Linear System Theory and Desing PDF

Linear System Theory
and Design
Li near
-~-'
and
HRW
Series in
Electrical and
Computer Engineering
De~
:.
Chi-Tsong Chen
Professor, Department
State University of Nev
M. E. Van Valkenburg, Series Editor

L. S. Bobrow ELEMENTARY LINEAR CIRCUIT ANALYSIS
C. T. Chen LINEAR SYSTEM THEORY AND DESIGN
D. J. Comer. DIGITAL LOGIC AND STATE MACHINE DESIGN
C. H. Durney, L. D. Harris, C. L. Alley ELECTRIC CIRCUITS: THEORY AND
ENGINEERING APPLlCATIONS
G. H. Hostetter, C. J. Savant, Jr., R. T. Stefani DESIGN OF FEEDBACK
CONTROL SYSTEMS
S. Karni and W. J. Byatt MATHEMATICAL METHODS IN CONTINUOUS
AND DISCRETE SYSTEMS
B. C. Kuo DIGITAL CONTROL SYSTEMS
B. P. Lathi MODERN DIGITAL AND ANALOG COMMUNICATION
SYSTEMS
C. D. McGillem and G. R. Cooper CONTINUOUS AND DISCRETE
SIGNAL AND SYSTEM ANALYSIS Second Edition
A. Papoulis CIRCUITS AND SYSTEMS: A MODERN APPROACH
S. E. Schwarz and W. G. Oldham ELECTRICAL ENGINEERING:
AN INTRODUCTION
_ __
,.__ _
.
_ .~_; __
A. S. Sedra and K. C. Smith MICROELECTRONIC CIRCUITS
M. E. Van Valkenburg ANALOG FILTER DESIGN
HOLT, RINEHART /J
New York
Montreal
Chicagc
Toronto

and Design
g
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
This text is a major revision of IntroduClion lO Linear

System Theory by Chi-Tsong Chen, originally published
in 1970. 1970 by Holt, Rinehart and Winston
Copyright 1984 Holt, Rinehart and Winston, Inc.
AII rights reserved. No part of lhis publication may be reproduced or
transmiued in any form or by any means, electronic or mechanical, ineluding
photocopy, recording or any information storage and retrieval system, without
permission in writng from lhe publisher.
Requests for permission to make copies of any part of the work should be
mailed to: Permissions, Holl, Rinehart and Winston, Inc., Orlando,
Florida 32887
Library of Congress Cataloging in Publication Data

Chen, Chj-Tsong.
Linear system theory and designo
Bibliography: p.
Ineludes indexo
1. System analysis.
QA402.C442 1984
2. System designo 1. Title.

003
83-12891
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
*May be omitted without 10ss (
Contents
Preface
Xlll
Glossary of Symbols
Chapter 1
xvii
Introduction
1-1 The Study of Systems

1-2 The Scope of the Book
Chapter 2
Linear Spaces and Linear Operators

2-1
2-2
2-3
2-4
2-5
2-6
2-7
Introduction
6
Linear Spaces over a Field

7
Linear Independence, Bases, and Representations

12
Change of Basis
17
19
Linear Operators and Their Representations

Matrix Representations of a Linear Operator
21
Systems ofLinear Algebraic Equations

26
Eigenvectors, Generalized Eigenvectors, and Jordan-Form
Representations of a Linear Operator

33
*Derivation of a Jordan-Form Representation

38
45
Functions of a Square Matrix

45
Polynomials of a Square Matrix

51

Functions of a Matrix Defined by Means of Power
Series
54
*May be omitted without loss of continuity.
vii
viii
CONTENTS
4-4
*2-8 Nonns and Inner Product

57
2-9 Concluding Remarks

60
Problems
62
4-5
Chapter 3
Mathematical Descriptions of Systems
70
3-1 Introduction
70
3-2 The Input-Output Description

72
Linearity
73
Causality
76
Relaxedness
77
80
Time Invariance
Transfer-Function Matrix
81
3-3 The State-Variable Description

83
83
The Concept of State

86
Dynamical Equations
Linearity
87
Time Invariance
89
Transfer-Function Matrix
90
Analog and Digital Computer Simulations ofLinear Dyna

mical Equations
91
3-4 Examples
94
101
*Dynamical Equations for RLC Networks

3-5 Comparisons of the Input-Output Description and the State
Variable Description
106
3-6 Mathematical Descriptions of Composite Systems

108
Time-Varying Case
108
Time-Invariant Case
111
Well-Posedness Problem
114
*3-7 Discrete-Time Systems

121

124
Problems
125
Chapter 4
linear Dynamical Equations and Impulse-Response
Matrices
133
4-1 Introduction
133
4-2 Solutions of a Dynamical Equation

134
Time-Varying Case
134
134
Solutions of x = A(t)x
139
Solutions of the Dynamical Equation E

Time-Invariant Case 141
4-3 Equivalent Dynamical Eqations

146
Time-Invariant Case
146
*Time-Varying Case
151
Linear Time-Varying Dynamical Equation with
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
IUlations of Linear Dyna
letworks
101
lescription and the State
lposite Systems
5-5
108
*5-6
114
Impulse-Response Matrices and Dynamical

Equations
154
*Time-Varying Case
154
Time-Invariant Case
157
Concluding Remarks
161
Problems
162
Controllability and Observability 01 linear Dynamical

Equations
168
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
Irreducible Realizations, Strict System Equivalence, and

232
Identification
139
146
lical Equation with
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
Realization from the Hankel Matrix

245
*Jordan-Canonical-Form Realization
249
*Realization of Linear Time-Varying Differential
Equations
252
Realizations of Vector Proper Rational Transfer
Functions
253
Realization from the Hankel Matrix

257
Irreducible Realizations of Proper Rational Matrices: Hankel
Methods
265
Method 1. Singular Value Decomposition

268
272
Method II. Row Searching Method
Irreducible Realizations of G(s): Coprime Fraction
Method
276
Controllable-Form Realization
276
Realization ofN(s)D-1(s), Where D(s) and N(s) Are Not

282
Right Coprime
Column Degrees and Controllability Indices
284
Observable-Form Realization
285
Polynomial Matrix Description

287
Strict System Equivalence

292
Identification of Discrete-Time Systems from Noise-Free
Data
300
Persistently Exciting Input Sequences
307
Nonzero Initial Conditions

309
Concluding Remarks
313
317
Problems
State Feedback and State Estimators

7 -1
7-2
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
Canonical-Form Dynamical Equations

325
Single-Variable Case
325
*Multivariable Case
325
State Feedback
334
334
Stabilization
339
Effect on the Numerator of g(s)

339
Asymptotic Tracking Problem-Nonzero Set

Point
340
*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
Full-Dmensional State Estimator

355
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
*7-6 Decoupling by State Feedback

371

377
Problems
378
7-4
atrix
245
tion
249
19 Differential
lal Transfer
ix
257
tional Matrices: Hankel
268
IPosition
lId
272
prime Fraction
\ 276
I~e D(s) and N(s) Are Not
\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
Stability of Linear Systems
365
384
8-1 lntroduction
384
8-2 Stability Criteria in Terms of the Input-Output
Description
385
Time-Varying Case
385
Time-Invariant Case
388
8-3 Routh-Hurwitz Criterion

395
8-4 Stability of Linear Dynamical Equations

400
Time-Varying Case
400
Time-Invariant Case
407
*8-5 Lyapunov Theorem

412
A Proof of the Routh-Hurwitz Criterion

417
*8-6 Discrete-Tme Systems

419

425
Problems
425
284
I,tems from Noise-Free

linces
xi
Linear Time-Invariant Composite Systems: Characteriza

tion, Stability, and Designs
432
9-1 lntroducton
432
9-2 Complete Characterizaton of Single-Variable Composite
Systems
434
9-3 Controllability and Observability of Composte
Systems
439
440
Parallel Connection
Tandem Connection
441
Feedback Connection
444
9-4 Stability of.Feedback Systems

448
Single-Variable Feedback System

449
Multivariable FeedIJack System

451
9-5 Design of Compensators: Unity Feedback Systems

458
458
Single-Input or Single-Output Case

464
xii
CONTENTS
Multivariable Case-Arbitrary Pole Assignment

468
Multivariable Case-Arbitrary Denominator

Matrix Assignment
478
Decoupling
486
9-6 Asymptotic Tracking and Disturbance Rejection
488
488
Multivariable Case
495
Static Decoupling-Robust and Nonrobust
Designs
501
504
State-Variable Approach
9-7 Design of Compensators: Input-Output Feedback
Sytems
506
506
Multivariable Case
511
Implementations of Open-Loop Compensators

517
Implementation 1
517
Implementation 11
519
Applications
523
Decoupling
523
Asymptotic Tracking, Disturbance Rejection, and
Decoupling
526
Model Matching
528

534
Problems
536
Appendix A
Elementary Transformations
A-l
*A-2
A-3
*A-4
542
Oaussian Elimination
543
Householder Transformation
544
546
Row Searching Algorithm

Hessenberg Form
551
Problems
553
Appendix B
Analytic Functions of a Real Variable
Appendix C
Minimum Energy Control
Appendix O
Controllability alter the Introduction of Sampling

Problems
Appendix E
556
564
Hermitian Forms and Singular Value
Decomposition
565
Problems
Appendix F
554
570
On the Matrix Equation AM

Problems
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
rbance Rejection, and
554
of Sampling
572
Poles and Zeros
559
636
577
Coprimeness of Polynomials
577
Reduction of Reducible Rational Functions

584
Polynomial Matrices
587
Coprimeness of Polynomial Matrices

592
Column- and Row-Reduced Polynomial Matrices

599
Coprime Fractions of Proper Rational Matrices

605
Problems
618
Problems
References
Index
657
ompensators
Polynomials and Polynomial Ma~rices
xiii
635
623
This text is intended fo

linear systems and mu
pendent study and ref
mathematical backgrc
matrix manipulation ,
The unstarred section:
first graduate course (
York at Stony Brook.
during the last three ~
multivariable systems,
number of universities
Vlith the advancen
designing systems tha
Consequently, it is im:
one might unknowingl
investigation of aH the
procedures have evo IVI
study and the design ] .
control text per se, b<
optimization, and sens
This text is a revise
Theory which discusse
in 1970. Since then; se
system theory. Aman!
matrices in fractional
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
pertinent to the original text. The geometric approach is well covered in

W. M. Wonham's Linear Multivariable Control: A Geometric Approach, 2d ed.,
Springer-Verlag, New York, 1979 and is outside the scope ofthis text. Hence
the new material of this edition is mainly in the transfer-function matrix in
fractional formo Because of this addition, we are able to redevelop, probably
more simply in concepts and computations, the results of the state variable
approach and establish a fairly complete link between the state-variable
approach and the transfer-function approach.
We aim to achieve two objectives in the presentation. The first one is to
develop major results and design procedures using simple and efficient methods.
Thus the presentation is not exhaustive; only those concepts which are essential
in the development are introduced. For example, the Smith-McMillan form is
not used in the text and is not discussed. The second objective is to enable the
reader to employ the results developed in the text. Consequently, most results
are developed in a manner suitable for numerical computation and for digital
computer programming. We believe that solving one or two problems of each
topic by hand will enhance the understanding ofthe topic and give confidence
in the use of digital computers. With the introduction of the row searching
algorithm (Appendix A), which has been classroom tested, this is possible even
for multivariable systems, as long as their degrees are sufficiently smal!.
The level of mathematics used in this edition is about the same as that of the
original edition. If concepts and results in modern algebra more extensive than
those in Chapter 2 are introduced, sorne results in the text can be developed
more elegantly and extended to more general settings. For example, the
Jordan form can be established concisely but abstractly by using the concepts
of invariance subspaces and direct sumo Its discussion can be found in a large
number of mathematical texts and will not be repeated here. In view of om
objectives, we discuss the computation of the required basis and then develop
the Jordan formo By using sorne concepts in abstract algebra, such as ring,
principal ideal domain, and module, the realization problem (Chapter 6) can be
developed more naturally and sorne results in this text can be extended to delay
differential equations, linear distributed systems, and multidimensional systems.
These are extensively studed in AIgebraic System Theory, which was initiated
by R. E. Kalman in the late 1960s and has extended in recent years most of the
results in this text to linear systems over rings. The concepts used in algebraic
system theory are less familiar to engineering students and require more
mathematical sophistication and will not be discussed. All the results and
design procedures in this text are developed by using only elementary concepts
and results in linear algebra.
The results in this text may eventually be implemented on digital computers.
Because of the finite word lengih, the sensitivity of problems and the stability
of algorithms become important on computer computations. These problems
are complex and extensively discussed in texts on numerical analysis. In our
development, we will take note of these problems and remark briefly wherever
appropriate.
The arrangementof the topics in this text was not reached without any
difficulty. For example, the concepts of poles and zeros seem to be best intro
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
The material connecte,

The logical dependenc
various combinations (
1 teach a one-semestel
following chapters are
I
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
don digital computers.

blems and the stability
tions. These problems
;:rical analysis. In our
;:mark briefl.y wherever
t reached without any
, seem to be best intro-
xvii
duced in Chapter 4. However, their complete treatments require irreducible

realizations (Chapter 6) and coprime fractions of transfer-function matrices
(Appendix G). Moreover, the concept of zeros is used only in Section 9-6.
Hence it was decided to create an appendix for the topic. The coprimeness of
polynomials and polynomial matrices might be inserted in the main text.
This, however, will digress too much from the state-variable approach; thus
the topic was grouped in an appendix.
The logical sequences of various chapters and appendixes are as follows:
Chapler 1
Chapter 2
Secs.2-1
lo 2-5
!
-======
Chapler 3 (These orders can be inlerchanged.)
Chapter 4
.!
Chapler 5 ---- Appendix B

Appendix A -
Seco 5-8
~ Appendixes e and D
Chapler 6
Seco 6-2
Chapler 8
Charler 7
1 - - - - Hankel melhod
Secs. G-l
G-2
Sec.6-5 -Appendix E ....... Secs. 8-5-Appendix F

Appendix G _ Secs. 6-6
8-6
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
The material connected by a broken line is not essential in the development.

The logical dependencies among Chapters 6, 7, 8, ana 9 are loose, ana their
various combinations can b~ adopted in one- or two-semester courses. When
1 teach a one-semester course at Stony Brook, the unstarred sections of the
following chapters are covered:
Chapter 1
Chapter 2
Chapter 3 (Skip Theorem 3-1 and its corollary.)
Chapter 4 (Skip Theorem 4-11.)
Chapter 5 (Emphasize the time-invariant part by skipping Theorems 5-2,
5-5, and 5-11.)
Chapter 6
Chapter 7
Charter 8
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
and discussions has

matter. 1 am gratefl
F. Trace for typing va
and Winston and the ~
and to Professors S. ]
from the People's R
thanks go to m y wife
their support during t
--~ ----~~-----~_.-.-
PREFACE
mplications; hence the

ample, we prove only
; of Theorems 4-1, 4-2,
g:
m 6-5)
Section 6-5)
gn methods using the
d from Sections 2-1 to
7, or only their single-
They are designed to

ld results covered. In
important results are
'ailable from the pub
ive. The length of this
le significant results is
bted to many people in
Desoer's book Linear
n of this book.Rosen
ng the present edition.
or C. A. Desoer. Even
~stions, For this, 1 can
B. J. Leon, E. J. Craig,
M. E. Van Valkenburg,
x:press my appreciation
ike to thank President
versity of Science and
nstitute of Science and
liversity, and Professor
n, for providing oppor
er 9 and Appendix G.
versity to interact with
ang, from various uni
.siderably the presenta
late students, specially
'i. S. Lai, C. C. Tsuiand
le form of dissertations
xix
and discussions has clarified considerably my understanding of the subject

matter. 1 am grateful to Mrs. V. Donahue, C. LaGuardia, T. Marasco, and
F. Trace for typing variolls drafts of this text, to Mr. P. Becker of Holt, Rinehart
and Winston and the staffofCobbjDunlop for their assistance in the production
and to Professors S. H. Wang, K. W. You, and D. Z. Zheng, visiting scholars
from the People's Republic of China, for their proofreading. My special
thanks go to my wife Beatrice and my children Janet, Pauline, and Stanley for
their support during the writing of this text.
Chi-Tsong Chen
Q.E.D.
I
A,B,P, ...
u, y, ex, ..
u, y, ex, ...
!l'
u(s), y(s), G(s), C(s)
v(A), ...
b(G(S)), deg G(s)
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
u(s), y(s), G(s), C(s)
v(A), ...
(G(s)), deg G(s)
A',x', .
A*, x*, .
det A, ...
iC
~(s)
~[sJ
p(A), rank A
End of the proof of a theorem.

This symbol denotes the end of a statement or an
example.
Capital boldface letters denote matrices.
Lowercase boldface letters denote vectors.
Lowercase italic and Greek type denote scalar-valued
functions or scalars. Capital italic letters are also used
in Chapter 9 and Appendix G to denote scalars.
Laplace transformo
1f a letter is used in both time and frequency domains,
circumf1ex wilI be used to denote the Laplace transform
such as u(s) = fe[lJ{t)] and (;(.1') = ft[G(r)J 1r l letter is
used in only one domain, no circumf1ex wi\l be used,
for example, C(s).
The nuIlity of the constant matrix A.
The degree of the rational matrix G(s).
The transpose of the matrix A and the vector x.
The complex-conjugate transpose of the matrix A and
the vector x.
The determinant of A.
The field of complex numbers.
The field of real numbers.
The field of ratianal functions of s withcoefficie'nts in IR.
The set of polynomials of s with coefficients in ~.
The rank of A. lf A is a constant matrix, the rank is
xxi
xxii
GLOSSAR y OF SYMBOLS
defined over iC or IR. If A is a rational or polynomial

matrix, the rank is defined over lR(s).
The degree of the determinant of A(s).
A diagonal matrix with A, B, and e as block diagonal
elements as
deg det A(s)

diag {A, B, C}
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,
it means that the operator is applied to every entry of

the matrix or the vector.
ational or polynomial
~(s).
. A(s).
e as block diagonal

and Design
not necessarily square
a matrix or a vector,
Jlied to every entry of
1-1
The Study 01
The study and design

methods. We apply v.
responses. If the perfo!
eters or connect to it :
design is guided by pas
approach has undoubte
The empirical methe
performance become ve
if physical systems beco
to be experimented. Ir:
Analytical study of phy
development of matherr
distinction between ph)
fact, circuits or control:
systems. A resistor witl
of the resistor does not
inductance is again a 1
amount of current fiowi
for the success of the
properlymodeled.
A physical system rr
asked and the different
amplifier may be modell
I
1
Introduction
1 -1
The Study of Systems
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
may be modeled as a particle in the study of its trajectory; however, it must be

modeled as a rigid body in the maneuvering. In order to develop a suitable
model of a physical system, a thorough understanding of the physical system
and its operatinal range is essential. In this book, we shall refer to models of
physical systems as systems. Hence a physical system is a device or a collection
of devices existing in the real world; a system is a model of a physical system.
Once a system (a model) is found for a physical system, the next step in the
studyis to develop, by applying various physicallaws, mathematical equations
to describe the system. For example, we apply Kirchhoffs voltage and current
laws to electrical.systems and Newton's laws to mechanical systems. The
equations that describe systems may assume many forms; they may be linear
equations, nonlinear equations, integral equations, difference equations, differ
ential equations, or others. Depending on the question asked, one form of
equation may be preferable to another in describing the same system. In con
clusion, a system may have many different mathematical-equation descriptions,
just as a physical system may have many different models.
Once a mathematical description of a system is obtained, the next step in
the study involves analyses-quantitative and/or qualitative. In the quanti
tative analysis, we are interested in exact responses of systems due to the applica
tion of certain input signals. This part of the analysis can be easily carried out
by using a digital or an analog computer. In the qualitative analysis, we are
interested in the general properties of the system, such as stability, control
lability, and observability. This part of analysis is very important, because
desig~ techniques may often evolve from this study.
lf the response of a system is found to be unsatisfactory, the system has to
be improved or optimized. In some cases, the responses of systems can be
improved by adjusting certain parameters of the systems; in other cases, com
pensators have to be introduced. Note that the design is earried out on the
model of a physical system. However, if the model is properly chosen, the
performance of the physical system should be correspondingly improved by
introducing the required adjustments or compensators.
1-2
Ttae Scope of the
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
systems are mainly lil
Y(s)=
and
and their extensions

relationship between
and is called the inp
frequency domain. T
its elements are limitec
as ratios of two polyn
said to be in the poly
is called a dynamical (
set of first-order diffe:
it is called the dynan
or the internal descri
from the eoncepts of
examples, how they ca
The major portion
around Equations (1
and qualitative. The
puters; hence we empl
will be thoroughly in
number of design proc
The design problerr
control systems. In (
without any constraint
design objectives, such
mum degrees which ce
are also interested in (
te designo In the desi;
system to meet some
steady-state error, and
compensators. A des
from the one without
systems is not consideff
of control systems, alth
in its designo For a dis(
546, and 596. In this se
chapter.
We review in Chapt
The objective of thisc
transformations, to sob
THE SCOPE OF THE BOOK
ry; however, it must be

r to develop a suitable
of the physical system
shall refer to models of
a device or a collection
el of a physical system.
em, the next step in the
nathematical equations
,ffs voltage and current
:hanical systems. The
ms; they may be linear
:rence equations, diITer
on asked, one form of
: same system. In con
[-equation descriptions,
~ls.
tained, the next step in

itative. In the quanti
tems dueto theapplica
an be easily carried out
itative analysis, we are
:h as stability, control
~ry important, because
:tory, the system has to
nses of systems can be
1S; in other cases, com
n is carried out on the
IS properly chosen, the
londingly improved by
;: 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
systems are mainly limited to

y(s) = G(s)u(s) = N(s)D- 1(s)u(s) = F- 1 (s)H(s)u(s)
(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
of a matrix. These techniques are very important, if not indispensable, in

analysis and design of linear systems.
In Chapter 3 we develop systematically the input-output description and
the state-variable description of linear systems. These descriptions are devel
oped from the concepts of linearity, relaxedness, causality, and time invariance.
We also show, by examples, how these descriptions can be set up for systems.
Mathematical descriptions of composite systems and discrete-time equations
are also introduced. We also discuss the well-posedness problem in the feedback
systems.
In Chapter 4 we study the solutions of linear dynamical equations. We
also show that different analysis often leads to different dynamical-equation
descriptions of the same system. The relation between the input-output
description and the state-variable description is also established.
We introduce in Chapter 5 the concepts of controllability and observability.
The importance of introducing these concepts can be seen from the networks
shown in Figure 1-1. Their transfer functions are both equal to 1. There is no
doubt about the transfer function of the network shown in Figure 1-I(b);
however, we may ask why the capacitor in Figure 1-I(a) does not play any role
in the transfer function. In order to answer this question, the concepts of
controllability and observability are needed. These two concepts are also
essential in optimal control theory, stability studies, and the prediction or
filtering of signals. Various necessary and sufficient conditions for a dynamical
equation to be controllable and observable are derived. We also discuss the
canonical decomposition of a dynamical equation and introduce an efficient
and numerically stable method of reducing a dynamical equation to an irreduci
ble one.
In Chapter 6 we study irreducible realizations of rational transfer-function
matrices. The problem is to find a controllable and observable linear time
invariant dynamical equation that has a prescribed rational matrix. lts solution
is indispensable in analog and digital computer simulations. lt also offers a
method of synthesizing a rational matrix by using operational-amplifier circuits.
This result is also needed in establishing the link between the state-variable
approach and the transfer-function approach in the design of linear time
invariant systems.
The practical implications ofthe concepts of controllability and observability
11
(urrent
SOUfCC
1:
151
In
,.
I
11
In
In
___ L
lo)
Figure 1-1
lb)
-r
,.
are studied in Chaptel

the eigenvalues of the
feedback with a cons
its state can be gener
val ues. Various desi.
is also established.
We study in Chapl
under the heading of ~
the design of a system
output stability, stabi
total stability. Their
discuss the Lyapunov
criterion. Their cour
In the last chapter
invariant composite s~
zero cancellation of t
with the transfer functi
ways, as shown in Fig
be studied from its con
and (c) cannot. We a
feedback systems. We
function matrices in t
tions: unity feedback
design compensators '
ment and to achieve
reestablish in the tran
in the state-variable
approaches.
A total of eight ;
various chapters are el
(a)
------o:<
Two different networks withthe same transfer function, 1.
Figure 1-2 Thre differ
THE SCOPE Of THE BOOK
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
are studied in Chapter 7. We show that if a dynamical equation is controllable,

the eigenvalues of the equation can be arbitrarily assigned by introducing state
feedback with a constant gain matrix. lf a dynamical equation is observable,
its state can be generated by designing a state estimator with arbitrary eigen
values. Various design procedures are introduced. The separation property
is also established.
We study in Chapter 8 a qualitative property oflinear systems. This comes
under the heading of stability, which is always the first requirement to be met in
the design of a system. We introduce the concepts of bounded-input bounded
output stability, stability in the sense of Lyapunov, asymptotic stability, and
total stability. Their characterizations and relationships are studied. We also
discuss the Lyapunov theorem and then use it to establish the Routh-Hurwitz
criterion. Their counterparts in the discrete-time case are also studied.
In the last chapter we study various problems associated with linear, time
invariant composite systems. One of them is to study the implication of pole
zero cancellation of transfer functions. For example, consider two systems
with the transfer functions l/(s -1) and (s - 1)/(s + 1) connected in three different
ways, as shown in Figure 1-2. We show why the system in Figure 1-2(b) can
be studied from its composite transfer function, but the systems in Figure 1-2(a)
and (c) cannot. We also study stabilities of single-variable and multivariable
feedback systems. We then study the design of compensators by using transfer
function matrices in the fractional formo We study two feedback configura
tions: unity feedback and plant input-output feedback connections. We
design compensators with minimum degrees to achieve arbitrary pole place
ment and to achieve asymptotic tracking and disturbance rejection. We
reestablish in the transfer-function approach essentially the results developed
in the state-variable approach and complete the link between the two
approaches.
A total of eight appendixes are introduced. Their relationships with
various chapters are covered in the Preface.
(a)
(b)
GJ-r
~J
s - \
(c)
;tion, 1.
Figure 1-2 Three dilTerent connections 01" 1/(s -1) and (s - 1)/(s
+ 1).
Section 2-7, we study

and the Cayley-Ham
concepts of inner pro<
This chapter is in
have some basic knm
multiplication, and im
be used. Let A, B, (
respectively. Let 3i be
we have
Linear Spaces and

Linear Operators
CA=C[a
b1
rt
BD= b
bm
2-1
Introduction
In this chapter we shall review a number of concepts and results in linear

algebra that are essential in the study of this text. The topics are carefully
selected, and only those which will be subsequently used are introduced. The
purpose of this chapter is to enable the reader to understand the mechanism of
similarity transformation, to solve linear algebraic equations, to find Jordan
form representations of square matrices, and to compute functions of a matrix,
in particular, exponential functions of a matrix (see Section 2-9, Concluding
Remarks 1 ).
In Section 2-2 we introduce the concepts of field and linear space over a
field. The fields we shall encounter in this book are the field of real numbers,
the field of complex numbers, and the field of rational functions. In order to
have a representation of a vector in a linear space, we introduce, in Section 2-3,
the concept of basis. The relationship between different representations of the
same vector is established. In Section 2-4, we study linear operators and their
representations. The concept of similarity transformation is embedded here.
In Section 2-5, the solutions of a set of linear algebraic equations are studied.
The concepts of rank and nullity are essential here. ' In Section 2-6, we show that
every square matrix has a Jordan-form representation; this is achieved by
introducing eigenvectors and generalized eigenvectors as basis vectors: . in
These identities can b

the product of an n x
The material presel
38, 39,43 to 45, 77, 86
emphasize the differen
(2-12) and Definition
matrix representation
follow naturally.
2-2
In the study o mathe

forms the center of stu
For example, in arith
algebra, we study the
examples of sets includ
the set of aH polynomi
matrices. In this sech
one of those just menti
Consider the set 01
plication with the corr
set. The sum and pr<
set has elements O and
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
LINEAR SPACES OVER A FIELD
Section 2-7, we study functions of a square matrix. The minimal polynomial

and the Cayley-Hamilton theorem are introduced. In the last section, the
concepts of inner product and norm are introduced.
This chapter is intended to be self-contained. The reader is assumed to
have some basic knowledge of matrix theory (determinants, matrix addition,
multiplication, and inversion). The matrix identities introduced below will also
be used. Let A, B, C, D be n x m, m x r, 1 x n, and r x p constant matrices,
respectively. Let a be the th column of A, and let bj be the jth row of B. Then
we have
AB~ [a,
CA = C[al
a,
a.{~J ~ a,b, +a,b, +... +a.b.
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)
These identities can be easily checked. Note that ab i is an n x r matrix, it is

the product of an n x 1 matrix a and a 1 x r matrix b_
The material presented here is well known and can be found in References 5,
38,39,43 t045, 77, 86, and 116. z However our presentation is different. We
emphasize the difference between a vector and its representations [see Equation
(2-12) and Definition 2-7]. After stressing this distinction, the concepts of
matrix representation of an operator and of the similarity transformation
follow naturally.
2-2
Linear Spaces over a Field
In the study of mathematics we must first specify a collection of objects that

forms the center of study. This coHection of objects or eements is called a seto
For example, in arithmetic, we study the set of real numbers. In boolean
algebra, we study the set {O, 1}, which consists of only two elements. Other
examples of sets inelude the set of complex numbers, the set of positive integers,
the set of aH polynomials of degree less than 5, the set of aH 2 x 2 real constant
matrices. In this section when we discuss a set of objects, the set could be any
one of thosejust mentioned or any other the reader wishes to specify~
Consider the set of real numbers. The operations of addition and multi
plication with the commutative and associative properties are defined Tor the
set. The sum and product of any two real numbers are real numbers.. The
set has elements O and 1. Any real number ex has an additive inverse ( - ex) and
l mind, for they Pfovide the

:oduced in this chapter.
Numbers correspond to the References at the end of the book.
LINEAR SPACES AND LINEAR OPERATORS
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.
field are, respectively,
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' f3)' y = a' (/3' y)
4. Multiplication is distributive with respect to addition: For any a, 13, y in :F,
a . (13 + y) = (a . 13) + (a . y)
5. :F contains an element, denoted by O, and an element, denoted by 1, such

that a + O= a, 1 . a = a for every a in :F.
6. To every a in :F, there Is an element 13 in :F s\.1ch that a + 13 = 0. The element
13 is caBed the additive inverse.
7. To every a in :F which is not the element O, there is an element y in :F such
that a y = 1. The element y is caBed the multiplieative inverse.
I
We give some examples to illustrate this concept.
Example 1
Note that the set of all

From the foregoin:
field could be anythin~
objects. The fields W{
familiar ones: the field
field of rational functic
tions of these fields ar
show that they satisfy ;
to denote the field of rea
and use !R(s) to denote
with indeterminate s. 1
field because it has no :
nomials do not forro a
Before introducing
two-dimensional geom
the plane can be consi<
A vector can be shrunl
product of two points
matical terminology, is
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
1. To every pair of ve<

in :!f, called the SUD:
2. Addition is commu
3. Addition is associa
(x 2 +x 3 )
4. fj( contains a vecto
The vector Ois caB(
5. To every x in fI, th<
6. To every a in :F, an(
the sealar product o
7. Scalar multiplicatic
Example 2
Consider
th~
set of all 2 x 2 matrices of the form

X
-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
A sel wilh all properlies of

precisely, a cornrnulalive rir
does lhe sel of polynornials .
LINEAR SPACES OVER AFIELO
field are, respectively,
set with these proper

l in the fo11owing.
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,
11, denoted by 1, such
+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:
~he set {O, 1} does not

nultiplication, because
r, if we define 0+0=
. be verified that {O, 1}
J the conditions listed
ms forms a fie\d. It is
f![, there corresponds a vector Xl +X z

in 9(, ca11ed the sum ofx l and X;.
2. Addition is commutative: For any Xl' X2 in ~ Xl + X2 =X 2 +x l .
3. Addition is associative: For any x, x 2 , and x 3 in f![, (Xl +X 2 )+X 3 =X l +
1. To every pair of vectors Xl and x 2 in
(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.
the usual definitioris of

lements Oand 1 of the
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
-- - - -
--~-----
-~-
-~-----------
--_..
_--~-
8. Scalar multiplication is distributive with respect to vector addition: For

anya in ff and any x, x 2 in f!l', a(x l +X 2 ) = cxx l +cxx 2 .
9. Scalar multiplication is distributive with respect to scalar addition: For
any a, {3 in ff and anyx in f!l', (a + {3)x = cxx + {3x.
10. For any x in !!l, Ix =x, where 1 is the element 1 in ff.
I
Let the vector additioj
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)).
lt is easy to verify tha

not a linear space.
Example 5
Let !!l denote the set e

Then
scalar multiplication dI
homogeneous, then ( Pi
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
~~
Let (f!l', ~) be a linear ~

be a subspace of (!!l, ff:
space over ff.
We remark on the
the vector addition an
space (!!l, ff), they sati:
2-2. Hence we need t(
whether a set iJ}j is a su
is in iJ}j for any Yl, Y2 in
6 are satisfied. Henc~
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
ror any fixed real a is a
~O
Example 7
4
Note that IR (s), with parenlheses, denotes the field of rational functions with real coefficients;
whereas IR [s J, with brackets, denotes the set of polyti.o~ials with real coefficients.
The real vector space 1
LINEAR SPACES OVER A FIELD
, vector addition: For
Let the vector addition and the scalar multiplication be defined as
~2'
n-l
) scalar addition: For
Ir addition and scalar

ns in the fie1d. For
hat (iC, IR) is a vector
) and (rrl (s), rrl) are also
defined over (- 00, 00)

le addition and scalar
a function space.
I
"-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
Let fL denote the set of aH solutions of the homogeneous differential equation

x + 2x + 3x = O. Then (fL, rrl) is a linear space with the vector addition and the
scalar multiplication defined in the usual way. lfthe differential equation is not
homogeneous, then (gr, rrl) is not a linear space. (Why?)
11
We introduce one more concept to conclude this section.
Definition 2-3
as columns
(2-4)
of Xi and the second

Lddition and the scalar
(2-5)
lled the n-dimensional

lsional complex vector
sional rat ional vector
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
We remark on the conditions for a subset of fL to form a subspace. Since

the vector addition and scalar multiplication have been defined for the linear
space (fL, .?), they satisfy conditions 2, 3, and 7 through 10 listed in Definition
2-2. Hence we need to check only conditions 1 and 4 through 6 to determine
whether a set JI is a subspace of (fL,.?). 1t is easy to verify that if IX lY 1 + 1X 2 Y2
is in JI for any y i> Y2 in JI and any 1Xi> 1X 2 in .?", then conditions 1 and 4 through
6 are satisfied. Hence we conclude that a set JI is a subspace of (gr, .?") if
Cf. l Y1 + Cf. 2 Y2 is in JI, IfJl" an y y j , y 2 in JI and any o: 1, 0: 2 in .'!fr.
Example 6
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
s than n with real co

for any fixed real IX is a 's~bspace of (rrl 2 , rrl).
Example 7
tions wth real coefficients;
.1 coefficients.
The real vector space (rrl n, ~) is a subspace of the vector space (iC", rrl).
12
LINEAR SPACES AND LINEAROPERATORS
2-3
then the linear indep

lowing definition.
Linear Independence, Bases, and Representations
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'
A set of vectors Xl, Xz

only if the equation
implies or; = O, where (

ingly, Q[ can be consid
Observe that line
but also on the field.
(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
then Cl.1X l +(Xzx z =0.

the field of real numbe
zero, such that (X 1X1 +
pendent in (IRZ(s), IR), 1
lt is clear from tI
Xl?
x2
>
",
Xi
are HrJea
as a linear combinatio
every one of them can
Definition 2-5
The maximal number

is called the dimension
lf we introduce the notation
(2-7)
In the previous sec

(IR", IR). The meaning
thereare, at most, n lin
dimensional real vect(
pendent veCtors. (Try
LINEAR INDEPENDENCE, BASES, AND REPRESENTATIONS
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.
1ff, (fll', ff), is said to

, 1X2' ... ,IX" in ff, not
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
for IX =0, 1X 2 = O, ... ,

ce of the set, we have
for which (2-6) holds;
:he right-hand side of
( dependent, there are
atisfy Equation (2-6).
lo conclude the linear
. This set of vectors

Ci. 2
= 0,
Ci. 2
= 0, ... ,
Iill
e vector. The set of

fx1=-O, the only way
d.
~"Ja
(2-7)
X2
X,,]ex=
implies ex = O, where cvery component of ex is an element of ff or, correspond
ingly, ex can be considered as a vector in ff".

Observe that linear dependence depends not only on the set of vectors
but also on the field. For example, the set of vectors {Xl' X2}, where
s~ 1]
x ~
), ... , IX" = 0, then the

'nt.
I
= 1,
13
s+2
]
(S+1:S+3)
X2~
1
s+2
[
s+3
is linearly dependent in the field of rational functions with real coefficients.

lndeed, if we choose
IX = -1
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)).
lt is clear from the definition of linear dependence that if the vectors

"2, ... ,X" are linearly dependent, then at least one of them can be written
as a linear combination of the others. However, it is not necessarily true that
every one of them can be expressed as a linear combination of the others.
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).
In the previous section we introduced the n-dimensional real vector space

(IR", IR). The meaning of n-dimensional is now clear. lt means that in (lR'i, iR)
there are, at most, n linearly independent vectors (over the field IR). In the two
dimensional real vector space (IR 2; IR), one cannot find three linearly inde
pendent vectors. (Try 1)
14
Example 3
Consider the function space that consists of aH real-valued piecewise continuous

functions defined over ( - 00, 00). The zero vector in this space is the one which
is identicaHy zero on ( - 00, 00). The foHowing functions, with - 00 < t < 00,
of el> e z, ... , en' No,"

is another linear com
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
which, together with 1
L ai=O
This completes the pI
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
This theorem ha:

vector space (f!{, g-), i
represented by a set (
(2-7), we may write (2
where
~ctor
P = [{Jl> {Jz,
P can be cons
A set of linearly independent vectors of a linear space (f!{, g-) is said to be a

basis of q; if every vector in fE can be expressed as a unique linear combination
of these vectors.
one-to-one correspond
the same dimensional,
Theorem 2-1
Definition 2-7
In an n-dimensional vector space, any set of n linearly independent vectors

qualifies as a basis.
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
The geometric plane s1

real vector space. An
~-
(2-8)
We claim that a o =/=-0. Ir a o =0, Equation (2-8) reduces to

(2-9)
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)
This shows that every vector x in fE can be expressed as a linear combination
Figure 2-1 A two-dim<
LINEAR INDEPENDENCE, BASES, AND REPRESENTATlONS
piecewise continuous
:pace is the one which
, with - CIJ < t < CIJ,
:tions {t n, n= 1,2, ... }

mts, lX/S, not all zero,
15
of el> ez, . .. , en Now we show that this combination is unique. Suppose there
is another linear combination, say
x= /3l e + /3z e z + ... + /3nen
(2-11 )
Then by subtracting (2-11) from (2-10), we obtain
0= (/31 -/3)e l + (/3z -/3z)ez +
... + (/3n -/3n)en
which, together with the linear independence of {eJ, implies that

i = 1,2, ... , n
This completes the proof of this theorem.

:he dimension of this
er are of finite dimen
This theorem has a very important implication. In an n-dimensional

vector space (f![, .?"), if a basis is chosen, then every vector in f![ can be uniquely
represented by a set of n scalars /31, /3z, ... , /3n in.?". lf we use the notation of
(2-7), we may write (2-10) as
x= [el e z ... en]p

.?") is said to be a
le linear combination
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![,
The geometric plane shown in Figure 2-1 can be considered as a two-dimensional

real vector space. Any point in the plane is a vector. Theorem 2-1 states that
(2-8)
(2-9)
on of el' ez, ... , e,,,

the assumption that
or i = 1, 2, ... ,n, then
(2-10)
i a linear combination
Figure 2-1 A two-dimensional real vector space.
16
Table
2-1
'"
~s
Bases
Different Representations ofVectors
real vector space (IR

space (IRn(s), lR(s)); a
functions, written as
[~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
This array of n numl

such; that is, it is a ve
of a vector with res
numbers, unless it is
vector. However we :
vectors 5 :
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 ]
hence [3 2 - 2 10]' (where the "prime" denotes the transpose) is the

representation of x with respect to {e , e 2, e 3, e 4}. Ir we choose e = S3 - S2,
e2 = S2 - s, e3 = s -1, and e4 = 1 as the basis vectors, then
x: :::
+::' ~,~
:,;OW" -,') +
relationships are bet\

In this subsection, thi:
Let the representa1
and {e, e 2 , . ,en} be
5(,' -,)+ 3(, -1) + 131
x=[
In order to derivf
infOI:mation of the rel
Hence the representationofx with respectto {e, e 2, e 3, e 4} is [3
5 3
13]'. I
5
In this example, there is a sharp distinction between vectors and representa

tions. However, this is not always the case. Let us consider the n-dimensional
This set of vectors is c~lle

One migot be. tempte,
[e 1 e z ... en] - 1 may
LINEAR INOEPENOENCE BASES, ANO REPRESENTATIONS
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
:is. Observe that we

. the basis vectors (as
nitude (scale) of these
Jases we have different
epresentations of the
and the basis {e 1> ez }
~" symbol denotes the
ors b, el, ez' el' and e z
set of aH real poly

Let el =s3,e z =sz,
: linearly independent
ctors, the vector x =
This array of n numbers can be interpreted in two ways: (1) lt is defined as

such; that is, it is a vector and is independent of basis. (2) It is a representation
of a vector with respect to sorne fixed unknown basis. Given an array of
numbers, unless it is tied up with sorne basis, we shall always consider it as a
vector. However we shall also introduce, unless stated otherwise, the following
vectors 5 :
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 )
:he transpose) is the

e choose el = S3 - sZ,
L
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)
In order to derive the relationship between ~ and p, we need either the

information of the representations of ej, for i = 1, 2, ... , n, with respect to the
Hs [3 5 3 13]'. I
:ctors and representa
der the n-dimensional
5 This set or vectors is called an orthonormal set.

60ne might be tempted to write p=[e 1 ez '" en]-l[e ez ...
[el e z .. . en] - may not be defined as can be seen rrom Example 5.
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
Consider the two set

readily verified that
(2-16)
Using rnatrix nota
EPn]
and
(2-17)
which, by using (2-2), can be written as

[el
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)
Substituting (2-18) into (2-15), we obtain

x = [el
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
ith colurnn: the

P = representation of
e with respect to
[ i
{el' ez, ... , e/l}
(2-21 )
linear Opel
2-4
The concept of a func1

JI, ifwe assign to each
of assignments is can
Figure 2-2(a) is a fur
usually denoted by th(
to the element x of f![
defined is caBed the d
to sorne elernent of ~
This establishes the relationship between p and p. In (2-16), if the repre

sentation of e with respect to [e j , e z, . .. , e,J is used, then we shall obtain
(2-22 )
where
ith colurnn: the ]

Q = ~epr.esentation of
ei wth respect to
{e b e z,, en}
(2-23)
Different representations of a vector are related by (2-20) or (2-22). There

fore, given two sets of bases, if the representation of a vector with respect to
one set of bases is known, the representation of the sarne vector with respect
to the other set ofbases can be cornputed by using either (2-20) or (2-22). Since
p = Pp and p = QP, we have p = PQp,for aH p; hencewe conclude that
PQ = 1
or
P = Q - 1
(2-24 )
y =Lx
-1
(a)
Figure 2-2 Examples i

not representa function.
19
LINEAR OPERATORS AND THEIR REPRESENTATlONS
rations of ei' for i =

the representation of
Oni]'; that is,
Example 6
Consider the two sets of basis vectors of

readily verified that
(~4[SJ, IR)
in Exarnple 5. lt can be
(2-16)
2, ... , n
Using rnatrix nota
and
(2-17)
le. e, " '4] l'. " " '4]
l- ~
O
1
O
-1
1
O -1
~J~l"
""
'4]Q
Clearly, we have PQ = 1 and ~ = P~, or
PI]
rn
=Pl ~lJ::
ll~J ~~ U
P2n
Pnn
(2-18)
2-4
PJ~ rl~~J~l
linear Operators and Their Representations
(2-19)
, e2 , . . . , en} is unique,
(2-20)
(2-21 )
:n (2-16), if the repre

1 we shall obtain
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
20) or (2-22). There

'ector with respect to
.e vector with respect
2-20) or (2-22). Since
:conclude that
(2-24 )
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
Consider the transformation that rotates a point in a geometric plane counter

clockwise 90 with respect to the origin as shown in Figure 2-3. Given any
two vectors in the plane, it is easy to verify that the vector that is the sum of the
two vectors after rotation is equal to the rotation of the vector that is the sum
ofthe two vectors before rotation. Hence the transformation is a linear trans
formation. The spaces (gr, ff) and (J/, ff) of this example are all equal to
(IR Z , ~).
is a linear transforma1
all equal to (O/t, IR).
Matrix representa
two examples that the

defined may be of fini
every linear operatc
dimensional (J/, ff) h:
Ir (gr, ff) and (J/, ff)
operator can still be I
infinite order or a fon
and will not be discus
Theorem 2-2
Let (gr, ff) and (J/,
over the same field. ]
in gr. Then the line:
by the n pairs of m,
respect to the basis {x
be represented by an
column af A is the rep]
Proof
Let x be an arbitrary \
the set af vectors qua
pressed uniquely as (;(
of L, we have
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
which implies that fOj

i = 1, 2, ... ,n. This p
Let the represeI
[au aZi ... ami]';
(2-25)
where the aJs are ele!

2
.-------- -
~s
-3
------------
L[x
--Y2 -0.5
xI
1.5
Figure 2-3 The transformation that rotates a vector counterclockwise 90.
Xz
21
LINEAR OPERATORS AND THEIR REPRESENTA:rIONS
ive realline, the range

e entire realline qy.
o a restricted class of
inear operators, linear
with linear operators
gr, ff) and (qy, ff). A
'ords, L maps (gr, ff)
Je a linear operator if
f O.!/. The reason for

to ensure that cxlLx l
metric plane counter

gure 2-3. Given any
. that is the sum f the
vector that is the sum
ltion is a linear trans
nple are aH equal to
I
'unctions defined over

a linear space whose
~t g be a continuous
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"
[al a2; ...
' , . ,lml
be
amJ'; that is,
(2-25)
a,:
n.] a
']
amI
i= 1,2, ... , n
(2-26)
22
between IX and eX and 1

related by eX = Pez and
in the field 9' and the
the basis {el' e2, ... ,e
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)
where JI ~ [,81 ,82 . .. ,8mJ' and ex ~ [al a2

anJ' are the representations ofy and x, respectively. After the bases are chosen, there are no differences
between specifying x, y, and ex, JI; hence in studying y = Lx, we may just study
the relationship between JI and ex. By substituting (2-27) into (2-28), we obtain
[u l
u2
...
umJJI= [u l
U2
...
umJAex
Hence, by the uniqu

we have Aa=PAP-l
or
(2-29)
where Q ,6,. p-l.
Two matrices A a
matrix P satisfying (:
which, together with the uniqueness of a representation, implies that
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
Hence the representat
The representation of
1t is easy tQ verify that
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
Consider the linear o:

{x, x 2 } as a basis, the
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}
Figure 2-4 Relationships between dilTerent representations of the same operator.
or
If, instead of {Xl'

Yl
LXI = [Xl
Hence the representat
23
LINEAR OPERATORS AND THElR REPRESENTATlONS
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
ith column of A is the

:h respect to the basis
of (JI, ~), the linear
~= A(j
(2-28)
nJ' are the representa
PAIX= PAP-(j
Hence, by the uniqueness of a representation with respect to a specific basis,

we have A(j = PAP-(j. Since the relation holds for any (j, we conclude that
here are no differences

Lx, we may just study
into (2-28), we obtain
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
Jn between the repre

that A depends on the
ent representations of
Consider the linear operator L of Example 1 shown in Figure 2-3. If we choose

{Xl' Xz} as a basis, then
linear operators that

ff) -4 (El", ff). In this
laces. lf a basis of El",
liion A of the linear
For a different basis
\. of the same operator
A and A. Consider
e representations of x
e, ez, ... , en}, respec
; representations with
ld. The matrix repre
.-2. The relationships
,P
=[
Y=Lx=[x 1
xz][~J
and
Yz=Lxz=[x
X Z][
-~J
Hence the representation of L with respect to the basis {x, Xz} is
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
{el' e2,, en}
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
{el. e z,., en}
he same operator.
Y=LX~[~~.
X3]
[-~J
and
Y3=Lx3 =[x
X3J[-~J
Bence the representation of L with respect t6 the basis {x, X3} is
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.
It can be shown that tI
Since the set of vectors

basis. We compute TI
It is clear that
and
The last equation is ob
respect to the basis {b,
The matrix A can al:

of a matrix and n 3 mI
more easily, QA = AQ,
Example 4
Consider the following matrix with coefficients in IR :

3
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
This interpretation can be extended to nonsquare matrices.
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
Figure 2-5 Oifferent representations ofa matrix (an operator).
-iJ
qn] =p-
LINEAR OPERATORS AND THEIR REPRESENTATIONS
mple and verify
25
,f numbers. With the

now give a new inter
~fficients in a field ff,
Ir, we shall consider it
atrix A is independent
~ vectors "1' "2' ... ,"n
la matrix) itself. This
the representation is
lis {" l' n2 , . . . , "n}. lf
~esentation of a with
we conclude that the
:ct to the basis (2-13)
f can be modified as in
from the fact that the
le basis {n, "2,, Un}.
X A has a representa
esentation may be
I transforrnation. In
ways much easier to
ity transformation.
Then 8
It can be shown that the following relation holds (check!):
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
We extend Example 4 to the general case. Let A be an n x n square matrix

with real coefficients. lf there exists a real vector b such that the vectors b,
Ab, ... ,An-Ib are linearly independent and if Anb= -(;(nb-(;(n_IAb- ...
(;( 1 An - 1 b (see Section 2-7), then the representation of A with respect to the basis
{b, Ab, ... , A n - 1 b} is
O O
O -(;(n
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
A matrix of the form shown in (2-33) or its transpose is said to be in the

companionform. See Problem 2-26. This form will constantly arise in this text.
As an aid in memorizing Figure 2-5, we write A = Q - 1AQ as
D&finition 2-9
QA=AQ
Since Q= [ql
The range of a linear (
qnJ, it can be further written as
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
The range of a linear (

Proof
of A, it is easy to shO\
(X1X 1 + (X2 X 2 is an elem
( ffm, ff) (see the rema
Consider the set of linear equations:
+alnxn=Yl
+a2nXn = Y2
flI (A)= {all the elem(

vector x ir
lf Y1 and y 2 are elemel

x 2 in (ff n , ff) such th;
in ff, the vector (X1Y 1 +
Systems of Linear Algebraic Equations
a11x 1 +a12 x 2+
a21 x l + a22x 2 +
operator which maps

ff) that undergo
n
(ff ,
Clearly, A is an m x n matrix, x is an n x 1 vector, and y is an m x 1 vector.

No restriction is made on the integer m; it may be larger than, equal to, or
smaller thanthe integer n. Two questions can be raised in regard to this set of
equations: first, the existence of a solution and, second, the number of solutions.
More specifically, suppose the matrix A and the vector y in Equation (2-36) are
given; the first question is concerned with the condition on A and y under which
at leastone vectorx exists such that Ax = y. If solutions exist, then the second
question is concenled with the number of linearly independent vectors x such
that Ax == y. Inorder to answer these questions, the rankand the nullity of
the matrix A have to be introduced.
We have agreed in the previous section to consider the matrix A as a linear
Let the ith column

then the matrix equati
where Xi, for i = 1, 2, ..

range space fAlCA) is, b
(ff n , ff). lt is the sam
(2-37) ranging througl
that flI(A) is the set of
Since fAlCA) is a linear s:
number of linearly in
9j'(Aj s the maximurn n.
Definition 2-10
The rank of a matrix j

independent columns
of A.
Example 1
Consider the matrix
27
SYSTEMS OF LINEAR ALGEBRAIC EQUATlONS
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
The range of a linear operator A is the set Yf.(A) defined by

(2-34 )
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
ection: Since a linear
lOO se one set of basis
The answer is affirma
ar algebraic equations.
Theorem 2-3
The range of a linear operator A is a subspace of (ff m , ff).

Proof
lf YI and y z are elements of Yf.(A), then by definition there exist vectors XI and
Xz in (ff", ff) such that YI =Ax l , yz=Ax z. We claim that for any (XI and (X2
in ff, the vector (X IY I + (XzY z is also an element of fJJI(A). lndeed, by the linearity
of A, it is easy to show that (XIYI + (XzYz = A((Xlx l + (Xzx 2), and thus the vector
(X IX I + (X2XZ is an element of (ff", ff). Hence the range .<?l(A) is a subspace of
(ff m , ff) (see the remark following Definition 2-3).
Q.E.D.
(2-35 )
:s of a field ff and the

This set of equations
(2-36)
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
Let the ith column of A be denoted by a; that is, A = [al

then the matrix equati.on (2-36) can be written as
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
The rank of a matrix A, denoted by peA), is the maximum number of linearly

independent columns in A, or equivalently, the dimension of the range space
clA
Example 1
Consider the matrix
28
The range space of A is aH the possible linear combinations of aH the columns

of A, or correspondingly, all the possible linear combinations of the first two
columns of A, because the third and the fourth columns of A are linearIy depen
dent on the first two columns. Hence the rank of A is 2.

The rank of a matrix can be computed by using a sequence of elementary
transformations (see Appendix A). This is based on the property that the rank
of a matrix remains unchanged after pre- or postmultiplications of elementary
matrices (Theorem 2-7). Once a matrix is transformed into the upper triangular
form as shown in (A-6), then the rank is equal to the number of nonzero rows in
(A-6). From the form, it is also easy to verify that the number of linear inde
pendent columns of a matrix is equal to the number of independent rows.
Consequently, if A is an n x m matrix, then
rank A = no. of linear independent columns
= no. of linear independent rows
':::;;min (n, m)
(2-38)
The computation of the rank of a matrix on digital computers, however, is

not a simple problem. Because of limited accuracies on digital computers,
rounding errors always arise on numerical computations. Suppose a matrix,
after transformation, becomes
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
if and only if all the
rows and columns of the matrix are linearly independent.
With the concepts of range space and rank, we are ready to study the
existence problem of the solutions ofAx = y.
1. Given A and giver

Ax = Y if and only
2. Given A, for ever)
if and only if ~(A)

Proof
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
The design of con

reduced to the solutio
very important in OUl
then a11 rows of A are 1
lf A has a full row ranl
[A:y] cannot increase
spanned by the colun
combination of the co
there exists a x such ti
said to have a full coh
vector a will lie in the
be written as a linear c(
a 1 x m vector b such t
After we find out th
to ask how many solUl
'Ne discuss only the ha
Definition 2-11
The null space of a line

X(A)= {aH
The dimension of X(A

In other words, tht
It is easy to show tha
.;V(A), veA), is 0, then
o
Theorem 2-4
Considerthe matrix equationAx = y, where the m x n matrix A maps (!F", !F)

into (!F m , ,'1.
lt is also called the right nt

nu 11 space of A. See Probl
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,
arise from the given

Tors. lf e arises from
natrix has rank 1. lf
o, and the matrix has
msidered as a zero is a
lroblems encountered
Ices S181, S182, S200,
he largest order of all

. A has rank k if and
; not vanish and every
I and Definition 2-10
nple, in Reference 43.
mly ifthe determinant
1 matrix is nonsingular
inearly independent.
e ready to study the
atrix A maps (ff", ff)
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
1. lt follows immediately from the definition of the range space of A. lf the

vector y is not an element of ~(A), the equation Ax = Y has no solution and
is said to be inconsistent.
2. The rank of A, p(A), is by definition the dimension of ~(A). Since ~(A)
is a subspace of (ff m , ff), if p(A) = m, then ~(A)=(ffm, ff). lf ~(A)=
(ff m , ff), then for any y in (ff m , ff), there exists a vector x such that Ax = y.
lf p(A) < m, there exists at least one nonzero vector y in (ff m , ff), but not in
~(A), for which there exists no x such that Ax = y.
Q.E. D.
The design of compensators to achieve various design objectives can be
reduced to the solution of linear algebraic equations. Hence this theorem is
very important in our application. lf A is an m x n matrix and if p(A) = m,
then all rows of A are linearly independent and A is said to have afull row rank.
lf A has a full row rank, no matter what column y is appended to A, the rank of
[A:y] cannot increase and is equal to p(A). In other words, y lies in the space
spanned by the columns of A and, consequently, can be written as a linear
combination of the columns of A. Hence if A has a full row rank, for any y,
there exists a x such that Ax = y. Similarly, if p(A) = n, the m x n matrix A is
said to have a full column rank. lf A has a full column rank, then every 1 x n
vector a willlie in the space spanned by the rows of A and, consequently, can
be written as a linear combination of the rows of A. In other words, there exists
a 1 x m vector b such that bA = a.
After we find out that a linear equation has at least one solution, it is natural
to ask how many solutions it may have. Instead of studying the general case,
we discuss only the homogeneous linear equation Ax: = O.
Definition 2-11
The null space of a linear operator A is the set .IV(A) defined by
.IV(A) = {aH the elements x of (ff", ff)for which Ax = O}
The dimension of .IV(A) is called the nullity of A and is denoted by v(A).
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
solution ofAx = Ois x = O. Ir v(A) = k, then the equation Ax = Ohas k linearly

independent vector solutions.
Note that the null space is a subspace of the domain (ff n , ff), whereas the
range space is a subspace of (ff m , ff).
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
Consider the matrix
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)
Since the vectors [O 1 2J and [1 2 O]' are linearly independent, we

conclude from (2-39) that a vector x satisfies Ax = Oif and only if
The number of linear

n - p(A), where n is th
linearly independent c(
This corollary follc
null space of A. 1t is
is x = O, which is calle
find a nonzero vector x
then Ax = O has a nor
det (A) = O, where det s
We introduce three
Theorem 2-6 (Sylves
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
The composite matrix

applied successively to
of AB is ~(B) and the r
p(AB) S min (p(A), p(B)
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)
Figure 2"6 A composite
n Ax = Ohas k linearly
1
(ff n , ff), whereas the
31
be arbitrarily chosen. Consequently, there are n - peA) linearly independent

vectors in .;V(A). Hence we conclude that n - peA) = veA). We state this as a
theorem; its formal proof can be found, for example, in References 43 and 86.
Theorem 2-5
Let A be an m x n matrix. Then

peA) + veA) = n
Corollary 2-5
the last three columns

A. Hence the rank of
Then
[=iJ
X,)[i]
x,
(2-39)
:arly independent, we
Id only if
The number of linearly independent vector solutions ofAx = O is equal to

n - peA), where n is the number of columns in A, and peA) is the number of
linearly independent columns in A.
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)
Let A, B be q x n and n x p matrices with coefficients in the same field. Then

peA) + p(B) - n.:5: p(AB).:5: min (p(A), p(B))
fA, peA). The solution

. two equations; hence
ed. Let X3 = 1, X4=0,
~ 5 = O; then Xl =
and
"2 = 1. lt is clear that
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(
ons that the vectors of

! components in every
e vectors of .;V(A) can
I
-T---,t-IR(AB'L
I.L --r- --- .
------~--
R{B)
(5P,5)--B-----..-
(]",5)
Figure 2-6 A composite transformation.
R{A)
32
l.1NEAR SPACES AND LINEAR OPERATORS
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
lf B is an n x n matrix and nonsingular, then
Hence Cl*A*ACl = O
which, in turn, imp
conclude that p(A *
2. This part can be Sil
using the fact p(A)
p(A) + p(B) - n = p(A) S p(AB) smin (p(A), n) = p(A)
Hence we have the following important theorem:

Theorem 2-7
Let A be an m x n matrix. Then
p(AC) = p(A)
and
p(DA) = p(A)
for any n x n and m x m nonsingular matrices e and D.
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
1. p(A) = n if and only if p(A *A) = n, or equivalently,

det (A*A) 1=0
2. p(A) = m if and only if p(AA *) = m, or equivalently,

det (AA*) 1=0
Note that A*A is an n x n matrix and AA* is an m x m matrix. In order to
have p(A) = n, it is necessary to have n s m. This theorem will be proved by
using Definition 2-4'. More specifically, we use the fact that if AFiJ. = ~ implies
Cl = O, then a11 the columns of A are linearly independent, and p(A) = n, where n
is the number of the columns of A.
Proof
1. Sufficiency: p(A*A)=n implies p(A)=n.
We show that ACl=O implies

Cl=O under the assumption of p(A*A)=n, where Cl is an n x 1 vector. lf
ACl = O, then A*ACl = O; which, with the assumption p(A *A) = n, implies Cl = O.
Hence we conclude that p(A) = n. N ecessity: p(A) = n implies p(A *A) = n.
Let Cl be an n x 1 vector. We show that A*ACl = O implies Cl = O under the
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
where 1 is the unit mat]

(2-40) is a set of hom
n x n square matrix. f
nontrivial solution if a
an eigenvalue oJA if ar
is a polynomial ofdegr
Since ~(l) is of degree
a11 distinct).
10
The intersection of two linear spaces is a linear space.
lIt is also called a right eige

y is called a left eigellvecto
REPRESENTATIONS OF A LINEAR OPERATOR
~(B)
33
assumption of p(A) = n. The equality A*Aa. = O implies a.*A*Aa. = O. Let

Aor:=[.81 .82 ... .8mJ'. Then
and .;V(A).10 The

onsequently, p(AB) ~
Q.E.D.
a.*A* = [,B!
and
.8!
.,. .8::'t]
2
a.*A*Aor: = 1.81\2 +1.821 + ... +1.8mI 2
Hence a.*A*Aa. = Oimplies .8 i = 0, for i = 1, 2, ... , m; or, equivalently, Aa. = O,

which, in turn, implies a. = Ofrom the assumption of p(A) = n. Therefore we
conclude that p(A*A) = n.
2. This part can be similarly proved or directIy deduced from the foregoing by
Q.E. D.
using the fact p(A) = p(A *).
n)= p(A)
<\)
2-6 Eigenvectors, Generalized Eigenvectors, and Jordan

Form Representations of a Linear Operator
the pre- or postmulti

erty, gaussian elimina
ndix A can be used to
let A * be the complex
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
that Aa. = O implies

is an n x 1 vector. lf
<\ *A) = n, implies (X = O.
=n implies p(A*A)=n.
mplies (X = O under the
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
Equation (2-42) implie:
Consider the matrix
(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.
By assumption the },'s
+1
which has no real-valued solution. Consequently, the matrix A has no eigen

value in IR.
Since the set of real numbers is a part of the field of complex numbers, there
is no reason that we cannot consider the matrix A in (2-41) as a linear operator
that maps (iC Z , iC) into itself. In so doing, then the matrix A has eigenvalues
+i and -i where Z ~ -1.
The constant matrices we shaH encounter in this book are aH real-valued.

However in order to ensure the existence of eigenvalues, we shaH consider them
as linear operators under the field of complex numbers.
With these preliminaries, we are ready to introduce a set of basis vectors
such that a linear operator has a diagonal or almost diagonal representation.
We study first the case in which aH the eigenvalues of A are distinct; the case
where A has repeated eigenvalues will then be studied.
implies Cl. 1 = O. This is

is linearly independent
Let A be the repre:
RecaH from Figure 2-5
Av with respect to {v l'
A is located at the ith
{v, vz, ... , v,,} is
This can also be (

Figure 2-5, we have
Case 1: AII the eigenvalues of A are distinct
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
+azv z ... +a"v" =0
=0, we may reorder Ai in such a way that
(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
REPRESENTATIONS OF A LINEAR OPERA1OR
Equation (2-42) implies that

(A - AzI)(A - A3 1) ... (A - AJ)
(2-41 )
ter Definition 2-12 can

)S (~", IR) into (IR", IR).
in IR is an eigenvalue of
early Ais an eigenvalue
)w
Since
(A-Ajl)v=(A-A)V
and
(A - A;I)v = O
Ct
35
(2-43)
<XiV) = O
ifj+i
the left-hand side of (2-43) can be reduced to

<Xl V'l - Az)(A l - A3 ) . (Al - AII)V 1= O
By assumption the A;'S, for i = 1,2, ... , n, are all distinct; hence the equation
.z + 1
n (Al -A)V l =0
11
<Xl
natrix A has no eigen

omplex numbers, there
H) as a linear operator
trix A has eigenvalues
I
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.
.ok are aH real-valued.

we shal1 consider them
O
Az O
O ..
.
O A3
: a set of basis vectors

agonal representation.
\. are distinct; the case
o
o.
.
O.]
O
(2-44)
..
O .
II
This can also be checked by using a similarity transformation.

Figure 2-5, we have
e an eigen vector of A
We shall use the set of
jo so, we have to show
siso
et Vi be an eigenvector
'1' vz,, VII} is linearly
Vz, ... ,V" are linearly

C such that
(2-42)
lch a way that <Xl 1-0.
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
We conclude that if the eigenvalues of a linear operator A that maps (C", C)

into itself are all distinct, then by choosing the set of eigenvectors as a basis,
the operator A has a diagonal matrix representation with the eigenvalues on
the diagonal:
Example 2
Consider
-IJ
-1
36
The characteristic polynomial of A is A? + 1. Hence the eigenvalues of A are

+ i and - i. The eigenvector associated with )'1 = i can be obtained by solving
the following homogeneous equation:
(A-AII)VI=[l~i
Example 4
Consider
-~~J[~::J=o
Clearly the vector VI = [1

1 - i]' is a solution. Similarly, the vector v2 =
[1
1 + i]' can be shown to be an eigenvector of A2 = - i. Hence the repre
sentation of Awith respect to {v 1> v2} is
The eigenvalues of A al
with Al = 1 can be fom
A=[iO O.J
-1
The reader is advised to verify this by a similarity transformation.
Case 2: The eigenvalues of A are not al! distinct
Unlike the previous case, if an operator A has repeated eigenvalues, it is not

always possible to find a diagonal matrix representation. We shall use examples
to illustrate the difficulty that may arise for matrices with repeated eigenvalues.
Since the matrix (A - A

Consequently, we can
VI = [1 O O]', associ
A3 = 2 can be found as
sufficient to form abas
Example 3
Consider
The eigenvalues of A are Al = 1, A2 = 1, and A3 = 2. The eigenvectors associated

with Al can be obtained by solving the following homogeneous equations:
(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
From this example

values, it is not alway
Consequently, the A c:
it is possible to find a s
is almost a diagonal fa
eigenvalues of A on thl
example, if A has an ~
with multiplicity 1, ther
forms.
In this example, although A has repeated eigenvalues, it can still be diag

onalized. However, this is not always the case, as can be seen from the following
example.
Al
O
O
Al
O
O: O
1 :O
Al: O
O
O
O
o---6--o-~-il-:
1
L __
O: A2
Which form it will as~

discussed in the next su

'"
le eigenvalues of A are
be obtained by solving
37
Example 4
Consider
J=o
(2-46)
nilarly, the vector V 2 =

Hence the repre
, - i.
formation.
The eigenvalues of A are Al = 1, }o2 = 1, and }'3 = 2. The eigenvectors associated

with Al = 1 can be found by solving
O1 2J
[O O 1
(A-A 1 I)v= O O 3 v=O
:d eigenvalues, it is not
We shall use exampies
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_ _ _
les, it can still be diag

seen from the following
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;
Which form it will assunie depends on the characteristics of A and will be

discussed in the next subsection. The matrices in (2-47) are all ofblock-diagonal
38
formo The blocks on the diagonal are of the form

A 1 O
O A 1
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.
Th is set of vectors ['i

of length k.
Let JV denote tb!
that (A - AI)ix = o. lt
is a subspace of .;v+ 1
tion 2-13 is in .;Vk but
defined in (2-49) is in
(A - AI)i1
and
(A-AI)i-l,
hence Vi is in .;Vi but r

Let A be an n x n
discuss in the followin
vectors of A associatet
alized eigenvectors of
i =0,1,2, ... , until rar
notations, weassume n
are as shown in Table :
and is equal to, f
.;VO e ';vI e ';v2 e ....
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
Now we are read)

associated with A. Bet
only one linearly inde
(A -AI)kv = O
and
(A -AIt-lv +0
Note that if k = 1, Definition 2-13 reduces to (A - AI)v = Oand v O, which

is the definition of an eigenvector. Hence the term "generalized eigenvector"
is well-justified.
Let v be a generalized eigenvector of grade k associated with the eigenvalue
Table 2-2 Chains of Ge

p(A - AI)o = 10
p(A -AI)=7
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
REPRESENTATlONS OF A LINEAR OPERATOR
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
on the diagonal just

block associ
11 form, or the J ordan
s and the remaining
dan blocks associated
Jordan block associ
fthe Jordan form: all
1 Jordan
(A - AI)ivi = (A -AlhA -AI)k-i v = (A - AI)kv =0

and
jan-form representa
e10ping a number of
the remainder of this
(A -AI)i-' vi = (A - AI)i-'(A -AV-iV = (A - AI)k-' v 1=0
hence Vi is in .;Vi but not in ';vi-l'

Let A be an n x n matrix and have eigenvalue A with multiplicity m. We
discuss in the following how to find m linearly independent generalized eigen
vectors of A associated with A.. This is achieved by searching chains of gener
alized eigenvectors of various lengths. First we compute ranks of (A - AI)i,
i = O, 1,2, ... , until rank (A - AI)k = n-m. In order not to be overwhelmed by
. notations, weassume n = 10, m = 8, k =4,and the ranks of(A -AW, i =0,1,2,3,4,
are as shown in Table 2-2. The nullity Vi is the dimension ofthe null space ';vi'
and is equal to, following Theorem 2-5, n - peA -AI)i. Because of
fi o C .;v 1 C .;v 2 C . , we have O = Vo :-:;; V1 :>: V2 :-:;; . :-:;; Vk = 111.

Now we are ready to find m = 8 linearly independent eigenvectors of A
associated with A. Because of .;v 3 C .;v4 and V4 - V~ = 1, we can find one and
only one linearly independent vector u in .;v 4 but not in .;v 3 such that
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
we can generate a chHin al faur generalized
Table 2-2 Chains of Generalized Eigenvectors, where B ~ A - Al

p(A ,1.1)0 = 10
p(A -,1,1)=7
p(A _,1,1)2 =4
p(A _,1,1)' =3
p(A _,1,1)4 =2
with the eigenvalue
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)
Because of X z e X 3 and V3 - Vz = 1, there is only one linearly independent

vector in X 3 but not in X z. The u 3 in (2-50) is such a vector; therefore we
cannot find any other linearly independent vector in X 3 but not in X z. Con
sider now the vectors in X z but not in Xl' Because Vz -V1 =3, there are
three such linearly independcnt vectors. Since Uz in (2-50) is one of them, we
can find two vectors v and w such that {u z, v, w} are linearly independent and
and
BvjoO
BwjoO
From v and w, we can generate two chains of generalized eigenvectors oflength

2 as shown in Table 2-2. As can be seen from Table 2-2, the number of vectors
in Xl is equal to v 1 - Vo = 3, hence there is no need to search other vector in
X l ' This completes the search of eight generalized eigenvectors of A associated
with ..1..
Theorem 2-10
which implies, because (

FinaHy, we have C1 =Cs
This completes the prol
Theorem 2-11
The generalized eigem

linearly independent.
This theorem can t
(A-A.il)k(A-A.jlt Th
Now we discuss thf
u4 v 1 V 2 W 1 W Z
associated with other ei,
tion A are the represent
v1 , V z, w1 , wz,x,x}. Bec
(A - A.I)u 4 = u 3 , we ha"
Q[l A. O O ... O
AU 4 =U 3 +A.U 4 = Q[O
transpose. Proceeding
The generalized eigenvectors of A associated with A. generated as in Table 2-2

are linearly independent.
Proof
First we show that if {u z, v, w} is linearly independent, then {u, v1. W1} is linearly
independent. Suppose {u, v1 , w} is not linearly independent, then there
exist ci,i=1,2,3, not aH zero, such that C1U1 +C Zv1 +C3W1 =0. However,
we have
This is a Jordan-form m;
Since y is a vector in X z, the only way ta have Ey = {D is that y = Gi. 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 )
The application of B 3 = (A - A.lf to (2- 51) yields
which implies; because of B3 U4 jo O, C4 = O. Similarly, we can show c 3 =0 by

applying B2 to (2-51). With C3 = C4 = O, the application of B to (2-51) yields
14
rhis number is called the 9

multlplicity oos the number
orders of aH lordan blocks
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}.
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
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
, the number of vectors

search other vector in
.vectors of A associated
lerated as in Table 2-2
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
This is a Jordan-form matrix. Notethat the number of Jordan blocks associated

with ). is equal to v 1 = 3, the dimension af the null space Ol (A. - U).14.
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
VoIe can show C3 = O by

of B to (2-51) yields
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
Its eigenvalues are Al = 1, A2 = 1, A3 = 2. An eigenvector associated with

),3 = 2 is v3 = [5
3 1J'. The rank of (A - Al1) is 2; hence we can find only one
eigenvector associated with Al' Consequently, we must use generalized eigen
vectors. We compute
and
B@(A-A'l)~[~
O
O
(A~A,Il'~[~
O
O
Example 5
Transform the followi
:]
m~
:] ~[~ ]
1
O
O
O
O
1. Compute the eigen'
det (A -AI)=
Since pB 2 = 1 = n-m, we stop here. We search a v such that B 2 v = O and

Bv =/= O. Clearly, v = [O 1 O]' is such a vector. It is a generalized eigenvector
of grade 2. Let
Hence A has eigen'

plicity 1.
2. Compute (A - 21Y,
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:
where Xl = 1, }'3 =2. Hence we have
(A-21)'
A=[~O 0:2
~-~-~J
~l
(2-53)
__
Since peA - 21)3 = n

a generalized eigen'
This can also be obtained by usihg the similarity transformation

A=Q- 1 AQ
where
Q =[v 1
v2 v,]
~[i
15
O
1
~]
We use the ract that
where A and e are square
43
REPRESENTATlONS OF A LINEAR OPERATOR
vector associated with

ce we can find only one
t use generalized eigen
Example 5
Transform the following matrix into the lordan form:

1
3 -1
1
1
1 -1 -1
O
O
O
2
A=
2
O
O
O
O
O
O
O
O
O
O
O
, O5]
O 3
O
O
O
O
1
1
-1 -1
1
1
1
1
(2-54 )
1. Compute the eigenvalues of AY
O 1
det (A -Al) = [(3 -..1.)(1-..1.)
such that B 2 v =0 and

~eneralized eigenvector
+ 1](..1. -2)2[(1 _..1.)2 -1] =(..1. _2)5 A
Hence A has eigenvalue 2 with multiplicity 5 and eigenvalue O with multi

plicity 1.
2. Compute (A - 21Y, for i = 1, 2, ... , as follows:
1
O
1 -1
1
-1
-1
-1
O
1
1
O
O
O
O
B ~ (A-21)=
-1
O
O
O
O
O
O
O
O -1
1
O
O
O
O
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
Since p(A - 21)3 = n-m = 1, we stop here. Because v3 - V2 = 1, we can find

a generalized eigenvector u of grade 3 such that B3u ;= O and B 2 u 1= O. It is
rmation
[5
We use the ract that

de{:
~J = det A det e
where A and e are square matrices, not necessarily or the same order.
44
where
easy to verify that u = [O O 1 O O O] is such a vector. Define
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
This is a chain of generalized eigenvectors oflength 3. Because of Vz - VI = 2,

there are two linearIy independent vectors in .K z but not in .K l. The vector
U z is one of them. We search a vector v which is independent of U z and has
the property BZv = O and Bv i= O. It can be readily verified that v =
[O
O
1 -1
1
1]' is such a vector. Define
VI
O
O
1
-1
1
1
O
O
2
~ Bv=
-2
O
O
Now we have found five generalized eigenvectors of A associated with A. = 2.

3. Compute an eigenvector associated with A. z =0. Let w be an eigenvector
of A associated with A. z =0; then
3
-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};
This is also called a J,

(2-55), we see that thl
below the main diago
In this book, we use
everything discussed
given in Equation (2-:
A Jordan-form re
into itself is unique u:
Jordan blocks and ti
by A. However, bec~
different Jordan-form
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
In this section we shall

tion that maps (en, 1[:)
extensively, because il
function of a matrix c:
matrix, and then defil
matrix.
(2-55)
Polynomials of a s
(C", C) into (1[:",1[:). If
5. This may be checked by using

A=Q-AQ
or
Functions (
and
FUNCTlONS OF A SQUARE MATRIX
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
In this example, if we reorder the basis
{Ul>
associated with A. = 2.
:t w be an eigenvector
o
O
1
w=o
-1
u z, u3, v1, V Z, w} and use
{w, v z, Vl' U 3 , U z, U 1 } as a new basis, then the representation will be
A=
O
O
1
1
1
1
l
Q=[U 1
45
(2-56)
This is also called a lordan-form representation. Comparing it with Equation

(2-55), we see that the new lordan block in (2-56) has l's on the diagonal just
below the main diagonal as a result ofthe different ordering ofthe basis vectors.
In this book, we use mostly the lordan block of the forro in (2-55). Certainly
everything discussed for this form can be modified and be applied to the form
given in Equation (2-56).
A lordan-forro representation of any linear operator A that maps (en, C)
into itselfis unique up to the ordering of lordan blocks. That is, the number of
lordan blocks and the order of each lordan block are uniquely determined
by A. However, because of different orderings of basis vectors, we may have
different lordan-form representations of the same matrix.
1
1
2-7
ution.
the following lordan
In this section we shall study functions ofa square matrix or a linear transforma
tion that maps (en, C) into itself. We shall use the lordan-form representation
extensively, because in terms of this representation almost all properties of a
function of a matrix can be visualized. We study first polynomials of a square
matrix, and then define functions of a matrix in terms of polynomials of the
matrix.
(2-55 )
Polynomials of a square matrix. Let A be.a square matrix that maps

(en, C) into (C n , C). Ifk is a positive integer, we define
Ak.~AAA
and
AO
(k terms) .'
(2:57a)
(2-57b)
46
where 1 is a unit matrix. Let f(A) be a polynomial in A of finite degree; then

f(A) can be defined in terms of(2-57). For example, if f(A) = A3 +2A z +6, then
polynomial of A is
f(A) ~ A 3 +2A z +61
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
Assume that aJorda:
where the
we have
f(A)
Qf(A)Q-I
or
f(A) = Q-l f(A)Q
(2-58)
where Al and A z are square matrices, then
[f(OAd
Definition 2-15
index of A in A.
(2-59)
x n mat
The largest order of
for any polynomial f(A).
One of the reasons to use the Jordan-form matrix is that if
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
The minimal polynol
This can be easily vetified by observing that

k
A=
[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.\
where is the index

Proof
Since the matrices A
as showing that tf(A)
'"f/e -6.rst shovv that tb
A consists of r Jorda
A=di
47
FUNCTIONS OF A SQUARE MATRIX
of finite degree; then

A)=A 3 +2A z +6,then
polynomial of A is
m
~(A) ~
n (A - AJ"
det (Al - A) =
(2-61 )
= 1
square matrix A that

or equivalently, there
QAQ-I with A in a
Assume that a Jordan-form represemation of A is
.(2-62 )
where the n x n matrix

f(A)Q
(2-58)
A denotes aB the Jordan blocks associated with A.
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)
The multiplicity of A is denoted by ni, the index of Ai is denoted by ni' For

the matrix in (2-52), nl=8,nl=4; for the matrix in (2-53), nl=n l =2,nz=
n z = 1; for the matrix in (2-55), ni = 5, nI = 3, nz = nz = 1. lt is clear that
nsn.
(2-60)
Theorem 2-12
The minimal.polynomial of A is
n (A - A)
m
lj(A)
= 1
where ni is the index of A in A.

Proof
nomiaP 6 lj(A) of least
. whose entries are aH

and only if feA) = O.
nimal polynomial, or
lynomial. Computing
imple job; however, if
:>le, its minimal poly-
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 = diag (A il , Az, ... ,Ar)

and
A with multiplcities
hat the characteristic
(Ail - Alt'
.=
t power is 1.
, is.
For example,
..
f.
?.
(Aiz ~ A;I)"'
..
(2-63)
48
LINEAR SPACES ANO LINEAR OPERATORS
Ifthe matrix (Aij - A;I) has dimension
(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
they have, respecth

minimal polynomiah
(Aij - AJ)k = O
~
O
oll
Corollary2-12
(2-64b)
(C,
Let i1(A) ~ det (J,l - .

polynomial of A. Th(
~(A
(2-64c)
for any integer k ;::: nij
Because the chaf

by the minimal polyr
Theorem 2-12.
(2-64d)
By definition, is the largest order of the Jordan blocks in A, or equivalently,

=max(nij,j=1,2, ... ,r). Hence (Aij-AI),=O for j=1,2, ... ,r. Con
sequently, 1jJ(A;) =0. It is easy to see from (2-63) and (2-64) that ifljJ(A)=(A-o:)k
with either o: I=A or k < , then ljJ (A;) 1=0. Hence we conclude that 1jJ = (A - A)
is the minimal polynomial o'f Aj . Now we claim that feA;) = Oif and only if f
is divisible without remainder by ljJ , denoted as ljJ d.r Inde:ed, if ljJ ;\f; then
f can be written as f = ljJ eh, where h is the quotient polynomial, and feA;) =
1jJ(A;)h(A;) = O' h(A;) = O. If f is not divisible without remainder by ljJ , then
lcan be written as f = ljJh +g where gis a polynomial of degree less than .
Now f(A) = O implies g(A) = O. This contradicts the assumption that ljJ is
the minimal polynomial of , for 9 is a polynomial of degree less than that of
ljJ i and g( A) = O. With these preliminaries, the theorem can be readily proved.
From (2-62), we have IjJ() = diag (IjJ() IjJ( 2 ) IjJ(rn )). Since IjJ() = O
i[ and only ifljJ eontains the [actor (il- ),);, we conclude that the mnima1 poly
nomial of and, correspondingly, of A is
. .
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
Let the charaeteristic
-_._--_....::_..__ .:..~.~,----------"---~----_.:..._=--.=--.....:....-~=--_:...._-_:...-::~_.,---...:_-,--,~_:._-----
49
they have, respectively, (A - 3)(A - 1), (A - 3f(A -1), and (A - W(A -1) as
minimal polynomials.
I
(2-64a)
Because the characteristic polynomial is always divisible without remainder

by the minimal polynomial, we have the following very important corollary of
Theorem 2-120
Corollary 2-12
(2-64b)
(Cayley- Hamilton theorem)
Let Ll(A)~det(AI-A)~An+txl;,n-l +'0' +txn-IA+tx ll be the characteristic

polynomial of A. Then
Ll(A)=An+txIAn-l +00. +txll_IA+txJ=O
(2-64c)
The Cayley-Hamilton theorem can also be proved directly without using

Theorem 2-12 (see Problems 2-39 and 2-40).
The reason for introducing the concept of minimal polynomial will be seen
in the following theoremo
(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
for 1= 0, 1,2, ooo, n - 1; i = 1,2, o.. , m

and
(2-65 )
g( )(A) is similarly defined.
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
In_order to apply this theorem, we must know the minimal polynomial of A.

The minimal polynomial can be obtained by transforming A into aJordan
form or by direct computation (Problem 2-42)0 Both methods are complicated.
Therefore it is desirable to modify Theorem 2-13 so that the use of the minimal
polynomial can be avoided.
Corollary 2-13
Let the characteristic polynomial of A be
n (A - A)n
m
- W(A -1); however,
Ll(A) ~ det (Al - A) =
i= 1
SO
Let
f and 9 be two arbitrary polynomials.
for l = 0, 1,2,
, ni - 1
i= 1, 2,
,m
<l)(A) = g(l)(A;)
Now, froro Corollary

On the spectrum of A
If
(2-66)
then f(A)=g(A).
Solving these two eqt

This follows immediately from Theorem 2-13 by observing that the condition
(2-66) implies (2-65). The set of numbers <L)(A;), for i = 1,2, ... ,m and l =
0, 1,2, ... , ni -1 (there are totally n = I7= 1 n) are called the values of f on
the spectrum of A. Corollary 2-13 implies that any two polynomials that have
the same values on the spectrum of A define the same matrix function. To
state it in a different way: Given n numbers, if we can construct a polynomial
which gives these numbers on the spectrum of A, then this polynomial defines
uniquely a matrix-valued function of A. It is well known that given any n
numbers, it is possible to find a polynomial g(A) of degree n - 1 that gives these
n numbers at sorne preassigned A. Hence if A is of order n, for any polynomial
f(A), we can construct a polynomial of degree n - 1,
g(A)=a o +a 1A + oo. +a n_ 1An- 1
(2-67)
such that g(A) = f(A) on the spectrum of A. Hence any polynomial of A can
be expressed as
f(A) =g(A) =aol +a 1 A + ... +a n_ 1An-1
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
Obviously A 100 can a

different g(A) such as 1
Functions of a Sql
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
which is equal to f(A

know that every funct
Ni
n-l
or
g(A) =
Pi
i=O
f1 (A -
Aj )
j= 1
In conclusion, the form of g(A) can be chosen to facilitate the computation.
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)
ting thatthe condition

= 1, 2, ... , m and 1=
ed the values of f on
Jolynomials that have
matrix function. To
mstruct a polynomial
ris polynomial defines
)wn that given any n
: n - 1 that gives these
n, for any polynomial
(2-67)
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.)
(1)100 =exO +ex 1

100'(1)99=ex 1
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
which is equal to f(2) on the spectrum of A. Hence from this definition we

know that every function of A can be expressed as
f(A) =exol +ex 1A + ... +ex n _An-l
:e the computation.
We summarize the procedure of computing a function of a matrix: Given

an n x n matrix A and a function f(A.), we first compute the characteristic
polynomial of A, say
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,
! -~]
~ [~
=(,(0 +(,(1 +(,(2

tel =Ci l +2(,(2
f(2)
e21
g(2)
2(,( 1
= (,(0 +
21
+4(,(2
2 - 2e
O
el
O
[ _el +e21
Hence,
(note that the derivative is with respect

to A, not t)
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
Then the conditions i
Compute eA". Or equivalently, if feA) = eM, what is f(A}?

The characteristic polynomial of Al is (A -lf(A - 2). Let g(A) = Cio +CilA +
2
1X2/"
Then
f(l) = g(l)
f'(I)=g'(I)
(;(0
Here in the last step v

If feA) = eJ..I, then
2
']
e AI
2e 21 _el
re~'1
t;
lo (
Example 4
Note that the derivati
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
for any function f tt

(2-71), any function o
mediately.
Example 6
Given
Consider
~ ~~
[ A,
(n
n)
1
Al
O
O
O
O
1
(2-68)
The characteristic polynomial of is (A - Alr.
1.]
Let the polynomial g(A) be
FUNcnoNs OF A SQUARE MATRIX
53
of the form
g(A) = Cto +Ct 1(A - A} +Ctz(A - Ad
+ ... +Ct n-1(A - A)"-l
Then the conditions in (2-66) give immediately

,
f(n-1)(A}
Ct O =f(A 1), Ct 1=f(A), ... , Ctn- 1= (n-l)!
Hence,
<\)?
. Le g(A) = Cto +Ct 1A +
f(A) =g(A) = f(AI)1

f(A1)
O
erivative is with respect
=?.
[ .
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
<n-1)(A 1)/(n _1)']

f(n-Z)(A)/(n-2)!
f"(A 1)/2!
f'(A 1)/1!
f(1)
Here in the last step we have used (2-64).

If f(A) = e AI , then
e AIl teA,l t Ze AIl/2! ...
Al _
O eA,1
te Al
...
e - ..
..
.
.
(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
A function of a matrix is defined through a polynomial of the matrix;

therefore, the relations that hold for p'olynomials can also be applied to functions
of a matrix. F or example, if A = QAQ -1, then
f(A)
lich is the same as the

15 in Example 3. Con
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)
If the absolute values

be shown that the in
Instead of provin:
matrix [unction, we ~
(2-69).
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
Functions of a matrix defined by means of power series.

We
have used a polynomial of finite degree to define a function of a matrix. We
shall now use an infinite series to give an alternative expression of a function
of a matrix.
The exponential funct
converges for all finit
Definition 2-17
Let the power series representation of a function
be
ce
f(A)
a)i
i=O
with the radius of convergence p. Then the function

is defined as
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. \
This definition is meaningfulonly if the infinite series in (2"76) converges.

lf A k = O for some positive integer k, then (2-76) reduces to .
In (2 e 79), if we choose s
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
Consider the lordan-form matrix

f(A)
A given in (2-68).
f(A l ) + f'(Al)(A - Al)
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
The exponential function
ower series. We
:tion of a matrix. We
pression of a function
Antn
ih 2
1 + At +-- + ... + - + ...

2!
n!
e).! =
converges for all finite Aand t. Hence for any A, we have

1
L
<Xl
eA! =
_tkAk
k=O
(2-78)
k!
I
(2-75 )
. of a square matrix A
(2-76)
lller -ihan p, the radius

;= O for sorne positive
A remark is in order concerning the computation of eAl . If e Al .s computed

by using Definition 2-16, a closed-form matrix can be obtained. However, it
requires the cornputation of the eigenvalues of A. This step can be avoided if
the infinite series (2-78) is used. ClearIy, the disadvantage of using (2-78) is that
the resulting matrix may not be in a clbsed formo However, since the series (2-78)
converges very fast, the series is often used to compute eA! on a digital computer.
We derive sorne important properties of expone;ntial functions of matrices
to close this section. Using (2-78), it can be shown that .
eO
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)
if and only if AB =BA
(2-80)
t, then frorn the fact that e = 1, we have

[eA!J - 1 =
e - Al
(2-81 )
56
By differentiation, term by term, of (2-78), we have
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)
It is easy to show that

ff'
[tJ
Norms and
All the concepts intre

over the field of comp
for convenience in th
space (C", C).
The concept of th!
idea of length. Any r
as a norm if it has the
1. Ilxll?: and Ilxll
=1
2. IlctX11 = Ictlllxll.
3ll x l+ x zll::;llx 1 11+
The last inequality is e
Letx =[Xl Xz .
S-(k+ 1)
By taking the Laplace transform of (2-78), we have

ff'{e At ) =
00
S-(k+l)A k =s-1
k=O
00
{s-IA)k
(2-84)
or
It is well known that the infinite series
f{Je)={l-Je)-1
or
k=O
1 +Je +Je z + ... =
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
lt is easy to verify that

norm 11'llz is called th
used mainly in the sta
if al! the eomponents 01
The concept of nor
into itself, or equivalel
norm of a matrix A is
(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).
where "sup" stands for

least upper bound of
is, for any x in (C"
IIAII
The norm of A is de
norm. For different 11"
then
17
May be skipped without Ir

and its study may be coup
...
-- _..
__....._--,------_._----_._--
-_ .. - .. - - - -----------_.
_-..
NORMS ANO INNER PROOUCT
2-8
(2-82)
natrix commute (see
X)) is defined as
(2-83)
57
Norms and Inner Product 17
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
1. Ilxll ~O and Ilxll =0 if and only ifx =0.

2ll ax ll=l a lllxll
3. Ilx I
+ xzll ::;llxlll + Ilxzll.
The last inequality is called the triangular inequality.

Let x = [Xl Xz .. X n ]'. Then the norm ofx can be chosen as
11
Ilxlld~
(2-84)
i;
~ i~l
(
(2-87)
IXil
11
or
Ilxllz
or
Ilxll oo ~ m~x IXil
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,
if al! the eomponents ofx are finite.

The concept of norm can be extended to linear operators that map (C", C)
into itself, or equivalently, to square matrices with complex coefficients. The
norm of a matrix A is defined as
(2-85)
IIAII
(2-86)
tly large s. However,

86) does hold for all s
) be established from
lcation ofthe Laplace
73) and (2-74), we can
~ ~~~ I\~III = 1I~~el IIAxII
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
May be ski~ped without los~ (lC continuity The ~~terial in tbi's

and its study may be coupled with that chapter.
secti~-n
is used only in Chapter 8,

.
58
and
where aij is the ijth eleroent of A. lfllXl12 is used, then

IIAI12 = (Amax(A*A))1/2
where A* is the coroplex conjugate transpose of A and AmaxCA*A) denotes the
largest eigenvalue of A*A (see Appendix E). Ir Ilxll", is used, then
IIAII",
=ro~x (tl laijl)
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,:
These norms are aH different, as can be seen froro Figure 2-7.

The norro of a roatrix has the foHowing properties
IIA +BII ~IIAII +IIBII
(2-91 )
IIABII ~IIAIIIIBII
(2-92)
These inequalities can be readily verified by observing

li(A +B)xll =IIAx +Bxll~IIAxII +IIBx\I~(IIAII +IIBII)llx\\
where the "overbar"

property iroplies that
that (x, ay) = ex (x, y).
In the coroplex ve<
where x* is the coro

roatrix A, we have
and
Consequently we have
The sum of these two
magnitudes gives the
norm of A.
This magni tude gives

the norm of A.
(a)
The inner product

In fact, this is the norn
(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
The inequality is obvio
-2
O~(x
This magnitude gives

the norm of A.
+exy,
for any ex. Let ex = -
<:
(el
Figure
2'-7
(a) IIAIll =4.
Solid lines denote x.; broken lines denote Ax, where A

(b)
IIAllz =3.7. (e) tlAII", = 5.
= [_
~ ~J
which gives the Schwat
. _ - . -
and
*A) denotes the

used, then
Amax(A
59
IIABxII :$IIAIIIIBxII :$IIAIIIIBllllxll
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
(X, y) =x*y = LXY
(2-93)
i=l
where x* is the complex conjugate transpose of x.

matrix A, we have
Hence, for any square
(x, Ay) =x*Ay

(A*x, y) = (A*x)*y =x*Ay
and
Consequently we have
This magnitude gives
the norm of A.
(b)
(x, Ay) = (A*x, y)
(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)
If we define Ilxll =(x,x)1/2, then
I(x, y)1 :$llxllllyll

Proof
The inequality is obviously true if y
O. Assume now y i= O. Clearly we have
0:$ (x +ay,x +ay) = (x,x) +a(y,x) +a(x, y) +aa(y, y)
(2-95)
for imy CJ. . Let a = - (y, x)j (y, y); then (2-95) becomes
(, where
A= [ 3. 2J.
-1 O
(x x) > (x, y)(y, x) J(x, y)i2

,
. . (y, y)
(y, y)
which gives the Schwarz inequality.
Q.E.D.
----
- -
60
LINEARSPACES'AND LINEAR OPERATORS
2-9
Concluding Remarks
In this chapter we have reviewed a number of concepts and results in linear

algebra which are useful in this book. The following three main topics were
covered:
1. Similarity transformation. The basic idea of similarity transformation
(change of basis vectors) and the means of carrying out the transformation
are summarized in Figures 2-4 and 2-5. Similarity transformations can be
carried out (a) by computing A =PAP- 1 = Q-l AQ, 2r (b) more easily, by
using the concept of representation: the ith column of A is the representation
of Aq with respect to the basis {q, q2,' .. , qn}' The second method will be
constantly employed in the remainder of this book.
2. Jordanform representation of a matrix. For each eigenvalue of A with
multiplicity m, there are m linearly independent generalized eigenvectors.
Using these vectors as a basis, the new representation of A is in a lordan
canonical formo A systematic procedure for searching these generalized
eigenvectors can be found in Reference S43. If aH eigenvalues are distinct,
then a lordan form reduces to a diagonal formo
lordan-form representation is usually developed by introducing invariant
subspaces and their direct sumo We bypass these concepts and concentrate
on the search of the required basis vectors. The interested reader may
correlate our derivation with the concepts of invariant subspaces and their
direct sumo
3. Functions of a square matrix. Three methods of computing a function of a
matrix f(A), where A is an n x n constant matrix, were introduced. (a) lJse
Definition 2-16: First compute the eigenvalues of A, and then find a poly
nomial g(A.) of degree n -1 that is equal to f(A.) on the spectrum of A, then
f(A)=g(A). (b) Use the lordan canonical form of A: Let A=QAQ-l.
Then f(A) = Qf(A)Q -1, where A is in the lordan form and f(A) is computed
in (2-69) and (2-71). (c) Use Definition 2-17.
Remarks are in order regarding the computer computation of the topics
covered in this chapter. A problem is said to be ill-conditioned ifsmall changes
in data willlead to large changes in solutions. In engineering terminology, if a
problem is very sensitive to the variations of data, then the problem is ill
conditioned; otherwise, it is weH-conditioned. A procedure for solving a
problem, or an algorithm, is said to be numerically stable if numerical errors
inside the procedure will not be amplified; otherwise, the algorithm is said to be
numerically unstable. The condition of a problem and the numerical stability
of an algorithm are two independent concepts. Clearly, whenever possible,
numerically stable methods should be used to solve a problem. There are a
large number of texts on the subject of computer computations. The reader is
referred to, e.g., References S138, S181, S182, S200, and S212. Several weH
tested computer programs such as LINPACK and EISPACK are discussed in
References S82, S103, and S182.
The solution of linear algebraic equations is a basic topic in computer
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
Roughly speaking, balan,

or (he norrns of sorne col
can often be irnproved. :
..
._-_._-~_-~._---_.
---
CONCLUDlNG REMARKS
s and results in linear

hree main topics were
ilarity transformation
mt the transformation
ransformations can be
or (b) more easily, by
Ais the representation
second method will be
eigenvalue of A with
leralized eigenvectors.
)n of A is in aJordan
ling these generalized
igenvalues are distinct,
y introducing invariant
lcepts and concentrate
interested reader may
nt subspaces and their
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
computation. The gaussian elimination without any pivoting is numerically

unstable and should be avoided. The gaussian elimination with partial pivoting
is a numerically stable method. Its numerical stability can be further improved
by using complete pivoting or by using the Householder transformation.
However, according to Reference S138, from the point of view of overall per
formance, which inc1udes efficiency, accuracy, reliability, generality, and ease
of use, the gaussian elimination with partial pivoting is satisfactory for most
general matrices.
The gaussian elimination with partial pivoting can be used to transform a
matrix into a triangular form (see Appendix A), and the rank of the matrix can
then be determined. However, because of the difficulty in determining how
small a number should be considered as a zero, ambiguity always occurs. The
numerical property of this process can be improved by using Householder
transformations. It can be further improved if Householder transformations
are employed together with pivoting (see Appendix A). The most reliable
method of computing the rank of a matrix is by using the singular value de
composition (see Appendix E). Although the problem of determining whether
a small number is a zero or not remains, the decomposition does provide a
"distance" to a matrix of lower rank. A subroutine to achieve the singular
value decomposition is available in LINPACK. The singular value decom
position however is quite expensive, and according to Reference S82, its use is
often oversimplified; the results depend highly on balancing. 18 Furthermore,
comparable results are often obtainable with less cost by using Householder
transformations with pivoting.
The eigenvalues of A are the roots of the characteristic polynomial of A.
Once the characteristic polynomial is known, the roots can be solved by using
subroutines for solving the roots of polynomials. However a polynomial
may be iII-conditioned in the sense that small changes. in coefficients may
cause large changes in one or more roots. Hence this procedure of computing
the eigenvalues may change a well-conditioned prob1em into an iII-conditioned
problem and should be avoided. The most reliable method of computing the
ejgenvalues is the so-called QR method. The method transforms A into the
Hessenberg form and then carries out a sequence of QR factorizations (see
Problems A-S and A-6 and References S181 and S2). lf it is necessary to
compute the characteristic polynomial of a matrix, several methods, inc1uding
the Leverrier algorithm (Problem 2-39), are discussed in References S82 and
S181 (see also Reference S28).
The Jordan canonical form is useful in developing a number of concepts
and results (see, e.g., Problems 2-22, 2-23, 2-37, and 2-4S). It is also useful in
solving Riccati equations (Problem 2-46). Computer computation of aJordan
canonical form, however, is an iII-conditioned problem. For a discussion of
this problem, see Reference S108. For a computer program,see Reference S124.
18
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?
a. The set of integers
2-12 Consider Table

respect to the basis {el' e
b, el' ez, el' and e 2 with r
2-13 Show that simil
quently, the same set of <
b. The set of rational numbers
c. The set of all 2 x 2 real matrices

d. The set of polynomial of degree less than n with real coefficients
2-14 Find the P matri
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)?
what are the representati

{h, Ah, A2 b, A3 h}, respee
2-5 Let lR(s) denote the set of all rational functions with real coefficients. Show that
2-16 What are the ranl
(IR (s), lR(s)) and (IR (s), IR) are linear spaces.
2-6 Which of the following sets of vectors are linearly independent?
. Hl [:!l HJ
b.
2-18 Are the matrices
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.
~[:
2-17 Find the bases oftt

2-16.
;n (R'. R)
[21++31iJ. , [104 -+2iJ1. ' [-iJ3
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))?
nonsingular in the field o

the matrices become nur
matrices nonsingular in t
2-19 Does there exist a :
2-8 What are the dimensions of the following linear spaces?

a. (IR, IR)
b. (C, C)
c. (e, IR)
d. (IR (s), lR(s))
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
See Equation (G-31) in Appendix G.
2-10 .Show that the set of a1l2 x 2 matrices with real coefficients forms a linear space over
IR with dimension 4.
If so, find one.
2-20 Consider the set o

x(n) = A"x(O) -t
where A is an n x n cons!;
x(O), under what conditi(
the equation? Hint: Wri
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
hich of the following sets
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
such that the set {O, 1, 2}
1 multiplication such that
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.
2-18 Are the matrices
~J
vise continuous functions
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)
If so, find one.
in (IRn(s), lR(s)) if and only

at
2-20 Consider the set of linear equations
forms a linear space over
x(n)=A"x(O) +An-lbu(O) +An-lbu(l) + ... + Abu(n -2) +bu(n -1)

where A is an n x n constant matrixand b is an n x 1 column vector. Given any x(n) and
x(O), under what conditions on A and b will there exist u(O), u(l), ... , u(n -1) satisfying
the equation? Hint: Write the equation in the form
u(n
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
Find the Jordan-canonical-fonn representations of the following matrices:
A, ~[ : -!J
A, ~[ l~]
3]
A'~[: -I~] A'~[:

A,{t 1'] A'~[~
]
4
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
Let A for i = 1,2, ... ,n be the eigenvalues of an n x n matrix A. Show that

detA=
TI }'i
j=l
2-23 Prove that a square matrix is nonsingular if and only if there is no zero eigenvalue.
2-29 Show that the dete
2-24 'Under what condition will AH = AC imply B' = C? (A is assumed to be a square

matrix.)
2-25
Show that the Vandermonde determinant
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
Consider the matrix
Show that the characteristic polynomial of A is

6(A)
tJ
det(AI-A)=A n +CX1An-l +CX2An-2 + ... +CXn-1A +cx n
lf Al is an eigenvalue of A [that is, 6(A) =0], show that [1 Al AI .,., A~-lris an

eigenvector associated with Al' [The matrix A is called the companion matrix of the poly
nomial 6(A), lt is said to be in the Frobenius form in the numericaL analysis literature.]
19
20
-al1"1
-a 11 (nl-l)
O
O
O
O
See Reference 6.
See Reference SI7.
-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)
~ (11 - 1)(11- 2)' .. (n - i)
i ;?: 1
123i
are generalized eigenvectors of A associated with XI'

Show that the matrix A in Problem 2-26 is nonsingular if and only ir a" f O. Veriry
that its inverse is given by
2-28
_a"i l/a"
:ix A. Show that
A- I
[
:re is no zero eigenvalue.
2-29 Show that the determinant of the m x m matrix
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)
are arbitrary polynomials.
Show that [he characterislic polynornial al lhe rnalrix
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
D.i}s) =aijls("i- I ) + ... + aij("rI)S + aij"i

Note that the submatrices on the diagonal are of the companion form (see Problem 2-26);
the submatrices not on the diagonal are aH zeros except the last row.
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?
2-32 Show that if 1 is an eigenvalue of A with eigenvector x, then f(l) is an eigenvalue
of f(A) with the same eigenvector x.
Find a matrix B such th

for any nonsingular mat]
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,
g(l)=a ol +a l (1-1) and g(1)=a o(1-1) +al(1-1f(1-2) +a z(1-2).

2-34 Given
[~
~ 1~l~J [~O -15~

-12~1
200
160
O O
by using Definition 2-16 and by using the Jordan-form representation.
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,
2-35 Compute eA! for the matrices
tr ARo
=- -
= -
tr AR z
3
~~-
tr AR"_I
(X'l=- ~--
where tr stands for the tr

matrix. This procedure (
The right-hand-side equa
Rn-zs+R,,_d. Forade
2-40 Prove the Cayley

R"_I' R,,-z, ... , from 0=,
2-41 Show, by using Pn
2-37 Let
(sI-A)-1 =
L
2-42 Let
Find a matrix B such that e
= C. Show that if 1, = O for sorne i then the matrix B does
-,
..
~.
- _.. -
_.
-------------- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - .------------
PROBLEMS
67
not exist. Let
c=
ni
fonn (see Problem 2-26);

ow.
'nomials of the following
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.]
2-40 Prove the Cayley-Hamilton theorem.

R n- I, R n- 2, ... , from 0= AR n_ 1 +anI.)
(Hint: Use Problem 2-39 and eliminate
2-41 Show, by using Problem 2-39,

1
.
(si - A)-I = ~(s) [An-I +(s +adA"- 2 +(S2 +als +(2)An- 3 + ...
+(s"-I +a l s n- 2 + ... +an_)I]
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).
q ualifies as an inner prol

space. What is the forrr
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
2-50 Show that an n >

has eigenvalue O with m
called a nilpotent matrix.
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
Consider the matrix equation
PEP +DP +PF +G=O

where all matrices are./) x n constant matrices. lt is called an algebraic Riccati equation.
Define
-F
M= [
G
Let
Q=[Q1
Q3
Q2J
Q4
consist of all generalized eigenvectors of M so that Q -1 MQ = J is in aJordan canonical

formo We write
Show that if Q 1 is nonsingular, then P = Q3Q 1 1 is a solution of the Riccati equation.

2-47 Give three different norms of the vector x
= [1 -4 3J.
2-48 Verify the three norms of A in Figure 2-7.

