Académique Documents
Professionnel Documents
Culture Documents
and Design
Li near
-~-'
and
HRW
Series in
Electrical and
Computer Engineering
De~
:.
Chi-Tsong Chen
Professor, Department
State University of Nev
HOLT, RINEHART /J
New York
Montreal
Chicagc
Toronto
Chi-Tsong Chen
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
HOLT, RINEHART AND WINSTON, INe.
New York
Montreal
Chicago
Toronto
San Francisco
Philadelphia
London
Sydney
Tokyo
Ineludes indexo
1. System analysis.
QA402.C442 1984
ISBN 0-03-060289-0
Printed in lhe United States of America
9 038 987
Holt, Rinehart and Winston
The Dryden Press
Saunders College PUblishing
To
Beatrice
and
Janet, Pauline, Stanley
cluding
without
d be
Preface
XIII
Glossary of Symbol:
Chapter 1
Introe
1-1 ".
1-2 l
j
Chapter 2
linea
2-1 1
2-2 1
2-3
2-4 1
2-5
2-6
2-7
Contents
Preface
Xlll
Glossary of Symbols
Chapter 1
xvii
Introduction
Chapter 2
2-3
2-4
2-5
2-6
2-7
Introduction
6
Change of Basis
17
19
45
Series
54
vii
viii
CONTENTS
4-4
Problems
62
4-5
Chapter 3
70
3-1 Introduction
70
Relaxedness
77
80
Time Invariance
Transfer-Function Matrix
81
83
Dynamical Equations
Linearity
87
Time Invariance
89
Transfer-Function Matrix
90
3-4 Examples
94
101
Time-Varying Case
108
Time-Invariant Case
111
Well-Posedness Problem
114
Problems
125
Chapter 4
Matrices
133
4-1 Introduction
133
Time-Varying Case
134
134
Solutions of x = A(t)x
139
Time-Invariant Case
146
*Time-Varying Case
151
153
Periodic A(o)
Chapter 5
Conti
Equa
5-1
5-2
5-3
5-4
5-5
]
]
*5-6
*5-7
*5-8
*5-9
(
(
Chapter 6
Irredu
Identi1
6-1
6-2
6-3
It
T
R
Ir
F
CONTENTS
4-4
4-5
70
ms
Chapter 5
5~3
83
5-4
letworks
101
lescription and the State
lposite Systems
5-5
108
*5-6
114
5-1
5-2
90
ix
*5-7
*5-8
*5-9
Introduction
168
Linear Independence of Time Functions
170
175
Controllability of Linear Dynamical Equations
Time-Varying Case
175
*Differential Controllability, Instantaneous Controllabil
180
ity, and Uniform Controllability
Time-Invariant Case
183
*Controllability Indices
187
192
Observability of Linear Dynamical Equations
Time-Varying Case
192
*Differential Observability, Instantaneous Observabil
196
ity, and Uniform Observability
197
Linear Time-Invariant Dynamical Equations
*Observability Indices
198
Canonical Decomposition of a Linear Time-Invariant Dyna
mical Equation
199
Irreducible Dynamical Equations
206
Controllability and Observability of lordan-Form Dynamical
Equations
209
Output Controllability and Output Function
Controllability
214
Computational Problems
217
Concluding Remarks
226
Problems
227
I Impulse-Response
Chapter 6
n
134
6-1
6-2
'34
~quation
139
146
6-3
Introduction
232
The Characteristic Pplynomial and the Degree of a Proper
Rational Matrix
234
Irreducible Realizations of Proper Rational
Functions
237
Irreducible Realization of ~1D(s)
237 .
240
Irreducible Realizations of g(s) = N(s)lD(s)
Observable Canonical-Fonn Realization
240
Controllable Canonical-Fonn Realization
243
CONTENTS
6-4
*6-5
*6-6
*6-7
*6-8
*6-9
6-10
Chapter 7
Equations
252
Functions
253
Methods
265
272
Method II. Row Searching Method
Irreducible Realizations of G(s): Coprime Fraction
Method
276
Controllable-Form Realization
276
Data
300
Persistently Exciting Input Sequences
307
Concluding Remarks
313
317
Problems
7-3
7-4
7-5
*7-6
7-7
Chapter 8
8-1
8-2
8-3
8-4
*8-5
*8-6
8-7
324
Introduction
324
Single-Variable Case
325
*Multivariable Case
325
State Feedback
334
Single-Variable Case
334
Stabilization
339
*Multivariable Case
341
Method I
341
Method II
345
Method III
347
Nonuniqueness of Feedback Gain Matrix
348
Assignment of Eigenvalues and Eigenvectors
351
Effect on theNumerator Matrix of G(s)
352
Computational Problems
353
Stab
Chapter 9
Linea
tion, ~
9-1
9-2
9-3
9-4
9-5
CONTENTS
State Estimators
354
Method I
357
358
Method II
Reduced-Dimensonal State Estimator
361
Method 1
361
363
Method II
7-5 Connection of State Feedback and State Estimator
Functional Estimators
369
Problems
378
7-4
atrix
245
tion
249
19 Differential
lal Transfer
ix
257
268
IPosition
lId
272
prime Fraction
\ 276
\lbility Indices
\285
1287
Chapter 8
307
IJ9
324
325
Chapter 9
339
I-Nonzero Set
s)
rain Matrix
348
351
352
~d Eig~nvectors
ix of G(s)
53
365
384
8-1 lntroduction
384
Description
385
Time-Varying Case
385
Time-Invariant Case
388
Time-Varying Case
400
Time-Invariant Case
407
Problems
425
284
xi
9-1 lntroducton
432
Systems
434
Systems
439
440
Parallel Connection
Tandem Connection
441
Feedback Connection
444
Single-Variable Case
458
xii
CONTENTS
Single-Variable Case
488
Multivariable Case
495
Designs
501
504
State-Variable Approach
9-7 Design of Compensators: Input-Output Feedback
Sytems
506
Single-Variable Case
506
Multivariable Case
511
Implementation 1
517
Implementation 11
519
Applications
523
Decoupling
523
Decoupling
526
Model Matching
528
Problems
536
Appendix A
Elementary Transformations
A-l
*A-2
A-3
*A-4
542
Oaussian Elimination
543
Householder Transformation
544
546
Problems
553
Appendix B
Appendix C
Appendix O
Appendix E
556
564
Decomposition
565
Problems
Appendix F
554
570
576
MB = N
572.
559
Appendix G
Poly
0-1
0-2
0-3
0-4
0-5
0-6
Appendix H
Pole
ProbJ
References
Index
657
636
CONTENTS
,Je Assignment
468
Appendix G
~nominator-
ce Rejection
G-l
G-2
G-3
G-4
G-S
G-6
488
ronrobust
Appendix H
Jut Feedback
517
554
of Sampling
572
559
636
577
Coprimeness of Polynomials
577
Polynomial Matrices
587
Problems
618
Problems
References
Index
657
ompensators
xiii
635
623
Preface
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
duced in Chapter 4.
realizations (Chapter
(Appendix G). Mono
Hence it was decided
polynomials and poi
This, however, will d
the topic was grouped
The logical sequen
Chapter 2
Secs. 2-]
to 2-5
Appendix A
Secs. G-I
Appendix G
("'
Single-variable
cases of
Secs. 9-5 to 9- 7
Secs. 9-5
9-6
9-7
Chapter 1
Chapter 2
Chapter 3 (Skip. Th
Chapter 4 (Ski.p Th
Chapter 5 (Empha~
5~5, and
Chapter6
Chapter 7
Charter 8
PREFACE
eh is well covered in
etric Approach, 2d ed.,
pe of this text. Hence
fer-function matrix in
o redevelop, probably
, of the state variable
~en the state-variable
Jll. The first one is to
: and efficient methods.
:pts which are essential
nith-McMillan form is
ljective is to enable the
,equently, most results
utation and for digital
: two problems of each
lic and give confidence
1 of the row searching
:d, this is possible even
lfficiently small.
the same as that of the
ra more extensive than
text can be developed
~s.
For example, the
by using the concepts
~an be found in a large
l here. In view of our
Jasis and then develop
algebra, such as ring,
tlem (Chapter 6) can be
m be extended to delay
ltidimensional systems.
ry, which was initiated
ecent years most of the
cepts used in algebraic
nts and require more
l. AII the results and
Iy elementary concepts
xvii
Chapter 2
Secs.2-1
lo 2-5
!
-======
Chapter 4
.!
Seco 5-8
~ Appendixes e and D
Chapler 6
Seco 6-2
Chapler 8
Charler 7
1 - - - - Hankel melhod
Secs. G-l
G-2
lo 6-9
----+-
Seco 7-4
Single-variable
cases of
Secs. 9-5 lo 9-7
Chapler 9
Secs. 9-1 /
Appendix H
Secs. 9-5
9-2
9-6 --------9-3 --------Sec.9-4
9-7
xviii
PREFACE
We emphasize the exact meanings of theorems and their implications; hence the
proofs of a number of theorems are skipped. For example, we prove only
Theorems 2-1 and 2-2 in Chapter 2. We skip the proofs ofTheorems 4-1, 4-2,
and others. In the second course, we cover the following:
Appendix A
Section 5-8, controllability and observability indices
Hankel method (Section 6-4 and method II of Section 6-5)
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
g:
m 6-5)
Section 6-5)
xix
Q.E.D.
I
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
Q.E.D.
I
A,B,P, .
u, y, 0:, .
u, y, (l.,
fe
v(A), ...
A',x', .
A*, x*, .
det A, ...
iC
~(s)
~[sJ
p(A), rank A
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
dt te,
= (~a.)
dt
1)
ff! [A] ~ (ff! [aij]), . " When an operator is applied to a matrix or a vector,
ational or polynomial
~(s).
. A(s).
e as block diagonal
a matrix or a vector,
Jlied to every entry of
1-1
The Study 01
1
Introduction
1 -1
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
INTRODUCTION
1-2
Bool~
The study of systems may be divided into four parts: modeling, setting up
mathematical equations, analysis, and designo The development of models for
physieal systems req uires knowledge ofeach particular field and sorne measuring
devices. For example, to develop models for transistors requires the knowledge
of quantum physics and sorne laboratory setup. Developing models for auto
mobile suspension systeins iequires actual testing and measurements; it cannot
be achieved by the use of pencil and paper alone. Thus, the modeling problem
should be studied in connection with each specific field and cannot be properly
covered in this text. Hence, we shall assume that models of physical systems,
or systems, are available to us in this text.
The mathematical equations which will be used in this text to describe
Y(s)=
and
;: modeling, setting up
~lopment of models for
~ld and sorne measuring
requires the knowledge
oping models for auto
leasurements; it cannot
, the modeling problem
and cannot be properly
ds of physical systems,
n this text to describe
(1-1 )
and
x(t) = Ax(t) + Bu(t)
(1-2a)
+ Eu(t)
(1-2b)
y(t) = Cx(t)
and their extensions to the time-varying case. Equation (1-1) describes the
relationship between input u and output y in the Laplace-transform domain
and is called the input-output dfscription or the external description in the
frequency domain. The matrix G(s) is called the trans!erl!1nction matrix, and
its elements are limited to rational functions. In the design, G(s) will be factored
as ratios of two polynomial matrices, such as N(s)D- 1 (s) or F- 1 (s)H(s), and is
said to be in the polynomial fraction formo The set of two equations in (1-2)
is called a dynamical equation or state-variable equation. Equation (1-2a) is a
set of first-order differential equations. Ir (1-2) is used to describe a systern,
it is called the dynamical-equation description, the state-variable description,
or the internal description. These two types of equations will be developed
from the concepts of linearity and time invariance. We shall then show, by
examples, how they can be used to describe systems.
The major portion of this text is devoted to the analysis and design centered
around Equations (1-1) and (1-2). Analysis can be divided into quantitative
and qualitative. The former .can now be delegated to analog or digital com
puters; hence we emphasize the latter. Various properties of these equations
will be thoroughly investigated. Their relationships will be established. A
number of design procedures will then be developed from the study.
The design problem studied in this book is not exactly the design of feedback
control systems. In our design, we are interested in the exact conditions,
without any constraints on the complexity of compensators, to achieve certain
design objectives, such as stabilization or poIe placement: What are the mini
mum degrees which compensators must possess to achieve such design? We
are also interested in developing simple and efficient procedures to carry out
the designo In the design of control systems, the ultimate purpose is to design a
system to meet sorne performance criteria, such as the rise time, overshoot,
steady-state error, and others, with or without constraints on the degrees of
compensators. A design problem with constraints is significantly different
from the one without constraints. In this text, the performance of control
systems is not considered. Hence our study is not a complete study ofthe design
of control systems, although most of the results in this text are basic and useful
In its designo For a discussion ofthe design of c~ntr~l systems, see References 3,
S46, and S96. In this section we give a brief descripfion of the contents of each
chapter.
We review in Chapter 2 a number ofconcepts and results in linear algebra.
The objective of this chapter is to enable the reader to carry out similarity
transformations, to solve linear algebraic equations, and to compute functions
INTRODUCTION
11
(urrent
SOUfCC
1:
151
In
,.
I
11
In
In
___ L
lo)
Figure 1-1
lb)
-r
,.
(a)
------o:<
5
;
not indispensable, in
Itput description and
escriptions are devel
, and time invariance.
be set up for systems.
screte-time equations
'oblem in the feedback
nical equations. We
t dynamical-equation
~en
the input-output
blished.
lity and observability.
m from the networks
lual to 1. There is no
wn in Figure l-l(b);
loes not play any role
tion, the concepts of
vo concepts are also
nd the prediction or
itions for a dynamical
We also discuss the
mtroduce an efficient
luation to an irreduci
)nal transfer-function
lservable linear time
Ll matrix. lts solution
ions. lt also offers a
Illal-amplifier circuits.
~en the state-variable
esign of linear time
ility and observability
(a)
(b)
GJ-r
~J
s - \
(c)
;tion, 1.
Figure 1-2 Three dilTerent connections 01" 1/(s -1) and (s - 1)/(s
+ 1).
rt
BD= b
bm
2-1
Introduction
2-2
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
linear Spaces
Numbers correspond to th
AB~ [a,
CA = C[al
a,
a z ' - . a m ] = [Cal
Caz'" Cam ]
BD= (blJ
bt D= (blDJ
b~D
bm bmD
; and results in linear
he topics are carefully
i are introduced. The
land the mechanism of
ltions, to find Jordan
: functions of a matrix,
:ction 2-9, Coneluding
Id linear space over a
: field of real numbers,
"unctions. In order to
:roduce, in Section 2-3,
representations of the
:ar operators and their
ion is embedded here.
equations are studied.
ction 2-6, we show that
1; this is achieved by
as basis vectors. In
(2-1 )
(2-2)
(2-3)
2-2
a multiplicative inverse (l/a, except a = O) in the set. Any set with these proper
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
operations caBed addition "+" and multiplication "'''; the two operations
are defined over :F such that they satisfy the following conditions:
1. To every pair of elements a and 13 in :F, there correspond an element a + [3
in :F called the sum of a and [3, and an element a . [3 or af3 in :F, called the
produet of a and [3.
2. Addition and multiplication are respective1y commutative: F or any a, [3 in :F,
a+f3=[3+a
3. Addition and multiplication are respectively associative: For any a, 13, y in :F,
(a + fJ) + y = a + (fJ + y)
a . (13 + y) = (a . 13) + (a . y)
spaee.
Definition 2-2
A linear space over a fie
of elements caBed vecte
and sealar multiplieatioj
that they satisfy all the
Consider the set of numbers that consists of O and 1. The set {O, 1} does not
form a fie1d ifwe use the usual definition of addition and multiplication, because
the e1ement 1 + 1 = 2 is not in the set {0,1}. However, if we define 0+ O =
1 + 1 =0, 1 + O = 1 and 01 =0,0 =0,1'1 = 1, then it can be verified that {O, 1}
with the defined addition and multiplication satisfies aB the conditions listed
for a field. Hence the set {O, 1} with the defined operations forms a field. It is
called the field of binary numbers.
I
Example 2
Consider
th~
-YJ
a(f3x) = (af3)x .
.[ y x .
where x and y are arbitrary real numbers. The set with the usual definitions of
matrix addition and multiplication forms a fie1d. The elements O and 1 of the
and
alled scalars and two
'; the two operations
:onditions:
lond an element ex + 13
:>r rxf3 in ?F, called the
:ive: For any ex, 13 in ?F,
e: For any a, /3, '}' in ?F,
(13 . y)
For any ex, 13, y in ?F,
+13=0. Theelement
Note that the set of a11 2 x 2 matrices does not form a field.
From the foregoing examples, we see that the set of objects that forms a
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
familiar ones: the field of real numbers, the field of complex numbers, and 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
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
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
nomials do not form a field because they have no multiplicative inverse. 3
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.
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
matical terminology, is ca11ed a linear space, or a vector space, or alinear vector
space.
n element y in ?F such
Je nverse.
Defnition 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:
(X 2 +X 3 )
f![ contains a vector, denoted by O, such that O +x = X for every x in f![.
The vector O is ca11ed the zero vector or the origino
5. To every x in ~ there is a vector x in f![, such that x +x =0.
6. To every rx in ff, and every J!: in f![, there corresponds a vector r:xx in f![ ca11ed
the scalar product of rx and x.
7. Scalar multiplication is associ:ive: For any rx, f3 in fF ap<:l any x in f![,
a(f3x) = (rxf3)x.
4.
A set with aB properties of a field except property 7 in Definilion 2-[ is caBed a ril1g or, more
precisely, a commutative ring with (multiplicative) identity. The set of integers forms a ring, as
does the set of polynomials with real" codficients.
-- -
-----
-------._~-----_._-_._---
10
-- - - -
--~-----
-~-
-~-----------
--_..
_--~-
Example 1
A field forms a vector space over itself with the vector addition and scalar
multiplication defined as the corresponding operations in the field. For
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
I
vector spaces, but not (~, ~(s)).
Example 5
x+2x + 3x =0.
Example 2
The set of all real-valued piecewise continuous functions defined over ( - 00, 00)
forms a linear space over the field of real numbers. The addition and scalar
I
multiplication are defined in the usual way. lt is called a function space.
We introduce one 1
Example 3
Definition 2-3
Given a field ff, let ff" be all n-tuples of scalars written as columns
Xlil
(2-4)
X= xt'
[ x.,
where the first subscript denotes various components of X and the second
subscript denotes different vectors in ff". lf the vector addition and the scalar
multiplication are defined in the following way:
Xli + X1Jl
x +Xj=
X2i
X2]
XIIi +Xllj~
aXli
cxx.=
!
aX2i
ra~lli-,
(2-5)
then (ff", ff) is a vector space. lf ff =~, (~", ~) is called the n-dimensional
real vector space; if ff = ic, (e, ic) is caHed the n-dimensional complex vector
space; if ff = ~(s), (W(s), ~(s)) is called the n-dimensional rational vector
~~
Example 6
In the two-dimensiona
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
asi
~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.
~2'
n-l
"-1
n-l
lXsi
i=O
11
f3iSi =
i=O
IX
Ct~
lXiS)
(IX
+ f3)S
;=0
~t~
(IXIX)S
1t is easy to verify that (rrln[S], rrl) is a linear space. Note that (rrln[s], rrl(s)) is
not a linear space.
I
Example 5
as columns
(2-4)
(2-5)
Let (fL, .?") be a linear space and let JI be a subset of fL. Then (JI, .?") is said to
be a subspace of (gr, .?") if under the operations of (fL, .?), JI itself forms a vector
space over !/F
In the two-dimensional real vector space (1R 2 , IR), every straight line passing
through the origin is a subspace of(rrl 2 , rrl). That is, the set
Example 7
tions wth real coefficients;
.1 coefficients.
The real vector space (rrl n, ~) is a subspace of the vector space (iC", rrl).
12
2-3
Every geometric plane has two coordinate axes, which are mutually perpendic
ular and of the same scale. The reason for having a coordinate system is to
have sorne reference or standard to specify a point or vector in the planeo In
this section, we will extend this concept of coordinate to general linear spaces.
In linear spaces a coordinate system is called a basis. The basis vectors are
generally not perpendicular to each other and have different scales. Before
proceeding, we need the concept of linear independence of vectors.
Definition 2-4
A set of vectors Xl' Xz, ... ,XII in a linear space over a field ff, (Er, ff), is said to
be linearly dependent if and only if there exist scalars (Xl' (Xz, . .. , (XII in ff, not
all zero, such that
Definition 2-4'
(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, ... ,
(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
(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
dependence of the set of vectors.
Example 1
Consder the set of vectors Xl' Xz, ... , XII in which Xl =0. This set of vectors
is always linearly dependent, because we may choose (Xl = 1, (Xz = 0, (Xz = 0, ... ,
Cl. = O, and Equation (2-6) holds.
id!
H
Example 2
Consider the set of vector Xl which consists of only one vector. The set of
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
is linearly dependent
Indeed, if we choose
x2
>
",
Xi
are HrJea
as a linear combinatio
every one of them can
Definition 2-5
(2-7)
then the linear independence of a set of vectors can also be stated in the fol
lowing definition.
'esentations
e mutually perpendic
.ordinate system is to
:ctor in the planeo In
general linear spaces.
The basis vectors are
ferent scales. Before
)f vectors.
Definition 2-4'
A set of vectors x , x 2, ... , x" in (fll', 9') is said to be linearly independent if and
only if the equation
[Xl
= 0,
Ci. 2
= 0, ... ,
Iill
~"Ja
(2-7)
X2
X,,]ex=
s~ 1]
x ~
= 1,
13
s+2
]
(S+1:S+3)
X2~
1
s+2
[
s+3
and
Ci. 2
s+3
s+2
=-
then IXX +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)).
X,
Definition 2-5
The maximal number of linearly independent vectors in a linear space (fl, ff)
is called the dimension of the linear space (fil, ff).
14
Example 3
Then by subtracting (
O=(p
are clearly elements ofthe function space. This set of functions {t n , n = 1,2, ...}
is linearly independent, because there exist no real constants, a/s, not aH zero,
such that
L ai=O
i= 1
There are infinitely many of these functions; therefore, the dimension of this
space is infinity.
We assume that aH the linear spaces we shaH encounter are of finite dimen
sions unless stated otherwise.
Definition 2-6
where
~ctor
P = [{Jl> {Jz,
P can be cons
one-to-one correspond
the same dimensional,
Theorem 2-1
Definition 2-7
In an n-dimensional
then every vector x iJ
called the re presentatl
Proof
Example 4
Let el> e z, ... , en be any n linearly independent vectors in f!{, and let x be an
arbitrary vector in!!( Then the set of n + 1 vectors x, el> e z, ... ,en is linearly
dependent (since, by the definition of dimension, n is the maximum number of
linearly independent vectors we can have in the space). Consequently. there
exist a o, al, ... , an in ff, not all zero, such that
aoX+alel +azez + ... +anen=O
~-
(2-8)
which, together with the linear independence assumption of el' ez, ... , en,
implies that al = O, az == O, ... , a = O. This contradicts the assumption that
not aH ao, al>' .. , a n are zero. Ir we define {Ji ~ - aJa o, for i = 1, 2, ... , n, then
(2-8) becomes
ll
(2-10)
piecewise continuous
:pace is the one which
, with - CIJ < t < CIJ,
15
of el> ez, . .. , en Now we show that this combination is unique. Suppose there
is another linear combination, say
(2-11 )
f![,
Q.E.D.
(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
one-to-one correspondence between any n-dimensional vector space (f![, .?") and
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
and let x be an
L, ez, ... , el! is linearly
maximum number of
Consequently, there
n
f![,
(2-8)
(2-9)
i a linear combination
16
Table
2-1
'"
~s
Bases
[~J
[~J [~J
[~J
[~J
any set of two linearly independent vectors forms a basis. Observe that we
have not only the freedom in choosing the directions of the basis vectors (as
long as they do not lie in the same line) but also the magnitude (scale) of these
vectors. Therefore, given a vector in (IR 2, IR), for different bases we have different
representations of the same vector. For example, the representations of the
vector b in Figure 2-1 with respect to the basis {e,e 2} and the basis {e,e 2}
are, respectively, [1 3]' and [-1 2]' (where the "prime" symbol denotes the
transpose). We summarize the representations ofthe vectors b, el, e 2, el, and e 2
with respect to the bases {el, e 2}, {el, e2} in Table 2-1.
Example 5
Consider the linear space (1R 4 [s], IR), where 1R 4 [s] is the set oC all real poly
nomials of degree less than 4 and with indeterminate s. Let e = S3, e 2 = S2,
e 3 =s, and e 4 = 1. Clearly, the vectors e, i = 1,2,3,4, are linearly independent
and qualify as basis vectors. With this set of basis vectors, the vector x =
3s 3 + 2s 2 - 2s + 10 can be written as
1
O
O
O
O
as the basis of(lRn, IR),
can be interpreted as
the basis {O, 2 " . , JI
representation and th,
Changa of basis. ,
representations with ]
x: :::
+::' ~,~
:,;OW" -,') +
x=[
In order to derivf
infOI:mation of the rel
5 3
13]'. I
5
17
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
01 =
1
O
O
O
1
O ,
, Oz=
O
O
nn-l =
O
O
O
O
O
, o n
O
O
O
(2-13)
O
1
1
O
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
the basis {o, 0z, ... , on}, because with respect fo this particular set of bases, the
representation and the vector itself are identical; that is,
(2-14 )
s-1)+13-1
Change of basis. We have shown that a vector x in (~, ff) has different
representations with respect to different bases. it is naturalto ask Vvhi: the
relationships are between these different representations of the same vector.
In this subsection, this problem will be studied.
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
x=[e
e z ._.
en]~=[e
ez
en]p
(2-15)
Hs [3 5 3 13]'. I
:ctors and representa
der the n-dimensional
cn]jl.
However,
18
basis {e b ez, ... ,en}, or the inforrnation of the representations of el' for i =
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,
i= 1,2, ... , n
where E ~ [el
tion, we write
ez
en], Pi ~ [Pli
[el e z ...
PZi
en] = [EpI
...
PnJ'
Epz
...
Example 6
(2-16)
EPn]
and
(2-17)
e z ...
en] = E[PI
pz
...
P/I]
Clearly, we have PQ .
en]
~ [el
ez
PII
PIZ
Pfl
pzz
PI]
PZn
pnl P/lZ
;/ln
en]P
(2-18)
e z ...
en]Pp = [el
(2-19)
Since the representation of x with respect to the basis {e b ez, ... ,en} is uniqtie,
(2-19) irnplies
p=Pp
(2-20)
where
(2-21 )
linear Opel
2-4
where
(2-23)
or
P = Q - 1
(2-24 )
y =Lx
-1
(a)
19
Example 6
(~4[SJ, IR)
in Exarnple 5. lt can be
(2-16)
2, ... , n
and
(2-17)
l- ~
O
1
O
-1
1
O -1
~J~l"
""
'4]Q
PI]
rn
=Pl ~lJ::
ll~J ~~ U
P2n
Pnn
(2-18)
2-4
PJ~ rl~~J~l
(2-19)
, e2 , . . . , en} is unique,
(2-20)
(2-21 )
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
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
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
1{
(2-22)
(2-23)
--"-0f----\---r----+--------'L----.X
y =Lx
-1
(a)
(b)
Figure 2-2 Examples in which (a) the curve represents a function and.(b) the curve does
not represent a functlon.
20
domain ofthe function shown in Figure 2-2(a) is the positive realline, the range
of the function is the set [ -1, 1], which is a subset of the entire real Hne J/.
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
mappings, or linear transformations. The sets associated with linear operators
are required to be linear spaces over the same field, say (gr, ff) and (J/, ff). A
linear operator is denoted by L:( gr, ff) ---> (J/, ff). In words, L maps (gr, ff)
into (J/, ff).
Definition 2-8
A function L that maps (gr, ff) into (J/, ff) is said to be a linear operator if
and only if
L((;(x + (;(zx z ) = (;(Lx + (;(zLx z
for any vectors Xl' X z in grand any scalars (;(, (;(z in ff.
Note that the vectors Lx and Lx z are elements of 0Jj. The reason for
requiring that J/ be defined over the same field as gr is to ensure that (;(Lx
and (;(zLx z be defined.
Example 1
is a linear transforma1
all equal to (O/t, IR).
Matrix representa
Example 2
Let O/t be the set of all real-valued piecewise continuous functions defined over
[O, T] for sorne finite T > O. lt is clear that (O/t, IR) is a linear space whose
dimension is infinity (see Example 3, Section 2-3). Let 9 be a continuous
function defined over [0, T]. Then the transformation
y(t)=
g(l-r)u(r)dr
(2-25)
.-------- -
~s
-3
------------
L[x
--Y2 -0.5
xI
1.5
Xz
21
Je a linear operator if
is a linear transformation. The spaces (gr, ff) and (qy, ff) of this example are
aH equal to (Ii/l, IR).
I
Matrix representations of a linear operator. We see fram the aboye
two examples that the spaces (gr, ff) and (qy, ff) on which a linear operator is
defined may be of finite or infinite dimensiono We show in the foHowing that
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
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
Let (gr, ff) and (qy, ff) be n- and m-dimensional vector spaces, respectively,
over the same field. Let Xl' X2, ... ,XII be a set of linearly independent vectors
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 }.
Proof
Let X be an arbitrary vector in gr. Since Xl' X2, ... , X" are linearly independent,
the set of vectors qualifies as a basis. Consequently, the vector X can be ex
pressed uniquely as CXIX l +CX2X2 + ... +CX"x1l (Theorem 2-1). By the linearity
of L, we have
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
.i = 1, 2, ... ,n. This praves the first part of the theorem.
Let the representation of y with respect to {1Ul 1" lffi2"
' , . ,lml
be
(2-25)
a,:
n.] a
']
amI
i= 1,2, ... , n
(2-26)
22
Note that the elements of A are in the field 9' and the ith column of A is the
representation of y i with respect to the basis of 1fJ/. With respect to the basis
{x,x 2 , ... ,x n} of (gr, 9') and the basis {u,u 2 , ... ,u m} of (1fJ/, 9'), the linear
operator y = Lx can be written as
[u l
U2
...
umJJI= L[x l
xnJex
X2
(2-28)
[u l
u2
...
umJJI= [u l
U2
...
umJAex
or
(2-29)
Two matrices A a
matrix P satisfying (:
JI=Aex
(2-30)
similarity transformat.
to different bases) oftJ
Hence we conclude that if the bases of (gr, 9') and (1fJ/, 9') are chosen, the
operator can be represented by a matrix with coefficients in 9'.
Q.E.D.
Example 3
We see from (2-30) that the matrix A gives the relation between the repre
sentations ex and p, not the vectors x and y. We also see that A depends on the
basis chosen. Hence, for different bases, we have different representations of
the same operator.
We study in the following an important subclass of linear operators that
maps a linear space (g[, 9') into itself; that is, L:(fI, 9')-+(gr, 9').. In this
case, the same basis is always used for these two linear spaces. If a basis of g[,
say {el e2, ... , en}, is chosen, then a matrix representation A of the linear
operator L can be obtained by using Theorem 2-2. For a different basis
{e1, e2, ... , en}, we shall obtain a different representation of the same operator
L. We shall now establish the relationship between A and . Consider
Figure 2-4; x is an arbitrary vector in g[; ex and eX are the representations of x
with respect to the basis {e, el>' .. ,en} and the basis {e, e2' ... ,en}, respec
tively. Since the vector y = Lx is in the same space, its representations with
respect to the bases chosen, say JI and p, can also be found. The matrix repre
sentations A and can be computed by using Theorem 2-2. The relationships
Independent
of basis
Basis
[el
x ----+y(=Lx)
[el
The representation of
ith
ith column:
]
column:
Representation of
Representation of
A = ~ wit? respect to ,P = e~ Wi!h respect to
{el' e 2, ... , en}
[ the basls
{el' e b .. , en}
'th
Basis
column:
]
llh coiumn:
Representation of
Representation of
= Le with respect to ,Q = e with respect to
the basis
{el' e 2, ... , en}
{el' ez, .. ,en}
or
LXI = [Xl
23
between IX and (j and between p and ~ have been established in (2-20); they are
related by (j = Prx and ~ = pp, where Pis a nonsingular matrix with coefficients
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
~= A(j
(2-28)
PAIX= PAP-(j
A = P AP -
or
(2-29)
lm]AIX
~= Pp=
and
Q - AQ
(2-31a)
A = P - AP = QAQ -
(2-31b)
where Q ~ P- .
Two matrices A and A are said to be similar if there exists a nonsingular
matrix P satisfying (2-31). The transformation defined in (2-31) is caBed a
similarity transformation. Clearly, al! the matrix representations (with respeet
to different bases) of the same operator are similar.
1
implies that
(2-30)
~, ff) are chosen, the
Q.E.D.
; in ff.
Example 3
,P
=[
Y=Lx=[x 1
xz][~J
and
Yz=Lxz=[x
X Z][
-~J
The representation of x 3 is
1.5J
[ 0.5
It is easy to verify that the representation af)' ~ with respect te (~( 1
~
ith column:
Representation of
e with respect to
O
[1
or
-)"(2]
is equal to
-IJ_[1.5J=[-0.5J
O 0.5
1.5
Y3=[X
x z] [ -0.5J
1.5
If, instead of {x, XZ}, we choose {Xl' X3} as a basis, then from Figure 2-3,
"th column:
Representation of
, Q = e, with respect to
he same operator.
Y=LX~[~~.
X3]
[-~J
and
Y3=Lx3 =[x
X3J[-~J
24
The reader is advised to find the P matrix for this example and verify A =
PAP- 1 .
In matrix theory, a matrix is introduced as an array ofnumbers. With the
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
as a linear operator that maps (ff n , ff) into itself. 7 The matrix A is independent
of the basis chosen for (ff n , ff). However, if the set of the vectors 01' 02, ... ,on
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
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
tion A. From Figure 2-5, we see that the matrix representation A may be
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.
and
The last equation is ob
respect to the basis {b,
Example 4
Ab=
-~ -~l
L=A= [ -~ ~
b~ m
Let
1
;::xample 5
We extend Example 4
with real coefficients.
Ab, ... , A"- 1b are Hne;
1X1An-lb(see Section 2
{b,Ab, ... ,An- 1b} is
L~A
lndependent of basis
Basis [o
Basis [q
q2 .. ,
'"
~.
n]
A
~
IJI
IQ
Qr
qn]
,JI
'Ji
rroumn
= Representation of Aq
with respeet to
{q, ih, ... , qn}
=Q-AQ
Q=[q q2
-iJ
qn] =p-
25
Then 8
A 3 b = 5A 2b -15Ab + 17b
(2-32)
Since the set of vectors b, Ab, and A2b are linearly independent, it qualifies as a
basis. We compute now the representation of A with respect to this basis.
It is clear that
A(b)~ [b
A(Ab)~ [b
A(A'b)~ [b
and
Ab
A'bJ
Ab
A'bJ
Ab
A'bJ
[!J
m
[-:~J
The last equation is obtained from (2-32). Hence the representation of A with
respect to the basis {b, Ab, A2b} is
0:0:
17]
A= 1 ;0: -15
[o 1
5
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 = Q -1 AQ, or
more easily, QA = AQ, to check our resulto The reader is asked to verify
[~
-1
O
1
-~l[~ ~ _~~l
-iJ
O 1
~J
= [_;
~
3
-1
O
1
-~]
-3
Example 5
~~ Aql
~} J
of
1 O
O -(;(n-I
O 1
O -(;("-2
(2-33)
A=
O O
O O
8
A 2 ~ AA, A 3 ~ AAA.
-(;(2
-(;(1
26
D&finition 2-9
QA=AQ
Since Q= [ql
q2
[ql
q2
qnJ A = [Aql
Aq2
AqnJ
(2-34)
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}.
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
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.
2-5
(2-35)
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
can be written in matrix form as
Ax=y
where
A~
(2-36)
a m1
a m2
Theorem 2-3
of A, it is easy to shO\
(X1X 1 + (X2 X 2 is an elem
( ffm, ff) (see the rema
+alnxn=Yl
+a2nXn = Y2
a11x 1 +a12 x 2+
a21 x l + a22x 2 +
n
(ff ,
27
se is said to be in the
tant1Yarise in this tex t.
-IAQ as
operator which maps (ff", ff) into (ff m , ff). RecaB that the linear space
(ff", ff) that undergoes transforrnation is caBed the domain of A.
D&finition 2-9
fJJI (A)= {aH the elements y of (ff m , ff) for which there exists at least one
vector x in (ff", ff) such that y = Ax}
!II
the representation of
ection: Since a linear
lOO se one set of basis
The answer is affirma
ar algebraic equations.
Theorem 2-3
(2-35 )
i Y is an m x 1 vector.
~ger than, equal to, or
in regard to this set of
le n umber of sol utions.
. in Equation (2-36) are
In A and y under which
s exist, then the second
Jendent vectors x such
ank and the nullity of
he matrix A as a linear
az
(2-37)
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
(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.
Definition 2-10
Example 1
28
(2-38)
where e is a very small number, say, 10 -1 0. This e may arise from the given
data (assuming no rounding errors) or from rounding errors. lf e arises from
rounding errors, we should consider e as zero, and the matrix has rank 1. lf
e is due to the given data, we cannot consider it as a zero, and the matrix has
rank 2. To determine what value is small enough to be considered as a zero is a
complicated problem in computer computations. For problems encountered
on matrix computation, the reader is referred to References S181, S182, S200,
and S212.
In matrix theory, the ranle of a matrix is defined as the largest order of aH
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
minor of order higher than k vanishes. This definition and Definition 2-10
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
ofthe matrix is different from zero; or correspondingly, a matrix is nonsingular
With the concepts of range space and rank, we are ready to study the
existence problem of the solutions ofAx = y.
lo It follows immedia
vector y is not an e
is said to be incons,
2. The rank of A, p(J
is a subspace of (
(!F m , ff), then for a
lf p(A) < m, there e;
~(A), for which the
Theorem 2-4
ms of aH the columns
ltions f the first two
i A are linearly depen
quence of elementary
lroperty that the rank
cations of elementary
o the upper triangular
ler of nonzero rows in
umber of linear inde
of independent rows.
mns
(2-38)
tmputers, however, is
)n digital computers,
s. Suppose a matrix,
29
1. Given A and given a vector y in (ff m , ff), there exists a vector x such that
Ax = Y if and only if the vector y is an element of P4(A), or equivalently,
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.
Proof
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
null space of A. See Problem 2-51.
30
be arbitrarily chosen.
vectors in Y(A). Hen
theorem; its formal pn
Theorem 2-5
Example 2
Let A be an m x n mat
h[! i i
~ =iJ
Corollary 2-5
which maps (~5,~) into (~3, ~). lt is easy to check that the last three columns
of A are linearly dependent on the first two col umns of A. Hence the rank of
A, p(A), is equal to 2. Let x = [Xl Xz X3 X4 xs]'. Then
(2-39)
Let A, B be q x n and
XI +X3 +xs =0
xz+X3+ 2x 4- XS=0
Notethat the number of equations is equal to the rank of A, p(A). The solution
x ofAx = Ohas five components but is governed by only two equations; hence
three of the five components can be arbitrarily assigned. Let X3 = 1, X 4 = O,
X5 = O; then XI = - 1 and xz = - 1. Let X 3 = O, X 4 = 1, X s = O; then Xl = Oand
Xz = -2. Let X3 = O, X 4 = O, Xs = 1; then XI = -1 and Xz = 1. lt is clear that
the three vectors
p(A)
Proof
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
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
()?,
J)
n Ax = Ohas k linearly
1
31
[=iJ
X,)[i]
x,
(2-39)
:arly independent, we
Id only if
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
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,
det (A) = 0, where det stands for the determinant.
We introduce three useful theorems to conclude this section.
Theorem 2-6 (Sylvester's inequality)
Proof
The composite matrix AB can be considered as two linear transformations
applied successively to (ffP, ff) as shown in Figure 2-6. Since the domain
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) =
r
__ L-_'J7(_B)---.J--l-j IL--'J7(_A)~ 1
--T-- --
,iB)
= O must be a linear
vectors form a basis of
1(
I
-T---,t-IR(AB'L
I.L --r- --- .
------~--
R{B)
(5P,5)--B-----..-
(]",5)
R{A)
32
p(B) - d, where d is the dimension of the intersection of ~(B) and .K(A).10 The
dimension oL/nA) is n - p(A); hence, d S n - p(A). Consequently, p(AB) 2:
p(B)-n+p(A).
Q.E.D.
assumption of p(A
ACl=[!31
f32
and
Hence Cl*A*ACl = O
which, in turn, imp
conclude that p(A *
2. This part can be Sil
using the fact p(A)
and
p(DA) = p(A)
In words, the rank of a matrix will not change after the pre- or postmulti
plication of a nonsingular matrix. Because of this property, gaussian elimina
tion or the row-searching algorithm discussed in Appendix A can be used to
compute the rank of a matrix.
Theorem 2-8
Let A be an m x n matrix with coefficients in a field, and let A * be the complex
conjugate transpose of A. Then
2-6 Eigenvectol
Form Represent.
With the background
posed at the end of Sec
that map (en, iC) into i1
to any operator that r
The reason for restrict
immediately.
Let A be an n x n n
consider A as a linear
Definition 2-12
Let A be a linear oper
is ca11ed an eigenvalue
Ax = lx. Anynonzerc
associated with the eig
In order to find an
10
~(B)
33
a.*A* = [,B!
and
.8!
.,. .8::'t]
2
a.*A*Aor: = 1.81\2 +1.821 + ... +1.8mI 2
n)= p(A)
<\)
With the background of Section 2-5, we are now ready to study the problem
posed at the end of Section 2-4. We discuss in this section only linear operators
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
consider A as a linear operator that maps (en, C) into (en, C).
Definition 2-12
m matrix. In order to
rem wil\ be proved by
t that ir Aa. = (} implies
, and peA) = n, where n
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
associated with the eigenvalue A. 11
I
In order to find an eigenvalue of A, we write Ax = A.x as
where I is the unit matrix of order n. We see that for any fixed), in e, Equation
(2-40) is a set of homogeneous linear equations. The matrix (A - Al) is an
n x n square matrix. From Corollary 2-5, we know that Equation (2-40) has a
nontrivial solution if and only if det (A - Al) = O. lt follows that a sea/ar A is
an eigenva/ue of A if and on/y if it is a so/ution of ~(A) ~ det (Al - A) = O. ~(A)
is a polynomial of degn~e n in Aand is called the eharaeteristie po/ynomia/ of A.
Since ~(A) is of degree n, the n x n matrix A has n eigenvalues (not necessarily
all distinct).
11
It is also caBed a right eigenveor of A.. Ifa 1 x n nonzero vector y exists'such thi yA = J.y, then
y is caBed a left eigenveclOr of A associated with J.
34
Example 1
(A-.
-1J
(2-41 )
-1
which maps (lR z, IR) into itself. We like to check whether Definition 2-12 can
be modified and applied to a linear operator that maps (IR", IR) into (IR", IR).
A modified version of Definition 2-12 reads as a scalar }, in IR is an eigenvalue of
A if there exists a nonzero vector x such that Ax = Ax. Clearly A is an eigenvalue
of A if and only if it is a solution of det (Al - A) = O. Now
det(AI-A)=det [ ..1.-1
-2
1 ]
..1.+1
=A z
Since
and
the left-hand side of (2
CI.
+1
Let Al' }'z, ... , A" be the eigenvalues of A, and let Vi be an eigenvector of A
associated with A, for i = 1,2, ... , n; that is, Av = Av i' We shaH use the set of
vectors {v , V z, ... , v,,} as a basis of (e", iC). In order to do so, we have to show
that the set is linearly independent and qualifies as a basis.
Since
Theorem 2-9
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
independent (over ic).
Proof
We prove the theorem by contradiction. Suppose VI' V z, ... , v.. are linearly
dependent; then there exist CI., Cl.z, ,!X" (not aH zero) in ic such that
Cl.IV I
We assume
Cl. 1
fO.
If
Cl. 1
(2-42)
al
fO.
we have
We conclude that if
into itself are aH distin
the operator A has a d
the diagonal.
Example 2
Consider
)w
Since
(A-Ajl)v=(A-A)V
and
(A - A;I)v = O
Ct
35
(2-43)
<XiV) = O
ifj+i
.z + 1
n (Al -A)V l =0
11
<Xl
i= Z
implies <Xl = O. This is a contradiction. Thus, the set of vectors {v, Vz, ... , VII}
is linearly independent and qualifies as a basis.
Q.E.D.
Let A be the representation of A with respect to the basis {v, Vz, ... , VII}.
Recall from Figure 2-5 that the ith column of A is the representation of Av =
Aviwithrespectto{vl,vZ, ... ,vlI}-thatis,[O
O Ai O
O]',where
A is located at the ith entry. Hence the representation of A with respect to
{v,vz, ... ,v lI } is
oo.
O
Az O
O ..
.
O A3
o
o.
.
O.]
O
(2-44)
..
O .
II
et Vi be an eigenvector
'1' vz,, VII} is linearly
From
Q = [v 1 VZ .. . VII]
Since
AQ=A[v l "'z ... VII] = [Av l Av z
= [A1V l AzVz ... AllvlI]=QA
we have
...
Av n]
A=Q-1AQ
Consider
-IJ
-1
36
(A-AII)VI=[l~i
Example 4
Consider
-~~J[~::J=o
The eigenvalues of A al
with Al = 1 can be fom
A=[iO O.J
-1
Example 3
Consider
(A-A'I)v~[~ ~ -nv~o
(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
A= [O1 O0J
1 O
O O 2
Al
O
O
Al
O
O: O
1 :O
Al: O
O
O
O
o---6--o-~-il-:
1
L __
O: A2
'"
le eigenvalues of A are
be obtained by solving
37
Example 4
Consider
J=o
(2-46)
, - i.
formation.
O1 2J
[O O 1
:d eigenvalues, it is not
h repeated eigenvalues.
eigenvectors associated
equations:
~eneous
Snce the matrix (A - ..1. 11) has rank 2, the null space of (A - ,111) has dimension 1.
ConsequentIy, we can find only one linearly independent eigenvector, say
VI = [1
O O]', associated with Al =..1. 2 = 1. An eigenvector associated with
. 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
From this example, we see that if an n x n matrix A has repeated eigen
values, it is not always possible to find n linearly independent eigenvectors.
Consequently, the A cannot be transformed into a diagonal formo However,
it is possible to find a special set of basis vectors so that the new representation
is almost a diagonal form, called aJordan canonical formo The form has the
eigenvalues of A on the diagonal and either O or 1 on the superdiagonal. For
example, if A has an eigenvalue Al with multiplicity 4 and an eigenvalue . 1. 2
with multiplicity 1, then the new representation will assume one of the following
forms.
(2-45)
Al : O
linearly independent
ry 2-5). Clearly, '"'1=
lependent eigenvectors
ed with ,13 = 2 can be
; {v, v2 , v3 } is linearly
m of A with respect to
{O
o--:-i
L
-:
1 .1.
__
O
O
O
O
O O
O: Al' O
'
l~ ~ :_~---~'-:-tJ
- - - ..! - - -.
O
O
O
O
Al
O
O
_ _ _ 1---:
O:
Al
:0
:0
O
O
Al : O
O
O
O---O--O-~-il-:
A2
O
1 ,
1_ _ _
O:!e l
L __
,
O , ,12
A1 1 : O ,O O
)'1:L O
O , O
O
______
, O O: l 1 ,O ,
O
O
)'1
O
O
Al
O
1
Al
O O
O O
1 ,
O
Al : O
(2-47)
0- - -O -- 0- - -O -:-l;
38
O O
O O
O O O
O O O
O O O
1 O
A 1
O A
A. Define
(2-48)
and
with the same eigenvalue on the main diagonal and l's on the diagonal just
aboye the main diagonal. A matrix of this form is caBed al ordan block associ
ated with A. A matrix is said to be in the lordan canonicalform, or the lordan
form, if its principal diagonal consists of Jordan blocks and the remaining
elements are zeros. The fourth matrix in (2-47) has two Jordan blocks associated
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
of its Jordan blocks are of order 1.
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
concepts and results; hence it will be extensively used in the remainder of this
chapter.
Derivation of a Jordan-form representation. 12 In this subsection,
we discuss how to find a set of basis vectors so that the representation of A with
respect to this set of basis vectors is in aJordan formo The basis vectors to be
used are called the generalized eigenvectors.
(A-AI)i-l,
Defihition 2-13
A vector v is said to be a generalized eigenvector of grade k of A associated with
A if and only if 13
(A -AI)kv = O
and
(A -AIt-lv +0
p(A-AI)2=4
p(A-AI)3=3
PlA-AI)4",i
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.
12
Vo =0
=3
v2 =6
v3 =7
v4 =8
VI
39
A.. Define
Vk
Vk- l ~ (A-AI)v=(A-AI)v k
Vk- 2 ~ (A-AI)2v =(A-AI)vk_'
(2-48)
(2-49)
and
This set of vectors {V, v2, ... , vk} is called a chain of generalized eigenvectors
of length k.
Let .;Vi denote the null space of (A - AI)i, that is, .;Vi consists of all x such
that (A-AI)ix=O. It is clear that ifx is in .;Vi' then it is in ';vi+l' Hence.;Vi
isasubspaceof .;Vi+"denotedas ';viC .;Vi +l' Cleariy,thevdefinedin Defini
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
defined in (2-49) is in .;Vi but not in ';vi-l' Indeed, we have
1 Jordan
jan-form representa
e10ping a number of
the remainder of this
are as shown in Table 2-2. The nullity Vi is the dimension ofthe null space ';vi'
In this subsection,
,resentation of A with
le basis vectors to be
, of A associated with
B 4 u =0
where El ~ A -,ll From this
= O and v 1= O, which
eralized eigenvector"
j
Illi,
and
B 3 u 1=0
I
s suggested that the reader
esentaon.
1'0
=0
VI
=3
=6
V, =7
1'4=8
1'2
. w2 =w.
v2 =v,
=Bw
v, =Bv
WI
No. of independent
vectors in Xi but
not in _I
TWO chains with
}
length 2
One chain with
length 4
)(2--~-
40
eigenvectors as
(2-50)
BvjoO
BwjoO
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. Sice {;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}.
Now we show that the generalized eigenvectors {u i , i = 1, 2,3,4; vj' Wj,j = 1,
2} are linearly independent. Consider
Example 4 (Continued
Consider
(2-51 )
14
41
which implies, because ofthe linear independenceof {uzo Vz, Wz}, Cz = C6 = Cs =0.
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.
(2-50)
;: linearly independent
a vector; therefore we
; but not in j{ z. Con
: Vz - v1 = 3, there are
50) is one of them, we
early independent and
Theorem 2 -11
The generalized eigenvectors of A associated with different eigenvalues are
linearly independent.
I
This theorem can be proved as in Theorem 2-10 by applying repetitively
(A - AI)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
assoqated with other eigenvalues. The first four columns ofthe new representa
tion A are the representations of Au, i = 1, 2, 3, 4, with respect to {u, u z, u 3 , u 4 ,
Vlo Vz, W1, Wz, X, x}. Because (A - AI)u1 = O, (A -AI)1z = u 1, (A - AI)u 3 =u z, and
(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
1eigenvectors oflengfh
A=
fu
m lo v lo W} is linearly
dependent, then there
+ C3W1 =0. However,
A 1 O 0:0
O A 1 0:0
O O A 1:O
O O O ,,1.:0
O O O
O
O O O
O O O
O-----O-:-~--(: O
O
,
,
O O O 0:0 A: O
O O O O - 0- -6: ),- - i -:,
O O O O O 0:0
A:
,
(2-52)
_ _ _ _ _ _1
at y =O.Since {uz, v, w}
c=O,i=1,2,3. This
ldent, so is {U l , '1 1, w}.
= 1, 2, 3,4; Vj, Wj,j = 1,
+cswz=O
Example 4 (Continued)
Consider
(2-51 )
14
l.
This number is caBed the geometric multiplicity of ), in Reference 86. In oLher words, geometric
multiplicity is the number of lordan block S, and the (algebraic) multiplicity is the sum of the
orders of aH lordan blocks associated with .l..
42
and
B@(A-A'l)~[~
O
O
(A~A,Il'~[~
O
O
Example 5
:]
m~
:] ~[~ ]
1
O
O
O
O
det (A -AI)=
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
Av l' Av 2 , and AV 3 with respect to the basis {v 1 , v2 , v 3 } are, respectively,
f:::,.
(A -21) =
(A-21)2 =
r:
(A-21)'
A=[~O 0:2
~-~-~J
~l
(2-53)
__
where
Q =[v 1
v2 v,]
~[i
15
O
1
~]
43
Example 5
, O5]
O 3
O
O
O
O
1
1
-1 -1
1
1
1
1
(2-54 )
O 1
linearly independent.
at of [v 1 V 2 v3 ]. Ir
ith column of the new
to the basis {v 1 , v2 , v3 }.
le representations of
re, respectively,
O
O
1
-1
1
-1
p(A-21)=4
v1 =6-4=2
O
O
O
O
-2
2
p(A-2If =2
V2 =4
O
O
O
(A -21)2 =
O
O
O
O
O
O
O
O
O
2
2
O
O
O
O
2
2
O
O
O
O
r~
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O -4
4 -4
O
(A-2Ij'
(2-53)
~ l~
O
O
O
O
O
O
O
2
-2
~l
~J
p(A -21iY =
v3 =5
rmation
[5
~J = det A det e
where A and e are square matrices, not necessarily or the same order.
44
where
UI
2
2
O
BZu=
O
O
1
-1
O
U z ~ Bu =
O
O
O
O
O
1
u 3 ~ u=
O
O
O
VI
O
O
1
-1
1
1
O
O
2
~ Bv=
-2
O
O
-1
1
O
(A A.zI)w =
O
O
O
O
O
O
O
1
1
O
O
-1 -1
O
O
2
1 w=O
1
O
O
2 -1 -1
O
O
1
1
1
O
O
1
= [O
-1]' is a solution
In this example,
{w, V z, V, U 3, U z, u};
2-7
Clearly,
4. With respect to the basis {u I , u z, u 3 , VI, vz, w}, A has the following Jordan
form representation:
2 1 0:0 O O
021:0 O O
O O 2:0 O O
A= 0- -6 - -6 ~ 2- - -: O
O O 0:0 2:0
O O O - O--6 - 0
(2-55)
Polynomials of a s
(C", C) into (1[:",1[:). If
or
Functions (
and
where
a vector. Define
Uz
u3
v1
Vz
w]
2 , 1 , O' O: O: O
,
2 : -1 , O , O', O: O
,
,
O
1 , 2: 1
O , O
,
O
O
O , O : -2 :, -1
O,
1
O : O
1
O
1 -1
O
O
O
O
O
O
~ u= 1
O
O
O
Becauseofvz -V 1 =2,
10t in j( l' The vector
:pendent of U z and has
lily verified that v =
efine
{Ul>
associated with A. = 2.
:t w be an eigenvector
o
O
1
w=o
-1
A=
O
O
1
1
1
1
l
Q=[U 1
45
(2-56)
1
1
2-7
ution.
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 )
and
AO
(k terms) .'
(2:57a)
(2-57b)
46
polynomial of A is
We have shown in the preceding section that every square matrix A that
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
where the
we have
f(A)
Qf(A)Q-I
or
(2-58)
[f(OAd
Definition 2-15
index of A in A.
(2-59)
x n mat
f(A)
The multiplicity (
the matrix in (2-52),
nz = 1; for the matr
n'::;; ni.
O ]
f(A z )
(2-60)
Theorem 2-12
[A1O
Definition 2-14
The minimal polynomial of a matrix A is the monic polynomial 16 if(A) of least
degree such that if(A) =0.
I
Note that the O in if(A) =0 is an n x n square matrix whose entries are all
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
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
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.
. 16.\
A consists of r Jorda
A=di
47
polynomial of A is
m
~(A) ~
n (A - AJ"
det (Al - A) =
(2-61 )
= 1
.(2-62 )
(2-58)
Definition 2-15
The largest order of the Jordan blocks associated with A in A is caBed the
index of A in A.
11
hat if
(2-59)
(2-60)
Theorem 2-12
The minimal.polynomial of A is
n (A - A)
m
lj(A)
= 1
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\'::
A consists of r Jordan blocks associated with A. Then
A with multiplcities
hat the characteristic
(Ail - Alt'
.=
t power is 1.
, is.
For example,
..
f.
?.
(Aiz ~ A;I)"'
..
(2-63)
48
(Aij - AI)
(nij x ni)
nj,
then we have
=[~ ~
... O]
O
...
(2-64a)
O O O ... 1
O O O
O
(Aij-A;I?=
(Aj - AI)nj- =
O O 1 O
O O O 1
O
O
O O O O
O O O O
O O O O
1
O
O
[~ ~ ~
O O O
and
(Aij - AJ)k = O
~
O
oll
Corollary2-12
(2-64b)
(C,
(2-64c)
(2-64d)
The Cayley-Ham
Theorem 2-12 (see Pr
The reason for in1
in the following theoJ
Theorem 2-13
Let A, Az, . .. , Am be
Let f and 9 be two pe
1. feA) = g(A).
2. Either f =h1jJ +f
A, and h and h 2 a
3.
pe)(A) = g('
where j(l)(A;)
~ d~
{
Proof
The equivalence of sta
Statements 2 and 3 ar
rn
TI (A _A)
Q.E.D.
i= 1
Example 1
The matrices
r~ ~ i ~J r~ ~ i ~J r~ ~ I~J
al! have the same characteristic polynomial i1(A) = (A - 3)3(A -1); however,
In order to apply 1
The minimal polynol
form or by direct com
Therefore it is desirab
polynomial can be av
Corollary 2-13
-_._--_....::_..__ .:..~.~,----------"---~----_.:..._=--.=--.....:....-~=--_:...._-_:...-::~_.,---...:_-,--,~_:._-----
49
they have, respectively, (A - 3)(A - 1), (A - 3f(A -1), and (A - W(A -1) as
minimal polynomials.
I
(2-64a)
(2-64b)
(2-64c)
(2-64d)
Theorem 2-13
:J
k 2: nij
in A, or equivalently,
j=1,2, ... ,r. Con
.) that ifi/J(A)=(A-txt
lude that lj; = (A - A)n
~) = O if and only if f
Indced, if lj;
then
ynomial, and f(A) =
:emainder by lj; , then
of degree less than ni'
issumption that 0lj; is
:gree less than that of
~n be readily proved.
(Am))o Since lj;(A) = O
that the minimal poly-
dI,
Let Al, A2"'" Am be the distinct eigenvalues of A with indices nI' ii 2, ... , iimo
Let f and 9 be two polynomialso Then the following statements are equivalent.
1. feA) = g(A).
2. Either f =hllj; +g or 9 =hi.lj; + f, where lj; is the minimal polynomial of
A, and h 1 and h 2 are sorne polynomials.
3.
f([)(A) = gO)(A;)
where f
(1)
(Ad
6.
=
d1(A)i
~ A=A
(2-65 )
Proof
Theequivalence ofstatements 1 and 2 follows directly from the fact that lj;(A) = O.
Statements 2 and 3 are equivalent following
n (A - A)
m
lj;(A) =
Q.E.D.
;~ ~l
) 3
O 1
Q.EoD.
i= 1
n (A - A)n
m
i= 1
SO
Let
for l = 0, 1,2,
, ni - 1
i= 1, 2,
,m
<l)(A) = g(l)(A;)
If
(2-66)
then f(A)=g(A).
This fact can also be deduced directly from Corollary 2-12 (Problem 2-38).
Corollary 2-13 is useful in computing any polynomial and, as will be dis
cussed, any function of A. lf A is of order n, the polynomial g(A) can be chosen
as in (2-67) or as any polynomial of degree n - 1 with n independent parameters.
For example, if all eigenvalues, Ai' i = 1,2, ... , n, of A are distinct, then g(A)
can be chosen as
n -1
g(A)= I
Pi
i=O
n (A-Aj)
j= 1
A 100 =g(A) =c
Definition 2-16
Let f(A) be a functie
spectrum of A. If g(A
spectrum of A, then th,
This definition is :
be precise, functions e
(2-65). The condition
is easier to obtain the
willlead to the same J
If A is an n x n me
can find a polynomial
Ni
n-l
or
g(A) =
Pi
i=O
f1 (A -
Aj )
j= 1
We summarize th,
an n x n matrix A al
polynomial of A, say
Example 2
Compute A 100, where
A=[~ ~J
In other words, given f(A) = A 00, compute f(A). The characteristic poly
nomial bf A is ~(A) = det (Al - A) = (A - 1)Z. Let g(A) be a polynomial of degree
n -1 = 1, say
Let
wher a o, al'" ., ex n - 1
to compute these ex/s
we have f(A)=g(A).
n independent param
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)
polynomial of A can
n-l
.2 (Problem 2-38).
al and, as will be dis
rial g(A.) can be chosen
dependent parameters.
He distinct, then g(A.)
f(l) = g(l)
1'(1) = g'(l)
Solving these two equations, we obtain ex 1 = 100 and exo = -99. Hence
[1 2J [1
1 0J
AI00=g(A)=exoI+exlA=-99 [ O
1 +100 O 1 = O 200J
1
Obviously A 100 can also be obtained by multiplying A 100 times or by using a
different g(A.) such as g(A.) = exo +ex 1(A. -1) (Problem 2-33).
Functions of a square matrix
Definition 2-16
Let f(A.) be a function (not necessarily a polynomial) that is defined on the
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).
This definition is an extension of Corollary 2-13 to include functions. To
be precise, functions of a matrix should be defined by using the conditions in
(2-65). The conditions in (2-66) are used because the characteristic polynomial
is easier Oto ohtain than the minimal polynomial. Of course, both conditions
will lead to the same resulto
If A is an n x n matrix, given the n values of f(A.) on the spectrum of A, we
can find a polynomial of degree n - 1,
g(A.) = exo +ex 1A. + ... +ex n_12n-l
:e the computation.
n (A. - A.Jn
m
.1\(2) =
i= 1
Let
g(2)=exo +ex 1A. + ... +ex n_ 12 n- 1
e characteristic poly
polynomial of degree
where exo, ex 1, ... ,ex n-1 are n unknowns. Next we use the n equations in (2-66)
to compute these ex/s in terms of the values of f on the spectruID of A. Then
we have f(A) = g(A). We note that other polynomial g(A.) of degree n-1 with
n independent parameters cari also be used.
52
Example 3
ofthe form
Let
g(A) =
A,
! -~]
~ [~
f(2)
e21
g(2)
2(,( 1
= (,(0 +
21
+4(,(2
2 - 2e
O
el
O
[ _el +e21
Hence,
Solving these equations, we obtain (,(0 = -2te' +e21'(,(1 =3tel +2el -2e 21 ,
and (,(2 = e21 - el - tel. Hence, we have
eA" = g(A ) = ( - 2te' +e 21 )I +(3tel + 2e' - 2e 21 )A 1 +(e21 - e' - tel)A ~
2e' - e
(;(0
2
']
e AI
2e 21 _el
re~'1
t;
lo (
Example 4
Let
A function of a n
therefore, the relations
of a matrix. F or exan
A'~[~ j -~]
lts characteristic polynomial is ~(A) = (A - 1)2(A - 2), which is the same as the
one of Al in Example 3. Hence we have the same g(A) as in Example 3. Con
sequently, we have
e A21 = g(A 2 ) =
and if
then
2e' -e2'
O
[ 21
e -
Example 5
Given
Consider
~ ~~
[ A,
(n
n)
1
Al
O
O
O
O
1
(2-68)
1.]
Let the polynomial g(A) be
53
of the form
g(A) = Cto +Ct 1(A - A} +Ctz(A - Ad
<\)?
=?.
[ .
f '(A )
+---yf- (A -
f'(A 1)/1!
f(A 1)
A11) + ...
l
.
f(n-3)(A~)/(n-3)!
..
.
. .
.
O
O
..
O
O
...
f(lI- 1)(' )
(n -1;t (A -A 11)n-1
f"(A 1)/2!
f'(A 1)/1!
f(1)
(2-69)
f(A)
t n- 1e A,I/(n-l)!l
t n-ZeA,l/(n - 2)!
..
.
(2-70)
eA,1
Note that the derivatives in (2-69) are taken with respect to Al' not to t.
11
and if
then
Qf(A)Q-1
A [AlO
O]
A2
[(A) = [f(:)
f(~2)J
(2~1'1 ~
for any function f that is defined on the spectrum of A. Using (2-69) and
(2-71), any function of a Jordan-canonical-form matrix can be obtained im
mediately.
Example 6
Consider
Al
(2-68)
O: O O
OA 1 '1 : O
A=O O A1 :O
-------~-~-----
he polynomial g(A) be
O O O :A z
O O O: O A2
(2-72)
54
lf feA)
eAt , then
f(A)=e At =
lf feA) = (s -
(2-73)
Af 1, where s is a complex
feA)
Example 7
variable, then
Consider the Jordan-
(si - A)-l
1
S-Al
1
(S-A l )2
1
(S-A l )3:
1
S -Al
1
,
(S-A l )2:
,
,
,
,
,
1
s -Al
then
(2-74 )
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ .l. _ _ _ _ _ _ _ _ _ _ _ _ _
O
O
1
(s -A2)2
: s -A 2
Since (A -AlI)i is of
immediately to (2-69).
1
s-A2
Example 8
I
Definition 2-17
Let the power series representation of a function
be
ce
f(A)
a)i
i=O
of a square matrix A
ce
f(A)~
L aA
(2-76)
i=O
if the absolute values of all the eigenvalues of A are smaller than p, the radius
of convergence; or the mafrix A has the property Ak.= O for some positive
integer k.
A remarl< is in ord
by using Definition 2
requires the computat
the infinite series (2-78
the resulting matrix m:
converges very fast, th(
We derive some irr
to close this section. \
k-l
f(A)
i=O
ai A
(2-73)
55
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
(2-69).
Example 7
A given in (2-68).
Let
+f'i~l) (A -
AIf + ...
then
o
(2-77)
(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
ower series. We
:tion of a matrix. We
pression of a function
Antn
ih 2
e).! =
L
<Xl
eA! =
_tkAk
k=O
(2-78)
k!
I
(2-75 )
. of a square matrix A
(2-76)
es in (2-76) converges.
to
=1
oA(!+s) =eA!eAs
e(A+B)1
In (2-79), if we choose s =
=eAie Ht
(2-79)
(2-80)
e - Al
(2-81 )
56
2-8
~eAt=
dt
I _1_tk-1Ak=A(I ~tkAk)
k=l{k-l)!
k=ok!
=AeAl=eAtA
(2-82)
Here we have used the fact that functions of the same matrix commute (see
Problem 2-36).
The Laplace transform of a function f defined on [O, (0) is defined as
(2-83)
[tJ
Norms and
=1
2. IlctX11 = Ictlllxll.
3ll x l+ x zll::;llx 1 11+
The last inequality is e
Letx =[Xl Xz .
S-(k+ 1)
00
S-(k+l)A k =s-1
k=O
00
{s-IA)k
(2-84)
or
f{Je)={l-Je)-1
or
k=O
00
Jek
k=O
converges for 1,1.1 < 1. Now if s is chosen sufficiently large, the absolute values
of all the eigenvalues of S-1 A are smaller than 1. Hence from Definition 2-17,
we have
{1-s- 1A)-1
00
{s-IA)k
(2-85)
k=O
Hence from (2-84) we have
(2-86)
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
(2-82). Because of ff'[dh{t)/dt] = sff'[h{t)] - h{O), the application of the Laplace
transform to (2-82) yield~
sff'{e At ) - eO = Aff'{eA1 )
..(sl - A)ff'{e A1 ) = I
or
which yields imrridiate\y (2-86).For the matrices in{2-73) and (2-74), we can
also readily establish (2-86).
IIAII
The norm of A is de
norm. For different 11"
then
17
...
-- _..
__....._--,------_._----_._--
-_ .. - .. - - - -----------_.
_-..
2-8
(2-82)
X)) is defined as
(2-83)
57
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,
for convenience in the discussion, we restrict ourself to the complex vector
space (C", C).
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
as a norm if it has the properties that for any x in (C", C) and any a in e
3. Ilx I
Ilxlld~
(2-84)
i;
~ i~l
(
(2-87)
IXil
11
or
Ilxllz
or
Ixd
}IZ
(2-88)
(2-89)
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
froro Definition 2-17,
(2-85)
IIAII
(2-86)
where "sup" stands for supremum, the largest possible number of IIAxII or the
least upper bound of IIAxII. An immediate consequence of the definition of
IIAII is, for any x in (C", C),
(2-90)
IIAxII ::;IIAllllxll
The norm of A is defined through the norm ofx; hence it is called an indueed
norm. For different Ilxll, we have different IIAII. For example, if Ilxlll is used,
then
17
secti~-n
58
and
IIABxII
for any x.
The norro is a fun
two vectors, caBed th
two vectors x and y iJ
the properties that fo
(x,
(ex
(x,:
(2-91 )
IIABII ~IIAIIIIBII
(2-92)
and
Consequently we have
The sum of these two
magnitudes gives the
norm of A.
(a)
(b)
Theorem 2 -14
Ilx 11
!.~
----
If we define
~ 1
(Sch'
Ilxll = ( (x, ,
.....:;:-----
--l--'=~/--/--':::::""k-l--l--+--
-4 -3 -2
~Q..
Xl
Proof
1 ------;.::",j
2 r-4_ 5
Ax
-2
O~(x
+exy,
<:
(el
Figure
2'-7
= [_
~ ~J
which gives the Schwat
. _ - . -
and
Amax(A
59
for any x.
The norm is a function of a vector. Now we shall introduce a function of
two vectors, called the scalar product or inner product. The inner product of
two vectors x and y in (C", C) is a complex number, denoted by (x, y), having
the properties that for any x, y in (C", C) and any al' a 2 in C,
(x, y) = (y,x)
(alx l +a 2x 2, y) =a l (Xl' y) +a 2(x 2, y)
(x,xO
for all xi=O
: 2-7.
(2-91 )
(2-92)
\ +IIBII)\\xll
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
n
(2-93)
i=l
and
Consequently we have
This magnitude gives
the norm of A.
(b)
(2-94)
The inner product provides a natural norm for a vector x: Ilxll = (x, x) )1 /2
In fact, this is the norm defined in Equation (2-88).
Theorem 2-14
(Schwarz inequality)
(2-95)
for imy CJ. . Let a = - (y, x)j (y, y); then (2-95) becomes
(, where
A= [ 3. 2J.
-1 O
Q.E.D.
----
- -
60
2-9
Concluding Remarks
computation. The g
unstable and should b
is a numerically stabh
by using complete 1
However, according 1
formance, which inel
of use, the gaussian t
general matrices.
The gaussian elim
matrix into a triangul
then be determined.
small a number shoul
numerical property e
transformations. It (
are employed togeth,
method of computin~
composition (see App
a small number is a
"distance" to a matri
value decomposition
position however is q
orten oversimplified; 1
comparable results al
transformations with
The eigenvalues o
Once the characteristi
subroutines for solviJ
may be ill-conditione
cause large changes in
the eigenvalues may el
problem and should t
ejgenvalues is the SO-(
Hessenberg form and
Problems A-S and A
compute the charactel
the Leverrier algorith
S181 (see also Referen
The lordan canon
and results (see, e.g., I
solving Riccati equatic
canonical form, howe'
this problem, see Refer
18
..
._-_._-~_-~._---_.
---
CONCLUDlNG REMARKS
Iputing a function of a
,e introduced. (a) lJse
and then find a poly
le spectrum of A, then
'A: Let A=QQ-l.
1 and feA) is computed
putation of the topics
tioned if small changes
:ering terminology, if a
en the problem is ill
;)cedure for solving a
ble if numerical errors
algorithm is said to be
the numerical stability
'Iy, whenever possible,
problem. There are a
.tations. The reader is
d S212. Several .well
lACK are discussed in
sic topic in computer
61
Roughly speaking, balancing or scaling is lo rnake al! enlries of a rnalrlx be ofcorriparable size
or Ihe notrns of sorne colurnn and row lo be equal. Afler balancing, Ihe. resulls of cornpulalion
can oflen be irnproved. See References S82, S84, SI 03 and S202.
62
Problems
2-1 With the usual definition of addition and multiplication, which of the following sets
forms a field?
2-15 Given
2-2 Is it possible to define rules of addition and multiplication such that the set {O, 1, 2}
forms a field?
A
2-3 Given the set {a, b} with a f= b. Define rules of addition and multiplication such that
{a, b} forros a field. What are the and 1 elements in this field?
2-4 Why is (e, IR) a linear space but not (IR, C)?
2-5 Let lR(s) denote the set of all rational functions with real coefficients. Show that
(IR (s), lR(s)) and (IR (s), IR) are linear spaces.
. Hl [:!l HJ
b.
in (C 2 , IR)
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)
e.
~[:
;n (R'. R)
A,
3s 2 -124s 5 +s 3 -2s-1
1
, - - - - - - , ---::---- in (IR (s), IR)
3
2s +4s - 1
S2 +s-l
2-7 Is the set in Problem 2-6b linearly independent in (C 2 , C)? Are the sets in Problem
2-6d and e linearly independent in (IR (s), lR(s))?
b. (C, C)
c. (e, IR)
e. (lR(s), IR)
2-9 Show that the vectors Xl' X 2, ... , X k are linearly dependent in (IR "(s), lR(s)) if and only
if there exist polynomials c(s), i = 1, 2, ... , k, not all zera, such that
CI(S)X I +C2(S)X 2 + ...
+Ck(S)X k =0
2-10 .Show that the set of a1l2 x 2 matrices with real coefficients forms a linear space over
IR with dimension 4.
x(n) .
2-11
In an n-dimensional vector space (!!l", $'), given the basis el' e 2, ... , en, what is the
representation of e with respect to the basis?
PROBLEMS
63
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}.
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
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!)
coefficients. Show that
2-16 What are the ranks and nullities of the fol\owing matrices?
Al =
lent?
[~ ~ =~]
Al
130
=[~ ~~]
001
A3
=[~ ~ ~
;
;]
34500
2-17 Find the bases ofthe range spacesand the null spaces ofthe matrices given in Problem
2-16.
~J
nonsingular in the field of rational functions with real coefficients .1R(s)? For every s in e,
the matrices become numerical matrices with elements in C. For every s in e, are the
matrices nonsingular in the field of complex numbers e?
Are the sets in Problem
2-19 Does there exist a solution for the fol\owing linear equations"!
IR (s), IR)
x(n)-A"x(O)=[b
el' el, ... , e", what is the
Ah ...
-1)1
An-lbJ u(n:-2) .
[
u(O)
64
2-21
A, ~[ : -!J
A, ~[ l~]
3]
1
O
2
O
-150
200
20
-25
1
O
O
O
O
ll.
2
_J.
2
-"2
A, ~r!
2-22
-4
-2
160
1
O
O
-4
-2
-2
O
1
O
-3
19
2_27
Consider the m
of the matrix with multiI
following k vectors,
1
1
O
O
O
16
-20
1
1
1
O
O
!l
where
are generalized eigenvect
2-28 Show that the ma
that its inverse is given b
TI }'i
j=l
2-23 Prove that a square matrix is nonsingular if and only if there is no zero eigenvalue.
2-25
O
f3m(s
is equal to
s'
where n=k l +k 2
is eq ua] to
TI
+ ...
(J'r-q
1 s.i<jS."
2-26
tJ
19
20
-al1"1
-a 11 (nl-l)
O
O
O
O
See Reference 6.
-t
65
PROBLEMS
lowing matrices:
-!]
2-27
19
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
fol1owing k vectors,
I~]
O
O
O
O
O
-20
(n-1) X'-k
'.k-1
~ ~]
1 1
O 1
11-
where
O O
1)
i ;?: 1
123i
2-28
_a"i l/a"
:ix A. Show that
A- I
[
:re is no zero eigenvalue.
assumed to be a square
Sk m
-1
O
O
Skm-l
O
~m(s)
O
-1
O
O
O
O
O
O
Skm-Z
Sk2
-1
~m-l(S)
~m - zls)
~2(S)
SkI +~I(S)
is equal to
Sn +~I(S)S"-k, +~2(S)sn-kl-k,
where 11 = k l +k 2
2_J0
+km and
~i(S)
+~",(s)
O
O
O
O
O
O
O
O
O
O
O
O
-a121
-a 122
---------------------------------~----------------~--------------,
O
O
a"_I A +a n
[( AI .. ,
20
+ ...
+ ...
A~-IJ'
is an
Janion matrix of the poly
:rical analysis literature.J
19
20
O
O
O
O
-a21n!
-a21(",-I)
-a212
See Reference 6.
See Reference S17.
O
O
O
O
-a2!! : - a22n 2
1
O
O
O
-a22(", -1)
-a222
-a221
66
is given by
where
not exist.
\(s) = sn i
+aiils(ll-l)
+ ...
+aji(n-1)S
Let
+aiill
2-31
Find the characteristic polynomials and the minimal polynomials of the foHowing
matrices:
l~' ~J l~'
1 O
11 1
O 11
O O
1
}'I
O
O
O
1
11
O
~J l~'
1 O
11 O
O 11
O O
U l~'
1 O
11 O
O 11
O O
1]
What are the multiplicities and indices? What are their geometric multiplicities?
2-38 Let A be an /1 x I~
k
A with k ~n can be writ
the minimal polynomial
2-39 Define
(sI-Ar l
where D.(s) ~ det (sI -A)
stant matrices. This defir
most /1-1. Verify that
al
2-33 Repeat the problems in Examples 2 and 3 of Section 2-7 by choosing, respectively,
[~
O O
2-36 Show that functions of the same matrix commute; that is,
f(A)g(A) = g(A)f(A)
Consequently, we have Ae A' =eA!A.
tr AR I
az=- -
2
a,
tr ARo
=- -
= -
tr AR z
3
~~-
tr AR"_I
(X'l=- ~--
2-37 Let
(sI-A)-1 =
L
2-42 Let
Find a matrix B such that e
-,
..
~.
- _.. -
_.
-------------- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - .------------
PROBLEMS
67
c=
ni
1 O]
A.
O }, O
[
O O A.
Find a matrix B such that eH = C. [Hint: Let I(J.) = log}, and use (2-69).] Is it true that
for any nonsingular matrix C, there exists a matrix B such that f!l =C?
2-38 Let A be an n x n matrix. Show by using the Cayley-Hamilton theorem that any
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?
len j(A) is an eigenvalue
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
most n - 1. Verify that
tr ARo
a l =- - -
Ro=I
JY choosing, respectively,
.-2).
tr AR I
a2=- - 2
tr AR 2
.
3
......................................................................................
trAR"_2
a,=- - -
ct.'l-l ==
n-l
tr AR"_J
(/.,,=---
:ation.
O=AR n_ 1 +anl
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 + ... +
R,,_ 2S + R,,_ d. For a derivation of the left-hand-side equations, see Reference SI85.]
2-42 Let
i then the matrix B does
1
.
(si -Al-! = - Adjoint (si -A)
~(s)
68
and let m(s) be the monic greatest common divisor of all elements of Adjoint (sI - A).
Show that the minimal polynomial of A is equal to /'.i(s)/m(s).
2-43 Let all eigenvalues of A be distinct and let q be a (right) eigenvector of A associated
with A, that is, Aq = Aq. Define Q ~ [q1 q2 ... q"J and define
p~ ~ ~~
Q'
where p is the ith row of P. Show that p is a left eigenvector of A associated with A,
that is, pA = Ap.
2-44 Show that if all eigenvalues of A are distinct, then (sI - A)-l can be expressed as
1
(sI-Ar 1 =-qp
s -A
where q and p are right and left eigenvectors of A associated with A.
2-45 A matrix A is defined to be cyclic if its characteristic polynomial is equal to its
minimal polynomial. Show that A is cydic if and only if there is only one Jordan block
associated with each distinct eigenvalue.
2_46 21
Let
Q=[Q1
Q3
Q2J
Q4
= [1 -4 3J.
21
~ IX> g*(t)h(t) dt
PROBLEMS
69
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
iefine
th .
lolynomial is equal to its
is only one Jordan block
r is in a Jardan canonical
3
Mathematical Descriptions
of Systems
3-1
Introduction
The very first step in the analytical study of a systerri is t set up mathematical
equations to describe the system. Because of different analytical methods used,
or because of different questions asked, we may often set up different mathe
matical equations to describe the same system. For example, in network
analysis, if we are interested in only the terminal properties, we may use the
'impedance or transfer fundion to describe the network; if we want to know
the current and voltage of each branch of the network, then loop analysis or
node analysis has to be used to find a set of differential equations to describe
the network. The transfer function that describes only the terminal property
of a system may be called the external or input-output description of the system,
The set of differential equations that describes the internal as well as the terminal
behavior of a system may be called the internal or state-variable description
of the system.
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
developed from the concepts of linearity, relaxedness, time invariance, and
causality. Therefore they will be applicable to any system, be it an electrical,
a mechanical, or a chemical system, provided the system has the aforementioned
properties.
The class of systems studied in this book is assumed to have someinput
terminals and output terminals. The inputs, or the causes, or the excitatioris u
are applied at the input terrinals; the outputs, or the effects, or the responses y
are measurable at theoutput terminals. In Section 3-2 we show that if the
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
y(t) =
(3-1 a)
If the input and output have, in addition, the causality property, then (3-1a)
can be reduced to
y(t) =
feo G(t,r)u(r)dr
(3-1 b)
y(t) =
r G(t, r)u(r) dr
Jto
(3-1 e)
if the system is relaxed at too Equation (3-1) describes the relation between the
input and output of a system and is caBed the input-output description or the
external description of the system. We also introduce in Section 3-2 the con
cepts of time invariance and the transfer function. In Section 3-3 the concept
of state is introduced. The set of equations
o set up mathematical
alytical methods used,
et up different mathe
example, in network
~rties, we may use the
:; if we want to know
then loop analysis or
equations todescribe
the terminal property
,cription of the system.
as well as the terminal
te-variable description
t description and the
setting. They will be
time invariance,. and
em, be it an electrical,
Las the aforementioned
d to have sorne input
es, or the excitations u
~cts, or the responses y
2 we show that if the
(3-2a)
(3-2b)
that relates the input u, the output y, and the state x is then .introduced. The
set of two equations of the form (3-2) is called a dynamical equation. If it is
used to describe a system, it is caBed the dynamical-equation description or
state-variable description ofthe system. We give in Section 3-4 many examples
to illustrate the procedure of setting up these two mathematical descriptions.
Comparisons between the input-output description and the dynamical-equation
description are given in Section 3-5. We study in Section 3-6 the mathematical
descriptions of parallel, tandem, and feedback connections of two systems.
FinaBy, the discrete-time versions of Equations (3-1) and (3-2) are introduced
in the last section.
We are concerned with descriptions of systems that are models of actual
physical systems; hence all the variables and functions in this chapter are
assumed to be real-valued. Before proceeding, we classify a system as a single
variable or a multivariable system according to the following definition.
Definition 3-1
A system is said to be a single-variable system if and only if it has only one input
. terminal and only one output terminal. A system is said to be a multivariable
system if and only if it has more than one input terminal or more than one
output terminal.
The references for this chapterare 24, 27, 31, 53, 60, 68, 70, 73, 92, 97, 109,
and 116. The main objective of Sections 3-2 and 3-3 is to introduce the concepts
of linearity, causality, time invariance, and the state, and to illustrate their
72
importance in developing the linear equations. They are not introduced very
rigorously. For a more rigorous exposition, see References 60, 68, 109, and
116.
3-2
Linearity. We intro
exactly the same as thl
Definition 3-2
A relaxed system is saie
Figure 3-1
_ _ __ 0
_ .
--~,--,_
. -_._.-...-_ ..
.... .. _._."
_~-----.-"---~
._-_. ,_._-_.. __ .
_'_"._._ .._.-,-'-"..'---""---
-~~_._---
__
__ ._ ..
_._'._.
... _---_ ..
will be excited solely and uniquely by u. Hence, under the relaxedness assump
tion, it is legitimate to write
y=Hu
73
(3-3)
Definition 3-2
A relaxed system is said to be linear if and only if
H(a l Ul + a2 u 2) = alHu l
+ a 2 Hu 2
(3-4)
for any inputs u l and u 2 and for any real numbers al and a 2 . Otherwise the
relaxed system is said to be nonlinear.
I
In engineering literature, the condition of Equation (3-4) is often written as
H(u l + U2) = HUI
+ HU 2
H(au l ) = aHu l
(3-5)
(3-6)
for any u l , u 2 and any real number a. It is easy to verify that the condition given
in (3-4) and the set of conditions in (3-5) and (3-6) are equivalen1. The relation
ship in (3-5) is caBed the property of additivity, and the relationship in (3-6) is
caBed the property of homogeneity. If a relaxed system has these two properties,
the system is said to satisfy the principIe ofsuperposition. The reader may wonder
whether or not there is redundancy in (3-5) and (3-6). GeneraBy, the property
of homogeneity does not imply the property of additivity, as can be seen from
the following example.
Example 1
Consider a single-variable system whose input and output are related by
if u(t -1) 1=0
ifu(t-1)=0
>
terminals.
for aB t. It is easy to verify that the input-output pair satisfies the property of
homogeneity but not the property of additivity.
74
/i
Figure 3-2
If the input-output pe
then we have 2
U 2 implies that H(IXU 1 ) =IXHu, for any rational number IX (see Problem 3-9).
Since any real number can be approximated as closely as desired by a rational
number, if a relaxed system has the continuity property that Un -4 U implies
H Un -4 Hu, then the property of additivit)' implies the property of homogeneity.
We shall develop in the following a mathematical description for a linear
relaxed system. Before proceeding we need the concept of the delta function
or impulse function. We proceed intuitively because a detailed exposition
would lead us too far astray.1 First let (j t,(t - t 1) be the pulse function defined
in Figure 3-2; that is,
for t < t 1
Now if H 6(t - r) is km
puted from (3-9). The
the relaxed system due
all~.
As
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
a(t - t 1) dt
-ce
111 +E 3(t -
t) dt = i
ti-E;
(3-7)
For a rigorous development of the subsequent material, the theory of distributions is needed;
see Reference 96.
T~
u (1)
I
=U(I,)/i0:'
I
1
I
I
I
I
1
I
1
I
Figure 3-3
I 11
I 11
I 11
I~.
I
I
:1
I
Pulse-func
~._------_.-_.,-_._._-,_
.. __ .-
_.
-"'~'-
_-
.. __ ._ ... ..
_-_
... "---'-
75
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}~
If the input-output pairs of the relaxed system satisfy the linearity property,
then we have 2
r ex (see Problem 3-9).
; desired by a rational
ty that U n - u implies
perty ofhomogeneity.
escription for a linear
: of the delta function
a detailed exposition
pulse function defined
(3-8)
(3-9)
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)
or simply b{unction.
The condilion lor inlerchanging the order of H and lhe Slil11mation is disregarded.
cussion of this problem, see Reference S125, pp. 2-6.
u( J6{',.(t - ti) {',.
(3-7)
to develop a mathe~
.ss first single-variable
ariable systems. Con-
(3-10)
T~
u (ti)
1
= u (t) -.{',.
,{',.
I 1I
I
I
I
I
I
I
I 11
I
\
I
I
I
I
I 1 I
of distributions is needed;
Figure 3-3
(3-11 )
For a dis
_._- .-
76
Hence if g(', r) for aH r is known then for any input u, the output can be com
puted fram (3-11). In other words, a linear relaxed system is completely
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.
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
y(t) =
where
gql(t,r)
gqz(t,r)
(3-12)
Consequently, the in
becomes
Relaxedness. Rec
the system is--relaxed
solely and uniquely b
Definition 3-3
glP(t, r)l
g2P~t,r)
A system is said to be
and uniquely excited
gqp(t,r)
and gij(t, r) is the response at time t at the ith output terminal due to an impulse
function applied at time r at the jth input terminal, the inputs at other terminals
being identically zero. Equivalently, gij is the impulse response between the
jth input terminal and the ith output terminal. Hence G is called the impulse
response matrix of the system.
A1though the input and output of a relaxed linear system can be related by
an equation ofthe form in Equation (3-12), the equation is not readily applicable
because the integration required is from - 00 to 00 and because there is no
way of checking whether or not a system is initially relaxed at - oo. These
difficulties will be removed in the subseqtient development.
It is clear that if ~
system is still relaxed ;
are linearly related, it
for a system to be re.
oo,!]
(3-13)
(3-14 )
Hence, if a system is dI
:m.
(3-12 )
~,
y(t) =
Loo
,)
aid to be noncausal or
1s not only on the past
lplies that a noncausal
n the future. For real
an intrinsic property of
(3-15 )
G(t, ,)u(,) d,
Definition 3-3
and uniquely excited by
~: :~j
77
u[to,oo)'
Y[to,oo)
is solely
Example 2
A unit-time-delay system is a device whose output is equal to the input delayed
one unit of time; that is, y(t) = u(t - 1) for aH t. The system is relaxed at t o if
u[ro- I.to) == 0, although u( - 00,10 - 1) 1= O.
~
Now if a system whose inputs and outputs are linearly related is known
to be relaxed at to, then the input-output description reduces to
y(t) =
roo G(t,
,)u(,) d,
(3-17)
we know that the input-output pairs of the system satisfy the linearity property
and that the system is causal and is relaxed at to.
A legitimate question lllay beraised at this point: Given a system at time
to, how do we know thatthe system is relaxed? For a system whose inputs and
78
outputs are linearly related, this can be determined without knowing the
previous history of the system, by the use of the foHowing theorem.
Proof
If u[to,oo)
== O, the
out:
lo
y(t) = Loo 4
Theorem 3-1
A system that is describable by
Since
y(t) =
=o.
Proof
1
00
G(t, ,)u(,) d,
lo
=O
1~ G(t, ,)u(,) d,
r
J-oo
In words, the net effect ofu(-oo,lo) on the output y(t) for t ?t o is zero, and hence
the system is relaxed at to.
Q.E. D.
An implication of Theorem 3-1 is that given a system at time to, if the system
is known to be describable by
is a constant vector, t
on [to, (0). Conseq
(see Appendix B), anl
This is an import
Corollary 3-1, its rela
over any nonzero intl
interval, then the sys
next chapter that the
function matrices 01
satisfies the condition
able.
We give an examp
whose inputs and out
Example 3
Consider the system
the output is the volt.
stored in the capacitor
zero if no voltage is a:
relaxed at to, because
depending on which i
roo
Loo G(t,,)u(,)dT
the relaxedness of the system can be determined from the behavior of the
system after t o without knowing the previous history of the system. Certainly
it isimpractical or impossible to observe the output from time to to infinity;
fortunately, for a large class of systems, it is not necessary to do so.
1--------
Corollary 3-1
IL
R = I 11
I
I
I
I
I
I
I
+ +
(a)
__
.~
_,_. _ _. , _ . _ ' _ 0
I
i
79
Proof
Ir u[co,oo) == O, the output y(t) of the system is given by
y(t) = roo G(t,t)u(t)dt=M(t) roo N(t)u(t)dt
for t 2: t o
Since
roo N(t)u(t) dt
t 2: t o is given by
G(t, t)u(t) dt
Q.E.D.
at time to, if the system
Consider the system shown in Figure 3-4; the input is a voltage source, and
the output is the voltage across the nonlinear capacitor. Ir the electric charge
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
relaxed at to, because if an input is applied, we may obtain different outputs
. be decomposed into
:; (see Appendix B) on
'or some fixed positive
~}
1-------------1
R=I.l1
+ +q
(a)
(b)
Figure 3-4 A nonlinear network. (a) The network. (b) The characteristic o(the non
linear capacitor C.
. '
80
Time invariance. Ifthe characteristics ofa system do not change with time,
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
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.
Definition 3-4
A relaxed system is said to be time invariant (or stationary or fixed) if and only if
HQ"u=Q"Hu
(3-18)
for any input u and any real number \1.. Otherwise the relaxed system is said
to be time varying.
I
The relation HQ"u = Q"Hu can also be written as HQ"u = Q"y, which implies
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
input is applied to a relaxed time-invariant system, the waveform of the output
is always the same.
If a relaxed linear system is known to be time invariant, what condition
does that impose on its impulse response? The impulse response g(', -r) is the
output due to a J-function input applied at time -r; that is, g(', -r) = HJ(t - -r).
Ifthe system is time invariant, then we have
Q~(', -r) = Q"HJ(t = HJ(t - (-r
Now by the definition of Q", the equation Q~(', -r) = g(', -r +\1.) implies g(t, -r) =
g(t +\1., -r +\1.), which holds for any t, -r, and \1.. By choosing \1. = - -r, we have
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
for all
y(s)
Figure 3-5
_._ ...
.----
_-",,_
..... _-----_
.. _.. ..
--' .. _..
-_ ..... -
__ . _ - ;
._.
--:-_
".
_ _ ..
_~
_-_. - - _ .... _
__
~
_ . . ._ - _ . . .
L __ _ . . .
o._ _.
~~
~~
_ _
..__
_~
~._
_ _ .. _
- ._,_
._.
_ -
. - -_ _ . _ .
._.~-----_
y(t) =
rr G(t -,)u(,) d,
(3-19 )
we know that its input-output pairs satisfy the linearity, causality, and time
invariance properties; furthermore, the system is relaxed at time to. In the
time-invariant case, the initial time t o is always chosen, without loss of generality,
to be Oand the time interval of interest is [O, (0). Note that t o =0 is the instant
we start to consider the system or to apply the input u. Ir t o =0, Equation
(3-19) becomes
y(t) =
81
G(t-,)u(,)d,=
J~ G(,)u(t-,)d,
(3-20)
The second equality of (3-20) can be easily verified by changing the variables.
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
time " G(t) represents the responses at time t due to Ci-function inputs applied
at, =0. Following (3-14), a linear time-invariant system is causal if and only
if G(l) =0 for all l < O.
Transfer-function matrix. In the -study of the class of systems that are
describable by convolution integrals, it is of great advantage to use the Laplace
transform, because it will change a convolution integral in the time domain into
an algebraic equation in the frequency domain. Let y(s) be the Laplace transform
of y; that iS,4
y(s)
5t'(y) =
too y(t)e- SI dt
Since G(t - ,) =0 for ,> t, the upper limit of the integration in (3-20) can be
set at 00; hence, from (3-20), we have
y(s) =
~ G(s)(s)
3
(3-21 )
Note thatG(t,"t) and G(t -"t) are two <;lirrerent runctions, However. ror convenience, the same
symbol G is used.
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
Here we have changed the order of integration, changed the variables, and used
the fact that G(l) = Ofor l < O. As defined in (3-21), C(s) is the Laplace transform
of the impulse-response matrix; that is,
A
rational functions al
proper rational funcl
rro
G(s) = Jo G(t)e - SI dt
3-3
U (relaxedatl=O
y(s) I
= -;;;--()
U S
relaxed
atl=O
(3-22)
where the circumf[ex (.. ) over a variable denotes the Laplace transform of the
salIle variable; for example,
We see that the familiar transfer functions are the input-output descriptions
of systems. It is important to note that this input-output description is obtained
under the relaxedness assumption of a system; hence, if the system is not relaxed
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
at t =0.
A transfer function is not necessarily a rltional function of s. For example,
the impulse response g(t) ofthe unit-time-delay system introduced in Example 2
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
rational functions ofs. In fact, we study only a special class ofrational functions.
Definition ::S-5
A. rational function g(s) is said to be proper if g(oo) is a finite (zero al" nonzero)
constant. [(s) is said to be strictly proper if [( (0) = O.. A rational matrix C(s)
is said to be proper if C(oo) is a finite (zero or nonzero) constant matrix. C(s)
is said to be strictly proper if C( (0) = O.
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
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
matrix is proper if andortly if al1 of its elements are proper. A rational matrix
.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
amplified and will overwhelm information-bearing signals. Hence improper
The State
The concept of
51
_-_._~--,---,~."._-----
"
_.-._-~_.~_.-~-----_.-
..._.. _---,. __ ..
~---_
_----_._~---~--_._._-
--_...
~,._--
.. _- -_ ...
_._._----~._.
__ - _. __ ._._ --_._--------.
_. __ _----. ----..
.._.'--'-"
(3-22)
lt-output descriptions
lescription is obtained
;.: system is not relaxed
:d. Thus whenever a
1ssumed to be relaxed
)n of S. For example,
roduced in Example 2
rational function of S.
book are exclusively
s of rational functions.
83
rational functions are hardly used in practice. 5 In this book, we study only
proper rational functions and proper rational matrices.
3-3
. F or single-variable
r function. Hence the
?sponse; it can also be
...
[)efinition 3-6
The state of a system at time t o is the amount of information at to that, together
with u[to. ,.,), determines uniquely the behavior of the system for all t '2:. too
The exceptions are transfer functions ,of sorne transducers such as tachorneters and accelera
rneters. See Reference s46.
'
__
..
..
~'-'--'._'
_..
_._.-.~--
~-_._._._--~
~._
_'
~.
__ _
_ . _ . -
-_._.~_
__ _
_.~--
- _ _.
-.-.'-~'--"-"-'-'--""
- - - _ _ . _ , __ o
. _ . _
__
~'._~
.~~~~~~~~~~~~~~--~~~~
84
VOltag:
source
E3i
~ ~
~_-_
A network.
Figure 3-6
y(t)
known, then for any driving voltage the behavior of the network can be deter
mined uniquely. Benee the inductor current and the capacitor voltage qualify
as the state of the network.
