6), dp(6) have
no common roots. The roots of np(6) and dp(6) are called zero6 and polel of
P(6) respectively. One can also give the input-output description of such a sy-
stem in the time domain in terms of a convolution operator. Systems described
in an input-output form are normally assumed to be initially "at rest". It is
straightforward to obtain the transfer function from the state space represen-
tation. This is done by taking the Laplace transform of (2.1) (assuming zero
initial conditions), and performing algebraic manipulations. One can easily
verify that P(I) = C(61 - A)-lB. Going the other direction is somewhat
more complicated, and the process is referred to as "realization". There are
a number of realization procedures that have been suggested in the control
literature [18].
Consider the single revolute link robotic manipulator shown in Figure 2.1.
y
d2 . d
I dt 2 8(t) + I dt 8(t) u(t) (2.3)
1
y(s) (Is f) u(s) (2.5)
6 +
In the state space representation (2.1), we have assumed that the input affects
the output indirectly through the state. In some cases the input does affect the
output directly through a scalar D, and the state space description becomes:
in Figure 2.2. The plant P and the controller C become the "subsystems"
which comprise the interconnection.
---
r +
+
e
C
u
P
y
y
C P
Definition 2.1 A dynamic sy6tem with tramfer function P(s) is called well
formed if P( s) is a proper rational function.
Y2
Y3
G3
r3(S) + Y2(S)
(2.7)
Let r, u, y be the 3-vectors [rl' r2, r3]', [Ull U2, U3]', and [Yl' Y2, Y31' respectively.
Let the constant interconnection matrix E be:
System Dynamics 21
Gl(S)
G(s) = ~
[
One has to insure that the overall system is well formed. There are two transfer
function matrices associated with this interconnection that describe the overall
system dynamics. Since G1(s), G2(S), G3 (s) are single input, single output
systems these are 3 X 3 matrices. The one from exogenous signals r to u is
denoted as H... (s), and the one from r to y is denoted as H,.. (s). These are
given by:
Definition 2.2 The interconnected system in Figure 2../ is called well formed
if the transfer function matrices H...(s) and H,.. (s) are proper.
2.3 STABILITY
A most fundamental concept associated with every dynamic system is the no-
tion of stability. It has been defined in many ways, but in every case it describes
dynamic characteristics of a system in a qualitative manner. One concept of
stability applies to systems which are described by an input-output relationship
and are initially "at rest". For example, a system is called bounded input, boun-
ded output stable, (BIBO stable) ,iffor every input which is a bounded function
of time, the system output is also a bounded function of time. Consider a sy-
stem described in the frequency domain by equation (2.2). In the case when the
numerator and denominator polynomials of the rational function ::f: ~, have
no roots in common, the following result is wellltnown [18]:
22 CHAPTER 2
Theorem 2.1 Let P be A .yltem with A proper traB6fer function P(.) = ::f:~
where ",,(8), dp(.) Are polynomitd6 with no common root.. The .yltem P i6
BIBO ItAble if And only if the root. of dp(.) lie drictly in the left hAlf complez
plAne.
Frequently, the adjective "BIDO" is dropped from the label and such a system is
just called "stable". Since BIDO stability of some system P is strongly related
to its transfer function P(8), the transfer function itself is called "stable".
Occasionally it is called "Hurwitz stable" and the denominator of its transfer
function dp(8) is referred to as a "Hurwitz" polynomial A system which is not
stable is called uB6tAble. From this result, it follows that the system in example
2.1 is clearly not BIDO stable, as one of the roots of the denominator polynomial
is at the origin of the complex plane. One can immediately reach the same
conclusion, working directly with the definition of BIDO stability. Suppose
that the. input u(t) to the system is a unit step, a perfectly bounded signal.
After expanding G(8)U(") in partial fraction expansion form, it is evident that
the resulting sum will include a term proportional to -!t. The inverse Laplace
transform of such a term is the function t, an unbounded signal.
For such a system, with initial condition Xo = x(to) at time to, let x(t, :1:0, to)
denote the solution of the differential equation for time t ~ to. A state X e
is called an equilibrium "tAte of this system, if x(t, :1:0, to) = x e , 'V t ~ to.
Clearly, for such a system the zero state is always an equilibrium state, since
with u(t) = 0 if the system is at the zero state at some time to, it will remain
there for all time thereafter. Other equilibrium states may exist, as determined
by the structure of the A matrix.
Lyapunov stability, characterizes the evolution of the system state when the
initial condition is "close" to some equilibrium state. Let the Euclidean norm
of x be given by II x II = (:[:=1:1:1) 1/2. The definition of a .tAble equilibrium
da.te is given as follows:
System Dynamics 23
In other words, if the system starts close enough to an equilibrium state, the
system "eventually" reaches it. Combining the two notions of stability and
convergence we have the concept of 46ymptotic ItAbility.
For LTI systems there is a very close connection between the notions of BIBO
stability and stability in the Lyapunov sense. Consider a system described by
a strictly proper transfer function as in:
24 CHAPTER 2
where np("), 4 (,,) are polynomials with no common roots and 8(4("» n.
The system can also be described by:
Theorem 2.3 The 6y6tem in (2.11) i8 BIBO "table if and only if the zero
equilibrium point of the free 6y.dem cOrTe6ponding to (2.12) i8 a6ymptotically
"table.
system equilibrium points. In view of the results in the previous section, for
LTI systems one would expect that these notions would be equivalent. Assume
that the interconnection consisted of Ie subsystems, each with transfer func-
tion G,(6) = ::f:~, where the numerator and denominator polynomials have
no roots in common. Let D(6) be the Ie x Ie diagonal matrix containing the
denominator polynomials:
In other words the interconnected system is called stable if the closed loop
characteristic polynomial is Hurwitz stable. One can prove that this definition
of stability corresponds to BIBO stability. One can also show that for LTI
systems this is equivalent to Lyapunov stability of the overall system if each
individual system is described in a minimal state space realization.
26 CHAPTER 2
2.5 V-STABILITY
The notion of stability can be extended so as to reflect other dynamic charac-
teristics. Consider again the single link robotic manipulator of Example 2.1 in
a "tracking" task. We would like to have the tip follow a "step function". In
other words go from the position 9(t) = 0 to the position 9 = 91"" where
91,n is some given angle. Clearly, this is somewhat unrealistic in that if we in-
sist on a step reference input we require this move to take place in "zero time".
Consequently, there will be "errors" involved due to physical limitations. In an
actual implementation this reference input is modified appropriately. However,
the step input is still useful in analysis because it can be used to characterize
dynamic behavior. Let us place the transfer function P( s) = .(1. \ j) in a
unity feedback configuration where the controller is assumed to be a positive
gain K. The scalar transfer function from r to y is given by:
K
7" (2.14)
This is the transfer function of a second order system with closed loop cha-
racteristic polynomial equal to ,,2 + ~" + ~. This is stahle since J > 0,
K > 0, f > O. The transfer function is in precisely the form considered in
any standard Classical Control text (see for example [57]). It is more frequently
expressed as:
(2.15)
The parameter ( is called the damping ratio and Wn the undamped natural
frequency . Both are assumed to take positive values. If this system is at rest
at time t = 0 and is subjected to a unit step input at this time (just like in
the robot example), it is well known that there is a direct relationship between
the values of (, W n and the output response of the system. In particular, if
o < ( < 1 the output response is called underdamped, if ( = 1 critically
damped and if ( > 1 overdamped. The poles of the system are at -(wn + jWdl
where Wd = W n VI - (2. In the underdamped case one can show that the
maximum overshoot is e- J~:(2 and the settling time (using the 2% criterion)
is ~. The larger the (, the smaller the overshoot and the larger the product
(w n~"'''
, the shorter the settling time. One can very easily visualize the implications
System Dynamics 27
of these concepts in the context of the single link robotic manipulator example.
Consider a plot of the pole locations of this system.
1m
Re
The ratio ,\_(2 is the tangent of the angle p. The smaller the angle p, the
larger becomes the corresponding , and so the maximum overshoot decreases.
Therefore, if the desired output response is to have maximum overshoot smaller
than a given value and the settling time is to be less than a given bound the
poles of the system must lie in a region such as the one given in Figure 2.6.
1m
Re
This is only one example of a situation where the poles of a dynamic system
may be required to lie in some region of the complex plane, other than the open
left half complex plane. Usually such a region, denoted as 'V, is a subset of the
28 CHAPTER 2
It is also evident that in addition to stability pole locations affect other per-
formance characteristics. We have just seen that the transient step response of
a second order system is "shaped" by pole locations. This property holds for
higher order systems as well even though the relationship cannot in general be
"analytically" expressed. However, assuming a higher order system is "second
order dominant" allows one to exploit this relationship.
2.6 PERFORMANCE
Even though stability is a fundamental property of any feedback system, it
seldom is the only desired performance objective. We have already seen how
some transient response performance characteristics can be "mapped" to requi-
rements on closed loop pole locations. However, this cannot be done for every
performance objective. Consider the single link manipulator of Example 2.1,
which was used to motivate the feedback configurations of Section 2.2. Recall
that the position of the tip was measured, and that the performance objective
was to have the link tip track a given reference trajectory r(t), for t 2: o. In
view of our discussion in Section 2.3, if r( t) is a bounded function of time, then
we would at least want the error e(t) to be bounded. This would impose a
stability requirement on the interconnected system. However, having e(t) be a
System Dynamics 29
bounded signal is clearly not acceptable for this application. Ideally, we would
want e(t) == 0, V t ~ O. This however is not possible in general, and one
has to seek other alternatives. One possibility is to try and make e(t) "small"
for all time. Another possibility may be to require that e(t) be "small" asym-
ptotically, i.e. uymptotic traclcing . Still another possibility is to require that
the error be zero asymptotically, lim t _ oo e(t) = O. Combining this asymptotic
requirement with a requirement on the transient response frequently results
in an acceptable performance. Let us focus now on the asymptotic tracking.
Consider the feedback system in Figure 2.7 case when the reference input r(t)
is the sinusoid Rsin(wt) of amplitude R and frequency w.
l_e_'1=-0 _--------_·I'------'_p T
Figure 2.T Unity Feedback Configuration
Let the plant be given by the transfer function P(s) = ~:f:?, and the controller
by the transfer function O(s) = ~;f:? For this interconnection we have that:
D( s) _- [ dc(s)
0
0 ]
dp(s)
N(s) = [ nco(s) 0 ]
np(s)
E _
-
[0 -1]
1 0
(2.16)
Assume that the interconnection is stable. Since the reference input r(t) is
a sinusoid of frequency w, it is well known [18], that asymptotically (i.e. in
steady state), both the output y(t) and the error e(t) will also be sinusoids of
the same frequency but with different amplitude and phase. This can be very
easily seen from Laplace transform arguments and the fact that iP( 8) is stable.
For simplicity assume that the n poles of the closed loop system are distinct,
Pi '# Pi for i '# i, 1 ~ i,i ~ n. Let iP(8) = (8 - pI)(8 - P2) ... (8 - Pn). With
zero initial conditions the Laplace Transform of the error signal is given by:
(2.17)
(2.18)
Using partial fraction expansion techniques we know that we can express e(8)
as:
a b ~ ~
+ 3W + iw + + ... +
e(8) = - - . (2.19)
8 S - 8 - Pl S - Pn
where a =- =
RSJIIoI) and b RSJr) and the Ci'S are complex numbers. It
follows immediately that e(t) is given by:
where 1·1 and arg(·) denote the magnitude and argument of a complex number
respectively. In steady state as t -+ 00 , the terms CiePit decay to zero since the
Pi'S have negative real parts. The steady state part of the error is given by:
System Dynamics 31
(2.22)
We have just seen the important role that the sensitivity transfer function plays
in the context of asymptotic tracking. It plays an equally important role in the
context of duturbance rejection for disturbance inputs which enter the feedb-
ack loop at the output of the plant P. It is also true that some "time domain"
specifications can be expressed approximately as requirements on the shape of
S(jw). Furthermore, since IS(}IoI)1 is the "distance" of C(jw)P(jw) from the
point -1 in the complex plane it characterises "relative" stability properties of
the closed loop system.
~,---c_
y
p
Assume that the reference input r(t) == 0, the disturbance input d(t) == 0
and that '7(t) is some sinusoid of frequency w. In this control problem one is
interested in eliminating the effect of '7(t) from y(t), for all t ~ O. As in the
tracking problem this is not possible in general, and one alternative is to relax
the noise attenuation requirement to asymptotic noise attenuation. In other
words require that in steady state the effects of noise on the output are "small".
Following the same reasoning as above one can identify the complementary
6en6itivity T(6), as the transfer function from -'7(t) to y(t), where
(2.23)
(2.24)
that must made in setting performance objectives and is reflected in the choice
of bounds for S(,,) and T(,,).
The performance objectives discussed in this Section address the steady state
behavior of a dynamic system. They are expressed as specifications set in the
frequency domain in terms of bounds on the sensitivity and complementary
sensitivity. It is true that many more performance objectives than the few di-
scussed here can be recast as requirements on the shape of these two transfer
functions. Actually some of these can be time domain requirements imposed
on the transient response. For instance, the overshoot in the step response of
a second order system is directly related to the peak of the complementary
sensitivity. For higher order systems this becomes only an approximate rela-
tionship, but it remains quite useful. Many more such "associations" can be
stated whose usefulness especially in design far outweigh their limitations. The
very successful Classical Control design techniques rely heavily on such formu-
lations. In this book we shall follow this trend and assume that performance
specifications can be posed either as requirements on the location of closed loop
poles or as requirements on the magnitude of certain transfer functions.
3
STABILITY TESTS
The study of polynomials and their roots has been carried out by mathemati-
cians and engineers for centuries and there exists extensive literature on this
important subject. Clearly, if one wants to determine whether a univariate
polynomial with complex coefficients has stable roots, one can compute them.
For polynomials of degree up to four, formulas exist that give the roots in terms
of the polynomial coefficients. For higher degree polynomials no such analyti-
cal descriptions exist so one may resort to a numerical computation of roots.
A better alternative is to employ analytical tests that involve only polynomial
coefficients. Further justification for such tests comes from problems in con-
troller synthesis. A fundamental problem in design is to find a controller that
makes the closed loop characteristic polynomial stable or V-stable. In such
problems the polynomial coefficients are no longer constants but functions of
the controller parameters. One would like to develop techniques for choosing
the controller so that the polynomial is stable or V-stable. Developing analyti-
36 CHAPTER 3
cal results for polynomial stability that possess an appropriate structure which
leads to efficient algorithms for design is of paramount importance in control.
In the next two sections we first recall two well known stability results, the
Routh-Hurwitz criterion and the Nyquist Theorem. Both have been extensively
used in control.
Construct the following table with n + 1 rows, called the Routh table,
h.-l
where the entries of row 3 through n + 1 are generated in the following manner:
Stability Tests 37
Theorem 3.1 The real polynomial peS) hQ.6 root,., with negative real part,., if
and only if the element,., in the firlt column of the Routh table are not zero and
have the "ame sign.
It is clear that in forming the Routh table some of the entries in the first column
may be lero, in which case care must be taken in completing the rest of the
table. Such issues are dealt with in most classical control textbooks (see [57]).
From a control perspective this is not important since we are mainly interested
in stability, and generating a lero in the first column during the construction
implies that the polynomial will not be stable. Therefore, proceeding with the
completion of the table is unnecessary.
(3.2)
The Routh table is:
1 11 6
5 13 0
8.4 6 0
9.4286 o 0
6 o 0
Since all entries in the first column are positive, the polynomial is stable. 0
generating the Routh table we use "division". Other stability tests have been
developed [14] that do not require "division". This issue is not important for
polynomials with constant coefficients. However, it does make a difference when
the polynomial under investigation has coefficients that are polynomic in some
parameters. The Routh-Hurwits test will then generate expressions that are
rational functions in these parameters. In this case it may be advantageous to
employ a stability test that involves expressions that are polynomic in these
parameters.
1m 1m
Ie
- + - - - 1 - 4 - - - - - Re --~-4-4--J---I-- Re
(a) (b)
left half complex plane, and r a curve encompassing the left half complex plane
which is made up of the jw axis and a semicircle of "infinite" radius. Such a
curve r is frequently referred to as the Nyqui8t path. However, the region 'D
could be some other region in the complex plane as we have seen in Section
2.5, and it may be that we require V-stability with respect to this region.
To apply the Nyquist Theorem, the user chooses a number of points on rand
plots their image in the complex plane. "Enough" points are plotted in order
to make the determination about encirclements of the origin. In applications,
this "sacrifice" of mathematical rigor is not a serious limitation as trial and
error, or a very fine grid of r will yield an answer. However, it would be very
beneficial from both a theoretical point of view and for applications, if one
could say with certainty how many points are enough. This is precisely the
contribution made in [47. 49]. with the Finite Nyquist Theorem.
40 CHAPTER 3
1m
Sm-l
int r
r(O) = 51
Re
Theorem 3.3 (Finite Nyqui8t Theorem, FNT) Let p(6) = Ei=OOl;si where
01; E C, and let r c C be a do,ed Jordan curve 6uch that int r i,
n ;::: 0 and
convez. Then, p i8 of degree n (i.e., 01" i- 0) and ha6 all ill zero, in int r if
1 Portions reprinted, with permission, from Proceedings 32nd IEEE CDC, San Antonio,
Texas, pp. 508-518, © 1993 IEEE
Stability Tests 41
(3.3)
(3.4)
(3.5)
(3.6)
Proof: The proof requires several preliminary definitions. Firstly, let [ = [0,1)
and define the total variation [52] of a function I: [ -+ R to be:
Secondly, let us parameterize the d06ed Jordan curve r (see Figure 3.2 by a
continuous function r: [ -+ C where r(O) = 61' For convenience both the curve
and function will be denoted by r, and the left-hand limit limt..... 1- r(t) will be
denoted by r(r). The above adjective d06ed refers to the property r(O) =
r(1-) while the adjective Jordan refers to the property V z 'I y, r(z) 'I r(y)
(i.e., r does not intersect itself). Note, an intuitive - but difficult to prove -
fact is that any closed Jordan curve r partitions the plane into two regions: a
bounded, open region int r and an unbounded, open region ext r.
Lastly, let us define the notion of a pha6e function ~: [ -+ R which, simply put,
describes the phase of the first degree polynomial s - z as s traverses r. We
42 CHAPTER 3
denote by arg • the argument of z E C \ {O}, and Arg • the principle argument
of z E C \ {O} with range (-11",11"]. Specifically, for z rt. r define 4>: 1 -+ R to
be any continuoul function such that:
For z E r\ {r(O)} take this same definition but at t' E (0,1) where r(t')-z = 0,
only require 4> to be right continuou!l and limt...t'- 4>(t) - 4>(t') E [0,211"). This
minor complication arises, of course, because (3.7)'s right-hand side is undefined
at t = t'. For technical reasons 4> will be left undefined in the z = r(O) case.
From Figure 3.3 it should be apparent that when int r is convez at least one
phase function 4> - and actually, infinitely many phase functions differing by
multiples of 211" - exists for any given z :j; r(O). Further, 4> will have certain
properties that depend only on whether z is inside, on, or outside r:
1m
Re
Now, from these observations we can prove the "if" part of Theorem 3.3. Let
p be of degree 0 :$ d:$ n and have (not necessarily distinct) zeros {Zk}i so:
n d
p(a) = La;"; = adII(a-zk).
;=0 k=1
Substituting r(t) for a and letting ~k be an arbitrary phase function associated
with Zk, we see:
II
argp(r(t» == argad + L arg(r(t) - Zk) == ~(t) (mod21r)
k=1
d
where ~(t) = Argad + L ~k(t)
k=1
Let 0 = tl < t2 < ... < t m < 1 be such that r(tk) = ak, 1 :::; Ie :::; m, and note
(3.5) implies {r(tk)} n {Zk} = 0i so, (3.6) implies for all Ie and some (k E Z,
~(tk) = 8k + 21r(k. Thus, ~(1-) = ~(O) + 21rN = 81 + 21r(N + (d where N is
the number of zeros Zk lying in int r. Now, observe:
m-l
21rn = 21rn + 81 - 8m +L 8A:+l - 8k (3.8)
k=1
m-l
:::; 121rn + 81 - 8m I + L 18 k +! - 8k I (3.9)
k=1
44 CHAPTER 3
m-l
1~(I-) - ~(tm)1 + L I~(tl:+l) - ~(tl:)1 (3.11)
1:=1
tl
~ V~ ~ L V~I: ~ 2~d ~ 2~n (3.12)
1:=1
where in going from (3.9) to (3.10) we used (3.3), (3.4) and the fact n, N, and (I:
are integers. Specifically, we use the fact that if Xis an angle -~ < X < ~ and
i is some integer then Ixl ~ Ix + 2~il. Since the left-hand side of (3.8) and right-
hand side of (3.12) are equal, we would have a contradiction if the inequality
between (3.9) and (3.10) were strict. Hence, necessarily N + (I - (m - n = 0
and (I:+l - (I: = 0, 1 ~ k < m. Summing these equations shows N = n. Thus,
as we set out to prove, p is of degree n and has all its leros in int r.
It now simply remains to prove the "only if" part of Theorem 3.3. Let p be of
degree n and have all of its leros in int r. H n =
0, trivially p(,,)=Qo =I 0,
and Theorem 3.3 is satisfied for m =I, fh =
ArgQo, and any "1 E r.
So, suppose n > O. By Nyquist's theorem, the point p(,,) will revolve hn
radians counterclockwise about the origin as " traverses r. Hence, if we choose
m = 2n + 1 and (JI:= ~(1 - 1/(3n»(A: - 1), 1 ~ A: ~ m, by the continuity of
p(,,) we will be able to determine {"I:} satisfying the theorem. 0
suitably large bounded region. The following corollary, specific to the open left
half plane, illustrates this point. (see [46] for a direct proof).
Proof: Consider a contour r comprising part of the imaginary axis and a large
half circle of radius R ~ 1 (see Figure 3.4).
1m
jR = Sm+!
jWm-l = Sm-l
Re
-jR = am
define sufficiently many points 61:, m < Ie ~ m, along the circular part of r the
"if" part of Corollary 3.1 should follow from Theorem 3.3 Indeed, if we define
= =
m m + n + 2 and (JI: 1rn/2 + 1rn(1e - m - l)/(n + 1) + £1:, m < Ie ~ m, by
the continuity of p( 6) we can determine 1£1: I ~ 1 and 61:, m < Ie ~ m, along the
circular part of r such that Theorem 3.3 is satisfied (by identifying the m in
Theorem 3.3 with m). So, p's leros all lie in the open left half plane as claimed.
The "only if" part of Corollary 3.1 is trivial when n E {O,l}. For example, one
may take m =
I, (J1 = =
Argao, and W1 O. So, suppose n > I, and note that
=
when p(6) a" n:=1(6 - ZI:) has all its zeros {ZI:} in the open left half plane:
"
~(w) = I: Arg(jw - ZI:) == argp(jw) (mod27r)
1:=1
is a continuous function that monotonically increases from -7rn/2 to 1rn/2 as
w ranges from -00 to 00. Thus, we can define:
(3.13)
Theorem 3.4 Let tf>(6) be a real, degree n polynomial with po6itive leading
coefficient. Then tf>( 6) U Hurwitz 6table if there ezuu m ~ 1 intervau (CI:' dl:) C
R, 1 ~ Ie ~ m, and real number6 0 ~ W1 ~ W2 ~ ••• ~ W m 6uch that
7r 7r
--2 <
-
C1 < d1 <-
- 2
Stability Tests 47
~n ~n
--~<Cm<dm<-+~
2- - 2
V 1:$ Ie < m max{dA:+l - CA:, dA: - CA:+l} :$ ~
V 1:$1e:$ m t/>(jWA:) C SA: = {rei 9 lr > 0, 8 E (cA:,dA:)}
(3.14)
We already know from from the Routh-Hurwitz criterion that this is a sta-
ble polynomial. Since the polynomial has real coefficients we can also apply
Theorem 3.4 to check for stability.
From the data in Table 3.1 it follows that p(s) is Hurwitz stable. The two sectors
satisfy the conditions of the Theorem, and at each of the two frequencies listed
the value of the polynomial lies in its respective sector. 0
An immediate question arises as to how these frequencies and sectors are chosen.
Several options are available. One can choose a set of sectors that satisfy the
theorem and then "judiciously" pick a number of frequencies so as to fit the
polynomial values in their respective sectors. Add more frequencies as they
are needed. Another option is to pick a number of frequencies and evaluate
the polynomial. Based on these values try to choose sectors that satisfy the
Theorem and contain the respective polynomial values. It is fair to say that this
seems to be an "open ended" procedure as we do not have at the present time
a definitive way of choosing frequencies and sectors. Even though the Finite
Nyquist Theorem is a very interesting theoretical result it must be applied with
prudence.
48 CHAPTER 3
,,~ < "1 < ,,; < "2 < ... < ,,~ < ",. < 0 (3.15)
, ,
(or "1 < "1 < "2 < ... < ",.-1 < ",. < 0)
and their highest coefficients are of like sign. One can prove (see [39]):
Let us consider a specific real polynomial of degree three and examine the
implications of the Hermite-Biehler stability conditions. In doing so, we will use
a "complex plane" interpretation of Hermite-Biehler. Consider the polynomial:
p(z) = z3 + 9z 2 + 4z + 1 (3.16)
where we have that h(z2) = 9z 2 + 1 and g(z2) = z2 + 4 are the "even" and
"odd" parts respectively. For Hurwitz stability 9" + 1 and" + 4 must be a po-
sitive pair, which in this case can be easily seen. This property has immediate
ramifications of the plot of p(jw) as w ranges from 0 to 00. The "real" part of
p(jw) is -9w 2 + 1 and the "imaginary" part is w(-w 2 + 4). The fact that the
9" + 1 has a root at -1/9 implies that the "real" part has a root at w = 1/3.