2-49 Show that the set of all piecewise continuous complex-valued functions defined
over [O, co) formsa linear space over C. Show that
.. (g, h)
21
See Reference S4.
~ IX> g*(t)h(t) dt
2-51 Let A beanm xn

iorms a linear space, can
PROBLEMS
69
nents of Adjoint (sI - A).
q ualifies as an inner product of the space, where 9 and h are two arbitrary functions of the
space. What is the form of the Schwarz inequality in this space?
1genvector of A associated
iefine
2-50 Show that an
A has the property A k = O for k> m if and only if A

has eigenvalue O with multiplicity n and index m. A matrix with the property A k =0 is
caBed a ni/potent matrix. [Hint: Use Equation (2-64) and Jordan canonical fonu.]
n x n matrix
2- 51 Let A be an m x n matrix. Show that the set of aB! x m vectors y satisfying yA = O

forms a linear space, caBed the left null space of A, of dimension In - peA).
r of A associated with Ai,
A)-l can be expressed as
th .
lolynomial is equal to its
is only one Jordan block
..Igebraic Riccati equation.
r is in a Jardan canonical
the Riccati equation.
valued functions defined
input u and the outp

be related by an equ
3
Mathematical Descriptions
of Systems
If the input and out

can be reduced to
which can be further
if the system is relaxe

input and output of
external description (
cepts of time invarim
of state is introduced
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
that relates the input

set of two equations
used to describe a s:
state-variable descript
to illustrate the proce
Comparisons betweer
description are given.
descriptions of paral
Finally, the discrete-t
in the last section.
We are concernec
physical systems; hel
assumed to be real-va
variable or a multivaJ
Definition 3-1
A system is said to be.
terminal and only om
system if and only if
output terminal.
The references for
and116. The main ob
of linearity, causality,
._-_...
_ ...
.. _.. -_._.- _.... -
_-_.~
.---~-------_._-.- ,
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) =
J~eo G(t, r)u(r) dr
(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)
which can be further reduced to

t
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
x(t) = A(t)x(t) + B(t)u(i)
(3-2a)
y(t) = C(t)x(t) + E(t)u(t)
(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
MA'rHEMATlCAL DfSCRIPTlONS OF SYSTEMS
importance in developing the linear equations. They are not introduced very
rigorously. For a more rigorous exposition, see References 60, 68, 109, and
116.
will be excited solely al

tion, it is Iegitimate to
3-2
where H is Some oper

terms of the input u of
initially relaxed at - (
short. Note that Equ
this section, whenever
mean only those input
The Input-Output Description
The input-output description of a system gives a mathematical relation between

the input and output of the system. In developing this description, the know
ledge of the internal structure of a system may be assumed to be unavailable to
us; the only access to the system is by means of the input terminals and the
output terminals. Under this assumption, a system may be considered as a
"black box," as shown in Figure 3-1. Clearly what we can do to a black box is
to apply all kinds of inputs and measure their corresponding outputs, and then
try to abstract key properties of the system from these input-output pairs.
We digress at this point to introduce sorne notations. The system shown
in Figure 3-1 is assumed to have p input terminals and q output terminals.
The inputs are denoted by Ul' U2' ... ,up or by a p x 1 column vector
u = [Ul U2 ... upJ'. The outputs or responses are denoted by Y10 Y2"'" Y q
or by a q x 1 column vector y = [Yl Y2 ... yqJ'. The time interval in which
the inputs and outputs will be defined is from - 00 to +00. We use u or u()
to denote a vector function defined over ( - 00, 00); u(t) is used to denote the
value ofu at time t. Ifthe function u is defined only over [to, tI)' we write u[to,tt!.
If the output at time tI of a system depends only on the input applied at time
t 10 the system is called an instantaneous or zero-memory system. A network
.that consists of orily resistors is sueh a system. Most systems of interest,
however, have memory; that is, the output at time tI depends not only on the
input applied at tI' but also on the input applied before and/or after tI' Hence,
if an input u[tl>oo) is applied to a system, unless we know the input applied
before tI' the output Y[tl.oo) is generally not uniquely determinable. In fact,
for different inputs applied before t lo we will obtain different output Y[tl> 00)'
although the same input U[II,OO) is applied. It is clear that such an input-output
pair, which lacks a unique relation, is ofno use in determining the key properties
of the system. Hence, in developing the input-output description, before an
input is applied, the system must be assumed to be relaxed or at rest, and that
the output is excited solely and uniquely by the input applied thereafter. If the
concept of energy is applicable to a system, the system is said to be relaxed at
time tI if no energy is stored in the system at that instant. As in the engineering
literature, we shall assume that every system is relaxed at time - oo. Con
sequently if an input u( _ 00, 00) is applied at t = - 00, the corresponding output
Linearity. We intro
exactly the same as thl
Definition 3-2
A relaxed system is saie
for any inputs U 1 and ti

relaxed system is said 1<
In engineering litera
for any u 1 , U 2 and any re

in (3-4) and the set of COi
ship in (3-5) is called th(
caBed the property of ha
the system is said to satis
whether or not there is 1
of homogeneity does no
che ollowing exampe.
Example 1
Consider a single-variab
y
for all l. It is easy to vel

homogeneity but not the
Figure 3-1
A system with p input terminals and q output terminals.
The property of addi

geneity. To be precise, 1
_ _ __ 0
_ .
--~,--,_
. -_._.-...-_ ..
.... .. _._."
_~-----.-"---~
._-_. ,_._-_.. __ .
_'_"._._ .._.-,-'-"..'---""---
-~~_._---
__
__ ._ ..
_._'._.
... _---_ ..
THE INPUT-OUTPUT DESCRIPTION
:e not introduced very

ences 60, 68, 109, and
will be excited solely and uniquely by u. Hence, under the relaxedness assump
tion, it is legitimate to write
y=Hu
atical relation between

description, the know
~d to be unavailable to
Lput terminals and the
lay be considered as a
an do to a black box is
ding outputs, and then
nput-output pairs.
1S. The system shown
Jd q output terminals.
p x 1 column vector
enoted by Yl,Y2, ,Yq
;: time interval in which
+00. We use u or u()
) is used to denote the
. [to, tI)' we write U[lo,I,)
he input applied at time
'ry system. A network
)st systems of interest,
epends not only on the
andJor after tI. Hence,
:now the input applied
determinable. In fact,
different output Y[I,OO)'
at such an input-output
lining the key properties
t description, before an
lxed or at rest, and that
'Jplied thereafter. lf the
is said to be relaxed at
1. As in the engineering
ed at time - oo. Con
e corresponding output
73
(3-3)
where H is sorne operator or function that specifies uniquely the output y in

terms of the input u of the system. In this book, we shall call a system that is
initially relaxed at - 00 an initially relaxed system-or a relaxed system, for
short. Note that Equation (3-3) is applicable only to a relaxed system. In
this section, whenever we talk about the input-output pairs of a system, we
mean only those input-output pairs that can be related by Equation (3-3).
Linearity. We introduce next the concept of linearity. This concept is
exactly the same as the linear operators introduced in the preceding chapter.
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.
The property of additivity, however, almost implies the property of homo

geneity. To be precise, the condition H(u l + u 2) = HUI + Hu 2 for any u l and
74
MATHEMATICAL DESCRIPTlONS OF SYSTEMS
sider a relaxed single-
As shown in Figure 3-.

by a series of pulse fUJ
u(t)(jt,(t - tJ~, we can
/i
Figure 3-2
A pulse functan (l/l(t-t l ).
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 as t,. tends to z(

summation becomes
toward a 6-function.
Now if H 6(t - r) is km
puted from (3-9). The
the relaxed system due
for t 1.::;; t < t 1 + ~
Note that (jt,(t - t 1 ) has unit area for

"function"
all~.
As
approaches zero, the limiting
Note that g is a functi

time at which the 6-fun
at which the output is
function, it is called th
write the output at tim
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;
Th;,::; condiUon fer inlerchai

cussion of this problem, see
for any positive c and that

U()O:,(I-I).
(3-7)
for any function f that is continuous at t l'

With the concept of impulse functin, we are ready to develop a mathe
rilatical description for relaxed linear systems. We discuss first single-variable
systems; the result can then be easily extended to multivariable systems. Con
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)
Now as ~ tends to zero, the approximation tends to an exact equality, the

summation becomes an integration and the pulse function b,,(t - t} tends
toward a b-function. Consequentiy, as ~-+O, Equation (3-8) becomes
y=
J~oo (Hb(t -'l:))u('l:)d'l:
(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)
ches zero, the limiting
Note that g is a function of two variables-the second variable denoting the

time at which the b-function is applied and the first variable denoting the time
at which the output is observed. Since g(', r) is the response of an impulse
function, it is called the impulse response of the system. Using (3-10), we can
write the output at time t as
y(t) =
or simply b{unction.
roo g(t, r)u(r) dr
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
Pulse-furiction approximation of an input function.
(3-11 )
For a dis
_._- .-
76
MATHEMATICAL DESCRIPTIONS OF SYSTEMS
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
roo G(t, r)u(r) dr
gll(t, r) gdt, r) ...

G(t,r)= g21~t,r) g22~t,r) .,.
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)
This concept can

If all the voltages ac
are zero at time to, the
at t o and if an input u
by the initial conditio
will be excited by the
at to, then its input-ol
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.
Causality. A system is said to be causal or nonanticipatory if the output of

the system at time t does not depend on the input applied after time t; it depends
only on the input applied before and at time t. In short, the past affects the
future, but not conversely. Hence, if a relaxed system is causal, its input and
output relation can be written as
y(t) = Hu( -
oo,!]
(3-13)
for all t in ( - 00, 00). If a system is not causal, it is said to be noncausal or

anticipatory. The output of a noncausal system depends not only on the past
input but also on the future value of the input. This implies that a noncausal
system is able to predict the input that will be applied in the future. For real
physical systems, this is impossible. Hence causality is an intrinsic property of
every physical system. Since in this book systems are models ofphysical systems,
all the systems we study are assumed to be causal.
If a relaxed system is linear and causal, what can we say about the impulse
response matrix, G(t, r), ofthe system? Every element of G(t, r) is, by definition,
theoutput due to a D-function input applied at time r. Now if a relaxed system
is causal, the output is identically zera before any inputis applied.. Hence a
linear system is causal if and only if
.
G(t,r)=O
for aH r and all t < r
(3-14 )
words, if the net efTec1

system is relaxed at t o
Example 2
A unit-time-delay syst.
one unit of time; that
uf'" - ! . te) == O, although
Now if a system v
to be relaxed at to, the
Hence, if a system is dI
we know that the inpul

and that the system is 1
A legitimate q uesti.
to, how do we know th
THE INPUT-OUTPUT DESCRIPTlON
he output can be com

system is completely
g(', -r) is the impulse
Iled by direct measure
:m.
,nals, and if the system
-11) can be extended to
(3-12 )
~,
y(t) =
Loo
,)
response between the
; is caHed the impulsestem can be related by

s not readily applicable
ld because there is no
:laxed at -~. These
lent.
'patory if the output of
after time t; it depends

>rt, the past affects the
is causal, its input and
(3-13)
aid to be noncausal or
1s not only on the past
lplies that a noncausal
n the future. For real
an intrinsic property of
1els ofphysical systems,

say about the impulse
. G(t, ,) is, by definition,
"ow if a relaxed system
It is applied. Hence a
(3-14 )
(3-15 )
G(t, ,)u(,) d,
Relaxedness. RecaH that the equation y = Hu of a system holds only when

the system is-relaxed at - 00, or equivalently, only when the output y is excited
solely and uniquely by u(-oo,oo)' We may apply this concept to an arbitrary to.
Definition 3-3
and uniquely excited by
lnal due to an impulse

puts at other terminals
?
Consequently, the input-output description of a linear, causal, relaxed system

becomes
A system is said to be relaxed at time t o if and only if the output
~: :~j
77
u[to,oo)'
Y[to,oo)
is solely
This concept can be easily understood if we consider an RLC network.

If aH the voltages across capacitors and aH the currents through inductors
are zero at time to, then the network is relaxed at to. If a network is not relaxed
at t o and if an input u[to,oo) is applied, then part of the response will be excited
by the initial conditions; and for different initial conditions, different responses
will be excited by the same input u[to,oo)' If a system is known to be relaxed
at to, then its input-output relation can be written as
(3-16)
It is clear that if a system is relaxed at - 00, and if u( _ oo,to) == 0, then the

system is still relaxed at to. For the class of systems whose inputs. and outputs
are linearly related, it is easy to show that the necessary and sufficient condition
for a system to be relaxed at t o is that y(t) = HU(-oo,to)=0 for all t~to' In
words, if the net effect of u( _ oo,to) on the output after to is identicaHy zero, the
system is relaxed at to.
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,
Hence, if a system is described by

y(t) =
r' 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
MATHEMATlCAL DESCRIPTIONS OF SYSTEMS
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
J~oo G(t, ,)u(,) d,
y(t) =
is relaxed at t o if and only if U[lo, 00) == O implies Y[lo, 00)
=o.
Proof
N ecessity: If a system is relaxed at to, the output y(t) for t ? t o is given by
1
00
G(t, ,)u(,) d,
lo
Hence, if U[lo,oo) = o, then Y[lo, 00) = O. Sufficiency: We show that if U[lo,oo)

implies Y['o,oo) =0, then the system is relaxed at to. Since
y(t) = roo G(t, ,)u(,) d,
= roo G(t, ,)u(,) d, +
=O
1~ G(t, ,)u(,) d,
the assumptions U[lo. 00) O, Y[lo,oo) O imply that

lo
for aH t ? to
G(t, ,)u(,)d, =0
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
+ +
Ir the impulse-response matrix G(t, ,) of a systerncan be decomposed into

G(t, ,) = M(t)N(,), and ir every element of M is analytic (see Appendix B) .on
( - 00, (0), then ihe system is relaxed at t o if and only if for sorne fixed positive
e, U[lo,lo+<) = O implies Y[lo,lo+<) = O.
(a)
Figure 3-4 A nonline

linear capacitor C.
._..~... ._-_.._.._._--_.__..__.."--_ .....__..... _._--:~:::::~::-. . ::::..-==~~-:::::=:.=.=-:~-=.:-~=-...:.:=.:.-=:..::..:..====::.~---------;------:----------
__
.~
_,_. _ _. , _ . _ ' _ 0
I
i
79
THE INPUT-OUTPUT DESCRIPTlON
without knowing the

g theorem.
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
how that if U[lo, 00) == O
G(t, t)u(t) dt
is a constant vector, the analyticity assumption of M implies that y(t) is analytc

on [to,oo). Consequently, if Y[to,lo+e) ==0 and if u[Co.oo) ==0, then Y[lo,oo)==O
(see Appendix B), and the corollary follows from Theorem 3-1.
Q.E.D.
This is an important result. For any system that satisfies the conditions of
Corollary 3-1, its relaxedness can be easily determined by observing the output
over any nonzero interval of time, say 10 seconds. Ir the output is zero in this
interval, then the system is relaxed at that moment. It will be shown in the
next chapter that the class of systems that are describable by rational transfer
function matrices or linear time-invariant ordinary differential equations
satisfies the conditions of Corollary 3-1. Hence Corollary 3-1 is widely applic
able.
We give an example to illustrate that Theorem 3-1 does not hold for systems
whose inputs and outputs are not linearly related.
Example 3
2:t o is zero, and hence
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
depending on which initial charge the capacitor has.

.\
n the behavior of the

the system. Certainly
Jm time t o to infinity;
'y to do so.
. 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
-r) = HQ"J(t --r)

+\1.)) = g(', -r +\1.)
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
this to the multivaria
for all
and -r. Heno
we know that its inp

invariance properties
time-invariant case, tI:
to be Oand the time ir
we start to consider
(3-19) becomes
y(t)
The second equality (

The integration in (3
G(t, -r) represents the
time -r. G(t) represent~
at -r =0. Following (
if G(t) = O for all t < O
Transfer-function
describable by convol
transform, because it \
an algebraic equation :
of y; that iS,4
Since G(t - -r) = O for

set at 00 ; herree, from
y(s)
Figure 3-5
The effect of a shifting operator on a signa\.
Note thilt G(e, r) and G(t

symbol G is used.
If Y contains delta functic
inc1ude the delta function:
_._ ...
.----
_-",,_
..... _-----_
.. _.. ..
--' .. _..
-_ ..... -
__ . _ - ;
._.
--:-_
".
_ _ ..
_~
_-_. - - _ .... _
__
~
_ . . ._ - _ . . .
L __ _ . . .
o._ _.
~~
~~
_ _
..__
_~
~._
_ _ .. _
- ._,_
._.
_ -
. - -_ _ . _ .
._.~-----_
10t change with time,

tionary. In order to
tor Q(/.. The elfect of
output of Q(/. is equal
defined as ~ Q(/.u if
this to the multivariable case, we obtain 3
G(t, ,) = G(t -', O) = G(t -,)

for all t and ,. Hence, if a system is described by
y(t) =
)r fixed) if and only if

(3-18)
elaxed system is said

I
'iant, what condition

response g(', ,) is the
is, g(,,)=HCi(t-,).
; +ex) implies g(t, ,) =

ing ex = - " we have
e gel, ,) 01 a relaxed,
01 l and ,. Extending
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) =
= Q(/.y, which implies
le output remains the

latter at what time an
tveform of the output
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) =
tOO (tOO G(t-,)U(,)d,)e-S1dt
too (tOO G(t _,)e-S(I-t) dt) u(,)e- St d,
tOO G(v)e- SV dv tOO u(,)e- St d,
~ 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
MATHEMATlCAL DESCRlPTIONS OF SYSTEMS
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
and is called the transfer-function matrix of the system. For single-variable

systems, C(s) reduces to a scalar and is cal1ed the transfer function. Hence the
transfer function is the Laplace transform of the impulse response; it can also be
defined, following (3-21), as
..
;t'[y(t)] I
g( s) = ;t'[ ( )] the system is
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
able only when the S)

say at time t o, then tb
output y [lo, ro) depend
ditions at too Henc
addition to the input
initial conditions is ca
together with the in
example, in newtoni
a particle (system) al
not uniquely determi
known. How the p~
immaterial in determ
two numbers, the po~
state of the system at t
the momentum at lo,
Definition 3-6
The state of a system ;

with U[IO, ro)' determinl
By the behavior o
the system. If the sysl
branch of the networ
be computed. If the
U[IO'''')' in computing 1
[1 summarizes the ess
needed in determining
input ui- 00,11)' i = 1, 2,
though ui - 00,11)' i = 1,
the system are identic;
We give sorne exal
Example 1
Consider the network

current through the i
5
The exceptions are transl

metelS. See Reference 54
-_.. __ ._ .. - -_._---- .. "---._....
_-_._~--,---,~."._-----
"
_.-._-~_.~_.-~-----_.-
..._.. _---,. __ ..
~---_
.. _.- ._--_.. ,._- ..
_----_._~---~--_._._-
--_...
~,._--
.. _- -_ ...
_._._----~._.
__ - _. __ ._._ --_._--------.
_. __ _----. ----..
.._.'--'-"
THE STATE-VARIABLE DESCRIPTION
he variables, and used

the Laplace transform
(3-22)
place transform of the
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.
nite (zero or nonzero)

\. rational matrix G(s)
;onstant matrix. G(s)
/(s2- s +2) is proper

9(5) = N(s)jD(s), 9(s) is
proper if and only if
llynomial. A rational
er. A rational matrix
tly proper.
noises wifl be greatly
als. Hence improper
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
...
The State-Variable Description
The concept of state. The input-output description of a system is applic

able only when the system is initially relaxed. If a system is not initially relaxed,
say at time t o, then the equation Y[to.,.,) = HU[to.,.,) does not hold. In this case the
output Y[to.,.,) depends not only on the input u[to.,.,) but also on the initial con
ditions at too Hence, in order to determine the output Y[to.,.,) uniquely, in
addition to the input u[to.,.,) we need a set of initial conditions at too This set of
initial conditions is called the state. Hence the state at t o is the information that,
together with the input u[to.,.,)' determines uniquely the output Y[to.,.,)' For
example, in newtonian mechanics, if an external force (input) is applied to
a particle (system) at time t o, the motion (output) of the particle for t '2:. t o is
not uniquely determinable unless the position and velocity at time t o are also
known. How the particle actually attained the position and velocity at t o is
immaterial in determining the motion of the particle after too Hence the set of
two numbers, the position and velocity at time t o, is qualified to be called the
state ofthe system at time.t o. Note that the set oftwo numbers, the position and
the momentum al t o, is also qualified as the state.
[)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
By the behavior of a system, we mean aH responses, including the state, of

the system. If the system is a network, we mean the voltage and current of every
branch of the network. Hence from the state at t o, the state at all t> t o can
be computed. If the state at ti> to is known, we need u[l" ,.,), rather than
u[to.ool' in computing the behavior of the system for t '2:. ti' Hence the state at
tI summarizes the essentia! information about the past input Ql( - ,.,.') which is
needed in determining the future behavior ofthe system. We note that different
input ui_>.t),i=1,2,
,may yield the same state at ti' In this case, even
, are different, their effects on the future behavior of
though ui _ ,.,.t), i = 1, 2,
the system are identical.
We give sorne examples to illustrate the concept of state.
Example 1
Consider the network shown in Figure 3-6.. It is well known that if the initial
, current through the inductor and the initialvltage across the capacitor are
,5
The exceptions are transfer functions ,of sorne transducers such as tachorneters and accelera
rneters. See Reference s46.
'
__
..
..
~'-'--'._'
_..
_._.-.~--
~-_._._._--~
~._
_._. __ _.. _. _._._._-
_'
~.
__ _
_ . _ . -
-_._.~_
__ _
_.~--
- _ _.
-.-.'-~'--"-"-'-'--""
- - - _ _ . _ , __ o
. _ . _
__
~'._~
.~~~~~~~~~~~~~~--~~~~
84
VOltag:
source
E3i
~ ~
output after t ~ t o ex(

minable. From (3-24
Taking the derivative
~_-_
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
which, together with f
Solving for
Cl
and
C2
Example 2
We consider again the network in Example 1. The transfer function from

u to y of the network can be easily found as
Bence if the network
222
y(t) = (2
g(5)=(5+1)(5+2)=5+1- 5+2
Bence the impulse response of the network is
-j
g(t) =2e- 1 -2e- 21
(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
We see that if y(to) an,

determined even if thE
y(to) and y(to) qualifie~
also qualifies as the st
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)=
Loo g(t-L)U(L)dL= roo g(t-L)U(L)dL +
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
Consider the network

voltages are known, tl
for any applied input.
as the state of the net...
6
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
are independent of t. Bence if C l and
C2
are known, the
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)
Taking the derivative of (3-24) with respect to t yields6

1
y(t) = -2e-lcl +4e- Zlc 2 +g(O)u(t) +
1
lO
letwork can be deter

acitor voltage qualify
which, together with g(O) =
o, implies that
-o g(t-"l:)u("l:)d"l:
t
y(to) = -2e- IO c l +4e- 2IO cz
Solving for
Cl
and
Cz
(3-27)
from (3-26) and (3-27), we obtain

Cl =
ansfer function from
Cz
0.5e'O (2y(to)
0.5e 210 (y(to)
+ y(to))
+ y(to))
Hence if the network is not relaxed at t o, the output y(t) is given by

y(t) = (2y(t o) + y(to))e-(t-t o) - (y(t o) + y(to))e- 2 (t-to)
t
+
(3-23)
nction 9(S). Now we

laxed at to, the output
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
for t "2: t o (3-24)

Hhave sorne effect on
pacitor and inductor.
inputu(_oo"o)' From
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
are known, the
Figure 3 c 7
A network with a loop that consists of capacitors only.
.. _.
__ o
86
_._. _ _
.. _._ _ . _ . _ _. _
. _ .
._0 __ -
'0_"
__ ._._. , __
....
"'._. __ .
_ .
_ _
_...
_ _
_ _
. '
_. _ _ _
__
...
. __ .
MATHEMATICAL DESCR1PT10NS OF SYSTEMS
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
or, more explicitly,

xl(t) =J
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:
to denote that the statt:

state x(t), for t '2: to, al
output pair of a system
Definition 3-7
A system is said to bell .
----,._._---~
_.
... __.__ .. _-_._.-._---
--
._ ... __ ...._----
---- _._'-.-'."- ----.--_.... -.- ....._-_.-. __.-.-_ ..
~.,-_
..-----_.- --.. ,-----_ .. _-_.-._
THE ST ATE-VARIABLE DESCRPTION
.ts of three capacitors,

t if any two of Xl> Xz,
lentiy, ifany two ofthe
ork is uniquely deter
s, any two of the three
capacitor voltages are
100sing the state of a
,f the least number of
)f variables for general
87
dynamical equation. In this book, we study only the dynamical equations of
the form
i(t) = h(x(t), u(t), t)
(state equation)
(3-28a)
y(t) = g(x(t), u(t), t)
(output equation)
(3-28b)
or, more explicitly,

x(t) =h(x(t), xz(t), ... , xn(t), u(t), uz(t),.., up(t), t)
xz(t)=hz(x(t), xz(t),, xn(t), u(t), U2(t),, up(t), t)
(3-29a)
Xn(t) = hn(x (t), xz(t), ... , xn(t), u(t), uz(t), ... , up(t), t)
is equal to u(t -1) for

iquely from u[to,oo)' we
U[lo-,lo) is qualified to
I
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
y(t) =g(x(t), xz(t),, xn(t), u(t), uz(t),, up(t), t)

yz(t) =gz(x(t), xz(t), ... , xn(t), u(t), uz(t), ... , up(t), t)
(3-29b)
yq(t) =gq(x(t), xz(t), ... , xn(t), u(t), uz(t), ... , up(t), t)
where x = [x Xz ... xnJ' is the state, y = [y Yz '" yqJ' is the output,

and u = [u Uz .. . upJ' is the input. The input u, the output y, and the
state x are real-valued vector functions of t defined over ( - ro, (0). In order
for (3-28) to qualify as a dynamical equation, we must assume that for any
initial state x(t o) and any given u, Equation (3-28) has a unique solution. A
sufficient condition for (3-28) to have a unique solution for a given u and a given
initial state is thath and ahJaxjare continuous functions oft for i,j = 1, 2, ... , n;
see References 24, 77, and 92. lf a uniquesolution exists in (3 c 28), it can be shown
that the solution can be solved in terms of x(to) and U[lo,l)' Hence x serves as
the state at to, as expected. Equation (3-28a) governs the behavior of the state
and is called a state equation. Equation (3-28b) gives the output and is called
an output equation. Note that there is no loss of generality in writing the output
equation in the form of (3-28b), because by the definition ofthe state, the know
ledge ofx(t) and u(t) suffices to determine y(t).
The state space of (3-28) is an n-dimensional real vector space; hence the
set of equation (3-28) is called an n-dimensional dynamical equation.
linearity. We use the notation
(U[lo,oo), x(to)} -+ {x[to, 00)' Y[to, oo)}
to denote that the state x(t o) and the input U[lo,oo) excite the output y(t) and the
state x(t), for t ~to, and call it an input-state-output pair. An input-state
output pair of a system is called admissible ifthe system can generate such a pairo
Definition 3-7
A system is said to be linear if and only if for any two admissible pairs
{x(t o), utto,oo)} -+ {xtto,oo), ytt,oo)}
{XZ(t o), ufro,oo)} -+ {xfro,oo), Yfto,oo)}
88
MATHEMATlCAL DESCRIPTlONS OF SYSTEMS
and any real numbers al and a2 , the following pair

2
{alxl(tO) +a 2x (t o), alutto,ro) +a 2 u fro,ro)}

----> {alxtto,ro) +a 2x fro,ro), alytto,ro) +a 2 Yfto,ro)}
(3-30)
is also admissible. Otherwise, the system is said to be nonlinear.
Similar to Equations (3-4) to (3-6), if al = a 2 = 1, the relationship in (3-30) is

caBed the property of additivity; if a 2 = 0, it is caBed the property of homogeneity.
The combination of these two properties is caBed the principie of superposition.
In this definition, the superposition property must hold not only at the output
but also at aB state variables; it must hold for zero initial state as weB as non
zero initial state. Hence this definition is much more stringent than the one in
Definition 3-2. Consequently, a system may not be linear according to Defini
tion 3-7 but may be linear according to Definition 3-2. For example, the system
in Figure 3-S (a) has a nonlinear capacitor C. Ifthe voltage across the capacitor
is zero at t o = 0, it will remain to be zero for aB t;::: t o no matter what input
waveform is applied. Hence, as far as the behavior at the input and output
terminals is concerned, the nonlinear capacitor can be disregarded. Hence the
system is linear according to Definition 3-2 but not linear according to Definition
3-7. Consider the network in Figure 3-S(b) with a nonlinear capacitor C and a
nonlinear inductor L. Because the L-C loop is in a series connection with the
current source, its behavior will not transmit to the output y. Hence the system
in Figure 3-S(b) is linear according to Definition 3-2 but not linear according to
Definition 3-7.
We discuss now the implication of Definition 3-7. If al =a 2 = 1 and if
xl(t O)= -x 2 (t O)
and
utto,ro) = -ufro,ro)
Responses due to {x(t o), IDI[to, ro)}

= responses due to {x(t o), O} +responses due to {O, u[to, ro)}
(3-31 )
Ll
~)
Figure 3-8
Hence for linear syster

zero-state responses in
in the previous section I
If a system is linear,
h(x(t), u(t), t)
= 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
then the system is said

lowing. Let Q" be the
Definition 3-8
A system is said to be t
and any real number a,

{Q,,~
In
L .
.2 H
xl(to) =x(t o
Time invariance.
then the linearity implies that {O,O}---->{O[to,ro),O[to,ro)}' Hence a necessary

condition for a system to be linear is that if x(to) = and u[to, ro) == 0, then the
responses of the system are identically zero. A very important property of any
linear system is that the responses of the system can be decomposed into two
parts, as
The responses due to {x

exclusively by the non
are caBed zero-state re~
Equation (3-31) follows
3F
T
Y
1
~)
Two nonlinear systems which are linear according to Definition 3-2.
is also admissible. Ott

In words, for time-i
the waveforms of excita
will always be the same
For linear time-invariaJ
are independent of time
_ _
_ _ ,_u__ ,
,_""
,_ _ ,,
----,---_..- ..- ..
_-----~_.
__
._~------_._---_._--_.~--~.,~---~-----.-
89
x2 yfro. col}
(3-30)
IOnlinear.
relationship in (3-30) is
roperty of homogeneity.
'inciple of superposition.
not only at the output

ial state as well as non
:ringent than the one in
lar according to Defini
~or example, the system
1ge across the capacitor
, no matter what input
t the input and output
lisregarded. Hence the
according to Definition
inear capacitor e and a
ies connection with the
ut y. Hence the system
not linear according to
If IX I =0: 2
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:
x(t) = A(t)x(t) + B(t)u(t)

y(t) =C(t)x(t)
+ 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 )
and any real number IX, the pair

{Q"x(t o), Q"u[to. col} -+ {Q"xto. co)' Q"y[co. co)}
In
2H
3F
T
1
------(b)
ling lo Definilion 3-2.
is also admissible. Otherwise, the system is said to be time varying.
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:
x(t) = Ax(t) +Bu(t)
(3-33a)
y(t) = Cx(t) + Eu(t)
(3-33b)
90
where A, B, e and E are, respectively, n x n, n x p, q x n, and q x p real constant

matrices. This set of equations is called a linear time-invariant n-dimensional
dynamical equation and is denoted by FE (fixed equation). For linear time
invariant systems, the responses are independent of the initial time; hence it is
always assumed without loss of generality that t o =0. The time interval of
interest then becomes [0, ro).
The state space L of E or FE is an n-dimensional real vector space (IR", IR),
Hence we can think of the n x n matrix A as a linear operator which maps L
into L. As mentioned in the preceding chapter, it is very convenient to intro
duce the set of the orthonormal vectors {01' 2 " " , 0n}, where 0i is an n x 1
column vector with 1 at its ith component, and zero elsewhere, as the basis of
the state space. In doing so, we may also think of the matrix A as representing a
linear operator with respect to this orthonormal basis. Hence, unless other
wise stated, the basis of the state space of E nr FE is assumed to be the set of the
orthonormal vectors
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.
si(s) -x o = AX(s) +BU(s)
(3-34a)
y(s) =CX(s) +Eii(s)
(3-34b)
(sI - A) always exists.

at t =O-then (3-35b)
A comparison of this
Hence if a linear tirr.

(;(s) and the dynamil
related by (3-36). We
Every entry ofthe adj(

the degree of the deter
rational matrix. If E
rational matrix. Note
Analog and digit

equations. As willl
number of state variab:

ical equations. We sh(
dyriamical equation ca .
connecting integrators
grator, 8 summer, and
puter, and their functio
block diagram of analo
time-invariant dynamil
where the circumflex over a variable denotes the Laplace transform of the
same variable; for example,
From (3-34), we have

B
x(s) = (sI - A)-1 XO + (sI - A)-1Bu(s)
(3-35a)
y(s) =C(sI -A)-1 xo +C(sI -A)-lBu(s) + Eu(s)
(3-35b)
In practice, pure dilTerentia

noises. On the other hand.
They are algebraic equations. Ifx o and u are known, x(s) and y(s) can be com
puted from (3-35). Note that the determinant of (sI - A) is different from zero
(the zero ofthe field of rational functions of S)7; htmce, the inverse ofthe matrix
<al
7
That is, the determinant of (sI - A) is not identically equal to zero.

for sorne s is permitted.
Note that det (si - Al = O
Figure 3-9 Analog com

or attenuator.
~-'------'------'''-----,--------------
THE STATE-VAR1ABLE DESCR1PTlON
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
l vector space (IR", Iln

lerator which maps L
y convenient to intro
, where ni is an n x 1
:where, as the basis of
rix A as representing a
Hence, unless other
/'led to be the set 01 the
e-invariant dynamical
Taking the Laplace
(3-34a)
(3-36)
Hence if a linear time-invariant system is described by the transfer matrix

G(s) and the dynamical equation {A, B, C, E}, the two descriptions must be
related by (3-36). We write (3-36) as
A
G(s) = d
1
(1
et s -A
) C[,Adj (sI - A)]B + E
Every entry of the adjoint of (si - A) is a polynomial of degree strictly less than
the degree ofthe determinant of(sI - A); hence C(sI -A)-lB is a strictly proper
rational matrix. If E is a nonzero matrix, then C(sl - A)-l B + E is a proper
rational matrix. Note that we have
I can be solved, given
direction of negative
positive time. In the
future responses, the
Hence the dynamical
91
G(oo) =E
(3-37)
Analog and digital computer simulations of linear dynamical

equations. As will be illustrated in the next section, systems that have a finite
number of state variables can always be described by finite-dimensional dynam
ical equations. We show in this subsection that every finite-dimensionallinear
dynamical equation can be readily simulated on an analog computer by ihter
connecting integrators, summers, and amplifiers (or attenuators). The inte
grator,8 summer, and amplifier are three basic components of an analog com
puter, and their functions are illustrated in Figure 3-9. We give in Figure 3-10 a
block diagram of analog computer connections ofthefollowing two-dimensional
time-invariant dynamical equation
(3-34b)
aa22_-I[xX2(t)(t)] +[bb
12
lace transform of the
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
;) and y(s) can be com

) is different from zero
e inverse of the matrix
(a)
Note that det (sI - A) = O
(b)
(e)
Figure 3-9 Analog eomputer components. (a) Integrator. (b) Summer. (e) Amplifier
or attenuator.
---- -------_._-_.
92
can be easily preparel

directly from a dynam
We give a simple e
sider the dynamical eq
U
Xl
Find the output y and

condition Xl(O)= 1, X2('
The CSMP input s
DYNAMIC
PARAMETER
e21I-----------J
'-----------~
'-------------+1
Figure 3-10
e22 1---
------'
Block diagram oftwo-dimensional dynamical equation.
Note that for a two-dimensional dynamical equation, we need two integrators.

The output 01 every integrator ean be assigned as a state variable. We see that
even for a two-dimensional dynamical equation, the wiring of the block diagram
is complicated; hence, for the general case, we usually use a matrix block
diagram. The matrix block diagram of the dynamical equation E is shown in
Figure 3-11. If E is n-dimensional, the integration block in Figure 3-11 consists
of n integrators. The matrix E represents the direet transmission part from the
input u to the output y. Ifthe matrix A is a zero matrix, then there is no "feed
back" in the block diagram. For a discussion of analog computer simulations,
see, e.g., Reference S46.
Dynamical equations can be readily simulated on a digital computer.
There are many specialized subroutines for solving dynamical equations, such
as MIDAS (Modified Integration Digital Analog Simulator), MIMIC (an
improved version of MIDAS), CSMP (Continuous System M odeling Program),
TELSIM (Teletype Simulator), and others; see Reference 26. These programs
Figure 3-11
Matrix block diagram of the dynamical equation E.
X1DT=
X2DT=
X1 =IN
X2=IN
Y=X1
TIMER DELT,
PRTPLT X1,Y
STOP
END
The first part of the

step size. "FINTIM" i
interval in which the n
result, DELT is usually
print out every comput
larger than DELT. In
DELT and OUTDEL.
as well as plot the outp
Comparisons of ana
tions are in order. On
are limited to a range,
range, sorne componen
of the sim ulation will t
equations must be pro]
carried out by cut and t
of numbers which a dig
magnitude scaling genel
is often limited to 0.1 1
precision of eight or m
computer simulation is
putero The generation
93
can be easily prepared from the block diagram of a dynamical equation or

directly from a dynamical equation.
We give a simple example of the application of System/360 CSMP. Con
sider the dynamical equation
[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.
e need two integrators.

. variable. We see that
Ilg ofthe blockdiagram
ly use a matrix block
;:quation E is shown in
in Figure 3-11 consists
nsmission part from the
, then there is no "feed
computer simulations,
n a digital computer.
amical equations, such
mulator), MIMIC (an
:m Modeling Program),
~e 26. These programs
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
The first part ofthe program is self-explanatory. "DELT' is the integration

step size. "FINTIM" is the final time of computation. "OUTDEL" is the
interval in which the responses will be printed. In order to have an accurate
result, DELT is usually chosen to be very smal!. It is, however, unnecessary to
print out every computed result; therefore the printout interval is chosen much
larger than DELT. In employing CSMP the user has to decide the sizes of
DELT and OUTDEL. For this program we have asked the computer to print
as well as plot the output y and the state variable X l'
Comparisons of analog computer simulations and digital computer simula
tions are in order. On analog computer simulations, the magnitudes of signals
are limited to a range, typically 10 volts. Ir the magnitudes go over the
range, sorne components of the analog computer will saturate and the result
of the simulation will be erroneous. Hence on analog computer simulations,
equations must be properly scaled. This is a difficult problem and usually is
carried out by cut and try. On digital computer simulations, because the range
of numbers which a digital computer can manage is very large, the problem of
magnitude scaling generally does not arise. The accuracy of an analog computer
is often limited to 0.1 percent of its full scale; a digital computer may have a
predsion of eght or more decimal digits. Therefore the result from a digital
computer simulation is much moreaccurate than hat from an analog conb
puter. The generation of nonlinear functions is easier on a digital computer
94
than on an analog computer. However, the interaction between an analog

computer and the user is very good. By this we mean that the parameters of a
simulation can be easily adjusted, and the consequence of the adjustments can
be immediately observed. The interaction between a large-frame general
purpose digital computer and the user is generally not very satisfactory, but
this interaction has been improving in recent years with the introduction of
time sharing and remote terminals. With the increasing speed of computation
and graphic display, the interaction between the user and a special-purpose
digital computer or a general-purpose mid-frame digital computer is now very
good. Furthermore, digital computers are much more versatile; they can be
used to carry out the designo Hence digital computer simulations are gaining
more popularity in recent years.
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
fricUon is clearly not a l

we neglect the static a
friction. Let k be thl
f isgiven by f=kdy/,
3-14. It is a nonlinear
(y, yz) as shown, then I
constant. Hence the lin
and simplification.
Now we shal1 de vele
external force u (input)
Newton's law yields
EXAMPLES
>n between an analog

Lat the parameters of a
)f the adjustments can
l large-frame generalvery satisfactory, but
th the introduction of
speed of computation
and a special-purpose
computer is now very
versatile; they can be
imulations are gaining
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
how the input-output

ystems are developed.
lonents. For example,
:nts: one resistor, one
system is nonlinear or
me varying. Hence in
ical system, every com
invariant element.
ime invariant. A tele
annot function forever;
aging or other factors.
1 in the time interval of
n. be considered as time
ber of physical systems
hat for any admissible
xx[O, 00)' IXY[o. oo)} for any
:al system, ifthe applied
)lit or saturate. Hence
3-2 or 3-7. However,
ysical systems. This is
Iperate only in a certain
1 systems can often be
~d by linearlization or
nples.
12. The fdction force

ree distinct parts: static
vn in Figure 3-13. The
(a)
Figure 3-13
(b)
(a) Static and Coulomb friction. (b) Viscous friction.

Force
Break
roo
---+-----,..<::------1e---_ Displacement
Figure 3-14
Characteristic of the spring.
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
MATHEMAT1CAL DESCRIPTIONS OF SYSTEMS
Taking the Laplace transform and assuming the zero initial condition, we obtain
u(s) = (ms 2 +k 1 s +k 2 )y(s)
Hence the input-output description in the frequency domain of the system is
y(s) =
If m = 1, k 1 = 3, k 2
g(t) =
= 2,
ms
lS
k u(s)
2
For example, if the iJ

cart is moved to the
pendulum back to thf
as a model of a space
Let H and V be, rl
the cart to the pendul
linear movements yiel,
then the impulse response of the system is
~ [S2 +~s +2] = ~ [(S ~ 1) - ~2)] = e-t - e- 2t

-1
-1
(S
H=
and the input and output are related by

y(t) =
J>(t - r)u(r) dr
and
We next derive the dynamical-equation description of the system. Let the

displacement and the velocity of the mass m be the state variables; that is,
Xl = y, X2 = y. Then we have
mg- V=
The application of Ne'

yields
m
.
=
[Xl]
or
X2
[0-;; - 1][X1] +[0]

k2
y=[l
O]
k1
m
X2
;;;
[:J
This is a state-variabIe description of the system. By choosing a different set

of state variables, we may obtain a different state-variable description of the
system.
I
These are nonlinear eql

the pendulum at the Vf
small. U nder this ass
cos e= 1. By retainin
terms with e2 , ip, ee, a;
which imply
Example 2
Consider a cart with an inverted pendulum hinged on top of it as shown in

Figure 3-15. For simplicity, the cart and the pendulum are tssumed to me'!'"
in only one plane, and the friction, the mass of the stick, and the gusts of wind
are neglected. The problem is to maintain the pendulum at the vertical position.
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
A cart with an inverted pendulum.
and
gOll
EXAMPLES
11 condition, we obtain
main of the system is
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
)f the system. Let the

tate variables; that is,
d2
.
mg - V = m dt 2 (l cos e) = mi [ - sin e (j - cos e (efJ
The application of Newton's law to the rotational movement of the pendulum

yields
m1 2(j =mgl sin e + VI sin e- HI cos e
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
Mji = u -mji -mle'

ml 2lf =mgle + mgle - (mji + mllJ)1
hoosing a different set

1ble description of the
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,
(M + m)s2 (s) + mls 28(s) = D(s)
(21s 2 - 2g)8(s) + S2 y(S) = O
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
To develop a dynamical equation, we define the state variables as Xl = y,

= y, X3 = and X4 =
From (3-38) and (3-39), we can solve ji and (J as
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 )
The purpose of this problem is to maintain the distance between adjacent

vehicles at a predetermined value ha and to maintain the velocity of each
vehicle as clase as possibie to a desired velocity Va. Define
i=1,2,3
= 1, 2, 3, 4
-k
O
-1
Example 3 (Reference 512)
From (3-46) and (3-44
These equations can b
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
The term kVa is the f(

maintain their velocil
that Yi,i+ (t) and v(t)
we have, for i = 1, 2,3,
u (3-42)
O O O]x
Yi,i+l(t)=y(t)-Yi+l(t)-h a
and
mi
-k
This is the state equati

interested in the diste
therefore the absolute
on what will be consic
developed.
Example 4 (ReferencE
Consider a satel!ite of
altilude of le sateilite
(3-45)
(3-46)
4-c:J U3-c:J " u2-2 UI_c:J
'l"/./##$////#//./#//#//.//#/##//./$#####/###//#///////#/#///##//////////////#/#/////////////////.
-Y4~
Y3~
.
- - - - - - - - - - - - - - - - - . : . . . - - - . - - - - Y2
Figure 3-16
Four vehicles moving in a single lane.
YI-J
Figure 3-17
Satellite il
-" ---
EXAMPLES
lte variables as Xl = y,
m solve ji and (j as
u(t) = ult) - kvo
and
1,2,3,4
'.
-k
1
-k
1
V\t)=-Vi +-u=-v(t) +-(t)
m
m
m
m
These equations can be arranged in matrix form as

(3-42)
VI(t)
-1
YdO
(2M +m)l
der simplification and
Tn in Figure 3-16. Lef

mass and the applied
fficient and be the same
(3-43)
(3-44)
lnce between adjacent

n the velocity of each
fine
2,3
= y(t) - y+ I (l) = v(t) - Vi + I (t)
From (3-46) and (3-44), we have, for i
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)
Consider a satellite of mass m in earth orbit as shown in Figure 3-17. The

aititude of the satellite is specHled by r(e),O(tj, and cPU) as shown. The orblt
(3-45 )
(3-46)
U&
] UI-G
VI
I Salellite (mass m)
.
&
--
.
------ -
7//////'///'//$///'//&//////////,I/.
Figure 3-17
.....................
Sateilite in earth orbit.
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)=
lf the perturbation is '

can be linearized as
u)]
8(t)
3w 2
O
-2w
cP(t)
x(t) =
~(t)
ro
------------
Then the system can be shown to be described by
re 2COS 2cP +r~2 - k/r 2 +u,./m
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
where A =oh(x, u)/ox,

and u o. This is a sixth
lt can be used to descri
the circular orbit rema
The dashed lines in
two uncoupled parts, (
two parts independentl:
Dynamical equatio
where ro and w are related by r6w2 =k =a known physical constant. Once

the satellite reaches this orbit, it will remain in the orbit as long as there are
no disturbances. lf the satellite deviates from the orbit, thrusts must be applied
to push the satellite back to the orbit. Define
'ro
+ cX l(t) -,
t:X2(t)
u(t) =uo(t) +t:(t) =
-:-
(3-50)
O
O
x(t) =xo(t) +t:X:(t) =
- -- -- - -
O O O O: 1
One solution which corresponds to a circular, equatorial orbit is specified by

ro(t)
ro(t) .
--
_8 2 COS cP sin cP -2rcP/r +u.p/mr

and
l O O O: O
O O 1 O:, O
wt +t: X3(t)
w +6: 4 (t)
cXs(t)
cX6(t)
section a systematic prc

cal equations for genera
pendent voltage and Cl
currents and the capac
behavior of the networ
is not necessary to cho
state variables. This ca
lf we assign all the cap,
as shown, then we see tI
[:~~~~~]
t: U3(t)
and
ro +cXl(t)]
y(t)=yo(t) +t:y(t) = wt +t:x~U)
[
cXs(t)
Figure 3-18 Circuits wit:

consists ofinductors only.
- - - - - - - --
------- -
EXAMPLES
" and u,,(t). The state,
[~l)
8(t)
,]
cP(t)
If the perturbation is very small, or equivaIentIy, : is

can be linearized as
,
O
o
1
O
O ,, O
O
,
1
,
O
3w 2
O
O 2wro , O
,
m
,
O
O
O
O
1 ,, O
O
,
-2w
,
x(t) =
O
O
O x(t) + O
O
O
,
,
,
ro
------------------,------
,
l,/m
(3-49a)
/mrcoscjl
, O
I
O
O
O
,
O ,, O
1 ,
O
~,
mro:
U)
(3-51a)
,
----------
,
I
very small, h(x, u) in (3-49)
O ,, O
, 1
O
, mro
I O O O: O 0J
y(t)=
u<f/mr
101
~ __~ __1__ ~_~-~--~
(3-51b)
x(t)
0000:10
(3-49b)
rial orbit is specified by
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)
*Dynamical equations for RlC networks.
ysical constant. Once

)it as long as there are
thrusts must be applied
We introduce in this sub

section a systematic procedure for assigning state variables and writing dynami
cal equations for generallumped linear RLC networks which may contain inde
pendent voltage and current sources. lt is well known that if all the inductor
currents and the capacitor voltages of an RLC networks are known, then the
behavior of the network is uniquely determinable for any input. However, it
is not necessary to choose aH the inductor currents and capacitar voltages as
state variables. This can be seen from the simple circuits shown in Figure 3-18.
lf we assign all the capacitor voltages and inductor currents as state variables
as shown, then we see that Xl (t) = xz(t) for aH t. Clearly there is a redundancy
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.
The branches of the normal tree are

chosen in the order of voltage sources, capacitors, resistors, inductors, and
current sources. Hence, a normal tree consists of aB the voltage sources,
the maximal number of permissible capacitors (those that do not form a
loop), the resistors, and finaBy the minimal number of inductors. UsuaBy
it does not contain any current source.
2. Assign the charges or voltages of the capacitors in the normal tree and the
flux or current of the inductors in the linksas state variables. The voltages
or charges ofthe capacitors in the links and the flux or current ofthe inductors
in the normal tree need not be chosen as state variables.
3. Express the branch variables (branch voltage and current) ofaB the resistors,
the capacitors in the links, and the inductors in the normal tree in terms ofthe
state variables and the inputs by applying the Kirchhoff voltage or current
law to the fundamentalloops or cutsets of these branches.
4. Apply the Kirchhoff voltage or current law to the fundamental loop or
cutset of every branch that is assigned as a state variable.
(voltage
source)
Figure 3-19
A networ
By applying the Kirchl

we have immediately
the voltage across resis
are expressed in terms
to the fundamental cuts
The application of the I<
This application of the }
These three equations ca
The output equation can
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.
Find the input-output d{

equivalently, the transfer
i~puts, one output, the tn
[g(~) 9ds)], where 9
and gds) is the transfer J
network reduces to the Oll<
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
to consist oIonly inde

nts.
: tree, link, and cutset
A tree of a network is
les) containing aH the
rery branch in a given
~ee is caHed a link. A
ranches such that the
ling network to be un
nd some tree branches
.ree branch with some
. Hence every funda
tal cutset includes only
give a systematic pro
1 of any RLC network
11 sources9 :
of the normal tree are
esistors, inductors, and
aH the voltage sources,
)se that do not form a
of inductors. Usually
the normal tree and the
rariables. The voltages
.current ofthe inductors
)les.
[reut) of aH the resistors,
)rmal tree in terms of the
h.hoff voltage or current
.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)'
A network with voltage and current sources.
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
The application of the KCL to the fundamental cutset of yields
This application of the KVL to the fundamental loop of (J yields
X3
+X3 -Xl +X2
=0
These three equations can be rearranged in matrix form as
o
O
-1
-1] [1
1 x
+ O
-1
The output equation can be easily found as

--
.1'"
. . /.,
Example 6
he normal tree is chosen
rce, two capacitors, and
'mal tree and the current
es. Next we express the
ariables and inputs. By
iamental loop of branch
e its current is (u - Xl)'
~apacitors or with a cutset that

.use in this case its dynamical-
Find the input-output description of the network shown in Figure 3-19, or

equivalently, the transfer-function matrix of the network. Since there are two
inputs, one output, the transfer-function matrix is a 1 x 2 matrix. Let G(s) =
[ (s) gds)], where g11 is the tninsfer function from U to y with U2 = O,
and gds) is the transfer function from U2 to y with U =0. With U2 ""'0, the
network reduces to the one in Figure 3-20(a). By loop analysis, we have
(1 +D
L
-~11ls)+
(s)
~.~ 2(S) ~ u(s)
2)A
( l+s+~ 1 2 (s)=0
104
MATHEMATlCAL DESCRPTlONS OF SYSTEMS
1.11
1
S
=
Ca)
Figure 3-20
Solving ror
Cb)
Two reduced networks.
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
Ir U = 0, the network in Figure 3-19 reduces to the one in Figure 3-20(b).

By node analysis,
1
)
( S +s
Figure 3-21
Network w
These can be rearranged

A
V(s) -sV2(s) =0
-sV(s) +(1 +2s)V2(s) -(1 +s)V3(s) =U2(S)

-(1 +s)V2(s) +(2 +s)V;,(s) = -uz{s)
Solving ror V (s), we obtain
Hence
V(s)
u z{s
This nonlinear time-inv

the general case. N ow il
then h(x(t can be appr
work in Figure 3-21(a) Cl
scribed by the linear tim
gds)=~) u,~o
initially
re1axed
Consequently, the input-output description in the rrequency domain or the

network is given by
S2
y(s)= [ S3 +2s 2 +3s +1
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
..
X2(t) = Cx(t) +i(t) = Cx(t) +h(x(t))

LX2(t) E - RX2(t) - x (t)
Now ir x(t) is known

duce. the variables x (t) ==
as h(x (t)) = i o - x (t)/R 2.
[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)
he one in Figure 3-20(b).
Network with a tunnel diode.
These can be rearranged as

=0
0=
0=
. ) - h(x 1(t))
x 1 (t =
e
uz(s)
uz(s)
(3-52)
. () -X1(t)-Rxz(t)
Xz t =
L
+I
This nonlinear time-invariant dynamical equation describes the network for

the general case. Now ir x 1 (t) is known to operate only inside the range (a, b),
then h(x 1(t)) can be approximated as h(x 1(t))=x 1(t)jR 1. In this case, the net
work in Figure 3-21(a) can be reduced to the one in Figure 3-21(b) and is de
scribed by the linear time-invariant state equation
~(s)
3s
Xz(t)
+c
+1
freq uency domain of the
T is a tunnel diode with

the capacitor and Xz
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
MATHEMATlCAL DFSCRIPTIONS OF SYSTEMS
where E = E - Vo - Rio. This linear equation is applicable only if v(t) is limited

inside the range (e, d). An equivalent network of this linear equation is shown
in Figure 3-21(c).
I
In
u
3-5 Comparisons of the Input-Output Description

and the State Variable Description
We compare in this section the input-output description and the state-variable
description of systems.
1. The input-output description of a system describes only the relationship
between the input and the output under the assumption that the system is
initial1y relaxed. Hence, if the system is not initial1y relaxed, the description
is not applicable. A more serious problem is that it does not reveal what
will happen ifthe system is not initiaHy relaxed nor does it reveal the behavior
inside the system. For example, consider the networks shown in Figures 3-22
and 3-23. The networks are assumed to have a capacitor with -1 farad. Ir
the initial voltage of the capacitor is zero, the current will distribute equally
among the two branches. Hence the transfer function of the network in
Figure 3-22 is 0.5; the transfer function of the network in Figure 3-23 is 1.
Because of the negative capacitance, if the initial voltage across the capacitor
is difIerent from zero, then the voltage Yl in Figure 3-22 will increase with
time, whereas the voltage Y2 in Figure 3-23 remains equal to u(t) for aH t.
It is clear that the network in Figure 3-22 is not satisfactory, because the
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
cannot be detected from its transfer function. Consequently, the input-output
description sometimes does not characterize a system completely.
The dynamical-equation descriptions of the networks in Figures 3-22

and 3-23 (see Problem 3-18) can be found, respectively, as
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
For the time-invaria
In
In
-IF
In
A network with linear, time-varying elements.
Thus the impulse re~

(3-55) or (3-56). See
For linear time
functions or impuls<
frequency s = jw can
COMPAR1SONS OF THE lNPUT-OUTPUT DESCR1PTlON
lle only if v(t) is limited

lear equation is shown
1n
scription
and the state-variab1e
only the reltionship

lion that the system is
elaxed, the description
: does not reveal what
:s it reveal the behavior
; shown in Figures 3-22
:itor with - 1 farad. Ir
will distribute equally
ion of the network in
,rk in Figure 3-23 is 1.
ge across the capacitor
3-22 will increase with
equal to u(t) for all t.
tisfactory, because the
dition is different from
lre 3-23 behaves well,
ges in branches 1 and 2
letwork wil\ eventually
0-23 can never function
s not known, this fact
lently, the input-output
n completely.
works in Figures 3-22
Iy, as
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
A dynamical equation describes not only the relationship between the

input and output but also the behavior inside a system under any initial
condition; hence a dynamical equation characterizes a system completely.
2. For ~xtremely complicated linear systems, it might be very involved to find
the dynamical-equation descriptions. In these cases, it might be easier to
find the input-output descriptions by direct measurements. We apply at
each input terminal a very sharp and narrow pulse; the responses of the
output terminals give us immediately the impulse-response matrix of the
system. In practice, we may have difficulty in generating a very sharp and
narrow pulse, or ideally a b-function; however, this can be avoided by using a
.
unit step function. A unit step function b 1 (t -' t o) is defined as
{l
U1(t-t O)= O
for t to
fort<t
o
( 5 )
3- 3
The response of a linear causal system due to a unit-step-function input is

given by
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 )
For the time-invariant case, (3-55) reduces to
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
quency response analyzers. After measuring gUw) at a number of fre

quencies, we can then find a g(s) to match the measured gUw). This method
of determining transfer functions is often used in practice.
3. Prior to 1960, the design of control systems had been mostiy carried out by
using transfer functions. However, the design had been limited to the single
variable case; its extension to the multivariable case had not been successful.
The state-variable approach was developed in the 1960s. In this approach,
the formulations in the single- and multivariable cases are the same, and a
number of results were established. These results were not available in the
transfer-function approach at that time; consequentiy, interest in this
approach was renewed in the 1970s. Now the results in the state-variable
approach can also be obtained in the transfer-function approach. It also
appears that the latter approach is simpler in concepts and computations.
4. The dynamical equation can be extended to the time-varying case; the
extension of the transfer function to the time-varying case has not been
successful. In the optimal design, dynamical equations can be used to study
finite terminal time problems; this is not possible by using transfer functions.
5. In rhe study of nonlinear systems, depending on the approach taken, either
description can be used. For example, in the study of the stability problem,
we use the input-output description in the functional analysis and operator
approaches. See References 28, 95, 117, and S79. Ir Lyapunov's second
method is employed, then we must use the dynamical-equation description.
6. Ir the dynamical-equation description of a system is available, the system
.can be readily simulated on an analog or a digital computer.
From the foregoing discussion, we see that the input-output and the state
variable descriptions have their own merits. In order to carry out a design
efficientiy, a designer should make himself familiar with these two mathematical
descriptions. In this book, these two descriptions will be developed equally
and their relationships will be explored.
3-6
where u i and y are 1

of the system Si' 1
the impulse-respon:
3-24 that in the pa
tandem connection,
we have u 1 =u - Y2'
SI and S 2 have com:
cannot be properly
effect in the connect
unchanged after COI
matrix of the paralle
For the tandem e
We prove (3-59) for ti

definition, the respon
response at Y1 due to 1
gl(t,r)is
Mathematical Descriptions of Composite Systems 10
Time-varying case. In engineering, a system is often built by interconnect

ing a number ofsubsystems. Such a system is called a composite system. Whether
or not a system is a composite system depends on what we consider as building
blocks. For example, an analog computer can be looked upon as a system,
or as a composite system that consists of operational amplifiers, function
generators, potentiometers, and others as subsystems. There are many forms
of composite systems; however, mostly they are built from the following three
basic connections: the parallel, the tandem, and the feedback connections, as
shown in Figure 3~24. In this section we shall study the input-output and
state-variable descriptions of these three basic composite systems.
We study first the input-output description of composite systems. Consider
10
two multivariable s:
The material in this section is not used until Chapter 9; thus its study may be postponed.
(al
u
Figure 3.24 Composit.
Tandem connection. (e)
109
MATHEMATICAL DESCRIPTIONS Of COMPOSITE SYSTEMS
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
two multivariable systems Si, which are described by
roo G(t,r)u;("r)dr
Yi(t) =
G(t, r) = G 1(t, r)
+ Gz(t, r)
G(t, r) =
Gz(t, v)G 1(v, r) dv
(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
J: gz(t, v)g I(V, r) dv

y =
Y + Y2
U=U;{~
O-~/
(a)
(b)
u
(e)
Figure 3.24 Composite eonneetions of two systems.

may be postponed.
(3-58)
For the tandem connection shown in Figure 3-24(b), we have
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
Tandem eonneetion. (e) Feedback eonneetion.
(a) Parallel eonneetion.
(b)
110
Hence,
g(t, r) =
gz(t, v)g[(v, r) dv
description is
The same procedure can be used to prove the multivariable case.

For the feedback connection shown in Figure 3-24(c), the impulse-response
matrix is the solution of the integral equation
G(t, r) = G[(t, r) -
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
[A~(t) A~(t)J[:J +[::~:~J u
y = [C[ (t)
Cz(t)]
[:J
+(E[ (t) + Ez(t))u
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
Let G l (s) and Gz(s) b

(not necessarily prope
(3-62a)
(3-62b)
The dynamical equation of the tandem connection of S [ and Sz is given by

(3-63a)
(3-63b)
which can be easily obtainedby observing o[ = U, Y1 = U z, y = y z.

For the feedback connection shown in Figure 3-24(c), its dynamical-equation
Observe that the [

while the matrix on tl
of order m. The eler
since the rational fUn<
be applied.
Proof of "Theprem 3-:
1t is well knowti that (

det (J
MATHEMATICAL DESCRIPTIONS OF COMPOSITE SYSTEMS
111
description is
.le case.
the impulse-response
r) ds dv
(3-64a)
(3-60)
:uation can be easily

~ is a general iterative
l.
lposite systems. Let
(3-61 a)
(3-61 b)
. A, Di, Ci and E are

; function of t defined
o linear spaces. The
[ and ~2' written as
, where x[ is a vector
le sum of those of ~ [
,nd S 2; its state space

~2' For the parallel
lamical equation is
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)
Observe that the matrix on the right-hand side o (3-65) is a q x q matrix,

while the matrix on the left-hand side is a p x p matrix. 1m is the unit matrix
of order m. The elements of these matrices are rational functions of s, and
since the rational functions form a field, standard results in matrix theory can
be applied.
Proof of Theorem 3-2
It is well known that det (NQ) = det N det Q. Hence we have
,Y=Y2'
:s dynamical-equation
det (NQP) = det N det Q det P= det (PQN)
(3-66)
112
MATIlEMATlCAL DESCRPTlONS OF SYSTEMS
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
They are square matrices of order (q + p). lt is easy to verify that

NQP=[I p
O
the feedback system
From Figure 3-24(c),

O
I q +G(s)G 2 (s)
A
:J
and
pQN=ep+(;2~)(;(S)
hence
det (NQP) =det (I q + 91(S)(;~(s))

det(PQN) =det(I p + (;2(s)(;(s))
and the theorem follows immediately from (3-66).

Problem 3-32.
which, together with
For a different proof, see

Q.E. D.
Theorem 3-3
Note that the COI

system to be defined.
meaningless in the se
Equation (3-69).
Example 1
Consider a feedback:
lf det (Iq + (;1 (S)(;2(S)) + 0, then
GI(s)(I p +G 2(S)G I(s))-1 =(I q + G I(S)G 2(sW IGI(S)

Proof
Note that the zero in det(I q +G I(S)G 2 (S))+0 is the zero element in the field
of rational functions. Hence, it can be written more suggestively as
det(Iq +G I(s)G 2 (s))+O for sorne s. The ~ondltion det(Iq +G(s)G 2 (s))+O
implies that the inverse of the matrix (I q +G I(S)G 2 (s)) exists. From Theorem
3-2, we have
lt is easy to verify tha
det(I q +G I(s)G 2 (s))=det(I p +G 2 (S)G 1(S)) +0

Hence, both (Iq +G 1(S)G 2 (s))-1 and (I p +G 2 (S)G 1(s))-1 exist. Consider the
identity
which can be written as
(I q +G 1(s)G 2 (S))G 1(s)(I p +G 2 (S)G 1(S))-1 = G1(s)
(3-67)
Premultiplying (I q + G 1(s)G 2 (S))-1 on both sides of (3-67), we obtain the desired

equality.
Q.E. D.
Knowing Theorem 3-3, we are ready to investigate the transfer-function
matrix of the feedback connection of SI and S2'
Corollary 3-3
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
MATHEMATICAL DESCRIPTIONS Of COMPOSLTE SYSTEMS
order. Let us choose
113
the feedback system is given by
(;(s) = (;l(s)(I p +(;iS)(;l(S))-l =(I q +(;l(s)(;is))-l (;l(S)
(3-68)
Proof
erify that
From Figure 3-24(c), we have (;l(s)(ii(s) - (;2(S)Y(S)) = y(s), or
(I q
+ (;1(S)(;2(S))Y(s) =
(3-69)
(;l(s)ii(s)
Q.E.D.
which, together with Theorem 3-3, implies this corol1ary.
Note that the condition det (I q + (; 1(S)(;2(S)) O is essential for a feedback

system to be defined. Without this condition, a feedback system may become
meaningless in the sense that for certain inputs, there are no outputs satisfying
Equation (3-69).
a different proof, see
Q.E.D.
Example 1
Consider a feedback system with
-s
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
lt is easy to verify that det (1 2
u(s)
Consider the
[1 0J
G 2 (s)= O 1
A
Let us choose
1]
s +2
[~
(s
1)2
Then (3-69) becomes
+
r
(3-67)
, we obtain the desired
(3-70)
s +1
Q.E.D.
Obviously there is no y(s) satisfying (3-70). In matrix theory, Equation (3-70)

is said to be inconsistent. Hence, in the feedback connection, we require
det (I q + (; 1 (s) (;2(S)) O for some s.
: 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 +
Let (;l(S) and (;2(S)

[atrices of Sl and S2,
fer-function matrix of
I
I
ji
114
MATHEMATICAL DESCRPTIO~S OF SYSTEMS
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.
Well-posedness problem. We showed in Corollary 3-3 that if G(s) and

are proger ratipnal fun~tion .matrices and if det (1 + G1(S)G 2 (S))+-0 for
sorne s, then G(s) = G(s)(1 + G 2 (s)G(s)r 1 is well defmed. However, nothing
has been said regarding whether or not G(s) is proper. In this subsection, this
problem will be studied.
1
s
G2 (s)
Example 2
1
s+
Uz
F
+
Consider the feedback system shown in Figure 3-24(c) with
-1
G1(s)= _1_
[
s +1
-JJ
Figure 3-25
s+lJ
The rational function matrix
s
1
-1
s +1
is clearly nonsin~ular in the field of rational functions. The overall transfer

function matrix G(s) is given by
-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
of the signa!. Hence,

although all of its su
design of feedback sy~
proper transfer functi<
rational matrix. In th,
Before proceeding, we
Theorem 3-4
Let M(s) be a square r
where Mp(s) is a poI)

matrix. Then M- i(s)
Proof
We first show that if 1\
s +1
which is not a proper rational function matrix. The block diagram of this
feedback system is plotted in Figure 3-25.
-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
A sufficient condition for

Appendx G. For necess
3-4 of Reference S34.
115
MATHEMATICAL DESCRIPTIONS Of COMPOSITE SYSTEMS
:hat is, the condition

). Hence, a feedback
In without having the
)Ived in the following
u,
-1
G1(s) and
(+ (;1(S)(;2(S)) #=0 for
'1 3-3 that if
:d. However, nothing

[n this subsection, this
+o-~
+
'-
s+T
E
Uz
ith
Figure 3-25
=1
-s + 2
s+1
yz
Feedback system with an improper transfer-function matrix.
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)
The overall transfer
+ Msp(s)
(3-71 )
where Mis) is a polynomial matrix and Msp(s) is a strictly proper rational

matrix. Then M - i(s) is proper if and only if M; 1(S) exists and is proper. 11
s +1
Proof
-s
We first show that if M; 1(S) is proper, so is M- 1(s). We write (3-71) as

M- 1(s) = {Mp(s)[1 +M; 1(s)Msp(S)]} -1 = [1 + M; 1(s)Msp(S)] - 1M; I(S)
block diagram of this
If M; 1(S) is proper, then M; 1((0) is a finite constant matrix. Since Msp(s) is

strictly proper, Msk~)=O. 'Hence we have M-1(00)=M;1(OO), a finite
constant matrix. Thus M -1(s)'is proper.
From M- 1 (s)=[M p(s)+M sj,(s)]-1, we have, as S~OO,M-1(S)~
(Mp(S)+O)-1~[Adj Mp(s)]fdet Mp(s). Hence if Mp(s) is singular, M- 1(oo) is
ignals u(t) is corrupted

,tied by the improper
n(t) =0.01 sin l000t no
L the amplitude of the
at the input terminals,
times larger than that
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.
Hence the transfer mal

Gf(s) =
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
Ip view of Theorem 3-:

G o(00 ) is nonsingular, .
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.
Then we have Go(co)=
Proof
lf I +G 2 (00)G 1(00) is nonsingular, [1 +G 2 (S)G 1(S)]-1 is proper.

product of two proper rational matrices; hence it is proper.
lf G(s) is proper, so are G 2 (s)G(s) and 1- G 2 (s)G(s). Since
G(s) is the
which is proper.
[1 + G2 (s)G 1(s)] -1 = 1- G 2 (S)G 1(s)[1 + Gz(s)G 1(s)] -1 = 1- G 2 (s)G(s)

which c~n be readily verified by postIllultiplying 1+ Gz(s)G 1{s), we ~conclude
that if G(s) is proper, so is [1+G 2 (s)G 1(S)]-I. Hence [1+G 2 (oo)G 1(00)]-1
's:finite and 1+ G 2 (00)G 1(00) is nonsingular. This establishes the theorem.
Q.E.D.
12
For adifferenl proof of this sta temen l. see Problein3A4.
Figure 3-26
A feedbad
117
MATHEMATICAL DESCRIPTIONS OF COMPOSITE SYSTEMS
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.
The condition det [1 + (;1(00)(;2(00)J =det [1 + (;z(oo)(;I(oo)J =/=-0 in this

theorem is the same as the condition det (1 + E 1 E 2 ) =/=- O for the existence of the
dynamical description of the feedback system shown in Figure 3-24(c). This
resol ves the discrepancy between these two descriptions.
Before discussing the physical implications of the condition
det [1 + (;2(00)(;{00)] =0, we study first the composite system shown in
Figure 3-26. This system is a generalization of the system in Figure 3-24(c)
and will be developed in Chapter 7 and then extensively studied in Chapter 9.
The transfer matrices (;(s) in Figure 3-26 are assumed to be proper rational
matrices of appropriate orders. Clearly, we have
depends only on the

tllowing.
e(s) = o(s) - (; 3(S)(;1 (s)e(s) - (;4(S)(;z(S)(; 1 (s)e(s)
which implies
e(s) = [1
Jer if and only if M( 00 )
+ (;3(S)(;I(S) + (;4(S)(;2(S)(;1 (s)] -I U(S)
Hence the transfer matrix from u to y is given by

(;(s) = (;z(S)(;1 (s)[1
M(oo). Hence M-'(s)

Q.E.D.
Ion posed at the begin-
+ (;3(S)(;I(S) + (;4(S)(;2(S)(;, (s)]-'
(;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 =
O, which is singular: however, we have
(;(s) = (;2(s)[1 - 1 + (;2(S)J -1
-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
Jet (;1(S) and (;z(s) be

d 52' Then the overall
s proper.
;:Jer.
Sin ce
(3-72)
=I
which is proper.
y
u
i-----r-r----.,/]
= I - (;2(S)(;(S)
2(S)(; lis), we...c onclude

[1 + G 2(oo )G 1 (e)J - 1
tab\ishes the theorem.
Q.E.D.
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.
from any point to an

establishes the first p,
the polynomial part of
proper if and only if
Theorem 3-6
We discuss now the

and more generally, de
A loop of a block diag
direction of the path b
For example, the feedl
is indicated by the hea'
of all transfer function
In our discussion, we ~
Clear1y if a loop gain i
zero at s = oo. lf aH t
has a nonzero loop ga
loop are exactIy prope
degree of its numerato
only loops with nonze
Consider the feedb:
not well posed becaus
loop with loop gain +
improper transfer func
conc1ude that if a systel
+ 1 at s = 00 which is
function.
The system in FigUJ
Go(00)=1-1 +2#0,
from the existence of 2
loop will offset the 100
The system in Figure 3-26, where G(s) are rational transfer matrices of appropri
ate orders, is weH posed if and only if Gi(s), i = 1, 2, 3,4, are proper and the
rational matrix
~~-
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.
e(s)= ---, [1 + G 3 (S)G 1 (s) + G 4 (s)Gz(s)G 1 (S)]-l (S)
w(s) = - G 3 (S)G 1 (s)[1 + G 3 (S)G 1(s) + G 4 (S)G 2 (S)G 1(s)] -l r (S)

and
(3-74)
y(s) = - Gi(s)G 1 (s)[1 + G 3(s)Gi(s) + G 4 (S)G 2(S)G 1(s)] -l r (S)
We note that in computing these transfer functions, no branch in the system

should be disconnected. Hence we compute c1osed-loop transfer functions.
lf we disconnect any branch, then the system becomes a different system.
'---.J
(3-75)
exists and is proper, or equivalentIy, the constant matrix

Go(oo)= I +G 3 (00)G 1 (00) +G 4 (00)G 2 (00)G 1(00)
is nonsingular.
(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
MATHEMATlCAL DESCRIPTlONS OF COMPOSITE SYSTEMS
singular. Will such a

tr, the transfer matrix
I contains a part which
mplitude of e(t) will be
1m out. Rence, in the
overall transfer matrix
put pairs of the system
fied, and in sorne sense,
preserved.
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)
o branch in the system

10P transfer functions.
dilTerent system.
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)
er if and only if G I(S)

per, the transrer matrix
: onlyto linear time-invariant

teferences S8, S207 and S214.
119
Figure 3-27
Feedback systems which are not well po sed.
120
or an improper transfe
to be well posed.
}--E..-_ _---+I 2s + 1
-s
3-7
(b)
(a)
Discrete-TiI
The inputs and outpul

defined for all t in ( - e
invariant case. Theya
study a different class
outputs of discrete-tirr
For example, a digital
of variables at discrete
continuous-time syster
responses are of intere:
For convenience, ti
output appear will be ;
time interval T is cal
{y(k) ~ y(kT)}, k = O, j
In the following, we di
the continuous-time S)
called a relaxed discre,
discrete-time system sa
I
(d)
(e)
Figure 3-28 (a) A well-posed system.

systems of (a).
(b) An ill-posed system.
(e), (d) Equivalent
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
where g(k, m) is called t

due to the application (
with the system relaxe

that is, the output doe
ylk?
= Q [ol' k < ;Ti. "
relaxed at ka, then (3-7(
lf a linear causal relaxe

g(k, m) = g(k - m) for al]
ka = Oand the set of tim

linear time-invariant cal
y(k) =
14
This section maybe skippe<
D/SCRETE-TIME SYSTEMS
121
or an improper transfer function. In practice, every system should be designed

to be well posed.
3-7
(b)
(d)
temo
(e), (d) Equivalent
lonzero loop gains at

)wn in Figure 3-28(a),
. net loop gain equal to
in Figure 3-28(a), the
is different fram + 1;
~m is wel1 posed. The
) gain 3 - 2 = 1; hence
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) =
;: consider the system

through the point H.
L loop between points
ned loop, the transfer
ause of the two strictly
Benee, the combined
gain 1 at s = ro. This
combined loop which
transfer function from
;:r. F or the system in
le system has no com
,lent to an infinite gain
(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) =
; a combined loop with

nt to a gain of infinity
,es the problem is the
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
lf a linear causal relaxed discrete-time system is time invariant, then we have

g(k, m) = g(k - m) for all k ~m. In this case, the initial time is ehosen to be
k o = and the set of time of interest is the set of positive integers. Bence for a
linear time-invariant causal relaxed discrete-timesystem, we have
y(k) =
g(k - m)u(m)
m~O
14
This section may be skipped withoutloss in continuity.
k =0, 1, 2, ...
(3-78)
122
Comparing with continuous-time systems, we see that in discrete-time

systems, we use summation instead of integration; otherwise, all the concepts
are the same. In the continuous-time case, if we apply the Laplace transform,
a convolution integral can be transformed into an algebraic equation. We
have the same situation here, the transformation which will be used is cal1ed
the z-transform.
Definition 3-10
The z-transform of the sequence {u(k), k =0,1,2, ...} is defined as
omitted. We introd
time-varying, discrete
where x is the state VI

time dynamical equa
set of first-order diffe
If A(k), B(k), C(k),
00
u(z) ~ Q[u(k)] ~
U(k)Z-k
k=O
where z is a complex variable.
Example 1
which is called a fin

x(z) be the z-transforr
lf u(k) = 1 for k = O, 1,2, ... , then

~
u(z) =
L.,
k=O
z- =
That is,
1
z
=-
1-z- 1 z-l
Example 2
Then
lf u(k) =e- 2k for k =0,1,2, ... , then

00
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
Hence the application
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)
g(k - m)u(m)z-(k-lII l z -111
y(z) =
L L
g(k - m)z-(k-lIIlu(m)z-m
00
k=O
00
They are algebraic eql

the same as (3-35) and
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
is called the sampled ,

identical to the coritin
Similar to the co
practical interest are p
----------,-----
123
OISCRETE-TIME SYSTEMS
that in diserete-time
rwise, aH the concepts
~he Laplace transform,
;ebraie equation. We
will be used is caBed
~fined
omitted. We introduce now the diserete-time dynamical equation. A linear,

time-varying, discrete-time dynamical equation is defined as
DE:
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)
which is caHed a linear, time-invariant, discrete-time dynamical equation. Let

x(z) be the z-transform of
{x(k)}k'=o
That is,
<X)
x(z) ~ .?l'[x(k)J ~
X(k)Z-k
k=O
Then
<X)
.?l'[x(k+ l)J =
I
x(k +l)z-k=z x(k +l)z-<k+l)

k=O
k=O
=Z[
causal, relaxed system,
(3-78) as
<X)
k=~
1 x(k
+ l)z-(k+ 1) -X(O)] = z[x(z) -x(O)J
Hence the application of the z-transform to (3-80) yields

zx(z) - zx o = Ax(z) + Bu(z)
Hz) =Cx(z) + Eu(z)
where x(O) =
Xo
Equation (3-81)
CHD
(3-81 )
be arranged as
x(z) = (zI - Af 1 zx o + (zI - Af 1 Bu(z)
(3-82)
y(z) = C(zI - A)- 1 Zxo + C(zI - A)-l Bo(z) + Eu(z)
(3-83)
They are algebraic equations. The manipulation of these equations is exaetly

the same as (3-35) and (3-36). Ifx(O) = 0, then (3-83) reduces to
(3-79)
where
thefact that g(k - m) =
le weighting sequence
'd tri:msfer function.
ngle-variable discrete
'd and its discussion is
y(z) = [C(zI -A)-lB +EJu(z)~ G(z)o(z)

G(z)=C(zI-A)-lB+E
(3-84)
IS called the sampled transferlunction matrix of (3~80). We see that (3-84) IS

identical to the continuous-time case in (3-36) if z is replaced by s.
Similar to the continuous-timecase,. the sampled transfer functions of
practical interest are proper rational functions. Consider the improper transfer
124
function
A
+2
z-l
Z2
g(z)=--=z+1+3z- +3z- +3z-'+'"
The inverse z-transform of 9(z) is g( -1) = 1, g(O) = 1, g(k) = 3, k = 1, 2, 3, ....

Its impulse sequence is not zero for g(k) 1= Ofor k < O; hence {g(k)} is not causal.
Consequently, an improper rational function does not describe a causal system.
AH physical systems are causal; hence we are interested in only proper sampled
transfer functions. This reason is different from the continuous-time case
where differentiations due to improper transfer functions are to be avoided
because of high-frequency noises.
The topics discussed in Sections 3-5 and 3-6 are directly applicable to the
discrete-time case without any modification.
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
In this chapter we have developed systematically the input-output description

and the state-variable description of linear systems. Although the input
output description can be obtained by analysis, it can also be developed from
the measurements at the input and output terminals without knowing the
internal structure ofthe system. For linear, time-invariant systems, the transfer
function description can also be used. Whenever transfer functions (including
impedances and admittances in network theory) are used, the systems are
implicitly assumed to be relaxed at t =0.
The condition for a set of first-order differential equations x =f(x, u, t),
y =g(x, u, t)-in particular, x = A(t)x +B(t)u, y =C(t)x + E(t)u-to be qualified
as a dynamical equation is that for any initial state X o and any input u[to.co)'
there is a unique solution Y[to. co) satisfying the equations. This uniqueness
condition is essential in the study of solutions of a dynamical equation. The
dynamical equations studied in this book are of the form x = A(t)x + B(t)u,
y =C(t)x +E(t)u. They can be extended to include derivatives ofu in the output
equation, such as
3-3 The impulse respol

and T. Is this system e:
3-4 The impulse respor
where w and t o are const:

ideallow-pass filter in tho
3-5
Consider a relaxed
for any u, where C( is a fix

ehops off the input after t
3-6 Let P a be the trunca
relaxed system y = Hu. w(
yU) =C(t)x(t) +E(t)u +EU)uU) +E 2 U)U) + ...
However, these extensions are of limited interest in practice.

One may wonder why we study a set of first-order differential equations
instead of studying higher-order differential equations. The reasons are as
follows: (1) every higher-order differential equation can be written as a set of
first-order differential equations, (2) the notations used to describe first-order
equations are compact and very simple, and (3) first-order differential equations
can be readily simulated on an analog or a digital computer.
In addition to transfer functions and dynamical equations, one may' also
encounter the following description .
.
P(s);(s) = Q(s)i(s)
y(s)=R(sj~(s) +W(s)u(s)
This faet is often used in

y
---,.......':----.
(a)
Figure
P3~1
II
PROBLEMS
7(k) = 3, k = 1,2,3, ....

lce {g(k)} is not causal.
scribe a causal system.
1 only proper sampled
continuous-time case
IllS are to be avoided
~ctly
applicable to the
125
where P, Q, R and W are polynomial matrices, in the analyses. This description

is called the polynomial matrix system description and will be studied in Chapter 6.
Both the input-output description and the state-variable description studied
in this chapter are useful in practice. Which description should be used depends
on the problem, on the data available, and on the question asked.
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
t and r. Is this system causal? Is it time invariant?
3-4 The impulse response of an ideallow-pass filter is given by

sin 2w(t - to)
g(t)=2w---
2w(t - to)
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.
This fact is often used in stability studies of nonlinear feedback systems.
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
MATHEMATICAL DF5CRIPTIONS OF SYSTEMS
In Problem 3-6, is it true that
3-14
The input u and t
(P.y)(t) = (P.Hu)(t) = (H p.u)(t)
for al! t in (- 00, oo)? Ir not, find the interval of time in which the equation holds.
What is the transfer fune
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?
3-15 Consider a multi'

the ijth element of (;(s) e
a. lfu 3 =u +uz, then Y3 = Y +Yz

b. lfU3=1(u +uz), thenY3 =1(Y +Yz).
c. lfu 3 =u -uz, then Y3 = Y - Yz
where Yi is the ith compe

3-16
Consider a multiv
Which are true if x(O) = O? (Answers: No, yes, no, for x(O) 1= O; all yes, if x(O) = O.)
3-9 Show that if H(u + u z ) = Hu
number rx and any u.
+ Hu z for any u, u z, then Hrxu =rxHu for any rational
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?
where the N;]s and D;]s:

of the system?
3-10
3-17 Find the dynamic

lf 8,8 , and 8 z are very srr
functions and dynamical
3-11 The causality of a relaxed system may also be defined as follows: A relaxed system
is causal if and only if u(t) = uz(t) for all 15,10 implies (Hu)(t) = (Huz)(t) for aH 15,1 0 ,
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.]
3-19 Find the dynamic:

network in Figure P3-19.
3-13
Consider a relaxed system that is described by

y(t)= Ig(t-L)U(L)dL
Find the dynamic

and 3-23.
\
,,(,,~
[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
The input u and the output y of a system is described by

ji+2i+3y=2+u
e equation holds.
'ee experiments are per
t 2': O. In each case, the
,d outputs are Yl(t), yz(t),
O) 1= O?
What is the transfer function of the system?

3-15 Consider a multivariable system that is describable by y(s) = G(s)I(s). Show that
the jth element of G(s) can be defmed as
A
9 ij\S) =
If[Yi(t)]
ro-[
()J
.;L
where Yi is the ith component of y and

3-16
= rxH u for any rational
igure 3-5 is a linear time

ansfer function? Is this
iniLially relaxed
Consider a multivariable system whose inputs and outputs are described by
LIl yes, if x(O) = O.)

u
Uj
t and u,~Ofork-l=j
is the jth component of u.
Uj
N 11 (P)Yl(t)
N ZI (p )Yl (t)
+ N 12(p)Yz(t) = D 11 (p)u 1(t) + D z(p )uz(t)
+ N zz(p)Yz(t) = D 21 (p)u 1(t) + Dzz(p)uz(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
Jl!ows: A relaxed system
)= (HuzXt) for al! 1:::; too

:sa.
3-18 Find the dynamical-equation descriptions of the networks shown in Figures 3-22
and 3-23.
't, Y = t -'t, then g(t, 't) =
3-19
we may conclude that if

)ll
Find the dynamical-equation description and the transfer-function matrix of the

network in Figure P3-19.
t - 't.]
: output due to the input

12
(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
where gl(l, t o ) is the step r

[Hint: This can be proved
(3-55).J
3-23 Find the transfer fu

in Figure P3-23. Do you
3-24 Show that single-va
in the sense that the order
true for the time-varying s
3-25 The impulse and Stl
relaxed system are, by del
function and i5 I is a step I

using the property given in
h~
:C--~----
3-26 Verify that the impl

given by Equation (3-60).
~12~11r
3-27
Verify the identity
Figure P3-20
where El is a q x p matrix.
3-28
Verify (3-64) by usin
3-29
Show that
Th rus! = kli
det(I"
Lun"ar surlce
Figure P3-21
Note that the matrix in the

rank 1, show that det(I n +
elements of G. [lf G is not
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
the moon can be modeled

ated is proportional to r,
;ribed by mji = - klil- mg,
the state variables of the
:al-equation description of
le to the input u(t)O(t - to)
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
Show that single-variable, linear, time-invariant, causal, relaxed systems commute

in the sense that the order of the tandem connection of two systems is irnmaterial. ls this
true for the time-varying systerns?
3-24
The impulse and step responses of a single-variable, linear tirne-invariant, causal,

relaxed systern are, by deftntion, given by 9 =Hb(t) and 9 = Hb(t), where b is a delta
d
fllnction and 15 is a step function. lt can be shown that b(t) =-b(!). Verify (3-56) by
d!
lIsing the property given in Problem 3-24.
3-25
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
Verify the identity

I p - Ez(Iq + EE z )- 1 El = (lp + EzE)-
where E is a q x p matrx and E z is a p x q matrix.

3-28
Verify (3-64) by using the identity in Problem 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
Find the dynamical-equation description of the feedback system in Figure 3-24(c),

where SI and Sz are, respectively, described by
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
A function h(x(t), u((

al h(xl(t), uI(t+
for any real numbers a, a

linear function ofx(t) and [:
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?
for sorne A and B. If h, x,

does the assertion hold? I
3-37 Consider the multiv
following composite lransfe
e(s)J
[ u(s) =
[(1 +
C(s)(
The transfer matrix (

have a different form of Gf(
as in Equation (3-68)?
3-38
3"39 Show that the systen

transfer function in the syst,
Figure P3-34
3-40 Show that the parall

posed if and only if I(s) al:
PROBLEMS
shown in Figure 3-24(c),

Iy,
131
: system in Figure 3-24(c),
Figure P3-35
d
:k system.
Q =detPQ.
Figure P3-37
own in Figure 3-24(c)with
3-36
A function h(x(t), uU)) is said to be a linear function ofxU) and u(t) if and only if
(;( h(x (t), U (t)) + (;(z h(xz(t), uz(t)) = h(;( x (t) + (;(zxz(t), (;( u 1 (t) + (;( zuz(t))
for any real numbers (;(, (;(z, any x(t), xz(t), and any u(t), uz(t). Show that h(x(t), u(t)) is a
linear function ofx(t) and u(t) ifand only ifh is of the form
h(x(e), u(e)) = A(e)x(e)
erall transfer functions?

\ the system shown in Fig.
+ 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:
e(s)] [(1 + (;(s)C(s)t

[ u(s) = C(s)(1 + (;(s)C(s))- 1
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
(;I(s) = (;(s)(I + K(;(s)r 1

is proper for almost aH p x ~ const~nt matrix K (or, in mathematical terminology, for a
generic K). [Hint: Express G(s) as G(s) = N(s)D-1(s), where N(s) and D(s) are polynomial
matrices, and use Theorem G-9. See Reference SI96.]
3-44
Let M(s) be a nonsingular rational matrix decomposed uniquely as
4-1
IntroductioDl
M(s) = Mp(s) + Msp(s)

where Mp(s) is a polynomial matrix and Msp(s) is a strictly proper rational matrix. Show
that if Mp(s) is singular, M-1(s) is not proper. [Bint: Write Msp(s)=N(s)D-1(s)=
(N(s)U(s))(D(s)U(s)r 1, where N(s), D(s), and U(s) are polynomial matrices, and write
M-1(s) = (D(s)U(s))(Mp(s)D(s)U(s) + N(s)U(S))-1
If Mp(s) is singular, so is Mp(s)D(s). Consequently, there exists a nonsingular U(s) such
that at least one column, say the jth column, of Mp(s)D(s)U(s) is a zero column. Then
apply Theorem G-9.]
Linear systems can be e

output description and
are obtained, the next
of analyses: qualitative
interested in the gener
stability, of the equati<
the quantitative analysl
due to sorne excitations
can be used to carry Ol
hO\iVeVer are not in eJe,::;
from these solutions. I

of the input-output de~
tionships between these
be discussed are applic:
descriptions are first co
If the irnpulse-respc
input, the output"y can
by direct compuation c
the equation Y(s) = G(s
matrix G(s) is a rationa
lS-1+C 2S- 2 + ....
Show
,(1 +CoG o) is nonsingular.
rem 3-5.
stant matrix and improper
:ments valid?
j.
Let G(s) be 2 x 2 proper
Linear Dynamical Equations

and Impulse-Response
Matrices
,er. Show that the rational
~matical terminology, for a

i(s) and D(s) are polynomial
uniquelyas
'per rational matrix. Show

Trite Msp(s)=N(s)D-1(s)=
tI matrices, and write
J(s)) -
sts a nonsingular U(s) such

s) is a zero column. Then
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
by direct computation or by a graphical method. In the time-invariant case,

the equation y(s) = G(s)i(s) can also be used. Unless the transfer-function
matrix G(s) is a rational matrix, general1y it is easier to compute y directly in
133
134
LINEAR DYNAMICAL EQUATlONS AND IMPULSE-RESPONSE MATRICES
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
Hence the set of solut

space of (4-2). We n
el' e 2, ... , en be any li
tions of (4-2) with th
show that W, for i = L
of(4-2) can be written
assertion is proved. \
independent. Suppos.
definition, there exists
4-2 Solutions of a Dynamical Equation

Time-varying case.
Consider
the
n-dimensional
linear
time-varying
dynamical equation
E:
x(t) = A(t)x(t) + B(t)u(t)

y(t) = C(t)x(t) + E(t)u(t)
[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
Note that the O in the

space; therefore, it is r
(4-2)
W2 1
In particular, we have
[WI(t O)
W;
which implies that e j ,

dicts the hYDothesis: h
( - 00, 00).
Let Wbe any solut

linearly independent "
be written as a unique:
as
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
SLUTIONS OF A DYNAMICAL EQUATION
\TRICES
135
Theorem 4-1
~lementary
to compute
e discussed further.
in Section 4-2. Solu
x. <1>(t, o), which is the
The set of aH solutions of x(t) = A(t)x(t) forms an n-dimensional vector space

over the field of real numbers.
Proof
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
lrious methods for the

tion 4-3 the concept of
llent dynamical equa
space. We show that
equivalent linear dy
o establish the theory
lr d ynamical equations
lfY and sufficient con
bY a linear dynamical
r rational matrix. has a
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
109, 114, and 116.
(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
for a11 t in (-00,00)
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
; a unique solution for
ly many possible initial
:ions. This set of solu
arly independent initial

on n. This fact will be
i = 1
is a solution of (4-2) with the initial condition
i= 1
IXiV(tO) =e
136
LINEAR OYNAMICAL EQUATIONS ANO IMPULSE-RESPONSE MATRICES
Proof
Hence, from the uniqueness of the solution, we conclude that
\jiU = "
Before we prove the th
1X\jIk)
*= A(t)x and if \jI(t o) =
i;:;; 1
This completes the proof that the solutions of (4-2) form an n-dimensional
vector space.
Q.E. D.
Definition 4-1
An n x n matrix function 'P is said to be a fundamental matrix of = A(t)x if

and only if the n columns of'P consist of n linearly independent solutions of
= A(t)x.
I
that is, \jI(') == O. It is ob

Again, froro the unique
solution with \jI(t o) =0.
We shall now pro"
that det 'JI(to) = det [~
set of n constant col um
(IR", IR). It follows that
which, together with th
Example 1
Consider the dynamical equation
*=[~ ~Jx
is a solution of
It actually consists of two equations: Xl =0, X2 = tXI. Their solutions are

XI(t) = x (to) and X2(t) = 0.5t 2XI(t O) - 0.5t6X(to) + X2(tO), wbich are obtained by
first solving for x(t) and then substituting x(t) into X2 = tx. Now two linearly
independent solutions \jII = [O 1]' and \jI2 = [2 t 2]' can be easily obtained
by setting to=O, x(to) =0, X2(tO) = 1 and x(to)=2, X2(tO) =0. Hence, the
matrix
= A(t)
This contradicts the aS5

pendent. Hence, we ce
Definition 4-2
Let 'PO be any fundam<

is a fundamental matrix.
Each column of'P, by definition, satisfies the differential equation

hence, it is evident that 'P satisfies the matrix equation
..y =
is said to be the state trl

A(t)x;
A(t)1Jf'
with 'P(t o) = H, where H is sorne nonsingular real constant matrix. Conversely,

if a matrix M satisfies (4-4) and if M(t) is nonsingular for sorne t, then from the
proof of Theorem 4-1 we know tbat all the columns of M are linearly indepen
dent. Hence, the matrix function M qualifies as a fundamental matrix. Thus
we conclude that a matrixfunction"qt is afundamental matrix of* = A(t)x ifand
only if'P satisfies (4-4) and 'P(t) is nOlisingular for some t.
An important property of a fundamental matrix 'PO is thatthe inverse of
'P(t) exists for each t in ( - 00, 00). This fllows from the following theorem.
Theorem 4-2
Every fundamental matrix 'P is nonsingular for al! t in ( - 00, 00).
(t, to)~
The physical meanir

for all t, its inverse is .
immediateiy the follow
matrix:
for any t, to, ti' and t 2 ir

Note that CI>(t, to) is
particular 'P chosen. Ll
*= A(t)x. Since the col
basis vectors, there exist
.matrix P such that 'JI 2 =
of the itb column of 'P 2 1
SOLUTlONS Of A DYNAMICAL EQUATlON
\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
", "
which, together with the fact that

n
IX'!Ik)
i= 1
is a solution ofx =A(t)x, implies that

. Their solutions are
which are obtained by
tXI' Now two linearly
:an be easily obtained
"zCto) =0. Hence, the
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)
is said to be the state transition matrix of x = A(t)x.

.tial equation x
ACt)X;
(4-Ll. )
ant matrix. Conversely,

r sorne t, then from the
VI are linearly indepen
lamental matrix. Thus
latrix ofx = A(t)x if and
~.) is that the inverse of
following theorem.
(- 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)
for any t, to, ti' and t z in ( - 00,(0).

Note that <l>(t, to) is uniquely determined by A(t) and is ndependent of the
particular 'P chosen. Let 'P 1 and 'P z be two different fundamental matrices of
x = A(t)x. Snce the columns of 'P l' as well as the columns of 'P 2' qualify as
basis vectors, there exists, as shown in Section 2-3, a nonsingtilar real constant
matrix P such that 'P z = 'P 1 P. In fact, the ith column of P is thereprese~tation
of the ith column of 'P 1 with respect to the basis that consists of the columns of
... .... ... ..... ==..-'....::c.....-
138
LINEAR DYNAMICAL EQUATIONS AND IMPULSE-RESPONSE MATRICES
'P 1('). By definition, we have

(t, to) = q z(t)(q z(to)) - 1 = q 1(t)PP -1 (q 1 (t O)) = q 1(t)(q l(t O))-1
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 (t, to). From Equation (4-4), it is evident that
(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 (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 (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 (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
where (t, '[) is the s

unique solution of
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
(l, to)=exp [ ( A('[)d'[]
(4-9)
(Problem 4-31). Ir A(t) is a diagonal matrix or a constant matrix, then it meets

the commutative property, and its transition matrix is given by (4-9). For other
special cases, see Problems 4-14 and 4-15. See also References S229 to S231.
From the concept of state transition matrix, the solution of Ji = A(l)x
follows immediately. To be more informative, we use ,p(t; t o, 1<0' u) lo denote
the solution of x = A(l)x at time l due to the initial condition x(lo) =x o. The
fourth argument of cP denotes the fact that u == O. The solution of Ji = A(l)x
with x(lo) =x o is given by
X(l)~ cP(t; lo, x o, O) = (l, lo)xo
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. (l, lo) is a linear
transformation that maps the state X o at lo nto the state x at time l.
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
In other words, (4-10)

We consider again
(4-10) reduces to .
ata J''o f(l, e) de = f(l, e)I,=,
SOLUTlONS OF A DYNAMICAL EQUATION
TRICES
139
Solutions of the dynamical equation E
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.
(4-4), it is evident that
Theorem 4-3
The s01ution of the state equation
x =A(t)x +B(t)u
(4-8)
is given by
x(t)~cP(t;
(4) and (4-8). If A(t)

,mtinuous1y differenti
Is1y differentiable in t,
\:ial in t; see References
form is generaUy very

[U A(t) is a triangular
t of scalar differential
I,tained (Problem 4-1).
l'
to, xo, u)=(t, to)xo +
(t, r)B(r)u(r) dr
(4-10)
to
= (t, to{x o +
(to, r)B(r)u(r) dr]
where (t, r) is the state transition matnx of

unique solution of
at (t, r) = A(t)(t, r)
x=
(4-11 )
A(t)x; or, equiva1ent1y, the
(r,r)=1
Proof
1)
Equation (4-11) is obtained from (4-10) by using (t, r) = (t, t o)<1>(to, r). We
first shw that (4-10) satisfies the state equation by direct substitution 2 :
[bY
el
a (t, to)xo +a
-x(t)
=dt
at
at
(4-9)
matrix, then it meets

ven by (4-9). For other
ences S229 to S231.
solution of x = A(t)x
I(t; to, X o, O) to denote
ldition x(to) =x o. The
s01ution of x= A(t)x
l'
'o
(t, r)B(r)u(r) dr
=A(t)<1>(t, to)xo +(t, t)B(t)u(t) +
= A(t) [ (t,to)x o +
At t
,a
1
-(t, r)B(r)u(r) d!
'o at
(t, r)B(r)u(r) dr] +B(t)u(t)
t o, we have
'0
x(to) = (t o, to)xo -r
(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. (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)
': exists and is continuous.
-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
Ifx o =0, Equation (4-10) reduces to

t
tr (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)
The matrix function
This is a very important property; it says that the response 01a linear state equa
tion can always be decomposed into the zero-state response and the zero-input
response. This is consistent with Equation (3-31).
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
is called the impulse-

the input-output re1a1
We see from (4-H
the solution of the d:
is the unique solution
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
(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 (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 (t, r)B(r)u(r) dr
where A, B, e, and E
respectively. Since th
dynamical equatian E
applied here. We ha,
We give now the solution of the entire dynamical equation E.

Corollary 4-3
The solution of the dynamical equation E in (4-1) is given by

t
y(t) = C(t)<D(t, to)xo +C(t)
~'C(t)<D{t, to{x o +
r $(t, r)B(r)u(r) dr + E(t)u(t)

t
(to, r)B(r)U(r)dr] +E(t)u(t)
I
(4-16)
and e At is nonsingula
In fact, e At is nonsingl
tion matrix of x = Ax
SOLUTIONS OF A DYNAMICAL EQUATlON
rRICES
(4-13)
-input response, and

lation. It is clear that
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)
m the fact that it is a

lS the one in deriving
ition, the solution of

1e input u into small
(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
The matrix function

G(t, r)~ C(t)<II(t, r)B(r)
+ E(t)b(t -
r)
(4-18 )
is called the impulse-response matrix of the dynamical eguation E. It governs

the input-output relation of E if E is initially in the zero state.
We see from (4-16) that if the state transition matrix <II(t, r) is known, then
the solution of the dynamical eguation can be computed. Recall that <II(t, r)
is the unique solution of
ata <II(t, r) = A(t)<1l(t, r), <II( r, r) = 1

Unfortunately, there is in general no simple relation between <II(t, r) and A(t),
and except for very simple cases, state transition matrices <II(t, r) cannot be
easily found. Therefore, Eguations (4-10) and (4-16) are used mainly in, the
theoretical study of linear system theory. If we are required to find the solution
of a dynamical equation due to a given Xo and u, it is unnecessary to compute
<II(t, r). The solution can be easily computed by using existing subroutines on a
digital computer.
Time-invariant case. In this subsection, we study the linear time-invariant
(fixed) dynamical eguation
FE:
x =Ax +Bu
(4-19a)
y =Cx +lElUl
(4-19b)
where A, B, e, and E are n x n, n x p, q x n, and q x p real constant matrices,

respectively. Since the equation FE is a special case of the linear time-varying
dynamical eguation E, all the results derived in the preceding subsection can be
applied here. We have shown in (2-82) that
ation E.
(4-14 )
,ofa linear state equa

lse and the zero-input
1411
by
+ E(t)u(t)
+ E(t)u(t)
(4-16)
and eAI is nnsingular at t = O; hence eA' is a fundamental n1atr~x of x = Ax.

In fact, eA' is non'singular for all t and (eA') - 1 = e- A'; therefore, the state transi
tion matri", .af X =x. is, by the use of (2-81) and (2-79),
<II(t, to)
= eA'(eA'O) -1 = eA(t-' 'o) = <II(t - to)
142
It follows from (4-10) that the solution of (4-19a) is

t
cP(t;to,xo,u)=eA(t-to)xo+
eA(t-<JBu(,)d,
We may apply them

(4-20)
If t o = O, as is usually assumed in the time-invariant equation, then we have the

fol1owing theorem.
Theorem 4-4
The solution ofthe linear time-invariant dynamical equation FE given in (4-19)
is
(4-21 )
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
The impulse response matrix of FE is

G(t, ,) = G(t -e) =CeA(t-<JB + El5(t -e)
or, as more commonly written,

G(t) = CeAtB + El5(t)
(4-23)
The solution of a linear time-invariant dynamical equation can also be

computed in the frequency domain. Taking the Laplace transform of (4-21)
and (4-22), and using g' [e Al ] = (si - At 1 [see Eq uation (2-86)], we obtain
and
x(s) = (si - A)-lX(O) +(sl - A)-l B(s)
(4-24)
y(s) = C(sl - A)-lX(O) +C(sl - A)-l Bn(s) + En(s)
(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.
Taking the inverse

Using Definition 2
Using (si - A)-l =
Using Definition 2
Taking the Laplacl
In addition, there is a
Example 2
We use methods 1 anl
where the circumflex denotes the Laplace transform 01 a variable. 1 nese

equations have been derived in Section 3-3 directly from the dynamical equa
tion. As is defined there, the rational-function matrix
1. (sl-A)-l
(;(s) = C(sl - A)-l B + E
(4-26)
s
-1
S
is cal1ed the transfer-function matrix of the dynamical equation FE. It is the

Laplace transform ofthe impulse-response matrix given in (4-23). The transfer
function matrix governs the zero-state response of the equation FE.
We give now sorne remarks concerriing the computation of e Al . 3 We
introduced in Section 2-7 three methods of computing functions of a matrix.
3See also Reference S159.
=[
(s
=
[
4 A squarematrix
;J
(s +J
is said lo be
-----------------,------~.-.-~~~_
SOLUTIONS OF A DYNAMICAL EQUATION
[RICES
We may apply them to compute

r) dr
(4-20)
:ion, then we have the
ion FE given in (4-19)
r) dr
(4-21 )
eA!:
1. Using Definition 2-16: First, compute the eigenvalues of A; next, find a

polynomial g(A) of degree n - 1 that is equal to eA! on the spectrum of A;
then eA! = g(A).
2. Using the Jordan canonical form of A: Let A = QQ-l; then e At = QitQ-l,
where is of the Jordan formo i t can be obtained immediately by the use of
(2-70).
3. Using the infinite series eA! = L:'=o tkAk/k!: This series will not, generally,
give a closed-form solution and is mainly used on digital computer compu
tation.
We introduce one more method of computing eA'. Since
we have
eAt =
-1(sI - At 1
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
We use methods 1 and 2 to compute (sI - A)-I, where
. 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
=[(; ~~' (, ::)'1
luation FE. lt is the
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)-1 =0:0 -0: 1
(s +1)-2=0: 1
g(A) = [(s + 1)- 1 +(S +
1)- 2J +(S + 1)- 2A
(si - A)-I = g(A) = [(s + 1)- 1 +(s + 1)-2JI +(s + 1)- 2A
and
J(: ~~' (, ::)~

l(S + 1)2
(s
+ 1)2J
Example 3
Figure 4-1
Consider the state equation
The solution is given by

x(t) = !x(O) +
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
know that every elen

2, ... ,m. Hence ev(
We caH tke Ai! a mode
The responses of
mainly by its modes,
a negative real part,
approaches zero eitt
its imaginary part is
mode will approach
zero real part and ha
its index ni is 2 or h
t; -1.
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.
where q and Pi are, I

with A. In the time dI
If Xo s chosen so tha
F or this il1itial state, '

direction of the eigen
formula: similar to (4
eigenvectors of A. 1
discussed.
'RICES
SOLUTIONS Of A DYNAMICAL EQUATlON
145
A. If f(A)6 (5 - A)-I
8
-----!L--+---+--+--+-+--+--1--1--- xl
Figure 4-1
(4-28)
rse Laplace transform
-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.
If all eigenvalues of A are distinct, the response of x = Ax due to x(O) =x o

can be written as, by using Problem 2-44,
'e
')Je -(t -
-(t-r)
rl
x(s) = (sI - A)
A
J U(T) dT
-1"
o = L,
X
for the initial states

at each point of the
1 Qi1ll1X
o
s -- /"'i
--1
where q and p are, respectively, a right and a left eigenvector of A associated

with Aj In the time domain, (4-29) becomes
{ , for i = 1,2, ... , m
; a linear combination
,et . be a lordan-form
l.
From (2-69) we
Ir an eigenvalue has multi

and Definition 2-15.
(4-29)
(4-30)
lf Xo is chosen so that
pXO
=0 for all iexcept i = j, then (4-30) reduces to

x(t) = (p }'o)e Ajt qj
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
Equivalent Dynamical Equations
We introduce in this section the concept of equivalent dynamical equations.

This concept in the time-invariant case is identical to the one of change of basis
introduced in Figure 2-5. Hence, we study first the time-invariant case and then
the time-varying case.
Time-invariant case.
Consider the linear time-invariant (fixed) dynamica\
matrix E is the dire

Since it has nothing 1
lence transformation
We explore noW
Recall that the state
unique relation betv
initiaUy relaxed. n
analysis often lead te
equation
FE:
X =Ax +Bu
y=Cx +Eu
(4-31a)
(4-31b)
where A, B, C, and E are, respectively, n x n, n x p, q x n, and q x p real constant

matrices; u is the p x 1 input vector, y is the q xl output vector, and x is the
n x 1 state vector. The state spaCe L of the dynamical equation is an n-dimen
sional real vector space and the matrix A maps L into itself. We have agreed in
Section 3-3 to choose the orthonormal vectors {n, n z, ... , n,.} as the basis
vectors ofthe state space L, where ni is an n x 1 vector with I at its ith component
and zeros elsewhere. We study now the effect of changing the basis of the state
space. The dynamical equations that result from changing the basis of the state
space are called equvalent dynamical equations. In order to allow a broader
class of equivalent dynamical equations, we shall in this section extend the field
of real numbers to the field of complex numbers and consider the state space as
an n-dimensional complex vector space. This generalization is needed in
order for the dynamical equation FEto have an equivalent lordan-form dy
namical equation.
Definition 4-3
Let P be an n x n nonsingular matrix with coefficients in the field of complex
numbers e, and let x = Px. Then the dynamical equation
FE;
where
A =PAP-
x=x +Bu
(4-32a)
y=Cx+Eu
(4-32b)
e =CP-
B =PB
Consider the networl

inductor X and the v
ables, then the dynan
lf, instead, the loo

the dynamical eq uati
The dynamical eq
describe the same sys
formation between ti
clear that X =x. S
we have xz=(x-x
(4-33)
is said to be equivalent to the dynamical equation FE in (4-31), and P is said to

be an equivalence transformation.
The dynamical equation F E in (4-32) is obtained from (4-31) by the substitu

tion of x = Px. In this substitution we have changed the basis vectorsof the
state space from the orthonormal vectors to the columns ofP- (see Figure 2-5).
Observe that the matrices A and are similar; they are different representations
of the same operator. If we let Q = P - = [q qz ... qn], then the i.th
column of A is therepresentation of Aqi with respect to the basis {q, qz, ., q,.}.
From the equation B == PB or B = P - B = [q qz ... qn]B, we see that
the ith column of B is the representation of the th column of B with respect to
the basis {q, qz, ... , qn}' The matrix is to be computed from CP ~ . The
Example 1
or
Figure 4-2
A networl
"RiCES
EQUlVALENT DYNAMICAL EQUATIONS
iynamical equations.
,ne of change ofbasis
variant case and then
ant (fixed) dynamical

(4-31 a)
(4-31b)
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
the field of complex

I
(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)
f-31), and P is said to
(4-36)
or
4-31) by the substitu

~
basis vectors of the

'p-I (see Figure 2-5).
erent representations
.. qtl], then the th
basis {ql, qz, ... , qtl}.
qtl]B, we see that
1 of B with respect to
ed from Cp-l. The
Xl
Figure 4-2
A network with two different choices of state variables.
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
Consider the two net

capacitor is zero, the
impulse responses are
are zero-state equivale
tions are zero-state eq
any initial voltage in 1
u = O are all identicall~
A necessary condit
they have the same d
network in Figure 4-31
ofthe network in FigUl
Theorem 4-5
Example 3
Two equivalent linear time-invariant dynamical equations are zero-state equi

valent and zero-input equivalent.
The two networks in ]

descriptions are zero-
if there is a nonzero ir
of Figure 4-4(b) is no]
is identically zero.
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)
The impulse-response matrix of FE is

G(t) = CeA1
+ Eb(t)
If the equations FE and FE are equivalent, we have = PAP- 1 ,

C = CP-l, and E =E. Consequently we have
eA' =
and
B = PB,
PeA'P- 1
1n
1
y
u(
CeA'B + Eb(t) = Cp- 1 PeA'p- 1 PB + Eb(t) = CeAtB + Eb(t)
Hence, two equivalent dynamical equations are zero-state equivalent.

