Professional Documents
Culture Documents
and Design
-~-'
Li near ~
and De~
HRW
Series in
¡
:.
Electrical and
Chi-Tsong Chen
Computer Engineering
Professor, Department
State University of Nev
M. E. Van Valkenburg, Series Editor
HOLT, RINEHART /J
Chi-Tsong Chen
g
Professor, Department of Electrical Engineering
State University ofNew York at Stony Brook
ANALYSIS
=:SIGN
;HINE OESIGN
~IRCUITS: THEORY AND
DESIGN OF FEEDBACK
HODS IN CONTINUOUS
~OMMUNICATION
US ANO DISCRETE
Id Edition
)ERN APPROACH
L ENGINEERING:
NIC CIRCUITS
;N
Requests for permission to make copies of any part of the work should be
mailed to: Permissions, Holl, Rinehart and Winston, Inc., Orlando,
Florida 32887
Chen, Chj-Tsong.
Linear system theory and designo
Bibliography: p.
Ineludes indexo
cluding
without
d be
Preface XIII
Glossary of Symbol:
Chapter 1 Introe
1-1 ".
j
1-2 l
Chapter 2 linea
2-1 1
2-2 1
2-3 !
2-4 1
2-5 S
2-6 I
2-7 F
Preface Xlll
Chapter 1 Introduction 1
2-1 Introduction 6
Change of Basis 17
Series 54
vii
viii CONTENTS
Problems 62
4-5
3-1 Introduction 70
Linearity 73 Equa
Causality 76
Relaxedness 77
5-1 ]
Time Invariance 80
5-2 ]
Transfer-Function Matrix 81
5-3 e
3-3 The State-Variable Description 83
Dynamical Equations 86
Linearity 87
Time Invariance 89
Transfer-Function Matrix 90
5-4 (
Analog and Digital Computer Simulations ofLinear Dyna
mical Equations 91
3-4 Examples 94
Problems 125
*5-8 (
*5-9 (
F
Chapter 4 linear Dynamical Equations and Impulse-Response
Matrices 133
Chapter 6 Irredu
4-1 Introduction 133
Identi1
4-2 Solutions of a Dynamical Equation 134
CONTENTS ix
Equations 252
Functions 253
Method 276
Realization ofN(s)D-1(s), Where D(s) and N(s) Are Not Chapter 8 Stab
Right Coprime 282 8-1
Column Degrees and Controllability Indices 284 8-2
Observable-Form Realization 285
7 -1 Introduction 324
Method I 341
Method II 345
9-4 S
Method III 347
Nonuniqueness of Feedback Gain Matrix 348
Assignment of Eigenvalues and Eigenvectors 351 9-5 D
Effect on theNumerator Matrix of G(s) 352
Computational Problems 353
CONTENTS xi
atrix 245
7-4 State Estimators 354
tion 249
Full-Dímensional State Estimator 355
19 Differential
Method I 357
Method II 358
Method 1 361
ix 257
Method II 363
IPosition 268
*7-6 Decoupling by State Feedback 371
lId 272
7-7 Concluding Remarks 377
¡prime Fraction
Problems 378
\ 276
Chapter 8
Stability of Linear Systems 384
1287
Time-Varying Case 385
linces 307
Time-Varying Case 400
324
Problems 425
325
Chapter 9
Linear Time-Invariant Composite Systems: Characteriza
tion, Stability, and Designs 432
Systems 434
s) 339
9-3 Controllability and Observability of Composíte
~d Eig~nvectors 351
9-5 Design of Compensators: Unity Feedback Systems 458
ix of G(s) 352
Single-Variable Case 458
53
Single-Input or Single-Output Case 464
xii CONTENTS
Designs 501
Sytems 506
Implementation 1 517
Implementation 11 519
Applications 523
Decoupling 523
Decoupling 526
Problems 536
Problems 553
Problems 564
Decomposition 565
Problems 570
CONTENTS xiii
ce Rejection 488
G-4 Coprimeness of Polynomial Matrices 592
Jut Feedback
Problems 635
References 636
Index 657
ompensators 517
e 554
of Sampling 559
N 572
This text is intended for use at the senior-graduate level in university courses on
linear systemS and multivariable system designo It may also be used for inde
pendent study and reference by engineers and applied mathematicians. The
mathematical background assumed for this book is a working knowledge of
matrix manipulation and an e1ementary knowledge of differential equations.
The unstarred sections of this book have been used, for over a decade, in the
first graduate course on linear system theory at the State University of New
York at Stony Brook. The majority of the starred sections were developed
during the last three years for a second course on linear systems, mainly on
multivariable systems, at Stony Brook and have been c1assroom tested at a
number of universities.
With the advancement of technology, engineers have become interested in
designing systems that are not merely workable but also the best possible.
Consequently, it is important to study the Iimitations of a system; otherwise,
one might unknowingly try to design an impossible system. Thus, a thorough
investigation of all the properties of a system is essential. In fact, many design
procedures have evolved from such investigations. This text is devoted to this
study and the design procedures developed thereof. This is, however, not a
control text per se, because performance criteria, physical constraints, cost,
optimization, and sensitivity problems are notconsidered.
This text is a revised and expanded edition of Introduction to Linear System
Theory which discussed mostly the state variable approach and was published
in 1970. Since then, several important developments have been made in linear
system theory. Among them, the geometric approach and the transfer-function
matrices in fractional forms, called the matrix-fraction description, are most
xv
xvi PREFACE
pertinent to the original text. The geometric approach is well covered in duced in Chapter 4.
W. M. Wonham's Linear Multivariable Control: A Geometric Approach, 2d ed., realizations (Chapter
Springer-Verlag, New York, 1979 and is outside the scope ofthis text. Hence (Appendix G). Mono
the new material of this edition is mainly in the transfer-function matrix in Hence it was decided
fractional formo Because of this addition, we are able to redevelop, probably polynomials and poi
more simply in concepts and computations, the results of the state variable This, however, will d
approach and establish a fairly complete link between the state-variable the topic was grouped
approach and the transfer-function approach. The logical sequen
We aim to achieve two objectives in the presentation. The first one is to
develop major results and design procedures using simple and efficient methods.
Thus the presentation is not exhaustive; only those concepts which are essential
in the development are introduced. For example, the Smith-McMillan form is Chapter 2
not used in the text and is not discussed. The second objective is to enable the Secs. 2-]
to 2-5
reader to employ the results developed in the text. Consequently, most results
are developed in a manner suitable for numerical computation and for digital
computer programming. We believe that solving one or two problems of each
topic by hand will enhance the understanding ofthe topic and give confidence
in the use of digital computers. With the introduction of the row searching
algorithm (Appendix A), which has been classroom tested, this is possible even Appendix A
design procedures in this text are developed by using only elementary concepts
and results in linear algebra. Chapter 1
The results in this text may eventually be implemented on digital computers. Chapter 2
Because of the finite word lengih, the sensitivity of problems and the stability Chapter 3 (Skip. Th
of algorithms become important on computer computations. These problems Chapter 4 (Ski.p Th
are complex and extensively discussed in texts on numerical analysis. In our Chapter 5 (Empha~
development, we will take note of these problems and remark briefly wherever 5~5, and
appropriate. Chapter6
The arrangementof the topics in this text was not reached without any Chapter 7
difficulty. For example, the concepts of poles and zeros seem to be best intro Charter 8
PREFACE xvii
eh is well covered in duced in Chapter 4. However, their complete treatments require irreducible
etric Approach, 2d ed., realizations (Chapter 6) and coprime fractions of transfer-function matrices
pe of this text. Hence (Appendix G). Moreover, the concept of zeros is used only in Section 9-6.
fer-function matrix in Hence it was decided to create an appendix for the topic. The coprimeness of
o redevelop, probably polynomials and polynomial matrices might be inserted in the main text.
, of the state variable This, however, will digress too much from the state-variable approach; thus
~en the state-variable the topic was grouped in an appendix.
The logical sequences of various chapters and appendixes are as follows:
Jll. The first one is to
: and efficient methods. Chapler 1
:pts which are essential
nith-McMillan form is Chapter 2
!
-====== Chapler 3 (These orders can be inlerchanged.)
ljective is to enable the Secs.2-1
lo 2-5
,equently, most results
utation and for digital 1
Chapter 4
: two problems of each
lic and give confidence
.!
Chapler 5 ---- Appendix B
1 of the row searching
:d, this is possible even Appendix A - Seco 5-8 ~ Appendixes e and D
lfficiently small.
the same as that of the
ra more extensive than
1
Chapler 6
text can be developed Seco 6-·2 Chapler 8 Charler 7
~s. For example, the 1 - - - - Hankel melhod
Sec.6-5 -Appendix E ....... Secs. 8-5-Appendix F ----+- Seco 7-4
We emphasize the exact meanings of theorems and their implications; hence the and discussions has
proofs of a number of theorems are skipped. For example, we prove only matter. 1 am gratefl
Theorems 2-1 and 2-2 in Chapter 2. We skip the proofs ofTheorems 4-1, 4-2, F. Trace for typing va
and others. In the second course, we cover the following: and Winston and the ~
and to Professors S. ]
Appendix A
from the People's R
Section 5-8, controllability and observability indices
thanks go to m y wife
Hankel method (Section 6-4 and method II of Section 6-5)
their support during t
Appendix E
Singular value decomposition method (Method l of Section 6-5)
Appendix G
Sections 6-6 to 6-9
Starred sections of Chapter 7
Appendix H
Chapter 9
Those who are interested in quick access to the design methods using the
transfer-function matrix in fractional form may proceed from Sections 2-1 to
2-5, Appendixes A and G, and then to Sections 9-5 to 9-7, or only their single
variable cases.
The problem sets form an integral part of the book. They are designed to
help the reader understand and utilize the concepts and results covered. In
arder to retain the continuity of the main text, sorne important results are
stated in the problem sets. A solutions manual is available from the pub
Iisher.
The Iiterature on linear system theory is very extensive. The length of this
text, however, is limited. Hence the omission of sorne significant results is
inevitable and l would Iike to apologize for il. l am indebted to many people in
writing this book. Kalman's work and Zadeh and Desoer's book Linear
System Theory form the foundation ofthe original edition ofthis book. Rosen
brock's and Wolovich's works are essential in developing the present edition.
l have benefited immensely in my learning from Professor C. A. Desoer. Even
to this date, 1 can always go to him whenever l have questions. For this, I can
never express enough of my gratitude. To Professors B. J. Leon, E. J. Craig,
I. B. Rhodes, P. E. Barry (first edition) and to Professors M. E. Van Valkenburg,
W. R. Perkins, D. Z. Zheng (present edition), l wish to express my appreciation
for their reviews and valuable suggestions. I would like to thank President
F. Zhang and Professor K. W. You of Chengdu University of Science and
Technology, Professor S. B. Park of Korea Advanced lnstitute of Science and
Technology, Professor T. S. Kuo of National Taiwan University, and Professor
S. K. Chow of National Sun Yat-Sen University, Taiwan, for providing oppor
tunities for me to lecture on an earlier draft of Chapter 9 and Appendix G.
I especially appreciate the opportunity at Chengdu University to interact witb
several faculty members, especially Professor L. S. Zhang, from various uni
versities in China; their suggestions have improved considerably the presenta
tion of the text. I am grateful to many of "my graduate" students, speciaUy
C. Waters, C. H. Hsu(the first edition), I. S. Krishnarao, Y. S. Lai, C. C. Tsui and
S. Y. Zhang (the present edition) whose assistance in the form of dissertations
--~ ----~~-----~_.-.-
PREFACE xix
A,B,P, ...
u, y, ex, •..
u, y, ex, ...
!l'
u(s), y(s), G(s), C(s)
v(A), ...
A',x', .
A*, x*, .
det A, ...
iC
~
~(s)
~[s]
p(A), rank A
Glossary of Symbols
xxi
xxii GLOSSAR y OF SYMBOLS
Linear Sys
and Desig
where A, B, and e are matrices, not necessarily square
and of the same order.
~ Equals by definition.
~A = (~a.)
dt te, dt 1) ,
ff! [A] ~ (ff! [aij]), . " When an operator is applied to a matrix or a vector,
it means that the operator is applied to every entry of
the matrix or the vector.
ational or polynomial
~(s).
. A(s).
d e as block diagonal
a matrix or a vector,
Jlied to every entry of
1-1 The Study 01
properlymodeled.
A physical system rr
asked and the different
amplifier may be modell
1
Introduction
The study and design of a physical system can be carried oui by "empirical
methods. We apply various signals to the physical system and measure its
responses. If the performance is not satisfactory, we adjust sorne of its param
eters or connect to it sorne compensator to improve its performance. This
design is guided by past experience, if any, and proceeds by cut and try. This
approach has undoubtedly succeeded in designing many physical systems.
The empirical method may become unsatisfactory if the specifications on the
performance become very precise and stringent. lt may also become inadequate
if physical systems become very complicated or too expensive or too dangerous
to be experimented. In these cases, analytica! methods become indispensable.
Analytical study of physical systems roughly consists of four parts: modeling,
development of mathematical-equation description, analysis, and designo The
distinction between physical systems and models are basic in engineering. In
fact, circuits or control systems studied in any textbook are models of physical
systems. A resistor with a constant resistance is a model; the power limitation
of the resistor does not appear in the resistance. An inductor with a constant
inductance is again a model; in reality, the inductance may varv with the
amount of current ftowing through it. Modeling is a very important problem,
for the success of the design dependsupon whether the physical system is
properly modeled.
A physicalsystem may have different models depending on the questions
asked and the different operational ranges used. For example, an electronic
amplifier may be modeled differently at high and low frequencies. A spaceship
2 INTRODUCTION
may be modeled as a particle in the study of its trajectory; however, it must be systems are mainly lil
modeled as a rigid body in the maneuvering. In order to develop a suitable
model of a physical system, a thorough understanding of the physical system Y(s)=
and its operatiúnal range is essential. In this book, we shall refer to models of and
physical systems as systems. Hence a physical system is a device or a collection
of devices existing in the real world; a system is a model of a physical system.
Once a system (a model) is found for a physical system, the next step in the
studyis to develop, by applying various physicallaws, mathematical equations
to describe the system. For example, we apply Kirchhoffs voltage and current and their extensions
laws to electrical.systems and Newton's laws to mechanical systems. The relationship between
equations that describe systems may assume many forms; they may be linear and is called the inp
equations, nonlinear equations, integral equations, difference equations, differ frequency domain. T
ential equations, or others. Depending on the question asked, one form of its elements are limitec
equation may be preferable to another in describing the same system. In con as ratios of two polyn
clusion, a system may have many different mathematical-equation descriptions, said to be in the poly
just as a physical system may have many different models. is called a dynamical (
Once a mathematical description of a system is obtained, the next step in set of first-order diffe:
the study involves analyses-quantitative and/or qualitative. In the quanti it is called the dynan
tative analysis, we are interested in exact responses of systems due to the applica or the internal descri¡
tion of certain input signals. This part of the analysis can be easily carried out from the eoncepts of
by using a digital or an analog computer. In the qualitative analysis, we are examples, how they ca
interested in the general properties of the system, such as stability, control The major portion
lability, and observability. This part of analysis is very important, because around Equations (1
desig~ techniques may often evolve from this study. and qualitative. The
lf the response of a system is found to be unsatisfactory, the system has to puters; hence we empl
be improved or optimized. In some cases, the responses of systems can be will be thoroughly in
improved by adjusting certain parameters of the systems; in other cases, com number of design proc
pensators have to be introduced. Note that the design is earried out on the The design problerr
model of a physical system. However, if the model is properly chosen, the control systems. In (
performance of the physical system should be correspondingly improved by without any constraint
introducing the required adjustments or compensators. design objectives, such
mum degrees which ce
are also interested in (
1-2 Ttae Scope of the Bool~ túe designo In the desi;
system to meet some
The study of systems may be divided into four parts: modeling, setting up steady-state error, and
mathematical equations, analysis, and designo The development of models for compensators. A des
physieal systems req uires knowledge ofeach particular field and sorne measuring from the one without
devices. For example, to develop models for transistors requires the knowledge systems is not consideff
of quantum physics and sorne laboratory setup. Developing models for auto of control systems, alth
mobile suspension systeins iequires actual testing and measurements; it cannot in its designo For a dis(
be achieved by the use of pencil and paper alone. Thus, the modeling problem 546, and 596. In this se
should be studied in connection with each specific field and cannot be properly chapter.
covered in this text. Hence, we shall assume that models of physical systems, We review in Chapt
or systems, are available to us in this text. The objective of thisc
The mathematical equations which will be used in this text to describe transformations, to sob
THE SCOPE OF THE BOOK 3
of a matrix. These techniques are very important, if not indispensable, in are studied in Chaptel
analysis and design of linear systems. the eigenvalues of the
In Chapter 3 we develop systematically the input-output description and feedback with a cons·
the state-variable description of linear systems. These descriptions are devel its state can be gener
oped from the concepts of linearity, relaxedness, causality, and time invariance. val ues. Various desi.
We also show, by examples, how these descriptions can be set up for systems. is also established.
Mathematical descriptions of composite systems and discrete-time equations We study in Chapl
are also introduced. We also discuss the well-posedness problem in the feedback under the heading of ~
systems. the design of a system
In Chapter 4 we study the solutions of linear dynamical equations. We output stability, stabi
also show that different analysis often leads to different dynamical-equation total stability. Their
descriptions of the same system. The relation between the input-output discuss the Lyapunov
description and the state-variable description is also established. criterion. Their cour
We introduce in Chapter 5 the concepts of controllability and observability. In the last chapter
The importance of introducing these concepts can be seen from the networks invariant composite s~
shown in Figure 1-1. Their transfer functions are both equal to 1. There is no zero cancellation of t
doubt about the transfer function of the network shown in Figure 1-I(b); with the transfer functi
however, we may ask why the capacitor in Figure 1-I(a) does not play any role ways, as shown in Fig
in the transfer function. In order to answer this question, the concepts of be studied from its con
controllability and observability are needed. These two concepts are also and (c) cannot. We a
essential in optimal control theory, stability studies, and the prediction or feedback systems. We
filtering of signals. Various necessary and sufficient conditions for a dynamical function matrices in t
equation to be controllable and observable are derived. We also discuss the tions: unity feedback
canonical decomposition of a dynamical equation and introduce an efficient design compensators '
and numerically stable method of reducing a dynamical equation to an irreduci ment and to achieve
ble one. reestablish in the tran
In Chapter 6 we study irreducible realizations of rational transfer-function in the state-variable
matrices. The problem is to find a controllable and observable linear time approaches.
invariant dynamical equation that has a prescribed rational matrix. lts solution A total of eight ;
is indispensable in analog and digital computer simulations. lt also offers a various chapters are el
method of synthesizing a rational matrix by using operational-amplifier circuits.
This result is also needed in establishing the link between the state-variable
approach and the transfer-function approach in the design of linear time
invariant systems.
The practical implications ofthe concepts of controllability and observability (a)
11
t 1: In
,.
+
11
t In
-r
,.
+
------o:<
J
(urrent
I
SOUfCC 151 In
___ L
lo) lb)
Figure 1-1 Two different networks withthe same transfer function, 1. Figure 1-2 Thrée differ
THE SCOPE Of THE BOOK 5
;
not indispensable, in are studied in Chapter 7. We show that if a dynamical equation is controllable,
the eigenvalues of the equation can be arbitrarily assigned by introducing state
Itput description and feedback with a constant gain matrix. lf a dynamical equation is observable,
escriptions are devel its state can be generated by designing a state estimator with arbitrary eigen
, and time invariance. values. Various design procedures are introduced. The separation property
be set up for systems. is also established.
screte-time equations We study in Chapter 8 a qualitative property oflinear systems. This comes
'oblem in the feedback under the heading of stability, which is always the first requirement to be met in
the design of a system. We introduce the concepts of bounded-input bounded
nical equations. We output stability, stability in the sense of Lyapunov, asymptotic stability, and
t dynamical-equation total stability. Their characterizations and relationships are studied. We also
~en the input-output discuss the Lyapunov theorem and then use it to establish the Routh-Hurwitz
blished. criterion. Their counterparts in the discrete-time case are also studied.
lity and observability. In the last chapter we study various problems associated with linear, time
m from the networks invariant composite systems. One of them is to study the implication of pole
lual to 1. There is no zero cancellation of transfer functions. For example, consider two systems
wn in Figure l-l(b); with the transfer functions l/(s -1) and (s - 1)/(s + 1) connected in three different
loes not play any role ways, as shown in Figure 1-2. We show why the system in Figure 1-2(b) can
tion, the concepts of be studied from its composite transfer function, but the systems in Figure 1-2(a)
vo concepts are also and (c) cannot. We also study stabilities of single-variable and multivariable
nd the prediction or feedback systems. We then study the design of compensators by using transfer
itions for a dynamical function matrices in the fractional formo We study two feedback configura
We also discuss the tions: unity feedback and plant input-output feedback connections. We
mtroduce an efficient design compensators with minimum degrees to achieve arbitrary pole place
luation to an irreduci ment and to achieve asymptotic tracking and disturbance rejection. We
reestablish in the transfer-function approach essentially the results developed
)nal transfer-function in the state-variable approach and complete the link between the two
lservable linear time approaches.
Ll matrix. lts solution A total of eight appendixes are introduced. Their relationships with
ions. lt also offers a various chapters are covered in the Preface.
Illal-amplifier circuits.
~en the state-variable
esign of linear time
(a)
·ility and observability
(b)
GJ-r
~J s - \
(c)
;tion, 1. Figure 1-2 Three dilTerent connections 01" 1/(s -1) and (s - 1)/(s + 1).
Section 2-7, we study
and the Cayley-Ham
concepts of inner pro<
This chapter is in
have some basic knm
multiplication, and im
2
be used. Let A, B, (
respectively. Let 3i be
we have
rt
BD= b
bm
These identities can b
2-1 Introduction the product of an n x
The material presel
In this chapter we shall review a number of concepts and results in linear 38, 39,43 to 45, 77, 86
algebra that are essential in the study of this text. The topics are carefully emphasize the differen
selected, and only those which will be subsequently used are introduced. The (2-12) and Definition
purpose of this chapter is to enable the reader to understand the mechanism of matrix representation
similarity transformation, to solve linear algebraic equations, to find Jordan follow naturally.
form representations of square matrices, and to compute functions of a matrix,
in particular, exponential functions of a matrix (see Section 2-9, Concluding
Remarks 1 ). 2-2 linear Spaces
In Section 2-2 we introduce the concepts of field and linear space over a
field. The fields we shall encounter in this book are the field of real numbers, In the study oí mathe
the field of complex numbers, and the field of rational functions. In order to forms the center of stu
have a representation of a vector in a linear space, we introduce, in Section 2-3, For example, in arith
the concept of basis. The relationship between different representations of the algebra, we study the
same vector is established. In Section 2-4, we study linear operators and their examples of sets includ
representations. The concept of similarity transformation is embedded here. the set of aH polynomi
In Section 2-5, the solutions of a set of linear algebraic equations are studied. matrices. In this sech
The concepts of rank and nullity are essential here. ' In Section 2-6, we show that one of those just menti
every square matrix has a Jordan-form representation; this is achieved by· Consider the set 01
introducing eigenvectors and generalized eigenvectors as basis vectors: . in plication with the corr
set. The sum and pr<
set has elements O and
1 It is recommended that the reader keeps the concludin~ remarks in mind, for they provide the
reader with' motivations for studying the mathematical theorems introduced in this chapter.
2 Numbers correspond to th
6
LINEAR SPACES OVER A FIELD 7
BD= (blJ
bt D= (blDJ
b~D (2-3)
bm bmD
These identities can be easily checked. Note that a¡b i is an n x r matrix, it is
the product of an n x 1 matrix a¡ and a 1 x r matrix b¡_
The material presented here is well known and can be found in References 5,
; and results in linear 38,39,43 t045, 77, 86, and 116. z However our presentation is different. We
he topics are carefully emphasize the difference between a vector and its representations [see Equation
i are introduced. The (2-12) and Definition 2-7]. After stressing this distinction, the concepts of
land the mechanism of matrix representation of an operator and of the similarity transformation
ltions, to find Jordan follow naturally.
: functions of a matrix,
:ction 2-9, Coneluding
2-2 Linear Spaces over a Field
Id linear space over a
: field of real numbers, In the study of mathematics we must first specify a collection of objects that
"unctions. In order to forms the center of study. This coHection of objects or eíements is called a seto
:roduce, in Section 2-3, For example, in arithmetic, we study the set of real numbers. In boolean
representations of the algebra, we study the set {O, 1}, which consists of only two elements. Other
:ar operators and their examples of sets inelude the set of complex numbers, the set of positive integers,
ion is embedded here. the set of aH polynomials of degree less than 5, the set of aH 2 x 2 real constant
equations are studied. matrices. In this section when we discuss a set of objects, the set could be any
ction 2-6, we show that one of thosejust mentioned or any other the reader wishes to specify~
1; this is achieved by Consider the set of real numbers. The operations of addition and multi
as basis vectors. In plication with the commutative and associative properties are defined Tor the
set. The sum and product of any two real numbers are real numbers.. The
set has elements O and 1. Any real number ex has an additive inverse ( - ex) and
l mind, for they Pfovide the
a multiplicative inverse (l/a, except a = O) in the set. Any set with these proper field are, respectively,
ties is caBed afield. We give a formal definition of a field in the foBowing.
Definition 2-1
A field consists uf a set, denoted by :F, of elements caBed sealars and two Note that the set of all
operations caBed addition "+" and multiplication "'''; the two operations From the foregoin:
are defined over :F such that they satisfy the following conditions: field could be anythin~
objects. The fields W{
1. To every pair of elements a and 13 in :F, there correspond an element a + [3 familiar ones: the field
in :F called the sum of a and [3, and an element a . [3 or af3 in :F, called the field of rational functic
produet of a and [3. tions of these fields ar
2. Addition and multiplication are respective1y commutative: F or any a, [3 in :F, show that they satisfy ;
a+f3=[3+a to denote the field of rea
and use !R(s) to denote
3. Addition and multiplication are respectively associative: For any a, 13, y in :F, with indeterminate s. 1
(a + fJ) + y = a + (fJ + y) (a' f3)' y = a' (/3' y) field because it has no :
nomials do not forro a
4. Multiplication is distributive with respect to addition: For any a, 13, y in :F, Before introducing
a . (13 + y) = (a . 13) + (a . y) two-dimensional geom
the plane can be consi<
5. :F contains an element, denoted by O, and an element, denoted by 1, such A vector can be shrunl
that a + O= a, 1 . a = a for every a in :F. product of two points
6. To every a in :F, there Is an element 13 in :F s\.1ch that a + 13 = 0. The element matical terminology, is
13 is caBed the additive inverse. spaee.
7. To every a in :F which is not the element O, there is an element y in :F such
that a· y = 1. The element y is caBed the multiplieative inverse. I Definition 2-2
A linear space over a fie
We give some examples to illustrate this concept.
of elements caBed vecte
and sealar multiplieatioj
Example 1 that they satisfy all the
Consider the set of numbers that consists of O and 1. The set {O, 1} does not 1. To every pair of ve<
form a fie1d ifwe use the usual definition of addition and multiplication, because in :!f, called the SUD:
the e1ement 1 + 1 = 2 is not in the set {0,1}. However, if we define 0+ O = 2. Addition is commu
1 + 1 =0, 1 + O = 1 and 0·1 =0,0 =0,1'1 = 1, then it can be verified that {O, 1} 3. Addition is associa
with the defined addition and multiplication satisfies aB the conditions listed (x 2 +x 3 )·
for a field. Hence the set {O, 1} with the defined operations forms a field. It is 4. fj( contains a vecto
called the field of binary numbers. I The vector Ois caB(
5. To every x in fI, th<
Example 2 6. To every a in :F, an(
the sealar product o
Consider th~ set of all 2 x 2 matrices of the form
7. Scalar multiplicatic
X -YJ
.[ · y · x .
a(f3x) = (af3)x .
Note that the set of a11 2 x 2 matrices does not form a field. I
alled scalars and two
'; the two operations From the foregoing examples, we see that the set of objects that forms a
:onditions: field could be anything so long as the two operations can be defined for these
objects. The fields we sha11 encounter in this book are fortunately the most
lond an element ex + 13 familiar ones: the field of real numbers, the field of complex numbers, and the
:>r rxf3 in ?F, called the field of rational functions with real coefficients. The additions and multiplica
tions of these fields are defined in the usual ways. The reader is advised to
:ive: For any ex, 13 in ?F, show that they satisfy aH the conditions required for a field. We use IR and iC
to denote the field of real numbers and the field of complex numbers, respectively,
and use lR(s) to denote the field of rational functions with real coefficients and
e: For any a, /3, '}' in ?F, with indeterminate s. Note that the set ofpositive real numbers does not form a
field because it has no additive inverse. The set of integers and the set of poly
(13 . y)
nomials do not form a field because they have no multiplicative inverse. 3
For any ex, 13, y in ?F, Before introducing the concept of vector space, let us consider the ordinary
two-dimensional geometric planeo lf the origin is chosen, then every point in
the plane can be considered as a vector: it has direction as we11 as magnitude.
11, denoted by 1, such A vector can be shrunk or extended. Any two vectors can be added, but the
product of two points or vectors is not defined. Such a plane, in the mathe
+13=0. Theelement matical terminology, is ca11ed a linear space, or a vector space, or alinear vector
space.
n element y in ?F such
Je ínverse. I Defínition 2-2
A linear space over a field ?F, denoted by (f![, ?F), consists of a set, denoted by f![,
of elements ca11ed vectors, a field fF, and two operations called vector addition
and scalar multiplication. The two operations are defined over f![ and ff such
that they satisfy a11 the fo11owing conditions:
~he set {O, 1} does not f![, there corresponds a vector Xl +X z
1. To every pair of vectors Xl and x 2 in
nultiplication, because in 9(, ca11ed the sum ofx l and X;.
r, if we define 0+0= 2. Addition is commutative: For any Xl' X2 in ~ Xl + X2 =X 2 +x l .
. be verified that {O, 1} 3. Addition is associative: For any x¡, x 2 , and x 3 in f![, (Xl +X 2 )+X 3 =X l +
J the conditions listed (X 2 +X 3 )·
ms forms a fie\d. It is 4. f![ contains a vector, denoted by O, such that O +x = X for every x in f![.
3 A set with aB properties of a field except property 7 in Definilion 2-[ is caBed a ril1g or, more
the usual definitioris of precisely, a commutative ring with (multiplicative) identity. The set of integers forms a ring, as
lements Oand 1 of the does the set of polynomials with real" codficients.
-- - - - --_..
-- - ----- -------._~-----_._-_._--- --~-----
-~- -~-----------
_--~-
8. Scalar multiplication is distributive with respect to vector addition: For Let the vector additioj
anya in ff and any x¡, x 2 in f!l', a(x l +X 2 ) = cxx l +cxx 2 .
9. Scalar multiplication is distributive with respect to scalar addition: For
any a, {3 in ff and anyx in f!l', (a + {3)x = cxx + {3x.
10. For any x in !!l, Ix =x, where 1 is the element 1 in ff. I
Example 1
A field forms a vector space over itself with the vector addition and scalar lt is easy to verify tha
multiplication defined as the corresponding operations in the field. For not a linear space.
example, (~,~) and (iC, iC) are vector spaces. Note that (iC,~) is a vector
space but not (~, ic). (Why?) We note that (~(s), ~(s)) and (~(s), ~) are also Example 5
vector spaces, but not (~, ~(s)). I
Let !!l denote the set e
Example 2 x+2x + 3x =0. Then
scalar multiplication dI
The set of all real-valued piecewise continuous functions defined over ( - 00, 00) homogeneous, then ( Pi
forms a linear space over the field of real numbers. The addition and scalar
multiplication are defined in the usual way. lt is called a function space. I
We introduce one 1
Example 3
Given a field ff, let ff" be all n-tuples of scalars written as columns Definition 2-3
Let (f!l', ~) be a linear ~
l
whether a set iJ}j is a su
Xli + X1Jl
x¡ +Xj= X2i ~ X2] cxx.= aX2i (2-5) is in iJ}j for any Yl, Y2 in
[ XIIi +Xllj~ !
ra~lli-,
then (ff", ff) is a vector space. lf ff =~, (~", ~) is called the n-dimensional
6 are satisfied. Henc~
real vector space; if ff = ic, (e, ic) is caHed the n-dimensional complex vector Example 6
space; if ff = ~(s), (W(s), ~(s)) is called the n-dimensional rational vector
In the two-dimensiona
~~ I
through the origin is a
Example 4
Consider the set !R"[s] of aH polynomials of degree less than n with real co
effiCients 4
n- 1
ror any fixed real a is a
L a¡si
¡~O
Example 7
4 Note that IR (s), with parenlheses, denotes the field of rational functions with real coefficients;
whereas IR [s J, with brackets, denotes the set of polyti.o~ials with real coefficients. The real vector space 1
LINEAR SPACES OVER A FIELD 11
, vector addition: For Let the vector addition and the scalar multiplication be defined as
~2' n-l n-l "-1
) scalar addition: For ¿ lX¡si + ¿ f3iSi = ¿ (IX¡ + f3¡)S¡
i=O i=O ;=0
¡;
• IX Ct~ lXiS¡) = ~t~ (IXIX¡)S¡
Ir addition and scalar 1t is easy to verify that (rrln[S], rrl) is a linear space. Note that (rrln[s], rrl(s)) is
ns in the fie1d. For not a linear space. I
hat (iC, IR) is a vector
) and (rrl (s), rrl) are also Example 5
• Let fL denote the set of aH solutions of the homogeneous differential equation
x + 2x + 3x = O. Then (fL, rrl) is a linear space with the vector addition and the
scalar multiplication defined in the usual way. lfthe differential equation is not
defined over (- 00, 00) homogeneous, then (gr, rrl) is not a linear space. (Why?) 11
le addition and scalar
a function space. I
We introduce one more concept to conclude this section.
A set of vectors Xl' Xz, ... ,XII in a linear space over a field ff, (Er, ff), is said to
Observe that line
be linearly dependent if and only if there exist scalars (Xl' (Xz, . .. , (XII in ff, not
but also on the field.
all zero, such that
(2-6)
X
lf the only set of (Xi for which (2-6) holds is (Xl = 0, (Xz = O, ... , (XII = 0, then the
set of vectors xl> Xz, ... ,XII is said to be linearly independent. I
Given any set of vectors, Equation (2-6) always holds for (Xl = 0, (Xz = 0, ... ,
is linearly dependent
(XII = O. Therefore,in order to show the linear independence of the set, we have
°
to show that (Xl = 0, (Xz = 0, ... , (XII = is the only set of (Xi for which (2-6) holds;
that is, if any one of the (X;'s is different fmm zero, then the right-hand side of
Indeed, if we choose
(2-6) cannot be a zero vector. lf a set of vectors is linearly dependent, there are
generaHy infinitely many sets of (x¡, not aH zero, that satisfy Equation (2-6).
However, it is sufficient to find one set of (x¡, not aH zero, to conclude the linear then Cl.1X l +(Xzx z =0.
dependence of the set of vectors. the field of real numbe
zero, such that (X 1X1 +
Example 1 pendent in (IRZ(s), IR), 1
ConsÍder the set of vectors Xl' Xz, ... , XII in which Xl =0. This set of vectors
lt is clear from tI
is always linearly dependent, because we may choose (Xl = 1, (Xz = 0, (Xz = 0, ... ,
Xl? x2 ·> ", Xi¡ are HrJea
Cl. = O, and Equation (2-6) holds.
H id!
as a linear combinatio
every one of them can
Example 2
Consider the set of vector Xl which consists of only one vector. The set of
Definition 2-5
vector Xl is linearly independent if and only if Xl #= O. lf Xl #=0, the only way
tohaveCl.1X l =OiS(Xl =0. Ifxl=O,wemaychooseCl.l=l. I The maximal number
is called the dimension
lf we introduce the notation
In the previous sec
(IR", IR). The meaning
thereare, at most, n lin
dimensional real vect(
(2-7)
pendent veCtors. (Try
LINEAR INDEPENDENCE, BASES, AND REPRESENTATIONS 13
then the linear independence of a set of vectors can also be stated in the fol
'esentations
lowing definition.
e mutually perpendic
.ordinate system is to Definition 2-4'
:ctor in the planeo In
A set of vectors x ¡, x 2, ... , x" in (fll', 9') is said to be linearly independent if and
general linear spaces.
only if the equation
The basis vectors are
ferent scales. Before [Xl X2 X,,]ex=ü
)f vectors.
implies ex = O, where cvery component of ex is an element of ff or, correspond
ingly, ex can be considered as a vector in ff". •
[ 1 [
'nt. I
s+2 s+3
for IX¡ =0, 1X 2 = O, ... ,
is linearly dependent in the field of rational functions with real coefficients.
ce of the set, we have
lndeed, if we choose
for which (2-6) holds;
:he right-hand side of s+3
( dependent, there are IX¡ = -1 and Ci. 2 =-
s+2
atisfy Equation (2-6).
lo conclude the linear then IX¡X¡ +Ci. 2X2 =0. However, this set of vectors is linearly independent in
the field of real numbers, for there exist no Ci.¡ and Ci. 2 in IR that are different from
zero, such that Ci.¡x¡ +Ci.2X2 =0. In other words, x¡ and X2 are linearly inde
pendent in (1R 2(s), IR), but are linearly dependent in (1R 2(s), IR (s)). •
There are infinitely many of these functions; therefore, the dimension of this This theorem ha:
• space is infinity. • vector space (f!{, g-), i
represented by a set (
We assume that aH the linear spaces we shaH encounter are of finite dimen (2-7), we may write (2
sions unless stated otherwise.
which, together with the linear independence assumption of el' ez, ... , en,
implies that al = O, az == O, ... , a = O. This contradicts the assumption that
ll
not aH ao, al>' .. , a n are zero. Ir we define {Ji ~ - aJa o, for i = 1, 2, ... , n, then
(2-8) becomes
(2-10)
This shows that every vector x in fE can be expressed as a linear combination Figure 2-1 A two-dim<
LINEAR INDEPENDENCE, BASES, AND REPRESENTATlONS 15
of el> ez, . .. , en· Now we show that this combination is unique. Suppose there
piecewise continuous is another linear combination, say
:pace is the one which x= /3l e ¡ + /3z e z + ... + /3nen (2-11 )
, with - CIJ < t < CIJ,
Then by subtracting (2-11) from (2-10), we obtain
0= (/31 -/3¡)e l + (/3z -/3z)ez + ... + (/3n -/3n)en
:tions {t n, n= 1,2, ... }
mts, lX/S, not all zero, which, together with the linear independence of {eJ, implies that
i = 1,2, ... , n
This completes the proof of this theorem. Q.E.D.
:he dimension of this This theorem has a very important implication. In an n-dimensional
• vector space (f![, .?"), if a basis is chosen, then every vector in f![ can be uniquely
represented by a set of n scalars /31, /3z, ... , /3n in.?". lf we use the notation of
er are of finite dimen (2-7), we may write (2-10) as
x= [el e z ... en]p (2-12)
where p = [/31' /3z, ... , /3nJ' and the prime denotes the transpose. The n x 1
vector p can be considered as a vector in (.?"n, .?"). Consequently, there is a
f![,.?") is said to be a
one-to-one correspondence between any n-dimensional vector space (f![, .?") and
le linear combination
I
the same dimensional linear space (.?"n, .?") if a basis is chosen for (f![, .?").
Definition 2-7
independent vectors In an n-dimensional vector space (f![, .?"), if a basis {el> ez, . .. ,en} is chosen,
then every vector x in f![ can be uniquely written in the fonn of (2-12). P is
called the representation of x with respect to the basis {el, ez' ... , en}. •
Example 4
n and let x be an
f![,
L, ez, ... , el! is linearly
The geometric plane shown in Figure 2-1 can be considered as a two-dimensional
maximum number of real vector space. Any point in the plane is a vector. Theorem 2-1 states that
Consequently, there
(2-8)
(2-9)
(2-10)
Changa of basis. ,
representations with ]
hence [3 2 - 2 10]' (where the "prime" denotes the transpose) is the relationships are bet\
representation of x with respect to {e ¡, e 2, e 3, e 4}. Ir we choose e ¡ = S3 - S2, In this subsection, thi:
e2 = S2 - s, e3 = s -1, and e4 = 1 as the basis vectors, then Let the representa1
x: ::: +::' ~,~ :,;OW" -,') + 5(,' -,)+ 3(, -1) + 131
and {e¡, e 2 , .•• ,en} be
x=[
In order to derivf
infOI:mation of the rel
real vector space (IR n, IR), complex vector space (en, C), or rational vector
space (W(s), lR(s)); a vector is an n-tuple of real, complex, or real rational
functions, written as
as the basis of(lR n, IR), (en, C), and (IRn(s), lR(s)). In this case, an array ofnumbers
can be interpreted as a vector or the representation of a vector with respect to
~ set of aH real poly the basis {o¡, 0z, ... , on}, because with respect fo this particular set of bases, the
Let el =s3,e z =sz, representation and the vector itself are identical; that is,
: linearly independent
ctors, the vector x =
(2-14 )
Change of basis. We have shown that a vector x in (~, ff) has different
representations with respect to different bases. it is naturalto ask Vvhái: the
:he transpose) is the relationships are between these different representations of the same vector.
e choose el = S3 - sZ, In this subsection, this problem will be studied.
L Let the representations of a vector x in (,qr, ff) with respect to {e¡, e z, ... , en}
and {el' ez, ... , en} be ~ and p, respectively; that iS,6
s-1)+13-1
x=[e¡ e z ._. en]~=[e¡ ez en]p (2-15)
Hs [3 5 3 13]'. I
5 This set or vectors is called an orthonormal set.
:ctors and representa 60ne might be tempted to write p=[e 1 ez '" en]-l[e¡ ez ... cn]jl. However,
der the n-dimensional [el e z .. . en] - ¡ may not be defined as can be seen rrom Example 5.
18 LINEAR SPACES AND LINEAR OPERATORS
basis {e b ez, ... ,en}, or the inforrnation of the representations of el' for i = Example 6
1,2, ... , n, with respect to the basis {e b ez, ... , en}. Let the representation of
e i with respect to {el, e z,···, en} be [PI¡ PZi P3¡ ... Pni]'; that is, Consider the two set
readily verified that
where E ~ [el ez en], Pi ~ [Pli PZi ... Pn¡J'· Using rnatrix nota and
tion, we write
[el e z ... en] = [EpI Epz ... EPn] (2-17)
l
PII PIZ
en] Pfl pzz PI]
PZn
Since the representation of x with respect to the basis {e b ez, ... ,en} is uniqtie, The concept of a func1
(2-19) irnplies úJI, ifwe assign to each
p=Pp (2-20)
of assignments is can
Figure 2-2(a) is a fur
ith colurnn: the
P = representation of
J usually denoted by th(
to the element x of f![
where (2-21 )
e with respect to defined is caBed the d
[ i
{el' ez, ... , e/l} to sorne elernent of ~
This establishes the relationship between p and p. In (2-16), if the repre
sentation of e¡ with respect to [e j , e z, . .. , e,J is used, then we shall obtain
(2-22 )
2, ... , n (2-16)
(2-17)
~
le. e, " '4] l'. " " '4] l- ~ -1
o
1
O
O
O -1
1 ~J~l" "" '4]Q
PI]
P2n
Pnn
(2-18)
rn =Pl ~lJ::
ll~J ~~ U 1
PJ~ rl~~J~l
, e2 , . . . , en} is unique, The concept of a function is basic to all parts of analysis. Given two sets f![ and
11JJ, ifwe assign to each elernent of f![ one and only one elernent of 11JJ, then the rule
(2-20)
of assignrnents is called a function. For exarnple, the rule of assignrnents in
Figure 2-2(a) is a function, but not the one in Figure 2-2(b). A function is
usually denoted by the notation f : f![ ---> 11JJ, and the elernent of l1JJ that is assigned
(2-21 ) to the elernent x of f![ is denoted by y = f(x). The set f![ on which a function is
defined is called the domain o! the function. The subset of l1JJ that is assigned
to sorne elernent of f![ is called the range of the function. For exarnple, the
:n (2-16), if the repre
1 we shall obtain 1{
(2-22)
(2-23)
--"-0f----\---r----+--------'L----.X
·20) or (2-22). There y =Lx
'ector with respect to -1
.e vector with respect
2-20) or (2-22). Since
(a) (b)
:conclude that
Figure 2-2 Examples in which (a) the curve represents a function and.(b) the curve does
(2-24 ) not represent a functlon.
20 LINEAR SPACES AND LINEAR OPERATORS
domain ofthe function shown in Figure 2-2(a) is the positive realline, the range is a linear transforma1
of the function is the set [ -1, 1], which is a subset of the entire real Hne úJ/. all equal to (O/t, IR).
The functions we shall study in this section belong to a restricted class of
functions, called linear functions, or more often called linear operators, linear Matrix representa
mappings, or linear transformations. The sets associated with linear operators
two examples that the
are required to be linear spaces over the same field, say (gr, ff) and (úJ/, ff). A defined may be of fini
linear operator is denoted by L:( gr, ff) ---> (úJ/, ff). In words, L maps (gr, ff)
every linear operatc
into (úJ/, ff). dimensional (úJ/, ff) h:
Ir (gr, ff) and (úJ/, ff)
Definition 2-8 operator can still be I
A function L that maps (gr, ff) into (úJ/, ff) is said to be a linear operator if infinite order or a fon
and only if and will not be discus
L((;(¡x¡ + (;(zx z ) = (;(¡Lx¡ + (;(zLx z
Theorem 2-2
for any vectors Xl' X z in grand any scalars (;(¡, (;(z in ff. I
Let (gr, ff) and (úJ/, ¿
Note that the vectors Lx¡ and Lx z are elements of 0Jj. The reason for over the same field. ]
requiring that úJ/ be defined over the same field as gr is to ensure that (;(¡Lx¡ in gr. Then the line:
and (;(zLx z be defined. by the n pairs of m,
respect to the basis {x
Example 1 be represented by an
column af A is the rep]
Consider the transformation that rotates a point in a geometric plane counter
clockwise 90° with respect to the origin as shown in Figure 2-3. Given any Proof
two vectors in the plane, it is easy to verify that the vector that is the sum of the
Let x be an arbitrary \
two vectors after rotation is equal to the rotation of the vector that is the sum
the set af vectors qua
ofthe two vectors before rotation. Hence the transformation is a linear trans
pressed uniquely as (;(
formation. The spaces (gr, ff) and (úJ/, ff) of this example are all equal to
of L, we have
(IR Z , ~). I
Example 2
which implies that fOj
Let O/t be the set of all real-valued piecewise continuous functions defined over i = 1, 2, ... ,n. This p
[O, T] for sorne finite T > O. lt is clear that (O/t, IR) is a linear space whose
Let the represeI
dimension is infinity (see Example 3, Section 2-3). Let 9 be a continuous
[au aZi ... ami]';
function defined over [0, T]. Then the transformation
------------
---
.-------- -
~s -3 Y2 -0.5 xI 1.5
Figure 2-3 The transformation that rotates a vector counterclockwise 90°.
LINEAR OPERATORS AND THEIR REPRESENTA:rIONS 21
ive realline, the range
is a linear transformation. The spaces (gr, ff) and (qy, ff) of this example are
e entire realline qy.
aH equal to (Ii/l, IR). I
o a restricted class of
inear operators, linear
Matrix representations of a linear operator. We see fram the aboye
with linear operators
two examples that the spaces (gr, ff) and (qy, ff) on which a linear operator is
gr, ff) and (qy, ff). A
defined may be of finite or infinite dimensiono We show in the foHowing that
'ords, L maps (gr, ff)
every linear operator that maps finite-dimensional (gr, ff) into finite
dimensional (qy, ff) has matrix representations with coefficients in the field ff.
lf (gr, ff) and (qy, ff) are of infinite dimension, a representation of a linear
operator can still be found. However, the representation will be a matrix of
Je a linear operator if infinite order or a form similar to (2-25). This is outside the scope of this text
and will not be discussed.
Theorem 2-2
I
Let (gr, ff) and (qy, ff) be n- and m-dimensional vector spaces, respectively,
f O.!/. The reason for over the same field. Let Xl' X2, ... ,XII be a set of linearly independent vectors
to ensure that cxlLx l in gr. Then the linear operator L:( gr, ff) ~ (qy, ff) is uniquely determined
by the n pairs of mappings Y¡ = Lx¡, for i = 1,2, ... ,n. Furthermore, with
respect to the basis {Xl' X2, ... , XII} of gr and a basis {u l , U2, ... , um} of qy, L can
be represented by an m x n matrix A with coefficients in the field ff. The ith
column of A is the representation OfYi with respect to the basis {Ul' U2,.··, u m }.
metric plane counter
gure 2-3. Given any Proof
. that is the sum úf the
vector that is the sum Let X be an arbitrary vector in gr. Since Xl' X2, ... , X" are linearly independent,
ltion is a linear trans the set of vectors qualifies as a basis. Consequently, the vector X can be ex
nple are aH equal to pressed uniquely as CXIX l +CX2X2 + ... +CX"x1l (Theorem 2-1). By the linearity
of L, we have
I
Lx = cxlLx l +cx2Lx2 + ... + cx"Lx"
=CX1Yl +CX2Y2 + ... +CX"YII
which implies that for any X in fI, Lx is uniquely determined by Y¡ = LXi' for
'unctions defined over
.i = 1, 2, ... ,n. This praves the first part of the theorem.
a linear space whose
Let the representation of y¡ with respect to {1Ul 1" lffi2" ' , . ,íl¡¡ml be
~t g be a continuous
[al ¡ a2; ... am¡J'; that is,
l
l
amI
22 LINEAR SPACES AND LINEAR OPERATORS
Note that the elements of A are in the field 9' and the ith column of A is the between IX and eX and 1
representation of y i with respect to the basis of 1fJ/. With respect to the basis related by eX = Pez and
{x¡,x 2 , ... ,x n} of (gr, 9') and the basis {u¡,u 2 , ... ,u m} of (1fJ/, 9'), the linear in the field 9' and the
operator y = Lx can be written as the basis {el' e2, ... ,e
[u l U2 ... umJJI= L[x l X2 xnJex (2-28)
where JI ~ [,81 ,82 . .. ,8mJ' and ex ~ [al a2 anJ' are the representa- Hence, by the uniqué
tions ofy and x, respectively. After the bases are chosen, there are no differences we have Aa=PAP-l
between specifying x, y, and ex, JI; hence in studying y = Lx, we may just study
the relationship between JI and ex. By substituting (2-27) into (2-28), we obtain
or
[u l u2 ... umJJI= [u l U2 ... umJAex (2-29)
Independent
of basis
L
x ----+y(=Lx) ith column:
Representation of
]
ith
f
column:
Representation of
J
A = ~¡ wit? respect to ,P = e~ Wi!h respect to or
1
~
'th
column:
Representation of
] llh coiumn:
Representation of Yl = LXI = [Xl
Basis Á = Le¡ with respect to ,Q = e¡ with respect to
the basis {el' e 2, ... , en} Hence the representat
[el
{el' ez,· .. ,en}
Figure 2-4 Relationships between dilTerent representations of the same operator.
LINEAR OPERATORS AND THElR REPRESENTATlONS 23
ith column of A is the between IX and (j and between p and ~ have been established in (2-20); they are
:h respect to the basis related by (j = Prx and ~ = pp, where Pis a nonsingular matrix with coefficients
of (óJI, ~), the linear in the field ff and the ith column of P is the representation of e i with respect to
the basis {el, ez, ... , en}. From Figure 2-4, we have
(2-28) ~= A(j and ~= Pp= PAIX= PAP-¡(j
nJ' are the representa Hence, by the uniqueness of a representation with respect to a specific basis,
here are no differences we have A(j = PAP-¡(j. Since the relation holds for any (j, we conclude that
Lx, we may just study A = P AP - ¡ = Q - ¡ AQ (2-31a)
into (2-28), we obtain
or A = P - ¡AP = QAQ - ¡ (2-31b)
lm]AIX (2-29)
where Q ~ P- . 1
implies that
Two matrices A and A are said to be similar if there exists a nonsingular
(2-30) matrix P satisfying (2-31). The transformation defined in (2-31) is caBed a
similarity transformation. Clearly, al! the matrix representations (with respeet
~, ff) are chosen, the
to different bases) of the same operator are similar.
; in ff. Q.E.D.
Example 3
Jn between the repre
that A depends on the Consider the linear operator L of Example 1 shown in Figure 2-3. If we choose
ent representations of {Xl' Xz} as a basis, then
"th column:
Representation of
~ Y¡=LX¡~[~~. X3] [-~J and Y3=Lx3 =[x¡ X3J[-~J
r
, Q = e, with respect to
{el. e z,··., en} Bence the representation of L· with respect t6 the basis {x¡, X3} is
he same operator.
24 LINEAR SPACES AND LINEAR OPERATORS
The reader is advised to find the P matrix for this example and verify A =
PAP- 1 . • Ab=
In matrix theory, a matrix is introduced as an array ofnumbers. With the It can be shown that tI
concepts of linear operator and representation, we shall now give a new inter
pretation of a matrix. Given an n x n matrix A with coefficients in a field ff,
if it is not specified to be a representation of sorne operator, we shall consider it Since the set of vectors
as a linear operator that maps (ff n , ff) into itself. 7 The matrix A is independent basis. We compute TI
of the basis chosen for (ff n , ff). However, if the set of the vectors 01' 02, ... ,on It is clear that
in Equation (2-13) is chosen as a basis of (ff n , ff), then the representation of
the linear operator A is identical to the linear operator A (a matrix) itself. This
can be checked by using the fact that the ith column of the representation is
equal to the representation of Ao¡ with respect to the basis {o 1o 02, ... , 0n}. lf
a¡ is the ith column of A, then AO i = a i . Now the representation of a¡ with
respect to the basis (2-13) is identical to itself. Therefore we conclude that the
representation of a matrix (a linear operator) with respect to the basis (2-13)
is identical to itself. For a matrix (an operator), Figure 2-4 can be modified as in
Figure 2-5. The equation Q = [q 1 q2 ... qnJ follows from the fact that the and
ith column ofQ is the representation of q¡ with respect to the basis {0 1o "2,"" Oll}'
If a basis {ql' q2, ... , qn} is chosen for (!F", ff), a matrix A has a representa The last equation is ob
tion A. From Figure 2-5, we see that the matrix representation A may be respect to the basis {b,
computed either from Theorem 2-2 or from a similarity transformation. In
most of the problems encountered in this book, it is always much easier to
compute A from Theorem 2-2 than from using a similarity transformation.
Qr IQ Á=Q-¡AQ
q2 .. ,
Á
Basis [q¡ qn] ~
'Ji Q=[q¡ q2 qn] =p-¡
• Then 8
Ir, we shall consider it Since the set of vectors b, Ab, and A2b are linearly independent, it qualifies as a
atrix A is independent basis. We compute now the representation of A with respect to this basis.
~ vectors "1' "2' ... ,"n It is clear that
A(b)~ [b
the representation of
la matrix) itself. This
the representation is
Ab A'bJ [!J
lis {" l' n2 , . . . , "n}. lf
~esentation of a¡ with
we conclude that the
:ct to the basis (2-13)
A(Ab)~ [b Ab A'bJ m
A(A'b)~ [b [-:~J
f can be modified as in
from the fact that the and Ab A'bJ
le basis {n¡, "2,···, Un}.
X A has a representa
The last equation is obtained from (2-32). Hence the representation of A with
esentation Á may be respect to the basis {b, Ab, A2b} is
I transforrnation. In
ways much easier to 0:0: 17]
ity transformation. A= 1 ;0: -15
[o 1 5
I I
I •
The matrix A can also be obtained from Q - 1 AQ, but it requires an inversion
of a matrix and n 3 multiplications. However, we may use Á = Q -1 AQ, or
more easily, QA = AQ, to check our resulto The reader is asked to verify
[~ -~l[~ ~ _~~l ~
-1
-~]
-1
O = [_; O
1 -iJ O 1 ~J 4 3 1 -3
Example 5
We extend Example 4 to the general case. Let A be an n x n square matrix
with real coefficients. lf there exists a real vector b such that the vectors b,
Ab, ... ,An-Ib are linearly independent and if Anb= -(;(nb-(;(n_IAb- ...
(;( 1 An - 1 b (see Section 2-7), then the representation of A with respect to the basis
{b, Ab, ... , A n - 1 b} is
~~ Aql
of
O O
1 O
O -(;(n
O -(;(n-I
~} J A=
O 1 O -(;("-2
(2-33)
O O O -(;(2
O O 1 -(;(1
•
8 A 2 ~ AA, A 3 ~ AAA.
26 LINEAR SPACES AND LINEAR OPERATORS
A matrix of the form shown in (2-33) or its transpose is said to be in the operator which maps
companionform. See Problem 2-26. This form will constantly arise in this text. n
(ff ,ff) that undergo
As an aid in memorizing Figure 2-5, we write A = Q - 1AQ as
D&finition 2-9
QA=AQ
The range of a linear (
Since Q= [ql q2 qnJ, it can be further written as
flI (A)= {all the elem(
[ql q2 qnJ A = [Aql Aq2 AqnJ (2-34)
vector x ir
From (2-34), we see that the ith column of A is indeed the representation of
Aqi with respect to the basis {ql' q2, ... , qn}. Theorem 2-3
We pose the following question to conclude this section: Since a linear
operator has many representations, is it possible to choose one set of basis The range of a linear (
vectors such that the representation is nice and simple? The answer is affirma
tive. In order to give a solution, we must first study linear algebraic equations. Proof
where the given aJs and y/s are assumed to be elements of a field ff and the
unknown x/s are also required to be in the same field ff. This set of equations where Xi, for i = 1, 2, ..
can be written in matrix form as range space fAlCA) is, b
Ax=y (2-36) (ff n , ff). lt is the sam
(2-37) ranging througl
[
~~: ~~~ :::
that flI(A) is the set of
Since fAlCA) is a linear s¡:
where A~ : : number of linearly in
9j'(Aj ¿s the maximurn n.
a m1 a m2
Clearly, A is an m x n matrix, x is an n x 1 vector, and y is an m x 1 vector. Definition 2-10
No restriction is made on the integer m; it may be larger than, equal to, or The rank of a matrix j
smaller thanthe integer n. Two questions can be raised in regard to this set of independent columns
equations: first, the existence of a solution and, second, the number of solutions. of A.
More specifically, suppose the matrix A and the vector y in Equation (2-36) are
given; the first question is concerned with the condition on A and y under which Example 1
at leastone vectorx exists such that Ax = y. If solutions exist, then the second
question is concenled with the number of linearly independent vectors x such Consider the matrix
that Ax == y. Inorder to answer these questions, the rankand the nullity of
the matrix A have to be introduced.
We have agreed in the previous section to consider the matrix A as a linear
SYSTEMS OF LINEAR ALGEBRAIC EQUATlONS 27
se is said to be in the operator which maps (ff", ff) into (ff m , ff). RecaB that the linear space
tant1Yarise in this tex t. (ff", ff) that undergoes transforrnation is caBed the domain of A.
-IAQ as
D&finition 2-9
The range of a linear operator A is the set Yf.(A) defined by
fJJI (A)= {aH the elements y of (ff m , ff) for which there exists at least one
(2-34 )
vector x in (ff", ff) such that y = Ax} !II
the representation of
Theorem 2-3
ection: Since a linear
lOO se one set of basis The range of a linear operator A is a subspace of (ff m , ff).
The answer is affirma
ar algebraic equations. Proof
lf YI and y z are elements of Yf.(A), then by definition there exist vectors XI and
Xz in (ff", ff) such that YI =Ax l , yz=Ax z. We claim that for any (XI and (X2
in ff, the vector (X IY I + (XzY z is also an element of fJJI(A). lndeed, by the linearity
of A, it is easy to show that (XIYI + (XzYz = A((Xlx l + (Xzx 2), and thus the vector
(X IX I + (X2XZ is an element of (ff", ff). Hence the range .<?l(A) is a subspace of
(ff m , ff) (see the remark following Definition 2-3). Q.E.D.
(2-35 ) Let the ith column of A be denoted by a¡; that is, A = [al az
then the matrix equati.on (2-36) can be written as
(2-37)
:s of a field ff and the
This set of equations where Xi, for i = 1,2, ... , n, are components of x and are elements of ff. The
range space fJJI(A) is, by definition, the set of y such that y = Ax for some x in
(2-36) (ff", ff). lt is the same as sayingthat fJJI(A) is the set ofy with XI' X2"'" X" in
(2-37) ranging through aB the possible values of ff. Therefore we conclude
that fJJI(A) is the set of all the possible linear combinations of the columns of A.
Since fJJI(A) is a linear space, its dimension is defined and is equal to the maximum
number of linearly independent vectors in fJJI(A). Hence, the dimension of
fJJI(A) is the maximum number of linearly inde pendent columns in A.
he matrix A as a linear
28 LINEAR SPACES AND LINEAR OPERATORS
The computation of the rank of a matrix on digital computers, however, is The design of con
not a simple problem. Because of limited accuracies on digital computers, reduced to the solutio
rounding errors always arise on numerical computations. Suppose a matrix, very important in OUl
after transformation, becomes then a11 rows of A are 1
lf A has a full row ranl
[A:y] cannot increase
spanned by the colun
where e is a very small number, say, 10 -1 0. This e may arise from the given combination of the co
data (assuming no rounding errors) or from rounding errors. lf e arises from there exists a x such ti
rounding errors, we should consider e as zero, and the matrix has rank 1. lf said to have a full coh
e is due to the given data, we cannot consider it as a zero, and the matrix has vector a will lie in the
rank 2. To determine what value is small enough to be considered as a zero is a be written as a linear c(
complicated problem in computer computations. For problems encountered a 1 x m vector b such t
on matrix computation, the reader is referred to References S181, S182, S200, After we find out th
and S212. to ask how many solUl
In matrix theory, the ranle of a matrix is defined as the largest order of aH 'Ne discuss only the ha
nonvanishing minors of A. In other words, the matrix A has rank k if and
only if there is at least one minor of order k in A that does not vanish and every Definition 2-11
minor of order higher than k vanishes. This definition and Definition 2-10 The null space of a line
are, in fact, equivalent; the proof can be found, for example, in Reference 43.
A consequence is that a square matrix has full rank ifand only ifthe determinant X(A)= {aH
ofthe matrix is different from zero; or correspondingly, a matrix is nonsingular The dimension of X(A
if and only if all the rows and columns of the matrix are linearly independent.
With the concepts of range space and rank, we are ready to study °the
In other words, tht
o
1. Given A and given a vector y in (ff m , ff), there exists a vector x such that
ms of aH the columns
Ax = Y if and only if the vector y is an element of P4(A), or equivalently,
ltions óf the first two
i A are linearly depen p(A) = p([A:y])
• m
2. Given A, for every y in (ff , ff), there exists a vector x such that Ax = Y
if and only if ~(A) = (ff m , ff), or equivalently, p(A) = m.
quence of elementary
lroperty that the rank
Proof
cations of elementary
o the upper triangular 1. lt follows immediately from the definition of the range space of A. lf the
ler of nonzero rows in vector y is not an element of ~(A), the equation Ax = Y has no solution and
umber of linear inde is said to be inconsistent.
of independent rows. 2. The rank of A, p(A), is by definition the dimension of ~(A). Since ~(A)
is a subspace of (ff m , ff), if p(A) = m, then ~(A)=(ffm, ff). lf ~(A)=
(ff m , ff), then for any y in (ff m , ff), there exists a vector x such that Ax = y.
mns
lf p(A) < m, there exists at least one nonzero vector y in (ff m , ff), but not in
~(A), for which there exists no x such that Ax = y. Q.E. D.
(2-38)
tmputers, however, is The design of compensators to achieve various design objectives can be
)n digital computers, reduced to the solution of linear algebraic equations. Hence this theorem is
s. Suppose a matrix, very important in our application. lf A is an m x n matrix and if p(A) = m,
then all rows of A are linearly independent and A is said to have afull row rank.
lf A has a full row rank, no matter what column y is appended to A, the rank of
[A:y] cannot increase and is equal to p(A). In other words, y lies in the space
spanned by the columns of A and, consequently, can be written as a linear
I arise from the given combination of the columns of A. Hence if A has a full row rank, for any y,
Tors. lf e arises from there exists a x such that Ax = y. Similarly, if p(A) = n, the m x n matrix A is
natrix has rank 1. lf said to have a full column rank. lf A has a full column rank, then every 1 x n
o, and the matrix has vector a willlie in the space spanned by the rows of A and, consequently, can
msidered as a zero is a be written as a linear combination of the rows of A. In other words, there exists
lroblems encountered a 1 x m vector b such that bA = a.
Ices S181, S182, S200, After we find out that a linear equation has at least one solution, it is natural
to ask how many solutions it may have. Instead of studying the general case,
he largest order of all we discuss only the homogeneous linear equation Ax: = O.
. A has rank k if and
; not vanish and every Definition 2-11
I and Definition 2-10 The null space of a linear operator A is the set .IV(A) defined by
nple, in Reference 43.
mly ifthe determinant .IV(A) = {aH the elements x of (ff", ff)for which Ax = O}
1 matrix is nonsingular
The dimension of .IV(A) is called the nullity of A and is denoted by v(A). •
inearly independent.
e ready to study the Inother words, the null space fi(A) is the set of all solutions ofAx =0. 9
lt iseasy to show that .A!(A) is indeed a linear space. lf the dimension of
fi(A), v(A), is O, then fi(A) consists of only the zero vector, and the only
9 It
is also called the right null space or A. The set of all y satisfying y A = O will be called the lert
atrix A maps (ff", ff)
null space of A. See Problem 2-51.
30 LINEAR SPACES AND LINEAR OPERATORS
solution ofAx = Ois x = O. Ir v(A) = k, then the equation Ax = Ohas k linearly be arbitrarily chosen.
independent vector solutions. vectors in Y(A). Hen
Note that the null space is a subspace of the domain (ff n , ff), whereas the theorem; its formal pn
range space is a subspace of (ff m , ff).
Theorem 2-5
Example 2
Let A be an m x n mat
Consider the matrix
h[! i i ~ =iJ
Corollary 2-5
The number of linear
which maps (~5,~) into (~3, ~). lt is easy to check that the last three columns n - p(A), where n is th
of A are linearly dependent on the first two col umns of A. Hence the rank of linearly independent c(
A, p(A), is equal to 2. Let x = [Xl Xz X3 X4 xs]'. Then
This corollary follc
null space of A. 1t is
is x = O, which is calle
find a nonzero vector x
then Ax = O has a nor
(2-39) det (A) = O, where det s
We introduce three
Since the vectors [O 1 2J and [1 2 O]' are linearly independent, we Theorem 2-6 (Sylves
conclude from (2-39) that a vector x satisfies Ax = Oif and only if Let A, B be q x n and 1
are linearly independent, and that ~very solution ofAx = O must be a linear
combination of these three vectors. Therefore the set of vectors 'forro a: basis of
.A'~(A) and v(A)= 3. I
We see fram this example that thenumber of equations that the vectors of
()?, J)
JIf(A) should obey is equal to p(A) and that there are n components in every
vector of X(A). Therefore n - p(A) components of the vectors of X(A) can Figure 2"6 A composite
SYSTEMS OF LINEAR ALGEBRAIC EQUATlONS 31
n Ax = Ohas k linearly be arbitrarily chosen. Consequently, there are n - peA) linearly independent
vectors in .;V(A). Hence we conclude that n - peA) = veA). We state this as a
1 (ff n , ff), whereas the theorem; its formal proof can be found, for example, in References 43 and 86.
Theorem 2-5
Let A be an m x n matrix. Then
peA) + veA) = n
Corollary 2-5
The number of linearly independent vector solutions ofAx = O is equal to
the last three columns n - peA), where n is the number of columns in A, and peA) is the number of
A. Hence the rank of linearly independent columns in A. •
Then
This corollary follows directly from Theorem 2-5 and the definition of the
null space of A. lt is clear that if peA) = n, then the only solution ofAx = O
x, [=iJ is x = O, which is called the trivial solution. lf peA) < n, then we can always
find a nonzero vector x such that Ax = O. In particular, if A is a square matrix,
then Ax = O has a nontrivial solution if and only if peA) < n, or equivalently,
X,)[i] (2-39)
det (A) = 0, where det stands for the determinant.
We introduce three useful theorems to conclude this section.
Proof
fA, peA). The solution
. two equations; hence The composite matrix AB can be considered as two linear transformations
ed. Let X3 = 1, X4=0, applied successively to (ffP, ff) as shown in Figure 2-6. Since the domain
~ 5 = O; then Xl = °
and
"2 = 1. lt is clear that
of AB is ~(B) and the range of AB is a subspace of ~(A), we have immediately
p(AB) .:5:min (p(A), p(B)) by using (2-38). From Figure 2-6 we have p(AB) =
-T---,t-- I ------~--
r I.L --r- --- .
--T-- -- IR(AB'L R{A)
,iB) R{B)
p(B) - d, where d is the dimension of the intersection of ~(B) and .K(A).10 The assumption of p(A
dimension oL/nA) is n - p(A); hence, d S n - p(A). Consequently, p(AB) 2: ACl=[!31 f32
p(B)-n+p(A). Q.E.D.
In words, the rank of a matrix will not change after the pre- or postmulti With the background
plication of a nonsingular matrix. Because of this property, gaussian elimina posed at the end of Sec
tion or the row-searching algorithm discussed in Appendix A can be used to that map (en, iC) into i1
compute the rank of a matrix. to any operator that r
The reason for restrict
Theorem 2-8 immediately.
Let A be an n x n n
Let A be an m x n matrix with coefficients in a field, and let A * be the complex
consider A as a linear
conjugate transpose of A. Then
1. p(A) = n if and only if p(A *A) = n, or equivalently, Definition 2-12
det (A*A) 1=0 Let A be a linear oper
2. p(A) = m if and only if p(AA *) = m, or equivalently, is ca11ed an eigenvalue
Ax = ílx. Anynonzerc
det (AA*) 1=0 associated with the eig
~(B)and .;V(A).10 The assumption of p(A) = n. The equality A*Aa. = O implies a.*A*Aa. = O. Let
onsequently, p(AB) ~ Aor:=[.81 .82 ... .8mJ'. Then
Q.E.D. a.*A* = [,B! .8!.,. .8::'t]
and a.*A*Aor: = 1.81\2 +1.821 + ... +1.8mI 2
2
n)= p(A) Hence a.*A*Aa. = Oimplies .8 i = 0, for i = 1, 2, ... , m; or, equivalently, Aa. = O,
which, in turn, implies a. = Ofrom the assumption of p(A) = n. Therefore we
conclude that p(A*A) = n.
2. This part can be similarly proved or directIy deduced from the foregoing by
using the fact p(A) = p(A *). Q.E. D.
<\)
2-6 Eigenvectors, Generalized Eigenvectors, and Jordan
I Form Representations of a Linear Operator
the pre- or postmulti With the background of Section 2-5, we are now ready to study the problem
erty, gaussian elimina posed at the end of Section 2-4. We discuss in this section only linear operators
ndix A can be used to that map (en, C) into itself with the understanding that the results are applicable
to any operator that maps a finite-dimensional linear space over Cinto itself.
The reason for restricting the field to the field of comp1ex numbers will be seen
immediately.
Let A be an n x n matrix with coefficients in the field C. We have agreed to
let A * be the complex
consider A as a linear operator that maps (en, C) into (en, C).
Definition 2-12
Let A be a linear operator that maps (en, C) into itself. Then a scalar A in e
is called an eigenva/ue of A if there exists a nonzero vector x in en such that
Ax = AX. Any nonzero vector x satisfying Ax = AX is called an eigenveetor of A
I associated with the eigenvalue A. 11 I
Theorem 2-9
we have
Let Al' Az, ... , A.. be the distinct eigenvalues of A, and let Vi be an eigenvector
of A associated with A¡, for i = 1, 2, ... , n. Then the set {v l' Vz, ... , v,,} is linearly We conclude that if
independent (over ic). into itself are aH distin
the operator A has a d
Proof the diagonal.
We prove the theorem by contradiction. Suppose VI' V z, ... , v.. are linearly
Example 2
dependent; then there exist CI.¡, Cl.z, ••• ,!X" (not aH zero) in ic such that
Cl.IV I +azv z ... +a"v" =0 (2-42) Consider
We assume Cl. 1 fO. If Cl. 1 =0, we may reorder Ai in such a way that al fO.
REPRESENTATIONS OF A LINEAR OPERA1OR 35
(2-41 )
(A - AzI)(A - A3 1) ... (A - AJ) Ct <XiV) = O (2-43)
i= Z
natrix A has no eigen
implies <Xl = O. This is a contradiction. Thus, the set of vectors {v¡, Vz, ... , VII}
omplex numbers, there is linearly independent and qualifies as a basis. Q.E.D.
H) as a linear operator
trix A has eigenvalues Let A be the representation of A with respect to the basis {v¡, Vz, ... , VII}.
I Recall from Figure 2-5 that the ith column of A is the representation of Av¡ =
A¡viwithrespectto{vl,vZ, ... ,vlI}-thatis,[O O Ai O O]',where
oo. o ••
O O . o o A II
(A-AII)VI=[l~i -~~J[~::J=o
Case 2: The eigenvalues of A are not al! distinct Since the matrix (A - A
Unlike the previous case, if an operator A has repeated eigenvalues, it is not Consequently, we can
always possible to find a diagonal matrix representation. We shall use examples VI = [1 O O]', associ
to illustrate the difficulty that may arise for matrices with repeated eigenvalues. A3 = 2 can be found as
sufficient to form abas
Example 3
From this example
Consider
values, it is not alway
Consequently, the A c:
it is possible to find a s
is almost a diagonal fa
eigenvalues of A on thl
The eigenvalues of A are Al = 1, A2 = 1, and A3 = 2. The eigenvectors associated
example, if A has an ~
with Al can be obtained by solving the following homogeneous equations:
with multiplicity 1, ther
(A-A'I)v~[~ ~ -nv~o
forms.
(2-45)
Note that the matrix (A - AII) has rank 1; therefore, two linearly independent
vector solutions can be found for (2-45) (see Corollary 2-5). Clearly, VI =
[1 O O]' and V2 = [O 1 O]' are two linearly independent eigenvectors
associated with Al = A2 = 1. An eigenvector associated with A3 = 2 can be
found as v 3 =[ -1 O 1]'. Since the set of vectors {V I,V 2 ,V 3 } is linearly
independent, it qualifies as a basis. The representation of A with respect to
{v 1> V2, V3} is Al O: O O
O Al 1 :O O
A= [O1 O0J
1 O I
O O Al: O
o---6--o-~-il-:
O
O
O O 2 1 L __
O O O O: A2
In this example, although A has repeated eigenvalues, it can still be diag
onalized. However, this is not always the case, as can be seen from the following
Which form it will as~
example.
discussed in the next su
¡
'"
J=o
(2-46)
O1 2J
(A-A 1 I)v= O O 3 v=O
formation. I [O O 1
Sínce the matrix (A - ..1. 11) has rank 2, the null space of (A - ,111) has dimension 1.
:d eigenvalues, it is not
ConsequentIy, we can find only one linearly independent eigenvector, say
We shall use exampies
VI = [1 O O]', associated with Al =..1. 2 = 1. An eigenvector associated with
h repeated eigenvalues.
. 1. 3 = 2 can be found as V 3 = [5 3 1]'. Clearly the two eigenvectors are not
sufficient to form a basis of (C 3 , C). I
{O
(2-45)
linearly independent
Al : O
o--:-iL
-:
1 .1.
__
O
O
¡
O
O
O
O
f
A1 1 : O ,O O
O )'1:L O
______ O , O
l
ry 2-5). Clearly, '"'1= O O:!e l O O' , O O: l¡ 1 ,O ,
l~ ~ :_~---~'-:-tJ
- - - ..! - - -.
lependent eigenvectors O O O: Al' O _ _ _ 1---:
O Al 1 , :0 O
O Al 1 O O
O O Al : O O O O Al 1 ,
O (2-47)
I
O---O--O-~-il-: O
O O O Al : O
L __
O O O
1_ _ _
,
O , ,12 0- - -O -- 0- - -O -:-l;
les, it can still be diag
seen from the following
Which form it will assunie depends on the characteristics of A and will be
discussed in the next subsection. The matrices in (2-47) are all ofblock-diagonal
38 LINEAR SPACES AND LINEAR OPERATORS
A 1 O O O
O A 1 O O
(2-48)
O O O 1 O
O O O A 1 and
O O O O A Th is set of vectors ['i
of length k.
with the same eigenvalue on the main diagonal and l's on the diagonal just
Let JV¡ denote tb!
aboye the main diagonal. A matrix of this form is caBed al ordan block associ
that (A - AI)ix = o. lt
ated with A. A matrix is said to be in the lordan canonicalform, or the lordan
is a subspace of .;v¡+ 1
form, if its principal diagonal consists of Jordan blocks and the remaining
tion 2-13 is in .;Vk but
elements are zeros. The fourth matrix in (2-47) has two Jordan blocks associated
defined in (2-49) is in
with Al (one with order 3, the other with order 1) and one Jordan block associ
ated with A2' A diagonal matrix is clearly a special case ofthe Jordan form: all (A - AI)i1
of its Jordan blocks are of order 1. and (A-AI)i-l,
Every matrix which maps (C n , C) into itself has a Jordan-form representa
tion. The use of Jordan form is very convenient in developing a number of hence Vi is in .;Vi but r
concepts and results; hence it will be extensively used in the remainder of this Let A be an n x n
chapter. discuss in the followin
vectors of A associatet
alized eigenvectors of
Derivation of a Jordan-form representation. 12 In this subsection, i =0,1,2, ... , until rar
we discuss how to find a set of basis vectors so that the representation of A with notations, weassume n
respect to this set of basis vectors is in aJordan formo The basis vectors to be are as shown in Table :
used are called the generalized eigenvectors. and is equal to, f
.;VO e ';vI e ';v2 e ....
Defihition 2-13 Now we are read)
associated with A. Bet
A vector v is said to be a generalized eigenvector of grade k of A associated with only one linearly inde¡
A if and only if 13
(A -AI)kv = O
and (A -AIt-lv +0
Note that if k = 1, Definition 2-13 reduces to (A - AI)v = Oand v O, which + Table 2-2 Chains of Ge
is the definition of an eigenvector. Hence the term "generalized eigenvector"
is well-justified. p(A - AI)o = 10 Vo =0
Let v be a generalized eigenvector of grade k associated with the eigenvalue p(A -AI)=7 VI =3
p(A-AI)2=4· v2 =6
p(A-AI)3=3 v3 =7
PlA-AI)4",i v4 =8
12 This section may be skipped withollt loss of continllity. However, it is sllggested that the reader
glances throllgh it to gain a better feeling aboLlt the Jordan-form representation.
13 (A - AI)k ~ (A- AI)(A - Al)' .. (A - AI)(k terms), (A _ .U)O ~ 1.
REPRESENTATlONS OF A LINEAR OPERATOR 39
A.. Define
Vk ~ V
Vk- l ~ (A-AI)v=(A-AI)v k
Vk- 2 ~ (A-AI)2v =(A-AI)vk_' (2-49)
(2-48)
and
This set of vectors {V¡, v2, ... , vk} is called a chain of generalized eigenvectors
of length k.
on the diagonal just Let .;Vi denote the null space of (A - AI)i, that is, .;Vi consists of all x such
1 Jordan block associ
that (A-AI)ix=O. It is clear that ifx is in .;Vi' then it is in ';vi+l' Hence.;Vi
11 form, or the J ordan isasubspaceof .;Vi+"denotedas ';viC .;Vi +l' Cleariy,thevdefinedin Defini
s and the remaining tion 2-13 is in ';vk but nbt in .;v k-l' In fact, for i = 1, 2, ... , k, Vi = (A - AI)k-i v
dan blocks associated
defined in (2-49) is in .;Vi but not in ';vi-l' Indeed, we have
Jordan block associ
fthe Jordan form: all (A - AI)ivi = (A -AlhA -AI)k-i v = (A - AI)kv =0
and (A -AI)i-' vi = (A - AI)i-'(A -AV-iV = (A - AI)k-' v 1=0
jan-form representa
e10ping a number of hence Vi is in .;Vi but not in ';vi-l'
the remainder of this Let A be an n x n matrix and have eigenvalue A with multiplicity m. We
discuss in the following how to find m linearly independent generalized eigen
In this subsection, i = O, 1,2, ... , until rank (A - AI)k = n-m. In order not to be overwhelmed by
,resentation of A with . notations, weassume n = 10, m = 8, k =4,and the ranks of(A -AW, i =0,1,2,3,4,
le basis vectors to be are as shown in Table 2-2. The nullity Vi is the dimension ofthe null space ';vi'
I )(2--~-
length 4
L
40 LINEAR SPACES AND LINEAR OPERATORS
Proof
First we show that if {u z, v, w} is linearly independent, then {u¡, v1. W1} is linearly
independent. Suppose {u¡, v1 , w¡} is not linearly independent, then there
exist ci,i=1,2,3, not aH zero, such that C1U1 +C Zv1 +C3W1 =0. However,
we have
This is a Jordan-form m;
Since y is a vector in X z, the only way ta have Ey = {D is that y = Gi. Siúce {;nz, '1, w}
is linearly independent by assumption, y = O implies Ci = O, i = 1,2,3. This
is a contradiction. Hence if {u z, v, w} is linearly independent, so is {u¡, Vl' w¡}. Example 4 (Continued
Now we show that the generalized eigenvectors {u i , i = 1, 2,3,4; vj' Wj,j = 1, Consider
2} are linearly independent. Consider
(2-51 )
which implies; because of B3 U4 jo O, C4 = O. Similarly, we can show c 3 =0 by 14 rhis number is called the 9
multlplicity o¡os the number
applying B2 to (2-51). With C3 = C4 = O, the application of B to (2-51) yields
orders of aH lordan blocks
REPRESENTATIONS OF A LINEAR OPERATOR 41
which implies, because ofthe linear independenceof {uzo Vz, Wz}, Cz = C6 = Cs =0.
(2-50) Finally, we have C1 =C s =C7 =0 following the linear independence of {U 1, '1 1, w 1}.
This completes the proof of this theorem. Q.E. D.
;: linearly independent
a vector; therefore we Theorem 2 -11
; but not in j{ z. Con
: Vz - v1 = 3, there are The generalized eigenvectors of A associated with different eigenvalues are
·50) is one of them, we linearly independent. I
early independent and
This theorem can be proved as in Theorem 2-10 by applying repetitively
(A - A¡I)k(A - Ajlf The proof is left as an exercise.
Now we discuss the representation of A with respect to Q ~ [u 1 Uz u 3
u 4 V1 VZ W1 WZ x x]. The last two vectors are the eigenvectors of A
1eigenvectors oflengfh assoqated with other eigenvalues. The first four columns ofthe new representa
, the number of vectors tion A are the representations of Au¡, i = 1, 2, 3, 4, with respect to {u¡, u z, u 3 , u 4 ,
search other vector in Vlo Vz, W1, Wz, X, x}. Because (A - AI)u1 = O, (A -AI)1z = u 1, (A - AI)u 3 =u z, and
.vectors of A associated (A-AI)u 4 =u 3, we have AU1=AU1=Q[A O O ... 0J',Au Z =u 1 +Au z =
Q[l A O O ... O]', AU 3 =U z +Au 3 =Q[O 1 A O ... OJ', and
AU 4 =u 3 +Au 4 =Q[O O 1 A O ... OJ', where the" prime denotes the
transpose. Proceeding similarly, the new representation A can be obtained as
A 1 O 0:0 O O O
°
A= O--Ó--Ó-O-:-~--(: O
, , O (2-52)
O O O 0:0 A: O
fu
m lo v lo W¡} is linearly
dependent, then there O O O O - 0- -6: ),- - i -:,
1
°
+ C3W1 =0. However, O O O O O 0:0 , A: _ _ _ _ _ _1
L +cswz=O (2-51 )
14 This number is caBed the geometric multiplicity of ), in Reference 86. In oLher words, geometric
VoIe can show C3 = O by
multiplicity is the number of lordan block S, and the (algebraic) multiplicity is the sum of the
of B to (2-51) yields orders of aH lordan blocks associated with .l..
l.
42 LINEAR SPACES AND LINEAR OPERATORS
Example 5
Its eigenvalues are Al = 1, A2 = 1, A3 = 2. An eigenvector associated with
),3 = 2 is v3 = [5 3 1J'. The rank of (A - Al1) is 2; hence we can find only one Transform the followi
eigenvector associated with Al' Consequently, we must use generalized eigen
vectors. We compute
B@(A-A'l)~[~
1 A
O
O :]
(A~A,Il'~[~ m~
:] ~[~ ¡]
1 1 O
O O O 1. Compute the eigen'
and
O O O det (A -AI)=
Since pB 2 = 1 = n-m, we stop here. We search a v such that B 2 v = O and Hence A has eigen'
Bv =/= O. Clearly, v = [O 1 O]' is such a vector. It is a generalized eigenvector plicity 1.
of grade 2. Let 2. Compute (A - 21Y,
B f:::,. (A -21) =
Theorems 2-10 and 2-11 imply that v1 , v2 , and v 3 are linearly independent.
This can also be checked by computing the determinant of [v 1 V2 v 3]. If
we use the set of vectors {v 1, V2' V3} as a basis, then the ith col umn of the new
representation A is the representation of Av; with respect to the basis {v 1 , v 2 , v3 }.
Since AV 1 =A 1V¡,Av 2 =v 1 +A 1V2 , and AV 3 =A 3 V3 , the representations of (A-21)2 =
Av l' Av 2 , and AV 3 with respect to the basis {v 1 , v2 , v 3 } are, respectively,
r:
where Xl = 1, }'3 =2. Hence we have
(A-21)' ~l
A=[~O 0:2
~-~-~J
__
(2-53)
Since peA - 21)3 = n
a generalized eigen'
This can also be obtained by usihg the similarity transformation
A=Q- 1 AQ
15 We use the ract that
~[i ~]
O
where Q =[v 1 v2 v,] 1
where A and e are square
O I
REPRESENTATlONS OF A LINEAR OPERATOR 43
, O5]
O 3 1. Compute the eigenvalues of AY
O 1 det (A -Al) = [(3 -..1.)(1-..1.) + 1](..1. -2)2[(1 _..1.)2 -1] =(..1. _2)5 A
such that B 2 v =0 and Hence A has eigenvalue 2 with multiplicity 5 and eigenvalue O with multi
~eneralized eigenvector plicity 1.
2. Compute (A - 21Y, for i = 1, 2, ... , as follows:
1 -1 1 1 O O
1 -1 -1 -1 O O
O O O O 1 1 p(A-21)=4
B ~ (A-21)=
O O O O -1 -1 v1 =6-4=2
O O O O -1 1
• linearly independent. O O O O 1 -1
at of [v 1 V 2 v3 ]. Ir
ith column of the new O O 2 2 O O
to the basis {v 1 , v2 , v3 }. O O 2 2 O O
le representations of O O O O O O p(A-2If =2
(A -21)2 =
re, respectively, O O O O O O V2 =4
O O O O 2 -2
O O O O -2 2
r~ ~l
O O O O
O O O O
~J
n
~ l~
v O O O p(A -21iY = 1
(A-2Ij' O O O O v3 =5
O O O -4
(2-53) O O O 4 -4
Since p(A - 21)3 = n-m = 1, we stop here. Because v3 - V2 = 1, we can find
a generalized eigenvector u of grade 3 such that B3u ;= O and B 2 u 1= O. It is
rmation
2 1 O
2 -1 O
~
O O 1
UI BZu= U z ~ Bu = u 3 ~ u=
O O O
O O O
O O O
This is a chain of generalized eigenvectors oflength 3. Because of Vz - VI = 2, In this example,
there are two linearIy independent vectors in .K z but not in .K l. The vector {w, V z, V¡, U 3, U z, u¡};
U z is one of them. We search a vector v which is independent of U z and has
the property BZv = O and Bv i= O. It can be readily verified that v =
[O O 1 -1 1 1]' is such a vector. Define
O O
O O
2 1
VI ~ Bv=
-2 -1
O 1 This is also called a J,
O 1 (2-55), we see that thl
below the main diago
Now we have found five generalized eigenvectors of A associated with A. = 2. In this book, we use
3. Compute an eigenvector associated with A. z =0. Let w be an eigenvector everything discussed
of A associated with A. z =0; then given in Equation (2-:
A Jordan-form re
3 -1 1 1 O O into itself is unique u¡:
1 1 -1 -1 O O Jordan blocks and ti
O O 2 O 1 1 w=O by A. However, bec~
(A A.zI)w = different Jordan-form
O O O 2 -1 -1
O O O O 1 1
O O O O 1 1
2-7 Functions (
Clearly, W = [O O O O -1]' is a solution
4. With respect to the basis {u I , u z, u 3 , VI, vz, w}, A has the following Jordan In this section we shall
form representation: tion that maps (en, 1[:)
extensively, because il
2 1 0:0 O O function of a matrix c:
021:0 O O
matrix, and then defil
O O 2:0 O O
matrix.
A= 0- -6 - -6 ~ 2- -¡ -: O (2-55)
O O 0:0 2:0
O O O - O--6 - 0 Polynomials of a s
(C", C) into (1[:",1[:). If
5. This may be checked by using
A=Q-¡AQ or and
FUNCTlONS OF A SQUARE MATRIX 45
Becauseofvz -V 1 =2, In this example, if we reorder the basis {Ul> u z, u3, v1, V Z, w} and use
10t in j( l' The vector {w, v z, Vl' U 3 , U z, U 1 } as a new basis, then the representation will be
:pendent of U z and has
lily verified that v =
efine
O A= (2-56)
O
1
1
1 This is also called a lordan-form representation. Comparing it with Equation
1 (2-55), we see that the new lordan block in (2-56) has l's on the diagonal just
below the main diagonal as a result ofthe different ordering ofthe basis vectors.
l associated with A. = 2. In this book, we use mostly the lordan block of the forro in (2-55). Certainly
:t w be an eigenvector everything discussed for this form can be modified and be applied to the form
given in Equation (2-56).
A lordan-forro representation of any linear operator A that maps (en, C)
o into itselfis unique up to the ordering of lordan blocks. That is, the number of
O lordan blocks and the order of each lordan block are uniquely determined
1 by A. However, because of different orderings of basis vectors, we may have
w=o different lordan-form representations of the same matrix.
-1
1
1
2-7 Functions of a Square Matrix
ution.
the following lordan In this section we shall study functions ofa square matrix or a linear transforma
tion that maps (en, C) into itself. We shall use the lordan-form representation
extensively, because in terms of this representation almost all properties of a
function of a matrix can be visualized. We study first polynomials of a square
matrix, and then define functions of a matrix in terms of polynomials of the
matrix.
(2-55 )
where 1 is a unit matrix. Let f(A) be a polynomial in A of finite degree; then polynomial of A is
f(A) can be defined in terms of(2-57). For example, if f(A) = A3 +2A z +6, then
f(A) ~ A 3 +2A z +61
We have shown in the preceding section that every square matrix A that Assume that aJorda:
maps (1[", C) into itself has a Jordan-form representation, or equivalently, there
exists a nonsingular constant matrix Q such that A = QAQ - I with A in a
Jordan canonical formo Since
index of A¡ in A.
One of the reasons to use the Jordan-form matrix is that if
The multiplicity (
(2-59) the matrix in (2-52),
nz = 1; for the matr
where Al and A z are square matrices, then n¡'::;; ni.
A=
k [A1O
where ñ¡ is the index
Definition 2-14
Proof
The minimal polynomial of a matrix A is the monic polynomial 16 if¡(A) of least
degree such that if¡(A) =0. I
Since the matrices A
as showing that tf¡(A)
'"f/e -6.rst shovv that tb
Note that the O in if¡(A) =0 is an n x n square matrix whose entries are all
A¡ consists of r Jorda
zero. A direct consequence of(2-58) is that f(A)=O ifand only if f(A)=O.
Consequently, the matrices A and A have the same minimal polynomial, or A¡=di
more general1y, similar matrices have the same minimal polynomial. Computing
the minimal polynomial of a matrix is general1y not a simple job; however, if
the Jordan-form representation of the matrix is available, its minimal poly
nomialcan be readily foundo
Let Al, Az, . oo,A m be the distirict eigenvalues of A with multiplicities
nI> nz, ooo,n m , respectivelyo It is .the same as saying that the characteristic
. 16.\ monic polynomial is. a polynomial the coefficient of whose highest power is 1. For example,
3x +1 aTid -x 2 +2x +4 are not monicpolynomials, but x 2 -4x +7 is.
FUNCTIONS OF A SQUARE MATRIX 47
where the n¡ x n¡ matrix A¡ denotes aB the Jordan blocks associated with A¡.
n (A - A¡)ñ¡
m
lj¡(A) =
¡= 1
Proof
nomiaP 6 lj¡(A) of least
Since the matrices A and A have the same minimal polynomial, it is the same
• as showing that lj¡(A) is the polynomial with least degree such that lj¡(A) = o.
We first show that the minimal polynomial of A¡ is ¡J¡¡()c) = (.>1-- AS'. Suppos\'::
. whose entries are aH
and only if feA) = O.
A¡ consists of r Jordan blocks associated with A¡. Then
nimal polynomial, or A¡ = diag (A il , A¡z, ... ,A¡r)
lynomial. Computing
imple job; however, if
:>le, its minimal poly- and
A with multiplícities
hat the characteristic
(Ail - A¡lt' A O _
.= .. ?. (Aiz ~ A;I)"'
.. (2-63)
t power is 1. For example, .
, is.
f. O O
48 LINEAR SPACES ANO LINEAR OPERATORS
Ifthe matrix (Aij - A;I) has dimension n¡j, then we have they have, respecth
=[~ ¿ ~
... O]
... O
minimal polynomiah
[~ ~ ~ ~
The Cayley-Ham
Theorem 2-12 (see Pr
(A¡j - A¡I)n¡j-¡ = °oll (2-64c)
The reason for in1
O O O O J in the following theoJ
and (Aij - AJ)k = O for any integer k ;::: nij (2-64d)
Theorem 2-13
By definition, ñ¡ is the largest order of the Jordan blocks in A¡, or equivalently,
Let A¡, Az, . .. , Am be
ñ¡=max(nij,j=1,2, ... ,r). Hence (Aij-A¡I)ñ,=O for j=1,2, ... ,r. Con
Let f and 9 be two pe
sequently, 1jJ¡(A;) =0. It is easy to see from (2-63) and (2-64) that ifljJ¡(A)=(A-o:)k
with either o: I=A¡ or k < ñ¡, then ljJ ¡(A;) 1=0. Hence we conclude that 1jJ¡ = (A - A¡)ñ¡ 1. feA) = g(A).
is the minimal polynomial o'f Aj . Now we claim that feA;) = Oif and only if f 2. Either f =h¡1jJ +f
is divisible without remainder by ljJ ¡, denoted as ljJ d.r Inde:ed, if ljJ ;\f; then A, and h¡ and h 2 a
f can be written as f = ljJ eh, where h is the quotient polynomial, and feA;) = 3.
1jJ¡(A;)h(A;) = O' h(A;) = O. If f is not divisible without remainder by ljJ ¡, then pe)(A¡) = g('
lcan be written as f = ljJ¡h +g where gis a polynomial of degree less than ñ¡.
Now f(A¡) = O implies g(A¡) = O. This contradicts the assumption that ljJ ¡ is
where j(l)(A;) ~ d~
{
the minimal polynomial of Á¡, for 9 is a polynomial of degree less than that of Proof
ljJ i and g( A¡) = O. With these preliminaries, the theorem can be readily proved.
From (2-62), we have IjJ(Á) = diag (IjJ(Á¡) IjJ(Á 2 ) ••• IjJ(Árn )). Since IjJ(Á¡) = O The equivalence of sta
i[ and only ifljJ eontains the [actor (il- ),¡)ñ;, we conclude that the mínima1 poly Statements 2 and 3 ar
nomial of Á and, correspondingly, of A is . .
rn
TI (A _A¡)ñ¡ Q.E.D.
i= 1 In order to apply 1
Example 1 The minimal polynol
form or by direct com
The matrices Therefore it is desirab
r~ ~ i ~J r~ ~ i ~J r~ ~ I~J
polynomial can be av
Corollary 2-13
Let the charaeteristic
al! have the same characteristic polynomial i1(A) = (A - 3)3(A -1); however,
-_._--_....::_..__ .:..~.~,----------"---~----_.:..._=--.=--.....:....-~=--_:...._-_:...-::~_.,---...:_-,--,~_:._-----
they have, respectively, (A - 3)(A - 1), (A - 3f(A -1), and (A - W(A -1) as
minimal polynomials. I
i= 1
Q.EoD.
Q.E.D.
In_order to apply this theorem, we must know the minimal polynomial of A.
The minimal polynomial can be obtained by transforming A into aJordan
form or by direct computation (Problem 2-42)0 Both methods are complicated.
;~ ~l
Therefore it is desirable to modify Theorem 2-13 so that the use of the minimal
polynomial can be avoided.
) 3
) O 1
° Corollary 2-13
Let the characteristic polynomial of A be
n (A - A¡)n
m
- W(A -1); however, Ll(A) ~ det (Al - A) = i
i= 1
SO LINEAR SPACES AND LINEAR OPERATORS
then f(A)=g(A). I
Solving these two eqt
This follows immediately from Theorem 2-13 by observing that the condition
(2-66) implies (2-65). The set of numbers ¡<L)(A;), for i = 1,2, ... ,m and l = A 100 =g(A) =c
0, 1,2, ... , ni -1 (there are totally n = I7= 1 n¡) are called the values of f on
the spectrum of A. Corollary 2-13 implies that any two polynomials that have
Obviously A 100 can a
the same values on the spectrum of A define the same matrix function. To
different g(A) such as 1
state it in a different way: Given n numbers, if we can construct a polynomial
which gives these numbers on the spectrum of A, then this polynomial defines
uniquely a matrix-valued function of A. It is well known that given any n Functions of a Sql
numbers, it is possible to find a polynomial g(A) of degree n - 1 that gives these
n numbers at sorne preassigned A. Hence if A is of order n, for any polynomial Definition 2-16
f(A), we can construct a polynomial of degree n - 1,
Let f(A) be a functie
g(A)=a o +a 1A + oo. +a n_ 1An- 1 (2-67) spectrum of A. If g(A
spectrum of A, then th,
such that g(A) = f(A) on the spectrum of A. Hence any polynomial of A can
be expressed as
This definition is :
f(A) =g(A) =aol +a 1 A + ... +a n_ 1An-1 be precise, functions e
This fact can also be deduced directly from Corollary 2-12 (Problem 2-38). (2-65). The condition
Corollary 2-13 is useful in computing any polynomial and, as will be dis is easier to obtain the
cussed, any function of A. lf A is of order n, the polynomial g(A) can be chosen willlead to the same J
as in (2-67) or as any polynomial of degree n - 1 with n independent parameters. If A is an n x n me
For example, if all eigenvalues, Ai' i = 1,2, ... , n, of A are distinct, then g(A) can find a polynomial
can be chosen as
n -1 n
g(A)= I
i=O
Pi n (A-Aj)
j= 1
which is equal to f(A
know that every funct
Ni
n-l i
or g(A) = I Pi f1 (A - Aj ) We summarize th,
i=O j= 1
an n x n matrix A al
In conclusion, the form of g(A) can be chosen to facilitate the computation.
polynomial of A, say
Example 2
Compute A 100, where
Let
A=[~ ~J
l
In other words, given f(A) = A 00, compute f(A). The characteristic poly wheré a o, al'" ., ex n - 1
nomial bf A is ~(A) = det (Al - A) = (A - 1)Z. Let g(A) be a polynomial of degree to compute these ex/s
n -1 = 1, say we have f(A)=g(A).
n independent paramé·
FUNCTIONS OF A SQUARE MATRIX 51
Now, from Corollary 2-13, if f(A.) = g(A.) on the spectrum of A, then f(A) = g(A).
On the spectrum of A, we have
,-1
(2-66) f(l) = g(l) (1)100 =exO +ex 1
100'(1)99=ex 1
• 1'(1) = g'(l)
Solving these two equations, we obtain ex 1 = 100 and exo = -99. Hence
ting thatthe condition
= 1, 2, ... , m and 1=
ed the values of f on
1 0J
AI00=g(A)=exoI+exlA=-99 [ O 1 +100 O 1 = O 200J
1 [1 2J [1
Jolynomials that have
Obviously A 100 can also be obtained by multiplying A 100 times or by using a
matrix function. To
different g(A.) such as g(A.) = exo +ex 1(A. -1) (Problem 2-33). •
mstruct a polynomial
ris polynomial defines
)wn that given any n Functions of a square matrix
: n - 1 that gives these
n, for any polynomial Definition 2-16
Let f(A.) be a function (not necessarily a polynomial) that is defined on the
(2-67) spectrum of A. If g(A.) is a polynomial that has the same values as f(A.) on the
spectrum ofA, then the matrix-valued function f(A) is definedas f(A)~g(A). •
polynomial of A can
n (A. - A.Jn¡
m
.1\(2) =
i= 1
Let
g(2)=exo +ex 1A. + ... +ex n_ 12 n- 1
e characteristic poly where exo, ex 1, ... ,ex n-1 are n unknowns. Next we use the n equations in (2-66)
polynomial of degree to compute these ex/s in terms of the values of f on the spectruID of A. Then
we have f(A) = g(A). We note that other polynomial g(A.) of degree n-1 with
n independent parameters cari also be used.
52 LINEAR SPACES AND LINEAR OPERATORS
Solving these equations, we obtain (,(0 = -2te' +e21'(,(1 =3tel +2el -2e 21 ,
and (,(2 = e21 - el - tel. Hence, we have Here in the last step v
If feA) = eJ..I, then
eA" = g(A ¡) = ( - 2te' +e 21 )I +(3tel + 2e' - 2e 21 )A 1 +(e21 - e' - tel)A ~
=
2e' - e
O
21
O
el
2é - 2e
O
2
']
e AI = re~'1 t;
[ _el +e21 O 2e 21 _el I
lo (
Example 4 Note that the derivati
Let A function of a n
therefore, the relations
A'~[~ j -~]
of a matrix. F or exan
lts characteristic polynomial is ~(A) = (A - 1)2(A - 2), which is the same as the and if
one of Al in Example 3. Hence we have the same g(A) as in Example 3. Con
sequently, we have
2e' -e2' then
e A21 = g(A 2 ) = O
[ 21 for any function f tt
e -é I
(2-71), any function o
mediately.
Example 5
Example 6
Given
Consider
1 O
~ ~~
[ A, Al 1
(2-68)
(n n) O
O
O
O
1.]
The characteristic polynomial of Á is (A - Alr. Let the polynomial g(A) be
FUNcnoNs OF A SQUARE MATRIX 53
of the form
g(A) = Cto +Ct 1(A - A¡} +Ctz(A - Ad + ... +Ct n-1(A - A¡)"-l
Then the conditions in (2-66) give immediately
, f(n-1)(A¡}
Ct O =f(A 1), Ct 1=f(A¡), ... , Ctn- 1= (n-l)!
<\¡)?
Hence,
. Leí g(A) = Cto +Ct 1A +
f '(A ) f(lI- 1)(' )
f(A) =g(A) = f(AI)1 +---yf- (A - A11) + ... + (n -1;t (A -A 11)n-1
•
Al _
l
e - ..
.
O
O eA,1
O
..
.
..
O
.
...
te Al ... t n-ZeA,l/(n - 2)!
eA,1
..
.
Note that the derivatives in (2-69) are taken with respect to Al' not to t.
(2-70)
11
Example 6
Consider
Al 1 O: O O
(2-68) OA 1 '1 : O O
A=O O A1 :O O (2-72)
-------~-~-----
O O O :A z 1
O O O: O A2
he polynomial g(A) be
54 LINEAR SPACES AND LINEAR OPERATORS
Example 7
lf feA) = (s - Af 1, where s is a complex variable, then
Consider the Jordan-
feA) = (si - A)-l
1 1 1
O O
S-Al (S-A l )2 (S-A l )3:
then
1 1 ,
O O O
S -Al (S-A l )2:
,
1 ,
O O , O O (2-74 )
,
s -Al ,
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ .l. _ _ _ _ _ _ _ _ _ _ _ _ _
Since (A -AlI)i is of
immediately to (2-69).
1
O O O
: s -A 2 (s -A2)2
Example 8
1
O O O O The exponential funct
s-A2 I
If the absolute values of aH the eigenvalues of A are smaller than p, it can also
be shown that the infinite series converges. For a proof, see Reference 77.
Instead of proving that Definitions 2-16 and 2-17 lead to exactIy the same
matrix function, we shall demonstrate this by using Definition 2-17 to derive
(2-73) (2-69).
Example 7
Consider the lordan-form matrix A given in (2-68). Let
o (2-74)
Since (A-A1I)i is of the form of (2-64), the matrix function (2-77) reduces
immediately to (2-69). •
1
Example 8
ih 2 Antn
ower series. We e).! = 1 + At +-- + ... + - + ...
2! n!
:tion of a matrix. We
pression of a function converges for all finite Aand t. Hence for any A, we have
1
eA! = L
<Xl
_tkAk (2-78)
k=O k!
I
00
lt is easy to verify that
f{Je)={l-Je)-1 = 1 +Je +Je z + ... = L Jek
norm 11'llz is called th
k=O
used mainly in the sta
converges for 1,1.1 < 1. Now if s is chosen sufficiently large, the absolute values if al! the eomponents 01
of all the eigenvalues of S-1 A are smaller than 1. Hence from Definition 2-17, The concept of nor
we have into itself, or equivalel
00
norm of a matrix A is •
{1-s- 1A)-1 = L {s-IA)k (2-85)
k=O
Hence from (2-84) we have
where "sup" stands for
(2-86) least upper bound of
IIAIIis, for any x in (C"
In this derivation, Equation (2-86) holds only for sufficiently large s. However,
it can be shown by analytic continuation that Equation (2-86) does hold for all s
except at the eigenvalues of A. Equation (2-86) can also be established from The norm of A is de
(2-82). Because of ff'[dh{t)/dt] = sff'[h{t)] - h{O), the application of the Laplace norm. For different 11"
transform to (2-82) yield~ then
sff'{e At ) - eO = Aff'{eA1 )
or ..(sl - A)ff'{e A1 ) = I
which yields imrrüidiate\y (2-86).For the matrices in{2-73) and (2-74), we can 17 May be skipped without Ir
also readily establish (2-86). and its study may be coup
-_ .. - .. - -- _..
- - - -----------_. __....._--,------_._----_._--
...
- _-..
All the concepts introduced in this section are applicable to any linear space
over the field of complex numbers or over the field of real numbers. However,
(2-82) for convenience in the discussion, we restrict ourself to the complex vector
space (C", C).
natrix commute (see The concept of the norm of a vector x in (C", C) is a generalization of the
idea of length. Any real-valued function of x, denoted by Ilxll, can be defined
X)) is defined as as a norm if it has the properties that for any x in (C", C) and any a in e
1. Ilxll ~O and Ilxll =0 if and only ifx =0.
(2-83)
2·ll ax ll=l a lllxll·
3. Ilx I + xzll ::;llxlll + Ilxzll.
The last inequality is called the triangular inequality.
Let x = [Xl Xz ..• X n ]'. Then the norm ofx can be chosen as
11
Ilxlld~ ¿
i; 1
IXil (2-87)
Ixd
}IZ
z (2-88)
It is easy to verify that each of them satisfies all the properties of a norm. The
norm 11'11z is called the euelidean norm. In this book, the concept of norm is
used mainly in the stability study; we use the fact that Ilxll is finite if and only
~, the absolute values if al! the eomponents ofx are finite.
froro Definition 2-17, The concept of norm can be extended to linear operators that map (C", C)
into itself, or equivalently, to square matrices with complex coefficients. The
norm of a matrix A is defined as
(2-85)
IIAII ~ ~~~ I\~III = 1I~~el IIAxII
where "sup" stands for supremum, the largest possible number of IIAxII or the
(2-86) least upper bound of IIAxII. An immediate consequence of the definition of
IIAII is, for any x in (C", C),
tly large s. However, (2-90)
IIAxII ::;IIAllllxll
·86) does hold for all s
) be established from The norm of A is defined through the norm ofx; hence it is called an indueed
lcation ofthe Laplace norm. For different Ilxll, we have different IIAII. For example, if Ilxlll is used,
then
73) and (2-74), we can 17 May be ski~ped without los~ (lC continuity The ~~terial in tbi's secti~-n is used only in Chapter 8,
and its study may be coupled with that chapter. .
58 LINEAR SPACES AND LINEAR OPERATORS
where aij is the ijth eleroent of A. lfllXl12 is used, then and IIABxII
and
Consequently we have
The sum of these two This magni tude gives
magnitudes gives the the norm of A.
norm of A. The inner product
(a) (b) In fact, this is the norn
3 Ilx 11 ~ 1
If we define Ilxll = ( (x, ,
2 !.~
----
.....:;:-----
--l--'¡=~/--/--':::::""k-l--l--+--
-4 -3 -2 ~Q.. 1 ------;.::",j
2 r-4_ 5
Xl
Proof
Ax
-2 The inequality is obvio
Figure 2'-7 Solid lines denote x.; broken lines denote Ax, where A = [_ ~ ~J
(a) IIAIll =4. (b) IIAllz =3.7. (e) tlAII", = 5. which gives the Schwat
. _ - . - 0
-
- -
----
(x, y) = (y,x)
: 2-7. (alx l +a 2x 2, y) =a l (Xl' y) +a 2(x 2, y)
(x,x»O for all xi=O
(2-91 )
(2-92) where the "overbar" denotes the complex conjugate of a number. The first
property implies that (x, x) is a real number. The first two properties imply
that (x, ay) =a(x, y).
In the complex vector space (C n , C), the inner product is always taken to be
\ +IIBII)\\xll
n
Consequently we have
for imy CJ. • . Let a = - (y, x)j (y, y); then (2-95) becomes
CONCLUDlNG REMARKS 61
2-5 Let lR(s) denote the set of all rational functions with real coefficients. Show that 2-16 What are the ranl
(IR (s), lR(s)) and (IR (s), IR) are linear spaces.
2-6 Which of the following sets of vectors are linearly independent? A, ~[: i
. Hl [:!l HJ ;n (R'. R)
2-17 Find the bases oftt
2-16.
b. [21++31iJ. , [104 -+2iJ1. ' [-iJ3 in (C 2 , IR) 2-18 Are the matrices
c. e- l , te -1, e - 21 in (a¡¡, IR), where a¡¡ denotes the set of all piecewise continuous functions
defined on [0, (0).
d. 3s 2 +s-10, -2s+3,s-5 in(1R 3 [s],IR) nonsingular in the field o
the matrices become nur
3s 2 -124s 5 +s 3 -2s-1 1
e. , - - - - - - , ---::---- in (IR (s), IR) matrices nonsingular in t
3
2s +4s - 1 S2 +s-l
2-12 Consider Table 2-1. Suppose the representations of b,e¡,el,e l , and el with
respect to the basis {el' el} are known, use Equation (2-20) to derive the representations of
b, e¡, el, el' and el with respect to the basis {el' el}.
hich of the following sets
2-13 Show that similar matrices have the same characteristic polynomial, and conse
quently, the same set of eigenvalues. [H int: det (AB) = det A det B.J
2-14 Find the P matrix in Example 3, Section 2-4, and verify A = PAP -1 .
nts
2-15 Given
such that the set {O, 1, 2}
what are the representations of A with respect to the basis {b, Ab, Alb, A3 b} and the basis
{b, Ab, Alb, A3 b}, respectively? (Note that the representations are the same!)
2-16 What are the ranks and nullities of the fol\owing matrices?
coefficients. Show that
lent?
Al = [~ ~ =~]
130
Al =[~ ~~]
001
A3 =[~ ~ ~;
34500
;]
2-17 Find the bases ofthe range spacesand the null spaces ofthe matrices given in Problem
2-16.
2-19 Does there exist a solution for the fol\owing linear equations"!
Are the sets in Problem
x(n)-A"x(O)=[b Ah ...
u(n
An-lbJ u(n:-2) .
-1)1
el' el, ... , e", what is the [
u(O)
64 LINEAR SPACES AND LINEAR OPERATORS
4
2
O -2
1
O
-4
of the matrix with multiI
following k vectors,
4 4
-150 20 16
200 160 -25 -20
¡]
_J.
2
2
-"-
2
-2
-2
O
O
O
O
O
O
1
O
O
where
A, ~r!
1 O
!l
O 1 2-28 Show that the ma
O O
that its inverse is given b
-4 -3
2-22 Let A¡ for i = 1,2, ... ,n be the eigenvalues of an n x n matrix A. Show that
detA= TI }'i
j=l
2-23 Prove that a square matrix is nonsingular if and only if there is no zero eigenvalue. 2-29 Show that the dete
is eq ua] to TI (J'r-q·
1 s.i<jS."
O 1
2-26 Consider the matrix O O
1 O
1
O O
tJ
O
-al1"1 -a 11 (nl-l)
O O
O O
O O
Show that the characteristic polynomial of A is
6(A) ~ det(AI-A)=A n +CX1An-l +CX2An-2 + ... +CXn-1A +cx n O O
nomial 6(A), lt is said to be in the Frobenius form in the numericaL analysis literature.]
20 See Reference SI7.
PROBLEMS 65
lowing matrices: 19
2-27 Consider the matrix shown in Problem 2-26. Suppose that }'l is an eigenvalue
ofthe matrix with multiplicity k; that is, ~(A) contains (le _XI)k as a factor. Verify that the
I~]
O O O
1 O O
-20
(n-1) X'¡-k
~ ~]
'.k-1
1 1
O 1 where (
11- 1) ~ (11 - 1)(11- 2)' .. (n - i)
i ;?: 1
i 1·2·3···i
O O
are generalized eigenvectors of A associated with XI'
2-28 Show that the matrix A in Problem 2-26 is nonsingular if and only ir a" f O. Veriry
that its inverse is given by
_a"i l/a"
:ix A. Show that A- I = O
[
O O
¡:re is no zero eigenvalue. 2-29 Show that the determinant of the m x m matrix
Sk m -1 O O O
assumed to be a square O Skm-l -1 O O
O O Skm-Z O O
O O O Sk2 -1
~m(s) ~m-l(S) ~m - zls) ~2(S) SkI +~I(S)
is equal to
Sn +~I(S)S"-k, +~2(S)sn-kl-k, + ... +~",(s)
O O O O O O O O
]
O O O O O O O
-a 122 -a121
---------------------------------~----------------~---
, -------------
O O O O O 1 O O
O O O O O O O O
O O O O O O O
a"_I A +a n
-a21n! -a21(",-I) -a212 -a2!! : - a22n 2 -a22(", -1) -a222 -a221
[( AI .. , A~-IJ' is an
Janion matrix of the poly 19 See Reference 6.
:rical analysis literature.J 20 See Reference S17.
66 LINEAR SPACES AND LINEAR OPERATORS
1]
}'I
O 11 O 11 O 11 O 11
O O O O O O O O (sI-Ar l
What are the multiplicities and indices? What are their geometric multiplicities?
where D.(s) ~ det (sI -A)
stant matrices. This defir
2-32 Show that if 1 is an eigenvalue of A with eigenvector x, then f(l) is an eigenvalue
most /1-1. Verify that
of f(A) with the same eigenvector x.
tr ARo
al =- -
2-33 Repeat the problems in Examples 2 and 3 of Section 2-7 by choosing, respectively,
1
az=- -
2
2-34 Given
tr AR z
a, = - ~~-
- 3
tr AR"_I
2-35 Compute eA! for the matrices (X'l=- ~--
n
[~ O O
~ 1~l~J [~O -15~
200
-12~1
160
where tr stands for the tr
matrix. This procedure (
The right-hand-side equa
by using Definition 2-16 and by using the Jordan-form representation. Rn-zs+R,,_d. Forade
2-36 Show that functions of the same matrix commute; that is, 2-40 Prove the Cayley
f(A)g(A) = g(A)f(A) R"_I' R,,-z, ... , from 0=,
2-42 Let
B
Find a matrix B such that e = C. Show that if 1, = O for sorne i then the matrix B does
-, - .. ~. - _.. - _.
-------------- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - .------------
PROBLEMS 67
c=
A.
O }, O
1 O]
[
¡ni O O A.
Find a matrix B such that eH = C. [Hint: Let I(J.) = log}, and use (2-69).] Is it true that
fonn (see Problem 2-26);
for any nonsingular matrix C, there exists a matrix B such that f!l =C?
·ow.
2-38 Let A be an n x n matrix. Show by using the Cayley-Hamilton theorem that any
'nomials of the following
Ak with k ~ n can be written as a linear combination of {I, A, ... , A" -I}. Ir the degree of
the minimal polynomial of A is known, what modification can you make?
2-39 Define
Tic multiplicities? where ~(s) ~ det(sl-A) ~ s" +als"-I +a2s"-2 + ... +anand Ro, R I, ... , R n- I are con
stant matrices. This definition is valid because the degree in s ofthe adjoint of (sI - A) is at
len j(A) is an eigenvalue most n - 1. Verify that
tr ARo
a l =- - - Ro=I
JY choosing, respectively, 1
.-2). tr AR I
a2=- - 2
tr AR 2
a,=- - -
. 3
......................................................................................
trAR"_2
ct.'l-l ==
n-l
tr AR"_J
(/.,,=--- O=AR n_ 1 +anl
n
where tr stands for the trace and is defined a~ the sum of all the diagonal elements of a
matrix. This procedure of computing ai and R¡ is called the Leverrier algorithm. [H int:
The right-hand-side equations can be verH-ied from L\(s)Jl = (sI - A)(iRos"·· J + iR IS" - 2 + ... +
:ation. R,,_ 2S + R,,_ d. For a derivation of the left-hand-side equations, see Reference SI85.]
2-40 Prove the Cayley-Hamilton theorem. (Hint: Use Problem 2-39 and eliminate
R n- I, R n- 2, ... , from 0= AR n_ 1 +anI.)
and let m(s) be the monic greatest common divisor of all elements of Adjoint (sI - A). q ualifies as an inner prol
Show that the minimal polynomial of A is equal to /'.i(s)/m(s). space. What is the forrr
2-43 Let all eigenvalues of A be distinct and let q¡ be a (right) eigenvector of A associated 2-50 Show that an n >
with A¡, that is, Aq¡ = A¡q¡. Define Q ~ [q1 q2 ... q"J and define has eigenvalue O with m
called a nilpotent matrix.
p~ ~ ~~ Q'
2-51 Let A beanm xn
iorms a linear space, can
where p¡ is the ith row of P. Show that p¡ is a left eigenvector of A associated with A¡,
that is, p¡A = A¡p¡.
2-44 Show that if all eigenvalues of A are distinct, then (sI - A)-l can be expressed as
1
(sI-Ar 1 =¿-q¡p¡
s -A¡
where q¡ and p¡ are right and left eigenvectors of A associated with A¡.
2-49 Show that the· set of all piecewise continuous complex-valued functions defined
over [O, co) forms·a linear space over C. Show that
nents of Adjoint (sI - A). q ualifies as an inner product of the space, where 9 and h are two arbitrary functions of the
space. What is the form of the Schwarz inequality in this space?
1genvector of A associated 2-50 Show that an n x n matrix A has the property A k = O for k> m if and only if A
iefine has eigenvalue O with multiplicity n and index m. A matrix with the property A k =0 is
caBed a ni/potent matrix. [Hint: Use Equation (2-64) and Jordan canonical fonu.]
¡th Á¡.
r is in a Jardan canonical
3 can be reduced to
of Systems
if the system is relaxe
input and output of
external description (
cepts of time invarim
of state is introduced
3-1 Introduction
that relates the input
The very first step in the analytical study of a systerri is tó set up mathematical set of two equations
equations to describe the system. Because of different analytical methods used, used to describe a s:
or because of different questions asked, we may often set up different mathe state-variable descript
matical equations to describe the same system. For example, in network to illustrate the proce
analysis, if we are interested in only the terminal properties, we may use the Comparisons betweer
'impedance or transfer fundion to describe the network; if we want to know description are given.
the current and voltage of each branch of the network, then loop analysis or descriptions of paral
node analysis has to be used to find a set of differential equations to describe Finally, the discrete-t
the network. The transfer function that describes only the terminal property in the last section.
of a system may be called the external or input-output description of the system, We are concernec
The set of differential equations that describes the internal as well as the terminal physical systems; hel
behavior of a system may be called the internal or state-variable description assumed to be real-va
of the system. variable or a multivaJ
In this chapter we shall introduce the input-output description and the
state variable description of systems from a very general setting. They will be Definition 3-1
developed from the concepts of linearity, relaxedness, time invariance, and
causality. Therefore they will be applicable to any system, be it an electrical, A system is said to be.
a mechanical, or a chemical system, provided the system has the aforementioned terminal and only om
properties. system if and only if
The class of systems studied in this book is assumed to have someinput output terminal.
terminals and output terminals. The inputs, or the causes, or the excitatioris u
are applied at the input terriúnals; the outputs, or the effects, or the responses y The references for
are measurable at theoutput terminals. In Section 3-2 we show that if the and116. The main ob
of linearity, causality,
70
._-_... _ ... .. _.. -_._.- _.... -
_-_.~
.---~-------_._-.- ,
lNTRODUCTlON 71
input u and the output y of a system satisfy the linearity property, then they can
be related by an equation of the form
If the input and output have, in addition, the causality property, then (3-1a)
can be reduced to
that relates the input u, the output y, and the state x is then .introduced. The
o set up mathematical set of two equations of the form (3-2) is called a dynamical equation. If it is
alytical methods used, used to describe a system, it is caBed the dynamical-equation description or
et up different mathe state-variable description ofthe system. We give in Section 3-4 many examples
example, in network to illustrate the procedure of setting up these two mathematical descriptions.
~rties, we may use the Comparisons between the input-output description and the dynamical-equation
:; if we want to know description are given in Section 3-5. We study in Section 3-6 the mathematical
then loop analysis or descriptions of parallel, tandem, and feedback connections of two systems.
equations todescribe FinaBy, the discrete-time versions of Equations (3-1) and (3-2) are introduced
the terminal property in the last section.
,cription of the system. We are concerned with descriptions of systems that are models of actual
¡as well as the terminal physical systems; hence all the variables and functions in this chapter are
te-variable description assumed to be real-valued. Before proceeding, we classify a system as a single
variable or a multivariable system according to the following definition.
t description and the
setting. They will be Definition 3-1
time invariance,. and
A system is said to be a single-variable system if and only if it has only one input
em, be it an electrical,
. terminal and only one output terminal. A system is said to be a multivariable
Las the aforementioned
system if and only if it has more than one input terminal or more than one
output terminal. •
d to have sorne input
es, or the excitations u
The ·references for this chapterare 24, 27, 31, 53, 60, 68, 70, 73, 92, 97, 109,
importance in developing the linear equations. They are not introduced very will be excited solely al
rigorously. For a more rigorous exposition, see References 60, 68, 109, and tion, it is Iegitimate to
116.
:e not introduced very will be excited solely and uniquely by u. Hence, under the relaxedness assump
ences 60, 68, 109, and tion, it is legitimate to write
y=Hu (3-3)
I
1
/i
As shown in Figure 3-.
by a series of pulse fUJ
u(t¡)(jt,(t - tJ~, we can
is called the impulse function or the Dirac delta function or simply (jlunction.
Thus the delta function (j(t - t 1) has the properties that
oo 111 +E 3(t -
J -ce
a(t - t 1) dt =
ti-E;
tí) dt = i 2 Th;,::; condiUon fer inlerchai
cussion of this problem, see
for any positive c and that
U(¡¡)O¡:,(I-I¡).
(3-7) /
for any function f that is continuous at t l' T~
u (1) I
1
I
I
I 11
I 11
With the concept of impulse functión, we are ready to develop a mathe 1
I
1 I I 11
=U(I,)/i0¡:' 1 I I~.
rilatical description for relaxed linear systems. We discuss first single-variable I 1 I :1
systems; the result can then be easily extended to multivariable systems. Con
J I
I
I
:
I
:
I
: I
1 For a rigorous development of the subsequent material, the theory of distributions is needed;
see Reference 96. Figure 3-3 Pulse-func
~._------_.-_.,-_._._-,_ .. __ .- _. -"'~'- _-
.. __ ._ ..•. .. _-_ ... "---'- _
.. _•.. .•..._. __ ._---_. __.. __ ..... _._- .-
sider a relaxed single-variable system whose input and output are related by
y=Hu
As shown in Figure 3-3, every piecewise continuous input can be approximated
by a series of pulse functions. Since every pulse function can be described by
u(tJb,,(t - tJ~, we can write the input function as
u =i= ¿ u(t¡}b,,(t - t¡}~
i
If the input-output pairs of the relaxed system satisfy the linearity property,
then we have 2
r ex (see Problem 3-9). (3-8)
; desired by a rational
ty that U n - u implies Now as ~ tends to zero, the approximation tends to an exact equality, the
perty ofhomogeneity. summation becomes an integration and the pulse function b,,(t - t¡} tends
escription for a linear toward a b-function. Consequentiy, as ~-+O, Equation (3-8) becomes
: of the delta function
a detailed exposition y= J~oo (Hb(t -'l:))u('l:)d'l: (3-9)
pulse function defined
Now if Hb(t - r) is known for aH r, then for any input, the output can be com
puted from (3-9). The physical meaning of Hb(t - r) is that it is the output of
the relaxed system due to an impulse function input applied at time r. Define
Hb(t - r) = g(', r) (3-10)
2 The condilion lor inlerchanging the order of H and lhe Slil11mation is disregarded. For a dis
cussion of this problem, see Reference S125, pp. 2-6.
(3-7) /
T~u (ti)
I
I
I
I
I 1I
I 1 I
to develop a mathe~ 1
= u (t) -.{',. I
I
I
I
I 11
I
.ss first single-variable ,{',. I I \
I I I
ariable systems. Con- I I I
of distributions is needed;
Figure 3-3 Pulse-furiction approximation of an input function.
76 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
Hence if g(', r) for aH r is known then for any input u, the output can be com Consequently, the in
puted fram (3-11). In other words, a linear relaxed system is completely becomes
described by the superposition integral (3-11), where g(', r) is the impulse
response of the system, and theoreticaHy it can be obtained by direct measure
ments at the input and the output terminals of the system.
Relaxedness. Rec
If a system has p input terminals and q output terminals, and if the system
is initially relaxed at - 00, the input-output description (3-11) can be extended to the system is--relaxed
solely and uniquely b
y(t) = roo G(t, r)u(r) dr (3-12)
Definition 3-3
he output can be com Consequently, the input-output description of a linear, causal, relaxed system
system is completely becomes
g(', -r) is the impulse
y(t) = Loo G(t, ,)u(,) d, (3-15 )
Iled by direct measure
:m.
Relaxedness. RecaH that the equation y = Hu of a system holds only when
,nals, and if the system
the system is-relaxed at - 00, or equivalently, only when the output y is excited
-11) can be extended to
solely and uniquely by u(-oo,oo)' We may apply this concept to an arbitrary to.
(3-12 )
Definition 3-3
A system is said to be relaxed at time t o if and only if the output is solely
~: :~j
Y[to,oo)
and uniquely excited by u[to,oo)' •
outputs are linearly related, this can be determined without knowing the Proof
previous history of the system, by the use of the foHowing theorem.
If u[to,oo) == O, theout¡:
Theorem 3-1
lo
r
y(t) = Loo 4
A system that is describable by Since
N ecessity: If a system is relaxed at to, the output y(t) for t ? t o is given by This is an import¡
Corollary 3-1, its rela
1 00
lo
G(t, ,)u(,) d, over any nonzero intl
interval, then the sys
Hence, if U[lo,oo) = o, then Y[lo, 00) = O. Sufficiency: We show that if U[lo,oo) =O next chapter that the
implies Y['o,oo) =0, then the system is relaxed at to. Since function matrices 01
satisfies the condition
y(t) = roo G(t, ,)u(,) d, = roo G(t, ,)u(,) d, + 1~ G(t, ,)u(,) d, able.
We give an examp
= =
the assumptions U[lo. 00) O, Y[lo,oo) O imply that whose inputs and out
r
lo
J-oo
G(t, ,)u(,)d, =0 for aH t ? to
Example 3
Consider the system
In words, the net effect ofu(-oo,lo) on the output y(t) for t ?t o is zero, and hence the output is the volt.
the system is relaxed at to. Q.E. D. stored in the capacitor
zero if no voltage is a¡:
An implication of Theorem 3-1 is that given a system at time to, if the system relaxed at to, because
is known to be describable by depending on which i
roo
Loo G(t,,)u(,)dT
the relaxedness of the system can be determined from the behavior of the 1--------
I R = I 11
system after t o without knowing the previous history of the system. Certainly
I
it isimpractical or impossible to observe the output from time to to infinity; I
I + +
fortunately, for a large class of systems, it is not necessary to do so. I
I
e
I
I
Corollary 3-1 IL _
Ir the impulse-response matrix G(t, ,) of a systerncan be decomposed into
G(t, ,) = M(t)N(,), and ir every element of M is analytic (see Appendix B) .on (a)
( - 00, (0), then ihe system is relaxed at t o if and only if for sorne fixed positive Figure 3-4 A nonline¡
e, U[lo,lo+<) = O implies Y[lo,lo+<) = O. linear capacitor C.
._..~... ._-_.._.._._--_.__..__.."--_ .....__..... _._--:~:::::~::-. . ::::..-==~~-:::::=:.=.=-:~-=.:-~=-...:.:=.:.-=:..::..:..====::.~---------;------:----------
- • __ .~ •• _,_. _ _. , _ . _ •• ' _ 0 • • . '•••
I
i
Since
roo N(t)u(t) dt
t 2: t o is given by This is an important result. For any system that satisfies the conditions of
Corollary 3-1, its relaxedness can be easily determined by observing the output
over any nonzero interval of time, say 10 seconds. Ir the output is zero in this
interval, then the system is relaxed at that moment. It will be shown in the
how that if U[lo, 00) == O next chapter that the class of systems that are describable by rational transfer
function matrices or linear time-invariant ordinary differential equations
satisfies the conditions of Corollary 3-1. Hence Corollary 3-1 is widely applic
G(t, t)u(t) dt able.
We give an example to illustrate that Theorem 3-1 does not hold for systems
whose inputs and outputs are not linearly related.
Example 3
Consider the system shown in Figure 3-4; the input is a voltage source, and
2:t o is zero, and hence the output is the voltage across the nonlinear capacitor. Ir the electric charge
Q.E.D. stored in the capacitor at time t ois either O, q b or qz, the output wil\ be identically
zero if no voltage is applied at the input. However, the system is not necessarily
at time to, if the system relaxed at to, because if an input is applied, we may obtain different outputs
depending on which initial charge the capacitor has. •
.\ ~}
1-------------1
. be decomposed into
:; (see Appendix B) on (a) (b)
'or some fixed positive Figure 3-4 A nonlinear network. (a) The network. (b) The characteristic o(the non
linear capacitor C.
80 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
Time invariance. Ifthe characteristics ofa system do not change with time, this to the multivaria
then the system is said to be time invariant, fixed, or stationary. In order to
define it precisely, we need the concept of a shifting operator Q". The effect of
the shifting operator Q" is illustrated in Figure 3-5. The output of Q" is equal for all t and -r. Heno
to the input delayed by \1. seconds. Mathematically it is defined as ii ~ Q"u if
and only if ii(t) = u(t - \1.) or ii(t +\1.) = u(t) for all t.
The relation HQ"u = Q"Hu can also be written as HQ"u = Q"y, which implies y(t)
that if an input is shifted by \1. seconds, the waveform of the output remains the
same except for a shift by \1. seconds. In other words, no matter at what time an The second equality (
input is applied to a relaxed time-invariant system, the waveform of the output The integration in (3
is always the same. G(t, -r) represents the
If a relaxed linear system is known to be time invariant, what condition time -r. G(t) represent~
does that impose on its impulse response? The impulse response g(', -r) is the at -r =0. Following (
output due to a J-function input applied at time -r; that is, g(', -r) = HJ(t - -r). if G(t) = O for all t < O
Ifthe system is time invariant, then we have
Transfer-function
Q~(', -r) = Q"HJ(t - -r) = HQ"J(t --r) describable by convol
= HJ(t - (-r +\1.)) = g(', -r +\1.) transform, because it \
Now by the definition of Q", the equation Q~(', -r) = g(', -r +\1.) implies g(t, -r) = an algebraic equation :
g(t +\1., -r +\1.), which holds for any t, -r, and \1.. By choosing \1. = - -r, we have of y; that iS,4
g(t, -r) = g(t - -r, O) for all t, -r. Hence the impulse response g(t, -r) of a relaxed,
linear, time-invariant system depends only on the difference of t and -r. Extending
y(s)
..__ •
_ _ .. _ - ._,_
._. _ •• - . - -_ _ . _ . • ~ ~
latter at what time an The second equality of (3-20) can be easily verified by changing the variables.
tveform of the output The integration in (3-20) is called the convolution integral. Since G(t - ,) =
G(t, ,) represents the responses at time t due to Ci-function inputs applied at
'iant, what condition time " G(t) represents the responses at time t due to Ci-function inputs applied
response g(', ,) is the at, =0. Following (3-14), a linear time-invariant system is causal if and only
is, g(·,,)=HCi(t-,). if G(l) =0 for all l < O.
3 Note thatG(t,"t) and G(t -"t) are two <;lirrerent runctions, However. ror convenience, the same
symbol G is used. -
4 Ir y contilins delta runctions at t =0, the lower ¡imil or the integration should start rrom O-to
inelude the delta runctions in the transrorm.
82 MATHEMATlCAL DESCRlPTIONS OF SYSTEMS
Here we have changed the order of integration, changed the variables, and used rational functions al
the fact that G(l) = Ofor l < O. As defined in (3-21), C(s) is the Laplace transform proper rational funcl
of the impulse-response matrix; that is,
A rro
G(s) = Jo G(t)e - SI dt 3-3 The State·
and is called the transfer-function matrix of the system. For single-variable The concept of 51
systems, C(s) reduces to a scalar and is cal1ed the transfer function. Hence the able only when the S)
transfer function is the Laplace transform of the impulse response; it can also be say at time t o, then tb
defined, following (3-21), as output y [lo, ro) depend
ditions at too Henc
.. ;t'[y(t)] I y(s) I (3-22) addition to the input
g( s) = ;t'[ ( )] the system is = -;;;--() relaxed
U (relaxedatl=O U S atl=O initial conditions is ca
together with the in
where the circumf[ex (.. ) over a variable denotes the Laplace transform of the
example, in newtoni
salIle variable; for example,
a particle (system) al
not uniquely determi
known. How the p~
immaterial in determ
We see that the familiar transfer functions are the input-output descriptions two numbers, the po~
of systems. It is important to note that this input-output description is obtained state of the system at t
under the relaxedness assumption of a system; hence, if the system is not relaxed the momentum at lo,
at t = O, the transfer function cannot be directly applied. Thus whenever a
transfer function is used, the system is always implicitly assumed to be relaxed Definition 3-6
at t =0.
A transfer function is not necessarily a r¡ltional function of s. For example, The state of a system ;
the impulse response g(t) ofthe unit-time-delay system introduced in Example 2 with U[IO, ro)' determinl
is (j(t -1), and its transfer function is e-s, which is not a rational function of s.
However, the transfer functions we shall study in this book are exclusively By the behavior o
rational functions ofs. In fact, we study only a special class ofrational functions. the system. If the sysl
branch of the networ
be computed. If the
Definition ::S-5
U[IO'''')' in computing 1
A. rational function g(s) is said to be proper if g(oo) is a finite (zero al" nonzero) [1 summarizes the ess
constant. [¡(s) is said to be strictly proper if [¡( (0) = O.. A rational matrix C(s) needed in determining
is said to be proper if C(oo) is a finite (zero or nonzero) constant matrix. C(s) input ui- 00,11)' i = 1, 2,
is said to be strictly proper if C( (0) = O. • though ui - 00,11)' i = 1,
the system are identic;
We give sorne exal
For example, [¡(s) = s2/(s -1) is not proper; [¡(s) = s2/(S2 - s + 2) is proper
and [¡(s) = s2/(S3 - s) is strictly proper. It is clear that if [¡(s) = N(s)/D(s), [¡(s) is
Example 1
proper if and only if degN(s).:s;degD(s); [¡(s) is strictly proper if and only if
deg N(s) < deg D(s), where deg stands for the degree of a polynomial. A rational Consider the network
matrix is proper if andortly if al1 of its elements are proper. A rational matrix current through the i
.is strictly proper if and only if all of its elements are strictly proper.
Ifatransfer function is not proper, high-frequency noises will be greatly 5 The exceptions are transl
amplified and will overwhelm information-bearing signals. Hence improper metelS. See Reference 54
-_.. __ ._ .. - -_._---- .. "---._.... _-_._~--,---,~."._----- - _.-._-~_.~_.-~-----_.-
" ..._.. _---,. __ .. ~---_ .. _.- ._--_.. ,._- .. _----_._~---~--_._._- --_... ~,._-- .. _- -_ ... _._._----~._. __ - _. __ ._._ --_._--------.