The fact that :I + 4 has a root at -4 implies that the "imaginary" part has a
root at w = 2. Therefore, for the frequency range 0 < w < 1/3 both "real" and
"imaginary" parts are positive (i.e. the point p(jw) is in the first quadrant).
For frequencies in the range 1/3 < w < 2 the "real" part is negative and the
"imaginary" part positive (i.e. the point p(jw) is in the second quadrant). In
Stability Tests 49
the frequency range 2 < w both are negative (i.e. the point p(jw) is in the third
quadrant). In other words there exist three frequencies (say 1/4, 1 and 4) for
= = =
which p(jw) lies in the sectors 51 (0,7(/2), 52 (7(/2,7(), and 53 (7(,37(/2)
respectively. It is immediately clear that the converse is also true. Specifically,
if there exist three frequencies WI < W2 < W3 that place p(jw) in the three
90° sectors 5 1,52,53 respectively then 96 + 1 and 6 + 4 form a positive pair.
Therefore, the existence of a finite number of frequencies for which the value of
the polynomial lies in appropriate 90° sectors guarantees stability. Actually, in
the Russian literature this frequency domain interpretation of Hermite-Biehler
is referred to as the Mikhailov Criterion [65].
This is precisely the flavor of the Finite Nyquist Theorem. Even though this
interpretation ofthe Hermite-Biehler Theorem is closely related to FNT, it will
be seen in later chapters that the Finite Nyquist Theorem is more powerful as it
allows for greater flexibility in choosing sectors. The Hermite-Biehler Theorem
seems to require consecutive nonoverlapping sectors that are 90° wide whereas
the Finite Nyquist Theorem can handle consecutive sectors that are up to 180°
wide with no restrictions on overlapping. This flexibility will be crucial in ro-
bust analysis and the development of robust controller synthesis methods.
The topic of polynomial stability has been extensively studied in the literature.
Our intention here is not to provide a complete exposition of all stability re-
sults that have been generated. Our viewpoint is quite narrow as we are mainly
interested in presenting the new approach for the investigation of stability and
1>-stability, made possible by the Finite Nyquist Theorem. Furthermore, we
are eager to begin to examine the effect of parameter uncertainty on dynamic
system models and its consequences on the stability of families of polynomi-
als. The next Chapter first motivates the presence of parameter uncertainty in
system models and then generalizes the notions ofstability, 1>-stability and per-
formance for families of systems (the notion of robtutne66). This immediately
points to the need for the study of root locations of families of polynomials.
4
UNCERTAINTY AND ROBUST
STABILITY
functions. In particular the single link robotic manipulator of Example 2.1 may
more appropriately be modelled by the family of transfer functions :
1
P(s, J,j) (4.1)
s(Js + j)
where J- , J+ ,f- ,f+ , are given values. This is an example of a system model
with parameter uncertainty. We could have used state space equations to
express the dynamics in which case we would have a family of state and output
equations describing the system.
represent the normalised plant deviation from unity. One possibility is to iden-
tify a stable transfer function W(a) that is an upper bound ofthese normalised
values for all Wi and 1 ~ k ~ n. We can then assume that the actual plant
belongs to the family of plants P(s) defined by:
P(a)
(4.4)
1 + ~W(a)
54 CHAPTER 4
~w
p p
a) b)
----EJ---~ + }-----r----
~w
c)
Figure 4.1 a) multiplic:ative, b) additive, c:) diviaive
Even though our motivation for using models with unstructured uncertainty
came from the experimental determination of transfer functions, this is not
the sole reason. Other dynamic system characteristics can also be modelled
by employing such perturbational models. Taking into account unmodelled
dynamic, is another example. For the single link robotic manipulator, we have
assumed that it is perfectly rigid. Clearly, this is only an approximation which
is valid at low frequencies. 1£ the link is excited by high frequency inputs,
there is no question that some "flexible modes" will be excited, effects of which
are clearly visible during operation. One way of taking into account neglected
dynamics is by the use of some form of unstructured uncertainty.
np(",a)
P(", a) (4.5)
d,,(", a)
T
feedback configuration
with plant P(s,a) = ~:~;::~ and controller C(s) = ~;~;~. The closed loop
characteristic polynomial from (2.16) has real coefficients and is given by:
It follows that the system in Figure 4.2 will be robustly stable if this polynomial
family is stable.
56 CHAPTER 4
When one uses the adjective "robust" one always assumes some type of uncer-
tainty to which the term applies. We have just seen one example of parameter
uncertainty but the notion can be extended to systems with unstructured per-
turbations. Consider the family of plants P(6), involving the multiplicative
perturbation (4.2), in the unity feedback interconnection:
~W
y
C(6) P(6) I----!.---I
If we assume that the "nominal" plant P(6) is strictly proper, one can immedia-
tely recognise that this feedback configuration is a well formed interconnected
system, as defined in Section 2.2. The interconnection of this family of plants
is stable if each interconnected system is stable for each allowable ~.
A very important fact about such an interconnected system is that if the nomi-
nal (~ = 0) feedback system is stable and there exists a ~ that destabilizes
the system then there exists a con.ftant complex ~ that destabilizes the system.
In other words if we can guarantee that the interconnection is robustly stable
for all comtant complex ~ then it would be robustly stable for all allowable
~. Suppose that the complex number ~ with magnitude less than or equal to
unity, is written as: b = b1 + jb 2 , where b1 , b2 are real parameters taking
values in the unit disc. We again see that robust stability of an interconnected
system with unstructured uncertainties can be reduced to checking the stability
of a polynomial, detD g (6, b), with "complex" coefficients, which are functions
of the real parameters b1 , b2 • Consider the feedback interconnection of Figure
4.3 where the multiplicative uncertainty is expressed as a single subsystem with
transfer function 1 + ~W where W(6) = ~:f:l, the controller has transfer
function C(6) = ~:f:l and the nominal plant P(6) = ~:f:l. Assume further
that ~ is replaced by b1 + jb2 • For this interconnection we have:
D(6) = [ ~
de
t
0
1]
Uncertainty and Robust Stability 57
00-1]
E=
[o1 0
1
0
0
The closed loop characteristic polynomial is iP( II) det(Dg(II)), and is equal
to:
Robust stability of this polynomial which has real parameter uncertainty but
complex coefficients guarantees robust stability of the feedback interconnection
in Figure 4.3.
(4.10)
Statement (4.11), continuity and the fact that dw1(6) and cP(6) are stable imply
the stability of the polynomial family:
(4.12)
--
r +
-
+
e
C(s) P(s,a)
y
For this interconnection we know that the closed loop characteristic polynomial
is: (>(6, a) = dc (6)dp(6, a) + n c (6)n,(6, a), and the sensitivity transfer function
is given by: 5(6,a) = de ~ ~.a,·a . Assume that (>(6, a) is robustly stable. The
robust performance objective is then given by:
where, Wl (6) = t:f:lis a given strictly proper, stable, minimum phase, trans-
fer function. This requirement can be equivalently stated as:
where e(6) is any stable proper transfer function with 1E>(jw)1 s 1 for all
w E [0,00). In other words, for no plant in the family and no frequency can we
have:
(4.16)
r l ei a1 n.l(jw)dc(jw)dp(jw, a) + d.l(jW)iP(jW, a) i= 0
V 0 ~ rl ~ 1, al E [0,211"), a E Oil' wE [0,00) (4.17)
for all al E [0,211"), rl E [0,1] and a E 011' Furthermore, the robust stability
of .,pS(6, a, rl, ad implies the robust stability of iP(6, a) and guarantees the va-
lidity of (4.17). Therefore, robust stability of the polynomial .,ps(6,a,rl,al)
is a necessary and sufficient condition for robust asymptotic tracking. Robust
asymptotic tracking has thus been expressed as robust polynomial stability of
a 6ingle polynomial family with real parameter uncertainty and complex coef-
ficients. As in the case of no plant uncertainty the polynomial .,pS(6, a, rl, al)
can be.thought of as the closed loop characteristic polynomial of a specific feed-
back loop (replace ~:f:~ in Figure 4.4 by ~f:::~).
r + y
C(,,) P(",a)
(4.19)
(4.20)
r + y
This immediately implies that for the feedback system in Figure 4.6 the mul-
tiobjective performance requirement of robust asymptotic tracking and robust
asymptotic noise attenuation will be achieved if and only if the two polynomials
,ps(6,8,r2,Q2) and ,pT(6,8,r2,Q2) are 6imultaneou6ly robustly stable.
(4.21)
and has coefficients that are polynomic in the parameter a. Robust stability of
the loop will follow from the stability of this polynomial family. If for the same
plant family one also requires robust asymptotic tracking as the performance
objective, one would need to check the robust stability of the polynomial family:
Since we are not dealing with a single polynomial but rather a family of poly-
nomials, the image of a point on the Nyquist path will not be a point but a set
of points. These sets are referred to as v41ue leu . In a Nyquist Theorem ba-
sed, polynomial value set approach one is faced with having to construct value
sets of polynomial families. Specifically, one needs to be able to construct the
value set 4J(II,O o ), (or 4JS(II, 0 0 )) at some point II on the Nyquist path. In the
general case when the polynomial coefficients are polynomic in a, no efficient
analytical representations exist for the value set, and its construction may be
quite cumbersome. However, in some important special cases the shape of the
value set can be described very simply [29, 30]. One such case is when the
polynomial 4J(II, a) has coefficients that are affine in a. In particular, let
Uncertainty and Robust Stability 65
Let
At a specific frequency w the mapping T : (at, a2, ... , at) --+ (z, y) is linear:
and as such the value set -q(jw,OIl)' which is the image of 011 under T is a
polytope. The shape of this polytope depends on T, and this is not a constant
map but depends on frequency. In particular, if at some frequency the rank of
T is 2, then this will be a 2-dimensional set. If T has rank 1 at some frequency,
this polytope degenerates to a straight line segment, and if T happens to be
lero at some frequency the value set is the origin of the complex plane. More
can be said about the structure of this polytope at each frequency as one can
identify its 2k edges. This can be done by eliminating the lIi parameters from
equations (4.26) one at a time. Consider the parameter ai, and assume that
E4>l (jw) i- O.
Let
Yl (W) = max Yl(W, a:z, ... , al:) attained at ii:z, ii3, ... , iil:
." ...,a.
Yl (W) min Yl(w,a:z,,,.,al:) attainedatii:z, ii3 , .", iil:
·2'·"1·.
One can immediately identify two edges of the polytope which are the parallel
straight line segments:
Proceeding in a similar fashion eliminating in turn parameters a:z, a3, ... , al: in
(4.26) one can generate 21 line segments:
This means that 2k edges of the -q(jW, 0 0 ) value set can be identified. As a
matter of fact the next result proves that these are all the edges.
Theorem 4.1 Let 4>(6, a) be (U in (4.24). Let all pair6 of column6 of the
matriz T be linearly independent at 60me frequency w. The line 6egment6
itl, it21 1 ~ i ~ 1 are the edge6 of the -q(jw, 0 0 ) value 6et.
Proof: At any frequency W I- 0, the mapping T: (al' a2,"" aA:) -+ (z, y) can
be viewed as being accomplished by first fixing a2, a3, ... , aA: and then allowing
al to vary. All these are straight lines parallel to lu, 112' We know that lu,l12
are both on the -q(jw,O o ) value set. The point 0 must lie between them by
68 CHAPTER 4
construction, (or perhaps lie on one ofthem). No point ofthe -q(jw, Oil) value
set can lie "outside" the space between these parallel lines. Suppose it did,
then it would lie on one suth parallel line, which had a y axis intercept grea-
ter than Yl_. or less than Yl....... But this is a contradiction. Furthermore,
no point of the -q(jw, Oil) value set can lie outside the line segment itself,
as this would also lead to a contradiction. These arguments can be repeated
for all other line segments lo 1 , lo2. Let i; be the unit vector in the complex
plane parallel to line ljll 1 ~ j ~ le. From the assumption that all pairs
of columns in T are linearly independent at this frequency, 'Yu,,(w) '# 0 for
all u '# v. All line segments loll lo2 will be edges with loll ljl not parallel to
each other when i '# j. We have therefore identified 2le edges. Our claim is
that these are all the edges of the q(jw, Oil) value set. To show this we shall
generate the value set in the following manner: Let S. be the subset of Oil
where tIi+l = tII+2 = ... aA:= 0 and al, a2,"" til are allowed to vary.
Let T(S.) denote its image under T. Clearly T(Sl) is a line segment in the
direction i l . T(S2) is obtained by translating T(S.) in the direction i2 as a2 is
allowed to vary. T(S2) is thus a parallelogram since i l '# i2. T(S3) is obtained
by translating T(S2) in the direction i; as a3 is allowed to vary. T(S3) is thus
a polygon with 6 edges since ~ '# i3 , i l '# i;. Assume that T(S.) is such
a polygon with 2i sides for 2 ~ i ~ m. The image of T(Sm+d is obtained
by translating T(Sm) in the direction ~+l as am+l is allowed to vary. Since
i:n+l is different from all previous direction vectors, this will add two additional
edges. The number of edges in T(Sm+d is 2(m + 1). The induction proof is
complete and therefore the -q(jw, Oil) value set has 2le edges. 0
tP(s,a) (4.30)
It is clear that for w = 0 the rank of line is equal to 1 and that for all other
frequencies the rank of line is equal to 2. From the above analysis this means
that at frequency zero the value set -"'(6,0 4 ) is a straight line segment on the
real axis and for all other frequencies it is a rectangle with sides parallel to the
real and imaginary axes [21]. As a matter of fact one can immediately identify
the four vertices of this rectangle. However, rather than reporting the vertices
of -"'(6, 0 4 ) we report the vertices of "'(6,04 ) in order to make connection with
earlier results.
2 3 a-6 4 5
1: 1(6) at + a2"6 +aj6 + a+6
4 + 5 + a-6
6 +
1:2(6) a+1 + a+6 2 a-6 3 a+6 4 a+6 5
2 +aj"6 + 4 + 5 + 6 +
A: 3(6) = a+1 + a2"6 +ai'62 + at 63 + at6 4 + a6"6 5 +
2 a-6 3 a-6 4 5
1:4(6) at + a+6
2 +aj6 + 4 + 5 + a+6
6 +
(4.31)
to provide value set descriptions that facilitate robust control system analysis
and synthesis. In doing so one should be prepared to give up some "accuracy"
and introduce "overbounding" if this can lead to analytical results or more effi-
cient algorithms. Uncertainty representation in many cases is the "key" for the
development of efficient methods for robust analysis and synthesis. A perfect
example of this practice for a family of plants with parameter uncertainty in
the numerator and denominator of the transfer function, is to overbound by
say multiplicative "norm bounded" uncertainty. The parameter uncertainty is
thus overbounded by a due in the complex plane. This formulation allows one
to exploit the powerful Hoo tools for robust analysis and synthesis. One can
argue that the benefits gained by overbounding far outweigh its disadvantages.
When the polynomial family has coefficients that are multiaffine in 8, one can
use a different overbounding scheme. In this case the mapping T : (alt a2, ... , al:)
--+ (z, y) at each frequency is multiaffine. The Mapping Theorem of Zadeh
and Desoer [73] immediately implies that at any frequency w the set T(O.. ) is
included in its convex hull. In other words, the value set at any frequency w
can be overbounded by a set which is analytically represented by its vertices.
This is a very efficient characterization, and is another example of value set
overbounding in the complex plane.
;(s, al) = hns" + hn_lS,,-l +.,. + ho +al U"s" + f"_l S"-l +... + fo)
" ".. " 'Y .,
4>0(') 4>,(,)
(4.32)
where al E 0 111 = {al E RI- 1 ~ al ~ I}, is a single real parameter and ;o(s),
;1(1), have complex coefficients. Let us first "augment" the given family by in-
troducing "fictitious" uncertainty. Define the new family (referred to as "family
~"):
It is clear that the set of polynomials ;(1, 0 111 ) is included in the set of polyno-
mials ~(s, 0111 ), Therefore if ~(I, 0111 ) is a stable family so will family ;(,.0111 )
[25].
2Portiona reprinted, with permission, Crom Proceedings 30th CDC, Brighton, England,
December 1991, pp. 427-432, @ 1991 IEEE
72 CHAPTER 4
Lemma 4.1 The value 6et q(jw, ( 01 ) at each frequency w u a rectangle, who6e
vertice6 corre6pond to the polynomia16 tPt(jw), -tPt(jw), ~t(jw), -~t(jw).
Proof: Consider the edge of 001 in the first quadrant. Since Re(~t(jw)) =
Re(tPt(jw)),Im(~t(jw)) = -Im(tPt(jw)), its image in the complex plane at
frequency w is:
This is a straight line segment parallel to the imaginary axis between the points
(Re(tPt(jw)), -Im(tPt(jw))) and (Re(tPt(jw)),Im(tPt(jw))) which correspond
to polynomials ~t (jw) and tPt (jw) respectively. The image of the edge at =
1 + at is:
This segment is a straight line parallel to the real axis between the points
(Re(tPt(jw)), -Im(tPt(jw))) and (-Re(tPt(jw)), -Im(tPt (jw))) which correspond
to polynomials ~t(jw) and -tPt(jw) respectively. The image of the edge at =
-1- at is:
This is a straight line segment parallel to the imaginary axis between the
points (-Re(tPt(jw), Im(tPt(jw))) and (-Re(tPt(jw)), -Im(tPdjw))) which cor-
respond to polynomials -~t(jw) and -tPt(jw) respectively. Finally the image
of the edge at = -1 + at is:
Uncertainty and Robust Stability 73
This is a straight line segment parallel to the real axis between the points
(-Re(<Pl(jW), Im(<Pl(jW))) and (Re(<Pl(jW)), Im(<Pl(jW))) which correspond to
polynomials -~1 (jw) and <PI(jw) respectively. All other points of OGI get map-
ped to the interior of the rectangle shown in Figure 4.10, where for illustrative
purposes we assume Re(<Pl(jW) > 0, Im(l/>l(jw)) > O.
1m
- f--
Re
In our discussion thus far we have explored how to overbound value sets in the
complex plane. Value set overbounding can also be accomplished indirectly in
parameter space. Consider the case of polynomials families with coefficients
that are quadratic in al. For example the term a~ can be written as Qlq2,
where Ql, Q2 are thought of as independent parameters both taking values
over the same range as al. Clearly, this is a very "crude" overbound and the
representation is by no means unique. As a matter of fact, al -+ aQI + (1- a )q2
and a~ -+ qlq2 is a whole family of substitutions accomplishing this task,
parameterised bya. Note, of all the possible substitutions al -+ aql + (1- a )q2
74 CHAPTER 4
2
U {(al' aD E R 1 a1 ~ al ~ at }
Ua {(aql + (1 - a)q2' qlq2) E R 2 1 ai ~ qi ~ at ,1 < i < 2,}
forefront of control theory research and practice. The literature on the subject
is extensive, and the few remarks that follow are narrowly focused on problems
with real parameter uncertainty.
In the context of linear time invariant systems, several analysis and design
methodologies have been proposed and among the most successful and widely
used are the H oo methods [35, 36] and LQG/LTR [37]. Even though, these
methods have been developed with unstructured (norm bounded) uncertainties
in mind, they can be applied to systems with real parameter uncertainty. The
success ofthe H oo methodology stems from the fact that it addresses important
control problems, it is based on a firm theoretical foundation, and it is insightful.
Furthermore, analytical results have been developed and efficient computer
algorithms have been suggested for automated controller design. Consider the
case when the plant family is defined in terms ofthe multiplicative perturbation:
(4.34)
with P(,,) strictly proper. One can show [36] using the Nyquist criterion,
that if the controller of a unity feedback configuration makes the nominal clo-
sed loop (i.e. ~ = 0) stable, the system is robustly stable if and only if
IW(jw)T(jw) I < 1 'V wE [0,(0). The transfer function T(,,) is the nominal
complementary sensitivity. For analysis this is an almost trivial computation.
Even more remarkable is the fact that the problem of finding a controller that
stabilizes the nominal feedback loop and minimizes the H oo norm of W(,,)T(,,)
can be very elegantly solved by H 00 methods (i.e. interpolation theory or the
two Riccati equation method).
Over the last decade there has been a great deal of interest in the analysis of
systems with real parameter uncertainty, which was "fueled" in part by Kha-
ritonov's result [50]. Attempts have been made to develop design methods for
parametrically uncertain systems (based on these recent developments), but
thus far have had only limited success [7, 10, 28, 41]. A promising alternative
technique for single input single output systems was presented in [51]. This
machinery however is limited to the single parameter case. An interesting ap-
proach is presented in [60] that addresses synthesis problems for a certain class
of systems with parameter uncertainty in terms of infinite-dimensional convex
optimization.
It has been shown in Sections 4.2, 4.3 and 4.4, that many control problems
can be formulated as robust polynomial stability problems. In some cases the
polynomial families have real coefficients, but in general the polynomial families
have complex coefficients. Therefore, if one wishes to employ this formulation
in the analysis and design of control systems, one has to develop appropriate
tools for robust polynomial stability and robust polynomial stabilization.
Consider the problem of robust stability in Section 4.2. Suppose that the plant
involves parameter uncertainty where 0 4 is some hypercube in parameter space
and that a controller has been chosen. The closed loop characteristic polynomial
is given by:
One needs to check if this polynomial is robustly stable, in other words, that it
is stable for all values of the parameter a E 0 4 • Since the coefficients are real,
Uncertainty and Robust Stability 77
both the Routh-Hurwitz criterion and the Nyquist criterion can be employed.
For the Routh-Hurwitz criterion, the Routh table can be constructed with
the use of a computer with symbolic computation capability, but the entries
will now be functions of the uncertain parameter a. If the coefficients of the
characteristic polynomial are polynomic in a, then the entries in the Routh
table will be rational functions in a. One is now faced with having to check
whether the entries in the first column are positive for all a E 0 0 , One could
have used a different criterion [14] whereby the positivity of a finite number of
polynomial expressions in a need to be checked. These are both problems in
algebraic geometry which in general are very difficult to solve, see [4, 3].
The Nyquist Theorem (i.e. the argument principle) can also be applied to
rational function families. The standard problem is determining stability of
a unity feedback loop with the family of transfer functions P(6, a)C(6) in the
forward path. Assume that the number of unstable poles of P( I, a) does not
change. At each point on the Nyquist path, one would now have a different value
set, and would have to determine the number of encirclements of the -1 point.
Having the appropriate number of encirclements of the -1 point for a single
parameter value, and the "exclusion" of -1 by each rational function value
set, would guarantee the robust stability of the unity feedback loop. Rational
function value set construction is the cornerstone ofthe QFT methodology [43].
important special cases value sets can be easily described analytically. However,
in general value sets cannot be described analytically in an efficient manner and
performing value set computations is not an easy undertaking even with the
use of powerful software.
Using this Nyquist Theorem approach for synthesis makes the procedure even
more complex in that the value sets are now parameterized in terms of the
controller parameters. Even though this formulation is straightforward and
easy to explain it makes development of "automated" design procedures very
difficult. One alternative is to sacrifice mathematical rigor by considering only a
finite number of points on the Nyquist path (as in QFT). Another alternative is
to use approximations and rely heavily on experience. Such practices have been
used in the past with success, but one should always strive for new theoretical
results that lead to improvements.
Other value set overbounding techniques can also be explored. The rectangular
value set overbounding introduced is one possibility. We will see in Chapter
10 that such endeavors do lead to efficient analytical value set representations
and to conditions for robust analysis and synthesis. The disadvantage is that
these are only sufficient conditions and result in robust analysis and synthesis
methods that are conservative.
5
SOME ROBUST STABILITY TESTS
In this Chapter we will first recall (without proof) some earlier results and then
present in detail two robust stability results. The first is the Finite Frequency
Test which asserts that robust stability of a polynomial family with affine un-
certainty can be determined from a finite number of frequency checks. In other
80 CHAPTER 5
The most general family of polynomials under investigation in this book is given
by:
where the ;i(6), 0 ~ i ~ u are given polynomials. Suppose that the parame-
ter a takes values in the hypercube 0/1 = {a E RI: la; ~ Oi ~ at, 1 ~ i ~ Ie},
a; < 0, at > 0, 1 ~ i ~ Ie. The leading coefficient of ;(6, a) is positive
and the family has degree n. Further assume that Qi(a),l ~ i ~ u are
polynomic in a and such that Qi(O) = O. Both the Routh-Hurwitz criterion
and the Nyquist Theorem can be used for determining robust stability of this
polynomial family. If some of the coefficients are complex or we are interested
in general1>-stability the Routh-Hurwitz criterion cannot be used.
Theorem 5.1 A6Iume that at lefUt one polynomial in the family; i8 'D-dable.
The entire family; i8 'D-,table if and only if 0 fJ. ;(,,04 ), for all , on the
boundary of 'D.
The result states that a polynomial family will be 'D-stable if one of its members
is stable and the origin ofthe complex plane is excluded from the ;(" 0 4 ) value
set for all , on the boundary of 'D. To apply this result one must be able to
construct the value sets and carry out a "sweep" of the boundary of 'D. It
would be very desirable if more efficient analytical solutions could be provided.
(5.3)
Theorem 5.2 The inte1'1lal polynomial family in (5.9) u Itable if and only if
it. four Kharitonov polynomiau are Itable.