The zero-input response of FE is
y(t) = CeA(t - 'Olx(t o)
-_Y-,
(a)
Figure 4-3 Two netw<

zero-input equivalent wit
The zero-input response of FE is

y(t)
= CeA(t -IO)X(tO) = Ce A(' -Iolp - 1x(to)
Hence, for any x(to), if we choose x(t o) = Px(t o), then FE and FE have the same
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)
Figure 4-4 Two netwo
TRICES
and (4-35) are indeed

I
'quivalent if and only if

ame transfer-function
: zera-input equivalent
ists a state in the other
two equations due to
I
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
ns are zero-state equi-
The two networks in Figure 4-4 or, more precisely, their dynamical-equation
descriptions are zero-state equivalent but not zero-input equivalent. Indeed,
if there is a nonzero initial condition in the capacitor, the zero-input response
of Figure 4-4(b) is nonzero, whereas the zero-input response of Figure 4-4(a)
is identically zero.
Ifthe matrix A in a dynamical equation is in the Jordan form, then the equa
tion is said to be a Jordan-formdynamicaJ equation. We have shown in Section
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)
mdFE have the same
Q.E.D.
0.5
0.5
t
I F
equivalence and zero

twa linear dynamical
valent withaut being
zero-state equivalent
(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.
This can be written as
Example 4
This equality holds fOl
E +CBS-l +CABs- 2

This establishes the th,
FE:
(4-37a)
*Time-varying caSE
(4-37b)
lt is easy to verify that if we choose
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:
[;:]= [~--U-.~][~:] [~ 1~][~:J

X3
O 0, 2
X3
varying dynamical eqt

AH the discussion and
well. The only conCf
invariant case are fixe<
time-varying case may
Consider the linear
Definition 4-5 6
Let P() be an n x n rr
and P(t) are nonsingul
dynamical equation
-1-2
[;}[: ~ 4:j
where
We give the following important theorem to concJude this subsection..
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
The dynamical eql

by the substitution of ji
matrix of E. Then we
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
6This definition reduces to O
151
~ATRICES
, a J ordan-form matrix
This can be written as, by using (2-85),
amical equation has an
E +CBs- 1 +CABs- 2 +CA 2 Bs- 3 + ...
E+ CBs - 1 + CABs - 2 + CA 2Bs - 3 +

This equality holds for every s if and only if E = E and
=
...
i =0,1,2, ...
This establishes the theorem.
(4-37a)
(4-37b)
Q.E.D.
*Time-varying case. In this subsection we study equivalent linear time

varying dynamical equations. This is an extension of the time-invariant case.
AH the discussion and interpretation in the preceding subsection applies here as
well. The only conceptual difference is that the basis vectors in the time
invariant case are fixed (independent of time), whereas the basis vectors in the
time-varying case may change with time.
Consider the linear time-varying dynamical equation
-5]
E:
x= A(t)x + B(t)u
y = C(t)x + E(t)u
-3
(4-38a)
(4-38b)
where A, B, C, and E are n x n, n x p, q x n, and q x p matrices whose entries are

continuous functions of t.
: method for finding the

e x = Px into (4-37), we
m dynamical equation:
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)
(P(t)A(t) + P(t))P -l(t)

P(t)B(t)
C(t)P-l(t)
E(t}
is said to be equivalent to the dynamical equation E in (4-38), and

be an equivalence transformation.
,E} and {A, B, C, E}, not
'alent or have the same
(4-40a)
(4-40b)
(4.-40c)
(4-4001)
PO is said to
The dynamical equation E in Equation (4-39) is obtained from (4-38)

by the substitution of x = P(t)x and = P(t)x + P(t)x. Let'JI be a fundamental
matrix of E. Then we claim that
q(t)~ P(t)'JI(t)
(4-41 )
is a fundamental matiix of E. The matrix 'JI is,
i:Jy assumption, a fundamental
6This definition reduces to Definition 4-3 ir P is independent of time.
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)
= CP(t) + P(t)A(t))P -l(t)P(t)'P(t)

= A(t)(P(t)'P(t))
Hence, P(t)'P(t) is a fundamental matrix of
x= Ax.
Theorem 4-7 7
Let Ao be an arbitrary constant matrix. Then the dynamical equation in (4-38)

is equivalent to the one in (4-39) with A(t) = A o.
Proof
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
P(t) = eAo1'P - 1(t)
Clearly P(t) is nonsingular and continuously differentiable for aH t and qualifies

as an equivalence transformation. We compute
A(t) = (P(t)A(t) + P(t))P- 1(t)
= (e A01'P- 1(t)A(t) + AoeAo1'P - l(t) +eAol'iJ-1(t))'P(t)e -
Aot
and for sorne
A(-) is a periodic funct
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
Consider the linear tir
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 .
The matrix function '1

Colum_l1s of 'f'(f) Etild e
solution space; hence, 1
2-3) such that
I::or the nonsingular J

AT
= Q (Problem 2-37
e
ECt) ~=========;l
Define
+ y
We show that
Figure 4-5 Matrix block diagraril 6f the dynamical equation
E in (4-42).
PO is a 1
P(t +T) =e
rRICES
lar for a11 t. Conse

Theorem 2-7). Now
:t)x. Indeed,
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)
ical equation in (4-38)
P(t) is nonsingular for

. 'P- 1(t)'I'(t) =1 yields
- 'P- 1(t)'f(t)'I'-1(t) =
.Because of (4-43).and the boundedness of P(t), P- 1(t) is b.ounded. Because

of P - l(t) = - P -1(t)P(t)P- 1(t) and the boundedness of P(t), P -1(t) is bounded.
Consequently, we can show that if PO is a Lyapunov transformation, so is
P -lO. Clear1y a nonsingular constant matrix is a Lyapunov transformation.
In Definition 4-5, if PO is a Lyapunov transformation, the dynamical equations
are said to be equivalent in the sense of Lyapunov. A Lyapunov transformation
preserves, as will be discussed in Chapter 8, the stability properties of a dy
namical equation, but an equivalence transformation does not. lf PO is
required to be a Lyapunov transformation, Theorem 4-7 does not hold in
general. In other words, not every time-varying dynamical equation can be
equivalent in the sense of Lyapunov to a dynamical equation with a constant A.
This is possible for the class of time-varying equations with periodic ACt) as will
be discussed in the following.
Linear time-varying dynamical equations with periodic A(')
e for aH t and qualifies
Consider the linear time-varying dynamical equation in (4-38). We assume

A(t
Q.E.D.
(4-42)
le in Figure 3-11, there
'fCt +T)=A(t +T)'PCt +T)=A(t)'P(t +T)

The matrix function 'P(t) is nonsingular for al! t; consequent\y, so is 'P(t + T).
Co\umns of ~(t) and columns of qJ(t + T) forrn two sets of basis vectors in the
solution space; hence, there exists a nonsingu\ar constant matrix Q (see Section
2-3) such that
'P(t
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
llly not known, there

Aa is chosen as zero,
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)
We show that PO is a periodic function with period T:

: in (4-42).
PCt +T) = eA{I+ T)'P -1(t +T) = eA1i T e- AT'P-1(t) = P(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:
x(t) = x(t) +P(t)B(t)u(t)

y"(t) = C(t)P -l(t)X(t) + E(t)u(t)
where is a constant matrix, are equivalent in the sense of Lyapunov.
The matrix PO in (4-46) is periodic and nonsingular; hence it is bounded.

Its derivative is clearly continuous and bounded. Hence, PO is a Lyapunov
transformation. The rest of the theorem follows directly fram Theorem 4-7.
The homogeneous part of this theorem is the so-called theory of Floquet.
It states that if x = A(t)x and)f A(t + T) = A(t) for all t, then its fundamental
matrix is ofthe form P- 1 (t)e At , where P- 1 (t) is a periodic function. Further
more, x =A(t)x is equivalent in the sense of Lyapunov to =x.
4-4 Impulse-Response Matrices and Dynamical Equations

Time-varying case.
In this section we shall study the relation between the

impulse-response matrix and the dynamical equation. Let the input-output
description of a system with p input terminals and q output terminals be
y(t) = rtG(t, T)U(T) dT
Jo
(4-47)
where y is the q x 1 output vector, u is the p x 1 input v(ctor, and G is the

q x p impulse-response matrix of the system. We have implicitly assumed in
(4-47) that the system is initially relaxed at too The ijth (ith row, jth column)
element of G(', T) is the response at the ith output terminal due to a b-function
input applied at time T at the jth input terminal. Suppose now that the internal
structure of the same system is accessible, and that analysis of this system
leads to a dynamical equation of the form
x(t) = A(t)x(t) + J3(t)..(t)
y(t) = C(t)x(t) + E(t)u(t)
E:
(4-,fJ.a~)
(4-48b)
where x is the n x 1 state vector of the system, and A, B, e, and E are n x n,

n x p, q x n, and q x p matrices whose entries are continuous functions of t
defined over (- 00, (0). Since Equations (4-47) and (4-48) are two different
descriptions of the same system, they should give the same input-output pairs
ifthe system is initially relaxed. The solution ofthe dynamical equation E with
x(t o) = O is given by
t
y(t) = C(t) r (t, T)B(T)U(T) dr + E(t)u(t)
Jo
t
r [C(t)(t, T)B(T) +E(t)b(t-T)]U(T) dT
Jo
(4-49)
where <Il(t, T) is the sta1

(4-49), we immediatel:
G(t, T) =
That G(t, T) = O for t <

itly embedded in writi
Ir the state-variab
description of the sys
problem-to find the
tion of a system-is n
two prablems: (1) Is j
fram the impulse-resp
state-variable descript
the first problem in th
studied in Chapter 6.
Consider a system
a linear finite-dimensil
response matrix, then
equation E, or more
G(t, T). The terminol!
dynamical equation, '
generate G(t, T). Not
impulse-response matl
not have any physical
tion gives only the Sal
equation is not in the
system.
Ir the realization (
dimensional linear dyr
not every G(t, T) is rea
form (4-48) that will gel
he impUlse response 1
ficient condition for G
Theorem 4-9
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.
IMPULSE-RESPONSE MATRICES AND DYNAMICAL EQUATlONS
TRICES
(4-38) is periodic with

amical equation E in
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).
: vector, and G is the

implicitly assumed in
1 (ith row, jth column)
lal due to a b-function
;e now that the internal
lllalysis of this system
(4-4881 )
(4-48b)
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 (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)(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 )
where E is a q x p matrix and M and N are, respectively, q x n and n x p con

tinuous matrices of t.
156
Proof
response matrix may

networks in Figure 4
o(t - r). For further r
Necessity: Suppose the dynamical equation

X = A(t)x. + B(t)u
E:
y = C(t)x. +E(t)u
Time-invariant ca~
case and see what cal

G(t,r)=G(t-r). Ce
response G(t - r) is re~
equation if and only if
is a realization of G(t, r); then

G(t, r) = E(t)O(t - r) + C(t)<ll(t, r)B(r)
= E(t)o(t - r) +C(t)'P(t)'P- 1 (r)B(r)
where'P is a fundamental matrix ofx = A(t)x.. The proof is completed by identi

fying
M(t) = C(t)'P(t)
and
G(t -r
N(t) = 'P- 1 (t)B(t)
Sujficiency: Let
G(t, r) =E(t)o(t -r) +M(t)N(r)
where M and N are q x n and n x p continuous matrices, respectively. Then the

following n-dimensional dynamical equation
E:
(4-52a)
(4-52b)
x(t) = N(t)u(t)
y(t)=M(t)x.(t) +E(t)u(t)
There are two objectio

terms of G(t - r) inste
to be realizable by a l
desirable is to have Cl
invariant dynamical-e<
we may also study as'
condition of realizatio~
terms of G(t).
Consider a system '
is a realization of G(t, r). Indeed, the state transition matrix of E is an n x n

identity matrix; hence,
Q.E.D.
G(t, r) = M(t)IN(r) + E(t)O(t - r)
We note that the dynamical equation in (4-52) can be simulated without

using feedback as shown in Figure 4-5 with B(t) = N(t) and C(t) = M(t).
Example 1
or, in the frequency do
where G(o) is the imp'

matrix of the system.
is found to be
Consider g(t, r) = g(t - r) = (t - r )e W - <). It is easy to verify that

g(t-r)=(t-r)eA(t-<)=[e At
r -re-A<l
teAtll
~
.L'
Le'" J
Hence, the dynamical equation

Since (4-53) and (4-55) <
E:
is a realization of g(t, r).
All the equivalent dynamical equations have the same impulse-response

matrix; hence, ifwe find a realization of G(t, r), we may obtain different realiza
tions of G(t, r) by applying equivalence transformations. Note that an impulse-
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~
IMPULSE-RESPONSE MATRICES AND DYNAMICAL EQUATIONS
[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
or, in thefreqency domain,

(4-53)
be simulated without
ld tU) =M(t).
where GO is the impulse-response matrix and G(s) is the transfer-function

matrix of the system. Suppose now a state-variable description of the system
is found to be
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)
Since (4-53) and (4-55) describe the same system, we have

G(s) = C(sI - A)-l B
ame impulse-response
,btain difIerent realiza
Note that an impulse-
+ E
(4-56)
As discussed in (3-36) and (3-37), C(sI - At 1 B is a strictly proper rational

matrix, and C(sI - A) -1 B +E is a ptoper rational matrix. Hence, the transfer
function matrix ofthe dynamical equation in (4-54) is a proper rational matrix.
Theorem 4-10
A transfer-function matrix G(s) is realizable by a finite-dimensional linear
time-invariant dynamical equation ifand only if G(s) is a proper rational matrix.
158
Proof
forward paths and loo
If (;(s) is realizable by a finite-dimensional linear time-invariant dynamical

equation, then from (4-56), we know that (;(s) is a proper rational matrix.
Before proving that every proper rational matrix (;(s) is realizable, we first
prove that every scalar (1 x 1) proper rational function is realizable. The most
general form of a proper rational function is
A()
gS = e
/3I S/l-I+"'+/3/1-IS+/3/1
s/l
We claim that the n-dimensional linear time-invariant dynamical equation

O
O
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)
+---,,-''--.:------,-,------;---'---''-----=-------''--''---+O:IS/l 1 + ... +O:II-IS +a/l
X2
and Li = 1, for i = 1,2,

input u to the output y
This proves the assertic

We are now ready
In order to avoid cumb
Let
O
1
(4-58a)
(4-58b)
is a realization of g(s). What we have to show is that the transfer function of

(4-58) is g(s). We shall demonstrate this by using Mason's formula for a signal
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
Signal-flow graph of the dynamical equation in (4-58).
is a realization of (;(s).
159
IMPULSE-RESPONSE MATRICES AND DYNAMICAL EQUATlONS
.TRICES
forward paths and loops have common nodes, we have

al
~-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+
{J IIs + {J 21S2 + ... + {J niSn

{JI Sn - 1
=e+
n
2
1 +aIls +a2/s + ... +anls
s" +a 1s"
+ ... + (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
r'J ~b" b'j[::J
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
oducts of two nontouching

not touching the ith forward
touching ir they do not have
Dn of (4-58). This is done in
~
~Y
is a realization of (;(s). Indeed, the transfer-function matrix of (4-61) is
[C~l C~2 C~l c~J

O
(sI-Ad- 1
O
O
!ln
(4-62)
4-58).
----
--------~---
------- -- =--=---=--=-=---=----=--==-=---,.--,,-=--=--=--=-----=--=-
=--~-,-===========-~--
160
-=-=-=--=-'==--=.. _=--='=---=---=--=' -----
~-~--
Thus every proper rational matrix is realizable by a finite-dimensional linear

time-invariant dynamical equation,
Q.E.Do
The realization procedure discussed in (4-60) and (4-61) is simple and straight
forwardo We realize every element of a transfer-function matrix independently
and then connect them from the input and the output as shown in Figure 4-70
The resulting realization, however, is generalIy not satisfactory for the folIowing
two reasons: First, the realization is internalIy not coupled as can be seen from
Figure 4-7. The coupling or interacting of alI variables, however, is a feature of
most physical multivariable systemso Second, the dimension of the realization
is generalIy unnecessarily large. These two problems will be resolved in Chapter
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
Hence, Corollary 4-10 follows directly from Theorem 4-100
,-------------,
I
"':
iT~l
:
:
;
L
:
I ~r
-~.~
I
, 1---
:""'
lA
g,2(S)
i
I
I
I
fl'
Y,
"2
I
r~
:Y2
~
~
+rc--'--
:
...J
Figure 4-7 lnternally uncoupled realization of (4-60),
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
The solutions of !ine;

The solution hinges on
ties <1>(1, t) = 1, <1>- l(t,
varying case, <1>(t, -r) is
<1>(t, -r) is equal to eA(l
duced in Section 2-7.
may bypass <1>(t, -r) an<
integration.
Oifferent analyses l
of a system. Mathema
basis chosen for the st
nothing to do with the
to the same input-outPl
description over the stc
Every proper ratior
dimensional!inear time
synthesis problem in ne
was also constructedo
generally it is possible t
be discussed in Chapter
bility,
To conclude this el
equationso The state
A(k)A(k -lJx(k -1) = .
A(m) (Problem 4-27). 1
putable. In terms of <1>(,

can be obtained as in Pr,
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
Since a realizable G(t) can be decomposed into G(t - r) = M(t)N(r), and

since entries of "G(t) are linear combinations of tket, for k =0,1, ... , and
i = 1,2, ... , the entries of M(t) and N{t) must be linear combinations of tke it .
Consequently, the matrices M(t) and N(t) are analytic functions of t on the entire
realline. As a consequence of this fact, we have the following theorem.
Theorem 4-11
A system that has a proper rational transfer function matrix is relaxed at to if and
only if u[lo.lo+'l = O implies Y[to.to+'l = O for sorne positive real s.
I
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
)lace transform of the

ion is of the form
lbination of the terms
10.
Concluding Remarks
The solutions of linear dynamical equations were studied in this chapter.

The solution hinges on the state transition matrix (t, r), which has the proper
ties (t, t)=I, -l(t, r)=(r, t), and (t, r)<l>(r, to)=(t, tolo For the time
varying case, (t, !) is very difficult to compute; for the time-invariant case,
(t, !) is equal to eA(t-t l , which can be computed by using the methods intro
duced in Section 2-7. In both cases, if only a specific solution is of interest, we
may bypass (t, !) and compute the solution on a digital computer by direct
integration.
Oifferent analyses often lead to different dynamical-equation descriptions
of a system. Mathematically, it means that dynamical equations depend on the
basis chosen for the state space. However, the input-output description has
nothing to do with the basis; no matter what analysis is used, it always leads
to the same input-output description. This is an advantage of the input-output
description over the state-variable description.
Every proper rational function has been shown to be realizable by a finite
dimensional linear time-invariant dynamical equation. This corresponds to the
synthesis problem in network theory. A realization of a proper rational matrix
was also constructed. However, the realization is not satisfactory, because
general1y it is possible to construct a lesser-dimensional realization. This will
be discussed in Chapter 6 after the introduction of controllability and observa
bility.
To conclude this chapter, we remark briefly the solution of discrete-time
equations. The state transition matrix (k, m) of .x(k + 1) = A(k)x(k) =
A(k)A(k -l),x(k - 1) = ... can be readily computed as (k, m) = A(k -1)'"
A(m) (Problem 4-27). Unlike the continuous-tim~ case, (k, m) is easily com
putable. [n terms of (k, m), the solution ofa discrete-time dynamical equation
can be obtained as in Problem 4-29. For the time-invariimt case, the equivalent
162
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
where o<1l(t, tol/Ot = A(t)lIJ
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
4-2 Show that ol1>(t, ')/0' = -11>(t, ,)A(,).

4-3 Find the solution of
x=[~ ~ ~]x+[~ ~]u

-2
-4
-3
-1
Consider
and its adjoint equation
shown in Figure P4-9. N,

. Ga(t, ,) be their impulse re:
1
2
Show that if A, R, and e are

matrices, then
with
uU) =
4-4
[~J
for t 2':0
Let
Find eA' by using the formula 2'['] =(sI-A)-l.
4-5 lf T- l(t) exists ~nd is <:ifferentiable for aH t, show that
d
-"[Tl(t)]
de .
= -
l(t) [el
- TU) ] r
4-6 From l1>(t, ), show how to compute AU).
elt ..
l(t)
Figure P4-9
PROBLEMS
.RICES
identical to the con

ion matrix (;(z) has a
mly if G(z) is a proper
m 4-10 is also directly
163
4-7 Given
show that
det (t, to)=exp [ [ (at(r) +adr))dr]
where 8<1>(t, to)lat = A(t)(t, to) and (to, to) = I. Hint: Show that
- det (t, to) = (a(t) +a 22 (t)) det (t, to)
latrices of the following
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 (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
and its adjoint equation
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)
Show that ir A, B, and C are constant matrices and G(s) and

matrices, then
G(s) = - G:( -s)
Figure P4-9
Gis) are their transfer-function
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
be the state transition matrix of
X=[A~
and the initial capacitor ,

immediately of the form ,
:::}
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"
Verify that B(t) = <1>(t, t o)Bo<1>*(t, t o) is the solution of

ti
-B(t)=A(t)B(t) +B(t)A*(t)
tlt
B(to)=B o
where <1>(t, to) is the state transition matrix of x = A(t)x.

4-13
Verify that X(t) = eA'CeBl is the solution of

ti
-X=AX +XB
tlt
4-14
X(O)=C
Show that if (t) =AA(t) -A(t)A, then

A(t) = eA,lA(O)e- A"
4-20
Find an equivalen!
X i (11 -t
Show also that the eigenvalues of A(t) are independent of t.

4-15
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
Can you transform

equivalence transformatioJ
4-21
4-22 Find a time-varying

realization of the impulse r
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
: response of some input
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
Find an equivalent time-invariant dynamical equation of
x= (cos t sin t)x

4-19
Find an equivalent Jordan-canonical-form dynamical equation of
1~][::]+[-~]
20
-25
4-20
X3
Find an equivalent discrete-time Jordan-canonicai-form dynamical equation of

XI(n
1)] [O
x(n+l) =
[
natrix of x= A(t)x is given
-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
of the network in Figure

Jr are zero and if an input
that ;1(1) ~.nd ;1(t) contain
you explain this?
he initial inductor current
Can you transform a time invariant {A, B, C} into {O, B, C} by a time-varying

equivalence transformation?
4-21
4-22 Find a time-varying dynamical equation realization and a time-invariant dynamical

realization of the impulse response g(t) = t 2 e!I.
4-23 Find a dynamical-equation realization of g(t, e) = sin te - (r -<) cos e. ls it possible to
find a linear time-invariant dynamical-equation realization for it?
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"
+ ... +CI."_IS +0:"

Verify also that
Realize the proper rational matrices
S+2
s +1
s +3
1
Ss
s +1
+1
s +2
2+2
+1
+3
1
5
+1
if and only if A(t) and A((
+1
52
+2
4-31
into continuous-time i:\nd discrete-time dynamical equations.
Show that if
A(e
lO
(
4-26 Consider the equivalent dynamical equations

Ax + Bu
{x ~x + Bu
{*y=Cx
y=Cx
=
is
where x = Px. Their adjoint equations are, respectively,
[Hint: Use Problem 4-30 a
: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
<1>(k, m)~ A(k - 1)A(k - 2)A(k - 3) ... A(m)

<1>(m, m)~ 1
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
Show that the solution of x(k
+ 1) =A(k)x(k) + B(k)u(k) is given by

k-l
x(k) = <1>(k, m)x(m)
+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
Let A(t) ~ (aij(t)). Then, by definition,
(d
- A(t) ~ -'--- adt)

dt
dt J
III
Bu(m)
PROBLEMS
ArRICES
Verify that
~ (A(t)B(t)) = A(t)B(t)
+ A(t)B(t)
dt
Verify also that
- [A(tW ~- (A(t)A(t)) = 2A(t)A(t)

dt
dt
;j
if and only if A(t) and A(t) commute; that is, (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
- <1>(t, to) = A(t)<l>(t, to)
is
<1>(t, to)=exp
f
lO
[Hint: Use Problem 4-30 and Equation (2-78).J

(la)
(lb)
(2a)
(2b)
d f,., respectively. Show
an
-1 *
edbyZ==(P )Z.
for k>m
.IS g'ven by x(k) = <1>(k, m)xo
trans ition matrix <1>(t, r) is

.
.....(k, m) is defined only?
Ilion,....
<:.111) to be defined for k < 111.
lis given by
)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
System analyses generally consist of two parts: quantitative and qualitative.

In the quantitative study we are interested in the exact response of the system
to certain input and initial conditions, as we studied in the preceding chapter.
In the qualitative study we are interested in the general properties of a systern.
In this chapter we shall introduce two qualitative properties oflinear dynamical
equations: controllability and observability. We shall first give the reader sorne
rough ideas of these two concepts by using the network shown in Figure 5-1.
The input u ofthe network is a current so urce. It is c1ear that ifthe initial voltage
in the capacitor e 2 in loop II is zero,no rnatter what input u is applied, the
mode e- t in II can never be excited. Hence the mode e- t in H is said te be y::
controllable by the input u. On the other hand, the rnode e - t in loop 1 can be
excited by the application of the input u; hence the mode in I is controllable by
the input u. Although the mode e -1 in 1 can be excited by the input, its presence
can never be detected frorn the output terminal y. Hence it is said to be not
observable from the ou

in II can be detected fre
This iIIustration, thou
cepts of controllability
The concepts of cor
study of control and fi\
systern shown in Figun
of the platform are sUPI
The rnass of the platfon
rnents of the two spring
both ends of the platfe
vibrate. If no force is
come back to rest. No~
to apply a force to bring
this question, the conce]
This chapter is orga
rnatical background is
for linear independence
results in controllabilit,
three theorerns. The c:
Necessary and sufficien
tions and linear time-il
derived. The concept o
to the concept of contrc
theorern is also devel0l
which are uncontrolIabh
theorern is developed.
function rnatrix of a dy
equation that is control
controllability and obser
cal equation. Their COI
inspection. In Section :
function controllabHitv :
of the input-output d~sc
sorne cornputational proi
Although elements 01
mathernatical convenienc
Xl
er'
Dampi.ng
coefficien!
u
(curren!
service)
Figure5-1
,11
Spring
constan!
A simple network.
Figure 5-2 A platform sys
168
INTRODUCTION
ltitative and qualitative.

:t response of the system
n the preceding chapter.
II properties of a system.
~rties oflinear dynamical
first give the reader sorne
)rk shown in Figure 5-l.
.r that ifthe initial voltage
lt input u is applied, the
e- t in 11 is said to be not
node e -1 in loop 1 can be
)de in 1 is controllable by
by the input, its presence
-lence it is said to be not
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
Figure 5-2 A platform system.
Spring
constant
1
170
CONTROLLABILlTY ANO OBSERVABILITY OF LINEAR OYNAMICAL EQUATlONS
plex numbers. Consequently, the state space of an n-dimensional dynamical

equation will be taken as an n-dimensional complex vector space (C", iC).
The references for this chapter are 2, 8,11,13,14,20,21,48,55,56,60,61,
69, 71, 98, 103, and 105 to 107.
The reader who is interested in only the time-invariant case may skip
Theorems 5-2, 5-5, and 5-6.
5-2
The concept of lin

tions. Let f j , for i = 1,
1 x p complex-valued
if there exist complex
alfl(t) +a
Otherwise, the f's are
in Equation (5-2) is a
we may also state tha
only if
Linear Independence of Time Functions
The concept of linear independence of a set of vectors of a linear space was

introduced in Section 2-3. We shall now apply this concept to a set offunctions
ofa real variable. A set of complex-valued functions Jl,f2> ... , Jn is said to be
linearly dependent on the interval 1 [t 1> t 2] over the field of complex numbers
if there exist complex numbers al' a2' ... , a m not all zero, such that
aI!I(t)+al!2(t)+ +allfn(t) =0
foral/tin [tI' t 2]
(5-1)
implies
IX =
0, where
Otherwise, the set of functions is said to be linearly independent on [t 1> t 2 ] over

the field of complex numbers. In this definition, the specification of time
interval is crucial.
Example 1
Consider the two continuous functions
11 and 12, defined by
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
Let f j , for i = 1, 2, ... , t

on [/,. I J. Let F be t
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
2 The matrix W(t ,. e2) is in ri
in this chapter.
.__
._~.:=~---.--_.-':::_
LINEAR INDEPENDENCE OF TIME FUNCTlONS
ICAL EQUATIONS
limensional dynamical
:tor space (Cn, C).
),21, 48,55, 56, 60, 61,
171
The concept of linear independence can be extended to vector-valued func

tions. Let f, for i = 1, 2, ... , n, be 1 x p complex-valued functions of t; then the
1 x p complex-valued functions f l , f l , ... , f n are linearly dependent on [t l' tlJ
if there exist complex numbers al' al" .. ,a", not all zero, such that
variant case may skip
(5-2)
Otherwise, the f;'s are linearly independent on [t 1> t l ]. Note that the zero vector
in Equation (5-2) is a 1 x p row vector [O
O]. As in Definition 2-4',
we may also state that f, f l , ... , f n are linearly independent on [ti' tlJ if and
only if
...
'Os
; of a linear space was

:ept to a set offunctions
,fl' ... , fn is said to be
Id of complex numbers
J, such that
'tin [ll,tlJ
(5-1)
for aH t in [tI' tlJ
(5-3)
implies t:x = O, where
)endent on [t l' t lJ over

~ specification of time
...
an
F"r~:l
lf..J
Clearly, t:x is a constant 1 x n row vector and F is an n x p matrix function.

The linear independence of a set of functions is a property associated with
an nterval; hence in testing for linear independence, we have to consider the
entire interva\. Let F*(t) be the complex conjugate transpose of F(t).
ed by
Theorem 5-1
ldent on [0, lJ, since ir

)r all I in [0, 1]. The
,0]. However, fl and
Let f j , for i = 1, 2, ... ,n, be 1 x p complex-valued continuous functions defined

on [1 l' I J. Let F be the 11 x (! matrix with f; as its ith row. Define
rlctions is linearly inde

linearly independent on
a subinterval on which
e 1 the functions fl and
for any positive 6. On
ltonan interval [tI' tlJ,
V interval that contains
re is no substantia\ difference
Every [t,. t,] is assumed to
W(t].
12)~ r"
Jt
F(t)F*(t) dI
Then f l f 2 , ... f n are linearly independent on [1 \.

constant matrix W(t l' ( 2 ) is nonsingular. 2
ir and only ir the
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
, The matrix W(t ,.

in this ehapter.
t ,)
is in faet positive definite(see Definition 8-6). This property is not needed

.
..
_-~---~
=. . .=-.--=..
--'="--==.C'-...
172
=----C-.~._."" . ,_. -._.-~ -" ... _.-.-.
,_U
. , , _ . _ _. _ _ - '_ _-_ _ - , - _
_ "~
CONTROLLABILITY AND OBSERVABILITY OF LINEAR DYNAMCAL EQUATIONS
implies exW(t, t 2 )ex* =0, or
and the C;s are linear

nonzero 1 x n row v'
'2
exW(t, t 2 )ex* =
1'1
(exF(t))(exF(t))* dt =0
(5-4)
Since the integrand (exF(t))(exF(t))* is a continuous function and is nonnegative

for all t in [t , t 2 J, Equation (5-4) implies that
for all t in [tI' t 2 J
exF(t) =0
This implies that

exF(k)( t) =
Hence we have
This contradicts the linear independence assumption of the set of f;, i = 1, 2,

... , n. Hence ifthef's are linearly independent on [tI' t 2 J, then det W(t l , t 2 ) f O.
We next prove the sufficiency of the theorem. Suppose that W(t l , t 2 ) is
nonsingular, but the lis are linearly dependent on [t l' t 2]. Then, by definition,
there exists a nonzero constant 1 x n row vector ex such that exF(t) = O for a11
t in [tI' t 2 ]. Consequently, we have
r exF(t)F*(t)dt=O
ex[F(t) : F
in particular,
which implies that al
t2
exW(t l ,t 2 )=
(5-5)
J"
which contradicts the assumption that W(t l , t 2 ) is nonsingular. Hence, if

W(t l' t 2 ) is nonsingular, then the f's are linearly independent on [t , t 2 ].
Q.E.D.
The determinant of W(t l' t 2) is called the Gram determinant of the f's. In
applying Theorem 5-1, the functions f i , for i = 1,2, ... , n, are required to be
continuous. lf the functions f;, for i = 1, 2, ... , n, have coIitinuous derivatives
up to order (n -1), then we may use the fo11owing theorem.
Thearem 5-2
are linearly depender
has rank n. Hence ti

The condition of
functions to be linea
example.
Example 2
Consider the two fun(
Assume that the 1 x p complex-valued functions f l , f 2 , . . . , f" have continuous

derivatives up to order (n - 1) on the interval [t , t 2]. Let F be the n x p matrix
with f i as its ith row, and let F(k) be the kth derivative of F. lf there exists some
t o in [tI' t 2J such that the n xnp matrix
defined ayer l -1, 1].
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
, Ir lo is al eilher 1, or 12 , lhe end points of an interval, then F<l(to)

approaching lo from inside the interval.
IS
denned as F1k)(I). with I
and
To check the linea

continuous, we can em
interval. If the functit
then Theorem 5-2 can 1
lheorem 5-1; however
LINEAR INDEPENDENCE OF TIME FUNCTIONS
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
for all t in [t l> t 2] and k = 1, 2, ... , n - 1
Hence we have
ex[F(t) : F(l)(t) : ... : F(llll(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 )
are linearly dependent. This contradicts the hypothesis that

[F(t o) : F(l)(t o) : ... : F(ll-l)(t O)]
)nsingular. Hence, if
lependent on [ti' t 2 J.
Q.E.D.
Q.E.D.
has rank n. Hence the f;s are linearly independent on [ti' t 2 ].
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.
rminant of the f's. In

, n, are required to be
;;ontinuous derivatives
~rn.
Example 2
Consider the two functions

fl(t)
f2(t)
.. , 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
t 2 ] over the field of
) 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
for all t in [-1, O)

at t =0
To check the linear independence of a set of functions, if the functions are

continuous, we can employ Theorem 5-1, which requires an integration over ario
interval. If the functions are continuously differentiable up to certain order,
then Theorern 5-2 can be used. lt is clear that Theorern 5~2 IS easier to use than
Theorem 5-1; however, it gives only sufficient conditions. lf the functions are
._ ... ._ .. -.- -_ - ._._.
"
..
---~------"'---------------
174
CONTROLLABILlTY AND OBSERVABILlTY OF LINEAR DYNAMICAL EQUATlONS
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
This can be seen fron
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
The sufficiency of the theorem can be proved as in Theorem 5-2. Now we

prove by contradiction the necessity of the theorem. Suppose that
p[FU o) :
F(l)UO) : ... : F(n-l)(to) : ..
'J < n
It is easy to verify th
Then the rows of the infinite matrix

[F(to): F(I)(t o): ... : F(II-l)U O):
p[F(t) : flll(t) : F(2 l (t) : F(
"'J
are linearly dependent. Consequently, there exists a nonzero l x n row vector

a. such that
(5-8)
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
for al! t in [to -6, t o +6J
(5-10)
Since the sum of analytic functions is an analytic function, the analyticity

assumption of the f/s implies that a.F(t) as a row vector function is analytic over
[t 1 , t 2 ]. Consequently, Equation (5-10) implies that
a.F(t) =0
for al! t in [t 1, t 2 J
or, equivalently, the f/s are linearly dependent on [t1>

diction.
t 2 ].
[F(t) : 1
This is a contra
Q.E.D.
A direct conseque'nce of this theorem is that if a set of anal ytic functions is

linearly independent on [t1> t 2 J, then
p[F(t): f'(I)(t) : ... : F(n-1)(t)=--'J =n

for al! t in [t l' t 2]. Itfollows that if asid of analytic functions is linearly inde
pendent on [t l' t 2 J, then the set of analytic functions is linearly independent on
=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
then we have the follo
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
CONTROLLABILlTY OF LINEAR DYNAMICAL EQUATlONS
on the fact that if a

pletely determinable
ion at that point are
175
every subinterval of [tI' t 2 ]. In this statement, the analyticity assumption is

essential. The statement does not hold without it, as we have seen in Example 1.
Note that Theorem 5-3 is not true if the infinite matrix in (5-7) is replaced by
p[F(ro) : F(l)(ro) : ... : Fin - l)(tO)J
This can be seen from the following example.

the n x p matrix with
t o be any fixed point
Jif and only if
=n
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
sin lOOOt: 10 3 cos lOOOt

[ sm 2000t :' 2 x 10 3 COS 20001
=.
It is easy to verify that p[F(t): F(l)(t)] < 2 at t
O,
10 - 3n;, . . ..
However,
p[F(t) : Fl)(t) : F(2)(t) : F( 3 )(t)]

=p [
zero 1 x n row vector
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 exists an e> O such

a Taylor series about
[F(t) : F(l)(t) : ... : F(n -l)(t)]
e, t o +eJ
(5-9)
then we have the following corollary.
e obtain
(5-10)
Corollary 5-3
ction, the analyticity

lction is analytic over
Assume that, for each i, f is analytic on [t l' t 2]. Then f l' f 2 ,

independent on [t 1, t 2 J if and only if
p[F(t) : F(l)(t) : ... : F(n -l)(t)]
... ,
f n are linearly
for almos! all t in [tI' t 2 ].

This corollary will not be used in this book, and therefore its proof is
omitted.
2]. This is a contra

Q.E.D.
.f analytic functions is
5-3
=n
Controllability of Linear
Dyn~mical
Equations
Time-var\'ng case. In this secti~n, we sh~ll int~oduce the conceptof con

trollability of linear dynamical equations. Tobe more precise, we study the
state controllability of linear state equations. As will be seen immediately,
ctions is linearly inde

nearly independent on
176
CONTROLLABILlTY AND OBSERVABILlTY OF LINEAR DYNAMICAL EQUATIONS
the state controllability is a property of state equations only; output equations

do not play any role here.
Consider the n-dimensional, linear state equation
E:
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
precisely, the dynami

at any to.
Example 2
Consider the system ~

X2 in the system, T
cannot transfer Xl ay.
then no matter what
Hence the equation t
The solution of tn
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;
where (t, t o) = 'P(t)\

nonsingular for aH t.
Theorem 5-4
The state equation E
ti> t o such tqat the 11
independent on [t o, ti
Proof
Sufficiency: If the row

Theorem 5-1 the n x y.
W(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) =
will transfer X o to the
In
In
+
11
~.
rv
In
In
Figure 5-3 An uncontrollable network.
IF;::
1.11:
Figure 5-4 An uncont
CONTROLLABILITY OF LINEAR DYNAMICAL EQUATIONS
:AL EQUATlONS
[lly; output equations
(5-11 )
tor, and A and B are,

continuous functions
III is an n-dimensional
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) =
ime to' if there exists a

md any x in ~, there
:he state x at time ti.
at time too
le of moving any state
lt trajectory the state
;traint imposed on the
ve sorne examples to
abk x ofthe system is

lor aH t?: t o no matter
letwork, and the input
the system-or, more
177
x(t) = cP(t; t o, xo, u) = 1I>(t, to)x o +
l'
Xo
is given by
<1>(t, L)B(-r)U(L) dL
[o
1I>(t, t o{ Xo +
lI>(t o, L)B( L)u(L) dLJ
where <1>(t, t o) = 'I'(t)'I' -(t o); '1' is a fundamental matrix of i

nonsingular for all 1.
(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)
.te is said to be co/ltrolfable.

r definition does not make
c. tl1e equation E is said lo
:rb "completely" is dropped
js nonsingular. Given any x.(t o) =
~~o
and any x j", '.ve claii1J that tb.e input

(5-14 )
will transfer X o to the state
1F
Xl
at time ti. Indeed, by substituting (5-14) into
+tz
X
-=.t
Figure 5-4 An uncontrollable network.
178
CONTROLLABILITY AND OBSERVABILITY OF LINEAR DYNAMICAL EQUATIONS
(5-12), we obtain
X(t) = <1>(t, tO){X O- [
<1>lt o r)B(r)B*(r)<1>*(to, r) elr W-(t o, I)
. [x o -(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
Thus we condude that the equation E is controllable. Necessity: The proof is

by contradiction. Suppose E is controllable at to, but the rows of (to,' )B(')
are Iinearly dependent on [1 (l. (IJ for aH (1 > (o. Then there exists a nonzero,
constant IXIl row vector a. such that
a.<1>(t (l. t)B(t) = O

Let us choose x(t(l)~x(l = a.*
for all 1 in [lo. IIJ
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 )
Then Equation (5-12) beco mes
<1>U (l. tl)xU 1 ) = a.* +
r"
10
<1>(1 (l. r )B( r )u( r) dr
(5-16 )
and, in general,
Premultiplying both sides of (5-16) by a.. we obtain

a.<1>(t(l.t l )x(tI) =a.a.*
r"
10
Ok
ot k $(t o, c:
a.<1>(to. r)B(r)u(r) dr
(5-17 )
Theorem 5-5
By hypothesis, E is controHable at t o ; hence for any state-in particular,

XI =O-there exists ullO,ld such that X(I)=O. Since a.<1>(10, t)B(t) =0, for allt
in [to, tJ, Equation (5-17) reduces to
Assume that the matri(
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
Then, from (5-19) and
r Ilu(1)111 dt
Jea
Since $(t o, t ) is nonsir
Define
[$(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}
In the following we shall give a controllability criterion based on the matrices

A and B without solving the state equation. However, in order to do so, we
need sorne additional assumptions on A and B. Assume that A and B are (n - 1)
times continuously differentiable. Define a sequence of n x D matrices M o('),
M(), ... , by the equation
d
M k+ (t) = -A(t)Mk(t) + dt Mk(t)
with
lecessity: The proof is

:he rows of C1>(to,' )B(')
here exists a nonzero,
179
k = O, 1,2, ... , n - 1
Mo(t)=B(t)
(S-18a)
(S-18b)
Observe that
(to, t)B(t) = (to, t)Mo(t)
(S-19a)
:t (to, t)B(t) = 'P(to) {[:t '11 -
(t)J B(t) +'11 - (t) :t B(t)}
(S-1S)
;omes
) dr
'P(to)'P-(t) [-A(t)B(t) + :t B(t)J
(to, t)M (t)
(S-19b)
(S-16)
and, in general,
ak
k=O,I,2, .. ,n-l
at k (to, t)B(t) = (to, t)Mk(t)

(r) dr
(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
nput u(t) that transfers

lption of A and B, the
lre generally many dif
the trajectory between
lputS that achieve the
jing to sorne criterion.
Define
a t"', (e.0' " )1R1-
"'\
- ,-,,1.
""\lo, ',)1lJl(
j;]) LJ
at
L
~'a'---
1=1,
0'
'
, M\Li
Then, from (5-19) and using (2-2), we have:
an
[(t o, t)B(t) : 'Jt;(t o, t)B(t) : ... : at1- (t o, t)B(t)
= (t o, t)[Mo(t} : M(t): ., . : Mn-1(t)]
Since (to, ti) is nonsingular, the assumption

the minimal :energy in
Appendix c.'
( '11 orthe state ttansi~
we mentioned earlier,
not readily applicable.
p[Mo(t} : M(t) : ... : Mn-(t 1 )] =n

implies
(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
If the matrices A and E

equation E is different
for any fixed to in ( - ex
Consider
Xl]
[X3 =
~z
10J [1] +[0]
[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 linear dynamical

in ( - 00, 00) if and onl)
Differential eontrollability. instantaneous eontrollability,

and uniform eontrollability
where the M/s are as d
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
In lhe engineering lileraluro

firsl used in Reference S6 lo
we adopt lhe terminology "
MICAL EQUATIONS
CONTROLLABILJTY OF LINEAR DYNAMICAL EQUATlONS
are linearly independent

e conclude that the state
Q.E.D.
sufficient but not neces
181
controllability. The condition for differential controllability can be easily

obtained by slight modifications of Theorems 5-4 and 5-5. However, if the
matrices A and B in the state equation E are analytic on ( - 00,00), then we have
the following theorem.
Theorem 5-6
Ifthe matrices A and B are analytic on ( - 00, CJ:)), then the n-dimensional state
equation E is differentially controllable at every t in ( - 00, 00) if and only if,
for any fixed t o in ( - CJ:), (0),
11
(5-22)
If the matrix A is analytic on (- 00, 00), it can be shown that the state transi
tion matrix (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 (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
- aH t 1=0, the dynamical
The linear dynamical equation E is said to be instantaneously controllable 5

in (- 00, (0) if and only if
for aH t in (-00, (0)
ilitv,
where the M's are as defined in Equation (5-18).
~ee distinct types of caD

s of continuity.
U a dynamical equation is instantaneously controHable, then the transer o

the states can be achieved instantaneously at any time by using an input that
consists of b-functions and their derivatives up to an order of n - 1. It is clear
that instantaneous controllability implies differential controllability. The
most important implication of instantaneous controllability is that in the case
of a single input, the matrix [Mo(t) MI(t) ... Mn-l(t)] qualifies as an
equivalence transformation (Definition 4-5). Consequently, many canonical
form equivalent dynamical equations can be obtained for instantaneously
controHable dynamical equations. See References 10,99, 103, and 110.
y) controllable at time t o
: Xl in L, there exists an
itrarily small interval of
I
state in a finite time (no

itrollable. If this can be
he state equation is said
IcontroHability implies
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
0< cx((Je)I ,:::;;W(t, t +(Je) ~cxz(aJI

and
0< cx 3 ((Je)I ~<l>(t +(J" t)W(t, t +(Je)<l>*(t +(J" t)':::;;cx 4 ((Je)I
-3
for all t in ( - 00, (0), where <l> is the state transition matrix and W is as defined
in Equation (5-13).
I
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
Consider the one-dimensionallinear dynamical equation
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)
for all t > O.
Example 5
Consider the one-dimensional dynamicai equation
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
Figure 5-5 The function
achieved in the time intt

will not be arbitrarily lar
W - ]. In optimal cont
sometimes required to e
n-dimensional linear tim
where x is the n x 1 statt:

and n x p real constant .
equations. the time inte
.
that is, [O, 00).
The condition for a t
that there exists a fini te ti
in t. on [t o, tI]. In the
As discussed in Chapter
terms of form t k (!/.'; hencl
sequently, if its rows ar
independent on [too t lJ f
time-invariant state eqm
and the transfer of any st
time intervaL Hence th,
trol1ability study of linea
Theorem 5-7
The n-dimensional linear

if and only if any of the fe
1. Al1 rows of e-AIB (an
[0, (0) over e, the fiel,
1'. All rows of (sI - A)-I
6
Although aH the entries of A al

of the field of complex numbe:
nCAL EQUATlONS
183
b (1)
'lable if and only if there

that
t)saiaJI
-3
-2
-1
trix and W is as defined

I
Figure 5-5 The functian b(t).
.tive definite matrix (see

le transfer of the states
.uniform controllability
:ms. See References 56
ability both imply con
neither implies nor is
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
ion is not instantaneous

; uniformly controllable
re =
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)
where x is the n x 1 state vector, u is the p x 1 input vector; A and B are n x n

and n x p real constant matrices, respectively. For time-invariant dynamical
equations, the time interval of interest is from the present time to infinity;
that is, [O, (0).
The condition for a time-varying state equation to be controllable at t o is
that there exists a finite t 1 such that aH rows of cl>(t o' t)B(t) are linearly independent,
in t, on [to, ti]. In the time-invariant case, we have eJ>lt o , t)B(t)=eA(to-l)B.
As discussed in Chapter 4, al! elements of eA(to-tlB are linear combinations of
terms of form tke 1 ; hence they are analytic on [O, 00) (see Appendix B). Con
sequently, if its rows are linearly independent on [O, 00), they are Iinearly
independent on [t o' ti] for any t o and any ti> t o 1n other words, if a linear
time-invariant state equation is controllable, it is controllable at every t o 2:0
and the transfer of any state to any other state can be aclieved in any nonzero
time interval. Hence the reference of t o and ti is often dropped in the con
trollability study of linear time-invariant state equations.
Theorem 5-7
The n-dimensional linear time-invariant state equation in (5-23) is control1able
if and only if any of the fol1owing equivalent conditions is satisfied:
1. All rows of e-AtB (and consequently of eA/B) are linearly independent on

[O, (0) over ic, the field of complex numbers. 6
1'. All rows of (sI - At 1 B are Iinearly independent over iC.
6 Althollgh
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
CONTROLLABILITY ANO OBSERVABlLITY Of LINEAR OYNAMICAL EQUATlONS
2. The controllability grammian

W el ~
Example 6
Consider the inverted r

equation is developed il
eA'BB* *, dr
2g(M +m)/(2M +m)l=
is nonsingular for any t> 0. 7

3. The n X (np) controllability matrix
U~[B:
AB: A 2 B:"': A"-lBJ
(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
for any t in [0, (0). Let t = O; then the equation reduces to

p[B: - AB: . ..; ( - 1)" - lA" - 1 B: ( -1)" A"B : ...] = n
From the Cayley-Hamilton theorem, we know that Am with m n can be written

as a linear combination of 1, A, ... , A"-l; hence the columns of AmB with
m "2: n are linearly dependent on the columns of B, Aa, ... ,A"-1B. Conse
quently,
p[B: -AB: ... : (-l)"-IA"-IB:' "J=p[B: -AB:"': (-I)"-1A"-lBJ
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
This matrix can be read

Thus, if X3 = eis differer
to bring it back to zero.
derivatives back to zer
balancing a broom on (
Example 7
Consider the platform s

of the platform is zero ;
systems. The spring c(
coefficients are assumed
X2 +X2 =u, or
This is the state-variabh

Now if the initial di~
platform will oscillate,
platform to come to re~
and X2(O) = -1, is it po~
.seconds? The answer d,
applied to thetwo sprin
For the state equatic
p
The matrix is in raet positive definite. See Problem E-II.
s This eondition implies that (sI -
Al and B are left eoprime. See Appendix G.
hence the state equation
~AL
CONTROLLABILlTY OF LINEAR DYNAMICAL EQUATIONS
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
5-3 implies that

nd only if
l:-"J=n
to
:'-'J=n
:h m n can be written
::olumns of AmB with
, ... ,An-IB. Conse
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
erator, if the rows of

complex numbers, so
-5 (page206). Q.E.D.
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~)
This is the state-variable description of the system.

Now if the initial displacements XI (O) and X2(0) are different from zero, the
platform will oscillate, and it will take, theoretically, an infinite time for the
platform to come to rest. Now we pose the following problem: If Xl (O) = 10
and X2(0) = - 1, is it possible to apply a force to bring the platform to rest in 2
seconds? The answer does not seem to be obvious because the same force is
applied to the two spring systems.
For the state equation in (5-26), we compute
p[B : ABJ = p
:ndix G.
[~.5=~.25J = 2
hence the state equation is controllable. Consequently, the displacements can
186
be brought to zero in 2 seconds by applying a proper input u. Using Equations

(5-13) and (5-14), we have
[(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
for t in [0, 2]. If a for

at t = 2. The behaviol
Hnes in Figure 5-6.
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
Consider again the plal

the viscous friction coe
a11 equal to 1. Then t
-40
xl (1)
Clearly we have
10
............
........
............
........... 2
2.5
...------- ....
3
1, seconds
and the state equation

an input to transfer x(C
xz(O), no input can tfar
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
Behavior of x) and X2(t) and the waveformof u..
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
CONTROLLABILlTY Of LINEAR DYNAMICAL EQUATlONS
;AL EQUATlONS
6.33J
27
and
ul(t)
=-
[0.5
l][e~'St ~] W-I(O, 2) [ ~~J = -44.1eo. St +20.7e
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
Let A and B be n x n and n x p constant matrices.

U k = [B :. AB : ... : AkB]
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
CONTROLLABlLITY AND OBSERVABILITY OF LINEAR DYNAMICAL EQUATIONS
explicitly as
U k = [b
bz
ro
(5-28)
/"k
(no. of dependent columns)
Now we search linearly independent columns of U k in order from left to right;

that is, if a column can be written as a linear combination of its left-hand-side
columns, the column is linearly dependent; otherwise, it is lintlarly independent.
This process of searching can be carried out by using, for example, the column
searching algorithm discussed in Appendix A. Let ri be the number of linearly
dependent columns in AiB, for i = O, 1, ... ,k. If B has a full column rank, then
ro = O. We note that if a column, say Ab z, is linearly dependenton its left-hand
side columns, then all Aib z, with j = 2, 3, ... , will be linearly dependent on its
left-hand side columns. lndeed, if
then
Ab z = IX b +lXzb z + ... +lXpb p +lX p+Ab

AZb z =1XAb +lXzAb z + ... +lXpAb p +lXp+Azb
We see that A zb z is a linear combination of its left-hand-side columns. Pro

ceeding similarly, we can show that Aib z, for j = 3, 4, ... ,are linearly dependent
on its left-hand-side columns. Because of this property, we have
O'::;ro'::;r'::;rz'::;'"
.::;p
Since there are at most n linearly independent columns in U oo , there exists an

integer fJ. such that
(5-29 a)
O.::;ro'::;r'::; '::;rL_ <P
(5-29b)
rL=rL+="'=P
and
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
where p is the rank (

singular.
Example 9
Consider the sateHite
equation is developed
the control of Xi' i = 1
are entirely independel
of (3-51):
Equivalently, fJ. is the integer such that

pU o < pUl < ... < pU/._
= pUl' =
pUl' +1
= ...
(5-30)
In other words, the rank of Uk increases monotonically until k reaches fJ. - 1;

thereafter, all P col umns of AiB will be linearly dependent of their left-hand-side
columns. Hence the controllability of {A, B} can be checked from UL- and fJ.
is called the controllability indexo We claim that the controllability index
satisfies the inequalities
n
(5-31 )
- '::;fJ. .::;min(, n - p+ 1)
P
where is the degree of the minimal polynomial of A and Ji is the rank of B.
From the definition of minimal polynomial, we have
A =1XA - +1X2A-Z + ... +lX l
y=
where we have assum<

bars. The rank of B is
the matrix
[B: AB
for sorne lX i , which implies

AB =IXA -B +IX Z A - zB + ... +IX B
has rank 4. This is thc
LINEAR INDEPENDENCE OF TIME FUNCfIONS
ICAL EQUATIONS
(5-28)
Irder from left to right;

on of its left-hand-side
is lin6arly independent.
r example, the column
the number of linearly
full column rank, then
endenton its left-hand
learly dependent on its
Hence, all columns of AB are linearly dependent on their left-hand-side

columns. Consequently, we have fl ~ . From (5-30), we see that the rank of
Uk must increase by at least 1 as k inereased by 1; otherwise, it will cease to
inerease. Sinee the rank of U", is at most n, if the rank of B is p, it is suffieient
to add at most n - p number of AjB. Hence, we eonclude from (5-30) that
fl- 1 ~ n - p or fl::; n - p + 1. The eombination of fl ~ and fl ~ n - p + 1
yields the upper bound in (5-31).
In order to have rank n, the matrix U 1"-1 must ha ve more columns than
rows. Hence, we need flP ~ n or fl ~ nI p. This establishes the lower bound in
(5-31).
The degree of the minimal polynomial of A is not easy to compute; hence

min (, n - p + 1) in (5-31) is often replaeed by n - p + 1. From (5-30) and
Theorem 5-7, we have immediately the following eorollary.
Corollary 5- 7
The state equation FE is controllable if and only if the n x (n - p + l)p matrix
Un - p = [B
\.b 1
ld-side columns. Pro

, are linearl y dependent
we have
; in U"" there exists an

(5-29a)
(S-29b)
.. , 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
y until k reaches fl- 1;

lt of their
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)
where we have assumed, for simplicity, w = m = ro = 1 and dropped the over

bars. The rank of B is 2; hence, the state equation is controllable if and only if
the matrix
and p is the rank of B.

1
AB
where p is the rank of B, has rank n, or the n x n matrix lJ n _ pU~_ p is non

singular.
xp + A 2 b
'1"+1
189
[B' AB: A'B]

.
~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
For this problem, we can readily show
the controllability inc

1, ... , /1}.
pU o =2<pU=4=pU 2 =pU 3 = ...

hence, the controllability index of (5-32a) is 2.
We study more the controllability matrix U. We assume that the linearly

independent columns of U in order from left to right have been found. We
now rearrange these independent columns as
b, Ab, ... , Al'l - b, b 2, Ab 2, ... , AI'2- b 2, ... , b p, Ab p, ... , Al'p- b p
Theorem 5-8
The set of the controll~
transformation and ar
Proof
Define
The integer /1 is the number of linearly independent columns associated with b i

in the set, or is the length of the chain associated with b. Clearly we have
/1
where A=PAP-, .
easily verified that
max{/1, /12, ... , /1 p}
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
r - r_ = no. of {bk , k = 1, 2, ... , p} with controllability index i
(5-33)
with l' _ d~O. For example, r - ro = 1, and b 2 has controllability index 1;

1'2 - r =0 and no b has controllability index 2; 1'3 - 1'2 = 2, and b and b4 have
controllability index 3; 1'5 -1'4 = 1, and b 3 has controllability index 5. Hence,
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
Figure 5-7 Crate diagram of Ai-1bj .
which implies
Hence, the r defined

indices are invariant Ul
The rearrangement
where M is a p x p eler
forward to verify that
-
f:,
Uk=[B:
where diag {M, M, ...

singular. Hence, we h
Consequently. we cane
the ordering ol" the eoll
Now we discuss a (
umns in U = [B AB
and then search its linl

In terms af the crate di
are searched in order fr,
column and so forth. 1
:AL EQUATIONS
LINEAR INDEPENDENCE OF TIME FUNCTIONS
191
the controllability indices of {A, B} are uniquely determinable from tri, i =0,
1, ... , Ji}.
I
ume that the linearly

ave been faund. We
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
nns associated with b

]early we have
k=[B: AB:"': AkBJ

where A= PAP- ,
easily verified that
l
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=
PB and P is any nonsingular matrix. Then it can be
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
: AB : ... : AkBJ = U k diag{M, M, ... , M}

where diag {M, M, ... , M} consists of k + 1 number of M, and is clearly non
singular. Hence, we have
for k =0, 1,2, ...
Consequently, we conclude that the controllability indices are independent of
the ordering of the columns of B.
Q.E.D.
Now we discuss a different method of searching linearly independent col
umns in U = [B AB ... An-lBJ. We first rearrange the columns of U as
b l , Ab l , AZb, ... , An-b; b z, Ab z, ... , An-lb z ; ... ; b p, Ab p, ... , An-lb p
(5-34)
and then search its linearly independent columns in order from left to right.
In terms of the erate diagram in Figure 5-7, the linearly independent columns
are searched in order from top to bottom in the firstcolumn, then in the second
column and so forth. Let
.
b l' Ab b
... ,
Aii,-lb l'. b 2' Ab 2,
... ,
Aii2- l b"
2,
... ,
b P' Ab P'
... ,
Aiip-lb p
...:.=---_._-_._----
-----~ -----=-------=-:..:::...-.::--:..:'--- - - - - - - - - - - ' -.." ' - - - , - - - - - -- - - ----- _.
192
-------------------,~=~
be the resulting linearly independent columns. If {A, B} is controllable, we have

ji + ilz + . .. + il p = n. The lengths of these chains are {il, ilz,, il p }
Unlike the controllability indices, these lengths of chains depend highly on the
ordering of {b, i = 1,2, ... , p}.
Consider the n-dirr
Example 10
where A, R, e, and E a
continuous functions (
Consider the state equation in (5-32). If we search linearly independent columns

from left to right in
Definition 5-5
b, Ab, AZb, A 3 b; b z, Ab z, AZbz, A3 b z

the resulting linearly independent columns are
b, Ab, AZb, b z
Its lengths are {3, 1}. If we search from left to right in
b z, Ab z, AZb z, A 3 b z ; b, Ab, AZb, A3 b
the resulting linearly independent columns are
b z, Ab z, AZb z, A3 b z
Its lengths are {4, O}. The lengths are indeed different for different ordering of
b and b z. The controllability indices of (5-32) can be computed as {2, 2} and
are independent of the ordering of b and b z.
I
Let il = max{il, ilz, ... ,ilp }. It is clear that il can never be smaller than /1,
the controllability indexo Since /1 = max{/1, /1z, ... , /1p}, we may conclude
that /1 is the smallest possible maximum length of chains obtainable in any
search of linearly independent columns of U. This controllability index will
play an important role in the design of state feedback in Chapter 7 and the design
of compensators in Chapter 9.
The dynamical equati

if there exists a finite tI
of the input U[lo,l,] and
to determine the state
unobservable at too
Example 1
Consider the network ~

the initial voltage acros
the output is identicall:
cally zero), but we are n
hence the system, or m
system, is not observab
Example 2
Consider the network sl

reduces to the one sho'
In
In
lF
urv
5-4
Observability of linear Dynamicai EqiUia-:imu$
Time-varying case. The concept of observability is dual to that of control

lability. Roughly speaking, controllability studies the possibility of steering
the state from the input; observability studies the possibility of estimating the
state from the output. If a dynamical equation is controllable, all the modes of
the equation can be excited from the input; if a dynamical equation is observ
able, all the modes of the equation can be observed at the output. These two
concepts are defined under the assumption that we have the complete knowledge
of a dynamical equation; that is, the matrices A, B, e, and E are known before
hand. Hence, the problem of observability is different from the problem of
realization or identification. The problem of identification is a problem of
estimating the matrices A, B, e, and E from the information collected at the
input and output terminals.
Figure 5-8 An unobserv

lH
(a)
Figure 5-9 An unobserv
ICAL EQUATIONS
193
OBSERVABILlTY OF LINEAR DYNAMICAL EQUATIONS
is controllable, we have
are {P1' z, ... , lp}.
IS depend highly on the
Consider the n-dimensional linear dynamical equation
x= A(t)x(t) + B(t)u(t)
E:
(S-3Sa)
(S-3Sb)
y = C(t)x(t) + E(t)u(t)
where A, B, e, and E are n x n, n x p, q x n, and q x p matrices whose entries are

continuous functions of t defined over (- 00, 00).
ly independentcolumns
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
for different ordering of

computed as {2, 2} and
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
never be smaller than /1,

, /1p}, we may conclude
lains obtainable in any
ontrollability index will
::::hapter 7 and the design
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
Figure S-S An unobservable network.

IH
IH
In.
1
' -_

_---4-----_
F
(a)
Figure S-9 An unobservable network.
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
Let us choose x(to) = cx.
The response of the dynamical equation (5-35) is given by

y(t) = C(t)<1>(t. to)x(to) +C(t)
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
y(t) = C(t)<1>{t. t o)x(t o)
(5-37)
y(t)~ y{t) -
(5-38)
C(t)
Ir l1>{t. r)B(r)u(r) dr - E(t)u(t)

lO
is a known function. Consequently. the observability problem is a problem of

determining x(to) in (5-37) with the knowledge of y, C, and l1>(t. to). Note that
the estimated state x(t o) is the state not at time t, but at time too However, if
x(to) is known, the state after t o can be computed from
x(t) =l1>(t, to)x(to)
Irlo l1>(t, r)B(r)u(r) dr
(5-39)
The dynamical equation E is observable at t o if and only if there exists a finite

ti> t o such that the n columns of the q x n matrix function C(' )<1>(', to) are linearly
independent on [t o, ti].
Proof
Sufficiency: Multiplying l1>*(t, to)C*(t) on both sides of (5-37) and integrating

from t o to tI' we obtain
l1>*(t, to)C*(t)y(t) dt
=[[ l1>*(t, to)C*(t)C(t)l1>(t, to) dt] X o ~ Veto, t l)xO

(5-40)
1,
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
Hence the initial statex(

tion that E is observa!
ti> t o such that the col
We see from this th~

tion depends only on e
can also be deduced [re
servability study, it is s
x = A(t)x, Y = C(t)x.
The controllability
independence of the rol1
by the linear independ
between these two conc
Theorem 5-10 (Theort
Theorem 5-9
Prove by contradictio
ti> t o such that the co
Then there exists an n
Consider the dynamiea

defined by
where A*, B*, C*, and E

E in E. The equation
equation E* is observab
Proof
From Theorem 5-4, the
rows of l1>(t o, t)B(t) are 1
5-9, the dynamical equ.
B*(t)l1>a(t, to) are linearly
[B*(t)a(t, to)J* = ;:'(t, t
<l>a is the state transition r
l1>Uo, t) (see Problem 4-8)
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. ~
OBSERVABILITY 01' LINEAR DYNAM1CAL EQUATIONS
CAL EQUATIONS
Prove by contradiction. Suppose E is observable at to, but there exists no

ti> t o such that the columns of COP(', to) are linearly independent on [to, ti].
Then there exists an n x 1 nonzero constant vector a such that
terminal. Therefore,
Iductor from the input
lmical equation is not
C(t)<.l>(t, tola = O
+ E(t)u(t)
Y(t) = C(t)<.l>(t, to)a = O

(S-36)
for al1 t> t o
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.
[n the study of observ
;nown, the initial state

(S-37 )
We see from this theorem that the observability of a linear dynamical equa
tion depends only on C(tl and (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 (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 (t, to). Note that
t time to. However, if
d,
for al1 t> t o
Let us choose x(to) = a; then
:n by
11;
195
(5-39 )
Theorem 5-10 (Theorem of duality)

Consider the dynamical equation E in (5-35) and the dynamical equation E*
defined by
y if there exists a finite

IC(' )<1>(', t o) are linearly
E*:
Proof
dtJxo~ Veto, t1)xo
From Theorem 5-4, the dynamical equation E is controllable if and only if the
rows of (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)a(t, to) are linearly independent, in t, on [to, t 1J, or equivalently, the rows of
[B*(t)a(t, to)]* = :(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 :(t, to) =
(t o, t) (see Problem 4-8); hence E is control1able if and only if E* is observable.
Q.E.D.
(S40)
(541)
lmns of q.)(', t o) are

lo, t d is nonsingular.
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
(5-37) and integrating
) dt
z= - A*(t)z +C*(t)v
1 =B*(t)z +E*(t)v
We list in the following,for observability, Theorems 5-11 to 5-14 and Defini

tions 5-6 lo 5-8, which are dual to Theorems5-5 to 5-8 and Definitions 5-2 to
5-4 for control1ability. Theorems 5-11 to 5-14 can be proved either directly or
(S-42)
rom(5-42). Necessity:
i
I
L
196
CONTROLLABILlTY ANO OBSERVABILlTY OF LINEAR OYNAMICAL EQUATIONS
by applying Theorem 5-10 to Theorems 5-5 to 5-8. The interpretations in the

controlIability part also apply to the observability part.
Definition 5-7
The linear dynamical

( - 00, 00) if and only i
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 the N;'s are as d

Definition 5-8
where
k = 0, 1, 2, ... , n-1
(5-45a)
The linear dynamical e

if and only if there exist
(5-45b)
I
WIth
*Differential observability, instantaneous observability,

and uniform observability
0< {31(O"o)I:
O < {33(O"o)I:
for alI t, where el> is the
Differential, instantaneous, and uniform observabilities can be defined by using

the theorem of duality; for example, we may define {A, C} to be differentialIy
observable if and only if { -A*, C*} is differentialIy controlIable. However,
for ease of reference we shalI define them explicitIy in the folIowing.
Time-invariance ca
equation
Definition 5-6
The dynamical equation E is said to be difJerentially observable at time t o if,

for any state x(t o) in the state space L, the knowledge of the input and the output
over an arbitrarily smalI interval of time suffices to determine x(t o).
I
Theorem 5-12
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
The n-dimensional linea:

able if and only if any 01
1. AH columns of CeA!
complex numbers.
1'. AlI columns of C(sI
2. The observability gra
=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
*Sections noted with an asterisk may be skipped withollt loss or continllity.
is nonsingular for an;
_---~~,
..
~
_
..
-~----~
'-.-.e-_
OBSERVABILlTY OF LINEAR DYNAMICAL EQUATIONS
(lCAL EQUAUONS
197
Definition 5-7
le interpretations in the
The linear dynamical equation E is said to be instantaneously observable in

( - 00, ro) if and only if
nal dynamical equation
dynamical equation E
.t
l~;~:;
l=n
Nn~l(t)J
for all t in ( -
00,
(0)
where the N/s are as defined in (5-45).
(5-44)
11
Definition 5-8
~,
... ,n-1
The linear dynamical equation E is said to be uniformly observable in ( - ro, (0)

if and only if there exist a positive (Jo and positive f3i that depends on (Jo such that
(5-45a)
(5-45b)
I
0< f31((Jo)I.:::::;V(t, t +(Jo).:::::;f3z((Jo)1

0< f33((Jo)I ':::::;$*(t, t +(Jo)V(t, 1 +(Jo)$(t, t
+ (Jo) .:::::;f34((Jo)I
Ity,
for all t, where $ is the state transition matrix and V is as defined in (5-41).
: can be defined by using

~, C} to be differentially
controllable. However,
he following.
Time-invariance case.
observable at time to if,
the input and the output
termine x(to).
I
:n the n-dimensional dy
t in (-ro, (0) ifand only
11
Consider the linear time-invariant dynamical
equation
FE:
x=Ax +Bu
y=Cx +Eu
(5-46a)
(5-46b)
where A, B, C, and E are n x n, n x p, q x n, and q x p constant matrices. The

time interval of interest is [O, (0). Similar to the controllability part, if a linear
time-invariant dynamical equation is observable, it is observable at every
t o ::::: O, and the determination of the initial state can be achieved in any nonzero
time interval. Hence, the reference of t o and t 1 is often dropped in the observa
bility study of linear time-invariant dynamical equations.
Theorem 5-13- _
The n-dimensional linear time-invariant dynamical equation in (5-46) is observ

able if and only if any of the foHowing eq uivalent conditions is satisfied:
_ontnuity.
1. AH columns of CeA! are linearly independent on [O, ro) over

complex numbers.
1'. AH columns of C(sI - A)-l are lineady independent over C.
2. The observability grammian
Wot~
is nonsingular for any 1 > O.
eA''C*Ce A < dI
e, the field of
-------------------------
198
..
------~_...:-_~~~~-=--.----------
CONTROLLABILlTY ANO OBSERVABILlTY Of LINEAR OYNAMICAL EQUATlONS
CANONICAL OECOMPO~
3. The nq x n observability matrix
and
Equivalently, v is the
pV o <
(5-47)
The integer v is called

have
has rank n.
4. For every eigenvalue Aof A (and consequently for every A in C), the (n +q) x n
complex matrix
where i5 the degree (

Corollary 5-13
has rank n, or equivalently, (sI - A) and
e are right coprime.
The dynamical equatif

V n _ q , where q is the n
V:_qVn_q is nonsingul
Observability Indices
Let A and
e be n x n and q x n constant matrices.

l
C2
C:
Define
Consider the matri)

in order from top to bo
independen t rows asso
the observability indice!J
ro (no. of dependent rows)
Cq
Theorem 5 -14
(5-48)
The set of observability

transformation and any
5-5
CalT1lonical Dre
DYl1lllmicai Equatic
lt consists of k + 1 block rows of the form CA and is of order (k + l)q x n. The

matrix V = V n _ 1 is the observability matrix. The q rows of e are denoted by
ej, i = 1, 2, ... ,q. Let us search linearly independent rows of Vk in order from
top to bottom. Let r be the number of linearly dependent rows in CA,
i = O; 1, ... , k. Similar to the controllability part, we have
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)
In the remainder of thi

dynamical equations. e
where A,B, e, and E ar

Weintroduced in the p
observability. The cond
able are also derived. A ql
said if the eq ua tion is un<
shall study this problem.
transformation. Let x =
CANONICAL DECOMPOSITION OF A LINEAR TIME-INVARIANT OYNAMICAL EQUATION
CAL EQUATIONS
rv=r v + = ... =q
and
199
(5-49b)
Equivalently, v is the integer such that
pVo<pV< ... <pVV_=pVv=pV v +='"
(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)
:y .A. in (:), the (n + q) x n
where is the degree of the minimal polynomial of A and
q is the rank
of C.
Corollary 5-13
The dynamical equation FE in (5-46) is observable if and only if the matrix
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.
Consider the matrix V n _ . lt is assumed that its linearly independent rows

in order from top to bottom have been found. Let Vi be the number of linearly
independent rows associated with e i. The set of {Vi' i = 1,2, .. , , q} is cal1ed
the observability indices of {A, C}. Clearly we have
ine
l rows)
v = max [Vi, i = 1, 2, ... , q l
and VI +V 2 + ... +\'q Sil. The equality holds ir [A, C) is observable.

Theorem 5-14
(5-48)
The set of observability indices of [A, el is invariant under any equivalence

transformalion and any ordering or the rows of C.
I
S-5
Canonical Decomposition of a Linear Time-llI1Ivariall1lt
Dyn<llmical IEquatioU1J
In the remainder of this chapter, we study exclusively linear time-invariant

dynamical equations. Consider the dynamical equation
I order(k + l)q x n. The

)WS of C are denoted by
,ows of Vk in order from
i
iependent rowS in CA ,
have
roo, there exists an integer

(5-49a)
FE:
x=Ax +Bu
y=Cx+Eu
(5-52a)
(5-52b)
where A, B, C. and E are n x n, n x p, q x n, and q x p real constant matrices.

We introduced in the previos sections the concepts of controllability and
observability. The conditions for' the equation to be controllable and observ
able are also .d~rived. A questi()ll that m;w be raised at this point is: What can be
saidif the equation is uncontrol1able and/or unobservable? In this section we
shall study this problem. Before proceeding, we review briefly the equivalence
transformation. Let x = Px, where P is a constant nonsingular matrix. The
200
CONTROLLABILlTY ANO OBSERVABILlTY Of LINEAR OYNAMICAL EQUATIONS
CANONICAL OECOMPOSI1
substitution of X = Px into (5-52) yields
Proof
X=Ax +Bu
+Eu
FE:
(5-53a)
(5-53b)
If the dynamical equati

have
where A = PAP-, B = PB, C = CP-, and E = E. The dynamical equations

FE and FE are said to be equivalent, and the matrix P is called an equivalence
transformation. Clearly we have
pU~
y = Cx
V~[B:
AB: ... : An-B] =[PB: PAB:: PAn-lB]

=P[B: AB: ... : An-B]~PU
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
The controllability and observability of a linear time-invariant dynamical

equation are invariant under any equivalence transformation.
I
This theorem is in fact a special case of Theorems 5-8 and 5-14. For easy
reference, we have restated it as a theorem.
In the following, e will be used to stand for controllable, efor uncontrollable,
o for observable, and (5 for unobservable.
Let q, qz, ... , qn, be an

each i = 1, 2, . . . , n , .
{q, qz, ... ,qnJ (Why
p-
where the last n - n coi
Q is nonsingular. We c
into the form of (5-54).
x = Px we are actually u:
state space. The ith coll
Aq with respect to {q, ~
linearly dependent on th
given in (5-54a). The c(
B with respect to {q, (
{q, qz, .. , q!!.l}; hence
Let U and U be the co
we have pU = pU = n (s
- [i
=[
U=
Theorem 5-16
Consider the n-dimensional linear time-invariant dynamical equation FE.

If the controllability matrix of FE has rank n (where n < n); then there exists
an equivalence transformation X = Px, where P is a constant nonsingular
matrix, which transforms FE into
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
Thus the transfer-functiOl
and the nI-dimensional subequation of FE
Xc = Acx c + Bcu
y=Ccx c +Eu
where
AcB wlth k
(5-55a)
(5-55b)
is controllable 9 and has the same transfer function matrix as FE.
- _][sI - Ac
[ Ce
Ce
O
=
sI
[Ce Ce] I
=CcCsI -Ac
9
It is easy to show that fhe equation

(Try)
FE s observable. then its subequation FE, is also observable.
which is the transfer-func
CANONICAL DECOMPOSITlON OF A LINEAR T1ME-INVARIANT DYNAMICAL EQUATION
IlCAL EQUATlONS
201
Proof
(5-53a)
(5-53b)
e dynamical equations
is caBed an equivalence
: PAn-lB]
~n-B]~ PU
If the dynamical equation FE is not controllable, then from Theorem 5-7 we
have
pU~p[B
: AB : ... : An-BJ =n < n
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.
;-8 and 5-14. For easy

lle, efor uncontrollable,
namical equation FE.

~ < n), then there exists
constant nonsingular
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
A~'B: ... A~-BeJ}n rows

O :
}(n-n)rows
where e represents the controBability matrices of F Eco
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.
We show now that the dynamical equations FE and FEe , or correspondingly,

It is easy to verify 1:h3.t
FE and FEc have the same transfer-functon matrix.
(5-57)
(5-54b)
Thus the transfer-function matrix of FE is

(5-55a)
(5-55b)
'ix as FE.
[Ce
CcJ[sl
~Ae
s~~~cTT~eJ +E
= [Ce C-J [(Si - Aer

e
O
(si -AJ-l A:?lSI - Ac)-J [Be]

(si - Acr
O
+E
=Ce(sl -AJ-Be +E
ualion
FE, is <j.lso observable.
which is the transfer-function matrix of FEe .
Q.E.D.
202
CANONICAL DECOMPO;
In the equivalence transformation X = Px, the state space L of FE is divided

into two subspaces. One is the nl-dimensional subspace of L, denoted by Lb
which consists of all the vectors
[~c
and
the other is the (n - n 1 )-dimensional
subspace, which consists of all the vectors
[:J
Since F Ec is controllable, all
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
Hence, the reduced ce

Xc =
Dual to Theorem
dynamical equations.
Theorem 5-17
Consider the n-dime

If the observabilitv m:
valence transform~tio:
FE
Consider the three-dimensional dynamical equation

1
y=[1
l]x
(5-58)
and the nrdimensiom
The rank of B is 2; therefore, we need to check U 1

the controllability of the equation. Since
[B : AB] in determining
1
1
1
is observable and has
the state equation is not controllable.

Let us choose, as in (5-56),
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[~
B~Ph[~ ~ ~][~o ~] r~lb __~]o
This theorem can l

The first n2 rows of P i
the observability matr
arbitrary so long as P
Xo does not appear eh
vector Xi; is not observ:
Combining Theore
theorem.
Theorem 5-18 (Cane
Consider the linear tirr
By equivalence transf<
-1
10
This is simplified versior

References 57, 60. and 116
ICAL EQUATIONS
CANONICAL DECOMPOSITION OF A LINEAR TIME-INVARIANT DYNAMICAL EQUATION
pace ~ of FE is divided
;e of~, denoted by L. l ,
and
1) shows that the state

. indirectly through the
llable and is dropped in
:mt dynamical equation
:e vector can be decom
r uncontrollable. By
tain a controllable dy
tate equivalent to the
O 1
1] 1 O
[
O 1
le (n - nl)-dimensional
FE c is controllable, all
Hence, the reduced controllable equation is

Xc =
G ~]
Xc
+ [~
~] u
y = [1 2Jx c
(5-59 )
Dual to Theorem 5-16, we have the following theorem for unobservable

Theorem 5-17
Consider the n-dimensional linear time-invariant dynamical equation FE.

If the observability matrix of FE has rank nZ(n2 < n), then there exists an equi
valence transformation X = Px that transforms FE into
FE:
1Jx
203
(5-60a)
y=[C o O][::J+EU
(5-58)
(5-60b)
and the n2-dimensional subequation of FE

I : AB] in determining
Xo= Aox o+Bou

y =Cox o +Eu
(5-61 a)
(5-61 b)
is observable and has the same transfer-function matrix as FE.
: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
Consider the linear time-invariant dynamical equation

FE:
x=Ax +Bu
y=Cx +Eu
By equivalence transformations, FE can be transformed into the following

lO
This is a simplified version of the canonical decomposilion theorem. For the general form, see
References 57, 60, and 116. See also Reference S127.
---~._---------
204
CANONICAL DECOMPC
CONTROLLABILlTY AND OBSERVABILITY OF LINEAR DYNAMICAL EQUATIONS
canonical form
(5-62 a )
Cco : Ce]x
y = [O
+Eu
(5-62b)
ca
co
where the vector x is controllable but not observable, x is controllable

and observable, and Xc is not controllable. Furthermore, the transfer function
of FE is
Cco(sI -
Aco )-) Bco +E
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
This theorem cal

which the uncontror
observable parts. V
response matrix of a
and observable part
matrix (the input-out
controllable and obse
input-output descripl
us why the input-ou
system, for the uncor
appear in the transfel
Example 2
Consider the networl

source, the behavior
detected from the ou
L 1 are not observabl{
the state variable aSSt
metry, the state varial
By dropping the state
the network in Figun
the transfer function
co
+ Eu
Combining these two transformations, we immediately obtain Equation (5-62).

Following directly from Theorems 5-16 and 5-17, we conclude that the transfer
function of FE is given by
A
B E.
Q.E. O.
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
Figure 5-10 Canonical decomposition of a dynamical equation. (e stands for control

lable, e for uncontrollable, o for observable, and o for unobservable.)
Figure 5-11 An uncon
CANONICAL DECOMPOSITION Of A LINEAR T1ME-INVARIANT DYNAMICAL EQUATION
CAL EQUATIONS
(5-62a)
(5-62b)
205
This theorem can be illustrated symbolically as shown in Figure 5-10, in

which the uncontrollable part is further decomposed into observable and un
observable parts. We see that the transfer-function matrix or the impulse
response matrix of a dynamical equation depends solely on the controllable
and observable part of the equation. In other words, the impulse-response
matrix (the input-output description) describes only the part of a system that is
controllable and observable. This is the most important relation between the
>le, xco is controllable

le, the transfer function
input-output description and the state-variable description. This theorem tells

us why the input-output description is sometimes insufficient to describe a
system, for the uncontrollable and/or unobservable parts of the system do not
appear in the transfer-function matrix description.
le part of the equation
Example 2
e transformed into the

FE c which is the con
ru be '<ausfmmed tuto
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.
lbta;,., Equattou (5-62)

Ilclude that the transfer
Q.E.D.
Cl
Li
Lz
"
In
In
c2
In
In
(a)
In
In
y
In
In
(b)
Figure 5-11 An uncontrollable and unobservablc systcm with transfer function 1.
206
CONTROLLABILlTY ANO OBSERVABILlTY OF LINEAR OYNAMICAL EQUATlONS
Proof of statement 4 of Theorem 5-7

The matrix (si - A) is nonsingular at every s in iC except at the eigenvalues of A;
hence the matrix [si - A : B] has rank n at every s in iC except possibly at the
eigenvalues of A. Now we show that if {A, B} is controllable, then
p[sl - A: B] =n at every eigenvalue of A. lfnot, then there exist an eigenvalue
}, and a 1 x n vector ex f. O such that
ex[AI -A : B] =0
exA = exA
and
exB = O
or
CANONICAL OECOMPO
equation still has the

ing definition.
Definition 5-9
A linear time-invaria
only if there exists a l
sion that has the sam
state equivalent to FE
which imply
Theorem 5-19
and, in general,
A linear time-invariar
is control1able and ob
i = 1,2, ...
Hence we have
Proof
This contradicts the controllability assumption of {A, B}. Hence, if {A, B} is

controllable, then p[ si - A : B] = n at every eigenvalue of A and, conse
quently, at every s in iC.
Now we show that if {A, B} is not controllable, then p[AI -A : B] < n
for sorne eigenvalue A of A. If {A, B} is not controllable, there exists an equi
valence transformation P that transforms {A, B} into {A, R} with
=[~c ~c2]
A=PAP- 1
B=PB=[~C]
Let A be an eigenvalue of Ac. We choose pf.O such that PAc=AIJ. Now we

form a 1 x n vector ex = [O JI]. Then we have
ex[AI-A; 8]=[0
p][AI~Ac
-A 12
'
Al -Ac '
R
oc] =0
Ifthe dynamical egual

is reducible (Theorem
FE is control1able an,
contradiction. Supp<
and observable and th
FE,
of lesser dimension, sa
Theorem 4-6, we have
Consider now the pro<
which implies
o=et[P(Ali -
A)p-l : PBJ =~'p[(},~ - A)p-l : RJ
Since ex f.O, we have ii~exP f.O. Because p- 1 is nonsingular, ii(AI - A)P- 1 =0

implies ii(AI - A) = O. Hence we have
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
Irreducible dynamical equation: We have seen from Theorems 5-16 and

5-17 that if a linear time-invariant dynamical equation is either uncontrollable
orunobservable, then there exists a' dynamical equation of lesser dimension
that has the same transfer-functioh rilatrix as the original dynamical equation.
In other words, if a linear time-invariant dynamical equation is either uncontrol
lable orunobservable,its dimension caribe reduced such that the reduced
where \'''_1 and Un -

lable and observable"
2-6 that p(VU) = n. N<
matrices; hence the ID
---------------
207
CANONICAL DECOMPOSITION OF A LINEAR TIME-INVARIANT DYNAMICAL EQUATION
: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
A linear time-invariant dynamical equation FE is said to be reducible if and

only if there exists a linear time-invariant dynamical equation of lesser dimen
sion that has the same transfer-function matrix as FE or, equivalently, is zero
state equivalent to FE. Otherwise, the equation is said to be irreducible.
Theorem 5-19
A linear time-invariant dynamical equation FE is irreducible if and only if FE
is controllable and observable.
Proof
"-la.BJ=O
f. Hence, if {A, B} is
lue of A and, conse
t1 p[ Al - A : BJ < n
:, there exists an equi
S} with
IJA c = AfI. Now we
If the dynamical equation FE is either uncontrollable or unobservable, then FE

is reducible (Theorems 5-16 and 5-17). Hence what we have to prove is that if
FE is controllable and observable, then FE is irreducible. We prove this by
contradiction. Suppose that the n-dimensional equation FE is controllable
and observable and that there exists a linear time-invariant dynamical equation
FE,
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)
Consider now the product
-1: BJ
ular, a,(AI - A)P-l =0
CB
CAB
~
: BJ does not have a
proof.
Q.E.D.
CAn-lB
CAnB
CAn-lB
CAnB
(5-65)
CA2(~-1)B
By (5-64), we may replace CAkB in (5-65) by CAkB; consequently, we have

1m Theorems 5-16 and
either uncontrollable
n of lesser dimension
II dynamical equation.
Ion is either uncontrol
:uch that the reduced
'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
PeY n _ 1 n _ ) = n > ni' This is a contradiction. Hence, if

FE is controllable and observable, then FE is irreducible.
Q.E. O.
(5-66) implies that
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
Let the dynamical equation {A, B, C, E} be an irreducible realization of a

q x p proper rational matrix G(s). Then {A, B, C, E} is also an irreducible
realization of G(s) if and only if {A, B, C, E} and {A, B, C, E} are equivalent;
that is, there exists a nonsingular constant matrix P such that A = PAP-1,
B = PB, C = CP -1, and E = E.
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)
where the matrices A

The n x n matrix A is
A.m' A denotes aH tb
the number of lordan
A = diag
(i!,
Table 5-1 Jordan-Fu
The irreducibility assumption implies that pV = n; hence the matrix (V*V) is

nonsingular (Theorem 2-8). Consequently, from (5-67), we have
= eY*V) - 1 V*VU ~ PU
(5-69)
A
(n x 11
where P ~ eY*V)-l V*V. From (5-69) we have p .=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.
*5-6 Controllability and Observability of Jordan-Form

Dynamical Equations
The controllability and observability of a linear time-invariant dynamical
equation are invariant under any equivalence transformation; hence it is con
ceivable that we may obtain simpler controllability and observability criteria
by transforming the equation into a special formo Indeed, if a dynamical
equation is in a Jordan forro, the conditions are very simple and can be checked
almost by inspection. In this section, we shall derive these conditions.
Consider the n-dimensional linear time-invariant Jordan-form dynamical
equation
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)
nin (pP, pU), which,

ifies as an equivalence
= PE. Since pU = n,
ions of (;(5) have the

equa
te the equation in an
Hage elements (if the
iven transfer-function
~ible dynamical
osition of linear time-
lA'
J (n~p) ~[J
A2
C=[C I C 2
A
(ni x n)
erive from (5-67) and

CP and A=PAP- l .
Q.E.D.
A = diag (Al' Az, ... , A m)
Table 5-1 Jordan-Form Oynamical Equation
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
CONTROLLABILITY AND OBSERVABILITY OF JORDAN-FORM DYNAMICAL EQUATIONS
{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
observable are that tI

dent, the set {C IZ1 }'
Hence J F E is not ob:
be the order of A and A ij , respectively; then

m
n=
ni =
i=I
r{i)
I I
nij
=lj=l
Corresponding to A and Aj, the matrices B and C are partitioned as shown.

The first row and the last row of Bij are denoted by blj and b lij , respectively.
The first column and the last column of Cij are denoted by Clij and c lij .
Theorem 5-21
The n-dimensional linear time-invariant Jordan-form dynamical equation J FE

is controllable if and only if for each i = 1, 2, ... , m, the rows of the r(i) x p
matrix
(5-71 a)
are linearly independent (over the field of complex numbers). J FE is observable

if and only if for each i = 1, 2, ... , m, the columns of the q x r(i) matrix
(5-71b)
are linearly independent (over the field of complex numbers).
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
Consider the Jordan-form dynamical equation
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
CONtROLLABILITY AND OBSERVABILITY OF JORDAN-FORM DYNAMICAL EQUATIONS
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'
amical equation JFE

~ rows of the r(i) x p
(5-71 a)
;). J FE is observable
>( r(i) matrix
(5-71 b)
11
:rs).
The conditions for controllability and observability in Theorem 5-21

required that each of the m set of vectors be individually tested for linear inde
pendence. The linear dependence of one. set on the other set is immaterial.
Furthermore, the row vectors of B excluding the bli/s do not play any role in
determining the controllability of the equation.
The physical meaning of the conditions of Theorem 5-21 can be seen from
the block diagram of the Jordan-form dynamical equation JFE. Instead of
studying the general case, we draw in Figure 5-12 a block diagram for the
Jordan-form dynamical equation in (5-72). Observe that each block consists of
an integrator and a feedback path, as shown in Figure 5-13. The output of each
block, or more precisely the output of e.ach integrator, is assigned as a state
variable. Each chain of blocks corresponds to a Jordan block in the equation.
Consider the last chain of Figure 5-12. We see that if b l21 #=0, then aH state
variables in that chain can be controHed; if C l21 #=0, then al! state variables in
that chain can be observed. Ir there are two or more chains associated with the
same eigenvalue, then we require the linear. independence of the first gain
vectors of these chains. The chains associated with different eigenvalues can
be studied separately.
~b111
~b112
1 U~b!13
2
(5-72a)
u
y
);
1J
~b121
u
y
(5-72b)
tere are three Jordan

Jordan block associ
controlable are that
tions for J FEto be
Figure 5-12
Block diagram of the Jordan-form equation (5-72).
212
CONTROLLABIl.ITY
A~
theorem can be estab
Figure 5-13 Analog computer simulation of 1/(5 -).).

We use statement 4 of Theorem 5-7 to prove the theorem. In order not to be
overwhelmed by notation, we assume [sI - A : B] to be of the form
-1
S-Al
O
O
S -
Al
O
b ll1
-1:
b Zll
b111
b 11Z
S-Al:
----------------------:- -. .: X; ------= i -- -,
~
s -Al
bl1z
- - - - - - - - - - - - -: -s- ~ Jc~ -----~-( --:- -l-Z-l

s - Az : blZ1
(5-73)
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)
By a sequence of elementary column operations, the matrix in (5-74) can be

transformed into
,
-1
O
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
Observe that in <

independent, it is ne<
that is, p = 1-it is n,
form dynamical eqm
block associated with
to a vector. Thus we
O
O
o :
o (
r~- -~
There are two distinc

responds to the last
is zero; therefore, the
corresponding to the fl
therefore, the equatior
Example 3
Consider the followin!
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
That the state eqmi.ti,

Equation (5-77) is a t
matrix is in the Jordar
ICAL EQUATIONS
CONTROLLABlUTY AND OBSERVABILlTY OF JORDAN-FORM DYNAMICAL EQUATlONS
213
linearly independent. By proceeding similarly for each distinct eigenvalue, the

theorem can be established.
Q.E.D.
Observe that in order for the rows of an r(i) x p matrix B to be linearly
independent, it is necessary that r(i) s p. Hence in the case of single input
that is, p = l-it is necessary to have rU) = 1 in order for the rows of B to be
linearly independent. In words, a necessary condition for a single-input Jordan
form dynamical equation to be controllable is that there is only one Jordan
block associated with each distinct eigenvalue. For p = 1, the matrix B reduces
to a vector. Thus we have the following corollary.
m. In order not to be
be of the form
, b lll
b l21
A single-input linear time-invariant Jordan-form dynamical equation is con

trollable if and only if there is only one Jordan block associated with eaeh
distinct eigenvalue and all the components of the column vector B that corre
spond to the last row of each Jordan block are different from zero.
A single-output linear time-invariant Jordan-form dynamical equation is
observable if and only if there is only one Jordan block associated with eaeh
distinct eigenvalue and all the components ofthe row vector e that correspond
to the first column of each Jordan block are different from zero.
(5-73)
Example 2
b 211
b L11
b l12
: b L12
:t~-
--- --=--( - -:- -b l-2 1

s-
"'-2
Corollary 5-21
There are two Jordan

becomes
Consider the single-variable Jordan-form dynamical equation
=,1.1' (5-73)
.
x=
(5-74)
bL12
-T-----
:
: :
b l21
b 121
aatrix in (5-74) can be
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
Consider the following two Jordan-form state equations:
[~J=[-~ _~Jx +GJ u
(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,
mly if blll and bl12 are
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
,=.c:::..:c..:==-C=",.c.'="'='_'="='--=----==='-''''=--=''------ ,_, _--'
214
OUTPUT CON
CONTROLLABILlTY AND OBSERVABILlTY Of LINEAR DYNAMICAL EQUATlONS
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
where A, R, and e are

response matrix of the
[ - '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.
It is clear that G(s) is a
*5-7 Output Controllability and Output Function

Controllability
A system whose transfe

output controllable if a
over the field of comple
Similar to the (state) controllability of a dynamical equation, we may define

controllability for the output vector of a system. Although these two concepts
are the same except that one is defined for the state and the other for the output,
the state controllability is a property of a dynamical equation, whereas the
output controllability is a property of the impulse-response matrix of a system.
Consider a system with the iilput-output description
has rank q.
y(t) =
roo G(t, r)u(r) dr
where u is the p x 1 input vector, y is the q x 1 output vector, G(t, r) is the q x p

impulse-response matrix of the system. We assume for simplicity that G(t, r)
does not contain o-functions and is continuous in t and r for t> r.
Corollary 5-22
The proof of Theore

trivial consequence of tl
controllable. We see
can also be stated in te
indeperidence of G(s), tl
The state controlla
the output controllabili
these two concepts are
Definition 5-10
A system with a continuous impulse-response matrix G(t, r) is said to be output

conlrollable at time lo if, for any y l ' there exist a finite l I > lo and an input
U[Io,',] that transfers the output from y(to) = Oto Y(l 1) = y l '
I
Example 1
Consider the network sI

observable, but it is out
Theorem 5-22
A system with a continuous G(t, r) is output controllable at t o if and only if there

exists a finite ti > t o such that all rows of G(t l , r) are linearly independent in r
on [to, ti] over the field of complex number.
I
The proof of this theorem is exactly the same as the one of Theorem -5-4
and is therefore omitted.
We study in the following the class of systems that also have linear time
invariant dynamical-equation descriptions. Consider the system that is
describable by
FE:
x=Ax +Bu
y=Cx
ID.
ID.
u
IF
+
-
Ir
1\
ID.
ID.
Figure 5-14 A network ,

observable.
OUTPUT CONTROLLABILITY AND OUTPUT FUNCTION CONTROLLABILITY
ICAL EQUATIONS
215
where A, B, and C are n x n, n x p, q x n real constant matrices. The impulse

response matrix of the system is
(5-77) is controHable.
t o, we have
-fo ]
= [ :- 2 10
The transfer-function matrix of the system is

mdent in t. Hence the
(5-78 )
rem, aH the conditions

ain an erroneous con
It is clear that (;(s) is a strictly proper rational function matrix.

Corollary 5-22
A system whose transfer function is a strictly proper rational-function matrix is

output controllable if and only if aH the rows of (;(s) are linearly independent
over the field of complex numbers or if and only if the q x np matrix
runction
luation, we may define

.ugh these two concepts
the other for the output,
equation, whereas the
mse matrix of a system.
1
ector, G(t, r) is the q x p

Ir simplicity that G(t, r)
lrfort>r.
[CB: CAB: ... : CAn-lB]

has rank q.
(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.
,(t, r) is said to be output
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
at to if and only ifthere

nearly independent in r
I
:he one of Theorem 5-4

t also have linear time
er the system that is
In
In
'1;
1\
In
In
Figure 5-14 A network which is output controllable but neither (state) controllable nor
observable.
-_ .._.. __ ._--_. __._._._._. __
._--_._._-~--_._._-
216
....
"
..-
...._..
_._- .._-"._--
CONTROLLABILITY ANO OBSERVABILITY Of LINEAR OYNAMICAL EQUAll0NS
1-----------------,
I
I
I
I
I
I
:
I
range of G(s), for wh

YI
I
I
l-----+:--.-Y2
I
Figure 5-15 A system which is controllable and observable but not output controllable.
Example 2
Consider the system shown in Figure 5-15. The transfer-function matrix of

the system is
,;1]
If the input is rest

then the given outpu
computed from (5-81
given output has som
needed to generate th
The condition fOl
terms of the matrices,
interested reader is re
Dual to the output
These problems are 1
q x p proper rational
exists a p x q rational
s +1
the rows of which are linearly dependent. Hence the system is not output
controllable. The dynamical equation of the system is
x= -x +u
which is controllable and observable.
If a system is output controllable, its output can be transferred to any

desired value at certain instant of time. A related problem is whether it is
possible to steer the output following a preassigned curve over any interval of
time. A system whose output can be steered over any interval of time is said
to be output function control/able or functional reproducible.
Theorem 5-23
A system with a q x p proper rational-function matrix (;(s) is output function

controllable if and only if p(;(s) = q in lR(s), the field of rational functions with
real coefficients.
Proof
If the system is initially relaxed, then we have
y(s) 7' (;(s)o(s)
(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)
then Equation (5-80) is satisfied. Consequently, if pG(s) = q, then the system is

output function controllable. If pG(s)< q, we can always find a y(s), not in the
A system is said to ha"

A necessary and suffi
pG(s) = q in IR (s). Tl
controllability; Man~
unique? Is it a prop
stable? What are its
problems will not be ;
References SI72, S185
"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.
1t not output controllable.
:fer-function matrix of
system is not output
be transferred to any
oblem is whether it is
ve over any intervalof
interval of time is said
ble.
(;(s) is output function

rational functions with
(5-80)
independent over the

5)(;*(S) is nonsingular
(5-81 )
= q, then the system is
IS
find a y(s), not in the
Ir the input is restricted to the class of piecewise continuous functions of t,

then the given output function should be very smooth; otherwise, the input
computed from (5-81) will not be piecewise continuous. For example, if the
given output has sorne discontinuity, an input containing b-functions may be
needed to generate the discontinuity.
The condition for output function controllability can also be stated in
terms of the m:ttrices A, B, and C. However, it is much more complicated. The
interested reader is referred to Reference 8.
Dual to the output function controllability is the input function observability.
These problems are intimately related to the inverse problem. A system with a
q x p proper rational matrix (;(s) is said to have a right (left) inverse if there
exists a p x q rational matrix GR1(S) [Gu(s)] such that
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.
*5-8 Computational Problems

In this section, we discuss sorne computational problems encountered in this
chapter. As discussed in Section 2-9, a problem may be well conditioned or ill
conditioned; a computational method may be numerical1y stable or numerical1y
unstable. Ir we use a numerically stable method to solve a wel-conditioned
problem, the result wil\ generally be good. Ir a problem is i1l conditioned, even
if we use a numerical1y stable method to solve it, there is no guarantee that the
result will be correct. Ir we use a numerical1y unstable method to solve a
problem, well or in conditioned, the result must be carefully scrutinized. Given
a problem, if we must use an unstable method because of nonexistence of any
stable method, the unstable method should be applied at a stage, as late as
possible, in the computation.
As discussed in Theorem 5-7, there are several ways of checking the con
trollability of a state equation. Among them, statements 3 and 4 appear to be
moresuitable for computer computation. However, they may encounter sorne
computationalproblems, as will be discussed in the fol1owing.
The computation of the control1ability matrix U = [B AB ... A"- 1 B]
is straightforward. Let Ko~ B. We compute K = AK_l; i = 1, 2, ... , n-1.
- ... '
._...
_-~
---~~.
218
CONTROLLABILITY AND OBSERVABILITY OF LINEAR DYNAMICAL EQUATlONS
At the end, we have U = [K o K 1

K"-l]. The rank of U can then be
computed by using the singular value decomposition (Appendix E) which is a
numerically stable method. If the dimension n of the equation is large, this
process requires the computation of AkB for large k and may transform the
problem into a less well conditioned problem. For convenience in discussion,
we assume that all eigenvalues Ai of A are distinct and B is an n x 1 vector.
We also arrange Ai so that IA 11 IA21 ... IA"I. Clearly, we can write B as
For example, consider
where v is an eigenvector associated with eigenvalue Ai; that is, Av = AiVi. It is

straightforward to verify
(see Reference S212). "

on the superdiagonal a
except the 1:: at the (20, 1
computed as
If IA 11is much larger than all other eigenvalues, then we have
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)
We compute U = [B AB '.. A9 BJ and then compute its singular values.

The three smallest singular values are 0.712 x 10- 7 , 0.364 X 10- 9 , and
0.613 x 10- 12 . If we use a digital computer with a relative precision of 10-10
(or a number will be considered as a zero if its absolute value is smaller than
10-10), then the rank of U is smaller than 10, and we will conclude that the
equation is not controllable, although the equation is clearly controllable
following Corollary 5-21.
If we use the condition rank [sI - A BJ = n, for all eigenvalues of A, to
check the controllability of {A, B}, we must compute first the eigenvalues of A.
The QR method is a stable and reliable method of contputing the eigenvalues of
a matrix. However, the eigenvalues ofsome matrices can be very illconditioned.
lf e =0, the eigenvalue:

the migration of the eil
5-16, the root locus of
indicate the migration (
eigenvalues except Al a
i
11
This lnalysis was suggest(
Figure 5-16
The root
MICAL EQUATIONS
COMPUTAT10NAL PROBLEMS
~ rank of U can then be

(Appendix. E) which is a
e equation is large, this
and may transform the
mvenience in discussion,
nd B is an n x 1 vector.
Lrly, we can write B as
For example, consider the matrix

20
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
:ctor, Vi> as k increases.
large.
different argument. The
~~ IIAlbllA -lllz = (J/(J"

Les of A. It can be shown
is a very large number.
ix with a large condition
md shouldbe avoided.
controHability of {A, B}
If 6 =0, the eigenvalues of A are clearly equal to A = i, i = 1, 2, ... ,20. To see

the migration of the eigenvalues of A as 6 increases from O, we plot, in Figure
5-16, the root locus of Ll(s) (see Reference S46),u The heavy hnes with arrows
indicate the migration of the eigenvalues as 6 increases from O. We see that aH
eigenvalues except Al and AzO become complex if 6 is sufficiently large (actually
11
This analysis was suggested by Proressor D. Z. Zheng.
equation (see Reference
u
El
lpute its singular values.

)-7, 0.364 x 10- 9 , and
lative precision of 10-10
lte value is smaller than
re will conclude that the
1 is clearly controllable
. all eigenvalues of A, to
lrst the eigenvalues of A.
puting the eigenvalues of
m be very ill conditioned.
10
Figure 5-16
The root locus of ~(s).
11
220
CONTROLLABILITY ANO OBSERVABILITY Of LINEAR OYNAMICAL EQUATIONS
if e is larger than 7.8 x 10- 14 ). We list in the following Alo and Al! for sorne e:
el
7.8
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
They are computed directly from A by usingthe QR method. [Thecomputation

of the eigenvalues from fl(s) is not advisable because the characteristic poly
nomial may be more sensitive to 8 than the eigenvalues and the roots of a poly
nomial are very sensitive to the variations of the coefficient.J We see that the
eigenvalues are very sensitive to e. Thus the eigenvalues of A are very ill
conditioned. For this reason, the use of the criterion rank [sI - A BJ = n
to check the controllability of {A, B} may yield an erroneous result.
It turns out that the use of Theorem 5-16 is the best way of checking the
controllability of a state equation. Furthermore, if the equation is not control
lable, the computation also yields a reduced controllable equation. The proof
of Theorem 5-16 provided a procedure of computing the required equivalence
transformation. The procedure, however, uses the controllability matrix and is
not satisfactory from the computational point of view. In the following, we
shall introduce an efficient and numerically stable method to transform a state
equation into the form in (5-54). For the single-input case, the method is
essentially the procedure of transforming the matrix A into the Hessenberg
form (see Appendix A). We discuss in the following the general case.
Let PIbe an orthogonal matrix,12 that is, P l 1 = P'I' such that
x x x
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:''';
has rank nI + n 2 (whel

lable and can be red!
is transformed, by a se
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
where P2AWP~ has bl

if A~2d =0, the control
x x
and
(5-84 )
where pB\I) = nI' pA 2 ]

smallest integer such
A k ,k-l has dimension
UnI +n 2 + ... +nk=
By this process, the co:
The in (5-89) is in
formation of {A, B} ir
follows (see Reference
Step 1. P =Im Ao =A,
Step 2. Find an ortho~
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
The form of is identical t
COMPUTATlONAL PROBLEMS
ICAL EQUATIONS
tIa and ,1,11 for sorne E:
{AW,
B~l)}.
221
Ir AW f. O, we find an orthogonal matrix P 2 such that
P2AW=[A~2lJ
(5-85)
where AW is a n2 x nI matrix with n2 = rank AW and n2 snl' We compute

(5-86)
~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)
1 matrix, and so forth.
!~~l~--:~~J
C=CP'=[C l
(5-82)
.rank B. This step can

~rmations.13 We then
(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]
where pB~I)=n, pA 21 =n2' pA 32 =n 3, ... , pA k,k-l =nk' The integer k is the

smallest integer such that either nk=O or nI +n2 +- .. +nk=n. Note that
A k,k-l has dimension nk x nk-l and has a full row rank. Clearly, we have
p nI n2 n3 ... nk O
If nI +n2 + ... +nk = n, {A, B} is controlabie; otherwise, it is no controllable.

By this process, the controllability of {A, B} can be determined.
The in (5-89) is in the block Hessenberg form (see Appendix A). The trans
formation of {A, B} into the form in (5-89) can be carried out recursively as
follows (see Reference S203):
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
where.rankZ j = nj' Ir nj =0, go to step 7.