.. _. __ _----. -----
.._.'--'-" ... __ .. ..
he variables, and used rational functions are hardly used in practice. 5 In this book, we study only
the Laplace transform proper rational functions and proper rational matrices.
VOltag:
source
Figure 3-6
E3i
~ ~
A network.
Y
l
~_-_
output after t ~ t o ex(
minable. From (3-24
y(t) = -2e
known, then for any driving voltage the behavior of the network can be deter which, together with f
mined uniquely. Benee the inductor current and the capacitor voltage qualify
as the state of the network. •
Solving for Cl and C2 I
Example 2
We consider again the network in Example 1. The transfer function from
u to y of the network can be easily found as
Bence if the network
A 222
g(5)=(5+1)(5+2)=5+1- 5+2 y(t) = (2
r g(t-L)U(L)dL
l
(3-25)
fx:
r----'\AII~------JE
R
for t ~ to, where
----=t
XI
Note that Cl and C2 are independent of t. Bence if C l and C2 are known, the Figure 3-7 A netwo
••. - • • • • ~ .. _. __ o • • _
output after t"2: t o excited by the unknown input u{ _ 00.10) is completely deter
minable. From (3-24) and (3-25) we have
y(t o)=2e- IO c l -2e- Ztoc 2 (3-26)
r g(t -"l:)u("l:)d"l:
t
+ for t "2:t o
(3-23)
t
We see that if y(to) and y(to) are known, the output after t "2:t o can be uniquely
nction 9(S). Now we determined even if the network is not relaxed at too Hence the set of numbers
laxed at to, the output y(to) and y(to) qualifies as the state of the network at too Clearly, the set {Cl' C2}
also qualifies as the state of the network. •
In this example, we see that the effect of the input over the infinite interval
(- co, t o) is summarized into two numbers {y(to), y(to)} or {Cl> C2}; hence the
omputed froro concept of state is very efficient and powerful.
'10
6 d
d). r g(t-T)u(r)dT=g(t-T)U(T) I
t
+ J' -g(t-T)U(T)dT
a
t lo t=t roat
(3-25 )
Lnd C2 are known, the Figure 3 c 7 A network with a loop that consists of capacitors only.
• _. _ •• _ _ .• ' •• _ •••• _ •• •.•••• •• •• _ _ _... • •• _ _ •••• _ .•• •• _._. _ _ ••• •••..• _._ ••• _ . _ . _ _. _ • • • • •.•• _ . ••• ._0 __ - '0_" • __ ._._•. , __ .... • "'._. __ •. • __ •• ... • . __ . • •••• _
~
_
.
~
~
~
.
_
~
_
~
~
_
,
~
86 MATHEMATICAL DESCR1PT10NS OF SYSTEMS
apply the Kirchhoff voltage law to the loop that consists of three capacitors, dynamical equation.
we have Xl(t) +xz(t) +X3(t)=Ü for all t. lt implies that if any two of Xl,X Z, the form
and X 3 are known, the third one is also known. Consequently, if any two of the
capacitor voltages are known, the behavior of the network is uniquely deter X(l
minable for any input applied thereafter. In other words, any two of the three y(t
capacitor voltages qualify as the state. If all the three capacitor voltages are
chosen as the state, then there is a redundancy. In choosing the state of a or, more explicitly,
system, it is desirable to choose a state that consists of the least number of xl(t) =J
variables. How to pick the state with the least number of variables for general xz(t) = J
RLC networks will be studied in the next section. •
xnCt)=J
Example 4
Yl(t) =6
A unit-time-delay system is a device whose output y(t) is equal to u(t -1) for
yzCt) = 6
all t. For this system, in order to determine Y[lo,CO) uniquely from U[lo,CO)' we
need the information U[lo-l,lo)' Hence the information U[lo-l,lo) is qualified to
yq(t) =9
be called the state of the system at time too I
where x = [Xl Xz ..
From these examples, we may have the following observations concerning and u = [Ul U z
the state of a system. First, the choice of the state is not unique. For the state x are real-valued
network shown in Figure 3-6, the state may be chosen as the inductor current for (3-28) to qualify a
and the capacitor voltage, or chosen as y(to) and y(to) or Cl and Cz. For the initial state x(to) and :
network shown in Figure 3-7, any two of the three capacitor voltages can be sufficient condition for
chosen as the state. Different analyses often lead to different choices of state. initial state is that h¡ an
Second, the state chosen in Example 1 is associated with physical quantities, see References 24,77, al
whereas in Example 2 the state is introduced for mathematical necessity. that the solution can 1
Hence, the state of a system is an auxiliary quantity that may or may not be the state at to, as expe<
easily interpretable in physical terms. Finally the state at each instant may and is caBed a state eq
consist of only a finite set of numbers, as in Examples 1,2, and 3, or consist of an an output equation. N(
infinite set of numbers, as in Example 4. Note that there is an infinite number of equation in the form 01
points between [to -1, to); hence the state of this example consists of an infinite ledge ofx(t) and u(t) su
set of numbers. The state space of
In this book we study only the class of systems whose states may be chosen set of equation (3-28) ¡¡
to consist of a finite number of variables. The state of a system can then be
represented by a finite-dimensional column vector x, called the state vector. linearity. We use tI:
The components of x are called state variables. The linear space in which the
state vector ranges is denoted by L. Since state variables are usually real
valued, and since we study only systems with a finite number of state variables, to denote that the statt:
the state spaces we encounter in this book are the familiar finite-dimensional state x(t), for t '2: to, al
real vector space(IW, IR). output pair of a system
Dynamical equations. In addition to the input and output of a system we Definition 3-7
have now the state of the system. The state at time t o is, by definition, the
required information at t o that, togetherwith input U[lo,CO)' determines uniquely A system is said to bell .
the behavior (output and state) ofthe system for all t '2: too The set 01 equations
that describes the unique relations between the input, output, and state is caUed a
----,._._---~ _.
... __.__ .. _-_._.-._--- -- ._ .•.. __ ...._---- ---- _._'-.-'."- ----.--_.•.•.. -.- ....._-_.-. __.-.-_ .. ~.,-_ ..-----_.- --.. ,-----_ .. _-_.-._
.ts of three capacitors, dynamical equation. In this book, we study only the dynamical equations of
t if any two of Xl> Xz, the form
lentiy, ifany two ofthe
i(t) = h(x(t), u(t), t) (state equation) (3-28a)
ork is uniquely deter
s, any two of the three y(t) = g(x(t), u(t), t) (output equation) (3-28b)
capacitor voltages are
or, more explicitly,
100sing the state of a
,f the least number of x¡(t) =h¡(x¡(t), xz(t), ... , xn(t), u¡(t), uz(t),.·., up(t), t)
)f variables for general xz(t)=hz(x¡(t), xz(t),···, xn(t), u¡(t), U2(t),···, up(t), t)
I (3-29a)
Xn(t) = hn(x¡ (t), xz(t), ... , xn(t), u¡(t), uz(t), ... , up(t), t)
and any real numbers al and a2 , the following pair The responses due to {x
2 exclusively by the non¡
{alxl(tO) +a 2x (t o), alutto,ro) +a 2 u fro,ro)}
are caBed zero-state re~
----> {alxtto,ro) +a 2x fro,ro), alytto,ro) +a 2 Yfto,ro)} (3-30)
Equation (3-31) follows
is also admissible. Otherwise, the system is said to be nonlinear. • xl(to) =x(t o
Similar to Equations (3-4) to (3-6), if al = a 2 = 1, the relationship in (3-30) is
Hence for linear syster
caBed the property of additivity; if a 2 = 0, it is caBed the property of homogeneity.
zero-state responses in
The combination of these two properties is caBed the principie of superposition.
in the previous section I
In this definition, the superposition property must hold not only at the output
If a system is linear,
but also at aB state variables; it must hold for zero initial state as weB as non
zero initial state. Hence this definition is much more stringent than the one in h(x(t), u(t), t) = j.
Definition 3-2. Consequently, a system may not be linear according to Defini
where A, B, e, and E :
tion 3-7 but may be linear according to Definition 3-2. For example, the system
(Problem 3-36). Heno
in Figure 3-S (a) has a nonlinear capacitor C. Ifthe voltage across the capacitor
form
is zero at t o = 0, it will remain to be zero for aB t;::: t o no matter what input
waveform is applied. Hence, as far as the behavior at the input and output E: x(
terminals is concerned, the nonlinear capacitor can be disregarded. Hence the
yt
system is linear according to Definition 3-2 but not linear according to Definition
3-7. Consider the network in Figure 3-S(b) with a nonlinear capacitor C and a A sufficient condition I
nonlinear inductor L. Because the L-C loop is in a series connection with the of A(') be a continuous
current source, its behavior will not transmit to the output y. Hence the system the entries of B(' ), C(' ), ~
in Figure 3-S(b) is linear according to Definition 3-2 but not linear according to Since the values of A(
Definition 3-7. equation E in (3-32) is 1
equation.
We discuss now the implication of Definition 3-7. If al =a 2 = 1 and if
xl(t O)= -x 2 (t O) and utto,ro) = -ufro,ro) Time invariance. 1
then the system is said
then the linearity implies that {O,O}---->{O[to,ro),O[to,ro)}' Hence a necessary
°
condition for a system to be linear is that if x(to) = and u[to, ro) == 0, then the
responses of the system are identically zero. A very important property of any
lowing. Let Q" be the
Definition 3-8
linear system is that the responses of the system can be decomposed into two
parts, as A system is said to be t
Responses due to {x(t o), IDI[to, ro)}
= responses due to {x(t o), O} +responses due to {O, u[to, ro)} (3-31 )
and any real number a,
e {Q,,~
u In
is also admissible. Ott
Ll t
L .
.2 H 3F
T Y
In words, for time-i
the waveforms of excita
will always be the same
~) ~)
1 For linear time-invariaJ
are independent of time
Figure 3-8 Two nonlinear systems which are linear according to Definition 3-2.
•••• _ _ • ••• .•• _ _ ,_u__ , ,_"" ,_ _ ,,
----,---_..- ..•- .. __
.
~
~
~
_
~
.
_
~
~
,
~
.
_
,
~
_-----~_. ._~------_._---_._--_.~--~.,~---~-----.-
The responses due to {x(to), O} are caBed zero-input responses; they are generated
exclusively by the nonzero initial state x(to). The responses due to {O, u¡to.co)}
are called zero-state responses; they are excited exclusively by the input u[to.oo)'
x2 yfro. col} (3-30)
Equation (3-31) foBows directly from (3-30) if we choose IX 1 = IX 2 = 1 and
IOnlinear. • xl(to) =x(to) u 1 =0 x 2 (t O) =0 ufro.CO) =U[lo.oo)
relationship in (3-30) is
Hence for linear systems, we may consider the zero-input responses and the
roperty of homogeneity.
zero-state responses independently. The input-output description discussed
'inciple of superposition.
in the previous section describes only the zero-state responses of linear systems.
not only at the output
If a system is linear, the h and g in (3-29) become linear functions ofx and u as
ial state as well as non
:ringent than the one in h(x(t), u(t), t) = A(t)x(t) + B(t)u(t), g(x(t), u(t), t) =C(t)x(t) + E(t)u(t)
lar according to Defini
where A, B, e, and E are, respectively, n x n, n x p, q x n, and q x p matrices
~or example, the system
(Problem 3-36). Hence an n-dimensional linear dynamical equation is of the
1ge across the capacitor
form
, no matter what input
t the input and output E: x(t) = A(t)x(t) + B(t)u(t) (state equation) (3-32a)
lisregarded. Hence the
y(t) =C(t)x(t) + E(t)u(t) (output equation) (3-32b)
according to Definition
inear capacitor e and a A sufficient condition for (3-32) to have a unique solution is that every entry
ies connection with the of A(') be a continuous function of t defined over (- 00, 00). For convenience,
ut y. Hence the system the entries of B('), C('), and E(') are also assumed to be continuous in ( - 00, 00).
not linear according to Since the values of A('), B('),C('), and E(') change with time, the dynamical
equation E in (3-32) is more suggestively caBed a linear time-varying dynamical
equation.
If IX I =0: 2 = 1 and if
- ufro. col Time invariance. Ifthe characteristics ofa system do not change with time,
then the system is said to be time invariant. We define it formally in the fol
" Hence a necessary
lowing. Let Q" be the shifting operator defined in Figure 3-5.
md U[to. 00) == O, then ihe
í'ortant property of any
Definition 3-8
~ decomposed into two
A system is said to be time invariant if and only if for any admissible pair
(3-31 )
and any real number IX, the pair
{Q"x(t o), Q"u[to. col} -+ {Q"x¡to. co)' Q"y[co. co)}
u In
is also admissible. Otherwise, the system is said to be time varying. •
2H
T
1 In words, for time-invariant systems, if the initial states are the same and
the waveforms of excitations are the same, then the waveform of the responseS
will always be the same no matter at what instant the excitations are applied.
-------
3F
T For linear time-invariant systems, the matrices A('), B('), C('), and E(') in (3-32)
are independent of time, and the equation reduces to
FE: x(t) = Ax(t) +Bu(t) (3-33a)
(b)
ling lo Definilion 3-2. y(t) = Cx(t) + Eu(t) (3-33b)
90 MATHEMATlCAL DESCRIPTlONS OF SYSTEMS
where A, B, e and E are, respectively, n x n, n x p, q x n, and q x p real constant (sI - A) always exists.
matrices. This set of equations is called a linear time-invariant n-dimensional at t =O-then (3-35b)
dynamical equation and is denoted by FE (fixed equation). For linear time
invariant systems, the responses are independent of the initial time; hence it is
always assumed without loss of generality that t o =0. The time interval of A comparison of this
interest then becomes [0, ro).
The state space L of E or FE is an n-dimensional real vector space (IR", IR),
Hence we can think of the n x n matrix A as a linear operator which maps L Hence if a linear tirr.
into L. As mentioned in the preceding chapter, it is very convenient to intro (;(s) and the dynamil
°
duce the set of the orthonormal vectors {01' 2 " " , 0n}, where 0i is an n x 1 related by (3-36). We
column vector with 1 at its ith component, and zero elsewhere, as the basis of
the state space. In doing so, we may also think of the matrix A as representing a
linear operator with respect to this orthonormal basis. Hence, unless other
wise stated, the basis of the state space of E nr FE is assumed to be the set of the Every entry ofthe adj(
orthonormal vectors the degree of the deter
rational matrix. If E
rational matrix. Note
The dynamical equations in (3-29), (3-32), and (3-33) can be solved, given
x(to) and u('), in the direction of positive time or in the direction of negative
time. Clearly we are interested in only the direction of positive time. In the Analog and digité
positive-time direction, the input u(t 1) affects only the future responses, the equations. As willl
responses for t ~ t 1; it does not affect the past responses. Hence the dynamical number of state variab:
equations are aH causal. ical equations. We sh(
dyriamical equation ca .
Transfer-function matrix. In the stuoy oflinear time-invariantdynamical
connecting integrators
equations, we may also apply the Laplace transformo Taking the Laplace grator, 8 summer, and ¡
transform of FE and assuming x(O) =x o, we obtain puter, and their functio
si(s) -x o = AX(s) +BU(s) (3-34a) block diagram of analo
time-invariant dynamil
y(s) =CX(s) +Eii(s) (3-34b)
where the circumflex over a variable denotes the Laplace transform of the
same variable; for example,
They are algebraic equations. Ifx o and u are known, x(s) and y(s) can be com
puted from (3-35). Note that the determinant of (sI - A) is different from zero
(the zero ofthe field of rational functions of S)7; htmce, the inverse ofthe matrix
<al
7 That is, the determinant of (sI - A) is not identically equal to zero. Note that det (si - Al = O Figure 3-9 Analog com
for sorne s is permitted. or attenuator.
~-'------'------'''-----,-------------- ---- -------_._-_.
nd q x p real constant (si - A) always exists. If the initial state X o is O-that is, the system is relaxed
variant n-dimensional at 1 =O-then (3-35b) reduces to
)n). For linear time Y(s) = [C(sl - A)-l B + E]u(s)
nitial time; hence it is
The time interval of A comparison of this equation with (3-21) yields
G(s)=C(sI-A)-lB +E (3-36)
l vector space (IR", Iln
lerator which maps L Hence if a linear time-invariant system is described by the transfer matrix
y convenient to intro G(s) and the dynamical equation {A, B, C, E}, the two descriptions must be
, where ni is an n x 1 related by (3-36). We write (3-36) as
:where, as the basis of A 1
rix A as representing a G(s) = d (1 ) C[,Adj (sI - A)]B + E
et s -A
Hence, unless other
/'led to be the set 01 the Every entry of the adjoint of (si - A) is a polynomial of degree strictly less than
the degree ofthe determinant of(sI - A); hence C(sI -A)-lB is a strictly proper
rational matrix. If E is a nonzero matrix, then C(sl - A)-l B + E is a proper
rational matrix. Note that we have
I can be solved, given G(oo) =E (3-37)
direction of negative
positive time. In the Analog and digital computer simulations of linear dynamical
future responses, the equations. As will be illustrated in the next section, systems that have a finite
Hence the dynamical number of state variables can always be described by finite-dimensional dynam
ical equations. We show in this subsection that every finite-dimensionallinear
dynamical equation can be readily simulated on an analog computer by ihter
e-invariant dynamical connecting integrators, summers, and amplifiers (or attenuators). The inte
Taking the Laplace grator,8 summer, and amplifier are three basic components of an analog com
puter, and their functions are illustrated in Figure 3-9. We give in Figure 3-10 a
(3-34a) block diagram of analog computer connections ofthefollowing two-dimensional
time-invariant dynamical equation
(3-34b)
~::][::~~G +[:::
8 In practice, pure difTerentiators are not used for the reason that they wil\ amplify high-frequency
(3-35a) noises. On the other hand, integrators wil\ smooth or suppress noises.
tEu(s) (3-35b) Xl
Note that det (sI - A) = O Figure 3-9 Analog eomputer components. (a) Integrator. (b) Summer. (e) Amplifier
or attenuator.
92 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
U¡ Xl
'-----------~ e21I-----------J
DYNAMIC
PARAMETER
'-------------+1 e22 1--- ------'
X1DT=
Figure 3-10 Block diagram oftwo-dimensional dynamical equation.
X2DT=
X1 =IN
X2=IN
Note that for a two-dimensional dynamical equation, we need two integrators. Y=X1
The output 01 every integrator ean be assigned as a state variable. We see that TIMER DELT,
even for a two-dimensional dynamical equation, the wiring of the block diagram PRTPLT X1,Y
is complicated; hence, for the general case, we usually use a matrix block STOP
diagram. The matrix block diagram of the dynamical equation E is shown in END
Figure 3-11. If E is n-dimensional, the integration block in Figure 3-11 consists
of n integrators. The matrix E represents the direet transmission part from the The first part of the
input u to the output y. Ifthe matrix A is a zero matrix, then there is no "feed step size. "FINTIM" i
back" in the block diagram. For a discussion of analog computer simulations, interval in which the n
see, e.g., Reference S46. result, DELT is usually
Dynamical equations can be readily simulated on a digital computer. print out every comput
There are many specialized subroutines for solving dynamical equations, such larger than DELT. In
as MIDAS (Modified Integration Digital Analog Simulator), MIMIC (an DELT and OUTDEL.
improved version of MIDAS), CSMP (Continuous System M odeling Program), as well as plot the outp
TELSIM (Teletype Simulator), and others; see Reference 26. These programs Comparisons of ana
tions are in order. On
are limited to a range,
range, sorne componen
of the sim ulation will t
equations must be pro]
u y carried out by cut and t
+ of numbers which a dig
magnitude scaling genel
is often limited to 0.1 1
precision of eight or m
computer simulation is
Figure 3-11 Matrix block diagram of the dynamical equation E. putero The generation
THE STATE-VARIABLE DESCRIPTION 93
[Xl]
X2 =
[2 -lJ[XIJ +[O ]
1 5_ X2 1.5 u
y = [ 1 0.6J [ :J
Find the output y and state variable Xl from O to 20 seconds due to the initial
condition Xl (O) = 1, X2(0) = O, and a unit-step-function input.
The CSMP input statements for this problem are listed in the following:
DYNAMIC
PARAMETER U = 1.0
X1 DT=2.0*X1-X2
X2DT=X1 +5.0*X2+1.5 *U
~quation.
X1 = INTGR(1.0,X1 DT)
X2 = INTG R(0.0,X2DT)
e need two integrators. Y=X1 +0.6 ·X2
. variable. We see that TIMER DELT =0.001 ,FINTIM =20.0,OUTDEL=0.10
Ilg ofthe blockdiagram PRTPLT X1,Y
ly use a matrix block STOP
;:quation E is shown in END
in Figure 3-11 consists
nsmission part from the The first part ofthe program is self-explanatory. "DELT' is the integration
, then there is no "feed step size. "FINTIM" is the final time of computation. "OUTDEL" is the
computer simulations, interval in which the responses will be printed. In order to have an accurate
result, DELT is usually chosen to be very smal!. It is, however, unnecessary to
n a digital computer. print out every computed result; therefore the printout interval is chosen much
amical equations, such larger than DELT. In employing CSMP the user has to decide the sizes of
mulator), MIMIC (an DELT and OUTDEL. For this program we have asked the computer to print
:m Modeling Program), as well as plot the output y and the state variable X l'
~e 26. These programs
Comparisons of analog computer simulations and digital computer simula
tions are in order. On analog computer simulations, the magnitudes of signals
are limited to a range, typically ± 10 volts. Ir the magnitudes go over the
range, sorne components of the analog computer will saturate and the result
of the simulation will be erroneous. Hence on analog computer simulations,
equations must be properly scaled. This is a difficult problem and usually is
carried out by cut and try. On digital computer simulations, because the range
of numbers which a digital computer can manage is very large, the problem of
magnitude scaling generally does not arise. The accuracy of an analog computer
is often limited to 0.1 percent of its full scale; a digital computer may have a
predsion of eíght or more decimal digits. Therefore the result from a digital
computer simulation is much moreaccurate than íhat from an analog conb
ion E. puter. The generation of nonlinear functions is easier on a digital computer
94 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
---+o·-t--- v,
3-4 Examples
-----.::------. Coulol11 b
In this section we shal1 give sorne examples to illustrate how the input-output
descriptions and state-variable descriptions of linear systems are developed.
(a)
A system generally consists ofmany subsystems or components. For example,
the network in Figure 3-6 consists of three components: one resistor, one Figure 3-13 (a) Stati(
inductor and one capacitor. Ir any component of the system is nonlinear or
time varying, then the overall system is nonlinear or time varying. Hence in • Force
order to obtain a linear time-invariant model for a physical system, every com
ponent of the system must be modeled as a linear time-invariant element.
Stricdy speaking, no physical system is linear and time invariant. A tele
vision set, an automobile, or a communication satel1ite cannot function forever;
its performance will deteriorate with time because of aging or other factors. Y2
However, if the changes of characteristics are very smal1 in the time interval of
interest-say, one year-then these physical systems can be considered as time
invariant. Hence over finite time intervals, a great number of physical systems
can be modeled by time-invariant systems.
A necessary condition for a system to be linear is that for any admissible Figure 3-14 Charact
pair {O, u¡o.ro)} ---+ {x¡o.ro), Y[O,ro)}, the pair {O, au[o.ro)} ---+ {ax¡o.OO)' ay[o.GO)} for any
a, even very very large, is also admissible. For any physical system, ifthe applied
signal is larger than certain limit, the system will burn out or saturate. Hence fricUon is clearly not a l
no physical system is linear according to Definition 3-2 or 3-7. However, we neglect the static a
linear models are often used in practice to represent physical systems. This is friction. Let k¡ be thl
possible because most physical systems are designed to operate only in a certain f isgiven by f=k¡dy/,
operational ranges. Limited to these ranges, physical systems can often be 3-14. It is a nonlinear
approximated by linear models. This is accomplished by.linearlization or (y¡, yz) as shown, then I
simplification, as will be discussed in the following examples. constant. Hence the lin
and simplification.
Example 1 Now we shal1 de vele
external force u (input)
Consider the mechanical system shown in Figure 3-12. The friction force Newton's law yields
between the floorand the mass generaUy consists of three distinct parts: static
friction, Coillomb friction, and viscous friction as shown in Figure 3-13. The
EXAMPLES 95
Coulomb
how the input-output
ystems are developed. (a) (b)
lonents. For example,
Figure 3-13 (a) Static and Coulomb friction. (b) Viscous friction.
:nts: one resistor, one
system is nonlinear or
me varying. Hence in Force
Taking the Laplace transform and assuming the zero initial condition, we obtain For example, if the iJ
cart is moved to the
u(s) = (ms 2 +k 1 s +k 2 )y(s)
pendulum back to thf
Hence the input-output description in the frequency domain of the system is as a model of a space
Let H and V be, rl
A 1 A
y(s) = 2 k k u(s) the cart to the pendul
ms + lS + 2
linear movements yiel,
If m = 1, k 1 = 3, k 2 = 2, then the impulse response of the system is
or [Xl]
. =
X2
[0-;;¡ - 1][X1] +[0]
-
k2 k1
m X2
1
;;; u
These are nonlinear eql
the pendulum at the Vf
small. U nder this ass
cos e= 1. By retainin¡
y=[l O] [:J terms with e2 , ip, ee, a;
and
~Y-----1
Figure 3-15 A cart with an inverted pendulum. and A
gOll
EXAMPLES 97
11 condition, we obtain For example, if the inverted pendulum is falling in the direction shown, the
cart is moved to the right and exerts a force, through the hinge, to push the
pendulum back to the vertical position. This simple mechanism can be used
main of the system is as a model of a space booster on takeoff.
Let H and V be, respectively, the horizontal and vertical forces exerted by
the cart to the pendulum as shown. The application of Newton's law to the
linear movements yields
ystem is d2
M---l=u-H
dt 2
J
- =e -1 -e -2/
) d
2
J i . 2
H =m dt 2 (y +1 sin e) =mji +ml cos e!J -mi sin e (e)
d2 . •
and mg - V = m dt 2 (l cos e) = mi [ - sin e (j - cos e (efJ
)f the system. Let the The application of Newton's law to the rotational movement of the pendulum
tate variables; that is, yields
(M+m)ji+mllJ=u (3-38)
-1
and (3-41 )
gou(s) = (2M +m)/s 2 - 2g(M +m)
98 MATHEMATlCAL DESCRIPTlONS OF SYSTEMS
Xl O O O Xl O
-2mg 2
X2 O O O X2 These equations can b
2M+m 2M+m u (3-42)
O O
+ O -k
X3 O X3 VI (t) O O
2g(M +m) -1 mi
X4 O O O X4 y 12(t) 1 O -1
(2M +m)l (2M +m)l
-k
V2(t) O O
y =[1 O O O]x m2
ydt) O O 1
Note that Equations (3-39) to (3-42) are obtained l;lnder simplification and
linearization and are applicable only for small e and e. I v3(t) O O O
Consider four vehicles moving"in a single lane as shown in Figure 3-16. Let V4(t) O O O
Yi, Vi' mi' and Ui be, respectively, the position, velocity, mass and the applied
force of the ith vehicle. Let k be the viscous friction coefficient and be the same This is the state equati
for a11 four vehicles. Then we have, for i = 1,2,3,4, interested in the diste
therefore the absolute
V¡=Yi (3-43) on what will be consic
(3-44 ) developed.
Ui =kv¡ +m¡vi
The purpose of this problem is to maintain the distance between adjacent Example 4 (ReferencE
vehicles at a predetermined value ha and to maintain the velocity of each Consider a satel!ite of
vehicle as clase as possibie to a desired velocity Va. Define altilude of lúe sateilite
Yi,i+l(t)=y¡(t)-Yi+l(t)-h a i=1,2,3 (3-45)
U
'l"/./##$////#//./#//#//.//#/##//./$#####/###//#///////#/#///##//////////////#/#/////////////////.
-Y4~ Y3~ ~
.
- - - - - - - - - - - - - - - - - . : . . . - - - . - - - - Y2 YI-J
Figure 3-16 Four vehicles moving in a single lane. Figure 3-17 Satellite il
-" --- ~--"--_=:...::-
EXAMPLES 99
Y34(t) O O O O 1 O -1 Y34(t) O O O O
-k o, 1
V4(t) O O O O O O v 4(t) O O (3-48)
Tn in Figure 3-16. Lef m4 m4
mass and the applied
This is the state equation description of the system. In this problem, we are
fficient and be the same
interested in the distances between adjacent vehicles and their velocities;
therefore the absolute distances do not appear in the equations. Depending
(3-43) on what will be considered as outputs, an output equation can be similarly
developed.
(3-44)
Example 4 (Reference 596)
lnce between adjacent
n the velocity of each Consider a satellite of mass m in earth orbit as shown in Figure 3-17. The
fine aititude of the satellite is specHled by r(e),O(tj, and cPU) as shown. The orblt
2,3 (3-45 )
(3-46)
U&
-
] UI-G
VI
y -
I Salellite (mass m)
I
1",
.
.
& --
.....................
1""
'"
7//////'///'//$///'//&//////////,I/. --- ---- - .""v
can be controlled by three orthogonal thrusts U,.(t), ua(t), and u.p(t). The state, lf the perturbation is '
input, and output of the system are chosen as can be linearized as
r(t) O 1 O
I\t)
8(t) U,.(t)] r
u)] 3w 2 O O
x(t) = u(t)= ua(t) y(t)= 8(t)
e(t) [ uq,(t) [ O O O
cP(t) cP(t)
-2w
x(t) = O O
~(t) ro
Then the system can be shown to be described by ------------
O O O
r
re 2COS 2cP +r~2 - k/r 2 +u,./m O O O
x= h(x u) =
,
. .. e
-2r8/r +284> sin <PIcos 4> +ua/mr cos cP
(3-49a)
l O O O: O
. ~. y-( t ) = O O 1 O:, O
-:-
_8 2 COS cP sin cP -2rcP/r +u.p/mr
1 O O O O O]
f
-- - -- -- - -
O O O O: 1
where A =oh(x, u)/ox,
and Y= ex = O O 1 O O o x (3-49b)
[ and u o. This is a sixth
o o o o 1 o
lt can be used to descri
One solution which corresponds to a circular, equatorial orbit is specified by the circular orbit rema
ro(t) ro The dashed lines in
ro(t) . O two uncoupled parts, (
~o(t) wt two parts independentl:
xo(t) = uo(t) =0 (3-50)
8 0 (t) W
<ro(t) O °Dynamical equatio
cPo(t) O section a systematic prc
cal equations for genera
where ro and w are related by r6w2 =k =a known physical constant. Once pendent voltage and Cl
the satellite reaches this orbit, it will remain in the orbit as long as there are currents and the capac
no disturbances. lf the satellite deviates from the orbit, thrusts must be applied behavior of the networ
to push the satellite back to the orbit. Define is not necessary to cho
'ro + cX l(t) -, state variables. This ca
t:X2(t) lf we assign all the cap,
as shown, then we see tI
x(t) =xo(t) +t:X:(t) = wt +t: X3(t)
w +6: 4 (t)
cXs(t)
R
cX6(t)
+
ro +cXl(t)]
and y(t)=yo(t) +t:y(t) = wt +t:x~U) Figure 3-18 Circuits wit:
[ consists ofinductors only.
cXs(t)
- - - - - - - -- ------- -
EXAMPLES 101
" and u,¡,(t). The state, If the perturbation is very small, or equivaIentIy, é: is very small, h(x, u) in (3-49)
can be linearized as
,
o 1 O O ,, O O O O O
, 1
,
[~l)
8(t)
,] 3w 2
O
O
O
O 2wro , O
O
,
,
1 ,, O
O
O
m
O
O
,
O ,, O
O
cP(t) -2w , 1 ,
x(t) = O O O , O O x(t) + O O üU)
, ~,
(3-51a)
ro , mro:
, ,
------------------,------
, ----------
, I
O O O O , O
I 1 O O ,, O
, 1
O O O O O
l,/m , mro
I
(3-49a)
/mrcoscjl I O O O: O 0J
y(t)= ~ __~ __1__ ~_~-~--~ x(t) (3-51b)
u<f¡/mr
0000:10
(3-49b)
where A =oh(x, u)/ox, B =oh(x, u)/ou, and computed at the circular orbit at Xo
and uo. This is a sixth-dimensionallinear time-invariant dynamical equation.
lt can be used to describe and control the satellite so long as the deviation from
rial orbit is specified by the circular orbit remains smal!.
The dashed lines in (3-51) show that the equation can be decomposed into
two uncoupled parts, one involving r and 8, the other <p. By studying these
two parts independently, the analysis and design can be considerably simplified.
(3-50)
*Dynamical equations for RlC networks. We introduce in this sub
section a systematic procedure for assigning state variables and writing dynami
cal equations for generallumped linear RLC networks which may contain inde
ysical constant. Once pendent voltage and current sources. lt is well known that if all the inductor
)it as long as there are currents and the capacitor voltages of an RLC networks are known, then the
thrusts must be applied behavior of the network is uniquely determinable for any input. However, it
is not necessary to choose aH the inductor currents and capacitar voltages as
state variables. This can be seen from the simple circuits shown in Figure 3-18.
lf we assign all the capacitor voltages and inductor currents as state variables
as shown, then we see that Xl (t) = xz(t) for aH t. Clearly there is a redundancy
R LI
+
¡-
X2 R
XI
C
c l C2
d L2
Figure 3-18 Circuits witha: loop which consists of capacitors only or a cutset which
consists ofinductors only.
102 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
here. Hence the state nf an RLC network can be chosen to consist of only inde
pendent capacitor voltages and independent inductor currents.
Before proceeding, we review briefly the concepts of tree, link, and cutset
of a network. We consider only connected networks. A tree of a network is
defined as any connected graph (connection of branches) containing aB the
nodes of the network and not containing any loop. Every branch in a given
tree is caBed a tree branch. Every branch not in the tree is caBed a link. A
cutset of a connected network is any minimal set of branches such that the (voltage
source)
removal of aB the branches in this set causes the remaining network to be un
connected. With respect to any fixed tree, every link and sorne tree branches
form a unique loop caBed a fundamental loop; every tree branch with sorne
links forro a unique cutset caBed a fundamental cutset. Hence every funda
Figure 3-19 A networ
mental loop ineludes only one link, and every fundamental cutset ineludes only
one tree branch. With these concepts, we are ready to give a systematic pro
By applying the Kirchl
cedure for developing a dynamical-equation description of any RLC network we have immediately Ü
that may contain independent voltage sources or current sources 9 : the voltage across resis
1. Choose a tree caBed a normal tree. The branches of the normal tree are are expressed in terms
chosen in the order of voltage sources, capacitors, resistors, inductors, and to the fundamental cuts
current sources. Hence, a normal tree consists of aB the voltage sources,
the maximal number of permissible capacitors (those that do not form a
loop), the resistors, and finaBy the minimal number of inductors. UsuaBy The application of the I<
it does not contain any current source.
2. Assign the charges or voltages of the capacitors in the normal tree and the
flux or current of the inductors in the linksas state variables. The voltages This application of the }
or charges ofthe capacitors in the links and the flux or current ofthe inductors
in the normal tree need not be chosen as state variables.
3. Express the branch variables (branch voltage and current) ofaB the resistors, These three equations ca
the capacitors in the links, and the inductors in the normal tree in terms ofthe
state variables and the inputs by applying the Kirchhoff voltage or current
law to the fundamentalloops or cutsets of these branches.
4. Apply the Kirchhoff voltage or current law to the fundamental loop or
cutset of every branch that is assigned as a state variable. The output equation can
Example 5
Consider the linear network shown in Figure 3-19. The normal tree is chosen Example 6
as shown (heavy lines); it consists of the voltage source, two capacitors, and Find the input-output d{
one resistor. The voltages of the capacitors in the normal tree and the current equivalently, the transfer
of the inductor in the link are chosen as state variables. Next we express the
i~puts, one output, the tn
variables of resistors CD and (2) in terms of the state variables and inputs. By
[g¡¡(~) 9ds)], where 9¡
applying the Kirchhoff voltage law (KVL) to the fundamental loop of branch and gds) is the transfer J
CD, the voltagé across CD is found as '(u¡ ~ Xl); hence its current is (u¡ - xJ. network reduces to the Oll<
9 A network with a loop that consists of only vo'ltage sources and ~apacitors or with a cutset that
consists or onlycurrent sources and inductors is excluded, because in this case its dynamical
L.
equation description cannot be or the form in (3-33). - -1
s
EXAMPLES 103
/ I
A tree of a network is
les) containing aH the
rery branch in a given lF
~ee is caHed a link. A (voltage
ranches such that the source) G)'
ling network to be un
nd some tree branches
.ree branch with some
. Hence every funda Figure 3-19 A network with voltage and current sources.
tal cutset includes only
give a systematic pro
By applying the Kirchhoff current law (KCL) to the fundamental cutset of (Z),
1 of any RLC network
we have immediately that the current through resistor (2) is X3' Consequently
11 sources9 :
the voltage across resistor (2) is X3' Now the characteristics of every branch
are expressed in terms of the state variables as shown. If we apply the KCL
of the normal tree are
to the fundamental cutset of branch Q), we have
esistors, inductors, and
o
O
-1
-1] [1
-1
1 x + O
O
.nches.
le fundamental loop or
The output equation can be easily found as
·iable.
-- .1'"
. . /.,
Example 6
he normal tree is chosen
1.11
1 R
T,--~[_':1_
S
= E c
Ca) Cb)
1s
1+
_ (l/s)u¡(s)
2( )-(1 +l/s)(l +s +2/s)-1/s 2 L LCT
r
Hence
A y(s)
g¡¡(S)=-A- u,~o
sI 2(s)
=-A-= ~ 32
S2
1
El__ ~
u¡(s) inilially u¡(s) s- +2s + s +
relaxed Cb)
Ir U¡ = 0, the network in Figure 3-19 reduces to the one in Figure 3-20(b). Figure 3-21 Network w
By node analysis,
1 )
These can be rearranged
( S +s
A A
V¡(s) -sV2(s) =0
-llR l
L i = h(u)
E=.
(h)
T
(a)
L L Xl
xl
+ +
R R
1+ J+ C
Xl e u Rl
Xl ¡¡ -R 2
i=
-L
+2s z +3s +1 (h) (e)
•
Now ir x 1(t) is known to operate only inside the range (e, d), we may intro
duce the variables Xl(t)=Xl(t) -vo, xz(t) =xz(t) - io and approximate h(Xl(t))
ash(x l(t)) = i o - Xl(t)/ R z. The substitution or these into (3-52) yields .
~ T is a tunnel diode with
~J[~l(t)J +l~JE
~oss the capacitor and Xz
1· .. R xz(t)
__ L·
L L
106 MATHEMATlCAL DFSCRIPTIONS OF SYSTEMS
We compare in this section the input-output description and the state-variable Figure 3-23 Active:
description of systems.
output will increase without bound if the initial condition is different from
zero. Although the output of the network in Figure 3-23 behaves well,
the network is still not satisfactory, because the voltages in branches 1 and 2
The response of a I
will increase with time (in different polarity), and the network will eventuaJJy given by
burn out. Hence the networks in Figures 3-22 and 3-23 can never function
properly. Ir the internal structure of the network is not known, this fact
u i¡
In In
r Y¡
For the time-invaria
u
t
iz -IF
+"
~
X Thus the impulse re~
In In (3-55) or (3-56). See
For linear time
functions or impuls<
Figure 3-22 A network with linear, time-varying elements. frequency s = jw can
COMPAR1SONS OF THE lNPUT-OUTPUT DESCR1PTlON 107
In ;n
+
1 Y2
n completely. where ~lí(t, to) is called the step response (dlle to a unit step function applied
works in Figures 3-22 at to). Differentiating (3-54) with respect to to, we obtain
Iy, as
(3-65 )
Thus the impulse response can be obtained from the step response by using
(3-55) or (3~56). See Problem 3-25.
For linear time-invariant systems, we may measure either transfer
functions or impulse responses. The transfer function g(s) of a system at
frequency s = jw can be measured easily and accurately by employing fre
108 MATHEMATlCAL DESCRIPTIONS OF SYSTEMS
quency response analyzers. After measuring gUw) at a number of fre two multivariable s:
quencies, we can then find a g(s) to match the measured gUw). This method
of determining transfer functions is often used in practice.
3. Prior to 1960, the design of control systems had been mostiy carried out by
using transfer functions. However, the design had been limited to the single where u i and y¡ are 1
variable case; its extension to the multivariable case had not been successful. of the system Si' 1
The state-variable approach was developed in the 1960s. In this approach, the impulse-respon:
the formulations in the single- and multivariable cases are the same, and a 3-24 that in the pa
number of results were established. These results were not available in the tandem connection,
transfer-function approach at that time; consequentiy, interest in this we have u 1 =u - Y2'
approach was renewed in the 1970s. Now the results in the state-variable SI and S 2 have com¡:
approach can also be obtained in the transfer-function approach. It also cannot be properly
appears that the latter approach is simpler in concepts and computations. effect in the connect
4. The dynamical equation can be extended to the time-varying case; the unchanged after COI
extension of the transfer function to the time-varying case has not been matrix of the paralle
successful. In the optimal design, dynamical equations can be used to study
finite terminal time problems; this is not possible by using transfer functions.
5. In rhe study of nonlinear systems, depending on the approach taken, either For the tandem e
description can be used. For example, in the study of the stability problem,
we use the input-output description in the functional analysis and operator
approaches. See References 28, 95, 117, and S79. Ir Lyapunov's second
method is employed, then we must use the dynamical-equation description. We prove (3-59) for ti
6. Ir the dynamical-equation description of a system is available, the system definition, the respon
.can be readily simulated on an analog or a digital computer. response at Y1 due to 1
gl(t,r)is
From the foregoing discussion, we see that the input-output and the state
variable descriptions have their own merits. In order to carry out a design
efficientiy, a designer should make himself familiar with these two mathematical
descriptions. In this book, these two descriptions will be developed equally
and their relationships will be explored.
10 The material in this section is not used until Chapter 9; thus its study may be postponed. Tandem connection. (e)
-equation description. We prove (3-59) for the single-variable case. The impulse response g(t, r) is, by
available, the system definition, the response at yz due to a b-function applied at time r at Ul' The
nputer. response at Y1 due to this o-function is g 1(t, !). The output of Sz due to the input
gl(t,r)is
-output and the state
to carry out a design
lese two mathematical
J: gz(t, v)g I(V, r) dv
be developed equal1y
y =
u Y¡ + Y2
site Systems 10 U=U;{~
O-~/
+
built by interconnect
Josite system. Whether
'e consider as building (a) (b)
Figure 3.24 Composite eonneetions of two systems. (a) Parallel eonneetion. (b)
Tandem eonneetion. (e) Feedback eonneetion.
I may be postponed.
110 MATHEMATICAL DESCRIPTlONS OF SYSTEMS
r
matrix is the solution of the integral equation
where G[ and G z are known and G is unknown. This equation can be easily y = [Y [(t)C[(t)
verified from the definition of G(t, r) (Problem 3-26). There is a general iterative
method for solving Equation (3-60), but it is very involved.
where y [(t)=(I+E l ,
Now we study the state-variable descriptions of composite systems. Let
order for (3-64) to be
the systems S [ and S z in Figure 3-24 be described by
and (1 + E z(t)E l (t)) eJ
X¡ = A¡(t)x¡ + B¡(t)u¡ i= 1, 2 (3-61 a) verified by observing
y¡ =C¡(t)x¡ +E¡(t)u¡ (3-61b)
Time-invariant ca
where x¡ is the state, o¡ is the input, and y¡ is the output; A¡, B¡, C¡ and E¡ are
applied to the time-;
matrices of compatible order whose entries are continuous function of t defined
discuss the transfer-J
over l - 00, (0). The state space of S ¡ is denoted by L¡.
Gz(s) be the proper
Let us introduce the concept of the direct sum of two linear spaces. The
tively; then the transl
linear space L is the direct sum of two linear spaces L[ and L2> written as
is G l (s) +Gz(s). Th
L = L[ EB LZ' if every vector in L is of the form [x{ xí]', where x[ is a vector
SI followed by Sz is
in L[ and X z is a vector in LZ' The dimension of Lis the sum of those of L[
reversed. It is clear
and L z.
G Z(s)G l (s). It is im
It is clear that the composite vector
compatible in both c;
In order to discus
shown in Figure 3-24
qualifies as the state of any composite connection of S [ and S z; its state space
Theorem 3-2
is the direct sum of the state spaces of S [ and S z, L [ EB LZ' F or the parallel
connection, we have o[ = U z = o, y = y [ +y z; hence its dynamical equation is Let G l (s) and Gz(s) b
(not necessarily prope
[::] == [A~(t) A~(t)J[:J +[::~:~J u (3-62a)
(3-63b)
Proof of "Theprem 3-:
description is
.le case.
the impulse-response
(3-64a)
r) ds dv (3-60)
l. where Y 1 (t) = (1 + E 1(t)E 2(t»- 1 and Y 2(t) = (1 + E 2 (t)E 1 (t))-1. It is clear that in
lposite systems. Let order for (3-64) to be defined, we must assume that the inverses of (1 + El (t)E 2 (t»
and (1 + E 2 (t)E 1 (t» exist for all t. The dynamical equation (3-64) can be easily
(3-61 a) verified by observing U1 = U-Y2' Y1 =U2' Y=y [. (See Problem 3-28.)
(3-61 b)
Time-invariant case. All the results in the preceding subsection can be
. A¡, Di, Ci and E¡ are applied to the time-invariant case without any modification. We shall now
; function of t defined discuss the transfer-function matrices of composite systems. Let G1 (s) and
G2(s) be the proper rational transfer-function matrices of SI and S 2, respec
o linear spaces. The
tively; then the transfer-function matrix of the parallel connection of SI and S 2
[ and ~2' written as
is G[(s) + Gz(s). The transfer-function matrix of the tandem connection of
, where x[ is a vector
SI followed by S2 is G 2(5)G 1 (5). Note that the order of G 2(S)G[(5) cannot be
le sum of those of ~ [
reversed. It is clear that if Gi (5), i = 1, 2, are proper, so are G1(5) + G2(s) and
G 2 (5)(;[(5). It is implicitly assumed that the orders of G[(5) and G 2 (5) are
compatible in both connections.
In order to discuss the transfer-function matrix of the feedback connection
shown in Figure 3-24(c), we need sorne preliminary results.
liI
¡(t)Ju (3-62b)
Observe that the matrix on the right-hand side oí (3-65) is a q x q matrix,
while the matrix on the left-hand side is a p x p matrix. 1m is the unit matrix
S [ and S2 is given by
of order m. The elements of these matrices are rational functions of s, and
[(t)
)E[(t) u
J (3-63 a )
since the rational functions form a field, standard results in matrix theory can
be applied.
where N, Q, and Pare any square matrices of the same order. Let us choose the feedback system
N - [Ip 0J Q _ [I p - (;2(S)J P = [I p (;2(S)J G(s) = G I,
- - (;¡(s) I q - (;¡(s) Iq O Iq
They are square matrices of order (q + p). lt is easy to verify that Proof
From Figure 3-24(c),
NQP=[I p AO A ]
O I q +G¡(s)G 2 (s)
Proof
Note that the zero in det(I q +G I(S)G 2 (S))+0 is the zero element in the field
of rational functions. Hence, it can be written more suggestively as
lt is easy to verify tha
det(Iq +G I(s)G 2 (s))+O for sorne s. The ~ondltion det(Iq +G¡(s)G 2 (s))+O
implies that the inverse of the matrix (I q +G I(S)G 2 (s)) exists. From Theorem
3-2, we have
det(I q +G I(s)G 2 (s))=det(I p +G 2 (S)G 1(S)) +0
Hence, both (Iq +G 1(S)G 2 (s))-1 and (I p +G 2 (S)G 1(s))-1 exist. Consider the
identity
Proof
erify that
From Figure 3-24(c), we have (;l(s)(ii(s) - (;2(S)Y(S)) = y(s), or
(I q + (;1(S)(;2(S))Y(s) = (;l(s)ii(s) (3-69)
+
Note that the condition det (I q + (; 1(S)(;2(S)) O is essential for a feedback
system to be defined. Without this condition, a feedback system may become
meaningless in the sense that for certain inputs, there are no outputs satisfying
Equation (3-69).
a different proof, see
Q.E.D.
Example 1
Consider a feedback system with
-s
A s +1
G 1(s)= _1_
s +2
1 ]
G 2 (s)= O 1
A [1 0J
[
s +1
~s~1
o element in the field
aor( suggestively as
lt is easy to verify that det (1 2 + (;1(S)(;2(S)) =0. Let us choose
et(I q + (;1 (S)(;2(S)) +0
dsts. FroID Theorem
+O
A
=
s +21]
s))
exist. Consider the
u(s)
[~
(s 1)2
(3-67)
Recal1 from (3-64) that the condition for the existence of the dynamical
Let (;l(S) and (;2(S) equationdescription ofthe feedback system in Figure 3-24(c) is det (1 + E 1 E 2 ) O +
[atrices of Sl and S2, where Eiis the direct transmission part of the dynamical equation of Si'
fer-function matrix of +
Because of(;¡(co) = E¡ [see Equation (3-37)J, if det (1 + E 1 E 2 ) O, then det(1 +
I
I
ji
G¡ (s)G 2 (s)) +-2. However, the converse is not true; that is, the condition
det(1 +G¡(s)G 2 (s))+-0 may not imply det(1 +E¡EJ+-O. Hence, a feedback
system may have the transfer-function matrix description without having the U¡
state-variable description. This discrepancy will be resolved in the following +
H
subsection.
Example 2 F
Uz + J -s .¡
Consider the feedback system shown in Figure 3-24(c) with s+
[
-1
G1(s)= _1_
s +1
-JJ
s+lJ
Figure 3-25
l
rational matrix. In th,
Before proceeding, we
s
1 Theorem 3-4
-1
s +1 Let M(s) be a square r
s +1 s+l s+l s +1
M-¡(s) = {Mp(s)[1
which is not a proper rational function matrix. The block diagram of this If M; ¡(s) is proper, tr
feedback system is plotted in Figure 3-25. • strictly proper, M sp( G
constant matrix. Thu:
For thefeedback system in Figure 3-25, if the input signals u(t) is corrupted From M-1(s)=[
by high-frequenS?' noises o(t), then the noises will be amplified by the improper (Mp(S)+O)-l-.[Adj l.\
rational matrix G(s). For example, let u(t)= sin t U o and o(t) =0,01 sin 1000t 0 0
where U o and oo' are 2 x 1 constant vectors. Although the amplitude of the 11 A sufficient condition for
noise is only one hundredth ofthe amplitude ofthe signa! atthe inpútterminals, Appendíx G. For necess
the ainplitude of the noise at the output terminals is 10 times larger than that 3-4 of Reference S34.
MATHEMATICAL DESCRIPTIONS Of COMPOSITE SYSTEMS 115
Theorem 3-4
Let M(s) be a square rational matrix and be decomposed uniquely as
M(s) = Mp(s) + Msp(s) (3-71 )
The overall transfer
where Mis) is a polynomial matrix and Msp(s) is a strictly proper rational
matrix. Then M - i(s) is proper if and only if M; 1(S) exists and is proper. 11
s +1
Proof
-s We first show that if M; 1(S) is proper, so is M- 1(s). We write (3-71) as
M- 1(s) = {Mp(s)[1 +M; 1(s)Msp(S)]} -1 = [1 + M; 1(s)Msp(S)] - 1M; I(S)
block diagram of this If M; 1(S) is proper, then M; 1((0) is a finite constant matrix. Since Msp(s) is
• strictly proper, Msk~)=O. 'Hence we have M-1(00)=M;1(OO), a finite
constant matrix. Thus M -1(s)'is proper.
ignals u(t) is corrupted From M- 1 (s)=[M p(s)+M sj,(s)]-1, we have, as· S~OO,M-1(S)~
,¡¡tied by the improper (Mp(S)+O)-1~[Adj Mp(s)]fdet Mp(s). Hence if Mp(s) is singular, M- 1(oo) is
n(t) =0.01 sin l000t no
L the amplitude of the l' A sufficient condition for M; '(s) to be proper is Mp(s) column-reduced or row-reduced. See
at the input terminals, Appendix G. For necessaryj'and sufficiént conditions, see Problem G.l7 and Sections 2-4 and
times larger than that 3-4 of Reference 534.
116 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
not a finite constant matrix and M- 1(s) is not proper. 12 Now we assume The condition de
Mis) to be nonsingular and show that if M- 1(s) is proper, so is M; 1(s). From theorem is the same a
M(s) = Mp(s)[1 + M; 1(s)Msp(s)], we have dynamical descriptior
resolves the discrepan
Mo(s)~ [1 + M; 1(s)Msp(s)] -1 = M- 1(s)Mis) = M- 1(s)[M(s) - Msp(s)]
Before discussin
= 1- M -1(s)M sp(s)
det [1 +G 2 (00)G 1(oof
which, together with the finiteness ofM -1(00) and Msp(oo) =0, implies Mo(oo) = Figure 3-26. This sy~
1. Hence we have M- 1(00) = Mo(oo)M; 1(00) = M; 1(00), and M;I(OO) is a and will be developed
finite constant matrix. Thus M; I(S) is proper. Q.E.D. The transfer matrices
matrices of appropriat
This theorem shows that the properness of M- 1(s) depends only on the
e(s)=í
polynomial part of M(s). We give a special case in the following.
which implies
Corollary 3-4
e(s) = I
If M(s) is a square proper rational matrix, M -1(S) is proper if and only if M( (0)
Hence the transfer mal
is nonsingular.
Gf(s) = G2
Proof
Ip view of Theorem 3-:
lf M(s) is proper, the polynomial part of M(s) is Mis) = M(oo). Hence M- 1(s) G o(00 ) is nonsingular, .
is proper if and only if M( (0) is nonsingular. Q.E. D.
Go(.
Using these results, we are ready to answer the question posed at the begin lf G¡(s) are proper, the
ning of this subsection: the properness of G(s) in (3-68). nonsingular, Corollary
in (3-73) is proper. H{
Theorem 3-5 condition for Gf(s) to b
Consider the feedback system shown in Figure 3-24. Let G 1(s) and G 2 (s) be
q x pand p x q proper rational transfer matrices of SI and S2' Then the overall
transfer matrix
G(s)=G 1(s)[1 +G 2 (S)G 1(S)]-1
is proper if and only ifl + G 2 ( oo)G1(00) is nonsingular.
Then we have Go(co)=
Proof
lf I +G 2 (00)G 1(00) is nonsingular, [1 +G 2 (S)G 1(S)]-1 is proper. G(s) is the which is proper.
product of two proper rational matrices; hence it is proper.
lf G(s) is proper, so are G 2 (s)G(s) and 1- G 2 (s)G(s). Since
[1 + G2 (s)G 1(s)] -1 = 1- G 2 (S)G 1(s)[1 + Gz(s)G 1(s)] -1 = 1- G 2 (s)G(s)
which c~n be readily verified by postIllultiplying 1+ Gz(s)G 1{s), we ~conclude
that if G(s) is proper, so is [1+G 2 (s)G 1(S)]-I. Hence [1+G 2 (oo)G 1(00)]-1
'¡s:finite and 1+ G 2 (00)G 1(00) is nonsingular. This establishes the theorem.
Q.E.D.
12 For adifferenl proof of this sta temen l. see Problein3A4. Figure 3-26 A feedbad
MATHEMATICAL DESCRIPTIONS OF COMPOSITE SYSTEMS 117
r. 12 Now we assume The condition det [1 + (;1(00)(;2(00)J =det [1 + (;z(oo)(;I(oo)J =/=-0 in this
~r, so is M; 1 (s). From theorem is the same as the condition det (1 + E 1 E 2 ) =/=- O for the existence of the
dynamical description of the feedback system shown in Figure 3-24(c). This
resol ves the discrepancy between these two descriptions.
(s) [M(s) - Msis)] Before discussing the physical implications of the condition
det [1 + (;2(00)(;¡{00)] =0, we study first the composite system shown in
)= 0, imp\ies M 0(00 ) = Figure 3-26. This system is a generalization of the system in Figure 3-24(c)
X»), and M;I(oo) is a and will be developed in Chapter 7 and then extensively studied in Chapter 9.
Q.E.D. The transfer matrices (;¡(s) in Figure 3-26 are assumed to be proper rational
matrices of appropriate orders. Clearly, we have
1 depends only on the e(s) = o(s) - (; 3(S)(;1 (s)e(s) - (;4(S)(;z(S)(; 1 (s)e(s)
tllowing.
which implies
e(s) = [1 + (;3(S)(;I(S) + (;4(S)(;2(S)(;1 (s)] -I U(S) (3-72)
Jer if and only if M( 00 ) Hence the transfer matrix from u to y is given by
(;¡(s) = (;z(S)(;1 (s)[1 + (;3(S)(;I(S) + (;4(S)(;2(S)(;, (s)]-' (3-73)
Ip view ofTheorem 3-5, one may wonder whether (;¡(s) is proper ifand only if
M(oo). Hence M-'(s) G o( (0) is nonsingular, where
Q.E.D. (;o(s) ~ I + (;3(S)(;,(s) + (;4(S)(;2(S)(;I(S)
Ion posed at the begin- If Gj(s) are proper, the polynomial part of (;o(s) is (;0(00). Hence, if Ca((0) is
nonsingular, Corollary 3-4 implies that (;0 '(s) is prop~r. Consequently, (; ¡(s)
in (3-73) is proper. However, the nonsingularity of G o( (0) is not a necessary
condition for (;¡(s) to be proper. For example, consider
Jet (;1(S) and (;z(s) be
l
1
d 52' Then the overall ... s
Gz(s) = O
For this system. GAs) is proper; wh...e reas Go( (0) is singular. Will such a from any point to an
system be acceptable in practice? lf Go(oo) is singular, the transfer matrix establishes the first p,
from u(s) to c(s) in (3-72) is improper. In this case, if u(t) contains a part which the polynomial part of
has a high-frequency spectrum or is discontinuous, the amplitude of e(t) will be proper if and only if é
very large or infinite, and the system may saturate or bum out. Hence, in the
design of a feedback system, we shall require not only the overall transfer matrix We discuss now the
but also all transfer functions from all possible input-output pairs of the system and more generally, de
to be proper. In this case, no signal will be unduly amplified, and in sorne sense, A loop of a block diag
the smoothness of signals throughout the system will be preserved. direction of the path b
For example, the feedl
Definition 3_9 13 is indicated by the hea'
of all transfer function
Let every subsystem of a composite system be describable by a rational transfer
In our discussion, we ~
function. Then the composite system is said to be well posed if the transfer
Clear1y if a loop gain i
function of every subsystem is proper and the c1osed-loop transfer function
zero at s = oo. lf aH t
from any point chosen as an input terminal to every other point along the
has a nonzero loop ga
directed path is well defined and proper. I
loop are exactIy prope
degree of its numerato
In this definition, if a point, say E in Figure 3-26, is chosen as an input
only loops with nonze
terminal, then we must add a fictitious input as shown. Then the c1osed-loop
Consider the feedb:
transfer functions from r to e, W, and y are, respectively.
not well posed becaus
e(s)= ---, [1 + G 3 (S)G 1 (s) + G 4 (s)Gz(s)G 1 (S)]-l (S) r loop with loop gain +
w(s) = - G 3 (S)G 1 (s)[1 + G 3 (S)G 1(s) + G 4 (S)G 2 (S)G 1(s)] -l r (S) (3-74) improper transfer func
conc1ude that if a systel
and y(s) = - Gi(s)G 1 (s)[1 + G 3(s)Gi(s) + G 4 (S)G 2(S)G 1(s)] -l r (S)
+ 1 at s = 00 which is
We note that in computing these transfer functions, no branch in the system function.
should be disconnected. Hence we compute c1osed-loop transfer functions. The system in FigUJ
lf we disconnect any branch, then the system becomes a different system. Go(00)=1-1 +2#0,
from the existence of 2
Theorem 3-6 loop will offset the 100
The system in Figure 3-26, where G¡(s) are rational transfer matrices of appropri
ate orders, is weH posed if and only if Gi(s), i = 1, 2, 3,4, are proper and the
rational matrix ~~- '---.J
(3-75)
Proof
The transfer matrices in (3-72) to (3-74) are c1early proper if and only if G l(S) o
o
is proper. Similarly, it can be shown that if G l(S) is proper, the transfer matrix
(e)
13 This definition is similar to the one in Reference S34 and is applicable only to linear time-invariant
lumped systems. For a more general definition and discussion, see References SS, S2Ü7 and S214. Figure 3-27 Feedback
MATHEMATlCAL DESCRIPTlONS OF COMPOSITE SYSTEMS 119
singular. Will such a from any point to any other point along the directed path is proper. This
tr, the transfer matrix establishes the first part of the theorem. If a11 Gi(s), i = 1, 2, 3, 4, are proper.
I contains a part which the polynomial part of Go(s) is Goloo). Hence (;0 I(S) is, fo11owing Corollary 3-4,
mplitude of e(t) will be proper if and only if Go((0) is nonsingular. Q.E. D.
1m out. Rence, in the
overall transfer matrix We discuss now the implications ofthe conditions det [1 + G2 ( CIJ )G 1 (CIJ)J f- O
put pairs of the system and more genera11y, det Go(oo) f- O. Before proceeding, we need sorne concepts.
fied, and in sorne sense, A loop of a block diagram is a closed path which travels from a point along the
preserved. direction of the path back to the same point and does not pass any point twice.
For example, the feedback system in Figure 3-25 has three loops; one of them
is indicated by the heavy lines. The loop gain of a loop is defined as the product
of a11 transfer functions along the loop including the signs at summing points.
e by a rational transfer
In our discussion, we are interested in only the value of the loop gain at s = oo.
II po sed if the transfer
Clearly if a loop gain is a strictly proper rational function, then its loop gain is
loop transfer function
zero at s = CIJ. If a11 transfer functions of a block diagram are prdper, a loop
other point along the
has a nonzero loop gain at s = 00 if and only if a11 transfer functions along the
I
loop are exactly proper (a rational function is said to be exactly proper if the
degree of its numerator is equal to that of its denominator). In the following,
is chosen as an input
only loops with nonzero loop gains at 5=00 will come into the discussion.
Then the closed-Ioop
Consider the feedback systems shown in Figure 3-27(a) and (b). Theyare
not well posed because det Go(oo) =0 in both systems. Each system has a
'(5)
loop with loop gain + 1 at s = 00, which is equivalent to a gain of infinity or an
)G 1(5)]-11-(5) (3-74)
improper transfer function as shown in Figure 3-27(c) and (d). Hence, we may
)G 1(5)] -11-(5)
conclude that if a system is not well posed, the system has a loop with loop gain
+ 1 at s = CIJ which is equivalent to an infinite gain or an improper transfer
o branch in the system function.
10P transfer functions. The system in Figure 3-28(a) has a loop with loop gain 1; however, we have
dilTerent system. Go(oo) = 1-1 +2 f-O,and the system is we11 posed. This discrepancy arises
from the existence of another loop with a nonzero loop gain at s = oo. This
loop will offset the loop with loop gain 1, as can be seen from Figure 3-28(c)
:r matrices of appropri
4, are proper and the
(3-75)
(a) (b)
(e) (d)
: onlyto linear time-invariant
teferences S8, S207 and S214. Figure 3-27 Feedback systems which are not well po sed.
120 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
or an improper transfe
+ }--E..-_ _---+I 2s + 1 to be well posed.
-s
3-7 Discrete-TiI
outputs of discrete-tirr
For example, a digital
of variables at discrete
continuous-time syster
responses are of intere:
For convenience, ti
output appear will be ;
(e) (d) time interval T is cal
{y(k) ~ y(kT)}, k = O, j
Figure 3-28 (a) A well-posed system. (b) An ill-posed system. (e), (d) Equivalent
In the following, we di
systems of (a).
the continuous-time S)
called a relaxed discre,
and (d). Henee, if there are two or more loops with nonzero loop gains at discrete-time system sa
s = 00 passing the same point, such as the point E shown in Figure 3-28(a),
these loops must be grouped into a combined loop with a net loop gain equal to
the sum of all individual loop gains. For the system in Figure 3-28(a), the
combined loop has a net loop gain 1 - 2 = -1 which is different from + 1; where g(k, m) is called t
henee the loop will not cause any problem, and the system is well posed. The due to the application (
system in Figure 3-28(b) has a combined loop with loop gain 3 - 2 = 1; hence
the system is not well posed (Problem 3-39).
From these examples, we conclude that if a system has a combined loop with
net loop gain 1 at s = 00, the combined loop is equivalent to a gain of infinity with the system relaxe
or an improper transfer function. Note that what causes the problem is the that is, the output doe
combined loop instead of individual loops. ylk? = Q [ol' k < ;Ti. "
To further verify the aforementioned statement, we consider the system relaxed at ka, then (3-7(
shown in Figure 3-25. There are two loops passing through the point H.
Along the loop indicated by the heavy lines, there is a loop between points
E and F. In computing the loop gain of the heavy-lined loop, the transfer
function from E to F should be computed. However because of the two strictly lf a linear causal relaxe
proper transfer functions, the loop gain is zero at s = oo. Hence, the combined g(k, m) = g(k - m) for al]
loop which passes through the point H has a net loop gain 1 at s = oo. This ka = Oand the set of tim
loop is equivalent to again of infinity. Similarly, the combined loop which linear time-invariant cal
passes the point J has anet loop gain 1 at s = oo. The transfer function from
Uz to yz through the loop is (s + 2)/1, which is improper. For the system in y(k) =
Figure 3-25, wehave det [1 +GkD)G¡(OO)] =0.
In conclusion; a system is well posed if and onlyif the system has nocom
bined loop with net loop gain 1 at s = 00, which is equivalentto an infinite gain 14 This section maybe skippe<
D/SCRETE-TIME SYSTEMS 121
The inputs and outputs of the systems we studied in the previous sections are
defined for all t in ( - 00, 00) for the time-varying case or in [0, ro) for the time
(b) invariant case. They are called continuous-time systems. In this section we shall
study a different class of systems, called discrete-time systems. The inputs and
outputs of discrete-time systems are defined only at discrete instants of time.
For example, a digital computer reads and prints out data that are the values
of variables at discrete instants of time; hence it is a discrete-time system. A
continuous-time system can also be modeled as a discrete-time system if its
responses are of interest or measurable only at certain instants of time.
For convenience, the discrete instants of time at which the input and the
output appear will be assumed to be equally spaced by an amount of T The
(d)
time interval T is called the sampling periodo We use {u(k) ~ u(kT)} and
{y(k) ~ y(kT)}, k =0, ± 1, ±2, ... , to denote the input and output sequences.
temo (e), (d) Equivalent In the following, we discuss only single-variable discrete-time systems. As in
the continuous-time system, a discrete-time system that is initially relaxed is
called a relaxed discrete-time system. lf the inputs and outputs of a relaxed
lonzero loop gains at diserete-time system satisfy the linearity property, then they can be related by
)wn in Figure 3-28(a), .00
{~
) gain 3 - 2 = 1; hence i=m
u(i) =
ifm
; a combined loop with
nt to a gain of infinity with the system relaxed at time m-o lf the discrete-time system is causal
,es the problem is the that is, the output does not depend on the future values of the input-then
g(le, m) = O for le < m. Consequentíy, if a discrete-time system is causal and is
;: consider the system relaxed at ko, then (3-76) reduces to
k
through the point H.
L loop between points
y(k) = L g(k, m)u(m) (3-77)
m~ko
ned loop, the transfer
ause of the two strictly lf a linear causal relaxed discrete-time system is time invariant, then we have
g(k, m) = g(k - m) for all k ~m. In this case, the initial time is ehosen to be
Benee, the combined
gain 1 at s = ro. This °
k o = and the set of time of interest is the set of positive integers. Bence for a
linear time-invariant causal relaxed discrete-timesystem, we have
combined loop which
k
transfer function from
;:r. F or the system in y(k) = ¿ g(k - m)u(m) k =0, 1, 2, ... (3-78)
m~O
Example 2 Then
lf u(k) =e- 2k for k =0,1,2, ... , then
00 ~[x(k+
z
u(z) = L e- 2kz- k I
k=O
Now we shall apply the z-transform to (3-78). For a causal, relaxed system, Hence the application
we have g(k - m) = O for k < m; hence we may also write (3-78) as
00
Consequently, we have
00 00 00
x(z)
y(z) = L y(k)Z-k = L L g(k - m)u(m)z-(k-lII l z -111
= (
that in diserete-time omitted. We introduce now the diserete-time dynamical equation. A linear,
rwise, aH the concepts time-varying, discrete-time dynamical equation is defined as
~he Laplace transform,
DE: x(k + 1) = A(k)x(k) + B(k)u(k)
;ebraie equation. We y(k) = C(k)x(k) + E(k)u(k)
will be used is caBed
where x is the state vector, u the input, and y the output. Note that a diserete
time dynamieal equation is a set of first-order difference equations instead of a
set of first-order differential equations, as in the continuous-time case.
~fined as If A(k), B(k), C(k), and E(k) are independent of k, then DE reduces to
{x(k)}k'=o
That is,
• x(z) ~ .?l'[x(k)J ~
<X)
¿ X(k)Z-k
k=O
Then
<X) <X)
3-3 The impulse response of a relaxed linear system is found to be g(t, r) = e- h -d for al!
mt-output description t and r. Is this system causal? Is it time invariant?
Although the input
Iso be developed from 3-4 The impulse response of an ideallow-pass filter is given by
without knowing the sin 2w(t - to)
g(t)=2w--- for aH t
tt systems, the transfer 2w(t - to)
er functions (including where w and t o are constants. Is the ideallow-pass filter causal? ls it possible to build an
used, the systems are ideal low-pass filter in lhe real world?
tice. This fact is often used in stability studies of nonlinear feedback systems.
y y y
differential equations
The reasons are as
be written as a set of
to describe tirst-order _ _ _-,//c-,-- u
----:*"------u --07"''''---+:,----~ u
r differential equations
uter.
llations,one may also
Figúre P3-1
126 MATHEMATICAL DF5CRIPTIONS OF SYSTEMS
3-8 Consider a linear system with input u and output y. Three experiments are per 3-15 Consider a multi'
formed on this system using the inputs u¡(t), uz(t), and U3(t) for t ;::::0. In each case, the the ijth element of (;(s) e
initial state at t =0, x(O), is the same. The corresponding observed outputs are y¡(t), h(t),
and Y3(t). Which of the following three predictions are true if X(O) 1= O?
a. lfu 3 =u¡ +uz, then Y3 = Y¡ +Yz· where Yi is the ith compe
b. lfU3=1(u¡ +uz), thenY3 =1(Y¡ +Yz).
c. lfu 3 =u¡ -uz, then Y3 = Y¡ - Yz· 3-16 Consider a multiv
Which are true if x(O) = O? (Answers: No, yes, no, for x(O) 1= O; all yes, if x(O) = O.)
3-9 Show that if H(u¡ + u z ) = Hu¡ + Hu z for any u¡, u z, then Hrxu =rxHu for any rational
number rx and any u. where the N;]s and D;]s:
of the system?
3-10 Show that for a fixed rx, the shifting operator Q. defined in Figure 3-5 is a linear time
invariant system. What is its impulse response? What is its transfer function? Is this 3-17 Find the dynamic
transfer function a rational function? lf 8,8 ¡, and 8 z are very srr
functions and dynamical
3-11 The causality of a relaxed system may also be defined as follows: A relaxed system
is causal if and only if u¡(t) = uz(t) for all 15,10 implies (Hu¡)(t) = (Huz)(t) for aH 15,1 0 , 3-18 Find the dynamic
Show that this definition implies that y(t) = Hu(_ 00,1]' and vice versa. and 3-23.
3-12 Let g(t, L)=g(t +rx, L +rx) for all t, L, and rx. Define x =t +L,y=t -L, then g(t, L)= 3-19 Find the dynamic:
g«x + y)/2, (x - y)/2). Show that ag(t, L)/aX = O. [From this fact we may conclude that if network in Figure P3-19.
g(t, L) = g(t + rx, L + rx) for all t, L, and rx, then g(t, L) depends only on t - L.]
y(t)= Ig(t-L)U(L)dL \ 1
,,(,,~
If the impulse response 9 is given by Figure P3-13(a), what is the output due to the input
shown in Figure P3-13(b)? (Use graphical method.)
[m g
u(tJ (a)
Figure P3-17
g(tJ
E-
c¡ Cz
o 2
bl-----t-I
R1
L
1 - - - + - 1-----,.-,' t
-1 + t II Z
11 1 rv Rz
o '2 3 4
(a) (b)
ji+2¡i+3y=2ú+u
'ee experiments are per 3-15 Consider a multivariable system that is describable by y(s) = G(s)íI(s). Show that
t 2': O. In each case, the the íjth element of G(s) can be defmed as
,d outputs are Yl(t), yz(t),
O) 1= O?
A
9 ij\S) =
If[Yi(t)]
ro-[
()J
I
iniLially relaxed
.;L Uj t and u,~Ofork-l=j
where Yi is the ith component of y and Uj is the jth component of u.
3-16 Consider a multivariable system whose inputs and outputs are described by
LIl yes, if x(O) = O.) N 11 (P)Yl(t) + N 12(p)Yz(t) = D 11 (p)u 1(t) + D ¡z(p )uz(t)
't, Y = t -'t, then g(t, 't) = 3-19 Find the dynamical-equation description and the transfer-function matrix of the
we may conclude that if network in Figure P3-19.
)ll t - 't.]
(a) (b)
Figure P3-17
(b)
Figure P3-19
128 MATHEMATlCAL DESCRIPTlONS OF SYSTEMS
3-20 Consider the simplified model of an aircraft shown in Figure P3-20. lt is assumed 1F
that the aircraft is dynamically equivalent at the pitched angle eo, elevator angle uo, altitude
ha, and cruising speed vo. lt is assumed that small deviations of e and u from eo and Uo
E-
e
generate forces JI = k I and J2 = k 2u, as shown in the figure. Let m be the mass of the
-I.Q
aircraft, 1 the moment ofinertia about the center of gravity P, be the aerodynamic damping,
and h the deviation of the altitude from h o. Show that the transfer function from u to h Currenl
is, by neglecting the effect of 1, source t
3-21 The soft landing phase of a lunar module descending on the moon can be modeled Figure P3-23
as shown in Figure P3-21. lt is assumed that the thrust generated is proportional to til,
where m is the mass of the module. Then the system can be described by my = - krh - mg,
where 9 is the gravity constant on the lunar surface. Define the state variables of the
where gl(l, t o ) is the step r
system as XI = y, X2 = y, X, = m, and u = rh. Find the dynamical-equation description of
[Hint: This can be proved
lhe system.
(3-55).J
3-22 Show that the output of a linear causal relaxed system due to the input u(t)i5 I(t - to)
3-23 Find the transfer fu
is given by
in Figure P3-23. Do you
y(t) = u(to)gl(t, to) + l'
'o
U(T)gl(t, T) dT for t ~to
3-24 Show that single-va
in the sense that the order
h true for the time-varying s
l
I
Th rus! = kli¡
det(I"
I
~l
IH
le to the input u(t)O¡(t - to) 3-23 Find the transfer function and the dynamical-equation description of the network
in Figure P3-23. Do you think the transfer function is a good description of this system?
t"2:to 3-24 Show that single-variable, linear, time-invariant, causal, relaxed systems commute
in the sense that the order of the tandem connection of two systems is irnmaterial. ls this
true for the time-varying systerns?
3-25 The impulse and step responses of a single-variable, linear tirne-invariant, causal,
relaxed systern are, by deftnítion, given by 9 =Hb(t) and 9 ¡ = Hb¡(t), where b is a delta
d
fllnction and 15¡ is a step function. lt can be shown that b(t) =-b¡(!). Verify (3-56) by
d!
lIsing the property given in Problem 3-24.
3-26 Verify that the impulse-response rnatrix of the feedback systern in Figure 3-24(c) is
given by Equation (3-60).
Note that the rnatrix in theleft-hand side isan.n xn rnatrix~ Let'G be an 11 x 11 matrix of
rank 1, show that det(ln+G)= 1 + trace G, where trace G is thesumof aH diagonal
elements of G. [H Gis not of rank 1, seeEquation (9-19) of Chapter 9.]
130 MATHEMATlCAL DESCRIPTlONS OF SYSTEMS
3-30 Find the transfer-function matrix of the feedback system shown in Figure 3-24(c),
where the transfer-function matrices of SI and Sz are, respectively,
ss +1:2J 2.5
s +2
'8
Y-B
3-31 Find the dynamical-equation description of the feedback system in Figure 3-24(c),
where SI and Sz are, respectively, described by
0.5
[~IIJ=[-2O
XIZ
_IJ[XuJ
1 XIZ
+[_ 4
lJU I
1 2
YI =[0 lJx l +[1 -IJu I
Figure P3-35
and [~::J=[~}z
YZ=[~ _~JXZ
Draw a block diagram of a computer simulation of this feedback system.
3-32 Prove Theorem 3-2 by using det N = det P = 1 and det NQ = det PQ.
Figure P3-37
3-33 Find the overall transfer matrix ofthe feedback system shown in Figure 3-24(c)with
l
-1
'i' J
A s +1 al h(xl(t), uI(t»+
GI(s)= O Gz(s)= _1_
[ for any real numbers a¡, a¿
s +1 linear function ofx(t) and [:
Can you find a combined loop with a net loop gain 1 at s = oo?
for sorne A and B. If h, x,
3-34 Which of the systems in Figure P3-34 have improper overall transfer functions? does the assertion hold? I
3-35 Can you find a combined loop with a net loop gain 1 in the system shown in Fig. 3-37 Consider the multiv
P3-35? What is its overall transfer matrix? following composite lransfe
e(s)J
[ u(s) =
[(1 +
C(s)(
Figure P3-35
+ e
:k system.
Q =detPQ.
Figure P3-37
own in Figure 3-24(c)with
3-36 A function h(x(t), uU)) is said to be a linear function ofxU) and u(t) if and only if
(;( ¡ h(x ¡ (t), U ¡ (t)) + (;(z h(xz(t), uz(t)) = h«(;( ¡ x ¡ (t) + (;(zxz(t), (;( ¡ u 1 (t) + (;( zuz(t))
for any real numbers (;(¡, (;(z, any x¡(t), xz(t), and any u¡(t), uz(t). Show that h(x(t), u(t)) is a
linear function ofx(t) and u(t) ifand only ifh is of the form
h(x(e), u(e)) = A(e)x(e) + B(e)u(e)
for sorne A and B. If h, x, and u are square matrix functions instead of vector functions,
erall transfer functions? does the assertion hold? If not, what modification do you need?
3-37 Consider the multivariable feedback system shown in Figure P3-37. Verify the
\ the system shown in Fig.
following composite transfer matrix:
e(s)] [(1 + (;(s)C(s)t ¡ - (;(s)(1 +C(S)(;(S))-¡][f(S)]
[ u(s) = C(s)(1 + (;(s)C(s))- 1 (1 +C(s)G(s))-¡ 3(s)
3-38 The transfer matrix of the system in Figure 3-26 is given in (3-73). Is it possible to
have a different form of (; ¡(s) with (;¡(s) or (;z(s) on the right-hand side of the parentheses
as in Equation (3-68)?
3-39 Show that the system in Figure 3-28(b) is not well posed by finding an improper
transfer function in the system.
3-40 Show that the parallel andtandem connections in Fi6ure 3-24(a) and (b) are wel1
posed if and only if (; ¡(s) and (;z(s) are proper.
132 MATHEMATlCAL DESCRIPTIONS OF SYSTEMS
3-41 Let (;(s) = G o +G1s- 1 +G 2s- 2 + ... and C(S) =C o +CIS- I +C 2S- 2 +.... Show
that (1 +C(s)(;(s)) -1 is proper if and only if (1 +C( <X) )(;( co)) = (1 +CoG o) is nonsingular.
Prove it directly by using the power series without using Theorem 3-5.
3-42 A rational matrix (;(s) is proper if (;(00) is a finite constant matrix and improper
if (;( (0) is not a finite constant matrix. Are the following statements valid?
1. Ir (;I(S) and (;2(S) are proper, (;2(S)(;I(S) is proper.
2. If (;I(S) and (;2(S) are improper, (;2(S)(;I(S) is impropero
Answer: Yes; no. Consider (;2(S)(;¡(S) = (;2(S)U(s)U- 1(S)(;I(S). Let (;¡(s) be 2 x 2 proper
rational matrices and let l
S" +1
U(s) = [ 1
3-43 Let (;(s) be a q x p rational matrix, not necessarily proper. Show that the rational
matrix
(;I(s) = (;(s)(I + K(;(s)r 1
is proper for almost aH p x ~ const~nt matrix K (or, in mathematical terminology, for a
generic K). [Hint: Express G(s) as G(s) = N(s)D-1(s), where N(s) and D(s) are polynomial
matrices, and use Theorem G-9. See Reference SI96.]
by direct compuíation c·
the equation Y(s) = G(s
matrix G(s) is a rationa
lS-1+C 2S- 2 + .... Show
,(1 +CoG o) is nonsingular.
rem 3-5.
uniquelyas
4-1 Introduction
the time domain than from the frequency domain. It is elementary to compute Theorem 4-1
y from the input-output description; hence it will not be discussed further.
Solutions of linear dynamical equations are studied in Section 4-2. Solu The set of all solutio
tions are stated in terms of the state transition matrix <.I>(t, r), which is the over the field of real 1
unique solution of
Proof
o
ot <.I>(t, r) = A(t)<l>(t, r) <.I>(r,r)=I Let WI and W2 be tW(
solution of (4-2) for al
In the time-invariant case, we have <D(t, r) = eA(!-t). Various methods for the d
computation of eAI and (sI - A)-I are discussed. In Section 4-3 the concept of dt (ex l W¡ +ex
equivalent dynamical equations is introduced. Equivalent dynamical equa
tions are obtained by changing the basis of the state space. We show that
every time-varying linear dynamical equation has an equivalent linear dy
namical equation with a constant A matrix. We al~o establish the theory Hence the set of solut
space of (4-2). We n
01 Floquet. In the last section, the relation between linear dynamical equations
el' e 2, ... , en be any li
and impulse-response matrices is studied. The necessary and sufficient con
tions of (4-2) with th
dition for an impulse-response matrix to be realizable by a linear dynamical
equation is established. We also show that every proper rational matrix has a show that W¡, for i = L
linear time-invariant dynamical-equation realization. of(4-2) can be written
The references for this chapter are 24,31,60,68, 77, 109, 114, and 116. assertion is proved. \
independent. Suppos.
definition, there exists
Solutions of X =A(t)x
1t is clear that
The set of first-order differential equations in (4-2) has a unique solution for
every initial state X o in (W,~). Since there are infinitely many possible initiaJ
states, Equation (4~2) has infinitely many possible solutions. This set of SOlU7
tions forms a linear spaceover~. There are only n linearly independent initial is a solution 0((4-2) wi
states in(W, ~); hence the linear space is of dimension n. This fact will be
formally established in the following theorem.
\TRICES SüLUTIONS OF A DYNAMICAL EQUATION 135
Note that the O in the right-hand side of (4-3) is the zero vector of the solution
. linear time-varying space; therefore, it is more informative to write (4-3) as
[Vl(t) V2(t) Vn(t)]ot=O for a11 t in (-00,00)
uation) (4-1 a) In particular, we have
:quation) (4-1 b)
:l q x p matrices whose
which implies that ei, for i = 1, 2, ... , n, are linearly dependent. This contra
over ( - 00, 00). Since
dicts the hypothesis; hence \ji" for i = 1, 2, ", n, are Ijnearly i.ndependenl ,,'!"'"r
) and any VI, there exists (-00,00).
t will be used frequently
Let V be any solution of (4-2), and let V(to) =e. Since el' e 2, ... , en are n
lre dynamical equation
linearly independent vectors in the n-dimensional vector space (~m ~), e can
of E; namely,
be wriUen as a unique linear combination of ei , for i = 1, 2, ... , n-for example,
(4-2) as n
e= ¿ IX¡e i
¡= 1
It is clear that
i = 1
:ions. This set of solu
¿ IXiV¡(tO) =e
i= 1
\
I
136 LINEAR OYNAMICAL EQUATIONS ANO IMPULSE-RESPONSE MATRICES
\jiU = ¿"
Before we prove the th
1X¡\jIk)
i;:;; 1 *= A(t)x and if \jI(t o) =
that is, \jI(') == O. It is ob
This completes the proof that the solutions of (4-2) form an n-dimensional
Again, froro the unique
vector space. Q.E. D.
solution with \jI(t o) =0.
We shall now pro"
Definition 4-1 that det 'JI(to) = det [~
*
An n x n matrix function 'P is said to be a fundamental matrix of = A(t)x if
set of n constant col um¡
(IR", IR). It follows that
and only if the n columns of'P consist of n linearly independent solutions of
*= A(t)x. I
*=[~ ~Jx
is a solution of * = A(t)
: that Proof
Before we prove the theorem, we need the following fact: Ir '!lO is a solution of
x = A(t)x and if '!ICto) = O for sorne to, then the solution '!lO is identicaHy zero;
that is, '!lO =:= O. It is obvious that '!lO =:= Ois a solution ofx = A(t)x with '!ICto) = O.
Jrm an n-dimensional Again, from the uniqueness ofthe solution, we conclude that '!lO =:= O is the only
Q.E.D. solution with '!I(to) =0.
We shaH now prove the theorem; we prove it by contradiction. Suppose
that det 'PCto) = det ['!IICtO) '!IzCto) .. , '!InCtO)] =0 for sorne too Then the
set ofn constant column vectors '!I1(t O), '!IlCtO)"'" '!InCtO) is linearly dependent in
! matrix of x = A(t)x if
(~n, ~). It follows that there exist real IX¡, for i = 1,2, ... , n, not aH zero, such that
iependent solutions of
n
I
¿ IXi'!l¡CtO) = O
", "
i= 1
tXI' Now two linearly This contradicts the assumption that '!Ik), for i = 1, 2, ... , n, are linearly inde
:an be easily obtained pendent. Hence, we conclude that det 'P(t) 1=0 for aH t in (- 00, (0). Q.ED.
"zCto) =0. Hence, the
Definition 4-2
Let 'PO be any fundamental matrix of x = A(t)x. Then
for all t, t o in (-00, (0)
I
is said to be the state transition matrix of x = A(t)x. I
.tial equation x = ACt)X;
The physical meaning of<l>Ct, to) will be seen latero Since 'P(t) is nonsingular
for aH t, its inverse ís well defined for each t. From the definition we have
(4-Ll. )
ímmediateíy the following very important properties ol the state transition
ant matrix. Conversely, matrix:
r sorne t, then from the <l>(t, t) = 1 (4-5)
VI are linearly indepen <l> - l(t, to) = 'PCto)'P - 1 (t) = <l>Cto, t) (4-6)
lamental matrix. Thus <l>(t z, to) = <l>Ctz, t ¡}<l>(t 1> to) (4-7)
latrix ofx = A(t)x if and
for any t, to, ti' and t z in ( - 00,(0).
~.) is that the inverse of Note that <l>(t, to) is uniquely determined by A(t) and is índependent of the
following theorem. particular 'P chosen. Let 'P 1 and 'P z be two different fundamental matrices of
x = A(t)x. Sínce the columns of 'P l' as well as the columns of 'P 2' qualify as
basis vectors, there exists, as shown in Section 2-3, a nonsingtilar real constant
matrix P such that 'P z = 'P 1 P. In fact, the ith column of P is thereprese~tation
(- 00,00). of the ith column of 'P 1 with respect to the basis that consists of the columns of
... .... ... ..... ==..-'....::c.....-
which shows the uniqueness of <I>(t, to). From Equation (4-4), it is evident that Theorem 4-3
<I>(t, to) is the unique solution of the matrix equation
The solution of the Sl
(4-8)
is given by
with the initial condition <I>(t o, to) = l.
Remarks are in order concerning the solutions of (4-4) and (4-8). Ir A(t)
is a continuous function of t, then <I>(t, to) and q¡(t) are continuously differenti
able 1 in t. More generally, if A(t) is n times continuously differentiable in t,
then <I>(t, to) and q¡(t) are n + 1 times continuously differential in t; see References
24 and 77.
The computation of the solution of (4-8) in a closed form is generally very where <I>(t, '[) is the s
difficult, if not impossible, except for sorne special cases. Ir A(t) is a triangular unique solution of
matrix, then its solution can be reduced to solving a set of scalar differential
equations, and its closed-form solution can be readily obtained (Problem 4-1).
If A(t) has the following commutative property
Proof
Equation (4-11) is ob
first show that (4-10) ~
for all t and lo, then the unique solution of (4-8) is given by
d a
dl X(l) = al <1>(
<I>(l, to)=exp [ ( A('[)d'[] (4-9)
=A(l)q
(Problem 4-31). Ir A(t) is a diagonal matrix or a constant matrix, then it meets
the commutative property, and its transition matrix is given by (4-9). For other
special cases, see Problems 4-14 and 4-15. See also References S229 to S231.
From the concept of state transition matrix, the solution of Ji = A(l)x
=A(l) I
follows immediately. To be more informative, we use ,p(t; t o, 1<0' u) lo denote
the solution of x = A(l)x at time l due to the initial condition x(lo) =x o. The At l = lo, we have
fourth argument of cP denotes the fact that u == O. The solution of Ji = A(l)x
with x(lo) =x o is given by
X(l)~ cP(t; lo, x o, O) = <I>(l, lo)xo
In other words, (4-10)
which can be verified by direct substitution. The physical meaning ofthe state
. trarisition matrix <D(l, lo) isnow clear. It governs the motion of the state vector
in the time interval in which the input is identically zero. <I>(l, lo) is a linear
We consider again
transformation that maps the· state X o at lo ¡nto the state x at time l. (4-10) reduces to .
lA function is said to be continuous/y differentlab/e if its first derivative exists and is conUnuous. 2
ata J''o f(l, e) de = f(l, e)I,=,
TRICES SOLUTlONS OF A DYNAMICAL EQUATION 139
We use cP(t; to, x o, u) to denote the state resulted at time t due to the initial state
I(,on- ' x(to) =x o and the appUcation of the input u.
Is1y differentiable in t,
\:ial in t; see References
= <I>(t, to{x o + L <I>(to, r)B(r)u(r) dr] (4-11 )
Proof
I
1) Equation (4-11) is obtained from (4-10) by using <I>(t, r) = <I>(t, t o)<1>(to, r). We
first shüw that (4-10) satisfies the state equation by direct substitution 2 :
[bY
el
-x(t) a <I>(t, to)xo +a
=- l' <I>(t, r)B(r)u(r) dr
(4-9) dt at at 'o
,a
=A(t)<1>(t, to)xo +<I>(t, t)B(t)u(t) +
1 -<I>(t, r)B(r)u(r) d!
'o at
t
J matrix, then it meets
ven by (4-9). For other
= A(t) [ <I>(t,to)x o + <I>(t, r)B(r)u(r) dr] +B(t)u(t)
ences S229 to S231.
solution of x = A(t)x
I(t; to, X o, O) to denote
ldition x(to) =x o. The At t = t o, we have
s01ution of x= A(t)x '0
x(to) = <I>(t o, to)xo -r
1
'o
<I>(to, r)B(r)u(r) dr = Ix o +0 =X o
In other words, (4-10) a1so meets the initia1 condition. Hence it is the s01ution.
11 meaning of the state Q.E.D.
ltion of the state vector
We consider again the so1ution given by Eqtiation (4-10). U u == O, then
I~o. <I>(t, to) is a linear (4-10) reduces to .
x at time t.
(4-12)
tr <I>(t, r)B('r)u(r) dr
cjl(t; to, O, u) = (4-13) zero-input response.
Equatian (4-16) beco
For obvious reasons, cjl(t; to, x o, O) is called the zero-input response, and
cjl(t; to, O, u) is called the zero-state response of the state equation. It is clear that y(t)
cjl(t; to, x o, O) and cjl(t; to, O, u) are linear functions of X o and u, respective1y.
Using (4-12) and (4-13), the solution given by Equation (4-10) can be written as
cjl(t; to, xo, u) = cjl(t; to, x o, O) +cjl(t; to, O, u) (4-14)
The matrix function
This is a very important property; it says that the response 01a linear state equa
tion can always be decomposed into the zero-state response and the zero-input
response. This is consistent with Equation (3-31).
is called the impulse-¡
Note that Equation (4-13) can be derived directly from the fact that it is a
the input-output re1a1
linear function of u. The procedure is exactly the same as the one in deriving
We see from (4-H
t
r G(t, r)u(r) dr
Jto
the solution of the d:
is the unique solution
in Section 3-2. The response cjl(t; to, O, u) is, by definition, the solution of
x= A(t)x +B(t)u with O as the initial state. If we cut the input u into small
pulses, say
Unfortunately, there
and except for very :
u= L u[t¡,t¡+Ó) easily faund. Theref
theoretical study af lil
then we have
of a dynamical equati
(4-15) <I>(t, r). The solution (
digital camputer.
where we have used the fact that if ~ is very sma11, the solution of x= A(t)x
+ B(t)u due to the input u[t¡,t¡H) with Oas the initial state is approximately equal
Time-invariant ca:
to B(t¡)u(t¡)~. The input u[t¡,t¡H) outside the time interval [ti, ti +~) is identica11y
(fixed) dynamical equ.
zero; hence, the response between t¡ +~ and t is governed by <I>(t, ti +~).
Summing up (4-15) for a11 i and taking the limit ~-+O, we immediately obtain
the equation
t
tr <I>(t, r)B(r)u(r) dr
cjl(t; to, O, u) = where A, B, e, and E
respectively. Since th
We give now the solution of the entire dynamical equation E. dynamical equatian E
applied here. We ha,
Corollary 4-3
The solution of the dynamical equation E in (4-1) is given by
t
y(t) = C(t)<D(t, to)xo +C(t) r $(t, r)B(r)u(r) dr + E(t)u(t)
t I and e At is nonsingula
In fact, e At is nonsingl
•
tion matrix of x = Ax
rRICES SOLUTIONS OF A DYNAMICAL EQUATlON 1411
By substituting (4-10) and (4-11) into (4-1 b), we immediately obtain Corollary
4-3. The output y can also be decomposed into the zero-state response and the
(4-13) zero-input response. If the dynamical eguation is initially in the zero state,
Eguation (4-16) becomes
¡-input response, and t
1 by
It follows from (4-10) that the solution of (4-19a) is We may apply them
t
cP(t;to,xo,u)=eA(t-to)xo+ r eA(t-<JBu(,)d, (4-20)
1. Using Definition
t polynomial g(A) o
then e At = g(A).
If t o = O, as is usually assumed in the time-invariant equation, then we have the
fol1owing theorem. 2. Using the Jordan (
where Á is of the J,
(2-70).
Theorem 4-4
3. Using the infinite
The solution ofthe linear time-invariant dynamical equation FE given in (4-19) give a closed-form
is tation.
We introduce one me
(4-21 )
we have
and
Hence, to compute l
(4-22) inverse Laplace tran
inverse of a matrix i¡
I
angular 4 or of order l
the inverse of a trian~
The impulse response matrix of FE is
Note that(sl-A)
G(t, ,) = G(t -e) =CeA(t-<JB + El5(t -e) many methods to con
1. Taking the inverse
or, as more commonly written,
2. Using Definition 2
G(t) = CeAtB + El5(t) (4-23) 3. Using (si - A)-l =
4. Using Definition 2
The solution of a linear time-invariant dynamical equation can also be 5. Taking the Laplacl
computed in the frequency domain. Taking the Laplace transform of (4-21)
In addition, there is a
and (4-22), and using g' [e Al ] = (si - At 1 [see Eq uation (2-86)], we obtain
x(s) = (si - A)-lX(O) +(sl - A)-l Bú(s) (4-24)
Example 2
We use methods 1 anl
and y(s) = C(sl - A)-lX(O) +C(sl - A)-l Bn(s) + En(s) (4-25)
S +
is cal1ed the transfer-function matrix of the dynamical equation FE. It is the
Laplace transform ofthe impulse-response matrix given in (4-23). The transfer
function matrix governs the zero-state response of the equation FE.
= (s ;J
We give now sorne remarks concerriing the computation of e Al . 3 We [
(s +J
introduced in Section 2-7 three methods of computing functions of a matrix.
(4-21 )
We introduce one more method of computing eA'. Since se [e
At
] = (sI - A)-1 ,
r) dr we have
eAt = -1(sI - At 1 se (4-27)
Hence, to compute eA!, we first invert the matrix (sI - A) and then take the
u(t) (4-22) inverse Laplace transform of each element of (sI - A)-I. Computing the
inverse of a matrix is generally not an easy jobo However, if a matrix is tri
I
angular 4 or of order less than 4, its inverse can be easily computed. Note that
the inverse of a triangular matrix is again a triangular matrix.
Note that (sl- A)-1 is a function of the matrix A; therefore, again we have
r) many methods to compute it:
1. Taking the inverse of (sI - A).
2. Using Definition 2-16.
(4-23) 3. Using (sI - A)-1 = Q(sI - Á)-IQ -1 and (2-74).
4. Using Definition 2-17.
~quation can also be 5. Taking the Laplace transform of eA!.
;e transform of (4-21)
In addition, there is an iterative scheme to compute (sI - A) - 1 (Problem 2-39).
(2-86)], we obtain
Example 2
(4-24 )
We use methods 1 and 2 to compute (sI - A)-I, where
+Eu(s) (4-25)
. 10 -1l
of a variabie. These Fl=Ll -2J
the dynamical equa
(4-26)
1. (sI-A)-1 =[ - 1 s
s
+2
1 J-l = + S2
1
2s +1
[s+2
1
-sI]
luation FE. lt is the
1 (4-23).The transfer
uation FE.
4 A square matrix is said to be triangular if al! the elements below or aboye the main diagonal are zero.
144 LINEAR DYNAMICAL EQUATIONS AND IMPULSE-RESPONSE MATRICES
2. The eigenvalues of A are -1, -1. Let g(A) =0:0 +0: 1,1. Iff(A)~(S-A)-l
= g(A) on the spectrum of A, then
8
f( -1) =g( -1): (s +1)-1 =0:0 -0: 1
j'(-1)=g'(-l): (s +1)-2=0: 1
Hence g(A) = [(s + 1)- 1 +(S + 1)- 2J +(S + 1)- 2A
and (si - A)-I = g(A) = [(s + 1)- 1 +(s + 1)-2JI +(s + 1)- 2A
(s~w
tñ; -1.
[
If aH eigenval ues
can be written as, by
and
)=L1(1 +t)e-
t
x(t te -/
where q¡ and Pi are, I
The sol utions with u == O are plotted in Figure 4-1 for the initial states with A¡. In the time dI
x(O) = [8 8]' and x(O) = [8 5]'. Note that the velocity at each point of the
trajectories in Figure 4-1 is equal to Ax. •
If the matrix A has m distinct eigenvalues A¡ with index ni, for i = 1, 2, ... , m If Xo ís chosen so tha
(see Definition 2"15),5 we claim that every element of eA! is a linear combination
ofthe factors tkeA¡t, for k = 0,1, ... ,ni - 1; i = 1,2, ... , m. Let Á be a Jordan-form
representation of A and A = QÁQ - 1. Then eA! = QeA!Q - l. From (2-69) we F or this il1itial state, '
direction of the eigen
formula: similar to (4
5If aH eigenvalues are distinct, then the index of every eigenvalue is 1. If an eigenvalue has mullí-o eigenvectors of A. 1
plicily k, then ilsindex could be 1,2, ... , k -1, or k. See Seclíon 2-6 and Definition 2-15. discussed.
'RICES SOLUTIONS Of A DYNAMICAL EQUATlON 145
A. If f(A)6 (5 - A)-I
-----!L--+---+--+--+-+--+--1--1--- xl
o 4 8
know that every element of e Al is ofthe form tkeA¡t, for k =0,1, ... , ñ¡ - 1; i = 1,
2, ... ,m. Hence every element of e At is a linear combination of these factors.
We call tkeA¡t a mode of the dynamical equation FE in (4-19).
The responses of a linear time-invariant dynamical equation are dictated
(4-28) mainly by its modes, or equivalently, the eigenvalues of A. If an eigenvalue has
a negative real part, its mode will approach zero exponentially as t --> 00; it
rse Laplace transform approaches zero ·either monotonically or oscillatorily depending on whether
its imaginary part is zero or not. If an eigenvalue has a positive real part, its
mode will approach infinity exponentially as t --> oo. If an eigenvalue has a
zero real part and has index 1,5 its mode is a constant or a pure sinusoidal; if
-te-
(l-t)e- t
t
J its index ñ¡ is 2 or higher, then its mode will approach infinity at the rate of
tñ¡-I.
If all eigenvalues of A are distinct, the response of x = Ax due to x(O) =x o
can be written as, by using Problem 2-44,
'e
-(t-r) l x(s) = (sI - A)
A -1"
o = L,
X --1
1 Qi1ll1¡X
o (4-29)
')Je -(t - rl J U(T) dT i s -- /"'i
where q¡ and p¡ are, respectively, a right and a left eigenvector of A associated
for the initial states with Aj • In the time domain, (4-29) becomes
I at each point of the
• (4-30)
{ ñ¡, for i = 1,2, ... , m lf Xo is chosen so that p¡XO =0 for all iexcept i = j, then (4-30) reduces to
; a linear combination
,et Á. be a lordan-form x(t) = (p }'o)e Ajt qj
r l. From (2-69) we For this initial state, only the mode e Ajt is excited and x(t) will travel along the
direction of the eigenvector qj. lf A has eigenvalues with indices 2 or higher,a
formula similar to (4-29) can be derived by using (2-74) and the genera:Jized
Ir an eigenvalue has multi eigenvectors of A. The situationis much more complicated and will not be
and Definition 2-15. discussed.
146 LINEAR DYNAMICAL EQUATIONS AND IMPULSE-RESPONSE MATRICES
matrix E is the direct transmission part between the input and the output.
Since it has nothing to do with the state space, it is not affected by any equiva
iynamical equations. lence transformation.
,ne of change ofbasis We explore now the physicaL meaning of equivalent dynamical equations.
variant case and then RecaIl that the state of a system is an auxiliary quantity illtroduced to give a
unique relation between the input and the output when the system is not
initially relaxed. The choice of the state is not unique; different methods of
ant (fixed) dynamical analysis often lead to different choices of the state.
Example 1
(4-31 a)
(4-31b) Consider the network shown in Figure 4-2. If the current passing through the
inductor Xl and the voltage across the capacitor Xl are chosen as the state vari
Id q x p real constant
ables, then the dynamical equation description of the network is
t vector, and x is the
lation is an n-dimen
(4-34a)
r. We have agreed in
... , "ti} as the basis
Lat its ith component (4-34b)
; the basis of the state
~ the basis of the state If, instead, the loop currents Xl and Xl are chosen as the state variables, then
r to aIlow a broader the dynamical equation is
ction extend the field
der the state space as (4-35a)
ization is needed in
ent Jordan-form dy (435b)
The dynamical equations in (4-34) and (4-35) have the same dimension and
describe the same system. Hence they are equivalent. The equivalence trans
the field of complex formation between these two equations can be found from Figure 4-2. 1t is
I clear that Xl =Xl' Since Xl is equal to the voltage across the l-ohm resistor,
we have Xl =(Xl -Xl)' Thus
(4-32a)
(4-32b)
(4-33)
(4-36)
f-31), and P is said to or
•
4-31) by the substitu
~ basis vectors of the
'p-I (see Figure 2-5). Xl
erent representations
.. qtl], then the ¡th
basis {ql, qz, ... , qtl}.
qtl]B, we see that
1 of B with respect to
ed from Cp-l. The Figure 4-2 A network with two different choices of state variables.
148 LINEAR DYNAMICAL EQUATIONS AND IMPULSE-RESPONSE MATRICES
It is easy to verify that the dynamical equations in (4-34) and (4-35) are indeed Example 2
related by the equivalence transformation (4-36). 11 Consider the two net
capacitor is zero, the
Definition 4-4 impulse responses are
Two linear dynamical equations are said to be zero-state equivalent if and only if are zero-state equivale
they have the same impulse-response matrix or the same transfer-function tions are zero-state eq
matrix. Two linear dynamical equations are said to be zero-input equivalent any initial voltage in 1
if and only if for any initial state in one equation, there exists a state in the other u = O are all identicall~
equation, and vice versa, such that the outputs of the two equations due to A necessary condit
zero input are identical. 11 they have the same d
network in Figure 4-31
Note that this definition is applicable to linear time-invariant as well as ofthe network in FigUl
linear time-varying dynamical equations.
Two equivalent linear time-invariant dynamical equations are zero-state equi The two networks in ]
valent and zero-input equivalent. descriptions are zero-¡
if there is a nonzero ir
Proof of Figure 4-4(b) is no]
is identically zero.
The impulse-response matrix of FE is
G(t) = CeA1B + Eb(t) lfthe matrix A in a
tion is said to be a lord
The impulse-response matrix of FE is
G(t) = CeA1Ü + Eb(t)
If the equations FE and FE are equivalent, we have Á = PAP- 1 ,
C = CP-l, and E =E. Consequently we have
and
eA' = PeA'P- 1
CeA'B + Eb(t) = Cp- 1 PeA'p- 1 PB + Eb(t) = CeAtB + Eb(t)
B = PB,
u t 1n
1 y u(
ns are zero-state equi- The two networks in Figure 4-4 or, more precisely, their dynamical-equation
descriptions are zero-state equivalent but not zero-input equivalent. Indeed,
if there is a nonzero initial condition in the capacitor, the zero-input response
of Figure 4-4(b) is nonzero, whereas the zero-input response of Figure 4-4(a)
is identically zero. •
Ifthe matrix A in a dynamical equation is in the Jordan form, then the equa
tion is said to be a Jordan-formdynamicaJ equation. We have shown in Section
+ Eb(t)
u
t In
1ty u
In
+
x
e
In)'
y
~
AIB In In
e equivalent. -'-- ---"'~-
-
---
(a) (b)
Figure 4-3 Two networks whose dynamical equations are zero-state equivalent and
zero-input equivalent without being equivalent.
o)
2-6 that every operator that maps (Cm C) into itself has a lordan-form matrix This can be written as
representation. Hence every linear, time-invariant dynamical equation has an
E +CBS-l +CABs- 2
equivalent Jordan1orm-dynamical equation.
5]-1 = [1
3 O
O
1
-5]
-3
1 O O 1 where A, B, C, and E al
continuous functions o
then the new A matrix will be in the lordan formo The method for finding the
matrix P was discussed in Section 2-6. If we substitute x = Px into (4-37), we Definition 4-5 6
obtain immediately the following equivalent lordan-form dynamical equation:
Let P(·) be an n x n rr
E:
y = C(t)x + E(t)u
(4-38a)
(4-38b)
1
where A, B, C, and E are n x n, n x p, q x n, and q x p matrices whose entries are
continuous functions of t.
: method for finding the
e x = Px into (4-37), we Definition 4.5 6
·m dynamical equation:
Let PO be an n x n matrix defined over (- ro, ro). It is assumed that P(t)
and P(t) are nonsingular and continuous for aH t. Let x = P(t)x. Then the
dynamical equation
E: x=Á(t)x+"8(t)u (4-39a)
y = qt)x + lt(t)u (4-39b)
,E} and {A, B, C, E}, not The dynamical equation E in Equation (4-39) is obtained from (4-38)
'alent or have the same
x
by the substitution of x = P(t)x and = P(t)x + P(t)x. Let'JI be a fundamental
matrix of E. Then we claim that
q¡(t)~ P(t)'JI(t) (4-41 )
matrix of E; hence 'iJ(t) = A(t)'P(t), and 'P(t) is nonsingular for all t. Conse Definition 4-6
quently, the matrix P(t)'P(t) is nonsingular for aH t (see Theorem 2-7). Now
x
we show that P(t)'P(t) satisfies the matrix equation =A(t)x. Indeed, A matrix P(·) is calle,
tinuous and bounded
d • •
di (P(t)'P(t)) =P(t)'P(t) + P(t)'P(t)
= CP(t) + P(t)A(t))P -l(t)P(t)'P(t)
= A(t)(P(t)'P(t))
Because of (4-43) a
Hence, P(t)'P(t) is a fundamental matrix of x= Ax. of P- 1(t) = - P-l(t)PI
Consequently, we car
Theorem 4-7 7 P- 1 (.). Clearly a nor
Let Ao be an arbitrary constant matrix. Then the dynamical equation in (4-38) In Definition 4-5, if P(
is equivalent to the one in (4-39) with A(t) = A o. are said to be equivale
preserves, as will be (
Proof namical equation, bu
required to be a Lya
Let 'P(t) be an fundamental matrix of x = A(t)x. That is, 'P(t) is nonsingular for
general. In other wo
~1I t and satisfies 'iJ(t) = A(t)'P(t). The differen~iation of '11 - 1 (t)'PJt) = 1 yields equivalent in the sense
'P- 1(t)'P(t) + '11 - 1 (t)'P(t) = O, which implies '11 -l(t) = - 'P- 1(t)'P(t)'P- 1(t) =
This is possible for the
- 'P- 1(t)A(t). We define, in view of (4-41),
be discussed in the foL
P(t) = eAo1'P - 1(t)
Linear time-varying e
Clearly P(t) is nonsingular and continuously differentiable for aH t and qualifies
as an equivalence transformation. We compute Consider the linear tir
A(t) = (P(t)A(t) + P(t))P- 1(t)
= (e A01'P- 1(t)A(t) + AoeAo1'P - l(t) +eAol'iJ-1(t))'P(t)e - Aot
for aH t and for sorne
=A o A(-) is a periodic funct
This establishes the theorem. Q.E.D. matrix of x = A(t)x. 1
Indeed we have
In this theorem, '11(.) and consequently PO are generaHy not known, there
fore nothing is really gained in this transformation. If Ao is chosen as zero, 'P(t .
then P(t) = 'P- 1 (t) and (4-40) becomes The matrix function '1
A(i) = ir» B(t) = q;-l (t)B(i) CU) = qt)\f(t) E(t) = lE(i) (4-42) Colum_l1s of 'f'(f) Etild e
solution space; hence, 1
Its block diagram is plotted in Figure 4-5. Unlike the one in Figure 3-11, there 2-3) such that
is no feedback in Figure 4-5.
7 This theorem was pointed out to the author by Professor T. S. Kuo in 1972. I::or the nonsingular J
AT
e = Q (Problem 2-37
ECt) ~=========;l
Define
+ y
+ We show that PO is a 1
Figure 4-5 Matrix block diagraril 6f the dynamical equation E in (4-42). P(t +T) =e
rRICES EQUlVALENT DYNAMICAL EQUATIONS 153
'P(t) •
.Because of (4-43).and the boundedness of P(t), P- 1(t) is b.ounded. Because
of P - l(t) = - P -1(t)P(t)P- 1(t) and the boundedness of P(t), P -1(t) is bounded.
Consequently, we can show that if PO is a Lyapunov transformation, so is
P -lO. Clear1y a nonsingular constant matrix is a Lyapunov transformation.
ical equation in (4-38) In Definition 4-5, if PO is a Lyapunov transformation, the dynamical equations
are said to be equivalent in the sense of Lyapunov. A Lyapunov transformation
preserves, as will be discussed in Chapter 8, the stability properties of a dy
namical equation, but an equivalence transformation does not. lf PO is
required to be a Lyapunov transformation, Theorem 4-7 does not hold in
P(t) is nonsingular for general. In other words, not every time-varying dynamical equation can be
. 'P- 1(t)'I'(t) =1 yields equivalent in the sense of Lyapunov to a dynamical equation with a constant A.
- 'P- 1(t)'f(t)'I'-1(t) = This is possible for the class of time-varying equations with periodic ACt) as will
be discussed in the following.
~or the nonsingular matrix Q there exists a constant matrix Á such that
n 1972.
eAT = Q (Prob\em 2-37). Hence, (4-44) can be written as
'P(t + T) = 'P(t)e AT (4-45)
Define
P(t)~eAl'P-l(t) (4-46)
where <I>(t, T) is the state transition matrix ofi = A(t)x. By comparing (4-47) and
(4-49), we immediately obtain
(4-38) is periodic with
amical equation E in
G(t, T) = {~(t)<I>(t, T)B(T) + E(t)(j(t - T) for t ¿ T
That G(t, T) = O for t < T fol\ows from the causality assumption which is implic
of Lyapunov. • itly embedded in writing (4-47); that is, the integration is stopped at t.
If the state-variable description of a system is available, the input-output
; hence it is bounded. description of the system can be easily obtained from (4-50). The converse
e, PO is a Lyapunov problem-to find the state-variable description from the input-output descrip
from Theorem 4-7. tion of a system-is much more complicated, however. It actual\y consists of
led theory of Floquet. two problems: (1) Is it possible at al\ to obtain the state-variable description
then its fundamental from the impulse-response matrix of a system? (2) If yes, how do we obtain the
ic function. Further state-variable description from the impulse-response matrix? We shal\ study
x
to =Ái. the first problem in the remainder of this section. The second problem will be
studied in Chapter 6.
¡mical Equations Consider a system with the impulse-response matrix G(t, T). If there exists
a linear finite-dimensional dynamical equation E that has G(t, T) as its impulse
e relation between the response matrix, then G(t, T) is said to be realizable. We cal\ the dynamical
Let the input-output equation E, or more specifical\y, the matrices {A, B, e, E}, a realization of
G(t, T). The terminology "realization" is justified by the fact that by using the
out terminals be
dynamical equation, we can build an operational amplifier circuit that will
(4-47). generate G(t, T). Note that the state of a dynamical7equation realization of the
impulse-response matrix of a system is purely an auxiliary variable and it may
: vector, and G is the not have any physical meaning. Note also that the dynamical-equation realiza
implicitly assumed in tion gives only the same zero-state response of the system. If the dynamical
1 (ith row, jth column)
equation is not in the zero-state, its response may not have any relation to the
lal due to a b-function system.
;e now that the internal If the realization of an impulse response G(t, T) is restricted to a finite
lllalysis of this system dimensional linear dynamical equation of the form (4-48), it is conceivable that
not every G(t, T) is realizable. For example, there is no linear equation of the
form (4-48) that will generate the impulse response of a unit-time-delay system or
(4-4881 ) the impulse response l/U - ,). We give in the [oliowing the necessary ana suf
(4-48b) ficient condition for G(t, T) to be realizable.
B, e, and E are n x n,
tinuous functions of t Theorem 4-9
~-48) are two different
.me input-output pairs A q x p impulse-response matrix G(t, T) is realizable by a finite-dimensional
lmical equation E with linear dynamical equation of the form (4-48) if and only if G(t, T) can be de
composed into . ..
g(t-r)=(t-r)eA(t-<)=[e At teAtll
r -re-A<l
.L'
~ Le'" J
Hence, the dynamical equation
E: Since (4-53) and (4-55) <
As discussed in (3-}
matrix, and C(sI - At 1
is a realization of g(t, r). I function matrix of the d
Theorem 4-10
All the equivalent dynamical equations have the same impulse-response
matrix; hence, ifwe find a realization of G(t, r), we may obtain different realiza A transfer-funciion ma
tions of G(t, r) by applying equivalence transformations. Note that an impulse- time-invariant dynamic~
[RICES IMPULSE-RESPONSE MATRICES AND DYNAMICAL EQUATIONS 157
response matrix may have different dimensional realizations; for example, the
networks in Figure 4- ~ are two different dimensional realizations of g(t, r) =
b(t - r). For further results in realization, see Reference S128.
+ u
x/I_ I O O O O X/I_ I O
and let
X/I -CJ." -ct. n - 1 -0:/1-2 -0: 2 -(;(1 X/I 1 (4-58a)
8 Mason's gain formula for signal-flow graph is as follows: The transfer fllnction of a signal-flow
graph is
where Ó = l-(I: all individual loop gains) +(I: aH possible gain prodllcts of two nontouching
loops)-' .. ; g¡ = gain of the ith forward path, and ó; = the part of Ó nol louching lhe ith forward is a realization of (;(s).
path. Two loops or two parts of a signal-flow graph are said to be nontollching if they do '10l have
any poinl in common. See, e.g., Reference S46.
9This can also be pro ved by computing algebraicaHy the transfer fllnction of (4-58). This is done in
Chapter 6.
e
{JI
{J2'-___
X2
~-invariant al a2 an
dynamical /!,. = 1 +- +- + ... +- (4-59)
·oper rational matrix. s S2 sn
is realizable, we first and /!,.i = 1, for i = 1, 2, ... , n. Hence, the transfer function of (4-58) from the
realizable. The most input u to the output y is
{J IIs + {J 21S2 + ... + {J niSn {JI Sn - 1 + ... + (Jn A
e+ 2 n=e+ =g(s)
(4-57) 1 +aIls +a2/s + ... +anls s" +a 1s" 1 + ... +a n
This proves the assertion that every scalar proper rational function is realizable.
namical equation We are now ready to show that every proper rational matrix is realizable.
Xl o In order to avoid cumbersome notations, we assume that (;(s) is a 2 x 2 matrix.
Let
X2 O
+ u (4-60)
O
and let
1 (4-58a)
(4-58b) be a realization of g¡j, for i,j = 1,2; that is, gJs) =cJsI -A¡F 1bu +eu' Note
he transfer function of that the bds are column vectors and the cL/s are row vectors. Then the com
posite dynamical equation
[~"j C'
O O
There are n loops with
~cept the direct trans:
.. , (Jjs". Since all the
X12 _
X21 -
O
O
A 12
O
O
A 21 O
O
O
X 12
X 21
+ O
b 21
b 22
(4-61 a)
i 22 O O O A 22 X 22 O
C
["]
er runction or a signal-flow
[~J=[C~l
O O el!
12
O C21 C22 J X 12
X 21 +[e 21 e
12 1
J[u J
e 22 U2
(4-61 b)
X 22
oducts of two nontouching
not touching the ith forward is a realization of (;(s). Indeed, the transfer-function matrix of (4-61) is
touching ir they do not have
~
O
~Y !ln
(4-62)
4-58).
---- --------~--- ------- -- =--=---=--=-=---=----=--==-=---,.--,,-=--=--=--=-----=--=- =--~-,-===========-~-- -=-=-=--=-'==--=.. _=--='=---=---=--=' -----
~-~--
Thus every proper rational matrix is realizable by a finite-dimensional linear Since a realizable
time-invariant dynamical equation, Q.E.Do since entries of 'G(l)
i = 1, 2, ... , the entri<
The realization procedure discussed in (4-60) and (4-61) is simple and straight ConsequentIy, the mal
forwardo We realize every element of a transfer-function matrix independently real lineo As a consec
and then connect them from the input and the output as shown in Figure 4-70
The resulting realization, however, is generalIy not satisfactory for the folIowing Theorem 4-11
two reasons: First, the realization is internalIy not coupled as can be seen from
Figure 4-7. The coupling or interacting of alI variables, however, is a feature of A system that has a pr,
most physical multivariable systemso Second, the dimension of the realization only if U[lo,lo+EJ =0 in
is generalIy unnecessarily large. These two problems will be resolved in Chapter
6. This is a restateme
The condition ofTheorem 4-10 is stated in terms of transfer-function matri have rational transfer
ces. We may translate it into the time domain as folIows: linear time-invariant e
an interval (no matter
Corollary 4-10 interval.
iT~l
lA
:""' I ~r :
Y,
Every proper ratior
dimensional!inear time
-~.~
synthesis problem in ne
: i was also constructedo
generally it is possible t
: I
, 1---
I I
A
g,2(S)
I
f-
l'
I
I
I
be discussed in Chapter
bility,
; \ I
~
"2 • I r~
~ :Y2
+rc--'--
To conclude this el
equationso The state
: 1 . :
I
...J A(m) (Problem 4-27). 1
putable. In terms of <1>(,
Figure 4-7 lnternally uncoupled realization of (4-60), can be obtained as in Pr,
TRICES CONCLUDING REMARKS 161
te-dimensional linear Since a realizable G(t) can be decomposed into G(t - r) = M(t)N(r), and
Q.E.O. since entries of "G(t) are linear combinations of tkeÁ¡t, for k =0,1, ... , and
i = 1,2, ... , the entries of M(t) and N{t) must be linear combinations of tke Áit .
is simpleand straight Consequently, the matrices M(t) and N(t) are analytic functions of t on the entire
matrix independently realline. As a consequence of this fact, we have the following theorem.
: shown in Figure 4-7.
~tory for the following Theorem 4-11
:d as can be seen from A system that has a proper rational transfer function matrix is relaxed at to if and
owever, is a feature of
only if u[lo.lo+'l = O implies Y[to.to+'l = O for sorne positive real s. I
sion of the realization
be resolved in Chapter
This is a restatement of Corollary 3-1. Hence, for the class of systems that
have rational transfer-function matrices-or equivalently, are describable by
ansfer-function matri
linear time-invariant dynamical equations-if the output is identically zero in
s: an interval (no matter how smal1), then the system is relaxed at the end of that
interval.
itecdimensional linear
ntry of G(t) is a linear 4-5 Concluding Remarks
., and i = 1, 2, ...) and
I
The solutions of linear dynamical equations were studied in this chapter.
The solution hinges on the state transition matrix <I>(t, r), which has the proper
)lace transform of the ties <I>(t, t)=I, <I>-l(t, r)=<I>(r, t), and <I>(t, r)<l>(r, to)=<I>(t, tolo For the time
ion is of the form varying case, <I>(t, !) is very difficult to compute; for the time-invariant case,
<I>(t, !) is equal to eA(t-t l , which can be computed by using the methods intro
duced in Section 2-7. In both cases, if only a specific solution is of interest, we
may bypass <I>(t, !) and compute the solution on a digital computer by direct
lbination of the terms integration.
Oifferent analyses often lead to different dynamical-equation descriptions
of a system. Mathematically, it means that dynamical equations depend on the
10. basis chosen for the state space. However, the input-output description has
nothing to do with the basis; no matter what analysis is used, it always leads
to the same input-output description. This is an advantage of the input-output
description over the state-variable description.
Every proper rational function has been shown to be realizable by a finite
dimensional linear time-invariant dynamical equation. This corresponds to the
synthesis problem in network theory. A realization of a proper rational matrix
was also constructed. However, the realization is not satisfactory, because
general1y it is possible to construct a lesser-dimensional realization. This will
be discussed in Chapter 6 after the introduction of controllability and observa
bility.
To conclude this chapter, we remark briefly the solution of discrete-time
equations. The state transition matrix <I>(k, m) of .x(k + 1) = A(k)x(k) =
A(k)A(k -l),x(k - 1) = ... can be readily computed as <I>(k, m) = A(k -1)'"
A(m) (Problem 4-27). Unlike the continuous-tim~ case, <I>(k, m) is easily com
putable. [n terms of <I>(k, m), the solution ofa discrete-time dynamical equation
can be obtained as in Problem 4-29. For the time-invariimt case, the equivalent
162 LINEAR DYNAMICAL EQUATlONS AND IMPULSE-RESPONSE MATRICES
dynamical equations and the realization problem are identical to the con 4-7 Given
tinuous-time case. For example, a sampled transfer-function matrix (;(z) has a
finite-dimensional realization of the form in (3-80) if and only if G(z) is a proper
rational matrix in z. The realization procedure in Theorem 4-10 is also directly
applicable to the discrete-time case. show that
4-1 Find the fundamental matrices and the state transition matrices of the following
o
homogeneous equations: o,
4-8 Givenx(t)=A(t)x.
where A* is the complex '
transition matrioes of x =
and
4-4 Let
o 1
1 O
O 1
O O
d
-"[T- l(t)] = - r l(t) [el
- TU) ] r l(t)
de . elt ..
4-6 From l1>(t, ¡), show how to compute AU). Figure P4-9
PROBLEMS 163
.RICES
where 8<1>(t, to)lat = A(t)<I>(t, to) and <I>(to, to) = I. Hint: Show that
a
latrices of the following - det <I>(t, to) = (a¡¡(t) +a 22 (t)) det <I>(t, to)
ot
4-8 Given x(t) =A(t)x. The equation z= -A*z is called the adjoint equation ofx =A(t)x,
where A* is the complex conjugate transpose of A. Let <I>(t, to) and <l>a(t, to) be the state
transition matrices of x= A(t)x and z= - A*z, respectively. Verify that
4-9 Consider
G(t, r) = G:(r, t)
Show that ir A, B, and C are constant matrices and G(s) and Gis) are their transfer-function
matrices, then
G(s) = - G:( -s)
Figure P4-9
164 LINEAR OYNAMICAL EQUATIONS ANO IMPULSE-RESPONSE MATRICES
4-10 Every e1ement of <1>(t, to) can be interpreted as the impulse response of sorne input
output pairo What is the input and the output of the·ijth element of <1>(t, to)?
4-11 Let
In
Figure P4-17
be the state transition matrix of
and the initial capacitor ,
X=[A~¡ :::}
immediately of the form ,
[
tlt
where <1>(t, to) is the state transition matrix of x = A(t)x.
where A2~A(O)-A¡.
4-16 Find the impedance (the transfer function from u to i) of the network in Figure
4-21 Can you transform
P4-16. lf the initial conditions of the inductor and the capacitor are zero and if an input
equivalence transformatioJ
voltage u(t)=e- c is applied, what are ;(t), ;¡(t) and iz(t)? Note that ;j(t) and ;2lt) contain
sorne exponential functions that do not appear in ¡(t). How do you explain this? 4-22 Find a time-varying
realization of the impulse r
4-17 Consider the network shown in Figure P4-17. Find the initial inductor current
4-23 Find a dynamical-e(
find a linear time-invarianl
4H
Figuré P4-16
y=[l
.TRICES PROBLEMS 165
tt of <1>(t, to)?
Figu re P4-17
and the initial capacitor voltage such that for the input u(t) = e- 41 the output y(t) wil1 be
immediately of the form e- 41 without containing any transient.
1~][::]+[-~]
4
20 u
-25 -20 X3 O
XI(n +
x¡(n+l) =
1)] [O
O 20
4
3][X1(n)] + [-1]
16 x¡(n) 3 u(n)
[
natrix of x= A(t)x is given x3(n + 1) O -25 -20 x3(n) O
XI(nl]
y(n)=[ -1 3 OJ x¡(n) +4u(n)
[x (n)
3
of the network in Figure 4-21 Can you transform a time invariant {A, B, C} into {O, B, C} by a time-varying
Jr are zero and if an input equivalence transformation?
that ;1(1) ~.nd ;1(t) contain
you explain this? 4-22 Find a time-varying dynamical equation realization and a time-invariant dynamical
realization of the impulse response g(t) = t 2 e!·I.
he initial inductor current
4-23 Find a dynamical-equation realization of g(t, e) = sin te - (r -<) cos e. ls it possible to
find a linear time-invariant dynamical-equation realization for it?
4-24 Use a signal-f1ow graph to show that the transfer function of the following single-
variable, linear time-invariant dynamical equation
X~ [¡
O O
O O
1 O
O
O -"o
-ct,,_l j
13U-l .
O -~n-¡ x + P";-2 u
[~.]
O O -!XI PI
y=[O O O O lJx+eu
166 LINEAR OYNAMICAL EQUATIONS ANO IMPULSE-RESPONSE MATRICES
is Verify tha t
A {31 S ,,-1 + ... +{3"
g(s)=e + S" +O:IS"-1 + ... +CI."_IS +0:"
1
S+2 2+2
ls +1
5
s +1
s +3
Ss
1
s +2
+1
1
l
into continuous-time i:\nd discrete-time dynamical equations.
2
2
+1
5
+1
2
52
2
+3
1
+1
+2
if and only if A(t) and A((
(
4-26 Consider the equivalent dynamical equations
{*y=Cx
Ax + Bu
= {x ~x + Bu =
é
y=Cx is
where x = Px. Their adjoint equations are, respectively,
[Hint: Use Problem 4-30 a
:i= -A*z+C*u (la)
{ y=B*z (lb)
z= ::-A*z + C*u (2a)
{y=B*z (2b)
where A* and Á* are the complex conjugate transposes of A and Á, respectively. Show
that Equations (1) and (2) are equivalent and that they are related by z= (p-l )*z.
<1>(m, m)~ 1
Show that, given the initial state x(m) = xo, the state at time k is given by x(k) = <1>(k, m)x o·
If A is independent of k, what is <1>(k, m)?
4-28 For continuous-time dynamical equations, the state transition matrix <1>(t, e) is
defined for aH t, T. However, in discrete-time dynamical equation, <1>(k, m) is defined only
for k 2: m. What condition do we need on A(k) in order for <1>(k, m) to be defined for !< < m?
[This can be easily verified by considering B(l)u(l) as an initial state at time (l + 1).J Show
that if A(k) and B(k) are independent of k, then the solution becomes
k-l
x(k) = Akx(O) +I Ak - 1
-
III
Bu(m)
m=O
d (d
- A(t) ~ -'--- adt)
dt dt J
)
ArRICES PROBLEMS 167
Verify that
d . •
~ (A(t)B(t)) = A(t)B(t) + A(t)B(t)
dt
Verify also that
d d .
- [A(tW ~- (A(t)A(t)) = 2A(t)A(t)
dt dt
;j
if and only if A(t) and A(t) commute; that is, Á(t)A(t) = A(t)Á(t).
4-31 Show that if r A(r) dr and A(t) commute for all t, then the unique solution of
'o o
- <1>(t, to) = A(t)<l>(t, to) <1>(to, to) = 1
oc
is <1>(t, to)=exp flO
A(r) dr
for k>m
lis given by
)11(1)
Xl ¡ er'
u
Dampi.ng
coefficien!
,11 Spring
constan!
(curren!
2
1
service)
observable from the output y. On the other hand, the presence ofthe mode e- l
in II can be detected from the output y; hence the mode is said to be observable.
This illustration, though not very accurate, may convey the ideas of the con
cepts of controllability and observability.
The concepts of contrQllability and observability are very important in the
study of control and filtering problems. As an example, consider the platform
system shown in Figure 5-2. The system consists of one platform; both ends
of the platform are supported on the ground by means of springs and dashpots.
The mass ofthe platform is, for simplicity, assumed to be zero; hence the move
ments ofthe two spring systems are independent. Ifthe initial displacements of
both ends of the platform are different from zero, the platform will start to
vibrate. If no force is applied, it will take an infinite time for the platform to
come back to rest. Now we may ask: For any initial displacements, is it possible
to apply a force to bring the platform to rest in afinite time? In order to answer
this question, the concept of controllability is needed.
This chapter is organized as follows. In Section 5-2 the required mathe
matical background is introduced. Three theorems that give the conditions
for linear independence of a set of vector functions are presented. All the
results in controllability and observability follow almost directly from these
three theorems. The concept of controllability is introduced in Section 5-3.
Necessary and sufficient conditions for linear time-varying dynamical equa
tions and linear time-invariant dynamical equations to be controllable are
derived. The concept of observability is introduced in Section 5-4. It is dual
to the concept of controllability; hence its discussion is rather brief.· Duality
ltitative and qualitative.
theorem is also developed. In Section 5-5, we study dynamical equations
:t response of the system
which are uncontrollable and/or unobservable. The canonical decomposition
n the preceding chapter.
theorem is developed. A consequence of this theorem is that the transfer
II properties of a system.
function matrix of a. dynamical equation depends solely on the part of the
~rties oflinear dynamical
equation that is controllable and observable. In Section 5-6, we study the
first give the reader sorne
controllability and observability of linear time-invariant Jordan-form dynami
)rk shown in Figure 5-l.
cal equation. Their conditions are very simple and can be checked almost by
.r that ifthe initial voltage
inspection. In Section 5-7, the concepts of output controllability and output
lt input u is applied, the
function controllability are introduced. It is shown that they are properties
e- t in 11 is said to be not
of the input-output description of a system. In the last section, we discuss
node e -1 in loop 1 can be
sorne computational problems encountered in this chapter.
)de in 1 is controllable by
Although elements of the matrices A, B, e, and E are all real-valued, for
by the input, its presence
mathematical convenience they are considered as elements of the field of com
-lence it is said to be not
¡2U
Xl
I .'":';'
¡ x2
Damping Spring Damping Spring
coefficient constant coefficient constant
2 1 J 1
plex numbers. Consequently, the state space of an n-dimensional dynamical The concept of lin
equation will be taken as an n-dimensional complex vector space (C", iC). tions. Let f j , for i = 1,
The references for this chapter are 2, 8,11,13,14,20,21,48,55,56,60,61, 1 x p complex-valued
69, 71, 98, 103, and 105 to 107. if there exist complex
The reader who is interested in only the time-invariant case may skip
alfl(t) +a
Theorems 5-2, 5-5, and 5-6.
Otherwise, the f¡'s are
in Equation (5-2) is a
5-2 Linear Independence of Time Functions we may also state tha
only if
The concept of linear independence of a set of vectors of a linear space was
introduced in Section 2-3. We shall now apply this concept to a set offunctions
ofa real variable. A set of complex-valued functions Jl,f2> ... , Jn is said to be
linearly dependent on the interval 1 [t 1> t 2] over the field of complex numbers
if there exist complex numbers al' a2' ... , a m not all zero, such that
aI!I(t)+al!2(t)+··· +allfn(t) =0 foral/tin [tI' t 2] (5-1) implies IX = 0, where
Otherwise, the set of functions is said to be linearly independent on [t 1> t 2 ] over
the field of complex numbers. In this definition, the specification of time
interval is crucial.
Example 1
Clearly, Ot is a constan
Consider the two continuous functions 11 and 12, defined by The linear indepen
J~(t)=t for / in [ -1, 1] an interval; hence in t
for t in [0, 1] entire interval. Let F'
12(t) = { t
-t for t in [ - 1, O]
Theorem 5-1
lt is clear that the functions(1 and.f2 are linearly dependent on [O, 1], since if Let f j , for i = 1, 2, ... , t
we choose al = 1, a2 = -1, then a¡j"l(t) +ad1.(t) =0 for aH I in [0, 1]. The on [/,. I J. Let F be t
functions JI and j~ are also linearly dependent on [ -1, O]. However, JI and
.f2 are linearly independent on [ -1, 1]. I
From this example, we see that although a set of functions is linearly inde Then f l , f 2 , .. o. f are
ll
pendent on an interval, it is not necessary that they are linearly independent on constant matrix W(t l ' I
1 TÍJe functions we study are mostly continuous functions: hence there ¡s·no sUDstantia\ difreorence
between using the open intervallt ,. e2) and the closedinterval [e l' e2] Every [e ,. e2 ] ís assumed to 2 The matrix W(t ,. e2) is in ri
be a nonzero interval. in this chapter.
.__ ._~.:=~---.--_.-':::_ .. _-~---~
limensional dynamical The concept of linear independence can be extended to vector-valued func
:tor space (Cn, C). tions. Let f¡, for i = 1, 2, ... , n, be 1 x p complex-valued functions of t; then the
),21, 48,55, 56, 60, 61, 1 x p complex-valued functions f l , f l , ... , f n are linearly dependent on [t l' tlJ
if there exist complex numbers al' al" .. ,a", not all zero, such that
variant case may skip (5-2)
Otherwise, the f;'s are linearly independent on [t 1> t l ]. Note that the zero vector
in Equation (5-2) is a 1 x p row vector [O ° ...
O]. As in Definition 2-4',
we may also state that f¡, f l , ... , f n are linearly independent on [ti' tlJ if and
'Os
only if
; of a linear space was
:ept to a set offunctions
,fl' ... , fn is said to be for aH t in [tI' tlJ (5-3)
Id of complex numbers
J, such that
F"r~:l
)endent on [t l' t lJ over
~ specification of time
... J
an
lf..J
Clearly, t:x is a constant 1 x n row vector and F is an n x p matrix function.
ed by
The linear independence of a set of functions is a property associated with
an ¡nterval; hence in testing for linear independence, we have to consider the
entire interva\. Let F*(t) be the complex conjugate transpose of F(t).
Theorem 5-1
ldent on [0, lJ, since ir Let f j , for i = 1, 2, ... ,n, be 1 x p complex-valued continuous functions defined
)r all I in [0, 1]. The on [1 l' I J. Let F be the 11 x (! matrix with f; as its ith row. Define
,0]. However, fl and
I W(t]. 12)~ r" F(t)F*(t) dI
Jt l
rlctions is linearly inde Then f l • f 2 , ...• f n are linearly independent on [1 \. I J ir and only ir the 11 x 11
re is no substantia\ difference
Every [t,. t,] is assumed to , The matrix W(t ,. t ,) is in faet positive definite(see Definition 8-6). This property is not needed
in this ehapter. .
=. . .=-.--=..
·--'="--==.C'-... =----C-.~._."" .• ,_•.• -._.-~ •• -" ..••. _.-.-.
,_U . , , _ . _ _. _ •• _ - • • •'_ •• _-_ •• _ - , - _ ••••• _ ••• "~ ••
r exF(t)F*(t)dt=O
t2
exW(t l ,t 2 )= (5-5)
J" are linearly depender
which contradicts the assumption that W(t l , t 2 ) is nonsingular. Hence, if
W(t l' t 2 ) is nonsingular, then the f¡'s are linearly independent on [t ¡, t 2 ].