-2e
Solving for
Cl
and
C2
Example 2
222
y(t) = (2
g(5)=(5+1)(5+2)=5+1- 5+2
-j
(3-23)
which is the inverse Laplace transform of the transfer function g(5). Now we
apply an input U[lo,oo) to the network. If the network is relaxed at to, the output
is given by'
.
r g(t-L)U(L)dL
Jl
y(t)=
for t ~to
In this example, w
is summari2
concept of state is ver:
(-
If the network is not relaxed at to, the output must be computed from
y(t)=
Example 3
g(t-L)U(L)dL
for
t ~ to
(3-24)
because the input that had been applied before t o may stHI have sorne effect on
the output after t o through the energy stored in the capacitor and inductor.
We consider now the effect on Y[lo, OO) due to the unknown input u(_ 00,10)' From
(3-23) we have
roo
t
21
g(t-L)U(L)dL=2e- roo e u(L)dL-2el
roo e
2t
00, t o)
u(L)dL
d
-d
t
J' g(t -
r)u(r) dr
= g(t
'o
fx:
(3-25)
r----'\AII~------JE
for
t ~ to,
Cl
Note that
where
Cl
----=t
~ Looo etu(L) dL
and
C2
and
C2
~ LOoo
XI
e 2t u(L) dL
C2
Figure 3-7
A netwo
. -
85
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)
1
lO
o, implies that
-o g(t-"l:)u("l:)d"l:
t
Solving for
Cl
and
Cz
(3-27)
Cz
0.5e'O (2y(to)
0.5e 210 (y(to)
+ y(to))
+ y(to))
+
(3-23)
r g(t -"l:)u("l:)d"l:
t
for t "2:t o
We see that if y(to) and y(to) are known, the output after t "2:t o can be uniquely
determined even if the network is not relaxed at too Hence the set of numbers
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
concept of state is very efficient and powerful.
omputed froro
g(t - "l: )U("l:) d"l:
Example 3
Consider the network shown in Figure 3-7. It is clear that if aH the capacitor
voltages are known, then the behavior of the network is uniquely determinable
for any applied input. Hence the set oI capacitor voltages Xl' X2, and X3 qualifies
as the state of the network. Let us examine the network more carefully. If we
r g(t-T)u(r)dT=g(t-T)U(T) I
t
'10
d
d).
t
lo
t=t
J' -g(t-T)U(T)dT
a
roat
(3-25 )
Lnd
C2
Figure 3 c 7
.. _.
__ o
86
_._. _ _
.. _._ _ . _ . _ _. _
. _ .
._0 __ -
'0_"
__ ._._. , __
....
"'._. __ .
_ .
_ _
_...
_ _
_ _
. '
_. _ _ _
__
...
. __ .
apply the Kirchhoff voltage law to the loop that consists of three capacitors,
we have Xl(t) +xz(t) +X3(t)= for all t. lt implies that if any two of Xl,X Z,
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
minable for any input applied thereafter. In other words, any two of the three
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
system, it is desirable to choose a state that consists of the least number of
variables. How to pick the state with the least number of variables for general
RLC networks will be studied in the next section.
dynamical equation.
the form
X(l
y(t
xz(t) =
xnCt)=J
Example 4
=6
yzCt) = 6
Yl(t)
A unit-time-delay system is a device whose output y(t) is equal to u(t -1) for
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
be called the state of the system at time too
I
From these examples, we may have the following observations concerning
the state of a system. First, the choice of the state is not unique. For the
network shown in Figure 3-6, the state may be chosen as the inductor current
and the capacitor voltage, or chosen as y(to) and y(to) or Cl and Cz. For the
network shown in Figure 3-7, any two of the three capacitor voltages can be
chosen as the state. Different analyses often lead to different choices of state.
Second, the state chosen in Example 1 is associated with physical quantities,
whereas in Example 2 the state is introduced for mathematical necessity.
Hence, the state of a system is an auxiliary quantity that may or may not be
easily interpretable in physical terms. Finally the state at each instant may
consist of only a finite set of numbers, as in Examples 1,2, and 3, or consist of an
infinite set of numbers, as in Example 4. Note that there is an infinite number of
points between [to -1, to); hence the state of this example consists of an infinite
set of numbers.
In this book we study only the class of systems whose states may be chosen
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.
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,
the state spaces we encounter in this book are the familiar finite-dimensional
real vector space(IW, IR).
Dynamical equations. In addition to the input and output of a system we
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
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
yq(t) =9
where x = [Xl Xz ..
and u = [Ul U z
state x are real-valued
for (3-28) to qualify a
initial state x(to) and :
sufficient condition for
initial state is that h an
see References 24,77, al
that the solution can 1
the state at to, as expe<
and is caBed a state eq
an output equation. N(
equation in the form 01
ledge ofx(t) and u(t) su
The state space of
set of equation (3-28)
linearity. We use tI:
----,._._---~
_.
--
._ ... __ ...._----
~.,-_
87
the form
i(t) = h(x(t), u(t), t)
(state equation)
(3-28a)
(output equation)
(3-28b)
(3-29a)
Xn(t) = hn(x (t), xz(t), ... , xn(t), u(t), uz(t), ... , up(t), t)
lservations concerning
not unique. For the
IS the inductor current
)r c and Cz. For the
acitor voltages can be
ferent choices of state.
th physical quantities,
athematical necessity.
at may or may not be
e at each instant may
, and 3, or consist of an
is an infinite number of
e consists of an infinite
e states may be chosen
r a system can then be
~alled the state vector.
Lear space in which the
ables are usually real
nber of state variables,
iliar finite-dimensional
l output of a system we
~o is, by definition; the
o), determines uniquely
o. The set of equations
Jut, and state is caUed a
88
(3-30)
and
utto,ro) = -ufro,ro)
(3-31 )
Ll
~)
Figure 3-8
= j.
where A, B, e, and E :
(Problem 3-36). Heno
form
E:
x(
yt
A sufficient condition I
of A(') be a continuous
the entries of B(' ), C(' ), ~
Since the values of A(
equation E in (3-32) is 1
equation.
1
In
L .
.2 H
xl(to) =x(t o
Time invariance.
3F
T
Y
1
~)
_ _
_ _ ,_u__ ,
,_""
,_ _ ,,
----,---_..- ..- ..
_-----~_.
__
._~------_._---_._--_.~--~.,~---~-----.-
89
x2 yfro. col}
(3-30)
IOnlinear.
relationship in (3-30) is
roperty of homogeneity.
'inciple of superposition.
1 and if
- ufro. col
" Hence a necessary
md U[to. 00) == O, then ihe
'ortant property of any
~ decomposed into two
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, uto.co)}
are called zero-state responses; they are excited exclusively by the input u[to.oo)'
Equation (3-31) foBows directly from (3-30) if we choose IX 1 = IX 2 = 1 and
xl(to) =x(to)
u 1 =0
x 2 (t O) =0
ufro.CO) =U[lo.oo)
Hence for linear systems, we may consider the zero-input responses and the
zero-state responses independently. The input-output description discussed
in the previous section describes only the zero-state responses of linear systems.
If a system is linear, the h and g in (3-29) become linear functions ofx and u as
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)
where A, B, e, and E are, respectively, n x n, n x p, q x n, and q x p matrices
(Problem 3-36). Hence an n-dimensional linear dynamical equation is of the
form
E:
+ E(t)u(t)
(state equation)
(3-32a)
(output equation)
(3-32b)
A sufficient condition for (3-32) to have a unique solution is that every entry
of A(') be a continuous function of t defined over (- 00, 00). For convenience,
the entries of B('), C('), and E(') are also assumed to be continuous in ( - 00, 00).
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.
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
lowing. Let Q" be the shifting operator defined in Figure 3-5.
Definition 3-8
A system is said to be time invariant if and only if for any admissible pair
(3-31 )
In
2H
3F
T
1
------(b)
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.
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:
(3-33a)
(3-33b)
90
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
positive-time direction, the input u(t 1) affects only the future responses, the
responses for t ~ t 1; it does not affect the past responses. Hence the dynamical
equations are aH causal.
In the stuoy oflinear time-invariantdynamical
equations, we may also apply the Laplace transformo Taking the Laplace
transform of FE and assuming x(O) =x o, we obtain
Transfer-function matrix.
(3-34a)
(3-34b)
A comparison of this
where the circumflex over a variable denotes the Laplace transform of the
same variable; for example,
(3-35a)
(3-35b)
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
~-'------'------'''-----,--------------
nd q x p real constant
variant n-dimensional
)n). For linear time
nitial time; hence it is
The time interval of
(si - A) always exists. If the initial state X o is O-that is, the system is relaxed
at 1 =O-then (3-35b) reduces to
Y(s) = [C(sl - A)-l B + E]u(s)
A comparison of this equation with (3-21) yields
G(s)=C(sI-A)-lB +E
e-invariant dynamical
Taking the Laplace
(3-34a)
(3-36)
G(s) = d
1
(1
et s -A
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
direction of negative
positive time. In the
future responses, the
Hence the dynamical
91
G(oo) =E
(3-37)
(3-34b)
aa22_-I[xX2(t)(t)] +[bb
12
ll
21
12][U1(t)]
b
b 22
U2(t)
~::][::~~G +[:::
(3-35a)
tEu(s)
In practice, pure difTerentiators are not used for the reason that they wil\ amplify high-frequency
noises. On the other hand, integrators wil\ smooth or suppress noises.
(3-35b)
Xl
(b)
(e)
Figure 3-9 Analog eomputer components. (a) Integrator. (b) Summer. (e) Amplifier
or attenuator.
---- -------_._-_.
92
Xl
e21I-----------J
'-----------~
'-------------+1
Figure 3-10
e22 1---
------'
Figure 3-11
X1DT=
X2DT=
X1 =IN
X2=IN
Y=X1
TIMER DELT,
PRTPLT X1,Y
STOP
END
93
[Xl]
X2
[2 -lJ[XIJ +[O ]
1
5_
y = [ 1 0.6J [
X2
1.5 u
: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:
~quation.
ion E.
DYNAMIC
PARAMETER U = 1.0
X1 DT=2.0*X1-X2
X2DT=X1 +5.0*X2+1.5 *U
X1 = INTGR(1.0,X1 DT)
X2 = INTG R(0.0,X2DT)
Y=X1 +0.6 X2
TIMER DELT =0.001 ,FINTIM =20.0,OUTDEL=0.10
PRTPLT X1,Y
STOP
END
94
y =0
(a)
Figure 3-12
A mech
Force
t .
Stallc
---+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 system generally consists ofmany subsystems or components. For example,
the network in Figure 3-6 consists of three components: one resistor, one
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
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.
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
pair {O, uo.ro)} ---+ {xo.ro), Y[O,ro)}, the pair {O, au[o.ro)} ---+ {axo.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
no physical system is linear according to Definition 3-2 or 3-7. However,
linear models are often used in practice to represent physical systems. This is
possible because most physical systems are designed to operate only in a certain
operational ranges. Limited to these ranges, physical systems can often be
approximated by linear models. This is accomplished by.linearlization or
simplification, as will be discussed in the following examples.
Example 1
Consider the mechanical system shown in Figure 3-12. The friction force
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
(a)
Figure 3-13
(a) Stati(
Force
Y2
Figure 3-14
Charact
EXAMPLES
95
y=o
y (s)
ti (s)
(b)
(a)
Figure 3-12
A mechanical system.
Force
Force
Viscous friction
Sta tic
----+-o-+-----
---~iL------
Velocity
Velocity
Coulomb
(a)
Figure 3-13
(b)
roo
---+-----,..<::------1e---_ Displacement
Figure 3-14
friction is clearly not a linear function of the velocity. To simplify the analysis,
we neglect the static and Coulomb frictions and consider only the viscous
friction. Let k be the viscous friction coefficient. Then the friction force
f is given by f = k1dy/dt. The characteristic of the spring is shown in Figure
3-14. lt is a nonlinear element. However if the displacement is limited to
(Yl, Y2) as shown, then the spring force is equal to k 2y, where k 2 is the spring
constant. Hence the linear model in Figure3-12(a) is obtained by linearization
and simplification.
Now we shall develop the input-output description of the system from the
external force u (input) to the displacement y (output). The application of
Newton's law yields
96
Taking the Laplace transform and assuming the zero initial condition, we obtain
u(s) = (ms 2 +k 1 s +k 2 )y(s)
y(s) =
If m = 1, k 1 = 3, k 2
g(t) =
= 2,
ms
lS
k u(s)
2
-1
(S
H=
J>(t - r)u(r) dr
and
mg- V=
.
=
[Xl]
or
X2
y=[l
O]
k1
m
X2
;;;
[:J
which imply
Example 2
and
With these linearized eq
variable descriptions of
to (3-38) and (3-39) yiel<
and
From these two equatioJ
function 90..(s) from u to
~Y-----1
Figure 3-15
and
gOll
EXAMPLES
11 condition, we obtain
97
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
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
- =e -1 -e -2/
d
J i . 2
H =m dt 2 (y +1 sin e) =mji +ml cos e!J -mi sin e (e)
and
d2
.
mg - V = m dt 2 (l cos e) = mi [ - sin e (j - cos e (efJ
These are nonlinear equations. Since the purpose ofthe problem is to maintain
the pendulum at the vertical position, it is reasonable to assume e and to be
small. Under this assumption, we may use the approximation sin e= e and
cos e= 1. By retaining only the linear terms in e and , that is, dropping the
terms with e 2, 2, ee, and ee', we obtain V = mg and
which imply
(M+m)ji+mllJ=u
top of it as shown in
are assumed to move
, and the gusts of wind
at the vertical position.
2{J - 2ge + ); = O
and
(3-38)
(3-39)
With these linearized equations, we can now derive the input-output and state
variable descriptions of the system. The application of the Laplace transform
to (3-38) and (3-39) yields, assuming zero initial conditions,
and
From'these two equations, the transfer function gyu(s) from u to y and the transfer
function gouCs) from u to e can be readily obtained as
21s 2 - 2g
(3-40)
gy,,(s) s2[(2M +m)/s 2 - 2g(M +m)]
A
and
-1
gou(s) = (2M +m)/s 2 - 2g(M +m)
(3-41 )
98
X2
e.
e,
..
2gm
2
Y= - 2M +m e +2M +m u
(J=2g(M+m)e_
1
u
(2M +m)l
(2M +m)l
and
From these two equations and the definition of Xi' the state-variable equation
description of the system can be readily obtained as
Xl
X2
X3
O
O
X4
y =[1
O
-2mg
2M+m
Xl
X2
X3
2g(M +m)
(2M +m)l
O
2
X4
2M+m
VI (t)
(2M +m)l
y 12(t)
O -1
V2(t)
ydt)
m2
1
v3(t)
Y34(t)
V4(t)
Consider four vehicles moving"in a single lane as shown in Figure 3-16. Let
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
for a11 four vehicles. Then we have, for i = 1,2,3,4,
V=Yi
(3-43)
Ui =kv +mvi
(3-44 )
= 1, 2, 3, 4
-k
O
-1
Note that Equations (3-39) to (3-42) are obtained l;lnder simplification and
I
linearization and are applicable only for small e and e.
v(t) = Vi(t) - Va
u (3-42)
O O O]x
Yi,i+l(t)=y(t)-Yi+l(t)-h a
and
mi
-k
Consider a satel!ite of
altilude of le sateilite
(3-45)
(3-46)
'l"/./##$////#//./#//#//.//#/##//./$#####/###//#///////#/#///##//////////////#/#/////////////////.
-Y4~
Y3~
.
- - - - - - - - - - - - - - - - - . : . . . - - - . - - - - Y2
Figure 3-16
YI-J
Figure 3-17
Satellite il
-" ---
EXAMPLES
lte variables as Xl = y,
m solve ji and (j as
and
1,2,3,4
'.
-k
1
-k
1
V\t)=-Vi +-u=-v(t) +-(t)
m
m
m
m
VI(t)
-1
YdO
(2M +m)l
(3-44)
o
2
2M +m
u
(3-47)
The term kvo is the force needed to overcome the friction for the vehicles to
maintain their velocities at va. Now the problem reduces to find (t) such
that Y,+l(t) and viCt) are as close as possible to zero for all t. From (3-45),
we have, for i = 1, 2,3,
Yi,+ 1(t)
;tate-variable equation
99
-k
mi
1
0-1
V2(t)
Ydt)
-k
m2
O 1
V3(t)
Y34(t)
V4(t)
O -1 O
-k
m}
O 1
O
1
mI
O
O
1
O
V2(t)
m2
ydt) +
O O
l(t)
2(t)
}(t)
4(t)
VI (t)
jidt)
v}(c)
O -1
Y34(t)
O
1
O
m}
O O
v 4(t)
o,
-k
m4
O
1
(3-48)
O
m4
This is the state equation description of the system. In this problem, we are
interested in the distances between adjacent vehicles and their velocities;
therefore the absolute distances do not appear in the equations. Depending
on what will be considered as outputs, an output equation can be similarly
developed.
Example 4 (Reference 596)
(3-45 )
(3-46)
U&
] UI-G
VI
I Salellite (mass m)
.
&
--
.
------ -
7//////'///'//$///'//&//////////,I/.
Figure 3-17
.....................
I
1",
1""
'"
.""v
~--"--_=:...::-
----
100
------~-----~--
--- -----
~---- ---~-
-_:::.._~.
can be controlled by three orthogonal thrusts U,.(t), ua(t), and u.p(t). The state,
input, and output of the system are chosen as
x(t) =
r(t)
I\t)
8(t)
e(t)
cP(t)
U,.(t)]
u(t)= ua(t)
[ uq,(t)
y(t)=
u)]
8(t)
3w 2
O
-2w
cP(t)
x(t) =
~(t)
ro
------------
x=
h(x u) =
.
..
,
-2r8/r +284> sin <PIcos 4> +ua/mr cos cP
(3-49a)
~.
y-( t ) =
1 O O O O O]
Y= ex = O O 1 O O o x
[
o o o o 1 o
xo(t) =
~o(t)
(3-49b)
8 0 (t)
<ro(t)
cPo(t)
ro
O
wt
W
uo(t) =0
+ cX l(t) -,
t:X2(t)
-:-
(3-50)
O
O
- -- -- - -
O O O O: 1
--
l O O O: O
O O 1 O:, O
wt +t: X3(t)
w +6: 4 (t)
cXs(t)
cX6(t)
[:~~~~~]
t: U3(t)
and
ro +cXl(t)]
y(t)=yo(t) +t:y(t) = wt +t:x~U)
[
cXs(t)
- - - - - - - --
------- -
EXAMPLES
[~l)
8(t)
,]
cP(t)
,
,
,
ro
------------------,------
,
l,/m
(3-49a)
/mrcoscjl
, O
I
O
O
O
,
O ,, O
1 ,
O
~,
mro:
U)
(3-51a)
,
----------
,
I
O ,, O
, 1
O
, mro
I O O O: O 0J
y(t)=
u<f/mr
101
(3-51b)
x(t)
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
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)
LI
X2
c l C2
XI
L2
Figure 3-18 Circuits witha: loop which consists of capacitors only or a cutset which
consists ofinductors only.
102
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
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
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
cedure for developing a dynamical-equation description of any RLC network
that may contain independent voltage sources or current sources 9 :
1. Choose a tree caBed a normal tree.
(voltage
source)
Figure 3-19
A networ
Example 5
Example 6
Consider the linear network shown in Figure 3-19. The normal tree is chosen
as shown (heavy lines); it consists of the voltage source, two capacitors, and
one resistor. The voltages of the capacitors in the normal tree and the current
of the inductor in the link are chosen as state variables. Next we express the
variables of resistors CD and (2) in terms of the state variables and inputs. By
applying the Kirchhoff voltage law (KVL) to the fundamental loop of branch
CD, the voltag across CD is found as '(u ~ Xl); hence its current is (u - xJ.
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
equation description cannot be or the form in (3-33).
- -1
L.
s
103
EXAMPLES
11 sources9 :
of inductors. Usually
)les.
.nches.
le fundamental loop or
iable.
IFundamental cutset of
I
(3)
0)
J F~ndamental cutset of
(U -xl)
lF
(voltage
source)
Figure 3-19
G)'
By applying the Kirchhoff current law (KCL) to the fundamental cutset of (Z),
we have immediately that the current through resistor (2) is X3' Consequently
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
to the fundamental cutset of branch Q), we have
(u-x)-x +U2
-X3=0
X3
=0
o
O
-1
-1] [1
1 x
+ O
-1
.1'"
. . /.,
Example 6
(1 +D
L
-~11ls)+
(s)
2)A
( l+s+~ 1 2 (s)=0
104
1.11
1
S
=
Ca)
Figure 3-20
Solving ror
Cb)
I 2(s), we obtain
1s
(l/s)u(s)
2( )-(1 +l/s)(l +s +2/s)-1/s 2
y(s)
g(S)=-A- u,~o
u(s) inilially
sI 2(s)
u(s)
Hence
T,--~[_':1_
=-A-= ~
1+
LCT
El__ ~
S2
3
1
s- +2s + s +
2
Cb)
relaxed
Figure 3-21
Network w
V(s) -sV2(s) =0
Hence
V(s)
u z{s
gds)=~) u,~o
initially
re1axed
Example 7
'Consider the network shown in Figure 3-21(a), where T is a tunneI diode with
the characteristic shown. Let x be the voltage across the capacitor and X2
thecurrent through the inductor. Then we have
..
[x~(t)J
(t)
2
EXAMPLES
105
-llR l
i = h(u)
R
u
E=.
(h)
(a)
xl
Xl
+
R
+
R
1+
Xl
-L
+2s z +3s
+1
J+ C
Xl
Rl
i=
Figure 3-21
-R 2
(e)
(h)
. ) - h(x 1(t))
x 1 (t =
e
uz(s)
uz(s)
(3-52)
. () -X1(t)-Rxz(t)
Xz t =
L
+I
~(s)
3s
Xz(t)
+c
+1
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
.
~oss
1
L
~J[~l(t)J +l~JE
.. R
__
L
xz(t)
106
In
u
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
will increase with time (in different polarity), and the network will eventuaJJy
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
In
In
-IF
iz
+"
Figure 3-22
t
In
Figure 3-23
Active:
A dynamical equ~
input and output 1
condition; hence a
2. For ~xtremely com
the dynamical-equ;
find the input-out:
each input termim
output terminals g
system. In practic<
narrow pulse, or ide
unit step function.
The response of a I
given by
at to). DifIerentiatiJ
In
In
-IF
In
1n
scription
1n
-1 F
CD1
1
Y2
In ;n
Figure 3-23
107
Active networks.
x=x
{Y1 =O.5x + O.5u
x=x
{Yz,=u
{l
U1(t-t O)= O
for t to
fort<t
o
( 5 )
3- 3
9 1(t, to) =
r' g(t, r) dr
J,o
(3-54 )
where ~l(t, to) is called the step response (dlle to a unit step function applied
at to). Differentiating (3-54) with respect to to, we obtain
(3-65 )
d
g(t)= dtg1(t)
(3-56)
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
3-6
two multivariable s:
The material in this section is not used until Chapter 9; thus its study may be postponed.
(al
u
109
at a number of fre
:d gUw). This method
tice.
mostly carried out by
n limited to the single
l.d not been successful.
,Os. In this approach,
~s are the same, and a
re not available in the
:ntly, interest in this
s in the state-variable
on approach. lt also
>ts and computations.
lme-varying case; the
ng case has not been
lS can be used to study
;ing transfer functions.
l.pproach taken, either
. the stability problem,
analysis and operator
1f L yapunov's second
-equation description.
available, the system
nputer.
-output and the state
to carry out a design
lese two mathematical
be developed equal1y
roo G(t,r)u;("r)dr
Yi(t) =
G(t, r) = G 1(t, r)
+ Gz(t, r)
G(t, r) =
(3-69)
We prove (3-59) for the single-variable case. The impulse response g(t, r) is, by
definition, the response at yz due to a b-function applied at time r at Ul' The
response at Y1 due to this o-function is g 1(t, !). The output of Sz due to the input
gl(t,r)is
U=U;{~
O-~/
(a)
(b)
u
(e)
(3-58)
site Systems 10
(3-57)
where Ui and Yi are the input and the output, G i is the impulse-response matrix
of the system Si. Let u, y, and G be, respectively, the input, the output, and
the impulse-response matrix of a composite system. We see from Figure
3-24 that in. the paral1el connection, we have u 1 =u z =u, Y=Yl +Yz; in the
tandem connection, we have u =u 1 , Y1 = u z, Yz = y; in the feedback connection,
we have U1 = u - Yz, Y= Yl' Here we have implicitly assumed that the systems
SI and Sz have compatible numbers of input and output; otherwise, the systems
cannot be properly connected. lt is also assumed that there is no loading
effect in the connection; that is, the impulse response matrices G, G z remain
unchanged after connection. 1t is easy to show that the impulse response
matrix ofthe parallel connection of SI and Sz shown in Figure 3-24(a) is
built by interconnect
Josite system. Whether
'e consider as building
:ed upon as a system,
I amplifiers, function
rhere are many forms
,m the fol1owing three
jback connections, as
. the input-output and
e systems.
ite systems. Consider
i= 1,2
(b)
110
Hence,
g(t, r) =
gz(t, v)g[(v, r) dv
description is
G[(t, v)
Gz(v, s)G(s, r) ds dv
(3-60)
where G[ and G z are known and G is unknown. This equation can be easily
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.
Now we study the state-variable descriptions of composite systems. Let
the systems S [ and S z in Figure 3-24 be described by
X = A(t)x + B(t)u
y =C(t)x +E(t)u
i= 1, 2
(3-61 a)
qualifies as the state of any composite connection of S [ and S z; its state space
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
y = [C[ (t)
Cz(t)]
[:J
where y [(t)=(I+E l ,
order for (3-64) to be
and (1 + E z(t)E l (t)) eJ
verified by observing
(3-61b)
where x is the state, o is the input, and y is the output; A, B, C and E are
matrices of compatible order whose entries are continuous function of t defined
over l - 00, (0). The state space of S is denoted by L.
Let us introduce the concept of the direct sum of two linear spaces. The
linear space L is the direct sum of two linear spaces L[ and L2> written as
L = L[ EB LZ' if every vector in L is of the form [x{ x]', where x[ is a vector
in L[ and X z is a vector in LZ' The dimension of Lis the sum of those of L[
and L z.
It is clear that the composite vector
[::] ==
y = [Y [(t)C[(t)
Time-invariant ca
applied to the time-;
discuss the transfer-J
Gz(s) be the proper
tively; then the transl
is G l (s) +Gz(s). Th
SI followed by Sz is
reversed. It is clear
G Z(s)G l (s). It is im
compatible in both c;
In order to discus
shown in Figure 3-24
Theorem 3-2
(3-62a)
(3-62b)
(3-63b)
111
description is
.le case.
the impulse-response
r) ds dv
(3-64a)
(3-60)
, where x[ is a vector
le sum of those of ~ [
y = [Y 1 (t)C1(t)
- Y 1 (t)E 1(t)C2(t)]
[::J
+ Y1(t)E 1(t)u
(3-64b)
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
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
verified by observing U1 = U-Y2' Y1 =U2' Y=y [. (See Problem 3-28.)
Time-invariant case. All the results in the preceding subsection can be
applied to the time-invariant case without any modification. We shall now
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
tively; then the transfer-function matrix of the parallel connection of SI and S 2
is G[(s) + Gz(s). The transfer-function matrix of the tandem connection of
SI followed by S2 is G 2(5)G 1 (5). Note that the order of G 2(S)G[(5) cannot be
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.
Theorem 3-2
Let G[(s) and G 2 (s) be, respectively, q x p and p x q rational function matrices
(not necessarily proper). Then we have
liI
(t)Ju
(3-62b)
S [ and S2 is given by
[(t)
)E[(t) u
(3-63 a )
(3-63b)
,Y=Y2'
:s dynamical-equation
(3-66)
112
where N, Q, and Pare any square matrices of the same order. Let us choose
N - [Ip
-
- (;(s)
0J
I q
Q _ [I p
-
P = [I p (;2(S)J
O
Iq
- (;2(S)J
Iq
(;(s)
G(s) = G I,
Proof
:J
and
pQN=ep+(;2~)(;(S)
hence
Theorem 3-3
Consider a feedback:
(3-67)
Consider the feedback system shown in Figure 3-24(c).. Let G1(s) and G 2 (s)
be q x pand p x q proper rational transfer-function matrices of SI and S2'
respectively_ lf det (I q + G 1(s)G 2 (s)) + 0, then the transfer-function matrixof
l;
Obviously there is no
is said to be inconsi~
det (I q + G 1(S)G 2 (s)) +
Recall from (3-64)
equation description o .
where E is the direc
Because of G(oo) = E
113
(3-68)
Proof
erify that
(I q
+ (;1(S)(;2(S))Y(s) =
(3-69)
(;l(s)ii(s)
Q.E.D.
Q.E.D.
Example 1
Consider a feedback system with
-s
s +1
G 1(s)= _1_
s +2
1
[
o element in the field
aor( suggestively as
et(I q + (;1 (S)(;2(S)) +0
dsts. FroID Theorem
s))
+ (;1(S)(;2(S)) =0.
A
+O
exist.
~s~1
s +1
u(s)
Consider the
[1 0J
G 2 (s)= O 1
A
Let us choose
1]
s +2
[~
(s
1)2
+
r
(3-67)
(3-70)
s +1
Q.E.D.
: the transfer-function
Recal1 from (3-64) that the condition for the existence of the dynamical
equationdescription ofthe feedback system in Figure 3-24(c) is det (1 + E 1 E 2 ) O
where Eiis the direct transmission part of the dynamical equation of Si'
Because of(;(co) = E [see Equation (3-37)J, if det (1 + E 1 E 2 ) O, then det(1 +
I
I
ji
114
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 +EEJ+-O. Hence, a feedback
system may have the transfer-function matrix description without having the
state-variable description. This discrepancy will be resolved in the following
subsection.
1
s
G2 (s)
Example 2
1
s+
Uz
F
+
-1
G1(s)= _1_
[
s +1
-JJ
Figure 3-25
s+lJ
s
1
-1
s +1
G(s)= _1_
[
s +1
~ lJl
s
-s-2
s+l
O
1
s+l
~ Jl-l =[-s-s+1
-1
For thefeedback system in Figure 3-25, if the input signals u(t) is corrupted
by high-frequenS?' noises o(t), then the noises will be amplified by the improper
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
noise is only one hundredth ofthe amplitude ofthe signa! atthe inptterminals,
the ainplitude of the noise at the output terminals is 10 times larger than that
Feedbal
s +1
which is not a proper rational function matrix. The block diagram of this
-s .
s+
M-(s) = {Mp(s)[1
If M; (s) is proper, tr
strictly proper, M sp( G
constant matrix. Thu:
From M-1(s)=[
(Mp(S)+O)-l-.[Adj l.\
11
115
u,
-1
G1(s) and
(+ (;1(S)(;2(S)) #=0 for
+o-~
+
'-
s+T
E
Uz
ith
Figure 3-25
=1
-s + 2
s+1
yz
of the signa\. Hence, the system in Figure 3.25 is only of limited use in practice,
although al1 of its subsystems have proper rational functions. Thus in the
design of feedback systems, we shal1 require not only that al1 subsystems have
proper transfer f.unctions but also that the resulting overal1 system has a proper
rational matrix, In the remainder ofthis subsection, we shal1 study.this problem.
Before proceeding, we need a preliminary result.
Theorem 3-4
Let M(s) be a square rational matrix and be decomposed uniquely as
M(s) = Mp(s)
+ Msp(s)
(3-71 )
Proof
-s
l' A sufficient condition for M; '(s) to be proper is Mp(s) column-reduced or row-reduced. See
Appendix G. For necessaryj'and sufficint conditions, see Problem G.l7 and Sections 2-4 and
3-4 of Reference 534.
116
not a finite constant matrix and M- 1(s) is not proper. 12 Now we assume
Mis) to be nonsingular and show that if M- 1(s) is proper, so is M; 1(s). From
M(s) = Mp(s)[1 + M; 1(s)Msp(s)], we have
Mo(s)~ [1 + M; 1(s)Msp(s)] -1 = M- 1(s)Mis) = M- 1(s)[M(s) - Msp(s)]
= 1- M -1(s)M sp(s)
which, together with the finiteness ofM -1(00) and Msp(oo) =0, implies Mo(oo) =
1. Hence we have M- 1(00) = Mo(oo)M; 1(00) = M; 1(00), and M;I(OO) is a
finite constant matrix. Thus M; I(S) is proper.
Q.E.D.
The condition de
theorem is the same a
dynamical descriptior
resolves the discrepan
Before discussin
det [1 +G 2 (00)G 1(oof
Figure 3-26. This sy~
and will be developed
The transfer matrices
matrices of appropriat
This theorem shows that the properness of M- 1(s) depends only on the
polynomial part of M(s). We give a special case in the following.
e(s)=
which implies
Corollary 3-4
e(s) =
If M(s) is a square proper rational matrix, M -1(S) is proper if and only if M( (0)
is nonsingular.
Proof
lf M(s) is proper, the polynomial part of M(s) is Mis) = M(oo). Hence M- 1(s)
is proper if and only if M( (0) is nonsingular.
Q.E. D.
Using these results, we are ready to answer the question posed at the begin
ning of this subsection: the properness of G(s) in (3-68).
Theorem 3-5
G2
Go(.
lf G(s) are proper, the
nonsingular, Corollary
in (3-73) is proper. H{
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.
Proof
G(s) is the
which is proper.
12
Figure 3-26
A feedbad
117
r. 12 Now we assume
~r, so is M; 1 (s). From
(s) [M(s) -
Msis)]
)= 0, imp\ies M 0(00 ) =
X), and M;I(oo) is a
Q.E.D.
which implies
e(s) = [1
(;o(s) ~ I
+ (;3(S)(;,(s) + (;4(S)(;2(S)(;I(S)
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
...
Gz(s) =
Then we have
1
s
O
Co( (0) =! -! +0 =
-1
(;(s) is the
(3-73)
Ip view ofTheorem 3-5, one may wonder whether (;(s) is proper ifand only if
G o( (0) is nonsingular, where
s proper.
;:Jer.
Sin ce
(3-72)
=I
which is proper.
y
u
i-----r-r----.,/]
= I - (;2(S)(;(S)
Figure 3-26
A feedbacksystem.
G (s) I--
~,.______./
118
For this system. GAs) is proper; wh...e reas Go( (0) is singular. Will such a
system be acceptable in practice? lf Go(oo) is singular, the transfer matrix
from u(s) to c(s) in (3-72) is improper. In this case, if u(t) contains a part which
has a high-frequency spectrum or is discontinuous, the amplitude of e(t) will be
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
but also all transfer functions from all possible input-output pairs of the system
to be proper. In this case, no signal will be unduly amplified, and in sorne sense,
the smoothness of signals throughout the system will be preserved.
Theorem 3-6
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
~~-
Definition 3_9 13
Let every subsystem of a composite system be describable by a rational transfer
function. Then the composite system is said to be well posed if the transfer
function of every subsystem is proper and the c1osed-loop transfer function
from any point chosen as an input terminal to every other point along the
directed path is well defined and proper.
I
In this definition, if a point, say E in Figure 3-26, is chosen as an input
terminal, then we must add a fictitious input as shown. Then the c1osed-loop
transfer functions from r to e, W, and y are, respectively.
(3-74)
'---.J
(3-75)
(a)
Proof
The transfer matrices in (3-72) to (3-74) are c1early proper if and only if G l(S)
is proper. Similarly, it can be shown that if G l(S) is proper, the transfer matrix
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, S27 and S214.
(e)
Figure 3-27
Feedback
e by a rational transfer
II po sed if the transfer
loop transfer function
other point along the
I
is chosen as an input
Then the closed-Ioop
'(5)
)G 1(5)]-11-(5)
)G 1(5)] -11-(5)
(3-74)
from any point to any other point along the directed path is proper. This
establishes the first part of the theorem. If a11 Gi(s), i = 1, 2, 3, 4, are proper.
the polynomial part of Go(s) is Goloo). Hence (;0 I(S) is, fo11owing Corollary 3-4,
proper if and only if Go((0) is nonsingular.
Q.E. D.
We discuss now the implications ofthe conditions det [1 + G2 ( CIJ )G 1 (CIJ)J f- O
and more genera11y, det Go(oo) f- O. Before proceeding, we need sorne concepts.
A loop of a block diagram is a closed path which travels from a point along the
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.
In our discussion, we are interested in only the value of the loop gain at s = oo.
Clearly if a loop gain is a strictly proper rational function, then its loop gain is
zero at s = CIJ. If a11 transfer functions of a block diagram are prdper, a loop
has a nonzero loop gain at s = 00 if and only if a11 transfer functions along the
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,
only loops with nonzero loop gains at 5=00 will come into the discussion.
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
loop with loop gain + 1 at s = 00, which is equivalent to a gain of infinity or an
improper transfer function as shown in Figure 3-27(c) and (d). Hence, we may
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
function.
The system in Figure 3-28(a) has a loop with loop gain 1; however, we have
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)
119
Figure 3-27
120
or an improper transfe
to be well posed.
}--E..-_ _---+I 2s + 1
-s
3-7
(b)
(a)
Discrete-TiI
(d)
(e)
and (d). Henee, if there are two or more loops with nonzero loop gains at
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;
henee the loop will not cause any problem, and the system is well posed. The
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
or an improper transfer function. Note that what causes the problem is the
combined loop instead of individual loops.
To further verify the aforementioned statement, we consider the system
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
proper transfer functions, the loop gain is zero at s = oo. Hence, the combined
loop which passes through the point H has a net loop gain 1 at s = oo. This
loop is equivalent to again of infinity. Similarly, the combined loop which
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
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
y(k) =
14
D/SCRETE-TIME SYSTEMS
121
3-7
(b)
(d)
temo
Discrete-Time Systems 14
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
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
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.
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
diserete-time system satisfy the linearity property, then they can be related by
.00
y(k) =
(3-76)
where g(k, m) is called the weighting sequence and is the response of the system
due to the application of the input
u(i) =
L g(k, m)u(m)
-00
{~
i=m
ifm
with the system relaxed at time m-o lf the discrete-time system is causal
that is, the output does not depend on the future values of the input-then
g(le, m) = O for le < m. Consequenty, if a discrete-time system is causal and is
relaxed at ko, then (3-76) reduces to
k
y(k) =
g(k, m)u(m)
(3-77)
m~ko
y(k) =
g(k - m)u(m)
m~O
14
k =0, 1, 2, ...
(3-78)
122
omitted. We introd
time-varying, discrete
00
u(z) ~ Q[u(k)] ~
U(k)Z-k
k=O
Example 1
u(z) =
L.,
k=O
z- =
That is,
1
z
=-
1-z- 1 z-l
Example 2
Then
u(z) =
e- 2kz- k
~[x(k+
k=O
Now we shall apply the z-transform to (3-78). For a causal, relaxed system,
we have g(k - m) = O for k < m; hence we may also write (3-78) as
00
y(k) =
g(k - m)u(m)
m=O
Consequently, we have
00
y(z) =
00
y(k)Z-k =
k=O
00
00
L L
k=O 111=0
x(z)
y(z) =
L L
g(k - m)z-(k-lIIlu(m)z-m
00
k=O
00
m=Ok=O
=
= (
00
g(k)Z-k
u(m)z-III =g(z)u(z)
(3-79) .
m=O
Here we have changed the order of summations and used the fact that g(k - m) =
O for k < m. The function g(z) is the z-transform of the weighting sequence'
{g(k)}k'=o and is called the z-'ansfer function or sampled transfer function.
The extension of the input~output description of single-variable discrete
time systems to the multivariable case is straightforward and its discussion is
y(z
where
G(z
----------,-----
123
OISCRETE-TIME SYSTEMS
that in diserete-time
rwise, aH the concepts
~he Laplace transform,
;ebraie equation. We
will be used is caBed
~fined
x(k
+ 1) =
A(k)x(k) + B(k)u(k)
y(k) = C(k)x(k) + E(k)u(k)
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.
If A(k), B(k), C(k), and E(k) are independent of k, then DE reduces to
as
DFE:
x(k
+ 1) =
Ax(k)
y(k) = Cx(k)
+ Bu(k)
+ Eu(k)
(3-80)
That is,
<X)
x(z) ~ .?l'[x(k)J ~
X(k)Z-k
k=O
Then
<X)
.?l'[x(k+ l)J =
I
=Z[
causal, relaxed system,
(3-78) as
<X)
k=~
1 x(k
Xo
Equation (3-81)
CHD
(3-81 )
be arranged as
(3-82)
(3-83)
where
thefact that g(k - m) =
le weighting sequence
'd tri:msfer function.
ngle-variable discrete
'd and its discussion is
(3-84)
124
function
A
+2
z-l
Z2
3-8
whereP, Q, R and W
is called the polynom
Both the input-ou
in this chapter are use
on the problem, on tl
Problems
3-1 Consider the merr
whieh u denotes the in:
possible to introduce a n
3-2 We may define th
response due to a b-fun<
describing the system?
Concluding Remarks
Consider a relaxed
---,.......':----.
(a)
Figure
P3~1
II
PROBLEMS
applicable to the
125
Problems
3-1 Consider the memoryless systems with characteristics shown in Figure P3-1, in
which u denotes the input and y the output. Which of them is a linear system? ls it
possible to introduce a new output so that the system in Figure P3-1(b) is linear?
3-2 We may define the impulse response of a relaxed nonlinear system gL r) as the
response due to a -function input applied at time r. ls this impulse response useful in
describing the system? Why?
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
Although the input
Iso be developed from
without knowing the
tt systems, the transfer
er functions (including
used, the systems are
equations X =f(x, u, t),
E(t)u-to be qualified
and any input u(to,oo)'
ms. This uniqueness
3.mical equation. The
form X = A(t)x + B(t)u,
ttives ofu in the output
for aH t
where w and t o are constants. Is the ideallow-pass filter causal? ls it possible to build an
ideal low-pass filter in lhe real world?
U(t)
y(t) = (p.u)(t) ~ { O
for t > ex
for any u, where ex is a fixed constant. In words, this system, called a truncation operator,
chops off the input after time ex. ls this system linear? ls it time invariant? ls it causal?
) + ...
Let Pa be the truncation operator defined in Problem 3-5. Show thal for any causal
relaxed system y = Hu, we have
p.Y =P.Hu =PaHPau
tice.
differential equations
The reasons are as
be written as a set of
to describe tirst-order
r differential equations
uter.
llations,one may also
3-6
_ _ _-,//c-,--
----:*"------u
(a)
Figre P3-1
--07"''''---+:,----~
(b)
(e)
126
3-7
3-14
for al! t in (- 00, oo)? Ir not, find the interval of time in which the equation holds.
3-8 Consider a linear system with input u and output y. Three experiments are per
formed on this system using the inputs u(t), uz(t), and U3(t) for t ;::::0. In each case, the
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?
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
number rx and any u.
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
transfer function a rational function?
3-10
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 ,
Show that this definition implies that y(t) = Hu(_ 00,1]' and vice versa.
3-18
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)=
gx + y)/2, (x - y)/2). Show that ag(t, L)/aX = O. [From this fact we may conclude that if
g(t, L) = g(t + rx, L + rx) for all t, L, and rx, then g(t, L) depends only on t - L.]
3-13
\
,,(,,~
[m
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.)
u(tJ
(a)
Figure P3-17
E-
g(tJ
bl-----t-I
'2
(a)
Figure P3-13
R1
-1
- - - + - 1-----,.-,' t
11 1
Cz
+
rv
(b)
Figure P3-19
II Z
Rz
PROBLEMS
3-14
127
e equation holds.
'ee experiments are per
t 2': O. In each case, the
,d outputs are Yl(t), yz(t),
O) 1= O?
9 ij\S) =
If[Yi(t)]
ro-[
()J
.;L
iniLially relaxed
Uj
t and u,~Ofork-l=j
is the jth component of u.
Uj
N 11 (P)Yl(t)
N ZI (p )Yl (t)
where the No/s and DJs are polynomials of p ~ d/dt. What is the transfer-function matrix
of the system?
3-17 Find the dynamical-equation descriptions of the systems shown in Figure P3-17.
l' and z are very smal!, can you consider the two systems as linear? Find the transfer
If
functions and dynamical equations to describe the Iinearized systems.
e, e
3-18 Find the dynamical-equation descriptions of the networks shown in Figures 3-22
and 3-23.
3-19
t - 't.]
(a)
Figure P3-17
(b)
Figure P3-19
(b)
128
3-20 Consider the simplified model of an aircraft shown in Figure P3-20. lt is assumed
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
generate forces JI = k I and J2 = k 2u, as shown in the figure. Let m be the mass of the
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
is, by neglecting the effect of 1,
1F
E-
3-21 The soft landing phase of a lunar module descending on the moon can be modeled
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
system as XI = y, X2 = y, X, = m, and u = rh. Find the dynamical-equation description of
lhe system.
3-22 Show that the output of a linear causal relaxed system due to the input u(t)i5 I(t - to)
is given by
y(t) = u(to)gl(t, to)
l'
U(T)gl(t, T) dT
'o
for t ~to
-I.Q
Currenl
source
Figure P3-23
h~
:C--~----
~12~11r
3-27
Figure P3-20
where El is a q x p matrix.
3-28
3-29
Show that
Th rus! = kli
det(I"
Lun"ar surlce
Figure P3-21
PROBLEMS
129
~ure
P3-20. lt is assumed
.eleva tor angle uo, altitude
::>f 8 and u from 80 and Uo
Let m be the mass of the
he aerodynamic damping,
nsfer function from u to h
-In
Curren!
source
In
y
f
IH
t"2:to
Figure P3-23
~l
where 9 1(t, to) is the step response and b is the step function defined in (3-53) and (3-54).
[Hin!: This can be proved either by decomposing u into a sum of step functions or by using
(3-55).]
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?
3-23
3-24
3-25
Verify that the impulse-response rnatrix of the feedback systern in Figure 3-24(c) is
given by Equation (3-60).
3-26
3-27
3-29
Show that
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
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,
3-30
:2J
ss +1
2.5
s +2
3-31
[~IIJ=[-2O
XIZ
YI =[0
_IJ[XuJ
1 XIZ
lJx l +[1
+[_
Y-B
0.5
lJU I
1 2
-IJu I
[~::J=[~}z
and
'8
Figure P3-35
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.
3-33
Find the overall transfer matrix ofthe feedback system shown in Figure 3-24(c)with
Figure P3-37
3-36
GI(s)=
s +1
O
'i' J
-1
_1_
Gz(s)=
s +1
Can you find a combined loop with a net loop gain 1 at s = oo?
3-34
Which of the systems in Figure P3-34 have improper overall transfer functions?
3-35 Can you find a combined loop with a net loop gain 1 in the system shown in Fig.
P3-35? What is its overall transfer matrix?
e(s)J
[ u(s) =
[(1 +
C(s)(
3-38
PROBLEMS
131
Figure P3-35
d
:k system.
Q =detPQ.
Figure P3-37
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,
does the assertion hold? If not, what modification do you need?
3-37 Consider the multivariable feedback system shown in Figure P3-37.
following composite transfer matrix:
Verify the
- (;(s)(1 +C(S)(;(S))-][f(S)]
(1 +C(s)G(s))-
3(s)
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)?
3-38
Show that the system in Figure 3-28(b) is not well posed by finding an improper
transfer function in the system.
3-39
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
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
S"
U(s) =
+1
1
3-43 Let (;(s) be a q x p rational matrix, not necessarily proper. Show that the rational
matrix
4-1
IntroductioDl
by direct compuation c
the equation Y(s) = G(s
matrix G(s) is a rationa
Show
,(1 +CoG o) is nonsingular.
rem 3-5.
stant matrix and improper
:ments valid?
j.
uniquelyas
4-1
Introduction
Linear systems can be described, as shown in the previous chapter, by the input
output description and the state-variable description. Once these descriptions
are obtained, the next step is natural1y to analyze them. There are two types
of analyses: qualitative and quantitative. In the qualitative analyses, we are
interested in the general properties, such as controllability, observability, and
stability, of the equations. These wil\ be discussed in Chapters 5 and 8. In
the quantitative analyses, we are interested in the exact responses of equations
due to some excitations. Digital computers are now widely available, and they
can be used to carry out these analyses (see Section 3-3). Computer solutions
however are not in closed forms. and it is difficl11t to extrapolate any propert\es
from these solutions. In this chapter we shal1 study the closed-form solutions
of the input-output description and the state-variable description. The rela
tionships between these descriptions will also be studied. We note that what will
be discussed are applicab1e to composite systems if their overal1 mathematical
descriptions are first computed.
If the impulse-response matrix G(t, r) of a system is known, then for any
input, the output y can be obtained from
y(t)=
G(t, r)u(r) dr
134
the time domain than from the frequency domain. It is elementary to compute
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
tions are stated in terms of the state transition matrix <.I>(t, r), which is the
unique solution of
o
ot <.I>(t, r) = A(t)<l>(t, r)
Theorem 4-1
The set of all solutio
over the field of real 1
Proof
Let WI and W2 be tW(
solution of (4-2) for al
<.I>(r,r)=I
In the time-invariant case, we have <D(t, r) = eA(!-t). Various methods for the
computation of eAI and (sI - A)-I are discussed. In Section 4-3 the concept of
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
01 Floquet. In the last section, the relation between linear dynamical equations
and impulse-response matrices is studied. The necessary and sufficient con
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
linear time-invariant dynamical-equation realization.
The references for this chapter are 24,31,60,68, 77, 109, 114, and 116.
d
dt (ex l W +ex
Consider
the
n-dimensional
linear
time-varying
dynamical equation
E:
[WI(t)
(state equation)
(output equation)
(4-1 a)
(4-1 b)
where A('), B('), C('), and E(') are n x n, n x p, q x n, and q x p matrices whose
entries are real-valued continuous functions of t defined over ( - 00, 00). Since
A(') is assumed to be continuous, for any initial state x(to) and any Ql!, there existe:
a unique solution in the dynamical equation E. This fact will be used frequently
in the following development. Before studying the entire dynamical equation
E, we study first the solutions of the homogeneous part of E; namely,
X =A(t)x
(4-2)
W2 1
In particular, we have
[WI(t O)
W;
Solutions of X =A(t)x
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
states in(W, ~); hence the linear space is of dimension n. This fact will be
formally established in the following theorem.
1t is clear that
is a solution 0((4-2) wi
\TRICES
135
Theorem 4-1
~lementary
to compute
e discussed further.
in Section 4-2. Solu
x. <1>(t, o), which is the
Let '!tI and '!t2 be two arbitrary solutions of (4-2). Then al VI +a2V2 is also a
solution of (4-2) for any real al and IX 2 . We prove this by direct verification:
d
d
- (IX IVl +a2V2) =IX 1 dt
dt
VI
d
+a 2 dt
V2 =a 1 A(t)Vl
+ IX 2A (t)V2
= A(t)(IX 1 VI + IX 2V2)
Hence the set of solutions forms a linear space over~. It is called the solution
space of (4-2). We next show that the solution space has dimension n. Let
el> e 2, ... , en be any linearly independent vectors in (IR n, ~) and Vi be the solu
tions of (4-2) with the initial condition Vi(t O) = e, for i = 1, 2, ... ,n. If we
show that Vi, for i = 1,2, ... , n, are linearly independent and that every solution
of (4-2) can be written as a linear combination of V;, for i = 1, 2, ... , n, then the
assertion is proved. We prove by contradiction the fact that the Ijfs are linearly
independent. Suppose that Vi, for i = 1,2, ... , n, are linearly dependent; then, by
definition, there exists a nonzero n x 1 real vector ot such that
(4-3)
Note that the O in the right-hand side of (4-3) is the zero vector of the solution
space; therefore, it is more informative to write (4-3) as
. linear time-varying
[Vl(t)
uation)
:quation)
(4-1 a)
(4-1 b)
V2(t)
Vn(t)]ot=O
In particular, we have
:l q x p matrices whose
over ( - 00, 00). Since
) and any VI, there exists
t will be used frequently
lre dynamical equation
of E; namely,
which implies that ei, for i = 1, 2, ... , n, are linearly dependent. This contra
dicts the hypothesis; hence \ji" for i = 1, 2, ", n, are Ijnearly i.ndependenl ,,'!"'"r
(-00,00).
Let V be any solution of (4-2), and let V(to) =e. Since el' e 2, ... , en are n
linearly independent vectors in the n-dimensional vector space (~m ~), e can
be wriUen as a unique linear combination of ei , for i = 1, 2, ... , n-for example,
as
n
(4-2)
e=
IXe i
= 1
It is clear that
i = 1
i= 1
IXiV(tO) =e
136
Proof
\jiU = "
1X\jIk)
i;:;; 1
This completes the proof that the solutions of (4-2) form an n-dimensional
vector space.
Q.E. D.
Definition 4-1
Example 1
*=[~ ~Jx
is a solution of
= A(t)
..y =
A(t)1Jf'
<I>(t, to)~
\TRICES
137
Proof
: that
Jrm an n-dimensional
Q.E.D.
! matrix of x = A(t)x if
iependent solutions of
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.
Again, from the uniqueness ofthe solution, we conclude that '!lO =:= O is the only
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
(~n, ~). It follows that there exist real IX, for i = 1,2, ... , n, not aH zero, such that
n
IXi'!lCtO) = O
i= 1
", "
IX'!Ik)
i= 1
IX'!I(')=:=O
i= 1
This contradicts the assumption that '!Ik), for i = 1, 2, ... , n, are linearly inde
pendent. Hence, we conclude that det 'P(t) 1=0 for aH t in (- 00, (0). Q.ED.
Definition 4-2
Let 'PO be any fundamental matrix of x
= A(t)x. Then
for all t, t o in (-00, (0)
ACt)X;
(4-Ll. )
(- 00,00).
The physical meaning of<l>Ct, to) will be seen latero Since 'P(t) is nonsingular
for aH t, its inverse s well defined for each t. From the definition we have
mmediatey the following very important properties ol the state transition
matrix:
<l>(t, t) = 1
<l> - l(t, to) = 'PCto)'P - 1 (t) = <l>Cto, t)
<l>(t z, to) = <l>Ctz, t }<l>(t 1> to)
(4-5)
(4-6)
(4-7)
138
Solutions of the dy
We use cj)(l; lo, x o, u)
X(t o) =x o and the apl
which shows the uniqueness of <I>(t, to). From Equation (4-4), it is evident that
<I>(t, to) is the unique solution of the matrix equation
Theorem 4-3
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
difficult, if not impossible, except for sorne special cases. Ir A(t) is a triangular
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
<I>(l, to)=exp [ ( A('[)d'[]
(4-9)
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
lA function is said to be continuous/y differentlab/e if its first derivative exists and is conUnuous.
dl X(l) = al <1>(
=A(l)q
=A(l)
At l
lo, we have
TRICES
139
I(,on- '
We use cP(t; to, x o, u) to denote the state resulted at time t due to the initial state
x(to) =x o and the appUcation of the input u.
Theorem 4-3
x =A(t)x +B(t)u
(4-8)
is given by
x(t)~cP(t;
l'
<I>(t, r)B(r)u(r) dr
(4-10)
to
= <I>(t, to{x o +
at <I>(t, r) = A(t)<I>(t, r)
x=
(4-11 )
<I>(r,r)=1
Proof
1)
Equation (4-11) is obtained from (4-10) by using <I>(t, r) = <I>(t, t o)<1>(to, r). We
first shw that (4-10) satisfies the state equation by direct substitution 2 :
[bY
el
a <I>(t, to)xo +a
-x(t)
=dt
at
at
(4-9)
l'
'o
<I>(t, r)B(r)u(r) dr
= A(t) [ <I>(t,to)x o +
At t
,a
1
-<I>(t, r)B(r)u(r) d!
'o at
t o, we have
'0
<I>(to, r)B(r)u(r) dr = Ix o +0 =X o
'o
In other words, (4-10) a1so meets the initia1 condition. Hence it is the s01ution.
11 meaning of the state
ltion of the state vector
I~o. <I>(t, to) is a linear
x at time t.
Q.E.D.
We consider again the so1ution given by Eqtiation (4-10). U u == O, then
(4-10) reduces to
.
(4-12)
-o
'1' (e, rl dr
oC 'o
= f(t,
rl
\ + l' -a (e, rl dr
,~,
10
0C
140
By substituting(4
4-3. The output y ca
zero-input response.
Equatian (4-16) beco
tr <I>(t, r)B('r)u(r) dr
cjl(t; to, O, u) =
(4-13)
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
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)
y(t)
(4-14)
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).
Note that Equation (4-13) can be derived directly from the fact that it is a
linear function of u. The procedure is exactly the same as the one in deriving
r G(t, r)u(r) dr
Jto
in Section 3-2. The response cjl(t; to, O, u) is, by definition, the solution of
A(t)x +B(t)u with O as the initial state. If we cut the input u into small
pulses, say
x=
u=
Unfortunately, there
and except for very :
easily faund. Theref
theoretical study af lil
of a dynamical equati
<I>(t, r). The solution (
digital camputer.
L u[t,t+)
then we have
(4-15)
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,tH) with Oas the initial state is approximately equal
to B(t)u(t)~. The input u[t,tH) outside the time interval [ti, ti +~) is identica11y
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
Time-invariant ca:
(fixed) dynamical equ.
cjl(t; to, O, u) =
tr <I>(t, r)B(r)u(r) dr
where A, B, e, and E
respectively. Since th
dynamical equatian E
applied here. We ha,
~'C(t)<D{t, to{x o +
I
(4-16)
and e At is nonsingula
In fact, e At is nonsingl
tion matrix of x = Ax
rRICES
(4-13)
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
zero-input response. If the dynamical eguation is initially in the zero state,
Eguation (4-16) becomes
t
y(t) =
, and u, respectively.
-10) can be written as
, u)
(4-15 )
solution of x = A(t)x
; approximately egual
-t,t i +8)isidentically
erned by <1>( t, ti + 8).
'e immediately obtain
[C(t)<II(t, r)B(r)
I'
G(t, r)u(r) dr
+ E(t)b(t -
r)]u(r) dr
(4-17)
'o
+ E(t)b(t -
r)
(4-18 )
x =Ax +Bu
(4-19a)
y =Cx +lElUl
(4-19b)
ation E.
(4-14 )
1411
by
+ E(t)u(t)
+ E(t)u(t)
(4-16)
142
1. Using Definition
polynomial g(A) o
then e At = g(A).
2. Using the Jordan (
where is of the J,
(2-70).
3. Using the infinite
give a closed-form
tation.
We introduce one me
we have
and
(4-22)
I
(4-23)
and
(4-24)
(4-25)
Hence, to compute l
inverse Laplace tran
inverse of a matrix i
angular 4 or of order l
the inverse of a trian~
Note that(sl-A)
many methods to con
1.
2.
3.
4.
5.
In addition, there is a
Example 2
We use methods 1 anl
(4-26)
s
-1
S
=[
(s
=
[
4 A squarematrix
;J
(s +J
is said lo be
-----------------,------~.-.-~~~_
[RICES
(4-20)
r) dr
(4-21 )
eA!:
se
se
u(t)
(4-22)
I
r)
(4-23)
~quation can also be
;e transform of (4-21)
(2-86)], we obtain
(4-24 )
+Eu(s)
(4-25)
143
[e
At
= (sI - A)-1 ,
(4-27)
Hence, to compute eA!, we first invert the matrix (sI - A) and then take the
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
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
many methods to compute it:
1. Taking the inverse of (sI - A).
2. Using Definition 2-16.
3. Using (sI - A)-1 = Q(sI - )-IQ -1 and (2-74).
4. Using Definition 2-17.
5. Taking the Laplace transform of eA!.
In addition, there is an iterative scheme to compute (sI - A) - 1 (Problem 2-39).
Example 2
. 10 -1l
Fl=Ll -2J
of a variabie. These
the dynamical equa
1. (sI-A)-1
(4-26)
=[
- 1 s
J-l = +
+2
S2
2s
+1
[s+2
-sI]
1
The transfer
uation FE.
utation of e At . 3 W'e.
:unctions of a matrix.
1 (4-23).
(s
+ 1)2
(s
+ 1)2J
A square matrix is said to be triangular if al! the elements below or aboye the main diagonal are zero.
..
144
2. The eigenvalues of A are -1, -1. Let g(A) =0:0 +0: 1,1.
= g(A) on the spectrum of A, then
f( -1) =g( -1):
j'(-1)=g'(-l):
Hence
Iff(A)~(S-A)-l
8
(s +1)-2=0: 1
and
(s
+ 1)2J
Example 3
Figure 4-1
eA(t-tlBu(r) dr
(4-28)
The matrix function eA! can be obtained by taking the inverse Laplace transform
of (si - A)- l. which is computed in Example 2. Hence
S
(s
[
+2
+ 1)2
(s~w
and
x(t
)=L1(1 +t)ete -/
TrajectoI
If aH eigenval ues
can be written as, by
The sol utions with u == O are plotted in Figure 4-1 for the initial states
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
(see Definition 2"15),5 we claim that every element of eA! is a linear combination
ofthe factors tkeAt, for k = 0,1, ... ,ni - 1; i = 1,2, ... , m. Let be a Jordan-form
representation of A and A = QQ - 1. Then eA! = QeA!Q - l. From (2-69) we
5If aH eigenvalues are distinct, then the index of every eigenvalue is 1. If an eigenvalue has mull-o
plicily k, then ilsindex could be 1,2, ... , k -1, or k. See Seclon 2-6 and Definition 2-15.
If Xo s chosen so tha
'RICES
145
A. If f(A)6 (5 - A)-I
8
-----!L--+---+--+--+-+--+--1--1--- xl
Figure 4-1
(4-28)
-te(l-t)e- t
Trajectories.
know that every element of e Al is ofthe form tkeAt, for k =0,1, ... , - 1; i = 1,
2, ... ,m. Hence every element of e At is a linear combination of these factors.
We call tkeAt a mode of the dynamical equation FE in (4-19).
The responses of a linear time-invariant dynamical equation are dictated
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
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
its index is 2 or higher, then its mode will approach infinity at the rate of
t-I.
')Je -(t -
-(t-r)
rl
x(s) = (sI - A)
A
J U(T) dT
-1"
o = L,
X
1 Qi1ll1X
o
s -- /"'i
--1
; a linear combination
,et . be a lordan-form
l.
From (2-69) we
(4-29)
(4-30)
lf Xo is chosen so that
pXO
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
eigenvectors of A. The situationis much more complicated and will not be
discussed.
146
4-3
equation
FE:
X =Ax +Bu
y=Cx +Eu
(4-31a)
(4-31b)
FE;
where
A =PAP-
x=x +Bu
(4-32a)
y=Cx+Eu
(4-32b)
e =CP-
B =PB
The dynamical eq
describe the same sys
formation between ti
clear that X =x. S
we have xz=(x-x
(4-33)
Example 1
or
Figure 4-2
A networl
"RiCES
iynamical equations.
,ne of change ofbasis
variant case and then
Id q x p real constant
t vector, and x is the
lation is an n-dimen
r. We have agreed in
... , "ti} as the basis
Lat its ith component
; the basis of the state
~ the basis of the state
r to aIlow a broader
ction extend the field
der the state space as
ization is needed in
ent Jordan-form dy
(4-32a)
147
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
lence transformation.
We explore now the physicaL meaning of equivalent dynamical equations.
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
analysis often lead to different choices of the state.
Example 1
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
ables, then the dynamical equation description of the network is
(4-34a)
(4-34b)
If, instead, the loop currents Xl and Xl are chosen as the state variables, then
the dynamical equation is
(4-35a)
(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
formation between these two equations can be found from Figure 4-2. 1t is
clear that Xl =Xl' Since Xl is equal to the voltage across the l-ohm resistor,
we have Xl =(Xl -Xl)' Thus
(4-32b)
(4-33)
(4-36)
or
Xl
Figure 4-2
148
It is easy to verify that the dynamical equations in (4-34) and (4-35) are indeed
related by the equivalence transformation (4-36).
11
Definition 4-4
Two linear dynamical equations are said to be zero-state equivalent if and only if
they have the same impulse-response matrix or the same transfer-function
matrix. Two linear dynamical equations are said to be zero-input equivalent
if and only if for any initial state in one equation, there exists a state in the other
equation, and vice versa, such that the outputs of the two equations due to
zero input are identical.
11
Note that this definition is applicable to linear time-invariant as well as
linear time-varying dynamical equations.
Example 2
Theorem 4-5
Example 3
Proof
The impulse-response matrix of FE is
G(t) = CeA1B
lfthe matrix A in a
tion is said to be a lord
+ Eb(t)
+ Eb(t)
eA' =
and
B = PB,
PeA'P- 1
1n
1
y
u(
-_Y-,
(a)
Hence, for any x(to), if we choose x(t o) = Px(t o), then FE and FE have the same
zero-input response.
Q.E. D.
We note that although equivalence implies zero-state equival~nce and zero
input equivalence, the conve;seis. not true. That is, two .linear dynamical
equations can be zero-state equivalimt and ze~o-input equivdlent witho~t being
equivalent. Furthermore, two linear equations can be zero-st~e equivalent
without being zero-input equivalent.
0.5
t
0.5
u(
(a)
TRICES
e-invariant as wel! as
149
Example 2
Consider the two networks shown in Figure 4-3. If the initial state in the
capacitor is zero, the two networks have the same input-output pairs; their
impulse responses are al! equal to b(t), the Dirac b-function. Therefore, they
are zero-state equivalent, or more precisely, their dynamical-equation descrip
tions are zero-state equivalent. Because of the symmetry of the network, for
any initial voltage in the capacitor, the outputs of these two networks due to
u =0 are al! identical!y zero. Hence, they are zero-input equivalent.
A necessary condition for two dynamical equations to be equivalent is that
they have the same dimensiono The dynamical-equation description of the
network in Figure 4-3(a) has dimension O; the dynamical-equation description
ofthe network in Figure 4-3(b) has dimension 1. Hence, they are not equivalent.
Example 3
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
u
AIB
+ Eb(t)
In
e equivalent.
1t
y
In
-'--
In)'
In
x
In
---"'~-
(a)
---
(b)
Figure 4-3 Two networks whose dynamical equations are zero-state equivalent and
zero-input equivalent without being equivalent.
o)
Q.E.D.
0.5
0.5
t
I F
(a)
In
In
(b)
Figure 4-4 Two networks whose dynamical equations are zero-state equivalent.
150
2-6 that every operator that maps (Cm C) into itself has a lordan-form matrix
representation. Hence every linear, time-invariant dynamical equation has an
equivalent Jordan1orm-dynamical equation.
Example 4
E +CBS-l +CABs- 2
FE:
(4-37a)
*Time-varying caSE
(4-37b)
5]-1 = [1
-5]
-3
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
obtain immediately the following equivalent lordan-form dynamical equation:
O 0, 2
X3
Definition 4-5 6
Let P() be an n x n rr
and P(t) are nonsingul
dynamical equation
-1-2
[;}[: ~ 4:j
where
is said to be equivalent
be an equivalence tran~
Theorem 4-6
Two linear time-invariant dynamical equations {A,B, L, E} and {A, B, C,E}, not
necessarily of the same dimension, are zero-state equivalent or have the same
transfer-function matrix if and only if E = E and
i =0,1,2, ....
is a fundamental matri:
Proof
Theyare zero-state equivalent if and .only if
E + C(sI - A)-1 B = E + C(sI---'A)-IB
151
~ATRICES
, a J ordan-form matrix
...
i =0,1,2, ...
This establishes the theorem.
(4-37a)
(4-37b)
Q.E.D.
-5]
E:
x= A(t)x + B(t)u
y = C(t)x + E(t)u
-3
(4-38a)
(4-38b)
Definition 4.5 6
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:
ie this subsection.
where
(4-39a)
x=(t)x+"8(t)u
y = qt)x + lt(t)u
A(t) =
B(t) =
~(t) =
E(t) =
(4-39b)
(4-40a)
(4-40b)
(4.-40c)
(4-4001)
PO is said to
q(t)~ P(t)'JI(t)
(4-41 )
152
matrix of E; hence 'iJ(t) = A(t)'P(t), and 'P(t) is nonsingular for all t. Conse
quently, the matrix P(t)'P(t) is nonsingular for aH t (see Theorem 2-7). Now
we show that P(t)'P(t) satisfies the matrix equation =A(t)x. Indeed,
di (P(t)'P(t)) =P(t)'P(t)
Definition 4-6
A matrix P() is calle,
tinuous and bounded
+ P(t)'P(t)
x= Ax.
Theorem 4-7 7
Let 'P(t) be an fundamental matrix of x = A(t)x. That is, 'P(t) is nonsingular for
~1I t and satisfies 'iJ(t) = A(t)'P(t). The differen~iation of '11 - 1 (t)'PJt) = 1 yields
'P- 1(t)'P(t) + '11 - 1 (t)'P(t) = O, which implies '11 -l(t) = - 'P- 1(t)'P(t)'P- 1(t) =
- 'P- 1(t)A(t). We define, in view of (4-41),
Because of (4-43) a
of P- 1(t) = - P-l(t)PI
Consequently, we car
P- 1 (.). Clearly a nor
In Definition 4-5, if P(
are said to be equivale
preserves, as will be (
namical equation, bu
required to be a Lya
general. In other wo
equivalent in the sense
This is possible for the
be discussed in the foL
Aot
Q.E.D.
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,
then P(t) = 'P- 1 (t) and (4-40) becomes
B(t) = q;-l (t)B(i)
CU) = qt)\f(t)
E(t)
= lE(i)
(4-42)
Its block diagram is plotted in Figure 4-5. Unlike the one in Figure 3-11, there
is no feedback in Figure 4-5.
7
for aH
=A o
A(i) = ir
Linear time-varying e
This theorem was pointed out to the author by Professor T. S. Kuo in 1972.
matrix of x = A(t)x. 1
Indeed we have
'P(t .
e
ECt) ~=========;l
Define
+ y
We show that
E in (4-42).
PO is a 1
P(t +T) =e
rRICES
153
Definition 4-6
A matrix PO is called a Lyapunov transformation if (1) PCt) and P(t) are con
tinuous and bounded on [to, 00) and (2) there exists a constant m such that
O<m<\detP(t)1
forallt2t o
(4-43)
'P(t)
Q.E.D.
(4-42)
n 1972.
A(t)
for all t and for sorne positive constant T. This means that every element of
AO is a periodic function with the same period T. Let 'PCt) be a fundamental
matrix of x = A(t)x. Then 'P(t + T) is also a fundamental matrix of x= A(t)x.
Indeed we have
E(O = JE(t)
+ T) =
+ T) =
'PCt)Q
(4-44)
~or the nonsingular matrix Q there exists a constant matrix such that
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)
_........
,'-
., ..
_~
---
"-~---~-
154
Theorem 4-8
Assume that the matrix A in the dynamical equation E in (4-38) is periodic with
period T. Let P be defined as in (4-46). Then the dynamical equation E in
(4-38) and the dynamical equation
E:
Jo
(4-47)
E:
(4-,fJ.a~)
(4-48b)
Jo
t
r [C(t)<I>(t, T)B(T) +E(t)b(t-T)]U(T) dT
Jo
(4-49)
G(t, T) =
A q x p impulse-respo
linear dynamical equa
composed into
G(t, T)
where E is a q x p mat
tinuous matrices of t.
TRICES
of Lyapunov.
; hence it is bounded.
e, PO is a Lyapunov
from Theorem 4-7.
led theory of Floquet.
then its fundamental
ic function. Further
to =i.
mical Equations
e relation between the
Let the input-output
out terminals be
(4-47).
B, e, and E are n x n,
tinuous functions of t
~-48) are two different
.me input-output pairs
lmical equation E with
where <I>(t, T) is the state transition matrix ofi = A(t)x. By comparing (4-47) and
(4-49), we immediately obtain
G(t, T) = {~(t)<I>(t, T)B(T)
(4-49)
+ E(t)(j(t -
T)
for t
for t <
T
T
(4-50)
That G(t, T) = O for t < T fol\ows from the causality assumption which is implic
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
description of the system can be easily obtained from (4-50). The converse
problem-to find the state-variable description from the input-output descrip
tion of a system-is much more complicated, however. It actual\y consists of
two problems: (1) Is it possible at al\ to obtain the state-variable description
from the impulse-response matrix of a system? (2) If yes, how do we obtain the
state-variable description from the impulse-response matrix? We shal\ study
the first problem in the remainder of this section. The second problem will be
studied in Chapter 6.
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
response matrix, then G(t, T) is said to be realizable. We cal\ the dynamical
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
dynamical equation, we can build an operational amplifier circuit that will
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
not have any physical meaning. Note also that the dynamical-equation realiza
tion gives only the same zero-state response of the system. If the dynamical
equation is not in the zero-state, its response may not have any relation to the
system.
If the realization of an impulse response G(t, T) is restricted to a finite
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
the impulse response l/U - ,). We give in the [oliowing the necessary ana suf
ficient condition for G(t, T) to be realizable.
Theorem 4-9
A q x p impulse-response matrix G(t, T) is realizable by a finite-dimensional
linear dynamical equation of the form (4-48) if and only if G(t, T) can be de
composed into
.
..
G(t, T) = E(t)<5(t - T)
T) dT
155
+ M(t)N(T)
for al\ t T
(4-51 )
156
Proof
E:
y = C(t)x. +E(t)u
Time-invariant ca~
and
G(t -r
Sujficiency: Let
G(t, r) =E(t)o(t -r) +M(t)N(r)
E:
(4-52a)
(4-52b)
x(t) = N(t)u(t)
y(t)=M(t)x.(t) +E(t)u(t)
r -re-A<l
teAtll
~
.L'
Le'" J
E:
As discussed in (3-}
matrix, and C(sI - At 1
function matrix of the d
Theorem 4-10
A transfer-funciion ma
time-invariant dynamic~
[RICES
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.
Time-invariant case. We shall first apply Theorem 4-9 to the time-invariant
case and see what can be established. For the time-invariant case, we have
G(t, r) = G(t - r). Consequently, Theorem 4-9 can be read as: An impulse
response G(t - r) is realizable by a finite-dimensional (time-varying) dynamical
equation if and only if there exist continuous matrices M and N such that
T)
; completed by identi
G(t - r) = M(t)N(r)
)B(t)
~spectively.
Then the
(4-52a)
forallt;:::r
r)
There are two objections to using this theorem. First, the condition is stated in
terms of G(t - r) instead of G(t). Second, the condition is given for G(t - r)
to be realizable by a linear time-varying dynamical equation. What is more
desirable is to have conditions on G(t) under which G(t) has a linear time
invariant dynamical-equation realization. Since, in the time-invariant case,
we may also study a system in the frequency domain, we shall first derive the
condition of realization in terms of transfer-function matrix, and then state it in
terms of G(t).
Consider a system with the input-output description
(4-52b)
y(t) =
latrix of E is an n x n
Q.E.D.
+ E(t)(t -
G(t - r)u(r) dr
be simulated without
ld tU) =M(t).
FE:
lfy that
,-A<l
t
x=Ax +Bu
y=Cx +Eu
(4-54a)
(4-54b)
By taking the Laplace transform and assuming the zera initial state, we obtain
y(s) = [C(sI - At 1 B
+ E]u(s)
(4-55)
ame impulse-response
,btain difIerent realiza
Note that an impulse-
+ E
(4-56)
158
Proof
/3I S/l-I+"'+/3/1-IS+/3/1
s/l
O
1
O
O
x/I_ I
-ct. n - 1
-0:/1-2
XI
O
O
O
O
XI
X2
+
X/I
-CJ."
y=[/3/1
/3/1-1
/311-2
-0: 2
/32
/3IJX
X/I_ I
-(;(1
X/I
+eu
e +~s !/3z/s
1 +afs +(;(2.
(4-57)
X2
O
1
(4-58a)
(4-58b)
and let
be a realization of 9j, f(
that the bJs are colum
posite dynamical equati
f10w graph. 8 ,9 Let us choose XI, XI, X2' X2"'" X/I' X/I as nodes; then the signal
f10w graph of (4-58) is of the form shown in Figure 4-6. There are n loops with
loop gain -afs, -a 2/s 2, ... , -aJs/l; and there are, except the direct trans
mission path e, n forward paths with gains /3ds, /32/S2, ... , /3Js/l. Since all the
8
Mason's gain formula for signal-flow graph is as follows: The transfer fllnction of a signal-flow
graph is
where = l-(I: all individual loop gains) +(I: aH possible gain prodllcts of two nontouching
loops)-' .. ; g = gain of the ith forward path, and ; = the part of nol louching lhe ith forward
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
Figure 4-6
is a realization of (;(s).
159
.TRICES
~-invariant
dynamical
oper rational matrix.
is realizable, we first
realizable. The most
(4-57)
namical equation
Xl
X2
/!,. = 1 +-
a2
an
++ ... +S2
sn
(4-59)
and /!,.i = 1, for i = 1, 2, ... , n. Hence, the transfer function of (4-58) from the
input u to the output y is
e+
+ ... + (Jn
=g(s)
1 + ... +a n
A
This proves the assertion that every scalar proper rational function is realizable.
We are now ready to show that every proper rational matrix is realizable.
In order to avoid cumbersome notations, we assume that (;(s) is a 2 x 2 matrix.
Let
u
O
1
(4-60)
(4-58a)
(4-58b)
he transfer function of
's formula for a signal
lodes; then the signal
There are n loops with
~cept the direct trans:
.. , (Jjs". Since all the
er runction or a signal-flow
and let
be a realization of gj, for i,j = 1,2; that is, gJs) =cJsI -AF 1bu +eu' Note
that the bds are column vectors and the cL/s are row vectors. Then the com
posite dynamical equation
[~"j C'
12 _
X21 -
O
O
22
[~J=[C~l
O
A 12
O
O
O
O
A 21
O
12
C21
O
O
A 22
O
C22
X 12
X 21
X 22
["]
X 12
X 21
O
b 21
O
el!
+[e 21
(4-61 a)
b 22
1
12
e
e 22 U2
J[u J
(4-61 b)
X 22
~
~Y
!ln
(4-62)
4-58).
----
--------~---
------- -- =--=---=--=-=---=----=--==-=---,.--,,-=--=--=--=-----=--=-
=--~-,-===========-~--
160
~-~--
6.
The condition ofTheorem 4-10 is stated in terms of transfer-function matri
ces. We may translate it into the time domain as folIows:
Corollary 4-10
An impulse-response matrix G(t) is realizable by a finite-dimensional linear
time-invariant dynamical equation if and only if every entry of G(t) is a linear
combination of terms of the form tke Ail (for k = 0, 1, 2, o.. , and i = 1, 2, o. o) and
possibly contains a b-function at t =0.
I
The impulse-response matrix G(t) is the inverse Laplace transform of the
transfer-function matrix (;(s). If a proper rationaI function is of the form
N(s)
g .(s) = d +
.,.---:..---,----,-
2
(s - Ad'(s - A2
IJ
then its inverse Laplace transform %(t) is a linear combination of the terms
,-------------,
I
"':
iT~l
:
:
;
L
:
I ~r
-~.~
I
, 1---
:""'
lA
g,2(S)
i
I
I
I
fl'
Y,
"2
I
r~
:Y2
~
~
+rc--'--
:
...J
Since a realizable
since entries of 'G(l)
i = 1, 2, ... , the entri<
ConsequentIy, the mal
real lineo As a consec
Theorem 4-11
A system that has a pr,
only if U[lo,lo+EJ =0 in
This is a restateme
have rational transfer
linear time-invariant e
an interval (no matter
interval.
4-5
Concluding
CONCLUDING REMARKS
TRICES
te-dimensional linear
Q.E.O.
is simpleand straight
matrix independently
: shown in Figure 4-7.
~tory for the following
:d as can be seen from
owever, is a feature of
sion of the realization
be resolved in Chapter
161
ansfer-function matri
s:
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
linear time-invariant dynamical equations-if the output is identically zero in
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
., and i = 1, 2, ...) and
4-5
10.
Concluding Remarks
162
dynamical equations and the realization problem are identical to the con
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.
4-7 Given
show that
d
Problems
o
o,
Find the fundamental matrices and the state transition matrices of the following
homogeneous equations:
4-1
4-8 Givenx(t)=A(t)x.
where A* is the complex '
transition matrioes of x =
and
4-9
-4
-3
-1
Consider
1
2
with
uU) =
4-4
[~J
for t 2':0
Let
d
-"[Tl(t)]
de .
= -
l(t) [el
- TU) ] r
elt ..
l(t)
Figure P4-9
PROBLEMS
.RICES
163
4-7 Given
show that
det <I>(t, to)=exp [ [ (at(r) +adr))dr]
where 8<1>(t, to)lat = A(t)<I>(t, to) and <I>(to, to) = I. Hint: Show that
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
:X
A(t)x
+ B(t)u
y =C(t)x
z= -A*(t)z +C*(t)v
w =B*(t)z
shown in Figure P4-9. Note the reversal of the flow direction of signals. Let G(t, r) and
Gu(t, r) be their impulse response matrices. Show that
G(t, r)
= G:(r, t)
Figure P4-9
164
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 theijth element of <1>(t, to)?
4-11
In
Let
Figure P4-17
X=[A~
:::}
Show that <1>2(t, to) =0 for all t, t o and that (a/at)<J>;;(t, to) = A;<1>;(t, to) for i = 1, 2.
4-12
4-18
Find an equivalen
4-19
Find an equivalen"
tlt
B(to)=B o
4-14
X(O)=C
4-20
Find an equivalen!
X i (11 -t
x 2(11
[
Show that if (t) = AA(t) - A(t)A, then a fundamental matrix of x = A(t)x is given
-t
X 3(11 -t
by
where A2~A(O)-A.
Find the impedance (the transfer function from u to i) of the network in Figure
P4-16. lf the initial conditions of the inductor and the capacitor are zero and if an input
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?
4-16
4-17
4-21
Consider the network shown in Figure P4-17. Find the initial inductor current
4-23 Find a dynamical-e(
find a linear time-invarianl
2n
Use a signal-ftow g
variable, linear time-invari
4-24
u 'V
4H
Figur 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.
(t, to) for ; = 1,2.
4-18
1~][::]+[-~]
20
-25
4-20
X3
1)] [O
x(n+l) =
[
-20
x3(n
+ 1)
3][X1(n)] + [-1]
20
16
x(n)
3 u(n)
-25
-20
x3(n)
y(n)=[ -1
OJ
XI(nl]
x(n) +4u(n)
[x (n)
3
4-21
Use a signal-f1ow graph to show that the transfer function of the following singlevariable, linear time-invariant dynamical equation
4-24
X~ [
y=[O
O O
O O
1 O
[~.]
O -"o
-ct,,_l
13U-l .
O -~n- x + P";-2 u
O O
-!XI
O O
O lJx+eu
PI
166
is
Verify tha t
{31 S ,,-1
g(s)=e
4-25
+ S" +O:IS"-1
+ ... +{3"
S+2
s +1
s +3
1
Ss
s +1
+1
s +2
2+2
+1
+3
1
5
+1
+1
52
+2
4-31
Show that if
A(e
lO
(
is
:i= -A*z+C*u
{ y=B*z
(la)
(lb)
z= ::-A*z + C*u
{y=B*z
(2a)
(2b)
where A* and * are the complex conjugate transposes of A and , respectively. Show
that Equations (1) and (2) are equivalent and that they are related by z= (p-l )*z.
4-27
Consider x(k
+ 1) = A(k)x(k).
Define
for k>m
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?
4-29
+I
<1>(k, 1 + 1)B(l)u(l)
l=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 -
m=O
4-30
(d
III
Bu(m)
PROBLEMS
ArRICES
Verify that
~ (A(t)B(t)) = A(t)B(t)
+ A(t)B(t)
dt
;j
if and only if A(t) and A(t) commute; that is, (t)A(t) = A(t)(t).
4-31
Show that if
A(r) dr and A(t) commute for all t, then the unique solution of
'o
o
oc
<1>(to, to) = 1
is
<1>(t, to)=exp
f
lO
(2b)
d f,., respectively. Show
an
-1 *
edbyZ==(P )Z.
for k>m
.IS g'ven by x(k) = <1>(k, m)xo
)11(1)
state at time (1
ecorn es
+ l).J Show
A(r) dr
167
5
Controllabi Iity and
Observability of Linear
Dynamical Equations
5-1
Introduction
Xl
er'
Dampi.ng
coefficien!
u
(curren!
service)
Figure5-1
,11
Spring
constan!
A simple network.
168
INTRODUCTION
169
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
theorem is also developed. In Section 5-5, we study dynamical equations
which are uncontrollable and/or unobservable. The canonical decomposition
theorem is developed. A consequence of this theorem is that the transfer
function matrix of a. dynamical equation depends solely on the part of the
equation that is controllable and observable. In Section 5-6, we study the
controllability and observability of linear time-invariant Jordan-form dynami
cal equation. Their conditions are very simple and can be checked almost by
inspection. In Section 5-7, the concepts of output controllability and output
function controllability are introduced. It is shown that they are properties
of the input-output description of a system. In the last section, we discuss
sorne computational problems encountered in this chapter.
Although elements of the matrices A, B, e, and E are all real-valued, for
mathematical convenience they are considered as elements of the field of com
2U
.'":';'
Damping
coefficient
Spring
constant
1
Xl
x2
Damping
coefficient
J
Spring
constant
1
170
5-2
foral/tin [tI' t 2]
(5-1)
implies
IX =
0, where
J~(t)=t
for / in [ -1, 1]
12(t) = { t
-t
for
for
in [0, 1]
in [ - 1, O]
Clearly, Ot is a constan
The linear indepen
an interval; hence in t
entire interval. Let F'
Theorem 5-1
lt is clear that the functions(1 and.f2 are linearly dependent on [O, 1], since if
we choose al = 1, a2 = -1, then aj"l(t) +ad1.(t) =0 for aH I in [0, 1]. The
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
pendent on an interval, it is not necessary that they are linearly independent on
any subinterval. However, it is true that there exists a subinterval on which
they are linearly independent. For example, in Example 1 the functionsJl and
J2 are linearly independent on the subinterval [ -e, e] for any positive eo On
the other hand, if a set of functions is linearly independent on an interval [t l ' t 2],
then the set of functions is linearly independent on any interval that contains
[1 l ' t 2 ].
Then f l , f 2 , .. o. f are
constant matrix W(t l '
TJe functions we study are mostly continuous functions: hence there sno sUDstantia\ difreorence
between using the open intervallt ,. e2) and the closedinterval [e l' e2] Every [e ,. e2 ] s assumed to
be a nonzero interval.
ll
Proof
The proof of this theore
is nonsingular ir and (
We prove first the ne
Assume that the f/s are
Then there exists a nOI
in this chapter.
.__
._~.:=~---.--_.-':::_
ICAL EQUATIONS
limensional dynamical
:tor space (Cn, C).
),21, 48,55, 56, 60, 61,
171
(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
only if
...
'Os
(5-1)
(5-3)
...
an
F"r~:l
lf..J
ed by
Theorem 5-1
re is no substantia\ difference
Every [t,. t,] is assumed to
W(t].
12)~ r"
Jt
F(t)F*(t) dI
11
11
Proof
The proof ofthis theorem is similar to that ofTheorem 2-8. RecaH that a matrix
is nonsingular if and only if all the columns (rows) are linearly independent.
We prove first the necessity of the theorem; we prove it by contradiction.
Assume that the f/s are linearly independent on [t l' t lJ, but W(t l ' t 2) is singular.
Then there exists a nonzero 1 x n row vector t:x such that t:xW(t, tl)=O. This
t ,)
..
_-~---~
=. . .=-.--=..
--'="--==.C'-...
172
,_U
. , , _ . _ _. _ _ - '_ _-_ _ - , - _
_ "~
'2
exW(t, t 2 )ex* =
1'1
(exF(t))(exF(t))* dt =0
(5-4)
exF(t) =0
Hence we have
r exF(t)F*(t)dt=O
ex[F(t) : F
in particular,
t2
exW(t l ,t 2 )=
(5-5)
J"
has rank n, then the f's are linearly independent on [t J, t 2J over the field of
complex numbers. 3
Proof
We prove the theorem by contradiction.' Suppose that there exists some t o in
[tI' t 2 J such that
IS
and
CAL EQUATIONS
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
(5-4 )
=0
exF(t) =0
This irnplies that
on and is nonnegative
exF(kl(t) = O
Hence we have
ex[F(t) : F(l)(t) : ... : F(ll- ll(t)]
f the set of f j , i = 1, 2,
], then det W(t l , t 2) 1= O.
)pose that W(t l' t 2) is
l. Then, by definition,
1 that exF(t) =0 for aH
in particular,
ex[F(to): F(ll(t O) : ... : F(ll-l)(t O)] =0
which irnplies that aH the n rows of
[F(to) : F(l)(t o) : ... : Fln-l)(to)]
(5-5 )
)nsingular. Hence, if
lependent on [ti' t 2 J.
Q.E.D.
Q.E.D.
The condition of Theorern 5-2 is sufficient but not necessary for a set of
functions to be linearly independent. This can be seen from the following
exarnple.
Example 2
.. , f n have continuous
:t F be the n x p matrix
F. If there exists sorne
to)]
l'
IJIU) f"?l(tll
(S-S)
p Lf2(t)
fl(t)
p [ f2(t)
there exists sorne t o in
defined as
F(k)(t),
with
t3
It 3 1
defined over [ -1, 1]. They are linearly independent on [ -1, 1]; however,
and
,rt
J?)(t)J = ; Lt 3
) IS
f?)(t)J= [ t
fPl(t)
p _t 3
fl(t)
p [ f2(t)
3t 2 1
3t 2 = 1
3t J=1
-3t 2
j<l)(t)J
fP)(t) =0
rO"
aH L in lO,
1J
"
..
---~------"'---------------
174
analytic, then we can use Theorem 5-3, which is based on the fact that if a
function is analytic on [tI' t 2 J, then the function is completely determinable
from a point in [t 1 , t 2 J if aH the derivatives of the function at that point are
known. (See Appendix B.)
every subinterval of
essential. The staten
Note that Theore
Theorem 5-3
Assume that for each i, f i is analytic on [t l' 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. Let t o be any fixed point
in [t l' t 2]. Then the f/s are linearly independent on [t l' t 2J if and only if
Example 3
Let
(5-7)
Proof
Then
p[FU o) :
'J < n
It is easy to verify th
"'J
The 's are analytic on [t 1 , t 2 J by assumption; hence there exists an 6>0 such
that, for aH t in [to -6, t o +6J, F(t) can be represented as a Taylor series about
the point t o:
~o)" F(II)(t O)
for aH t in [to - 6, to +6J
n.
Premultiplying a. on both sides of (5-9) and using (5-8), we obtain
F(t) =
(t -
(5-9)
11=0
a.F(t) =0
(5-10)
a.F(t) =0
for al! t in [t 1, t 2 J
t 2 ].
[F(t) : 1
This is a contra
Q.E.D.
=p [
sin lO't
'
sin 2 x lO't : 2 x 1
for al! t.
The matrix in (5-~
many cases, it is not no
Example 3, we check
Corollary 5-3
Assume that, for each
independent on [t b t 2
for almost all t in [t 1> j
This corollary wil
omitted.
5-3
Controllabil
Time-varying case
trollability of linear d
state controllability o
o\L EQUATIONS
175
Example 3
Let
(5-7)
_ [sin 1000tJ
F(t) - sin 2000t
Then
~orem
5-2.
'pose that
Now we
[F(r) : F(l)U)]
<n
=.
O,
10 - 3n;, . . ..
However,
sin 10 3 r
sin 2 x 10 3 {
10 3 cos 10 3 r
2 x 10 3 cos 2 X 10 3 r
-10sinI0 3 r
'
-10cosI0 3 r
3
-4xI0.sin2xl0 r :-8xI0 9 cos2xI0 3 r =2
for all t.
=O
(5-8)
The matrix in (5-7) has n rows but infinitely many columns. However, in
many cases, it is not necessary to check aH the derivatives of F. For instance, in
Example 3, we check only up to F(3 l . lf we use the matrix
e, t o +eJ
(5-9)
e obtain
(5-10)
Corollary 5-3
... ,
f n are linearly
5-3
=n
Controllability of Linear
Dyn~mical
Equations
176
x=
A(t)x(t)
+ B(t)u(t)
(5-11 )
where x is the n x 1 state vector, u is the p x 1 input vector, and A and B are,
respectively, n x n and n x p matrices whose entries are continuous functions
of t defined over ( - 00,00). The state space of the equation is an n-dimensional
complex vector space and is denoted by ~.
Definition 5-1
The state equation E is said to be (state) controllabl 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
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
should take is not specified. Furthermore, there is no constraint imposed on the
input. Its magnitude can be as large as desired. We give some examples to
illustrate this concept.
Example 1
Consider the network shown in Figure 5-3. The state variabk x ofthe system is
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
has no effect on the voltage across the capacitor. Hence the system-or, more
x(t) = I(t;
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
+
11
~.
rv
In
In
IF;::
1.11:
:AL EQUATlONS
(5-11 )
precisely, the dynamical equation that describes the system-is not controllable
at any too
Example 2
Consider the system shown in Figure 5-4. There are two state variables X and
Xz in the system. The input can transfer X al' Xz to any value; however, it
cannot transfer x and Xz to any values. For example, if X(t o) 0:=0, xz(to)=O,
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
The solution of the state equation E with x(to) =
177
l'
Xo
is given by
<1>(t, L)B(-r)U(L) dL
[o
1I>(t, t o{ Xo +
(5-12)
A(t)x and is
Theorem 5-4
The state equation E is controllable at time t o if and only if there exists a fmite
ti> t o such that the n rows of the n x p matrix function <1>(t o,')B(') are linearly
independent on [to, ti].
Proof
Sufflciency: Ir the rows of <1>(t o, )B(-) are linearly independent on [t o, ti], from
Theorem 5-1 the n x n constant matrix
(5-13)
~~o
1F
Xl
+tz
X
-=.t
178
(5-12), we obtain
. [x o -<I>(t o, t)XJ}
= <1>(t, (0){Xo - WU o I)W-1U o. ()[x o -<1>(to, ()xJ}
= <1>(1 , 10)<1>(t O (I)x I
=x
with
In the following we sl
A and B without solv
need sorne additional ;
times continuously di
M(), ... , by the equ:
Mo(t) = I
Observe that
$(t o, t)B
ot $(t o, t)B
(5-15 )
r"
10
(5-16 )
and, in general,
r"
10
Ok
ot k $(t o, c:
a.<1>(to. r)B(r)u(r) dr
(5-17 )
Theorem 5-5
n - 1 times continuom
lable at t o if there exist
a.a.* =0
which, in tum, implies that ex = O. This is a contradiction.
p[r
Q.E.D.
Proof
In the proof of this theorem, we also give in (5-14) an input u(t) that transfers
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 o 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
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.
If the total energy
r Ilu(1)111 dt
Jea
Define
[$(to, t)B(t) :
O~l $(t
ll
is used as a criterion, the input given in (5-14) wiII use the minimal energy in
transferring x(t o) to Xl at time tI' Thisisestablished in Appendix C.
In order to apply Theorem 5-4, a fundamental matrix 'P or the state transi
tion matrix <l>{t, r) of:x: = A(t)x has to be computed. As we mentioned earlier,
this is generally a difficult task, Hence, Theorem 5-4 is not readily applicable.
p[~
implies
p [$(t o, tl)B(t) :
CAL EQUATIONS
(t o, t )
. [X O- C1>(tO, t }XJ}
179
k = O, 1,2, ... , n - 1
Mo(t)=B(t)
(S-18a)
(S-18b)
Observe that
<I>(to, t)B(t) = <I>(to, t)Mo(t)
(S-19a)
(S-1S)
;omes
) dr
(S-19b)
(S-16)
and, in general,
ak
k=O,I,2, .. ,n-l
(S-19c)
(5-17 )
Theorem 5-5
state-in particular,
I(tO, t)B(t) = O, for aH t
Assume that the matrices A(') and B(-) in the n-dimensional state equation E are
-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
n
(S-20)
n.
Q.E.D.
Proof
Define
"'\
- ,-,,1.
""\lo, ',)1lJl(
j;]) LJ
at
L
~'a'---
1=1,
0'
'
, M\Li
an
(S-21 )
180
It follows from Theorem 5-2 that the rows of $(t o, .)B(-) are linearly independent
on [t o, t 1 ] for any t 1 > t. Thus, from Theorem 5-4, we conclude that the state
equation E is controllable.
Q.E.D.
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
Consider
Xl]
[X3 =
~z
[t
Xz
X3
1 u
1
controllability. The (
obtained by slight mo
matrices A and B in the
the following theorem.
p[Mol
(5-22)
If the matrix A is an
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).
Defintion 5-3
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
The remainder of this subsection will be devoted to three distinct types of Gon
trollability. This material may be omitted with no loss of continuity.
Ir a dynamical equa
the states can be achie'
consists of b-functions :
that instantaneous co
most important implic:
of a single input, the
equivalence transforma
form equivalent dynaI
controllable dynamical
Definition 5-2
A state equation E is said to be differentially (completely) controllable at time t o
if, for any state x(to) in the state space L and any state Xl in L, there exists an
input u that will transfer x(t o) to the state Xl in an arbitrarily small interval of
~~
If every state in L can be transferred to any other state in a finite time (no
matter how long), the state eqation is said to be controllable. If this can be
achieved in an arbitrarily small interval of time, then the state equation is said
to be differentially controllable. Clearly, differential controllability implies
MICAL EQUATIONS
181
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),
11
(5-22)
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
ilitv,
y) controllable at time t o
: Xl in L, there exists an
itrarily small interval of
I
11
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
we adopt the terminology"instantaneous eontrollability."
182
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
-3
for all t in ( - 00, (0), where <l> is the state transition matrix and W is as defined
in Equation (5-13).