Theorem 5.3 The polynomial family tP(6, a) with affine uncertainty u .!table
if and only if all the edge polynomia16 of the hypercube 0,. are .!table.
Along with Kharitonov's Theorem this result generated a great deal of excite-
ment in the control research community during the 1980's. Even though the
Edge Theorem is a very valuable analytical tool, it carries with it a heavy
computational burden when it is applied. Clearly superior to "gridding", it
nevertheless involves the stability determination of an exponential number of
single parameter families of polynomials.
Some Robust Stability Tests 83
°
where the ~i(6), ~ i ~ Ie are given polynomials with ~O(6) monic of degree n
and the rest of degree less than n and Oil = {a E R11a; ~ at ~ at, 1 ~ i ~ Ie},
a; < 0, at > 0, 1 ~ i ~ Ie. If we are to follow a Nyquist Theorem approach
involving the sero exclusion condition, we are required to compute the value set
~(jw, Oil)' for each nonnegative frequency, (one can show that examining the
entire Nyquist path is not required but only the jw axis). In view of the fact
that we are dealing with a polynomial family of degree n, we need not consider
the entire nonnegative axis. One need only guarantee sero exclusion over some
finite interval "M" of frequencies w E [0, M]. We already know that at each
frequency the value set is a 2k convex parpolygon. In applications, one would
"grid" the frequency axis and check for sero exclusion at a finite number of
points. One would then hope that no critical frequency was "missed" and that
the test results are correct. Certainly, it would be desirable to have rigorous
analytical results and not to have to deal with questions about the outcome
of some test that involves gridding. One can actually show that this is indeed
possible. A stability test for the polynomial family in (5.5) exists, that requires
only a finite number of frequency checks [30]. This number is polynomial in le,
0(1c 3 ), and these critical frequencies correspond to the real nonnegative roots
of some polynomials that can be constructed from given data.
Let us consider first the case when only two parameters are present in the
polynomial family:
1 Portions reprinted, with permi••ion, Cram IEEE Trana. on AC, Vol. 34, No.9, 1989, pp.
982-986, @ 1989 IEEE
84 CHAPTER 5
These are valid expressions for the edge lines as long as E 4>1 (jw), E 4>2 (jw),
04>1 (jw) or 04>2 (jw) are not zero for some w. Since the polynomial family
contains at least one stable polynomial, then it contains only stable polyno-
mials if and only if the point tPo(jw) lies outside the value set -q(jw, 0 0 )
at each frequency. Given that this value set has a boundary described by
the lines ill,i12,i2lti22, tPo "enters" the -q(jw,Oo) value set if it crosses
one of these lines at some frequency. Evaluating these lines at the point
(z, y) = (E4>o(jw), 04>0 (jw)), generates the four polynomials:
Let g,(w), i = 1,2,3,4 be those g,(w) which are not identically zero and define
90(W) = W(E4>I 0 4>O - 04>IE4>O)' Let
CF1 =
{wlw = 0, or a real po"itive zero of g" i = 1,2,3,4, E4>I' E4>3' 04>1' 04>2' "Y12}
CF2=
{wlw = 0, or a real po"itive zero of go, E4>lI E4>3' 04>1' 04>2}
Some Robust Stability Tests 85
Theorem 5.4 (Finite Frequency Te.tt, two parameter ctue) Let E"'l' E"'2' 0"'1'
0"'2 be nonzero polynomiaZ,. The family (5.6) contaim only Itable polynomiaZ,
if and only if ;o(jw) lie' outlide the value ,et -q(jw, Oil) at the following
critical frequencie, :
Proof: Let us first consider the case when 1'12 # O. Now if ;0 is outside the
value set -q(jw, Oil) at some frequency (say w = 0) it can enter it through an
edge. This implies that the point ;0 must lie on one of the edge lines at some
frequency. Specifically, we must have:
These expressions are valid at all frequencies except the zeros of E"'l (jw),
E"'2(jW), O"'l(jW) or 0"'2(jW). Suppose that none of the 9i polynomials are
identically zero, and let C F 1 be the set defined above. It is then clear that ;0
is outside the -q(jw, Oil) value set for all frequencies if and only if ;0 is outside
the -q(jw, Oil) value set at the finite number of frequencies in CF1 • The set
C F 1 contains the potentially troublesome "critical frequencies" that need to
be checked. Suppose that exactly one of the polynomials in (5.9) is identically
zero. This implies that the point ;0 lies on an edge line of the -q(jw, Oil) value
set for all frequencies. Then;o will be outside the -q(jw, Oil) value set for all
frequencies if and only if it is outside the -q(jw, Oil) value set at frequencies
86 CHAPTER 5
in CFt. This is true because the only way the <1>0 point can enter is through
a different edge (one other than that corresponding to the identically lero g,).
If more than one g, is identically lero, then either 112 = 0 (a contradiction
to our assumption), or the frequencies in CFt are adequate to determine loca-
tion. When 112 = 0 we have the following simplification. This is because the
-q(jw, (}Il) value set reduces to a line segment since:
E;2 )
z = -E;1 (at + -a2 = -E;I At
E;I ,
...
Al
0;2 )
y -wO;I(at + Oa2 At -O;I At
;1
which implies:
(5.10)
(5.U)
Now 90 cannot be identically lero because this would imply that [29]:
(5.12)
where g( 6) is arbitrary and et, e2 are even polynomials. Since <1>0 is stable e2
must be equal to a nonzero constant otherwise <1>0 would have at least one root
in the right half complex plane. But then g(6) would be a polynomial of degree
n. However, by assumption <l>t can have degree only up to n - 1. The set of
critical frequencies is thus C F2 • 0
Remark: In the Theorem we have assumed that E;., 0;. are not identically
lero. The Theorem holds in these cases as well as we now demonstrate. Let
112 f. 0, E;I = 0 and E;2' 0;1,0;2 nonzero. The gi polynomials are:
Some Robust Stability Tests 87
If none of these are identically zero the set of critical frequencies becomes:
CFf = {wlw = 0, or a real po6itive zero 0/ gi, i = 1,2,3,4, E.2, 0." O.2}
Similarly if E.2 = 0, or 0.. = 0, or 0.2 = o. If say E.. = 0 and 0.2 = 0
then:
parameter. If E. 2 = 0 then the -q(jw, nil) value set lives always on the
imaginary axis and thus we need only check frequencies which make E. o = 0
(a finite number).
88 CHAPTER 5
Let us now consider the k-parameter case. We already know that the -q(jw, Oil)
value set is in general a 2k convex parpolygon. However, one can show that
under certain conditions the number of sides can actually be less than 2k. In
Section 4.5 we considered the polynomial family with affine uncertainty:
We say that polynomials IPu and IPlI satisfy a ,haping condition [29], if and
only if 'YUlI (jw) is identically lero. One can easily visualise the implication of
such a condition on say polynomials IPl and IP2. If 'Y12(jW) = 0, this implies
that 0 ~1 (jw )E.", (jw) - E~l (jw)O ~2 (jw) is the lero polynomial. Therefore, the
columns in T that correspond to polynomials IPl and IP2 are linearly dependent
for all frequencies. This implies that the direction vectors ii, i2 are collinear
for all frequencies and that the resulting parpolygon has two less sides.
In the next result we will use the fact that the -q(jw, Oil) value set is a parpo-
lygon with at most 2k sides to show that assuming the family of polynomials
with affine uncertainty contains at least one stable polynomial, robust stability
can be determined from a finite number of frequency checks. This number is
shown to be O(k 3 ) and not exponential in k as one might have expected.
Proof: First consider the case when no shaping conditions are satisfied (i.e.
'YulI(jw) # 0, 1 ~ u,v ~ k,u # v). Let
Some Robust Stability Tests 89
Cl..... (W) =
max (11202
·2,.·.,0,11
+ .. ,+ 111: 0 1:) = 1124 + 1130 3" + 1144 + 115 0 5" + ... + 111: 0 t
Cl......(W) =
mm . (11202 + ... +
111: 0)
1: = 11202-+ 11303
+ + 11404- + 11505+ + ... + 1a01:-
·2, ... ,011
This implies that at frequencies in M ll (which are not leros of E~l' O~l) the
two edge lines are:
This procedure can be repeated for all other parameters as well (i.e. 02,03, •.. ,01:),
Specifically, for polynomials 1"., 1 ~ v ~ le, u f. v let
V. =
{wlw 0 reol po6itive zero 0/ 60me 1"., 1 ~ v ~ le, u f. v in the intervol M}
k-1
z ~ n --k
2
+ 1 (5.14)
What we now have is that within each interval Mi the 2k edge lines are:
E. 1 Y wO. 1 Z = WClimaz
E. , Y wO., Z WClimi"
E. 2 Y wO. 2 Z WC2imaz
E. 2 Y wO. 2 Z WC2imi"
where the -Cjima:C1 -Cjimi" are solely functions offrequency and do not depend
parametrically on the tit '5. It is then clear that the point ifJo will be on anyone
of these lines for w E Mi if anyone of the 2k polynomials:
= w((E.,O.O Climaz)
= w((E.,O.O Climi,,)
has a real positive zero in Mi' Assuming that none of the polynomials gti are
identically zero define:
Ui = {wlw a real positive zero of some gti,l ~ t ~ 2k, in the interval Mi}
and let U = U 1 U U 2 U ••• U•• We are now in a position define the set of
"critical frequencies":
Some Robust Stability Tests 91
This is clearly a finite set of frequencies. It contains all the frequencies that
are potentially troublesome in the sense that it is at one of these that the tPo
point has a chance of entering the -q(jw, (}/I) value set. If then tPo lies outside
the -q(jw, (}/I) value set at these frequencies it must be outside this value set
for all w.
tion 1..
Suppose that some (not all) of the gei are identically equal to zero. Our assump-
:f:. 0 for all u, v, is still in effect. This corresponds to the situation
where tPo lies on one or more of the gti for all frequencies. Now if the -q(jw, 0/1)
value set continues to be of dimension 2 for most frequencies (except perhaps
at positive real zeros of 1•• ) and tPo lies outside the -q(jw, 0/1) value set for
some w (say w = 0) it can enter only through a different edge. We now modify
the sets Ui and U by excluding these zero polynomials. Specifically, let
Now suppose that some shaping conditions are satisfied. The consequence of
In the case we are considering (not all 1•• = 0) we have k* > 1. The required
modifications to the set of critical frequencies are: First define for 1 :5 u :5 k*,
92 CHAPTER 5
If some of the gei now become zero the set of critical frequencies should be
suitably modified (as in the 2 parameter case) and set equal to CF4 •
Finally let us direct our attention to the situation when all '1.11 = 0, but ;0
does not lie on the -q(jw, 0 4 ) value set for all frequencies. The -q(jw, 0 4 )
value set is a straight line segment given by:
(5.17)
We have just presented the Finite Frequency Test, a robust stability result
for families of polynomials with affine parameter uncertainty. The basic idea
behind the proof was the fact that the only way the point ;o(jw) could enter
the -q(jW,04 ) is through some edge. This happens at the roots of specific
Some Robust Stability Tests 93
Consider the state space model for a "crane" [1] where me is the crab mass, mt
is the load mass, l the rope length, g the gravitational constant, Zl the crab
position, Z2 the crab velocity, Z3 the rope angle, Z4 the rope velocity, and u
the input force accelerating the crab.
o
~
x me
o y =Z (5.18)
(me+mt)g
mel
Suppose that all states are available for feedback purposes and that we employ
a constant state feedback controller C, (u = -CZ), in order to robustly
stabilise the system. For this Example assume that the rope length land
the load mass mt are only known to lie in given intervals. Let l = 1/1 with
l = alO + al, mt =
a20 + a2 where alO =
1/10 corresponds to some
nominal rope length and a20 = 1 is the nominal load mass. Let 0 ~ al ~ .01,
and 0 ~ a2 ~ 10. Let g =
10 and me =
10. One can verify that the
controller C = [240 500 - 990 4000] places the poles of the nominal closed
loop characteristic polynomial at -1, -2, -3, -4. The closed loop characteristic
polynomial det(,I - (A - BC)) is given by:
+ + + +
4>(" a) ,4 10,3
..
35,2 + 50, 24,
"'0(')
,,2
+ + + + + --.-.-
,,2 at
.,.,(...
(-400,3 110,,2 500" 240), at a2
--.-.-
a2
, 10
) .,..(.)
.,.,(.)
94 CHAPTER 5
We would like to use a variant of the Finite Frequency Test in order to de-
termine whether this polynomial family is robustly stable. Even though the
uncertainty enters the polynomial coefficients in a multiaffine manner, one can
show [31] that the -q(jw, 0/1) value set at each frequency is a polygon (actually
a trapezoid). It is then clear that the ideas presented earlier in this Section ap-
ply and that we should expect to be able to verify robust stability from checking
the location of the -q(jw,O/l) value set at a finite number of frequencies. Let
One can identify the four edge lines of the trapezoid as:
-(1l0w 2 - 240)
y = 0, y = -w(4w 2 + 5), 100w(4w 2 + 5) y,
= (-120w 2 + 240) + w2
:z: lOOw(4w 2 + 5) Y
91(W) wO.o(jw)
92(W) wO.o(jw) + w(4w 2 + 5)
93(W) wO.o(jw)(1l0w 2 240) + 100w(4w
2
+ 5)E.o(jw)
94(W) wO.o(jw)(120w 2 - 240) + 100w(4w 2 + 5)(E.o(jw) - w2)
where
V24 v'55
CF = {O, r.-:t' v'S, ID ,6.041523009,6.144102841}
yll y6
For these frequencies we need to construct the -q(jW,04 ) value set. For this
example a check reveals that the point 4Io(jw) is excluded in each case and the
polynomial family is thus stable. 0
Assume that arg z (the argument (phase) of a complex number z), takes non-
negative values. For polynomials p(a), q(a), let again the polynomial 'Yqp(lI)
be:
(5.19)
96 CHAPTER 5
Theorem 5.6 (Finite Matched PJuue Theorem, FMPT) Let q(6) be a degree n
monic polynomial with real coefficienu. Then q( 6) u ,table if and only if there
ezuU a degree n monic ,table polynomial p( 6) ,uch that for the nonnegative
rooU Wi of w'Yqp(jw), we have erg p(jWi) = erg q(jwi) (mod 2,..).
Proof: Necessity comes from the fact that since q(6) is stable then there exists
a small enough ( > 0 such that for all the frequencies Wi which are the
nonnegative roots of OIJ(jw), (0 < Wl < W2 < W3 < ...), we have:
and P(6) = =
q(6) + E is stable. Then 'Yqp(8) 01J(6)E,,(6) - EIJ(8)0,,(6) =
0'1(1)(. By construction erg p(jWi) = erg q(jw,) (mod 2,..) for all the nonne-
gative roots of w'Yqp(jw).
ofthe polynomial w'Yqp(jw). For a degree n polynomial q(6) there are at most
n such frequencies. Therefore, if for these frequencies the value set ;(6,OA)
excludes the origin the family ;(6, OA) will be stable. This is guaranteed if at
these frequencies Wi, we have that arg P(jwi) = arg q(jWi) (mod 2'll'). Since
the family includes polynomial q(6), then it also will be stable. 0
Actually, in Theorem 5.6 we have proved more than just stability of the q(6)
polynomial. The entire polynomial family (5.20) will also be stable, so FMPT
is actually a robust stability result.
It is interesting to compare the Finite Nyquist Theorem and the Finite Matched
Phase Theorem. Even though both results have to do with the value structure
of polynomials at a finite number of points on the imaginary axis, they ne-
vertheless are different in some respects. Whereas in FNT the frequencies are
chosen so that no two consecutive points are more than 1800 apart, no such
requirement is imposed in FMPT. The frequencies that need to be checked are
simply the real roots of some polynomial w'Y"q(jw). This fact also has implicati-
ons on the number of frequencies that need to be checked. In FNT the number
r
of frequencies cannot be less than n/21 whereas in FMPT it could be just the
single frequency w = 0 if 'Yqp(w) has no real positive roots. Furthermore, in
FMPT the frequencies can be computed apriori from given data whereas in
FNT one needs to find frequencies (and sectors). The following corollary to
Theorem 5.6 sheds some more light on this relationship.
Corollary 6.1 Let q(6) be II degree three polynomilliwith q(O) po,itive, which
lit frequency W1 lie, on II line of po,itive ,lope 1:1 in the fir6t qUlldrant lind lit
frequency W2, W2 > W1 lie, on II line of po6itive 610pe 1: 2 in the third quadrant.
Then q(,) i6 6table if 1:2 /1: 1 < W2/W1.
Proof: In view of Theorem 5.6 all we need to show is that the conditions given
guarantee the existence of a stable polynomial p(,,) that matches the phase of
q(,,) at the appropriate finite number of frequencies. Let p(,,) = ,,3 + P2,,2 +
P1" + Po, with E,,(jw) = -P2w 2 + Po and O,,(jw) = _w 2 + Pl. Let P1 be
positive with wJ < P1 < w~ and choose P2,Po as the solution to the linear
equations:
98 CHAPTER 5
k2W1(P1 - wn +
k1W2(W~ - pd
=
k1k2(W~ - wn
k2W1WHp1 - wn + k1W2WHw~ P1)
Po
k1k2(W~ - wD
One can immediately see that if k 1l k 2, W1, W2 satisfy the conditions of the
Corollary, then such a p( 6) is stable since Po > 0, P1 > 0, P2 > 0 and P2P1 - Po >
0. We can therefore construct a stable polynomial whose phase matches that
of q(6) at frequencies 0, W1,W2' By Theorem 5.6, q(6) is stable. 0
It is interesting to note that with the data given in the previous example one
would not have been able to conclude that q(,,) is a stable polynomial if one
employed the Finite Nyquist Theorem. The reason is that the points (2,;1) and
(-115/2,-;115) are more than 1800 apart. However, stability can be deduced
from the Finite Matched Phase Theorem. One would expect that results such
as Corollary 5.1 can be stated in a broader context for higher degree polynomi-
als. It is also true that one can generate robust stability results for polynomial
families based on this concept. However, at the present time we do not have
strong results to present. An additional reason why we have not pursued more
research along this direction is that the Finite Matched Phase Theorem does
not seem to lead to efficient methods for controller synthesis.
'(8, al) = hn8" + hn_18,,-1 +... + h.o +al (1,,8" + f"_1 8"-1 +... + fo)
, v ,,, ". ,
·M·) .,(.)
(5.22)
Lemm~ 5.1 As,ume that hn + ad" + ad" ::P 0 for all (al' (h) E OGI' The
family, i8 a ,table if and only if the four polynomial8 tPO(8) + tP1(8), tPo(s)-
'1(8), '0(8) + ~1(8), '0(8) - ~1(8) are 8table.
Proof: Necessity is clear. Suppose then that the four given polynomials are
stable. Note that the entire family ¢ is of degree n. From [ll] it is clear
that since '0(8) + '1(8) and '0(8) + ~1(8) are stable then the following convex
combination, 0 :5 a1 :5 1 is also stable (real parts are the same when 8 = iw):
But this set of polynomials is derived from the first quadrant edge of OG" and
they are all stable. Consider the convex combination, 0 :5 a1 :5 1:
Some Robust Stability Tests 101
Again from [11] we conclude these are stable polynomials, and these are derived
from the second quadrant edge of 001 , Consider now the convex combination,
0:$ iii :$ 1:
These are again stable polynomials, derived from the third quadrant edge of
001 , Finally consider the convex combination, 0 :$ al :$ 1:
These are again stable polynomials, derived from the fourth quadrant edge of
001 • Since the "ed~e" [9] polynomials are stable and the leading coefficient
nonzero, the entire tP family is stable. It is evident, that Lemma 5.1 is thus a
sufficient condition for the stability of (5.22). 0
The result for an affine family with Ie parameters follows in a natural way. Two
options are available to us. We can employ the single parameter analysis to
each of the edges and overbound each of them with a rectangular overbound or
impose a rectangular overbound on the entire value set. We choose to pursue
the second alternative. A similar result would be obtained if we proceeded with
each edge. Consider the Ie-parameter affine family of monic polynomials:
1
~=1+4i
a..=-l-aa
Let
Lemma 5.2 The value 6et q(jw, ( 0 ) at each frequency w u a rectangle, who6e
vertice6 come from 60me four polynomial6 taken from the 6et:
Some Robust Stability Tests 103
Proof: From Lemma 5.1 the value set for each term a;f;(jw) + a.;~;(jw) is a
rectangle whose vertices correspond to polynomiaJa ,p;1 (jw), ,p;2(jW), ,p;.(jw),
,p;.(jw). The a; parameters are independent and thus the value set q(jw,Ota)
is the sum of k rectangles with sides parallel to the axes. The overall set
q(jw,Ota) is thus a rectangle with sides parallel to the axes. It is interesting
to note that summing hon.ontal sides one from each of the k rectangles, ge-
nerates a hon.ontalline segment in q(jw, Ota), (there are up to 21 such lines).
The hon.ontal sides of the q rectangle will be generated from two such com-
binations. At this point we do not know from which two, and at different
frequencies these may change. Similarly, summing vertical sides one from each
of the k rectangles generates a vedicalline segment in q(jw, Ota), (there are up
to 21 such lines). The vertical sides of the q rectangle are generated from two
such combinations. The four vertices of q(jw, Ota) clearly come from the set
,pl'1 (jw) + ,p2'2 (jw) + ... + ,pt,. (jw), 1 $ i l $ 4, 1 $ i 2 $ 4, 1 $ i3 $
4, 1 $ i" $ 4. There may be frequencies at which this rectangle collapses to a
line segment on one of the axes, or even to the origin. 0
Remark We know the the q(jw, (}ta) value set at each frequency is in general
a 2k convex parpolygon. By construction we know that it is included in the
rectangular value set in q(jw, Ota). We illustrate the rectangular overbound for
the case when A: = 3. The "thick" lines of Figure 5.2 show the q(jw, Ota) value
set and the "thin" lines the q(jw, Ota) value set. It is interesting to note that for
this frequency the "edge-wise" rectangular overbound is identical to the "value
set-wise" rectangular overbound.
In view of this result one can generalize Lemma 5.1 to the k parameter case.
104 CHAPTER 5
Lemma 5.3 The family ~ iI ,table if and only if the 4A: polynomiall
are dable.
Proof: Necessity is clear. Consider the family of polynomials which are gene-
rated from the first quadrant edges of the k parameter rectangles fiG.:
We claim that this family is stable. To show this fix parameters a2 through aA:
at either +1 or 0 and let al vary. This is precisely the situation encountered
in Lemma 5.1, where a2tP2(") + ... (1 - aA:)~A:(") is now lumped with tPo(")
and is labeled ~o(")' Therefore, one can write:
All these polynomials families will also be stable. Now since these polynomial
families form the set from which the four sides of the q(jw,OG) value set are
generated at each frequency, the entire family ~ is stable by the "zero exclusion
condition". 0
In this chapter we present a very powerful tool for robust V-stability, namely
the Finite Incl'IUionl Theorem [47, 49]. It is actually a Corollary of the Fi-
nite Nyquist Theorem presented in Chapter 3. Recall that the Finite Nyquist
Theorem guarantees that one can determine the V-stability of a lingle poly-
nomial, from only the approximate knowledge of the polynomial's phase at
finitely many points. For stability, at each frequency the polynomial evaluates
to a complex number and these points must lie in appropriately chosen sectors
in the complex plane. It is important to mention that the required number of
108 CHAPTER 6
frequencies is not high and could be as low as the degree of the polynomial.
Therefore, one should not interpret this as some form of "gridding". The Finite
Nyquist Theorem can be immediately extended to robust 1>-stability for poly-
nomial families. Specifically, if one is not dealing with a single polynomial but
rather a polynomial family, the Finite Inclusions Theorem guarantees robust
1>-stability if the polynomial value let lies in appropriately defined sectors at a
finite number of boundary points.
iP(", a)
Proof: For each a E 011, condition (6.4) implies that for allie, 4>(SI:, a) -# 0 and
that there exists a 81: E (CI:' dl:) such that arg4>(sl:' a) == 81:(mod211"). Further,
from the inequality
we see conditions (6.2), (6.3) imply that for all 1 $ Ie < m, 181;+1 - 81: I < 11" and
1211"n + 81 - 8m l < 11". Hence, for each a E 011 Theorem 3.3 implies ph a) is of
degree n and has all its zeros in int r. 0
The Finite Inclusions Theorem (FIT) basically says the following for the case
when 'D is the convex interior of some Jordan curve r: If one can find a finite
number of open sectors SI: which are consecutively "spaced" no more than 11"
radians apart and revolving a net 211"n radians about the origin of the complex
plane, and if we can find a counterclockwise sequence of points Sl: on the bo-
undary r such that the 4>(SI:, 011) value set lies in the corresponding sector SI:
for each Ie, then each polynomial in the family 4> is of degree n and is 'D-stable.
For polynomial families with polynomic parameter dependence FIT provides
only a sufficient condition for robust 'D-stability. It is clear that the Theorem
would apply to more general classes of polynomial families such as polynomials
with coefficients that are rational functions in a which are bounded in 011' Ho-
wever, the next theorem shows that in the important case of an affine family
of polynomials (the Qi(a) are affine expressions in a) the result is necessary as
well [47, 49].