Q.E.D. has rank n. Hence ti
The determinant of W(t l' t 2) is called the Gram determinant of the f¡'s. In The condition of
applying Theorem 5-1, the functions f i , for i = 1,2, ... , n, are required to be functions to be linea
example.
continuous. lf the functions f;, for i = 1, 2, ... , n, have coIitinuous derivatives
up to order (n -1), then we may use the fo11owing theorem.
Example 2
Thearem 5-2 Consider the two fun(
Assume that the 1 x p complex-valued functions f l , f 2 , . . . , f" have continuous
derivatives up to order (n - 1) on the interval [t ¡, t 2]. Let F be the n x p matrix
with f i as its ith row, and let F(k) be the kth derivative of F. lf there exists some
defined ayer l -1, 1].
t o in [tI' t 2J such that the n xnp matrix
has rank n, then the f¡'s are linearly independent on [t J, t 2J over the field of
complex numbers. 3
Proof
and
We prove the theorem by contradiction.' Suppose that there exists some t o in
[tI' t 2 J such that
To check the linea
continuous, we can em
interval. If the functit
, Ir lo is al eilher 1, or 12 , lhe end points of an interval, then FÚ<l(to) IS denned as· ·F1k)(I). with I then Theorem 5-2 can 1
approaching lo from inside the interval. lheorem 5-1; however
CAL EQUATIONS LINEAR INDEPENDENCE OF TIME FUNCTIONS 173
and the f;s are linearly dependent on [t b t2]. Then by definition, there exists a
nonzero 1 x n row vector ex such that
=0 (5-4 )
exF(t) =0
This irnplies that
on and is nonnegative
exF(kl(t) = O for all t in [t l> t 2] and k = 1, 2, ... , n - 1
Hence we have
f the set of f j , i = 1, 2, ex[F(t) : F(l)(t) : ... : F(ll- ll(t)] = O
], then det W(t l , t 2) 1= O.
in particular,
)pose that W(t l' t 2) is
l. Then, by definition, ex[F(to): F(ll(t O) : ... : F(ll-l)(t O)] =0
1 that exF(t) =0 for aH
which irnplies that aH the n rows of
[F(to) : F(l)(t o) : ... : Fln-l)(to)]
(5-5 )
are linearly dependent. This contradicts the hypothesis that
)nsingular. Hence, if [F(t o) : F(l)(t o) : ... : F(ll-l)(t O)]
lependent on [ti' t 2 J. Q.E.D.
has rank n. Hence the f;s are linearly independent on [ti' t 2 ].
Q.E.D.
The condition of Theorern 5-2 is sufficient but not necessary for a set of
rminant of the f¡'s. In functions to be linearly independent. This can be seen from the following
, n, are required to be exarnple.
;;ontinuous derivatives
~rn.
Example 2
Consider the two functions
.. , f n have continuous fl(t) = t3
:t F be the n x p matrix f2(t) = It 3 1
F. If there exists sorne defined over [ -1, 1]. They are linearly independent on [ -1, 1]; however,
rOí" aH L in lO,
l' t 2 ] over the field of 3 2
fl(t) j¡<l)(t)J
analytic, then we can use Theorem 5-3, which is based on the fact that if a every subinterval of
function is analytic on [tI' t 2 J, then the function is completely determinable essential. The staten
from a point in [t 1 , t 2 J if aH the derivatives of the function at that point are Note that Theore
known. (See Appendix B.)
Proof
Then
The sufficiency of the theorem can be proved as in Theorem 5-2. Now we
prove by contradiction the necessity of the theorem. Suppose that [F(t) : 1
p[FU o) : F(l)UO) : ... : F(n-l)(to) : .. 'J < n
It is easy to verify th
Then the rows of the infinite matrix
p[F(t) : flll(t) : F(2 l (t) : F(
[F(to): F(I)(t o): ... : F(II-l)U O): "'J sin lO't '
=p [
are linearly dependent. Consequently, there exists a nonzero l x n row vector sin 2 x lO't : 2 x 1
a. such that for al! t.
(5-8)
The matrix in (5-~
The Í¡'s are analytic on [t 1 , t 2 J by assumption; hence there exists an 6>0 such many cases, it is not no
that, for aH t in [to -6, t o +6J, F(t) can be represented as a Taylor series about Example 3, we check
the point t o:
~o)" F(II)(t O)
n.
F(t) = ffor aH t in [to - 6, to +6J
11=0
(t - (5-9)
then we have the follo
Premultiplying a. on both sides of (5-9) and using (5-8), we obtain
a.F(t) =0 for al! t in [to -6, t o +6J (5-10) Corollary 5-3
Since the sum of analytic functions is an analytic function, the analyticity Assume that, for each
assumption of the f/s implies that a.F(t) as a row vector function is analytic over independent on [t b t 2
[t 1 , t 2 ]. Consequently, Equation (5-10) implies that
a.F(t) =0 for al! t in [t 1, t 2 J
for almost all t in [t 1> j
or, equivalently, the f/s are linearly dependent on [t1> t 2 ]. This is a contra This corollary wil
diction. Q.E.D. omitted.
A direct conseque'nce of this theorem is that if a set of anal ytic functions is
linearly independent on [t1> t 2 J, then
5-3 Controllabil
p[F(t): f'(I)(t) : ... : F(n-1)(t)=--'J =n
Time-varying case
for al! t in [t l' t 2]. Itfollows that if asid of analytic functions is linearly inde trollability of linear d
pendent on [t l' t 2 J, then the set of analytic functions is linearly independent on state controllability o
o\L EQUATIONS CONTROLLABILlTY OF LINEAR DYNAMICAL EQUATlONS 175
on the fact that if a every subinterval of [tI' t 2 ]. In this statement, the analyticity assumption is
pletely determinable essential. The statement does not hold without it, as we have seen in Example 1.
ion at that point are Note that Theorem 5-3 is not true if the infinite matrix in (5-7) is replaced by
p[F(ro) : F(l)(ro) : ... : Fin - l)(tO)J = n
Then
~orem 5-2. Now we
sin lOOOt: 10 3 cos lOOOt J
'pose that [F(r) : F(l)U)] =.
[ sm 2000t :' 2 x 10 3 COS 20001
<n It is easy to verify that p[F(t): F(l)(t)] < 2 at t = O, ± 10 - 3n;, . . .. However,
p[F(t) : Fl)(t) : F(2)(t) : F( 3 )(t)]
=p [
sin 10 3 r 10 3 cos 10 3 r -10·sinI0 3 r
3
' -10·cosI0 3 r J
zero 1 x n row vector sin 2 x 10 3 { : 2 x 10 3 cos 2 X 10 3 r -4xI0.sin2xl0 r :-8xI0 9 cos2xI0 3 r =2
for all t. I
=O (5-8)
The matrix in (5-7) has n rows but infinitely many columns. However, in
:e exists an e> O such many cases, it is not necessary to check aH the derivatives of F. For instance, in
a Taylor series about Example 3, we check only up to F(3 l . lf we use the matrix
[F(t) : F(l)(t) : ... : F(n -l)(t)]
e, t o +eJ (5-9)
then we have the following corollary.
e obtain
] (5-10) Corollary 5-3
ction, the analyticity Assume that, for each i, f¡ is analytic on [t l' t 2]. Then f l' f 2 , ... , f n are linearly
lction is analytic over independent on [t 1, t 2 J if and only if
p[F(t) : F(l)(t) : ... : F(n -l)(t)] = n
for almos! all t in [tI' t 2 ].
2]. This is a contra This corollary will not be used in this book, and therefore its proof is
Q.E.D. omitted.
.f analytic functions is
1
176 CONTROLLABILlTY AND OBSERVABILlTY OF LINEAR DYNAMICAL EQUATIONS
the state controllability is a property of state equations only; output equations precisely, the dynami
do not play any role here. at any to.
Consider the n-dimensional, linear state equation
Example 2
E: x= A(t)x(t) + B(t)u(t) (5-11 )
Consider the system ~
where x is the n x 1 state vector, u is the p x 1 input vector, and A and B are,
X2 in the system, T
respectively, n x n and n x p matrices whose entries are continuous functions
cannot transfer Xl ay.
of t defined over ( - 00,00). The state space of the equation is an n-dimensional
then no matter what
complex vector space and is denoted by ~.
Hence the equation t
Definition 5-1
The solution of tn
The state equation E is said to be (state) controllablé at time to, if there exists a
finite ti > t o such that for any x(to) in the state space ~ and any Xl in ~, there x(t) = «I»(t; i
exists an input u[to,ld that will tIansfer the state x(t o) to the state Xl at time ti'
Otherwise, the state equation is said to be uncontrollable at time to.
This definition requires only that the input u be capable ofmoving any state
in the state space to any other state in afinite time; what trajectory the state where <I>(t, t o) = 'P(t)\
should take is not specified. Furthermore, there is no constraint imposed on the nonsingular for aH t.
input. Its magnitude can be as large as desired. We give some examples to
illustrate this concept. Theorem 5-4
The state equation E
Example 1 ti> t o such tqat the 11
Consider the network shown in Figure 5-3. The state variabk x ofthe system is independent on [t o, ti
the voltage across the capacitor. Ir x(to) = O, then x(t) = Ofor all t '2:. t o no matter
what input is applied. This is due to the symmetry ofthe network, and the input Proof
has no effect on the voltage across the capacitor. Hence the system-or, more Sufficiency: If the row
Theorem 5-1 the n x y.
W(t
4 In lhe lileralure, if a slale can be lransferred lo lhe zero state O, the slate is said to be controllable.
If a state can be reached frarn O, lhe state is said to be reachable. Our definition does not rnake
this distinction to simplify the subsequent presentatiol1. Furthermorc. lhe cquation E is said lo
be, in the literature, completely controllable. For conciseness, the adverb "cornpletely" is dropped
in this book. u(t) =
In In
1 Y
+ ~.
11 rv IF;::
In In 1.11:
.
Figure 5-3 An uncontrollable network. Figure 5-4 An uncont
:AL EQUATlONS CONTROLLABILITY OF LINEAR DYNAMICAL EQUATIONS 177
[lly; output equations precisely, the dynamical equation that describes the system-is not controllable
at any too
Example 2
(5-11 )
Consider the system shown in Figure 5-4. There are two state variables X¡ and
tor, and A and B are, Xz in the system. The input can transfer X¡ al' Xz to any value; however, it
continuous functions
cannot transfer x¡ and Xz to any values. For example, if X¡(t o) 0:=0, xz(to)=O,
III is an n-dimensional
then no matter what input is applied, x¡(t) is always equal to xz(t), for all t> too
Benee the equation that describes the system is not controllable at any too iI
at time too
= 1I>(t, t o{ Xo + L lI>(t o, L)B( L)u(L) dLJ (5-12)
(5-13)
.te is said to be co/ltrolfable.
r definition does not make
js nonsingular. Given any x.(t o) = ~~o and any x j", '.ve claii1J that tb.e input
c. tl1e equation E is said lo
:rb "completely" is dropped
(5-14 )
will transfer X o to the state Xl at time ti. Indeed, by substituting (5-14) into
1F
+tzX
-=.t
Thus we condude that the equation E is controllable. Necessity: The proof is Observe that
by contradiction. Suppose E is controllable at to, but the rows of <I>(to,' )B(')
$(t o, t)B
are Iinearly dependent on [1 (l. (IJ for aH (1 > (o. Then there exists a nonzero,
constant IXIl row vector a. such that o
ot $(t o, t)B
a.<1>(t (l. t)B(t) = O for all 1 in [lo. IIJ (5-15 )
<1>U (l. tl)xU 1 ) = a.* + r" <1>(1 (l. r )B( r )u( r) dr (5-16 )
10 and, in general,
Premultiplying both sides of (5-16) by a.. we obtain Ok
ot k $(t o, c:
a.<1>(t(l.t l )x(tI) =a.a.* + r" a.<1>(to. r)B(r)u(r) dr (5-17 )
10 Theorem 5-5
By hypothesis, E is controHable at t o ; hence for any state-in particular,
Assume that the matri(
XI =O-there exists ullO,ld such that X(I¡)=O. Since a.<1>(10, t)B(t) =0, for allt
n - 1 times continuom
in [to, t¡J, Equation (5-17) reduces to
lable at t o if there exist
a.a.* =0
p[r
which, in tum, implies that ex = O. This is a contradiction. Q.E.D.
Proof
In the proof of this theorem, we also give in (5-14) an input u(t) that transfers
Define
x(t o) to x¡ at time tI' Because of the continuity assumption of A and B, the
input u in (5-14) is a continuous function oí t in [t o, tI]. ¿
If a linear dynamical equation is controllable, there are generally many dif
ferent inputs u that can transfer x(t o) to x¡ at time t ¡, for the trajectory between Then, from (5-19) and 1
x(t o) and Xl is not specified. Among these possible inputs that achieve the
same mission, we may ask which input is optimal according to sorne criterion. [$(to, t¡)B(t¡) : O~l $(t
If the total energy
r Ilu(1)111 dt
ll
) dr (S-16)
and, in general,
ak
k=O,I,2, .. ,n-l
at k <I>(to, t)B(t) = <I>(to, t)Mk(t) (S-19c)
(r) dr (5-17 )
Theorem 5-5
state-in particular,
Assume that the matrices A(') and B(-) in the n-dimensional state equation E are
I(tO, t)B(t) = O, for aH t
n -1 times continuously differentiable. Then the state equation E is control
lable at t o if there exists a finite t¡ > t o such that
(S-20)
n. Q.E.D.
Proof
nput u(t) that transfers
Define
lption of A and B, the
a ""\lo, ¡',)1lJl(
- ,-,,1. "'\
j;]) LJ
L a t"', (e.0' " )1R1-
~'a'--- 0' ,¡ M\L¡i
'
at 1=1, t¡
lre generally many dif
the trajectory between Then, from (5-19) and using (2-2), we have:
lputS that achieve the
jing to sorne criterion.
a an -¡
[<I>(t o, t¡)B(t¡) : 'Jt;<I>(t o, t¡)B(t¡) : ... : at1-¡ <I>(t o, t¡)B(t¡)
J
= <I>(t o, t¡)[Mo(t¡} : M¡(t¡): ., . : Mn-1(t¡)] (S-21 )
Since <I>(to, ti) is nonsingular, the assumption
It follows from Theorem 5-2 that the rows of $(t o, .)B(-) are linearly independent controllability. The (
on [t o, t 1 ] for any t 1 > tá. Thus, from Theorem 5-4, we conclude that the state obtained by slight mo
equation E is controllable. Q.E.D. matrices A and B in the
the following theorem.
As in Theorem 5-2, the condition of Theorem 5-5 is sufficient but not neces
sary for the controllability of a state equation.
Theorem 5-6
Example 3 If the matrices A and E
Consider equation E is different
for any fixed to in ( - ex
Xl]
~z
[X3 =
10J° [1] +[0]
[t° Xz 1 u (5-22) p[Mol
°° t
Z
X3 1 If the matrix A is an
From (5-18), we have tion matrix $(t o, .) of x
of two analytic functior
implies that $(t o, -)B(-)
is that a set of analytic I
if the set of analytic fm
matter how small) of(
from Theorems 5-3 an
A and B is controllable
t in ( - 00, 00).
Definítion 5-3
The linear dynamical
in ( - 00, 00) if and onl)
Since the matrix [Mo(t) : M 1 (t) : Mz(t)] has rank 3 for all t -/=0, the dynamical
equation is controllable at every t. I
If every state in L can be transferred to any other state in a finite time (no
matter how long), the state eqúation is said to be controllable. If this can be
5 In lhe engineering lileraluro
achieved in an arbitrarily small interval of time, then the state equation is said
firsl used in Reference S6 lo
to be differentially controllable. Clearly, differential controllability implies we adopt lhe terminology "
MICAL EQUATIONS
CONTROLLABILJTY OF LINEAR DYNAMICAL EQUATlONS 181
are linearly independent controllability. The condition for differential controllability can be easily
e conclude that the state obtained by slight modifications of Theorems 5-4 and 5-5. However, if the
Q.E.D. matrices A and B in the state equation E are analytic on ( - 00,00), then we have
the following theorem.
sufficient but not neces
Theorem 5-6
Ifthe matrices A and B are analytic on ( - 00, CJ:)), then the n-dimensional state
equation E is differentially controllable at every t in ( - 00, 00) if and only if,
for any fixed t o in ( - CJ:), (0),
(5-22) 11
If the matrix A is analytic on (- 00, 00), it can be shown that the state transi
tion matrix <I>(t o, .) of x= A(t)x is also analytic on (- 00, 00). Since the product
oftwo analytic functions is an analytic fundion, the assumption ofTheorem 5-6
implies that <I>(t o, ')BO is an analytic function. An implication of Theorem 5-3
is that a set of analytic functions is linearly independent on ( - 00,(0) if and only
if the set of analytic functions is linearly independent on every subinterval (no
matter how smaH) of ( - 00, (0). With this fact, Theorem 5-6 foHows immediately
from Theorems 5-3 and 5-4. Consequently, if a state equation with analytic
A and Bis controllable at any point at all, it is differentially controllable at every
tin(-oo,oo).
Definition 5-3
The linear dynamical equation E is said to be instantaneously controllable 5
in (- 00, (0) if and only if
- aH t 1=0, the dynamical
I for aH t in (-00, (0)
~ee distinct types of caD U a dynamical equation is instantaneously controHable, then the transÍer oí
s of continuity. the states can be achieved instantaneously at any time by using an input that
consists of b-functions and their derivatives up to an order of n - 1. It is clear
that instantaneous controllability implies differential controllability. The
most important implication of instantaneous controllability is that in the case
y) controllable at time t o of a single input, the matrix [Mo(t) MI(t) ... Mn-l(t)] qualifies as an
: Xl in L, there exists an equivalence transformation (Definition 4-5). Consequently, many canonical
itrarily small interval of form equivalent dynamical equations can be obtained for instantaneously
I
controHable dynamical equations. See References 10,99, 103, and 110.
state in a finite time (no
itrollable. If this can be
he state equation is said s In the engineering literature, it is called uni(orm controllability. However, this terminology was
first used in Reference 56 to define a difTerent kind of eontrollability (see Definition 5-4); hence
IcontroHability implies we adopt the terminology"instantaneous eontrollability."
182 CONTROLLABILlTY AND OBSERVABILlTY OF LINEAR DYNAMICAL EQUATIONS
Definition 5-4
The dynamical equation E is said to be uniformly controllable if and only if there
exist a positive (J e and positive CXi that depend on (J e such that
0< cx¡((Je)I ,:::;;W(t, t +(Je) ~cxz(aJI
and 0< cx 3 ((Je)I ~<l>(t +(J" t)W(t, t +(Je)<l>*(t +(J" t)':::;;cx 4 ((Je)I -3
for all t in ( - 00, (0), where <l> is the state transition matrix and W is as defined
in Equation (5-13). I
b (1)
t)saiaJI -3 -2 -1 o 2 3
W el ~ t eA'BB*é *, dr
Consider the inverted r
equation is developed il
2g(M +m)/(2M +m)l=
is nonsingular for any t> 0. 7
(3-42) becomes
3. The n X (np) controllability matrix
1
U~[B: AB: A 2 B:"': A"-lBJ (5-24 )
O
has rank n. O
4. For every eigenvalue A of A (and consequently for every ), in C), the n x (n + p)
O
complex matrix [Al -A : BJ has rank n. s
We compute
Proof
The equivalence of statements 1 and 2 follows directly from Theorems 5-1 and
5-4. Since the entries of e- AIB are analytic functions, Theorem 5-3 implies that u= [B Al
the rows of e - AIB are linearly independent on [0, (0) if and only if
p[e-AIB: -e-AIAB:"': (-l)"-le-AIA"-lB: "'J=n
This matrix can be read
for any t in [0, (0). Let t = O; then the equation reduces to Thus, if X3 = eis differer
to bring it back to zero.
p[B: - AB: . ..; ( - 1)" - lA" - 1 B: ( -1)" A"B : ...] = n
derivatives back to zer
From the Cayley-Hamilton theorem, we know that Am with m ¿ n can be written balancing a broom on (
as a linear combination of 1, A, ... , A"-l; hence the columns of AmB with
m "2: n are linearly dependent on the columns of B, Aa, ... ,A"-1B. Conse
quently, Example 7
p[B: -AB: ... : (-l)"-IA"-IB:' "J=p[B: -AB:"': (-I)"-1A"-lBJ Consider the platform s
of the platform is zero ;
Since changing the sign will not change the linear independence, we conclude
systems. The spring c(
that the rows of e-AIB are linearly independent if and only if p[B : AB : ...
coefficients are assumed
: A n -1 BJ = n. This proves the equivalence of statements 1 and 3. In the
X2 +X2 =u, or
foregoing argument we also proved that the rows of e - AIB are linearly inde
pendent if and only if the rows of eAIB are linearly independent on [0, (0) over
the field of complex numbers. Next we show the equivalence of statem.eEt:::
and 1'. Taking the Laplace transform of eAIB, we have
This is the state-variabh
5t'[e AI BJ = (si - A)-1 B Now if the initial di~
platform will oscillate,
Since the Laplace transform is a one-to-one linear operator, if the rows of
platform to come to re~
eAIB are linearly independent on [0, (0) over the field of complex numbers, so
are the rows of (si - A) -1 B, and vice versa. and X2(O) = -1, is it po~
.seconds? The answer d,
The proof of statement 4 will be postponed to Section 5-5 (page 206). Q.E. O.
applied to thetwo sprin
For the state equatic
p
7 The matrix is in raet positive definite.· See Problem E-II.
s This eondition implies that (sI - Al and B are left eoprime. See Appendix G. hence the state equation
~AL EQUATlONS CONTROLLABILlTY OF LINEAR DYNAMICAL EQUATIONS 185
Example 6
Consider the inverted pendulum system studied in Figure 3-15. Its dynamical
equation is developed in (3-42). For convenience, we assume 2mg/(2M +m) = 1,
2g(M +m)/(2M +m)l = 5, 2/(2M +m) = 1, and 1/(2M +m)l = 2. Then Equation
(3-42) becomes
x~l~
1 O
!}+[j}
(5-24 )
O -1
y=[1 O O OJx (5-25)
O O
O 5
We compute
r
(5-13) and (5-14), we have and u 1 (t)=-[0.5
In Figure 5-6 we a]
40
xl(O)=lO and X2(O) =
in transferringx(O) to 2
of the input. If no re:
x(O) to zero in an arbi
20
the input may become
u 2/ imposed, we might nc
3 ..... /4 interval of time. For
o
might not be able to tr
--- 1, seconds
Example 8
-20
Consider again the plal
the viscous friction coe
-40 a11 equal to 1. Then t
xl (1)
10 Clearly we have
............
........
............
1 ........... 2
...------- ....
3 4
o
2.5 1, seconds
and the state equation
an input to transfer x(C
X (1) xz(O), no input can tfar
2
* Controllabilíty índicE
10 Let A and B be n x
2 3 4
ol---l--+--I----+-+----t--+-----:,f--------.
-5 -- ,.,ti' 1, seconds It consists of k + 1 bloc
The matrix U ~ Un - 1
thenU n_ 1 has a rank 01
Note that there are a to
ways to choose these ni
ing the most natural anc
Figure 5-6 Behavior of x¡Ü) and X2(t) and the waveformof u.. Let b , ; = 1, 2, ... , p, be
;AL EQUATlONS CONTROLLABILlTY Of LINEAR DYNAMICAL EQUATlONS 187
Jt u. Using Equations
and ul(t) =- [0.5 l][e~'St ~] W-I(O, 2) [ ~~J = -44.1eo. St +20.7e t
r
1.6
- [ 6,33 27
6.33J for t in [0,2]. If a force of the form U I is applied, the platform will come to res t
at t = 2. The behavior of Xl' X2 and of the input u I are plotted by using solid
lines in Figure 5-6. •
In Figure 5-6 we also plot by using dotted lines the input U2(t) that transfers
XI(O) = 10 and X2(0) = -1 to zero in 4 seconds. We see from Figure 5-6 that,
in transferringx(O) to zero, the smaller the time interval the larger the magnitude
of the input. If no restriction is imposed on the input u, then we can transfer
x(O) to zero in an arbitrarily small interval of time~ however, the magnitude of
the input may become very large. If sorne restriction on the magnitude of u is
imposed, we might not be able to transfer x(O) to zero in an arbitrarily small
interval of time. For example, if we require iu(t)\.::; 5 in Example 5, then we
might not be able to transfer x(O) to zero in less than 4 seconds.
Example 8
Consider again the platform system shown in Figure 5-2. Now it is assumed that
the viscous friction coefficient and the spring constant ofboth spring systems are
all equal to 1. Then the state-variable description of the platform system is
Clearly we have
-lJ
-1
=1<2
and the state equation is not controllable. If XI(O) = X2(0), it is possible to find
an input to transfer x(O) to the zero state in a finite time. However, if X1 (O) 1=
X2(0), no input can transfer both Xl (O) and X2(O) to zeíO in a finite time. íill
*Controllability índices
Let A and B be n x n and n x p constant matrices. Define
It consists .of k + 1 block columns of the form AiB and is of order n x (k + l)p.
The niatrix U~ U _¡ is the controllability matrix. If {A, B} is controllable,
II
e controllability index
o 1
~l
O :, O
~l!
1
O O 2 : -1
1
[B' AB: A'B] O,, O 1 : -2
. O 1 : -2 O' O -4
has rank 4. This is the case; hence, the state eqilation in (5-32) is eontrollable.
190 CONTROLLABILlTY AND OBSERVABILlTY OF LINEAR DYNAMICAL EQUATIONS
We study more the controllability matrix U. We assume that the linearly The set of the controll~
independent columns of U in order from left to right have been found. We transformation and ar
now rearrange these independent columns as
Proof
b¡, Ab¡, ... , Al'l -¡ b¡, b 2, Ab 2, ... , AI'2-¡ b 2, ... , b p, Ab p, ... , Al'p-¡ b p
Define
The integer /1¡ is the number of linearly independent columns associated with b i
in the set, or is the length of the chain associated with b¡. Clearly we have
where A=PAP-¡, Ü.
/1 = max{/1¡, /12, ... , /1 p}
easily verified that
and /1¡ +/12 + ... + /1p::;; n. The equality holds if {A, B} is controllable. The
set {/1¡, /12, ... ,/1 p} will be called the controllability indices of {A, B}.
Now we shall establish the relationship between the controllability indices which implies
and the r¡'s defined in (5-28). In order to visualize the relationship, we use an
example. We assume p = 4, /1¡ = 3, /12 = 1, /13 = 5, and /14 = 3. These indepen
dent columns are arranged in a crate diagram as shown in Figure 5-7. The Hence, the r¡ defined
(i, j)th cell represents the column Ai - ¡bj- A column which is linearly indepen indices are invariant Ul
dent of its left-hand-side columns in (5-28) is denoted by "x"; otherwise denoted The rearrangement
by "O". The search of linearly independent columns in (5-28) from left to right
is equivalent to the search from left to right in each row and then to the next
row in Figure 5-7. Hence, the number of zeros in the ith row of Figure 5-7 is where M is a p x p eler
equal to ri-¡ as shown. From the crate diagram, we can deduce that r¡ is equal forward to verify that
to the number of {bk , k = 1, 2, ... ,p} with controllability indices equal to or - f:, - .
smaller than i. Hence, we conclude that Uk=[B: 1
r¡ - r¡_¡ = no. of {bk , k = 1, 2, ... , p} with controllability index i (5-33) where diag {M, M, ...
singular. Hence, we h¡
with l' _ d~O. For example, r¡ - ro = 1, and b 2 has controllability index 1;
1'2 - r¡ =0 and no b¡ has controllability index 2; 1'3 - 1'2 = 2, and b¡ and b4 have
controllability index 3; 1'5 -1'4 = 1, and b 3 has controllability index 5. Hence, Consequently. we cane
the ordering ol" the eoll
b¡ b2 b3 b4
Now we discuss a (
1 x x x x ro =0 umns in U = [B AB
A x O x x r¡ = 1
A2 X O X X 1'2 = 1
A3 O O X O 1'3 =3
A4 O O X O 1'4= 3 and then search its linl
AS O O O O In terms af the crate di
1'5 =4
the controllability indices of {A, B} are uniquely determinable from tri, i =0,
1, ... , Ji}.
I Theorem 5-8
The set ofthe controllability indices of {A, B} is invariant under any equiva1ence
ume that the linearly
transformation and any ordering of the columns of B.
ave been faund. We
Proof
Ab p, ... , AJ.Lp-l b p
Define
nns associated with b¡
]early we have
Ük=[B: AB:"': AkBJ
where A= PAP- , l
B= PB and P is any nonsingular matrix. Then it can be
easily verified that
is controllable. The
for k = O, 1, 2, ...
es of {A, B}.
controllability indices which implies
~lationship, we use an
for k =0,1,2, ...
. = 3. These indepen
11 in Figure 5-7. The Hence, the r¡ defined in (5-28) and, consequently, the set of controllability
;;h is linearly indepen indices are invariant under any equivalence transformation.
x"; otherwise denoted The rearrangement of the columns of B can be represented by
;-28) from left to right
B=BM
, and then to the next
h row of Figure 5-7 is where M is a p x p elementary matrix and is nonsingular. Again it is straight
deduce that ri is equal forward to verify that
ty indices equal to or
iJd~~ [B
: AB : ... : AkBJ = U k diag{M, M, ... , M}
where diag {M, M, ... , M} consists of k + 1 number of M, and is clearly non
ability index i (5-33)
singular. Hence, we have
mtrollability index 1;
for k =0, 1,2, ...
2, and b¡ and b 4 have
)i1ity index 5. Hence, Consequently, we conclude that the controllability indices are independent of
the ordering of the columns of B. Q.E.D.
and then search its linearly independent columns in order from left to right.
In terms of the erate diagram in Figure 5-7, the linearly independent columns
are searched in order from top to bottom in the firstcolumn, then in the second
column and so forth. Let .
b l' Ab b ... , Aii,-lb l'. b 2' Ab 2, ... , Aii2- l b"
2, ... , b P' Ab P' ... , Aiip-lb p
...:.=---_._-_.•_----
-----~ -----=-------=-:..:::...-.::--:..:'--- - - - - - - - - - - ' -.." ' - - - , - - - - - -- - - ----- _. -------------------,~=~
be the resulting linearly independent columns. If {A, B} is controllable, we have Consider the n-dirr
ji ¡ + ilz + . .. + il p = n. The lengths of these chains are {il¡, ilz,···, il p }·
Unlike the controllability indices, these lengths of chains depend highly on the
ordering of {b¡, i = 1,2, ... , p}.
where A, R, e, and E a
Example 10 continuous functions (
Consider the state equation in (5-32). If we search linearly independent columns
from left to right in Definition 5-5
b¡, Ab¡, AZb¡, A 3 b¡; b z, Ab z, AZbz, A3 b z The dynamical equati
if there exists a finite tI
the resulting linearly independent columns are of the input U[lo,l,] and
b¡, Ab¡, AZb¡, b z to determine the state
unobservable at too
Its lengths are {3, 1}. If we search from left to right in
Example 1
b z, Ab z, AZb z, A 3 b z ; b¡, Ab¡, AZb¡, A3 b¡
Consider the network ~
the resulting linearly independent columns are the initial voltage acros
b z, Ab z, AZb z, A3 b z the output is identicall:
cally zero), but we are n
Its lengths are {4, O}. The lengths are indeed different for different ordering of hence the system, or m
b¡ and b z. The controllability indices of (5-32) can be computed as {2, 2} and system, is not observab
are independent of the ordering of b¡ and b z. I
Example 2
Let il = max{il¡, ilz, ... ,ilp }. It is clear that il can never be smaller than /1, Consider the network sl
the controllability indexo Since /1 = max{/1¡, /1z, ... , /1p}, we may conclude reduces to the one sho'
that /1 is the smallest possible maximum length of chains obtainable in any
search of linearly independent columns of U. This controllability index will
play an important role in the design of state feedback in Chapter 7 and the design
of compensators in Chapter 9. In In
lF
+
urv +
Example 1
Consider the network shown in Figure 5-8. lfthe input is zero, no matter what
the initial voltage across the capacitor is, in view ofthe symmetry ofthe network,
the output is identically zero. We know the input and output (both are identi
cally zero), but we are not able to determine the initial condition ofthe capacitor;
for different ordering of hence the system, or more precisely, the dynamical equation that describes the
computed as {2, 2} and system, is not observable at any to.
111
Example 2
never be smaller than /1, Consider the network shown in Figure 5-9(a). Ir no input is applied, the network
, /1p}, we may conclude reduces to the one shown in Figure 5-9(b). Clearly the response to the initial
lains obtainable in any
ontrollability index will
::::hapter 7 and the design
:luations
y
md E are known before
nt from the problem of
ication is a problem of
' -_
F
_---4-----_
1
1 L _ _ ~------JIF 1
(a)
~mation collected at the (b)
current in the inductor can never appear at the output terminal. Therefore, Prove by contradictio
there is no way of determining the initial current in the inductor from the input ti> t o such that the co
and the output terminals. Hence the system or its dynamical equation is not Then there exists an n
observable at any too I
The response of the dynamical equation (5-35) is given by Let us choose x(to) = cx.
where l1>(t. r) is the state transition matrix of X = A(t)x. In the study of observ tion that E is observa!
ability. the output y and the input u are assumed to be known, the initial state ti> t o such that the col
x(t o) is the only unknown: hence (5-36) can be written as
y(t) = C(t)<1>{t. t o)x(t o) (5-37) We see from this th~
tion depends only on e
where y(t)~ y{t) - C(t) Ir l1>{t. r)B(r)u(r) dr - E(t)u(t) (5-38) can also be deduced [re
lO
servability study, it is s
is a known function. Consequently. the observability problem is a problem of x = A(t)x, Y = C(t)x.
determining x(to) in (5-37) with the knowledge of y, C, and l1>(t. to). Note that The controllability
the estimated state x(t o) is the state not at time t, but at time too However, if independence of the rol1
x(to) is known, the state after t o can be computed from by the linear independ
between these two conc
x(t) =l1>(t, to)x(to) + Irlo l1>(t, r)B(r)u(r) dr (5-39)
Theorem 5-10 (Theort
Theorem 5-9 Consider the dynamiea
The dynamical equation E is observable at t o if and only if there exists a finite defined by
ti> t o such that the n columns of the q x n matrix function C(' )<1>(', to) are linearly
independent on [t o, ti].
(5-41 )
Xo = V- 1 (t o, tI) 1
1
'
lo
l1>*(t, to)C*(t)y(t) dt (5-42) We list in thefollowir
tions 5~6 to 5-8, which a
Thus, ifthe function y[lo, Id is known,x o can be computed from (5-42). Necessity: 5-4 for controllability. ~
OBSERVABILITY 01' LINEAR DYNAM1CAL EQUATIONS 195
CAL EQUATIONS
terminal. Therefore, Prove by contradiction. Suppose E is observable at to, but there exists no
Iductor from the input ti> t o such that the columns of COP(', to) are linearly independent on [to, ti].
lmical equation is not Then there exists an n x 1 nonzero constant vector a such that
:n by
Let us choose x(to) = a; then
Y(t) = C(t)<.l>(t, to)a = O for al1 t> t o
11; + E(t)u(t) (S-36)
Hence the initial statex(t o) = a cannot be detected. This contradicts the assump
[n the study of observ
tion that E is observable. Therefore, if E is observable, there exists a finite
;nown, the initial state ti> t o such that the columns of C(· )<.1>(', t o) are linearly independent on [t o' ti].
Q.E.D.
(S-37 ) We see from this theorem that the observability of a linear dynamical equa
tion depends only on C(tl and <I>(t, to) or, equivalently, only on C and A. This
(S-38) =
can also be deduced from Definition 5-5 by choosing u O. Hence in the ob
servability study, it is sometimes convenient to assume u =0 and study only
'oblem is a problem of X= A(t)x, Y= C(t)x.
nd <I>(t, to). Note that The controllability of a dynamical equation is determined by the linear
t time to. However, if independence of the rows of <I>(to, .)B('), whereas the observability is determined
by the linear independence of the columns of q. )<1>(', t o). The relationship
between these two concepts is established in the following theorem.
d, (5-39 )
Theorem 5-10 (Theorem of duality)
Consider the dynamical equation E in (5-35) and the dynamical equation E*
defined by
y if there exists a finite
IC(' )<1>(', t o) are linearly E*: z= - A*(t)z +C*(t)v (S-43a)
1 =B*(t)z +E*(t)v (S-43b)
where A*, B*, C*, and E* are the complex conjugate transposes of A, B, C, and
E in E. The equation E is control1able (observable) at t o if and only if the
(5-37) and integrating equation E* is observable (controllable) at too
Proof
1 dtJxo~ Veto, t1)xo From Theorem 5-4, the dynamical equation E is controllable if and only if the
(S40) rows of <I>(to, t)B(t) are linearly independent, in t, on [to, ti]. From Theorem
5-9, the dynamical equation E* is observable if and only if the columns of
) dt (541) B*(t)<I>a(t, to) are linearly independent, in t, on [to, t 1J, or equivalently, the rows of
[B*(t)<I>a(t, to)]* = <I>:(t, to)B(t) are linearlY independent, in t, on [to, tlJ, where
lmns of q.)<I>(', t o) are <l>a is the state transition matrix ofi = - A*(t)z. lt is easy to show that <I>:(t, to) =
<I>(t o, t) (see Problem 4-8); hence E is control1able if and only if E* is observable.
lo, t d is nonsingular.
Q.E.D.
(S-42)
We list in the following,for observability, Theorems 5-11 to 5-14 and Defini
dt
tions 5-6 lo 5-8, which are dual to Theorems5-5 to 5-8 and Definitions 5-2 to
·rom(5-42). Necessity: 5-4 for control1ability. Theorems 5-11 to 5-14 can be proved either directly or
i
I
L
196 CONTROLLABILlTY ANO OBSERVABILlTY OF LINEAR OYNAMICAL EQUATIONS
by applying Theorem 5-10 to Theorems 5-5 to 5-8. The interpretations in the Definition 5-7
controlIability part also apply to the observability part.
The linear dynamical
Theorem 5-11 ( - 00, 00) if and only i
Assume that the matrices A(') and C(-) in the n-dimensional dynamical equation
E are n - 1 times continuously differentiable. Then the dynamical equation E
is observable at t o if there exists a finite ti> t o such that
Definition 5-8
Definition 5-6
The dynamical equation E is said to be difJerentially observable at time t o if, where A, B, C, and E al
for any state x(t o) in the state space L, the knowledge of the input and the output time interval of interest
over an arbitrarily smalI interval of time suffices to determine x(t o). I time-invariant dynamic
t o ~ 0, and the determino
Theorem 5-12 time interval. Hence, tI:
bility study of linear tirr.
Ir the matrices A. and iC are analytic on (-oo. o')), then the i1-o.imensioi1cJ ay·
namical equation E is differentialIy observable at every t in ( - 00,00) if and only
if, for any fixed t o in ( - 00, 00), Theorem 5 -1 3
No(to) The n-dimensional linea:
NI (to) able if and only if any 01
1. AH columns of CeA!
p =n complex numbers.
N n - 1(to) 1'. AlI columns of C(sI
2. The observability gra
I
.
~
_
(lCAL EQUAUONS OBSERVABILlTY OF LINEAR DYNAMICAL EQUATIONS 197
l~;~:;
( - 00, ro) if and only if
Definition 5-8
The linear dynamical equation E is said to be uniformly observable in ( - ro, (0)
~, ... ,n-1 (5-45a)
if and only if there exist a positive (Jo and positive f3i that depends on (Jo such that
(5-45b) 0< f31((Jo)I.:::::;V(t, t +(Jo).:::::;f3z((Jo)1
I
0< f33((Jo)I ':::::;$*(t, t +(Jo)V(t, 1 +(Jo)$(t, t + (Jo) .:::::;f34((Jo)I
Ity, for all t, where $ is the state transition matrix and V is as defined in (5-41). 11
: can be defined by using Time-invariance case. Consider the linear time-invariant dynamical
~, C} to be differentially equation
controllable. However,
he following. FE: x=Ax +Bu (5-46a)
y=Cx +Eu (5-46b)
t o ::::: O, and the determination of the initial state can be achieved in any nonzero
time interval. Hence, the reference of t o and t 1 is often dropped in the observa
bility study of linear time-invariant dynamical equations.
:n the n-dimensional dy·
t in (-ro, (0) ifand only
Theorem 5-13- _
The n-dimensional linear time-invariant dynamical equation in (5-46) is observ
able if and only if any of the foHowing eq uivalent conditions is satisfied:
1. AH columns of CeA! are linearly independent on [O, ro) over e, the field of
complex numbers.
Equivalently, v is the
(5-47) pV o <
The integer v is called
have
has rank n.
4. For every eigenvalue Aof A (and consequently for every A in C), the (n +q) x n
complex matrix where ñ i5 the degree (
Corollary 5-13
has rank n, or equivalently, (sI - A) and e are right coprime. The dynamical equatif
V n _ q , where q is the n
·Observability Indices V:_qVn_q is nonsingul¡
1
in order from top to bo
l
independen t rows asso
C2
C: ro (no. of dependent rows) the observability indice!J
Cq
Theorem 5 -14
(5-48)
The set of observability
transformation and any
~svsmin(ñ,n-q+l) (5-51)
q
:y .A. in (:), the (n + q) x n
where ñ is the degree of the minimal polynomial of A and q is the rank of C.
Corollary 5-13
The dynamical equation FE in (5-46) is observable if and only if the matrix
coprime. Vn _ q , where q is the rank of C, is of rank n, or equivalently, the n.X n matrix
V:_qVn_q is nonsingular. I
ine Consider the matrix V n _ ¡. lt is assumed that its linearly independent rows
in order from top to bottom have been found. Let Vi be the number of linearly
independent rows associated with e i. The set of {Vi' i = 1,2, .. , , q} is cal1ed
the observability indices of {A, C}. Clearly we have
l rows)
v = max [Vi, i = 1, 2, ... , q l
Theorem 5-14
(5-48) The set of observability indices of [A, el is invariant under any equivalence
transformalion and any ordering or the rows of C. I
Since the rank of a matrix does not change after multiplication of a nonsingular where the last n - n¡ coi
matrix (Theorem 2-7), we have rank U = rank Ü. Consequently FE is con Q is nonsingular. We c
trollable if and only if FE is controllable. A similar statement holds for the into the form of (5-54).
observability part. x = Px we are actually u:
state space. The ith coll
Theorem 5-15 Aq¡ with respect to {q¡, ~
linearly dependent on th
The controllability and observability of a linear time-invariant dynamical
given in (5-54a). The c(
equation are invariant under any equivalence transformation. I
B with respect to {q¡, (
{q¡, qz,· .. , q!!.l}; hence
This theorem is in fact a special case of Theorems 5-8 and 5-14. For easy
Let U and U be the co
reference, we have restated it as a theorem.
we have pU = pU = n¡ (s
In the following, e will be used to stand for controllable, efor uncontrollable,
o for observable, and (5 for unobservable. - [i
U=
Theorem 5-16
Consider the n-dimensional linear time-invariant dynamical equation FE.
If the controllability matrix of FE has rank n ¡ (where n¡ < n); then there exists
=[
an equivalence transformation X = Px, where P is a constant nonsingular where
-k-
Ve
represents the
.
matrix, which transforms FE into AcB wlth k ~n¡ are line
pU =n¡ implies pUc = nI'
FE:
[;:J = [:' t J[::J
z
+ [~'J u (5-54a)
We show now that the
FE and FiL. hav~ the ss
is controllable 9 and has the same transfer function matrix as FE. = [Ce Ce] I
=CcCsI -Ac
9 It is easy to show that ¡f¡he equation FE ís observable. then its subequation FE, is also observable.
(Try) which is the transfer-func
IlCAL EQUATlONS CANONICAL DECOMPOSITlON OF A LINEAR T1ME-INVARIANT DYNAMICAL EQUATION 201
Proof
(5-53a) If the dynamical equation FE is not controllable, then from Theorem 5-7 we
(5-53b) have
¡e dynamical equations pU~p[B : AB : ... : An-¡BJ =n¡ < n
is caBed an equivalence Let q¡, q2' ... , qn, be any n¡ linearly independent columns of U. Note that for
each i = 1, 2, ... , n¡, Aq¡ can be written as a linear combination of
: PAn-lB] {q¡, Q2,"" qnJ (Why?) Define a nonsingular matrix
~n-¡B]~ PU (5-56)
¡cation of a nonsingular where the last n - n¡ columns of Q are entirely arbitrary so long as the matrix
msequently FE is con Q is nonsingular. We claim that the transformation ji: = Px will transform FE
;tatement holds for the into the form of (5-54). Recall from Figure 2-5 that in the transformation
ji: = Px we are actuaBy using the columns of Q ~ p-¡ as new basis vectors of the
state space. The ith column of the new representation A is the representation of
Aq¡ with respect to {q¡, q2' ... , qn}' Now the vectors Aq¡, for i = 1,2, ... , n 1> are
linearly dependent on the set {q¡, q2' ... , qn,}; hence the matrix A has the form
le-invariant dynamical given in (5-54a). The columns of B are the representations of the columns of
¡ation. • B with respect to {q¡, q2"'" qn}' Now the columns of B depend only on
{q¡, q2' ... , qn,}; hence B is of the form shown in (5-54a).
;-8 and 5-14. For easy Let U and Ü be thecontrollability matrices of FE and FE, respectively. Then
we have pU = pÜ = n¡ (see Theorem 5-15). It is easy to verify that
lle, efor uncontrollable,
- .[BeJ~ AeBe !I • . • !I A~-¡BeJ
U=
O: O: 1 O
=[Üe! A~'B:¡ ... ·¡A~-¡BeJ}n¡ rows
namical equation FE. O¡ O : : O }(n-n¡)rows
~¡ < n), then there exists
where Üe represents the controBability matrices of F Eco Since columns of
constant nonsingular
A~B with k ?:::n¡ are linearly dependent on the columns of Ü e, the condition
pU = n¡ implies pÚe = ni' Hence the dynamical equation F Ee is controllable.
We show now that the dynamical equations FE and FEe , or correspondingly,
FE and FEc • have the same transfer-functíon matrix. It is easy to verify 1:h3.t
(5-54b) (5-57)
'ix as FE.
= [Ce C-J [(Si - Aer 1 (si -AJ-l A:?lSI - Ac)-¡J [Be] +E
e O (si - Acr ¡ O
=Ce(sl -AJ-¡Be +E
ualion FE, is <j.lso observable.
which is the transfer-function matrix of FEe . Q.E.D.
202 CONTROLLABILlTY AND OBSERVABILlTY OF LINEAR DYNAMICAL EQUATIONS CANONICAL DECOMPO;
subspace, which consists of all the vectors [:J Since F Ec is controllable, all
Hence, the reduced ce
Xc =
the vectors Xc in Ll are controllable. Equation (5-54a) shows that the state
variables in Xc are not affected directly by the input u or indirectly through the
state vector xc; therefore, the state vector Xc is not controllable and is dropped in
the reduced equation (5-55). Thus, if a linear time-invariant dynamical equation Dual to Theorem
is not controllable, by a proper choice of a basis, the state vector can be decom dynamical equations.
posed into two groups: one controllable, the other uncontrollable. By
dropping the uncontrollable state vectors, we may obtain a controllable dy Theorem 5-17
namical equation of lesser dimension that is zero-state equivalent to the
original equation. See Problems 5-22 and 5-23. Consider the n-dime
If the observabilitv m:
valence transform~tio:
Example 1
Consider the three-dimensional dynamical equation FE
1
1
y=[1 1 l]x (5-58)
1
and the nrdimensiom
The rank of B is 2; therefore, we need to check U 1 = [B : AB] in determining
the controllability of the equation. Since
1
is observable and has 1
1
1 This theorem can l
the state equation is not controllable. The first n2 rows of P i
Let us choose, as in (5-56), the observability matr
arbitrary so long as P
ro 1 I 1l Xo does not appear eh
p-l = Q = 11 o¡oJ vector Xi; is not observ:
Lo 1 lo Combining Theore
theorem.
The first two columns of Q are the first two linearly independent columns of
U 1; the last column of Q is chosen arbitrarily to make Q nonsingular. Let
Theorem 5-18 (Cane
x=Px. We compute
m~
Consider the linear tirr
pace ~ of FE is divided O 1
;e of~, denoted by L. l , and 1 1] 1 O
[
le (n - nl)-dimensional O 1
Hence, the reduced controllable equation is
FE c is controllable, all
1) shows that the state
. indirectly through the
Xc = G ~] Xc + [~ ~] u y = [1 2Jx c (5-59 )
FE: (5-60a)
lO This is a simplified version of the canonical decomposilion theorem. For the general form, see
References 57, 60, and 116. See also Reference S127.
---~._---------
which depends solely on the controllable and observable part of the equation Example 2
FE.
Consider the networl
source, the behavior
Proof
detected from the ou
If the dynamical equation FE is not controllable, it can be transformed into the L 1 are not observabl{
form of (5-54). Consider now the dynamical equation FE c which is the con the state variable aSSt
trollable part of FE. If FE c is not observable, then FE c can be transformed into metry, the state varial
the form of (5-60), which can also be written as By dropping the state
the network in Figun
[~caJ [AOco
X co
= ~12J[~caJ
A co X
+[~caJ
co B
u
co
the transfer function
LI In
!l
J/
u
ca
y
I
I ni
1------1
I I
I I
I 1
I I
I I
:e :
10
1
I
I
I
CO
1
I
1
I
L -.J
Before moving to the next topic, we use Theorem 5-16 to prove statement 4
of Theorem 5-7.
lbta;,., Equattou (5-62)
Ilclude that the transfer Cl
R
Q.E.D.
+
Li Lz
In In
" f c2
y
In In
(a)
In In
y
In In
(b)
equation still has the same zero-state response, This fact motivates the follow
ing definition.
, the eigenvalues of A;
~xcept possibly at the
Definition 5-9
,s controllable, then
:re exist an eigenvalue A linear time-invariant dynamical equation FE is said to be reducible if and
only if there exists a linear time-invariant dynamical equation of lesser dimen
sion that has the same transfer-function matrix as FE or, equivalently, is zero
state equivalent to FE. Otherwise, the equation is said to be irreducible.
Theorem 5-19
A linear time-invariant dynamical equation FE is irreducible if and only if FE
is controllable and observable.
Proof
"-la.BJ=O If the dynamical equation FE is either uncontrollable or unobservable, then FE
is reducible (Theorems 5-16 and 5-17). Hence what we have to prove is that if
f. Hence, if {A, B} is FE is controllable and observable, then FE is irreducible. We prove this by
lue of A and, conse
contradiction. Suppose that the n-dimensional equation FE is controllable
and observable and that there exists a linear time-invariant dynamical equation
t1 p[ Al - A : BJ < n
FE,
:, there exists an equi
S} with FE: x=Ax +Su (5-63a)
y =Cx +Eu (5-63b)
of lesser dimension, say ni < n, that is zero-state equivalent to FE. Then, from
Theorem 4-6, we have E = E and
tí IJA c = AfI. Now we
CAkB=CAkS k=O,I,2,.,. (5-64)
l
-1: BJ
ular, a,(AI - A)P-l =0
CB CAn-lB
CAB CAnB
(5-65)
[
~ : BJ does not have a = CAn-lB CAnB CA2(~-1)B
proof. Q.E.D.
By (5-64), we may replace CAkB in (5-65) by CAkB; consequently, we have
1m Theorems 5-16 and (5-66)
'VU=Vn-1Ü n- l
¡ either uncontrollable
n of lesser dimension where V"-l and Ü"-l are defined as in (5-48) and (5-27)., Since FE is control
II dynamical equation. lable and observable, we have p U = n and pV = n, lt follows 'from Theorem
Ion is either uncontrol 2-6 that p(VU) = n. Now Vn - l and Ü n - 1 are, respectively, qn xn¡ and ni x np
:uch that the reduced matrices; hence the matrix V"-lÜ"-l has a rank of at most ni' However,
208 CONTROLLABILlTY AND OBSERVABILlTY OF LINEAR DYNAMICAL EQUATIONS CONTROLLABILlTY A1
(5-66) implies that PeY n _ 1 Ü n _ ¡) = n > ni' This is a contradiction. Hence, if invariant dynamical e
FE is controllable and observable, then FE is irreducible. Q.E. O. is referred to Referen<
Recal1 from Section 4-4 that if a dynamical equation {A, B, C, E} has a pre
scribed transfer-function matrix G(s), then the dynamical equation {A, B, C, E} *5-6 Controllabi
is cal1ed a realizatian of G(s). Now if {A, B, C, E} is control1able and observable, Dynamical Equa
then {A, B, C, E} is cal1ed an irreducible realizatian of G(s), In the following we
shall show that all the irreducible realizations ofG(s) are eql..livalent. The controllability
equation are invarial
Theorem 5-20 ceivable that we ma~
by transforming the
Let the dynamical equation {A, B, C, E} be an irreducible realization of a
equation is in a lord:;
q x p proper rational matrix G(s). Then {A, B, C, E} is also an irreducible
almost by inspection
realization of G(s) if and only if {A, B, C, E} and {A, B, C, E} are equivalent;
Consider the n-d
that is, there exists a nonsingular constant matrix P such that A = PAP-1,
equation
B = PB, C = CP -1, and E = E.
Proof
The sufficiency fol1ows directly from Theorems 4-6 and 5-15. We show now the where the matrices A
necessity of the theorem. Let U, V be the controllability and the observability The n x n matrix A is
matrices of {A, B, C, E}, and let Ü, V be similarly defined for {A, B, C, E}. If A.m' A¡ denotes aH tb
{A, B, C, E} and {A, B, C, E} are realizations of the same G(s), then from (5-64) the number of lordan
and (5-65) we have E = E,
A¡ = diag (Ái!, A
VU=VÜ (5":67)
and VAU=VAÜ (5-68) Table 5-1 Jordan-Fu
The irreducibility assumption implies that pV = n; hence the matrix (V*V) is
nonsingular (Theorem 2-8). Consequently, from (5-67), we have
Ü = eY*V) - 1 V*VU ~ PU (5-69) A
(n x 11
where P ~ eY*V)-l V*V. From (5-69) we have p Ü .=s; min (pP, pU), which,
together with pÜ = n, implies that pP = n. Hence P qualifies as an equivalence (
This theorem implies that al1 the irreducible realizations of G(s) have the
"same dimensiono Physically, the dimension of an irreducible dynamical equa Aij
tion is the minimal number of integrators (if we simulate the equation in an (l1¡j x nij
" analog computer) or the minimal number of energy-storage elements (if the
system" Is an RLC n"etwork) required to generate the given transfer-function
matrix.
We studied in this section only the canonical decomposition of linear time-
~AL EQUATIONS CONTROLLABILITY AND OBSERVABILITY OF JORDAN-FORM DYNAMICAL EQUATIONS 209
tradiction. Hence, if invariant dynamical equations. For the time-varying case, the interested reader
Q.E.D. is referred to References 106 and 108.
A, B, e, E} has a pre
equation {A, B, e, E} *5-6 Controllability and Observability of Jordan-Form
llable and observable, Dynamical Equations
1, In the following we
15. We show now the where the matrices A, B, and e are assumed of the forms shown in Table 5-1.
and the observability The n x n matrix A is in the Jordan form, with m distinct eigenvalues A¡, ,,1,2' ..• ,
d for {A, D, e, E}. If Am • Ai denotes all the Jordan blocks associated with the eigenvalue Ai; r(i) is
(;(5), then from (5-64) the number of Jordan blocks in A¡; and A¡j is thejth Jordan block in A¡. Clearly,
A i = diag (A¡l' A i2 , ... , Air(i») and A = diag (Al' Az, ... , A m)
(5-67)
(5-68) Table 5-1 Jordan-Form Oynamical Equation
e the matrix W*V) is
ve have
ifies as an equivalence
(n
A
x n)
lA'
C=[C I C 2
A2
J (n~p) ~[J
Cm]
-,
= PE. Since pU = n,
A¡
(ni x n¡)
{U A'2
A,j
I
B. _ Bi2
(n, "p) -
lB" 1J
,
.
BHj )
erive from (5-67) and
Cj =[C i1 C'2 Cir(il]
CP and A=PAP- l .
(q x n¡)
Q.E.D.
Let n¡ and nij be the order of A¡ and A ij , respectively; then observable are that tI
m m r{i) linearly independent.
n= I
i=I
ni = I I
¡=lj=l
nij dent, the set {C IZ1 }'
Hence J F E is not ob:
Corresponding to A¡ and A¡j, the matrices B and C are partitioned as shown.
The first row and the last row of Bij are denoted by bl¡j and b lij , respectively. The conditions 1
The first column and the last column of Cij are denoted by Clij and c lij . required that each of
pendence. The linee:
Theorem 5-21 Furthermore, the rov
determining the cont]
The n-dimensional linear time-invariant Jordan-form dynamical equation J FE The physical mea
is controllable if and only if for each i = 1, 2, ... , m, the rows of the r(i) x p the block diagram o:
matrix studying the general
Jordan-form dynamic
an integrator and a fe.
(5-71 a) block, or more preci:
variable. Each chain
Consider the last cha
are linearly independent (over the field of complex numbers). J FE is observable variables in that chail
if and only if for each i = 1, 2, ... , m, the columns of the q x r(i) matrix that chain can be obs(
same eigenvalue, the
(5-71b) vectors of these chair
are linearly independent (over the field of complex numbers). 11 be studied separately.
Example 1
Consider the Jordan-form dynamical equation
Al 1: O O O O O O O O
O Al: O O O O O 100 ~bl11
------
-0- - 6- :-l~: O O O O O 1 O ~bl1z
------
JFE: x= O O -ü-:-il -:
___ .L
O O O x
_
+ O O 1 U~bl13 (5-72a)
------
O O O O: Az 1 O 1 1 2
lo O °O °O O: O
O: O
Az
O
1 J'
Az l~ ~ ~J
11;2:0;020]
y= [ 110:1:2:02121 x (5-72b)
0:2:3:0
i i i i
The matrix A has two distinct eIgenvalues Al and Az- There are three Jordan
blocks associated with Al; hence r(l) = 3. There isonly one Jordan block associ
ated with Az ; hence r(2) = 1. The conditions for J FEto be controllable are that
the set {blll , b l1z , b ll3 } and the set {bIZl } be, individually, linearly independent.
This is the case; hence JFE is controllable. The conditions for JFE to be Figure 5-12 Block di¡
AL EQUATIONS CONtROLLABILITY AND OBSERVABILITY OF JORDAN-FORM DYNAMICAL EQUATIONS 211
observable are that the set {C III , C l1 2> CI 13 } and the set {C 12 ¡} be, individually,
linearly independent. Although the set {C III , C 112 , C I13 } is linearly indepen
dent, the set {C 12 ¡}, which consists of a zero vector, is linearly dependent.
Hence J FE is not observable. E
mrtitioned as shown.
and b 1ij, respectively. The conditions for controllability and observability in Theorem 5-21
/ C I ij and Clij' required that each of the m set of vectors be individually tested for linear inde
pendence. The linear dependence of one. set on the other set is immaterial.
Furthermore, the row vectors of B excluding the bli/s do not play any role in
determining the controllability of the equation.
amical equation JFE The physical meaning of the conditions of Theorem 5-21 can be seen from
~ rows of the r(i) x p
the block diagram of the Jordan-form dynamical equation JFE. Instead of
studying the general case, we draw in Figure 5-12 a block diagram for the
Jordan-form dynamical equation in (5-72). Observe that each block consists of
an integrator and a feedback path, as shown in Figure 5-13. The output of each
(5-71 a) block, or more precisely the output of e.ach integrator, is assigned as a state
variable. Each chain of blocks corresponds to a Jordan block in the equation.
Consider the last chain of Figure 5-12. We see that if b l21 #=0, then aH state
;). J FE is observable variables in that chain can be controHed; if C l21 #=0, then al! state variables in
>( r(i) matrix
that chain can be observed. Ir there are two or more chains associated with the
same eigenvalue, then we require the linear. independence of the first gain
(5-71 b) vectors of these chains. The chains associated with different eigenvalues can
:rs). 11 be studied separately.
)
y
) ~b111
) ~b112
-
1 U~b!13 (5-72a) u
y
2
);
1J ~b121 u
y
(5-72b)
u
y
tere are three Jordan
Jordan block associ
controlÍable are that
linearly independent.
tions for J FEto be Figure 5-12 Block diagram of the Jordan-form equation (5-72).
212 CONTROLLABILITY AND OBSERVABILITY OF LINEAR DYNAMICAL EQUATIONS
CONTROLLABIl.ITY A~
linearly independent.
theorem can be estab
(5-73) Example 2
The matrix A has two distinct eigenvalues Al and Az. There are two Jordan Consider the single-v~
blocks associated with Al' one associated with A z. If s =A b (5-73) becomes
O -1 O b ll1 o :
O O -1:
. b Zl1
x=
. o (
O O O:
- - - - - - - - - - - - - - - : - Ó- - - - =---( - -:
,
: O O:
,
b'12 I
b 111
b 11Z (5-74)
r~- -~
There are two distinc
-----------T----------------T-----
, Al -A z I
- 1 , b 1Z1 I
responds to the last
A1 -A z : b lZ1 _i is zero; therefore, the
I
corresponding to the fl
By a sequence of elementary column operations, the matrix in (5-74) can be therefore, the equatior
transformed into
, Example 3
O -1 O: O
O O -1: O Consider the followin!
I
O O O: b l11
- ------------- -: - ó-- -- =---( - -: O (5-75)
~ _? 0 \ :__ ~l!! _
I Al -A z O, O and
I •
A1 -A z : O
That the state eqmi.ti,
Note that Al -A z is different from zero. The matrixin (5-75), or equivalently, Equation (5-77) is a t
the matrix in (5- 73) at s = Al' has a full row rank if and only if b 1ll aild b 11Z are matrix is in the Jordar
ICAL EQUATIONS CONTROLLABlUTY AND OBSERVABILlTY OF JORDAN-FORM DYNAMICAL EQUATlONS 213
(5-73) Example 2
There are two Jordan Consider the single-variable Jordan-form dynamical equation
becomes
=,1.1' (5-73)
~r--H\~l ~1}
I
b lll
O 1 O: O 10
1
b 211
b l11
x=
. O O 1 0
I x+ 9
y = [1 O O: lJx
b l12
(5-74)
, bL12
There are two distinct eigenvalues O and 1. The component of B which cor
-T-----
: b l21 responds to the last row of the Jordan block associated with eigenvalue O
: : b 121 _
is zero; therefore, the equation is not controllable. The two components of e
corresponding to the first column of both Jordan blocks are different from zero;
aatrix in (5-74) can be therefore, the equation is observable. 11
Example 3
o
O Consider the following two Jordan-form state equations:
., b
o
,
--------
, O
o
l12
and [XIJ
X2 =
[-1
O 0J x + [e- J
-2 e-21
I
u (5-77)
l·
o
o
O
That the state equation (5-76) is controllable follows from Corollary 5-21.
(5-75), or equivaleritly, Equation (5-77) is a time-varying dynamical equation; however, since its A
mly if blll and bl12 are matrix is in the Jordan form and since the components of B are different fram
,=.c:::..:c..:==-C=",.c.'="'='_'="='--=----==='-''''=--=''------ ,_, _--'
zero for all l, one might be tempted to conclude that (5-77) is controllable. where A, R, and e are
Let us check this by using Theorem 5-4. For any fixed to, we have response matrix of the
- ('o -
et>(to-t)B(t)= [ e O
t) O J[:-2'
e- 2('0-'l_
t ]
= :-21
[ - 'o ]
0
The transfer-function n
It is clear that the rows of et>(to - t)B(l) are linearly dependent in t. Hence the
state equation (5-77) is not controllable at any to. I
From this example we see that, in applying a theorem, all the conditions It is clear that G(s) is a
should be carefully checked; otherwise, we might obtain an erroneous con
clusion. Corollary 5-22
Theorem 5-22
A system with a continuous G(t, r) is output controllable at t o if and only if there
exists a finite ti > t o such that all rows of G(t l , r) are linearly independent in r
on [to, ti] over the field of complex number. I
ID. ID.
+ IF
The proof of this theorem is exactly the same as the one of Theorem -5-4
and is therefore omitted.
u
-
- Ir
1\
We study in the following the class of systems that also have linear time ID. ID.
invariant dynamical-equation descriptions. Consider the system that is
describable by
FE: x=Ax +Bu Figure 5-14 A network ,
y=Cx observable.
ICAL EQUATIONS OUTPUT CONTROLLABILITY AND OUTPUT FUNCTION CONTROLLABILITY 215
(5-77) is controHable. where A, B, and C are n x n, n x p, q x n real constant matrices. The impulse
t o, we have response matrix of the system is
-fo ]
= [ :- 2 10
The transfer-function matrix of the system is
mdent in t. Hence the
I (5-78 )
rem, aH the conditions It is clear that (;(s) is a strictly proper rational function matrix.
ain an erroneous con
Corollary 5-22
1
nearly independent in r
I In
In
-
t also have linear time
er the system that is
In In
J
Figure 5-14 A network which is output controllable but neither (state) controllable nor
observable.
- - -_ ..•_.. __ ._--_. __._._._._. __ ._--_._._-~--_._._- .... " ..- ...._.. _._- .._-"._--
1-----------------,
range of G(s), for wh
I I
I I YI
I I
u I I
I I
If the input is rest
: l-----+:--.-Y2
I I then the given outpu
L ~
[,;1]
q x p proper rational
exists a p x q rational
s +1
the rows of which are linearly dependent. Hence the system is not output A system is said to ha"
controllable. The dynamical equation of the system is A necessary and suffi
pG(s) = q in IR (s). Tl
x= -x +u controllability; Man~
unique? Is it a prop
which is controllable and observable. I stable? What are its
problems will not be ;
If a system is output controllable, its output can be transferred to any References SI72, S185
desired value at certain instant of time. A related problem is whether it is
possible to steer the output following a preassigned curve over any interval of
time. A system whose output can be steered over any interval of time is said
to be output function control/able or functional reproducible. "5-8 Computatic
In this section, we dis
Theorem 5-23 chapter. As discussed
A system with a q x p proper rational-function matrix (;(s) is output function conditioned; a comput
controllable if and only if p(;(s) = q in lR(s), the field of rational functions with unstable. If we use a
real coefficients. problem, the result wil
if we use a numericall)
Proof
result will be correcto
If the system is initially relaxed, then we have problem, well or ill cor
a problem, if we must
y(s) 7' (;(s)o(s) (5-80)
stable method, the un
If p(;(s) = q-that is, all the rows of (;(s) are linearly indepe~dent over the possible, in the compu
field of rational functions-then theq x q matrix (;(s)(;*(s) is nonsingular As discussed in TI
(Theorem 2-8). Consequently, for any y(s), if we choose trollability of astate el
more suitable for comI
o(s) = (;*(s)(;(s)G*(S))-ly(S) . (~-81)
computational probler
then Equation (5-80) is satisfied. Consequently, if pG(s) = q, then the system is The computation (
output function controllable. If pG(s)< q, we can always find a y(s), not in the is sttaightforward. L(
CAL EQUATIONS COMPUTATIONAL PROBLEMS 217
range of G(s), for which there exists no solution u(s) in (5-80) (Theorem 2-4).
Q.E.D.
system is not output A system is said to have an inverse if it has both a right inverse and a left inverse.
A necessary and sufficient condition for (;(s) to have a right inverse is that
p(;(s) = q in rR(s). This condition is identical to that of the output function
controllability; Many questions may be raised regarding a right inverse. Is it
unique? Is it a proper rational matrix? What is its minimal degree? Is it
I stable? What are its equivalent conditions in dynamical equations? These
problems will not be stu<;lied in this text. The interested reader is referred to
be transferred to any References S172, S185, S218, and S239.
oblem is whether it is
ve over any intervalof
interval of time is said
ble. *5-8 Computational Problems
In this section, we discuss sorne computational problems encountered in this
chapter. As discussed in Section 2-9, a problem may be well conditioned or ill
(;(s) is output function conditioned; a computational method may be numerical1y stable or numerical1y
rational functions with unstable. Ir we use a numerically stable method to solve a weíl-conditioned
problem, the result wil\ generally be good. Ir a problem is i1l conditioned, even
if we use a numerical1y stable method to solve it, there is no guarantee that the
result will be correct. Ir we use a numerical1y unstable method to solve a
problem, well or in conditioned, the result must be carefully scrutinized. Given
a problem, if we must use an unstable method because of nonexistence of any
(5-80) stable method, the unstable method should be applied at a stage, as late as
independent· over the possible, in the computation.
5)(;*(S) is nonsingular As discussed in Theorem 5-7, there are several ways of checking the con
trollability of a state equation. Among them, statements 3 and 4 appear to be
moresuitable for computer computation. However, they may encounter sorne
(5-81 ) computationalproblems, as will be discussed in the fol1owing.
= q, then the system is The computation of the control1ability matrix U = [B AB ... A"- 1 B]
IS find a y(s), not in the is straightforward. Let Ko~ B. We compute K¡ = AK¡_l; i = 1, 2, ... , n-1.
- ... '
---~~.
._... _-~
At the end, we have U = [K o K 1 K"-l]. The rank of U can then be For example, consider
computed by using the singular value decomposition (Appendix E) which is a
numerically stable method. If the dimension n of the equation is large, this
process requires the computation of AkB for large k and may transform the
problem into a less well conditioned problem. For convenience in discussion,
we assume that all eigenvalues Ai of A are distinct and B is an n x 1 vector.
We also arrange Ai so that IA 11 ¿IA21 ¿ ... ¿IA"I. Clearly, we can write B as
where v¡ is an eigenvector associated with eigenvalue Ai; that is, Av¡ = AiVi. It is (see Reference S212). "
straightforward to verify on the superdiagonal a
except the 1:: at the (20, 1
computed as
for k large
lf e =0, the eigenvalue:
In other words, AkB tends to approach the same vector, vl' as k increases. the migration of the eil
Hence, it will be difficult to check the rank of U if n is large. 5-16, the root locus of i
The same conclusion can also be reached by using a different argument. The indicate the migration (
condition number of a matrix A may be defined as cond A~ IIAI1211A -1112 = a¡JaS' eigenvalues except Al a
where al and as are the largest and smallest singular values of A. It can be shown
that asslA,,1 < IA 11 sal. Hence, if IA 11 ~ IA"I, Cond A is a very large number.
In computer computation, the multiplication of a matrix with a large condition 11 This ¡lnalysis was suggest(
number will introduce large computational error and should be avoided.
Hence, the use of [B AB ... A"-lBJ to check the controllability of {A, B}
is not necessarily a good method.
As an example consider the 10-dimensional state equation (see Reference
S169)
; that is, Av¡ = A¡V¡. lt is (see Reference S212). The diagonal elements range from 1 to 20; the elements
on the superdiagonal are aH equal to 20. The rest of the matrix are aH zeros
except the 6 at the (20, l)th position. The characteristic polynomial of A can be
computed as
:have Ll(s) = n
20
i= 1
(s -i) -(20)19 6
El
lpute its singular values.
)-7, 0.364 x 10- 9 , and 10 11
lative precision of 10-10
lte value is smaller than
re will conclude that the
1 is clearly controllable
. all eigenvalues of A, to
lrst the eigenvalues of A.
puting the eigenvalues of
m be very ill conditioned. Figure 5-16 The root locus of ~(s).
220 CONTROLLABILITY ANO OBSERVABILITY Of LINEAR OYNAMICAL EQUATIONS
if e is larger than 7.8 x 10- 14 ). We list in the following Alo and Al! for sorne e:
el = 7.8 x 10- 14 10.5 ±jO
10
82=10- 10.5±j2.73
83 = 10- s 10.5 ±j8.05 where AW is a nz x n
e4 = 1 10.5 ±j16.26
They are computed directly from A by usingthe QR method. [Thecomputation
of the eigenvalues from fl(s) is not advisable because the characteristic poly
nomial may be more sensitive to 8 than the eigenvalues and the roots of a poly (
and
nomial are very sensitive to the variations of the coefficient.J We see that the p
eigenvalues are very sensitive to e. Thus the eigenvalues of A are very ill
conditioned. For this reason, the use of the criterion rank [sI - A BJ = n where P2AWP~ has bl
to check the controllability of {A, B} may yield an erroneous result. if A~2d =0, the control
It turns out that the use of Theorem 5-16 is the best way of checking the
controllability of a state equation. Furthermore, if the equation is not control
lable, the computation also yields a reduced controllable equation. The proof
of Theorem 5-16 provided a procedure of computing the required equivalence
transformation. The procedure, however, uses the controllability matrix and is has rank nI + n 2 (whel
not satisfactory from the computational point of view. In the following, we lable and can be red!
shall introduce an efficient and numerically stable method to transform a state is transformed, by a se
equation into the form in (5-54). For the single-input case, the method is
essentially the procedure of transforming the matrix A into the Hessenberg A 11
form (see Appendix A). We discuss in the following the general case. A 21
Let PIbe an orthogonal matrix,12 that is, P l 1 = P'I' such that Á~PAP'= O
x x x x
o x x o
P 1 B= O O x : =[B~I] (5-82) C=CP'=[C 1
where pB\I) = nI' pA 2 ]
O O O O smallest integer such
A k ,k-l has dimension ¡
where B\I) is an nI x P upper triangular matrix and nI = rank B. This step can
be accomplished by a sequence of Householder transformations,u We then
compute UnI +n 2 + ... +nk=
A(I) : A(l)] By this process, the co:
- J :_'- - : ~
PI AP'I = A(l) (5-83) The Á in (5-89) is in
[ 2I!I A(l) 22 formation of {A, B} ir
where AW is an nI x nI matrix, AW is an (n -nI) x nI matrix, and so forth. follows (see Reference
If AW = O, the controllability matrix
Step 1. P =Im Ao =A,
U=[B AB:'''; An-IBJ=Pl[~Y}--~W~~~_!~!D~~~~_~~~J
O O O ... (5-84 )
Step 2. Find an ortho~
has rank nI < n and {A, B} is not controllable and can be reduced to controllable
where rank Zj
12If {A, S} has complex elements. we use a unitary matrix, that is. P,' = P¡'
13The P [ can also be obtained by using gaussian elimination with partial pivoting. In this case.
P [ is not orthogonal. However. the method is stiB numericaBy stable. 14 The form of Á is identical t
COMPUTATlONAL PROBLEMS 221
ICAL EQUATIONS
(5-86)
~d. [The computation
he characteristic poly
md the roots of a poly
(5-87)
ient.] We see that the
Llues of A are very ill
rank [sI -A B] =n Now
leous result.
,t way of checking the
:quation is not control
e equation. The proof
.e required equivalence
:J PI
[
B(l) A(1)B(l)
- _-_-
O O
x x
O O ...
"']
~_- ~}j~y~- _- _~_- _- ~_-_-~-~-~- (5-88)
'ollability matrix and is has rank nI + n2 (where x denotes nonzero matrices), and {A, B} is not control
In the following, we lable and can be reduced. Ir AW f.O, we continue the process until {A, B}
od to transform a state is transformed, by a sequence of orthogonal transformations, into 14
lt case, the method is
A 11 A 12 A 13 A 1,k-l AIk B\l)
\. into the Hessenberg
: general case. A 2l An A23 A 2,k-l A2k O
o such that Á~ PAP' = O A 32 A 33 A 3,k-l A 3k , B~ PB = O (5-89)
O O O Ak,k-l O
'artial pivating. In this case.· 14 The farin of Á is identical to the one in (A-17).
lble.
222 CONTROLLABILITY AND OBSERVABILITY OF LINEAR DYNAMICAL EQUATIONS
))- ) [ _!
B A
J..
j j dependent; hence}J.2 =
where X j is a nj x nj matrix. not appear in the subst
Step 4. Update
p=p
[1- O]
n
01' Á, both columns an
Á; hence A 2b¡ and .~21
O Pj
Á, the first column is 1
Step5. ñ=ñ+nj. Ifñ=n,gotostep8.
Á 3 b¡ is linearly depenc
Step 6. j = j + 1 and go to step 2.
column is linearly indo
Step 7. {A, B} is not controllable and can be reduced to an ñ-dimensional
linearly independent o
controllable equation. these indices are equal
Step 8. {A, B} is controllableo triangles as shown in (:
The {A, B} in (5-90)
This is a numerically stable method 01' checking the controllability 01' a state
tion into the following
equation. This algorithm was first proposed by Rosenbrock [S185] and then
discussed in References S6 and S155. These papers used the gaussian elimina 1 x
tion with partial pivoting in finding Pj' Van Dooren [S203] and Patel [S170] O
suggested to use orthogonal transformations to improve the numerical stability
O O
01' the algorithm. The singular value decomposition was also suggested in
Reference S170 to determine the rank 01' B i . O O
The algorithm actually reveals more than the controllability 01' {A, B}; it O O
also reveals its controllability indices. A comparison 01' (5-84) and (5-88) with O O
(5-28) yields immediately O O
i =0,1,2, ... O O
where r; is the number 01' linearly dependent columns in AiB and ni+ ¡ is the The matrix Á is said to
number 01' linearly independent columns in AiB. Since the set of controllability formo This is achieved
indices is uniquely determinable from {r;, i=O, 1,2, ... } [see Figure 5-7 or
Equation (5-33)], it is also uniquely determinable from {ni, i = 1, 2, ... }.
Actually we can saya little more about the controllability indices. In order 1
not to be overwhelmed by notation, we assume that {A, B}, where A and B
are respectively 8 x 8 and 8 x 3 matrices, has been transformed into
, ,
~:(~ x
í 1
v x A X ,
, X X
.,, X X
,X
,
, ,
O : 1: x x x x ,x x ,x x ,x
A , ,
O O /!\
------- :x.... x x'x
0J x'x x'x L 1 _
O O O
A= :..t: x x: x x: x x: x
(5-90)
O O O O O ,ti\.: x x: x x: x Then we have
-------
0- -ó- -Ü Xü-~-: ~- -;;-:-;
- - I I 1
O O O
O O O
O O O: Oo '1' :b1 e1: d 1 I L -)"t I
-------
O O O 0--0- -Ü-:-6-0-:0-:'E:f¡
~ '--v--' "-v--'
1st 2nd 3rd
block co\umn
Let B = fb¡ b2 b31; that is, b; is the ith calumn ofB. Thcn from the structure
01' Á and B in (5-90), we can readily veril'y that 111=3,jJ.2 = 1, and}J.3 =4. These
CAL EQUATlONS COMPUTATlONAL PROBLEMS 223
controllability indices can actually be read out directly from {A, B} without any
computation. The second column of the first block column of A is linearly
dependent on its left-hand-side columns. This implies that Ab 2 is linearly
dependent; hence /12 = 1. Once a column becomes linearly dependent, it will
not appear in the subsequent block columns of A. In the second block column
of A, both columns are linearly independent of their left-hand-side columns of
A; hence A2 b1 and A2 b3 are linearly independent. In the third block column of
A, the first column is linearly dependent on its left-hand-side columns; hence
A3"b 1 is linearly dependent, and /11 = 3. The second column of the third block
column is linearly independent, and there is no other column of A which is
to an ñ-dimensional linearly independent of its left-hand-side columns; hence /13 = 4. We see that
these indices are equal to the numbers of 1 enclosed by circles, squares, and
triangles as shown in (5-90).
The {A, B} in (5-90) can be further transformed by equivalence transforma
ontrollability of a state tion into the following form:
brock [S185J and then , ,
i the gaussian elimina 1 x x x x x. x x,, x x,x , ·
;203J and Patel [S170J
the numerical stability
O 1 x
O O 1
x x x: x x', x x:x
x x x. x x. x x' x
_______
,
·
~ _____ L
. ____ ~ __
------
was also suggested in O O O 1 O O ,, O O; O 0:0
B= O O O Á= O O 1 :0 0:0 O: O
(5-91 )
ollability of {A, B}; it ------
_______ ~ _____ L ____ ~ __
~
~
,
O O O :0 o' O 1 , O .
n AiB and l1i+ 1 is the The matrix Á is said to be in the block companionform or the block Frobenius
le set of controllability formo This is achieved by choosing
. } [see Figure 5-7 or
/'ji, i = 1,2, ... }.
LJ
Jility indices. In order 1
A, B}, where A and B 1
ormed into 1 1
1
l ~.j
( x', x
x: x: x 1
l( x', x
"------
l( x: x
1
O -\.J
~ - - ~ _:->:
(5-90) Then we have
x x x:x x:x
,
x.x
.
l( x:x x x x:x x.x . x:x
\ _':.1_: d 1_ x
- .. -
x
- .. -
X'x
- ..' - - -
X'X
- - ~ - - -
x:x
- _1 .. _ _
) /! :.1'1'.1
1 x x:x x:x x:x
--v--' (5-92)
3rd O O 1:, x x:, 'x x:, x
0- -Ó- -6 : -( - :-x- : :-x- - ~ -:-;
~hcn rrom the structure O O O: O 1 : b2 C2 :d 2
= 1, and f13 = 4. These 0- -Ó- -O : -0- - b- :- 0- - i -:"0
224 CONTROLLABILITY AND OBSERVABILITY OF LINEAR DYNAMICAL EQU.\ 11\ ' ' ' '
Proceeding upward, {A, B} in (5-90) can be transformed into the form in (5-91). are left coprime. Sirr
This process can be easily programmed on a digital computer (see Reference SG).
The process of transforming {A, B} into the block Hessenberg form in (5-89) -A B: O
is numerically stable. The process of transforming the block Hessenberg form 1 O:-A
into the block companion form in (5-91), however, is not numerically stable. T)l= O O: I
The matrix P¡I in (5-92) carries out gaussian elimination on columns of A
• • I •
[ : :
,
I :
without any pivoting. lf pivoting is used, the form of A will be altered, and we O O : O
can never obtain the form in (5-91). Hence, P¡I must be chosen without any '-v
The {A, B} in (5-90) or, equivalently, the {A, B} in (5-91) has the con_troJ
lability indices /ll = 3, /lz = 1, and /l3 = 4. Based on these indices, the {A, B} There are /l + 1 block
can be transformed into A and I and p columns
, B columns. Now we:
1 x x x X x,x'x
, , x x X left to right. Because
O O O 1 O O ,' O ,' O O O O independent of their 1<
O O O O 1 O : O L: __________
O O O O dependent B columns i
_ _ _ _ _ _ _ ..I _ _
------- of the structure of T )l' .
O 1 x
B=PB= - - - - - - -
X x x:x:x x x x
A=PAP- I = -------'--r--------- (5-93)
O O 1 x x x,x'x
, , x x x
O O O O O 0:0: 1 O O O
O O O O O 0:0:0 1 O O Let /l be the least inte~
, ,
O O O O O 0:0:0 O 1 O
and
by the equivalence transformation 15
It can be shown directl
{):~ O O O O O O O in (5-94) are the same 2
O O O (O O O O O
O O O O O (Ü O O rank U = rank [H A
= total numb<
P= O ~! ~ O
O O O O O
O O ,¡. O OL _ , O O O Consequently, we coo(
A
O O O O L '1'
_ , O O O number of linearly inde
O O O O O O ,¡.l. _ .1
O independent columns e
O O O O O O O -1- J
L _ >
position of [he rows oi
apply Householder tr
The positions of 1 in this matrix are determined from the positions of 1 in pivoting on the rows o
(5-90) with the same encirclements. The transformation of {A, B} into (A, B} Appendix A). Once in
can also be easily programmed without first determining the controllability readily determined.
indices (see Reference SG). The form in (5-93) is said to be in the controllable There are two disa
fotm and is very useful in the design of state feedback as will be seen in Chapter 7. {A, B}. First, the size (
We mentian one more method of checking controllability to conclude this larger than the size of
section. Frómstatement 4 ofTheorem 5-7 and Theorem 0-8', we may conclude controlla1:?le eguation c
that {A, B} iscontrollable if and only ir the polynomial matrices si - A and B of {A, B} into a block H
the controllability of {A
The discussion· of t
15 This performs only permutations of columns and rows. part and will not be rep
:reAL EQU.\ 11' '''' COMPUTATIONAL PROBLEMS 225
into theform in (5-91). are left coprime. Similar to Theorem G-14, we form the matrix
lUter (see Reference S6). -A B: O O:
Ltion on columns of A [ :
,
: I :
, :'
It can be shown directly (Problem 5-35) or deduced from Chapter 6 that the ri
O in (5-94) are the same as the ri in (5-28). Hence we have
O
rankU=rank[B AB ... AI'-IB]=(p-ro)+(p-r¡)+'" +(p-rl')
O =total number oflinearly independent B columns in (5-94)
O
O Consequently, we conclude that {A, B} is controllable if and only if the total
O number of linearly independent B columns in (5-94) is n. Note that the linearly
O independent columns of TI' are to be searched in order from left to right. The
position of the rows of TI" however, can be arbitrary altered. Bence, we may
'1-', apply Householder transformations or gaussian eliminations with partial
TI the positions of 1 in pivoting on the rows of TI' to transform TI' into an upper triangular form (see
on of {A, B} into (A, B} Appendix A). Once in this form, the linearly independent B columns can be
ning the controllability readily determined.
o be in the controllable There are two disadvantages in using TI' to check the controllability of
will be seen in Chapter 7. {A, B}. First, the size ofTI' is (2 +/l)n x (n +p)(¡.¡. +1), which is generally much
lability to conclude this larger than the size of A. Second, ir{ A, B} is not controllable, its reduced
n G-8', we may conclude controllable equation cannotbe readily obtained. Bence, the transformation
tI matrices si - A and B of {A, B} into a block Hessenberg formseems to be a better method of checking
the controllability of {A, B}.
The discussion of the observability part is similar to the controllability
part and will not be repeated.
226 CONTROLLABILlTY AND OBSERVABILlTY Of LINEAR DYNAMICAL EQUATlONS
TYT T sr r I~::,;em
Controllability: Theorems the time-varying case,
discussed.
(Chapter 6) and the stability study oflinear systems (Chapter 8). Sorne pr:,tctic:;¡l
implications of these concepts will be given in Chapter 7. ) = L -- J. 3 üJ Ji.
The computational problems of the various controllability and observability
5-3 Show that a linear dyr
conditions are also discussed. A1though the conditions can be stated nicely in finite ti> to such that for an
terms of the ranks of [B AB '" An- 1 B] and [51 -A BJ in the con Hint: Use the nonsingularil
trollability case, they are not suitable for computer computations. An efficient
and numerically stablemethod is introduced to transform a dynamical equation 5-4 Show that ir a linear,
into the form in (5-54) or (5-60), and its controllability or observability can then at any t < too ls it true tha
be determined. The algorithm can also be used to reduce a reducible dyilamical controllable at any t> t o?
equation to an irreducible one.
Before conc1uding this chapter, we remark on the controllability of the "6(1) the Ií!erature, ifx o =0, it
n-dimensional linear time-invariant descrete-time equation encompasses both and doe:
abilit)i and die condition o
x(k + 1) = Ax(k) + Bu(k) identical.
CAL EQUATlONS
PROBLEMS 227
Similar to Definition 5-1, we may define {A, B} to be controllable ifand only if,
given any X o and any Xl' there exists an input sequence {u(k)} of finite length to
ollability and observa transfer Xo to x1.16 The condition of controllability is that rank [B AB
to be controllable and A"-1B] =n (see Problem 2-20). This condition is identical to statement 3 of
lations among sorne of Theorem 5-7. Hence, most of the results in the time-invariant continuous-time
ual to each other: case are applicable to the discrete-time case without any modification. For
the time-varying case, the situation is different but is simpler. This will not be
-8 51-16 lrheorem discussed.
1
,-14 5-17
5-10
: controllability of the 161n the líterature, ifxa =0, it is caBed reachable:irx,=O, il is called contro[[able. Ollr definilion
encompasses both and does not make this distinclion. Ir A is singular, the condítion of reach
tion ability and the condition of controllability are slightlY dirferenl. If A is nonsinglllar, they are
idenlical.
L
228 CONTROLLABILITY AND OBSERVABILITY OF LINEAR DYNAMICAL EQUATIONS
5-5 Is it true that p[B: AB:···: An- 1BJ =p[AB: A2B:··· : AnBJ? If not, under
Is it possible to choose an
what condition wi1l it be true? is of the form y(t) = te-' f
5-6 Show that if a linear time-invariant dynamical equation is control1able, then it is
5-16 Consider the dyna
uniformly control1able. state ofthe equation is no
yiste-'fort~l?
5-7 Check the observability of the dynamical equations given in Problem 5-2.
5-17 Show that the state
5-8 State (without proof) the necessary and sufficient condition for a linear dynamical
equation can be determÍl
equation E to be differential1y control1able and differentia11y observable at too
11 -1 order. [Hin!: Com
5-9 Check the contro11ability of the fo11owing state equations: 5-18 Show that controll;
c. x =[~ ~J x +L~2,] u
to a controllable one.
5-10 Check the control1ability and observability of
5-20 Reduce the equatior
~ ~J +[~ ~J
2 -1
x
O O
u
5-21 Reduce the fol1owin.
y = [1 O 1J x
by using statement 3 ofTheorems 5-7 and 5-13.
5-11 What are the contro11ability indices and the observability index of the equation in
Problem 5-1O?
J.
5-12 Compute the contro11ability indices and the contro11ability index of the state equa to a contro11able and obser
tion in (3-48).
5-22 Consider the n-dimer
5-13 Given a COl1trollable linear time-invariant single-input dynamical equation
x=Ax +bu
where A is an 11 x 11 matrix and b is an 11 x 1 column vector. What is the equivalent dy The rank of its controllabili
namical equation if the basis {b, Ab, ... , An-1 b} is chosen for the state space? Or, equi
valently, if i = Px and if p-1 = [b Ab ... A"-l bJ, what is the new state equation?
is assumed to be n1 « n). 1
5-14 Find the dynamical equations for the systems shown in Figures 5-3, 5-4, and 5-1\ independent columns ofU.
and check the contro11ability and observability of these equations. n1 x n1 unit matrix. Show
... : A"B]? lf not, under Is it possibleto choose an initial state at t =0 such that the output ofthe dynamical equation
is ofthe form y(t)=te- r for t>O?
1 is controllable, then it is 5-16 Consider the dynamical equation in Problem 5-15. 1t is assumed that the initial
state ofthe equation is not known. Is it possible to find an input u¡O,,,,) such that the output
y is te -, for t :?1 ?
n in Prob\em 5-2.
5-17 Show that the state of an observable, n-dimensional linear time-invariant dynamical
ion for a linear dynamica\ equation can be determined instantaneous]y from the output and its derivatives up to
.bservable at too n - 1 order. [Hint: Compute y(t), j;(t), . .. , yI"-¡)(t).J
5-18 Show that controllability and observability of linear time-varying dynamical equa
tions are invariant under any equivalence transformation x = P(t)x., where Pis nonsingular
for all t and continuously differentiable in 1.
y=[1 lJx
to a controllable one.
O
},[~}
y=[O O lJ x
5-23 In Problem 5-22, the reduction procedure reduces to solving for p¡ in p¡Q¡ = I,,¡.
Find a method to solve P ¡ in P ¡Q¡ = 1"1
230 CONTROLLABILITY AND OBSERVABILITY OF LINEAR DYNAMICAL EQUATIONS
5-24 Develop a similar statement as Problem 5-22 for an unobservable linear time-
O
2 1
2
O
O
O
O
O
O
O
O
O
O
2 1 1
2 1 1
e, ~r
O O 2 O O O O 1 1 1 are they output controlla
x= O O O 2 O O O x+ 3 2 1 u
O O O O 1 1 O -1 O O 5-33 AA q X P ra tional rr
m!-trix G l f(S) such that 1
O O O O O 1 O 1 O 1 pG(s) = p in lR(s).
O O O O O O 1 1 O O
5-34 Let P be a nonsing
y~[:
2 1 3 -1 1
1
1
1
1
2
1
O
O
O
1 i}
where B I is an nI x p matl
PE
5-26 Is it possible to find a set of bij and a set of cij such that the following Jordan-form sions (n - n¡) x nI and (n
equation ifand only if {A 22 , A 21 } is
.{j
1 O O
5-35 Show that
1 O O O b 21 b
O
O
1
O
1
1
0] ["
O x + b 31
O b41
b'j
b
22
32
b42
u pU =p[B AB '" A"
y=
['" C2l
C31
CI2
C22
C32
Cl3
C23
C33
CI4
C24
C34
";]
C25
C35
X
is controllabIe? Observable?
5-27 Show that A is cyclic (see Problem 2-45) if and only ifthere exists a vector b such that
5-36 Show that {A, C} is o
{A, b} is controllable. (Hint: Use Corollary 5-21.)
are, respectively, n x n and q
pose ofe.
5-28 Show that if {A, !3} is controllable and A is cyclic. then there exists 8 fJ x 1 solumn
vector r such that {A, Br} is controllable. (H int: Use Theorem 5-21.)
5-37 Showthat{A, B}isc
that is orthogonal to aH coll
5-29 Show that a necessary condition for {A, B} where Bis an n x p matrix, to be control
lable is p ?m, where m is the largest number of Jordan blocks associated with the same left eigenvector O( of A such ti
eigenvalue of A. (See Problem 5-26 and use Theorem 5-21.)
This is called the Popov-Befe
5-30 In Corollary 5-22 we have that a dynamical equation with E = O is output of Theorem 5-7.)
controllable if and only if p[CB: CAB: . --: CA" - lB] = q. Show that if E F O, then the
dynamical equation is output controllable ir and only ir
p[eB : CAB: ... : CA" - lB: E] = q
5-3i Consider a linear time-inváriant dynamical equation with E =0. Under ·what
condition on e will the (state) controllability imply output cóntrollability?
.... _..--:..:=.
PROBLEMS 231
CAL EQUATIONS
Inobservable linear time- 5-32 Consider two systems with the transfer-function matrices
PB=[~IJ I
PAP- =[ ~~:
where B I is an ni x P matrix and pB = pB I = ni' The matrices A21 and A 22 have dimen
sions (n - n ¡) x ni and (n - n¡) x (n - n d, respectively. Show that {A, B} is controHable
the foHowing Jardan-form
if and only if {A n , A2 ¡} is controllable.
re exists a vector b such that 5-36 Show that {A, c} is observable if and only if {A, C*C} is observable, where A and C
are, respectively, 11 x n and q x /l constant matrices and C* is the complex conjugate trans
pose of C.
1 there exists a p x 1 col uml1
m 5-21.) 5-37 Show that {A, B} is controllable if and only if there exists no left eigenvector of A
that is orthogonal to aH columns of B, that is, there exist no eigenvalue A. and nonzero
n n x p matrix, to be control left eigenvector a of A such that
ks associated with the same ).a=aA and IXB=O
This is called the Popov-Belevitch-Hautus test in Reference S125. (Hint: See statement 4
of Theorem 5-7.)
ation with E =0 is output
Show that if E ~O, then the
_
.
~
_
.
_
~
.
~
function is the only a\
tion is obtained, by a~
system can be readily
operational amplifier (
link between the state
For every realizat
!
!
INTRODUCTlON 233
i
function is the only available description. After a dynamical-equation realiza
tion is obtained, by assigning the outputs of integrators as state variables, the
system can be readily simulated. The realization can also be built by using
operational amplifier circuits. Final1y, the results can be used to establish the
link between the state-variable approach and the transfer-function approach.
For every realizable transfer-function matrix (;(s), there is an unlimited
number of linear time-invariant dynamical-equation realizations. Therefore
a major problem in the realization is to find a "good" realization. lt is c1ear
that a dynamical-equation realization with the least possible dimension is a
good realization. We c1aim that a realization of (;(s) with the least possible
dimension must be a control1able and observable dynamical equation. lndeed,
if a linear time-invariant dynamical-equation realization of (;(s) is found, and
,I , if the equation is uncontrol1able or unobservable, then fol1owing from Theorem
5-19 it is \?ossible to reduce the realization to a lesser-dimensional equation that
still has G(s) as its transfer-function matrix. This reduction is impossible only
if the equation is control1able and observable. Therefore, we conclude that a
realization of (;(s) with the least possible dimensioñ is a control1able and obser
vable dynamical equation, or equivalently, an irreducible dynamical equation.
Such a realization is cal1ed a minimal-dimensional or irreducible realization.
In this chapter we study mainly irreducible realizations for the following reasons:
(1) A rational transfer-function matrix describes only the control1able and
observable part of a dynamical equation; hence a faithful realization should be
an irreducible one. (2) When an irreducible realization is used to synthesize a
network, the number of integrators needed will be minima\. This is desirable for
reasons of economy and sensitivity. Note that if an irreducible realization is
; in electrical engineering.
found, any other irreducible realization can be obtained by applying an equiva
,r an active network that
lence transformation (Theorem 5-20).
le subject matter we shall
This chapter is organized as fol1ows. In Section 6-2, we introduce the con
t is, to determine a linear
cept of the degree for proper rational matrices. 1ts significance is demonstrated
scribed rational transfer
in Theorem 6-2. In Section 6-3, various realizations are introduced for scalar
Ddern version of network
rational functions. Hankel theorem is also introduced. The realization
methods are then extended to vector rational functions in Section 6-4. Two
has a prescribed transfer
different irreducible realization methods, one based on Hankel matrices and
1 "realization" is justified
the other on coprime fractions, for proper rational matrices are discussed in
: system with the transfer-
Sections 6-5 and 6-6. In Section 6-7 we introduce a new mathematical descrip
by using an operational
tion, cal1ed polynomial matrix description, for linear time-invariant systems. lts
-em 4-10 that every proper
relationships with transfer functions and dynamical equations are also estab
ime-invariant dynamical
lished. In Section 6-8 the concept ofequivalent dynamical equation is exteDded
questions. In this chapter
to strict system equivalence for polynomial matrix description. lt is shown that,
under the coprimeness assumption, al1 polynomial matrix descriptions which
_g reasons: First, there are
have the same transfer matrix are strictly system equivalent. In the last section,
algorithms developed ex
we study the identification of discrete-time systems from arbitrary input-output
'ply these techniques and
pairs. The:concept of persistent exciting is introduced.
zed into dynamical equa
This chapter is based mainly o.n references 15,42,47,60,62,67,68,83,89,
ways desirable to simulate
98, 115, S27: S48, S52, S126, S158, SI61,SI85, S187, S209; and S218. For the
~ck its performance before
realization ofimpulse-response matrices G(t, T), the interested reader is referred
i efficiently if the transfer
to References 32, 100, 101, 114, and S128.
234 IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICAnON THE CHARACTERlSTIC F
be a realization of the proper rational function g(s). Then FE 1 is irreducible Definition 6-1
(controllable and observable) if and only if
The characteristic poZ
det(sI - A) = k[ denominator of g(s)] (6-1 ) the least common de
denoted by D(;(S), is (
or dim A = deg g(s) of (;(S).1
where k is a nonzero constant.
Example 1
Consider the rational-
Proof
Let Ll(s)~ det (sI - A) and let
First we show that {A, b, c} is irreducible if and only if Ll(s) and N I(S) are co The minors of order
prime. Indeed, if {A, b, c} is not irreducible, then Theorems 5-16 and 5-17 The minor of order 2
imply the existence of a {A, b, e} such that dim A < dim A and (;1(S) is s +1 and b(;
1/(s + 1), 1/(s + 1), and
Hence the characteristi
~1(S) ~c(sI _A)-1 b =c(sI -A)-lb = N I(S)
Ll(s) Ll(s)
From this example
where J,(s) = det(sI - A). Since deg J,(s) = dim Á < dim A = deg Ll(s), we con general different from
elude that Ll(s) and NI (s) have common factors. Reversing the aboye argument,
we can show that if Ll(s) and N I(S) are not coprime, then {A, b, c} is not irredu lItis also caBed lhe McMilL
cible. Hence we have established that {A, b, c} is controllable and observable only lo proper ralional mal
if and only if Ll(s) and NI (s) are coprime. thepoles .al s = oo. See Re
:, AND IDENTIFICAnON
THE CHARACTERISTIC POLNOMIAL AND THE DEGREE Of A PROPER RATlONAL MATRIX 235
Example 1
Consider the rational-function matrices
A
--
5+1
1
--
s+1
1 1 A
~-2-s+1
square matrix] and different from the least common denominator of al1 the We remark that 1
entries of (;(S). Ir (;(S) is scalar (a Ixl matrix), the characteristic polynomial of nonzero constant [se,
(;(s) reduces to the denominator of (;(s). the characteristic po:
6-1' can be establisr
Example 2 reader is referred to 1
Consider the 2 x 3 rational-function matrix
*Theorem 6-2
S
Let the multivariable
A S +l (s +1)(s +2)
G(s) =
[
-1 1
s +1 (s + 1)(s +2)
be a realization of ti
The minors of order 1 are the entries of (;(s). There are three minors of order 2. (controllable and obs
Theyare det (sI
s 1 s+1 1 or dir
(s+lf(s+2)+(s+lf(s+2)=(s+lf(s+2)=(s+l)(s+2) (6-3)
where k is a nonzero ,
sil s +4
--'-+ =----
The irreducibility
s +1 s (s +1)(s +3) (s +1)(s +3)
lenominator of all the We remark that the polynomials det D,.(s) and det D/s) differ at most by a
Lcteristic polynomial of nonzero constant [see Equation (6-189a)]; hence either one can be used to define
the characteristic polynomial of G(s). The equivalence of Definitions 6-1 and
6-1' can be established by using the Smith-McMillan formo The interested
reader is referred to References 15, S125, and S185.
*Theorem 6-2
Let the multivariable linear time-invariant dynamical equation
FE: x=Ax +Bu
y=Cx +Eu
be a realization of the proper rational matrix (;(s). Then FE is irreducible
:hree minors of order 2. (controllable and observable) if and only ir
det (sI - A) = k [characteristic polynomial of G(s)]
1 or dimA=degG(s)
(6-3)
: + 1)(s +2) where k is a nonzero constant. 11
up to the (n - l)th arder qualify as state variables. Define (6-8). To show that (
to Figure 6-1 to ShOVi
r
r
A()D. 2(s) D.
X S=:
XI(S~
x
=:
y(s)
sy(s) = J [1 S
:
] A
y(s) (6-6a)
equal to g(s). A diffe
. . .
Xn(s) Sn-¡y(S) S"-¡ where
or, in the time domain,
X¡(t)~ y(t)
The first n - 1 equations of (6-8a) are obtained directly from (6-6b). Theyare
the consequence of the definition of Xi, i = 1, 2, ... , n, and are independent of
the given transfer function. The transfer function g(s) comes into (6-8a) only
at its last equation through (6-7). We draw in Figure 6-12 a block diagram of
2 Nole Ihal ir f3 is moved lo Ihe OUlpUI, Ihen b and e in (6-8) beco me b' = [O O ... O 1] and
e = [f3O O ... O). and g(s) = c(s]
f3
¿l(s
ine (6-8). To show that (6-8) is a realization of g(s), we may apply Mason's formula
to Figure 6-1 to show, as in Figure 4-6, that the transfer function from u to y is
equal to g(s). A different way is to verify
(s) (6-6a) g(s) =c(sI - A)-l b
where
-1
s
o
O
O - I 0l
O]
··· O
...
~~,
(6-9)
O s
(6-6b)
s O J
It is clear that (sI - A)-l b is equal to the last column of (sI - A)-l 13 or, equi
valently, to the cofactors of the last row of (131 ~(s))(sI - A), where
~(s)~ det (sI - A) = sn +CtlSn-1 + ... +Ct n
Xn + f3u (6-7)
(see Problem 2-26). In view of the form of si - A, the cofactors of the last row
of (f3/~(s))(sI - A) can be easily computed as
O
13 S S2 ...
O
~(s) [1
O
+ u (6-Sa)
Hence,
In
O -1 O
(SI_A)-'b~[l
13 s O
~ ~ 13 s.
(6-Sb)
O:
-1
: = ~(s)
O
[1 : ]
sn-2
rom (6-6b). They are O s
ld are independent of Ct n Ct n_ 1 Ct 2 S +Ct 1 13 Sn- I
:omes into (6-8a) only (6-10)
12 a block diagram of
13 13 (6-11 )
~(s) sn +Ct1s n- 1 + ... +Ct n
This verifies that (6-8) is indeed a realization of g(s) in (6-4). Since deg g(5) =
dim A, (6-8) is an irreducible realization (Theorem 6-1). This can also be verified
by showing that (6-8) is controllable (except for the trivial case 13 = O) and
observable. Note that the realization (6-8) can be obtained directly from the
coefficients of g(s) in (6-4).
J
.. _~_ •• ~ .~~ .. ~ _ _ •• ,~~ • • " , ' _ _ •• ~~_,._,_'~'_'._ _ ••· _ •• 0_ ••' _ . ' ,_~ •• "~
-------~-----~------------
.
i
-1
I
240 IRREDUCIBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE AND lDENTIFICATION 11 !
!
(6-12)
in the right-hand sic
remainder gives the r
where Ci i and Pi,
for i =0, 1, 2,. _., n, are real constants. It is assumed that coefficients associatet
Ci o i= O. By long division, 91 (s) can be written as u a unique y can be d
A f31 Sn - 1 +f3 zs n-Z + ... +f3n 6A
as
gl(S)= sn + a sn-l + . . . + a - s + a +e=g(s) +e (6-13)
1 n 1 n Xn(t)~y(t)
X n -1(t)~y< 1)(t)
where e = 91( 00) = Po/Ci o. Since the constant e gives immediately the direct X n _ Z(t)~ y<Z)(t)
transmission part of a realization, we need to consider in the following only the
strictly proper rational function
X 1(t)~ y<n -1)(
N(s) 13 sn - 1 + 13 sn - Z + ... + 13
A( )
g s-6
_ -
)6 _ 1 Z n
(6-14 ) then x = [Xl X z
D(s - sn +a 1sn 1 + ... +a n in (6-17) yields
Let u and y be the input and output of g(s) in (6-14). Then we have
where D(p) = pn + a1p n-l + .. , -i-a mN(p) = f3 1pn-l + f3 zpn- Z + ... + 13m and pi
stands for di/dt i . Clearly, the transfer function of(6-15) is g(s) in (6-14). In the
following we introduce several different realizations of (6-15). The foregoing equatio
"ND IDENTIFlCATlON
IRREDUCIBLE REALIZA nONs OF PROPER RA T10NAL FUNcnONs 241
(6-14 )
then x = [Xl Xz XII]' qualifies as the state vector. The set of equations
in (6-17) yields
Controllable canonie
We shall now introdu(
form realization, of g(.
a new variable v(t) del
and
Equation (6-21) has t
variables as, similar te
Figure 6-2 Block diagram of the observable canonica\-form dynamica\ equation (6-18).
x(,)J:~i:~~
The observability matrix of (6-18) is lxn(s}J
o
which, together with lX =1=0, implies that the controllability matrix of (6-18) has a
-el"
rank less than n. . Hence if D(s)and N(s) are not coprime, the realization in
(6-18) is not controllable. e = [f3n
AND IDENTIFICATION IRREDUCIBLE REALlZATlONS OF PROPER RATIONAL FUNCTlONS 243
Equation (6-21) has the same form as (6-5a); henee we may define the state
variables as, similar to (6-6a),
O O O .. 1 O
ymatrix of (6-18) has a
-a" -(Xn - I -a.- l -!Xi 1
ime, the realization in
e = [.8n .8n-1 .8n - 2 ... .81]
.J
244 IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTlFICATION IRREO"
This is a realizalion of 9(s) in (6-14). Unlike (6-17), there are no simple relation Realization from the Ha
ships between Xi, u, and y. The dynamical equation in (6-26) is always control Consider the proper rati(
lable no matter whether D(s) and N(s) are coprime or not and is said to be in the
controllab/e canonica/fonn. Ir D(s) and N(s) are coprime, then (6-26) is observ
able as weJl; otherwise, it is not observable. This assertion is dual to that of the
observable canonical form, and its proof is left as an exercise. A block diagram We expand it into an inf
of (6-26) is drawn in Figure 6-3.
9(S
Example 1 The coefficients {hU), i =
Consider the proper irreducible transfer function These parameters can be
3 2 h(O) = {Jo
A(S) = 4s + 25s +45s + 34
g 2s 3 + 12s 2 + 20s + 16 h(l) = -(1.
h(2) = -(1.
By long division, 9(S) can be written as
A) 0.5s 2 +2.5s + 1 h(n) = -(1.
g(s = S3 +6s2 +10s +8 +2
h(n +0 = -(1.
Hence its controllable canonical-form realization is
i= 1, 2
h(2)
H(cx, f3)~ h\3)
1 -6 0.5 [
y=[O O lJ x + 2u h(cx)
lt is caJled a Hanke/ me
{hU), i = 1, 2, 3, ... }.
invo/ved in H«(1., f3).
Theorem 6-3
The proper transfer funl
pH(n, n)=pH
where pdenotes the rar
Proof
Figure6-3 Block diagram of the controllable canonical-form dynamical equation We show first that if del
(6-26). lfdegg(s)=n, then(6-3C
'CE, AND IDENTIFICATION IRREDUCIBLE REALIZATIONS OF PROPER RATIONAL FUNCTIONS 245
The coefficients {hU), i=O, 1,2, ... } will be called the Markov parameters.
These parameters can be obtained recursively from a i and /3i as
h(O) = /30
h(l) = -a¡h(O) +/31
h(2) = -a 1h(1)-a2h(O) +/32 (6-29)
Theorem 6-3
The proper transfer function g(s) in (6-27) has degree n if and only if
pH(n, n)=pH(n +k, n +l)=n for every k,l = 1,2,3, . . . (6-32)
Proof
\-rorm dynamical equation We showfirst that if deg g(s) = n, then pH(n, n) = pH(n + 1, IX)) = pH(oo, (0) = n.
If degg(s) = n, then (6-30) holds, alld n is the smallest integer having this property.
Because of (6-30) the (n+ l)th row of H(n+ 1, (0) can be written as a linear Note that there is on
combination of the n rows of H(n, CX)). Hence we have pH(n, (0) = up to h(2n) are used ir
pH(n+ 1, (0). Furthermore, we have pH(n, oo)=n; otherwise, there would be algorithm 3 discussed i
an integer ñ smal!er than n with the property (6-30). Because oflhe structure of Hin (6-35) in order fn
H. the matrix H(n + 2,(0) without the first row reduces to the matrix H(n + 1, ro) pendent and the (O" + 1
without the first column. Hence the (n +2)th row of H(n +2, (0) is linearly Then Theorem 6-3 iml
dependent on its previous n rows and, consequently, on the first n row of dependent on their pre
H(n + 2, (0). Proceeding in this manner, we can establish pH(n, CX) ) = pH( cx), (0) = Iinearly dependent ro\\
n. Again using (6-30), the (n+1)th column of H(n, (0) is linearly dependent cal! the (O" + l)th row
on the columns of H(n, n). Proceeding similarly, we have pH(n, n) = (a + k)th row, k = 2, 3,
p H(n + k, n + 1) = n for every k, 1= 1, 2 .... D(s) and N(s) are coprir
Now we show that if (6-32) holds, then g(s) =h(O) +h(l)s-I + ... can be ing algorithm wilI also
reduced to a proper rational function of degree n. The condition pH(n, n) =
pH( 00, CX)) = n implies the existence of {CXi' i = 1,2, ... , n} to meet (6-30). Using [al l
(6-29) we can compute {/3¡, i =0,1,2, ... , n}. Hence we have This equation expresse
/3os· +/3I S·-1 + ... +/3. combination of its pre'
g(s)=h(O) +h(l)s-1 +h(2)s-2 + ... dependent row. Note
- s· +cxls·- 1 + ... +cx.
we do not have a¡ = (J.._
Since the n is the smal!est integer having this property, we have deg g(s) = n.
This completes the proof of this theorem. Q.E. D.
with
Consider the dynamical equation
O 1 O
FE: x=Ax +bu O O 1
be less than n. We expand 9(s) as in (6-28) by using the recursive equations in lh(O" + 1)
(6-29) and (6-30). We then form the Hankel m'atrix The effect of the multip
h(l) h(2) h(n) 1, or equivalently, shifts
fl(2) h(3) h(n + 1) c, cAkb just picks up thé
H(n +1, n)= (6-35) cb=,
h(n) h(n + 1) h(2n -1)
f¡(n + 1) 'h(n +2) h(2n) 3 For eomputer eOmpul'l.!ion.
:E, AND IDENTIFICATION IRREDUCIBLE REALIZATION OF PROPER RATIONAL FUNCTIONS 247
L be written as a linear Note that there is one more row than column, and the Markov parameters
we have pH(n, 00) = up to h(2n) are used in forming H(n + 1, n). Now we apply the row-searching
herwise, there would be algorithm 3 discussed in Appendix A to search the linearly independent rows of
:cause ofthe structure of Hin (6-35) in order from top to bottom. Let the first (J rows be linearly inde
o the matrix H(n + 1, oc!) pendent and the ((J + l)th row of H be linearly dependent on its previous rows.
. H(n + 2, oc!) is linearly Then Theorem 6-3 implies that the ((J +k)th rows, k = 1, 2, 3, ... , are alllinearly
, on the first n row of dependent on their previous rows and the rank ofH(n + 1, n) is (J. Hence once a
pH(n, 00) = pH(oo , 00) = linearly dependent row appears in H(n + 1, n), we may stop the search. We shall
)) is linearly dependent call the ((J + l)th row of H(n + 1, n) the primary linearly dependent row; the
, we have pH(n, n) = ((J + k)th row, k = 2, 3, ... , nonprimary linearly dependent rows. Note that if
D(s) and N(s) are coprime, then (J = n; otherwise, we have (J < n. The row scarch
) +h(l)s- 1 + ... can be ing algorithm will also yield {a¡, i = 1, 2, , } such that
le condition pH(n, n) =
[al az .,. a" (Il O O]H(n+l,n)=O (6-36)
!} to meet (6-30). Using
lave This equation expresses the primary linearly dependent row as a unique linear
sn- 1 + +f3n combination of its previous rows. The element ({l corresponds to the primary
,-1 + +lXn dependent row. Note that if (J = n, then a¡ = IX n _;, i = 1, 2, ... ,n. If (J < n, then
we do not have a¡ = IX n -;. We claim that the (J-dimensional dynamical equation
y, we have deg g(s) = n.
Q.E.D. x=Ax +bu y=cx +eu (6- 37)
with
A~[ O
O 1 O O
[h(ll ¡
n
O O 1 O h(2)
b- : (6-38)
O O O - h((J -1)
-al -az -a3 -a,,_ I -a" h((J)
e=[ O O O O] e= h(O)
+ ... (6-33) is a controllable and observable realization of g(s). Because of (6-36) and
Theorem 6-3, we have
s a realization of g(s) in
h((J + i)= -alh((J + i -l)-azh((J+i - 2)- ... - a"h(i) i= 1,2,3, ...
: a different realization.
¡ith deg D(s) = n. Here
;e the degree of g(s) may
Ab =[~i~~ j, =[~;~: l. . , ~[~~~ :~n, . .
AZb A'b (6-39)
The effect of the multiplication of A simply increases the atgumént ¡in h(i) by
h(n) 1, or equivalently, shifts the elements up by one position. Because ofthe form of
e, cAkb just picks up the first element of Akb as
h(n + 1)
(6-35 ) eb =h(I), eAb =h(2), eAzb =h(3), (6-40)
h(2n -1)
h(2n) 3 For computer compullttion, numerically stable methods should be used. See Appendix A.
248 IRREDUCIBLE REALlZATlONS, STRlcr SYSTEM EQUlVALENCE, AND IDENTIFlCATlON IR
This shows that (6-37) is indeed a realization of g(s). The controllability matrix We note that this
of (6-37) is nominator of g(s) are
smaller than the degrel
[b Ab ... A"-lbJ=H(u,u)
Dual to the introd
The Hankel matrix H(u, u) has rank u; hence {A, b} in (6-37) is controHable. pendent columns of H(
The observability matrix of (6-37) is irreducible realization.
e 100 O
•Jordan-canonical-fo
cA O 1 O O
cAz O O 1 O We use an example to i
a lordan-form dynami
general case. Assume
eA"-1 o O O and assume that D(s) c~
Clearly {A, e} is observable. Hence (6-37) is an irreducible realization of g(s), assume that g(s) can be
A( ) = -el
gs
Example 2
(s -;
Consider
The block diagrams of
ficients el!, e lZ , e 13 , ez,
" Z ,
I 4
I
Z
1
H(S, 4) = ,--
3 I
I
1- 1 _1
,__ 1-1 Z 8
1- 1 _1- _2
Z 8 4
1 _1- _2 _J.
8 4 8
X~U5
1
O
-1
y=[ O OJ x+O.Su
The last row ofthe companion-form matrix A consists ofthe first three elements
of k with the signs reversed. The b vector consists of the first three Markov
parameters of g(s) [excluding h(O)]. The form of e is fixed and is independent
of g(s). • Figure 6-4 Two block di;
, AND IDENTIFICATION IRREDUCIBLE REALlZATION OF PROPER RATIONAL FUNCTIONS 249
: controllability matrix We note that this procedure also reveals whether the numerator and de
nominator of 9(S) are coprime. If the rank of the Hankel matrix of 9(S) is
smal1er than the degree of its denominator, then 9(S) is not irreducible.
Dual to the introduced procedure, we may also search the linearly inde
I (6-37) is controHable. pendent columns of H(n, ri + 1) in order from left to right and obtain a different
irreducible realization. The procedure wil1 not be repeated.
•Jordan-canonical-form realization
We use an example to illustrate the procedure to realize a transfer function into
a Jordan-form dynamical equation. The idea can be easily extended to the
general case. Assume that D(s) consists of three distinct roots Al, ..1. 2 , and ..1. 3 ,
and assume that D(s) can be factored as D(s) = (s - ..1. 1)3(S - . 1. 2 )(s - ..1. 3), We also
;ible realization of 9(S). assume that 9(S) can be expanded by partial fraction expansion into
1=0 e13 \ - - - - - - - - - - - - - - - ,
they are associated with the input We note that every block in Figure 6-4 can numbers. In this ca
be viewed as consisting of an integrator, as shown in Figure 6-5. Hence the puter, for complex m
output of each block qualifies as a state variable. By assigning the output of this can be taken car
each block as a state variable, and referring to Figure 6-5, we can readily obtain will be demonstratec
the dynamical equation for each block of Figure 6-4 as shown. By grouping assumed to be reaL if
the equations in Figure 6-4la), we obtain /:, is also a pole of 9(5).
following subequatioi
1.[ O : O O O
O ;'1 1: O O O
O O 1'1: O O x + 1 u (6-42a)
-o'---6---6-:--~;-: O 1
O O O· ""0"" : ;.~ 1
(6-42b) where Al is the Jordar
Equation (6-42) is in the Jordan canonical farm. There is one Jardan block (no transpose) of Al'
associated with each eigenvalue. The eq uation is clearly controllable (Carollary introduce the equivale
(5-21); it is also observable, except for the trivial cases e ll = O, ez = O, or e 3 = O.
Therefore the dynamical equation (6-42) is an irreducible realization of the
g(5) in (6-41).
If the block diagram in Figure 6-4(b) is used and if the state variables are and
chosen as shown, then the dynamical equation is
XII
.1, O O: O 01 ell Then it can be easily
-
~xu
s-a
---+--
11 I
I
I
b
x l\'
1
x=
O
O
- - -
1 +2i :
- -
O ,'
- - - - l.
I
I
O O ,:
I
L ~
O '0 .
1 u
1
1
y=[c¡ cIJ[::J
(6-43b)
(6-42b) where Al is the Jordan block associated with A and Al is the complex conjugate
(no transpose) of A¡. Clearly, Al is the Jordan block associated with 1 Let us
re is one Jordan block introduce the equivalence transformation x = Px, where
controllable (Corol!ary
11 =0, ez =0, or e3 =0.
Then it can be easily verified that the dynamical equation in (6-43) can be
transformed into
l ~ll=[
x2
ReAl
-1m A¡
ImA1J[~¡J+[
Re Al x2
2Reb l
-21m b¡
Ju (6-44a)
of g(s). Note the dif where Re A and 1m A denote the real part and the imaginary part of A, respec
Hes 6-4(a) and6-4(b). tivel)'. Since al! the coefficients in (6-44) are reaL this equation can be used on
on intoaJordancanoni analog computer simulations. Another convenient way to transform a Jordan
on must be factored. 01' form dynamical equation into an equation with real coefficients is to use the
lUSt be computed. This transformation introduced in Problems 6-14 and 6-15.
Second. if the transfer
will consist of complex Example 3
Consider the Jordan-form equation with complex eigenvalues.
" 1 +2i 1 O D O 2-3i
I
Ix
O 1 +2i () O
- - - - - - - - - - - - - - '- - - - - - ... - - - - - - - ... -.
e 1
----¡-----
I
x= O o 1 -2; 1 O x + 2 +3i u (6-45)
l
O O O 1-2;: O 1
I
I
O o '----0-------0---:--2 2
I
.~
y=[ -; i 2J x
252 IRREDU('IBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATION REALIZA
p= O -i O ,, O general formula, we gi
,
O
_____
i O -i ,
__ , _ _ _ _ _ _ _ _ _ 1___
O
O O O O ,, 1 Example 4
y
whose coefficients are all real. I and then verify this by
terms of the coefficien
A remark is in order regarding these realizations of y(s). Clearly, the con we obtain
trollable canonical form and observable canonical form realizations are the
easiest to obtain. These realizations. however. are not necessarily irreducible y=PY=X2+'
.. 2
unless the given 9(s) = N(s)/D(s) is known to be irreducible [N(s) and D(s) are y=py=-a;
coprime]. The realization obtained from the Markov parameters and the Substituting these into
Jordanform realization are always controllable and observable no matter the obtain
given 9(S) is irreducible or not. Because of the requirement of computing the
poles of 9(S). the Jordan-form realization is generally more difficult to compute. e(t) = 130(1)
The Jordan-form realization however is. as discllssed in References 23, 81 b¡(t) = fi¡(t)
and 87, least sensitive to parameter varialions among all realizations. The b 2(t) = fi2(t)
sensilivity is defined as the shifting of the eigenvallles dlle to parameter varia Since the time functions
lions. In practice, in order to reduce sensHivity. a transfer function 9(S) is the differential equatior
ractored as ofthe form in (6-48). \\
9(s) =g¡(S)92(.':)··· the relations between b,
or 9(S)=91(S)+92(S)+'"
where 9¡(s) and y¡(s) are transfer functions of degree I or 2. We then realize
6-4 Realizations,
each Ms) and y¡(s), and then connect them together. Thc first one is called a
Transfer Function:
tandem realization; the second one, a parallel realizatioll. This typc of realiza
tion is often used in the design of digital filters (see Refercncc S.t7).
In this section realizatio
studied. By a vector rati
function matrix. Consi<
·Realization of linear time-varying differehtial equations. Before
conc1uding this section, we shall briefly discuss the se"tup of dynam ical eqllations
by choosing y(t), y(t), ... , yln -¡)(t) as state variables, a dynamical equation of
exactly the same form as (6-8) can be set up. However, if the right-hand side of
(6-46) consists of the derivatives of u, although it can still be realized into a
dynamical equation, the situation becomes very involved. lnstead of giving a
general formula, we give an example to iI1ustrate the procedure.
Example 4
Consider the following second-order time-varying differential equation
[p2 + IX¡ (t)p + IX2(t)]y(t) = [,80(t)p2 + ,8¡ (t)p + ,8 2(t)]U(t) (6-47)
1
-IX¡(t)
J[Xl] +
X2
[b¡(t)]
b 2(t) u
(6-48a)
I and then verify this by computing the unknown time functions b¡, b 2 , and e in
terms of the coefficients of (6-47). Differentiating (6-48b) and using (6-48a),
. gIs). Clearly, the con we obtain
rm realizations are the y= py = X2 +b¡(t)u(t) + é(t)u(t) + e(t)u(t)
t necessarily irreducible
ji = p2 y = -1Y.2X¡ -1Y.¡X2 +b 2u +b¡u +b¡u + eu + 2éú + eii
cible [N(s) and D(s) are
ov parameters and the Substituting these into (6-47) and equating the coefficients of u, u, and Ü, we
bservable no malter the obtain
:ment of computing the
e(t) = f3 o(t)
ore difficult to compute.
b ¡ (t) = ,8 ¡ (t) - IY.¡ (t),8 o(t) - 2~ o(t) (6-49)
;ed in References 23, 81
b 2(t) = ,8 it) - b ¡(t) - IX¡ (t )b¡ (t) - IX ¡(t)~o(t) ~ 1Y.2(t ),8o(t) - fio
g all realizations. The
due to pararneter varia Since the time functions b¡, b 2 and d can be solved from (6-49), we conclude that
transfer function g(s) is the differential equation (6-47) can be transformed into a dynamical equation
of the form in (6-48). We see that even for a second-order differential equation,
the relations bet ween !J, e. and lhe 1:J.¡'s and f3¡'s are verv complic'olled. tjj
J)u(t) (6-46)
=========",-=',='-=,."., .
REALIZI
254 IRREDUCIBLE REALIZAT10NS, STRicr SYSTEM EQUIVALENCE, AND IDENTIFICAT10N
Example 1
It is assumed that every g;(S) is irreducible. We first expand Ginto
Consider
l
(;(s) =
e ] rgl(S)]
e
t
eq
+ ~z:(S)
gq(s)
(6-51 )
A
G(s) =
(8 +s1)(s
+3
+2)j =
[ 5 +4
where e¡ =g;(oo), and gi(S)~g;(s) -e¡ is a strictly proper rational function. 5 +3
We compute the least common denominator of g¡, for i = 1,2, ... , q, say sn +
al s" - 1 + ... + a", and then express G(5) as
(6-52 )
Hence a minimal-dime
o 1 O O O
O O 1 O O
x= x+ u (6-53a)
O O O 1 O
-Cf.
n -Cf. n- l -Cf. n- Z -Cf.
l 1 We study now the
Since its development i
the result. Consider th
YI]
~2
[Pln PI(n-l)
Ptn PZ\n-l)
Pll] [el]
P~l x+ e u
G(5) = [g'I(S) : g'z(5) : ..
[Yq
=
tq (6-53b)
=[el : ez:··· : el
Pqn Pq(n-l) Pq" e
= [el: ez : ... : el
is a realization of (6-52). This can be proved by using the controllable-form
realization of g(s) in (6-14). By comparing (6-53) with (6-26), we see that the
transfer function from u to y¡ is equal to
form and the observable-fonn realizations. But for column ration'al functions
it is not possible to have the observable-form realization.
REALlZATIONS Of VECTOR PROPER RATIONAL TRANSfER FUNCTIONS 255
:, AND IDENTIFICATION
Example 1
and G into
Consider
(6-51 )
A
G(s)-
_
¡(S s +3 +2)
+ 1)(s
s +4
j =
[OJ
1
+
[s +3j
(s + 1)(s
1
+2)
-- --
per rational function. s+3 s+3
= 1,2, ... , q, say Sil +
OJ 1 [ (s+3f J
= [ 1 +(s+1)(s +2)(s +3) (s +1)(5 +2)
,,]
e u
the result. Consider the 1 x p proper rational matrix
G(s) = [g'l (s) : g~(s) : ... : g'p(s)]
t
eq
(6-53b)
= [el: e2 : ... : e p] +[gl(S) : g2(S) : ... : gp(s)]
= [ e .. e .....
. 'e] +----.----- 1
1, 2· . p Sn + a 1 Sn 1 + ' .. + a n
the control1able-form
n sn -
X [ 1-'11- 1 n sn - 2 + . .. + 1-'1/1
+ 1-'12 n sn - 1 + 1-'22
n .: 1-'21' n sn - 2 +. . .
(6-26), we see that the
+{J2n:"': {JpIS"-¡ +{Jp2 S 2 + .. - + (Jpn]
/l-
(6-55)
Then the dynamical eguation
;ertion. Since gas) for XI O O O -a n f3 In P2n Ppn
e of G(s) is egual to n. X2 1. O O -a n- I PI(n-l) P2(n-l) Pp(n-I)
it is a minimal-dimen X3 O 1 O -a n - ., x+ PI (n- 2) P2(n-2)
.~ Pp(n-2) U
e or al1 of g;(s) are not
emains to be control-
xn -al
lave both control1able y
°
= [O
O
O O
PII
] x + [el
P21
.. e2
PPI
ep ] u
lmn raÍional functions
(6-56)
¡
·,_._• • • • > __ • U _ •. •.• --.". __._. " . ~ _ ••• _~ , __ ~ _ . ' - . . _ •.• _ ••. ~ ••••
.-._-- -~ ..... _----- ..- _... -_ .. '- ..
~---
~._." ..
is a realization of (6-55). The realization is always observable whether 9;(S), Realization from the
i = 1, 2, ... , p, are irreducible or noto If they are all irreducible, then the realiza Consider a q x 1 prop<
tion is controllable as well.
A
It is also possible to find lordan-form realizations for vector proper rational 9
functions. The procedure is similar to the one for the scalar case. We use an
example to iIIustrate the procedure. A
G(s) = ª
"
[
'Example 2 g,
r "¡(cx I -+
(6-57)
T~l:;~X
Note that ni, i = 1, 2, .
we draw in Figure 6-6 a block diagram of(6-57). With the state variables chosen number of rows. The i
as shown, we canobtain the lordan-form equation is equal to or lare:er t
{g¡{s), i = 1,2, ... , q} .
U ,
.ucible, then the realiza- Consider a q x 1 proper rational matrix G(s) expanded as
lction
¡
91(Sj
(;(s) = 92.(S)
gq(~)
=
[h1(0)+h1(l)S_1 +h 1(2)s-2
The Markov parameters h¡U) can be obtained recursively as in (6-29) and (6-30).
+"'J
hz{?) +h 2(l)s - 1 + h 2(2)s - 2 + ...
+... .
(6-58)
For each 9;(s), we form a Hankel matrix H¡ defined in (6-31). We then form the
composite matrix
[Hi~/+--( J}
~1_(~~ ~1: p)l) u 1 (no. of Iinearly independent rows)
T~ ~2S~~¿-_1_,p)ll U 2 (6-59)
- jJ) uq
(6-57)
Note that H¡, i = 1, 2, ... , q, have the same number of columns but different
number of rows. The integer IX; is the degree of the denominator of ?j¡(s) and f3
he state variables chosen
is equal to or laflzer th::ln the degree of the least cnmmon cip.nominator of
{9¡(S), i = 1,2, ... ,q}. If the least common denominator of t9¡{s), i = 1, 2, 3,
... , q} is not available, we may choose f3 to be equal to the sum of the degrees
of the denominator of 9¡(S), i = 1, 2, ... , q. Note that in this method, 9;(S) need
not be irreducible. .
Now we shall apply the row searching algorithm to search the linear
independent rows of T in order from top to boltom. Because of the structure of
T and Theorem 6-3, if one row in H¡ is Iinearly dependent on its previous rows of
T in (6-59), then all subsequent rows in H¡ will also be Iinearly dependent. Let
eadily shown to be con U¡ be the number of Iinearly independent rows in H¡. If all 17¡(s) are irreducible,
~nce it is an irreducible then the first IX¡ rows of H¡ will be Iinearly independent in H¡ (Theorem 6-3).
However these linear independent rows ofH¡ may not be alllinearly independent
in T because they may become dependent on the rows orH j for j < i; hence we
have (J¡:S;CI.¡, i=2, 3, ... " q. Note that if g1\S) is irreducible, then we do have
u 1 = 1X 1 · The row of H¡ which first becomes Iinearly dependent on its previous
rows of T is caBed the primar y Iinearly dependent row of H¡. If u ¡ is the num ber
of Iinearly independent rows in H¡, then the (u ¡ + 1)th row of H¡ is the primary
Iinearly dependent row. Corresponding to these q primary dependent rows,
the row-searching algorithm will yield
-Y1
<x, +1 (x, +I <X q +1
'-------y---"" '----y------""
U¡ a2
(6-60)
such that kiT = O. Note that the row vector k j has, except element 1, only Ir (Ji = O for some i, say
((JI +(J2 + ... +(JJ possible nonzero elements. This is a consequence of the
1
row searching algorithm. The k i expresses the primary dependent row of H¡
as a unique linear combination of its previous linearly independent rows of O
T. See Appendix A. e= -(131(1)'" -(1,
Now we c1aim that (;(s) in (6-58) has the following irreducible realization: O
x = Ax + bu (6-61 a)
y = ex + eu (6-61 b) o
with
hl(O)j
h 2 (0)
e= . The dimension of this
[ the rank of H.
hq(O) The assertion that (1
by using the procedure
where, for i = 1, 2, ... , q,
forward, though tediow
O 1 O o
O O 1 o
A¡¡= ((Ji x (J¡) matrix
O O
-a¡¡(2) -a¡¡(3)
for i > j,
O O O
O O O
b. =
I
h¡IU)
h¡I(2)
•
j We show next that t!
see that lb Ab .".
in (6-59); hence its rank
verified that the matrix
i
[
h¡I«(J¡)
o o 0:0
,
o 0"
: : 1 O O
'-v---" ~ '-v---"
except element 1, only If (Ji =0 for sorne i, say i = 3 for convenience of il1ustration, then
s a consequence of the
y dependent row of H¡
o : O o O: ;00"'0
:O O
O :, 1 O :0 O .". O: :0 O O
Y independent rows of
-a31«J¡) :-a32(1) ... -ad(J2): O O ... O:,.. ':0
, O O
c= O O O :1 O "·0: :00'''0
~educible realization:
(6-61 a) .: ,,
(6-61b) o o o O :0 O .". O:
'--------.....--'
(J¡ (J2
h(0)J
1 (6-62b)
e=
l
h2 (0)
hq(O)
The dimension of this realization is n = (J 1 +(J 2 + .. , +(J q which is equa] to
the rank of H.
The assertion that (6-61) is a realization of G(s) in (6-58) can be established
by using the procedure used in the scalar case. Similar to (6-39), it is straight
forward, though tedious, to verify
h¡(k + 1)
hl(k +2)
hq(k +(Jq)
Let c¡ be the ith row of C. Then we have
c¡Akb=h¡(k+\), k=O, \,2, ... ;i=\,2, ... ,q
This establishes, fol1owing (6-34), the assertion.
We show next that the realization in (6-61) is irreducible. From (6-63), we
see that [h Ah ... A" - 1 bJ consists of ali iinear independent rows of H
in (6-59); hence its rank is n and lA, b} is control1able. Similarly it can be
verified that the matrix
o
O 0J}"
"O
(6-62a)
·· .. q
· '.
O O cq A C1
q- \
~
(Jq is the identity matrix of order n; hence {A, C} is o~servable. This completes
the proof that (6-61) is an irreducible realization of G(s) in (6-58).
'-~ " '.,__ ~ _ ~ _ ~ _ ~ ~ _ ' _ ' ~ ~ _ ' ~ ' ~ ' _ O _ ' " - ~ "__ ~ _ , __
" _ ~ ~ , _ " '_'_~' __' _ " '__
. ' _ ' - ' ~ "~_' ~"~"_ -~_.,,~._ ".__,,'.'-0._ , ., - __.,__
~ ~~~~-',
260 ,,,,"U"" , RCWZA no"" """ ""CM ",U,"A"NCC ANO "''',UACA UON ReAWA I
Example 3
The leftmost matrix is t
Consider we have u I = 4 and u 2 '
using the formula in (F
Gs
()
=[.~':,:-~)2] =[s'J +s'~-JS'5
1 '\'''_,,'')+,,'6_,\'7+ S ' 8 _ ..
6
+5s' -7s' +9s· 8 -\1,,'9
7
+"'J k¡ = [O
k 2= [_1
J
O
_.1
9
'---'--
s3(~+I) uI
We form the Hankel matrix. with 2¡ =4.2 2 =4. and fi =5. The elements Corres O::
an irreducible realizati~
:(f:
, O 1 1 - J
;0: 1 1 -,3 S o
:, 1:, I -,3 5 -7 O
'\ -,3 5 -7 9 X= O
'- '
O
T=[HdS.5)J= -,3 5
-7 9 -11
H 2 ( 5. S) - : ó: -- -O" - - 6--- -1 - - -:..:: 1 l'
3
y=[~
O O 1 -1 \
O 1 -1 1 -1
1 -1 1 -1 1 T~e (u I ~ 4)th and (u 1 +0
-1 1 -\ 1 -\ wlth the slgns reversed. T
If we apply the row-searching algorithm to T, we will fina\ly obtain of9 l(s)andthefirstu 2 Mi
depends only on u 1 and u
1 It is possible to searci
-1 1 order. First we expand (;
3 -\ 1
(;(s) = H(O
-S 3 -1 1
where HU) ha ve the same o
7 -5 3 2 1
--6---0- ---6--- -3¡~ - - O ex i are defined in (6-59). W
r
-1_____________
O O _l
3i> O O 1 _ ~(!~ ----~(~! - I
- 1 - 1 O 6
J6 O O O H(2) H(3) - - - I
-\ I -1 1 O O O O
T=:
. :
,
1 -\ \ 1
2
O O O O O
l H(ex)
H(ex +1)
H(ex+1)
H(ex +2)
Ii
H
O O :Il 1 There are ex + 1 block rows i
O C[: O -3 } -4
8 U I =4 block row is associated with '
:}) O O \2 -24 ordering of rows, the Hankei
O O O -36 72 .(6-59). Now We shall apply 1
O O O O O mdependent roWS of t in ord
0- ---O ---0- ---6 --- m U2 = 1
block row is linearIy depende:
subsequent block rows are Ji¡
o o o o o
o o o o o
o o o o o 4S~e Ihe discussion in (he Subsecliol
Chaplcr 5. Sce also Schcmes I an(
o o o o o lrollabilily malrix in the second half
REALIZATIONS OF VECTOR PROPER RATIONAL TRANSFER FUNCTIONS 261
E, AND lDENT1FICAT10N
The leftmost matrix is the matrix F defined in (F-9). From the rightmost matrix,
we have (J 1 = 4 and (J 2 = 1. The fifth and seventh rOws of K can be obtained by
using the formula in (F-l1) as
k l =[ O O 1 2 ~:(~ :, O O O O OJ
k 2 =[-.l3 _19 -.1.-
18
_-.1.