1 = Pf.
1
'artial pivating. In this case.

lble.
"']
~_- ~}j~y~- _- _~_- _- ~_-_-~-~-~-
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
The farin of is identical to the one in (A-17).
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.
Step 6. j = j + 1 and go to step 2.
Step 7. {A, B} is not controllable and can be reduced to an -dimensional
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
The matrix is said to

formo This is achieved
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
____
____
__
__
ollability of {A, B}; it

f (5-84) and (5-88) with
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
n AiB and l1i+ 1 is the

le set of controllability
. } [see Figure 5-7 or
/'ji, i = 1,2, ... }.
Jility indices. In order
A, B}, where A and B
ormed into
(
The matrix is said to be in the block companionform or the block Frobenius

formo This is achieved by choosing
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
rrom the structure

1, and f13 = 4. These
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 .. _ _
x:x x:x x:x

1:, x
0- -- -6 : -(
1
~.j
x:, 'x x:, x

- :-x- : :-x- - ~ -:-;
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
There are /l + 1 block

A and I and p columns
B columns. Now we:
left to right. Because
independent of their 1<
dependent B columns i
of the structure of T )l' .
(5-93)
Let /l be the least inte~

and
by the equivalence transformation 15

{):~ O O O O O O O
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 _
>
The positions of 1 in this matrix are determined from the positions of 1 in

(5-90) with the same encirclements. The transformation of {A, B} into (A, B}
can also be easily programmed without first determining the controllability
indices (see Reference SG). The form in (5-93) is said to be in the controllable
fotm and is very useful in the design of state feedback as will be seen in Chapter 7.
We mentian one more method of checking controllability to conclude this
section. Frmstatement 4 ofTheorem 5-7 and Theorem 0-8', we may conclude
that {A, B} iscontrollable if and only ir the polynomial matrices si - A and B
15
are left coprime. Sirr
This performs only permutations of columns and rows.
It can be shown directl

in (5-94) are the same 2
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
:reAL EQU.\ 11' ''''
into theform in (5-91).

lUter (see Reference S6).
:ssenberg form in (5-89)
block Hessenberg form
not numerically stable.
Ltion on columns of A
will be altered, and we
be chosen without any
(5-91) has the control
lese indices, the {A, B}
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
are left coprime. Similar to Theorem G-14, we form the matrix

-A
B:
O:
I O:, -A B:
,
0 0 '' 1 0 ' ," ,
T=
l'
[
,
O O :
:
'-.--'
(5-94 )
,
O O:
:'
'----v--'
(no. of dependent columns)
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)
Let /l be the least integer such that

O~rO~rl ~
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
... ~rl'-l <p

rl'=rl'+l ='" =p
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
rankU=rank[B AB ... AI'-IB]=(p-ro)+(p-r)+'" +(p-rl')

=total number oflinearly independent B columns in (5-94)
Consequently, we conclude that {A, B} is controllable if and only if the total
number of linearly independent B columns in (5-94) is n. Note that the linearly
independent columns of TI' are to be searched in order from left to right. The
position of the rows of TI" however, can be arbitrary altered. Bence, we may
apply Householder transformations or gaussian eliminations with partial
pivoting on the rows of TI' to transform TI' into an upper triangular form (see
Appendix A). Once in this form, the linearly independent B columns can be
readily determined.
There are two disadvantages in using TI' to check the controllability of
{A, B}. First, the size ofTI' is (2 +/l)n x (n +p)(.. +1), which is generally much
larger than the size of A. Second, ir{ A, B} is not controllable, its reduced
controllable equation cannotbe readily obtained. Bence, the transformation
of {A, B} into a block Hessenberg formseems to be a better method of checking
the controllability of {A, B}.
The discussion of the observability part is similar to the controllability
part and will not be repeated.
226
CONTROLLABILlTY AND OBSERVABILlTY Of LINEAR DYNAMICAL EQUATlONS
5-9
Concluding Remarks
In this chapter we have introduced the concepts of controllability and observa

bility. Various theorems for linear dynamical equations to be controllable and
observable were derived. We shall discuss briefly the relations among some of
these theorems. We list first those t'.leorems which are dual to each other:
Controllability:
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
Which of the followil
a. {t, t 2 ,e', e 2 ',te'}

b. {e', te', t 2 e', te", te 3 ']
c. {sin t, cos t, sin 2t}
5-2
b.
Check the controllal
x=[~
-2
2~ l~l
-25
) = L --
J.
-20-,
J Ji.
5-3 Show that a linear dyr

finite ti> to such that for an
Hint: Use the nonsingularil
5-4 Show that ir a linear,
at any t < too ls it true tha
controllable at any t> t o?
the I!erature, ifx o =0, it
encompasses both and doe:
abilit)i and die condition o
identical.
"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.
ollability and observa

to be controllable and
lations among sorne of
ual to each other:
-8
51-16 lrheorem
,-14
5-17
227
5-10
Problems
ierived from the con

.uality), and vice versa.
results of this chapter.
r), and hence they are
continuous difTerenti
-5 (or 5-11), which give
le have the analyticity
ne-varying dynamical
12). Theorem 5-7 (or
5-4, gives the necessary
lamical equation to be
5-1
a.
Which of the following sets are linearly independent over ( - ro, ro)?
{t, e2 , el, e 2l, te l }
b. {e', te', t 2 e', te 2 ', te 3t }

c. {sin t, cos t, sin 2t}
5-2
Check the controllability of the following dynamical equations:
a. [~J=[~
y=[O
ix and the linear time

ter. This was achievd
;: (1) controllable and
trollable and unobser
ansfer-function matrix
h.e dynamical equation.
controllable and not
rvable one.
;:ssential in the study of
1 of a rational matrix
)ter 8). Sorne practica!
b.
:I::l+[~lu
lJ
[:J
x=[ ~ ~ ~Jx+[ ~ ~]u

-2
-4
-3
-1
y=[~ ~ -~lx
c.
=[~O 2~ I~J x + [- ~l u
-25
y = [-1
-20
3 OJ
~J
bility and observability

can be stated nicely in
-A
B] in the con
Jutations. An efficient
il a dynamical equation
observability can then
~ a reducible dynamical
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
Show that if a linear time-invariant dynamical equation is control1able, then it is

uniformly control1able.
5-6
Check the observability of the dynamical equations given in Problem 5-2.
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
Check the contro11ability of the fo11owing state equations:
a. x=[~ ~Jx +CJu

[~ ~J x +L~'] u
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?
5-17 Show that the state

equation can be determl
11 -1 order. [Hin!: Com
Show that controll;
tions are invariant under a
for a11 t and continuously
5-18
5-19
~J x +L~2,] u
Reduce the dynami,
to a controllable one.
Check the control1ability and observability of
~ ~J +[~ ~J
x
y = [1
-1
5-20
Reduce the equatior
5-21
Reduce the fol1owin.
O O
O 1J x
by using statement 3 ofTheorems 5-7 and 5-13.

What are the contro11ability indices and the observability index of the equation in
Problem 5-1O?
5-11
Compute the contro11ability indices and the contro11ability index of the state equa
tion in (3-48).
5-12
J.
to a contro11able and obser

5-22
5-13
x=Ax +bu
where A is an 11 x 11 matrix and b is an 11 x 1 column vector. What is the equivalent dy

namical equation if the basis {b, Ab, ... , An-1 b} is chosen for the state space? Or, equi
valently, if i = Px and if p-1 = [b Ab ... A"-l bJ, what is the new state equation?
Find the dynamical equations for the systems shown in Figures 5-3, 5-4, and 5-1\
and check the contro11ability and observability of these equations.
5-14
5-15
Consider the n-dimer
Given a COl1trollable linear time-invariant single-input dynamical equation

The rank of its controllabili
is assumed to be n1 n). 1
independent columns ofU.
n1 x n1 unit matrix. Show

is control1able and is zero-st
Hint: Use Theorem 5-16.
In Problem 5-22, the
Find a method to solve P 1 ir
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
Show that controllability and observability of linear time-varying dynamical equa

tions are invariant under any equivalence transformation x = P(t)x., where Pis nonsingular
for all t and continuously differentiable in 1.
5-18
5-19
Reduce the dynamical equation

y=[1
lJx
to a controllable one.
5-20
Reduce the equation in Problem 5-19 to an observable one.
5-21
Reduce the following dynamical equation
.~
lity index of the equation in
l'
O
1
}'l
O A
O O
O O
5-22
in Figures 5-3,5-4, and 5-11

ions.
},[~}
to a controllable and observable equation.

Consider the n-dimensional linear time-invariant dynamical equation
F ; .
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
i1ity index of the state equa
O
O
O
y. = Av" + 131.1
y=Cx+Eu
The rank of its controllability matrix,

U=[B: AB:"': A"-BJ
is assumed to be n n). Let Q be an n x n matrix whose columns are any ni linearly
independent columns ofU. Let Pibe an ni x n matrix such that P Q = 1,,\, where 1"1 is the
n x n unit matrix. Show that the following ni-dimensional dynamical equation
FE:
x =PAQx + PBu
y=CQx +Eu
is controllable and is zero-state equivalent to FE. In other words, FE is reducible to FE.

Hint: Use Theorem 5-16
5-23 In Problem 5-22, the reduction procedure reduces to solving for p in pQ = I,,.
Find a method to solve P in P Q = 1"1
230
Develop a similar statement as Problem 5-22 for an unobservable linear time-
invariant dynamical equation.
5-24
5-25
5-32
Consider two SYSl
Is the following Jordan-form dynamical equation controllable and observable?
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
are they output controlla
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"
wh\?re Iinearly independent

Hint: Premultiply T~ by
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
In Corollary 5-22 we have that a dynamical equation with E = O is output

controllable if and only if p[CB: CAB: . --: CA" - lB] = q. Show that if E F O, then the
dynamical equation is output controllable ir and only ir
5-30
p[eB : CAB: ... : CA" - lB: E] = q

5-3i Consider a linear time-invriant dynamical equation with E =0. Under what
condition on e will the (state) controllability imply output cntrollability?
Show that {A, C} is o

are, respectively, n x n and q
pose ofe.
5-36
5-37 Showthat{A, B}isc

that is orthogonal to aH coll
left eigenvector O( of A such ti
This is called the Popov-Befe

of Theorem 5-7.)
.... _..--:..:=.
PROBLEMS
CAL EQUATIONS
Inobservable linear time-
5-32
Consider two systems with the transfer-function matrices
G[--rS~11
able and observable?
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)
are they output controllable? utput function controllable?

u
the foHowing Jardan-form
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
Let P be a nonsingular matrix such that
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
...
A"-IB] = total number of linearly independent B columns ofT"

in (5-94)
where linearly independent columns of TI' are to be searched in order from left to right.
Hint: Premultiply TI' by
)5
re exists a vector b such that
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.
there exists a p x 1 col uml1
5-21.)
n n x p matrix, to be control
ks associated with the same
ation with E =0 is output

Show that if E ~O, then the
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
Network synthesis is one of the important disciplines in electrical engineering.

lt is mainly concerned with determining a passive or an active network that
has a prescribed impedance or transfer function. The subject matter we shall
introduce in this chapter is along the same line-that is, to determine a linear
time-invariant dynamical equation that has a prescribed rational transfer
matrix. Hence this chapter might be viewed as a modern version of network
synthesis.
A lillear time-invariant dynamical Aequation that has a prescribed transfer
matrix G(s) is called a realization of G(s). The term "realization" is justified
by the fact that, b)' using the dynamical equation, the system with the transfer
function matrix G(s) can be built in the real world by using an operational
amplifier circuit. Although we have proved in Theorem 4-10 that every proper
rational matrix has a finite-dimensional linear time-invariant dynamical
equation realization, there are still many unanswered questions. In this chapter
we shal1 study this and other related problems.
We study the realization problem for the following reasons: First, there are
many design techniques and many' computational algorithms developed ex
clusively for dynamical equations. In order to apply these techniques and
algorithms, transfer-function matrices must be realized into dynamical equa
tions. Second, in the design of complex system it is always desirable to simulate
the system on an analog or a digital computer to check its performance before
the system is built. A systemcannbt be simulated efficiently if the transfer
232
function is the only a\

tion is obtained, by a~
system can be readily
operational amplifier (
link between the state
For every realizat
number of linear timt
a major problem in n
that a dynamical-eqm
good realization. We
dimension must be a c(
if a linear time-invaria
if the equation is uncor
5-19 it is possible to rec
stil1 has G(s) as its trar
if the equation is contr
realization of (;(s) with
vable dynamical equati
Such a realization is ,
In this chapter we study
(1) A rational transfer
observable pan of a dy!
an irreducible one. (2)
network, the number of.
reasons of economy an
found, any other irredUl
lence transformation (T
This chapter is orga
cept of the degree for prt
in Theorem 6-2. In Se(
rational functions. H
methods are then exten
different irreducible rea
the other on coprime f1
Sections 6-5 and 6-6. Ir
tion, cal1ed polynomial rr.
relationships with trans
lished. In Section 6-8 th
to strict system equivaler
under the coprimeness ,
have the same transfer rr
we study the identificatic
pairs. The concept of P'
This chapter is baseG
98, 115, S27, S48, SS2, S
realization of impulse~re~
to References 32, 100, 10
=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
_g reasons: First, there are

algorithms developed ex
'ply these techniques and
zed into dynamical equa
ways desirable to simulate
~ck its performance before
i efficiently if the transfer
233
function is the only available description. After a dynamical-equation realiza

tion is obtained, by assigning the outputs of integrators as state variables, the
system can be readily simulated. The realization can also be built by using
operational amplifier circuits. Final1y, the results can be used to establish the
link between the state-variable approach and the transfer-function approach.
For every realizable transfer-function matrix (;(s), there is an unlimited
number of linear time-invariant dynamical-equation realizations. Therefore
a major problem in the realization is to find a "good" realization. lt is c1ear
that a dynamical-equation realization with the least possible dimension is a
good realization. We c1aim that a realization of (;(s) with the least possible
dimension must be a control1able and observable dynamical equation. lndeed,
if a linear time-invariant dynamical-equation realization of (;(s) is found, and
if the equation is uncontrol1able or unobservable, then fol1owing from Theorem
5-19 it is \?ossible to reduce the realization to a lesser-dimensional equation that
still has G(s) as its transfer-function matrix. This reduction is impossible only
if the equation is control1able and observable. Therefore, we conclude that a
realization of (;(s) with the least possible dimensio is a control1able and obser
vable dynamical equation, or equivalently, an irreducible dynamical equation.
Such a realization is cal1ed a minimal-dimensional or irreducible realization.
In this chapter we study mainly irreducible realizations for the following reasons:
(1) A rational transfer-function matrix describes only the control1able and
observable part of a dynamical equation; hence a faithful realization should be
an irreducible one. (2) When an irreducible realization is used to synthesize a
network, the number of integrators needed will be minima\. This is desirable for
reasons of economy and sensitivity. Note that if an irreducible realization is
found, any other irreducible realization can be obtained by applying an equiva
lence transformation (Theorem 5-20).
This chapter is organized as fol1ows. In Section 6-2, we introduce the con
cept of the degree for proper rational matrices. 1ts significance is demonstrated
in Theorem 6-2. In Section 6-3, various realizations are introduced for scalar
rational functions. Hankel theorem is also introduced. The realization
methods are then extended to vector rational functions in Section 6-4. Two
different irreducible realization methods, one based on Hankel matrices and
the other on coprime fractions, for proper rational matrices are discussed in
Sections 6-5 and 6-6. In Section 6-7 we introduce a new mathematical descrip
tion, cal1ed polynomial matrix description, for linear time-invariant systems. lts
relationships with transfer functions and dynamical equations are also estab
lished. In Section 6-8 the concept ofequivalent dynamical equation is exteDded
to strict system equivalence for polynomial matrix description. lt is shown that,
under the coprimeness assumption, al1 polynomial matrix descriptions which
have the same transfer matrix are strictly system equivalent. In the last section,
we study the identification of discrete-time systems from arbitrary input-output
pairs. The:concept of persistent exciting is introduced.
This chapter is based mainly o.n references 15,42,47,60,62,67,68,83,89,
98, 115, S27: S48, S52, S126, S158, SI61,SI85, S187, S209; and S218. For the
realization ofimpulse-response matrices G(t, T), the interested reader is referred
to References 32, 100, 101, 114, and S128.
234
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICAnON
THE CHARACTERlSTIC F
Ir FE 1 is a realiz,
6-2 The Characteristic Polynomial and the Degree

of a Proper Rational Matrix
In this section we shall introduce the concepts of degree and characteristic
polynomial for proper rational matrices. These concepts are the extension of
the denominator of a proper rational function and its degree to the matrix
case. Consider a proper rational function g(s) = N(s)/D(s). It is assumed that
N(s) and D(s) are coprime (have no nontrivial common factor). Then the
denominator of g(s) is defined as D(s), and the degree of g(s) is defined as the
degree of D(s). Without the assumption of coprimeness, the denominator and
the degree of g(s) are not well defined. In the following, we use det, deg, and dim
to stand, respectively, for the determinant, degree, and dimensiono
Let the single-variable linear time-invariant dynamical equation

X = Ax + bu
y =cx +eu
be a realization of the proper rational function g(s). Then FE 1 is irreducible

(controllable and observable) if and only if
det(sI - A) = k[ denominator of g(s)]
(6-1 )
dim A = deg g(s)
or
where k is a nonzero constant.
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
With this theorem

dynamical equation (
function. It is desira
case. The answer is a
matrix.
Theorem 6-1
FE 1:
It is easy to show th;

+ N Ils) are coprime.
if and only if
(sI - A) and let

c(sI _ A)-I b~ N I(S)
-
(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
~1(S) ~c(sI _A)-1 b =c(sI -A)-lb = N I(S)

Ll(s)
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.
The minors of order

The minor of order 2
(;1(S) is s +1 and b(;
1/(s + 1), 1/(s + 1), and
Hence the characteristi
From this example
general different from
lItis also caBed lhe McMilL
only lo proper ralional mal
thepoles .al s = oo. See Re
THE CHARACTERISTIC POLNOMIAL AND THE DEGREE Of A PROPER RATlONAL MATRIX
:, AND IDENTIFICAnON
235
lf FE 1 is a realization of g(s), then we have
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
Denominator of g(s) =L'1(s) = det(sI - A)

deg g(s) = deg L'1(s) = dim A
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
With this theorem, the irreducibility of a single-variable linear time-invariant

dynamical equation can be easily determined from the degree of its transfer
function. It is desirable to see whether this is possible for the multivariable
case. The answer is affirmative if a "denominator" can be defined for a ratianal
matrix.
'hen F El is irreducible
Definition 6-1
:)]
(6-1 )
The characteristic polynomial of a proper rational matrix (;(s) is defined to be

the least common denominator of aH minors of (;(s). The degree of (;(s),
denoted by 15(;(5), is defined to be the degree of th.e characteristic polynomial
of (;(S).1
. '
Example 1
Consider the rational-function matrices

--
Gl(S)=.s~l
(6-2)
f L'1(s) and N I(S) are co

heorems 5-16 and 5-17
1 A and
N I(S)
- L'1(s)
A = deg L'1(s), we cQn

mg the above argument,
1 {A, b, e} is not irred.u~ .
rollable and observable'
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.
From this example, we see that the characteristic polynomial Of(;(5) is in

general dilTerent from the denominator of the determinant of (;(s) [if (;(s) is a
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
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTIFICAnON
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
Let the multivariable
(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
Note that in computing the characteristic polynomial of a rational matrix,

every minor of the matrix must be reduced to an irreducible one as we did in
(6-3); otherwise we will obtain an erroneous result.
In the following, we introduce a different but equivalent definition of the
characteristic polynomial and degree of a proper rational matrix. This defini
tion is similar to the scalar case and requires the concepts developed in Appendix
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
Irreducible real iza1

study first the transfer
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)
where P and (Xi, for i'

coefficient of D(s) is 1.
or, in the time dornain
and thedegree of (;(s) is defined as

deg (;(s) = deg det D,.(s) = deg detD(s)
where deg det stands for the degree of the determinant.
* Maybe skipped without
loss of continuity_
where pi stands for di/

- known that -in an nth
solution for any u, we
vector will consist of n (
IRREDUCIBLE REALIZATION OF PROPER RATIONAL FUNCTIONS
, 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.
lenominator of all the

Lcteristic polynomial of
*Theorem 6-2
Let the multivariable linear time-invariant dynamical equation
FE:
be a realization of the proper rational matrix (;(s). Then FE is irreducible

(controllable and observable) if and only ir
:hree minors of order 2.

1
: + 1)(s +2)
or
(6-3)
det (sI - A) = k [characteristic polynomial of G(s)]

dimA=degG(s)
where k is a nonzero constant.
11
The irreducibility of the realizations in this chapter for the multivariable

case will be established by using the conditions of controllability and observa
bility without relying on this theorem. Hence this theorem will not be proved
here. In fact, this theorem will be a direct consequence of the irreducible
realization discussed in Section 6-6. 1t is stated in this section because of its
analogy to Theorem6-1 and its bringing out the importance of the concepts
of the characteristic polynomial and degree of G(s).
3)
t- 2)(s
x=Ax +Bu
y=Cx +Eu
+ 3) and I>G(s) =4.
ial of a rational matrix,
ucible one as we did in
6-3
valent definition of the

lal matrix. This defini
;developed in Appendix
Irreducible Realizations of Proper Rational Functions
Irreducible realization of (3/D(s). Before considering the general case, we

study first the transfer function
g(s) =
A
/)
s" +as"-
,
~_.
L> _ _
+ ... +all_s +a ll = D(s)
(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
G(s) = N.(s)D.- ' (s) =

. coprime, and D(s) and
(lial of (;(s) is defined as
lS
D(s)y(s) = (3u(s)
(6-5a)
or, in the time domain,

(p" +aIP"- I
+ ... +all)y(t) = (3u(t)
(6-5b)
where pi stands for i/de. This is an nth-order differential equation. It is well

known that in an nth-order differential equation, in order to have a unique
solution for any u, we needn number of initial conditions. Hence the state
vector wi\l consist of n components. In this case the output y and its derivatives
238
IRREDUCIBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTlFICATION
up to the
(n -
l)th arder qualify as state variables.
XI(S~
x
J [1
y(s)
A()D. 2(s) D.
sy(s)
=
=:
=:
.
.
Xn(s)
Sn-y(S)
X S
Define
IF
(6-8). To show that (

to Figure 6-1 to ShOVi
equal to g(s). A diffe
A
y(s)
(6-6a)
where
S"-
X(t)~ y(t)
X2(t)~y(t) = py(t) = x(t)
X3(t)~ y(t) = p2y(t) =X2(t)

Xn(t)~ y<n-
(6-6b)
It is c1ear that (sI - A

valently, to the cofact
I)(t) = pn - 1 y(t) = Xn_ (t)
Differentiating xn(t) once and using (6-5) we obtain

xn(t) =pny(t) = -(XnX -(Xn-X2 - ... -(XX n +f3u
(see Problem 2-26). 1

Of(f3/l(s))(sI-A) can
These equations can be arranged in matrix form as
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
The blockdiagram of Equation (6-8).
This verifies that (6-8)
dim A, (6-8) is an irredl
, by showing that (6-8)
obserVable. Note tha
coefficients ofg(s) in (6
_-_
....
...
...
--_.... ,_._
.;;.:...:::...~..::.::.:.:.:.::::......,,_.- - '' : : " : : : " - ~ _ . _ - - - - ~ - , - - - - - _ .
~-_
_-~-_
239
IRREDUCIBLE REALIZATIONS OF PROPER RATIONAL FUNCTIONS
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
g(s) =c(sI - A)-l b
(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)
rom (6-6b). They are

ld are independent of
:omes into (6-8a) only
12 a block diagram of
o"
...
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
sn +Ct1s n- 1 + ... +Ct n
(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
IRREDUCIBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE AND lDENTIFICATION
11
!
!
IrreduciblR realization of g(s)=N(s)/D(s).

proper transfer function
Consider the following scalar
(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
gl(S)= sn + a sn-l + . . . + a - s + a +e=g(s) +e

n 1
1
n
(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
in the right-hand sic

remainder gives the r
coefficients associatet
u a unique y can be d
as
Xn(t)~y(t)
X n -1(t)~y<
X n _ Z(t)~
X 1(t)~
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
The foregoing equatio
Observable canonical-form realization
Consider the nth-order differential equation D(p)y(t) = N(p)u(t). It is well knowh

that if we have n initial conditions-for exam pie, y(to), y(l)(t o), ... , y<n - ll(t 0)
then for any input UUO.l,j' the output Y[CO,I,j is completely determinabk. in tbi:
case, however, if we choose y(t), y(l)(t), . .. , y<n-l)(t) as state variables as we did
in (6-6), then we cannot obtain a dynamical equation of the form x =Ax+ bu,
y = ex. Instead, we will obtain an equation ofthe form x = Ax + bu, y = ex +elU
+ezu(l) +e3ulZ) + .. '. Hence in order to realize N(s)/ D(s) in the form
x = Ax + bu, y = ex, a different set of state variables has to be chosen.
Taking the Laplace transform of(6-15) and grouping the terms associated
with the same power of s, we finally obtain
y(s) = N(S)) u(s) +~) {y(O)sn- 1 + [y< 1l(0) +a y(O) _ t11 U(O)]Sn - z + ...
D(s
D(s
.
.
n
1
n
2
3
+ [i - l(0) +ain~Z)(o) - f3ul - )(0) +adn - ){0)
-f31uln-3)(0)+'" +C(n-ly(O)-f3n~IU(O)]]
(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), \
.=.-=... =.
__ .=~=
.. =;=="..=..=~_.=.-.=.-=
... = = = = = . =......-.... =-=.:..-=.=..
~~=~
=.
IRREDUCIBLE REALIZA nONs OF PROPER RA T10NAL FUNcnONs
"ND IDENTIFlCATlON
. the following scalar
241
The term
N(s)
--) u(s) =g(s)u(s)

D(s
(6-12 )
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)
nmediately the direct

the following only the
(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
- f3dl)(t) +:Xzy(t) - f3zu(t)
(6-17)
XII]' qualifies as the state vector. The set of equations

y=X"
Then we have
X" +CI.x,,- f3u

x,,_z=x,,_ +<X zx -f3 zu
X,,_I =
(6-15a)
lI
Differentiating x I in (6-17) once and using (6-15), we obtain

)II-Z + ... +/3",and pi
.s 9(S) in (6-14). In the
;-15).
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
The foregoing equations can be arranged in matrix form as
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
-..-~"-z~ 1~3 I1 ~"-z1

-Cl. 1I
Xz
x,
-o:z
-o:
/311-
P.
1+
JlX"-J
XII
,u
lf3z
/3
]x
(6-18b)
A dynamical equation in the form of (6-18) is said to be in the observable

canonical formo The block diagram of (6-18) is shown in Figure 6-2. Since
(6-18) is derived from (6-15), the transfcr function of (6-18) is 9(S) in (6-14).
This can be verified by either applying to Figure 6-2 Mason's formula for signal
flow graph or computing
e(sI - A)- 1 b = [e(sI - A)- lb]' = b'(sI -::- A')- le'
- 3)(0)
(6-16)
(6-18a)
with the aid of (610), where the "prime" symbol denotes the transpose.
242
IR.
IRREDUCIBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE, AND lDENTIFlCATION
Controllable canonie
We shall now introdu(

form realization, of g(.
a new variable v(t) del
and
Equation (6-21) has t
variables as, similar te
Figure 6-2
x(,)J:~i:~~
Block diagram of the observable canonica\-form dynamica\ equation (6-18).
The observability matrix of (6-18) is
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)
which becomes, in the

O
1
where x denotes possible nonzero elements. The matrix V is nonsingular for

any (Xi and f3i; hence (6-18) is observable no matter D(s) and N(s) in (6-14) are
coprime or not, and is thus called the observable canonicalform realization.
The dynamical equation in (6-18) is controllable as well if D(s) and N(s) are
coprime. lndeed, if D(s) and N(s) are coprime, then deg 9(S) = deg D(s) = dim A
and (6-18) is controllable and observable following Theorem 6-1. If D(s) and
N(s) are not coprime, the observable dynamical equation (6-18) cannot be con
trollable as well; otherwise, it would violate Theorem 6-1. In the following we
shall show directly that if D(s) and N(s) are not corprime, then (6-18) is not
controllable. Let s - A be a common factor of N(s) and D(s); then we have
and
D(A)=An+!XIA n- I + ... +!Xn-IA+!Xn=O

N(A)=f3iA n- 1 +f32 An - 2 + ... +f3n=O
Using (6-23), Equation
These equations can
be
(6-19)
(6-20)
Define the 1 x n constant vector lX as lX~ [1 A A2 . . . An-lJ. Then (6-20)

can be writtenas lXb=O, where b isdefined in (6-18a). Using(6-19), it is easy to
verify that lXA =AlX, lXA 2 =A 2 lX,
, lXAn- 1 =An-1lX. Hence the identity lXb =0
implies that lXAib =0, for i = 0,1,
, n - 1, which can be written as
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
IRREDUCIBLE REALlZATlONS OF PROPER RATIONAL FUNCTlONS
AND IDENTIFICATION
243
Controllable canonical-form realization
We shall now introduce a different realization, called the control1able canonical

form realization, of g(s) = N(s)D- 1 (s) 01' .v(s) = N(s)D- 1(s)U(s). . Let us introduce
a new variable v(t) defined by ves) = D- 1(s)(s). Then we have
D(s)v(s) =(s)
.v(s) = N(s)v(s)
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'
This definition implies Xl =X l , Xl =X3'" . 'X n- I =Xn- From (6-21), we have
SXn(s)=snv(s) = -a1Sn-1v(s)-alsn-lJ(s)_'" -a 1v(s) +(s)

=[-a n -a n- 1
-al -a1]X(S)+U(S)
which becomes, in the time domain,

x
x
(6-24)
Using (6-23), Equation (6-22) can be written as
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',
in the time domain,

y(t) = [.8"
.8" - 1 .. , .8 l]X(t)
(6-25)
These equations can be arranged in matnx form as
(6-19)
X=Ax +bu
(6-20)
y=cx
(6-26)
where
. .1.,,-1]. Then (6-20)

sing (6-19), it is easy to
nce the identity cxb = O
: written as
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"
IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTlFICATION
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
Consider the proper irreducible transfer function

3
2
A(S) = 4s + 25s +45s + 34
g
2s 3 + 12s 2 + 20s + 16
Realization from the Ha
Consider the proper rati(
We expand it into an inf

9(S
The coefficients {hU), i =

These parameters can be
h(O) = {Jo
h(l) = -(1.
h(2) =
By long division, 9(S) can be written as

A)
0.5s 2 +2.5s + 1
g(s = S3 +6s2 +10s +8 +2
h(n) =
-(1.
+0 =
-(1.
h(n
Hence its controllable canonical-form realization is
l:}U -1~ -l}l~}
y=[
2.5
0.5J x +2u
i= 1, 2
These equations are obt
(f3 os" + f3 1s" -
h(l)
h(2)
X~[! ~ -~~]x +(~.5]U
y=[O
-6
lJ x
+ . .. + f3
and then equating the c

matrix
Its observable canonical-form realization is
-(1.
0.5
H(cx, f3)~ h\3)

[
+ 2u
h(cx)
lt is caJled a Hanke/ me
{hU), i = 1, 2, 3, ... }.
invo/ved in H(1., f3).
Theorem 6-3
The proper transfer funl
pH(n, n)=pH
where pdenotes the rar
Proof
Figure6-3
(6-26).
Block diagram of the controllable canonical-form dynamical equation
We show first that if del

lfdegg(s)=n, then(6-3C
IRREDUCIBLE REALIZATIONS OF PROPER RATIONAL FUNCTIONS
'CE, AND IDENTIFICATION
ore are no simple relation

(6-26) is always control
tot and is said to be in the
me, then (6-26) is observ
rtion is dual to that of the
(ercise. A block diagram
245
Realization from the Hankel matrix
Consider the proper rational function
/3 os + /31 s - 1 + . .. + /3.
A()
9s =
s +al s
+ ...
(6-27)
+a.
We expand it into an infinite power series of descending power of s as

g(s)=h(O) +h(l)s-1 +h(2)s-2
+ ...
(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(l) = -ah(O) +/31

(6-29)
h(2) = -a 1h(1)-a2h(O) +/32
h(n
h(n)= -a 1h(n-l)-a 2h(n-2)- ... -a.h(O)
+/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
+ ...
+a.)(h(O) +h(l)s-1 +h(2)s-2
+ ...)
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(a, /3)~ h~3)
h(4)
h(5)
h(/3
+ 2)
h(a
+/3 -1)
]
(6-31 )
[
h(a)
h(a
+ 1)
h(a
+ 2)
lt is called a Hankel matrix of order a x {3. It is formed from the coefficients

{hU), i = 1, 2, 3, ... }. lt is important to note that the coefficient h(O) is not
involved in H(a, /3).
Theorem 6-3
The proper transfer function g(s) in (6-27) has degree n if and only if
pH(n, n)=pH(n +k, n +l)=n
for every k,l = 1,2,3, . . .
(6-32)
where pdenotes the rank.
Proof
\-rorm dynamical equation
We showfirst that if deg g(s) = n, then pH(n, n) = pH(n + 1, IX)) = pH(oo, (0) = n.
If degg(s) = n, then (6-30) holds, alld n is the smallest integer having this property.
246
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIV ALENCE, AND IDENTIFICATION
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
+ ...
/3os +/3I S-1 + ... +/3.

- s +cxls- 1 + ... +cx.
IR
Note that there is on

up to h(2n) are used ir
algorithm 3 discussed i
Hin (6-35) in order fn
pendent and the (O" + 1
Then Theorem 6-3 iml
dependent on their pre
Iinearly dependent ro\\
cal! the (O" + l)th row
(a + k)th row, k = 2, 3,
D(s) and N(s) are coprir
ing algorithm wilI also
[al
This equation expresse

combination of its pre'
dependent row. Note
we do not have a = (J.._
Since the n is the smal!est integer having this property, we have deg g(s) = n.
Q.E. D.
with
x=Ax +bu
FE:
lts transfer function is clearly equal to

9(s) = e +c(sI - A)-I b =e +s-lc(I -S-I A)-I b
+cA bs-
+...
(6-33)
From (6-28) and (6-33), we conclude that {A, b, c, e} is a realization of g(s) in

(6-27) if and only if e =h(O) and
h(i)=cAi-lb
i=1,2,3,...
(6-34)
With this background, we are ready to introduce a different realization.

Consider a proper transfer function 9(s) = N(s)/D(s) with deg D(s) = n. Here
we do not assume that D(s) and N(s) are coprime; hence the degree of g(s) may
be less than n. We expand 9(s) as in (6-28) by using the recursive equations in
(6-29) and (6-30). We then form the Hankel m'atrix
h(l)
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 ,
h(tJ-i- i)= -'Qlh(a
Using this, we can read
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)
H(n +1, n)=

h(n)
f(n + 1)
-al
c=[
which can be expanded as, by using (2-85),

9(S) =e +cbs- l +cAbs-
1
O
h[ O
y =cx +eu
O
O
For eomputer eOmpul'l.!ion.
:E, AND IDENTIFICATION
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
pH(n, 00) = pH(oo , 00) =
)) 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)
+ i)= -alh((J + i -l)-azh((J+i -
(6- 37)
Because of (6-36) and
2)- ... - a"h(i)
i= 1,2,3, ...
Using this, we can readily show

Ab
=[~i~~
j, =[~;~: l. . , ~[~~~ :~n, . .
h((J + 1)
AZb
A'b
h((J +2)J
(6-39)
h(k +(JJ
The effect of the multiplication of A simply increases the atgumnt in h(i) by

1, or equivalently, shifts the elements up by one position. Because ofthe form of
e, cAkb just picks up the first element of Akb as
h(n)
h(n
O
O
is a controllable and observable realization of g(s).

h((J
(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
y, we have deg g(s) = n.

Q.E.D.
+ ...
247
IRREDUCIBLE REALIZATION OF PROPER RATIONAL FUNCTIONS
+ 1)
eb =h(I),
(6-35 )
eAb =h(2), eAzb =h(3),
(6-40)
h(2n -1)
h(2n)
For computer compullttion, numerically stable methods should be used. See Appendix A.
248
IRREDUCIBLE REALlZATlONS, STRlcr SYSTEM EQUlVALENCE, AND IDENTIFlCATlON
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
We note that this

nominator of g(s) are
smaller than the degrel
Dual to the introd
pendent columns of H(
irreducible realization.
Jordan-canonical-fo
O O
Clearly {A, e} is observable. Hence (6-37) is an irreducible realization of g(s),
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
The block diagrams of

ficients el!, e lZ , e 13 , ez,
We form the Hankel matrix
:- -6-;
_1-
, Z
_1- _J.
" Z ,
4
H(S, 4) =
,--
_1
_1-
_2
_1-
_2
_J.
1-
1-
,__ 1-1
Z
J.
U
~.
Using the row searching algorithm discussed in Appendix A, we can readily

show that the rank of H(S, 4) is 3. Hence an irreducible realization of g(s) has
dimension 3. The row searching algorithm also yields
kH(S, 4)~[~ :I~ 0JH(S,4)=0
Hence an irreducible realization of g(s) is
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).
Figure 6-4 Two block di;
IRREDUCIBLE REALlZATION OF PROPER RATIONAL FUNCTIONS
, 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
We use an example to illustrate the procedure to realize a transfer function into

a Jordan-form dynamical equation. The idea can be easily extended to the
general case. Assume that D(s) consists of three distinct roots Al, ..1. 2 , and ..1. 3 ,
and assume that D(s) can be factored as D(s) = (s - ..1. 1)3(S - . 1. 2 )(s - ..1. 3), We also
assume that 9(S) can be expanded by partial fraction expansion into
;ible realization of 9(S).
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),
:ndix A, we can readily

le realization of 9(s) has
1=0
(a)
e13 \ - - - - - - - - - - - - - - - ,
XlI=AXI\+el\u
x12=Axi+e12u
x13=Ax13+ e 13 u
Xl!
fthe first three elements

. the first three Markov
ixed and is independent
I
(b)
Figure 6-4 Two block diagniffis of g(5) in Equalion (6-41).
. _=._=-._="=---=-=
..._=._=-. -'='-=-'-=-=--='--=-'====-.=--=.--=--=-=---== ... --
~~"~.
_.~.-=.=.-~._=.-_.=-_.=--=_
250
--C'=~="-,"=c----c-.=c<=="--:=--==_=-------='=='C~'O'
.._..... ---- --- ..--_._..... _......-_ .. :'
...... - _ .....
11
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTlFICATION
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
_~_ _~l ;~ 1.. ~ _, 6 x +
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
where Al is the Jordar

(no transpose) of Al'
introduce the equivale
1] x
which is another irreducible Jordan-form realization of 9(5). Note the dif

ferences in the assignment of the state variables in Figures 6-4(a) and 6-4(b).
There are two difficulties in realizing a transfer function into aJordan canoni
cal formo First the denominator of the transfer funcUon must be factored. o.
correspondingly, the poles of the transfer function g(5) must be computed. This
is a difficult task if the degree of g(5) is larger than 3. Second, if the transfer
function has complex poles, the matrices A, b, and e will consist of complex
~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
Consider the Jordan-fo
11 I
l\'
1
x=
I
I
Figure 6-5 Internal structure of block b/(s - a).
y=[
1 +2i
O
O
O
O
1
1 +2i :
O ,'
-
O
'0
-i
- l.
,:
.
IRREDUCIBLE REALlZATION I Of PROPER RATIONAL FUNCTlONS
251
, ANO IDENTIFlCATlON
)Iock in Figure 6-4 can

~igure 6-5. Hence the
lssigning the output of
i, we can readily obtain
lS shown. By grouping
numbers. In this case. the equation cannot be simulated on an analog com

puter, for complex numbers cannot be generated in the real world. However,
this can be taken care of by introducing sorne equivalence transforrnation, as
will be demonstrated in the fol!owing. Since all the coefficients of g(s) are
assumed to be real, if a complex number A is a pole of g(s). its complex conjugate
l., is also a pole of g(s). Hence in the Jordan-form realization of g(s), we have the
following subequa tion:
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)
re is one Jordan block

controllable (Corol!ary
11 =0, ez =0, or e3 =0.
:ible realization of the
. the state variables are
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
of g(s). Note the dif

Hes 6-4(a) and6-4(b).
on intoaJordancanoni
on must be factored. 01'
lUSt be computed. This
Second. if the transfer
will consist of complex
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
Consider the Jordan-form equation with complex eigenvalues.

"
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
IRREDU('IBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATION
by choosing y(t), y(t),

exactly the same form
(6-46) consists of the
dynamical equation, t:
general formula, we gi
Let X = Px. where
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
Consider the followin
Then t6-45) can be transformed into
[p2
+ o:((
Tbe procedure of f<

We first assume that (E
y
I
whose coefficients are all real.
A remark is in order regarding these realizations of y(s). Clearly, the con

trollable canonical form and observable canonical form realizations are the
easiest to obtain. These realizations. however. are not necessarily irreducible
unless the given 9(s) = N(s)/D(s) is known to be irreducible [N(s) and D(s) are
coprime]. The realization obtained from the Markov parameters and the
Jordanform realization are always controllable and observable no matter the
given 9(S) is irreducible or not. Because of the requirement of computing the
poles of 9(S). the Jordan-form realization is generally more difficult to compute.
The Jordan-form realization however is. as discllssed in References 23, 81
and 87, least sensitive to parameter varialions among all realizations. The
sensilivity is defined as the shifting of the eigenvallles dlle to parameter varia
lions. In practice, in order to reduce sensHivity. a transfer function 9(S) is
ractored as
9(s) =g(S)92(.':)
or
and then verify this by

terms of the coefficien
we obtain
y=PY=X2+'
..
2
y=py=-a;
Substituting these into
obtain
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).
Realization of linear time-varying differehtial equations. Before
conc1uding this section, we shall briefly discuss the se"tup of dynam ical eqllations
for linear, time-varying dirrerential eq uations. Ir an ~th~ordcr. linear. time
varying dirrerential equatiori is or.the form
(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<
REALlZATIONS OF VECTOR PROPER RATIONAL TRANSFER FUNCTIONS
:, 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
Consider the following second-order time-varying differential equation

[p2
+ IX (t)p + IX2(t)]y(t) =
[,80(t)p2
+ ,8 (t)p + ,8 2(t)]U(t)
(6-47)
The procedure of formulating a dynamical equation for (6-47) is as follows:

We first assume that (6-47) can be set into the form
1
-IX(t)
Clearly, the con

rm realizations are the
t necessarily irreducible
cible [N(s) and D(s) are
ov parameters and the
bservable no malter the
:ment of computing the
ore difficult to compute.
;ed in References 23, 81
g all realizations. The
due to pararneter varia
transfer function g(s) is
. 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
Substituting these into (6-47) and equating the coefficients of u,

obtain
e(t) = f3 o(t)
b (t) = ,8 (t) - IY. (t),8 o(t) - 2~ o(t)
b 2(t) = ,8 it) - b (t) - IX (t )b (t) -
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
6-4 Realizations of Vector Proper Rational

Transfer Functions
In this section realizations of vector proper ralional transfer functions will be
studied. Bya vector rational function we mean ther a 1 x por a q x 1 rational
function matrix. Consider the q x 1 proper rational-function matrix
(6-50)
=========",-=',='-=,.".,
254
REALIZI
IRREDUCIBLE REALIZAT10NS, STRicr SYSTEM EQUIVALENCE, AND IDENTIFICAT10N
It is assumed that every g;(S) is irreducible. We first expand
(;(s)
Consider
l
e ] rgl(S)]
+ ~z:(S)
eq
Example 1
Ginto
(6-51 )
(8
G(s) =
gq(s)
where e =g;(oo), and gi(S)~g;(s) -e is a strictly proper rational function.

We compute the least common denominator of g, for i = 1,2, ... , q, say sn +
al s" - 1 + ... + a", and then express G(5) as
(6-52 )
+3
+s1)(s
+2)
5
+4
+3
Hence a minimal-dime
lt is c1aimed that the dynamical equation
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
We study now the

Since its development i
the result. Consider th
G(5) = [g'I(S) : g'z(5) : ..
(6-53b)
=[el : ez: : el
[el: ez : ... : el
is a realization of (6-52). This can be proved by using the controllable-form

realization of g(s) in (6-14). By comparing (6-53) with (6-26), we see that the
transfer function from u to y is equal to
Then the dynamical eql

which is the ith component of G(s). This ilroves the assertion. Since g;(s) for
i = 1, 2, ... , q are assumed to be irreducible, the degree of G(s) is equal to n.
The dynamical equation (6-53) has dimension n; hence it is a minimal-dimen
sional realization of (;(5) in (6-52). We note that if some or all of g;(s) are not
irreducible, then (6~53) is not observable, although it remains to be control
lable.
For single-input..single-output transfer functions, we have both controllable
form and the observable-fonn realizations. But for column ration'al functions
it is not possible to have the observable-form realization.
x3
O
1
O
O
O
1
Xl
Xz
= [O
REALlZATIONS Of VECTOR PROPER RATIONAL TRANSfER FUNCTIONS
255
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
per rational function.

= 1,2, ... , q, say Sil +
: : .... 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
Hence a minimal-dimensional realization of G(s) is given by
+ ... +Pqn
Xl]
[ ~2
[OO
-6
X3
o
,+ u
y= [
1
O
-11
6
3
(6-54 )
(6-53a)
I
We study now the realizations of 1 x p proper rational-function matrices.

Since its development is similar to the one for the q x 1 case, we present only
the result. Consider the 1 x p proper rational matrix
,,]
e
G(s) = [g'l (s) : g~(s) : ... : g'p(s)]

(6-53b)
eq
[el: e2 : ... : e p] +[gl(S) : g2(S) : ... : gp(s)]
'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)
Then the dynamical eguation

;ertion. Since gas) for
e of G(s) is egual to n.
it is a minimal-dimen
e or al1 of g;(s) are not
emains to be control-
XI
X2
X3
O
1.
O
O
O
xn
lave both control1able

lmn raional functions
= [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
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATION
is a realization of (6-55). The realization is always observable whether 9;(S),

i = 1, 2, ... , p, are irreducible or noto If they are all irreducible, then the realiza
tion is controllable as well.
It is also possible to find lordan-form realizations for vector proper rational
functions. The procedure is similar to the one for the scalar case. We use an
example to iIIustrate the procedure.
Realization from the
Consider a q x 1 prop<
A
A
G(s)
=
"
[
'Example 2
g,
The Markov paramete

For each g(s), we form
composite matrix
Find a lordan-form realization of the 2 x 1 rational function
"(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,)
k,~ [a" (1 )a2' (2)' .. a2,(IT, )

-2 f-------.;.-{+ } - - - - - Y2
U ,
Figure 6-6
Block diagram of Equation (6-57).
257
REALIZATIONS OF VECTOR PROPER RATIONAL TRANSFER fUNCTIONS
E, AND IDENTIFICATION
Realization from the Hankel matrix
Jservable whether 17;(S),
.ucible, then the realiza-
Consider a q x 1 proper rational matrix G(s) expanded as
91(Sj
r vector proper rational

scalar case. We use an
(;(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(~)
hq(O) +hq(l)s-l +h q(2)s-2
+...
(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
~1_(~~ ~1: p)l) u 1 (no. of Iinearly independent rows)

T~ ~2S~~-_1_,p)ll U 2
(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
he state variables chosen
eadily shown to be con

~nce it is an irreducible
-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]
O :.lIq,(I) .. (/q,(o-,) O '" O O:

'----y------""
a2
(6-60)
258
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTlFICATION
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"
Ir (Ji = O for some i, say

1
e=
-(131(1)'"
-(1,
with
hl(O)j
h 2 (0)
The dimension of this

the rank of H.
The assertion that (1
by using the procedure
forward, though tediow
e= .
hq(O)
where, for i = 1, 2, ... , q,
O
O
o
o
1
O
-a(2)
-a(3)
((Ji x (J) matrix
A=
for i > j,
O
O
O
O
((J x (JJ matrix
Let
Cj
be the ith row of (

ciAkb =h{.
and, for i= 1, 2, ... , q,

hIU)
hI(2)
b. =
I
This estabJishes, followiI

We show next that t!
see that lb Ab .".
in (6-59); hence its rank
verified that the matrix
hI(J)
The matrix
e is given by, if (Ji 1=0, for i = 1, 2, ... , q,
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)
is the identity matrix of (

the proof that (6-61) is an
--_.. _._._--------_._-----------
REALlZATIONS OF VECTOR PROPER RATlONAL TRANSFER FUNCTIONS
0, ANO IDENTIFICATION
except element 1, only

s a consequence of the
y dependent row of H
Y independent rows of
~educible realization:
(6-61 a)
(6-61b)
----
259
If (Ji =0 for sorne i, say i = 3 for convenience of il1ustration, then
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)
The dimension of this realization is n = (J 1 +(J 2 + .. , +(J q which is equa] to

the rank of H.
The assertion that (6-61) is a realization of G(s) in (6-58) can be established
by using the procedure used in the scalar case. Similar to (6-39), it is straight
forward, though tedious, to verify
hq(O)
h(k + 1)
hl(k +2)
J x (J) matrix
k =0,1,2, "..
(6-63)
hq(k +(Jq)
Let c be the ith row of C. Then we have

cAkb=h(k+\),
k=O, \,2, ... ;i=\,2, ... ,q
This establishes, fol1owing (6-34), the assertion.

We show next that the realization in (6-61) is irreducible. From (6-63), we
see that [h Ah ... A" - 1 bJ consists of ali iinear independent rows of H
in (6-59); hence its rank is n and lA, b} is control1able. Similarly it can be
verified that the matrix
"O
0J}"
o
O
..
'.
(Jq
(6-62a)
cq A
C1
q- \
is the identity matrix of order n; hence {A, C} is o~servable. This completes

the proof that (6-61) is an irreducible realization of G(s) in (6-58).
'-~
260
"
,,,,"U"" , RCWZA no""
'.,__
~ _
~ _ ~ _ ~ ~ _ ' _ ' ~ ~ _ ' ~ ' ~ ' _ O _ ' "
__
- ~ "
~ _ ,
__
" _ ~ ~ , _ "
'_'_~'
__' _ " '__
. ' _ ' - ' ~
"~_'
~"~"_
-~_.,,~._
".__,,'.'-0._ , .,
=[.~':,:-~)2] =[s'J +s'~-JS'5
The leftmost matrix is t