I
By A> B, we mean that the matrix (A - B) is a positive definite matrix (see
Section 8-5). Uniform controllability ensures that the transfer of the states
can be achieved in the time interval (J e' The concept of uniform controllability
is needed in the stability study of optimal control systems. See References 56
and 102.
Instantaneous controllability and uniform controllability both imply con
trollability. However, instantaneous controllability neither implies nor is
implied by uniform controllability.
Example 4
x=e-1rlu
Since p(Mo(t)) = p(e- Itl ) = 1 for al1 t, the dyilamical equation is instantaneously
controllable at every t. However, the dynamical equation is not uniformly
control1able because there exists no CX that depends on (Je but not on t such that
,+
W(t,t+(Je)= 1
J.
e-2Idt=0.5e-21(I-e-2Uecx((Je)
Example 5
x= b(t)u
with b(t)defined as in Figure 5-5. The dynamical equation is not instantaneous
Iy control1able in the interval (-1, 1). However, it is uniformly control1able
I
in ( - 00, oo). This can be easily verified by choosing (J e = 5.
A remark is in order concerning controllability, differential control1ability,
and uniform controllability. lf a dynamical equation is differentially con
trollable, a state can be transferred to any other state in an arbitrarily smal1
interval of time. However, the magnitude of the input may become very large;
in the extreme case, a delta-function input is required. lfadynamical equation
is merely controllable, the transfer of the states may take a very long interval of
time. However, if it is uniformly controllable, the transfer of the states can be
nCAL EQUATlONS
183
b (1)
t)saiaJI
-3
-2
-1
achieved in the time interval a e ; moreover, the magnitude of the control input
will not be arbitrarily large [see (5-14) and note that the input is proportional to
W-l]. In optimal control theory, the condition of uniform controllability is
sometimes required to ensure the stability of an optimal control system.
Time-invariant case. In this subsection, we study the controllability
n-dimensional linear time-invariant state e<tuation
FE:
)n
ation is instantaneously
lation is not uniformly
re but not on t such that
5.
Jerential controllability,
m is differentially con
: in an arbitrarily small
may become very large;
Ir a dynamical equation
~e a very long interval of
lsfer of the states can be
x= Ax + Bu
oC the
(5-23)
all the enfres of A and B are real mmbers. we have agreed to consider them as elements
of the field of complex nllmbers.
184
Example 6
eA'BB* *, dr
(3-42) becomes
1
O
O
O
(5-24 )
has rank n.
4. For every eigenvalue A of A (and consequently for every ), in C), the n x (n
complex matrix [Al -A : BJ has rank n. s
+ p)
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
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
Since changing the sign will not change the linear independence, we conclude
that the rows of e-AIB are linearly independent if and only if p[B : AB : ...
: A n -1 BJ = n. This proves the equivalence of statements 1 and 3. In the
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
5t'[e AI BJ = (si - A)-1 B
Since the Laplace transform is a one-to-one linear operator, if the rows of
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.
The proof of statement 4 will be postponed to Section 5-5 (page 206). Q.E. O.
u=
[B
Al
Example 7
~AL
EQUATlONS
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
(5-24 )
x~l~
1
O
O -1
O
O
5
O
!}+[j}
y=[1
O O OJx
O
1
O
2
-2
O
O -10
-l~l
(5-25)
We compute
am Theorems 5-1 and
U=[B
~orem
AB
A2B A'B]
!
-2
This matrix can be readily shown to have rank 4. Hence (5-25) is controllable.
Thus, if X3 = e is different from zero by a small amount, a control u can be found
to bring it back to zero. In fact, a control exists to bring XI = y, X3 = e, and their
derivatives back to zero. This is certainly consistent with our experience of
balancing a broom on our hand.
Ii
Example 7
endence, we conclude
anly if p[B : AB. : ...
nts 1 and 3. In the
AtB are linearly inde
~ndent on [0, (0) over
llence of statements 1
Consider the platform system shown in Figure 5-2. lt is assumed that the mass
of the platform is zero and the force is equally divided among the two spring
systems. The spring constants are assumed to be 1, and the viscous friction
coefficients are assumed to be 2 and 1 as shown. Then we have XI +2xI = u and
X2 +X2 =u, or
(5-2~)
:ndix G.
[~.5=~.25J = 2
186
[(l0~5'
W(O, 2) =
J[0.5J [0.5
(1'
O5
. '
1] [ O
0]
(1'
dr =
[1.6
6.33
6.33J
27
and
u 1 (t)=-[0.5
u (1)
40
20
u 2/
..... /4
---
1, seconds
In Figure 5-6 we a]
xl(O)=lO and X2(O) =
in transferringx(O) to 2
of the input. If no re:
x(O) to zero in an arbi
the input may become
imposed, we might nc
interval of time. For
might not be able to tr
Example 8
-20
-40
xl (1)
Clearly we have
10
............
........
............
........... 2
2.5
...------- ....
3
1, seconds
X (1)
2
* Controllabilty
10
Let A and B be n x
2
ol---l--+--I----+-+----t--+-----:,f--------.
-5
Figure 5-6
ndicE
--
,.,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
Let b , ; = 1, 2, ... , p, be
Jt u. Using Equations
1.6
- [ 6,33
187
;AL EQUATlONS
6.33J
27
and
ul(t)
=-
[0.5
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 zeO in a finite time.
ill
*Controllability ndices
Define
k =0,1,2, ...
(5-27)
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,
then Un_has a rank of n and consequently has n linearly independent columns.
Note that there are a total of np colurrins.in U ; hence there are many possible
ways to choose these n linearly ind.ependeiltcolumns. We discuss in the follow
ing the most natur~1 and also the most important way of choosingthese columns.
Let b . i = 1, 2, ... ,p, be the ith column ofB. Then the matrix Uk can be written
II
II
188
explicitly as
U k = [b
bz
ro
(5-28)
/"k
.::;p
Hence, aH columns
columns. Consequer
Uk must increase by
increase. Since the r
to add at most n - p
fJ. -l'::;n - p or J.'::;/
yields the upper boun
In order to have
rows. Hence, we neec
(5-31).
The degree of th
min (, n - p + 1) in (
Theorem 5-7, we havt
Corollary 5- 7
The state equation FJ
= pUl' =
pUl' +1
= ...
(5-30)
y=
[B: AB
ICAL EQUATIONS
(5-28)
Un - p = [B
\.b 1
.. , An-PB]
Example 9
Consider the satellite system studied in Figure 3-17. Its linearized dynamical
equation is developed in (3-51). As can be seen from the dotted lines in (3-51),
the control of Xi> i = 1,2, 3, 4, by and 2 and the control of X5 and X6 by 3
are entirely independent; hence, we may consider only the following subequation
of(3-51):
o
o
o
1
(5-30)
='"
.
3
x= o
ro
o
o
LO
-2
left-hand-side
ecked froro U 1 and fl
e controllability index
J. -
(5-31 )
y=[~
~l I~o ~lo
1 x+
OJ
LO
(S-32a)
1j
~J x
(S-32b)
AB
xp + A 2 b
'1"+1
189
~l!
O
O,
O :,
O
O,,
O
O
2 : -1
1 : -2
1 : -2
O'
~l
-4
has rank 4. This is the case; hence, the state eqilation in (5-32) is eontrollable.
190
Theorem 5-8
The set of the controll~
transformation and ar
Proof
Define
where A=PAP-, .
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
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
(i, j)th cell represents the column Ai - bj- A column which is linearly indepen
dent of its left-hand-side columns in (5-28) is denoted by "x"; otherwise denoted
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
equal to ri- as shown. From the crate diagram, we can deduce that r is equal
to the number of {bk , k = 1, 2, ... ,p} with controllability indices equal to or
smaller than i. Hence, we conclude that
(5-33)
1
A
A2
A3
A4
AS
/1i
b
x
x
X
O
O
O
3
b2
x
O
O
O
O
O
1
b3
x
x
b4
x
x
O
O
O
3
O
5
ro =0
r = 1
1'2 = 1
1'3 =3
1'4= 3
1'5 =4
which implies
where M is a p x p eler
forward to verify that
-
f:,
Uk=[B:
Consequently. we cane
the ordering ol" the eoll
Now we discuss a (
umns in U = [B AB
:AL EQUATIONS
191
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
transformation and any ordering of the columns of B.
Proof
Ab p, ... , AJ.Lp-l b p
Define
is controllable. The
es of {A, B}.
controllability indices
~lationship, we use an
. = 3. These indepen
11 in Figure 5-7. The
;;h is linearly indepen
x"; otherwise denoted
;-28) from left to right
, and then to the next
h row of Figure 5-7 is
deduce that ri is equal
ty indices equal to or
ability index i
(5-33)
mtrollability index 1;
2, and b and b 4 have
)i1ity index 5. Hence,
B=
for k = O, 1, 2, ...
which implies
for k =0,1,2, ...
Hence, the r defined in (5-28) and, consequently, the set of controllability
indices are invariant under any equivalence transformation.
The rearrangement of the columns of B can be represented by
B=BM
where M is a p x p elementary matrix and is nonsingular. Again it is straight
forward to verify that
iJd~~ [B
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
... ,
... ,
Aii2- l b"
2,
... ,
b P' Ab P'
... ,
Aiip-lb p
...:.=---_._-_._----
192
-------------------,~=~
Example 10
where A, R, e, and E a
continuous functions (
Definition 5-5
In
In
lF
urv
5-4
(a)
ICAL EQUATIONS
193
is controllable, we have
are {P1' z, ... , lp}.
IS depend highly on the
x= A(t)x(t) + B(t)u(t)
E:
(S-3Sa)
(S-3Sb)
y = C(t)x(t) + E(t)u(t)
Definition 5-5
The dynamical equation E is said to be (completely state) observable at t o
if there exists a finite t 1 > t o such that for any state X o at time to, the knowledge
of the input ulto,tl! and the output Ylto,tI! over the time interval [to, t 1J suffices
to determine the state xo. Otherwise, the dynamical equation E is said to be
unobservable at to.
111
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;
hence the system, or more precisely, the dynamical equation that describes the
system, is not observable at any to.
111
Example 2
Consider the network shown in Figure 5-9(a). Ir no input is applied, the network
reduces to the one shown in Figure 5-9(b). Clearly the response to the initial
:luations
, dual to that of control
e possibility of steering
;ibility of estimating the
'ollable, all the modes of
lical equation is observ
the output. These two
the complete knowledge
md E are known before
nt from the problem of
ication is a problem of
~mation collected at the
IH
In.
1
' -_
_---4-----_
F
(a)
In.
L _ _
(b)
~------JIF
194
__
-
-~' ~~--~-~.
_,o
.~--~_.~- --~~-~--~--~~-_
..
__._-------
__
current in the inductor can never appear at the output terminal. Therefore,
there is no way of determining the initial current in the inductor from the input
and the output terminals. Hence the system or its dynamical equation is not
observable at any too
I
l1>(t, r)B(r)u(r) dr
+ E(t)u(t)
y(t
(5-36)
where l1>(t. r) is the state transition matrix of X = A(t)x. In the study of observ
ability. the output y and the input u are assumed to be known, the initial state
x(t o) is the only unknown: hence (5-36) can be written as
where
(5-37)
y(t)~ y{t) -
(5-38)
C(t)
(5-39)
where
Veto, t }~
1 l1>*(t, to)C*(t)C(t)l1>(t,to) dt
(5-41 )
lo
From Theorem 5-1 and the assumption that all the eolumns of C(')l1>(', to) are
linearly independent on [to, t 1J, we concIude that Veto, tI) is nonsingular.
Hence, from (5-40) we have
Theorem 5-9
Prove by contradictio
ti> t o such that the co
Then there exists an n
Xo =
V- 1 (t o, tI)
'
l1>*(t, to)C*(t)y(t) dt
(5-42)
lo
Thus, ifthe function y[lo, Id is known,x o can be computed from (5-42). Necessity:
We list in thefollowir
tions 5~6 to 5-8, which a
5-4 for controllability. ~
CAL EQUATIONS
terminal. Therefore,
Iductor from the input
lmical equation is not
C(t)<.l>(t, tola = O
+ E(t)u(t)
Hence the initial statex(t o) = a cannot be detected. This contradicts the assump
tion that E is observable. Therefore, if E is observable, there exists a finite
ti> t o such that the columns of C( )<.1>(', t o) are linearly independent on [t o' ti].
Q.E.D.
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
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
X= A(t)x, Y= C(t)x.
The controllability of a dynamical equation is determined by the linear
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.
(S-38)
'oblem is a problem of
nd <I>(t, to). Note that
t time to. However, if
d,
:n by
11;
195
(5-39 )
E*:
Proof
From Theorem 5-4, the dynamical equation E is controllable if and only if the
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
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
<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.
Q.E.D.
(S40)
(541)
dt
(S-43a)
(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
equation E* is observable (controllable) at too
) dt
z= - A*(t)z +C*(t)v
1 =B*(t)z +E*(t)v
(S-42)
rom(5-42). Necessity:
i
I
L
196
Definition 5-7
Theorem 5-11
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
(544)
where
k = 0, 1, 2, ... , n-1
(5-45a)
(5-45b)
I
WIth
0< {31(O"o)I:
O < {33(O"o)I:
for alI t, where el> is the
Time-invariance ca
equation
Definition 5-6
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),
No(to)
NI (to)
Theorem 5 -1 3
=n
where A, B, C, and E al
time interval of interest
time-invariant dynamic
t o ~ 0, and the determino
time interval. Hence, tI:
bility study of linear tirr.
N n - 1(to)
I
_---~~,
..
~
_
..
-~----~
'-.-.e-_
(lCAL EQUAUONS
197
Definition 5-7
le interpretations in the
.t
l~;~:;
l=n
Nn~l(t)J
for all t in ( -
00,
(0)
(5-44)
11
Definition 5-8
~,
... ,n-1
(5-45a)
(5-45b)
I
+ (Jo) .:::::;f34((Jo)I
Ity,
for all t, where $ is the state transition matrix and V is as defined in (5-41).
Time-invariance case.
termine x(to).
I
:n the n-dimensional dy
t in (-ro, (0) ifand only
11
equation
FE:
x=Ax +Bu
y=Cx +Eu
(5-46a)
(5-46b)
_ontnuity.
Wot~
is nonsingular for any 1 > O.
eA''C*Ce A < dI
e, the field of
-------------------------
198
..
------~_...:-_~~~~-=--.----------
CANONICAL OECOMPO~
and
Equivalently, v is the
pV o <
(5-47)
Observability Indices
Let A and
Define
Cq
Theorem 5 -14
(5-48)
5-5
CalT1lonical Dre
DYl1lllmicai Equatic
ro '::;/"1
'::;/"2'::;"
'::;q
Since there are at most n linearly iridependent rows in V ro' there exists an integer
v such that
(5-49a)
CAL EQUATIONS
rv=r v + = ... =q
and
199
(5-49b)
(5-47)
(5-50)
The integer v is called the observability index of {A, C}. Similar to (5-31), we
have
~svsmin(,n-q+l)
(5-51)
q is the rank
of C.
Corollary 5-13
The dynamical equation FE in (5-46) is observable if and only if the matrix
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
coprime.
ine
l rows)
S-5
Dyn<llmical IEquatioU1J
FE:
x=Ax +Bu
y=Cx+Eu
(5-52a)
(5-52b)
200
CANONICAL OECOMPOSI1
Proof
X=Ax +Bu
+Eu
FE:
(5-53a)
(5-53b)
pU~
y = Cx
V~[B:
Since the rank of a matrix does not change after multiplication of a nonsingular
matrix (Theorem 2-7), we have rank U = rank . Consequently FE is con
trollable if and only if FE is controllable. A similar statement holds for the
observability part.
Theorem 5-15
- [i
=[
U=
Theorem 5-16
FE:
[;:J
t J[::J
z
= [:'
y = [Ce Ce]
[::J +
[~'J u
Eu
(5-54a)
(5-S4b)
Ve
represents the
~n are line
pU =n implies pUc = nI'
We show now that the
FE and FiL. hav~ the ss
-k-
-A~ZJ [(S]
sI -Ae
Xc = Acx c + Bcu
y=Ccx c +Eu
where
AcB wlth k
(5-55a)
(5-55b)
- _][sI - Ac
[ Ce
Ce
O
=
sI
[Ce Ce] I
=CcCsI -Ac
9
IlCAL EQUATlONS
201
Proof
(5-53a)
(5-53b)
e dynamical equations
is caBed an equivalence
: PAn-lB]
~n-B]~ PU
have
pU~p[B
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
{q, Q2,"" qnJ (Why?) Define a nonsingular matrix
(5-56)
cation of a nonsingular
msequently FE is con
;tatement holds for the
le-invariant dynamical
ation.
where the last n - n columns of Q are entirely arbitrary so long as the matrix
Q is nonsingular. We claim that the transformation ji: = Px will transform FE
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
given in (5-54a). The columns of B are the representations of the columns of
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).
Let U and be thecontrollability matrices of FE and FE, respectively. Then
we have pU = p = n (see Theorem 5-15). It is easy to verify that
- .[BeJ~ AeBe !I . !I A~-BeJ
U=
O: O:
1 O
=[e!
O
}(n-n)rows
Since columns of
A~B with k ?:::n are linearly dependent on the columns of e, the condition
pU = n implies pe = ni' Hence the dynamical equation F Ee is controllable.
(5-57)
(5-54b)
'ix as FE.
[Ce
CcJ[sl
~Ae
s~~~cTT~eJ +E
+E
=Ce(sl -AJ-Be +E
ualion
Q.E.D.
202
CANONICAL DECOMPO;
[~c
and
[:J
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
is not controllable, by a proper choice of a basis, the state vector can be decom
posed into two groups: one controllable, the other uncontrollable. By
dropping the uncontrollable state vectors, we may obtain a controllable dy
namical equation of lesser dimension that is zero-state equivalent to the
original equation. See Problems 5-22 and 5-23.
Example 1
Dual to Theorem
dynamical equations.
Theorem 5-17
FE
y=[1
l]x
(5-58)
[B : AB] in determining
1
1
1
ro
p-l = Q
11
Lo
1l
lo
ooJ
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
x=Px. We compute
A~PAP-' ~[~
o
o
m~
J[~
By equivalence transf<
-1
10
ICAL EQUATIONS
pace ~ of FE is divided
;e of~, denoted by L. l ,
and
O 1
1] 1 O
[
O 1
le (n - nl)-dimensional
FE c is controllable, all
G ~]
Xc
+ [~
~] u
y = [1 2Jx c
(5-59 )
FE:
1Jx
203
(5-60a)
y=[C o O][::J+EU
(5-58)
(5-60b)
(5-61 a)
(5-61 b)
:3
ldependent columns of
~e Q nonsingular. Let
This theorem can be readily established by using Theorems 5-16 and 5-10.
The first n2 rows of P in Theorem 5-17 are any n2 linearly independent rows of
the observability matrix of {A, C}; the remaining n - n2 rows of Pare entirely
arbitrary so long as P is nonsingular. Equation (5-60) shows that the vector
x does not appear directly in the output y al' indirectly through %0' T-lence the
vector Xi; is not observable and is dropped in the reduced equation.
Combining Theorems 5-16 and 5-17, we have the following very important
theorem.
Theorem 5-18 (Canonical decomposition theorem)
10
x=Ax +Bu
y=Cx +Eu
This is a simplified version of the canonical decomposilion theorem. For the general form, see
References 57, 60, and 116. See also Reference S127.
---~._---------
204
CANONICAL DECOMPC
canonical form
(5-62 a )
Cco : Ce]x
y = [O
+Eu
(5-62b)
ca
co
Cco(sI -
which depends solely on the controllable and observable part of the equation
FE.
Proof
If the dynamical equation FE is not controllable, it can be transformed into the
form of (5-54). Consider now the dynamical equation FE c which is the con
trollable part of FE. If FE c is not observable, then FE c can be transformed into
the form of (5-60), which can also be written as
[~caJ [AOco
=
X co
-y = [O
~12J[~caJ
+[~caJ
u
A
X
B
co
Cco]i
co
co
+ Eu
Before moving to
of Theorem 5-7.
CcoCsI - co)-1 co +
LI
!l
J/
I
ca
In
ni
1------1
I
I
I
I
1
I
:e
10
1
I
I
I
CO
1
I
1
I
-.J
CAL EQUATIONS
(5-62a)
(5-62b)
205
Example 2
Consider the network shown in Figure 5-11(a). Because the input is a current
so urce, the behavior due to the initial conditions in C I and L I can never be
detected from the output. Hence the state variables associated with el and
L I are not observable (they may be controllable, but we don't care). Similarly
the state variable associated with L 2 is not controllable. Because of the sym
metry, the state variable associated with C2 is uncontrollable and unobservable.
By dropping the state variables that are either uncontrollable or unobservable,
the network in Figure 5-11(a) is reduced to the form in Figure 5-11 (b). Hence
11
the transfer function of the network in Figure 5-11(a) is g(s) = 1.
Before moving to the next topic, we use Theorem 5-16 to prove statement 4
of Theorem 5-7.
Cl
Li
Lz
"
In
In
c2
In
In
(a)
In
In
y
In
In
(b)
206
ex[AI -A : B] =0
exA = exA
and
exB = O
or
CANONICAL OECOMPO
which imply
Theorem 5-19
and, in general,
A linear time-invariar
is control1able and ob
i = 1,2, ...
Hence we have
Proof
=[~c ~c2]
A=PAP- 1
B=PB=[~C]
ex[AI-A; 8]=[0
p][AI~Ac
-A 12
'
Al -Ac '
R
oc] =0
FE,
of lesser dimension, sa
Theorem 4-6, we have
which implies
o=et[P(Ali -
a[AI -A : B] =0
In other words, if {A, B} is not controllable, then [si-A: B] does not have a
rank of n for sorne eigenvalue of A. This completes the proof.
Q.E. O.
l,
By (5-64), we may replc
---------------
207
:AL EQUATIONS
, the eigenvalues of A;
~xcept possibly at the
,s controllable, then
:re exist an eigenvalue
equation still has the same zero-state response, This fact motivates the follow
ing definition.
Definition 5-9
Proof
"-la.BJ=O
f. Hence, if {A, B} is
lue of A and, conse
t1 p[ Al - A : BJ < n
:, there exists an equi
S} with
FE:
x=Ax +Su
y =Cx +Eu
(5-63a)
(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
CAkB=CAkS
k=O,I,2,.,.
(5-64)
-1: BJ
ular, a,(AI - A)P-l =0
CB
CAB
~
proof.
Q.E.D.
CAn-lB
CAnB
CAn-lB
CAnB
(5-65)
CA2(~-1)B
'VU=Vn-1 n- l
(5-66)
where V"-l and "-l are defined as in (5-48) and (5-27)., Since FE is control
lable and observable, we have p U = n and pV = n, lt follows 'from Theorem
2-6 that p(VU) = n. Now Vn - l and n - 1 are, respectively, qn xn and ni x np
matrices; hence the matrix V"-l"-l has a rank of at most ni' However,
.~,~"~~~.
208
CONTROLLABILlTY A1
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}
is cal1ed a realizatian of G(s). Now if {A, B, C, E} is control1able and observable,
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.
Theorem 5-20
invariant dynamical e
is referred to Referen<
*5-6 Controllabi
Dynamical Equa
The controllability
equation are invarial
ceivable that we ma~
by transforming the
equation is in a lord:;
almost by inspection
Consider the n-d
equation
Proof
The sufficiency fol1ows directly from Theorems 4-6 and 5-15. We show now the
necessity of the theorem. Let U, V be the controllability and the observability
matrices of {A, B, C, E}, and let , V be similarly defined for {A, B, C, E}. If
{A, B, C, E} and {A, B, C, E} are realizations of the same G(s), then from (5-64)
and (5-65) we have E = E,
and
VU=V
VAU=VA
(5":67)
(5-68)
(i!,
(5-69)
A
(n x 11
where P ~ eY*V)-l V*V. From (5-69) we have p .=s; min (pP, pU), which,
together with p = n, implies that pP = n. Hence P qualifies as an equivalence
transformation. The first p columns of (5-69) give B = PB. Since pU = n,
Equation (5-69) implies that
P = (U*)(UU*)-l
With P=eY*V)-lV*V=(U*)(UU*)-l, it is easy to derive from (5-67) and
(5-68) that V = VP and PA = AP, which imply that C = CP and A = PAP - 1.
Q.E.O.
This theorem implies that al1 the irreducible realizations of G(s) have the
"same dimensiono Physically, the dimension of an irreducible dynamical equa
tion is the minimal number of integrators (if we simulate the equation in an
" 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-
A
x 11;)
(n
c
(q xn
Aij
(l1j
x nij
~AL
EQUATIONS
tradiction.
Hence, if
Q.E.D.
A, B, e, E} has a pre
equation {A, B, e, E}
llable and observable,
1, In the following we
quivalent.
cible realization of a
is also an irreducible
C, E} are equivalent;
lch that A = P AP - 1,
invariant dynamical equations. For the time-varying case, the interested reader
is referred to References 106 and 108.
X =Ax +Bu
y=Cx +Eu
JFE:
15. We show now the
and the observability
E}. If
d for {A, D,
(;(5), then from (5-64)
e,
and
(n
A
x n)
~ible dynamical
lA'
J (n~p) ~[J
A2
C=[C I C 2
A
(ni x n)
W*V) is
(5-69)
(5-70a)
(5-70b)
where the matrices A, B, and e are assumed of the forms shown in Table 5-1.
The n x n matrix A is in the Jordan form, with m distinct eigenvalues A, ,,1,2' .. ,
Am Ai denotes all the Jordan blocks associated with the eigenvalue Ai; r(i) is
the number of Jordan blocks in A; and Aj is thejth Jordan block in A. Clearly,
A i = diag (Al' A i2 , ... , Air(i)
(5-67)
(5-68)
e the matrix
ve have
209
{U
Cj
=[C i1
Cm]
-,
I
A'2
A,j
Cir(il]
C'2
(q x n)
. AJ
(nij
x ni)
=[A'~; [ ]
Al
Ai
B.
lB" 1J
_ Bi2
(n, "p) -
BHj )
210
Let
CONTROLLABILITY Al'
and
nij
n=
ni =
i=I
r{i)
I I
nij
=lj=l
(5-71 a)
11
The conditions 1
required that each of
pendence. The linee:
Furthermore, the rov
determining the cont]
The physical mea
the block diagram o:
studying the general
Jordan-form dynamic
an integrator and a fe.
block, or more preci:
variable. Each chain
Consider the last cha
variables in that chail
that chain can be obs(
same eigenvalue, the
vectors of these chair
be studied separately.
Example 1
JFE:
x=
Al 1: O O
O Al: O O
-0- - 6- :-l~: O
O O --:-il -:
O O O O:
___ .L
lo
O O
O
O
O
O
O
O
O
O
O
O x
O
Az
O: O
O: O
Az
O
11;2:0;020]
y= [ 110:1:2:02121 x
0:2:3:0
1 J'
Az
O O O
100
~bl11
O 1 O
------
+ O O 1
U~bl13
------
------
~bl1z
(5-72a)
1 1 2
l~ ~ ~J
(5-72b)
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
211
AL EQUATIONS
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.
/ C I ij and Clij'
(5-71 a)
;). J FE is observable
>( r(i) matrix
(5-71 b)
11
:rs).
~b111
~b112
1 U~b!13
2
(5-72a)
u
y
);
1J
~b121
u
y
(5-72b)
Figure 5-12
212
CONTROLLABIl.ITY
A~
linearly independent.
theorem can be estab
S-Al
O
O
S -
Al
O
b ll1
-1:
b Zll
b111
b 11Z
S-Al:
----------------------:- -. .: X; ------= i -- -,
~
s -Al
bl1z
The matrix A has two distinct eigenvalues Al and Az. There are two Jordan
blocks associated with Al' one associated with A z. If s =A b (5-73) becomes
O
O
O
-1
O
O
O
-1:
O:
b ll1
b Zl1
b 111
b 11Z
- - - - - - - - - - - - - - - : - - - - - =---( - -:
,
O:
-1:
Example 2
Consider the
(5-74)
A single-input linear
trollable if and only
distinct eigenvalue an
spond to the last row
A single-output li
observable if and onl
distinct eigenvalue an
to the first column of
single-v~
O
O
Corollary 5-21
x=
: O
O:
b'12
-----------T----------------T-----
, Al -A z
- 1 , b 1Z1
A1 -A z : b lZ1 _i
I
O
O
o :
o (
r~- -~
O:
- ------------- -: - -- -- =---(
~ _?
b l11
O
- -:
:__ ~l!! _
\
I
(5-75)
Al -A z
O,
and
A1 -A z : O
Note that Al -A z is different from zero. The matrixin (5-75), or equivalently,
the matrix in (5- 73) at s = Al' has a full row rank if and only if b 1ll aild b 11Z are
ICAL EQUATIONS
213
b l21
(5-73)
Example 2
b 211
b L11
b l12
: b L12
:t~-
"'-2
Corollary 5-21
=,1.1' (5-73)
.
x=
(5-74)
bL12
-T-----
:
: :
b l21
b 121
O O 1 0
I
10
x+
y = [1
O O: lJx
There are two distinct eigenvalues O and 1. The component of B which cor
responds to the last row of the Jordan block associated with eigenvalue O
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;
therefore, the equation is observable.
11
Example 3
(5-75)
l12
,
, O
., O
., b
-------o
~r--H\~l ~1}
I
O 1 O: O
b lll
b 211
b l11
b l12
(5-75), or equivaleritly,
and
X2
[XIJ
[-1
0J x + [e- J
-2
(5-76)
e-21
(5-77)
That the state equation (5-76) is controllable follows from Corollary 5-21.
Equation (5-77) is a time-varying dynamical equation; however, since its A
matrix is in the Jordan form and since the components of B are different fram
214
OUTPUT CON
zero for all l, one might be tempted to conclude that (5-77) is controllable.
Let us check this by using Theorem 5-4. For any fixed to, we have
- ('o -
et>(to-t)B(t)= [ e O
t)
J[:-2'
O
e- 2('0-'l_
:-21
[ - 'o ]
t ]
It is clear that the rows of et>(to - t)B(l) are linearly dependent in t. Hence the
I
state equation (5-77) is not controllable at any to.
The transfer-function n
From this example we see that, in applying a theorem, all the conditions
should be carefully checked; otherwise, we might obtain an erroneous con
clusion.
has rank q.
y(t) =
Corollary 5-22
Definition 5-10
Example 1
Theorem 5-22
ID.
ID.
u
IF
+
-
Ir
1\
ID.
ID.
ICAL EQUATIONS
215
(5-77) is controHable.
t o, we have
-fo ]
= [ :- 2 10
runction
(5-79)
I
The proof of Theorem 5-7 can be applied here with slight modification. A
trivial consequence of this corollary is that every single-output system is output
controllable. We see that although the condition of output controllability
can also be stated in terms of A, B, and C, compared with checking the linear
independence of (;(s), the condition (5-79) seems more complicated.
The state controllability is defined for the dynamical equation, whereas
the output controllability is defined for the input-output description; therefore,
these two concepts are not necessarily related.
Example 1
ite
Consider the network shown in Figure 5-14. It is neither state controllable nor
observable, but it is output controllable.
y\.
tI> t o
and an input
I
In
In
'1;
1\
In
In
Figure 5-14 A network which is output controllable but neither (state) controllable nor
observable.
._--_._._-~--_._._-
216
....
"
..-
...._..
_._- .._-"._--
1-----------------,
I
I
I
I
I
I
:
I
I
I
l-----+:--.-Y2
I
Figure 5-15 A system which is controllable and observable but not output controllable.
Example 2
,;1]
s +1
the rows of which are linearly dependent. Hence the system is not output
controllable. The dynamical equation of the system is
x= -x +u
which is controllable and observable.
(5-80)
If p(;(s) = q-that is, all the rows of (;(s) are linearly indepe~dent over the
field of rational functions-then theq x q matrix (;(s)(;*(s) is nonsingular
(Theorem 2-8). Consequently, for any y(s), if we choose
o(s) = (;*(s)(;(s)G*(S))-ly(S)
. (~-81)
"5-8 Computatic
In this section, we dis
chapter. As discussed
conditioned; a comput
unstable. If we use a
problem, the result wil
if we use a numericall)
result will be correcto
problem, well or ill cor
a problem, if we must
stable method, the un
possible, in the compu
As discussed in TI
trollability of astate el
more suitable for comI
computational probler
The computation (
is sttaightforward. L(
COMPUTATIONAL PROBLEMS
CAL EQUATIONS
217
range of G(s), for which there exists no solution u(s) in (5-80) (Theorem 2-4).
Q.E.D.
:fer-function matrix of
be transferred to any
oblem is whether it is
ve over any intervalof
interval of time is said
ble.
(5-80)
(5-81 )
= q, then the system is
IS
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
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
References S172, S185, S218, and S239.
- ... '
._...
_-~
---~~.
218
for k large
In other words, AkB tends to approach the same vector, vl' as k increases.
Hence, it will be difficult to check the rank of U if n is large.
The same conclusion can also be reached by using a different argument. The
condition number of a matrix A may be defined as cond A~ IIAI1211A -1112 = aJaS'
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
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)
11
Figure 5-16
The root
MICAL EQUATIONS
COMPUTAT10NAL PROBLEMS
20
19
20
18
20
; that is, Av =
AV.
lt is
219
20
1
(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
n
20
:have
Ll(s) =
(s -i) -(20)19 6
i= 1
large.
11
u
El
10
Figure 5-16
11
220
if e is larger than 7.8 x 10- 14 ). We list in the following Alo and Al! for sorne e:
el
7.8
82=1083 =
e4 =
10- 14
10
10- s
1
10.5 jO
10.5j2.73
10.5 j8.05
10.5 j16.26
where AW is a nz x n
P 1 B= O O x
O O O
=[B~I]
A 11
~PAP'=
An-IBJ=Pl[~Y}--~W~~~_!~!D~~~~_~~~J
O
O
O
...
A 21
C=CP'=[C 1
(5-82)
A(I) : A(l)]
PI AP'I = A(l)
- J :_'- - : ~
(5-83)
[ 2I!I A(l)
22
where AW is an nI x nI matrix, AW is an (n -nI) x nI matrix, and so forth.
If AW = O, the controllability matrix
AB:''';
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
U=[B
x x
and
(5-84 )
has rank nI < n and {A, B} is not controllable and can be reduced to controllable
where rank Zj
If {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.
12
14
COMPUTATlONAL PROBLEMS
ICAL EQUATIONS
{AW,
B~l)}.
221
P2AW=[A~2lJ
(5-85)
~d.
[The computation
he characteristic poly
md the roots of a poly
ient.] We see that the
Llues of A are very ill
rank [sI -A B] =n
leous result.
,t way of checking the
:quation is not control
e equation. The proof
.e required equivalence
'ollability matrix and is
In the following, we
od to transform a state
lt case, the method is
\. into the Hessenberg
: general case.
o such that
(5-87)
Now
:J
A 11
A 2l
~ PAP' = O
(5-83)
!~~l~--:~~J
C=CP'=[C l
(5-82)
(5-88)
O O ...
A 12
An
A 32
A 13
A23
A 33
A 1,k-l
A 2,k-l
A 3,k-l
Ak,k-l
C3
Ck -
AIk
B\l)
A2k
O
A 3k , B~ PB = O (5-89)
O
C k]
Step 1. P = 1m Aa = A, B a = B, = 0, j = 1.
Step 2. Find an orthogonal transformation P j such that
(5-84)
P.B.
= [Zj]}n
) )-1
O
reduced to controllable
"']
has rank nI + n2 (where x denotes nonzero matrices), and {A, B} is not control
lable and can be reduced. Ir AW f.O, we continue the process until {A, B}
is transformed, by a sequence of orthogonal transformations, into 14
B(l) A(1)B(l)
- _-_[
O
O
PI
14
222
Step 3. Compute
P A pi. =
))- )
o
where X j is a
nj
nj
Step 4. Update
matrix.
p=p
Xo Yo]
J..
B A
[ _!
j
[1- O]
n
O Pj
Step5. =+nj. If=n,gotostep8.
controllable equation.
Step 8. {A, B} is controllableo
This is a numerically stable method 01' checking the controllability 01' a state
equation. This algorithm was first proposed by Rosenbrock [S185] and then
discussed in References S6 and S155. These papers used the gaussian elimina
tion with partial pivoting in finding Pj' Van Dooren [S203] and Patel [S170]
suggested to use orthogonal transformations to improve the numerical stability
01' the algorithm. The singular value decomposition was also suggested in
Reference S170 to determine the rank 01' B i .
The algorithm actually reveals more than the controllability 01' {A, B}; it
also reveals its controllability indices. A comparison 01' (5-84) and (5-88) with
(5-28) yields immediately
i =0,1,2, ...
where r; is the number 01' linearly dependent columns in AiB and ni+ is the
number 01' linearly independent columns in AiB. Since the set of controllability
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
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
O : 1: x
O O /!\
------O O O
O O O
------O O O
O O O
------O O O
A
A=
x
x
:x....
,
,
.,,
x ,x
x'x
x'x x'x
0J
,
,
x ,x
,
,X
,
,
x ,x
:..t: x x: x x: x x: x
O O
,ti\.:
x: x x: x
(5-90)
0- -- - X-~-: ~- -;;-:-;
-
O O O: Oo '1' :b
1 e1: d 1
I
-)"t
0--0- --:-6-0-:0-:'E:f
~
1st
'--v--'
"-v--'
2nd
3rd
block co\umn
Let B = fb b2 b31; that is, b; is the ith calumn ofB. Thcn from the structure
01' and B in (5-90), we can readily veril'y that 111=3,jJ.2 = 1, and}J.3 =4. These
controllability indices,
computatian. The se,
dependent on its left
dependent; hence}J.2 =
not appear in the subst
01' , both columns an
; hence A 2b and .~21
, the first column is 1
3 b is linearly depenc
column is linearly indo
linearly independent o
these indices are equal
triangles as shown in (:
The {A, B} in (5-90)
tion into the following
1 x
O
O
O
O
O
O
O
O
O
O
O
O
Then we have
COMPUTATlONAL PROBLEMS
CAL EQUATlONS
ontrollability of a state
brock [S185J and then
i the gaussian elimina
;203J and Patel [S170J
the numerical stability
was also suggested in
_______
_______
_____ L
_____ L
____
____
__
__
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
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
tion into the following form:
,
,
1 x x
x x x. x x,, x x,x
,
O 1 x
x x x: x x', x x:x
,
O O 1
x x x. x x. x x' x
------
O O O
1 O O ,, O O; O 0:0
(5-91 )
B= O O O
= O O 1 :0
0:0 O: O
------
O O O
O O O 1 0;0 0;0
,
O O O
1 , ____
O O:J_._
O
O
O O :0
_______
_____
,
- - --6
O O O :0 o' O 1 , O
~
to an -dimensional
223
x', x
x
x:
x:
1
1
O
x', x
l(
"------
x: x
l(
(5-90)
~ - - ~ _:->:
x:x
l(
\ _':.1_: d 1_
) /! :.1'1
'.1
--v--'
3rd
~hcn
LJ
Then we have
.. -
-\.J
x
x
x
1
O
x
x
x
x
O
-
.. -
..' -
x:x
x:x
X'x
x.x
x:x
x:x
x:x
x.x
X'X
_1 .. _ _
0- -- -6 : -(
1
~.j
O O O: O 1 : b2 C2 :d 2
0- -- -O : -0- - b- :- 0- - i -:"0
(5-92)
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).
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)
is numerically stable. The process of transforming the block Hessenberg form
into the block companion form in (5-91), however, is not numerically stable.
The matrix PI in (5-92) carries out gaussian elimination on columns of A
without any pivoting. lf pivoting is used, the form of A will be altered, and we
can never obtain the form in (5-91). Hence, PI must be chosen without any
pivoting and the process is numerically unstable.
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}
can be transformed into
1 x x
O O O
O O O
------O 1 x
B=PB= - - - - - - O O 1
O O O
O O O
O O O
x X x,x'x
x x X
, ,
1 O O ,' O ,' O O O O
O O O O
O 1 O : O L: __________
X
x
x x x
x:x:x
A=PAP- I = -------'--r---------
x x x,x'x
x x x
, ,
O O 0:0: 1 O O O
O O 0:0:0
1 O O
, ,
O O 0:0:0 O 1 O
_ _ _ _ _ _ _ ..I _ _
-A
1
T)l=
B: O
O:-A
O:
O :
O
'-v
1"0
1"1
(5-93)
O O O (O O
O O O O O
O O
P= O ~! ~ O
O O ,. O O
O O O O L '1'
_ ,
O O O O O
O O O O O
L _ ,
O O
( O
O O
O O
O O
O ,.
O O
l. _ .1
O
O
O
O
O
O
-1- J
L _
>
15
rank U = rank [H A
= total numb<
Consequently, we coo(
number of linearly inde
independent columns e
position of [he rows oi
apply Householder tr
pivoting on the rows o
Appendix A). Once in
readily determined.
There are two disa
{A, B}. First, the size (
larger than the size of
controlla1:?le eguation c
of {A, B} into a block H
the controllability of {A
The discussion of t
part and will not be rep
COMPUTATIONAL PROBLEMS
x x x x
O O O O
O O O O
---------x x x x
---------x x x x
1 O O O
O 1 O O
O O 1 O
225
B:
O:
I O:, -A B:
,
0 0 '' 1 0 ' ," ,
T=
l'
[
,
O O :
:
'-.--'
(5-94 )
,
O O:
:'
'----v--'
rl
There are .. + 1 block columns in TI'; each consists of n col umns formed from
A and I and p columns formed from B. We call the former A columns, the latter
B columns. Now we search linearly independent columns of TI' in order from
left to right. Because of the unit matrix 1, all A columns in TI' are linearly
independent of their left-hand-side columns. Let ri be the number of linearly
dependent B columns in the (i + 1)th block column as shown in (5-94). Because
of the structure of TI" we have
(5-93)
and
O
O
O
O
O
O
O
'1-',
TI the positions of 1 in
on of {A, B} into (A, B}
ning the controllability
o be in the controllable
will be seen in Chapter 7.
lability to conclude this
n G-8', we may conclude
tI matrices si - A and B
It can be shown directly (Problem 5-35) or deduced from Chapter 6 that the ri
in (5-94) are the same as the ri in (5-28). Hence we have
226
5-9
Concluding Remarks
Theorems
Observab ility:
Theorems
TYT T sr
5-9
5-11
5-12
5-13
5-14
r I~::,;em
5-17
The theorems in the observability part can be easily derived from the con
trollability part by applying Theorem 5-10 (theorem of duality), and vice versa.
Theorems 5-1 and 5-4 (or 5-9) are two fundamental results of this chapter.
They are derived with the least assumption (continuity), and hence they are
most widely applicable. If additional assumptions (continuous differenti
ability) are introduced, then we have Theorems 5-2 and 5-5 (or 5-11), which give
only sufficient conditions but are easier to apply. If we have the analyticity
assumption (the strongest possible assumption) on time-varying dynamical
equations, then we have Theorems 5-3 and 5-6 (or 5-12). Theorem 5-7 (or
5-13), which follows directly from Theorems 5-1, 5-3, and 5-4, gives the necessary
and sufficient conditions for a linear time-invariant dynamical equation to be
controllable.
The relationship between the transfer-function matrix and the linear time
invariant dynamical equation was established in this chapter. This was achieved
by decomposing a dynamical equation into four parts: (1) controllable and
observable, (2) controllable but unobservable, (3) uncontrollable and unobser
vable, and (4) uncontrollable but observable. The transfer-function matrix
depends only on the controllable and observable part of the dynamical equation.
If a linear time-invariant dynamical equation is not controllable and not
observable, it can be reduced to a controllable and observable one.
The concepts of controllability and observability are essential in the study of
Chapters 6 to 8. They will be used in the realization of a rational matrix
(Chapter 6) and the stability study oflinear systems (Chapter 8). Sorne pr:,tctic:;l
implications of these concepts will be given in Chapter 7.
The computational problems of the various controllability and observability
conditions are also discussed. A1though the conditions can be stated nicely in
terms of the ranks of [B AB '"
An- 1 B] and [51 -A
BJ in the con
trollability case, they are not suitable for computer computations. An efficient
and numerically stablemethod is introduced to transform a dynamical equation
into the form in (5-54) or (5-60), and its controllability or observability can then
be determined. The algorithm can also be used to reduce a reducible dyilamical
equation to an irreducible one.
Before conc1uding this chapter, we remark on the controllability of the
n-dimensional linear time-invariant descrete-time equation
x(k
+ 1) =
Ax(k)
+ Bu(k)
Similar to Definition :
given any X o and any)
transfer X o to Xl' 16 T
A"-lB] =n (see Prob
Theorem 5-7. Hence,
case are applicable to
the time-varying case,
discussed.
Problems
5-1
5-2
b.
x=[~
-2
2~ l~l
-25
) = L --
J.
-20-,
J Ji.
"6(1)
PROBLEMS
CAL EQUATlONS
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
transfer Xo to x1.16 The condition of controllability is that rank [B AB
A"-1B] =n (see Problem 2-20). This condition is identical to statement 3 of
Theorem 5-7. Hence, most of the results in the time-invariant continuous-time
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
discussed.
51-16 lrheorem
,-14
5-17
227
5-10
Problems
5-1
a.
Which of the following sets are linearly independent over ( - ro, ro)?
5-2
a. [~J=[~
y=[O
b.
:I::l+[~lu
lJ
[:J
-4
-3
-1
y=[~ ~ -~lx
c.
=[~O 2~ I~J x + [- ~l u
-25
y = [-1
-20
3 OJ
~J
5-3 Show that a linear dynamical equation is controllable at to if and only if there exists a
finite ti> ro such that for any x o, there exists a u that transfers X o to the zeTo state at time ti'
R/nt: Use the nonsingularity of the state transition matrix.
: controllability of the
tion
161n the lterature, 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 condtion of reach
ability and the condition of controllability are slightlY dirferenl. If A is nonsinglllar, they are
idenlical.
5-4 Show that if a linear dynamical equalion is controllable al to, lhen t is controllable
at any t < to, ls it true that if a linear dynamcal equation is controllable at to, then it is
1::ontrollable at any t> to? Why?
228
Is it true that p[B: AB:: An- 1BJ =p[AB: A2B: : AnBJ? If not, under
what condition wi1l it be true?
5-5
5-6
5-7
State (without proof) the necessary and sufficient condition for a linear dynamical
equation E to be differential1y control1able and differentia11y observable at too
5-8
5-9
b.
x=
c.
x =[~
5-10
Is it possible to choose an
is of the form y(t) = te-' f
5-16 Consider the dyna
state ofthe equation is no
yiste-'fort~l?
5-19
~J x +L~2,] u
to a controllable one.
~ ~J +[~ ~J
x
y = [1
-1
5-20
5-21
O O
O 1J x
5-11
Compute the contro11ability indices and the contro11ability index of the state equa
tion in (3-48).
5-12
J.
5-13
x=Ax +bu
5-14
5-15
5-23
y=[l
1J x
PROBLEMS
.fICAL EQUATIONS
... : A"B]?
229
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?
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 uO,,,,) such that the output
y is te -, for t :?1 ?
n in Prob\em 5-2.
ion for a linear dynamica\
.bservable at too
5-17 Show that the state of an observable, n-dimensional linear time-invariant dynamical
equation can be determined instantaneous]y from the output and its derivatives up to
n - 1 order. [Hint: Compute y(t), j;(t), . .. , yI"-)(t).J
5-19
lJx
to a controllable one.
5-20
5-21
.~
lity index of the equation in
l'
O
1
}'l
O A
O O
O O
5-22
},[~}
iynamical equation
What is the equivalent dy
.r the state space? Or, equi
is the new state equation?
..1. 2
O lJ x
y=[O
O
O
O
y. = Av" + 131.1
y=Cx+Eu
FE:
x =PAQx + PBu
y=CQx +Eu
230
5-24
5-25
5-32
O
O
x= O
O
O
O
y~[:
1
2
O
O
O
O
O
O
O
1
1
1
O
O
O
O
1
1
O
O
O
O
O
O
O
1
O
O
O
O
O
O
1
1
O
-1
O
O
1
O
1
O
O
O
3
2
e,
2 1 1
O
2 1 1
O
O
1 1 1
3 2 1 u
O x+
O
-1 O O
O
1 O 1
1
1 O O
AA q X P ra tional rr
m!-trix G l f(S) such that 1
pG(s) = p in lR(s).
5-33
5-34
i}
Is it possible to find a set of bij and a set of cij such that the following Jordan-form
equation
.{j
y=
['"
O
O
1
O
O
O
O
1
1
O
CI2
Cl3
CI4
C2l
C22
C23
C24
C31
C32
C33
C34
O
b 21
O x + b 31
O
b41
1
b 51
0] ["
";]
C25
Let P be a nonsing
PE
5-26
1
1
O
O
O
~r
where B I is an nI x p matl
sions (n - n) x nI and (n
ifand only if {A 22 , A 21 } is
5-35
b
b'j
b
u
Show that
22
pU =p[B
32
b42
b5
AB
'"
A"
C35
is controllabIe? Observable?
Show that A is cyclic (see Problem 2-45) if and only ifthere exists a vector b such that
{A, b} is controllable. (Hint: Use Corollary 5-21.)
5-27
5-28 Show that if {A, !3} is controllable and A is cyclic. then there exists
vector r such that {A, Br} is controllable. (H int: Use Theorem 5-21.)
fJ x 1 solumn
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
eigenvalue of A. (See Problem 5-26 and use Theorem 5-21.)
5-29
.... _..--:..:=.
PROBLEMS
CAL EQUATIONS
5-32
G[--rS~11
211
211
11 1
321
11 OO O1
1O O
(s +1)
(S~S2;s3~I)J G-l(S~2)J
(s +2)
(s+1)
(s +2)
5-33 A q x p rational matrix G(s) is said to have a left inverse if there exists a rational
matrix Gu(s) such that GLI(s)G(s) =11'. Show that G(s) has a left inverse if and only if
pG(s) = p in lR(s).
5-34
12J
U
'42
~~:
Show that
pU = p[B AB
'21
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
if and only if {A n , A2 } is controllable.
5-35
...
where linearly independent columns of TI' are to be searched in order from left to right.
Hint: Premultiply TI' by
)5
2-
(s +3)(s +1)
PB=[~IJ
'32
231
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.
5-21.)
n n x p matrix, to be control
ks associated with the same
O U nder what
.n with E =.
:ontrollability?
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
left eigenvector a of A such that
).a=aA
and
IXB=O
This is called the Popov-Belevitch-Hautus test in Reference S125. (Hint: See statement 4
of Theorem 5-7.)
_do
. _ _ _ _
_."-"
.-.-
'"
-----
6
Irreducible Realizations,
Strict System Equivalence,
and Identification
6-1
Introduction
=C.C--:_~~I
INTRODUCTlON
,I
; in electrical engineering.
,r an active network that
le subject matter we shall
t is, to determine a linear
scribed rational transfer
Ddern version of network
has a prescribed transfer
"realization" is justified
: system with the transferby using an operational
-em 4-10 that every proper
ime-invariant dynamical
questions. In this chapter
1
233
234
THE CHARACTERlSTIC F
Ir FE 1 is a realiz,
(6-1 )
or
Ll(s)~ det
Ir the denominator 01
then we need a nonze
Definition 6-1
The characteristic poZ
the least common de
denoted by D(;(S), is (
of (;(S).1
Example 1
Consider the rational-
Proof
Let
or
Theorem 6-1
FE 1:
(6-2)
Ll(s)
First we show that {A, b, c} is irreducible if and only if Ll(s) and N I(S) are co
prime. Indeed, if {A, b, c} is not irreducible, then Theorems 5-16 and 5-17
imply the existence of a {A, b, e} such that dim A < dim A and
Ll(s)
where J,(s) = det(sI - A). Since deg J,(s) = dim < dim A = deg Ll(s), we con
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
cible. Hence we have established that {A, b, c} is controllable and observable
if and only if Ll(s) and NI (s) are coprime.
:, AND IDENTIFICAnON
235
le Degree
A( )
g s -e
-
~ree
and characteristic
)ts are the extension of
; degree to the matrix
I(S). lt is assumed that
lon factor). Then the
)f g(s) is defined as the
>, the denominator and
'Ale use det, deg, and dim
limension.
I(S)
+ NL'1(s)
eL'1(s) + N I(S)
L'1(s)
lt is easy to show that L'1(s) and N 1 (s) are coprime if and only if L'1(s) and eL'1(s)
+ N Ils) are coprime. Consequently, we have established that FE is irreducible
if and only if
or
If the denominator of g(s) is not monic (its leading coefficient is not equal to 1),
then we need a nonzero constant k in (6-1).
Q.E.D.
~quation
'hen F El is irreducible
Definition 6-1
:)]
(6-1 )
Gl(S)=.s~l
(6-2)
Ll
5+1
1
-s+1
1
s~l
~-2s+1
G2(S)=,_s~_1
The minors of order 1 of (;1 (s) are 1/(s + 1), 1/(s + 1), 1/(s + 1), and 1/(s + 1).
The minor of order 2 of (;1(S) is O. Hence the characteristic polynomial of
(;1(S) is s +1 and (;I(S)=1. The minors of order 1 of (;2(S) are 2/(s +1),
l/(s +1), 1/(s +1), and 1/(s +1). The minar of order 2 of (;2(S) is 1/(s +1)2.
Hence the characteristic polynomial of (;2(S) is (s + 1)2 and 15 (;2(S) = 2.
11 is also called lhe McMillan degree or lhe Smilh McMillan degree. The definition is applicable
on.ly lo proper ralional malr,ices. If G(s)'S nol proper, lhe d'efinition musl be modified lo include
lhe poles al s = oo. See References 34.62.83. and S185.
236
square matrix] and different from the least common denominator of al1 the
entries of (;(S). Ir (;(S) is scalar (a Ixl matrix), the characteristic polynomial of
(;(s) reduces to the denominator of (;(s).
We remark that 1
nonzero constant [se,
the characteristic po:
6-1' can be establisr
reader is referred to 1
Example 2
Consider the 2 x 3 rational-function matrix
*Theorem 6-2
S
S
G(s) =
[
+l
-1
s +1
(s +1)(s +2)
1
(s
+ 1)(s +2)
The minors of order 1 are the entries of (;(s). There are three minors of order 2.
Theyare
s
1
s+1
1
(s+lf(s+2)+(s+lf(s+2)=(s+lf(s+2)=(s+l)(s+2)
(6-3)
sil
s +4
--'-+
=----
s +1 s (s +1)(s +3) (s +1)(s +3)
(s+ 1)(s+2)s
(s+ 1)(s+2)(s+3)
s(s+ 1)(s+2)(s+3)
Hence the characteristic polynomial of (;(s) is s(s + 1)(s + 2)(s + 3) and b(;(s) =4.
I
be a realization of ti
(controllable and obs
or
det (sI
dir
where k is a nonzero ,
The irreducibility
case will be establishe
bility without relying
here. In fact, this tl
realization discussed '
analogy to Theorem I
of the characteristic p,
6-3
Irreducible I
G.
9(S)
*Definition 6-1'
Consider a proper rational matrix (;(s) factored as (;(s) = N r (s)D r- 1 (s) =
Dl(S)N,(s). lt is assumed that Dr(s) and Nr(s) are right coprime, and D{s) and
N(s) are left coprime. Then the characteristic polynomial of (;(s) is defined as
det Dr(s)
or
det D(s)
loss of continuity_
, AND IDENTIFICATlON
237
We remark that the polynomials det D,.(s) and det D/s) differ at most by a
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:
: + 1)(s +2)
or
(6-3)
11
3)
t- 2)(s
x=Ax +Bu
y=Cx +Eu
6-3
g(s) =
A
/)
s" +as"-
,
~_.
L> _ _
(6-4)
where (3 and ai, for i = 1, 2, .. _,n, are real constants. Note that the leading
coefficient of D(s) is 1. Let u and y be the input and output; then we have
lS
D(s)y(s) = (3u(s)
(6-5a)
(6-5b)
238
up to the
(n -
XI(S~
x
J [1
y(s)
A()D. 2(s) D.
sy(s)
=
=:
=:
.
.
Xn(s)
Sn-y(S)
X S
Define
IF
A
y(s)
(6-6a)
where
S"-
X(t)~ y(t)
(6-6b)
x=
O
O
O
O
1
O
O
O
O
-(Xn -1
-ct n -
O
-(Xn
y= [
l(s
(6-7)
x+
O
O
O
(6-8a)
Hence,
O
f3
-(X
O ]x
(6-8b)
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).
e = [f3
[O O ...
O 1] and
and
g(s) = c(s]
f3
l(s
Figure 6-1
coefficients ofg(s) in (6
_-_
....
...
...
--_.... ,_._
.;;.:...:::...~..::.::.:.:.:.::::......,,_.- - '' : : " : : : " - ~ _ . _ - - - - ~ - , - - - - - _ .
~-_
_-~-_
239
ANO IDENTIFICATION
~.
(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
ine
(6-6a)
(s)
where
(6-6b)
-1
O]
0l
O
...
~~,
(6-9)
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
Xn
+ f3u
~(s)~ det
(6-7)
(sI - A) = sn +CtlSn-1
+ ...
+Ct n
(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
O
O
13
~(s) [1
(6-Sa)
(SI_A)-'b~[l
13
(6-Sb)
...
Hence,
11'=[0
S S2
Ct n
-1
s
O
O
O
Ct n_ 1
s
Ct 2
In
:
O
-1
S
+Ct 1
13
: = ~(s)
O
13
s.
: ]
[1
sn-2
Sn- I
(6-10)
I]and
and
g(s)=c(sl-A)-lb=[l
13
~(s)
13
O .. , O
(6-11 )
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).
..
_~_
.~~
..
_ _
,~~
" , ' _ _
_ _
~~_,._,_'~'_'._
0_ ' _ . '
,_~
"~
-------~-----~------------
i
-1
240
11
!
!
(6-12)
Pi,
where Ci i and
for i =0, 1, 2,. _., n, are real constants. It is assumed that
By long division, 91 (s) can be written as
Ci o i= O.
A
f31 Sn - 1 +f3 zs n-Z + ... +f3n
6A
(6-13)
where e = 91( 00) = Po/Ci o. Since the constant e gives immediately the direct
transmission part of a realization, we need to consider in the following only the
strictly proper rational function
N(s) 13 sn - 1 + 13 sn - Z + ... + 13
A( )
1
Z
n
g
s-6
_
)6
_
D(s
sn +a 1sn 1 + ... +a n
(6-14 )
The term
1)(t)
y<Z)(t)
y<n -1)(
then x = [Xl X z
in (6-17) yields
Let u and y be the input and output of g(s) in (6-14). Then we have
D(s)y(s)
N(s)u(s)
(6-15a)
D(p)y(t) = N(p)u(t)
(6-15b)
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).
Differentiating
Xl
in (1
(6-16)
A dynamical equal
canonical formo The 1
(6-18) is derived from
This can be verified by
f10w graph or computi
e(sI - A
withthe aid of (6-10), \
.=.-=... =.
__ .=~=
.. =;=="..=..=~_.=.-.=.-=
... = = = = = . =......-.... =-=.:..-=.=..
~~=~
=.
"ND IDENTIFlCATlON
241
The term
N(s)
s. It is assumed that
~g(s)
+e
(6-13 )
in the right-hand side of (6-16) gives the response due to the input u(s); the
remainder gives the response due to the initial conditions. Therefore, if all the
coefficients associated with Sil - 1, Sil - z, ... , SO in (6-16) are known. then for any
u a unique y can be determined. Consequently. if we choose the state variables
as
XII(t)~y(t)
XII_(t)~y<I)(t)
(6-14 )
+Cl. l y(t)-f3u(t)
XII_Z(t)~y<Z)(t) +CI.y(\)(t)
then x = [Xl Xz
in (6-17) yields
(6-17)
Then we have
(6-15a)
lI
lt is well knowh
o), ... , yI"-)(to)
'J )u(t).
(\)(t
determinable. ln this
He variables as we did
. the form x =Ax+ bu,
= Ax + bu, y = ex + e I u
J(s)/D(s) in the form
lo be chosen.
19 the terms associated
6U(0)]S"-Z
+ ...
X = -o:"X" +/3I1U
x~
Xz
X3
x,J
y=
O O O
1 O O
O 1 O
O
O
O O O
O O O
[O O O
Xz
x,
-o:z
-o:
/311-
P.
1+
JlX"-J
XII
,u
lf3z
/3
]x
(6-18b)
- 3)(0)
(6-16)
(6-18a)
with the aid of (610), where the "prime" symbol denotes the transpose.
242
IR.
Controllable canonie
x(,)J:~i:~~
V~
e
cA
cA 2
o
O
O
O
O
O
lxn(s}J
This definition implies
O
1 -!Xl
-!Xl -!X 2 +!Xi
O
O
O
O
O
1
sxn(s) = snv(s)
be
(6-19)
(6-20)
where
O
O
O
A=
which, together with lX =1=0, implies that the controllability matrix of (6-18) has a
rank less than n. . Hence if D(s)and N(s) are not coprime, the realization in
(6-18) is not controllable.
o
-el"
e = [f3n
AND IDENTIFICATION
243
and
(6-21 )
(6-22)
Equation (6-21) has the same form as (6-5a); henee we may define the state
variables as, similar to (6-6a),
ynamical equation (6-18).
01'
(6-24)
x V is nonsingular for
and N{s) in (6-14) are
'a/form rea/izaton.
lel1 if D{s) and N{s) are
1{s) = deg D{s) = dim A
orem 6-1. lf D{s) and
1 (6-18) cannot be con
\. In the fol1owing we
me, then (6-18) is not
D(s); then we have
=0
=0
y(s) = [.8n
.8n-1
.81]
...
ves)
sv(s)
=[.8n
S-l V(s)
01',
.8" - 1 .. , .8 l]X(t)
(6-25)
(6-19)
X=Ax +bu
(6-20)
y=cx
(6-26)
where
A=
ymatrix of (6-18) has a
ime, the realization in
O
O
O
1
O
O
O
1
.. 1
-a"
e = [.8n
.J
-(Xn - I -a.- l
.8n-1
.8n - 2 ...
-!Xi
.81]
b=
O
O
1
244
IRREO"
This is a realizalion of 9(s) in (6-14). Unlike (6-17), there are no simple relation
ships between Xi, u, and y. The dynamical equation in (6-26) is always control
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
of (6-26) is drawn in Figure 6-3.
Example 1
h(2) =
h(n) =
-(1.
+0 =
-(1.
h(n
y=[
2.5
0.5J x +2u
i= 1, 2
These equations are obt
h(l)
h(2)
y=[O
-6
lJ x
+ . .. + f3
-(1.
0.5
+ 2u
h(cx)
lt is caJled a Hanke/ me
{hU), i = 1, 2, 3, ... }.
invo/ved in H(1., f3).
Theorem 6-3
pH(n, n)=pH
where pdenotes the rar
Proof
Figure6-3
(6-26).
245
/3 os + /31 s - 1 + . .. + /3.
A()
9s =
s +al s
+ ...
(6-27)
+a.
+ ...
(6-28)
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(n
+/3.
+ i) =
a.hU)
-a 1h(n
+i -1) -
a2h(n
+ i - 2) _ ... -
(6-30)
i = 1,2, ...
These equations are obtained by equating (6-27) and (6-28) as
(/3os +/31 S-1
+ ... +/3.)
=(s +als-1
+ ...
+ ...)
and then "equating the coefficients of the same power of s. We form the a x {3
matrix
h(1)
h(2)
h(3)
h(2)
h(3)
h(4)
h(/3)
h(/3 + 1)
h(4)
h(5)
h(/3
+ 2)
h(a
+/3 -1)
]
(6-31 )
[
h(a)
h(a
+ 1)
h(a
+ 2)
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
(6-32)
Proof
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.
246
Because of (6-30) the (n+ l)th row of H(n+ 1, (0) can be written as a linear
combination of the n rows of H(n, CX)). Hence we have pH(n, (0) =
pH(n+ 1, (0). Furthermore, we have pH(n, oo)=n; otherwise, there would be
an integer smal!er than n with the property (6-30). Because oflhe structure of
H. the matrix H(n + 2,(0) without the first row reduces to the matrix H(n + 1, ro)
without the first column. Hence the (n +2)th row of H(n +2, (0) is linearly
dependent on its previous n rows and, consequently, on the first n row of
H(n + 2, (0). Proceeding in this manner, we can establish pH(n, CX) ) = pH( cx), (0) =
n. Again using (6-30), the (n+1)th column of H(n, (0) is linearly dependent
on the columns of H(n, n).
Proceeding similarly, we have pH(n, n) =
p H(n + k, n + 1) = n for every k, 1= 1, 2 ....
Now we show that if (6-32) holds, then g(s) =h(O) +h(l)s-I + ... can be
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
(6-29) we can compute {/3, i =0,1,2, ... , n}. Hence we have
g(s)=h(O) +h(l)s-1 +h(2)s-2
+ ...
IR
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
x=Ax +bu
FE:
+cA bs-
+...
(6-33)
i=1,2,3,...
(6-34)
fl(2)
h(2)
h(3)
h(n)
h(n + 1)
-a 2
-a,
h(n + 1)
'h(n +2)
h(2n -1)
h(2n)
O
1
is a controllable and ,
Theorem 6-3, we have
h(tJ-i- i)= -'Qlh(a
AJ ::~:
J,
Ab,
2
lh(O" + 1)
The effect of the multip
1, or equivalently, shifts
c, cAkb just picks up th
cb=,
(6-35)
-al
c=[
1
O
h[ O
y =cx +eu
O
O
be written as a linear
we have pH(n, 00) =
herwise, there would be
:cause ofthe structure of
o the matrix H(n + 1, oc!)
. H(n + 2, oc!) is linearly
, on the first n row of
)) is linearly dependent
, we have pH(n, n) =
) +h(l)s- 1 + ... can be
le condition pH(n, n) =
!} to meet (6-30). Using
lave
sn- 1 +
+f3n
,-1 +
+lXn
Note that there is one more row than column, and the Markov parameters
up to h(2n) are used in forming H(n + 1, n). Now we apply the row-searching
algorithm 3 discussed in Appendix A to search the linearly independent rows of
Hin (6-35) in order from top to bottom. Let the first (J rows be linearly inde
pendent and the ((J + l)th row of H be linearly dependent on its previous rows.
Then Theorem 6-3 implies that the ((J +k)th rows, k = 1, 2, 3, ... , are alllinearly
dependent on their previous rows and the rank ofH(n + 1, n) is (J. Hence once a
linearly dependent row appears in H(n + 1, n), we may stop the search. We shall
call the ((J + l)th row of H(n + 1, n) the primary linearly dependent row; the
((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
ing algorithm will also yield {a, i = 1, 2,
, } such that
[al
az
.,.
a" (Il O
O]H(n+l,n)=O
x=Ax +bu
y=cx +eu
with
O
O
1
O
A~[ O
-al
s a realization of g(s) in
: a different realization.
ith deg D(s) = n. Here
;e the degree of g(s) may
le recursive equations in
[h(ll
h(2)
-az
-a3
O
-a,,_
(6-38)
h((J -1)
-a"
h((J)
O]
e= h(O)
(6- 37)
i= 1,2,3, ...
=[~i~~
h((J + 1)
AZb
A'b
h((J +2)J
(6-39)
h(k +(JJ
h(n)
h(n
O
O
(6-34)
O
1
b-
e=[
(6-33)
(6-36)
This equation expresses the primary linearly dependent row as a unique linear
combination of its previous rows. The element ({l corresponds to the primary
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
+ ...
247
+ 1)
eb =h(I),
(6-35 )
(6-40)
h(2n -1)
h(2n)
For computer compullttion, numerically stable methods should be used. See Appendix A.
248
This shows that (6-37) is indeed a realization of g(s). The controllability matrix
of (6-37) is
[b
Ab
...
A"-lbJ=H(u,u)
The Hankel matrix H(u, u) has rank u; hence {A, b} in (6-37) is controHable.
The observability matrix of (6-37) is
e
cA
cAz
O 1 O
eA"-1
O
O
O
100
O O 1
IR
O O
We use an example to i
a lordan-form dynami
general case. Assume
and assume that D(s) c~
assume that g(s) can be
A( ) = -el
gs
(s -;
Example 2
Consider
:- -6-;
_1-
, Z
_1- _J.
" Z ,
4
H(S, 4) =
,--
_1
_1-
_2
_1-
_2
_J.
1-
1-
,__ 1-1
Z
J.
U
~.
X~U5
y=[
1
O
-1
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).
, AND IDENTIFICATION
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
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.
: controllability matrix
249
(6-37) is controHable.
Jordan-canonical-form realization
9(S)=
ll 3 +
(s-..1.)
e12
(s-..1.)
e l3
(S-Al)
e2
(s-..1. 2 )
+_e_3_
(6-41)
(s-..1. 3 )
The block diagrams of 9(s) are given in Figure 6-4. In Figure 6-4(a), the coef
ficients e 11' e 12' e 13' e2, and e3 are associated with the output. In Figure 6-4(b),
(a)
e13 \ - - - - - - - - - - - - - - - ,
XlI=AXI\+el\u
x12=Axi+e12u
x13=Ax13+ e 13 u
Xl!
(b)
. _=._=-._="=---=-=
..._=._=-. -'='-=-'-=-=--='--=-'====-.=--=.--=--=-=---== ... --
~~"~.
_.~.-=.=.-~._=.-_.=-_.=--=_
250
--C'=~="-,"=c----c-.=c<=="--:=--==_=-------='=='C~'O'
...... - _ .....
11
they are associated with the input We note that every block in Figure 6-4 can
be viewed as consisting of an integrator, as shown in Figure 6-5. Hence the
output of each block qualifies as a state variable. By assigning the output of
each block as a state variable, and referring to Figure 6-5, we can readily obtain
the dynamical equation for each block of Figure 6-4 as shown. By grouping
the equations in Figure 6-4la), we obtain
O : O
1: O
1'1:
O
1.[
;'1
O
O
O
-o'---6---6-:--~;-:
O
O
O x
O
O
O
1 u
1
1
O ""0"" : ;.~
Equation (6-42) is in the Jordan canonical farm. There is one Jardan block
associated with each eigenvalue. The eq uation is clearly controllable (Carollary
(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
chosen as shown, then the dynamical equation is
XII
.1,
O: O
01
XlZ
-X 1 3
Xz
,X3
..
__
O
O
O
O
y=[O
o'
Az ' O
O: ..1. 3_1
'- __ 1 __
ell
e lZ
el3 u
ez
e3
---+--
and
Then it can be easily
transformed into
y=[
3.. : ;: ax + bu
,--------------------1
s-a
1] x
~x
-
numbers. In this ca
puter, for complex m
this can be taken car
will be demonstratec
assumed to be reaL if
/:, is also a pole of 9(5).
following subequatioi
(6-42a)
(6-42b)
'-"--'-'-'-,,-~--'"
where Re A and 1m A e
tively. Since aH the co,
analog computer simul,
iorm dynamical equati,
transformation introdu
Example 3
11 I
l\'
1
x=
I
I
y=[
1 +2i
O
O
O
O
1
1 +2i :
O ,'
-
O
'0
-i
- l.
,:
.
251
, ANO IDENTIFlCATlON
o
O
1 u
u
[ ~lJz =[AIO A~ J[X1J +[~J
b
(6-43a)
cIJ[::J
(6-43b)
(6-42a)
1
1
XZ
y=[c
(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
introduce the equivalence transformation x = Px, where
p= [ 1
il
and
-il
_~
[1 -;IJ
-2 1
il
Then it can be easily verified that the dynamical equation in (6-43) can be
transformed into
~ll=[
x2
ReAl
-1m A
y = [Re e
1m
ImA1J[~J+[
Re Al
2Reb l
-21m b
x2
Ju
clJ [::]
(6-44a)
(6-44b)
where Re A and 1m A denote the real part and the imaginary part of A, respec
tivel)'. Since al! the coefficients in (6-44) are reaL this equation can be used on
analog computer simulations. Another convenient way to transform a Jordan
form dynamical equation into an equation with real coefficients is to use the
transformation introduced in Problems 6-14 and 6-15.
Example 3
Ix
---------
I
x=
.~
y=[
1 +2i
1
O
D
O
O
1 +2i
()
O
e
- - - - - - - - - - - - - - '- - - - - - ... - - - - - - - ... -.
O
o
1 -2;
1
O x
O
O
O
1-2;: O
O
o '----0-------0---:--2
-;
2J x
2-3i
1
2 +3i u
1
2
(6-45)
252
REALIZA
O
p=
O
O
O ,
1
O
1
O -i
O ,,
,
i
O -i ,
O
O
O ,,
_____
_, -
O
O
O
O
-1- -
__ , _ _ _ _ _ _ _ _ _ 1___
Example 4
[p2
+ o:((
y
I
e(t) = 130(1)
b(t) = fi(t)
b 2(t) = fi2(t)
Since the time functions
the differential equatior
ofthe form in (6-48). \\
the relations between b,
9(S)=91(S)+92(S)+'"
where 9(s) and y(s) are transfer functions of degree I or 2. We then realize
each Ms) and y(s), and then connect them together. Thc first one is called a
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).
conc1uding this section, we shall briefly discuss the se"tup of dynam ical eqllations
(pn +a(t)pn-l
+ ...
+an(t))y(t)=fi(t)u(t)
(6-46)
6-4 Realizations,
Transfer Function:
In this section realizatio
studied. By a vector rati
function matrix. Consi<
:, AND IDENTIFICATION
253
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
+ IX (t)p + IX2(t)]y(t) =
[,80(t)p2
+ ,8 (t)p + ,8 2(t)]U(t)
(6-47)
-IX(t)
. gIs).
or 2. We then realize
Thc lirst one is called a
111. This typc of realiza
Crcncc S'+7).
JI equations. Befo re
) of dynamical equations
nth-ordcr. \lnear, time
J)u(t)
(6-46)
X2
[b(t)]
b 2(t) u
(6-48a)
O ] x + e(t)u
y=f
I
J[Xl] +
(6-48b)
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),
we obtain
y=
py =
X2
ji = p2 y =
+b(t)u(t)
+ (t)u(t) + e(t)u(t)
-1Y.2X -1Y.X2
+b 2u +bu +bu
+ eu + 2 + eii
u, and , we
(6-49)
IX (t)~o(t)
~ 1Y.2(t ),8o(t) -
fio
Since the time functions b, b 2 and d can be solved from (6-49), we conclude that
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
(6-50)
=========",-=',='-=,.".,
254
REALIZI
IRREDUCIBLE REALIZAT10NS, STRicr SYSTEM EQUIVALENCE, AND IDENTIFICAT10N
(;(s)
Consider
l
e ] rgl(S)]
+ ~z:(S)
eq
Example 1
Ginto
(6-51 )
(8
G(s) =
gq(s)
(6-52 )
+3
+s1)(s
+2)
5
+4
+3
Hence a minimal-dime
O
O
O
O
x=
x+
O
-Cf.
YI]
~2
[Yq
-Cf.
n- l
n- Z
[Pln PI(n-l)
Ptn PZ\n-l)
Pqn Pq(n-l)
-Cf.
Pll]
[el]
P~l x+ e u
Pq"
(6-53a)
O
1
O
-Cf.
tq
=[el : ez: : el
[el: ez : ... : el
x3
O
1
O
O
O
1
Xl
Xz
= [O
255
:, AND IDENTIFICATION
and
Example 1
G into
Consider
(6-51 )
A
G(s)-
(S
+3 +2)
+ 1)(s
[OJ
s +4
--
[s
+3j
(s + 1)(s
+2)
1
s+3
--
s+3
: : .... 7pI2nn,]
(6-52)
OJ
1
[ (s+3f J
[ 1 +(s+1)(s +2)(s +3) (s +1)(5 +2)
OJ
1
[S2 +6s +9J
[ 1 +s3+6s 2 +11s+6 s2+3s+2
+ ... +Pqn
Xl]
[ ~2
[OO
-6
X3
o
,+ u
y= [
1
O
-11
6
3
(6-54 )
(6-53a)
I
,,]
e
eq
'e] +----.-----
1
.
= [ e .. e .....
1,
2
.
p
Sn + a 1 Sn 1 + ' .. + a n
the control1able-form
(6-26), we see that the
n sn -
X [ 1-'11-
+ 1-'12
n sn - 2 + . .. + 1-'1/1
n .: 1-'21'
n sn - 1 + 1-'22
n sn - 2 +. . .
+{J2n:"': {JpIS"- +{Jp2 S 2 + .. - + (Jpn]
/l-
(6-55)
XI
X2
X3
O
1.
O
O
O
xn
= [O
O -a n
P2n
f3 In
O -a n- I
PI(n-l) P2(n-l)
O -a n - ., x+ PI (n- 2) P2(n-2)
.~
-al
O
PII
] x + [el
P21
.. e2
Ppn
Pp(n-I)
Pp(n-2)
PPI
ep
] u
(6-56)
,_._ > __
256
. --.".
__._.
" . ~ _
_~
, __
. _ . _ .
~ _ . ' - .
.-._--
-~
..- _...
..... _-----
-_ .. '- ..
~._."
..
~---
REALIZA
Consider a q x 1 prop<
A
A
G(s)
=
"
[
'Example 2
g,
"(cx I -+
(6-57)
we draw in Figure 6-6 a block diagram of(6-57). With the state variables chosen
as shown, we canobtain the lordan-form equation
1J [- 1
X
[
~:
- -
1:
0] \01
~ ~- +~~ x +t~-J ~
- _:-
y=[ ~
2:-IJ [oJ
'
-2 :
x+
to describe the block diagram. The equation can be readily shownto be con
trollable and observable by using Theorem 5-21. Hence it is an irreducible
lordan-form realization of (6-57).
r------j-------c~
} - - - - Y
T~l:;~X
Note that ni, i = 1, 2, .
number of rows. The i
is equal to or lare:er t
{g{s), i = 1,2, ... , q} .
. . . , q} is not available,
of the denominator of 9
not be irreducible.
Now we shalI app
independent rowsofT ii
T and Theorem 6-3, if Ol
T in (6-59), then all sub~
(J be the number of Iim
then the first (X rows 01
However these linear inc
in T because they may 1
hav~ (Ji.:s;,C(j, i=2. 3.,
(J =Cl.
The row of H
rows ofT is called the pr
of Iinearly independent
Iinearly dependent row.
the row-searching algori
(x, +1
u --f-------.I
k, =[a,,(I)a,,(2)"'a, ,(u,)
U ,
Figure 6-6
257
E, AND IDENTIFICATION
91(Sj
(;(s) = 92.(S)
+"'J
[h1(0)+h1(l)S_1 +h 1(2)s-2
hz{?) +h 2(l)s - 1 + h 2(2)s - 2 + ...
gq(~)
+...
(6-58)
The Markov parameters hU) can be obtained recursively as in (6-29) and (6-30).
For each 9;(s), we form a Hankel matrix H defined in (6-31). We then form the
composite matrix
lction
(6-57)
[Hi~/+--( J}
jJ)
(6-59)
uq
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
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
U be the number of Iinearly independent rows in H. If all 17(s) are 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,
k'=[UlI(I)UlI(l)"'all(o-) 1: O
k;=[u,,(I)a,,(l)"'a,,(o-) O:
k q =.[a q,(I)uql (l) "'aql(o-I)
.'
..
'-------y---""
U
ad')
+I
... O
<X q
O 0 O:
: O
"'0
: O
"''',,(0-,) 1 0
+1
O
O
O]
O]
a2
(6-60)
258
such that kiT = O. Note that the row vector k j has, except element 1, only
((JI +(J2 + ... +(JJ possible nonzero elements. This is a consequence of the
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
T. See Appendix A.
Now we c1aim that (;(s) in (6-58) has the following irreducible realization:
x = Ax + bu
(6-61 a)
y = ex + eu
(6-61 b)
REALIZA"
e=
-(131(1)'"
-(1,
with
hl(O)j
h 2 (0)
e= .
hq(O)
O
O
o
o
1
O
-a(2)
-a(3)
A=
for i > j,
O
O
O
O
Let
Cj
b. =
I
hI(J)
The matrix
c=
1 O ... O: O O ... O:
,
"
O O ... O' 1 o ... O""
[ ::
1
1..
o o
0:0
,
'-v---"
(JI
0j}
1:
0"
: : 1
(J2
O
.
.
1
I
: I
'o
: 1::
:O
I
l.
'-v---"
(Jq
q (6-62a)
--_.. _._._--------_._-----------
0, ANO IDENTIFICATION
(6-61 a)
(6-61b)
----
259
O: ;00"'0
.".
O:
:0 O
O
O :, 1
O
:0 O
O
-a31J) :-a32(1) ... -ad(J2): O O ... O:,.. ':0
, O
O
O
O :1 O "0: :00'''0
o
c=
o
:O O
.: ,,
:0 O .". O:
'--------.....--'
(J
(J2
h(0)J
(6-62b)
e=
---
h2 (0)
hq(O)
h(k + 1)
hl(k +2)
J x (J) matrix
k =0,1,2, "..
(6-63)
hq(k +(Jq)
"O
0J}"
o
O
..
'.
(Jq
(6-62a)
cq A
C1
q- \
'-~
260
"
'.,__
~ _
__
- ~ "
~ _ ,
__
" _ ~ ~ , _ "
'_'_~'
"~_'
~"~"_
-~_.,,~._
".__,,'.'-0._ , .,
'\'''_,,'')+,,'6_,\'7+ S ' 8 _ ..
~~~~-',
ReAWA
Consider
()
__.,__
Example 3
Gs
k = [O
k 2= [_1
+"'J
O
_.1
9
'---'--
s3(~+I)
uI
O::
:(f:
,
- J
;0:
:, 1:,
1
I
-,3
-,3
5
-7
S
-7
9
'\
-,3
'- '
-,3
5
The elements
Corres
an irreducible realizati~
fi =5.
o
O
X= O
O
l'
T=[HdS.5)J=
-7
9 -11
H 2 ( 5. S)
- : : -- -O" - - 6--- -1 - - -:..:: 1
O
O
1
O
1
-1
1
-1
1
-1
1
-1
\
-1
1
-\
-\
-1
y=[~
T~e (u I ~ 4)th and (u 1 +0
wlth the slgns reversed. T
of9 l(s)andthefirstu 2 Mi
depends only on u 1 and u
It is possible to searci
order. First we expand (;
-1
3
-S
7
1
-\
3
-5
1
-1
3
1
2
(;(s) = H(O
1
O
-1_____________
O
O
_l
3i> O O
6
- 1 - 1
O
J6
O
O
-\
1
I
-\
-1
\
1
1
O
O
O
O
1
O
O
O
~(!~ ----~(~! -
H(2)
T=:
.
O
O
H(ex)
H(ex +1)
O
O
:Il
C[:
:})
O
O
O
1
-4
-3 }
8 U I =4
O
\2 -24
O -36
72
O O
O
m
o
o
o
o
U2 =
I
H(3) - - - I
:
,
H(ex+1)
H(ex +2)
Ii
H
4S~e
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
O
k l =[ O
k 2 =[-.l3 _19
~:(~
_-.1.
18
'(JI
\8
:,
O
O
O
O
O
~:l}
O
O
OJ
OJ
'-v--'
~4
(J2
=1
=5,
1
3
5
7
9
1
-.1.-
-3
5
-7
9
-11
x=
O
O
O
1
O
O
O
1
O'
O
O
1
O -1 -2
1 - - - -[ - - -
"3
1 - - -...= i
1
1
1
-1
-1
1
1
-1
-1 - - -
-18
1-
i -
O
O
O x+
O
0
18:
Hence
o
O
1
1
O
O:
0J
O : 1 x
The (JI = 4)th and (J 1 +(J2 = 5)th rows of A are taken from the k l and k 2 rows
with the signs reversed. The vector b consists of the first (J 1 Markov parameters
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
It. is possible to search the Iinearly independent rows of T in a dilTerent
order. First we expand (;(s) as
t finally obtain
O
~}:~
1
-4
-]
O
8 (Jl =4
O 12 -24
72
O -36
O
O
O
-6
--(J2 = 1
O
O
O
O
O
O
O
O
O
O
O
O
O
-cF
H(ct. + fi
.,.
'DW (&-64)
-l)J
H(ct.+f3)
There are ct. + 1 block rows in T; each block has q rows. The ith row of every
block row is associated with i)(s). Except from having more rows and dilTerent
ordering of rows, the Hankel matrix in (6-64) is basically the same as the one in
(6-59). Now we shall apply the row searching algorithm to search the Iinearly
independent rows of t in order from top to bottom. Clearly, if the ith row of a
block row is linearlydependent on its previous rows, then aH the ith rows in the
subsequent block rows are. linearly dependent on their previous rows. 4 After
4See the discussion in the subsections on the controllability indices and observabilily indices in
Chapter 5. See also Shemes I ano 2 of lhe scarch of Iinearly inoependent columns of lhe. con
trollability malrix in the secano half of Section 7-2.
262
the completion of this search, we then rearrange the rows of tinto the form of
(6-59) as
__
H q( a
+ 1, [3)
\'
Example 4
G(S)=[~J)-3 +G}
(6-65)
J \' q
1
O
1
-1
4
1
-1
2
O
3 -8 -1
6
1 -2 -1
3
-5
12
3 -12
-1
2
1 -3
7 - 16 -5
18
1 -2 -1
3
T..
Now from the row searching algorithm and the rearrangement of the coef
ficients of combinations according to the rearrangements from t to 1', we can
obtain, similar to (6-60),
~+I
. k, =[ulIll) u,,(2)
k, =[u"ll) (/,,12)
~+I
u"I",) 1 o
u"I",) o o
o:
o:
o o ... o:
~+I
o o .. o: : (/,.i1~'--O:
u,,(I',) I o ... o: : (/,.111'" ",.(",,) o o ... O]
u,,(1)
u"t",)
(/,,(1)
o o .. ' o: .. : (/,,,11)'"
~
\'}
u".!".) 1
----v-'
\'"
o .. O]
(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
x=Ax +Bu
y=Cx
(6-67)
A2 ...
with
~l
A 2
A=
.
lA"
A 22
...
Aq
Aq2
... A qq
k2
A"l
A 2q
(
(
k row corres
where Ai' B, and are given as in (6-61) and (6-62) with (J replaced by V and
aJk) taken from (6-66). The proof of this statement is similar to the one in
(6-60); it is more tedious, however, because of the necessity of tracking the
rearrangements of coefficients. The proof does not involve any riew idea and
will be omitted. Insteadwe give an example to iIIustrate the procedure.
= [
= [-1
T and k as
l'
r.~JT~[ -16
Lk
2
263
Example 4
+ ...
1
O
-1
-1
3
1
4 and
-8 -1
1
6
-2 -1
-5
12
3 -12
2
1 -3
7 -16 -5
18
1 -2 -1
3
-1
1
O
-1
-1
3
1
*
*
1
O
O
O
O
1
O 1
O O 1
O O O
O O
1 -3
O O O 1-1
-----------------O 1 1 -3
5
O O 1 -1
1
-----------------1
1 -3
5-7
O 1 -1
1-1
-----------------1 -3
5 -7
9
1 -1
1 -1
1
------------------3
5 -7
9-11
-1
1 -1
1 -1
,,+1
) : . ..
~)00'01
: (/ I q~ 1)'
l/ 111\ "l/
) ':
:.
."'
1\)'"
(/21/
",,1",)
-
o o ..
.
01
o .'
01
(6-67)
:~:1
'qq
(j)
-3
:I,-1
O
~2;
O
O
O
O
O
------------------O
O
O
O
O
O
O
O
O
O
------------------O
O
O
O
O
O
O
O
O
O
(6-66)
O
O
.~
1',
O
O
kl =[ 6 O:
kz
[-1
4
0:-1
-18 : -1 O : 1 O: O O]
3: O 1 : O O: O O]
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
f and k as
[~Jt~[_~ _~ -~ :~;
. ~
VI
0:0
O: O
-V~
;)
'--v--'
Vz
264
,,O,
:0:
'1'
-3
1
1
1 -3
-3
5
1 - 3
5 - 7
- 3
5 - 7
9
x -:6:- -- -0- - - - ----(
5
-7
9
-11
1- _1
=[~J
~Primary
dependent
row of H 1
-- i
-~
'0'.!
O
1
O
1 - 1
1 -1
1
-1
1 -1
1-
-1
1
- 1
1
1
-1 ~ Primary dependent
1
row of Hz
-1J
IRREDUCIBLE REALIZ,
introduced in CAkB.
i = 1, 2, ... , q, on digito
which is q(r:t. + 1) x 13, is
x 13; however, in the se;
row, say the ith row ir
block rows can be skir
thaJ1 the T in (6-59), th
independent rows in t
on computer computat
We discuss now an i:
Consider
C(s)
= [gl(S)
TJ:g
y=[
O:
1
O
O
O
O
O
O
1
-3
::>
The v1th and (v 1 +vz)th rows of A are taken from the k1 and kz rows with
the signs reversed. The vector b consists of the first V1 and Vz Markov para
meters of gl and gz The form of C is fixed and depends only on V1 and Vz. I
gzl
lH(a
It is a la + 1) x fip matrix
in order from top to bot
independent and the (v +'
Then there exist a(i), i' 1
[a(l)
Remarks are in order regarding the use of the Hankel matrices in (6-59)
and (6-64) in the realizations. First we show that the v in (6-65), which are
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
X~,
L- a(1)
Hence the largest order of A in (6-67) is smaller than'or equal to the onein
(6-60). This implies that the data up to CAV+P+1B are used i~ the realization
in (6-67); whereas, the data up to CA"+P+ 1B are used in therealization in (6-60).
As discussed in Section 5-8, the larger (he power of A, the more errors may be
y= [
-(;
is an irreducible realizatio
proof of this statement is s
1, AND IDENTIFICATION
~Primary
dependent
rOW ofH I
II
265
~ Primary dependen!
row ofH z
...
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
T=
f~(ex+1)
[a(1)
H({3)
H({3 + 1)
H(ex +2)
~(ex + {3)
a(2)
a(v)
1 O ...
iJ ~
I
O
L- a (1)
~he postmultiplication of
: the linearly independent
; the linearly independent
Itly, we conclude that the
irreducible realization of
xample4.
dices, we have
y=[ 1
O
-a(2)
O
-'l(3)
O]T =0
l l
O
O
:
H(2)
f"(l)
-a(v)J
O ]x
+:
!
(6-68)
H(v - 1)!
LH(v)
+ H(O) u
q}
H(3)
VI
= 1, 2, ... ,
H(2)
H(l)
H(2)
[.
"
~t
266
IRREDUCIBLE REALI2
+R implies
and
y =Cx +eu
where A, b, C, and e are of the form shown in (6-53), u is the ith component
of u and y is the q x 1 output vector due to the input U, then the composite
dynamical equation
l}}l1'
Y
[C
sVm(s) =
(sm
fIJ
A2
O
C2
and
Consider
C(sI-A)-B
,.,
(6-70)
G(,
Op
Op
OP]
[OP]
x+:
u
Op
Op
-lXm-I p -lX m_ 2 I p
Ip
-1X1I p
R
Op
Ip
J x+
+i)=
-1XH(m
(6-71 a)
G(oo)u
(6-71 b)
HI
;, AND IDENTIFICATlON
267
+B implies
SV I(S)=V 2(S)
sV 2(S) = V 3(S) = S2V I(s)
and
Ui
:-j[uJ
bp
e p]
+<XIS
m-
+ ... +<xm)V1(s)=!/J(s)V(s)=I p
. _si-Ilp
V ,(s)- !/J(s)
and
(6-72)
Consider
cture of A, b it can be
,le. lt is however gener
irreducible one requires
This shows that Equation (6-71) is a realization of G(s). Because of the forms
of A and B, it is easy to verify that the relization is controllable. lt is, however,
generally not observable.
We note that the realization in (6-71) is a generalization ofthe controllable
form realization for a scalar transfer function shown in (6-26). lt is also possible
to find a generalization ofthe one in (6-38). Let
(6-69)
(6-70)
: }+[r}
ti I p
(6-73)
where H(i) are q AX P constant matrices. Let lf(s) be the monic least common
denominator of G(s) and of the forro shown in (6-69). Then similar to (6-30),
we can show
i = 1, 2, ...
(6-74)
(6-71.1
lp
R ] x+ (;(oo)u
(6-71b)
(6-75)
Fro~ (6-73) and (6-75), we may conclucle, similar to (6-34), that {A, B,
is a reaiization of G(s) in (6-73) if and only if E = H(O) and
i=O,I,2, ...
e,
E}
(6-76)
268
IRREDUCIBLE REAU
Hankel matrices
Iq
Oq
Oq
H(1)
H(2)
Oq
Iq
Oq
Oq
Oq
Oq
Iq
-C(m1q
-G(m_I1q
-G(m-2 I q
-G( 1 I q
x=
X+
y= [ I q
H(2)]
AB
= ~(3)
[
AZB
] X+ H(O)u
(6-77b)
[H(3)]
= ~(4)
~(m + 1)
(6-77a)
H(m -1)
H(m)
.. "' A'B =
H(m +2)
(6-78)
HU +m)
and, in general,
M~
Kq
-IXm-Il q
-IX m -
zl q
Ip
O
O
-IXm_Il p
N~ O
Ip
-IXm-zl p
Ip
-IXll p
c=
A=
18=
IE=
(6-79a)
-:m1q
lq
T~[::
1)]
HU +
H(i+2)
.nd
:. J
-IXll q
-IXml p
is a pm-dimensional re:
realization is a unit mat
realization however is g(
Now we shall use TI
an irreducible realizatio
existence of qm x qm anc
(6-79b)
, ANO IDENTIFICATION
269
Hankel matrices
[ H(l)
:J( 1)
T~ ~m
~(2)
H(m)
(6-77 a)
H(2)
H(3)
H(m)
H(m +1)
H(m +1)
Il(2m-l)
H(3)
H(4)
H(m
+1)1
H(m
+2)
H(m +2)
H(2m)
(6-80)
H(m -1)
H(m)
(6-77b)
[(O)u
0J
+
~U + 2)
HU
iB =
and
H(2)
T~ ~(3)
[
H(m + 1)
(6-81 )
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
verified that
(6-78)
T=MT=TN
(6-82)
and, in general,
H(i +m)
(6-83)
i=O, 1,2, ...
Note that the left-upper-corner element of MiT = TN i is HU + 1). Let Ik,l be a
k x 1(1 > k) constant matrix of the form Ik,l = [I k OJ, where I k is the unit matrix
of order k, and O is t he k x (1 - k) zero matrix. Then the comer element H(i + 1)
can be removed from MiT = TN i as
(6-84)
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
realization of G(s) in (6-73). Note that this realization is the one in (6-77) and is
observable but not necessarily controllable. Similarly, the dynamical equation
e = I q . qmT = [H(1)
H(2)
...
A=N
R=
I'p,pm
= [I p
O]
H(m)]
(6-85)
lE = IH(O)
A
(6-79a)
(6-79b)
I:
T=K [ O
0J'
O L
(6-86)
. where I: =diag P'l, Al, , .. , An}and Ai , i = 1,2, .' .. , n, are the positive square
roots of the eigenvalues of T*T, where theasterisk stands for the complex
270
IRREDUCIBLE REALl
Consider
A 2 =(Y+Tl
Kl~Ll
Kl~1/2~1/2Ll ~ YU
=(Y+Ml
=Y+MT
=Y+MT
=Y+MM
=Y+M 21
(6-87)
(6-88)
and
=~1/2LIL1*~-1/2
= 1"
(6-89)
Consider (6-84):
(6-90)
HU
Consider a q x p... proper rational matrix (;(s) expanded as (;(s) = L:::o H(i)s-i.
We form T and T as in (6-80) and (6-81), and factor T as T = YU, and pT = p y =
pU, where p denotes the rank, by using the singular value decomposition.
Then the {A, B, C, E} defined by
C In
=CAiB,
i,
This shows that (6-91)
The dimension of
Am-1B] and V'=[C'
T=V and'n=Tsmi
matrix, we conclude
controllable and obser'
(6-91 a )
(6-91b)
(6-91c)
Y*Y==
A realization whose e
property V*y = UU* i
speaking, the signal tral
state to the output are s
(see Problems 6-22 and
In the application, it
model for a given systeJ
to this problem. For e)
Define
T+~U+Y+
(6-92)
1q,qm MiTI'P..
Iq,qmMiTT +
= Iq,qm TNiT+'
= Iq,qm TT + TJ'~
= Iq,qm YUU+,
=
Theorem 6-4
A=Y+TU+
B = first p columns of U = UI'p,pm
e = first q rows of Y = Iq,qm y
E=H(O)
+ 1) =
(6-93)
AND IDENTlFICATION
Consider
A 2 =(Y+TU+)2 = (Y+MTU+f
=(Y+MTU+)(Y+MTU+)
= Y+MTT+TNU+
=Y+MTNU+
=Y+MMTU+
=Y+M 2TU+
qm x n matrix, and
271
[ using (6-82)J
[using (6-92) and (6-83)]
[using (6-93)J
[using (6-83)J
(6-88)
1 (2
Ai=Y+MiTU+
(6-89)
(6-90)
i = 1, 2,3, ...
Consider (6-84):
H(i
+ 1) = Iq,qmMTI'p,pm
= Iq,qmMiTT+Tl'p,pm
ling theorem.
= Iq,qm TNiT+Tl'p,pm
= Iq,qmTT+TNiT+Tl'p,pm
= 1 YUU+Y+MiTU+Y+YUI'
lS
G(s)
= I;':o H(i)s- .
~ '--v-' '-v-----''-v--'~
=CAiB,
II!
Ai
II!
i=O, 1,2, ...