\8 O O ~:l} O O OJ
'- '-v--'
(JI ~4 (J2 =1
The elements :I: correspond to the primary linearly dependent rows. Hence
an irreducible realization of (;(s) is given by
=5,
1 -3
o 1 O O O o
O O 1 O' O O
3 5
5 -7 x= O O O 1 O x+ 1 u
O O -1 -2 O 1
7 9 1 - - - -[ - - - -1 - - - 1- i - 0
9 -11 "3 9 -18 18: O
1 - - -...= i o o O: 0J
O : 1 x
1 1 O O
1 -1 The «(JI = 4)th and «(J 1 +(J2 = 5)th rows of A are taken from the k l and k 2 rows
-1 1 with the signs reversed. The vector b consists of the first (J 1 Markov parameters
1 -1 of 91(s) and the first (J 2 Markov parameter of 92(S). The form of e is fixed and
depends only on (J 1 and (J 2
t finally obtain
It. is possible to search the Iinearly independent rows of T in a dilTerent
order. First we expand (;(s) as
(;(s)=H(O)+H(1)S-1 +H(2)s-2 +H(3)s-3 + ...
where HU) have the same order as (;(s). Let ct. = max {ct.h i = 1,2, ... , q }, \Where
ct.¡ are defined in (6-59). We then form the Hankel matrix
1 H(ct.)
H(ct.+l)
H(ct. + 1) H(o: +2)
H(ct.+2) H(0:+3) .,.
H(ct. + fi
H(ct.+f3)
-l)J
O
There are ct. + 1 block rows in T; each block has q rows. The ith row of every
~
~}:~ 1 block row is associated with i)¡(s). Except from having more rows and dilTerent
-]
-4
O 8 (Jl =4 ordering of rows, the Hankel matrix in (6-64) is basically the same as the one in
O 12 -24 (6-59). Now we shall apply the row searching algorithm to search the Iinearly
O -36 72· independent rows of t in order from top to bottom. Clearly, if the ith row of a
O O O block row is linearlydependent on its previous rows, then aH the ith rows in the
O
-6 --- -cF
(J2 = 1
subsequent block rows are. linearly dependent on their previous rows. 4 After
O O O
O O O 4See the discussion in the subsections on the controllability indices and observabilily indices in
O Chapter 5. See also Séhemes I ano 2 of lhe scarch of Iinearly inoependent columns of lhe. con
O O
trollability malrix in the secano half of Section 7-2.
O O O
262 IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, AND lDENTlFICATION
REALIZ.
the completion of this search, we then rearrange the rows of tinto the form of
Example 4
(6-59) as
\'¡
T
__ H 2 (a + 1, [3)
(6-65) G(S)=[~J)-3 +G}
H q( a + 1, [3) J \' q We form í with Ct=I
Clearly the total numbers of independent rows of t in (6-64), t in (6-65), and independent rows as
T in (6-59) are aH the same. Let \'¡ be the number of linearly independent
rows of H¡(a + 1, [3) in (6-65). Then we have VI +V2 + ... +v q = (T¡ +(T2 + ... 1
+(Tq, where (T¡ are defined in (6-59). Note that generally we have (T¡ ¡'v¡, for sorne O 1
or all i. Clearly the primary linearly dependent row of H¡ in T is its (v¡ + l)th -1 4 1
row. The major difference between T in (6-59) and t in (6-65) is the orders in -1 2 O
which the linearly independent rows are searched. In 1', the linearly indepen 3 -8 -1
gerit rows are searched in the order of the first row of H¡, i = 1, 2, ... , q; the 6
1 -2 -1 3
secQnd row of H¡, i = 1, 2, ... , q, and so forth. For convenience of searching,
we search the rows of t in order from top to bottom and then rearrange it as -5 12 3 -12
T.. -1 2 1 -3
Now from the row searching algorithm and the rearrangement of the coef 7 - 16 -5 18
ficients of combinations according to the rearrangements from t to 1', we can 1 -2 -1 3
obtain, similar to (6-60),
~+I ~+I ~+I
k.=[u.¡(l) (/.,(2) ... u.,I",) o o ... o: u.,II) ... u.,I",) o o .. ' o: .. ·: (/,,,11)'" u".!".) 1 o .. · O]
'-v---- ~ ----v-'
\'1 \'} \'"
(6-66)
such that kiT = O. The major difference between (6-66) and (6-60) is that aJk),
for j > i, are generaHy different from zeros in (6-66). In other words, the (v ¡ + 1)th
row in H¡ in (6-65) depends not only on the.first V¡ rows of H¡ but also on the
first VI rows ofH 2, H 3 , . . . , H q in (6-65); whereas the «(T¡ +l)th row ofH¡ in
(6-59) depends only on the first ()¡ rows of H¡. Similar remarks apply to the
(Vi + l)th row ofH¡ in (6-65). Now we c1aim that the dynamical equation T~e rows of K correspondí
x=Ax +Bu y=Cx (6-67) tamed, by using the formul~
A¡2 ... ~l = [ 6 (
... = [-1
with A= .lA"
A 2¡ A 22
A"l
A 2q
Note that the
k2
k¡ row corres
(
I
r.~¡JT~[ -16
Lk2
will be omitted. Insteadwe give an example to iIIustrate the procedure.
=E, AND IDENTIFICATl0N REALIZATlON OF VECTOR PROPER RATIONAL TRANSFER FUNCTIONS 263
We form f with o: = max {4, 4} = 4 and f3 = 5, and then search its linearly
independent rows as
in (6-64), t in (6-65), and
: of linearly independent 1 O O 1 -3
t- ... +Vq=O"l +O"z + ... O 1 O O O 1-1
------------------
Iy we have"O"¡ f.v¡, for sorne -1 4 1 O 1 1 -3 5
of H- in T is its (v¡ + l)th
r in (6-65) is the orders in
-1
3
2
-8 -1
O 1
6 1
*
*
O O 1 -1
1 1 -3 5-7
1
------------------
ln t, the linearly indepen
ofH¡,I= . 1, 2, ... , q', the 1 -2 -1 3 O 1 O 1 -1 1-1
------------------
convenience of searching, -5 12 3 -12 -1 O 1 1 -3 5 -7 9
n and then rearrange it as -1 2 1 -3 -1 O O 1 1 -1 1 -1 1
------------------
7 -16 -5 18 3 O O O 1 -3 5 -7 9-11
rearrangem~t of"the coef
ments from T to T, we can 1 -2 -1 3 1 O O O O -1 1 -1 1 -1
O O (j) 1 -3
,,+1 O O O :I,-1
~)00'01 6-- -::(:---6--- -6 --- 4
) : . .. : (/ I q~ 1)' l/ 111\ "l/
) ': ."'
:. 1\)'"
(/21/
",,1",)
-
o o ..
.
01 O O O O ~2;
: .. ' : "',(1) .. , ",,11',) I o .' 01
(ti -- -0- - - - 6--- -0- - - - 6
O .~
O O O O O ~
1', (6-66) -------------------
O O O O O ~
;-66) and (6-60) is that aJk),
O O O O O
In otherwords, the (v 1 + l)lh -------------------
roWS ofH l but also on the O O O O O
~ the (O" 1 + 1)th row of H 1 in O O O O O
.imilar remarks apply to the The rows of K corresponding to the zero rows indicated by arrows can be 00
le dynamical equation tained, by using the formula in (F-11), as
(6-67)
kl =[ 6 O: 4 -18 : -1 O : 1 O: O O]
kz = [-1 0:-1 3: O 1 : O O: O O]
:~:1 Note that the k l row corresponds to the first row of H(k) which first becomes
linearly dependent or, equivalent\y, the primary dependent row of H l , and the
kz row corresponds to the primary dependent row of Hz. Now we rearrange
'qq
f and k¡ as
62) with O"¡ replaced by vi a~d
ment is similar to the one ¡n
the necessity of tracking the
not involve any new idea and
[~Jt~[_~ _~ -~ :~; 0:0
O: O
-V~
;)
. ~ '--v--'
Hustrate the procedure. VI Vz
264 IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATION IRREDUCIBLE REALIZ,
, I
introduced in CAkB.
=[~J
,,O·, O 1 1 -3
i = 1, 2, ... , q, on digito
:0: 1 1 -3 5 which is q(r:t. + 1) x 13, is
, I
'1'
1- _1
-31 5 -7 x 13; however, in the se;
1 - 3 5 - 7 9 ~Primary dependent row, say the ith row ir
- 3 5 - 7 9 -11 row of H 1 block rows can be skir
x -:6:- -- -0- - - - ó----( -- i
-~ thaJ1 the T in (6-59), th
, I
independent rows in t
'0'.! O 1 -1 1
1-
on computer computat
O 1 - 1 1 -1 ~ Primary dependent We discuss now an i:
1 -1 1 - 1 1 I
row of Hz Consider
-1 1 -1 1 -1J
C(s) = [gl(S) gzl
Hence an irreducible realization of G(s) is given by where H(k) are 1 x p con
O 1 O O O O denominator of aH elem,
O O 1: O O O of g¡(s). Define fi = max
I
-6 -4
O
1
O
1: O 18 x+ 1 u
---------------------
O: O
1
1 -O
O:
I
O -3 ::>
TJ:g
y=[ 1 O O O lH(a
O O O 1 It is a la + 1) x fip matrix
The v1th and (v 1 +vz)th rows of A are taken from the k1 and kz rows with in order from top to bot
the signs reversed. The vector b consists of the first V1 and Vz Markov para independent and the (v +'
meters of gl and gz· The form of C is fixed and depends only on V1 and Vz. I Then there exist a(i), i' 1
[a(l) (
Remarks are in order regarding the use of the Hankel matrices in (6-59) 1t is claimed that the dyn¡
!
and (6-64) in the realizations. First we show that the v¡ in (6-65), which are
X~,
obtained by searching the linearly independent rows of T in order from top to
bottom, are the observability indices of any irreducible realization of G(s).
Indeed, from (5-64) and H(k) = CAk-1 B, we have
L- a(1) -(;
II
~ Primary dependen!
row ofH z
We discuss now an irreducible realization for a 1 x p proper ralional matrix.
Consider
G(s)= [91(S) 92(S) ... 9 p (S)]=H(0)+H(1)s-I+H(2)s-z+ ...
where H(k) are 1 x p constant matrices. Let ex be the degree ofthe least common
denominator of aH elements of G(s) and let {3i be the degree of the denominator
of 9¡(S). Define {3 =max{{3i' i = 1, 2, ... ,p}. We form the Hankel matrix
H(l)
H(2)
H(2)
H(3)
H({3)
H({3 + 1)
1
T=
f~(ex+1) H(ex +2) ~(ex + {3)
It i5 a (ex + 1) x f3p matrix. ; TOW we search the linearly independent rows of T
in order from top to bottom. 1t is assumed that the first V rows are linearly .
the k1 and kz rows with independent and the (v + l)t& row is linearly dependent on its previous rows.
VI and V 2 Markov para- Then there exist a(i), i' 1,2, , v, such that
Ids only on VI and V2' • [a(1) a(2) a(v) 1 O ... O]T =0
iJ ~ ~ ~ O
O
:
l l x +:
f"(l)
H(2)
u (6-68)
I O O O 1 ! ! H(v - 1)!
L- a (1) -a(2) -'l(3) -a(v)J LH(v) J
~he postmultiplication of y=[ 1 O O O ]x + H(O) u
: the linearly independent
is an irreducible realization of the 1 x p proper rational transfer matrix. The
; the linearly independent
proof of this statement is similar to the column case and is left as an exercise.
Itly, we conclude that the
irreducible realization of
xample4. *6-5 Irreducible Realizations of Proper Rational
dices, we have Matrices: Hankel Methods
= 1, 2, ... , q}
There are many approaches to find irreducible realizations for q x p proper.
an or equal to the one in ¡ rational matrices. One approach is to first finda reducible realization and then
lre used in the realization [.
apply the reduction procedure discussed in Section 5-8 to reduce itto an ir
"
~t
in the realization in (6-60). reducible one. We discuss this approach first.
~, the more errors may be
266 IRREDUCIBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTlFICATlON IRREDUCIBLE REALI2
Given a q x p proper rational matr~x G(s) = 19¡/s)), if we first find an irre +R implies
ducible realization for every element g¡j(s) of G(s), and then combine them
together as in (4-61) or Figure 4-7, then the resulting realization is generally not
controllable and not observable. To reduce this realization requires the
application of the reduction procedure twice.", If we Jind the controllable
canonical-form realization for the ith column, G¡{s), of G(s), say
X¡ = A¡x¡ + b¡u¡ y¡ =C¡x¡ +e¡u¡ and
where A¡, b¡, C¡, and e¡ are of the form shown in (6-53), u¡ is the ith component sVm(s) =
of u and y¡ is the q x 1 output vector due to the input U¡, then the composite
These equations impl)
dynamical equation
(sm +
O
l}}l1'
Y = [C¡
A2
O
C2
fIJ and
Consider
C(sI-A)-¡B
A
is a realization of G(s). (Pro ve.) Because of the structure of A¡, b¡ it can be which becomes, after tl
readily verified that the realization is always controllable. It is however gener
ally not observable. To reduce the realization to an irreducible one requires
C(sI -A)
the application of the reduction procedure only once. A
Let G(s) = G(s)+G(oo), where G(s) is strictly proper. Let tItes) be the monic This shows that Equati
least common denominator of (;(s) and of the form of A and B, it is easy to
generally not observabJ
tItes) = sm +IX¡ sm - ¡ +1X 2 S m- 2 + ... +lX m (6-69) We note that the reé
Then we can write (;(s) as form realization for a sc
to find a generalization
1
G(s) = t{t(s) [R¡sm-¡ +R 2 sm- 2 + ... +RmJ
A
(6-70) A
G(,
where R¡ are q x p constant matrices. Let I p be the p x p unit matrix and Op where H(i) are q AX P co
be the p x p zero matrix. Then the dynamical equation denominator of C(s} an
we can show
OP]
Op [OP]
Op
x+: u (6-71 a)
H(m +i)= -1X¡H(m +
Op Op Ip Op
-lXm-¡I p -lX m_ 2 I p -1X1I p Ip This is a key equation ir
. Let {A, R, C, E} be a
R¡ J x+ G(oo)u (6-33),
(6-71 b)
G(s) = E +C(sI - A)-¡ R
is a realization of (;(s). To show this,itis sufficient to show C(sI -A)-¡ B = (;(s).
From (6-73) and (~-75),
Define V(s)"~ {sI '-'- A)-IB or (sI - A)V(s) = B. Ves) is a mp x p matrix. If
is arealization of G(s) in
we partition it as V'(s) = [V'l (s) Vi(s)'" V~(s)], where the prime denotes
the transpose and V¡(s) is a p x p matrix, then (sI - A)V(s) = B or sV(s) = AV(s)
¡ HI
IRREOUCIBLE REALIZATlONS OF PROPER RATlONAL MATRICES: HANKEL METHOOS 267
;, AND IDENTIFICATlON
+B implies
if we first find an irre
Id then combine them SV I(S)=V 2(S)
lization is generally not sV 2(S) = V 3(S) = S2V I(s)
:alization requires the
: find the controllable
G(s), say
and
¡Ui
Consider
V ,(s)- !/J(s) i=I,2, ... ,/1'l
bp p
p x p unit matrix and Op where H(i) are q AX P constant matrices. Let lf¡(s) be the monic least common
)ll
denominator of G(s) and of the forro shown in (6-69). Then similar to (6-30),
we can show
: }+[r}
ti I p lp
(6-71.1
This is a key equation in the following ~evelopment.
i = 1, 2, ...
R¡ ] x+ (;(oo)u (6-33); .
(6-71b) (6-75)
.how C(sI -A( ¡B =G(s). Fro~ (6-73) and (6-75), we may conclucle, similar to (6-34), that {A, B, e, E}
is a mp x p matrix. Ir is a reaiization of G(s) in (6-73) if and only if E = H(O) and
where the prime denotes
)V(s) = B or sV(s) = AV(s) i=O,I,2, ... (6-76)
268 IRREDUCIBLE REAUZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATION IRREDUCIBLE REAU¡
Oq Iq Oq Oq H(1)
Oq O Iq Oq H(2)
x= X+ U (6-77a)
Oq Oq Oq Iq H(m -1)
-C(m1q -G(m_I1q -G(m-2 I q -G( 1 I q H(m)
y= [ I q O O
-IX m - zl q
:. J
-IXll q
realization is a unit mat
realization however is g(
Now we shall use TI
O O O -IXml p an irreducible realizatio
Ip O O -IXm_Il p existence of qm x qm anc
N~ O Ip O -IXm-zl p (6-79b)
O O Ip -IXll p
where 1: = diag "Ir Al' A2,·
where 1" denotes the n x n unit matrix. We also define the following two roots of the eigenvalués
, ANO IDENTIFICATION
IRREDUCIBLE REALlZATIONS OF PROPER RATIONAL MATRICES: HANKEL METHODS 269
Hankel matrices
:J( 1)
~(2)
[ H(l)
T~ ~m
H(2)
H(3)
H(m)
H(m +1)
J (6-80)
u (6-77 a) H(m) H(m +1) Il(2m-l)
H(m -1)
H(m)
H(2) H(3) H(m +1)1
[(O)u (6-77b)
and T~ ~(3) H(4) H(m +2)
(6-81 )
[
(6-74), we can readily H(m + 1) H(m +2) H(2m)
Since T and T consist of m block raws and m block columns, and since H(i)
¡
are q x p matrices, T and tare of order qm x pm. Using (6-74), it can be readily
iB =
HU +
~U + 2)
0J (6-78)
verified that
T=MT=TN (6-82)
H(i +m) and, in general,
where the prime denotes the transpose. From this equation and (6-76), we
conclude that ,,.{ A = M, B = TI'p,pm' e = I q .qm , E = H(Ol} is a qm-dimensional
nsider the q x p proper realization of G(s) in (6-73). Note that this realization is the one in (6-77) and is
l Slll - 1 + ... + rt. 1II be the
observable but not necessarily controllable. Similarly, the dynamical equation
)efine the qm x qm and e = I q . qmT = [H(1) H(2) ... H(m)]
A=N (6-85)
R= I'p,pm = [I p o O]
lE = IH(O)
A
. where I: =diag P'l, Al, , .. , An}and Ai , i = 1,2, .' .. , n, are the positive square
efine the following two roots of the eigenvalues of T*T, where theasterisk stands for the complex
270 IRREDUCIBLE REALlZATlONS, STRICT SYSTEM EQUlVALENCE, AND IDENTlFICATlON IRREDUCIBLE REALl
conjugate transpose. Clearly n is the rank of T. Let K 1 denote the first n Consider
columns of K and L 1 denote the first n rows of L. Then we can write T as
A 2 =(Y+Tl
T= Kl~Ll = Kl~1/2~1/2Ll ~ YU (6-87) =(Y+Ml
=Y+MT
where ~1/2 =diag {JAl, JA2>"" JAn}, y ~Kl~1/2 is a qm x n matrix, and =Y+MT
U = ~1/2Ll is an n x pm matrix. Defines =Y+MM
and (6-88) =Y+M 21
Because of KfK 1 = 1". L¡Lf = 1" (see Problem E-7), we have Repeating the proces~
Consider a q x p... proper rational matrix (;(s) expanded as (;(s) = L:::o H(i)s-i.
= Iq,qm YUU+,
'---v--' '---v-' '
We form T and T as in (6-80) and (6-81), and factor T as T = YU, and pT = p y = C In
pU, where p denotes the rank, by using the singular value decomposition. =CAiB, i,
Then the {A, B, C, E} defined by
This shows that (6-91)
A=Y+TU+ (6-91 a ) The dimension of
B = first p columns of U = UI'p,pm (6-91b) Am-1B] and V'=[C'
e = first q rows of Y = Iq,qm y (6-91c) T=VÜ and'n=Tsmi
E=H(O) matrix, we conclude ¡
A
We give first a justification of (6-91) before giving a formal proof. We see The crux of this 1
from (5-65) that Y and U are, respectively, the observability matrix and the Numerically stable co
controllability matrix of {A, B, C}. Consequently, we have (6-91b) and (6-91c). Hence this realization (
Again from (5-65), we can easily verify that t = y AV. The pre- and postmulti The decomposition
plication of y+ and U+ to T = y AU and the use of (6-89) and (6~90) yield choose Y = K 1 and U =
y
'f and ~J def]ned in (6-8
(6-91a). Since the dimension of A is equaí io the rank ofT, the irreducibiíiiy of
(6-91a) follows from pY = pU =dim A.
Y*Y==
Now we give a formal proof of the theorem.
A realization whose e
Proof of Theorem 6-4 property V*y = UU* i
speaking, the signal tral
Define
state to the output are s
T+~U+Y+ (6-92) (see Problems 6-22 and
Then we have, by using (6-87), (6-89), and (6-90), In the application, it
model for a given systeJ
TT+T =VUU+Y+YU = YU =T (6-93) to this problem. For e)
Consider
K1 denote the first n
Ne can write T as A 2 =(Y+TU+)2 = (Y+MTU+f [ using (6-82)J
=(Y+MTU+)(Y+MTU+)
(6-87)
= Y+MTT+TNU+ [using (6-92) and (6-83)]
1 qm x n matrix, and =Y+MTNU+ [using (6-93)J
=Y+MMTU+ [using (6-83)J
=Y+M 2TU+
1 (2 (6-88)
Repeating the process we can show
,ave
Ai=Y+MiTU+ i = 1, 2,3, ...
(6-89)
Consider (6-84):
(6-90)
H(i + 1) = Iq,qmMíTI'p,pm
ling theorem. = Iq,qmMiTT+Tl'p,pm [using (6-93)J
= Iq,qm TNiT+Tl'p,pm [using (6-83)]
= Iq,qmTT+TNiT+Tl'p,pm [using (6-93)J
= 1 YUU+Y+MiTU+Y+YUI' [using (6-87), (6-92), and (6-83)J
lS G(s) = I;':o H(i)s- í. ~ '--v-' '-v-----''-v--'~
C II! Ai II! B
value decomposition. A
where 1: = diag {A¡, A2 , • • . An} and Al ;:O:A 2 ;:o: •.. ;:o: Ano If the system is to be bottom. After the Cl
approximated by an m-dimensional model with m < n, then the reduced model into the form
can be obtained from
H¡¡(a+1',B) ]
T = K I diag{A¡, A2 , . .. ,Am}L¡
T= ~2¡(a+1,,B) ]
where K¡ is the first m columns of K and L¡ is the first m rows ofL. See Refe
rences S141 and S162. [
Hq¡(a +1, f3) ]
To conclude this subsection, we remark that the dimension ofthe irreducible
r~alization is equal to the rank of T. Becallse of Theorem 6-2, we have deg as in (6-65). Using th
G(s) = rank T. ronsequently, the degree of G(s) can also be computed from T. q primary linear1y de
Ifthe degree of Gis known a priori, then the m in (6-80) can be chosen to be the (6-67) can be readily
least integer such that the rank ofT is n. Because ofT = VU, from the definitions servability indices of ~
of controllability index /l and the observability index v, we may conclude that illustrate the procedu
the 1east integer to have pT = n is equal to
Example 1
max{/l, v}
Consider the proper r
Note that /l and v can also be ~ obtained directly from the column and row
degrees of coprime fractions of G(s). See Theorem 6-6 on page 284.
where Hij(a¡ + 1, {3j) is the Hankel matrix of g¡j(s) defined as in (6-35). Now
we search the linearly independent rows ofT in order from top to bottom. Then
corresponding to the primary linearly dependent rows of [H i ¡ H i2 •.• H ip ],
i = 1, 2, ... , q, we may obtain an irreducible realization similar to the one in
(6-61). Another way is to form the Hankel matrix
ANO IDENTIFICATION IRREDUCIBLE REALIZAnONs OF PROPER RATIONAL MATRICES: HANKEL METHODS 273
If the system is to be bottom. After the completion of the search, we then rearrange the rows of T
en the reduced model into the form
Hll(O:+l,{3) Hdo:+1,{3) ... HIP(O:+l'{3)]}Vl~nO'OflinearlY
.. H 21 (0: + 1, {3) H 22 (0: +1, {3) .,. H 2io: + 1, {3) mdependent rows)
T= . . .
rowS ofL. See Refe-
[ : : : (6-96)
H q1 (0:+1,{3) Hdo:+1,{3) H qp(o:+l,{3) }v q
Ision ofthe irreducible
as in (6-65). Using the coefficients of linear combinations corresponding to the
rem 6-2, we have deg
q primary linearly dependent rows of T, an irreducible realization similar to
I be computed from T.
(6-67) can be readily obtained. Similar to (6-65), the Vi in (6-96) yield the ob
an be chosen to be the
servability indices of any irreducible realization of G(s). We use an example to
U, from the definitions
illustrate the procedures.
we may conc1ude that
Example 1
l
Consider the proper rational matrix
the column and row -2S 2 -3s-2
l
)n page 284.
G(s) = (s + 1)2
S2 (6-97)
ection we shall search 4s +5 -3s-5
to find an irreducible s +1 s +2
~ctor case discussed in We compute the Markov parameters:
ion before proceeding.
the degree of the least -2s 2 -3s-2
the degree of the least (s + 1)2
heir computations are
e denominator degrees
;o:¡, i = 1, 2, ... , q} and
)rm the Hankel matrix 4s+5
- - = 4 +S-1_ S-2 +S-3 -S-4 + ...
p~l}
, {3 O" 1 (no. of
s +1
, {3 p) } 0"2 linearly -3s-5
and - - - = -3 +S-l -2s- 2 +4s- 3 -8s- 4 +16s- 5 - •..
independent s +2
Pp) J} O"p rows) It is clear that 0:1 =4, 0: 2 =2, {31 =2 and {32 =3. We form
(6-94)
T=[H l l (O:I +1,{31) HdO:l +1, {32)J
tled as in (6-35). Now
H 21 (0:2 +1, {3¡) Hd0:2 +1, {32)
:n top to bottom. Then
[H i1 H i2 ... H¡pJ, .,--f: -2 :0
I I
,
I 1 O
In similar to the one in
.:-2: , 3 , ,
,1 , O O
,
, 3: -4
,
:0 ,, O O
~~) + 1)]
:-4:
1 ___ I
5 :0
...
_~
, O O
,
5 -6 , O O O
(6-95) ------- - - - - ., ------- - - - - - - -
: 1: -1
---
.; 1: -2
1--
4
[(o: +{3) -1 1 :-2 4 -8
ws in order from top to 1 -1 4 -8 16
........__ ... _=
.....=.....=.. ==
274 IRREDUCIBLE REALIZATIONS, STRICf SYSTEM EQUIVALENCE, AND IDENTIFICATION IRREDUCIBLE REAL!
We note that hJO) are not used in forming T. The application ofthe row-search Next we use í ar
ing algorithm to T yields
1
G(s) =
A [-2 0J
+ 4 -3
O 1
O O 1
O O .±
3
1
O O ~ 2 1 :T We hav~ IX =max{IX I ,
2-- f -- ----=-"4- - - -ó-- - -1- -
~ ~~ matrix T and apply t
3
42390 210 1
-8- - --=4---=- -17 - ~~ f9- ---O-- =--4- ---0- ---f 2 1
3
1 O-2 1 O O -1 1
-2
3 O
1
-4
3 O
O
O
O
}., ~4 -4
O
2
O
O
O
O 1. O O O 8 -4 O
3
O O O O O O O O
-------------------- -16 8 O
O O O O 4 } ~2 = 1
O O O
O O O O O
32 -16 O
O O O O O
(6-98) 1 O: -2
Clearly we haye 0"1 =4 and 0"2 = 1. Hence the dimension of any irreducible 1 1: -1 -L
realization of G(s) is 5. Corresponding to the fifth and seventh row ofthe right -~ 2--- -f -;-- 3--- -e
hand-side matrix of (6-98), we can compute, by using (F-l1), from the leftmost -1 -2: 1 4
matrix in (6-98): : --:3 --- 0- -;--=4 ---()
kIT=[O O 1 2 :I~: O O O]T=O 1 4 :'-1 -8
~ 4--- -ó -; -- 5----<i
and k 2 T=[0
~
O 1 1 O :2
'-v-J
()) O]T=O
---
-1 -8: 1 16 1 __
O Corresponding to the 1
~----~--~~--=--~--l-~ _~~ ~
as shown, we can com]
x= x+ __ u kl = [O 2
O O -1 -1 :-2 1 1
k 2 = [O 2
Y= [ ~ ~ ~ ~ ~ \ ] x +[-~ _ ~ ] u
The elements (Ü corres
rearrange tinto t anc
TheO"lth and (0"1 +0"2)th rows of A are taken from k l and k 2 with the signs
reversed. The matrix B consists of the first O" I Markov parameters of 9I l (s) [~IJt~[O,O
2 .
O1((
O O O
and 9rz(s) and the first 0"2 Markov parameters of 921(S) and 922(S). The form of '-----v----'
ANO IDENTlFICATION IRREDUCIBLE REALlZATIONS OF PROPER RATIONAL MATRICES: HANKEL METHOOS 275
tion ofthe row-search Next we use t and t to find a different irreducible realization. We write
G(s) =[-2
4
0J +[1 0J
-3 1 1
S-l +[-2-1 -21J s -z
+[3 1
0J
4 s
-3 +[-4-1 -8
°JS-4+ ...
We have a=max{a 1 , az} =4 and p=max{Pb pz} =3. We form the Hankel
matrix t and apply the row-searching algorithm to yield
1
2 1
O -1 1
I~ ~ ~ }(J1
O O O
=4
-4
O
2
O
O 1
O -3 1
8 -4 O 3 O 1
lO O O
O O O 4 1 O 1
lO O O
l· - -- - ---- - -16 8 O -7 O O O 1
10 O 4} ~z = 1 O O O -5 -1 O O O 1
10 O O 32 -16 O 15 O O O O O
O O O
(6-98) 1 O -2 1 3 O
3 1 -5 O 7 4
ion 'of any irreducible --------------------------
:venth row ofthe right -5 O 8 O -11 -4
1-11), from the leftmost 1 O -1 O 1 O
--------------------------
O O -1 O 2 O
O O O O O O~
=0 --------------------------
O O O O O O <
=0 O O O O O O
--------------------------
O O O O O O
- Q 1}
0-'
1 Corresponding to the primary Iinearly dependent rows indicated by the arrows
O u as shown, we can compute, by using (F-1l), the coefficients of combinations as
O 2 O 1 1 O (n O O O] (6-99)
1 2 O 3 O 'I: O O O O]
The eleme~ts :X: c,?rrespond to theyrimary linearly dependent rows. Nowwe
rearrange Tinto T and rearrange k¡ accordingly to yield
and kz with the signs
v parameters of 911 (s) [kz~lJt~[O
- ,O
O 1 0.) 0;2 1 O 0J
O O O O: 2 3: f: O O
'-v------" ~
?
.
,nd gds). The form of
V1 =3 vz=2
1.
... ~:::":,:::,=----::..:'..:.,:",:,, _::...:...,. ::;::' ._--",_..... -::::...7"....:.:....:::.:=;:.• ..
, .~~, .. u·,,_,. _•.. ,_~.,.,'~""" .•• ~'.~ __ "_n-"'.~ __ ~·~ . :.:.:::::.':.:.._._J'~:"_":::':'~:"':'::=--'--....:'.:::._~:'::"":'::":'':':::::::-'':':':'''''''':~::'::::'':';;~:,:":,:_::::_::: ..:::.~:.::::::::~~..::-::::~~;-~~=----=---:::=:.::::::.::.:::.:..:::....::.o=~~:::.....:..::.::;::_:
. . .:.::~_ _~~.-
x ~:t-~{----~-~:---~--:~~---~ =[~J
I 1 I I
'-1'
1
1 1
-1 :'-2' 4 -8 1 ----1
'
1 -1 1: 4 -8 16 ~
- 1 1 - 1 :, - 8 16 - 32 O
1 -1 1: 16 - 32 64 H(s)= O
O O 1: O O -2 1
x=
O O -1 :-2 -1 x+ 3 O
u
-----------~------- -------
O 00:01 11 The matrix H(s) is a p x p
O O O: -2 -3 -1-2 on the diagonal. The m
That is, n is the sum of al! the column degrees of D(s). Let us write
D(s) = DhcH(s) + DlcL(s)
with
,
, o :0
=[~J
s
,
, o
,
,
:0
,
,, :.
,
,
Sil! O O O Slll -1 , O , :0
O Sllz O O
_ _ _ _ _ _1 _ _ _ _ _
O
,
, 1
J
., L
:0
_
~ct
transmission part of a Since D(s) is column reduced, the constant matrix D hc is nonsingular. Hence
m. Using the procedure (6-105) can be written as 6
H(s)v(s) = - Dh;, 1 DfcL(s)v(s) + Dh;, 10(S) (6-107)
Let us define
matrices and are right x(s) = L(s)v(s) '. (6-108)
nd ¿¡ciN(s) <¿¡ciD(S) = ¡..ti,
6This step does not appear in the scalar case because D(s) i~ normalized to be ~~nic beFare' ;he
realization.
1
- ... - ...
-.--------
~ , ~.-
x12 ¡j1
V<¡" -1)
Xl!"
------
XZ1 Vz
= : .(6-109)
V~2 -1)
Xz!' 2
-------
-------
X p1 Vp
V(!'p-1)
XP!'p p
where Vi is the ith component of v, and V\k) = dkVi/dt. From this definition we
have
i = 1, 2, ... , p; k = 1,2, ... , Jli-1 (6-110)
r xP!'p
where al!'; denotes the ¡th row of Dh;, 1 D re , and bl!'i denotes the ith row of Dh;, 1.
ap!'p bp!,p
Figure 6-7 Block diagn
A block diagram of(6-11O) and (6-111) is shown in Figure 6-7. It consists of p
chains of integrators. The output of each integrator is chosen as a state variable
as shown. This block diagram is a generalization of the one in Figure 6-3.
From this block diagram or, equivalently, from (6-110) and (6-111), we can
where "x" denotes poss
readily obtain the following dynamical equation.
-al!", the (Jl1 + Jlz)th
O
o
1 O
O ]
O
O ,'
o o
l 10
·0 1 (6-112b) follows directl
We note that if P'i = 0, t;
, 6-7 and the ¡th row of ]J
,
O O O , O
x x x x :x
(-a,",) X ... X: :X x x ... x b,",
- - - - - - - - - - - - -: -0- - I - -0- - - e - - .-. : - ó: ------------ -0 - Example 1
:O O 1 O: , O Consider the strictly pro
o l'
I .
:
•
l ...
I
I
I o x(t) + u(t)
'o O O 1" O -6s-2
X
- - -
wX
- -
X
-. - -
...
- - -
x:
- -4 -
X
- -
...
-
(-a1¡t) ... x: l·
:X
1- - -
X
- -
X
-: - - - - -
X
_~1"_' A
_3S2
(S +1)3 (S-;
\J(S) =
-------------1---------------- I :0- -1- - 0- - - - -O ---
, O [
. S
, (s-~
,
, ,'o O 1 O O (s+I)3
o ,
, O , ,:
= [(S + 1)03(S - 2)
, , :0 O O 1
, , (s +1
xpp. X X X X : ..x ( -ap"p) x, ,x x x x bp "
(6-112a)
ANO IDENTlFICATlON
IRREDUCIBLE REALIZATIONS OF G(S)-COPRIME FRACTlON METHOD 279
XllI100'0-1~SI
xII
_~ - (/11 integrators)
(6-109) u
'00-'"
[}Y
1.
s
PI
)
(/1 p integrators)
.. , Jl¡-1 (6-110) y
I'!]
l
1
b
(t) + b~l'l u(t) (6-111)
bpl'p
.. -1
Figure 6-7 Block diagramofG(s)=N(s)D-1(s).
)tes the ith row of Ohc .
ure 6-7. It consists of p
hosen as a state variable
le one in Figure 6-3. y(t)= Cx(t) = N1cx(t) (6-112b)
110) and (6-111), we can where "x" denotes possible nonzero elements. The Jllth row of A is equal to
- 3 11',' the (JlI + Jl2)th row of A is equal to - 3 21'2 and so forth. Equation
(6-112b) follows directly from (6-106) by the substitution of x(s) =L(s)v(s).
1
We note that if Jli =0, then the ith chaLa of integrators wíll not appear in Fig.
o 6-7 and the ith row of O¡;;,IOIe and O¡;;,I wil1 not appear in Equation (6-112).
x X .. , X Example 1
----------
Consider the stricdy proper rational matrix
o(t)
o +
x .~ .~
x(t)
A
_3S 2 -6S-2
(s + 1)' (s -
s' -3.1-1
2)(.1 + 1)' (s -
1
2)(.1 + 1)2
J
",(s) =
.:----- [
s
100 O (s + 1)3 (s -2)(.1 +1)'· .: (s -2);.1 +1)2
O 1 O O
.
=.[(.1 + 1)'(.1 -
ü·
2)
(s
O ..
+ 1)'(.1 - 2)
J-'[(
.
~3S2 -6.1 :-2)(.1 -2)
. .1(.1 - 2)
s3.-:3s-1 s +1 ]
s(s + 1)
O O 1
x x x
(6-113)
(6-112a)
-------'- ' . . • ~ . .""c"'
•....:.:c." •. ~ - - - "' .. " " "" _ --.:=
"=:.""=.,,=::. . .
This fraction is obtained by using the least common denominator of each is a real ization of G(
row and is generally not irreducible. From the coefficients of this fraction, showing that the A, ]
we can obtain, by using the procedure discussed in Appendix G, the following
C(sI -A)-I]
irreducible fraction:
or equivalently,
~Jr +3<t'+1
-1 -l -1
~
G(s) =
[-3S 2
-6s-2 1
s-2 -3
(sI - A)
s O
O J
1
This equation implie~
BD(s)
Bence we have /11 = 3, /12 = 1, and /13 = O and From (6-112), wesee Ü
n=/11 +/12 +/13=4 are zero rows. Furth
(I~= 1 /1m)th row of
We write O(s) and N(s) in the form of(6-100) and (6-102) as
-2 -6 O O
N(s) = [ O -x -x
1
(si A)L(s) =
Note that, beca use of /13 =0, the last colun;m of L(s) is a zero column. We
compute
-x -x
3
O
O
3
O
o
3 -2
O
O
-ll O
L-x -x
(6-114a)
-2 -6 -3
y= [ \6-114b)
O 1 O
Note that, because of /13 = O, the third row of O~ 1 and Ol~ 1Olcdo nbt appear in
(6-114).
The dynamical equation in (6-112) is developed from the block diagram in
Figure 6-7 which, in turn, is developed from (;(s)=N(S)O-I(S); hence(6-112)
, AND IDENTIFICATION IRREDUCIBLE REALIZATIONS OFG(S)-COPRIME FRACTION METHOD 281
denominator of each is a realization of G(s) = N(s)D -I(S). We shall now formally establish this by
cients af this fraction,
showing that the A, B, and e in (6-112) satisfies
endix G, the fallowing C(sI - A)-I B = N1c(sI - A)-IB = N(s)D-I(s) = N¡J,(s)D- I(S)
or eguivalently,
-1
s-2
-IJ-l
-3
(sI - A) - lB
This eguation implies
O 1
BD(s) = BDh/H(s) + D¡;/ D¡cL(s~J =(sl - A)L{s) (6-116)
From (6-112), we see that all rows om, except the {L~ = 1 J.!m)th rows, i = 1, 2, ... , p,
are zero rows. Furthermore, these nonzero rows are egua! to D~ 1. Hence the
(L~ = 1 t'm)th row of
!) as BDhc(H{s) +Dh;, 1 D¡~{s))
is egua! to the ith row ofH(s) + Dh;,1 D1cL(s) for i = 1,2, ... , p, and the remainder
are all zero rows. To establish the eguality in (6-116), we write
o o: ,
s
o
-\
-\ o:
• I
.,
,
I
: I ,
, ,
o O o ... -1 : , ,
, , ,
~X -x -x S-XI -x -x o" -XI l-X -x ... -x
- - - - - ~ - - - - - - - - - - - - - - - -:- - - ~ - - =-.- f ------- 0-: :- -- - - - - - - - - -
I • ,
, O
, 0 , ,,
(sI - A)L(s) = '
I : I I
, I
is a zera column. We o o -1: :
-
-x
- - -
-x
.. - - - -
-x .,.
- - - - -. - -
-x I -x
- - - - -1- - - - -
-x
- - - - - - - - - -
s-X:
-
:-x
- - •• - - - - -
-x ...
_.- - - - - -
-x
- -
. .
- - - - - - - - - - - - - - - - - - - - - -:- - - - - - - - - - - - - - - - - -: :- -5- - --:.! - -.-.-. - - --o
:
,
: : o
1 I •
O
I I I :
-ll
I ,
-x -x -x ... -x' -x -x ... -x' I-X -x s-x
3
3
O O -2 I \
, <;
O
~ l
O 0, I
S/JI-I: O:
--o-~---(-:
, ,
o , I S
,
I
(6-114a) x ,
I :
. ,
I
,
I ,
,
O : S,J¡-l: O
\6-114b)
From this equation, we can readily verify that all rows, except the eL¡
m= 1 J.!m)th Proof
row, for i = 1, 2, ... , p, of (sI - A)L(s) are zero rows. Because the (¿~ = 1 J.!m)th We first show that if
row of A is equal to the ith row of - D hc 1 D le, as can be seen from (6-111), the
observable. Under t:
(¿~=1 J.!m)th row of (sI-A)L(s)=sL(s)-AL(s) is equal to the ith row of
det D(s) such that G(~ I
H(s) + D,;/ DlcL(s). This establishes the equality in (6-116). Hence the dy
find a realization of G
namical equation in (6-112) is a realization of (;(s) = N(s)D- 1 (s).
(6-112) is not irreducn
We discuss now the relationship between det(sI -A) and det D(s). From
ever, (6-112) is known
(sI - A)-IB = L(s)D-l(S) in (6-115), we have
shalI establish this st;
Realization of N(s)D-l(S) where D(s) and N(s) are not right coprime
or
The realization procedure discussed from (6-100) to (6-112) is applicable
whether G(s) = N(s)D- 1 (s) is a coprime fraction or not as long as D(s) is column Since L(A) consists of a
reduced and (\¡N(s) < (\¡D(s). Now we shall show that the dynamical equation is the n x 1 vector L(A)a
in (6-112) is controllable whether D(s) and N(s) ;lre right coprime or not. This does not have a full ran
can be established by employing the condition rank [B AB ... An - 1 B] = Now we show that
n or the left coprimeness of (sI - A) and B. For this problem, it is easier to right coprime. Under 1
employ the latter. From (6-115), we have on page 231, an eigenva
6-119)
Since L(s) consists of, as can be seen from (6-101), a unit matrix of order p as a
submatrix, we have Now we show that t.hen
rank [AI-A
Theorem 6-5
which implies, because {,
The controIlable realization of (;(s) = N(s)D - I(S) in (6-112) is observable if and
only if D(s) and N(s) are right coprime.
IRREDUCIBLE REALlZATIONS OF G(S}---COPRIME FRACTlON METHOD 283
, AND IDENTlFICATlON
Proof
(cept the (I. (11= 1 JLm)th
i
:cause the (I~= 1 JLm)th We first show that if O(s) and N(s) are not right coprime, then (6-112) is not
•seen from (6·1-11), the observable. Under the premise, we can find a 6(s) with deg det 6(s) < deg
lal to the ith row of det D(s) such that (;(s) = Ñ(s)D - 1 (s) = N(S)O-I(S). Usin~ Ñ(s)D -1(S), we can
-116). Hence the dy find a realization of (;(s) with a dimension smaller than that of (6-112). Hence
5)0-1(S). (6-112) is not irreducible (controllable and observable) (Theorem 5-19). How
~) and det D(s). From ever, (6-112) is known to be controllable; hence (6-112) is not observable. We
shall establish this statement once again by using a different argumen1. If
O(s) and N(s) are not right coprime, then there exists at least one s, say s = A,
s) L(s) Adj O(s) such that
I (6-112) is applicable
s long as D(s) is column Since L(A) consists of a unit matrix of order p as a submatrix, if lX is nonzero, so
the dynamical equation is the n x 1 vector L(A)lX. Hence we conclude from (6-121) that [(sI-A)' CJ'
t coprime or no1. This does not have a full rank at s =A, and {A, C} is not observable.
: AB .,. An- 1 BJ = Now we show that if {A, C} is not observable, then D(s) and N(s) are oot
problem, it is easier to right coprime. Under the premise, there exists, from the dual of Problem 5-37
on page 231, an eigenvalue Aof A and its n x 1 eigenvector e such that
6-119)
[AI~AJe=o (6-122)
.t matrix of order p as a
Now we show that there exists a p x 1 vector el! such that
e = L(A)lX and O(A)lX=O (6-123)
(6-120)
To show this, consider (6-116) or
. From the realization
[AI-A L(A)J (6-124)
O(s). Hence Corollary -BJ [ D(A) =0
nk[sl-A BJ=n,for
leorem 5-7). Up to this The composite matrix [H-A - BJ is of dimension n x (n + p); hence
sed. Theorem 2-5 implies
Consequently, the p linear independent columns of [V(A) D' (},)]' form a Proof
basis of the null space of [Al - A - BJ. Because of
Controllability indice
similarity transformat
[U-A
singular matrix. eOll
properties of (;(s). He
the vector [e' O']' is in the null space. Hence there exists a p x 1 vector IX
theorem. Ir we rearra
such that
[OeJ = [L(A)
0(,,1.)
J
IX
where D hc is nonsingl
becomes
This establishes (6-123). The substitution of e = L(A)ex into (6-122) yields (6-121)
"
~
o
which implies, by using (6-116), (6-102), and (6-112b),
O .
8'=
O(A)J ex = O
[ N(A)
O
Hence O(s) and N(s) are not coprime. This completes the proof ofthis theorem.
Q.E.D. l'
where the prime denote~
Theorem 6-5 implies that the realization in (6-112) is irreducible ifand only readily show that the (
if (;(s) = N(S)O-l(S) is a right-coprime fraction. Ir (;(s) = N(s)D-1(s) is a right column degrees of D(s).
coprime fraction, then we have, according to Definition 6-1 ,
Observable-Form RI
deg (;(s)=deg det O(s) realization discussed ir
observable form realiza1
Consequently, the realization in (6-112) is irreducible if and only if deg (;(s) = extensively in the next s
dim A. Since all irreducible realizations of (;(s) are equivalent (Theorem 5-20), Consider a strictly p
we conclude that a realization of (;(s) is irreducible if and only if its dimension is
equal to the degree of (;(s). This establishes essentially Theorem 6-2.
The realization in (6-112) is a generalization of the controllable canonical where the polynomial Ir
form realization discussed in (6-26). Hence we shall call (6-112) a multivariable reduced. We take the tl
controllable-form realization. We note that the realization depends on the
fraction N(S)O-l(S) used. For example, ifwe use (N(s)U(s»(O(s)U(S»-l, where
U(s) is a unimodular matrix, and D(s)U(s) remains íO be cnturrn reducr;c1 th,,;'-,
we will obtain a different multivariable controllable-form realization. Hence, Furthermore, N'(s)and [
unlike the scalar case, the controllable-form realization in the multivariable
case is not unique. Consequently, the adjective "canonical" is oot used.
of the form in (6-112) ca
Column degrees and controllability indices equation
Proof
:V(}.) D'(A)]' form a
Controllability indices are, as shown in Theorem 5-8, invariant under any
similarity transformation and any rearrangement of input vectors by a non
singular matrix. Column degrees are, as discussed in Theorem G-15', intrinsic
properties of C(s). Hence we may use the realization in (6-112) to establish this
~xists a p x 1 vector a. theorem. If we rearrange the input vector u(t) by
o(t) = Dhcu(t)
where Ohc is nonsingular by assumption, then the new B matrix in (6-112)
becomes
to (6-122) yields (6-121) O"-01:0"'00: :O"'OOj
O O O: O O 1: : O .. , O O
'8'=
r
I ' ••• 1
: : I : : I 1: :
• • I • • I l· ••
0'''00:0'''00::0 01
'---v-----' ~ '---v---'
e proof of this theorem. J1.1 J1.1 J1.p
Q.E.D.
where the prime denotes the transpose_ Using this Üand the A in (6-112), we can
readily show that the controllability indices are J1.l> J1.1' ... , J1. p, which are the
; irreducible if and only column degrees of D(s). This establishes the theorem. Q.E.D.
== N(S)0-1(S) is a right
6-] , Observable-Form Realization. Similar to the observable canonical-form
realization discussed in (6-18) for scalar transfer functions, we can develop
observable form realizations for transfer matrices. This realization will be used
and only if deg C(s) == extensively in the next section.
ivalent (Theorem 5-20), Consider a strictly proper q x p rational matrix (;(s). We factor it as
i only ifits dimension is C(s) = 0-1(s)N(s) (6-125)
Theorem 6-2.
controllable canonical where the polynomial matrices O(s) and N(s) are left coprime and D(s) is row
1 (6-112) a multivariable reduced. We take the transpose of C(s):
zation depends on the C'(s) == N'(s)(D- 1(s))' = N'(s)(D'(s))-1 (6-126)
LJ(s))(D(s)U(S))-I, where
e column reduced. then The rOVlS 01 D(s) beco me the columns oÍ [}J'(s). tlence D'(s) 1S C01UIIlll (¿;clUCé>:L
rm realization. Hence, Furthermore, N'(s) and D'(s) are right coprime. Hence an irreducible realization
)O in the multivariable
z=Az +Bv w=Cz
lical" is not used.
of the form in (6-112) can be developed for N'(s)(D'(s))-1. Consequently, the
equation
x=A'x +C'u y=B'x
mn degrees and control
is an irreducible Tealization of D -1(s)N(s), and the equation is said to be in a
multivariable observable form_
af course, an observable form realization can also be obtained directly from
(6-125). Let the row degrees of D(s) be Vi, that iS,b,.¡O(s) = Vi, i = 1, 2, ... , q, and
~ realization bf a strictly let
)Iumn degrees of D(s) in
)0-1(S). (6-127)
"'_' .... __ __._ _ .... _.• ,. __ .. .... __ c __
------------
~
~
~
~
.
"
'
_
n
_
.
_
.
~
~
_
~
_
~
_
~
.
~
.
_
.
286 IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATION
We write C(sI
D(s) = H(s)D h,. + L(s)D¡,. = (H(s) + L(s)D¡,.D,;;.1 )D h ,. (6-128)
which implies
where
SVI O or
_
H(s) =
l O
~
O
l
1 s ... SVI - 1 : O : : O
_ ------O-------:- i --S- - -. ~ ~ - ;;2:'--1 - -: :- - - - - - - - Ó- - - - - - observable-form realiza
L(s) = - - - - - -:- - - - - - - -' - - - - - - -. ~ - - - - - - - -~ '- - - - - - - - - - - - - - - whether D(s) and N(s) ~
.
. . . '. then the realization is e
------O-------;- ----- -'0 --------: :--1- -5- - .- .-. - - ~ vq =- í
(6-129)
*6-7 Polynomial
and N(s) = L(s)N¡c (6-130)
Note that D h ,. is a q x q matrix and is nonsingular, for D(s) is row reduced. The Consider the network si
matrix L(s) is a q x n matrix. DI,. and D¡rD';;' 1 are n x q matrices. Thus a and inductor currents
realization of (6-125) is given by Section 3-4, then a d}
network. There are, 1
O O O ,
x: .x', x matical equations to de~
1 O O x', x , x currents ~1(t) and ~2(t) a
O 1 O x: O x O x
. ,
,
O O 1 x: x, x
-------------- ... --------------,
x:OO"'Ox' x 3
,
x , 1 O O x x 2
, , 6s +1
, or
x(t) = O x , O 1 O x , O x x(t) [ -1
,
, to describe the network.
,
x , O O x , x
--------------1-_-------------.,.
,
x , x ,, :- 0- -O------6 - ~x
, , ,
x , x , : 1 O O x This set of equations can
O x ,, O x , , O 1 O x
, ,
, ,
, ,
,
x , x ,, O O 1 x 2h JI
,----rml'-----.-_ ~
+ Nteu(t) (6-131a)
~r~
y= [O O OC h : O O ... Oc : ... : O O ... O
1, 2 V2 I I .
(6-131b)
wherec¡y¡ is the ith column of D,;;.1 and the (L~~ 1 vm)th column ofthe matrix A In
is the ith column of - DlrD,;;.l. Since (6-131) is a realization of (6-125), we have Figure 6-8 A network.
POLYNOMIAL MATRIX DESCRIPTlON 287
;, AND IDENTIFICATlON
This is a key equation in this realization, and will be used in the next section.
Eq uation (6-131) will be called a multivariable observable-form realization of
(;(s).
Dual to a multivariable controllable-form realization, a multivariable
observable-form realization of (;(s) = D-1(s)N(s) is always observable no matter
whether D(s) and N(s) are left coprime or nol. If D(s) and N(s) are left coprime,
then the realization is controllable as well; otherwise, it is not controllable.
x 1
A
--¿l(S)+
3s 3s s
(1 1
-+-+2s+1 )A~2(S)=0
x
6s2;31s+4JU:~:~J=[~JU(S)
2
x
O x x(t) or [6s _+/
to describe the network. The output y is equal to y(s) = 2S~2(S) or
r-~ (S'1
-----------
)I(s)=[O 2s]l~:(s;j +O'u(s)
O O x
This set of equations can be written in a general form as
O O x
1 O x P(s)~(s) = Q(s)u(s) (6-135a)
y(s) = R(s)~(s) + W(s)u(s) (6-135b)
O 1 x
+ Nrcu(t) «6-131a)
2h II
--r
O .. , O Cqvq]x
(6-131b)
~ 31
~2h
~-
y
In
1 column of the matrix A Figure 6-8 A network.
~ation of(6-125), we have
288 IRREDUCIBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATlON
where P{S), Q(S). R(s). and W{s) are, respectively, m x m, m x p, q x m, and q x p (6-131) such that
polynomial matrices; u{s) is the p x 1 input vector, and y(s) is the q x 1 output
vector. The variable ~(s) is a m x 1 vector and will be called the pseudostate.
In arder to insure a unique solution in (6-135), the square matrix Pis) is assumed lf deg det Pr(s) = n, th¡
to be nonsingular. The set of equations in (6-135) is called the polynomial matrices. Similar to (1
matrix description of a system. In the time domain. this set of equations
becomes and
where P,.~ M(s)P(s) and Q,.(s)~ M(s)Q(s). Note that deg det Pr(s) = deg det
where p =djdt, is a realiz
Pis). In general P,:-l(S)Q,.(s) is not strictly proper. If so, we carry out the divi
~r the transfer matrix G(~
sion:
Gis) in (6-137) is strictIy pI
Qr(s) = P,.(s)Y(s) +Q,.(s) (6-140) E(s) defined in (6-149). I
written explicitly as
so that P;I(S)Q..(s) is strictly proper (Theorem G-12'). The substitution of
(6-140) into (6-139) yields y(t) =O
~(s) = P,:- 1(s:Q,(s)u(s) + Y(s)fi(s) (6-141) From (6-144) and Pr(s)
have
Now P,:- 1(s)Qr(s) is strlctly proper; he~ce the realization procedure discussed
from (6-125) to (6-131) can be applied to find{A, B, Ca} of the feirm shown in
POLYNOMIAL MATRIX DESCRIPTION 289
, AND IDENTlFICATlON
1W(s) = E. The fraction This is notexactly in the state variable form becausethe matrix R(s)Co(sI - A)-1 B
1tify R(s) = Nr(s), P(s) = may not be strictly proper. Ir so, we can apply Corollary G-12 to R(s)C o to
(s)N¡(s) is also a special obtain
¡) = N¡(.s), and W(s) = O.
R(s)C o = X(s)(sI - A) +C (6-147)
:ransfer function matrix where X(s) is a polynomial matrix and C = Rr(A)C o is a q x n constant matrix
37). Note that the com defined as in Corollary G-12. Using (6-147), Equation (6-146) becomes
we discuss a method of
~r first (6-135a). If P(s)
y(s) = C(sI - A)-IBu(s) +(X(s)B + R(s)Y(s) + W(s))u(s) (6-14S)
r matrix M(s) such that in which C(sI - A)- 1 Bis a q x p strictly proper rational matrix and
ultiplication of M(s) on
E{s)~ X(s)B + R{s)Y{s) + W{s) (6-149)
. deg det P,.(s) = deg det where p =d/dt, is a realization of the polynomial matrix description in (6-135)
;0, we carry out the divi or the transfer matrix G{s) in (6-137). Note that W(s) =0 does not imply that
G(s) in (6-137) is strictly proper. Whether or not G(s) is proper is determined by
(6-140) E(s) defined in (6-149). If E(p) = E ó +E¡p +E 2p 2 + .. " then (6-150b) can be
written explicitly as
2'). The substitution of
y(t) = Cx(t) + Eou{t) +E¡ü(t) +E 2ü(t) + ...
(6-141 ) From (6-144) and Pr(s) = M(s)P(s), where M(s) is a unimodular matrix, we
~~ . .
:ion procedure discussed
:::o} of the form shown in det (sI - A) = k det P(s)
290 IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATION
for sorne nonzero constant k. This relation is a consequence of the realization Recall that P,.(s) = M
procedure and is independent of the coprimeness arrrong P(s), R(s), and Q(s). The premultiplicatior
The {A, Ca} in (6-144) is of the multivariable observable form and is always yields
observable. The observability of {A, Ca}, however, does not imply the obser
vabili ty of {A, C} in (6-150). In the following, we shall establish -Ul 1(S) UI2(S)
P(s) M-l(S)L(.
[
{P(s), Q(s) left coprime} if and only if {A, B controllable} (6-151) -0-- - ------=.X(;)
(P(s), R(s) right coprime} ifand only if {A, C observable} (6-152)
L(s) 0J[SI-A
[ -X(s) I q -C
B J=[ Pr(s) Q,,(S)J[C a -Y(S)J (6-153) =L:T!~~)-
E(s) -R(s) W(s) O Ip
where the leftmost and
They are, respectively, (m + q) x (n + q), (n + q) x (n + p), (m + q) x (m + p), and equation, we are ready
(m +p) x (n +p) matrices. Because of the right coprimeness of {P,.(s), L(s)} and From the first two
the left coprimeness of {si -A, Ca}, (6-145) and Problem G-ll imply the exist
-U 11 (S) Uds)
ence of U 11 (s) and Uds) such that [ P(s) M-l(S)L(s)
[
-U l l (S) ~12(S)J[ -CaJ=[I"J (6-154 )
where 8(s) denotes the r
P,.(s) L(s) si A O most matrices in (6-158
~--~-~'
unimodular 1
[ 0m (sI
where the leftmost matrix is a unimodular matrix. Using this matrix, we expand
(6-153) as if and only if
o
P(s)
which implies that
[(si - A)
(6-155)
if and only ir [l
orequivalently, [A, lB} is
This is obtained by augmenting the first three matrices in (6-153) as shown 1,;
Simiiariy, we can establi
(6-155) and then searching the fourth matrix in (6-155) to complete the equality. are right coprime.
Since the left-upper-corner matrix in the leftmost matrix in (6-155) is unimodular,
so is theentire leftmost matrix in (6-155). From (6-154), we have U ds)(sl - A) =
1" - U II(S)C O and Definition 6-2
[
-U 11 (S)
1m
UdS)(SI-A)J=[-Ull(S)
Co 1m
1" - U 11 (S)C O
Ca
J The polynomial matrix
only if P(s) and Q(s) are I
Recall that P,(s) = M(s)P(s) and Q,.(s) = M(s)Q(s), where M(s) is unimodular.
uence of the realization
The premultiplication ol' diag (In> M- 1 (s), I q }, a unimodular matrix, to (6-155)
ng P(s), R(s), and Q(s).
yields
ble form and is always
es not imply the obser
-Uld S) UI2(S) : 0J[lm: O O ]
establish _~!~) ~~l_(~)~J~)_:_~_ -Ó-:-(~i~Á)---B-
[
; control1able} (6-151) O -X(s): Iq O: -C E(s)
C observable} (6-152)
where S(s) denotes the rightmost matrix in (6-157). Since the leftmost and right
: Ip or equivalentl)', [A, B} is controllable if and only if 1"(5) and Q(s) are left coprime
~s in (6-153) as shown in Similarly, ...ve can eSlablish thal (A, C; is observable ir and oníy il' íP(s) and R(s)
to complete the equality. are right coprime.
,in (6-155) is unimodular,
. we have U 12(S)(sl - A) =
Definition 6-2
The polynomial matrix description in (6-135) is said to be irreducible il' and
only il' P(s) and Q(s) are left coprime and P(s) and R(s) are right coprime.
Consider the polynomial matrix description in (6-135) with n = deg det P(s).
1imodular, so is the first .Then an n-dimensional dynamical equation realization ol' (6-135) is irreducible
rix in (6-155) is also uni or, equivalently, control1ableand observable ifimd ónly if (6-135) is irreducible.
I
(6-155) by P(s) and Q(s).
292 IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTIFICATION
This theorem follows directly from Definition 6-2 and the statements in
given a transfer mat
(6-151) and (6-152). To conclude this section, we discuss the situation where
procedures in Section
(6-135) is not irreducible. If PIs) and Q(s) are not left coprime, then there exists
relationships among
a m x m polynomial matrix H(s) with deg det H(s) > 1 such that
questions can stiJl be
PIs) = H(s)P(s) and Q(s) = H(sQ(s) ~ider a transfer matrix
G(s) = N .(05)0- 1(05)
Consequently. at the roots, l., of det H(s) = 0, we have A ' " or
<.: 1 is strictly proper (
pepe),) Q(A)] = pH(A)[P(),) Q(),)] < m G(s) = NI (05)0 1 1(05) + E
betw~en {N,.(s), D,.(s)} (
where p stands for the rank. lf an n-dimensional dynamical equation is de
q uestIOns. we rewri te ((
veloped for {P(s), Q(s), R(s), W(s)} with n = deg det PIs), then this n-dimensional
state equation will not be controllable. lfthis equation is decomposed into the
form in (5-54), then the eigenvalues associated with A. c, the uncontrollable part,
wil! be equal to the roots of det H(s). Hence the roots of det H(s) or, equivalently,
[
where ~ is called the PSI
those A in e with p[P(A) Q(A)] < m wil! be called the uncontrollable mode of
{Pis), Q(s)}. These uncontrollable modes are called the input-decoupling
zeros in Reference S185.
Similarly, if Pis) and R(s) are not right coprime, then the roots of the deter
is called the system l11{{tr
minant of their greatest common right divisor wil! be called the unobservable
modes of {P(s), R(s)}. These roots are called the output-decoupling zeros in
Reference S 185. Hence an irreducible {P(s), Q(s), R(s), W(s)} does not have any
Ir we identify R(s) = e
uncontrollable or unobservable modes.
.identify R(s) = N,.(s), P(s:)
The discussion of the polynomial matrix description in this section is not
meludes (6-159) and (6-1,
complete. We discuss· only its realization problem. For a more detailed
be used to describe any (
discussion, see References S34 and S218. In S218, the description is called the
Consider the system 1
differential operator description.
and the statements in given a transfer matrix or polynomial matrix description, we may use the
lSS the situation where
procedures in Section 6-7 to deve\op a state variable description. Hence the
prime, then there exists relationships among them have been essentia\\y established. Even so, sorne
questions can still be Ilosed regarding these descriptions. For example. con
:;uch that
sider a transfer matrix G(s) which is not necessarily proper. It can be factored as
9(s)=N,.(s)D,:-I(S) or it can be decomposed as G(s)=GI(s)+E(s), where
<}:1 is strictly proper and E(s) is a polynominal matrix and then factored as
G(s) = N ¡(S)D¡I (s) + E(s). The question is then: what is the re\ationship
between {N,.(s), D,.(s)} and {N 1 (s), D 1(s), E(s)} ?To answer this and other related
namical equation is de questions. we re"'Tite (6-161) as
then this n-dimensional
1 is decomposed into the
P(s)
[ -R(s)
Q(s) -Il
~(S)J
W(st -ú(s) =
[ O ]
-y(s)
(6-162 )
the uncontrollable part,
det H(s)or, equiva\ent\y, where ¡; is called the pseudostate, and the matrix
uncontro\\ab\e mode of
S(S)=[ P(s) Q(S)] (6-163 )
:d the input-decoup\ing -R(s) W(s)
en the roots of the deter is ca\\ed the system matrix. 1ts transfer function from u to y is
cal\ed thc unobservab\e G(s)=R(s)P-!(s)Q(s) +W(s) (6-164)
tput-decoupling zeros in
, W(s)} does not have any Ir we identify R(s)=C,P(s)=(sl-A),Q(s)=B, and W(s)=E+E 1 s+ "', or
identify R(s) = N,.(s), P(s) = DAs), Q(s) = 1, and W(s) =0, then the system matrix
ion in this section is not ineludes (6-159) and (6-160) as special cases. Hence the system matrix S(s) can
1. F or a more detai\ed
be used to describe any of the three descriptions Ín (6-159)-(6-161). .
e description is called the
Consider the system matrix S(s) in (6-163). We extend it to
Se(s) =
1
O
O :O
P(s) : Q(s)
J~ [ P e(s) Qe(S)]
[ ---------,----- -Re(s) W(s) (6-165)
O - R(s) , W(s)
1ematica\ descriptions for where I is a unit matrix of any order so that the order of P e(s) is equal to or larger
than deg det P(s). lt is c\ear that
re state-variable equation
'~Let :F'(.{s) = det :2\:;)
and Ge(s) =Rels)P(:-I(S)Q,,(s) +W(s) =R(s)P-I(s)Q(s) +W = G(.s)
Hence the input-ou tpu t behavior of S(s) and that oC Se(s) are ident icaL In fact,
(6-160)
1 (s)N,(s)ú(s) their entire dynamical behaviors are, as will be shown latee equivalent.
Consider two system matrices
Definition 6-3
Hence we have
Two system matrices SI (s) and Sz(s) are said to be strictly system equivalent if
and only if there exist m x In unimodular polynomial matrices Ves) and Ves) and
q x m and m x p polynomial matrices X(s) and Ves) such that which, together with 1,
Theorem 6-8 ~
Two system matrices. which are strictly system equivalent have the same where Ves) is unimodular, a,
transfer matrix and det p¡ (s) = k det P z(s), where k is a nonzero constant.
t(s)=
Proof
We see that ~¡ and ~z areTe":
We multiply (6-167) out toyie1d ~nd (6-174). Hence, ifS 1 (s)a
VP¡V VP¡Y +VQ, ] [ P z
~n~~t u(t), and any set of i~i
[ -(R¡ -XP,)V (XP¡ -R¡)Y +(XQ¡ +W¡) - -R z
lO'tIal
'd .
conditions in S·
z, an
d'
, entIeal and their pseudosta
:E, AND IDENTIFlCATION
STRICT SYSTEM EQUlVALENCE 295
Hence we have
notes strict system eqUlv In order to establish the relationship between the pseudostates ~l and ~2'
we use (6-162) to writeS¡ as
6-3 can be seen from the
ttions (6-153) and (6-155). (6-170)
(6-167)may not besquare
¡¡imodularity property, we
and (6-171 )
J[ -otJ = ['-R
-168) and (6-155) reveals
lization {A, B, e, D} de U
[X
0J[
1 -R
PI
I
Q1
W1
P 2 Q2
2 W2
J[VO 1YJ[ -o; J 1
1 equivalent are eq uivalent which can be written as, by the substitution of (6-170),
ix' their pseudostates are
n~mical-equation realiza
Section 6-7' are zero-input
realization is controllable
[U0J[ -y0J =[-yJ=
X 1
ol [ ]32
-R 2
Q2JrlV~I--:,'YM---'J'
W2 -u
(6-172)
We see that ; 1 and ;2 are related by the invertible transformation pair in (6-173)
and (6-174). Hence, ifS 1(s) and S2(S) are strictIy system equivalent, then for any
input u(t), and any set of initial conditions iriS¡, there exists a unique set of
initial conditions in S2, and vice versa, such that the outputs of SI and S2 are
identieal and their pseudostates are related by (6-173) and (6-174). In short, if
296 ¡RREDUCIBLE REALIZATIONS, STR¡CT SYSTEM EQU¡VALENCE, AND IDENTIF¡CAnON
Si, i = 1,2, are strictly system equivalent, there is no difference in their dynamical
Proof
behaviors.
Consider two system matrices S¡(s) and S2(S). lf they are realized by using If {A, B, C, E} and {A, i
the procedure developed in Section 6-7, then the realizations {A¡, B¡. C¡, E¡} constant matrix P such
have the property these, it is straightforwa:
det (sI - A;) = k¡ det p¡(s) i = 1, 2
lf S¡(s) and S2(S) are strictly system equivalent, then det P¡(s)=k det Pz(s).
Hence the dynamical-equation realizations of S¡(s) and S2(S) have the same
dimension and the same characteristic polynomial. Hence their system matr
Now we assume that
Theorem 6-9 valent. Then from (6-16
Coprimeness, controllability, and observability are invariant under the trans U(S) O][SI
formation of strict system equivalence. [ X(s) I -
q
Since U(s) and V(s) are unimodular, we have, for every s in ic,
which implies P(s)(sI - A
~roper rational matrix, s'
P(s) is a constant matrix.
and to yield
where p denotes the rank in the field of complex numbers. Hence P ¡ (s) and
Q ¡ (s) are left coprime if and only if P 2(S) and Q z(s) are lei'L coprirne. ¿ ;]22\;;) = v¡!hich lmplies if' = jp o and .
sI - A and Q2(S) =E, then P ¡(s) and Q¡(s) are left coprime if and only if {A, E}
is controllable [see Equation (6-151)]. The rest ofthe theorem can be similarly P(sI -A)
proved. Q.E.D. Now we shalI show tl
unimodular, U- ¡(s) is also
In Section 4-3, we introduced the concept of equivalent dynamical equa
l
tions. Now we shall show that strict system equivalence is a generalization of
this concept. for some polynomial matri
plication of(6~179) and (6-1
Theorem 6-10 U(s)U- 1 (s) = (sI'..:.. A)U(s)(sJ
c,
Two dynamical equations {A, B, C, E} and {A,R, E} are equivalent ifand only which becomes, because of I .
if their system matrices are strictly system equivalent.
I-Pp¡=(sI-;
,._..-._-""-- ...--._,' , .. .,.--" .. -_ ... , .•.. ~ .. _'-" ", ......
for sorne unimodular U(s) and Ves). Since {A, B, C, E} and {A, B, E} have e,
the same transfer matrix (Theorem 6-8), we have E =:E. Equation (6-177)
implies
) zero matrices, the first
U(s)(sl - A) = (sI - A)V(s) (6-178)
where Pis a constant matrix. The substitutíon of (6cI79) into (6-178) yields
(6-176) P(sl A) = (sI - A)[V(s) - U(s)(sl - A)]
~ (sI - A)P(s) (6-180)
.¡ s in e, which implies P(s)(sl-A)-l=(sl-A)-lp. Since (sl-A)-lP is a strictly
s)] proper rational matrix, so must be P(s)(sl - A)-l. This is possible only if
pes) is a constant matrix. Hence we may replace pes) in (6-180) by P(s) = Po
to yield
sP-PA=sPo-AP o
nbers. Hence P 1(s) and
which implies JP' = Po and PA = APo. Hence we have
: left coprime. lf P 2(S) =
,rime if and only if {A, B} P(sl - A) = (sI - A)P and PA=AP (6-181)
theorem can be similarly
Q.E.D. Now we shall show that the P in (6-179) is nonsingular. Since U(s) is
unimodular, U- 1 (s) is also unimodular and can be expressed as
or D 2(s) = D 1(s)T(s) an
(sl-AT 1(I-PP¡)=Ü(s)(sl-A)U 1(s)+Ü(S)P 1 +PU 1(s)
[~
(6-183)
Since its left-hand side is a strictly proper rational matrix, whereas its right-hand
side is a polynomial matrix, (6-183) holds only if both sides are identically which implies the st
equal to zero. Hence we have I-PP 1 =0 or PP 1 =1. Consequently, P is Now if G(s) is no
nonsingular and is qualified as an equivalence transformation. (;(s) = N(s:
To complete the proof ofthis theorem, we must show 8 = PE and C = CP- 1
From (6-177), we have where Ñ(S)D-1(S)is ~
of
U(s)B =(sl -A)Y(s) +8 D,
which becomes, by using (6-179), [ -N,
System matrices which are strictly system equivalent have the same transfer where the leftmost m
matrix. System matrices which have the same transfer matrix, however, are not D/N,. = NID,., we forn
U U 12 : O] ~I__
necessarily strictly system equivalent. Example 2 of Section 4-3 is an example
of this statement. A different example will be G(s) = N(s)D -l(S) = N(s)D- 1(s),
11
where N(s) and D(s) are right coprime, but N(s) and D(s) are not right coprime. [-1DI 0:1
-
NI 'O O'
- - -
O:
- - - - - _1- _ I
Clearly the two system matrices {D(s), 1, N(s), O} and {D(s), 1, N(s), O} have the
same transfer matrix. However, because of det D(s) 1= det D(s), they are not The second and thirc
strictly system equivalent (Theorem 6-8). matrices of {D,., 1, N r
Although system matrices {p¡(s), Q¡(s), Ri(s), W¡(s)}, which have the same unimodular because i
transfer matrix, are generally not strictly system equivalent, they become (6-186). We show in
strictly system equivalent if the system matrices are irreducible: that is, {F';(s), unimodular. From (6
Q¡(s)} are left coprime and {P¡(s), Ri(s)} are right coprime. We establish first a U 12 D
special case.
N,.
Since the two matrice
Theorem 6-11 the first matrix in (6- j
matrix in (6-187). (
All coprime fractions of G(s) are strictly system equivalent where G(s) is a
rational matrix, not necessarily proper. {D,., 1, N,., O} and {D h 1
proof of this theorem.
Proof
.7The proor orTheorem G-1
Consider the two right coprime fractionsG(s)=N1(S)Dl1(s)=N2(S)D21(s). to improper G(s).
Theorem G-13 implies the existence of a unimodular matrix T(s) such that 8Th!: procedure is similar to
where N(S)O-l(S) is strictly proper and coprime; N(s) = N(s) + E(s)O(s). Because
of
(6-186). We show in the fol1owing that the rightmost matrix in (6-187) is also
luivalent, they become
'The proorof Theorem G-13 does nol use lhe properness of (;(5); hence lhe lheorem: is applicable
(s)D1l(s) = Nz(s)Dzl(s).
lo improper (;(5). . .
1
__ .• ~'~_J~'._" .. ,". _ _ ','._'
.. ~._-"---~-
Theorem 6-12
polynomial matrix 1
AH irreducible polynomial matrix descriptions {PieS), Q¡(s), R¡(s), W¡(s)} which These results are di]
have the same transfer matrix are sti-ictly system equivalent. out the foHowing tra
Proof
Every irreducible {Pi, Q¡, R¡, W;} has an irreducible realization {A, B, C, E(p)}
with dim A = deg det PieS). AH irreducible {A, B, C, E(p)} which have the same Transfer matrix
transfer matrix are equivalent (Theorem 5-20) and, consequently, strictly
system equivalent (Theorem 6-10). Hence by the transitivity property, we Dynamical equation
conclude that aH irreducible {Pi Q¡ R¡ W;} of the same transfer matrix
are strictly system equivalent. Q.E. D.
Polynomial matrix
AH irreducible dynamical equations, all coprime fractions and aH irreducible description
polynomial matrix descriptions which have the same transfer matrix are
strictly system equivalent and consequently have, following Theorems 6-2 and System matrix
6-8 and Definition 6-1', the following properties:
~(G(s)) ~ det(sI - A)~ det D(s) ~ det pes) (6-189)
where ~(G(s)) is the characteristic polynomial of C(s) and denotes the Hence the discussion
equality of two polynomials modulo a nonzero constánt factor. Conversely, if time case. Note that
all descriptions have the same transfer matrix and satisfy (6-189), then they the discrete-time case
must be all irreducible. Under the irreducibility assumption, any one of the equivalently, s- I is re]
descriptions can be used, without loss of any essential information, to study There is, however
and design a system. special discussion. e
To conclude this section, we mention that system matrices which are strictly time system expanded
system equivalent to
pes) Q(s) ]
Ses) = [ -R(s) W(s) As in the continuous-l
parameters of g(z). In
where P, Q, R, and W are, respectively, m x m, m x p, q x m, and q x p polynomial
be computed from di g
matrices, can be generated by the foHowing elementary operations:
response of the system
1. Multiplication of any of the first m rows or columns by a nonzero constant. of an impulse as an in
2. Interchange of any two of the first m rows or columns. of g(t) are again impra(
3. Addition of the multiple of any of the first m rows (coiumns) by a poly time case are not real
nomial to any of the m +q rows (the m + p columns). entirely different. If v
initiaHy relaxed linear
These operations can be readily derived from the unimodular matrices
data at the output term
Ves) O] [ves) y(S)] 2, . . .. Hence the real
[ X(s) I q ' O Ip case can be considered.
mathematical descripti
used in the definition of strict system equivalence. ment at the input and (
corrupted by noises.
requires the concepts o
*6-9 Identification of Discrete,.Time Systems
from Noise-Free Data
In the previous sections, we introduced various realization methods for con
~-----
s) (6-189)
and denotes the Hence the discussion of these problems will not be repeated for the discrete
actor. Conversely, if time case. Note that a1l block diagrams in this chapter are also applicable to
¡fy (6-189), then they the discrete-time case if every integrator is replaced by a unit-delay element or,
ption, any one of the equivalently, S-1 is replaced by Z-I.
.nformation, to study There is, however, one problem in the discrete-time case which deserves
special discussion. Consider a sampled transfer function, g(z), of a discrete
~ices which are strictly time system expanded as
g(z) = h(O) +h(l)z- I +h(2)z- 2 + ...
As in the continuous-time case, we sha1l ca1l {hU), i = 0, 1, 2, ... } the Markov
parameters of g(z). In the continuous-time case, the Markov parameters must
,and q x p polynomial be computed from dig(t)/dt i , i = 0, 1, 2, ... , at t = 0, where g(t) is the impulse
.perations: response of the system or the inverse Laplace transform of g(s). The generating
of an impulse as an input is not possible in practice; repetitive differentiations
>y a nonzero constant.
of g(t) are again impractical. Hence the Markov parameters in the continuous
;.
time case are not really available. 9 In the d.iscrete-time case. the situation 1S
(columns) by a poly-
entirely different. If we apply the input {u(O) = 1, u(i) = 0, i = 1, 2, ... } to an
initia1ly relaxed linear time-invariant discrete-time system, then the measured
modular matrices data at the output terminal are the Markov parameters, that is, y(i) = h(i), i = 0,1,
2, . . .. Hence the realization from the Markov parameters in the discrete-time
case can be considered as an identification problem-a problem of determining a
mathematical description of a system from the data obtained by direct measure
ment at the input and output terminals. In actual measurement, all data will be
I corrupted by noises. A study of the identification problem. with noisy data
requires the concepts of probability and statistics and is outside the scope of this
:ems
I
~
ltion methods for con 9Consequently methods are developed lo find realizations by lIsing the moments defined by
We also introduced M k = S~ tkg(t) dt./< =0.1,2, . (see References 528 and 5146). .
l
·., ··--'·····c·······_·--·· ' _.'.' =
'=.'..c='•.""'.,,,,'===.=--=.. =.._='=. -- _.
text. Hence we assume in this text that all data are free of noise and call the In the following, 'i
problem the deterministic identificatían. an arbitrary input-~Ue
Consider a single-variable linear time-invariant discrete-time system with introduced. The prot
transfer function 9(Z). Ir we apply the impulse sequence {u(O) = 1, u(i) = 0, Consider a linear
í = 1, 2, 3, ... }, then the zero-state response yields the Markov parameters transfer matrix (;(z).
h(i)=y(i), i=O, 1,2, .... Clearly the z-transform of {h(i)} yields the transfer
function 9(Z). However, this approach requires the use of an infinite number of
{hU)}. Ir we form a Hankel matrix from {hU)}, and if the system is known to where D(z) and N(z)
have a degree bounded by N,1 o then we need only 2N + 1 of the hU), i = 0, 1, ... , Let the highest degree
2N. From the Hankel matrix, we can readily obtain a dynamical-equation expressed as
description of the system as shown in (6-35) to (6-38). Ir the transfer function
description is desired, we have
and
f3 oza +(J 1 za -1 + ... + f3 a
gA()
Z = where Di> Ni' i = 0, 1, .
. za +(XI Za 1 + . . . + (Xa
apply the p x 1 input SI
where (Xi are obtained from the Hankel matrix as in (6-36) and (Ji can be com
puted from
°
l
h(O) to the initially relaxed ~
h(1) h(O)
~1(<T) :h(<T-l)
The substitution of I
This matrix equation is just the set of equations in (6-29) arranged in matrix
formo Consequently, the transfer function description of a system can be ob
tained from a finite number of Markov parameters. This assertion also applies Equating the coefficien
to the multivariable case.
In order to obtain Markov parameters, the system must be initially relaxed.
In a multivariable system with p inputs and q outputs, we apply an impulse
sequence to the first input terminal and no input to all other input terminals,
then the responses at the outputs yield {h k1 (i), k = 1, 2, ... , q; i =0,1,2, ...}.
After the system is at rest again [in theory, after an infinite time; in practice,
after h(i) is practically zero or becomes almost periodic 11 ], we then repeat the
process for the second input terminal and so forth. Hence the measurement of
Markov parameters is possible only ií the system. is at OUí disposal. Ií a system
is in continuous operation, its Markov parameters cannot be measured.
where M = [-No Do
10lfno bOllnd 01 the degree of a system is available. then it is theoretically impossible to identify the matrix Si -Y, 00) has (
system. For example, ifwe have [h(O) =0. h(i) = 1, i = 1,2..... 200) and ifthe degree ofthe system columns. This equatio!
is bOllnded by 10, then we have g(z) = I/(z -1). However the system I/(z -1) + l/z 1oOO may also
generate the given seqllence. Hence if no bOllnd is available. there is no way to identify a system
Hence, given an arbitra
from a finite seqllence of {hU), i =0,1,2, ... , N}. The problem of finding a transfer fllnction to to the search of M to ro
match a finite sequence of Markov parameters is caBed the partial realizatioll problem (see Refe rows of nontrivial solu
rences 68, S 126 and S239). Sv( -Y, (0). Since it is,
II ¡·f ~ ·system is not BI BO stable (see Chapter 8), then h(i) will approach infinity or l'emain oscil
as possible, we search,
Jatory (including approach a nonzero constant). In the former case. the system will satllrate Ol'
bum out and th~ linear model is no longer applicable. In the latter case, the system can be brollght Sv( - Y, (0). Hence the
to rest by resetting. In thebry, thereaÚzati~n ~~ identification is applicable no maller the system dependent rows of Sv('
is stable or not so long as the data are available. similar to the coprime I
--------~---~---_._._. _ ..~-~~._._ ...-._.. __ ...... "
of noise and call the In the following, we shall discuss a method of identifying a system from
an arbitrary input-output pair. The concept of persistently exciting will be
~te-time system with introduced. The problem of nonzero initiai conditions will also be discussed.
ce {u(O) = 1, u(i)=O, Consider a linear time-invariant discrete-time system with q x p proper
Markov parameters transfer matrix G(z). Let G(z) be factored as
')} yields the transfer G(z) = D-l(z)N(z) (6-190)
an infinite number of
where D(z) and N(z) are, respectively, q x q and q x p polynomial matrices.
: system is known to
Let the highest degree of all entires of D(z) be v. Then D(z) and N(z) can be
)f the hU), i = O, 1, .... ,
expressed as
dynamical-equation
the transfer function D(z) = Do + D 1z + ... + Dvzv (6-191)
and N(z) = No +N 1z + '" +Nvzv (6-192 )
where Di, Ni' i =0, 1, ... , vare q x q and q x p real constant matrices. If we
apply the p x 1 input sequence
;) and f3i can be como.
u(z) = u(O) + U(l)Z-1 + u(2)z- 2 + . . . (6-193)
to the initially relaxed system, then the output is a q x 1 seq uence given by
y(z) = G(z)u(z) = y(O) + y(l)z -1 +y(2)z- 2 + ... (6-194)
.1..
304 IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFlCATION
IDENTlFi
.sal and to have proper form an space of an infinite dimension, almost aH input sequences are per
hen we have y(k) =0 for sistently exciting. It means that if an input is generated randomly, the prob
ability for the input to be persistently exciting is almost lo
Consider So(k, 00).12 We use, for example, the row searching algorithm to
(6-197 ) search the linearly dependent rows of So(k, 00) in order from top to bottom.
(6-198) Because ú is persistently exciting by assumption, aH dependent rows will appear
in y(k, 00). Let Yo be the number of dependent rows in y(k, 00). Clearly Yo'::;; q,
and q - Yo is the number of linearly independent rows in y(k, 00). Next, we
apply the row-searching algorithm to SI(k, 00). Let y[ be the number oflinearly
dependent rows in y(k + 1, 00). Let Yi(k, 00) be the ith row of y(k, 00). If
Yi(k, 00) is linearly dependent in So(k, 00), then y¡(k+ 1,00) will also be linearly
dependent in SI(k, 00). This foHows from the fact that ü(k, 00) =
,2 ... (6-199)
[ u(k) : ü(k + 1, oo)J and y(k, 00) = [y(k): y(k + 1, 00)]. Because of this property,
we have Yl '2:Yo. We continue this process until Yv =q as shown.
ü(k, 00)
~~k.! .ex:} _____ } Yo [no. of dependent rows in y(k, 00)]
nsists of p rows of ü and ü(k +1, 00)
:(l-k+1). WecaHthe y(k + 1, 00)
-----------
} Y[
l y Y rowS. Sv(k, 00) = (6-200)
.s much more than neces
ü(k +v -1,00)
y(k+v-1,00)
}Y v -l
(6-205 )
Because of the assumption VI = 3, V2 = 1, and V 3 = 3, we have df¡ = 1, d~2 = 1,
(6-203) d~3 = 1, and aH elements in (6-205) on the right-hand sides of dfl' dL, and d~3
are zeros. In addition, the columns associated with dit 1, d~;Z, d%;l are zero
for 1= 1,2, 3, .... Hence the Di reduce to 14
3 ' d 111 O dI1 3 '' (11
do11 do12 dIo' O d 21 3 '' :~)
'1"' O O]
D3J=[d~1 d~2 d~3:
[2
[Do DI D 2 dLQ)O : O O O : O O O
q l ' d2
do31 do32 do' , 31 O d 33'' O O'1-'
1 2
~
33 ' d 31 O d 33
1 I I
: ••'
and
d~3 +dt3 Z +d"f3 Z2 J
[
.gonal, as discuss ed in d~1 +dLz+dLz2+z3 d~2
wehavevl =2,v 2 =0, D(z)= d~1 +d~lZ d~2 +z d~3 (6-206)
,index 2. d~1 +d11Z+d~IZ2 d~2 d~3 +d13Z +d~3Z2 +z3
)endent on its previous
example, the y2 in the This D(z) is clearly row reduced. It is also column reduced. (This property,
1; the Y2 in the second
however, is not needed here.) Therefore we have
~d block-row of (6-203) deg det D(z) = VI +v 2 + v3 = n (6-207)
rimary dependent rowS Because aH elements bn the right-hand side of d~2' di l' and d%3 in (6-205) are
nary dependent y¡ row zeros, the row degree of N(z) is at most equal to the corresponding row degree of
D(z). Hence D- 1 (z)N(z) is a proper transfer matrix. We claim that N(z) and
D(z) are left coprime. Suppose not, then there exists a q x q polynomial matrix
R(z) with deg det R(z) > Osuch that
sistently exciting input N(z) = R(z)N(z) D(z) = R(z)D(z) (6-208)
lependent roWS in order This implies deg det D(z) > deg det D(z). However, this is not possible because
r triangular matrix with the n = LVi computed in the algorithm is unique and smaHest possible. There
y-rows in the last block fore we conclude that N(z) and D(z) are left coprime, and the degree of the trans
.: -:N, D J be the q fer matrix is equal to deg det D(z) = Vi. L¡
t rows of S,. Then the
That D(z) is in the polynomial echelon form foHows from the definition given
N(z), where in Appendix G. This completes the proof of this theorem. Q.E. D.
(6-204a)
(6-204b)
An example will be given, after the diseussion of nonzero initial conditions,
to illustrate the identification procedure. We note th.at once the transfer
=deg det D(z) and D(z) is matrix, in the coprime fractional fom, of a system is identified, a dynamical
olynomial echelon form equation description can be readily obtained by using the procedure discussed
in Section 6-6.
SAk, (0). In practice, we use the matrix Sv(k, l) for a finite l. Clearly, 1must be has a full row rank. '
larger than (v + 1)(P + q) to ensure that there are more columns than rows. In full row rank. The CO]
actual computation, once y¡(k + Vi, l) is Iinearly dependent in SV¡(k, 1), then y¡ is the Hankel matrix o
may be deleted in forming SV¡+ j, j = 1, 2, .... tions for all rows of U(
If u is persistently exciting. then the Ni and Di computed by using the row from the results in Seco
searching algorithm have the properties that D(z) is row reduced and, con
sequently, nonsingular and D - 1 (z)N(z) is proper. Ifü is not persistently exciting,
then the computed D-I(z)N(z) may not be proper. For example, consider and
g(.:) = 1/(.:: + 1). If we apply u(z) = 1 +.:: - 1, then the output )";(z) is equal to z - l.
We form where u¡(z) is the ¡th COi
tions state that the inp
ifwe useS v (l, 00) in the i
Ü(O, (0)] [1 1 O O O ...]
Hankel matrix of u(z)/"
SI = ht~~-:~ = ~--Ó--Ó-~--6- -:-:~ (6-209) to be linearly independ¡
[
y( 1, (0) 1 O O O O ...
Clearly ü(l, 00) is linearly dependent on its previous rows: hence u(z) is not and
persistently exciting. If we solve (6-209) by using the first linearly dependent
row ofS¡, then the solution is The conditions in (6-21.
ample, one of u¡(z) may
[-1 1: 1 O]SI =0 (6-213). See also Proble
A 1 -z A
case, we may choose an
and y(z) = -1- u(z) (6-210) If the chosen or given i
appear as linear depenc
We see that this g(z) is not proper and is erroneous. Fortunately, this problem automatically whether o
can be automatically detected in the search of linearly independent rows of Si'
1t is possible to obtain a different but equivalent equation of (6-196) by Nonzero initial cond
grouping all u rows at the upper half of S. and all y rows at the lower half. For ofsystems which are not
this example, the equation becomes and observable discrete-I
y(l,oolJ
and the solution is also equal to g(z)=(l-z)/l. Since all u rows of SI in y(z) = C(z]
(6-211) are linearly independent by themselves, there is no way to check from
which can be written as
SI that g(z) is erroneous. Hence the persistent exciting of {u(n)} cannot be
defined solely on {u(n)}.
y(z)
A necessary condition for an input sequence to be persistently exciting is
that the matrix where R(z) is a q x 1 poly
U(k' (0)
_ u(k+l,oo)
J to the corresponding ro",
We rewrite(6~217) as
ú(k, (0) - . (6-212)
[
u(k +v, (0) [ -R¡
lDENTlFICATION Of DISCRETE-T1ME SYSTEMS FROM NOISE-FREE DATA 309
, AND lDENTlFICATlON
el. Clearly, lmust be has a ful1 row rank. We give sorne sufficient condition for U(k, 00) to have a
olumns than rows. In full row rank. The condition depends on the value of k. If k = 1, then U(l, 00)
ent in Sv¡(k, 1), then J¡ is the Hankel matrix of u(z). Hence, if u(z) is a rational vector, then the condi
tions for all rows of U(l, 00) to be linearly independent are, as can be deduced
'uted by using the row from the results in Section 6-4,
ow reduced and, con
bu(z)::?:p(v +1) (6-213a)
ot persistently exciting,
and bí/¡(z)::?: \' + 1 i = 1,2, ... , p (6-213b)
For example, consider
,ut y(:) is equal to Z-I. where u¡(z) is the ith component of u(z) and b denotes the degree. These condi
tions state that the input signals must be more complicated than the system
ifwe useS v(l, 00) in the identification. If k = - v, then the matrix U( - v, 00) is the
Hankel matrix of u(z)/zv+ l. Therefore the conditions for al1 rows of (6-212)
.] (6-209) to be linearly independent are
equation of (6-196) by
Nonzero ¡nitial conditions. In this subsection we study the identification
s at the lower half. For
of systems which are not necessarily initia\ly relaxed. Consider the control1able
and observable discrete-time equation
This equation is similar to (6-195) and implies hence the Ni, Di compl
o
O 1 O with the Ni, Di computt
puted by solving success
O ... O u(O) u(l)
In other words the solut
O ... O y(O) y(l)
-----------------
...
--------- MS v( - v, oc!) = O. Hent
O 1 O O and the ones computed J
O ... u(O) u(l) u(2)
[ -Ro -No D o·: ... :. -R v -N v D,.] lar polynomial matrix (1
O .. . y(O) y(1) y(2) matrices are the same.
---------------- ----------
---------------- ---------- Example 1
1 O O O
u(O) u(v -1) u(v) u(v + 1) ... Consider a system with t
y(O) y(v -1) y(v) y(v + 1) .. ,
= MSi - v, (0) = O (6-219)
hence the Ni, Di computed from (6-219) will also satisfy (6-196). Conversely,
with the Ni' Di computed from (6-196), a set of R i, i=O, 1, ... , v, can be com
1
UrO) u(1)
°
puted by solving successively the first v + 1 column equations ofMS v( -v, (0) =0.
In other words the solution M of MSJO, (0) = O is also a part of the solution of
y(O) ?'~ 1) - MS)-v, 00)=0. Bence the D(z) and N(z) computed from MSv(O, 00)=0,
------
O and the ones computed from MS v( - v, (0) = O will differ at most by a nonsingu
O
u(1) u(2)
lar polynomial matrix (dual of Theorem G-13). Consequently, their transfer
matrices are the same. This completes the proof of the theorem. Q.E.D.
y( 1) y(2)
----------
______ 6 _
Example 1
o o
Consider a system with transfer matrix
-1) u(v) u(v +1) .. ,
y( v + 1) ...
z~ 1]
- 1) y( v)
(-v, (0)=0 (6-219)
(6-220)
and Sv( - v, (0) denotes 2z
~e the input [1 u'(z)]', z-1
lly exciting. Thi~ input
we cannot use Sv(k, 1) An irreducible realization can be found as
ter may meet the condi
+ 1) =[~
ith its initial conditions
u¡(k) O O -1 1 O O O 1 -1 O O O
'1) and y(n) are no longer
U1(k) O O 1 O O O 1 O O O O 1
e system. Since (6-196)
.r I (k)
:lOt have the information
.\·,(/d
0.5
,
3.5 0.5 3.5
()
3 3 3 4 5
-¡
2.5 55
7
25
1i
"
')
(¡ (: n u 1';,':
O -1 O O 1
-3° -4
° O
-1 °
_1
O 1
o O -1 (1) O O O 1 -1 O O x (ü O O
O O 1 O O (~(~
O O O O O x O 0-1
O O ¡
0.5 3.5 0.5 3.5 Gt: 3 3 4 5 2.5 5.5 x
:5 ; 3 6 6 6 6 8 9 7 9 7 x 0.5 3.5 O.
** Ó- - --.:: ¡ -----f - -- -0- - - - - 0- ---6.. -- -1 - f:; --0- ----ó-----ó
t~-
x * O 5 3 6
* .T·"--ü- ---ó----6
O :}) O O O 1 O O O O 1 x ,_J _,
3.5 0.5 3.5 3 3 3 4 5 2.5 5.5 2.5 O O 0:-11 1
3 6 6 6 6 8 9 7 9 7 11 O O O '-'
1
0.5 3.5 0.5 3.:
5 3 6 6
where the leftmost matrix is the F defined in (A-9). Its first column is chosen to
make the fourth column, except the first element, of S2(0, 11) a zero column. The last two rows of E
Its second column will make the seventh column, except the first two elements,
of K IS 2(0, 11) a zero column. Note that the location ofthe pivot element in
each row is encircled. Since the computation is carried out by hand, the pivot
element is chosen for convenience in computation. For this problem, we have
ro =0, rl =2 =q, and VI =V 2 = 1. The last two rows of K, which correspond
=[
to the linearly dependent rows ofSz(O, 11), can be readily computed by using the Hence we have
formula in (F-11) as
-1 :-1
2
1 --34 ',,
_.1 •
2 ,
O
-1
O
-2
1
O ~J which'are the same as
For this example, we c;
Hence we have
N(Z)=L!} ;}+1]
D(z)= [ Z_¡2
+1
:~}J (6-222) Hence the result is corr
o o x
o :}) O O O O O O O O O O x
o o x
O O O -1 eD
O O O 1 -1 O O x
O O O 1 O O O (f: O O O O x
2.5 5.5 x
O 0.5 3.5 0.5 3.5 el: 3 3 4 5 2.5 5.5 x
9 7 x
----Ó-----Ó x
** o
-_ ..
5 3 6 6 6 6 8 9 7 (9.~ 7
_-------------------------------------------------
x
O 1 x
.'Í "
' ... "
O o
-,
o o o o o O o o o x
O O (- 11 1 O O O 1 1 O O O x
5 5.5 2.5 o
O
..
o -'1 :}l~ o o 1 O O O O 1 x
7 11 O
0.5 3.5 0.5 3.5 3 3 3 4 5 2.5 5.5 2.5 O
5 3 6 6 6 6 8 9 7 9 7 11 O
rst column is chosen to
O :_l O O:,
~J
l l 3 t 1 I
1
out by hand, the pivot
= I 4 2, 2 -4 : -Z ,
. this problem, we have
[ O'
, l2 -1 : -1 -1':-5:-1 -2: O
f K, which correspond
Hence we have
computed by using the
o O 1
N(z) =[ ¡
z_l
2
-1 -2 O which are the same as (6-222) obtained by disregarding the initial conditions.
For this example, we can also verify that
n-1(z)R(z) = C(zI - A)-l zx(O)
1
2 (6-222)
Hence the result is correct.
•
6-10 Concluding Remarks
d. In this case we must
rithm yields In this chapter, we discussed three approaches to find irreducible realizations
for proper rational matrices. The first approach is to find a reducible one and
then use the procedure in Section 5-8 to reduce it to an irreducible one. The
second approach computes the Markov parameters by using the recursive
formula in (6-29) and (6-30) and form the Hankel matrix. We can then use the
singular value decomposition or the row searching algorithm to find irreducible
realizations. In the last approach, irreducible realizations are obtained from
coprime fractions of strictly proper rational matrices. The majar effort in
this approach is the computation of coprime fractions and the procedure in
Appendix G can be employed. Allhough we have intro.duced a number of
irreducible realization methods;the treatment is not exhausÜve. For example,
the Jordan-form realization IS not discussed for the general case.' The interested
reader is referred to References S38, S59, and S145. For other methods, see,
1 for example, References, S62, S63, and S191.
,,,., .-y ••• _ - - _ .•. ,. " •• ~._,., • , •. '._.,_.".",~ • . • ."_. _ , .• ~_ . ' • • • . • ~ .'_' ' ." ~ " 0 ' ' ' _ _ "... • ~ •• ~ •• _ • •
.............-
.~.~~.' -~
.
- - - - ~ - ~_------~--'--~--~----,------~--_
Now we compare briefly the numerical efficiencies of these three approaches. introduce errors due te
Given a q x p strictly proper rational matrix. Let a¡(s) and flj(s) be respectively will not be discussed.
the (least)common denominators ofthe ith row and jth column of G(s). Because In this chapter, we
the computation ofthe least common denominator is not a simple task, we may established its relation
simply multiply all denominators of the ith row and the jth column of G(s) We showed that if the 1
to yield a¡{s) and flis). For simplicity, we assume p = q and a = deg a¡(s) = deg irreducible, then they a
!3j(s), for all i,j. Then we can find a controllable but not necessarily observable them can be used in the
(or vice versa) realization of dimension ap. To reduce this reducible realiza tion.
tion, we must append B and e to A; hence the matrix under operation is roughly The degree of a prc
of order (a + l)p. In the Hankel matrix, the order of the Hankel matrix is degree can be compute
(a + l)p x ap. In the coprime fraction method, if we use the method discussed
in Appendix G to find a coprime fraction, we must form a generalized resultant. 1. Compute the least c
we have deg G(s) = (
Each block row has (p + q) = 2p rows, and we need roughly (a + 1) block rows.
2. Compute the Mark,
Hence the generalized resultant has 2(a + l)p number of rows. Thus we con
shown in (6-80). n
elude that the matrices used in the first two approaches are roughly of the same
order; the matrix used in the third approach is about twice the sizes of those computed by using t
used in the first two approaches. 3. Find a right fraction
The operation in the first approach requires the triangularization of a matrix deg G(s) = deg det D
by similarity transformations (both column and row operations). In the second coefficient matrices e
and third approaches, we require the triangularization of a matrix by either row We then search the liJ
or column operations. Hence the first approach requires twice the computa Then we have
tions of the second and third approaches. Thus we conclude that the Hankel deg G(s) = too
method is probably the most efficient. For a more detailed comparison, see
. Reference S202. We note that all three approaches can.be implementeqby using Similar results can b(
numerically stable methods. 4. Consider the realizat
It is of interest to compare the singular value decomposition and the row is controllable but nc
searching method discussed in Section 6-5. They are implemented on UNIVAC QIO =0, Qlj = R,
1110 at Stony Brook in Reference S202. The row searching algorithm is carried QiO =0, Qij =Q(i-lH.
out by applying Householder transformations and the gaussian elimination
with partial pivoting on the columns of Hankel matrices and then computing Using Qij, the observ
the coefficients of combinations by back substitution. The scaling problem is
included in the programo The examples in Reference S202 do not show any
substantial difference between Householder transformations and the gaussian
eiimination with partiai pivoting. The singuíar vaiue decomposition is most
reliable. However, its reliability over the row searching method is not over
whelming in the examples in Reference S202. This seems to be consistent with Then we have deg (;(5
the remarks in Reference S82, p. 11.23. 5. Find a realization of
The identification of linear time-invariant discrete-time systems is studied. Section 5-8, to an irred
The identification is carried out from a set of arbitrary input-output pair so
long as the input sequence is persistently exciting. Whether the system is These methods can be
initially relaxed or not, the transfer functíon matrix in coprime fractional parison of these method~
form can be obtained by solving MSv(O, ex)) = O. The realization problem can stability seems to be una\
be considered as a special case of the identification problem in which the system To conclude this chal
.. is to be identified from a particular input-output pair, the impulse response or realization in Equation (f
the Markov parameters. The identification method introduced is applicable in algebraic system theor
to continuous-time systeins ir they are first discretized. This, however, will q x p strictly proper ratic
AND IDENTIFICATION
CONCLUDING REMARKS 3]5
hese three approaches. introduce errors due to discretization. This is ou tside the scope of this text and
nd f3 J.(s) be respectively will not be discussed.
A
,lumnofG(s). Because In this chapter, we also introduced the polynomial matrix description and
t a simple task, we ~ay established its relationships with dynaroical equations and transfer matrices.
he jth column of G(s) We showed that if the three descriptions have the same transfer matrix and are
and ex = deg o:;(s) = deg irreducible, then they are aH strictly system equivalent. In thiscase, any one of
necessarily observable them can be used in the analysis and design without loss of any essential informa
this reducible realiza tion.
ler operation is roughly The degree of a proper rational matrix is introduced in this chapter. The
, the Hankel matrix is degree can be computed by using any of the following methods:
;: the method discussed 1. Compute the least common denominator, L\(s), of all minors of (;(s). Then
a generalized resultant. we have deg (;(s) = deg L\(s).
ghly (o: + 1) block rows. 2. Compute the Markov parameters of (;(s) and form the Hankel matrix T
)f rows. Thus we con shown in (6-80). Then we have deg (;(s) = rank T. The rank of T can be
are roughly of the same computed by using the singular value decomposition.
twice the sizes of those 3. Find a right fraction G(s) = Ñ(s)D -1(S). Ir the fraction is right coprime, then
deg G(s) = deg det D(s). Ir the fraction is not right coprime, we use the
gularization of a matrix coefficient matrices of D(s) and N(s) to form the matrix Sk shown in (G-67).
:rations). In the second We then search the linear independent rows ofS k in order from top to bottom.
•f a matrix by either row Then we have
ires twice the computa
Illclude that the Hankel deg G(s) = total number of linear independent N roWS in Soo.
etaHed comparison, see Similar results can be stated for a left fraction of G(s).
)e implemented by using 4. Consider the realizatíon of (;(s) shown in Equation (6-71). The realization
is controllable but not necessarily observable. Define
mposition and the row
plemented on UNIVAC QIO=O, QIj=R m- j + 1 j=I,2, ... ,m
hing algorithm is carried QiO =0, Qij=Q(i-I)(j-l) -O:m-j+1Q(i-l)rn i =2, 3, ... , m~ j = 1, 2, ... , m
he gaussian elimination Using Q¡j, the observability matrix of (6-71) can be computed as
ces and then computing
The scaling problem is [QH Q12
QI~l
: S202 do not show any
lations and the gaussia n
V~. ~21 Q22 ~2m
. !
5. Find a realization of (;(s) and reduce it, by using the method discussed in
-time systeros is studied. Section 5-8, to an irreducible one. Then we have deg (;(s) = dim A.
uy input-output pair so
Whether the system is These methods can be readily programmed on a digital computer. A com
x in coprime fracUona\ parison of these methods in terms of computational efficiency and numerical
realization problem can stability seems to be unavailable at present.
blero in which the system To conclude this· chapter, we give a different derivation of the observable
. the impulse response or realization in Equation(6-131). lt wil\ be derived from three maps developed
introduced is applicable in algebraic systeffi theory. We· discuss the discrete-time case. Cons"ider the
~ed. This, however, will q x p strictly proper ratianal matrix G(z) = D-1(z)N(z), where D(z) and N(z)
316 IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, AND IDENTlFICATlON
are respectively q x q and q x p real polynomial matrices. It is assumed that If i = 1, j = 1, then D - 1(.
D(z) is row-reduced and its row degree is Vi' that is, b,.iD(S) = Vi' Clearly we
have b,.¡N(z) < Vi'
Let IK D denote the set of all q x 1 real polynomial vectors with row degrees
smaller than Vi' It can be readily verified that (IK D , IR) is a linear space over the
field of real numbers IR. The dimension of (IK D , IR) is n~ VI + Vz + ... + Vq.
(See Example S, on page 16.) In this polynomial vector space, we shall choose
the n columns of the q x n matrix:
1 z z v, - 1 O O ... This is the first column
O O O
columns of A except tr
J,]
_ O O O 1 z .. . Z "2 - 1 O O
(6-223) (6-131a). Consider now
L(z) = :: :::
.. ...
[ 0D(.H(z)) = D(z)O(
00 O 00 O 1z
Because Dh;.1 is a constal
as a basis. Its columns will be denoted by qij, i = 1, 2, ... , q; j = 1,2, ... , Vi'
O. Because D -1(Z)[(Z~
Let f(z) be a q x 1 rational vector. Define the operator O as
D- 1(z)[(z)D¡,.D';;.I. HeI
O(f(z)) = strictly proper part of f(z) (6-224)
OD(H(z)) =
If f(z) is a polynomial vector, then O(f(z)) = O; if f(z) is strictly proper, then
Note that the ith columl
O(f(z)) = f(z). It is a linear operator which maps (lRq(z), IR) into itself. Next
given by the ith column
we define, for any q x 1 real polynomial,
Now we show that th
OD(h(z)) = D(z)O(D -1(z)h(z)) (6-225) From (6-226), we have
It is a linear operator which maps (lRq[ zJ, IR) into itself. Because the ith row D-l(Z) =Dh;.I[H(z)(1 + f
degree of 0D(h'(z)) is at most V¡- 1, ihe range space of OD(h(z)) is (IK D, IR). =Dh~.I(I+E1Z-l
In algebraic system theory, a realization can be expressed by the following
three maps: for some constant E¡, i =
that if j < Vi, then (D -1(;
B D : IRP----* IK D ; u t-+N(z)u This completes the deriv;
A D : IK D----* IK D; x t-+OD(ZX) (6-226) three maps in (6-226). l
C D : IK D ----* IRq; x----*(D- 1(z)X)_1 maps in (6-226) are useful
SS5 and Section 9-8.
where, ifwe expand D- 1(z)x =lX 1Z- 1 + lXzZ- z + "', then (D- 1(z)X)_1 ~lXl' If
the columns qu in (6-223) are chosen as a basis of IK D , then the map R D is the
representation of N(z) with respect to the basis, that is,
N(z) = [(z)B 6-1 Find the degrees and Ü
matrices.
as shown in (6-130). The ith column of A is the representation of OD(Zqij) with
"ll'~:)'
s +3
respect to the basis in (6-223). Indeed, if we rewrite (6-12S) here as ~-
s +2 s+
D(z) = H(z)D h ,. + [(z)D¡,. (6-227 ) s +1
Consider
H(z) = D(z)D,;;.1 - [(z)D¡I.D,;;.1
b.
l ' ~ l' +:)~d2)J
1
(s +2)
1)'
(s+1)(s+2)
_._-_.. -_._---._..- ..
----,-------------_.
---_•... --~----
_
AND IDENTIFICATION
PROBLEMS 317
~
O columns of A except the (¿~ = 1 vm)th columns can be obtained as shown in
O (6-131a). Consider now
] (6-223)
TID(H(z» = D(z)TI(D- 1(z)H(z» = D(z)TI[Dh~l - D- 1(z)L(z)D .D';;.1 ]
z z Vq - 1 "
Because D';; 1 is a constant matrix (a polynomial of degree O), we have TI(D¡~.l) =
... , q; j = 1, 2, ... , Vi' O. Because D-1(Z)L(z)D1J.Dh~1 is strictly proper, TI[D -1(z)L(z)D 1r Dh";.1 ] =
tor TI as D- 1(z)L(z)D l.. D,;;'1. Hence we have
z) (6-224 ) nDCH(z)) = - D(z)D-1(Z)L(z)Dl.. D,~1 = - L(z)D 1,.D¡;;' 1
) is strictly proper, then Note that the ith column of H(z) is zq¡v,. Hence the (2:::n= l)th column of A is
z), IR) into itself. Next given by the ith column of -Dl,.D/~.1, as shown in (6-131a).
Now we show that the ijth column of e in (6-131a) is equal to (D-1(z)qij)_ l'
(6-225) From (6-226), we have
f. Because the ith row D-1(Z) =D¡~.l[H(z)(I + H- 1(z)L(z)Dl")] -1 ~D/~.l(I + H-1(Z)L(zp¡.. t 1iI- 1(z)
D(h(z)) is(IK D , IR). =D,;;.1(I+E 1z- 1 +E 2z- 2 + ·.. )fi- 1(z)
lressed by the following for sorne constant E¡, i = 1, 2, . . .. From this equation, we can readily show
that if j<v¡, then (D-1(Z)qij)_1 =0 and (D-l(z)q¡v¡}=ihe ith column ofD,;;.l.
This completes the derivation of the observable realization in (6-131) from the
(6-226) three maps in (6-226). For a derivation of (6-226), see Reference Si02. The
maps in (6-226) are useful in solving polynomiaJ matrix equations; see Reference
SS5 and Section 9-S.
en (D- 1(z)X)_1 ~Cl1' If
then the map RD is the
6-1 Find the degrees and the characteristic polynomials of the following proper rational
matrices.
~ntation of TID(zq¡J with
bt ;1)'
(s +2)
Id:~d2)1
(s +1)(s +2) .'
318 IRREDUCIBLE REALIZATlÜNS, STRICf SYSTEM EQUIVALENCE, AND lDENTlFICATlüN
S +1
6-7 Find the controll~
6-2 Find the dynamical-equation description of the block diagram with state variables lordan canonical-form d:
chosen as shown in Figure P6-2.
6-8 Set up a linear time
(p3 +7. 1(t),
where pi ~ d/dti.
Figure P6-2
6-9 Find irreducible ce
matrices
6-3 Find the dynamical-equation realizations of the transfer functions
6-4 Find lordan-canonical-form dynamical-equation realizations of the transfer func 6-10 Find irreducible re2
tions
a.
(S + 1)(s + 2)(s +3)
S2 + 1
b. - - "
(s +2)3
Use two different methods
S2 +1
c. S2 +2s +2
6-11 Find irreducible rea
lf the lordan-form realizations consist of complex numbers, find their equivalenl dynamical
equations that do not contain any complex numbers.
6-5 Write a dynamical equation for the feedback system shown in Figure P6-S. First
find an overall transfer function and then realize it. Second, realize the open-loop transfer
function and then make the necessary connection. Which realization is more convenient in
computer simulalions if the gain k is to be varied? Use two different methods.
Figure P6-5 I
I
[See Problem 6-3(a).]
I
1
AND IDENTIFICATION PROBLEMS 319
6-7 Find the controllable canonical-form, the observable canonical-form, and the
agram with state variables Jordan canonical-form dynamical-equation realizations of l/sO..
6-8 Set up a linear time-varying dynamical equalion for the differential equation
(p3 +a 1(t)p2 +C(z(t)p +C(3(t))y(t) = (f30(t)p2 + f31 (t)p + f3z(t))u(t)
-~
~
6-9 Find irreducible controJlable or observable canonical-form realizations for the
matrices
'unctions
~ :3J
Use two different methods.
ls: s +1
1
5
s +2
+1
6-12 Find a linear time-invariant discrete-time dynamical eq uation whose sampled trans
fer Function is
i
I [See Problem 6-3(a).]
1
l.
320 IRREDUCIBLE REALlZATlüNS, STRICT SYSTEM EQUIVALENCE, ANO IDENTlFICATlüN
6-13 Find an irreducible, discrete-time dynamical-equation realization of the sampled 6-17 Show that (6-77) i~
transfer-function matrix k=O, 1,2, ....
Z~3J
Z+2
6-18 Show that (6-77) is
Z +1
Z Z +1 6-19 Consider the q x p
[
Z +1 Z +2
x =[)"
O)"
~Jx +[~¡J
b¡
u
KoH(v+i)= -K¡H(
or [K
where the overbar denotes the complex conjugate. Verify that by using the transformation where f is defined as in (6
x=Q¡x, where cal equation
-Xb¡ ~¡J _KVKÜ¡ I
Ql = [ -),E¡ b¡
. -K,,_¡K ü ¡ (
the equation can be transformed into x= :
[ -KzK ü ¡ (J
-KIK ü ¡ O
where
Y=[ Kü ¡ O
~xJ b=[~J
_ [ O
Á= cl =[ -2 Re (Xb¡cd 2 Re (b¡c¡)J
-AX A is a realization of G(s). Sh
controllable.
6-15 Verify that the Jordan-form dynamical equation
6-20 Use Problem 6-19 a
A 1 O O O O b¡
irreducible one. (Hint: Shil
O A 1 O O O bz
O O A O O O b3 O 2:
x= x+ u [O O'
O O O X 1 O El I
6-16 Write an irreducible dynamical equation for the following simultaneous differential
equation:
ealization of the sampled 6-17 Show that (6-77) is a realization of G(s) in (6-73) by establishing H(k +l)=CA k B,
k=0,1,2, ....
6-18 Show that (6-77) is always observable but not necessarily controllable.
1t is assumed that there exist q x q constant matrices Ki, i =0,1,2, ... , v, such that
: x +: :
-K 2 K ü 1 O O 1 K3 K 4
-K 1 Kü l O O O K2 K3
O O OJ x + H(O) u
is a realization of (;(s). Show that the realization is always observable but not necessarily
controllable.
6-20 Use Problem 6-19 and (6-99) to find a realization of (;(s) in (6-97). Reduce it to an
irreducible one. (Him: Shift k 2 in (6-99) to right to forro
O 2:01:10:1 0J
[ O O: O 2: O 3: O 1 = [K¡ K2 K3 KoJ
u , , ,
and then proceed. This procedure can be used to find an irreducible realization of the form
in (6-131) except the rearrangements of state variables as discussed in Section 5-8. See
References S63 and S201.)
6-21Let N(s)D- l (s) be a q x,n strictly proper rational matrix and let
N(s)=N o +N l s+'" +N._ l s·- l
O
-DI -D 2
O
1
O
. ] ["
O
:
-~I'-l'
x+:
O
~,' I
u
y=[ No NI N2 N.- 1 Jx
is a control1able realization of N(s)D-1(s).
'----------_.~-~~~._-
6-22 Consider the equivalent time-invariant dynamical equations {A, B, q and {A, B, Cj It is assumed that all eiger
with A=PAP- I, B=PB, and c=cp- 1. Let wer and W Ol be the contr01lability and system
observability grammians defined in Theorems 5-7 and 5-13. Show that
W C1 = PWerp* where
and \VOl =(P-1)*WOlP- 1
6-23 Consider the irreducible dynamical equation lA, B, q. Using Theorem E-4 to
is also internally balanced.
write its grammians as
(Hint: Write W=diag {W 1
and
6-26 Show that ror any sin.
where R*R = I and L 2 =diag [/,i, ;.~, ... , l.;':. Define
is used in the identification
H~L:R:RcLc nonzero input sequence is al
is used?
Using the singular value decomposition (Theorem E-S), we write
w = w coo = W 000 = LH
where W is a diagonal matrix with positive diagonal entries. Show that W is the unique
solution of
AW +WA* = -BB* or WA +A*W = -C*C
,ns {A, B, C} and [A, B, c: Jt is assumed that all eigenvalues of A have negative real parts. Show that the reduced
Je the controllability and system
ow that x= Ax +Bu, y= ex
where A=A¡¡ -A 12 Ai}A 21
B= B¡ -A¡zA:;jB z
C=C¡ -C2A221A21
Using Theorem E-4 to is also internally balanced. This result is useful in syslem reduction. See Reference S93.
(Hint: Write W =diag {W¡, W 2 } and use Problem 6-24.)
1*
'o
6-26 Show that for any single-variable system with a proper transfer function, ifS v( - \', IX))
is used in the identification, the impulse sequence l urO) = L u(i) = O, i = 1. 2, L .. ~ or any
nonzero input sequence is always persistently exciting. ls the statement still true ifS,.(O, IX))
is used?
e
6-27 ldentify a linear time-invariant system from the following input-output sequence:
2 is chosen as k . O 1 2 3 4 5 6 7 8 9 !O 11 12 13 14 ..
_
u\(k) 1 1 -1 O O 1 O -\ O \ O -\ -l -2 \
u2(k) \ O O -\ O O O \ -\ O -\ \ O 1 -2
)'I(k) -\ \ 2 1 O O \ \ \ O \ O O -\ -2
)'2(k) -2 -0.5 \ -0.5 -0.5 -0.5 \ \-0.5 -0.5 \ - \ -2.5 -4 -7
ving correspondíng gram -----
First identify the system by disregarding the initial conditions and then identify the system
as well as lts initial conditions.
6-28. Lel {A, B, C} be irreducible and let G(s)=C(sl -A¡- IB= N(s)O-\(s), where O(s)
and N(s) are righl coprime. Show lhat for any C I , there exists a polynomial matrix N 1 (s)
ld internally balal1ced on such that C ¡(si -A)-I B = N I(S)O - I(S). Show also lhal, conversely, for any strictly proper
N 1(5)0-I(S), there exisls a C I sllch that N ¡(5)0-I(S) = C¡(sl - A)- ¡ B.
= -C*C
form u=r+ Kx on a <
the linear state feedba<
of an uncontrollable d
In state feedback,
outputs. This assump
want to introduce stal
l
qn D. b
V= cA
e J
qn-l ~ A
eA~-l q"-2 D. A
ql ~ A
and if FE l and FE l are observable, then
is linearly independen1
or (7-6) from Figure 2-5 that i
¡th column ofthe new
Thus, if two same-dimensionallinear time-invariant, single-variable dynamical
to the basis {ql' q2' ..
equations are known to be equivalent and if they are either controllable or
observable, then the equivalence transformation P between them can be com Aql=(
puted by the use of either (7-5) or (7-6). This is also true for multivariable
controllable or observable time-invariant dynamical equations. For multi
variable equivalent equations, the relation Ü =}FU still holds. In this case U
is not a square matrix; however, from Theorem 2-8, we have ÜU* =PUU*
and P = ÜU*(UU*)-l.
Let the characteristic polynomial of the matrix A in (7-1a) be
Theorem 7-1
If the n-dimensional linear time-invariant, single-variable dynamical equation
FE l is controllable, then it can be transformed, by an equivalence transforma
tion, into the form
)btained by introducing
;:mstant matrix. Then
eFE¡: O 1 O O O O
O O 1 O O O
O O O O O O
x= x+ u (7-7a)
O O O O O
~tively,
-IX" -IX" _¡ -IX"_ Z -lX z -IX¡ 1
(7-3) y=[ /31l f3n- 1 /3" - Z /3z /3¡]x+eu (7-7b)
where IX¡, IXZ' . •. , ct" are the coefficients of the characteristic polynomial of A,
(7-4 )
and the /3¡'s are to be computed from FE¡. The dynamical equation (7-7) is
, FE¡ are control1able,¡
said to be in the controllable canonical formo The transfer function of FE ¡ is
llar. Hence, from (7-4) /3 s"- ¡ +/3 s" - z + ... +/3
(7-5b)
.,1 ~ l
we have ÜU* =PUU*
<llll_01X"J
1 (7-la) be
Aq,~q, -,q.~[q,
- •• q, ... q.] lLJ
ble dynamical equation
.................................................
~quivalence transforma
1
Aqll = qll-¡ -1X¡q" = [ql qz ... q"] [ : ]
I
I _-IX¡
I
j
328 STATE FEEDBACk AND STATE ESTlMATORS
Hence if we choose {q1' qz, ... ,qn} as a new basis of the state space, then A as one might expect.
and b have new representations of the form
o 1 O O O O
O O 1 O O O
O O O O O O
A= b=
(7-10)
O O O O O This can be directly v
-eJ. n -eJ. n - 1 -eJ.n-z -eJ. z -eJ. 1 1 tions FE 1 and CFE 1
hence Q = UÜ- 1 . I
The matrices A and b can also be obtained by using an equivalence transforma columns of Q as new
tion. Let Q ~ [q1 qz ... qnJ ~ p- 1, and let x =Px or x·=Qx, then the puting UD-1, we see
dynamical eq uation FE 1 can be transformed into that is,
Q =[q1 q
x=Q- 1AQx +Q- 1bu
We have the folle
y=cQx+eu
observable dynamical
The reader is advised to verify that Q - 1AQ = A or AQ = QA and Q -1 b = b.
The vector e is to be computed from cQ as Theorem 7-2
If the n-dimensional l
(7-11 )
FE 1 is observable, thel
Hence the controllable dynamical equation F El has the equivalent controllable into the form
canonical-form dynamical eq uation C F E l ' This proves the first part of the
theorem.
The dynamical equations FE 1 and eFE 1 are equivalent; hence they have
the same transfer function. lt has been shown in Sections 4-4 and 6-3 that the
transfer function of CFE 1 is equal to g(s) in Equation (7-8). Q.E.D.
One may wonder how we obtain the set of basis vectors given in (7-9). The dynamical eq uati
This is derived in the following. Let D be the controllability matrix of the moreover, the transfer
controllable canonical-form dynamical equation, then
g(s) =
ü ~ [Ri A1bJ ... An- 1 bl J
O O O
O O O el
This theorem can l
theorem of duality (Th
(7-12)
O O 1 en- 3
alence transformation
O 1 el en-z
Equation (7-6): P = V·
1 el ez en- 1
of the observable cano
matrix in (7-12). Hen
where and (7-15) is given by
k-1
ek = - ¿
i=O
eJ.i+ 1e k-i-1 k = 1, 2, ... , n - 1; ea = 1
The controllability matrix Ü is nonsingular for any eJ. 1 , eJ.z, . .. ,eJ.1I" Therefore
the controllable canonical-form dynamical equation is always controllable,
CANONICAL-FORM DYNAMICAL EQUATlONS 329
he state space, then A as one might expect. The inverse of Ü has the following very simple form:
o
a n_ 1
a,,-z
an-Z
a ll - 3
...
••.
al
1
1]
O
O
Ü- I =:: :: ~ A (7-13 )
[ 1 . "
al O O
O
b= (7-10)
1 O O O
This can be directly verified by showing that DÜ- I = 1. The dynamical equa
O
tions FE I and eFE I are related by the equivalence transformation x=QX;
1
hence Q = uD - l. In the equivalence transformation x = Qx, we use the
columns of Q as new basis vectors of the state space (see Figure 2-S). By com
:¡uivalence transforma
puting UÜ-l, we see that the columns of Q are indeed those given in (7-9);
)x or X·= Qx, then the
that is,
Q=[ql qz qll]=[b Ab
oO. An-Ib]ü-l=UA (7-14 )
oO'
Theorem 7-2
lf the n-dimensional linear time-invariant, single-variable dynamical equation
(7-11 ) FE lis observable, then it can be transformed, by an equivalence transformation,
equivalent controllable into the form
es the first part of the
1 O
O O -a ll -
n ] .~n]
f3n-l
l
alent; hence they have
X= ~ 1 O -~n-z X + ~n-z u (7-15a)
Ins 4-4 and 6-3 that the [
'-8). Q.E.D. O O 1 -al I
y=[O O O 1 ]x+eu (7-15b)
vectors given in (7-9). The dynamical equation (7-1S) is said to be in the observable canonical form;
Jllability matrix of the moreover, the transfer function of FE¡ is
f3 ¡ sn - ¡ + f3 ZSil - Z + ... + f3 n-¡ s + f3 n +
gA( s) = . 1 e
, :/' -1-0:. 1sl1 - -1-·' +CX¡¡_lS -~-(j~i1
[a"~, i]~~t,
a ll - z a¡
1; ea = 1 an - z (Xn-3 1
P= : =AV (7-16 )
, az,··· ,.a no Therefore al 1 O
is always controllable, 1 O O O cAn-¡
330 STATE FEEDBACK AND STATE ESTIMATORS
y=cQx+eu
This equation has the same A matrix as the one in (7-15) and is easier to obtain
because its equivalence transformation Q is simpler. However, the usefulness The controllable
of an equation of this form is not known at present. also be obtained by
equation. The coeff
Example 1 canonical-form eq uati
Transform the following controllable and observable single-variable dynamical The controllable
equation dynamical equations a
They are also useful iJ
x~ [~
2
computer. For exan
-1 (7-17a)
generally need n Z am]
2
and 2n amplifiers and
y=[O O (7-17b) transformed into the c(
into the controllable and observable canonical-form dynamical equations. of (7-16), then the num
The characteristic polynomial of the matrix A in (7-17) is . from n Z + 2n to 2n.
A-1 -2 -lO]
=..1. -9..1.+2 *Multivariable cas(
!l(A)=det [ -~ A~~
3
multivariable dynamic
U=[~1 : 1~]
2 12
where A, B, e, and E .
respectively. Let b¡ bf
Ifthe dynamical eq
From Equations (7-5b) and (7-13), we have U = [b¡ b z ... bp
Q=UA =
2 4 16] [-9 O1]
[11 62 128 O1 O1 OO =
[-2 4 2l
-1 6 1J has rank n. Conseque
3 2 1
U. There are many v
The matrix Q can also be obtained from (7-9). From(7-11), we have [rom toe n x np compos
for choosing n linearly
[/33 /3z /3IJ=cQ=[3 2 lJ state space. These two
Hence the equivalent controllable canonical-form and observable canonical-
CANON1CAL-FORM DYNAMICAL EQUATIONS 331
X~[!
y= [O
The transfer function of (7-17) is thus equal to
A S2 +2s +3
g(s) = S3 -9s +2
and is easier to obtain
The controllable or observable canonical-form dynamical equation can
:owever, the usefulness
also be obtained by first computing the transfer function of the dynamical
equation. The coefficients of the transfer function give immediately the
canonical-form equations. I
4 2l
62 lJ1
has rank n. Consequently, there are n linearly independent column vectors in
U. There are many ways to choose n linearly independent column vectors
1 (7-11), we have
¡rom the n x np composite matrix U. In the following, we shall give two schemes
for choosing n linearly indepeildent column vectors to forro new bases for the
state space. These two schemes were first discussed in Section S-3.
observable canonical-
-------_...._._--_._-------_. __.__ ._----..__ .__ .-_.. ~ .. _-~_._ ... _._---~--_._ .. _-_._-_._. __._.._._-_. __.. _-------------_._-_._.... _-_.._------.-._ .. -._._------_ .. _._---_._-,----_ .. _._._----_._--~---_ .. __ .__._._ ..._-._._.- _.. _..__ . -----._.. __ ... ~--
Seheme 1 OOO"'Ox
We start with the vector b 1 and then proceed to Ab 1, AZb 1, up to AIi,-l b 1 until 100"'O x
the vector AIi'b 1 can be expressed as a linear combination of {b¡, ... , AIi¡-l bd. 010···0 x '
If {1.1 = n, the equation can be control1ed by the first column of B alone. lf
jil < n, we select b z, Ab z, up to Aliz -1 b z, until the vector Alizb z can be expressed OOO···lx·
---- J
as a linear combination of {b 1, ... , AIi¡ -1 b 1, b z, ... ,Aliz -1 b z} (see Problem 7-5). (PI X {1.1) ,
If {1.1 + {1.z < n, we proceed to b 3, Ab 3, ... ,AIi' -1 b 3 and so forth. Assume that
jil +{1.z + {1.3 = n, and the n vectors
A=
are linearly independent. An important property of this set is that the vector
Aliib i can be expressed as a linear combination of the preceding vectors; for
example, Ailzb z can be expressed as a linear combination of {b 1, Ab 1, ... ,
Ail¡-l b 1, b z, Ab z, ... , Ailz -1 bz}.
Seheme 2
The linearly independent vectors are selected in the order of (7-19); that is, we where the X's denote'
start from b 1, b z, ... , b p and then Ab 1, Ah z, ... , Ab p, and then AZb 1, AZb z, zeros. This can be ea~
and so forth, until we obtain n linearly independent vectors. Note that if a representation of Aqi ,
vector, say Ab z , is skipped because of linear dependence on the vectors If the set ofvectors
{b 1, b z, ... , b p, Ab 1}, then aH vectors of the form Akb z, for k ~ 1, can also be will be of the form 3
skipped becatise they must also be dependent on the previous columns. After ,
choosing n linearly independent vectors in this order, we rearrange them as O O x;
1 O x:,
{b 1"·" AJl,-lb'b
1,2,'··'
AJlz-1b"b
2,"" p, ... ,
AJlp-1b}
P (7-21 ) O 1 x,
,
,
where /11 + /1z + ... + /1p = n. Note that the main difference between this O O x :
scheme and Scheme 1 is that in Equation (7-20) Ail'b 1 can be expressed as a
linear combination of {b 1, Ab 1, ... , Ail¡ -1 b 1}, whereas the vector AJl¡ b 1 in
-(~~ ""X ~:f ~- O' - :
(7-21) cannot be expressed as a linear combination of {b 1, ... , AJl¡-l bd; AJl¡ b 1 x :1
x :O
is general1y linearly dependent on al1 vectors in (7-21).z Similar remarks
apply to AJlib¡, for i = 1, 2, ... , p.
Now if the set of vectors in (7-20) is chosen as a basis of the state space of Á= :'(_ _0_ i
FE or, equivalently, let x = Q-1x where (1
Ailz- 1b z
b 3 ... Ail, -1 b 3J
--------------
x,
,
.
then the matrices A and B will be of the forms
x,
x'
.,
,
,
x:
2 More can besaid by using the relative sizes of l1i. For example, if 1115,1125,'" 5,l1p, ihen A!"b 1
depends on{Akb¡, i= 1,.2, ,p; k =0,1, ... ,111 -l}; A"2b z depends on {Akb¡,k=O, 1, ... ;0 111 - 1;
Akb¡,i = 2, 3, ... , p; k =0, 1,0 ,112 -l} and so forth.
3 Using footnote 2, sorne elen
----- --- - ----- - - -
OOO"'Ox x: x 1 O O
:olumn of B alone. lf I : 1
:010"'Ox, x O O O
A= ,
l' ••
B= b4 "'bp
:OOO"'lx x OOO
is set is that the vector - - -(~~ -x-fi;)- -- ~ 0- 0- Ó-.-. ~ -Ó -.; OO 1
preceding vectors; for :100"'Ox OOO
.ation of {bl> Ab¡, ... , :010"'Ox OOO
1: : :
I •••
:OOO"'lx OOO
(j].3 x j].3)
(7-22)
ler of (7-19); that is, we where the X's denote possible nonzero elements and unfilled positions are all
" and then AZb¡, AZb z, zeros. This can be easily verified by observing that the ith column of A is the
ectors. Note that if a representation of Aq¡ with respect to the basis vectors [q¡ qz '" qn].
¡dence on the vectors lfthe set of vectors in (7-21) is used as a basis, then the new matrices Á and B
z, for k;:::l, can also be will be of the form 3
'evious columns. After
ve rearrange them as O O x ,, .x', x 1 O O
x ,,
(7-21 )
1 O
O 1 x ,
,
.
x;
x,
,
. x
x
O
O
O
O
O
O
, ,
, ,
lifference between this O O x'! x'J , x
------- O O O
can be expressed as a -----------
x : ,,
1 - __ - - - -
(111 X )J.l) x: O O x O 1 O
.s the vector AJL, b¡ in
I1¡-lb}' x :1 O X: ,, x O O O
'b ... , A l ' Al1'b 1 ,
x : O 1 x: x O O O
~ 1). z Similar remarks . ,, . ,
, ,
x , X , :O O
I
x O O O
5,/1" then A"'b!
.,5,/125,'" (7-23)
on{Akb"k=O, 1, ... ,/1,-1;
3 Using footnote 2, sorne elements denoted bY X can be replaced by' zeros.
I
J
334 STATE FEEDBACK AND STATE ESTIMATORS
where the X'S again denote possible nonzero elements and unfilled positions are
all zeros. The matrices Aand B can be verified by inspection. The matrix C
in both cases are to be computed from CQ.
By comparing (7-22) with (7-23), we see immediately the differences in the
matrices A and B due to the different choices of basic vectors according to
Schemes l and 2. The matrix Á has three blocks in the diagonal, whereas the
matrix Á has p blocks. The first three columns of B are very simple, whereas
every column of Bconsists of only one nonzero element.
The usefulness of the forms in (7-22) and (7-23) is not known at present.
The purpose of introducing these two forms is to show that there is no additional
conceptual difficulty in developing canonical forms for multivariable dynamical
equations. By rearranging the vectors in (7-20) or (7-21), different dynamical
equations can be obtained. One of them will be discussed in Section 7-3. For
a survey of various canonical forms, see References S154. Figure 7-1 A state f
where X is the n x l state vector, u is the scalar input, y is the scalar output, We see that in the
A is an n x n real constant matrix, b is an n x 1 real constant column vector, invariance are not use
and c is a 1 x n real constant row vector. In state feedback, every state variable
is multiplied by a gain and fed back into the input terminal. Let the gain
between the ¡th state variable and the input be k¡. Define k~ [k l k 2 .•. knJ. Corollary 7-3
Then the dynamical equation of the state-feedback system shown in Figure 7-1 is
The con troI1ability af
FE{: x= (A + bk)x + br (7-25 a ) is invariant under any
y=(c+ek)x+er (7-25b)
which is obtained by replacing u in (7-24) by r +kx, where r is the reference Note that Theoren
input. Note that the dynamical equations (7-24) and (7-25) have the same
dimension and the same state space. Now we shall show that the controllability p[b Ab
of a linear time-invariant dynaniical equation is invariant under any linear
state feedback. for any 1 x n real cons
A1though state fe
Theorem 7-3 equation, it is alway:
dynamical equation b
The state feedback dynamical equation FE{. in (7-25) is controllabie .ror any ( -1/e)c, then the stat(
1 x n real vector k if and only ifthe dynamical equation FE I in (7-24) is con is observable. If e =0
trollable. feedback dynamical ec
Problem 7-14).
STATE FEEDBACK 335
unfilled positions are
:ction. The rnatrix e
ot known at present.
.there is no additional
lltivariable dynarnical
e
), different dynarnical
d in Section 7-3. For
Figure 7-1 A state feedback system.