we have u I = 4 and u 2 '
using the formula in (F
6
+5s' -7s' +9s 8 -\1,,'9
'\'''_,,'')+,,'6_,\'7+ S ' 8 _ ..
~~~~-',
ReAWA
Consider
()
__.,__
""" ""CM ",U,"A"NCC ANO "''',UACA UON
Example 3
Gs
k = [O
k 2= [_1
+"'J
O
_.1
9
'---'--
s3(~+I)
uI
We form the Hankel matrix. with 2 =4.2 2 =4. and
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 (;
If we apply the row-searching algorithm to T, we will fina\ly obtain

1
-1
3
-S
7
1
-\
3
-5
1
-1
3
1
2
--6---0- ---6--- -3~ - -
(;(s) = H(O
where HU) ha ve the same o

ex i are defined in (6-59). W
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
0- ---O ---0- ---6 --o o o o

o o o o
o o o o
o o o o
m
o
o
o
o
U2 =
I
H(3) - - - I
:
,
H(ex+1)
H(ex +2)
Ii
H
There are ex + 1 block rows i

block row is associated with '
ordering of rows, the Hankei
.(6-59). Now We shall apply 1
mdependent roWS of t in ord
block row is linearIy depende:
subsequent block rows are Ji
4S~e
Ihe discussion in (he Subsecliol

Chaplcr 5. Sce also Schcmes I an(
lrollabilily malrix in the second half
REALIZATIONS OF VECTOR PROPER RATIONAL TRANSFER FUNCTIONS
261
E, AND lDENT1FICAT10N
The leftmost matrix is the matrix F defined in (F-9). From the rightmost matrix,
we have (J 1 = 4 and (J 2 = 1. The fifth and seventh rOws of K can be obtained by
using the formula in (F-l1) as
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
The elements :I: correspond to the primary linearly dependent rows.

an irreducible realization of (;(s) is given by
=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
(;(s)=H(O)+H(1)S-1 +H(2)s-2 +H(3)s-3 + ...

where HU) have the same order as (;(s). Let ct. = max {ct.h i = 1,2, ... , q }, \Where
ct. are defined in (6-59). We then form the Hankel matrix
=[n:~\ ---nm----~~l- --_:::- -~rrc ii --\ lon, block

H(ct.)
H(ct.+l)
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. + 1) H(o: +2)

H(ct.+2) H(0:+3)
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
IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, AND lDENTlFICATION

REALIZ.
the completion of this search, we then rearrange the rows of tinto the form of
(6-59) as
H(C( +1, [3)J]

H 2 (a + 1, [3)
__
H q( a
+ 1, [3)
\'
Example 4
Consider the transfer-f
(no. oflinearly independent rows)
G(S)=[~J)-3 +G}
(6-65)
J \' q
We form with Ct=I

independent rows as
Clearly the total numbers of independent rows of t in (6-64), t in (6-65), and

T in (6-59) are aH the same. Let \' be the number of linearly independent
rows of H(a + 1, [3) in (6-65). Then we have VI +V2 + ... +v q = (T +(T2 + ...
+(Tq, where (T are defined in (6-59). Note that generally we have (T 'v, for sorne
or all i. Clearly the primary linearly dependent row of H in T is its (v + l)th
row. The major difference between T in (6-59) and t in (6-65) is the orders in
which the linearly independent rows are searched. In 1', the linearly indepen
gerit rows are searched in the order of the first row of H, i = 1, 2, ... , q; the
secQnd row of H, i = 1, 2, ... , q, and so forth. For convenience of searching,
we search the rows of t in order from top to bottom and then rearrange it as
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
k.=[u.(l) (/.,(2) ... u.,I",)

'-v---\'1
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)
u.,II) ... u.,I",)
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
T~e rows of K correspond

tamed, by using the formul~
(6-67)
A2 ...
with
~l
A 2
A=
.
lA"
A 22
...
Aq
Aq2
... A qq
k2
A"l
A 2q
Note that the
(
(
k row corres
~nearly dependent 01', equivi

~2 row_corresponds to the p
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
REALIZATlON OF VECTOR PROPER RATIONAL TRANSFER FUNCTIONS
=E, AND IDENTIFICATl0N
Example 4
IWS of tinto the form of
Consider the transfer-function matrix (;(5) in Examp\e 3. We write

?endent rowS)
(6-65)
We form f with o: = max {4, 4}

independent rows as
in (6-64), t in (6-65), and
: of linearly independent
t- ... +Vq=O"l +O"z
+ ...
Iy we have"O" f.v, for sorne

of H- in T is its (v + l)th
in (6-65) is the orders in
ln t, the linearly indepen
. 1, 2, ... , q', the
ofH,I=
convenience of searching,
n and then rearrange it as
rearrangem~t of"the coef

ments from T to T, we can
1
O
-1
-1
3
1
4 and
f3 = 5, and then search its linearly
-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
: .. ' : "',(1) .. , ",,11',) I
(6-67)
:~:1
'qq
62) with O" replaced by vi a~d

ment is similar to the one n
the necessity of tracking the
not involve any new idea and
Hustrate the procedure.
(j)
-3
:I,-1
O
~2;
(ti -- -0- - - - 6--- -0- - - - 6
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)
;-66) and (6-60) is that aJk),

In otherwords, the (v 1 + l)lh
roWS ofH l but also on the
~ the (O" 1 + 1)th row of H 1 in
.imilar remarks apply to the
le dynamical equation
O
O
6-- -::(:---6--- -6 --- 4
.~
1',
O
O
The rows of K corresponding to the zero rows indicated by arrows can be 00

tained, by using the formula in (F-11), as
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
Hence an irreducible realization of G(s) is given by

O
O
1
O
O
O
O
O
1: O
O
O
-6 -4
1: O 18 x+ 1 u
---------------------O
O
O
O: O
1
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)
where H(k) are 1 x p con

denominator of aH elem,
of g(s). Define fi = max
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
1t is claimed that the dyn
X~,
L- a(1)
where Vk and U k are defined in (5-48) and (5-27). The postmultiplication of

U p - 1 on V~ operates only on the columns of V~, hence the linearly independent
rows of T in order from top to bottom are the same as the linearly independent
rows of V~ in order from top to bottom. Consequently, we conclude that the
Vil i = 1, 2, ... , q, are the observability indicesof any irreducible realization of
G(s). This can also be easily verified for the G(s) in Example 4.
Because Vi' i = 1,2, ... , q, are the observability indices, we have
v~max{v;, i = 1, 2, ... , q} su = max{uiJ= 1,2, ... , q}
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
*6-5 Irreducible Re,

Matrices: Hankel ME
There are many approachl
rational matrices. One app
apply the reduction proced
reducible one. We discuss
IRREDUCIBLE REALIZATIONS OF PROPER RATIONAL MATRICES: HANKEL METHODS
1, AND IDENTIFICATION
introduced in CAkB. Hence it is more difficult to determine the correct (Ji,

i = 1, 2, ... , q, on digital computer computation. We note that the order of t,
which is q(ex + 1) x {3, is larger than the order of T, which is (ex l +rx2 + ... +ex q )
x {3; however, in the search of linearly independent rows of t, once a dependent
row, say the ith row, in a block row is found, aH the ith rows in the subsequent
block rows can be skipped. Hence, although the order of t in (6-64) is larger
thaJ1 the T in (6-59), the highest power of CAkB used in the search of linearly
independent rows in t is smaller. As a result, it is easier to determine Vi than (Ji
on computer computation, and t should be used.
We discuss now an irreducible realization for a 1 x p proper ralional matrix.
Consider
~Primary
dependent
rOW ofH I
II
265
~ Primary dependen!
row ofH z
G(s)= [91(S) 92(S)
...
9 p (S)]=H(0)+H(1)s-I+H(2)s-z+ ...
where H(k) are 1 x p constant matrices. Let ex be the degree ofthe least common
denominator of aH elements of G(s) and let {3i be the degree of the denominator
of 9(S). Define {3 =max{{3i' i = 1, 2, ... ,p}. We form the Hankel matrix
T=
f~(ex+1)
the k1 and kz rows with

and V 2 Markov paraIds only on VI and V2'
[a(1)
H({3)
H({3 + 1)
H(ex +2)
~(ex + {3)
a(2)
a(v)
1 O ...
It is claimed that the dynamical equation
lankel matrices in (6-59)

le Vi in (6-65), which are
)f T in order from top to
;ible realization of G(s).
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
is an irreducible realization of the 1 x p proper rational transfer matrix. The

proof of this statement is similar to the column case and is left as an exercise.
*6-5 Irreducible Realizations of Proper Rational

Matrices: Hankel Methods
q}
an or equal to the one in

lre used in the realization
in the realization in (6-60).
~, the more errors may be
H(3)
It i5 a (ex + 1) x f3p matrix. ; TOW we search the linearly independent rows of T

in order from top to bottom. 1t is assumed that the first V rows are linearly .
independent and the (v + l)t& row is linearly dependent on its previous rows.
Then there exist a(i), i' 1,2,
, v, such that
VI
= 1, 2, ... ,
H(2)
H(l)
H(2)
[.
"
~t
There are many approaches to find irreducible realizations for q x p proper.

rational matrices. One approach is to first finda reducible realization and then
apply the reduction procedure discussed in Section 5-8 to reduce itto an ir
reducible one. We discuss this approach first.
266
IRREDUCIBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTlFICATlON
IRREDUCIBLE REALI2
Given a q x p proper rational matr~x G(s) = 19/s)), if we first find an irre

ducible realization for every element gj(s) of G(s), and then combine them
together as in (4-61) or Figure 4-7, then the resulting realization is generally not
controllable and not observable. To reduce this realization requires the
application of the reduction procedure twice.", If we Jind the controllable
canonical-form realization for the ith column, G{s), of G(s), say
X = Ax + bu
+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) =
These equations impl)
(sm
fIJ
A2
O
C2
and
Consider
C(sI-A)-B
is a realization of G(s). (Pro ve.) Because of the structure of A, b it can be

readily verified that the realization is always controllable. It is however gener
ally not observable. To reduce the realization to an irreducible one requires
A
the application of the reduction procedure only once.
lto. is possible to obtain difTerent controllable realizations of a proper G(s).
Let G(s) = G(s)+G(oo), where G(s) is strictly proper. Let tItes) be the monic
least common denominator of (;(s) and of the form
tItes) = sm +IX sm - +1X 2 S m- 2 + ... +lX m
(6-69)
A
which becomes, after tl

C(sI -A)
This shows that Equati
of A and B, it is easy to
generally not observabJ
We note that the re
form realization for a sc
to find a generalization
,.,
Then we can write (;(s) as

1
G(s) = t{t(s) [Rsm- +R 2 sm- 2 + ... +RmJ
A
(6-70)
G(,
where R are q x p constant matrices. Let I p be the p x p unit matrix and Op

be the p x p zero matrix. Then the dynamical equation
Op
Op
OP]
[OP]
x+:
u
Op
Op
-lXm-I p -lX m_ 2 I p
Ip
-1X1I p
R
where H(i) are q AX P co

denominator of C(s} an
we can show
H(m
Op
Ip
J x+
+i)=
-1XH(m
(6-71 a)
This is a key equation ir

. Let {A, R, C, E} be a
(6-33),
G(oo)u
(6-71 b)
is a realization of (;(s). To show this,itis sufficient to show C(sI -A)- B = (;(s).

Define V(s)"~ {sI '-'- A)-IB or (sI - A)V(s) = B. Ves) is a mp x p matrix. If
we partition it as V'(s) = [V'l (s) Vi(s)'"
V~(s)], where the prime denotes
the transpose and V(s) is a p x p matrix, then (sI - A)V(s) = B or sV(s) = AV(s)
G(s) = E +C(sI - A)- R

From (6-73) and (~-75),
is arealization of G(s) in
HI
IRREOUCIBLE REALIZATlONS OF PROPER RATlONAL MATRICES: HANKEL METHOOS
;, AND IDENTIFICATlON
if we first find an irre

Id then combine them
lization is generally not
:alization requires the
: find the controllable
G(s), say
267
+B implies
SV I(S)=V 2(S)
sV 2(S) = V 3(S) = S2V I(s)
and
Ui
is the ith component

U, then the composite
sVm(s)=-<xmV(S)-O:"'_IV2(S)- ... -<xIV",(s)+I p
These equations imply

(sm
:-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)
i=I,2, ... ,/1'l
Consider
C(sI - A( 1B = CV(s) = RmV I(s) + R m-1 Vis) + ... + R 1Vmis)
cture of A, b it can be
,le. lt is however gener
irreducible one requires
which becomes, after the substitution of (6-72),
:ations of a proper G(s).
Let lf(s) be the monic
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)
p x p unit matrix and Op

)ll
: }+[r}
ti I p
C(sI _ ArlB = Rm+sRm _ 1 + ... + sm-I R 1 = G(s)

!/J(s)
G(s) = H(O) + H(l)s- I + H(2)s-2 + ...
(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
This is a key equation in the following ~evelopment.

Let {A, B, e, E} be a realization of G(s) in (6-73). Then we have, similar to
(6-33);
.
lp
R ] x+ (;(oo)u
(6-71b)
.how C(sI -A( B =G(s).

is a mp x p matrix. Ir
where the prime denotes
)V(s) = B or sV(s) = AV(s)
(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
IRREDUCIBLE REAUZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATION
Hankel matrices
Now we claim that the dynamical equation

Oq
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
is a qm-dimensional realization of G(5).

verify that
H(2)]
AB
= ~(3)
[
AZB
] X+ H(O)u
(6-77b)
lndeed. using (6-74), we can readily
[H(3)]
= ~(4)
~(m + 1)
(6-77a)
H(m -1)
H(m)
.. "' A'B =
H(m +2)
(6-78)
HU +m)
and, in general,
Consequently we have CAiB = HU + 1). This establishes the assertion. The

observability matrix of (6-77) is the unit matrix of order qm; hence (6-77) is
always observable. lt is, however, generally not controllable.
lt is possible to introduce other either controllable or observable realiza
tions. Their introduction however would not introduce any new concept and
will be skipped. Instead, we shall discuss in the following two methods which
wi11 yield directly irreducible realizations. Both m'ethods are based on the
Hankel matrices.
Method 1. Sin~ular value decomposition. Consider the q xp proper
rational matrix (;(s) given in (6-73). Let Jt(s) =j" + IX I S",-1 + ... + IX", be the
least common denominator of all elements of (;(s). Define the qm x qm and
pm x pm matrices
M~
Kq
-IXm-Il q
-IX m -
zl q
Ip
O
O
-IXm_Il p
N~ O
Ip
-IXm-zl p
Ip
-IXll p
where 1" denotes the n x n unit matrix.
Note that the left-uppe

k x !(!>k) constant ma
of order k, and O is the t
can be removed from i\
"HU + 1) = 1q.
where the prime denot
conclude that lA = M,
realization of G(s) in (6
observable but not nece
c=
A=
18=
IE=
(6-79a)
-:m1q
lq
T~[::
Since T and t consist

are q x p matrices, T ar
verified that
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)
We also define the following two
where 1: = diag "Ir Al' A2,

roots of the eigenvalus
IRREDUCIBLE REALlZATIONS OF PROPER RATIONAL MATRICES: HANKEL METHODS
, 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
(6-74), we can readily
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)
les the assertion. The

ier qm; hence (6-77) is
,llable.
or observable realiza
e any new concept and
ing two methods which
lods are based on the
nsider the q x p proper

l Slll - 1 + ... + rt. 1II be the
)efine the qm x qm and
(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
i=O, 1,2, ...
(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)
efine the following two
is a pm-dimensional realization of (;(5). The controllability matrix of this

realization is a unit matrix; hence the realization is always controllable. The
realization however is generally not observable,
Now we shall use Theorem E-S, the singular value decomposition, to find
an irreducible realization directly fram T and t. Theorem E-S implies the
existence of qm x qm and pm x pm unitary matriCes K and L such that
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 REALlZATlONS, STRICT SYSTEM EQUlVALENCE, AND IDENTlFICATlON
IRREDUCIBLE REALl
Consider
A 2 =(Y+Tl
conjugate transpose. Clearly n is the rank of T. Let K 1 denote the first n

columns of K and L 1 denote the first n rows of L. Then we can write T as
T=
Kl~Ll
Kl~1/2~1/2Ll ~ YU
=(Y+Ml
=Y+MT
=Y+MT
=Y+MM
=Y+M 21
(6-87)
where ~1/2 =diag {JAl, JA2>"" JAn}, y ~Kl~1/2 is a qm x n matrix, and

U = ~1/2Ll is an n x pm matrix. Defines
and
(6-88)
Repeating the proces~
Because of KfK 1 = 1". LLf = 1" (see Problem E-7), we have

Y+Y =~-1/2K!Kl~1/2 = In
UU+
and
=~1/2LIL1*~-1/2
= 1"
(6-89)
Consider (6-84):
(6-90)
HU
With these preliminaries we can now establish the following theorem.
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
'---v--' '---v-' '
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)
is an irreducible realization of G(s).
The crux of this 1

Numerically stable co
Hence this realization (
The decomposition
choose Y = K 1 and U =
y
'f and ~J def]ned in (6-8
We give first a justification of (6-91) before giving a formal proof. We see

from (5-65) that Y and U are, respectively, the observability matrix and the
controllability matrix of {A, B, C}. Consequently, we have (6-91b) and (6-91c).
Again from (5-65), we can easily verify that t = y AV. The pre- and postmulti
plication of y+ and U+ to T = y AU and the use of (6-89) and (6~90) yield
(6-91a). Since the dimension of A is equa io the rank ofT, the irreducibiiiy of
(6-91a) follows from pY = pU =dim A.
Now we give a formal proof of the theorem.
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)
Then we have, by using (6-87), (6-89), and (6-90),

TT+T =VUU+Y+YU = YU =T
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)
5V+ is caBed the pseudoinverse ofV. See Reference 116.
IRREDUCIBLE REALIZATlONS OF PROPER RATIONAL MATRICES: HANKEL METHODS
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+
K1 denote the first n

Ne can write T as
(6-87)
1
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
Repeating the process we can show

,ave
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- .
. =YU, and pT= pY =

value decomposition.
~ '--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)
formal proof. We see

ability matrix and the
ave (6-91b) and (6-91c).
ihe pre- and postmulti
(6-89) and (6-90) yield
fT, the irreducibility o
(6-92)
(6-93)
This shows that (6-91) is a realization of (;(s).

The dimen~ion of A is equal to n, the rank of T. Let = [B AB ...
Am-1BJ and V'=[C' A'C' ... (A')"'-lC'J. Then from (5-65), we have
T = V and n =T :o;min(pV, p). 'Since is a n x mp matrix and V is a mq x n
matrix, we conc1ude p =n and pV =n. Hence the realization in (6-91) is
controllable and observable. This completes the proof of the theorem. Q.E.D.
The crux of this realization is the singularvalue decomposition of T.
Numerically stable computer programs are available for this computation.
Hence this realization can be readily carried out on a digital computer.
The decomposition ofT = YU in (6-87) is not unique. For example, we may
choose Y = K1 and U = :EL l , and the subsequent derivation still applies. The
V and U defined in (6-87), however, have the property
y*y = :E 1/2 Kt K l :E 1/2 =:E = :El/2LlLt:El/2 = UU*
A realization whose controllability and observability matrices having the
property y*y = UU* is called an internally balanced realization. Roughly
speaking, the signal transfer effect from the input to the state and that from the
state to the output are similar or balanced in an internally balanced realization
(see Problems 6-22 and 6-23).
In t)Je application, it is often desirable to flnd a simplifled or reduced-order
model fora given system. This realization procedurecan be directiy applied
to this problem. For example, consider
.
272
IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATlON
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 2(a2 +.1, {3)

T:
[
H q(a q +I,f3)
H. 22 .(a 2 + 1, {32)
:
H.2p(a 2 +.1, {3 p) }. a 2 linearly

:
independent
...
H qp(a q +l,{3p)J}a p rows)

(6-94)
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)
bottom. After the

into the form
T=
[
Cl
H(a+1',B)
~2(a+1,,B)
Hq(a +1, f3)
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
Consider the proper r
We compute the Mark
-2s 2 -3s-2
(s + 1)2
4s+5
---=4
s +1
and
-3s-5
s +2
It is clear that a = 4, a
IRREDUCIBLE REALIZAnONs OF PROPER RATIONAL MATRICES: HANKEL METHODS
ANO IDENTIFICATION
If the system is to be
en the reduced model
rowS ofL. See RefeIsion ofthe irreducible

rem 6-2, we have deg
I be computed from T.
an be chosen to be the
U, from the definitions
we may conc1ude that
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
the column and row

)n page 284.
ection we shall search
to find an irreducible
~ctor case discussed in
ion before proceeding.
the degree of the least
the degree of the least
heir computations are
e denominator degrees
;o:, i = 1, 2, ... , q} and
)rm the Hankel matrix
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:
---
ws in order from top to
(6-97)
+ ...
- - - = -3 +S-l -2s- 2 +4s- 3 -8s- 4 +16s- 5
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
We compute the Markov parameters:
p~l}
, {3
O" 1 (no. of
, {3 p) } 0"2 linearly
independent
Consider the proper rational matrix

-2S 2 -3s-2
HdO:l +1, {32)J

Hd0:2 +1, {32)
,
:0
I
I I
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 REALIZATIONS, STRICf SYSTEM EQUIVALENCE, AND IDENTIFICATION
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
2-- f -- ----=-"4- - - --- - -1- -
~~
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
: --:3 --- 0- -;--=4 ---()
---
O: -2
1: -1
-~ 2--- -f -;-- 3--- -e
Hence an irreducible realization of (;(s) is given by
1
O
O
O
O
O
O
O
2
2
The elements ( corres

rearrange tinto t anc
O1((
[~IJt~[O,O
2
O O O
'-----v----'
VI
=3
.....= =..
=
..
_~~
_= _=.=_.. =--=
~
~=.=--~.
=.=
.. =--=
..=..=.= =._=._=
~
=.
275
IRREDUCIBLE REALlZATIONS OF PROPER RATIONAL MATRICES: HANKEL METHOOS
ANO IDENTlFICATION
Next we use
tion ofthe row-search
t and t to find a different irreducible realization. We write
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)
ion 'of any irreducible

:venth row ofthe right
1-11), from the leftmost
--------------------------
-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-'
Corresponding to the primary Iinearly dependent rows indicated by the arrows

as shown, we can compute, by using (F-1l), the coefficients of combinations as
1
O u
O
O
O
1
O
'I:
(n
O]
O]
(6-99)
The eleme~ts :X: c,?rrespond to theyrimary linearly dependent rows. Nowwe

rearrange Tinto T and rearrange k accordingly to yield
[kz~lJt~[O
- ,O
and kz with the signs

v parameters of 911 (s)
,nd gds). The form of
O 1 0.) 0;2 1
O 0J
O O O O: 2 3: f: O O
'-v------"
V1
1.
=3
vz=2
~~,.~==.
.~~,
.. u,,_,.
_.. ,_~.,.,'~""" . ~'.~ __ "_n-"'.~ __ ~~
276
... ~:::":,:::,=----::..:'..:.,:",:,,
:.:.:::::.':.:.._._J'~:"_":::':'~:"':'::=--'--....:'.:::._~:'::"":'::":'':':::::::-'':':':'''''''':~::'::::'':';;~:,:":,:_::::_:::
..:::.~:.::::::::~~..::-::::~~;-~~=----=---:::=:.::::::.::.:::.:..:::....::.o=~~:::.....:..::.::;::_:
_::...:...,. ::;::' ._--",_..... -::::...7"....:.:....:::.:=;:.
.:.::~_
:-2:
:__ ~_: -4
-4
5
: O:
-4
: 1:
5, :0:
-6
O
O
O
O
O
That is, n is the sum of al
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
Hence a different irreducible realization of G(s) is
x=
_~~.-
IRREDUCIBJ
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUlVALENCE, AND IDENTIFICATION
:" -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.)
*6-6 Irreducible Realizations of G(s): Coprime

Fraction Method
Controllable-form realization. Consider a q x p proper &rational matrix
G(s). By long division of each element al :8(5)., 'He can "llfite rG(s; ~1.S
~
G(s) =G( (0)
+ 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
The matrix H(s) is a p x p

on the diagonal. The m
responding column is a Z'
p x p and p x n constan
expressed as
where N Ic is a q x n coos
The procedure to be
lable canonical-form re;
difference is that the D(s
here. Let v be a p x 1 v,
and
The substitution of (6- H
and
Since D(s) is column re(
(6-105) can be written a~
H(s
Let us define
6This step does not appear il

rcalization.
CE, AND IDENTIFICATION
IRREDUCIBLE REALIZATIONS OFG(S)-COPRIME FRACTION METROD
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
and k2 with the signs re

IV parameters (excluding
larameters of 921(S) and
~pend on VI and V2' The
Verify them.)
(6-102)
where N lc is a q x n constant matrix.

The procedure to be discussed follows the one for developing the control
lable canonical-form realization discussed from (6-21) to (6-26). The only
difference is that the D(s) in the scalar case is normalized to be monic, but not
here. Let v be a p x 1 variable defined by ves) = D- 1 (s)u(s). Then we have
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
(DhcH(s) + DlcL(s))v(s) = u(s)

y(s) = N1cL(s)v(s)
(6-105)
(6-106)
Since D(s) is column reduced, the constant matrix D hc is nonsingular. Hence

(6-105) can be written as 6
~ct
transmission part of a
m. Using the procedure
H(s)v(s) = - Dh;, 1 DfcL(s)v(s)
+ Dh;, 10(S)
(6-107)
Let us define
x(s) = L(s)v(s)
matrices and are right

nd ciN(s) <ciD(S) = ..ti,
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
We express this explicitly in the time domain:

X 11
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
is the ith component of v, and
Vi
From this definition we
V\k) = dkVi/dt.
i = 1, 2, ... , p; k = 1,2, ... , Jli-1
(6-110)
From the definition of H(s) and (6-107), we have
a1
X:"
Xl!',]
- ni;,n ,x(t) + ni; 'u(t) ~
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!",
Consider the strictly pro
O
X
-
X
-
-: -
X
-
_~1"_'
:0- -1- - 0- - - - -O
---
,'o
,:
:0
,
where "x" denotes poss

the (Jl1 + Jlz)th
(6-112b) follows directl
We note that if P'i = 0, t;
6-7 and the th row of ]J
Example 1
-0 -
I
I
10
Figure 6-7 Block diagn
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
IRREDUCIBLE REALIZATIONS OF G(S)-COPRIME FRACTlON METHOD
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
..
Figure 6-7 Block diagramofG(s)=N(s)D-1(s).
-1
)tes the ith row of Ohc .

ure 6-7. It consists of p
hosen as a state variable
le one in Figure 6-3.
110) and (6-111), we can
o
x
.. ,
y(t)= Cx(t) = N1cx(t)

(6-112b)
where "x" denotes possible nonzero elements. The Jllth row of A is equal to
- 3 11',' the (JlI + Jl2)th row of A is equal to - 3 21'2 and so forth.
Equation
(6-112b) follows directly from (6-106) by the substitution of x(s) =L(s)v(s).
We note that if Jli =0, then the ith chaLa of integrators wll not appear in Fig.
6-7 and the ith row of O;;,IOIe and O;;,I wil1 not appear in Equation (6-112).
Example 1
----------
Consider the stricdy proper rational matrix
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)
+1)' .: (s -2);.1 +1)2
O ..
(s
+ 1)'(.1 -
2)
J-'[(
.
~3S2 -6.1 :-2)(.1 -2)

.
.1(.1 - 2)
s3.-:3s-1
s +1 ]
s(s + 1)
(6-113)
"' .. "
-------'....:.:c."
' . . ~ . .""c"'
. ~ - - -
280
"" _
--.:=
"=:.""=.,,=::. . .
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTlFICATlON
IRRED
This fraction is obtained by using the least common denominator of each

row and is generally not irreducible. From the coefficients of this fraction,
we can obtain, by using the procedure discussed in Appendix G, the following
irreducible fraction:
~
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
Bence we have /11 = 3, /12 = 1, and /13 = O and

n=/11 +/12 +/13=4
We write O(s) and N(s) in the form of(6-100) and (6-102) as
is eq ual to the ith row

are all zero rows. To
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
Hence a realization of (;(s) is given by
(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
IRREDUCIBLE REALIZATIONS OFG(S)-COPRIME FRACTION METHOD
, AND IDENTIFICATION
denominator of each
cients af this fraction,
endix G, the fallowing
-1
s-2
O
281
is a realization of G(s) = N(s)D -I(S). We shall now formally establish this by

showing that the A, B, and e in (6-112) satisfies
C(sI - A)-I B
= N1c(sI - A)-IB = N(s)D-I(s) = NJ,(s)D- I(S)
or eguivalently,
-IJ-l
(sI - A) - lB
-3
1
L(s)D - I(S) = L(s)(DhcH(s) + DcL(s))- 1
(6-115)
This eguation implies
BD(s) = BDh/H(s) + D;/ DcL(s~J =(sl - A)L{s)
(6-116)
From (6-112), we see that all rows om, except the {L~ = 1 J.!m)th rows, i = 1, 2, ... , p,
are zero rows. Furthermore, these nonzero rows are egua! to D~ 1. Hence the
(L~ = 1 t'm)th row of
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
- -
--:.! - -.-.-. - - --o
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 '
)m the block diagram in

~(S)D-l(S); hence (6-112)
I
I
(6-117)
I
I
282
IRREDUCIBLE REALlZATlONS, STRICf SYSTEM EQUIVALENCE, AND IDENTlFlCATlON
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
(s~ _ A) (Adj (sI -
A))B = det
~(s) L(s) Adj D(s)
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
which implies the exisl
det(sI - A) = k det D(s)

for some constant k. Det (sI -A) has a leading coefficient of 1, and det D(s)
has a leading coefficient of det Dhc> hence we have k = (det D hJ- 1 and
det(sI - A) = (det D hJ- 1 det D(s)
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
Now we show that t.hen
(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.
To show this, consider ((
The composite matrix

Theorem 2-5 implies
rank [AI-A
which implies, because {,
IRREDUCIBLE REALlZATIONS OF G(S}---COPRIME FRACTlON METHOD
, AND IDENTlFICATlON
(cept the (I. (11= 1 JLm)th

:cause the (I~= 1 JLm)th
seen from (61-11), the
lal to the ith row of
-116). Hence the dy
5)0-1(S).
~) and det D(s). From
i
s) L(s) Adj O(s)

11.
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
51- A) and det O(s) are
which implies the existence of a p x 1 nonzero constant vector lX such that

O(A)lX =0
:ient of 1, and det O(s)
let Dhcr 1 and
Because of (6-116), (6-102) and (6-112b), we have
(6-118)
lot right coprime
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 )
Since L(A) consists of a unit matrix of order p as a submatrix, if lX is nonzero, so

is the n x 1 vector L(A)lX. Hence we conclude from (6-121) that [(sI-A)' CJ'
does not have a full rank at s =A, and {A, C} is not observable.
Now we show that if {A, C} is not observable, then D(s) and N(s) are oot
right coprime. Under the premise, there exists, from the dual of Problem 5-37
on page 231, an eigenvalue Aof A and its n x 1 eigenvector e such that
[AI~AJe=o
6-119)
(6-122)
.t matrix of order p as a
Now we show that there exists a

e = L(A)lX
(6-120)
x 1 vector
el!
and
such that
O(A)lX=O
(6-123)
To show this, consider (6-116) or

. From the realization
O(s). Hence Corollary
nk[sl-A BJ=n,for
leorem 5-7). Up to this
sed.
[AI-A
The composite matrix [H-A
Theorem 2-5 implies
rank [AI-A
L(A)J
-BJ [ D(A) =0
(6-124)
- BJ is of dimension n x (n + p); hence
-BJ +nullitY[Ai-A
which implies, because {A, B} is cootrollable,

112) is obserVable if and
nullity[.. - A
-BJ=p
-BJ=n +p
_.~_.
284
.-
'.~-
...
~-
'.
"
....
. " .... "".'.'
._a'.
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTIFlCATION
Consequently, the p linear independent columns of [V(A)

basis of the null space of [Al - A
- BJ. Because of
IRREDU(
D' (},)]' form a
[U-A
the vector [e'

such that
O']' is in the null space. Hence there exists a p x 1 vector
[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'
where the prime denote~

readily show that the (
column degrees of D(s).
Observable-Form RI
realization discussed ir
observable form realiza1
extensively in the next s
Consider a strictly p
where the polynomial Ir

reduced. We take the tl
Furthermore, N'(s)and [
of the form in (6-112) ca

equation
is an irreducible realizati
multivariable obser~able
Of course, an observa
(6-125). Let the row d~gI
let
=.. ==e-=..
=====o===.=,
....=.-c=
....=---=
.._-=_
...=....._=
.... =_=_
....=.. =.-~. .=....
_ _.=-_
..
IRREDUCIBLE REALlZATlONS OF G(S)-COPRIME FRAcnoN METHOD
, 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)
where Ohc is nonsingular by assumption, then the new B matrix in (6-112)

becomes
to (6-122) yields (6-121)
'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-----'
e proof of this theorem.

Q.E.D.
; irreducible if and only
== N(S)0-1(S) is a right
6-] ,
and only if deg C(s) ==

ivalent (Theorem 5-20),
i only ifits dimension is
Theorem 6-2.
controllable canonical
1 (6-112) a multivariable
zation depends on the
LJ(s))(D(s)U(S))-I, where
e column reduced. then
rm realization. Hence,
)O in the multivariable
lical" is not used.
mn degrees and control
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
of the form in (6-112) can be developed for N'(s)(D'(s))-1. Consequently, the

equation
x=A'x +C'u
y=B'x
is an irreducible Tealization of D -1(s)N(s), and the equation is said to be in a

multivariable observable form_
af course, an observable form realization can also be obtained directly from
(6-125). Let the row degrees of D(s) be Vi, that iS,b,.O(s) = Vi, i = 1, 2, ... , q, and
let
(6-127)
===-
286
c __
__
....
__ ..
,.
'
"
_ .... _.
__._
.... __
"'_'
------------
We write
C(sI
D(s) = H(s)D h,.
+ L(s)D,. = (H(s) + L(s)D,.D,;;.1 )D h ,.
(6-128)
which implies
where
SVI
_
H(s) =
1 s ...
SVI - 1 :
------O-------:- i --S- - -.
_
L(s) =
or
O
O
O
~ ~
: :
;;2:'--1 - -: :- - - - - - - - - - - - - -
- - - - - -:- - - - - - - -' - - - - - - -. ~ - - - - - - - -~ '- - - - - - - - - - - - - - -

.
.
'.
.
.
------O-------;- ----- -'0 --------: :--1- -5- - .- .-. - -
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
This is a key equation

~quation (6-131) will t
G(s).
Dual to a multiv:
observable-form realiza
whether D(s) and N(s) ~
then the realization is e
Consider the network si

and inductor currents
Section 3-4, then a d}
network. There are, 1
matical equations to de~
currents ~1(t) and ~2(t) a
. ,
,
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
to describe the network.
~x
O x
O x
O O
1 x
+ Nteu(t)
: ... : O O ...
I
x(t)
,
,
x ,,
O O ...
3
2
This set of equations can
(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
Figure 6-8 A network.
POLYNOMIAL MATRIX DESCRIPTlON
;, AND IDENTIFICATlON
(6-128)
287
C(sI - A)-lB = C(sI -A)-lN 1c

= D-1(s)N(s) = D -l(s)L(s)N lc
(6-132)
D-1(s)L(s) = C(sI - A)-l

L(s)(sI - A) = D(s)C
(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
Polynomial Matrix Description
(6-130)
I(s) is row reduced. The

[ x q matrices. Thus a
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)
to describe the network. The output y is equal to y(s) = 2S~2(S) or

r-~ (S'1
)I(s)=[O
-----------
O x
O
1
O x
O x
1 x
+ Nrcu(t)
O .. , O
This set of equations can be written in a general form as

P(s)~(s) = Q(s)u(s)
y(s) = R(s)~(s) + W(s)u(s)
6-131a)
Cqvq]x
--r
~2h
II
2h
(6-131b)
31
In
column of the matrix A

~ation of(6-125), we have
2s]l~:(s;j +O'u(s)
Figure 6-8 A network.
~-
(6-135a)
(6-135b)
288
IRREDUCIBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATlON
where P{S), Q(S). R(s). and W{s) are, respectively, m x m, m x p, q x m, and q x p

polynomial matrices; u{s) is the p x 1 input vector, and y(s) is the q x 1 output
vector. The variable ~(s) is a m x 1 vector and will be called the pseudostate.
In arder to insure a unique solution in (6-135), the square matrix Pis) is assumed
to be nonsingular. The set of equations in (6-135) is called the polynomial
matrix description of a system. In the time domain. this set of equations
becomes
P(p )~(t) = Q(p )u(t)
(6-136a)
y{t)=R{p)~{t) +W(p)u{t)
(6-136b)
where p stands for d/dt and pi for (idt i . From (6-135a) we have ~(s) =
P-1(s)Q(s)u(s). The substitution of ~(s) into (6-135b) yields
y(s) = [R(s)P- 1 (s)Q(s) + W(s)]u(s)
Hence the transfer matrix of (6-135) from u to y is equal to
G(s)=R(s)P-1(s)Q(s) +W(s)
(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)
~(s) = (M(s)P(S))-1 M(s)Q(s)u(s)~ P,:-I(s)Q,(s)fi(s)
(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)
so that P;I(S)Q..(s) is strictly proper (Theorem G-12'). The substitution of

(6-140) into (6-139) yields
~(s) = P,:- 1(s:Q,(s)u(s) + Y(s)fi(s)
(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
(6-131) such that
lf deg det Pr(s) = n, th

matrices. Similar to (1
and
or
where L(s) is defined m

unit matrix of order m,
p .. (s) and L(s) are lefl
Pr(s), CoroIlary G-S in:
sequently, {A, Ca} is ot
of {Pis), Q(s)} or {P(s), 1
into (6-135b) yields
y(s)
= R(s
This is not exactly in the ~

may not be strictly pro
obtain
where X(s) is a poIyriom

defined as in Corollary (
y(s) = C(sI
in which qsI - A)-l Bis
is a q x p poIynomiaI m.
dynamical equation
where p =djdt, is a realiz

~r the transfer matrix G(~
Gis) in (6-137) is strictIy pI

E(s) defined in (6-149). I
written explicitly as
y(t)
=O
From (6-144) and Pr(s)

have
POLYNOMIAL MATRIX DESCRIPTION
, 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)
::m is a special case of

1W(s) = E. The fraction
1tify R(s) = Nr(s), P(s) =
(s)N(s) is also a special
) = N(.s), and W(s) = O.
:ransfer function matrix
37). Note that the com
we discuss a method of
~r first (6-135a). If P(s)
r matrix M(s) such that
ultiplication of M(s) on
(6-138)
)Q,.(s)u(s)
(6-140)
2').
(6-131) such that

(6-142)
If deg det Pr(s) = n, then A, B, and C o are, respectively, n x n, n x p, and m x n

matrices. Similar to (6-130), (6-133), and (6-134), we have
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
y(s) = R(s)Co(sI - A) -1 B(s) +(R(s)Y(s) + W(s))u(s)
R(s)C o = X(s)(sI - A) +C
(6-147)
where X(s) is a polynomial matrix and C = Rr(A)C o is a q x n constant matrix

defined as in Corollary G-12. Using (6-147), Equation (6-146) becomes
y(s) = C(sI - A)-IBu(s) +(X(s)B + R(s)Y(s) + W(s))u(s)
(6-14S)
in which C(sI - A)- 1 Bis a q x p strictly proper rational matrix and

E{s)~ X(s)B
+ R{s)Y{s) + W{s)
(6-149)
is a q x p polynomial matrix. From this development, we conc1ude that the

dynamical equation
x{t) = Ax(t)
y{t) = Cx{t)
+ BiJI(t)
+ E(p)u(t)
(6-150a)
(6-15Ob)
where p =d/dt, is a realization of the polynomial matrix description in (6-135)

or the transfer matrix G{s) in (6-137). Note that W(s) =0 does not imply that
G(s) in (6-137) is strictly proper. Whether or not G(s) is proper is determined by
E(s) defined in (6-149). If E(p) = E +Ep +E 2p 2 + .. " then (6-150b) can be
written explicitly as
y(t) = Cx(t) + Eou{t) +E(t) +E 2(t) + ...
From (6-144) and Pr(s) = M(s)P(s), where M(s) is a unimodular matrix, we

~~
:ion procedure discussed

:::o} of the form shown in
(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)
where L(s) is defined as in (6-129) and is a m x n matrix. Since L(s) contains a

unit matrix of order m, the rank of [L(s) : Pr(s)] is m for every s in iC. Hence
Pr(s) and L(s) are left coprime (Theorem G-8'). Because dim A = deg det
P,.(s), Corollary G-8 implies that C o and (sI - A) are right coprime, and con
sequently, {A, C o} is observable. Note that no assumption of the coprimeness
of {P(s), Q(s)} or {P(s), R(s)} is used here. The substitution of (6-141) and (6-142)
into (6-135b) yields
(6-139)
. deg det P,.(s) = deg det

;0, we carry out the divi
289
det (sI - A) = k det P(s)
290
for sorne nonzero constant k. This relation is a consequence of the realization

procedure and is independent of the coprimeness arrrong P(s), R(s), and Q(s).
The {A, Ca} in (6-144) is of the multivariable observable form and is always
observable. The observability of {A, Ca}, however, does not imply the obser
vabili ty of {A, C} in (6-150). In the following, we shall establish
{P(s), Q(s) left coprime} if and only if {A, B controllable}
(P(s), R(s) right coprime} ifand only if {A, C observable}
Before doing so, we need sorne preliminary development.
(6-145), (6-140), (6-143), (6-147), and (6-149) to form
L(s) 0J[SI-A
[ -X(s) I q -C
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 )
where the leftmost and

equation, we are ready
From the first two
-U 11 (S)
P(s)
Uds)
M-l(S)L(s)
where 8(s) denotes the r

most matrices in (6-158
~--~-~'
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[
-Ul 1(S) UI2(S)

P(s)
M-l(S)L(.
We combine
ence of U 11 (s) and Uds) such that

-U l l (S)
P,.(s)
(6-151)
(6-152)
They are, respectively, (m + q) x (n + q), (n + q) x (n + p), (m + q) x (m + p), and

(m +p) x (n +p) matrices. Because of the right coprimeness of {P,.(s), L(s)} and
the left coprimeness of {si -A, Ca}, (6-145) and Problem G-ll imply the exist
Recall that P,.(s) = M

The premultiplicatior
yields
(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
orequivalently, [A, lB} is

Simiiariy, we can establi
are right coprime.
Definition 6-2
The polynomial matrix

only if P(s) and Q(s) are I
- 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
Consider the polynomia

Then an n-dimensional d
or, equivalently, controll~
POLYNOMIAL MATRIX OESCRIPTlON
, ANO IDENTlFICATlON
uence of the realization

ng P(s), R(s), and Q(s).
ble form and is always
es not imply the obser
establish
; control1able}
C observable}
>pment.
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)
ll'-:- __ ~-----~-] [-UI1(S) Uds)(sl-A): U 12

Q(s)
1m
Co
: - Y(s)
O : - R(s) W(s)
- - -O ---------- - - - - : - - -I~ - -
(S)BJ
We combine
= O : p(s)
\[~o -~p(S)J
(6-153)
,(m+q)x (m-tp), and

~ness ol' {p,.(s), L(s)} and
m G-ll imply the exist
:J
-U l1 (S)
P(s)
U12(s) J[l m
M- 1(s)L(st O
(6-158)
~ A) ~]
has rank n +m for every s in C
il' and on!y if
Q~S)J has rank n +m for every s in C
which implies that

[(sI - A) : B] has rank n for every s in C
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
[;
Ig this matrix, we expand
(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
to complete the equality.

,in (6-155) is unimodular,
. we have U 12(S)(sl - A) =
if and only if [P(s) : Q(s)] has rank
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)
1imodular, so is the first

rix in (6-155) is also uni
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
(6-155) by P(s) and Q(s).
292
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTIFICATION
This theorem follows directly from Definition 6-2 and the statements in
(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)
Consequently. at the roots, l., of det H(s) = 0, we have
pepe),) Q(A)]
= pH(A)[P(),)
Q(),)] < m
where p stands for the rank. lf an n-dimensional dynamical equation is de

veloped for {P(s), Q(s), R(s), W(s)} with n = deg det PIs), then this n-dimensional
state equation will not be controllable. lfthis equation is decomposed into the
form in (5-54), then the eigenvalues associated with A. c, the uncontrollable part,
wil! be equal to the roots of det H(s). Hence the roots of det H(s) or, equivalently,
those A in e with p[P(A) Q(A)] < m wil! be called the uncontrollable mode of
{Pis), Q(s)}. These uncontrollable modes are called the input-decoupling
zeros in Reference S185.
Similarly, if Pis) and R(s) are not right coprime, then the roots of the deter
minant of their greatest common right divisor wil! be called the unobservable
modes of {P(s), R(s)}. These roots are called the output-decoupling zeros in
Reference S 185. Hence an irreducible {P(s), Q(s), R(s), W(s)} does not have any
uncontrollable or unobservable modes.
The discussion of the polynomial matrix description in this section is not
complete. We discuss only its realization problem. For a more detailed
discussion, see References S34 and S218. In S218, the description is called the
differential operator description.
*6-8
given a transfer mat

procedures in Section
relationships among
questions can stiJl be
~ider a transfer matrix
G(s) = N .(05)0- 1(05)
or
<.: 1 is strictly proper (
G(s) = NI (05)0 1 1(05) + E
betw~en {N,.(s), D,.(s)} (
q uestIOns. we rewri te ((
' "
[
where ~ is called the
PSI
is called the system l11{{tr
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
Strict System Equivalence
In this text, we have introduced three types of mathematical descriptions for

linear time-invariant multivariable system~. They are state-variable equation
where 1 is a unit matrix 0[,

than deg det P(s). It is ele
transfer matrix in fractional forms
and
(6-160)
Hence the input-output b

their en tire dynamical beh~
Consider two system m
(6-161 )
S(s)
and polynomial matrix description
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.
where p(s), Q(s), R(s), an
q x p polynomial matrices.
ever, because of the discussi
to make them egual and re,
this reg uirement, the subseg
S186, and S187.
STRlCT SYSTEM EQUIV ALENCE
, AND lDENTlF\CATlN
and the statements in

lSS the situation where
prime, then there exists
:;uch that
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
e description is called the
1.
given a transfer matrix or polynomial matrix description, we may use the

procedures in Section 6-7 to deve\op a state variable description. Hence the
relationships among them have been essentia\\y established. Even so, sorne
questions can still be Ilosed regarding these descriptions. For example. con
sider a transfer matrix G(s) which is not necessarily proper. It can be factored as
9(s)=N,.(s)D,:-I(S) or it can be decomposed as G(s)=GI(s)+E(s), where
<}:1 is strictly proper and E(s) is a polynominal matrix and then factored as
G(s) = N (S)DI (s) + E(s). The question is then: what is the re\ationship
between {N,.(s), D,.(s)} and {N 1 (s), D 1(s), E(s)} ?To answer this and other related
questions. we re"'Tite (6-161) as
P(s)
[ -R(s)
where
S(S)=[
(s)N,(s)(s)
(6-160)
(6-161 )
ldied in earlier sections by

sfer matrix can be extended
uation,the transfer function
,)-IB--i-(E+E 1S+}.. lf
{nomia\ matrix descnptlOn
,and W(s) =0. Converse\y,
O ]
(6-162 )
-y(s)
(6-163 )
is ca\\ed the system matrix. 1ts transfer function from u to y is

G(s)=R(s)P-!(s)Q(s) +W(s)
(6-164)
Ir we identify R(s)=C,P(s)=(sl-A),Q(s)=B, and W(s)=E+E 1 s+ "', or

identify R(s) = N,.(s), P(s) = DAs), Q(s) = 1, and W(s) =0, then the system matrix
ineludes (6-159) and (6-160) as special cases. Hence the system matrix S(s) can
be used to describe any of the three descriptions n (6-159)-(6-161). .
Consider the system matrix S(s) in (6-163). We extend it to
:O
P(s) : Q(s)
J~
[ ---------,----O - R(s) , W(s)
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) =
; is called the pseudostate, and the matrix
Se(s) =
1ematica\ descriptions for

re state-variable equation
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
IRREDUCIBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTIFICATION
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
Since Ves) and Ves) are unimodulars, so are

Ves)
[ X(s)
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)
Qz(S)][V(S) .Y(S)J (6-168)

Wz(s) O
Ip
where V(s)=V-(s) and Y(s) = -V-(s)Y(s). Clearly Ves) is unimodular and

Y(s) is a polynomial matrix. A comparison of (6-168) and (6-155) reveals
immediately that {pes), Q(s), R(s), W(s)} and its realization {A, B, C, D} de
veloped in Section 6-7 are strictly system equivalent.
Mathematical descriptions which are strictly system equivalent are equivalent
in every sense. They have the same transfer matrix; their pseudostates are
related by an invertible transformation. Their dynamical-equation realiza
tions developed by using the procedure discussed in Section 6-7 are zero-inpul
and zero-state equivalent (see Section 4-3). If one realization is controllable
(observable), so is the other, and conversely. These properties will be established
in the following.
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
Note that 8 1 (s) and Sz(s]

hence u and y in (6-170) al
~ and ~z are different in '
0J[ P ~
[U
X 1 -R: "
which can be written as , b)
r1lT
L~
~"""1 r
"
;Jl-;_
A comparison of (6-171) an,
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
which, together with
VPY +VQ,
] [ P z
(XP -R)Y +(XQ +W) - -R z
where Ves) is unimodular, a,

t(s)=
We see that ~ and ~z areTe":

~nd (6-174). Hence, ifS 1 (s)a
~n~~t u(t), and any set of i~i
lO'tIal
conditions in S
d'
'd .
z, an
, entIeal and their pseudosta
295
STRICT SYSTEM EQUlVALENCE
:E, AND IDENTIFlCATION
Hence we have
'ctly system equivalent if
atrices u\s) and V(s) and
1 that
P2(S) Q2(S) (6-167)

- R 2(s) W 2(S)j
U(S)P I(s)V(s) = P 2(S)

which, together with the unimodularity ofU(s) and Ves), implies
k det PI(S) =det P 2(s)
(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
Iy verified that this equiv
oSl(S), then S2(S)~SI(S)],
tivity property [if S I(S).~
notes strict system eqUlv
GAs) = R1 Pi 1Q2 + W 2 =
R I PIQl + W 1 =
G1(s)
Q.E.D.
In order to establish the relationship between the pseudostates ~l and ~2'

we use (6-162) to writeS as
6-3 can be seen from the

ttions (6-153) and (6-155).
(6-167)may not besquare
imodularity property, we
(6-170)
(6-171 )
and
(s) 11[Y\s)
2(S)J O
Y(sq (6-168)
lp
y V\s) is unimodular and

-168) and (6-155) reveals
lization {A, B, e, D} de
1 equivalent
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
J[VO 1YJ[ -o; J
which can be written as, by the substitution of (6-170),
ol [
[U0J[ -y0J =[-yJ=
X
]32
-R 2
Q2JrlV~I--:,'YM---'J'
W2
-u
(6-172)
A comparison of (6-171) and (6-172) yields

;2(S) = V(S);I(S) - Y(s)u(s)
where
equivalent have the same

.a nonzero constant.
(6-173)
Ves) is unimodular, and

;1(S) = V- 1 (S);2(S) +Y- 1 (s)Y(s)u(s)
= V(S);2(S) - Y(s)o(s)
(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
RREDUCIBLE REALIZATIONS, STRCT SYSTEM EQUVALENCE, AND IDENTIFCAnON
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
Coprimeness, controllability, and observability are invariant under the trans

formation of strict system equivalence.
Proof
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
=
If {A, B, C, E} and {A, i

constant matrix P such
these, it is straightforwa:
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 )
Hence their system matr

Now we assume that
valent. Then from (6-16
U(S) O][SI
[ X(s) I
q
for sorne unimodular U(s
the same transfer matrb
implies
Using CorolIary G-12,
Wf
and the first block column of (6-167) can be written as

U(S)' O][p(S)]V(S)_[ P 2(S)]
[ X(s) I
R(s)
- -R 2(s)
L
where P is a constant mal

(6-176)
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.
v!hich lmplies if' =
jp o and .
P(sI -A)
Now we shalI show tl
unimodular, U- (s) is also
l
for some polynomial matri

plication of(6~179) and (6-1
U(s)U- 1 (s) = (sI'..:.. A)U(s)(sJ
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.
which becomes, because of I .

I-Pp=(sI-;
,._..-._-""-- ...--._,' , .. .,.--" .. -_ ... ,
STRIGr SYSTEM EQUIV ALENCE
... ~
.. _'-"
", ......
297
:ence in their dynamical
Proof
ey are realized by using

izations {A, B, C i , EJ
If {A, B, C, E} and {A, B, E} are equivalent, then there exists a nonsingular

constant matrix P such that A = PAP - \ B = PB, = CP - 1 and E = E. With
these, it is straightforward to verfy
,2
e,
det PieS) = k det P 2 (s).

nd Sz(s) have the same
variant under the trans
O][sl-A
o I
-C
B-![P-
O]=[sl-}
-C
Hence their system matrices are strictly system equivalent.

Now we assume that {A, B, C, D} and {A, B, } are strictly system equi
valent. Then from (6-168), we have
e,
U(S) O
][SI-C
-A EBJ [SI-C-}
[X(s)
I
=
~][V(S)
Y(S)]
I
(6-177)
e,
) zero matrices, the first
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)
Using Coro])ary G-12, we can wrte U(s) as
Q2(S)]
U(s) = (sI - A)U(s) + P
(6-179)
where Pis a constant matrix. The substituton of (6cI79) into (6-178) yields
(6-176)
. s in
e,
s)]
P(sl A) = (sI - A)[V(s) - U(s)(sl - A)]

~ (sI - A)P(s)
(6-180)
which implies P(s)(sl-A)-l=(sl-A)-lp. Since (sl-A)-lP is a strictly

proper rational matrix, so must be P(s)(sl - A)-l. This is possible only if
pes) is a constant matrix. Hence we may replace pes) in (6-180) by P(s) = Po
to yield
sP-PA=sPo-AP o
nbers. Hence P 1(s) and

: left coprime. lf P 2(S) =
,rime if and only if {A, B}
theorem can be similarly
Q.E.D.
Jivalent dynamical equa
mce is a generalization of
which implies JP' = Po and PA = APo. Hence we have

P(sl - A) = (sI - A)P
and
PA=AP
Now we shall show that the P in (6-179) is nonsingular. Since U(s) is

unimodular, U- 1 (s) is also unimodular and can be expressed as
(6-182)
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
are equivalent if and only
(6-181)
+ P(sl- A)U l(S) + PP 1
whichbecomes, because of (6-181) and U(s)U- 1(s) == 1,

1- PP 1 =(sl - A)[U(s)(sl - A)U 1 (s) + U(s)P 1 + PU l(S)]
298
.IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTIFICATION
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
which implies the st

Now if G(s) is no
(;(s) = N(s:
where (S)D-1(S)is ~
of
U(s)B =(sl -A)Y(s) +8
D,
[
which becomes, by using (6-179),
We conclude that {
SimilarIy we can she
alent.
What remains to
left-coprime fraction
D I- 1N/. Tben there
(si -A)U(s)B + PB = (si - A)Y(s) +8

or
-N,
(s)B - Y(s) = (si - A)-1(8 - PB)
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.
where the leftmost m

D/N,. = NID,., we forn
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- _
The second and thirc

matrices of {D,., 1, N r
unimodular because i
(6-186). We show in
unimodular. From (6
U 12 D
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
Consider the two right coprime fractionsG(s)=N1(S)Dl1(s)=N2(S)D21(s).

Theorem G-13 implies the existence of a unimodular matrix T(s) such that
299
STRICT SYSTEM EQUIVALENCE
AND lDENTIFICATION
Dz(s) = Dl(s)T(s) and Nz(s) = N l (s)T(s).7 Hence we have,

PI
+ PUl(s)
1 0J[ Dl(s) I][T(S)

[ O 1_ -Nl(s) O O
(6-183)
whereas its right-hand

h sides are identically
L Consequently, P is
lation.
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)
+ E(s) = (N(s) + E(s)O(s))O -l(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
[-~~:~ ~]=[ -~(S) ~I -Zi;~ ~(s)I~~]
(6-185)
We conc1ude that {O, 1, N, E} and {O, 1, N, O} are strictly system equivalent.

Similarly we can show that all left-coprime fractions are strictly system equiv
alent.
What remains to be shown 8 is the strict system equivalence of right- and
left-coprime fractions of G(s). Consider the coprime fractions (;(s) = N r D; 1 =
DI-lN. Then there exist U l l and U lZ such that
l-hand side is a rational

we conc1ude B = PB.
Q.E.O.
Ull
[ DI
Ulz-I[- N"J = [IJ

NI _
D,.
O
(6-186)
where the leftmost matrix is a unimodular matrix (Problem G-ll). Now using
DN,. = NI O,., we form the identity
. have the same transfer

natrix, hwever, are not
cUon 4-3 is an example
i(s)D-l(s) = N(s)O-l(s),
:) are not right coprime.
O(s), 1, (s),O} have the
!o det O(s), they are not
~~l ~:z~~]r~-~--~~.--i]=r-ri---~l--~~[~ll U~,D< ~lZ]
[ =--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
}, which have the same
luivalent, they become
'educible; that is, {p(s),
neo We establish first a
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
. Q.E. O.
ivalent where G(s) is a
'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).
r matrix T(s) such that
__ .
..
300
~'~_J~'._"
.. ,". _ _ ','._'
~._-"---~-
IRREDUCIBLE REALIZATIONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTIFICATION

IDENT
Theorem 6-12
polynomial matrix 1
These results are di]
out the foHowing tra
AH irreducible polynomial matrix descriptions {PieS), Q(s), R(s), W(s)} which

have the same transfer matrix are sti-ictly system equivalent.
Proof
Every irreducible {Pi, Q, R, W;} has an irreducible realization {A, B, C, E(p)}
with dim A = deg det PieS). AH irreducible {A, B, C, E(p)} which have the same
transfer matrix are equivalent (Theorem 5-20) and, consequently, strictly
system equivalent (Theorem 6-10). Hence by the transitivity property, we
conclude that aH irreducible {Pi Q R W;} of the same transfer matrix
are strictly system equivalent.
Q.E. D.
Transfer matrix
Dynamical equation
Polynomial matrix
description
AH irreducible dynamical equations, all coprime fractions and aH irreducible

polynomial matrix descriptions which have the same transfer matrix are
strictly system equivalent and consequently have, following Theorems 6-2 and
6-8 and Definition 6-1', the following properties:
~(G(s)) ~ det(sI - A)~ det D(s) ~ det pes)
System matrix
(6-189)
where ~(G(s)) is the characteristic polynomial of C(s) and

denotes the
equality of two polynomials modulo a nonzero constnt factor. Conversely, if
all descriptions have the same transfer matrix and satisfy (6-189), then they
must be all irreducible. Under the irreducibility assumption, any one of the
descriptions can be used, without loss of any essential information, to study
and design a system.
To conclude this section, we mention that system matrices which are strictly
system equivalent to
pes)
Ses) = [ -R(s)
Hence the discussion

time case. Note that
the discrete-time case
equivalently, s- I is re]
There is, however
special discussion. e
time system expanded
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
where P, Q, R, and W are, respectively, m x m, m x p, q x m, and q x p polynomial

matrices, can be generated by the foHowing elementary operations:
1. Multiplication of any of the first m rows or columns by a nonzero constant.
2. Interchange of any two of the first m rows or columns.
3. Addition of the multiple of any of the first m rows (coiumns) by a poly
nomial to any of the m +q rows (the m + p columns).
These operations can be readily derived from the unimodular matrices
Ves) O] [ves) y(S)]

[ X(s) I q ' O
Ip
used in the definition of strict system equivalence.
*6-9 Identification of Discrete,.Time Systems

from Noise-Free Data
In the previous sections, we introduced various realization methods for con
tinuous-time systems described by transfer matrices. We also introduced
~-----
9Consequently melhods ar
M k =I~ tkg(r)dt, k =0, l,2
,D IDENTIFICATION
polynomial matrix description and the concept of strict system equivalence.

These results are directly applicable to the discrete-time systems if we carry
out the following transformations:
, Ri(s), W;(S)} which

1.
Continuous-time systems
ition {A, B, e, E(p)}

.vhich have the same
msequently, strictly
;itivity property, we
ame transfer matrix
Q.E.D.
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)
ns and a1l irreducible

transfer matrix are
19 Theorems 6-2 and
s)
301
IDENTIFICATION OF OISCRETE-TIME SYSTEMS FROM NOISE-FREE DATA
which are strictly
g(z) = h(O) +h(l)z- I +h(2)z- 2
,and q x p polynomial
.perations:
>y a nonzero constant.
;.
(columns) by a polymodular matrices
+ ...
As in the continuous-time case, we sha1l ca1l {hU), i = 0, 1, 2, ... } the Markov

parameters of g(z). In the continuous-time case, the Markov parameters must
be computed from dig(t)/dt i , i = 0, 1, 2, ... , at t = 0, where g(t) is the impulse
response of the system or the inverse Laplace transform of g(s). The generating
of an impulse as an input is not possible in practice; repetitive differentiations
of g(t) are again impractical. Hence the Markov parameters in the continuous
time case are not really available. 9 In the d.iscrete-time case. the situation 1S
entirely different. If we apply the input {u(O) = 1, u(i) = 0, i = 1, 2, ... } to an
initia1ly relaxed linear time-invariant discrete-time system, then the measured
data at the output terminal are the Markov parameters, that is, y(i) = h(i), i = 0,1,
2, . . .. Hence the realization from the Markov parameters in the discrete-time
case can be considered as an identification problem-a problem of determining a
mathematical description of a system from the data obtained by direct measure
ment at the input and output terminals. In actual measurement, all data will be
corrupted by noises. A study of the identification problem. with noisy data
requires the concepts of probability and statistics and is outside the scope of this
:ems
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.
In the following, 'i

an arbitrary input-~Ue
introduced. The prot
Consider a linear
transfer matrix (;(z).
where D(z) and N(z)
Let the highest degree
expressed as
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
to the initially relaxed
h(O)
The substitution of I
Equating the coefficien
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
IDENTIFICATION OF OISCRETE-TIME SYSTEMS FROM NOlSE..FREE DATA
.ND IDENTIFICATION
In the following, we shall discuss a method of identifying a system from

an arbitrary input-output pair. The concept of persistently exciting will be
introduced. The problem of nonzero initiai conditions will also be discussed.
Consider a linear time-invariant discrete-time system with q x p proper
transfer matrix G(z). Let G(z) be factored as
of noise and call the

~te-time
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)
where D(z) and N(z) are, respectively, q x q and q x p polynomial matrices.

Let the highest degree of all entires of D(z) be v. Then D(z) and N(z) can be
expressed as
and
D(z) = Do + D 1z + ... + Dvzv

N(z) = No +N 1z + '" +Nvzv
(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)
The substitution of (6-190) into (6-194) yields D(z)y(z) = N(z)u(z) or

') arranged in matrix
. a system can be ob
assertion also applies
(6-195)
- N(z)u(z) + D(z)Y(z) =0
Equating the coefficient of z, i = v, v - 1, ... , - 00, to zero yields
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)
where M = [ - No Do ... - N v D v] is a q x (p + q)(v + 1) matrix and the

matrix SJ - v, 00) has (p +q)( v + 1) number of rows and an infinite number of
columns. This equation is applicable no matter what input sequence is applied.
Hence, given an arbitrary input-output pair, the identification problem reduces
to the search of M to meet (6.. 196). There are q rows of M. In order to have q
rows of nontrivial s~lutioris in (6-196), we !leed q linearly dependent rows in
Sv( - v, 00). Since it is desirable to have v, the degree of D(z) and N(z), as small
as possible, we search, roughly speaking, the first q linearly dependent rows of
S.{-v, 00). Hence the ldentification problem reduces to the search of linearly
dependent rows of Sv( - v, 00) in order from top tobottom. This problem is
similar to the coprime fractiori problem discussed in Section G~4 .
impossible to identiry the

1ifthe degree of the system
/(z - 1) + l/z ' OO may also
lO way to identify a system
ding a transfer function to
,lization problell1 (see Refe..
:h infinity or remain oscil..

lhe system will salurate or
. the system can be broughl
cable nomatter the system
.1..
304
IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFlCATION

IDENTlFi
The systems to be identified are assumed to be causal and to have proper

rational transfer matrices. Hence if u(k) = O for k < 0, then we have y(k) =0 for
k<O. Define
(k,/)~[u(k)
y(k, I)~ I[y(k)
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)
----------
()( =0,1,2 ...
(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
y(k +IX, I +IX)
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-+~-~
An input sequence {u(n)} is called persistently exciting if every u row in (6-199)

is linearly independent of its previous rows in S,,(k, 1).
This definition, as such, is not well defined because of its dependence on k, 1,

and Ci.. The integer k is usually chosen as 0, although other value is also per
mitted. In theory, the integer I should be infinity; in practice, it is chosen so that
S,,(k, 1) has more columns than rows. The integer IX should be equal to or larger
than, as will be discussed later, the observability index of any irreducible realiza
tion of G(z) or the largest row degree of row reduced D(z) in any left-coprime
fraction of G(z) = D -1(z)N(z).
The persistent exciting of an input sequence is not defined solely on the input
signal; it also depends on the output it generates. In other words, whether or
not an input is persistently exciting depends on the system to be identified.
An input sequence which is persistently exciting to a systein may not be so to a
different system. Roughly speaking, an input sequence must be sufficiently
random or .rich in order to identify a system. Since the space spanned by the
rows of Si k; 1) is of a finite dimension, whereas al! possible input sequences
i: ~
Y(k +v -1,
(k +-v: 00)
y(k
with
Definition 6-4
~~k_ ~}~ 00)
+v,
00)
0::$ ro :::$ l' 1 ::$ ... ::$1
j ~ v. Note also that alll
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
12The subsequentanalysis is smil;
_~~
__.__..
..
,~._~~._~_~.~c.~==~~==========
IDENTIFlCATION OF DlSCRETE-TIME SYSTEMS FROM NOlSE-FREE DATA
E, AND IDENTIFICATION
form an space of an infinite dimension, almost aH input sequences are per

sistently exciting. It means that if an input is generated randomly, the prob
ability for the input to be persistently exciting is almost lo
Consider So(k, 00).12 We use, for example, the row searching algorithm to
search the linearly dependent rows of So(k, 00) in order from top to bottom.
Because is persistently exciting by assumption, aH dependent rows will appear
in y(k, 00). Let Yo be the number of dependent rows in y(k, 00). Clearly Yo'::;; q,
and q - Yo is the number of linearly independent rows in y(k, 00). Next, we
apply the row-searching algorithm to SI(k, 00). Let y[ be the number oflinearly
dependent rows in y(k + 1, 00). Let Yi(k, 00) be the ith row of y(k, 00). If
Yi(k, 00) is linearly dependent in So(k, 00), then y(k+ 1,00) will also be linearly
dependent in SI(k, 00). This foHows from the fact that (k, 00) =
[ u(k) : (k + 1, oo)J and y(k, 00) = [y(k): y(k + 1, 00)]. Because of this property,
we have Yl '2:Yo. We continue this process until Yv =q as shown.
.sal and to have proper

hen we have y(k) =0 for
(6-197 )
(6-198)
,2 ...
305
(6-199)
(k, 00)
nsists of p rows of and

:(l-k+1). WecaHthe
l y Y rowS.
.s much more than neces
'rom MS.(O, 1) = O, with 1
The range of data {k, l}
s initiaHy relaxed. How
nonzero initial conditions
~~k.! .ex:} _____

(k +1, 00)
} Yo [no. of dependent rows in y(k, 00)]
+ 1, 00)
-----------
} Y[
y(k
Sv(k, 00) =
(k
(6-200)
+v -1,00)
y(k+v-1,00)
}Y v -l
-----------(k +v, 00)
y(k +v, 00)
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
.g if every u roW in (6-199)

I
~ of its dependence on k, 1,
'h other value is also per

:ractice, it is chosen so that
lOuld be equal to or larger
. of any irreducible realiza
d D(z) in any left-coprime
defined solely on the input
n other words, whether or
.e system to be identified.
system may not be so to a
jence must be sufficiently
: the space spanned by the
1 possible input sequences
assumption of persistently exciting.

y rows in Sv(k, 00) is clearly equal to
n~ number
The number of linearly independent
of linearly independent y rows

v-l
=(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
12The subsequent analysis is similar to the one in Section G-4.
(6-202)

--~-
---
, ,.. ,._"-~
,~~""-~
306
",._..,
,_.
-.~-'''-~'~.''.'_.~
IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTIFICATlON

IDENTII
As an example, suppose q = 3 and

(k, 00)
Yl(k, (0)
Y2(k, 00)
y3(k, 00)
--------(k +1, 00)
Yl(k +1, 00)

Y2(k+1,00) =
Y3(k+l,00)
x
--------u(k +2, 00)
x
Yl(k +2, (0)
Y2(k +2, 00)
Y3(k +2, 00)
~ }r
Because of the assum

d~3 = 1, and all elemei
are zeros. In additiOJ
for 1= 1, 2, 3, .. " Hel
(6-203)
and
where K is a lower triangular matrix with 1 on the diagonal, as discussed in

(A-7) of Appendix A; and x denotes nonzero rows. Then we have VI = 2, V2 = O,
and V 3 = 2, and y 1 has index 2, y2 has index 0, and y3 ha$ index 2.
The Y, i = 1, 2, ... , q, which first becomes linearly dependent on its previous
rows in SI' are caBed the primary dependent rows. For example, the Y2 in the
first block row of (6-203) is the primary dependent row; the Y2 in the second
and third block rows are not. The y1 and y 3 in the third block-row of (6-203)
. are .the primary dependent rows. The positions of the primary dependent rows
are clearly determinable from the index of y. The primary dependent yi row
appears in the (Vi + l)th block row of SI"
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) >
Consider an initiaBy relaxed system excited by a persistently exciting input

sequence. We form Sa(k, 00) and search its linearly independent rows in order
[rom top to bottom to yield KS v =Sv, where K is a lower triangular matrix with
1 on the diagonal, and V is the first integer such that aH y-rows in the last block
row of SI' are linearly dependent. Let [ - No Do:' .. : - N v D.,,] be the q
rows of K corresponding to the q primary dependent rows of SI" Then the
transfer matrix of the system is given by G(z) = D- 1 (z)N(z), where
D(z)=D o +D 1 z+'" +Dvzv
N(z) = No +N 1 z + ... +Nvzv
1\
This implies deg det D(z

the n =
Vi computed in
rore we conclude that N(
fer matrix is equal to del
That D(z) is in the pol
in Appendix G. This co
(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
(see Appendix G).
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~
This D(z) is clearly ro,

however, is not needed
Theorem 6-13
and
+ d~ lZ + di
D(z)= dg 1 +dLz
Persistently exciting
variable system can b~ a
14 The matrix is essenlially in
307
. IDENTIFICATlON OF DISCRETE-TlME SYSTEMS FROM NOlSE-FREE DATA
.ND IDENTIFlCATION
(6-205 )
(6-203)
Because of the assumption VI = 3, V2 = 1, and V 3 = 3, we have df = 1, d~2 = 1,

d~3 = 1, and aH elements in (6-205) on the right-hand sides of dfl' dL, and d~3
are zeros. In addition, the columns associated with dit 1, d~;Z, d%;l are zero
for 1= 1,2, 3, .... Hence the Di reduce to 14
[2
O d 21 3 '' :~)
'1"' O O]
do11 do12 dIo'
3 ' d 111 O dI1 3 '' (11
[Do DI D 2
dLQ)O : O O O : O O O
2
1
l ' d2
do31 do32 do'
O d 33
, 31 O d 33'' O O'1-'
: '
33 ' d 31
and
d~1 +dLz+dLz2+z3
d~2
d~3 +dt3 Z +d"f3 Z2
D(z)= d~1 +d~lZ
d~2 +z
d~3
(6-206)
d~1 +d11Z+d~IZ2
d~2
d~3 +d13Z +d~3Z2 +z3
This D(z) is clearly row reduced. It is also column reduced. (This property,
however, is not needed here.) Therefore we have
D3J=[d~1 d~2 d~3:
.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
sistently exciting input

lependent roWS in order
r triangular matrix with
y-rows in the last block
.: -:N, D J be the q
t rows of S,. Then the
N(z), where
(6-204a)
(6-204b)
=deg det D(z) and D(z) is
olynomial echelon form
. an example to prove the
deg det D(z) = VI +v 2 + v3 = n
(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.
An example will be given, after the diseussion of nonzero initial conditions,

to illustrate the identification procedure. We note th.at once the transfer
matrix, in the coprime fractional fom, of a system is identified, a dynamical
equation description can be readily obtained by using the procedure discussed
in Section 6-6.
Persistently exciting input seq.uences. The identification of a multi
variable system can be achieved by searching the linearly dependent rows of
14
The matrix is essenlially in the echelon formo See Appendix G-l
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)
y(z) = -1- u(z)
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
has a full row rank. '

full row rank. The CO]
is the Hankel matrix o
tions for all rows of U(
from the results in Seco
and
where u(z) is the th COi
tions state that the inp
ifwe useS v (l, 00) in the i
Hankel matrix of u(z)/"
to be linearly independ
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
ofsystems which are not

and observable discrete-I
(6-211 )
y(l,oolJ
and the solution is also equal to g(z)=(l-z)/l. Since all u rows of SI in

(6-211) are linearly independent by themselves, there is no way to check from
SI that g(z) is erroneous. Hence the persistent exciting of {u(n)} cannot be
defined solely on {u(n)}.
A necessary condition for an input sequence to be persistently exciting is
that the matrix
U(k' (0)
_ u(k+l,oo)
(k, (0) .
[
u(k +v, (0)
y(z) = C(z]
y(z)
where R(z) is a q x 1 poly
to the corresponding ro",
We rewrite(6~217) as
(6-212)
[ -R
lDENTlFICATION Of DISCRETE-T1ME SYSTEMS FROM NOISE-FREE DATA
, AND lDENTlFICATlON
el. Clearly, lmust be

olumns than rows. In
ent in Sv(k, 1), then J
'uted by using the row
ow reduced and, con
ot persistently exciting,
For example, consider
,ut y(:) is equal to Z-I.
.]
(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)
lrtunately, this problem
independent roWS of Si'
equation of (6-196) by
s at the lower half. For
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
x(k + 1) = Ax(k) + Bu(k)

y(k) = Cx(k) + Eu(k)
O]SI
(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]
rows; hence /(z) is not

FIrst linearly dependent
309
(6-215a)
(6-215b)
(6-211)
The application of the z-transform to (6-215) yields

y(z) = C(zI - A)-lZX(O) + [C(zI - A)-IB +E]u(z)
lce al1 u rows of SI in
s no way to check from
ing of {u(n)} cannot be
(6-216)

y(z) = O -l(z)R(z) + O -l(z)N(z)u(z)
(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
IRREDUCIBLE REALlZATIONS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATION
This equation is similar to (6-195) and implies
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 -1) u(v)

y(v -1) y(v)
u(v
y(v
MSi - v, (0) =
+ 1) ...
+ 1) .. ,
IDENTIFlC
hence the Ni, Di compl

with the Ni, Di computt
puted by solving success
In other words the solut
MS v( - v, oc!) = O. Hent
and the ones computed J
lar polynomial matrix (1
matrices are the same.
Example 1
Consider a system with t
(6-219)
whereM=[-R o -No Do:"': -R,. -N v DvJ, andSv(-v, (0) denotes

the v(p + 1 + q) x 00 matrix in (6-219). Now we require the input [1 u'(z)]',
where the prime denotes the transpose, to be persistently exciting. This input
[1 u'(z)]' violates the conditions in (6-213b); hence we cannot use Sv(k, 1)
with k;:,:: 1, in the identification. This new input however may meet the condi
tions in (6-214), hence by using Sv( -v, 1), the system with its initial conditions
can be identified.
ls it really necessary to identify the initial conditions? In the design, it seems
that initial conditions are never required. What is essential is the transfer
function 01' dynamical equation. Furthermore, if the system is not relaxed at
n =0, and ifthe system is controllable and observable, then we can always find an
input sequence {u(n), n = - no, - no + 1, ... , -1} so that the system is relaxed
at n = - no and has the given initial state at n = O. Since the initial time is not
critical in the time-invariant case, a system with nonzero initial conditions at
n = Ocan be viewed as a system with zero initial conditions at 11 = - no. Con
sequently, (6-196) with a proper modification [since u(n) and y(n) are no longer
zero for n < OJ can be used in the identification of the system. Since (6-196)
contains more than enough equations and since we do not have the information
of o(n) and y(n) for n < 0, we simply use MSik, ro) = O, fOl" any k 2: 0, in th;;
identification. Consequently, given a system and given {u(n), y(n), i = O, 1,2, ...},
no matter the initial conditions of the system are zero 01' not, its transfer matrix
can be obtained by solving MSiO, (0) = O.
Theorem 6-14
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
With the initial statex(O) =

pair
k
u,(k)
li2(k)
.I',(k)
O
05
O
3.5
1
05
-0--0--=-1-
Now we shall try to id

{u(n), y(n)}.First we aSSurr
row-searching algorithm to
1
O
_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
IDENTlFICATION OF DISCRETE-TIME SYSTEMS FROM NOISE-FREE DATA
:, 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
Consider a system with transfer matrix
z~ 1]
(6-219)
and Sv( - v, (0) denotes

~e the input [1 u'(z)]',
lly exciting. Thi~ input
we cannot use Sv(k, 1)
ter may meet the condi
ith its initial conditions
;? In the design, it seems

essential is the transfer
;;ystem is not relaxed at
en we can always find an
lat the system is relaxed
lce the initial time is not
,ero initial conditions at
tions at n = - no. Con
'1) and y(n) are no longer
e system. Since (6-196)
:lOt have the information
=0, for any k O, in the
{u(n), y(n), i =0,1,2, .. '}'
)r not, its transfer matrix
y(n), n=O, 1,2, .. '}' the
by assuming zero initial

Jt the assumption of zero
as MSJO, (0) = O. First

l, (0) reduces to Sv(1, (0);
(6-220)
2z
z-1
An irreducible realization can be found as
x(k
+ 1) =[~
With the initial statex(O) = [0.5

pair
O
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
IRREDUCIBLE REALIZATIONS, STRICf SYSTEM EQUIVALENCE, ANO IDENTIFICATION
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
f:; --0- ---------

t~-
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)
It can be readily verified that D-I(z)N(z) = G(z).

Now we assume that the system is not initially relaxed. In this case we must
use (6-219). The application of the row-searching algorithm yields
1
O
O
-0.5
-5
O
O
O
:-3.5
-3
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
The last two rows of E
=[
Hence we have
which'are the same as

For this example, we c;
Hence the result is corr
6-10
Concluding
In this chapter, we disc

for proper rational mah
then use the procedure
second approach com:
formula in (6-29) and (6
singular value decompm
realizations. In the las!
coprime fractions of st]
this approach is the COI
Appendix G can be em
irreducible realization m
the Jordan-form realizati
reader is referred to Ref
for example, References,
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
rst column is chosen to
!(O, 11) a zero column.
. the first two elements,
Jf the pivot element in
out by hand, the pivot
. this problem, we have
f K, which correspond
computed by using the
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
The last two rows of K can be computed as
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)
Hence the result is correct.
6-10
d. In this case we must
rithm yields
Concluding Remarks
In this chapter, we discussed three approaches to find irreducible realizations

for proper rational matrices. The first approach is to find a reducible one and
then use the procedure in Section 5-8 to reduce it to an irreducible one. The
second approach computes the Markov parameters by using the recursive
formula in (6-29) and (6-30) and form the Hankel matrix. We can then use the
singular value decomposition or the row searching algorithm to find irreducible
realizations. In the last approach, irreducible realizations are obtained from
coprime fractions of strictly proper rational matrices. The majar effort in
this approach is the computation of coprime fractions and the procedure in
Appendix G can be employed. Allhough we have intro.duced a number of
irreducible realization methods;the treatment is not exhausve. For example,
the Jordan-form realization IS not discussed for the general case.' The interested
reader is referred to References S38, S59, and S145. For other methods, see,
for example, References, S62, S63, and S191.
,,,.,
.~.~~.'
.-y
-~
_ - - _ .. ,.
"
~._,.,
, . '._.,_.".",~ . ."_.
_ , . ~_ . ' .
.'_'
'
."
"
0 ' ' ' _ _ "...
- - - - ~ - ~_------~--'--~--~----,------~--_
314
~ _
.............-
IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, ANO IDENTIFICATlON
Now we compare briefly the numerical efficiencies of these three approaches.

Given a q x p strictly proper rational matrix. Let a(s) and flj(s) be respectively
the (least)common denominators ofthe ith row and jth column of G(s). Because
the computation ofthe least common denominator is not a simple task, we may
simply multiply all denominators of the ith row and the jth column of G(s)
to yield a{s) and flis). For simplicity, we assume p = q and a = deg a(s) = deg
!3j(s), for all i,j. Then we can find a controllable but not necessarily observable
(or vice versa) realization of dimension ap. To reduce this reducible realiza
tion, we must append B and e to A; hence the matrix under operation is roughly
of order (a + l)p. In the Hankel matrix, the order of the Hankel matrix is
(a + l)p x ap. In the coprime fraction method, if we use the method discussed
in Appendix G to find a coprime fraction, we must form a generalized resultant.
Each block row has (p + q) = 2p rows, and we need roughly (a + 1) block rows.
Hence the generalized resultant has 2(a + l)p number of rows. Thus we con
elude that the matrices used in the first two approaches are roughly of the same
order; the matrix used in the third approach is about twice the sizes of those
used in the first two approaches.
The operation in the first approach requires the triangularization of a matrix
by similarity transformations (both column and row operations). In the second
and third approaches, we require the triangularization of a matrix by either row
or column operations. Hence the first approach requires twice the computa
tions of the second and third approaches. Thus we conclude that the Hankel
method is probably the most efficient. For a more detailed comparison, see
. Reference S202. We note that all three approaches can.be implementeqby using
numerically stable methods.
It is of interest to compare the singular value decomposition and the row
searching method discussed in Section 6-5. They are implemented on UNIVAC
1110 at Stony Brook in Reference S202. The row searching algorithm is carried
out by applying Householder transformations and the gaussian elimination
with partial pivoting on the columns of Hankel matrices and then computing
the coefficients of combinations by back substitution. The scaling problem is
included in the programo The examples in Reference S202 do not show any
substantial difference between Householder transformations and the gaussian
eiimination with partiai pivoting. The singuar vaiue decomposition is most
reliable. However, its reliability over the row searching method is not over
whelming in the examples in Reference S202. This seems to be consistent with
the remarks in Reference S82, p. 11.23.
The identification of linear time-invariant discrete-time systems is studied.
The identification is carried out from a set of arbitrary input-output pair so
long as the input sequence is persistently exciting. Whether the system is
initially relaxed or not, the transfer functon matrix in coprime fractional
form can be obtained by solving MSv(O, ex)) = O. The realization problem can
be considered as a special case of the identification problem in which the system
.. is to be identified from a particular input-output pair, the impulse response or
the Markov parameters. The identification method introduced is applicable
to continuous-time systeins ir they are first discretized. This, however, will
introduce errors due te

will not be discussed.
In this chapter, we
established its relation
We showed that if the 1
irreducible, then they a
them can be used in the
tion.
The degree of a prc
degree can be compute
1. Compute the least c
we have deg G(s) = (

2. Compute the Mark,
shown in (6-80). n
computed by using t
3. Find a right fraction
deg G(s) = deg det D
coefficient matrices e
We then search the liJ
Then we have
deg G(s) = too
Similar results can b(
4. Consider the realizat
is controllable but nc
QIO =0,
Qlj = R,
QiO =0, Qij =Q(i-lH.
Using Qij, the observ
Then we have deg (;(5

5. Find a realization of
Section 5-8, to an irred
These methods can be
parison of these method~
stability seems to be una\
To conclude this chal
realization in Equation (f
in algebraic system theor
q x p strictly proper ratic
CONCLUDING REMARKS
AND IDENTIFICATION
hese three approaches.

nd f3 J.(s) be respectively
,lumnofG(s). Because
t a simple task, we ~ay
he jth column of G(s)
and ex = deg o:;(s) = deg
necessarily observable
this reducible realiza
ler operation is roughly
, the Hankel matrix is
;: the method discussed
a generalized resultant.
ghly (o: + 1) block rows.
)f rows. Thus we con
are roughly of the same
twice the sizes of those
A
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
is controllable but not necessarily observable. Define

mposition and the row
plemented on UNIVAC
hing algorithm is carried
he gaussian elimination
ces and then computing
The scaling problem is
: S202 do not show any
lations and the gaussia n
: decomposition is most
ing method is not over
:ms to be consistent with
-time systeros is studied.
uy input-output pair so
Whether the system is
x in coprime fracUona\
realization problem can
blero in which the system
. the impulse response or
introduced is applicable
~ed. This, however, will
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
Then we have deg (;(s) = rank V (why?).
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
IRREDUCIBLE REALIZATlONS, STRICT SYSTEM EQUIVALENCE, AND IDENTlFICATlON
are respectively q x q and q x p real polynomial matrices. It is assumed that

D(z) is row-reduced and its row degree is Vi' that is, b,.iD(S) = Vi' Clearly we
have b,.N(z) < Vi'
Let IK D denote the set of all q x 1 real polynomial vectors with row degrees
smaller than Vi' It can be readily verified that (IK D , IR) is a linear space over the
field of real numbers IR. The dimension of (IK D , IR) is n~ VI + Vz + ... + Vq.
(See Example S, on page 16.) In this polynomial vector space, we shall choose
the n columns of the q x n matrix:
1 z
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)) =
Note that the ith columl

given by the ith column
Now we show that th
D-l(Z) =Dh;.I[H(z)(1 + f
=Dh~.I(I+E1Z-l
for some constant E, i =

that if j < Vi, then (D -1(;
This completes the deriv;
three maps in (6-226). l
maps in (6-226) are useful
SS5 and Section 9-8.
6-1 Find the degrees and

matrices.
as shown in (6-130). The ith column of A is the representation of OD(Zqij) with

respect to the basis in (6-223). Indeed, if we rewrite (6-12S) here as
D(z) = H(z)D h ,. + [(z)D,.
Because Dh;.1 is a constal

Because D -1(Z)[(Z~
D- 1(z)[(z)D,.D';;.I. HeI
O.
(6-224)
If f(z) is a polynomial vector, then O(f(z)) = O; if f(z) is strictly proper, then

O(f(z)) = f(z). It is a linear operator which maps (lRq(z), IR) into itself. Next
we define, for any q x 1 real polynomial,
OD(h(z)) = D(z)O(D -1(z)h(z))
This is the first column

columns of A except tr
(6-131a). Consider now
0D(.H(z)) = D(z)O(
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
l ' ~ l' +:)~d2)J

1
1)'
(s +2)
(s+1)(s+2)
_._-_
----,-------------_.
---_...
.. -_._---._..- ..
PROBLEMS
AND IDENTIFICATION
es. It is assumed that

iD(S) = Vi' Clearly we
Ir i = 1,j = 1, then D- 1(s)Zqll is strictly proper and
o
1
ctors with row degrees

a linear space over the
; n~v1+V2+'" +vq
space, we shall choose
O
O
z Vq -
(6-223)
... , q; j = 1, 2, ... , Vi'
tor TI as
z)
(6-224 )
) is strictly proper, then

z), IR) into itself. Next
(6-225)
f. Because the ith row

D(h(z)) is(IK D , IR).
lressed by the following
(6-226)
317
TID(Zqll) = D(s)D- 1(s)Zq11 = zqll
= q12 = L(s) O
O
This is the first column of A as shown in (6-131a). Proceeding similarly, all

columns of A except the (~ = 1 vm)th columns can be obtained as shown in
(6-131a). Consider now
TID(H(z = D(z)TI(D- 1(z)H(z = D(z)TI[Dh~l - D- 1(z)L(z)D .D';;.1 ]
"
Because D';; 1 is a constant matrix (a polynomial of degree O), we have TI(D~.l) =
O. Because D-1(Z)L(z)D1J.Dh~1 is strictly proper, TI[D -1(z)L(z)D 1r Dh";.1 ] =
D- 1(z)L(z)D l.. D,;;'1. Hence we have
nDCH(z)) = - D(z)D-1(Z)L(z)Dl.. D,~1 = - L(z)D 1,.D;;' 1
Note that the ith column of H(z) is zqv,. Hence the (2:::n= l)th column of A is
given by the ith column of -Dl,.D/~.1, as shown in (6-131a).
Now we show that the ijth column of e in (6-131a) is equal to (D-1(z)qij)_ l'
D-1(Z) =D~.l[H(z)(I + H- 1(z)L(z)Dl")] -1 ~D/~.l(I + H-1(Z)L(zp.. t 1iI- 1(z)
=D,;;.1(I+E 1z- 1 +E 2z- 2 + .. )fi- 1(z)
for sorne constant E, i = 1, 2, . . .. From this equation, we can readily show
that if j<v, then (D-1(Z)qij)_1 =0 and (D-l(z)qv}=ihe ith column ofD,;;.l.
This completes the derivation of the observable realization in (6-131) from the
three maps in (6-226). For a derivation of (6-226), see Reference Si02. The
maps in (6-226) are useful in solving polynomiaJ matrix equations; see Reference
SS5 and Section 9-S.
en (D- 1(z)X)_1 ~Cl1' If

then the map RD is the
Find the degrees and the characteristic polynomials of the following proper rational
matrices.
6-1
~ntation of TID(zqJ with
-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
+1)(s +2) .'
--~----
318
IRREDUCIBLE REALIZATlNS, STRICf SYSTEM EQUIVALENCE, AND lDENTlFICATlN
+3l
+ 1
+1
6-6 Find an irreducibll

an unobservable dynami(
dynamical-equation reali
6-7 Find the controll~

lordan canonical-form d:
Find the dynamical-equation description of the block diagram with state variables
chosen as shown in Figure P6-2.
6-2
6-8
Set up a linear time

(p3
+7. 1(t),
where pi ~ d/dti.
Figure P6-2
Find irreducible ce
matrices
6-9
Find the dynamical-equation realizations of the transfer functions
6-3
S4
S2
b. SS
_ S4
a'l l' +% ~2X' +3)J
+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
Find lordan-canonical-form dynamical-equation realizations of the transfer func
6-4
+2s +2
s(s +1)2(s +4)
2s +3
(5 +W(s +2)
s(s
Find irreducible re2
tions
a.
(S
+ 1)(s + 2)(s +3)
S2
+ 1
b. - - "
(s +2)3
S2
c.
S2
Use two different methods
+1
+2s +2
6-11
Find irreducible rea
lf the lordan-form realizations consist of complex numbers, find their equivalenl dynamical
equations that do not contain any complex numbers.
6-5 Write a dynamical equation for the feedback system shown in Figure P6-S. First
find an overall transfer function and then realize it. Second, realize the open-loop transfer
function and then make the necessary connection. Which realization is more convenient in
computer simulalions if the gain k is to be varied?
Use two different methods.

6-12 Find a linea((ime-in
fer function is
(s + 1) (s + 2)(s + 3)
Figure P6-5
I
I
[See Problem 6-3(a).]
PROBLEMS
AND IDENTIFICATION
319
6-6 Find an irreducible realization, an uncontrollable dynamical-equation realization,

an unobservable dynamical-equation realization, and an unobservable and uncontrollable
dynamical-equation realization of 1/(s3 + 1).
Find the controllable canonical-form, the observable canonical-form, and the

Jordan canonical-form dynamical-equation realizations of l/sO..
6-7
agram with state variables
6-8
Set up a linear time-varying dynamical equalion for the differential equation

(p3 +a 1(t)p2 +C(z(t)p +C(3(t))y(t) = (f30(t)p2 + f31 (t)p
-~
+ f3z(t))u(t)
Find irreducible controJlable or observable canonical-form realizations for the

matrices
6-9
'unctions
a.
'l(S +l)(s ~2)(s +3)j

S2
+2s +2
s(s + 1)2(5 +4)

2s
+3
b. [ (s+1)2(s+2)
lalog computer simulations
6-10
ltions of the transfer func
+ 2s +2 ]
s(s+V
S2
Find irreducible realizations of the ralional matrix
S+2
ls:
s +1

6-11
j
~ :3J
5
+1
s +2
Find irreducible realizations of the rational matrix
their equivalent dynamical
lown in Figure P6-S. First

~alize the open-loop transfer
zation is more convenient in

Find a linear time-invariant discrete-time dynamical eq uation whose sampled trans
fer Function is
6-12
I
1
l.
[See Problem 6-3(a).]
320
IRREDUCIBLE REALlZATlNS, STRICT SYSTEM EQUIVALENCE, ANO IDENTlFICATlN
Find an irreducible, discrete-time dynamical-equation realization of the sampled

transfer-function matrix
6-13
Z~3J
Z+2
+1
[
Z
6-14
+1
+1
+2
6-17 Show that (6-77)

k=O, 1,2, ....
6-18
Show that (6-77) is
6-19
Consider the q x p
lt is assumed that there el
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
the equation can be transformed into
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.
Verify that the Jordan-form dynamical equation
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
Use Problem 6-19 a

irreducible one. (Hint: Shil
6-20
O 2:
[O O'
and then proceed. This pre

in (6-131) except the rearr;
References S63 and S21.)
can be transformed into

N(s) = N o
D(s) = Do'
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
where D" is assumed to be

dynamical equation
Write an irreducible dynamical equation for the following simultaneous differential

equation:
6-16
2(p+l)y +(p+l)yz=pu +uz

(p
wherep~d/dl.
+ l)y
+(p +l)yz
(p - l)u
y= [
l'
is a controllable realization
PROBLEMS
AND IDENTIFICATION
ealization of the sampled
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
Consider the q x p proper rational matrix

(;(s) = H(O) + H(l)s- 1 + H(2)s- 2 + ...
1t is assumed that there exist q x q constant matrices Ki, i =0,1,2, ... , v, such that
i= 1, 2, 3, ...
KoH(v + i) = - KH(v + i -1) -K 2H(v + i - 2) - ... -K,.HU)
]x
IY using the transformation
or
[K
Kv
K2 .. ,
Ko O .,. OJT=O
where T is defined as in (6-95). lf Ko is nonsingular, show that the vq-dimensional dynami

cal equation
=[= ~:~,~", ~ ~ ...
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
N(s)=N o +N l s+'" +N._ l s- l
D(s) = Do + D l s +. " +D.s = D.(D o + Dl s + ... + Is")
where D. is assumed to be nonsingular and Di = D; 1 Di' Show that the pp-dimensional

dynamical equation
lit matrix of order 2. [Hint:
to [XI X4 X2 Xs X3 X6J'
Q=diag(Q1> Q2' Q3)J
6]'
'ing simultaneous differential
1
O
-Do
-DI
-D 2
y=[ No
NI
N2
hU
is a control1able realization of N(s)D-1(s).
. ] ["
O
:
O
x+:
-~I'-l'
N.- 1 Jx
~,'
'----------_.~-~~~._-
322
IRREDUCIBLE REALIZATlNS, STRICT SYSTEM EQUIVALENCE, AND IDENTIFICATIN
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
6-23 Consider the irreducible dynamical equation lA, B,

write its grammians as
It is assumed that all eiger

system
q.
Using Theorem E-4 to
is also internally balanced.

(Hint: Write W=diag {W 1
and
where R*R = I and L 2 =diag [/,i, ;.~, ... , l.;':. Define
H~L:R:RcLc
Show that ror any sin.

is used in the identification
nonzero input sequence is al
is used?
6-26
Using the singular value decomposition (Theorem E-S), we write

H=RHLHQH
where Rt;R H = I and QHQt; = 1. Show that if P in Problem 6-22 is chosen as
i. Pi" = QHL c- I R
ii. POUl = Rt;L:R:
iii. P ib = Pi"LF = LFp OUl
Then its equivalent dynamical equation {A,
mians:
B, C} has the fol\owing corresponding gram
They are ca1led, respectively, input-normal, output-normal, and interna//y balanced on

[0, l] (see References SI61 and SI62).
Consider the irreducible dynamical equation
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
(Hinl: Use Corollary F-Ib.)

6-25
Identify a linear tim
O [
2
3
-----[
u dk)
[
-[ O
u2(k)
[
O
O -[
k
[
Yl(k) -[
Y2(k) -2 -0.5
2
[
[ -0.5
First identify the system by d

as welI as its initial conditior:
i. W" = 1, W01 = L~
ii. W" = L~, Wo, = I
ii. Wer = Wo , = LH
6-24
6-27
Consider the interna1ly balanced system partitioned as
6-28. Let {A, B, q be irre(

and N(s) are right coprime.
such that C(sl-A- lB = NI
N1(s)D-I(s), there exists a C I
PROBLEMS
AND IDENTlFICATlON
c:
,ns {A, B, C} and [A, B,

Je the controllability and
ow that
Jt is assumed that all eigenvalues of A have negative real parts. Show that the reduced
system
y= ex
x= Ax +Bu,
where
Using Theorem E-4 to
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
ldentify a linear time-invariant system from the following input-output sequence:

k .
_
u\(k)
u2(k)
)'I(k)
ving correspondng gram
)'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
irts and that the equation
.how that W is lhe unique
-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
In engineering, design techniques are often developed from qualitative analyses

of systems. For example, the design of feedback control systems by using
Bode's plot was developed from the stability study of feedback systems. In
Chapter 5 we introduced two qualitative properties of dynamical equations:
controllability and observability. In this chapter we shall study their practical
implications and develop sorne design techniques from them.
Ir an n-dimensional linear time-invariant dynamical equation is con
trollable, the controllability matrix [B AB ... An -1 BJ has n linearly
independent columns. By using these independent columns or their linear
combinations as basis vectors of the state space, various canonical forms can
be obtained. We introduce in Section 7-2 the most usefuI one: the controllable
canonical formo We a!so introduce there the observable canonjeal form
dynamical equation. These two canonical forms are very useful in the designs
of state feedback and state estimators.
Consider a system, called a plant, and a desired or reference signal; the
control problem is to find a control signal or an actuating signal so that the
output of the plant will be as close as possible to the reference signa1. Ir a
control signal is predetermined and is independent of the actual response of
the plant, the control is called an open-loop control. This type of control is nol
satisfactory if there are disturbances or changes in the system. Ir a control
signal depends on the actual response of the system, it is called a feedback or
closed-loop control. Since the state of a system contains all the essential in
farmation of the system,.if a control signal is designed to be a function of the
state and the reference signal, a reasonably good control can be achieved. In
Section 7-3, we study the effects of introducing a linear state feedbackof the
324
form u=r+ Kx on a <

the linear state feedba<
of an uncontrollable d
In state feedback,
outputs. This assump
want to introduce stal
mated from the availal
state estirnators undei
asymptotic state estim:
Both full-dimensional:
In Section 7-5 we a
We shall establish the :
and state estimator ca
not affect their intendl
not appear in the trans
output.
The inputs and ou
that is, every input COl
trolled by more than 01
input controls one and
to be decoupled. We:
system by state feedbac
system can be decouple
We' study in this el
The material is based il
78 to 80,84,99,111,11
can be extended to lir
stantaneously controlla
to References 10, 98, 99
7-2
Canonical-F<
Single-variable cas
single-variable dynamic
where A, b, e, and e are,

matrices. Note that th(
this section we shall dI
under the controllabilit)
Let the following dy
CANONCAL-FORM DYNAMCAL EQUATlONS
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
form u = r + Kx on a dynamical equation. We show what can be achieved by

the linear state feedback under the assumption of controllability. Stabilization
of an uncontroHable dynamical equation is also discussed.
In state feedback, aH the state variables are assumed to be available as
outputs. This assumption generally does not hold in practice. Therefore if we
want to introduce state feedback, the state vector has to be generated or esti
mated from the available information. In Section 7-4 we introduce asymptotic
state estimators under the assumption of observability. The outputs of an
asymptotic state estimator give an estimate of the state of the original equation.
Both full-dimensional and reduced-dimensional state estimators are introduced.
In Section 7-5 we apply the state feedback to the output of a state estimator.
We shall establish the separation property which shows that the state feedback
and state estimator can be designed independently and their connection w-[
not affect their intended designs. We also show that the state estimator wiH
not appear in the transfer-function matrix from the reference input to the plant
output.
The inputs and outputs of a multivariable system are generally coupled;
that is, every input controls more than one output, and every output is con
trolled by more than one input. If a compensator can be found such that every
input controls one and on]y one output, then the multivariable system is said
to be decoupled. We study in Section 7-6 the decoupling of a multivariable
system by state feedback. The necessary and sufficient condition under which a
system can be decoupled by linear state feedback is derived.
We study in this chapter only linear time-invariant dynamical equations.
The material is based in part on References 1, 4, 7, 17, 18, 36, 41, 46, 51, 52, 68,
78 to 80, 84, 99, 111, 112, S17, S23, S69, and SllO. The results in this chapter
can be extended to linear time-varying dynamical equations if they are in
stantaneously controllable (see Definition 5-3). The interested reader is referred
to References 10,98,99,110, and S119.
7-2
Canonical~Form Dynamical
Equations
Single-variable case. Consider the n-dimensional linear time-invariant,

single-variable dynamical equation:
X=Ax +bu
y =cx +eu
(7-1 a)
(7-1 b)
where A, b, c, and e are,respectively, n x n, n x 1, 1 x n, and 1 x 1 real constant

matrices. Note that the subscript 1 in FE stands for single variableness. In
this section we shall derive various equivalent dynamical equations of FE
under the controllability or observability assumption.
Let the followingdynamical equation FE
X=Ax +bu
y=cx+eu
(7-2a)
(7-2b)
326
STATE FEEDBACK AND STATE ESTlMATORS
be an equivalent dymmical equation of FE l , which is obtained by introducing

Px = Q - lX, where P ~ Q - 1 is a nonsingular constant matrix. Then
from (4-33) we have
A=PAP- l
h=Pb
c=eP-l
x=
eFE l
O
O
O
x=
O
The controllability matrices of FE l and FE l are, respectively,

Ab
Ah
=P[b Ab
U~[b
~[h
An-lb]
An-lh]
An-lb] =PU =Q-lU
-C/.."
(7-3)
(7-4)
Now if the dynamical equations FE l and, consequently, FE l are controllable,1

then the controllability matrices U and are nonsingular. Hence, from (7-4)
we have
P=U- l
Q=U- l
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)
Thus, if two same-dimensionallinear time-invariant, single-variable dynamical

equations are known to be equivalent and if they are either controllable or
observable, then the equivalence transformation P between them can be com
puted by the use of either (7-5) or (7-6). This is also true for multivariable
controllable or observable time-invariant dynamical equations. For multi
variable equivalent equations, the relation =}FU still holds. In this case U
is not a square matrix; however, from Theorem 2-8, we have U* =PUU*
and P = U*(UU*)-l.
Let the characteristic polynomial of the matrix A in (7-1a) be
Theorem 7-1
If the n-dimensional linear time-invariant, single-variable dynamical equation
FE l is controllable, then it can be transformed, by an equivalence transforma
See Theorem 5-15.
where (;( lo (;(2, .. ,Cl n a

and the f3's are to be
said to be in the contr
Proof
The dynamical equa1
of n x 1 column vecl
quently, the following
ql ~ A
and if FE l and FE l are observable, then
f3n
qn D. b
qn-l ~ A
q"-2 D. A
eA~-l
tion, into the form
y=[
is linearly independen1
from Figure 2-5 that i
th column ofthe new
to the basis {ql' q2' ..
Aql=(
CANONICAL-FORM DYNAMICAL EQUATIONS
)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,
llar. Hence, from (7-4)
g(s) =
/3 s"- +/3 s" - z + ... +/3
z
" +e
s" +IXS" + ... +IX"_S +IX"
(7-8)
Proof
(7-5a)
(7-5b)
The dynamical equation F E is control1able by assumption; hence the set

of n x 1 column vectors b, Ab, ... , A"- b is Iinearly independent. Conse
quently, the fol1owing set of n x 1 vectors
q" ~ b
qll- ~ Aq" +1Xq" =Ab +1Xb
q"-z ~ Aqll- +lX zqll=A zb +1XAb +lXzb
=VQ
(7-9)
q ~ Aqz +1X"_q"=A"-b +IXA"-zb + ... +1X"_b

is linearly independent and qualifies as a basis of the state space of FE . Recall
from Figure 2-5 that if the vectors {q, qz, ... , qll} are used as a basis, then the
th column of the new representation . is the representation of Aq with respect
to the basis {q, qz, ... , qll}' Observe that
(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)
Aq=(A"+ IX A"-+ ... +lY. ll _A+IX"I)b-lX"b
.,1 ~ l
<llll_01X"J
be
Aq,~q, -,q.~[q,
-
q,
... q.]
lLJ
.................................................
ble dynamical equation
~quivalence transforma
1
Aqll = qll- -1Xq" = [ql
qz
...
q"] [ : ]
_-IX
328
STATE FEEDBACk AND STATE ESTlMATORS
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
as one might expect.
The matrices A and b can also be obtained by using an equivalence transforma

tion. Let Q ~ [q1 qz ... qnJ ~ p- 1, and let x =Px or x=Qx, then the
dynamical eq uation FE 1 can be transformed into
This can be directly v

tions FE 1 and CFE 1
hence Q = U- 1 . I
columns of Q as new
puting UD-1, we see
that is,
Q =[q1
x=Q- 1AQx +Q- 1bu

y=cQx+eu
We have the folle

observable dynamical
The reader is advised to verify that Q - 1AQ = A or AQ = QA and Q -1 b = b.

The vector e is to be computed from cQ as
(7-11 )
Theorem 7-2
If the n-dimensional l
FE 1 is observable, thel
Hence the controllable dynamical equation F El has the equivalent controllable

canonical-form dynamical eq uation C F E l ' This proves the first part of the
theorem.
The dynamical equations FE 1 and eFE 1 are equivalent; hence they have
the same transfer function. lt has been shown in Sections 4-4 and 6-3 that the
transfer function of CFE 1 is equal to g(s) in Equation (7-8).
Q.E.D.
into the form
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
The dynamical eq uati

moreover, the transfer
g(s) =
[Ri A1bJ ... An- 1 bl

J
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,
This theorem can l

theorem of duality (Th
alence transformation
Equation (7-6): P = V
of the observable cano
matrix in (7-12). Hen
and (7-15) is given by
CANONICAL-FORM DYNAMICAL EQUATlONS
he state space, then A
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'
We have the following theorem, which is similar to Theorem 7-1, for an

observable dynamical equation.
Theorem 7-2
lf the n-dimensional linear time-invariant, single-variable dynamical equation

(7-11 )
equivalent controllable
es the first part of the
FE lis observable, then it can be transformed, by an equivalence transformation,
into the form
1 O
O
alent; hence they have

Ins 4-4 and 6-3 that the
'-8).
Q.E.D.
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 +
:/' -1-0:. 1sl1 -
(7-15b)
-1-'
+CX_lS
-~-(j~i1
This theorem can be proved either by a dircct verification or by using the

theorem of duality (Theorem S-lO). Its proof is left as an exercise. The equiv
alence transformation x = Px between (7-1) and (7-1S) can be obtained by using
Equation (7-6): P = V-IV. It is easy to verify that the observability matrix V
of the observable canonical-form dynamical equation OF E is thesame as the
matrix in (7-12). Hence the equivalence transformation x =Px between (7-1)
and (7-IS) is given by
1; ea = 1
[a"~,
an - z
P=
, az, ,.a no Therefore

is always controllable,
-~n-z X + ~n-z u
The dynamical equation (7-1S) is said to be in the observable canonical form;

moreover, the transfer function of FE is
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
STATE FEEDBACK AND STATE ESTIMATORS
form dynamical equa

We see from (7-9) that the basis ofthe controllable canonical-form dynamical
equation is obtained from a linearcombination ofthe vectors {b, Ab, ... , An-I b}.
One may wonder what form we shall obtain if we choose {b, Ab, ... , An - lb} as
the basis. Define ql ~ b, qz ~ Ab, ... , qn ~ An - lb and let x = Q - IX, where
Q ~ [ql qz .. , qnJ; then we can obtain the following new representation
(see Problem 7-2):
i~[! : g =~::} {r}
e,
01
The transfer function
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
Transform the following controllable and observable single-variable dynamical

equation
x~ [~
y=[O
-1
(7-17a)
2
O
(7-17b)
into the controllable and observable canonical-form dynamical equations.

The characteristic polynomial of the matrix A in (7-17) is
.
!l(A)=det
A-1
[ -~
Hence CX 3 =2, CX z = -9, and
(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
=0. The controllability matrix. U ls
U=[~1
: 1~]
where A, B, e, and E .
respectively. Let b bf
Ifthe dynamical eq
12
From Equations (7-5b) and (7-13), we have

Q=UA =
The controllable
also be obtained by
equation. The coeff
canonical-form eq uati
2 4 16] [-9 O1]

[11 62 128 O1 O1 OO
[-2 4 2l
-1 6 1J
3 2 1
The matrix Q can also be obtained from (7-9). From(7-11), we have
[/33 /3z
/3IJ=cQ=[3
lJ
Hence the equivalent controllable canonical-form and observable canonical-
U = [b
b z ...
bp
has rank n. Conseque

U. There are many v
[rom toe n x np compos
for choosing n linearly
state space. These two
CANON1CAL-FORM DYNAMICAL EQUATIONS
331
form dynamical equations are, respectively,

lonical-form dynamical
tors {b,Ab, ... ,A"- l b}.
~ {b, Ab, ... , A"- l b} as
.d let X = Q - x, where
ing new representation
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
and is easier to obtain

:owever, the usefulness
gle-variable dynamical
(7-17a)
(7-17b)
namical
7)is
-9).,
equ~tions.
+2
natrix U s
The controllable or observable canonical-form dynamical equation can

also be obtained by first computing the transfer function of the dynamical
equation. The coefficients of the transfer function give immediately the
I
canonical-form equations.
The controllable canonical-form and the observable canonical-form
dynamical equations are useful in the study of state feedback and state estimator.
They are also useful in the simulation of a dynamical equation on an analog
computer. For example, in simulating the dynamical equation FE l , we
generally need n 2 amplifiers and attenuators (potentiometers) to simulate A,
and 2n amplifiers and attenuators to simulate b and c. However, if FE is
transformed into the controllable or observable canonical form or into the form
of (7-16), then the number of components required in the simulation is reduced
from n 2 +2n to 2n.
*Multivariable case. Consider the n-dimensional linear time-invariant,

multivariable dynamical equation
FE:
X=Ax +Bu
y=Cx +Eu
(1-18a)
(7-18b)
where A, B, C, and E are n x n, n x p, q x n, and q x p real constant matrices,

respectively. Let b be the th column ofB; that is, B= [b b 2 b p ].
If the dynamical equation FE is controllable, then the controllability matrix
2l
62 lJ1
(7-11), we have
observable canonical-
has rank n. Consequently, there are n linearly independent column vectors in

U. There are many ways to choose n linearly independent column vectors
rom the n x np composite matrix U. In the following, we shall give two schemes
for choosing n linearly indepeildent column vectors to forro new bases for the
state space. These two schemes were first discussed in Section S-3.
-------_...._._--_._-------_.
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
where the X's denote'

zeros. This can be ea~
representation of Aqi ,
If the set ofvectors
will be of the form 3
,
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 ...
then the matrices
O 1
(7-21 )
O O
x;
x:,
x,
,
,
x :
- :
-(~~ ""X ~:f ~- O'

x :1
x :O
:'(_ _0_
(1
.
.,
-------------x,
,
x,
Ail, -1 b 3J
A and B will be of the forms
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.
Using footnote 2, sorne elen
-----
---
----- - -
CANONICAL-fORM DYNAMICAL EQUATIONS
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
- - -(~~ -x-fi;)- -- ~ 0- 0- -.-. ~ - -.;
is set is that the vector

preceding vectors; for
.ation of {bl> Ab, ... ,
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)
ler of (7-19); that is, we

" and then AZb, AZb z,
ectors. Note that if a
dence on the vectors
z, for k;:::l, can also be
'evious columns. After
ve rearrange them as
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
, ,
. ,, .
lifference between this

can be expressed as a
.s the vector AJL, b in
I1-lb}'
'b ... , A
l ' Al1'b 1
~ 1). z Similar remarks
__ - - - -
x:
_________ !_l ?__ ?
;is of the state space of
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,/1" then A"'b!

on{Akb"k=O, 1, ... ,/1,-1;
.,5,/125,'"
Using footnote 2, sorne elements denoted bY X can be replaced by' zeros.
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
First we show that t

FE{. Let X o and Xl 1
trollabili ty assumptio
in a finite time. No",
r(t) = u(t) - kx(t), then
that FE{ is controllac
We see from Figur
it generates u to contr
other words, if FE I is
(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)
which is obtained by replacing u in (7-24) by r +kx, where r is the reference

input. Note that the dynamical equations (7-24) and (7-25) have the same
dimension and the same state space. Now we shall show that the controllability
of a linear time-invariant dynaniical equation is invariant under any linear
state feedback.
Theorem 7-3
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.
We see that in the

invariance are not use
Corollary 7-3
The con troI1ability af

is invariant under any
Note that Theoren
p[b
Ab
for any 1 x n real cons

A1though state fe
equation, it is alway:
dynamical equation b
( -1/e)c, then the stat(
is observable. If e =0
feedback dynamical ec
Problem 7-14).
STATE FEEDBACK
335
unfilled positions are

:ction. The rnatrix e
the differences in the
vectors according to
iiagonal, whereas the
very simple, whereas
ot known at present.
.there is no additional
lltivariable dynarnical
), different dynarnical
d in Section 7-3. For
Figure 7-1
A state feedback system.
Proof
First we show that the controllability of FE implies the controllability of
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)
We see that in the proof, the assumptions of single-variableness and time

invariance are not used. Therefore we have the following corollary.
is the scalar output,

lstant column vector,
k, every state variable
~minal. Let the gain
k~ [k kz ... knJ.
shown in Figure 7-1 is
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)
Note that Theorem 7-3 can also be proved by showing that
lere r is the reference

(7-25) have the sarne
.hat the controllability
ant under any linear
p[b
Ab ...
An-lb] =p[b
(A+ bk)b
...
(A+ bk)n-b] (7-26)
for any 1 x n real constant vector k (see Problern 7-10)..

Although state feedback preserves the controllability of a dynamical
eq uation, it is always possible to destroy the observabiiity property of a
dynarnical equation by sorne choice of k. For example, if e =/=-0 and if k =
( - l/e)c, then the state-feedback equation (7-25) is notobservable even if FE
is observable. If e = 0, it is still possible to choose sorne k such that the state
feedback dynarnical equation will not preserve the observability property (see
Problern 7-14).
; controllable for any

FE in (7-24) is con-
336
STATE FEEDBACK ANO STATE ESTIMATORS
Example 1
nomial of the matrix (

eigenvalues be
Consider the controllable and observable dynamical equation
x=G
n +[~Ju
y=[1
2Jx
If k is chosen as
If we introduce the state feedback
then the sta te-feedback '
u=r +[ -3
-IJx
x=[~ ~Jx +[~}-
FE{:
CFE{:
then the state-feedback equation is
x=
O
-e
2Jx
y=[1
O
O
which is controllable but not observable.
An important property of state feedback is that it can be used to control the

eigenvalues of a dynamical equation.
Theorem 7-4
Since the characteristic

we conclude th
values.
... +am
If the single-variable dynamical equation FE 1 given in (7-24) is controllable,

then by the state feedback u =:= r +kx, where k is a 1 x n real vector, the eigen
values of (A + bk) can be arbitrarily assigned, provided that complex conjugate
eigenvalues appear in pair.
Proof
Ifthe dynamical equation FE 1 is controllable, by an equivalence transformation

FE I can be transformed into the following controllable canonical
form (Theorem 7-1):
x = Px,
O
O
O
-a"_1
O
-a,,-z
O
-az
i~
l!
y=
[fi" fi,,- 1 fi,,- 2
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
.. u = r +kx = r.+kp-Ix ~ r +kx
(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
The gain vector k ir

Therefore,'
k = kP, where the colurr
The matrices A, b, and k
is assigned to the matrix
The procedure of choosi
therefore it can be appl
linear time-invariant dy
choosing Rr re the fo!1ovvi
u = r +kx.
Algorithm
Given a controllable {A
1 x n real vector k such
its eigenvalues.
1.
2.
3.
4.
5.
6.
7.
Find the characteristic

Compute (s -I)(s -~ .
Compute k = [a" ~. a"
Compute q"-i =Aq,,_
FormQ=[ql qz '
Find P ~ Q - l.
k =kP.
--==c::===___
-~~~,,===~===-
~-_-_-~~.:__-=_-=-=---=-~----------------------------~--------
------.--_.-
--
---------------_ ..-.
..
----,_.~--------_._
337
STATE FEEDBACK
nomial of the matrix (A + bk) or, correspondingly, of (A + bk) with desired

eigenvalues be
ation
If k is chosen as
(7-29)
then the state-feedback dynamical equation becomes

CFE{:
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)
y=[I1" +e(cx"-X") 11"- +e(cx"_-X"_) ... 11 +e(cx-iX)]x +er

be used to control the
(7-24) is controllable,
real vector, the eigen
mt complex conjugate
(7-30b)
Since the characteristic polynomial of the A matrix in (7-30) is s" +iXs"- +

... +ii.., we conclude that the state-feedback equation has the desired eigen
values.
Q.EoD.
The gain vector k in (7-29) is chosen with respect to the state x; that is,
Therefore, with respect to the original state x, we have to use
k=kP, where the columns ofP- =Q are those vectors introduced in (7-9).
The matrices A, b, and k are assumed to be real; hence, if a complex eigenvalue
is assigned to the matrix (A + bk), its complex conjugate must also be assignedo
The procedure of choosing k has nothing to do with the number of outputs;
therefore it can be applied to any controllable single-input, multiple-output,
linear time-invariant dynamical equationo We summarize the procedure of
choosing k in the followingo
u =1' +kx.
alence transformation
ontrollable canonical
o
o
o
-IX
eu
rO
'+l~
(7-27 a)
(7-27b)
p-, b = Pbo Because

;omes
(7-28)
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
This algorithm can also be developed directly without transforming the

equation explicitly into the controllable canonical formo Let ~(s) be the
characteristic polynomial of A + bk. Then we have
~(s) = det (sI - A - bk) = det [(sI - A)(I - (sI - A)-l bk)J
Stabi Iization
= det (sI - A) det [1 -(sI - A)-lbkJ
This can be written as, by using (3-65) and defining
~(s) =
det (sI - A),
~(s) = ~(s)(1 - k(sl - A)-l b)

~(s) - ~(s) = ~(s)k(sl - A)-l b
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
Consider the inverted pendulum problem studied in Example 4 on page 185.

Its dynamical equation is, as derived in (5-25),
X~l~
y = [1
O -1
It is controllable; hence its eigenvalues can be arbitrarily assigned. Let the

desired eigenvalues be - 1, - 2, - 1 j. Theri we have
~(s)
= (s + 1)(s + 2)(s + 1 + j)(s + 1 - j) = S4 + 5s 3 + lOs 2 + lOs +4
The characteristic polynomial of
A+bkJ ~
l-2k
where k = [k l
k2
k3
det (sI - A - bk) =
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
Hence the introduction of the state feedback

u(t)=r(t)+[i
lf
liJx
will place the eigenvalues of the resulting equation at - 1, - 2, -1 j.
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
Consider the dyna

c(sl - A) - 1 b +e. Si
7-4 we conclude tha
duction ofstate feec
g(s) are shifted by st:
. _---_.
-_._._"~-~-----~,~._---.=.",.-===-====--=-::==-=----~'-':'~=-:::"-:::=:':'''....--::::'''~==-.==::::--==-----~
-':::'--:::.='::::-- -==--=--=-.-.:.......===...-=--===-:::..:=....::.::=.::~=='::::.====--------_
STATE FEEDBACK
out transforming the

m. Let li(s) be the
339
case, a different method of computing k by solving a Lyapunov equation

(Appendix F) wilI be introduced. The method is directly applicable to the
single-variable case. To avoid repetition, it will not be discussed here.
Stabilization
1= det (sI - A),

(7-31 )
(7-13) and (7-14), we
Ir a dynamical equation is controlIable, then alI the eigenvalues can be arbitrarily

assigned by the introduction of state feedback. If a dynamical equation is not
controlIable, one may wonder how many eigenvalues can be controlIed. It is
shown in Theorem 5-16 that if a linear time-invariant dynamical equation is not
controllable, by a proper choice of basis vectors, the state equation can be
transformed into
x=Ax +bu
lmple 4 on page 185.
ily assigned.
Let the
10s 2 +10s +4
'S2
+ 3k 2 s + 3k 1
id those of li(s) yields
:z
where
-_[AO A
A-
11
12 ]
(7-32)
(7-33)
22
and the reduced equation Xl = A11 Xl + b1 u is controllable. Because of the

form of A, the set of eigenvalues of A is the union of the sets of the eigenvalues
of A11 and of A22 . In view of the form of b, it is easy to see that the matrix
A22 is not affected by the introduction of any state feedback of the form u =
r +kx. Therefore alI the eigenvalues of A22 cannot be controlIed. On the
other hand, aH the eigenvalues of A11 can be arbitrarily assigned because
{A 11 , bd is controllable. Hence we conclude that the eigenvalues of(A +bk)
can be arbitrarily assigned if and only if {A, b} is controlIable. Since k is
required to be a real vector, complex conjugate eigenvalues must be assigned
in pairs.
In the design of a system, sometimes it is required only to change unstable
eigenvalues (the eigenvalues with nonnegative real parts) to stable eigenvalues
(the eigenvalues with negative real parts). This is calIed stabilizaton. From
the aboye discussion we see that ifthe matrix A 22 has unstable eigenvalues, then
the equation cannot be stabilized. Hence we may conclude that if {A, b} is
transformed into the form in (7-33), and if {Al 1> bd is controlIable and alI the
eigenvalues of .4. 22 have negative real parts, then {A, b} is stabilizable. Whether
a dynamical equation is stabilizable can also be seen from its equivalent Jordan
form dynamical equation. If alI the Jordan blocks associated with unstable
eigenvalues are controllable, then the equation is stabilizable. From the proof
of Statement 4 of Theorem 5-7 on page 206, we can conclude that }, is in A22
if and only if the matrix [..1.1 - A
B] does not have a full row rank. We note
that Al! and A22 in (7-33) may have same eigenvalues.
Effec(on th numerator of g(s)
1, -2, -1 j.
:t (si - A - bk) is very

.on. Hence, for large
In the multivariable
Considerthe dynamical equation (7-24). Its transfer function is g(s)=

c(sl - Af 1 b +e. Since every pole. of g(s) is an eigenvalue of A, from Theorem
7-4 we conclude that the poles of g(s) can be arbitrarily assigned by the intro
duction of state feedback. It is of interest to note that although the poles of
g(s) are shifted by state feedback, the zeros of g(s) are not affected. That is, the
340
zeros of g(S) remain unchanged after the introduction of state feedback. We

prove this by showing that the numerator of the transfer function of FE 1 in
(7-24) is equal to the numerator of the transfer function of eFE 1 in (7-30).
The transfer function of FE 1 is
A()
gs
131S,,-1 +132S,,-2 + ... +f3" +

e
s" +O:I S" 1 + ... +0:"
es" +(131 +eO:l)s,,-1 +
s" +O:IS,,-1 +
+(13" +eo:,,)
+0:"
(7-34 )
The transfer function of (7-30) is

g(s)
[131 +e(O:I -&I)]S,,-1 + ... +[13" +e(o:" - IX,)]

+e
s" +O:I S" -1 + ... +0:"
es" +(131 +e0: 1)s,,-1 +
s" +O:I S" 1 +
+(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)]
Ir r(t) = rd, for t ~ O, then
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
and the design is COI

{A, b} is not con trol
have complete contr,
Much more can b
It appears that the d
function and will be
discussed further in t
*Multivariable ca:
multivariable dynam
where A, B, e, and l
real matrices. In sta
where r stands for a r

called the feedback g
(7-37)
The substitution of the identity
(si - A - bk)-I S-1 =( -A - bk)-lS-1 +(sl -A - bk)-I(A + bk)-1
In the following,
trollable, then the ei
proper choice of K.
[see also Equation (H-I)] and

g(s)=(e +ek)(sl-A-bk):-lb +e
into (7-37) yields
;(s) = (e +ek)(sl - A - bk)-1 [x(O)1 +(A + bk)-1 bprd] +g()prds-1
(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)
ves that the zeros of

Jack.
dback may alter the
poles are shifted to
5) is less than n, and
his case, (7-25) must
:m so that its output

d = 0, the problem is
e asymptotic tracking
n of a potentiometer
n.put u is chosen as
(7-36)
)1".
The substitution
ace transform to the
ores)]
+ epr(s)
341
which becomes, in the time domain,

y(t)=(c +ek)eIA+bk)1 [x(O)I +(A +bk)-lbprd] +g(O)prd
(7-39)
for t 2: O. If aH eigenvalues of (A + bk) have negative real parts, then e lA +1>1<),

approaches zero as t->oo and consequently y(t) approaches g(O)prd as t->C.
If {A, b} is controHable, aH eigenvalues of (A +bk) can be arbitrarily assigned
by a proper choice of k. Clearly, if the eigenvalues of (A + bk) are shifted
deeper inside the left-half s plane, the faster the output approaches g(O)prd'
However, in this case the required feedback gain wil1 be larger and may cause
saturation. Furthermore, the system may be more susceptible to noises. Thus
the choice of proper eigenvalues is not a simple task and is generally reached
by a compromise between conflicting criteria.
As discussed in the previous subsection, the zeros of g(s) are not affected
by state feedback; hence we have g(O) f.O if and only if g(O) f.O. If g(O) f. O,
we may choose p = l/gJ(O). In this case, we have
as
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 A, B, C, and E are, respectively, n x n, n x p, q x n, and q x p constant

real matrices. In state feedback, the input u in FE is replaced by
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
x =(A +BK)x +Br

y=(C+EK)x+ Er
(7-41 a )
(7-41b)
In the following, we shall show that if the dynamical equation FE is con

trollable, then the eigenvalues of (A + BK) can be arbitrarily assigned by a
proper choice of K. This wil1 be established by using three different methods.
(7-38) .
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
_._-------_.-,_._----------~------------_._----_
..
_---~--_._-_.
__.--_._---_.
---~-_._-----_
.. _------_.
__..--_._ .._--_._._._._-_._. __ _-_._._-_.

..
A matrix A is called cyclic if its characteristic polynomial is equal to its

minimal polynomial. From Theorem 2-12 or Example 1 on page 48, we
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
Let the characteristic
where a, i= 1, 2, ... ,
tion of X(s) with respt
L1'1
If L1(s) has repeated re
sufficient condition fe
Proof
Controllability is invariant under any equivalence transformation, hence we

may assume A to be in the Jordan canonical formo To see the basic idea, we
use the following example:
det
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.
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
IS controllable. However, there is no v such that {A, Bv} is controllable.

lf allthe eigenvalues of A are distinct, then there is onlyone Jordan block
associated with each eigenvalue. Hence a sufficient condition for A to be cyclic
is that all the eigenvalues of A are distinct
an
O
an-
O
2a n_
--------.
an - I
O
O
a'l-
(See Appendix G). V

nomial of kij. There i
of {ku} as a vector in 1
clear that the solutio
almost all K, we hay.
establishes the theoreI
V,;/e are
D.OV/
ready
Theorem 7-7
lfthe dynamical equat

of the form u = r + Kx.
of (A + BK) can be arbi
appear in pairs.
This theorem can be extel

real constant matrix H, th
cyclic. The matrix A + BI
Theorem 7-6 is equally apl
--_._--~-----
--_ .. _--_._--_._._.-----------
------------- --_._----------------- -
~-
---~-- -~--
-=.~~::=.:----=-~~----~--~--------_._-------_._~
STATE FEEDBACK
lmial is equal to its

: 1 on page 48, we
le Jordan canonical
i with each distinct
1 the property that if
)rs b, Ab, ... ,An - 1 b
s controllable. This
--.-_.--
--~-----
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
Let the characteristic polynomial of A
+ BK be
where a, i = 1, 2, ... , n, are functions of aH elements, kij, of K. The differentia

tion of Li(s) with respect to s yields
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
(See Appendix G). We note that y(kij) is generally a nonhomogeneous poly

nomial of kij. There is a total of p x n number of kij in K. Hence we may think
of {kij} as a vector in the (p x n)-dimensional real vector space (IR P )(", IR). It is
clear that the solution, kij, of y(kij) = O is a subset of (lIF )(n, IR). Hence for
almost aH K, we have y(k) +0 and all the roots of Li(s) are distinct. This
establishes the theorem. See Problem 7-8.
Q.E. D.
We are now ready to establish the following main theorem.
Theorem 7-7
Ir the dynamical equation FE in (7-40) is controllable, by a linear state feedback

of the form u = r + Kx, where K is a p x n real constant matrix, the eigenvalues
of (A + BK) can be arbitrarily assigned provided complex conjugate eigenvalues
appear in pairs.
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
.. __..
~------_
..-----_..
__.__.._._--_._--_._------_._-_..
~--"_._
_---~
_.~-
.. _.. _ ..
"-~
..
_~
_~---_._.,._----_.-
__ _.__._-_._._._.,_ _---_._._.__
..
._--_._-----~-_
.. _.. _-_._-_._.- .. - ..'.

,
__ .',_.'---_ __
If {A, B} is not cont
Proof
If A is not cyclic, we introduce u = w + Kx such that ~ A + BK! in
i =(A +BK)x +Bw
(7-43)
is cyclic. Since {A, B} is controllable, so is {A, B} (Corollary 7-3). Hence

there exists a p x 1 real vector v such that {A, Bv} is controllable. 5 Now we
introduce another state feedback w = r +Kzx, as shown in Figure 7-2, with K z
chosen of the form
and the eigenvalues of ~

conclude that the eiger
only if {A, B} is control:
Method 11
In this method, we shall
in (5-93) and (6-112) al
similar to the single-inpl
we assume n =9 and p =
are Jl = 3, Jlz = 2, and Jl
where k is a 1 x n real vector, then (7-43) becomes

i=(A +BKz)x +Br =(A +Bvk)x +Br
Since {A, Bv} is controllable, the eigenvalues of (A + Bvk) can be arbitrarily

assigned by a proper choice of k (Theorem 7-4). By combining the state feed
back u = w + Kx and the state feedback w = r + Kzx as
M =[b
Ah!
(see Section 5-3) is nonsi
u =r +(K +Kz)x ~ r +Kx
Q.E.D.
the theorem is proved.
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
State feedback of a multivariable dynamical equation.
6Tills mat~ix can be obtained e

computing explicitly M - l. ~
.-
----.:.-==--=-:::.::----===-====- -=-=-=---====-.-====-=..=.===-..:==:=::--==-------
---_. __
--------- -- - _._--------_.__.__. -._-_.
.
- - - _ - - - - _ _ _. - - - _ _ . _ - - - - _ . _ - - - - -
-~
----_._----_._-----_.
..
.-----_ ... -._--
345
STATE FEEDBACK
Ir {A, B} is not controllable, they can be transformed into
(7-43)
Corollary 7-3). Hence

:ontrollable. 5 Now we
i1 in Figure 7-2, with K z
[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
(see Section 5-3) is nonsingular. We compute M- and name its rows as

ell
e lZ
e l3
eZ l
e zz
e 3l
e 3Z
e3 3
e34
Q.E.D.
it;arily, and the probability is

Reference S38, a procedure is
tion required however is very
~ell'l
~eZ1'2
~e31'J
From M-lM=I, we can obtain sorne relationships such as e 13 Ak b l =O,

k = 0,1; e 13A z b l = 1; e 13 A k b Z= O, k= O, 1;e 13 A k b 3 = O, k= O, 1,2,3. Note that
we do not have e 13 A z b z = Obecause AZb z is a linear combination of {bj , Abj,j=
1,2,3 and AZb l } and e 13 A2b l = 1. Now we use eil'i' i = 1,2,3, to form the square
matrix 6
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
Using the aforementioned relationship eijAkb =0 or 1, it is straightforward,

though very tedious, to verify that P is nonsingular and 7
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
possible to obtain sorne

form on the diagonal o
c~mp~~ion form on tl
(A +BK) in (7-47) are
( s3
and
Since a, b, C, and d. can

Theorem 7-7.
1
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
x=(A +BK)x +Br

Because of the form of B, all rows of A except the three rows denoted by strings
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 :
.
.
--------:-0- -i -:-: - - ------

:b
bz :
l
-------~-----~----------
:
:
:,
:O 1 O O
:O O 1 O
:O O O 1
or
.... -
_1- _
.. _ _ .. 1_ ..
........ __
-6--6 -0- ~ -o -0-: -- -( - --O
We justify first the algori

AT-TF= -Bi( implies
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
;5;JIz ;5;JI3, they become
001
.
In our example, JIl = 3, JIz = 2, and JI3 = 4, and the three nonzero royvs ofO are
o~~J
1. Choose an arbitrary n >

with those of A.
2. Choose an arbitrary p x
3. Solve the unique T in thc
4. If T is nonsingular the
~igenvalues as those'of F.
K and repea t the process.
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
A simple pattern can be deduced from these examples.
Notethat we have d(T)"'AT -TI

d(T) are ,litA) - JIAF).
STATE FEEDBACK
1, 1't
d7
l'S
straightforward,
O O
O O O
1 x O
S=PB= O 1
(S3 - a3 s2 - a2 s - al)(s2 - b 2 s - b)(S4 - C4S3 - C3S2 - C2S -
O O O
O O O
O O 1
(7-46)
{A, B} is_said to be_i~ a

If u = r + Kx, where K lS a
:e rows denoted by strings

se the three nonzero rows
enoted by string~ of ! _can
ose K so that (A +BK) is
0 0 : 0 0 00
00:000 0
, 1 0 : 0 0 00
:- 0- -i -:- 0- -- - - 0
:00:100 O
:--o-~---(-o--
:00:0010
: O O: O O O 1
: d 4 d s : d 6 d7 ds d9
(7-47)
anion form on the diagonal;

lon formo Of course, it is
xl
.~ J
ir Il ::;; 112 ::f 1l3'
they become
three nonzero roWS ofB are
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
Since a, b, C, and di can be arbitrarily assigned, we have established once again

Theorem 7-7.
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
that we have d(T) =AT - TF rather than d(T) = AT +TF.

d(T) are A;(A) - Il}F).
Hence the eigenvalues of
348
Nonuniqueness of feedback gain matrix K

From the design procedures discussed in the previous subsections, we see
that many different feedback gain matrices can yield the same eigenvalues of
(A + BK). Hence it is natural to ask which K should we choose? Before dis
cussing this problem, we give an example.
Exan'lple 3
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
The equation is in a multivariable controllable formo The problem is to find a

are -1, -2 j, and -1 2j.
We compute
Ll(5) =5 5 +75 4 +245 3 +485 2 +555 +25
K so that the eigenvalues of the resulting matrix
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
where /Xlmax is its largest t

gains are also proportion
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
)us subsections, we see

the same eigenvalues of
we choose? Before dis
"j~u
1
(7-48)
The problem is to find ~
-1, - 2j, and -1 2}.
O
O
.-1]
-5
O
24
elements of 1(2' which preserves the original blocks of companion formo If

most of the assigned eigenvalues have magnitudes larger than one, then it is
clear that the larger the order of a companion-form block, the larger the magni
tudes of the feedback gains.
Next we compare the responses of the two systems. Although 1(1 and 1(2
yield the same eigenvalues, the transfer matrix C(sI - A- 81(1)-18 and the
transfer matrix C(sI - A- 81(2) - 1 B are generally different. Thus, their zero
input and zero-state responses are generally different. For convenience, we
compare only their zero-input responses. We plot in Figure 7-3(a) and
(b) the responses of =(A+8K 1 )x and =(A+8K 2 )x due to x'(O)=
[2
1
O -1 -2]. We see that the largest magnitude in transient in
1(1 is roughly three times larger than the one in 1(2' Hence 1(2 has smalier
I
gains and yields smaller magnitudes in transient and is preferred.
F or a companion-form matrix of arder m, if all its eigenval ues are distinct, it
is shown in Reference 517 that the largest magnitude in transient is roughly
proportional to
(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- _ _ _ _ _ _ _ _ _ _ _ _ _ _
The K can be uniqueJy

This is outside the scop
and S4, or any other bo,
Assignment of eigenva
We see that the feedback gains are very large.

_
Now we use the procedure discussed in Problem 7-20 to find a K so that
the resulting +:DK is of the form
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
which implies that
BK[e
STATE FEEDBACK
n-form block should be

nined by the lengths 01'
110sen by using Scheme 2
Ilocks are equal to the
est order is equal to the
i in (7-20) or any other
uent1y, the order 01' the
:st equal to, fl. Hence in
t we shall use Scheme 2
J .,. An-1B] and the
the con
ll, fl2' . . . , fl P'
e order 01' a companion
ease it. Hence in finding
companion-form blo_cks.
K. We may choose K so
In block triangular form!.
les will yield different K.
lal and group the desired
p\max)"p-l}
e in the th companion
K is considerably reduced.
he orders 01' companion
~t the desired eigenvalues
. not possib1e to choose a
'onal or a block triangular
~s into one, then the result-
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
The K can be uniquely determined by solving an algebraic Riccati equation.

This is outside the scope 01' this text and the reader is referred to References 3
and S4, or any other book on optimal control.
Assignment of eigenvalues and eigenvectors
m 7-20 to find a K so that

1
351
(7~50)
In this subsection, we discuss the assignment 01' eigenvalues as well as the

eigenvectors 01' the resulting A +BK. For convenience, we assume that the
assigned eigenvalues, Xi' are all distinct. Let e i be an eigenvector 01' (A + BK)
associated with Xi' that is
or
which implies that
(A +BK)ei= Xie i
i=1,2, ... ,n
BKe = (XiI - A)e ~ f i
,_._-_.- ....
_-_
....
_~~--~."
.__
._._._._~_._---~-~_._,-_
352
..
_~---_
..
_-
__
. _ __ _
, ... _ . ~ ,
~ , ~ _ .__ ~ .
__
~.
_ _ , u . . . ~ ~
~_g
' _ ' _ _
~~~~_'
'
~'_'._'~_.~'
_ _ ~ ' _ _
'_~AA_".
----- - . - . - - - . - - . - ..... - - - - - .._ ....- _ . __ -
._~"-
._~.~,--~--
Now if {e i , i
1,2, ... , n} are linearly independent, then we have

BK = [f 1
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
The effect '
which implies
u(s) = (1- ~
The substitution of (7-5:
y(s) = (N(,
Hence the transfer matr
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
D(s) = D'.eH(s) + DlcL(s)

N(s) =CL(s)
where H(s) and L(s) are defined as in (6-101). We call N(s) +ED(s) the numer
ator matrix and D(s) denominator matrix of G(s). If we plot G(s) as in Figure
7-4 and call the output of D -les) the pseudostate a(s), then the state x(s) is, as
established in (6-108), equal to x(s) = L(s );(s). Hence the state feedback o(s) =
r(s) + Kx(s) can be achieved by applying Kas shown in Figure 7-4. From
Figure 7-4 we have
u(s) = r(s) + KL(s)~(s) = r(s) + KL(s)D-l(S)U(S)
In this subsection, we dis<

III for state feedback. 1
{A, B} into a multivariabl,
to achieve this are given i
AkB for k = 1,2, ... , n - )
ditioned problem, as discu
not satisfactory from a co
This observation was providec
STATE FEEDBACK
we have
>
-n
]-1
353
(7-51 )
of (7-51) are inside the

mce the assignment of
ignment of eigenvalues.
1at form an eigenvector
forms shown in (7-47),
O O O O O O]',
1 X" F X3]' for the
~ Xr' Ip Xi~ IP]' for
Hence for the {A, B}
nvectors are ehosen as
~ solution is unique ane
7-46).
n arbitrary assignment.
. three companion-form
)ck, the eigenvectors are
lVectors are ehosen as
le assignment of eigen
iscussed in the previous
i eigenvectors, the reader
.tor of g(s) is not affeeted

be arbitrarily assigned.
riable case. Before pro
ion 6-6, especially Equa
(7-46), by reversing the
(s) +ED(s))D- 1 (s)
+ y(s)
+
Figure 7-4
The effect of state feedback.
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
y(s) = (N(s) + ED(s))D-l(S)D(s)(D(s) - KL(s))-l r (S)

Hence the transfer matrix from r(s) to y(s) is
G(s) = (N(s) + ED(s))(D(s) - KL(s))-1
= (N(s) + ED(s))[DheH(s) + (DIe - K)L{s)] -1
(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
N(s) +ED(s) the numer

we plot G(s) as in Figure
), then the state X(S)Ais, as
the state feedback u(s) =
vn in Figure 14 From
'-I(S)O(S)
In this subsection, we diseuss sorne computational problems in Methods Il and

III for state feedback. The major step in Method Il is the transformation of
{A, B} into a multivariable controllable formo The equivalence transformations
to achieve this are given in (7-14) and (7-45). They require the computation of
AkB for k = 1, 2, ... , n - 1 and may change the problem into a less well-con
ditioned problem, as discussed in Section 5-8. Hence this procedure is generally
not satisfactory frorri a computer computational point of view,
This observation was provided by Professor D. Z. Zheng ofTsinghua University, Beijing, China
- . - ._-
----.. ". -_. ,.".-
._,u ._._.__
".~
..
_.~L_
~._.,,
..
"_~._.
_._
.__ .__.._..
~_
~_
..__ .__
._._~
.~
.. __ .
.__ _.__
~~
__ .__._ . _._._. _ .._. __
~~_
__.._.___ .__ .._ . _ M
.,,"
._ _
__._.__ ..
--~----~-------------~~~~~
354
In Section 5-8, we discussed a different method of transforming an equation

into a controllable form without explicitIy computing AkB. The method is first
to transform A into a Hessenberg form as in (5-89), and then transform it into
a controllable form as in (5-93). (Note the different ordering of the state vari
ables.) The first step is carried out by a sequence of orthogonal transforma
tions or gaussian eliminations with partial pivoting and is numerical1y stable.
The second step must be carried out by gaussian eliminations without any
pivoting and is numerically unstable. Hence this transformation method is
again not satisfactory.
In Section 6-6, we introduced an irreducible realization of G(s) = N(S)O-l(S)
in a controllable formo Hence, if N(S)O-l(S) can be computed from {A, B, C},
then a controlIable form dynamical equation can be obtained. In Reference
S17l, an algorithm is presented to compute a coprime N{S)O-l{S) from an
Hessenberg-form dynamical equation. However in the design, we still need
the equivalence transformation. Thus it is not clear whether this design pro
cedure will be satisfactory. In References S156 and S157 a dynamical equation
is first transformed into the Hessenberg form, the pole assignment is then
carried out by using the QR algorithm. The methods used in References S156
and S157 are alI numericalIy stable. Once a dynamical equation is transformed
into a Hessenberg form, the Hyman method in Reference S2l2 can also be
employed to carry out the pole-assignment. For the 20 x 20 matrixon page 219
with e =0, the Hyman method yields a comparable result as the QR method.
See Reference S202.
The procedure in Method 111 does nqt require the transformation of an
equation into a controlIable formo It requires merely the solution of a Lyapunov
equation. The solution of Lyapunov equations has been extensively studied
in the literature, see, for example, References S9, S18, and Sl07. The algorithm
in Reference Sl07 is claimed to be efficient and numericalIy stable and may be
employed in our designo
A limited comparison between Method n and Method nI in terms of
operational efficiency and numerical stability was carried out in Reference
S202. Because of the complexities of the problem, no clear-cut concl usion was
reached in Reference S202.
state observer. In thi

full-dimensional and
dual to the design 01
single-variable case \\
special case of the mu
In this section we
ofthe variable. For e
Full-dimensional ~
time-invariant dynam
where A, R, and e are,

For simplicity, the dir
assume now that the s
state variables are nol
completely known. J
x(t) from the availablt
matrices A, B, and C.
original system as shc
estimator. Now if the
initial state and are dr
will be equal to the act
is how to find the initi:
that state. This proble
dynamical equation Fj
from its input and ou
an open-loop estimatOl
There are, howevel
First, the initial state m
lJ
7-4
State Estimators
In the previous section we introduced state feedback under the assumption

that all the state variables are available as outputs. This assumption often does
not hold in practice, however, either because the state variables are not acces
sible ror direct measurement or because the number of measuring devices is
limited. Thus, in order to apply state feedback to stabilize, to optimize, or to
decouple (see Section 7-6) a system, a reasonablesubstitute for the state vector
often has to be found. In this sedion we shall show how the available inputs
and outputs of a dynamical equation can be used to drive a device so that the
outputs of the device will a)proximate the state vector. The device that con
strucs an approximation of the state vector is calIed a state estimato,. or a
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
Full-dimensional state estimator. Consider the n-dimensional linear

time-invariant dynamical equation
:ion of G(s) = N(s)D-1(s)

)mputed from {A, B, C},
obtained. In Reference
.me N(s)D - 1 (s) from an
the design, we still need
whether this design pro
57 a dynamical equation
pole assignment is then
; used in References Sl56
1equation is transformed
erence S2t2 can also be
Ox 20 matrixon page 219
~esult as the QR method.
FE:
X=Ax +Bu
y=Cx
(7-54a)
(7-54b)
where A, B, and C are, respectively, n x n, n x p, and q x n real constant matrices.

For simplicity, the direct transmission part has been assumed to be zero. We
assume now that the state variables are not accessible. Note that although the
state variables are not accessible, the matrices A, B, and C are assumed to be
completely known. Hence the problem is that of estimating or generating
x(t) from the available input u and the output y with the knowledge of the
matrices A, B, and C. If we know the matrices A and B, we can duplicate the
original system as shown in Figure 7-5. We called the system an open-Ioop
estimator. Now if the original equation FE and the estimator have the same
initial state and are driven by the same input, the output x(t) of the estimator
will be eq.ual to the actual state x(t) for al1 t. Therefore, the remaining question
is how to find the initial state of FE and set the initial state of the estimator to
that state. This problem was solved in Section 5-4. It is shown there that if the
dynamical equation FE is observable, the initial state of FE can be computed
from its input and output. Consequently, if the equation FE is observable,
an open-Ioop estimator can be used to generate the state vector.
There are, however, two disadvantages in using an open-Ioop estimator.
First, the initial state must be computed and set each time we use the estimator.
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
lck under the assumption