[using (6-93)J
[using (6-83)]
[using (6-93)J
[using (6-87), (6-92), and (6-83)J
(6-91 a)
(6-91b)
(6-91 e)
(6-92)
(6-93)
272
where 1: = diag {A, A2 , . An} and Al ;:O:A 2 ;:o: .. ;:o: Ano If the system is to be
approximated by an m-dimensional model with m < n, then the reduced model
can be obtained from
T = K I diag{A, A2 ,
. ..
,Am}L
where K is the first m columns of K and L is the first m rows ofL. See Refe
rences S141 and S162.
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
G(s) = rank T. ronsequently, the degree of G(s) can also be computed from T.
Ifthe degree of Gis known a priori, then the m in (6-80) can be chosen to be the
least integer such that the rank ofT is n. Because ofT = VU, from the definitions
of controllability index /l and the observability index v, we may conclude that
the 1east integer to have pT = n is equal to
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.
Method 11. Row Searching method. In this subsection we shall search
'" to find an irreducible
the linear independent rows of a Hankel matrix of G(s)
realization. This method is a direct extension of the vector case discussed in
Section 6-4 and the reader is advised to review the section before proceeding.
Let (;(s) be a q x p proper rationl1 matrix. Let a be the degree of the least
common denominator of the th row of (;(s) and {3j be the degree of the least
common denominator of the jth column of (;(s). If their computations are
complicated, they may simply be chosen as the sums ofthe denominator degrees
of the ith row andjth column of (;(s). Define a = max {a, i = 1,2, ... , q} and
{3 =max{{3j,j = 1, 2, ... , p}. Similar to (6-59), we may form the Hankel matrix
H(a +1,{3)
Hda +1,{32)
Hp(a +1,{3p)1}a(no.Of
H. 22 .(a 2 + 1, {32)
:
where Hij(a + 1, {3j) is the Hankel matrix of gj(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
[:~~~
~(a+1)
T=
[
Cl
H(a+1',B)
~2(a+1,,B)
as in (6-65). Using th
q primary linear1y de
(6-67) can be readily
servability indices of ~
illustrate the procedu
Example 1
max{/l, v}
H q2 la q +i,{32)
IRREDUCIBLE REALI
::~;
:~)
H(a+2)
H(a+{3)
+1 1]
(6-95)
as in (6-64). We then search its linearly independent rows in order from top to
-2s 2 -3s-2
(s + 1)2
4s+5
---=4
s +1
and
-3s-5
s +2
It is clear that a = 4, a
ANO IDENTIFICATION
If the system is to be
en the reduced model
273
bottom. After the completion of the search, we then rearrange the rows of T
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= .
.
.
: :
(6-96)
[ :
H q1 (0:+1,{3) Hdo:+1,{3)
H qp(o:+l,{3) }v q
as in (6-65). Using the coefficients of linear combinations corresponding to the
q primary linearly dependent rows of T, an irreducible realization similar to
(6-67) can be readily obtained. Similar to (6-65), the Vi in (6-96) yield the ob
servability indices of any irreducible realization of G(s). We use an example to
illustrate the procedures.
Example 1
G(s) =
Pp)
J}
O"p
rows)
(6-94)
tled as in (6-35).
:n top to bottom.
[H i1 H i2 ...
In similar to the
~~) + 1)]
Now
Then
HpJ,
one in
4s+5
- - = 4 +S-1_ S-2 +S-3 -S-4
s +1
and
-3s-5
s +2
It is clear that 0:1 =4, 0: 2 =2, {31 =2 and {32 =3. We form
T=[H l l (O:I +1,{31)
H 21 (0:2 +1, {3)
.,--f:
-2
3
,
,
, 3: -4
5
5 -6
------- - - - : 1: -1
-1
1
1 -1
.:-2:
---
(6-97)
+ ...
1 ___ I
[(o: +{3)
S2
-3s-5
s +2
-2s 2 -3s-2
(s + 1)2
:-4:
(6-95)
+ 1)2
4s +5
s +1
(s
p~l}
, {3
O" 1 (no. of
, {3 p) } 0"2 linearly
independent
O
1
O
O
,
:0 ,, O O
, O
:0
O
...
,
, O
O
O
., ------- - - - - - - -
.; 1: -2
4
:-2
4 -8
4 -8
16
, ,
,1 ,
_~
1--
..
........__ ... _=
.....=.....=.. ==
274
IRREDUCIBLE REAL!
We note that hJO) are not used in forming T. The application ofthe row-search
ing algorithm to T yields
Next we use
G(s) =
A
O
O
ar
[-2 0J
+
4
-3
We hav~ IX =max{IX I ,
matrix T and apply t
:T
~~
42390
210
-8- - --=4---=- -17 - ~~ f9- ---O-- =--4- ---0- ---f
3
1
-2
-2
3
-4
O
1
O
1
O
O
}.,
3
O
O
O
1.
O
O
O
O
3
O
O
O
O
O
-------------------O
O
O
O
4 } ~2 = 1
O
O
O
O
O
O
O
O
O
O
~4
1
2
1
O
-1
-4
2
O
O
-4
8
O
O
-16
8
O
O
32 -16
(6-98)
Clearly we haye 0"1 =4 and 0"2 = 1. Hence the dimension of any irreducible
realization of G(s) is 5. Corresponding to the fifth and seventh row ofthe right
hand-side matrix of (6-98), we can compute, by using (F-l1), from the leftmost
matrix in (6-98):
and
kIT=[O
O 1 2
:I~: O
O O]T=O
k 2 T=[0
O 1
:2
()) O]T=O
'-v-J
1
1
~~:~.
O -1
~----~--~~--=--~--l-~
x=
Y=
-1
x+
-1 :-2
~ ~
\
__
] x
4 :'-1
-8
-8: 1 16
O: -6
O
16 : - 1 - l ?
1 __
". '
Corresponding to the 1
as shown, we can com]
kl = [O
k 2 = [O
+[-~ _ ~ ]
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)
and 9rz(s) and the first 0"2 Markov parameters of 921(S) and 922(S). The form of
is fixed and depends only on O" I and O" 2'
-2:
4--- - -; -- 5----<i
-1
5
li -~ .~--,j
_~~ ~
-L
---
O: -2
1: -1
1
O
O
O
O
O
O
O
2
2
O1((
[~IJt~[O,O
2
O O O
'-----v----'
VI
=3
.....= =..
=
..
_~~
_= _=.=_.. =--=
~
~=.=--~.
=.=
.. =--=
..=..=.= =._=._=
~
=.
275
ANO IDENTlFICATION
Next we use
0J +[1 0J
-3
1 1
G(s) =[-2
4
S-l
+[-2-1 -21J s
+[3
-z
0J
4 s
-3
+[-4
-1
JS-4+ ...
-8
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
I~ ~ ~ }(J1
1
2
1
O -1
-4
2
O
O
-4
8
O
O
-16
8
O
O
32 -16
=4
O
lO
lO
O
O
O
O
O
O
10
O
O
O
4} ~z = 1
O
O
l - -- - ---- -
10
O
1
O
1
O -3
O
3
O
4
O -7
O -5
O 15
1
O
1
O
-1
O
1
O
O
O
O
1
O
O
O
1
O
O
1
O
1
O -2
1
3
O
3
1 -5
O
7
4
--------------------------5
O
8
O -11 -4
1
O -1
O
1
O
(6-98)
--------------------------
-1
O
O
O
O
O
O~
O
O
O
O
O
O <
O
O
O
O
O
O
-------------------------O
O
O
O
O
O
=0
--------------------------
=0
1}
- Q
0-'
1
O u
O
O
O
1
O
'I:
(n
O]
O]
(6-99)
[kz~lJt~[O
- ,O
O 1 0.) 0;2 1
O 0J
O O O O: 2 3: f: O O
'-v------"
V1
1.
=3
vz=2
~~,.~==.
.~~,
.. u,,_,.
276
... ~:::":,:::,=----::..:'..:.,:",:,,
:.:.:::::.':.:.._._J'~:"_":::':'~:"':'::=--'--....:'.:::._~:'::"":'::":'':':::::::-'':':':'''''''':~::'::::'':';;~:,:":,:_::::_:::
..:::.~:.::::::::~~..::-::::~~;-~~=----=---:::=:.::::::.::.:::.:..:::....::.o=~~:::.....:..::.::;::_:
.:.::~_
:-2:
:__ ~_: -4
-4
5
: O:
-4
: 1:
5, :0:
-6
O
O
O
O
O
with
~:t-~{----~-~:---~--:~~---~ =[~J
'-1'
1
1 -1
- 1
1
1 -1
1
-1 :'-2' 4 -8
1: 4 -8
16
- 1 :, - 8
16 - 32
1: 16 - 32 64
1 ----1
'
H(s)= O
x=
_~~.-
IRREDUCIBJ
:" -i -: - 2
..
O
O
O
1
O: O
1: O
O
O -1 :-2
y=[~
~Jx+[-~ _~Ju
The vth and (v +v2)th rows of A are taken from k and k 2 with the signs re
versed. The matrix B consists of the first V Markov parameters (excluding
hij(O)) of gIleS) and gds) and the first V2 Markov parameters of g2(S) and
g22(S). The form of e is fixed and the positions of 1 depend on V and V2' The
observability indices of this realization are V and V2 (Verify them.)
+ G(s)
where G(s) is strictly proper. Since G( (0) gi ves the direct transmission part of a
realization, we may exclude it in the following discussion. Using the procedure
discussed in Appendix G, we can factor G(s) as
G(s)=N(s)D~(s)
where N(s) and D(s) are q x p and p x p polynomial matrices and are right
coprime. Furthermore, D(s) is column reduced and 0cN(s) < 0cD(s) = JJ.,
for i= 1, 2, ... , p.. Define
.
n=JJ. +JJ.2
+ ...
O
u
11
-1-2
00:01
O
O: -2 -3
-------
-----------~-------
O
O
O
1
O
-2
-1 x+
3
+JJ. p
277
That is, n is the sum of al! the column degrees of D(s). Let us write
D(s) = DhcH(s)
+ DlcL(s)
with
=[~J
H(s)
Sil!
O
O
Sllz
O
O
Sil
O
O
O
Slll -1
,
, o
,
, o
,
,
,
,
, O
,
, 1
_ _ _ _ _ _1 _ _ _ _ _
L(s)
,
,
,
:0
:0
,
,, :.
,
,
,
,
,
Sllz-1
O
------------
O
1
2
3
1
O u
1
-2
Sllp
~----
,
I
(6-101 )
:0
_----
'O
....
------
:0
:0
, .
.,
The matrix H(s) is a p x p diagonal matrix with S to the powers of column degrees
on the diagonal. The matrix L(s) is an n x p matrix; if ..ti is zero, then the cor
responding column is a zero column. The matrices D hc and D le are, respectively,
p x p and p x n constant matrices. Because of bciN(s) < bciD(s), N(s) can be
expressed as
2
4
(6-102)
lrime
and
] proper rational matrix
" as
n write G(s)
D(s)V(s) = u(s)
YCs) = N(s)v(s)
(6-103)
(6-104)
The substitution of(6100) and (6-102) into (6-103) and (6-104) yields
and
(6-105)
(6-106)
~ct
transmission part of a
m. Using the procedure
+ Dh;, 10(S)
(6-107)
Let us define
x(s) = L(s)v(s)
6This step does not appear in the scalar case because D(s)
realization.
I
'. (6-108)
i~
normalized to be
~~nic
beFare' ;he
- ... - ...
-.--------
~
278
~.-
IRREDl
V1
x12
j1
V<" -1)
------
Vz
Xl!"
XZ1
=
Xz!'
.(6-109)
V~2 -1)
-------
-------
where
have
X p1
Vp
XP!'p
V(!'p-1)
p
Vi
V\k) = dkVi/dt.
(6-110)
a1
X:"
Xl!',]
r- !"] {b1!"]
~ ~'" x(t)
xP!'p
b~" u(t)
ap!'p
(6-111 )
bp!,p
where al!'; denotes the th row of Dh;, 1 D re , and bl!'i denotes the ith row of Dh;, 1.
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
readily obtain the following dynamical equation.
O
1 O
O ]
O ,'
O O O
,
,
,
x x x
x :x
...
(-a,",) X
- - - - - - - - - - - - -: -0- - I - -0- - - e - - .-. :
:O
wX
-
-. -
l'
I
.
...
-
x:
-
-4 -
:X
x x ... x
------------
O:
X:
-
'o
...
-
O
(-a1t)
l ...
I
1"
...
x:
-------------1---------------- I
o
xpp.
,
,
,
X : ..x
O
( -ap"p)
,
,
,
,
,
,
,
x,
:X
1- -
O
b,",
x(t)
u(t)
-al!",
O
X
-
X
-
-: -
X
-
_~1"_'
:0- -1- - 0- - - - -O
---
,'o
,:
:0
,
-0 -
I
I
10
O O
,x x x
_3S2
A
\J(S) =
(S-;
(s-~
(s+I)3
=
-6s-2
(S +1)3
[(S
+ 1)03(S -
2)
(s
bp "
(6-112a)
+1
279
ANO IDENTlFICATlON
XllI100'0-1~SI
xII
_~
'00
(6-109)
(/11 integrators)
-~
(/12 integrators)
[}Y
(/1 p integrators)
'00-'"
1.
PI
~rom
this definition we
(6-110)
.. , Jl-1
l
b
(t)
I'!]
+ b~l'l
u(t)
(6-111)
bpl'p
..
-1
o
x
.. ,
----------
o
x
.~
x(t)
o(t)
_3S 2 -6S-2
.~
.:-----
",(s) =
100
O O
x x
+ 1)'
(s
+ 1)3
=.[(.1
s' -3.1-1
(s -
2)(.1
+ 1)'
2)(.1
(s -
+ 1)2
(6-112a)
(s
+ 1)'(.1
(s -2)(.1
2)
O ..
(s
+ 1)'(.1 -
2)
J-'[(
.
.1(.1 - 2)
s3.-:3s-1
s +1 ]
s(s + 1)
(6-113)
"' .. "
-------'....:.:c."
' . . ~ . .""c"'
. ~ - - -
280
"" _
--.:=
"=:.""=.,,=::. . .
IRRED
~
G(s) =
"
[-3S
~Jr +3<t'+1
1
O
-6s-2
s
-1
s-2
-l
-1
(sI - A)
This equation implie~
BD(s)
From (6-112), wesee
are zero rows. Furth
(I~= 1 /1m)th row of
O0J
s O +[1O
O 1
C(sI -A)-I]
or equivalently,
-3
is a real ization of G(
showing that the A, ]
N(s) = [
-6
-2
-x
-1
O
-x
(si A)L(s) =
Note that, beca use of /13 =0, the last colun;m of L(s) is a zero column.
compute
We
-x
-ll
O
O
3 -2
O
(6-114a)
y= [
-2
-6
-3
\6-114b)
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)
-x
L-x -x
, AND IDENTIFICATION
denominator of each
cients af this fraction,
-1
s-2
O
281
or eguivalently,
-IJ-l
(sI - A) - lB
-3
1
(6-115)
(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
BDhc(H{s)
!) as
+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
s
o ...
O
-x
.,
o:
-\
~X
,
,
o:
-\
-x
,
, ,
, ,
, ,
-1 :
S-XI
-x
-x
- - - - - ~ - - - - - - - - - - - - - - - -:- - - ~ - - =-.- f
-XI
o"
, O
,
(sI - A)L(s) =
-x
-
-x
-
.. - - -
-x .,.
-
- -
- -
-. - -
-x
-x
-x
- - - - -1- - - - -
I
,
I
I
- -
- -
-1: :
:-x
s-X:
- -
- -
- - -
- - - - - - - - - - - - - - - - - - - - - -:- - - - - - - - - - - - - - - - - -:
3
O
-ll
-x
-x
-x
...
: : o
1
I
I
I
,
...
-x'
3
O -2
:- -5- -
-x
-x ...
- - - -
:
-x' -x
...
-x
0 , ,,
'
is a zera column. We
-x
l-X
------- 0-: :- -- - - - - - - - - -
_.- -
-x
I \
0,
-x
- -
I-X
- - - -
s-x
<;
S/JI-I:
O:
--o-~---(-:
(6-114a)
o ,
I
,
I
,
,
I
: S,J-l:
\6-114b)
D hc 1 D lc do oot appear in
o '
I
I
(6-117)
I
I
282
IRREDUC
eL
From this equation, we can readily verify that all rows, except the
m= 1 J.!m)th
row, for i = 1, 2, ... , p, of (sI - A)L(s) are zero rows. Because the (~ = 1 J.!m)th
row of A is equal to the ith row of - D hc 1 D le, as can be seen from (6-111), the
(~=1 J.!m)th row of (sI-A)L(s)=sL(s)-AL(s) is equal to the ith row of
H(s) + D,;/ DlcL(s). This establishes the equality in (6-116). Hence the dy
namical equation in (6-112) is a realization of (;(s) = N(s)D- 1 (s).
We discuss now the relationship between det(sI -A) and det D(s). From
(sI - A)-IB = L(s)D-l(S) in (6-115), we have
det
A))B = det
where "det" stands for the determant and "Adj" the adjoint. Since (Adj (sI - A))B
and L(s) Adj D(s) are polynomial matrices and since det (sI - A) and det D(s) are
polynomial of degree n, we must have
Proof
We first show that if
observable. Under t:
det D(s) such that G(~
find a realization of G
(6-112) is not irreducn
ever, (6-112) is known
shalI establish this st;
D(s) and N(s) are not
such that
and
Because of (6-116), (6-1
(6-118)
Realization of N(s)D-l(S) where D(s) and N(s) are not right coprime
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
reduced and (\N(s) < (\D(s). Now we shall show that the dynamical equation
in (6-112) is controllable whether D(s) and N(s) ;lre right coprime or not. This
can be established by employing the condition rank [B AB ... An - 1 B] =
n or the left coprimeness of (sI - A) and B. For this problem, it is easier to
employ the latter. From (6-115), we have
or
Since L(A) consists of a
is the n x 1 vector L(A)a
does not have a full ran
Now we show that
right coprime. Under 1
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
rank
[~~;)J =p
for every s in iC
(6-120)
Hence D(s) and L(s) are right coprime (Theorem G-S). From the realization
procedure, we have dim A = deg det(sI - A) = deg det D(s). Hence Corollary
G-S implies that sI - A and B are left coprime or rank [sI - A B] = n, for
every s in iC. ConsequentIy, {A, B} is controIlable (Theorem 5-7). Up to this
point, the coprimeness of D(s)and N(s) has not been used.
Theorem 6-5
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.
, AND IDENTlFICATlON
283
Proof
We first show that if O(s) and N(s) are not right coprime, then (6-112) is not
observable. Under the premise, we can find a 6(s) with deg det 6(s) < deg
det D(s) such that (;(s) = (s)D - 1 (s) = N(S)O-I(S). Usin~ (s)D -1(S), we can
find a realization of (;(s) with a dimension smaller than that of (6-112). Hence
(6-112) is not irreducible (controllable and observable) (Theorem 5-19). How
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,
such that
O(A)J
rank [ N(A) <p
Since(Adj(sl-A))B
(6-118)
N~~=O
ud
(AI-A)L(A)lX=O
or
(6-112) is applicable
s long as D(s) is column
the dynamical equation
t coprime or no1. This
: AB .,. An- 1 BJ =
problem, it is easier to
[Al
CL(A)lX=O
~AJ L(A)lX=O
(6-121 )
[AI~AJe=o
6-119)
(6-122)
.t matrix of order p as a
(6-120)
x 1 vector
el!
and
such that
O(A)lX=O
(6-123)
[AI-A
The composite matrix [H-A
Theorem 2-5 implies
rank [AI-A
L(A)J
-BJ [ D(A) =0
(6-124)
-BJ +nullitY[Ai-A
-BJ=p
-BJ=n +p
_.~_.
284
.-
'.~-
...
~-
'.
"
....
._a'.
IRREDU(
[U-A
[OeJ
[L(A)
0(,,1.)
IX
IX
Proof
Controllability indice
similarity transformat
singular matrix. eOll
properties of (;(s). He
theorem. Ir we rearra
where D hc is nonsingl
becomes
This establishes (6-123). The substitution of e = L(A)ex into (6-122) yields (6-121)
which implies, by using (6-116), (6-102), and (6-112b),
O(A)J ex = O
[ N(A)
Hence O(s) and N(s) are not coprime. This completes the proof ofthis theorem.
Q.E.D.
Theorem 6-5 implies that the realization in (6-112) is irreducible ifand only
if (;(s) = N(S)O-l(S) is a right-coprime fraction. Ir (;(s) = N(s)D-1(s) is a right
coprime fraction, then we have, according to Definition 6-1 ,
deg (;(s)=deg det O(s)
Consequently, the realization in (6-112) is irreducible if and only if deg (;(s) =
dim A. Since all irreducible realizations of (;(s) are equivalent (Theorem 5-20),
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
form realization discussed in (6-26). Hence we shall call (6-112) a multivariable
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 O be cnturrn reducr;c1 th,,;'-,
we will obtain a different multivariable controllable-form realization. Hence,
unlike the scalar case, the controllable-form realization in the multivariable
case is not unique. Consequently, the adjective "canonical" is oot used.
Column degrees and controllability indices
Before proceeding, we give a relationship between column degrees and control
lability indices.
Theorem 6-6
The set of the controllability in dices of any irreducible realization of a strictly
proper rational matrix G(s) is equal to the set of the column degrees of O(s) in
any column-reduced right coprime fraction (;(s) = N(S)O-l(S).
8'=
"
O .
O
l'
Furthermore, N'(s)and [
is an irreducible realizati
multivariable obser~able
Of course, an observa
(6-125). Let the row d~gI
let
=.. ==e-=..
=====o===.=,
....=.-c=
....=---=
.._-=_
...=....._=
.... =_=_
....=.. =.-~. .=....
_ _.=-_
..
, AND IDENTIFlCATION
:V(}.)
~xists
D'(A)]' form a
a p x 1 vector a.
c,~ ..
285
Proof
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
theorem. If we rearrange the input vector u(t) by
o(t) = Dhcu(t)
'8'=
O"-01:0"'00:
O
O O: O
O 1:
:O"'OOj
: O .. , O O
I
I
'
I
1
1:
0'''00:0'''00::0
'---v-----'
realization bf a strictly
)Iumn degrees of D(s) in
)0-1(S).
J1.1
J1.1
01
'---v---'
J1.p
where the prime denotes the transpose_ Using this and the A in (6-112), we can
readily show that the controllability indices are J1.l> J1.1' ... , J1. p, which are the
Q.E.D.
column degrees of D(s). This establishes the theorem.
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
extensively in the next section.
Consider a strictly proper q x p rational matrix (;(s). We factor it as
C(s) = 0-1(s)N(s)
(6-125)
where the polynomial matrices O(s) and N(s) are left coprime and D(s) is row
reduced. We take the transpose of C(s):
C'(s) == N'(s)(D- 1(s))' = N'(s)(D'(s))-1
(6-126)
The rOVlS 01 D(s) beco me the columns o [}J'(s). tlence D'(s) 1S
C01UIIlll (;clUC>:L
Furthermore, N'(s) and D'(s) are right coprime. Hence an irreducible realization
z=Az +Bv
w=Cz
y=B'x
===-
286
c __
__
....
__ ..
,.
'
"
_ .... _.
__._
.... __
"'_'
------------
We write
C(sI
(6-128)
which implies
where
SVI
_
H(s) =
1 s ...
SVI - 1 :
------O-------:- i --S- - -.
_
L(s) =
or
O
O
O
~ ~
: :
;;2:'--1 - -: :- - - - - - - - - - - - - -
vq =-
(6-129)
N(s) = L(s)Nc
and
x:
O
,
O x',
O x:
x
x
.x',
x ,
*6-7
Polynomial
(6-130)
Note that D h ,. is a q x q matrix and is nonsingular, for D(s) is row reduced. The
matrix L(s) is a q x n matrix. DI,. and DrD';;' 1 are n x q matrices. Thus a
realization of (6-125) is given by
O O
1 O
O 1
. ,
,
1 x:
O O
x
x
x
x,
x(t) =
x:OO"'Ox'
,
x , 1 O
O x
,
,
,
x , O 1
O x ,
,
,
,
x , O O
y= [O O
,
,
,
,
x ,
OC h
:
1,
x ,
--------------1-_-------------.,.
,
x ,
,
x ,
x ,,
x ,,
,
x ,
x ,
,
,
Oc
2 V2
:- 0- -O------6 -
: 1 O
O 1
or
6s +1
-1
~x
O x
O x
O O
1 x
+ Nteu(t)
: ... : O O ...
I
x(t)
,
,
x ,,
O O ...
3
2
(6-131a)
(6-131b)
wherecy is the ith column of D,;;.1 and the (L~~ 1 vm)th column ofthe matrix A
is the ith column of - DlrD,;;.l. Since (6-131) is a realization of (6-125), we have
2h
JI
,----rml'-----.-_ ~
~r~
In
;, AND IDENTIFICATlON
(6-128)
287
(6-132)
(6-133)
(6-134)
which implies
or
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.
6-129)
*6-7
(6-130)
x
x
x
Consider the network shown in Figure 6-8. If we assign the capacitor voltages
and inductor currents as state variables and use the procedure discussed in
Section 3-4, then a dynamical equation can be developed to describe the
network. There are, however, other methods available to develop mathe
matical equations to describe the network. For example, if we assign the loop
currents l(t) and z{t) as shown and use the loop analysis, then we can obtain
(2s
1
3s
~2(S)=U(s)
(1 1
A
--l(S)+
-+-+2s+1 )A~2(S)=0
3s
3s s
x
x
x
x
+ 3s) ~l(S) -
x(t)
[6s _+/
or
6s2;31s+4JU:~:~J=[~JU(S)
)I(s)=[O
-----------
O x
O
1
O x
O x
1 x
+ Nrcu(t)
O .. , O
6-131a)
Cqvq]x
--r
~2h
II
2h
(6-131b)
31
In
2s]l~:(s;j +O'u(s)
~-
(6-135a)
(6-135b)
288
(6-137)
From this, we can see that the state-variable equation is a special case of
(6-135) ifwe identify R(s) = C, P(s) = sI - A, Q(s) = B and W(s) = E. The fraction
G(s) = N,.(s)D,:-I(S) is a special case of (6-135} if we identify R(s) = N,.(s), P(s) =
DAs), Q(s) = 1, and W(s) =0. The fraction G(s) = D-I(S)N(s) is aIso a special
case of (6-135) if we identify R(s) = 1, P(s) = D1(s), Q(s) = N(s), and W(s) =0.
Hence this description is the most general one.
Given a polynomial matrix description (6-135), its transfer function matrix
G(s) from u to y can be readily computed by using (6-137). Note that the com
puted G(s) is not necessarily proper. In the following we discuss a method of
finding a dynamical equation for (6-135). We consider first (6-135a). If Pis)
is not row reduced, we can find an m x m unimodular matrix M(s) such that
M(s)P(s) is row reduced (Theorem G-ll). The premultiplication of M(s) on
both sides of (6-135) yields,
M(s)P(s)~(s) = M(s)Q(s)u(s)
(6-138)
(6-139)
which implies
where P,.~ M(s)P(s) and Q,.(s)~ M(s)Q(s). Note that deg det Pr(s) = deg det
Pis). In general P,:-l(S)Q,.(s) is not strictly proper. If so, we carry out the divi
sion:
Qr(s) = P,.(s)Y(s) +Q,.(s)
(6-140)
(6-141)
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
and
or
= R(s
is a q x p poIynomiaI m.
dynamical equation
=O
, AND IDENTlFICATlON
m x p, q x m. and q x p
y(s) is the q xl output
called the pseudostate.
matrix P(s) is assumed
called the polynomial
this set of equations
(6-136a)
(6-136b)
135a) we have~(s)=
ields
l to
(6-137)
)Q,.(s)u(s)
(6-140)
2').
Qr(s) = L(s)B
Co(sI - A)- 1 = p r- 1(S)L(s)
Pr(S)C O = L(s)(sI - A)
and
or
R(s)C o = X(s)(sI - A) +C
(6-147)
(6-14S)
+ R{s)Y{s) + W{s)
(6-149)
y{t) = Cx{t)
+ BiJI(t)
+ E(p)u(t)
(6-150a)
(6-15Ob)
(6-146)
This is notexactly in the state variable form becausethe matrix R(s)Co(sI - A)-1 B
may not be strictly proper. Ir so, we can apply Corollary G-12 to R(s)C o to
obtain
The substitution of
(6-141 )
(6-143)
(6-144 )
(6-145)
(6-139)
289
290
B J=[ Pr(s)
E(s)
-R(s)
Q,,(S)J[C a
W(s) O
L(s)
-0-- - ------=.X(;)
-Y(S)J
Ip
r- I,,:
-CaJ=[I"J
si A
O
(6-154 )
-U 11 (S)
P(s)
Uds)
M-l(S)L(s)
~--~-~'
1
[ 0m
unimodular
where the leftmost matrix is a unimodular matrix. Using this matrix, we expand
(6-153) as
=L:T!~~)-
(6-153)
~12(S)J[
We combine
(6-151)
(6-152)
(sI
if and only if
o
P(s)
which implies that
[(si - A)
(6-155)
This is obtained by augmenting the first three matrices in (6-153) as shown 1,;
(6-155) and then searching the fourth matrix in (6-155) to complete the equality.
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
-U 11 (S)
[
1m
UdS)(SI-A)J=[-Ull(S)
Co
1m
1" - U 11 (S)C O
Ca
=[1"
0
- U 11 (S)J[0
1111
if and only ir [l
1111
I"J
Co
(6-156)
Since. the two matrices afteT the last equality are unimodular, so is the first
natrix ih (6-15"6). Consequently, the rightmost matrix in (6-155) is also uni
modular.
As the last step, we shall replace P,.(s) and Q/s) in (6-155) by P(s) and Q(s).
Theorem 6-7
, ANO IDENTlFICATlON
Recall that P,(s) = M(s)P(s) and Q,.(s) = M(s)Q(s), where M(s) is unimodular.
The premultiplication ol' diag (In> M- 1 (s), I q }, a unimodular matrix, to (6-155)
yields
-Uld S)
[
(6-151)
(6-152)
_~!~)
UI2(S)
: 0J[lm:
O ]
~~l_(~)~J~)_:_~_ --:-(~i~)---B-
-X(s): Iq
O:
-C
E(s)
(S)BJ
We combine
= O : p(s)
\[~o -~p(S)J
(6-153)
:J
-U l1 (S)
P(s)
U12(s) J[l m
M- 1(s)L(st O
(6-158)
~ A) ~]
l) : U 12(S)8]
, - Y(s)
._J
(sI
O
P(s)
-~t]
O 0J
[II!O P(s)
O Q(s)
OJ S(s)
B
(si -A)
where S(s) denotes the rightmost matrix in (6-157). Since the leftmost and right
most matrices in (6-158) are unimodular, we conclude that
[;
(6-157)
wherc the leftmost and rightmost matrices remain to be unimodular. With this
eq uation, we are ready to esta blish (6-151) and (6-152).
From the first two block rows of(6-157), we can readily obtain
[
(6-154)
291
(6-155 )
Ip
~s in (6-153) as shown in
In
for every s in C
or equivalentl)', [A, B} is controllable if and only if 1"(5) and Q(s) are left coprime
Similarly, ...ve can eSlablish thal (A, C; is observable ir and ony il' P(s) and R(s)
are right coprime.
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.
1,:,1 (S)J[:'"
(6-156)
Theorem 6-7
Consider the polynomial matrix description in (6-135) with n = deg det P(s).
.Then an n-dimensional dynamical equation realization ol' (6-135) is irreducible
or, equivalently, control1ableand observable ifimd nly if (6-135) is irreducible.
I
292
This theorem follows directly from Definition 6-2 and the statements in
(6-151) and (6-152). To conclude this section, we discuss the situation where
(6-135) is not irreducible. If PIs) and Q(s) are not left coprime, then there exists
a m x m polynomial matrix H(s) with deg det H(s) > 1 such that
PIs) = H(s)P(s)
and
Q(s) = H(sQ(s)
pepe),) Q(A)]
= pH(A)[P(),)
Q(),)] < m
*6-8
[
where ~ is called the
PSI
Ir we identify R(s) = e
.identify R(s) = N,.(s), P(s:)
meludes (6-159) and (6-1,
be used to describe any (
Consider the system 1
and
(6-160)
(6-161 )
S(s)
P(s)~(s) = Q(s)u(s)
y(s) = R(s)~(s) +W(s)u(s)
Ge(s)=R,,(S)P,,-1
Equation (6-159) is more general than the one studiedin earlier sections by
including derivatives ofu in y. By so doing, its transfer matrix can be extended
to include improper case. Given a state-variable equation, the transfer function
description can be obtained as G(s)=C(sI-A)-lB+(E+E1s+). lf
G(s) is factored as G(s) = N,.(S)O,:-l(S), then a polynomial matrix description
can be obtained as Q(s) = 1, P(s) = 0,.(05), R(s) = N,.(s), and W(s) = O. Conversely.
q x p polynomial matrices.
ever, because of the discussi
to make them egual and re,
this reg uirement, the subseg
S186, and S187.
, AND lDENTlF\CATlN
namical equation is de
then this n-dimensional
1 is decomposed into the
the uncontrollable part,
det H(s)or, equiva\ent\y,
uncontro\\ab\e mode of
:d the input-decoup\ing
en the roots of the deter
cal\ed thc unobservab\e
tput-decoupling zeros in
, W(s)} does not have any
ion in this section is not
F or a more detai\ed
1.
S(S)=[
(s)N,(s)(s)
(6-160)
(6-161 )
O ]
(6-162 )
-y(s)
(6-163 )
(6-164)
:O
P(s) : Q(s)
J~
P e(s)
-Re(s)
Qe(S)]
W(s)
(6-165)
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
'~Let :F'(.{s) = det :2\:;)
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,
their entire dynamical behaviors are, as will be shown latee equivalent.
Consider two system matrices
S(s)
:)
Q(S)]
W(s)
P(s)
-R(s)
and
1
-Il
Q(s)
~(S)J
W(st -(s) =
Se(s) =
293
=[
p(s)
-R(s)
Q(s) ]
W(s)
i = 1,2
(6-166)
where PieS), Q(s), R(s), and Wi(s) are respectively mi x mi, m x p, q x mi, and
q x p polynomial matrices. Note that mi is not necessarily equal to 1112' How
ever, because ofthe discussion in (6-165), we may extend either mi or m20r both
to make them equal and require m =m!'=m2 2':deg det PieS), i=l, 2. Without
this requirement, the su bsequent discussion may not hold. See References S125,
S186, and S187.
294
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
ves) 0J[ PIes) Q(S)][V(S)
[ X(s) I q -R(s) W(s) O
Y(S)]=[ Pz(s)
Ip
-Rz(s)
Qz(s) J (6-167)
Wz(s)
I
O]
[ves)
I
O
q '
Y(S)]
Ip
and their inverses. Using this property, it can be readily verified that this equiv
alence relation has the symmetry property [if S(s)~ Sz(s), then Sz(S) ~ S(s)],
the rel1exitivity property [S (s) ~ S(s)] and the transitivity property [if S(s) ~
Sz(s) and Sz(S)~S3(S), then S(S)~S3(S), where ~ denotes strict system equiv
alence].
The reason of using extended SeS) in Definition 6-3 can be seen from the
realization developed in Section 6-7, especially Equations (6-153) and (6-155).
Without extending Pi (s), the matrices Ves) and Ves) in (6-167) may not be square
and, certainly, may not be unimodular. Using the unimodularity property, we
can write (6-167) as
. [ves) 0J[ PIes)
X(s) 1 -R(s)
Q(S)]=[ Pz(s)
W(s)
-Rz(s)
1,
where k is a nonzero c
We compute
Gz(s)=R 2 P-Q
+W
z
Z
= (R - XP dV(U:
which, by simple manip
Gz(s) = R
This completes the proo
In order to establish
we use (6-162) to writeS.
1
[-:
[-:
and
0J[ P ~
[U
X 1 -R: "
which can be written as , b)
r1lT
L~
~"""1 r
"
;Jl-;_
Theorem 6-8
Two system matrices. which are strictly system equivalent have the same
transfer matrix and det p (s) = k det P z(s), where k is a nonzero constant.
Proof
We multiply (6-167) out toyie1d
VPV
[ -(R -XP,)V
VPY +VQ,
] [ P z
295
Hence we have
'ctly system equivalent if
atrices u\s) and V(s) and
1 that
(6-169)
where k is a nonzero constant. Hence we have deg det P 1(s)=deg det P 2(s).
We compute
G2(s)=R 1 Pi l Ql +W 1
=(Rl-XPl)V(UPIVrlU(PIY +Ql)+(XP 1 -R I )Y +(XQl +W I )
which, by simple manipulation, becomes
GAs) = R1 Pi 1Q2 + W 2 =
R I PIQl + W 1 =
G1(s)
Q.E.D.
(6-170)
(6-171 )
and
(s) 11[Y\s)
2(S)J O
Y(sq (6-168)
lp
are eq uivalent
ix' their pseudostates are
n~mical-equation realiza
Section 6-7' are zero-input
realization is controllable
'operties wi\l be established
Note that SI(S) and S2(S) are two different descriptions of the same system,
hence u and y in (6-170) and (6-171) are the same. However, their pseudostates
~1 and ;2 are different in general. From (6-168), we have
U
[X
0J[
1 -R
PI
I
Q1
W1
J[ -otJ = ['-R
P 2 Q2
W2
2
ol [
[U0J[ -y0J =[-yJ=
X
]32
-R 2
Q2JrlV~I--:,'YM---'J'
W2
-u
(6-172)
(6-173)
(6-174)
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
Si, i = 1,2, are strictly system equivalent, there is no difference in their dynamical
behaviors.
Consider two system matrices S(s) and S2(S). lf they are realized by using
the procedure developed in Section 6-7, then the realizations {A, B. C, E}
have the property
det (sI - A;) = k det p(s)
Theorem 6-9
Consider (6-167). Because of the presence of the two zero matrices, the first
block row of(6-167) can be written as
Q(s)]
[v(S)
y(S)]
O
I
=
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.
U(S)[p(s)
Proof
[P 2(s)
Q2(S)]
(6-175 )
Wf
P(sI
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.
to yield
and
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\;;) =
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
Q.E.D.
proved.
In Section 4-3, we introduced the concept of equivalent dynamical equa
tions. Now we shall show that strict system equivalence is a generalization of
this concept.
jp o and .
P(sI -A)
Now we shalI show tl
unimodular, U- (s) is also
l
Theorem 6-10
c,
Two dynamical equations {A, B, C, E} and {A,R, E} are equivalent ifand only
if their system matrices are strictly system equivalent.
:, AND IDENTIFICATION
... ~
.. _'-"
", ......
297
Proof
,2
e,
O][sl-A
o I
-C
B-![P-
O]=[sl-}
-C
e,
U(S) O
][SI-C
-A EBJ [SI-C-}
[X(s)
I
=
~][V(S)
Y(S)]
I
(6-177)
e,
for sorne unimodular U(s) and Ves). Since {A, B, C, E} and {A, B,
E} have
the same transfer matrix (Theorem 6-8), we have E =:E. Equation (6-177)
implies
U(s)(sl - A) = (sI - A)V(s)
(6-178)
Q2(S)]
(6-179)
where Pis a constant matrix. The substituton of (6cI79) into (6-178) yields
(6-176)
. s in
e,
s)]
(6-180)
and
PA=AP
for sorne polynomial matrix U 1 (s) and sorne constant matrix P 1 - The multi
plication of (6~179) and (6-182) yields
U(s)U- 1 (s) =(sl - A)U(s)(sl - A)U 1 (s) +(sl - A)U(s)P 1
(6-181)
298
or
D 2(s) = D 1(s)T(s) an
(sl-AT 1(I-PP)=(s)(sl-A)U 1(s)+(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
equal to zero. Hence we have I-PP 1 =0 or PP 1 =1. Consequently, P is
nonsingular and is qualified as an equivalence transformation.
To complete the proof ofthis theorem, we must show 8 = PE and C = CP- 1
From (6-177), we have
where (S)D-1(S)is ~
of
D,
[
We conclude that {
SimilarIy we can she
alent.
What remains to
left-coprime fraction
D I- 1N/. Tben there
-N,
Its left-hand side is a polynomial matrix, whereas its right-hand side is a rational
matrix. This is possible only if B - PB = O. Hence we conclude 8 = PB.
Similarly, we can show C=cp- 1. This completes the proofofthis theorem.
Q.E.D.
System matrices which are strictly system equivalent have the same transfer
matrix. System matrices which have the same transfer matrix, however, are not
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),
where N(s) and D(s) are right coprime, but N(s) and D(s) are not right coprime.
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
strictly system equivalent (Theorem 6-8).
Although system matrices {p(s), Q(s), Ri(s), W(s)}, which have the same
transfer matrix, are generally not strictly system equivalent, they become
strictly system equivalent if the system matrices are irreducible: that is, {F';(s),
Q(s)} are left coprime and {P(s), Ri(s)} are right coprime. We establish first a
special case.
12 : O] ~I__
U U
NI 'O O'
[-1DI 0:1
O:
11
_1- _
N,.
Since the two matrice
the first matrix in (6- j
matrix in (6-187). (
{D,., 1, N,., O} and {D h 1
proof of this theorem.
Theorem 6-11
All coprime fractions of G(s) are strictly system equivalent where G(s) is a
rational matrix, not necessarily proper.
Proof
.7The proor orTheorem G-1
to improper G(s).
8Th!: procedure is similar to
299
AND lDENTIFICATION
+ PUl(s)
(6-183)
B= PBand C=Cp-l
0J = [ Dz(s) OIJ
I
-Nz(s)
(6-184)
which implies the strict system equivalence of {DI' 1, NI' O} and {D z, 1, Nz, O}.
Now if (;(s) is not strictly proper, then we may also factor (;(s) as
(;(s) = N(s)O -l(S)
where N(S)O-l(S) is strictly proper and coprime; N(s) = N(s) + E(s)O(s). Because
of
(6-185)
Ull
[ DI
(6-186)
where the leftmost matrix is a unimodular matrix (Problem G-ll). Now using
DN,. = NI O,., we form the identity
[ =--C---o---:-i
L~
-N
Lo:
-1
o] -----O---:-I-
(6-187)
The second and third matrices from the left in (6-187) are extended system
matrices of {D.., 1, N,., O} and {DI, NI' 1, O}. The leftmost matrix in (6-187) is
unimodular because its left-upper-corner matrix is unimodular fol1owing from
(6-186). We show in the fol1owing that the rightmost matrix in (6-187) is also
unimodular. From (6-186), we have - U 11 N, + U 12D.. = 1 and
UIll
UlzD"J=[Ull
N,.
I
UllNI+I]=[U ll
N..
I
C[I N..]
0_ O I
(6-188)
Since the two matrices after the last equality in (6-188) are unimodular, so is
the first matrix in (6-188). This implies the unimodularity of the rightmost
matrix in (6-187). Consequently, we conc1ude from Definition 6-3 that
{D,., 1, N,., O} and {DI> NI> 1, O} are strictly system equivalent. This completes the
proof of this theorem.
. Q.E. O.
'The proorof Theorem G-13 does nol use lhe properness of (;(5); hence lhe lheorem: is applicable
lo improper (;(5).
.
.
8The procedure is similar lo lhe one from(6-154) lo (6-156).
(s)D1l(s) = Nz(s)Dzl(s).
__ .
..
300
~'~_J~'._"
.. ,". _ _ ','._'
~._-"---~-
Theorem 6-12
polynomial matrix 1
These results are di]
out the foHowing tra
Transfer matrix
Dynamical equation
Polynomial matrix
description
System matrix
(6-189)
Q(s) ]
W(s)
As in the continuous-l
parameters of g(z). In
be computed from di g
response of the system
of an impulse as an in
of g(t) are again impra(
time case are not real
entirely different. If v
initiaHy relaxed linear
data at the output term
2, . . .. Hence the real
case can be considered.
mathematical descripti
ment at the input and (
corrupted by noises.
requires the concepts o
~-----
9Consequently melhods ar
M k =I~ tkg(r)dt, k =0, l,2
,D IDENTIFICATION
Continuous-time systems
Transfer matrix
Dynamical equation
Polynomial matrix
description
System matrix
= Cx(t) +Eu(t)
G(s)=E +C(sI-A)-IB
P(s);(s) = Q(s)u(s)
y(s) = R(s );(s) + W(s)u(s)
P(s)
S(s) = [
-R(s)
Discrete-time systems
G(z) = N,(z)D,:-l(Z)
D-1(Z)N1(z)
x(k+ l)=Ax(k)+Bu(k)
y(k) = Cx(k) + Eu(k)
G(z)=E +C(zI-A)-IB
P(z);(z) = Q(z)u(z)
Y(z) = R(z);(z)
Q(s) ]
W(s)
S(z) = [
P(z)
-R(z)
+W(z)u(z)
Q(z) ]
W(z)
(6-189)
Hence the discussion of these problems will not be repeated for the discrete
time case. Note that a1l block diagrams in this chapter are also applicable to
the discrete-time case if every integrator is replaced by a unit-delay element or,
equivalently, S-1 is replaced by Z-I.
There is, however, one problem in the discrete-time case which deserves
special discussion. Consider a sampled transfer function, g(z), of a discrete
time system expanded as
and
denotes the
actor. Conversely, if
fy (6-189), then they
ption, any one of the
.nformation, to study
~ices
G(s) = N,.(s)D,:-l(S)
= D-I(S)N(s)
x(t) = Ax(t) + Bu(t)
y(t)
301
,and q x p polynomial
.perations:
>y a nonzero constant.
;.
+ ...
:ems
ltion methods for con
We also introduced
I
~
9Consequently methods are developed lo find realizations by lIsing the moments defined by
M k = S~ tkg(t) dt./< =0.1,2, . (see References 528 and 5146).
.
.,
302
--'c_--
' _.'.'
'=.'..c='.""'.,,,,'===.=--=.. =.._='=.
--
_.
text. Hence we assume in this text that all data are free of noise and call the
problem the deterministic identificatan.
Consider a single-variable linear time-invariant discrete-time system with
transfer function 9(Z). Ir we apply the impulse sequence {u(O) = 1, u(i) = 0,
= 1, 2, 3, ... }, then the zero-state response yields the Markov parameters
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
have a degree bounded by N,1 o then we need only 2N + 1 of the hU), i = 0, 1, ... ,
2N. From the Hankel matrix, we can readily obtain a dynamical-equation
description of the system as shown in (6-35) to (6-38). Ir the transfer function
description is desired, we have
f3 oza +(J 1 za -1
gA()
Z =
. za +(XI Za 1 +
+ ... + f3 a
. . . + (Xa
IDENTIF.
and
where Di> Ni' i = 0, 1, .
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
h(O)
h(1)
~1(<T)
:h(<T-l)
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
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 i the system. is at OU disposal. I a system
is in continuous operation, its Markov parameters cannot be measured.
lfno bOllnd 01 the degree of a system is available. then it is theoretically impossible to identify the
system. For example, ifwe have [h(O) =0. h(i) = 1, i = 1,2..... 200) and ifthe degree ofthe system
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
from a finite seqllence of {hU), i =0,1,2, ... , N}. The problem of finding a transfer fllnction to
match a finite sequence of Markov parameters is caBed the partial realizatioll problem (see Refe
rences 68, S 126 and S239).
II f ~ system is not BI BO stable (see Chapter 8), then h(i) will approach infinity or l'emain oscil
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
to rest by resetting. In thebry, thereazati~n ~~ identification is applicable no maller the system
is stable or not so long as the data are available.
10
h(O)
The substitution of I
where M = [-No Do
matrix Si -Y, 00) has (
columns. This equatio!
Hence, given an arbitra
to the search of M to ro
rows of nontrivial solu
Sv( -Y, (0). Since it is,
as possible, we search,
Sv( - Y, (0). Hence the
dependent rows of Sv('
similar to the coprime I
--------~---~---_._._.
..~-~~._._ ...-._..
__ ......
"
303
.ND IDENTIFICATION
system with
ce {u(O) = 1, u(i)=O,
Markov parameters
')} yields the transfer
an infinite number of
: system is known to
)f the hU), i = O, 1, .... ,
dynamical-equation
the transfer function
G(z) = D-l(z)N(z)
(6-190)
(6-191)
(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)
(6-195)
- N(z)u(z) + D(z)Y(z) =0
st be initially relaxed.
we apply an impulse
)ther input terminals,
, ... , q; i =0,1,2, ...}.
nite time; in practice,
L], we then repeat the
:e the measurement of
disposal. If a system
t be measured.
O
O
[-No Do -NI DI .. , -N v DvJ O
u(O)
iLJ,(n
v,
O
O
u(O)
y(O)
u(O)
y(O)
u(l)
y(l)
u(l)
y(l)
u(2)
y(2)
u( v - 1) u( v) u( v + 1) ...
y(v -1) y(v) y(v
+ 1)
MSA-v,oo)=O
...:
(6-196)
.1..
304
u(k+1)
y(k + 1)
u(l)]
y(l)]
(6-197)
(6-198)
and define
(k, 1)
y(k, 1)
---------(k + 1, 1+1)
S,,(k, I)~ y(k + 1, I + 1)
----------
(6-199)
.'
(k +IX, I +IX)
form an space of an
sistently exciting. It r
ability for the input to
Consider So(k, 00).
search the linearly der
Because is persistentI'
in y(k, 00 ). Let ro be tb
and q-r o is the num!:
apply the row-searchin!
~ependent rows in y(k
Yi(k, 00) is Iinearly depe
dependent in SI(k, a:.
[u(k): (k + 1,00)] and.
we have 1'1 ~ro. We co
There are ()( + 1 block rows in S". Each block row consists of p rows of and
q rows of y, hence S,,(k, l) is of order (p +q)(1X + 1) x (1- k + 1). We call the
rows formed from u rows and the rows formed from y y rows.
There are infinitely many equations in (6-196). lt is much more than neces
sary to solve for M. In fact, M can often be solved from MSy(O, l) =0, with I
chosen so that SvC0, 1) has more columns than rows. The range of data {k, I}
used in the identification is not critical ir. the system is initially relaxed. How
ever, it becomes critical, as will be discussed later, ifthe nonzero initial conditions
are to be identified as well.
(k,oo)
Y(k, (0)
(k +-( 00)'
Sy(k, 00) =
C"k-+~-~
i: ~
Y(k +v -1,
(k +-v: 00)
y(k
with
Definition 6-4
+v,
00)
assumption of persisten1
y rows in S y(k, 00) is clcar
n~number ,
=(q-ro)+
. Let ~\, i = 1, 2, ... , q, I
mdependent Yi in Sy(k, 00).
... , y(k +v i -1, (0) are lir
and Yi(k +1,00), I=v v.'
Then Yi is said to have ;~d~),
and
_~~
__.__..
..
,~._~~._~_~.~c.~==~~==========
E, AND IDENTIFICATION
,2 ...
305
(6-199)
(k, 00)
+ 1, 00)
-----------
} Y[
y(k
Sv(k, 00) =
(k
(6-200)
+v -1,00)
y(k+v-1,00)
}Y v -l
with
O'::;;YO'::;;Yl'::;;'"
} Yv =q
::::;Yv-I'::;;Yv=q.
Note that if
Yv=q,
then
Yj=q
for aH
j '2: v. Note also that aH the rows in S.(k, 00) are linearly independent by the
~ of its dependence on k, 1,
=(q-'o)+(q-t'}+ +(q-t'v_)=vq-
I.
Yi
i=O
Let Yi' i = 1, ;2, ... , q, be the ith row of y. Let Vi be the number of linearly
independent Yi in Sv(k, 00). By this, we mean that the rows y(k, 00), y{k + 1, 00),
... , y(k + Vi - 1, 00) are linearly independent of their previous rows in Sv(k, 00),
and Yi(k + 1, 00), 1= Vi, Vi + 1, ... , V- 1 are linearly dependent in S.(k, 00).
Then Yi issaid to have index Vi' Cleuly, we have Vi'::;; V, V= max{ Vi' i = 1,2, ... , q},
and
n = VI +V 2
+ ...
+Vq
(6-202)
--~-
---
, ,.. ,._"-~
,~~""-~
306
",._..,
,_.
-.~-'''-~'~.''.'_.~
~ }r
(6-203)
and
d? 1
[
Because aH elemen ts or
zeros, the row degree of
D(z). Hence D -l(z)N(;
D(z) are left coprime. S
R(z) with deg det R(z) >
1\
(6-204a)
(6-204b)
An example will be g:
to iIIustrate the identifi,
matrix, in the coprime fi
equation description can
in Section 6-6,
Furthermore, D(z) and N(z) are left coprime, deg (;(z) = deg det D(z) and D(z) is
row reduced, column reduced, and actually in the polynomial echelon form
Proof13
In order not to be overwhelmed by notations, we use an example to prove the
13The proof is identical to the proof of Theorem G-14.
d~ 1 + d!31 z + d~
Theorem 6-13
and
+ d~ lZ + di
D(z)= dg 1 +dLz
Persistently exciting
variable system can b~ a
307
.ND IDENTIFlCATION
(6-205 )
(6-203)
.gonal, as discuss ed in
wehavevl =2,v 2 =0,
,index 2.
)endent on its previous
example, the y2 in the
1; the Y2 in the second
~d block-row of (6-203)
rimary dependent rowS
nary dependent y row
(6-207)
Because aH elements bn the right-hand side of d~2' di l' and d%3 in (6-205) are
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
N(z) = R(z)N(z)
D(z) = R(z)D(z)
(6-208)
This implies deg det D(z) > deg det D(z). However, this is not possible because
the n = LVi computed in the algorithm is unique and smaHest possible. There
fore we conclude that N(z) and D(z) are left coprime, and the degree of the trans
Vi.
fer matrix is equal to deg det D(z) =
That D(z) is in the polynomial echelon form foHows from the definition given
Q.E. D.
in Appendix G. This completes the proof of this theorem.
308
SAk, (0). In practice, we use the matrix Sv(k, l) for a finite l. Clearly, 1must be
larger than (v + 1)(P + q) to ensure that there are more columns than rows. In
actual computation, once y(k + Vi, l) is Iinearly dependent in SV(k, 1), then y
may be deleted in forming SV+ j, j = 1, 2, ....
If u is persistently exciting. then the Ni and Di computed by using the row
searching algorithm have the properties that D(z) is row reduced and, con
sequently, nonsingular and D - 1 (z)N(z) is proper. If is not persistently exciting,
then the computed D-I(z)N(z) may not be proper. For example, consider
g(.:) = 1/(.:: + 1). If we apply u(z) = 1 +.:: - 1, then the output )";(z) is equal to z - l.
We form
(O, (0)]
SI
ht~~-:~
y( 1, (0)
[1
=
1 O O O ...]
~-----~--6- -:-:~
(6-209)
1: 1 O]SI =0
1 -z
and
(6-210)
We see that this g(z) is not proper and is erroneous. Fortunately, this problem
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
grouping all u rows at the upper half of S. and all y rows at the lower half. For
this example, the equation becomes
[ -No
1 O O O O ...
Clearly (l, 00) is linearly dependent on its previous rows: hence u(z) is not
persistently exciting. If we solve (6-209) by using the first linearly dependent
row ofS, then the solution is
[-1
lDENTIFI
DI][~~-:I~
[-1
and
The conditions in (6-21.
ample, one of u(z) may
(6-213). See also Proble
case, we may choose an
If the chosen or given i
appear as linear depenc
automatically whether o
Nonzero initial cond
(6-211 )
y(l,oolJ
U(k' (0)
_ u(k+l,oo)
(k, (0) .
[
y(z) = C(z]
which can be written as
y(z)
where R(z) is a q x 1 poly
to the corresponding ro",
We rewrite(6~217) as
(6-212)
[ -R
, AND lDENTlFICATlON
.]
(6-209)
has a ful1 row rank. We give sorne sufficient condition for U(k, 00) to have a
full row rank. The condition depends on the value of k. If k = 1, then U(l, 00)
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
from the results in Section 6-4,
bu(z)::?:p(v +1)
i = 1,2, ... , p
b/(z)::?: \' + 1
and
(6-210)
equation of (6-196) by
and
b[u(Z)/ZV+l]::?:v+1
~p(v
+ 1)
i=1,2, ... ,p
=O
(6-214a)
(6-214b)
The conditions in (6-214) are less stringent than the ones in (6-213). For ex
ample, one of u(z) may have degree Oand satisfies (6-214), but wil1 not satisfy
(6-213). See also Problem 6-26. In practice, we do not have v a priori. In this
case, we may choose an upper bound v* ~ v and use v* in (6-213) or (6-214).
Ir the chosen or given input sequence is not persistently exciting, then will
appear as linear dependent rows in Sao Therefore the procedure will check
automatical1y whether or not the input sequence is persistently exciting.
Nonzero nitial conditions. In this subsection we study the identification
of systems which are not necessarily initia\ly relaxed. Consider the control1able
and observable discrete-time equation
(6-213a)
(6-213b)
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)
to be linearly independent are
b[U(Z)/ZV+I]
309
(6-215a)
(6-215b)
(6-211)
(6-216)
(6-217)
;: persistently exciting is
where R(z) is a q x 1 polynomial matrix with row degree smaller than or equal
to the corresponding row degree of O(z). Clearly, R(z) is dependent on x(O).
We rewrite (6-217) as
(6-212)
(6-218)
310
1
O
O ... O
u(O)
O ... O
y(O)
----------------...
O
1
O
... u(O)
u(l)
O
[ -Ro -No D o: ... :. -R v -N v D,.]
y(1)
O .. . y(O)
O
u(l)
y(l)
--------O
u(2)
y(2)
----------------
----------
----------------
----------
1
u(O)
y(O)
u(v
y(v
MSi - v, (0) =
+ 1) ...
+ 1) .. ,
IDENTIFlC
(6-219)
Given a linear time-invariant system and given {u(n), y(n), n =0,1,2, ...}, the
transfer matrix of the system obtained from Sv(O, (0) by assuming zero initial
condition and the one obtained from Sv( - v, (0) without the assumpqn of zero
initial conditions are the same.
Proof
Consider (6-196) and (6-219) with (6-196) modified as MSiO, (0) =0. First
we note that, by deleting the zero rows, the matrix Sv(l, (0) reduces to Sv(l, (0);
An irreducible realization
+1
x(k
y(k
O
05
O
3.5
1
05
-0--0--=-1-
1
-3
-6
O
-8
- 1
-2
O
-1
-3
-6
-4
-9
1
O
-2
1
O
O
1
O
O
-1
UrO) u(1)
?'~ 1)
-
O
O
u(1) u(2)
y( 1) y(2)
---------y(O)
------
______
-1) u(v)
u(v +1) .. ,
- 1) y( v)
y( v + 1) ...
(-v, (0)=0
311
:, AND IDENTlFlCATION
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
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
MS)-v, 00)=0. Bence the D(z) and N(z) computed from MSv(O, 00)=0,
and the ones computed from MS v( - v, (0) = O will differ at most by a nonsingu
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.
Example 1
z~ 1]
(6-219)
(6-220)
2z
z-1
An irreducible realization can be found as
x(k
+ 1) =[~
lJX(k)+[
[l
y(k) =
1
-0.5
(6-221 )
1JU(k)
1
~J x(k) + [~ ~] iJ(k)
3.5]', we can obtain the following input-output
10
1I
-~---
u(k)
U1(k)
O
O
O
O
.r I (k)
0.5
3.5
.\,(/d
"
1
O
O
O
O
O
0.5
3.5
()
(:
-1
1
O
1
3
')
O
4
-1
O
5
-
O
O
O
O
2.5
55
7
25
1';,':
1
1i
Now we shall try to identify the system from the input-output sequence
(u(n), y(n)}. First we assume that the system is initially relaxed and apply the
row-searching algorithm to S2(0, 11) as
1
O
_1
-6
-3
-8
I
-2
-1
O
O
-3 -4
-1
_1
-6
-2
_1
-9
O
O
-1
-2
312
o
O
0.5
:5 ;
-1
1
0.5
6
O
O
3.5
3
** - - --.:: -----f
O
3.5
3
:})
O
O
O
O
3
6
(~(~
O
3
3.5
Gt:
8
6
6
- -- -0- - - - - 0- ---6.. -- -1 -
O
3.5
6
0.5
6
(1)
O
3
6
1
3
1
O
4
9
O
O
2.5
9
O
O
5.5
7
O
5
7
O
4
9
-1
O
5
7
O
2.5
9
O
5.5
7
1
2.5
11
Do : -NI
DIJ=[
-t
.1
--34 ',,
O
O
-1
-2
2 ,
2
2
-1 :-1
_.1
1
O
x
x
x
x
O
O
~J
Hence we have
N(Z)=L!}
+1
D(z)= [ Z_2
;}+1]
:~}J
(6-222)
1
O
_1
1
-3
2
-6 -8
O
O
O -1
O
O
-3 -4
-6
-9
1
-2
O
O
O
-1
-2
O
O
O
_1. '
4
_1
1 .
.0
O' ,
1
2
-5
1
1
O
1
1
O
-2
O
O
0.5
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.
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
formula in (F-11) as
[ -No
O
O
0-1
O
3.5
O.
* O
5
3
6
* .T"--- -------6
,_J
O
O
0.5
5
_,
0:-11
O '-'
1
3.5
0.5
3
6
3.:
6
=[
Hence we have
6-10
Concluding
CONCLUDING REMARKS
AND IDENTIFICATION
o
o
o
o
2.5
5.5
9
7
---------
5
x
x
x
x
O
5.5
1
2.5
11
f K, which correspond
o
O
O
O
** o
-_ ..
.' "
' ... "
O
O
0.5
5
:})
O
O
O
O
O
O
O
O
O
O
O
O -1
O
O
O
1 -1
O
O
O
O
1
O
O
O (f:
O
O
O
O
0.5
3.5
0.5
3.5 el:
3
3
4
5
2.5
5.5
5
3
6
6
6
6
8
9
7
(9.~
7
_-------------------------------------------------
O
o
o
o
o
o
o
O
o
o
o
-,
O (- 11
1
O
O
O
1
1
O
O
O
o -'1 :}l~ o o 1 O O O O 1
3.5
0.5
3.5
3
3
3
4
5
2.5
5.5
2.5
3
6
6
6
6
8
9
7
9
7
11
eD
..
-1
-2
1
O
x
x
x
x
x
x
x
x
O
O
O
=
:_l
I
[ O'
,
l2
2,
-1 : -1
=[
-4 : -Z ,
-1':-5:-1
Hence we have
N(z)
313
O:,
-2:
1
O
~J
z_l
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)
(6-222)
6-10
d. In this case we must
rithm yields
Concluding Remarks
,,,.,
.~.~~.'
.-y
-~
_ - - _ .. ,.
"
~._,.,
, . '._.,_.".",~ . ."_.
_ , . ~_ . ' .
.'_'
'
."
"
- - - - ~ - ~_------~--'--~--~----,------~--_
314
~ _
.............-
Qlj = R,
CONCLUDING REMARKS
AND IDENTIFICATION
gularization of a matrix
:rations). In the second
f a matrix by either row
ires twice the computa
Illclude that the Hankel
etaHed comparison, see
)e implemented by using
3]5
introduce errors due to discretization. This is ou tside the scope of this text and
will not be discussed.
In this chapter, we also introduced the polynomial matrix description and
established its relationships with dynaroical equations and transfer matrices.
We showed that if the three descriptions have the same transfer matrix and are
irreducible, then they are aH strictly system equivalent. In thiscase, any one of
them can be used in the analysis and design without loss of any essential informa
tion.
The degree of a proper rational matrix is introduced in this chapter. The
degree can be computed by using any of the following methods:
1. Compute the least common denominator, L\(s), of all minors of (;(s). Then
we have deg (;(s) = deg L\(s).
2. Compute the Markov parameters of (;(s) and form the Hankel matrix T
shown in (6-80). Then we have deg (;(s) = rank T. The rank of T can be
computed by using the singular value decomposition.
3. Find a right fraction G(s) = (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
coefficient matrices of D(s) and N(s) to form the matrix Sk shown in (G-67).
We then search the linear independent rows ofS k in order from top to bottom.
Then we have
deg G(s) = total number of linear independent N roWS in Soo.
Similar results can be stated for a left fraction of G(s).
4. Consider the realizaton of (;(s) shown in Equation (6-71). The realization
QIO=O,
QIj=R m- j + 1
j=I,2, ... ,m
QiO =0, Qij=Q(i-I)(j-l) -O:m-j+1Q(i-l)rn
i =2, 3, ... , m~ j = 1, 2, ... , m
Using Qj, the observability matrix of (6-71) can be computed as
[QH
V~. ~21
LQml
Q12
Q22
QI~l
~2m
.
Qm2
QmJ
5. Find a realization of (;(s) and reduce it, by using the method discussed in
Section 5-8, to an irreducible one. Then we have deg (;(s) = dim A.
These methods can be readily programmed on a digital computer. A com
parison of these methods in terms of computational efficiency and numerical
stability seems to be unavailable at present.
To conclude this chapter, we give a different derivation of the observable
realization in Equation(6-131). lt wil\ be derived from three maps developed
in algebraic systeffi theory. We discuss the discrete-time case. Cons"ider the
q x p strictly proper ratianal matrix G(z) = D-1(z)N(z), where D(z) and N(z)
316
1 z
O O
L(z) =
::
..
[
00
z v, - 1
O
O O ...
1 z .. .
O
Z "2 - 1
O O
O O
:::
...
O
00
1z
J,]
(6-223)
(6-225)
It is a linear operator which maps (lRq[ zJ, IR) into itself. Because the ith row
degree of 0D(h'(z)) is at most V- 1, ihe range space of OD(h(z)) is (IK D, IR).
In algebraic system theory, a realization can be expressed by the following
three maps:
B D : IRP----* IK D ; u t-+N(z)u
A D : IK D----* IK D; x t-+OD(ZX)
C D : IK D ----* IRq; x----*(D- 1(z)X)_1
(6-226)
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
OD(H(z)) =
O.
(6-224)
as a basis. Its columns will be denoted by qij, i = 1, 2, ... , q; j = 1,2, ... , Vi'
Let f(z) be a q x 1 rational vector. Define the operator O as
O(f(z)) = strictly proper part of f(z)
If i = 1, j = 1, then D - 1(.
(6-227 )
"ll'~:)'
(s+3j2
Then we have
H(z) = D(z)D,;;.1 - [(z)DI.D,;;.1
Consider
b.
s +3
~-
s +2
s+
s +1
s +4
1)'
(s +2)
(s+1)(s+2)
_._-_
----,-------------_.
---_...
.. -_._---._..- ..
PROBLEMS
AND IDENTIFICATION
o
1
z Vq -
(6-223)
tor TI as
z)
(6-224 )
(6-226)
317
= q12 = L(s) O
O
6-1
-128) here as
(6-227)
'"ll' ~ :)'
(s+W
bt ;1)'
(s +2)
s +3
s +2
s +1
s +4
~:51
Id:~d2)1
(s
--~----
318
+3l
+ 1
+1
Find the dynamical-equation description of the block diagram with state variables
chosen as shown in Figure P6-2.
6-2
6-8
+7. 1(t),
where pi ~ d/dti.
Figure P6-2
Find irreducible ce
matrices
6-9
6-3
S4
S2
b. SS
_ S4
+1
a. 4s4 + 2s3
+ 2s + 1
S2
-s + 1
+ S3 _
S2
+s- 1
b. [
Are these realizations irreducible? Find block diagrams for analog computer simulations
of these transfer functions.
6-10
6-4
+2s +2
2s +3
(5 +W(s +2)
s(s
tions
a.
(S
S2
+ 1
b. - - "
(s +2)3
S2
c.
S2
+1
+2s +2
6-11
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?
(s + 1) (s + 2)(s + 3)
Figure P6-5
I
I
PROBLEMS
AND IDENTIFICATION
319
6-7
6-8
-~
+ f3z(t))u(t)
6-9
'unctions
a.
+2s +2
+3
b. [ (s+1)2(s+2)
6-10
+ 2s +2 ]
s(s+V
S2
S+2
ls:
s +1
~ :3J
5
+1
s +2
6-12
I
1
l.
320
6-13
Z~3J
Z+2
+1
[
Z
6-14
+1
+1
+2
6-19
Consider the q x p
Consider
~Jx +[~J
u
b
KoH(v+i)= -KH(
x =[)"
O)"
Ql
where f is defined as in (6
cal equation
~J
-Xb
_KVK
-),E
where
6-15
O
-AX
b=[~J
~xJ
[K
or
where the overbar denotes the complex conjugate. Verify that by using the transformation
x=Qx, where
_ [
=
i~
cl
=[ -2 Re (Xbcd
2 Re (bc)J
.
-K,,_K
x=
:
[ -KzK
-KIK
Y=[
(J
is a realization of G(s). Sh
controllable.
A
O
O
x=
O
O
O
1
A
O
O
O
O
O
1
A
O
O
O
O
O
O
X
O
O
O
O
O
1
X
O
O
O
O
x+
O
1
[
b
bz
b3
El
Ez
E3
6-20
O 2:
[O O'
where , b, and c are defined in Problem 6-14 and I z is the unit matrix of order 2. [Hint:
Change the order ofstate variables from [Xl X2 X3 X4 Xs X6J' to [XI X4 Xz Xs X3 X6J'
and then apply the equivalence transformation x = Qx, where Q = diag (Q, Qz, Q3)J
6-16
+ l)y
+(p +l)yz
(p - l)u
y= [
l'
is a controllable realization
PROBLEMS
AND IDENTIFICATION
321
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.
6-19
1t is assumed that there exist q x q constant matrices Ki, i =0,1,2, ... , v, such that
i= 1, 2, 3, ...
]x
IY using the transformation
or
[K
Kv
K2 .. ,
Ko O .,. OJT=O
O]
O
:
[Ko
~v
O
~o
+:
O
O
-K 2 K 1
O O
K3
K 4
-K 1 K l
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
u
O 2:01:10:1 0J
[ O O: O 2: O 3: O 1 = [K
,
,
,
K2 K3 KoJ
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
6]'
1
O
-Do
-DI
-D 2
y=[ No
NI
N2
hU
. ] ["
O
:
O
x+:
-~I'-l'
N.- 1 Jx
~,'
'----------_.~-~~~._-
322
6-22 Consider the equivalent time-invariant dynamical equations {A, B, q and {A, B, Cj
with A=PAP- I, B=PB, and c=cp- 1. Let wer and W Ol be the contr01lability and
observability grammians defined in Theorems 5-7 and 5-13. Show that
C1
and
\VOl
PWerp*
=(P-1)*WOlP- 1
where
q.
and
where R*R = I and L 2 =diag [/,i, ;.~, ... , l.;':. Define
H~L:R:RcLc
x=Ax +Bu
y=Cx
[t is assumed that a1l the eigenvalues of A have negative real parts and that the equation
is internally balanced, that is.
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
O [
2
3
-----[
u dk)
[
-[ O
u2(k)
[
O
O -[
k
[
Yl(k) -[
Y2(k) -2 -0.5
2
[
[ -0.5
i. W" = 1, W01 = L~
ii. W" = L~, Wo, = I
ii. Wer = Wo , = LH
6-24
6-27
PROBLEMS
AND IDENTlFICATlON
c:
Jt is assumed that all eigenvalues of A have negative real parts. Show that the reduced
system
y= ex
x= Ax +Bu,
where
323
A=A -A 12 Ai}A 21
B= B -AzA:;jB z
C=C -C2A221A21
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
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?
6-26
e
6-27
2 is chosen as
u\(k)
u2(k)
)'I(k)
)'2(k)
1
\
1
O
\
-1
O
O
1
O
O
O -\
-\
2 1
O
O
-2 -0.5
\ -0.5 -0.5 -0.5
\
O -\
O
O
O \ -\
\
O
\
\
\
\-0.5 -0.5
!O
11
12
13
14 ..
-2
O -\ -l
\
-\
\
O
1 -2
\
O O -\ -2
\ - \ -2.5 -4 -7
-----
First identify the system by disregarding the initial conditions and then identify the system
as well as lts initial conditions.
ld internally balal1ced on
-C*C
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)
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.
7
State Feedback
and State Estimators
7-1
Introduction
7-2
Canonical-F<
Single-variable cas
single-variable dynamic
om qualitative analyses
ntrol systems by using
, feedback systems. In
f dynamical equations:
all study their practical
1 them.
lical equation is con
\."-1B] has n linearly
:olumns or their linear
us canonical forms can
'til one: the controllable
rvable canonical forro
~ry useful in the designs
)r reference signal; the
ating signa1 so that the
~ reference signa\. If a
. the actual response of
Lis type of control is not
e system. If a control
: is called a feedback or
.ns all the essential in~
to be a function of the
'0] can be achieved. In
ir state feedback of the
325
7-2
Canonical~Form Dynamical
Equations
X=Ax +bu
y =cx +eu
(7-1 a)
(7-1 b)
X=Ax +bu
y=cx+eu
(7-2a)
(7-2b)
326
x=
eFE l
O
O
O
x=
O
U~[b
~[h
An-lb]
An-lh]
An-lb] =PU =Q-lU
-C/.."
(7-3)
(7-4)
or
(7-5a)
(7-5b)
Similarly, ifV and Vare the observability matrices of FE l and FE l , that is,
V=
cA
e
or
(7-6)
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
Proof
The dynamical equa1
of n x 1 column vecl
quently, the following
ql ~ A
f3n
qn D. b
qn-l ~ A
q"-2 D. A
eA~-l
y=[
is linearly independen1
from Figure 2-5 that i
th column ofthe new
to the basis {ql' q2' ..
Aql=(
)btained by introducing
;:mstant matrix. Then
eFE:
O
O
O
1
O
O
O
1
O
O
O
O
O
-IX"
O
-IX" _
O
-IX"_ Z
/31l
f3n- 1
/3" - Z
O
O
O
~tively,
(7-3)
O
O
O
x+
x=
y=[
-IX
z
/3z
-lX
327
(7-7a)
O
1
/3]x+eu
(7-7b)
where IX, IXZ' . . , ct" are the coefficients of the characteristic polynomial of A,
and the /3's are to be computed from FE. The dynamical equation (7-7) is
said to be in the controllable canonical formo The transfer function of FE is
(7-4 )
, FE are control1able,
g(s) =
z
" +e
s" +IXS" + ... +IX"_S +IX"
(7-8)
Proof
(7-5a)
(7-5b)
=VQ
(7-9)
(7-6)
19le-variable dynamical
~ either control1able or
ween them can be com
1 true for multivariable
equations. For multi
I holds. In this case U
we have U* =PUU*
1 (7-la)
.,1 ~ l
<llll_01X"J
be
Aq,~q, -,q.~[q,
-
q,
... q.]
lLJ
.................................................
~quivalence transforma
1
qz
...
q"] [ : ]
_-IX
328
Hence if we choose {q1' qz, ... ,qn} as a new basis of the state space, then A
and b have new representations of the form
o
A=
O
O
1
O
O
O
O
O
-eJ. n
-eJ. n - 1
-eJ.n-z
-eJ. z
O
O
O
b=
O
O
O
(7-10)
O
-eJ. 1
Theorem 7-2
If the n-dimensional l
FE 1 is observable, thel
One may wonder how we obtain the set of basis vectors given in (7-9).
This is derived in the following. Let D be the controllability matrix of the
controllable canonical-form dynamical equation, then
g(s) =
O O
O O
O
O
el
O O
O 1
en- 3
el
en-z
ez
en- 1
el
(7-12)
where
k-1
ek = -
eJ.i+ 1e k-i-1
k = 1, 2, ... , n - 1; ea = 1
i=O
The controllability matrix is nonsingular for any eJ. 1 , eJ.z, . .. ,eJ.1I" Therefore
the controllable canonical-form dynamical equation is always controllable,
o
b=
as one might expect. The inverse of has the following very simple form:
a n_ 1
an-Z
a,,-z
a ll -
=::
...
.
1]
al
(7-13 )
:: ~ A
1
.
"
O
O
[
1
O
O O
This can be directly verified by showing that D- I = 1. The dynamical equa
tions FE I and eFE I are related by the equivalence transformation x=QX;
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
puting U-l, we see that the columns of Q are indeed those given in (7-9);
that is,
Q=[ql qz
qll]=[b Ab
An-Ib]-l=UA
(7-14 )
- I
O
O
329
al
(7-10)
O
1
:uivalence transforma
)x or X= Qx, then the
oO.
oO'
equivalent controllable
es the first part of the
1 O
O
X= ~
O O
1
O
y=[O O
vectors given in (7-9).
Jllability matrix of the
f3 sn -
.
-al
(7-15a)
]x+eu
+ f3 ZSil - Z + ... + f3 n- s + f3 n +
(7-15b)
-1-'
+CX_lS
-~-(j~i1
1; ea = 1
[a"~,
an - z
P=
-~n-z X + ~n-z u
gA( s) =
(7-12 )
n ] .~n]
f3n-l
O -a ll -
a ll - z
(Xn-3
al
i]~~t,
O
O O cAn-
=AV
(7-16 )
330
e,
01
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
of an equation of this form is not known at present.
Example 1
x~ [~
y=[O
-1
(7-17a)
2
O
(7-17b)
A-1
[ -~
(Xi
-2 -lO]
=..1. -9..1.+2
A~~
The controllable
dynamical equations a
They are also useful iJ
computer. For exan
generally need n Z am]
and 2n amplifiers and
transformed into the c(
of (7-16), then the num
from n Z + 2n to 2n.
*Multivariable cas(
multivariable dynamic
U=[~1
: 1~]
where A, B, e, and E .
respectively. Let b bf
Ifthe dynamical eq
12
The controllable
also be obtained by
equation. The coeff
canonical-form eq uati
[-2 4 2l
-1 6 1J
3 2 1
[/33 /3z
/3IJ=cQ=[3
lJ
U = [b
b z ...
bp
331
j=[
-2
y= [
X~[!
y= [O
The transfer function of (7-17) is thus equal to
S2 +2s +3
g(s) = S3 -9s +2
A
gle-variable dynamical
(7-17a)
(7-17b)
namical
7)is
-9).,
equ~tions.
+2
natrix U s
X=Ax +Bu
y=Cx +Eu
(1-18a)
(7-18b)
2l
62 lJ1
(7-11), we have
observable canonical-
-------_...._._--_._-------_.
332
.-_..
..
_-~_._
...
_._---~--_._
..
_._._----_._--~---_
..
__ .__._._ ..._-._._.-
~--
Seheme 1
OOO"'Ox
100"'O x
0100 x '
We start with the vector b 1 and then proceed to Ab 1, AZb 1, up to AIi,-l b 1 until
the vector AIi'b 1 can be expressed as a linear combination of {b, ... , AIi-l bd.
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
as a linear combination of {b 1, ... , AIi -1 b 1, b z, ... ,Aliz -1 b z} (see Problem 7-5).
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
OOOlx
---J
(PI X {1.1)
,
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
start from b 1, b z, ... , b p and then Ab 1, Ah z, ... , Ab p, and then AZb 1, AZb z,
and so forth, until we obtain n linearly independent vectors. Note that if a
vector, say Ab z , is skipped because of linear dependence on the vectors
{b 1, b z, ... , b p, Ab 1}, then aH vectors of the form Akb z, for k ~ 1, can also be
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
1 O
AJlz-1b"b
{b 1"" AJl,-lb'b
1,2,''
2,""
p,
... ,
AJlp-1b}
P
where /11 + /1z + ... + /1p = n. Note that the main difference between this
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
(7-21) cannot be expressed as a linear combination of {b 1, ... , AJl-l bd; AJl b 1
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
FE or, equivalently, let x = Q-1x where
Ailz- 1b z
b 3 ...
O 1
(7-21 )
O O
x;
x:,
x,
,
,
x :
- :
:'(_ _0_
(1
.
.,
-------------x,
,
x,
Ail, -1 b 3J
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.
-----
---
----- - -
x:
x:
. ,
OOO"'Ox
0100x
I
A=
x
x
x;
100Ox'
b, up to Aiit - b until
nof{b, ... ,Aii'-b}.
:olumn of B alone. lf
Aii2 b z can be expressed
b z} (seeProblem 7-5).
so forth. Assume that
x
_
x
x
x
O O O
O 1 O
O O O
O O O
B=
:100"'Ox
:010"'Ox
1: : :
I
b4 "'bp
OOO
OO 1
OOO
OOO
333
OOOl.x:
x:J
------------J(j]. x j].)
: O O O ... O x :
:100"'Ox
:010"'Ox,
,
:OOO"'lx
l'
1 O O
O O O
O O O
:OOO"'lx
(j].3 x j].3)
OOO
(7-22)
where the X's denote possible nonzero elements and unfilled positions are all
zeros. This can be easily verified by observing that the ith column of A is the
representation of Aq with respect to the basis vectors [q qz '"
qn].
lfthe set of vectors in (7-21) is used as a basis, then the new matrices and B
will be of the form 3
O O
x ,,
x
.x',
1 O
O
x ,,
1 O
x
x;
O
O
O
,
O 1
x ,
x,
x
O
O
O
,
,
,
,
,
O
O
x'!
x'J ,
x
O O
O
-------
----------(111 X )J.l) x: O O
x : ,,
O 1
O
x
x :1 O
X: ,,
O O
x
O
,
x : O 1
x
O
O
O
, ,
. ,, .
__ - - - -
x:
Aii2- 1 b z
b 3
(7-21 )
-: ~ ;
::
B=
()J.z x )J.z)
-----------"'j-----------
Aii3 - 1 b 3 ]
X,
x,
x ,,
,
,
x ,
x,
,
x',
()J.p x )J.p)
:0--0------;
:1 O
x , :0 1
, ,.
, ,
X , :O O
I
O O
-----------
x
x
----------O O
1
O O
O
O O
O
O O
O
(7-23)
.,5,/125,'"
334
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 has three blocks in the diagonal, whereas the
matrix 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
7-3
Proof
State Feedback
In this section we study the effect of the linear state feedback u = r +Kx on a
linear time-invariant dynamical equation. The variable r denotes the reference
input and the K is the feedback gain matrix and is required to be real. In this
study, we implicitly assume that all state variables are available for the feedback.
Single-variable case. Consider the single-variable, linear time-invariant
dynamical equation
FE!:
x=Ax +bu
y=cx +eu
(7-24a)
x= (A + bk)x + br
y=(c+ek)x+er
(7-24b)
where X is the n x l state vector, u is the scalar input, y is the scalar output,
A is an n x n real constant matrix, b is an n x 1 real constant column vector,
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.
Then the dynamical equation of the state-feedback system shown in Figure 7-1 is
FE{:
A state f
(7-25 a )
(7-25b)
The state feedback dynamical equation FE{. in (7-25) is controllabie .ror any
1 x n real vector k if and only ifthe dynamical equation FE I in (7-24) is con
trollable.
Corollary 7-3
p[b
Ab
STATE FEEDBACK
335
Figure 7-1
Proof
First we show that the controllability of FE implies the controllability of
FE{. Let Xo and X be two arbitrary states in the state space L. By the con
trollability assurnption of FE , there exists an input u that will transfer X o to Xl
in a finite time. Now for the state-feedback dynamical equation, if we choose
r(t) = u(t) - kx(t), then the input r will transfer X o to Xl. Therefore we conclude
that F E{ is controllable.
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
other words, if FE is not controllable, neither is FE{.
Q.E.D.
,back u = f +Kx on a
denotes the reference
ed to be real. In this
lable for the feedback.
linear time-invariant
(7-24a)
(7-24b)
Corollary 7 - 3
The controllability o a multivariable linear Unle-,arying dynamical cCjuation.
is invariant under any state feedback of the form u(t) = r(t) +K(t)x(t).
(7-25a)
(7-25b)
p[b
Ab ...
An-lb] =p[b
(A+ bk)b
...
336
Example 1
x=G
n +[~Ju
y=[1
2Jx
If k is chosen as
u=r +[ -3
-IJx
FE{:
CFE{:
x=
O
-e
2Jx
y=[1
O
O
... +am
x = Px,
O
O
O
-a"_1
O
-a,,-z
O
-az
i~
l!
y=
CFE!:
-Cl.fT
...
f3z
filJx +eu
r]i+[!]U
-al
1
(7-27 a)
(7-27b)
Let A and b denote the matrices in (7-na), then A = PAP- 1, b = Pb. Because
. of the equivalence transformation, the state feedback becomes
(7-28)
where k ~kP-I. It is easy to see that the ser Of the eigenvalues of (A + bk) is
equal to the set of the eigenvalues of (A + bk). Let the characteristic poly
Algorithm
Given a controllable {A
1 x n real vector k such
its eigenvalues.
1.
2.
3.
4.
5.
6.
7.
--==c::===___
-~~~,,===~===-
~-_-_-~~.:__-=_-=-=---=-~----------------------------~--------
------.--_.-
--
---------------_ ..-.
..
----,_.~--------_._
337
STATE FEEDBACK
ation
If k is chosen as
(7-29)
x=
O
O
O
O
-Cl. n
1
O
O
O
- ii"_
O
O
O
x+
O
O
O
l'
-IX" - 2
-ii 2
-IX
1
(7-30a)
(7-24) is controllable,
real vector, the eigen
mt complex conjugate
(7-30b)
alence transformation
ontrollable canonical
o
o
o
-IX
eu
rO
'+l~
(7-27 a)
(7-27b)
envalues of (A + bk) is
le characteristic poly
Algorithm
Given a controllable {A, b} and a set of eigenvalues Xl> X2 , o.. , X... Find the
1 x n real vector k such that the matrix (A +bk) has the set {Xl' X2 , o.. , X"} as
its eigenvalues.
1.
2.
3.
4.
5.
6.
7.
_.
--_._---------_ .. _._.-------
-----------------.---
Find the characteristic polynomial of A: det (sl- A) = Sil +IX s" - + ... +IX...
Compute (s - X)(s -' X2 ) (s - Xn ) = s" +iX s" - -+: ., +ii...
Computek=[cx"-iX" IXn-1-iX n - o.. IXI-~il
Compute q.. _ i = Aqn-i+ + (Xq", for i = 1,2, . o. , (n-l), with q" = bo
Form Q = [q Q2--' o. qnl
Find P ~ Q - l.
k=kP.
I
._._----~-----------_._-----
338
Stabi Iization
~(s) =
or
case, a differen t rr
(Appendix F) will
single-variable case
(7-31 )
From this equation and using Problem 2-39, Equations (7-13) and (7-14), we
can readily establish k =kUA =kQ or k =kP.
Ir a dynamical equa
assigned by the intr
controllable, one m
shown in Theorem :
controllable, by a
transformed into
Example 2
X~l~
y = [1
O -1
~(s)
A+bkJ ~
l-2k
where k = [k l
k2
k3
1
k2
O
-2k 2
k 4 J can be computed as
S4
+(2k 4
k 2 )S3 +(2k 3
k1 -
5)S2
+ 3k 2 s + 3k l
The comparison of the coefficients of det (sI - A - bk) and those of ~(s) yields
k1=t
k 2 =lf
k 3 =s,
k 4 =li
lf
liJx
where
and the reduced ec
form of A, the set 01
of All and of A22 .
A22 is not affected
r +kx. Therefore;
other hand, all the
{A lb lid is control
can be arbitrarily l
required to be a re:
in pairs.
In the design of
eigenvalues (the eig(
(the eigenvalues wit
the aboye discussior
the equation canno
transformed into th,
eigenvalues of Azz h
a dynamical equatio
form dynamical eqt
eigenvalues are cont
of Statement 4 of T
if and only ifthe ma
that Al l and An in
Effect on the nume
Even for this simple {A, b}, direct computation of det (sI - A - bk) is very
complicated and is not suitable foi computer computation. Hence, for large
n, the use of the algorithm is much more convenient. In the multivariable
. _---_.
-_._._"~-~-----~,~._---.=.",.-===-====--=-::==-=----~'-':'~=-:::"-:::=:':'''....--::::'''~==-.==::::--==-----~
-':::'--:::.='::::-- -==--=--=-.-.:.......===...-=--===-:::..:=....::.::=.::~=='::::.====--------_
STATE FEEDBACK
339
x=Ax +bu
lmple 4 on page 185.
ily assigned.
Let the
10s 2 +10s +4
'S2
+ 3k 2 s + 3k 1
:z
where
-_[AO A
A-
11
12 ]
(7-32)
(7-33)
22
1, -2, -1 j.
340
gs
+(13" +eo:,,)
+0:"
(7-34 )
+(13" +eo:,,)
+"
(7-35 )
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.
This property can be used to explain why a state feedback may alter the
observability property of an equation. Ir sorne of the poles are shifted to
coincide with zeros of g(s), then the degree of g(s) in (7-35) is less than n, and
the dynamical equation in (7-25) is not irreducible. In this case, (7-25) must
be unobservable beca use it is controllable (Theorem 7-3).
Asymptotic tracking problem-nonzero set point
In this subsection we use state feedback to design a system so that its output
will track the step reference input r(t) =rd, for t ~O. Ir rd =0, the problem is
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
u(t) = pr(t) +kx(t)
(7-36)
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
;(s) = (e + ek)[(sl - A - bkt l X (O) + (si - A - bk)-1 bpr(s)]
res) = rJs.
+ epr(s)
which becomes, in t
y(t) = (e
for t ~ O. If all eige
approaches zero as
Ir {A, b} is controlla
by a proper choice
deeper inside the le
However, in this cas
saturation. Further
the choice of proper
by a compromise be
As discussed in .,
by state feedback; h
we may choose p = l
multivariable dynam
where A, B, e, and l
real matrices. In sta
(7-37)
In the following,
trollable, then the ei
proper choice of K.
(7-38)
Method 1
In this method we e
problem and then ap!
STATE FEEDBACK
state feedback. We
r function of FE! in
1 of eFE! in (7-30).
(7-34 )
--ex,,)]
+
- e
(7-35)
ores)]
+ epr(s)
341
(7-39)
t-> 00
and the design is completed. We note that the design can also be achieved if
{A, b} is not control1able but is stabilizable. In this case, however, we do not
have complete control over the rate of convergence of y(t) to rd'
Much more can be said regarding the design ofasymptotic tracking problem.
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.
*Multivariable case. Consider the n-dimensional linear time-invariant,
multivariable dynamical equation
FE:
x=Ax+Bu
y=Cx+Eu
(7-40a)
(7-40b)
where r stands for a reference input vector and K is a p x n real constant matrix,
caBed the feedback gain matrix; and Equation (7-40) becomes
(7-37)
FEJ:
le identity
(7-41 a )
(7-41b)
Method 1
In this method we change the multivariable problem into a single-variable
problem and then apply the result in the previous subsection.
342
_._-------_.-,_._----------~------------_._----_
..
_---~--_._-_.
__.--_._---_.
---~-_._-----_
.. _------_.
can conclude immediately that A is cyclic if and only if the Jordan canonical
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
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
property can be easily deduced from Corollary 5-21.
Theorem 7-5
lf {A, B} is controllable and if A is cyclic, then for almost any p x 1 real vector v,
the single-input pair {A, Bv} is controllable.
Theorem 7-6 4
If {A, B} is controlla
the eigenvalues of A
Proof
where a, i= 1, 2, ... ,
tion of X(s) with respt
L1'1
If L1(s) has repeated re
sufficient condition fe
Proof
O]
2 21 O:
O O
O
1:0
A= O O 2: O O
~--~--~-r~--~
There is only one Jordan block associated with each distinct eigenvalue; hence
A is cyclic. The condition for {A, B} to be controllable is that the two rows in
B encircled by dotted lines in (7-42) are nonzeros (Theorem 5-21).
The necessary and sufficient condition for {A, Bv} to be controllable is that
ex 1=0 and f3 1=0 in (7-42). ex = Oand f3 =0 if and only if VI =0 and vtfvz = - 2/1.
Hence any v other than VI =0 and vI/vz = - 2 will make {A, Bv} controllable.
This establishes the theorem.
Q.E.D.
The cyclicity assumption in this theorem is essential. Without this assump
tion, the theorem does not hold. For example, the {A, B} in
an
O
an-
O
2a n_
--------.
an - I
O
O
a'l-
V,;/e are
D.OV/
ready
Theorem 7-7
--_._--~-----
--_ .. _--_._--_._._.-----------
------------- --_._----------------- -
~-
---~-- -~--
-=.~~::=.:----=-~~----~--~--------_._-------_._~
STATE FEEDBACK
--.-_.--
--~-----
343
Theorem 7 _6 4
If {A, B} is controllable, then for almost any p x n real constant matrix K, all
the eigenvalues of A + BK are distinct and consequently (A +BK) is cyclic.
Proof
+ BK be
IY P x 1 real vector v,
Li'(s)=nsn-l +(n- 1)ad'-z
+ ...
+an-l
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
ee the basic idea, we
O
1
O
O
1
o
O
a
an-I
det ------------------ n------------------ ~ y(kij) =0
O
O
O
n
a,,-l 2a n-2
O
O
O
(n - l)al
n
an-I
O
lct eigenvalue; hence
that the two rows in
n 5-21).
e controHable is that
=0 and Vl/VZ = - 2/1.
{A, Bv} controllable.
Q.E.D.
Vithout this assump
in
is controllable.
lly one Jordan block
lion for A to be cyclic
an-l
4This theorem can be extended as follows: If {A, B,C} is irreducible, then for almost any p x q
real constant matrix H, the eigenvalues of A + BHC are distinct and consequently (A+ BHC) is
cyclic. The matrix A + BHC r~sults from the constant output feedback u = r + Hy. The proofor
Theorem 7-6 is equally applicable here.
-._._-_.~
..
_,_..
__
~-~~_
_----_._-~
344
.. __..
~------_
..-----_..
__.__.._._--_._--_._------_._-_..
~--"_._
_---~
_.~-
.. _.. _ ..
"-~
..
_~
_~---_._.,._----_.-
__ _.__._-_._._._.,_ _---_._._.__
..
._--_._-----~-_
__ .',_.'---_ __
Proof
If A is not cyclic, we introduce u = w + Kx such that ~ A + BK! in
i =(A +BK)x +Bw
(7-43)
M =[b
Ah!
Q.E.D.
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
Figure 7-2
.-
----.:.-==--=-:::.::----===-====- -=-=-=---====-.-====-=..=.===-..:==:=::--==-------
---_. __
--------- -- - _._--------_.__.__. -._-_.
.
- - - _ - - - - _ _ _. - - - _ _ . _ - - - - _ . _ - - - - -
-~
----_._----_._-----_.
..
345
STATE FEEDBACK
(7-43)
[A~l
i::J
[~lJ
and the eigenvalues of A22 are not affected by any state feedback. Hence we
conclude that the eigenvalues of (A + BK) can be arbitrarily assigned if and
only if {A, B} is controllable.
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
are MI = 3, Mz = 2, and M3 = 4. Then the matrix
+-Br
.Bvk) can be arbitrarily
Jmbining the state feed-
(7-44)
l
Q.E.D.
~ell'l
~eZ1'2
~e31'J
en
e 13 A
e 13 A z
e 22
P = e 22 A
~elI'IA/ll-l
(7-45)
e 3 4
e 34 A
e 34 A 2
e 34 A3.
[uation.
_... _-----_._----
.
..
_~----._----------_
This matrix can be obtained directly from the coefficients of linear combinations of A~' b without
computing explicitly M-l. See Reference S178.
_---._~--~
_~
346
o 1 O
O O 1:
I
X'X
X: X
x'x
.. __ ...... __ J ........ __ 1_
A= PAP- 1 =
:O
:
:
.. _
........ _
...... _
--------~-----~---i--6--0-
:
:
: O O
: O O
x:x
x:x
1 O
O 1
O O
O O O
1 X O
------
O O O
B=PB= O 1 O
------
O O O
O O O
O O O
1:
X: X
and
Method '"
O O 1
We introduce a method o
forming A into a controlI~
equation. See Appendix j
(7-46)
where the unfilled positions are all zeros. The pair {A, B} is said to be in a
multivariable controllable formo The introduction of u = r + Kx, where K is a
3 x 9 real constant matrix, yields
Algorithm
Consider a controlIable {
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
be arbitrarily assigned. For example, we may choose K so that CA +BK) is
of the form
O 1 O:
O O 1 :
al az a3 :
.
.
bz :
l
-------~-----~----------
:
:
:,
:O 1 O O
:O O 1 O
:O O O 1
or
.... -
_1- _
.. _ _ .. 1_ ..
........ __
000;00;0010
000:00:0001
! ' 6
11(7 l..-1.3
A a9
l
L C.,11 d
'2 d 3 ' d
~4 {ts
(A+ BKT-I)l
Hence A + BK and F are sin
F can be arbitrarily chosen so
the eigenvalues of A +BK cal
As discussed in Appendix
solution T always exists in Al
F lIave common eigenvalues
BK
'
. . T o remove this uncert~
elgenvalues. A necessary co
controlI~ble and {F, K} obser
for the smgle-variable case (p =
Theorem 7-10; the proof will
Hence the assertion will not bl:
.,
(7-47)
The left-hand-side matrix has three blocks of companion form on the diagonal;
the right-hand-side matrix has only one companion formo Of course, it is
x x]
lf JI, > JIz > JI3, the three nonzero rows of O become O 1
lOO]
O I
[ O O 1
r.~l
~
0'1"
x.
lf
JII
001
.
In our example, JIl = 3, JIz = 2, and JI3 = 4, and the three nonzero royvs ofO are
o~~J
01000:0000
00100:0000
000.10:0000
000:01:0000
O O O: O O: 1 O O O
S9
-a 3 s Z - azs .