Proof
First we show that the controllability of FE¡ implies the controllability of
,back u = f +Kx on a FE{. Let Xo and X¡ be two arbitrary states in the state space L. By the con
denotes the reference trollability assurnption of FE ¡, there exists an input u that will transfer X o to Xl
ed to be real. In this in a finite time. Now for the state-feedback dynamical equation, if we choose
lable for the feedback. r(t) = u(t) - kx(t), then the input r will transfer X o to Xl. Therefore we conclude
that F E{ is controllable.
linear time-invariant We see from Figure 7-1 that the input r does not control the statex directly,
·it generates u to control x. Therefore, ir u cannot control x, neither can r. In
(7-24a) other words, if FE¡ is not controllable, neither is FE{. Q.E.D.
(7-24b)
I is the scalar output, We see that in the proof, the assumptions of single-variableness and time
lstant column vector, invariance are not used. Therefore we have the following corollary.
k, every state variable
~minal. Let the gain
k~ [k¡ kz ... knJ. Corollary 7 - 3
shown in Figure 7-1 is The controllability oí a multivariable linear Unle-,¡arying dynamical cCjuation.
(7-25a) is invariant under any state feedback of the form u(t) = r(t) +K(t)x(t). •
(7-25b)
lere r is the reference Note that Theorem 7-3 can also be proved by showing that
(7-25) have the sarne
p[b Ab ... An-lb] =p[b (A+ bk)b ... (A+ bk)n-¡b] (7-26)
.hat the controllability
ant under any linear
for any 1 x n real constant vector k (see Problern 7-10)..
Although state feedback preserves the controllability of a dynamical
eq uation, it is always possible to destroy the observabiiity property of a
dynarnical equation by sorne choice of k. For example, if e =/=-0 and if k =
; controllable for any ( - l/e)c, then the state-feedback equation (7-25) is notobservable even if FE ¡
FE ¡ in (7-24) is con- is observable. If e = 0, it is still possible to choose sorne k such that the state
feedback dynarnical equation will not preserve the observability property (see
Problern 7-14).
1
336 STATE FEEDBACK ANO STATE ESTIMATORS
x=G
y=[1
n +[~Ju
x
2Jx
If k is chosen as
l!
CFE!: O O O
i~
O 1 O
O O O
-Cl.fT
O
-a"_1
O
-a,,-z
O
-az
r]i+[!]U
-al 1
Algorithm
Given a controllable {A
1 x n real vector k such
(7-27 a)
its eigenvalues.
y= [fi" fi,,- 1 fi,,- 2 ... f3z filJx +eu (7-27b)
1. Find the characteristic
Let A and b denote the matrices in (7-na), then A = PAP- 1, b = Pb. Because 2. Compute (s -I¡)(s -~ .
. of the equivalence transformation, the state feedback becomes 3. Compute k = [a" ~. a"
.. u = r +kx = r.+kp-Ix ~ r +kx (7-28) 4. Compute q"-i =Aq,,_
5. FormQ=[ql qz '
where k ~kP-I. It is easy to see that the ser Of the eigenvalues of (A + bk) is 6. Find P ~ Q - l.
equal to the set of the eigenvalues of (A + bk). Let the characteristic poly 7. k =kP.
---------------_ ..-. .. _. --_._---------_ .. _._.-------
--==c::===_-
-~~~,,===~===- __ ~-_-_-~~.:__-=_-=-=---=-~----------------------------~--------
------.--_.- -- ----,_.~--------_._
-----------------.---
If k is chosen as
(7-29)
CFE{: O 1 O O O O
O O 1 O O O
O O O O O O
x= x+ l'
O O O O 1 O
-Cl. n - ii"_ ¡ -IX" - 2 -ii 2 -IX¡ 1
I (7-30a)
y=[I1" +e(cx"-éX") 11"-¡ +e(cx"_¡-éX"_¡) ... 11¡ +e(cx¡-iX¡)]x +er
be used to control the (7-30b)
'+l~
o
u Algorithm
Given a controllable {A, b} and a set of eigenvalues Xl> X2 , o.. , X... Find the
-IX¡ 1 x n real vector k such that the matrix (A +bk) has the set {Xl' X2 , o.. , X"} as
(7-27 a) its eigenvalues.
eu (7-27b)
1. Find the characteristic polynomial of A: det (sl- A) = Sil +IX ¡s" - ¡ + ... +IX...
p-¡, b = Pbo Because 2. Compute (s - X¡)(s -' X2 )· •• (s - Xn ) = s" +iX ¡ s" -¡ -+: ., +ii...
;omes 3. Computek=[cx"-iX" IXn-1-iX n -¡ o.. IXI-~il
(7-28)
4. Compute q.. _ i = Aqn-i+ ¡ + (X¡q", for i = 1,2, . o. , (n-l), with q" = bo
5. Form Q = [q¡ Q2--' o. qnl
envalues of (A + bk) is 6. Find P ~ Q - l.
le characteristic poly 7. k=kP. I
._._----~-----------_._-----
This algorithm can also be developed directly without transforming the case, a differen t rr
equation explicitly into the controllable canonical formo Let ~(s) be the (Appendix F) will
characteristic polynomial of A + bk. Then we have single-variable case
~(s) = det (sI - A - bk) = det [(sI - A)(I - (sI - A)-l bk)J
Stabi Iization
= det (sI - A) det [1 -(sI - A)-lbkJ
Ir a dynamical equa
This can be written as, by using (3-65) and defining ~(s) = det (sI - A),
assigned by the intr
~(s) = ~(s)(1 - k(sl - A)-l b) controllable, one m
or ~(s) - ~(s) = ~(s)k(sl - A)-l b (7-31 ) shown in Theorem :
controllable, by a
From this equation and using Problem 2-39, Equations (7-13) and (7-14), we
transformed into
can readily establish k =kUA =kQ or k =kP.
Example 2
where
Consider the inverted pendulum problem studied in Example 4 on page 185.
Its dynamical equation is, as derived in (5-25),
and the reduced ec
X~l~
1 O form of A, the set 01
O -1 of All and of A22 .
O O A22 is not affected
O 5
r +kx. Therefore;
y = [1 O O
other hand, all the
It is controllable; hence its eigenvalues can be arbitrarily assigned. Let the
{A lb lid is control
desired eigenvalues be - 1, - 2, - 1 ±j. Theri we have can be arbitrarily l
required to be a re:
~(s) = (s + 1)(s + 2)(s + 1 + j)(s + 1 - j) = S4 + 5s 3 + lOs 2 + lOs +4 in pairs.
The characteristic polynomial of In the design of
eigenvalues (the eig(
A+bkJ ~
1 (the eigenvalues wit
k2 the aboye discussior
O the equation canno
l-2k l -2k 2 transformed into th,
eigenvalues of Azz h
where k = [k l k2 k3 k 4 J can be computed as a dynamical equatio
det (sI - A - bk) = S4 +(2k 4 - k 2 )S3 +(2k 3 - k1 - 5)S2 + 3k 2 s + 3k l form dynamical eqt
eigenvalues are cont
The comparison of the coefficients of det (sI - A - bk) and those of ~(s) yields of Statement 4 of T
k1=t k 2 =lf k 3 =s,¡ k 4 =li if and only ifthe ma
that Al l and An in
Hence the introduction of the state feedback
u(t)=r(t)+[i lf ~ liJx Effect on the nume
will place the eigenvalues of the resulting equation at - 1, - 2, -1 ±j. I Consider the dyna
c(sl - A) - 1 b +e. Si
Even for this simple {A, b}, direct computation of det (sI - A - bk) is very 7-4 we conclude tha
complicated and is not suitable foi computer computation. Hence, for large duction ofstate feec
n, the use of the algorithm is much more convenient. In the multivariable g(s) are shifted by st:
-_._._"~-~-----~,~._---.=.",.-===-====--=-::==-=----~'-':'~=-:::"-:::=:':'''....--::::'''~==-.==::::--==-----~ . _---_.
-':::'--:::.='::::-- -==--=--=-.-.:.......===...-=--===-:::..:=....::.::=.::~=='::::.====--------_
out transforming the case, a different method of computing k by solving a Lyapunov equation
m. Let li(s) be the (Appendix F) wilI be introduced. The method is directly applicable to the
single-variable case. To avoid repetition, it will not be discussed here.
Stabilization
Ir a dynamical equation is controlIable, then alI the eigenvalues can be arbitrarily
1= det (sI - A),
assigned by the introduction of state feedback. If a dynamical equation is not
controlIable, one may wonder how many eigenvalues can be controlIed. It is
(7-31 ) shown in Theorem 5-16 that if a linear time-invariant dynamical equation is not
controllable, by a proper choice of basis vectors, the state equation can be
(7-13) and (7-14), we
transformed into
x=Ax +bu (7-32)
where -_[AO A
A-
A
11 12 ]
(7-33)
lmple 4 on page 185. 22
zeros of g(S) remain unchanged after the introduction of state feedback. We which becomes, in t
prove this by showing that the numerator of the transfer function of FE 1 in
(7-24) is equal to the numerator of the transfer function of eFE 1 in (7-30). y(t) = (e
The transfer function of FE 1 is for t ~ O. If all eige
A() 131S,,-1 +132S,,-2 + ... +f3" + approaches zero as
gs = e Ir {A, b} is controlla
s" +O:I S" 1 + ... +0:"
by a proper choice
es" +(131 +eO:l)s,,-1 + +(13" +eo:,,) deeper inside the le
(7-34 )
s" +O:IS,,-1 + +0:" However, in this cas
The transfer function of (7-30) is saturation. Further
the choice of proper
[131 +e(O:I -&I)]S,,-1 + ... +[13" +e(o:" - IX,)] by a compromise be
g¡(s) +e
s" +O:I S" -1 + ... +0:" As discussed in .,
es" +(131 +e0: 1)s,,-1 + +(13" +eo:,,) by state feedback; h
(7-35 ) we may choose p = l
s" +O:I S" 1 + +á"
which has the same numerator as g(s) in (7-34). This proves that the zeros of
g(s) remain unchanged after the introduction of state feedback. and the design is COI
This property can be used to explain why a state feedback may alter the {A, b} is not con trol
observability property of an equation. Ir sorne of the poles are shifted to have complete contr,
coincide with zeros of g(s), then the degree of g¡(s) in (7-35) is less than n, and Much more can b
the dynamical equation in (7-25) is not irreducible. In this case, (7-25) must It appears that the d
be unobservable beca use it is controllable (Theorem 7-3). function and will be
discussed further in t
Asymptotic tracking problem-nonzero set point
In this subsection we use state feedback to design a system so that its output *Multivariable ca:
will track the step reference input r(t) =rd, for t ~O. Ir rd =0, the problem is multivariable dynam
called a regulator problem; if rd =1= O, it is a special case of the asymptotic tracking
problem. A step reference input can be set by the position of a potentiometer
and is often referred to as set point. In this problem, the input u is chosen as
where A, B, e, and l
u(t) = pr(t) +kx(t) (7-36) real matrices. In sta
where p is a constant gain and k is the feedback gain vector. The substitution
of (7-36) into (7-24) and then the application of the Laplace transform to the
resulting eq uation yields where r stands for a r
called the feedback g
;(s) = (e + ek)[(sl - A - bkt l X (O) + (si - A - bk)-1 bpr(s)] + epr(s) (7-37)
Ir r(t) = rd, for t ~ O, then res) = rJs. The substitution of the identity
(si - A - bk)-I S-1 =( -A - bk)-lS-1 +(sl -A - bk)-I(A + bk)-1 In the following,
trollable, then the ei
[see also Equation (H-I)] and
proper choice of K.
g¡(s)=(e +ek)(sl-A-bk):-lb +e (7-38)
Method 1
into (7-37) yields
In this method we e
;(s) = (e +ek)(sl - A - bk)-1 [x(O)1 +(A + bk)-1 bprd] +g¡(Ü)prds-1 problem and then ap!
STATE FEEDBACK 341
Jack. and the design is completed. We note that the design can also be achieved if
dback may alter the {A, b} is not control1able but is stabilizable. In this case, however, we do not
poles are shifted to have complete control over the rate of convergence of y(t) to rd'
5) is less than n, and Much more can be said regarding the design ofasymptotic tracking problem.
his case, (7-25) must It appears that the design can be carried out more easily by using the transfer
function and will be discussed in Section 9-6. Hence this problem will not be
discussed further in this chapter.
:m so that its output *Multivariable case. Consider the n-dimensional linear time-invariant,
multivariable dynamical equation
d = 0, the problem is
A matrix A is called cyclic if its characteristic polynomial is equal to its Theorem 7-6 4
minimal polynomial. From Theorem 2-12 or Example 1 on page 48, we If {A, B} is controlla
can conclude immediately that A is cyclic if and only if the Jordan canonical the eigenvalues of A
form of A has one and only one Jordan block associated with each distinct
eigenvalue (Problem 2-45). The term "cyclicity" arises from the property that if Proof
A is cyclic, then there exists a vector b such that the vectors b, Ab, ... ,An -1 b
span the n-dimensional real space or, equivalently, {A, b} is controllable. This Let the characteristic
property can be easily deduced from Corollary 5-21.
O O
2 21 O: O]
O O
det --------.
O 1:0 an - I 2a n_
A= O O 2: O O
O a'l-
[
~--~--~-r~--~ O O
The cyclicity assumption in this theorem is essential. Without this assump Theorem 7-7
tion, the theorem does not hold. For example, the {A, B} in lfthe dynamical equat
of the form u = r + Kx.
of (A + BK) can be arbi
appear in pairs.
IS controllable. However, there is no v such that {A, Bv} is controllable. 4 This theorem can be extel
lf allthe eigenvalues of A are distinct, then there is onlyone Jordan block real constant matrix H, th
associated with each eigenvalue. Hence a sufficient condition for A to be cyclic cyclic. The matrix A + BI
is that all the eigenvalues of A are distinct Theorem 7-6 is equally apl
--_ .. _--_._--_._._.----------- ------------- --_._----
-------------- --.-_.--
-=.~~::=.:----=-~~--
--_._--~-----
---~--~--------_._-------_._~
---~-- -~-- - - ~- --~-----
Theorem 7 _6 4
lmial is equal to its
: 1 on page 48, we If {A, B} is controllable, then for almost any p x n real constant matrix K, all
le Jordan canonical the eigenvalues of A + BK are distinct and consequently (A +BK) is cyclic.
i with each distinct
1 the property that if Proof
If Li(s) has repeated roots, then Li(s) and Li'(s) are not coprime. A necessary and
sufficient condition for Li(s) and Li'(s) to be not coprime is
'ormation, henee we
O O
ee the basic idea, we
1 O
o O a an-I 1
det ------------------ n------------------ ~ y(kij) =0
a,,-l 2a n-2 n O O O
O an-I (n - l)al n O O
O O O an-l n
Theorem 7-7
Vithout this assump
in Ir the dynamical equation FE in (7-40) is controllable, by a linear state feedback
of the form u = r + Kx, where K is a p x n real constant matrix, the eigenvalues
of (A + BK) can be arbitrarily assigned provided complex conjugate eigenvalues
appear in pairs.
is controllable. 4This theorem can be extended as follows: If {A, B,C} is irreducible, then for almost any p x q
lly one Jordan block real constant matrix H, the eigenvalues of A + BHC are distinct and consequently (A+ BHC) is
lion for A to be cyclic cyclic. The matrix A + BHC r~sults from the constant output feedback u = r + Hy. The proofor
Theorem 7-6 is equally applicable here.
-._._-_.~ .. _,_.. ~-~~_ __ _----_._-~ .. __.. ~------_ ..-----_.. __.__.._._--_._--_._------_._-_.. ~--"_._ _---~ _.~- .. _.. _ .. "-~ .. _~ _~---_._.,._----_.- __ _.__._-_._._._.,_ _---_._._.__
. .. ._--_._-----~-_ .. _.. _-_._-_._.- .. - ..'.
, __ .',_.'---_ __ .-
5 The choices of K , and vare not unique. They can be choserr arbitrarily, and the probability is
almost I that they will meet the requirements. In Theorem 7-5 of Reference S38, a procedure is
given to choose K , and v without any uncertainty. The computation required however is very
complicated.
From M-IM= 1, w
k = O, 1; e 13 Az h 1 = 1; el.
we do not have e 13 A2 bz =
1,2,3 and A 2b¡} and e l3
matrix 6
+ y
(7-43)
[A~l i::J [~lJ
and the eigenvalues of A22 are not affected by any state feedback. Hence we
Corollary 7-3). Hence conclude that the eigenvalues of (A + BK) can be arbitrarily assigned if and
:ontrollable. 5 Now we only if {A, B} is controllable.
i1 in Figure 7-2, with K z
Method 11
In this method, we shall transform {A, B} into the controllable form discussed
in (5-93) and (6-112) and then compute the required K. The procedure is
similar to the single-input case. In order not to be overwhelmed by notations,
we assume n =9 and p = 3. lt is also assumed that the controllability indices
+-Br are MI = 3, Mz = 2, and M3 = 4. Then the matrix
.Bvk) can be arbitrarily (7-44)
Jmbining the state feed- l
(see Section 5-3) is nonsingular. We compute M- and name its rows as
ell
e lZ
Q.E.D. e l3 ~ell'l
eZ l
e zz ~eZ1'2
e 13 A
e 13 A z
~elI'IA/ll-l
e 22
P = e 22 A
(7-45)
e 3 4
e 34 A
e 34 A 2
e 34 A3.
óThis matrix can be obtained directly from the coefficients of linear combinations of A~' b¡ without
[uation. computing explicitly M-l. See Reference S178.
346 STATE FEEDBACK AND STATE ESTIMATORS
: : O O 1 O O O O
: : O O O 1 O O O Method '"
X X x:x x:x X X X O O 1
We introduce a method o
(7-46)
forming A into a controlI~
where the unfilled positions are all zeros. The pair {A, B} is said to be in a equation. See Appendix j
multivariable controllable formo The introduction of u = r + Kx, where K is a
3 x 9 real constant matrix, yields Algorithm
x=(A +BK)x +Br Consider a controlIable {
Because of the form of B, all rows of A except the three rows denoted by strings n x P constant matrices. 1
values.
of x are not affected by the state feedback. Because the three nonzero rows
of B are linearly independent, the three rows of A denoted by strings of x can 1. Choose an arbitrary n >
be arbitrarily assigned. For example, we may choose K so that CA +BK) is with those of A.
of the form 2. Choose an arbitrary p x
O 1 O: 01000:0000 3. Solve the unique T in thc
O O 1 : .
.
00100:0000 4. If T is nonsingular the
000.10:0000 ~igenvalues as those'of F.
al az a3 :
--------:-0- -i -:-: - - ------ - - -
000:01:0000
- .... - _1- _ .. _ _ .. 1_ .. ........ __
l x x] elgenvalues. A necessary co
7 lf JI, > JIz > JI3, the three nonzero rows of O become O 1 x. lf JII ;5;JIz ;5;JI3, they become controlI~ble and {F, K} obser
lOO] 001 . for the smgle-variable case (p =
O I O In our example, JIl = 3, JIz = 2, and JI3 = 4, and the three nonzero royvs ofO are Theorem 7-10; the proof will
[ O O 1 Hence the assertion will not bl:
r.~l
~ 0'1" o~~J A simple pattern can be deduced from these examples. B Notethat we have d(T)"'AT -TI
d(T) are ,litA) - JIAF).
STATE FEEDBACK 347
possible to obtain sorne other forms, for example, two blocks of companion
1, 1't l'S straightforward,
form on the diagonal or a block triangular form with two or three blocks of
d7 companion form on the diagonal. The characteristic polynomials of the
o O O (A + SK) in (7-47) are
O O O (S3 - a3 s2 - a2 s - al)(s2 - b 2 s - b¡)(S4 - C4S3 - C3S2 - C2S - Cl)
1 x O and S9 - d 9sS - dsS 7 - ... - d2s - d1
0- - Ó- - Ó
S=PB= O 1 O Since a¡, b¡, C¡, and di can be arbitrarily assigned, we have established once again
O--O--Ó Theorem 7-7.
O O O
Method 1I1
O O O
O O 1 We introduce a method of computing the feedback gain matrix without trans
(7-46) forming A into a controllable formo It will be achieved by solving a Lyapunov
equation. See Appendix F.
{A, B} is_said to be_i~ a
If u = r + Kx, where K lS a
Algorithm
Consider a controllable {A, B}, where A and B are, respectively, n x n and
n x p constant matrices. Find a K so that A + BK has a set of desired eigen
¡:e rows denoted by strings values.
se the three nonzero rows
1. Choose an arbitrary n x n matrix F which has no eigenvalues in common
enoted by string~ of ! _can
with those of A.
ose K so that (A +BK) is
2. Choose an arbitrary p x n matrix K such that {F, K} is observable.
3. Solve the unique T in the Lyapunov equation AT - TF = - BK.
0 0 : 0 0 00 4. Ir T is nonsingular, then we have K =KT- l , and A +BK has the same
00:000 0 eigenvalues as those of F. Ir T is singular, choose a different F or a different
, 1 0 : 0 0 00 K and repeat the process. I
:- 0- -i -:- 0- -Ó- -Ó - 0 We justify first the algorithm. Ir T is nonsingular, the Lyapunov equation
:00:100 O AT - TF = - BK implies
:-ó-o-~-ó--(-o--Ó
(A+BKT-l)T=TF or A+BK=TFT-l
:00:0010
Hence A + BK and F are similar and have the same set of eigenvalues. Since
: O O: O O O 1 F can be arbitrarily chosen so long as its eigenvalues are distinct from those of ¡le,
: d 4 d s : d 6 d7 ds d9
the eigenvalues of A + BK can almost be arbitrarily assigned.
" (7-47)
As discussed in Appendix F, if A and F have no common eigenvalues,s a
anion form on the diagonal; solution T always exists in AT - TF = - BK for any K and is unique. Ir A and
lÍon formo Of course, it is F have common eigenvalues, a solution T may or may not exist depending on
BK. To remove this uncertainty, we require A and F to have no common
eigenvalues. A necessary condition for T to be nonsingular is that {A, B}
.~ J
xl ir Il¡ ::;; 112 ::f 1l3' they become
controllable and {F, K} observable. The condition becomes sufficient as well
for the single-variable case (p = 1). The dual of this assertion will be proved in
Theorem 7-10; the proof will again be used in the proof of Theorem 7-11.
three nonzero roWS ofB are Hence the assertion will not be proved here.
8 Notethat we have d(T) =AT - TF rather than d(T) = AT +TF. Hence the eigenvalues of
mples. d(T) are A;(A) - Il}F).
348 STATE FEEDBACK AND STATE ESTlMATORS
x=[L_Ltjj] +[LJ]u
O
O
00:01
O O :-1 -2
x
O
O
O
1
(7-48)
(h) the responses of
[2 1 O -1 _
K¡ is roughly three tim
gains and yields smaller
Y = [1 -1 3: 2 O) x For a companion-for
is shown in Reference S
The equation is in a multivariable controllable formo The problem is to find a proportional to
K so that the eigenvalues of the resulting matrix are -1, -2 ±j, and -1 ± 2j.
We compute
where /Xlmax is its largest t
Ll(5) =5 5 +75 4 +245 3 +485 2 +555 +25
gains are also proportion
If we choose K as
then we have
-
K¡=
.
[-2
-25
O
-55
O
-48
-1
-23
-lJ
-5 10
1 O O
U j]
O
A +BK¡ = O
O
1
O
O
O
1
O
-55 -48 -24
If we choose K as
7 -s
1(2 =l-
r
O
-9
O O
-1
-4 -~J
then we have
..
A+BK2=[_~~ ~tJjj] O
__
00:01
. .. O O O :-5
, -2
This matrix has twó. blocks of companion formo One has eigenvalues
:.... 1, - 2 ±j; the "otherhas -1 ± 2j. These two blocks are noninteracting.
Now we are ready to compare these two K's.. The eIements of Kr , which
yields one block of companion form, are much larger in magnitude than the Figure 7-3 Transient respe
__________________ _ .:::::=--=.:.:.:.._:::..---=.__ ==-_ ..:..-::.:::=-__---::::...----::::..--==-::. ---=-=-=-- - ---------."
v ~ __ ~ _~ ,_~ ~, "_~ ~~
"j~u
[2 1 O -1 -2]. We see that the largest magnitude in transient in
1(1 is roughly three times larger than the one in 1(2' Hence 1(2 has smalier
(7-48)
gains and yields smaller magnitudes in transient and is preferred. I
1 O
) 1 F or a companion-form matrix of arder m, if all its eigenval ues are distinct, it
is shown in Reference 517 that the largest magnitude in transient is roughly
proportional to
The problem is to find ~
-1
·23
.-1]
-5
10
24
1
O
O j] O
-5
~2J:;t~~¿~5~~~4;====t"""'---i6---------- sec
-10 (a) K¡
,J]
, 1
-2
-5
and small transient, the order of the largest companion-form block should be and of eigenvalues
kept as small as possible. gains required in (7-5
The orders of companion-form blocks are determined by the lengths of shown by a computer
chains ofvectors b¡, Ab¡, ... , Akb¡ in (7-44). Ifthey are chosen by using Scheme 2 (7-50) is much much s
discussed in (7-21), the orders of companion-form blocks are equal to the We recapitulate t]
controllability indices of {A, B}. In this case, the largest order is equal to the formation of {A, B}
controllability index f1. If we use Scheme 1 discussed in (7-20) or any other Scheme 2 discussed i
scheme, the length of the longest chain, and consequently, the order of the [H AH ... An- 1 B
largest companion-form block will be larger than, at best equal to, f1. Hence in {Á, B} are equal to t
order to have small feedback gains and small transient, we shall use Scheme 2 the order ofthe larges1
to search the linearly independent columns in [B AB ... An-lB] and the among aH transforma!
orders of the resulting companion-forrn blocks are f11, f12, ... , f1 p , the con diagonal form with c<
trollability indices. this is not possible du
The introduction of state feedback can increase the order of a companion Problem 7-20 to choo
form block, as shown in Example 3, but can never decrease it. Hence in finding K =KP may yield a sy
a K:, we should preserve the orders of the original companion-form blocks. The design procedt
Even so, there are stiB a lot of flexibility in choosing K. We may choose K so single-input system; h<
that the resulting matrix is in block diagonal form or in block triangular form3. the method should not
Furthermore different grouping of desired eigenvalues will yield different K. However, the method 1
If we require the resulting matrix to be block diagonal and group the desired be discussed in Section
eigenvalues to minimize Up to this point, w
mas {(IXll max )!'! - \ (iX 2Imax)!'2 - \ ... , (IXplmax)!'r l} eigenvalues. This dep
where IXdmax is the largest eigenvalue in magnitud~ in the ith companion such as the rise time, se1
form block of order f1¡, then the flexibility bf choosing K is considerablyreduced. signals, and so forth. )
A problem may sometimes arise in preserving the orders of companion simpIe answer to the I
form blocks. For example, let f1¡ = 3, i = 1, 2, and let the desired eigenvalues simulations. Of cours
be - 3 ± 3j, - 4 ± 2j, and - 3 ±j. In this case, it is not possible to choose a not be unique. The 01
real Kso that A + BK preserve f1¡ and has a block diagonal or a block triangular consequently, a unique
formo Ifwe combine the two companion-form blocks into one, then the result formance index
ing A +BK is
J
I ~l
O 1 O O O
O O 1 O O
-3J O O
-4800
We see that the feedback gains are very large. _
O O O
-2804 -920 -180
and S4, or any other bo,
Assignment of eigenva
In this subsection, we ,
Now we use the procedure discussed in Problem 7-20 to find a K so that eigenvectors of the resu
the resulting Á +:DK is of the form assigned eigenvalues, Xi,
O 1 O O O O associated with Xi, that i
O O 1 O O O
-54 -36 -9 : -20 -8 -1
(A
- -
O
-
O
- -
O , O
,
- -
1
- - - - - - -
.
_1- _ _ _ _ _ _ _ _ _ _ _ _ _ _
O
(7-50) or BI
O O O , O O 1
which implies that
18 6 1 : -60 -44 -11
BK[e
STATE FEEDBACK 351
n-form block should be and 01' eigenvalues - 3 ± 3j, - 4 ± 2j, and - 3 ±j. We see that the feedback
gains required in (7-50) are much smaller than those in (7-49). lt can also be
nined by the lengths 01' shown by a computer simulation that the largest magnitude in the transient of
110sen by using Scheme 2 (7-50) is much much smaller than the one in (7-49). See Reference S17.
Ilocks are equal to the We recapitulate the preceding discussion in the following. In the trans
est order is equal to the formation 01' {A, B} into a multivariable controllable form, we should use
i in (7-20) or any other Scheme 2 discussed in (7-21) to search the linearly independent columns in
uent1y, the order 01' the [B AB ... An-l B]. The orders 01' companion-form blocks in the resulting
:st equal to, fl. Hence in {A, B} are equal to the controllability indices, fl¡, 1, 2, ... , p, 01' {A, B} and
t we shall use Scheme 2 the order 01' the largest companion-form block in {A, B} is the smallest possible
¡J .,. An-1B] and the among all transformations. We then choose a K so that (A +BK) is 01' block
ll, fl2' . . . , fl P'
the con diagonal form with companion-form blocks 01' order fli on the diagonal. lf
this is not possible due to complex eigenvalues, we may use the procedure in
e order 01' a companion Problem 7-20 to choose a K. This process 01' choosing K and consequently
ease it. Hence in finding K = KP may yield a system with small feedback gains and small transient.
companion-form blo_cks. The design procedure discussed in Method I transforms the problem into a
K. We may choose K so single-input system; hence the order 01' the companion-form block is n. Thus
In block triangular form!. the method should not be used to avoid large feedback gains and large transient.
les will yield different K. However, the method may require a smaller-dimensional state estimator as will
lal and group the desired be discussed in Sections 7-5 and 9-5.
Up to this point, we have not yet discussed how to choose a set 01' desired
eigenvalues. This depends highly on the performance criteria 01' the design
p\max)"p-l}
such as the rise time, settling time, overshoot, largest magnitude 01' the actuating
e in the ¡th companion
signals, and so forth. Even if these criteria are precisely specified, there is no
K is considerably reduced.
simple answer to the posed problem.One way to proceed is by computer
he orders 01' companion
simulations. 01' course, the set 01' eigenvalues obtained by this process will
~t the desired eigenvalues
not be unique. The only known systematic way 01' finding a unique K and,
. not possib1e to choose a
consequently, a unique set 01' eigenvalues is by minimizing the quadratic per
'onal or a block triangular
formance index
~s into one, then the result-
Unless all columns of the matrix on the right-hand side of (7-51) are inside the ~(s)
range space of B, no K will satisfy the equation. Hence the assignment of
eigenvectors cannot be entirely independent from the assignment of eigenvalues.
For a general matrix A, there is no way to predict what form an eigenvector
will assume. However if a matrix is of the companion forms shown in (7-47),
then its eigenvectors are of the form [1 Xi x¡ O O O O O O]',
[O O O 1 Xi O O O O]', or [O O O O O 1 Xi x¡ Xn for the
Figure 7-4 The effect '
left-hand-side matrix in (7-47) and [1 Xi Xl "5.r "5.t "5.( "5.P "5.? "5.~]' for
the right-hand-side matrix in (7-47) (see Problem 2-26). Hence for the {A, S}
in (7-46), after the eigenvalues are chosen, if the eigenvectors are chosen as which implies
discussed, then a solution K exists in (7-51). In fact, the solution is unique an¿
u(s) = (1- ~
can be solved by using the three nonzero rows of S in (7-46).
This process of choosing eigenvectors is far from an arbitrary assignment. The substitution of (7-5:
Once we choose the structure of A +SK (one, two, or three companion-form
y(s) = (N(,
blocks), and once we assign the eigenvalues for each block, the eigenvectors are
practically uniquely determined. Hence if the eigenvectors are chosen as Hence the transfer matr
discussed, there is no essential difference between the assignment of eigen
vectors and the assignment of structure of (Á + SK) discussed in the previous
Gf (
subsection. For further discussion of the assignment of eigenvectors, the reader
is referred to References S91, S125, S160. We see that the numera
state feedback. The col
Effect on the Numerator matrix of G(s) matrix of the denominó
any state feedback. Th(
In the single-variable case, we showed that the numerator of 9(s) is not affected assigned by a proper ch,
by any state feedback, whereas its denominator can be arbitrarily assigned. case.
Now we establish a similar statement for the multivariable case. Before pro
If G(s) is factored a1
ceeding, the reader should review the material in Section 6-6, especially Equa affected by any state feed
tions (6-100) through (6-117). ~tor of every element of
Once {A, B, C} is transformed into the form in (7-46), by reversing the G(s) wiJI be affected by st
procedure from (6-100) to (6-117), we can obtain feedback gain matrices I
C(s) =C(sIT - A)-l 10 +E = N(s)lD -1 (s) +E = (N(s) + ED(s))JD - j (5) differel1t traj].sient respon
we have E
>
-n
]-1 (7-51 )
A
of (7-51) are inside the + y(s)
mce the assignment of
+
ignment of eigenvalues.
1at form an eigenvector
forms shown in (7-47),
O O O O O O]',
1 X" F X3]' for the Figure 7-4 The effect of state feedback.
~ Xr' Ip Xi~ IP]' for
Hence for the {A, B}
which implies
nvectors are ehosen as
~ solution is unique ane o(s) = (1- KL(s)D-l(S))-l r (S)= D(s)(D(s)- KL(s)tlr(s) (7-52)
7-46). The substitution of (7-52) into y(s) = G(s)ñ(s) = (N(s) + ED(s))D- 1(s)íi(s) yields
n arbitrary assignment.
. three companion-form y(s) = (N(s) + ED(s))D-l(S)D(s)(D(s) - KL(s))-l r (S)
)ck, the eigenvectors are
Hence the transfer matrix from r(s) to y(s) is
lVectors are ehosen as
le assignment of eigen G¡(s) = (N(s) + ED(s))(D(s) - KL(s))-1
iscussed in the previous = (N(s) + ED(s))[DheH(s) + (DIe - K)L{s)] -1 (7-53)
i eigenvectors, the reader
We see that the numerator matrix N(s) +ED(s) of G¡(s) is not affected by any
state feedback. The column degrees of H(s) and the column-degree-coefficient
matrix of the denominator"matrix D¡(s) á D(;)-KL(s) are not affected by
any state feedback. The lower degree part of D ¡(s) however can be arbitrarily
.tor of g(s) is not affeeted assigned by a proper choice of K. This result is similar to the single-variable
be arbitrarily assigned. case.
riable case. Before pro If G(s) is factored as N(s)D -1(S), the numerator matrix N(s) will not be
ion 6-6, especially Equa affected by any state feedback. This, however, does not imply that the numer
ator of every element of G(s) will be unaffected. In fact, zeros of element of
(7-46), by reversing the G(s) will be affected by state feedback. This is one of the reasons why different
feedback gain matrices K, which yield the same set of poles, may yield vastly
different transient responses. 9
(s) +ED(s))D- 1 (s)
Computational problems
In this subsection, we diseuss sorne computational problems in Methods Il and
III for state feedback. The major step in Method Il is the transformation of
N(s) +ED(s) the numer {A, B} into a multivariable controllable formo The equivalence transformations
we plot G(s) as in Figure to achieve this are given in (7-14) and (7-45). They require the computation of
), then the state X(S)Ais, as AkB for k = 1, 2, ... , n - 1 and may change the problem into a less well-con
the state feedback u(s) = ditioned problem, as discussed in Section 5-8. Hence this procedure is generally
vn in Figure 14 From not satisfactory frorri a computer computational point of view,
'-I(S)O(S) 9 This observation was provided by Professor D. Z. Zheng ofTsinghua University, Beijing, China
- .• - ._- ----.. ". -_•. ,.•".- ·._·,u· •._._.•__ ••. ".~ •..• ·••· ·_.~L_ ~._.,, .. "_~._. ~ _._ .__ .__.._.. ~_ ~_ ..__ .__ ._._~ .~ .. __ . .__ _.__ ~~ __ .__._ . _._._. _ .._. __ ~~_ __.._.___ .__ .._ . _ M .,," ._ _ __._.__ ..
--~----~-------------~~~~~
In Section 5-8, we discussed a different method of transforming an equation state observer. In thi
into a controllable form without explicitIy computing AkB. The method is first full-dimensional and
to transform A into a Hessenberg form as in (5-89), and then transform it into dual to the design 01
a controllable form as in (5-93). (Note the different ordering of the state vari single-variable case \\
ables.) The first step is carried out by a sequence of orthogonal transforma special case of the mu
tions or gaussian eliminations with partial pivoting and is numerical1y stable. In this section we
The second step must be carried out by gaussian eliminations without any ofthe variable. For e
pivoting and is numerically unstable. Hence this transformation method is
again not satisfactory. Full-dimensional ~
In Section 6-6, we introduced an irreducible realization of G(s) = N(S)O-l(S) time-invariant dynam
in a controllable formo Hence, if N(S)O-l(S) can be computed from {A, B, C},
then a controlIable form dynamical equation can be obtained. In Reference
S17l, an algorithm is presented to compute a coprime N{S)O-l{S) from an
Hessenberg-form dynamical equation. However in the design, we still need where A, R, and e are,
the equivalence transformation. Thus it is not clear whether this design pro For simplicity, the dir
cedure will be satisfactory. In References S156 and S157 a dynamical equation assume now that the s
is first transformed into the Hessenberg form, the pole assignment is then state variables are nol
carried out by using the QR algorithm. The methods used in References S156 completely known. J
and S157 are alI numericalIy stable. Once a dynamical equation is transformed x(t) from the availablt
into a Hessenberg form, the Hyman method in Reference S2l2 can also be matrices A, B, and C.
employed to carry out the pole-assignment. For the 20 x 20 matrixon page 219 original system as shc
with e =0, the Hyman method yields a comparable result as the QR method. estimator. Now if the
See Reference S202. initial state and are dr
The procedure in Method 111 does nqt require the transformation of an will be equal to the act
equation into a controlIable formo It requires merely the solution of a Lyapunov is how to find the initi:
equation. The solution of Lyapunov equations has been extensively studied that state. This proble
in the literature, see, for example, References S9, S18, and Sl07. The algorithm dynamical equation Fj
in Reference Sl07 is claimed to be efficient and numericalIy stable and may be from its input and ou
employed in our designo an open-loop estimatOl
A limited comparison between Method n and Method nI in terms of There are, howevel
operational efficiency and numerical stability was carried out in Reference First, the initial state m
S202. Because of the complexities of the problem, no clear-cut concl usion was
reached in Reference S202.
lJ
ansforming an equation state observer. In this section we shal1 introduce two types of state estimators:
.kB. The method is first ful1-dimensional and reduced-dimensional. The design procedure is basical1y
d then transform it into dual to the design of state feedback; hence its díscussion will be brief. The
'dering of the state vari single-variable case will not be discussed separately and will be included as a
orthogonal transforma special case of the multivariable case.
Id is numerically stable. In this section we use the circumflex over a variable to denote an estimate
íminations without any x
of the variable. For example, is an estimate ofx; i is an estimate of X.
msformation method is
Full-dimensional state estimator. Consider the n-dimensional linear
:ion of G(s) = N(s)D-1(s) time-invariant dynamical equation
)mputed from {A, B, C}, FE: X=Ax +Bu (7-54a)
obtained. In Reference y=Cx (7-54b)
.me N(s)D - 1 (s) from an
the design, we still need where A, B, and C are, respectively, n x n, n x p, and q x n real constant matrices.
whether this design pro For simplicity, the direct transmission part has been assumed to be zero. We
57 a dynamical equation assume now that the state variables are not accessible. Note that although the
pole assignment is then state variables are not accessible, the matrices A, B, and C are assumed to be
; used in References Sl56 completely known. Hence the problem is that of estimating or generating
1equation is transformed x(t) from the available input u and the output y with the knowledge of the
erence S2t2 can also be matrices A, B, and C. If we know the matrices A and B, we can duplicate the
Ox 20 matrixon page 219 original system as shown in Figure 7-5. We called the system an open-Ioop
~esult as the QR method. estimator. Now if the original equation FE and the estimator have the same
initial state and are driven by the same input, the output x(t) of the estimator
the transformation of an will be eq.ual to the actual state x(t) for al1 t. Therefore, the remaining question
he solution ofa Lyapunov is how to find the initial state of FE and set the initial state of the estimator to
been extensively studied that state. This problem was solved in Section 5-4. It is shown there that if the
and St07. The algorithm dynamical equation FE is observable, the initial state of FE can be computed
~rically stable and may be from its input and output. Consequently, if the equation FE is observable,
an open-Ioop estimator can be used to generate the state vector.
Method IIIin terms of There are, however, two disadvantages in using an open-Ioop estimator.
carried out ih Reference First, the initial state must be computed and set each time we use the estimator.
) clear-cut conclusion was
This is very inconvenient. Second, and more seriously, if the matrix A has which can be written
eigenval ues with positive real parts, then even for a very smal1 difference between
x(t o) and x(to) at sorne to, which may be caused by disturbance or incorrect
estimation of the initial state, the difference between the actual x(t) and the
estimated x(t) will increase with time. Therefore an open-loop estimator is, in The asymptotic e;
general, not satisfactory. either by a block diag
Another possible way to generate the n-dimensional state vector is to
differentiate the output and the input n - 1 times. If the dynamical equation is
observable, then from u(t), y(t), and their derivatives, the state vector can be Clearly x is the error
computed (see Problem 5-17). However, pure differentiators are not easy to tracting (7-56) from (7
build. Furthermore, the estimated state might be severely distorted by noises
if pure differentiators are used.
We see from Figure 7-5 that although the input and the output of FE are
available, we use only the input in the open-loop estimator. It is conceivable If the eigenvalues of (A
that if both the output and input are utilized, the performance of an estimator error x can be control
can be improved. negative real parts sm
Consider the state estimator shown in Figure 7-6. The estimator is driven zero at rates faster thar
by the input as wel1 as the output of the original system. The output of FE, x(to) and x(t o) at initi2
Y=Cx, is compared with y ~ Cx, and their difference is used to serve as a the eigenvalues of A
correcting termo The difference of y and Cx, y -Cx, is multiplied by an n x q is much more desirabl
real constant matrix L and fed into the input of the integators of the estimator. In the fol1owing \
This estimator wil1 be cal1ed an asymptotic state estimator, for a reason to estimator.
be seen later.
The dynamical equation of the asymptotic state estimator shown in Figure Method I
7-6 is given by
In this method, we ap
i =Ax +Bu +L(y-Cx) (7-55 ) state estimators.
l
,------------------1
I + I
I ,---------
I
I
I
I
x
I I
I I
I I
I A I
I L
:
L- --.JI I
L _
Figure 7-6 An asymptotic state estimator. Figure 7-7 An asympt
STATE ESTlMATORS 357
,----------------------,
I I
I
I I
I I
I I ~
I
I
I
I
IL I
~
Proof
is observable, its state Define e ~ z-Tx
ltor Then we have
°
Ir the three conditions in (7-60) are met, then e(t) = eFte(O) -+ for any x(O), z(O),
and u(t). Hence z(t) is an estimate of Tx(t).
ned, provided complex Now we show the necessity of the conditions. Ir 3 is not met, then for
• x(O) =0 and u(t) =0, we have e(t)=eF1z(0)+0 as t-+oo. IfHi=TB, we can
find a u(t) to make e(t)+O as t-H/J. Ir TA - FT i= GC and if {A, B} is con
f {A, e} is observable, trollable, we can find a u(t) to generate a x(t) which makes e(t)+O as t-+ oo.
'assigned. Indeed, if This establishes the necessity part of the theorem. Q.E. D.
notes the transpose, is
genvalues, we can find With this theorem, we can now propose a design algorithm.
(Theorem 7-7). Since
le same, the theorem is Algorithm
edback can be used to 1. Choose an F so that all of its eigenvalues have negative real parts and are
not be repeated. We disjoint from those of A.
and small transient in 2. Choose a G so that {F, G} is controllable.
cks in the transformed 3. Solve the unique T in TA - FT = Gc.
tate-feedback part, the 4. Ir T is nonsingular, computeH =TB. The equation in (7-59) with these
F, G, and H is an estimate ofTx(t) or x(t)=T- 1 z(t). Ir T is singular, choose
different F and/or G and repeat the process. •
Proof
n real constant matrix T Let the characteristic polynomial of A be
O), and u(t) if and only if
Ll(s)=det(sI-A)=sn +a¡sn-¡'+azsn- z + ..'. +an
fl-60) °
Clearly we have Ll(A) = (Cayley-Hamilton theorem). Ir A¡ is aneigenvalue of
F, then Ll(AJ is an eigenvalue of Ll(F) (Problem 2-32). Since A and F have no
360 STATE FEEDBACK ANO STATE ESTIMATORS
common eigenvalue, we have Ó,(Ai) +0 for al1 eigenvalues Ai of F and dynamical equation
n
no means simple nUD
det Ó,(F) = TI ó,(},¡) + O It is not clear at presen
i= 1 efficiency and numeril
(Problem 2-22). Hence Ó,(F) is nonsingular.
The substitution of T A = FT + GC into TA 2 - F 2T yields Reduced-dimensio
T A2 - F 2T = (FT + GC)A - F 2T = F(TA - FT) + GCA = FGC + GCA Method I
Proceeding similarly, we can obtain the fol1owing set ofequalities: Consider the n-dimen~
TI-IT=O
TA-FT=GC
TA 2 - F 2T = GCA + FGC where A, B, and C are,
TA 3 - F 3 T = GCA 2 + FGCA + F 2GC
In this section we assUl
TAn-FnT=GCAn-1 +FGCAn-2 + ... +FI- 2GCA+F"- I GC
We multiply the first equation by !X n , the second equation by !X n - I , ... , the last
equation by 1, and then sum them up. After sorne manipulation, we finally where R is an (n -q) x
obtain as P is nonsingular. V
for T to be nonsingular. y =[
Ir q = 1, UF' A q, and V A are n X n square matrices. A q is always nonsingular.
Hence T is nonsingular if and only ifU F and VA are nonsingular or, equivalently, where XI consists of thl
{F, G} is control1able and {A, C} is observable. This establishes the theorem. A12 , A2!> and A22 are,
Q.E.D. (n -q) matrices;.B¡ and
that y = Xl ' Hence on
Although the unique solution T in TA - FT = GC can be computed from Consequehtly we need 01
(7-62), the methódmay not be desirable because it requires the .computation n-dimensional estimatol
of UF and VA (see Section S-S). The computational problem of TA - FT = GC Usiilg XI =Y,we wri
has been studied exterisivelyin the literature; see, for example, References S9,
S18, and S107.
The design procedure in Method 1 requires the transformation of a
STATE ESTIMATORS 361
GCA +F"-IGC
P~ [~J (7-64 )
on by C1. n - l " .. , the last
lanipulation, we finally
where R is an (n - q) x n real constant matrix and is entirely arbitrary so long
as P is nonsingular. We compute the inverse ofP as
(7-65 )
o
O
III=PQ=[~J[Ql Qz]=[~~: ~~:J=[~q ~-J (7-66)
lblem ofTA - FT = GC
Using Xl = y, we write (7-67a) as .
.J
362 STATE FEEDBACK AND STATE ESTIMATORS
We note that ü and w are functions of known signals u and y. Now if the This gives an estimate
dynamical equation in (7-68) is observable, an estimator of x 2 can be constructed. ~iag.ram ofthe (n - q)-d
In Figure 7-8.
Theorem 7 -11
A comparison betw
The pair {A, C} in Equation (7 -63) or, equivalently, the pair {A, C} in Equation estimators is in order.
(7-67) is observable if and only if the pair {A 22 , A 12 } in (7-68) is observable. • alence transformation d
required in the reduced
The controllability part of this theorem was implicitIy established in Section dimensional case. In th
5-8. See also Problem 5-34. Thus the proof ofthis theorem is left as an exercise. of integrators. In the I
Ir {A, C} is observable, then {A 22 , A¡2} is observable. Consequently, there the constant matrix Q t
.h ¡
exists an (n - q)-dimensional state estimator of x 2 of the form Wlt noises, the nbises VI
dimensional estimator t1
(7-69 ) noises in y will be sup~n
such that the eigenvalues of (A 22 - LA 12) can be arbitrarily assigned by a Method 11
proper choice ofL (Theorem 7-8). The substitution ofü and w into (7-69) yields
Consider the n-dimensior
(7-70)
(7-68b)
X=Q~=[Q¡ Q2J[L/+Z]=[Q¡ Q2J[~ I~-JGJ (7-74 )
itrarily assigned by a
Method 11
md w into (7-69) yields
Consider the n-dimensional dynamical equation
(7-70) FE: x=Ax +Bu (7-75a)
y=Cx (7-75b)
eliminated by defining
where Á, B, and C are, respectively, n x n, n x p, and q x n real constant matrices.
(7-71 ) 1t is assumed that FE is irreducible and rank C = q. Let
z=Fz+Gy +Hu (7-76)
-LAd{z + LX 1)
--~
~A12X2 - LB 1"
(7-73)
h
X
rm
The procedure of computing F, G, and H is similar to Method 11 of the full Now we consider tl
dimensional case. theorem by showing th;
Algorithm
1. Choose an (n - q) x (n - q) real constant matrix F so that al! of its eigen Define Il = Al VAf = LB]
values have negative real parts and are di~tinct from those of A.
2. Choose a G so that {F, G} is controllable, that is
rank [G FG ... Fn-q-1G]=n-q
P=[~J
1 Method II of the full- Now we consider the single-output case. We show the sufficiency of the
theorem by showing that r = O is the only possible solution in
C ] r-O (7-79)
[ UFAIV A
so that aH of its eigen ¡1n]', ar
Ithose of A.
l
~~ IlllXn-
2 =
1 . IX
IX n.- 2
n- 2
IX n.- 3
IXI
1
n-q .
. .. ..
is an (n - q) x n matrix. fJn 1 O O
where x denotes nonzero elements. From this equation, we see that if Cr = O,
then we have ~II=O. From UFA 1 VAr=UdJ=O, we have, by using ~n =0,
0-77)
" n- 1
Uf·A. V Ar= L fi¡F i - 1
G= L fiiFi- 1 G =0 (7-80)
eat the process. If P is j-= 1 i= I
ion (7 -76) is an estimate which together with the controllability of {F, G} imply Pi = O, i = 1, 2, ... , n - 1.
Hence (7-79) and rank UF = n -1 imply p = O.
Because Al is nonsingular, p=AIVAr=O implies VAr=O, which together
(7-78) with the observability of {A, C} imply r = O. This establishes that if {A, C} is
Theorem 7 -13
If the n-dimensional dynamical equation FE is observable, an (n -q)-dimen
sional state estimator, as given in (7-72)and (7-74)or in (7-76) and (7-78), with any
ollable. For the single-
desired eigenvalues (provided complex conjugate eigenvalues appear in pair)
can be constructed, where q is the rank of matrix C in FE. 11
present case. The only 7-5 Connection of State Feedback and State Estimator
Id UF is an (n - q) x nq
Consider the n-dimensional dynamical equation
FE: x=Ax +Bu (7-81 a)
(F)][U:qVJ y=Cx
where A, B, and C are, respectively, n x n, n x p, and q x n real constant matrices.
(7-81 b)
The eigenvalues of a
Ir the eigenvalues of F are chosen to have negative real parts, the estimated
transformation; henc
state x(t) will approach the actual state x(t) exponential1y.
system in Figure 7-9
The state feedback is designed with respect to the actual state of FE. Ir the
that the eigenvalues (
actual state x(t) is not available for feedback, it is natural to apply the feedback
as far as the eigenval i
gain K to the estimated state x(t), that is,
from the estimated st:
u =r +Kx (7-84) of state feedback ane
pendently and the eig'
as shown in Figure 7-9. Three questions may be raised in this connection:
feedback and those (
(1) In the state feedback" u = r + Kx, the eigenvalues ofthe resulting equation are
separation property.
given by the eigenvalues of A + BK. In the estimated state feedback u = r + Kx,
We discuss now Ü
do we stil1 have the same set of eigenvalues? (2) Will the eigenvalues of the
shown in (5-54). He]
state estimator be affected by the feedback u = r + Kx? (3) What is the effect
7-9 can be computed
of the estimator on the transfer-function matrix from r to y? To answer these
questions, we form the composite dynamical equation of the system in Figure
7-9 as as
[i] [Ai -
+BKQ1C
GC +HKQ1C
BKQ2
F +HKQ2
][x] [B]
z + H r
(7-85a)
(Theorem 5-16). Thi~
the use of a state estim
and does not appear
y =[C OJ [:] (7-85b) explanation. In the c,
to be zero. Consequ
This is obtained by the subsütution of (7-84) and (7-820) into (7-81) aad (¡-¿La). aH t. Hence, aS fu; do
difference whether a ~
Ir x(O) f- x(O), the
values of the estimate
parts of these eigenva
cause larger gains in
state feedback case.
mator will act as a dil
and large transient W.
simple answer to the
suggested in the liten
two or three times fas
ofthe negative real pa
Figure 7-9 Feedback rrom the estimated state. of those of A + BK. '
CONNECTION OF STATE FEEDBACK ANO STATE ESTIMATOR 367
(7-83)
y=[C OJ [:] (7-86b)
Example 1
Now we may apply th.
Consider the inverted pendulum problem studied in Exmple 2 on page 338.
Its dynamical equation is
l~ :
d-!---
~ I~- -~ -~l + r-! l
oJO J 5 -2
This completes the des
Functional estimat
an estimator of dimens
not necessary to recom
y = [1 : O O OJ x some functions of the
As discussed there, if we introduce the state feedback In these cases, the dim
In the following, we d
u(t) = r(t) + \j lf- ~ ~Jx kx, where k is an arbitr
then the resulting equation has eigenvalues -1, -2, and -1 ±j. functional estimator or
Now if the state x is not available for feedback, we must design a state esti Consider the dynaI
mator. We use Method 1 of the reduced-dimensional state estimator to design is to design an estimat
a three-dimensional estimator. Since the equation is already in the form of we transform (7-63) in
(7-67), the step of the equivalence transformation in (7-64) can be skipped.
Clearly we have
-1 y=
O
5 !] where x = Px with P ~
output w(t) will be cho:
Let us choose arbitrarily the eigenvalues of the estimator as - 3 and - 3 ± 2j.
Then we have to find a L such that the eigenvalues of A22 - LA l2 are - 3 and
- 3 ± 2j. For this problem, we shall solve L by brute force. Let L' = [11 /2 /3J,
where the prime denotes the transpose. Then we have where F is a m x m m
problem is to design Sl
A ~ A22 -L~~12 = =~~ r -6 ~] that w(t) will approach I
7-9, we can show inaé i
_ L-/3 5 O
and det(51 - A) = 53 +/152 -(5 +/2)5 - (/3 + 5/ 1 )
By equating its coefficients with those of
and a11 eigenvalues of F
(5 +3)(5 +3 +2j)(5 +3 -2j)=S3 +95 2 +315 +39 matrix, then z(t) appro
we obtain to meet
/2 = - 36
Hence a three-dimensional state estimator is, following (7-72) and (7-74), then w(t) approaches k,
lS a stochastic process, 1 O O
x= -36
r -84
O
O
Now we may apply the state feedback from
1
O
xas
mple 2 on page 338.
u= r + [~ Jf- \:1. -1f-Jx
This completes the designo I
and all eigenvalues of F have negative real parts, where T is a m x n real constant
ls +39 matrix, then z(t) approaches Tx(t) as t~ oo. Hence, ir M and N are designed
to meet
(1-88c)
- ¿]u 'o [lq O] can be replaced by [C, O], where C, is.aq x q nonsingularmatrix, and the subsequent
-2 development, with slight modification, still holds. This su bsection follows closely Reference SIlO.
370 STATE FEEDBACK AND STATE ESTIMATORS
Now we partition T as T = [T¡ T 2]' where T ¡ and T 2 are, respectively, The mq x mq ma trix
m x q and m x (n - q) matrices. Then (7-88a) can be written as, using = e Theorem 2-4 that, fe
[I q O], if and only if the mq
T¡A¡¡ +T 2A 2¡ -FT¡ =G (7-89a)
T¡A¡2 +T 2A 22 - FT 2 =0 (7-89b)
! 1 O Once T ¡ is known, v
O 1 N from (7-89a), (7-8:
JJ
estimator of kx. Th;
F= [
o o
-rx. m Theorem 7-14
M=[ O O O ] lf the dynamical eqt
then for any 1 x n r
where rx.i can be arbitrarily assigned. Note that {F, M} is observable. Let
functional estimator
tij,j = 1, 2, ... , m, be the jth row of Ti, i = 1, 2. Then because of the forms of F
approaches kx(t) exp
and M, (7-89d) and (7-89b) imply
t2 ¡ = k2 The dimension (lO
t ll A 12 +t 2¡A 22 =t 22
example, if n = 12 an
t 12 A 12 +t 22 A 22 =t 23 (7-90a)
index v could be ¡; =
then n - q = v - 1.
lf a p x n feedbac
where vis ap xl veCl
These equations, except the last, can be written as
7-2, then Kx can als.
t 2 ¡ =k 2 estimator. For a ge
t 22 =k 2A 22 +t¡¡Á¡2 eigenvalues of the fu
t 23 = k2Á~2 + t¡¡Á¡2 Á 22 + t¡2 Á ¡2 (7-9Ob) arbitrarily chosen. 1
lt mc~Y be possible í_(
In the design, if we ;
dimension of estimat
The substitution of these t 2j ,j = 1,2, ... ,m, into the last equation of (7-90a)
5131 and 5184.
yields, after sorne straightforward manipulation,
Id T 2 are, respectively, The rnq x rnq matrix in (7-91) is nonsingular for any (Xi' Thus we conclude from
written as, using = e Theorem 2-4 that, for any ka solution [t 1m t 1,m-1 . .. t l l ] exists in (7-91)
if and only if the mq x (n - q) matrix
(7-89a)
(7-89b)
V= A12
A1Z:A 22 J
are, respectively, 1 x q
r Á1zAz21
has rank n -q. Now {A, C} in (7-63) is observable if and only if {A 22 , Á 12 }
(7-89c)
is observable. Furthermore, the observability index of {A zz , A1Z } is v - 1,
(7-89d)
where vis the observability index of {A, C} (Problem 7-27). Hence, if m = v - 1,
then V has rank n - q, and for any k and any (Xi, a solution T 1 exists in (7-91),
Once T 1 is known, we compute T z from (7-9üb) and then compute G, H, and
N from (7-89a), (7-88b), and (7-89c). The resulting estimator is a functional
I,J
estimator of kx. This is stated as a theorem.
Theorem 7 -14
The dimension (v -1) could be much smaller than n -q for large q. For
example, if n = 12 and q = 4, then we have n - q = 8; whereas the observabili(y
(7-90a) 1;
index v could be =3 [see Equation (5-51)], and we have v -1 =2. If q = 1,
then n - q = v - 1.
lf a p x n feedback gain matrix K is of rank 1, it can be written as K = vk,
where v is a p x 1 vector and k is a 1 x n vector as the K z in the design in Figure
7-2, then Kx can also be generated by using a (v -1)-dimensional functiona1
estimator. For a general K, the situation is much more complicated. The
eigenvalues of the functional estimators discussed aboye are permitted to be
(7-9Ob) arhitrarily chosen. For a given set of eigenvalues, if they are chosen properly,
it may be possible to further reduce the dimension of a fl.lnctional estim;;üo;
\.ZZ+tl. m-
In the design, if we augument the matrix C in (7-63b) as [e' K'T, then the
1Á 1z
dimension of estimators may also be reduced. See References S92, S97, Sl10,
ast equation of (7-9üa) Sl31 and S184.
; given and (Xi are chosen. where u is the p x 1 input vector, y is the p x 1 output vector; A, B, and e are
...... _
_._- _._-_.... .... • • __ . '.• _~'_'._ • • • _ •••• _ .~n
••
: :_
._. :.~_.::..:.
_ _::..:_.~ •• . . . . _.-~._.
• •_ - - - . __ - __
• • •••• -:- _
n X n, n X p, and p X n real constant matrices, respectively. It is assumed that The transfer function
p :::; n. The transfer function of the system is
G(s)=C(sI-A)-lB (7-93)
We shall derive in th,
Clearly G(s) is a p X P rational-function matrixo If the system is initially in can be decoupled by s
the zero state, its inputs and outputs are related by d¡ ~ min (the differen
Y1 (s) = 911 (S)U1 (s) + gds)uz(s) + o. o + glP(S)Up(s) of each entr
yz(s) = gZl (S)U1 (s) + gds)uz(s) + . o o + gzp(s)uP(S) and the 1 X P constan
(7-94)
. Yp(s) =gp1 (S)U1(S) + gpz (s)uz (s) + ... +gpp(s)up(s)
where gij is the ijth element of G(s). We see from (7-94) that every input controls where G¡(s) is the ¡th
more than one output and that every output is controlled by more than one
i,lPUt. Because of this phenomenon, which is called coupling or interacting, Example 1
it is generally very difficult to control a multivariable systemo For example,
suppose we wish to control Y1 (s) without affecting yAs), Y3(S), .. o, Yp(s); the Consider
required inputs U1(S), uz(s), .. o, u/s) cannot be readily foundo Therefore, in
sorne cases we like to introduce sorne compensator so that a coupled multi
variable system may become decoupled in the sense that every input controls
only one output and every output is controlled by only one input. Conse
quently, a decoupled system can be considered as consisting of a set of inde
pendent single-variable systemso It is clear that if the transfer-function matrix
of a multivariable system is diagonal, then the system is decoupled. The differences in deg
Definition 7-1
A multivariable system is said to be decoupled if its transfer-function matrix is
The differences in deg
diagonal and nonsingular. I
is nonsingular.
~:~1
found. Therefore, in
that a coupled multi
Lt every input controls (;(s) = [ " "+; +2]
Iy one input. Conse
isting of a set of inde s2+~s+1 s2+ s +4
ansfer-function matrix
; decoupled. The differences in degree of the first row of (;(s) are l and 2; hence dI =0 and
El = lim s [
s'" 00 S2
s +2
+s +1 S2
l
+s +2
J= [1 OJ
sfer-function matrix is
• The differences in degree of the second row of (;(s) are 2 and 2, hence d 2 = 1 and
E2 = hm
•
s2 [ 1
~2----'---- 3J 11
lpling ofmultivariable s""oo s +2s +1
(7-95) Theorem 7 -1 5
A
i constant nonsingular A system. with the transfer-function rnairix G(s) can be decoupted by state
e feedback is shown in feedback of the form u = Kx + Hr if and only if the constant matrix
(7-96a)
(7-96b)
y
=>
is nonsingular.
•
We see from this theorem that wh~ther or not a system can be decoupled
is a property of its transfer-function matrix. The dynamical"equation descrip
tion comes into playa role onlywhen thegain matrix K is to be found. There
fore the controlIability and observability of the dynamical-equation description
of the system are immaterial here. Let (; ¡(s) and C¡ be the ith row of (;(s) and C,
.L
374 STATE FEEDBACK AND STATE ESTlMATORS
In the following, we shall establish the relations between the integer di, the
vector E¡, and the matrices C¡, A, and B. First, we expand (7-98) into (see
then
Prablem 2-39)
A 1 CBH=O, (
G¡(s) = Ll(s) Ci[sn-11+R1Sn- Z + ... + Rn-1]B and
1 Weestablish in the
= Ll( ) [CBsn- 1 + CR1Bsn- z + ... + CRd¡Bsn-d¡-l + ... + CRn-1B] E¡andE i.
s (7-99)
Theorem 7 -16
where
Ll(s) ~ det (si - A) ~ sn +CX1S n- 1 + ... +cx n For any K and any no
(7-100)
11 E¡ and C¡Rd¡B may dirrer by a constant because ó>(s) is factored out in (7-99). For convenience,
this diITerence, if ~ny, i~ neglected.
•. __ '
"_~'_W'_"._.'_' "~ •• J ••• •• _ _ •• _ _ , . ~ . •__ .~ __ ••• _ _ •• _"_ •• __ • __ ..
w_._~._.,, _··_"··~·_"'_ -_.. - . __.- .. _.~ '.'- ·...••••__ •. _.W __ ··_··_ _' .• ~ ,,_._ ~y . . . . _ •• _. __ ••••• __ • • • . • • • • • ~ " _ " _ •• - _ . _ . - - • • • •" ••• ~ .•••..• - - _ . . • ••• _ _ . _ . _. . -
+0: 11
(7-100)
Proof
It is easy to verify that, for each i, the conditions C¡B = O, C¡AB = O, ..
(7-101 )
C¡Ad¡-l B = Oimply that
. +O:n-l 1
CitA + BK)k =CiAk k = 0,1,2, ... , di (7-111 )
he numerator and the and CitA +BK)k =CiAd¡(A +BK)k-d¡ k =d¡ +l,d¡ +2, ... (7-112)
19) that the coefficients
the coefficient matrix Consequently, we have
have C(A +BK)kBH =0 k =0,1, ... , di~ 1
(7-102) and CitA +BK)d¡BH=CiAd¡BH =E¡H
(7-103)
Since H is nonsingular by assumption, if E i is nonzero, so is E¡H. Therefore
Iy that we conclude that di = di and Ei = KH. Q.E. D.
'unction matrix Gf in
(7-113)
in (7-99). For convenience,
376 STATE FEEDBACK ANO STATE ESTlMATORS
We show that if K = - E - 1F and H = E - 1, then the system can be decoupled e¡, we immediately obta.
and the transfer function of the decoupled system is decoupled by using K =
l/S(d1 + 1) O
or, equivalently,
1 1
G¡(s, -E- F,E- )=
l 6O 1/s (d2
O
+ 1) Although a system c
the resulting system is n
system are at the origino
to move these poles to ti-
If the dynamical equ~
values by state feedback.
(7-114)
the number of eigenvall
where e¡ is a row vector with 1 in the ith place and zeros elsewhere. First we reduced. For a complete
show that C¡(A +BK)d¡+l =0. From (7-105), (7-112) and (7-113) and using compute K and H to have
K= -E- 1F and E¡E- 1 =e¡, we obtain 36,41, 112, S218, and S2:
CiCA +BK)d¡+l =C¡Ad'(A +BK)=C¡Ad¡+l +C¡Ad¡BK
=F¡ -E¡E- 1F = F¡ -e¡F =0
7-7 Concluding Rf
where F¡ and E¡ are the ith row ofF and E, respectively. Hence we conclude that
In this chapter we studi,
CiCA +BK)d¡+k =0 for any positive integer k (7-115)
observability. We showe
Since di =d¡, Equation (7-106) reduces to is controllable, we can, t
eigenvalues of the resultin
1 - 1
C¡(sl - A - BK)-lBH ==- [CRd¡Bsn-d¡ - + C¡Rd¡+ 1Bs,,-d¡-2 + ... of controllable canonicaJ
. Ms) + Cít-1 BJH (7-116) state feedback, all the sta
asymptotic state estimatoJ
Now, from (7-105), (7-108), (7-112), (7-115), and the fact CiCA +BK)kB =0; for
observable, a state estima
k =0,1, ... , d¡ -1, which follows from (7-109) and the nonsingularity of H,
structed. The constructic
it is straightforward to verify that
servable canonical-form d)
Cj~.d¡B =C¡(A + BK)d¡B = C¡Ad¡B = E¡ TA-FT=GC.
CR d¡+ 1B= C[(A + BK)d¡ +l + !Y. 1(A+ BK)d¡JB = &l E ¡ We studied in this ch,
Section 7-2, 7-3, and 7-4 c,
equation is instantaneousl
trollability assumption, th
Consequently, Equation (7-116) becomes
dynamical equation has a fl
C¡(sl-A-BK)-lBH=L\~S)[sn-d¡-l +&lSn-d¡-2 + ... +&n-1-dJE¡H (7-117)
transformation, a controll
obtained. Consequently
remarks apply to the obser"
What is left to be shown is that
We also studied the der..
L\(s)=sn +éi 1s n- 1 +&2Sn-2 + ... +&n state feedback. The cone
= Sd¡ +l(sn-d¡-l + &¡sn-d¡- 2+ ... + &n-d¡- ¡) (7-118) required in this part of Stl
decoupling also used the st.
From the Cayley-Hamiltonthebrem, we have The combination of stat.
(A +BK)n +&¡(A +BKr 1 + ... +&nl =0 (7-119) in Section 9-5 directly fram t
the state-variable approad
By multiplying CiCA +BK)d¡ to(7-119) and using (7-115), we can obtain discussed there.
&nC¡(A+BK)d¡=O which implies that &n=O. Next by multiplying The introduction of con
CiCA + BK)d¡ -1, we can show that&" _1= O. Proceeding similarly, we can prove constant gain output feedbac
Equation (7-118). By substituting (7-118) into (7-117) and using E¡H = E¡E- 1 = cannot arbitrarily assign all
---:~---~--~~-----~------~. ------~----------.------------- - ~
stem can be decoupled e¡, we immediately obtain Equation (7-114). Consequently, the system can be
decoupled by using K = - E- I F and H = E-l. Q.E.D.
In other words, the estimated x(k + 1) is reconstructed from {y(O), y(l), ... , y(k)}.
If x(k + 1) is reconstructed from {y(O), y(1), ... , y(k), y(k + 1)} and modeled as
x(k + 1) = Ax(k) +Bu(k) + L(y(k + 1) -Cx(k + 1))
then the situation wil1 be different. The reader is referred to Reference S215.
Figure P7-7
Problems
7-8 Let
7-1 The equivalence transformatíon Q ofi = Px = Q - IX in Theorem 7-1 can be obtained
either by using Q = [ql ql ... qn], where the q¡'s are given in (7-9), or by using Q =
UÜ- l , where U is the controllabílíty matrix of FE I and Ü- I is given in (7-13). From a
computatíonal point of view, which one is easier to use? Fínd the values of k = [k l I
Fínd the values of h such thé
7-2 Consider the controllable dynamical equatíon concIude that for almost all k
7-11
7-4 Transform the equatíon
if~ =~ ]x+[~}
.
x= [-~~
y= [1 1 O]x
into the controllable canonical-form dynamical equation. What is its transfer functíon?
whích has an unstable eígenvaJw
7-5 Let f.l.l be the largest integer such that {b[, Ab[, ... , A~l - Ib l } is a linearly independent subequation associated w¡'th .
seto Let f.l.l be the largest integer such that {b[, Ab[," .. , A~l -1 b[, b1 , Ab 1 , " .. , A~2 -1 b1 } st a bI"l" Ize the equatl'on by " elge
. usmg sto
is a Iinearly independent seto Show that Anb 1 for all n ¿f.l.l is linearly dependent on closed-Ioop equatíon has eigen~a':
{b[, ... , A~l -Ib[, b1 , ... , A~2 -·1 b1 }.
7~12 Gíven
7-6 Transform the dynamical equation in Problem 7-4 into the observable canonical
form dynamical equation.
PROBLEMS 379
gain output feedback 7-7 Consider the state feedback and the output feedback systems shown in Figure P7-7.
B and q = rank C. In Show that for any constant matrix H, there exists a constant matrix K such that Kx = HCx.
ed. Furthermore, the Under what condition on C will there exist a matrix H such that K = HC for any K? (C, K
See References S31, and H are q x n, p x n, mul p x q constant matrices, respectively. It is general1y assumed
that n¿q.) (Answer: A solution H exists in K=HC for a given K if and only if p [~J=Pc.
results in tbis cbapter
tbe discrete-time case. Consequently, a solution H exists in K = HC for any K if and only if e is square and non
singular.)
.Id be modeled as
Ix(k))
Ilom {y(O), y(l), ... , y(k)}.
I + 1)} and modeled as
I:x(k + 1))
\red to Reference S215. Figure P7-7
7-8 Let
7-9 Use state feedback to transfer the eigenvalues of the dynamical equation in Problem
7-4 to - 1, - 2, and - 2. Oraw a block diagram for the dynamical equation in Problem 7-4
and then add the required state feedback.
tat is its equivalent dynamical
1'lVhat is its transfer functíon? which has an unstable eigenvalue 1. The dynamical equation is not controllable, but the
subequation associated' with eigenvalue 1 is control1able. Do you think it is possible to
1, -1 b l } is a linearly independent stabilize the equation by using state feedback? Ir yes, find the gain vector k such that the
IA",-Ib¡, bz, Ab z, ... . '. A"' -1 b z} closed-loop equation has eigenvalues - 1, - 1, - 2, - 2, and - 2.
¿!lz is linearly dependent on
7-12 Given
Find the gain vector [k l k 2 ] such that the state-feedback system has - 1 and - 2 as its
7-21 Verify that the ma
eigenvalues. Compute k l , k 2 directly without using any equivalence transformation.
7-22 Let
7-13 Consider the uncontrollable state equation
1 O
2 O
O -1
O O
Is it possible to find a gain vector k such that the equation with state feedback u = kx + r Find two different real co
has eigenvalues -2, -2, -1, -1? Is it possible to have eigenvalues -2, -2, -2, -1? values - 4 ± 3j and - 5 ±
Howabout -2, -2, -2, -2? (Answers: Yes; yes; no.)
7-23 Show that the state
7-14 The observability of a dynamical equation is not invariant under any state feedback.
Where does the argument fail if the argument used in the proof of Theorem 7-3 is used to 7-24 Show that the (11 - <:¡
prove the observability part? 7-9.
7-15 Find a three-dimensional state estimator with eigenvalues - 2, - 2, and - 3 for the 7-25 Establish the sepan
dynamical equation in Problem 7-4. Use two different methods.
7-26 Consider the dynarr
7-16 Find a two-dimensional state estimator with eigenvalues - 2 and - 3 for the dynam
ical equation in Problem 7-4. Use two different methods.
7-18 In Problem 7-17, if the state is not available for feedback, can you carry out the
design by using a state estimator? Choose the eigenvalues ofthe estimator as -4 and - 4. 7-27 Prove Theorem 7-11.
What is its overall transfer function? observability índices of (A,
l _~
.~.
i=l
PROBLEMS 381
:m has -1 and -2 as it, 7-21 Verify thatthe matrix in (7-50) has eigenvalues - 3 ± ~i, - 3 ±j and - 4 ± 2j.
i1ence transformation.
7-22 Let
A= [~ ¿ ~ ~l
-3 1 2 3
2100
~~ ~l
B= 1 2
02
Find two different real constant 2 x 4 matrices K such that the matrix (A + BK) has eigen
1 state feedback u = kx + r
values -4±3j and -5 ±4j.
lValues -2, -2, -2, -1?
7-23 Show that the state estimator in Theorem 7-8 meets Theorem 7-9.
t under any state feedback.
7-24 Show that the (11 - q)-dimensional state estimator in (7-72) and (7-74)meets Theorem
¡ of Theorem 7-3 is used to
7-9.
111 - 1
and note that if linearly independent rows of the observability matrices are searched in .
TI (s -A¡) order from top to bottom, the locations of the linearly independent rows of both observ
¡= 1 ability matrices are identica!.)
~
52 +1
y=[~ ~J x
-1
2
be decoupled by state feedback?
-1 -2
x(k +1)= ~ -1
[ O
y(k) = [ 1 1
Find a feedback gain k so that all the eigenvalues of the resulting equation are zeros.
Show that, for any initial sta te, the zero-input response of the equation becomes identicany
zero for k "2.3. This is caBed a dead bea! control. See Problem 2-50.
7-32 Show that the controllability and observability indices of {A + BHC, B, C} are the
same as those of.{A, B, C} for any H. Consequently, the controllability and observability
indices are invariant under any constant gain output feedback.
7-33 Given
~::':=:-~:_:;'--=-==:":::="===;::_'_'_" ' "'__ ~'L _ _ ' _ •• _ . .~ •_ _• • • ._ . _•• _ ~ __ _ __ ~ • • • _
._-----_._-----------~-----,-_._._-~----~._~-----~-~.--,------"
PROBLEMS 383
ion
ete-time case.
8-2 Stabilityel
Descri ption
Time-varying case
systems with only on
shown in Chapter 3 t
y(t) =
Stability of Linear Systems
where g(t, T) is the im
output measured at tir
In the qualitative ~
perhaps the only gues
properties, under wha
For example, if the inp
\u(t
under what condition e
output y satisfies
8-1 Introduction
for allt in ( - 00, oo)?
Controllability and observability, introduced in Chapter 5, are two important
gualitative properties of linear systems. In this chapter we shall introduce
another qualitative property of systems-namely, stability. The concept of
stability is extremely important, because almost every workable system is
designed to be stable. Ir a system is not stable, it is usually ofno use in practice. does there exist a consí.
We have introduced the input-output description and the dynamical-equation
description of systems; hence it is natural to study stability in terms of these
two descriptions separately. In Section 8-2 the bounded-input-bounded
output (BIBO) stability of systems is introduced in terms of the input-output If the bpul approache
description. We show that if the impulse response of a system is absolutely output approach anotl
integrable, then the system is BIBO stable. The stability condition in terms of input approaches a cons
rational transfer functions is also given in this section. In Section 8-3 we ing to these various pro
introduce the Routh-Hurwitz criterion which can be used to check whether or for the system. We sha;
not all the roots of a polynomial have negative real parts. We study in Section used in linear systems: t
8-4 the stability ofa system in terms ofthe state-variable description. We intro Recall that the inpl
duce there the concepts of equilibrium state, stability in the sense of Lyapunov. when the system is initia
asymptotic stability, and total stability. The relationships between these relaxed system. Hence t
concep.ts.are established. In Section 8-5, we introduce a Lyapunov theorem description is applicable
and then use it to establish the Routh-Hurwitz criterion. In the last section,
the stability of linear discrete-time systems is discussed. Definition 8-1
The references for this chapter are 58; 59, 65, 76, 90, 102, 116, S79, S 111,
S135, and S153. A relaxed system is said
if and only if for any bOL
384
J
STABILITY CRITERIA IN TERMS OF THE INPUT-OUTPUT DESCRIPTION 385
s where g(t, "C) is the impulse response of the system and is, by definition, the
output measured at time t due to an impulse function input applied at time "c.
In the qualitative study of a system from the input and output terminals,
perhaps the only question which can be asked is that if the input has certain
properties, under what condition will the output have the same properties?
For example, if the input u is bounded, that is,
under what condition on the system does there exist a constant kz such that the
output y satisfies
ly(t)I.:::; kz < OC!
Definitíon 8.-1.
:d.
We illustrate the importance of the qualification "relaxedness" by showing output. Let us choo
that a system, being BIBü stable under the relaxedness assumption, may not
be BIBü stable if it is not initially relaxed.
where
Example 1
Consider the network shown in Figure 8-1. lf the system is initially relaxed,
that is, the initial voltage across the capacitor is zero, then y(t) = u(t)/2 for all t.
Therefore, for any bounded input, the output is also bounded. However, if Clearly u is bounded.
the ·initial voltage across the capacitor is not zero, because of the negative
capacitance, the output will increase to infinity even if no input is applied. I
Proof G(t,
Sujficiency: Let u be an arbitrary input and let lu(t)1 =:;;, k l for all t in ( - 00, 00).
Then
Then 9ii is the impul~
ly(t)1 = \ Loo g(t, r)u(r) dr I=:;;, Loo Ig(t, r)llu(r)i dr =:;;, k l Loolg(t, r)1 dr =:;;, kk l
output terminal. Sim
system is defined to be
for all t in ( - 00, 00). Necessity: A rigorous proof of this part is rather involved. the output vector is ba
ponent of the vector
We shall exhibit the basic idea by showing that if
pair of input and outp
number of bounded fu
Loo Ig(t, r)1 dr = 00
theorem.
for sorne t, say tI' then we can find a bounded input that excites an unbounded
Theorem 8-2
A relaxed multivariable
In In
¡y
~
+
u"-'
-i F
is BIBü stable ir and e
In
entry of G,
In
Figure 8-1 A system whose output is bounded for any bounded input if the initial
voltage across the capacitor is zero. for all t in ( - 00, (0).
STABILlTY CRITERIA IN TERMS Of THE INPUT-OUTPUT DESCRIPTION 387
={ ~
if x =0
sgn x if x>O
em is initially relaxed, -1 if x <O
en y(t) =u(t)/2 for all t.
Clearly u is bounded. However, the output excited by this input,
Jounded. However, if
ecause of the negative
o input is applied. •
y(t 1 ) = Loo g(t 1 , r)u(r) dr = Loo \g(t 1, r)1 dr = 00
Then gij is the impulse response between the jth input terminal and the ith
output terminal. Similar to single-variable systems, a relaxed multivariable
l [00 Ig(t, r)i dr ~ kk 1 system is defined to be BIBO stable if and only if for any bounded-input vector,
the output vector is bounded. By a bounded vector, we mean that every com
lis part is rather involved. ponent of the vector is bounded. Applying Theorem 8-1 to every possible
pair of input and output terminals, and using the fact that the sum of a finite
number of bounded functions is bounded, we have immediately the following
theorem.
._-
where g(t) is the impulse response of the system. Recall that in order to have a
description of the form (8-7), the input-output pairs of the system must satisfy
linearity, causality, and time-invariance properties. In addition, the system is
assumed to be relaxed at t = O.
234
Corollary 8-1
A relaxed single-variable system which is described by
I Ig(t, L)I do = I Ig(t - L)I dL = I Ig(o:)1 do: s 100 Ig(o:)1 do: (8-8) Consider a relaxed si!
related by
For the time-invariant case, the initial time is chosen at = O; hence the inte
°
gration in (8-8) starts [rom instead of - oo.
t
Q.E.D.
4
-1
n +(. t - n)n
for - 3 < (t -n)sO
f(t-n)= .
n
{ is nondecreasing as ex increa
1
n -(t - n)n 4 forOs(t-n)s3
n
for n = 2,3,4, . . .. The function .r
is depicted in Figure 8-2. lt is easy to Consequently,
verify that f is absolutely integrable; however, f is neither bounded on [0, (0)
nor approaching zero ast -> oo.
STABILlTY CRITERIA IN TERMS OF THE INPUT-OUTPUT DESCRIPTION 389
Id, (8-7)
n
that in order to have a
he system must satisfy
addition, the system is
II
2 3 4
Figure 8-2 An absolutely integrable function that is neither bounded nor tending to
zero as (---+00.
(8-9)
as a ----+ oo. We shall use this faet in the proof of the following theorem.
at Theorem 8-3
Consider a relaxed single-variable system whose input u and output y are
100 Ig(a)1 da (8-8)
related by
it t = O; henee the inte
Q.E.D. y(t) = J~ g(t - ,)u(,) d,
0, (0) if
l Since
Ig(/)1 dI
-n)~O I Ig(t)1 d(
liro
a- 00
Jao Ig(/)\ dt ..... J.'o" 19(1)1 dt
gure 8-2. ltis easy to Conseq uen tI y.
ither bounded on [0, (0)
390 STABILITY OF LINEAR SYSTEMS STABIl
for sorne constant k, then we have the following: Let UM ~ rnax lu(t)
I
1. If u is a periodic function with period T -that is, u(t) = u(t + T) for all
t ~O-then the output y tends to a periodic function with the sarne period
(not necessarily of the sarne waveforrn).
Ir g(r)u(
2. If u is bounded and tends to a constant, then the output will tend to a constant. which appraaches .
3. Z If u is of finite energy, that is, ciently large, (8-12)
then the output is also of finite energy; that is, there exists a finite kz that
for all t ~ t l. As car:
depends on k l such that
is approxirnately eq
becornes, for all t ~
(IX> [y(t)i2 dt}/z =::; kz < 00
Proof
1. We shall show that if u(t) = u(t + T) for all t ~O, then y(t)-4 y(t + T) as t -4 oo. which is independen
It is c1ear that 3. It can be shown tha
with the property
y(t) = J>(r)u(t - r) dr (8-10)
and
r + l T
r,+ T for sorne constant k
y(t+T)= Jo g(r)u(t+T-r)dr= Jo g(r)u(t-r)dr (8-11 ) space. In this space
(see Prablern 2-49):
Subtracting (8-10) frarn (8-11), we obtain
IIJII ~ (
Ir+
T
¡y(t +T) - y(t)1 = g(r)u(t -r)drl
The Schwartz inequa
r'+T rl + T
=::; J, ig(r)llu(t -r)\ dr =::;UM J. Ig(-r)\ dr
~ 1"' ü'-'
11 J(t)gi
where UM ~ rnax lu(t)l. It follows frarn (8-9) that Iy(t) - y(t + T)i-4 O as
OSIST
t-4 00, or y(t)-4 y(t + T) as t -4 oo.
2 This can be eXlended lo as fol1ows: For any real number p in [1. co]. if . ,
(J '"
g (r)
)I/P
o lu{t)lpdt ::;;k¡<co lJl -¡-\- .,..-__
then there exists a finite k 2 .such that
y(t) =
1',
o
g('1:)u(t - '1:) d'1: + J' g('1:)u(t - '1:) d'1:
'1
(8-12 )
Let UM ~ max
, Iu(t)l· Then we have, with t > t b
which approaches zero as t 1 -'> 00, following from (8-9). Hence if t 1 is suffi.
r Ig('1:)\ d'1:
: exists a finite k'}. that for aH t 2: t ,. As can be seen from Figure 8-3, if t is m uch larger than t l, u(t - '1:)
is approximately equal to a constant, say (f., for aH '1: in [0, t ,]. Hence (8-13)
becomes, for all t ~ t 1 ~ 0,
for some constant k forms a linear space over IR, It is called the L'}. function
space. In this space we may define a norm and an inner product as follows
)u(t - -c) d'1: (8-11 )
(see Problem 2-49):
r)\ d'1:
I[} f(t)g(t) dt Is (t
dl
]. ir u (t - T)
tY:+-:~-....,....----------
---f-::--~----;I'''''''''--t------~-------I-_T
i
392 STABlLITY Of LINEAR SYSTEMS STABlI
With these preliminaries, we are ready to proceed with the proof. Consider lf
Proof
Since sin (t - r) = sin t
Consider now
y(t) = I g(r)u(t-r)dr
too.1u(t - r)12 dt ~k 1
Hence, from (8-9) we (
~kk1
By definition, g(s) is th
too ly(t)12 dt too Ig(r)1 dr =k 2k 1
We see from Corollary 8-1 and Theorem 8-3 that if a system that is describ g(iw) = roo g(¡
able by Jo
is BIBO stable, then if the input has certain property, the output will have the
same property. For the time-varying case, this is not necessarily true. The
following is a very important corollary of Theorem 8-3.
Hence (8-17) becomes
Corollary 8-3 y(t) -> sin wt (R
Consider a relaxed single-variable system whose input u and output y are where () = tan -1 (1m g(i,
related by
.",
Proof
ig('r)l[u(t _,)1 2 d,
o Since sin (t - ,) = sin t cos ¡; - cos I sin ¡;, we have
y(t) = t g(,)u(t - ¡;) d¡; = I g(,)[ sin wt cos w¡; - cos WI sin w¡;] d¡;
d¡; dI
=sinwt to g(,)cosw,d¡;-coswt t'" g(¡;)sinw¡;d¡;- f" g(¡;)sinw(t-¡;)d-c
(8-15)
It is clear that
ly assumption, u is of Ir g(,) sin w(t - ,) d¡; \::; r [g(¡;)[[sin w(t - ')1 d,::; r Ig(,)1 d¡;
y(t)---> ~in wt t'" g(,) cos w, d¡; - COS WI t'" g(,) sin w'"[(h (8-17)
inusoidal output, the is BIBO stable if and only if there exists a finite number k such that, for every
ilency can be read out entry of G,
: transfer function of a
[' \gJt)\ dt .:s;k< 00
'Y transfer functions;
; of transfer functions. If the Laplace transform of G(t), denoted by G(s), is a proper rational
mdition can be easily functíon matrix, then the BIBO stability condition can also be stated in terms
, then the situation is of G(s).
Theorem 8-5
A relaxed multivariable system that is described by y(s) = G(s)u(s), where G(s)
oper rational function is a proper rational-function matrix, is BIBO stable if and only if all'the poles
e in the open left-half of every entry of G(s) have negative real parts. I
~ real parts. •
Corollary 8-2 and Theorem 8-S follow immediately from Theorems 8-2
s plane excluding the and 8-4 if we consider every entry of G as the impulse response of a certain
input-output pair.
s plane is the left-half
:y of a system is inde
8-3 Routh-Hurwitz Criterion
= ao n
k
(s +ctd n
j
(S2 + 2f3 jS + f3; +yJ) (S-20)
_. _.. . .._ ... __, . . _ .. _.J_ ..__. .. .__ . .. . . . .
~
~
.
~
.
,
.
~
.
~
_
~
_
.
_
.
.
.
,
.
.
.
-
~
-
-
.
~
_
=
-
-
-
-
.
-
-
-
~
-
-
_
_
=
:
:
:
:
.
:
_
.
:
:
_
.
.
.
.
:
=
=
_
=
:
=
=
=
'
~
=
:
:
:
:
:
:
_
_
:
:
-
-
-
'
:
.
:
-
-
396 STABILITY Of LINEAR SYSTEMS
where Cf.k > 0, fJ j > Ú, and iZ = - 1. Since aH the coefficients of the factors in
the right-hand side of (S-20) are positive, we conclude that if D(s) is a H urwitz
polynomial, its coefficients ai> i = 1, 2, ... ,n must be al! positive. Hence given a
polynomial with a positive leading coefficient, if some oIits coefficients are negative
or zero then the polynomial is not a Hurwitz polynomial. The condition that aH
coefficients of a polynomial be positive is only a necessary condition for the
polynomial to be Hurwitz. A polynomial with positive coefficients may still Since the four numbe
not be a Hurwitz polynomial; for example, the polynomial with positive co conclude from Theoren
efficients .
In order to ohtain o
S3 + SZ + lis + 51 = (s + 3)(s - 1 +4i)(s -1 - 4i) These processes can b{
is not a Hurwitz polynomial. an even number and
Consider the polynomial D(s) given in (S-19). We form the fol!owing with n' ~ n/2,
polynomials
Do(s) =aos n +azsn - Z + ... (S-21a)
D1(s) = als n - 1 +a3s"- 3 + o •• (S-21b)
Note that the number
that is, if n is even, Do(s) consists of the even part of D(s) and D1(s) consists of is even; they are equa
the odd part of D(s); if n is odd, Do(s) consists of the odd part of D(s) and D1(s) shown in Table S-1.
consists of the even part of D(s). Observe that the degree of Do(s) is always one two rows are just the
degree higher than that of D1(s). Now we expand Do(s)/D1(s) in the foHowing remaining rows are 01
Stieljes continued fraetion expansion:
1
1
Cf.ZS +-------------
1 (S-22)
Cf.3 S +---------
Table S.1 Continuous-
s
11-1
a~l a\l)
-----~---
This theorem will be proved in Section S-5. Here we give an example to a\j'-2l a·1(II-2)t,
.
illustrate its application.
Si a\j'-ll:
Exarilple 1
-'--------:
SO a~l:
,
Consider D(s) = S4 + 2s + 6s z + 4s + 1.
3 We have Do(s) = S4 + 6s" + 1, D1(s) =
2s 3 +4s,and t Ir n is odd, the first two ro
the pattern is identical.
ROUTH-HURWITZ CRITERION 397
ients of the factors in
at if D(s) is a Hurwitz Do(s) 1 1 ¡
--=-s+ 3 ="2s+----
)sitive. Hence given a D¡(s) 2 2s +4s . 1
oejJicients are negative
ts+--- (8-23)
4s 2 + 1 .B. 1
The condition that al! 7 S +"7
"2s
;ary condition for the
: coefficients may still Since the four numbers 1X¡=~,1X2=~'(.(3=~' and 1X 4 are all positive, we =f
nial with positive co conclude from Theorem 8-6 that the polynomial D(s) is a Hurwitz polynomial. I
(8-22)
----------------- ----
,
I if and only if the n 5"-3 aff) a\3) a~3) a(3)
n'- 2
,
I
,
'e give an exampIe to a(M-Z) ,
1 ,
, ag'-Z)
(n-l)1
ao , 0("_1 =-'
ag'-I)
Cf.,,=~)
ao
.t Ir n is odd, the first two rows have the same number of coefficients. Otberwise,
the pattern is idenlical.
-- _ _- _._---~-._~_ .. .
_-------_.~--_._---_._-------_ .. _-----_._--_.~-_._-_
~ .. _ - - - - - - - - - - - - - -
_--_._--_._----_._._------_._---._--_._--------"~-.~._._------------------_._--._-_._-.-._--
----_.~--~-_ ... ~._-_._-_
which implies
AH the coefficients ar'
(8-27) Example 3
Consider the polynor
form
This equation holds for k=O, 1, ... , n-2. For k=n-1, we haveDn(s)jDn-¡(s)=
a~)ja~-¡)s = 1j(J.ns. Using (8-27), we can write Do(s)jD¡(s) into the form in (8-22).
This establishes that the (J.¡ in (8-22) can indeed be computed from the coefficients
in the first column of Table 8-1 as shown.
Proof
The assumption ao = abO) > Oin (8-19) and the relations (J.¡ = ag - 1 )jag), i = 1,2, ... ,
n, imply (J.¡ > O, i= 1, 2, ... , n if and only if ag) > O, i = 1,2, ... , n. Hence the
A negative number ap;
first part of the theorem foHows directly from Theorem 8-6. To show the
polynomial.
second part, we show that aW > O, i = 1, 2, ... , n if and only if aH the coefficients
in Table 8-1 are positive. The sufficient part is obvious. To show the necessary
We note that in co
part, we write (8-25) as
- are not áffecüid if a ro'
a(k) - a(k, +2) + (J.k +1 a(k1 +1+1)
1+1 - the cbmputation of Ta
ROUTH-HURWITZ CRITERION 399
Example 2
Consider the polynomial in Example 1. We form
1 6 1
\-26), 2 4
4 1
3.5
1
(8-27) Example 3
Consider the polynomial D(5) = 35 7 +25 6 + 25 5 + 54 + 35 3 + 52 + 1.55 + 1. We
form
e haveD,,(s)/D,,-¡(s)=
into the form in (8-22). 3 2 3 1.5
:d from the eoeffieients 2 1 1 1
1 1 O
2 2
The response of the state equation (8-28a) can always be decomposed into A state Xe of a dynam
the zero-input response and the zero-state response as only if
y(t) = l'
lO
C(t)<Il(t, r)B(r)u(r) d'l:
definition, an equilibr
xe =
= l'
lO
G(t,r)u(r)dr for all t 2. too Clearly,
STABILITY OF LINEAR DYNAMICAL EQUATIONS 401
Theorem 8-8
n of 5)
n of 21) The zero-state response ofthe dynamical equation E is BIBO stable if and only
if there exists a finite number k such that
t o, O, u)
~ zero-input response
for all t ;::: to. •
hese results, we shall We see fram this definition that if a trajectory reaches an equilibrium state
ynamical equation. and ifno input is applied, the trajectory will stay .at the equilibrium state forever.
nse. The response of Hence at any equilibrium state, xe(t) = O for .all t;::: too Consequently, an
equilibrium state of x =A(t)x is a solution of A(t)x';"O for aH t ::::(0' Or fram
definition, an equilibrium state of x = A(t)x is a solution of
or
for all t ;::: to. Clearly, the zera state, O, is always an equilibrium state ofx = A(t)x.
402 STABlLITY OF LINEAR SYSTEMS
We shall now define the stability of an equilibrium state in terms of the Definition 8-4
zero-input response.
An equilibrium stat<
i.S.L. at t o and if eve
Definition 8-3
t --> oo. More precise I
An equilibrium state X e is said to be stable in the sense of Lyapunov at t o if and any e> 0, there exist
only if for every ¡; > 0, there exists a positive number Dwhich depends on ¡; and
t o such that if Ilx o- xell ~ D, then
for all t :2:t l +T lf
1I<I»(t; to, xo,O) - xell ~¡; pendent of ti> the eql
over [to,co).
for aH teto. It is said to be uniformly stable i.s.L. (in the sense of Lyapunov)
over [t o, (0) if and only if for every ¡; > 0, there exists a positive Dwhich depends
The concept of a
on ¡; but not on t o such that if Ilxo- xell ~ D, then
that the stabilities e
1I<I»(t; ti> xo,O) - xell ~6 because we do not kl
should be chosen. 1
for any t 1 e t o and for all t e t l' • property, ¿¡ and y can
rium state of a linear
If an equilibrium state is uniformly stable i.s.L., then it is stable i.s.L. How in the large.
ever, the converse may not be true. For example, the zero state of
Theorem 8-9
x(t) = (6t sin t - 2t)x(t)
Every equilibrium st¡ ,
is stable i.s.L., but not uniformly stable i.s.L. See Reference S206. Roughly and only if there exis
speaking, an equilibrium state X e is stable i.s.L. if the response due to any initial
state that is sufficiently near to X e will not move far away from X e . For time
invariant systems, there is no difference between stability and uniforrn stability. for all t :2: too Ir k is i
Example 1 Proof
Consider the pendulum system shown in Figure P3-17 (page 127). The applica Sufficiency: Let Xe be
tion of Newton's law yields
u(t) cos e- mg sin e = mil}
Let Xl = e, and X2 = e. Then we have
Xl =X2
. = (g)
X2 -¡ sin Xl
COS Xl
+-----;;:;¡ U
tí +T
miformly stable Ls.L.
.ivever, are not stable
Figure 8-4 Asymptotic stability.
.J
404 STABILlTY OF LINEAR SYSTEMS
Proof
We prove only the second part of the theorem.
Sufficiency: lf
then
¡lx(t)11 = 11cI>(t, to)x(to)11 ~ 11cI>(t, to)llllxoll ~kle-k2(1-'o)llxol¡ lO + í.
which implies that Ilx(t)II-4O at t-4oo, uniformlyin too Necessity: Ifthe zero
state is asymptotical1y stable, then by definition it is stable in the sense of Figure 8-5 11<I>(t, t 0)11
STABILITY Of LINEAR DYNAMICAL EQUATIONS 405
Lyapunov. Consequently, there exists a finite number k 3 such that 11 <I>(t, to)11 :::;;k 3
for any t o and all t ~ t o (Theorem 8-9). From Definition 8-4, there is sorne
(8-30)
y > O, and for every e> Othere exists a positive T such that
Ilx(to +T)II = 11<I>(to +T, to)xoll:::;;e (8-32)
-xell for aH Ilxoll sy, and for any to. Now choose an Xo such that Ilxoll ='1 and
- xe \\:::;;6 implies that Il<I>(to +T, to)xoll = 11<I>(to +T,to)lll!xoll and choose f. ='1/2; then (8-32) implies
that
1I <I>(t o + T, to)ll :::;; ~ for any t o
1. We prove this by
>(t, to) is not bounded, This condition and 11 <I>(t, to)11 s k3 imply that
say c/J¡/t, lo), becomes I! <I>(t, to)!1 :::;; k 3 for aH t in [to, t o + T)
lf the zero state of X: = A(t)x is uniformly asymptotically stable, then for any
at to if and only if initial state, the zero-input responsewi1l tend to zero exponentially. Hence,
ero state is uniformly for linear equations, if the zero state is uniformly asymptotically stable, it is
xist positive numbers also said to be exponentially stable.
We have discussed separately the stability ofthe zero-input response and the
stability of the zero-state response. By combining these results, we may give
(8-31 )
......
IIk3/2 - - - - - -
~k-4--- .
[ 31 ---- k 18
I 3 -
k2(HOl\lx o ll lO +T i o + 2T lO + 3T
L
::==::::--=.:.~.::-:::".:.-===::::::.=====-===:-._-~--'.'--_ .. _-_._-----"---~--~. __ ._._-----,-_._-----------~-- -~ _-_
..- -... _--_._--_._------_. __ ._-------_._~,-~. __ .-.-.--.~_
------------_.~_.~_._--~---.---.-_.--'-------~~------.
.... _._._-----_.
_.-------.._--
various definitions and theorems for the stability of the entire dynamical of the zero-state resp
equation E. Before doing SO, we shall discuss the relation between the stability study the entire respl
of the zero-state response and the stability of the zero-input response.
The necessary and sufficient condition for the zero-state response of E Definition 8-5
to be BIBO stable is that, for sorne finite k,
A linear dynamical e
L IIC(t)<1>(t, e)B(e)11 de :5; k < 00 if and only if for any
as all the state variab
for any t o and for all t ~ to. It has been shown by the function given in Figure
We see that the e
8-2 that an absolutely integrable function is not necessarily bounded. Con
BIBO stability; they
versely, a bounded function need not be absolutely integrable. Hence, the
of all state variables;
stability Ls.L. of an equilibrium state, in general, does not imply nor is implied
also for any initial ~
by the BIBO stability of the zero-state response. A function that approaches
function properly, bel
zero as t--+ 00 may not be absolutely integrable; hence asymptotic stability
and the system will 1
may not imply BIBO stability. If a system is uniformly asymptotically stable,
every system is requil
then $(t, e) is bounded and absolutely integrable as implied by (8-31); hence
with sorne conditions on B and C, uniformly asymptotic stability may imply Theorem 8 -13
BIBO stability.
A system that is descr
Theorem 8-11 stable if and only ifC
Consider the dynamical equation E given in Equation (8-28). If the matrices
B and C are bounded on (- 00,00), then the uniformly asymptotic stability
of the zero state implies the BIBO stability of the zero-state response. I
for an~ t o and for all i
Proof Proof
This theorem fol1ows directly from the fact that The response of (8-28;
where IIB(t)11 ~kl' 11C(t)11 ~k2 for all t. Q.E.D. Hence we conclude f
initial state and any
The converse problem of Theorem 8-11~that of determining the COD <l>(t, e)B(T) is absolutel
ditions under which the BIBO stability of the zero-state response implies the bounded if and oniy ii
asymptotic stability of the zero state-is much more difficult. In order to solve
this problem, the concepts of uniform controllability and uniform observ Total stability cle~
ability (Definitions 5-4 and 5-8) are needed. We state only the result; its proof true. A systell,I may
can be found in Reference 102. If B and C are bour
stability. The situatic
*Theorem 8-12
rime~invariant ca~
Consider the dynamical equation E given in (8-28). If the matrices A, B, and C . tained for linear, time
are bounded on (- 00,00) and if E is uniformly controllable and uniformly because aH the conditi
observable, then thé zero state of E is asymptotically stable (under the zero- . are very difficult, if TI(
input respOnse) if and only if its zero-state response is BIBO s t a b l e . 1 time-invatiant .dynam
matrix is, however, no
We have studied the stability of the zero-input response and the stability the matrix A.
STABILlTY Of LINEAR OYNAMICAL EQUATIONS 407
:he entire dynamical of the zero-state response. Their relations are also established. We shall now
between the stability study the entire response.
mt response.
,-state response of E Definition 8-5
A linear dynamical equation is said to be totally stable, or T-stable for short,
if and only if for any initial state and for any bounded input, the output as well
as all the state variables are bounded. •
lction given in Figure We see that the conditions of T-stability are more stringent than those of
Hily bounded. Con BIBO stability; they require not only the boundedness of the output but also
tegrable. Hence, the of all state variables; the boundedness must hold not only for the zero state but
t imply nor is implied also for any initial state. A system that is BIBO stable sometimes cannot
~tion that approaches function properly, because sorne of the state variables might increase with time,
: asymptotic stability and the system will bum out or at least be saturated. Therefore, in practice
asymptotically stable, every system is required to be T-stable.
plied by (8-31); hence
c stability may imply Theorem 8-13
A system that is described by the linear dynamical equation E in (8-28) is totally
stable if and only if C(to) and <l>(t, t o) are bounded and
Proof
The response of (8-28a) is
r <Il(t, ,)B(,)u(,) d,
t
llamical equation
(8-33a)
(S-33b)
Its equilibrium states are [x le X2e O]', for any Xl e -1=0, X2e -1=0. In other
es,respectively. As in words, every point in the Xl - X2 plane is an equilibrium state. The eigenvalues
llse and the zero-input of the matrix are -1, O, and O. Its minimal polynomial is s(s + 1). The eigen
;ponse of FE is charac value O, which has a zero real part, is a distinct root of the minimal polynomial.
Hence every equilibrium state is stable i.s.L. •
Example 3
, stable if and only if alI
he zero-input response Consider
lt an equilibrium state
r
property X e = e Al x e 101'
:y in the time-invariant
~O -1~] x(t)
It has eigenvalues -1, O, and O. Its minimal polynomial is S2(S + 1). The
eigenvalue O is not a distinct root of the minimal polynomial. Hence the
equilibrium states of the equation are not stable i.s.L. I
fLyapunov ifand only
:ro) real parts and those
lynomial of A. 3 Theorem 8-15
The zero state ofi = Ax is asymptotically stable if and only if all the eigenvalues
of A have negative real parts.
Hence
• = eA'(x o - x e ).
~e is a constant k such Proof
?;ular matrix such that In order for the zero state to be asymptotically stable, in addition to the bounded
Alp.-l, then
ness of JIeAlII, it is required that IleAl11 tends to ~ero as t-+ ro, 01' equivalently,
that IleA'II-+ Oas ! -+ ro. Since every entry of eA' is of the form tkeajl +':j" we
conclude that I\eAlII-+O as t -+ ro if and only if all the eigenvalues of A, and
sely, from the equation consequently of A have negative real parts. Q.E.D.
'11. I-!ence w~ conc1ude
if Ile A '\ I i~ bounded on Ir a linear time-invariant system is asymplOtically stable, its zero input
Iltry of eA' is bounded. response will approach zero exponentially; thus it is also said to be exponentialIy
e form tkeajl +iOJj" where
stable. This is consistent with Theorem S-lO because asymptotic stability in
legative, it is easy to see the time-invariant case implies uniformly asymptotic stability.
If aj=O, the function The eigenvalues of A are the roots of the characteristic equation of
.er of the J ordan block A, det (sI - A) = O. We have introduced in the previous section the Routh
Q.E.D.
Hurwitz criterion to check whether 01' not all the roots of a polynomial have
negative real parts. Hence the asymptotic stability of the zero-input response
of i = Ax can be easily determined by first forming the characÚ:ristic poly
nomial of A ando then applying the Routh-Hurwitz criterion.
TheBIBO stability of the linear time-invariant dynamical equation FE is
determined by the poies of G(s). Since
very lordan blocks associated
every pole of (;(s) is an eigenvalue of A (the converse is not true). Consequently, It is clear that tol
if the zero state of FE is asymptotically stable, the zero-state response of FE ever may not imply 1
will also be BIBO stable. (This fact can also be deduced directly from Theorem the BIBO stability.
8-11.) Conversely, the BIBO stability of the zero-state response in general asymptotic stability
does not imply the asymptotic stability of the zero state, because the zero asymptol ic stability is
state response is determined by the transfer function, which, however, describes If a linear time-in
only the controllable and observable part of a dynamical equation. able, then the charac
?olynOmial 2f G(s) (1
Example 4 IS a pole of G(s), and (
we have the fOllowing
Consider a system with the following dynamical equation description:
If a linear time-invarian
y=[1 lJx
then the following state
Its transfer function is
1. The dynamical eq ua
g(s) = [1 lJ [S-1
-1 s+1
° J-l [OJ =s+11
1
2.
3.
The zero-state respol
The zero state of FE j
4. AH the poles of the tI
Hence the zero-state response of the dynamical equation is BIBO stable;
5. AH the eigenval ues of
however, the zero state is not asymptotically stable, because there is a positive
eigenvalue. I
A system is said to h
Theorem 8-16 function matrix if the d
troHable' and observable.
Let pletely characterized by i,
x= [~e ~;2 Jx +[~eJ u (8-34a)
the system can be detern
need of considering the d
y = [ Ce Ce J x (8-34b) A remark is in order 1
is independent of t and if
be an equivalent dynamical equation of the dynamical equation in (8-33) the zero state is asymptoti
with {Ae, Be} controllable. Then the dynamical equation in (8-33) is totally that if fo~ each t, aH the eig
stable if and only if al! the eigenvalues of Ae have negative real parts and all the state of x = A(t)x is asym r
eigenvalues of Al' have negative or zero real parts and those with zero real parts the folIowing exampk.
are distinct roots of the minimal polynomial of Al"
Example 5
Proof
Consider the linear time-v:
The application of the Laplace transform to (8-34a) yields
[Si -A
x(s) = e - A~? J- \(0) +[Si -Ae - A~2 -1[BeJu(s) J-
O si - Al' O si - Ac O
[(SI -AJ-l ...~ -IJX(O) +[(Sl - AJ- BeJu(s) The characteristic polynom
1
=
O l$.-Ac) O
where M= (si- Ae)-IA I2 (sl -Áe)-I. From this eql,lation, we may conclude det [,1.1- A
that x is bounded for ahy initialstate and any bounded u if and only if the
conditions in the theorem hold. lf x is bounded, so is y. This establishes the Hence the eigenvalues of A
theorem. Q.E.D. ofthe equation is neither asy
STABILITY OF LINEAR DYNAMICAL EQUATIONS 411
It is clear that total stability implies BISO stability. BIEO stability how
,t tme). Consequently,
ever may not imply total stability because no condition is imposed on Ae in
o-state response of FE the BIEO stability. A comparison of Theorems 8-15 and 8-16 yields that
directly from Theorem asymptotic stability implies total stability but not conversely. Hence
.te response in general asymptotic stability is the most stringent among these three different stabilities.
tate, because the zero If a linear time-invariant dynamical eguation is controllable and observ
LÍch, however, describes able, then the characteristic polynomial of A is egual to the characteristic
al equation. polynomial of G(s) (Theorem 6-2). This implies that every eigenvalue of A
is a pole of G(s), and every pole of G(s) is an eigenvalue of A. Consequently,
we have the following theorem.
on description:
Theorem 8-17
If a linear time-invariant dynamical equation FE is controllable and observable,
then the following statements are equivalent:
1. The dynamical equation is totally stable.
2. The zero-state response of FE is BIEO stable.
1
3. The zero state of FE is asymptotically stable (under the zero-input response).
- s +1
4. AH the poles of the transfer function matrix of FE have negative real parts.
uation is BIEO stable;
5. AH the eigenvalues of the matrix A of FE have negative real parts. I
::cause there is a positive
lI>(t'O)=[e~' , '\'¿ 1-
)~min
n
I
Xii
i= 1
(as in Problem 4-1 or by direct verification), whose norm tends to infinity
as t----> OCJ. •
The fact that p- I =P
A min ~
, 1-1 2::S;X *Mx -__x*Mx -
1... Xi A __ '\-\2 ::s; A'
1... A¡ X¡
"
max (8-36)
i=l i= 1
norm tends to infinity
I
The fact that p-l = P* implies that
n
IlxW=x*x=x*x= I IX¡1 2
i= 1
" where
M = j:::: ;::~ ;::~lJ'
Lm.31 m32 nt:33
k
414 STABlLITY OF LINEAR SYSTEMS
AlI the eigenvalues of A have negative real parts or, equivalently, the zero state where N ~ -(A*M·
of x = Ax is asymptotically stable if and only if for any given positive definite along any trajectory
hermitian matrix N, the matrix equation - x*(t)Nx(t) is always
A*M+MA= -N (8-37) cally with time along
approach zero as t -> (
has a unique hermitian solution M and M is positive definite. I only at x =0; hence v
M and N that are relat
Corollary 8-20 the zero state as t -> (
AlI the eigenvalues of A have negative real parts, or equivalently, the zero state x=Ax. A Lyapuno'
of x = Ax is asymptotically stable, if and only if for any given positive semi concept of distance or
definite hermitian matrix N with the property {A, N} observable, the matrix of x = Ax decreases w
equation
Proof of Theorem 8-;
A'!'M+MA= -N
Sufficienc y: Consider .
has a unique hermitian solution M and M is positive definite. I
will be bowl shaped, as shown in Figure 8-6. Consider now the successive
values taken by V along a trajectory of x= Ax. We like to know whether the
:oof of condition 2 can value of V wil\ increase or decrease with time as the state moving along the
l proof of condition 3, trajectory. Taking the derivative of V with respect to t along any trajectory of
Q.E.D. x= Ax, we obtain
the Lyapunov theorem :t V(x(t)) = {~t (x*(t)Mx(t)) = (:t x*(t)) Mx(t) +x*(t)M (:t X(t))
:h-Hurwitz criterion.
=x*(t)A*Mx(t) +x*(t)MAx(t)=x*(t)(A*M +MA)x(t)
= -x*(t)Nx(t) (8-39)
ivalently, the zero state where N ~ - (A*M + MA). This equation gives the rate of change of V(x)
. given positive definite along any trajectory of x= Ax. Now if N is positive definite, the function
- x*(t)Nx(t) is always negative. This implies that V(x(t)) decreases monotoni
cally with time along any trajectory of x = Ax; hence V(x(t)) will eventually
(8-37)
approach zero as t -> oo. Now since V(x) is positive definite, we have V(x) = O
efinite. I only at x = O; hence we conclude that if we can find positive definite matrices
M and N that are related by (8-37), then every trajectory ofi = Ax will approach
the zero state as t -> oo. The function V(x) is caBed a Lyapunov function of
i = Ax. A Lyapunov function can be considered as a generalization of the
Livalently, the zero state
concept of distance or energy. Ir the "distance" of the state along any trajectory
ny given positive semi
of i = Ax decreases with time, then x(t) must tend to O as t-> oo.
observable, the matrix
Proof of Theorem 8-20
Sufficienc y: Consider V(x) = x*Mx. Then we have
efinite. I
. d
V(x) ~ dt V(x) = - x*Nx
8-20 is that if A is
r positive semidefinite, along any trajectory of i = Ax. From Theorem 8-18, we have
However, it does not V x*Nx (AN)min
,itive definite, then the -=---<---- (8-40)
V x*Mx - (AM)max
positive semidefinite.
few comments. Since
V(x)
n matrix N, the matrix
ti is a hermitian matrix,
. M is a real symmetric
I be solved. Hence the
1gebraic equations. To
for M, and then check
y task. Hence Theorem
rmining the stability of
>ility study of nonlinear
method of Lyapunov..
itz criterion.
punov theorem. lfthe
)
Figure 8-6 A Lyapunov function V(x).
(8-38)
\
!
\
J
416 STARII.ITY (lF I.INEAR SYSTEMS
where (AN)min is the smal1est eigenvalue of N and (AM)max is the largest eigenvalue t -" oo. This shows t:
of M. From Theorem 8-19 and from the assumption that the matrices M and similarly proved as ir
°
N are positive definite, we have (AN)min > and (AM)max > O. lf we define Theorem 8-20 ar
asymptotic stability
CI. ~ (AN)min right and are basic
- (AM)inax provide a simple pro.
then inequality (8-40) becomes V:::;; -Cl.v, which implies that V(t):::;;e-atV(O).
lt is clear that CI. > O; hence V decreases exponentially to zero on every trajectory A proof of the Ro
°
of x = Ax. Now V(x) = only at x = O; hence we conclude that the response of
x = Ax due to any initial state X o tends to Oas t -" oo. This proves that the zero
D(s) = (
state of x = Ax is asymptotical1y stable. N ecessity: If the zero state of x = Ax with real coefficients .
is asymptotically stable, then all the eigenvalues of A have negative real parts.
Consequently, for any N, there exists a unique matrix M satisfying
A*M +MA=-N and compute
and M can be expressed as Do(s)
D (s) =CI.¡:
M = L'" eA 'INe A' dt (8-41 ) 1
Since H is nonsingular and e At is nonsingular for all t, we have HeAtx o =1=- O for
°
all t unless X o =0. Hence we conclude that xóMxo > for all X o =/=-0, and M
is positive definite. This completes the proof of this theorem. Q.E.D.
For convenience, we s
Theorem 8-6
In order to establish Corollary 8-20, we show that if N is positive semidefinite The polynomial D(s) i
CI.¡, Cl. z , .. . , Cl. are posit
and if {A, N} is observable, then x*(t)Nx(t) cannot be identically zero along any 1I
nontrivial trajectory ofx = Ax (any solution due to any nonzero initial state xo)'
First we use Theorem 8-19 to write N as N = H*H. Then the observability ProoF
of {A, N} implies the observability of {A, H} (Problem 5-36). Consider First we assume that ~
x*(l)Nx\t) ,.~ x(¡cH"ik';x¡", lilleA(xoi¡2 function
Since {A, H} is observable, all rows of He At are linearly independent on [0, 00). g(s) ~
Hence we have that HeA1x o = O for all t if and only if X o = O. Because e A1 is
analytic over [0,00), we conclude that for any X o =/=- O, HeAlxo can never be The assumption Cl. i =/=-1
identically zero over any finite interval, no matter how small; otherwise it factor between Do(s)
would be identically zero over [0, 00). See Theorem B-l. Note that HeA1x o = O, between D¡(s) and D(s;
at sorne discrete instants of time, is permitted. Cohsider the block
With the preceding discussion, we are ready to establish Corollary 8-20. function from u to y is !
Consider the Lyapunov function V(x) defined in (8~38)and dV(x)/dt = equivalently, from the
- x*(t)Nx(t) in (8-39). If X o =1=- O, dV(x)/dt:::;;O and theequality holds only at
sorne discrete instants of time; hence V(x(t)) will decrease with time, not neces
sarily monotonic at every instant of time, and will eventually.approach zero as 7 This rollows Reference 90.
.. =~
t --+ oo. This shows the sufficiency of the corollary. The necessary part can be
the largest eigenvalue
3.t the matrices M and similarly proved as in Theorem 8-20 by using (8-42).
O. If we define
Theorem 8-20 and its corollary are generally not used in checking the
asymptotic stability of x = Ax. However, they are important by their own
right and are basic in the stability study of nonlinear systems. They also
provide a simple proof of the Routh-Hurwitz criterion.
s that V(t) s: e-'" V(O).
ero on every trajectory A proof of the Routh-Hurwitz criterion. Consider the polynomial
je that the response of ao>O
lis proves that the zero
with real coefficients a¡, i =0, 1,2, ... , n. We form the polynomials
le zero state of X = Ax
ve negative real parts. Do(s)=aos n +azs n - Z + ...
[ satisfying D 1(s)=a1 Sn - 1 +a3 Sn - 3 + ...
and compute
Do(s)
--=<XlS +-----------:-------
D 1 (s)
(8-41 ) <XzS +----------:-----
1
<X3 S + - - - - - - - - - -
:finite, so is M. Let H
1
8-19). Consider +--
(842)
+-
ve have HeAlxo O for For convenience, we shall restate the theorem here.
) for all Xo =/= O, and M
orem. Q.E. D. Theorem 8-6
The polynomial D(s) is a Hurwitz polynomial if and only if aH the n numbers
is positive semidefinite <X l' a z , ... , <XII are positive.
ntically zero along any
onzero initial state xo)·
ProoF
rhen the observability
)-36). Consider First we assume that a11 the iX/S are different from zero. Consider the rationa!
function
"011' 1
indepeodent 00 [O, el»).
X o = O. Because e
Al'IS
',HeA'xo can oever be The assumption <Xi =/=0, for i = 1, 2, ... , n, implies that there is no common
)w smaH; otherwise it factor between Do(s) and D 1(s). Consequent1y, there is no common factor
Note that HeAtx o = O, between Dl(s) and D(s); in other words, g(s) is irreducible.
Consider the block diagram shownin Figure 8-7. Weshow that the transfer
:ablish Corollary 8-20. function from u to y is 9(s). Let h 1(s) be the transfer function from X n to X n -1 or
(8-38) and dV(x)/dt = equivalently, from the terminal E to the terminal F, as shown in Figure 8-7.
equality holds only at
;e with time, not neces
7 This follows Reference 90.
ually approach zero as
418 STABILITY Of LINEAR SYSTEMS
Then
A*M -t
y(s) 1/a l s 1
D(s) = 1 +[1 +hl(s)J/als = 1 +als +hl(s)
Let hz(s) be the transfer function from X,,-l to x,,_ z, then h l (s) can be written as It is clear that N is .
verify that {A, N} is o
the zero state of x =)
hl(s) = Yazs = 1~ positive definite, or eq
1 + hz(s)/azs a 2 s +hz(s)
the zero state of x = A
Proceeding forward, we can show easily that the transfer function frorn u to y of A, or equivalently
is indeed g(s). With the state variables chosen as shown, we can readily write words, D(s) is a Hurwi
the dynamical equation of the block diagram as are positive.
Consider now the
In other words, sorne
Xl O O O O O Xl O equal to zero. Suppo
an
8-1 are equal to zero,
-1 factor. The cornmon
Xz O O O O Xz
O
an- 1 an- 1 say fes). Then D(s) c
-1 even function or an (
X3 O
an- Z
O O O O X3
+
O u Hurwitz polynornial.
I : zero, we may replace
l:
-1
Xn-l O O O Ir sorne a i is negative, ~
az by e) has a positive re
-1 functions of its coeffi.1
xn O O O
positive or zero real p.
al
(8-43)
y=[ O O O O O lJx
Irreducibility of (8-43) can be verified either by showing that it is controUable *8-6 Linear Tim
and observable or by the fact that its dimension is equal to the degree of the
denorninator of g(s). Consequently, the characteristic polynomial of the The stability concepts
rnatrix A in (8A3) is ~qual to the denorninator of g(s) (Theorem 6-2). Now we 'applicable'to the disCl
shall derive the condition. for the zero state of (8-43) to be asymptotically st¡¡.ble. quite different. In thi
DISCRETE-TIME SYSTEMS 419
M~ ~'
O O
Ci. 1I _¡ O
O
O
Ci. z
O J] (8-44)
O O O O O
O O O O O
O O O O O
A*M +MA= ~-N (8-45)
O O O O O
¡ (s)
O O O O 2
l X:-¡ If sorne Ci.¡ is negative, at least one root of the modified D(s) (since abZ ) is replaced
by ¡¡) has a positive real part. Now the roots of a polynornial are continuous
functions of its coefficients. Hence, as ¡¡--->O, at least one root of D(s) has a
XII 1J positive or zero real part. Q.E. D.
(8-43)
]x
Consider a relaxed linear time-invariant di serete-time system described by Table 8.2 Discrete-tim
k
y(k) = ¿ g(k - m)u(m) (8-46) Cl~) a\O)
m=O _) kOCl~,O) koa~,o~ 1
Then for any bounded-input sequence {u(k)}, (that is, there exists a finite h Clbl ) a\l)
such that lu(k)1 < h for k =0, 1,2, ...), the output sequence {y(k)} is bounded, if -)klCl~,1~l kla~IJ.2
2 a\2)
and only if Clb )
C()
a~l-l) aY'-l)
¿ Ig(k)1 < ca (8-47)
_)k"_laY'-I)
k=O
ag')
that is, {g(k)} is absolutely summerable. 8 The proof of (8-47) is similar to the
continuous-time case and is left as an exercise. The z-transform of (8-46) yields
Y(z) = g(z)u(z) is the difference of its t
Ir g(z) is a rational function of z, then the system is BlBO stable if and only if all by the same procedu
the poles of g(z) have magnitudes less than 1, or equivalently, aH the poles of define (Xi = agJ, i = 0,1.
g(z) lie inside the unit circle of the z planeo This can be readily proved by
noting the z-transform pair Theorem 8-21
All the roots of D(z) i
k =0,1,2, ... <:> g(z) =fl[g(k)] =~ numbers (Xi ~ ag), i = 1
z-b
where bis a real or a complex number. lf Ibl < 1, then We shall prove thi:
discrete-time systenis.
ical equation
Otherwise, it diverges.
lf g(z) is irreducible, the poles of g(z) are equal to the roots ofits denominator.
lf the degree of the denominator is three or higher, the computation of the roots The concepts of equi
is complicated. We introduce in the following a method of checking whether asymptotic stability ~
or not aH the roots of a polynomial are inside the unit circle without computing Theorem 8-14, every ~
explicitly the roots. The method is a counterpart of the Routh-Hurwitz only if aH the eigenvall
criterion. \vith magnitudes eq ua
Consider the polynomial with real coefficients Similar to Theorem 8
stable if and only if al1
D(z) = aoz n +alz"- 1 + ... +an-1z +a ll ao>O (8-48)
Lyapunov theorem fOl
We define ala) = ai, i = 0, 1, ... , n, and form the table in Table 8-2. The first
row is just the coefficients of D(z). The constant k o is the quotient of its last Theorem 8-22
and first elements. The second "row is obtained by multiplying k o on the first
row, except the first element, and then reversing its order. The third row All the eigenvalues of 1
positive definite herm
hermitian matrix N wi
8 An absolutely integrable function is neither rrecessarily bounded nor necessarily"approaches zero as
(-+ 00as shown in Figure 8-2. An absolutely summerable sequence however is always bOllnded
and approaches zero as k-+ oo. Thus the stability problem in the discrete-time case is simpler
than the one in the continuing time case. has a unique hermitiar
OISCRETE-TIME SYSTEMS 421
(8-47)
z AH the roots of D(z) in (8-48) have magnitudes less than I if and only if the n
]=- numbers (J.¡ ~ ag), i = 1, 2, ... , n, computed in Table 8-2 are aH positive. I
z-b
We shall prove this after the establishment of the Lyapunov theorem for the
diserete-time systems. Consider the linear time-invariant diserete-time dynam
ieal equation
and computing
V(x(k» = x*(k)Mx(k) 1
~ V(x(k»~ V(x(k + 1»
- V(x(k» = x*(k)A*MAx(k) - x*(k)Mx(k) 2 +
cxi _ [
= x*(k)(A*MA - M)x(k)
= - x*(k)Nx(k)
with Do(z) = D(z). Note that the coefficients of Di(z) are the a~¡), j =0,1, ... , n - i c=
defined in Table 8-2. It can be verified that the reciprocal of Di(z) can be
ex pressed as b=
-kak,,_¡CJ./l_¡/CJ.¡ -k¡
Simple manipulation among (8-51), (8-52), and (8-53) yields -kaCJ..!cx¡ -k¡
- IX· ¡
Di· 1(z) = (zD ¡( z) + k i _ 1 D;(z) ---'-=-- (8-54 )
a¡
This state equation is ob1
- - IX· ¡
the dimension of A; henc
D i - ¡ (z) = (ki_¡zD¡(z) + D;(z)) -'-- (S-55 )
lXi F or this ma trix A, if we e
These recursive equations are valid for i = 1,2, ... ,n. Note that D/l(z) = D/l(z) =
a\í'l. Consider now the transfer function
D i - 1 (Z) k i - 1 zD i(z) + Di(z) (1 - kt_l)D¡(z) then it is straightforward
--- = = k· 1 +----.:...--.:...-=..:.~,..:
D¡-I(Z) zD¡(z) +k i - 1 Di(z) 1 zDi(z) +k i - 1 D i(z)
which becomes, by usiri.g.(8~53), A*MA-
- x*(k)Mx(k)
8-21. Define, l'or i = l'or i = 1, 2, ... , n. The block diagram ol' (8-56) is shown in Figure 8-8. lf we
apply this diagram repetitively, we can final1y obtain the block diagram ol'
<¡ = a~:~ Ja~) (S-49) Do(z)/Do(z) shown in Figure 8-9. Note that Dn(z)/Dn(z) = 1. In this block
+a~) (8-50) diagram, there is a total ol' n unit delay elements. lf we assign the output ol'
Z-I as a state variable as shown, then we will obtain the l'ollowing state equation
tials can be defined re
x(k + 1) = Ax(k) + bu(k)
y(k) =cx(k) +kou(k)
!, ... , n (8-51) with
e = [1 O O .. , O O]
thea)i), j =0,1, ... , n - i
iprocal ol' Di(z) can be b=[~6
<xo
<xnJ'
<XO
and
2, ... , n (S-52)
-kok, 1 O O O
k"f) or - k ok 2rx 2/rx¡ -k¡k2 1 O O
- k ok 3rx 3 /rx¡ -k¡k 3rx 3/rx 2 -k 2k 3 O O
(S-53) A=
- kok n ¡rx" - ¡/rx ¡ -k¡k,,_¡IY.,,_ Jrx 2 - k 2 k" _ ¡rx" - ¡/rx 3 -k n 2 k,,-¡
-kolY.,jlY.¡ -k¡IY.,j1Y. 2 - k2rx ,jrx 3 -k Zcx,ja n
ll - 1
ields
(S-57)
(8-54·)
This state equation is obtained l'rom D(z)/D(z), and the degree 01' D(z) is equal to
the dimension ol' A; hence the roots ol' D(z) are identical to the eigenvalues ol' A.
- 1 (8-55) For this matrix A, if we choose M as
8-7 Concluding
r-------------~
--------------------l
I In this chapter we in
I and the stability i.s.L
I
i
zero-input response.
+ I being uniformly stab
I-~ -..-{ ,. I
I not available.
I
For the time-inva
I
I
I
the transfer funetion (
I
I
I
the Routh~Hurwitzcrj
I
I
the Lyapunov theorer
I
I form by a numerica11:
I
I
I then be more easily Cl
I
I I
form is also used in tl
I
I I
I
I I
I Reference SI 07. One,
: iT-
I i dJi 1
Lyapunov equation a
I
: L""r-J ... i ¡.~
1 I
; --:
1
-- I
I .
J-~
I
l .••
""-] : 1
Routh-Hurwitz methc.
in checking the stabili
the former may also b
The coneepts of s1
I I : _ 1;1;
: I :
1
I L J'Q<Q'I I
The stability conditio
I I I
I
function He inside the '
1 - I I
BIBO stable; whereas
I I
I
I
I
. - _,Jo} I
L__---1~tr- I~I~:
I
I
, I
whieh maps the left~ha
I
I
may transform a dise
I
I
using the bilinear trar
I
I
IL
_ on diserete-time equati
I
------------ ...J
1
I
I ~
1 _o I
~
:
1
I ?:J
1
T
?:J ;1; :;
" I el Problems
.~ - - ~-- ~ ~~ ~ -l
o
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - u..
8-1 1s a system with th
How about g(t, r) = sin te
424
I~:
zero-input response, For the time-varying case, a system may be stable withou t
being uniformly stable. For the time-invariant case, there is, however, no
distinction between uniform stabilities and (nonuniform) stabilities. Although
1----, I
I
i
necessary and sufficient eonditions are established for the time-varying case,
they can hardly be employed because state transition matrices are generally
II I I
not available.
I I
F or the time-invariant case, the stability can be checked from the poles of
I
the transfer function or fram the eigenvalues of the matrix A. Whether or not
I
I
all the eigenvalues of A have negative real parts can be cheeked by applying
I
the Routh-Hurwitz criterion to the characteristic polynomial of A or by applying
I
the Lyapunov theorem. The characteristic polynomial of A can be eomputed
I
I
by using the Leverrier algorithm (Problem 2-39), which however is sensitive to
I
computational errars. If the matrix A is first transformed into a Hessenberg
I
form by a numerieally stable method, the characteristic polynomial of A can
I
1
then be more easily eomputed. See Referenees S90 and S2l2. A Hessenberg
I
form is also used in the efficient method of solving the Lyapunov equation in
I
tion needed in the Routh-Hurwitz method is mueh less than that in solving the
I
Lyapunov equation and eheeking the positive definiteness of M. Henee the
I
Routh-Hurwitz method is simpler eomputationally than the Lyapunov method
I
I
The stability eonditions, however, are different. lf all the poles of a transfer
I
function lie inside the open left-half s plane, then the continuous-time system is
I
I
BIBü stable; whereas if all the poles of a transfer function lie inside the open
I
unit circle of the z plane, then the discrete-time system is BIBO stable. Their
I
1"
relationship can be established by using the bilinear transformation
:~
! \--1 .2- 1
¡ s =_._
I
I 'Q z +1
lO
I
I
I 6
which maps the left-half s plane into the unit circle in the z planeo Although we
I
I f':
I
1 ~ may transform a diserete-time prablem into a eontinuous-time problem by
,;.o
~I~ I 1..><: using the bilinear transformation, it is simpler to check the stability directly
1
I~~:--.1_ _ I g
1:;5
on discrete-time equations.
l<e
I
1CJ'l
I ,
\~
IClO
~I~
~ ~ I :J Problems
IQ Q I en
___________ ...1 u.
8-' ls a system with the impulse responses g(l,r)=e-21t1-hl, for l'2:r, B[BO stable?
How about g(l. r)=sin te-(I-<) cos r?
426 STABlLlTY OF LINEAR SYSTEMS
8-2 Is the network shown in Figure P8-2 BIBO stable? Ir not, find a bounded input that 8-11 Prove parts ! ane
will excite an unbounded output. systems describable by r
ti 1F
I
Y
8-12 Consider the dyn,
~
Figure P8-2
8-3 Consider a system with the transfer function g(s) that is not necessarily a rational Y'
function of S, Show that a necessary condition for the system to be BIBO stable is that Is the zero state asympto
19(s)1 is finite for all Re s ~ O. stable? Is the equation 1
8-4 Consider a system with the impulse response shown in Figure P8-4. Ir the input 8-13 Consider
u(t) = sin 2nt, for t ~O, is applied, what is the waveform of the output? After how many
seconds will the output reach its steady state?
g(t}
8-5 Is a system with the impulse response g(t) = 1/(1 + t) BIBO stable?
8-6 Is a system with the transfer function g(s) = e-s/(s + 1) BIBO stable?
8-15 Find the rangcs of I
8-7 Use the Routh-Hurwitz criterion to determine which of the rollowing polynomials
are Hurwitz polynomials.
a. S5 +4s 4 + !Os 3 +2,,2 +5, +6
b. SS +S4 +2s 3 +2s 2 +5s +5 is BIBO stable.
8-8 Can you determine without solving the roots that the real parts of all the roots of
S4 +14s3 +71s 2 +154s + 120 are smaller than -1? (Hint: Lets=s' -1.)
8-9 Give the necessary and sufficient conditions for the following polynomials .lo be
Hurwitz polynomials:
8-10 Find the dynamical-ecjuation description of the· network shown in Problem 8-2.
Find the equilibrium states of the equation. Is the equilibrium state stable in the sense of 8-17 Showthatx=Oísti
nd a bounded input that 8-11 Prove parts 1 and 2 ofTheorem 8-3 by using the Laplace transform for the c1ass of
systems describable by rational transfer functions.
X-l O
-1
¡ -1
O
O
O
-2
1
O
O
O
1
O
- 10
!}+W
-4 -IJ
y=[ O O O O OJx
ot necessari\y a rational Is the zera state asymptotically stable (for the case u =,O)? Is its zero-state response BIBO
o be BIBO stable is that stable? Is the equation totally stable?
)"=[1 1Jx
Find all the equilibrium states Of the equation. Is every equilibrium state stablc í.s.L.?
Is it asymptoticaIly stable? Is its zero-state response BlBO stable? Is the equation totany
stable?
8-14 Check the BlBO stability of a system with the transfer function
25 2 -1
stable?
o stable')
8-15 Find the ranges of k¡ and k 2 such that the system with the transfer function
he foIlowing polynomials 5 +k¡
is BIBO stable.
,owing polynomials to be
is B1BO stable. Can we conclude that the real part of A is negative? Why?
:k shown in Problem 8-2.
state stable in the sense Of 8-17 Show that x = O is the only solution satisfying (<1>(t, to) - I)x =,0 if 1I <1>(t, to)"":"O as
t---+OCi.
428 STABlllTY Of LINEAR SYSTEMS
E : * =2ex +u
y=x
Sho\\l that the zero state of E is nol stable i.s.l.(under the zero-input response).
8-19 Consider the equivalent equation of E in Problem 8-18 obtained by the equivalence
transformation.;; = PU)x. where P(C) =e- r ': Figure P8-26
E: .\: =(2ee- r' _2ee- r'jer'.;; +e-r'u =0 +e-r'u
y=el:!.x
Show that the zero state of Eis stable i.s. L. (under the zero-input response). From Problems 8-27 Consider a discrete-t
8-18 and 8-19. we conclude that (//1 equivalelJce eralJsforll1ac ionlleecl'lOe preserre che scabilitr
af che zero scaee. Does an equivalence transformation preserve the BI BO stability of the
zero-state response?
Show that any bounded-in
8-20 Show that stability i.s.L. and asymptotic stability of the zero state of Ji: = A(C)x are if and only if
invariant under any Lyapunov transformation (see Definition 4-6). Is the transformation
P(c) = e-/ in Problem 8-19 a Lyapunov transformation?
8-21 Show that if Ji: = A(t)x is stable i.s. L. at ca' then it is stable i.s. L. at every tI "2: ca·
8-28 Consider a discrete-
[Hinl: Use <1>(1, e¡)= «1>(1,10)<1>-'(1" lo) and note the boundedness of <I>-I(e" to) for any
finite 1, and 10.J
x{i
-1
1
O
if and only if
y=[ 2 3 'J x
Is the zero state asymptotically stable? Is the zero-state response BIBO stable? Is the 8-29 Consider a system \
system T-stable?
and
X =Ax +Bu
W(T)= I T
e-A'BB*eAo'dr T is an arbiti-ary ,positive number
have proper rationa\ tran
8-26 Are the networks shown in Figure P8-26 totally stable? AIJSlrers: No; yeso
uation:
)-input response).
Show that any bounded-input sequenee {u(n)} excites a bounded-output sequenee {y(n)}
zero state of x = A(I)x are if and only if
-ó). ls the transformation
for all n
m=-oo
I Ig(m)1 ~k < 00
m=O
mse BIBO stable? Is the 8-29 Consider a system with the impulse response
ltion
r 19¡(I)1 dI ~kl < 00
and I la,J~k2 < 00
i=O
8-30 Prove Corollary 8-3by usíng partía\ fraetion expansion for the class of systems that
have proper rationa\ transfercfunetion deseriptions.