(6.5)
(6.6)
110 CHAPTER 6
(6.7)
Proof: Sufficiency follows directly from Theorem 6.1. Suppose then that r/>(II, a)
is an affine polynomial family and that for all a E 0/1, r/>('" a) is a degree n
polynomial with all its leros in V, the interior of r. First note, r/>(", 0/1) is a
compact, convex set that revolves n times counterclockwise about the origin
without touching the origin as" traverses r (since for arbitrary a E 0/1, r/>(",a)
revolves in this fashion). By the above facts and the continuous deformation of
r/>(II,O/l) as " traverses r, we may define two continuous functions a, b: [0,1]-
R, such that for all tEl = [0, I], {re ill I r > 0, 8 E [a(t), b(t)]} is the smallest
sector containing r/>(r(t), 0/1)' and so, 0 ~ b(t) - a(t) < '11", a(I-) = a(O) + 2'11"n,
and b(I-) = b(O) + 2'11"n. Since a continuous function defined on a compact
set achieves its supremum and is uniformly continuous on that set, we see
that c = sup(b(t) - a(t» = maz (b(t) - a(t» < 'II" and that a and bare
uniformly continuous. Thus, for an arbitrary 0 < £ < ('II"-c)/2 we can choose a
sufficiently fine partition {tl;} of I such that conditions (6.5), (6.6) are satisfied
with CI; = a(t1:) - £ and dl; = b(tl;) + £, and by construction, condition (6.7) is
satisfied with 81; = r(tl;). 0
For the real affine polynomial family case and Hurwitz stability one can compare
the Finite Inclusions Theorem (FIT) and the Finite Frequency Test (FFTEST)
presented in Chapter 5. Both are necessary and sufficient conditions for robust
stability and both require that a "test" be performed at a finite number of
frequencies. In FFTEST, stability of a single polynomial must also be checked
in addition to the test at these frequencies. A major difference between them
is the fact that in FFTEST the frequencies where the test will be made can be
identified apriori from given data. They are shown to be the roots of specific
polynomials. At these frequencies the value set, a 2k convex parpolygon in
general, must be shown to exclude the origin. In FIT one does not have these
frequencies apriori but needs to find them along with the corresponding set of
sectors which must include the value sets. So, even though FIT is an analytical
result for robust stability in this sense it is not "constructive". From this, one
would be inclined to say that FFTEST is superior to FIT for robust analysis.
On the other hand in FFTEST one needs to find roots of polynomials which
in general must be done numerically. This inevitably involves approximations.
It would be interesting to explore the possibility of combining both results in
order to generate a new more computationally efficient test for robust stability.
The Finite Inclusions Theorem 111
It will also be seen in later chapters that FIT has a distinct advantage over FF-
TEST in robust synthesis. Whereas FIT possesses a structure which facilitates
the development of methods for controller synthesis, the same is not true for
FFTEST.
The same ideas used to derive Theorem 6.1 from Theorem 3.3 (that is, intr<r
ducing value sets and sectors) can be used to generate robustness results. For
example, the following two theorems on robust Hurwitz stability are derived
from Corollary 3.1 [49].
Theorem 6.3 Let ;(6, a) = Ei=o Qj(a)";, a E {lel1 n~ 0, Qj: {l4 --+ C,
Q,,: {l4 --+ R+. Then, for all a E Q, ;(6, a) hu all iu zero6 in the open left half
plane if there ezut m ~ 1 interva16 (Ci, di ) C R, 1 ~ lc ~ m, and real number6
WI ~ W2 ~ ••• ~ W m 6uch that
Many problems encountered in control are characterised by value sets that have
the complex conjugate symmetry ;( -jWj (la) =
[;(jWj {la)]*. This symmetry,
allows one to get away with using roughly half the number of sectors that would
otherwise be needed. One situation when this occurs is in the robust stability
determination of a unity feedback system where the plant is described by a
transfer function with real parameter uncertainty. The resulting closed loop
characteristic polynomial will have coefficients that are real expressions of the
parameter a [47, 49].
Theorem 6.4 Let ;(6, a) = Ei=o Qj(a)";, a E {la, n~ 0, Qj: {la --+ C,
Q,,: {la --+ R+, and for all W E R, ;( -jw, (la) = [;(jw, {la)]*. Then, for all
a E Q, ;( 6 , a) hu all iu zero6 in the open left half plane if there ezut m ~ 1
interva16 (Ci, d i ) C R, 1 ~ Ie ~ m, and real number6 0 ~ WI ~ W2 ~ ... ~ W m
luch that
'If 'If
--2 <
-
CI < d l <-
- 2
112 CHAPTER 6
~n ~n
--~<Cm<dm<-+~
2 - - 2
V 1 ~ Ie < m max{dl:+l- CI:, dl: - CI:+l} ~ ~
V 1 ~ Ie ~ m ~(jWI:, no) c 51: = {re'" 6 Ir > 0, 8 E (CI:, dl:)}
Several versions of the Finite Inclusions Theorem have been presented and their
usefulness in robust 1'-stability of polynomial families with parameter uncer-
tainty is immediately apparent. The Finite Inclusions Theorem requires that
a finite number of points on the boundary of l' be found such that the corre-
sponding polynomial value sets lie in appropriate sedors. One alternative in
implementing the result is to choose a set of frequencies and plot the correspon-
ding value sets. One then seeks to find sectors that satisfy the conditions of
the theorem that "fit' the value sets. One may not be able to conclude robust
1'-stability from this data and additional frequencies and sectors may be requi-
red. Another alternative is to pick a set of sedors that satisfy the theorem and
then choose frequencies that "fit' value sets in them. Again, one may need to
modify the initial choice with additional frequencies and sectors. Clearly, this
is an iterative process and at the present time no definitive procedure exists for
choosing frequencies and sectors.
To apply FIT one must also be concerned with the computation of the value
sets at each required point. For polynomial families with polynomic parameter
dependency this is in general is not an easy task. The boundary of each value
set will not come from the image of the edges or boundary of no' For affine
parameter dependence since the value sets will be convex polygons we would
just need the value set vertices in order to determine sector inclusion. This
corresponds to a dramatic reduction in complexity. In the multiaffine parameter
case one can employ the Mapping Theorem to overbound each value set and
again work with value set vertices [46].
Consider the situation where the stability of the following polynomial family is
to be determined:
in an affine manner, it is known from Chapter 4 that the value sets at each
frequency will be parallelograms. From these plots it can be seen that Theorem
6.4 holds with m = 2 and the Wt, Ct, and d~" of Table 6.1. Therefore, the
family is stable. Data from only two frequencies are shown. One should not
get the impression that an arbitrary choice of two frequencies would work. As
mentioned the procedure is iterative. 0
400
300
200
q,(j7, 0 0 ) q,(j2.6,O o ) _
~-----------
----
100
~
.51 0
f
.5 ---
.... " -
-100
-200
·300
--- -
-400 L....-~~~_~..L-~~~~~-'
-500 400 -300 -200 -100 0 100 200 300 400 500
real
In Example 6.1 the polynomial family had parameter uncertainty which appea-
red in an affine manner in the coefficients. It is interesting to note that in this
case one could determine the location of the value sets with respect to the sec-
tors by plotting only value set vertices and not the entire value set. This is due
to the fact that for affine uncertainty a value set will lie in a sector if and only
if its vertices lie in the sector. This is generally not true for more complicated
uncertainty structures. In the next example parameter uncertainty appears in
a multiaffine manner. The value sets at each frequency will not in general be
polygons but more complicated shapes [47, 49].
where a E 0 0 =
[-1,1] x [-1,1] and a =
(al' a2)' Figure 6.2 is a plot of
~(jWi 0 0 ) for two values of w. Note that the value sets are not polygons. From
these plots it can be seen that Theorem 6.4 holds with m = 2 and the WI;, Cl,
and d l 6 of Table 6.2. 0
3..--------------,....------.
X
I
I
I
2 I
~(j, 0 0 ) ~(j.• 7, 0 0 )
~
:; 1 :< ",
I ",,'
..
'.",~
....., 0 1- , ' . -: ••••••
.=
~
-1
" ",
"" ... " . . . :
-3 +---l~-+-+---lf---+-+---lf---+--l
-7 -6 -5 -4 -3 -2 -1 0 2
real
k Wl Cl dl
1 .47 27(/17 7(/2
2 1 7(/2 197(/17
As in the case of FIT a special version of this theorem can be stated for the
case when value sets possess complex conjugate symmetry. As we saw earlier
the required number of sectors is roughly halved.
where P(6, B), BE Oell is a stable family ofrational transfer functions and W
is a stable, minimum phase, rational "weighting" transfer function. Recall that
a transfer function is &table (minimum phase), if all its poles (zeros) are in the
open left half plane.
FIT is helpful here in that it provides a way to verify that a given number 'Y+
is indeed an upper bound to 'Y. For example, suppose in (6.10) that W(s) = 1
and
(6.11)
where G1 E [4,6] and G2 E [3,5]. From the plot of IP(jw, a)1 vs. w for a =
(G1G2) E {4,6} X {3,5} we see 'Y- = 20.4 is a lower bound for 'Y.
lL...---,-~~-'-""""'~-_-"'---'---'-~~'-'-:'
.1 1 10
frequency
and only if "y+ is a strict upper bound on "y. So, "y < "y+ if and only if the
dosed loop characteristic polynomial ;(6,a) has all its roots in the open left
T
half plane for all a E 011'
l' ,'_°_-·13
P(.,a)
For "y+ = 25 this is indeed the case as shown by Theorem 6.4 (with m = I,
"'1 = 1.75, =0, and b =
01 1 1r /2), multiaffine uncertainty, and Figure 6.5. The
symbol EztO Il is used to denote the extreme points of the set 011' Recall that
an extreme point of some convex set q is a point which is not contained in any
open line segment contained in q. Thus, with complete certainty we can say
20.4 $ "y < 25.
10 , . . . . - - - - - - - - - - - - - - - - . . . ,
·i.s o ----------------------
-5 +----+---+----+---+----i
-10 -5 o 5 10 IS
real
(6.12)
with 4>0(") monic of degree n, 4>1(") of degree less than n and a1 E [-1,1].
We presented a sufficient condition for the robust stability by introducing the
rectangular value set overbound. Assume for simplicity that 4>0(,,),4>1(") have
real coefficients. The result was proved by overbounding the polynomial family
(6.12), by the polynomial family:
(6.13)
In Lemma 5.1 we proved that the polynomial family (6.13) is stable if and only
if the four polynomials
are stable. This is a necessary and sufficient condition for robust stability of
family (6.13). Since parameter uncertainty appears in an affine manner one can
also obtain such a characterization from FIT. The following result clarifies this
relationship and at the same time makes a connection with the Hermite-Biehler
Theorem.
120 CHAPTER 6
Lemma 6.1 Let the four polynomia18 tPi(6) 1:::; i :::; 4, be monic and degree
n. Define n non-overlapping, cOn6ecutive, 90° 6ector, in the complez plane a6:
The four tPi(6) polynomia18 in (6.1,j) are ,table if and only if there ezut po6itive
frequencie6 Wl < W2 < ... < Wn 6uch that tPi(jWIc) E Sic for 1:::; i ~ 4 and each
1:, 1:::; Ie ~ n.
Proof:
Sufficiency follows directly from Theorem 6.4.
We shall prove necessity for n even and m = n/2. The case when n is
odd follows in a similar way. Assume then that tPl' tP2' tP3' tP4 are stable. By
construction we have:
Let Ul < U2 < U3 ••• < Urn be the positive real roots of rl (w) and u~ < u~ ••• u:n
be those of r2(w), Let Vl < V2 ... < Vm-l be the positive real roots of tt(w)
and v~ < v~ ... < v:n_l those oft2(w), This fact as well as the ordering below
follows from the Hermite-Biehler theorem. To simplify notation assume that
Ul :::; U~,U2 :::; u~, ... ,Urn ~ u:n and Vl ~ V~,V2 :::; v~, ... ,Vm-l ~ v:n-l' We
now have:
Ul < Vl < U2 < V2 ... Vm-l < Urn, u~ < v~ < u~ < v~ ... V:n_l < u:n
(6.15)
Vm-l < u mI
(6.16)
The importance of this result will become evident in Chapter 10, when we will
use the Simultaneous Polynomial Stability result for robust synthesis. One can
easily envision how this will be accomplished. Think of the family in (6.12)
as the closed loop characteristic polynomial of some feedback loop where the
plant contains the single parameter 41 and the controller parameters affect the
coefficients of tPo(") and tPl(")' It is clear that the feedback loop will be robustly
stabilised if a controller can be found that 6tabilize6 the four 1/1,(") polynomials.
In this context Lemma 6.1 is helpful in two ways: First, it provides a way for
designing such controllers by using the Finite Inclusions Theorem. Second, it
shows that this technique may be more conservative than FIT as value sets are
forced to be rectangles and sectors are only 90 0 wide, whereas sectors in FIT
can be up to 180 0 and no such requirement is imposed on the value sets!
After presenting some background material and pertinent definitions the first
six chapters of the book dealt with control system analysis for systems with
parameter uncertainty. We stated a number of important control problems and
proposed solution methods by formulating them as robust polynomial stability
problems. This motivated the development of several tools for robust analysis:
the Finite Frequency Test, the Finite Matched Phase Theorem, the Simulta-
neous Polynomial Stability result, and the Finite Inclusions Theorem. These
results complement existing literature on robust analysis for systems with para-
meter uncertainty. However, we do not view the development ofrobust analysis
tools as our main objective but rather the "means" towards the development
of robust synthesis methods.
Our task in the second part of the book is to demonstrate that the Finite
Inclusions Theorem and the Simultaneous Polynomial Stability result are par-
ticularly suited for the development of robust synthesis methods for systems
122 CHAPTER 6
with parameter uncertainty. We will show first how the Finite Inclusions Theo-
rem provides the foundation for powerful new robust synthesis methods. We
call these methods FIT Iynthelil and use them to solve control problems posed
earlier in the book. A key element in the design approach is that the formu-
lation leads to iterative robust synthesis algorithms where at each iteration
a controller is designed by solving a set of linear inequalitiel. We will then
show that the Simultaneous Polynomial Stability theorem can be used to ge-
nerate analytical results for robust synthesis. In a typical result, a controller is
shown to satisfy some specific robust performance objective for a plant family
with parameter uncertainty if the controller simultaneously stabilizes a finite
number of polynomials. These design methods are referred to as SimultaneoUl
Polynomial Stabilization.
7
FIT BASED V-STABILIZATION
[1] and Quantitative Feedback Theory [43] have been employed for systems
with parameter uncertainty. More recently Linear Quadratic methods [54] and
pi H oo techniques [36] were developed which are particularly helpful in multi-
variable system design. All of these design techniques have their foundation on
results from analysis and it is a fact that some analysis results are better suited
for design than others. A good example is provided by the Routh-Hurwitz test
for stability and the Nyquist Theorem. They both are good analysis tools but
the Nyquist Theorem has proved much more valuable in developing methods
for controller design.
Our major thrust in this book is the development of new tools for robust con-
trol synthesis. During the eighties there was a great deal of work that was done
on systems with parameter uncertainty. Most of the emphasis was on robust
analysis issues and [6] contains a comprehensive exposition of many of these
results. In view of our emphasis on synthesis it is rather disappointing that
these efforts did not lead to the development of comprehensive methods for
robust design.
We would like to suggest that the Finite Nyquist Theorem and the Finite
Inclusions Theorem do provide the foundation for the development of methods
for controller synthesis for systems with parameter uncertainty. In contrast to
several of the existing methods the technique works directly with polynomials
and polynomial value sets. Essentially, it is a method of robust V-stabilization,
but has much broader applicability as other performance specifications can
be recast in this setting. The method can best be described as a rob1Utneu
enhancement technique in that an initial controller needs to be provided. The
initial controller can be constructed by employing one's favorite technique and
it mayor may not have any robustness properties apriori. Then in an iterative
manner the region ofrobustness is enlarged with the hope that it will encompass
the given region. Controllers are described in terms of a finite number of
parameters and at each iteration a new controller is designed by solving a linear
let of inequ.alitiel in these parameters. This provides a great computational
advantage. Even though graphics can be used to display pertinent data, the
technique is analytical and thus can be au.tomated. Main elements of the theory
behind the methodology have been developed as well as specific algorithms
for implementation. Some of these have been programmed in the MATLAB
environment and used to perform the computations in the examples that follow.
FIT based 'V-Stabilization 125
--
r +
-
+
e
C(s) P(s, a)
y
_ n,(s, a)
P( s, a ) - dp(s, a) (7.1)
where dp(s, a) is monic of degree n and the parameter a takes values in some
hypercube 0/1. We assume that the uncertain parameters enter in a multiaffine
manner in the coefficients of n,(s, a), dp(s, a). The controller is proper of order
q and given as:
The controller parameters are grouped in the vector x = (:1:1, :1:2,· •• , :l:2q+l)
E R d where d = 2q + 1 and the closed loop characteristic polynomial is given
by:
To the nontrivial problem of computing a set of points 81: and sectors 51: for
which a solution x exists, we propose the following sequential procedure: The
basic idea is to design an initial controller x(l) by employing some design tech-
nique, such as pole placement, classical loop shaping, jJ/ Hoo synthesis, etc..
This initial controller may robustly 'D-stabilise for a "small" region of uncer-
tainty, say {l~l) C (lo. This controller is then iteratively improved upon using
FIT. This can be done by choosing a new set of points 81: that "center" the
value sets in their corresponding sectors, or by choosing new sectors. When
(if) a controller x(i) is obtained that guarantees 'D-stability over the entire set
(la, the procedure terminates with x = x(i) and the desired set of points 81:
and sectors 51: having been found.
It should be emphasized that at each step of the procedure (with points 81: and
sectors 51: fixed), the solution of the linear inequalities is a convez problem.
However, the entire procedure is not a convex problem and there should be no
expectation of a "globally optimal" solution. Clearly, the solution will depend
on the "initial" controller and therefore care should be taken at this step. One
should use available knowledge and insight to guide the initial design. As will
seen in the examples that follow, a variety of such designs have been used such
as HOCJ' loopshaping and root locus.
FIT based V-Stabilization 127
If this algorithm terminates, then a robust V-stabilizer would have been found
and the controller parameters would be x = x(j). In order to guard against
unnecessarily long computations in the event that the "growth" of the uncer-
tainty hypercube is very slow, one may introduce some early stopping condi-
tions. One possible variation of this algorithm is to use a single set of sectors
= =
throughout (Le., si;) Sl: and m(;) m for all j).
[46, 49]. The system is comprised of two masses and a spring conneded as in
Figure 7.2. A control force u(t) ads on mass m1, and the position :112(t) of
mass m2 is measured.
(7.4)
=
where a (k, m1, ffl2) and k, mtt m2 take values in the set 0 .. = {(k, m1, m2)1
.4 ~ k ~ 2, .8 ~ m1 ~ 3, .8 ~ m2 ~ 2}.
The specific problem we will solve is the following: design a controller for this
mass-spring-mass system that places the closed loop poles of the system in the
region 'D = {" E CIRe(,,) < -.2} for all a EO... The position :112 is the con-
troller's input, and the force u(t) is the controller's only output.
Using the standard change ofvariable, v = ,,+.2, the above robust 'D-stabilization
problem can be turned into a robust stabilisation problem. In the following im-
plementation of the FIT algorithm we execute the steps in a "manual" fashion.
In other words the choice of frequencies that center value sets and the en-
largement of the uncertainty region are done by inspection. This makes the
exposition clearer. Later on, we will demonstrate implementations of this al-
gorithm where the entire procedure is carried out "automatically" in software.
FIT based V-Stabilization 129
STEP 1
Our first priority is to design an initial controller. In this case we choose to use
the H oo controller designed for this problem in [20]. This controller is given by:
It can be easily checked that this controller does not meet our robust 'D-
stabilization objective because for a = (04,3,2) the closed loop characteristic
polynomial does not have all its roots in 'D. As a matter of fact the polynomial
is not even stable. However, this controller does guarantee robust 'D-stability
for the "smaller" set, with k taking values in the range .5 ~ k ~ 2, and
m1 = ~ = 1. In view of this initial design we will also use a 4th order proper
controller for the robustness enhancement:
Our design has to find a controller that makes this polynomial robustly stable
for the entire region of interest.
STEP 2
Our next priority is to find and sectors SI: that satisfy the conditions of FIT
and points 61:, which in this case are "frequencies" WI:, that place the value
sets ,p(WI:, oil), XCl») in SI:. In view of the conjugate symmetry present in this
example we can use Theorem 604 to reduce the number of sectors to roughly
130 CHAPTER 7
half. The sectors are chosen as follows: Sector 51 has boundaries C1 = -11:/4,
d1 = 11:/4, and 5" C" = =
1I:(k-2}/4, d" 1I:(k-2}/4+311:/4, for 2 ~ k ~ m 15. =
Note that with the exception of the first one the sectors are 135° wide and that
they overlap. The frequencies that roughly center the corresponding value sets
in these sectors are given in Column 1 of table 7.1. In addition to this data
Column 1 gives the uncertainty set oil) = [.5,2.0] x [1,1] x [1,1] as well as the
controller coefficients x( 1) •
STEP 3
STEP 4
The set of linear inequalities is now solved to obtain a new controller z(Z) that
guarantees robust stability for O~2) = [.39,2.01] x [.79,2] x [.79, 1.5]. In this
example the projection method was used [49]. The parameters of such a con-
troller and the frequencies that center the new value sets in their corresponding
sectors are given in Column 2 of Table 7.1. Clearly, the size of O~2) may be too
large in this step and a solution may not be found. Some trial and error might
be required in identifying an appropriate uncertainty set.
STEP 5
The set O~2) does not encompass the required uncertainty range so the enlar-
gement procedure must be repeated followed by the solution of a new set of
linear inequalities. The set oi3
) = [.39,2.01] x [.79,3.01] x [.79,2.01] is shown
in Column 3 of Table 7.1, along with the new controller parameters and the
corresponding frequencies. It is evident that in this example the FIT algorithm
was successful in finding a robustly stabilizing controller for the required region
of uncertainty.
x) _
R C ('"
dc(",x) -
901,250,,4 + 29,690,000,,3 + 34,645,000,,2 + 17,090,000" + 5, 067, 500
,,4 + 88.058,,3 + 4502.7,,2 + 278,440" + 1,619,300
132 CHAPTER 7
where m is the bus's "virtual mass" in kilograms and v its velocity in meters
per second. A complete explanation of how this transfer function is developed
from the nonlinear dynamics is given in [1]. It is also important to note that
unlike the previous example, parameter uncertainty enters the numerator and
denominator of P(6, m, v) in a polynomic manner.
Assume that this transfer function is placed in the unity feedback configuration
of Figure 7.1. The goal here is to design a controller C(6) = ~ to place all
closed loop poles in the hyperbolic region {cr + iwlcr < -.35Jl + (w/1.75)2}
for all (m, v) E 0 4 = [9950,32000] x [3,20]. The above hyperbolic region was
chosen to assure passenger comfort (see [1]), and to further this aim and simplify
implementation we would also like both the loop's bandwidth and controller's
order to be as low as practically possible. We impose this requirement by
insisting that the closed loop poles not be more than a specified distance from
the origin of the complex plane. In particular the region 1) is defined as:
The three goals of robust pole placement (for polynomic parameter uncer-
tainty), low loop bandwidth, and low controller order make this a very challen-
ging robust synthesis problem.
The design carried out in [1] began with the realization that for fixed m and v
Root Locus arguments show that a suitable controller has the structure:
FIT based V-Stabilization 133
(7.8)
This is so because the system's two zeros lie in the hyperbolic region. See-
ing this, the author of [1] rightly conjectured that a rolnut controller of the
same form could be found. The design was then carried out using Parame-
ter Space Methods. In particular, the parameters Wo = 100 and k l = 2.5
were fixed, and via a graphical method the set of all k 2 and k 3 's was de-
termined that robustly placed the closed loop poles in the above region for
(m, v) E EztO o = {9950, 32000} x {3,20} (recall that EztO o denotes the ez-
treme point" or "vertices" of ( 0 ), After choosing the point k 2 =.7 and k3 = .15
from this set, which turns out to be nonempty, he further adjusted the para-
meters Wo and k l to lower the controller's bandwidth. The final controller he
obtained was:
.6 + .78 + .158
2
C( ) (7.9)
$ = (1 + 8/25 + 82 /25 2 )(1 + 8/25)
which guaranteed robustness for only the "vertex plants" (m, v) E EztO o CO o '
After checking, it so happens that this controller solves the original problem
(i.e., it is robust for all (m, v) E ( 0 ), This was verified graphically via a plot
of stability boundaries in parameter space (see [1] for details).
One indication that solving this problem through FIT synthesis may be diffi-
cult is that for the controller in (7.9) the value set of the loop's characteristic
polynomial
t/J(8j m, v) = (.6 + .78 + .158 2 )(607876mv 2 82 + 387· 109 v8 + 48· 10 9 v 2 )
+(1 + $/25 + 82 /25 2 )(1 + 8/25)8 3 (m 2 v 2 $2 + 106 mv8 + 16630mv2 + 269.109 )
evaluated at a point on the stability region's boundary can exceed 1800 in width
(see Figure 7.3). Hence, FIT cannot be directly used to verify this controller
134 CHAPTER 7
01-···················· ; ; '''~~::-.: ; -I
1.0.5 1- ; ·_········;··········11"1-""1"-+-.1=-·"
«I
.5
·11-···_·······_··········_····+··············.. ········~'Y'V:
_2L~~~::E:~--L_----l_~
.0.5 0 0.5 1.5 2
real x10 13
Figure T.I Plot oCthe Value Set 1/>( -0.7471 +;3.3, n..)
design a controller that will place the closed loop poles in 'V for all a E 00 •
Clearly, this plant family encompasses the original plant family. The region 'V
is given in Figure 7.4.
70 "T'""-------:,-T"'""-------...,
, ,
'- - ...