- dgs S _ dsS 7
STATE FEEDBACK
1, 1't
d7
l'S
straightforward,
O O
O O O
1 x O
S=PB= O 1
O O O
O O O
O O 1
(7-46)
: O O: O O O 1
: d 4 d s : d 6 d7 ds d9
(7-47)
xl
.~ J
they become
dsS 7
... -
Cl)
d2s - d1
Method 1I1
We introduce a method of computing the feedback gain matrix without trans
forming A into a controllable formo It will be achieved by solving a Lyapunov
equation. See Appendix F.
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
values.
1. Choose an arbitrary n x n matrix F which has no eigenvalues in common
with those of A.
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.
4. Ir T is nonsingular, then we have K =KT- l , and A +BK has the same
eigenvalues as those of F. Ir T is singular, choose a different F or a different
K and repeat the process.
I
We justify first the algorithm. Ir T is nonsingular, the Lyapunov equation
AT - TF = - BK implies
(A+BKT-l)T=TF
or
A+BK=TFT-l
Hence A + BK and F are similar and have the same set of eigenvalues. Since
F can be arbitrarily chosen so long as its eigenvalues are distinct from those of le,
the eigenvalues of A + BK can almost be arbitrarily assigned.
As discussed in Appendix F, if A and F have no common eigenvalues,s a
solution T always exists in AT - TF = - BK for any K and is unique. Ir A and
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}
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.
Hence the assertion will not be proved here.
8 Note
mples.
S9 - d 9sS
O--O--
"
possible to obtain sorne other forms, for example, two blocks of companion
form on the diagonal or a block triangular form with two or three blocks of
companion form on the diagonal. The characteristic polynomials of the
(A + SK) in (7-47) are
and
0- - - -
347
348
x=[L_Ltjj] +[LJ]u
O
O
Y = [1
00:01
O
O :-1 -2
-1
3: 2
O)
O
O
(7-48)
O
1
If we choose K as
[-2
K=
.
-25
O
-55
O
-48
-1
-23
-lJ
1
O
O
O
-55
O
1
O
O
-48
O
O
1
O
-24
j]
-5
elements of K2 , which
most of the assigned e
cIear that the larger the
tudes of the feedback g;
Next we compare tI
yield the same eigenva
transfer matrix C(sI _)
input and zero-state re
compare only their ze
(h) the responses of
[2
1
O -1 _
K is roughly three tim
gains and yields smaller
For a companion-for
is shown in Reference S
proportional to
10
then we have
A +BK
If we choose K as
=lr
1(2
-9
-s
-1
-4
-~J
then we have
A+BK2=[_~~ ~tJjj]
__
..
..
00:01
O :-5
,
-2
This matrix has tw. 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
~
__ ~
_~
,_~
349
STATE FEEDBACK
"j~u
1
(7-48)
O
O
.-1]
-5
O
24
(IIlmaxt -1
where IIlmax is its largest eigenvalue in magnitude. The magnitudes of feedback
gains are also proportional to m. Hence, in order to have small feedback gains
s +25
-1
23
j]
10
~2J:;t~~~5~~~4;====t"""'---i6----------
-5
(a)
-10
,J]
,
~,
-2
lrm. One has eigenvalues
lcks are nonintera0ing. .
The elements of K 1 , whlch
uger in magnitude than the
-5
Figure 7-3
Transient responses.
sec
"_~
~~
350
and small transient, the order of the largest companion-form block should be
kept as small as possible.
The orders of companion-form blocks are determined by the lengths of
chains ofvectors b, Ab, ... , Akb in (7-44). Ifthey are chosen by using Scheme 2
discussed in (7-21), the orders of companion-form blocks are equal to the
controllability indices of {A, B}. In this case, the largest order is equal to the
controllability index f1. If we use Scheme 1 discussed in (7-20) or any other
scheme, the length of the longest chain, and consequently, the order of the
largest companion-form block will be larger than, at best equal to, f1. Hence in
order to have small feedback gains and small transient, we shall use Scheme 2
to search the linearly independent columns in [B AB ... An-lB] and the
orders of the resulting companion-forrn blocks are f11, f12, ... , f1 p , the con
trollability indices.
The introduction of state feedback can increase the order of a companion
form block, as shown in Example 3, but can never decrease it. Hence in finding
a K:, we should preserve the orders of the original companion-form blocks.
Even so, there are stiB a lot of flexibility in choosing K. We may choose K so
that the resulting matrix is in block diagonal form or in block triangular form3.
Furthermore different grouping of desired eigenvalues will yield different K.
If we require the resulting matrix to be block diagonal and group the desired
eigenvalues to minimize
mas {(IXll max )!'! - \ (iX 2Imax)!'2 - \ ... , (IXplmax)!'r l}
where IXdmax is the largest eigenvalue in magnitud~ in the ith companion
form block of order f1, then the flexibility bf choosing K is considerablyreduced.
A problem may sometimes arise in preserving the orders of companion
form blocks. For example, let f1 = 3, i = 1, 2, and let the desired eigenvalues
be - 3 3j, - 4 2j, and - 3 j. In this case, it is not possible to choose a
real Kso that A + BK preserve f1 and has a block diagonal or a block triangular
formo Ifwe combine the two companion-form blocks into one, then the result
ing A +BK is
I
l
O
O
O
1
O
O
O
O
O
1
O
O
O
O
-2804 -920 -180
1
O
O
O
O
O
-4800
O
O
-3J
~l
O'
-lJ
(1-49)
O
O
-1
O
1
-11
_1- _ _ _ _ _ _ _ _ _ _ _ _ _ _
O
1
O
O
O
O
1
O
O
O
-8
-54 -36 -9 : -20
1
O , O
O
O
,
,
O
O
O
O
O
1 : -60 -44
18
6
and of eigenvalues
gains required in (7-5
shown by a computer
(7-50) is much much s
We recapitulate t]
formation of {A, B}
Scheme 2 discussed i
[H AH ... An- 1 B
{, B} are equal to t
the order ofthe larges1
among aH transforma!
diagonal form with c<
this is not possible du
Problem 7-20 to choo
K =KP may yield a sy
The design procedt
single-input system; h<
the method should not
However, the method 1
be discussed in Section
Up to this point, w
eigenvalues. This dep
such as the rise time, se1
signals, and so forth. )
simpIe answer to the I
simulations. Of cours
not be unique. The 01
consequently, a unique
formance index
(7-50)
In this subsection, we ,
eigenvectors of the resu
assigned eigenvalues, Xi,
associated with Xi, that i
or
(A
BI
BK[e
STATE FEEDBACK
e in the th companion
K is considerably reduced.
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
shown by a computer simulation that the largest magnitude in the transient of
(7-50) is much much smaller than the one in (7-49). See Reference S17.
We recapitulate the preceding discussion in the following. In the trans
formation 01' {A, B} into a multivariable controllable form, we should use
Scheme 2 discussed in (7-21) to search the linearly independent columns in
[B AB ... An-l B]. The orders 01' companion-form blocks in the resulting
{A, B} are equal to the controllability indices, fl, 1, 2, ... , p, 01' {A, B} and
the order 01' the largest companion-form block in {A, B} is the smallest possible
among all transformations. We then choose a K so that (A +BK) is 01' block
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
Problem 7-20 to choose a K. This process 01' choosing K and consequently
K = KP may yield a system with small feedback gains and small transient.
The design procedure discussed in Method I transforms the problem into a
single-input system; hence the order 01' the companion-form block is n. Thus
the method should not be used to avoid large feedback gains and large transient.
However, the method may require a smaller-dimensional state estimator as will
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
such as the rise time, settling time, overshoot, largest magnitude 01' the actuating
signals, and so forth. Even if these criteria are precisely specified, there is no
simple answer to the posed problem.One way to proceed is by computer
simulations. 01' course, the set 01' eigenvalues obtained by this process will
not be unique. The only known systematic way 01' finding a unique K and,
consequently, a unique set 01' eigenvalues is by minimizing the quadratic per
formance index
J=
O
O
O
O
1
-20
O
O
1
O
-180
(7-49)
O
O
~
-1
(----
)
~
-11
10'"
[x*(t)Qx(t) +u*(t)Ru(t)] dt
351
(7~50)
(A +BK)ei= Xie i
i=1,2, ... ,n
BKe = (XiI - A)e ~ f i
,_._-_.- ....
_-_
....
_~~--~."
.__
._._._._~_._---~-~_._,-_
352
..
_~---_
..
_-
__
. _ __ _
, ... _ . ~ ,
~ , ~ _ .__ ~ .
__
~.
_ _ , u . . . ~ ~
~_g
' _ ' _ _
~~~~_'
'
~'_'._'~_.~'
_ _ ~ ' _ _
'_~AA_".
._~"-
._~.~,--~--
Now if {e i , i
f2
. ..
f n] [e 1
e2
. . . e n] -
(7-51 )
Unless all columns of the matrix on the right-hand side of (7-51) are inside the
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
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
discussed, then a solution K exists in (7-51). In fact, the solution is unique an
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.
Once we choose the structure of A +SK (one, two, or three companion-form
blocks), and once we assign the eigenvalues for each block, the eigenvectors are
practically uniquely determined. Hence if the eigenvectors are chosen as
discussed, there is no essential difference between the assignment of eigen
vectors and the assignment of structure of ( + SK) discussed in the previous
subsection. For further discussion of the assignment of eigenvectors, the reader
is referred to References S91, S125, S160.
Effect on the Numerator matrix of G(s)
In the single-variable case, we showed that the numerator of 9(s) is not affected
by any state feedback, whereas its denominator can be arbitrarily assigned.
Now we establish a similar statement for the multivariable case. Before pro
ceeding, the reader should review the material in Section 6-6, especially Equa
tions (6-100) through (6-117).
Once {A, B, C} is transformed into the form in (7-46), by reversing the
procedure from (6-100) to (6-117), we can obtain
C(s) =C(sIT - A)-l 10
+E =
N(s)lD -1 (s)
+E =
(N(s)
+ ED(s))JD -
(5)
with
~(s)
Figure 7-4
which implies
u(s) = (1- ~
The substitution of (7-5:
y(s) = (N(,
Gf (
We see that the numera
state feedback. The col
matrix of the denomin
any state feedback. Th(
assigned by a proper ch,
case.
If G(s) is factored a1
affected by any state feed
~tor of every element of
G(s) wiJI be affected by st
feedback gain matrices I
differel1t traj].sient respon
Computational problem
STATE FEEDBACK
we have
>
-n
]-1
353
(7-51 )
+ y(s)
+
Figure 7-4
which implies
(7-52)
o(s) = (1- KL(s)D-l(S))-l r (S)= D(s)(D(s)- KL(s)tlr(s)
The substitution of (7-52) into y(s) = G(s)(s) = (N(s) + ED(s))D- 1(s)i(s) yields
(7-53)
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) D(;)-KL(s) are not affected by
any state feedback. The lower degree part of D (s) however can be arbitrarily
assigned by a proper choice of K. This result is similar to the single-variable
case.
If G(s) is factored as N(s)D -1(S), the numerator matrix N(s) will not be
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
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
Computational problems
'-I(S)O(S)
This observation was provided by Professor D. Z. Zheng ofTsinghua University, Beijing, China
- . - ._-
._,u ._._.__
".~
..
_.~L_
~._.,,
..
"_~._.
_._
.__ .__.._..
~_
~_
..__ .__
._._~
.~
.. __ .
.__ _.__
~~
~~_
.,,"
._ _
__._.__ ..
--~----~-------------~~~~~
354
lJ
7-4
State Estimators
Figure 7-5
An open-Io
STATE ESTIMATORS
355
state observer. In this section we shal1 introduce two types of state estimators:
ful1-dimensional and reduced-dimensional. The design procedure is basical1y
dual to the design of state feedback; hence its dscussion will be brief. The
single-variable case will not be discussed separately and will be included as a
special case of the multivariable case.
In this section we use the circumflex over a variable to denote an estimate
of the variable. For example, is an estimate ofx; i is an estimate of X.
ansforming an equation
.kB. The method is first
d then transform it into
'dering of the state vari
orthogonal transforma
Id is numerically stable.
minations without any
msformation method is
FE:
X=Ax +Bu
y=Cx
(7-54a)
(7-54b)
the transformation of an
he solution ofa Lyapunov
been extensively studied
and St07. The algorithm
~rically stable and may be
Method IIIin terms of
carried out ih Reference
) clear-cut conclusion was
L
Figure 7-5
~~~~
356
This is very inconvenient. Second, and more seriously, if the matrix A has
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
general, not satisfactory.
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
computed (see Problem 5-17). However, pure differentiators are not easy to
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
that if both the output and input are utilized, the performance of an estimator
can be improved.
Consider the state estimator shown in Figure 7-6. The estimator is driven
by the input as wel1 as the output of the original system. The output of FE,
Y=Cx, is compared with y ~ Cx, and their difference is used to serve as a
correcting termo The difference of y and Cx, y -Cx, is multiplied by an n x q
real constant matrix L and fed into the input of the integators of the estimator.
This estimator wil1 be cal1ed an asymptotic state estimator, for a reason to
be seen later.
The dynamical equation of the asymptotic state estimator shown in Figure
7-6 is given by
(7-55 )
The asymptotic e;
either by a block diag
If the eigenvalues of (A
error x can be control
negative real parts sm
zero at rates faster thar
x(to) and x(t o) at initi2
the eigenvalues of A
is much more desirabl
In the fol1owing \
estimator.
Method I
In this method, we ap
state estimators.
l
,------------------1
I
I
,---------
--.JI
L-
Figure 7-6
I
I
I
I
Figure 7-7
An asympt
STATE ESTlMATORS
357
(7-56 )
x-x
Clearly x is the error between the actual state and the estimated state. Sub
tracting (7-56) from (7-54), we obtain
i=(A-LC)x
(7-57)
If the eigenvalues of (A - Le) can be chosen arbitrarily, then the behavior of the
error x can be control1ed. For example, if al1 the eigenvalues of (A - Le) have
negative real parts smal1er than - CJ, then al1 the elements of x will approach
zero at rates faster than e-a'. Consequently, even ifthere is a large error between
x(to) and x(t o) at initial time t o, the vector will approach x rapidly. Thus, if
the eigenvalues of A - Le can be chosen properly, an asymptotic state estimator
is much more desirable than an open-loop estimator.
In the following we discuss two different methods of designing a state
estimator.
Method I
In this method, we apply the design procedures for state feedback to design
state estimators.
,----------------------,
I
I
I
I
I
I
IL
Figure 7-7
==_._._
~ ~
358
__..
_~.~.
..
..-.=--::::="-::::""=_.::::::::.:: .._:::,.::::
:_ __ ,:,::
__",._" ._"__..
~~"_~
-.~._ =-~ =--~"-="::::
.~:::,
__". __..__
._._~
..
.. _._
.~.
~~
__
~~
__
__
._.~
_.~_~.
..,._.
.~_u
_ ~
__ . . .._.
_~
._,_.u~_
-;
Theorem 7-8
Proof
Define
Then we have
~ = (A LC)x + Ly + Bu
x =x xgoverned by
FE:
X=Ax +Bu
y=Cx
(7-58a)
(7-58b)
(7-59)
Theorem 7-9
Proof
1. TA-Ff=GC
2. H=TB
3. All the eigenvalues of F have negative real parts.
(7-60)
STATE ESTIMATORS
359
Proof
is observable, its state
ltor
Define
e ~ z-Tx
Then we have
=:i - Ti = Fz + Gy + Hu - TAx - TBu
f {A, e} is observable,
'assigned. Indeed, if
notes the transpose, is
genvalues, we can find
(Theorem 7-7). Since
le same, the theorem is
edback can be used to
not be repeated. We
and small transient in
cks in the transformed
tate-feedback part, the
fl-58a)
(7-58b)
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).
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
trollable, we can find a u(t) to generate a x(t) which makes e(t)+O as t-+ oo.
This establishes the necessity part of the theorem.
Q.E. D.
fl-59)
1 real
constant matrices.
Proof
360
dynamical equation
no means simple nUD
It is not clear at presen
efficiency and numeril
det ,(F) =
TI ,(},) + O
i= 1
Reduced-dimensio
Method I
Consider the
n-dimen~
where R is an (n -q) x
as P is nonsingular. V
In =PQ
O
(7-61 )
=[
Now we transform
x=Px,
From this equation we conclude that if ranJe UF < n or rank V A < n, then rank
T < n (Theorem 2-6) and T is singular. However, rank UF = n and rank V A = n
do DOt imply the nonsingularity ofT. Hence they are only necessary conditions
for T to be nonsingular.
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,
{F, G} is control1able and {A, C} is observable. This establishes the theorem.
Q.E.D.
';[1 r
T=_,-I(F)UFAqV A
Lx J=l
2
=[
STATE ESTIMATORS
les
}.
361
of F and
yields
;CA = FGC
Method I
+ GCA
f equalities:
x=Ax +Bu
y=Cx
....... -
(7-63a)
(7-63b)
P~ [~J
GCA +F"-IGC
(7-64 )
(7-65 )
o
O
(7-66)
(7-61 )
V A is the o bservability
d is similar to the one in
on (7-61) implies
j =PAP-Ix +PBu
y=CP-lx=CQx=[Iq O]x
which can be partitioned as
(7-62)
(1'-67 a )
(7-67b)
quiresthe' computation
lblem ofTA - FT = GC
i z =A 22 x Z +AzlY +Bzo
e transformation of a
.J
(7-68a)
362
We note that and w are functions of known signals u and y. Now if the
dynamical equation in (7-68) is observable, an estimator of x 2 can be constructed.
Figure 7-8.
A comparison betw
estimators is in order.
alence transformation d
required in the reduced
dimensional case. In th
of integrators. In the I
the constant matrix Q t
.h
Wlt noises, the nbises VI
dimensional estimator t1
noises in y will be sup~n
In
Theorem 7 -11
The pair {A, C} in Equation (7 -63) or, equivalently, the pair {A, C} in Equation
(7-67) is observable if and only if the pair {A 22 , A 12 } in (7-68) is observable.
The controllability part of this theorem was implicitIy established in Section
5-8. See also Problem 5-34. Thus the proof ofthis theorem is left as an exercise.
Ir {A, C} is observable, then {A 22 , A2} is observable. Consequently, there
exists an (n - q)-dimensional state estimator of x 2 of the form
(7-69 )
Method 11
Consider the n-dimensior
(7-70)
elir~lnated by
defining
z =i 2 -Ly
(7-71)
(7-72)
be an (n - q)-dimensional ,
(n - q) x (n - q), (n - q) x q,
or
ni
e(t) approaching zero or, equivalentIy, the rate of (z + Ly) approaching x 2 can
Figure 7-8
An(n -q)-dimen:
=='--__
- ---- _
=:::_=-:c-_--~
STATE ESTIMATORS
(7-68b)
itrarily assigned by a
X=Q~=[Q
(7-71 )
Qx or
Q2J[L/+Z]=[Q
Q2J[~ I~-JGJ
(7-74 )
(7-70)
eliminated by defining
FE:
x=Ax +Bu
y=Cx
(7-75a)
(7-75b)
z=Fz+Gy +Hu
-(B 2 - LBd"
(7-72)
(7-76)
l U
-----------------------,
--~
-LAd{z + LX 1)
~A12X2 - LB 1"
363
(7-73)
h
.. _- -_ .. _...
--.--~._-----_
__._..-_ .._--_.
364
_._-_.-.-._--_._----- ----_.
__ .._.__ .__...-
Now we consider tl
theorem by showing th;
Algorithm
1. Choose an (n - q) x (n - q) real constant matrix F so that al! of its eigen
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
...
Define
Il = Al VAf = LB]
Fn-q-1G]=n-q
P=[~J
[7-77)
P=[~J
is nonsingular are {A,e} observable and {F, G} controllable. For the single
output case (q = 1), the conditions are sufficient as well.
If the n-dimensional d:
sional state estimator, a:
cles1,ed :igenva!ues (pr
can be constructed, Whl
Proof
First we note that Equation (7-62) still holds for the present case. The only
differences are that F is an (n - q) x (n - q) matrix and UF is an (n - q) x nq
matrix. We write
P=
[~J
=[
-Ll
J=
-l(~UFAqV [~
-Ll
~l(F)J[UF~qV
7-5
Connection.
365
C ] r-O
[ UFAIV A
1n]', ar
n-q
is an (n - q) x n matrix.
0-77)
(7-78)
;h that z is an estimate
nmediately (7"78).
(7-79)
~~ IlllXn=
n-
..
..
.
.
fJn
IXI
1 . IX
2
IX n.- 2
IX n.- 3
1
O
"
UfA. V Ar=
L fiF i j-=
G=
fiiFi- 1 G =0
(7-80)
i= I
(F)][U:qVJ
Hence we have rank
e, there .exists ~onzero
and Pr = O. Hence P is
7-5
FE:
x=Ax +Bu
y=Cx
(7-81 a)
(7-81 b)
---_
366
_ __ .
"'_._._-..
_._--_
_-~
.._ ,_ .. __ _._. __ ..
_------".~---_._.-
-.-.--" ..
..
~-_
_-_ ..
_~.~
..
(7-82a)
To simplify (7-85), w
(7-82b)
[Q1
Q2J
(7-83)
Ir the eigenvalues of F are chosen to have negative real parts, the estimated
state x(t) will approach the actual state x(t) exponential1y.
The state feedback is designed with respect to the actual state of FE. Ir the
actual state x(t) is not available for feedback, it is natural to apply the feedback
gain K to the estimated state x(t), that is,
u =r +Kx
(7-84)
[i] [A
i
+BKQ1C
GC +HKQ1C
y =[C
BKQ2
F +HKQ2
OJ [:]
][x]
[B]
z + H r
(7-85a)
(7-85b)
This is obtained by the substution of (7-84) and (7-820) into (7-81) aad (-La).
Figure 7-9
The eigenvalues of a
transformation; henc
system in Figure 7-9
that the eigenvalues (
as far as the eigenval i
from the estimated st:
of state feedback ane
pendently and the eig'
feedback and those (
separation property.
We discuss now
shown in (5-54). He]
7-9 can be computed
as
(Theorem 5-16). Thi~
the use of a state estim
and does not appear
explanation. In the c,
to be zero. Consequ
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
of those of A + BK. '
367
Ion n or n - q can be
wing, we consider only
~ is left as an exercise.
'ed in (7-76) and (7-78):
(7-82a)
(7-82b)
[:]=[A
y=[C
(7-83)
ed in this cannection:
: resulting equation are
te feedback u = r + Kx,
the eigenvalues of the
(3) What is the effect
o y? To answer these
lf the system in Figure
(7-85b)
Q2][:] +[:]r
BK
F
OJ [:]
(7-86a)
(7-86b)
We discuss now the transfer matrix from r to y. Equation (7-86) is ofthe form
shown in (5-54). Hence the transfer matrix of (7-86) or of the system in Figure
7-9 can be computed from
x = (A + BK)x + Br
as
(7-85a)
~BK
y = Cx
G(s) =C(sI -A -BKt lB
(Theorem 5-16): This is the transfer matrix ofthe perfect state feedback without
the use ofa state estimator. In other words, the estimator is completely canceled
and does not appear in the transfer matrix from r to y.This has a simple
explanation. In the computation oftransfer matrix, aH initial states are assumed
to be zero. Consequently, we have x(O) = x(O) =0, which implies x(t) =x(t) for
al! t. Hence, as rar as the transfer matrix fram r to y is concemed, tnere is no
difference whether a state estimator is employed or no1.
If x(O) i=- x(O), the rate of x(t) approaching x(t) is determined by the eigen
values of the estimator. Clear1y, the larger the magnitudes of the negative real
parts of these eigenvalues, the faster x(t) approaches x(t). However, these will
cause larger gains in L and generate larger magnitudes in transient, as in the
state feedback case. As the eigenvalues approach negative infinity, the esti
mator will act as a differentiator and will be susceptible to noises. Large gains
and large transient will also cause the estimator to saturate. Hence there is no
simple answer to the question of what are the best eigenvalues. Ithasbeen
suggested in the literature that the eigenvalues of A - LC or F be chosen to be
two or three times faster than the eigenvalues of A + BK; that is, the magnitudes
ofthe negative real parts ofthe eigenvalues of A - LC or F betwo or three times
of those of A +BK. This seems to be a simple and reasonable guideline. lf a
368
d-!---
y = [1
: O
-2
OJ x
u(t) = r(t) + \j
lf-
~Jx
-1
O
5
Functional estimat
an estimator of dimens
not necessary to recom
some functions of the
In these cases, the dim
In the following, we d
kx, where k is an arbitr
functional estimator or
Consider the dynaI
is to design an estimat
we transform (7-63) in
y=
!]
r -6
and
~]
5
O
det(51 - A) = 53 +/152 -(5 +/2)5 - (/3 + 5/ 1 )
where x = Px with P ~
output w(t) will be cho:
where F is a m x m m
problem is to design Sl
that w(t) will approach I
7-9, we can show ina i
/2
= -
36
1
O
Z2
Z3
240 y +
O u
576
-2
o [I q oJ can be replaced by
developmenl, with slight m
a stochastic process,
square error. The
"
-84
O
O
x= -36
xas
+ [~ Jf- \:1.
-1f-Jx
I
Id -1 j.
O]
r as - 3 and - 3 2j.
LA 12 are - 3 and
. LetL' =[/1 /2 /3]'
369
[:J
y = [Iq O]x ~ Cx
where x = Px with P given in (7-64). The functional estimator with a single
output w(t) will be chosen of the form
!2 -
w = Mz
+ Ny
(1-87a)
(1-81b)
TA-FT=GC
H=TB
-5/)
ls +39
(7-88a)
(7-88b)
and all eigenvalues of F have negative real parts, where T is a m x n real constant
matrix, then z(t) approaches Tx(t) as t~ oo. Hence, ir M and N are designed
to meet
(1-88c)
- ]u
-2
t~oo.
'o [lq O] can be replaced by [C, O], where C, is.aq x q nonsingularmatrix, and the subsequent
development, with slight modification, still holds. This su bsection follows closely Reference SIlO.
370
[I q O],
TA +T 2A 2 -FT =G
TA2 +T 2A 22 - FT 2 =0
The mq x mq ma trix
Theorem 2-4 that, fe
if and only if the mq
(7-89a)
(7-89b)
and
(7-89c)
(7-89d)
F= [
JJ
-rx. m
M=[
Theorem 7-14
O ]
where rx.i can be arbitrarily assigned. Note that {F, M} is observable. Let
tij,j = 1, 2, ... , m, be the jth row of Ti, i = 1, 2. Then because of the forms of F
and M, (7-89d) and (7-89b) imply
t2 =
k2
t ll A 12 +t 2A 22 =t 22
t 12 A 12 +t 22 A 22 =t 23
(7-90a)
(7-9Ob)
The substitution of these t 2j ,j = 1,2, ... ,m, into the last equation of (7-90a)
yields, after sorne straightforward manipulation,
= - [tm t.m-
t2
rx.~1
rx.~1
tll]
*7-6
.I
rx.1
Decouplin
Consider a p-input p
eq uation description
[
rx.1II_1
(7-91 )
The vector k is a 1 x (n - q) vector and is known oncek is giveri and rx. i are chosen.
where u is the p x 1 iJ
The rnq x rnq matrix in (7-91) is nonsingular for any (Xi' Thus we conclude from
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
Id T 2 are, respectively,
written as, using =
(7-89a)
(7-89b)
V=
are, respectively, 1 x q
A12
A1Z
:A 22
1zAz21
(7-89c)
(7-89d)
I,J
Theorem 7 -14
lf the dynamical equation in (7-63) is observable with observability index v,
then for any l x n real constant vector k, there exists a (v - l)-dimensional
functional estimator with arbitrarily assignable eigenvalues so that its output
approaches kx(t) exponentially.
o ]
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
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
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;;o;
In the design, if we augument the matrix C in (7-63b) as [e' K'T, then the
dimension of estimators may also be reduced. See References S92, S97, Sl10,
Sl31 and S184.
1;
(7-90a)
(7-9Ob)
\.ZZ+tl. m-
371
1 1z
*7-6
FE:
(7-91 )
x=Ax+Bu
y=Cx'
(7-92a)
(7-92b)
......
372
__
. '. _~'_'._
_ _
.~n
: :_
._
. :.~_.::..:.
_ _::..:_.~
_ - - - . __
. . . . _.-~._.
-
__
-:-
G(s)=C(sI-A)-lB
(7-93)
o.
+ glP(S)Up(s)
+ gzp(s)uP(S)
(7-94)
+ ...
P constan
+gpp(s)up(s)
where gij is the ijth element of G(s). We see from (7-94) that every input controls
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,
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
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.
where
G(s) is the th
Example 1
Consider
Definition 7-1
Ez = 1
(7-95 )
x=(A+BK)x+BHr
y=Cx
Theorem 7 -15
feedback of the form u
(7-96a)
(7-96b)
is nonsingular.
H
Fig u re 7 -10
.....
~~~
Decoupling by state feedback_
__._----- .- --
_-
--._--
------:==::::::::::=::::-._--
373
(7-97)
We shall derive in the following the condition on (;(s) under which the system
can be decoupled by state feedback. Define the nonnegative integer d as
system is initiaHy in
di ~ min (the difference ofthe degree in s ofthe denominator and the numerator
of each entry of the ith row of (;(s))-I
p(s)Up(s)
p(s)up(s)
(7-94)
E ~ lim
pp(s)uP.(s)
Sd;+ l
(;(s)
S-oo
Consider
(;(s) = [ "
~:~1
s2+~s+1
"+; +2]
s2+ s +4
The differences in degree of the first row of (;(s) are l and 2; hence dI =0 and
El = lim s [
s'" 00
sfer-function matrix is
s +2
S2
+s +1
S2
J=
+s +2
[1
OJ
The differences in degree of the second row of (;(s) are 2 and 2, hence d 2 = 1 and
E2
lpling ofmultivariable
= hm
s2 [
s""oo
1
~2----'----
s +2s +1
3J
11
Theorem 7 -1 5
(7-95)
i constant nonsingular
e feedback is shown in
(7-96a)
(7-96b)
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
and
(7-98)
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
Prablem 2-39)
CBH=O,
and
1
= Ll( ) [CBsn- 1 + CR1Bsn- z + ... + CRdBsn-d-l + ... + CRn-1B]
s
(7-99)
Weestablish in the
EandE i.
Theorem 7 -16
where
Ll(s) ~ det (si - A) ~ sn +CX1S n- 1 + ... +cx n
and
(7-100)
and 11
(7-102)
(7-103 )
CiB=O,
CiAB=O,
... ,
CAd-lB=O
(7-104)
(7-106)
where
~(s) ~ det (si - A - BK) ~ Sil +~lSn-1 + ... +~n
(7-107)
E and CRdB may dirrer by a constant because >(s) is factored out in (7-99).
For convenience,
It is easy to verify th
CAd-lB=Oimply th:
and
C(,
C(A + BK)'
Consequently, we havt
and
Since H is nonsingulaJ
we conclude that di = (.
With these prelimir
Proof of Theorem 7-1
(7-105)
...
R1=A+cx11
Rz=AR 1 +cxzI=A z +cx1A+cxzI
(7-101 )
11
then
is a diagonal constant
is nonsingular by assu;
Sufficiency: If the mah
Define
__ '
"_~'_W'_"._.'_'
"~
_ _ _ _ , . ~ .
__
.~
__
_ _ _"_ __ __
..
w_._~._.,,
__"~_"'_
-_.. - . __.- ..
_.~
'.'-
,,_._
~y
. . . . _ _. __ __ . ~ " _ " _ - _ . _ . -
and
(7-98)
(7-108)
then
CBH=O,
and
375
(7-109)
(7-110)
(7-99 )
Theorem 7 -16
+0:
11
(7-100)
di = d and Ei = EH.
Proof
(7-101 )
. +O:n-l 1
he numerator and the
19) that the coefficients
the coefficient matrix
have
(7-102)
(7-103)
Iy that
=0
(7-104)
and
k = 0,1,2, ... , di
CitA + BK)k =CiAk
CitA +BK)k =CiAd(A +BK)k-d
k =d +l,d +2, ...
(7-111 )
(7-112)
Consequently, we have
and
C(A +BK)kBH =0
k =0,1, ... , di~ 1
CitA +BK)dBH=CiAdBH =EH
(7-105)
Necessity: Suppose that there exists K and H such that G(s, K, H) is diagonal
<md nonsingular. Then
'unction matrix Gf in
is a diagonal constant matrix and nonsingular. Since E = EH and since H
is nonsingular by assumption, we conclude that E is nonsingular.
Sufficiency: Ir the matrix E is nonsingular, then the system can be decoupled.
Define
(7-113)
in (7-99). For convenience,
- "
... - - _ . .
_ _ . _ . _. . -
376
or, equivalently,
l/S(d1 + 1)
O
e, we immediately obta.
decoupled by using K =
1/s
(d2
+ 1)
(7-114)
where e is a row vector with 1 in the ith place and zeros elsewhere. First we
show that C(A +BK)d+l =0. From (7-105), (7-112) and (7-113) and using
K= -E- 1F and EE- 1 =e, we obtain
CiCA +BK)d+l =CAd'(A +BK)=CAd+l +CAdBK
=F -EE- 1F = F -eF =0
Although a system c
the resulting system is n
system are at the origino
to move these poles to tiIf the dynamical equ~
values by state feedback.
the number of eigenvall
reduced. For a complete
compute K and H to have
36,41, 112, S218, and S2:
7-7
Concluding Rf
where F and E are the ith row ofF and E, respectively. Hence we conclude that
CiCA +BK)d+k =0
Since
(7-115)
1
1
C(sl - A - BK)-lBH ==- [CRdBsn-d - + CRd+ 1Bs,,-d-2 + ...
.
Ms)
+ Ct-1 BJH
(7-116)
Now, from (7-105), (7-108), (7-112), (7-115), and the fact CiCA +BK)kB =0; for
k =0,1, ... , d -1, which follows from (7-109) and the nonsingularity of H,
it is straightforward to verify that
Cj~.dB =C(A + BK)dB = CAdB = E
CR d+ 1B= C[(A + BK)d +l + !Y. 1(A+ BK)dJB = &l E
C(sl-A-BK)-lBH=L\~S)[sn-d-l +&lSn-d-2
(7-118)
(7-119)
---:~---~--~~-----~------~. ------~----------.-------------
CONCLUDING REMARKS
(7-114)
os elsewhere. First we
and (7-113) and using
decoupled by using K =
(7-115)
integer k
BS n -
d -2
+ ...
_CRn-BJH
(7-116)
in
. + !XII-d-)
-riJ=O
(7~1l5),
(7-119)
we canobtain
Next by multiplying
ng similatly, we can prove
and usingEH=EE- =
E- I F and H = E-l.
Q.E.D.
7-7
Hence we conclude that
377
Concluding Remarks
378
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.
Problems
7-8
Figure P7-7
FE I :
x=Ax +bu
y=cx
Ir the vectors {b, Ab, ... , An -1 b} are used as basis vectors, what is its equivalent dynamical
equation? Does the newc bear a simple relation to the coefficients ofthe transfer function
of FE I as it does in the controllable canonical form? (Hint: See Problem 2-15.)
Let
7-3 Find the matrix Q which transforms (7-17) into the observable canonical-form
dynamical equation.
.. " An-l
if~ =~ ]x+[~}
y= [1
Ab
.
x=
O]x
[-~~
into the controllable canonical-form dynamical equation. What is its transfer functon?
7-5 Let f.l.l be the largest integer such that {b[, Ab[, ... , A~l - Ib l } is a linearly independent
seto Let f.l.l be the largest integer such that {b[, Ab[," .. , A~l -1 b[, b1 , Ab 1 , " .. , A~2 -1 b1 }
is a Iinearly independent seto Show that Anb 1 for all n f.l.l is linearly dependent on
{b[, ... , A~l -Ib[, b1 , ... , A~2 -1 b1 }.
7-6 Transform the dynamical equation in Problem 7-4 into the observable canonical
form dynamical equation.
Gven
PROBLEMS
Consider the state feedback and the output feedback systems shown in Figure P7-7.
Show that for any constant matrix H, there exists a constant matrix K such that Kx = HCx.
Under what condition on C will there exist a matrix H such that K = HC for any K? (C, K
and H are q x n, p x n, mul p x q constant matrices, respectively. It is general1y assumed
7-7
that nq.) (Answer: A solution H exists in K=HC for a given K if and only if p
379
[~J=Pc.
Consequently, a solution H exists in K = HC for any K if and only if e is square and non
singular.)
Ix(k))
Ilom {y(O), y(l), ... , y(k)}.
I + 1)} and modeled as
I:x(k
+ 1))
Figure P7-7
7-8
Let
A=[~ ~J
b=[~J
c=[1
O]
Find the values of k = [k 1 kz] such that the matrix (A + bk) has repeated eigenvalues.
Find the values of h such that the matrix (A + bhc) has repeated eigenvalues. Can you
conclude that for almost al1 k and h, the two matrices have distinct eigenvalues?
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.
7-9
7-10
Show that
7-11
(A -
bk)b
1, -1 b l } is a linearly independent
. '. A"' -1 b z}
IA",-Ib, bz, Ab z, ...
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
stabilize the equation by using state feedback? Ir yes, find the gain vector k such that the
closed-loop equation has eigenvalues - 1, - 1, - 2, - 2, and - 2.
7-12
lnto the observable canonical-
Given
------.;=:::::-.--._.==:-.::__
.-_--.. ---=-=~.~-~.:::::.-.-:::::--:::::-.--:::::.---:..:..:. -:::::--::::::::::--:::::.. -:::::
. --::::::: : .- : : :-.-: : : : : "': : :"-:: : "-', : : ---:: : -- ------.---"'-'--'--'"--.---.----.- - - - - - - - - . - - . - . ------.---.---.-.--.----_.. '~"'--'" - - - 1
380
Find the gain vector [k l k 2 ] such that the state-feedback system has - 1 and - 2 as its
eigenvalues. Compute k l , k 2 directly without using any equivalence transformation.
7-13
1
2
7-21
7-22
Let
O
O
O -1
Is it possible to find a gain vector k such that the equation with state feedback u = kx + r
has eigenvalues -2, -2, -1, -1? Is it possible to have eigenvalues -2, -2, -2, -1?
Howabout -2, -2, -2, -2? (Answers: Yes; yes; no.)
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
prove the observability part?
7-23
7-24
7-14
<:
7-9.
7-15
7-25
7-26
Find a two-dimensional state estimator with eigenvalues - 2 and - 3 for the dynam
ical equation in Problem 7-4. Use two different methods.
7-16
7-17
(s
+ l)(s- 2)(s + 3)
y=
Find a full-dimensional es
estimator for kx(t) with
(s+2)(s+3)
7-19
7-27
_~
Consider the matrix in Problem 2-30. Show that the eigenvalues of the matrix are
Al, A2" .. , AII , - , CI. jj3, ')'1' ')'20' .. , ')'112 _ 1 if aij are chosen as
7-20
"1- 1
-1
+ ...
II
i=Ol
8ds)=aI21 Sn2 - 1
+... +aI2112=fi
/12 -
TI (s-y;)
i=l
7-28
(5 - ')') .
I
,
.~.
where A is an n x n matrix, b i:
function. Show that gis) ha~
PROBLEMS
7-21
7-22
Let
A=
state feedback u = kx + r
lValues -2, -2, -2, -1?
[~ ~ ~l
381
~~ ~l
B= 1 2
02
-3 1 2 3
2100
Find two different real constant 2 x 4 matrices K such that the matrix (A
values -43j and -5 4j.
7-23
Show that the state estimator in Theorem 7-8 meets Theorem 7-9.
7-24
Show that the (11 - q)-dimensional state estimator in (7-72) and (7-74)meets Theorem
7-9.
7-25
7-26
o
O
1
O
O
O
1
O
1
1
1
1
O
k=[O
)ack, can you carry out the
he estimator as - 4 and - 4.
TI
(s -A)
= 1
Prove Theorem 7-11. Let Vi be the observabilty indices of {A, C}. Show that the
observabilty indices of {A n , A12 } are v-1. (Hint: Compute
-A l l
and note that if linearly independent rows of the observability matrices are searched in .
order from top to bottom, the locations of the linearly independent rows of both observ
ability matrices are identica!.)
7-28
/12 -
TI
i= 1
(s-Y,)
-IJ
7-27
111 -
x=Ax +bu
y=cx
where A is an 11 x 11 matrix, bis an n.x 1 vector, ande is a 1 x 11 vector. Let gbe its transfer
function. Show that g(s) has m zeras-in other words, that the numerator of g(s) has
382
Find a k so that (A + bk
for i=O, 1,2, ... , n-m-2
and cAn-m-1 b f= O. Or equivalently, the difference between the degree of the denominator
and the degree of the numerator of g(5) is J = n-m if and only if
cAd-1bf=O
cAib=O
(Hin!: Show that cAib is invariant under any equivalence transformation and then use the
controllable canonical-form dynamical equation.)
7-29
52
+1
"['J
25 +1
53 +5 +1
X~[~
y=[~
-~} +[~
1
O
1 -1
~J x
-1
2
~}
-1
x(k +1)=
y(k) = [
-2
-1
O
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.
7-31
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
-1
F= [ ~
if
.,d-2
formation and then use the
T= -3
[
-1/2
ion
12
-4
-1/3
-51 ]
-1/4
20J, then
383
STABIl
8-2 Stabilityel
Descri ption
Time-varying case
systems with only on
shown in Chapter 3 t
satisfy the homogenei
then the input u and t
y(t) =
\u(t
under what condition e
output y satisfies
8-1
Introduction
for allt in ( -
00,
oo)?
384
385
y(t) =
(8-1)
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,
for all t in (- co, co)
(8-2)
under what condition on the system does there exist a constant kz such that the
output y satisfies
ly(t)I.:::; kz <
OC!
:d.
(8-3)
lf the input approaches a periodic function . under Wh3.t concilic);} \vill the
output approach another periodic function with the same period? If the
input approaches a constant, will the output approach sorne constant? Accord
ing to these various properties, we may introduce different stability definitions
for the system. We shall introduce here only the one which is most commonly
used in linear systems: the bounded-input-bounded-output stability.
Recall that the input-output descl'iption of a system is applicable only
when the system is initially relaxed. A system that is initially l'elaxed is called a
relaxed system. Hence the stability that is defined in terms of the input-output
description is applicable only to relaxed systems.
Definiton 8.-1.
A relaxed system is said to be BIBO (bounded-input-bounded-output) stable
if and only ir fol' any bounded input, the output is bounded.
386
STAB!
output.
Let us choo
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
the initial voltage across the capacitor is not zero, because of the negative
I
capacitance, the output will increase to infinity even if no input is applied.
Clearly u is bounded.
Theorem 8-1
is not bounded.
We consider now
terminals and q outPl
is BIB stable if and only if there exists a finite number k such that
where u is the p x 1 in
q x p impulse-responso
G(t,
Proof
Sujficiency: Let u be an arbitrary input and let lu(t)1 =:;;, k l for all t in ( Then
ly(t)1 = \
Loolg(t,
00, 00).
r)1 dr =:;;, kk l
for all t in ( - 00, 00). Necessity: A rigorous proof of this part is rather involved.
We shall exhibit the basic idea by showing that if
tI'
In
In
u"-'
-i
In
00
A relaxed multivariable
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).
axedness" by showing
; assumption, may not
387
em is initially relaxed,
en y(t) =u(t)/2 for all t.
Jounded. However, if
ecause of the negative
o input is applied.
={ ~
-1
if x =0
if x>O
if x <O
Loo \g(t
1,
r)1 dr = 00
Q.E.D.
is oot bouoded.
(8-5 )
where u is the p x 1 input vector, y is the q x 1 output vector, and G(t, r) is the
q x p impulse-response matrix of the system. Let
9l1(t,r) glZ(t,r)
G(t, r) = gZl~t, r) gzz\t, r)
gql(t, r)
gdt, r)
glP(t,r)l
gzp\t, r)
. (8-6)
gqp(t, r)
Then gij is the impulse response between the jth input terminal and the ith
output terminal. Similar to single-variable systems, a relaxed multivariable
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
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.
Theorem 8-2
[00
Ig(t, r)i dr
~ kk 1
y(t) =
is BIBO stable if and only if there exists a finite number k such that, for every
entry of G,
--_ ..
388
_-----------~------_._-------_
.. _-----
----------_
._-
..- --
-.---~---.-_._-_
STABIL
Time-invariant case.
input-output description:
(8-7)
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
g(t - L)U(L) dL
00
Proof
Ig(t, L)I do =
Ig(t - L)I dL =
Ig(o:)1 do: s
Theoreril 8-3
We shall u:
= O;
(8-8)
00
I
Graphically, it says that the total area under Igl is finite. The fact that 9 is
absolutely integrable does not imply that 9 is bounded on [0,(0) nor that g(t)
approaches zero as t -> oo. Indeed, consider the function defined by
n +(. t - n)n
-1
for - 3 < (t -n)sO
n
n -(t - n)n 4
forOs(t-n)s3
f(t-n)=
{
.r
1
n
Since
and since
is nondecreasing as ex increa
Consequently,
389
Id,
(8-7)
n
II
2
Figure 8-2 An absolutely integrable function that is neither bounded nor tending to
zero as (---+00.
as a ----+ oo. We shall use this faet in the proof of the following theorem.
at
100 Ig(a)1 da
Theorem 8-3
(8-8)
Q.E.D.
y(t) =
J~ g(t -
,)u(,) d,
0, (0) if
Since
r
o
\g(/)1 dI =
r r
o
Ig(/)1 dI +
Ig(/)1 dI
and since
-n)~O
Ig(t)1 d(
Conseq uen tI y.
00
390
STABIl
If
00
Let
UM
~ rnax lu(t)
I
Ir
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).
2. If u is bounded and tends to a constant, then the output will tend to a constant.
3. Z If u is of finite energy, that is,
(rO
which appraaches .
ciently large, (8-12)
00
then the output is also of finite energy; that is, there exists a finite kz that
depends on k l such that
g(r)u(
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.
It is c1ear that
(8-10)
y(t) = J>(r)u(t - r) dr
and
r +
y(t+T)= Jo
l
,+
r
g(r)u(t+T-r)dr= Jo
which is independen
3. It can be shown tha
with the property
g(r)u(t-r)dr
(8-11 )
Ir+
g(r)u(t -r)drl
r'+T
=::;
J,
IIJII ~ (
rl + T
J.
11
Ig(-r)\ dr
J(t)gi
OSIST
+ T) as t
-4
oo.
This can be eXlended lo as fol1ows: For any real number p in [1. co]. if
(J '"
o lu{t)lpdt
)I/P
::;;k<co
g (r)
lJl
--\-
.,..-__
(r
and the system is said to be Lp-stable.
. tI
Figure 8--3:
The convolut
391
y(t) =
1',
Let UM ~ max
, Iu(t)l Then we have, with t > t b
\r
(8-12 )
'1
g('1:)u(t -'1:)d'1: \ s
Ig('1:)\ d'1:
which approaches zero as t 1 -'> 00, following from (8-9). Hence if t 1 is suffi.
ciently large, (8-12) can be approximated by
y(t) =i=
(8-13 )
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,
g('1:) d'1:
3. It can be shown that the set of aH real-valued functions defined over [O, (0)
with the property
(ro
(8-10)
sk < ro
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
(see Problem 2-49):
(8-11 )
(1, g) {
ro
f(t)g(t) dt
I[} f(t)g(t) dt Is (t
dl
r)\ d'1:
t Iy(t) - y(t
(I"
Ig(t)i2 dtY
+ T)\ --> o as
]. ir
u (t - T)
tY:+-:~-....,....------------f-::--~----;I'''''''''--t------~-------I-_T
Figure
8-3~
392
STABlI
With these preliminaries, we are ready to proceed with the proof. Consider
y(OI
I f~ g(r)u(t -
r) dr \
~ [ ' ig(r)llu(t -
r)1 dr
for sorne k and if u(t)
lf
too (ig(r)11/2)(ig(r)/1/2Iu(t - rm dr
(8-14 )
Proof
Since sin (t - r) = sin t
Consider now
too ly(t)i2 dt
~k
y(t) =
g(r)u(t-r)dr
(8-15 )
1t is clear that
where we have changed the order of integration. By assumption, u is of
fini te energy; hence there exists a finite k 1 such that
too.1u(t - r)12 dt
~k
11
00
g(r)sinw(
~kk1
too Ig(r)1 dr =k 2k 1
f>(t -
By definition, g(s) is th
roo g(
Jo
r)u(r) dr
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
f>(t -
y(t)
r)u(r) dr
If
-,JI eh
I ::::: 0,
00
then
)d,
.",
ig('r)l[u(t
(8-14)
_,)1 2 d,
393
as t ---> 00
(8-16)
where 8 = tan -1 (1m g(iw)(Re g(iw)) and g(s) is the Laplace transform of g('); Re
and 1m denote the real part and the imaginary part, respectively.
Proof
Since sin (t - ,) = sin t cos ; - cos I sin ;, we have
y(t) =
g(,)u(t - ;) d;
d; dI
=sinwt
(8-15)
ly assumption, u is of
to
g(,)cosw,d;-coswt
It is clear that
Ir
t'"
g(;)sinw;d;-
[g(;)[[sin w(t -
t'"
t ---> 00,
g(,) cos w, d; -
f"
')1 d,::;
g(;)sinw(t-;)d-c
Ig(,)1 d;
we obtain
COS WI
t'"
(8-17)
r'" g(t)e-
Jo
iwr
dt =
Jo
Re g(iw) =
:le output will have the
necessarily true. The
1m g(iw) = -
wl
di
JO
g(t)coswtdt
(8-18a)
10'" g(t)sinwtdt
(8-18b)
Q.E.D.
394
Theorem 8-4
A relaxed single-variable system that is described by a proper rational function
9(S) is BIBO stable if and only if all the poles of 9(S) are in the open left-half
A relaxed multivariab
is a proper rational-fu
of every entry of C(s) 1
s plane or, equivalently, all the poles of 9(S) have negative real parts.
By the open left-half s plane, we mean the left-half s plane exc1uding the
imaginaryaxis. On the other hand, the c10sed left-half s plane is the left-half
s plane inc1uding the imaginary axis. Note that stabilty of a system is inde
pendent of the zeros of 9(S).
Proof of Theorem 8-4
f3
(s-Af
and possibly of a constant, where A is a pole of 9(s). Consequently, g(t) is a
sum of finite number of the term t k - ) eA,! and possibly of a 6-function. lt is
easy to show that t k - ) eA,! is absolutely integrable if and only if A has a negative
real part. Hence we conc1ude that the relaxed system is BIBO stable if and
only if all the poles of g(s) have negative real parts.
Q.E.D.
Example 2
Consider the system with a transfer function g(s) = l/s. The pole of 9(S) is on
the imaginary axis. Hence the system is not BIBO stable. This can also be
shown from the definition. Let the input be a unit step function, then u(s) = l/s.
Corresponding to this bounded input, the output is fE -) [g(s)u(s)] =
fE -) [1/s 2 ] = t, which is not bounded. Hence the system is not BIB stable.
For multivariable systems, we have the following results.
Corollary 8-2
A relaxed multivariable .system that is c;lescribed .by
y(t) =
LG(t - .)u(r) d.
8-3
Routh-Hurv
D(s) =aos n +
where the a;'s are real nu
we shall give some nec
Hurwitz polynomial-t
then D(s) can be factore
D(s) =c
=(,
ROUTH-HURWITZ CRITERION
395
is BIBO stable if and only if there exists a finite number k such that, for every
entry of G,
Theorem 8-5
Corollary 8-2 and Theorem 8-S follow immediately from Theorems 8-2
and 8-4 if we consider every entry of G as the impulse response of a certain
input-output pair.
8-3
ed by partal fraction
:m
::::onsequently, g(t) is a
of a b-functon. It is
mly if A has a negative
is BIBO stable if and
Q.E.D.
Routh-Hurwitz Criterion
ao>O
(S-19)
where the a's are real numbers. Before developing the Routh-Hurwitz criterion,
we shall give sorne necessary condition for D(s) to be Hurwitz. If D(s) is a
Hurwitz polynomial-that is, if all the roots of D(s) have negative real partsthen D(s) can be factored as
..
D(s) =ao
= ao
(s +ctd
(S2
(S-20)
396
where Cf.k > 0, fJ j > , 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
not be a Hurwitz polynomial; for example, the polynomial with positive co
.
efficients
S3
+ SZ + lis + 51 =
(s
+ 3)(s -
1 +4i)(s -1 - 4i)
that is, if n is even, Do(s) consists of the even part of D(s) and D1(s) consists of
the odd part of D(s); if n is odd, Do(s) consists of the odd part of D(s) and D1(s)
consists of the even part of D(s). Observe that the degree of Do(s) is always one
degree higher than that of D1(s). Now we expand Do(s)/D1(s) in the foHowing
Stieljes continued fraetion expansion:
1
Cf.ZS
+-------------
Cf.3 S
+---------
(S-22)
Table S.1 Continuous-
+---
s"
s
11-1
a~O)
a)O)
a~l
a\l)
-----~---
Theorem 8-6
The polynomial D(s) in (S-19) is a Hurwitz polynomial if and only if the n
numbers Cf.l> Cf.z, , Cf. n in (S-22) are all positive.
Sll- 2
a\il
a\2)
Sn-3
ab"l
a\3)
a\j'-2l a1(II-2)t,
Si
Exarilple 1
SO
Consider D(s) = S4 + 2s + 6s z + 4s + 1.
3
2s 3 +4s,and
a\j'-ll:
-'--------:
We have Do(s) = S4 + 6s" + 1, D1(s) =
a~l:
'
'
.. .
.__
..
. _ .. _.J_ ..__.
_. _.. .
ROUTH-HURWITZ CRITERION
Do(s) 1
1
--=-s+ 3
="2s+----
D(s) 2
2s +4s
.
1
4s 2
ts+---
+ 1
(8-23)
1
7 S +"7
"2s
.B.
=f
+a~Ols"- 2
D(s)=a~)s"-
(8-21a)
(8-21b)
+ ...
+dl ls"-3
+a~9~ S2 +a~9)
+ ...
(8-24 )
+a~~~s
Note that the number of coefficients in D(s) is one less than that of Do(s) if n
is even; they are equal if n is odd. Using these coefficients we form the table
shown in Table 8-1. lt is caBed the Routh tableo The coefficients in the first
two rows are just the coefficients of Do(s) and D(s). The coefficients in the
remaining rows are obtained from its previous two rows by using the formula
a(k+ la(k) _ a(k)a(k + )
a\k+2)=
<+~k+lo <+1 =a\~ -1Xk+a\\+/l
(8-25)
ao
(8-22)
Table 8.1 Continuous-Time Stability Table (for n even)t
n-S
Sil
abO)
a\O)
a~O)
a~?~
abl )
a\l)
a~l)
a(l)
IX"S
11-1
-----------------
5'1-2
397
5"-3
ab2 )
a~3)
a\3)
aff)
"'-1
----
a(Z)
a~Z)
a\Z)
a~9) :
"'-1
a(3)
n'- 2
:
I
a(M-Z) ,
1
,
(n-l)1
ao
0("_1
ag'-Z)
=-'
ag'-I)
Cf.,,=~)
ao
.t Ir n is odd, the first two rows have the same number of coefficients.
the pattern is idenlical.
Otberwise,
--
_-
_._---~-._~_
..
_-------_.~--_._---_._-------_
398
..
_-----_._--_.~-_._-_
.. _ - - - - - - - - - - - - - -
_--_._--_._----_._._------_._---._--_._--------"~-.~._._------------------_._--._-_._-.-._--
----_.~--~-_
...
~._-_._-_
where
(8-26)
There is a simple pattern in (8-25). Using this pattern, the Routh table can be
readily computed. We note that the number of coefficients will decrease by
one at every odd power of s at the leftmost column of Table 8-1. The (J. are
defined as the ratios of two subsequent coefficients in the first column of the
table. We shall show that these (J. are equal to the (J. in (8-22). In order to do so,
we use the coefficients in Table 8-1 to form the polynomials, for k = O, 1, ... , 17,
n-k
where
J2
-l
k'-
n-~-l
if n-k is even
Example 2
if n-k is odd
which implies
AH the coefficients ar'
(8-27)
This equation holds for k=O, 1, ... , n-2. For k=n-1, we haveDn(s)jDn-(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.
Example 3
a~)ja~-)s =
Consider 2s 4
+ 2s 3 + s
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
first part of the theorem foHows directly from Theorem 8-6. To show the
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
part, we write (8-25) as
- a(k, +2) + (J.k +1 a(k1 +1+1)
a(k)
1+1 -
ROUTH-HURWITZ CRITERION
(8-26)
399
Using this equation we can show that if a~) > 0, a~-I) > and (/.n-l > 0, then
O. Similarly we can show that the coefficients associated with 53 in
Table 8-1 are aH positive. Proceeding upward, we can show that if ag) > 0,
i = 1,2, ... , n, then aH the coefficients in Table 8-1 are positive. This establishes
the theorem.
Q.E.D.
a~'- 2) >
1
2
\-26),
6 1
4
4
1
3.5
1
e haveD,,(s)/D,,-(s)=
into the form in (8-22).
:d from the eoeffieients
Example 3
1.5
1 1
1 O
We
Example 4
Consider 25 4 +25 3
+S2
A negative number appears in the table; henee the polynomial is not a Hurwitz
polynomial.
We note that in eomputing Table 8-1, the signs of the numbers in the table
are not affected if a rowis multiplied bY'a positive number. By using this fact,
the computation of Table 8-1 may often be simplified.
400
Example 5
Consider D(s) = 2s 4
Theorem 8-8
5
There are four positive numbers in the first column (excluding the first number);
hence D(s) is a Hurwitz polynomial.
To conclude this section, we mention that the Routh table can also be used
to determine the number of roots in the right-half s planeo If ag> f.O, i =
1,2, ... ,n, then the number of roots in the right-half s plane is equal to the
number of sign changes in {ab1l , ab2 J, ... , ag'l}. See, for example, References
S41 and S44. If ag> =0 for sorne i, the situation becomes complicated. See
Reference S37 and a number of papers discussing this topic in Reference S239.
Time-varying case.
namical equation
x=
E:
A(t)x
y =C(t)x
+ B(t)u
(8-28a)
(8-28b)
The response of the state equation (8-28a) can always be decomposed into
the zero-input response and the zero-state response as
A state
only if
l'
l'
lO
Definition 8-2
Xe
of a dynam
xe =
lO
G(t,r)u(r)dr
401
,lIowing table:
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
(8-29)
00
The norm used in (8-29) is defined in terms ofthe norm ofu. At any instant
of time, Ilu(OII can be chosen as
lu(t)l, max lu(t)l, or
lu(t)12)l/2. Corre
sponding to these different norms of u,
15
has different values. However, as far as stability is concerned, any norm can
be used.
Next we study the stability of the zera-input response. More specifically,
we study the response of
x=A(t)x
ear time-varying dy
(8-28a)
(8-28b)
Li
(Li
Ilqt)Q)(t, T)B(T)II
initial state
Definition 8-2
; be decomposed into
t o, O, u)
zero-input response
hese results, we shall
ynamical equation.
nse. The response of
402
e t l'
Definition 8-4
An equilibrium stat<
i.S.L. at t o and if eve
t --> oo. More precise
any e> 0, there exist
Theorem 8-9
Every equilibrium st ,
and only if there exis
is stable i.s.L., but not uniformly stable i.s.L. See Reference S206. Roughly
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.
Example 1
Proof
Consider the pendulum system shown in Figure P3-17 (page 127). The applica
tion of Newton's law yields
Sufficiency: Let
u(t) cos
Let
Xl
e, and X2 = e.
Ir k
Xe
is i
be
e- mg sin e = mil}
Then we have
Xl
=X2
. = (g)
- sin
X2
Xl
COS Xl
+-----;;:;
U
k=O, 1, 2, ...
2, 4,
1, 3,
, are uniformly
stable i.s. L.
, however, are not stable
Figure 8-4
Asympto
403
Definition 8-4
Lyapunov at t o if and
jch depends on f. and
e sense of Lyapunov)
itive j which depends
in the large.
Theorem 8-9
Every equilibriurn state of x= A(t)x is stable in the sense of Lyapunov at to if
and only if there exists sorne constant k which depends on t o such that
00
Proof
ge 127). The applica
Sujficiency: Let
Xe
I>(t, tO)x e
for all t c. t o
t
miformly stable Ls.L.
.ivever, are not stable
Figure 8-4
.J
Asymptotic stability.
+T
404
Then we have
x(t) - Xe = cI>(t, tO)XO - Xe = cI>(t, tO)(X O- Xe)
(8-30)
Lyapunov. Consegu
for any t o and aH t.
}' > 0, and for every E
forallt~to
Q.E.D.
k3
The zero state is, as mentioned earlier, an eguilibrium state. If the zero
state is asymptotically stable, then the zero state is the only equilibrium state of
= A(t)x. lndeed, if there were another eguilibrium state different from the
zero state, then by choosing that eguilibrium state as an initial state, the response
wouldnot approach the zero state. Hence, if x = A(t)x is asymptotically
stable, the zero state is the only eguilibrium state.
Theorem 8-10
for any t o ~
(8-31 )
~22
for
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
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
lO
Figure 8-5
+ .
11<I>(t, t 0)11
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
y > O, and for every e> Othere exists a positive T such that
(8-30)
-xell
- xe \\:::;;6 implies that
that
1I <I>(t o
1. We prove this by
>(t, to) is not bounded,
say c/J/t, lo), becomes
:uch that aH the com
, which is equal to rx.
to)'rx, which becomes
;e X e is not stable Ls.L.
n CI>(t, lo) is bounded.
Q.E.D.
+ T, to)ll :::;; ~
for any t o
for aH t in [to, t o + T)
\\<I>(t, t o)\\ = \I<I>(t, t o +T)<I>(t o +T, t o)\\ :::;;1\<I>(t, t o +T)I\\\<I>(to +T, t o)\\
:::;;23
and so forth, as shown in Figure 8-5. Let us choose k2 such that e- k,T =~ and
let k 1 = 2k 3 . Then from Figure 8-5, we see immediately that
t ~ too
Q.E.D.
lf the zero state of X: = A(t)x is uniformly asymptotically stable, then for any
initial state, the zero-input responsewi1l tend to zero exponentially. Hence,
for linear equations, if the zero state is uniformly asymptotically stable, it is
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
at to if and only if
ero state is uniformly
xist positive numbers
(8-31 )
......
IIk3/2 - - - - - -
~k-4---
k2(HOl\lx o
405
ll
lO
Figure 8-5
+T
[ 31
----
i o + 2T
lO
k 18
3
.
-
+ 3T
::==::::--=.:.~.::-:::".:.-===::::::.=====-===:-._-~--'.'--_
406
..
_-_._-----"---~--~.
__
._._-----,-_._-----------~--
-~
_-_
._-------_._~,-~.
__
.-.-.--.~_
....
_._._-----_.
_.-------.._--
------------_.~_.~_._--~---.---.-_.--'-------~~------.
various definitions and theorems for the stability of the entire dynamical
equation E. Before doing SO, we shall discuss the relation between the stability
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
to be BIBO stable is that, for sorne finite k,
IIC(t)<1>(t,
for any t o and for all t ~ to. It has been shown by the function given in Figure
8-2 that an absolutely integrable function is not necessarily bounded. Con
versely, a bounded function need not be absolutely integrable. Hence, the
stability Ls.L. of an equilibrium state, in general, does not imply nor is implied
by the BIBO stability of the zero-state response. A function that approaches
zero as t--+ 00 may not be absolutely integrable; hence asymptotic stability
may not imply BIBO stability. If a system is uniformly asymptotically stable,
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
BIBO stability.
Theorem 8-11
for
Proof
Proof
an~ t o
Q.E.D.
XI
Hence we conclude f
initial state and any
<l>(t, e)B(T) is absolutel
bounded if and oniy ii
Total
true. A
If B and
stability.
stability cle~
systell,I may
C are bour
The situatic
rime~invariant ca~
407
of the zero-state response. Their relations are also established. We shall now
study the entire response.
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.
We see that the conditions of T-stability are more stringent than those of
BIBO stability; they require not only the boundedness of the output but also
of all state variables; the boundedness must hold not only for the zero state but
also for any initial state. A system that is BIBO stable sometimes cannot
function properly, because sorne of the state variables might increase with time,
and the system will bum out or at least be saturated. Therefore, in practice
every system is required to be T-stable.
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
el)
r <Il(t, ,)B(,)u(,) d,
Jo
t
',kz
JI1cD(t, ')11 do
Q.E.D.
Le matrices A, B, and C
oHable and uniformly
table (under the zero
BO stable.
Hence we conclude frum Theorems 8-8 and 8-9 that x is bounded for any
initial state and any bounded input if and only if <l>(t, to) is bounded and
<l>(t, ,)B(,) is absolutely integrable. From y(t) = C(t)x(t), we conclude that y is
bounded if and only if C(t) is bounded.
Q.E. D.
Total stability clearly implies BIBO stability; the converse however is not
true. A systeIl} may be totaHy stable without being asymptotically stable.
If B and C are bounded, then uniformly asymptotic stability implies total
stability. The situation is similar to Theorem 8-11.
Time-invariant case. Although various stability conditions have been ob
tained for linear, time-varying dynamical equations, they can hardly be used,
because aH the conditions are stated in terms of state transition matrices, which
are very difficult, if not impossible, to obtaio. In the stability study of linear
time-invariant dynamical equatioos, the knowledge of the state transition
matrixis, however, not needed. The stability can be determined directly from
the matrix A.
408
FE:
x=Ax +Bu
y=Cx
(S-33a)
(S-33b)
Consider
It has eigenvalues -1
eigenvalue O is not a
equilibrium states of th
Theorem 8 -1 5
Proof
Consider
Example 3
Equivalently, if A is transformed nto the Jordan form, the order of every Jordan blocks associated
wh eigenvalues with zero real parts is l. Note that this condition does not imply that the eigen
values of A with zero real parts are distinct
Proof
lf a linear time-inv,
response will approach z
stable. This is consiste]
the time-invariant case i
The eigenvalues of
A, det(sI - A) =0. We
Hurwitz criterion to che
negative real parts. Her
of x = Ax can be easily
nomial of A and then ap
The BIB stability (
determined by the pales
G(s) =C(sI
409
llamical equation
(8-33a)
(S-33b)
es,respectively. As in
llse and the zero-input
;ponse of FE is charac
Its equilibrium states are [x le X2e O]', for any Xl e -1=0, X2e -1=0. In other
words, every point in the Xl - X2 plane is an equilibrium state. The eigenvalues
of the matrix are -1, O, and O. Its minimal polynomial is s(s + 1). The eigen
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
Consider
~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
I
equilibrium states of the equation are not stable i.s.L.
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
~e is a constant k such
?;ular matrix such that
Alp.-l, then
= eA'(x o - x e ).
Proof
In order for the zero state to be asymptotically stable, in addition to the bounded
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
consequently of A have negative real parts.
Q.E.D.
Ir a linear time-invariant system is asymplOtically stable, its zero input
response will approach zero exponentially; thus it is also said to be exponentialIy
stable. This is consistent with Theorem S-lO because asymptotic stability in
the time-invariant case implies uniformly asymptotic stability.
The eigenvalues of A are the roots of the characteristic equation of
A, det (sI - A) = O. We have introduced in the previous section the Routh
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 charac: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
G(s) =C(sI -A( lB
410
x=G _~Jx+[~Ju
y=[1
Theorem 8-17
If a linear time-invarian
then the following state
lJx
[1
lJ
[S-1
-1
1. The dynamical eq ua
s+1
Let
(8-34a)
Ce J x
(8-34b)
y = [ Ce
2.
3.
4.
5.
A system is said to h
function matrix if the d
troHable' and observable.
pletely characterized by i,
the system can be detern
need of considering the d
A remark is in order 1
is independent of t and if
the zero state is asymptoti
that if fo~ each t, aH the eig
state of x = A(t)x is asym r
the folIowing exampk.
Example 5
Proof
Consider the linear time-v:
J-
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
theorem.
Q.E.D.
411
It is clear that total stability implies BISO stability. BIEO stability how
ever may not imply total stability because no condition is imposed on Ae in
the BIEO stability. A comparison of Theorems 8-15 and 8-16 yields that
asymptotic stability implies total stability but not conversely. Hence
asymptotic stability is the most stringent among these three different stabilities.
If a linear time-invariant dynamical eguation is controllable and observ
able, then the characteristic polynomial of A is egual to the characteristic
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.
,t tme). Consequently,
o-state response of FE
directly from Theorem
.te response in general
tate, because the zero
Lch, however, describes
al equation.
on description:
Theorem 8-17
If a linear time-invariant dynamical equation FE is controllable and observable,
then the following statements are equivalent:
- s +1
(8-34a)
(8-34b)
. [-1 J
2l
ields
sl-AcJ
Lo
_~c)-lBc1
u(s)
x=
e
O -1 x
+1
det [ Al ~ A] .= det [A. O
Hence the eigenvalues of A are - 1 and - 1 for aH l. However, the zero state
of the equation is neither asymptotically stable nor stable i.s.L., because the state
412
lI>(t'O)=[e~'
, '\' 1-
)~min
Xii
i= 1
*8-5
=P
Lyapunov Theorem
A hermitian matrix l\
for all nonzero x in en
or nonnegative definite
holds for sorne nonzer
Theorem 8-19
A hermitian matrix IV
any one of the fol1owiJ
l. AIl the eigenvalues
2. AH the leading prin
of M are nonnegati
3. There exists a nom
N*N. 6
Theorem 8-18
}'max
(8-35)
Ixd2
m,
are m,,_ m22_ m))_ det [
m2
i= I
and
Xi
Proof
Note that x*Mx is a real number fO( any x in en. Let P be the nonsingular
matrix such that P - I = P* and M =PMP*, where M is a diagonal matrix
with eigenvalues of M on the diagonal. Let x = Px or x = P - IX = P*x, then
n
i= 1
AJX'2
LYAPUNOV THEOREM
413
A min ~
1... Xi
"
(8-36)
max
i= 1
i=l
1... A X
__
IlxW=x*x=x*x=
I IX1
i= 1
Q.E.D.
Definition 8-6
A hermitian matrix M is said to be positive definite if and only if x*Mx > O
for all nonzero x in 1[:". A hermitian matrix M is said to be positive semidefinite
or nonnegative definite if and only if x*Mx::::::O for all x in 1[:", and the equality
holds for some nonzero x in C".
I
Theorem 8-19
A hermitian matrix M is positive definite (positive semidefinite) if and only if
any one ofthe following conditions holds:
" where
j::::
;::~ ;::~lJ'
Lm.31
m32
m!2'J'], det[m
m22
11
m31
nt:33
m,,], det[m n
m.:n
m33
m:n
minors whose diagonal elements are also diagonal elements of the matrix. The leading principal
~t
P be the nonsingular
~ is a diagonal matrix
,
1
414
Proof
dt
AlI the eigenvalues of A have negative real parts or, equivalently, the zero state
of x = Ax is asymptotically stable if and only if for any given positive definite
hermitian matrix N, the matrix equation
(8-37)
I
Corollary 8-20
AlI the eigenvalues of A have negative real parts, or equivalently, the zero state
of x = Ax is asymptotically stable, if and only if for any given positive semi
definite hermitian matrix N with the property {A, N} observable, the matrix
equation
A'!'M+MA= -N
has a unique hermitian solution M and M is positive definite.
where N ~ -(A*M
along any trajectory
- x*(t)Nx(t) is always
cally with time along
approach zero as t -> (
only at x =0; hence v
M and N that are relat
the zero state as t -> (
x=Ax. A Lyapuno'
concept of distance or
of x = Ax decreases w
Proof of Theorem 8-;
Sufficienc y: Consider .
I
dt
=x*
A*M+MA= -N
V(x(t =
(8-38)
Xl
Figure 8-6
A Lyapun,
LYAPUNOV THEOREM
415
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
value of V wil\ increase or decrease with time as the state moving along the
trajectory. Taking the derivative of V with respect to t along any trajectory of
x= Ax, we obtain
:t V(x(t)) =
efinite.
.
d
V(x) ~ dt V(x) = - x*Nx
8-20 is that if A is
r positive semidefinite,
However, it does not
,itive definite, then the
positive semidefinite.
few comments. Since
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
)
x*Nx
x*Mx -
(AN)min
(AM)max
-=---<----
V(x)
Figure 8-6
(8-38)
\
!
\
(8-40)
416
where (AN)min is the smal1est eigenvalue of N and (AM)max is the largest eigenvalue
of M. From Theorem 8-19 and from the assumption that the matrices M and
N are positive definite, we have (AN)min > and (AM)max > O. lf we define
CI.
~ (AN)min
- (AM)inax
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
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
state of x = Ax is asymptotical1y stable. N ecessity: If the zero state of x = Ax
A proof of the Ro
D(s) =
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
L'" eA 'INe
Do(s)
D (s) =CI.:
A'
dt
(8-41 )
xMxo =
(8-42)
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 xMxo > for all X o =/=-0, and M
is positive definite. This completes the proof of this theorem.
Q.E.D.
For convenience, we s
Since {A, H} is observable, all rows of He At are linearly independent on [0, 00).
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
identically zero over any finite interval, no matter how small; otherwise it
would be identically zero over [0, 00). See Theorem B-l. Note that HeA1x o = O,
at sorne discrete instants of time, is permitted.
With the preceding discussion, we are ready to establish Corollary 8-20.
Consider the Lyapunov function V(x) defined in (8~38)and dV(x)/dt =
- 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
Theorem 8-6
1I
ProoF
First we assume that
function
g(s) ~
.. =~
LYAPUNOV THEOREM
417
t --+ oo. This shows the sufficiency of the corollary. The necessary part can be
similarly proved as in Theorem 8-20 by using (8-42).
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.
and compute
Do(s)
--=<XlS
D 1 (s)
(8-41 )
+-----------:-------
<XzS
+----------:-----
1
<X3 S + - - - - - - - - - -
:finite, so is M. Let H
8-19). Consider
+--
(842)
+-
"011'
indepeodent 00 [O, el).
Al'IS
X o = O. Because e
',HeA'xo can oever be
)w smaH; otherwise it
Note that HeAtx o = O,
:ablish Corollary 8-20.
(8-38) and dV(x)/dt =
equality holds only at
;e with time, not neces
ually approach zero as
iX/S
The assumption <Xi =/=0, for i = 1, 2, ... , n, implies that there is no common
factor between Do(s) and D 1(s). Consequent1y, there is no common factor
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
function from u to y is 9(s). Let h 1(s) be the transfer function from X n to X n -1 or
equivalently, from the terminal E to the terminal F, as shown in Figure 8-7.
7
418
Figure.8-7
Then
A*M -t
1/a l s
1
D(s) = 1 +[1 +hl(s)J/als = 1 +als +hl(s)
y(s)
Let hz(s) be the transfer function from X,,-l to x,,_ z, then h l (s) can be written as
hl(s) =
Yazs
=
1~
1 + hz(s)/azs a 2 s +hz(s)
Proceeding forward, we can show easily that the transfer function frorn u to y
is indeed g(s). With the state variables chosen as shown, we can readily write
the dynamical equation of the block diagram as
Xl
Xz
X3
an
-1
an-
-1
an- Z
an- 1
O
Xl
Xz
X3
O u
I :
Xn-l
xn
l:
y=[ O
-1
az
-1
al
It is clear that N is .
(8-43)
lJx
*8-6
Linear Tim
DISCRETE-TIME SYSTEMS
419
M~ ~'
O
Ci. 1I _
O
O
O
O
Ci. z
J]
(8-44)
A*M +MA=
(s)
:r function from u to y
n, we can readily wrte
X3
X:-
XII
]x
O u
.
~ \
1J
O O O
O O O
O O O
O O
O O
O O
O O O
O O O
O O
O 2
~-N
(8-45)
(8-43)
g that it is controllable
lal to the degree of the
tic polynomial of the
'heorem 6-2). Now we
e asyrnptotically stable.
*8-6
The stability concepts introduced for the cootinuous-tirne systems are directly
applicable to the discrete-timecase. However, the conditions of stability are
quite different. In this section we shall discuss sorne of these conditions.
420
y(k) =
g(k - m)u(m)
_)
Then for any bounded-input sequence {u(k)}, (that is, there exists a finite h
such that lu(k)1 < h for k =0, 1,2, ...), the output sequence {y(k)} is bounded, if
and only if
C()
Ig(k)1 < ca
(8-47)
k=O
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)
Ir g(z) is a rational function of z, then the system is BlBO stable if and only if all
the poles of g(z) have magnitudes less than 1, or equivalently, aH the poles of
g(z) lie inside the unit circle of the z planeo This can be readily proved by
noting the z-transform pair
z-b
ao>O
(8-48)
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
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
8
a\O)
koa~,o~
Clbl )
-)klCl~,1~l
a\l)
kla~IJ.2
2
Clb )
a\2)
a~l-l)
aY'-l)
_)k"_laY'-I)
ag')
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
is complicated. We introduce in the following a method of checking whether
or not aH the roots of a polynomial are inside the unit circle without computing
explicitly the roots. The method is a counterpart of the Routh-Hurwitz
criterion.
Consider the polynomial with real coefficients
ll
kOCl~,O)
Cl~)
(8-46)
m=O
(-+ 00
OISCRETE-TIME SYSTEMS
le system described by
(8-46)
abo l
_) koa~,O)
ab )
1
421
- )kla~i~
ab
a\O)
a~?~ 2
Q~IO~ 1
koa~,o~ 1
a\l)
koaS-)
koa\O)
a~,l~
a~,l~
kla~1]2
k 1a\1)
a\2)
a~12~
a(O)
"
k 2= d,2] 2/a62 )
(8-47)
is the difference of its two previous rows. The remainder of the table is obtained
by the same proeedure until n numbers {abl >, ab2l , ... , a~)} are obtained. We
define (J. = agl, i = O, 1, ... , n.
Theorem 8-21
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
x(k
ao >0
(8-48)
+ 1) = Ax(k) + Bu(k)
y(k) =Cx(k)
Theorem 822
All the eigenvalues of A have magnitudes less than 1 if and only if for any given
positive definite hermitian matrix N or for any given positive semidefinite
hermitian matrix N with the property {A, N} observable, thematrix equation
A*MA-M= -N
has a unique hermitiansolution M and M is positive definite.
422
V(x(k = x*(k)Mx(k)
and computing
~ V(x(k~ V(x(k
+ 1
cxi _
Figure S-8
and
(S-49)
(S-50)
A block di
Define, for i =
i=I,2, ... ,n
(S-51 )
with
c=
with Do(z) = D(z). Note that the coefficients of Di(z) are the a~), j =0,1, ... , n - i
defined in Table 8-2. It can be verified that the reciprocal of Di(z) can be
ex pressed as
i = 1, 2, ... , n
kia::~i
(S-52)
and
(S-53)
-kakI
-k ak 2 cx 2 /CJ. I
A= - k ak 3cx 3/CJ.
= ag)(1-kf) or
(i+ 1)
IX
k2=1-~=I-~
ag)
lXi
1
D i - (z)
IX
IX
+ D;(z)) -'--
(ki_zD(z)
-k
-k
-kak,,_CJ./l_/CJ. -k
-kaCJ..!cx
-k
b=
(8-54 )
(S-55 )
lXi
These recursive equations are valid for i = 1,2, ... ,n. Note that D/l(z) = D/l(z) =
a\'l. Consider now the transfer function
D i - 1 (Z)
D-I(Z)
--- =
which becomes, by
k i - 1 zD i(z) + Di(z)
= k
zD(z) +k i - 1 Di(z)
then it is straightforward
A*MA-
usiri.g.(8~53),
Di-(z);"'k _
i
Di_(z)
9
(1 - kt_l)D(z)
zDi(z) +k i - 1 D i(z)
1 +----.:...--.:...-=..:.~,..:
IXDi(Z)/lXi-1D(z)
z+ki-Di(z)jDi(Z)
(S~56)
By assumption, we have a
it has the property that
eigenvalues of A and,cons
if and only if M is positi ve (
423
OISCRETE-TIME SYSTEMS
- x*(k)Mx(k)
Figure S-S
8-21.
Define, l'or i =
(S-49)
(8-50)
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'
Do(z)/Do(z) shown in Figure 8-9. Note that Dn(z)/Dn(z) = 1. In this block
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
!, ... , n
(8-51)
x(k
e = [1
(S-52)
O O .. , O O]
b=[~6
<xo
<xnJ'
<XO
and
k"f) or
(S-53)
Ax(k) + bu(k)
y(k) =cx(k) +kou(k)
with
+ 1) =
A=
-kok,
- k ok 2rx 2/rx
- k ok 3rx 3 /rx
-k 2k 3
ields
O
O
O
O
1
-kk2
-kk 3rx 3/rx 2
-k n 2 k,,-
-k Zcx,ja n
ll -
O
O
(S-57)
(8-54)
- 1
(8-55)
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.
For this matrix A, if we choose M as
al
()(2
()(n
.kt- I)D(z)
) +k_ )(z)
ao
z)
:(z)
(S-56)
.'
. ..-._..._-
-====:=====
_.~---------------~._ ~
------=-_==::::_~.::_c==c::.:_.:.=:
8-7
--------------------l
r-------------~
In this chapter we in
and the stability i.s.L
zero-input response.
being uniformly stab
distinetion between u
not available.
the Routh~Hurwitzcrj
the Lyapunov theorer
by using the Leverriel
eomputational errors.
form by a numerica11:
then be more easily Cl
form is also used in tl
Reference SI 07. One,
tion needed in the ROl
Lyapunov equation a
Routh-Hurwitz methc.
in checking the stabili
the former may also b
The coneepts of s1
The stability conditio
function He inside the '
BIBO stable; whereas
unit circle of the z pla
relationship ean be es
I
I
I
i
""
I-~
I
I
I
--------------,
,.
-..-{
I
I
I
I
I
I
I
I
I
I
: iT-
I i dJi
-;- ~I,,--..~
; --:
--
: L""r-J ... i .~
I .
J-~
IL
""-] :
1
J'Q<Q'I
-;:1-;:
l .
1;1;
I-~~
. - _,Jo}
'--J
L__---1~tr- I~I~:
_
------------
Concluding
...J
_o
1
?:J
.~ - 424
- - - - - - - - - - - -
~--
?:J
- - - - - - - - - - - - - - - - - -
I
I
~
~
;1; :;
"
~~ ~ -l
o
el
Problems
u..
8-1 1s a system with th
How about g(t, r) = sin te
CONCLUDING REMARKS
8-7
I----------l
I
I~:
1----,
II
I
I
1"
:~
!
Concluding Remarks
In this chapter we introduced the BIBO stability for the zero-state response
and the stability .s.L. and asymptotie stability for the equilibrium state of the
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
necessary and sufficient eonditions are established for the time-varying case,
they can hardly be employed because state transition matrices are generally
not available.
F or the time-invariant case, the stability can be checked from the poles of
the transfer function or fram the eigenvalues of the matrix A. Whether or not
all the eigenvalues of A have negative real parts can be cheeked by applying
the Routh-Hurwitz criterion to the characteristic polynomial of A or by applying
the Lyapunov theorem. The characteristic polynomial of A can be eomputed
by using the Leverrier algorithm (Problem 2-39), which however is sensitive to
computational errars. If the matrix A is first transformed into a Hessenberg
form by a numerieally stable method, the characteristic polynomial of A can
then be more easily eomputed. See Referenees S90 and S2l2. A Hessenberg
form is also used in the efficient method of solving the Lyapunov equation in
Reference S107. Once A is transformed into a Hessenberg form, the computa
tion needed in the Routh-Hurwitz method is mueh less than that in solving the
Lyapunov equation and eheeking the positive definiteness of M. Henee the
Routh-Hurwitz method is simpler eomputationally than the Lyapunov method
in checking the stability of A. Although no comparison has been carried out,
the former may also be more stable numerically than the latter.
The concepts of stability are equally applicable to the diserete-time case.
The stability eonditions, however, are different. lf all the poles of a transfer
function lie inside the open left-half s plane, then the continuous-time system is
BIB stable; whereas if all the poles of a transfer function lie inside the open
unit circle of the z plane, then the discrete-time system is BIBO stable. Their
relationship can be established by using the bilinear transformation
.2- 1
\--1
'Q
s =_._
z +1
lO
~I~
f':
,;.o
I~~:--.1_ _
1..><:
I g
1:;5
425
which maps the left-half s plane into the unit circle in the z planeo Although we
may transform a diserete-time prablem into a eontinuous-time problem by
using the bilinear transformation, it is simpler to check the stability directly
on discrete-time equations.
l<e
I
1CJ'l
I
,
IClO
~I~
~ ~
\~
:J
IQ Q I en
___________ ...1 u.
Problems
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?
8-'
426
Is the network shown in Figure P8-2 BIBO stable? Ir not, find a bounded input that
will excite an unbounded output.
8-2
1F
ti
8-11
8-12
Figure P8-2
8-3 Consider a system with the transfer function g(s) that is not necessarily a rational
function of S, Show that a necessary condition for the system to be BIBO stable is that
19(s)1 is finite for all Re s ~ O.
Consider a system with the impulse response shown in Figure P8-4.
8-4
Ir the input
Y'
Is the zero state asympto
stable? Is the equation 1
8-13
Consider
u(t) = sin 2nt, for t ~O, is applied, what is the waveform of the output? After how many
o
Figure P8-4
+ t)
8-5
8-6
8-14
8-15
BIBO stable?
+ 1) BIBO stable?
a.
S5
is BIBO stable.
8-16
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.)
8-8
Give the necessary and sufficient conditions for the following polynomials .lo be
Hurwitz polynomials:
8-9
8-10
8-17 Showthatx=Osti
t --> OO.
PROBLEMS
427
Prove parts 1 and 2 ofTheorem 8-3 by using the Laplace transform for the c1ass of
systems describable by rational transfer functions.
8-11
8-12
X-l
O
-1
y=[ O
1
O
O
O
-1
O
1
O
O
-2
O
O
1
O
- 10
!}+W
-4
-IJ
OJx
ot necessari\y a rational
o be BIBO stable is that
Is the zera state asymptotically stable (for the case u =,O)? Is its zero-state response BIBO
stable? Is the equation totally stable?
8-13
Consider
)"=[1
1Jx
Find all the equilibrium states Of the equation. Is every equilibrium state stablc .s.L.?
Is it asymptoticaIly stable? Is its zero-state response BlBO stable? Is the equation totany
stable?
8-14
stable?
o stable')
8-15
Find the ranges of k and k 2 such that the system with the transfer function
he foIlowing polynomials
+k
is BIBO stable.
8-16
,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
t---+OCi.
Show that x = O is the only solution satisfying (<1>(t, to) - I)x =,0 if 1I <1>(t, to)"":"O as
428
8-18
E : * =2ex +u
y=x
Sho\\l that the zero state of E is nol stable i.s.l.(under the zero-input response).
Consider the equivalent equation of E in Problem 8-18 obtained by the equivalence
transformation.;; = PU)x. where P(C) =e- r ':
8-19
Figure P8-26
E:
y=el:!.x
Show that the zero state of Eis stable i.s. L. (under the zero-input response). From Problems
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?
8-20 Show that stability i.s.L. and asymptotic stability of the zero state of Ji: = A(C)x are
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
[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
8-22
x{i
y=[
1
-1
Consider a discrete-t
8-28
Consider a discrete-
'J x
Is the zero state asymptotically stable? Is the zero-state response BIBO stable? Is the
system T-stable?
8-23
8-27
8-29
Consider a system \
8-24
Prov~
8-25
~~-1
9.
1-] inL.
U:)(; f/rr
:==
and
X =Ax +Bu
Show that if u =
W(T)=
e-A'BB*eAo'dr
Is the function
then the overall system is asymptotically stable. Furthermore, V(x(c)) = x*(e)W- '(T)x(e)
8-26
Are the networks shown in Figure P8-26 totally stable? AIJSlrers: No; yeso
PROBLEMS
429
uation:
)-input response).
btaned by the equivalence
Figure P8-26
8-27
y(n) =
g(n, m)u(m)
Show that any bounded-input sequenee {u(n)} excites a bounded-output sequenee {y(n)}
if and only if
for all n
m=-oo
8-28
>n
description:
g(n - m)u(m)
m=O
Show that any ~ounded~input sequence {u(n)} ~xeites a bounded-output sequenee {y(n)}
if and only if
Ig(m)1 ~k <
00
m=O
8-29
C/o(l -T)
i=O
19(I)1 dI
~kl < 00
ltion
and
la,J~k2 <
00
i=O
Prove Corollary 8-3by usng parta\ fraetion expansion for the class of systems that
have proper rationa\ transfercfunetion deseriptions.
8-30
;itive number
8-31
Is the funetion
[Xl
AnslI'ers: No; yeso
Xl
X 3]
[~ ~
2 2
n6J r~0
G~
430
8-32 Which of the following hermitian (symmetric) matrices are positive definite or
positive semidefinite?
a.
[ ~]
[! ~
3
c.
O
O
O
b.
d.
[",a,
aloz
a l a3
~]
8-38
and W = diag [W I , W 2 ]
W 2 ha ve no diagona [ er
alaz
ozaz aZ a 3
aZ a 3 a3 a 3
",a:l
--'
8-33
af
2A 1
alaz
---}'I +},z
azal
M= - - - },z + Al
a3 a l
---Al + A3
ala3
---Al + )'3
aZ a 3
a2
2A 2
---
Az }'3
a3 a Z
---},z + }'3
a 3Z
and the matrix A in (8
Verify that Ac=Q-1AQ
21 3
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
8-34
(Due to Y. P. Ham)
Consider tbe systc
Eill i .
Use lb
scription of the system \'
is i.he TuHr\j; A, of tb..i.s r:
v\lhmls LOe eondlLlon for
that lhe system be BIBO
8-40
y= C;x+
11.
Let e, i = !, 2..
.11,
,,~.
8-36
8-37
x(k
+ 1):=
O
O
I
O
0.008
O
0.008
be a set of linearly
Li '
.:~ i:;:;';;;
. O
.O
l ' x(k)
O
-0.79 -0.8 .
(1) by computing itsch~racterislic polynomial and the'n' ap'plyingTheo;:em 8"21 and (2)
by solving.the Lyapunov equation in Theorem 8-22.
Figure P8-40
PROBLEMS
8-38
431
[ x~lJ
[Al!
A 22 _ x2
A 21
y = [CI
C2]
82
[::J
and W = diag {W [, W 2}' Show that if the eguation is asymptotically stable and ifW 1 and
W 2 have no diagonal entries in common. then the subsystems
Xi =Ax +8u
y=Cx
See
~~[~
'3
'3
-a
ll
-~J
O
1
O
O
-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 ]
O
O
ag,-l)
aY,:-I)
ag')
(Due to Y. P. Harn.)
Consider the system shown in Figure P8-40, where S is described by Xi = Xi+ Ru,
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.
15 the matrix of this descriplion the salTle ':\s the ane in the tjm~-inwlrajll. c,:,.se of 1}.fA\')
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?
8-40
y = c,x + Eu.
. ,11,
laton
f[
.u
+
~
(k)
YI
SI
I
u2
Y2
.. f2
S2
Figure P8-40
(al
9
Linear Time-I nvariant
Composite Systems:
Characterization, Stability,
and Designs
9-1
Introduction
This terminology is by no means universal. Another possible name is the Luenberger-- or state
estimator-type configuration_
432
Figure 9-1
(a) Unity
INTRODUCTlON
433
Ca)
Cb)
Figure 9-1
434
COMPLETE (
with state vectors Xi, i = 1, 2. Then the state vector of any connection of SI
and S2 will be defined as x' = [X'1 x~]. With this definition, there will be no
confusion in the state-variable description of any composite system. The
state-variable descriptions of the tandem, parallel, and feedback connections
of two systems are derived in (3-62) to (3-64).
All the subsystems that form a composite system will be assumed to be
completely characterized by their transfer-function matrices. This assumption,
however, does nat imply that a composite system is completely characterized
by its composite transfer-function matrix. In Section 9-2 we study the con
ditions of complete characterization of composite systems. For single-variable
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
nections, then the composite system is completely characterized by its transfer
function. In Section 9-3 we extend the results of Section 9-2 to the multi
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
design problem. In Section 9-5 we study the design of compensators in the
unity feedback system to achieve arbitrary pole placement and arbitrary
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
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
general than those obtained in the state-variable approach discussed in
Chapter 7.
The references for this chapter are S2, Sil, S19, S34, S35, S4, S49 to S51, S54,
S55,S64toS66, S75,S81,S85,S93, S94,S98,SI74,SI85, SI99,S218,S237,and
S23?,
Composite Systems
(9-1 a)
(9-1b)
where Xi, U, and y are, respectively,the state, the input, and the output of the
system Si. A, B, C,.and E are real constant matrices. The transfer-function
matrix of S is
(9-2)
It is assumed that th
transfer-function m[
equations (9-1) are c
that the transfer-fun
(;1(S) +G 2 (s); the tI
followed by S2 is (
connection of SI W'
([ + Gl (S)G 2 (S))-IGi
and
Their transfer functio
It is easy to check
observable; hence frOl
charaderized by its ce
In the. following
transfer functions co
problem directly fre
dynamical equations.
function matrix, then
of the systemis equal t
my connection of SI
tion, there will be no
lposite system. The
eedback connections
ill be assumed to be
es. This assumption,
lpletely characterized
-2 we study the con
. For single-variable
on pole in the parallel
nd the feedback con
terized by its transfer
.on 9-2 to the multi
tion replaced by the
stability problem of
1 terms of the transfer
lter is devoted to the
compensators in the
ement and arbitrary
and that of arbitrary
They are, however,
uired for the latter is
;r. In Section 9-6 we
:king and disturbance
sed; both robust and
le study the design of
fhe results are more
,proach discussed in
5, S40, S49 to S51, S54,
SI99,S218,S237,and
435
lt is assumed that the systel'Ils SI and SJ. are completely characterized by their
transfer-function matrices G l (s) and G 2 (s); 01', equivalently, the dynamical
eq uations (9-1) are controllable and observable. It was shown in Section 3-6
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
followed by S2 is 6 2(5)6 1(5); the transfer-function matrix of the feedback
connection of SI with S2 in the feedback path is 6 1(5)(1+6 2 (5)(;1(5))-1=
(1 +6 1 (s)6 2(s))-1(;I(S). Although (;1(S) and (;2(S) completely characterize
the systems SI and S 2, respectively, it does not follow that a composite tnmsfer
function G(s) completely ~haracterizes a composite system.
Example 1
and
FE1:
Xl =X l +Ul
FE!:
Yl =X 1 +Ul
X2=X2- U2
Y2 =X2
-1
s-
and
s-1
92(S) =-1
8-1
-1
+- = 1
s-1
It is clear that g(5) = 1 does not characterize completely the composite system,
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
nection, we have u 1 = U 2 = u and y = y 1 + y 2; hence the composite d ynamical
equation is
y=[1
Variable
equation descriptions
lJx +u
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
characterized by its composite transfer function.
I
(9-1 a )
(9-1b)
436
COMPLETE (
Consider two systems SI and S2, which are completely characterized by their
proper transfer-function matrices Gl (s) and G2(s), respectively. Any composite
connection of SI and S2 is completely characterized by its composite transfer
function matrix G(s) if and only if
il
....
15 91 +(592. We pr
there is at least o
there is a comm(
assumption that ti
there is a common
tion that 91 is irre(
in common, then
connection of S 1 a
2. The proof of this 1
3. The transfer funct
By the irreducibili
have D 2 and N 2.
if and only if D 2 al
Example 2
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).
We assume thal there are no capacitors-only loops and inductors-only cutsets in the network.
Figure 9-2
A single-vi
437
g=gl+g2=
N I D2 +N 2 D I
.
DI D2
We show now that if 91 and g2 do not have any pole in common, then lJg =
We prove this by contradiction. Suppose 8g <8g 1 +8g 2, then
there is at least one common factor between N ID 2 +N 2D I and D ID 2. If
there is a common factor, say, between N ID 2 + N 2DI and DI' then the
assumption that there is no common factor between DI and D 2 implies that
there is a common factor between NI and DI' This contradicts the assump
tion that 91 is irreducible. Hence we conclude that if g 1 and g2 have no pole
in common, then 9(5) = 91(5) +92(5) characterizes completely the paral1el
connection of SI and 52
2. The prQofof this part is obvious and is omitted.
3. The transfer function of the feedback system shown in Figure 9-2 is
lJ91 +lJ92'
haracterized by their
vely. Any composite
s composite transferi
~quation
descriptions
ce of any composite
of the state spaces of
: dynamical equation
lescriptions of SI and
;)-1 and Theorem 6-2.
: the transfer-function
ticularly in the design
as of G1and (;,Z for G
of SI and S2 will be
systems. The multiscalar, and its degree
)n is irreducible. We
:ducible; that is, their
D 1 D 2 +N IN 2
By the irreducibility assumption, DI and NI have no common factor, nor
have D z and N 2 Hence D 2 N] and D I D 2 +NIN z have common factors
if and only if D 2 and NI have common factors.
Q.E.D.
Example 2
1
-1
and
re completely charac
ldg z (5).
y cutsets in he netwof.k.
5-1
s + 1
Figure 9-2
9z
438
ce
(a)
Consider a spec
= k, where k is a
by gI (s), the transfer I
g2(S)
(b)
Figure 9-3 Tandem connection of 5 1and 52' which s not characterized completely by
=9291 = 1/(5 + 1).
g(5)
always characterizes
S+
does not completely characterize the tandem connection. This can be seen by
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
III
time, as shown in Figure 9-3(b).