;itive number
8-31 Is the funetion
a.
[¡ ~]
O
3
1
b.
[¡
O
O
O ~]
[",a, and W = diag [W I , W 2 ]
[! ~
c. O d. alaz
O
1 aloz
a l a3
",a:l
ozaz aZ a 3
aZ a 3 a3 a 3
W 2 ha ve no diagona [ er
--'
a3 a l a3 a Z a 3Z
---- ----
Al + A3 },z + }'3 21 3 and the matrix A in (8
Verify that Ac=Q-1AQ
is a positive definite matrix. [Hint: Let A = diag (A ¡, Az, Az).J
8-34 A real matrix M (not necessarily symmetric) is defined to' be, as in Definition 8-6,
positive definite if x'Mx > O for all nonzero x in IR". Is it tme that the matrix M is positive
definite if all the eigenvalues of M are positive real or if all the leading principal minors are
positive? If not, how do you check its positive definiteness? Hil1l: Try
(Due to Y. P. Ham)
l
8-37 Determine the asymptotic stability of the discrete-time equation
~
O I O
. O O 1
. O
x(k + 1):= O O O .O
l ' x(k)
(1) by computing itsch~racterislic polynomial and the'n' ap'plyingTheo;:em 8"21 and (2)
Figure P8-40
by solving.the Lyapunov equation in Theorem 8-22.
PROBLEMS 431
8-38 Consider the internally balanced system in Problems 6-23 to 6-25 partitioned as
are positive definite or
I2 J[X 1J+[8 IJU
[ x~lJ
2
= [Al!
A 21
A
A 22 _ x2 82
~~[~ -~J
I O
'3 O 1
O O
'3 -a ll -Q"-l -Q"-2
and the matrix A in (8-57) obtained from the discrete-time stability table in Table 8-2.
Verify that A c = Q - [AQ or QA c = AQ, wherc
Q~[l
all) a~) a!i1 2
ab11 a\2) a~1.!3 al.".
a~12J 2 ]
be, as in Definition 8-6,
the matrix M is positive O O
ag,-l) aY,:-I)
ling principal minors are O O O ag')
int: Try
(Due to Y. P. Harn.)
8-40 Consider the system shown in Figure P8-40, where S¡ is described by Xi = Á¡Xi+ R¡u¡,
y¡ = c,x¡ + E¡u¡. Use the composite state x' = [X'1 x~J to develop a sta te variable de
scription of the system with [r'l r2]' as the input and [Y'I y'l u'¡ U'l]' as the output.
. ,n, be a set of linearly 15 the matrix Á of this descriplion the salTle ':\s the ane in the tjm~-inwlr¡ajll. c,:,.se of 1}.fA\')
. ,11,then pv:iI is posilive What is the condition for the system to be asymptotically stable'! Will thecondition ensure
that the system be BIBO stable from any input-output, pair?
latíon
f[ + .u¡ YI
SI
~
I I
(k) Y2 u2 + .. f2
S2
+
Ca)
Cb)
Figure 9-1 (a) Unity feedback system. (b) lnput-output feedback system.
with state vectors Xi, i = 1, 2. Then the state vector of any connection of SI It is assumed that th
and S2 will be defined as x' = [X'1 x~]. With this definition, there will be no transfer-function m[
confusion in the state-variable description of any composite system. The equations (9-1) are c
state-variable descriptions of the tandem, parallel, and feedback connections that the transfer-fun
of two systems are derived in (3-62) to (3-64). (;1(S) +G 2 (s); the tI
All the subsystems that form a composite system will be assumed to be followed by S2 is (
completely characterized by their transfer-function matrices. This assumption, connection of SI W'
however, does nat imply that a composite system is completely characterized ([ + Gl (S)G 2 (S))-IGi
by its composite transfer-function matrix. In Section 9-2 we study the con the systems SI and S
ditions of complete characterization of composite systems. For single-variable function G(s) comple
systems, the conditions are very simple; ifthere is no common pole in the parallel
connection, or no pole-zero cancellation in the tandem and the feedback con Example 1
nections, then the composite system is completely characterized by its transfer
Consider the paralle
function. In Section 9-3 we extend the results of Section 9-2 to the multi
whose dynamical-eql
variable case, with the condition of pole-zero cancellation replaced by the
condition of coprimeness. In Section 9-4 we study the stability problem of
composite systems. The stability conditions are stated in terms of the transfer
matrices of the subsystems. The remainder of the chapter is devoted to the and
design problem. In Section 9-5 we study the design of compensators in the
unity feedback system to achieve arbitrary pole placement and arbitrary
Their transfer functio
denominator matrix. The problem of pole placement and that of arbitrary
denominator are identical in the single-variable case. They are, however,
different in the multivariable case; the compensator required for the latter is
much more complicated than that required for the former. In Section 9-6 we
The composite transf(
design robust control systems to achieve asymptotic tracking and disturbance
rejection. The static decoupling problem is also discussed; both robust and
nonrobust designs are considered. In the last section, we study the design of
compensators in the input-output feedback system. The results are more
1t is clear that g(s) = 1
general than those obtained in the state-variable approach discussed in
beca use g(5) does not
Chapter 7.
be checked from the
The references for this chapter are S2, Sil, S19, S34, S35, S4Ü, S49 to S51, S54,
nection, we have u I =
S55,S64toS66, S75,S81,S85,S93, S94,S98,SI74,SI85, SI99,S218,S237,and
equation is
S23?,
Composite Systems
It is easy to check Ü
observable; hence frOl
Consider two systems Si> for i = 1,2, with the dynamical-equation descriptions
charaderized by its ce
x¡=Á¡x¡ +B¡"¡ (9-1 a)
y¡ =C;x¡ + E¡"¡ (9-1b) In the. following
transfer functions co
where Xi, U¡, and y¡ are,· respectively,the state, the input, and the output of the
problem directly fre
system Si. A¡, B¡, C¡,.and ·E¡ are real constant matrices. The transfer-function
dynamical equations.
matrix of S¡ is
function matrix, then
(9-2) of the systemis equal t
COMPLETE CHARACTERIZATlON OF SINGLE-VARIABLE COMPOSITE SYSTEMS 435
my connection of SI lt is assumed that the systel'Ils SI and SJ. are completely characterized by their
tion, there will be no transfer-function matrices G l (s) and G 2 (s); 01', equivalently, the dynamical
lposite system. The eq uations (9-1) are controllable and observable. It was shown in Section 3-6
eedback connections that the transfer-function matrix of the parallel connection of SI and S2 is
6 1(s) +6 2(s); the transfer-function matrix of the tandem connection of Si
ill be assumed to be followed by S2 is 6 2(5)6 1(5); the transfer-function matrix of the feedback
es. This assumption, connection of SI with S2 in the feedback path is 6 1(5)(1+6 2 (5)(;1(5))-1=
lpletely characterized (1 +6 1 (s)6 2(s))-1(;I(S). Although (;1(S) and (;2(S) completely characterize
-2 we study the con the systems SI and S 2, respectively, it does not follow that a composite tnmsfer
. For single-variable function G(s) completely ~haracterizes a composite system.
on pole in the parallel
nd the feedback con Example 1
terized by its transfer Consider the parallel connection of two single-variable systems SI and Sl
.on 9-2 to the multi whose dynamical-equation descriptions are, respectively,
tion replaced by the
stability problem of FE1: Xl =X l +Ul
1 terms of the transfer Yl =X 1 +Ul
lter is devoted to the and FE!: X2=X2- U2
compensators in the Y2 =X2
ement and arbitrary Their transfer functions are
and that of arbitrary
They are, however, A s A -1
91(S)=- and 92(S) =-1
uired for the latter is s-1 s-
;r. In Section 9-6 we The composite transfer function of the parallel connection of SI and S 2 is
:king and disturbance
sed; both robust and A A 5 -1
g(s) =91(S) +g2(S) = - - +- = 1
le study the design of 8-1 s-1
fhe results are more It is clear that g(5) = 1 does not characterize completely the composite system,
,proach discussed in because g(s) does not reveal the unstable mode e' in the system. This can also
be checked from the composite dynamical equation. In the paral1el con
5, S40, S49 to S51, S54, nection, we have u 1 = U 2 = u and y = y 1 + y 2; hence the composite d ynamical
SI99,S218,S237,and equation is
y=[1 lJx +u
Variable
It is easy to check that the composite equation is not controllable and not
observable; hence from Definition 9-1, the composite system is not completely
·equation descriptions characterized by its composite transfer function. I
(9-1 a )
(9-1b) In the followíng we shall study the cOl)ditions under which composite
transfer functioIls completely describe composite systems. We study this
and the output of the problem directly from transfer-function. matrices without looking into
The transfer-function dynamicalequations. If a system is c6rilpl~tely chara~ierized by its transfer
function matrix, then the dimension. of the dynamical-equation description
(9-2) of the system isequal to the degree of its transfer-function matrix (Theorem 6-2).
COMPLETE (
436 LINEAR TlME-INVARIANT COMPOSITE SYSTEMS
Therefore, whether or not a system is completely characterized by its transfer 1. The parallel conr
function matrix can be checked from the number of state variables of the 91(S) +92(S) ifano
system. Ir a system is an RLC network,2 then the number of state variables is 2. The tandem conr
equal to the number of energy-storage elements (inductors and capacitors); 92(s)9 1 (s) if and 01
hence an RLC network 2 is completely characterized by its transferlunction 3. The feedback cor
matrix if and only if the number oi energy storage elements is equal lO the degree characterized by !
oi its transJer-Junct ion matrix. Consider now two RLC networks SI and S2, canceled by any z
which are completely characterized by their -transfer-function matrices Gl (s) Pfoof
and G2(s), respectively. The number of energy-stora~e elemellts in any c~m
posite connection of SI and S2 is clearly equal to 15G l (s) +15G 2(s). Let G(s) 1. It is obvious tha
be the transfer-function matrix of the composite connection of SI and S2. 15 9 < bg l +1592. 1
Now the composite s),stem consists of(15 G1(s) +15G 2(s)) energy-storage elements;
hence, in order for G(s) to characterize the composite system completely, it is
necessary and sufficient to have 15G(s) =15G l (s) +15G 2(s). This is stated as a
theorem. We show now tha
15 91 +(592. We pr
TheQrem 9-1 there is at least o
there is a comm(
Consider two systems SI and S2, which are completely characterized by their assumption that ti
proper transfer-function matrices Gl (s) and G2(s), respectively. Any composite there is a common
connection of SI and S2 is completely characterized by its composite transfer tion that 91 is irre(
function matrix G(s) if and only if in common, then
il connection of S 1 a
2. The proof of this 1
.... 3. The transfer funct
This theorem can also be verified from the dynamical-equation descriptions
of systems. Recall from Section 3-6 that the state space of any composite
connection of SI and S2 is chosen to be the direct sum of the state spaces of
SI and S2; consequently, the dimension of the composite dynamical equation
is the sum of the dimensions of the dynamical-equation descriptions of SI and By the irreducibili
S2- Hence Theorem 9-1 follows directly from Definition 9-1 and Theorem 6-2. have D 2 and N 2.
In order to apply Theorem 9-1 we must first compute the transfer-function if and only if D 2 al
matrix of a composite system. This is not desirable, particularly in the design
of feedback control systems. Hence the conditions in terms of G1 and Gzfor G Example 2
to characterize completely the composite connections of SI and S2 will be Consider the tandem
studied. We study in this section only single-variable systems. The multi
variable systems will be studied in the next section.
The transfer function of a single-variable system is a scalar, and its degree
is just the degree of its denominator if the transfer function is irreducible. We as shown in Figure 9-.
assume in this section that all transfer functions are irreducible; that is, their
denominators and numerators are coprime.
Theorem 9-2
Consider two single-variable systems SI and S2, which are completely charac
terized by their proper rational transfer functions 91(S) and 92(S).
re completely charac
ldg z (5).
Consider a specü
g2(S) = k, where k is a
by gI (s), the transfer I
(b)
Figure 9-3 Tandem connection of 5 1and 52' which ís not characterized completely by
=9291 = 1/(5 + 1).
g(5)
always characterizes 1
Hence the composite transfer function
A A A 1
g(s) = g2(S)g I(S) =--1
S+
9-3 Controllabil
does not completely characterize the tandem connection. This can be seen by of Composite Sy
applying a unit step input to the composite system; although the output of the
tandem connection is bounded, the output of SI increases exponentially with Let gj(s) = N¡(s)D j- I (s)
time, as shown in Figure 9-3(b). III N j and Di, Theorem (
SI and S2 is completE
Example 3 D 2 (s) are coprime. l
Consider the feedback connections shown in Figure 9-4. In Figure 9-4(a), the g2(S)g¡(s) if and only
pole of the transfer function in the feedback path is canceled by the zero of the coprime. In this secti
transfer function in the forward path. Hence the transfer function of the composite systems. B
feedback system does not completely describe the feedback system. Indeed, 6-8, on strict system ec
its transfer function is A system can be de
s-1 matrix in fractional fo
s+l ,')-1
g(5) = - - - -
s-1 1 s+2
1 +---
s+ls-1 where .u and y are th,
Its degree is smaller than 2. descriptions describe t
If {A, B, C} is not irred
{D¡, NI}' and {pes), Q
strictly system equivale
system equivalent, and
For our problem, it tI
convenient. Hence th
We recall from (6-151
{pes), Q(s)} is left COpl
(a) (b) right coprime. These
Figure 9-4 Feedback systems. complete characterizat
.. .• _---_ .. _----
_,._~
-------~._-~-_ .. _-~-~-_ .. ---
On the other hand, although the pole ofthe transfer function in the forward
path of Figure 9-4(b) is canceled by the zero of the transfer function in the feed
back path, the transfer function of the feedback system still completely charac
terizes the feedback system. Its transfer function is
A s +1
lIlI
g(s) = (s _ 1)(s +2)
-l
Consider a special case of the feedback system shown in Figure 9-2 with
g2(S) = k, where k is a real constant. Since there is no pole in g2(S) to be canceled
g
by ¡(s), the transfer function
A
g(s)= [ A ] ¡A
1 +kg[(s) - g[(s)= N[k
acterized completely by D[ + N[
always characterizes the feedback system completely.
Let G¡(S) = Nr¡(s)Dri ¡(S) = D¡¡ ¡(s)N¡¡(s) be, respectively, right and left co
O,.¡(s) and D..z(s) are l.
prime fraction of (;¡(s). Then system Si can also be described by
nection is controllable
D,.¡(s)
[ - N,.¡(s) O
1][ - t(S)] [ O]
u¡(s) - -Yi(S)
i = 1,2 (9-3) Ir we use (9-4). then
has a full rank for every s in e Because of the block triangular form, the O"o(s)
second matrix has a full rank for every s in e if and only if the matrix
[D,.¡(s) - D..z(s)] has a full rank for every s in e or, following Theorem G-8',
r
- N,.¡(s)
----Ó-----=
SIZ
u
,-.,------------
I
+
.'UI. 'YI Uz Y;
I .
s, . .
S,
IL _
ly, right and left co O"I(S) and 0"2(S) are left coprime. Hence we conc1ude that the paraHel con
'ibed by nection is controllable if and only if Or[ (s) and 0,.2(S) are left coprime.
lf we use (9-4), then the system matrix of the parallel connection is given by
1,2 (9-3)
= OII(S) O: NII(S)]
O 012(S) : N I2 (S)
j¡,l] (9-5)
nection is observable if and only if OII(S) and 012(S) are right coprime.
H).
Iil
The roots of det O"j.,(s) or det O¡¡(s) are ca)led the potes of G¡(s) (see Appendix
lf the poles of G I (s) and those of G 2 (s) are disjoint, then the matrix
irreducible by assump
ible. In order to obtain [Orl(s) 0"2(S)] and the matrix [Oíl(S) D;2(S)]' have a full rank for every s in t:
;tem equivalence trans (why?). Hence a sufficient condition for the parallel connection to be con
trollable and observable is that GI(s) and (;2(S) have no pole in common.
This condition, hüwever, is not a necessary condition (see Problem 9-4).
Or[(S)
O
-O'2(S): 0J
O,.z(s) : 1
If (;I(S) and (;2{S) are 1 xl rational functions, then this theorem reduces to
the following: The parallel connection is controllable and observable if and
=--Ñ:I(s) -- ~-Ñ'~2(;) : 0 only if theír denominators have no roots in common. This provides a different
proof of statement 1 ofTheorem 9-2. By combining the conditions in Theorem
lock row of the system
9-3, we have the necessary and sufficient conditions for (;\(s) + (;2(S) to charac
lability and coprimeness
terize completely the parallel connection of the two systems.
TI equivalence (Theorem
:cription of the parallel Tandem connection. For the tandem connection of SI followed by S2
shown in Figure 9-6, we have q¡ = P2, u(s) = M] (s), an 2 \s) = YI (s), aod y(s) = Y·/el.
~~2_(~)_ ~ ?-I (9-6)
Using lhese reiations and (9-3), we can obtaín
D'2{S) : lJ
ock triangular form, the
and only if the matrix (9-8)
following Theorem 0-8/,
Sl2
r----------~--l
I I
u IU¡ Y2Ly
I I
IL I
--J
to describe the tandem connection. Note that Yl(S)=U 2 (s) is a part of the fol1owing three pain
pseudostate of the tandem connection. It is clear that the tandem connection and N'2(S), or D,.¡(s) ,
is controllable if and only ir. for every s in e, the matrix
Let G;(s) be a q¡ x
o N'i(S)D,:¡ l(S)= Di! ¡(5)
O'2(S) (9-9)
called the po/es of G¡(
o '3!in (Pi, q J, where p ,
has a ful1 rank for every s in C. By adding the second block row to the third Gj(s) (see AppendixH
block row,3 (9-9) becomes
Corollary 9-4
O,,(s) O: O IJ A sufficient condition
O D'2(S) : 1 O
[ ------~-----~-~---- P"l ?q¡, q2 ?P2' and q
- Nrt (s) O,-z(.~)' O O
G2~S) is a transmissiol
which implies that the matrix in (9-9) has a full rank for every s in e if and only of G 2 (s)].
if [ - N,¡(s) D,z(sl] has a full rank for every s in C. Hence we conclude that
the tandem conneclion is control1able if and only if O'2(S) and N'.I(S) are, Proof
fol1owing Theorem G-8', left coprime.
Let (;l(S) = Nr¡(s)D,:¡¡
lt _
If we use (9-3) to describe SI and (9-4) to describe 52, then the tandem con
polynomial matrix ane
nection is described by
P2 at evelY s except the
Orl(s)
-NrI(s)
O
O
0:11
-1:0
Theorem 9-4 yields the
tandem connection. 1
Theorem 9-4 reduces te
o:
r - N I2 (S)N'.I(S) D I2 (S)
- - - - -ó-- - - - - - - i - ---- 0-:- ó
Hence we conc1ude that the tandem connection is controllable if and only if
O
is controllable (observaI
of g¡(s) [no pole of g¡(
characterizes complete!
N l2 (S)N r1 (s} and D l2 (S} are lefi coprir(~~:;. Ey!:1. :..:lrrülar ~lrgU;:.G'~;:;';'1 -,:"j~:. ~:;(.'~~'" :;:::'.":\;:: ~nd. (Y;;:!~i ~~ :.:~~~.-::; ):'
) This is a-lransformation of slriclsyslem equivalence_ See lhe slalemenl al lhe end ofSeclion 6-8.
CONTROLLABILITY AND OBSERVABILITY Of COMPOSITE SYSTEMS 443
following three pairs of polynomial matrices, D 11 (S) and Nds), D/[(s)D.. 2 (s)
02(S) is a part of the
and Nds), or D,.¡(s) and Nds)N¡-[(s), are right coprime. Ií
he tandem connection
Let G¡(s) be a q¡ x p¡ rational matrix and have the coprime fractions G¡(s)=
N .. ¡(s)D,~ 1(S)= D/i 1 (s)Nu(s). Then the roots of det D ..i(s) or det Du(s) are
(9-9) called the poles of (;¡(s) and those s for whích pN.. ¡(s) < min (p¡. q¡) or pI"'Us) <
mín (p¡, qJ, where p stands for the rank, are called the tral1smissiol1 zeros of
(;¡(s) (see Appendix H).
block row to the third
Corollary 9-4
A sufficient condition for the tandem connection of SI followed by S2 with
PI "2:.Q1>Q2 "2:.P2' and ql =P2 to be control1able (observable) is that no pole of
(;2(S) ís a transmíssíon zero of (;1(S) [no pole of (;I(S) ís a transmíssíon zero
of (;2(S)].
every s in e if and only
lence we conc1ude that Proof
D"2(S) and N.. 1 (s) are,
Let (;1(S) = N¡-[(s)D,:¡I(S) and G2(s) = N"2(S)D,:;I(S), where N"1(S) is ql X PI
~, thcn the tandem con polynomial matrix and Dds) is a P2 x P2 polynomial matrix. D"2(S) has rank
P2 at every s except the roots of det D,As). If no pole of (;2(S) ís a transmissíon
zero of (; I(S), we have pN"I(S) = ql = P2 at the roots of det Dds). Hence the
Jl-~y(s)i 1J (9-10)
matrix [Dds) N r1 (s)] has rank P2 at every s in e and, consequently, is left
coprime. Hence ifno pole of (;2(S) ís a transmission zero of (;1(S), the tandem
connection of S1 followed by S2 is, fol1owing TheQrem 9-4, controllable. The
observability part can be similarly proved. ' Q.E. D.
to the third block row,3
The combination of the controll~bílitx and observability condítions in
Theorem 9-4 yields the condition for G 2 (S)G 1 (s) to completely characterize the
tandem connection. If (;1(S) and G2 (s) are 1 xl rational functions, then
Theorem 9-4 reduces to that the tandem connection of g1(S) fol1owed by g2(S)
is control1able (observable) if and only if no pole of g2(S) is canceled by any zero
of gl(S) [no pole of gl(S) is canceled by any zero of g2(S)]. Hence g2(S)gl(S)
characterizes completely the tandem connection of gl(S) followed by 92(S) if
ntrollable if and only if é1.nd only if there is no pole-zcro cancellgu.c~~.' '/";';:'1,''-:''': ..: ,~::; ( ". ~l.~í
; I ";'-".'
. 'l'
~s-l
s
G 1 (o5)= _Í_ . j'
emenl al lhe end o[ Section 6-8. 05-1
444 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS Cal
then b(G 2(s)G I (S))=O < bG 2(s)+ bGI(s)= 1 and there is one pole-zero cancel Proof
lation in G 2(S)G I(s). Clearly, the existence of pole-zero cancellations in
We shall prove this ti'
G 2(s)G 1(s) does not imply the existence of pole-zero cancellations in GI (s)G 2(s),
ability. Let x be the e
as can be verified from the example. Unlike the scalar case, pole-zero cancel
Xi is the state of Si' :
lations in the multivariable case may not involve actual cancellations of poles
composite state space
and zeros.
time. Let Y2 be the Oi
Let ~(s) and ~i(S) be, respectively, the characteristic polynomials of G(s)
S J, this input will tran
and G¡(s). Ir G(s) = G 2(S)G I (s) has no pole-zero cancellation, then we have
is SJ' Conversely, if
~(s)= ~1(S)~2(S), as can be easily seen from Equation (3-63a). Ir (;2(S)(;I(S)
uI=u-Y· HenceSJ;
has pole-zero cancellations, then deg ~(s) < deg ~ 1(s) + deg ~2(S), and the roots
Now we consider
of ~1(S)~2(S)/~(S) are called the cancelled poles. If A is a cancelled pole, then
S J due to sorne unkno
from Theorem 9-4, we have the following equivalent conditions:
if - Yo is applied to S
1. [DdA) N"l(A)] and/or [DNll(~))] have no full rank. S21 is Yo· If S21 is ob
output Yo, the initial sl
12(;"
able. lfS 21 isnotobs~
D ll V,)D,z(A)] output Yo. Consequer
2. [Dll(A)DdA) N1I(A)] and/or [ Nd},) have no ful1 rank.
its zero-input response
this theorem.
3. [DdA) NdA)N"I(A)] and/or [ N D"I(A)]
')N have no ful1 rank.
dA ,.¡(A)
In general, the COI
For a direct proof of this statement, see Reference S2. See also the discussion condition of observab
of the input-decoupling zero on page 292. If al1 cancelled poles have negative two conditions are ider
real parts, then G 2(S)G I (s) is said to have no unstable pole~zero cancellation. of 91(S), then 91(s) follo
is observable. Hence \\o
Feedback connection. Consider the feedback connection of SI and S2 connection in Figure <;
shown in Figure 9-7. It is assumed that S¡ is completely characterized by e¡(s). controllable or, equiva
It is also assumed that det (1 + el
(s)Gz(s)) =1= oat s = <Xl to ensure that the feed This checks with what '
back transfer matrix is wel1 defined and proper (Theorem 3-5). Let S 17. denote We consider a speci:
the tandem connection of SI followed by S2; $21, the tandem connection of le We cal1 this a consl,
S2 followed by SI; and SJ' the feedback syslem in Figure 9-7. any new state variable,
observable, so is K fol
Theorem 9-5 \Nith a constant gain ¡¡¡:,
ano only ii SI is con1
Consider two systems Si which are completely characterized by their transfer observability are invari,
)e
matrices (;¡(s), i = 1, 2. It is assumed that det (1 + GI( <Xl 2(00)) 1= O. Then with the constant state
the feedback system S J is controllable (observable) if and only if S 12 is con is preserved under com
trollable (S21 is observable). preserved.
We discuss further !
be described by irred uc
r .
,A i , Bi ¡ C¡, E;}, where G
u + ut
rationa,1 matrices. By a
~[G¡(sr
and deg G¡(s)"" ~
where ~[.] denotes the
Figure 9-7 Feedback conneclion of SI and S2' denotes equality of polyr
CONTROLLABIUTY AND OBSERV ABIUTY Of COMPOSITE SYSTEMS 445
and deg (;¡(s) = deg de.t Dr¡(s) == deg det Dli(S) = dím A¡= 11¡
transfer matrix ofthe feedback system in Figure 9-7 is, as derived in (3-68), hence, we have
GJ(s) = G¡(s)(1 + G 2 (s)G¡(s))-¡
= N,.¡(s)D,~ ¡(s)[1 + Dii ¡(s)Nds)N,o¡(s)Dri ¡(s)] -¡
= N,o¡ (s)[Dds)D,.¡ (s) + Nds)N,.¡(s)] -¡Dds) (9-11)
The state equatior
The matrix A of its state-variable description is, as derived in (3-64),
__ [A¡-B¡Y 2 E 2 C¡ -B¡Y2C2 ]
AJ ( 9-12)
B 2 Y¡C¡ A 2 -B 2Y¡E¡C 2
with Y¡ =(1 + E¡E z )-¡ and Y z =(1 + E 2E¡)-¡. Ir the feedback system is not
completely characterized by GJ(s), then +
+
L\.[ GJ(s)] det (s 1- A J) lts A matrix is identic
and deg GJ(s) < dim AJ=n¡ + nz
Theorem 9-6
where +
denotes not equal modulo a nonzero constant factor. Conversely, if
the equalities hold, then the (n¡ + n2)-dimensional dynamical equation de Consider the system in
scription of the feedback system in Figure 9-7 is controllable and observable. O. Then we have
Note that we have
L\.(H(s))
L\.[ GJ(s)] +det [Dds)D,.¡(s) + Nds)N,.¡(s)]
because the factorization in (9-11) is not necessarily irreducible.
The controllability and observability of a system depends on the assignment
of the input and output. Although the system in Figure 9-7 may be uncon
trollable or unobservable, if we assign the additional input and output as shown
in Figure 9-8, then it is easy to show, by using the argument in the proof of
Theorem 9-5, that the feedback system is always controllable and observable
from the input [r'l r~]' and the output [Y'I y'zJ. Similariy, the system is
always controllable and observable from the input [r') r~]' and the output
[u'¡ u~J. We study now the transfer matrix H(s) from [r'l r'J' to [u'¡ u'zJ.
From u¡(s)=r l (S)-Y2(s)=r,(s)-G 2(S)U 2 (s) and lJ 2 (s)=r 2 (s)+YI(s)=r 2 (s)+ where ~ denotes the
G¡(s)u¡(s), we have factor.
!.....;; I ~:. J
+ . [. ~1
fl
H(s)=
- G ¡ (s)
Figure 9-8 Feedback system with additional input and output. which can be easily veri
CONTROLLABILITY AND OBSERVABILITY OF COMPOSITE SYSTEMS 447
(9-11)
The state equation of the system in Figure 9-8 can be readily computed as
ved in (3-64),
(9-12 ) [ ~IJz
X
= [Al - B I Y ~EzCI
BzYIC[
- B I Y zC, ]
Az-BzYIEICz
[Xl]
Xz
Theorem 9-6
factor. Conversely, if
rnamical equation de Consider the system in Figure 9-8 wilh lhe assumpl ion det [1 + G[( ex:;. )G 2 ( ex:;. JJ =F
)l1able and observable. O. Then we have
We remark first that the condition det [1 + GI (ex:;. )(;z( ex:;.)] =F Oensures, as shown
in Section 3-6, the properness of al! elements of H(s). Without the assumption,
H(s) may have poles at s = ex:;. and the theorem may not hold. Consider the
right fraction
[
='
0 .. [(.1)
O
0][ ~
Ods)
o,[(s)·.N,.z(S)]-1
N,.¡ (s). Uds)
which can be easily verified. The fraction is right coprime as canbeseen from,
tpul.
448 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
l
S if the zero-state res po;
O"I(S) Nd )] O
-N'I(S) Ods) -+ -Nr¡(s) asymptotically stable,
0"1 (S) O Or¡ (S) However, ir a system
O Ods) O asymptotic stability ir
Consider two lineé
and the right coprime assumption of N..¡(s)O,~ I(S) (Theorem G-8). Thus, we
characterized by thei
have established, following (6-189), the equality in (9-13c).
transfer matrix G(s) (
Again by direct verification, we have
stability or the compe
H(s) = [ 1 N,.z(S)O,:-/(S)]-1 asymptotic stability (
- Oil 1(s)N l1 (s) 1 system is known to \;
1 O][1
- [O O,.z(s) - N l1 (s)
NdS ) ]
0 11 (s)O,z(s)
1 [1 O]
O O'I(S)
1 N,z(S)] ]
osite systems from e
~(H(s))=det [ -N (s) DI 1(s)Ods) =det[Ol1(s)O,z(s)+N l1 (s)N,z(s) function matrices. W(
l1 For the parallel an
where we have used, in the last step, the formula simple. The parallel
stable if and only ir 5
det[~ ~]=det[-dP-l ~][~ ~]=det[~ V-~p-IU] asymptotic stability is
(3-62) and (3-63), we ~
=det P det (V - QP- 1U) nection of SI and 52 a
assertion follows.
This establishes (9-13a). The equality ~[H(s)] ~ det (sI - A f ) follows directly
For the feedback
from (6- 189) and the fact that the state-variable equation is controllable and feedback system may
observable from the input [r'l r~]' and the output [U'1 u~]'. back system may be un
From (9-13a), we have
variable case and then
det {Ol1[I + 0 11 l(s)Nl1 (s)N"2(S)Or2 1(s)]O,z{s)}
totically stable if the response due to every initial state approaches the zero
state as t-+ oo. A system is bounded-input-bounded-output (BIBO) stable
if the zero-state response due to every bounded input is bounded. If a system is
'S)l asymptotically stable, then the system is also BIBO stable, but not conversely.
However, if a system is completely characterized by its transfer matrix, then
(s) asymptotic stability implies and is implied by BIBO stability.
Consider two linear time-invariant systems SI and 52 which are completely
:)[em G-8). Thus, we characterized by their transfer matrices G1 (s) and G2 (s). If the composite
). transfer matrix (;(s) of any connection of S, and S2 is computed, the BIBO
stability of the composite system can be determined from (;(s). However, its
asymptotic stability cannot be determined from (;(s) unless the composite
system is known to be completely characterized by (;(s). It is not a simple
task to compute composite transfer matrices, especially, in the feedback con
nection. Even if the composite transfer function is obtained, it is still quite
tedious to check whether or not the composite system is completely charac
ence wehave terized. Therefore, it is desirable to be able to determine the stability of comp
osite systems from GI(s) and (;2(S) without computing composite-transfer
function matrices. We shall study this problem in this section.
For the parallel and tandem connections of S, and S2' the problem is very
simple. The parallt::1 or tandem connection of S, and S 2 is asymptolically
stable if and only if S, and S 2 are asymptotically stable. The condition for
asymptotic stability is that all eigenvalues of A have negative real parts. From
(3-62) and (3-63), we see that the eigenvalues of the parallel or tandem con
nection of SI and S 2 are the union of those of S, and those of S2- Hence the
assertion follows.
[- A f) follows directly
For the feedback system, the situation is much more complicated. A
.on is controllable and
feedback system may be stable with unstable subsystems; conversely, a feed
"2, J'. back system may be unstable with stable subsystems. We study first the single
variable case and then the multivariable case.
Ods)}
Single-variable feedback systems. Consider the single-variable feed
O"2(S)
back system shown in Figure 9-9. The transfer function ofthe feedback system is
Q.E.D.
g (s) = : ~ (:; = i'f ' Q ?JI;s;
ng det(sI - A f ) and is f 1 +g,(S)g2(S) D,(s)D 2(s) +N,(s)N 2(s)- Df(s)
there is no distinction where g¡(s) ~ N¡(s)/D¡(s) for i = 1, 2 and N f and Df are assumed to be coprime.
;:s to Recall that every transfer function is assumed to be irreducible; therefore, when
¡(s) we speak of g¡(s), we mean gf(S) = Nf(s)/D¡(s). It was shown in Theorem 9-2
Thebr€m.9.-7.
:fore it is important to
~. recall from Chapter 8 Consider the feedback system shown in Figure 9-9. lt is assumed that S, and
;e of a system is asymp- 52 are completely characterized by their proper transfer functions g,(s) and
450 LINEAR TlME-INVARIANT COMPOSITE SYSTEMS
system is asymptoti.
N ¡(s)N 2(5) have nega
Because of possil
D¡(s)D 2(s) + N¡(s)N z(
the theorem is only
Consequently, the S1
Figure 9-9 A feedback system. asymptotically stable
gzCs). Itisalsoassumedthatl +g¡(<Xl)92(<Xl) 7'=0. Letg¡(s)= N¡(s)/D¡(s). Then
the feedback system is asymptotically stable if and only if all the roots of This theorem car
D¡(s)Dz(s) + N¡(s)N z(s) have negative real parts. The condition is sufficient but proved here, howeve
Prom (9-15) we see t
not necessary for the systern to be BIBO stable.
independent of wheth
Proof tion of the feedback
°
First we note that the condition 1 + g1 (<Xl)9 2 (<Xl) -=/=- will ensure the properness
of gf(s). Let x = [x'¡ X'2J' be the composite state vector of the feedback
whether the feed ba(
g
I + ¡(<Xl )g z (<Xl) = O, e
real parts, the thec
system, where X¡ is the state vector of Si, and let x = Afx + bfu, y = efx +efU be (- S2 + s + 1)/S 2 , 92 s,
A ( •
the dynamical equation describing the feedback system [see Equation (3-64)]. (- Sz + s + I )/(s + 1) \1
Then the characteristic polynomial, det (s 1 - A f), of Af is a polynomial of N1(s)Nz(s)=(s+ 1) h,
degree n¡ +nz, where degD¡(s) =n¡, i = 1,2. input sin (2 will excitt
Nowwe show that ifl +g¡(<Xl )g 2(<Xl )7'=0, then deg [D¡(s)Dz(S) + N ¡(s)N2(S)] = ofthe feed back system
n¡ +n2' In other words, the term sn, -r n 2 in D¡(s)D 2(s) will not be canceled by 0, Theorem 9-7 does r
any term in N¡(s)N 2(s). Let If yz(s) in Figure 9-'
b· sn¡ +b· sn¡-¡ + ... output feedback systt
g.(s) = 'o I¡ controllable and obse
1 a¡Osn, +an sn ¡ ¡ + ...
implied by BIBO stab
Then we have
Corollary 9-7
· [1 +A ()A (')]-1 +b¡ob 20 _Q¡OQ20+ b lO b 20 (9-14 )
l 1m 91 s 92 S -
Consider the feedback
,~O() Q¡OQ20 Q¡OQ20
D¡(s)D 2 (s) +N¡(s)N2(s)=aIOa20Sn,+n2 + ... +bIOb20Sn,+rt2 + ... that S ¡ is completely
and
=(QIOQ20 +b lO b 20 )sn, +n, I + kg ¡( <Xl) -=/=- O. Then
(9-15 }
stable if and only if al
+terms of lower degrees
From (9-14) and (9-15) we see immediately that if 1 +g¡(<Xl)92(<Xl) 7'= 0, then
deg [D¡(s)D 2(s) + NI (s)N 2 (s)] = n¡ + n z = deg det (sl- A f ). Using this fact and Multivariable feed
the relation case is much more ca
Consider the multivari
g¡(s)=c¡(sI-Af)-¡bf+ef=d (; A ) [cfAdj(sI-A f )bf ]+ef S ¡ is assumed to be COI
et s - J
G¡(s). The assumptioJ
connection. Let Gf(S
NI (s)Dz(S)
u
we conclude that
det (sI - A f ) =k(D¡"(s)D 2(s) + N¡(s)Nz(s)) (9-16)
for sorne constant k. In other words, the set of the eigenvalues of A f and the
set of the roots of D¡(s)Dls)+ N ¡(s)N2(S) are the same. Hence the feedback Figure 9"10 A multiv,
STAB1UTY OF FEEDBACK SYSTEMS 451
system is asymptotically stable if and only if all the roots of D¡ls)D 2ls) +
N ¡ ls)N 2ls) have negative real parts.
Because of possible cancellations between N¡(s) and D 2(s), some roots of
D¡(s)D 2(s) + N¡(s)N 2(s) may not appear as poles of g¡(s); hence the condition in
the theorem is only a sufficient condition for the system to be BIBü stable.
Consequently, the system in Figure 9-9 can be BIBü stable without being
asymptotically stable. Q.E.D.
,) = N ¡(s)/D ¡(s). Then This theorem can in fact be deduced directly from Theorem 9-6. It is
Iy if all the roots of proved here, however, by using explicitly the condition I +g¡(00)g2(00)+0.
dirion is sufficient but
From (9-15) we see that (9-16) holds if and only if I +g¡(00)g2(00)+0; it is
independent of whether the (n¡ + n 2 )-dimensional dynamical equation descrip
tion of the feedback system is irreducible or not. It is also independent of
whether the feedback system is completely characterized by g¡(s). If
ensure the properness I +g¡(00)g2(00)=0, even if all roots of D¡(s)D 2(s) + N¡(s)N 2(s) have negative
ctor of the feedback real parts, the theorem stili does not hold. For example, if g¡(s)=
I-b¡u, Y=cfx +e¡u be (-s2+ s +1)/s2, g2(s)=I, then we have l+g¡(00)g2(00)=0 and 9¡(S)=
:see Equation (3-64)]. (-S2+ S+ I)/(s+ 1) which is improper. Although the root of D¡(s)D 2(s)+
~ is a polynomial of N¡{s)N 2(s)=(s+ 1) has a negative real part, the application of the bounded
input sin {2 will excite a term of 2{ cos {2, which is not bounded, at the output
)D2(S) + N ¡ (s)N 2 (s)] = ofthe feedback system. Thus without imposing thecondition I + g¡ (00 )g2(00) +
ill not be canceled by 0, Theorem 9-7 does not hold in general.
If y2(S) in Figure 9-9 is a consta~nt k, the feedback system reduces to a constant
output feedback system and is, as discussed in the previous section, always
controllable and observable. In this case, asymptotic stabiiity implies and is
implied by BIBO stability.
Corollary 9-7
(9-14 )
Consider the feedback system shown in Figure 9-9 with {¡](s) = k. lt is assumed
Q2Ü
that S ¡ is completely characterized by its proper transrer runction yI (.1') and
b 2 0s"¡ +"2 +
I +kg¡(oo)+O. Then the feedback system is BIBü stable and asymptotically
stable ir and only if all zeros of 1 +kg¡(s) or, equivalently, all roots of D,(s)+
s J,,; t'll (.;-;) h8_V~ nega tiv~ ·~·~2.1 :-::'.:;"X í::::..
N 2(S)) , .
nvalues of A¡ and the
Hence the feedback
(9-16)
~r
Figure 9-10 A multivariable feedback system.
452 LINEAR TlME-INVARIANT COMPOSITE SYSTEMS
l
where p is the order (
"
G¡(s) =
s-1
1 s +1
-2
J (;¡(s) of order i; for
elements of (;¡(s). TI
s +1 Now 6.¡(s) is the least
lt can be easily verified that F(s) is at most equal
minors of (;¡(s) folle
2S
,1¡(S), is the least common denominator of al! the minors of G¡(s). lf G¡(s) has
be tne dynamical-eq ua
the coprime fraction N¡(s)D¡¡(s), then we also have 6.¡(s)=kdetD¡(s), for
some nonzero constant k.
Note that the dynam
Theorem 9-8 its knowledge is not
assumption that S¡ is
Consider the feedback system shown in Figure 9-10. It is assumed that S ¡
control!able and obst
is completely characterized by its proper rational matrix G¡(s) and that
6.¡ (s) of G ¡ (s) is equal t
det [1 + (; ¡ (00)] :¡b O. Let (; ¡ (s) = N,.¡ (s)D,:¡ ¡ (s) = D¡l~(s)Nl1 (s) be coprime frac
tions and 6.¡ (s) be the characteristic polynomial of G¡ (s). Then the feedback
system is asymptotical!y stable and BIBO stableif and only if al! the roots of
By the substitution 0[.
any of the three polynomials 6.¡(s)det[I+·G¡(s)],det[Ó,.;(s)+N rl (s)J and
det [D1¡(s) + N¡¡(s)] have negative real parts. I
STABILITY OF FEEDBACK SYSTEMS 453
where p is the order of eleS) and (Xi is the sum of all the principal minors of
eleS) of order i; for example, (Xp(s)=det (;[(s), (X[(s)=sum of the diagonal
elements of eleS). The verification of (9-19) is straightforward and is omitted.
Now ~[(s) is the least common denominator of all the minors of G 1 (s), whereas
F(s) is at most equal to the least common denominator of aH the principal
minors of (;¡(s) following (9-19); hence we conc1ude that ~¡(s) is divisible
without remainder by F(s) and that ~¡(s)det [1+ (;¡(s)] is a polynomial.
Theorem 9-8 can be deduced from Theorem 9-6; however, because of its im
portance, we shall prove it directly in the following.
Note that the dynamical equation FE is used only implicitly in the proof;
its knowledge is not required in the application of the theorem. Now the
assumption that S ¡ iscompletely characterized by e ¡(s) implies that FE is
1t is assumed that S ¡ controllable and observable. Consequently, the characteristic polynomial
natrix G¡ (s) and that ~ ¡(s) of e 1(s) is equal to the characteristic polynomial of A¡; that is,
IN l ¡ (s) be coprime frac
:s). Then the feedback ~¡ (s) ~ ~[G¡(s)J = det(sl - A;) . (9-22)
only if all the roots of By the substitutionof ti 1.= u - y ¡, y = y[ and the use óf the identity
:t[D, ¡(s) + Nrt(s)J and
• 1-'---(1 +Etl-[E[ =(1 +Etl-¡
"-.
the dynamical-equation description 01' the feedback system can be obtained as Example 2
Since X(s) and Y(s) - X(s) are both polynomial matrices, Theorem G-8 implies
the rightcoprimeness 01' D'1(s)+ N.. ¡(s) and Nr¡(s). Conseq.uently, G¡(s)=
N,.¡(s)[D d (s) + N .. 1(s)J -¡ is a coprime fraction and ó[ Gf(S)J~ det [D,.¡(s) +
. N,.¡(s)] ~ Ó¡(s). Hence the feedback system is BIBü and asymptotically stable 11' we do not cancel o
if and only if all the roots 01' detCD,.¡ (s) + N"I (s)] have negative real parts. half-plane zero. Thw
Q.E.D. not stable.
STABIUTY Of FEEDBACK SYSTEMS 455
(9-23a)
(9-23b)
le fol1owing Theorem
) and asymptotical1y
A
GI(s)=
l l
Determine the stability of the feedback system in Figure 9-10, with
-s
s-1 s
5+1
-2
s +1
¡)-le l have negative
f the eigenvalues of The characteristic polynomial of G¡(s) is oÓ¡(s)=(s-l)(s +1). It is easv to
oÓ¡(s)det [1 + (;1(5l]. verify that det [1 + (;¡(s)] = -1. Hence the feedback system is stable if'and
on1y if al1 the roots of oÓ ¡ (s) det [1 + (; ¡(s)J = - (s - 1)(s + 1) have negative real
parts. This is not the case; hence the feedback system is neither BlBO stab1e
nor asymptotically stable. ~
Example 3
(9-27)
Consider the feedback system shown in Figure 9-10 with
1 1
·e Dr¡(s) + N,¡(.s) and
istence of polynomial s-0.5
1
s -0.5 J
lt is straightforward to verify that
["l(S)J=
condition is sufficient but not necessarily for the system to be BIBO stable.
Ir Gz(s) = K, a p x q constant matrix, the condition is necessary as we11 for the "z(s)
system to be BIBO stable. 1
By direct computatior
This theorem can be deduced directly from Theorem 9-6. It can also be
proved by using its dynamical eguation description to establish, as in (9-26),
H(s) = [~" Gz(é
-G 1 (s) t¡
L1 (s)
f
~ det (si - A ) = L1 (s)L1 (s) det [1 +A G1(s)?z(S)]
f 1 z (9-28) [1" + G2 (s)d
det [1 + G1(oo)Gzioo)] = [ ¿l(s)[i p +G z(
riecause the cnaracter;
referred to Rererence 49. Similar to the proof in (9-14) and (9-15), it is possible egual to det (si - A f ),
to establish, by assuming the column reducedness or Drz(s) and the row re sufficient for the systen
d ucedness of 0 11 (s), that that there exist G 1 (s) ¡
deg det [Dll(s)D,.z(s) + N 11 (S)N.. z(s)] = deg ¿l(S) + deg Gz(s) = n1 + nz stable but the fourth er
instability of the en trito
In the design of fee
+ section ofwe11-posedn(
pairs to be proper. SiJ
al! possible input-outp¡
is not BIBO stable, an
input at rz wi11 cause th
in practice to reguire e'
Figure 9~11 A multivariable feedback system. input-output pairs.
STABILITY OF FEEDBACK SYSTEMS 457
and consequently.
(9-29 )
-0.5) if and only if
det [1 + GI(co)GAcoJ] +O
To save space, this assertion wil! not be proved here. We note that the validity
4'(.1') = (s +1.5)(s +0.5), of (9-29) is independent of whether 01' not the feedback systern is completely
~able. 11 characterized by
G ¡(s) = [1 + G¡(s)Gz(s)] - I GI(s)
ppears in the feedback = [1 + Di¡ I(s)N II (s}N,.z(s)D,:;I(SJ] - I Di¡l(s)N lI (s)
lerally not completely = Dds)[DlI(s)Dds) + NII(s)N,.z(s)] - I NlI(s) (9-30)
situation here is more
sure 9-10. The factorization in (9-30) is not necessarily irreducible; hence general!y we have
~[G ¡(s)] f det [DI I(s)Dds) + NII(s)Nds)] ~ det (s 1- A¡) (9-31 )
For this reason, the condition in Theorem 9-9 is only sufficient for G¡(s) 01' the
, assumed that SI and system in Figure 9-11 to be BIBO stable. Ir Gz(s) = K, a p x q constant matrix,
oper transfer m~trices the feedback system is, as discussed in the previous section, completely charac
(co)J '1'=0. Let G¡(s) = terized by G¡(s) and ~[G¡(s)] ~det [DII(s)+ N¡¡(s)KJ; hence the condition in
ns and ~i(S) be the Theorem 9-9 becomes necessary as wel! for G¡(s) to be BIBO stable.
,tem is}syrnytotical!y Now we consider the same system in Figure 9-11 but with additional input
t [1 + GI(s)GZ(s)J, or and Olltput as shown in Figure 9-8. As derived in (9-12), we have
sative real parts. The
n to be BIBO stable. GZ(s)J-I [~I(S)J
::essary as wel! for the Iq 1'z(s)
1 By direct computation, we have
Consider again the system in Figure 9-8 with [r'l ra as the input and
[Y'I y'2T as the output. Clearly, we have
YI(S)J (¿ Hy(s)
[ Y2(S) -
[~I(S)J [~I(S)~I(S)J= [G
= 1 (S)
I'~~lll
--j
k
~"
----... ces) ~i g(s) =.V(s)/D{sl l---;--..
1- ~L! I !
L __._. .=.J
FigurE> 9-12 Unily feedback syslem.
1 +C(s)g(s)
From g¡(s) +g¡(s)C(s)9(s) = kC(s)g(s), and using gj(s) = N¡(s)/O¡(s) and g(s) =
(9-36) N(s)fD(s), we can write C(s) as
In the following, we discuss a method of computing C(s) which will not Equating the coeffici
involve any undesirable pole-zero cancel1ations. However, we must paya
price for achieving this. The zeros of N¡(s) of the resulting system cannot be
arbitrarily chosen; they are dictated by the zeros of C(s) and g(s). If we wish to with
control the poles as well as the zeros of gAs), then we must choose a different
Do DI ...
N¡(s)
D¡(s)
kNc(s)N(s)
Dc(s)D(s) + Nc(s)N(s)
(9-39) SnI=l:~ );::::
-- - --------
In the arbitrary pole placement, we study only the polynomial equation O O .
(9-40) O O .
In the fol1owing, we shall study the condition on D(s) and N(s) under which a This is a set oflinear ~
set of solution {Dc(s), Nc(s)} exists in (9-40) for any DAs). We sha1l also study between the polynorr
the degrees of Dc(s) and Nc(s) in order to achieve arbitrary pole placement. 4 hence it is permissit
By arbitrary pole placement, we always assume implicitly that complex con Sm consists of m + 1 1
jugate poles appear in pair. Otherwise, the compensator will have complex coefficients of g(s) anl
coefficients and cannot be implemented in practice. right by one column.
The polynomial equation in (9-40) is called the Diophantine equation in application of Theor
honor of Diophantus of the third century in Reference S139. 1t is ca1led the (9-44) has a solution
compensator equation in Reference S34. This is the most important equation S", has a full column
in the remainder of this chapter. ·Instead of studying· it directly (see Problem rank is that 2(m + 1);
G-14), we shall translate itinto a set of linear algebraic equations. From the compensator to achi
set of algebraic equations, we will then develop all properties of the Diophantine G-4, we have that S"
equation. The set of algebraic equations can also be used to compute the In fact, from the proo
compensator C(s). Let us define for m:2:n -1 ifand on
D(s)=Do+Dls+··· +D"sn Dn-l=0 (9-41 a) the following theoren
N(s)=N o +NIs + +N"s" (9-41 b)
Dc(s)=D co +Dels + + Dc",s"' (9-42a) Theorem 9-10
Nc(s)=N co +N c1s + +Ncmsm 19-42b \
and define
(9-43)
with deg N(s)~deg D
where Di, Ni, Dci> N ci , and F¡ are real constant, not necessarily a1l nonzero. The for every DAs) of degI
substitution of (9-41) to (9-43) into (9-40) yields and only if D(s) and !'
F o + F IS + ... + Fn+msn+m =(D co + DeiS + +Dcmsm)(D o + D1s + +Dnsn)
+(N co + NeiS + +Ncmsm)(N o + N IS + +N"s") This theorem statl
to meet (9-46). Noth
4
00 Matbemalieally, Ihe finding of polynomial solutions in (9-40) is equivalent to the finding of inleger
solutions x, y in ax +by = f, wherea,b, and r are giOven integerso This is a topie in number theory 5 Ir D(s)and N(s) are note.
or continued ofraetlonso only if the grealest como;
DESIGN OF COMPENSATORS: UNITY FEEDBACK SYSTEMS 461
· C(s) which will not Equating the coefficients of the same power of s yields
:ver, we must paya [DeoNeo:Dc1Nc1:"':DemNem]Sm=[Fo F¡ Fz (9-44 )
Img system cannot be
ld g(s). If we wish to with
I 1St choose a different
i N(s) under which a This is a set of linear algebraic equations. There is a one-to-one correspondence
We shalI also study between the polynomial equation in (9-40) and the algebraic equations in (9-44);
lry pole placement. 4 hence it is permissibk to study the former by using the latter. The matrix
y that complex con Sm consists of m + I block rows; each block row has two rows formed from the
JI' will have complex coefficients of g(s) and can be obtained by shifting its previous block row to the
right by one columno lt is clear that Sm is a 2(m + 1) x (n +m + 1) matrixo The
phantine equation in application of Theorem 2-4 to the transpose of (9-44) reveals that for every F,
,139. 1t is called the (9-44) has a solution {De;, N ei , i = 0,1,. o. ,m} if and only if pSm = n +m + 1 or
t important equation Sm has a fuIl column rank. A necess~ry condition for Sm.to have a fuIl column
jirectly (see Problem rank is that 2(m + l);::::n +m + 1, 01' m;:::: n -1. Hence the smallest degn~e o(a
quationso From the compensator to achieve arbitrary pole placement is n -1. From CoroIlary
es of the Diophantine G-4, we have that S"-1 is nonsingular if and only if D(s) and N(s) are coprime.
lsed to compute the In fact, from the proof of CoroIlary G-4, we have that Sm has a full column rank
for m ~n - 1 if and only if D(s) and N(s) are coprimeo Hence we haveestablished
(9-41 a)
the foIlowing theoremo s
O
(9-41 b)
(9A2a) Theorem 9-10
(9-42b 1
:) +D¡s + .. +D"s")
o
:) +N¡s + .... +N,¡S") This theorem states only the conditions for the existence of DC<s) and NC<s)
'. to meet (9-46). Nothing hasbeen said regarding the properness of Ne(s)/De(s).
This question will be answered in the fol1owing two theorems by removing the The employment
condition deg D¡(s) < n + m fram Theorem 9-10. set of linear algebraic
whether g(s) is strictl
Theorem 9-11 required compensato
Consider the feedback system shown in Figure 9-12 with g(s) = N(s)/D(s) and
deg N(s) < deg D(s) = n. Then for any D ¡(s) of degree n +m, a praper compen Example 1
sator C(s) of degree m exists so that the feedback system has transfer function Consider a plant witl
N(s)Dj 1(s)Ne(s) if and on1y if D(s) and N(s) are coprime and m e. n - 1.
Proof
If we show that D em in (9-44) is different from zero, then this theorem follows
Let C(s) be chosen as
directly from Theorem 9-10. If g(s) is strictly proper, then N" =0. Conse
quently, fram the last column of (9-44), we have Dem=Fn+,JD n. Hence if
F n +m is different from zero, so is Dem . This proves that the compensator C(s)
is proper. Q.E.D.
Then the equation in
Theorem 9-11'
Consider the feedback system shown in Figure 9-12 with 9(S) = N(s)/D(s) and
deg N(s)~deg D(s)=n. Then for any D¡(s) of degree n+m, a strictly proper
compensator C(s) of degree m exists so that the feedback system has transfer
function N(s)Dj l(s)Ne(s) if and only if D(s) and N(s) are coprime and m c.n.
Since the square ma
Proof solution exists in (9-4~
For a proper 9(S), the solution {Dei, NeJ in (9-44) exists if and only if N(s) and sator of degree 1 alwa:
D(s) are coprime and m c..n -1. However, if m = n -1, the solution is unique, (9-48) may become
and there is no guarantee that Dem i=- O, and the compensator may become the s01ution in (9-48)
or C(s) = (3s + 1)/(-(
improper. If m c.n, the number of unknown {De;, NeJ in (9-44) is larger than
the number of equations. Since the last row of Sm, with m c.n, is linearly function.
Now we choose
dependent on its previous rows (see Section G-2), we may choose N e ", = O.
For this choice, we have De", = Fn+,JDn- Hence the compensator is strictly
proper. This establishes the theorem. Q.E.D.
rems by removing the The employment of Theorems 9-11 and 9-11' is very simple. We form the
set of linear algebraic equations in (9-44) with m = n - 1 or m = n, depending on
whether g(s) is strictly proper or proper. Its solution yields immediately the
required compensator. This is illustrated by an example.
h g(s)=N(s)/D(s) and
Example 1
-m, a proper compen
has transfer function Consider a plant with transfer function
mdm¿n-l.
A +1
S2
g\s) .2
s +2 s_ 2
. this theorem follows
Let C(s) be chosen as
then N n = O. Conse
= Fn+",/D". Hence if
the compensator C(s) (9-47 )
Q.E.D.
Then the equation in (9-44) becomes
In Theorem 9-11, n - 1 is the smallest degree which a compensator must scribed by the q x 1
have in order to achieve arbitrary pole placement. In other words, for every
D ¡(s), a compensator C(s) of degree n - 1 always exists so that the unity feedback
system in Figure 9-12 has D¡(s) as its denominator. lf Df(s) is prechosen, then
it may be possible to find a compensator of degree smaller than 11 - 1 to achieve
the designo For example, if Df(s) is chosen as
D f(S) = 4s 2 + 2s + 1
in the example, then we have
[DeO Neo]So = [1 2 4] where D(s) the least (
and DeO = 1, NeO = 3. In other words, we can find a compensator C(s) =f of D(s)
degree O to achieve this particular pole placement. Hence, if the degree of a N(s)
compensator is extremely important, we_ may proceed as follows: First, we
where Dí are constar
form So and check whether the chosen D f(S) of degree n is in the row space
find a compensator
ofS o. lfyes, the design can be completed by using a compensator of degree O.
number of poles of 1
lfnot, we formSI and increase the degree of Df(s) by one to Df(s) = 15 f (s)(s +k).
assigned. Furtherml
Ir Df(s) is in the row space of SI for sorne k, and if the pole s + k of this k is
small as possible. FI
acceptable, then the design can be accomplished by using a compensator of
degree 1. Ir D¡(s) is not in the row space of SI' then we must increase m by 1
and repeat the process. For a study of this problem, the reader is referred to
Hence the transfer rr
References S97, S184, S21O, and S223.
equal to
The larger the degree of a compensator, the larger the number of parameters
available for adjustment to achieve design purposes. Ir the number of para
meters is larger than the minimum required for arbitrary pole placemeni, the
Let us write the com~
spared parameters can be used to achieve other design objectives such as the
assignment of zeros or the minimization of sensit ivity functions. The interested 1
C(s)=-I
reader is referred to References S25 and S83 and Problem 9-29. DJ5)
In Theorems 9-11 and 9-11', if D¡(s) is chosen as a Hurwitz polynomial, then with
the feedback system is stable. Hence we conc1ude that every plant with a
proper transfer function can be stabilized by using a compensator with a proper
transfer function in the unity feedback configuration shown in Figure 9-12. where Dí are scalars
19-50) and (9-53) i'1tc (
5~g-og~e-Üli'1pijJr~ or sfirugoe-QlLf[Put case. In this section vveüiscüss i.·i-lé¿C.s¡5~·,;
rE; f(s) = N(s),
of compensators to achieve pole placement for single-input multiple-output and
multiple-input single-output systems. The general case will be postponed to
= [De(s
the next subsection. Because N(s) and N e (
Consider the feedback system shown in Figure 9-13(a). The plant is de and N(s)Ne(s) is a q x
Di(' .
(al (b)
Figure 9-13 Unity reedbacksystemswith single-input or single-output plant. 6 Note lhat, beca use of pos
DESIGN OF COMPENSATORS: UNITY FEEDBACK SYSTEMS 465
N~(s)
D~(s)
where D(s) the least common denominator of all elements of C(s). We assume
LS follows: First, we
where Di are constants and Ni are q x 1 constant vectors. The problem is to
! is in the row space
find a compensator with a proper transfer matrix of degree m so that n +m
pensator of degree O. number of poles of the feedback system in Figure 9-13(a) can be arbitrarily
o D fes) =15 f(s)(s +k). assigned. Furthermore, the degree m of the compensator is required to be as
,ole s +k of this k is small as possible. From Figure 9-13(a) we have
19 a compensator of
nust increase m by 1 u(s) = (1 + C(s)G(s))- I C(s)r(s)
reader is referred to Hence the transfer matrix of the overall feedback system in Figure 9-13(a) is
equal to
umber ofparameters
the number of para- G fes) = G(s)[1 +C(s)G(s)] - IC(S) (9-52)
pole placement, the Let us write the compensator C(s) as
bjectives such as the
we d iscuss the design Gfes) = N(s)D- I(s)[1 '+ D e- I (s)Ne(s)N(s)D - I (s)] - ID; I (s)Ne(s)
multiple-output and
= [DcCs)D(s) + NcCs)N(s)] -1 N(s)NcCs) (9-55)
'Viii be postponed to
Because N(s) and Ne(s) are q x 1 and 1 x q vectors, Ne(s)N(s) is a 1 x 1 matrix
1). The plant is de- and N(s)Ne(s) is a q x q matrix. Hence Gfes) is a q x q rational matrix. Define 6
Df(s) = DcCs)D(s) + Ne(s)N(s) (9-56)
Hence the problemof pole placement reduces to the solving of Equation (9-56).
'This equationis a generalization ofthe Diophantineequation in (9-40). Similar
. to (9-40), we shall translate it into a set of linear algebraic equations. Let
(9-57)
(b)
~-Olltpllt plan!. 6 Note thal, because of possible pole-zero cancellalions, nol all rools óf D¡(s) are poles of G¡(s).
466 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
The substitution of (9-51), (9-54), and (9-57) into (9-56) and equating the co Proof
efficients of the same power of s yield
The design problen
F,,+mJ ~ F application of Theo
(9-58) there exists a set of
if the matrix Sm has
with definition of r¡ and t
Do DI D" O O
} one block row
No NI N" O O
----------------------------
O Do ... O Hence the conditior
D"_I D"
S",= O .. . ... O (9-59) This is the case if al
No N"_I NI!
---------------------------- (Corollary G-14). 1
---------------------------- any {F;}.
O O O Do DI Dn lf (;(s) is strictly
O O O No NI ...
N" solved as De", = FIl + m
is proper and if m =
We call the rows formed from {DJ D rows and the rows formed from {Ni} m;::: v, then the last q
N rows. Then every block row in (9-59) has one D row and q N rows. The Consequently, we ID;
matrix S", consists of m + 1 block rows; every block row is the shifting to the Fn+mlD". Hence the
right by one column of its previous block row. The matrix S,. is c1early a proof of this theorem
(1 + q)(m+ 1) x (n + 1 + m) matrix. Now we search the Iinearly independent
rows ofS,. in order from top to bottom by using, for example, the row-searching The application o
algorithm discussed in Appendix A. Then as discussed in Theorem G-13, aH in (9-50) have no COI
D rows in Sm wiH be linearly independent, and sorne N rows will be Iinearly form S", by using the
dependent on their previous rows in Sm' Let r¡ be the number of Iinearly independent rows in
dependent N rows in the i + 1 block row ofS",. Then because of the structure algorithm. Once v is
ofS",. we have O:::::; ro:::::; 1'1 :::::; :::::;q. Let l' be the integer such that,.o:::::; 1'1 :::::; •.. :::::; (;(s) is strictly proper.
r,. _ 1 < q and 1', = 1',+ 1 = = q. We cal1 l' the row index of C(.~). Then l' is Its solution yields th
the largest row degree of A(s) in any left-coprime fractíon of C(s) = A - 1 (s)B(s) solution of (9-58) can
with A(s) row reduced [see (G-80) and (G-81)J or the observability index of any
irreducible realization of (;(s) (dual ofTheorem 6-6).
Consider the feedback system shown in Figure 9-13(a) with the plant described ability index, the con
by a q x 1 strictly proper (proper) rational matrix C(s) = N(s)D-I(s) with hinges on the search
deg D(s) =n. Then for any D¡(s) of degree n +m, there exists a 1 xq proper S,. is a (l +q)(m + 1)
(strictly proper) compensator C(s) = D; I(s)Ne(s) with deg De (s) = m so that the we need (1 +q)(m + 1
feedback system has q x q transfer matrix N(s)Dj I(s)Ne(s) if and only if D(s) may start from the SI
and N(s) are right coprime and m;::: v - 1 (m;::: v), where v is the row index of note that aH remarb
(;(s), or the observai?i1ity index of any ir~educible realization of (;(s), or the For example, for a giv,
largest row degree of A(s) in any left~coprime fraction of (;(s) = A -1(s)B(s) with degree smaller than v
A(s)row reduced. of a compensator, it
other design objectivt
Dual to Theorem
71[ q = 1, this theorem reduces to Theorems 9-11 and 9-1\' and l' is equal.to 11. system shown in Figl
DESIGN OF COMPENSATORS: UNITY FEEDBACK SYSTEMS 467
Gl(s)= ·lS_1_
largest column degree of A(s) in any right-coprime factorization of C(s) = A +1 5+
B(s)A -les) with A(s) column reduced. III
o
s +1
The polynomial equation arises in this theorem is of the form then it can be readil
Every 1 x por q xiI
D¡(s) = D(s)De(s) + N(s)Ne(s) (9-61 )
(;(s) = tjJ - 1 (s)N(s)N e(.
Note that De and N e are on the right-hand side of D(s) and N(s), rather than the matrices, then C(s) i:
!eft-hand side as in (9-56). Equation (9-61) can be solved indirectly by taking its elements gij(S) of G(.I
transpose to become the form of (9-56) or solved directly as fol1ows: Using Consider the cycl
the coefficient matrices of D(s), N(s), Dcls), Ne(s), and D¡(s), we form the linear
algebraic equation
, ,
Do No O O O O
Deo , Deo Fo
DI NI , Do No , O O
Neo Neo Fl
,o
Del Del F2
,o O O We see that if al =
N el , , N cl
Tm ~ DII Nn - l··
, Do
°1
No (9-62) nomials of C2 (s) and
N" Dn -
: 1 1 ,
,
O O Dn N" DI NI teristic polynomials, l
D em Dem
the characteristic poI:
N em , N em Fn + m shows that in any liJ
O O O O D II
N II
to right. Let /l be the least integer such that the last p N columns of TI' are all
Consider a q x p cycl
linearly dependent of their left-hand-side columns. The /l will be called the
and 1 x q real consta
column index of C(s). It is equal to the controllability index of any irreducible
realization of C(s) or the largest column degree of the column-reduced A(s)
in any right-coprime factorization of C(s) = B(s)A -l(S). The proof ofTheorem
where 6.(. ) denotes tt
9-12' is similar to the one of Theorem 9-12 and will not be repeated.
.general proper ratiOnlJ.l matrices. We extend it .first to a special class of rational polynomial of A. Comp.
matrices;called cydic ratiónaJmatrices, ando thentothe general case. matrix in Prob\em 2-45.
Consider a q x pproper rational matrix G(s). Let tjJ(s) be the least common 9 Compare this theorem w
l
í_1_
~1 -J
:torization of C(s) =
:le form
I
A
l S
GI(s)= _1_
s
+1
+ 1
A
Gz(s)=
s+1
_1_
s
s +2
s +1 s +1
then it can be readily verified that Cz(s) and C,(s) are cyclic, but C¡(s) is noL
Every 1 x p or q x 1 proper rational matrix is cy;lic. If C(s) can be expressed as
(9-61 )
C(s) = lf¡ -1(s)N(s)NcCs), where N(s) and Nc(s) are q x 1 and 1 x p polynomial
N(s), rather than the matrices, then C(s) is cyc1ic (why?). Ir no A. appears as a pole of two or more
idirectly by taking its elements gij(s) of C(s), then C(s) is cycli<l (Problem 9-14).
Iy as fol!ows: Using Consider the cyc1ic rational matrix Gz(s). We form
), we form the linear
o
o
Fo
F¡ .
(;z(s)a
.
~
-
Gz(s) [al]
az
= l a~ :~z J
azs +(2az
s +1
+al~
Fz
We see that if al = -az, then Gz(s)a = [O -az]' and the characteristic poly
(9-62) nomials of Gz(s) and Gz(s)a are differenL For a with al i= - az, their charac
teristic polynomials, however, are eq ua!. Recall that for cyc1ic rational matrices,
the characteristic polynomial is eqlial to the minimal polynomia!. This example
Fn + m shows that in any linear combination of the columns of Gz(s), the chance of
canceling a pole is very smal!. This is true in general as state in the following
theorem.
Theorems G-14 and
Tm in order fram lelt
~olumns of T" are all
Consider a q x p cyc1ic proper rational matrix C(s). Then for almost al! p x 1
Ji. will be called the
and 1 x q real constant vectors t ¡ and t z, we have
ex of any irreducible
:olumn-reduced A(s) Li[C(s)] = Li[C(s)t¡] = Li[tzC(s)] (9-63)
he proof of Theorem
where Li(') denotes the characteristic polynomial of a rational matrix. 9
repeated.
~
N ,IS)t¡]
'\. _ ~ N 2 (s)t 1
GI.\)t¡ - ,',
'I'(s) :
Nq(s)t l
have no common factor. Hence l{I(s) is the charactcristic polynomial of G(s)t¡.
This establishes 8[ G(s)] = 8[ G(s)t lJ for the case where aH roots of l{I(s) are real.
ir sorne of ;'j, i = 1, 2, ... , in, are not real, the preceding proof stil! holds if we
use N(A¡} +N(},t). This completes the proof of .1[G(s)J =.1[G(s)t¡J.J:"óe i:
other part of the theorem can be similarly proved.
Now we show that (9-63) holds for almost al! tI in ~p. For easy visualization,
we assume p = 2 and consider only tI with IIt l l1 2 = 1. Then from the preceding
proof, we see that at most 2m points on the unit circle of ~ 2 will not meet (9-63),
the rest ofthe unit circle (infinitely many points) will meet (9-63). Hence almost
every vector tI in ~2 has the property 8[(;(S)] =8[G(s)t l J. Hence ifa vector
tI is. chosen arbitrarily or generated randomly, th~ probability of having
8[G(s)] == 8[G(s)tlJ is almost équal to 1. This establishes the theorem. Q.E.D.
The cy~licity of (;(s) is essenltal in this theorem. Ir (;(5)' is not cyclic, the
theüi-em d~esnot h61d in general (tr"y G 4 (s) in Problem 9'-13).. UsingTheorem
9-13, we can now extend the design procedure in Theorems9-12 and 9-12' to
cyclic rational matrices. Figure 9-14 Design (
DESIGN Of COMPENSATORS: UNITY FEEDBACK SYSTEMS 471
Theorem 9-14
Consider the feedback system shown in Figure 9-14 with the plant described
by a q x p cyc1ic strictly proper (proper) rational matrix G(s) of degree n. The
compensator is assumed to have a p x q proper (strictly proper) rational matrix
C(s) ofdegree m. lf m ~min (jl-l.l' - 1) [m ~min (}l, v)], then aH n +m poles
of the unity feedback system can be arbitrarily assigned, where jJ. and vare,
respectively, the controllability and observability indices of any irreducible
realization of G(s} or the coJumn index and row index of G(s).
:nts of G(S); N(s} is a
s). We assume that
Proof
ption of (;(s) implies
)ts of ¡Jt(s). First we Since Gis) is cyc1ic by assumption. there exists a f' x l constant vector tI such
that Ll[ G(s)] = Ll[ G(s)t 1 J. Lel us wrile l he q x I ral ional matrix e(s)t I as
GIs)t¡ = NIs)D-11s)
u\! space of N(A.¡) has Then Theorem 9-12 imp\ies lhe exislence of a 1 XI/ proper rational matrix
ds, the set of vectors t (:(s) = D c- 1 (s)N,(s) with deg Ós) = 111 ~ l' - I if G(.~) is strict\y proper. such that
¡)t = O isa linear space all n +m roots of
will be discussed. we
L, is at most a straight D Il~) = D.(s)D(s) + Nrl-~)N(s)
:s on which N(A.i)t = O can be arbitrary assigned. Now we show that the roots of Df(s) give the poles
aperty of the feedback system in Figure 9-14(a). Indeed, from Figure 9-14(a), we have
h(s) = (:(s)C(s) =C(s)[r(s) - (;(s)t¡h(s)]
,----------1
¡ ¡
e I I Y
I
1
polynomial of G(s)t¡. I
L
I
-.J
roats af 1/1 (s) are rea\.
)roof sti\\ ho\ds if we
sl] = Ll[G(S)t¡J. The iI
(al
¡:< ar easy visualization,
~n from the preceding
2 will not meet (9-63),
The substitution of G(s)t¡ = N(s)D- ¡(s) and C(s) = D,-¡(s)N,(s) into (9-64) yields G(s) =
G feS) = N(s)D- ¡(s)[l + D,- ¡(s)N,(s)N(s)D- ¡(s)] -¡ D c- ¡(s)NcCs) is proper (strictly pro~
= N(s)[D,(s)D(s) + Nc(s)N(s)] -¡ Nc(s) (9-65)
10
Proof
where we have used the fact that Dc(s) and D(s) are l xl polynomial matrices.
From (9-65) we conclude that the roots of D feS) give the poles of (9-65). Hence We show that the roc
the qxp compensator defined by cts)=t¡C(s)=D;¡(s)t¡Nc(s) can achieve distinct for almost all ]
Now we show that the observability index of G(s)t¡ is equal to the one of
C(s). If G(s) is factored as G(s) = A -l(s)B(s), where A(s) and B(s) are left coprime where Gi, i = 1, 2, ... ,n
and A(s) is row reduced, then the observability index of G(s) is equal to the tion of 3.(s) with respe(
largest row degree of A(s) (see Theorem 6-6). Consider G(s)t 1 = A -¡(s)B(s)t¡.
The condition ~[G(s)] = Ll[G(S)t1] =det A(s) implies that A(s) and B(s)t¡ are 3.' (s)
left coprime. Hence the observability index of G(S)t1 is also equal to the largest If ~(s) has repeated roc
row degree of A(s). This establishes that the observability index of G(s) and sufficient condition for
that of G(s)t 1 are the saine. Hence we have deg C(s) ¿ v - 1, where v is the
observability index of G(s) or C(s)t 1. Since cts) = t¡ C(s) is cyclic, we have
deg cts) = deg C(s) ¿ v - 1. This completes the proof of one part of the theorem.
The rest can be similarly proved. Q.E.D.
det O
From (9-65), we see that the transfer matrix from r to y is of the form
N(s)Dc-¡(s)Nets), as in the vector case. However, the N(s) and Nc(s) in this
design are not unique; they depend on the choice oft¡. Although the degree of o
compensators in Theorems 9-12 and 9-12' are minimal to achieve pole place (See Appendix G). W
ment, the degree in Theorem 9-14 may not be minima\. In other words, it may nomial of kij. There is
be possible to design a compensator of degree less than min (J1- 1, v -1) to of {kij} as a vector in
achieve arbitrary pole placement for a q x p cyclic proper rational matrix. clear that the solution,
What is the minimum degree seems to be a difficult problem. almost every k¡j, we ha'
With Theorem 9-14, we can now discuss the design of compensators for rrhis Ís Si!:.~.712.r t0 th2J
general proper rational matrices. The procedure consists of two steps: First are distinct.) Hence G
change a noncyclic rational matrix into a c}'clic one and then apply Theorem If G(s) is strictly prc
9-14. Consider a proper rational matrix G(s). Let Ll(s) be its characteristic the condition for G( "
polynomia\. We claim that if all roots of Ms) are distinct, then G(s) is cyclic. det (1 + C( 00 )K) '" O. fl
Let if¡(s) be the least common denominator of all elements of G(s). Then we proper, so is G(s) " for al
have Ll(s) = if¡(s)h(s) for sorne polynomial hes). If all roots of Ll(s) are distinct,
thell we have Ll(s)=if¡(s)k for sorne constant k, forAif¡(s) must contain every root
of G(s). Hence G(s) iscyclic. Note that a cyclic G(s) may have repeated poies. With this theorem,
Hence the condition thatall roots of ~(s) are distinct is a sufficient but not placement for general
necessary condition for G(s) to be cyclic. This properiy ~v.i11 be used to establish gajn output feedback to
that every noncyclic proper rational matrix can be transformed into a cyclic .
one by introducinga constant giliri feedback from the outpUt tothe input, as
stated in the following theorem. ¡o The proof is identical lo lh
DESIGN OF COMPENSATORS: UNITY FEEDBACK SYSTEMS 473
Theorem 9-15
Consider a q x p proper (strictly proper) rational matrix G(s). Then for almost
} (9-64) every p x q constant matrix K, the q x p rational matrix
'lAs) into (9-64) yields G(s) =[1 +G(s)K]-IG(s)=G(s)[1 +KG(S)]-l
1D c- 1(s)Nc(s) is proper (strictly proper) and cyclic.
(9-65)
Proof'O
polynomial matrices. A
oles of (9-65). Hence We show that the roots of the characteristic polynomial, ~(s), of G(s) are all
5)t l Nc(s) can achieve distinct for almost all K. Let
~(s)=aOsn +als"-l + ... +a"
s equal to the one of
d B(s) are left coprime where a¡, i = 1, 2, ... , n, are functions of all elements, kij, of K. The differentia
G(s)t l =A -1(s)B(s)tl'
ill be used to establish . gain outP\lt feedback to make G(s) = [1 + G(s)K] -IG(S) cyclic. We then apply
sformed :intoa cyclic
utput to the input, as
10 The proof is identical lo the one of Theorem 7-6.
474 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
C(s) and the configuration reduces to the unity feedback system shown in X D
Figure 9-15(b). = [D(s).
By a similar manipulal
Theorem 9-16
can be computed as
Consider the feedback system shown in Figure 9-15ib) with the plant described
Gfes) = [D(s)De(s) + N(~
by a q x p strictly proper (proper) rational matrix G(s) of degree n. The com
pensator C(s) is assumed to have a p x q proper (strictly proper) rational matrix From (9-66) and (9-67),
of degree m. If m =?:min (J.1-1, v -1)[m =?:min (J.1, v)], then al1 n +m poles ll of have the same set of po!
the unity feedback system in Figure 9-15(b) can be arbitrarily assigned, where J.1 assigned by a proper ch
and vare, respectively, the control1ability and observability indices of any C(s) + K, the poles of n
irreducible realization of G(s), or the column index and row index of C(s). Now we c1aim that d,
be an irreducible realiza
Proof x = Ax +Bu, y = ex + (J
Hence we have deg de
First we show that if C(s) = C(s) + K and if C(s) can be written as C(s) =
proved by using Theo rel
De-l(s)tl Nis), where t l is a p x 1 constant vector and NcCs}is a 1 x q polynomial implies that the parallel (
matrix, then the poles of the system in Figure 9-15(a) and those in Figure 9-15(b) Hence we ha ve deg det «
are the same. The transfer matrix of the feedback system in Figure 9-15(a) is
What remains to bé
indices of C(s) and (;(s)
11 Because of possible pole-zero cancellations, not all n + m of these roots will be the poles of the for almost every consta
resulting system. However, for convenience, we shall call all of them poles. The strictly proper N(s) and D(s) are right c
part of this theorem was first established in Reference S26 in the state variable approach and in lability index of any irree
Reference S51 in the transfer-function approach. degree of D(s) (see Theol
'" = l\!(s)lDJ-
G(s) !r·
Since N(s) and D(s) are
(;(5)
A
and Ves) such that X(s)O
¡----------l
I + I y
X(s)[DI
Hence Theorem G-S imp
I
I
column degrees of D(s) al
I for all K if N(s)D -les) is
I
I I
proper. .Hence the contr
L -.J
facl <;an also be proved b
can besimilarly proved.
(al (b)
The tr:ansfer matrix G
Figure 9-15 Design of compensator.
Unlike the cases of single
DESIGN OF COMPENSATORS: UNITY FEEDBACK SYSTEMS 475
:actly a unity feedback The substitution of G(s)=D- 1(s)N(s) and C(S)=D;I(S)t 1 N e (s) yields
10Ies of GAs), not the GJI (S) = {D - 1 (s)[D(s)De(s) + N(s)KDe(s)+ N (s)t 1 Nc(s)]D c' 1 (s)} -1
ion of C(s) and K into X D-I(s)N(s)D e- 1 t 1 N e(s)
ack system shown in = [D(s)De(s) + N(s)KDe(s) + N(s)t 1 Nb)]' 1 N(S)tl Nls) (9-66)
ith the plant described G J(s) = [D(s)De(s) + N(s)KDe(s) + N(s)t 1Ne(s)] -1 N(s)[t 1NAs) + KDe(s)] (9~67)
)f degree n. The com From (9·66) and (9-67), we conclude that the systems in Figure 9-15(a) and (b)
)roper) rational matrix have the same set of poJes. 11 Since the poles in Figure 9-15(a) can be arbitrarily
11
;en all n+m poles of assigned by a proper choice of C(s) and K, we conclude that by choosing (:(s) =
'arily assigned, where J1 C(s) +K, the poles of the system in Figure 9-15(b) can be arbítrarily assigned.
ability indice~ of any Now we claim that deg det C(s) = deg det Ces). Let x= Ax + Bu, y=Cx +Eu
row index of G(s). be an irreducible realization of C(s); then we have deg det C(s) = dim A. Clearly
x = Ax + Bu, y = Cx +(E +K)u is an irreducible realization of (:(s) = C(s) +K.
Hence we have deg det C(s) = dim A = deg det (:(s). This fact can also be
be written as C(s) = proved by using Theorem 9-3. If we :vrite K = KI.¡-I = 1; lK, then Theorem 9-3
s) is al x q' polynomial implies that the paral1el connection of C(s) and K is controllable and observable.
those in Figure 9-15(b) Hence we have deg det (:(s) = deg det C(s) +deg det K = deg det C(s).
em in Figure 9·1~(a) is What remains to be proved is that the control1ability and observability
indices of (;(s) and (;(s) = [1+ G(s)K] - 1G(s)= G(s)[I+KG(s)] - 1 are the same
for almost every constant K. Ir G(s) is factored as G(s) = N(s)D-I(S), where
'oots wíll be the poles of the N(s) and D(s) are right coprime and D(s) is column reduced, then the control
m poles. The strictiy prop~r
lte varíable approach and ID
lability index of any irreducible realization of 6(s) is equfll to the largest column
degree of D(s) (see Theorem 6-6). Using this fraction, (;(s) becomes
G(s) = N(s)D-1(s)[I +KN(s)D- 1 (s)] -1 = N(s)[D(s) +KN(s)]-\
Since N(s) and D(s) are right coprime, there exist polynomial matrices X(s)
the form of N(s)D¡ 1 (s)NAs) with scalar D f(s). However, from the design We compute
procedure, we may conclude that G¡(S) in (9-67) iscyclic. See also Problem 9-16.
We recapture the design procedure in the following:
(;(s) = [1 + G(s)K
Step 1. Find a K such that G(s) = (1 + G(s)Kr 1 G(s) is cyclic.
The solutions of
(;(s)
.
l ~ ~1
= si,
.O O s
lt is a noncyclie rational matrix of degree3. Rather arbitrarily, we choose the are D co = 1, Mo ==1, N~
constant ·output feedback gain as 9-IS(b) is given by .
C(s) = C(s)t 2 -t
DESIGN OF COMPENSATORS: UNITY FEEDBACK SYSTEMS 477
2"- F'I+V-¡] Hence the characteristic polynomial of (;(s) is S3 + 3, which is equal to the
minima1 po1ynomial. Hence (;(s) is cyclic. The system has three inputs and
two outputs; hence the controllability index is probab1y smaller than the
observability indexo We choose t 2 = [1 O]. Then we have
... 1
t 2 G(s)=-,-[s+1 s(s+l) 1]
;+ ... + Nc(v-IlS .-1] s' +3
I and L1[ (;(s)] = L1[t 2 (;(s)]. We form, as in (9-62),
], then we must modify
earch its linear1y inde
: the 1east integer such T=l~~~ ~
o 0:0 1 O b
le column index or the
lear a1gebraic equation
1:O O O J
t 2 + K. The matrix T o has a full row rank; hence J1 = 1, and the degree of compensator
:gree n. Ir its control is J1 - 1 = O. Let
1engths, then we have
NJO]
[ N~o
Hence if p '2:q, we use C(S)=f
n {J1-1, v -l} without cO N3
co
wh¡;re [be; superscrlfils denote the componems oí Nc(.s).iile n +UJ. - l} = j
The solutions of
-1] r2 ., . , 1J
O = 3 O
1 O 1
478 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
G¡(s)=[I+G(s)C(s)]
A N _lA c-.J
G(S)C(s)=(s+1)2(S+2) O (s+1)(s+2) G/(s) = N(s)I
= N(s)[
Its poles are -1, -1, and -2. We note that G¡ is cyclic (verify it). I
Define the polynomis
To conclude this subsection, we mentíon that although we can place arbitrary
the poles, the resulting G¡(s) is always cyclic. 12 In other words, the structure
Then we have G¡(s) =
of GAs) is restricted; we have no control over it. This restriction will be re Given D(s) and N(s) a
moved in the next subsection. This is the matrix ver
(9-40), we shall transla
Multivariable case: Arbitrary denominator matrix assignment. 1 3
In this subsection, we study the design of compensators to achieve arbitrary
denominator matrices. If we can assign an entire denominator matrix, then
certainly we can achieve arbitrary pole assignment. Hence this problem
accomplishes more than the pole assignment discussed in the previous subsec We also write
tion. Consequentiy, the degrees of compensators required for arbitrary denomi
nator matrix are generally much larger than the ones required for pole assign
mento
Consider the unity feedback system in Figure 9-16. The plant is described and
by a q x p proper rational matrix G(s). The compensator to be designed is
required to have a p x q proper rational matrix C(s). Let G ¡(s) be the transfer The substitution ofthe
matrix of the overall system. Then we have
G ¡(s) = [1 + G(s)C(s)] - 1G(s)C(s)
= G(s)C(s)[1 + G(S)C(S)]-1
where
= G(s)[1 + C(s)G(s)] - 1C(S) (9-68)
The first equality is obtained from y(s) = G(s)C(s)[r(s) - y(s)]; the second one Do D1 .
from e(s) = res) - G(s)C(s)e(s); and the third one from o(s) = C(s)[r(s) - G(s)u(s)]. No NI .
(Verify them. Compare also with Theorem 3-3.) In the single-variable case, S
if we assign both poles and zeros of the unity feedback system, then the design
= [ O Do'" [
O No'" l'
----- ---------
ni
13
12 Compare with Method 1 of the multivariable case in the design of state feedback.
l
O
O
The matrix Sm has m +
'"
...
O
The plant is described and D¡(s) = F o + F 1s+ F 2S 2 + ... + F mw S '" +1' (9-73)
ator to be designed is
~t G¡(s) be the transfer The substitution of these into (9-70) yields
[Deo Neo: Dc¡ Nc¡: ... : D em Nem]Sm=[F o F 1 F2 ... Fmw ] ~ F
(9-74)
where
(9-68)
r
linearly depertdent N rows in the (i + 1)th block. Then because of the structure
ofS m, we have ro S;rl S;'" S;r m S;q. Let v be the least integer such that r~ =q.
14 The integer J1 is thelargest colunm degree of D(s). It is different from deg det D(s) = n. In the
single-input case, however we have J1 = n.
480 LINEAR TIME-INV ARIANT COMPOSITE SYSTEMS
where n ~ L;~~ (q - rj) is the degree of C(s). We call v the r~w index of C(s).
It is the 1argest rowdegree of A(s) in any leftcoprime fraction of G(s) = A -l(s)B(s)
with A(s) row reduced. It is also equal to the observability index of any irreduci This equality holds if
ble realization of C(s). See Equation (G-8l) and Theorem 6-6. O(s) and N(s) are rig
rank of Sk and the n
Theorem 9-17 from v - 1, we conc1'
only if D(s) and N(s).
Consider a q x p proper rational matrix with the fraction G(s) = N(S)O-I(S).
If Df(s) is of colt
Let /l¡, i = 1, 2, ... , p, be the column degrees of O(s), and let v be the row index zero columns in the
ofC(s). Ifm 2.v -1, then for any D¡(s)withcolumndegrees m +/l¡, i = 1, 2, .. . ,p, zero columns coincid
or less, there exist Dc(s) and N/s) of row degree m or less to meet
must be inside the ro'
o fes) = 0c(s)D(s) + Nc(s)N(s) (9-74), or equivalentl
completes the proof e
if and only if D(s) and N(s) are right coprime and O(s) is column reduced. 15
This theorem stal
Proof meet (9-70), but state~
To study this questic
Let /l =max {/l¡, i = 1,2, ... , p}. Since C(s) is proper, the column degrees of proper or proper. Th
N(s) are equal to or smaller than the corresponding column degrees of D(s). 9-11 and 9-11'). Ho
Consequently, the matrix coefficient matrix D/L (
Before proceeding
S _[0 0 01 °/L-1 O/L]
0- No NI N/L-I NI'
¿;
has at least a total of = 1 (/l - /l¡) zero columns. In the matrix SI' some new and
zero column will be created in the rightmost block column; however, sorne
zero colulnns in So 'NHl cHsappeó~r froro. § 1- I-Iencc the nurnb~r oi zere cJI~;,,_:..L_' lheorem 9-Hs
in SI remains to be
p p
Consider a q x p stric
C(s) = N(s)D- 1 (s). L
¿ (/l - /l¡) = P/l - ¿ /l¡
i= 1 i=l v be the row indeJ
In fact, this is the minimum number of zero columns in Si, i = 2,3, . . .. Let m¡ 2. v - 1 (m¡ 2. v) for :
8. _1 be the matrix S. _ 1 after deleting these zero columns. Since the number of
columns inS miS(/l + 1 +m)p, the number of columns in 8.- 1 is equal to
p p
(/l +v)p-(P/l- L /lJ=vp +¿ /li (9-77 ) exists and is nonsing
(;;1 i=l meet
15 This theorem reduces to Theorem 9-10 for the single variable case. In the single-variable case.
we have J1 = v=degD(s) = 11, and the search of the row 'index becomesunnecessary. if and only if D(s) and
DESIGN OF COMPENSATORS: UNITY FEEDBACK SYSTEMS 481
The rank of S"-1 is clearly equal to the rank of S"-I' Hence we have, from
(9-76),
for m?:. v - 1 A comparison of (9-78) with (9-77) reveals immediately that S,. _1 has a full
column rank if and only if
(9-76 )
p
(9-82)
In the single-variable case,
.les unnecessary. ir and only if D(s) and N(s) are right coprime and D(s) is column reduced.
482 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
Proof
proper) rational mal
feedback system in I
Let m = max {mi} and J1 = max {J1;}. Consider the ith row equation of (9-74):
are right coprime ano
lim diag(s··\l,.
It is important to note that if mi;::: l' for some i, the solutions of (9-74) 01' s .... 00
(9-82) are not unique and there may be improper solutions. However, there
exists) unde~· f.b.e 2.ssunl.ption of {9-81)., at least ODe 3,,:1: ')f ~~:~.. :::y.>~: :)'.. :.;t:~s'!:I~/ exists !:-tnd lS llf)7lsi '1'; 1
proper solutions. Ir O[(s) meets (9-81), it is said to be row-column-reduced in (strictly proper) ratio~
Reference S34. Every O [(s) can be transformed, by elementary operations, r to y in Figure 9-1 7 i,
into a row-column-reduced one. are left coprime and [
Now we may apply Theorem 9-18 to the design problem.
Theorem 9-19 e
1-----
Consider a q x p strictly proper (proper) rational matrix (;(s) with the fraction
(;(s) N(s)D-l(S). Letllí, i= 1,2,.'.. ,p be thecolumn deg~ees of O(s) and let v
0;= _ _o -'/~'-J-_--l-;-/~ Dc 1 ~
be lbe row index of G(s). Let the row degrees of Dc(s} be mi, i;=; 1,2, ... , p. Ir L _
mi;::: l' ~.l (mi;::: v) for al1 i, thenfor any O [(s) with the property thilt
proper) rational matrix D: l(s)N c(s) such that the transfer matrix of the unity
feedback system in Figure 9-16 is N(S)O¡l(S)N c (s) ifand only if O(s) and N(s)
>w equation of (9-74): are right coprime and O(s) is column reduced. il
This theorem follows directly from Theorem 9-18 and Equations (9-69) and
li 2: v - 1, the resultant (9-70). Several remarks are in order regarding this theorem. We discuss only
.Jll column rank. The the case where (;(s) is strictly proper. Ir mi = v - 1, for all i, and if we choose
, has column degrees O [(s) = P(s)O(s), then the unique solution of (9-70) is Oc(s) = pes) and Nc(s) = O.
zero e1ements whose In other words, if the denominator matrix of the plant is not to be altered, the
de that for any Dr(s) compensator C(s)= O,:"l(s)N c(s) is O. This is a degenerated case. The second
most mi exist in (9-82). remark concerns the possible existence of common divisors between N(s) and
: MI(s)é M¡Js)H: 1 (s) O[(s) and between D[(s) and Nc(s) in (;r(s) = N(s)D¡ l(s)N c(s); and between
¡lady, we write Nc(s) and Dc(s) in C(s) = Oc- 1 (s)N c(s). In the first case, the pole-zero cancella
tions involve the poles which the designer has the freedom in placing, therefore
these cancellations are permitted in practice. From (9-70), we see that if O,.{s)
and Nc(s) have a common left divisor, then it is also a left divisor of D [(s).
(9-83) Hence the pole-zero cancel1ations between Dc(s) and Nc(s) involve again only
assignable poi es. The final remark concerns the well posedness of the feedback
system. Since (;(s) is strictly proper and C(s) is proper, we have 1 + (;( 00)C( ro)=
)per rational matrices.
I. Hence the unity feedback system in Theorem 9-18 is well posed. Note that
matrix of D(s) and is
ifDe(s) and Ne(s) are left coprime, thedegreeofthecompensator is ¿mi 2:p(v- 1),
) into (9-82) yields, at
which is much larger than the one required for arbitrary pole assignment.
We state the dual of Theorem 9-19 as a corollary.
(9-84 )
of proper or strictly
(strictly proper) rational matrix Nc(s)D c- 1 (s) such that the transfer matrix from
)w-column-reduced in
r to y in Figure 9-17 is equal to 1- Dc(s)D¡ I (s)O(s) if and only if O(s) and N(s)
lementary operations,
are left coprime and D(s) is row reduced.
,Iem.
The substitution of G(S) = D-l(S)N(s) and C(S) = N c(s)D c- l (s) into the first
equality of (9-68) yields
Define
(9-86)
This shows that the transfer matrix from r to y in Figure 9-17 is equal to
[1 - Dc(s)D¡ 1 (s)D(s)]. The desígn ín Corollary 9-19 hínges on solvíng (9-86).
Note that the transpose of (9-86) becomes (9-70); left copríme and row reduced
become ríght copríme and column reduced. Hence Theorem 9-18 can be
applied directly to the transpose of (9-86). Of course Equation (9-86) can also We see that Dc(s) is :
be solved directly. We use the coefficient matrices of D(s) and N(s) to form Hence if G(s) ís prc
T k as shown in Equation (9-62) and then search linearly independent columns compensator.
in order from left to right. The least integer Jl such that all N columns in the If we choose mi =
last block column of T¡.¡ ·are linearly dependent, is called the column index of
(;(s). Dual to the row index, the column index is equal to the largest column D
degree of the column reduced Dr(s) in any right coprime fraction of G(s)=
N r (s)D,:-l(S). It is also equal to the controllability index of any irreducible
then the compensato
realization of (;(s). The proof ofCorollary 9-19 is similar to the one ofTheorem
9-19 and will not be repeated.
Example 3
We gíve an example to iJlustrate the appíicatíon oí Theorem 9-i9. o....:onsiuer
the proper rational matrix Remarks are in 01
and the design of ar
~ = N(s)D- l (s) = [S2 O+ 1
G(s)
S2
S J[s2-1
+s + 1 O S2
O
-1
J-l strictly proper, the
min (Jl- 1, v - 1) (Th,
p(v -1) (Theorem 9-1
We form So, Si' ... , and search their linearly dependeni rows in order from top ment assigns only the
to bottom. For this example, we can readily obtain that v = 2. Clearly we always yields a cycL
have Jll = Jl2 = 2. Let /111 =/112 = V - 1 = 1. We choose mairix yields general
·of thecyclic and non
applicable here, the
·placement design are
denominatormatrix (
Then the compensator is the solution of of these two designs i
DESIGN OF COMPENSATORS: UNITY FEEDBACK SYSTEMS 485
uation (9-86) can also We see that Dc(s) is singular and the compensator Dc-l(s)Nc(s) is not defined.
[)(s) and N(s) to form Hence if G(s) is proper, the choice of mi = v - 1 may not yield the required
independent columns eompensator.
: all N eolumns in the Ir we choose mi = v and choose
1 the colúmn index of
to the largest column
me fraction of (;(s) =
lex of any irreducible then the compensator can be computed as
to the one ofTheorem
DJs)=l O
rs2 +3 4~s - 1)/3J NJs) =[
4S +4
O
-4(s + 1)/3J
s- +s + 1 2(s + 1)
S2
O
-1
J-l and the design of arbitrary assignment of denominator matrices. Ir (;(s) is
strictly proper, the minimal degree of compensator for the former is
min (J1-1, v -1) (Theorem 9-16); whereas the minimal degree for the latter is
p(v -1) (Theorem 9-18) or q(J1-1) (Theorem 9-19). The design ofpole assign
ows ín order from top ment assigns only the poles or the determinant of the denominator matrix and
lat v= 2. Clearly we always yields a eyc!icoverall transfer matrix. The design of denominator
matrix yields generaHy· a noncyclie overall transfer matrix. if the diseussion
of the cyclic and noncyc!ic designs in the multivariable case of Section 7-3 is
[O30J [1O 0J1 s applicable here, the transient responses of the, system obtained by the pole
t - ' s 2' + 3
ptacement design areprobablyworse than those of the system obtained by the
2
denominator matrix design with the same set of poles. A detailed comparison
of these two designs is not available at present.
486 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
1 "'1
----." '-/--~l I ---'-.,r¡
1of the compensator for i = 1, 2, ... ,p, then C(s) is proper. That is, if the pole-zera excess 17 of the
P(s) = P(s). Because ith column al P(s) is equal to 01' larger. than the largest zero-pole excess of the
9-18 reduces to a set ¡th colum~ of (; -1(S), the~ C(s)' is proper. Note that the left-hand side of (9-89)
system is decoupled. ~ also equal to the pole-zera excess of G¡(s). Ir p = 1, the zero-pole excess 01
nsator G -1(S) is equal to the pole-zera excess of G(s), and (9-89) reduces to the poIe
zera excess inequality discussed in (9-38) for the single-variable case.
(9-87)
The poles of G¡(s) are the zeras of (;(¡(s) + P¡(s). Once f3¡(s) and the degrees of
(;(¡(s) are determined fram (9-88) and (9-89), fram the assignment of the poles of
G¡(s), we can readily compute o::¡(s). Using these cx¡(s), thc compensator C(s) in
(9-87) is prape! and the unity feedback system G¡(s) in Figure 9-18 is de~oupled.
The poles of G f(S) are assignable by the designer. The zeros, p¡(s), of G f(s) are
9- J 8 win not involvE: dictatecl by the closecl right-half p!ane mots of del I''i(.», cailed [he nomninimum
phase zeros of (;(5). They are chosen as in (9-88) to avoid unstable pole-zera
cancellations. In other words, nonminimum phase zeras of (;(5) should not be
canceled and should be retained in Gf(5).
If a plant is not stable, it must be stabilized before the application of the
decoupling pracedure. We see that the decoupling is achieved by exact cancel
lations, although all stable ones. If there are any perturbations in (;(s) and
C(s), the property af decoupling will be destroyecl. Furthermore, the degree of
compensator is usually very large for an exact decoupling. Hence decoupling
is very sensitive to parameter variations and is expensive to implement.
then the feedback system achieves asymptotic tracking and disturbance rejec
Theorem 9-20
tion. In this section we shall study the design of this problem.
Consider the feedbac
-:;omplete!y characteri:
slgnai r(t) ano disturí
w(s) = N w(s)/Dw(s). Le
poles of res) and w(s).
sator with a proper 1
(a)
asymptotically stable
rejection.
Proof
[f no root óf cP,(s) is a
transfer function.l/ cjJ(s
. able (Theorem 9~2).
(b)
coprime, and there exi
Figure 9-19 Design or control systems. back system shown in
ASYMPTOTlC TRACKING AND DISTURBANCE REJECTION 489
Rejection Befóre proceeding, we discuss first the nature of the signals r(t) and w(t).
Ir r(t) and w(t) both go to zera as t -+ 00, then (9-90) will be automatically rnet if
ptrol systern shown in the feedback system in Figure 9-19(b) is asymptotical1y stable. If either r(t)
ven, the prablem is to or w(t) does not go to zera, and if we have no knowledge whatsoever about its
¡'so thatthe fuedback nature, then it is not possible to achieve asyrnptotic tracking and disturbance
oecifications. One of rejection. Hence we need sorne information of r(t) and w(t) before carrying out
I:he plant y(t) to track the designo We aSSume that the Laplace transforrns of r(t) and w(t) are given by
1;, it is not possible to
lerence 546). The best A) =.;Lro[r(t )J = -
res N,.(S)
- (9-91 )
D,.(s)
and
A
w(s) = 2 [w(t)
J = -(-)
N w(s) (9-92)
Dws
11 that if r(t) is a step
Iystem, that is, C(s)g(s) where the polynomials D,(s) and Dw(s) are known and the polynomials N,.(s)
1I in Figure 9-19(a) is and N,Js) are however arbitrary so long as res) and w(s) are proper. This is
1\:) asyrnptotical1y. In equivalent to the assumption that r(t) and w(t) are generated by
established.
x,.=A,X,. (9-93a)
y(t)
Proof
J
[f no rootof cjJ(s) is a zero of g(s), the tandem connection of the systern with
transfer. function llc/J(s) followed by 9(S) = N(s)/Q(s) iscontroIlable artd observ
able (Theore"m 9-2). Consequently, the polynomials N(Is) and D(s)cjJ(s) are
coprinie, and there exists a compensator C(s) = Nc(s)/Dc(s) such that the feeo
back system shown in Figure 9-20 is asymptotically stable (Theorems 9-11 and
490 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
Figure 9-20 Asymptotic tracking and disturbanee rejection. Figure 9-21 Asympll
ASYMPTOTIC TRACKING AND DISTURBANCE REJECTION 491
y(t)
y{t)
We see that k 1 i:
then we have r(t) - y(t) -> O as t -> oo. In order to have a finite k, we need exact, then k 1 is 2
No =1= Oand Neo =1= O. In the design of compensators, it is always possib1e to Ir k 1 is nonz~
find a C(s) with Neo =1= O. Ir Do =/=-0, we have No =/=- O if and only if <fJ(s) = s k ¡ea'. If a is po>
is not a zero of g(s). Hence we condude that, under the condition stated in output y(t) will r
Theorem 9-20, ifr(s)= l/s, il is also possible to design a feedback system, track r(t) but wit
without introducing the internal model, to achieve asymptotic tracking. with m > 1, and i
This design, however, is Dot robust. Ir there are perturbations in ?ny ;." again not possibl.
No, Do, Neo, and Deo, we do nol have r(t)-y(t)->O as t->oo. Hence the This, howevet
design which does not employ the internal model principIe is not robust. interest. In the d
We emphasize that the condition that no root of </J(s) is a zero of 9(S) is result will be evet
needed in both robust and nonrobust designs. this case, if the inl
5. From the discussion in item 4, we may condude that the condition stated in errors in impleml
Theorem 9-20 is necessary as well. If we introduce the interm!.l model and but with a finite S1
if anúoot of <fJ(s) is a zero of g(s), then the root, which is unstable, becomes a of <fJ(s), the smalh:
hidden mode and will not be affected by any compensation. Hence the 7. The design of N k
unity feedback system can never 'be asymptotically stable. Ir nointernal of the polynomia
model is employed, even though g(s) is of the form </J(s)Ñ(s)/D(s), it is straight-
forwúd to show from (9-95) and (9-96) ttiat <fJ(s) will not appear as anumer~
atar oC the transfer functions f~om w to' y and r to e; hence asymptotic Under theconditi
tracking and disturbance rejection cannot be achieved.
ASYMPTOTIC TRACKING AND D1STURBANCE REJECTION 493
6. Unlike Db), Nc(s), D(s), and N(s), the variation of the coefficients of <f;(s) is
not permitted in robust design, because tracking and disturbance rejection
are achieved by exact cancellation of the unstable modes of r(t) and w(t)
by the roots of <f;(s). In practice, exact cancellations are very difficult
and expensive to achieve; hence inexact cancellations often occur. We
study in the following the effect of inexact cancellation. For simplicity, we
assume
-(DoD co +NoNco ) ~ 1 ~
r(s) = - - ro(s)
= l/s, the application s-a
and <f;(s) = (s - a +e)<f;o(s)
where a:2:0 and e is a small real number which denotes the amollnt of
perturbation or inexact implementation of <f;(s). The transfer function
from r to e in Figure 9-20 then has the form
(9-97)
g,.(s) =(s -a +e)g,.o(s)
l zero of g(s), then we where all poles of g,.o(s) have negative real parts. The signal e(s) due to
: 1, then r(t) -y(t)-->O r(s) and w(s) =0 is equal to
¡ to be asymptotically
~ ~ 1 ~
parameter perturba e(s) = (s - a +e)g,·o(s) - - ro(s)
s-a
e internal model l/s. ~
k¡ +terms due to t he poles of g,.o(s) ~ )
choose k in (9-97) as =-- and ro(s
s-a
with k ¡
(s - a +e)
s-a
A ~ I A
use Theorem 9-10, the degree of DJ~) is deg D(s) +deg cjJ(s) -1 and the total differential eguat
degree of the compensator N /s)/Dc(s)cjJ(s) is deg D(s) + 2 deg cjJ(s) - 1. A discussion of this
different procedure is to solve the eguation
'Multivariable ca~
DJ(s) = Dc(s)D(s) + N c(s)N(s) system shown in Figl
with the constraints that Dc(s) contains the factor cjJ(s) and Nc(s)/Dc(s) is matrix G(s) factored f
proper. By so doing, the degree of compensator can be considerably q xq and q xp poly
reduced. See References S34 and S238. signal r(t) and the q )<
8. We give a state-variable interpretation of the disturbance w(t) in the p!ant
shown in Figure 9-20. The output _v(.~) is egual to and
y(s) = D - 1 (s)~¡;(s) + D - I (s)N(s)u(s) where D,.(s) and D w (!
1
Let D(S)=S"+DI_1S"- + ... +D o and N(s)=NI_IS"- 1 +N I _2 S"- 2 + q x 1 polynomial ma
... + No· Then YlI(S) = N(s)D- 1 (s)u(s) can be realized as any N,.(s) and any N,
with A and e of the forms shown in (6-38). Similarly, y",(s) = D- I(S)W(S) This is the problem e
can be realized as, by using (6-8), Following the sil
x= Ax + bww(t) y=ex (9-99) introduction of an il
There is, however, on
where A and e are identiGal to the A, e in (9-98). Hence we may combine is critical in the m u lt
(9-98) and (9-99) to yield w(t) by assuming r(t)'
shown in Figure 9-23
x= Ax + bllu(l) + bww(t) y=ex
ew(s)
From the eguation, we see that the disturbance in Figure 9-19 can be
considered as imposed on the state of the system. This eguation can be or en.(s)
generalized to
x= Ax + bllu(t) + b",w(t)
and w(t) is callcd an additiue disturbance. This is the type of disturbancc ,. + ~-----Jr""
~L1Jll'IlS;
most often studied in dynamical eguations.
, I
9. There are two types of specifications in the design of control systems. One
is calíed the steady-state performance, the other the transient performance.
The steady-state performance is specified for the response as t -> (jJ. Hence
asymptotic tracking and disturbance rejection belong to this type of
lb---
Figure 9-22 Multiv8
specification. The transient performance is specified for the response right
after the application of the reference signal and disturbance. Typical
specifications are rise time, settling time, and overshoot. They are governed
mainly by the location of the poles of the overall systems or, eguivalently,
the roots of Dc(s)D(s)<f>.(s) + Nc(s)N(s). The relationship between these poles
. and the transient performance is generally complicated. For a discussion
of this problem, séeReference S46.
10. To conclude this subsection, we remark that disturbances can be roughly
. classified as nóíse-type aÍld waveform-structured disturbances. The former
L-__,
reglJires the statisticaJ description and is studied in stochastic control
theory. See, for example, Reference S10. The latter is describable by Figure 9-23 Placem(
ASYMPTOTlC TRACKING AND DISTURBANCE REJECnON 495
.c/J(s) - 1 and the total differential equations such as the ones in (9-93) and (9-94). For an extensive
;) +2 degc/J(s)-l. A discussion of this type of disturbances, see References S121 and S122.
where O,(s) and O",(s) are q x q polynomial matrices and Nr(s) and Nw(s) are
_ 1s" - 1 +N" _2S " - 2 + q x 1 polynomial matrices. The problem is to find a compensator so that, for
as any N,(s) and any N,)s),
-Iy, Yw(s) = D -1 (s)w(s) This is the problem of asymptotic tracking and disturbance rejection.
Following the single-variable case, the design will consist of two steps:
(9-99) introduction of an internal model and stabilization of the feedback system.
There is, however, one important difference: the location of the internal model
nce we may combine is critica! in the multivariable case. To iIIustrate this point, we consider only
w(t) by assuming r(t) = O. The model D:; I(S) of the disturbance w(t) is placed as
shown in Figure 9-23. Let ew(s) be the input of D - 1 (s) as shown. Then we have
" ew(s) = w(s) - N(s)O,~ l(s)NAs)D; l(s)D- 1(s)ew(s)
Figure 9-19 can be
fhis equation can be 01' e,)s) = [1 + N(s)O;;: l(s)N c(s)D; I(S)O-I(SW 1-W(S)
) +eww(t) -------1
[
'-if~"'"''''''
I
: type of disturbance
f----------'-___/)
)ntrol systems. One : i \
3.nsient performance.
Ii - - - - - - - - - - - - - - - - - - 1I
•iI ' .
nse as t -> co. Hence
mg to this type of
Figure 9-22 Multivariable feedback system.
01' the response right
isturbance. Typical
. They are governed ~(S)
:::ms 01', equivalently,
I between these poles
~
y.(S)
d. For a discussion I (.n .
~.~~~-I
nces can be roughly
·bances. The former
1 stochastic control
:::1' is describable by Figure 9-23 Placement or internal model for disturbance rejection.
496 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
Theorem 9-21
The taQdem connection of cjJ -l(s)l p followed by the q x p proper rational
Figure 9-24 M ullivari
matrixC(s)( C(s) followed by cjJ - 1 (s)I q ) is controllable and observable if and only
ASYMPTOTlC TRACKING AND DlSTURBANCE REJECTlON 497
lX P proper rational
Figure 9-24 Multivariable feedback system.
observable if and only
498 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
res) = D,:-I(S)N,.(s) and w(s) = D;;; 1(s)Nw(S). Let <jJ(s) be the least common The design consi~
denominator of the unstable poles of every element of D,:-I(S) and D\~ I(S). If and stabilization of
no root of <jJ(s) is a transmission zero of G(s) and p ¿q or, equivalentIy, Hence the total comr
model <jJ(s) appears ~
rank [JJ--CA BJE = n +q for every root 1, of <jJ(s) are called, accordin¡
modes of r(t) and w
there exists a compensator with a p x q proper rational matrix such that the established in Theon
feedback system is asymptotical1y stable and achieves asymptotic tracking appear in e(t). In ere
and disturbance rejection. zero of the plant G(
observable mode or (
Proof will make the system
tracking and disturb~
Let G(s) = D- 1(s)N(s) be a left-coprime fraction and let <jJ -1(s)l q be the internal
C(s) and all of <jJ(s) m
mode\. The compensator Nc(s)D c- 1(s) is to be designed to stabilize the unity
We note that the j
feedback system. From Figure 9-24, we have of the reference and
e(s) =r(s) - D- 1(s)N(s)NcCs)D; l(s)(<jJ(s)l )-le(S)
q
be brought out of the
the blocking zeros of
Hence the transfer matrix from r to e is equal to De(s), N(s) and Ne(s)
Ge,.(s) = [1 + D- 1(s)N(s)N c(s)D; 1(s)(<jJ(s)lqr 1] - 1 (9-107) and (9-108). 1
= {D- 1(s)[D(s)(<jJ(s)l q )D c(s) + N(s)Nc(s)]D; l(s)(<jJ(s)I q )-I} -1 (9-107) of robustness, see Re!
= <jJ (s) DcCs)[<jJ(s)D(s)Dc(s) + N(s)Ne(s)] -1 D(s) We discuss in the
single copy of the ref(
Similarly, the transfer matrix from w to y can be computed as wil1 be general1y not
Gyw(s) = <jJ(s)De(s)[<jJ(s)D(s)DcCs) + N(s)N~(s)] -1 (9-108) 0-I(s)N(s)NcCs)D e- 1(1
Then we have
If p ¿ q and if no root of <jJ (s) is a transmission zero of G(s), then G(s) = D - 1(s) N (s)
followed by <jJ -1(s)Iq is controllable and observable (Theorem 9-21). Hence e(s) = [1 -+
we have deg [<jJ-l(s)lqD-l(S)N(s)] =deg [<jJ-l(s)I q ] +deg det D(s)(Theorem 9-1) We assume, without I(
which implies that <jJ(s)D(s) and N(s) are left coprime. Consequently the roots real parts. Since alll
of the determinant of of e(s) have negative
D ¡(s) ~ <jJ(s)D(s)De(s) + N(s)Ne(s) matrix. If D,.(s) and !"
matrix if and only if]
can be arbitrarily placed. in particular, placed in the open left-half plane by a t.hat D(s) = D(s)j[}l(sl 1,
proper choice af JDc{s) and NAs) (Theorem 9-1ó Oi 9-19). i-lence llii;; wúilj is possible if and only
feedback system is asymptotically stable.
Al1 roots of <jJ(s) have zero or positive real parts, hence no cancel1ation G(s)C(s)
between the roots of <jJ(s)D(s)Dcls) + N(s)Ne(s) and the roots of <jJ(s) will occur.
<';:onsequentIy, <jJ(s) will app~ar as a nUlperator of every element of Ge,.(s) and This is consistent witl
Gyw(s). !'i0te that wAe have Gew(s) = - Gyw(s). placed at the point wl
With Ge,·{s) and Gew(s), the response e(t) due to r(t) and w(t) can be written as this design, we req uil
However, the comput
e(s) = Ge,.(s)r(s) + Gew(s)w(s) = Ge,.(s)D; 1(s)N,.(s) + Gew(s)D; l(s)N w(s)
C(s) is found fora gi
(9-109)
subtracted and move
Since cjJ(s) appearsas zeros of every element of Ge,.(s) and Gew(s), áll unstable turbatiol1s inG(s).. H
poles of D,:-I(S) and D;;; I(S) are canceled by <jJ(s). What remains in e(s) are all design, the design' pr
stable poles. Hence, for any Nw(s) and N,.(s), we have e(t)-->O as t--> 00: This straightforward and ti
proves the theorem. Q.E.D. is m uch larger.
ASYMPTOTlC TRACKING AND DlSTURBANCE REJECTION 499
~ the least common The design consists of two steps: introduction of the internal model cP -1(s)lq
),:-I(S) and O:I(S). Ir and stabilization of the feedback system by the compensator N c(s)D c- 1 (s).
:, equivalently, Hence the total compensator is given by Nc(s)(Oc(s)p(S))-I. B~cause the internal
model cP(s) appears as zeros of every element of Ge,.(s) and Gew(s), these zeros
-t ), of cP(s) are called, according to Definition H-4, the blocking zeros. Ir the unstable
modes of r(/) and W(/) are the blocking zeros of Ce,(s) and Cew(s), then, as
matrix such lhal the established in Theorem H-7, for any initial state these unstable modes will not
asyrnptotic tracking appear in e(/). In creating these blocking zeros if any of them are a transmission
zero of the plant (;(s), then this zero will become an uncontro!lable 01' un
observable mode 01' a hidden mode of the system. This unstable hidden mode
will make the system useless in practice. Hence, in order to achieve asymptotic
tracking and disturbance rejection, no root of cP(s) can be a transmission zero of
-1(s)l q be the internal (;(s) and aH of cP(s) must be the blocking zeros of C ...(s) and CeAs).
to stabilize the unity We note that the internal model consists of q copies cP -1(s)l q of the dynamic
of the reference and disturbance signals. Because of these q copies, cP(s) can
be brought out of the parenthes~s, as shown in (9-107) and (9-108), and become
the blocking zeros of Ge,·(s) and Gyw(s). Consequently,the perturbations of O(s),
0cCs), N(s) and NcCs) will not affect these blocking zeros as can be seen from
(9-107) and (9-108). Hence the design is robust. For a more detailed discussion
(9-107) of robustness, see Reference S81.
We discuss in the following that the design can also be achieved by using a
single copy of the reference signal as an internal mode!. The design, however,
ed as wil1 be generally not robust. Consider Figure 9-22. Define (;(s) = C(s)C(s) =
sn -1 (9-108) D -1(s)N(s)NcCs)D c- 1 (s)é O-I(s)N(s), where O(s) and N(s) are left coprime.
Then we have
hen (;(s) = 0- 1(s)N(s)
leorern 9-21). Hence c(s) = [1 + G(s)] -lr(S) = [O(s) + N(s)ll D(S)O,-I (s)N,(s)
det O(s)(Theorern 9-1) We assume, without loss of generality, that aH poles of 0,:- 1 (s) have nonnegative
onsequently the roots real parts. Since all poles of (O(s) + N(s)) - 1 have negative real parts, all poles
of c(s) have negative real parts if and only if O(S)D,:-I(S)N,(s) is a polynomial
matrix. Ir D,.(s) and N,.(s) are left coprime, then O(s)D,:- 1(s)N,(s) is a polynomial
matrix if and only if 0,.(5) is a right divisor of 0(5), 01' there exists a D(s) such
~n left-half plane by a that D(s) = 6(s)O,.(s) (see Problems 9-19 lo 9-21). Henc~ asymptoOr. tn3cJdng
9). Hence the unily is possible if and only if a compensator C(s) can be found such that
Similar to the single-variable case, the perturbation of the internal model Consider the feedbac
is not permitted in the robust designo All other remarks in the single-variable
case are equally applicable here and will not be repeated.
Before moving to the next topic, we remark on the necessity of the condition
p ?:.q; that is, the number of plant inputs be larger than or equal to the number
of plant outputs. If q > p, the tandem condltion of <jJ -l(s)l p followed by G(s)
is controllable and observable, although G(s) followed by <jJ -l(s)lq is not
It is a plant with one
(Theorem 9-21). Consequently, we may introduce the internal model as shown
No matter how the f~
in Figure 9-25. The transfer matrix from r to e in Figure 9-25 can be computed as
at the outputs, the in
Ge,.(s) = [1 + 0- 1(s)N(s)<jJ - 1(s)lpNc(s)O; 1(S)J -1 (9-111 )
A k
If we write 0-I(s)N(s)<jJ-1(S)l p =0-1(S)<jJ-l(S)N(s)=(<jJ(S)0(S))-IN(s), then u(s) =- +(t
s
(9-111) can be simplified as
Hence we have, by u:
(9-112 )
This equation is similar to (9-107). Although <jJ(s) in (9-107) may become lirr
blocking zeros of Ge,.(s), the <jJ(s) in (9-112) can never be all blocking zeros of
We see that, unless 1
G.,.(s). This can be seen from the step of writing
achieve asymptotic f
0-1(s)N(s)<jJ -1(s)l p = 0-1(S)<jJ -1 (s)IqN(s) = [<jJ(s)O(s)J -1 N(s) asymptotic tracking,
equal to the q-dimen:
We see that p copies of <jJ(s) is increased to q copies. Because <jJ -1(s)I p fol1owed
q x p matrix, in order
by G(s) is controllable and observable, we have deg (O -1(s)N(s)<jJ - 1(s)I p ) =
of plant inputs must
deg det O(s) + p deg <jJ(s) (Theorem 9-1). The degree of det <jJ(s)O(s) is equal to
Using this argurr
q deg <jJ(s) +deg det O(s), which is, because q > p, .larger· than
Theorem 9-22. If r(
deg «<jJ(s)0(s))-1 N(s)). Hence <jJ(s)O(s) and N(s) are not left coprime. Con
tracking, we need ran
sequently, not al1 the roots of the determinant of
for every root ). of ,
<jJ(s)O(s)Oc(s) + N(s)Nc(s) and if }, is not a pole e
is not a transmission
can be arbitrarily assigned (Corollary 9-19). Furthermore sorne roots of <jJ(s)
will appear as unassignable roots, and there is always cancellation between
Static decoupling
<J¡(s)O(s)Oc(s) + N(s)Nc(s) and 4) (s). Hence not all roots of <jJ(s) will be the
feedback system shO\
blocking zeros of Ger(S) in (9-112). Consequently, if q > p, asymptotic tracking
terized by its q x p prc
is not always possible. A similar remark applies to disturbance rejection.
~¡1t l~Jr~J.i -?{.J'} == ;~:" ..
We use a different argument to establish once again the necessity ofp ?:.q.l J
reference signals are s
output as t -HfJ is cal
18 This argument was provided to the author by Proressor C. A. Desoer. application of r(t) is
matrix from r to y t
Figure 9-25 Multivariable feedback system with q> p. Figure 9-26 Design (
ASYMPTOTIC TRACKING AND DISTURBANCE REJECnON 501
;)f the internal model Consider the feedback system shown in Figure 9-25 with
in the single-variable
+ y
~
-~-~
+
L.
502 LINEAR TIME-INVARIANT COMPOSlTE SYSTEMS
response due to r(s1= ds- 1 can becomputed as, by using the final-value theorem, discussed in Figure S
hence we set P = 1 in]
lim y(e) = lim sG¡(s)((s) = lim sG ¡(s)ds- ¡ = G¡(O)d (9-113)
We find a C(s) so tha
t-cc s-o s-o
Now if G¡(O) is diagonal and nonsingular, in particular, a unit matrix, then
the feedback system is said to be seaeically decoupled. Indeed, if G¡(O) =
have negative real pa
diag {h¡, h z , ... , h q }, then (9-113) implies
computed, by using ('
lim h(t) =hid i i = 1, 2, ... , q
t-"" G¡(s)
where Yi and di are the ith components of y and d. Hence the steady-state
response at the ¡th output of a statically decoupled system depends solely on the Hence we have
ith reference input and is decoupled from the other inputs. Ir there is a change
in the magnitude of the ith reference step input, it will cause responses at all
output terminals. However, as time approaches infinity or as the transient Since al! the poles of
dies out, it will cause only a change at the ith output and no change at other is nonsingular. Ir s
outputs. Hence in a staticalIy decoupled system, only the steady-state responses Now we may design
are decoupled, but not the transient responses. This differs from the decoupled constant matrix N(O
system discussed in Section 7-6 where all the responses, transient as well as we may choose
steady state, are decoupled. Furthermore, decoupling is defined for any
reference signal; whereas, static decoupling is defined only for step reference
inputs. The class of step reference signals, however, is very important and is
often encountered in practice. For example, temperature and humidity controls
and G¡(O) becomes
of a room are this type of reference signals. Maintaining an aircraft at a fixed
altitude is another example.
Hence the system in f
Asymptotic tracking actually achieves decoupling as e---+ 00; the steady state
to the one in Figure (
of Yi(e) tracks r¡(e) and is independent of r ie), for j i= i. Hence the design for
Nc(O) and D ¡(O), then
asymptotic tracking can be directly applied to static decoupling. In this case,
marize the preceding
we have <b(s) = s. Let K = Iq and P = <b - l(s)I q = S-1 I q in Figure 9-26. We then
design a compensator C(s) in Figure 9-26 to stabilize the feedback system. As
shown in Theorem 9-22, if P'2:q and if s is not a transmission zero of G(s), then s Corollary 9-22
will appear as a zero of every element of (;er(S), the transfer matrix from r to e. Consider the feedback
Hence we have (;er(O) =0 and charact~~ized by its q
~.~:~r:' ·~:-f f;J ,,\ ~~.~'
lim e(i) = lim [¡r(ei - Y(é)] = Hm sGe.. (S)~S . . \ = Ge ,(O);1 = 0 ri --
"
t- 00 t- 00 s-o
Because of e(s) = r(s) - y(s) or Y(s) = r(s) - (;e,.(s)r(s), the transfer matrix, G¡(s),
[rom r to y is equal to
G¡(s) = 1 - (;er(S) then there exists a com
feedback system is ¡
Consequently, we have (;¡(O'....= Iand the feedback system is statically decoupled. {A, B, e, E} is any im
Note that every element of G¡(s); except those on the diagonal, has s as a zeJ,"o
in its numerator. We remark once
Because of the presence of the internal model, the design is robust. .Tl:iat internal model or.nonl
is, the system remains to be statically decóu~led with perturbations, even large the remark in (5) on ti
perturbations, of the parameters of C(s) and G(s), so long as the feedback system· To check· the conditi
remains to be asymptotically stable. In the following, we introduce a design of (;(s) and is complica
which is not robust. The design is an extension of the single-variable case 9-22 is eq uivalent to r
ASYMPTOTIC TRACKING AND D1STURBANCE REJECTION 503
:final-value theorem, discussed in Figure 9-21. In this desilin, no internal model will be employed;
hence we set P = 1 in Figure 9-26. Let G(s) = N(s)D- 1(s) and C(s) = D e- 1(s)N e(s).
(9-113) We find a C(s) so that all the roots of the determinant of the polynomial matrix
then there exists a compensator with a q x p proper rational matrix such that the
feedback system is asymptotically stable and statically decoupled, where
s statical\y decoupled. {A, B, C, E} is any irreducible realization of G(s). . •
gonal, has s as a zero
We remark once again that the design can be robust by introducing an
:sign is robusto That internal model or nonrobust without introducing an internal model. Similar to
turbations, even large the remark in (5) on the single~variable case, the condition is necessary as well.
LS the feedback system
To check the condition in Corollary 9-22, we must find a copriine fraction
ve introduce a design of G(s)and is complicated. If G(s)hasno pole at s =0, thecondition in Corollary
e single-variable case 9-22 is equivalent to rank G(O) = q, which can be easily checked.
504 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
with some unknown initial state. The problem is to design a robust control
system so that the output of the plant will track asymptotica11y the reference
signal r(t) generated by Figure 9-27 Design of r
X,.(t) = A,.x,.{t) r(t) = C,.x,.(t)
with some unknown initial state. Let </>,)05) and </>,.(s) be the minimal poly This connection was sh(
nomials of A w and A,., respectively, and let if and only if p 2: q and r
we reestablish directly t
be the least common multiple of the closed right-half s plane roots of </>,.(05) and ,11
rank
</>,.(05). Thus a11 roots of </>(05) have nonnegative real parts. The internal model [ -(
</>-l(05)I q can be realized as
Theorerh 5-7 implies th~
Xc = Aex e + 8 ee
Ye =X e rank V(s) ~
w
lall discuss the design
¡
dng and disturbance
Icussion will be brief
lpproach. Consider a
,q X 11, q X p, 11 X q. and
lmtrollable and {A, C}
l1al w(t) is generated by
be the minimal poly This connection was shown in Theorem 9-21 to be controllable and observable
if and only if p 2.q and no root of </J(s) is a transmission zero of the plant. Now
we reestablish directly that (9-116) is controllable if
,Iane roots of </Jw(s) and rank [ AI-A BJ =/l +q for every root ). of </J{s) (9-117)
s. The internal model -C E
Theorem 5-7 implies that (9-116) is controllable if and only if
(b)
;tem in Figure 9-27 can
, we can design a state
ut of the state estimator. (3)
to achieye asymptolic
Figure 9-28 lnpul-olllPlll feedback syslem.
:d discussion, see Refer-
(9-119 ) For any De(s) of degree m and any D ¡(5) of degree n + mor less, the polynomial
D /(05) - D/s)O(s) is of degree n +111 or less. The application of Theorem 9-10
to (9-122) yields that (9-122) has solutions L(s} and M(s} of degrees mor lcss if
(9-120)
and only if D(s} and N(s) are eoprime and m '2.11 - 1. This proyes the theorem.
Q.ED.
re 9-1(b}can be redrawn
1 x 2 rational matrÍz We compare first this theorem with Theorems 9-1 1 and 9-11'. First, in
using m integrators. 1f order to ensure the praperness of the compensators, we m lIst consider separately
ingle-loop system shown in Theorems 9-11 and 9-11' the cases where the plant is proper or strictly praper
mpensators are different. and require the degree of O /(s) to be exactly equal to n +m. Since the D,(s) in
IOtegrators as the one in Theorem 9-23' is chosen a priori to haye degree m, the compensators are always
:ions in Figure 9-28 are proper whether D¡(s) is of degree n + 111 or not. Second, the oyerall transfer
1 implementation. function of the unity feedback system in Figure 9-12 is N(s)D¡ l(s)N ,(s), where
can be computed as we can contraJ only D¡(s). The oyerall transfer function of the input-output
feedback system in Figure 9-28(a) is N(s)D¡ 1 (s)De(s), where we can control
IDe(s) (9-121 )
D¡(s) as well as De(s). The reason for haYing this extra freedom can be seen from
JD(s) + M(s)N(s)
Figures 9- 12 and 9-28(a).. Although the compensator D; 1(s)NAs) in Figure 9-12
and the compensator D e- l(s)[L(s) M(s)] in Figure9-28(a) have the same
degree, the foimer has two setsof parameters Del' N el' whereas the lalter has
N(s)
N(s) (9-122 ) three sets üf paraineters Dei' L¡, M j. Since ítreq lIires only two sets ofparameters
S08 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
DE!
to meet (9-40) and (9-122), either De(s) or L(s) in (9-122) can be arbitrarily
+
assigned. We choose to assign Dc(s) because it dictates the poles of the com
pensator.
The employment of Theorem 9-23' requires the solving of (9-122) or
E(s) ~ D¡(s) - Dc(s)D(s) = L(s)D(s) + M(s)N(s) (9-124 )
This is the Diophantine equation in (9-40) and, hence, can be translated into a
set of linear algebraic eq uations. Let
Figure 9-29 An input
D(s)=D o +D¡s + +D"s"
N(s)=N o +N¡s + +N"s"
Dc(s)=D co +Dc¡s + +Dcmsm
L(s)=L o +L¡s + +Lmsm (9-125) Although al! comI
M(s)=Mo+M¡s+'" +Mms m serious problem. Th.
D¡(s)=F o +F¡s + ... +F"+ms"+ m function and is, as dis
and E(s) = E o + E¡s + . _. + E Il + msrz+m feedback systems, we
the resulting system t
Then Equation (9-124) is equivalent to the algebraic equation posed system and sho
[Lo M o : L¡ M ¡ : ... : L m MmJSm = [E o E¡ ... EIl+mJ ~ E (9-126)
Theorem 9-23
where SIIl is defined as in (9-45). Ir m:::::n - 1, then SIIl has a ful! column rank.
Hence for any E, solutions {L i , M;} exist in (9-126). The solutions yield im Consider a plant with 1
mediately the proper compensators Co(s) and C¡(s). n. For any Dc(s) of dI
compensators L(s)/D c(:
Example 1 9-28(a) is well posed an
only if D(s) and N(s) al
Consider a plant with transfer function g(s) = N(s)/D(s) = (s -1)/s(s - 2).
Clearly, we have n = 2 and m "2:: n - 1 = 1. Let us choose arbitrarily Dc(s) = s + 1, We see that by re
and D¡(s) = S2 + 2s + 2. Note that the degree of D¡(s) is smal!er than n + m = 3. then assert the wel! pos
We compute feedback system is wel
E(s) =D¡(s)-Dc(s)D(s) =S2 +2s +2-(s +1)s(s-2) from O at s = OC! if deg
= 2 + 4s + 2s - S3
2
De l(s)D f (s)[;
and form
":-:.':" '.;;:; ~-;<~ ':"/.'. ¡--" ~~:>. ,"
value of De-¡(s)D¡(s)D
[Lo Mo: L¡ 4 2 -1J for S-' 00,
D; ¡(s)D¡(s)D-¡(s)·
Its solutions are Lo = - 6, L¡ = - 1, M 0= - 2 and M ¡ = 6. Hence the com Hence if FIl+m=O or (
pensators are the system is not well]
Co(s) = L(s) = -6 -s 1 +Co(oo) +C¡(oo)g(oo
D/s) 1 +s . note that if deg D¡(s) < j
We consider now a
and C¡(s) = M(s) = - 2 +6s
Dc(s) 1 +s .. Corollarv 9-23
They are al! proper rational functions. The block diagram of the feedback Consider a plarit with tr
system is shown in Figure 9-29. • /1. For any i5 ¡(s) of d(
DESIGN OF COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS 509
Ing of (9-122) or
1
I;)N(S) (9-124)
Iw be translated into a
(9-125 )
Although al! compensators in Figure 9-29 are proper, the system has one
serious problem. The system has a loop which yields an improper transfer
function and is, as discussed in Section 3-6, not well-posed. In the design of
feedback systems, we require not only al! compensators to be proper but also
the resulting system to be well-posed. Theorem 9-23' may not yield a well
posed system and should be replaced by the fol!owing theorem.
En+mJ ~ E (9-126)
Theorem 9-23
,¡aSa ful! column rank.
Consider a plant with transfer functiong(s) = N(s)/D(s) and deg N(s):::; deg D(s) =
I[he solutions yield im
n. For any Dc(s) of degree m and any D[(s) of degree n +m, there exist proper
compensators L(s)/Dc(s) and M(s)/Dc(s) such that the feedback system in Figure
9-28(a) is wel! posed and has a transfer function N(s)Dj l(s)Dc(s) from r to y if and
only if D(s) andN(s) are coprime and m ~ n - 1 . . 11
s)/D(s) = (s -1)/s(s - 2).
arbitrarily Dc(s) = s + 1, We see that by removing deg DAs) < n +m from Theorem 9-23', we can
smaller than n +m = 3. then assert the well posedness in Theorem 9-23. We show that the input-output
feedback system is well posed by showing that 1 + Co(s) + C,(s)g(s) is different
from O at s = 00 if deg D¡(s) = n +m. From (9-122), we have
-1)s(s - 2)
D; '(s)D¡(s)D - '(s) = 1 + D,- '(s)L(s) + Dc-'(s)M(s)N(s)D- '(s)
=1+Co(s)+C,(s)g(s) (9-127)
Corollar·y 9-23
iagram of the feedback Consider a plant with transfer function g(s) = N(s)/ D(s) and deg N(s) :S deg D(s) =
I n. For any i5 As) of degree n, there exist proper compensators L(s)jD/s) and
510 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS DE~
M(s)/Dc(s) of degree n - 1 and with arbitrarily assignable poles such that the system to minimize Ü
feedback system in Figure 9-28(a) is wel1 posed and has transfer function
N(s)i5 ¡ 1 (s) if and only if D(s) and N(s) are coprime. Ili
This design always involves the canceIlation of Dc(s), which can be chosen by
where i5 f (s) is a Hur'
the designer, however. The degree of Dc(s) in the coroIlary can be larger than
N(s) in the numerator
n - 1; however, it does not seem to serve any design purpose because DcCs) is
function 9(s). See R
completely canceled in the designo
requires the solution (
We compare now CoroIlary 9-23 with the result obtained in the state
variable approach. By state feedback, we can assign the eigenvalues of A or
Multivariable case
the poles of g(s) = N(s)/D(s) as the roots of i5 As) without affecting N(s). An will be extended to the
(n -1)-dimensional state estimator with arbitrary eigenvalues can be con
system shown in Fig
structed to generate an estimate of the state. The connection of the feedback rational matrix (;(s) ,
gain from the output of the estimator yields the overal1 transfer function p x p proper rational
N(s)ji5As) (see Section 7-5). Hence Corollary 9-23 establishes essentially the matrix C¡(s)= Dc-l(s)1
result of state feedback and state estimator. However, the result in Corollary be computed as
9-23 is slightly more general: the plant is permitted to have a proper transfer
function. In Chapter 7, we design state estimators only for strictly proper plants Gis) = (;(s)[1
or dynamical equations wi.th the direct transmission parts equal to zero. In = N(s)D
Corol1ary 9-23, we require deg i5¡(s) = deg D(s); in the design of state feedback, = N(s)[I
we require deg i5fes) = deg D(s) and i5 f" = D" (that is, their leading coefficients
are equal). Ir deg i5¡(s) = deg D(s) and i5 f" = D", the compensator Co(s) is Define
always strictly proper for g(s) strictly proper. This can be verified from the last
column .equation of (9-126) (Problem 9-28). Hence we always have D
1 +Co(oo) +CI(oo)§(oo) #=0 for the class ofsystems studied in the state-variable or E
approach, and consequently, the well-posedness problem does not arise in the
approach. Hence the result in Corollary 9-23 is more general than the result of
state feedback and state estimator.