~
"60
.§
0
l.:f""\
+-------Ji;i;i§~~~;;:;+----~--___I
~r:::::::::~::::j:::~:~::~:r:::~::~:~:~I:~; :
STEP 1
In view of the Root Locus arguments mentioned earlier it makes sense to choose
a controller with the structure:
(7.10)
(7.11)
The initial controller places the closed loop poles of the "nominal" system
= =
m 20975, v 11.5 in the desired region. The closed loop poles of the nominal
design are shown in Figure 7.5. The coefficients of this controller are: :1:1 =
15630, Z2 = 1250, Z3 = 50, Z4 =
24620, Z5 =
18850, Z6 = 4546. This controller
guarantees 'D stability for the "modified" system for O~I) = [20975,20975] x
[20975,20975] x [11.5,11.5) x [11.5,11.5). This information is also displayed in
Column 1 of Table 7.2.
55 -r------------,
}.----
~
.~ 0
.5
-55
,----
+-----+------1
-55 o 55
real
Figure T.5 Clolcd Loop Poles For Nominal DClign
FIT based V-Stabilization 137
STEP 2
(7.12)
where -47 ~ Ie ~ 48. In this implementation these sectors will remain fixed
throughout the procedure. By the Finite Inclusions Theorem and the above
observations x will be a robust controller if there exists a counterclockwise
sequence of points (SA:) along the boundary of V such that ifJ(SA:! Ezt0 4 , x) C SA:
for all Ie. This condition is easily checked since ifJ(s, Ezt0 4 , x) is a finite set,
which deforms continuously as S traverses V's boundary. Further, because of
the complex conjugate symmetries ifJ(so, .) = ifJ(S, .)*, S-A: = Si, and V = V* ,
we need only check roughly half of these sectors. Namely, if (SA:), 0 ~ Ie ~ 48,
lie on or above the real axis and satisfy the above inclusions, all the inclusions
will be satisfied upon setting SA: = s:'l: for -47 ~ Ie < O. We therefore compute
a counterclockwise sequence of points (s~» along the upper half of V such
that ifJ(s~),EztO~),x(j» is roughly centered (angularly) within SA: for each
o ~ Ie ~ 48. Some of these points are depicted in Figure 7.4.
STEP 3
Next we choose a set O~+1) :J O~).
138 CHAPTER 7
STEP 4
We now compute a new vector of controller coefficients x(i+l) such that:
(7.13)
It may be that the above system of inequalities has no solution. In this case,
one must go back to STEP :I and try a smaller set fi!1+1).
STEP 5
stabilising. In fact, it turns out that this controller is V-stabilising for all
(m, v) E 0 0 = [9950,32000] X [3,20]. This was verified by partitioning 0 0 into
several smaller sets and then by using FIT to verify V-stability over each set
individually.
k si5
) k ,;p)
0 -.3500+jO 25 -8.251 +j41.22
1 -.3526+j.2130 26 -10.39+j48.91
2 -.3663+j.50411 27 -12.78+j48.34
3 -.3915+j.8769 28 -15.08+j47.67
4 -,4197+j1.158 29 -17.19+j46.95
5 -,4523+j1.432 30 -19.54+j46.02
6 -,4874+j1.696 31 -22.32+j44.74
7 -.5179+j1.909 32 -25.00+j43.30
8 -.5480+j2.108 33 -27,43+j41.80
9 -.5806+j2.317 34 -29.98+j40.01
10 -.6236+j2.581 35 -32,40+j38.08
11 -.6754+j2.888 36 -34.70+j36.00
12 -.7692+j3.425 37 -36.76+j33.90
13 -.9493+j4.412 38 -39.01+j31.28
14 -1.269+j6.097 39 -41.04+j28.56
15 -1.731+j8,477 40 -42.79+j25.87
16 -2.309+jl1.41 41 -44.55+j22.70
17 -2.925+jI4.52 42 -45.88+jI9.87
18 -3.755+jI8.69 43 -47.19+jI6.52
19 -4.850+j24.19 44 -48.10+jI3.66
20 -5.799+j28.94 45 -48.94+jl0.22
21 -6.861+j34.26 46 -49.55+j6.688
22 -7.625+j38.09 47 -49.88+j3,409
23 -7.846+j39.19 48 -50.00+jO
24 -7.932+j39.62
Plots of the closed loop poles for both the design carried out in [1] and our
controller over (m, v) E {9950, 20975, 32000} x {3, 11.5,20} are given in Figures
7.6 and 7.7, respectively. These figures show the controller from [1] is slightly
better than ours at keeping the system's poles close in to the origin and hence,
may provide passengers with a slightly smoother ride by better attenuating high
frequency disturbances. Our controller, however, is nearly as good and was ob-
tained from a general synthesis procedure. While our FIT synthesis procedure
140 CHAPTER 7
55 T"'------------,
.'-~r::~.~ - -- .........."
,,
,
\
\
,
\
\
~ I
.S 0
,,
I
~
.S ,
,, ,
,
,,
-55 +-------+------1
-55 o 55
real
55
..:~::r::J - - - -",
~
'i 0
.S
'j----
-55
-55 o 55
real
P( ) _ .:..("_+_1...:.)..:..("_+_1_0~)+_a~I(=-=-5._5"_+---:-:-7)_+_a~2...:..(,,_2_+_0_.5_"_+~2) (7.14)
", a - (" + 2)(" + 4)(" - 1)
where a = (al,a2) lies in thereetangle 0 .. {(al,a2)I-Q < al < Q,=
-Q < a2 < Q}. The positive parameter Q regulates the "sise" of the rec-
tangle. The plant is in the feedback configuration of Figure 7.1. The objective
is to find a robustly stabilizing controller for as large a value of Q as possi-
ble. An additional requirement is to try and keep the order of the controller
as low as possible. Note that unlike the previous two examples, the family
of plants can contain non-minimum ph/Ue systems for values of Q larger than
= =
unity (e.g. al -I,a2 -1.1), so that a "high gain" solution is not possible.
Furthermore, the family is unstable.
This is a "stability margin optimisation" problem and we know that such pro-
blems have been considered in the past and solved in appropriate settings. Re-
call that [51] provides a complete solution in the case when P(", al) = aIPO(")'
It provides a constructive solution for the controller that robustly stabilises
such a system in the feedback configuration of Figure 7.1 for the largest range
of values of al. Several "single parameter" problems can be cast into such
a formulation, however our example does not fall in this category because of
the presence of two uncertain parameters. The authors in [36] consider the
problem of multiplicative perturbations and norm bounded uncertainty. Let
the plant family be given by P e (,,) = P(")(1 + ~W2("» where ~ is any sta-
142 CHAPTER 7
ble proper transfer function with 1I~lIoo < E, P(s) a nominal transfer function
and W2 (,) a stable and proper weighting function. One seeks to find E,."the
least upper bound on E such that a robustly stabilising controller can be found.
The solution to this problem comes from a "model matching problem" and
in [36] it is solved using Nevanlina-Pick Interpolation. When compared with
these "analytical" solutions our approach, even though it is based on a sound
theoretical foundation, is more "algorithmic". Furthermore, since the overall
problem is not convex we cannot make any statements about global optimality
of the solution.
The algorithm requires that an initial controller be supplied. In view of the fact
that the plant family can be expressed in a multiplicative perturbation form,
it makes good sense to exploit the existing theory. Specifically, we will design
an H oo controller assuming a multiplicative, norm bounded uncertainty struc-
ture. In order to employ this methodology, our first priority will be to find an
appropriate weighting function W 2 (,), for the parameter uncertainty:
It is reasonable to expect that the better the "fit" of the overbound W2 the
less conservative would be the norm bounded approximation. It then follows
that the higher the order of the transfer function W2 , the better becomes the
approximation. In turn, the higher the order of W2 the higher the order of the
controller. Figure 7.8 shows two second order W 2 (,)'s called the "suboptimal"
101 ........-.....--..,.-..,.--r-........T"T"T--....--.---r-.........................----,..---.--.-.........,..,....,
':.:::::::::::::::::: . . . .. . ::t:):: ~::j::;:;.. . + ~ ~ +.. +.+. +.:
.....•.••••.••c.••.•.... c••••.•e.••..•.••(.•. i..(.. c•.•••.•••••••••.••·;...•.••.•; i (. i j. ••; •• i.(················(.··.·····.(.·.·..(.·.. ·j.,.(. •• i...;.(
..............1 i. i i j i.. ~.. ~ ..i ~ .i i i i.. l..i..i.i l i l i l.. ~..l.;
..............1. ~ i i ~ i.. ~..i..i i i ; 1. ;.. 1..i..i.1. i ~ l i 1 i...~.;
..............L L L L..j LLLL L L L..L.Ll..LLl. L ~ i. L..L.Ll.;
. . . . . . . . . . . . t.:.: j:.' ~:.' .. ·!:.'· ..1..·.. 1+ · · . I:' ;:.' L.. t·++H-f.... · ·.. l l....··.. ++H·i
: : : :::: .....
.. ··········r······T·····r....1.....:.. ~ . ~..: ·....T· ·+···r···~··~··~··:·~···.. ······f···..t····~···~··~··~ ;
1 ! : !·1. . !Ill i
1.1336608,,2 + 6.8857303" + 9
(7.17)
(" + 1)(" + 10)
These were developed in [53] where more details of the procedure can also
be found. Using these bounds one can now employ the Boo methodology to
compute the controller that maximises the value of Q for which robust stability
can be guaranteed. Using Nevanlina-Pick Interpolation [36] the Boo controller
computed for the bound (7.16) was found to be:
which guaranteed robust stability for Q as large as Q:'IIZ = 1.273. Observe that
this controller is improper! In this case an "easy" way of obtaining a proper
144 CHAPTER 7
For this design we chose to add the factor (17100~. + I) and use:
as the initial controller for the FIT software package. The FIT software routine
generated the controller:
order W Z(6). For the one labelled "optimal" in (7.17) one can show [53] that
again an improper controller gives a:GZ
= 1.2926, a number smaller than a~~~.
Better results could of course be obtained with higher order WZ(6)'S but these
would inevitably lead to higher order controller designs. Since low controller
order is a design requirement this would require some form of "reduction" after
the design. This example demonstrated that th.e combination of H00 and FIT
synthesis became a powerful tool for controller synthesis.
Justifiably, the problem of simultaneous plant family stabilization has not re-
ceived as much attention in the literature [70]. Its importance is undeniable.
Recall that in Chapter 4 the problem of asymptotic tracking was posed as a ro-
bust polynomial stabilization problem. SpecificallYI a controller C(6) =
~;f:~
was needed to be found that robustly stabilizes the polynomial family:
(7.21)
In other words the controller would need to robustly stabilize the "plant family":
(7.22)
Suppose now that this same controller needs to guarantee asymptotic noise
attenuation as discussed in Chapter 4. This requirement can be expressed as a
robust stabilization requirement for the polynomial family:
(7.23)
146 CHAPTER 7
In other words the controller must robustly stabilise the "plant family":
np(6)(r2~a2n.2(6) + d.2(6))
dp(6)d.2(6)
(7.24)
where P(6) is some "nominal" system transfer function, W(6) is some given mi-
nimum phase stable transfer function and Ii. is any arbitrary stable and proper
transfer function with IIi. (jw) I :5 I, for all wE [0,00). The transfer function
W(6) would need to be appropriately shaped. Then a necessary and sufficient
condition for robust stability is IW(jw)T(jw) I < 1 for all w E [0,00), where
T is the nominal complementary sensitivity. This elegant characterization of
robust stability allows for controller synthesis as a solution to a Nevanlina-Pick
interpolation problem. This solution may lead to conservative designs because
of excessive overbounding.
In Chapter 6 SSFIT was presented, a result which is very well suited for the
development of algorithms for simultaneous plant family stabilization. We will
concentrate on the development of such an algorithm for plant families with
parameter uncertainty. Consider the following canonical stabilization problem:
We are given the unity feedback system
l_ e
_·I'---_C-=-(6)-====,I=p._,-"(6,a:) y
Figure T.8 Unity Feedb.c:k Configuration, Parametric Uncertainty
where Pi(6, a), 1 ::; i :5l is a family of strictly proper plants given by:
where all the d,n(6, a) are of degree 7ii and the parameters a take values in some
hypercube 0 0 and with positive leading coefficients. Uncertain parameters
enter polynomial coefficients in a multiaffine manner. For now we will also
assume the following:
148 CHAPTER 7
Assumption
We can construct a controller that simultaneously stabilises the finite number
nominal of plants, P1 (", 0), P3 (", 0), ... Pt(", 0).
For the two motivating paradigms of simultaneous plant stabilisation this as-
sumption is true. The development of the more general algorithm will be
considered at the end of the Section. We use a proper, order q controller given
by:
Controller parameters are grouped in the vector x = (:1:1, :1:3,' .. , :l:2q+d and
the closed loop characteristic polynomials are given by:
(7.27)
The ~i('" a) have coefficients that are multiaffine in a and affine in x. The
design problem is now to find (if possible) a controller of order q that stabilises
~'('" a, x) for all 1 ~ i ~ l and values of a E 011'
employing some design technique, such as pole placement, classical loop sha-
ping, lSI H oo synthesis, etc.. This initial controller may simultaneously stabwe
for a "small" region of uncertainty, say oil) COo • This controller is then ite-
ratively improved upon using SSFIT. This can be done by choosing a new set
of frequencies Wil: that "center" the individual value sets ;i(jW, O~j>, xCi») in
their corresponding sectors, or by choosing new sectors. When (if) a controller
x(j) is obtained that simultaneously stabwes over the entire set 0 0 , the pro-
cedure terminates with x = xCi) and the desired set of points Wil: and sectors
Sil: having been found.
With frequencies Wil: and sectors Sil: fixed, the solution of the linear inequalities
is a convez problem. However, as in the case of the FIT algorithm, the overall
problem is not convex. The algorithm we propose below is called the SSFIT
Synthe,iI Algorithm.
STEP 1: Let X(l) E ad and oil) C 0 0 be such that ;i(" oil), x(l»)
are stable and set j:=l.
STEP 2: Determine m~j)1_ II:' 1 <
> 1 sectors ~;) _k_ < m~j)
II 1 <_i_ <l
and frequencies W!~) along the iw axis such that
;i(jw!~),EztO~j>,xCi»)C S!~). By FIT, ;i("o~),x(i») are
stable. Each w!!)should roughly center (angularly) the set
;i (iw!~), EztO~), xCi») in S!t).
STEP 3: Choose a slightly larger set O~+l) :::> O~).
STEP 4: Compute a new vector of controller parameters xU+l) such
that ;i(jW!~), Ezt O~+l), x Ci + I ») C s~t) for allk and i.
Note this is equivalent to solving a system of linear
inequalities in x(i+l). If no solutions exist to this system of
inequalities, return to Step 3 and choose a smaller O~+l).
STEP 5: Let i := i + 1, and if O~) :::> 0 0 , stop; otherwise, go to Step 2.
2Or----r----r----,r----r-----r-----,--.----,
10 : ; i : ; ( ; .
··· .. .
. . . .
· .
.
. ... ..
. .
...
.
.
.
.
.
.
.
.
.
.
5 ; : :....... : :
. ...~ ...
:J 0 , .:'/ : ...
,
)(
.
:
,
; : , .
.. .
-5 ~ : : : : ~ ~ .
: : : : : : :
· : ~ : : : :
....
; :
• ••• ~ .......•.. ;, .. , ••••...:..•••••.••. i
: : :
., .: .•••••.... <....
·10
-15 : :
.
······i·········
..
: ·f··· ~ :. .
: :. .:
·~--.1L,-5- -....,.10,..---.~5-~0'---75----:1"::-0---!:15~---::!20
real
One would be inclined to say that by "breaking up" the plant family into a
number of "sub-families" and then using SSFIT one can design a controller
that guarantees robust stability for a larger uncertainty set.
FIT based V-Stabilization 151
Consider the plant family in a unity feedback configuration of Figure 7.1 with
transfer function:
P(..,a) = np(..,a) =
dp(.., a)
(18a1aZ - 12.5a2 - 12.5a1 + 3.583333).. + 8a1aZ - 6az - 6a1 + 1.083333
..2 + (- 37a1az + 24.5az + 23.5a1 - 5.916667).. - 3az + a1 + 4.916667
(7.28)
It should mentioned that our techniques do not guarantee that the largest
computed a is the largest attainable. Rather, FIT can be used to compute
lower bounds. Our purpose here is to demonstrate that SSFIT can be used
to compute improved lower bounds. This is done by exploiting the fact that
SSFIT in essence allows one to work with sectors that are wider than 180 0 •
This in turn can lead to controllers that guarantee robust stability for larger a.
Let us first design a controller using FIT. The initial controller is chosen by
t
pole placement and is given by: z,' 11. It places the poles of the nominal
closed loop system at: -0.3189 ± j1.9026, -1.61216. The FIT software [53]
was run with this initial controller and with equally saed sectors 311"/4 radians
152 CHAPTER 7
wide. There is one such sector symmetrically positioned about the positive real
axis and the others follow by rotating 45° at a time. The procedure terminated
with the controller:
2.01736 + .96559
6 + 1.0032
10,.---,....--,....--,....--,....--,....--,....--,....--.---.-----.
8 , . . ..;...
.. .
6 ; ; : ; : : :- , , .
··· .
..
. .
.
· .
.: : ~
·· .
~
..
~ .
.;
.: :.
· .
;
.
: .
·
· ... ..
~ ··· .. ...
.r
.E .., Q ; ; : .
.
..
.
..
.
.
....
-2 ; ; , : : : , , .
: : .
-4 ; ~ : : : + : i ; .
·6 , , -: .
·8 .
. 1~~0--.8~-.6~--4L--.2~-0~-2~-4L--6~-8~---!10·
real
From this value set it would seem that as Q increases the convex hull will include
the origin but the actual value set would not. We must caution here that a
single value set at a specific frequency does not tell the whole story, because this
may not be the "critical" frequency. It could happen that as the uncertainty
range increases some vertex or other plant family member goes unstable first,
and this could occur at a different frequency. However, this seems to be a
FIT based V-Stabilization 153
good candidate for SSFIT. The geometry of this value set suggests that we
"partition" the family along the 02-axis. We will partition the given plant
family in two sub-families and run the SSFIT software. The first plant family
is obtained from the original by substituting 02 - .25 for 02. This results in:
n,1('" a)
Pl(6,a) = d, ( ) =
1 ",a
(180102 - 12.502 - 1701 + 6.708333)6 + 80102 - 602 - 801 + 2.583333
62 + (-370102 + 24.502 + 32.7501 - 12.041667)6 - 302 + 01 + 5.666667
The second plant family is obtained by substituting 02 + .25 for 02. The
transfer function is:
n,2('" a)
P2 (",a) = d, ( ) =
2 ",a
(180102 - 12.502 - 801 + .45833)" + 80102 - 602 - 401 - .4166667
62 + (-370102 + 24.502 + 14.2501 + .208333)" - 302 + 01 + 4.166661
We now seek to find a controller that simultaneously stabilizes these two plant
families for parameters in the range [-a ::5 01 ::5 a] and [-a/2 ::5 02 ::5 a/2],
for the largest possible a. The manner in which the SSFIT software is written
allows one to use different a's for each plant family. Using the same initial
controller as in the running of FIT and with the same sector assignment for
both families, the SSFIT software generated the controller:
2.01526 + .9735
(7.29)
6 + 1.0001
a larger region of stability would have been guaranteed. Figure 7.12 shows the
value set for this controller at frequency w = 2.
154 CHAPTER 7
15r---"'T-"--~---r----.----,.----,
10 ..
5 .
:i
~
0 .. · ·9 · ..
-5 . ..
-10
-1_i':-5---....J.1o::-----~5---0~----!5:----..Ll0-----l15
,eal
The value set 1/>(j2, O~SFIT), lies in a sector which is 188.60 wide! In essence
the SSFIT algorithm gave us the ability to manipulate value sets that are wider
than 1800 , guaranteeing robust stability for a larger uncertainty region. It is
important to point out that such a value set could never have appeared in the
execution of the FIT algorithm.
We reiterate, that one should distinguish between what is guaranteed at the end
of a design procedure and what can be shown to be true after further analysis
of the results. It could happen, because of the overbounding introduced, that
the designed controller does actually robustly stabilize for a larger uncertainty
range than what the procedure guarantees. This is a matter of chance, and
one cannot speculate apriori that it will occur. Furthermore, this example by
no means proved in a theoretical sense, that SSFIT can design controllers for
larger uncertainty ranges than what would be possible with FIT. It may very
well be that had a different initial controller been chosen in FIT, it could have
been shown to guarantee robust stability for the region O~SFIT. We simply
do not know. However, the example did demonstrate that for this case SSFIT
with little effort generated a robust controller for a larger uncertainty set.
FIT based V-Stabilization 155
Consider the situation when a finite number oCstrictly proper plants Pl (6), P:Z(6),
... , Pt(6) of the same order are to be simultaneously stabilDed. The basic idea
of an algorithm is the following: Identify one of these plants as the "center"
plant, say Pl. If these plants are thought of as points in plant coefficient space,
the center plant should correspond to the plant closest to the geometric cen-
ter. A "robust" controller is then designed for the center plant, for as large
a range of an uncertainty region as possible. Clearly, if all the other plants
fall inside this ball our job is done. If not, then "auxiliary" plants are defined
Piou = :;...... inside this ball and closest to the actual plants. In an iterative
.....
manner, the choice of auxiliary plants is improved so as to get closer and clo-
Ier to the actual plants. Associated with these auxiliary plants are families of
"auxiliary" characteristic polynomial families formed in the following manner:
156 CHAPTER 7
4>iou:I:(S,~)=
dc (s)((1 - ~)~OU:I:(s) + ~d,n(s)) + n c (s)((1 - ~)~OU:I:(s) + ~7lpi(s))
where ~ takes values in the interval A C [0,1] and 2 :S i :S i. These are used
in the iterative procedure in order to identify "new" auxiliary plants that are
increassingly closer to the actual plants .
We should again mention that this is simply an iterative algorithm for stabili-
zing a finite number of plants. It has been inspired by SSFIT, but there is no
guarantee that it will find a simultaneous stabilizer if one exists.
FIT based V-Stabilization 157
A version of this algorithm can also be stated for directly performing simulta-
neous plant famil1l stabilisation for families Pt(", a), P2 ('" a), ... Pt(", a) with
multiafline uncertainty. Assume that each family involves the parameter a
which takes values in the set 0/1. In the same spirit as above, a "center" plant
family is chosen (say Pt('" a)) and a controller computed that robustly stabili-
les this family for as large an uncertainty region 0/1 as possible (0/1 :J 0/1). If
this controller also robustly stabilises the rest of the plant families our job is
done. Otherwise, "auxiliary" plant families are defined PI(", a) inside 0/1 for
2 ~ i ~ l in the "direction" of the actual plant families. Using SSFIT, in an
iterative manner these auxiliary plant families are "enlarged" and a controller
computed that simultaneously robustly stabilizes Pt ('" a), P~('" a), ... PI(", a).
I! the original plant families are included in this set the process terminates. I!
not, new auxiliary plant families are defined inside the "old" auxiliary plant
families (in the "direction" of the actual plant families) and the "enlargement"
procedure is repeated. I! during the execution of the algorithm the actual
plant families are found to be included in a set of auxiliary plant families the
algorithm terminates successfully.
8
FIT SYNTHESIS FOR ROBUST
PERFORMANCE
Having the ability to design controllers that achieve the desired performance
objectives is of paramount importance to control. Throughout the years a great
deal of effort has been directed in the development of such methodologies and
a number of successful formulations and solutions have been proposed. Among
the earlier ones is the classical technique of "loop-shaping" which is primarily
a single input, single output technique. As the label denotes it is a formulation
that takes performance objectives and "maps" them on the shape of the loop
160 CHAPTER 8
One specific .example for which existing H oo techniques giYe an elegant cha-
racterization for robust stability, is the case of SISO plants with multiplicative
uncertainty. Consider the family:
(8.1)
where P(,,) is some "nominal" system transfer function, W2 (,,) is some given
minimum phase stable transfer function and ti. is any arbitrary stable and
proper transfer function with lti.(jw)1 :5 1, for all w E [0, (0). Let the con-
troller C(,,) and the plant P(s) be in the forward path of a unity feedback
configuration and assume that C(,,) stabilises the nominal plant P(,,). Then a
necessary and sufficient condition for robust stability is IW2 (jw)T(jw) 1 < 1
for all w E [0, (0), where T(,,) is the nominal complementary sensitivity [36].
This elegant characterization of robust stability allows for controller synthesis
as a solution to a Nevanlina-Pick interpolation problem [36].
For this feedback system let us in addition require robust asymptotic tracking as
a performance requirement. In other words we insist that IW1(jw)S(jw)1 < 1
for all w E [0, (0), where S(,,) is the sensitivity transfer function of the plant
family, 5(,,) the nominal sensitivity and WI (,,) is some minimum phase stable
transfer function providing the "weight" for S(,,). A characterization of robust
stability and robust performance (assuming nominal stability) is [36]:
Unfortunately, even though this is a necessary and sufficient condition this for-
mulation does not lend itself directly to controller synthesis, unlike the case of
mere robust stability. One is forced to introduce conservatism by overbounding.
Specifically, rather than working directly with I IWI(jW)S(jW) 1 + IW2 (jw)T(jw)ll,
one employs the expression 1 IW1(jw)S(jwW + IW2 (jw)T(jwW 11/ 2 • Using
geometry one can show that:
FIT Synthesis for Robust Performance 161
Working with this expression does lead to Riccati equation solution methods.