Example 3
Consider the feedback connections shown in Figure 9-4. In Figure 9-4(a), the
pole of the transfer function in the feedback path is canceled by the zero of the
transfer function in the forward path. Hence the transfer function of the
feedback system does not completely describe the feedback system. Indeed,
its transfer function is
s-1
s+l
,')-1
g(5) = - - - -
s-1 1
s+2
1 +---
s+ls-1
Its degree is smaller than 2.
(a)
Figure 9-4
Feedback systems.
(b)
9-3 Controllabil
of Composite Sy
Let gj(s) = N(s)D j- I (s)
N j and Di, Theorem (
SI and S2 is completE
D 2 (s) are coprime. l
g2(S)g(s) if and only
coprime. In this secti
composite systems. B
6-8, on strict system ec
A system can be de
matrix in fractional fo
..
. _---_ .. _----
_,._~
-------~._-~-_
..
.. ---
_-~-~-_
439
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
s +1
g(s) = (s _ 1)(s +2)
A
-l
lIlI
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
by (s), the transfer function
A
[
A ] A
g(s)=
1 +kg[(s)
- g[(s)=
acterized completely by
N[k
D[ + N[
pes) Q(s)
S(s)l
L-R(s) w(stlL -i(s)J =
r [b!
L-y(sU
where u and y are the input and output and ~ is the pseudostate. If these
descriptions describe the same system, they all have the same transfer matrix.
If {A, 8, e} is not irreducible (oot controllable or not observable) or if {D" Nr },
{D" NI}, and {pes), Q(s), R(s)} are oot coprime, then they are generally not
strictly system equivalent. Howevcr, if they are all irreducible, they are strictly
system equivalent, and any one of them can be used in the analysis and designo
For our problem, it turnsout that the use of the system matrix is the most
convenient. Hence the system matrix will be used extensively inthis section.
We recall from (6-151) aod (6-152) that {A,B} is controllable ifand only ir
{P(s), Q(s)} is left coprime; {A, C} is observable if aod only if {pes), R(s)} is
fight coprime. These properties will be used to establish the conditions for
complete characterization for the multivariable case.
440
Let G(S) = Nr(s)Dri (S) = D (s)N(s) be, respectively, right and left co
prime fraction of (;(s). Then system Si can also be described by
1][ - t(S)]
[
O]
u(s) - -Yi(S)
D,.(s)
[ - N,.(s) O
i = 1,2
Di(S)
[ -1
(9-3)
1, and qi x 1 vectors, or
i = 1,2
(9-4 )
By a similar argument i
if and only if D l1 (s) a~(
results as a theorem.
l-N,.(s)
-N..z(s):O
-u(s)
= [ __ __ ]
-y(s)
(9-5)
This transformation merely substracts the second block row of the system
matrix in (9-5) from the first block row. Since controllability and coprimeness
are invariant under the transformation of strict system equivalence (Theorem
6-9), we conclude that the dynamical equation description of the parallel
connection is controllable if and only if
lJi)r\(S)
O..z(s) : 1
or
(9-1Oi )
Theorem 9-3
Consider lwo systems 1
matrices G(s) and G2
fractions of G(s). Ther
trollable if and only if 1
nection is observable if a
~
Tandem connection.
sb.oyvn. in Figu.re 9(; i;;I~ !~.
using these relations and I
has a full rank for every s in e Because of the block triangular form, the
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',
O"o(s)
- N,.(s)
---------=
SIZ
,-.,------------
.'UI.
I .
'YI
s,
IL
Figure 9-5
Figure gc6
Y;
Uz
S,
O"I(S) and 0"2(S) are left coprime. Hence we conc1ude that the paraHel con
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
(9-3)
1,2
OII(S)
(9-4)
md q x 1 vectors.
lown in Figure 9-5, we
'e u and y denote the
quations and (9-3), we
j,l]
O
-1
-1:
(9-7 )
(9-5)
irreducible by assump
ible. In order to obtain
;tem equivalence trans
Or[(S)
O:
NII(S)]
012(S) : N I2 (S)
[---------- -- -- -
and q x 1 vector s, or
i = 1, 2
441
0J
-O'2(S):
O,.z(s) : 1
=--:I(s) -- ~-'~2(;) : 0
~~2_(~)_ ~ ?-I
lJ
(9-6)
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
[Orl(s) 0"2(S)] and the matrix [Ol(S) D;2(S)]' have a full rank for every s in t:
(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, hwever, is not a necessary condition (see Problem 9-4).
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
only if ther denominators have no roots in common. This provides a different
proof of statement 1 ofTheorem 9-2. By combining the conditions in Theorem
9-3, we have the necessary and sufficient conditions for (;\(s) + (;2(S) to charac
terize completely the parallel connection of the two systems.
H).
Tandem connection.
D'2{S) :
ock triangular form, the
and only if the matrix
following Theorem 0-8/,
(9-8)
Sl2
r----------~--l
I
IU
IL
Figure 9-6
Y2Ly
--J
442
ce
to describe the tandem connection. Note that Yl(S)=U 2 (s) is a part of the
pseudostate of the tandem connection. It is clear that the tandem connection
is controllable if and only ir. for every s in e, the matrix
o
(9-9)
O'2(S)
Corollary 9-4
O: O IJ
O,,(s)
D'2(S) : 1
A sufficient condition
O,-z(.~)'
O
O O
[ ------~-----~-~----
- Nrt (s)
G2~S) is a transmissiol
of G 2 (s)].
which implies that the matrix in (9-9) has a full rank for every s in e if and only
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,
fol1owing Theorem G-8', left coprime.
If we use (9-3) to describe SI and (9-4) to describe 52, then the tandem con
nection is described by
Ijr ~(S)j
~ ~'-: ~s~ _~I~~S) __ ~~t~ j_~ _-J:~:~ =
O,(s)
O:
O: O . - u(s)
lt _
O
__
Proof
Let (;l(S) = Nr(s)D,:
polynomial matrix ane
P2 at evelY s except the
zero of G (s), we have
matrix [D,.2(s) N,.(sf
coprime. Hence if no
connection of S .follolA
observability part can t
(9-10)
- y~s)
If we add the product of the second block row and N l2 to the third block row,3
then the system matrix in (9-10) becomes
Orl(s)
-NrI(s)
- N I2 (S)N'.I(S)
O
O
D I2 (S)
The combination (
Theorem 9-4 yields the
tandem connection. 1
Theorem 9-4 reduces te
is controllable (observaI
of g(s) [no pole of g(
characterizes complete!
0:11
-1:0
o:
O
Theorem 9-4
Consider two systems Si which are completeIy characterized by their transfer
matrices G;(s), i = 1, 2. Let Gi(s) = Di 1 (s)N(s) = N,i(S)D,: l(S) be coprime
fractions of G;(s). Then the tandem connection of SI fol1owed by 52 is con
trol1able if and only if any one of the following three pairs of polynomial
matrices, 0ds) and NrI(s), 0 11 (slDds) and N l1 (s), or O/2(s) and Nz(s)N'I(S), are
left coprime. The tandem connection is observable if and onfy if apy one of the
) This is a-lransformation of slriclsyslem equivalence_ See lhe slalemenl al lhe end ofSeclion 6-8.
I
I !
(9-9)
443
following three pairs of polynomial matrices, D 11 (S) and Nds), D/[(s)D.. 2 (s)
and Nds), or D,.(s) and Nds)N-[(s), are right coprime.
I
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
called the poles of (;(s) and those s for whch pN.. (s) < min (p. q) or pI"'Us) <
mn (p, qJ, where p stands for the rank, are called the tral1smissiol1 zeros of
(;(s) (see Appendix H).
Corollary 9-4
Jl-~y(s)i 1J
(9-10)
";'-".'
<'5 (;(s) = <'5(G 2 (s)G I(S)) < <'5.G 2(s) + <'5 G l (s)
In this case, we may define that there are pol-zera cancellations in G 2 (S)G 1 (s).
For example, if
. 'l'
~s-l
j'
s
G 1 (o5)= __
.
emenl al lhe end o[ Section 6-8.
05-1
444
Cal
then b(G 2(s)G I (S))=O < bG 2(s)+ bGI(s)= 1 and there is one pole-zero cancel
lation in G 2(S)G I(s). Clearly, the existence of pole-zero cancellations in
G 2(s)G 1(s) does not imply the existence of pole-zero cancellations in GI (s)G 2(s),
as can be verified from the example. Unlike the scalar case, pole-zero cancel
lations in the multivariable case may not involve actual cancellations of poles
and zeros.
Let ~(s) and ~i(S) be, respectively, the characteristic polynomials of G(s)
and G(s). Ir G(s) = G 2(S)G I (s) has no pole-zero cancellation, then we have
~(s)= ~1(S)~2(S), as can be easily seen from Equation (3-63a). Ir (;2(S)(;I(S)
has pole-zero cancellations, then deg ~(s) < deg ~ 1(s) + deg ~2(S), and the roots
of ~1(S)~2(S)/~(S) are called the cancelled poles. If A is a cancelled pole, then
from Theorem 9-4, we have the following equivalent conditions:
1. [DdA)
N"l(A)] and/or
2. [Dll(A)DdA)
3. [DdA)
N1I(A)] and/or [
D ll V,)D,z(A)]
Nd},)
have no ful1 rank.
For a direct proof of this statement, see Reference S2. See also the discussion
of the input-decoupling zero on page 292. If al1 cancelled poles have negative
real parts, then G 2(S)G I (s) is said to have no unstable pole~zero cancellation.
Feedback connection. Consider the feedback connection of SI and S2
shown in Figure 9-7. It is assumed that S is completely characterized by e(s).
It is also assumed that det (1 +
(s)Gz(s)) =1= oat s = <Xl to ensure that the feed
back transfer matrix is wel1 defined and proper (Theorem 3-5). Let S 17. denote
the tandem connection of SI followed by S2; $21, the tandem connection of
S2 followed by SI; and SJ' the feedback syslem in Figure 9-7.
el
Theorem 9-5
Consider two systems Si which are completely characterized by their transfer
matrices (;(s), i = 1, 2. It is assumed that det (1 + GI( <Xl 2(00)) 1= O. Then
the feedback system S J is controllable (observable) if and only if S 12 is con
trollable (S21 is observable).
)e
ut
Proof
We shall prove this ti'
ability. Let x be the e
Xi is the state of Si' :
composite state space
time. Let Y2 be the Oi
S J, this input will tran
is SJ' Conversely, if
uI=u-Y HenceSJ;
Now we consider
S J due to sorne unkno
if - Yo is applied to S
S21 is Yo If S21 is ob
output Yo, the initial sl
able. lfS 21 isnotobs~
output Yo. Consequer
its zero-input response
this theorem.
In general, the COI
condition of observab
two conditions are ider
of 91(S), then 91(s) follo
is observable. Hence \\o
connection in Figure <;
controllable or, equiva
This checks with what '
We consider a speci:
le We cal1 this a consl,
any new state variable,
observable, so is K fol
\Nith a constant gain :,
ano only ii SI is con1
observability are invari,
with the constant state
is preserved under com
preserved.
We discuss further !
be described by irred uc
r
.
,A i , Bi C, E;}, where G
rationa,1 matrices. By a
and
Figure 9-7
~[G(sr
deg G(s)"" ~
e polynomials of G(s)
:l1ation, then we have
(3-63a). Ir Gz(s)G I(s)
eg i!2(S), and the roots
a cancelled pole, then
lditions:
full rank.
445
Proof
We shal1 prove this theorem from the definition of controllability and observ
ability. Let" be t he composite state of SI and S2; that is, X = [X'I x~J', where
Xi is the state of Sj. If S12 is controllable, then for any Xo and any XI in the
composite state space, an input u I to SI2 exists to transfer Xo to Xl in a finite
time. LetY2 betheoutputofS 12 duetou 1. Nowifwechooseu=u l +Y2for
S f' this input will transfer Xo to x l' This proves that if 5 12 is controllable, so
is Sf' Conversely, if Sf is control1able, corresponding to any u, we choose
"1
y. Hence 5 12 is controllable.
Now we consider the observability part. Let the zero-input response of
Sf due to sorne unknown state "o be Yo. If S21 is in the same initial state and
if - Yo is applied to S21, then from Figure 9-7, we can see that the output of
S21 is Yo. If 5 21 is observable, from the knowledge of the input - Yo and the
output Yo, the initial state Xo can be determined. Consequently, S f is observ
able. If S21 is not observable, we cannot determine Xo from its input -Yo and
output Yo. Consequently, we cannot determine the initial state "o of 5f from
its zero-input response, and Sf is not observable. This completes the proof of
this theorem.
Q.E. D.
=" -
) full rank.
See also the discussion
ed poles have negative
.Ie-zero cancellation.
mection of SI and S2
characterized by (;(s).
to ensure that the feed
n 3-5). Let SI? denote
tandem connection of
Ire9-7.
446
COI
transfer matrix ofthe feedback system in Figure 9-7 is, as derived in (3-68),
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)
__ [A-BY 2 E 2 C
B 2 YC
-BY2C2
]
A 2 -B 2YEC 2
( 9-12)
and
where
denotes not equal modulo a nonzero constant factor. Conversely, if
the equalities hold, then the (n + n2)-dimensional dynamical equation de
scription of the feedback system in Figure 9-7 is controllable and observable.
Note that we have
L\.[ GJ(s)]
!.....;; I ~:.
~ (S)J
[ uz(s)
By definition,
H(s) [:
I(S)J
r z(s)
. [.
H(s)=
. 1I2
Figure 9-8
fl
hence, we have
r..,
~1
- G (s)
=[D,o~(s)
which can be easily veri
447
; derived in (3-68),
hence, we have
(9-11)
The state equation of the system in Figure 9-8 can be readily computed as
ved in (3-64),
[ ~IJz
(9-12 )
[Al +[
B I Y ~EzCI
BzYIC[
[Xl]
Xz
J[r[J
BY z
- B I Y zC, ]
Az-BzYIEICz
-B 1 Y zE z
BzY\ _ r2
BzY,E t
Consider the system in Figure 9-8 wilh lhe assumpl ion det [1 + G[( ex:;. )G 2 ( ex:;. JJ =F
O. Then we have
6(H(s)) ~ del [D/[(s)Dds) + N Il (s)N,z(sIJ
~ del [Ods)D,[(s) + Nz(s)N,.(s)J
rl(s)]
~ det [
ducible.
ends on the assignment
.lre 9-7 may be uncon
ut and output as shown
"ument in the proof of
~l1able and observable
;imilarly, the system is
'1 r'2]' and the outpul
1[r'l r'z]' t o [u' u'z]'.
l=r z(.I)+YI(s)=r 2 (s)+
~ det [
'1(.1)
- N,.(s)
NdS)]
ds)
N/z(s)]
D/1(s)
(9-13a)
(9-13b)
(9-13c)
Dds)
- NIl(s)
(9-13d)
det (.1'1- A j )
~ 6[ G 1 (s)J6[ G z(s)J del [1
where
factor.
(9-13e)
(9-13f)
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
H(s) = [
. -G(s)
[
='
0 .. [(.1)
O
(;2(S)J-['=[.
l.
0][ ~
Ods)
1_[
-N,.[(s)O"I(s)
NdS)O,:;I(S)J- [
.
o,[(s).N,.z(S)]-1
N,. (s). Uds)
which can be easily verified. The fraction is right coprime as canbeseen from,
tpul.
448
O"I(S)
-N'I(S)
0"1 (S)
S
Nd )]
O
Ods) -+ -Nr(s)
O
Or (S)
Ods)
O
1
- Oil 1(s)N l1 (s)
N,.z(S)O,:-/(S)]-1
1 O][1
- [O
O,.z(s)
Thus, we
- N l1 (s)
NdS ) ]
0 11 (s)O,z(s)
[1 O]
O O'I(S)
~(H(s))=det [ -N (s)
l1
N,z(S)]
]
DI 1(s)Ods) =det[Ol1(s)O,z(s)+N l1 (s)N,z(s)
Q.E.D.
9-4
Single-variable fee
back system shown in F
9f (s) =1-+
where 9(s) ~ Ni(S)/Di(~
Recall that every transf
we speak of 9f(S)' we n
that if NI and D 2 havt
syst~m js completely el
tween N and D 2 , ther
system is not completel
Theorem 9-7
Consider the feedback ~
52 arecompletely char
'S)l
(s)
:)[em G-8).
Thus, we
).
ence wehave
[- A f) follows directly
"2, J'.
Ods)}
O"2(S)
Q.E.D.
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
asymptotically stable, then the system is also BIBO stable, but not conversely.
However, if a system is completely characterized by its transfer matrix, then
asymptotic stability implies and is implied by BIBO stability.
Consider two linear time-invariant systems SI and 52 which are completely
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
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.
For the feedback system, the situation is much more complicated. A
feedback system may be stable with unstable subsystems; conversely, a feed
back system may be unstable with stable subsystems. We study first the single
variable case and then the multivariable case.
Single-variable feedback systems. Consider the single-variable feed
back system shown in Figure 9-9. The transfer function ofthe feedback system is
g (s) =
ng det(sI - A f ) and is
there is no distinction
;:s to
(s)
449
: ~ (:;
1 +g,(S)g2(S)
:fore it is important to
~. recall from Chapter 8
;e of a system is asymp-
Consider the feedback system shown in Figure 9-9. lt is assumed that S, and
52 are completely characterized by their proper transfer functions g,(s) and
450
Figure 9-9
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
asymptotically stable
A feedback system.
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
system, where X is the state vector of Si, and let x = Afx + bfu, y = efx +efU be
the dynamical equation describing the feedback system [see Equation (3-64)].
Then the characteristic polynomial, det (s 1 - A f), of Af is a polynomial of
degree n +nz, where degD(s) =n, i = 1,2.
Nowwe show that ifl +g(<Xl )g 2(<Xl )7'=0, then deg [D(s)Dz(S) + N (s)N2(S)] =
n +n2' In other words, the term sn, -r n 2 in D(s)D 2(s) will not be canceled by
any term in N(s)N 2(s). Let
b sn +b sn- + ...
g.(s) = 'o
I
1
aOsn, +an sn + ...
Then we have
[1 +A ()A (')]-1 +bob 20 _QOQ20+ b lO b 20
l 1m
91 s 92 S
,~O()
QOQ20
QOQ20
and
Corollary 9-7
(9-14 )
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
the relation
g(s)=c(sI-Af)-bf+ef=d
NI (s)Dz(S)
we conclude that
det (sI - A f ) =k(D"(s)D 2(s) + N(s)Nz(s))
Multivariable feed
case is much more ca
Consider the multivari
S is assumed to be COI
G(s). The assumptioJ
connection. Let Gf(S
u
(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,
451
system is asymptotically stable if and only if all the roots of Dls)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 BIB stable.
Consequently, the system in Figure 9-9 can be BIB stable without being
asymptotically stable.
Q.E.D.
,) = N (s)/D (s). Then
Iy if all the roots of
dirion is sufficient but
Corollary 9-7
(9-14 )
Q2
b 2 0s" +"2 +
s
g(00)g2(00) 1=0, then
). Using this fact and
sI - A) bf
+ e
Consider the feedback system shown in Figure 9-9 with {](s) = k. lt is assumed
that S is completely characterized by its proper transrer runction yI (.1') and
I +kg(oo)+O. Then the feedback system is BIB stable and asymptotically
stable ir and only if all zeros of 1 +kg(s) or, equivalently, all roots of D,(s)+
J,,; t'll (.;-;) h8_V~ nega tiv~ ~~2.1 :-::'.:;"X ::::..
Multivariable feedback systems. The situation in the multivariable
case is much more complex, hence we study first the simplest possible case.
Consider the multivariable feedback system shown in Figure 9-10. The system
~ is assumed to be completel)' characterized by its p x p proper rational matrix
G(s). The assumption that G(s) is square is necessary for the proper feedback
connection. Let G(s) be the transfer matrix of the feedback system. Then
SI
N 2(S)) , .
(9-16)
~r
Figure 9-10
452
First we note
det [1+ G1 (<Xl)J:b (
L1 1 (s)det [1+ (; I (s)] i
we have
Example 1
Consider the multivariable feedback system shown in Figure 9-10 with
-s
"
G(s)
lt can be easily verified that
l
l
s-1
1
2S
"
G(s)
== s +1.
-1
_s_l
s +1
-2
s +1
s~Sl
J
s-2
s-l
which is unstable. However, we have det [1 +G(s)J = -1, which does not
have any right-half-plane zeros. Hence the stability of the feedback system
cannot be determined solely from the zeros of det [1 + G (s)].
!!l
In order to study the stability of multivariable feedback systems, the concept
of the characteristic polynomial of a proper rational matrix ~ needed. Recall
f~_
"o':; ".' .
;: 1 thn" t'D~ r.hnrnnro-'si'r rnj .. ..r_r.' .. ;
.r. r ..' ,:,~,,-,_
itvl
L ...".,nfoon 0:H
le ..... i c..~a.''-'L'..... ! >'i" !?V~Y,L:'Jl':'''..i.c.;.: h c ';':"':;1\~1):,
. j~.' .' ' .
,1(S), is the least common denominator of al! the minors of G(s). lf G(s) has
the coprime fraction N(s)D(s), then we also have 6.(s)=kdetD(s), for
some nonzero constant k.
.!:, ...
'p
Proof of Theorem 9
We shal! use the dyn.
first part of the theore
.J'
Theorem 9-8
Consider the feedback system shown in Figure 9-10. It is assumed that S
is completely characterized by its proper rational matrix G(s) and that
det [1 + (; (00)] :b O. Let (; (s) = N,. (s)D,: (s) = Dl~(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
any of the three polynomials 6.(s)det[I+G(s)],det[,.;(s)+N rl (s)J and
det [D1(s) + N(s)] have negative real parts.
I
be
tne dynamical-eq ua
453
is completely charac
e feedback system in
order of el (s) reduces
stability depends only
ted to assert that the
)nly on the zeros of
(9-18)
F(s)
To show that ~[(s) det [1 + e[ (s)J is a polynomial is the same as showing that
is divisible without remainder by F(s). This will be shown by using the
following identity:
~[(s)
det [1 +G[(s)J =
(X(s) +1
(9-19)
i; 1
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.
III
1t is assumed that S
natrix G (s) and that
IN l (s) be coprime frac
:s). Then the feedback
only if all the roots of
:t[D, (s) + Nrt(s)J and
x = Ax + 81I
y =Cx + E"
(9-20a)
(9-20b)
lhal
(9-21 )
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
controllable and observable. Consequently, the characteristic polynomial
~ (s) of e 1(s) is equal to the characteristic polynomial of A; that is,
~ (s) ~ ~[G(s)J = det(sl - A;) .
(9-22)
"-.
454
X =[A-BI(I+E)-lCJx +BI(I+E)-lu
y = (1+ E1r IClx + (1 + E)- IElu
(9-23a)
(9-23b)
(9-24 )
Example 2
The characteristic pI
verify that det [1 + G
only if aH the roots o
parts. This is not th
nor asymptotically st
lt is 01' interest to
(9-25)
lt is shown in Theorem 3-2 that det [1 + G(s)G 2(S)] = det [1 + (; 2(S)(; I(s)J and,
since (.~) =det (sI - At!. Equaton (9-25) implies that
ls)=ls)det[1 +CI(sl-A}-IBI(1 +E}-IJ
=ls)det [[(t +E) +C(sl -A)-IBIJ(I +E}-I}
=I(s)det [1 + (;(s)J det (1 + E)-
(9-26)
Since det (1 + E) - = det [1 + Gl co)] - I =FO. we conclude from (9-26) that the
set of the eigenvalues of Al - BllI + E I )- le l is eguaI to the set 01' the zeros 01'
lls) det [1 + G(s)]. This proves the first part 01' the theorem.
Now we show the second part. Let G{s)= N,. I(s)D':-1 I (s). Then we have
(;(s) =
=
G1(.1')[1 + (; 1(.1')] - I = N.. (s) 0':-1 I (s)[1 + N"I(S)O,~ I (s)] N,. (s)D,:j I(S){ [D"I (s) + N .. I (s)]D,:j I (s)] - 1
Example 3
(9-27)
= Nr(s)[D,.(s)+N'I(s)]-1
We show that if D"1 (s) and Nr (s) are right coprime.. so are iD,. l.,.. -: ]'>1 i ,) ?n-j
NI[ (s). The coprimeness of D q (s) and N.. I(s) imples the existence 01' polynomia
matrices X(s) and Y(s) such that
X(s)Dr (s) + Y(s)N,. (s) = I
It is straightforward t,
+ [Y(s) -
X(s)JNr (s);' I
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 [ Gf(S)J~ det [D,.(s) +
. N,.(s)] ~ (s). Hence the feedback system is BIB and asymptotically stable
if and only if all the roots 01' detCD,. (s) + N"I (s)] have negative real parts.
Q.E.D.
m can be obtained as
(9-23a)
(9-23b)
Example 2
(9-24 )
GI(s)=
[1 +
-s
s-1
s
5+1
-2
s +1
The characteristic polynomial of G(s) is o(s)=(s-l)(s +1). It is easv to
verify that det [1 + (;(s)] = -1. Hence the feedback system is stable if'and
on1y if al1 the roots of o (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.
~
lt is of interest to give sorne interpretation of the roots of the p01ynomial
~comes
E1)-IC1J
l
l
le fol1owing Theorem
) and asymptotical1y
)-le l have negative
f the eigenvalues of
o(s)det [1 + (;1(5l].
455
E(s)
F(s)
(9-25)
(9-26)
where E(s) and F(s) have no common factor. The roots of E(s) are poles of the
feedback system introduced by feedback, whereas the roots of o 1(s)j F(s) are
poles which are possessed by the open-loop system S las well as by the feedback
system. For example, in Example 1 we have E(s) = - 1,F(s) = 1, o(s)jF(s) =
'1 + 1)(s -1); hence no new pole is introduccd in the feedback system G(s);
G (s) and G 1(s) have the same poles 1 and -1.
Roughly speaking, the polynomial o1(s)/ F(s) takes care of the missing zeros
of det [1 + (;I(S)]. Therefore one rnight suggest that if we do not cancel out
any cornmon factors in det [1 +6(s)J, then the stability ofthe feedback system
might be determinable from the zeros of det [1 + 6 1(s)]. This is however not
necessarily correct, as can be seen from the fo\lowing example.
Example 3
(9-27)
s-0.5
1
s -0.5
)=1
(s + 0.S)(S2 + s + 0.25)
(s+ 1.5)(s-0.5)2
s+0.5
456
det G 1(s) =
and conseguently,
(s
-1
det (.
=---:-
if and only if
G1(s) is
Since we have E(s)/ F(s) = (s +0.5)/(s - 0.5) and L1 1(s)E(s)/ F(s) = (s + 1.5)(s +0.5),
the feedback system is asymptotica11y stable and BIBO stable.
111
We consider now the case where a dynamical system appears in the feedback
path as shown in Figure 9-11. This feedback system is generally not completely
characterized by its composite transfer matrix; hence the situation here is more
complicated than the unity feedback system shown in Figure 9-10.
of (9-29) is independ
characterized by
G(s) =
The factorization in (9
L1[G(s)]
Theorem 9-9
Consider the feedback system shown in Figure 9-11. It is assumed that S 1 and
~z are com,Pletely characterized by their q x p an~p x q proper transfer m~trices
G 1(s) and Gz(s). It is also assumed that det [1 + G 1(00 )G z(00)] O. Let G(s) =
N..(s)D,: l(S)= Di 1(s)Nli(s), i = 1,2, be coprime fractions and L1,(s) be the
characteristic polynomial of (s). Then the feedback system is!symytotically
stable if and only if a11 the roots of L1 1(s)L1 z(s)det[I+G 1(s)G Z (s)], or
DI 1(s)Drz(s) + N 11 (S)N,z(s), or any other of (9-13) have negative real parts. The
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
system to be BIBO stable.
1
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),
L1 (s)
f
~ det (si -
Figure 9~11
["l(S)J=
"z(s)
By direct computatior
H(s) =
referred to Rererence 49. Similar to the proof in (9-14) and (9-15), it is possible
to establish, by assuming the column reducedness or Drz(s) and the row re
d ucedness of 0 11 (s), that
deg det [Dll(s)D,.z(s) + N 11 (S)N.. z(s)]
(9-28)
det [1 + G1(oo)Gzioo)]
n1
+ nz
[~"
-G 1 (s)
Gz(
[1" + G2 (s)d
[ l(s)[i p +G z(
457
and consequently.
(9-29 )
-0.5)
if and only if
det [1 + GI(co)GAcoJ]
4'(.1') = (s
To save space, this assertion wil! not be proved here. We note that the validity
of (9-29) is independent of whether 01' not the feedback systern is completely
characterized by
+1.5)(s +0.5),
~able.
11
(9-28)
co )(;z(co)]
he interested reader lS
nd (9-15), it is possible
>rz(s) and the row re-
(9-30)
+O
(9-31 )
For this reason, the condition in Theorem 9-9 is only sufficient for G(s) 01' the
system in Figure 9-11 to be BIBO stable. Ir Gz(s) = K, a p x q constant matrix,
the feedback system is, as discussed in the previous section, completely charac
terized by G(s) and ~[G(s)] ~det [DII(s)+ N(s)KJ; hence the condition in
Theorem 9-9 becomes necessary as wel! for G(s) to be BIBO stable.
Now we consider the same system in Figure 9-11 but with additional input
and Olltput as shown in Figure 9-8. As derived in (9-12), we have
GZ(s)J-I
Iq
[~I(S)J
1'z(s)
[~p
-GI(s)
GZ(S)J-I
Iq
[lp + Gz(s)G(s)] - 1
[
= _GI(s)[l p + GZ(s)GI(S)]-1
- Gz(s)[I'!- + G~(s)Gz(s)] -
[I q + GI(s)GZ(S)]-1
(9-32)
458
[Y'I
[~I(S)J [~I(S)~I(S)J= [G
=
YI(S)J ( Hy(s)
[ Y2(S) r 2(s)
G 2(S)U 2(s)
(S)
G~(s)G2(s)[I~+ G~(S)G2(S)J
G 2(s)[I q + G 1 (s)G 2(s)] -
-IJ
= [1 + G(s)GAS)]-1
for i,j= 1, 2
Unity fee
Figurb 9-12
(9-34)
and defining
(9-35)
::;-
rom gf s +gf(S)C(s);
N(s)/D(s), we can write
A
()
(9-36)
9-5
f\.
C(s) (
-[
Compensator
459
Plan!
I'~~lll
~"
k
----... ces) ~i g(s) =.V(s)/D{sl l---;--..
--j
1-
FigurE> 9-12
'01'
i,j= 1,2
(9-34)
(9-35)
(9-36)
;ingularity of J, every
lis BIBO stable if and
n Figure9-8. Because
ed, so are y(t), i= 1,2.
~L!
L __._.
.=.J
kC(s)g(s)
~O.
Thus, if a feedback
system is BIBO stable
d, if a system is BIBO
:table hidden mode al',
:n the system is asymp
uld be designed to be
back Systems
problem of state feed
s of the original system
or. In this section, we
ns in fractional forms.
le corifiguration devel
re 9-12. The compen-
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) =
N(s)fD(s), we can write C(s) as
O(s)N (s)
(9-37)
N(s)(kO(s) - N(s
Now we claim that for any gj(s) = Nj(s)/Dj(s) meeting the pole-zero excess
inequality
(jD j(s) -(jN(s)
';2: (jD(s)
- (jN(s)
(9-38)
where (j denotes the degree, a constant gain k and a proper compensator C(s)
exist to achieve the designo lndeed, by a proper choice of k, we have
(j[kO (s) - N (s)] = (jD j(s)
';2: O
Hence C(s) is proper. Thus for any ij(s) meeting the pole-zero excess inequality,
a proper compensator exists in Figure 9-12 to meet the designo We note that
not only the poles but also the zeros of the overall system can be arbitrarily
assigned.
There is, however, a very serious problem in the designo From (9-37), we
see that C(s) contains the factor l/g(s) = O(s)/N(s). Hence the tandem con
nectiori. of .C(s) followed by g(s) always involves poie-zero cancellations.
Furthe'rmore, the canceled poles are dictated by the given plant; the designer
has no cO'n'trol ayer these poles. Ir the 'given 5(s) has polesor zeros in the open
right-half splane, then the design wm" cOntain unstable pole-zero cancellations.
Hence thisdesign is not always permissible inpractice. In oth.er words, if
undesirable pole~zero cancellations are notpermitted, ai-bitrary assignments of
polesand zeros are not always possible in Figure 9-12 by computing C(s)
from (9-37).
460
N(s)
D(s)
kNc(s)N(s)
Dc(s)D(s) + Nc(s)N(s)
(9-39)
with
Do
No
--
O
O
(9-40)
D(s)=Do+Dls+ +D"sn
Dn-l=0
N(s)=N o +NIs +
+N"s"
+ Dc",s"'
Dc(s)=D co +Dels +
Nc(s)=N co +N c1s +
+Ncmsm
(9-41 a)
...
NI
SnI=l:~ );::::
In the fol1owing, we shall study the condition on D(s) and N(s) under which a
set of solution {Dc(s), Nc(s)} exists in (9-40) for any DAs). We sha1l also study
the degrees of Dc(s) and Nc(s) in order to achieve arbitrary pole placement. 4
By arbitrary pole placement, we always assume implicitly that complex con
jugate poles appear in pair. Otherwise, the compensator will have complex
coefficients and cannot be implemented in practice.
The polynomial equation in (9-40) is called the Diophantine equation in
honor of Diophantus of the third century in Reference S139. 1t is ca1led the
compensator equation in Reference S34. This is the most important equation
in the remainder of this chapter. Instead of studying it directly (see Problem
G-14), we shall translate itinto a set of linear algebraic equations. From the
set of algebraic equations, we will then develop all properties of the Diophantine
equation. The set of algebraic equations can also be used to compute the
compensator C(s). Let us define
DI
--------
O
O
.
.
(9-41 b)
(9-42a)
Theorem 9-10
19-42b \
and define
(9-43)
where Di, Ni, Dci> N ci , and F are real constant, not necessarily a1l nonzero. The
substitution of (9-41) to (9-43) into (9-40) yields
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")
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
or continued ofraetlonso
[DeoNeo:Dc1Nc1:"':DemNem]Sm=[Fo
'"
Do
No'"
..
D,,_
-O - --O -
lOmial equation
(9-40)
phantine equation in
t important equation
jirectly (see Problem
quationso From the
es of the Diophantine
lsed to compute the
(9-41 a)
(9-44 )
Fz
D"
...
o_.~ __ ~_
011
v 1
D,,-z
D,,_
D"
..
N,,-z
N,,_
N"
...
O~
--------------------------------------
(9-39)
with
Do
461
-'-0-0- -
...
No
(9-45)
N"
(9-41 b)
(9A2a)
Theorem 9-10
(9-42b 1
(9-46)
(9-43)
:) +Ds + .. +D"s")
:) +Ns + .... +N,S")
with deg N(s) ::;deg D(s) = n. Let De(s) and Nc(s) be of degree m 01' lesso Then
for every D(s) of degree n + mor less, there exist DC<s) and N ls) to meet (9-46) if
and only ir D(s) and N(s) are coprime and m ~n -1.
I
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).
lf D(s) ami 'N(s) arenot eoprime, 'then a solution {D,(s), N~(s)} exists in (9-46) for any DAs) if and
only if the greatest comrrion divisor of D(s) and N(s) is a divisor of D /(s). See Problem G-14.
462
This question will be answered in the fol1owing two theorems by removing the
condition deg D(s) < n + m fram Theorem 9-10.
Theorem 9-11
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
sator C(s) of degree m exists so that the feedback system has transfer function
N(s)Dj 1(s)Ne(s) if and on1y if D(s) and N(s) are coprime and m e. n - 1.
The employment
set of linear algebraic
whether g(s) is strictl
required compensato
Example 1
Consider a plant witl
Proof
If we show that D em in (9-44) is different from zero, then this theorem follows
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.
h g(s)=N(s)/D(s) and
-m, a proper compen
has transfer function
mdmn-l.
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.
Example 1
g\s)
463
+1
+2
s
s_ 2
S2
.2
F2
F3
F 4J
(9-49)
The first five rQws of the 6 x 5 matrix in (9-49) are linearly independent and have
full column rank; hence for any F,i=O, 1, ... ,4, a solution {Dci,N ci with
Dc2 = F~ ad N c2 = O} exists in (9A9).ln other words, the compensator has a
464
scribed by the q x 1
[DeO
Neo]So
[1
2 4]
or sfirugoe-QlLf[Put case.
D(s)
N(s)
1
C(s)=-I
DJ5)
with
where D are scalars
19-50) and (9-53) i'1tc (
Di(' .
(al
Figure 9-13
(b)
compensator must
her words, for every
at the unity feedback
:.'1) is prechosen, then
than n -1 to achieve
465
G (s) =
A
N~(s)
1
6.
I
D~(s) - D(s) N 2(S) = N(s)D - (s)
(9-50)
N~(s)
D~(s)
where D(s) the least common denominator of all elements of C(s). We assume
=t
of
lpensator C(s)
::e, if the degree of a
LS follows: First, we
! is in the row space
pensator of degree O.
o D fes) =15 f(s)(s +k).
,ole s +k of this k is
19 a compensator of
nust increase m by 1
reader is referred to
umber ofparameters
the number of parapole placement, the
bjectives such as the
:1 9-29.
D(s) = Do +DIs +D 2 s2 +
N(s) = No + NIs + N 2 s 2 +
D n #=0
(9-51 )
+ C(s)G(s))- I C(s)r(s)
Hence the transfer matrix of the overall feedback system in Figure 9-13(a) is
equal to
G fes) = G(s)[1 +C(s)G(s)] - IC(S)
(9-52)
= Deo + Deis +
NcCs) = Neo + N eI S +
with
De(s)
+ Demsm
+ Nemsrn
(9-54 )
'Viii be postponed to
1). The plant is de-
+ DlIs n
+ Nnsn
Gfes) =
+ NcCs)N(s)] -1 N(s)NcCs)
= [DcCs)D(s)
(9-55)
(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!.
Note thal, because of possible pole-zero cancellalions, nol all rools f D(s) are poles of G(s).
466
The substitution of (9-51), (9-54), and (9-57) into (9-56) and equating the co
efficients of the same power of s yield
F,,+mJ ~ F
(9-58)
with
O
O
D"
} one block row
O
O
No NI
N"
---------------------------S",=
Do
DI
O
O
Do
No
D"_I
.. .
N"_I
D"
NI!
...
...
O
O
(9-59)
----------------------------
---------------------------O O
O
Dn
Do
DI
...
O
O O
No NI
N"
We call the rows formed from {DJ D rows and the rows formed from {Ni}
N rows. Then every block row in (9-59) has one D row and q N rows. The
matrix S", consists of m + 1 block rows; every block row is the shifting to the
right by one column of its previous block row. The matrix S,. is c1early a
(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
algorithm discussed in Appendix A. Then as discussed in Theorem G-13, aH
D rows in Sm wiH be linearly independent, and sorne N rows will be Iinearly
dependent on their previous rows in Sm' Let r be the number of Iinearly
dependent N rows in the i + 1 block row ofS",. Then because of the structure
ofS",. we have O:::::; ro:::::; 1'1 :::::;
:::::;q. Let l' be the integer such that,.o:::::; 1'1 :::::; .. :::::;
r,. _ 1 < q and 1', = 1',+ 1 =
= q. We cal1 l' the row index of C(.~). Then l' is
the largest row degree of A(s) in any left-coprime fracton of C(s) = A - 1 (s)B(s)
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).
TheOl'em S- - T
Consider the feedback system shown in Figure 9-13(a) with the plant described
by a q x 1 strictly proper (proper) rational matrix C(s) = N(s)D-I(s) with
deg D(s) =n. Then for any D(s) of degree n +m, there exists a 1 xq proper
(strictly proper) compensator C(s) = D; I(s)Ne(s) with deg De (s) = m so that the
feedback system has q x q transfer matrix N(s)Dj I(s)Ne(s) if and only if D(s)
and N(s) are right coprime and m;::: v - 1 (m;::: v), where v is the row index of
(;(s), or the observai?i1ity index of any ir~educible realization of (;(s), or the
largest row degree of A(s) in any left~coprime fraction of (;(s) = A -1(s)B(s) with
A(s)row reduced.
71[ q =
l'
is equal.to
11.
Proof
md equating the co
Fn+mJ Do F
(9-58)
467
Proof
The design problem is equivalent to the solving of Equation (9-58). The
application of Theorem 2-4 to the transpose of (9-58) yields that for every F,
there exists a set of {Dei> NeJ or, equivalently, a compensator C(s) if and only
if the matrix Sm has a fu11 column rank, that is, pSm = n +m + 1. From the
definition of ri and the linear independence of a11 D rows, we have
m
pSm = (m
le block row
t")
(9-60)
i~O
(9-59)
.\ to 11.
+ 1) + I (q -
468
Theorem 9-12'
Consider the feedback system shown in Figure 9-13(b) with the plant described
by a strictly proper (proper) 1 xp rational matrix C(s)=D-l(s)N(s) with
deg D(s) = n. Then for any D(s) of degree n + m, there exists a p x 1 proper
(strictly proper) compensator CCs)=Ne(s)D;l(S) with degDe(s)=m so that the
feedback system has 1 x 1 transfer function N(s)Dj 1 (s)Ne(s) if and only if D(s)
and N(s) are left coprime and m /l-l(m /l), where /l is the column index of
C(s), or the controllability index of any irreducible realization of C(s), or the
largest column degree of A(s) in any right-coprime factorization of C(s) =
B(s)A -les) with A(s) column reduced.
III
denominator of all
of C(S).8 Let 6.(s) b(
common denominat
Definition 6-1). C[
nomial h(s). lf -(s)
rational matrix. Fo
1_1_
+1
Gl(s)= lS_1_
D(s) = D(s)De(s)
+ N(s)Ne(s)
(9-61 )
Note that De and N e are on the right-hand side of D(s) and N(s), rather than the
!eft-hand side as in (9-56). Equation (9-61) can be solved indirectly by taking its
transpose to become the form of (9-56) or solved directly as fol1ows: Using
the coefficient matrices of D(s), N(s), Dcls), Ne(s), and D(s), we form the linear
algebraic equation
Tm
Deo
Neo
Del
N el
D em
N em
O
No
NI , Do
No
DII
O
,
:
N" Dn O
Dn
Nn N"
Do
DI
,
,o
,o
,
1
O
O
O
O
O
, Do
DI
O
D
m + 1 block columns
O
II
No
NI
N
Deo
Neo
Del
N cl
Dem
N em
Fo
Fl
F2
(9-62)
Fn + m
II
5+
s +1
The polynomial equation arises in this theorem is of the form
We see that if al =
nomials of C2 (s) and
teristic polynomials, l
the characteristic poI:
shows that in any liJ
canceling a pole is ve
theorem.
This equation is essential1y the transpose of (9-58) (see Theorems G-14 and
i : ':,;;"....:: .... ~.; .
to right. Let /l be the least integer such that the last p N columns of TI' are all
linearly dependent of their left-hand-side columns. The /l will be called the
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
9-12' is similar to the one of Theorem 9-12 and will not be repeated.
Multivariable case: Arbitrary pole assignment.
In this section, the
. design technique de'leloped in the orevious silb~ection will be exten1ed to
.general proper ratiOnlJ.l matrices. We extend it .first to a special class of rational
matrices;called cydic ratinaJmatrices, ando thentothe general case.
Consider a q x pproper rational matrix G(s). Let tjJ(s) be the least common
.._
Consider a q x p cycl
and 1 x q real consta
polynomial of A. Comp.
~s)=D-(s)N(s)
l
S
(9-61 )
Gz(s)=
+ 1
Fo
F
Fz
(9-62)
Fn + m
s+1
_1_
~1 -J
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
C(s) = lf -1(s)N(s)NcCs), where N(s) and Nc(s) are q x 1 and 1 x p polynomial
matrices, then C(s) is cyc1ic (why?). Ir no A. appears as a pole of two or more
elements gij(s) of C(s), then C(s) is cycli<l (Problem 9-14).
Consider the cyc1ic rational matrix Gz(s). We form
(;z(s)a
o
_1_
+1
GI(s)= _1_
:le form
469
~
-
Gz(s) [al]
az
l a~ :~z J
azs +(2az
s +1
+al~
We see that if al = -az, then Gz(s)a = [O -az]' and the characteristic poly
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
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.
Consider a q x p cyc1ic proper rational matrix C(s). Then for almost al! p x 1
and 1 x q real constant vectors t and t z, we have
Li[C(s)] = Li[C(s)t]
Li[tzC(s)]
(9-63)
Let {A, R, C, El be an irreducible realization of G(5). Then it can be shown that t/J(s) is the m"nmal
polynomial of A. Compare the defintion of cydic.raonal matrix with theone for'a constan! A
matrix in Problem 2-45.
9 Compare this theorem with Theorem 7-5.
470
Proof
Theorem 9-14
Let us write
l~~i~~l
l{I(s):
Ni s)
where l{I(s) is the least common denominator of all elements of (;(s); N(s) is a
q x p polynomial matrix, and N(s) is the ith row of N(s). We assume that
every element of G(s) is irreducible. The cyclicity assumption of G(s) implies
8[G(s)J = l{I(s). Let A, i = 1, 2, ... ,m, be the distinct roots of l{I(s). First we
assume that A, i= 1,2, ... , m, are real. Then we have
N(A} 1=0
Proof
Since Gts) is cyclis. b
that 8[ G(s)J = 8[ G(.'
fori=1,2, ... ,m
Hence the rank of N(A) is at least 1. Consequently, the null space of N(A) has
a dimension of at most p - l (Theorem 2-5). In other words, the set of vectors t
in the p-dimensional real vector space ~ P which satisfy N(A}t = Ois a linear space
of dimension at most [J - 1. To help to visualize what will be discussed. we
assume p = 2. Then those t satisfying N(A)t = O, for each i, is at most a straight
line in ~ 2. Consequently, there are at most m straight lines on which N(A)t = O
for sorne i. The tI not on these straight Iines have the property
'I'(s)
N ,IS)t]
N 2 (s)t 1
:
Nq(s)t l
i:
Figure 9-14
Design (
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).
Proof
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)
Then Theorem 9-12 imp\ies lhe exislence of a 1 XI/ proper rational matrix
(:(s) = D c- 1 (s)N,(s) with deg s) = 111 ~ l' - I if G(.~) is strict\y proper. such that
all n +m roots of
D Il~) = D.(s)D(s)
Nrl-~)N(s)
can be arbitrary assigned. Now we show that the roots of Df(s) give the poles
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)th(s)]
,----------1
e
polynomial of G(s)t.
roats af 1/1 (s) are rea\.
)roof sti\\ ho\ds if we
sl] = Ll[G(S)tJ. The
I
L
-.J
iI
(al
.J
lb)
Figure 9-14
472
which implies
Theorem 9-15
hes) = [1 + C(s)G(s)t] - C(s)r(s)
G(s)=G(s)t[1 +C(s)G(s)t]-C(s)
(9-64)
Consider a q x p pro pI
every p x q constant rr
The substitution of G(s)t = N(s)D- (s) and C(s) = D,-(s)N,(s) into (9-64) yields
G(s) =
and
is proper (strictly
pro~
(9-65)
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
the qxp compensator defined by cts)=tC(s)=D;(s)tNc(s) can achieve
arbitrary pole placement.
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
and A(s) is row reduced, then the observability index of G(s) is equal to the
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
left coprime. Hence the observability index of G(S)t1 is also equal to the largest
row degree of A(s). This establishes that the observability index of G(s) and
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.
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
compensators in Theorems 9-12 and 9-12' are minimal to achieve pole place
ment, the degree in Theorem 9-14 may not be minima\. In other words, it may
be possible to design a compensator of degree less than min (J1- 1, v -1) to
achieve arbitrary pole placement for a q x p cyclic proper rational matrix.
What is the minimum degree seems to be a difficult problem.
With Theorem 9-14, we can now discuss the design of compensators for
general proper rational matrices. The procedure consists of two steps: First
change a noncyclic rational matrix into a c}'clic one and then apply Theorem
9-14. Consider a proper rational matrix G(s). Let Ll(s) be its characteristic
polynomia\. We claim that if all roots of Ms) are distinct, then G(s) is cyclic.
Let if(s) be the least common denominator of all elements of G(s). Then we
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.
Hence the condition thatall roots of ~(s) are distinct is a sufficient but not
necessary condition for G(s) to be cyclic. This properiy ~v.i11 be used to establish
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.
Proof
10
3.' (s)
If ~(s) has repeated roc
sufficient condition for
det
o
(See Appendix G). W
nomial of kij. There is
of {kij} as a vector in
clear that the solution,
almost every kj, we ha'
rrhis s Si!:.~.712.r t0 th2J
are distinct.) Hence G
If G(s) is strictly prc
"
the condition for G(
det (1 + C( 00 )K) '" O. fl
" for al
proper, so is G(s)
473
Theorem 9-15
Consider a q x p proper (strictly proper) rational matrix G(s). Then for almost
every p x q constant matrix K, the q x p rational matrix
(9-64)
Proof'O
polynomial matrices.
oles of (9-65). Hence
5)t l Nc(s) can achieve
We show that the roots of the characteristic polynomial, ~(s), of G(s) are all
distinct for almost all K. Let
~(s)=aOsn +als"-l + ... +a"
If ~(s) has repeated roots, then ~(s) and ~'(s) are not coprime. A necessary and
sufficient condition for ~(s) and ~'(s) to be not coprime is
a n- l
an
al
az
a
an-l
O
2a n- z
a"
O
det
ao
al
O
O
ao
a u -: 1
O
2a,,_z
nao
=y(k) =0.
u.,
t~i..~
'((";'0 1',3
~f(x~:,'2 +C~2:.'+"'
"
are dis}inct.) Hence G(s) is cyclic"
lf G(s) is strctly PIoper, so is G(s) for every K. (Prove it.) If G(s) is proper,
the condition for G(s) to be proper, as discussed in Section 3-6, is
det(1 + G( 00 )K) 1= O. Almost all K satisfy this condition. Hence if G(s) is
'" for almost all K. This completes the proof of this theorem.
proper, so is G(s)
Q.E.D.
10
474
Gn
(s)
= [1 +
=[1 +1
=[1 +,
G(
The substitution of
o,
l.
(s)1-- int.iL...l'
X
Consider the feedback system shown in Figure 9-15ib) with the plant described
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
of degree m. If m =?:min (J.1-1, v -1)[m =?:min (J.1, v)], then al1 n +m poles ll of
the unity feedback system in Figure 9-15(b) can be arbitrarily assigned, where J.1
and vare, respectively, the control1ability and observability indices of any
irreducible realization of G(s), or the column index and row index of C(s).
Proof
First we show that if C(s) = C(s) + K and if C(s) can be written as C(s) =
De-l(s)tl Nis), where t l is a p x 1 constant vector and NcCs}is a 1 x q polynomial
matrix, then the poles of the system in Figure 9-15(a) and those in Figure 9-15(b)
are the same. The transfer matrix of the feedback system in Figure 9-15(a) is
Because of possible pole-zero cancellations, not all n + m of these roots will be the poles of the
resulting system. However, for convenience, we shall call all of them poles. The strictly proper
part of this theorem was first established in Reference S26 in the state variable approach and in
Reference S51 in the transfer-function approach.
= [D(s).
Gfes) =
[D(s)De(s)
+ N(~
'" = l\!(s)lDJG(s)
(;(5)
----------l
I +
X(s)[DI
I
I
I
-.J
(al
Design of compensator.
!r
Figure 9-15
By a similar manipulal
can be computed as
Theorem 9-16
11
given by
(b)
m ~ min (ji - 1, V- 1)
oper or proper, w~ere
lbility indices of G(s).
{a) can be arbitrarily
:actly a unity feedback
10Ies of GAs), not the
ion of C(s) and K into
ack system shown in
475
given by
GJI (S) =
be written as C(s) =
s) is al x q' polynomial
those in Figure 9-15(b)
em in Figure 91~(a) is
'oots wll be the poles of the
m poles. The strictiy prop~r
lte varable approach and ID
G J(s)
(9~67)
From (966) and (9-67), we conclude that the systems in Figure 9-15(a) and (b)
have the same set of poJes. 11 Since the poles in Figure 9-15(a) can be arbitrarily
assigned by a proper choice of C(s) and K, we conclude that by choosing (:(s) =
C(s) +K, the poles of the system in Figure 9-15(b) can be arbtrarily assigned.
Now we claim that deg det C(s) = deg det Ces). Let x= Ax + Bu, y=Cx +Eu
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
proved by using Theorem 9-3. If we :vrite K = KI.-I = 1; lK, then Theorem 9-3
implies that the paral1el connection of C(s) and K is controllable and observable.
Hence we have deg det (:(s) = deg det C(s) +deg det K = deg det C(s).
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
N(s) and D(s) are right coprime and D(s) is column reduced, then the control
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)
Hence Theorem 0-8 implies that N(s) and D(s) +KN(s) are right coprime. The
column degrees of D(s) are dearly equal to the column degrees of D(s) +KN(s)
for a11 K if N(s)D-I(s) is strictly proper, and Joralrno~t ~.ll K if N(s)D-1(s) is
proper. Hence the controllability ndices of G(s) and G(s) are the same. This
fact CilO also be proved by using dynamical equations:, The observability part
Q.E.D. ,
can besimilarly proved. This completes the proof of this'iheorem.
Thetransfer matrix G(s) of the resulting feedback system isglven in (9-:67).
Unlike the cases of single input, single output, and cyclic plant, G/s) is not in
476
the form of N(s)D 1 (s)NAs) with scalar D f(s). However, from the design
procedure, we may conclude that G(S) in (9-67) iscyclic. See also Problem 9-16.
We recapture the design procedure in the following:
Step 1. Find a K such that G(s) = (1 + G(s)Kr 1 G(s) is cyclic.
Step 3. Write G(s)t I = N(s)D-I(s) =(N o + NIs+'" + Nllsn)(D o + DIs + ... +D"s")- 1
and form the SIlI in (9-59). Find the least integer \' such that Sv- has a [ull
column rank. The integer vis called the row index or the obseruability indexo
Step 4. Choose n +1.' - 1 number of poles and compute
We compute
(;(s) = [1 + G(s)K
(S3
+ W [(s-+
1
~(s) =
(;"3+W
F1
[s(s-
Example 2
~f(S)
Consider
(;(s)
l ~ ~1
The solutions of
= si,
.O
O s
are D co = 1, Mo ==1, N~
9-IS(b) is given by .
C(s) = C(s)t 2 -t
We compute
(;(s) =
clic.
.Do+D1s+"'+ D "s")-1
,ch that S. _ 1 has a fun
the observability indexo
[1
+ G(S)K]-IG(S) =
~3 [s + 1
S3 +
s(s + 1)
- 2s
_2
477
1]
S2 - s + 1
[2
('
s'
1
s
W[s(s+1)(s2- s + 1)+2s]= 3-3)2 S(S3+ 3 )=_,-
+
(s +
(s' +3)
1
2"-
F'I+V-]
T=l~~~ ~
o
0:0 1 O
1:O O O
The matrix T o has a full row rank; hence J1 = 1, and the degree of compensator
is J1 - 1 = O. Let
C(S)=f
cO
NJO]
[ N~o
N3
co
The solutions of
-1] r2 ., . , 1J
O = 3
478
G(s)=[I+G(s)C(s)]
c-.J
1
[(S+1)(3S+2)
G(S)C(s)=(s+1)2(S+2)
O
_lA
Its poles are -1, -1, and -2. We note that G is cyclic (verify it).
(9-68)
The first equality is obtained from y(s) = G(s)C(s)[r(s) - y(s)]; the second one
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)].
(Verify them. Compare also with Theorem 3-3.) In the single-variable case,
if we assign both poles and zeros of the unity feedback system, then the design
will generally involve undesirable pole-zero cancellations. In arder t avoict
"i:hese canceilatios, v<l Clssigi only t~le. poles and leave the ZfOS UGSpCIi(;u.
13
Compare with Method 1 of the multivariable case in the design of state feedback.
We also write
and
The substitution ofthe
where
Do
No
S
=
ni
[ O
O
D1
NI
.
.
Do'"
No'"
[
l'
----- --------l
! - - - - --------
'"
...
Figure 9-16
N(s)I
= N(s)[
12
G/(s) =
(s+1)(s+2)
479
3s +2)
G(s) =
ic (verify it).
-1
D e- 1 (s)N e(s)
(9-69)
trix assignment.'
+ Ne(s)N(s)
D(s) = Dcls)D(s)
(9-70)
Then we have G(s) = N(s)Dj 1 (s)Ne(s). Hence the design problem becomes:
Given D(s) and N(s) and an arbitrary D(s), find De(s) and Ne(s) to meet (9-70).
This is the matrix version of the Diophantine equation in (9-40). Similar to
(9-40), we shall translate it into a set oflinear algebraic equations. Let us write 14
to achieve arbitrary
Jminator matrix, then
Hence this problem
n the previous subsec
j for arbitrary denomi
quired for pole assign-
'S
(9-71 a )
(9-71b)
We also write
De(s) = Deo + Dcls +
+ Demsm
Ne(s) = Neo + Ncls +
+ Nemsm
D(s) = F o + F 1s+ F 2S 2 + ... + F mw
and
(9-72a)
(9-72b)
S '" +1'
(9-73)
Neo: Dc
Nc: ... : D em
Nem]Sm=[F o F 1
F2
...
Fmw ] ~ F
(9-74)
(9-68)
where
Do
DI
No
S,.,= [
~ ~o
--
l-~
DI'
NI'
NI
--
:::
~1'-1 ~I' ~
1'- 1
l'
...
.,.
:::
}ro(numberofdependent
}r/.o WS)
(9-75)
The rnatrix Sm has m + 1 block rows; each block row consists of p rows formed
from Di and q rows formed frorn Ni' We call the former D rows and the latter
N rows. This rnatrixS mis studied in Appendix G. It is shown in Theorem G-13
that if N(s)D- 1 (s)is proper, then all D rows in Sm, m =0,1,2, ... , are Iinearly .
independent of their previous rows. Sorne N rows in each block, however,
rnay be linearly dependent on their previous rows. Let r be the number of
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
(9-76),
rank Sm = (m + l)p +
(q -1)
j=O
v- 1
and
j=
for m 2. v-1
(q -I)=(m +l)p +n
o
(9-76)
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
ble realization of C(s). See Equation (G-8l) and Theorem 6-6.
Theorem 9-17
Consider a q x p proper rational matrix with the fraction G(s) = N(S)O-I(S).
Let /l, i = 1, 2, ... , p, be the column degrees of O(s), and let v be the row index
ofC(s). Ifm 2.v -1, then for any D(s)withcolumndegrees m +/l, i = 1, 2, .. . ,p,
or less, there exist Dc(s) and N/s) of row degree m or less to meet
S _[0 0
0- No
01
NI
/L-1
N/L-I
O/L]
i= 1
NI'
has at least a total of = 1 (/l - /l) zero columns. In the matrix SI' some new
zero column will be created in the rightmost block column; however, sorne
zero colulnns in So 'NHl cHsappe~r froro. 1- I-Iencc the nurnb~r oi zere cJI~;,,_:..L_'
in SI remains to be
p
A comparison of (9
co1umn rank if and (
/l
i=l
In fact, this is the minimum number of zero columns in Si, i = 2,3, . . .. Let
and
lheorem 9-Hs
Consider a q x p stric
C(s) = N(s)D- 1 (s). L
v be the row indeJ
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
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.
L
(;;1
15
(9-77 )
(/l +v)p-(P/l-
/lJ=vp
/li
i=l
(9-78)
for m?:. v - 1
481
S,. _1
has a full
(9-76 )
p
J1i=n
i= 1
This equality holds if and only if, following Corollary G-14 and Definition G-4,
D(s) and N(s) are right coprime and D(s) is column reduced. Since both the
rank of Sk and the number of columns in Sk increase by p as k increases by 1
from v -1, we conclude that S"" for m?:.v -1, has a full column rank if and
only if D(s) and N(s) are right coprime and D(s) is column reduced.
Ir DAs) is of column degree m +J1i, then there are at least I~=1 (fl-flJ
zel'o columns in the F matl'ix in (9-74). Furthel'more, the positions of these
zel'o columns coincide with those ofS",. Since S", has a full column l'ank, the F
must be inside the row space of Sm' Hence a set of solutions {Dei, N ei } exists in
(9-74), 01' equivalently, a set of solutions {De(s), Ne(s)} exists in (9-70). This
completes the proof of the theol'em.
Q.E.O.
\ is column reduced. 15
]
e matrix SI' sorne new
,Iumn; however, sorne
1mber ofzero columns
This theorem states the condition fol' the existence of De(s) and Nis) to
meet (9-70), but states nothing l'egarding whethel' D; 1 (s)Ne(s) is Rl'oper or not.
To study this question, we consider separately the case where G(s) is strictly
proper 01' proper. The situation is similar to the single-variable case (Theorems
9-11 and 9-11'). However, the proof must be modified because the leading
coefficient matrix DI' of D(s) is generally not nonsingular.
Before proceeding, we define
and
, Sil,,)
, s",p}
(9-79)
(9-S0)
Theorem 9-' 8
Consider a q x p strictly proper (proper) rational matrix G(s) with the fl'action
G(s) = N(s)D- l (s). Let fli, i= 1,2, ... ,p, be the column degrees of D(s) and let
v be the row index of G(s). Let mi be the row degrees of D,(s). [f
mi ~ v - 1 (mi?:. v) for all i, then for an y D f(S) with the property that
lim He-1(s)D(s)H-l(S)=J
s .....
00
(9~S1)
ir and only if D(s) and N(s) are right coprime and D(s) is column reduced.
482
Proof
Let m = max {mi} and J1 = max {J1;}. Consider the ith row equation of (9-74):
where D,cj denotes the ith row of D cj and so forth. Since mi;::: l' - 1, the resultant
Sl1li' excluding (J1- J1) number of zero columns, has a full column rank. The
assumption of (9-81) implies that the ith row of D [(s) has column degrees
(J1- Ilj) number of zero elements whose
at most mi + Ilj. Hence F has
positions coincide with those o S"", Hence we conclude that for any D [(s)
meeting (9-81), solutions Dc(s) and Nc(s) of row degrees at most m exist in (9-82).
We write (G-57) as M(s) = [M lJc + M(s)JHc(s), where M(s)~ M1c(s)H c- 1 (s)
is a strictly proper rational matrix and M( (0) = O. Similarly, we write
D(s) = [D h + D/(s)JH(s)
N(s) = [N h + N(s)JH(s)
(9-83)
where D/(s), DcI(s), N 1(s), Ncl(s) and D [1(S) are strictly proper rational matrices.
The constant matrix D IJ is the column degree coefficient matrix of D(s) and is
nonsingular by assumption. The substitution of (9-83) into (9-82) yields, at
s=oo,
Oel,OIJ + NeI,N" = D[IJ
(9-84)
Ir G(s) is strictly proper, N,,=O and D ch = D[hD,;-1. By assumption, D[IJ=J
is nonsingular; hence Dc(s) is row reduced and 0c- l(s)N c(s) is proper. Ir G(s)
is proper, we have N h i= O. However, the row of Sm corresponding to N cll
are linearly dependent if m;::: v, Hence we may choose N clJ = O. For this choice,
we have Dich=D[hDh-l and D ch =D[I,D,;-1. Hence Dc(s) is row reduced.
Because of Ni'h = O, Dc- 1 (s)N c(s) is strictly proper.
Q.E. D.
It is important to note that if mi;::: l' for some i, the solutions of (9-74) 01'
(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~/
proper solutions. Ir O[(s) meets (9-81), it is said to be row-column-reduced in
Reference S34. Every O [(s) can be transformed, by elementary operations,
into a row-column-reduced one.
Now we may apply Theorem 9-18 to the design problem.
lim diag(s\l,.
s ....
00
Theorem 9-19
1-----
Consider a q x p strictly proper (proper) rational matrix (;(s) with the fraction
(;(s) N(s)D-l(S). Letll, i= 1,2,.'.. ,p be thecolumn deg~ees of O(s) and let v
be lbe row index of G(s). Let the row degrees of Dc(s} be mi, i;=; 1,2, ... , p. Ir
mi;::: l' ~.l (mi;::: v) for al1 i, thenfor any O [(s) with the property thilt
0;=
_ _
o
-'/~'-J-_--l-;-/~
Dc 1
Figure 9-17
Unity feec
s- 00
483
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)Ol(S)N c (s) ifand only if O(s) and N(s)
are right coprime and O(s) is column reduced.
il
This theorem follows directly from Theorem 9-18 and Equations (9-69) and
2: v - 1, the resultant
.Jll column rank. The
, has column degrees
zero e1ements whose
de that for any Dr(s)
most mi exist in (9-82).
: MI(s) MJs)H: 1 (s)
lady, we write
li
(9-83)
(9-84 )
( assumption, D [11 = J
[e(S) is proper. Ir (;(s)
~orresponding to N icl1
el = O. For this choice,
Oe(s) is row reduced.
Q.E.D.
solutions of (9-74) 01'
ions. However, there
of proper or strictly
)w-column-reduced in
lementary operations,
(9-70). Several remarks are in order regarding this theorem. We discuss only
the case where (;(s) is strictly proper. Ir mi = v - 1, for all i, and if we choose
O [(s) = P(s)O(s), then the unique solution of (9-70) is Oc(s) = pes) and Nc(s) = O.
In other words, if the denominator matrix of the plant is not to be altered, the
compensator C(s)= O,:"l(s)N c(s) is O. This is a degenerated case. The second
remark concerns the possible existence of common divisors between N(s) and
O[(s) and between D[(s) and Nc(s) in (;r(s) = N(s)D l(s)N c(s); and between
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).
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)=
I. Hence the unity feedback system in Theorem 9-18 is well posed. Note that
ifDe(s) and Ne(s) are left coprime, thedegreeofthecompensator is mi 2:p(v- 1),
which is much larger than the one required for arbitrary pole assignment.
We state the dual of Theorem 9-19 as a corollary.
Corollary 9-19
Consider a q x p strictly proper (proper) rational matrix G(s) with the fraction
C(s) = O -l(s)N(s). Let Vi, i = 1, 2, ... , q, be the row degrees of D(s) and let J1
be the column index of C(s). Let thecolumn degrees of De(s) be mi, i = 1,2, ... , q.
If mi 2: 1I - 1 (mi :?-/I) for aH i, then for any D [(s) with the property that
liro diag {s - \\ s- \\ ... , s - "q} D [(s) diag [s m"
S -m>,
... , S-m(/} = J
S""" 00
exists ano is nons~nguh1f. th~~:,:: :~x.js~~ '~\ r::r)n:;~:-=:":SF:t~;~- ~//itb :.0. ';.' ','
-~_
(strictly proper) rational matrix Nc(s)D c- 1 (s) such that the transfer matrix from
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)
are left coprime and D(s) is row reduced.
""
,Iem.
a p x q proper (strictly
..: ,
484
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
Dc(s)[D(s)Dls)
(9-85)
Define
(9-86)
Then
G(s) becomes
G(s) =
which yields
= 1- Dc(s)D 1 (s)D(s)
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 desgn n Corollary 9-19 hnges on solvng (9-86).
Note that the transpose of (9-86) becomes (9-70); left coprme and row reduced
become rght coprme and column reduced. Hence Theorem 9-18 can be
applied directly to the transpose of (9-86). Of course Equation (9-86) can also
be solved directly. We use the coefficient matrices of D(s) and N(s) to form
T k as shown in Equation (9-62) and then search linearly independent columns
in order from left to right. The least integer Jl such that all N columns in the
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
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
realization of (;(s). The proof ofCorollary 9-19 is similar to the one ofTheorem
9-19 and will not be repeated.
Example 3
S J[s2-1
S2
+s + 1
S2
O
-1
J-l
We form So, Si' ... , and search their linearly dependeni rows in order from top
to bottom. For this example, we can readily obtain that v = 2. Clearly we
have Jll = Jl2 = 2. Let /111 =/112 = V - 1 = 1. We choose
Remarks are in 01
and the design of ar
strictly proper, the
min (Jl- 1, v - 1) (Th,
p(v -1) (Theorem 9-1
ment assigns only the
always yields a cycL
mairix yields general
of thecyclic and non
applicable here, the
placement design are
denominatormatrix (
of these two designs i
;)Nc(s)D; I(S)
,; I(S)
-1
O
O -1
O
O
O, O
1, O
O
O
100
110
O
1010
1,0
O
O
O
O
O
O
1
O
NclJ - -6 ---0- : ~
(9-85)
(9-86)
=
O
O,
O
O
O
1
O
485
O -1
1
O
O
1
O
O
O
O
1
1
O
1
O
tJ
[~ ~ ~ ~ ~ ~ ~ ~J
which yields
_2s;21
s+d
We see that Dc(s) is singular and the compensator Dc-l(s)Nc(s) is not defined.
Hence if G(s) is proper, the choice of mi = v - 1 may not yield the required
eompensator.
Ir we choose mi = v and choose
rs2 +3
DJs)=l O
1)/3J
4~s s- +s + 1
NJs)
=[
4S +4
O
-4(s +
1)/3J
2(s + 1)
lIi
O
S2
-1
J-l
[O30J
2
t - ' s 2'
[1O 0J1 s
486
GJ(s) = N(s)D i
l(s)N c(s)
= N(s)(D(s)N(sW 1 Nc(S) = fi j
1 (s)Nc(S)
p
P( s) =d'lag {f31(S) , f32(S) , ... , f3 (S)}
(;(l(S) (;(2(S)
(;(p(s)
deg iX(S
We see that the transfer matrix of the tandem connection of the compensator
followed by the plant is G(s)C(s) = N(s)D- \s)D(s)N -l(S)p(S) = P(s). Because
G(s)C(s) = P(s) is diagonal, the feedback system in Figure 9-18 reducesto a 'set
of p number of decoupled single-loop systems. Hence the system is decoupled.
Now the matrix P(s) is to be chosen so that the compensator
C(s)
(9-87)
G ()-d"
Js -
{131(5)
132(S)
~)_}
is BIBO stable. !f det N(s) is H urwitz, the design in Figure 9-18 wil! n.ot
16
Compcnsator
,----------------l
\
1--,---1----::-"<1
1--'---'
I ---'-.,r
rT------'-'""' I
1 "'1
----." '-/--~l
-l
j,_y:-:1'i~
l7
Figure 9-18
Decoupling of a plant.
'Ll:.t ,:;1U;1r:;c:'
ed to design a unity
,nal and nonsingular.
If the plant transfer
may choose D(s) =
n the overall transfer
=
Dj 1 (s)Nc(s)
487
(9-89)
of the compensator
P(s) = P(s). Because
9-18 reduces to a set
system is decoupled.
nsator
(9-87)
for i = 1, 2, ... ,p, then C(s) is proper. That is, if the pole-zera excess 17 of the
ith column al P(s) is equal to 01' larger. than the largest zero-pole excess of the
th colum~ of (; -1(S), the~ C(s)' is proper. Note that the left-hand side of (9-89)
~ also equal to the pole-zera excess of G(s). Ir p = 1, the zero-pole excess 01
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.
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
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.
We cal! dg D(s) - deg N(s) he po\e-zero excess and deg N(s) - dg D(s)
he zero-pole excess of gIs).
488
9-6
and
l-co
This is called the asymptotic tracking. It is well known that if r(l) is a step
function, or res) = 2[r(t)] = l/s and if C(s)g(s) is a type 1 system, that is, C(s)g(s)
has one pole at the origin, and if the feedback system in Figure 9-19(a) is
asymptotically stable, then the output y(t) wi1l track r(t) asymptotically. In
this section, this statement and its generalization will be established.
Consider the feedback system shown in Figure 9-19(b). The plant is written
as g(s) = D-1(s)N(s). At the input of D-1(s), a signal w(t) is injected into the
plant. The signal W(l) will be called the disturbance signa!. Now in addition
to asymptotic tracking, we also require that the affect of w(t) at the output
a pproaches zero as t -> 00 ; that is, lim y w(t) -> Oas t -> 00, where y w{t) is the out put
of the feedback system in Figure 9-19(b) due to the application of w(t) and
r(t) == O. This is called the disturbance rejection. Hence, if we succeed In finding
a compensator C(s) = Nc(s)/Dc(s) in Figure 9-19(b) so that for any r(t) and any
w(t), we have
lim e(t) = lim [r(t) - y(t)] = O
{-co
Befare proceedin~
If r(t) and w(t) both g<
the feedback system
or w(t) does not go to
nature, then it is not
rejection. Hence we J
the designo We assun
(9-90)
and
with some unknown ir
A,. and A", are D,(s) al
t-> 00 have no effect e
and D",(s) have zero o
nominator of the unsl
zero or positive real p;
l-CO
then the feedback system achieves asymptotic tracking and disturbance rejec
tion. In this section we shall study the design of this problem.
Theorem 9-20
(b)
Figure 9-19
[f no root f cP,(s) is a
transfer function.l/ cjJ(s
. able (Theorem 9~2).
coprime, and there exi
back system shown in
Rejection
ptrol systern shown in
ven, the prablem is to
'so thatthe fuedback
oecifications. One of
I:he plant y(t) to track
1;, it is not possible to
lerence 546). The best
Befre 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
the feedback system in Figure 9-19(b) is asymptotical1y stable. If either r(t)
or w(t) does not go to zera, and if we have no knowledge whatsoever about its
nature, then it is not possible to achieve asyrnptotic tracking and disturbance
rejection. Hence we need sorne information of r(t) and w(t) before carrying out
the designo We aSSume that the Laplace transforrns of r(t) and w(t) are given by
A) =.;Lro[r(t )J = N,.(S)
res
D,.(s)
and
nd disturbance rejec
blern.
489
w(s)
J = -(-)
N w(s)
= 2 [w(t)
Dws
(9-91 )
(9-92)
where the polynomials D,(s) and Dw(s) are known and the polynomials N,.(s)
and N,Js) are however arbitrary so long as res) and w(s) are proper. This is
equivalent to the assumption that r(t) and w(t) are generated by
x,.=A,X,.
and
r(t) = C,.X,
xw(t) = Awx w
w(t) =CwX w
(9-93a)
(9-93b)
(9-94a)
(9-94b)
with some unknown initial states x,(O) and xw(O). The minimal polynomials of
A, and A w are D,(s) and Dw(s). The parts of t(t) and w(t) which go to zera as
t --+ 00 have no effect on y as t -+ 00, hence we assume that sorne roots of D,(s)
and Dw(s) have zera or positive real parts. Let cjJ(s) be the least common de
nominator of the unstable poIes of res) and w(s). Then aH roots of cjJ(s) have
zera or positive real parts.
Theorem 9-20
Consider the feedback system shown in Figure 9-19(b), where the plant is
complete1y characterized by its proper transfer function g(s). The referenr;<>:
signa! r(l) and disturbance signal w(t) are modeled as res) = N,.(s)fD,.(s) and
w(s) = N w(s)/Dw(s). Let cjJ(s) be the least common denominator of the unstable
poles of res) and w(s). If no root of cjJ(s) is a zera of g(s), there exists a compen
sator with a proper transfer function so that the unity feedback systern is
asymptotical1y stable and achieves asymptotic tracking and disturbance
rejection.
y(t)
Proof
[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
+ Ncls)N(s)
N ",(s)
+ Ncls)N(s)
D",(s)
Dcls)Nw(s)
4> (s)
D c(s)D(s)4>(s) + N c(s)N(s) D",(s)
(9-95)
Since aH unstable roots of D,/s) are canceled by 4>(s), all the poles of Yw(s) have
negative real parts. Hence we have yw(t) = -ew(t)->O as t->oo. We see that
even though w(t) does not go to zero as t ........ 00, its effect on y(t) diminishes as
(->OO.
y (s) =
,.
(1 _
Nc(s)N(s)/Dc(s)D(s)c/J(s) ) r(s)
1 +Nc(s)N(s)/D c(s)D(s)4>(s)
Dc(s)D(s)c/J(s)
N ,(s)
D c(s)D(s)4>(s) + Nc(s)N(s) D,.(s)
Dcls)D(s)N,.(s)
D c(s)D(s)4>(s)
+ Nc(s)N(s)
4>(s)
D..(s)
(9-96)
Again, aH the poles ofr(s) - y,.(s) have negative real parts; hence we have
r(t)-y,.(t)->O as t->oo. Because of linearity, we have y(t)=y)t) +y,.(t) and
e(t) = r(t) - y(t)->O as t ........ oo. This establishes the theorem.
Q.E.D.
The design procedure developed in this proof consists of tvvo steps: ntrc
duction of 1/4>(s), a rnodei of the reference and disturbance signals, inside ihe
loop, and the stabilization of the feedback system by introducing the compen
sator Nc(s)/Dc(s). The duplication of the dynamic or rnodel, 114>(s), inside the
loop is often referred as the internal model principie. As will be discussed later,
rernain to have n
is not permitted.]
The robustnes
case where r(s) = 1
with transfer fune
A
g
and consider the 1
Figure 9-20
Figure 9-21
Asympll
(9-95)
(9-96)
hence we have
y(t) = yw(t) + y,.(t) and
n.
Q.E.D.
:s of two steps: intro
lce signals, inside the
oducing the compen
ldel, lN(s), inside the
'1i1l be discussed later,
491
+ NJs)N(s)
remain to have negative real parts. [Note that the perturbation of <jJ(s)
is not permitted.] Hence this design is insensitive to pH"ameter perturb2c
lion ana is said io be rooust.
The robustness is due to the presence ofthe internal model. We use the
case where r(s) = l/s and w(s) = O to ilIustrate this point. Consider a plant
with transfer function
Nnsn+NII_sn-
g(s) = Dns" +Dn_sn 1
A
+
+
+Ns+N o
+Dls +D o
and consider the feedback system shown in Figure 9-21, where a constant
v(t)
y(t)
y{t)
Figure 9~21
492
.
.
+ NoN eo )
+(DoDeo+NoNeo)
and
where a ~ O and
perturbation or
from r to e in Fi
t-+ ce
1-
kNoN eo
DoD eo + N oN eo
(9-97)
If C(s) contains the internal model l/s and if s is not a zero of g(s), then we
have Deo =0, Neo f.O and N O f. O. In this case, if k = 1, then r(1) - y(t)->O
as t -> oo. As long as Deo = O, Neo =1= Oand 9(s) remains to be asymptotically
stable, we always have r(t) - y(t)-> O as t-> 00 for all parameter perturba
tions of g(s) and C(s). Hence the design is robust.
Now we assume that g(s)C(s) does not contain the internal mode! l/s.
Then we have Do f.O and Deo f.O. In this case, if we choose k in (9-97) as
s
with
then we have r(t) - y(t) -> O as t -> oo. In order to have a finite k, we need
No =1= Oand Neo =1= O. In the design of compensators, it is always possib1e to
find a C(s) with Neo =1= O. Ir Do =/=-0, we have No =/=- O if and only if <fJ(s) = s
is not a zero of g(s). Hence we condude that, under the condition stated in
Theorem 9-20, ifr(s)= l/s, il is also possible to design a feedback system,
without introducing the internal model, to achieve asymptotic tracking.
This design, however, is Dot robust. Ir there are perturbations in ?ny ;."
No, Do, Neo, and Deo, we do nol have r(t)-y(t)->O as t->oo. Hence the
design which does not employ the internal model principIe is not robust.
We emphasize that the condition that no root of </J(s) is a zero of 9(S) is
needed in both robust and nonrobust designs.
5. From the discussion in item 4, we may condude that the condition stated in
Theorem 9-20 is necessary as well. If we introduce the interm!.l model and
if anoot of <fJ(s) is a zero of g(s), then the root, which is unstable, becomes a
hidden mode and will not be affected by any compensation. Hence the
unity feedback system can never 'be asymptotically stable. Ir nointernal
model is employed, even though g(s) is of the form </J(s)(s)/D(s), it is straight-
forwd 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
tracking and disturbance rejection cannot be achieved.
k =
We see that k 1 i:
exact, then k 1 is 2
Ir k 1 is nonz~
k ea'. If a is po>
output y(t) will r
track r(t) but wit
with m > 1, and i
again not possibl.
This, howevet
interest. In the d
result will be evet
this case, if the inl
errors in impleml
but with a finite S1
of <fJ(s), the smalh:
7. The design of N k
of the polynomia
Under theconditi
-(DoD co +NoNco )
=
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
~
1 ~
r(s) = - - ro(s)
s-a
and
(9-97)
493
<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
g,.(s) =(s -a +e)g,.o(s)
where all poles of g,.o(s) have negative real parts. The signal e(s) due to
r(s) and w(s) =0 is equal to
~
~
1 ~
e(s) = (s - a +e)g,o(s) - - ro(s)
s-a
=--
with
'e a finite k, we need
is aIways possible to
f and only if <f;(s) = s
e condition stated in
n a feedback system,
lsym ptotic tracking.
turbations in any of
s t----> oo. Hence the
nciple is not robust.
;) is a zero of g(s) is
e condition stated in
: internal model and
unstable, becomes a
nsation. Hence the
able. If no internal
(s)/D(s), it is straight
t appear as a numer
?; hence asymptotic
(s - a +e)
s-a
~
I
g,o(s)ro(s)(s-a) s=a =eg,.o(a)ro(a)
A
+ Nc(s)N(s)
494
differential eguat
discussion of this
use Theorem 9-10, the degree of DJ~) is deg D(s) +deg cjJ(s) -1 and the total
degree of the compensator N /s)/Dc(s)cjJ(s) is deg D(s) + 2 deg cjJ(s) - 1. A
different procedure is to solve the eguation
DJ(s) = Dc(s)D(s)
+ N c(s)N(s)
with the constraints that Dc(s) contains the factor cjJ(s) and Nc(s)/Dc(s) is
proper. By so doing, the degree of compensator can be considerably
reduced. See References S34 and S238.
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
y(s) = D - 1 (s)~;(s)
+ D - I (s)N(s)u(s)
x=
Ax
+ bllu(t)
YlI = ex
Ax
+ bww(t)
y=ex
(9-99)
x=
Ax
+ bllu(l) + bww(t)
y=ex
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
generalized to
x=
Ax
and
where D,.(s) and D w (!
q x 1 polynomial ma
any N,.(s) and any N,
(9-98)
x=
'Multivariable ca~
system shown in Figl
matrix G(s) factored f
q xq and q xp poly
signal r(t) and the q )<
en.(s)
or
+ bllu(t) + b",w(t)
,.
+ ~-----Jr""
~L1Jll'IlS;
, I
lb---
Figure 9-22
Multiv8
L-__,
Figure 9-23
Placem(
495
differential equations such as the ones in (9-93) and (9-94). For an extensive
discussion of this type of disturbances, see References S121 and S122.
r(s) = O,:-l(s)N,(s)
w(s) = O,~ l(s)N,)s)
and
_ 1s"
as
where O,(s) and O",(s) are q x q polynomial matrices and Nr(s) and Nw(s) are
q x 1 polynomial matrices. The problem is to find a compensator so that, for
any N,(s) and any N,)s),
e(t)
(9-98)
"
Figure 9-19 can be
fhis equation can be
(9-100)
(9-101 )
(9-102)
01'
) +eww(t)
-------1
[
: type of disturbance
)ntrol systems. One
3.nsient performance.
nse as t -> co. Hence
mg to this type of
01' the response right
isturbance. Typical
. They are governed
:::ms 01', equivalently,
I between these poles
d. For a discussion
nces can be roughly
bances. The former
1 stochastic control
:::1' is describable by
'-if~"'"''''''
f----------'-___/)
: i \
iI '
Ii - - - - - - - - - - - - - - - - - - 1I
.
Figure 9-22
~(S)
~
(.n
y.(S)
~.~~~-I
Figure 9-23
496
if no root of cjJ(s) is a
(9-103 )
Although there is D,~ I{S) inside the parentheses, because ofthe noncommutative
of N{s) and D,~ 1 (S), it cannot be brought out of the parentheses to cancel D,~ 1 (s).
Hence there is no guarantee that the internal model D,~ 1 located as shown will
accomplish the disturbance rejection.
Now we shall place the internal model D;;: l(S) at the input of the summing
point where w{t) enters the system. In other words, we interchange the positions
ofthe blocks D;;: l{S) and N(s) in Figure 9-23, or replace N(s)D;;: 1 (s) by D,~ 1 (s)N(s)
in (9-I03). Then we have
y".{s) = Dc{s)[D(s)Dc(s) +D;;: l(s)N(s)N c(S)] -1 D,~ 1 (s)Nw(S)
= Dc{s){D;;: 1(s)[Dw(s)D(s)DAs) + N(s)Nc(s)J} -1 D;;: l(s)N w (s)
= Dc(s)[D,..(s)D(s)Dc(S) + N{s)Nc{s)J - l Dw(s)D;;: 1 (s)N",(S)
(9-104)
ran
k[IJ-A
-c
where {A, R, C, E} is
of A or the degree of
Proof
Using Theorem 9-4.
lowed by C(s) = DI"':
trollable if and only
Since D(s) and N/(s)
is always controllabh
The tandem conr
N 12 (s) ~ N/(s) are ri
matrix
Dw(s), then the modes of w(t) are completely canceled by the internal model.
Therefore, in the multivariable case, the position of the internal model is very
important; it must be placed at the summation junction where the disturbance
signal enters the loop. In the scalar case, because of the commutative property,
the position of the internal model is not critical, so long as it is not in the forward
paths from r and e and from w to y.
By a similar argument, in orderto achieve asymptotic tracking, we must
place D,:- fes) at the input of the summing point where the reference signal r
enters the loop. In other words, we must place D r- 1 (s) after D- 1 (s) or in the
feedback path. Ir the internal model is required to locate inside the compen
sator shown in Figure 9-22, then it must commute with N(s) to achieve distur
bance rejection and commute with N(s) and D- 1 (s) to achieve asymptotic
tracking. For a general q x p rational matrix C(s) = D- 1 (s)N(s), the internal
model which has these commutative properties must be very restricted. Ir the
internal model is chosen of the form a(s)I, where a{s) is a polynomial and 1 is a
unit matrix, then it has all the commutative properties. Furthermore, these
commutative properties hold even if there are parameter perturbations in
C(s). Hence, by choosing the internal model of the form a{s)l, the design will
be robust.
Consider the reference signal and disturbance signal given in (9-100) ancl
(9-101). Let cjJ(s) be the least common denominator of the unstable poles of
every element of D; 1 (s) and D;;; l(S). Let C(s) = D-l(S)N(s) = N(s)D- 1{s),
where D(s) and N/(s) are left coprimeand N(s) and D(s) are right coprime.
Then is called a transmissian zera of' C{s) ir rank N() < min (p, q) or rank
N/(X) < min (p, q). See Appendix H.
Theorem 9-21
The taQdem connection of cjJ -l(s)l p followed by the q x p proper rational
matrixC(s)( C(s) followed by cjJ - 1 (s)I q ) is controllable and observable if and only
Figure 9-24
M ullivari
(9-103)
fthe noncornmutative
leses to cancel O,~ 1 (s).
located as shown wil\
rank
IJ-A
[
-c
:]=Il+P
(l1+q)
01',
497
equivalently,
(9-105)
(s)N,.(5)
N",(s)
(9-104)
lX P proper rational
observable if and only
Using Theorem 9-4, the tandem connection of lp(cP(s)I p)- 1 = (cP(s)I p)-1 I p fol
lowed by (;(s) = O-I(s)N(s), where O(s) and N(5) are left coprime, is co,,
trol\able if and only if Ods) ~ O(s) and N1Z(s)N,. ~ N1(s)I p are left coprime.
Since 0(5) and N(s) are left coprime by assumption, the tandem connection
is always controllable.
The tandem connection is observable if and only if 011(S) ~ (cP(s)I p) and
Nz(s) ~ N(s) are right coprime 01', equivalently, the (p +q) x p polynomial
matrix
M(s)
[cP(S)I p]
(9-106)
N(s)
has rank p for every s in iC (Theorem G-8). Ir A is not a root of cP(s) = 0, then
cP(I~) i= 0, and rank M()~) = p. If), is a root of cP(s) = O, then <f(A) = O and rank
M(A)=rank N(A). Hence the tandem conn~ction is observable.ifand only.if
rank N(A) = p. By Oefinition H-3, rank N(A) = p if and only if A is not a trans
mission zero of G(s). This together with Theorem H-6 completes the proof of
this part of the theorem. The proof of the tandem connection of (;(s) followed
by [<f(s)I q ] - 1 is similar and is omitted.
Q.E.D.
With this result, we are ready to develop the condition for the existence of
compensators to achieve tracking and disturbance rejection. The condition
is a generalization of the scalar case in Theorem 9-20.
Theore-n 9-22
Consider the feedback system shown in figure 9-24 where the plant is com
pletely characterized by its q x p proper rational matrix (;(s). It is assumed
that the reference signal r(t) and the disturbance signal w(t) are modeled as
w
Figure 9-24
498
res) = D,:-I(S)N,.(s) and w(s) = D;;; 1(s)Nw(S). Let <jJ(s) be the least common
denominator of the unstable poles of every element of D,:-I(S) and D\~ I(S). If
no root of <jJ(s) is a transmission zero of G(s) and p q or, equivalentIy,
rank
[JJ--CA BJE
= n +q
there exists a compensator with a p x q proper rational matrix such that the
feedback system is asymptotical1y stable and achieves asymptotic tracking
and disturbance rejection.
Proof
Let G(s) = D- 1(s)N(s) be a left-coprime fraction and let <jJ -1(s)l q be the internal
mode\. The compensator Nc(s)D c- 1(s) is to be designed to stabilize the unity
feedback system. From Figure 9-24, we have
e(s) =r(s) - D- 1(s)N(s)NcCs)D; l(s)(<jJ(s)l )-le(S)
q
(9-107)
(9-108)
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
we have deg [<jJ-l(s)lqD-l(S)N(s)] =deg [<jJ-l(s)I q ] +deg det D(s)(Theorem 9-1)
which implies that <jJ(s)D(s) and N(s) are left coprime. Consequently the roots
of the determinant of
D (s) ~ <jJ(s)D(s)De(s)
+ N(s)Ne(s)
Since cjJ(s) appearsas zeros of every element of Ge,.(s) and Gew(s), ll unstable
poles of D,:-I(S) and D;;; I(S) are canceled by <jJ(s). What remains in e(s) are all
stable poles. Hence, for any Nw(s) and N,.(s), we have e(t)-->O as t--> 00: This
proves the theorem.
Q.E.D.
of cP(s)
(9-107)
ed as
sn
-1
(9-108)
~n
9).
left-half plane by a
Hence the unily
lence no cancellation
ots of cP(s) will occur.
elernenl of Ce,.(s) and
. w(t) can be written as
499
The design consists of two steps: introduction of the internal model cP -1(s)lq
and stabilization of the feedback system by the compensator N c(s)D c- 1 (s).
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
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
established in Theorem H-7, for any initial state these unstable modes will not
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
(;(s) and aH of cP(s) must be the blocking zeros of C ...(s) and CeAs).
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
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,
wil1 be generally not robust. Consider Figure 9-22. Define (;(s) = C(s)C(s) =
D -1(s)N(s)NcCs)D c- 1 (s) O-I(s)N(s), where O(s) and N(s) are left coprime.
Then we have
c(s) = [1
We assume, without loss of generality, that aH poles of 0,:- 1 (s) have nonnegative
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
that D(s) = 6(s)O,.(s) (see Problems 9-19 lo 9-21). Henc~ asymptoOr. tn3cJdng
is possible if and only if a compensator C(s) can be found such that
C(s)C(s) = 0- l(s)N(s)NcCs)Oc-1(s) = (D(s)O,(S))-1 (s)
=0,-I(s)D-1(s)N(s)
(9-110)
This is consistent with the statements that thc internalmodel, O,:-I(S), must be
placed at the point where the reference signal enters the loop. We see that, in
this design, we require only one copy of the dynamic of the reference signa!.
However, the computing of C(s) to meet (9-110) is complicated. Even if such a
C(s) is found fora given C(s), there is no guarantee that O,:-I(S) can still be
subtracted and moVed to the leftmost posilion as in (9-110) if there are per
turbations in C(.~) ..Herice this c1esign will not be robust. In contrast with this
design, the design procedure of employing cP - 1 (s)lq asan internalmodel is
straightforward and the design is robust. However, the degree ofcompensators
is much larger.
500
(9-111 )
A
k
s
u(s) =- +(t
Hence we have, by u:
(9-112 )
<jJ(s)O(s)Oc(s) + N(s)Nc(s)
can be arbitrarily assigned (Corollary 9-19). Furthermore sorne roots of <jJ(s)
will appear as unassignable roots, and there is always cancellation between
<J(s)O(s)Oc(s) + N(s)Nc(s) and 4) (s). Hence not all roots of <jJ(s) will be the
blocking zeros of Ger(S) in (9-112). Consequently, if q > p, asymptotic tracking
is not always possible. A similar remark applies to disturbance rejection.
We use a different argument to establish once again the necessity ofp ?:.q.l J
18
lirr
We see that, unless 1
achieve asymptotic f
asymptotic tracking,
equal to the q-dimen:
q x p matrix, in order
of plant inputs must
Using this argurr
Theorem 9-22. If r(
tracking, we need ran
for every root ). of ,
and if }, is not a pole e
is not a transmission
Static decoupling
feedback system shO\
terized by its q x p prc
~1t l~Jr~J.i -?{.J'} == ;~:" ..
reference signals are s
output as t -HfJ is cal
application of r(t) is
matrix from r to y t
Figure 9-25
Figure 9-26
Design (
G(J;~,~l
ofthe condition
L+3J
:s)J
-1
501
It is a plant with one input and two outputs and is BIBO stable. Let r(s) = ro/s.
No matter how the feedback system is designed, in order to have step functions
at the outputs, the input to the plant must be of the form
(9-111)
k
s
(9-112)
t-oo
We see that, unless ro is inside the range space of (;(0), it is not possible to
achieve asymptotic tracking. Hence for any ro in IRq, in order to achieve
asymptotic tracking, it is necessary to have the range space, gjl(G(O)), of G(O)
equal to the q-dimensional real space (IR q, IR) ol' rank (;(0) = q. Since (;(s) is a
q x p matrix, in order to have rank (;(0) = q, we need p ?:.q; that is, the number
of plant inputs must be greater than or equal to the number of plant outputs.
Using this argument, we can now establish intuitively the conditions in
Theorem 9-22. Ir r(s) = ro/(s - A) with A?:.O, in order to achieve asymptotic
tracking, we need rank (;(..1.) = q. For the general case, we require rank (;(..1.) = q
for every root A of </>(s). Ir we use the copl'ime fraetion G(s) = N(s)D- 1 (s)
and if Ais not a pole of (;(s), then rank (;(..1.) = q if and only if rank N(A) = q, or }.
. is not a transmission zero of (;(s). This is essential\y Theorem 9-22.
)D(s)] -1 N(s)
LUse 4> -1(s)I p followed
r l(s)N(s)4>-I(S)I p )
[-O.5Jk
et 4>(s)D(s) is equal to
> p,
larger than
t left coprime. Con
[.
~
-~-~
Figure 9-26
L.
502
response due to r(s1= ds- 1 can becomputed as, by using the final-value theorem,
lim y(e) = lim sG(s)((s) = lim sG (s)ds- =
s-o
t-cc
s-o
G(O)d
(9-113)
t-
00
[r(ei - Y()]
i = 1, 2, ... , q
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
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
dies out, it will cause only a change at the ith output and no change at other
outputs. Hence in a staticalIy decoupled system, only the steady-state responses
are decoupled, but not the transient responses. This differs from the decoupled
system discussed in Section 7-6 where all the responses, transient as well as
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
of a room are this type of reference signals. Maintaining an aircraft at a fixed
altitude is another example.
Asymptotic tracking actually achieves decoupling as e---+ 00; the steady state
of Yi(e) tracks r(e) and is independent of r ie), for j i= i. Hence the design for
asymptotic tracking can be directly applied to static decoupling. In this case,
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
will appear as a zero of every element of (;er(S), the transfer matrix from r to e.
Hence we have (;er(O) =0 and
t-
discussed in Figure S
hence we set P = 1 in]
We find a C(s) so tha
= Hm
s-o
G(s) = 1 -
and
G(O) becomes
Because of e(s) = r(s) - y(s) or Y(s) = r(s) - (;e,.(s)r(s), the transfer matrix,
[rom r to y is equal to
Hence we have
ri
--
"
G(s),
(;er(S)
:final-value theorem,
(9-113)
--> OC! ;
503
Since all the potes of D j I(S) have negative real parts, the constant matrix D(O)
is nonsingular. If s is not a transmission zero of G(s), then rank N(O) = q.
Now we may design a compensator C(s) so that rank Ne(O) = q and the q x q
constant matrix N(O)Dj I(O)N e(O) is nonsingular. Under these assumptions,
we may choose
(9-115 )
s statical\y decoupled.
gonal, has s as a zero
:sign is robusto That
turbations, even large
LS the feedback system
ve introduce a design
e single-variable case
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
{A, B, C, E} is any irreducible realization of G(s).
.
504
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
X,.(t) = A,.x,.{t)
r(t) = C,.x,.(t)
with some unknown initial state. Let </>,)05) and </>,.(s) be the minimal poly
nomials of A w and A,., respectively, and let
be the least common multiple of the closed right-half s plane roots of </>,.(05) and
</>,.(05). Thus a11 roots of </>(05) have nonnegative real parts. The internal model
</>-l(05)I q can be realized as
Xc = Aex e + 8 ee
Ye =X e
q-tuple
~
with
lJ..