In the transfer~function approach, we require only the concept of coprime
ness. In the state-variable approach, we require the concepts of control1ability
and observability. In the former approach, the design consists of forming a
linear algebraic equation and its solutions yield immediately the required
compensators. In the latter approach, the design requires one similarity
transformation to compute the feedback gain, and requires one similarity
transfonnation. or one solution of a Lyapunov matrix equation to find a state
estimator. Hencefor the single-variable case, it appears that the design in
the transfer-function approach is simpler conceptua:Ily and computationaIly
than the one in the state-variable approach.
To conclude this subsection, we remarkthat Corollary 9-23 can be used in
the design of optimal systems. Consider a plant with transfer function g(s) =
N(s)jD(s) with input u(t) and output y(t). ltis required to design an overal1 Figure 9-30 Input-out]
DESIGN OF COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS 5U
where q> O is a weighting factor and r(t) is the reference signa\. lf r(l) is a step
l by choosing D ¡(s) = function. then the optimal system which has the smallest J is of the forro
~ (s) = ~(s)
1= N(s)i5¡ 1(S) g¡ D¡(s)
Ihich can be chosen by where 15¡(s) is a Hurwitz polynomial and has the same degree as D(s). The
lary can be larger than N(s) in the numerator of 9¡(s) is the same as the numerator of the plant transfer
urpose because D/s) is function 9(s). See Reference S46. Hence the design of the optimal system
requires the solution of Corollary 9-23.
obtained in the state
:he eigenvalues of A or
Multivariable case. In this su bsection, the results in the single-variable case
out affecting N(s). An
wil1 be extended to the multivariable case. Consider the input-output feedback
7envalues can be con
system shown in Figure 9-30. The plant is described by the q x p proper
~ection of the feedback
rational matrix (;(s) = N(s)D- 1(s). The compensators are denoted by the
..eran transfer function
p x p proper rational matrix Co(s) = D; 1(s)L(s) and the p x q proper rational
:ablishes essential1y the
matrix C 1(s)= D;-I(S)M(s). The transfer matrix from r to y in Figure 9-30 can
. the result in Corollary
be computed as
have a proper transfer
Define
: compensator Co(s) is
shown in (9-75). We t
D c- l(s)D¡(s)D -I(S)= 1 + D,- I (s)L(s)+ D e- l(s)M(s)N(s)D-I(S)
from top to bottom. I
= 1 + Co(s)+ C¡(s)C(s) (9-137)
block row of S, are lir
We write, similar to (9-83), convenience, we assum
L(s) =L o + L
D(s) = [D /I + D/(s)]H(s) (9-138)
De(s) = H,(s)[D eI, + DcI(s)] (9-139 ) M(s) = M o +~
and E(s)=O/s)
where D/(s) and Dcl(s) are strictly proper. Theri we have
where J1 = max {,ui, i =
lim D; 1 (s)D ¡(s)D(s) = lim [Del. + DcI(:s)] -1 He- l(s)D¡(s)H - 1 (s)[D" + O/(S)]-I (9-131).yields
s-+ co s-+ 00 .
(9-140)
[Lo Mo : L I MI
which is nonsingular by the assumptions of O,(s), D(s), and (9-135). Henee the
DESIGN OF COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS 513
input-output feedback system in Figure 9-30 is, following Theorem 3-6, well
posed. Q.E.D.
(9-132)
where dei(s) are arbitrary Hurwitz polynomials of degrees mi. Ir some O!' al1
mi are odd integers, then we may not be able to assign complex conjugate roots.
;(s) = N(s)D -les). Let In this case, if we choose DcCs) to be of the form shown in Problem 2-29, then
v be the row index of
the difficulty of assigning complex conjugate roots will not arise.
)e(S) of row degrees mi
Corollary 9-24
Consider a plant with q x p proper rational matrix G(s)= N(s)D-'(s). Let
(9-135 )
Il¡, i = 1,2, ... ,p, be the column degrees ofD(s) and let v be the row index of G(s).
Then for any De(s) of row degrees aH equal to v - 1 and row reduced, and any
roper rational matrices fi ¡(s) of column degrees Ili, i = 1, 2, ... , p, and column reduced, there exist
n in Figure 9-30 is well compensators with proper rational matrices D; '(s)L(s) and D; l(s)M(s) such
11y ir D(s) and N(s) are that the feedback system in Figure 9-30 is well posed and has transfer matrix
N(s)fi¡ '(s) if and only if D(s) and N(s) are right coprime and D(s) is column
reduced.
Proof
Let Df(s) = De(s)D¡(s). Clearly the degree requirements ofD¡(s) in Theorem
9-24 ~re met under the assumptions ofDcCs)and DAs). With this D¡(s), (9-131)
and D" is the column and G ¡(s) become
üsts, we conclude from DcCs)[fi¡(s) - D(s)] = L(s)D(s) + M(s)N(s) (9-142)
of row degrees at most and G feS) = N(s)[DcCs)D ¡(s)] -, De(s) = N(s)D¡ '(s) (9-143)
(s)M(s) are proper fol
- D dl D iI is nonsingular This establishes the corollary. Q.E.D.
ving that 1 + C o(ro) + The application of Theorem 9-24 and its corol1ary is straightfol'W8.cd.
ilar to (9- i 27), First we use the coefficient matrices 01 D(s) and N(s) to form the matrix Sm
shown in (9-75). We then search the linearly independent rows ofS m in arder
l(s)N(s)D- '(s) from top to bottom. Let v be the least integer such that aH N rows in the last
(9-137) block row of S. are linearly dependent. This v is the row index of G(s). For
convenience, we assume mi = v - 1 for aH i in Theorem 9-24. Let
L(s)=Lo+L,s+'" +L._,s·-' (9-144)
(9-138)
M(s)=M o +M,s + ... +M._,s·-' (9-145)
(9-139)
arid E(s) = D ¡(s) - DcCs)D(s) =Eo + Els + ... + El' h ' _ ¡S" +v -l (9-146)
where Il =max {Il¡, i = 1,2, ... , p}. The substitufion of (9-144) to (9-146) into
IH - '(s)[D" + D¡(s)]-' (9-131) yields
(9-140) [Lo Mo:L l Ml:"<L'- 1 M,-,]S'-l=[E o El ... El'h-l]
(9-147)
and (9-135). Hence the
514 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS DI
The solution of this set of linear algebraic equations yields the required com Using G(s)= N(s)D
pensators. be written as
It is ofinterest to compare the result in Corollary 9-24 with the one developed
G¡(s) = N
in the state-variable approach. If D(s) and N(s) are right coprime and D(s) is
column reduced, then an irreducible realization can be readily found for (;(s) = Ir we define
N(s)D- 1(s) (Section 6-6). By state feedback, we can achieve N(s)Dj I(S), where
D¡(s) and D(s) have the same column degrees and the same column-degree
coefficient matrix. Note that the latter condition is not required in Corollary and E(s) ~
9-24. If the state is not available for feedback, we may design a state estimator
then G¡(s) = N(s)D ¡
with arbitrary eigenvalues, which are equivalent to the roots of det Dc(s), to
Theorem 9-24 to acb
generate an estimate of the state. Now the application of the state feedback
ever, the row degrees
from the output of the state estimator yields N(s)Dj I(S). Note that the eigen
degrees of Dc(s) will t
values of the estimator are not control1able as can be seen from (7-86) and will
introductionofQ-l(:
not appear in the transfer matrix from r to y.
A remark is in order regarding the degreesof compensators. In the state
Theorem 9-25
variable approach, the dimension ofthe state estimator is n -q (Theorem 7-13).
In the transfer-function approach, the degree of the compensators is p(v -1). Consider a plant wil
Note that the row index v of (;(s) is equal to the observability index of any assumed that D(s) an<
irreducible realization of (;(s), and has the property v 2.n/q. If p =q = 1, then column degrees f1.i, i =
n = v and the dimension of the state estimator is equal to the degree of com P x P arbitrary polynl
pensators. Ir p 2.q, then p(v -1) 2.n - q; if p < q, p(v -1) can be greater than, and be row reduced.
equal to, or less than n - q. Hence the results in the state-variable and transfer
function approaches are not exactly identical. Similar to the single-variable
case, the design procedure in the transfer-function approach appears. to ·be is a proper rational
simpler, conceptual1y and computationally, than the one in the state-variable D c-l(S)M(s), and Q
approach. posed and has trans
In the fol1owing, the input-output feedback system in Figure 9-30 will be computed from
extended to the one shown in Figure 9-31, in which Q(s) is a polynomial matrix
and Q -1(S) is required to be proper. The transfer matrix from r to y in Figure
9-31 can be readily computed as with (SciR 1(s) < ()ci O(s)
G¡(s) = (;(s)Q-l(s)[I +C O(S)Q-l(S) +C 1(s)(;(S)Q-l(S)]-1 (9-148)
Proof
ds the required com Using (;(s) = N(s)D-I(S), Co(s) = D,:-I(S)L(s) and C1(S)= Dc-I(S)M(s), G¡(s) can
be written as
'ith the one developed G¡(s) = N(s)[Dc(s)Q(s)D(s) + L(s)O(s) + M(s)N(s)] -11\(s) (9-149 )
t coprime and 'P(s) is
Ldily found for G(s) = Ir we define
:ve N(s)Dj 1(s), where D¡(s) ~ Dc(s)Q(s)D(s) +L(s)D(s) +M(s)N(s) (9-150)
,ame coluron-degree
required in Corol1ary and E(s) ~ D¡(s)- Oc(s)Q(s)D(s)= L(s)O(s)+ M(s)N(s) (9-151)
sign a state estimator then Gj(s) = N(s)Dj l(s)Dc(s) and has the same form as (9-132). The design in
roots of det Dc(s), to Theorem 9-24 to achieve (9-132) is accomplished without using Q-I(S); how
of the state feedback ever, the row degrees of Dc(s) are general1y difTerent. In the fol1owing, al1 row
Note that the eigen degrees of Dc(s) wil1 be required to be the same. This is possible because of the
n froro (7-86) and wil\ introduction ofQ - 1(s). In this case, R(s) in (9-134) becomes diag {Slll, Slll, ... ,s"'}.
with o,.¡ 1f«2{S) < o,.¡ Dc(s) = m, and íL(s) and IVK(sj are soiutlons 01
Proof
The column degrees of R 2 (s)D(s) + R 1(s) are c1early at most 111 + ¡Ji. Hence L(s)
and M(s) of row degrees at most m exist in (9-155) (Theorem 9-17). Thus
Dc-I(S)L(s) and Dc-I(S)M(s) are proper.
Next we show that D(s)D ¡ 1(s) Dc(s) is proper under theassumption of
(9-152). We use (9-138) and (9-139) with m = mi, for all i, to write
D(s)D¡ 1(s)Dc(s) = [O" + D¡(s)]H(s)D¡ 1(s)Hc(s)[D c" + Dcl(s)] (9-156)
Since D" + D¡(s)and Dc/,+ D,¡(s) are prope"r, if(9-152) holds, then D(s)D¡ 1(s) Dc(s)
IS proper.
516 LINEAR TIME-INVARIANT COMPOSlTE SYSTEMS
where
Using (9-153) and (9-154), we have
pes) = 0c- 1(s)[OAs)Q(s)O(s) + R2(S)0(S) + R 1(s)] O - 1(s)Q -I(S) and D ¡(s) and Nfes)
= 1+ D c- 1(s)R 2 (s)Q -'1 (s) + Oc- 1(s)R 1(sjO - 1(s)Q - I(S) (9-162) shown in Figure 9-16
Because O; 1(s)R 2(s) and RI(s)D - 1(s) are strictly proper, D c-1(S) and Q - I(S) are is to be solved from a
proper, the polynomial part ofP(s) is I. Hence P- I(S) is, following Theorem 3-4, hence the unity feedb
proper. This establishes the well-posedness of the feedback system. Q.E.D. the open-loop system
9-31, the overall tran
We remark the condition in (9-152). One way to check the properness of assignable. A compa
H(s)O¡ l(s)Rc(s) is by direct computation. The computation of the inverse 01 bility of implementinl:
Of(S) is, however, complicated. Instead, we may compute and O f(S) are meto .
introdüce a differcnt ¡
which can be obtained by inspection because of the forms of Rc(s) and H(s). lmplementation l. e
Clearly Y(s) is much simpler than D feS). Now H(s)O¡ 1(s)Hc(s) is proper if and Figure 9-31. Its oyera
only if y-I(S) is proper (Theorem 3-4). This is a simpler way of checking the Gfes) = G(s)Q-I(s)[1
condition in (9-152). Because of Hc(s) = diag {sm, s"', ... ,Slll}, a sufficient con
dition for y-I(S) to be proper is that Df(s) has column degrees m+ /li and is
column reduced. Even if O feS) is not column reduced, it is still possibl~ for 19 This probIem was first 1
Reference SSO.
Y(s) to have a proper inverse. See the footnote on page 115.
The designs in Theorems 9-24 and 9-25 yield
G feS) = ~(s)D ¡ ](;)'Oc(s) = N(s)O - 1(s)D(s)D¡ 1(s)Oc(s)
= G(s)T(s) (9-163)
~)] - 1 D,(s) (9-159) where d¡(s) are Hurwitz polynomials of smallest possible degrees to make
T(s) proper, then the resulting system is decoupled. Thus the problem to be
(9-159) yields
discussed may find several applications in the design of multivariable systems.
(S-15D) LetC(s)= N(s)D-l(s) be a coprime fractíon with D(s) column reduced.
Then the open-loop transfer matrix is
(;o(s) = (;(s)T(s) = N(s)D- l (s)T(s) ~ N(s)Dj l(s)N ¡(s) (9-165 )
(9-161 )
where D- l (s)T(s) = D¡ l(s)N ¡(s) (9-166)
[)-I(S)Q-l(S) and D ¡(s) and N ¡(s) are left coprime. If we use the utlity feedback system
~)Q-l(S) (9-162) shown in Figure 9-16, the overall transfer matrix is N(s)D¡ l(s)Nis) and Nc(s)
Dc- 1 (s)and Q-l(s)are
is to be solved from a Diophantine equation and cannot be arbitrarily chosen;
hence the unity feedback system in Figure 9-16 cannot be used to implement
lllowing Theorem 3-4,
the open-loop system. Ir we use the input-output feedback systemin Figure
ck system. Q.E.D.
9-31, the overall transfer matrix is N(s)D¡I(s)D c (s), where Dc(s) is arbitrarily
assignable. A comparison of this with (9-165) reveals immediately the possi
eck the properness of
bility of implementing Co(s) in Figure 9-31 as long as the conditions on O,(s)
ltíon of the inverse of
and D¡(s) are met. This will be done in Implementation l. We shall also
te
introduce a different implementation in Implementation n.
y)H - ¡ (s)
Implementation l. Consider the input-output feedback system shown in
ms of R(s) and H(s).
Figure 9-31. Its overall transfer matrix is, from (9-148) and (9-149),
.)Hc(s) is proper if and
r way of checking the (;¡(s) = G(s)Q -1(s)[1 + D - l (s)L(s)Q - I(S) + D; 1 (s)M(s)N(s)D -1 (s)Q -1(S)] - 1
c
,sm}, a sufficient con (9-167)
degrees m + J1¡ and is
it is still possible for 19 This problem was firsl f¿rmulaled and solved in Reference S218. This presentation fo\lows
Reference SSO.
115.
(9-163)
where Dc(s) can be arbitrarily chosen. Hence this configuration may be used The substitution of(
to implement G(s)T(s) = N(s)D; l(s)N ¡(s). In this implementation, we require G¡(S) = 1
T(s) to be nonsingular. The nonsingularity ofT(s) implies the nonsingularity of
Nfes). Consequently we can always find a unimodular U(s) such that U(s)N¡(s) Hence G¡(S) implem
is row reduced and U(s)D¡(s) and U(s)N¡(s) remain to be let coprime (Theorem Equations (9-169
G-11). Thus we assume without 10ss of generality that N¡(s) is row reduced. T(s) =
A In the following, we present a procedure so th,Ú G[(s) will implement
G(s)T(s). Note that Dc(s)D ¡(¿
U sing (9-138) and (9·
Step 1. Compute a fraction (;(s) = N(s)D-l(s), where D(s) ann N(s) are right
coprime and D(s) is column reduced. Compute [D1.+ D¡(s)]
D-l(s)T(s) = D; 1 (s)N ÁS) (9-169) Since Oh and Del,
[Del. + Dcl(s)] - ¡ are,
where D ¡(s) and N ¡(s) are left coprime and N Ás) is row reduced. Let the condition in (9
DriN ¡(s) = mi, i = 1,2, ... ,p Theorem 9-25.
where Dri denotes the ith row degree. We discuss now .
Step 2. Compute the row index, v, of (;(s). Define depends on T(s). Ifl
det D¡(s) in (9-169) is
will also be BIBO s1<
Let det Dc(s) which are 1
.o/s) = diag {Cl. l (s), Cl.2(S), ... , Cl.p(s)} Hurwitz, then the fee
where lXi(S) is an arbitrary polynomial of degnie m - mi' Then the matrix The decompositi(
algorithms in Referel
Dc(s) = .o/s)N¡(s) (9-170)
design, the condition
has row degrees all equal to m:;:::: v - 1 and is row reduced. (9-172) by
Step 3. lf
i= 1, 2, ... ,p (9-171 )
with Dc¡R¡(s):::;;DciD(s
set Q(s) = 1 and go to step 4. If not, compute example, from the un
D ¡(s) = Ql(s)D(s) +Rl(s) (9-172) Q¡{
with C\¡Kl(S) < ÓciÜ(s) = j1i, for aH i, and compute ¡Ol" dúy COllstalil rnatr
with Dri R 2 (s) < DriN ¡(s) = mi, i = 1, 2, ... ,p. These decompositions are with D,J{2(S) :::;;Dr¡N¡l
unique. proper. However, f(
Step 4. Solve L(s) and M(s) from Furthermore, the res'
degree conditions in
.oc(s)[D fes) - N ¡(s)D(s)] ;= L(s)D(s) + M(s)N(s) (9-174a)
degrees at most m stil
or .o/s)[R 2 (s)D(s) + R 1 (s)] = L(~)D(s) + M(s)N(s) (9-174b) .
1mplementatioll 11.
Theorem 9-26 section. In this iinph
The input-output feedback system in Figure 9-31 with Dc(5), Q(s), L(s), andM(s), . square and may not t
cAomputed from (9-170) and(9-174) implements the open-Ioop system (;o(s) = in FigLire 9-31 excep
G(s)T(s) = N(s)D; l(s)N fes) and is well posed. . terminal as shúwhin
DESIGN OF COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS 519
(9-168) Proof
The substitution of (9-170) to (9-174) ¡nto (9-168) yields
guration may be used
nentation, we require G¡(s) = N(s) [fic(s)D¡(s)J -lfic(s)N¡(s) = N(s)Dj l (s)N ¡(s) (9-175 )
s the nonsingu\arity of
(s) such that U(s)N ¡(s)
Hence G¡(s) implements the open-Ioop system.
let coprime (Theorem Equations (9-169) and (9-170) imply
~¡(s) is row reduced. T(s) = D(s)O¡ l(s)N¡(s) = O(s)[fic(s)D¡(s)] - 1 Oc(S) (9-176 )
G¡(s) will implement
Note that Dc(s)D¡(s) in (9-175) corresponds to the D¡(s) in Theorem 9-25.
Using (9-138) and (9-139) with mi = m, for all i, we can write (9-176) as
t(s) ann N(s) are right
[D h+ D/(s)J - lT(s)[D ch + Dcl(s)J - l = H(s)[fic(s)D¡(s)J - l Hc(s) (9-177)
(9-169) Since D" and D,.!> are nonsingular by assumption, [Oh + D,(s)J - 1 and
[Och + DcI(s)J - l are, following Corollary 3-4, proper. Hence, if T(s) is proper,
ow reduced. Let the condition in (9-152) is satisfied and the theorem fol1ows directly from
Theorem 9-25. Q.E.D.
101 any constanl matrix ¿, LO meeL (9- i 78 j. 3íl11liélnj, (9-1 ¡ j) can be replaced by
(9-180)
(9-173)
decompositions are with (j,.¡Rz(s)::S:;¿¡riN¡(s). In this case, we cannot expect every Q¡I(S) to be
proper. However, for almost all QI(S) so computed, Q¡I(S) is still proper.
Furthermore, the resulting D,(s)[D ¡(s) - N¡(s)Q(s)D(s)] still meets the column
degree conditions in Theorem 9-17; hence solutions L(s) and M(s) of row
s)N(s) (9-174a)
degrees at most m still exist in (9-174).
s)N(s) (9-174b)
lmplementatiori jI. We introduce a different implementation in this sub
section. In this implemen.tation, the op~n-IoopcompensatorT(s) may be non
(s), Q(s), L(s), and M(s), squáre and inay n6t haveá full rank. The configuration is identícal to the one
n-loop system Go(s) = in Figure 9-31except that acorripensator Ole. l(s)K(s) is placedat the input
terminal as shown in Figure 9-33. The transfer matrix of the system is clearly
520 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS DE
Step 4. Define
Ir J: - f.J..¡ - In 2. O, ~
Hurwitz polynorr
b,.¡W(s) ~m.
Step 5. If
20 Because theproperness of T(s) = D(s)D¡ '(s)N¡(s) does no! imply the properness of D(s)D¡ 1 (s)
(see' Problem 3-42), we cannot conclude that f; ":?:.J1i. Consequently, we must introduce v¡(s) to Henee under the assu
conclude ¡; ~Jl¡. However, in most cases, for example, when T(s)has a fult rank, we have v¡(s) = 1, reduced anci m?:. v - 1
for al! i. (9-191). Hence the eo
DESIGN OF COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS 521
Step 4. Define
W(s) = diag {Wl (s), W2(S), > • > , Wp(s)} (9-186 )
If J: - ¡.t¡ - m:?: O, set w¡(s) = 1. IfJ: - ¡.t¡ - m < O, set w¡(s) as a monic arbitrary
Hurwitz polynomial of degree (¡.t¡ +m - J:). Because of ¡; :?:¡.t¡, we have
ór¡W(s)':::::m.
Step 5. If
i = 1,2, .. >, P (9-187 )
))-l(V(s)N¡(s))
GJ(s) = N(s)[W(s)D¡(s)J - 1W(s)Ñ ¡(s) = N(s)Df l(S)Ñ ¡(s)
;-_:: l\I(s)[j- J·(,))T(:;} ~-;~ ·:G(:;')1'(:;:
(9-185)
This shows that the feedback system in Figure 9-33 does implement the open
C¡DJ(S)=!i. Because loop compensator T(s).
I
I{S)Q-l{S)
witb N I{S) = diag {/311 (5), /31 is), ... , /31 p{S)}
where 13 ¡¡(S) is the greatest common divisor of the ith row of N(s). Let 13 2i(S) -
be the least common denominator of the unstable poles of the ith column of D(s)N 2 (s) = L'r,
N Z1(s). Define N2d(s)~diag {f321(S), f3ds), .. . , f32P(S)} and define
Hence we may choos
N2(S)~ N Z1(s)N 2d (S)
It is a rational matrix with only stable poles, that is, poles with negative real
parts. Then we have
The implementation
N 2(s)N 2(s) = N 2d (S) = diag {f321(S), f3ds), ... , f32P(S)} matrix
Now we choose an open-loop compensator T(s) as
T(s) = D(s)N 2(s)D,-1(S) (9-198)
with D,(s) =diag {1X 1(S), 1X 2(S), ... , IXp(S)}
where IX¡(S) are Hurwitz polynomials of minimum degrees to make T(s) proper.
Then we have
Implementation I. V<
G(s)T(s) = N [(s)N 2(s)D -1(s)D(s)N 2(s)D,- l(S) = N [(s)N 2d(S)D t- 1(s)
N(s)=Nl(S)N2(S)=I~ . . .J
0,_,1ls,2 11
...J
Wecompute
Thus, the L(s) and M(
_[
N 2 (s) = -------: 1 [1
(s+l)(s-1) -1
Hence we have
and 1
'N 2(s) = Ni (s)N 2As) = [_ ~ ~21 ] s ~ 1
Now we choose T(s) =D(s)N 2(s)D,-1(S), withD,(s) = diag {1X 1(S), 1X 2(S)}, where
IX¡(S) are Hurwitz polynomials of minimumdegrees to make T(s) proper. We = [-3//4 ~27/4
compute 54 13/4 '
DESIGN OF COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS 525
-
Dc(s) = DC<s)NAs) =
[s +3
O
1
-i- J
0J[ 1 s-
-1 '4
1 = r3
_
+
1
which has row degrees all equaJ to m = 1 and is row reduced. Clearly we havc
.y factor N(s) as ¿¡c¡(f>c(s)DJ(S»~¡Li+l11,i= 1,2, hence we may set Q(s)= l. We compute
~l
lj Js +';' J
Thus, the L(s) and M(s) in (9-174) can be solved from
1 1 O O 1 O
O O O 1 O O
O· . 1- 6- ·0- -1- -Ó
1 1 1 1 O O
---------------------
(9-200)
1 1 O O 1 O
1 000100
+1
0--(-6--0--1--0
1 1 1 1 O O
19 {CC1(S),<xZ(s)}, where .
iake T(s) proper. We =[-3//4 -27/4: 13/2 -9/4: 9/4 O: O 00J
54 13/4 : - 1/4 3 :1 O: O
DE~
526 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
unstable pole-zero ca
as
L(s)= [
(3l/2)S+33/4
(-7/2)s- 2
27/4
-1/4
J _ [( -31/2)s- 6
M(s)- (1/2)s+2
-9J
13/4
This completes the first implementation. where #¡(s) are uniql
to be chosen to make
Implementation JI. We implement G(s)T(s) in Figure 9-33. We have to be decoupled onl)'
D¡/e =1 and J¡ ?:.p¡; hence V(s) =1, D¡(s) =D¡(s), and Ñj(s) = N¡(s). We choose designed, in additior
m = v -1 = 1and rejection, iX¡(S) cannot
the additional feedba
bance rejection.
In order to achie'v
be robust with respe
Because JI - PI - m = 2 - 2 - 1 = -1, we choose WI(S) arbitrarily as s +2, of discussed in Section 9
degree 1. Because.f2 - P2 - m = 2 - 1 - 1 = O, we set w 2 (s) = 1. Hence we have Ifwe introduce a diagl
with h¡(s)/cjJ(s) proper
~J
S +2
W(s)= [ O reduces to p number e
transfer function of th
Clearly, we have 6ci (W(s)O ¡(s)):::; Il¡ +m, i = 1, 2; hence we set Q(s) = 1 and
s+2 3(s
-~~ +2)J
~ Its denominator is a 1
r
K(s) = W(s)N ¡(s) =
-1 s--,¡ h¡(s)/</>(s) to be strictly
h¡{s); there are deg a¡{.
The L(s) and M(s) in (9-191) can be solved as and #¡(s) are coprime,
proper choices of h¡(s) ¡
L(s) = [llS + 21/4 (-3/4)S+3J -1IS-4 -27/4J and disturbance reject
(- 5/2)s - 2 -5/2 M(s) = [ (5/2)s + 2 9/4 or, beca use N(s) in G
#¡(s) is a factor of det r
This completes the second implemental ion. 111
For this example, the total degrees of compensators of these two imple
+
mentations are the same. !n general, the total degree of com1Jensators in the ·r~
----/I~-
second implementation is less than or equai to the one in the rirst implementa
tion.
The design procedure discussed in this subsection can be modified in
several ways. For example, if a stab1e root of det N 2 (s) is very close to the
imaginary axis, it may be retained in #2;(S), instead of being canceled. Instead
of decoupling the plant for each pair of input and output, we may decouple it
for a group of inputs and a group of olitputs. In this case, the plant is to be
decoupled into a block diagonal mat~ix. These modifications ar~ straight
forward and will not be discussed.
'? E
Asymptotic Tracking, DistUl'bance Rejection, and Decoupling . . 111 this .sub
section, we shall design a robust system· foachieve decoüpling,asyú1ptotic I
compensators in the
can be modified in
19 canceled. Instead
, we may decouple it
(a)
pfillg. In this su b
:oupling, asymptotic
zero of (;(s), then <fJ(s) and f3¡(s) are coprime, and the roots of f¡(s) can be arbi loop compensator.
trarilyassigned. From the assignment of f¡(s), we can compute h¡(s) and lXi(S).
We then design an input-output feedback system to decouple the plant as
diag {f31(S)/1X 1(S), f32(S)/1X 2(S), .. " f3 p (S)/lX p (S)}. The resulting system is asymp Now ir a solution T(s
totical1y stable and is decoupled. It will also track asymptotically the reference design can be accoml
input and reject plant-entered disturbances. In this design, if there are param configuration shown
eter perturbations in the plant and compensators (excluding the internal on the solution of (9
model), the decoupling property will be destroyed. However, if the overall Let T¡(s) and Cm;1
system remains to be asymptotica11y stable, the property of asymptotic tracking be written, following
and disturbance rejection will be preserved. is a linear algebraic e
real rational functior
Example 3 necessary and sufficie
Consider the plant in Example 2 on page 524. We have f31 (s) =s - 1, f3 2(S) =S2 - 1,
deg IXI(S) = I and deg Cf.2(S) = 2. Suppose the plant is to be designcd to track
step inputs l/s and reject plant-entered disturbance ofthe form e2', then we have over lR(s). The soluti
In the model matchir
<fJ(s) = s(s - 2) proper rational matri
Since <fJ(s) and f3 ¡(s) are coprime, h¡(s) and IX¡(S) can be found to stabilize the sma11 as possible. Tb
to meet (9-204) is call
system in Figure 9-34(b). Rather arbitrarily, we choose 11(S) =(s +2)3. Then
SI72 and S210.
the solutions of
Before proceedin¡
<fJ(SP1(S) + f31 (s)h l(S) = 11 (s) = (s + 2)3 A(s) and B(s) are left
are h 1(s) =36s -8 and 1X 1(S) =s -28. Ir 12(S)=(S +2)4, the solutions of is unimod ular, that is,
is, unimodular. Folle
<fJ(S)1X 2(S) + f32(s)h 2(s) = 12(S) = (s + 2)4 matrix M(s) tb be roH'
are h 2(S)=1;2 S -16 and 1X2(S)=S2_1;2S_1;4. Next we replace (9-201) and if the R(s) in any fa
Theorem G-8', M(s) i
(9-202) by
in iC. We may also ex
O,(s) = [Cf.l(S)
O
O J=[S-28
1X2(S) O
O
S2_1~2S-~J
l A square polynomial
sum of a11 row degree:
if and only if its row-,
and (;(s)T(s) =diag {s
s-
1
-28'
s
2 s~:~ 1 124 1í
--3-S--3-
we define a q x n poi)
coefficient matrix is e
ii' iVK(s) is 01' ¡"uii row (¡
Once G(sr¡C(s) is implemented as an input-output feedback system inside the
U(s) such that U(s)1V
box in Figure 9-34(a), the design is comp1eted. This part of design is similar
irreducibility and co
to the one in Example 2 and will not be repeated. •
With this preliminary
Model matchillg. The design problcms discussed so far concern only with the We discuss first th
assignment of poles and denominator matrix; the numerator matrix is left un Ir (;(s) is square and n
only if (; -1(S)(;",(s) Ü
specified. In this subsection, we discuss the assignment of denominator matrix
~ore complicated. . L
as well as the numerator matrix 01', equivaJently, rhe entire overall transfer
G(s) and Gm(s). The I
'matrix. This problem i~ often referred to as the exact módel matching problem.
, Consider a plant with q x:p proper rational matrix (;(s). The desired mQdel
is assumed to have the q x r pr.oper rational matrix(;m(.~)' The problem is to
find a'configuration and co'mperrsators lor the plailt so that the resulting overall n If Gm(s)= 1, then the solut
reduces also to the minim
systcm has (;m(S) as its transfermatrix. We study theproblem by usingan open-
DESIGN OF COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS 529
ts of f¡(s) can be arbi loop compensator. Let T(s) be a p x r open-Ioop compensator so that
)mpute h¡(5) and íX¡(5).
G(s)T(s) = G",(s) (9-204)
ecouple the plant as
ng system is asymp Now if a solution T(s) exists in (9-204) and is a proper rational matrix, then the
,totically the reference design can be accomplished by implementing T(s) in the input-output feedback
~n, if there are param configuration shown in Figure 9-31 or 9-33. Thus the design problem hinges
tcluding the internal on the solution of (9-204).
;)wever, if the overall Let Ti(s) and Gmi(s) be the ith column of T(s) and Gm(s). Then (9-204) can
)f asymptotic tracking be written, fol1owing (2-2), as G(s)T¡(s) = Gmi(s), i = 1,2, ... ,r. Each equation
is a linear algebraic equation studied in (2-36) with entries in lR(s), the field of
real rational functions. Hence Theorem 2-4 is directly applicable. Thus the
necessary and sufficient condition for the existence of a solution T(s) in (9-204) is
,)=s-1,P2(s)=s2-1, rank G(s) = rank [G(s) G",(s)J
be designed to track
form él, then we have over lR(s). The solution T(s) is general1y a rational matrix, proper or improper.
In the model matching problem, we are interested in only solutions which are
proper rational matrices. Furthermore, we require the degree of T(s) to be as
found to stabilize the small as possible. This problem of finding a proper T(s) with a minimal degree
11(5) =(s +2)3. Then to meet (9-204) is cal1ed the minimal design problem 22 in References S95, S125,
SI72 and S21 O.
Before proceeding, we digress to introduce some concepts. Recall that
A(s) and B(s) are left coprime if and only if their greatest common left divisor
he solutions of is unimodular, that is, the R(s) in any factorization [A(s) B(s)] = R(s)[ Á(s) B(s)J
is unimodular. Following this, we may define a q xn, with n ?q, polynomial
matrix M(s) to be row irreducible, or its column submatrices to be left coprime,
'e replace (9-201) and if the R(s) in any factorization M(s) = R(s)M(s) is unimodular. Fol1owing
Theorem G-8', M(s) is row irreducible if and only if rank M(s) = q for every s
in C. We may also extend the concept of row reducedness to nonsquare matrix.
A square polynomial matrix A(s) is row reduced if deg det A(s) is equal to the
sum or all row degrees or A(s). This definition implies that A(s) is row reduced
if and only if its row-degree-coeflicient matrix is nonsingular. Following this,
we define a q x n polynomial matrix M(s) to be row reduced if its row-degree
coefficient matrix is of rank q or of ful1 row rank. Similar to Theorem G-l1,
lck system inside the ir M(s) is 01' fun row rank in iR (S), [hefe ex;::;,::; 8. unimodular polynoiHici< maÜiX
rt of design is similar U(s) such that U(s)M(s) is row reduced. We may similarly define column
I irreducibility and column reducedness for nonsquare polynomial matrices.
With this pre!iminary, we are ready to study the minimal design problem.
concern only with the We discuss first the condition for the solution T(s) in (9-204) to be proper.
ator matrix is left un Ir G(s) is square and nonsingular, the answer is very simple: T(s) is proper if and
f denominator matrix only ir G-1(S)G m (s) is proper. Ir G(s) is not square, the situation is slightly
ntire overal1 transfer ~ore complicated. Let 4J(s) be the least coml1)on denominator of al1 entries of
del matching problem. G(s) and Gm(s). The inultiplication of 4J(s) to (9-204) yields
l The desired model A(s)T(s)=B(s) (9-205)
l The pro blem is to
.t theresulting overall 22 Ir Gm(s) ~ I~ then the solution '-i(s) in (9-204; is a rig~t invers~of ~(sj.· Hencetheinverse problem
lem by usirig an open- reduces also to the minimum design problem. .
530 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS [
where A(s) = </J(S)G(S) and B(s) = </J(s)G", (S) are q x p and q x r polynomial does have a full ce
matrices. If G(s) and, consequently, A(s) are of ful1 row rank, there exists a algorithm.
unimodular polynomial matrix U(s) such that U(s)A(s) is row reduced. Consider
U(s)A(s)T(s)= U(s)B(s) (9-206) Theorem 9 -28
IrY(s) in (9-209) is a rr
We assert that if G(s) is of full row rank in lR(s), then T(s) is proper if and only if G-14', then NT(s)Di
b..¡(U(s)A(s)) :::O:b'i(U(s)BCs)) i= 1, 2, .. "q (9-207)
Proof
where b,.¡ denotes the ith row degree. Ir A(s) is square, this assertion is essenti
Ir Y'¡'e has rank r, ther
ally Theorem G-IO. The proof for the nonsquare case is similar to the one of
hence N T(s)Di 1 (s) is
Theorem G-l Oand will be omitted.
The properness condition in (9-207) does not tell us how to find a T(s) with a T(s) =
minimum degree. In the following, we shall introduce such a method. The
method also gives the properness condition of T(s); hence the condition in IrY'¡'e has no ful1 COIUl
(9-207) is not really needed. Let T(s) = NT(s)Di 1 (s). Then (9-205) becomes algorithm, at least on
A(s)NT(s) = B(s)DT(s) or and Y/,,) = 0, where Y'
basis, Yhe has a fui! ce
[A(s) B(s)] [ - ~~i;~J = O (9-208)
jth column degree of
This is the polynomial equation studied in Equation (G-90); hence all discussion
in Section G-6 is directly applicable. Let (W, lR(s)) denote the right null space IfT(s) is proper, as s
of (9-208). Its dimension is, following Theorem 2-5, equal to Hence ir Y¡'e has no
the theorem.
r~ p +r - rank [A(s) B(s)] = p + r - rank [G(s) Gm(S)]
In this theorem,
Now we may apply Theorem G-14' to solve (9-208). We form the generalized minimal basis is eSSé
resultant T k from the coefficient matrices of A(s) and B(s) as in Theorem G-14'. 9-33). The theorem i
We then search its linearly dependent columns in order from left to right by case. Ir r < r, clearl:
using the column searching algorithm. There will be exactly r primary depend If ;::::O:r, Y(s) and \'¡"
ent columns in T k • Let the (p + r) x;: polynomial matrix Y(s) be the solutions holds. In this case,
corresponding to these r primary dependent columns. Then Y(s) is, as in have increasing colur
Theorem G-14', column irreducible and column reduced, and is a minima! the 11rst 1" coiumns VI
polynomial basis of the right nuií space oí" (9-208). of (9-204). This com
Let Y he be the column-degree-coefficient matrix of Y(s). For convenience
r
of discussion, we assume = r and partition Yhe and Y(s) as Example 4
Find a 3 x 2 minima
- NT(S)] (9-209)
Y(s) = [ DT(s)
where Y/le and NT(s) are p x r and Y'hc and DT(s) are r x r matrices. " We note
that, because of the properness assumption of G(s) = A - l(s)B(s), the
[ - N'(s) D'(s)]' in Theorem G-14' always has the properties that bciN(S):::;;
be; D(s) and D(s) is column reduced orD he has a full column rank. In the minimal
design problem, A(s) in (9-208) is not necessarily square and A - 1 (s)B(s) may not Yhe is not computed 1
23"if
be defined; therefore there is no guarantee that DT/le, the coiumn-degree sorne j. However, ¡fral
coefficient matrix of DT(s) in (9-209), is of full columri rank. We note that Yhe elernentary colurnn ope
DESIGN OF COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS 531
1d q x r polynomial does have a ful1 column rank as a consequence of the coiumn searching
Irank, there exists a algorithm.
w reduced. Consider
(9-206) Theorem 9-28
UY(s) in (9-209) is a minimal polynomial basis of (9-208) obtained as in Theorem
; proper if and only if
G-l4', then NT(s)D·¡ les) is proper if and only if Y¡'e has rank r.
(9-207)
Proof
s assertion is essenti
U Y¡'e has rank r, then D'r(s) is column reduced and (jei N T(S) :S;(jei DT(s), for al1 i;
similar to the one of
hence NT(s)Di les) is proper. To show the converse, we consider
N to find a T(s) with a or (9-21 O}
uch a method. The
.1ce the condition in UY¡'e has no full column rank, there is, as a consequence ofthe column searching
hen (9-205) becomes algorithm, at least one column, say thejth column, such that óejNT(s) > (jej DT(s)
and Yhej = O, where Y¡'ej denotes the jth colu!!1n of Y¡'c- 23 §ince Ves) is a minimal
basis, V he has a ful1 column rank. Hence ifY hej =0, then Yhej=/=O. Let skj be the
(9-208) jth column degree of Ves). Consider the jth column equation of(9-21O):
T(s)DTj{s)S-k j = NTAs)S-k j
); hence al1 discussion
e the right null space UT(s) is proper, as s---> 00, we have T(ca )Y/,ej = O = Yhej =/= O. This is not possible.
al to Hence if Y/,e has no ful1 column rank, then T(s) is improper. This establishes
the theorem. Q.E.D.
In this theorem, the condition that Ylre is of ful1 column rank or Ves) is a
form the generalized minimal basis is essential; otherwise the theorem does not hold (see Problem
as in Theorem G-14'. r
9-33). The theorem is developed for the case = r. We discuss now the general
from left to right by r
case. U < r, c1early no solution T(s), proper or improper, exists in (9-204).
tly r primary depend If r? r, Ves) and Y¡'e are (p + r) x r matrices of rank r, and Theorem 9-28 still
Ves) be the solutions holds. In this case, in order to obtain a minimal solution, we arrange Ves) to
Then Ves) is, as in have increasing column degrees, that is, (jcl Y(s).::s; (je2 Y(s):S; ... :s;(jcf Ves). Then
d, and is a minimal the nrst r columns with a nonsigular \\e wil1 give the minimal proper so!ution
of (9-204). This compíetes the discussion oí' the minunai design proolelÚ.
s). For convenience
lS Example 4
(9-209)
"matrices. We note
(s) = A -l(s)B(s), the
.. lerties that ÓciN(s).::s;
Find a 3 x 2 minimal proper solution of
-s
s+3
-3s-7
s+3
l
rank. In the minimal
jA -:-.1 (s)B(s) may not
23 Ir Yh,· i~ not computed by· using the column searching algorithm, we may not have Yh,j = O, for
the column-degree can
sorne j. H·owever, ifrank Y,,, < r, bne of its cciluinn be transformed into a zero colurnn by
:k. We note that Ylle elementary coluinn opeiations.
532 LINEAR TIME"INVARIANT COMPOSITE SYSTEMS D
The multiplication of S(S + 1)(S + 3) and the substitution of T(s) = NT(s)Di 1 (S) to the three primar)
yields to compute
o O O: O O:
o
3 O O: O O:
:3 - - - 0- -- -0- -: - 6---- -0- -:- - ó--- -0- - - - Ó- : - -0- - - - - Ó which yields immedi;
4 -6 -3 : 1 -7: 3 O O: O
-----------L---------~------------l--------
O r
y(,)~ 1
1 O -3 : 1 - 1 : 3 O O, O O
1 -8 -4 : 2 -10 :L 4 -6 -3 ,: 1
___________ L
-7_
O O -1 : 1 -1: 1 O -3 ,1 -1
O -2 -1 :
___________ L 1 - 3 : L 1 -8 -4 : 2 -10_ ~-----
• O O -1 1 -1
o ,
"., O -2 -1 1 -3
Y(s) has rank 3 for ev¡
reduced, for Y,,~ is 01
100 O O O O O O O basis of the right nul
rank 2; hence we ha"
O 1 -1/2 1/6 -7/6: 1/2 O O O O
O O 1 1 -1 1 O -3 1 -1
O O O O 1,0 O O O O
'l O
O~
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
1
O
O
O
O
1 O
01
O O
O -1
1
-0 J'
-1
[x x x: x O: O x x: O O]
i i i This is a minimal pro
"Wherexdenotes thenonzero col~mn, and"O thezerócolumn. Thiscomputation
is very easy to carry out by hand. The notation : is explained in Appendix A. The numerator m
There are four linearly depe~dent cblumns in TI"; three of them are primary discussed in Section í
dependent columns as indicated by thearrows. Note that in Theorein G-14', the numerator matrix
"linearly dependent columns will appear only in B-rows (see dual of Theorem to achieve pole or den
G-13). In the minimal design problem,dependent columns mayhowever introduce any undesil
appear in B- as well as A-columns as shown in this example. Corresponding numerator matrices ir
DESIGN OF COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS 533
to the three primary dependent columns, we use the dual formula of (A-II)
to compute
o O O
-1 O O
O I 1
103
1 O O
--------
O 1 O
=3. We form T l as O O O
Igorithm to compute O O I
O O I
O O O
1 O
O -1
-~J
O
I and
O -1
n. This computation
lÍned in Appendix A. The numerator matrix of a plant cannot be affected by state feedback, as
of them are primary . discussed in Section 7-3, nor by output feedback; hence the only W'á,/ toaffect
:ü in Theorem G-l.4', . the numerator matrix is by direct cancel1ation. In the design of compeü.sators .
·See dual of The'orem to achieve pole or denominator matrix assign~e~t, we have beeÍl careful ilot to
lumns may however introduce any undesirablepole-zero cancel1ation. This is possible becausethe
lple. Corresponding numerator matrices in these designs are not specified. In the model matching,
534 LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
because the numerator matrix is also specified, undesirable pole-zero cancella Then it can be showr
tion may be unavoidable. Therefore in the model matching, the choice of Reference S85, whe
Gm(s) must be very careful. Otherwise, even if we can find a proper T(s) and D-l(S)N(s); the metl
implement it in a well-posed input-output configuration, the design is still rings. The method I
unacceptable. S51,andS218. Wefir:
[Lo Mo : L l
9-8 Concluding Remarks Then a solution exisl
solution of the equati<
In this chapter, we studied a number of topics: characterization of composite ical comparisons of '
systems by their transfer matrices, stability of feedback systems, and the design In addition to Eq\
of compensators in the unity feedback systems and the plant input-output
feedback systems to achieve arbitrary pole placements, asymptotic tracking,
disturbance rejection, and decoupling. We also studied the implementation of also arises in the desig
open-loop compensators by using plant input-output feedback systems. from the Diophantil1E
All the design problems in this chapter hinge essentially on the solution of See References S88, S
the Diophantine equation: asymptotic tracking a
the internal model is c
L(s)O(s)+ M(s)N(s)= F(s) (9-211 a)
bypasses the equation
or O(s)L(s)+ N(s)M(s)= F(s) (9-211b)
this texto
Several methods are available to solve this equation. First, we may apply a In this chapter we
sequence of polynomial elementary matrices to transform [O'(s) N'(s)]' into a plant input-output fe
Hermite row form or a upper triangular form as in (G-33): compensator is not ~
N(s)O j 1(s)Nc(S). In t
U 11 (s) U 12(S)J [O(S)J = [R(S)J assignable, and the n:
[ Uds) Uds) N(s) O Dc(s) in N(s)Oj l(s)D c
Then a solution exists in (9-211a) if and only if R(s) is a right divisor of F(s). A completely concel1ed :
general solution L(s) and M(s) can then be obtained as in Problem G-14o See Dc(s) in the plant inI
References S34, S139, and S1890 Another method is to find an observable feedback configuratio
realization {A, B, C} of O-l(s)N(s) as in (6-131). We then use the operator [lD L(s), M(s), and Oc(s) il
defined in (6-225) to compute ITv(F(s)). Ifwe use the method in Problem G-15 Dc(s) in the latter, and
or any other method to compute In the design, the
fracUonal formo Tw
[I(s) == D(s)Q J. (s) -:-- f 1 {,~')
ll1e Ílrst method, we
with briFl(S) <briO(s), then we have ITv(F(s» = F1(s). We then express Fl(s) sequence of polynomi
in terms o[ the basis in (6-223) as one as in (G-33). The
In this case, we must '
Fl(s)=L(s)E 1
reduced one. In the
Then a solution L(s) and M(s) exists in (9-211 b) if and only if A -l(s)B(s). We fom
ofA(s) and B(s), and n
ránk[B AB ... An-1B] = rank [B AB ... An-1B El] left to right. From the
tainaright-coprime fr
or, equivakntly, El lies in the space spanned by the column of
reducéd; in fact, it is a
[B AB . 0'0 An-1B]. The solution of El =[B ABo" A"-1B]P is the
Once a right-cop
represent"ation.of M(s) in (9-211 b)with respect to the basis L(s), that is, M(s) =
, step'is to co"mpute th<
L(s)P. Next we compute
necessary because we
N(s)M(s) = D(s)Q2(s) + F 2(S) search linearly indepe
CONCLUDING REMARKS 535
e pole-zero cance\la Then it can be shown that L(s) in (9-211 b) is given by L(s) = Ql(S) - Q2(S). See
ching, the choice of Reference S85, where the method is developed only for strictly proper
Id a proper T(s) and D-l(S)N(s); the method, however, has been extended to linear systems over
n, the design is still rings. The method presented in this chapter fo\lows the line of References 18,
S51, and S218. We first translate (9-211) into a set oflinear algebraic equations as
[Lo Mo:L 1 M 1 :-·<L/II MmJSm=[F o F 1 .. , FU+/IIJ~F
Then a solution exists in (9-211) if and only if F is the row space of S",. The
solution of the equation yields immediately the required L(s) and M(s). N umer
ization of composite ical comparisons of these three approaches are not available at presento
;tems, and the design In addition to Equation (9-204), the equation
: plant input-output L(s)D(s) + N(s)M(s) = F(s) (9-212)
asym ptotic tracking,
le implementation of also arises in the design of multivariable systems. This equation is quite different
dback systems. from the Diophantine equation, and its solutions are much more complicated.
Iy on the solution of See References S88, S 139 and sno. The equation is essential in the study of
asymptotic tracking and disturbance rejection (without robust). In our design,
the internal model is chosen as c/J -1(s)lp. This results in a robust design and also
(9-211 a)
bypasses the equation. Consequently, the equation in (9-212) does not arise in
(9-211b) this texto
irst, we may apply a In this chapter we studied two feedback configurations: unity feedback and
[D'(s) N'(s)]' into a plant input-output feedback. In the former, the denominator matrix of the
compensator is not assignable, and the resulting overall transfer matrix is
N(s)Djl(S)N c(s). In the latter, the denominator matrix ofthe compensatorsis
assignable, and the resulting overall transfer matrix is N(s)Dj 1(s)Dc(S). The
Dc(s) in N(s)Dj 1(s)Dc(S) can be used to generate the desired numerator or be
~ht divisor of F(s). A completely concelled by O f(S). The reason that we have freedom in assigning
Problem G-14. See Oc(s) in the plant input-output feedback configuration but not in the unity
) find an observable feedback configuration is very simple. There are three sets of parameters
use the operator IT D L(s), M(s), and Dc(s) in the former, but only two sets of parameters N,(s) and
lOd in Problem G-15 Dc(s) in the latter, and two sets of parameters are needed to achieve the designo
In the design, the transfer matrix G(s) must be expressed in a coprime
fractional formo Two such procedures are developed in Appendix G. In
Lhe nrsi method, we íirst find a noncoprime righi fraciion anci then use el
fe then express F 1(s) sequence of polynomial elementary operations to reduce it to a right coprime
one as in (G-33). The D(s) obtained by this process may not be column reduced.
In this case, we must find a unimodular matrix to transform D(s) to a column
reduced one. In the second method, we first find a noncoprime left fraction
A - 1 (s)B(s). We form a generalized resultant from the coefficient matrices
of A(s) andB(s), and then search its linearly independent columns in order from
left to right. From the primary linearly-dependent columns, we can readily ob
tain a right-coprime fraction: Furthermore,the resulting O(s) is always column
)y the column of reduced; in fact, it is also row reduced and -inthe polynomial echelon formo
A/- 1 BJP is the
Once a right-coprime fraction G(s) = N(s)D - 1(.s:) is obtained, the next
sL(s)~that ¡s, M(s) =
step is to compute the -row index of G(s). In the scalar case, this step is un
necessary because we have.v = n =deg D(s). In the multivariable case, we
search linearty ilidependent rows of Sm in order from top to bottom. Once all
536 LINEAR T1ME-INVARIANT COMPOSITE SYSTEMS
N rows of the last block row ofS m become Iinearly dependent, the mis the row 9-5 Are the feedback sys
indexo The total number of Iinearly independent N rows in Sv should be equal
to deg det D(s). This can be used to check the correctness of the computation.
-J----
+ I 1 1_
With the computed v, the chosen F will be in the row space of Sv- 1> and the - t L- -
sr, + I J f
l~
solutions of (9-211) yield immediately the required compensators.
The search of linearly dependent rows or columns of a matrix is basically
an ill-conditioned problem, beca use a zero row can be made to be nonzero by
introducing a very small number. Consequently, no matter what method,
singular value decomposition or Householder transformations with pivoting, is Figure P9-5
used, difficulties remain. In om design, the order of rows cannot be changed.
Hence the singular value decomposition or Householder transformations with 9-6 Is the feed back systel
pivoting cannot be used. However we may use Householder transformations
without pivoting and back substitution. These two methods are numerically
stable and can be used in the designo We emphasize that the numerical stability
of an algorithm and the ill- or well-condition of a problem are two distinct
properties. Since the search of dependent rows is an ill-conditioned probiem,
even though we use only numerically stable methods, difficulties may still
arise in the computation. 8180 stable? asymptotic
Ll¡(s)Ll 2 (s)det (1 + (;ds)G 2 (.
Problems
9-7 Given the plant 9(05)
9-1 Consider system so that the poles of
~. 05+2
and . g2(S)=-· 9-8 Given the plant 9(05)
05+3 degree 1 so that the poles of
Is the tandem connection of 9¡(S) followed by 92(S) completely characterized by its overall proper compensator C(s) c
transfer function? Verify your result by showing the controllability and observability of -1, -2, -3, and -4.
its composite dynamical equation.
9-9 Consider the plant gi
9-2 1s the feedback system with 9¡(S), given in Problem 9-1, in the forward path and
pensator in the llnity feedb
92(05) in the feedback path completely characterized by its overall transfer fllnction?
com pensator of dcgrec 1 in
9-3 Consider
:t_ll
:~:j
f) l s !l·10 Consider the íeedb<\!
iúnctlOn"l Let N(s)=N¡(s)
~J. N¡(s)N¡(s)/D¡(s), wíth deg 1
compensator P(S)fDc(s) an!
05+2
Dc(s) be Hurwitz? lf Dc(s)
Are their paralIeI and tandem [(;¡(s) folIowed by (;2(05)] connections controllable? obser the system be acceptable?
vable? completely characterized by their transfer matrices? wilI be acceptable? See Prc
I 1
A
G¡(s) =
05+1
t
O and
A
G 2 (s) =
l
05+2
O
s
controlIable and observable? Note that (; ¡(s) and (;2(S) have the same set of poles.
O
1.
+l
J
Figure P9~10
PROBLEMS 537
ldent, the m is the row 9-5 Are the feedback systems shown in Figure P9-S BIBO stable? asymptotically stable?
; in Sv should be equal
,s of the computation.
,pace of Sv _ 1, and the
+
Ti~"(S-+-11
I I I ¡
1.,-- s h
_s_+_2_ I 5(S
I
+ 1)
·ensators.
fa matrix is basically
lade to be nonzero by
matter what method,
ltions with pivoting, is Figure P9-S
",s cannot be changed.
9-6 Is the feedback system shown in Figure 9-11 with
: transformations with
lolder transformations 1 I
l I
thods are numerically
the nurnerical stability
S2 -1
G¡(s)= _1_ G,(s) =
-
s +2
1
s +3
1 J
,blem are two distinct
-conditioned problem, s- 1 s +1 s ~3
;, difficulties may still
BIBO stable? asymptotically stable? Use both the coprime fraction formula and
+ (;1(S)(;2(S)). Which one is simpler to use for this problem?
6¡(s)6 2(s)det(I
9-7 Given the plant g(s) = (s - I)/s(s - 2). Find a compensator in the unity feedback
system so that the poles of the resulting system are - 1, - 2, and - 3.
9-9 Consider the plant given in Problem 9-8. Find a set of lwo poles so thal the com
in the forward path and pensator in the L1nity feedback system is a constant. Find a set of three poles so that (he
1 transfer fLlnction'J compensator of degree 1 in the uni(y feedback syslem is improper.
s +1l
9-10 Consider the feedback system shown in Figure P9-1O. What is iíS overall transrer
function'! Let N(s) = N ¡(s)N 2(s), where 1'/2(S) can be cancelied. Snow ihat, ¡Or any 9f(.I') =
;-7- 2j'
N[ (s)Nf(s)/Df(S), with deg DfeS) - deg (N 1 (s)Nf (s)) ~ deg D(s)- deg N(s), there exist proper
compensator P(s)/De(s) and Ne(S)/De(s) to achieve the designo If Df(s) is Hurwitz, will
s +2 De(s) be HLlrwitz? lf De(s) is not Hurwitz and if the system is implemented as shown, will
:tions eontrollable? obser the system be acceptable? Can you find a different but equivalent implementation which
will be aceeptable? See Problem 9-32 and Reference S34.
. \ PIs) ~.
ID;W~
Find a compensator so that the poles of the overall unity feedback system are - 1, - 2 ± j,
Find a compensator in 11
and the rest chosen as - 2.
9-13 Which of the following matrices are cyclic? 9-19 Consider (;(5) = N(.
~ ~
---.!_
5+ 1
5-2 1 l ~ 5 +2
(5-1X5+1)
Let N(5)O-1(5)=O-l(S)Ñ
M(5) are left coprime, so ¡
r
Gl(s)= 1 G 2(5) = 1 9-20 Lel D-l(S)N(s) be
nomial ~atrix ifand only
5+3 (5 - 1)2 5- 1
malrix M(s) such lhat M(~
~
G,(5)=
.
l5-1~;5+1) (5-1~1(5+1)J ~
2
r:
G 4 (5)= O
9-21 Consider lhe feedba
h'lck syslem is asymptol
where aH poles of r(s) He ¡
5-1 5-1 coprime. Show thal
9-14 Show that ir all elements of G(5) have no poles in common, then G(5) is cyclic. Note e(s)=(I +
that repeated po\es in individual g¡j(5) are, however, permitted.
has no cIosed righl-half-pla
9-15 Find the set of 2 x 1 vectors t such that 6«;(5» of6«;(5)t), where
5 +1
~
G(5) =
5-1
-2
=!'J
s
2
f
s- 1 s
9-16 Show that ifthe rational matrix C(5) is cyclic, so is (:(5) = C(5) + K for any constant K.
G(s)
A = l~ ~ °l [52 0J-I[1° °5 0J
lOs
=
1
°° s
:k system are -1, -2 ±j, Find a compensator in the unity feedback system so that the resulting denominator is
lS + 1)2
Dj(s) = [
_°
Is the resulting Gj(5) cyclic? Compare your result with the one in the text.
9-19 Consider G(s) = N(S)O-I(S)M(s), where N(s), O(s), and M(s) are polynomial matrices.
Let N(s)O - I(S) = O - 1 (s)N(s), where O(s) and Ñ[s) are left coprime. Show that if O(s) and
M(s) are left coprime, so are O(s) and Ñ(s)M(s).
9-20 Let O-I(s)N(s) be a left coprime fraction. Show that M(s)D-l(S)N(s) is a poly
nomial matrix if and only if O(s) is a right divisor of M(s), that is; there exists a polynomial
matrix M(s) such that M(s) = M(s)O(s).
9-21 Consider the feedback system shown in Figure P9-21. It is assumed that the feed
o'lck system is asymptotically stable. Let G(s) = O -1(s)N(s) and r(s) = D,~ l(s)N,(s)ro,
where all poles of r(s) He inside the closed rigbt-half s plane and D,.(.1) and N..(s) are left
coprime. Show that
e(s) =(1 +G(sW lr(s) = (O(s) + N(s)t ID(s)O,:-l(s)N..(s)ro
1, then G(s) is cyclic. Note
has no closed right-half-plane poles if and only if O,.(s) is a right divisor of O(s).
), where
~(S) ~(s)
figure PS-2'¡
:(s) + K for any constant K.
9-22 Consider the unity feedback system shown in Figure 9-16. Let G(s) = N(S)O-I(S) =
0-1(s)N(s) be coprime fractions of the plant transfer matrix. Let X(s) and Y(s) be poly
nomial matrices such that X(s)D(s) + Y(s)N(s) = I. Show that for any ralional matrix
H(s) with all poles inside the open left-haH s plane, the compensator
stabilizes the unity feedback system. lfH(s) is proper, will Cls) be proper'? [An5wer: No.
TryG(s) ==(s +1)/s(s +2) and H(s)=O.J
9-24 Consider the plant transfer matrix 9-32 Show that the tra
N(s)[De(s)D(s) + Ne(s)N(s
A
G(s) =
[sI
is nonsingular and can t
ditions, will there exist ¡:
Verify that the row index of G(s) is 2. Design an input-output feedback system to have design?
(s +1)2 O J
~
DJ(s) = [ O (s +1)3
[ ,~
s +2 s +2
as its denominator matrix. Is the system decoupled? Are all compensators proper? Is 1
the system well posed? ls there any unstable pole-zero cancellation?
s+2
9-27 Consider the plant given in Problem 9-24. Design a well-posed input-output
feedback system to decouple it without involving any unstable pole-zero cancellations. 9-35 Let Y(s) be a n x J
í = 1,. 2, ... , p, and let )
¡J.¡,
9-28 Consider Corollary 9-24 with G(s) strictly proper. Show that if Dcill¡(s) = De¡O(s) Consider
and OJ/¡e = D'K' that is, column degrees and column-degree-coefficient matrices of OJ(s)
and O(s) are all equal, then the compensator Co(s) = D; 1 (s)L(s) is strictly proper. The
compensator C 1 (s) = D; 1 (s)M(s) is generally proper. Show that
9-29 Considcr Equations (9-70) and (9-74). Show that if 111::::: v- 1, for any real k¡, í =
1, 2,. ,eL, the following
o •
[D co Neo: '0'; Dc", Nc",]+[k 1 k 2 o •• k.] [ -80 A o :"': -8", A",] if and only ir V(s) is colu
Rererence S95 and can be
meets (9-74), where (;(=(111+ l)q-n and [-Ro Ao : ... : -B", A",] is a basis of the ininin12..l. 1fin" t~t ,~¡ =
(J.-dimensional kIt ¡"¡uii space o[S", computcQ as in Appendix G. ThíS ,s;:, ¡),u<c¡,o,CIL¡ ''''~L'G" show the equaiity, we she
of all solutions of (9-74).
relation Yd = Y/KCX, where:
9-30 Consider G(s) = N(s)D - I(S), T(s) = D r- 1 Nr\s), and
Go(s) = G(s)T(s) = N(s)[O,(s)D(sW 1 N,(s)
Show that if Dr(s)D(s) and N,(s) are left coprime, any feedback implementation of Go(s)
reduces es'sentially to an open-loop system.
9-31 Consider
G(s)=
A
. I
[O I J[S2 + 1
s +1 1
Design a decoupled system to track step reference inputs and reject plant entered distur
bance of the type ero . .
PROBLEMS 541
9-32 Show that lhe transfer matrix of the feedback system in Figure P9-32 is GJ(s)=
N(s)[D,(s)D(s) + N,(s)N(sW 1 K(s). Let G(s) = N(s)D- '(.1')= Nl(S)N 2 (s)D- 1 (s), where N 2 (s)
is nonsingular and can be cancelled. Given GJ(s) = NI (s)D j 1 (s)NJ(s), under what con
ditions, will there exist proper compensators D,- 1 (s)K(s) and D,-l(S)Nc(s) to achieve the
design?
'eedback system to have
Figure P9-32
9-33 Consider G(s) = N(s)D-l(S) with óciN(s)'::;ÓciD(S), for aH i. Is it true that G(s) is
proper ir and only ir D(s) is column reduced? Hnot, find a counterexample.
[ s~ ::~] 1 [1.1'+.1'2::+ 3)
2(,+1)
(s-I~;s+3)
J
,(s)
:k implementation of Go(s)
Consider the n x m m:
A
First we assume a 11 =/=
Elementary Transformations the ith row, for i =o 2, ~
1 O O O O 1 O O O O
O 1 O O O O 1 O O O O O O x
E-l-
1 - O O 1 O O E21 = Ez E-l-
J - O O 1 O O O O O O
O O O l/e O O -d O 1 O
O O O O 1 O, O O O 1 O O O O
(A-2)
where X denotes possib
They are again elementary matrices. The premultiplicationof E¡ on A operates
called the gaussian elim
on the rows of A, whereas the postmultiplication of E¡ on· A operates on thé matrices. 1
columns· of A. For example, EzA interchanges the second and flfth rows;
EJA adds the product of the second row and d to the fouáh row of A; whereas
AE J adds the product of thefourth column and d to the second column ofA. 1 AII diagonal elemenls in (A
by llsillg additional column
542
GAUSSIAN ELIMINATION 543
A= ~Zl [a" a zz a Z3
a'-l
a Zm
(A-3)
anl a nz a n3 a nm
First we assume all ,p O. We add the product of the fmt row and (-a¡JÍall) to
s the ith row, for i = 2,3, ... , n; then we have
o O
1 O
°l
O
a
O
ll
:-~?- -aa- F --. ~ ~ -_aa-:
I 22 23 2m
T
··, a~3
I
O 1 O O O
E. 3 = O O 1 O O
O en 1 O 1 O O 1
1
a;m
O d O 1 O
X.
x
;(
x
f..
x xl
j,
"
O O O O O O x x x x
or
1 O O O O O O x x x
O 1 O O O O O O O O O O O x x x
-d O 1 O O O O O x x
O O O 1 O O O O O O
(A-6)
(A-2)
where x denotes possible nonzero elements. This process of transformations is
ion ofEjon Aoperates cal1ed the·gaussian eliminaríon. The matrices in (A-6) are upper right triangular
on A operates on th<; matrices. 1
econd and fifth rows; .
Jrth row of A; whereas
1 AII diagonal elements in (A-6) are nOt nccessarily nonzcro. Il is possiblc lO makc lhern nonzcro
e second column of A.
byusing additional coturnn operations. Sce also Equation (G-28).
544 ELEMENTARY TRANSFORMATIONS
This process will fail if any of a¡¡, ah aj3' or other diagonal element is zero. with x'x = 1. where x
Even if all of them are nonzero, the process may still encounter some com (1 - 2xx')' = 1 - 2xx' =
putational difficu\ty. For example, if a¡¡ is very small in magnitude, then
(- a¡da¡ ¡) = eil will be a very large number. lf there are errors in representing H'H=
([ Ij, these errors will be greatly amplified by e¡ ¡ in the operations. F urthermore,
we ha ve H 1 = H'; hl
these errors may propagate and be reamplified at \ater stages. Consequently,
singular values of H "
the final resu lt may be overwhelmed by errors. Because of this phenomenon, the
ratio of the largest sin
gaussian elimination is said to be numerically unstable.
the use of Householde
To overcome this difficulty, before carrying out the elimination. we first
of a problem lsee Prol
search the element with the largest absolute value 2 in the first column of A,
An important pro
and then interchange this row with the f¡rst row by using a matrix of the form of
two vectors a and b 1
E 2 . We then carry out the elimination for the first column. Before carrying
exists a Householder t
out the elimination for the second column, we search the largest element in
x = la - b)flla - b11 2 , th
magnitude in the second column of K¡A excluding the first element, and bring
it to the a~2 position. This process of elimination is called the gaussian elimina
lion with parlial piVOl ing. In this process, we have Ieijl.:; 1, and the errors will
not be amplified. Hence this process is more stable numerically than the
gaussian elimination. By using partial pivoting, we conclude that every matrix Since a'a = b'b by aSSl
A can be transformed into the form in (A-6) by using a sequence of elementary is invariant under trar
row operations.
The stability of the gaussian elimination can be further improved by using
complete pivoting. In this case, we must use elementary row operations as well
as elementary co\umn operations. We first search the largest e\ement in magni
tude among all elements·of A. We then move the element to the f¡rst column by This establishes. the ·a
an elementary column operation and then to the first row by an elementary row K¡ and K 2 in (AA) ane
operation. We then carry out the elimination of the first column of the trans Consider the n x m
formed A. After this elimination, we repeat the process for the remaining of the first colllmn al
matrix. This process is called the gaussian elimination with complete pivoting. there exists a Househ·
This process is more stab\e numerically than the one with partial pivoting; how of the right-hand-side I
ever, it is more costly becallse the search ofthe largest element in magnitllde is a row of H ¡ A and repea
time-consuming process. Hence complete pivoting is not used as often as partial matrix. Proceeding i:
pivoting. According to Reference Sl38, from the point of view of overal1 per sequence of Househol(
formance which ineludes efficiency, accuracy. reliability. generality. and ease Di The numerical sta!:;
iíiUUilion or columns.
use, the gaussian elimination with partiai pivoting is satisfactory for most
general matrices. we permute the columi
apply the Householde
process, the diagonal e
*A-2 Householder Transformation order of magnitude.
pivoling.
The numerical stability of the elimination can be further improved by using the We note that Hom
Householder trarisformation. Let 1 be a unít matrix. A HOLlseholder trans K¡ and K 2 in (A-4)anc
formation is a ~quare matrix of the form 3 after the application of
H=I -'--2xX:
'AIl argurnents still hold ir H is defined as H = [ - 2xx*, where x* is ¡he cornplex conjugate transpose 4Thesign rnay bechosen to b,
or x. In this case, wehave H - 1 = H* and H ís unitary. tion (see Problern A-S), we I
HOUSEHOLDER TRANSFORMAnON 545
agonal element is zero. with x'x = 1, where x is a column vector and x' is its transpose. Clearly H' =
encounter sorne com II - 2xx')' = 1- 2xx' = H; hence H is symmetric. Because, by using x'x = 1,
LlI in magnitude, then
H'H = (1 - 2xx')(I - 2xx') = 1 - 4x:x' + 4xx'xx' = 1
: errors in representing
rations. Furthermore, we have H-' =H'; hence H is orthogonal (see Appendix E). Because all the
stages. Consequently, singular values of H are 1 (Problem E-9), its condition number, defined as the
,fthis phenomenon, the ratio of the largest singular value and the smallest singular value, is 1. Hence
the use of Householder transformations will not impair the numerical property
e elimination, we first of a problem (see Problem E-lO).
the first colurnn of A, An important property of Householder transformations is that given any
a matrix of the form of two vectors a and b of equal euclidean norms, that is, Ila\l~ =a'a = b'b, there
umn. Before carrying exists a Householder transformation H such that Ha = b. 1ndeed, if we choose
the largest element in x =(a - b)flla - bll z, then we have
¡rst element, and bring
:d the gaussian elimina
:;; 1, and the errors will
Ha = [1 _ 2(a - b)(a - b)'] a =a _ 2(a - b)[(a - b)'a]
Ila - bll~ a'a -a'b - b'a +b'b
numerically than the
:\ude that every matrix Since a'a = b'b by assumption and a'b = b'a for this is a scalar quantity which
;equence of elementary is invariant under transposition, we have
Ide.
4TI1esign may be chosen lo beequal lO -sign (/11 10 reduce lhe roundoff errors. In lhe QR facloriza
complex conjugale lranspose
If a matrix is transformed into the form in (A-6), the rank of the matrix is is no nonzero eIernen
equal to the number of nonzero rows in (A-6). This is a simple way of computing to the next row. If""
the rank of a matrix. 5
where K ~ KIl-¡K,,_
A-3 Row-Searching AIgorithm 6 1 on its diagonal. so i:
[fthe;th rowof Á is a
In our application, it is often required to search the linearly independent rows of on its previous rows.
matrix A in order from top to bottom. By this, we mean that we first check
whether or not the first row of A is nonzero, or equivaIently, linearly independent rb j ¡
by itseIf. We then check whether or not the second row is linearly independent
with bjj = Lis just th
of the [¡rst row. In the kth step, we check whether or not the kth row of A is
The matrix K can
Iinearly independent of its previous k - 1 rows. If a row is Iinearly dependent
11 - 1. This is comp[
onits previous rows, the row is to be eliminated from subsequent consideration.
computing any row o
Furthermore, we may want to find the coefficients of the linear combination of of
the dependent row.
Let the (i, j)th element of the n x m matrix A be denoted by aij' Let a ¡k
be any nonzero eIement in the first row af A. This element will be called the
pivot element ar simply the pivoto Let K¡ be of the form shown in (A-4) with
ei¡ = - aik/alk, i = 2, 3, ... , n. Then the kth coIumn, except the first element, of
K¡A = (a¡)) IS a zero column, where aL
=aij +ei¡a¡j' Let aL be any nonzero
element in the second row of K¡A. Let K z be of the form shown in (A-5)
with eiZ = - aMaL. Then the jth column, except the firsl two elements, of To compute thejth re
K z K1A = (alj) is a zero column, where alj =ai) + eizaL. In this process, if there jth row of K under i1
o O
e21 1 O
SThe computation or the rank of a matrix is a difficult problem. For example, the matrix ----
e 3t : e 32 : 1
AS=\: -; =:
O O O
=:
1
-'1
-\
-1
-1
¡
fjl 1
1
('j2
bj2
1
1
: ..
ejJ
~j: _ _ _
_
O
'
O o o
j[ l1en il is straightiorw
is c1early or rank 5 and nonsingular. However, we have
4 bjj =
2 -1 2 -4 1
1] [2-
Asx ~ As 2- 2 = 2- 4] =2- 4 [1] 1 ~ 2- 4 e
[
2- 3
2- 4
2- 4
2- 4
[
1
b jk =
Thus,"iLn in A n is"very large, there exists a nonzero x such that A.x = 2 1 -'e--+O and A. is nearly
singular. This irtformation cannot be detected from the determinant of A. (which is equal to 1
1
ror all nI, norfrom Its eigenvalues [which are all equa\ to (see Problem 2,23)],
nor from the form
of A.. However, if we compute the singular valuesof A,,(see the singular value decomposition in
Appendix E}, tÍlen the smallest singular value can be shown to behave as 2-'; for large 11 (see
Problem E-13), and the rank degeneracy orAn can therefore be detected. Thus the singular value
decomposition is considered the most reliable methodof compúting the rankof a matrix. 7 Ki cannol be chosen as He
6This algorithm is a simp[¡fied version of the one in Reference S2Z r()WSof A.
ROW-SEARCHING ALGORITHM 547
~rank of the matrix is is no nonzero element in a row, we assign K¡ as a unít matrix and then proceed
7
np1e wayof computing to the next row. If we carry this process to the last row of A, then we have
Kn-¡K n- z -" KzKIA~KA=A (A-7)
O O O
e Z! I O O
'~e
L bj~epk
¡ e1k+ I)kj
bjj ] ~(k+Z)k
ejk
k = j - 1, j - 2, . . . , 1
(A-11 )
p;k-t 1
em 2-23)], fiar from ¡he form
guIar value decomposilion in
have as 2-', for large n (see
ted. Thus ¡he singular value
: ¡he rank of a níalrix: 7 .K; eannOI be ehosen as Householder lransrormalíons beeause lhey will seramble ¡he order br the
T()WS of A.
548 ELEMENTARY TRANSfORMATIONS
We see that bjk is just the inner product of the vector on its right-hand side and
the vector above it as shown in (A-10)8 Hence the coefficients of combination
in (A-8) or, equivalently, thejth row of K can be readily computed by this simple
procedure.
In the application of this algorithm in this text, the information of the entire
K = K II - 1 . . . K 1 is never needed. We need only a few rows of K. Therefore
it is better to store K¡ in F as in (A-9) and write (A-7) as
The last row of A is
to write
(A-12)
1
Whenever a row of K is needed we then use the procedure in (A-lO) and (A-11)
to compute that row of K from F. -2 1
F~A= -4 -1
2 -1
Example 1
1 -2
Find the linearly independent rows of
Note that the ith colu
A~H
-1 4 -2 4
four rows.
-3 8 -6 6
To find the coeffil
1 -4 10 1 by using (A-11), from
L7 1 -2 10 4
r -~
K1A = -4 O 1
1
A=
~:~=: - ~ :~: -~ ~
3 1 O 2 2 ~Al Hence we have
l
2 O O 1
10001
3
6
-1
O
O 6
085
3
Except the first element, the first column of K 1 is the third column of A divided
or
where a¡ is the ith rov
by - 2. We next chaase the (2, 1) element of A\ as the piVOl i:"w.d compute
1 -1 -1 :1~ -2 1
O 1 :):~ 1 O 2 2
KzA 1 = O -1 1 A1 = O O O O O ~Az
O -1 O 1 O -2 O 4 1
O -2 O O O -2 O 4 1
Since the third row of A z is a zt:ro row, we set K 3 = 1 and proceed to the next row.
The pivot element of the fourth row of A z is chosen as shown. We compute Hence we have .
ar
1 -1 -1 2 -2
~e in (A-lO) and (A-ll) 1
-2 1 3 1 O 2 2
FtA= -4 -1 1 iA = O O O O O ~A
2 -1 O 1 O -2 O 4 1
1 -2 O -1 O O O O O
Note that the ith column of F is the ith column of K¡, i = 1,2,3,4. Since the third
row of Á is a zero row, the third row of A is linearly dependent on its previous
two rows. Similarly, the fifth row of A is linearly dependent on its previous
four rows.
To find the coefficients of combination for the third row of A, we compute,
by using (A-ll), from the first three rows of Fas
~]
O
O
-2 1 O
,- - -; 1- ",-- -_:. - - - : ....-1
2 2 '-211:-1} ~1~'
I 11"'- - " ...... I
--- ---------
2 2 f'...A¡ Hence we have
6 3 [-2 -1 O OJA=O
8 5 or -23¡ -az +a 3 =0 (A-"i3)
d column of A divided where a¡ is the ith row of A.
ivot and compute To find the coefficients al cambination for the fifth í"OW 'Ji A.. w" r;i)nCr.)l)f'~
-2 1 1
2 2 -2 1
O O ~Az -4 -1
4 1 2 -1 ;-0-; 1
, ,
4 1 1 -2 :_0_: -1
roceed to t he next row. -1 O r ---f:
:~-=-
-------
own. We compute Hence we have
[1 - 1 O - liJA = O
or a¡ -3 z +03 3 -'a 4 +a 5 =0 (A-14)
550 ELEMENTARY TRANSfORMA TIONS
This algorithm of searching linearly independent rows of A in the order of To find the coefficien
the first row, the second row, and so forth will be referred to as the row-searching
algorithm. In this algorithm, if the ith row is linearly dependent, then the ith
row will not contribute to the linear combination for the jth linear dependent which implies
row for j > i. For example, since a 3 is linearly dependent, its coefficient is zero
in (A-I4); hence as is expressed as a linear combination of its previous linearly
independent rows. In general, in (A-8), if the ith row, with i <j, of A is linearly
dependent, then bji =0. Indeed, if the ith row is linearly dependent, then the and
ith row of A is a zero row, and the K¡ in (A-12) is a unit matrix. Consequently,
all elements below e¡i in the ith column of F are zero. That is, epi = O, for To find the coefficient
p = i + 1, i +2, ... ,j. Hence from (A-ll), we have bji =0. Because of this
property, the b jk computed in this algorithm are unique. Without bji=O, the
coefficients of combination may not be unique. For eX?'>lple, the addition of which im plies
(A-13) and (A-14) yields
t
-1 1JA=O
which does not have b S3 =0, and is a different combination. The property and
=0 if the ith row is dependent is essential in establishing Theorem G-14
b ji
in Appendix G. We see that the b ij a
A remark is in order regarding the numerical stability of the row-searching stable method (see R¡
algorithm. If the pivot element of each row is chosen as the leftmost nonzero formed iuto the forro
element, then this algorithm reduces essential1y to the gaussian elimination easily obtained. I n ~
(without any pivoting). Though the pivot is chosen as the largest element in linearly dependent ro
magnitude of each row, the pivot is no"t necessarily the larges't element in magni be identical to the 01
tude of its column. Hence the row-searching algorithm by choosing the largest this method of searcl
element in magnitude as the pivot is not exactly equivalent to the gaussian searching algorithm, 1
elimination with partial pivoting. Consequently, there is no guarantee that the The row searchinl
row-searching algorithm will be numerical1y stable. for the following two
In the search of linearly independent rows of A, it is essential not to alter the not complicate or ob~
order of rows. The column positions, however, can be arbitrarily altered. can be easily carricd o
Hence we may apply to the columns of A the gaussian elimination with partial problems by hand, or
pivoting or the Householder transformations to transform A into the form fidence in using a digi
a¡¡ O O O O' chosen for ihe conver
element with value +
aZ¡ azz O O O
zero elements as the p
AL¡L z '" ~ a 31 a 32 O O O ~A (A-15 )
is for pedagogical reas
a 41 a4Z a 43 O O use the method discus
a S¡ asz aS3 O O
s of A in the order of To find the coefficients of linear combination for the third row, we solve
o as the row-searching o OJA =0
:pendent, then the ith
: jth linear dependent which implies
, its coefficient is zero _ -a32
If its previous linearly b 32 ----
a22
h i <j, of A is linearly
{ dependent, then the _ -(031 +b 32 a21)
and b 31
atrix. Consequently, all
That is, epi=O, for To find the coefficients of linear combination for the fifth row, we solve
=0. Because of this
Without bji=O, the (A-16)
·..lple, the addition of which implies
lJA=O
important in computer computation and its generalized form will be used in Problems
Section 5-8.
Let A be an n x n matrix. Consider A-1 Use the row searcl
1 :0 O O ...
__ J _ _ _ _ _ _ _ _ _ _ _ _ _ _O X X X X X
O: X X X X X
O: O X X X X
PIA~ I
P ll A= Use the círc!ed elements,
O: O X X X X coefficients of its linear Co
,
A-2 Repeat Problem A
O: O X X X X
The matrix P II is chosen to make the first column of A, except the first two
elements, zeros as shown. This can be achieved by using gaussian elimination
with partial pivoting or a Householder transformation. The inverse of PI has
the same form as PI' The postmultiplication of PIA by P¡I will not operate
on the first column of P¡A; hence the pattern of zeros in the first column of
PI A is preserved in P IAP¡I. Next we find a P 2 so that Verify that b ji = O if the it
O o:, O X X X X
A-4 Show that if Q¡. i =
PIAP¡I= O O
O O: X X X so is Q~Q,"Qm-l'" Qz
,
,
, A-S Let A be an 11 xm n
O O: O O X X X
that
Again the inverse of P 2 has the same form as P 2 and the postmultiplication of
P 2 P¡AP¡1 by Pi l will not operate on the first two columns of P 2 P¡AP¡I.
Hence the pattern ofzeros in the first two columns of P 2 P IAP¡1 is preserved where R is an upper trian
in P 2 P IAP¡I Pi l. Proceeding similarly, we can transform a matrix, by a is the first nI columns of <:
sequence of similarity transformations, into the form holder transformations a
() N.. f8.ctorizatíon 8.::1.0 t~(
elements of JR may be diff¡
x x X X X X X
¡atter, they can be chosen .
X X X X X X X
O O O O x x x
Ü O O O O x x
an
This is called upper Hessenhergjorm and can be obtained by using numerically
stable methods. This form is very useftil in many computer computations.
PROBLEMS 553
A-1 Use the row searching algorith"l to find linearly dependenl rows of
x
x
x
x [~:r &~
s'j
4 1.5
x x Use the circled elements as pivot elements. Use the procedure in (A-9) to (A-12) to find the
x x coefficients of ¡ts linear combination.
3 -1
\, except the first two I :)) 2
g gaussian elimination
The inverse of PI has
I P1
1
will not operate
in the first column of
[ i
20
3
I
6
1
O
3
O -3
3
7
O
-1
Verify that bji =0 if the ith row. with i <jo is linearly dependent.
A-S Let A be an n x m matrix of rank m. Show that there exists a unitary matrix Q such
x x x
that
: postmultiplication of or
,lumns ofP 2 P¡AP¡I.
2 P 1 AP ti is preserved where R is an upper triangular matrix with nonnegalive clements 011 the diagonal and Q¡
.sform a matrix, by a is the first m columns of Q*. This is called the QR !actorizacion of A. (Hinc: Use House
holder transformations and Problem A-4. Note that the only difference between the
QR factorization and the I~t'::n~seholder tr8.nsform.8.~.~,C)'l. ~.~ th!::1) t'1~ sl~~"?.s ~.J ~r,:,; r:!~2.g.~·):n?~
elements of R may be differenl. In the former. the signs must be chosen as positive: in the
latter. they can be chosen as positive or negative to reduce the roundoff errors.)
A-6 Let A be a square matrix and let A = QR be its QR factorization. Show that the
matrix A~RQ has the same set of eigenvalues of A =QR. (Hinc: Show that A and A
(A-17)
are similar. The QR algorichm, which is the most reliable method of computing the eigen
values of A, ís based on this property. See References SI81 and S200.)
d by using numerically
Iter computations.
The sum, product
o at any point) of an
nomials, exponential
entire real lineo
If a function is kn
determinable from an
Proof
If the function is iden
Let D be an open interval in the realline IR and letf(') be a function defined on (to, ( 1 ), then the func!
D; that is, to each point in D, a unique number is assigned to f The function By analytic continua'
fU may be real-valued or complex-valued.
A function fO of a real variable is said to be an element of dass C" on D
if its nth derivative, ¡<"lO, exists and is continuous for aH t in D. ero is the class
of functions having derivatives of aH orders.
A function of a real variable, f('), is said to be analytic on D iff is anelement
of ero and if for each t o in D there exists a positive real number eo such that, for
aH t in (to - eo, t o +eo), f(t) is representable by a Taylor series about the point
to:
554
ANALYTIC FUNCTlONS OF A REAL VARIABLE 555
Theorem 8-1
If a function! is analytic on D and if! is known to be identically zero on an
arbitrarily small nonzero interval in D, then the function! is identically zero on
D.
Proof
:m D if! is an element
mber 80 such that, for
;eries about the point
(8-1 )
)f a complex variable
:hat f has continuous
{Ior-series expansiono
:d as analytic if it has
'al variable, even if a
IOt be analytic. For
Proof
The solution of the sI
e Defüle
x(t 1
Minimum-Energy Control 1 x~
Then the assumption
-i
X=
lO
l1
Theorem C-1
Let u{'o,ld be any control that transfers.(xo, to) to (x 1, t ¡), and let UO be the control
defined in (C-l) that accomplishes thesame transfer; then
lO
IluO(t)112 dt
556
MINIMUM-ENERGY CONTROL 557
Proof
The solution of the state equation E is
(C-2)
Define
x~<ll-I(tl' tO)x(tl)-x(tO)=<ll(t o, t¡)XI-XO
Then the assumptions that u l and UO transfer (X 0, to) to (xl, t¡) imply that
x=
i
"
lO
<ll(t o, r)B(r)ul(r) dr =
lation
which implies that
:tor, and A and B are, ( [ <I>(t o, r)B(r)(ul(r) -uO(r)) dr, W-l(t o, tl)X) =0
Jntinuous functions of
)llable at timeto, then By using (2-94), we can write this equation as
i"
state XI, there exists a
e Xo at time to to Xl at (u] (r) - uO(r), (4)(t o, r)B(r »*W -I(t o, ti }x) dr =o (C-3)
'o
With the use of(C-\), Equation (C-3) becomes
) dr 'l
i'o
(l.I l (r)-uO(r), uO(r» tlr =0 (C-4)
Consider now
+2 [ (u 1 (r)-uO(r},uO(t»dr
= r'¡ Ilul(r) -
Jto
u O(r)112 dr + r"
Jro
IluO(r)W elr
558 MINIMUM-ENERGY CONTROL L
Since
Q.E.D.
where x is the n x 1 st
vector; A, B, and e ¡;
The response of FE Ü
-energy control.
whert:: T is a posiÜve constant, caBed the sampling periodo The discrete times
O, t; 2T, . .. , are caBed sampling instants. The behavior of FE with the piecewise
559
560 CONTROLLABILlTY AFTER THE INTRODUCTION Of SAMPLlNG
o
1"
tjil
o
1
T
1
I
2T
I
1
I
L-J
l.
I
are directly applicab
Let ;,¡(A) denote;
imaginary part and t:
Theorem 0-1
The term in brackets in (0-4) is equal to x(k); the input U(T) is constant in the
(0-4) values of A, then {Xi,
Proof of Theorem D
interval (kT, kT + T) and is equal to u(k); hence (0-4) becomes, after the change
of variable a = kT + T -T, We assume, without I
shown in Table 5-1.
x(k +1)=e ATx(k) + ( f eAada) Bu(k)
B=(f e
At
dT) B~MB (0-7) A
C=C (0-8)
he input is piecewise - A-
I.
e
A
(1
,,1'
e
O
A12T
O
«] l I
anse at the sampling .A==q t-;:-::~
(0-5a) T(II;j-l)
T
(0-5b) where (nij-1)!
T(II¡j- Z l
(0-6)
O
(nij 2)!
Á-·6. A ijT -J.
(0-7) 1)= e - \"l T(lI i j - 3 l
O O (D-10)
(nij-3)!
(0-8)
x x x x O In the proof of T
O x x x O theorem.
O O x x O
Qij= (0-11 )
Theorem D-2
O O O x O If a continuous-tim~
O O O O 1 its discretized state ec
such that
This theorem is i
, by using an input de
II.¡ O O O
be controlled by an il
O Xi 1 O O
O O Xi O O Example
PijÁijQij = (D-12)
Consider the sample
O O O X¡ expansion, we have
O O O O X¡
where Pti~ Qij 1 and has exactly the same form a~ Qij' We see that the lordan (5 + 1)(5 + 1
form of Aij is the same as A ij with A i replaced by Ai' Define
(0-13)
with P Áp-l exactly in the form in Table 5-1 with A¡replaced by Xi- Now we
shall use this equation toestablish the theorem., ' '
First we note that, because of (0-9) and (0-10), the matrix M defined in
(0-7) is of block diagonal form, each blockis a triangularform of order n¡j
with all diagonal elements equal to (l-e-AiT)/A i if A;=I=O or T if A¡ = O. The Figure 0-2
CONTROLLABILITY AFTER THE INTRODUCTlON OF SAMPLING 563
(0-12) Example
Consider the sampled-data system shown in Figure D-2. By partial fractiol1
expansion, we have
(0-13)
3
2
T
s 1
(0-14)
~
8 y
(s + 1) (s + 1 + 2j) (s + \ - 2j) -3 -2 -\ 2 3
llaced by Xi' Now we
: matrix M defined in
ular form of order nij
O or T if Ai = O. The Figure 0-2
564 CONTROLLABILlTY AFTER THE INTRODUCTION OF SAMPLING
Problems
r:I
T
can be computed as E
Hermitian Forms and
-T- 2iT)
u(k) (D-15) Singular Value Decomposition
- T+ 2iT)
I
'¡,j~ ¡
mijxiXj
X2 -][:~: :~:
X Il • • :~'j'l[:~,]
" =x *M ¡x
6. (E-1 )
where M =~(M¡ + Mi). lt is clear that M = M*. Thus every hermitian form
11 equation in (D-S) to be
can be written as x*Mx with M = M*. A matrix M with the property M = M*
lnd 1m [},¡(A)-})A)] f
is ca\led a hermitian matrix.
r the single-output case
In the study of hermitian forms; it is convenient to use the notation of inner
product. .Observe that the hermitia'n matrix M can be considered as a linear
operator that maps the n-dimensi.ona\ Gomplex vector space (1[", C) into itself.
If the inner product of (C', C)is chosenas' .
(x, y)~x*y ( E-3)
565
566 HERMITIAN FORMS AND SINGULAR VALUE DECOMPOSITION
where X and y are any vectors in (iC", q, then the hermitian form can be written Now we show thal
as matrix can be chosen '
following theorem.
x*Mx = (X, Mx) = (M*x, X) = (Mx, X) (E-4 )
where, in the last step, we have used the fact that M* = M. Theorem E-3
The eigenvectors of a f:
Theorem E-1
are orthogonal.
All the eigenvalues of a hermitian matrix M are real.
Proof
Proof
Let e¡ and ej be the eig
Let il be any eigenvalue ofM and let e be an eigenvector ofM associated with il; ),¡ and }'j, respectively
that is, Me = ile. Consider
(e, Me) = (e, ile) = il(e, e) ( E-5)
and
Since (e, Me) is a real number and (e, e) is a positive real number, fram (E-S)
we conclude that il is a real number. Q.E.D. where we have used t
fram (E-7), we obtai
Theorem E-2 (e¡, e) =0.
Two vectors x,. y are said to be orthogonal if and only if (x, y) = O. A vector The procedúre for de
x is said to be no.rmalized if and only if (x, x) ~ IIxl1 2 = 1. It is' clear that every Schmidt oi,thonormalizl
vector x can benormalized by choosing x = (l/llxll)x. A set of basis vectors that (q¡, qj'j :=0 for i i=
{ql, Q2, .. . "qll} issaid to be an orchonormal basis if and only if Fram Theorem E·
conclude thatfor any r.
i 1=} with'respect lo which t
( E-6)
i=} or equivalently, for any
HERMITIAN FORMS AND SINGULAR VALUE DECOMPOSI1'ION 567
1 form can be written Now we show that the basis of a lordan-form representation of a hermitian
matrix can be chosen as an orthonormal basis. This is derived in part from the
following theorem.
(E-4 )
Theorem E-3
The eigenvectors of a hermi tian matrix M corresponding to different eigenval ues
are orthogonaL
Proof
is a diagonal matrix.
Since every eigenvector can be normalized and since eigenvectors of a
hermitian matrix M associated with distinct eigenvalues are orthogonal, then
the eigenvectors associated with different eigenvalues can be made to be ortho
laps (C", C)into'itself normaL We consider now the linearly independent eigenveCtors associated'with
.t give a Jordan-form the same eigenvalue. Let {e 1, ez, ... , em } be a set of linearly independent eigen
eigenvectors of the vectors associated with the same eigenvalue. Now we shall obtain a set of
ere is no generalized orthonormal vectors from the set {el, ez, . .. , e m }. Let
¡ion. Suppose there
.¡I)k - le +0for some
.. Hence there is no
:re is no lordan block
n representation of a
ere exists a nonsingu
,"-1
atrix with eigenvalues
Q.E.D.
um=e m- ¿ (qk,em)qk
k=l
, (x, y) = O. A vector The pracedure for defining q¡ is illustrated in Figure E-l. 1t is called the
1t is clear that every S,chmidt orthonormalization pr0cedure. By direct verification, it can be shown
\. set of basis vectors +
that (q¡, qj) =0 for i j.
nly if 'From Theorem E-3 and the Schmidt orthonormalization procedure we
condude that for any hermitian matrix there exists a set of orthonormal vectors
( E-6) with respectto which the hermitian matrix has adiagonal-form representation;
or equivalently,for any hermitian matrix M, there exists a nonsingular matrix Q
568 HERMITlAN FORMS AND SINGULAR VALUE DECOMPOSITION H
e
, If the rank of H is r. SI
3 A21 », z2 > . . . '''j'
; 2 > O a'ne
u ---
---- - -
the orthonormal eigenv
Q = [ql qz
Then Theorem E-4 imp
ql
~< ql,e 2
where :E 2 = diag{ A¡, },~,
Figure E-1
and
whose columns are orthonormal, such that
which implies
!VI = Q - lMQ g PMP - l
where M is a diagonal matrix and P g Q - l. Let Q = [ q l qz qll].
where 2: = diag {Al' Az,
Because of the orthonormal assumption, we have
Q*H*HQ=[~Z ~J (E-11)
where 1;z = díag{Ai, J,i, ... , An. Using Q = [Q¡ Qz], (E-lI) can be written as
Q!H*HQz =0 ( E-12)
Then (E-l3) becomes RtR¡ = 1 which implies that the columns of R¡ are ortho
normal. Let R z be chosen so that R = [R¡ R z] is unítary. Consider
DI H. ·{fH iS hermítian,
,f the eigenvalues of H. Although L is uniquely determined by H, the unitary matrices R and Q
.are not necessarily unique. Indeed let Al be a multiple eigenvalue of H*H; then
the correspqnding columns ofQ may be chósen as any orthonormal basis for the
( E-9) spacespanned by the eigenvectors of H*H corresponding to Al. Hence Q
570 HERMITlAN FORMS AND SINGULAR VALUE DECOMPOSITlON
is not unique. Once Q is chosen, R¡ can be computed from (E-14). The choice E-9 Show that aH the ~
of R z again may not be unique so long as [R 1 RzJ is unitary. are equal to 1.
The singular value decomposition has found many applications in linear
systems. In Section 6-S, we use it to find an irreducible realization from a E-10 What are the eiger
Hankel matrix. It can also be used to find simplified or approximated models
of systems. See References S141, S161, and S171. The singular value de
composition is also essential in the study of sensitivity and stability margin of
multivariable systems. See References S34 and S194. For computer pro
grams, see Reference S82. El may arise in gaussian e
The elements of matrices in this appendix are permitted to assume real or tion with partial pivoting.
complex numbers. Certainly all results still apply if they are limited to real largest and smallest sinDlI
numbers. For a real matrix M we have M* = M', where the prime denotes speaking, a condition ~u:
computation.
the transpose. A real matrix with M = M' is called a symmetric matrix; a real
matrix with M- 1 = M' is called an orthogonal matrix. With these modifications
E-11 Show that the conl
in nomenclature, all theorems in this appendix apply directly to real matrices.
mm UJ
Al =!
E-3 Show that if the rank of H is r, so are the ranks of H*H and HH* (Hint: Find a P
such that AP = [A O] and rank A = rank Á.)
10 011 : fl O Il
H¡=l~ v"
h2=l~ ~J
(j
O -IJ O
E-5 Find the singular value decompositions of the matrices in Problem E-4.
E-7 In Theorem E-S, we have Q*Q = QQ*= 1... If we write Q = [Q¡ Qz], do we have
Ql*Q¡ = 1,. and Q¡Qf = In? .
E-S Are aH the. eigenvalues ofa unitary matrix (including orthogonal matdx) real?
Show that all the eigenvalues of a unitary ma:trix have magnitudes equalto 1.
PROBLEMS 571
om (E-14). The choice E-9 Show that all the singular values of a unitary matrix (including orthogonal matrix)
.nitary. are equal to L
applications in linear
ible realization from a E-lO What are the eigenvalues and singular values ofthe elementary matrices
. approximated models
['he singular value de
and stabili ty margin of
El =[ ~ ~ ~]
-3 O 1
E 2 = 0.5
[0.8
1 O 0J
1 O
O 1
For eomputer pro
El may arise in gaussian elimination without any pivoting: E 2 may arise in gaussian elimina
tion with partial pivoting. lfthe condition number of a matrix is defined as the ratio ofthe
itted to assume real or
largest and smallest singular values, which matrix has a larger condition number'! Roughly
hey are limited to real speaking, a condition number gives the amplification factor of the relative errors in the
lere the prime denotes computation.
ymmetric matrix; a real
'ith these modificatiuns E-ll Show that the controllability grammian
irectly to real matrices.
W" = I eA'BB*e A', de
hat x*Mx is a real number E-12 Show that Theorem E-S reduces to Theorem E·4 if H is square, helmitian, and posi
with real coefficients, is it tive semidefinite. If H is square and hermitian (without being positive semidefinite),
'=M? what are the differences between Theorem E-S and Theorem EA?
A _
2
[1 -IJ
O 1
-1
1
-1]
-1
O 1
n Problem E-4
be the characteristic
Then we have
In this appendix we shall study the rnatrix equation AM + MB = N, where If Y/kI - A and B have
A, B, M, and N are n x n complex-valued matrices. We observe that all the
n x n complex-valued matrices, with the usual rules of multiplication and addi
tion, form a linear space (Problern 2-10). Let us denote this space by (X, C). is nonsingular. Inde(
The dimension of (X, C) is n 2 • Consider the operator d defined by eigenvalues of ~(B)
If Y/kI- A and B ha'
d(M)~AM +MB ror aH M in X det ~(B) +0. Hence.
It is clear that the operator d rnaps (X, C) into itself and is a linear operator. From (F-1), we ca
The equation AM + MB = N is often caHed a Lyapunov matrix equation.
Theorem F-1
Let.si!: (X, C)->(;c C) b~ th.e open~üor denned by <s4{~~) = fi:J\!R ~- fA!1], f~,:- ~:.F r~,~: .'.
X. Let A¡, for i = 1,2, ... , 1.:s; n, be the distinct eigenvalues of A and let 11 j' for j = 1, The surnmation of the
2, ... , m ~ n, be the distinct eigenvalues of B. Then (A¡ +11) is an eigenvalue of with IX o = 1, yields
d. Conversely, let Y/k> k = 1, 2, . " . , p.:s; n2 , be the distinct eigenvalues of d,
then ror each k,
which, together with (
d (M) = AM + MB = r¡kM
or (r¡kI - A)M = MB (F-1 )
We now show that the matrices I1kI - A and B have at least one eigenvalue in
common. We prove this by contradiction. Let
O= MÓ(B) (F-3)
Since Ó(B) is nonsingular, (F~3) implies M =0. This contradicts the assump
tion that Mi=- O. Hence the matrices 1] kI - A and B have at least one commori
and y be nonzero n x 1 eigenvalue. Now theeigenvalue of1]kI - A is ofthe form 17k - }e¡. Consequently,
for sorne iand forsome j,
or Q.E.D.
".!
Any matrix representation of the operator d is nonsingular if and only if Let s1: (X, C)->(X, 1[
A¡ +fJ-j=/=-O for aH i,j. M in 7... Let A¡, for i =
for j = 1, 2, ... , m ~
Proof eigenvalue of d. Co
values of .sd, then for
Since the linear operator .sd maps an n 2 -dimensional linear space into itself,
it has a matrix representation (Theorem 2-3). A matrix representation can be
easily obtained by writing the n 2 equations AM +MB =C in the form of
Aro = e, where ro is an n 2 x 1 column vector consisting of aH the n2 elements of for sorne i and sorne j
M. The coroHary follows directly from the fact that the determinant of A is the
product of its eigenvalues (Problem 2-22). Q.E. D. Corollary F-2a
Any matrix represenl
Corollary F-1 b
A¡fJ-j=/=-l for aH i,j.
Ir all the eigenvalues of A have negative real parts, then for any N there exists a
unique M that satisfies the matrix equation
Corollary F-2b
A*M +MA=-N (F-4)
lf aH the eigenvalues
Furthermore, the solution M can be expressed as exists a unique M tha
= - N
linear algebraic equ
1=0
a computing center
2. Compute the infinit
This establishes the coroHary. Q.E.D. have magnitudes Ir
infinite series can b,
Although the matrix M can be solved from Cf~5), the formula is not suitable
for computer computation. The solution of AM.+MB = N has been exten~' M,
sively studied. in ihe literature. The reader is referred to References S9, S 18; and
S107. . . . . .' . The computation m
We list in the rollowing the discrete-time version ofTheorem F-l. than a predeterrnim
ON THE MA TRIX EQUAnON 575
Theorem F-2
ngular ir and only ir Let stl: (x, C) -> (x, C) be the operator defined by d(M) = AMB - M for aH
M in X. Let )'i' for i = 1, 2, ... , /.::; n, be the distinct eigenvalues or A and let ¡lj'
for j = 1, 2, ... , m'::; n, be the distinct eigenvalues of B. Then (A¡!lj - 1) is an
eigenvalue of d. Conversely, let ¡lb k = 1, 2, ... ,p '::;n 2 , be the distinct eigen
values of d, then for each k,
near space into itself,
representation can be {:=-S)
~ = e in the form of
f all the n2 elements of for some i and some j.
leterminant of A is the
Q.E.D. Corollary F-2a
Any matrix representation of the operator d is nonsingular if and only if
A¡!lr/= 1 for all i, j.
)r any N there exists a
Corollary F-2b
(F-4)
Ir all the eigenvalues of A have magnitudes less than 1, then for any N there
exists a unique M that satisfies the matrix equation
A*MA-M= -N (F-7)
(F-5)
Furthermore, the solution M can be expressed as
00
M = L (A*)kNAk
k~O
If the eigenvalues of A,
parts. Consequentiy, I
esentation of d(M) =
>a unique M satisfying The procedure used to prove Theorem F -1 can be used to establish Theorem
F-2. Corollary F-2b can be readily verified by direct substitution. We discuss
(F -4). Because all the three methods of solving (F-7) to conclude this appendix:
al in (F-5) converges.
L By equating toe corresponding elF.ment:o. 0f A*M.<I:\ - ¡"lIT ~.Dd·- i"{ WP; r:2 n
obtain a set of linear algebraic equations. Ir A is an n x n matrix, the set
consists of n 2 equations. The number of equations can be reduced by using
the hermitian or symmetry property of M (see Reference S16). The set of
linear algebraic equations can then be solved by using existing subroutines in
a computing center.
2. Compute the infinite power series in (F-8) directly. Ir all the eigenvalues of A
Q.E.D. have magnitudes much less than 1, the series will converge rapidly.. The
infinite series can ·be computed recursively as follows:
formula is not suitable
8 = N has been exten with M o =0
References S9, S18, and
The computation ma) be stopped when IIMk+l-' Mkll first becomes snialler
fheorem F-1. than a predetermined number.
576 ON THE MATRIX EQUATION
3. Define
Pk+1=Pl P1=A
M.k+ 1= PtMkP k + M k
By this process, the convergence in method 2 can be speeded up (see Refer
ence S4).
Problems
F-2 Find the M to meet A*M +MA = - N with A given in Problem F-l and N = 1,
F-3 Solve Problem F-2 by using (F-S). Which method, by solving algebraic equations
F-9 Transform A*M +MA = - N into a discrete-time Lyapunov equation by using the
G-1 Coprimene!
transformation
A-->(Ad - 1)(Ad + 1)-1 Let D(s) be a polyoomi
D(~
G
+MB with
Polynomials and
Polynomial Matrices
In this appendix, we shall study the coprimeness of two polynomials, its exten
part of Corollary F-1 b. sion to the polynomial matrix case and the factorization of a rational matrix
into two coprime polynomial matrices. The coprimeness of po.1ynomials was
studied by the French mathematician E. Bezout in 1764; 'its application to linear
N be expressed as system problems, however, was quite recent. The concept of coprimeness in
the fractions of transfer-function matrices is, roughly speaking, equivalent
to the concepts of controllability and observability in dynamical equations;
hence its importance cannot be overstated. In this appendix, the concepts and
results in the scalar case will be developed in such a way that they can be readily
\ and B given in Problem extended to the matrix case. The material will be presented in a manner suitable
for numerical computation. The row-searching algorithm discussed in Appen
dix A will be constantly used.
'In this appendix, capital letterswithoutboldface are also used lo denote scalars.
577
578 POLYNOMIALS AND POLYNOMIAL MATRICES
5. Go to step 2.
Jen there exist unique At the end, the resulting N(s) is the remainder R(S).2
In Equation (G-l) if R(s) =0, N(s) is said to be divisible (without remainder)
by D(s), and D(s) is afactor, or a divisor, of N(s). lf R(s) is a divisor of D(s) and
a divisor of N(s), then R(s) is called a common divisor or common factor of D(s)
(G-1 ) and N(s). Note that a nonzero constant is a common factor of every nonzero
D(s) and N(s), and is called a trivial common factor bf D(s) and N(s). A non
trivial common factor wíll be a polynomial of degree 1 or higher.
:t deg N(s) ::?:deg D(s).
h Q(s) as the quotient Definitíon G-1
or example, if N(s) = A polynomial R(s) is a greatest cammon divisor (gcd) of D(s) and N(s) if R(s) is a
Ne have common divisor of D(s) and N(s), and is divisible by every common divisor of
D(s) and N(s). Ir a gcd of D(s) and N(s) is a nonzero constant (independent of s),
then D(s) and N(s) are said to be relatively prime or coprime. 1
In other words, two polynomials are coprime if they have only trivial com
mon factors. Ir they have nontrivial common factors, they are not coprime.
The gcd is unique only up to a constant; that is, if R(s) is a gcd, then cR(s),
for any nonzero number e, is also a gcd. Ir we require a gcd to be monic, then
the gcd is unique.
Given two polynomials D(s) and N(s), by a sequence of long divisions, often
~ there are other Q¡(s) caBed the euclidean algorithm, we can write
N(s)=Q¡(s)D(s) +R¡(s) deg R¡ <deg D
(G-3)
(G-2)
R p_2(s)=Qp(s)R p_¡(s) + Rp(s)
ríght-hand-side poly
;sible. Hence we have This process will eventually stop because the degree of R¡(s) decreases at each
step. \"1-/ e claim. that .R p(s) is ;3 gC(~ el .9(.-:;) aD.(~ /''1(.:.'>
From the last equation of (G-3) we see that Rp_¡(s) is divisible, without
Q.E.D. remainder, by Rp(s). The next to the last equation can be written as R p_2(S) =
(1 + QpQp + ¡)Rp(s); hence Rp- 2(S) is also divisible by Rp(s). Proceeding up
ward, it can be shown that R p - 3 , . . . , R¡, D(s), and N(s) are all divisible by
into a programmatic Rp(s). Hence Rp(s) is a common divisor of D(s) and N(s).
Now we claim that for each R¡(s) in (G-3), there exist polynomials X¡(s)
and Y;(s) such that R¡(s) = X ¡(s)D(s) + Y;(s)N(s). ·This is clearly true for R 1(s)
with X ¡ (s) = - Q¡ (s) and Y¡ (s) = 1. The substitution of R ¡ (s) into the s~cond·
D(s) with m> n: equation shows that the claim holds for Rz(s). Proceeding downward, ·the
. .. .
'Ir D" has a very small absolute value, large errors may·arise on a digital eomputer implementation
or this algorithm. Henee thismethod may not be nllmerie~lIy stable. This situation is similar to
the gaussian elimination without any pivoting.
580 POLYNOMIALS AND POLYNOMIAL MATRICES
claim can be verified for every R í , i = 1, 2, ... ,p. Hence there exist polynomials Proof
X(s) and Y(s) such that 1. If D(s) and N(s) an
Ris) = X(s)D(s) + Y(s)N(s) (G-4) Hence the matrix
are not coprime, ti
This equation implies that every common divisor of D(s) and N(s) divides Ris). and the rank of [
Indeed if C(s) is a common factor, that is, D(s) = D(s)C(s) and N(s) = Ñ(s)C(s), has rank 1 at ever~
then we have Rp(s) = [X(s)D(s) + Y(s)Ñ(s)]C(s). Hence Rp(s) is divisible by its rank only at th
every common divisor of D(s) and N(s). Consequently, Ris) is a gcd. AH gcd 2. This follows direci
differ at most by a constant; hence every gcd can be expressed in the form of 3. If D(s) and N(s) ¡
(G-4). This is stated as a theorem. divisor, we obtain
coprime, there exi~
Theorem G-2 = 1. The substitu
Every gcd of the polynomials D(s) and N(s) is expressible in the form
O O O
3i!' can be shown th~1 deg X(s) < deg N(s) and deg Y(s) < deg D(s)' (see Reference. SI25). This
properlY is nOl needed in lhis téxt.
COPRIMENESS Of POLYNOMIALS 581
and let
)(S)J =0
les)
(G-7) S ~
-
: '__ :_:__ :_:__ :::
No N¡ Nz
_::~: :~-_.
Nm
__
O
__ ::__: ,__ :::
O O
_:_.c_.__ :._ }
O O}
:w,
O No NI Nm - I Nm O O O O n
: rows
O 00 O O No NI Nm - I Nm
see Reference S125). This
( G-12)
582 POLYNOMIALS AND POLYNOMIAL MATRICES
l
(G-13)
Corollary G-3
The polynomials D(s) and N(s) are coprime if and only if their Sylvester matrix Note that the matri~
S defined in (G-12) is nonsingular. 11 There are five nonzel
singular. Correspon,
six rows of F, by usir
Whether or not D(s) and N(s) are coprime hinges on the existence of a non
trivial solution in (G-13). Mariy numerically stable meihods and "canned"
subroutines are available in the literature and computing centers to solve this
problem. However, we are interested in only the special solution which yields Using (G-I0), (G-ll),
the smallest degree in A(s). The row searching algorithm discussed in Appendix
A turns out to yield such a solution. This is illustrated by an example
Example 1 and
Consider the polynomials D(s) = - 2s 4 + 2s 3 - S2 - s +1 and N(S)=S3 +2s 2
-2s -1- 3. V;fe forrn thefr resultant~
1 -1 -1 2 -2 O O
O 1 -1 -1 2 -2 O
We note that the j
O O 1 -1 -1 2 -2
first linearly depender
S= ----------------------------- (G-14)
right-hand-side matri
3 -2 2 1 O O O
O O O -0.5 O 1
---------------------------------;..
luations. We see that O O 0.5 -1.5 -0.5 O
leen transformed into
then the only solution
(Ü -1 -1 2 -2
O O
words, there exists no
ence D(s) and N(s) are, O tI,
'"
-1 -1 2 -2 O
singular, a nontrivial O O 1 -1 -1 2 "'.. -=-- 2-'
- ~I
polynomials A(s) of O O 6 -4 4 (~~ O ~S (G-15)
~(s)/A(s), and D(s) and O O (~}D 17 -17 O O
lowing eorollary. O O O O O O O
O O O O O O O
their Sylvester matrix Note that the matrix F is the one defined in Equation (A-9) of Appendix A.
I There are five nonzero rows in S; henee the resultant S has a rank of 5 and is
singular. Corresponding to the first zero row ofS, we can obtain from the first
six rows of F, by using the recursive formula in (A-ll), the equation
he existence of a non
ethods and "eanned" [1.5 0.5 O: -0.5 O 1 0]8 =0
g centers to solve this
solution whieh yields Using (G-I0), (G-ll), and (G-13), we have
discussed in Appendix
>y an example A(s) =S2 -0.5 B(s) = -0.5s -1.5
Nls) B(s) -0.5s -1.5 s +3
and ( G-16)
D(s) A{.S) s2-0.5 -2s 2 +1
2
1 and N(s) =S3 +2s
Hence D(s) ana N(s) are not coprime.
O
O We note that the B(s) and A(s) eomputed from [-B A]S =0 by using the
-2 first linear1y dependent row of S, which corresponds to the first zero row of the
(G-14 )
O right-hand-side matrix of (G-15), are coprime. Suppose not, then there exist
O B(s) andA(s) ofsmaller degrees to meet [-B A]8=0. This implies that a
linear1Y dependenirowofSwill appear before the appearance ofthe zero rows in
O
(G-15). Thisis nat possible. Hence we conclude that the B(s) and A(s) in
1 (G-,16) are cbprime.
:ndix A to searchthe In theSylvester matrix.in (G-12), the coefficients of D(s) and N(s) are arranged
losen as shown. The in the'ascending power of s. Clear1Y' we can also arrange the eoefficients in the
descendingpowef of s as
584 POLYNOMIALS AND POLYNOMIAL MATRICES
~l
DIl o o ... O It is a 2(k + 1) x (n + k
has two rows formed
and N(s). We note tI
~ ~ In
- - - - - - -:
m-l
~ ~ -~-o -~ _~n_ P~II~_I- - !J~1/----~ ~ ~ -~ ~o-J
- - - - - 1- -
right by one column.
Sk is a square matrix;
:: ::::: :: We search now lin
o o o o o Nm Nm- I NI No For convenience of di
(G-17) rows formed from Ni
linearly independent (
Using this Sylvester matrix, Corollary 0-3 still holds. However, if we use Dn ¡. O and the structü
(0-17) to compute A(s) and B(s), care must be taken in determining their degrees. elements aboye Dn in
To see this, we write is linearly independen
[-B A]S=[-B m _ 1 ..• -Bo:AI _1 ... Ao]S=O may not be linearly iI
Sk' we can readily see
In general, B m - I ¡. O. Hence the degree of B(s) is m -1, and consequently, previous rows, then al
the degree of A(s) is n - 1. However, if the degree of the gcd of D(s) and N(s) Hence the total numbe
is 2 or higher, then B(s) and A(s) have a common factor of form Sk. No such tonically as k increase~
problem will arise in using (0-12). Hence we shal1 arrange the coefficients of number of linearly in<
polynomials in ascending order throughout this appendix. how many more blod
Let v be the total
G-2 Reduction of Reducible Rational Functions words, the v N rows in
and al! N rows, not in
Consider a rational function N(s)jD(s) = N(s)D -I(S), where D(s) and N(s) on their previous row~
are polynomials.1f D(s) and N(s) are coprime, the rational fun"ction N(s)/D(s) dent; hence we have
is said to be irreducible. Otherwise, it is said to be reducible. In application, it
is often required to reduce a reducible rational function to an irreducible one.
Clearly the Sylvester matrix in (0-12) can be used for this reduction, as shown
in the example of the previous section. In this section, we shall modify the Since Sk is a 2(k + 1) x
procedure to improve its computability and to lay ground for the extension to A necessary condition
the matrix case. or k ::; n - 1. Hence w
Consider the two polynomials linearly independent 1
Let A(s) and B(s) b
( G-18~
~1
O It is a 2(k + 1) x (n + k + 1) matrix and consists of k + 1 block rows. Each block
has two rows formed from the coefficients, in the ascending power of s, of D(s)
and N(s). We note that each block is the shifting of its previous block to the
.,. DI DO
._--------- right by one column. If k < n -1, Sk has more columns than rows; if k = n -1,
... O O Sk is a square matrix; if k > n - 1, then Sk has more rows than columns.
We search now linearly independent rows of Sk in order from top to bottom.
NI No J For convenience of discussion, the rows formed froin Di are calJed D rows; the
(G-17)
rows formed from Ni are called N rows. First we note that all D rows in Sk are
linearly independent of their previous rows. This follows from the assumption
However, if we use D" f.O and the structure of Sk' For example, if a new block is added to Sk> all
~rmining
their degrees. elements aboye D" in the last column of Sk+l are zeros; hence the new D row
is linearly independent of its previous rows. The new N row, however, may or
. Ao]S=O may not be linearly independent of its previous rows. From the structure of
Sk> we can readily see that once an N row becomes linearly dependent on its
-1, and consequently, previous rows, then all N rows in subsequent blocks wi.ll be linearly dependent.
e gcd of D(s) and N(s) Hence the total number of linearly independent N rows in Sk will increase mono
. of forro Sk. No such tonically as k increases. However, once the number ceases to increase, the total
nge the coefficients of number of linearly independent N rows will remain to be the same no matter
ix. how many more block rows are added to Sk'
Let v be the total number of linear independent N rows in Soo' In other
:;tions words, the v N rows in SV-l are alllinearly independent of their previous rows,
and all N rows, not in the first v block rows of Sk> k;::: v, are linearly dependent
where D(s) and N(s) on their previous rows. Note that all (k + 1) D rows in Skare linearly indepen
loal f~nction !v (s)/D(s) dent; hence we have
'ble. In application, it 2(k+1) fork':::;;v-1
rank S.= { (G-21 )
to an irreducible one. k (k +1) +v for k;:::v
is reduction, as shown
1, we shall modify the
Since Sk is a 2(k + 1) x (n +k + 1) matrix, if k.:::;; v - 1, then Sk has a full row rank.
A necessary condition for Sk to have a ful\ row rank is that 2(k + 1)':::;; n + k + 1
Id for the extension to
or k.:::;; n - 1. Hence we conclude that v':::;; n or, eq uivalently, the total number of
linearly independent N rows in Sk is at most equal to n.
Let A(s) and B(s) be two polynomials defined as
( G-18)
li(5; = Aa +- Al.) -l·' '"¡"·/~/{'·)
( G-19) (G-22)
and B(s)=B o +B 1 s+'" +Bkl
)f N(s) can be smaller Then from the equation
- B(s)D(s)+ A(s)N(s) =0 (G-23)
k + l block rows IfS k has a full row rank, the only solution in (G-24) or, equivalently, (G-23) is
the trivial solutionA(s)=O; B(s) =0. Ask increases fromO,I, 2, ... ,the first
nontrivial solutionwill appear atk = v. Hence the v in (G-21) yields thesmallest
degree amongall A(s) and B(s) which satisfy (G-23). In other words, the smal\est
degree of A(s) to meet (G-23) is equal to the total number of linear independent
( G-20) N rows in Sv.
586 POLYNOMIALS AND POLYNOMIAL MATRICES
We note that if k ;?:n, the resultant Sk has more rows than columns and sol u with e f= O and d(s) is
tions always exist in (G-24). For example, if k =n, then A(s) =D(s) and B(s) = elementary matrices ~
N(s) are solutions of (G-24) and (G-23); if k = n + 1, then A(s) =D(s)(s + e) and inverses are
B(s) = N(s)(s + e), for any real e, are solutions of (G-24). Clearly these solutions
1 O O O
are of no interest to uso
O 100
E¡-l = O O 1 O
Corollary G-4
O O O c- 1
The two polynomials D(s) and N(s) with deg N(s)::s: deg D(s) = n are coprime if O O O O
and only if the square matrix SIl- 1 of order 2n defined in (G-20) is nonsingular
or if and only ifthe total number of linear independent N rows in Sn-1 is equal
to n. III
They are again elem
operares on lÍle row~
This corollary follows directly from Theorem G-4, and its proof is left as an operates on the colum
exercise. From Theorem G-4, we see that the reduction of N(s)/D(s) hinges on and fifth row. E 3 A(s)
the search of the first linearly dependent row in Sk. The row searching algorithm fourth row of A(s); w]
in Appendix A is developed exactly for this purpose. It is illustrated in Example A(s) and d(s) to the S(
1 of Section G-1 and will not be repeated. operations and A(s)E i
The matrixSk has a special structure: Every block row is a shift ofits previous The set of polynor
block row. By using this shifting property, a very efficient method is developed plication is not a poly!
in Reference S140 to search tije first lin~arly dependent row of Sk' To fully tions, then the set beco
utilize the shifting property, theelimination :must be carried out from left to rational functions, the
right and the meihod is generally not numericálly stable. For hand calculation, developed for matrices
the method can definitely be used'in "place bf the ro\\' s~arching itlgorithm. equally applicable. 1
The result, however, may not be in an echelon form in the matrix case. ' field of rational functic
-~ - ---_ .. ------:---_._---~.- .. _.~-------~._ .. _-----------_.~-------,-------------_ .. _-~----_._-.-._----------_.-.~~ .. -- -.
;).:::;;deg D(s) = n. The A matrix A(s) with polynomials as elements is called a polynomial matrix.
ly dependent row of Sk Similar to matrices with elements in IR 01' e, we may introduce the following
elementary operations on A(s):
1. Multiplication of a row 01' column by a nonzero real 01' complex number.
2. Interchange any two rows 01' two columns.
nearly dependent ofits 3. Addition of the product of one row 01' column and a polynomial to another
ent, this N row is the row 01' column.
jing to this dependent These operations can be carried out by using the elementary matrices, for
s)such that N(s)/D(s)= n = 5, of the form
exist A(s) of a degree
1 O O O O 1 O O O O I O O O O
O I O O O O O O O I O 1 O O O
EI = O O 1 O O Ez = O O 1 O O E3 = O O 1 O O
O O O e O O O O 1 O O des) O 1 O
fore the last N row of O O O O 1 O 1 O O O O O O O I
me. Q.E.D.
(G-25)
han columns and solu with e 1=0 and des) is a polynomial. We note that the determinants of these
A(s) =D(s) and B(s) = elementary matrices are nonzero constants and are independent of s. Their
nA(s)=D(s)(s+c) and inverses are
Clearly these solutions I O O
I O O O O O O
O 1 O O O, O 1 O O O
E-I I -
O O 1 O O E; I =E z E-l-
3 O O I O O
O O O c- 1 O O -des) O 1 O
D(s) = n are coprime if O O O O 1 O O O O 1
I (G-20) is nonsingular
(G-26)
¡ rows in Sn -1 is equal
• They are again elementary matrices. The premultiplication. of 1E¡ DIi A(s)
operates Gil the rows 01' A(s), whereas túe postrnuiUpiicalioii 01 Ji<;¡ on A(s)
Id its proof is left as an operates on the columns of A(s). For example, EzA(s) interchanges the second
of N(s)/D(s) hinges on and fifth row. E 3 A(s) adds the product of the second row of A(s) and des) to the
)W searching algorithm
fourth row of A(s); whereas A(s)E 3 adds the product of the fourth column of
illustrated in Example A(s) and des) to the second column of A(s). We cal! EiA(s) elementary row
operations and A(s)E¡ elementary column operations.
is a shift of i ts previous The set of polynomials does not form a field beca use its inverse in multi
t method is developed ' plication is not a polynomial. Ir we extend the set to include all rational func
t row of Sk' To fully . tions, then the set becomes a field. Fo'r matrices with elements in 'thefield of
,nried out from left to rational functions, the concepts of linear independence, rank, and singulafity
For hand calcu¡'ation? ' developed for matrices with elements in the field of real 01' complex nUJ;nbers are
, searching algorithm. equally applicable. Hence,' ifwe c6nsiderp61ynomials as eiementsof the
le matrix case. field of rational functíons; then we may apply the concept of linear dependence
588 POLYNOMIALS AND POLYNOMIAL MATRICES
and rank to polynomial matrices. For example, the determinant of the poly k¡th
nomial matrix coturnn
J
... O al,k l
+2 s-I~J
al,k¡-l
S
[:
... O O O
[ s-1 S +2
O \1 O
is (s + 2)2 - (s - 1)2 = 6s + 3 which is not the zero element in the field of rational
O O O
functions. Hence, the matrix is nonsingular and has a full rank." The non
singularity of the matrix in the field of rational functions does not imply that the O O O
matrix is nonsingular for all s in iC. For example, the matrix has rank 1, rather
than 2 at s = - 0.5.
Conversely, if the determinant of a polynomial matrix, which is a special
lf O O O
case of rational matrices, is equal to the zero element of lR(s), the field of rational The first r rows are
functions, then the polynomial matrix is singular. For example, the polynomial nOnzero element, a¡.ki
of ai,k; must be on th,
matrix
is, k¡ <k 2 < ... <k,.. ~
the degree of a¡,k;, tha
S +2
lf deg ai,k¡=O, then l
[ S2 +3s +2 these properties is sa
of (G-28) without th,
has a determinant of (s + 2)(S2 - 1) - (s - 1)(S2 + 3s + 2) = O. Hence the matrix just a upper right tria
is singular. Consequently, there exist rational functions a¡(s) and az(s) such lf A is a matrix v
that echelon (row)form. JI
row is 1 and all elen
(G-27) words, the leftmost J
column. By reversir
For this example, we may choose a¡(s) = 1 and a 2(s) = -1/(s + 1). obtain a different, bl
Let a(s) be the least common denominator of a ¡ (s) and a 2(s), and let IX ¡ (s) = appear in the remain(
a(sp¡(s) and &2(S) = a(s)cx2(S). Then (G-27) implies Every polynomia
ementary row operat
Triangularization Pn
where IX! (s) and 1X 2(S) are polynomials. Hence we conclude that polynomja 1
veetors are linearly dependent in the field of rational funetions if and only ir ;jtep 1. íVli(s)=Á\s) al
they ean be made dependent by using only polynomials as eoeffieients 4 (see nonzero.
Problem 2-9). Step 2. Ir all e1ement:
Let A(s) be a polynomial matrix with rank r in the field of rational funetions zero, go to step 6;
IR (s). We show that, by using exclusively elementary row operations, A(s) can Step 3. Search the ele
be transformed into the form and bring it to the I
element monie. Ce
S.tep 4. Compute j
Add the produet
M¡(s), i=2, 3, ...
4In DefinifiQn 2-2 ira field is replaced by a ring ~; (see Footnote 3 or chapter 2), then·(E{, 1Ki;l is
called·al1lod4l.e over.the ringo A fl X 1 or I x fl polynomial vector can be considered"as an demenl
~dditional column ope!
5 [[
or therational. vector space (W(s), ~(s)), or an elementar the modllle (IR "[s], ~[s]). A set or poly
ioto a.diagonal matrix. ca
nomíal veciors is Hnearly independentover the field ~(s) ir and only ir the set ís linearly indeperident
In (G-28), we set, ror red\,
over thering ~[s]. See Relerence S34.
POLYNOMIAL MATRICES 589
t_28l
a3,kr G3,k,.+ 1
O
"'1
ltrix has rank 1, rather
Triangularization Procedure
lude that polynomial
nctions if and only if Step L M(s) = A(s) and oeieie lts cotUülllS ll'Oll1 leíl unlil Lfle úrsl COIUfIlll ¡S
ls as coefficients4 (see nonzero.
Step 2. If all elements, except the first element, of the first column of M(s) are
i of rational functions zero, go to step 6; otherwise, go to step 3.
'" operations, A(s) can Step 3. Search the element with the smallest degree in the first column of M(s)
and bring it to the (1, 1) position by the interchange of two rowS. Make the
element monic. Call the resulting matrix M I(S) = (mb(s))
Step 4. Compute m¡\(s)=qd(s)mll(s)+mh(s) with degmh(s)<degmll(s).
. Add the product of the first row of M 1 (s) and -Q¡I(S) to the ith row of
M 1(s), i = 2,3, ... ,n. Callthe resulting matrix Mz(s) = (m¡z¡(s)).
r chapter 2), then (.'r, IR ,) is
le cons'idered as an elemen(
~n[s], !R[s]). A sel of poly
5 Ir addi;io~al column operations are employed, ~'e can always have k¡ = ¡ and lransfonu the matrix
into a diagonal matr.ix caUed the Smith formo The interested reader is referred lo References S34.
e set is linearly independent
In (G-28), we set, ror reducing lhe size of the matrix, k 3 =k , + l.
590 POLYNOMIALS AND POLYNOMIAL MATRICES
Steps 2 to 5 will reduce the degree of the (1, 1) element of M(s) by at least 1;
hence, after a finite number of iterations, we will go to step 6. We repeat the Hence, we have
process for the submatrices of M, and eventually we will transform A(s) into
the form in (G-28). This completes the triangularization of the matrix. ro -s-1:
In order to have the property deg aj.k, < deg ai,k¡, for j = 1, 2, ... , i - 1, we
need sorne additional row operations. Let aj,k¡ = qj(s)ai,k; + a],k; with deg a],k; <
deg ai,k¡. We add the product of the ith row and - qj(s) to the jth row, the
l~ ---;2~~ -~ i-:
resulting matrix will retain the form of (G-28) and has the property deg aj,k¡ < In this example,
deg ai,k;, j = 1, 2, ... , i - 1. Thus we have established the following theorem. stop at step 4. Dual
formed, by a seq uel
Theorem G-5 triangular matrix or I
Since the determi
Every polynomial matrix can be transformed into the Hermite row form in independent of s, so
(G-28) by a sequence of elementary row operations. 1 matrices are called UI~
Definition G-2
Example 1
A square polynomial
We give an example to illustrate this theorem. Consider is nonzero and indep(
A(')~[-~
-1
3dl
S2 +s - 2
S2 +25 -1
l -4 [-1s
-1
,'+,-2]
3s +1
52 +2s -1
~
[1O S3-,'-,+2
O
+s + 1
+S2
5 +1
] Theorem G-6
A square polynomial
nomial matrix.
[1
[~ [~ ~ +11 (G~9)
-s'-s+2 l -,' -s+2 ]
Proof
-J. O s+ 1 ~ 5 +1 ~
O S3 + 52 + 5 + 1 J O
Let M(s) be unimod
[Adj M(s)J/det M(s) i
In the first step, we interchange the first and second row. This is achieved by polynomial matrices.
the multiplicati.on of A(s) by the element8.ry matrix derroted by 1 listed 2.t th" eDr4 nomials. Sinee M(s)
of this paragraph. In the second step, we multiply the first row by - 1, add the The oníy way lor poly:
product of the first row and s to the second row and add the product of the first are both nonzero con:
row and - 1 to the third row. This is achieved by the matrix denoted by 2.
In the proof, we ha
In the third step, we interchange the second and third row. We then add the
unimodular matrix.
product of the second row and - (S2 + 1) to the third row in step 4. In the last
written as a product o
step, we add the product of the second row and s to the first row so .that the Unimodular matr
degree of a12(s) is smaller than that of all(s). The corresponding elementary tions. For every 5 in
~] [~ ~ ~]O [- ~ ~ ~] [~O O~ ~]
5 O The ránk of a poly
1 1 or postmultiplied by
O _(S2 + 1) 1 O 1 -1 O 1 1 modular matrices.
POLYNOMIAL MATRICES 59!
[~
I
( G-30)
t of M(s) by at least 1;
;tep 6. We repeat the Hence, we have
ill transform A(s) into
1 of the matrix. 3s + 1 ] I 1 2 l
r j = 1,2, ... , i - 1, we s2+ s -2 =1 __ ~ __ ~~I_-J (G-31)
+ a},k, with deg aJ.k; < S2 +2s -1 loo
!!l
(s) to the jth row, the
le property deg a¡,k; < In this example, if we are interested in only the triangularization, we may
he following theorem. stop at step 4. Dual to Theorem G-S, every polynomial matrix can be trans
formed, by a sequence of elementary column operations, into a lower left
triangular matrix or the Hermite column form which is the transpose of (G-28).
Since the determinants of elementary matrices are nonzero constants and
Hermite row form in independent of s, so is the polynomial matrix in (G-30). Such polynomial
I matrices are called unimodular.
Definition G-2
A square polynomial matrix M(s) is called a unimodular matrix ifits determinant
is nonzero and independent of s. ii
2 Theorem G-6
-5 -S+2 ]
s3+ s2+ s +1
A square polynomial matrix is unimodular if and only if its inverse is a poly
s+l nomial matrix.
•[i ~ + 1]
Proof
(G-29) Let M(s) be unimodular. Then det M(s) is a constant. Hence M - 1(.~) =
[Adj M(s)]/det M(s) is c1early a polynomial matrix. Let M(s) and M -l(S) be
polynomial matrices. Then, det M(s) = a(s) and det M -l(S) = bes) are poly
'. This is achieved by
nomials. Since M(s)M - I (s) = 1, we have det M(s)' det M- 1 (s) = a(s)b{s) = J
:d by lUsted at the ene!
'st row by - 1, add the The oniy way for polynomials a(s) and bes) to meet a(s)b(s) = 1 is that a(s) and b(s)
are both nonzero constants. Hence, M(s) and M - l(S) are unimodular. Q.E. D.
the product of the first
matrix denoted by 2.
In the proof, we have shown that the inverse of a unimodular matrix is also a
ow. We then add the
unimodular matrix. It can be shown that every unimodular matrix can be
v in step 4. In the last
written as a product of elementary matrices in (G-2S).
e first row so that the Unimodular matrices are c1early nonsingular in the field of rational func
'espónding elementary tions. For every s in e, they are also rionsingular in the field of complex num
bers. In general, a nonsingular polynomial matrix is nonsingular in the field
of complex numbers for almost aIl S in iC. It becomes singular only at those s
2 1 which are the roots of its determinant. . .
,! ~][! HJ
O -s-1
R(s) = X(s)O(s) + Y(s)N(s)
where X(s) and Y(s) are p x p and p x q polynomial matrices, respectively.
Proof
l O
-( -. -1
-sy+~-+
We form the composite polynomial matrix [O'(s) N'(s)]', where the prime de
notes the transpose. Then Theorem G-S implies that there exists aunimodular
matrix U(s) such that
which is not a unimód
p[U 11 (s) U 12(S)][D(S)] = [R(S)]}P (G-33)
q U Z1 (s) Uds)_ N(s) O}q Let W(s) be any P.
N(s). Then W(s)R(s)
U(s) premultiplication of t
COPRIMENESS Of POLYNOMIAL MATRICES 593
J'l º j
'ices, respectively.
l 9_ ----:=-_1_ - - - - ~ --- -- ! --- IL~~:~] = [~~~)] =
;0
1
-s-1
S2 +s + 1
Hence a gcrd of D(s) and N(s) is
:
: - (S2
,
s
+ 1)
'1
__
O
L
X±- L
O
the gcrd of 0(.'1) and N(s) is not unique. For example, the polynomial matrix
W(S)=[i .'1:
1
n has rank p (in the
2. There exist polyr
is unimodular for any positive integer k; hence such that
R¡(S)=W(S)R(S)=[l+1
. Sk
lIJ[OI 2J=[l+1
.'1+1 i
k
2s k+S+3J
2s +s+l
This is cal1ed the 1
is also a gcrd of the 0(.'1) and N(s) in the example. We see that the degrees of 3. There exist no po
the elements of R¡(s) may be larger than those ofO(s) and N(s). This phenome such that B(s)D(s)
non can never arise in the scalar case.
Let R¡(s) and R 2 (s) be two different gcrds of 0(.'1) and N(s). Can they always -B(
be related by a unimodular matrix? The answer is affirmative if the matrix
[O'(s) N'(s)]' is of full column rank. and
Corollary G -7
Ir [0'(.'1) N'(s)]' is of ful1 column rank, in particular, if 0(.'1) is nonsingular, Proof
then aH gcrds of 0(.'1) and N(s) are nonsingular and are related by unimodular
matrices. l. For convenience \\
Proof
From (G-33), if [0'(.'1) N'(s)]' is of full column rank, so is [R'(s) O')'. Hence,
R(s) is nonsingular. Let R¡(s) be any gcrd ofO(s) and N(s). Then by definitlon, For every s in e, t
we have two polynomial matrices W ¡(s) and W 2(.'1) such that complex numbers.
same as the rank ,
R(s) = W ¡(s)R¡(s) R¡(s) = W 2(s)R(s). unimodular and h.
rank p for every s i
which imply is a polynomial o
following the nonsi
R(s) = W ¡(s)W2(s)R(s) (G-37)
such that det R(s)
Since R(s) is nonsingular, we have W ¡(s)W 2(.'1) = l. Hence, both W ¡(s) and Hence, weconclud.