~his assum ption often does
.te variables are not acces
~r of measuring devices is
.tabilize, to optimize, or to
)stitute for the state vector
!/ .how the available inputs
) drive a device so that the
:tor. The devicethatcon
lled a state estimator or a
L
Figure 7-5
~~~~
An open-loop state estimator.
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
=Ax +Bu +L(y-Cx)
(7-55 )
which can be written
The asymptotic e;
either by a block diag
Clearly x is the error

tracting (7-56) from (7
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
An asymptotic state estimator.
I
I
Figure 7-7
An asympt
STATE ESTlMATORS
y, if the matrix A has

nal1 difference between
sturbance or incorrect
he actual x(t) and the
~n-Ioop estimator is, in
Iilal state vector is to
: dynamical equation is
he state vector can be
.tiators are not easy to
~ely distorted by noises
d the output of FE are
.ator. It is conceivable
rmance of an estimator
The estimator is driven
:m. The output of FE,
e is used to serve as a
: multiplied by an n x q
:gators of the estimator.
'mator, for a reason to
imator shown in Figure
(7-55)
357
i. =(A -Le)x +Ly +Bu
(7-56 )
The asymptotic estimator in Figure 7-6 can be redrawn as in Figure 7-7

either by a block diagram manipulation or from (7-56). Define
x~
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
An asymptotic state estimator.
==_._._
~ ~
--===-_ _.__ _._._._.__ __"

~.
358
__..
_~.~.
..
..-.=--::::="-::::""=_.::::::::.:: .._:::,.::::
:_ __ ,:,::
__",._" ._"__..
~~"_~
-.~._ =-~ =--~"-="::::
.~:::,
__". __..__
._._~
..
.. _._
.~.
~~
__
~~
__
__
._.~
_.~_~.
..,._.
.~_u
_ ~
__ . . .._.
_~
_..._ ...._.__ .._. __ u, " "
._,_.u~_
-;
Theorem 7-8
Proof
If the n-dimensional dynamical equation FE in (7-54) is observable, its state

can be estimated by using the n-dimensional state estimator
Define
Then we have
~ = (A LC)x + Ly + Bu
with the error
x =x xgoverned by
and all the eigenvalues of A - Le can be arbitrarily assigned, provided complex

conjugate eigenvalues appear in pairs.
To show the theorem, it is sufficient to show that if {A, C} is observable,

then all the eigenvalues of A - Le can be arbitrarily assigned. lndeed, if
{A, C} is observable, then {A', C}, where the prime denotes the transpose, is
controllable (Theorem 5-10). Hence for any set of n eigenvalues, we can find
a real K so that A' +CK has the set as its eigenvalues (Theorem 7-7). Since
the eigenvalues of A + K'C and those of A' +C'K are the same, the theorem is
established by choosing L = - K'.
It is clear that all the design procedures for state feedback can be used to
design state estimators. Hence their discussions will not be repeated. We
mention only that, in order to have small gains in L and smal1 transient in
i = (A Le)x, the largest order of companion-forrn blocks in the transformed
A matrix' should be as smal1 as possible. Dual to the state-feedback part, the
smallest such order is the observability index of {A, C}.
Method 11
Consider the n-dimensional dynamical equation
FE:
X=Ax +Bu
y=Cx
(7-58a)
(7-58b)
where A, B, and e are, respectively, n x n, n x p, and q x n real constant matrices.

It is assumed that FE is irreducible. Define the n-dimensional dynamical
equation
If the three conditions

and u(t). Hence z(t) is
Now we show the
x(O) = O and u(t) == O, \
find a u(t) to make e(t
trollable, we can find ;
This establishes the ne(
With this theorem,

Algorithm
1. Choose an F so tha
disjoint from those (

2. Choose a G so that
3. Solve the unique Ti
4. lf T is nnsingular,
F, G, and H is an est
different F and/or (
We give sorne remar J
eigenvalue, then a solu
(Appendix F). lf A and
may not exist in TA
and A to have no comn
The controllability (
as will be established in
Theorem 7 -1 O
z=Fz +Gy +Hu
(7-59)
where F, G, and H are, respectively, n x n, n x q, and n x p real constant matrices.
lf A and F have no com

of a nonsingular T in '
trollable. For the singlt
Theorem 7-9
Proof
Thestate z(t) in (7-59) is anestimateofTx(t) for sornen x n. real constant matrix T

in the sense that z(t) Tx(t)-+ O as t -+ 00, for any x(O), z(O), nd u(t) if and only if
Let the characteristic p.

(s) = det
1. TA-Ff=GC
2. H=TB
3. All the eigenvalues of F have negative real parts.
(7-60)
Clearly we have (A) =

F, then (A.) is an eiger
STATE ESTIMATORS
359
Proof
is observable, its state
ltor
Define
e ~ z-Tx
Then we have
=:i - Ti = Fz + Gy + Hu - TAx - TBu
= Fe +(FT - TA + GC)x +(H - TB)u
ned, provided complex
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)
real constant matrices.

limensional dynamical
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.
With this theorem, we can now propose a design algorithm.

Algorithm
1. Choose an F so that all of its eigenvalues have negative real parts and are
disjoint from those of A.
2. Choose a G so that {F, G} is controllable.
3. Solve the unique T in TA - FT = Gc.
4. Ir T is nonsingular, computeH =TB. The equation in (7-59) with these
F, G, and H is an estimate ofTx(t) or x(t)=T- 1 z(t). Ir T is singular, choose
different F and/or G and repeat the process.
We give sorne remarks regarding this algorithm. Ir A and F have no common

eigenvalue, then a solution T always exists in TA - FT = GC and is unique
(Appendix F). Ir A and F have common eigenvalues, then a solution T may or
may not exist in TA - FT = GC. To remove this uncertainty, we require F
and A to have no common eigenvalues.
The controllability of {F" G} is needed for the existence of a nonsingular T
as will be established in the following theorem.
Theorem 7 -1 O
fl-59)
1 real
constant matrices.
Ir A and F have no common eigenvalues, necessary conditions for the existence

of a nonsingular T in TA-FT=GC are {A,C} observable and {F, G} con
trollable. For the single-output case (q = 1), the conditions are sufficient as well.
Proof
n real constant matrix T

O), and u(t) if and only if
fl-60)
Let the characteristic polynomial of A be

Ll(s)=det(sI-A)=sn +asn-'+azsn- z + ..'. +an
Clearly we have Ll(A) = (Cayley-Hamilton theorem). Ir A is aneigenvalue of

F, then Ll(AJ is an eigenvalue of Ll(F) (Problem 2-32). Since A and F have no
360
dynamical equation
no means simple nUD
It is not clear at presen
efficiency and numeril
common eigenvalue, we have ,(Ai) +0 for al1 eigenvalues Ai of F and

n
det ,(F) =
TI ,(},) + O
i= 1
(Problem 2-22). Hence ,(F) is nonsingular.

The substitution of T A = FT + GC into TA 2 - F 2T yields
T A2 - F 2T = (FT + GC)A - F 2T = F(TA - FT) + GCA = FGC + GCA
Reduced-dimensio
Method I
Consider the
Proceeding similarly, we can obtain the fol1owing set ofequalities:

TI-IT=O
TA-FT=GC
TA 2 - F 2T = GCA + FGC
TA 3 - F 3 T = GCA 2 + FGCA + F 2GC
n-dimen~
where A, B, and C are,

In this section we assUl
TAn-FnT=GCAn-1 +FGCAn-2 + ... +FI- 2GCA+F"- I GC

We multiply the first equation by !X n , the second equation by !X n - I , ... , the last
equation by 1, and then sum them up. After sorne manipulation, we finally
obtain
where R is an (n -q) x
as P is nonsingular. V
where QI and Q2 are r.
In =PQ
O
(7-61 )
=[
Now we transform
x=Px,
where UF is the control1ability matrix of {F, G}, and VA is the observability

matrix of {A, C}. The matrix A q is an nq X nq matrix and is similar to the one in
(7-13). Since ,(A) =0 and ,(F) is nonsingular, Equation (7-61) implies
(7-62)
which can be partitione
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
Although the unique solution T in TA - FT = GC can be computed from

(7-62), the methdmay not be desirable because it requires the .computation
of UF and VA (see Section S-S). The computational problem of TA - FT = GC
has been studied exterisivelyin the literature; see, for example, References S9,
S18, and S107.
The design procedure in Method 1 requires the transformation of a
Lx J=l
2
=[
where XI consists of thl

A12 , A2!> and A22 are,
(n -q) matrices;.B and
that y = Xl ' Hence on
Consequehtly we need 01
n-dimensional estimatol
Usiilg XI =Y,we wri
STATE ESTIMATORS
les
}.
361
dynamical equation into a multivariable observable formo This step is by

no means simple numerically. 5ee the discussions in the state-feedback part.
It is notclear at present which method, 1 or I1, is better in terms of computational
efficiency and numerical stability.
of F and
Reduced-dimensional state estimator
yields
;CA = FGC
Method I
+ GCA
f equalities:
x=Ax +Bu
y=Cx
....... -
(7-63a)
(7-63b)

In this section we assume that C has a full row rank, that is, rank C = q. Define
P~ [~J
GCA +F"-IGC
on by C1. n - l " .. , the last

lanipulation, we finally
(7-64 )
where R is an (n - q) x n real constant matrix and is entirely arbitrary so long

as P is nonsingular. We compute the inverse ofP as
(7-65 )
where Ql and Qz are n xq and n x (n - q) matrices. Clearly we have
III=PQ=[~J[Ql Qz]=[~~: ~~:J=[~q ~-J
o
O
(7-66)
Now we transform Equation (7-63) into, by the equivalence transformation

x=Px,
(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)
rank V A < n, then rank

UF = n and rank V A = n
lly necessary conditions
(1'-67 a )
(7-67b)
'q is always nonsingular.

ingular or, equivalently,
~stablishes the theorem.
Q.E.D.
~here_Xl consis!s of the first q elements of X and Xz the remainder of x; ll ,

A IZ ' A ZI ' and A 22 are, respectively, q x q, q x (n- q), (n - q) x q, and (n - q) x
(n - q) matrices; B l and Bz are partitioned accordingly. We see from (7-67b)
that y = x. Hence only the last n --' q elements of x need to be estimated.
Consequently we need only an (n - q)-diinensional state estimator rather than an
n-dimensional estimator as discussed in ihe' previous subsection.
can be computed from
quiresthe' computation
lblem ofTA - FT = GC
~xample, References 59,
Using Xl = y, we write (7-67a) as
y = At1 y + A12 x 2 + BID
i z =A 22 x Z +AzlY +Bzo
e transformation of a
.J
(7-68a)
362
which become, by defining = A 2y + B 2u and w =y - Ay - B U,

x 2 =A 22 x 2 +
W =A 12x2
Since x = Px, We have

(7-68b)
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.
This gives an estimate
~iag.ram ofthe (n - q)-d
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 )
such that the eigenvalues of (A 22 - LA 12) can be arbitrarily assigned by a

proper choice ofL (Theorem 7-8). The substitution of and w into (7-69) yields
Method 11
Consider the n-dimensior
(7-70)
This equation contains the derivative of y. This can be
elir~lnated by
defining
z =i 2 -Ly
(7-71)
where A, B, and C are res;

It is assumed that
is ir
(7-72)
be an (n - q)-dimensional ,
(n - q) x (n - q), (n - q) x q,
Then (7-70) becomes
z =(A 22 -LAd(z +Ly) +(A 2 -LA l1 )y +(B 2 -LBdu

=(A 22 -LAdz + [(A 22 -LAdL +(A 2 -LA l1 )]y +(B 2 -LBdu
This is an (n - q)-dimensional dynamical equation with u and y as the inputs

and can be readily constructed by using, for example, operational amplifier
circuits. From (7-71), we see that z +Ly is an estimate of x 2. Indeed, if we
define e = x 2 - (z +Ly), then we nave
or
ni
=i 2 -(z +Li) =A 2x +A 22 x 2 +B 2u -(A 22 -LAd(z +Lx)

-(A 2 -LA l1 )x -(B 2 -LBdu-LA l1 x -LA12x2 -LBu
= (A 22 - LA12)(X 2 - Z - Lx)
= (A 22 - LA 12)e
(7-73)
Since the eigenvalues of (A 22
LA 12) can be arbitrarily assigned, the rate of
e(t) approaching zero or, equivalentIy, the rate of (z + Ly) approaching x 2 can
be determined by thedesigner. Hence z +Ly yields an estimate of X2'

Nowwe combine x =y ;=i with i 2 =z +Ly to form
Figure 7-8
An(n -q)-dimen:
=='--__
- ---- _
=:::_=-:c-_--~
STATE ESTIMATORS
Since x =Px, we have x = p- 1 x =
(7-68b)
" and y. Now if the

'X 2 can be constructed.
air {A,C} in Equation

7-68) is observable.
. established in Section
~m is left as an exercise.
. Consequently, there
form
(7-69)
itrarily assigned by a
md w into (7-69) yields
X=Q~=[Q
(7-71 )
Qx or
Q2J[L/+Z]=[Q
Q2J[~ I~-JGJ
(7-74 )
This gives an estimate of the original n-dimensional state vector x. A block

diagram of the (n - q)-dimensional state estimator in (7-72) and (7-74) is plotted
in Figure 7-8.
A comparison between an (n - q)-dimensional and an n-dimensional state
estimators is in order. The reduced-dimensional estimator requires the equiv
alence transformation discussed in (7-64). Excluding this step, the computation
required in the reduced-dimensional case is c1eady less than the one in the full
dimensional case. In the implementation, the former also requires less number
of integrators. In the reduced-dimensional case, the signal y is fed through
the constant matrix Q 1 to the output of the estimator. Hence, if y is corrupted
with noises, the noises wil\ appear in the output of the estimator. In the full
dimensional estimator, the signal y is integrated or filtered; hence high-frequency
noises in y will be suppressed in the full-dimensional estimator.
Method 11
(7-70)
eliminated by defining
FE:
x=Ax +Bu
y=Cx
(7-75a)
(7-75b)
where , B, and C are, respectively, n x n, n x p, and q x n real constant matrices.

1t is assumed that FE is irreducible and rank C = q. Let
z=Fz+Gy +Hu
-(B 2 - LBd"
(7-72)
(7-76)
be an (n -q)-dimensional dynamical equation. F, G, and H are, respectively,

(n - q) x (n - q), (n - q) x q, and (n - q) x p real constant matrices to be designed.
and y as the inputs

, operational amplifier
:e of x2 . Indeed, if we
l U
-----------------------,
--~
-LAd{z + LX 1)
~A12X2 - LB 1"
363
(7-73)
h
ly assigned, the rate of

~y) approaching X2 can
[1 estimate of X2'
rm
Figure 7-8
An (n -q)-dimensional state estimator.
.. _- -_ .. _...
--.--~._-----_
...._--- ---_.. _..
__._..-_ .._--_.
364
_._-_.-.-._--_._----- ----_.
__ .._.__ .__...-
The procedure of computing F, G, and H is similar to Method 11 of the full

dimensional case.
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
3. Solve the unique T in TA - FT = GC. Note that T is an (n - q) x n matrix.

4. If the square matrix of order n
P=[~J
[7-77)
is singular, go back to step 1 and/or step 2 and repeat the process. If P is

nonsingular, compute H =TB. Then the z in Equation (7-76) is an estimate
of Tx and the original state x can be estimated by
(7-78)
I
The proof of Theorem 7-9 can be used to establish that z is an estimate

ofTx. The combination ofy =ci and z =Ti yields immediately (7-78).
Theorem 7 -1 2
If A and F have no common eigenvalues, necessary conditions for the existence

of a full rank T in TA - FT = GC such that
where x denotes nonze:

then we have f31l=O. F
which together with the

Hence (7-79) and rank 1
Because Al is nons:
with the observability (
observable and {F, G}
the proof of this theore:
The discussion in ]
computational probler
dimensional case regar.
equally applicable here
To conclude this se(
Theorem 7 -1 3
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
If {F, G} is not controllable, thenrank UF < n - q. Hence we have rank

T < n - q and P is singular. If {A, C} is not observable, there exists a nonzero
n x 1 vector f such that VAf = O; This implies Cf = Oand Pr = O. HenceP is
singular. This establishes the necessity of the theorem.
7-5
Connection.
Consider the n-dimensi
where A, B, and C are, r'

It was shown in Sectio
the state feedback u =
desired positions. It '"
365
CONNECTIN OF STATE FEEDBACK AND STATE ESTIMATOR
Method II of the full-
Now we consider the single-output case. We show the sufficiency of the

theorem by showing that r = O is the only possible solution in
C ] r-O
[ UFAIV A
1n]', ar
so that aH of its eigen

those of A.
n-q
is an (n - q) x n matrix.
0-77)
eat the process. If P is

ion (7 -76) is an estimate
(7-78)
;h that z is an estimate
nmediately (7"78).
ditions for the existence
(7-79)
~~ IlllXn=
n-
..
..
.
.
fJn
IXI
1 . IX
2
IX n.- 2
IX n.- 3
1
O
where x denotes nonzero elements. From this equation, we see that if Cr = O,

then we have ~II=O. From UFA 1 VAr=UdJ=O, we have, by using ~n =0,
n- 1
"
UfA. V Ar=
L fiF i j-=
G=
fiiFi- 1 G =0
(7-80)
i= I
which together with the controllability of {F, G} imply Pi = O, i = 1, 2, ... , n - 1.

Hence (7-79) and rank UF = n -1 imply p = O.
Because Al is nonsingular, p=AIVAr=O implies VAr=O, which together
with the observability of {A, C} imply r = O. This establishes that if {A, C} is
observable and {F, G} is controllable, then P is nonsingular. This completes
Q.E. D.
the proof of this theorem.
The discussion in Method II of the full-dimensional case regarding the
computational problem and the discussion in Method 1 of the reduced
dimensional case regarding the comparison with the full-dimensional case are
equally applicable here; hence they will not be repeated.
To conclude this section, we summarize the preceding results as a theorem.
Theorem 7 -13
ollable. For the single-
present case. The only

Id UF is an (n - q) x nq
(F)][U:qVJ
Hence we have rank
e, there .exists ~onzero
and Pr = O. Hence P is
If the n-dimensional dynamical equation FE is observable, an (n -q)-dimen

sional state estimator, as given in (7-72)and (7-74)or in (7-76) and (7-78), with any
desired eigenvalues (provided complex conjugate eigenvalues appear in pair)
can be constructed, where q is the rank of matrix C in FE.
11
7-5
Connection of State Feedback and State Estimator
FE:
x=Ax +Bu
y=Cx
(7-81 a)
(7-81 b)

It was shown in Section 7-3 that if {A, B} is controHable, the introduction of
the state feedback u = r + Kx can place the eigenval ues of (A+ BK) in any
desired positions. It was shown in Section 7-4 that if {A,C} is observable, a
---_
366
_ __ .
"'_._._-..
_._--_
_-~
.._ ,_ .. __ _._. __ ..
_. ---_ .. __ .. _-_. ._..__ .. _

__
_------".~---_._.-
-.-.--" ..
..
~-_
_-_ ..
_~.~
..
STATE FEEDBACK ANDSTATE ESTIMATORS
state estimator, with arbitrary eigenvalues, of dimension n or n - q can be

constructed by using u and y as the inputs. In the fol1owing, we consider only
the reduced-dimensional case; the ful1-dimensional case is left as an exercise.
Consider the (n - q)-dimensional state estimator developed in (7-76) and (7-78):
z=Fz +Gy +Hu
(7-82a)
To simplify (7-85), w
After sorne manipu\.

final1y obtain the fol1
(7-82b)
with TA -FT = GC, H =TB and
[Q1
Q2J
[~] =Q1 C +Q2T =1
(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)
as shown in Figure 7-9. Three questions may be raised in this connection:

(1) In the state feedback" u = r + Kx, the eigenvalues ofthe resulting equation are
given by the eigenvalues of A + BK. In the estimated state feedback u = r + Kx,
do we stil1 have the same set of eigenvalues? (2) Will the eigenvalues of the
state estimator be affected by the feedback u = r + Kx? (3) What is the effect
of the estimator on the transfer-function matrix from r to y? To answer these
questions, we form the composite dynamical equation of the system in Figure
7-9 as
[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
Feedback rrom the estimated state.
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
CONNECTION OF STATE FEEDBACK ANO STATE ESTIMATOR
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)
To simplify (7-85), we introduce the following equivalence transformation
After sorne manipulation and using T A - FT = Ge, H = TB and (7-83), we

finally obtain the following equivalent dynamical equation
(7-82b)
[:]=[A
y=[C
(7-83)
11 parts, the estimated

y.
:ual state of FE. If the

1to apply the feedback
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)
into (7-81) and (7-82a).
Q2][:] +[:]r
BK
F
OJ [:]
(7-86a)
(7-86b)
The eigenvalues of a dynamical equation are invariant under any equivalence

transformatian; hence we conclude from (7-86a) that the eigenvalues of the
system in Figure 7-9 are the unian of those of A +BK and those of F. We see
that the eigenval ues of the state estimator are not affected by the feedback and,
as far as the eigenvalues are concerned, there is no difference in state feedback
from the estimated state xor from the actual state x. Consequently, the design
of state feedback and the design of a state estimator can be carried out inde
pendently and the eigenvalues of the entire system are the union of those of state
feedback and those of the state estimator. This property is often called the
separation property.
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
STATE FEEDBACK ANO STATE ESTlMATORS
dynamical equation is corrupted by noises and modeled as a stochastic process,

the matrix L can be computed by minimizing the mean square error. The
interested reader is referred to References S4 and S10.
Example 1
Now we may apply th.
Consider the inverted pendulum problem studied in Exmple 2 on page 338.
Its dynamical equation is
d-!---
~ I~- -~ -~l + r-! l

l~ :
oJ
J
y = [1
: O
-2
OJ x
As discussed there, if we introduce the state feedback
u(t) = r(t) + \j
lf-
~Jx
then the resulting equation has eigenvalues -1, -2, and -1 j.

Now if the state x is not available for feedback, we must design a state esti
mator. We use Method 1 of the reduced-dimensional state estimator to design
a three-dimensional estimator. Since the equation is already in the form of
(7-67), the step of the equivalence transformation in (7-64) can be skipped.
Clearly we have
-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=
!]
Let us choose arbitrarily the eigenvalues of the estimator as - 3 and - 3 2j.

Then we have to find a L such that the eigenvalues of A22 - LA l2 are - 3 and
- 3 2j. For this problem, we shall solve L by brute force. Let L' = [11 /2 /3J,
where the prime denotes the transpose. Then we have
A ~ A22 -L~~12 = =~~

_
L-/3
r -6
and
This completes the des
~]
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
By equating its coefficients with those of

(5 +3)(5 +3 +2j)(5 +3 -2j)=S3 +95 2 +315 +39
we obtain
/2
= -
36
Hence a three-dimensional state estimator is, following (7-72) and (7-74),
O] [ZI] [-45] [1]
1
O
and a11 eigenvalues of F

matrix, then z(t) appro
to meet
Z2
Z3
240 y +
O u
576
-2
then w(t) approaches k,
o [I q oJ can be replaced by
developmenl, with slight m
CONNECTION OF STATE FEEDBACK AND STATE ESTIMATOR

lS
LO
a stochastic process,
square error. The
"
-84
O
O
x= -36
Now we may apply the state feedback from

mple 2 on page 338.
u= r
xas
+ [~ Jf- \:1.
-1f-Jx
I
This completes the designo

u
Id -1 j.
ust design a state esti
lte estimator to design
lready in the form of
7-64) can be skipped.
O]
r as - 3 and - 3 2j.
LA 12 are - 3 and
. LetL' =[/1 /2 /3]'
369
Functional estimators. If all the n state variables are to be reconstructed,

an estimator of dimension at least n - q is required. In many cases, however, it is
not necessary to reconstruct al1 the state variables; what is needed is to generate
sorne functions of the state, for example, to generate Kx in the state feedback.
In these cases, the dimensions of the estimators may be reduced considerably.
In the following, we discuss a method of designing an estimator to generate
kx, where k is an arbitrary 1 x n constant vector. Such an estimator is caBed a
functiona/ estimator or observer.
Consider the dynamical equation in (7-63) with rank e =q. The problem
is to design an estimator so that its output wil\ approach kx as t -H:JJ. First
we transform (7-63) into the form in (7-67), rewritten here for convenience,lO
[:J
[i~: i~:][:J +[:J u ~ Ax +Bu
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
z=Fz +Gy +Hu
!2 -
w = Mz
+ Ny
(1-87a)
(1-81b)
where F is a m x m matrix; G, m x q; H, m x p; M, 1 x m and N,l x q. The

problem is to design such an estimator of a dimension as smal1 as possible so
that w(t) will approach kx = kP-1x =kx, where k = kP-1. Following Theorem
7-9, we can show that if
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)
7-72) and (7-74),
- ]u
-2
then w(t) approaches kx(t) as
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
Now we partition T as T = [T T 2]' where T and T 2 are, respectively,

m x q and m x (n - q) matrices. Then (7-88a) can be written as, using =
[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)
lf we also partition k as k = [k kz], where k and k 2 are, respectively, 1 x q

and 1 x (n -q) vectors, then (7-88c) becomes
MT +N=k
MT 2 =k 2
and
(7-89c)
(7-89d)
Now we choose F and M to be of the forms
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)
These equations, except the last, can be written as

t 2 =k 2
t 22 =k 2A 22 +t2
t 23 = k2~2 + t2 22 + t2 2
has rank n -q. No

is observable. Furt
where v is the observ;
then V has rank n
Once T is known, v
N from (7-89a), (7-8:
estimator of kx. Th;
(7-9Ob)
The substitution of these t 2j ,j = 1,2, ... ,m, into the last equation of (7-90a)
yields, after sorne straightforward manipulation,
lf the dynamical eqt

then for any 1 x n r
functional estimator
approaches kx(t) exp
The dimension (lO

example, if n = 12 an
index v could be ; =
then n - q = v - 1.
lf a p x n feedbac
where vis ap xl veCl
7-2, then Kx can als.
estimator. For a ge
eigenvalues of the fu
arbitrarily chosen. 1
lt mc~Y be possible _(
In the design, if we ;
dimension of estimat
5131 and 5184.
k~k2''2 +rx.k 2''; + ... +rx.m-k222 +rx.m k2
= - [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
DECOUPLING BY STATE FEEDBACK
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
has rank n -q. Now {A, C} in (7-63) is observable if and only if {A 22 , 12 }

is observable. Furthermore, the observability index of {A zz , A1Z } is v - 1,
where vis the observability index of {A, C} (Problem 7-27). Hence, if m = v - 1,
then V has rank n - q, and for any k and any (Xi, a solution T 1 exists in (7-91),
Once T 1 is known, we compute T z from (7-9b) and then compute G, H, and
N from (7-89a), (7-88b), and (7-89c). The resulting estimator is a functional
estimator of kx. This is stated as a theorem.
(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 ]
vi} is observable. Let

:cause of the forms of F
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
ast equation of (7-9a)
*7-6
Decoupling by State Feedback
Consider a p-input p-output system with the linear, time-invariant, dynamical

equation description
FE:
(7-91 )
where u is the p x 1 input vector,
; given and (Xi are chosen.
x=Ax+Bu
y=Cx'
(7-92a)
(7-92b)
y is the p x 1 output vector; A, B, and e are
......
_._- _._-_.... ....
372
__
. '. _~'_'._
_ _
.~n
: :_
._
. :.~_.::..:.
_ _::..:_.~
_ - - - . __
. . . . _.-~._.
-
__

-:-
STATE FEEDBACK AND STATE ESTIMAlDRS
n, n X p, and p X n real constant matrices, respectively. It is assumed that
The transfer function
p :::; n. The transfer function of the system is
G(s)=C(sI-A)-lB
(7-93)
Clearly G(s) is a p X P rational-function matrixo If the system is initially in

the zero state, its inputs and outputs are related by
Y1 (s) = 911 (S)U1 (s) + gds)uz(s) +
yz(s) = gZl (S)U1 (s) + gds)uz(s) + .
o.
+ glP(S)Up(s)
+ gzp(s)uP(S)
(7-94)
. Yp(s) =gp1 (S)U1(S) + gpz (s)uz (s)
+ ...
We shall derive in th,

can be decoupled by s
d ~ min (the differen

of each entr
and the 1
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
The differences in deg
Definition 7-1
A multivariable system is said to be decoupled if its transfer-function matrix is

diagonal and nonsingular.
I
The differences in deg
Ez = 1
In this section we shall study the problem of the decoupling ofmultivariable

systems by linear state feedback of the form
u(t)=Kx +Hr
(7-95 )
where K is a p x n real constant matrix, H is a p >< p real constant nonsingular

matrix, and r denotes the p x 1 reference input. The state feedback is shown in
Figure 7-10. Substituting (7-95) into (7-92), we obtain
FE!:
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_
We see from this 1

is a property of its tra
tion comes into playa
forethe controllabilit y
ofthe system are imm
... -_----_ _._--_.. _.. ------
__._----- .- --
_-
--._--
------:==::::::::::=::::-._--
DECOUPLlNG BY STATE FEEDBACK
The transfer function of the state-feedback system is
Iy. It is assumed that
(;(s, K, H) ~ C(sI - A -BK)-IBH

(7-93)
~
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)
and the 1 x p constant row vector E as
(7-94)
E ~ lim
pp(s)uP.(s)
Sd;+ l
(;(s)
S-oo
at every input controls

led by more than one
oupling or interacting,
ystem. For example,
(s), )\(s), . .. , yp(s); the
found. Therefore, in
that a coupled multi
Lt every input controls
Iy one input. Conse
isting of a set of inde
ansfer-function matrix
; decoupled.
where (;(s) is the ith row of (;(s).

Example 1
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)
A system. with the transfer-function rnairix G(s) can be decoupted by state

feedback of the form u = Kx + Hr if and only if the constant matrix
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
respectively. Then, fram (7-93) and using (2-3), we have

G(s) =C(sl - A)-lB
(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,
G(s) = Ll(s) Ci[sn-11+R1Sn- Z + ... + Rn-1]B
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)
R n- 1 = AR n- z +cxll-11 = An-1 +cx1A"- z + ... +cxn-11

Since di + 1 is the smallest difference in degree between the numerator and the
denominator of each entry of Gi(s), we conclude from (7-99) that the coefficients
associated with s"-d,sn-d+l, ... ,and sn-1 vanish, but the coefficient matrix
associated with sn-d-l is different fram zero. Hence we have
CiB=O, CiR1B=0, ... , CRd_lB=O
Ei=CiRdB",O
and 11
(7-102)
(7-103 )
Fram (7-101) it is easy to see that conditions (7-102) imply that
CiB=O,
CiAB=O,
... ,
CAd-lB=O
(7-104)
and Equation (7-103) becomes
Therefore, di can also be defined from a dynamical equation as the sma!1est

integer such that CAdB "'0. If CAkB =0, for all k ::;n, then di ~ n -1. (See
Prablem 7-28.)
Similarly, we may ~define d and E i for the transfer-function matrix G J in
(7-97). If we expand GJi into
G Ji = ~(s) [CBHsn- 1 +CiR1BHs"- z + ... +CRII-1BH]
(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,
this diITerence, if ~ny, i~ neglected.
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
(7-105)
...
For any K and any no

Proof
R1=A+cx11
Rz=AR 1 +cxzI=A z +cx1A+cxzI
(7-101 )
11
then
Necessity: Suppose th~

and nOllsingular The:
is a diagonal constant
is nonsingular by assu;
Sufficiency: If the mah
Define
__ '
"_~'_W'_"._.'_'
"~
_ _ _ _ , . ~ .
__
.~
__
_ _ _"_ __ __
..
w_._~._.,,
__"~_"'_
-_.. - . __.- ..
_.~
'.'-
...__ . _.W __ __ _' .
,,_._
~y
. . . . _ _. __ __ . ~ " _ " _ - _ . _ . -
DECOUPLING BY STATE FEEDBACK
R1 =(A +BK) +&11

R2 =(A +BK)2 +&l(A +BK) +&21
and
(7-98)
:en the integer di, the

.pand (7-98) into (see
(7-108)
then
CBH=O,
and
+ ... +CR n- 1B]
375
C(A +BK)BH=O, ... , C(A +BK)d-lBH=O

E =C(A +BKfBH feO
(7-109)
(7-110)
We establish in the following the relationships between d and Ji and between

E i and E.
(7-99 )
Theorem 7 -16
+0:
11
(7-100)
For any K and any nonsingular H, we have
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)
It is easy to verify that, for each i, the conditions CB = O, CAB = O, ..

CAd-l B = Oimply that
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
Since H is nonsingular by assumption, if E i is nonzero, so is EH. Therefore

we conclude that di = di and Ei = KH.
Q.E. D.
With these preliminaries, we are ready to prove Theorem 7-15.
Proof of Theorem 7-1 5
(7-105)
lation as the smallest

then d ~ n -1. (See
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
STATE FEEDBACK ANO STATE ESTlMATORS
We show that if K = - E - 1F and H = E - 1, then the system can be decoupled

and the transfer function of the decoupled system is
1
1
G(s, -E- F,E- )=
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
for any positive integer k
(7-115)
di =d, Equation (7-106) reduces to
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
Consequently, Equation (7-116) becomes
C(sl-A-BK)-lBH=L\~S)[sn-d-l +&lSn-d-2
+ ... +&n-1-dJEH (7-117)
What is left to be shown is that

L\(s)=sn +i 1s n- 1 +&2Sn-2 + ... +&n
= Sd +l(sn-d-l + &sn-d- 2+ ... + &n-d- )
(7-118)
From the Cayley-Hamiltonthebrem, we have

(A +BK)n +&(A +BKr 1 + ... +&nl =0
(7-119)
By multiplying CiCA +BK)d to(7-119) and using (7-115), we can obtain

&nC(A+BK)d=O which implies that &n=O. Next by multiplying
CiCA + BK)d -1, we can show that&" _1= O. Proceeding similarly, we can prove
Equation (7-118). By substituting (7-118) into (7-117) and using EH = EE- 1 =
In this chapter we studi,

observability. We showe
is controllable, we can, t
eigenvalues of the resultin
of controllable canonicaJ
state feedback, all the sta
asymptotic state estimatoJ
observable, a state estima
structed. The constructic
servable canonical-form d)
TA-FT=GC.
We studied in this ch,
Section 7-2, 7-3, and 7-4 c,
equation is instantaneousl
trollability assumption, th
dynamical equation has a fl
transformation, a controll
obtained. Consequently
remarks apply to the obser"
We also studied the der..
state feedback. The cone
required in this part of Stl
decoupling also used the st.
The combination of stat.
in Section 9-5 directly fram t
the state-variable approad
discussed there.
The introduction of con
constant gain output feedbac
cannot arbitrarily assign all
---:~---~--~~-----~------~. ------~----------.-------------
CONCLUDING REMARKS
stem can be decoupled
(7-114)
os elsewhere. First we
and (7-113) and using
e, we immediately obtain Equation (7-114). Consequently, the system can be
decoupled by using K =
(7-115)
integer k
BS n -
d -2
+ ...
_CRn-BJH
(7-116)
lct C(A +-BK)kB =0, for

he nonsingularity of H,
E
BK)d'JB= rilK
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.
Although a system can be decoupled by using K = - E - 1 F and H = E - 1,

the resulting system is not satisfactory because all the poles of the decoupled
system are at the origino However, we may introduce additional state feedback
to move these poles to the desired location.
If the dynamical equation is controllable, then we can control all the eigen
values by state feedback. Now, if in addition we like to decouple the system,
the number of eigenvalues that can be controlled by state feedback will be
reduced. For a complete discussion of this problem and others, such as how to
compute K and H to have a decoupled system with desired poles, see References
36,41, 112, S218, and S227.
7-7
Hence we conclude that
377
Concluding Remarks
In this chapter we studied the practical implications of controllability and

observability. We showed that if a linear time-invariant dynamical equation
is controllable, we can, by introducing state feedback, arbitrarily assign the
eigenvalues of the resulting dynamical equation. This was achieved by the use
of controllable canonical-form dynamical equations. When we introduce
state feedback, all the state variables must be available. If they are not, an
asymptotic state estimator has to be constructed. If a dynamical equation is
observable, a state estimator with a set of arbitrary eigenvalues can be con
structed. The construction of state estimators is achieved by the use of ob
servable canonical-form dynamical equation or by the solution of the equation
TA-FT=GC.
We studied in this chapter only the time-invarlant case. The results in
Section 7-2, 7-3, and 7-4 can be extended to the linear time-varying case if the
equation is instantaneously controllable since, under the instantaneous con
trollability assumption, the controllability matrix of the linear time-varying
dynamical equation has a full rank for all t. Hence by using it in the equivalence
transformation, a controllable canonical-form dynamical equation can be
obtained. Consequently the state feedback can be introduced. Similar
remarks apply to the observability part.
We also studied the decoupling of a transfer function matrix by introducing
state feedback. The concepts of controllability and observability are not
required in this part of study. It was discussed in this chapter because the
decoupling also used the state feedback as in Section 7-3.
The corilbination of state feedback and state estimator will again be studied
in Section 9-5 directly from the transfer-function matrix. Comparisons between
the state-variable approach and the transfer-function approach will also be
discussed there.
The introduction of constant gain feedback from the output is called the
constant gain output feedback. Unlike the state feedback, the output feedback
cannot arbitrarily assign all the eigenvalues of the resulting equation. Roughly
378
speaking, the number of eigenvalues which a constant gain output feedback

can assign is equal to min {n, p + q - 1}, where p = rank B and q = rank C. In
this assignment, no repeated eigenvalues are permitted. Furtherrnore, the
assignment cannot be exact hut only arbitrarily close. See References S31,
S71, S129, Sl30, and Sl31. See also Reference S165.
To conclude this chapter, we mention that al1 the results in this chapter
can he applied direct1y, without any modification, to the discrete-time case.
We mention that the discrete-time state estimator should be modeled as
7-7 Consider the state fe

Show that for any Constan
Under what condition on 1
and H are q x 11, P X 11, al1(
that n 'g.) (Answer: A solu
Consequently, a soluton 1
singular.)
.
x(k + 1) = Ax(k) + Bu(k) + L(y(k) - Cx(k))
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
7-1 The equivalence transformaton Q ofi = Px = Q - IX in Theorem 7-1 can be obtained

either by using Q = [ql ql ... qn], where the q's are given in (7-9), or by using Q =
U- l , where U is the controllablty matrix of FE I and - I is given in (7-13). From a
computatonal point of view, which one is easier to use?
7-2
Figure P7-7
Consider the controllable dynamical equaton
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
Fnd the values of k = [k l I

Fnd the values of h such th
concIude that for almost all k
7-9 Use state feedback to tra
7-4to -1, -2,and -2. Drav
and then add the requred stat,
7-10 Show that

p[b
7-3 Find the matrix Q which transforms (7-17) into the observable canonical-form
dynamical equation.
.. " An-l
for any 1 x n constant vector k.

7-11
7-4 Transform the equaton
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.
whch has an unstable egenvaJw

subequation associated w'th .
st a bI"l" Ize the equatl'on by " elge
.
usmg sto
closed-Ioop equaton has eigen~a':
7~12
Gven
PROBLEMS
gain output feedback

B and q = rank C. In
ed. Furthermore, the
See References S31,
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
results in tbis cbapter

tbe discrete-time case.
.Id be modeled as
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))
\red to Reference S215.
Figure P7-7
7-8
[heorern 7-1 can be obtained

en in (7-9), or by using Q =
'1 is given in (7-13). From a
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
tat is its equivalent dynamical
\cien ts of the transfer function
!See Problem 2-15.)
7-10
I,e observable canonical-form
for any 1 x n constant vector k.
Show that
p[b Ab .,. An-1b] =p[b
7-11
(A -
bk)b
Consider the lordan-form dynamical equation
1'lVhat is its transfer functon?
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.
!lz is linearly dependent on
7-12
lnto the observable canonical-
Given
------.;=:::::-.--._.==:-.::__
.-_--.. ---=-=~.~-~.:::::.-.-:::::--:::::-.--:::::.---:..:..:. -:::::--::::::::::--:::::.. -:::::
. --::::::: : .- : : :-.-: : : : : "': : :"-:: : "-', : : ---:: : -- ------.---"'-'--'--'"--.---.----.- - - - - - - - - . - - . - . ------.---.---.-.--.----_.. '~"'--'" - - - 1
380
STATE FEEDBACK AND STATE ESTIMATRS
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
Consider the uncontrollable state equation
1
2
7-21
Verify that the ma
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.)
Find two different real co

values - 4 3j and - 5
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
Show that the state
7-24
Show that the (11 -
7-14
<:
7-9.
Find a three-dimensional state estimator with eigenvalues - 2, - 2, and - 3 for the

dynamical equation in Problem 7-4. Use two different methods.
7-15
7-25
Establish the sepan
7-26
Consider the dynarr
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
Consider a system with the transfer function

(s - l)(s +2)
(s
+ l)(s- 2)(s + 3)
ls it possible to change the transfer function to

s-l
y=
Find a full-dimensional es
estimator for kx(t) with
(s+2)(s+3)
by state feedback? 1f yes, how?

In Problem 7-17, if the state is not available for feedback, can you carry out the
design by using a state estimator? Choose the eigenvalues ofthe estimator as -4 and - 4.
What is its overall transfer function?
7-18
7-19
Prove Theorem 7-11.

observability ndices of (A,
l
7-27
_~
Prove Theorem 7-11.

1
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
and note that if linearly ind

order from top to bottom, tI
ability matrices are identical.
811(S)=s"l+alllSn,-1 + ... +alln,=(s-IX)TI (s-A)

i=l
"1
-1
821(s)=a21ISn,-I+ ... +a21I1,=-j3 TI (s-}ct)'

i=l
"2 -
8 22 (s) =S"2 +a221Sn2 -
+ ...
+a22112 =(s -,- IX)
II
i=Ol
8ds)=aI21 Sn2 - 1
+... +aI2112=fi
/12 -
TI (s-y;)
i=l
7-28
(5 - ')') .
Consider the irreducib
I
,
.~.
where A is an n x n matrix, b i:
function. Show that gis) ha~
PROBLEMS
:m has -1 and -2 as it,

i1ence transformation.
7-21
Verify thatthe matrix in (7-50) has eigenvalues - 3 ~i, - 3 j and - 4 2j.
7-22
Let
A=
state feedback u = kx + r
lValues -2, -2, -2, -1?
t under any state feedback.

of Theorem 7-3 is used to
:s - 2, - 2, and - 3 for the

1s.
[~ ~ ~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.
+ BK) has eigen
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
Establish the separation property by using a full-dimensional state estimator.
7-26
Consider the dynamical eguation
_ 2 and - 3 for the dynam
o
O
1
O
O
O
1
O
1
1
1
1
O
Find a full-dimensional estimator, a reduced-dimensional cstimator, and a functional

estimator for kx(t) with
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
:igenvalues of the matrix are
111 -
Consider the irreducible single-variable dynamical eguation
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
degree m-if and only if
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
fori=0,1,2, ... ,d-2
(Hin!: Show that cAib is invariant under any equivalence transformation and then use the
controllable canonical-form dynamical equation.)
7-29
a. Can a multivariable system with the transfer function
52
+1
"['J
25 +1
53 +5 +1
be decoupled by state feedback?
b. Can a system with the dynamical-equation description
X~[~
y=[~
-~} +[~
1
O
1 -1
~J x
-1
2
~}
be decoupled by state feedback?

7-30
Consider the linear time-invariant discrete-time equation
-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
Establish a counterpart of Theorem 7-9 for the discrete-time case.
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
Hin!: Use the method or
~::':=:-~:_:;'--=-==:":::="===;::_'_'_"
' "'__
~'L
_ _ ' _ _ .
.~
_ _
._
. _
_ ~
__
__
._-----_._-----------~-----,-_._._-~----~._~-----~-~.--,------"
PROBLEMS
Find a k so that (A + bk) is similar to

2
degree ofthe denominator
-1
F= [ ~
if
.,d-2
formation and then use the
Hint: Use the method on page 347. Choose k = [6

1
T= -3
[
-1/2
ion
resulting equation are zeros.

e equation becomes identical1y
oblem 2-50.
ete-time case.
;es of {A + BRe, B, C} are the
ontrollability and observability
ack.
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) =
Stability of Linear Systems
where g(t, T) is the im

output measured at tir
In the qualitative ~
perhaps the only gues
properties, under wha
For example, if the inp
\u(t
under what condition e
output y satisfies
8-1
Introduction
for allt in ( -
Controllability and observability, introduced in Chapter 5, are two important

gualitative properties of linear systems. In this chapter we shall introduce
another qualitative property of systems-namely, stability. The concept of
stability is extremely important, because almost every workable system is
designed to be stable. Ir a system is not stable, it is usually ofno use in practice.
We have introduced the input-output description and the dynamical-equation
description of systems; hence it is natural to study stability in terms of these
two descriptions separately. In Section 8-2 the bounded-input-bounded
output (BIBO) stability of systems is introduced in terms of the input-output
description. We show that if the impulse response of a system is absolutely
integrable, then the system is BIBO stable. The stability condition in terms of
rational transfer functions is also given in this section. In Section 8-3 we
introduce the Routh-Hurwitz criterion which can be used to check whether or
not all the roots of a polynomial have negative real parts. We study in Section
8-4 the stability ofa system in terms ofthe state-variable description. We intro
duce there the concepts of equilibrium state, stability in the sense of Lyapunov.
asymptotic stability, and total stability. The relationships between these
concep.ts.are established. In Section 8-5, we introduce a Lyapunov theorem
and then use it to establish the Routh-Hurwitz criterion. In the last section,
the stability of linear discrete-time systems is discussed.
The references for this chapter are 58; 59, 65, 76, 90, 102, 116, S79, S 111,
S135, and S153.
00,
oo)?
does there exist a cons.
If the bpul approache

output approach anotl
input approaches a cons
ing to these various pro
for the system. We sha;
used in linear systems: t
Recall that the inpl
when the system is initia
relaxed system. Hence t
description is applicable
Definition 8-1
A relaxed system is said
if and only if for any bOL
384
STABILITY CRITERIA IN TERMS OF THE INPUT-OUTPUT DESCRIPTION
385
8-2 Stability Criteria in Terms of the Input-Output

Description
Time-varying case. We conslder first single-variable systems, that is,
systems with only one input terminal and only one output terminal. It was
shown in Chapter 3 that if the input and output of an initially relaxed system
satisfy the homogeneity and additivity properties, and if the system is causal,
then the input u and the output y of the system can be related by
y(t) =
Loo g(t, "C)U("C) d,
for aH t in ( - co, oc!)
(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!
for aH t in ( - co, oc!)? Ir the input is of finite energy, that is,

,ter 5, are two important
lpter we shall introduce
ability. The concept of
'ery workable system is
lal1y ofno use in practice.
d the dynamical-equation
tability in terms of these
lounded-input-bounded
erms of the input-output
of a system is absolutely
.litY condition in terms of
lion. In Section 8-3 we
used to check whether or
lrts. We study in Section
,le description. We intro
in the sense of Lyapuno v ,
ltionships between these
uce a Lyapunov theorem
rion. In the last section,
:d.
" 90, 102, 116, S79, Slll,
(8-3)
does there exist a constant k 4 such that
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
STABlLITY OF LINEAR SYSTEMS
STAB!
We illustrate the importance of the qualification "relaxedness" by showing

that a system, being BIB stable under the relaxedness assumption, may not
be BIB stable if it is not initially relaxed.
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.
A relaxed single-variable system that is described by

y(t) =
We consider now
terminals and q outPl
Loo g(t, r)u(r) dr
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
Loo Ig(t, r)\ dr =:;;,k< 00

for all t in (- 00,00).
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 = \
Loo g(t, r)u(r) dr I=:;;, Loo Ig(t, r)llu(r)i dr =:;;, k
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
Loo Ig(t, r)1 dr =

for sorne t, say
tI'
In
then we can find a bounded input that excites an unbounded
In
u"-'
-i
In
00
Then 9ii is the impul~

output terminal. Sim
system is defined to be
the output vector is ba
ponent of the vector
pair of input and outp
number of bounded fu
theorem.
Theorem 8-2
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.
is BIB stable ir and e

entry of G,
for all t in ( -
00,
(0).
STABILlTY CRITERIA IN TERMS Of THE INPUT-OUTPUT DESCRIPTION
axedness" by showing
; assumption, may not
387
output. Let us choose

(8-4 )
u(t) =sgn [g(t 1 , t)]

where
sgn x
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
Clearly u is bounded. However, the output excited by this input,

y(t 1 )
= Loo g(t 1 , r)u(r) dr =
Loo \g(t
1,
r)1 dr = 00
Q.E.D.
is oot bouoded.
We consider now multivariable systems. Consider a system with p input

terminals and q output terminals, which is described by
y(t)=
r k such that
foo G(t, r)u(r)dr
(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)
k 1 for all t in (-00,00).
gdt, r)
glP(t,r)l
gzp\t, r)
. (8-6)
gqp(t, r)
lis part is rather involved.
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.
)at excites an unbounded
Theorem 8-2
[00
Ig(t, r)i dr
~ kk 1
A relaxed multivariable system that is described by
y(t) =
foo G(t, r)u(r) dr
is BIBO stable if and only if there exists a finite number k such that, for every
entry of G,
bounded input if the intia\
for all t in ( - 00, co).
--_ ..
388
_-----------~------_._-------_
.. _-----
----------_
._-
..- --
-.---~---.-_._-_
STABIL
STABlLlTY Of LINEAR SYSTEMS
Consider a single-variable system with the following
input-output description:
(8-7)
y(t) = J>(t-L)U(L)dL= Ig(L)U(t-L)dL
where g(t) is the impulse response of the system. Recall that in order to have a
description of the form (8-7), the input-output pairs of the system must satisfy
linearity, causality, and time-invariance properties. In addition, the system is
assumed to be relaxed at t = O.
234
Corollary 8-1
A relaxed single-variable system which is described by

y(t) =
Figure 8-2 An absolul
zero as t -> oo.
g(t - L)U(L) dL
is BIB stable if and only if

If gis absolutely ir
tOO [g(t)\ dt sk <
00
for sorne constant k.

as o: -> oo.
Proof
This follows directly from Theorem 8-1 by observing that
Ig(t, L)I do =
Ig(t - L)I dL =
Ig(o:)1 do: s
Theoreril 8-3
100 Ig(o:)1 do:
For the time-invariant case, the initial time is chosen at

gration in (8-8) starts [rom instead of - oo.
We shall u:
= O;
(8-8)
Consider a relaxed si!

related by
hence the inte

Q.E.D.
A function 9 is said to be absolutely integrable on [0, (0) if
rOO Ig(t}1 dt sk <

Jo
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
for n = 2,3,4, . . .. The function

is depicted in Figure 8-2. lt is easy to
verify that f is absolutely integrable; however, f is neither bounded on [0, (0)
nor approaching zero ast -> oo.
Since
and since
is nondecreasing as ex increa
Consequently,
STABILlTY CRITERIA IN TERMS OF THE INPUT-OUTPUT DESCRIPTION
389
tem with the following
Id,
(8-7)
n
that in order to have a

he system must satisfy
addition, the system is
II
2
Figure 8-2 An absolutely integrable function that is neither bounded nor tending to
zero as (---+00.
If g is absolutely integrable on [0, (0), then 1

(8-9)
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)
Consider a relaxed single-variable system whose input u and output y are

related by
it t = O; henee the inte
Q.E.D.
y(t) =
J~ g(t -
,)u(,) d,
0, (0) if
Lite. The faet that g is

l on [0,(0) nor that g(t)
)ll defined by
Since
r
o
\g(/)1 dI =
r r
o
Ig(/)1 dI +
Ig(/)1 dI
and since
-n)~O
Ig(t)1 d(
is nondecreasing as fY. increases,

liro
a-
gure 8-2. ltis easy to

ither bounded on [0, (0)
Conseq uen tI y.
00
Jao Ig(/)\ dt ..... J.'o" 19(1)1 dt
390
STABILITY OF LINEAR SYSTEMS
STABIl
If
2. We prove this par

[ ' [g(t)1 dt =::; k <
00
for sorne constant k, then we have the following:
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
\u(t)lz dt}/Z =::;k l <
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
(IX> [y(t)i2 dt}/z =::; kz <
g(r)u(
for all t ~ t l. As car:

is approxirnately eq
becornes, for all t ~
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
for sorne constant k

space. In this space
(see Prablern 2-49):
g(r)u(t-r)dr
(8-11 )
Subtracting (8-10) frarn (8-11), we obtain

y(t +T) - y(t)1
Ir+
g(r)u(t -r)drl
r'+T
=::;
J,
IIJII ~ (
ig(r)llu(t -r)\ dr =::;UM
where UM ~ rnax lu(t)l.
rl + T
J.
The Schwartz inequa

~ 1"' '-'
11
Ig(-r)\ dr
It follows frarn (8-9) that Iy(t) - y(t + T)i-4 O as
J(t)gi
OSIST
t-4 00, or y(t)-4 y(t
+ 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
--\-
.,..-__
then there exists a finite k 2 .such that
(r
and the system is said to be Lp-stable.
. tI
[y(t)!P dt}'P ::;;k 2 < co
Figure 8--3:
The convolut
STABILITY CRITERIA IN TERMS OF THE INPUT-OUTPUT DESCRIPTION
391
2. We prove this part by an intuitive argumento Consider
y(t) =
1',
g('1:)u(t - '1:) d'1:
J' g('1:)u(t - '1:) d'1:
Let UM ~ max
, Iu(t)l Then we have, with t > t b
\r
u(t) = u(t + T) for aH

with the same period
(8-12 )
'1
g('1:)u(t -'1:)d'1: \ s
Ig('1:)llu(t -'1:)1 d'1: SUM
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
wil1 tend to a constant.
y(t) =i=
(8-13 )
g('1:)u(t - '1:) d'1:
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,
: exists a finite k'}. that
y(t) =i= (f.
g('1:) d'1:
which is independent of t. This completes the proof.
y(t)--> y(t + T) as t --> oo.
3. It can be shown that the set of aH real-valued functions defined over [O, (0)
with the property
(ro
(8-10)
)u(t - -c) d'1:
If(t)i2 dt) l/'}.
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 )
Ilfll ~ ([' If(t)i2 dt),l').
(1, g) {
ro
f(t)g(t) dt
The Schwartz inequality (Theorem 2-14) reads as
I[} f(t)g(t) dt Is (t
dl
r)\ d'1:
t Iy(t) - y(t
[fU)\2 dt) 1/2
(I"
Ig(t)i2 dtY
+ T)\ --> o as
]. ir
u (t - T)
tY:+-:~-....,....------------f-::--~----;I'''''''''--t------~-------I-_T
Figure
8-3~
The convolution of g and u.
392
STABlLITY Of LINEAR SYSTEMS
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)

ly(t}1
lf
too (ig(r)11/2)(ig(r)/1/2Iu(t - rm dr
(8-14 )
where () = tan -1 (lm

and 1m denote the re~
Applying the Schwartz inequality, (8-14) becomes
Proof
Since sin (t - r) = sin t
Consider now
too ly(t)i2 dt
~k
y(t) =
g(r)u(t-r)dr
too too Ig(r)llu(t -r)i2 dr dt
= k tOO (too lu(t - r)i2 dt) Ig(r)l dr
sin wt too g(r) ce
(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(
Hence, from (8-9) we (

y(t)->sin w
Hence (8-15) implies that

too ly(t)12 dt
~kk1
too Ig(r)1 dr =k 2k 1
In other words, if g is absolutely integrable on [O, 00) and if u is of finite

energy, then the output y is also of finite energy.
Q.E. O.
We see from Corollary 8-1 and Theorem 8-3 that if a system that is describ
able by
y(t) =
f>(t -
By definition, g(s) is th
which implies that

g(iw) =
roo g(
Jo
Since g(t) is real-valuec
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
y(t) -> sin wt (R
Consider a relaxed single-variable system whose input u and output y are

related by
where () = tan -1 (1m g(i,
f>(t -
This corol1ary sho\\

system, ifthe input is a si
y(t)
r)u(r) dr
STABILlTY CRlTERIA IN TERMS Of THE INPUT-OUTPUT DESCRIPTION
the proof. Consider
If
IX> [g(t)1 dt ::; k <
-,JI eh
for sorne k and if u(t) = sin wt, for
I ::::: 0,
00
then
y(t)---> \g(iw)\ sin (WI +8)
)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(,)[ sin wt cos w; - cos WI sin w;] d;
g(,)cosw,d;-coswt
It is clear that
Ir
g(,) sin w(t - ,) d; \::;
t'"
g(;)sinw;d;-
[g(;)[[sin w(t -
Hence, from (8-9) we concl ude that as

y(t)---> ~in wt
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)
g(,) sin w'"[(h
By definition, g(s) is the Laplace transform of g('); that is,
'fJ) and if u is of finite

Q.E.D.
system that is describ
which implies that

g(iw)=
r'" g(t)e-
Jo
iwr
dt =
f'" g(t) cos wt lit - i r'" g(t) sin
Jo
Since g(t) is real-valued, we have
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)
Hence (8-17) becomes

y(t)---> sin wt (Re gUro)) +cos wt (1m g(iro)) = Ig(iw)\ sin (wt +8)
lt u and output y are
where 8 = tan - 1 (Img(iw)(Re g(iro)).
Q.E.D.
This corollary shows that for a BIBO-stable, .linear time-invariant relaxed

system, ifthe input is a sinusoidal function, after the transient dies out, the output
394
STABILITY Of LINEAR SYSTEMS
is also a sinusoidal function. Furthermore, from this sinusoidal output, the

magnitude and phase of the transfer function at that frequency can be read out
directly. This fact is often used in practice to measure the transfer function of a
linear time-invariant relaxed system.
Linear time-invariant systems are often described by transfer functions;
hence it is useful to study the stability conditions in terms of transfer functions.
If g(s) is a proper rational function of s, the stability condition can be easily
stated in terms of g(s). If q(s) is not a rational function, then the situation is
much more complicated; see Reference S32.
is BIBO stable if and

entry of G,
If the Laplace tra

function matrix then
ofC(s).
'
Theorem 8-5
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).
If g(s) is a proper rational function, it can be expanded by partial fraction

expansion into a sum of finite number of terms of the form
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.
Corollary 8-2 and

and 8-4 if we consideJ
input-output pairo
8-3
Routh-Hurv
If the transfer functior

stability of the system
is irreducible, that is, ti
nator and numerator, t
nator of g(s). Hence I
system is determined b
polynomial is called a
parts. Hence a systen
irreducible transfer fun
a polynomial is Hurwi
are computed. Howev
computa tia n of the roo
the exact locations of
Therefore it is desirab
without solving for thl
method: the Routh-Hu
Consider the polyn(
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
inusoidal output, the

ilency can be read out
: transfer function of a
'Y transfer functions;

; of transfer functions.
mdition can be easily
, then the situation is
395
is BIBO stable if and only if there exists a finite number k such that, for every
entry of G,
[' \gJt)\ dt .:s;k< 00

If the Laplace transform of G(t), denoted by G(s), is a proper rational
functon matrix, then the BIBO stability condition can also be stated in terms
of G(s).
Theorem 8-5
oper rational function

e in the open left-half
~ real parts.
s plane excluding the

s plane is the left-half
:y of a system is inde
A relaxed multivariable system that is described by y(s) = G(s)u(s), where G(s)

is a proper rational-function matrix, is BIBO stable if and only if all'the poles
of every entry of G(s) have negative real parts.
I
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.
The pole of g(s) is on

.ble. This can also be
mcton, then u(s) = l/s.
t is ;t' - 1 [g(s )u(s)] =
is not BIBO stable.
.ults.
Routh-Hurwitz Criterion
If the transfer function of a system is a rational function of s, then the BIBO

stability of the system is completely determined by the poles of g(s). If g(s)
is irreducible, that is, there is no nontrivial common factor between its denomi
nator and numerator, then the poles of g(s) are equal to the roots of the denomi
nator of g(s). Hence under the irreducibility of g(s), the BIBO stability of a
system is determined by the roots of a polynomial, the denominator of g(s). A
polynomial is called a Hurwitz polynomial if all of its roots have negative real
parts. Hence a system is BIBO stable if and only if the denominator of its
irreducible transfer function is a Hurwitz polynomial (Theorem 8-4). Whether
a polynomial is Hurwitz or not can be readily determined once all of its roots
are computed. However, if the degree of the polynomial is three or higher, the
computation of the roots is not a simple task. Furthermore, the knowledge of
the exact locations of the roots is not needed in determining the stability.
Therefore it is desirable to have sorne method of determining the stability
without solving for the roots. In this section we shall introduce one such
method: the Routh-Hurwitz criterion.
Consider the polynomial
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
n (s +ctdn(s +f3j +iy)(s +f3j - iy) ..

k
= ao
(s +ctd
(S2
+ 2f3 jS + f3; +yJ)
(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)
is not a Hurwitz polynomial.

Consider the polynomial D(s) given in (S-19).
polynomials
Do(s) =aos n +azsn - Z + ...

D1(s) = als n - 1 +a3s"- 3 +
We form the fol!owing

(S-21a)
(S-21b)
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:
Since the four numbe

conclude from Theoren
In order to ohtain o
These processes can b{
an even number and
with n' ~ n/2,
Note that the number

is even; they are equa
shown in Table S-1.
two rows are just the
remaining rows are 01
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.
This theorem will be proved in Section S-5.

illustrate its application.
Sll- 2
a\il
a\2)
Sn-3
ab"l
a\3)
Here we give an example to
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:
Ir n is odd, the first two ro

the pattern is identical.
'
'
.. .
.__
..
. _ .. _.J_ ..__.
.._ ... __,
_. _.. .
ients of the factors in

at if D(s) is a Hurwitz
)sitive. Hence given a
oejJicients are negative
The condition that al!
;ary condition for the
: coefficients may still
nial with positive co
l-4i)
:
form the fol!owing
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
Since the four numbers 1X=~,1X2=~'(.(3=~' and 1X 4

are all positive, we
conclude from Theorem 8-6 that the polynomial D(s) is a Hurwitz polynomial. I
In order to ohtain IX , 1X 2 , ... ,IX", a series oflong division must be performed.
These processes can be carried out in tabular formo First we assume that n is
an even number and rename the coefficients of Do(s) and D(s) in (8-21) as,
with n' ~ n/2,
Do(s) = abO)s"
+a~Ols"- 2
D(s)=a~)s"-
(8-21a)
(8-21b)
) and D(s) consists of

part of D(s) and D(s)
of Do(s) is always one
'D(s) in the following
+ ...
+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
I if and only if the n
397
5"-3
ab2 )
a~3)
a\3)
aff)
"'-1
----
a(Z)
a~Z)
a\Z)
a~9) :
"'-1
a(3)
n'- 2
:
I
'e give an exampIe to
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
Using this equation

a\n- 2) > O. Similarly
Table 8-1 are all po
i = 1, 2, ... ,17, then al
the theorem.
In forming the R
zero appears, we ma)
is not a Hurwitz po]
Hurwitz polynomial,
coefficients are positi
if n-k is even
Example 2
if n-k is odd
Consider the polynor
Then it is straightforward to verify, by using (8-25) and (8-26),
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
Consider the polynor

form
a~)ja~-)s =
Theorem 8-7 (Routh-Hurwitz criterion)

The polynomial D(s) of degree n in (8-19) 1s a Hurwitz polynmnial if and onl:;
if the n terms ag), i = 1, 2, ... , n, in the first column of Table 8-1 are aH positive
or if and only if aH the coefficients aj) in Table 8-1 are positive.
A zero appears in the

Example 4
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 -
A negative number ap;

polynomial.
We note that in co
- are not ffecid if a ro'
the cbmputation of Ta
(8-26)
le Routh table can be

ents will decrease by
'able 8-1. The (/. are
le first column of the
~2). In order to do so,
als, for k = 0, 1, ... , n,
n- 2k'
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) >
In forming the Routh table for a polynomial, if a negative number or a

zero appears, we may stop the computation and conclude that the polynomial
is not a Hurwitz polynomial. However, to conclude that a polynomial is a
Hurwitz polynomial, we must complete the table and then check whether aH
eoefficients are positive.
Example 2
Consider the polynomial in Example 1. We form
1
2
\-26),
6 1
4
4
1
3.5
1
AH the coefficients are positive; hence the polynomial is Hurwitz.

(8-27)
e haveD,,(s)/D,,-(s)=
into the form in (8-22).
:d from the eoeffieients
Example 3
Consider the polynomial D(5) = 35 7 +25 6 + 25 5 + 54 + 35 3 + 52 + 1.55 + 1.

form
3
1.5
1 1
1 O
== ag-1l/ag), i = 1,2, ... ,

l, 2, . _. , n: Hence the
:ro 8-6. To show the
Iy if aH the coefficients
To show the necessary
We
A zero appears in the table; henee the polynomial is not Hurwitz.

,olynomial if and only
ble 8-1 are aH positive
ositive.
Example 4
Consider 25 4 +25 3
+S2
+35 +2. We have
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
where G(t, r) ~ C(t),

E. What we have d
applied here. We s
(see Section 2-8).
Example 5
Consider D(s) = 2s 4
+ 5s 3 + 5s 2 + 2s + 1. We forrn the following table:

2 5
5 2
21 5
17
Theorem 8-8
(after the multiplication of 5)

(after the multiplication of 21)
The zero-state respol

if there exists a finite
5
There are four positive numbers in the first column (excluding the first number);
hence D(s) is a Hurwitz polynomial.
for any t o and for all
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.
The norm used in

of time, 11 u(t)11 can b(
sponding to these dif
8-4 Stability of Linear Dynamical Equations
has different values.

be used.
Next we study th
we study the responsl
Time-varying case.
namical equation
Consider the n-dimensional linear time-varying dy
x=
E:
A(t)x
y =C(t)x
+ B(t)u
(8-28a)
(8-28b)
due to any initial st~

x(to) = X o is given by
where x is the n x 1 state vector, u is the p x 1 input vector, y is the q x 1 output

vector; and A, B, and C are n x n, n x p, and q x n matrices, respectively. It is
assumed that every entry of A, B, and C are continuous functions of t in ( - 00,(0).
No assumption as to boundedness of these entries is made. We have assumed
that there is no direct transmission part in (8-28) (that is, E = O) because it does
not play any role in the stability study.
m ust introduce diffen

of equilibrium state.
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
x(t)=$(t; t o, x o, u)=$(t; t o, xo,O)+$(t; t o, O, u)
Hence it is very convenient to study the stabilities of the zero-input response

and of the zero-state .response separately. Combining these results, we shall
immediately obtain the stability properties of the entire dynamical equation.
First we consider the stability of the zero-state response. The response of
E with Oas the initial state at time t o is given by . . .
y(t) =
l'
l'
C(t)<Il(t, r)B(r)u(r) d'l:
lO
Definition 8-2
Xe
of a dynam
for all t 2. too

We see fram this (
and if no input is appli
Hence at any equili
equilibrium state of x
definition, an equilibr
xe =
lO
It is clear that the B:
G(t,r)u(r)dr
for all t 2. too Clearly,
STABILITY OF LINEAR DYNAMICAL EQUATIONS
401
where G(t, T) ~ C(t)$(t, T)B(T) is, by definition, the impulse-response matrix. of

E. What we have discussed in Section 8-2 (in particular, Theorem 8-2) can be
applied here. We shall rephrase Theorem 8-2 by using the notion of norm
(see Section 2-8).
,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
ling thefirst number);
IIC(t)$(t, T)B(T)II dT sk<
(8-29)
00
for any t o and for all t ;::: to.
:able can also be used

plane. If agl 1= 0, i =
plane is equal to the
example, References
les complicated. See
}ic in Reference S239.
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)
" Y is the q x 1 output

es, respectively. It is
;tions of t in ( - 00, 00).
e. We have assumed
1<:=0) becauseit does
Li
(Li
Ilqt)Q)(t, T)B(T)II
due to any initial state. The response of

x(to) = Xo is given by
x = A(t)x(t) due to the
initial state
x(t) ~ cjJ(t; to, xo, O) = $(t, to)iX o

It is c1ear that the BIBO stability can no longer be applied here. Hence we
must introduce different kinds of stability. Before doing so we need the concept
of equilibrium state.
Definition 8-2
; be decomposed into
t o, O, u)
zero-input response
hese results, we shall
ynamical equation.
nse. The response of
A state X e of a dynamical equation is said to be an equilibrium state at t o if and

only if
for all t ;::: to.
We see fram this definition that if a trajectory reaches an equilibrium state

and ifno input is applied, the trajectory will stay .at the equilibrium state forever.
Hence at any equilibrium state, xe(t) = O for .all t;::: too Consequently, an
equilibrium state of x =A(t)x is a solution of A(t)x';"O for aH t ::::(0' Or fram
definition, an equilibrium state of x = A(t)x is a solution of
or
for all t ;::: to. Clearly, the zera state, O, is always an equilibrium state ofx = A(t)x.
402
We shall now define the stability of an equilibrium state in terms of the

Definition 8-3
An equilibrium state X e is said to be stable in the sense of Lyapunov at t o if and
only if for every ; > 0, there exists a positive number Dwhich depends on ; and
t o such that if Ilx o- xell ~ D, then
1I<I(t; to, xo,O) - xell ~;

for aH teto. It is said to be uniformly stable i.s.L. (in the sense of Lyapunov)
over [t o, (0) if and only if for every ; > 0, there exists a positive Dwhich depends
on ; but not on t o such that if Ilxo- xell ~ D, then
1I<I(t; ti> xo,O) - xell ~6

for any t 1 e t o and for all t
e t l'
If an equilibrium state is uniformly stable i.s.L., then it is stable i.s.L. How

ever, the converse may not be true. For example, the zero state of
Definition 8-4
An equilibrium stat<
i.S.L. at t o and if eve
t --> oo. More precise
any e> 0, there exist
for all t :2:t l +T lf

pendent of ti> the eql
over [to,co).
The concept of a
that the stabilities e
because we do not kl
should be chosen. 1
property, and y can
rium state of a linear
in the large.
Theorem 8-9
x(t) = (6t sin t - 2t)x(t)
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.
for all t :2: too
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
Its equilibrium states are the solutions of Xl =

yield
and X2 =0 with u{t) =0 which
k=O, 1, 2, ...
The equilibrium states [kn O]', k = 0,

The equilibrium states [kn O]', k =
i.s.L. (Why?)
2, 4,
1, 3,
, are uniformly
stable i.s. L.
, however, are not stable
Figure 8-4
Asympto
403
Definition 8-4
state in terms of the
An equilibriurn state X e is said to be asymptotically stable at t o if it is stable

i.s.L. at t o and if every rnotion starting sufficiently near Xe converges to Xe as
t-"oo. More precisely, there is sorne y>O such that ifllx(tl)-xell.=s;y, then for
any e> 0, there exists a positive T which depends on e, y, and ti such that
Lyapunov at t o if and
jch depends on f. and
11P(t; t1> x(t 1), O) -xell.::;e

for aH t c. t 1 + T. If it is unifonnly stable Ls.L. over [to, 00) and if T is inde
pendent of ti' the equilibriurn state is said to be uniformly asymptotically stable
over [to, 00).
e sense of Lyapunov)
itive j which depends
The concept of asyrnptotic stability is illustrated in Figure 8-4. We see

that the stabilities defined in Definitions 8-3 and 8-4 are local properties,
because we do not know how srnall o in Definition 8-3 and y in Definition 8-4
should be chosen. However, for linear systerns, because of the hornogeneity
property, j and y can be extended to the entire state space. Hence, if an equilib
riurn state of a linear equation is stable at all, it will be globally stable or stable
in the large.
is stable i.s.L. How

) state of
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
ence S20G. Roughly

)Use due to any initial
y frorn X e. F or tirne
and uniforrn stability.
1\ I>(t, to)11 sk <

for aH t c. too
00
Ir k is independent of t o, it is uniforrnly stable Ls.L.
Proof
ge 127). The applica
Sujficiency: Let
Xe
be an equilibriurn state of x = A(t)x; that is

Xe =
I>(t, tO)x e
for all t c. t o
'o with u(t)=O which
t
miformly stable Ls.L.
.ivever, are not stable
Figure 8-4
.J
Asymptotic stability.
+T
404
STABILlTY OF LINEAR SYSTEMS
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
lt follows from (2-90) that
x(t) - xell ~ 11 cI>(t, to)llllx o - xell ~ kllxo - xell

for all t. Henee, for any e, if we ehoose b = e/k, then Ilx o - xell ~b implies that
Ilx(t)-xell~e
forallt~to
Necessity: If Xe is stable i.s.L., then cI>(t, to) is bounded. We prove this by

eontradietion. Suppose that Xe is stable i.s.L. and that cI>(t, to) is not bounded,
then for sorne to, say to, at least one element of cI>(t, to), say 4>ij(t, to), becomes
arbitrarily large as t-4 oo. Let us choose Xo at time to such that all the eom
ponents of (x o - x e) are zero exeept the jth component, whieh is egual to cx.
Then the ith component of (x(t) -xe) is egual to 4>Jt, to)'cx, whieh becomes
arbitrarily large as t-4 00, no matter how small cx is. Hence Xe is not stable i.s.L.
This is a eontradiction. Hence, if Xe is stable i.s.L., then cI>(t, t o) is bounded.
Q.E.D.
for allllxoll~'l', and

11 cI>(to + T, to)xoll = 11 e
that
This eondition and 11

11 cI>(t, to)11 ~ k 3
fOl
11 cI>(t, to)11 = 11 cI>(t, t o +
k3
~2
fOl
11 cI>(t, to)11 ~ 11 cI>(t, t o +
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
The zero state of x = A(t)x is asymptotically stable at t o if and only if

11 cI>(t, to)11 ~ k(t o) < 00 and 11 cI>(t, t o)ll-4 as t -400. The zero state is uniformly
asymptotieally stable over [0, (0) if and only if there exist positive numbers
k1 and k 2 sueh that
for any t o ~
11 <1>(t, t o)\ 1 ~ k e- k2 (1-lo)
(8-31 )
~22
for
and so forth, as showl

let k 1 = 2k 3 . Then fr,
for any t o and for a1l1

If the zero state of
initial state, the zero
for linear eguations, j
also said to be expone
We have diseussed
stability of the zero-sl
and for all t ~ too
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(t, t 0)11
STABILITY Of LINEAR DYNAMICAL EQUATIONS
Lyapunov. Consequently, there exists a finite number k 3 such that 11 (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)
Ilx(to +T)II = 11(to +T, to)xoll:::;;e

(8-32)
for aH Ilxoll sy, and for any to. Now choose an Xo such that Ilxoll ='1 and
Il(to +T, to)xoll = 11(to +T,to)lll!xoll and choose f. ='1/2; then (8-32) implies
-xell
- xe \\:::;;6 implies that
that
1I (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
This condition and 11 (t, to)11 s k3 imply that
I! (t, to)!1 :::;; k 3
for aH t in [to, t o + T)
\\(t, t o)\\ = \I(t, t o +T)(t o +T, t o)\\ :::;;1\(t, t o +T)I\\\(to +T, t o)\\
:::;;23
for al1 t in [to +T,to +2T)
1I(t, to)11 :::;;II (t, t o +2T)IIII(to +2T, t o +T)IIII(to +T, to)11

:::;;~;
Jm state. lf the zero

ly equilibrium state of
foralltin[t o +2T,t o +3T)
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
ate different from the

,tial state, the response
f)x is asymptotical1y
IICI>(t, to)11 :::;;ke- k, (r-ro)

for any t o and for al1
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
Necessity: lf the zero

,table in the sense of
Figure 8-5
+T
[ 31
----
i o + 2T
lO
k 18
3
.
-
+ 3T
II <1>(t, to)11 bounded by an exponentialIy decreasing function.
::==::::--=.:.~.::-:::".:.-===::::::.=====-===:-._-~--'.'--_
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,
e)B(e)11 de :5; k < 00
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.
of the zero-state resp

study the entire respl
Definition 8-5
A linear dynamical e
if and only if for any
as all the state variab
We see that the e
BIBO stability; they
of all state variables;
also for any initial ~
function properly, bel
and the system will 1
every system is requil
Theorem 8 -13
Theorem 8-11
A system that is descr

stable if and only ifC
Consider the dynamical equation E given in Equation (8-28). If the matrices

B and C are bounded on (- 00,00), then the uniformly asymptotic stability
I
of the zero state implies the BIBO stability of the zero-state response.
for
Proof
Proof
This theorem fol1ows directly from the fact that
The response of (8-28;
an~ t o
JIIC(t)$(t, e)B(e)11 de:5; JIIC(t)IIII<l>(t, e)IIII B(e)11 de ~klk2 JII$(t, e)il de

where
IIB(t)11 ~kl' 11C(t)11 ~k2 for all t.
Q.E.D.
The converse problem of Theorem 8-11~that of determining the COD

ditions under which the BIBO stability of the zero-state response implies the
asymptotic stability of the zero state-is much more difficult. In order to solve
this problem, the concepts of uniform controllability and uniform observ
ability (Definitions 5-4 and 5-8) are needed. We state only the result; its proof
can be found in Reference 102.
*Theorem 8-12
Consider the dynamical equation E given in (8-28). If the matrices A, B, and C .

are bounded on (- 00,00) and if E is uniformly controllable and uniformly
observable, then th zero state of E is asymptotically stable (under the zero- .
input respOnse) if and only if its zero-state response is BIBO s t a b l e . 1
We have studied the stability of the zero-input response and the stability
and for all
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~
tained for linear, time

because aH the conditi
are very difficult, if TI(
time-invatiant .dynam
matrix is, however, no
the matrix A.
STABILlTY Of LINEAR OYNAMICAL EQUATIONS
:he entire dynamical

between the stability
mt response.
,-state response of E
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.
lction given in Figure

Hily bounded. Con
tegrable. Hence, the
t imply nor is implied
~tion that approaches
: asymptotic stability
asymptotically stable,
plied by (8-31); hence
c stability may imply
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
8-28). If the matrices

y asymptotic stability
tate response.
[II<Il(t, ,)B(,)II d, sk<
el)
for any t o and for aH t '2: to.

Proof
The response of (8-28a) is
r <Il(t, ,)B(,)u(,) d,
Jo
t
',kz
JI1cD(t, ')11 do
Q.E.D.
determining the con

response implies the
cult. In order to solve
and uniforrn observ
lly the result; its proof
~
Le matrices A, B, and C
oHable and uniformly
table (under the zero
BO stable.
lonse and the stability
x(t) =<Il(t, to)x o +
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
Consider the n-dimensional linear time-invariant dynamical equation
FE:
x=Ax +Bu
y=Cx
(S-33a)
(S-33b)
where A, B, C are n x n, n x p, q x n real constant matrices, respectively. As in

the time-varying case, we study first the zero-state response and the zero-input
response and then the entire response. The zero-state response of FE is charac
terized by
Its equilibrium states

words, every point in tl
of the matrix are - 1, (
value O, which has a ze:
Hence every equilibriu
G(s) =C(sI - A)-lB

From Theorem 8-5, the zero-state response oj FE is BIBO stable if and only if al!
the poles ofevery entry oj(;(s) have negative real parts. The zero-input response
of FE is governed by x = Ax or x(t) =eA1xO' Recall that an equilibrium state
of x = Ax is a solution of x = O or Ax = O, and has the property X e = eAtx e for
all t ~O. Note also that stability implies uniform stability in the time-invariant
case.
Theorem 8-14
Every equilibrium state of x = Ax is stable in the sense of Lyapunov if and only

if all the eigenvalues of A have nonpositive (negative or zero) real parts and those
with zero real parts are distinct roots of the minimal polynomial of A.3
Consider
It has eigenvalues -1
eigenvalue O is not a
equilibrium states of th
Theorem 8 -1 5
The zero state of x = Ax

of A have negative real
Proof
Let X e be an equilibrium state of x = Ax. Then x(t) - X e = eAI(x o - xJ Hence

every equilibrium state is stable .S.L. if and only if there is a constant k such
that IleAtll.::; k < <X) for aH t ~ O. Let P be the nonsingular matrix such that
A= PAp-l and Ais in the lordan formo Since e AI = PeAlp-l, then
IleAtll.::;llplllleAtllllp-lll
Consequen,tly, if I\eAtll is bounqed, so is IleAtll. Conversely, from the equation
e At =p-1eAtp, we see that if IleAtl1 is bounded, so is IIeAtll. f!ence W'Y conclude
that every equili~rium state is stable .s.L. if and only if IIeAtl1 i~ bounded on
[O, 00). Now IleAtl1 is bounded if and only if every entry of e A1 is bounded.
Since is in the lordan form, every entry of e At is of the form leajt +irojt, where
el. j + iro j is an eigenval ue of A(see Section 2-7). Ir el. j is negative, it is easy to see
that leajt +irojt is bounded on [O, 00) for any integer k. Ir el.j = O, the function
le irojt is bounded if and only if k = O-that is, the order of the lordan block
associated with the eigenvalue with el. j = O is 1.
Q.E. D.
Example 2
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
In order for t he zero statl

ness of JleAtll, it is requi
that IleAtll~O as t---+oo.
conclude that IleAtl1 ~O
consequently of A have
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
STABILITY Of LINEAR DYNAMICAL EQUATIONS
409
llamical equation
(8-33a)
(S-33b)
es,respectively. As in
llse and the zero-input
;ponse of FE is charac
, stable if and only if alI

he zero-input response
lt an equilibrium state
r
property X e = e Al x e 101'
:y in the time-invariant
fLyapunov ifand only

:ro) real parts and those
lynomial of A. 3
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 ).
sely, from the equation

'11. I-!ence w~ conc1ude
if Ile A '\ I i~ bounded on
Iltry of eA' is bounded.
e form tkeajl +iOJj" where
legative, it is easy to see
If aj=O, the function
.er of the J ordan block
Q.E.D.
very lordan blocks associated

does not imply that the eigen-
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
det (s~ -A) e [Adj (sI -A)JB
410
every pole of (;(s) is an eigenvalue of A (the converse is not true). Consequently,

if the zero state of FE is asymptotically stable, the zero-state response of FE
will also be BIBO stable. (This fact can also be deduced directly from Theorem
8-11.) Conversely, the BIBO stability of the zero-state response in general
does not imply the asymptotic stability of the zero state, because the zero
state response is determined by the transfer function, which, however, describes
only the controllable and observable part of a dynamical equation.
Example 4
Consider a system with the following dynamical equation description:
x=G _~Jx+[~Ju
y=[1
Theorem 8-17
If a linear time-invarian
then the following state
lJx
Its transfer function is

g(s) =
[1
lJ
[S-1
-1
1. The dynamical eq ua
J-l [OJ =s+11
s+1
Hence the zero-state response of the dynamical equation is BIBO stable;

however, the zero state is not asymptotically stable, because there is a positive
eigenvalue.
I
Theorem 8-16
Let
x= [~e ~;2 Jx +[~eJ u
(8-34a)
Ce J x
(8-34b)
y = [ Ce
It is clear that tol

ever may not imply 1
the BIBO stability.
asymptotic stability
asymptol ic stability is
If a linear time-in
able, then the charac
?olynOmial 2f G(s) (1
IS a pole of G(s), and (
we have the fOllowing
be an equivalent dynamical equation of the dynamical equation in (8-33)

with {Ae, Be} controllable. Then the dynamical equation in (8-33) is totally
stable if and only if al! the eigenvalues of Ae have negative real parts and all the
eigenvalues of Al' have negative or zero real parts and those with zero real parts
are distinct roots of the minimal polynomial of Al"
2.
3.
4.
5.
The zero-state respol

The zero state of FE j
AH the poles of the tI
AH the eigenval ues of
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:
The application of the Laplace transform to (8-34a) yields
e - A~? J- \(0) +[Si -Ae - A~2 -1[BeJu(s)

[Si -A
O
si - Al'
O
si - Ac
O
[(SI -AJ-l
...~ -IJX(O) +[(Sl - AJBe
Ju(s)
O
l$.-Ac)
O
where M= (si- A
e)-IA I2 (sl -e)-I. From this eql,lation, we may conclude
x(s) =
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.
The characteristic polynom

det [,1.1- A
Hence the eigenvalues of A
ofthe equation is neither asy
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
uation is BIEO stable;
::cause there is a positive
(8-34a)
(8-34b)
mical equation in (8-33)

.lation in (8-33) is totally
ltive real parts and all the
those with zero real parts
1. The dynamical equation is totally stable.

2. The zero-state response of FE is BIEO stable.
3. The zero state of FE is asymptotically stable (under the zero-input response).
4. AH the poles of the transfer function matrix of FE have negative real parts.
5. AH the eigenvalues of the matrix A of FE have negative real parts.
I
A system is said to be completely or faithfully characterized by its transfer

function matrix if the dynamical-equation description of the system is con
troHable and observable. We see from Theorem 8-17 that if a system is com
pletely characterized by its transferjunction matrix, then asymptotic stability of
the system can be determined from its transfer-function matrix alane with no
need of considering the dynamical-equation description of the system.
A remark is in order concerning the stability of x = A(t)x. If the matrix A
is independent of t and if aH the eigenvalues of A have negative real parts, then
the zero state is asymptotically stable. Hence one might be tempted to suggest
that iffor each l, aH the eigenvalues of A(l) have negative real parts, then the zero
state of x= A(t)x is asymptotically stable. This is not so, as can be seen from
the following example
Example 5
Consider the linear time-varying equation
. [-1 J
2l
ields
A~.? llrBc1 u(s)
sl-AcJ
Lo
_~c)-lBc1
u(s)
quation, we may conclude

unded u if and only if the
) is y. This establishesthe
Q.E.O.
x=
e
O -1 x
The characteristic polynomial of the matrix A at each t i.s given by
+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
transition matrix of the equation is
where the ,l;'s are the
lI>(t'O)=[e~'
, '\' 1-
)~min
Xii
i= 1
(as in Problem 4-1 or by direct verification), whose norm tends to infinity

as t----> OCJ.
*8-5
The fact that p- I
=P
Lyapunov Theorem
The asymptotic stability of x = Ax can be determined by first computing the

characteristic polynomial of A and then applying the Routh-Hurwitz criterion.
If al1 the roots of the characteristic polynomial have negative real parts, then
the zero state of x = Ax is asymptotical1y stable. There is one more method of
checking the asymptotic stability of x = Ax without computing explicitly the
eigenvalues of A. We shal1 discuss such a method in this section and then
apply it to establish the Routh-Hurwitz criterion.
Before proceeding, we need the concept of positive definite and positive
semidefinite matrices. An n x n matrix M with elements in the field of complex
numbers is said to be a hermitian matrix if M* = M, where M* is the complex
conjugate transpose of M. If M is a real matrix, M is said to be symmetric.
The matrix M can be considered as an operator that maps (en, 1[:) into itself.
It is shown in Theorem E-l that al1 the eigenvalues of a hermitian matrix are
real, and that there exists a nonsinguIar matrix P, cal1ed a unitary matrix, such
that p-I = P* and M =PMP*, where M is a diagonal matrix with eigenvalues
on the diagonal (Theorem E-4). We shall use this fact to establish the following
theorem.
Hence, the inequality

Definition 8-6
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
Let M be a hermitian matrix and let Amin and

eigenvalues of M, respectively. Then
}'max
be the smallest and largest

4
AminllxW .::;;x*Mx '::;;AmxllxW
The prillcipal millors of the
(8-35)
for any x in the n-dimensional complex vector space 1[:/1, where
IlxW ~ (x, x) ~ x*x =
Ixd2
m,
are m,,_ m22_ m))_ det [
m2
i= I
and
Xi
is the ith component of x.
minors whose diagonal ele:

millors of M are m". det [
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
x*Mx = x*PMP*x = x*Mx =
i= 1
AJX'2
last k columns and the last

It is shown in Reference 39
the principal minors are pe

are nonnegative, then all
Problem 8-32b.
" Ir N is an upper triangular
able in LlNPACK and lB~
LYAPUNOV THEOREM
413
where the },'s are the eigenval ues of M. It fol1ows that
1-1 2::S;X *Mx -__x*Mx -
A min ~
1... Xi
"
(8-36)
max
i= 1
i=l
norm tends to infinity
'\-\2 ::s; A'
1... A X
__
The fact that p-l = P* implies that
IlxW=x*x=x*x=
I IX1
i= 1
Q.E.D.
Hence, the inequality (8-36) implies (8-35).

by first computing the
outh-Hurwitz criterion.
egative real parts, then
: is one more method of
)mputing explicitly the
1 this section and then
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
'e definite and positive

s in the field of complex
here M* is the complex
IS said to be symmetric.
maps (C", iC) into itself.
a hermitian matrix are
j a unitary matrix, such
matrix with eigenvalues
) establish the following
Theorem 8-19
A hermitian matrix M is positive definite (positive semidefinite) if and only if
any one ofthe following conditions holds:
1. Al! the eigenvalues of M are positive (nonnegative).

2. All the leading principal minors 4 of M are positive (al1 the principal minors
of M are nonnegative).s
3. There exists a nonsingular matrix N (a singular matrix N) such that M =
N*N. 6
the smallest and largest

4The principal minor, of lhe malrix
M
" where
j::::
;::~ ;::~lJ'
Lm.31
m32
are m". mn. m.'.,. det [m\l

m21
m!2'J'], det[m
m22
11
m31
nt:33
m,,], det[m n
m,,], and detM. that s, the
m.:n
m33
m:n
minors whose diagonal elements are also diagonal elements of the matrix. The leading principal
minors or M are m", det [m"

m21
~t
mIl], and det M, thal s, the minors obtained by deleling the

m22
last k columns and lhe last k rows, for k = 2, 1. and O.

5 lt is shown in Referenee 39 that if a11 the leading principal minors of a malrix are positive, then all
the principal minors are positive. However, it is nol true thal ir all the leading.prncipal minors
are nonnegative, then 'all the principal minors are nonnegative. For a counterexample, lry
Problem 8-32b.
.
(, Ir N is an upper lriangular malrix, i~ is called lhe Cholesk)' decomposilion. Subroulines arc avail
able in LlNPACK and IBMScientific Subroutine Package to carry out this decomposition.
P be the nonsingular
~ is a diagonal matrix
,r x = P-1x = P*x, then
,
1
414
Proof
Condition 1 follows directly from Theorem 8-18. A proof of condition 2 can

be found, for example, in References 5 and 39. For a proof of condition 3,
see Problem 8-24.
Q.E. O.
will be bowl shaped,

values taken by Val<
value of V will incre
trajectory. Taking ti
x = Ax, we obtain
d
With these preliminaries, we are ready to introduce the Lyapunov theorem

and its extension. They will be used to prove the Routh-Hurwitz criterion.
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
has a unique hermitian solution M and M is positive definite.
(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
The implication of Theorem 8-20 and Corollary 8-20 is that if A is

asymptotically stable and if N is positive definite or positive semidefinite,
then the solution M of (8-37) must be positive definite. However, it does not
say that if A is asymptotically stable and if M is positive definite, then the
matrix N computed from (8-37) is positive definite or positive semidefinite.
Before proving the Lyapunov theorem, we make a few comments. Since
Theorem 8-20 holds for any positive definite hermitian matrix N, the matrix
N in (8-37) is often chosen to be a unit matrix. Since M is a hermitian matrix,
there are n 2 unknown numbers in M to be solved. If M is a real symmetric
matrix there are n(n + 1)/2 unknown numbers in M to be solved. Hence the
matrix equation (8-37) actually consists of n 2 linear algebraic equations. To
apply Theorem 8-20, we first solve these n 2 equations for M, and then check
whether or not M is positive definite. This is not an easy task. Hence Theorem
8-20 and its corollary are generally not used in determining the stability of
x = Ax. However, they are very important in the stability study of nonlinear
time-varying systems by using the so-called second method of Lyapunov.
Furthermore, we shall use it to prove the Routh-Hurwitz criterion.
We give now a physical interpretation of the Lyapunov theorem. If the
hermitian matrix M is positive definite, the plot of V(x)
V(x) ~ x*Mx
dt
=x*
Theorem 8-20 (Lyapunov theorem)
A*M+MA= -N
V(x(t =
(8-38)
along any trajectory o
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
:oof of condition 2 can

l proof of condition 3,
Q.E.D.
the Lyapunov theorem
:h-Hurwitz criterion.
:t V(x(t)) =
{~t (x*(t)Mx(t)) = (:t x*(t)) Mx(t) +x*(t)M (:t X(t))
=x*(t)A*Mx(t) +x*(t)MAx(t)=x*(t)(A*M +MA)x(t)

= -x*(t)Nx(t)
(8-39)
where N ~ - (A*M + MA). This equation gives the rate of change of V(x)
along any trajectory of x= Ax. Now if N is positive definite, the function
- x*(t)Nx(t) is always negative. This implies that V(x(t)) decreases monotoni
cally with time along any trajectory of x = Ax; hence V(x(t)) will eventually
approach zero as t -> oo. Now since V(x) is positive definite, we have V(x) = O
only at x = O; hence we conclude that if we can find positive definite matrices
M and N that are related by (8-37), then every trajectory ofi = Ax will approach
the zero state as t -> oo. The function V(x) is caBed a Lyapunov function of
i = Ax. A Lyapunov function can be considered as a generalization of the
concept of distance or energy. Ir the "distance" of the state along any trajectory
of i = Ax decreases with time, then x(t) must tend to O as t-> oo.
ivalently, the zero state

. given positive definite
(8-37)
efinite.
Livalently, the zero state

ny given positive semi
observable, the matrix
Sufficienc y: Consider V(x) = x*Mx. Then we have

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
)
along any trajectory of i = Ax. From Theorem 8-18, we have

V
V
x*Nx
x*Mx -
(AN)min
(AM)max
-=---<----
V(x)
Figure 8-6
(8-38)
\
!
\
A Lyapunov function V(x).
(8-40)
416
STARII.ITY (lF I.INEAR SYSTEMS
t -" oo. This shows t:

similarly proved as ir
Theorem 8-20 ar
asymptotic stability
right and are basic
provide a simple pro.
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) =
with real coefficients .
is asymptotically stable, then all the eigenvalues of A have negative real parts.
Consequently, for any N, there exists a unique matrix M satisfying
A*M +MA=-N
and compute
and M can be expressed as

M
L'" eA 'INe
Do(s)
D (s) =CI.:
A'
dt
(8-41 )
(see Appendix F). Now we show that if N is positive definite, so is M. Let H

be a nonsingular matrix such that N = H*H (Theorem 8-19). Consider
xMxo =
L'" X eA *'H*HeA'xo dt = L'" IIHeAtxoW dt
(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
In order to establish Corollary 8-20, we show that if N is positive semidefinite

and if {A, N} is observable, then x*(t)Nx(t) cannot be identically zero along any
nontrivial trajectory ofx = Ax (any solution due to any nonzero initial state xo)'
First we use Theorem 8-19 to write N as N = H*H. Then the observability
of {A, N} implies the observability of {A, H} (Problem 5-36). Consider
The polynomial D(s) i

CI., Cl. z , .. . , Cl. are posit
x*(l)Nx\t) ,.~ x(cH"ik';x", lilleA(xoi2
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) ~
The assumption Cl. i =/=-1

factor between Do(s)
between D(s) and D(s;
Cohsider the block
function from u to y is !
equivalently, from the
7
This rollows Reference 90.
.. =~
LYAPUNOV THEOREM
the largest eigenvalue

3.t the matrices M and
O. If we define
s that V(t) s: e-'" V(O).

ero on every trajectory
je that the response of
lis proves that the zero
le zero state of X = Ax
ve negative real parts.
[ satisfying
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.
A proof of the Routh-Hurwitz criterion.
Consider the polynomial

ao>O
with real coefficients a, i =0, 1,2, ... , n. We form the polynomials

Do(s)=aos n +azs n - Z + ...
D 1(s)=a1 Sn - 1 +a3 Sn - 3 + ...
and compute
Do(s)
--=<XlS
D 1 (s)
(8-41 )
+-----------:-------
<XzS
+----------:-----
1
<X3 S + - - - - - - - - - -
:finite, so is M. Let H
8-19). Consider
+--
(842)
+-
ve have HeAlxo O for

) for all Xo =/= O, and M
orem.
Q.E. D.
is positive semidefinite
ntically zero along any
onzero initial state xo)
rhen the observability
)-36). Consider
For convenience, we shall restate the theorem here.

Theorem 8-6
The polynomial D(s) is a Hurwitz polynomial if and only if aH the n numbers
<X l' a z , ... , <XII are positive.
ProoF
First we assume that a11 the
function
"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
are different from zero. Consider the rationa!

1
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
This follows Reference 90.
418
Let the M rnatrix in (
Figure.8-7
A block diagram of g(s).
Then it is easy to veri
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
~:lj lxj l:J
It is clear that N is .
verify that {A, N} is o

the zero state of x =)
positive definite, or eq
the zero state of x = A
of A, or equivalently
words, D(s) is a Hurwi
are positive.
Consider now the
In other words, sorne
equal to zero. Suppo
8-1 are equal to zero,
factor. The cornmon
say fes). Then D(s) c
even function or an (
Hurwitz polynornial.
zero, we may replace
complete Table 8-1.
Ir sorne a i is negative, ~
by e) has a positive re
functions of its coeffi.1
positive or zero real p.
(8-43)
lJx
Irreducibility of (8-43) can be verified either by showing that it is controUable

and observable or by the fact that its dimension is equal to the degree of the
denorninator of g(s). Consequently, the characteristic polynomial of the
rnatrix A in (8A3) is ~qual to the denorninator of g(s) (Theorem 6-2). Now we
shall derive the condition. for the zero state of (8-43) to be asymptotically st.ble.
*8-6
Linear Tim
The stability concepts

'applicable'to the disCl
quite different. In thi
DISCRETE-TIME SYSTEMS
419
Let the M matrix in Corollary 8-20 be chosen as
M~ ~'
O
Ci. 1I _
O
O
O
O
Ci. z
J]
(8-44)
Then it is easy to verify that
A*M +MA=
(s)
h (s) can be written as
: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)
lt is clear that N is a positive semidefinite matrix. lt is straightforward to

verify that {A, N} is observable. Hence from Corollary 8-20 we conclude that
the zero state of x = Ax is asymptotically stable if and only if the matrix M is
positive definite, or equivalently, the n numbers Ci., Ci. z, ... , Ci. 1I are positive. Now
the zero state of x = Ax is asymptotically stable if and only if all the eigenvalues
of A, or equivalently all the roots of D(s), have negative real parts. In other
words, D(s) is a Hurwitz polynornial if and only if all the n numbers Ci., Ci. z , ... ,Ci. 1I
are positive.
Consider now the case in which not all the Ci.;'S are different from zero.
In other words, sorne of the coefficients in the first column of Table 8-1 are
eq ual to zero. Suppose abZ ) = O. If all the coefficients in the Sil - Z row of Table
8-1 are equal to zero, it irnplies that Do(s) and D(s) have at least one cornrnon
factor. The common factor is clearly either an even or an odd function of S,
say f(s). Then D(s) can be factored as f(s)i5(s). Since not all the roots of an
even function or an odd function can have negative real parts, D(s) is not a
Hurwitz polynomial. lf not all thecoefficients in the s"-Z row are equal to
zero, we may replace a\?l by a very small positive number ; and continue to
complete Table 8-1. In this case it can be seen that sorne Ci. will be negative.
If sorne Ci. is negative, at least one root of the modified D(s) (since abZ ) is replaced
by ) has a positive real part. Now the roots of a polynornial are continuous
functions of its coefficients. Hence, as --->O, at least one root of D(s) has a
positive or zero real part.
Q.E. D.
(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
Linear Time-Invariant Discrete-Time systems
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
Consider a relaxed linear time-invariant di serete-time system described by
Table 8.2 Discrete-tim
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
where bis a real or a complex number. lf Ibl < 1, then
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')
is the difference of its t

by the same procedu
define (Xi = agJ, i = 0,1.
Theorem 8-21
We shall prove thi:

discrete-time systenis.
ical equation
Otherwise, it diverges.
lf g(z) is irreducible, the poles of g(z) are equal to the roots ofits denominator.
lf the degree of the denominator is three or higher, the computation of the roots
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)
All the roots of D(z) i

numbers (Xi ~ ag), i = 1
k =0,1,2, ... <:> g(z) =fl[g(k)] =~
D(z) = aoz n +alz"- 1 + ... +an-1z +a
Cl~)
(8-46)
m=O
An absolutely integrable function is neither rrecessarily bounded nor necessarily"approaches zero as

as shown in Figure 8-2. An absolutely summerable sequence however is always bOllnded
and approaches zero as k-+ oo. Thus the stability problem in the discrete-time case is simpler
than the one in the continuing time case.
The concepts of equi

asymptotic stability ~
Theorem 8-14, every ~
only if aH the eigenvall
\vith magnitudes eq ua
Similar to Theorem 8
stable if and only if al1
Lyapunov theorem fOl
Theorem 8-22
All the eigenvalues of 1

positive definite herm
hermitian matrix N wi
(-+ 00
has a unique hermitiar
le system described by
(8-46)
Table 8.2 Discrete-time Stability Table
abo l
_) koa~,O)
ab )
1
there exists a finite h

:e {y(k)} is bounded, if
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)
(8-47) is similar to the

nsform of (8-46) yields
stable if and only if aH

llentIy, all the poles of
be readily proved by
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
,ots of its denominator.

mputation of the roots
d of eheeking whether
ele without eomputing
)f the Routh-Hurwitz
ao >0
(8-48)
Table 8-2. The first

the quotient of its last
tiplying ko on the first
lrder. The third row
+ 1) = Ax(k) + Bu(k)
y(k) =Cx(k)
The eoneepts of equilibrium state, stability in the sense of Lyapunov and

asymptotie stability are identical to the eontinuous-time case. Similar to
Theorem 8-14, every equilibrium state of x(k + 1) = Ax(k) is stable i.s.L. if and
only if all the eigenvalues of A have magnitudes equal to or less than 1 and those
with magnitudes equal to 1 are distinet roots af the minimal polynomial af A.
Similar to Theorem 8-15, the zero state of x(k + 1) = Ax(k) is asymptotical1y
stable if and only if all the eigenvalues of A have magnitudes less than l. The
Lyapunov theorem for the diserete-time case reads as:
ecessarily approaches zero as

however is always bounded
iiscrete-time case is simpler
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
This theorem can be proved by defining
V(x(k = x*(k)Mx(k)
and computing
~ V(x(k~ V(x(k
+ 1
- V(x(k = x*(k)A*MAx(k) - x*(k)Mx(k)

= x*(k)(A*MA - M)x(k)
= - x*(k)Nx(k)
cxi _
and is left as an exercise (Problem 8-36).

A Proof of Theorem 8-21 9
Figure S-8
Now we shall use Theorem 8-22 to prove Theorem 8-21.

0,1, ... , n.
and
Di(z) =aglz/l- i +a~)z/l-i- + ... +a~~ 1

k i =a~~Jag)
Di(z) =a~i~ Z"-i +a~i~_z"-i-l + ... +alilz +ag)
(S-49)
(S-50)
where D(z) is the reciprocal of D(z). These polynomials can be defined re

cursively by
D-I(Z)-ki-Di-(Z)
Di()
Z =-------
A block di
Define, for i =
i=I,2, ... ,n
(S-51 )
for i = 1,2, ... ,n. The

apply this diagram re
Do(z)/Do(z) shown in
diagram, there is a tot,
z- 1 as a state variable a
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
From Table 8-2, we have ag+ 1 ) =ag) -
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
(zD ( z) + k i _ 1 D;(z) ---'-=-a
IX
+ D;(z)) -'--
(ki_zD(z)
-k
-k
-kak,,_CJ./l_/CJ. -k
-kaCJ..!cx
-k
Simple manipulation among (8-51), (8-52), and (8-53) yields

Di 1(z)
b=
(8-54 )
(S-55 )
lXi
This state equation is ob1

the dimension of A; henc
F or this ma trix A, if we e
These recursive equations are valid for i = 1,2, ... ,n. Note that D/l(z) = D/l(z) =
a\'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 +----.:...--.:...-=..:.~,..:
This follows Rererence SIl!.
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
- x*(k)Mx(k)
Figure S-S
8-21.
Define, l'or i =
< = a~:~ Ja~)

+a~)
(S-49)
(8-50)
A block diagram of 15_ (z)/D_I(z).
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
tials can be defined re
!, ... , 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
thea)i), j =0,1, ... , n - i

iprocal ol' Di(z) can be
2, ... , n
+ 1) =
A=
-kok,
- k ok 2rx 2/rx
- k ok 3rx 3 /rx
-k 2k 3
- kok n rx" - /rx -kk,,_IY.,,_ Jrx 2 - k 2 k" _ rx" - /rx 3

-kIY.,j1Y. 2
- k2rx ,jrx 3
-kolY.,jlY.
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
M=diag{~, ~, ... ,~}
Note that Dn(z) = .5,,(z) =
al
()(2
()(n
then it is straightl'orward to veril'y
.kt- I)D(z)
A *MA'~ M = - N = - diag {~ , 0, O, ... , O}
) +k_ )(z)
ao
z)
:(z)
(S-56)
.'
By assumption, we haveao > O; hence N is positive seniidefinite. Furthermore

{A, N} is observable. Hence we conclude that aH
eigenvalues ol' A and, consequently, al1 rootsol' D(z)have magnitudes less than 1
il' and only il' Mis positive definite or, equivalently, ()(i > O, i = 1,2,. ., n. Q.E. D.
it has the property that
. ..-._..._-
-====:=====
_.~---------------~._ ~
===. . . . . -=.---= =----=-===
------=-_==::::_~.::_c==c::.:_.:.=:
8-7
--------------------l
r-------------~
In this chapter we in
and the stability i.s.L
being uniformly stab
distinetion between u
neeessary and suffieie
they ean hardly be e
not available.
For the time-inva

the transfer funetion (
a11 the eigenvalues of
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
whieh maps the left~ha

may transform a dise
using the bilinear trar
on diserete-time equati
...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
STABlLlTY OF LINEAR SYSTEMS
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
Prove parts ! ane

systems describable by r
8-11
8-12
Consider the dyn,
Figure P8-2
8-3 Consider a system with the transfer function g(s) that is not necessarily a rational
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
seconds will the output reach its steady state?

g(t}
Find all the equilibrium

Is it asymptotically stable'
stable?
o
Figure P8-4
+ t)
8-5
Is a system with the impulse response g(t) = 1/(1
8-6
Is a system with the transfer function g(s) = e-s/(s
8-14
Check the BIBO st
8-15
Find the rangcs of I
BIBO stable?
+ 1) BIBO stable?
Use the Routh-Hurwitz criterion to determine which of the rollowing polynomials

are Hurwitz polynomials.
8-7
a.
S5
+4s 4 + !Os 3 +2,,2 +5, +6
b. SS +S4 +2s 3 +2s 2 +5s +5

c. -2s4 -7s 3 -4s 2 -5s-10
is BIBO stable.
8-16
It is known that the
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
.a. aos 2 +as +a2

b. aos 3 +as2 +a 2s +a3
is BIBO stable. Can we e
Find the dynamical-ecjuation description of the network shown in Problem 8-2.

Find the equilibrium states of the equation. Is the equilibrium state stable in the sense of
Lyapunov? Is it asymptotically stable?
8-10
8-17 Showthatx=Osti
t --> OO.
PROBLEMS
nd a bounded input that
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?
igure P8-4. If the input

utput') After how many
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
Check the BlBO stability of a system with the transfer function

25 2 -1
stable?
o stable')
8-15
Find the ranges of k and k 2 such that the system with the transfer function
he foIlowing polynomials
+k
is BIBO stable.
8-16
It is known that the dynamical equation
iI parts of a1\ the roots of

s=s'-I.)
,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
STABlllTY Of LINEAR SYSTEMS
Consider the following linear time-varying dynamical equation:
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
.\: =(2ee- r' _2ee- r'jer'.;; +e-r'u =0 +e-r'u
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(e" to) for any
finite 1, and 10.J
8-22
Consider a sy'stem with the following dynamical-equation description:
x{i
y=[
1
-1
Consider a discrete-t
Show that any bounded-in

if and only if
8-28
Consider a discrete-
Show that any bounded-ii

if and only if
'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 \
Prove Theorem 8-13.
Show that the system is B
8-24
Prov~
8-25
Consider the linear time-invariant controllable state equation
condirion 3 o' Thcorerr.
~~-1
9.
1-] inL.
U:)(; f/rr
:==
and
X =Ax +Bu
Show that if u =
B*W - '(T)x, where
8-30 Prove Corollary 8

have proper rationa\ tran
W(T)=
e-A'BB*eAo'dr
T is an arbiti-ary ,positive number

8-31
Is the function
then the overall system is asymptotically stable. Furthermore, V(x(c)) = x*(e)W- '(T)x(e)
is a' suitable Lyapunov function for the closed-loop system.
8-26
Are the networks shown in Figure P8-26 totally stable? AIJSlrers: No; yeso
positive definite or semi
PROBLEMS
429
uation:
)-input response).
btaned by the equivalence
Figure P8-26
'esponse) From Problems

el! /101 preserve /e slabilily
: the BI80 stability of the
zero state of x = A(I)x are

-). ls the transformation
8-27
Consider a discrete-time system that is described by
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
ible i.s.L. at every 1 1 ;:>: lo.

less of -I(ll, lo) for any
8-28
Consder a diserete-time system that is deseribed by

y(n)
>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
mse BIBO stable? Is the
8-29
Consider a system with the impulse response

g(l)=g(I)
C/o(l -T)
i=O
Show that the system is BIBO stable ifand only if
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~
positivedefinite or semidefinite? [Hinl: Use Equation (E-2) in Appendix E.]
430
STABILlTY F LINEAR SYSTEMS
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
Consider the inI.
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
--'
where a, i = 1,2,3 are any real numbers.

Let Al = - 1, Az = - 2, A3 = - 3 and let a 1> az, a3 be arbitrary real numbers, not
zero. Prove by using Corollary 8-20 that the matrix
8-33
for i = 1, 2. are asympt

References SI62 and SI
8-39
af
2A 1
alaz
---}'I +},z
azal
M= - - - },z + Al
a3 a l
---Al + A3
Consider tbe poi:
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
is a positive definite matrix. [Hint: Let A = diag (A , Az, Az).J
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+
8-35 Let M be a hermitian matrix of order

jndependent vectcrs.
definite? (Answer: False.)
11.
Let e, i = !, 2..
.11,
,,~.
8-36
Prove Theorem 8-22.
8-37
Determine the asymptotic stability of the discrete-time equation
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
Consider the internally balanced system in Problems 6-23 to 6-25 partitioned as
are positive definite or
[ x~lJ
[Al!
I2 J[X 1J+[8 IJU
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
trary real numbers, not
for i = 1, 2, are asymptotical1y stable.

References S 162 and S 176.
8-39
This result is useful in system reducton.
See
Consider the polynomial in (8-48) with the following companion matrix
~~[~
'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
be, as in Definition 8-6,

the matrix M is positive
ling principal minors are
int: Try
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.
. ,n, be a set of linearly

then pv:iI is posilive
. ,11,
laton
f[
.u
+
~
(k)
YI
SI
I
u2
Y2
.. f2
S2
Ig Theorem 8-21 and (2)
Figure P8-40
(al
9
Linear Time-I nvariant
Composite Systems:
Characterization, Stability,
and Designs
9-1
Introduction
The design of control systems can be formulated as fol1ows. Giyen a plant,

design an oyerall system to meet certain design objectiyes_ Because of the
presence of noises, the compensators are required to have proper rational
functions or matrices. In order to reduce sensitiYity due to the plant yariations
and load disturbances, the configuration of the oyerall system must be of feed
back or closed-loop type. There are many possible feedback configurations.
In this chapter we study only the two configurations shown in Figure 9-1. The
one in Figure 9-1(a) will be called the unityfeedback system. The one in Figure
9-1(b) will be caHed the plant input-outputfeedback system or, simply, the input
output feedback system,! for the feedbacks are introduced from both the input
and output of the plant g(s). This configuration arises from the c'Jr.nbiD.<tUcr
of the state-feedback and state-estimator studied in Chapter 7. Before pro
ceeding, the reader should review the time-inyariant part of Section 3-6.
A control system is, as can be seen from Figure 9-1, basically a composite
system. By a composite system, we mean that the system consists of two or
more subsystems. Rather than plunging directly into the design problem, we
study first sorne basic problems associated with composite systems. We
assume thatthere is no loading effect in any connection of two suhsystems;
that is, the transfer function of each subsystem remains unchanged after the
connection (see Reference S46).
Although there are many forms of composite systems, theyare mainly built
1
This terminology is by no means universal. Another possible name is the Luenberger-- or state
estimator-type configuration_
432
Figure 9-1
(a) Unity
up from three basic e

Hence we shal1 restfi.
aH the subsystems tha
it is easy to show th.
system. For linear CI
can be directly appliel
linear time-inyariant :
and a transfer functio!
feedback connection,
output stability of the
many questions can:
ample, what is the il
91(s)andg 2 (s)? Is it
from 91 and 92 witho
questions for the sir
answered. Before prc
Definition 9-1
A system is said to be
matrix if and only if t
trollab!e and observal
The motivation of
description of a systerr
function matrix of th
controllab le from the
function does not desc:
equation description
information obtained
- transfer function of ti
case the system is saic
- matrix.
- Eyery dynamicale.
one; heilce when we a:
what its dynamical-eq
INTRODUCTlON
433
Ca)
Cb)
Figure 9-1
:>ws. Given a plant,

ves. Because of the
lave proper rational
:> the plant variations
;tem must be of feed
lback configurations.
n in Figure 9-1; The
1. The one in Figure
or, simply, the input
from both the input
'om the combination
Lpter 7. Before pro
:>f Section 3-6.
>asically a composite
m consists of two or
~ design problem, we
posite systems. We
i oftwo subsystems;
unchanged afier the
they are mainly built
the Luenberger- or state-
(a) Unity feedback system. (b) lnput-output feedback system.
up from three basic connections: parallel, tandem, and feedback connections.

Hence we shall restrict ourselves to the studies of these three connections. Ir
aH the subsystems that form a composite system are alllinear and time-invariant,
it is easy to show that the composite system is again a linear time-invariant
system. For linear composite systems, all the results in the previous chapters
can be directly applied. For example, consider the feedback connection of two