Our focus in this book is the development of robust control synthesis methods
for linear time invariant systems with parameter uncertainty. Even though
most of our emphasis is on SISO systems in the last Chapter we will discuss
the application of our techniques to MIM0 systems as well. Our method of
attack is to take the given performance requirements and first express them
as specifications in the frequency domain or on the location of system poles.
The frequency domain approach is a "standard" practice used by both clas-
sical control techniques as well as the more modern H oo synthesis methods.
We then take these performance requirements and pose them as robust poly-
nomial stability requirements or robust polynomial 1>-stability requirements.
We have already seen how this can be done in Chapter 4. With this formula-
tion the robust design problem then becomes a problem of polynomial family
stabilisation or 1>-stabilisation. The solution of this problem is then carried
out by using Finite Inclusions Theorem synthesis or Simultaneous Polynomial
Stabilisation. We will first focus on FIT synthesis and discuss Simultaneous
Polynomial Stabilisation in Chapter 10.
For some performance requirements one can immediately see how they can be
expressed as specifications on the magnitude of some transfer function. One
example is asymptotic tracking. In Chapter 4, we considered the unity feedback
system depicted in Figure 8.1
-
r +
-
+
e
C(6) P(6,a)
y
where the plant family is strictly proper and involves parameter uncertainty,
P(,,) = 'Zf:::I.
The parameter a lies in some hypercube O. C R i , the degree of
dp(", a) is n and its leading coefficient is positive. The controller C(,,) = ~:f:1 is
proper. The input is a sinusoid of some frequency w which belongs to some given
frequency range. For a robustly stable system the robust asymptotic tracking
requirement imposes a magnitude constraint on the sensitivity, namely:
is stable for all Ql E [0,211"), rl E [0,1] and a EO•. Let 1/Js('" a, rl, Ql) be a
degree n polynomial family where n = 8(d..l(8» + 8(dp(8, a» + 8(dc (8».
In view of the Finite Inclusions Theorem one can immediately formulate the
following sufficient condition for robust asymptotic tracking. From our work
in Chapter 7 we expect that this result will provide the foundation for a robust
performance synthesis procedure.
FIT Synthesis for Robust Performance 163
W1 ~ W2 ~ W3 ~ ••• ~ Wm
wn wn
-2 - w ~ C1 < d1 ~ 2 +w
wn wn
--w<em<dm<-+w
2 - - 2
V 1 ~ A: < m, maz{dl - cl+1,dl+ 1 - cll ~ w
1PS (jWl , 0/1, r1, (1) E {rei I r > 0, Cl < <
'(}
1 ~ A: ~ m, (J dll
Now there may be additional performance requirements that also lead to con-
straints on the magnitude of the sensitivity transfer function all of which can be
taken into account by a proper choice of the weighting function W 1 (.). Again,
all these would result in the stability of a ,ingle polynomial family.
In the case of time domain specifications on the step response an alternate for-
mulation would be to exploit the pole placement method and express these as
requirements on the location of the closed loop poles. In that case an appro-
priate 'D region would be defined and the closed loop characteristic polynomial
would need to be robustly 'D-stable. Again these requirements are translated
to the V-stability of a ,ingle polynomial family.
In this Chapter we show how FIT synthesis can be carried out based on Theo-
rem 8.1, in the case when performance objectives lead to the requirement of
stability of a single polynomial family. We will use the robust asymptotic
tracking paradigm as a representative of this class of problems and will develop
a FIT based algorithm for its solution. In the next Chapter we will consider the
more general case when performance requirements are translated to stability of
a finite number of polynomial families.
1m
--.Re
lisation which is based on the Finite Inclusions Theorem it will have the same
structure as the FIT algorithm presented in Chapter 7. It will be an iterative
algorithm where again some initial controller is redesigned in order to enhance
robustness. At each iteration a new controller is computed as the solution of a
set of linear inequalities.
Consider the feedback system in Figure 8.1 where the family of plants is strictly
proper given by:
where dp(", a) is of degree fl., the parameter a takes values in some hypercube (lei
and the leading coefficient of dp(", a) is positive. We assume that the uncertain
parameters enter in a multiaffine manner in the coefficients of n,(", a), dp(", a).
The controller is proper of order q and given by:
The controller parameters are grouped in the vector x = (Zl' :1:2,·· ., Z29+l)
E R II where d = 2q + 1 and the closed loop characteristic polynomial is given
by:
(8.11)
is robustly stabilised. This is true since the polynomial family tPs(" a, r1, (1)
is included in the polynomial family tPS(" a, b). We will denote by 06. the
"square" overbound and by 060 the "octagon" overbound. Due to the over-
bounding this is only a sufficient condition for robust asymptotic tracking. In
order to make the discussion more concrete let us assume that the square over-
bounding is used (i.e. b = b1 + jb2, bE 06.). Structurally, this problem is
no different than the one encountered in Chapter 7. The polynomial family has
complex coefficients and real parameter uncertainty which includes the plant
uncertainty a and the "performance" uncertainty b. Even though there is no
structural difference, parameters a and b have a different interpretation and
this distinction must be maintained in any algorithm. Furthermore, we are
only required to robustly stabilize. With these issues in mind, appropriate mo-
difications are made to the FIT algorithm of Chapter 7 in order to generate
the Performance FIT (PFIT) Synthesis Algorithm.
Just as in the case of the FIT algorithm, one needs to chose an initial controller
(1)
X • Let 046 = 04 X 06" an
d 1et 0(1)
46 = 04(1) X 0 6(1)
• C 046. T 18
h'
controller must stabilise tPs(" o~i), x(l» and can be designed by employing
any design technique. This controller is then iteratively improved upon using
FIT. At each iteration frequencies "'~) are found that place the value sets
tPs(j",~), o~/, xC;»~ in their corresponding sectors 51;). By FIT this guarantees
the robust stability of tPs(', o~/, x(j». At each iteration the uncertainty set
o~/ is enlarged. Initially only the b parameters are affected until (if possible)
for some j, the desired performance is attained o~;) = 0".. Once this is
accomplished the procedure continues this time with an enlargement of o~/
where the a parameters are affected and the b parameters remain unchanged.
FIT Synthesis for Robust Performance 167
The algorithm terminates when (if) n~~ ::> nail, with x(j) being the desired
controller.
STEP 1: Let x(l) E Ref and n~i) c nail be such that 4>(8, n~i), X(I»)
is stable and set j:=1.
STEP 2: Determine m(j) ~ 1 sectors sij), 1 ~ A: ~ m(j), and frequencies
w(j)
Ie
along the J'w axis such that Y'
.l.(w(j)
Ie ,
EztO(j)
all'
x(j») C S(j)
Ie
By FIT, 4>(8, n~~, x(j») is stable. Each wij) should
roughly center (angularly) the set 4>(w~j),EztO~!,,>,x(j») in sij).
STEP 3: Choose a slightly larger set n~+I) ::> n~). First this should
affect the b parameters. When (if) n~j) ::> nil" the
enlargement in the b-direction terminates and the enlargement
in the a-direction commences.
STEP 4: Compute a new vector of controller parameters x(i+1) such
that ; (jwij) , Ezt n~+1)I x(i+ I ») C sij) for allA:. Note, this is
equivalent to solving a system of linear inequalities in x(i +1).
If no solutions exist to this system of inequalities, return to
Step 3 and choose a smaller n~+l).
STEP 5: Let j := j + 1, and if n~b) ::> nail, stop; otherwise, go to Step 2.
This is one of several synthesis algorithms that can be proposed which are based
on the Finite Inclusions Theorem formulation.
(8.13)
where
The software package written in [53], has the capability of executing a version
of the PFIT synthesis algorithm. The first step in this algorithm is the design
of an initial controller. We have a number of alternatives at this point. One
FIT Synthesis for Robust Performance 169
8.3.1 H oo Synthesis
The plant family is already represented in terms of a multiplicative uncertainty
model where the nominal plant is:
(8.15)
(8.16)
In order to carry out an Boo design we need to take the parameter uncertainty
model and turn it into a norm bounded uncertainty model. In so doing, very
frequently overbounding takes place and care must be exercised so that this is
minimi.ed. This was done in Example 3, Chapter 7. Several such weighting
functions were developed from which the "best" from an Boo point of view (i.e.
it gave the largest Q for which robust stability could be guaranteed) was:
Beiter weighting functions can be developed, in the sense that the Boo design
will produce a controller for a larger Qmll%' if one uses higher order transfer
functions W 3 (,,) or if one chooses the appropriate shape for W 3 (,,). Picking
one of higher order is straightforward, but in general it would lead to a design
with higher order controllers. Since controller order is an issue in applications
it is prudent to try and keep the controller order as low as possible. Choosing
an appropriate shape for W 3 (-,) is quite more difficult and at this stage of
development it involves trial and error and reliance on "experience". Since one
170 CHAPTER 8
With W2(6) defined, the norm bounded uncertainty becomes A(6) = W~(6)9(6),
where 9(6) is any proper Itable tranlfer function with H oo norm less than or
equal to 1. Recall that the Hoo-norm of a stable and proper tranlfer function
9(6) is 6up",19(jw)l. We further ulume that no unltable poles of PO(6) are
canceled. It is clear that this norm bounded uncertainty includes the original
parametric plant family. Therefore, any H oo controller bued on this represen-
tation of uncertainty will guarantee robust stability and performance for the
given plant family. The design will bf! conservative since overbounding has been
introduced.
A necessary and sufficient condition for robust stability and robust performance
for the loop with norm bounded uncertainty (assuming nominal stability) is
[36]:
However, this expression does not ealily lend itself to Iynthesis and one works
instead with:
This is equivalent to finding a controller for which the infinity norm of the
transfer function:
(8.20)
In our case we want to find the largest Q for which robust stability and robust
performance can be achieved. This Wal accomplished by multiplying W~(6) by
Q, solving the corresponding H oo problem, and iteratively increasing Q until
FIT Synthesis for Robust Performance 171
condition (8.19) could no longer be satisfied. The controller for the largest a
(using our procedure) was found to be:
(8.21)
where
for
a: 4Z = 0.8399 (8.22)
Figures 8.3 and 8.4 show the magnitude plots of T'l t1 1 and IW1SI + IW2TI
respectively.
Clearly, one sees from Figure 8.4 that there is room for improvement, if one was
able within the context of H oo synthesis to work directly with IW1SI + IW2 TI.
!inii
:: l.::
: : ::::
: :::::
: : ::::
.... • j
:
.l::!!!
i i.. i.i.~.~.i
!::
~ j••.• i
. : . : : : : ::
i.~.i.i.~~ i
:..
~ ~ i..i.. i.~~~
. : ::
~
::;"
; i j ~·i.:
..
.
~ ~
=€I
>
Ul
-4.5 ···+····;···j··j·j·t·ii ·.. ·······+······;····~···t·i·t~·rr· ~ ~ ~ :.......
.:..:.+.;+ :..
: : I:::
: : I:::
-5 ..!..... ~.! "'i . .':..'I...':..':. :•.:•,....·.. ·.·.:.······1·· .~.. f... '•••••••• " •• '•• ··········1 ..... ~ ·+··!··j··j·nf········· .~....+..
· ··!H
·. ..
..
-5.5 ~-..i.......i-.L..i...i....i..l..i"":---'-....i-.L..i...i..i..i.i.L---'-....i-.L..i...' .....~_.L-i-i...i.... .....i.iJ
10-3 10-2 10-1 100 101
Frequency - Rad/Sec
. : :::::
......
-1.5 -····f·· · .. ~. ·····1 ·..1
-2 'L r..·f.. !·+,~;·;, .. ···..·····.. ··.!..·· .Ll.ll.L ,.., l,..L.U ~", , _ i..,.,_ ..
: i : iii::
-2.5 ··········!······~,···~···f··i.i·t·~·~···········+····
; ....:
·.! ..·.i···t·1·: :.~~ i······t····}···~··j··l·tit···········t······;····;···~··t·~·i.
; : : :::; : : ..
.. '.{..~~ ~ ..
·3.5
...
-4.5
; :::::
~· ..~·"·I·I·I· ··········i ..... ~... !.++ 1'~"~f ..;
•5.5
;: ::;;
L.....-I................................._-'---........................'---'--'-
..
_"'--
. .
........
10.3 10.2 10. 1 100 101
Frequency - RadlScc
FIT Synthesis for Robust Performance 173
One can verify that if a = I, this controller does not satisfy the robust
performance requirement. Specifically, for 01 =
-1,02 = -I, the sum of
the weighted sensitivity and complementary sensitivity is larger than unity at
low frequencies. The software was run using the "square" overbound for the
performance parameter and generated the controller
FIT
a maz = 1.1065 (8.25)
We should again point out that the values of a maz that are reported are those
generated by the particular synthesis procedure used. It may be that the con-
trollers constructed after testing are shown to actually provide robust perfor-
mance for a larger uncertainty range.
It is interesting to compare these results with those obtained from Hoc synthe-
sis. For this example, FIT synthesis generated a 8econd order controller that
satisfies the objectives for a considerably larger range of uncertainty. However,
one has to be careful not to draw general conclusions about the two synthesis
procedures. We are comparing olgorithTn6 and not theoTeTn6. At this stage,
all we can say is that for this example, the formulation used and the solution
method applied, FIT synthesis gave favorable results. Much more testing is
needed in order to draw concrete conclusions about FIT synthesis. Just like
any other synthesis method there will be classes of systems for which FIT syn-
thesis will be advantageous and some for which it will not. Our hope is that the
reader will be motivated enough by our exposition in order to use FIT synthesis
in control applications.
9
FIT SYNTHESIS FOR ROBUST
MULTIOBJECTIVE
PERFORMANCE
y
nt~
de ,
where the plant involves parameter uncertainty and robust performance requi-
rements include asymptotic tracking and noise attenuation. From the discus-
sion in Chapter 4, one can take the tracking requirement and turn it into a
robust stability requirement of the feedback configuration in Figure 9.2.
r y
for a EO.. , 0 ~ rl ~ 1, (ll E [0,2?l'). In addition, one can take the noise
attenuation requirement and turn into a robust stability requirement of the
feedback configuration in Figure 9.3.
exposition in Chapter 4 gives a hint for a different formulation. One can take the
two requirements and translate them into a 6ingle robust stability requirement
of the feedback configuration of Figure 9.4:
y
p
for all Q:z E [0, 2~), r:z E [0, I], Ql E [0, 2~), rl E [0,1] and a E no' Since po-
lynomials d.l(a), d.:z(a) are stable one can immediately see that if polynomial
family ,pST in (9.3) ia stable then polynomial families ,ps and ,pT will be sta-
ble. With this formulation the stability of a 6ingle polynomial family would
guarantee both requirements. However, since ,pST is a "larger" family there is
conservatism introduced. We may be able to simultaneously robustly stabilize
,ps and ,pT, but not be able to robustly stabilize ,pST.
Consider the unity feedback system of Figure 9.1 where the plant family is
strictly proper of order ti and involves parameter uncertainty, pea) = Zf;::~.
The parameter a lies in some hypercube 0/1 C R i , the leading coefficient
of dp(a, a) is positive and C(a) = ~:f;l is some proper order q controller.
Let W1(a) be an appropriate strictly proper, stable, minimum phase trans-
fer function which is the weight on the sensitivity and W2 (a) be a proper,
stable, minimum phase transfer function which is the weight on the comple-
mentary sensitivity. Let ns, AT be the degrees of ,pS,,pT respectively and let
rs = =
rl, Ots =
Otl, rT =
r2, OtT Ot2. The following result is a restatement of
the SSFIT result (Theorem 6.5) presented in Chapter 6, that applies specifi-
cally to robust asymptotic tracking and noise attenuation. It provides the basis
for the multiobjective synthesis algorithm presented in the next Section.
180 CHAPTER 9
Consider the feedback system in Figure 9.1 where the family of plants is strictly
proper given by:
_ np(s, a)
P( s, a ) - dp(s, a) (9.4)
The controller parameters are grouped in the vedor x = (Zt, Z2, ••. , Z2q+t)
E R d where d = 2q + 1 and the closed loop characteristic polynomial is given
by:
(9.6)
Our first priority in the PSSFIT algorithm, is to chose an initial controller x(1).
Let Oel" = (1)
0el X 0"" and let 0el" = (1) (1) •
Oel X 0", C Oel'" This controller must
simultaneously stabilise tPi(6, O~~), x(1») for i = 5, T. Such a controller could
be designed by loop shaping techniques applied to the "nominal" system. Ano-
ther possibility might be to employ "standard" Hoc techniques to the problem,
which we know are conservative. This controller is then iteratively improved
upon using SSFIT. At each iteration frequencies w~~) are found that place the
value sets tPi(jW~~), o~2, x(j») in their corresponding sectors ~~). By FIT this
guarantees the robust stability of tPi(6, O~,,>, x(i») , i = 5, T. At each iteration
the uncertainty set O~,,> is enlarged. Initially only the b parameters are affected
until (if possible) for some j, the desired performance is attained O~i) = 0",.
Once this is accomplished the procedure continues this time with an enlarge-
FIT Synthesis for Robust Multiobjective Performance 183
ment of O~~ where the a parameters are affected and the b parameters remain
unchanged. The algorithm terminates when (if) O~~ :::> OGb, with x(j) being
the desired controller .
(9.11)
where J D is the polar moment of inertia of the antenna dish, TM is the DC-
motor torque, KCA a cable stiffness gain and B a viscous friction gain. In this
simplified model the DC-motor dynamics are assumed to be a gain K t • A rate
sensor measures the angular rate WA = -lie
and is a second order system with
transfer function:
n (6)
G(6) = g
dg (6)
Table 9.1 provides the values for the system parameters where it can be seen
the J D and B are uncertain.
The primary performance objective for the seeker system is to have the actual
seeker rate WA follow a commanded seeker rate Wc. In addition to this tracking
requirement the system must also reject a disturbance torque Til at the input
of the plant. Specifically, there is a bound placed on the magnitude of the
e
transfer function from Til to for high frequencies. In view of the presence
of the parameter uncertainty the system must exhibit robust performance. A
block diagram of the overall system including the controller is shown in Figure
9.5.
FIT Synthesis for Robust Multiobjective Performance 185
we
1
JD'
WF
• +B
!lla
The robust performance objectives for the system if Figure 9.5 in more detail
are as follows:
• The system must exhibit a transient step response between the comman-
ded seeker rate We, and the achieved seeker rate WA, with the following
properties:
- The time to reach its peak value Tp , must be less than .013 seconds.
- The overshoot must be less than 20% (corresponding to a damping
ratio greater than .5 for a second order dominant system).
- The step response must reach a ±2% envelope by .05 seconds and
remain in this envelope up to .5 seconds.
• The bandwidth of the closed loop transfer function :~~:~ (with Td = 0),
must be less than the rate sensor's natural frequency.
• The magnitude of the transfer function from Td(-') to 9(-,) should be less
than -100 db for frequencies higher than 300 rps.
Even though some of the design specifications are given in the frequency do-
main, some also refer to the transient response. In general, there is no direct
method for taking such design specifications and posing them as frequency do-
main objectives or as requirements on the location of closed loop poles. This
must be done in an indirect manner which involves approximations. Here we
have two alternatives: The first, is try to identify a region 1) in the complex
plane such that if the closed loop poles, are in 1) the system has the desired
186 CHAPTER 9
transient response. The second, is to use a second order dominant model for the
system and employ existing results to relate transient response characteristics
to the shape of specific transfer functions. One could also use some "combina-
tion" of these alternatives. In view of the fact that some of the frequency design
specifications already involve the sensitivity and complementary sensitivity, in
this example we will exploit the second alternative.
The plant in this example (which includes the DC-motor gain) has a transfer
function:
P(,,) (9.12)
and other transfer functions of importance are the loop, sensitivity, comple-
mentary sensitivity and closed loop transfer functions respectively:
L(,,) = C(,,)P(,,)G(,,)
1
S(,,) = 1 + C(,,)P(,,)G(,,)
C(,,)P(,,)G(,,)
T(,,)
1 + C(,,)P(,,)G(,,)
C(,,)P(,,)
TCL(") 1 + C(,,)P(,,)G(,,)
120s
Po(S)
S2 + .55s + .75
This initial controller will then be used as input to the SSFIT software which
in essence performs "robustness enhancement".
The nominal transfer function from Td( s) to O(s) is"ifl7\l So( s). In order to
",.0\' )
satisfy the requirement placed on this transfer function we must have:
where Wi1(jw) takes values less than -100 db in the frequency range
w E [300,00). Such a WF(S) will be incorporated into W1(s).
The step response "speed" requirement can be related to the system bandwidth.
For a second order system the bandwidth WB is approximately related to the
undamped natural frequency w" by wB ~ 1.2w". And for a second order
system:
In view of the presence of 6 in the numerator of the plant, the step response
will eventually go to zero. Our goal is to have the response reach unity and
stay in the +2% envelope up to .5 seconds. This implies that Lo(jw) which at
low frequencies is zero must cross the 0 db line at a "small enough" frequency
7. The smaller the 7 the longer the response stays in the envelope. At frequen-
cies just above 7, ILo(w)1 is large and therefore ISo(jw)1 ~ ILo(jw)l- l begins
to decrease. The frequency 7 will thus be used as the crossover frequency for
W l (6) (at low frequencies).
Finally, the magnitude of the loop transfer function must attain a high enough
value in the low/mid frequency range so the the step response will reach the
value 1. This is reflected in the shape of W l (6) in this frequency range.
There are two more performance requirements that must be taken into account,
the restriction on the overshoot and the high frequency bandwidth and roll off
requirement imposed by the rate sensor. These naturally can be stated as
restrictions on the shape of TCLO(6). One additional requirement will also be
imposed which is not present in the specifications of the loop but is a sensible
thing to do. In an automated synthesis software package as SSFIT, one has to
guard against having the zero at the origin being "canceled" by the controller.
In actuality, since closed loop stability is guaranteed the cancellation will not
occur, but one might have a "near cancellation" which is not appropriate. A
W2(6) needs to be constructed which incorporates all these requirements.
The bandwidth limit and roll off characteristics of TCLO(6) are directly affected
by the those of W 2(6). In view of the relationship between TO(6) and TCLO(6)
this can be immediately reflected on To( 6).
Using again the second order approximation the requirement on the overshoot
is expressed as a restriction on the magnitude of TCLO(jW) to be less than 6 db.
Finally, the additional requirement placed to prevent the near pole-zero cancel-
lation is imposed as a low frequency crossover limit on Lo(jw) and consequently
TCLO(jW).
The design requirements (in addition to robust stability) can now be expressed
as:
From the previous section we have that the requirements will be robustly met
if a controller can be found that simultaneously stabilizes the following two
polynomial families:
190 CHAPTER 9
where
The following controller was designed employing loop shaping on the nominal
system and used as the initial controller for the PSSFIT algorithm:
e LS (6) =
5.418664 + 5.101 X 103 63 + 4.8161 X 106 62 + 2.6512 X 106 6 + 3.61 X 106
64 + 1.4194 X 103 63 + 1.0076 X 106 62 + 5.3936 X 106 6 + 1.444 X 107
One can check that this controller does not satisfy the robust performance
requirements, as there are parameter values for which the overshoot objective
is violated. Had this been an acceptable controller the design would have been
complete. At this stage one could go back and modify the weights such that
they reflect the parameter uncertainty in the plant and carry out a new loop
shaping design. Rather than following this avenue, the controller is input to the
SSFIT software package [13] (which implements the PSSFIT algorithm) where
sector size is set to \1:
rad. The resulting controller is:
F1T
C (6) =
5.96264 + 4.6132 X 103 63 + 4.407 X 106 6 2 + 4.1894 X 106 6 + 4.8806 X 106
64 + 1.4242 X 103 6 3 + 8.9185 X 1056 2 + 6.4548 X 106 6 + 1.4667 X 107
which guarantees robust performance for not only the given range but for
the parameter set 0/1 = {a E R 2 I at E [-1.3672,1.3672] , a2 E
[-1.3672,1.3672)}. The three plots that follow show the results of the de-
sign. Figure 9.6 shows the step response of the family of plants for the time
period up to .02 seconds. This allows one to see that the overshoot is less than
20%, the peak time is less than .013 seconds and that the response is inside the
±2% envelope by .05 seconds.
192 CHAPTER 9
1.4
1.2
•
!0.8
t
t o. 6
0.4
0.2
00 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
time in seconds
Figure 9.7 gives the step response of the family of plants up to time .5 seconds,
showing that the response remains in the ±2% envelope for up to this time.
1.4
1.2 _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._.
•
!0.8
t
t o. 6
0.4
0.2
00 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
time in seconds
Figure 9.8 is the magnitude plot for the family of transfer functions from Til to
6, showing that for frequencies above 300 rps values are less than -100 db.
50..---~-........-~--..---~- ........-~----,
o : : : ..
e.Ii ·50
«
i
ie ·1OO .
·150 ~ ; , .
··· ...:
···: ...
··· ...
· .
·200·L,------'-:,....-~--"":c------'-:,....-~-----l
10~ 10" 10· 10' 10'
lI8quency In rps
Finally, Figure 9.9 shows that the bandwidth of the closed loop transfer function
(we to WA) is below 942.5 rps.
10....-----.,-----..------.,------,
o : .._.-------~-,
,~
·10
·50L.:·----"-:·,....-~--""::·,....-~--1""::0·,....-~---'10·
10~ 10' 10
lI8quency In rps
T
Consider the standard robust stabilisation problem of Figure 10.1:
~lf--e-·I_C_(S)_ -~IIL--p_(s,a_)
n,(s, a)
P(s,a) (10.1)
d,,(s, a)
where d,,(s, a) is monic of degree n and the parameters a take values in some
hypercube 0 ... The controller is order q and described by some proper transfer
function C(s) = ~:f:~. The closed loop characteristic polynomial is given by:
and we seek to construct such a controller that stabilises (if possible) this
polynomial family.