,11
[ -(
th~
~
'-------v---'
ro
Design of r
rank V(s)
where
Figure 9-27
V(s) =
(9-116 )
[~n o
(11
+mq) )
505
,q X 11, q X p, 11 X q. and
lmtrollable and {A, C}
l1al w(t) is generated by
BJ =/l +q
E
(9-117)
O
:
:
si - A c
B
B
= n +mq
cE
The first factor has tank iz + mq by the irreducible realization of </J -1(s)I q The
second factor has rank 11 +q +mq at every root of </J(s),by the assumption of
(9-117). Hence 5ylvester's irtequality (Theorem 2-6) implies :
rank V(s)2.(n +mq) +(n +q +mq) - (n +q +mq) = n +mq
Since V(s) is an (n
(9-116 )
+q +mq
+ mq) x (n
506
Co(s) = L(s)
DAs)
(9-119 )
M(s)
C(s)=-
DAs)
(9-120)
Because of this restriction, the feedback system in Figure 9-l(b) can be redrawn
as shown in Figure 9-28(a). if deg DJ~) = m, the ;( 2 [",liana! ,{lB."'>
[Co(s) C I(S)] is of degree In and can be realized by using In integrators. Ir
the configuration in Figure 9-28(a) is reduced to the single-loop system shown
in Figure 9-28(b), then the denominators of the two compensators are different.
Hence their implementations require twice as many integrators as the one in
Figure 9-28(a). Hence, although the two configurations in Figure 9-28 are
equivalent mathematical\y, they are different in actual implementation.
The transfer functionfrom r to y in Figure 9-28(a) can be computed as
9 (s) =
9(S)
1 + Co(S) + C I(s)g(s)
. N(s)De(s)
DAs)D(s)
+ L(s)D(s) + M(s)N(s)
(9-121 )
.Define
or
+ M(s)N(s)
+ M(s)N(s)
Figure 9-28
Input-out
and
(a)
(9-122)
Theorem 9-23'
Consider a plant with tr:
For any DAs) of deg
proper Com pensators L
Figure 9-28(a) has tram
D(s) and N(s) are coprirr
11.
Proof
For any DAs) of degree l
D (s) - DAs)D(s) is of de
.;
~ry s in
e and (9-116) is
507
(b)
(3)
Figure 9-28
put
(9-123)
arbitrary denominator
Jensators to achieve the
lfiguration is deyeloped
iscussed in Chapter 7.
,t the case C(s) = 1. The
le same denominator as
(9-119 )
(9-120)
Theorem 9-23'
Consider a plant with transfer functiong(s) =N(s)/D(s) and deg N(s).:::;;deg D(s) =
n. For any DAs} of degree m and any D(s) of degree n +m or less, there exist
proper compensators L(s)/Dcls) and M(s)fDe(s) so that the feedback system in
Figure 9-28(a) has transrer function N(s)D l(s)D e(s) from r to y if and only if
D(s) and N(s) are coprime and m '2. n -1.
Proof
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
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 matrz
using m integrators. 1f
ingle-loop system shown
mpensators are different.
IOtegrators as the one in
:ions in Figure 9-28 are
1 implementation.
can be computed as
IDe(s)
JD(s)
N(s)
N(s)
(9-121 )
+ M(s)N(s)
(9-122 )
We compare first this theorem with Theorems 9-1 1 and 9-11'. First, in
order to ensure the praperness of the compensators, we m lIst consider separately
in Theorems 9-11 and 9-11' the cases where the plant is proper or strictly praper
and require the degree of O /(s) to be exactly equal to n +m. Since the D,(s) in
Theorem 9-23' is chosen a priori to haye degree m, the compensators are always
proper whether D(s) is of degree n + 111 or not. Second, the oyerall transfer
function of the unity feedback system in Figure 9-12 is N(s)D l(s)N ,(s), where
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
D(s) as well as De(s). The reason for haYing this extra freedom can be seen from
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
three sets f paraineters Dei' L, M j. Since treq lIires only two sets ofparameters
S08
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
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
D(s)=D o +Ds +
N(s)=N o +Ns +
Dc(s)=D co +Dcs +
L(s)=L o +Ls +
M(s)=Mo+Ms+'"
D(s)=F o +Fs + ...
E(s) = E o + Es + . _.
and
Figure 9-29
+D"s"
+N"s"
+Dcmsm
+Lmsm
+Mms m
+F"+ms"+ m
+ E Il + msrz+m
(9-125)
[Lo
M o : L
: ... :
L m MmJSm = [E o
...
EIl+mJ ~ E
An input
(9-126)
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
mediately the proper compensators Co(s) and C(s).
Example 1
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,
and D(s) = S2 + 2s + 2. Note that the degree of D(s) is smal!er than n + m = 3.
We compute
Theorem 9-23
Consider a plant with 1
n. For any Dc(s) of dI
compensators L(s)/D c(:
9-28(a) is well posed an
only if D(s) and N(s) al
We see that by re
then assert the wel! pos
feedback system is wel
from O at s = OC! if deg
De l(s)D f (s)[;
and form
[Lo
Mo: L
-1J
D; (s)D(s)D-(s)
Its solutions are Lo =
pensators are
6, L =
1, M 0=
2 and M
= 6.
Co(s) = L(s) = -6 -s
D/s)
1 +s
and
Hence if FIl+m=O or (
the system is not well]
1 +Co(oo) +C(oo)g(oo
. note that if deg D(s) < j
We consider now a
.. Corollarv 9-23
Consider a plarit with tr
/1.
For any i5 (s) of d(
509
Ing of (9-122) or
I;)N(S)
(9-124)
Iw be translated into a
Figure 9-29
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.
(9-125 )
En+mJ ~ E
(9-126)
Theorem 9-23
,aS
Consider a plant with transfer functiong(s) = N(s)/D(s) and deg N(s):::; deg D(s) =
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
We see that by removing deg DAs) < n +m from Theorem 9-23', we can
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)
= [2
4 2
- 1]
(9-127)
Hence if Fn +m=0 or deg DAs) < n +m, then 1 +Co(oo) +C,(oo)g(oo) =0 and
the system is not wel! posed. However, if F II + m 1=0 or deg D(s) = n +m, then
1 + Co( 00) + e, (oo)]( 00) 1= O and the system is wel! posed. It is of interest to
note that if deg DAs) < n +m, the pole-zero excess jnequality in (9-38) is violated.
We consider now a speciai case ofTheorem 9-23.. :
Corollary 9-23
Consider a plant with transfer function g(s) = N(s)/ D(s) and deg N(s) :S deg D(s) =
n. For any i5 As) of degree n, there exist proper compensators L(s)jD/s) and
510
M(s)/Dc(s) of degree n - 1 and with arbitrarily assignable poles such that the
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
DE~
system to minimize
This design always involves the canceIlation of Dc(s), which can be chosen by
the designer, however. The degree of Dc(s) in the coroIlary can be larger than
n - 1; however, it does not seem to serve any design purpose because DcCs) is
completely canceled in the designo
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
the poles of g(s) = N(s)/D(s) as the roots of i5 As) without affecting N(s). An
(n -1)-dimensional state estimator with arbitrary eigenvalues can be con
structed to generate an estimate of the state. The connection of the feedback
gain from the output of the estimator yields the overal1 transfer function
N(s)ji5As) (see Section 7-5). Hence Corollary 9-23 establishes essentially the
result of state feedback and state estimator. However, the result in Corollary
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
or dynamical equations wi.th the direct transmission parts equal to zero. In
Corol1ary 9-23, we require deg i5(s) = deg D(s); in the design of state feedback,
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
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
1 +Co(oo) +CI(oo)(oo) #=0 for the class ofsystems studied in the state-variable
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
Gis) =
=
=
(;(s)[1
N(s)D
N(s)[I
Define
or
Figure 9-30
D
E
Input-out]
11
l by choosing D (s) =
:he eigenvalues of A or
: compensator Co(s) is
= [' {q[y(t) -
r(l)] 2
(9-128)
~ (s) = ~(s)
D(s)
where 15(s) is a Hurwitz polynomial and has the same degree as D(s). The
N(s) in the numerator of 9(s) is the same as the numerator of the plant transfer
function 9(s). See Reference S46. Hence the design of the optimal system
requires the solution of Corollary 9-23.
Multivariable case. In this su bsection, the results in the single-variable case
wil1 be extended to the multivariable case. Consider the input-output feedback
system shown in Figure 9-30. The plant is described by the q x p proper
rational matrix (;(s) = N(s)D- 1(s). The compensators are denoted by the
p x p proper rational matrix Co(s) = D; 1(s)L(s) and the p x q proper rational
matrix C 1(s)= D;-I(S)M(s). The transfer matrix from r to y in Figure 9-30 can
be computed as
(;(s) = (;(s)[1 +Co(s) +C 1(s)G(S)]-1
Define
D (s) =Dc<s)D(s) + L(s)D(s) +M(s)N(s)
E(s) ~ D(s)- D,(s)D(s)= L(s)D(s) + M(s)N(s)
or
l~ry
+ u 2 (t)} dl
where q> O is a weighting factor and r(t) is the reference signa\. lf r(l) is a step
function. then the optimal system which has the smallest J is of the forro
1= N(s)i5 1(S)
Ihich can be chosen by
lary can be larger than
urpose because D/s) is
5U
Figure 9-30
(9-130)
(9-131 )
512
Then
DI
Define
H(s)= diag
[S"I,
su', ... , SU
S111"
,Slllp}
(9-133 )
(9-134 )
input-olltput feedbae
po sed.
In the design, tht
chosen as
Theorem 9 c 24
lim R:-1(s)D(s)H-I(s)=J
s~
(9-135)
00
exists and is nonsingular, there exist compensators with proper rational matrices
D; 1(s)L(s) and De-I(S)M(s) such that the feedback system in Figure 9-30 is well
posed and has transfer matrix N(s)D 1(s)D,(s) if and only if D(s) and N(s) are
right coprime and D(s) is column reduced.
Proof
Proof
00
where D,,, is the row degree coefficient matrix of D,(s) and D" is the column
degree coefficient matrix of D(s). Because J - O,,,D / exists, we conclude from
the proof of Theorem 9-18 that solutions L(s) and M(s) of row degrees at most
mi exist in (9-131). Consequently D,-I(S)L(s) and D,-I(S)M(s) are proper fol
lowing the assllmption of DcC.s). Note that whether J - D c " D h is nonsinglllar
or not is immaterial here.
We show that the system is well posed by showing that !+Co(co)+
el (co )C( ex)) is nOl1singuiar. F rorn (9- dO j, W.; navc:, sirnia LV Ci- iL."/j,
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)
= 1 + Co(s)+ C(s)C(s)
(9-137)
We write, similar to (9-83),
D(s) = [D /I + D/(s)]H(s)
De(s) = H,(s)[D eI, + DcI(s)]
and
This cstablishes the ca
The applic3.tion o
\r!t, use Le coed
shown in (9-75). We t
from top to bottom. I
block row of S, are lir
convenience, we assum
r,
lO; L
and
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
s-+ 00
(9-138)
(9-139 )
Corollary 9-24
(9-140)
which is nonsingular by the assumptions of O,(s), D(s), and (9-135). Henee the
L(s) =L o + L
M(s) = M o +~
E(s)=O/s)
[Lo
Mo : L I
MI
513
(9-132)
input-output feedback system in Figure 9-30 is, following Theorem 3-6, well
posed.
Q.E.D.
In the design, the denominator matrix De(s) of the compensators can be
chosen as
(9-133)
(9-141 )
(9-134)
O!'
al1
mi are odd integers, then we may not be able to assign complex conjugate roots.
Let
v be the row index of
)e(S) of row degrees mi
(9-135 )
In this case, if we choose DcCs) to be of the form shown in Problem 2-29, then
the difficulty of assigning complex conjugate roots will not arise.
Corollary 9-24
Consider a plant with q x p proper rational matrix G(s)= N(s)D-'(s). Let
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
fi (s) of column degrees Ili, i = 1, 2, ... , p, and column reduced, there exist
compensators with proper rational matrices D; '(s)L(s) and D; l(s)M(s) such
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
IH - '(s)[D" + D(s)]-'
(9-140)
(9-143)
Q.E.D.
(9-138)
(9-139)
(9-142)
arid
L(s)=Lo+L,s+'" +L._,s-'
M(s)=M o +M,s + ... +M._,s-'
E(s) = D (s) - DcCs)D(s) =Eo + Els + ... + El' h ' _ S" +v -l
(9-144)
of (9-144) to
(9-146) into
Mo:L l
Ml:"<L'- 1
M,-,]S'-l=[E o El
...
(9-145)
(9-146)
El'h-l]
(9-147)
514
DI
The solution of this set of linear algebraic equations yields the required com
pensators.
It is ofinterest to compare the result in Corollary 9-24 with the one developed
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) =
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
9-24. If the state is not available for feedback, we may design a state estimator
with arbitrary eigenvalues, which are equivalent to the roots of det Dc(s), to
generate an estimate of the state. Now the application of the state feedback
from the output of the state estimator yields N(s)Dj I(S). Note that the eigen
values of the estimator are not control1able as can be seen from (7-86) and will
not appear in the transfer matrix from r to y.
A remark is in order regarding the degreesof compensators. In the state
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).
Note that the row index v of (;(s) is equal to the observability index of any
irreducible realization of (;(s), and has the property v 2.n/q. If p =q = 1, then
n = v and the dimension of the state estimator is equal to the degree of com
pensators. Ir p 2.q, then p(v -1) 2.n - q; if p < q, p(v -1) can be greater than,
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
simpler, conceptual1y and computationally, than the one in the state-variable
approach.
In the fol1owing, the input-output feedback system in Figure 9-30 will be
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
G(s) = (;(s)Q-l(s)[I +C O(S)Q-l(S) +C 1(s)(;(S)Q-l(S)]-1
Ir we define
and
E(s) ~
is a proper rational
D c-l(S)M(s), and Q
posed and has trans
computed from
(9-148)
R
Proof
Since D 11 + D(s)and D
is proper.
515
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
(9-149 )
(9-150)
Ir we define
and
(9-151)
then Gj(s) = N(s)Dj l(s)Dc(s) and has the same form as (9-132). The design in
Theorem 9-24 to achieve (9-132) is accomplished without using Q-I(S); how
ever, the row degrees of Dc(s) are general1y difTerent. In the fol1owing, al1 row
degrees of Dc(s) wil1 be required to be the same. This is possible because of the
introduction ofQ - 1(s). In this case, R(s) in (9-134) becomes diag {Slll, Slll, ... ,s"'}.
Theorem 9-25
Consider a plant with q x p proper rational matrix (;(s) = N(s)D-1(s). It is
assumed that D(s) and N(s) are right coprime and D(s) is column reduced and of
column degrees J, i = 1,2, ... ,p. Let v be the row index of (;(s). Let Dc(s) be a
p x p arbitrary polynomial matrix of row degrees all equal to 111 with m;::: v - 1
and be row reduced. Then for any D(s) with the property that
H(s)D (s)H,(s)
(9-152 )
p,
(9-153)
and
QI (s) = Dc(s)Q(s) + R 2 (s)
with o,. 1f2{S) < o,. Dc(s) = m, and L(s) and IVK(sj are soiutlons
(9-154 )
01
(9-155 )
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
(9-157)
Implementations
which implies
O; 1(s)O f(s)O - 1(s) = Q(s) + Oc- l (s)R 2 (s)+ Oc- 1(s)R(s)O -I(S) (9-158)
Because O(s) is column reduced and Oci O(s) > ociRI(s), RI(S)O-I(S) is strictly
proper. Because Oc(s) is row reduced and o,- Oc (s) > o.-iR 2 (s), Oc- l (s)1 is proper
and Oc- l(s)R 2 (s) is strictly proper. Hence the polynomial part of
O;I(S)Of(S)O-I(s) is Q(s). Since [Oc- I(S)Of(S)O-I(S)]-l is proper, it follows
from Theorem 3-4 that Q-I(S) is proper.
The feedback system is well posed if and only if the matrix
P- I(S)
T(
(9-160)
(9-161 )
which implies
Go(s)
where
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
proper, the polynomial part ofP(s) is I. Hence P- I(S) is, following Theorem 3-4,
proper. This establishes the well-posedness of the feedback system.
Q.E.D.
We remark the condition in (9-152). One way to check the properness of
H(s)O l(s)Rc(s) is by direct computation. The computation of the inverse 01
Of(S) is, however, complicated. Instead, we may compute
which can be obtained by inspection because of the forms of Rc(s) and H(s).
Clearly Y(s) is much simpler than D feS). Now H(s)O 1(s)Hc(s) is proper if and
only if y-I(S) is proper (Theorem 3-4). This is a simpler way of checking the
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
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)
Gfes) = G(s)Q-I(s)[1
19
(9-163)
Figure9-32
An open
517
:.1')
(s),
is strictly
)olynomial part of
1 is proper, it follows
atrix
'I(S)Q-l(S)] -1
~)] - 1 D,(s)
(9-159)
(9-159) yields
(S-15D)
(9-161 )
[)-I(S)Q-l(S)
~)Q-l(S)
(9-162)
(9-164)
y)H - (s)
(9-167)
19
This problem was firsl frmulaled and solved in Reference S218. This presentation fo\lows
Reference SSO.
(9-163)
Figure 9-32
An open-Ioopsystem.
518
or
(9-168)
where Dc(s) can be arbitrarily chosen. Hence this configuration may be used
to implement G(s)T(s) = N(s)D; l(s)N (s). In this implementation, we require
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)
is row reduced and U(s)D(s) and U(s)N(s) remain to be let coprime (Theorem
G-11). Thus we assume without 10ss of generality that N(s) is row reduced.
In the following, we present a procedure so th, G[(s) will implement
G(s)T(s).
Proof
The substitution of(
G(S) = 1
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
D-l(s)T(s) = D; 1 (s)N S)
(9-169)
where D (s) and N (s) are left coprime and N s) is row reduced. Let
DriN (s) = mi,
= 1,2, ... ,p
Let
.o/s) = diag {Cl. l (s), Cl.2(S), ... , Cl.p(s)}
where lXi(S) is an arbitrary polynomial of degnie m - mi' Then the matrix
Dc(s)
= .o/s)N(s)
(9-170)
j1i,
We discuss now .
depends on T(s). Ifl
det D(s) in (9-169) is
will also be BIBO s1<
det Dc(s) which are 1
Hurwitz, then the fee
The decompositi(
algorithms in Referel
design, the condition
(9-172) by
(9-172)
Q{
Ol" dy
COllstalil rnatr
(9-173)
with DcR(s):::;;DciD(s
example, from the un
[D1.+ D(s)]
(9-171 )
(9-174a)
(9-174b) .
Theorem 9-26
The input-output feedback system in Figure 9-31 with Dc(5), Q(s), L(s), andM(s),
cAomputed from (9-170) and(9-174) implements the open-Ioop system (;o(s) =
G(s)T(s) = N(s)D; l(s)N fes) and is well posed.
.
1mplementatioll 11.
519
(9-168)
Proof
The substitution of (9-170) to (9-174) nto (9-168) yields
(9-175 )
(9-176 )
ow reduced. Let
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,
the condition in (9-152) is satisfied and the theorem fol1ows directly from
Theorem 9-25.
Q.E.D.
)}
(9-169)
.mi'
uced.
(9-178)
... , P
(9-171)
(9-172)
101
(9-173)
decompositions are
s)N(s)
(9-174a)
s)N(s)
(9-174b)
LO
R(s) - BO(s)
Fol'
(9-179)
520
DE
Step 4. Define
Ir J: - f.J.. - In 2. O, ~
Hurwitz polynorr
b,.W(s) ~m.
Step 5. If
or
(9-181 )
(9-182)
(9-183)
W(s
where DJ(s) and N J(s) are left eoprill).e. Furthermore, we require DAs)
to be eolumn redueed and to have D Jh = 1 (the eolumn-degree-eoeffieient
matrix is the unit matrix).
Step 2. Let 6ci D J(s) = J, and beD(s) = Ji, i = 1, 2,
, p. Define
V(s) =diag {vI(s), V2(S),
, vp(s)}
Theorem 9-27
The feedback system i
the open-Ioop system
(9-184)
Proof
Ir ;?:. {ti' set Vi(S) = 1. If J < {ti' set vi(s) as a monie arbitrary H urwitz
polynomial of degree Ji - J. We then write
D -I(s)T(s) =
~fij1(S)N(s)
(9-Hl5)
where f> J(s) ~ V(s)D J(s) and NJ(s) = V(s)NJ(s). Let bef> J(s) = f. Beeause
of the assumption D JI. = 1, we may eonclude Ji ?:.Jii. 20
Step 3. Compute the row index, v, of (;(s). Let De(s) be an arbitrary polynomial
matrix of row degrees aH egual to In and row redueed sueh that m?:. v - 1
and D; 1N(s) is proper.
20
The sllbstitlltion. of JK
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
conclude ; ~Jl. However, in most cases, for example, when T(s)has a fult rank, we have v(s) = 1,
for al! i.
W(s)D J
521
Step 4. Define
W(s) = diag {Wl (s), W2(S),
>
>
Wp(s)}
(9-186 )
Step 5. If
i = 1,2, .. >, P
(9-187 )
(9-188)
)Q - (s)J - D; (s)K(s)
(9-181 )
K(s)
Dc(s)Q(s)
+ R 2 (s)
(9-189)
(9-182)
(9-190)
(9-183)
(9-191)
Theorem 9-27
The feedback system in Figure 9-33 obtained from (9-184) to (9-191) implements
the open-Ioop system G(s)T(s) = N(s)Dj (s)N(s) and is wel! posed.
Define
(9-184 )
ic arbitrar y Hurwitz
Proof
Thc substitulion of K(s) = W(s)N J(s), (9-186), and (9-191) into (9-182) yields
))-l(V(s)N(s))
CDJ(S)=!i. Because
I
"arbitrary polynomial
d such that m:?: v - 1
This shows that the feedback system in Figure 9-33 does implement the open
loop compensator T(s).
From (9-188) and (9-189), we have
W(s)D J(s) - Dc(s)Q(s)D(s) = R 2 (s)D(s) + W(s)R (s)
>
1 fun
1,
for al! i
(9-192)
(9-193)
Hence under" the assumption that D(s) and N(s) right coprime, D(s) column
reduced and m:?: v -l,solutions L(s)"and M(s) of degrees at most m exist in
(9 c I9l).Hencethe compensators D c- l(s)L(s) and D; 1 (s)M(s) are proper.
522
DE
Because D e- (s)R 2 (s) and R 1 (s)D- (s) are strictly proper and D;I(S)W(S) is
proper, the polynomial part of D e- 1(s)W(s)D f (s)D- 1(s) is Q(s). To show
Q-l(S) is proper, it is sufficient to show, by using Theorem 3-4, that
(D e-l(S)W(s)D f (s)D- 1(s))-1 is proper. Using (9-133), (9-134), (9-138), (9-139),
and similar formulas for D f(S) and W(s), we have
D(s)D 1 (s)W - 1 (s)De(s) = [D h+ Dt(s)]H(s)H [;,1(.1')[1 + D fl(S)] - 1
[1+ W(s)]-IH:;I(S)H e(S)[D eh + DcI(s)]
(9-195)
Because D(s), D ft(S), W(S), and Dcl(s) are all strictly proper, we have, as s-> 00,
D(s)D 1 (s)W - l(s)D e(s) -> DhH(s)H[;/ (s)H
w(s)He(s)D eh
(9-196)
Applications. 21
+ N ft(s)]
(\ N(s):5: J - Jt
The row degree m of Dis) is chosen so that D; l(s)N (s) is proper. Now we
claim that D e- 1 (s)W(s)N ((s) remains to be proper. If J - Il - m 2::0, brW(s) = 0,
and br;(W(s)Nf(S)) = briN'f(s) :5:b,.D e(s). If J - Il - m < 0, b,W(s) = Ili + m - J,
and we have bri(W(s)Nf(s)):5:Il+m-];+J-Il=m=briDcls).
Hence
D e- l(s)K(s) is proper.
We have shown that all compensators are proper. What remains to be
shown is that the overaH system is well posed. The overal1 syst,,:,n ,~ '1,,,,11
if and only if the matrix
P- I(S) ~ [1 + D e- \s)L(s)Q - I(S) + D e- 1 (s)M(s)N(s)D - I(S)Q - I(S)] = Q(s)D(s)[De(s)Q(s)D(s) + L(s)D(s) + M(s)N(s)] - 1 Dcls)
A remark is in or
In general, the degre
previous subsection.
the same denominatl
to deg det De(s). He
deg det Q(s) + deg de
previous subsection.
the implementation i:
The remarks in (
previous subsection ,
section and will not t
In addition to the
back implementation
may use the single-lo
feedback system in I
The basic idea and (
mentation 1; however
steps in (9-172) and (~
see Reference S218.
(9-197)
11
will be employed to dI
problem and then de
bance rejection. Fim
Decoupling. Conside
N(s)D- 1 (s). Ifan Op
where f(s) = diag {el
minimum degrees to r
the input-output feedl
Dj 1(.1'). We note that
of det (D f(s)N(s)); hen(
Since the input-outou
the disappearance oi ]
cancellations. Thus j
involve unstable pe
D(s)N -1(s)D 1(.1') can
We discuss in the
without involving an
nonsingular proper ra
right coprime and D(s
with
21
1
N
f> fl{S)] - 1
t-
,+ Dcl(s)]
(9-195)
)[N fh + fl{s)]
use T{s) is proper, we
;) is proper. Now we
m 2:0, <\W{s) = O,
o,. W{s) = j.l + m - ;,
m = 6,.Ok;).
Hence
U -
1,
What remains to be
11 system is well posed
:s)Q - I{S)] 1- 1 D c{s)
(9-197)
I{S)Q-l{S)
witb
N{s) == N{s)NzCs)
Q.E.D.
523
21
524
DE
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
N Z1(s). Define N2d(s)~diag {f321(S), f3ds), .. . , f32P(S)} and define
N2(S)~ N Z1(s)N 2d (S)
D(s)N 2 (s) =
L'r,
It is a rational matrix with only stable poles, that is, poles with negative real
parts. Then we have
The implementation
matrix
with
(9-198)
where IX(S) are Hurwitz polynomials of minimum degrees to make T(s) proper.
Then we have
lag
{f31(S) f32(S)
f3 p(S)}
( )' 1X ()
, ... , IXpS
()
S
1X 1 S
2
Implementation I.
V<
(9-199)
S2 +2.s
4
_l s
4
~lJ[s2;1
~T1
N(s)=Nl(S)N2(S)=I~
. . .J
Wecompute
_[
0,_,1ls,2
6d (D c (s)D r (s)),:::; Ji + 17
11
...J
1 [1
N 2 (s) = -------:
(s+l)(s-1) -1
Hence we have
N 2As) =diag {(s -1), (s -1)}
and
1
'N 2(s) = Ni (s)N 2As) = [_
(9-200)
~ ~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
compute
[-3//4 ~27/4
54
13/4
'
525
~J[ -~
)W
(s
+ 1)2
(9-201 )
, f3 zP(s)}
(9-198)
. {s - s- 1}
=dlag - -1. -
s+2's+1
s)N zAs)D,-I(S)
(9-199)
r G(s)T(s)
will yield a
Iy the cancellations of
system will be asym
l)}
(9-202)
-IJ[(S+2)(S+I)
O J-l
O
(s+l?
-i-s-~ J-l[ 1 -i-J~D-l(S)N (s)
SZ +lls +2
-1
s _.1 -
D-11)TI)=()D-"II)=[
1
S
\s \s
2
,
\s,
-1
=[SZ+~S+~
_l s +.1
4
sZ_
where D J(s) and N J(s) are left coprime. This is computed by using the pro
cedure discussed in Appendix G. The row index of G(s) can be computed as
v = 2. Clearly we have /11 = max {v - l,ml, /11z} = max {2 -1, 0, 1} = 1 and, by
choosing Dc(s) arbitrarily as Dc(s) = diag {s + 3, I},
Dc(s) = DC<s)NAs) =
[s
+3
0J[ 1 s-
-i- J
1
-1
1
'4
r3
+
which has row degrees all equaJ to m = 1 and is row reduced. Clearly we havc
c(f>c(s)DJ(S~Li+l11,i= 1,2, hence we may set Q(s)= l. We compute
.y factor N(s) as
~l
lj
Js
O O 1 O
O O O 1 O O
O . 1- 6- 0- -1- -
1
(9-200)
0--(-6--0--1--0
+1
1 1 1 1 O O
19 {CC1(S),<xZ(s)}, where .
1 1 O O
1 1 O O 1 O
000100
---------------------
+';' J
526
DE~
unstable pole-zero ca
as
L(s)= [
(3l/2)S+33/4
(-7/2)s- 2
27/4
-1/4
_ [( -31/2)s- 6
M(s)- (1/2)s+2
-9J
13/4
= v -1 = 1and
W(s)= [
~J
+2
s+2
K(s) = W(s)N (s) =
-1
3(s
+2)J
-~~
Its denominator is a 1
h(s)/</>(s) to be strictly
h{s); there are deg a{.
and #(s) are coprime,
proper choices of h(s)
s--,
-1IS-4
M(s) = [ (5/2)s + 2
-27/4J
9/4
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
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.
Asymptotic Tracking, DistUl'bance Rejection, and Decoupling . . 111 this .sub
section, we shall design a robust system foachieve decopling,asy1ptotic
tracking and disturbance rejection. Let C(s) = N(S)O-I(S) be a JI x JI 110n
singular proper transfer matrix: It can be decoupled, without involving any
r~
----/I~-
'? E
I
Figure 9-34
Oesign of
527
-9J
d.
2)s- 6
1-2
13/4
{~l(S) ~2(S)
)=
rbitrarily as s + 2, of
1=1- Hence we have
where ~i(S) are uniquely determinable from N(s) and the degrees of iX;(S) are
to be chosen to make the open-Ioop compensator T(s) proper. Ir the plant is
to be decoupled only, CY.(s) can be arbitrarily assigned. Ir the plant is to be
designed, in addition to decoupling, to achieve tracking and disturbance
rejection, (X;(s) cannot be arbitrarily assigned. They must be used to stabiJize
the additional feedback connection to be introduced for tracking and distur
bance rejection.
In order to achieve asymptotic tracking and disturbance rejection, and to
be robust with respect to parameter perturbations, we must introduce, as
discussed in Section 9-6, an internal model <f> -'(s)I p as shown in Figure 9-34(a).
Ir we introduce a diagonal polynomial matrix H(s) = diag {h, (s), h 2(s), ... , hp(s)}
with h(s)/<f>(s) proper or strictly proper, then the system in Figure 9-34(a)
reduces to p number of single-variable systems shown in Figure 9-34(b). The
transfer function of the system in Figure 9-34(b) is clear1y equal to
h;(s)~(s)
6. h(s)~(s)
<f>(S)iX(S) +h(s)/J(s)
-4
+-2
~p(S)}
[(s)
(9-203)
/4J
-27
4
91
I
compensators in the
19 canceled. Instead
, we may decouple it
pfillg. In this su b
:oupling, asymptotic
l(S) be a p xp non
ithout ivolving any
(a)
(b)
Figure 9-34
528
zero of (;(s), then <fJ(s) and f3(s) are coprime, and the roots of f(s) can be arbi
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
totical1y stable and is decoupled. It will also track asymptotically the reference
input and reject plant-entered disturbances. In this design, if there are param
eter perturbations in the plant and compensators (excluding the internal
model), the decoupling property will be destroyed. However, if the overall
system remains to be asymptotica11y stable, the property of asymptotic tracking
and disturbance rejection will be preserved.
Example 3
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
<fJ(s) = s(s - 2)
Since <fJ(s) and f3 (s) are coprime, h(s) and IX(S) can be found to stabilize the
system in Figure 9-34(b). Rather arbitrarily, we choose 11(S) =(s +2)3. Then
the solutions of
<fJ(SP1(S)
(s
+ 2)3
are h 1(s) =36s -8 and 1X 1(S) =s -28. Ir 12(S)=(S +2)4, the solutions of
<fJ(S)1X 2(S) + f32(s)h 2(s) = 12(S) = (s
+ 2)4
and
(;(s)T(s) =diag
O J=[S-28
1X2(S)
O
{s
1
-28'
s-
S2_1~2S-~J
s~:~ 1 124 1
--3-S--3-
loop compensator.
he solutions of
(9-204)
Now if a solution T(s) exists in (9-204) and is a proper rational matrix, then the
design can be accomplished by implementing T(s) in the input-output feedback
configuration shown in Figure 9-31 or 9-33. Thus the design problem hinges
on the solution of (9-204).
Let Ti(s) and Gmi(s) be the ith column of T(s) and Gm(s). Then (9-204) can
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,
be designed to track
form l, then we have
529
G",(s)J
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
small as possible. This problem of finding a proper T(s) with a minimal degree
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
is unimodular, that is, the R(s) in any factorization [A(s) B(s)] = R(s)[ (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,
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,
ir M(s) is 01' fun row rank in iR (S), [hefe ex;::;,::; 8. unimodular polynoiHici< maiX
U(s) such that U(s)M(s) is row reduced. We may similarly define column
irreducibility and column reducedness for nonsquare polynomial matrices.
With this pre!iminary, we are ready to study the minimal design problem.
We discuss first the condition for the solution T(s) in (9-204) to be proper.
Ir G(s) is square and nonsingular, the answer is very simple: T(s) is proper if and
only ir G-1(S)G m (s) is proper. Ir G(s) is not square, the situation is slightly
~ore complicated. Let 4J(s) be the least coml1)on denominator of al1 entries of
G(s) and Gm(s). The inultiplication of 4J(s) to (9-204) yields
A(s)T(s)=B(s)
22
(9-205)
invers~of ~(sj.
.
Hencetheinverse problem
530
where A(s) = </J(S)G(S) and B(s) = </J(s)G", (S) are q x p and q x r polynomial
matrices. If G(s) and, consequently, A(s) are of ful1 row rank, there exists a
unimodular polynomial matrix U(s) such that U(s)A(s) is row reduced. Consider
(9-206)
U(s)A(s)T(s)= U(s)B(s)
We assert that if G(s) is of full row rank in lR(s), then T(s) is proper if and only if
i= 1, 2, .. "q
b..(U(s)A(s)) :::O:b'i(U(s)BCs))
(9-207)
Proof
where b,. denotes the ith row degree. Ir A(s) is square, this assertion is essenti
ally Theorem G-IO. The proof for the nonsquare case is similar to the one of
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
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
(9-207) is not really needed. Let T(s) = NT(s)Di 1 (s). Then (9-205) becomes
A(s)NT(s) = B(s)DT(s) or
[A(s)
B(s)] [ -
~~i;~J = O
(9-208)
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
of (9-208). Its dimension is, following Theorem 2-5, equal to
r~ p +r -
rank [A(s)
T(s) =
IrY''e has no ful1 COIUl
algorithm, at least on
and Y/,,) = 0, where Y'
basis, Yhe has a fui! ce
jth column degree of
IfT(s) is proper, as s
Hence ir Y'e has no
the theorem.
Gm(S)]
Now we may apply Theorem G-14' to solve (9-208). We form the generalized
resultant T k from the coefficient matrices of A(s) and B(s) as in Theorem G-14'.
We then search its linearly dependent columns in order from left to right by
using the column searching algorithm. There will be exactly r primary depend
ent columns in T k Let the (p + r) x;: polynomial matrix Y(s) be the solutions
corresponding to these r primary dependent columns. Then Y(s) is, as in
Theorem G-14', column irreducible and column reduced, and is a minima!
polynomial basis of the right nui space o" (9-208).
Let Y he be the column-degree-coefficient matrix of Y(s). For convenience
of discussion, we assume = r and partition Yhe and Y(s) as
- NT(S)]
Y(s) = [ DT(s)
In this theorem,
minimal basis is eSS
9-33). The theorem i
case. Ir r < r, clearl:
If ;::::O:r, Y(s) and \'"
holds. In this case,
have increasing colur
the 11rst 1" coiumns VI
of (9-204). This com
Example 4
Find a 3 x 2 minima
(9-209)
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
be defined; therefore there is no guarantee that DT/le, the coiumn-degree
coefficient matrix of DT(s) in (9-209), is of full columri rank. We note that Yhe
23"if
1d q x r polynomial
rank, there exists a
w reduced. Consider
(9-206)
531
(9-207)
Proof
s assertion is essenti
similar to the one of
to find a T(s) with a
uch a method. The
.1ce the condition in
hen (9-205) becomes
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;
hence NT(s)Di les) is proper. To show the converse, we consider
(9-208)
or
(9-21 O}
UY'e has no full column rank, there is, as a consequence ofthe column searching
algorithm, at least one column, say thejth column, such that ejNT(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
jth column degree of Ves). Consider the jth column equation of(9-21O):
T(s)DTj{s)S-k j = NTAs)S-k j
For convenience
UT(s) is proper, as s---> 00, we have T(ca )Y/,ej = O = Yhej =/= O. This is not possible.
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
minimal basis is essential; otherwise the theorem does not hold (see Problem
9-33). The theorem is developed for the case = r. We discuss now the general
case. U < r, c1early no solution T(s), proper or improper, exists in (9-204).
If r? r, Ves) and Y'e are (p + r) x r matrices of rank r, and Theorem 9-28 still
holds. In this case, in order to obtain a minimal solution, we arrange Ves) to
have increasing column degrees, that is, (jcl Y(s).::s; (je2 Y(s):S; ... :s;(jcf Ves). Then
the nrst r columns with a nonsigular \\e wil1 give the minimal proper so!ution
of (9-204). This competes the discussion o' the minunai design proolel.
Example 4
lS
"matrices. We note
(s) = A -l(s)B(s), the
.. lerties that ciN(s).::s;
rank. In the minimal
jA -:-.1 (s)B(s) may not
the column-degree
:k. We note that Ylle
-s
s+3
-3s-7
s+3
23
Ir Yh, i~ not computed by using the column searching algorithm, we may not have Yh,j = O, for
sorne j. However, ifrank Y,,, < r, bne of its cciluinn
be transformed into a zero colurnn by
elementary coluinn opeiations.
can
532
The multiplication of S(S + 1)(S + 3) and the substitution of T(s) = NT(s)Di 1 (S)
yields
O
O
:3 - - - 0-
O: O
O:
o
O: O
O:
-- -0- -: - 6---- -0- -:- - --- -0- - - - - : - -0- - - - -
4 -6 -3 : 1
-7: 3
O
O: O
O
-----------L---------~------------l-------1
O -3 : 1 - 1 : 3
O
O, O
O
1 -8 -4 : 2 -10 :L 4 -6 -3 ,: 1
-7_
___________ L
O
O -1 : 1
-1: 1
O -3 ,1
-1
O
-2
-1
:
1
3
:
1
-8
-4
:
2
-10
___________ L
L
_
O
O -1
1
-1
,
o
"., O -2 -1
1
-3
y(,)~ 1
~-----
100
1 -1/2
O
O
O
1
O
O
O
O
O
O
O
1/6
1
1
O
O
O
-7/6: 1/2
-1
1
1
O
1,0
O
O
O
O
O
O
O -3
1
O 9/2 -3/2
O
O
O
O
O
-1
3/2
-1
1
-0 J'
'l
O~
[x
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
1
O
01
O
O
x: x
O: O x
i i
-1
x: O O]
533
to the three primary dependent columns, we use the dual formula of (A-II)
to compute
O O
-1 O O
O I 1
103
1 O O
--------
O 1 O
O
O
O
O
=3. We form T l as
Igorithm to compute
O
O
O
O
O
I
I
O
Y(s)
O s
O
-1 O
O
O 1 s+1
- - --------1 O s+3
I O
O
with
Y/le
O 1 O
-1 O O
O O 1
-------1 O 1
1 O O
O
O
O
O
-1
-3/2
3/2
O
-1
O
O
O s+ 1
-------I s+ 3
O
O
-------O
O
-1
O
I
-~J
and
-1
534
[Lo
9-8
Concluding Remarks
or
(9-211 a)
(9-211b)
Several methods are available to solve this equation. First, we may apply a
sequence of polynomial elementary matrices to transform [O'(s) N'(s)]' into a
Hermite row form or a upper triangular form as in (G-33):
U 11 (s)
[ Uds)
Then a solution exists in (9-211a) if and only if R(s) is a right divisor of F(s). A
general solution L(s) and M(s) can then be obtained as in Problem G-14o See
References S34, S139, and S1890 Another method is to find an observable
realization {A, B, C} of O-l(s)N(s) as in (6-131). We then use the operator [lD
defined in (6-225) to compute ITv(F(s)). Ifwe use the method in Problem G-15
or any other method to compute
[I(s) == D(s)Q J. (s) -:-- f 1 {,~')
Fl(s)=L(s)E 1
Then a solution L(s) and M(s) exists in (9-211 b) if and only if
rnk[B
AB
...
An-1B] = rank [B
AB
...
An-1B
El]
Mo : L l
535
CONCLUDING REMARKS
e pole-zero cance\la
ching, the choice of
Id a proper T(s) and
n, the design is still
Then it can be shown that L(s) in (9-211 b) is given by L(s) = Ql(S) - Q2(S). See
Reference S85, where the method is developed only for strictly proper
D-l(S)N(s); the method, however, has been extended to linear systems over
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
ization of composite
;tems, and the design
: plant input-output
asym ptotic tracking,
le implementation of
dback systems.
Iy on the solution of
(9-211 a)
(9-211b)
~ht
divisor of F(s). A
Problem G-14. See
) find an observable
use the operator IT D
lOd in Problem G-15
fe
)y
the column of
A/- 1 BJP is the
sL(s)~that s, M(s) =
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
ical comparisons of these three approaches are not available at presento
In addition to Equation (9-204), the equation
L(s)D(s) + N(s)M(s) = F(s)
(9-212)
also arises in the design of multivariable systems. This equation is quite different
from the Diophantine equation, and its solutions are much more complicated.
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
bypasses the equation. Consequently, the equation in (9-212) does not arise in
this texto
In this chapter we studied two feedback configurations: unity feedback and
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
completely concelled by O f(S). The reason that we have freedom in assigning
Oc(s) in the plant input-output feedback configuration but not in the unity
feedback configuration is very simple. There are three sets of parameters
L(s), M(s), and Dc(s) in the former, but only two sets of parameters N,(s) and
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 irst find a noncoprime righi fraciion anci then use el
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
reduced; in fact, it is also row reduced and -inthe polynomial echelon formo
Once a right-coprime fraction G(s) = N(s)D - 1(.s:) is obtained, the next
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
N rows of the last block row ofS m become Iinearly dependent, the mis the row
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.
With the computed v, the chosen F will be in the row space of Sv- 1> and the
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
used, difficulties remain. In om design, the order of rows cannot be changed.
Hence the singular value decomposition or Householder transformations with
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.
Problems
9-1
~.
05+2
05+3
g2(S)=-
-J---- t
Lsr, + I J
1_
f
l~
Figure P9-5
9-6
1s the feedback system with 9(S), given in Problem 9-1, in the forward path and
92(05) in the feedback path completely characterized by its overall transfer fllnction?
9-2
Consider
f)
:t_ll
:~:j
~J.
05+2
9-8
Is the tandem connection of 9(S) followed by 92(S) completely characterized by its overall
transfer function? Verify your result by showing the controllability and observability of
its composite dynamical equation.
9-9
Are their paralIeI and tandem [(;(s) folIowed by (;2(05)] connections controllable? obser
vable? completely characterized by their transfer matrices?
9-4
Consider
and .
9-3
9-5
I
A
G(s) =
05+1
O
and
G 2 (s) =
05+2
O
O
1.
s
+l
controlIable and observable? Note that (; (s) and (;2(S) have the same set of poles.
Figure P9~10
PROBLEMS
9-5
537
Are the feedback systems shown in Figure P9-S BIBO stable? asymptotically stable?
Ti~"(S-+-11
I
1.,--
s
_s_+_2_
I
5(S
1)
Figure P9-S
9-6
S2 -1
G(s)= _1_
s- 1
G,(s) =
-
s +2
s +3
s +1
~3
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.
+2
haracterized by its overall
bility and observability of
9-9 Consider the plant given in Problem 9-8. Find a set of lwo poles so thal the com
pensator in the L1nity feedback system is a constant. Find a set of three poles so that (he
compensator of degree 1 in the uni(y feedback syslem is improper.
+3
+1l
2j'
;-7-
s +2
:tions eontrollable? obser
Consider the feedback system shown in Figure P9-1O. What is iS overall transrer
function'! Let N(s) = N (s)N 2(s), where 1'/2(S) can be cancelied. Snow ihat, Or any 9f(.I') =
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
De(s) be HLlrwitz? lf De(s) is not Hurwitz and if the system is implemented as shown, will
the system be acceptable? Can you find a different but equivalent implementation which
will be aceeptable? See Problem 9-32 and Reference S34.
9-10
. \
PIs)
~.
ID;W~
Figure P9-10
538
9-11
Consider
9-18
5+1
~
5(5-1)
1
G(5) =
f J
52 - 1
Find a compensator so that the poles of the overall unity feedback system are - 1, - 2 j,
and the rest chosen as - 2.
9-12
~
[ 5 +1
G(5)= -
52~J
1)
5(5 -
9-13
Find a compensator in 11
9-19
G,(5)=
.
1 l
~
---.!_
5+ 1
5-2
1
5+3
(5 - 1)2
Gl(s)=
5 +2
(5-1X5+1)
1
G 2(5) =
5- 1
l5-1~;5+1) (5-1~1(5+1)J ~
r:
G 4 (5)= O
5-1
5-1
Show that ir all elements of G(5) have no poles in common, then G(5) is cyclic. Note
that repeated po\es in individual gj(5) are, however, permitted.
Lel D-l(S)N(s) be
nomial ~atrix ifand only
malrix M(s) such lhat M(~
9-20
9-21
9-14
9-15
e(s)=(I
G(5) =
+1
5-1
-2
=!'J
s
s- 1
9-16
Show that ifthe rational matrix C(5) is cyclic, so is (:(5) = C(5) + K for any constant K.
9-17
G(5) = .
-1
.[ O
52
-1
5
9:22
52 +5 + 1
52 -
Find a compensator in the unity feedback system so that the poles of the resulting system
consists of the ro<its of 52 + s+ 1 and the resi fram - 1. Compr~ your result with the one
in the text.
Giveng(5)=(5-1)j5(.~
PROBLEMS
9-18
G(s)
A
l~ ~
l [52 0J-I[1 5 0J
=
lOs
s
539
Find a compensator in the unity feedback system so that the resulting denominator is
Dj(s) = [
lS
+ 1)2
Is the resulting
Gj(5) cyclic?
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 [s) are left coprime. Show that if O(s) and
M(s) are left coprime, so are O(s) and (s)M(s).
9-19
9-20
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
9-21
1,
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.
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
9-22
9-23
540
9-24
9-32
G(s) =
[s
I
Verify that the row index of G(s) is 2. Design an input-output feedback system to have
(s +1)2
DJ(s) = [
(s
O
+1)3
~ ILS;
11- ~
~
_ [s
DJ(s) =
Figure P9-32
+1
N(s)[De(s)D(s) + Ne(s)N(s
9-34
+2)
Fnd a mnimal pr
s +2J
s +2
DJ(s) = [
s +2
as its denominator matrix. Is the system decoupled? Are all compensators proper? Is
the system well posed? ls there any unstable pole-zero cancellation?
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.
,~
s+2
J.,
9-28 Consider Corollary 9-24 with G(s) strictly proper. Show that if Dcill(s) = DeO(s)
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, =
1, 2,. ,eL, the following
o
[D co
Neo:
k.] [ -80
A o :"': -8",
A",]
meets (9-74), where (;(=(111+ l)q-n and [-Ro Ao : ... : -B", A",] is a basis of the
(J.-dimensional kIt "uii space o[S", computcQ as in Appendix G. ThS ,s;:, ),u<c,o,CIL ''''~L'G"
of all solutions of (9-74).
9-30
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
[O
G(s)=
.
I
A
I
s +1
J[S2 + 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?
Figure P9-32
9-33
9-34
compensators proper? Is
ation?
well-posed input-output
pole-zero cancellations.
s~ ::~]
1
.1'+ 2
2(,+1)
T(s) =
[1.1'+.1'2::+ 3)
.1'(.1' - 1)
-
(.1'+ 2)2
(s-I~;s+3)
.1'+2
9-35 Let Vis) be a 11 x p polynomial matrix or rull column rank with column degrees
jJ., i = 1;2, ... , p, and lel x(s) = [Xl (s)
X2(S) ... xp(s)]' be any p x I polynomial vector.
Consider
y(s) = Y(s)x(s)
Show that
2': v - 1, rOl' any real k, i =
A o : .. ' : - B",
A",J
,(s)
:k implementation of Go(s)
FO
if and only if Y(s) is column reduced. This is called tbe prediclable degree properly in
Reference S95 and can be used to show the degrees of a minimal polynomial basis to be
111 ininla.l. Hin!: Let d =:: lllax [p; + deg .':f{S) ~
~~~h r->n ,..._1~? ri~~1 . . "le ha'J~ !:t~g ~'/(,r;) :.::; J
'T,~
show the equality, we show the y, in y(s)=YO+Y1S+'" +y,s' is nonzero by using the
relation Y, = Y,,,a, where -,(s) = aii -" + ... and a = [a 1 a2 ... a p ]'.
A-1
Gaussian E
Consider the n x m m:
Elementary Transformations
EI
1
O
O
O
O
O
1
O
O
O
O O O
O O O
1 O O
O e O
O O 1
Ez =
1
O
O
O
O
O
O
1
O
1 O
O
O
O
O
O
O
O
1
O
1
O
O
O O
EJ =
1
O
O
O
O
O O O
1 O O
O 1 O
d O L
O
O
O
O
O O O 1
where e1 = -ail/all.
matrixK I is the prodl
EJ. If a~ z is different 1
row of KIA and -a}z/,
KzKIA =
where eZ = -atz/a~z a
can be transformed int<
(A-1 )
IX
.0
where e and d are real or complex numbers and e =/=0. Elementary matrices are
squ.o.re an.d nODsingular. Their inverses Pcre
1
O
E-lO
1 O
O
O O O O
1 O O O
O 1 O O
O O l/e O
O O O 1
E21
Ez
1
O
O
1
E-lO
O
J O -d
O,
O
O
O
1
O
O
O
O
O
O
O
O
O
1 O
O 1
O e 3Z
\)
x
x
x
x
/'o
)\
O O O
O O O
x
O
O O O
(A-2)
543
GAUSSIAN ELIMINATION
A-1
Gaussian Elimination
a t3
A= ~Zl
a zz
a Z3
a Zm
anl
a nz
a n3
a nm
[a"
a'-l
(A-3)
First we assume all ,p O. We add the product of the fmt row and (-aJall) to
the ith row, for i = 2,3, ... , n; then we have
o O
1 O
O 1
ll
a
O
1 O
1
O
d
O
O O O
O O O
1 O O
O 1 O
O O 1
: a~z
O
O
KzK1A= O e 32
O
a ll
O
O
O KIA= O
O en 1
I~;0
1
O
-d
O
O
O
1
O
O
x x
x x
O X
(A-4)
a1m
a~3
a~m
a l 3 ... Qlm
a12 ab . .. a1m
'-----------2-
. .. a 3'11
O aZ
33
al2
, a~3
(A-5)
a;m
1
1
x x
x x
:l
fx
)(
xl
X.
'0
X.
;(
f..
"
j,
O O x
O O O
O O O
x x
O x
x
x
O
O
O
1
O
O
O
O
O O O
O O O
x x
O O
O O O
O O
or
(A-6)
(A-2)
2m
(A-1)
23
whereeil = -ail/all, i=2, 3, ... ,n and aIj=aij +eilalj' Note that the n x n
matrixK I is the product of n -1 number of elementary matrices of the form of
E 3 . lf a1z is different fram zero, then the addition of the product of the second
row of KIA and -alz/a1z to the ith row, i = 3,4, ... , n, yields
1 O
O 1
O
E. 3 = O
O
O
22
O O
Dwing operations on A
1)multiplyinga rowor
rchanging two rows or
)Iumn and a number to
. achieved by using the
AII diagonal elements in (A-6) are nOt nccessarily nonzcro. Il is possiblc lO makc lhern nonzcro
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.
Even if all of them are nonzero, the process may still encounter some com
putational difficu\ty. For example, if a is very small in magnitude, then
(- ada ) = eil will be a very large number. lf there are errors in representing
([ Ij, these errors will be greatly amplified by e in the operations. F urthermore,
these errors may propagate and be reamplified at \ater stages. Consequently,
the final resu lt may be overwhelmed by errors. Because of this phenomenon, the
gaussian elimination is said to be numerically unstable.
To overcome this difficulty, before carrying out the elimination. we first
search the element with the largest absolute value 2 in the first column of A,
and then interchange this row with the frst row by using a matrix of the form of
E 2 . We then carry out the elimination for the first column. Before carrying
out the elimination for the second column, we search the largest element in
magnitude in the second column of KA 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
A can be transformed into the form in (A-6) by using a sequence of elementary
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 elementsof A. We then move the element to the frst column by
an elementary column operation and then to the first row by an elementary row
operation. We then carry out the elimination of the first column of the trans
formed A. After this elimination, we repeat the process for the remaining
matrix. This process is called the gaussian elimination with complete pivoting.
This process is more stab\e numerically than the one with partial pivoting; how
ever, it is more costly becallse the search ofthe largest element in magnitllde is a
time-consuming process. Hence complete pivoting is not used as often as partial
pivoting. According to Reference Sl38, from the point of view of overal1 per
formance which ineludes efficiency, accuracy. reliability. generality. and ease Di
use, the gaussian elimination with partiai pivoting is satisfactory for most
general matrices.
*A-2
Householder Transformation
The numerical stability of the elimination can be further improved by using the
Householder trarisformation. Let 1 be a unt matrix. A HOLlseholder trans
formation is a ~quare matrix of the form 3
H'H=
we ha ve H 1 = H'; hl
singular values of H "
ratio of the largest sin
the use of Householde
of a problem lsee Prol
An important pro
two vectors a and b 1
exists a Householder t
x = la - b)flla - b11 2 , th
H=I -'--2xX:
2 For conciseness: we call such elernent the largest elernent in rnagnitude.
'AIl argurnents still hold ir H is defined as H = [ - 2xx*, where x* is he cornplex conjugate transpose
or x. In this case, wehave H - 1 = H* and H s unitary.
HOUSEHOLDER
Ide.
TRANSFORMAnON
545
with x'x = 1, where x is a column vector and x' is its transpose. Clearly H' =
II - 2xx')' = 1- 2xx' = H; hence H is symmetric. Because, by using x'x = 1,
[1 _
Since a'a = b'b by assumption and a'b = b'a for this is a scalar quantity which
is invariant under transposition, we have
2(a'a - b'a)
Ha =a - -----(a -b)= b
.
2(a'a - b'a)
This establishes the assertion. Now we use this property to show that the
K, and K z in (A-4) and (A-S) can be chosen as Householder transformations.
Consider the n x m matrix A given in (A-3). First we compute the norm (J
of the first column a, of A. We choose b[ as [(J O O .. , 0]'.4 Then
there exists a Householder transformation H, such that H[A is in the form
ofthe right-hand-side matrix in (AA). Next we delete the first column and first
row of H,A and repeat the process for the first column of the remaining sub
matrix. Proceeding in this manner, the matrix A can be transformed by a
sequence of Householder transformations into the form shown in (A-6).
The numerical stability of this process can be further improved by the per
muiation of columns. First we compute the i10rms of aii coiurnns of A. NeXL
we permute the column with the largest norm with the first column. We then
apply the Householder transformation to carry out the eliminaton. By this
process, the diagonal elements of the resulting matrix will be in the decreasing
order of magnitude. We call this process Householder lransformations wilh
pivoting.
We note that Householder transfor.mations are not triangular matrices as
K and K z in (A-4) and (A-S). Hence the rows of A are completely scrambled
after the application of Householder tfansformations.
4TI1esign may be chosen lo beequal lO -sign (/11 10 reduce lhe roundoff errors. In lhe QR facloriza
lion (see Problem A-S), we muSl choose howeverlhe positive signo
546
ELEMENTARY TRANSFORMATlONS
If a matrix is transformed into the form in (A-6), the rank of the matrix is
equal to the number of nonzero rows in (A-6). This is a simple way of computing
the rank of a matrix. 5
A-3
Row-Searching AIgorithm 6
is no nonzero eIernen
to the next row. If""
where K ~ KIl-K,,_
1 on its diagonal. so i:
[fthe;th rowof is a
on its previous rows.
rb
aL
To compute thejth re
jth row of K under i1
o
1
e21
SThe computation or the rank of a matrix is a difficult problem. For example, the matrix
-'1
=:
'
O
O
fjl
-1
-1
j[
: ..
ejJ
~j: _ _
bjj =
1
Asx ~ As 2- 2 = 2- 4] =2- 4
2- 3
2- 4
2- 4
1
1
l1en il is straightiorw
2 -4
1] [2[
('j2
bj2
2 -1
---e 32 : 1
-\
AS=\: -; =:
O
O
e 3t
[1]
2- 4
2- 4 e
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
ror all nI, norfrom Its eigenvalues [which are all equa\ to (see Problem
nor from the form
of A.. However, if we compute the singular valuesof A,,(see the singular value decomposition in
Appendix E}, tlen 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 compting the rankof a matrix.
6This algorithm is a simp[fied version of the one in Reference S2Z
2,23)],
Ki cannol be chosen as He
of A.
r()WS
ROW-SEARCHING ALGORITHM
(A-S)
o
F=
eZI
O
O
e31
e 3Z
enl
enz
e n3
O
O
O
(A-9)
To compute the jth row of K. we take the firstj rows of F and then arrange the
jth rw of K undet it as
O
I
e Z!
he matrix
(A-7)
aL
~xample,
547
e 31
ej l
: e32 :
'
ejZ :
O
O
O
First.i rows of F
ej3
l'j(j-l)
~ith
(A-10)
row of K
b jj = 1.
e1k+ I)kj
bjj ] ~(k+Z)k
'~e
(A-11 )
ejk
j .
p;k-t 1
bj~epk
k = j - 1, j - 2, . . . , 1
.K; eannOI be ehosen as Householder lransrormalons 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
(A-12)
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.
Example 1
\-1
A~H
L7
-1
(~:
-1
-2
-2
-3
1
8
-4
-2
-6
10
10
A is
-2
-4
2
1
-1
-1
-2
1
4
1
4
-~
K1A = -4 O 1
2 O O 1
~:~=: - ~ :~: -~ ~
1
A=
10001
1
-1
O
3
3
O
2
2 ~Al
O
6
3
085
Except the first element, the first column of K 1 is the third column of A divided
by - 2. We next chaase the (2, 1) element of A\ as the piVOl i:"w.d compute
1
1
O
KzA 1 = O -1 1
O -1 O 1
O -2 O O
-1
1
O
O
O -2
O -2
-1
:):~
A1 =
:1~ -2
O ~Az
1
1
O
O
O
O
4
4
Since the third row of A z is a zt:ro row, we set K 3 = 1 and proceed to the next row.
Hence we have
or
where a is the ith rov
Hence we have .
ar
549
ROW-SEARCHING ALGORITHM
K4 K 3 Az = K4Az =
(A-12 )
~e
1
O
O
O
O
-1 -1
(~) 1
Az =
O O
O -2
O O
1
O 1
O O 1
O 0-1
~A; -2
1
2
O 2
O O
O @
O O
O ~A
1
-2
FtA= -4
2
1
-1
-1
-2
-1
3
1
O
1
O -1
-1
1
O
O
O -2
O
O
iA =
O
O
O
O
We use (A-12)
-2
2
O
4
O
1
2
O ~A
1
Note that the ith column of F is the ith column of K, i = 1,2,3,4. Since the third
row of 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
-2
1
2
2 f'...A
3
5
2
2
6
8
d column of A divided
ivot and compute
-2
O
4
4
O ~Az
1
1
O
O
O
,- -
-; 1- ",-- -_:. -
[-2
or
: ....-1
~]
'-211:-1} ~1~'
--- --------I
Hence we have
11"'- - "
......
-1
O
OJA=O
-23 -az +a 3 =0
1
-2
-4
2
1
1
-1
-1
-2
-1
;-0-;
, ,
"OW
(A-"i3)
'Ji A..
w" r;i)nCr.)l)f'~
:_0_: -1
O
r ---f:
:~-=-------
Hence we have
or
[1 - 1 O - liJA = O
a -3 z +03 3 -'a 4 +a 5 =0
(A-14)
550
-1
which implies
and
To find the coefficient
which im plies
1JA=O
which does not have b S3 =0, and is a different combination. The property
=0 if the ith row is dependent is essential in establishing Theorem G-14
in Appendix G.
A remark is in order regarding the numerical stability of the row-searching
algorithm. If the pivot element of each row is chosen as the leftmost nonzero
element, then this algorithm reduces essential1y to the gaussian elimination
(without any pivoting). Though the pivot is chosen as the largest element in
magnitude of each row, the pivot is no"t necessarily the larges't element in magni
tude of its column. Hence the row-searching algorithm by choosing the largest
element in magnitude as the pivot is not exactly equivalent to the gaussian
elimination with partial pivoting. Consequently, there is no guarantee that the
row-searching algorithm will be numerical1y stable.
In the search of linearly independent rows of A, it is essential not to alter the
order of rows. The column positions, however, can be arbitrarily altered.
Hence we may apply to the columns of A the gaussian elimination with partial
pivoting or the Householder transformations to transform A into the form
and
b ji
a
aZ
ALL z '" ~ a 31
azz
a 32
O
O
a 43
a 41 a4Z
a S asz
aS3
O
O
O
O
O
O'
O
O ~A
O
O
(A-15 )
Householder transformations with pivoting {permutation of rows), however, are not permitted.
* A-4
Hessenber
HESSENBERG fORM
s of A in the order of
o as the row-searching
:pendent, then the ith
: jth linear dependent
, its coefficient is zero
If its previous linearly
h i <j, of A is linearly
{ dependent, then the
atrix. Consequently,
That is, epi=O, for
=0. Because of this
Without bji=O, the
..lple, the addition of
551
To find the coefficients of linear combination for the third row, we solve
o OJA =0
which implies
and
-a32
b 32
----
b 31
a22
-(031
+b 32 a21)
all
To find the coefficients of linear combination for the fifth row, we solve
(A-16)
which implies
lJA=O
~A
(A-15)
lnsformations. 9 It is
I implies that the first,
.f their previous roWs
[l their previous rows.
and
We see that the bij are obtained by back substitution. This is a numerically
stable method (see Reference S212). Hence we conclude that once A is trans
formed into the form in (A-lS), the coefficients of linear combination can be
easily obtained. In solving these coefficients, the elements corresponding to
linearly dependent rows are set to zero as in (A-16). In so doing, the result will
be identical to the one obtained by the row-searching algorithm. Although
this method of searching linearly dependent rows is less direct than the row
searching algorithm, the method is stable numerically.
The row searching algorithm, however, will be used exclusively in this text
for the following two reasons. First, the concept is very simple. Its use will
not complicate or obscure the basic issue of the problem. Second, the method
can be easily earried out by hand for simple problems. After solving one or lwo
problems by hand, one would understand the material better and gain a con
fidence in llsing a digital computer. For hand ea1clllation, the pivot should be
chosen rOl' [he convenienee 01 computation. For exampie, we may choose ene
element with value + 1 or the element whose column has the largest number of
zero elements as the pivoto In summary, the use of the row searching algorithm
is for pedagogical reasons. In actual digital eomputer computation, one should
use the method discussed in (A-l S) or other method which is numerically stable.
*A-4
Hessenberg Form
552
ELEMENTARY TRANSFORMATIONS
O O ...
1 :0
__ J _ _ _ _ _ _ _ _ _ _ _ _ _ _
PIA~
O:
O:
P ll
O:
,
A=
O:
O
O
9__ ! _: _~ _9
O o:,
~_
A-2
x x
PIAPI= O O
O O:
,
,
,
O O
O O:
O x x x
O O x x
O O O x
x x x
x x x
x x x
O O O O
O O O
x x x
O x x
an
Repeat Problem A
A-S
Let A be an
11
xm n
that
Again the inverse of P 2 has the same form as P 2 and the postmultiplication of
P 2 PAP1 by Pi l will not operate on the first two columns of P 2 PAPI.
Hence the pattern ofzeros in the first two columns of P 2 P IAP1 is preserved
in P 2 P IAPI Pi l. Proceeding similarly, we can transform a matrix, by a
sequence of similarity transformations, into the form
x x
A-1
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 PI will not operate
on the first column of PA; hence the pattern of zeros in the first column of
PI A is preserved in P IAPI. Next we find a P 2 so that
10:0 O
Problems
(A-17)
PROBLEMS
Problems
A-1
x
x
x
[~:r &~
s'j
x
x
x
553
1.5
Use the circled elements as pivot elements. Use the procedure in (A-9) to (A-12) to find the
coefficients of ts linear combination.
A-2
i
20
-1
-~]
:))
3
I
O -3
3
O
-1
Verify that bji =0 if the ith row. with i <jo is linearly dependent.
A-3 Is it possible to obtain a triangular form by using the procedure of obtaining a
Hessenberg form?
x x
x x
x x
x x
x
x ..
A-4 Show that if Qi' i = L 2. . .. m. are unitary matrices. that s. QtQi = QQt = I. then
so isQ~QmQm-l'" Q2QL
x x
A-S
that
: postmultiplication of
,lumns ofP 2 PAPI.
2 P 1 AP ti is preserved
.sform a matrix, by a
(A-17)
d by using numerically
Iter computations.
Let A be an n x m matrix of rank m. Show that there exists a unitary matrix Q such
or
where R is an upper triangular matrix with nonnegalive clements 011 the diagonal and Q
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
are similar. The QR algorichm, which is the most reliable method of computing the eigen
values of A, s based on this property. See References SI81 and S200.)
o at
B
Analytic Functions
of a Real Variable
Theorem 8-1
If a functionf is anal
arbitrarily smaH nonz
D.
Proof
f(t)= { ~ .
for t-l=O
for t =0
555
Theorem 8-1
(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
(lo, tI), then the function and its derivatives are all equal to zero on (to, tI)'
where Ilu(t)II~(u(t),
Section 2-8.)
Proof
The solution of the sI
x(t 1
Defle
Minimum-Energy Control 1
x~
-i
l1
X=
lO
Subtracting both side
x= A(t)x + B(t)u
E:
where x is the n x 1 state vector, u is the p x 1 input vector, and A and B are,
respectively, n x n and n x p matrices whose entries are continuous functions of
defined over (ro). lfthe state equation is control1able at time to, then
given any initial state Xo at time to and any desired final state xl, there exists a
finite tI > to and an input u[lo,ld that will transfer the state Xo at time to to Xl at
time t 1> denoted as (x o, to) to (x 1> t d. Define
0:<
W(t o,
\L
l!>(.
By using (2-94),we ca
lO
fO aH t in [too t,
U>1
will transfer (xo, to) to (Xl, td. This is proved in Theorem 5-4. Now we show
that the input UfrO,I,) consumes the minimal amount of energy, among all the
u's that can transfer (xo, to) to (Xl, t).
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
556
1,
lO
IluO(t)112 dt
Considcr now
By sorne manipulatio
557
MINIMUM-ENERGY CONTROL
(See
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
i
"
x=
<ll(t o, r)B(r)ul(r) dr =
i'l
lO
<ll(t o, r)B(r)u(r) dr
lO
lation
which implies that
:tor, and A and B are,
Jntinuous functions of
)llable at timeto, then
state XI, there exists a
e Xo at time to to Xl at
i"
'o
=o
(C-3)
'l
) dr
(C-4)
'o
Consider now
LII
ul 2
l1 elr
10
nd let UO be the control
n
= ['
+2 [
=
(u 1 (r)-uO(r},uO(tdr
r' Ilul(r) -
Jto
u O(r)112 dr +
r"
Jro
IluO(r)W elr
558
MINIMUM-ENERGY CONTROL L
Since
where x is the n x 1 st
vector; A, B, and e ;
The response of FE
where T is a positive
0, T, 2T, ... , are called
constant inputs given "
the" behavior at samp
dynamical equation (
Q.E.O.
izing
-energy control.
x=Ax +Bu
( 0-1a)
y=Cx
(0-1b)
eA(I-'lBu(r)dr
y(t) = Cx(t)
(0-2a)
( D-2b)
= u(k)
(0-3)
whert:: T is a posive constant, caBed the sampling periodo The discrete times
O, t; 2T, . .. , are caBed sampling instants. The behavior of FE with the piecewise
559
560
u(e)
.::::kJ
T~
Hold
u ~(e)
u(e)
u(
u(k)
"
1"
tjil
o
l.
2T
Theorem 0-1
L-J
Figure 0-1
First we remark (
is an eigenvalue o
Al =T +jf3 and Az =.
Xz = e(t+ i(fJ + ZnajT))T =
=eAT[eAkTxo+
J:
}'i
T
eA(kT-tlBU(T)dT]
(k+ 1)T
eA(kT+J;-tlBu(T) dT
(0-4)
kT
The term in brackets in (0-4) is equal to x(k); the input U(T) is constant in the
interval (kT, kT + T) and is equal to u(k); hence (0-4) becomes, after the change
of variable a = kT + T -T,
Proof of Theorem D
We assume, without I
shown in Table 5-1.
DFE:
where
A=e AT
B=(f e
C=C
(0-5a)
(0-5b)
where
(0-6)
At
dT)
B~MB
(0-7)
(0-8)
---
h
I
l.
2T
561
Theorem 0-1
L-J
u(r) dr
(0-4)
:)
he input is piecewise
anse at the sampling
:;an be rep\aced by the
quation
.A==q
A.
t-;:-::~
,,1'
(1
A12T
where
eA"'~'(m) J
T
(nij-1)!
T(IIj- Z l
(0-6)
O
A ijT -J.
-6.
1)= e
\"l
(0-7)
(0-8)
O O
and 1i =
eAT
(nij 2)!
T(lI i j - 3 l
O O
T(II;j-l)
(0-5a)
(0-5b)
(D-10)
(nij-3)!
------
562
--..
----~
.'~--
__ ._---------_._----._--_..
-.~------_.-
o
ij- XI =
x x
O O x
O O O
O O O
O O O
O x
O O
Qij=
x O
x O
x x x
O x x
O O x
In the proof of T
theorem.
(0-11 )
x O
O O O
O O O
such that
PijijQij =
O Xi
O O
Xi
O O O
O O O
O O
O O
O O
This theorem is i
by using an input de
be controlled by an il
(D-12)
Example
X
O
Theorem D-2
If a continuous-tim~
its discretized state ec
O 1
,
II.
where Pti~ Qij 1 and has exactly the same form a~ Qij' We see that the lordan
form of Aij is the same as A ij with A i replaced by Ai' Define
(5
+ 1)(5 + 1
(0-13)
+ 1) =
PP- Ix(k)
+ PMB~(k)
(0-14)
-;f;.-
1 ,Hold
~-L
u()
'
with P p-l exactly in the form in Table 5-1 with Areplaced 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 nj
with all diagonal elements equal to (l-e-AiT)/A i if A;=I=O or T if A = O. The
Figure 0-2
563
(0-11 )
Theorem D-2
lf a continuous-time linear time-invariant state equation is not controllable,
its discretized state equation for any sampling period is not controllable.
1m
This theorem is intuitively obvious. lf a state equation is not controllable
by using an input defined for all time, its discretized equation certainly cannot
be controlled by an input defined only at discrete instants of time.
(0-12)
Example
Consider the sampled-data system shown in Figure D-2. By partial fractiol1
expansion, we have
(s
(s
+ 1 + 2i)
(s
+ 1 - 2i)
(0-13)
3
(0-14)
llaced by
Xi'
Now we
: matrix M defined in
ular form of order nij
O or T if Ai = O. The
Figure 0-2
s
8
(s + 1) (s + 1 + 2j) (s + \ - 2j)
1
y
-3 -2 -\
564
o
O
-1-2i
O
y = [2
-1
][XI] +[1]
-1+2i
1 u
1
x?
X3
-I]x
1 -e -T
O
_1_(I_e- T -
e - T- 2iT
2iT )
+2i
Si r
u(k) (D-15)
O
_1_(1
_e-T+2iT)
1-2i
We conclude from Theorem 0-1 that the discrete-time state equation (0-15)
is controllable if and only if
2nrx
T op-- =nrx
A hermian form of
geneous polynomial ,
2nrx n
Top--=- rx
and
rx = 1, 2, ...
This fact can also be verified directly from (0-15) by using either the criterion
p[B AB A2 B] = 3 or Corollary 5-21.
Problems
0-1
o
- 1--- 2i
O
1\1; +l~
-1 +2;J
Lo
Show that its discretized state equation is always controllable ror any T including T = 27((44
for cx = 1, 2, .... This shows that the conditions on the eigenvalues in Theorem 0-1
are not necessary ror a multiple-input discretized equation to be controllable:.
Show that sufficient conditions for th discrete-time dynamical equation in (0-5) to be
observable are that the dynamical equation iri (O-I)1s observable and lni [A(A)-A)A)] i=
2ncx/T for cx'= 1, 2, ... , whenever Re [A-(A)-J,}A)]=O. For the single-output case
(q = 1), the cOndit-ions are necessary as well.
0-2
where
Mi is the com
x*M
where M =~(Ml +M
can be written as x*lV
is called a hermitian t
In thestudy of hel
product. Observe te
operatol" that maps t
lf the inner product (
can be computed as
r:I
T
-T- 2iT)
u(k) (D-15)
- T+ 2iT)
X,X2, ... , X Il
is a real-valued homo
11
',j~
01',
mijxiXj
in matrix form,
X2
-][:~: :~:
X Il
1~1"1 '~1"2
:~'j'l[:~,]
"
1~11l1l
6.
=x *M x
(E-1 )
x"
where the mi/s are any complex numbers and Xi is the comp\ex conjugate of Xi'
Since every hermitian form is assumed lo be real-valued, we have
lnd 1m [},(A)-})A)] f
r the single-output case
(E-2)
where M =~(M + Mi). lt is clear that M = M*. Thus every hermitian form
can be written as x*Mx with M = M*. A matrix M with the property M = M*
is ca\led a hermitian matrix.
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
where X and y are any vectors in (iC", q, then the hermitian form can be written
as
x*Mx = (X, Mx)
(M*x, X)
(Mx, X)
(E-4 )
Theorem E-3
Theorem E-1
The eigenvectors of a f:
are orthogonal.
Let il be any eigenvalue ofM and let e be an eigenvector ofM associated with il;
that is, Me = ile. Consider
(e, Me)
(e, ile)
il(e, e)
( E-5)
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.
Theorem E-2
Recall fram Section 2-6 that every square matrix which'maps (iC", iC) into itself
has a Jordan-form representation. The basis vectors that give a Jordan-form
representation consist of eigenvectors and generalized eigenvectors of the
matrix. We show that if a matrix is hermitian, then there is no generalized
eigenvector of grade k ~2; we show this by contradiction, Suppose there
exists a vector e such that (M - il1/e = O and (M - ill)k - le 1=0 for sorne
eigenvalue il of M. Consider now, for k ~2,
O= (M -il1)k-2 e, (M -il1)ke ) = (M -il1)k-I e, (M _ill)k-I e )
= Ii(M - il1)k - leW
and
where we have used t
fram (E-7), we obtai
(e, e) =0.
Since every eigen'
hermitian matrix M a
the eigenvectors assoc
normal. We consider)
the same eigenvalue.
vectors associated wil
orthonormal vectors f
Ll I
U2 =
i 1=}
i=}
( E-6)
um=e
567
M associated with A;
( E-5)
is a diagonal matrix.
and
( E-7)
( E-S)
where we have used the fact that the eigenvalues are reaL Subtracting (E-8)
from (E-7), we obtain (A - Aj)(ej, el) = O. Since A j, we conclude that
(e, e j ) =0.
Q.E.D.
+0
, (x, y)
= O. A vector
1t is clear that every
\. set of basis vectors
nly if
( E-6)
,"-1
um=e m-
(qk,em)qk
k=l
568
If the rank of H is r. SI
A21 ,
z2 >
. . . '''j'
; 2 > O a'ne
the orthonormal eigenv
---
----
Q = [ql
qz
~<
ql,e 2
Figure E-1
and
whose columns are orthonormal, such that
which implies
Q'Q
Ji::;:]
[ql
q2
qz
qll].
qll] = I
l(qll)*
Hence Q-I =Q*, and p- i =P*. A matrix Q with the property Q* =Q-I
is called a unitary matrix. We summarize what we have achieved in the fol!ow
ing.
Theorem E-4
tVA = PMP*
where !VI is a diagonal matrix with real eigenvalues ofM on the diagonal.
R*HQ=[
Ciearly, (E-14) implies
is orthonormal, we h:
becomes
In the fol!owing, we shal! utilize Theorem E-4 to develop the singular value
decompositions for matrices. Let H be an m x n matrix. Then the matrix
H*H is a square matrix of order n. Clearly H*H is hermitian; hence its eigen
values are al! real. BecauseH*His positive semidefinite; its eigenvalues are al!
non:negative (Theorem 8-19). Let il, i = 1, 2, ... ,n be the eigenvalues of H*H.
The set.{l ~O, i = 1,2, ... , n} is calledthe singular values of H. lfH is hermitian,
its singular values are equal to the absolute values of the eigenvalues of H.
Fot convenience, we arrange {,q such that
( E-9)
Every m x n matrix H o
R*HQ'
whereR*R = RR* = 1m ,
Al ~A2~'" ~Ar > O.
Although :E is uniql
arenot necessarily uniql
the corresponding colun
space spanned by the t
569
qz
...
q,.: q,+
(E-10)
Q*H*HQ=[~Z ~J
where 1;z = dag{Ai, J,i, ... ,
An.
Using Q = [Q
(E-11)
Q!H*HQz =0
and
QtH*HQ =1;2
which implies
1;-IQtH*HQ1;- =1
Q = [q
qz
...
qll].
( E-13)
(E-14 )
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 untary. Consider
.he property Q* = Q-
: acheved in the follow
. [Rt]
R*HQ=
R! H [ Ql
"] = [RtHQi
Qz
R*HQ
z
RtHQz]
R*HQ
z
z
( E-15)
Clearly, (E-14) mplies HQ = R1; and (E-12) mplies HQz =0. Because R
is orthonormal, we have RtR = I and R!R = O. Conseq uently, (E-15)
becomes
R*HQ=[~ ~J
)ll
the dagonal.
Every m x
R*HQ=[~~]
or
H=R[~ ~JQ*
where R*R = RR * = 1m , Q*Q = QQ* = 1m .and 1; = dag {A,A Z' . " . , A,} wth
A ~Az ... A,>O.
I
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 chsen as any orthonormal basis for the
spacespanned by the eigenvectors of H*H corresponding to Al. Hence Q
570
is not unique. Once Q is chosen, R can be computed from (E-14). The choice
of R z again may not be unique so long as [R 1 RzJ is unitary.
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
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.
The elements of matrices in this appendix are permitted to assume real or
complex numbers. Certainly all results still apply if they are limited to real
numbers. For a real matrix M we have M* = M', where the prime denotes
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
in nomenclature, all theorems in this appendix apply directly to real matrices.
Problems
E-11
H=l~
v"
011 :
-IJ
What arethesingul
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 .)
10
mm UJ
E-4
E-10
E-1 Ir Mis an n x n matrix with complex coefficients, verify that x*Mx is a real number
for any x in C" if and only if M* = M. Ir M is an n x n matrix with real coefficients, is it
true that x'Mx is a real number for any x in IR" if and only if M' = M?
E-2
fl
h2=l~
Il
(j
~J
E-5
E-G
Show that
= [Q
Qz], do we have
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
E-9 Show that all the singular values of a unitary matrix (including orthogonal matrix)
are equal to L
E-lO
What are the eigenvalues and singular values ofthe elementary matrices
El
=[
~ ~ ~]
-3 O 1
1 O
0J
1 O
E 2 = 0.5
0.8
O 1
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
largest and smallest singular values, which matrix has a larger condition number'! Roughly
speaking, a condition number gives the amplification factor of the relative errors in the
computation.
E-ll
W" =
eA'BB*e A', de
'=M?
E-12 Show that Theorem E-S reduces to Theorem E4 if H is square, helmitian, and posi
tive semidefinite. If H is square and hermitian (without being positive semidefinite),
what are the differences between Theorem E-S and Theorem EA?
E~13
[1 -IJ
A _
2
n Problem E-4
-1
1
-1]
-1
clearly x is a right ei
the operator d on 1
obtain
d(XY
Hence ), + J-lj is i
See Defini tion 2-12.
Next we prove tha
that Y/k is an eigenvalu
F
On the Matrix Equation
AM+MB=N
or
We now show tha"
cornrnon. We prove
be the characteristic
Then we have
is nonsingular. Inde(
eigenvalues of ~(B)
If Y/kI- A and B ha'
det ~(B) +0. Hence.
From (F-1), we ca
ror aH M in X
It is clear that the operator d rnaps (X, C) into itself and is a linear operator.
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~or 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,
2, ... , m ~ n, be the distinct eigenvalues of B. Then (A +11) is an eigenvalue of
d. Conversely, let Y/k> k = 1, 2, . " . , p.:s; n2 , be the distinct eigenvalues of d,
then ror each k,
Ax =AX
572
I
~
. .
573
d (M) = AM + MB = rkM
(rkI - 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
+ A).
(F-2)
AM + MB = N, where
le observe that aH the
ultiplication and addi
te this space by (X, C).
defined by
X
Id is a linear operator.
matrix equation.
is nonsingular. lndeed, if Ji i = 1, 2, ... , n, are eigen val ues of B, then (.L) are
eigenvalues of (B) (Problem 2-32) and det (B) = n(Ji) (Problem 2-22).
lf I1kI - A and B have no common eigenvalue, then (Ji) f.O, for aH i, and
det (B) f.O. Hence (B) is nonsingular.
From (F-l), we can develop the foHowing equalities
AM +-ME for aH M in
)f A and let Jij, for j = 1,
r- Ji) is an eigenvalue of
The summation ofthe products Of(I1kI- A)M = MB and IY., for i =0,1,2, ... , n,
with lY.o = 1, yields
inct eigenvalues of d,
which, together with (F-2), implies
O= M(B)
and y be nonzero n x 1
(F-3)
Since (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
eigenvalue. Now theeigenvalue of1]kI - A is ofthe form 17k - }e. Consequently,
for sorne iand forsome j,
or
Q.E.D.
".!
574
Corollary F-1 a
Theorem F-2
Proof
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
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-1 b
Ir all the eigenvalues of A have negative real parts, then for any N there exists a
unique M that satisfies the matrix equation
A*M +MA=-N
Corollary F-2b
(F-4)
lf aH the eigenvalues
exists a unique M tha
I'"
eA"Ne At dt
(F-5)
roo (A*eA'tNe AI
It=
+eA"NeA!A)dt=
roo ~(eA'INeAI)dt
Jo
00
=e tNeA'[
= - N
1=0
Q.E.D.
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~'
sively studied. in ihe literature. The reader is referred to References S9, S 18; and
S107.
. . . .
.'
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,
{:=-S)
(F-5)
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)
If the eigenvalues of A,
parts. Consequentiy,
esentation of d(M) =
>a unique M satisfying
(F -4). Because all the
al in (F-5) converges.
(A*)kNAk
k~O
Q.E.D.
~.Dd-
576
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-1
F-3 Solve Problem F-2 by using (F-S). Which method, by solving algebraic equations
F-6
t"
eA'NeB1dt
F-7 Find a matrix representation of d(M) = AMB- M, with A and B given in Problem
F-l and verify Theorem F-2.
!o-s
lo this appendix, we ,
sion to the polynomi
ioto two coprime poi
studied by the Freoch
system problems, ho\
the fractions of tran
lo the concepls of ce
hence its importance (
results in the scalar ca
extended to the matrix
for numerical comput
dix A will be constant
F-9 Transform A*M +MA = - N into a discrete-time Lyapunov equation by using the
transformation
A-->(Ad - 1)(Ad + 1)-1
G-1
Coprimene!
G
Polynomials and
Polynomial Matrices
+MB with
= 1,
N be expressed as
IOV
In this appendix, we shall study the coprimeness of two polynomials, its exten
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
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
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.
G-1
Coprimeness of Polynomials
'In this appendix, capital letterswithoutboldface are also used lo denote scalars.
577
578
Theorem G-1
5. Go to step 2.
Let D(s) and N(s) be two polynomials and let D(s) 1=0. Then there exist unique
polynomials Q(s) and R(s) such that
N(s)=Q(s)D(s) +R(s)
and
(G-1 )
Proof
If deg N(s) < deg D(s), then Q(s) =0 and R(s) = N(s). Let deg N(s) :::::deg D(s).
Then by direct division, we have the relation in (G-1) with Q(s) as the quotient
and R(s) as the remainder with deg R(s) < deg D(s). For example, if N(s) =
2s 3 + 3s - 1 and D(s) = S2 + s - 2, then by long division, we have
2s -2
S2 + s - 2)2s 3 + O + 3s - 1
2s 3 + 2s 2 - 4s
- 2s 2 + 7s-1
-2s2 - 2s +4
9s-5
and
Now we show that Q(s) and R(s) are unique. Suppose there are other QI(S)
and R 1(s) such that
N(s) =Q(s)D(s) +R(s)=Q(s)D(s) +R(s)
Definition G-1
A polynomial R(s) is :
common divisor of D
D(s) and N(s). Ir a gCI
then D(s) and N(s) an
In other words, t\\
mon factors. Ir they
The gcd is unique
for any nonzero numl
the gcd is uniq ue.
Given two polyno:
called the euclidean a
N(s) =
D(s) =
R(s)=
which implies
(Q(s) - Q (s))D(s) = R (s) - R(s)
(G-2)
and
R(s)
R(s)
Q.E.D.
1. Q(s) =0.
2. lf deg N(s) < deg D(s), stop.
3. Let N m and D;, be the leading coefficients of N(s) and D(s) with m> n:
R p - 2 (s) =
Rp_(s)=
N
'.
N(s)+-N(s) --:-'D m s'"'-"O(s)
"
579
5. Go to step 2.
Jen there exist unique
(G-1 )
N(s)=Q(s)D(s) +R(s)
D(S)=Q2(S)R(s) +R 2(s)
R (s) = Q3(s)R 2(s) + R 3(s)
1(s)
(G-2)
Q.E.D.
into a programmatic
R p_2(s)=Qp(s)R p_(s)
(G-3)
+ Rp(s)
Rp_1(s)=Qp+l(S)Rp(s) +0
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
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
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
equation shows that the claim holds for Rz(s). Proceeding downward, the
.
..
'Ir D" has a very small absolute value, large errors mayarise 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
claim can be verified for every R , i = 1, 2, ... ,p. Hence there exist polynomials
X(s) and Y(s) such that
Ris) = X(s)D(s)
+ Y(s)N(s)
(G-4)
This equation implies that every common divisor of D(s) and N(s) divides Ris).
Indeed if C(s) is a common factor, that is, D(s) = D(s)C(s) and N(s) = (s)C(s),
then we have Rp(s) = [X(s)D(s) + Y(s)(s)]C(s). Hence Rp(s) is divisible by
every common divisor of D(s) and N(s). Consequently, Ris) is a gcd. AH gcd
differ at most by a constant; hence every gcd can be expressed in the form of
(G-4). This is stated as a theorem.
Theorem G-2
Proof
Every gcd of the polynomials D(s) and N(s) is expressible in the form
R(s) = X(s)D(s)
+ Y(s)N(s)
which implies deg
Consider two polynomials D(s) and N(s) with D(s) 1=- O. Then D(s) and N(s) are
coprime if and only if any one of the foHowing conditions holds:
We give a remar~
gcd of D(s) and N(s).
which are roots of R(.
s in iC. Hence if D(s)
rank 1 for almost al! 5
s in iC.
In the following, v
D(s) =
N(s) =
1. For every s in iC, the field of complex numbers, or for every reol of D(s), the
2 x 1 matrix [D(S)] has rank 1.
N(s)
2. There exist lwo polynomials X(s) and Y(s) such that 3
X(s)D(s)
+ Y(s)N(s) =
and let
(G-5)
D(s)
with no assumption o
B(s)
----
A(s)
(G-6)
or, equivalently,
-B(s)D(s) +A(s)N(s)=[ -B(s)
.and
A(Sn[D(S)]=o
N(s)
(G-7)
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 txt.
Do
DI
Dz
Do
Di
No
NI
N z
No
Ni
S~ ------------
581
COPRIMENESS Of POLYNOMIALS
Proof
1. Ir D(s) and N(s) are coprime, there is no s in iC such that D(s) = Oand N(s) =0.
Hence the matrix [D(s) N(s)]' has rank 1 for every s in iC. lf D(s) and N(s)
are not coprime, there exists at least one s in iC such that D(s) = Oand N(s) = O
and the rank of [D(s) N(s)]' is zero at that s. The matrix [D(s) N(s)]'
has rank 1 at every s except at the roots of D(s); hence it is necessary to check
its rank only at the roots of D(s).
2. This foHows directly from Equation (G-4).
3. Ir D(s) and N(s) are not coprime, by canceling their nontrivial common
divisor, we obtain A(s) and B(s) with deg A(s) < deg D(s). Ir D(s) and N(s) are
coprime, there exist polynomials X(s) and Y(s) such that X(s)D(s) + Y(s)N(s)
= L The substitution of N(s) = B(s)D(s)/A(s) into it yields
in the form
[X(s)A(s)
+ Y(s)B(s)]D(s) = A(s)
which implies deg A(s) 2>:deg D(s). This completes the proor.
Q.E.D.
11
We give a remark concerning the rank of [D(s) N(s)]'. Let R(s) be the
gcd of D(s) and N(s). Then the matrix [D(s) N(s)]' has a rank ofO at those s
which are roots of R(s). The matrix however still has a rank of 1 at all other
s in iC. Hence if D(s) and N(s) are not coprime, the matrix [D(s) N(s)]' has
rank 1 for almost al! s in iC. If they are coprime, the matrix has rank 1 for al!
s in iC.
In the following, we discuss a method of solving Equation (G-7). Let
+DlIs n
+Nmsm
(G-S)
(G-9)
and let
A(s)=A o +As +
B(s)=B o +Bs +
(G-5)
with no assumption of A II _
(G-6)
)(S)J =0
les)
(G-7)
~
-
This
N
No
Nz
NI
00
f.O and Bm _
_::~: :~-_.
__
Nm
Nm -
+An_ls"-
+Bm_sm-
I
(G-10)
(G-11 )
f.O. Define
:w,
_:_.c_.__ :._ }
O O
O}
Nm
O
:
No
NI
Nm - I
n
rows
Nm
( G-12)
582
(G-13)
and
Example 1
+ 2s 3 -
S2 - s
+1
1 -1 -1
2 -2
O
O
1 -1 -1
O
2 -2
O
O
O
1 -1 -1
2 -2
S= ----------------------------3 -2
2
1
O
O
O
-2
O
3
2
1
O
O
.1
O
O
3 -2
2
O
O
O
3 -2
O
2.
1
(G-14)
We use the row searching algorithm discussed in Appendix A t() search the
linearly dependent rows bf s. The circled pivots are chosen as shown. The
result is as follows:
COPRIMENESS Of POLYNOMIALS
1
O
1
O
1
O
O
1
F:S= -3 -1
O -3
1
O
-3
O
O
-0.5
O
1
O
---------------------------------;..
O
O
0.5 -1.5 -0.5
O
he existence of a non
ethods and "eanned"
g centers to solve this
solution whieh yields
discussed in Appendix
>y an example
( -1
O
O
O
O
O
O
O
O
O
O
O
-1
-1
1
6
(~}D
O
O
:8
O
O
-2
2
O
-1
2 "'.. -=-- 2-'
- ~I
4
-4
O ~S
(~~
O
17 -17
O
O
O
O
O
O
O
O
O
-2
2
-1
-1
(G-15)
Note that the matrix F is the one defined in Equation (A-9) of Appendix A.
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
[1.5
0.5
O: -0.5
O 1
0]8 =0
and
1 and N(s) =S3 +2s
tI,
'"
583
Nls)
D(s)
B(s)
A{.S)
-0.5s -1.5
s2-0.5
+3
-2s 2 +1
( G-16)
O
O
-2
O
O
O
1
(G-14 )
:ndix A to searchthe
losen as shown. The
We note that the B(s) and A(s) eomputed from [-B A]S =0 by using the
first linear1y dependent row of S, which corresponds to the first zero row of the
right-hand-side matrix of (G-15), are coprime. Suppose not, then there exist
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
(G-15). Thisis nat possible. Hence we conclude that the B(s) and A(s) in
(G-,16) are cbprime.
In theSylvester matrix.in (G-12), the coefficients of D(s) and N(s) are arranged
in the'ascending power of s. Clear1Y' we can also arrange the eoefficients in the
descendingpowef of s as
584
Il
~ ~
- -
In
- - - - -:
m-l
...
~l
- -
1- -
::
:::::
::
NI
o o
Nm
Nm- I
No
(G-17)
Using this Sylvester matrix, Corollary 0-3 still holds. However, if we use
(0-17) to compute A(s) and B(s), care must be taken in determining their degrees.
To see this, we write
[-B A]S=[-B m _ 1
-Bo:AI _1
..
...
Ao]S=O
Consider a rational function N(s)jD(s) = N(s)D -I(S), where D(s) and N(s)
are polynomials.1f D(s) and N(s) are coprime, the rational fun"ction N(s)/D(s)
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
procedure to improve its computability and to lay ground for the extension to
the matrix case.
Consider the two polynomials
( G-18~
and
(G-19)
DI
N1
- - - D - - -6 o
No
N1
O
O
O
O
Dn - I
Nn - I
- - - - - - -
O
O
Dn
N Il
-_- - n I
Nn-
O
O
O
} 1 block row
O
D - - - - - - - - -
- -
'"0
Since Sk is a 2(k + 1) x
A necessary condition
or k ::; n - 1. Hence w
linearly independent 1
Let A(s) and B(s) b
and
Then from the equatio
Nn
It is a 2(k + 1) x (n + k
has two rows formed
and N(s). We note tI
right by one column.
Sk is a square matrix;
We search now lin
For convenience of di
rows formed from Ni
linearly independent (
Dn . O and the struct
elements aboye Dn in
is linearly independen
may not be linearly iI
Sk' we can readily see
previous rows, then al
Hence the total numbe
tonically as k increase~
number of linearly in<
how many more blod
Let v be the total
words, the v N rows in
and al! N rows, not in
on their previous row~
dent; hence we have
+ 1 block rows
O
O
( G-2)
_.
__
--~-
.------~--------~----~--
- --- --
- - ~ ~ - - ---------~
-- -- -
----
---~-~~--
~1
.,. DI DO
._---------
...
NI
No J
(G-17)
However, if we use
their degrees.
~rmining
. Ao]S=O
-1, and consequently,
e gcd of D(s) and N(s)
. of forro Sk. No such
nge the coefficients of
ix.
:;tions
where D(s) and N(s)
loal f~nction !v (s)/D(s)
'ble. In application, it
to an irreducible one.
is reduction, as shown
1, we shall modify the
Id for the extension to
For convenience of discussion, the rows formed froin Di are calJed D rows; the
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
D" f.O and the structure of Sk' For example, if a new block is added to Sk> all
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
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
previous rows, then all N rows in subsequent blocks wi.ll be linearly dependent.
Hence the total number of linearly independent N rows in Sk will increase mono
tonically as k increases. However, once the number ceases to increase, the total
number of linearly independent N rows will remain to be the same no matter
how many more block rows are added to Sk'
Let v be the total number of linear independent N rows in Soo' In other
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
on their previous rows. Note that all (k + 1) D rows in Skare linearly indepen
dent; hence we have
rank S.= {
k
li(5;
2(k+1)
(k +1) +v
fork':::;;v-1
for k;:::v
and
= Aa +- Al.)
-l'
B(s)=B o +B 1 s+'"
'""/~/{')
+Bkl
( G-20)
(G-23)
(G-22)
} l block row
(G-21 )
( G-18)
( G-19)
585
(G-24)
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
N rows in Sv.
586
Theorem G-4
G-3
Consider two polynomials D(s) and N(s) with deg N(s)::s:deg D(s)=n. The
A(s) and B(s) solved from (G-24) by using the first linearly dependent row of Sk
are coprime, where Sk is defined as in (G-20).
Proof
Let v be the least integer such that the last N row ofS, is linearly dependent of its
previous rows. Since all D rows are linearly independent, this N row is the
first linearly dependent row in Sk, for k ;?: v. Corresponding to this dependent
row, the solution of(G-24) yields an A(s)of degree v and B(s) such that N(s)/D(s) =
B(s)/ A(s). Now if A(s) and B(s) are not coprime, there exist ,.4(s) of a degree
smaller than v and 8(s) such that
N(s) _ B(s) _ 8(s)
This implies that a linearly dependent row appears before the last N row of
S,. This is not possible. Hence A(s) and B(s) are coprime.
Q.E. D.
We note that if k ;?:n, the resultant Sk has more rows than columns and sol u
tions always exist in (G-24). For example, if k =n, then A(s) =D(s) and B(s) =
N(s) are solutions of (G-24) and (G-23); if k = n + 1, then A(s) =D(s)(s + e) and
B(s) = N(s)(s + e), for any real e, are solutions of (G-24). Clearly these solutions
are of no interest to uso
Corollary G-4
The two polynomials D(s) and N(s) with deg N(s)::s: deg D(s) = n are coprime if
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
This corollary follows directly from Theorem G-4, and its proof is left as an
exercise. From Theorem G-4, we see that the reduction of N(s)/D(s) hinges on
the search of the first linearly dependent row in Sk. The row searching algorithm
in Appendix A is developed exactly for this purpose. It is illustrated in Example
1 of Section G-1 and will not be repeated.