linear time-invariant systems with a transfer function g1 (s) in the forward path
and a transfer function gz(s) in the feedback path. Ir the transfer function of the
feedback connection, g(s) = (l +g 1(s)gZ(S))-l g 1(s), is computed, then the input
output stability of the feedback system can be determined from g(s). However,
many questions can still be raised regarding the composite system. For ex
ample, what is the implication if there are pole-zero cancellations between
gl(S) and gz(s)? Is it possible to determine the stability of the feedback system
from g 1 and gz without computing g(s)? In the first part of this chapter, these
questions for the single-variable as well as multivariable systems will be
answered. Before proceeding, we introduce a definition.
Definition 9-1
A system is said to be completely characterized by its rational transfer-function
matrix if and only if the dynamical-equation description of the system is con
trollable and observable.
;;n
The motivation of this definition is as follows. Ir the dynamical-equation
description of a system is uncontrollable and/or unobservable, then the transfer
function matrix of the system describes only the part of the system which is
controllable from the input and observable at the output; hence the transfer
function does not describe the system fully. On the other hand, ifthe dynamical
equation desciiption of a system is controllable and observable, then the
information obtained from the dynamical equation and the one from the
transfer functionof the system will be essent:illy the same. Hence, in this
case the system is said to be completely charact~rized by its transfer-function
matrix.
Every dynamcal equation can be reduced to controiiab](~ and observable
one; hence when we apply Definition 9-1 to a composite system, we m,ist c1arify
what its dynamical-equation description is. Let Si, i = 1,2, be two subsystems
434
LINEAR TIME-INVARIANT COMPOSITE SYSTEMS
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?,
9-2 Complete Characterization of Single-Variable
Composite Systems
Consider two systems Si> for i
1,2, with the dynamical-equation descriptions

x=x +B"
y =C;x + E"
(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
the systems SI and S

function G(s) comple
Example 1
Consider the paralle

whose dynamical-eql
and
Their transfer functio
The composite transf(
1t is clear that g(s) = 1

beca use g(5) does not
be checked from the
nection, we have u I =
equation is
It is easy to check
observable; hence frOl
charaderized by its ce
In the. following
transfer functions co
problem directly fre
function matrix, then
of the systemis equal t
COMPLETE CHARACTERIZATlON OF SINGLE-VARIABLE COMPOSITE SYSTEMS
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
Consider the parallel connection of two single-variable systems SI and Sl

whose dynamical-equation descriptions are, respectively,
and
FE1:
Xl =X l +Ul
FE!:
Yl =X 1 +Ul
X2=X2- U2
Y2 =X2
Their transfer functions are

s
91(S)=-
A
-1
s-
and
s-1
92(S) =-1
The composite transfer function of the parallel connection of SI and S 2 is

A
g(s) =91(S) +g2(S) = - -
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
l~:J=l~ ~Jl::J +l-~Ju
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)
and the output of the

The transfer-function
(9-2)
In the followng we shall study the cOl)ditions under which composite

transfer functioIls completely describe composite systems. We study this
problem directly from transfer-function. matrices without looking into
dynamicalequations. If a system is c6rilpl~tely chara~ierized by its transfer
function matrix, then the dimension. of the dynamical-equation description
of the system isequal to the degree of its transfer-function matrix (Theorem 6-2).
436
COMPLETE (
LINEAR TlME-INVARIANT COMPOSITE SYSTEMS
Therefore, whether or not a system is completely characterized by its transfer

function matrix can be checked from the number of state variables of the
system. Ir a system is an RLC network,2 then the number of state variables is
equal to the number of energy-storage elements (inductors and capacitors);
hence an RLC network 2 is completely characterized by its transferlunction
matrix if and only if the number oi energy storage elements is equal lO the degree
oi its transJer-Junct ion matrix. Consider now two RLC networks SI and S2,
which are completely characterized by their -transfer-function matrices Gl (s)
and G2(s), respectively. The number of energy-stora~e elemellts in any c~m
posite connection of SI and S2 is clearly equal to 15G l (s) +15G 2(s). Let G(s)
be the transfer-function matrix of the composite connection of SI and S2.
Now the composite s),stem consists of(15 G1(s) +15G 2(s)) energy-storage elements;
hence, in order for G(s) to characterize the composite system completely, it is
necessary and sufficient to have 15G(s) =15G l (s) +15G 2(s). This is stated as a
theorem.
TheQrem 9-1
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
....
This theorem can also be verified from the dynamical-equation descriptions

of systems. Recall from Section 3-6 that the state space of any composite
connection of SI and S2 is chosen to be the direct sum of the state spaces of
SI and S2; consequently, the dimension of the composite dynamical equation
is the sum of the dimensions of the dynamical-equation descriptions of SI and
S2- Hence Theorem 9-1 follows directly from Definition 9-1 and Theorem 6-2.
In order to apply Theorem 9-1 we must first compute the transfer-function
matrix of a composite system. This is not desirable, particularly in the design
of feedback control systems. Hence the conditions in terms of G1 and Gzfor G
to characterize completely the composite connections of SI and S2 will be
studied. We study in this section only single-variable systems. The multi
variable systems will be studied in the next section.
The transfer function of a single-variable system is a scalar, and its degree
is just the degree of its denominator if the transfer function is irreducible. We
assume in this section that all transfer functions are irreducible; that is, their
denominators and numerators are coprime.
1. The parallel conr

91(S) +92(S) ifano
2. The tandem conr
92(s)9 1 (s) if and 01
3. The feedback cor
characterized by !
canceled by any z
Pfoof
1. It is obvious tha
15 9 < bg l +1592. 1
We show now tha
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
Consider the tandem
as shown in Figure 9-.
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
COMPLETE CHARACTERIZATlON OF SINGLE-VARIABLE COMPOSITE SYSTEMS
:rized by its transfer

:ate variables of the
r of state variables is
ors and capacitors);
it5 tran5ferjunction
i5 equal to the degree
networks SI and S2,
ction matrices (;1(5)
lements in any com
.) +<5G 2 (5). Let G(5)
:ction of SI and 52'
rgy-storage elements;
;tem completely, it is
This is stated as a
437
1. The parallel connection of SI and S2 is completely characterized by 9(5) =

91(5) +92(5) if and only if gl(5) and 92(5) do not have any pole in common.
2. The tandem connection of SI and 52 is completely characterized by 9(5) =
92(5191 (5) if and only if there is no pole-zero cancellation between 91 and g2.
3. The feedback connection of SI and 52 shown in Figure 9-2 is completely
characterized by 9(5) = (1 +9192) - 191 if and only if there is no pole o 92(5)
canceled by any zero of 91 (5 j.
Proof
1. It is obvious that if 91 and 92 have at least one pole in common, then
lJ9 < lJg 1 +lJ92' Let 9i = N (Di' for i = 1, 2; then
A
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
Consider the tandem connection of SI and S2 with transfer functions

5
1
-1
and
as shown in Figure 9-3(a). There is a pole-zera cancellation between 91 and
re completely charac
ldg z (5).
y cutsets in he netwof.k.
5-1
s + 1
Figure 9-2
A single-variable feedback system.
9z
438
ce
On the other han

path of Figure 9-4(b)
back path, the transf<
terizes the feedback ~
(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
Hence the composite transfer function

A
g(s) = g2(S)g I(S) =--1
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
where .u and y are th,

descriptions describe t
If {A, B, C} is not irred
{D, NI}' and {pes), Q
strictly system equivale
system equivalent, and
For our problem, it tI
convenient. Hence th
We recall from (6-151
{pes), Q(s)} is left COpl
right coprime. These
complete characterizat
..
. _---_ .. _----
_,._~
-------~._-~-_
..
.. ---
_-~-~-_
CONTROLLABIUTY AND OBSERVABIUTY Of COMPOSlTE SYSTEMS
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[
always characterizes the feedback system completely.
This can be seen by

ugh the output of the
~s exponentially with
I
In Figure 9-4(a), the

led by the zero of the
(lsfer function of the
,ack system. Indeed,
9-3 Controllabilityand Observability

of Composite Systems
Let g(s) = N (s)D-(s), i = 1, 2, and let N (s) and D(s) be coprime. In terms of
Ni and D, Theorem 9-2 can be stated as follows: The parallel connection of
Sand S2 is completely characterized by g(s) +g2(S) if and only if D(s) and
D2 (s) are coprime. The tandem connection is completely characterized by
g2(S)g(S) if and only if D(s) and N 2(s) are coprime and D 2(s) and N (s) are
coprime. In this section we shall extend these conditions to the multivariable
composite systems. Before proceeding, the reader is advised to review Section
6-8, on strict system equivalence.
A system can be described by a dynamical equation {A, B, C, E}, a transfer
matrix in fractional forms C(s) = N,.(s)D,:-(s) =D-(s)NI(s), or a system matrix
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

CON
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
where ~(s), Ui(S) and y(s) are, respectively, Pi x 1, Pi

Nl(S)][ ~(S)J = [
O]
O
-u(s)
-Yi(S)
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.
wheres(s), u(s), and y(s) are, respectively, q xl, p x 1, and qi xl vectors.

For the parallel connection shown in Figure 9-5, we
have P=P2,q=q2'U=U 2 =u, and Y=Y +Y2, where u and y denote the
input and output of the overall system. Using these equations and (9-3), we
can obtain
Parallel connection.
~~'~~s: I?;~~)__)] [t~;~-J
l-N,.(s)
-N..z(s):O
-u(s)
= [ __ __ ]
-y(s)
(9-5)
to describe the parallel connection. Although (9-3) are irreducible by assump

tion, the system matrix in (9-5) is not necessarily irreducible. In order to obtain
a simpler result, we carry out the following strict system equivalence trans
formation:
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
O,.(s) and D..z(s) are l.

nection is controllable
Ir we use (9-4). then
(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
~
The roots of det O,.,(s;

H). lf the poles of GI
[D,.(s) Dds)] and the r.
(why?). Hence a suffici<
trollable and observable
This condition however
Ir G(s) and G 2 (s) ar~
the folIowing: The parall
only if their denominators
proof of statement 1 of Th,
9-3, we have the necessary
terize completely the para
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
The parallel connection of SI and S 2'
Figure gc6
Y;
Uz
S,
S12. the tandern (
CONTROLLABIUTY AND OBSERVABILlTY OF COMPOSITE SYSTEMS
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
ly, right and left co

'ibed 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 )
By a similar argument, it can be shown that the parallel connection is observable

if and only if 011(S) and Dds) are right coprime. \Ve recapitulate the aboye
results as a theorem.
Theorem 9-3
Consider two systems which are completely characterized by their transfer
matrices GI(s) and G 2 (s). Let G(s) = 0i;- l(s)Nl(s) = N,.(s)O,~ I(S) be coprime
fractions of (;;(s). Then the parallel connection of these two systems is con
trollable if and only if O"I(S) and Ods) are left coprime. The parallel con
nection is observable if and only if OII(S) and 012(S) are right coprime.
Iil
(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
lock row of the system

lability and coprimeness
TI equivalence (Theorem
:cription of the parallel
~~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).
For the tandem connection of SI followed by S2

shown in Figure 9-6, we have q = P2, u(s) = M] (s), an 2 \s) = YI (s), aod y(s) = Y/el.
Using lhese reiations and (9-3), we can obtan
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
Sil' the tandem connection of SI followed by Sl'
442
ce
LINEAR TlME-INVARIANT COMI'OSITE SYSTEMS
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
fol1owing three pain

and N'2(S), or D,.(s) ,
Let G;(s) be a q x
N'i(S)D,: l(S)= Di! (5)
called the po/es of G(
'3!in (Pi, q J, where p ,
Gj(s) (see AppendixH
o
(9-9)
O'2(S)
has a ful1 rank for every s in C.

block row,3 (9-9) becomes
By adding the second block row to the third
Corollary 9-4
O: O IJ
O,,(s)
D'2(S) : 1
A sufficient condition
O,-z(.~)'
P"l ?q, q2 ?P2' and q
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
- - - - --- - - - - - - i - ---- 0-:-
Hence we conc1ude that the tandem connection is controllable if and only if

N l2 (S)N r1 (s} and D l2 (S} are lefi coprir(~~:;. Ey!:1. :..:lrrlar ~lrgU;:.G'~;:;';'1 -,:"j~:. ~:;(.'~~'" :;:::'.":\;::
the fol1owing theorem.
~nd. (Y;;:!~i ~~ :.:~~~.-::; ):'
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.
provides a different proc

The concept of pole
the matrix case if it is ca,
the tandem connection
<5G(s) = c5( G 2 (s)G (s)) = <5
tandem connection is eitl
<5G(s
I
I !
In this case, we may defir

For example, if
CONTROLLABILITY AND OBSERVABILITY Of COMPOSITE SYSTEMS
02(S) is a part of the

he tandem connection
(9-9)
block row to the third
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
every s in e if and only

lence we conc1ude that
D"2(S) and N.. 1 (s) are,
~,
thcn the tandem con
Jl-~y(s)i 1J
(9-10)
to the third block row,3
ntrollable if and only if

gument, we can establish
:terized by their transfer

~ri(s)D; 1(8) be coprime
\ followed by S2 is con
Tee pairs of polynomial
lds) and Nds)Nrl(s), are
md only if anyone of the
A sufficient condition for the tandem connection of SI followed by S2 with

PI "2:.Q1>Q2 "2:.P2' and ql =P2 to be control1able (observable) is that no pole of
(;2(S) s a transmsson zero of (;1(S) [no pole of (;I(S) s a transmsson zero
of (;2(S)].
Proof
Let (;1(S) = N-[(s)D,:I(S) and G2(s) = N"2(S)D,:;I(S), where N"1(S) is ql X PI
polynomial matrix and Dds) is a P2 x P2 polynomial matrix. D"2(S) has rank
P2 at every s except the roots of det D,As). If no pole of (;2(S) s a transmisson
zero of (; I(S), we have pN"I(S) = ql = P2 at the roots of det Dds). Hence the
matrix [Dds) N r1 (s)] has rank P2 at every s in e and, consequently, is left
coprime. Hence ifno pole of (;2(S) s a transmission zero of (;1(S), the tandem
connection of S1 followed by S2 is, fol1owing TheQrem 9-4, controllable. The
observability part can be similarly proved.
'
Q.E. D.
The combination of the controll~blitx and observability condtions in
Theorem 9-4 yields the condition for G 2 (S)G 1 (s) to completely characterize the
tandem connection. If (;1(S) and G2 (s) are 1 xl rational functions, then
Theorem 9-4 reduces to that the tandem connection of g1(S) fol1owed by g2(S)
is control1able (observable) if and only if no pole of g2(S) is canceled by any zero
of gl(S) [no pole of gl(S) is canceled by any zero of g2(S)]. Hence g2(S)gl(S)
characterizes completely the tandem connection of gl(S) followed by 92(S) if
1.nd only if there is no pole-zcro cancellgu.c~~.' '/";';:'1,''-:''': ..: ,~::; ( ". ~l.~
provides a different proof of statement 2 of Theorem 9-2.
The concept of pole-zero cancellation in the scalar case can be extended to
the matrix case if it is carefully defined. If G 2 (S)G 1 (s) characterizes completely
the tandem connection in Figure 9-6, then we have, following Theorem 9-1,
<'5G(s) = <'5(G 2 (s)G 1(s)) = <'5G 2 (s)+ <'5G 1 (s), where <'5 denotes the degree. Ir the
tandem connection is either uncontrollable or unobservable, then we have
;
";'-".'
<'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)
[DNll(~))] have no full rank.

12(;"
N1I(A)] and/or [
D ll V,)D,z(A)]
Nd},)
have no ful1 rank.
NdA)N"I(A)] and/or [ N D"I(A)]

')N
have no ful1 rank.
dA ,.(A)
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
Feedback conneclion of SI and S2'
~[G(sr
deg G(s)"" ~
where ~[.] denotes the

denotes equality of polyr
CONTROLLABIUTY AND OBSERV ABIUTY Of COMPOSITE SYSTEMS
one pole-zero cancel

zero cancellations in
el1ations in (;1(S)G 2(s),
;:;ase, pole-zero cancelcancellations of poles
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.
~rized by their transfer
(00)G 2(00 fO. Then

md only ir SI 2 is con-
In general, the condition of control1ability of 5 12 is different from the

candition of observability of S21' However, in the single-variable case the
two canditions are identical. Indeed if no pole of 92(S) is canceled by any zero
of 91(s), then 91(s) followed by 92(S) is controllable, and 92(S) followed by gl(s)
is observable. Hence we conclude that, for the single-variable case, the feedback
connection in Figure 9-7 is controllable and observable if and only if SI2 is
controllable or, equivalently, no pole of 92(S) is canceled by any zero of 91(S),
This checks with what we have in part 3 of Theorem 9-2.
We consider a special case of Figure 9-7 in which Gz{s) is a constant matrix
K We cal! this a constant outputfeedback system. Since K does not introduce
any new state variable, if SI is control1able, so is SI followed by K. If SI is
observable, so is K fol1owed by S l' Hence the feedback connection of SI
with a constant gain K in the feedback. oath 1s contro!Iabie 8.p.d I)bserv'~.bc; r
and oniy if SI is controilable and observable. Hence controllability and
observability are invariant under any constant output feedback. This contrasts
with the constant state feedback discussed in Chapter 7 where controllability
is preserved under constant state feedback, but observability is generally not
preserved.
We discuss further the feedback system in Figure 9-7. Let the system Si
be described by irreducibleG(s)= N..(s)D,~ I (s) = Di; I (s)Nti(s) and irreducible
{A j, B, e, E;}, where GI(s)and G 2(s) are, respectively,'q xp and p x q proper
rational matrices. By assumption, we have
.
and
i![ G(s)] ~ det D,.(s) ~ det Dli(S)~ det(sI - A;) .
deg (;(s) = deg de.t Dr(s) == deg det Dli(S) = dm A= 11
where i![' ] denotes the characteristic polynomial of a rational matrix and ~

denotes equality ofpolynomials modulo a nonzero constant factor. The overall
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)
The state equatior
The matrix A of its state-variable description is, as derived in (3-64),

AJ
__ [A-BY 2 E 2 C
B 2 YC
-BY2C2
]
A 2 -B 2YEC 2
with Y =(1 + EE z )- and Y z =(1 + E 2E)-.

completely characterized by GJ(s), then
( 9-12)
Ir the feedback system is not
L\.[ GJ(s)] det (s 1- A J)

deg GJ(s) < dim AJ=n + nz
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 ~:.
lts A matrix is identic

Theorem 9-6
Consider the system in
O. Then we have
L\.(H(s))
~ (S)J
[ uz(s)
where ~ denotes the

factor.
By definition,
H(s) [:
I(S)J
r z(s)
We remark first that th,

in Section 3-6, the pro::
H(s) may have poles a
right fraction
. [.
H(s)=
. 1I2
Figure 9-8
+det [Dds)D,.(s) + Nds)N,.(s)]
because the factorization in (9-11) is not necessarily irreducible.

The controllability and observability of a system depends on the assignment
of the input and output. Although the system in Figure 9-7 may be uncon
trollable or unobservable, if we assign the additional input and output as shown
in Figure 9-8, then it is easy to show, by using the argument in the proof of
Theorem 9-5, that the feedback system is always controllable and observable
from the input [r'l r~]' and the output [Y'I y'zJ. Similariy, the system is
always controllable and observable from the input [r') r~]' and the output
[u' u~J. We study now the transfer matrix H(s) from [r'l r'J' to [u' u'zJ.
From u(s)=r l (S)-Y2(s)=r,(s)-G 2(S)U 2 (s) and lJ 2 (s)=r 2 (s)+YI(s)=r 2 (s)+
G(s)u(s), we have
fl
hence, we have
r..,
Feedback system with additional input and output.
~1
- G (s)
=[D,o~(s)
which can be easily veri
447
CONTROLLABILITY AND OBSERVABILITY OF COMPOSITE SYSTEMS
; 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
feedback system is not
- B I Y zC, ]
Az-BzYIEICz
-B 1 Y zE z
BzY\ _ r2
BzY,E t
Its A matrix is idenlical lo lhe A r in (9-12).

Theorem 9-6
factor. Conversely, if
rnamical equation de
)l1able and observable.
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)
+ {;, (.I')G :!(sl]
denotes the egua lity of polynomials modulo a nonzero constanl
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
by carrying out elementary row operations,
O"I(S)
-N'I(S)
0"1 (S)
S
Nd )]
O
Ods) -+ -Nr(s)
O
Or (S)
Ods)
O
and the right coprime assumption of N..(s)O,~ I(S) (Theorem G-8).

have established, following (6-189), the equality in (9-13c).
Again by direct verification, we have
H(s) = [
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)
It is easy to show that the factorization is irreducible. Hence we have
~(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)
where we have used, in the last step, the formula
det[~ ~]=det[-dP-l ~][~ ~]=det[~ V-~p-IU]

=det P det (V - QP- 1U)
This establishes (9-13a). The equality ~[H(s)] ~ det (sI - A f ) follows directly
from (6- 189) and the fact that the state-variable equation is controllable and
observable from the input [r'l r~]' and the output [U'1 u~]'.
From (9-13a), we have
det {Ol1[I + 0 11 l(s)Nl1 (s)N"2(S)Or2 1(s)]O,z{s)}
= det D l1 (s) det [1 + (;1(S)(;2(S)] det D,z(s)
Q.E.D.
This implies (9-13[).
This theorem provides five formulas for computing det(sI-A f ) and is

useful in stability study. In the single-variable case, there is no distinction
between right and left fractions, and the theorem reduces to
det (sI - A f )=D(s)D 2 (s) + N 1 (s)N 2 (s)
where 9(S) = Ni(S)/Di(S), i= 1,2.
9-4
totically stable if the

state as t -+ oo. A s'
if the zero-state res po;
asymptotically stable,
However, ir a system
asymptotic stability ir
Consider two line
characterized by thei
transfer matrix G(s) (
stability or the compe
asymptotic stability (
system is known to \;
task to compute com]
nection. Even if the
tedious to check whei
terized. Therefore, it .
osite systems from
function matrices. W(
For the parallel an
simple. The parallel
stable if and only ir 5
asymptotic stability is
(3-62) and (3-63), we ~
nection of SI and 52 a
assertion follows.
For the feedback
feedback system may
back system may be un
variable case and then
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
Stability of Feedback Systems
Every control 'system is designecf to be stable. Therefore it is important to

study the stability problem of composite systems. We recall from Chapter 8
that a system or, mOre precisely, the zero-input response of a system is asymp
Theorem 9-7
Consider the feedback ~
52 arecompletely char
STABILITY OF FEEDBACK SYSTEMS
'S)l
(s)
:)[em G-8).
Thus, we
).
ence wehave
[- A f) follows directly
.on is controllable and
"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)
i'f ' Q ?JI;s;

D,(s)D 2(s) +N,(s)N 2(s)- Df(s)
where g(s) ~ N(s)/D(s) for i = 1, 2 and N f and Df are assumed to be coprime.
Recall that every transfer function is assumed to be irreducible; therefore, when
we speak of g(s), we mean gf(S) = Nf(s)/D(s). It was shown in Theorem 9-2
that if N, and D 2 have no common factor, then D f =D,D 2 +N I N 2, and the
. system is completely characterized by g(s). If there are common factors be
-tween.N, and D 2, then D j consists of only a part of D,D 2 +N I N 2, and the
system is not completely characterized by gf(S).

Thebrm.9.-7.
: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.
gzCs). Itisalsoassumedthatl +g(<Xl)92(<Xl) 7'=0. Letg(s)= N(s)/D(s). Then

the feedback system is asymptotically stable if and only if all the roots of
D(s)Dz(s) + N(s)N z(s) have negative real parts. The condition is sufficient but
not necessary for the systern to be BIBO stable.
Proof
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 + ...
This theorem car

proved here, howeve
Prom (9-15) we see t
independent of wheth
tion of the feedback
whether the feed ba(
I + (<Xl )g z (<Xl) = O, e
real parts, the thec
(- S2 + s + 1)/S 2 , 92 s,
(- Sz + s + I )/(s + 1) \1
N1(s)Nz(s)=(s+ 1) h,
input sin (2 will excitt
ofthe feed back system
0, Theorem 9-7 does r
If yz(s) in Figure 9-'
output feedback systt
controllable and obse
implied by BIBO stab
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 )
D(s)D 2 (s) +N(s)N2(s)=aIOa20Sn,+n2 + ... +bIOb20Sn,+rt2 + ...

=(QIOQ20 +b lO b 20 )sn, +n,
(9-15 }
+terms of lower degrees
From (9-14) and (9-15) we see immediately that if 1 +g(<Xl)92(<Xl) 7'= 0, then
deg [D(s)D 2(s) + NI (s)N 2 (s)] = n + n z = deg det (sl- A f ). Using this fact and
the relation
g(s)=c(sI-Af)-bf+ef=d
(; A ) [cfAdj(sI-A f )bf ]+ef

et s - J
NI (s)Dz(S)
D I"(s)D 2(s) + N(s)N 2(s)
we conclude that
det (sI - A f ) =k(D"(s)D 2(s) + N(s)Nz(s))
Consider the feedback

that S is completely
I + kg ( <Xl) -=/=- O. Then
stable if and only if al
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,
STAB1UTY OF FEEDBACK SYSTEMS
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
ensure the properness

ctor of the feedback
I-bu, Y=cfx +eu be
:see Equation (3-64)].
~ is a polynomial of
)D2(S) + N (s)N 2 (s)] =
ill not be canceled by
This theorem can in fact be deduced directly from Theorem 9-6. It is

proved here, however, by using explicitly the condition I +g(00)g2(00)+0.
From (9-15) we see that (9-16) holds if and only if I +g(00)g2(00)+0; it is
independent of whether the (n + n 2 )-dimensional dynamical equation descrip
tion of the feedback system is irreducible or not. It is also independent of
whether the feedback system is completely characterized by g(s). If
I +g(00)g2(00)=0, even if all roots of D(s)D 2(s) + N(s)N 2(s) have negative
real parts, the theorem stili does not hold. For example, if g(s)=
(-s2+ s +1)/s2, g2(s)=I, then we have l+g(00)g2(00)=0 and 9(S)=
(-S2+ S+ I)/(s+ 1) which is improper. Although the root of D(s)D 2(s)+
N{s)N 2(s)=(s+ 1) has a negative real part, the application of the bounded
input sin {2 will excite a term of 2{ cos {2, which is not bounded, at the output
ofthe feedback system. Thus without imposing thecondition I + g (00 )g2(00)
0, Theorem 9-7 does not hold in general.
If y2(S) in Figure 9-9 is a consta~nt k, the feedback system reduces to a constant
output feedback system and is, as discussed in the previous section, always
controllable and observable. In this case, asymptotic stabiiity implies and is
implied by BIBO stability.
Corollary 9-7
(9-14 )
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)
nvalues of A and the

Hence the feedback
~r
Figure 9-10
A multivariable feedback system.
452
First we note
det [1+ G1 (<Xl)J:b (
L1 1 (s)det [1+ (; I (s)] i
we have
where Adj(') denotes the adjoint of a matrix. Since SI is completely charac

terized by G (s), it fol!ows from Theorem" 9-5 that the feedbas.k system in
Figure 9-10 is completely characterized by G(s). lf the order of G (s) reduces
to 1 xl, then det [1 +G 1 (s)] becomes 1 +g(s) and the stability depends only
on the zeros of 1 +g(s). Hence, one might be tempted to assert that the
stability of multivariable feedback systems depends only on the zeros of
det [1 +G(s)].
where E(s) and F(s) a
To show that ,1(s) d<

6.(s) is divisible with
following identity:
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
where p is the order (

(;(s) of order i; for
elements of (;(s). TI
Now 6.(s) is the least
F(s) is at most equal
minors of (;(s) folle
without remainder b
Theorem 9-8 can be (
portance, we shall pro
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
Note that the dynam

its knowledge is not
assumption that S is
control!able and obst
6. (s) of G (s) is equal t
By the substitution 0[.
453
First we note that e As) is proper because of the assumption

det [1 + G1 (ro)] =1=- o (Theorem 3-5). Before proving the theorem, we show that
~[(s)det [1+ e[(s)J is a polynomial. Define
E(s)
det [1 +G[(s)] ~ F(s)
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)
where E(s) and F(s) are assumed to be coprime. Then we have
~[(s)det [1 +e[(s)1 = ~[ls)E(s)

-
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 =
gure 9-10 with
(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.
= - 1, which does not

)f the feedback system
;(s)].
III
ck systems, the concept

itrix is needed. Recall
1 o G1(s), denoted by
s of G(s). If G(s) has
~(s)=kdetD(s), for
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
Weshall use the dynamical-equation description of the system to pro ve the

first part of the theorem. Let
FE:
x = Ax + 81I
y =Cx + E"
be the dynamical-equation description o the system .)". it is ciear
(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;) .
By the substitutionof ti 1.= u - y , y = y[ and the use f the identity

1-'---(1 +Etl-[E[ =(1 +Etl-
(9-22)
"-.
454
the dynamical-equation description 01' the feedback system can be obtained as

FE:
X =[A-BI(I+E)-lCJx +BI(I+E)-lu
y = (1+ E1r IClx + (1 + E)- IElu
(9-23a)
(9-23b)
The dynamical equation FE is controllable and observable following Theorem

9-5. Hence we conclude that the feedback system is BIBO and asymptotically
stable if and only if all the eigenvalues 01' Al -B(I +E)-C ha ve negative
real parts. We show in the following that the set 01' the eigenvalues 01'
Al -BI(I +E}- IC I is equal to the set of the zeros of c,(s)det [1 +G(s)].
Consider
=det {(.d - Al )[1 +(sl- A )- 8 1(1 + El )-ICIJ:
(9-24 )
Example 2
Determine the stabil
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
It is well known that det AB =det A det 8: hence (9-24) becomes

f(s)
= det (sI - Ad det [1 +(sl - A}-BI(I +E)-CJ
(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
where E(s) and F(s) h

feedback system intr
poles which are posse
system. For exampl,
(1 + 1)(s - 1); hence n
G(s)and G1(s)have
Roughly speaking
of det [1 + (;(s)]. T
any common factors i
might be determinabl
necessarily correct, as
Consider the feedbacl
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,
(Theorem G-8). This equation implies

X(s)[D,. (s) + N.. (s)]
+ [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.
11' we do not cancel o

half-plane zero. Thw
not stable.
STABIUTY Of FEEDBACK SYSTEMS
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
Determine the stability of the feedback system in Figure 9-10, with
le fol1owing Theorem
) and asymptotical1y
)-le l have negative
f the eigenvalues of
o(s)det [1 + (;1(5l].
455
E(s)
o(s)det [1 +GI(sj] =o(s)
F(s)
(9-25)
6 2(s)(; I(s)J and,
(9-26)
: from (9-26) that the

he set of the zeros of
Drem.
I(S). Then we have
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)
Consider the feedback system shown in Figure 9-10 with

1
e Dr(s) + N,(.s) and

istence of polynomial
s-0.5
1
s -0.5
lt is straightforward to verify that
)=1
[heorem G-S implies

)nsequently, G(s) =
t; (s)] ~ det [D,. (s) +
asyrnptotical1y stable
legative real parts.
Q.E.D.
(s + 0.S)(S2 + s + 0.25)
(s+ 1.5)(s-0.5)2
s+0.5
(s+ 1.5)(s- 0.5)2
(s +0.5)(5 - 0.5)(s + 1.5)

(s - 0.5)(s - 0.5)(5 + 1.5)
If we do not cancel out the common factor, then det [1 + G(s)J has a right
half-plane zero. Thus we may suggest erroneously that the feedback system is
not stable.
456
The determinant of (;l(S) is given by

1-(s+0.5)
det G 1(s) =
and conseguently,
(s
-1
det (.
=---:-
+ 1.5)(s - O.W (s + 1.5)(s - 0.5)
Hence the characteristic polynomial, L1 1 , of
if and only if
G1(s) is
Ll 1 (s) =(s +1.5)(s -0.5)
Tosave space, this as
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 -
A ) = L1 (s)L1 (s) det [1 +A G1(s)?z(S)]

f
Figure 9~11
A multivariable feedback system.
deg l(S) + deg Gz(s) =
["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)]
For this reason, the c(

system in Figure 9-11
the feedback system i~
terized by G(s) and
Theorem 9-9 beco mes
Now we consider
and output as shown i
(9-28)
det [1 + G1(oo)Gzioo)]
n1
+ nz
[~"
-G 1 (s)
Gz(
[1" + G2 (s)d
[ l(s)[i p +G z(
riecause the cnaracter;

egual to det (si - A f ),
sufficient for the systen
that there exist G 1 (s)
stable but the fourth er
instability of the en trito
In the design of fee
section ofwe11-posedn(
pairs to be proper. SiJ
al! possible input-outp
is not BIBO stable, an
input at rz wi11 cause th
in practice to reguire e'
input-output pairs.
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
ppears in the feedback

lerally not completely
situation here is more
sure 9-10.
G (s) = [1 + G(s)Gz(s)] - I GI(s)

= [1 + Di I(s)N II (s}N,.z(s)D,:;I(SJ] - I Dil(s)N lI (s)
= Dds)[DlI(s)Dds) + NII(s)N,.z(s)] - I NlI(s)
, assumed that SI and

oper transfer m~trices
(co)J '1'=0. Let G(s) =
ns and ~i(S) be the
,tem is}syrnytotical!y
t [1 + GI(s)GZ(s)J, or
sative real parts. The
n to be BIBO stable.
::essary as wel! for the
1
n 9-6. It can also be

;tablish, as in (9-26),
(.1')(; z(s)J
(9-28)
co )(;z(co)]
he interested reader lS
nd (9-15), it is possible
>rz(s) and the row re-
(9-30)
The factorization in (9-30) is not necessarily irreducible; hence general!y we have

~[G (s)]
+O
f det [DI I(s)Dds) + NII(s)Nds)] ~ det (s 1- A)
(9-31 )
For this reason, the condition in Theorem 9-9 is only sufficient for G(s) 01' the
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)
By direct computation, we have

H(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)
Because (he characteristic polynomiai 01 H(s) is, as dcveloped in fleorem 9-6,

equal to det (sI - A), the condition in Theorem 9-9 is now necessary and
sufficient for the system in Figure 9-8 to be BIBO stable. It is important to note
that there exist GI(s) and Gz(s) so that any three entries of H(s) can be BIBO
stable but the fourth entry is noL For al! possible combinations of stability and
instability of the entries of H(s), see Reference S75.
In the design of feedback systems, we shl! require, as discussed in the sub
section of well-posedness, every transfer function from l! possible input-output
pairs to be proper. Similar1y we shall require a system tobe BIBO stable from
al! possible input-output pairs. For example, ifthe transfer matrix from r z tO"1
isnot BIBO stable, any load disturbances Which can be 11l0deleq as a nonzero .
input at l' z will cause the system in Figure9-8 to satura~. '. Hence it is rea"sonable
inpractice to require every feedback system to be BIBO stable from al! possible
input-outPllt pairs.
458
[Y'I
LINEAR T1ME-INVARIANT COMPOSITE SYSTEMS
Consider again the system in Figure 9-8 with [r'l

y'2T as the output. Clearly, we have
[~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)
ra as the input and
(S)
which mplies, by using (9-32),

HAS)=[A GI~)[lp+G2~)GI~)]-1
G 2(s)G 1(s)[l p + G 2(s)G 1(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
Using the identity

1- G(s)Gj(s)[1 + G(s)Gj(s)] -1
Figurb 9-12
(9-34)
and defining
sator C(s) = N,(s)/Dc(s

is a real constant. W,
without imposing any,
of C(s) is imposed to
Let gf(S) be the trar
(9-35)
we can readily establish
::;-
rom gf s +gf(S)C(s);
N(s)/D(s), we can write
A
()
(9-36)
From this relationship, we see that, because of the nonsingularity of J, every

pole of H(s) is a pole of HAs) and vice versa. Hence H(s) is BIBO stable if and
only if HAs) is BIBO stable. This fact can also be seen from Figure 9-8. Because
OfY2 = r l - U1 and Yl = U2 - r 2, if r(t) and U(t) are bounded, so arey(t), i= 1,2.
Corollary 9-9
The feedback system in Figure 9-8 is BIBO stable if and only if the feedback
I
system in Figure 9-11 is asymptotically stable.
This corollary can also be conc1uded from Problem 8-40. Thus, if a feedback
system is designed to be asymptotically stable, then the system is BIBO stable
from al1 possible input-output pairs. On the other hand, if a system is BIBO
stable from one input-output pair and if there is no unstable hidden mode or,
roughly speaking, no unstable pole-zero cancellation, lhen the syslem is asymp-
totically stable. In practice, every control system should be designed to be
asymptotically stable.
9-5
Design of Compensators: Unity Feedback Systems
Single-variabJe case. In Chapter 7, we studied the problem of state feed

back and state.estimator and showed that the eigenvalues of the original system
can be arbitrarily assigned withOut affecting its numerator. In this section, we
shall study a similar problem by using transfer functions in fractional forms.
We study" onlyunty feedbacksystems in this section; the configuration devel
opedin Chapter 7 wi11 be studj(~d inSection 9-7.
Consider the unityfeedback system shown in Figure 9-12. The compen-
f\.
C(s) (
-[
Now we c\aim that fo

inequality
where j denotes the de,

exist to achieve the desi
which, together with (9

(j D e(s) - j N e!
rlence C(s) is proper. Ti

a proper compensator e
not only the poles but ;
assigned.
There is, however, a
see that C(s) contains t
nection of C(s) fol1ow(
Furthermore, the cancel
has no control over thest
right-halfs plane, then tI
Hence this design is no
undesirable pole-zero ca
polesand zeros are no\
from (9-37).
DESIGN OF COMPENSATORS: UNITY FEEDBACK SYSTEMS
as the input and
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
Unily feedback syslem.
sator C(s) = N,(s)/Dc(s) is required to be a proper rational function. The gain k

is a real constant. We shall study what can be achieved in this configuration
without imposing any constraint on C(s) so long as it is proper. The properness
of C(s) is imposed to avoid amplification of high-frequency noises.
Let 9(S) be the transfer function from r to y in Figure 9-12. Then we have
g(.s)
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)
d only if the fedback
';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)
which, together with (9-37) and (9-38), imply

bDc(s) - bN ,(s) = [bO(s) - bN (s)] - [bD(s) - bN(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
In the following, we discuss a method of computing C(s) which will not

involve any undesirable pole-zero cancel1ations. However, we must paya
price for achieving this. The zeros of N(s) of the resulting system cannot be
arbitrarily chosen; they are dictated by the zeros of C(s) and g(s). If we wish to
control the poles as well as the zeros of gAs), then we must choose a different
configuration (see Problem 9-10 and Reference S34).
The substitution of g N /D, g N/D, and C NJD c into g(s) =
kC(s)(s)/[l + C(s)(s)] yields
N(s)
D(s)
kNc(s)N(s)
Dc(s)D(s) + Nc(s)N(s)
(9-39)
Equating the coeffici
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 arbitrary pole placement, we study only the polynomial equation
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
.
.
This is a set oflinear ~

between the polynorr
hence it is permissit
Sm consists of m + 1 1
coefficients of g(s) anl
right by one column.
application of Theor
(9-44) has a solution
S", has a full column
rank is that 2(m + 1);
compensator to achi
G-4, we have that S"
In fact, from the proo
for m:2:n -1 ifand on
the following theoren
(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
with deg N(s)~deg D

for every DAs) of degI
and only if D(s) and !'
This theorem statl
to meet (9-46). Noth
Ir D(s)and N(s) are note.

only if the grealest como;
C(s) which will not

:ver, we must paya
Img system cannot be
ld g(s). If we wish to
I 1St choose a different
Equating the coefficients of the same power of s yields
[DeoNeo:Dc1Nc1:"':DemNem]Sm=[Fo
rIt"_o__!!___ ~._._ J!::-: ~ __ ,,!_,,

D
'"
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"
...
1st block row
O~
--------------------------------------
(9-39)
,139. 1t is called the
with
Do
i N(s) under which a

We shalI also study
lry pole placement. 4
y that complex con
JI' will have complex
461
-'-0-0- -
...
--0- - - - -i>~ -- -i>~ -
No
(9-45)
-i>; --.-.~ --D,~ } (m+ l)th block row

N z ...
N"
This is a set of linear algebraic equations. There is a one-to-one correspondence

between the polynomial equation in (9-40) and the algebraic equations in (9-44);
hence it is permissibk to study the former by using the latter. The matrix
Sm consists of m + I block rows; each block row has two rows formed from the
coefficients of g(s) and can be obtained by shifting its previous block row to the
right by one columno lt is clear that Sm is a 2(m + 1) x (n +m + 1) matrixo The
application of Theorem 2-4 to the transpose of (9-44) reveals that for every F,
(9-44) has a solution {De;, N ei , i = 0,1,. o. ,m} if and only if pSm = n +m + 1 or
Sm has a fuIl column rank. A necess~ry condition for Sm.to have a fuIl column
rank is that 2(m + l);::::n +m + 1, 01' m;:::: n -1. Hence the smallest degn~e o(a
compensator to achieve arbitrary pole placement is n -1. From CoroIlary
G-4, we have that S"-1 is nonsingular if and only if D(s) and N(s) are coprime.
In fact, from the proof of CoroIlary G-4, we have that Sm has a full column rank
for m ~n - 1 if and only if D(s) and N(s) are coprimeo Hence we haveestablished
the foIlowing theoremo s
(9-41 b)
(9A2a)
Theorem 9-10
(9-42b 1
D(s) = Dls)D(s) + Ne(s)N(s)

I-m
(9-46)
(9-43)
rily aIl nonzero. The
:) +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
ent to he find ing of integer

is a topie in number theory
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.
Let C(s) be chosen as
Then the equation in

Theorem 9-11'
Consider the feedback system shown in Figure 9-12 with 9(S) = N(s)/D(s) and
deg N(s)~deg D(s)=n. Then for any D(s) of degree n+m, a strictly proper
compensator C(s) of degree m exists so that the feedback system has transfer
function N(s)Dj l(s)Ne(s) if and only if D(s) and N(s) are coprime and m c.n.
Proof
For a proper 9(S), the solution {Dei, NeJ in (9-44) exists if and only if N(s) and
D(s) are coprime and m c..n -1. However, if m = n -1, the solution is unique,
and there is no guarantee that Dem i=- O, and the compensator may become
improper. If m c.n, the number of unknown {De;, NeJ in (9-44) is larger than
the number of equations. Since the last row of Sm, with m c.n, is linearly
dependent on its previous rows (see Section G-2), we may choose N e ", = O.
For this choice, we have De", = Fn+,JDn- Hence the compensator is strictly
proper. This establishes the theorem.
Q.E.D.
In Theorem 9-11, ir m = n -1 and D(s) = P(s)D(s), where P(s) is an arbitrary
polynomial of degree m, then the unique solution of (9-46) is Dis) = P(s) and
Ne(s) = O. In this case, no pole of the plant is to be altered, and the compensator
is zero. This is a degenerated case. Similarly, in Theorem 9-11', if m = n and
D(s) = P(s)D(s), and if the compensator is required to be strictly proper, then
Ne(s)=O, and the compensator is zero. Clearly, Theorem 9-11' still holds if
the compensator is permitted too be .proper. As discussed in Section 3-6, the
well-posedness problem of feedback systems .should be checked in the designo
In Theorem 9~, 9(s) is strictly proper and C(s) is proper; hence 1 + C( 00 )g(oo) i=
O, and the feedback system is wdl psed. 'In .Theorem 9-11', 9(S) is proper, but
C(s) is strictly proper: hence 1 +C(co)g(co)i=-O, and the feedback system is
again well posed. the design mayinvolve polezer cancel1ations between
D(s) and N(s )Ne(s). H owever these poles are al1assignable by the designer;
hence they will not cause any problem in the designo
Since the square ma

solution exists in (9-4~
sator of degree 1 alwa:
(9-48) may become
the s01ution in (9-48)
or C(s) = (3s + 1)/(-(
function.
Now we choose
The first five rows of t

a ful1 column rank;
De2 = F 4 and N e2 = O
strictly proper transf(
DEsIGN OF COMPENSATORS: UNITY FEEDBACK SYSTEMS
rems by removing the
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
Consider a plant with transfer function

A
g\s)
. this theorem follows

then N n = O. Conse
= Fn+",/D".
Hence if
the compensator C(s)
Q.E.D.
463
+1
+2
s
s_ 2
S2
.2
Let C(s) be chosen as

(9-47 )
Then the equation in (9-44) becomes

th g(s) =N(s)/D(s) and
+ m, a strictly proper
~k system has transfer
coprime and m no
If and only if N(s) and

:he solution is unique,
lensator may become
n (9-44) is larger than
vith m n, is linearly
may choose N cm = O.
Jmpensator is strictly
Q.E.D.
;:re P(s) is an arbitrary
46) is Dc(s) = P(s) and
, and the compensator
em 9-11', ir m = n and
;: strictly proper, then
em 9-11' still holds if
ed in Section 3-6, the
;hecked in the designo
hence 1'+ C( 00 )g(oo) i=
11', g(.s)is proper, but
,e feedback system is
~anceations betv:.een
lable by the designer;
Since the square matrix in (9-48) is nonsingular, for any F, i = 0,1,2,3, a

solution exists in (9-48). In other w.ords, for any three arbitrary poles;a compen
sator of degree 1 always exists. However, for sorne set of chosen poles, the Del in
(9-48) may become zero. For example, if D(s) =2 +2s +0.5s 2 +3s 3 , then
the solution in (9-48) can be computed as Dco = - 0.5, Del = O, N cO = 1, N el = 3,
or C(s) = (3s + 0/( - 0.5). This compensator does not have a proper transfer
function.
N ow we choose
2
C(s) = N co +Nels +N c2 S
2
Dco +Dels +D c2 S
in this case, (c;-44) becomes
-2
2
1
O
O
O 100
1
--------------------O -2
2
1
O
O
1
O
1
O
--------------------O
O -2
2
1
O
O
1
.0
1
=[F o
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
strictly proper transfer function. This confirms with Theorem 9-11'.
464
In Theorem 9-11, n - 1 is the smallest degree which a compensator must

have in order to achieve arbitrary pole placement. In other words, for every
D (s), a compensator C(s) of degree n - 1 always exists so that the unity feedback
system in Figure 9-12 has D(s) as its denominator. lf Df(s) is prechosen, then
it may be possible to find a compensator of degree smaller than 11 - 1 to achieve
the designo For example, if Df(s) is chosen as
D f(S) = 4s 2 + 2s + 1
scribed by the q x 1
in the example, then we have
[DeO
Neo]So
[1
where D(s) the least (
2 4]
and DeO = 1, NeO = 3. In other words, we can find a compensator C(s) =f of

degree O to achieve this particular pole placement. Hence, if the degree of a
compensator is extremely important, we_ may proceed as follows: First, we
form So and check whether the chosen D f(S) of degree n is in the row space
ofS o. lfyes, the design can be completed by using a compensator of degree O.
lfnot, we formSI and increase the degree of Df(s) by one to Df(s) = 15 f (s)(s +k).
Ir Df(s) is in the row space of SI for sorne k, and if the pole s + k of this k is
acceptable, then the design can be accomplished by using a compensator of
degree 1. Ir D(s) is not in the row space of SI' then we must increase m by 1
and repeat the process. For a study of this problem, the reader is referred to
References S97, S184, S21O, and S223.
The larger the degree of a compensator, the larger the number of parameters
available for adjustment to achieve design purposes. Ir the number of para
meters is larger than the minimum required for arbitrary pole placemeni, the
spared parameters can be used to achieve other design objectives such as the
assignment of zeros or the minimization of sensit ivity functions. The interested
reader is referred to References S25 and S83 and Problem 9-29.
In Theorems 9-11 and 9-11', if D(s) is chosen as a Hurwitz polynomial, then
the feedback system is stable. Hence we conc1ude that every plant with a
proper transfer function can be stabilized by using a compensator with a proper
transfer function in the unity feedback configuration shown in Figure 9-12.
5~g-og~e-li'1pijJr~
or sfirugoe-QlLf[Put case.
D(s)
N(s)
where D are constar

find a compensator
number of poles of 1
assigned. Furtherml
small as possible. FI
Hence the transfer rr

equal to
Let us write the com~
1
C(s)=-I
DJ5)
with
where D are scalars
19-50) and (9-53) i'1tc (
In this section vveiscss i.i-lC.s5~,;
of compensators to achieve pole placement for single-input multiple-output and

multiple-input single-output systems. The general case will be postponed to
the next subsection.
Consider the feedback system shown in Figure 9-13(a). The plant is de
rE; f(s) = N(s),

= [De(s
Because N(s) and N e (
and N(s)Ne(s) is a q x
Hence the problem of

This equatiori is agem
to (9-40), we shall trar:
Di(' .
(al
Figure 9-13
(b)
Unity reedbacksystemswith single-input or single-output plant.
Note lhat, beca use of pos
compensator must
her words, for every
at the unity feedback
:.'1) is prechosen, then
than n -1 to achieve
465
scribed by the q x 1 proper rational matrix

N'I(s)
D'I(s)
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
ions. The interested
:1 9-29.
litz polynomial, then
every p\ant with a

:nsator with a proper
own in Figure 9-12.
D(s) = Do +DIs +D 2 s2 +
N(s) = No + NIs + N 2 s 2 +
D n #=0
(9-51 )
where Di are constants and Ni are q x 1 constant vectors. The problem is to

find a compensator with a proper transfer matrix of degree m so that n +m
number of poles of the feedback system in Figure 9-13(a) can be arbitrarily
assigned. Furthermore, the degree m of the compensator is required to be as
small as possible. From Figure 9-13(a) we have
u(s) = (1
+ 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)
Let us write the compensator C(s) as

(9-53)
= Deo + Deis +
NcCs) = Neo + N eI S +
with
De(s)
+ Demsm
+ Nemsrn
(9-54 )
where Di are scalars and N ci are 1 x q constant vectors. The substitution of

(9-50) and (9-53) into (9-52) yields
we d iscuss the design

multiple-output and
'Viii be postponed to
1). The plant is de-
+ DlIs n
+ Nnsn
Gfes) =
N(s)D- I(s)[1 '+ D e- I (s)Ne(s)N(s)D - I (s)] - ID; I (s)Ne(s)
+ NcCs)N(s)] -1 N(s)NcCs)
= [DcCs)D(s)
(9-55)
Because N(s) and Ne(s) are q x 1 and 1 x q vectors, Ne(s)N(s) is a 1 x 1 matrix

and N(s)Ne(s) is a q x q matrix. Hence Gfes) is a q x q rational matrix. Define 6
Df(s) = DcCs)D(s) + Ne(s)N(s)
(9-56)
Hence the problemof pole placement reduces to the solving of Equation (9-56).
'This equationis a generalization ofthe Diophantineequation in (9-40). Similar
. to (9-40), we shall translate it into a set of linear algebraic equations. Let
(9-57)
(b)
~-Olltpllt plan!.
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 =
1, this theorem reduces to Theorems 9-11 and 9-1\' and
l'
is equal.to
11.
Proof
The design problen

application of Theo
there exists a set of
if the matrix Sm has
definition of r and t
Hence the conditior

This is the case if al
(Corollary G-14). 1
any {F;}.
lf (;(s) is strictly
solved as De", = FIl + m
is proper and if m =
m;::: v, then the last q
Consequently, we ID;
Fn+mlD". Hence the
proof of this theorem
The application o
in (9-50) have no COI
form S", by using the
independent rows in
algorithm. Once v is
(;(s) is strictly proper.
Its solution yields th
solution of (9-58) can
v'/t; fpn.asizt: ((le

ability index, the con
hinges on the search
S,. is a (l +q)(m + 1)
we need (1 +q)(m + 1
may start from the SI
note that aH remarb
For example, for a giv,
degree smaller than v
of a compensator, it
other design objectivt
Dual to Theorem
system shown in Figl
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)
vs rormed from {Ni}

and q N rows. The
is the shifting to the
latrix Sm is clearly a
>Ie, the row-searching
in Theorem G-13, all
row's will be linearly
~ number of linearly
;;ause of the structure
;hthatro~"I:::;'" :::;
:x of G(.~). Then \' is
1 of (;(s) = A -1(s)8(s)
rvability index ofany
:h the plant described

(s) = N(s)D -I(S) with
exists a 1 x q proper
,De(s)= m so that the
(s) ir and only if D(s)
v is the row index of
~ation of C(s), or the
G(s) = A -1 (s)B(s) with
.\ to 11.
+ 1) + I (q -
Hence the condition for Sm to have a rank of n + m + 1 is n = ?~ o (q - r).

This is the case if and only if D(s) and N(s) are right coprime and m;? v - 1
(Coro11ary G-14). This establishes the existence of Dei and N ei in (9-58) for
any {FJ
lf C(s) is strictly proper, then N" = O. Consequently, Dem in (9-58) can be
solved as D em =F"+,,jD n, and the compensator C(s) in (9-53) is proper. lf (;(s)
is proper and if m = v -1, then Dem in (9-58) may become zero. However, ir
m;? v, then the last q rows of SOl are linearly dependent on their previous rows.
Consequently, we may choose N em = O. For this choice, D em becomes D em =
F"+,,jDI!" Hence the compensator C(s) is strictly proper. This completes the
Q.E.D.
The application of this theorem is very simple. If D(s), N (s), N 2(S), ... , N q(s)
in (9-50) have no common roots, then D(s) and N(s) are right coprime. We
form SOl by using the coefficients of D(s) and N(s), and then search its linearly
independent rows in order from top to bottom by using the row-searching
algorithm. Once v is found, we choose n + v - 1 poles ror the overa11 systems ir
C(s) is strictly proper. From these poles, we form Equation (9-58) with m = \'-1.
Its solution yields the compensator. We note that in hand calculation. the
solution of (9-58) can be obtained by continuing the row searching to
We emphasize that aithougnTneorem 'J-l L. iS stalCG in terms o the observ

ability index, the concept is not needed in applying the theorem. The theorem
hinges on the search of m so that Sm in (9-59) has a full column rank. Since
Sm is a (l +q)(m +1) x (n +m +1) matrix, in order to have a full column rank,
we need (l +q)(m + 1) ;?(n + m + 1) or m ;?(n/q) - 1. Hence in searching m, we
may start from the smallest integer m ;?(n/q) - 1 rather than from m = O. We
note that a11 remarks for the .single-variable case are equa11y applicable here.
For example, for a given set ofpoles, it may be possible to find a compensator ofa
degree smaller than v- 1tb achieve pole placement. By increasing the degree
of a compensator, it is possibJe to achieve, in addition to the pole placement,
other dcsign objectives.
.
Dual to Theorem 9-12, we have the fo11owing theorem ror the feedback
system shown in Figure 9-13(b).
468
LINEAR TI\1E-INVARIANT COMPOSITE SYSTEMS
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
then it can be readil

Every 1 x por q xiI
(;(s)
tjJ - 1 (s)N(s)N e(.
matrices, then C(s) i:

elements gij(S) of G(.I
Consider the cycl
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
where 6.(. ) denotes tt
Lel [A, B, C, E} be an rre
polynomial of A. Comp.
matrix in Prob\em 2-45.
9 Compare this theorem w
h the plant described

with
exists a p x 1 proper
Dc(s) = m so that the
.(s) if and only if D(s)
. the column index of
~ation of C(s), or the
:torization of C(s) =
~s)=D-(s)N(s)
denominator of all elements of C(s) or, equivalently, of all minors of order 1

of C(S)8 Let Li(s) be the characteristic polynomial of C(s) defined as tht least
common denominator of all minors of orders 1,2, ... , min (p, q), of C(.}) (see
Definition 6-1). Clearly, in general, we have Li(s) = lf(s)h(s) for sorne poly
nomial hes). Ir Li(s) = lf(s) k for sorne constant k, then C(s) is called a cyclic
rational matrix. For example, if
l
S
(9-61 )
N(s), rather than the

idirectly by taking its
Iy as fol!ows: Using
), we form the linear
Gz(s)=
+ 1
Fo
F
Fz
(9-62)
Fn + m
Theorems G-14 and

Tm in order fram lelt
~olumns of T" are all
Ji. will be called the
ex of any irreducible
:olumn-reduced A(s)
he proof of Theorem
repeated.
In this section, the
will be exten1ed to
e<;:ial c1ass ofrational
:neral case.
be the least common
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)
where Li(') denotes the characteristic polynomial of a rational matrix. 9
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
Consider the feedba

by a q x p cycIic stri,
compensator is aSSUI
C(s) of degree m. If
of the unity feedbac
respectively. the cor
realization of G(.s) 01
G(s) =_1 N(s) =_1

l{I(s)
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
Then Theorem 9-12

C(s) = D c- l(s)N..{s) wi
alln +m roots of
can be arbitrary assi

of the feedback systel
for all i = 1,2, ... , m

This implies that l{I(s) and N(s)t, i = 1, 2, ... , q, in
'\. _ ~
GI.\)t - ,',
'I'(s)
N ,IS)t]
N 2 (s)t 1
:
Nq(s)t l
have no common factor. Hence l{I(s) is the charactcristic polynomial of G(s)t.

This establishes 8[ G(s)] = 8[ G(s)t lJ for the case where aH roots of l{I(s) are real.
ir sorne of ;'j, i = 1, 2, ... , in, are not real, the preceding proof stil! holds if we
use N(A} +N(},t). This completes the proof of .1[G(s)J =.1[G(s)tJ.J:"e
other part of the theorem can be similarly proved.
Now we show that (9-63) holds for almost al! tI in ~p. For easy visualization,
we assume p = 2 and consider only tI with IIt l l1 2 = 1. Then from the preceding
proof, we see that at most 2m points on the unit circle of ~ 2 will not meet (9-63),
the rest ofthe unit circle (infinitely many points) will meet (9-63). Hence almost
every vector tI in ~2 has the property 8[(;(S)] =8[G(s)t l J. Hence ifa vector
tI is. chosen arbitrarily or generated randomly, th~ probability of having
8[G(s)] == 8[G(s)tlJ is almost qual to 1. This establishes the theorem. Q.E.D.
The cy~licity of (;(s) is essenltal in this theorem. Ir (;(5)' is not cyclic, the
thei-em d~esnot h61d in general (tr"y G 4 (s) in Problem 9'-13).. UsingTheorem
9-13, we can now extend the design procedure in Theorems9-12 and 9-12' to
cyclic rational matrices.
i:
Figure 9-14
Design (
DESIGN Of COMPENSATORS: UNITY FEEDBACK SYSTEMS
471
Theorem 9-14
:nts of G(S); N(s} is a

s). We assume that
ption of (;(s) implies
)ts of Jt(s). First we
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
u\! space of N(A.) has

ds, the set of vectors t
)t = O isa linear space
will be discussed. we
L, is at most a straight
:s on which N(A.i)t = O
aperty
= 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
:< ar easy visualization,

~n from the preceding
2 will not meet (9-63),
(9-63). Hence almost
1]. Hence if a vector
lrobability of having
the thearem. Q.E. D.
G(s) is not cyc1ic, the

-i3). Using Theorem
~ms 9-12 and 9-12' to
.J
lb)
Figure 9-14
Design of compensators for plant with cyclic proper rational matrix.
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
G feS) = N(s)D- (s)[l + D,- (s)N,(s)N(s)D- (s)] - D c- (s)NcCs)

= N(s)[D,(s)D(s) + Nc(s)N(s)] - Nc(s)
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
We show that the roc
distinct for almost all ]
where Gi, i = 1, 2, ... ,n

tion of 3.(s) with respe(
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)
With this theorem,

placement for general
gajn output feedback to
o The proof is identical lo lh
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)
G(s) =[1 +G(s)K]-IG(s)=G(s)[1 +KG(S)]-l
'lAs) into (9-64) yields

1D c- 1(s)Nc(s)
is proper (strictly proper) and cyclic.

(9-65)
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"
s equal to the one of
d B(s) are left coprime

f G(s) is eq ual to the
G(s)t l =A -1(s)B(s)tl'
It A(s) and B(s)t l are

;0 equal to the largest
ity index of (;(s) and
.v - 1, where v is the
(s) is cyclic, we have
le part ofthe theorem.
Q.E.D.
. to y is of the form
~(s) and N~(s) in this
.lthough the degree of
o achieve pole place
n other words, it may
1 min (.t - 1, v -1) to
)per rational matrix.
lem.
of compensators fOl"
ts of two steps: First
then apply Theorem
) be its characteristic
ct, thell (;(s) is cyclic.
ltS of G(s). Then we
ts of ~(s) are distinct,
1st contain every root
y have repeated poles.
s a sufficierit but not
ill be used to establish
sformed :intoa cyclic
utput to the input, as
where a, i = 1, 2, ... , n, are functions of all elements, kij, of K. The differentia
tion of ~(s) with respect to s yields
3.'(s) = naos"-l + (n - l )als"- Z + ... + a,,- 1
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.
- - - - - - - - - --- - - - - - -n- - - - - - - - - - - - - - - - - -ao

nao
tl -
(See Appendix G). We note that y(kij) is generally a nonhomogeneous poly

nomial of kij' There is a total of p x q number of kij in K. Hence we may think
of {kij} as a vector in the (p x q)-dimensional real vector space ~PXq It is
c1ear that the solution, kj, of y(kj) = O is a subset in ~ p x q. In other words, for
almost every kj, we have y(kij) f- O. Consequently, all roots of ~(s) are distinct.
(This is similar to tha!.for ?J.rnast
~dl C~:., C~~~, ~lr.-cl
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.
Wth ths theorem, the desgn of compensator to acheve arbitrary pole

. placement foro general (;(s) becomes obvious. We first introduce a constant
. gain outP\lt feedback to make G(s) = [1 + G(s)K] -IG(S) cyclic. We then apply
~.
10
The proof is identical lo the one of Theorem 7-6.
474
Theorem 9-14 to design a compensator;.. C(s) of degree m=?:min(1-1,v-l)

or In =?: min (1, v) depending on whether C(s) is strictly proper or proper, where
Ji. and vare, respectively, controlIability and observability indices of G(s).
Hence all poles of the feedback sy;tem in Figure 9-15(a) can be arbitrarily
assigned.
The feedback configuration in Figure 9-15(a) is not exactly a unity feedback
system. HO';Yever, if we are concerned only with the poles of CAs), not the
structure of GAs), we may combine the parallel connection of C(s) and K into
C(s) and the configuration reduces to the unity feedback system shown in
Figure 9-15(b).
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(~
From (9-66) and (9-67),

have the same set of po!
assigned by a proper ch
C(s) + K, the poles of n
Now we c1aim that d,
be an irreducible realiza
x = Ax +Bu, y = ex + (J
Hence we have deg de
proved by using Theo rel
implies that the parallel (
Hence we ha ve deg det
What remains to b
indices of C(s) and (;(s)
for almost every consta
N(s) and D(s) are right c
lability index of any irree
degree of D(s) (see Theol
'" = l\!(s)lDJG(s)
(;(5)
----------l
I +
X(s)[DI
Hence Theorem G-S imp

column degrees of D(s) al
for all K if N(s)D -les) is
proper. .Hence the contr
facl <;an also be proved b
can besimilarly proved.
I
I
I
-.J
(al
Design of compensator.
!r
Since N(s) and D(s) are

and Ves) such that X(s)O
Figure 9-15
By a similar manipulal
can be computed as
Theorem 9-16
11
given by
(b)
The tr:ansfer matrix G

Unlike the cases of single
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) = [1 + G(s)C(s)] - 1 G(s)C(s)

= [1 + (1 + G(s)K) - 1G(s)C(s)] - 1(1 + G(s) K) - 1G(s)C(s)
= [1 + G(s)K + G(s)C(s)]- 1G(s)C(s)
The substitution of G(s)=D- 1(s)N(s) and C(S)=D;I(S)t 1 N e (s) yields
GJI (S) =
{D - 1 (s)[D(s)De(s) + N(s)KDe(s)+ N (s)t 1 Nc(s)]D c' 1 (s)} -1

D-I(s)N(s)D e- 1 t 1 N e(s)
[D(s)De(s) + N(s)KDe(s) + N(s)t 1 Nb)]' 1 N(S)tl Nls)
(9-66)
X
By a similar manipulation, the transfer matrix of the system in Figure 9-1S(b)

can be computed as
ith the plant described
)f degree n. The com
)roper) rational matrix
11
;en all n+m poles of
'arily assigned, where J1
ability indice~ of any
row index of G(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)
[D(s)De(s) + N(s)KDe(s) + N(s)t 1Ne(s)] -1 N(s)[t 1NAs) + KDe(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)
and Y(s) such that X(s)D(s) + Y(s)N(s) = I (Theorem G-S). We modify it as
X(s)[D(s) + KN(s)] + [Y(s) - X(s)K]N(s) = I
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 2. Find a",tI so that ~[G(s)] = ~[G(S)tI].
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
The minimal polynon

teristic polynomial of
(S3
+ W [(s-+
1
~(s) =
(53 +3)1 [(s'
Fo +F IS + F lSl + ... +FIl+v_ISn+v-I
Step 5. Solve Dci and N ci from
(;"3+W
[DcoN co : ... : Dc(v-I)Nc(V-ll]Sv-1 =[F o F I
F1
[s(s-
Hence the characteri~

minimal polynomial.
two outputs; hence t
observability indexo \
Then the compensator is given by

C(s) =tlC(s) +K
whereC(s)=(Dco+DcS+'" +Dc(v- I)SV-I )-I[N co + Ncs+ ... + NC(V-llSV-I]

111
A
In step 2, if we find a t z such that ~[G(s)] = ~[tzG(s)], then we must modify

steps 3 and 5. In step 3, we form T m as in (9-62) and search its linearly inde
pendent eolumns in order from left to right. Let J1 be the least integer such
that Tt-I has a full row rank. The integer J1 is called the column index or the
controllability indexo In step 5, we solve C(s) from the linear algebraic equation
in (9-62). Then the compensator is given by C(s) = C(s)t 1 + K.
Consider a q xp proper rational matrix G(s) of degree n. Ir its control
lability and observability indices are roughly of equal lengths, then we have
J1 znjp and \' znjq (see Equations (5-31) and (5-51)). Hence if p "2:.q, we use
t 1 ; otherwise, use t l' By so choosing, we may achieve min {J1- 1, v -l} without
computing explicitly both~L and V.
The matrix T o has a fu

iSJ1-1=O. Let
vi dei'e .l.ll... s Lqjtt'SCC i Pl~

poles are eh osen arbitr
Example 2
~f(S)
Consider
(;(s)
l ~ ~1
The solutions of
= si,
.O
O s
lt is a noncyclie rational matrix of degree3. Rather arbitrarily, we choose the

constant output feedback gain as
are D co = 1, Mo ==1, N~
9-IS(b) is given by .
C(s) = C(s)t 2 -t
We compute
:ver, from the design

See a1so Problem 9-16.
(;(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
The minimal polynomi,l.l of (;(s) is clearly equal to S3 + 3. To find the charac

teristic po1ynomia1 of (;(s), we compute its minors of order 2 as
(S3
+ W [(s + 1)( -2s) - ses + 1)( -2)] =0

1
[2
(S3+W (s+1)(s -s+1)+2 =(S3+W(S +3)=(S3+3)
('
s'
1
s
W[s(s+1)(s2- s + 1)+2s]= 3-3)2 S(S3+ 3 )=_,-
+
(s +
(s' +3)
1
2"-
Hence the characteristic polynomial of (;(s) is S3 + 3, which is equal to the

minima1 po1ynomial. Hence (;(s) is cyclic. The system has three inputs and
two outputs; hence the controllability index is probab1y smaller than the
observability indexo We choose t 2 = [1 O]. Then we have
...
1
t 2 G(s)=-,-[s+1
s(s+l)
1]
s' +3
F'I+V-]
;+ ... + Nc(v-IlS .-1]

I
], then we must modify

earch its linear1y inde
: the 1east integer such
le column index or the
lear a1gebraic equation
t 2 + K.
:gree n. Ir its control
1engths, then we have
Hence if p '2:q, we use
n {J1-1, v -l} without
and L1[ (;(s)] = L1[t 2 (;(s)]. We form, as in (9-62),
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
wh;re [be; superscrlfils denote the componems o Nc(.s).iile n +UJ. - l} =

poles are chosen arbitrarily as -1, -1, and - 2. Hcnce we have
L1 f (s) =(s + lf(s + 2) = 2 + 5s + 4s 2 +S3
The solutions of
Jitrarily, wechoose the
Hence the compensator in Figure
-1] r2 ., . , 1J
O = 3
478
We shall do the same

Let G(5) = N(s)D
It is a compensator of degree O. As a check, we compute

A
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).
Multivariable case: Arbitrary denominator matrix assignment. 1 3
In this subsection, we study the design of compensators to achieve arbitrary

denominator matrices. If we can assign an entire denominator matrix, then
certainly we can achieve arbitrary pole assignment. Hence this problem
accomplishes more than the pole assignment discussed in the previous subsec
tion. Consequentiy, the degrees of compensators required for arbitrary denomi
nator matrix are generally much larger than the ones required for pole assign
mento
Consider the unity feedback system in Figure 9-16. The plant is described
by a q x p proper rational matrix G(s). The compensator to be designed is
required to have a p x q proper rational matrix C(s). Let G (s) be the transfer
matrix of the overall system. Then we have
G (s) = [1 + G(s)C(s)] - 1G(s)C(s)
= G(s)C(s)[1 + G(S)C(S)]-1
= G(s)[1 + C(s)G(s)] - 1C(S)
(9-68)
The first equality is obtained from y(s) = G(s)C(s)[r(s) - y(s)]; the second one
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.
Follows closely References 549 and 5237.
Define the polynomis
Then we have G(s) =

Given D(s) and N(s) a
This is the matrix ver
(9-40), we shall transla
We also write
and
The substitution ofthe
where
Do
No
S
=
ni
Multivariable feedback system.
[ O
O
D1
NI
.
.
Do'"
No'"
[
l'
----- --------l
! - - - - --------
'"
...
The matrix Sm has m +

from Di and q rows fon
N rows. This matrix Sn
that if N(s)D- 1 (s) is pn
independent of their pi
may be lineariy depenc
linearly dependent N ro
ofS m ,. we have ro :$1'1 ::'S:
14
Figure 9-16
N(s)I
= N(s)[
To conclude this subsection, we menton that although we can place arbitrary

the poles, the resulting G(s) is always cyclic. 12 In other words, the structure
of GAs) is restricted; we have no control over it. This restriction will be re
moved in the next subsection.
12
G/(s) =
(s+1)(s+2)
The integer J1 is the largest .

single-input case, however ,
479
We shall do the same for the multivariable case.

Let G(s) = N(s)D- I(S) and let C(s) = D; l(s)N e(s). Then (9-68) implies
3s +2)
G(s) =
N(s)D- 1 (s)[I + D; l(s)N e(s)N(s)D- 1 (s)]

=N(s)[De(s)D(s) +Ne(s)N(s)]-INclS)
ic (verify it).
-1
D e- 1 (s)N e(s)
(9-69)
Define the polynomial matrix
.we can place arbitrary

:r words, the structure
restriction wil\ be re
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-
D(s)=D o +D 1 s + ... +Dl'sl'

N(s)=N o +N 1 s+'" +Nl'sl'
'S
The plant is described

ator to be designed is
~t G(s) be the transfer
(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)
The substitution of these into (9-70) yields

[Deo
Neo: Dc
Nc: ... : D em
Nem]Sm=[F o F 1
F2
...
Fmw ] ~ F
(9-74)
(9-68)
. y(s)]; the secpnd one

= C(s)[r(s) - G(s)u(s)].
le single-variable case,
;ystem, then the design
ns. In order to avoid
the zeros unspecified.
,tate feedback.
where
Do
DI
No
S,.,= [
~ ~o
--
l-~
DI'
NI'
NI
--
:::
~1'-1 ~I' ~
1'- 1
l'
...
.,.
:::
}ro(numberofdependent
}r/.o WS)
(9-75)
-----~ - -e:---- ---:--~:~:-~:k

------------------------------
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
LINEAR TIME-INV ARIANT COMPOSITE SYSTEMS
The rank of S,. _1 is
See Equation (G-70). Then we have
(9-76),
rank Sm = (m + l)p +
(q -1)
for m < v-1
j=O
v- 1
and
rankS m=(m +l)p
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
o fes) = 0c(s)D(s) + Nc(s)N(s)

if and only if D(s) and N(s) are right coprime and O(s) is column reduced. 15
Proof
Let /l =max {/l, i = 1,2, ... , p}. Since C(s) is proper, the column degrees of
N(s) are equal to or smaller than the corresponding column degrees of D(s).
Consequently, the matrix
S _[0 0
0- No
01
NI
/L-1
N/L-I
O/L]
(/l - /l) = P/l -
i= 1
This equality holds if

O(s) and N(s) are rig
rank of Sk and the n
from v - 1, we conc1'
only if D(s) and N(s).
If Df(s) is of colt
zero columns in the
zero columns coincid
must be inside the ro'
(9-74), or equivalentl
completes the proof e
This theorem stal
meet (9-70), but state~
To study this questic
proper or proper. Th
9-11 and 9-11'). Ho
coefficient matrix D/L (
Before proceeding
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
exists and is nonsing

meet
This theorem reduces to Theorem 9-10 for the single variable case. In the single-variable case.
we have J1 = v=degD(s) = 11, and the search of the row 'index becomesunnecessary.
if and only if D(s) and
L
(;;1
15
(9-77 )
(/l +v)p-(P/l-
/lJ=vp
/li
i=l
The rank of S"-1 is clearly equal to the rank of S"-I'

(9-76),
Hence we have, from
rankS"_l =rankS"_I =vp +n
(9-78)
A comparison of (9-78) with (9-77) reveals immediately that

column rank if and only if
for m?:. v - 1
481
S,. _1
has a full
(9-76 )
p
the row index of G(s).

on of G(s) = A - 1 (s)B(s)
y index of any irreduci
:m6-6.
ion G(s) = N(s)D -1(S).

let v be the l'ow index
esm +fl, i = 1,2, .. . ,p,
to rneet
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
:he column degrees of

llumn degrees of D(s).
]
e matrix SI' sorne new
,Iumn; however, sorne
1mber ofzero columns
in Si' i =2,3,.... Let

;. Since the number of
,,_ 1 is equal to
(9-77 )
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
H(s) = diag {Sl'l, Sil',

He (s) = diag {s'" I , s"'z,
, 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)
exists and is nonsingular, there exists proper (strictly proper) De-:-1(s)Ne(s) to

meet
(9-82)
In the single-variable case,

.les unnecessary.
ir and only if D(s) and N(s) are right coprime and D(s) is column reduced.
482
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)
Dc(s) = He(s)[DclJ + DcI(s)J
Nc(s) = He(s)[NclJ + NcI(s)J

D [(s) = He(s)[D [h + D [1 (s)] H(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.
proper) rational mal

feedback system in I
are right coprime ano
This theorem fo1l
(9-70). Several rem
the case where G(S)
D [(s) = P(s)D(s), thel
In other words, if th(
compensator C(s) = ,
remark concerns the
D [(s) and between 1
Nc(s) and Dc(s) in q
tions involve the pol(
these cancel1ations al
and Nc(s) have a COI
Hence the poIe-zero
assignable poles. Th
system. Since (;(s) is,
I. Hence the unity fe
if Dc(s) and Nc(s) are \(:
which is much larger t
We state the dual
Corollary 9-19
~onsider a q x p stric
G(s) =D-l(S)N(s). L<
be thecolumn index 01
Ir mi ;:::l - 1 (mi ;:::Jl) fe
lim diag(s\l,.
s ....
00
exists !:-tnd lS llf)7lsi '1';

(strictly proper) ratio~
r to y in Figure 9-1 7 i,
are left coprime and [
1
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
lim Ht 1 (s)D[(s)H- 1(s)=J
s- 00
exists and is nonsingular, thereexists a compensator with a p x q proper (strictly
483
>w equation of (9-74):
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)
)per rational matrices.
matrix of D(s) and is
) into (9-82) yields, at
(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.
(;(.1') with the fraction

egrees of D(s) and let v
,e mi, i= 1,2, ... ,p. If
)perty that
'-'---------------------~
----~---_._-----------'
a p x q proper (strictly
Figure .9-17Unity feedback system.
..: ,
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
Gf(s) = [1 + D- l (s)N(s)N c(s)D c- l (s)] -1 D- l (s)N(s)N c(s)D; l(S)

=
Dc(s)[D(s)Dls)
+ N(s)Nc(s)] -1 N(s)N c(s)D c- l (s)
(9-85)
Define
(9-86)
Then
G(s) becomes
G(s) =
Dc(s)D l(s)[D(s) - D(s)Dc(s)]D; l(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.
We see that Dc(s) is :

Hence if G(s) s prc
compensator.
If we choose mi =
D
then the compensato
Example 3
We gve an example to iJlustrate the appicaton o Theorem 9-i9. o....:onsiuer

the proper rational matrix
~ = N(s)D- l (s) = [S2 O+ 1

G(s)
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
Then the compensator is the solution of
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
s)D c- 1 (s) into the first
;)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
--- 0- ---6 ---0- -- -( ---
O -1
1
O
O
1
O
O
O
O
1
1
O
1
O
tJ
[~ ~ ~ ~ ~ ~ ~ ~J
which yields
gure 9-17 is equal to

1ges on solving (9-86).
rime and row reduced
rheorem 9-18 can be
uation (9-86) can also
[)(s) and N(s) to form
independent columns
: all N eolumns in the
1 the colmn index of
to the largest column
me fraction of (;(s) =
lex of any irreducible
to the one ofTheorem
_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
then the compensator can be computed as
rs2 +3
DJs)=l O
1)/3J
4~s s- +s + 1
NJs)
=[
4S +4
O
-4(s +
1)/3J
2(s + 1)
and D c- l(s)Nc(s) is strictly proper. The degree of Dc(s) is equal to 4.
lIi
orem 9-19. Consider
O
S2
-1
J-l
ows n order from top

lat v= 2. Clearly we
[O30J
2
t - ' s 2'
[1O 0J1 s
Remarks are in order regarding the design of arbitrary assignment of poles

and the design of arbitrary assignment of denominator matrices. Ir (;(s) is
strictly proper, the minimal degree of compensator for the former is
min (J1-1, v -1) (Theorem 9-16); whereas the minimal degree for the latter is
p(v -1) (Theorem 9-18) or q(J1-1) (Theorem 9-19). The design ofpole assign
ment assigns only the poles or the determinant of the denominator matrix and
always yields a eyc!icoverall transfer matrix. The design of denominator
matrix yields generaHy a noncyclie overall transfer matrix. if the diseussion
of the cyclic and noncyc!ic designs in the multivariable case of Section 7-3 is
applicable here, the transient responses of the, system obtained by the pole
ptacement design areprobablyworse than those of the system obtained by the
denominator matrix design with the same set of poles. A detailed comparison
of these two designs is not available at present.
486
Decoupling. The results in Theorem 9-19 may be used to design a unity

feedback system so that its transfer matrix GJ(s) is diagonal and nonsingular.
Such a sxstem is said to be decoupled (see Definition 7-1). If the plant transfer
matrix G(s) = N(s)D-1(S) is square and nonsingular, we may choose DJ(s) =
DJ(s)N(s), where fiJ(s) is diagonal, in Theorem 9-19. Then the overall transfer
matrix in Figure 9-16 is
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)
Now if the degree of compensator is sufficiently large. we may be able to choose

a Nc(s) which is diagonal. In this case, G(s) becomes diagonal and the system
is decoupled. This design is achieved by canceling N(s) and may involve
undesirable pole-zero cancellations. Hence this decoupling method is not
always satisfactory.
In the following, we discuss a different method of achieving decoupling. 16
The method involves direct cancellations and is applicable only to stable plants.
Con~ider a p x p proper rational matrix G(s) = N(s)D -l(S). 1t is assumed that
G(s) is nonsingular and det D(s) is a Hurwitz polynomial. Let P(s) in Figure 9-18
be a diagonal rational matrix of the form
any unstable pole-z(

i = 1,2, ... , p, and th
consequently the nur
is proper. Once the
In this design, it is a
both (;((s) and (Ct.(s) -!
Ir det N(s) is not
pole-zera cancellatio
N- 1(5). [Recal! that
poles in 0-1(5).J To
has no open right-ha

to be the least comml
N- 1 (s). Once f3(s) i~
(9-87) is proper. Nc
proper. Let G-1(S)
that if
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)
D(s)N -1(S)P(S) = G-1(S)P(S)
(9-87)
is a proper rational matrix and the overall transfer matrix
G ()-d"
Js -
{131(5)
132(S)
~)_}
lag (;(1(5) +f3l(5)' (;(2(5) +/32(5)"'" iXp(S) +/3p(s)
is BIBO stable. !f det N(s) is H urwitz, the design in Figure 9-18 wil! n.ot
16
Follows Reference S76.
Compcnsator
,----------------l
\
1--,---1----::-"<1
1--'---'
I ---'-.,r
rT------'-'""' I
1 "'1
----." '-/--~l
-l
j,_y:-:1'i~
fol' i = 1,2, ... ,p, thel

ith column of P(s) is 1
ithcolumn of G- \s),
is also equal to the p
G-l(S) is equal to the
zero excess inequa1ity
The poles of GJ(s)
(;((s) are determined fr
GJ(s), we can readily ,
(9-87) is proper and th,
The poles of GJ(5) are
r;.I~~e:.t2d by
l7
Figure 9-18
Decoupling of a plant.
'Ll:.t ,:;1U;1r:;c:'
phase zeros of G(s). ~

cancel1ations. In othe
canceled and should b
Ir a plant is not st
decoupling procedure.
lations, although aH s
C(5), the property of d(
compensator is usuall~
is very sensitive to par
Lel gis) = N(s)( D(s). We
the zero-pole excess of g(i
DESIGN 01' COMPENSATORS: UNITY FEEDBA.CK SYSTEMS
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)
nay be able to choose

gonal and the system
(s) and may involve
pling method is not
1ieving decoupling. 16
only to stable plants.
). It is assumed that
~et P(s) in Figure 9-18
487
any unstable pole-zera cancellations. In this case, we may choose f3(s) = 1,

i = 1, 2, ... ,p, and the poles of Gf(S) becomes o::(s) + L The degree of o:(s) and
consequently the number of poles of GAs) are to be chosen so that C(s) in (9-87)
is praper. Once the poles of G(s) are specified, (;((s) can be readily computed.
In this design, it is also possible to design a stable compensator by requiring
both (;((s) and (o:(s) + 1) to be Hurwitz.
If det N(s) is not Hurwitz, then the preceding design will involve unstable
pole-zero cancellations. These cancel1ations arise from the unstable poles 01
N- 1(s). [Recall that (;(s) is assumed to be stab!e, hence there are no unstable
poles in D- 1 (s).] To avoid these, we choose f3(s) so that the rational matrix
(9-88)
has no open right-half s plane poles. This is accomplished by choosing f3(s)

to be the least common denominator af the unstable poles of the ith column of
N- 1(s). Once P{s) is ehosen, the degree of (;((s) is to be chosen so that C(s) in
(9-87) is praper. Note that although (;(s) is proper, G-1(S) is generally not
proper. Let G- I(S) ~ (}s)/dij(s)). Then fram (9-87), we can readily verify
that if
deg (;(J~) - deg f3(s) ~ max [deg i1}S) - deg dij(s)]
(9-89)
of the compensator
P(s) = P(s). Because
9-18 reduces to a set
system is decoupled.
nsator
(9-87)
9- J 8 win not involvE:
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.
17 Let gIs) = N(s)/D(s),
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
Asymptotic Tracking and Disturbance Rejection
Single-variable case. Consider the design of the control system shown in

Figure 9-19(a). The plant with transfer function g(s) is given, the problem is to
find a compensator with a proper transfer function C(s) so that the feedback
system is asymptotically stable and meets some other specifications. One of
the important specifications is to require the output of the plant y(l) to track
the reference signal r(t). Because of physical limitations, it is not possible to
design a feedback system so that y(t) = r(t) for all t (see Reference S46). The best
we can achieve is that
lim e(t) = lim [r(t) - y(t)] = O

' .... 00
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)
where the polynomiai

and Nw(s) are howeve
equivalent to the assu:
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
Consider the feedbac

-:;omplete!y characteri:
slgnai r(t) ano distur
w(s) = N w(s)/Dw(s). Le
poles of res) and w(s).
(a)
sator with a proper 1

asymptotically stable
rejection.
Proof
(b)
Figure 9-19
Design or control systems.
[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
ASYMPTOTlC TRACKING AND DISTURBANCE REJECTION
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
11 that if r(t) is a step

Iystem, that is, C(s)g(s)
1I in Figure 9-19(a) is
1\:) asyrnptotical1y. In
established.
. The plant is written
t) is injected into the
al. Now in addition
lof w(t) at the output
11ere yw(t) is the output
'Plication of w(t) and
we succeed in finding
l
t for any r(t) and any
(9-90)
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
9-11') or, equivalently, all roots of

D I(s) ~ Dc(s)D(s)4>(s)
+ Ncls)N(s)
have negative real parts.

Now we clairn that the feedback system in Figure 9-20 with C(s) =
N,ls)/Dc(s)cjJ(s) will achieve asyrnptotic tracKmg and disturbance rejection.
Indeed, the output y(t) excited by w(t) and r(t) == O is equal to
)~w(s) = - ew(s) = 1 + Nc(S)~~;~!(S)4>(S)D(S) l{S)

Dc(s)4>(s)
Dcls)4>(s)D(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.
Let y,(t) be the output excited exclusively by r(t). Then we have

r(s) -
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,
for sorne reference

tracking without int
of the internal mode
variations of 9(s) = JI
robusto This will be
1. It is weH known
have a zero stea(
input), the plant
has one poJe (tw
function), then w{
transfer function
designo
2. From (9-95) and
rejection are achi
reference and d
those modes 1/4>(.
transfer functiom
rnent of the inter
the undesirable rr
3. The location of th
case so long as 1/
from w to y. lf il
will not appear in
4. Because of aging
9(s) or, equivalent
This is caHed the
perturbation, eve
permitted so long
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
Asymptotic tracking and disturbanee rejection.
Figure 9-21
Asympll
ASYMPTOTIC TRACKING AND DISTURBANCE REJECTION
re 9-20 with C(s) =

Iisturbance rejection.
I to
(9-95)
:he poles of yw(s) have

s t-+ oo. We see that
on y(t) diminishes as
hen we have
-) r(s)
,)
(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
for sorne reference input and w(t) = 0, it is possible to achieve asymptotic

tracking without introducing an internal model. However, the employment
of the internal model will make the feedback system insensitive to parameter
variations of g(s) = N(s)/D(s) and N c(s)/Dc(s). Hence this design is said to be
robusto This will be expanded in the following remarks:
1. It is well known in the design of unity feedback systems that, in order to

have a zero steady-state error for a step reference input (a ramp reference
input), the plant must be of type 1 (type 2) transfer function, that is, g(s)
has one pole (two poles) at the origino Ir r(t) is a step function (a ramp
function), then we have cfJ(s) = s [cfJ(s) = S2] and (lN(s)g(s) is oftype I (type 2)
transfer function. Hence the well-known result is a special case of this
designo
2. From (9-95) and (9-96), we see that asymptotic tracking and disturbance
rejection are achieved by exact cancellations of the unstable modes of the
reference and disturbance signals. This is accomplished by duplicating
those modes 1N(s) inside the loop so that cfJ(s) appears as numerators of the
transfer functions from r to e and from w to y. In other words, the employ
ment of the internal model is to create the required numerator to cancel
the undesirable modes.
3. The location ofthe internal modeI11cP(s) is not critical in the single-variable
case so long as l/cfJ(s) does not appear in the forward paths from r to e and
from w to y. If it appears in the forward path from, say, w to y, then cfJ(s)
will not appear in tlle numerator of the transfer function from w to y.
4. Because of aging or the variation of the load, the plant transfer function
g(s) or, equivalently, the coefficients of N(s) and D(s) may change with time.
This is called the parameter perturbation. In this design, the parameter
perturbation, even large perturbation, of N(s), D(s), NJs), and Dc(s) are
permitted so long as all roots of
Dc(s)D(s)<jJ(s)
+ 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
Asymptotic tracking with and without an internal model.
492
gain k is placed at the input. Let

N sm +
C(s) = Dems m +
em
6. Unlike De(s), N e(s

not permitted in
are achieved by ,
by the roots of
and expensive te
study in the follo'
assume
.
.
Then the transfer function, 9(s), from r to y is given by

kg(s)C(s)
k(NIINems"+1lI +
g(s) = 1 +g(s)C(s) = (DIID em + NIINem)s"+ 111 +
~
+ NoN eo )
+(DoDeo+NoNeo)
lt is assumed that g(s) IS asymptotical1y stable. Ir r(s) = l/s, the application
and
of the final-value theorem yields
where a ~ O and
perturbation or
from r to e in Fi
lim (r(t)- y(t)) = lim s(r(s) - 9(s)r(s)) = 1 - 9(0)

.,,-+0
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
where all poles e

r(s) and w(s) =0 i:
e(s) = (
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
ASYMPTOTIC TRACKING AND D1STURBANCE REJECTION
-(DoD co +NoNco )
=
l/s, the application
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)
l zero of g(s), then we

: 1, then r(t) -y(t)-->O
to be asymptotically
parameter perturba
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
e internal model l/s.

choose k in (9-97) as
k +terms due to t he poles of g,.o(s)

~
~ )
and 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
We see that k is proportional to the deviation e. lf the cancellalion is

exact, then k is zero; otherwise, k is nonzero.
If k is nonzero, even though very small, then e(t) contains the term
k ea'. If a is positive, this term will approach infinity as t --> 00, and the
output y(t) will not track r[t) asymptotically. If a is zero, then y(t) will
track r(t) but with a finite deviation. If r(s) has a repeated pole (s - at,
with m> 1, and if cancellations are not exact, then asymptotic tracking ;:;
again nol possibie.
This, however, does not mean that Theorem 9-20 is only of theoretical
interest. In the design, if the internal model principIe is not empIoyed, the
resuIt will be even worse. In practice, most r(t) and w(t) are bounded. In
this case, if the internal model principIe is employed, even though there are
errors in implementing <f;(t), the output will still track the reference signal
but with a finite steady-sta: error. The more aceurate the implementation
of <f;(s), the smaller the errO"r.
7. The design of N c(s)/Dc(s) tostabilize the feedback system requires the solving
of the polynomial equation.
.
D(s)=Dc(s)D(s)<f;{s)
+ Nc(s)N(s)
Under the condition ofTheorem 9-20, D(s)<f;(s) and N(s)are coprime. If we
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)
Let D(S)=S"+DI_1S"- + ... +D o and N(s)=NI_IS"- 1 +N I _2 S"- 2 +

... + No Then YlI(S) = N(s)D- 1 (s)u(s) can be realized as
x=
Ax
+ bllu(t)
YlI = ex
Ax
+ bww(t)
y=ex
(9-99)
where A and e are identiGal to the A, e in (9-98). Hence we may combine

(9-98) and (9-99) to yield
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)
with A and e of the forms shown in (6-38). Similarly, y",(s) = D- I(S)W(S)

can be realized as, by using (6-8),
x=
'Multivariable ca~
system shown in Figl
matrix G(s) factored f
q xq and q xp poly
signal r(t) and the q )<
This is the problem e

Following the sil
introduction of an il
There is, however, on
is critical in the m u lt
w(t) by assuming r(t)'
ew(s)
en.(s)
or
+ bllu(t) + b",w(t)
and w(t) is callcd an additiue disturbance. This is the type of disturbancc

most often studied in dynamical eguations.
9. There are two types of specifications in the design of control systems. One
is caled the steady-state performance, the other the transient performance.
The steady-state performance is specified for the response as t -> (jJ. Hence
asymptotic tracking and disturbance rejection belong to this type of
specification. The transient performance is specified for the response right
after the application of the reference signal and disturbance. Typical
specifications are rise time, settling time, and overshoot. They are governed
mainly by the location of the poles of the overall systems or, eguivalently,
the roots of Dc(s)D(s)<f>.(s) + Nc(s)N(s). The relationship between these poles
. and the transient performance is generally complicated. For a discussion
of this problem, seReference S46.
10. To conclude this subsection, we remark that disturbances can be roughly
. classified as nse-type ald waveform-structured disturbances. The former
reglJires the statisticaJ description and is studied in stochastic control
theory. See, for example, Reference S10. The latter is describable by
,.
+ ~-----Jr""
~L1Jll'IlS;
, I
lb---
Figure 9-22
Multiv8
L-__,
Figure 9-23
Placem(
495
ASYMPTOTlC TRACKING AND DISTURBANCE REJECnON
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.
.c/J(s) - 1 and the total

;) +2 degc/J(s)-l. A
'>(s) and Nc(s)jDAs) is

can be considerably
'Multivariable case. Consider the design of the multivariable control

system shown in Figure 9-22. The plant is described by a q x p proper rational
matrix G(s) factored as G(s) = 0-I(s)N(s), where D(s) and N(s) are, respectively,
q x q and q x p polynomial matrices. lt is assumed that the q x 1 reference
signal r(t) and the q x 1 disturbance signal w(t) are modeled as
ance \.I'(t) in the plant
r(s) = O,:-l(s)N,(s)
w(s) = O,~ l(s)N,)s)
and
_ 1s"
+N" _2S " - 2 +
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)
-Iy, Yw(s) = D -1 (s)w(s)

(9-99)
nce we may combine
"
Figure 9-19 can be
fhis equation can be
(9-100)
(9-101 )
= lim (rU) - y(t)) = O
(9-102)
This is the problem of asymptotic tracking and disturbance rejection.

Following the single-variable case, the design will consist of two steps:
introduction of an internal model and stabilization of the feedback system.
There is, however, one important difference: the location of the internal model
is critica! in the multivariable case. To iIIustrate this point, we consider only
w(t) by assuming r(t) = O. The model D:; I(S) of the disturbance w(t) is placed as
shown in Figure 9-23. Let ew(s) be the input of D - 1 (s) as shown. Then we have
ew(s) = w(s) - N(s)O,~ l(s)NAs)D; l(s)D- 1(s)ew(s)
e,)s) = [1
01'
+ N(s)O;;: l(s)N c(s)D; I(S)O-I(SW 1-W(S)
) +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
Placement or internal model for disturbance rejection.
496
if no root of cjJ(s) is a
This equation and (9-101) imply

Y,..(s) = D - l(s)[1 + N(s)D~ l (s)Nc(s)D c- 1 (s)D - I(S)J -1 D;;: l(s)N\vC~)
= Dc(s)[D(s)Dc(s) + N{s)D,~ I (s) Nc(s)J -1 D;;: 1(s)N",{s)
(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)
We see that ir there is no cancellation between Dw(s)D{s)DAs) + N(s)Nc{s) and
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.
has rank p for every !

cjJ{X) =1= O, and rank M
M(A) = rank N(A). H
rank NI()c)=p. By DI
mission zero of G(s).
this part of the theorej
by [cjJ{s)I q ] - I is simila
With this result, \\
compensators to achi
is a generalization of
Consider the feedbacl

pletely characterized
that the reference sig
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
ASYMPTOTlC TRACKING AND DlSTURBANCE REJECTlON
if no root of c/J(s) is a transmission zero of (;(s) and q ~p (p ~ q)

l(s)N .,,(5)
(s)
(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,
for every root l of cP(5)
(9-105)
where {A, B, C, E} is any irreducible realization of (;(s) and n is the dimension

of A 01' the degree of (;(5).
Proof
input of the surnming

~rchange the positions
;)0;;: 1(5) by O,: 1(s)N(s)
s)
1
(s)N,.(5)
N",(s)
(9-104)
:)Oc(5) +N(s)Nc(s) and
'Y the internal model.

nternal model is very
where the disturbance
omrn utative property,
it is not in the forward
tic tracking, we tnust
the reference signal r
after O-l(S) 01' in the
te inside the compen
'Il(s) to achieve distur
) achieve asymptotic
-l(s)N(s), the internal
very restricted. Ir the
polynomial and 1 is a
Furthermore., these
eter perturbations in
1 a(s)I, the design will
given in (9-100) ancl
the unstable poles of
1(s)N(s) = N(S)O-l(S),
(s) are righ t coprime.
~j < min (p, q) 01' rank
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
for every root 1, of <jJ(s)
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
Hence the transfer matrix from r to e is equal to

Ge,.(s) = [1 + D- 1(s)N(s)N c(s)D; 1(s)(<jJ(s)lqr 1] - 1
= {D- 1(s)[D(s)(<jJ(s)l q )D c(s) + N(s)Nc(s)]D; l(s)(<jJ(s)I q )-I} -1
= <jJ (s) DcCs)[<jJ(s)D(s)Dc(s) + N(s)Ne(s)] -1 D(s)
(9-107)
Similarly, the transfer matrix from w to y can be computed as

Gyw(s) = <jJ(s)De(s)[<jJ(s)D(s)DcCs) + N(s)N~(s)] -1
(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)
can be arbitrarily placed. in particular, placed in the open left-half plane by a

proper choice af JDc{s) and NAs) (Theorem 9-1 Oi 9-19). i-lence llii;; wilj
feedback system is asymptotically stable.
Al1 roots of <jJ(s) have zero or positive real parts, hence no cancel1ation
between the roots of <jJ(s)D(s)Dcls) + N(s)Ne(s) and the roots of <jJ(s) will occur.
<';:onsequentIy, <jJ(s) will app~ar as a nUlperator of every element of Ge,.(s) and
Gyw(s). !'i0te that wAe have Gew(s) = - Gyw(s).
With Ge,{s) and Gew(s), the response e(t) due to r(t) and w(t) can be written as
e(s) = Ge,.(s)r(s) + Gew(s)w(s) = Ge,.(s)D; 1(s)N,.(s) + Gew(s)D; l(s)N w(s)
(9-109)
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.
The design consi~

and stabilization of
Hence the total comr
model <jJ(s) appears ~
are called, accordin
modes of r(t) and w
established in Theon
appear in e(t). In ere
zero of the plant G(
observable mode or (
will make the system
tracking and disturb~
C(s) and all of <jJ(s) m
We note that the j
of the reference and
be brought out of the
the blocking zeros of
De(s), N(s) and Ne(s)
(9-107) and (9-108). 1
of robustness, see Re!
We discuss in the
single copy of the ref(
wil1 be general1y not
0-I(s)N(s)NcCs)D e- 1(1
Then we have
e(s) = [1 -+
We assume, without I(
real parts. Since alll
of e(s) have negative
matrix. If D,.(s) and !"
matrix if and only if]
t.hat D(s) = D(s)j[}l(sl 1,
is possible if and only

G(s)C(s)
This is consistent witl
placed at the point wl
this design, we req uil
However, the comput
C(s) is found fora gi
subtracted and move
turbatiol1s inG(s).. H
design, the design' pr
straightforward and ti
is m uch larger.
ASYMPTOTlC TRACKING AND DlSTURBANCE REJECTION
the least common

),:-I(S) and O:I(S). Ir
:, equivalently,
-t ),
of cP(s)
matrix such lhal the

asyrnptotic tracking
-1(s)l q be the internal

to stabilize the unity
(9-107)
ed as
sn
-1
(9-108)
hen (;(s) = 0- 1(s)N(s)

leorern 9-21). Hence
det O(s)(Theorern 9-1)
onsequently the roots
~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
,.(s)D: l(s)N w(s)

(9-109)
d Cew(s), all unstable

:emains in e(s) are all
1).....,,0 as 1-+00. This
Q.E.D.
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
+ G(s)] -lr(S) = [O(s) + N(s)ll D(S)O,-I (s)N,(s)
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
Similar to the single-variable case, the perturbation of the internal model

is not permitted in the robust designo All other remarks in the single-variable
case are equally applicable here and will not be repeated.
Before moving to the next topic, we remark on the necessity of the condition
p ?:.q; that is, the number of plant inputs be larger than or equal to the number
of plant outputs. If q > p, the tandem condltion of <jJ -l(s)l p followed by G(s)
is controllable and observable, although G(s) followed by <jJ -l(s)lq is not
(Theorem 9-21). Consequently, we may introduce the internal model as shown
in Figure 9-25. The transfer matrix from r to e in Figure 9-25 can be computed as
Ge,.(s) = [1
+ 0- 1(s)N(s)<jJ - 1(s)lpNc(s)O; 1(S)J -1
Consider the feedbac
It is a plant with one

No matter how the f~
at the outputs, the in
(9-111 )
A
If we write 0-I(s)N(s)<jJ-1(S)l p =0-1(S)<jJ-l(S)N(s)=(<jJ(S)0(S))-IN(s), then

(9-111) can be simplified as
k
s
u(s) =- +(t
Hence we have, by u:
(9-112 )
This equation is similar to (9-107). Although <jJ(s) in (9-107) may become

blocking zeros of Ge,.(s), the <jJ(s) in (9-112) can never be all blocking zeros of
G.,.(s). This can be seen from the step of writing
0-1(s)N(s)<jJ -1(s)l p = 0-1(S)<jJ -1 (s)IqN(s) = [<jJ(s)O(s)J -1 N(s)

We see that p copies of <jJ(s) is increased to q copies. Because <jJ -1(s)I p fol1owed
by G(s) is controllable and observable, we have deg (O -1(s)N(s)<jJ - 1(s)I p ) =
deg det O(s) + p deg <jJ(s) (Theorem 9-1). The degree of det <jJ(s)O(s) is equal to
q deg <jJ(s) +deg det O(s),
which
is,
because
q > p,
.larger than
deg <jJ(s)0(s))-1 N(s)). Hence <jJ(s)O(s) and N(s) are not left coprime. Con
sequently, not al1 the roots of the determinant of
<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
This argument was provided to the author by Proressor C. A. Desoer.
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
Multivariable feedback system with q> p.
Figure 9-26
Design (
ASYMPTOTIC TRACKING AND DISTURBANCE REJECnON
Consider the feedback system shown in Figure 9-25 with
;)f the internal model

in the single-variable
~ssity
G(J;~,~l
ofthe condition
~ equal to the numper
L+3J
(s)I p fo\lowed by G(s)

j by cP-1(s)I q is not
ernal model as shown
25 can be compu ted as
: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
u(s) =- +(terms with poles inside the open left-half s plane)
p(s)D(s)t 1 N(s), then
Hence we have, by using the final-value theorem,

>(s)4>(s)
(9-112)
lim y(t)= limsG(s)u(s)= G(O)k =

s-o
(9-107) may become

al\ bloeking zeros of
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
)re sorne roots of 4>(s)

cancel\ation between
s of cP(s) wil1 be the
7, asymptotic traeking
rbance rejection.
le necessity of p ?:. q. 18
Static decoupling: Robust and nonrobust designs. Considel'

the
feedback system shown in Figure 9-26 where the plant is completely charac
tel'ized by its q x p proper rational matrix G(s). Let the reference signal r(s) be of
the fornl r(s)== ds- 1, where <rn is an arbitrary q)( i constant -'Ector; ~a~ :J:> ~~'l'-'
reference signals are step functions of various magnitudes. The response at the
output as t-" 00 is caBed the steady-state response. The response right after the
application of r(t) is caBed the transient response. Let the overall transfel'
matrix from r to y be G[(s). Ir G[(s) is BIBO stable, then the steady-state
[.
~
-~-~
Figure 9-26
L.
Design of unity feedback system.
502
LINEAR TIME-INVARIANT COMPOSlTE SYSTEMS
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)
Now if G(O) is diagonal and nonsingular, in particular, a unit matrix, then

the feedback system is said to be seaeically decoupled. Indeed, if G(O) =
diag {h, h z , ... , h q }, then (9-113) implies
lim h(t) =hid i
t-""
lim e(i) = lim

00
t-
00
[r(ei - Y()]
have negative real pa

computed, by using ('
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 -
Since al! the poles of

is nonsingular. Ir s
Now we may design
constant matrix N(O
we may choose
and
G(O) becomes
Hence the system in f

to the one in Figure (
Nc(O) and D (O), then
marize the preceding
Corollary 9-22
Consider the feedback
charact~~ized by its q
~.~:~r:' ~:-f f;J ,,\ ~~.~'
sGe.. (S)~S . . \ = Ge ,(O);1 = 0
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)
Consequently, we have (;(O'....= Iand the feedback system is statically decoupled.

Note that every element of G(s); except those on the diagonal, has s as a zeJ,"o
in its numerator.
Because of the presence of the internal model, the design is robust. .Tl:iat
is, the system remains to be statically decu~led with perturbations, even large
perturbations, of the parameters of C(s) and G(s), so long as the feedback system
remains to be asymptotically stable. In the following, we introduce a design
which is not robust. The design is an extension of the single-variable case
then there exists a com

feedback system is
{A, B, e, E} is any im
We remark once
internal model or.nonl
the remark in (5) on ti
To check the conditi
of (;(s) and is complica
9-22 is eq uivalent to r
ASYMPTOTIC TRACKING AND D1STURBANCE REJECTION
:final-value theorem,
(9-113)
a unit matrix, then

Indeed, if G(O) =
:nce the steady-state

jepends so lely on the
If there is a change
ause responses at al!
{ or as the transient
j no change at other
teady-state responses
'S from the decoupled
transient as well as
; is defined for any
lly for step reference
ery important and is
nd humidity controls
an aircraft at a fixed
the steady state
Hence the design for
IUpling. In this case,
~igure 9~26. We then
feedback system. As
Jn zero of (;(s), then s
"er matrix from r to e.
--> OC! ;
503
discussed in Figure 9-21. In this desilin, no internal model will be employed;

hence we set P = 1 in Figure 9-26. Let G(s) = N(s)D- 1(s) and C(s) = D e- 1(s)N e(s).
We find a C(s) so that all the roots of the determinant of the polynomial matrix
D(s) = Dis)D(s) + Ne(s)N(s)
have negative real parts. The transfer matrix from r to y in Figure 9-26 can be
computed, by using (9-68), as
G (s) = G(s)[I + C(s)G(s)] - lC(s)K
= N(s)[De(s)D(s) + Ne(s)N{s)] -1 Ne(s)K = N(s)Dj 1 (s)Ncls)K
Hence we have
(9-114 )
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 )
and G(O) becomes

G{O) = 1
Hence the system in Figure 9-26 is statically decoupled. This design is, similar
to the one in Figure 9-21, not robusto Ir there are any perturbations in N(O),
NiO) and D(O), then the system will not be statically decoupled. We sum
marize the preceding results as a corollary.
Corollary 9-22
Consider the feedback system shown in Figure 9-26 where the plant is completely
charactelized by its q x p proper rational matrix G(s). If s is not a transmission
zero of G(s) and p Cj or, eqElv2JeEt~.:~/:,
ransfer matrix, G(s),
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).
.
We remark once again that the design can be robust by introducing an

internal model or nonrobust without introducing an internal model. Similar to
the remark in (5) on the single~variable case, the condition is necessary as well.
To check the condition in Corollary 9-22, we must find a copriine fraction
of G(s)and is complicated. If G(s)hasno pole at s =0, thecondition in Corollary
9-22 is equivalent to rank G(O) = q, which can be easily checked.
504
State-variable approach. In this subsection, we shall discuss the design

of robust control systems to achieve asymptotic tracking and disturbance
rejection by using sta te-variable equations. The discussion wil\ be brief
becallse the procedure is similar to the transfer fllnction approach. Consider a
plant described by
X =Ax +Bu + Bww(t)
y =Cx +Eu +Eww(t)
where A, R C. K B"" and E w are. respectively, 11 x n. 11 x p. q x n, q x p. 11 X q. and

q x q constant matrices. It is assllmed that {A, B} is controllable and {A, e}
is observable. It is also assumed that the disturbance signar w{t) is generated by
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
~
Be = block diag {t, t, ... , t}
with
lJ..
This connection was sh(

if and only if p 2: q and r
we reestablish directly t
rank
,11
[ -(
Theorerh 5-7 implies
th~
~
for every S in e Ir fA,

every s in e Ir s is not
rank(05I-Ae)=mq. He
not a root of </>(05), then v
'-------v---'
ro
Design of r
rank V(s)
A e = block diag Ir, r, ... , r}
where
Figure 9-27
Next we show that unde

at the roots of </>(05). We
and e = r - y as shown in Figure 9-27. FEe is an mq-dimensional equation. This

internal model is cal\ed the o5ervocompensator in References S64 to S66. Note
tha.t theoutput of the servocompensator consists of al\ qm numbers of state
variables. This is possible if (he servocompensator- is implemented by using
operational amplifiers; resistors, and capacitors. -Now consider the tandem
connection of the plant fol\owed' by. the servocompensator. Its composite
dynamical equation is, as derivedin (3-63),
V(s) =
The first factor has rank

second factor has rank n
(9-117). Hence Sylvester
rank V(s) 2: (n
. Since V(s) is an
(9-116 )
[~n o
(11
+mq) )
ASYMPTOTIC TRACKING AND DISTURBANCE REJECTlON
505
lall discuss the design

dng and disturbance
Icussion will be brief
lpproach. Consider a
,q X 11, q X p, 11 X q. and
lmtrollable and {A, C}
l1al w(t) is generated by
::sign a robust control

,totically the reference
Figure 9-27 Design of robust system to achieve tracking and disturbance rejcction.
be the minimal poly
,Iane roots of </Jw(s) and

s. The internal model
This connection was shown in Theorem 9-21 to be controllable and observable

if and only if p 2.q and no root of </J(s) is a transmission zero of the plant. Now
we reestablish directly that (9-116) is controllable if
rank [ AI-A
-C
BJ =/l +q
E
for every root ). of </J{s)
(9-117)
Theorem 5-7 implies that (9-116) is controllable if and only if

sI -A
rank V(s) ~ rank [
BcC
O
:
:
si - A c
B
B
= n +mq
cE
for every s in e Ir {A, B} is controllable, we havc rank [s 1- A B] = n for

every s in e Ir s is not an eigenvalue of Ac or, equivalently, a root of P(A), then
rank (sI - A r ) = mq. Hence we concludc from the structure of V(s) that ir <; is
nol a root of </J(s), then we have
rankV(s)=n -!-:ne
Next we show that under the condition of (9-117), Equation (9-118) still holds
at the roots of </J(s). We write
~nsional eq uation. This
nces 564 to 566. Note

11 qm numbers of state
implemented by using
IJ consider the tandem
nsator. Its composite
VI') ~[~ :, ,[ ~AJ['T [~,. ~EJ fn
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
+mq +q) matrix, we conclude that
rank V(s) = n +mq
506

DE~
at every root of 1>(S). Consequently, (9-118) holds at every s in e and (9-116) is

control1able.
[f (9-116) is control1able, then the eigenvalues of the composite system can
be arbitrarily assigned by the state feedback
u=[K KeJ[:J=KX +Kexe

Hence by a proper choice of K and KC' the feedback system in Figure 9-27 can
be stabilized. If the state of the plant is not available, we can design a state
estimator and then apply the feedback gain K at the output ofthe state estimator.
This completes the design of a robust control system to achieve asymptotic
tracking and disturbance rejection. For a more detailed discussion, see Refer
ence S81.
9-7 Design 01' Compensators: Input-Output

Feed back Systems
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
D(s) = De(s)D(s) +L(s)D(s)

D(s) - De(s)D(s)= L(s)D(s)
+ M(s)N(s)
+ M(s)N(s)
Figure 9-28
Input-out
Then 9(5) becomes
Single-variable case. In this section, we study the design problem in

Section 9-5 for a different configuration. Specifical\y, we study the condition
on the plant for the existence of proper compensators in the input-output
feedback system shown in Figure 9-l(b), to achieve arbitrary denominator
matrix. We also study the minimum degrees ofcompensators to achieve the
designo We note that the input-output feedback configuration is deve\oped
frorn the design of state feedback and state estimator -discussed in Chapter 7.
Consider the system in Figure 9-1(b). We study first the case C(s) = 1. The
compensators Co(s) and C I(S) are required to have the same denominator as
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
to (9-122) yields that (9- i

and only if D(s) and N(s)
We compare first thi

order to enSure the prope
in Theorems 9-11 and 9-1
and require the degree of
Theorem 9-23' is chosen 8
proper whether D(s) is (
function of the unity feed
we can control only D r(s
feedback system in Figu
Dr(s) as well as De(s). The
Figures 9-12 and 9-28(a).
and the compensator D'
degree, the former has t~
three sets of parameters D el
.;
DESIGN OF COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS
~ry s in
e and (9-116) is
507
: composite system can
(b)
;tem in Figure 9-27 can

, we can design a state
ut of the state estimator.
to achieye asymptolic
:d discussion, see Refer-
(3)
Figure 9-28
lnpul-olllPlll feedback syslem.
Then 9/(s) becomes
put
(9-123)
the design problem in
we study the condition
)[s in the input-output
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
E(s) ~ D(s) - Dc(s)D(s)
L(s)D(s)
+ M(s)N(s)
(9-124 )
This is the Diophantine equation in (9-40) and, hence, can be translated into a
set of linear algebraic eq uations. Let
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)
Then Equation (9-124) is equivalent to the algebraic equation
[Lo
M o : L
: ... :
L m MmJSm = [E o
...
EIl+mJ ~ E
An input
Although al! comI

serious problem. Th.
function and is, as dis
feedback systems, we
the resulting system t
posed system and sho
(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
E(s) =D(s)-Dc(s)D(s) =S2 +2s +2-(s +1)s(s-2)

2
= 2 + 4s + 2s - S3
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
":-:.':" '.;;:; ~-;<~ ':"/.'. --" ~~:>. ,"

value of De-(s)D(s)D
for S-' 00,
D; (s)D(s)D-(s)
Its solutions are Lo =
pensators are
6, L =
1, M 0=
2 and M
= 6.
Hence the com
Co(s) = L(s) = -6 -s
D/s)
1 +s
and
C(s) = M(s) = - 2 +6s

Dc(s)
1 +s
They are al! proper rational functions.

system is shown in Figure 9-29.
The block diagram of the feedback
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
122) can be arbitrarily

the poles of the com
1
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
An input-output feedback system which is not well posed.
(9-126)
Theorem 9-23
a ful! column rank.

I[he solutions yield im
,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
s)/D(s) = (s -1)/s(s - 2).

arbitrarily Dc(s) = s + 1,
smaller than n +m = 3.
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)
D; '(s)D(s)D - '(s) = 1 + D,- '(s)L(s) + Dc-'(s)M(s)N(s)D- '(s)

=1+Co(s)+C,(s)g(s)
= [2
4 2
value of Dc-'(s)D(s)D-'(s) at s=oo. Using the notations of(9-125), we have,

for s -> 00,
- 1]
Dc- '(s)D(s)D - '(s) -> Dc~' s -m(FII+msn+m

, = 6.
(9-127)
Hence the com
+ Fn+m _ ,Sn+ m-, + ... )s-IID;;'
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
iagram of the feedback

I
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
This corol1ary foIlows directly from Theorem 9-23 by choosing D f(s) =

Dc(s)i5 As) and noting
9fes) = N(s)D 1(s)Dc(s) =
DE~
system to minimize
where q> O is a weigr

function, then the opt
N(s)i5 l(s)D; 1(s)Dc(s) = N(s)i5 l(S)
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
where i5 f (s) is a Hur'

N(s) in the numerator
function 9(s). See R
requires the solution (
Multivariable case
will be extended to the

system shown in Fig
rational matrix (;(s) ,
p x p proper rational
matrix C(s)= Dc-l(s)1
be computed as
Gis) =
=
=
(;(s)[1
N(s)D
N(s)[I
Define
or
Figure 9-30
D
E
Input-out]
lle poles such that the

has transfer function
system to minimize the quadratic performance index

J
11
l by choosing D (s) =
obtained in the state
:he eigenvalues of A or
out affecting N(s). An
7envalues can be con
~ection of the feedback
..eran transfer function
:ablishes essential1y the
. the result in Corollary
have a proper transfer
'or strictly proper plants
)arts equal to zero. In
lesign of state feedback,
heir \eading coefficients
: compensator Co(s) is
be verified from the last

:nce we a\ways have
lied in the state-variable
~m does not arise in the
;eneral than the result of
the concept of coprime
ncepts of controllability
n consists of forming a
mediltely the req uired
requires one similarity
requires one similarity
equation to find a state
lears thaJ the design in
Iy and computationally
= [' {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
= N(s)D- 1(s)[1 + D; 1(s)L(s) + D; 1(s)M(s)N(s)D - 1(s)] -1

= N(s)[Dc(s)D(s) + L(s)D(s) + M(s)N(s)] -1 Dc(s)
(9-129)
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
9-23 can be used in

transfer function g(s) =
ed to design an overall
+ 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
lnput-output feedback system.
(9-130)
(9-131 )
512
Then
DI
C(S) beco mes

(9-132 )
Define
H(s)= diag
[S"I,
He(s) = diag {Sil",
su', ... , SU
S111"
,Slllp}
(9-133 )
(9-134 )
input-olltput feedbae
po sed.
In the design, tht
chosen as
where dei (s) are arbit
Theorem 9 c 24
mi are odd integers, ti
Consider a plant with q x p proper rational matrix C(s) = N(s)D-1(s). Let