Let us begin with the simplest case of a single parameter uncertainty. In par-
ticular, let the plant family be described by:
It is important to note that the first two polynomials 1/11 (8), 1/12(_) come from
family "vertex plants", but that the other two come from "fictitious vertices".
At this point we have two concerns: Firstly, even though this is a very inte-
resting interpretation of robust stabilisability it does not provide any insight
into the problem of how to construct such a controller. The requirement here is
to find a controller that simultaneously stabilizes four polynomials. Well, this
is a problem very similar to simultaneous plant stabilization which we know is
unsolved for more than two plants. Secondly, in order to obtain this analyti-
cal result we had to resort to rectangular value set overbounding, a procedure
which is inherently conservative. We now focus our attention on these issues
and provide some answers.
Consider the plant family in (10.3) where n,o(8) is stable, and d,1(") = 0
(i.e. a single parameter affecting the numerator). Theorem 10.1 guarantees
that the family will be stabilised if a controller is found that stabilizes the four
polynomials:
r + e y
+ "cf·
dc•~ PiCa)
-
(1 + ne(6)npl(6»
Pl(a) = ".0(. }!lc(. )+".,(.)".(.)
"c(. jilPo (.) = ne(a)npo(6)
ne(a)~(6) ne(6)npo(6)
ne(6)npo(6) (1 _ ne(6)npl(6»
P:z(a) "po(. )"c(. )-",,(. )"c(.)
"c(.)lI,o(.) = ne(6)~(6) ne(6)npo(6)
ne(6)npo(6) (1 + ne(-a)npl(-a»
P3 (6) = ".0(. )"c(. ~+"Pl~-. )"c(-.)
ftc .)11,0 .) = nc (6)d,.o(6) nc (6)npo(6)
P.(6) "po(.}!lc(. ~-""~-. )nc( -.)
=
ne(6)npo(6) (1 _ ne(-6)npl (-"»
ftc .)11,0 .) nc(6)~(6) nc(")npo (,,)
(10.7)
Let We,,) be a stable, proper, minimum phase transfer function such that:
Since nee,,) is stable ":~~» is a Blaschke product and I".:~&t» I= 1 for all
wE [0,00). Furthermore, I":;~&~»I = 1:::f~:~1 for all w E [0,00). Let 6(,,) be
200 CHAPTER 10
any proper and stable transfer function with 1~(jw)1 ~ 1 for all w E [0,00).
We are now able to define via this norm bounded perturbation the family of
plants with multiplicative norm bounded uncertainty:
where Po(s) = ~:f:~. One can easily verify that the four plants pieS) in (10.7)
are members of this plant family. Since ne(s) and noes) are stable, choose the
4 ~(s)'s as:
~l(S)
_ W- 1 (s) ne(s)npl(S)
ne(s)npo(s)
~3(S) = W-l(s)n.(-.~n'lf-·) -W- 1 (s) ne (-s)npl( -s)
n.(. npo .) , n e (s)npo (s)
For the plant family (10.9), in the feedback loop of Figure 10.3:
~W
we know from [36] that a necessary and sufficient condition for robust stabiliz-
ability is the existence of a controller that stabilizes the nominal plant poes)
and satisfies:
Theorem 10.2 Let P("al) be a.fa.mily ofpla.nt6 til in (10.9) withal E 0/1,
dpl(') = 0 and npo(') dable. Let W(,) be a proper, minimum phtlle, da.ble
tranlfer function that ,atilfie, (10.8). Suppo,e that C(,,) = ~:f:l iI a. mi-
nimum pktlle controller tkat dabilize, poe,,) and ,atilfie, (10.10). Tken thil
controller ,imultaneoUlly ,tabilize, tke ,pi(S) in (10.6).
Let us now consider a second, single uncertainty strictly proper plant family
with uncertainty solely in the denominator:
(10.11)
Suppose that we now restrict ourselves to &table controllers (dc (") is stable).
Then it is evident that simultaneously stabilizing the four 1/1,(") in (10.12)
is equivalent to simultaneously stabilizing dc (")1/1,("), 1 ~ i ~ 4. However,
these four are the closed loop characteristic polynomials of the unity feedback
configurations as in Figure 10.1, where the four plants are:
+ dC(,,)dpl("))_1
Pl(") = npoC.)dt)
d.(. )dpo(.)+d•• )dpl(.) = dc(")npo(")(1
dc(")d,o(") dc(")d,o(")
Pz(") = .. ~o(.)dt\
d.(.)dpo .) d•• dpl(.) = dc(")npo(,,)
dc(")d,o(")
(1 _ dc(,,)dpl(") )-1
dc(")d,o(")
(1 + dc(-a)dp1( -6) )-1
P3 (,,) = n..oC.)d.C.\
d.(. )dpo(I)+d.( -. dpl(-.) = dc(")npo(")
dc(6)d,o(6) dc(6)d,o(6)
p.(,,) = npoC.)d.(.~
d.(.)dpo(. )-d.( -. dpl( -.) = dc(")npo(")
dc(")d,o(")
(1 _ dc(-")dpl(-,,) )-1
dc(")d,o(")
(10.13)
Let W(,,) be a stable, proper, minimum phase transfer function such that:
(10.14)
Robust Design via Simultaneous Polynomial Stabilization 203
Since d c(") is stable dd~~» is a Blaschke product and Idd~&~» I = 1 for all
I I I I
wE [0,00). Furthermore, 't~U~» = ~:f::~ for all wE [0,00). Let ~(,,) be
any proper and stable transfer function with 1~(jw)1 ~ 1 for all w E [0,00).
We are now able to define via this norm bounded perturbation the family of
plants with divisive norm bounded uncertainty:
(10.15)
where Po(") = ~:f;l. One can easily verify that the four plants Pi(") in
(10.13) are members of this plant family. Since d c (") and do(,,) are stable,
choose the 4 ~(")'s as:
For the plant family in (10.15) in the feedback loop of Figure 10.4:
~w
we know from [36] that a necessary and sufficient condition for robust stabilis-
ability is the existence of a controller that stabilises the nominal plant Po(")
and satisfies:
204 CHAPTER 10
where 5 0(6) = 1 +l epo ' This allows us to state a condition for simultaneous
stabilUation of the four polynomials in (10.12).
Theorem 10.3 Let P(6, al) be a family of planu a6 in (10.11) with al E 011,
7Ipl(6) = 0 and dpo(6) &table. Let W(,) be a proper, minimum pha6e, 6table
tranafer function that ,atilfie, (10.14). SUPpo6e that C(6) = ~:f:~ iI a 6ta-
ble controller that 6tabilize6 PO(6) and 6atilfie6 (10.16). Then thil controller
6imultaneowly 6tabilizu the 1/1,(6) in (10.1~).
As in the case of Theorem 10.2 this result provides concrete answers to questi-
ons of existence and computation of a controller for simultaneous polynomial
stabilUation. Since the controller is computed via H00 synthesis this also gives
us the opportunity to compare the two designs.
(10.17)
;(6, a) de (6)dpo(6)
...
+ ne (6)n,o(6),+a1 ...(ne (6)np1(6) + de (6)d,1(6)) + ",
v v
;0(.) ;1(.)
... + al: ...(ne (6)npl:(6) + de (6)d,1: (6)),
...
;.(.)
= ;0(6) + a1;1(6) + a2;2(") + ... + al:;I:(") (10.18)
Lemma 5.3 places a rectangular overbound on the entire value set and identifies
the set of polynomials which contains the four vertex polynomials (at each
frequency). An alternative overbounding approach would be to overbound
each value set "edge" by a rectangle. The value set overbound will then be a
union of rectangles. Simultaneous stabilisation of the set of vertex polynomials
from each of the individual rectangles will guarantee the robust stabilisability
of each edge, and by the Edge Theorem [9] the entire family would be robustly
stabilised.
P(6,a) (10.20)
206 CHAPTER 10
';,(-)
e
... '
';2(-)
= 4>0(6) + al4>1(6) + a24>2(6) (10.21)
From Theorem 10.4 we know that a controller robustly stabilizes the plant
family if it simultaneously stabilizes 16 polynomials. If we overbound each
edge there will only be 12 distind polynomials:
Suppose tlo.at we restrid the controller to be stable and minimum phase. The
twelve polynomials in (10.22) will be stable if and only if the twelve polynomials
n e (,,)de (")1/Ii (6), 1 $ i $ 12 are stable. These twelve polynomials are closed loop
characteristic polynomials ofthe unity feedback systems as in Figure 10.1 where
the twelve plants are:
Robust Design via Simultaneous Polynomial Stabilization 207
P1(I) =
(de(I)ne(I)npo(I) + de(I)ne(I)np1(I))(1 + de (I)dp2 (I) )-1
de(I)ne(I)~(I) de(I)ne(I)~(I) de(I)~(I)
P2(I) =
(de(I)ne(I)npo(I) + de(I)ne(I)np1(I))(I_ de(I)dp2(I))_1
de(I)ne(I)~(I) de(I)ne(I)~(I) de(I)~(I)
P3(I) =
(de(I)ne(I)npo(I) + de (I)ne(I)np1 (I) )(1 + de(-I)dp2(-I) )-1
de(I)ne(I)~(I) de(I)ne(I)~(I) de(I)~(I)
P.(,) =
(de(I)ne(I)npo(I) + de (I)ne(I)np1 (I) )(1 _ de( -I)dp2(-I) )-1
de(I)ne(I)~(I) de(I)ne(I)~(I) de(I)~(I)
(de(I)ne(I)npo(I) _ de(I)ne(I)np1(I))(I+ de(I)dp2(I))_1
P5(I) =
de(I)ne(I)~(I) de(I)ne(I)~(I) de(I)dpO(I)
(de(I)ne(I)npo(I) _ de(I)ne(I)np1(I))(1 _ de(I)dp2(I))_1
P6(I) =
de(I)ne(I)~(I) de(I)ne(I)~(I) de(I)dpO(I)
(de(I)ne(I)npo(I) _ de(I)ne(I)np1(I))(1 + de( -I)dp2(-I) )-1
P1(I) =
de( I )ne( I )~( I) de( I )ne( I )~( I) de( I )~( I)
(de(.)ne(I)npo(I) _ de(.)ne(I)np1(I))(1 _ de(-I)dp2(-I) )-1
P8(I) =
de(I)ne(I)~(I) de(I)ne(I)~(I) de(I)dpo(I)
(de(I)ne(I)npo(I) + de(I)ne(-I)np1(-I))(1 + de(I)dp2(I))_1
Pg (.) = de(I)ne(I)dpo(I) de(I)ne(I)dpo(I) de(I)~(I)
(de(I)ne(I)npo(I) _ de(I)ne(-I)np1(-I) )(1 + de(I)dp2(I) )-1
PlO (I) = de(I)ne(I)~(I) de(I)ne(I)~(I) de(I)~(I)
Pll (I)
(de(I)ne(I)npo(I) + de(I)ne(-I)np1(-I))(I_ de(I)dp2(I))_1
= de(I)ne(I)~(I) de(I)ne(I)~(I) de(I)~(I)
(de(I)ne(I)npo(I) _ de(I)ne(-')np1(-1))(1 _ de(I)dp2(I) )-1
P12 (I) =
de(I)ne(I)~(I) de(I)ne(I)~(I) de(I)~(I)
(10.23)
Let ..1..(,,),..1.4(") be arbitrary proper and stable transfer functions with 1..1..(;"')1
:5 1,1..1.4(;"')1 :5 1 for all '" E [0,(0). We can now define the following plant
family with additive/divisive norm bounded uncertainty:
(10.25)
where pOe,,) = Z:f:~. Consider the plant family in (10.25) and the feedback
configuration of Figure 10.5.
Figure 10.6 shows this system with the uncertainty blocks "pulled-out"
+ M
-
b..W. 0
0 b.4W 4
Assuming that C stabilises the nominal system Po and since the largest singular
value of diag[.:1., .:14] is leu than or equal to unity, we know from [34] that this
feedback: system is stable if:
where J.' is the structured singular value. Since n c (6), dc (6) are stable n·<n)
n. ,
and 4,i~(,» are Blaschke products and have unity magnitude on the imaginary
p1
<-.i"')I
axis and in addition l±n11,0('''') d,o ,'" and 1±d,2<-.i"'
= In'l~~"'~1 d,o(,"') )I = I~I
~
on
the jw axis we have that plants Pi(6), 1 $ i $ 12 included in plant family
(10.25). Therefore any controller that stabilises (10.25) will also simultaneously
stabilise the Pi'S. We have thus proved:
n e(, )(dpa(, )d e(,) ± dpm+1 ( , )d e(,) ± ... ± dpj( -, )de ( - , ) ± ... ± dpA:(')d e(,»
(10.28)
where a lingle term appears either in the numerator or denominator of the form
npo(-8)ne (-8), 1 ~ i ~ m, or dpj(-s)de (-8),m+l ~ j ~ Ie, respectively. With
the nominal plant PO(8) = ~o' no • ~ o' and employing an additive/divisive
c: • ftc' 1'0'
uncertainty model these plants can be expressed as:
(10.29)
(10.32)
= n
I4:(jw)I (10.33)
Now, let W ..(s), Wd(s) be proper, stable and minimum phase transfer functions
such that:
Let il.. (s) and ild(S) be arbitrary proper and stable transfer functions with
la.. (jw) I ~ 1, lild(jW)1 ~ 1 for all w E [0, (0). We can now define the
following plant family with additive/divisive norm bounded uncertainty:
(10.35)
where Po(s) = ~::f:l. By construction this family includes the plants Pi, Pi.
For this plant family in the feedback configuration of Figure 10.5 we have
conditions for stability in terms of the structured singular value as in the two
parameter case. This allows us to state:
Once this was done H oo methods would be used to design the controller. At
each frequency the expreuions in (10.36) would have given lower bounds for
the uncertainty blocks. However, one would not be able to say at this point
whether this would translate into an overall less conservative design than would
be p088ible with simultaneous polynomial stabilisation.
10.3.1 Example
The simultaneous polynomial stabilisation (SPS) approach provides interesting
analytical results for polynomial family stabilisation. More than this, it can
also be used as a practical design tool for systems with parameter uncertainty.
The idea we have been promoting throughout the Chapter is to take a plant
family and overbound it by norm bounded uncertainty in a way that we can
exploit the H 00 methodology. In several cases we showed that controllers obtai-
ned from the solution of specific H oo problems do simultaneously stabilise the
required number of polynomials. However, if one has to solve an H oo problem
in order to find a controller that simultaneously stabilizes the required number
of polynomials, one might question the value of the SPS approach as a prac-
tical design tool. Just as with FIT synthesis, we can view the SPS approach
as a "robustne88 enhancement" technique. Specifically, we can employ some
design technique to obtain an initial controller which we then modify in order
to enhance robustne88 properties. To demonstrate this technique we return to
Example 3, considered in Chapter 7. The plant family is given by:
where a = (ai, a,) lies in the rectangle Oil = {(ai, a,) I - < al < Q,
a,
Q
-Q < < The positive parameter Q regulates the "size" of the rec-
Q}.
tangle. This plant family is placed in a unity feedback system and a controller
is sought that robustly stabilizes for as larg~ a value of Q as possible. We can
represent the parameter uncertainty in a multiplicative form and are able to
then overbound the family with a norm bounded multiplicative perturbation
214 CHAPTER 10
. t erms 0 f th e Wetg
m . ht w:2 (8 ) -_ + 1)(.
.3 +
(. 1.2S.++lOr
9 In part'lC ular, WI'th t his w:2 (8 )
we have that:
and wE [0,00)
We can account for an arbitrary a by multiplying by this factor both sides of
this expression. The nominal system is minimum phase as would be required
by our procedure. We then employed a Nevanlina-Pick technique to compute
the controller C(8) = ::f:l !
= 1.213t.~1~5~~(: 1~i4545) which guaranteed robust
stability for a~z = 1.273. Note that the controller is minimum phase as would
be required for this case.
Employing the SPS approach for this example would require the simultaneous
stabiliJation of the twelve polynomials:
At this point one could take the H oo controller and modify it by introducing
two or three parameters and use graphical techniques to determine regions for
Robust Design via Simultaneous Polynomial Stabilization 215
which all twelve polynomials are stable for as large an a as possible. For low
order controllers and systems this may be feasible but one has to say that FIT
synthesis would provide a much better alternative. In this example, we simply
multiplied the controller by a single gain and performed this procedure. We
concluded that the controller:
P(6, 8, e, (, g)
npO(6) + npe(6) + 6npo(6) (10.39)
dpo(6) + dpe(6) + 6dpo(6)
where
216 CHAPTER 10
= el + e262 + e364 + .
It + 1262 + 1364 + .
= 91 + 9262 + 93 64 (10.40)
with 8(dpo) = n, 8(dpe) < n, 8(dpo) < n - 1, 8(npo) < n, 8(r&pe) < n
and 8(r&po) < n - 1. If this interval plant family is placed the unity feedback
configuration of Figure 10.1, the closed loop characteristic polynomial will be:
a+(6) =
a-(6) =
e+(6) =
e-(6)
/+(6) =
r(6) =
9+(6) =
9-(6) (10.41)
Theorem 10.T There ezuu II controller C(,,) thllt Itllbilize, the intervtJlfllmilJl
in (10.39) ifthu controller ,imulttJneoualJl "tllbilize" the 6~ polJlnomillu ;0(")+
,p.(,,) + X;(,,), 1 ~ i ~ 8, 1 ~ j ~ 8.
will limit our discussion to asymptotic tracking and noise attenuation. In this
Section we consider a plant as in Figure 10.7 with no uncertainty and state
conditions for nominal performance synthesis.
-
r + C(,,) P(,,)
y
Let us first consider the problem of nominal tracking (and nominal stability).
Suppose that the plant is described by a strictly proper transfer function P(,,) =
zf:~ where 8(dp(")) = n. From Chapter 8, we know that performance would
be guaranteed (IW1(jw)S(jw)1 < 1 V wE [0,00)), if a controller can be found
that stabilizes the polynomial family:
(10.42)
for all 0 ::; rl ::; 1 and Ql E [O,2?l'), q,(,,) being the closed loop characteristic
polynomial. The uncertainty r1ei a1 corresponds to the unit disc. The key
element that allowed us to state simultaneous polynomial stabilization results
was our ability to overbound straight line segments in the complex plane with
rectangles. The disc uncertainty present in (10.42) prevents us from directly
implementing this scheme. We are therefore forced to first overbound the disc
by a polygon and then apply the rectangular overbounding. Of course, both
these procedures will introduce conservatism in the design. This conservatism
can be reduced by employing tight polygonal overbounds of the disc. If the
polygon edges are "short" this will also help the second stage when the rectan-
gular overbound is employed. For simplicity (at the expense of conservatism)
we overbound the disc by a square:
Robust Design via Simultaneous Polynomial Stabilization 219
1m
1
1/
'-
"
./
1 Re
(10.43)
Recall from Chapter 4 that the "-,, operation on a complex polynomial stands
for: 4)1(6) = fn( _6)n + fn-1( -6) + ... + fo, It is the complex conjugate of
Ii, 1 ~ i ~ n. At each frequency the value set ~s(jw, 011 1 , 0112 ) is a rectangle.
Now it is important to realise that because 4>S2(6) = j4>Sl(6), at each frequency
w the value set ~s(jw, 011 1 , 0h) will have corners that are formed by the sum
of an "actual" vertex of the rectan$lc: bl 4>s(jw) + bl 4)Sl(6) and a "fictitious"
vertex of the rectangle b24>S2(6) + b24>S2(6) and vise versa. This implies that
the four vertices of the '¢ S (jw, 011 1 , 011 2 ) rectangle come from the set of eight
vertices:
Now if these eight polynomials are stable the t$S(6, 61 , b1, 62 , b2) family will be
stable [11] and consequently the 1/1S(6, 61 , 62 ) and 1/1S(6, rl, Ql) families will be
stable. Nominal performance will thus have been achieved. We have proved:
Suppose now that in addition to tracking we also have the noise attenuation
specification: IW2 (jw)T(jw) I < 1 V w E [0,00). We know that this will be
achieved if a controller can be found that stabilizes the polynomial family:
{or all Q2 E [0,211"), r2 E [0,1]. Let us again overbound the unit disc by a
square, in other words we replace r 2 ei a2 by the complex uncertainty b = 61 +
j6 2,-1 ~ 61 ~ I, -1 ~ 62 ~ 1. Clearly, if 1/1T(6, b) = b1l.v2(") n c (")npo(6) +
dw2(6)(dc (6)d,,0(6)+n c(6)npo(6», is stable {or b in the square then 1/1T(6, r2, Q2)
will be stable. Let ;TO(6) = dw2(6)(dc (6)d,,0(6) + nc (s)npo(6)), ;Tl(6) =
1l.v2(6)nc (6)npo(6), ;T2(6) = #T1(6), then:
(10.46)
Following the exact same procedure as above we generate the following 8 po-
lynomials:
Robust Design via Simultaneous Polynomial Stabilization 221
and state a result for the multiobjeetive problem of tracking and noise attenua-
tion.
(10.48)
222 CHAPTER 10
bnwl(S)dc(s)dpo + ~l(s)dc(s)dpo +
d,.,l(s)n c(s)(n"o(s) + a1n,,1(s) + ... + akn"k(s)))
(10.50)
is stable for b in the square and ft E no, then ,ps(s, ft, rl, ot} will be stable. Let
4>so(s) = d,.,l(s)(dc(s))dpo(s) + nc(s)n"o(s)), 4>Si(S) = d,.,l(s)n c(s)TLpi(s), i ~
i ~ Ie, 4>Sk+1(S) = n,.,l(s)dc(s)dpo(s), 4>Sk+2(S) = i4>Sk+I(S), and
.(iJS(s, ai, al," .ak, ak, b1, b1, b2, b2) = 4>so(s) + al4>sl(s) + al~Sl(S) ...
bl 4>sk+l(S) + b1, ~Sk+1(S) + b24>sk+2(S) + b2~Sk+2(S)
Theorem 10.10 (rob1Ut tUymptotic tracking) Let P(s, 8) be the strictly proper
family in (10.-18). A controller C(s) ezisu which guarantees rob1Ut tUymptotic
tracking (and rob1Ut stability) if this controller simultaneo1Uly stabilizes 8( A: +
1)21: polynomials.
An identical result can be stated for the feedback system in Figure 10.7 with
a family of plants with affine parameter uncertainty in the denominator of the
plant transfer function and robust noise attenuation as a performance requi-
rement. In the case of more complicated plant family uncertainty or multiob-
jective sensitivity and complementary sensitivity requirements similar results
can be stated. Consider the case when the plant family has affine uncertainty
in both numerator and denominator and robust tracking and robust noise at-
tenuation are the design requirements. Then 1/Is(s, 8, b) and 1h(s, 8, b) will
have multiaffine uncertainty in their coefficients. In order to use the rectangu-
lar overbounding techniques presented in this Chapter, the multiaffine terms
would need to be overbounded by affine terms. This would be done by first in-
troducing a polygonal overbound for the disc uncertainty (say a square). Then
the vertex plants of the corresponding polynomial family would need to be
identified and line segments introduced connecting them. These line segments
would form the edges of the convex hull of value sets in the complex plane. The
rectangular overbound procedure would then be applied to each of these edges
which would lead to a sufficient condition for robust performance. Clearly, all
this overbounding would lead to conservative designs.
Over the last twenty years extensive work has been done on multivariable sy-
stems and fundamental results have been produced for both robust analysis and
synthesis [63, 38, 10, 54, 67, 42, 55]. For the very successful design methodolo-
gies such as H oo , LQG/LTR,~, widely available software exist in MATLAB. In
the context of this very well developed theory, the material presented here can
only be characterised as "foundational". More work needs to be done in order
to completely develop the methodology. However, we felt that is was important
226 CHAPTER 11
to introduce these new ideas in order to help motivate future research on the
subject. We target multivariable systems with parameter uncertainty and focus
on robust 1>-stabwation and specific problems in robust performance design.
(11.2)
where P(6) has entries which are rational functions in 6. Using polynomial ma-
triz fraction repre,entation6 [15], one can express the transfer function matrix
in terms of a right matriz fraction repre6entation :
(11.3)
where NJW (6), DJW (8) are m x l,l x l polynomial matrices respectively, or a left
matriz fraction repre6entation :
(11.4)
involved one must distinguish between left and right multiplication. Polyno-
mial matrices whose determinant is a nonzero constant are called unimodular.
With elementary row operations [15] one can compute a greatest common right
divisor of N",.(6), D",.(")' By multiplying matrices N",.(,,), D",.(6) on the right
by the inverse of this greatest common right divisor one can obtain a right re-
presentation N",.(6), D",.(6) which is right coprime, (i.e. the greatest common
right divisor of N",.(6), D",.(,) is a unimodular matrix). One can also find a
unimodular matrix L(6) [15], such that D",.(6)L(,,) is column reduced (i.e. its
highest column coefficient matrix is invertible).
A similar procedure can be carried out with a left matrix representation N,,(,,),
Dpl(6) and elementary column operations to obtain a left representation N,,(,,),
D,'(6) which is left coprime (i.e. the greatest common left divisor of N,'(6),
Dpl(6) is a unimodular matrix). One can find a unimodular matrix M(6) [15],
such that M (6 )D" (,,) is row reduced (i.e. its highest row coefficient matrix is
invertible). The transfer function matrix is proper (,trictlll proper) if all its
entries are proper (strictly proper) rational functions. In particular, the transfer
function matrix of a system as in (11.1) is clearly proper and if D = 0, strictly
proper.