The matrixSk has a special structure: Every block row is a shift ofits previous
block row. By using this shifting property, a very efficient method is developed
in Reference S140 to search tije first lin~arly dependent row of Sk' To fully
utilize the shifting property, theelimination :must be carried out from left to
right and the meihod is generally not numericlly stable. For hand calculation,
the method can definitely be used'in "place bf the ro\\' s~arching itlgorithm.
The result, however, may not be in an echelon form in the matrix case. '
Polynomia
1. M ultiplication of
2. Interchange any t,
3. Addition of the pr
row or column.
These operations ca!
n = 5, of the form
1
O
El = O
O
O
O O O
100
O 1 O
O O e
O O O
O
O
O
O
1
E-l
O O O O
They are again elem
operares on lle row~
operates on the colum
and fifth row. E 3 A(s)
fourth row of A(s); w]
A(s) and d(s) to the S(
operations and A(s)E i
The set of polynor
plication is not a poly!
tions, then the set beco
rational functions, the
developed for matrices
equally applicable. 1
field of rational functic
-~
---_ ..
------:---_._---~.-
.. _.~-------~._ .. _-----------_.~-------,-------------_
.. _-~----_._-.-._----------_.-.~~ ..
--
-.
587
POlYNOMIAl MATRICES
G-3
;).:::;;deg D(s) = n. The
ly dependent row of Sk
Polynomial Matrices
These operations can be carried out by using the elementary matrices, for
n = 5, of the form
1 O O O O
I O O O
1 O O O O
O O O O I
O O 1 O O
O O O e O
O O O O 1
Ez = O O 1 O O
O O O 1 O
O 1 O O O
EI =
fore the last N row of
me.
Q.E.D.
E3
I O O
O 1 O
= O O 1
O des) O
O O
O
O
1
O O O O
O
O
O
I
(G-25)
with e 1=0 and des) is a polynomial. We note that the determinants of these
elementary matrices are nonzero constants and are independent of s. Their
inverses are
E-I I
D(s)
n are coprime if
I (G-20) is nonsingular
rows in Sn -1 is equal
I O O O O
O 1 O O O,
O O 1 O O
O O O c- 1 O
O O O O 1
E; I =E z
I
O
1
O
E-lO
O
3
O -des)
O
O
O
I
O
O
O
O
O
O
1
O
O
1
(G-26)
588
coturnn
+2
s-1
...
s-I~J
S
+2
is (s + 2)2 - (s - 1)2 = 6s + 3 which is not the zero element in the field of rational
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
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
case of rational matrices, is equal to the zero element of lR(s), the field of rational
functions, then the polynomial matrix is singular. For example, the polynomial
matrix
[ S2
kth
S +2
+3s +2
For this example, we may choose a(s) = 1 and a 2(s) = -1/(s + 1).
Let a(s) be the least common denominator of a (s) and a 2(s), and let IX (s) =
a(sp(s) and &2(S) = a(s)cx2(S). Then (G-27) implies
[:
...
J
O al,k l
O O
O \1
O O
lf
al,k-l
O
O
O O
O
O
O O
4In DefinifiQn 2-2 ira field is replaced by a ring ~; (see Footnote 3 or chapter 2), then(E{, 1Ki;l is
calledal1lod4l.e over.the ringo A fl X 1 or I x fl polynomial vector can be considered"as an demenl
or therational. vector space (W(s), ~(s)), or an elementar the modllle (IR "[s], ~[s]). A set or poly
nomal veciors is Hnearly independentover the field ~(s) ir and only ir the set s linearly indeperident
over thering ~[s]. See Relerence S34.
;jtep 1. Vli(s)=\s) al
nonzero.
Step 2. Ir all e1ement:
zero, go to step 6;
Step 3. Search the ele
and bring it to the I
element monie. Ce
S.tep 4. Compute j
Add the produet
M(s), i=2, 3, ...
~dditional column ope!
ioto a.diagonal matrix. ca
In (G-28), we set, ror red\,
5 [[
589
POLYNOMIAL MATRICES
klth
kzth
k 3 th
k,th
column
column
column
column
(G-27)
l/(s + 1).
d o:z(s), and let C(I(S) =
O
O
O
O al.k,
O O
O 0
O
O
O
O
O O
O O
O
O
O O
al,k 1
al ,k,
al,kz+ 1
al,k,
Gl,k r + 1
al,k 2
a2,kz+ 1
al.k,.
a2,k,.+ 1
a3.k,
a3,kr
G3,k,.+ 1
O
O
O
O
O
O
a,..k r
G',krT 1
al.k,-I
+1
O
O
...
t_28l
"'1
The first r rows are nonzero rows with polynomial elements. The left-most
nonzero element, a,k, of each row is a monic polynomial. The column position
of a,k must be on the right hand side of the column position of Q - \ ,k _\' that
is, k 1 < k z < ... < k,.. 5 The degrees of all elements aboye a,k are smaller than
the degree of a,k, that is, deg aj,k< deg a,k, for = 2,3, ... , r;j = 1, 2, ... , -1.
If deg a,k = O, then aj,k = O, for j = 1, 2, ... , - 1. The matrix in (G-28) with
these properties is said to be in the Hermte (row)form. A matrix in the form
of (G-28) without the property deg aj,k < deg a,k has no special name; it is
just a upper right triangular matrix.
If A is a matrix with real e\ements, the Hermite form is said to be in the
echelon (row)form. In this case, the leftmost nonzero e\ement of every nonzero
row is 1 and all elements aboye it are zero (becatise deg a,k = O). In other
words, the leftmost nonzero element 1 is the only nonzero element in that
column. By reversing the order of rows and the order of columns, we will
obtain a different, but equivalent echelon (row) formo This form will often
appear in the remainder of this appendix.
Every polynomial matrix can be triangularized by using exclusively el
ementary row operations. This is presented as an algorithm.
Triangularization Procedure
Step L M(s) = A(s) and oeieie lts cotUlllS ll'Oll1 lel unlil Lfle rsl COIUfIlll S
nonzero.
Step 2. If all elements, except the first element, of the first column of M(s) are
zero, go to step 6; otherwise, go to step 3.
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 -QI(S) to the ith row of
M 1(s), i = 2,3, ... ,n. Callthe resulting matrix Mz(s) = (mz(s)).
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.
In (G-28), we set, ror reducing lhe size of the matrix, k 3 =k , + l.
5 Ir
590
Every polynomial matrix can be transformed into the Hermite row form in
(G-28) by a sequence of elementary row operations.
1
Hence, we have
ro
-s-1:
l~ ---;2~~ -~ i-:
In this example,
stop at step 4. Dual
formed, by a seq uel
triangular matrix or I
Since the determi
independent of s, so
matrices are called UI~
Definition G-2
Example 1
A square polynomial
is nonzero and indep(
A(')~[-~
-1
-J.
3dl
S2 +s - 2
S2 +25 -1
[1
[-1
,'+,-2]
3s +1
s
-1
-s'-s+2 l
-4
s+ 1
S3
+ 52 + 5 + 1
]
[1O S3-,'-,+2
+s + 1
52 +2s -1
[~
-,' -s+2 ]
5 +1
Theorem G-6
+S2
A square polynomial
nomial matrix.
5 +1
[~ ~ +11 (G~9)
In the first step, we interchange the first and second row. This is achieved by
the multiplicati.on of A(s) by the element8.ry matrix derroted by 1 listed 2.t th" eDr4
of this paragraph. In the second step, we multiply the first row by - 1, add the
product of the first row and s to the second row and add the product of the first
row and - 1 to the third row. This is achieved by the matrix denoted by 2.
In the third step, we interchange the second and third row. We then add the
product of the second row and - (S2 + 1) to the third row in step 4. In the last
step, we add the product of the second row and s to the first row so .that the
degree of a12(s) is smaller than that of all(s). The corresponding elementary
matrices are
1
O
1
_(S2
+ 1)
~] [~ ~ ~]O
1
[- ~ ~ ~]
-1
O 1
[~O O~ ~]
1
Proof
Let M(s) be unimod
[Adj M(s)J/det M(s) i
polynomial matrices.
nomials. Sinee M(s)
The ony way lor poly:
are both nonzero con:
In the proof, we ha
unimodular matrix.
written as a product o
Unimodular matr
tions. For every 5 in
. bers. In general, a nI
POLYNOMIAL MATRICES
59!
[~
t of M(s) by at least 1;
( G-30)
Hence, we have
3s + 1 ]
s2+ s -2
S2 +2s -1
=1 __ ~ __ ~~I_-J
loo
(G-31)
!!l
Theorem G-6
-S+2 ]
s3+ s2+ s +1
s+l
-5
[i
~ + 1]
(G-29)
,! ~][! HJ
592
G-4
In this section, the concept of coprimeness for scalar polynomials will be extend
ed to polynomial matrices. Since the multiplication of matrices does not com
mute in general, the situation here is more complicated.
Consider A(s) = B(s)C(s), where A(s), B(s), and C(s) are polynomial matrices
of appropriate orders. We call C(s) a right divisor of A(s) and A(s) a left multiple
of C(s). Similar!y we call B(s) a left divisor of A(s) and A(s) a right multiple of B(s).
Consider two polynomial matrices N(s) and O(s). The square polynomial
matrix R(s) is called a common right divisor of N(s) and D(s) if there exist poly
nominal matrices N(s) and D(s) such that
N(s) = (s)R(s)
O(s) = D(s)R(s)
(G-32)
U 11
[ U Z1
Then we have
In this definition, N(s) and O(s) are required to have the same number of columns.
Their numbers of rows, however, can be different.
and
Definiton G-3
A square polynomial matrix R(s) is a greatest common right divisor (gcrd) of
N(s) and O(s) if R(s) is a common right divisor of N(s) and O(s) and is a left
multiple of every common right divisor of N(s) and O(s). Ir a gcrd is a uni
modular matrix, then N(s) and O(s) are said to be right coprime.
I!l
Theorem G-7
Consider the p x p and q x p polynomial matrices O(s) and N(s). Then they
llave a gcrd R(s) expressible In. th.e forrri
R(s)
X(s)O(s)
+ Y(s)N(s)
<;
[
"
-1
O
O
-( -.
-s-1
-1
-sy+~-+
Proof
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
p[U
11 (s)
q U Z1 (s)
U 12(S)][D(S)] =
Uds)_ N(s)
U(s)
[R(S)]}P
O}q
593
[V
D(S)] =
11 (s)
[ N(s)
V 2(S)
and
D(s) = V (s)R(s)
Then they
(G-35)
Hence, R(s) is a common right divisor of D(s) and N(s). From (G-33) we have
(G-36)
R(s)=Ull(s)D(s) +Vds)N(s)
Let R(s) be any common right divisor of D(s) and N(s); that is, D(s) = D(s)R(s)
and N(s) = N(s)R (s). The substitution of these into (G-36) yields R(s) =
[U 11 (s)D(s) + Vds)N(s)]R(s); that is, R(s) is a left multiple of R(s). Hence,
we conc1ude that R(s) is a gcrd.. This establishes the theorem.
Q.E.D.
Example 2
[-1
S2
+2s-1J
J'l
+
;0
'ices, respectively.
V 12(S)][R(S)]
Vds)
O
N(s)=V 2 (s)R(s)
D(s)=[_~
and N(s).
(G-34)
Then we have
le number of columns.
VileS)
[ v 2 (s)
-s-1
S2
+s + 1
: - (S2
,
1)
'1
__
X- L
O
R(s) =
[~
which is not a unimodular matrix. Hence D(s) and N(s) are not right coprime. I
p
q
(G-33)
Let W(s) be any p x p unimodular matrix, and let R(s) be a gcrd'of D(s) and
N(s). Then W(s)R(s) isalso a gcrd of N(s) and O(s). This can be proved by
premultiplication of the unimodular matrix diag{W(s), I q } to (G-33). Hence
594
the gcrd of 0(.'1) and N(s) is not unique. For example, the polynomial matrix
1
W(S)=[i .'1:
lIJ[OI
2J=[l+1
.'1+1
i
k
2s k+S+3J
2s +s+l
is also a gcrd of the 0(.'1) and N(s) in the example. We see that the degrees of
the elements of R(s) may be larger than those ofO(s) and N(s). This phenome
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
be related by a unimodular matrix? The answer is affirmative if the matrix
[O'(s) N'(s)]' is of full column rank.
3. There exist no po
such that B(s)D(s)
-B(
and
Corollary G -7
Proof
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,
we have two polynomial matrices W (s) and W 2(.'1) such that
R(s) = W (s)R(s)
R(s) = W 2(s)R(s).
which imply
R(s) = W (s)W2(s)R(s)
(G-37)
Since R(s) is nonsingular, we have W (s)W 2(.'1) = l. Hence, both W (s) and
W 2(s) are unimodular matrices. Consequently, RJ(s)=Wz(s)R(s) i5 also
nonsingular.
Q.E.u.
In application, we often have the condition that 0(.'1) is nonsingular. With
this condition, the condition in Corollary G-7 is always mel. In this case, the
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.
Theorem G-8
Let O(s) and N(s) be p x p and q x p potynomial matrices, and let 0(.'1) be
nonsingular. Then 0(.'1) and N(s) are right coptime if and only if any one of the
'following conditions holds:
1. For every s in
mtrix
e, or fofevery
For every s in e, t
complex numbers.
same as the rank ,
unimodular and h.
rank p for every s i
is a polynomial o
following the nonsi
such that det R(s)
Hence, weconclud.
of [D'(s) N'(s)]' i5
uJ. tt rsL Slaltll1(
condition is met at
check the rank onl
2. To show the secon
he polynomial matrix
595
D(s)J
[ N(s)
has rank p (in the field of complex numbers).
2. There exist polynomial matrices X(s) and Y(s) of order p x p and p x q
such that
2s +s+ 3J
2s k +s+ 1
X(s)D(s) +Y(s)N(s) = 1
and
f D(s) is nonsingular,
elated by unimodular
;).
(G-37)
A(s)]
[~~:Q =0
(G-39)
Proof
(G-38)
Uds)J[D(S)] = [R(S)]
Uds) N(s)
O
(G-40)
(G-41 )
596
which implies
The substitution o
+ Uds)N(s) = O
This is already in the form of (G-39) if we identify B(s) = - U 21 (s) and A(s) =
U 22 (s). Hence what remains to be proved is the following inequality:
deg det A(s) = deg det U 22 (s) < deg det D(s)
D(s) = V 11 (s)R(s)
Hence, if D(s) is nonsingular, so are V II(S) and R(s). Using the identity
[1
-V 21 (S)V/(s)
= [V
OII(S)
O][V
1
11 (s)
V 21 (S)
VdS)]
Vds)
VI2 (S)
]
V 22(S) - V21(S)V /(s)V ds)
(G-43)
Because of (G-47)
D(s).
Conversely, if dI
ment and conclude
In the following, w
left coprime. Given t\
of rows, then, similar
formations so that
(G-44)
q[Q(s
VdS)]-I[ 1
Vds)
-V 21 (S)V/(s)
0J-l
[V 11 (S)
= O
~dS)]-1
Theorem G-S'
l..}.
1. For every s in
matrix
e,
or
0IJ
has rank q in the fie
2. There exist polynon
that
The
597
(G-46)
Since Ves) is unimodular, we have deg det Ves) =0. Hence, (G-46) implies
deg det V 11 (s) = deg det Uds)
D(s)
(G-47)
This relation holds for any unimodu.lar matrix U(s) and its inverse Ves) =
U- 1(s). We shal1 now use (G-47) and (G-42) to establish statement 3 of
Theorem G-S. Ir D(s) and N(s) are not right coprime, there exists a gcrd
R(s) with deg det R(sO. From (G-42), we have det D(s)=det V 11 (S)
x det R(s), which implies
deg det D(s) > deg det V 11(S)
(G-42)
(G-43)
as
Because of (G-47) and A(s) = - Uds), we conclude deg det A(s) < deg det
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
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
(G-44)
q[Q(s)
,-.J'--..
,-.J'--..
q P
V 21(S)
Vds) }p
tr
fol1owing (G-44).
r and can be readily
-hand-side matrix of
Id side oftheequality
- V2l(s)V/(s)
0J
1
Theorem G-S'
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
ing condtions holds:
1. For every s in ic, or for every root of the determinant of A(s), the q x (q + p)
matrix
[A(s)
B(s)]
(G-48)
the fol1owing.
The
598
or, equivalently,
- A(s)N(s) + B(s)D(s) = [A(s)
B(s)] [-
~~;~] =0
(G-SO)
G-S
Corollary G-S
Column- al
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
U(S)[D(S)]~[U11(S)
N(s) -
U 21 (S)
or
U dS)][D(S)] = [R(S)]
Uds) N(s)
O
Then we have
1. U 22 (S) and U 21 (S) are left coprime.
2. Uds) is nonsingular and N(s)D- 1(s) = - Ui}(s)U 21 (s).
3. D(s) and N(s) are right coprime if and only if deg det D(s) =deg det Uds).
Proof
~ d~
d 11
l2
n21
d 21
r'
13
dn 13
n23
'
n 22
-
--
d 22
d23
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)
a(s)U 21 (S)D(s) =0
a(s)U 21 (S) =0
Hence we have a(s)[U 21 (S) Uds)] =0. This contradicts with the fact that
[U 21 U 22 ] has rank q in the field of rational functions. Hence we conclude
that U 22 (s) is nonsingular. Consequently, from U 21 D + U 22 N =0, we have
Part 3 ofthis theorem has been essentia\ly est~blished in the proof ofTheorem
G-S. Indeed, if D(s) and N(s) are right oprime; then R(s) in (G-42) is unimodular
and deg detD(s) =deg det V 11(S). This, together with (G-47), implies .deg det
D(s) = deg det U~2(S), Conversely; if deg det D(s) = deg det Uds), we. may
reverse the above argument to conclude degdet R(s) = O. SinceRes) is non
where
l-:<LI-.
11 "0
'-' l
bciM(~
briM(~
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
D(s)
(G-50)
Consider two polynomial matrices N(s) and D(s). Ir D(s) is square and non
singular, then the matrix N(s)D- 1 (s) is generally a rational matrix. Conversely,
given a q x p rational matrix C(s), we can always factor G(s) as
C(s) = N(S)0-1(S)
(G-51 )
C(s) = A- 1 (s)B(s)
(G-52)
o).
D(s) =deg det Uds).
l~
d ll
nZ1
d Z1
n1Z
d 12
n22
-
d22
l3 =~el
the proofofTheorem
n (G-42) is unimodular
G-47), implies degdet
:g det Uds), we may
o. Since R(s) is non-
~e3 =[~ll
n23
d Z3
n22
d ez
nZ1
-
del
n23
d e3
n21
n1Z
13
nZ3
nZZ
J
O
O
O
0l'
O .
d ez
22
l"
~l[dd Tl ~l3J
d,.z
d,z
1Z
ld"1
d"1
Z1
d,.z
=0 implies
nlZ
d ez
n Z3
d,z
ll
12
Z1
22
(.G-53)
dd
n23
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
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
3-5). In terms of the elements of C(s), the properness of C(s) can be easily
determined. For example, the rational matrix C(s) is proper if and only if
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.
Given a polynomial column or ro.w vector, its degree is definedas the highest
power of s in all entries of the vector.. We define
b~J"l(s) = the degree of the ith column of M(s) .
600
Definition G-4
and call b ei column degree and by row degree. For exarnple, for
M(s)=[S+l
s-l
we have b el
1, b e2
S3 +2s +1
S3
A nonsingular p x p poIy
(G-54)
= 3, b e3 = 1, and brl = 3, b r2 = 3.
Theorem G-9
brjB(s) ~brjA(s)
for j = 1, 2, ... ,q.
Proof
We write N(s) = (;(s)O(s). Let niis) be the ijth elernent of N(s). Then we have
p
nij(s) =
gik(S)dds)
= 1,2, ... , q
k= 1
Note that, for every element in the jth column of N(s), the sunrnation is carried
over the jth column of O(s). Ir G(s) is proper, the degree of ni)s), i = 1,2, ... , q,
is srnal1er than or equal to the highest degree in dkj(s), k = 1. 2, __ . ,p. Hence
we have
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\
[1
2]
S2
D(s)= [ s+l
[-2S1 -1
(G-55)
. M(s) =\
mple, for
Definition G-4
601
(G-54)
bciM(s)
i~1
b,.iM(S)
i= 1
)(s)J
_[3s2 +2s 3
M(s) -
l(S)J
2s
s +s -
1J
(G-56)
., q
.he summation is carried
~e of nij(s), i = 1, 2, .. _, q,
, k = 1. 2, ... , p. Hence
(G-57)
where Hc(s) = diag{sk ci, i = 1,2, _. _, p}. The constant matrixM hc will be called
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 Mc(s) conta-ins the
remaining terms and its ith column has a degree smaller than k ci ' For example,
the M(s) in (G-56) can be written as
Q.E_D.
(G-55)
~J
mcept.
(G-58)
where R(s) =diag{i'-, i = 1, 2, .. _, p} and k"i = b,_M(s) is the degree of the ith
row. M h ,. will be called the row-degree coefficient matrix; its ith row is the coef
ficients of the ith row of M(s) associated with Sk'i. The polynomial matrix
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
602
In terms of (G-58), we have that M(s) is row reduced if and only if its row-degree
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 only if
N(s) = NheHc(s)
+ Nds) =
[N lle + Nle(s)He-1(S)]He(s)
where 6eiDlc(S) < 6ci D(S) ~ J1.i, He(s) ~ diag {si", sl'" . .. , sl'P}, and 6 ei N1e(s) < J1.i.
Then we have
(;(s) ~ N(s)D- 1(s) = [N',e + Nte{s)He- 1(S)] [D he + D 1e (s)H e- 1(s)]-1
Ul(~
to M(s), we obtain
Clearly Nds)H e- 1 (s) and D1e(S)He- I(S) both approach zero as s-+ oo. Hence we
have
lim (;(s) = N',eD,;;, 1
M(s)U 1
s-oo
Example
Consider
+1
M(s)= 2s - 2
S
-s
52
+2s +1
~2S2
+1
5s 2 -2s
An algorithm is' av
column- or rbw~reducf
to the matrix case.
603
-2
5
and let 0(5) be column
'per (strictly proper) if
and is singular. Hence, M(s) is not column reduced. Since M llc is singular there
exist (;(1- (;(2' and (;(3 such that
~)]
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
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)
1(s)J Oc (s)
Ud')
>1c(s)Hc-1(s)] -1
oas s-oo.
~ [~
Cl. 1 S
1
Cl. 3 S
[' +1
M(s)UI(s)= ~~2
-2s 2
~]
to M(s), we obtain
Hence';'e
~] ~[~
3s
1
5s +1
-6s +1
-2s
~J~M'(S)
It can be readily
Iil
From the example, we see that by a proper elementary column operation, the
column degree can be reduced, whereas the determinantal degree remains
unchanged. Hence, by a sequence of eleYrt~Dtary s01~~n~~ ~:J~r~l~:~:J~).8, 2. ;..~<>Jy.
nomial matrix can be reduced to be column reduced. We summarize this with
the earlier statement as a theorm.
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.
604
Theorem G-1 2
Let D(s) and N(s) be p x p and q x p polynomial matrices and let D(s) be non
singular. Then there exist unique q x p polynomial matrices Q(s) and R(s)
such that
and
Ths theorem is du
of Theorem G-12 is c
compute Q(s) and R(
available in Reference
In the following, w
Consider p x p polyn<
A is a p x p constant m
1, i = 1, 2, ... , p. We
i = 1, 2, ... , p
Define
Proof
Consider the rational matrix (;(s) = N(s) D - I (s2. This rational matrix is not
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)
is a polynomial, then we can write G(s) as
G(s) = N(s)D-I(s) = Gsp(s) +Q(s)
(G-59)
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
N(s) = Q(s)D(s)
with
R(s)
=:'
Gsp(s)l)(s)
or
+ R(s)
Corollary G -12
and
This corollary can
Gsp(s) =R(s)D-1(s)
Q,(s) = N"s" -
Since R(s) is equal to the difference of two polynomial matrices [R(s) = N(s)
Q(s)D(s)], it must be a polynomial matrix.
To show uniqueness, suppose there are other Q(s) and R(s) such that
N(s) = Q(s)D(s) +R(s) = Q(s)D(s) + R(s)
and
+ (NIlA
(G-60)
G-6
Coprime Frr
Consider a q x p prop,
!S c:3.11eC ~J rfqh.t-('D.TJri'
A- 1(s)8(s) a leji-coprir.
will also be called an iI
many fractions, sorne e
related by the followin
Theorem G-12'
.Consider a q x p prop
C(s) = N(s)D - I (s). Th
.. p x p nonsingular poly
N(s)
i = 1, 2, ... ,q
Theorem G -1 3
- - - ---_._-----------
605
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
compute Q(s) and R(s). Ir O(s) is colurnn reduced, difTerent procedures are
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
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
o..
Define
rational matrix is not
decomposed as 9is) =
nal function and q;j(s)
(G-59 )
and
Corollary G-12
Let O(s) = sl- A, and let N(s) be an arbitrary polynomial matrix. Then there
exist unique polynornial matrices Qr(s) and Q(s) such that
is a polynomial rnatrix.
and
III
)D-1(s)
+ NoI
+IN o
(G-60)
G-6
O) implies
side is a strictiy
= Q(s) and R(s) = R(s).
~orem G-10.
Q.E.D.
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) =
A -1(s)8(s) a lefl-coprimefraclion if A(s) and 8(s) are ieft coprime. Either one
will also be called an irreduciblefraclion. Given a (;(s), it is possible to obtain
many fractions, sorne are irreducible and sorne are noto However, they are all
related by the following theorem.
. If A(s) is nonsingular,
R(s) such that
Theorem G-13
).
~ft-hand
. and
D(s) = D(s)T(s)
606
Proof 6
Let 0- l(S) = Adj O(s)/det O(s), where Adj stands for the adjoint and det stands
for ~he ~eterminant ofa matrix. Then, N(s)O- l(S) = (s)D- 1 (s) and O(S)O-I(S)
= O(s)O- l(S) = 1 imply
Then we have
and
(;(s) = N(s)D-
N(s) Adj O(s) det D(s) = (s) Adj D(s) det O(s)
= (s)O- J
O(s) Adj O(s) det D(s) = D(s) Adj D(s) det O(s)
Let R(s) be a gcrd of N(s)and D(s), and let R(s) be a gcrd of (s) and D(s). Then,
it is clear that R(s) Adj D(s) det D(s) is a gcrd of the two left-hand-side poly
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
matrix D(s) Adj D(s) det D(s) is nonsingular. lndeed, beca use of det D(s)D- l(S)
=det [D(s) Adj O(s)/det D(s)] = det [D(s) Adj D(s)]/(det D(s))P = 1, we have
det [D(s) Adj D(s) det D(s)] = [det D(s) det D(s)]P =/=0. Hence Corollary G-7
implies the existence of a unimodular matrix W(s) such that
R(s) Adj D(s) det D(s) = W(s)R(s) Adj D(s) det D(s)
which implies
R(s)D- l(S) = W(s)R(s)D - l(S)
or
(G-62)
and N(s)D-l(S)=V Z1
rght coprime fraction
of R(s) can be avoided
sarily column reduced
In this section, we
fraction from a right
procedure is similar to
Consider the q x p
N(s)D-I(s), where A(s)
and p x p polynomial
written as B(s)D(s) = A
(G-63)
If we consider polYI
!R(s), then Equation (G
sequently, aH 1 x (p +
nomials) satisfying
71l
607
Then we have
G(s) = N(S)O-I(S) = N(s)R(s)[D(s)R(s)] -1 = N(s)R(s)R -I(S)O-I(S)
= N(s)D-I(s)
t 0(.1')
t 0(.1')
(G-61 )
>fN(s)andD(s). Then,
left-hand-side poly
gcrd ofthe right-hand
: equalities in (G-61),
I different gcrds of N(s)
m that the polynomial
;ause of det O(s)O -I(S)
let D(sP = 1, we have
Hence Corollary G-7
that
'10
det D(s)
(G-62)
(G-63)
Q.E.D.
'ns of a proper rational
: irreducible fraction is
~enerated from a single
t irreducible, we may
I/(s)and D(s). We then
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
nomial matrix and is rather complicated. Ir in the process of generating (G-33),
we also compute Ves) in (G-35), then we have
O(S)J = Vis) [R(S)] = [V II(S)J R(s)
[ N(s)
O
V 21(S)
and N(S)0-I(S)=V 21 (s)V?(s). Since Ves) is unimodular, V21(s)V1II(s) is a
rght coprime fraction (why?). By this method, the computation of the inverse
of R(s) can be avoided. Note that V 11(.1') obtained in this process is not neces
sarily column reduced.
In this section, we shall introduce a method of obtaining a left-coprime
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
written as 8(s)0(s) = A(s)N(s) or
[-B(s)
A(S)][D(S)] =0
N(s)
(G-64)
;~(s) l_ N(sJ =
is a linear space over IR (s), denoted as (\1, !R(s. It is a subspace of (IR [J +q(s), !R(s.
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
dimensional null space (\1, lR(s, any set of q linearly independent vectors in W
qualifies as a basis (Theorem 2-1). In our study we are however interested in
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 polynmial basis, or a basis
in (W p, IR [s]V if every vector in \1 p can be expressed as a unique combination
71l is in rael a rrec module over lhe polynomial ring !REs] wilh dimension q. 5ee roolnole 4 arlhis
ehapter and Rererenee 534.
608
D(s)=D o +DIs +
and
N(s) = No +NIs +
A(s)=A o +A 1 s +
B(s) = Bo + BIs +
+Ddsd
+Ndsd
(G-65)
+Amsm
+Bmsm
0- - -~ -.-. ~ -i>~~ ~
[-80 A o : -B I A I :---: -Bm A m]
O O'" O
O O O
-i>~ - -.-. ~-
No ... N d I N d O ... O =0
..
..
O
Do
DI "Dd
NI ..... Nd
O ... No
- - -.- - - - - - - - - - - - - - - - - - - -.
O
O
(k
+ 1)th block {
(G-66)
'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),
there are q rows of unknown [ - B o A o ... - Bm A m ]. In order to have q
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,
BThe discussion'is briefand Ihere~der needs nol be concerned because the subsequenL development
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.
609
(G-65)
:onstant matrices. By
:ient of Si to zero yield,
O' O
O O
O"
O"
O
O
~-----------
-1
-1
Do O
No O
O
O
=0
DI Do
NI ..... Nd
(G-66)
the subsequenldevi:lopmenl
see Rererence S95, where
. D(s) 10 be nonsingularafld
the first q linearly dependent rows in the resultant in (G-66). In order to do so,
we define
1st block { Do D] ... Do
No NI'" No
O
O
O' .. O
O' .. O
~number of
dependent rows)
} f0
+ l)th block
0- -0- - -o ~ ~ -~ -- -i>~
{O O ...
} f 1
~ -.-.-. -D~J
NI' ..... No } fk
No
-'-0
The rows formed from Di will be called D rows; those from Ni, N rows. The
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.
Lemma G-1
Ir G(s) = N(s)D- 1 (s) is proper, aH D rows in Sb k =0, 1, ... ,are linearly inde
!l
pendent of their previous rows.
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
ample, a linearly dependent row will appear in the D rows of Sk formed for
[:2
~I:::~s
~Tl=[_:
S-:l]
even though we have bcN(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
l?---~!
Do DO-l ...
No N O - 1
S" ==! !Di
A
Do
No
O
O
0l
. ..
..
O
o
--------------------------
.
..
.
- - - -'-'-'- -
lf'\',C
TI""''l:
~ ~
~_1_
~_o_ - ~ '- '- - ?J
..
.
-
..
then the statement is again not valid even if N(s) D - I(S) is proper. For example,
consider
[1
1][~
.:Tl
610
Let
matrices
(G-70)
n=(q-ro)+(q-rl)+'"
of linearly independent N
once the number ceases to
are added, the number of
Define
+(cl-r v -)
(G-71 )
rO+(rl-rO)+(rz-r}+'"
+(rv-r,._}=r,.=q
'-[--(~l?.Al-a-?-z -a~f\ :
-B o a~1 a~2 a~3:
"On a digital compllter complltation, th; algorithm ShOllld be replaced by a nllmerically stable
method:' SecAppendix A. (fthe fOW searching aigorithm is not empioyed, theresllltis generally
not in the polynomial echelon formo
611
matrices
"
A(s) =
(G-70)
nearly independent N
the number ceases to
idded, the number of
me
(G-71 )
L Ais
and
18 v =0
(G-72)
Bis i
(G-73)
are left coprime, and A(s) is in a canonical form called the polynomial echelon
formo
Proof
In order not to be overwhelmed by notations, we assume p = q = 3 and
-D
x
x
x
D
x
X
KS3=~
},o~o
},'
O
--
D
x
x
1
(G-74 )
},' 1
O
D
O
O },'
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
study the structure of A(s), we write (G-72) as
~3
=r,,=q
L
i=O
in Sk for k 2: v -1. It
my irreducible realiza
al to 1'0+1'1 + ... +r.,
lly q primary linearly
T if the corresponding
ows. For example, all
uy linearly dependent
)Iock of S" are not all
:st block have already
:ond block of S., there
larly, there are r2 -r1
v' ProCeeding in this
rly dependent rows in
B(s) =
i=O
rrching algorithm. in
as in (A-7), correspon
(G-75)
~80
.
I
.\~~
.A
S3
S2
Si
r
(f~
,
,,
O O
O O O ,
O , -8 2 aL ai2 O : -B 3 :~ O
,
O ct~
O ,,
aL a~2 O ,
--
..
(G-76)
612
This is in an echelon formo The column positions ofthe three :1) are determined
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
searching algorithm, then (G-76) has the foIlowing properties:
ce ,
Hence we have
{set of row degrees o
Since A(s) is row redl
q
Property 3 is the same as saying that aIl columns which are on the right-hand
side and occupy the same positions in each block as those columns with the
((: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
side of the sixth column, with element:))' of the second block. We note that
the aboye three properties may overlap. For example, the rightmost column
of (G-76) is a zero column following 1 alone 01' foIlowing 2 alone. Because of
these properties, A(s) becomes
a?l +alls
i - - - - i - -2 - - 3'
A(s) = [ :~:1_~~3~~~~3~~ _~~_:
a~l +a1ls+a~ls2
, - <5 - - -
a?2 +al2s
a32+ai2s+a~2s2
,. <5 - - - -1- - - - -2- -2- - -
3'
(G-77)
j;:;:
v,
and
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
and every column has only one pivot element. Because of Property 1, the
degree of a pivot element is larger than the degree of every right-hand-side
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
of a pivot element is larger than the degree of every other element in the same
column. A polynomial matrix with these properties is said to be in the row
polynomial echelonform 01' Popov form. 10 It is clear that a polynomial matrix
in the echeleon form is column reduced and row reduced.
We discuss now the row degrees of A(s). We note that in each block row of
S" there are q N rows, and the ith N row, i = 1, 2, ... , q, appears v + 1 times in
S". Define, for i = 1, 2, ... , q,
Vi
I:
(G-78)
They will be caBed the row indices of G(s) = N(s)D-l(s). For the example in
(G-74), we have VI =3,1.'2=3, and 1.'3=1. Clearlythe n defined in (G-71) is
. also given by
Consider
(G-79)
s
\OThe nu'~erical"malrices [-B o Ao ... -,-B. A .]and [A o Al ... A,.] oblained by
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
G(s) =
+l
7
s-
[
S2
-1
s2
- - - - ---------- ----
613
'::;"2'::;
V3 '::;
responding to primary
Vi
(G-80)
i = 1,2, ... , q.
Hence we have
{set of row degrees of A(s)} = {Vi' i = 1, 2, " . " , q} = {l', i = 1, 2, ... , q}
(G-81)
g 2 alone. Because of
=
i
Vi
(G-82)
where deg det stands for the degree of the deterrninant. With this background,
we are ready to show that A(s) and B(s) are left coprime. Suppose they are not,
then there exists a polynomial matrix Q(s) such that
deg det Q(s) > O
(G-77)
A(s) = Q(s)A(s)
B(s) = Q(s)B(s)
A -1(s)B(s) = A - 1(s)B(s)
and
(G-83)
(G-S4)
(G-78)
Ir the linearly dependent rows ofS. appear only in the N rows, then frorn (G-76)
we can see that the row degrees of B(s) are smaller than or equal to those of A(s).
Since A(s) is row reduced, the computed A -l{s)B(s) is proper. Now ir a linearly
dependent row of . shows up in a D row, then ~) element in ,G-(6) wili appear
in the colurnn of - B i . Now because of property 1, the row degree of B(s) will
be larger than the corresponding row degree of A(s). This violates the assump
tion that C(s) is proper. Hence all linearly dependent rows of Sv must appear
in the N rows.
Q.E.D.
Example 1
Consider"
(G-79)
1"
C(s)"~
s. +1 -"
S2
S2 -1
,~~
s- d
s-1
S2 -1""
S~IJ=[S-1
lJ"[S2S2
2
s- 1
O
s-1
J-I
614
The fraetion N(s)D - ~(s) is earried out by taking the least eommon denominator
of eaeh eolumn of G(s) as the eorresponding diagonal e\ement of D(s). Ths
fraetion happens to be right eoprime. In general, a fraetion obtained by this
process will not be coprime.
We form S2 and apply the row-searehing algorithm:
1
O
-1
O
1
2
1:
- -- - -6 -- ---i
:- -1- - - - - - - - - - - - - - ~ - - - - - - - - - - - - - - - - - -
O O O O: O 1
O O 1 O: O O
O O O O ;-1
O
0.5
--------------,---------O O O 1: O O -1
FtS 2 =
1:
J
0;--(---------------
O O O: O 1
0.5
O O 1 :-1 -1
O
O O O: O -2
-1
O O 1 O
O 1 O O
{)) O O O
2
O 0(1)0.
O: O
O: O
0:-1
1
O
O
1
O
O ~I~
[~.5
A(s) =[-0.5
-1
and G(5) =A- 1 (s)B(s
linearly independent
the polynomial eehel
Consider a proper
right eoprime if and
linearly independent
In the following \\
properties of (;(5) anc
Hon. First, we note
altered without affee
and (G-84) by ehoos
O -\
:T~
-1
1
O 2
O
1
1
O
O
O
O
O ::~ I:
O
1
O
O
O - - - - -(j) - - -( - -~ i --- -0
O
8(5) = _
Corollary G-14
--------------~1---6--0-----(6-0
::-=- -(~
Combining Theo'
eorollary reduces to I
-------
O -1
**
[~.5 =~.5:
and
O
O
O
:~D O
O ~~~D
-1
1
O
Lemma G-2
,X:
:I~
O -\
The m x m polynom
the fie\d of rationa\ fl
TO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
00
O
(i~
:I~
O
O
v,~;
O
O
[
.~
615
common denominator
:::lement of D(s). This
ction obtained by this
Note that the pivots are chosen for the convenience of hand calculation. Since
r2 = 2 = q, we have v = 2. The primary dependent rows of S2 are indicated by
the arrows shown. Corresponding to these primary dependent rows, we use
the formula in (A-U) to compute
-0.5 -1:-1
[1.5 -2.5:
-1 :-1 :-1
Note that the solution is in the echelon formo
B(s) = _
[~.5
Hence, we have
-2.
5 -J - -l--11 0J s-[0 -J
-1
2.5
[s-0.5
-1
-J
S2 =
s-1
s +1
s 2_[0.5S-0.5
.---_.
:~~_:.JJ'
and
A(s ) -
1
O
s+
:_.~.2_ -=- ))
and G(s) = A -l(s)B(s). Note that deg det A(s) is equal to the total number of
lnearly independent N rows in S2, A(s) and B(s) are left coprime, and A(s) is in
the polynomial echelon formo
lilI
1
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
lnearly 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 '.,:Vh is irnportal1t 1~:; th set : ",';, ; :::";. ':.
2, ... , q} rather than the individual Vi' We need the following lemma to estab
lsh 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
V.~
.0
[ T,
O
...
T1
Tj
To
T j - i Tj
...
To
...
Ti
: ] k+ 1 block 'Ows
T j
-E-
1,2, ....
(G-85)
616
Proof
IfT(s) is singular, there exists a 1 x m nonzero polynomia\ vector
cx(s) = CX +cxs + ... +CXkS k
o
such that
cx(s)T(s)=0
This equation implies
[CX o CX 1
cxk]Vk= O
Hence, if T(s) is singular, Vk does not have a full row rank for every k.
By reversing the aboye argument, we can show that if Vk does not have a
flll.l row rank for every k, then T(s) is singular. This completes the proof of the
lemma.
Q.E.D.
With this lemma, we are ready to establish the following theorem.
Theorem G -1 5
form a q-dimensional
G-14 are left coprime:
nomial basis of the nu1
rows of Sk are seard
[ - B(s) A(s)] are the
that 6r B(s) .::s;6,.A(s) aj
degrees of [ - B(s) A
Since the row degrees
degrees is unique (The<
minimal polynomial bao
A(s) is nol row reduced
have deg det (s) = de~
(G-86)
Sk =SkV d+k
(G-87)
As implied by (G-8
Hence we condude tr
In other words. if G(~
coprime and A(s) is ro
As discussed follo\!
k =0,1,2,. . .
(G-88)
Consequently, I' =', i =0,1,2, ... , where Fare defiried as in (G-67). Hence
i =0, 1, 2, ... , vare independent of N(s) and D(s) used. in the computation.
Now we shall show that the set {Vi, i = 1, 2, ... ,q} is uniquely determinable
from {r, i = 0, 1,2, ... , v} .We bserve Ihat /' gives the number !V rows with
I',
of
r _1
= O. Then
/ =0, 1, ... , v
.
(G-89)
Allsolutions y(s) of [A
[f D(s) and N(s) are rig.
is a 'polynomial basis o
the set is a minimal po
coluriln searching algo
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,
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.
.al vector
.k for every k.
: if V k does not have a
lpletes the proa f of the
Q.E.D.
ving theorem.
Jtation.
( G-86)
'heorem G-13). If we
pUes, rOl k = 0, 3., 2.} ... "
(G-87)
he degree of D(s) as in
full row rank for every
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
coprime and A(s) is row reduced, then the set of row degrees of A(s) is unique.
As discussed following Equation (G-64), all solutions x(s) of
X(s{~~;~J =0
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
rows of Sk are searched in order from top to bottom, the row degrees of
[ - B(s) A(s)] are the smallest possible. The properness of N(s)D - 1(.1') ensures
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
minima! po!ynomia! basis. We show that if A(s) and B(s) are left coprime but
A(s) is not row reduced, then [ - B(s) A(s)] is not a minimal basis. Indeed, we
have deg det A(s) = deg det A(s) and
Thus the row degrees of [ - B(s) (s)] are not minimum. This establishes that
[ - B(s) A(s)] is a minimal polynomial basis if and only ir A(s) and 8(.1') are !eft
coprime and A(s) is row reduceo. 'l'ile basis obtained in Theeem 0-14 has une
additional nice property; it is in the polynomial echelon form. This is a conse
quence of the row searching algorithm. In conc\usion, the solution of (G-66)
obtained by using the row searching algorithm is a polynomial basis, a minimal
basis and in a canonical formo
We disc ..ss briefly the dual case of Equation (G-64). Consider
(G-88)
~d
as in (G-67). Hence
:d in the computation.
uniq uely determinable
Iuniber of N rows with
(G-89)
(),.A(s)
[A(s)
B(s)] [ -
~~;J =
All solutions y(s) of [A(s) B(s)]y(s)= O forro a p-dimensional right null space.
If D(s) and N(s) are right coprime, the set of the p columns of [ - N'(s) D'(s)]'
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
column searching algorithm (dual to the row searching algorithm), then tlle
618
Aa
Al
Am _
Bo
B1
1
T k = Am
O
Bm
B",
O
I :
O
Aa
O
Bo
Am z
Bm
Bm
B",
O
Am
O
:O
: . .. : O
, .
,, :
Z
I
O1
O
r
lo
: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
Di and N(s) =
Nis ,
least integer such that Ji = p. Then D(s) =
solved from
G-3
,:
,.
L;=o
L;=o
G-4
D(
G-5
a. D,(s)= [
+1
(s-lXs +2)
b. D 2(s)=D I(s)
c. D 3 (s)= D,(s)
d. D 4 (s) = D,(s)
;",;,;:~,,
hre
tne Vau's
N2(~
N3(~
N 4 ()
l j-'~j\S.
Problems
G' Apply the row searching algorithm to Skin(G-20) to reduce the following rationl
functions to irreducible ones:
G-8
-------..- - - - - _ . _ -
PROBLEMS
N(s)
D(s)
N(s)
D(s)
619
5'+5 2 _5+2
= --;::--=--
2s'-s2+ s +1
2s' +25 2 - s -1
25 4 +5 2 -1
G-2
G-3
(s + 1)2
-s-1
S2
+s+
s-1
_S2
+s+1
1]
"1'
s+1
O
G-4
52 +2s
O(s)= [ 15 2 -5
G-5
N(s)=[s
1]
a. 01(S)= [
1I
5+3
3s-2
+1
(s-I)(s+2)
b.0 2 (s)=D\(s)
N 2 (s)=[s-1
c.O,(s)=O(s)
N,(s)=[s+1
d.0 4 (s)=D(s)
N 4 (s)=[s s]
N(s) =[s +2
s
s +1]
+ 1]
s - 1]
Are the pairs l&i\s), J\sjj, = 1, 2, .\ 4, ieil copnie ir" \s) = ;Dii(.),/ anu 1>(.)j = l~il':;i
in Problem G-5? Are they left coprime if A(s) = O(s) and B(s) = N;(s)? (The prime de
notes the transpose.)
G-o
s) be column reduced.
of (;(s) and the set. of
2s + l'
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. .'
:e the following ratiqnal
G-B
Is the M(s) in Problem G-7 row reduced? Ir not, transform it to a row-reduced one.
._~---_._~~----- ------~--------------~--
620
-------~----------~-
Find, by applying the row searching algorithm to Sk in (G-67), left coprime fractions
from the following three different right fractions:
G-9
G(s) =
[.1'3 +.1'2 +s + 1
~I:2
7
1
-s+l
J-'
=fS2 +.1'
-s +1
bs+1
J-'
and
Are the results the same? Which right fraction, if there is any, is right coprime?
(soIUtiOn:
s2~J-r~1
[~
G(S)=[S2~1
.1'2 + 1
=[.1'
G-10
----------;
s~lIs~l
~:~J
Consider
s2~lTl)
N(s)W'(s)~r
=(
V 11 (s) V ds)J
[ B(s)
A(s)
such that
11(.1) V12(S)J[
[V
B(s)
A(s) _ -
O(S)J=['J
N(s)
or
= N(s
Q(s)~Q(s
D(s)~DI;;:
Q
Q
and U H and U 22 are left coprime and Vil V 21 = ND-'. Using the dual ofTheorem G-13,
there exists a unimodular matrix M such that MV 21 = B and MV 22 = A.)
Silow that ji N(.I')lJ}-l(.I') = I\J{s)D (sj are [wo right-coprime "ractions, lhea he
matrix V(s)=D-'(s)O(s) is unimodular. (Hint: The equation XN+YO=I implies
XiJ- D +YDD-' O =(X +YD)V = 1 which implies U-' to be a polynomial matrix.)
:':;;-")2
G-~3
G-14
+ Nc(s)N(s) = F(s)
PROBLEMS
3 J-l
2
621
and N(s), show that for any F(s), there exist solutions D/s) and Nc(s) to meet the Diophan
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
D~(s) = F(s)U11(s)
+25
are solutions of the Diophantine equation. Finally, sho\\' that, for any p x p polynomial
matrix T(s), the matrices
Dc(s) = D~(s) +T(s)U} 1(S)
and
s right coprime?
Consider
N(s)=Q(s)O(s) +R(s)
-lJ
-2
with bcR(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)
:tions.
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
Np(s) = (s)[N po + Np1s- 1 + ... + Np"s-']
Q(s)~Q(S)Ohc~H(s)[Qo+ QIS-I + ... + Q,05-''J
D(S)~0/;;,iOlcls)H-l(S)~0IS-I+02S-2+ .,. +O.s-
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
Qk = N pk -
I Q,D
k- 1
1=0
( p, and p x p polynomial
where Ol~ O if / > JI. From Q(s), we can compute Q(s) = Q(s)DI;;' 1 and lhen compute
R(s) = N(s) - Q(s)O(s). See Reference S137.
G-';<'; COi1sic( G(SI = h. '(s)J8;(s) = r<J(s)Lv
ralional function solutions of
'(S).
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) (s)J be left coprime. 'If [ - 8(s) A(s)J is a basis, there exists a
.
po\ynomia\ malrix TM S1Ich that
[- B(s) (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 Os)
622
- - [O(S)J
=0
[-B(s) A(s)]
N(s)
Then use Theorem G-13 to show the existence of a unique polynomial matrix T(s) such
hat [ - B(s) A(s)] = T(s)[ - B(s) A(s)] and [- bis) a(s)] = t(s)[ - B(s) A(s)].]
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 ~
oi the acnomina
not a pole of
mOL
although it is a root o
Consider a proper
tion
Pales
and Zeras
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.
Definiton 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
Ir a proper rational transfer function is irreducible (that is, there is no non
trivial common factor between its numerator and denominator), then it is clear
that every root of the denominator of 9(S) is a pole of 9(S) and every root of the
numerator of g(05) is a zero of 9(5). Without this irreducibility assumption, "
root oi tile Cienorlnawi oi Y(S} ilay not O:; <l pole o' Y(s). t'ur el..dnipie, - 1 iS
not a pole of
A
+1
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
with sorne constant k. If D(s) is monic (the coefficient associated with the highest
power of s is equal to 1), then k = 1. We discuss in the following the implica
tions of poles and zeros.
Theorem H-1
Consider a system w
tion {A, b, e, e}. Ir
is not a pole of g(s), t
and the input u = e;r
for all
Theorem H-2
t ~O
(sI -A)-l(S
Proof
(s
ar
y(t)=eve AC
[Al -AJ
vf.O
=e(sI -
Since (Al - A)v = O, we must have r = el' f.O; otherwise, it would have violated
the aboye condition. This competes the proof of :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).
Ir y(t) = re A', then we have
y(s) = c(sl - A)-I X (O) = r(s - ,,1.)-1
(s~ _
Proof of Theorem H
forallt~O
What remains lo be shown is that the constant r =cv is different from zero.
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
particular, the vector v, we have
c
~A
or
det
or
AY
y(s) = g(A)(s - A:
or
y(t) =
g(A )e A'
625
not a pole of 9(S), then this is not possible; the only way to generate
output is to apply e;" at the input.
e}"
at the
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,
is not a pole of 9(S), then the output due to the initial state x(O) = - (A - Al) - 1 b
and the input u = e}" is equal to y(t) = g(J,)eJ.' for t ~ O.
I:ii
To prave this theorem, we need the following identity:
(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,)
et v be an eigenvector
'erified (Problem 2-32)
eigenvalue (s - A) - 1.
( H-2)
Y(s)
or
= g(A)(S -
y(t) = g(A)eJ.'
A)-1
for t ~ O
-A)
,f A and, consequently,
Q.E.D.
e e;" can be generated
of any input. Ir A is
Q.E.D.
626
Cansider a system with proper transfer function 9(s) and an irreducible realiza
tian {A, b, e, e}. Ir}, is a zero of 9(s), then the response at the output due to the
initial state x(O) = - (A - ll) - 1b and the input u(t) = eA' is identically zero. 111
We shall now extend the concepts of poles and zeros to the multivariable
case. Consider a q x p proper rational matrix (;(s) with the following coprime
fraction
(;(s) = O-I(s)N(s) = N,.(s)O,:- I(S)
y=Cx +Eu
only if there exists an initial state X o such that the zero-input response al the
output of the system is egual to
1
'L.,...}
~f'
~_'.1 . "~'.
G(s) = D- 1 (s)N(s).
""\'J/
y(t) = reAl
<;;;onsider a q x p pr
Example 1
The proof of this theorem is identical to the one of Theorem H-l and will
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
ofaIi Ilinors 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
of rational fundions. By this, we mean that if q'~ p, then p(;(s) = q; if q> p,
G 1 (s)
s
s +2
O
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
p x 1 rational vector P(s) such that
an irreducible realiza
the output due to the
is identically zero. 11
M(s)(;(s) =0
or
(;(s)P(s) =0
( H-3)
s to the multivariable
the fol1owing coprime
M(s)(;(s) =0
or
(;(s)P(s) = O
where M(s) and P(s) are polynomial vectors. Since y(s) = G(s)u(s), we have
M(s)y(s) = M(s)G(s)u(s) =0
q x p, and p x p poly
for al1 possible inputs. This implies that the number of effective outputs is
smal1er than q. Similarly, if G(s)P(s) =0, and if we connect a precompensator
P(s) to the system, then for all possible inputs u(s) we have
G(s)P(s)u(s) =0
: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.
Consider (;(s) with the fol1owing coprime fractions:
(;(s) = D-1(S)NtCs) = N,.(s)D,:-1(S)
~m
to characterize the
(H-4)
If (;(s) is of full rank in the field of rational functions, so are NtCs) and Nr(s)..
This implies that for almost every A in ic, the q x p complex matrix NtCA) has
a rank equal to min (p, q) in the field of complex numbers.
Definition H-3
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),
in ic, the field of complex numbers.
~
Example 1
Consider the left coprime fraction
s
s +2
G 1 (s)= O
~
+11
s+2 =[s+2
+ 1
0J-1[S
S2
O
O
s+1
This N(s) has rank 2 ror eYery s in iC; hence G (s) has no transmission zero. I
'If (;(s) does no! have a full rank, lhen pN(A) < min(p, q) for every A in e. In olher words. every
i.in e is a lransmission zero of (;(s). This is a degenera le case and will not be considered.
628
Example 2
Case 1: q > P
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).
~
From ExampIe 2, we see that (;(s) may not be weIl defined at its transmission
zeras. Hence we cannot use p(;(},) < min(p, q) to define its transmission zera.
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
N(s) = U(s)Nh)V(s)
(H-5)
Hence the transmission zeros of (;(s) can be defined by using N(s) or Nr(s). We
note that if N(s) and Nr(s) are square, the transmission zeros of (;(s) are the roots
of det N(s) or det Nr(s), where det stands for the determinant.
Fram the aboye two examples, we see that the transmission zeros of (;(s)
exhibits some phenomena which do not exist.in the scalar case. A transmission
zera may appear as a pole of the same (;(s). Even though elements of (;(s)
have zeras, (;(s) may not have any transmission zera. The transmission zeros
of a square (;(s) may be different fram the zeras of det G(s). In spite of these
differences, the transmission praperties of the transmission zeras of (;(s) are
quite analogous to the scalar case. We shall establish these in the foIlowing.
and consequently,
y(r
This property is si
H-2. Note that if 1,
pN(}.)=p and N(J,) k ,
If, in addition, ), is ne
is also a nonzero vectl
Theorem H-4
is equal to
y(t) = G(A)ke At
fort;::o:O
(H-6)
I
The praof of this theorem is identical to the one of Theorem ,H-2 and will
not be repeated.
The substitution of (;(..1.) = D- 1 (A)N(A) into (H-6) yields
y(t) = D-I(A)N(A)ke At
(H-7)
"S'r
629
Case 1: q"2:.p
y(t) = D- l(X)N(),)k/'/
+O
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
N(X"- = O and
forall t"2:.0
( H-6)
I
( H-9)
Since this equation holds for every X in iC, it cannot be used to characterize the
transmission zero of C(s) as in the case of q "2:.p. For this reason, we must con
sider this case separately from the case q "2:. p.
Ir q < p, then X is a transmission zero of C(s) if and only if there exists a
nonzero 1 x q constant vector h such that
hN(A) =0
If, in addition, X is not a pole of G(s), the 1 x q vector f defined by f
is also a nonzero vector. Consider
( H-8)
f~l{t) == fG{.J..)~~,(!A(
hN(X)ke
-::-:
( H-10)
= hD{..1.)
A
'
( H-11 )
which is, because of (H-lO), identically zero for any k. This property, namely
fy(t) =0 for t "2:.0, exists only if X is a transmission zero of G(s). Hence this can
be used to characterize the transmission zero of G(s) for the case q < p.
ds
( H-7)
e following, we discuss
denote the rank.
630
for all t? O
x(O) =
-1
Sl-A
-c
( H-12)
0J[SI-A
Iq
-C
01
Iiil
This corollary is just a restatement of (H-8) and (H-IO). We see that this
corollary is quite similar to the one for the scalar case. Because of this trans
mission-blocking property, the zeros defined in Definition H-3 are called the
transmission zeros.
The transmission zero can also be defined by using dynamical equations.
Consider a system with a coprime fraction G(s) = D1-l(S)N1(s) and an irreducible
dynamical equation {A, B, C, E}. Consider the system matrix
S(s)~ [
BJ [Sl-A
E O
=[
B
J
C(sl-A)-lB+E
SI-A B
[SI-A
O
C(s) =
O
The combination of (
D-l(S)Nt(s)
or
we have
Sl-A
f[C(Ai
Define f 1 =fC(AI - A;
B =p(sI-A.)+pG(s)
G(s)
P [ -C
and
where p denotes the rank in the field of rational functions.
rank, that is, pG(s) = min(p, q), then
sl- 1\
p [ -C
:J
= n
+min(p, q)
Ir C(s) is of full
(H-14 )
Since [f 1 f] is a nom
the link between Coro
In the following, ti
will be removed. The
be skipped in the first
Theorem H-5
Theorem H-6
Consider a q x p proper rational matrix G(s) with full rank and with an irredu
cible realization {A, B, C, E}. Ir Ais not a pole of G(s), theri isa transmission
Consider a q x p pro::
dimensional irreducib
d if A is a transmission
{here k is an arbitrary
such that the output
- 1Bk has the property
631
pS(I,)=p
AI-A
[
: ] < n +min(p, q)
-C
(H-15 )
This theorem foilows directly from Definition H-3 and Equation (H-13).
In establishing (H-13), the nonsingularity of sI - A is implicitly assumed.
Hence the proof is applicable only if }, is not a pole of G(s). Before removing
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
nonzero k such that the output due to u(t) = ke At and the initial condition
x(O)=-(A-AI)-IBk
are identically zero.
Or
d q x p matrices, then
lunctions.
[A~-CA
( H-17)
:1 -:(O)]
=0
or
fG(A) =0
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)
ons.
+ Ek = O
Because of
lB
]
I,)-IB +E
(s)
(H-16)
(A-AI)x(O)+Bk=O
(H-13)
lf (;(5) is of fuil
fl(AI - A) = fC
flB +fE. =0
and
which can be combined as
[fl
(H-14)
ex numbers, say s == A,
1st be computed in C.
f][AI-A
-C
8J=0
E
( H-18)
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.
In the fo\!owing~ the restriction that A is not a pole of G(s) in Theorem H-S
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
Consider a q x p proper rational matrix G(s) with a fu\! rank and with an n
dimensional irreducible realizaton {A, B, C, E}. Then A is a transmission
632
Proof
:J
AI-A
-c
SI-A
p[
< n +min(p, q)
Let E =
(;( co).
We factor (;(s) as
(H-19)
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}
such that, similar to (6-133) and (6-130),
Co(sI - A o)- 1 = D -1(s)L(s)
N(s) = L(s)8 0
and
-C
(H-20)
(H-21)
Definition H-4
Consider a q x p prope
divisor (gcd) of the nur
are called the blocking :
Example 3
Consider
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
~(S)
[L(s)
Y(S)][SI - A o
D(s)
- Co
~(S)J =[I
Ves)
~(S)
[L(s)
V(s)J[SI - Ao
D(s)
- Co
:OJ=[1
fl.
=Lo
~(s)Bo +Y(S)EJ
L(s)B o + D(s)E
~{.(S)~G -l-\Y(s)}f,~
N(s)
In the last step, we have used N(s) = D(s)E + (s) = D(s)E + L(s)8 0 in (H-19)
and (H-21). Since the leftmost matrix in (H-23) is unimodular, we have, for
every s in e,
P [SI -Ao
-Co
BOJ= [In
E
p O
X(s)Bo +Y(S)EJ
N(s)
-n +pN(s)
e,
Since {A, B,
E} and {Ao, Bo, Co, E} are both irreducible realizationsof
C(s), they are sti-ictly system equivalent. Hence we have, for every s in e,
'This proof \vas provided by Professor L. S. Chang of China University of Sciences and Technology,
HoFei.
--~--
sI -A
p [ -C
l(p, q)
BJ
= p [SI -Aa
E
-Ca
BoJ
E
633
+pN(s)
(H-24 )
Q.E.D.
=11
(H-19)
Definition H-4
Consider a q x p proper rational matrix G(s). Let {3(s) be the greatest common
divisor (gcd) of the numerators of all elements of G(s). Then the roots of {3(s)
are called the blocking zeros of G(s).
Example 3
Consider
s +1
s +2
(s +2)(s
S2
(H-22)
+ 1)
+2s +2
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
- 2 (Problem H-S).
l!iI
)8 0 +Y(S)E]
)8 0
+ D(s)E
+- Y(s)ll.!'.l
)lB\o
N(s)
From this example, we see that the definition of transmission zero and that
of blocking zero are not equivalent. lf every element of G(s) is irreducible, a
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
zero of (;(s) will appear in every element of N,(s); hence every blocking zero is l
transmission zero. The converse is, of course, not necessarily true.
If G(s) is a scalar transfer function, then there is no difference between
~.ransmjsslon. 7,f;"(OS .3.n.et blor:k~.ng ~~r~s.
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
component of the input vector u(s). Let
i= 1,2, ... ,p
EJ = n +p N(s)
educible realizations of
ve, rOl' every s in iC,
( H-25)
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
is blocked from y(t).
.
Theorem H- 7 .
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
634
a mode at any output terminal if and only if every root of <jJ(s) is a blocking zero
of G(s).
Problems
H-1
H-2
Proof
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
state response of the system is given by
.
y(s) =
H-3
L gij(s)u}s)
( H-26)
j= 1
Let f3(s) be the gcd of the numerators of all gij(s). Then we can write 9ij(S) as
nij(s)
%(s) = f3(s) dij(s)
A
H-4
H-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
Yk(S) = f3(s) nk!(s)
<jJ(s) dkl(S)
H-7
and s - IX will appear as a mode of Yk(t). This completes the proof ofthis theorem
Q.E.D.
Le[ ;(s) = k/(s - ,t). H), is a lransmission zero o G(s), lflen e"; Wii! no apr ea ,'
as a mode at the output vector only for a certain k. Ir A. is a blocking zero of
G(s), then eA! will not appear at the output vector for any k. These concepts of
zeros will be used in the study of tracking control problems in Chapter 9.
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
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 presehtatin of this
appendix follows closely References S78 and S94.
Show thal
A-
rank [
zero-input response of
)f cP(s). Let Yi(S) be the
:gj(s)). Then the zero
( H-26)
635
Problems
H-1
H-2
y=[2
lJx
Find an initial state so that its zero-input response is y(c) = 5e- 1 for aH t zO.
H-3 Consider the equation in Problem H-2. Find the initial state so that the response
y(t) due to u(t) = e31 and the initia\ state is of the form e 31 for t zO.
H-4
H-S
What are the poles and transmission zeros of the transfer matrices:
.1'-1
G1(s)=
O
A
G 3 (s) =
A
.1'+10
2(s-W
.1'+1
.1'+2
(s +2)(s+l~
.1'+1
.1'2 +2.1' +2
(s - 1)2
.1'-1
.1'2
~ 1rT::~
OJ- 1[1
.1'-1
O
.1'.1'2
s -IJ
s +1
lJ
H-G Let N(s)D - 1(.1') = (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 (s) are the same.
H-7
Show that
rank [A -Al
-C
BJ
=rank
E
[T(A +BK-AI)r
-(C+EK)r l
under any state feedback and any equivalence transformation? Can you arrive at the same
conclusion fram the facts that a state feedback does not affect the numerator matrix of a
transfer-function matrix and that an eq uivalence transformation does not affect a transfer
function matrix?
H-S
What are the blocking zeras of the transfer matrices in Problem H-5?
n. Chang,
S. S.
McGraw-Hill.
"c
C. T..
Automatic Col1t
t3, Chen.
References
14.
--o
"Linear il
vol. 15, pp. 127
15,
--o
16.
--o
17.
--o
"Represel
Tra/1s. Automat
"Stabilit y
vol. 56. pp. 821
"A note (
AC-13, pp. 597
18. - - , "Design ,
vol. 57. pp. 46
19. - - , "Design o
Preprints 1970.
There are two lists of references. The first one was used in the preparation of
the original edition and the second one, the present edition. The lists are not
necessarily exhaustive nor do they indicate the original sources.
1. Anderson, B. D. O., and D. G. Luenberger, "Design of multivariable
feedback systems," hoc. lEE (London), vol. 114, 1967, pp. 295-399, 1967.
2. - - , R. W. Newcomb, R. E. Kalman, and D. C. Youla, "Equivalence of
linear time-invariant dynamical systems," 1. Franklin Inst., vol. 281,
4. Bass, R. W .. and I. Gura, "High order system dcsign via state-spacc con
siderations," Preprints 1965 JACC, pp. 311-318.
6. Brand, L., "The companion matrix and its properties," Am. Math
Monthly, vol. 71, pp. 629-634,1964; vol. 75, pp. 146-152,1968.
7 Brockett, R. W., "Poles, zeros, and feedback: State space interpretation,"
IEEE Trans. Automatic Control, vol. AC-lO, pp. 129-135, 1965.
9. Buck, R.
10. Bucy, R. S., "Canonical forms lor multivariable systems," IEEE Trans.
Automatic Control, vol. AC-lJ, pp. 567-569,1.968'
11. Chang, A., "Ari algebraic characterization of controllability:' JEEE
Ihms. Automatic Control, vol. AC 10, pp. 112-113, 1965.
636
21. - - , and - - ,
IEEE Tiam. Au
22. -.-, - - , and
and observabilil
Corztrol, vol. AC
23. Chi, H. H., an,
simulation," Pn
24. Coddington, E.
Equatiol1s. Ne\
25. DeBon, F., and
with inaccessib
26. Dertouzos, M. 1
block-diagram ~
1969.
27. DeRusso, P. M.
New York: WiJ
28. Desoer, C. A., '
Ti(ll1S. Circuit T
29. - , EECS 22:
1968.
30. - - , and C. 1
systems," IEEE
31.
--o
1969.
and E. S.
REFERENCES
637
New York:
30. _."-.,
ontrollability," IEEE
1965.
638
REFERENCES
canonical (phasl
AC-9, pp. 312-3
35. Duffin, R. J., and D. Hazony, "The degree of a rational matrix function,"
1. SIAM, vol. 11, pp. 645-658,1963.
lems," y,ans. AS
55. _._, "Gn '[he !
Congl'. Autom. e
56. - - , "Contribut
vol. 5, pp. 102-1
57. _._, "Canoniea
Acad, Sci. U.S., v
67. - - , and T. S. En
N ASA, Washingt,
50. Jacob, J. P., and E. Polak, "On the inverse ofthe operator (-) =A(J +8('),"
Am. Math. Monthly, vol. 73, pp. 388-390, 1966.
68. - - , P. L. Falb, ;
New York: MeGI.
69. ---, Y. C. Ho, al
systems," Contrib .
70. 'Kapian, W.,Opel
Addison-Westey,
49. Hsu, C. H., and C. T. Chen, "A proof of stability ofmultivariable feedback
systems," Proc: IEEE, vol. 56, pp. 2061-2062, 1968.
_.
__
~_L.'_.
__
~ _ .
_. __
~ .
_ _.,---,--._.
._.
- ' - - - - _ ~
~_~~~
REFERENCES
__
639
of a nonanticipative
. 15, pp. 754-763,1967.
3.
52. Joseph, P. D., and 1. T. Tou, "On linear control theory," AlEE 1Jans.
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
lems," Trans. AS ME, ser. O, vol. 82, pp. 35-45, 1960.
design ofmultivariable
)1. AC-12, pp. 651-659,
55. - - , "On the general theory of control systems," Proc. First Intern.
Congr. Autom. Control, Butterworth, London, pp. 481-493, 1960.
56. - - , "Contribution to the theory of optimal control," Bol. Soco Mat. Mex.,
vol. 5, pp. 102-119, 1960.
'd Mathematics.
New
58. - - , "On the stability of linear time-varying systems," 1RE 1J'ans. Circuit
Theory, vol. CT-9, pp. 420-423, 1962.
New York:
1 and 2.
1
mu ltivariable control
53.
t in muti-input con
; Control, vol. AC-13,
1 of linear
66. - - , and R. S. Bucy, "New results in linear filtering and prediction theory,"
Trans. ASME, ser. O, vol. 83, pp. 95-108,1961.
state variable
ton Conf, pp. 449-459,
. ~ _ . _ , ~
, IRETrans. Automatic
multivariable feedback
erator(')=A(') +B('),"
the transformation to
Reading, Mass,:
71. Kreindler, A., and P. E. Sarachik, "On the concepts of conlrollability and
640
REFERENCES
San Francisco:
New York:
the 'second me
pt. 4, pp. 694-
91. Polak, E., "An
system to state
577-579,1966.
92. Pontryagin, L.
Addison-Wesle
93. Rekasisu, Z. V.
variable feedba
94. Sandberg,1. W.
"On the
tions," Bell S YSi
96. Schwartz, L., Ti
95.
--o
78. Luenberger, D. G., "Observing the state of a linear system," IEEE 7Ians.
Military Electronics, vol. MIL-8, pp. 74-80, 1964.
Hill, 1965.
98. Silverman, L. 1\
Ph. D. dissertati
99. - - , "Transfo
variable) form,"
1966.
",
using
"":',
REFERENCES
approach to network
'orks.
San Francisco:
New York:
form realization of a
, vol. AC-14, pp. 66-69,
Lll
641
the 'second method' of Lyapunov," Proc. Cambridge Phil. Soc., vol. 58,
pt. 4, pp. 694-720, 1962.
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.
92. Pontryagin, L. S., Ordinary Differential Equations.
Addison-Wesley, 1962.
Reading, Mass.:
642
REFERENCES
of interconnec
vol. AC-28, pp
S9. Armstrong, E.
control. New
115. - - , and Plinio Tissi, "n-port synthesis via reactance extraction-part I,"
1966 IEEE Intern. Conu. Record, vol. 14, pt. 7, pp. 183-208.
116. Zadeh, L. A., and C. A. Desoer, Linear System Theory.
McGraw-Hill, 1963.
New York:
Supplementary References
SI. Anderson, B. D. O., "Internal and external stability of linear time
varying systems," SIAM 1. Control and Optimization, vol. 20, pp. 408-413,
1982.
S2. - - , and M. R. Gevers, "On multivariable pole-zero cancelIations and
the stability of feedback systems," IEEE Trans. Circuits and Syslems,
vol. CAS-28, pp. 830-833, 1981.
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,
pp. 551-556,1976.
S4. - - , and J. B. Moore, Optimal Filtering.
Prentice-HalI, 1979.
matic Control, .
unpublished ne
of rnultivariabl,
Atherton, Ed.),
S22. Birkoff, G. and
York: Macmill
S23. Bitmead, R. R.,
common diviso
Trans. Automat
S24. Bongiorno, J. J.
system," I nI. J.
pp. 183-190, 19
S25. - - , and - - o
back systems in
Contl'ol, vol. AC
S8. Araki, M., and M. Saeki, "A quantitative condition for the well-posedness
SUPPLEMENTARY REFERENCES
643
83-208.
Theory.
New York:
ne-varying nonlinear
ltomatic Control, vol.
'1,
N ew York:
n of linear systems,"
644
REFERENCES
826. Brasch, F. M., le, and 1. B. Pearson, "Pole assignment using dynamic
compensator," IEEE 1/ans. Automatic Control, vol. AC-15 (1), pp. 34-43,
1970.
827. Brockett, R. W., Finite-dimensional Linear Systems. New York: Wiley,
1970.
828. Bruni, c., A. Isidori, and A. Ruberts, "A method of realization based on
the moments of the impulse response matrix," J EEE Ttans. Automatic
Control, vol. AC-14, pp. 203-204, 1969.
829. Brunovsky, P., "A classification of linear control1able systems," Kyber
netika, vol. 3, pp. 173-188, 1970.
830. Budin, M. A., "Minimal realization of discrete systems from input
output observations," IEEE TIans. Automatic Control, vol. AC-16, pp.
395-401,1971.
831. Byrnes, C. 1., and P. K. Stevens, "Pole placement by static and dynamic
output feedback," Proc. 21st IEEE Con! Decision and Control, pp. 130
133, 1982.
832. Cal1ier, F. M. and C. A. Desoer, "An algebra of transfer functions for
distributed linear time-invariant systems" IEEE Trans. Crcuits and
Systems, vol. CAS-25, pp. 651-662, 1978. lts simplification and clarifica
tion, vol. CAS-27, pp. 320-323, 1980.
833. '--, and - - , "Stabilization, tracking and disturbance rejection in
multivariable convolution systems," 1978 IEEE Con! n jeCisi~n and
Control, San Diego, Calif., 1979.
834. - - , and - - , Multivariable Feedback Systems. New York: Springer
Verlag, 1982.
842. - - , "Minim;
puters. vol. C-2
S43. - - "An alg
Ins1., vol. 297,
844. - - , "lnertia
Appl., vol. 49, ~
845. - - , "Synthes
vol. 32, pp. 112
546. - - , Analysis G
Pond Woods, 1
547. - - , One-dime
1979.
848. - - , "Identifi(
Prac. 1982 IF A
849. - - , "A contril
systems," Pro{;.
850. - - , "Feedbac
Report, SUNY
851. - - , and C. H.
systems," Prepr
852. - - , and Y. S. f
from finite inp
AC-24, pp. 748
853. - - , and D. P
IEEE TI'ans. Au
854. - - , and S.
5ystem ami Con
855. Cheng, L., and j
variable linear 1
837. Chang, T. S., and C. T Chen, "On the Routh-Hurwitz criterion," IEEE
Trans. Automatic Control, vol. AC-19, pp. 250-251,1974.
857. Cheng, V. H. l
transfer functior
IEEE Trans. Au.
838. Chen, C. T., I ntroduction to Linear System Theory. New York: Halt,
Rinehart and Winston, 1970.
839. -.- , "Irreducibility af dynamical equation realizations of sets of dif
ferential equations," IEEE Trans. Automatic Control, vol. AC-15, p. 131,
1970. :
S40, -.- ., "A new look at transfer-function design," Proc. IEEE, vol. 59, pp.
1580--:1585, Nov. 1971.
S56. - - , and -
Trans. Automati,
SUPPLEMENTARY REFERENCES
tS.
)f realization based on
EEE 71ans. Automatie
645
--o
S46. - - , Analysis amI Synthesis 01linear control systems. Stony Brook, N.Y.:
Pond Woods, 1978.
,turbance rejection in
Con! dn Deeision and
646
REFERENCES
870.
~-,
576. - - , and M.
stable plant, ~
AC-26, pp. 4O~
577. - - , R. W. L
design: The fra
1EEE 7)ans. A
S78. - - , and J. D
and their dyn;
vol. CAS-21 (1
S79. - - , and M..
New York: Ac
580. - , and Y..
compensator,"
1976.
881. - - , and Y. 1
problem: A sel
vol 16 (C T. L,
882. Oongarra, J. J.,
U ser's Cuide. P
883. Emami-NaeinL
variable robust
884. - - , and P. V
transfer functio
55, 1982.
885. Emre, E., ''rhe
dynamic feedb~
620,1980.
886. - - , "Pole-zer
1EEE Con! De<
l
888. - - , and -
solutions, "SI A
873. Oesoer, C. A., Notesfor a Second Course on Linear Systems. Ne:w York:
Van Nostrand, 1970.
892. Fairman, F. W.
SUPPLEMENTARY REFERENCES
lVariant multivariable
,," IEEE u-ans. Auto
2; its correction, vol.
f linear multivariable
,," IEEE Trans. Auto
:oblem for linear time
utomatic Control, vol.
2, p. 283, 1977.
IS and 'squared-down'
28, pp. 221-222,1983.
leral servomechanism
11. 11 (5), pp. 461-471,
lvariantmultivariable
:dback," IEEE Trans.
i; its correction, vol.
:ransmission zeros o
pp. 643-658, 1974.
multivariable systems
Control, vol. AC-20,
.tion of transmission
dback," IEEE Trans.
647
)f stability coriditions
EE 1Ians.. Automatic
648
REFERENCES
order observers," Int. 1. Systems Sc., vol. 11, pp. 1083-1094, 1980.
S93. Fernando, K. V.~ and H. Nicholson, "Singular perturbatonal model
reduction of balanced systems," IEEE Trans. Automatic Control, vol.
AC-29, pp. 466-468, 1982.
S94. Ferreira, P. G., and S. P. Bhattacharyya, "On blocking zeros," IEEE
Trans. Automatic Control, vol. AC-22 (2), pp. 258-259, 1977.
S95. F orney, G. D., J f., "Minimal bases of rational vector spaces with applica
tions to multivariable linear systems," SIAM J. Control. vol. 13, pp. 493
520,1975.
S96. Fortmann, T. E., and K. L. Hitz, An lntroduction to Linear Control
Systems. New York: Dekker, 1977.
S97. - - , and D. Williamson, "Design of low-order observers for linear
feedback control laws," IEEE Trans. Automatic Control, vol. AC-17, pp.
301:-308,1972.
S98. Francis, B. A., "The multivariable servomechanism problem from the
input-output viewpoint," 1EEE y,ans. Automatic Control, vol. AC-22 (3),
pp. 322-328, 1977.
S99. - - , and W. M. Wonham, "The role of transmission zeros in linear
multivariable regulators," lnt. J. Control, vol. 22 (5), pp. 657-681, 1975.
al
of the Jordan no
M ath. Software, .
S125. Kailath, T., Linel
S126. Kalman, R. E.. "(
forros," ACTA, J
"--"'-'-'_."."---
_._--- - --
SUPJ>LEMENTARY REFERENCES
083-1094, 1980.
perturbational model
,utomatic Control, vol.
)Iocking zeros," IEEE
-259, 1977.
..
"--'---'----~~~ ~ ~
649
Sl09. Gopinath, B., "On the control of linear multiple input-output systems,"
Bell Syst. Tech. 1., vol. 50, pp. 1063-1081,1971.
SIlO. Gupta, R. D., F. W. Fairman, and T Hinamoto, "A direct procedure for
the design of single functional observers," IEEE 7i-ans. Circuits Systems,
vol. CAS-28, pp. 294-300, 1981.
SIll. Harn, Y. P., and e T Chen, "A proof of a discrete stability test via the
Liapunov theorem," IEEE 7ians. Automatc Control, vol. AC-26, pp.
733-734, 1981.
S1l2. Hayton, G. E., "The generalized resultant matrix," ni. J. Control, vo1.
32, pp. 567-579, 1980.
to Linear Control
SllS. Hikita, H., "Design of exact model matching systems and its applications
to output tracking problems," Int. J. Control, vol. 34, pp. 1095-1111, 1981.
S1l6. Householder, A. S., The Theory of Matrces n Numerical Analysis.
Waltham Mass.: Blaisdell, 1964.
S1l7. Hsia, Te, System Identificaton. Boston: Lexington, 1977.
Sl18. Hsu, e H., and e T. Chen, "Conversion of cyclicity for transfer function
matrices," Int. J. Control, vol. 16, no. 3, pp. 451-463,1972.
B. Moler, "Matrix
1S,"
Lecture Notes in
[-Verlag, 1977.
lation of controllability
--1026,1980.
, matrix polynomials,"
S 124. Kagstrom, B., and A. Rule, "An algorithm for numerical computation
of the Jordan normal form of a complex matrix [F2]," ACM Trans.on
Math. Software, vol. 6, pp. 437-443, 1980.
;enberg-Scheer method
matie.Control, vol. AC
'V.,
S 126. Kalman, R. E., "On partial realizations; transfer functions, and canonical
_-_ __..-.
.
650
REFERENCES
S128. Kamen, E. W., "New results in realization theory for linear time-varying
analytic systems," IEEE Trans. Automatic Control, vol. AC-24, pp. 866
878,1979.
S 129. Kimura, H., "Pole assignment by gain output feedback," IEEE y,ans.
Automatic Control, vol. AC-20, pp. 509-516,1975.
S 130. - - , "Pole assignment by gain output feedback," IEEE y,ans. Automatic
Control, vol. AC-22, pp. 458-463, 1977.
S131. - - , "Geometric structure of observers for linear feedback control
laws," IEEE Trans. Automatic Control, vol. AC22, pp. 846-854, 1977.
S132. - - , "On pole assignment by output feedback," Int. J. Control, vol. 28,
pp. 11-22, 1978.
S133. Klein, G., and B. e Moore, "Eigenvalue-generalized eigenvector assign
ment with state feedback," IEEE y,ans. Automatic Control, vol. AC-22,
pp. 140-141, 1977.
SI34. Klema, V. e, and A. J. Laub, "The singular value decomposition: Its
computation and sorne application," IEEE Trans. Automatic Control,
vol. AC-25, pp. 164-176, 1980.
S135. Krishnamurthi, V., "Implications of Routh stability criteria," IEEE
Trans. Automatica Control, vol. AC-25, pp. 554-555, 1980.
8136. Krishnarao, 1. S., and e T. Chen, "Properness of feedback transfer
function matrices," Int. J. Control, vol. 39, pp. 57-61, 1984.
S 137. - - , and - - , "Two polynomial matrix operations," IEEE Trans.
Automatic Control, vol. AC-29, No. 3, March, 1984.
S138. Kronsjo, L., AIgorithms - their complexity and efficiency. New York:
Wiley,1979.
S139. Kucera, V., Discrete Linear Control- The Polynomial Equation Approach.
Chichester: Wiley, 1979.
SI4(). Kung, S., "Multivariable and multidimensional systems: Analysis and
Oesign," Ph.O. dissertation, Stanford University, 1977.
"/-.,h,,
8141. - - , "A new iow-order approximation a!go!"ithr:0 "/1'"
decomposition," Proc. 12th Amer. Asilomar Con! Circuits, Systems alld
Computers, 1978.
SUPPLEMENTARY REFERENCES
651
8146. Lal, M., H. Singh, and K. A. Khan, "A simplified minimal realization
algorithm for symmetric impulse response matrix using moments,"
IEEE Tans., vol. AC-18, pp. 683-684, 1973.
r EEE
Trans. Automatic
S149. Liu, R., and L. C. Suen, "Minimal dimensional realzation and identi
fiability of input-output sequences," IEEE Tans. Automatic Control,
vol. AC-22, pp. 227-232,1977.
ue decomposition: Its
IS. Automatic Control,
S152. MacFarlane, A. G., and N. Karcanias, "Poles and zeros of linear multi
variable systems: A survey of the algebraic geometric and complex
variable theory," /nt. J. Control, vol. 24, pp. 33-74, 1976.
S153. Mansour, M., "A note on the stability of linear discrete systems and
Lyapunov method," / EEE Trans. Automatic Control, vol. AC-27, pp.
707- 708, 1982.
; of feedback transfer
-61, 1984.
lS
S159. Moler, C. B., and C. F. Van Loan, "Nineteen dubious ways to compute
the exponential of a matrix," SIAM Rev., vol. 20, pp. 801-836, 1978.
S160. Moore, B. c., "On the ftexibility offered by state feedback il'! multi
Trans. Automatic Control, vol. AC-21, pp. 689-692, 1976: its {;omments,
1977.'
.
S161. - - , "Singular value analysis of linear systems," in Proc. 17th / EEE
C0171 Decision Control, pp. 66-73,Jan. 1979.
652
REFERENCES
S168. Pace, 1. S., and S. Barnett, "Efficient algorithms for linear system calcula
tions Part 1-Smith form and common divisor of polynomial matrices,"
Int. J. Systems Sci., vol. 5, pp. 68-72, 1974.
S185. Rosenbrock, I
Wiley- Intersci
8187. - - , "The tn
vol. 25, pp. 11
New
01 Control Systems.
S188. - - , "Comm.
A survey of t
Int. J. Control
8189. - - , and G.
Int.1. Control.
8190. - - , and A. J
25, pp. 389-39
S191. Rozsa, P., and .
of a rational 1
comments by ,
326. 1978.
SI92. Saeks, R., and
disturbance re
Ae-26, pp. 203
S193. Sain, M. K., '
Academic, 198
S194. Safonov, M. G
of multivariab
. matrix," IEEE
S195. Schumacher, J
1EEE 7)ans. Al
SI%. Scott, R.W., al
feedback transf
vol. AC-21, pp.
SUPPLEMENTARY REFERENCES
stems: Control1ability,
" 1976.
New
, "Robust solutions to
ns. Automatic Control,
for linear regulators,"
),231-234,1974.
f eontrollers for linear
nd feedforward designo
~n, IEEE Trans: Auto
94, 1981.
invariant co ntrollable
972.
Jringer-Verlag, 1973.
653
New York:
654
REFERENCES
New York:
8201. Tsui, C. e, and e T. Chen, "An algorithm for companion form realiza
tion," Int. J. Control, vol. 38, pp. 769-777,1983.
8202. - - , "Computational aspects in linear system theory," Ph.D. Disserta
tion, SUNY, Stony Brook, 1983.
8203. Van Dooren, P. M., "The generalized eigenstructure problem in linear
system theory," IEEE Trans. Automatic Control, vol. AC-26, pp. 111-129,
1981.
8204. Varga, A., and V. Sima, "Numerically stable algorithm for transfer
function matrix evaluation," Int. J. Control, vol. 33, pp. 1123-1133, 1981.
8205. Vidyasagar, M., "Conditions for a feedback transfer function matrix to
be proper," IEEE Trans. Automatic Control, vol. AC-20, pp. 570-571,
1975.
8206. - - , Nonlinear Systems Analysis. Englewood Cliffs, N.J.: Prentice-Hall,
1978.
8207. - - , "On the well-posedness of large-scale interconnected systems,"
IEEE Trans. Automatic Control, vol. AC-25, pp. 413-421, 1980.
8208. Wang, J. W., and e T. Chen, "On the computation of the characteristic
polynomial of a matrix," IEEE Trans. Automatic Control, voL AC-27,
pp. 449-551, 1982.
8209. Wang, S. H., Design of Linear Multivariable Systems, Memo. No. ERL
Ivn09, University of 3.liforra, Berkeley, 1971
8210. - - , and E. J. Davison, "A minimization algorithm for the design of
linear multivariable systems," IEEE Trans. Automatic Control, vol.
AC-18, pp. 220-225, 1973.
8211. - - , and - - , "A new invertability criterion for linear multivariable
systems," IEEE Trans. Automatic Control, vol. AC-18, pp. 538-539, 1973.
8212. Wilkinson, J. H., The Algebraic Eigenvalue Problem. London: Oxford
University Press, 1965.
8213. - - , and .C. Reinsch (Eds.), Handbook for Automatic Computation,
Linear Algebra, Vol. 11. New York:' Springer-Verlag, 1971.
.
.
8214. Willems; J. c., The Analysis of FeedbackSystems. Cambridge: MIT
Press, 1971.
8215. Willems, J. L., "Designof state observerS for linear discrete-time systems,"
Int. J. Systems Sci., vol. 11, pp. 139-147,1980.
8216. W olovich, W
IEEE Ti'ans. A
8217. - - , "Multiv
IEEE Ti'al1s. 1
8218. - - , Linear
8219. - - , "Outpu
vol. AC-20, pr
S22{t - - , "Sket pr
vol. AC-23, Pf
8221. - - , "On the
Tians. Automa
8222. - - , "Multpl
Automatic Ca;
8223. - - , P. Ants
minimal and
Control, vol. A
8224. - - , and P. I
variable syster
465,1979.
8225. Wonham, W.
systems?" SI ~l
8226. - - , "Towan
System, Man
S227. - - , Linear
New York: Sp
8228. - - , and J. B.
multivariable s
8229. Wu, M. Y., "A
Trans. Automal
J~i(jl.
--,'':'ome
(il
1EEE Trans. A
8231. - - , "SolutiOI
J. Control, vol.
8232. - - , "A succe
time varying sy
8233. Wu, y, e, ane
dynamc systen
504, 1980.
8234. Yeung, K. S., "
Trans. Automat.
8235. Youla, D. e. J
design of optirr
SUPPLEMENTARY REFERENCES
655
class of multivariable
SEE Trans. Automatie
5218.
--o
n equivalence," Int. J.
New York:
ltations.
,n of the characteristic
S229. Wu, M. Y., "A note on stability of linear time-varying systems," IEEE
Trans. Automatie Control, vol. AC-19, p. 162,1974.
r linear multivariable
18, pp. 538-539, 1973.
~m.
London: Oxford
tomatie Computation,
,lag, 1971.
!s.
Cambridge: MIT
liscrete-time systems,"
23. - - , ...S ome new resuits in inear Lime-varying syslems" (lecn. conesp.),
IEEE y,-ans. Automatie Control, vol. AC-20, pp. 159-161, 1975.
S231. ---, "Solutions of certain classes of linear time varying systems," Int.
J. Control, vol. 31, pp. 11-20,1980.
S232. - - , "A successive decomposition method for the solution of linear
time varying systems," Int. J. Control, vol. 33, pp. 181-186, 1981.
S233. Wu, Y. c., and Z. V. Rekasius, "Deterministic identification of linear
dynamic systems," IEEE Trans. A~tomatie Control, vol. AC-25~pp. 501
504,1980.
S234.. Yeung, K. S., ~'A necessary condUn for H urwitz polynorIlials;'" IEEE ..
Trans. Automatie Control, vol. AC-27, pp. 251-252, 1982,
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
8236. Zhang, S. Y, and C. T. Chen, "An algorithm for the division of two poly
nomial matrices," IEEE Trans. Automatic Control, vol. AC-28, pp. 238
240,1983.
8237. - - , and - - , "Design of unity feedback systems for arbitrary denomi
nator matrix," IEEE Trans. Automatic Control, vol. AC-28, pp. 518-521,
1983.
8238. - - , and - - , "Design of robust asymptotic tracking and disturbance
rejection," Proc. 22nd IEEE Cotif. Decision and Control, Oec. 1983.
8239. IEEE Trans. Automatic Control 1956-1980 Cumulative Index, vol. AC
26, no. 4, August 1981.
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
Asym pto tic tracking p
495
Basis, 14
change of, 18, 22
minimal polynomial,
orthonormal, 17,90
polynomial, 607, 621
Bezout identity, 5.95
Black box, 72 .
Blocking zer6; 499, 633
Bounded-input bc>und
stabHty, 385, 394
Canonical dccomposilic
1976.
!ndex
Additivity, 73, 88
AIgorilhm
ElIclidean, 579
Levenier, 67
trianglllarizalion, 589
Jordan, 150,209
Causality, 76
Characteristic polynomial
block, 223
unifGnn!y, 403
495
Basis, 14
orthonormal, 17, 90
polynomial, 607,621
Siack box, 72
.
stabiiy, 385, 394
.
Controllability, 176
dilTerential, 180
grammian, 184,322
index, 188
ndices, 190
instantaneous, 181
output, 214
unifonn, 182
657
658
INDEX
Coprimeness, 579
left,592
right,592
Convolution integral, 81
Decoupling,372
static, 502
Degree,577
column,600
row,600
Degree of rational malrix, 235
Delta function, 74
Description of systern
external, 3, 71
input-outpul, 71, 72
internal, 3, 71
state-variable, 71, 83
tion, 560
additive,494
Divisor, 579
cornmon, 579
Iefl,592
right, 592
Domain, of function, 19
of linear operator, 27
equivalent, 146
input-normal, 322
lordan-form, 150,209
output~normal, 322
reducible, 207
(Dynamical equation)
Eigenvalue,33
assignment, 336
Eigenvector,33
generalized, 39
grade of, 38
Equivalence, 146,200,296
zero-inpu 1, 148
zero-state, 148
properties or, 55
inpllt-outpllt, 432
Feld,8
of complex numbers, 9
of rational functions; 9
of real numbers, 9
Fraction,599
coprime, 605
irreducible, 605
Frobenius form, 64
block,223
(Function)
integrable, 388
linear, 19
of matrix, 45
Function reproducbilt~
Functional estimator, 3
Fundamental rnatrjx, 13
Gaussian elimination, 61
Generalized eigenvector,
chain of, 39
grade of, 38
Grammian, controlIabili
observability, 197
nonnegative definite, 4
positive semidefinite, 4
Identification, determnisl
Impulse function, 74
Impulse response, 75
matrix,76
Inner product, 59
Input-output pair, 87
irreducible polynomal.1
tion,291
Irreducible realizaton, 20
INDEX
ation)
14
36
',41,66
(Function)
integrable, 388
lnear, 19
of matrix, 45
351
alion, 542,587
~, 401
,200,296
94
1, 146, 151,200
Generalized eigenvector, 38
chain of, 39
grade of, 38
observability, 197
281
f,143
)rm of, 56
lit y, 333
324
t,
377, 445
32
7,379,474
put,432
344, 379
of,118
ltrix, 341,348
Ibers,9
:ions,9
,9
,154
554
Homogeneity, 73, 88
Impulse funclion, 74
Impulse response, 75
matrix,76
Inner product, 59
Input-output pair, 87
659
60,209
Lcverrier algorithm, 67
Linear independence
Linear mapping, 20
Linear operator, 20
change of basis, 22
nullity, 29, 31
Linear space,9
basis of, 14
dimension of, 13
Linearity, 20,73,87
Link, 102
Matrix, 7, 24
;Cf~\i('.,
function of, 51
fundamental, 136
Hermite, 589
hermitian,412
nonsingular, 28, 32
norm of, 57
polynomial, 587
polynomial of, 45
leading,413
rank or, 27
660
INDEX
(Matrix)
symmetric, 412
Sylvester, 582
unitary, 220
Moment,301
Nilpotent matrix, 69
Nonanticipalive system, 76
Norm, of matrix, 57
Euclidean, 57
induct:d,57
of vector, 57
Nullity, 29
left, 29, 69
right,29
index,199
instantaneous, 197
matrix, 198
uniform, 197
unimodular, 591
Realization, 155
input-normal, 322
internally balanced, 2".
irreducible, 208
minimal-dimensional,
of time-varying diffe
252
output-normal, 322
partal, 302
reducible, 207
Reachability, 176,227
Regu\ator problem, 340
Relaxedness, 77,79, 161
Representa lion
of linear operator, 21
of vector, 15
Response, 72
impulse, 75, 126
steady-state, 494, SOl
step, 107, 129
lransient, 494, 50 I
zero-input, 89, 140
zero-state, 89, 140
ReSl1ltanl, 582
generalized, 608
Riccati equation, 68
Ring,9, 588
Robust design, 491, SOl
Routh-Hurwitz criterion,
Roulh table, 397
Row degree coefficient m~
Row searching algorithm,
QR algorithm, 553
QR faclorizatioD, 553
Ql1adratic form, 411.
Range
of function, 19
of linear operator, 27
Rank, of matrix, 27, 28, 32
computation of, 546
of product of. matrices, 31
Rational functian, 82
irreducible, 584
reducible, 584
Rational mati-x, 82
proper,82
strictly proper,82
INDEX
Jrbation, 491
!n,540
'ying system, 153
,1
lIation,437, 443,444
.459
:rix, 587
:ed,601
306,612
,589
512
601
i91
matrix, 171,182,184
'Inite matrix, 4 I3
O
413
matrix,82
of,468,618
315
l66,613
70
,293,295
53
,553
42
tor, 27
27,28,32
f,546
latriees, 31
\,82
r.
32
82
Realizaton, 155
input-normal, 322
irreducible, 208
minimal-dimensional, 233
252
output-normal,322
partal, 302
reducible, 207
Reachability, 176,227
Representation
of linear operator, 2 I
of vector, 15
Response, 72
zero-input,89, 140
Resultanl,582
generalized, 608
Riccati equation, 68
Ring, 9, 588
matrix, 123
S~nllplii1g insiarn.s,
550
scalar,8
Scalar product, 59
output, 121
weightng, 121
Servocompensator, 504
Similar matrices, 23
661
analog computer, 92
digital computer, 93
global, 403
Stabilization, 339,490
Statc, 83 -86
definiton of, 83
equilibrium,401
initial,89
pseudo-, 288
variable, 86
vector, 86
asymptolic, 356
full-dimensiollal,355
fUllctional, 369
open-Ioop,355
reduced-dimensiollal,361
Slate space, 90
Subspace, 11,27
Superposition principie, 73
Sylvester's inequality, 31
causal, 70
continuous-time,121
discrete-time, 121
linear, 73, 87
matrix,293
multivariable, 71
.reduced, 323
relaxed,385
single-variable,7I
time-invariant, 80, 89
zero-memory, 72
662
INDEX
Transfer function, 82
matrix, 82, 90
sampled, 122
Transformation, 143
Lyapunov, 153
orthogonal, 220,222
Tree,102
branch,102
normal, 102
equation, 87, 89
Vandermonde determinant, 64
Vector, 9
normalized, 566
representation of, 15
Vector space, 9
basis of, 14
complex,10
dimension of, 13
ralional, 10
real, la
Zero,623
blocking, 499,633
input-decoupling,292
output-decoupling, 292
Zero-memory system, 72
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
------_._._---'"_._--"----~._-~-----------_._-------~----~----_._---~--_._--------------~---------------------------~--'"._-----