W 2(s) are unimodular matrices. Consequently, RJ(s)=Wz(s)R(s) i5 also of [D'(s) N'(s)]' i5
nonsingular. Q.E.u. uJ. tút ·úrsL Slaltll1(
condition is met at
In application, we often have the condition that 0(.'1) is nonsingular. With check the rank onl
this condition, the condition in Corollary G-7 is always mel. In this case, the 2. To show the secon
gcrd of 0(.'1) and N(s) is unique in the sense that all gcrds can be obtained from
a single gcrd by premultiplying unimodular matrices.
Ir 0(.'1) and N(s) an
Theorem G-8 nomial matr.ix. T
Let O(s) and N(s) be p x p and q x p potynomial matrices, and let 0(.'1) be yields (G-38) with :
nonsingular. Then 0(.'1) and N(s) are right coptime if and only if any one of the Conversely, we
'following conditions holds: and N(s). Lét R(s
N(s) = N(s)R(s). T
1. For every s in e, or fofevery root of·thedeterminant of 0(.'1), the (p +q)x p
mátrix
COPRIMENESS OF POLYNOMIAL MATRICES 595
[1-V 21 (S)V¡/(s)
O][V
1
11 (s)
V 21 (S)
VdS)]
Vds)
ment and conclude
In the following, w
= [V
OII(S) VI2 (S) ] (G-43) left coprime. Given t\
V 22(S) - V21(S)V ¡/(s)V ds) of rows, then, similar
formations so that
we can write the determinant of V(s) defined in (G-34) as
det V(s) = det V II(s)det [V ds) - V 21(S)V ¡/(s)V ds)] "1=0 (G-44)
q[Q(s
for all s in C. Taking the inverse of (G-43) yields q
Based on this, we have
V 11 (S) VdS)]-I[ 1 0J-l [V 11 (S) ~dS)]-1 (G-45)
[ V 21 (S) Vds) -V 21 (S)V¡/(s) 1 = O l..}. Theorem G-S'
where t.=V 22 (S)-V 21 (S)V¡II(S)Vds) is nonsingular following (G-44). Let A(s) and 13(s) be q
The inverse of a triangular matrix is again triangular and can be readily singular. Then A(s) an
computed. After computing the inverse of the right-hand-side rnatrix of ing condition.s holds:
(G-45), we [hen move the second inverse on the left-hand side of the equaiíty 1. For every s in e, or
in (G-45) to the right-hand side: matrix
where X denotes elements which are not needed in the rollowing. The 3. There exists no poi:
comparison of this equation with (G-34)yidds such that
matrix
det Ves) = det V 11 (s) (G-46)
det Uds)
Hence, D(s) and N(s)
Since Ves) is unimodular, we have deg det Ves) =0. Hence, (G-46) implies
deg det V 11 (s) = deg det Uds) (G-47)
This relation holds for any unimodu.lar matrix U(s) and its inverse Ves) =
= - U 21 (S) and A(s) = U- 1(s). We shal1 now use (G-47) and (G-42) to establish statement 3 of
ing inequality: Theorem G-S. Ir D(s) and N(s) are not right coprime, there exists a gcrd
R(s) with deg det R(s»O. From (G-42), we have det D(s)=det V 11 (S)
D(s) x det R(s), which implies
deg det D(s) > deg det V 11(S)
(G-42)
Because of (G-47) and A(s) = - Uds), we conclude deg det A(s) < deg det
Jsing the identity D(s).
Conversely, if deg det A(s) < deg det D(s), we may reverse the aboye argu
ment and conclude that D(s) and N(s) are not right coprime. Q.E.D.
In the fol1owing, we develop a dual ofTheorem G-S for matrices which are
(G-43) left coprime. Given two polynomial matrices A(s) and B(s) ofthe same number
of rows, then, similar to (G-33), there exists a sequence of elementary trans
formations so that
as
q p .
,-.J'--.. ,-.J'--..
tr fol1owing (G-44). Let A(s) and B(s) be q x q and q x p polynomial matrices and let A(s) be non
singular. Then A(s) and B(s) are left coprime if and only if any ane of the fol1ow
r and can be readily
-hand-side matrix of ing condítions holds:
Id side oftheequality 1. For every s in ic, or for every root of the determinant of A(s), the q x (q + p)
matrix
1
- V2l(s)V¡/(s)
0J
1 [A(s) B(s)] (G-48)
A(s)X(s) +B(s)V(s) = 1
the fol1owing. The 3. There exists no polyriomial matrices N(s) and D(s) of order q x parid p x p
such that
A(s)N(s) = B(s)D(s)
598 POLYNOMIALS AND POLYNOMIAL MATRICES
and deg det D(s) < deg det A(s). G-S Column- al
We give the fo\lowing coro\lary to conclude this section. Consider two polyno
singular, then the mat
Corollary G-S given a q x p rational
Let D(s) and N(s) be p x p and q x p polynomial matrices and let D(s) be non
singular. Let U(s) be a unimodular matrix such that
or
U(S)[D(S)]~[U11(S) U dS)][D(S)] = [R(S)]
N(s) - U 21 (S) Uds) N(s) O where N(s), D(s), A(s:
Then we have polynomial matrices.
~
~ d~
1. U 22 (S) and U 21 (S) are left coprime.
r'
13
d23 J
.
Since U(s) is unimodular, it has rank p +q for every s in C. This implies that
for every s in e, its submatrix [U 21 (S) U 22 (S)], aq x (p + q) polynomial matrix,
has rank q. Hence U 22 (S) and U 21 (S) are, fo\lowingTheorem G-S', left coprime.
We show the nonsingularity of U 22 (S) by contradiction. Suppose U 22 (S) is
not nonsingular, then there exists a 1 x q polynomial vector a(s), not identically
zero, such that
a(s)Uds)=0
which, together with U 21 (S)D(s) + Uds)N(s) =0, implies
a(s)U 21 (S)D(s) =0 where nij and d ij are r
um¡-
l . . 1 of fe.
11 "0
l-:<LI-.
1
'-' l ~"
:uT\'s'J"I ah.
(;(S). These fractions .
a(s)U 21 (S) =0 not right coprime and
A rational functiol
Hence we have a(s)[U 21 (S) Uds)] =0. This contradicts with the fact that 3-5). In terms of th{
[U 21 U 22 ] has rank q in the field of rational functions. Hence we conclude determined. For exa
that U 22 (s) is nonsingular. Consequently, from U 21 D + U 22 N =0, we have the degree of the nun
to that of its denomil
situation is more com
Part 3 ofthis theorem has been essentia\ly est~blished in the proof ofTheorem Given a polynomia
G-S. Indeed, if D(s) and N(s) are right ¿oprime; then R(s) in (G-42) is unimodular power of s in a\l entrie
and deg detD(s) =deg det V 11(S). This, together with (G-47), implies .deg det bciM(~
D(s) = deg det U~2(S), Conversely; if deg det D(s) = deg det Uds), we. may
briM(~
reverse the above argument to conclude degdet R(s) = O. SinceRes) is non
COLUMN AND ROW-REDUCED POLYNOMIAL MATRICES 599
singuar and deg det R(s) =0, R(s) must be unimodular. Hence D(s) and N(s)
are right coprime. Q.E. D.
N(S)] =0 (G-50)
D(s)
n1Z
l~
nlZ
d ll d 12 l3 =~el -
d ez
~e3 =[~ll J
""1 [" ""1
o). n1Z
d ñ
13
D(s) =deg det Uds). nZ1 n22 n23 nZ1 n22 n23 n21 nZZ nZ3
- - - -
d Z1 d22 d Z3 del d ez d e3
O
iC. This implies that
q) polynomial matrix,
rem G-8', left coprime. l"
x O
O
d ez
O
0l'
O .
dd
(.G-53)
lññd"1 d"1
ñ 22 n Z3 O
O
d,·z
ll
ñ Z1 ñ 22
12
n23
Z1 -
d,.z d,.z d,·z
where nij and d ij are polynomials, dei is the least common denominator of the
ith column of (;(s), and el,., is the least common denominator of the ¡th row of
=0 implies C(s). These fractions are easy to carry out; however, N(s) and D(s) are generally
not right coprime and A(s) and B(s) are generally not left coprime.
A rational function C(s) is called strictly proper if C((0) < 00 (see Definition
icts with the fact that 3-5). In terms of the elements of C(s), the properness of C(s) can be easily
. Hence we conc1ude determined. For example, the rational matrix C(s) is proper if and only if
-UzzN =0, we have the degree of the numerator of every element of C(s) is smaller than or equal
to that of its denominator. In terrns of the fractions in (G-51) or (G-52) the
situation is more complicated. 'We shall study this problem in this section.
1 the proofofTheorem Given a polynomial column or ro.w vector, its degree is definedas the highest
n (G-42) is unimodular power of s in all entries of the vector.. We define
G-47), implies degdet
b~J"l(s) = the degree of the ith column of M(s) .
:g det Uds), we may
o. Since R(s) is non- b,.¡M(s) = the degree of the ith row of M(s)
600 POLYNOMIALS AND POLYNOMIAL MATRICES
Definition G-4
and call b ei column degree and by¡ row degree. For exarnple, for
S3 +2s +1 A nonsingular p x p poIy
M(s)=[S+l
s-l S3
(G-54)
Theorem G-9
Ir (;(s) is a q x p proper (strictly proper) rationalmatrix and if (;(s) = N(s)O -1(S)
=A -1(s)B(s), then The rnatrix O(s) in (1
0< b el O(s) +bezO(s) = 2
bciN(s) ~bciD(s)
colurnn reduced but not
for i = 1,2, ... , p, and
brjB(s) ~brjA(s)
for j = 1, 2, ... ,q. is colurnn reduced but
rnatrix is always both co]
Proof Let beiM(s) =k ei. Th,
We write N(s) = (;(s)O(s). Let niis) be the ijth elernent of N(s). Then we have
p
where Hc(s) = diag{lei, i'
nij(s) = L
k= 1
gik(S)dds) i = 1,2, ... , q the column-degree coeiTic
colurnn of M(s) associatel
Note that, for every element in the jth column of N(s), the suÍnrnation is carried rernaining terms and its i
over the jth column of O(s). Ir G(s) is proper, the degree of ni)s), i = 1,2, ... , q, the M(s) in (G-56) can be
is srnal1er than or equal to the highest degree in dkj(s), k = 1. 2, __ . ,p. Hence
we have
M(s)
j = 1, 2, ... , p
[n terrns of (G-57), we ha
The rest of the theorem can be sirnilarly proved. Q.E.D.
det M(s) = (det 1\
We showed in Theorem G-9 that if G(s) = N(s)KY- '(s) is strictly proper..
then column degrees of N(s) are smaller than the corresponding column degrees nence, we conclucie chat
of D(s). lt is natural to ask whether the converse is also true. In general, the coefficient matrix M',e is I
answer is negative, as can be seen from Similar to (G-57), we
S2
N(s) = [1 2] D(s)= [ s+l (G-55)
where H,.(s) = diag{sk r i , i'
row. MIli_ wil1 be called tt
where bciN(s) < beiO(s), i = 1; 2. However, we have
ficients of {he ¡th row c
N(S)O-l(S) = [-2S1 -1 M1r(s) contains the remail
For example, the matrix
bciM(s)
i~1
l(S)J M(s) -
2
_[3s2 +2s 3 2s + 1J (G-56)
s +s - s
is column redueed but not row reduced (verify!)_ A diagonal polynomial
matrix is always both column and row reduced.
Let bc¡M(s) = k ci - Then the polynomial matrix M(s) can be written as
t of N(s)_ Then we have (G-57)
where Hc(s) = diag{sk ci, i = 1,2, _. _, p}. The constant matrixM hc will be called
., q
the column-degree coefficient matrix; its ith column is the coefficients of the ith
column of M(s) associated with i ci . The polynomia-I matrix M¡c(s) conta-ins the
.he summation is carried
remaining terms and its ith column has a degree smaller than k ci ' For example,
~e of nij(s), i = 1, 2, .. _, q,
the M(s) in (G-56) can be written as
, k = 1. 2, ... , p. Hence
~J Mi.cs)contains the remaining terms and its ith row has adegree smaller than k r ¡·
For ex.ample, the matrix in (G-56) can be written as
M s)
(
=[S2O (j J[3
S2
0J +[ 2s
1 O s- 3
mcept.
602 POL YNOMIALS AND POLYNOMIAL MATRICES
In terms of (G-58), we have that M(s) is row reduced if and only if its row-degree The column degrees a
coefficient matrix M h , is nonsingular.
With the concept, we can now generalize Theorem G-9 to the following.
Theorem G -1 O
Let N(s) and D(s) be q x p and p x p polynomial matrices, and let D(s) be column
reduced. Then the rational function N(s)D- 1(s) is proper (strictly proper) if and is singular. Hencl
and only if exist (Xl_ (X1_ and (X3 SUC
Clearly Nds)H e- 1 (s) and D1e(S)He- I(S) both approach zero as s-+ oo. Hence we
have M(s)U 1
lim (;(s) = N',eD,;;, 1
s-oo
where D he is nonsingular by the column reducedness assumption ofD(s). Now if where the degree of n
6ciN(S) '::;6 ei D(s), N'Ie is a nonzero matrix and (;(s) is proper. Ir 6ciN(S) < 6ciD(S), verified that M 1(s) is Cl
Nhc is a zero matrix and (;(s) is strictly proper. Q.E.D.
From theexample,
The Hermite-form polynomial matrix shown in (G-28) is column reduced. column degree can b,
Since every polynomial matrix can b~ transformed ¡nto the Hermite f0rm b)' ~'.
sequence of elementary row transformations (Theorem G-5), we conclude that nomial matrix can be r
every nonsingular polynomial matrix can be transformed to be column reduced the earlier statement a~
by a sequence of elementary row operations. It turns out that the same can be
achieved by a sequence of columns operations. This will be illustrated by an
Theorem G -11
example.
For every nonsingular
Example U(s) and V(s) such th
reduced.
Consider
+1 S 52 +2s +1 An algorithm is' av
M(s)= 2s - 2 ~2S2 +1 column- or rbw~reducf
[
-s 5s 2 -2s to the matrix case.
COLUMN AND ROW-REDUCED POLYNOM1AL MATRICES 603
Id onlJI if its row-degree The column degrees are 1, 2, and O. The column-degree coefficient matrix is
~)]
We normalize the (;(i associated with the column with the highest column degree
to be 1. In this example, the second column has the highest degree; hence, we
·9. We now show the
choose Cl. 2 = 1, and (;(1 and (;(3 can be computed as Cl. 1 = 3, (;(3 = -2. Now if we
postmultiply the unimodular matrix
(s)]Hc(s)
~ [~ ~] ~[~ ~]
Cl. 1 S 3s
1(s)J Oc (s)
Ud') 1 1
2
Cl. 3 S -2s 2
oas s-oo.
~J~M'(S)
Hence';'e 5s +1
[' +1
M(s)UI(s)= ~~2 -6s +1
-2s
nption ofD(s). Now if where the degree of the second column is reduced by one. It can be readily
er. IU)ci N(s) < bciD(s), verified that MI (s) is column reduced. Iil
Q.E.D.
From the example, we see that by a proper elementary column operation, the
28) is column reduced. column degree can be reduced, whereas the determinantal degree remains
the Hermite form by a unchanged. Hence, by a sequence of eleYrt~Dtary s01~~n~~ ~:J~r~l~:~:J~).8, 2. ;..~<>Jy.
3-5), we conclude that nomial matrix can be reduced to be column reduced. We summarize this with
1to be column reduced the earlier statement as a theorm.
It that the same can be
ill be illustrated by an
Theorem G -11
For every nonsingular polynomial matrix M(s), there exist unimodular matriCes
U(s) and V(s) such that M(s)U(s) and V(s)M(s) are column reduced or row
reduced. •
Theorem G-1 2
Thís theorem is du
of Theorem G-12 is c
Let D(s) and N(s) be p x p and q x p polynomial matrices and let D(s) be non
available in Reference
such that
In the following, w
N(s) = Q(s)D(s) + R(s) Consider p x p polyn<
and R(s)D-I(s) is strictly proper A is a p x p constant m
1, i = 1, 2, ... , p. We
which can be replaced by, if D(s) is column reduced,
i = 1, 2, ... , p
Define
Proof
Consider the rational matrix (;(s) = N(s) D - I (s2. This rational matrix is not and
necessarily proper. If every element gij(s) of G(s) is decomposed as guCs) =
gijsP(S) +%(s), where gijsp(s) is a s~rictly proper rational function and %(s) Corollary G -12
is a polynomial, then we can write G(s) as
Let D(s) = sl- A, and
G(s) = N(s)D-I(s) = Gsp(s) +Q(s) (G-59) exist unique polynomi
where (;sp(s) is a strictly proper rational matrix and Q(s) is a polynomial matrix.
The postmultiplication of D(s) to (G-59) yields and
N(s) = Q(s)D(s) + R(s)
This corollary can
with R(s) =:' Gsp(s)l)(s) or Gsp(s) =R(s)D-1(s)
Q,(s) = N"s" - l + (NIlA
Since R(s) is equal to the difference of two polynomial matrices [R(s) = N(s)
and is left as an exercis
Q(s)D(s)], it must be a polynomial matrix.
To show uniqueness, suppose there are other Q(s) and R(s) such that their colurnn and row '
Let A(s) and B(s) be q x q and q x p polynomial matrices. IfA(s) is nonsingular, Theorem G -1 3
there exist unique q x p polynomial matrices Q(s) and R(s) such that
.Consider a q x p prop
B(s) = A(s)Q(s) + R(s) C(s) = N(s)D - I (s). Th
and A -1(s)R(s) is striCtly proper .. p x p nonsingular poly
which can be replaced by, if A(s) is row reduced, N(s)
This theorem is dual to Theorem G-12, and its proof is omitted. The proof
of Theorem G-12 is constructive in nature, and its procedure can be used to
;es and let O(s) be non
compute Q(s) and R(s). Ir O(s) is colurnn reduced, difTerent procedures are
matrices Q(s) and R(s)
available in References S34, S137, and S236 (see also Problem G-15).
In the following, we discuss two special cases of Theorems G-12 and G-12'.
Consider p x p polynomial matrices D(s) and N(s). Let D(s) = sI - A, where
o..
A is a p x p constant matrix. Clearly D(s) is nonsingular and bciD(s) = ¡D(s) =
1, i = 1, 2, ... ,p. We write N(s) as
p
Define
N,.(A) = N"A" + N"_IA"-1 + + NoI
rational matrix is not and N¡(A)=A"N" +A"-IN"_1 + +IN o
decomposed as 9¡is) =
nal function and q;j(s) Corollary G-12
Let O(s) = sl- A, and let N(s) be an arbitrary polynomial matrix. Then there
(G-59 ) exist unique polynornial matrices Qr(s) and Q¡(s) such that
is a polynomial rnatrix. N(s) = Q,.(s)(sl- A) + N..(A)
and N(s) = (sI - A)Q¡(s) + N¡(A) III
R(s) (G-60)
G-6 Coprime Fractiolls of Proper Rational Matrices
O) implies
). Consider a q x p proper rational matrix (;(s). The fraction (;(s) = N(s)D - I(S)
is caIled a right-coprim€ fracUon jr N(s) and JIJJ(s) are right coorime; G(s) =
~ft-hand
side is a strictiy A -1(s)8(s) a lefl-coprimefraclion if A(s) and 8(s) are ieft coprime. Either one
= Q(s) and R(s) = R(s). will also be called an irreduciblefraclion. Given a (;(s), it is possible to obtain
~orem G-10. Q.E.D. many fractions, sorne are irreducible and sorne are noto However, they are all
related by the following theorem.
Let 0- l(S) = Adj O(s)/det O(s), where Adj stands for the adjoint and det stands D
for ~he ~eterminant ofa matrix. Then, N(s)O- l(S) = Ñ(s)D- 1 (s) and O(S)O-I(S) Then we have
= O(s)O- l(S) = 1 imply
(;(s) = N(s)D- J
N(s) Adj O(s) det D(s) = Ñ(s) Adj D(s) det O(s) = Ñ(s)O- J
and O(s) Adj O(s) det D(s) = D(s) Adj D(s) det O(s) and N(s)D - l (s) is irre<
This procedure, how(
Let R(s) be a gcrd of N(s)and D(s), and let R(s) be a gcrd of Ñ(s) and D(s). Then,
nomial matrix and is r
it is clear that R(s) Adj D(s) det D(s) is a gcrd of the two left-hand-side poly
we also compute Ves)
nomial matrices in (G-61) and R(s) Adj D(s) det D(s) is a gcrd of the right-hand
side polynomial matrices in (G-6l). Because of the equalities in (G-61),
R(s) Adj D(s) det D(s) and R(s) Adj D(s) det D(s) are two different gcrds of N(s)
Adj D(s) det D(s) and D(s) Adj D(s) det D(s). We claim that the polynomial and N(s)D-l(S)=V Z1
matrix D(s) Adj D(s) det D(s) is nonsingular. lndeed, beca use of det D(s)D- l(S)
ríght coprime fraction
=det [D(s) Adj O(s)/det D(s)] = det [D(s) Adj D(s)]/(det D(s))P = 1, we have of R(s) can be avoided
det [D(s) Adj D(s) det D(s)] = [det D(s) det D(s)]P =/=0. Hence Corollary G-7 sarily column reduced
implies the existence of a unimodular matrix W(s) such that In this section, we
R(s) Adj D(s) det D(s) = W(s)R(s) Adj D(s) det D(s) fraction from a right
procedure is similar to
which implies
Consider the q x p
R(s)D- l(S) = W(s)R(s)D - l(S) N(s)D-I(s), where A(s)
and p x p polynomial
or D- l(S) = R - 1 (s)W(s)R(s)D- l(S) (G-62) written as B(s)D(s) = A
Since R(s) is unimodular following the irreducibility assumption of N(s)D - l(S),
the matrix
= N(s)D-I(s)
t 0(.1') (G-61 )
and N(s)D - 1(.1') is irreducible. This is one way to obtain an irreducible fraction.
This procedure, however, rcquires the computation of the inverse of a poly
>fN(s)andD(s). Then, nomial matrix and is rather complicated. Ir in the process of generating (G-33),
'10 left-hand-side poly we also compute Ves) in (G-35), then we have
gcrd ofthe right-hand
: equalities in (G-61), O(S)J = Vis) [R(S)] = [V II(S)J R(s)
I different gcrds of N(s)
[ N(s) O V 21(S)
m that the polynomial and N(S)0-I(S)=V 21 (s)V¡?(s). Since Ves) is unimodular, V21(s)V1II(s) is a
;ause of det O(s)O -I(S) ríght coprime fraction (why?). By this method, the computation of the inverse
let D(s»P = 1, we have of R(s) can be avoided. Note that V 11(.1') obtained in this process is not neces
Hence Corollary G-7 sarily column reduced.
that In this section, we shall introduce a method of obtaining a left-coprime
det D(s) fraction from a right fraction, not necessarily coprime, and vice versa. The
procedure is similar to the scalar case discussed in Section G-2.
Consider the q x p proper rational matrix (;(s). Let G(s) = A -1(s)8(s) =
N(S)O-I(S), where A(s), B(s), N(s), and D(s) are, respectively, q x q, q x p, q x p,
and p xp polynomial matrices. Theequality A-I(s)B(s)=N(s)O-I(s) can be
(G-62)
written as 8(s)0(s) = A(s)N(s) or
Jmption of N(s)O- I(S),
[-B(s) A(S)][D(S)] =0 (G-64)
N(s)
(G-63) If we consider polynomials as elements of the field of real rational functions
lR(s), thcn Equation (G-64) is a homogencous linear algcbraic cC]ualion. Con
(G-62) yields O-lis) = sequently, all 1 x (p + q) vectors x(s) with elements in lR(s) (including poly
-I(S) =T(S)O-l(S) into nomials) satisfying
f(s). The nonsingular i[})(s11
;), and R(s). ;~(s) l_ N(síJ = J
)dular. Consequently,
e proof of the theorem. is a linear space over IR (s), denoted as (\1, !R(s». It is a subspace of (IR [J +q(s), !R(s».
Q.E.D. Following Definition 2-11, we call it the left null space. Its dimension is equal
to (p + q) - rank [0'(.1') N'(s)]' = p + q - P = q (Problem 2-51). In this q
'ns of a proper rational dimensional null space (\1, lR(s», any set of q linearly independent vectors in W
: irreducible fraction is qualifies as a basis (Theorem 2-1). In our study we are however interested in
~enerated from a single only the polynomial solutions of (G-64). We use Wp to denote the polynomial
part of V A set of q vectors in Wp will be called a polynómial basis, or a basis
t irreducible, we may in (W p, IR [s]V if every vector in \1 p can be expressed as a unique combination
I/(s)and D(s). We then
71l is in rael a rrec module over lhe polynomial ring !REs] wilh dimension q. 5ee roolnole 4 arlhis
ehapter and Rererenee 534.
608 POLYNOMIALS AND POLYNOMIAL MATRICES
(G-65)
A(s)=A o +A 1 s + +Amsm Lemma G-1
and B(s) = Bo + BIs + +Bmsm
If (;(s) = N(s)D- 1 (s) i~
'We shail caD ¡he mairix formed trom lDl i and Ni the generafized resultant of D(s)
and N(s). Ir D(s) and N(s) are known, this equation can be used to solve B i
and A¡ and, consequently, B(s) and A(s). Conversely, if A(s) and B(s) are given,
a similar equation can be set up to solve for D(s) and N(s). In Equation (G-66), then the statement is al
there are q rows of unknown [ - B o A o ... - Bm A m ]. In order to have q consider
rows of nontrivial solutions, there must be, roughly speaking, q linearly de
pendent rows in the resultant in (G-66). Since it is desirable to have m, the
degree of A(s) and B(s), as small as possible, we shal1 try to use, roughly speaking,
If we use Sk' then a lin<
Because af Lernrn
only in the N rowS. L
BThe discussion'is briefand Ihe·re~der needs nol be concerned because the subsequenL development (i + 1)th block raw. B
is independent or the discussion. for a. complete discussion, see Rderence S95, where
[D'(s) N'(sl]' is assumedto have afu\l rank. Our problem assumes D(s) to. be nonsingular and
the devel~pment can be simplified slightly.
COPRIME FACTORIZATlONS OF PROPER RATlONAL MATRICES 609
the first q linearly dependent rows in the resultant in (G-66). In order to do so,
ltSo We note that every we define
converse however is not
oprime if and only if the 1st block { Do D] ... Do O O' .. O
-'-0
: it into a homogeneous
The rows formed from Di will be called D rows; those from Ni, N rows. The
Let
matrix Sk has k + 1 block rows; each block row has p D rows and q N rows.
Now it is assumed that the row-searching algorithm has been applied to Sk and
its linearly dependent rows in order from top to bottom have been identified.
(G-65)
Lemma G-1
Ir G(s) = N(s)D- 1 (s) is proper, aH D rows in Sb k =0, 1, ... ,are linearly inde
pendent of their previous rows. !l
:onstant matrices. By
:ient of Si to zero yield,
This lemma will be proved later. This lemma does not require that
N(s)D - 1 (s) be a right-coprime fraction nor D(s) be column reducedo lf
N(s)D -1 (s) is not proper, then the statement is not true in general. For ex
O' O O" O ample, a linearly dependent row will appear in the D rows of Sk formed for
O O O
~I:::~s ~Tl=[_:
O"
~-----------
-1
-1
Do O
No O
O
O =0
[:2 S-:l]
even though we have bc¡N(s) ~bciD(s), where b ci denotes the column degree.
Ir, instead ofSk , we arrange Di and Ni in the descending power of s as
DI ······Do
NI ..... Nd
¡
Do DO-l ...
No N O - 1 •••
Do
No
O
O
. ..
..o O
0l
(G-66) A --------------------------
)0 In Equation (G-66), then the statement is again not valid even if N(s) D - I(S) is proper. For example,
'm]. In order to have q consider
:>eaking, q linearly de
:sirable to have ni, the
. use, roughly speaking,
[1 1][~ .:Tl
.. If we use Sb then ~ linearly dependent row 'w\1i appear in a D row.
.. Because of Lemma G-l, the linearly dependent rows of Sk will appear
orily in the N rows. Let r¡ be the number .of linear1y dependent N rows in the
the subsequenldevi:lopmenl (i + 1)th block r6w. Because of the structure of Sk' we have
see Rererence S95, where
. D(s) 10 be nonsingularafld (G-G9)
610 POLYNOMIALS AND POLYNOMIAL MATRICES
(G-70)
are left eoprime, ane
This implies that as k increases, the total number of linearly independent N formo
rows in Sk will increase monotonically. However, once the number ceases to
increase, no matter how many more block rows are added, the number of Proof
linearly independent N rows will remain the same. Define In order not to be o'
rO+(rl-rO)+(rz-r¡}+'" +(rv-r,._¡}=r,.=q
where x denotes non;
Consider Equation (G-66) with m replaced by v: independent of their r
From (G-74), wc hay·
(G-72)
study the structure of
These A¡ and lB¡ are to 1]<:': obtained by using the row-searching algo,ithm. ,~
other words, they are the q rows of l(in iK§", computed as in (A-7), correspon
ding to the q primary dependent rows of SI"
Theorem G-14
The primary depel
Consider a q x p proper rational matrix (;(s) factored as (;(s) = N(s)D- 1 (s). shown. Correspondi
We form Sk and search its linearly dependent rows by using the row-searching (G-75) assume the for
algarithm. 9 Let [ - B o Aa ,.. -B,. Av] be the q rows of I< in KS v=Sv
SO
corresponding to the q primary de.pendent rows af SI" Then the palynomial
'-[--(~l?.Al-a-?-z -a~f\ :
-B o a~1 a~2 a~3:
"On a digital compllter complltation, thÚ; algorithm ShOllld be replaced by a nllmerically stable a~1 a~z a~3:
method:' SecAppendix A. (fthe fOW searching aigorithm is not empioyed, theresllltis generally
not in the polynomial echelon formo
COPRIME FACTORIZATIONS OF PROPER RATIONAL MATRICES 611
matrices
"
A(s) = L Ais i
and B(s) = L Bis i (G-73)
i=O i=O
(G-70)
are left coprime, and A(s) is in a canonical form called the polynomial echelon
nearly independent N formo
the number ceases to
idded, the number of Proof
me In order not to be overwhelmed by notations, we assume p = q = 3 and
(G-71 )
-D
x
in Sk for k 2: v -1. It
my irreducible realiza
x },o~o
x
al to 1'0+1'1 + ... +r., D
x
~
lly q primary linearly
T if the corresponding X
},' 1
ows. For example, all KS3=~ O (G-74 )
uy linearly dependent --
D
)Iock of S" are not all
~
x
:st block have already
},' 1
x
:ond block of S., there
larly, there are r2 -r1 O
¡v' ProCeeding in this D
~3
rly dependent rows in O ~
O },' ~
=r,,=q O
where x denotes nonzero row and O zero rows. Since all D rows are linearly
independent of their previous rows, they are not written out explicitly in (G-74).
From (G-74), we have ro = O, 1'1 = 1, r2 = 1, r3 = 3 = q, and v = 3. In order to
18 v =0 (G-72)
study the structure of A(s), we write (G-72) as
rrching algorithm. in
as in (A-7), correspon
(G-75)
(f~ ,
.\~~
,
~
aL al -- O O O , O O
~80
1\ . ·12 ,,
a21 ti~2 a~~:
O'
-, -8 1 aL ai2 ..
O , -8 2 aL ai2 O : -B 3 :~Ü O
,
:ed by a numerically slable
a~l a° 32 aO33 ,
I aL a12 O ,, aL a~2 O , O ct~
ioyeci, lhe resu il is generally
(G-76)
612 POLYNOMIALS AND POLYNOMIAL MATRICES
This is in an echelon formo The column positions ofthe three :1) are determined and the v defined in (
by the primary linearly dependent rows. For example, the last row of the
second block of S3 is a primary dependent row; hence the last column of the
second block in (G-76) has the element O). lf (G-76) is obtained by the row We shaIl cal! v the re
searching algorithm, then (G-76) has the foIlowing properties:
v3::; .... then (G- 76'
The it
1. AIl elements (including - BJ on the right-hand side of:D are zeros.
2. AIl elements, except element ce ,
of the columns corresponding to primary Hence we have
linearly dependent rows of S" are zeros.
{set of row degrees o
3. AIl elements ofthe column corresponding to nonprimary linearly dependent
rows of S,. are zeros. Since A(s) is row redl
q
Property 3 is the same as saying that aIl columns which are on the right-hand deg det A(s) = I: v,
side and occupy the same positions in each block as those columns with the j;:;: 1
((:elements are zero columns. For example, in (G-76), the sixth column ofthe
third and fourth blocks are zero columns because they are on the right-hand
where deg det stands
side of the sixth column, with element:))' of the second block. We note that
we are ready to show
the aboye three properties may overlap. For example, the rightmost column
then there exists a po
of (G-76) is a zero column following 1 alone 01' foIlowing 2 alone. Because of
these properties, A(s) becomes
The elements encircled by dotted Hnes will be called pivot ·elements. Their
positions are determined by the elements:I: in (G-76). We note that every row This implies that the
and every column has only one pivot element. Because of Property 1, the than n. This is not p(
degree of a pivot element is larger than the degree of every right-hand-side pletes the proof of thi
element in the same row and is larger than 01' equal to the degree of every left
hand-side element in the same roW. Because of properties 2 and 3, the degree Proof of Lemma G -'
of a pivot element is larger than the degree of every other element in the same
lf the linearly depem}(
column. A polynomial matrix with these properties is said to be in the row we can see that the 1'0\
polynomial echelonform 01' Popov form. 10 It is clear that a polynomial matrix Shl.ce A\{s) is rov! red 1!'
in the echeleon form is column reduced and row reduced. aependen t row ol' §,. si
We discuss now the row degrees of A(s). We note that in each block row of in the column of -B i
S" there are q N rows, and the ith N row, i = 1, 2, ... , q, appears v + 1 times in be larger than the con
S". Define, for i = 1, 2, ... , q, tion that (;(s) is propt
Vi ~ number of linearly independent ith N row in S" (G-78) in the N rows.
They will be caBed the row indices of G(s) = N(s)D-l(s). For the example in Example 1
(G-74), we have VI =3,1.'2=3, and 1.'3=1.· Clearly·the n defined in (G-71) is
. also given by Consider
(G-79) 1
s +l s-
\OThe nu'~erical"malrices [-B o Ao ... -,-B. A •.]and [A o Al ... A,.] oblained by G(s) = 7
[ s- 2
Ihe row searching algorilhm \vi\! always be in lhe echelon formo Hence Ihe corresponding poly
nomial rri~lrix [ - B(s) A(s)) and A(s) are said lo be in ¡lJi': polynomia\ echelon formo S2 -1 s
- - - - ---------- ----
,~~ S~IJ=[S-1
(G-79) 1"
s-1
S2 -1""
The fraetion N(s)D - ~(s) is earried out by taking the least eommon denominator Note that the pivots
of eaeh eolumn of G(s) as the eorresponding diagonal e\ement of D(s). Thís /"2 = 2 = q, we have l'
fraetion happens to be right eoprime. In general, a fraetion obtained by this the arrows shown.
process will not be coprime. the formula in (A-l1
We form S2 and apply the row-searehing algorithm:
1 [~.5 =~.5:
O
-1 1 Note that the solutic
O 2 O 1: ,
- -Ó- - -6 -- í ---i :- -1- - - - - - - - - - - - - - ~ - - - - - - - - - - - - - - - - - -
, ,
8(5) = _ [~.5
O O O O: O 1
and
FtS 2 = O O 1 O: O O
O O O O ;-1 O 0.5 1: A(s) =[-0.5
--------------,---------- J
O O O 1: O O -1 0;--(--------------- -1
O O O O: O 1 0.5 O: O 1 and G(5) =A- 1 (s)B(s
O O O 1 :-1 -1 O O: O O 1 linearly independent
the polynomial eehel
O O O O: O -2 -1 0:-1 O O
:~D O O O 1 O Combining Theo'
O ~~~D O 1 O O eorollary reduces to I
-1 1 {)) O O O
Corollary G-14
O 2 O 0(1)0.
- - - - - - - - ---= ¡ -- -6 --0- --6- -~:ü --ó----- Consider a proper
right eoprime if and
O -1 O ~I~ O O
** linearly independent
-1 (Ü 1 O O O
O 2 O O 1 O In the following \\
--·--------·----~1---6--0---ó--(6·-0
properties of (;(5) anc
O -\ O :T~ O O Hon. First, we note
altered without affee
-1 1 1 O O O
and (G-84) by ehoos
O 2 O O o J
::-=- -(~ O O O 1 O 2, ... ,q} rather than
O ::~ I: O 1 O O lish the main resulto
O - - Ó- - -(j) - - -( - -~ i --- -0
Lemma G-2
O O O 2 ,X: O
o 3 O O :I~ O The m x m polynom
the fie\d of rationa\ fl
O -\ O 1 O O
Ó- - - :l~ ---0- ---O . --0- -. -O
O
O
O
O
O
O
O
O
O
00
O
:I~ O
v,~;
[
TO
.
•
O O O O O· (i~ O O
0- ---O . --ó -- -O ---0- . - ·0--- O
O .~
is offull row rank (i
O O O O O O O O
COPRIME FACTORIZATIONS OF PROPER RATIONAL MATRICES 615
common denominator Note that the pivots are chosen for the convenience of hand calculation. Since
:::lement of D(s). This r2 = 2 = q, we have v = 2. The primary dependent rows of S2 are indicated by
ction obtained by this the arrows shown. Corresponding to these primary dependent rows, we use
the formula in (A-U) to compute
Combining Theorems G-8 and G-14, we have the following corollary. The
corollary reduces to Corollary G-4 for the scalar case.
Corollary G-14
Considero a proper rational matrix C(s). The fraction G(s) = N(s)D-i(s) is
right coprime if and only if deg det D(s) = n, where n is the total number of
línearly independent N rows in S" -1 or Sk for k 2. v - 1. lii1
In the following we show that the row indices defined in (G-78) are invariant
properties of G(s) and are independent ofthe N(s) and D(s) used in the computa
tion. First, we note that the order of the rows of the matrix in (G-76) can be
altered without affecting Equation (G-75). This can also be seen from (G-83)
and (G-84) by choosing Q(s) as an elementary matrix which interchanges the
rov:,¡ positions of ftl(S) and 1B(s). 1-If~nce '.,:VhéÜ is irnportal1t 1~:; thé set : ",';, ; :::";. ':.
2, ... , q} rather than the individual Vi' We need the following lemma to estab
lísh the main resulto
Lemma G-2
The m x m polynomial matrix T(s)=T o +Tis + ... +Tjsj is nonsingular (in
the field of rational functions) if and only if the numerical matrix
O
...
!
T1 Tj
[ T, ...
V.~
To T j - i Tj
.0 : ] k+ 1 block 'Ows (G-85)
O O To Ti ... T j
O -E-
is of full row rank (in the field of complex numbers) for k = 0, 1,2, ....
With this lemma, we are ready to establish the following theorem. form a q-dimensional
G-14 are left coprime:
Theorem G -1 5 nomial basis of the nu1
Let G(s) = A -1(s)B(s) be a left-coprime fraction and A(s) be row reduced. rows of Sk are seard
Then the row degrees of A(s) are intrinsic properties of G(s) and are independent [ - B(s) A(s)] are the
ofthe N(s) and D(s) used in Theorem G-14 in the computation. that 6r ¡B(s) .::s;6,.¡A(s) aj
degrees of [ - B(s) A
Proof Since the row degrees
degrees is unique (The<
Because of (G-81), it is sufficient to show that the row indices defined in (G-78) minimal polynomial bao
are independent of the N(s) and D(s) used in the computation. We recall that A(s) is nol row reduced
every right fraction N(s)D -1(S) of G(s) can be obtained from a single right have deg det Á(s) = de~
coprime fraction Ñ(s)D -1(S) by the relationship
Consequently, I'¡ ='¡, i =0,1,2, ... , where F¡are defiried as in (G-67). Hence
i =0, 1, 2, ... , vare independent of N(s) and D(s) used. in the computation.
I'¡,
Allsolutions y(s) of [A
Now we shall show that the set {Vi, i = 1, 2, ... ,q} is uniquely determinable [f D(s) and N(s) are rig.
of
from {r¡, i = 0, 1,2, ... , v} .We óbserve Ihat /'¡ gives the number !V rows with is a 'polynomial basis o
row index i or smaller. Define r _1 = O. Then . the set is a minimal po
/ =0, 1, ... , v (G-89) coluriln searching algo
COPRIME FACTORIZATIONS OF PROPER RATIONAl MATRICES 617
yields the number of N rows with row index exactly equal to i. For example,
consider (G-74). Since rO-I"-I=O, rl-ro=l, 1"2-rl=0, and r 3 -r2=2,
.al vector we have one N row with row index 1, no N row with row index 2, and two N rows
with row index 3. In other words, the row indices are {1, 3, 3}. This shows that
the row indices are uniquely determined by rj, i =0, 1, ... v. Hence the row
indices are also intrinsic properties of (;(.1') and are independent of the N(s) and
D(s) used in the computation. Q.E. D.
As implied by (G-81), the set of row indices is equal to the set of row degrees.
Hence we conc\ude that the row degrees are also intrinsic properties of (;(.1').
In other words, if (;(.1') is factored as A -1(s)B(s), where A(s) and B(s) are left
.k for every k. coprime and A(s) is row reduced, then the set of row degrees of A(s) is unique.
: if V k does not have a As discussed following Equation (G-64), all solutions x(s) of
lpletes the proa f of the
Q.E.D.
X(s{~~;~J =0
ving theorem. form a q-dimensionalleft null space. The A(s) and B(s) computed in Theorem
G-14 are left coprime; hence the set of the q rows of [ - B(s) A(s)] is a poly
nomial basis of the null space (Problem G-16). Because the linearly dependent
A(s) be row reduced. rows of Sk are searched in order from top to bottom, the row degrees of
:.1') and are independent [ - B(s) A(s)] are the smallest possible. The properness of N(s)D - 1(.1') ensures
Jtation. that b,.¡B(s)'::;b,.¡A(s) and that all pivot elements appear in A(s); thus the row
degrees of [ - B(s) A(s)] are equal to those of A(s) and A(s) is row reduced.
Since the row degrees of A(s) are smallest possible and since the set. of row
degrees is unique (Theorem G-15), the set of q rows of [ - B(s) A(s)] is indeed a
dices definedin (G-78) minima! po!ynomia! basis. We show that if A(s) and B(s) are left coprime but
tation. We recall that A(s) is not row reduced, then [ - B(s) A(s)] is not a minimal basis. Indeed, we
d from a single right have deg det A(s) = deg det A(s) and
~d as in (G-67). Hence
[A(s) B(s)] [ - ~~;J = O
:d in the computation. All solutions y(s) of [A(s) B(s)]y(s)= O forro a p-dimensional right null space.
uniq uely determinable If D(s) and N(s) are right coprime, the set of the p columns of [ - N'(s) D'(s)]'
Iuniber of N rows with is a polynomial basis of thenull space. rr, in addition, D(s) is column reduced,
the set is a minimal polynomial basis. If the solution is obtained by using the
(G-89) column searching algorithm (dual to the row searching algorithm), then tlle
618 POLYNOMIALS AND POLYNOMIAL MATRICES
Theorem G-14'
Consider a q x p proper rational matrix G(s) factored as G(5) = A -1(s)8(s) with
G-2 The polynomials [
A(s) = I~=o A¡Si and B(s) = I~=o B¡Si. We form
if the square matrix S of
Aa Bo O O , :O O1 matrix S" _, of arder Li.
Al B1 Aa Bo : . .. : O O statements.
, .
,, : G-3 Transform the mal
Am _ 1 Bm I : Am z Bm Z : Aa Bo k + 1 block columns (each
T k = Am B", Bm , :, Al B 1 block has q +p columns)
O O
Ám
Am
O
I
B",
O
I
: Az B z r
lo
,: into the Hermite row for
,.
O O O O :A
, m Bm
and search linearly dependent columns in order from Idt to right. Let r¡ be the
number of linearly dependent B columns in the (i + 1)th block, and let jJ. be the
i
r
least integer such that Ji = p. Then D(s) = Di and N(s) = L;=o Nis , L;=o G-4 Find a gcrd of the I
solved from
D(
polynomial echelon form, that is, the degree of every pivot element of D(s) is d. D 4 (s) = D,(s) N 4 ()
larger than the degree of every other element in the same row, larger than the
degree of every lov/er el~D1en.t in ~Jv~ same columt~. ;",;,;:~,¿, hre tne Vau's l j-'~j\S.
degree of every upper element in the same column. We call jJ. the column index in Problem O-S? Are th
of G(s). notes the transpose.)
G·' Apply the row searching algorithm to Skin(G-20) to reduce the following rationál
G-8 ls the M(s) in Probl
functions to irreducible ones:
·-------..- - - - - _ . _ -
PROBLEMS 619
• O
5-1
column 'reduced? Ir not, find unimodular matrices U(s) and V(s) such that M(s)U(s) and
V(s)M(s) are column reduced. .'
=fS2 +.1'
bs+1
o
-s +1
J-' and
Are the results the same? Which right fraction, if there is any, is right coprime? are solutions of the Dior
G-10 Find a right fraction from the left fraetion G-15 Consider
[V 11(.1) V12(S)J[
B(s) A(s) _ -
O(S)J=['J
N(s) O
or V 12(.1) V I I (S)J[ - N(S)J = [IJ
[ A(s) B(s) 0(.1') O
where the ith row of Qk i
compllted recursively as
(Hint: Ir O and N are right coprime, we have Q
and U H and U 22 are left coprime and Vil V 21 = ND-'. Using the dual ofTheorem G-13, whcre D[~O irl>/l. f
there exists a unimodular matrix M such that MV 21 = B and MV 22 = A.) R(s) = N(s) - Q(s)O(s). S
Silow that ji N(.I')lJ}-l(.I') = I\J{s)D ¡(sj are [wo right-coprime í"ractions, lhea ¡he
:':;;-")2 G·"hi Consicier G(s) = ¡.
matrix V(s)=D-'(s)O(s) is unimodular. (Hint: The equation XN+YO=I implies rational fUl1ction solutior
XÑiJ-¡ D +YDD-' O =(XÑ +YD)V = 1 which implies U-' to be a polynomial matrix.)
U 11(.1) Udsf\[O(s)J~[R(S)J
[ U2\(s) udst N(s) O Show that T(s) is unimodu
be any vector in Wp • Sho'
where the leftmost matrix is unimodular and R(s) is a gcrdof 0(.1) and N(s). Given O(s)
PROBLEMS 621
and N(s), show that for any F(s), there exist solutions D/s) and Nc(s) to meet the Diophan
·67), left coprime fractions
tine equation if and only if R(s) is a right divisor of F(s), that is, there exists a polynomia1
matrix F(s) such that F(s) = F(s)R(s). Show also that
3 J-l D~(s) = F(s)U11(s)
2 +25
are solutions of the Diophantine equation. Finally, sho\\' that, for any p x p polynomial
matrix T(s), the matrices
s right coprime? are solu tions of the Diophantine equation (see Reference S 139).
G-15 Consider
N(s)=Q(s)O(s) +R(s)
-lJ
-2 with bc¡R(s) < bciO(s) = /l¡. Let H(s) = diag {S·', SI", ... , Sl'p} and Jet 0(5) = OhcH(S) + D1c(s).
Ir D(s) is column reduced, then D{oc is nonsinguJar. We rewrite the equation as
-1 )
N(s)H-I(S)~ Np(s) + N,.(s) = Q(S)(OhcH(S) + D1c(s))H - I(S) + R(s)W I(S)
=Q(s)D"c(I +Dl~IOlcls)Wl(S)) +R(s)H-1(s)
:tions. Show that there where N p(s) is a polynomial matrix and N,(s) is strictly proper. Let Vi be the row degrees
of Np(s). Define H(s) = diag {s", S''', ... ,s'p} and /l = max {Jl i }, v = max (v,}. We write
where the ilh row of Qk is a zero row if k> Vi. Show that the nonzero rows of Qk can be
compuled recursively as
Qo = N po
k- 1
he dual ofTheorem G-13, where Ol~ O if / > JI. From Q(s), we can compute Q(s) = Q(s)DI;;' 1 and lhen compute
0 22 = A.) R(s) = N(s) - Q(s)O(s). See Reference S137.
)rime fractions, then the G-';<'; COi1siüc( G(SI = h. '(s)J8;(s) = r<J(s)íLv '(S). Lel W (¡eIlOl\;; ¡he leí, ¡luir ~fl<i¡;" \.)\ ,d,
n XN+YD=I imp1ies ralional function solutions of
be a polynomial matrix.)
x(s)[O(S)]=0
N(s)
and let Wp denote the polynomial part of W. A set of polynomial vectors is called a poly
nomial basis of W if every vector in Wp can be expressed as a unique combination of the
basis by usingonly polynomials as coefficients. Show thal the set of lhe rows
of [ - 8(s) A(s)l is a polynomial basis of W if and on\y if A(s) and 8(s) are left coprime.
[Hint: (~) Let [ ....:. B(s) Á(s)J be left coprime. 'If [ - 8(s) A(s)J is a basis, there exists a
( p, and p x p polynomial
po\ynomia\ malrix TM S1Ich that .
[- B(s) Á(s)].=T(s)[ ~ 8(.1') A(s)J
Show that T(s) is unimodular and hence A(s)andB(s) are left coprime.( <:= )Let [ - b(s) ¡¡(s)J
be any veclor in V p. Show that ¡¡(s) ¡lo O, and thenappend (q - 1) vectors in VI' lo it to form
s) and N(s). Given Oís)
622 POLYNOMIALS AND POLYNOMIAL MATRICES
G-17 Let Mp(s) be a polynomial matrix and let Vis) be a unimodular matrix such that
Mp(s)U(s) is column reduced. Show that M; 1 (s) is proper if and only if 6 d V(s).:s;
6 d (l\1 p (s)V(s)), for al! i.
In this appendix we ~
functions. We discw
case.
Definition H-1
A number Je (real or (
function g(s) if Ig(Je)1 =
If a proper ration:
trivial common facto¡
that every root of the
numerator of q(s) is ~
mOL oi the acnomina·
not a pole of
although it is a root o
Consider a proper
tion
In this appendix we shall introduce the concepts of pole and zero for transfer
functions. We discuss first the single-variable case and then the multivariable
case.
Definitíon H-1
A number A (real 01' complex) is said to be a pole of a proper rational transfer
function g(s) if ig(A)1 = oc. It is said to be a zero of g(s) if g(A) = O. ¡¡j
although it is a root of S2 + 3s + 2.
Consider a proper rational function g(s) with the following irreducible realiza
tion
X '7Ax +bu y =cx +eu
If we write g(s) = N(s)/ D(s) and N(s) an.O D(s) are coprime (have no nontrivial
" commoti facto'i), then we have
D(s) =k det (s 1 - A)
623
624 POLES AND ZEROS
with sorne constant k. If D(s) is monic (the coefficient associated with the highest not a pole of 9(s), th,
power of s is equal to 1), then k = 1. We discuss in the following the implica output is to apply e;'
tions of poles and zeros.
Theorem H-2
Theorem H-1 Consider a system w
Consider a single-variable system with proper transfer function 9(S) and an tion {A, b, e, e}. Ir
irreducible realization {A, b, e, e}. Then a number 1, is a pole of 9(S) if and only is not a pole of g(s), t
if there exists an initial state X o such that the zero-input response at the output and the input u = e;·r
of the system is equal to
To prove this the
y(t) = re A' for all t ~O
(sI -A)-l(S
for sorne nonzero constant r.
for any 1, that is not ;
the partial fraction e;
Proof
The zero-input response of the system is given by
(s
y(s) =e(sI -A(I X (O)
The identity in (H-l)
Ir ), is a pole of 9(S), then it is an eigenvalue of A. Let v be an eigenvector (A-I,I) and (sI-A)
of A associated with A; that is, Av = AV. Then it can be verified (Problem 2-32)
that v is an eigenvector of (si - A( 1 associated with eigenvalue (s - ,,1.)-1. Proof of Theorem H
Hence we have
The response of the
Y(s) = c(sl - A)-I V=cv(s - ,,1.)- 1 as derived in (4-25),
ar y(t)=eve AC forallt~O
y(s) = c(:
What remains lo be shown is that the constant r =cv is different from zero. =c(.
The realization {A, c} is observable by assumption; hence the matrix [sI' -A' :
cT has a full rank at every s in C. Consequently, for every nonzero vector, in The substitution of (f
particular, the vector v, we have
y(s) =c(sl- AY
[Al -AJc
vf.O
=e(sI - AY
Since (Al - A)v = O, we must have r = el' f.O; otherwise, it would have violated
the aboye condition. This compíetes the proof of í:he necessity of [he theorem.
To show the converse, we show that if y = re A', t ~ O, then A is a pole of 9(S). y(s) = g(A)(s - A:
Ir y(t) = re A', then we have
or y(t) = g(A )e A'
y(s) = c(sl - A)-I X (O) = r(s - ,,1.)-1
This completes the pr
or det (s~ _ A) e [Adj (sI - A)]x(O) = s ~A In this theorern,
Otherwise, lhe theore
or (s - A)c [Adj (sI - A)Jx(O) = r det (si - A) ing that the impedan
which implies det(Ai - A) = O. HenceA is an eigenvalue of A and, consequently, a positive real functi(
a pole o[g(s). This completes the proóf orthe theorem. Q.E.D. give a physicannterpl
g(s), that is,gv,) = O, t .
This theorem states that if 1, is a poIe of g(s), the mode eAC can be generated identically zero even :
at the output by an initial state without the application of any input. If A is blocked by the systerr
POLES AND ZEROS 625
,ciated with the highest not a pole of 9(S), then this is not possible; the only way to generate e}" at the
following the implica output is to apply e;" at the input.
Theorem H-2
Consider a system with proper transfer function 9(S) and an irreducible realiza
tion {A, b, c, e}. Ir the input u(t) is of the form e"', where }" real or complex,
function 9(S) and an
is not a pole of 9(S), then the output due to the initial state x(O) = - (A - Al) - 1 b
pole of 9(S) if and only
and the input u = e}" is equal to y(t) = g(J,)eJ.' for t ~ O. I:ii
response at the output
To prave this theorem, we need the following identity:
(sl-Ar1(s-A)-1 =(AI-A)-1(S-A)-1 +(sI-A)-1(A-,U)-1 (H-1)
for any A that is not an eigenvalue of A. Note the similarity of this identity to
the partial fraction expansion
1 1 1 1 1
---- =--'--+--'-
(s-a)(s-A) (A.-a) (s-),) (s-a) (a-J,)
The identity in (H-l) can be readily verified by post- and premultiplication of
et v be an eigenvector (A - ),1) and (si - A). (Problem H-1.)
'erified (Problem 2-32)
eigenvalue (s - A) - 1. Proot ot Theorem H-2
The response of the system due to the initial state x(O) and the input u(t) is,
as derived in (4-25),
Y(s) = c(sI - A)-I X (O) +c(sl - A) - 1bu(s) +eu(s)
is different fram zero. =c(sI - A)-1 X(0) +c(sI - A)-lb(s - A)-1 + e(s - ),)-1 ( H-2)
e the matrix [sI' -A' :
ery nonzero vector, in The substitution of (H-l) into (H-2) yields
Definition H-Z
Ir (;(s) is of full rank
A number íl, real or complex, is said to be a pote of a proper rational matrix This implies that for
(;(s) if and only if it is a root of det O(s) = O, where O(s) is the denominator a rank eq ual to min (
matrix of any right- or left-coprime fraction of C(s). 111
Definition H-3
Similar to Theorem H-l, we have the following theorem to characterize the
<;;;onsider a q x p pr
poles of (;(s). G(s) = D¡- 1 (s)N¡(s). j
only if there exists an initial state X o such that the zero-input response al the Example 1
output of the system is egual to
Consider the left eopr
y(t) = reAl
s
o
l
for some nonzero vector r. I
~ s +2
The proof of this theorem is identical to the one of Theorem H-l and will G 1 (s) = O
not be repeated. We note that every pole of C(s) must be a pole of sorne element
of(;(s),and every pole of every element of (;(s) must be a pole of (;(s). This
factfol1ows from the faet that det O(s) is equal to the least eommon denominator This N¡(s) has rank 2 I
ofaIi Iílinors of C(s) (see Definition 6-1and Theorem 6-2).
We shall now extend the concept of zeros to the inultivariable case. First,
we assume that the q x p proper rational matrix (;(s) has a full rank in the field 'If G(5) does not llave a fuI
of rational fundions. By this, we mean that if q'~ p, then p(;(s) = q; if q> p, ;. in C'is atransmission zen
POLES AND ZEROS 627
then p(;(s) = p, where p denotes the rank. If (;(s) does not have a full rank in
the field of rational functions, then there exists a 1 x q rational vector M(s) or a
an irreducible realiza p x 1 rational vector P(s) such that
the output due to the
is identically zero. 11 M(s)(;(s) =0 or (;(s)P(s) =0 ( H-3)
where M(s) and P(s) are polynomial vectors. Since y(s) = G(s)u(s), we have
:s) Hence the number of effective inputs is smaller than p. Hence, if G(s) has a full
rank, there are no redundant or noneffective input and output terminals.
;hip, we may give the Consider (;(s) with the fol1owing coprime fractions:
(;(s) = D¡-1(S)NtCs) = N,.(s)D,:-1(S) (H-4)
If (;(s) is of full rank in the field of rational functions, so are NtCs) and Nr(s)..
roper' rational matrix This implies that for almost every A in ic, the q x p complex matrix NtCA) has
s) is the denominator a rank equal to min (p, q) in the field of complex numbers.
• Definition H-3
~m to characterize the Consider a q x jJ proper rational function C(s) with the coprime fraction
C(s) = D¡-1(S)N¡(s). It is assumed 1 that C(s) and, consequently, N¡(s) have a
full rank (in the field of rational function). Then a number )" real or complex,
is said to be a transmission zero of (;(s) if and only if rank N¡(A.) < min (p, q),
latrix G(.~) ~nd an ir in ic, the field of complex numbers. ~
E~ pole af {~\s) ir and
input response at the Example 1
Consider the left coprime fraction
s
• o +11
l
~ s +2 s O
s+2 =[s+2 0J-1[S
G 1 (s)= O 1 ·0 O s+1
'heorem H-l and wíll s + 1
S2
l poleofsomeelement 7 S
a pole of (;(s). This
:ommon denominator This N¡(s) has rank 2 ror eYery s in iC; hence G ¡(s) has no transmission zero. I
).
:ivariable case. First,
a full rank in the field 'If (;(s) does no! have a full rank, lhen pN(A) < min(p, q) for every A in e. In olher words. every
en pG(s)=q; ir q > p, i.in e is a lransmission zero of (;(s). This is a degenera le case and will not be considered.
628 POLES AND ZEROS
and consequently,
This N ¡(s) has rank 1 at s = Oand s = - 2. Hence Oand - 2 are two transmission
zeras of (;(s). Note that O and - 2 are also poIes of (;2(S). ~ y(r
This property is si
From ExampIe 2, we see that (;(s) may not be weIl defined at its transmission H-2. Note that if 1,
zeras. Hence we cannot use p(;(},) < min(p, q) to define its transmission zera. pN(}.)=p and N(J,) k ,
We note that N¡(s) and Nr(s) in (HA) are both q x p poIynomiaI matrices. If X
is not a poIe of (;(s), then it is cIear that p N¡(X) < min(p, q) if and onIy if p Nr(X) <
min (P, q). In fact, it is shown in Reference S34 by using the Smith-McMillan
form that there exist unimodular matrices V(s) and U(s) such that Case 11: q < P
Case 1: q"2:.p
In thiscase, we have pN¡(s) = p in ~(s) by assumption. Hence we have pN¡(}.) = P
in iC for all 1, in iC except the transmission zeros of C(s). Ir l. is a transmission
zero, then pN¡(A) <p. Consequently, there exists a p x 1 nonzero constant
vector k such that
NLC).)k=O
and consequently,
2 are two transmission
s). I for all t"2:.0 ( H-8)
( H-5) In this case, we have p N¡(s) = q in ~(s) by assumption. ln this case, for any), in ic,
we have pN¡(X)::::;q <p in iC. Consequently, there exists a nonzero k such that
ing N¡(s) or N,(s). We N¡(X"- = O and
os of C(s) are the roots
nant. forall t"2:.0 ( H-9)
;mission zeros of G(s) Since this equation holds for every X in iC, it cannot be used to characterize the
.case. A trans'mission transmission zero of C(s) as in the case of q "2:.p. For this reason, we must con
>ugh elements of C(s) sider this case separately from the case q "2:. p.
'he transmission zeros Ir q < p, then X is a transmission zero of C(s) if and only if there exists a
](5), In spite of these nonzero 1 x q constant vector h such that
;ion zeros of G(s) are
e in the following. hN¡(A) =0 ( H-10)
input u(t) = ke J•t is identically equal to zero. (2) Ir q < p and if }, is a transmission zero of C(s) if and 01
zero but not a pole of G(s), then for the input u(t) = ke.l.t, where k is an arbitrary
p x 1 constant vector, there exists a nonzero 1 x q vector f such that the output
y(t) due to u(t) = ke.l. t and the initial state x(O) = - (A - i,I)- 1 Bk has the property
=[
SI-A B
O C(s) = O
J
[SI-A B
D¡-l(S)Nt(s)
J holds for every k. He
we have or f[C(Ai
Define f 1 =fC(AI - A;
Sl-A
P [ -C G(s) J
B =p(sI-A.)+pG(s) A
and
which can be combiné
where p denotes the rank in the field of rational functions. Ir C(s) is of full
rank, that is, pG(s) = min(p, q), then
sl- 1\
p [ -C :J
= n +min(p, q) (H-14 ) Since [f 1 f] is a nom
the link between Coro
Now if s is replaced by an element in e, the field of complex numbers, say s = A, In the following, ti
then SeA) is a matrix with elements in e and its rank must be computed in e will be removed. The
be skipped in the first
Consider a q x p proper rational matrix G(s) with full rank and with an irredu Consider a q x p pro¡::
cible realization {A, B, C, E}. Ir Ais not a pole of G(s), theri Áisa transmission dimensional irreducib
POLES AND ZEROS 631
This theorem foilows directly from Definition H-3 and Equation (H-13).
lO). We see that this
In establishing (H-13), the nonsingularity of sI - A is implicitly assumed.
Beca use of this trans
Hence the proof is applicable only if }, is not a pole of G(s). Before removing
Jn H -3 are ca\!ed the
this restriction, we shall reestablish Theorem H-S from Corollary H-4. If
q ~p, and if A is a transmission zero, Corollary H-4 states that there exists a
dynamical equations.
nonzero k such that the output due to u(t) = ke At and the initial condition
I(S) and an irreducible
Jatrix x(O)=-(A-AI)-IBk Or (A-AI)x(O)+Bk=O (H-16)
are identically zero. Since y(t)=Cx(t) +Eu(t)= Cx(t) +Eke A', we have
( H-12)
y(O) = Cx{O) + Ek = O ( H-17)
d q x p matrices, then The combination of (H-16) and (H-17) yields
l') are polynomials, we
lunctions. Because of [A~-CA :1 -:(O)] =0
fG(A) =0
or f[C(AI - A)-IB +E] =fC(AI- A)-IB HE=O
Define f l = fC(AI - A) - 1 Then we have
-A) +pG(s)
fl(AI - A) = fC
¡(s) (H-13)
and flB +fE. =0
which can be combined as
ons. lf (;(5) is of fuil
f][AI-A 8J=0 ( H-18)
[fl -C E
(H-14)
Since [f1 fJ is a nonzero vector, (H-18) im plies pS(A) < n +q. This completes
the link between Corollary H-4 and Theorem H-S.
¡ex numbers, say s == A, In the fo\!owing~ the restriction that A is not a pole of G(s) in Theorem H-S
1st be computed in C. will be removed. The proof reHes heavily on the result of Section 6-6 an'd may'
be skippedin the first reading.
Theorem H-6
Ik and with an irredu Consider a q x p proper rational matrix G(s) with a fu\! rank and with an n
Ilen A is a transmission dimensional irreducible realizatíon {A, B, C, E}. Then A is a transmission
632 POLES AND ZEROS
pS(A) = P [
AI-A
-c :J < n +min(p, q)
p[ -C
and the theorem follow
2
Proof The concept of tral
possible ways of definin
Let E = (;( co). We factor (;(s) as
duce one more definitic
(;(s)=E +D- 1 (s)Ñ(s)=D- 1 (s)[D(s)E +Ñ(S)]~D-l(S)N(s) (H-19)
1 Definition H-4
where D- (s)N(s) is strictly proper and left coprime and D(s) is row reduced.
We then apply the procedure in Section 6-6 to find an irreducible {Ao, Bo, Co} Consider a q x p prope
such that, similar to (6-133) and (6-130), divisor (gcd) of the nur
are called the blocking :
Co(sI - A o)- 1 = D -1(s)L(s) (H-20)
and N(s) = L(s)8 0 (H-21) Example 3
where L(s) is defined as in (6-129). Furthermore, we have {D(s), L(s)} left Consider
coprime and {sI -A o, Co} right coprime. Hence, Theorems G-S and G-S'
imply that there exist polynomial matrices X(s), - Ves), X(s), and Ves) such that
X(s)(sI - A o) - Y(s)C o = In
L(s)X(s) + D(s)V(s) = I q
which, together with D(s )C o = L(s )(s 1 - A o), can be written in matrix form as The gcd of the three nur
zero of (;(5)_ 1t can be ~
~(S)J =[I
~(S)
[L(s) Y(S)][SI - A o
D(s) - Co Ves)
n·
O
X(s)X(s) +Y(S)Y(S)J (i-1-22)
Iq
- 2 (Problem H-5).
'This proof \vas provided by Professor L. S. Chang of China University of Sciences and Technology,
Consider a system with
HoFei. form shown in (H~25).
--~-- _ .._._._-_. . ._. . ._--- - - - - - - - - - - - _ . _ - - . _...- - - -
sI -A
p [ -C
BJ
E
= p [SI -Aa
-Ca
BoJ
E =11 +pN(s) (H-24 )
l(p, q)
and the theorem follows from the definition of transmission zero. Q.E.D.
(s +2)(s + 1)
j
S2 +2s +2
ritten in matrix form as The gcd of the three numerators of (;3(S) is s + l. Hence -1 is the only blocking
zero ofG(s).. lt can be shown that the transmission zeros of G(s) are O, -;-1, and
X(s) +Y(S)Y(S)] - 2 (Problem H-S). l!iI
1 (H-22)
q
From this example, we see that the definition of transmission zero and that
lence the left-hand-side
of blocking zero are not equivalent. lf every element of G(s) is irreducible, a
t-hand-side matrices are
blocking zero can never be a pole of G(s). On the other hand, a transmission
zero can also be a poleo If we factor G(s) = N,.(s)D,:-l(S), then every blocking
)8 0 +Y(S)E] zero of (;(s) will appear in every element of N,(s); hence every blocking zero is él
)8 0 + D(s)E transmission zero. The converse is, of course, not necessarily true.
If G(s) is a scalar transfer function, then there is no difference between
)lB\o +- Y(s)ll.!'.l ·~.ransmjsslon. 7,f;"(OS .3.n.et blor:k~.ng ~~r~s.
N(s) J Similar to transmission zeros, blocking zeros also have the property of
blocking the transmission of certain mode. Let u¡{s), i = 1, 2, ... , p, be the ith
)(s)E +L(s)B o in (H-19)
component of the input vector u(s). Let
limodular, we have, [01'
i= 1,2, ... ,p ( H-25)
EJ = n +p N(s) where cP(s) and f¡(s) are polynomials. Let s -lX be a root of cP(s). lf e~t appears
in the output y¡(t), then S-lX is said to appear as a mode of y¡(t). Otherwise, it
educible realizations of is blocked from y¡(t). .
ve, rOl' every s in iC,
Theorem H- 7 .
ty of Sciences and Technology, Consider a system with proper transfer matrix G(s). Letthe input u(s) be of the
form shown in (H-2S). Then, for any initial state, no root of cP(s) will appear as
a mode at any output terminal if and only if every root of <jJ(s) is a blocking zero Problems
of G(s).
H-1 Prove the identity
Proof
H-2 Consider the dyna
Ir every root of <jJ(s) is a blocking zero of G(s), then G(s) contains no root of
<jJ(s) as a pole. Consequently, for any initial state, the zero-input response of
the system will not contain any mode due to the roots of <jJ(s). Let y¡(s) be the
ith component of the output vector Y<s) and let G(s) = (g ij(s». Then the zero Find an initial state so tI"
state response of the system is given by
. p H-3 Consider the egua
y¡(s) = L gij(s)u}s) ( H-26) = é' and ti
y(t) due to ult)
j= 1
Let f3(s) be the gcd of the numerators of all gij(s). Then we can write 9ij(S) as H-4 Prove Theorems 1-:
y¡(s) = ±
j= 1
f3(s) nij(s)h(s)
<jJ(s) d¡}s)
Now if every root of <jJ(s) is a blocking zero, then <jJ(s) will divide f3(s). Hence
no root of <jJ(s) will appear as a mode of y¡(t).
Now suppose the root, s - IX, of <jJ(s) is not a blocking zero. Then there
exists at least one nkl(s) which does not contain 's - d. as a factor; otherwise,
s - ex would be a blocking zero. We choose /;(s) = O for j = 1, 2, ... ,1 - 1,
1+ 1, ... , p andfi(s) = 1, then H-G Let N(s)D-1(s) = t'
set of transmission zeros
Yk(S) = f3(s) nk!(s)
<jJ(s) dkl(S)
H-7 Show thal
and s - IX will appear as a mode of Yk(t). This completes the proof ofthis theorem A-
Q.E.D. rank [ _
Le[ ¡;¡(s) = k/(s - ,t). H), is a lransmission zero oí G(s), lflen e"; Wii! no¡ apr ea ,'
under any state feedback a
as a mode at the output vector only for a certain k. Ir A. is a blocking zero of conclusion from the facts
G(s), then eA! will not appear at the output vector for any k. These concepts of transfer-function matrix a
zeros will be used in the study of tracking control problems in Chapter 9. function matrix?
The transmission zeros of a proper transfer matrix G(s) are defined from a
coprime fraction of G(s) in Definition H-3. We see from Theorem H-6 that H-S What are the blocki
they can also be defined from an irreducible realization of G(s) by using the
system matrix in (H-15). They can again be defined fro.m the Smith-McMillan
form of G(s), see References S34 and S185. F or discussions of v~rious definitioils
of zeros, see References S152 and S186; for their computation, see References
sn and S148. Ir all transmission zeros lie inside the open'left half s-plane,
G(s) is said to be minimum phase. See Reference S67.. The presehtatión of this
appendix follows closely References S78 and S94.
POLES AND ZEROS 635
H-3 Consider the equation in Problem H-2. Find the initial state so that the response
( H-26) y(t) due to u(t) = e31 and the initia\ state is of the form e 31 for t zO.
G1(s)=
.1'-1
O
l
1
.1'+10
2(s-W
.1'+1
(s - 1)2
J .1'+1
.1'+2
(s +2)(s+l~
.1'2 +2.1' +2
l
~ 1rT::~
O 1 s -IJ
.viii divide f3(s). Hence G 3 (s) =
A [
.1'2
.1'-1 O s +1
cing zero. Then there
as a [actor; otherwise,
OJ- 1[1 .1'- lJ
.1'-1 2 .1'2
) for j=l, 2, ... ,/-1,
H-G Let N(s)D - 1(.1') = Ñ(s)D -1 (s) be two right-coprime fractions of (;(.1'). Show that the
set of transmission zeros defined fram N(s) and the one fram Ñ(s) are the same.
-C E -(C+EK)r l
47. Ho, B. L., and R. E. Kalman, "Effeetive construction oflinear state variable 1I-ans. ASME, ser
models from inputfoutput data," Proc. Third Allerton Conf, pp. 449-459, 67. - - , and T. S. En
48. Ho, Y. c., "What constitutes a controllable system," 1RE Trans. Automatic 68. - - , P. L. Falb, ¡;
Control, vol. AC-7, p. 76, 1962. . New York: MeGI.
49. Hsu, C. H., and C. T. Chen, "A proof of stability ofmultivariable feedback 69. ---, Y. C. Ho, al
systems," Proc: IEEE, vol. 56, pp. 2061-2062, 1968. systems," Contrib .
50. Jacob, J. P., and E. Polak, "On the inverse ofthe operator (-) =A(J +8(')," 70. 'Kapian, W.,Opel
Am. Math. Monthly, vol. 73, pp. 388-390, 1966. Addison-Westey,
51. Johnson, C. D., and W. M. Wonham, "A note on the transformation to 71. Kreindler, A, and
• _·••.• __
~_L.'_. __
~ _ . ~ •. ~ _. __ ~ . _ _.,---,--._. ._. - ' - - - - _ ~ • ~_~~~ __ . ~ _ . _ , ~ _
REFERENCES 639
3.riant systems," IEEE 52. Joseph, P. D., and 1. T. Tou, "On linear control theory," AlEE 1J·ans.
3. Applications and Industry, vol. 80, pt. Ir, pp. 193-196,1961.
k linear time-invariant
53. Jury, E. l., Sampled-Data Control Systems. New York: Wiley, 1958.
.1968.
54. Kalman, R. E., "A new approach to linear filtering and prediction prob
onal matrix function,"
lems," Trans. AS ME, ser. O, vol. 82, pp. 35-45, 1960.
55. - - , "On the general theory of control systems," Proc. First Intern.
design ofmultivariable Congr. Autom. Control, Butterworth, London, pp. 481-493, 1960.
)1. AC-12, pp. 651-659, 56. - - , "Contribution to the theory of optimal control," Bol. Soco Mat. Mex.,
vol. 5, pp. 102-119, 1960.
.r control systems with 57. - - , "Canonical structure of linear dynamical systems," Proc. N atl.
utomatic Control, vol. Acad. Sci. U.S., vol. 48, no. 4, pp. 596-600, 1962.
58. - - , "On the stability of linear time-varying systems," 1RE 1J'ans. Circuit
'd Mathematics. New Theory, vol. CT-9, pp. 420-423, 1962.
59. - - , "Lyapunov functions for the problems of Lur'e in automatic control,"
1 and 2. New York: Proc. Natl. Acad. Sci. U.S., vol. 49, pp. 201-205, 1962.
60. - - , "Mathematical description of linear dynamical system," SI AM J.
1mu ltivariable control Control, vol. 1, pp. 152-192,1963.
53.
61. - - , "When is a linear control system optimal?" Trans. ASM E, ser. O,
state feedback," SIAM
vol. 86, pp. 51-60, 1964.
62. - - , "Irreducible realizations and the degree of a rational matrix,"
-red ucible J ordan form SIAM J., vol. 13, pp. 520-544,1965.
.utomatic Control, vol. 63. - - , "Toward a theory of difficulty of computation in optimal control,"
Proc. 4th IBM Sci. Computo Symp., pp. 25-43, 1964.
son-Wesley,1961.
64. - - , "On structural properties of linear constant multivariable systems,"
d ed. Princeton, N.J.: l'eprint of papel' 6A, Thil'd Congl'ess of the lnternational Federation of
Automatic Control, 1966.
.: Blaisdell, 1964. M. - - , and J. E. Bertram, "Control system a.nalysis and design vía the
t in muíti-input con 'second method' ofLyapunov," "li'ans. ,,;S!Vi E, S¡;;I. S', mi. 82, pp. ~./ i .:;~:,
; Control, vol. AC-13, 1960.
66. - - , and R. S. Bucy, "New results in linear filtering and prediction theory,"
1 of linear
state variable Trans. ASME, ser. O, vol. 83, pp. 95-108,1961.
ton Conf, pp. 449-459, 67. - - , and T. S. Engler, A User's Manualfor Automatic Synthesis Programs,
, IRETrans. Automatic 68. - - , P.L. Falb, and M. A. Arbib, Topics in MathematÜ';al System Theory.
New York: McGraw-Hill, 1969.
multivariable feedback 69. - - , Y. e. Ho; and K. S. Narendra, "Co~tiollability of linear dynamical
erator(')=A(') +B(')," 70. Kaplan, W" Operational Methods for Linear Systems. Reading, Mass,:
Addison-Wesley, 1962.
the transformation to 71. Kreindler, A., and P. E. Sarachik, "On the concepts of conlrollability and
640 REFERENCES
observability of linear systems," IEEE Trans. Automatic Control, vol. the 'second me
AC-9, pp. 129-136, 1964. pt. 4, pp. 694-í
72. Kuh, E. S., and R. A. Rohrer, "The state-variable approach to network 91. Polak, E., "An
analysis," Proc. IEEE, vol. 53, pp. 672-686, 1965. system to state
577-579,1966.
variable feedba
75. Kuo, F. F., and J. F. Kaiser, System Analysis by Digital Computer. New
York: Wiley, 1966. 94. Sandberg,1. W.
79. - - , "Observers for multivariable systems," IEEE Trans. Automatic 98. Silverman, L. 1\
Control, vol. AC-ll, pp. 190-197, 1966. Ph. D. dissertati
80. - - , "Canonical forms for linear multivariable systems," IEEE Trans. 99. - - , "Transfo
Automatic Control, vol. AC-12, pp. 290-293,1967. variable) form,"
1966.
81. Mantey, P.. E.,. "Eigenvalue sensitivity and state-variable selection,"
1EEE 1i·ans. Automatic Control, vol. AC-13, pp. 263-269, 1968. 100. - - , "Stable re:
Conv. Record, VI
82. Mayne, D. Q., "Computational procedure for the minimal realization of
transfer-function matrices," Proc. 1EE (London), vol. 115, pp. 1363-1368, 101. - - , "Synthesi
1968. passive realizati
1968.
83. McMillan, B., "Introduction to formal realizability theory," Bell System
Tech. 1., vol. 31, pp. 217-279,541-600,1952. 102. - - , and B. O.
01 linear system:
84. Morgan, B. S., Jr., "The synthesis of linear multivariable systems by state
variable feedback," Proc. 1964 lACC, pp. 468-472. 103. - - , and H. E
; •. ~.,.., 1
", "":', .
85. Narendra, K. S., and C. P. Neuman, "Stability oí a class 01 ulfíeremial
equations with a single monotone linearity," SI AM 1. Control, vol. 4, 104. Truxal, J. G., e
pp. 295-308, 1966. 105. Weiss, L., "The
86. Nering, E. D., Linear Algebraand Matrix Theory. NewYork: Wiley, 1963. observability," j
87. Newcomb, R. W., Active Integrated Circuit Synthesis. Englewood Cliffs, 106. - - , "On the st
N.J.: Prentice-Hall, 1968. Princeton Con!
659-680, 1968.
88. Ogata, K., State Space Analysis of Control Systems. Englewood Cliffs,
N.J.: Prentice-Hal1, 1967. 107. ---, "Lectures
590, University ,
89. Panda, S. P., and C. T. Chen, "Irreducible Jordan form realization of a
rational matrix," IEEE Trans. Automatic Control, voL AC-14, pp. 66-69; 108. ----:-:-, and P. L. I
1969. parametrized elf
90. Parks, P. c., "A new proof of the Routh-Hurwitz stability criterion using vol. 3, pp. 67-7':;
REFERENCES 641
utomatic Control, vol. the 'second method' of Lyapunov," Proc. Cambridge Phil. Soc., vol. 58,
pt. 4, pp. 694-720, 1962.
~ approach to network 91. Polak, E., "An algorithm for reducing a linear, time-invariant differential
system to state form," IEEE 7/·ans. Automatic Control, vol. AC-11, pp.
577-579,1966.
'orks. San Francisco:
92. Pontryagin, L. S., Ordinary Differential Equations. Reading, Mass.:
Addison-Wesley, 1962.
lfiable systems," M.S.
(,1969. 93. Rekasisu, Z. V., "Oecoupling of multivariable systems by means of state
variable feedback," Proc. Third Allerton Conf, pp. 439-447, 1965.
)igital Computer. New
94. Sandberg, I. W., "Linear multiloop feedback systems," Bell System Tech. J.,
vol. 42, pp. 355-382, 1963.
s Direct Method. New
95. - - , "On the Lz-boundness of solutions of nonlinear functional equa
tions," Bell System Tech. 1., vol. 43, pp. 1581-1599, 1964.
Theory. New York:
96. Schwartz, L., Théorie des distributions. Paris: Hermann & Cie, 1951, 1957.
Lr system," IEEE Trans. 97. Schwarz, R. J., and B. Friedland, Linear Systems. New York: McGraw
Hill, 1965.
EEE 7/·ans. Automatic 98. Silverman, L. M., "Structural properties of time-variable linear systems,"
Ph.O. dissertation, Oept. of Elec. Eng., Columbia University, 1966.
systems," IEEE Trans. 99. - - , 'Transformation of time-variable systems to canonical (phase
variable) form," IEEE 7/·ans. Automatic Control, vol. AC-ll, pp. 300-303,
1966.
ate-variable selection,"
53-269, 1968. 100. - - , "Stable realization of impulse response matrices," 1967 IEEE I ntern.
Conv. Record, vol. 15, pt. 5, pp. 32-37.
: minimal realization of
'01. 115, pp. 1363-1368, un. --, "Synthesis of impulse response matrices by internally stable and
passive realizations," IEEE Trans. Circuit Theory, vol. CT-15, pp. 238·-245,
1968.
ity theory," Bell System
102. --, and B. O. O. Anderson, "Controllability, observability and stability
of linear systems," SIAM J. Control, vol. 6, pp. 121-129,1968.
ariable systems by state
2. 103. - - , and H. E. Meadows, "Controllability and observability in time
)f a class of differentiaí
AM J. Control, vol. 4, 104. Truxal, J. G., Control System Synthesis. New York: McGraw-Hill, 1955.
105. Weiss, L., 'The concepts of differential controllability and differential
New York: Wiley, 1963. observability," J. Math. Anal. Appl., vol. 10, pp. 442-449, 1965.
esis. Englewood Cliffs, 106. - - , "On the structure theory of linear differential systems," Proc. Second
Princeton Conf, pp. 243-249, 1968; also SIAM 1. Control, vol. 6, pp.
?ms. Englewood Cliffs,
659-680, 1968.
107. --, "Lectures on controllability and observability," Tech. Note BN
Lllform realization of a 590, University of Maryland, Jan. 1969:· .
, vol. AC-14, pp. 66-69, 108. --,and P. L. Falp, "Oolezal's theOl:em, linear algebra with continuously
parametriZed elements,and time-varying systems," Math. System TheorJ',
stability criterion using vol. 3, pp. 67-75,1969.
642 REFERENCES
111. Wonham, W. M., "On pole assignment in multi-input controllable linear SlO. Astrom, K. J.,
systems," IEEE Trans. Automatic Control, vol. AC-12, pp. 660-665, 1967. Academic, 19T
112. - - , and A. S. Morse, "Decoupling and pole~assignment in linear multi SIL -.- , "Algebn
variable systems-a geometric approach," SI AM J. COl7trol, vol. 8, pp. Polytechnica S.
1-18,1970. S12. Athans, M., M
113. Wylie, e R., Jr., Aduanced Engineering Mathematics. New York: Networks ami
McGraw-Hill,1951. McGraw-Hill,
114. Youla, D. e, 'The synthesis of linear dynamical systems from prescribed S13. Balestrino, A.,
weighting patterns," SI AM 1. Appl. Math., vol. 14, pp. 527-549, 1966. variable systen
matic Control, .
115. - - , and Plinio Tissi, "n-port synthesis via reactance extraction-part I,"
1966 IEEE Intern. Conu. Record, vol. 14, pt. 7, pp. 183-208. S14. Bar-ness, Y., a
sampling," 1ni.
116. Zadeh, L. A., and C. A. Desoer, Linear System Theory. New York:
S15. Barnett, S., fI;.
McGraw-Hill, 1963.
Reinhold, 1971
117. Zames, G., "On the input-output stability of time-varying nonlinear
feedback systems," pts. I and n, IEEE y,·ans. Automatic Control, vol. S16. - - , "Regular
AC-ll, pp. 228-238 and 465~476, 1966. Proc. Cambrid[;
817. Barry, P. E., ":
unpublished ne
Supplementary References
S18. Bartels, R. H., ~
SI. Anderson, B. D. O., "Internal and external stability of linear time equation AX +
varying systems," SIAM 1. Control and Optimization, vol. 20, pp. 408-413, 826,1972.
1982. S19. Bengtsson, G.,
S2. - - , and M. R. Gevers, "On multivariable pole-zero cancelIations and dornain approa
the stability of feedback systems," IEEE Trans. Circuits and Syslems, S20. Bhattacharyya,
vol. CAS-28, pp. 830-833, 1981. Automatic COl1l
53, - - , and Jury, ]E. I., "Genera1.ized BezolJtian 2nd Syl,'/F.ste;· 'i'8t,ir:?,;;: ,
multivariable linear control," JEEE Trans. Automatic Control, vol. AC-21, of rnultivariabl,
pp. 551-556,1976. Atherton, Ed.),
S4. - - , and J. B. Moore, Optimal Filtering. Englewood Cliffs, NJ.: S22. Birkoff, G. and
Prentice-HalI, 1979. York: Macmill
SS. Antsaklis, P. J., and J. B. Pearson, "Stabilization and regulation in S23. Bitmead, R. R.,
linear multivariable systems," IEEE Trans. Automatic Control, vol. common diviso
AC-23, pp. 928-930, 1978. Trans. Automat
S6. Aplevich, J. D., "Direct computation of canonical forms oflinear systems S24. Bongiorno, J. J.
by elementary matrix operations," IEEE' y,"ans. Automatic Control, system," I nI. J.
vol. AC-19, pp. 124-126, 1974. . pp. 183-190, 19
S7. - - , "Tableau methods for analysis' an,d design of linear systems," S25. - - , and - - o
Automatica, vol. 15, pp. 419-429, 1979. back systems in
S8. Araki, M., and M. Saeki, "A quantitative condition for the well-posedness Contl'ol, vol. AC
SUPPLEMENTARY REFERENCES 643
;ystem theory," Int. 1. of interconnected dynamical systems," IEEE Trans. Automatic Control,
vol. AC-28, pp. 569-577, 1983.
)le systems," Preprints S9. Armstrong, E. S., ORACLS: A design system for linear multivariable
control. New York: Dekker, 1980.
)Ut controllable linear S10. Astrom, K. 1., Introduction to Stochastic Control Theory. N ew York:
12, pp. 660-665, 1967. Academic, 1970.
nment in linear multi SI L - - , "Algebraic system theory as a tool for regular design," ACTA
Polytechnica Scandinavica, Ma 31, pp. 52-65, 1979.
J. Control, vol. 8, pp.
SU. Athans, M., M. L. Destouzos, R. N. Spann, and S. 1. Mason, Systems,
emaUcs. New York: Networks and Computational Multivariable Methods. New York:
McGraw-Hill,1974.
,tems from prescribed S 13. Balestrino, A., and G. Celentano, "Pole assignment in linear multi
pp. 527-549,1966. variable systems using observers of reduced order," IEEE ll-cms. Auto
matic Control, vol. AC-24, pp. 144-146, 1979.
:e extraction-part 1,"
83-208. S14. Bar-ness, Y., and G. Langhotz, "Preservation of controllability under
sampling," Int. J. Control, vol. 22, pp. 39-47, 1975.
Theory. New York:
S15. Barnett, S., Matrices in Control Theory. London: Van Nostrand
Reinhold,1971.
ne-varying nonlinear
S16. - - , "Regular greatest common divisor of two polynomial matrices,"
ltomatic Control, vol.
Proc. Cambridge Philos. Soc., 72, pp. 161-165,1972.
SI7.. Barry, P. E., "State feedback, pole placement and transient response,"
unpublished notes.
S18. Bartels, R. H., and G. H. Stewart, "Algorithm 432, solution of the matrix
bility of linear time equation AX +XB =C," Commun. Ass. Comput Mach., vol. 15, pp. 820
vol. 20, pp. 408-413,
'1,
826,1972.
S19. Bengtsson, G., "Output regulation and internal models-a frequency
:ero cancellations and domain approach," Automatica, vol. 13, pp. 333-345, July 1977.
Circuits and Systems, S20. Bhattacharyya, S. P., "The Structure of Robust Observers," 1EEE 1l-cms.
Automatic Control, vol. AC-21 (4), pp. 581-588, 1976.
Syl'vesteT rD.atrices in.
ic Control, vol. AC-21, of multivariable systems," in Multivariable Technological Systems (D. P.
Atherton, Ed.), New York: Pergamon, 1979.
glewood Cliffs, N.J.: S22. Birkoff, G. and S. MacLane, A Survey of Modern Algebra, 3d ed. New
York: Macmillan, 1965.
)TI and regulation in S23. Bitmead,R. R., S. Y. Kung, B. D. O. Anderson, and T. Kailath, "Greatest
romatic Control, vol. common divisors via generalizedSylvester and Bezout matrices," IEEE
Trans. Automatic Control, vol. AC-23, pp. 1043-1047,1978.
xms oflinear systems S24. Bongiorrio, J. J.; andD. C. Youla, "On observers in multivariable control
. Automatic Control, system," 1m. 1. Control, vol. 8, pp. 221-243, 1968; its discussion; ·Yo.1.12,
pp. 183-190, 1970.
n of linear systems," S25. - - , and - - , "On the design ofsingle-loop single-input-oütput feed
back systems in thecomplex-frequency domain," 1EEE 1)·ans. Automatic
for the well-posedness Control, vol. AC-22, pp. 416-423,1977.
644 REFERENCES
842. - - , "Minimí;
826. Brasch, F. M., le, and 1. B. Pearson, "Pole assignment using dynamic
puters. vol. C-2
compensator," IEEE 1/·ans. Automatic Control, vol. AC-15 (1), pp. 34-43,
1970. S43. - - "An alg
Ins1., vol. 297, ¡
827. Brockett, R. W., Finite-dimensional Linear Systems. New York: Wiley,
1970. 844. - - , "lnertia
Appl., vol. 49, ~
828. Bruni, c., A. Isidori, and A. Ruberts, "A method of realization based on
the moments of the impulse response matrix," J EEE Tt·ans. Automatic 845. - - , "Synthes
Control, vol. AC-14, pp. 203-204, 1969. vol. 32, pp. 112
829. Brunovsky, P., "A classification of linear control1able systems," Kyber 546. - - , Analysis G
netika, vol. 3, pp. 173-188, 1970. Pond Woods, 1
830. Budin, M. A., "Minimal realization of discrete systems from input 547. - - , One-dime
output observations," IEEE TI·ans. Automatic Control, vol. AC-16, pp. 1979.
395-401,1971. 848. - - , "Identifi(
831. Byrnes, C. 1., and P. K. Stevens, "Pole placement by static and dynamic Prac. 1982 IF A
output feedback," Proc. 21st IEEE Con! Decision and Control, pp. 130 849. - - , "A contril
832. Cal1ier, F. M. and C. A. Desoer, "An algebra of transfer functions for 850. - - , "Feedbac
distributed linear time-invariant systems" IEEE Trans. Círcuits and Report, SUNY
Systems, vol. CAS-25, pp. 651-662, 1978. lts simplification and clarifica 851. - - , and C. H.
833. '--, and - - , "Stabilization, tracking and disturbance rejection in 852. - - , and Y. S. f
multivariable convolution systems," 1978 IEEE Con! ón ¡jeCisi~n and from finite inp
Control, San Diego, Calif., 1979. AC-24, pp. 748
834. - - , and - - , Multivariable Feedback Systems. New York: Springer 853. - - , and D. P
Verlag, 1982. IEEE TI'ans. Au
835. - - , and C. D. Nahum, "Necessary and sufficient conditions for the 854. - - , and S.
complete controllability and observability of systems in series using the 5ystem ami Con
coprime factorization of a rational matrix," IEEE 7¡·ans. Circuits and 855. Cheng, L., and j
Systems, CAS-22, pp. 90-95, 1975. variable linear 1
S3~ . Chang, F. C.~ ttThe inverse of gEn~n:1.Hz~d 'O/~1.:-~c!e:'~lT1Gndé /r.L:~·~:·L~ ~~·d·;)L~6~
the partíal fraction expansion," IEEE Trans. Automatic Control, vol. 19, S56. - - , and -
pp. 151-152, 1974. Trans. Automati,
837. Chang, T. S., and C. T Chen, "On the Routh-Hurwitz criterion," IEEE 857. Cheng, V. H. l
Trans. Automatic Control, vol. AC-19, pp. 250-251,1974. transfer functior
838. Chen, C. T., I ntroduction to Linear System Theory. New York: Halt, IEEE Trans. Au.
Rinehart and Winston, 1970. 858. Csaki, F. G., "S
839. -.- , "Irreducibility af dynamical equation realizations of sets of dif matrices," IEEE
ferential equations," IEEE Trans. Automatic Control, vol. AC-15, p. 131, 859. D' Alessandro, O
1970. : lordan form rea
S40, -.- ., "A new look at transfer-function design," Proc. IEEE, vol. 59, pp. 1971.
1580--:1585, Nov. 1971. . . 860. Daly, K. C.,"T
841. - - , "A generalizatian of the inertia theorem," SIAM J. Appl. Matiz., elementary sim
vol. 25, no. 2, pp. 158-161,1973. pp. 1-15, 1976.
SUPPLEMENTARY REFERENCES 645
;nment using dynamic 842. - - , "Minimization of linear sequential machines," IEEE 71·a/1s. Com
1. AC-15 (l), pp. 34-43, puters, vol. C-23, pp. 93-95,1974.
843. --o "An algorithm for Jordan form transformation," J. Fra/1klin
tS. New York: Wiley, Inst., vol. 297, pp. 449-455, 1974.
S44. - - , "Inertia Theorem for general matrix equations," 1. Math. Anal.
)f realization based on Appl., vol. 49, pp. 207-210,1975.
EEE 71·ans. Automatie 845. --o "Synthesis of linear sequential machines," In[ormation Control.
vol. 32, pp. 112-127, 1976.
able systems," Kyber S46. - - , Analysis amI Synthesis 01linear control systems. Stony Brook, N.Y.:
Pond Woods, 1978.
systems from input S47. - - , One-dimensional Digital Signal Proeessing. New York: Dekker,
lntrol, vol. AC-16, pp. 1979.
848. - - , "Identification of linear time-invariant multivariable systems,"
by static and dynamic Proe. 1982 1FA C Symposium on Ident Uleea ion, J une 1982.
and Control, pp. 130 849. - - , "A contribution to the design of linear time-invariant multivariable
systems," Proe. Am. Automatie Control Conf., June, 1982.
transfer fu nctions for 850. - - , "Feedback implementations of open-Ioop compensators." Tech.
~ Trans. Cireuits and Report, SUNY, Stony Brook, N.Y., 1983.
ification and c\arifica
851. - - , and C. H. Hsu, "Design of dynamic compensator for multivariable
systems," Preprint J A CC, pp. 893-900, 1971.
,turbance rejection in S52. - - , and y.. S. Kao, "Identification of two-dimensional transfer functions
Con! dn Deeision and from finite input-output data," IEEE 71·ans. Automatie Control, vol.
AC-24, pp. 748-752,1979.
New York: Springer S53. - - , and D. P. Mital, "A simplified irreducible realization algorithm,
IEEE 71·ans. Automatie Control, vol. AC-17, pp. 535-537,1972.
:nt conditions for the 854. - - , and S. Y. Zhang, "Multivariable multipurpose controllers."
:ms in series using the System and Control le tt as, vol. 3, pp. 73-76, 1983.
'E Trans. Cireuits and 855. Cheng, L., and J. B. Pearson, Jr., "Frequency domain synthesis of multi
variable linear regulators," IEEE Trans. Automatie Control, vol. AC-23,
:londc iTiali:ix Lhrougá
matie Control, vol. 19, 856. - - , and - - , "Synthesis of linear multivariable regulators," IEEE
71-ans. Automatie Control, vol. AC-26, pp. 194-202, 1981.
'witz criterion," IEEE 857. Cheng, V. H. L., and C. A. Desoer, "Limitations on the closed-Ioop
, 1974. transfer function due to right-half plane transmission zeros of the plant,"
'y. New York: Holt, IEEE Trans. Automatie Control, vol. AC-25, pp. 1218-1220, 1980.
858. Csaki, F. G., "Sorne noteson the inversion of confluent Vandermonde
zations of sets of dif matrices," IEEE Trans. AutomatieControl, vol. AC-20, pp. 154-157, 1973.
'01, vol. AC-15, p. 131, 859. D'Alessandro, O. P., anci S: .Guili;ínelli, "A direct procedure for irreducible
'Jordan form realization," Ricnerehe Di Automatiea, vol. 2, pp. 193-198,
·oe. IEEE, vol. 59, pp. 1971.
860. Daly, K. c., "The computationof Llienberger canonical forms using
'IAM 1. Appl. Math., elementary similarity transformations," Int. 1. Systems Sei., vol. 7,
pp. 1-15, 1976.
646 REFERENCES
576. - - , and M.
561. Oamen, A. A. H., and A. K. Hajdasinski, "Practical tests with different stable plant, ~
approximate realizations based on the singular value decomposition of AC-26, pp. 4O~
the Hankel matrix," Proc. 6th 1FAC Symposium on Identification,
Washington, O.C, 1982. 577. - - , R. W. L
design: The fra
862. Oatta, K. B., "An algorithm to compute canonical forms in multivariable 1EEE 7)·ans. A
control systems," IEEE Trans. Automatic Control, vol. AC-22, pp. 129
S78. - - , and J. D
132, 1977.
and their dyn;
863. - - , "Mínimal realization in companion forms," J. Frank./in Inst., vol. vol. CAS-21 (1
309, pp. 103-123, 1980.
S79. - - , and M..
864. Davison, E. J., "The output control of linear time-invariant multivariable New York: Ac
systems with unmeasurable arbitrary disturbances," IEEE Y¡·ans. Auto
580. - , and Y..
matic Control, vol. AC-17 (5), pp. 621-630, 1972; its correction, vol.
AC-20, p. 824, 1975. compensator,"
1976.
865. - - , "A generalization of the output control of linear multivariable
systems with unmeasurable arbitrary disturbances," IEEE Y¡·ans. Auto 881. - - , and Y. 1
problem: A sel
matic Control, vol. AC-20 (6), pp. 788-792, 1975.
vol 16 (C T. L,
866. - - , "The robust control of a servomechanism problem for linear time
invariant multivariable systems," IEEE Trans. Automatic Control, vol. 882. Oongarra, J. J.,
U ser's Cuide. P
systems," IEEE Trans. Automatic Control, vol. AC-28, pp. 221-222, 1983. variable robust
linear multivariable systems," Automatica, vol. 10, pp. 643-658, 1974. 1EEE Con! De<
using output feedback," IEEE Trans. Automatic Control, vol. AC-20, tion for relatiVt
pp. 516-518,1975. Control, vol. AC
872. - - , and - - , "An algorithm for the calculation of transmission 888. - - , and -
zeros of (C, A, B, O) using high-gain output feedback," IEEE Trans. solutions, "SI A
Automatic Control, vol. AC-23, pp. 738-741, 1978. 889. Eykhoff, P., SY5
873. Oesoer, C. A., Notesfor a Second Course on Linear Systems. Ne:w York: 890. Fadeev, O. K.,
Van Nostrand, 1970. Algebra. San F
874. - - , F. M. Callier, and W. S. Chan; '~Robustness of stability conditions 591. Fahmy, M. M.,
for linear time-invaríant feedback systems," 1EEE Y¡~ans. Automatic multivariable s:
Control, vol. AC-22, pp.. 586-590, 1977. . pp. 690-,-693, 19:
875..- - , and W. S.Chan, "The feedback interconnection of lumped linear 892. Fairman, F. W.
time-invariantsystems," J. Frank.lin Inst., vol. 300, pp. 335-351, 1975~
SUPPLEMENTARY REFERENCES 647
order observers," Int. 1. Systems Scí., vol. 11, pp. 1083-1094, 1980. SI09. Gopinath, B., ,
Bell Syst. Tech.
S93. Fernando, K. V.~ and H. Nicholson, "Singular perturbatíonal model
reduction of balanced systems," IEEE Trans. Automatic Control, vol. SIlO. Gupta, R. D., !
AC-29, pp. 466-468, 1982. the design of si
vol. CAS-28. Pi
S94. Ferreira, P. G., and S. P. Bhattacharyya, "On blocking zeros," IEEE
8111. Ham, Y. P., an
Trans. Automatic Control, vol. AC-22 (2), pp. 258-259, 1977.
Liapunov theo
S95. F orney, G. D., J f., "Minimal bases of rational vector spaces with applica 733- 734,1981
tions to multivariable linear systems," SIAM J. Control. vol. 13, pp. 493
SH2. Hayton, G. E.,
520,1975.
32, pp. 567-57~
S96. Fortmann, T. E., and K. L. Hitz, An lntroduction to Linear Control SI13. Heymann, M.,
Systems. New York: Dekker, 1977.
D ynamical Syst
S97. - - , and D. Williamson, "Design of low-order observers for linear S114. - - , 'The po
feedback control laws," IEEE Trans. Automatic Control, vol. AC-17, pp. Control, vol. A(
301:-308,1972.
S115. Hikita, H., "De~
S98. Francis, B. A., "The multivariable servomechanism problem from the to output tracki
input-output viewpoint," 1EEE y,·ans. Automatic Control, vol. AC-22 (3),
S116. Householder, J
pp. 322-328, 1977.
Waltham Mass.
S99. - - , and W. M. Wonham, "The role of transmission zeros in linear S117. Hsia, T. e., SYSI
multivariable regulators," lnt. J. Control, vol. 22 (5), pp. 657-681, 1975.
S118. Hsu, e. H., and
SI00. - - , and - - , "The internal model principie of control theory," Auto matrices," lnt. J
matica" vol. 12,pp. 457-465, 1976.
SU9. Ikeda, M., H. 1
S101. Franklin, G. F., and e. R. Johnson, "A condition for full zero assignment linear time-vary
in linear control systems," IEEE Trans. Automatic Control, vol. AC-26, Vol. 13, pp. 304·
pp. 521-523,1981. S120. Inouye, Y., "N
S102. Fuhrmann, P. A., "AIgebraic system theory: An analyst's point of view," discrete-time sy~
J. Franklin Inst., vol. 301, pp. 521-540, 1976. S121. Johnson, e. D..
8103. Garbow, B. S., J. M. Boyle, J. J. Dongarra, and e. B. Moler, "Matrix Regulator and
eigensystem routines-Eispack guide extensions," Lecture Notes in Control, vol. AC
Computer Sciences, vol. 51. New York: Springer-Verlag, 1977.
§¡~. Gentín, Y. V., ano S. Y. Kung, "A two-variabíe approach to the moaei amI Dynamic Sys
reduction problem with Hankel norm-criterion," IEEE y,·ans. Circuits pp. 287-489, 19~
Systems, vol. CAS-28, pp. 912-924, 1981.
S 123. Jury, E. 1., lnner
SI05. Gibson, J. A., and T. T. Ha, "Further to the preservation of controllability 1974.
under sampling," lnt. J. Control, vol. 31, pp. 1013-1026, 1980.
S 124. Kagstrom, B., al
SI06. Gohberg, 1. e., and L. E. Lerer, "Resultants of matrix polynomials," of the Jordan no
Bull. Amer. Math. Soc., vol. 82, pp. 565-567, 1976. M ath. Software, .
S107. Golub, G. H., S. Nash ande. Van Loan, "A Hessenberg-Scheer method
S125. Kailath, T., Linel
for the problem AX + XB ==C," 1EEEY,·ans. Automatic Control, vQ!; AC
24, pp. 909-913, 1979. . S126. Kalman, R. E.. "(
forros," ACTA, J
8108. - - , and J. H. Wilkinson, "Ill-conditioned eigensystems aI1d the com
putation oftheJordan canonical form," SIAM Rev., vol. 18, pp. 578-619. S127. - - , "On the COI
1976. SfAMJ. Controi
"--"'-'-'_."."--- _._--- - -- "--'---'----~~~ ~ ~.. _-_ __..-.
.•
Sl09. Gopinath, B., "On the control of linear multiple input-output systems,"
083-1094, 1980.
Bell Syst. Tech. 1., vol. 50, pp. 1063-1081,1971.
perturbational model
SIlO. Gupta, R. D., F. W. Fairman, and T Hinamoto, "A direct procedure for
,utomatic Control, vol.
the design of single functional observers," IEEE 7i-ans. Circuits Systems,
vol. CAS-28, pp. 294·-300, 1981.
)Iocking zeros," IEEE SIll. Harn, Y. P., and e T Chen, "A proof of a discrete stability test via the
-259, 1977.
Liapunov theorem," IEEE 7i·ans. Automatíc Control, vol. AC-26, pp.
or spaces with applica
733-734, 1981.
)I1trol, vol. 13, pp. 493 S1l2. Hayton, G. E., "The generalized resultant matrix," ¡ni. J. Control, vo1.
32, pp. 567-579, 1980.
~ion to Linear Control S1l3. Heymann, M., Structure and Realization Problems in the Theory oJ
Dynamical Systems. New York: Springer-Verlag, 1975.
:r observers for linear Sl14. - - , "The pole shifting theorem revisited," IEEE 7i-ans. Automatíc
:ontrol, vol. AC-17, pp. Control, vol. AC-24, pp. 479-480, 1979.
SllS. Hikita, H., "Design of exact model matching systems and its applications
ism problem from the to output tracking problems," Int. J. Control, vol. 34, pp. 1095-1111, 1981.
Control, vol. AC-22 (3), S1l6. Householder, A. S., The Theory of Matríces ín Numerical Analysis.
Waltham Mass.: Blaisdell, 1964.
nission zeros in linear S1l7. Hsia, Te, System Identificatíon. Boston: Lexington, 1977.
(5), pp. 657-681,1975. Sl18. Hsu, e H., and e T. Chen, "Conversion of cyclicity for transfer function
control theory," Auto- matrices," Int. J. Control, vol. 16, no. 3, pp. 451-463,1972.
S1l9. Ikeda, M., H. Malda, and S. Kodaina, "Estimation and Feedback in
or full zero assignment linear time-varying systems: A deterministic theory," SIAM 1. Control,
;c Control, vol. AC-26, Vol. 13, pp. 304-326, 1975.
S120. Inouye, Y., "Notes on controllability and constructability of linear
nalyst's point of view," discrete-time systems," I nt. 1. Control, vol. 35, pp. 1081-1084, 1982.
S121. Johnson, e D., "Accommodation of External Disturbances in Linear
e B. Moler, "Matrix Regulator and Servomechanism Problems," IEEE 7i-ans. Automatic
1S," Lecture Notes in Control, vol. AC-16 (6), pp. 635-644 (1971).
[-Verlag, 1977.
SJ.2!..? --., 4'Theory oí disturbance-aCCOIDJIJ.odadng controners~)) ~n ~~(}ntr(}r
lpproach to the model
and Dynamíc Systems, vol. 12 (e T Leondes, ed.). New York: Academic,
, IEEE 1i-ans. Circuíts
pp. 287-489, 1976.
S 123. Jury, E. L, Inners and Stability of Dynamic Systems. New York: Wiley,
lation of controllability
1974.
--1026,1980.
S 124. Kagstrom, B., and A. Rule, "An algorithm for numerical computation
, matrix polynomials,"
of the Jordan normal form of a complex matrix [F2]," ACM Trans.on
Math. Software, vol. 6, pp. 437-443, 1980.
;enberg-Scheer method
SUS. Kailath, T., Linear Syilems. EnglewoodCliffs,N.J.: Prentice-Hall, 1980.
'V.,
S 127. - - , "On the computation of the reachablejobservable canonical form,';
S128. Kamen, E. W., "New results in realization theory for linear time-varying 8146. Lal, M., H. Sí
analytic systems," IEEE Trans. Automatic Control, vol. AC-24, pp. 866 algorithm for
878,1979. IEEE y,-ans., \
S 129. Kimura, H., "Pole assignment by gain output feedback," IEEE y,·ans. S147. Laub, A. J., "
Automatic Control, vol. AC-20, pp. 509-516,1975. MIT Rep. ESI
S 130. - - , "Pole assignment by gain output feedback," IEEE y,·ans. Automatic 8148. - - , and B. t
Control, vol. AC-22, pp. 458-463, 1977. techniques, Al:
S131. - - , "Geometric structure of observers for linear feedback control S149. Líu, R., and 1.
laws," IEEE Trans. Automatic Control, vol. AC22, pp. 846-854, 1977. fiability of in~
S132. - - , "On pole assignment by output feedback," Int. J. Control, vol. 28, vol. AC-22, pp.
pp. 11-22, 1978. S150. Luenberger, O
S133. Klein, G., and B. e Moore, "Eigenvalue-generalized eigenvector assign matic Control, .
ment with state feedback," IEEE y,·ans. Automatic Control, vol. AC-22, S151. MacFarlane, A
pp. 140-141, 1977. Proc. 2 1st lEE.
SI34. Klema, V. e, and A. J. Laub, "The singular value decomposition: Its S152. MacFarlane, A
computation and sorne application," IEEE Trans. Automatic Control, variable systen
vol. AC-25, pp. 164-176, 1980. variable theory
S135. Krishnamurthi, V., "Implications of Routh stability criteria," IEEE 8153. Mansour, M.,
Trans. Automatica Control, vol. AC-25, pp. 554-555, 1980. Lyapunov met
8136. Krishnarao, 1. S., and e T. Chen, "Properness of feedback transfer 707- 708, 1982.
function matrices," Int. J. Control, vol. 39, pp. 57-61, 1984. S154. Maronlas, J., a
S 137. - - , and - - , "Two polynomial matrix operations," IEEE Trans. control system
Automatic Control, vol. AC-29, No. 3, March, 1984. Int. J. Systems:
vol. 10, pp. 33
S138. Kronsjo, L., AIgorithms - their complexity and efficiency. New York:
Wiley,1979. 8155. Mayne. O. Q.,
rea lization alg<
S139. Kucera, V., Discrete Linear Control- The Polynomial Equation Approach. pp. 306-307. 19
Chichester: Wiley, 1979.
S156. Miminis, G. S.,
SI4(). Kung, S., "Multivariable and multidimensional systems: Analysis and time-invariant l
Oesign," Ph.O. dissertation, Stanford University, 1977.
S157. --o and --
8141. - - , "A new iow-order approximation a!go!"ithr:0 "/1'" "/-.,h·,,
decomposition," Proc. 12th Amer. Asilomar Con! Circuits, Systems alld pp. 62-67, 1982
Computers, 1978. S158. Mital, O. T., a
S142. - - , and T. Kailath, "Fast projection methods for minimal design time-invariant s
problems in linear system theory," Automatica, vol. 16, pp. 399-403, 1980. 8159. Moler, e B., an
S143. - - , - - , and M. Morf, "A generalized resultant matrix for polynomial the exponential
matrices," Proc. IEEE Corif. on Decision and Control, pp. 892-895, 1976. S16O. Moore, B. e, '
S 144. - - , - - , and - - , "Fast and stable algorithms for minimal design variable system
problems," Proc. Fourth IFAC Int. Symposium on Multivariable Tech. Trans. Aútomati
Systems (D. P. Altherton, ed). Candon: Pergamon, pp. 97-104,1977. vol. AC-22:pp. ¡
1977. '.
S145. Kuo, Y. L., "On the im;ducible Jordan form realization and the degree
of a rational matrix," IEEE Trans. Circuir Theory, vol. CT-17, pp. 322 S161. - - , "Singular
332,1970. ConI Decision e
SUPPLEMENTARY REFERENCES 651
for linear time-varying 8146. Lal, M., H. Singh, and K. A. Khan, "A simplified minimal realization
r, vol. AC-24, pp. 866 algorithm for symmetric impulse response matrix using moments,"
IEEE T¡·ans., vol. AC-18, pp. 683-684, 1973.
edback," IEEE Trans. S147. Laub, A. 1., "Linear multivariable control: numerical considerations,"
MIT Rep. ESL-P-833, July 1978.
r EEE Trans. Automatic S148. - - , and B. C. Moore, "Calculation of transmission zeros using QZ
techniques, Automatica, vol. 14, pp. 557-566, 1978.
lear feedback control S149. Liu, R., and L. C. Suen, "Minimal dimensional realízation and identi
~2, pp. 846-854, 1977. fiability of input-output sequences," IEEE T¡·ans. Automatic Control,
I nI. J. Control, vol. 28, vol. AC-22, pp. 227-232,1977.
SISO. Luenberger, D. G., "An introduction to observers," IEEE T¡·ans. Auto
l.ed eigenvector assign matic Control, vol. AC-16, pp. 596-603, 1971.
ic Control, vol. AC-22, S151. MacFarlane, A. G. J., "Gains, phase and angles in multivariable systems,"
Proc. 21st IEEE Con! on Decision and Control, pp. 944-947, 1982.
ue decomposition: Its S152. MacFarlane, A. G., and N. Karcanias, "Poles and zeros of linear multi
IS. Automatic Control, variable systems: A survey of the algebraic geometric and complex
variable theory," /nt. J. Control, vol. 24, pp. 33-74, 1976.
bili ty critería," IEEE S153. Mansour, M., "A note on the stability of linear discrete systems and
;5, 1980.
Lyapunov method," / EEE Trans. Automatic Control, vol. AC-27, pp.
; of feedback transfer
707- 708, 1982.
-61, 1984.
S154. Maronlas, 1., and S. Barnett, "Canonical forms of time-invariant linear
'ations," IEEE Trans.
control systems: a survey with extension. PartI, single-input cases,"
Int. J. Systems Sci., vol. 9, pp. 497-514,1978; Part n, multivariable case,
t
) for minimal design time-invariant systems," Int. J. Control, vol. 18, pp. 881-887, 1973.
16, pp. 399-403, 1980. S159. Moler, C. B., and C. F. Van Loan, "Nineteen dubious ways to compute
matrix for polynomial the exponential of a matrix," SIAM Rev., vol. 20, pp. 801-836, 1978.
'01, pp. 892-895, 1976. S160. Moore, B. c., "On the ftexibility offered by state feedback il'! multi
lS for mínimal desígn variable ·systems beyond closed. loop eigenvalue assignment," IEEE
n Multivariable Tech. Trans. Automatic Control, vol. AC-21, pp. 689-692, 1976: its {;omments,
, pp. 97-104, 1977. voL AC-22, pp. 888-889 by Porterand D'Azzo, p. 889 by Gonrishankar,
1977.' .
zation and the degree
, vol. CT-17, pp. 322 S161. - - , "Singular value analysis of linear systems," in Proc. 17th / EEE
C0171 Decision Control, pp. 66-73,Jan. 1979.
652 REFERENCES
S162. - - , "Principal component analysis in linear systems: Controllability, S179. Porter, B., and
observability and model reduction," IEEE 7)·ans. Automatic Control, time tracking
vol. AC-26, pp. 17-32,1981. 1974.
S163. Morse, A. S., "Minimal solutions to transfer function equations," IEEE S180. Pugh, A. c.,
7)·ans. Automatic Control, vol. AC-21, pp. 131-133,1976. pp. 315-324,
SI64. - - , W. A. Wolovich and B. D. O. Anderson, "Generic pole assignment: 8181. Ralson, A., a
Preliminary results," IEEE Trans. Automatic Control, vol. AC-28, pp. New York: IV
503-506, 1983. S182. Rice, L Matl
S165. - - , and W. M. Wonham, "Decoupling and pole assignment com McGraw-Hill
pensation," SIAM 1. Control, vol. 8, pp. 317-337,1970. S183. Rissanen, J,
SI66. Owens, D. R, Multivariableand Optimal Systems. New York: Academic, Control, vol. S
1981. 8184. Roman, J. R.
S167. Pace, 1. S., and S. Barnett, "Comparison of algorithms of calculations of observers for
g.c.d. of polynomials," Int. J. Systems Sci., vol. 4, pp. 211-226, 1973. Trans. AutomG
S168. Pace, 1. S., and S. Barnett, "Efficient algorithms for linear system calcula S185. Rosenbrock, I
tions Part 1-Smith form and common divisor of polynomial matrices," Wiley- Intersci
Int. J. Systems Sci., vol. 5, pp. 68-72, 1974. S186. - - , 'The zeI
S169. Paige, C. c., "Properties of numerical algorithms related to computing 8187. - - , "The tn
controllability," IEEE 7)-ans. Automatic Control, vol. AC-26, pp. 130-138, vol. 25, pp. 11
1981. S188. - - , "Comm.
S170. Patel, R. Y., "Computation ofmatrix fraction descriptions of linear time A survey of t
invariant systems," IEEE Trans. Automatic Control, vol. AC-26, pp. 148 Int. J. Control
161,1981. 8189. - - , and G.
S17!. - - , "Computation of numerical-order state-space realizations and Int.1. Control.
observability indices using orthogonal transformation," Int. J. Control, 8190. - - , and A. J
vol. 33, pp. 227-246,1981. 25, pp. 389-39
SIn. - - , and N. Munro, Multivariable System Theoryand Design. New S191. Rozsa, P., and .
York: Pergamon, 1982. of a rational 1
S173. Pearson, J. B., R. W. Shields, and P. W. States, Jr., "Robust solutions to comments by ,
linear multivariable control problems," IEEE Trans. AuforYw(ic "~')!'i ¡'I}i 326. 1978.
vol. AC-19, pp. 508-517, 1974. SI92. Saeks, R., and
S174. - - , and P. W. Staats, Jr., "Robust controllers for linear regulators," disturbance re
IEEE Trans. Automatic Control, vol. AC-19 (3), pp. 231-234,1974. Ae-26, pp. 203
S175. Pernebo, L., "An algebraic theory for the design of controllers for linear S193. Sain, M. K., '
multivariable system. Part 1: Structure matrices and feedforward designo Academic, 198
Part 11: Feedback realizations and feedbackdesign, IEEE 7)-ans. Auto S194. Safonov, M. G
matic Control, vol. AC-26, pp: 171-172, pp. 1~3-194, 1981. of multivariab
S176. - - , and L. M. Silverman,:'Model reduction via balanced state space . matrix," IEEE
representation," 1EEE Trans. Automatic Control, vol. AC-2l, pp. 382-387, S195. Schumacher, J
1982. . 1EEE 7)·ans. Al
SI77.Popov, Y. M., "Invariant description oflinear time-in~aiiantcontrol1able SI%. Scott, R.W., al
systeITIs," SI AM 1. Control, vol. 10, pp. 252-264, 1972. feedback transf
S178. - - , Hyperstability 01 Control Systems. Berlin: Springer-Verlag, 1973. vol. AC-21, pp.
SUPPLEMENTARY REFERENCES 653
stems: Control1ability,
S179. Porter, B., and A. Bradshaw, "Design of linear multivariable continuous
lS. Automat ic Control,
time tracking systems," Int. J. Syste111s Sci., vo\. 5 (12), pp. 1155-1164,
1974.
tion equations," IEEE
S180. Pugh, A. c., "Transmission and system zeros," Int. 1. Contro!' vol. 26,
" 1976.
pp. 315-324, 1979.
:neric pole assignment:
S181\ Ralson, A., and P. Rabinowitz, A First Cow'se in Numerical Analysis.
,ntrol, vo\. AC-28, pp. New York: McGraw-Hil1, 1978.
S182. Rice, J., Matrix Computation and Mathematical Software. New York:
pole assignment com McGraw-Hil1,1981.
1970. S183. Rissanen, J., "Recursive identification of linear systems," SI AM J.
New York: Academic, Control, vol. 9, pp. 420-430,1971.
SI84. Roman, J. R., and T. E. Bullock, "Design of minimum order stab1e
thms of ca1culations of observers for linear functions of the state via realization theory," IEEE
>p. 211-226, 1973. Trans. Automatic Control, vol. AC-20, pp. 613-622,1975.
r linear system calcula S185. Rosenbrock, H. H., State-Space and Multivariable Theory. New York:
po1ynomial matrices," Wiley-Interscience, 1970.
S186. - - , "The zeros ofa system," Int. J. Control, vol. 18, pp. 297-299,1973.
; related to computing S187. - - , "The transformation of strict system equiva1ues," Int. J. Control,
)1. AC-26, pp. 130-138, vol. 25, pp. 11-19,1977.
S188. - - , "Comments on 'Poles and zeros of linear multivariable systems:
riptions of linear tíme A survey of the a1gebraic, geometric and complex variable theory,'"
1/, voL AC-26, pp. 148 Int. 1. Corarol, vol. 26, pp. 157-161, 1977.
S189. - - , and G. E. Hayton, "The general prob1em of po1e placement,"
pace realizations and Int.1. Control, vol. 27, pp, 837-852,1978.
lUon," Int. J. Control, S190. --', and A. J. 1. van der Weiden, "lnverse System," Int. J. Control, vol.
25, pp. 389-392, 1977.
)ry and Design. New S191. Rozsa, P., and N. K. Sinha, "Efficient algorithm for irreducible realizatían
of a rational matrix," Int. J. Control, vol. 20, pp. 739-748, 1974. lts
, "Robust solutions to comments by Y. S. Kao and C. T. Chen, Int. 1. Control, vol. 28, pp. 325
ns. Automatic Control, 326,1978.
SI92. Saeks, R., and 1. Murray, "Feedback system design: The tracking and
for linear regulators," disturbance rejectíon prob1ems," IEEE Ti·ans. Automatic Control, vol.
),231-234,1974. AC-26, pp. 203-217, 1981.
f eontrollers for linear S193. Sain, M. K., Introduction to Algebraic System Theory. New York:
nd feedforward designo Academic, 1981.
~n, IEEE Trans: Auto 8194. Safonov, M. G., A. 1. Laub, and G. L. Hartmann, "Feedback properties
94, 1981.
of multivariable systems. The role and use of the return diffetenee
l balanced state space
matrix," IEEE Trans., vol. AC-26, pp. 47-65, 1981.
,\. AC-21,pp. 382-387,
8195. Schumacher, J. M., "Compensator synthesis using (C, A, B)~p'airs,"
IEEE Trans. Automatic Control, vol. AC-25; pp. 1133-1137, 1980' .
·invariant co ntrollable
S196. Seott, R. W., and B. D. O. Anderson, "Comments on 'Conditíons fora
972.
feedback transfer matrix to be proper,''' IEEE Ti·ans. Automatic Control,
Jringer-Verlag, 1973.
voI.AC-21, pp. 632-634,1976.
654 REFERENCES
8197. Shieh, L. S., and Y. T. Tsay, "Transformations of a class of multivariable 8216. W olovich, W
control systems to block companion forms," IEEE Trans. Automatic IEEE Ti'ans. A
Control, vol. AC-27, pp. 199-203,1982. 8217. - - , "Multiv
8198. Silverman, L., "Realization of linear dynamical systems," IEEE Ti·ans. IEEE Ti'al1s. 1
Automatic Control, vol. AC-16, pp. 554-567, 1971. 8218. - - , Linear ~
8199. Smith, M. e, "Matrix fractions and strict system equivalence," Int. 1. 8219. - - , "Outpu
Control, vol. 34, pp. 869-884,1981. vol. AC-20, pr
8200. Stewart, G. W., Introduction to Matrix Computations. New York: S22{t - - , "Sket pr
Academic,1973. vol. AC-23, Pf
8201. Tsui, C. e, and e T. Chen, "An algorithm for companion form realiza 8221. - - , "On the
tion," Int. J. Control, vol. 38, pp. 769-777,1983. Ti·ans. Automa
8202. - - , "Computational aspects in linear system theory," Ph.D. Disserta 8222. - - , "Multípl
tion, SUNY, Stony Brook, 1983. Automatic Ca;
8203. Van Dooren, P. M., "The generalized eigenstructure problem in linear 8223. - - , P. Ants
system theory," IEEE Trans. Automatic Control, vol. AC-26, pp. 111-129, minimal and
1981. Control, vol. A
8204. Varga, A., and V. Sima, "Numerically stable algorithm for transfer 8224. - - , and P. I
function matrix evaluation," Int. J. Control, vol. 33, pp. 1123-1133, 1981. variable syster
8205. Vidyasagar, M., "Conditions for a feedback transfer function matrix to 465,1979.
be proper," IEEE Trans. Automatic Control, vol. AC-20, pp. 570-571, 8225. Wonham, W.
1975. systems?" SI ~l
8206. - - , Nonlinear Systems Analysis. Englewood Cliffs, N.J.: Prentice-Hall, 8226. - - , "Towan
1978. System, Man ú
8207. - - , "On the well-posedness of large-scale interconnected systems," S227. - - , Linear
IEEE Trans. Automatic Control, vol. AC-25, pp. 413-421, 1980. New York: Sp
8208. Wang, J. W., and e T. Chen, "On the computation of the characteristic 8228. - - , and J. B.
polynomial of a matrix," IEEE Trans. Automatic Control, voL AC-27, multivariable s
pp. 449-551, 1982. 8229. Wu, M. Y., "A
8209. Wang, S. H., Design of Linear Multivariable Systems, Memo. No. ERL Trans. Automal
Ivn09, University of ü3.liforrüa, Berkeley, 1971 J~iÍ(jl. --,'':'ome (il
8210. - - , and E. J. Davison, "A minimization algorithm for the design of 1EEE Trans. A
linear multivariable systems," IEEE Trans. Automatic Control, vol.
8231. - - , "SolutiOI
AC-18, pp. 220-225, 1973.
J. Control, vol.
8211. - - , and - - , "A new invertability criterion for linear multivariable
8232. - - , "A succe
systems," IEEE Trans. Automatic Control, vol. AC-18, pp. 538-539, 1973.
time varying sy
8212. Wilkinson, J. H., The Algebraic Eigenvalue Problem. London: Oxford
University Press, 1965. 8233. Wu, y, e, ane
dynamíc systen
8213. - - , and .C. Reinsch (Eds.), Handbook for Automatic Computation, 504, 1980.
Linear Algebra, Vol. 11. New York:' Springer-Verlag, 1971.
. . 8234. Yeung, K. S., "
8214. Willems; J. c., The Analysis of FeedbackSystems. Cambridge: MIT
Trans. Automat.
Press, 1971.
8215. Willems, J. L., "Designof state observerS for linear discrete-time systems," 8235. Youla, D. e. J
Int. J. Systems Sci., vol. 11, pp. 139-147,1980. design of optirr
SUPPLEMENTARY REFERENCES 655
lclass of multivariable S216. Wolovich, W. A., "On determining the zeros of state-space systems,"
SEE Trans. Automatie IEEE 7]-ans. Automatie Control, vol. AC-18 (5), pp. 542-544, 1973.
S217. - - , "Multivariable system synthesis with step disturbances rejection,"
;ystems," IEEE Trans. IEEE 7]-ans. Automatie Control, vol. AC-19 (1), pp. 127-130,1974.
5218. --o Linear Multivariable Systems. New York: Springer-Verlag, 1974.
n equivalence," Int. J. S219. - - , "Output feedback decoupling," IEEE 7]-ans. Automatie COl1trol,
vo!. AC-20, pp. 148-149,1975.
ltations. New York: S220. - - , "Sket prime polynomial matrices," IEEE Trans. Automatie Control,
vol. AC-23, pp. 880-887, 1978.
mpanion form realiza S221. - - , "On the stabilization of closed-Ioop stabilizable systems," IEEE
7]-ans. Automatie Control, vol. AC-23, pp. 1103-1104, 1978.
eory," Ph.D. Disserta S222. - - , "Multipurpose controllers for multivariable systems," IEEE y,-ans.
Automatic Control, vol. AC-26, pp. 162-170, 1981.
:ure problem in linear S223. --, P. Antsaklis, and H. Elliott, "On the stability of solutions to
.1. AC-26, pp. 111-129, minimal and nonminimal design problems," IEEE Trans. Automatie
Control, vol. AC-22, pp. 88-94, 1977.
.1gorithm for transfer S224. - - , and P. Ferreira, "Output regulation and tracking in linear multi
, pp. 1123-1133, 1981. variable systems," IEEE 7]-ans. Automatie Control, vol. AC-24, pp. 460
fer function matrix to 465, 1979.
AC-20, pp. 570-571, S225. Wonham, W. M., "Tracking and regulation in linear multivariable
systems," SIAM J. Control, vol. 11, pp. 424-437,1973.
Is, N.J.: Prentice-Hall, S226. - - , "Towards an abstract internal model principIe," IEEE y,-ans.
System, Man and Cybernetics, vol. SMC-6 (11), pp. 735-740, 1976.
:rconnected systems," 8227.. --, Linear Multivariable Control: A Geornetrie Approaeh, 2d ed.
.3-421, 1980.
New York: Springer-Verlag, 1979.
,n of the characteristic
S228. - - , and J. B. Pearson, "Regulation and internal stabilization in linear
. Control, vol. AC-27,
multivariable systems," SIAM J Control, vol. 12, pp. 5--8,1974.
S229. Wu, M. Y., "A note on stability of linear time-varying systems," IEEE
?ms, Memo. No. ERL
Trans. Automatie Control, vol. AC-19, p. 162,1974.
§23®. - - , ·...S ome new resuits in ¡inear Lime-varying syslems" (lecn. conesp.),
thm for the design of
IEEE y,-ans. Automatie Control, vol. AC-20, pp. 159-161, 1975.
tomatie Control, vol.
S231. ---, "Solutions of certain classes of linear time varying systems," Int.
J. Control, vol. 31, pp. 11-20,1980.
r linear multivariable
S232. - - , "A successive decomposition method for the solution of linear
18, pp. 538-539, 1973.
time varying systems," Int. J. Control, vol. 33, pp. 181-186, 1981.
~m. London: Oxford
S233. Wu, Y. c., and Z. V. Rekasius, "Deterministic identification of linear
dynamic systems," IEEE Trans. A~tomatie Control, vol. AC-25~pp. 501
tomatie Computation,
504,1980.
,lag, 1971.
S234.. Yeung, K. S., ~'A necessary condUíón for H urwitz polynorIlials;'" IEEE ..
!s. Cambridge: MIT
Trans. Automatie Control, vol. AC-27, pp. 251-252, 1982,
liscrete-time systems," S235. Youla, D. c., 1. J. Bongiorno, and H. A. labr, "Modern Wiener-Hopf
design of optima] controllers- Part 1 and Part n." IEEE y,-uns. Auto
656 REFERENCES
Additivity, 73,88
Adjoint equation, 163,
Algorithm
Euclidean, 579
Feedback implement
Leverrier,67
pole assignment, 476
row searching, 550
state estimator, 359, :
state feedback, 337, 3
triangularization, 58S
Analytic continuation, :
Analytic funclion, 78, 5:
Asymptolic stabilily, 4C
Basis, 14
change of, 18, 22
minimal polynomial,
orthonormal, 17,90
polynomial, 607, 621
Be·zout identity, 5.95
Black box,· 72 .
Blocking zer6; 499, 633
Bounded-input bc>undé
stabíHty, 385, 394
Canonical dccomposilic
1976.
Additivity, 73, 88
Canonical-form dynamical equation
AIgorilhm
Jordan, 150,209
ElIclidean, 579
observable, 241, 286, 329
Levenier, 67
Cayley-Hamilton theorem, 49, 67
trianglllarizalion, 589
Cholesky decomposilion, 413
unifGnn!y, 403
495
Compensator equation, 460
Basis, 14
Controllability, 176
dilTerential, 180
grammian, 184,322
orthonormal, 17, 90
index, 188
polynomial, 607,621
índices, 190
instantaneous, 181
Siack box, 72
output, 214
657
658 INDEX
Coprimeness, 579
(Dynamical equation)
(Function)
left,592
solution of, 134
integrable, 388
right,592
linear, 19
Convolution integral, 81
of matrix, 45
assignment, 336
Function space, 10, 391
Decoupling,372
index of, 47, 66
Fundamental loop, \02
static, 502
multiplicity or, 41, 66
Fundamental rnatrjx, 13
Degree,577 Eigenvector,33
row,600 generalized, 39
Gaussian elimination, 61
Delta function, 74
Elementary operation, 542, 587
Generalized eigenvector,
Description of systern
Equilibrium state, 401
chain of, 39
external, 3, 71
in the sense of Lyapunov (i.s.L.), 153
Gram determinant, 172,
input-outpul, 71, 72
strict system, 294
Grammian, controlIabili
internal, 3, 71
transrormation, 146, 151,200
observability, 197
state-variable, 71, 83
zero-state, 148
tion, 560
positive semidefinite, 4
Divisor, 579
outpul, 343, 377, 379,474
Hurwitz polynomial, 395
Iefl,592
unity, 432, 457
right, 592
wcll-posedness or, 118
Identification, determínisl
Domain, of function, 19
Feedback gain malrix, 341, 348
Impulse function, 74
of linear operator, 27
Fíeld,8
Impulse response, 75
matrix,76
Inner product, 59
Input-output pair, 87
equivalent, 146
Floquet, theory of, 154
input-normal, 322
Fraction,599
lordan-form, 150,209
irreducible, 605
irreducible polynomíal.1
tion,291
output~normal, 322
block,223
Irreducible realizatíon, 20
reducible, 207
Function, analytic 554
INDEX 659
(Function)
J ordan block, 38, 49
ation)
integrable, 388
Jordan-canonical-form matrix, 37,41,53,
14
línear, 19
60,209
of matrix, 45
36
',41,66
351
)
Linear mapping, 20
Linear operator, 20
change of basis, 22
alion, 542,587
Generalized eigenvector, 38
nullity, 29, 31
~, 401
chain of, 39
,200,296
grade of, 38
Linear space,9
94
Grammian, controllability, 184
basis of, 14
1, 146, 151,200
observability, 197
dimension of, 13
281
f,143
)rm of, 56
lit y, 333
324
Homogeneity, 73, 88
t, 377, 445
32
7,379,474
put,432
344, 379
function of, 51
of,118
ltrix, 341,348
Impulse funclion, 74
Hermite, 589
Impulse response, 75
hermitian,412
Ibers,9
matrix,76
nonsingular, 28, 32
:ions,9
Inner product, 59
norm of, 57
,9
Input-output pair, 87
orthogonal, 220, 570
,154
Internal model principIe, 490
polynomial, 587
tion,291
rank or, 27
554
660 INDEX
(Matrix)
Parameter perturbation, 491 Realization, 155
stale transition, 137
Parameterization, 540 input-normal, 322
symmetric, 412
Periodically varying system, 153 internally balanced, 2".
Sylvester, 582
Physica\ system, 1 irreducible, 208
system, 293, 439
Pole, 441, 623 minimal-dimensional,
trace of, 67,129
Pole-zero cancellation, 437,443,444 of time-varying diffe
unitary, 220
unstable, 444, 459 252
Minimal design problem, 529 Pole-zero excess inequality, 459, 487 output-normal, 322
Realizatíon, 155
Similarity transformation, 23, 60
Jrbation, 491
input-normal, 322
Simulation ofdynamical equation, 91
!n,540
internally balanced, 271, 322
analog computer, 92
irreducible, 208
digital computer, 93
,1
minimal-dimensional, 233
Singular value decomposition, 61,269,569
of time-varyíng differential equation,
Solution space, 135
lIation,437, 443,444
252
Stability, asymptotic, 403, 409
.459
output-normal,322
bounded-input bounded-output, 385
partíal, 302
exponential, 405, 409
:rix, 587
reducible, 207
global, 403
:ed,601
Reachability, 176,227
in the sense of Lyapunov (i.s.L.), 402
306,612
,589
512
Representation
Stabilization, 339,490
601
of linear operator, 2 I
Statc, 83 -86
i91
of vector, 15
definitíon of, 83
Response, 72
equilibrium,401
matrix, 171,182,184
'Inite matrix, 4 I3
O
zero-input,89, 140
State equation, 87, 89, 134
413
matrix,82
generalized, 608
asymptolic, 356
Riccati equation, 68
full-dimensiollal,355
of,468,618
Ring, 9, 588
fUllctional, 369
315
l66,613
Routh table, 397
State feedback, 334
70
,293,295
Subspace, 11,27
Superposition principie, 73
53
matrix, 123
Svivestcr nl.rJlrix; SR2
,553
4í2
scalar,8
continuous-time,121
Scalar product, 59
discrete-time, 121
tor, 27
27,28,32
f,546
output, 121
relaxed,·385
latriees, 31
\,82
weightíng, 121
time-invariant, 80, 89
r.
z transform of, 122
zero-memory, 72
Servocompensator, 504
32
Similar matrices, 23
Time varying, 80,89
82
662 INDEX
Transfer function, 82
representation of, 15
matrix, 82, 90
Vector space, 9
sampled, 122
basis of, 14
Transformation, 143
complex,10
Lyapunov, 153
see also Linear space
orthogonal, 220,222
Tree,102
Weighling sequence, 121
branch,102
normal, 102
Zero,623
blocking, 499,633
equation, 87, 89
Zero-memory system, 72
Vandermonde determinant, 64
z transform, 122
66
1 or, 1S
13
r space
118
once, 121
633
ng, 292
.ling,292
143,627
,tem,72
ilence, 148
-n,122
l
:j
1
1
------_._._---'"_._--"----~._-~-----------_._-------~----~----_._---~--_._--------------~---------------------------~--'"._-----