Let the normal rank of a matrix with entries that are rational functions in 6
be its rank over the field of rational functions. Let P(,,) = N",.(")D,,,.(,,)-1
= Dpl(6)-1 N,,(,,) be right coprime and left coprime representations of P(6)
respectively and r = min(m,l) be the normal rank of P(,,). The roots of
<let D",.(6) and <let D,,(,,) are identical [15] and are called the pole, of P(,,).
The zero, of P(s) are those complex" that make the rank of N",.(,,) less than
r. These are identical to complex numbers" that make the rank of N,,(,,) less
than r [15].
Consider a two link planar robotic manipulator with prismatic joints operating
in the horizontal plane (no gravity). One can think oHhis as a model for an
X - Y Plotter. A top view is shown in Figure 11.1.
Link I, has mass m1 and length l1 and is constrained to move along the y-
axis. Let h 1 (t) be the position ofits "tip" relative to the "base". Assume that
some actuator applies some force h (t) on link 1 and that there is a friction force
acting on it which is equal to d1h. 1(t) where d1 is some constant coefficient. Link
2, has mass mz and length lz and moves along the z-axis. Let hz(t) be the
228 CHAPTER 11
f1 y
h h
1
1 2
x V
m
2 t
Figure 11.1 Two Link Planar Manipulator with Priamatic Jointl
position of its tip. Assume that some actuator applies some force h(t) on link
2 and that there is a friction force acting on it which is equal to d2 h.2 (t) where
d2 is some constant coefficient. Since the motion in the :I: and y directions
is "decoupled", one need not resort to the Lagrange formulation in order to
obtain the equations of motion. One can simply apply Newton's law of motion
in the :I: and y directions to obtain:
d2
+ m2) dt 2 hl(t) =
d2
m2 dt 2 h2 (t) =
If we let:l:l =
h l ,:l:2 = Zl,:l:3 = h21 :1:4 = Z31 these two equations can be
expressed in state space form as:
1 0 0 0
[~
II, 1
-(m,+m2) 0 0 (m,+m2)
=
1. ]
:ic(t) ] x(.) + [ u(t),
0 0 1 0
0 0 _h. 0
m2
y(t) [~ 0
0
o 0] x(t)
1 0
(11.5)
FIT for Robust Multivariable Design 229
The system transfer function matrix has two inputs and two outputs and is
expected to be diagonal:
(11.6)
A consequence of the diagonality is the ease of generating right and left matrix
fraction representations. In particular, we can write:
P(6)
-
r +
-
+
e
0
u
P
l'
This is one specific system interconnection and one can expect that these can
be investigated using a similar approach [15] to that introduced in Chapter 2
for SISO systems. We again insist that systems be well formed .
Definition 11.1 A dynamic ,yltem with tranlfer function matriz P(,) iI cal-
led well formed if each entry of P(,) iI a proper rational function.
1'2
1'3
The only difference between this interconnection and a similar one presented in
Chapter 2 is that system blocks are represented by matrices rather than scalars
and the signal paths are vectors rather than scalars. For this interconnection
we have that:
(11.8)
r = [r1' r2, r3]', U= [U1,U2,U3]', and Y =[Y1' Y2, Y3]', where'" tI denotes
matrix transpose. Let the constant interconnection matrix E be:
U =r + Ey (11.9)
One has to insure that the overall system is well formed. There are two transfer
function matrices associated with this interconnection. The one from exogenous
signals r to u, denoted as H. r , and the one from r to y denoted as H,r'
Definition 11.2 The interconnected ,yltem in Figure 11.3 iI called well for-
med if the tmnlfer function matrice, H. r and H,r are proper.
232 CHAPTER 11
11.3 STABILITY
The notion of stability defined in Chapter 2 for SISO systems can be generalised
to the multivanable case. Recall that we presented two definitions, one for
systems described by input output relations and one for systems in state space
form. A multivanable system which is initially at rest is called BIBO ,table
if for every input which is a bounded function of time, the system output is
also a bounded function of time. Since inputs and outputs are vectors, a signal
is bounded if its norm is a bounded function of time. The Euclidean norm,
vi
IIxli = zt + z~ +... + z~, is a good example. In the SISO case conditions
for BIBO stability were given in terms of the pole locations of transfer functions.
The generalisation of Theorem 2.1 for multivanable systems becomes [18]:
Theorem 11.1 Let P(,) = NJW(,)DJW(,)-l be proper, where NJW (,), DJW(')
are right coprime. The ,y,tem P iI BIBO Itable if and only if the root. of
det DJW(') lie ,trictly in the left half complez plane.
One could have used a left coprime matrix fraction representation and related
BIBO stability to det D,,(,), which is a polynomial with identical roots.
The concept of Lyapunov stability remains exactly the same as it was stated
for systems in state space form with no input. The aero equilibrium state of
x(t) = A x(t), is asymptotically stable if the eigenvalues of A lie strictly in the
left half complex plane. The relationship between BIBO stability and Lyapunov
stability is immediate, (for the definition of minimal realisation see [18]).
If the region of stability is not the left half complex plane but some other region
'D, one deals with 'D-stability in exactly the same manner as for SISO systems.
For interconnections of multivariable systems one can again extend the no-
tions of BIBO stability and Lyapunov stability. An interconnected system
which is initially at rest is BIBO stable if for any bounded exogenous input
ri(t), all system inputs Uj(t) and outputs Yj(t) are bounded. Assume that
FIT for Robust Multivariable Design 233
o
o
o o
Let N r (_) be the the block diagonal matrix containing all the Nir(6) matrices:
o
o
o o
Let NI(6) be the block diagonal matrix containing all the Nil(6) matrices:
o
o
o o
Consider the feedback interconnection of Figure 11.2. Let the plant be de-
scribed by a the proper transfer function matrix P(s) with left coprime repre-
sentation D,r(6)-1 N,r(6) where D,,(6) is row reduced. Let the controller be
described by the proper transfer function matrix C( 6) with the right coprime re-
presentation C(6) = Ncr (6)Dcr (6)-1 where Dcr (6) is column reduced. Think
of C(6) as subsystem 1 and of P(6) as subsystem 2. We then have:
E= [10-1]
0 '
with the same determinant as Dg (6). It is then clear that det Dg (6) dif-
fers from det (Dp /(6)Da (6) + Np /(6)Na (6)) by a nonzero constant. There-
fore, in examining the stability or V-stability of the standard feedback in-
terconnection, one can work with det ~( ..) where ~(6) = (Dp /(6)D a (6) +
Npl(6)Na (6)). If the plant is given by a right matrix fraction representa-
tion Npr(6)DJI"(6)-t and the controller by a left matrix fraction representation
Dc:l(6)-t Nc:l(6) then one can show that the closed loop poles are the roots of
det(D c/(6)DJI"(6) + Nc/(6)NJI" (6)).
11.4 PERFORMANCE
Just as in the SISO case, stability or more generally V-stability is only the first
desired performance specification for a multivanable feedback system. Addi-
tional performance requirements are also imposed that address other aspects of
dynamic behavior. Some of these requirements involve time domain specificati-
ons while others are posed as frequency domain constraints. As we saw in single
input single output systems, some time domain specifications can be approxi-
mately expressed as frequency domain objectives. Many, such as asymptotic
tracking and noise attenuation are generalizations of SISO concepts. Others,
such as decoupling, are inherent to the multivanable character of the feedback
system. Let us first discuss asymptotic tracking and noise attenuation and see
what kind of conditions will guarantee this behavior for multivariable systems.
Consider the feedback system of Figure 11.4 where the plant is described by a
strictly proper transfer function matrix P(,,) and C(6) is a proper controller.
236 CHAPTER 11
--
r +
+
e
C(,) P(,)
y
+ '7
++
The transfer function from the reference input vector r to the error vector e is
given by the sensitivity transfer function 5 = (1 + pC)-l. If the system
is stable, '7 = 0 and the reference input is a vector of sinusoids of some
frequency w, in steady state all loop signals including e will tend to sinusoidal
signals. In particular, if r(t) = M IIII , where R is some complex vector,
then e(t) asymptotically will tend to S(jw)r(t) [15]. Therefore, if we desire
good tracking the "operator" S(jw), the sensitivity, must be "small" in the
frequency range of interest. Since 5 is a matrix, one can interpret "smallness"
in a variety of ways. A very appropriate interpretation is to make the ,pectral
norm u(S) of 5 small. The spectral norm u(A) of some complex matrix A is
defined to be the square root ofthe largest eigenvalue of A* A, where A* stands
for complex conjugate transpose. This has been a particularly useful measure
of "sise". Specifically, let Wll (,), W12 (,) be a strictly proper rational matrices
with stable and minimum phase entries which reflect desired requirements, then
asymptotic tracking will be guaranteed if:
The transfer function from the noise vector '7(t) to the output vector y(t) is
-T(,) = -PC(l + pC)-l. If the system is stable, for the same reasons as
above, the effect ofnoise on the output y(t) will be small if the operator T(jw),
the complementary sensitivity, is "small" over the frequencies of interest. Let
W21(')' Wz z(') be proper rational matrices with stable and minimum phase
entries chosen to shape T(,) appropriately. Then noise attenuation will be
achieved if:
In the multivariable case one can choose to define the "sise" of an operator
in several ways. Suppose we are interested in asymptotic tracking where our
goal is to make the error small. An alternative way of achieving this goal is to
insist that every entry 6i;(jW) of S(jw) is "small" over the appropriate range of
frequencies. In particular we may express the asymptotic tracking requirement
as:
(11.12)
where Wli;(6) is the (i,j)th element of some matrix W 1(6) which is strictly
proper and has stable and minimum phase entries. Similarly, in the case of
noise attenuation where our objective is to keep T(jw) small, we can insist that
every entry t.;(jw) is small in the desired frequency range and require that:
where W2i;(6) is the (i,j)th element of some matrix W 2(6) which is proper and
has stable and minimum phase entries. The idea of placing bounds on individual
entries of the sensitivity and complementary sensitivity transfer functions has
been used in the past [44, 66].
of the objects being picked up may lie in a range. This implies that the tran..
fer function matrices of plants will involve parameters. In this Chapter we will
assume that the plant family has parameter uncertainty that can be expressed
in a polynomial matrix fraction representation. Specifically, the plant transfer
function matrix is strictly proper and can be described by the polynomial left
matrix fraction representation:
P(6,a) (11.14)
where the polynomial entries of Np l(6 1 a), Dp l(6, a) have coefficients that are po-
lynomic in a and where a E 0 4 , some hypercube. Such parametric uncertainty
could also have been defined for plants with polynomial right matrix fraction
representations P(6 1 a) = Npr(6 1 a)Dpr(6 1 a)-I. One can easily envision how
such representations can be obtained. For some systems it may be that dy-
namic descriptions are already given in this form. This is the case with the
two link robotic manipulator example presented earlier. Another example is a
system in state space form where C = I and the" A" and" B" matrices have
entries that are polynomic in a. If the given plant matrix has rational function
entries which are polynomials in " with coefficients that are polynomic in a,
one can form the left or right polynomial matrix representation by carrying out
operations over the field of rational functions in a. This will result in a polyno-
mial matrix fraction representation where the coefficients are rational function6
in a. Using Taylor Series expansions on each of these terms will provide (at
any degree of accuracy) expressions that are polynomic in a.
of our polynomial value set approach. The key element of our polynomial ap-
proach to robust control design is the manipulation or shaping of polynomial
value sets. Specifically, in FIT synthesis for robust stabilization, we try to "fit"
or "squeeze" the value set in specified sectors at particular frequencies. In the
simultaneous polynomial stabilization approach to robust stabilization we do
more or less the same thing but are limited to sector mes that are 90 0 wide.
(see Lemma 6.1). We again manipulate polynomial value sets. This process
would have been greatly facilitated, if there were "enough" degrees of freedom
in the controller. One possibility could have been the existence of a controller
with enough degrees of freedom so that we could have been able to arbitrarily
pole assign the nominal system and at the same time eliminate the presence
of the uncertainty from the characteristic polynomial. This would have forced
the polynomial value set to be a single point at each frequency and this "ro-
bust pole assignment" method would have been a very effective technique for
robust V-stabilization. As nice as this sounds, in the SISO case there cannot
be enough degrees of freedom for such an approach. We explain why this is
true 1 • Let P(6, al) be some arbitrary strictly proper plant family with transfer
function:
fapO(6) + alfapl(6)
dpo(6) + aldpl(6)
where a E 0/1 and dpo(6) is monic with 8(dpo) = n. Suppose that we use a
proper order q controller C(6) = ::f:~ with monic dc (6), in a standard unity
feedback configuration. The closed loop characteristic polynomial is
2. Eliminate the parameter uncertainty from 4>(s, ad, i.e. make 4>l(S)
de(S)dpl(S) + ne(s)npl(S) = o.
The second condition will make the value set 4>(s, ( 4 ) be a single point at each
frequency regardleu of the size of 0 4 • One can show [32] that this is impossible
to accomplish. A fact which can be easily proved: If npl(S) = 0 then we
must have de(s) = o. If dpl(S) = 0 then we must have ne(s) = O. In
either case we cannot achieve our design objective. Suppose then that de and
n e are nonzero. If 8(dpd > 8(npd then we must have an improper controller
in order to satisfy the second condition. Suppose then that 8(dpd ~ 8(npd.
Then de ()S -- - fi.(')"p'(')
d,,(,)
-
- -
".(,)n,,(,)
d,,(')
h
were - ()
npl S, d-pi () •
S are copnme.
Since dpl(S) divides the product iipl(S)dpl(S) it must divide ne(s), ie ne(s) =
dpl(S)Y(S). So de(s) = Y(S)iipl(S). But this implies that:
v =y +
R 1
where we auume that the value of the resistance R is only known to lie in some
range. With the capacitor voltage Vc and indudor current iL as state variables
the dynamic equations for this system in state space form are:
Suppose that we use a first order controller in a standard unity feedback con-
figuration described by the left matrix fraction representation:
1 0 2 0 2 0
o 0 0 0 2 0
002 020
o 1 0 202
000 002
000 202
242 CHAPTER 11
be full column rank. It is clear that this condition is satisfied and that the
"robust pole assignment" scheme presented above can be carried out. In par-
ticular, the controller
9
C(I) -I -
2
3]
will make the closed loop characteristic polynomial equal to ,,3 + 6,,2 + 116 + 6.
Notice that the parameter R has been eliminated. This implies that for this
multivanable system controllers can be constructed that robustly 1>-stabilise
for any range of parameter values.
It seems rather remarkable that in this example one is able to guarantee robust
1>-stabiliJation with an "infinite" stability margin. It would be great if this
could be accomplished for an arbitrary multivanable system. This however, is
not true because in addition to "enough" degrees of freedom in the controller
the system itself must have an appropriate parameter uncertainty structure.
What can be shown [32], is that robust pole assignment is a "generic" property.
Even though it is good to have "generic" results, they are more of theoretical
significance. In applications one can easily encounter systems that do not have
the required parameter uncertainty structure. Therefore, one should develop in
addition "general purpose" robust 1>-stabilisation methods that can be applied
to any multivariable system.
where
FIT for Robust Multivariable Design 243
diag(6",,-1)N,hc(a) + Nplot(6,a)
diag(""")D,hc + D,'ot(6,a)
(11.15)
where
(11.17)
For this plant family the polynomial <p('" a) has coefficients that in general are
polynomic in a and multiaffine in X. Therefore, the problem of 'V-stabilization
of multivariable systems with parameter uncertainty is expressed in terms of
the 'V-stabilization of a single polynomial family. It is evident that this pro-
blem can be attacked using the Finite Inclusions Theorem in a manner similar
to that presented in Chapter 7. We will first assume that by introducing addi-
tional uncertain parameters (and consequently overbounding), the polynomial
coefficients are made multiaffine in a. This allows one to work with value set
vertices as a consequence of the Mapping Theorem and leads to computatio-
nally efficient synthesis algorithms.
The basic idea behind the algorithm remains the same: Pick an initial control-
ler, and an appropriate set of points and sectors that satisfy the conditions of
the Finite Inclusions Theorem (see Chapter 6). Expand the uncertainty range
and solve for a new controller that fits the new value sets in the sectors. Center
the value sets (i.e. pick new frequencies) and repeat the process. The only
"structural" difference with the SISO case is the fact that the coefficients of
a general <p('" a) depend on controller coefficients in a multiaffine manner rat-
her than in an affine manner. Consequently, the inequalities that need to be
!IOlved in a FIT based procedure are multia.ffine in the controller parameters.
Existing FIT based 'V-stabilization algorithms must therefore be modified to
address this issue. However, it is important to point out that if all but one
(say the it"), of the columns of X take specific values then the coefficients of
<p('" a) become affine in the remaining controller parameters Xi and so do the
FIT for Robust Multivariable Design 245
STEP 1: Let X(l) E R d and O~l) C 0 .. be such that 4>(8, O~l), X(l))
is V-stable and set j:=1.
STEP 2: Determine m(j) ~ 1 sectors sin, 1 ~ Ie ~ m(j), and points 8~)
along the boundary of V such that 4>(B~),EztO~j>,X(j)) c sin
By FIT, 4>( 8, O~), X(j») is V-stable. Each 8~) should
roughly center (angularly) the set 4>( 8~), E:z:tO~), X(j») in sin.
STEP 3: Choose a slightly larger set O~+l) :::) o~).
STEP 4: Compute a new vector of controller parameters X(i+ 1) such
that 4>(B~), E:z:t O~+l), X(i+ 1») c sin for all Ie. This is
equivalent to solving a system of multiaffine inequalities in
X(j+l). Fix all columns of X(i+ 1) but one at current values.
Solve the corresponding set of linear inequalities. Iterate over
the choice of column. If no solutions are found to this system
of inequalities, return to Step 3 and choose a smaller O~+l).
STEP 5: Let j := j + 1, and if O~) :::) 0 .. , stop; otherwise, go to Step 2.
(11.18)
with Npl('" a) and Dpl('" a) as in (11.15). The plant family is in the unity
feedback configuration of Figure 11.4. The controller is described in the right
matrix fraction representation:
5(", a) = (1 + P(",a)C("n- 1
FIT for Robust Multivariable Design 247
where Wlij(6) is the (i,i)th element of some matrix W1(6) which is strictly
proper and has stable and minimum phase entries. Since the plant transfer
function matrix is m x L there will be m 2 such conditions.
T(6,a) =I - 5(6,a)
Its diagonal entries are ~i(6, a) = 1 - 6ii(6, a) and the off diagonal entries are
~j(6, a) = -6ij(6, a). They too can be thought of as rational functions with
denominator ~(6, a). The robust noise attenuation requirement is expressed as:
where W2ij(6) is the (i, j)th element of some matrix W 2(6) which is proper
and has stable and minimum phase entries. There will be m 2 such conditions.
One can immediately see that the weights W 1(6) and W 2(6) cannot be chosen
independently because the diagonal entries 6ii(jW, a) and 1 - 6ii (jw , a) are
very related. In particular, both cannot be made "small" at the same frequency.
Fortunately, in many cases tracking and noise attenuation requirements are
placed at sufficiently disjoint frequency ranges.
248 CHAPTER 11
Lemma 11.1 The entriu of 5(", a) and T(", a) are rational junctiof1.8 in "
who"e numerator and denominator polllnomiaU are multiaffine ezpre",iof1.8 of
the controller parameter".
Proof: We know that 5(8, a) = Dcr (8)9(8, a)-1 Dp l(8, a). Write 9(8, a)-1 =
··k~~~»)) where adj(·) denotes adjoint of a matrix. This implies that:
5(8,a)
T(8, a)
The elements of the itll. column of 9(8, a) are affine expressions of the controller
parameters from the itll. column of X. As a result 4>(8, a) = det9(8, a) is a
polynomial in 8 with coefficients that are multiaffine expressions in controller
parameters. This also implies that the elements ofthe itll. row of adj( 9( 8, a))
do not contain controller parameters from the it'" column of X. These elements
are multiaffine expressions of controller parameters from columns of X other
than the itll.. Furthermore, when the product D cr (8)adj(9(8, a)) is formed, the
elements of this matrix are multiaffine expressions of the controller parameters.
Multiplying this product on the right by the matrix Dpl (8, a) does not change
the structure. As a result the elements n,ij(8, a) are polynomials in 8 with
coefficients that are multiaffine in controller parameters. Since T(8, a) = I -
5(8, a), the same holds for the numerator polynomials ntij(8, a). 0
Consider now any one of the m 2 conditions for robust tracking. If Wlij(8)
"vii·' it can be equivalently stated as:
d... H; ,
n.1ij(8)n,ij(jw, a) I< 1 V [ ) n.
I dwlij(8)4>(8, a)
w E 0,00, a E (11.21)
0('
\':7 JW
) n.lij (jw)n'ij (jw, a) 1< 1 \J
v W E [0 ,00,
) a En • (11.22)
1 dwlij(jW)4>(jw, a)
FIT for Robust Multivariable Design 249
where 9(,,) is any stable proper transfer function with 19(jw)1 ~ 1 for all
w E [0,00). In other words, for no plant in the family and no frequency can we
have:
° °
where ~ r ~ 1 and ~ a < 211". This is equivalent to saying (since d..,1i;(")
and ;('" a) are stable) that the polynomial:
(11.24)
is stable for all a E [0,211"), r E [0,1] and a E 00' then the robust performance
requirement corresponding to the (i, j)th. sensitivity transfer function element
will have been met.
One can express the performance requirements for all the other sensitivity trans-
fer function elements in a similar manner. The robust tracking requirement can
therefore be expressed as a simultaneous polynomial family stabilisation requi-
rement. In exactly the same manner one can express the robust noise attenua-
tion requirement as a simultaneous polynomial family stabilization problem
where polynomial families are of the form:
Just as in Chapters 8 and 9, we need to overbound the disc uncertainty rei a as-
sociated with performance by some polygonal overbound and in the following
algorithm we choose the square overbound Ob,. Now, each such polynomial
family will in general have coefficients that are polynomic in the plant uncer-
tain parameters a and multiaffine in the controller parameters X. In order to
250 CHAPTER 11
STEP 1: Let XCl) E R d and n~~) c nab be such that 4>i(6, n~~), XCl»)
are stable and set j:=1.
STEP 2: Determine m(j) ,-
> 1 sectors S~i)
,A:'
< lc -
1 - < m(j)
,,--
1 <i <l
and frequencies w~~) along the jw axis such that
Cj ) Eztn(j) X(j») c SCi) By FIT ,J,·(6 n(j) XCi») are
,J,·(]·wiA:'
.", tb , iA: • ,.", , lib'
stable. Each Wi~) should roughly center (angularly) the set
Cj
.", d.' Eztn ab') xCj») in ~j)
,J,·(l·W~j) ,A:'
STEP 3: Choose a slightly larger set n~+1) J n~). First this should
affect the b parameters. When (if) n~i) J nb" the
enlargement in the b-direction terminates and the enlargement
in the a-direction commences.
STEP 4: Compute a new vector of controller parameters X(j+1) such
that 4>i(jW~~), Ezt n~+l), XCj+1») c S~~) for alllc and i.
This is equivalent to solving a system of multiaffine inequalities
in XCH1). Fix all columns of X(j+1) but one at current values.
Solve the corresponding set of linear inequalities. Iterate over
the choice of column. If no solutions are found to this system
of inequalities, return to Step 3 and choose a smaller n~+l).
STEP 5: Let j := j + I, and if n~~ J nllb, stop; otherwise, go to Step 2.
It should be clear that this algorithm can be employed for both single objec-
tive and multiobjective robust performance synthesis. This formulation and
algorithms can also be used to solve a number of other robust multivariable
synthesis problems. One such is the problem of RobU6t Decoupling discussed in
the next Section.
FIT for Robust Mtdtivariable Design 251
The problem ofrobust decoupling for systems with parameter uncertainty in the
sense of making the closed loop system diagonally dominant can be attacked
using FIT based methods. We can pose the following decoupling problem:
Let the given plant family be P(8,a) = P(8,a) = Dp l(8,a)-1Np l(8,a)
with Np l(8, a), Dp l(8, a) as in (11.15) where a E 0 4 It is in the standard unity
feedback configuration of Figure 11.2. The closed loop transfer function is
T(8, a) = P(8, a)C(8)(1 + P(8, a)C(8»-1. Let W 2(8) be a proper transfer
function matrix with stable, minimum phase entries. It is chosen to achieve
diagonal dominance over the appropriate frequency range (i.e. its off diagonal
entries are "larger" than the diagonal entries). Find (if possible) a controller
C(8) such that robust stability is guaranteed and:
We have just presented a formulation that will allow FIT synthesis to be used for
the solution a number of robust performance design problems for multivariable
systems with parameter uncertainty. This was done without compromising the
multivariable structure of the plant family nor the structure of the controller.
252 CHAPTER 11
[49] R. D. Kaminsky, "A Novel Approach to the Analysis and Synthesis of Con-
trollers for Parametrically Uncertain Systems," PhD Thesis, University of
Massachusetts Amherst, 1993.
[50] V. L. Kharitonov, "Asymptotic Stability of an equilibrium position of a
family of systems of differential equations," Differential'nye Uravneniya,
Vol. 11, 1978, pp. 2086-2088.
[51] P. P. Khargonekar, A. Tannenbaum, "Non-Euclidean Metrics and the Ro-
bust StabiJUation of Systems with Parameter Uncertainty," IEEE Trani.
on AC, Vol. 30, No.10, 1985, pp. 1005-1013.