J.1i, i = 1,2, ... , p, be the column degrees of O(s), and let v be the row index of
C(s). Let mi ~ v - 1, for i = 1, 2, ... ,p. Then for any D,(s) of row degrees mi
and row red uced, and any D (s) of the property that
In this case, if we che

the d ifficu 1ty of assigr
lim R:-1(s)D(s)H-I(s)=J
Consider a plant wil

J1, i = 1, 2, ... ,p, be tt
Ihen for any De(s) of
D (s) of column deg
compensators with p:
that the feedback sysl
N(s)Dj I(S) if and onl
reduced.
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
Consider the E(s) defined in (9-131). Clearly we have

lim H;I(s)E(s)H-l(s)=J-Iim H;I(s)DcCs)D(s)H-I(s)=J-DchD h (9-136)
s~
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
Let D es) = De(s)D(s).

9-24 are met under thf
and G(s) become
(9-138)
(9-139 )
where D/(s) and Dcl(s) are strictly proper. Theri we have

s-+ co
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)
where J1 = max {,ui, i =

(9-131).yields
[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)
where dei(s) are arbitrary Hurwitz polynomials of degrees mi. Ir some
O!'
al1
mi are odd integers, then we may not be able to assign complex conjugate roots.
Let
v be the row index of
)e(S) of row degrees mi
;(s) = N(s)D -les).
(9-135 )
roper rational matrices

n in Figure 9-30 is well
11y ir D(s) and N(s) are
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
and D" is the column

sts, we conclude from
of row degrees at most
(s)M(s) are proper fol
- D dl D iI is nonsingular
ving that 1 + C o(ro) +
ilar to (9- i 27),
l(s)N(s)D- '(s)
(9-137)
Let Df(s) = De(s)D(s). Clearly the degree requirements ofD(s) in Theorem

9-24 ~re met under the assumptions ofDcCs)and DAs). With this D(s), (9-131)
and G (s) become
and
IH - '(s)[D" + D(s)]-'
(9-140)
and (9-135). Hence the
(9-143)
Q.E.D.
This establishes the corollary.
The application of Theorem 9-24 and its corol1ary is straightfol'W8.cd.

First we use the coefficient matrices 01 D(s) and N(s) to form the matrix Sm
shown in (9-75). We then search the linearly independent rows ofS m in arder
from top to bottom. Let v be the least integer such that aH N rows in the last
block row of S. are linearly dependent. This v is the row index of G(s). For
convenience, we assume mi = v - 1 for aH i in Theorem 9-24. Let
(9-138)
(9-139)
(9-142)
DcCs)[fi(s) - D(s)] = L(s)D(s) + M(s)N(s)

G feS) = N(s)[DcCs)D (s)] -, De(s) = N(s)D '(s)
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
where Il =max {Il, i = 1,2, ... , p}. The substitufion

(9-131) yields
[Lo
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
Using G(s)= N(s)D

be written as
G(s) = N
Ir we define
and
E(s) ~
then G(s) = N(s)D

Theorem 9-24 to acb
ever, the row degrees
degrees of Dc(s) will t
introductionofQ-l(:
Theorem 9-25
Consider a plant wil

assumed that D(s) an<
column degrees f1.i, i =
P x P arbitrary polynl
and be row reduced.
is a proper rational
D c-l(S)M(s), and Q
posed and has trans
computed from
with (SciR 1(s) < ()ci O(s)
(9-148)
R
Proof
The column degrees O

and M(s) of row deg
D c-l(S)L(s) and D c- I(S
Next we. show th
(9-152). We use (9-13:
D(s)DI(S)D
Figure 9-31
An input-output feedback system.
Since D 11 + D(s)and D
is proper.
515
ds the required com
Using (;(s) = N(s)D-I(S), Co(s) = D,:-I(S)L(s) and C1(S)= Dc-I(S)M(s), G(s) can
be written as
'ith the one developed

t coprime and 'P(s) is
Ldily found for G(s) =
:ve N(s)Dj 1(s), where
,ame coluron-degree
required in Corol1ary
sign a state estimator
roots of det Dc(s), to
of the state feedback
Note that the eigen
n froro (7-86) and wil\
G(s) = N(s)[Dc(s)Q(s)D(s) + L(s)O(s) + M(s)N(s)] -11\(s)
(9-149 )
D(s) ~ Dc(s)Q(s)D(s) +L(s)D(s) +M(s)N(s)
(9-150)
lsators. In the state

l1-q (Theorem 7-13).
lpensators is p(v -1).
"Vability index of any
l/q. lfp=q=1,then
lo the degree of com
) can be greater than,
-variable and transfer
to the single-variable
)roach appears to be
: in the state-variable
m Figure 9-30 will be

s a polynoroial matrix
( froro r to y in Figure
Ir we define
E(s) ~ D(s)- Oc(s)Q(s)D(s)= L(s)O(s)+ M(s)N(s)
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 )
is a proper rationa" matrix, there exist proper compensators Dc-l(s)L(s),

Dc-I(S)M(s), and Q-I(S) so that the feedback system in Figure 9-31 is well
posed and has transfer matrix N(s)D 1(s)Dc(s). Furthermore, Q(s) can be
computed from
D (s) = Q(s)D(s) + R1(s)
with c)ciR(s) < (\D(s) =
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
R 2 (s)D(s)+ R1(s)= L(s)D(s)+ M(s)N(s)
(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
LINEAR TIME-INVARIANT COMPOSlTE SYSTEMS
proper rational mat]

problem will be studi
From (9-153) and (9-154), we have

Of(S) = Oc(s)Q(s)O(s)+ R 2 (s)0(s)+ RI(s)
(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)
loop system shown

proper rational matr
this subsection, we d
feedback configurati
N(S)O-l(S) and ifT(s)
transfer matrix is G a
trary assignment of e
is square and nonsinf
T(
[1 + Oc- 1(s)L(s)Q - I(S) + Oc- l(s)M(s)N(s)O- 1(s)Q - 1(s)] - 1

(9-159)
= Q(s)O(s)[ Oc(s)Q(.I')O(s) + L(s) 0(.1') + M(s)N(s)] - 10c(s)
is proper (Theorem 3-6). The substitution of (9-150) into (9-159) yields

p-I(S)= Q(s)O(s)O l(s)Dc(s)
(9-160)
pes) = O; I(S)Of(s)D - 1(s)Q - 1(s)
(9-161 )
which implies
Go(s)
where
Using (9-153) and (9-154), we have

pes) = 0c- 1(s)[OAs)Q(s)O(s) + R2(S)0(S) + R 1(s)] O - 1(s)Q -I(S)
= 1+ D c- 1(s)R 2 (s)Q -'1 (s) + Oc- 1(s)R 1(sjO - 1(s)Q - I(S)
(9-162)
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)
where di(s) are Hun

T(s) proper, then the
discussed may find se
LetG(s)= N(s)O
Then the open-Ioop ti
and D (s) and Nfes)

is to be solved from a
hence the unity feedb
the open-loop system
9-31, the overall tran
assignable. A compa
bility of implementinl:
and O f(S) are meto .
introdce a differcnt
lmplementation l. e
Figure 9-31. Its oyera
Gfes) = G(s)Q-I(s)[1
19
This probIem was first 1

Reference SSO.
(9-163)
where T(s) ~ 0(.1')0l(S)Oc(.S) is, as proved in (9-156), a proper rational matrix.

Itturns out that every system expressible as G(s)T(s), where T(s) is an arbitrary
Figure9-32
An open
517
proper rational matrix, can be implemented as shown in Figure 9-31. This

problem wil1 be studied in the next section.
(9-157 )
:.1')
)R 1(s)D- 1 (s) (9-158)

~l(s)D-l(S)
(s),
is strictly
D;- 1(.1')1 is proper
)olynomial part of
1 is proper, it follows
atrix
Implementations of open-Ioop compensators. 19 Consider theopen

loop system shown in Figure 9-32, where the plant is denoted by the q x p
proper rational matrix (;(s) and T(s) is an arbitrary proper compensator. In
this subsection, we discuss the implementations of this open loop system by
feedback configurations. Befare proceeding, we mention that if (;(s) =
N(s)D - 1 (s) and if Tis) is ch~sen as D(s)D 1(s) and is proper, then the open-loop
transfer matrix is Go(s) = G(s)T(s) = N(s)Di I(S). This is the problem of arbi
trary assignment of denominator matrix discussed in Corollary 9-24. If C(s)
is square and nonsingular and if we choose T(s) as
T(s) = (; -1(S) diag {dI (s), d 2(s), ... ,dp(s)}
'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)
Dc- 1 (s)and Q-l(s)are
lllowing Theorem 3-4,

ck system.
Q.E.D.
eck the properness of
lton of the inverse of
te
(9-164)
where d(s) are Hurwitz polynomials of smallest possible degrees to make

T(s) proper, then the resulting system is decoupled. Thus the problem to be
discussed may find several applications in the design of multivariable systems.
LetC(s)= N(s)D-l(s) be a coprime fracton with D(s) column reduced.
Then the open-loop transfer matrix is
(9-165 )
(;o(s) = (;(s)T(s) = N(s)D- l (s)T(s) ~ N(s)Dj l(s)N (s)
(9-166)
where
D- l (s)T(s) = D l(s)N (s)
and D (s) and N (s) are left coprime. If we use the utlity feedback system
shown in Figure 9-16, the overall transfer matrix is N(s)D l(s)Nis) and Nc(s)
is to be solved from a Diophantine equation and cannot be arbitrarily chosen;
hence the unity feedback system in Figure 9-16 cannot be used to implement
the open-loop system. Ir we use the input-output feedback systemin Figure
9-31, the overall transfer matrix is N(s)DI(s)D c (s), where Dc(s) is arbitrarily
assignable. A comparison of this with (9-165) reveals immediately the possi
bility of implementing Co(s) in Figure 9-31 as long as the conditions on O,(s)
and D(s) are met. This will be done in Implementation l. We shall also
introduce a different implementation in Implementation n.
y)H - (s)
ms of R(s) and H(s).

.)Hc(s) is proper if and
r way of checking the
,sm}, a sufficient con
degrees m + J1 and is
it is still possible for
115.
Implementation l. Consider the input-output feedback system shown in

Figure 9-31. Its overall transfer matrix is, from (9-148) and (9-149),
(;(s) = G(s)Q -1(s)[1 + D - l (s)L(s)Q - I(S) + D; 1 (s)M(s)N(s)D -1 (s)Q -1(S)] - 1
c
(9-167)
19
This problem was firsl frmulaled and solved in Reference S218. This presentation fo\lows
Reference SSO.
(9-163)
roper rational matrix.

ere T(s) is ah arbitrary
Figure 9-32
An open-Ioopsystem.
518
G(S) = N(s)[Dc(s)Q(s)D(s) + L(s)D(s) +M(s)N(s)] -1 D/s)
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
Hence G(S) implem

Equations (9-169
T(s) =
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
where Dri denotes the ith row degree.

Step 2. Compute the row index, v, of (;(s). Define
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)
has row degrees all equal to m:;:::: v - 1 and is row reduced.

Step 3. lf
i= 1, 2, ... ,p
with C\Kl(S) < ci(s) =
j1i,
with Dri R 2 (s) < DriN (s) = mi, i = 1, 2, ... ,p.

unique.
Step 4. Solve L(s) and M(s) from
or
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)
These decompositions are
.oc(s)[D fes) - N (s)D(s)] ;= L(s)D(s) + M(s)N(s)

.o/s)[R 2 (s)D(s) + R 1 (s)] = L(~)D(s) + M(s)N(s)
Since Oh and Del,

[Del. + Dcl(s)] - are,
the condition in (9
Theorem 9-25.
with DcR(s):::;;DciD(s
example, from the un
for aH i, and compute
Ql(S)= N(s)Q(s) +R 2 (s)
[D1.+ D(s)]
(9-171 )
set Q(s) = 1 and go to step 4. If not, compute

D (s) = Ql(s)D(s) +Rl(s)
Note that Dc(s)D (

U sing (9-138) and (9
(9-174a)
with D,J{2(S) :::;;DrNl

proper. However, f(
Furthermore, the res'
degree conditions in
degrees at most m stil
(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.
section. In this iinph

. square and may not t
in FigLire 9-31 excep
terminal as shwhin
519
(9-168)
guration may be used

nentation, we require
s the nonsingu\arity of
(s) such that U(s)N (s)
let coprime (Theorem
~(s) is row reduced.
G(s) will implement
Proof
The substitution of (9-170) to (9-174) nto (9-168) yields
G(s) = N(s) [fic(s)D(s)J -lfic(s)N(s) = N(s)Dj l (s)N (s)
(9-175 )
Hence G(s) implements the open-Ioop system.

Equations (9-169) and (9-170) imply
T(s) = D(s)O l(s)N(s) = O(s)[fic(s)D(s)] - 1 Oc(S)
(9-176 )
Note that Dc(s)D(s) in (9-175) corresponds to the D(s) in Theorem 9-25.

Using (9-138) and (9-139) with mi = m, for all i, we can write (9-176) as
t(s) ann N(s) are right
[D h+ D/(s)J - lT(s)[D ch + Dcl(s)J - l
H(s)[fic(s)D(s)J - l Hc(s) (9-177)
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.
)}
We discuss now the stability of the implementation. Clear\y, the stability

depends on T(s). lfT(s) is chosen so that T(s) and G(s)T(s) are BIB stable, then
det D(s) in (9-169) is a Hurwitz po1ynomial, and the feedback implementation
wil1 also be BIB stable. The design involves the cancellation of the roots of
det fi,(s) which are however arbitrarily assignable. lf det Dc(s) is chosen as
Hurwitz, then the feedback implementation is asymptotically stable as well.
The decomposltions in (9-172) and (9-173) can be carried out by using the
algorithms in References S34, S 137 and S236. See also Problem G-15. In the
design, the conditions in (9-172) and (9-173) can be l'elaxed. We may replace
(9-172) by
(9-169)
.mi'
Then the matrix

(9-170)
uced.
(9-178)
... , P
(9-171)
with cRI(s) ::s:;ocD(s). In this case, the decomposition is not unique.

example, from the unique QI(S) and RI(s) in (9-172), we can obtain
RI(s)
(9-172)
101
(9-173)
decompositions are
s)N(s)
(9-174a)
s)N(s)
(9-174b)
(s), Q(s), L(s), and M(s),

n-loop system Go(s) =
any constanl matrix ,
LO
R(s) - BO(s)
Fol'
(9-179)
meeL (9- i 78 j. 3l11lilnj, (9-1 j) can be replaced by

(9-180)
with (j,.Rz(s)::S:;riN(s). In this case, we cannot expect every QI(S) to be

proper. However, for almost all QI(S) so computed, QI(S) is still proper.
Furthermore, the resulting D,(s)[D (s) - N(s)Q(s)D(s)] still meets the column
degree conditions in Theorem 9-17; hence solutions L(s) and M(s) of row
degrees at most m still exist in (9-174).
lmplementatiori jI. We introduce a different implementation in this sub
section. In this implemen.tation, the op~n-IoopcompensatorT(s) may be non
squre and inay n6t have full rank. The configuration is identcal to the one
in Figure 9-31except that acorripensator Ole. l(s)K(s) is placedat the input
terminal as shown in Figure 9-33. The transfer matrix of the system is clearly
520
DE
Step 4. Define
Ir J: - f.J.. - In 2. O, ~
Hurwitz polynorr
b,.W(s) ~m.
Step 5. If
set Q(s) = 1 and g.

Figure 9-33
A plant input-output feedback system.
equal to, fol\owing (9-167) and (9-168),

GAs) = G(s)Q -1(s)[1 + D e- 1(s)L(s)Q- I(S) + D e- l(s)M(s)G(S)Q-1(S)] - 1D e- 1(s)K(s)
or
G(s)= N(s)[De(s)Q(s)D(s) +L(s)D(s) +M(s)N(s)]-IK(s)
(9-181 )
(9-182)
with 8,.I~2(S) < 6,.

Step 6. Let
The implementation eonsists of the fol\owing steps:

Step 1. Compute a fraeton (;(s) = N(s)D- 1(s), where D(s) and N(s) are right
eoprime and D(s) is eolumn redueed. Compute
and salve L(s) and
(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) =
D;: 1 (s)V - l(s)V(s)N J(s) = (V(s)Dr(S))-l(V(s)N As))
~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.
This shows that the fe

loop eompensator T(s
From (9-188) and
W(s)D J
Beeause b,. R 2(s) < /11, (~
Henee under the assu

reduced anci m?:. v - 1
(9-191). Hence the eo
521
Step 4. Define
W(s) = diag {Wl (s), W2(S),
>
>
Wp(s)}
(9-186 )
If J: - .t - m:?: O, set w(s) = 1. IfJ: - .t - m < O, set w(s) as a monic arbitrary

Hurwitz polynomial of degree (.t +m - J:). Because of ; :?:.t, we have
rW(s)':::::m.
Step 5. If
i = 1,2, .. >, P
(9-187 )
set Q(s) = 1 and go to step 6. If not, compute

D(s) = Ql(s)D(s) +R(s)
(9-188)
with cR(s)<cD(s)=.t, and compute

W(s)Q(s)
)Q - (s)J - D; (s)K(s)
(9-181 )
K(s)
Dc(s)Q(s)
+ R 2 (s)
(9-189)
with ,R 2 (s) < ,>;Dc(s) = m.

Step 6. Let
(9-182)
K(s) = W(s) J(s) = W(s)V(s)N J(s)

s) and N(s) are right
(9-190)
and solve L(s) and M(s) fram

W(s)D(s) - Dc(s)Q(s)D(s) = L(s)D(s) + M(s)N(s)
(9-183)
)re, we require Df(s)

mn-degree-coefficient
(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
GJ(s) = N(s)[W(s)D(s)J - 1W(s) (s) = N(s)Df l(S) (s)
))-l(V(s)N(s))
;-_:: l\I(s)[j- J(,))T(:;} ~-;~ :G(:;')1'(:;:

(9-185)
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)
C[W(s)D I(s) - Dc(s)Q(s)D(sJ] < m + .ti

: prpperness of D(s)Dj 1(s)
wr:, must introduc~ v(s) \0
>
1 fun
rnk, we ha_e v,(s)
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
The polynomial equation in (9-192) can be written as

D; l(s)W(s)Df(s)D -1(S) = Q(s)
+ D; l(s)R 2(s) +D; l(s)W(s)R 1(s)D- I(S)

(9-194 )
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)
which approaches a zero or nonzero finite constant matrix. Hence

D(s)D 1(.1')'" -1(s)De(s) is proper, and consequently, Q - 1(.1') is proper.
Now we show that the compensator D; 1(s)K(s) = D e- 1 (s)W(s)N f(S) is
proper. Consider
T(s) = D(s)D j \s)N f(S)
= [D + D(s)]H(s)H[;/(s)[1 + D fl(S)] -1 H,v is)[Nfh
'I
Applications. 21
+ N ft(s)]
As .1'->00, we have T(s)-> D hH(s)H b1 (s)H(s)N fh . Because T(s) is proper, we

conclude
(\ 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)
is proper. The substitution of (9-191) into (9-197) yields

P-l(S) = Q(s)D(s)[W(s)D (s)] -ID e (s)
which implies,by using (9-192),
P(s) = De""' l(s)[D e(s)Q(s)D(s) + R 2(s)D(s) +W(s)R 1 (s)]D - I(S)Q - I(S) .
= I + D; l(s)R 2 (s)Q - I(S) + D e- 1 (s)W(s)R 1 (s)D - I(S)Q - I(S)
Because D e- 1 (s)R 2 (s), R 1(s)D-l(s) are strictlyproperand because Q-'I(S) and
D e- 1 (S)W(s) are proper, the polynomial part of P(s) is 1, which is nonsingular.
Hence P-l(S) is proper.
.
Q.E.D.
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
Follows Reference S54.
;)W{s)R 1{s)D- {s)

(9-194 )
1
er and 0c- {s)W{s) is

;) is Q{s). To show
~ Theorem 3-4, that
-134), (9-138), (9-139),
f> fl{S)] - 1
t-
,+ Dcl(s)]
(9-195)
ler, we have, as s -> 00,

(9-196 )
tant matrix. Hence

(s) is proper.
= D; 1{s)W{s)N f{s) is
)[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)
)]D - 1{s)Q - I{S)
I{S)Q-l{S)
A remark is in order regarding the degrees of compensators in this designo

In general, the degree of det Q{s) in this design is smaller than the one in the
previous subsection. Because D c- l{s)K{s), D c- 1{s)L{s) and D c- 1(s)M(s) all have
the same denominator matrix, they can be implemented with a degree equal
to deg det Dc{s). Hence the total degree of compensators in this design is
deg detQ(s) + deg det Dc{s), which is generally smaller than the ones in the
previous subsection. If det D(s), det V{s), det Dc{s) and det W(s) are Hurwitz,
the implementation is, similar to Implementation 1, asymptotically stable.
The remarks in (9- 17S) and (9-1 SO) regarding (9-172) and (9- 173) in the
previous subsection are equally applicable to (9-ISS) and (9-1S9) in this sub
section and will not be repeated.
In addition to the two implementations, it is possible to obtain other feed
back implementations of the open-Ioop compensator T{s). For example, we
may use the single-loop feedback system in Figure P9-32 or the input-output
feedback system in Figure 9-30 [without using Q -l{S)] to implement T{s).
The basic idea and design procedures will be similar to the ones in Imple
mentation 1; however, they may require less computation because the division
steps in (9-172) and (9-173) are not needed. For yet another implementation,
see Reference S21S.
Applications. ZI In this subsection, the results of the previous subsection
will be employed to design feedback systems. We consider first the decoupling
problem and then decoupling together with asymptotic tracking and distur
bance rejection. Finally, we study themodel matching problem.
Decoupling. Consider a p x p nonsingular proper rational matrix G{s) =

N{s)D- 1 {s). If an open-Ioop compensator T{s) is chosen as D{s)N- 1 {s)i>jl{S),
where O(s) = diag {dI (s), dz{s), . .. , dp{s)} and d{s) are H urwitz polyr;.omials of
minimum degrees to make T{s) proper, then the implementation of G{s)T{s) in
the input-output feedback system in Figure 9-31 will yield a decoupled system
Oj 1(s). We note that in this design, the roots of det D{s) are shi}ted to the roots
ofdet (O(s)N(s)); hence there are no cancellations involving the roots of det O(s).
Since the input-output feedback has no effect on the numerator matrix N(s),
the disappearance of N(s) from the decoupied system is accompished by exact
cancellations. Thus if det N{s) is not a Hurwitz polynomial, the design will
involve unstable pole-zero cancellations. Thus the choice of T(s)=
D{s)N - I(S)O jI (s) cannot always be employed to design decoupled systems.
We discuss in the following a method of designing decoupled systems
without involving any unstable pole-zero cancellations. Consider a p x p
nonsingular proper rational matrixG{s) = N{s)D-1{s), where D{s) and N{s) are
right coprime and D(s) is column reduced. We factor N(s) as
witb
N{s) == N{s)NzCs)
N I{S) = diag {/311 (5), /31 is), ... , /31 p{S)}
d because Q- 1 (s) and

which is nonsingular.
Q.E.D.
523
21
Follows Rererence S54.
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,
Hence we may choos
It is a rational matrix with only stable poles, that is, poles with negative real
parts. Then we have
The implementation
matrix
N 2(s)N 2(s) = N 2d (S) = diag {f321(S), f3ds), ... , f32P(S)}
Now we choose an open-loop compensator T(s) as

T(s) = D(s)N 2(s)D,-1(S)
D,(s) =diag {1X 1(S), 1X 2(S), ... , IXp(S)}
with
(9-198)
where IX(S) are Hurwitz polynomials of minimum degrees to make T(s) proper.
Then we have
G(s)T(s) = N [(s)N 2(s)D -1(s)D(s)N 2(s)D,- l(S) = N [(s)N 2d(S)D t- 1(s)

-d'
lag
{f31(S) f32(S)
f3 p(S)}
( )' 1X ()
, ... , IXpS
()
S
1X 1 S
2
Implementation I.
V<
(9-199)
S2 +2.s
4
where f3(s) = f3(S)f32i(S). Hence the implementation of G(s)T(s) will yield a

decoupled system. We note that this design involves only the cancellations of
the stable roots of det N 2(s), and the resulting feedback system will be asym
ptotically stable.
Example 2
_l s
4
where D res) and N r(~

ced ure discussed in A
v = 2, Clearly we ha'
choosing Dc(s) arbitrE
Consider the plant

G(S)=N(S)D- 1(S)=L s:l
~lJ[s2;1
~T1
which has row degree:
We compute f311(S) = 1, and f312(S) =s +1. Hence we may factor N(s) as
N(s)=Nl(S)N2(S)=I~
. . .J
Wecompute
_[
0,_,1ls,2
6d (D c (s)D r (s)),:::; Ji + 17
11
...J
Thus, the L(s) and M(
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
of N(s). Let ~Zi(S)

of the ith column of
,nd define
~J[ -~
)W
Hence we may choose

D,(s) = [s 0+
,les with negative real
(s
+ 1)2
(9-201 )
The implementation of G(s)T(s) will yield a decoupled system with transfer

matrix
, f3 zP(s)}
GIS)TIs) = dia {s - 1 (s + 1)(s

\ \
g S + 2' (s + 1)2
(9-198)
. {s - s- 1}
=dlag - -1. -
s+2's+1
to make T(s) proper.

Implementation I.
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)
We implement G(s)T(s) in Figure 9-31. We compute
-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
Thus, the L(s) and M(s) in (9-174) can be solved from

1
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 .
iake T(s) proper. We
1 1 O O
1 1 O O 1 O
000100
---------------------
=[-3//4 -27/4: 13/2

54
13/4 : - 1/4
-9/4: 9/4 O: O 00J

3 :1
O: O
+';' 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
This completes the first implementation.

Implementation JI.
We implement G(s)T(s) in Figure 9-33. We have
D/e =1 and J ?:.p; hence V(s) =1, D(s) =D(s), and j(s) = N(s). We choose
m
= v -1 = 1and
Because JI - PI - m = 2 - 2 - 1 = -1, we choose WI(S) arbitrarily as s +2, of

degree 1. Because.f2 - P2 - m = 2 - 1 - 1 = O, we set w 2 (s) = 1. Hence we have
S
W(s)= [
~J
+2
Clearly, we have 6ci (W(s)O (s)):::; Il +m, i
1, 2; hence we set Q(s) = 1 and
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--,
The L(s) and M(s) in (9-191) can be solved as

L(s) = [llS + 21/4 (-3/4)S+3J
(- 5/2)s - 2
-5/2
This completes the second implemental ion.
where #(s) are uniql

to be chosen to make
to be decoupled onl)'
designed, in additior
rejection, iX(S) cannot
the additional feedba
bance rejection.
In order to achie'v
be robust with respe
discussed in Section 9
Ifwe introduce a diagl
with h(s)/cjJ(s) proper
reduces to p number e
transfer function of th
-1IS-4
M(s) = [ (5/2)s + 2
and disturbance reject

or, beca use N(s) in G
#(s) is a factor of det r
-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
DESIGN Of COMPENSATORS: INPUT-OUTPUT FEEDBACK SYSTEMS
527
unstable pole-zero cancelIations, as
-9J
d.
2)s- 6
1-2
13/4
{~l(S) ~2(S)
ure 9-33. We have

N f(s). \Ve choose
)=
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
set Q(s) = 1 and
h;(s)~(s)
6. h(s)~(s)
<f>(S)iX(S) +h(s)/J(s)
-4
+-2
~p(S)}
lag CY.(s)' iX2(S)'' O:p(s)
[(s)
(9-203)
Its denominator is a polynomial of degree deg <f>(s) +deg iX(S). Ir we require

h(s)/<!>(s) to be strictly proper, there are deg <f>(s) number of free parameters in
hi(s); there are deg iX(S) + 1 number or free parameters in o:(s). Hence ir (s)
and /J(s) are coprime, the roots of f(s) in (9-203) can be arbitrarily assigned by
proper choices of h(s) and o:;(s). The condition for asymptotic stability, tracking
and disturbance rejection is that no root of <f>(s) is a transmission zero of (;(s)
or, because N(s) in (;(s) = N(s)O -1(S) is square, a root of det N(s). Because
~(s) is a factor of det N(s), we conclude that if no root of <1>(8) is a transmission
/4J
-27
4
91
I
of these two imple
compensators in the
the first implementa

can be modified in
is very close to the
19 canceled. Instead
, we may decouple it
se, the plant is to be
cations ate straight
pfillg. In this su b
:oupling, asymptotic
l(S) be a p xp non
ithout ivolving any
(a)
(b)
Figure 9-34
Design of robus! sys!em.
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)
+ f31 (s)h l(S) = 11 (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
are h 2(S)=1;2 S -16 and 1X2(S)=S2_1;2S_1;4. Next we replace (9-201) and

(9-202) by
O,(s) = [Cf.l(S)
O
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-
Once G(srC(s) is implemented as an input-output feedback system inside the

box in Figure 9-34(a), the design is comp1eted. This part of design is similar
to the one in Example 2 and will not be repeated.
The design problcms discussed so far concern only with the

assignment of poles and denominator matrix; the numerator matrix is left un
specified. In this subsection, we discuss the assignment of denominator matrix
as well as the numerator matrix 01', equivaJently, rhe entire overall transfer
'matrix. This problem i~ often referred to as the exact mdel matching problem.
, Consider a plant with q x:p proper rational matrix (;(s). The desired mQdel
is assumed to have the q x r pr.oper rational matrix(;m(.~)' The problem is to
find a'configuration and co'mperrsators lor the plailt so that the resulting overall
systcm has (;m(S) as its transfermatrix. We study theproblem by usingan openModel matchillg.
loop compensator.
Now ir a solution T(s

design can be accoml
configuration shown
on the solution of (9
Let T(s) and Cm;1
be written, following
is a linear algebraic e
real rational functior
necessary and sufficie
over lR(s). The soluti

In the model matchir
proper rational matri
sma11 as possible. Tb
to meet (9-204) is call
SI72 and S210.
Before proceedin
A(s) and B(s) are left
is unimod ular, that is,
is, unimodular. Folle
matrix M(s) tb be roH'
if the R(s) in any fa
Theorem G-8', M(s) i
in iC. We may also ex
A square polynomial
sum of a11 row degree:
if and only if its row-,
we define a q x n poi)
coefficient matrix is e
ii' iVK(s) is 01' "uii row (
U(s) such that U(s)1V
irreducibility and co
With this preliminary
We discuss first th
Ir (;(s) is square and n
only if (; -1(S)(;",(s)
~ore complicated. . L
G(s) and Gm(s). The I
n If Gm(s)= 1, then the solut

reduces also to the minim
ts of f(s) can be arbi

)mpute h(5) and X(5).
ecouple the plant as
ng system is asymp
,totically the reference
~n, if there are param
tcluding the internal
;)wever, if the overall
)f asymptotic tracking
loop compensator. Let T(s) be a p x r open-Ioop compensator so that

G(s)T(s) = G",(s)
found to stabilize the

11(5) =(s +2)3. Then
he solutions of
'e replace (9-201) and
lck system inside the

rt of design is similar
I
concern only with the

ator matrix is left un
f denominator matrix
ntire overal1 transfer
del matching problem.
l The desired model
l The pro blem is to
.t theresulting overall
lem by usirig an open-
(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
rank G(s) = rank [G(s)
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
Ir Gm(s) ~ I~ then the solution '-i(s) in (9-204; is a rig~t

reduces also to the minimum design problem.
(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))
does have a full ce

algorithm.
Theorem 9 -28
IrY(s) in (9-209) is a rr
G-14', then NT(s)Di
(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)
B(s)] = p + r - rank [G(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)
Ir Y''e has rank r, ther

hence N T(s)Di 1 (s) is
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
Yhe is not computed 1

sorne j. However, fral
elernentary colurnn ope
23"if
1d q x r polynomial
rank, there exists a
w reduced. Consider
(9-206)
; proper if and only if
531
does have a ful1 column rank as a consequence of the coiumn searching

algorithm.
Theorem 9-28
UY(s) in (9-209) is a minimal polynomial basis of (9-208) obtained as in Theorem
G-l4', then NT(s)D les) is proper if and only if Y'e has rank r.
(9-207)
Proof
s assertion is essenti
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
); hence al1 discussion

e the right null space
al to
form the generalized

as in Theorem G-14'.
from left to right by
tly r primary depend
Ves) be the solutions
Then Ves) is, as in
d, and is a minimal
s).
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
Find a 3 x 2 minimal proper solution of

(9-209)
"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
LINEAR TIME"INVARIANT COMPOSITE SYSTEMS
The multiplication of S(S + 1)(S + 3) and the substitution of T(s) = NT(s)Di 1 (S)
yields
to the three primar)

to compute
Clearly we have r = p + r - rank [G(s) Gm(s)] = 3 + 2 - 2 = 3. We form TI as

in Theorem G-14' and then apply the column searching algorithm to compute
O
O
:3 - - - 0-
O: O
O:
o
O: O
O:
-- -0- -: - 6---- -0- -:- - --- -0- - - - - : - -0- - - - -
which yields immedi;
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'
Y(s) has rank 3 for ev

reduced, for Y,,~ is 01
basis of the right nul
rank 2; hence we ha"
** 0- - -0- ---- - - - -- - - - -6--, --1- ----6---0- ---6-----6
'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]
"Wherexdenotes thenonzero col~mn, and"O thezercolumn. Thiscomputation

is very easy to carry out by hand. The notation : is explained in Appendix A.
There are four linearly depe~dent cblumns in TI"; three of them are primary
dependent columns as indicated by thearrows. Note that in Theorein G-14',
"linearly dependent columns will appear only in B-rows (see dual of Theorem
G-13). In the minimal design problem,dependent columns mayhowever
appear in B- as well as A-columns as shown in this example. Corresponding
This is a minimal pro

The numerator m
discussed in Section
the numerator matrix
to achieve pole or den
introduce any undesil
numerator matrices ir
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
which yields immediately
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
e, hence it is column irreducible. It is also column

reduced, for Y/le is of ful1 colu~n rank. Hence Y(;) is a minimal polynomial
basis of the right nul1 space. The first and third columns of Y/le yield a Ylle of
rank 2; hence we have
Y(s) has rank 3 for every s in
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
This is a minimal proper solution.

n. This computation
lned in Appendix A.
of them are primary .
: in Theorem G-l.4', .
See dual of The'orem
lumns may however
lple. Corresponding
The numerator matrix of a plant cannot be affected by state feedback, as

discussed in Section 7-3, nor by output feedback; hence the only W',/ toaffect
the numerator matrix is by direct cancel1ation. In the design of compe.sators .
to achieve pole or denominator matrix assign~e~t, we have beel careful ilot to
introduce any undesirablepole-zero cancel1ation. This is possible becausethe
numerator matrices in these designs are not specified. In the model matching,
534
because the numerator matrix is also specified, undesirable pole-zero cancella

tion may be unavoidable. Therefore in the model matching, the choice of
Gm(s) must be very careful. Otherwise, even if we can find a proper T(s) and
implement it in a well-posed input-output configuration, the design is still
unacceptable.
Then it can be showr

Reference S85, whe
D-l(S)N(s); the metl
rings. The method I
S51,andS218. Wefir:
[Lo
9-8
Concluding Remarks
In this chapter, we studied a number of topics: characterization of composite

systems by their transfer matrices, stability of feedback systems, and the design
of compensators in the unity feedback systems and the plant input-output
feedback systems to achieve arbitrary pole placements, asymptotic tracking,
disturbance rejection, and decoupling. We also studied the implementation of
open-loop compensators by using plant input-output feedback systems.
All the design problems in this chapter hinge essentially on the solution of
the Diophantine equation:
L(s)O(s)+ M(s)N(s)= F(s)
O(s)L(s)+ N(s)M(s)= F(s)
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)
U 12(S)J [O(S)J = [R(S)J

Uds) N(s)
O
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 {,~')
with briFl(S) <briO(s), then we have ITv(F(s = F1(s).

in terms o[ the basis in (6-223) as
We then express Fl(s)
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]
or, equivakntly, El lies in the space spanned by the column of

[B AB . 0'0 An-1B]. The solution of El =[B ABo"
A"-1B]P is the
represent"ation.of M(s) in (9-211 b)with respect to the basis L(s), that is, M(s) =
L(s)P. Next we compute
N(s)M(s) = D(s)Q2(s) + F 2(S)
Mo : L l
Then a solution exisl

solution of the equati<
ical comparisons of '
In addition to Eq\
also arises in the desig

from the Diophantil1E
See References S88, S
asymptotic tracking a
the internal model is c
bypasses the equation
this texto
In this chapter we
plant input-output fe
compensator is not ~
N(s)O j 1(s)Nc(S). In t
assignable, and the n:
Dc(s) in N(s)Oj l(s)D c
completely concel1ed :
Dc(s) in the plant inI
feedback configuratio
L(s), M(s), and Oc(s) il
Dc(s) in the latter, and
In the design, the
fracUonal formo Tw
ll1e lrst method, we
sequence of polynomi
one as in (G-33). The
In this case, we must '
reduced one. In the
A -l(s)B(s). We fom
ofA(s) and B(s), and n
left to right. From the
tainaright-coprime fr
reducd; in fact, it is a
Once a right-cop
, step'is to co"mpute th<
necessary because we
search linearly indepe
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)
irst, we may apply a

[D'(s) N'(s)]' into a
~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
then express F 1(s)
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
Is the feed back systel
8180 stable? asymptotic

Ll(s)Ll 2 (s)det (1 + (;ds)G 2 (.
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
Given the plant 9(05)

degree 1 so that the poles of
proper compensator C(s) c
-1, -2, -3, and -4.
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.
Consider the plant gi

pensator in the llnity feedb
com pensator of dcgrec 1 in
9-9
Consider the eedb<\!

inctlOn"l Let N(s)=N(s)
N(s)N(s)/D(s), wth deg 1
compensator P(S)fDc(s) an!
Dc(s) be Hurwitz? lf Dc(s)
the system be acceptable?
wilI be acceptable? See Prc
!l10
Are their paralIeI and tandem [(;(s) folIowed by (;2(05)] connections controllable? obser
vable? completely characterized by their transfer matrices?
9-4
Are the feedback sys
9-7 Given the plant 9(05)

system so that the poles of
Consider
and .
9-3
9-5
ls the paralIel connection of
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
ldent, the m is the row

; in Sv should be equal
,s of the computation.
,pace of Sv _ 1, and the
ensators.
fa matrix is basically
lade to be nonzero by
matter what method,
ltions with pivoting, is
",s cannot be changed.
: transformations with
lolder transformations
thods are numerically
the nurnerical stability
,blem are two distinct
-conditioned problem,
;, difficulties may still
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
Is the feedback system shown in Figure 9-11 with
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
9-8 Given the plant g(s)=(s2-1)/(s2'---3s+I). Find a proper compensator C(s)of

degree 1 so that the poles of the unity feedback system are - 1, - 2, and - 3. Find a strictly
proper compensator e(s) of degree 2 so that the poles of the unity feedback system are
-1, -2, -3, and -4.
in the forward path and

1 transfer fLlnction'J
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~
1e same set of poles.
Figure P9-10
538
9-11
Consider
9-18
Consider the sys!l
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
Repeat Problem 9-11 for
~
[ 5 +1
G(5)= -
52~J
1)
5(5 -
9-13
Find a compensator in 11
Is the resulting Gf(5) cycL

Consider (;(5) = N(.
Let N(5)O-1(5)=O-l(S)
M(5) are left coprime, so
9-19
Which of the following matrices are cyclic?
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
Consider lhe feedba

h'lck syslem is asymptol
where aH poles of r(s) He
coprime. Show thal
9-21
9-14
9-15
e(s)=(I
has no cIosed righl-half-pla
Find the set of 2 x 1 vectors t such that 6;(5 of6;(5)t), where

5
~
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
Consider the system in Example 3 on page 484; that is,

2+ 1
~
G(5) = .
-1
.[ O
52
-1
5
~onsider lhe unity fe

O-l(s)N(s) be coprime frac
nomial matrices such that
H(s) with aH poles inside lh
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.
stabilizes lhe Unity feedback

Try G(s) =(5 + 0/5(5 +2) an,
9-23
Giveng(5)=(5-1)j5(.~
feedback systein lo have gf(S
PROBLEMS
9-18
Consider the system in Example 2 on page 476, that is,
G(s)
A
l~ ~
l [52 0J-I[1 5 0J
=
lOs
s
:k system are -1, -2 j,
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?
Compare your result with the one in the text.
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
Let O-I(s)N(s) be a left coprime fraction. Show that M(s)D-l(S)N(s) is a poly

nomial matrix if and only if O(s) is a right divisor of M(s), that is; there exists a polynomial
matrix M(s) such that M(s) = M(s)O(s).
9-20
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,
e(s) =(1 +G(sW lr(s) = (O(s) + N(s)t ID(s)O,:-l(s)N..(s)ro
then G(s) is cyclic. Note
has no closed right-half-plane poles if and only if O,.(s) is a right divisor of O(s).
), where
~(S)
~(s)
+
figure PS-2'
:(s) + K for any constant K.
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
C(s)= [X(s)- H(s){sW1[Y(s)+ H(s) 0(5)1

stabilizes the unity feedback system. lfH(s) is proper, will Cls) be proper'? [An5wer: No.
TryG(s) ==(s +1)/s(s +2) and H(s)=O.J
lles of the resulting system
re your result with the one
Given 9(S) = (s -1)/s(s - 2) and 9j(S) = (s - 1)/(2s 2 +45

feedback system to have g(5) as its overall transfer function.
9-23
3), design an input-output
540
9-24
9-32
Consider the plant transfer matrix

A
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
is nonsingular and can t

ditions, will there exist :
design?
~ ILS;
11- ~
~
as its denominator matrix. Choose De(s) as diag {s +2, s + 2}.

9-25
Repeat Problem 9-24 to have
_ [s
DJ(s) =
Figure P9-32
+1
9-33 Consider G(s) = N

proper if and only ir D(s) i~
as its denominator matrix.

9-26
Show that the tra
N(s)[De(s)D(s) + Ne(s)N(s
Repeat Problem 9-24 to have

S(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
9-35 Let Y(s) be a n x J

= 1,. 2, ... , p, and let )
Consider
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:
'0'; Dc", Nc",]+[k 1 k 2
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
Consider G(s) = N(s)D - I(S), T(s) = D r- 1 Nr\s), and

Go(s) = G(s)T(s) = N(s)[O,(s)D(sW 1 N,(s)
Show that if Dr(s)D(s) and N,(s) are left coprime, any feedback implementation of Go(s)
reduces es'sentially to an open-loop system.
9-31
Consider
[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
.
.
if and only ir V(s) is colu

Rererence S95 and can be
ininin12..l. 1fin" t~t ,~ =
show the equaiity, we she
relation Yd = Y/KCX, where:
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
'eedback system to have
design?
Figure P9-32
Consider G(s) = N(s)D-l(S) with ciN(s)'::;ciD(S), for aH i. Is it true that G(s) is

proper ir and only ir D(s) is column reduced? Hnot, find a counterexample.
9-33
9-34
Find a minimal proper solution of
compensators proper? Is
ation?
well-posed input-output
pole-zero cancellations.
)w thal ir L\J> J(i) = bciD(s)
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
Jefficient matrices of Ois)

.(5) is strictly proper. The
y(s) = Y(s)x(s)
Show that
2': v - 1, rOl' any real k, i =
degy(s)= max {L+degx(sj}

i:x,(s)
A o : .. ' : - B",
A",J
B", A",J is a basis of tbe

This is a paramelerizalion
,(s)
:k implementation of Go(s)
reject plant entered distur-
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:
First we assume a 11 =/=

the ith row, for i =o 2, ~
Elementary Transformations
Consider a matrix A with e1ements in IR or iC. The following operations on A

are called the elementary operations or transformations: (1) multiplying a row or
a column by a nonzero real or complex number, (2) interchanging two rows or
two columns, and (3) adding the product of one row or column and a number to
another row or column. These transformations can be achieved by using the
elementary matrices, for n = 5, of the forms
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)
They are again elementary matrices. The premultiplicationof E on A operates

on the rows of A, whereas the postmultiplication of E on A operates on th
columns of A. For example, EzA interchanges the second and flfth rows;
EJA adds the product of the second row and d to the fouh row of A; whereas
AE J adds the product of thefourth column and d to the second column ofA.
542
where X denotes possib

called the gaussian elim
matrices. 1
1
AII diagonal elemenls in (A

by llsillg additional column
543
GAUSSIAN ELIMINATION
A-1
Gaussian Elimination
Consider the n x m matrix

a lz
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
lementary matrices are
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)
ion ofEjon Aoperates

on A operates on th<;
econd and fifth rows; .
Jrth row of A; whereas
e second column of A.
2m
where ei 2 = -aIz/aiz and at = a +ezaL. Proceeding in this manner, matrix A

can be transformed into a matrix whose elements below the diagonal are zero as
(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
~1J A=0O j ~1z a13
O O
Dwing operations on A
1)multiplyinga rowor
rchanging two rows or
)Iumn and a number to
. achieved by using the
:-~?- -aa- F --. ~ ~ -_aa-:
where x denotes possible nonzero elements. This process of transformations is

cal1ed thegaussian eliminaron. The matrices in (A-6) are upper right triangular
matrices. 1
1
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
with x'x = 1. where x

(1 - 2xx')' = 1 - 2xx' =
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
Since a'a = b'b by aSSl

is invariant under trar
This establishes. the a

K and K 2 in (AA) ane
Consider the n x m
of the first colllmn al
there exists a Househ
of the right-hand-side I
row of H A and repea
matrix. Proceeding i:
sequence of Househol(
The numerical sta!:;
iiUUilion or columns.
we permute the columi
apply the Householde
process, the diagonal e
order of magnitude.
pivoling.
We note that Hom
K and K 2 in (A-4)anc
after the application of
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.
4Thesign rnay bechosen to b,

tion (see Problern A-S), we I
HOUSEHOLDER
agonal element is zero.

encounter sorne com
LlI in magnitude, then
: errors in representing
rations. Furthermore,
stages. Consequently,
,fthis phenomenon, the
e elimination, we first
the first colurnn of A,
a matrix of the form of
umn. Before carrying
the largest element in
rst element, and bring
:d the gaussian elimina
:;; 1, and the errors will
numerically than the
:\ude that every matrix
;equence of elementary
her improved by using
row operations as well
~gest element in magni
,t to the first column by .
, by an e\ementary row
st column of the trans
:ess for the remaining
with complete pivoting.
1 parta\ pivoting; how
:ment in magnitude is a
t used as orten as partial
of view of overall per
generality. and ease of
. satisfactory for most
improved by using the

A Householder trans-
Ide.
complex conjugale lranspose
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,
H'H = (1 - 2xx')(I - 2xx') = 1 - 4x:x' + 4xx'xx' = 1

we have H-' =H'; hence H is orthogonal (see Appendix E). Because all the
singular values of H are 1 (Problem E-9), its condition number, defined as the
ratio of the largest singular value and the smallest singular value, is 1. Hence
the use of Householder transformations will not impair the numerical property
of a problem (see Problem E-lO).
An important property of Householder transformations is that given any
two vectors a and b of equal euclidean norms, that is, Ila\l~ =a'a = b'b, there
exists a Householder transformation H such that Ha = b. 1ndeed, if we choose
x =(a - b)flla - bll z, then we have
Ha =
[1 _
2(a - b)(a - b)'] a =a _ 2(a - b)[(a - b)'a]

Ila - bll~
a'a -a'b - b'a +b'b
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
In our application, it is often required to search the linearly independent rows of

matrix A in order from top to bottom. By this, we mean that we first check
whether or not the first row of A is nonzero, or equivaIently, linearly independent
by itseIf. We then check whether or not the second row is linearly independent
of the [rst row. In the kth step, we check whether or not the kth row of A is
Iinearly independent of its previous k - 1 rows. If a row is Iinearly dependent
onits previous rows, the row is to be eliminated from subsequent consideration.
Furthermore, we may want to find the coefficients of the linear combination of
the dependent row.
Let the (i, j)th element of the n x m matrix A be denoted by aij' Let a k
be any nonzero eIement in the first row af A. This element will be called the
pivot element ar simply the pivoto Let K be of the form shown in (A-4) with
ei = - aik/alk, i = 2, 3, ... , n. Then the kth coIumn, except the first element, of
KA = (a)) IS a zero column, where
=aij +eiaj' Let aL be any nonzero
element in the second row of KA. Let K z be of the form shown in (A-5)
with eiZ = - aMaL. Then the jth column, except the firsl two elements, of
K z K1A = (alj) is a zero column, where alj =ai) + eizaL. In this process, if there
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
with bjj = Lis just th

The matrix K can
11 - 1. This is comp[
computing any row o
of
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
is c1early or rank 5 and nonsingular. However, we have
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
rank of the matrix is

np1e wayof computing
is no nonzero element in a row, we assign K as a unt matrix and then proceed

7
to the next row. If we carry this process to the last row of A, then we have
Kn-K n- z -" KzKIA~KA=A
'ly independent rows of

~an that we fIrst check
,y, linearly independent
is linearly independent
IOt the kth row of A is
'" is linearIy dependent
Iseq uent consideration.
: linear combination of
:noted by aij' Let alk
ment will be caBed the
m shown in (A-4) with
;:pt the first element, of
~et
be any nonzero
~ form shown in (A-S)
first two e1ements, of
In this process, if there
(A-S)
with b jj = L is just thejth row of K.

The matrix K can be computed by direct multiplications of K, i = L 2..
f1 - 1.
This is complicated. We introduce in the following a recurs.ive way of
computing any row of K. First we store the ith column of K in the ith column
of
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)
where K ~ KII _ K II - z '" K 1 Since every K is a lower triangular matrix with

Ion its diagonal. so is K. The number of nonzero rows in A gives the rank of A.
Ifthejth row of is a zerO row vector, then thejth row of A is linearly dependent
on its previous rows. F urthermore, the coefficients of combination
aL
~xample,
547
e 31
ej l
: e32 :
'
ejZ :
O
O
O
First.i rows of F
ej3
l'j(j-l)
'- bj -z-:- b;;------b;j-_-I~ --b~;:

I
~ith
(A-10)
row of K
Then it is straightforward lo verify tha t
b jj = 1.
e1k+ I)kj
bjj ] ~(k+Z)k
'~e
(A-11 )
ejk
=2' -ne ..... O and A" is nearly

nI of A. (which is equal lO I
em 2-23)], fiar from he form
guIar value decomposilion in
have as 2-', for large n (see
ted. Thus he singular value
: he rank of a nalrix:
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.
The last row of

to write
1
F~A=
Example 1
Find the linearly independent rows of
\-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
Note that the ith colu

row of A is a zero ro
two rows. Similarly
four rows.
To find the coeffil
by using (A-11), from
1
4
1
4
We choose the (1, 3) element as the pivot and compute
-~
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.
The pivot element of the fourth row of A z is chosen as shown. We compute
Hence we have
or
where a is the ith rov
Hence we have .
ar
8This presentation was suggested by Proressor J. W. Wang.
549
ROW-SEARCHING ALGORITHM
ts right-hand side and

lcients of combination
,mputed by this simple
K4 K 3 Az = K4Az =
'ormation of the entire

~ows of K. Therefore
(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
The last row of is a zero row, and the search is completed.

to write
in (A-lO) and (A-ll)
-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
roceed to t he next row.

own. We compute
O
O
O
,- -
-; 1- ",-- -_:. -
[-2
or
: ....-1
~]
'-211:-1} ~1~'
--- --------I
Hence we have
11"'- - "
......
-1
O
OJA=O
-23 -az +a 3 =0
where a is the ith row of A.

To find the coefficients al cambination for the fifth
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
ELEMENTARY TRANSfORMA TIONS
This algorithm of searching linearly independent rows of A in the order of

the first row, the second row, and so forth will be referred to as the row-searching
algorithm. In this algorithm, if the ith row is linearly dependent, then the ith
row will not contribute to the linear combination for the jth linear dependent
row for j > i. For example, since a 3 is linearly dependent, its coefficient is zero
in (A-I4); hence as is expressed as a linear combination of its previous linearly
independent rows. In general, in (A-8), if the ith row, with i <j, of A is linearly
dependent, then bji =0. Indeed, if the ith row is linearly dependent, then the
ith row of A is a zero row, and the K in (A-12) is a unit matrix. Consequently,
all elements below ei in the ith column of F are zero. That is, epi = O, for
p = i + 1, i +2, ... ,j. Hence from (A-ll), we have bji =0. Because of this
property, the b jk computed in this algorithm are unique. Without bji=O, the
coefficients of combination may not be unique. For eX?'>lple, the addition of
(A-13) and (A-14) yields
-1
To find the coefficien
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 )
where Li are elementary matrices or the Householder transformations. 9 It is

assumed that al l 4=0, azz 4=0; and a43 4=0. Clearly (A-15) implies that the first,
second, and fourth rows of A are linearly independentof their previous rows
and the third and fifth rows of A are linearly dependent on their previous rows.
Householder transformations with pivoting {permutation of rows), however, are not permitted.
We see that the b ij a

stable method (see R
formed iuto the forro
easily obtained. I n ~
linearly dependent ro
be identical to the 01
this method of searcl
searching algorithm, 1
The row searchinl
for the following two
not complicate or ob~
can be easily carricd o
problems by hand, or
fidence in using a digi
chosen for ihe conver
element with value +
zero elements as the p
is for pedagogical reas
use the method discus
* A-4
Hessenber
Every square matrix

similarity transformal
values and eigenvectol
we shal1 use elemcnta
a matrix into a speci
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
ation. The property

shing Theorem G-14
of the row-searching
the leftmost nonzero
gaussian elimination
:he largest. element in
~est element in magni
y choosing the largest
alent to the gaussian
no guarantee that the
;ential nol to alter the
le arbitrarily altered.
mination with partial
)rm A into the form
~A
(A-15)
lnsformations. 9 It is
I implies that the first,
.f their previous roWs
[l their previous rows.
vever, are not permitted.
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
Every square matrix can be transformed into aJordan canonical form by a .

similarity transformation. However, this requites the computation of eigen-.
values and eigenvectors and is anumerically 1lI1stabh: problein~ In the followfng,
we shall use elementary matrices inthe similarity transformation to transform
a matrix into a special form, called the Hessenberg formo This form is very
552
ELEMENTARY TRANSFORMATIONS
important in computer computation and its generalized form will be used in

Section 5-8.
Let A be an n x n matrix. Consider
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
Use the row searcl
Repeat Problem A
Verify that b ji = O if the it

A-3 Is it possible to 01
Hessenberg form?
A-4 Show that if Q. i =
so is Q~Q,"Qm-l'" Qz
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
Use the crc!ed elements,

coefficients of its linear Co
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)
This is called upper Hessenhergjorm and can be obtained by using numerically

stable methods. This form is very useftil in many computer computations.
where R is an upper trian

is the first nI columns of <:
holder transformations a
() N.. f8.ctorizaton 8.::1.0 t~(
elements of JR may be diff
atter, they can be chosen .
A-6 Let A be a square
matrix A~ RQ has the s
are similar. The QR algo,
values oC A, is based 00 tI
PROBLEMS
I form will be used in
Problems
A-1
x
x
x
Use the row searching algorith"l to find linearly dependenl rows of
[~: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
Repeat Problem A-I for the matrix
\, except the first two

g gaussian elimination
The inverse of PI has
1
I P1
will not operate
in the first column of
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.)
The sum, product

any point) of an
nomials, exponential
entire real lineo
If a function is kn
determinable from an
known. This can be
derivatives at t o are
(t 0 - eo, t o +e o) Ne)
again, from (B-l), we
both directions, we c~
analytic continuation.
o at
B
Analytic Functions
of a Real Variable
Theorem 8-1
If a functionf is anal
arbitrarily smaH nonz
D.
Proof
Let D be an open interval in the realline IR and letf(') be a function defined on

D; that is, to each point in D, a unique number is assigned to f The function
fU may be real-valued or complex-valued.
A function fO of a real variable is said to be an element of dass C" on D
if its nth derivative, <"lO, exists and is continuous for aH t in D. ero is the class
of functions having derivatives of aH orders.
A function of a real variable, f('), is said to be analytic on D iff is anelement
of ero and if for each t o in D there exists a positive real number eo such that, for
aH t in (to - eo, t o +eo), f(t) is representable by a Taylor series about the point
to:
A remark is in order at this point. If f is a function of a complex variable

and if f has continuous derivative, then it can be shown that f has continuous
second derivative, third derivative, ... , and in fact a Taylor-series expansiono
Therefore, a function of a complex variable may be defined as analytic if it has
a continuous derivatVe. However, for functions of a real variable, even if a
function possesses derivatives of aH order, it may still not be analytic. For
example; the function
1/,2
f(t)= { ~ .
for t-l=O
for t =0
is not analytic at t = O,' even though it is infinitely differentiable at t = O; see

Reference 9.
554
If the function is iden

(to, ( 1 ), then the func!
By analytic continua'
ANALYTIC FUNCTlONS OF A REAL VARIABLE
555
The sum, product, or quotient (provided the denominator is not equal to

O at any point) of analytic functions of a real variable is analytic. AII poly
nomials, exponential functions, and sinusoidal functions are analytic in the
entire realline.
If a function is known to be analytic in D, then the function is completely
determinable from an arbitrary point in D ifall the derivatives at that point are
known. This can be argued as follows: Suppose that the value of ! and its
derivatives at t o are known, then by using (B-l), we can compute! over
(to.-8 0, t o +80)' Next we choose a point tI that is almost equal to t o +8 0 :
again, from (B-l), we can compute! over (to +80, t o +8 0 +8 1), Proceeding in
both directions, we can compute the function! over D. This process is called
analytic continuation.
Theorem 8-1
If a function! is analytic on D and if! is known to be identically zero on an

arbitrarily small nonzero interval in D, then the function! is identically zero on
D.
Proof
If the function is identically zero on an arbitrarily small nonzero interval, say

a function defined on
~d to f The function
lent of class C" on D
in D. CO is the class
:m D if! is an element
mber 80 such that, for
;eries about the point
(8-1 )
)f a complex variable
:hat f has continuous
{Ior-series expansiono
:d as analytic if it has
'al variable, even if a
IOt be analytic. For
entiable at t=O; see
(lo, tI), then the function and its derivatives are all equal to zero on (to, tI)'
By analytic continuation, the function can be shown to be identically zero.

Q.E.D.
where Ilu(t)II~(u(t),
Section 2-8.)
Proof
The solution of the sI
x(t 1
Defle
Minimum-Energy Control 1
x~
Then the assumption
-i
l1
X=
lO
Subtracting both side
Consider the n-dimensional linear time-varying state equation

which implies that
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,
t)~ l' <l>(t o, ,)B(,)B*(,)lI>*(to, ,) d,
\L
l!>(.
By using (2-94),we ca
With the use of(C-l),
lO
Then the input ufro,ld defined by

Iilio(t) = (<<!l(t o, t)R(t))*W - l(rO, tI )r$(i o. tdx! - Xo]
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
1', Ilul(t)112 dtc.l

'o
This appendix c10sely rollows Reference 69.
556
1,
lO
IluO(t)112 dt
Considcr now
By sorne manipulatio
557
MINIMUM-ENERGY CONTROL
where IIU(t)II~U(t), U(t)112~(U*(t)U(t))1/2, the euclidean norm of U(t)

Section 2-8.)
(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
Subtracting both sides, we obtain
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
( [ (t o, r)B(r)(ul(r) -uO(r)) dr, W-l(t o, tl)X) =0
By using (2-94), we can write this equation as
i"
'o
(u] (r) - uO(r), (4)(t o, r)B(r *W -I(t o, ti }x) dr
=o
(C-3)
With the use of(C-\), Equation (C-3) becomes
'l
) dr
(l.I l (r)-uO(r), uO(r tlr =0
(C-4)
'o
Consider now
LII
m 5-4. Now we show

energy, among all the
ul 2
l1 elr
By sorne manipulation and using (C-4), we obtain
r'l Ilul(r)W elr = Jtor" llul(r} -uo(r} +uo(r}11 2 dr
10
nd let UO be the control
n
= ['
Ilul(r} - uO(r}11 2 elr + [' IluO(r)W dr
+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
is always nonnegative, we conclude that

Q.E.D.
We see that in transferring (xo, to) to (Xl, t), if minimizing
is used as a criterion, the control defined in (C-I) is optimal. Since, in many

instances,
is related to energy, the control U o is caBed the minimum-energy control.
Consider the linear ti
where x is the n x 1 st
vector; A, B, and e ;
The response of FE
where Xo is the initial

We consider now
the input u changes v;
occur in sampled-dat
used to generate u.
sampler and a filter, e
u(t) = u
where T is a positive
0, T, 2T, ... , are called
constant inputs given "
the" behavior at samp
dynamical equation (
Q.E.O.
izing
Controllability after the

Introduction of Sampling
imal. Since, in many
-energy control.
Consider the linear time-invariant dynamical equation

FE:
x=Ax +Bu
( 0-1a)
y=Cx
(0-1b)
where x is the n x 1 state vector, u is the p x 1 input vector, y is the q x 1 output

vector; A, B, and e are n x n, n x p, and q x n constant matrices, respectively.
The response ofF E is given by
x(t)= eA1xo+
eA(I-'lBu(r)dr
y(t) = Cx(t)
(0-2a)
( D-2b)
where Xo is the initial state at t = o.

We consider now the case in which the input u is piecewise constant; that is,
the input u changes values only at discrete instants of time. Inputs of this type
occur in sampled-data systems or in systems in which digital computers are
used to generate u. A piecewise-constant funcHon is often generated by a
sampler and afilter, caBed zero-order hold, as shown in Figure 0-1. Let
u(t)
= u(k)
for kT 50t k + l)T; k =0,1,2, ...
(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
constant inputs.given in (0-3) can be computed from (0-2). However, if only
. the behavior at sam pling instants O, T, 2 T, ... , is of interest, a discrete-time
dynamicalequatio~ can be written to give the responseof x(k)~x(kT) at
559
560
CONTROLLABILlTY AFTER THE INTRODUCTION Of SAMPLlNG
u(e)
This pro blem is imp

See Problem 7-30.
A discrete-time e
for any initial state ~
q uence {u(n)} of finit<
tions in statements 3
are directly applicab
Let ;,(A) denote;
imaginary part and t:
.::::kJ
T~
Hold
u ~(e)
u(e)
u(
u(k)
"
1"
tjil
o
l.
2T
Theorem 0-1
L-J
Assume that the dYI

sufficient condition f
(0-5) to be controllat
whenever Re [A(A)
For the single-inp
Figure 0-1
k = O, 1,2, . . .. From (0-2a), we have

(k+ I)T
x(k+l)=e A (k+l)T xo + o
eA[kT+T-tIBu(T)dT
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)
tions of the theorem

conditions or the the(
values of A, then {Xi,
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.
x(k +1)=e ATx(k) + ( f eAada) Bu(k)
which is a discrete-time state equation. Therefore, ir the input is piecewise

constant over the same interva! T; and if only the respons"C ~t '-l'C s2xrpj",,.
instants is of interest, the dynamical equatio:1 FE in (0-1) can be replaced by the
following discrete-time linear time-invariant dynamical equation
DFE:
x(k + 1) = x(k) +Bu(k)

y(k) = Cx(k)
where
A=e AT
B=(f e
C=C
(0-5a)
(0-5b)
where
(0-6)
At
dT)
B~MB
(0-7)
(0-8)
lfthe dynamical equation FE iscontrollable, it is ofinterest to study whether

the system remains controllable after the introduction of sampling or, corre
spondingly, whether the discrete-time dynamical equation DF E is controllable.
CONTROLLABILlT'I"'!\FTER THE INTRODUCTlON OF SAMPLlNG
This problem is important in the design of dead-beat sampled-data systems.

See Problem 7-30.
A discrete-time dynamical equation DF E is defined to be controllable ir
for any nitial state Xo and any x 1 in the state space, there exists an input se
q uence {u(n)} offinite length that transfers X o to x l' The controllability condi
tions in statements 3 and 4 ofTheorem 5-7 and in Theorem 5-21 and its corollary
are direct\y applicab\e to the discrete-time case w1thout any modification.
Let ),(A) denote an eigenvalue of A, and let "1m" and "Re" stand for the
imaginary part and the real part, respectively.
---
h
I
l.
2T
561
Theorem 0-1
L-J
Assume that the dynamical equation FE given in (D-1) is controllable. A

sufficient condition for the discrete-time dynamical equation DFE given in
(D-5) to be controllable is that 1m [,.li(A) - }./A)] i= 2nr:t./T for IX = 1, 2, ... ,
whenever Re [A(A)-A/A)] =0.
For the single-input case (p = 1), the condition is necessary as well.
I
u(r) dr
(0-4)
u(r) is constant in the

Jmes, after the change
First we remark on the conditions of the theorem. Because of = eAT , ir

Ai is an eigenvalue of A, Xi ~ eAT is an eigenvalue of (Problem 2-32). Let
}'l =r +jf3 and Az =r +j(f3 +2nrt./T), for IX = 1, 2, .... Then we have
Xz = <+ (P + Zna.IT))T = e+ jPlT + jZna. = ei<+ jPlT = Xl' In other words, if the condi
tions of the theorem are violated, different Ai may yield the same Xi- Hence the
conditions of the theorem ensure that if {Ai' i = 1, 2, ... , m} are distinct eigen
values of A, then {Xi' i = 1,2, ... , m} are distinct eigenva\ues of .
Proof of Theorem D-1
We assume, without loss of generality, that A is in the Jordan canonical form

shown in Table 5-1. Then, from (0-6), (2-70), and (2-71), we have
:)
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)
~rest to study whether
O O
and 1i =
eAT
(nij 2)!
T(lI i j - 3 l
O O
If sampling or, corre

1 DFE is controllab\e
T(II;j-l)
(0-5a)
(0-5b)
(D-10)
(nij-3)!
Although A i is a Jordan block, ijis not in aJordan formo Now
------
562
--..
----~
.'~--
- ------ - - - - - - - - - - --- - - - - - - - - - - - - - - - - - - - _..
__ ._---------_._----._--_..
-.~------_.-
CONTROLLABILITY AFTER THE INTRODUCTION OF SAMPLING
we shall transform it into a lordan formo We compute
o
ij- XI =
x x
O O x
O O O
O O O
O O O
O x
O O
where x denotes possible nonzero element. Because of its special form, it is

easy to verify that (ij - XJ? has one more zero superdiagonal row than
( ij - XJ); and finally, similar to (2-64), (j - XJ)"i-l has only one nonzero
element at the upper right comer and (ij - XJ)"j is a zero matrix. Conse
quently, e = [O O ... O 1]' is a generalized eigenvector of grade nij (see
Section 2-6). From this vector we can obtain a similarity transformation of
the form
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
Consider the sample

expansion, we have
X
O
Theorem D-2
If a continuous-tim~
its discretized state ec
O 1
,
II.
condition of the the,

if T > O, M is nonsin
If is in the for[
b/ij, j = 1, 2, ... , r(i), i
tions on the eigenval
Hence the necessary
trollable is that the r,
rU), are liriearly inde
the case if and only i
conclude that under
the discrete-time equ
time equation in (0
In the single-inpl
there will be two or r
Equation (0-14). H
lary 5-21, not con tro
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)
then (O-Sa) can be transformed into

x(k
+ 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
CONTROLLABILITY AFTER THE INTRODUCTlON OF SAMPLING
. its special form, it is

lerdiagonal row than
h.as only one nonzero
zero matrix. Conse
~ctor of grade nij (see
rity transformation of
condition of the theorem implies Ai T +- 2mx, if Re )'i = O or e - i.T +- 1. Hence

if T > O, M is nonsingular.
lf A is in the form in Table S-1 and if {A, B} is controllable, then the rows
b1ij , j = 1, 2, ... , r(i), in (S-71a) are linearly independent. Now under the condi
tions on the eigenvalues stated in Theorem 0-1, Xi, i = 1, 2, ... , m, are distinct.
Hence the necessary and sufficient conditions for {PAP-, PMB} to be con
trollable is that the rows of PMB corresponding to the rows of buj , 1= 1, 2, ... ,
r(i), are linearly independent. Because of the special forros of P and M, this 1S
the case if and on!y if b1ij,j = 1, 2, ... , rU) are linearly independent. Hence we
conclude that under the conditions on the eigenvalues stated in Theorem 0-1,
the discrete-time equation in (O-Sa) is controllable if and only if the continuous
time equation in (O-la) is controllable.
In the single-input case, if the conditions on the eigenvalues are violated,
there will be two or more Jordan blocks associated with the same eigenvalue in
Equation (0-14). Hence the single-input equation (0-14) is, following Corol
lary S-21, not controllable. This establishes the theorem.
Q.E. D.
In the proof of Theorem 0-1, we have essentially established the following
theorem.
(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
We see that the Jordan

fine
(s
+ 1)(s + 1 + 2i)(s + 1 - 2i) s + 1
(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
CONTROLLABILlTY AFTER THE INTRODUCTION OF SAMPLING
Consequently, an irreducible Jordan-form realization of the system S can be

found as
o
O
-1-2i
O
y = [2
-1
][XI] +[1]
-1+2i
1 u
1
x?
X3
-I]x
By using (0-5) through (0-8), a discrete-time state equation can be computed as
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.
or, in matrix form,
Problems
0-1
Consde:r the: continuolls-time: state e:qllaton
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 the mJs are an

Sinc~ every hermitio.f
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 (
the system S can be
can be computed as
r:I
T
Hermitian Forms and

Singular Value Decomposition
-T- 2iT)
u(k) (D-15)
- T+ 2iT)
state equation (D-15)
A hermiLian form of n comp\ex variables

geneous polynomial of the form
X,X2, ... , X Il
is a real-valued homo
11
',j~
Ig either the criterion
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
x*Mx =(x*Mx)* =x*Mix

IY T including T = 2ncx/4
lValues in Theorem D-1
;on tro Hable.
11 equation in (D-S) to be
lnd 1m [},(A)-})A)] f
r the single-output case
where Mi is the complex conjugate transpose of M; hence
x*Mx =x*(~(M +Mj')x~x*Mx
(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
HERMITIAN FORMS AND SINGULAR VALUE DECOMPOSITION
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 )
Now we show thal

matrix can be chosen '
following theorem.
where, in the last step, we have used the fact that M* = M.
Theorem E-3
Theorem E-1
The eigenvectors of a f:
are orthogonal.
All the eigenvalues of a hermitian matrix M are real.

Proof
Proof
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
The Jordan-form representation of a hermitian matrix M is a diagonal matrix.

Proof
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 =
which implies (M - illl- le = O. This is a contradiction. Hence there is no

generalized eigenvector of grade k ~2. Consequently, there is no Jordan block
whose order is greater than one.' Hence the Jordan-form representation of a
hermitian matrix is a diagonal matrix. In other words, there exists a nonsingu
lar matrix P such that PMP - 1 = M and M is a diagonal matrix with eigenval ues
on the diagonal.
Q.E.D.
Two vectors x,. y are said to be orthogonal if and only if (x, y) = O. A vector
x is said to be no.rmalized if and only if (x, x) ~ IIxl1 2 = 1. It is' clear that every
vector x can benormalized by choosing x = (l/llxll)x. A set of basis vectors
{ql, Q2, .. . "qll} issaid to be an orchonormal basis if and only if
i 1=}
i=}
Let e and ej be the eig

), and }'j, respectively
( E-6)
um=e
The procedre for de

Schmidt oi,thonormalizl
that (q, qj'j :=0 for i i=

Fram Theorem E
conclude thatfor any r.
with'respect lo which t
or equivalently, for any
HERMITIAN FORMS AND SINGULAR VALUE DECOMPOSI1'ION
form can be written

(E-4 )
567
Now we show that the basis of a lordan-form representation of a hermitian

matrix can be chosen as an orthonormal basis. This is derived in part from the
following theorem.
Theorem E-3
The eigenvectors of a hermi tian matrix M corresponding to different eigenval ues

are orthogonaL
Proof
M associated with A;
( E-5)
l number, [ram (E-S)

Q.E.D.
is a diagonal matrix.
laps (C", C)into'itself

.t give a Jordan-form
eigenvectors of the
ere is no generalized
ion. Suppose there
.I)k - le
for some
Let e and ej be the eigenvectors of M corresponding to the distinct eigenvalues

A and Aj, respectively; that is, Me = Ae and Mej = Aje j. Consider
and
(ej, Me) = (ej, Ae) =A(ej, el)

(ej, Me) = (Mej, el) = (Ajej, el) = )'j (ej, el)
( 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.
Since every eigenvector can be normalized and since eigenvectors of a

hermitian matrix M associated with distinct eigenvalues are orthogonal, then
the eigenvectors associated with different eigenvalues can be made to be ortho
normaL We consider now the linearly independent eigenveCtors associated'with
the same eigenvalue. Let {e 1, ez, ... , em } be a set of linearly independent eigen
vectors associated with the same eigenvalue. Now we shall obtain a set of
orthonormal vectors from the set {el, ez, . .. , e m }. Let
+0
Uz =ez - (ql, eZ)ql

.. Hence there is no
:re is no lordan block
n representation of a
ere exists a nonsingu
atrix with eigenvalues
Q.E.D.
, (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
The pracedure for defining q is illustrated in Figure E-l. 1t is called the

S,chmidt orthonormalization pr0cedure. By direct verification, it can be shown
that (q, qj) =0 for i j.
'From Theorem E-3 and the Schmidt orthonormalization procedure we
condude that for any hermitian matrix there exists a set of orthonormal vectors
with respectto which the hermitian matrix has adiagonal-form representation;
or equivalently,for any hermitian matrix M, there exists a nonsingular matrix Q
568
HERMITlAN FORMS AND SINGULAR VALUE DECOMPOSITION
If the rank of H is r. SI
A21 ,
z2 >
. . . '''j'
; 2 > O a'ne
the orthonormal eigenv
---
----
Q = [ql
qz
Then Theorem E-4 imp

ql
~<
ql,e 2
where :E 2 = diag{ A, },~,
Figure E-1
and
whose columns are orthonormal, such that
which implies
!VI = Q - lMQ g PMP - l

where M is a diagonal matrix and P g Q - l. Let Q = [ q l
Because of the orthonormal assumption, we have
Q'Q
Ji::;:]
[ql
q2
qz
qll].
where 2: = diag {Al' Az,
Then (E-13) becomes Rj

normal. Let Hz be cho
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
A hermitian matrix M can be transformed by a unitary matrix into a diagonal

matrix with real elements; or, equivalently, for any hermitian matrix M there
exists a nonsingular matrix P with the property P - l = p* such that
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
This is stated as a theor

Theorem E-5 (Singula
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
HERMITiAN FORMS AND SINGULAR VALUE DECOMPOSITION
569
If the rank of H s r, so s the rank of H*H (Problem E-3). Hence we have

Ai ~Al ~ '" X; >0 and A;+ =A;+z = ... =A~ =0. Let q, i = 1,2, .. " ,n,be
the orthonormal eigenvectors of H*H assocated wth }.f. Define
Q = [q
qz
...
q,.: q,+
(E-10)
Then Theorem E-4 implies
Q*H*HQ=[~Z ~J
where 1;z = dag{Ai, J,i, ... ,
An.
Using Q = [Q
(E-11)
Qz], (E-lI) can be written as

( E-12)
Q!H*HQz =0
and
QtH*HQ =1;2
which implies
1;-IQtH*HQ1;- =1
Q = [q
qz
...
qll].
( E-13)
where 1; = diag {A, Az, ... , J.,.}. Define the m x r matrix R by

Rl~HQ1;-
(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
, matrix into a diagonal

rmtan matrx M there
~* such that
. [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
This s stated as a theorem.

Theorem E-5 (Singular value decomposition)
)ll
the dagonal.
velop the singular value

Ltrix. Then the matrix
mitian; hence its eigen
e, ts egenvalues are a11
:he eigenvilJues of H*H.
DI H. {fH iS hermtian,
,f the eigenvalues of H.
( E-9)
Every m x
matrix H of rank r can be transformed into the form
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
HERMITlAN FORMS AND SINGULAR VALUE DECOMPOSITlON
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
Find an orthonormal set from the vectors
What are the eiger
El may arise in gaussian e

tion with partial pivoting.
largest and smallest sinDlI
speaking, a condition ~u:
computation.
E-11
Show that the conl
Find the eigenvalues and singular values of the matrices
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
E-12 Show that Theorerr

tive semidefinite. lf H i:
what are the differences be
E-13
mm UJ
E-4
E-10
is positive definite if and e
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
E-9 Show that aH the ~

are equal to 1.
fl
h2=l~
Il
(j
~J
E-5
Find the singular value decompositions of the matrices in Problem E-4.
E-G
Show that
IIAlb =largestsingular value of A

= [)'max(A *Al] 1/2
E-7 In Theorem E-S, we have Q*Q = QQ*= 1... If we write Q
Ql*Q = 1,. and QQf = In?
.
= [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
om (E-14). The choice

.nitary.
applications in linear
ible realization from a
. approximated models
['he singular value de
and stabili ty margin of
For eomputer pro
itted to assume real or
hey are limited to real
lere the prime denotes
ymmetric matrix; a real
'ith these modificatiuns
irectly to real matrices.
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
Show that the controllability grammian
W" =
eA'BB*e A', de
is positive definite if and only if {A, B} is controlIable.

hat x*Mx is a real number
with real coefficients, is it
'=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
What are the singular values of the following matrices?
[1 -IJ
A _
2
md HH* (Hint: Find a P
n Problem E-4
orthogonaI matrix) real'!

tdes eq uaI to 1.
-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
In this appendix we shall study the rnatrix equation AM + MB = N, where

A, B, M, and N are n x n complex-valued matrices. We observe that all the
n x n complex-valued matrices, with the usual rules of multiplication and addi
tion, form a linear space (Problern 2-10). Let us denote this space by (X, C).
The dimension of (X, C) is n 2 Consider the operator d defined by
d(M)~AM +MB
If Y/kI - A and B 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,
The surnmation of the

with IX o = 1, yields
which, together with (

ror sorne i and sorne j.
Proof
We prove, first, that A + I1j is an eigenvalue of d. Let x andy be nohzero nx 1
and Ix n vectors such that
Ax =AX
572
I
~
. .
Since Ll(B) is nonsing

tion that M /=-0. Hen
eigen vaJue. N ow the I
ror sorne i and for son
573
ON THE MATRIX EQUATION
c1early x is a right eigenvector of A and y is a left eigenvector of B. Applying

the operator d on the n x n matrix xy that is c1early a nonzero matrix, we
obtain
d(xy) = Axy
+ xyB = },xy + Jijxy = (/' + Jij)xy
Hence }' + Jij is an eigenvalue of d for i = 1, 2, ... , 1, and j = 1, 2, ... , In.

See Definition 2-12.
Next \Ve prove that aH eigenvalues of d areofthe form }' + Jlj. Let us assume
that rk is an eigenvalue of d. Then, by definition, there exists a M f. Osuch that
or
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
be the characteristic polynomial o I1kI - A, that is Ms) = det(sI -11kI

Then we have
+ 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.
lf I1kI - A and B have no common eigenvalue, then the matrix

(B)
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
(r,J - A)"M = MBn
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
Any matrix representation of the operator d is nonsingular if and only if

A +fJ-j=/=-O for aH i,j.
Let s1: (X, C)->(X, 1[

M in 7... Let A, for i =
for j = 1, 2, ... , m ~
eigenvalue of d. Co
values of .sd, then for
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.
for sorne i and sorne j

Corollary F-2a
Any matrix represenl

AfJ-j=/=-l for aH i,j.
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
Furthermore, the solution M can be expressed as

M =
I'"
eA"Ne At dt
(F-5)
Furthermore, the solut

Proof
Since the eigenvalues of A * are the complex conjugates of the eigenvalues of A,
aH the eigenvalues of A and A* have negative real parts. Consequently,
A + fJ-j=/=-O for aH i, j, which implies that the matrix representation of d(M) =
A*M +MA is nonsingular. Hence, for any N there exists a unique M satisfying
d(M) = - N.
Now we show that the M in (F-5) is the solution of (F-4). Because all the
eigenvalues of A have negative real parts, the integral in (F-5) converges.
CQusider
A*M +MA=
roo (A*eA'tNe AI
It=
+eA"NeA!A)dt=
The procedure usel

F-2. CoroHary F-2b (
three methods of solvi
roo ~(eA'INeAI)dt
Jo
00
=e tNeA'[
= - N
1=0
This establishes the coroHary.
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.
. . . .
We list in the rollowing the discrete-time version ofTheorem F-l.
.'
] By eoua.ting the .--;:.

obtain a set of Hne
consists of n 2 equat
the hermitian or Sy
linear algebraic equ
a computing center
2. Compute the infinit
have magnitudes Ir
infinite series can b,
M,
The computation m
than a predeterrnim
ON THE MA TRIX EQUAnON
575
Theorem F-2
ngular ir and only ir
near space into itself,

representation can be
~ = e in the form of
f all the n2 elements of
leterminant of A is the
Q.E.D.
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)
for some i and some j.

Corollary F-2a
Any matrix representation of the operator d is nonsingular if and only if

A!lr/= 1 for all i, j.
)r any N there exists a
Corollary F-2b
(F-4)
(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)
Furthermore, the solution M can be expressed as

00
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
The procedure used to prove Theorem F -1 can be used to establish Theorem

F-2. Corollary F-2b can be readily verified by direct substitution. We discuss
three methods of solving (F-7) to conclude this appendix:
L By equating toe corresponding elF.ment:o. 0f A*M.<I:\ - "lIT
Q.E.D.
formula is not suitable

8 = N has been exten
References S9, S18, and
fheorem F-1.
~.Dd-
i"{ WP; r:2 n
obtain a set of linear algebraic equations. Ir A is an n x n matrix, the set

consists of n 2 equations. The number of equations can be reduced by using
the hermitian or symmetry property of M (see Reference S16). The set of
linear algebraic equations can then be solved by using existing subroutines in
a computing center.
2. Compute the infinite power series in (F-8) directly. Ir all the eigenvalues of A
have magnitudes much less than 1, the series will converge rapidly.. The
infinite series can be computed recursively as follows:
with M o =0
The computation ma) be stopped when
than a predetermined number.
IIMk+l-' Mkll first becomes snialler
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
Find a matrix representation of the operator $i'(M) = AM + MB with
A=[_~ -~J B=[~ ~J
and verify Theorem F-l.

F-2 Find the M to meet A*M +MA = - N with A given in Problem F-l and N = 1,
by setting up a set of linear algebraic equations.
F-3 Solve Problem F-2 by using (F-S). Which method, by solving algebraic equations
or by direct integration, is simpler?

F-4 Establish, without using Corollary F-la, the uniqueness part of Corollary F-lb.
(H int: Use d(e A"(M 1 - M 2 )e A')jdt = O.)
F-5
Under what conditions can the solution ofAM +MB = - N be expressed as

M
F-6
t"
eA'NeB1dt
Prove Theorem F-2 and its corollaries.
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
Salve (F-7) with A and N given in Problems F-l and F-2.
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!
Let D(s) be a polyoomi

D(~
The polynomial D(s)

associated with the h
polyoomial is called n
In this appendix, capital le'
speeded up (see Refer
G
Polynomials and
Polynomial Matrices
+MB with
Problem F-1 and N
= 1,
Iving algebraic equations
part of Corollary F-1 b.
N be expressed as
\ and B given in Problem
IOV
equation by using the
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
Let D(s) be a polynomial with real coefficients and indeterminate s expressed as 1

D(s) = Dns' +D._1s- 1 + ... +Ds +D
o
The polynomial D(s) is said to be of degree n if D. =1=0. The coefficient Dn

associated with the highest power. of s is caBed the leading eoe,ffieient. A
polynomial is caBed monie if its leadingcoefficient is equal to 1.
'In this appendix, capital letterswithoutboldface are also used lo denote scalars.
577
578
POLYNOMIALS AND POLYNOMIAL MATRICES
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
At the end, the result

In Equation (G-1:
by D(s), and D(s) is a.
a divisor of N(s), theI
and N(s). Note that
D(s) and N(s), and is
trivial common facto
N(s)=Q(s)D(s) +R(s)
and
deg R(s) < deg D(s)
(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
N(s) = (2s - 2)D(s) +(9s - 5)
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)
IrQ(s)-Q(s) 1=0, the degree of the left-hand-side polynomial of (G-2) is equal

to or larger than that of D(s), whereas the degree of the right-hand-side poly
nomial of(G-2) is smaller than that of D(s). This is not possible. Hence we have
Q(s) =Q(s)
and
R(s)
R(s)
Q.E.D.
This completes the proof of the theorem.
The division procedure in this theorem can be put into a programmatic

format as follows:
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)=
This process will even

From the last equ
remainder, by Rp(s). '
(1 + QpQp+ t}Rp(s); he
ward, it can be ShOWI
Rp(s). Hence Rp(s) is ;
Now we claim th~
and Y;(s) such that R.
with X(s) = -Q(s) a
equation shows that
N
'.
N(s)+-N(s) --:-'D m s'"'-"O(s)
"
'Ir D" has avery small absol

or this algorilhm. Hencc
the gaussian elimination w
COPRIMENESS OF rOL YNOMIAlS
579
5. Go to step 2.
Jen there exist unique
(G-1 )
:t deg N(s) ::?:deg D(s).

h Q(s) as the quotient
or example, if N(s) =
Ne have
there are other Q(s)
At the end, the resulting N(s) is the remainder R(S).2

In Equation (G-l) if R(s) =0, N(s) is said to be divisible (without remainder)
by D(s), and D(s) is afactor, or a divisor, of N(s). lf R(s) is a divisor of D(s) and
a divisor of N(s), then R(s) is called a common divisor or common factor of D(s)
and N(s). Note that a nonzero constant is a common factor of every nonzero
D(s) and N(s), and is called a trivial common factor bf D(s) and N(s). A non
trivial common factor wll be a polynomial of degree 1 or higher.
Definiton G-1
A polynomial R(s) is a greatest cammon divisor (gcd) of D(s) and N(s) if R(s) is a
common divisor of D(s) and N(s), and is divisible by every common divisor of
D(s) and N(s). Ir a gcd of D(s) and N(s) is a nonzero constant (independent of s),
then D(s) and N(s) are said to be relatively prime or coprime.
1
In other words, two polynomials are coprime if they have only trivial com
mon factors. Ir they have nontrivial common factors, they are not coprime.
The gcd is unique only up to a constant; that is, if R(s) is a gcd, then cR(s),
for any nonzero number e, is also a gcd. Ir we require a gcd to be monic, then
the gcd is unique.
Given two polynomials D(s) and N(s), by a sequence of long divisions, often
caBed the euclidean algorithm, we can write
deg R <deg D
deg R 2 < deg R
deg R 3 < deg R 2
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)
lmial of (G-2) is equal

rght-hand-side poly
;sible. Hence we have
Q.E.D.
into a programmatic
D(s) with m> n:
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
1. If D(s) and N(s) an

Hence the matrix
are not coprime, ti
and the rank of [
has rank 1 at ever~
its rank only at th
2. This follows direci
3. If D(s) and N(s)
divisor, we obtain
coprime, there exi~
= 1. The substitu
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
where X(s) and Y(s) are polynomials.

Although every gcd is expressible in the form of (G-4), the converse is not
true. That is, a polynomial expressible as in (G-4) is not necessarily a gcd of
D(s) and N(s) (why?).
Theorem G-3
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)
3. There exist no polynomials A(s) and B(s) such that

N(s)
--
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)
deg A(s) < deg D(s)
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
:here exist polynomials

(G-4)
and N(s) divides R/s).

(s) and N(s) =(s)C(s),
e Rp(s) is divisible by
R/s} is a gcd. AH gcd
pressed in the forrn of
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
4), the converse is not

ot necessarily a gcd of
fhen D(s) and N(s) are

IS holds:
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
D(s)=D o +DIs +Dzs z +

N(s)=N o +NIs +Nzs z +
every root of D(s), the
(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)
~
-
: '__ :_:__ :_:__ :::

No
O
O
see Reference S125).
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
It is a square matrix of order n + m and is called the Sylvester matrix or the

resultant of D(s) and N(s). The substitution of(G-8) to (G-l1) into (G-7) and
equating the coefficients of Si, i =0,1, ... , n + m - 1, yields
[-B o -B -B 2 ... -B m - : A o A A 2 '"
An-JS=[ -B: AJS=O
(G-13)
This is a set of n +m linear algebraic homogeneous equations. We see that

the polynomial equation -B(s)D(s) + A(s)N(s) =0 has been transformed into
the equation in (G-13). Ifthe resultant S is nonsingular, then the only solution
of (G-13) is the trivial solution A(s) = B(s) = O. In other words, there exists no
A(s) of degree n -1 or less to meet N(s)!D(s) = B(s)/A(s); hence D(s) and N(s) are,
following Theorem G-3, coprime. Ir the resultant S is singular, a nontrivial
solution [-B AJ exists in (G-13). In other words, polynomials A(s) of
degree n -1 or less and B(s) exist such that N(s)!D(s) = B(s)!A(s), and D(s) and
N(s) are not coprime. Hence we have established the following corollary.
Corollary G-3
The polynomials D(s) and N(s) are coprime if and only if their Sylvester matrix
S defined in (G-12) is nonsingular.
11
Whether or not D(s) and N(s) are coprime hinges on the existence of a non
trivial solution in (G-13). Mariy numerically stable meihods and "canned"
subroutines are available in the literature and computing centers to solve this
problem. However, we are interested in only the special solution which yields
the smallest degree in A(s). The row searching algorithm discussed in Appendix
A turns out to yield such a solution. This is illustrated by an example
Note that the matri~

There are five nonzel
singular. Correspon,
six rows of F, by usir
Using (G-I0), (G-ll),
and
Example 1
Consider the polynomials D(s) = - 2s 4

-2s -1- 3. V;fe forrn thefr resultant~
+ 2s 3 -
S2 - s
+1
and N(S)=S3 +2s 2
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:
We note that the j

first linearly depender
right-hand-side matri
B(s) and ,1(s) of sma\'
linearly dependent rO\
(G-15). Thisis noi 1
(G-16) are coprime.
In the Sylvester 1m
in the ascending pOwe
descending power of 5
COPRIMENESS Of POLYNOMIALS
'y/vester matrix or the

(G-ll) into (G-7) and
ls
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
18=[ -B: AJ8=0

(G-13)
luations. We see that

leen transformed into
then the only solution
words, there exists no
ence D(s) and N(s) are,
singular, a nontrivial
polynomials A(s) of
~(s)/A(s), and D(s) and
lowing eorollary.
their Sylvester matrix

I
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
Using (G-I0), (G-ll), and (G-13), we have

A(s) =S2 -0.5
and
1 and N(s) =S3 +2s
tI,
'"
583
Nls)
D(s)
B(s)
A{.S)
B(s) = -0.5s -1.5
-0.5s -1.5
s2-0.5
+3
-2s 2 +1
( G-16)
Hence D(s) ana N(s) are not coprime.
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
~ ~ -~-o -~ _~n_ P~II~_I- - !J~1/----~ ~ ~ -~ ~o-J

-
- -
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
In general, B m - I . O. Hence the degree of B(s) is m -1, and consequently,

the degree of A(s) is n - 1. However, if the degree of the gcd of D(s) and N(s)
is 2 or higher, then B(s) and A(s) have a common factor of form Sk. No such
problem will arise in using (0-12). Hence we shal1 arrange the coefficients of
polynomials in ascending order throughout this appendix.
G-2
Reduction of Reducible Rational Functions
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)
N o assumption of N n . O is imposed; hence the degree of N(s) can be smaller

than n. We define, for k =0,1,2, ... ,
Do
No
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
we can obtain, similar

[ -B o
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)
If Sk has a full row ral

the trivial. solution A(
nontrivial solution wil

degree among all A(s) 2
degree of A(s) to meet
N rows in S .
_.
__
--~-
.------~--------~----~--
- --- --
- - ~ ~ - - ---------~
-- -- -
----
---~-~~--
REDUCTION OF REDUCIBLE RATIONAL FUNCTIONS
~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
It is a 2(k + 1) x (n + k + 1) matrix and consists of k + 1 block rows. Each block

has two rows formed from the coefficients, in the ascending power of s, of D(s)
and N(s). We note that each block is the shifting of its previous block to the
right by one column. If k < n -1, Sk has more columns than rows; if k = n -1,
Sk is a square matrix; if k > n - 1, then Sk has more rows than columns.
We search now linearly independent rows of Sk in order from top to bottom.
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;
)f N(s) can be smaller
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)
we can obtain, similar to (G-13),
[-13 o A o : -B 1 Al:"': -B k Ak]Sk=O

k + l block rows
(G-22)
Then from the equation

- B(s)D(s)+ A(s)N(s) =0
} l block row
(G-21 )
Since Sk is a 2(k + 1) x (n +k + 1) matrix, if k.:::;; v - 1, then Sk has a full row rank.

A necessary condition for Sk to have a ful\ row rank is that 2(k + 1)':::;; n + k + 1
or k.:::;; n - 1. Hence we conclude that v':::;; n or, eq uivalently, the total number of
linearly independent N rows in Sk is at most equal to n.
Let A(s) and B(s) be two polynomials defined as
( G-18)
( 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).
A matrix A(s) with

Similar to matrices \
elementary operatior
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)
D(s) - A(s) - ,.4(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
with e f= O and d(s) is

elementary matrices ~
inverses are
O O O
O 100
O O 1 O
O O O c- 1
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
nearly dependent ofits

ent, this N row is the
jing to this dependent
s)such that N(s)/D(s)=
exist A(s) of a degree
Polynomial Matrices
A matrix A(s) with polynomials as elements is called a polynomial matrix.

Similar to matrices with elements in IR 01' e, we may introduce the following
elementary operations on A(s):
1. Multiplication of a row 01' column by a nonzero real 01' complex number.
2. Interchange any two rows 01' two columns.
3. Addition of the product of one row 01' column and a polynomial to another
row 01' column.
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)
han columns and solu

A(s) =D(s) and B(s) =
nA(s)=D(s)(s+c) and
Clearly these solutions
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
Id its proof is left as an

of N(s)/D(s) hinges on
)W searching algorithm
illustrated in Example
is a shift of i ts previous
t method is developed '
t row of Sk' To fully .
,nried out from left to
For hand calcu'ation? '
, searching algorithm.
le matrix case.
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)
They are again elementary matrices. The premultiplication. of 1E DIi A(s)

operates Gil the rows 01' A(s), whereas te postrnuiUpiicalioii 01 Ji<; on A(s)
operates on the columns of A(s). For example, EzA(s) interchanges the second
and fifth row. E 3 A(s) adds the product of the second row of A(s) and des) to the
fourth row of A(s); whereas A(s)E 3 adds the product of the fourth column of
A(s) and des) to the second column of A(s). We cal! EiA(s) elementary row
operations and A(s)E elementary column operations.
The set of polynomials does not form a field beca use its inverse in multi
plication is not a polynomial. Ir we extend the set to include all rational func
tions, then the set becomes a field. Fo'r matrices with elements in 'thefield of
rational functions, the concepts of linear independence, rank, and singulafity
developed for matrices with elements in the field of real 01' complex nUJ;nbers are
equally applicable. Hence,' ifwe c6nsiderp61ynomials as eiementsof the
field of rational functons; then we may apply the concept of linear dependence
588
and rank to polynomial matrices.

nomial matrix
S
[
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
For example, the determinant of the poly
S +2
+3s +2
has a determinant of (s + 2)(S2 - 1) - (s - 1)(S2 + 3s + 2) = O. Hence the matrix

is singular. Consequently, there exist rational functions a(s) and az(s) such
that
(G-27)
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
The first r rows are

nOnzero element, a.ki
of ai,k; must be on th,
is, k <k 2 < ... <k,.. ~
the degree of a,k;, tha
lf deg ai,k=O, then l
these properties is sa
of (G-28) without th,
just a upper right tria
lf A is a matrix v
echelon (row)form. JI
row is 1 and all elen
words, the leftmost J
column. By reversir
obtain a different, bl
appear in the remain(
Every polynomia
ementary row operat
Triangularization Pn
Hence we conclude that polynomja 1
where IX! (s) and 1X 2(S) are polynomials.

veetors are linearly dependent in the field of rational funetions if and only ir
they ean be made dependent by using only polynomials as eoeffieients 4 (see
Problem 2-9).
Let A(s) be a polynomial matrix with rank r in the field of rational funetions
IR (s). We show that, by using exclusively elementary row operations, A(s) can
be transformed into the form
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
terminant of the poly-
klth
kzth
k 3 th
k,th
column
column
column
column
in the field of rational

. full rank. The non
ioes not imply that the
ltrix has rank 1, rather
rix, which is a special
(s), the field of rational
ample, the polynomial
,O. Hence the matrix

s o:\(s) and 0:2(S) such
(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
lude that polynomial

nctions if and only if
ls as coefficients4 (see
i of rational functions
'" operations, A(s) can
r chapter 2), then (.'r, IR ,) is

le cons'idered as an elemen(
~n[s], !R[s]). A sel of poly
e set is linearly independent
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
Step 5. M(s)~M2(s) and go to step 2.

Step 6. Delete the first row and the first column of M(s) and rename it M(s);
go to step 1.
i
Steps 2 to 5 will reduce the degree of the (1, 1) element of M(s) by at least 1;
hence, after a finite number of iterations, we will go to step 6. We repeat the
process for the submatrices of M, and eventually we will transform A(s) into
the form in (G-28). This completes the triangularization of the matrix.
In order to have the property deg aj.k, < deg ai,k, for j = 1, 2, ... , i - 1, we
need sorne additional row operations. Let aj,k = qj(s)ai,k; + a],k; with deg a],k; <
deg ai,k. We add the product of the ith row and - qj(s) to the jth row, the
resulting matrix will retain the form of (G-28) and has the property deg aj,k <
deg ai,k;, j = 1, 2, ... , i - 1. Thus we have established the following theorem.
Theorem G-5
Every polynomial matrix can be transformed into the Hermite row form in
(G-28) by a sequence of elementary row operations.
1
The product of the
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(
We give an example to illustrate this theorem. Consider
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
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
.of complex numbers f
which are the roots of
The rnk of a poly

or postmultiplied by
modular matrices.
POLYNOMIAL MATRICES
59!
The product of these five matrices is

;) and rename it M(s);
[~
t of M(s) by at least 1;
;tep 6. We repeat the

ill transform A(s) into
1 of the matrix.
r j = 1,2, ... , i - 1, we
+ a},k, with deg aJ.k; <
(s) to the jth row, the
le property deg a,k; <
he following theorem.
Hermite row form in

I
( G-30)
Hence, we have
3s + 1 ]
s2+ s -2
S2 +2s -1
=1 __ ~ __ ~~I_-J
loo
(G-31)
!!l
In this example, if we are interested in only the triangularization, we may

stop at step 4. Dual to Theorem G-S, every polynomial matrix can be trans
formed, by a sequence of elementary column operations, into a lower left
triangular matrix or the Hermite column form which is the transpose of (G-28).
Since the determinants of elementary matrices are nonzero constants and
independent of s, so is the polynomial matrix in (G-30). Such polynomial
matrices are called unimodular.
Definition G-2
A square polynomial matrix M(s) is called a unimodular matrix ifits determinant
is nonzero and independent of s.
ii
Theorem G-6
-S+2 ]
s3+ s2+ s +1
s+l
-5
[i
~ + 1]
A square polynomial matrix is unimodular if and only if its inverse is a poly

nomial matrix.
Proof
(G-29)
'. This is achieved by

:d by lUsted at the ene!
'st row by - 1, add the
the product of the first
matrix denoted by 2.
ow. We then add the
v in step 4. In the last
e first row so that the
'espnding elementary
,! ~][! HJ
Let M(s) be unimodular. Then det M(s) is a constant. Hence M - 1(.~) =

[Adj M(s)]/det M(s) is c1early a polynomial matrix. Let M(s) and M -l(S) be
polynomial matrices. Then, det M(s) = a(s) and det M -l(S) = bes) are poly
nomials. Since M(s)M - I (s) = 1, we have det M(s)' det M- 1 (s) = a(s)b{s) = J
The oniy way for polynomials a(s) and bes) to meet a(s)b(s) = 1 is that a(s) and b(s)
are both nonzero constants. Hence, M(s) and M - l(S) are unimodular. Q.E. D.
In the proof, we have shown that the inverse of a unimodular matrix is also a
unimodular matrix. It can be shown that every unimodular matrix can be
written as a product of elementary matrices in (G-2S).
Unimodular matrices are c1early nonsingular in the field of rational func
tions. For every s in e, they are also rionsingular in the field of complex num
bers. In general, a nonsingular polynomial matrix is nonsingular in the field
of complex numbers for almost aIl S in iC. It becomes singular only at those s
which are the roots of its determinant.
.
.
The rank of a polynomial matrix in !R(s) will not challge if it is premultiplied
or postmultiplied by nonsingular polynomial matrices, in particular, uni
modular matrices.
592
G-4
Coprimeness of Polynomial Matrices
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)
where R(s) is an UPI

(G-28). The form of
q rows of the right-r
ract that the compc
Hence it has a rank 01.
it has at least q linear!:
row operations, it is (
Since U(s) is unimc
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
Dual to this definition, a square polynomial matrix Q(s) is called a greatest

. common left .divisor (gcld) of A(s) and B(s) if Q(s) is a common left divisor of A(s)
and B(s) [that is, there exist polynomial matrices A(s) and B(s) such that A(s) =
Q(s)A(s), 8(s) = Q(s)B(s)] and Q(s) is a right multiple of every common left
divisor Ql(S) of A(s) and B(s) [that is, there exists a polynomial matrix W(s)
such that Q(s) = Ql(S)W(S)].
Now we shall extend Theorem G-2 to the matrix case.
Hence, R(s) is a como
Let R(s) be any comr

and N(s) = (s)R 1 (s).
[U 11(S)0(S) + U ds)t\
we conclude that R(s)
Example 2
Find a gcrd of the po:

O(s) =
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)
where X(s) and Y(s) are p x p and p x q polynomial matrices, respectively.
<;
[
"
-1
From (G-30), we have
O
O
-( -.
-s-1
-1
-sy+~-+
Hence a gcrd of O(s) ,
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
which is not a unimd

(G-33)
Let W(s) be any P.

N(s). Then W(s)R(s)
premultiplication of t
593
COPRIMENESS Of POLYNOMIAL MATRICES
omials wil1 be extend

atrices does not com
: polynomial matrices
md A(s) a left multiple
a right multiple ofB(s).
he square polynomial
)(s) if there exist poly
(G-32)
where R(s) is an upper triangular polynomial matrix of the form shown in

(0-28). The form of R(s) is immaterial here. What is important is that the last
q rows of the right-hand-side matrix are zero rows. This fol1ows from the
fact that the composite matrix [D'(s) N'(s)]' is of dimension (q +p) x p.
Hence it has a rank of at most p (in the field of rational functions). Consequently,
it has at least q linearly dependent rows. Therefore, by a seq uence of elementary
row operations, it is always possible to achieve (0-33).
Since Ves) is unimodular, its inverse is a polynomial matrix. Let
[V
D(S)] =
11 (s)
[ N(s)
V 2(S)
and
I(S) is called a greatest

non left divisor of A(s)
:i B(s) such that A(s) =
)f every common left
.Iynomial matrix W(s)
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
Find a gcrd of the polynomial matrices

N(s) =
[-1
S2
+2s-1J
From (G-30), we have
J'l
+
9_ ----:=-_1_ - - - - ~ --- -- ! --- IL~~:~] = [~~~)] =
;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.
right divisor (gcrd) of

and D(s) and is a left
). lf a gcrd is a uni
:oprime.
I
UdS)]- ~[Vll(S) Vds)]~V(S)

Uds)
- v 2 (s) Vds)
VileS)
[ v 2 (s)
-s-1
S2
+s + 1
: - (S2
,
1)
'1
__
X- L
O
Hence a gcrd of D(s) and N(s) is

]', where the prime de
ore exists a unimodular
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:
has rank p (in the

2. There exist polyr
such that
is unimodular for any positive integer k; hence

R(S)=W(S)R(S)=[l+1
.
Sk
lIJ[OI
2J=[l+1
.'1+1
i
k
2s k+S+3J
2s +s+l
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.
This is cal1ed the
3. There exist no po
such that B(s)D(s)
-B(
and
Corollary G -7
Ir [0'(.'1) N'(s)]' is of ful1 column rank, in particular, if 0(.'1) is nonsingular,

then aH gcrds of 0(.'1) and N(s) are nonsingular and are related by unimodular
matrices.
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
root ofthedeterminant of 0(.'1), the (p +q)x p
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
Ir 0(.'1) and N(s) an

nomial matr.ix. T
yields (G-38) with :
Conversely, we
and N(s). Lt R(s
N(s) = N(s)R(s). T
COPRIMENESS OF POLYNOMIAL MATRICES
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
:ee that the degrees of

N(s). This phenome
This is cal1ed the Bezoui ideniiiy in References S34 and S125.

3. There exist no polynomial matrices B(s) and A(s) of order q x p and q x q
such that B(s)D(s) = A(s)N(s) or, equivalently,
lI(s). Can they always

lrmative if the matrix
- 8(s)D(s) + A(s)N(s) = [ - 8(s)
and
f D(s) is nonsingular,
elated by unimodular
;).
(G-37)
:nce, both W 1(s) and

) = W z(s)R(s) is also
Q.E.D.
is nonsingular. With
net. In this case, the
can be obtained from
ices, and let D(s) be

only if any one of the
)f O(s), the (p +q) x P
A(s)]
[~~:Q =0
(G-39)
deg det A(s) < deg det D(s).
Proof
1. For convenience we rewrite (G-33) in the following

U 11 (s)
[ U 11 (S)
s [R'(s) O'J'. Hence,

). Then by definitin,
hat
(G-38)
Uds)J[D(S)] = [R(S)]
Uds) N(s)
O
(G-40)
For every s in C, the unimodular matrix U(s) is .nonsingular in the field of

complex numbers. Hence, for every s in C, the rank of [D'(s) N'(s)]' is the
same as the rank of R(s). Ir D(s) and N(s) are right coprime, then R(s) is
unimodular and has rank p for every s in C; hence [D'(s) N'(s)J' also has
rank p for every s in C. Ir D(s) and N(s) are not right coprime, then det R(s)
is a polynomial of degree one or higher [Note that R(s) is nonsingular
following the nonsingularity of D(s)]. Therefore, there is at least one s in C
such that det R(s) =0 or, equivalently, the rank of R(s) is smaller than p.
Hence, we conc1ude that if D(s) and N(s) are not right coprime, then the rank
of [D'(s) N'(s)]' is smaller than p for sorne s in e This completes the proof
oi" ihe nrst statement. Since D(s) is nosigui( ay assmp~0li, ,fi ,il(
condition is met at every s in C except the roots of det D(s) =0. Hence we
check the rank only at the roots of det D(s) = O.
2. To show the second statement, we write the first p equations of (GAO) as
U 11 (s)D(s) + U ds)N(s) = R(s)
(G-41 )
Ir D(s)andN(s) are right coprime, R(s) is unimodular, and R -1(S) is a poly

nomial. matrix. The premultiplication of R - I(S) on both sides of (G-41)
yields (G~38) with X(s) = R -1(S)U ll (s) and Y(s) = R -1(S)U 12 (s).
Cnversely, we show that (G-38) implies the right copdmeness 'of D(s)
and N(s). Let R(s) be a gcrd of D(s) and N(s); that is, D(s) = D(s)R(s) and
N(s)=(s)R(s). The substitution of these into (G-38) yie\ds
[X(s)D(s) + Y(s)(s)] R(s) = 1
596
which implies
The substitution o
R -1(S) = X(s)D(s) + Y(s)N(s) = polynomial matrix

It follows rom Theorem G-6 that R(s) is unimodular. Hence, D(s) and N(s)
are right coprime.
3. We write the bottom q equations of (G-40) as
U 21 (S)D(s)
+ 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)
Since V(s) is unimc
This relation hold~

U- 1(s). We shall
Theorem G-S. Ir
R(s) with deg det
x det R(s), which ir
From (G-34) and (G-35), we have

(G-42)
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)
we can write the determinant of V(s) defined in (G-34) as

det V(s) = det V II(s)det [V ds) - V 21(S)V /(s)V ds)] "1=0
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
for all s in C. Taking the inverse of (G-43) yields

V 11 (S)
[ V 21 (S)
VdS)]-I[ 1
Vds)
-V 21 (S)V/(s)
0J-l
[V 11 (S)
= O
~dS)]-1
Based on this, we have

(G-45)
Theorem G-S'
l..}.
where t.=V 22 (S)-V 21 (S)VII(S)Vds) is nonsingular following (G-44).

The inverse of a triangular matrix is again triangular and can be readily
computed. After computing the inverse of the right-hand-side rnatrix of
(G-45), we [hen move the second inverse on the left-hand side of the equaity
in (G-45) to the right-hand side:
V 11 (S)
[ V (S)
21
VdS)]-1 = [Vi}(S) -V/(S)VdS)/}.-IJ[ 1

Vds)
o
/}.-1
-V 21 (S)Vl(s)
where X denotes elements which are not needed in the rollowing.

comparison of this equation with (G-34)yidds
Ui2(S) = /}.- 1 = [V 22(S) - V 21 (s)V /(s)V 12(S)] - 1
Let A(s) and 13(s) be q

singular. Then A(s) an
ing condition.s holds:
1. For every s in
matrix
e,
or
0IJ
has rank q in the fie
2. There exist polynon
that
The
3. There exists no poi:

such that
COPRIMENESS OF POLYNOMIAL MATRICES
597
The substitution of this equation to (G-44) yields

matrix
det Ves) = det V 11 (s)

det Uds)
Hence, D(s) and N(s)
(G-46)
Since Ves) is unimodular, we have deg det Ves) =0. Hence, (G-46) implies
deg det V 11 (s) = deg det Uds)
U 21 (S) and A(s) =

ing inequality:
= -
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)
Jsing the identity
(G-43)
as
5')V 12(S)] i=- O
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'--..
O] = q{[A(s) B(sn[V 11 (s) V 12(s)J}q
q P
V 21(S)
Vds) }p
Based on this, we have the fol1owing.

(G-45)
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)
has rank q in the field of complex numbers.

2. There exist polynomial matrices X(s) and Ves) of order q x q and px q such
ili~
.
A(s)X(s) +B(s)V(s) = 1
the fol1owing.
The
3. There exists no polyriomial matrices N(s) and D(s) of order q x parid p x p

such that
A(s)N(s) = B(s)D(s)
598
singuar and deg det

are right coprime.
or, equivalently,
- A(s)N(s) + B(s)D(s) = [A(s)
B(s)] [-
~~;~] =0
(G-SO)
and deg det D(s) < deg det A(s).
G-S
We give the fo\lowing coro\lary to conclude this section.
Consider two polyno

singular, then the mat
given a q x p rational
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
where N(s), D(s), A(s:

~ 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)
+ Uds)N(s) =0, implies
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
nij and d ij are r

um~"
l . . 1 of fe.
:uT\'s'J"I ah.
(;(S). These fractions .
not right coprime and
A rational functiol
3-5). In terms of th{
determined. For exa
the degree of the nun
to that of its denomil
situation is more com
Given a polynomia
power of s in a\l entrie
where
l-:<LI-.
11 "0
'-' l
bciM(~
briM(~
599
COLUMN AND ROW-REDUCED POLYNOMIAL MATRICES
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)
G-S Column- and Row-Reduced Polynomial Matrices

on.
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
s and let O(s) be nonor
C(s) = N(S)0-1(S)
(G-51 )
C(s) = A- 1 (s)B(s)
(G-52)
where N(s), D(s), A(s), and B(s) are, respectively, q x p, p x p, q x q, and q x p

polynomial matrices. For example, we have
o).
D(s) =deg det Uds).
iC. This implies that

q) polynomial matrix,
rem G-8', left coprime.
)fi.
Suppose Uzz(s) is
:or a(s), not identically
l~
d ll
nZ1
d Z1
n1Z
d 12
n22
-
d22
l3 =~el
icts with the fact that

. Hence we conc1ude
-UzzN =0, we have
the proofofTheorem
n (G-42) is unimodular
G-47), implies degdet
:g det Uds), we may
o. Since R(s) is non-
~e3 =[~ll
""1 [" ""1

d
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) .
b,.M(s) = the degree of the ith row 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.
[t is called row reduced iI
Theorem G-9
Ir (;(s) is a q x p proper (strictly proper) rationalmatrix and if (;(s) = N(s)O -1(S)

=A -1(s)B(s), then
bciN(s) ~bciD(s)
The rnatrix O(s) in (1

0< b el O(s) +bezO(s) = 2
colurnn reduced but not
for i = 1,2, ... , p, and
brjB(s) ~brjA(s)
for j = 1, 2, ... ,q.
is colurnn reduced but

rnatrix is always both co]
Let beiM(s) =k ei. Th,
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
where Hc(s) = diag{lei, i'

the column-degree coeiTic
colurnn of M(s) associatel
rernaining terms and its i
the M(s) in (G-56) can be
M(s)
j = 1, 2, ... , p
[n terrns of (G-57), we ha
The rest of the theorem can be sirnilarly proved.
Q.E.D.
det M(s) = (det 1\
We showed in Theorem G-9 that if G(s) = N(s)KY- '(s) is strictly proper..

then column degrees of N(s) are smaller than the corresponding column degrees
of D(s). lt is natural to ask whether the converse is also true. In general, the
answer is negative, as can be seen from
N(s) =
[1
2]
S2
D(s)= [ s+l
where bciN(s) < beiO(s), i = 1; 2. However, we have

N(S)O-l(S) =
nence, we conclucie chat

coefficient matrix M',e is I
Similar to (G-57), we
[-2S1 -1
whichis neither stritly proper norproper.

[n order to resolve this difficulty, we need a new concept.
(G-55)
where H,.(s) = diag{sk r i , i'

row. MIli_ wil1 be called tt
ficients of {he th row c
M1r(s) contains the remail
For example, the matrix
. M(s) =\
COLUMN AND ROW-REDUCED POLYNOMIAL MATRICES
mple, for
Definition G-4
A nonsingular p x p polynomial matrix M(s) is called column reduced if
601
(G-54)
deg det M(s) =
bciM(s)
i~1
It is called row reduced if

p
deg det M\s) =
b,.iM(S)
i= 1
and if (;(s) = N(s)D -1 (S)

The matrix D(s) in (G-55) is not column reduced because deg det D(s) =
neither is it row reduced. A matrix may be
column reduced but not row reduced or vice versa. For example, the matrix
o <bciD(s) +b c2 D(s) =2 +1 =3;
)(s)J
_[3s2 +2s 3
M(s) -
l(S)J
2s
s +s -
1J
(G-56)
is column redueed but not row reduced (verify!)_ A diagonal polynomial

matrix is always both column and row reduced.
Let bcM(s) = k ci - Then the polynomial matrix M(s) can be written as
t of N(s)_ Then we have
., 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
M(s)=[~ ~I~2 ~J+L~3 . ~J
Q.E_D.
In terms of (G-57), we have

det M(s) = (det MhcJSLkci +terms with degrees smaller than "L,k ci
-I(S) is strictly proper,

ponding column degrees
Iso true. In general, the
Menee, we concude that M(s) is coiumn-reauceli ij anc! omy ij its cOiumn-degree

coefficient matrix M hc is nonsingular.
Similar to (G-57), we can also write M(s) as
M(s) = H,_(s)M hr + M1r(s)
(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
POL YNOMIALS AND POLYNOMIAL MATRICES
In terms of (G-58), we have that M(s) is row reduced if and only if its row-degree
coefficient matrix M h , is nonsingular.
With the concept, we can now generalize Theorem G-9 to the following.
The column degrees a
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
and is singular. Hencl

exist (Xl_ (X1_ and (X3 SUC
for i = 1, 2, ... ,p.

Proof
The necessity part has been established in Theorem G-9.

sufficient parto Following (G-57), we write
We now show the
D(s) = DheHe(s) + Dle(s) = [Dile + D lc (s)H e- 1(s)]H e(s)
N(s) = NheHc(s)
+ Nds) =
We normalize the (Xi as

to be 1. In this exam
choose 0':1 = 1, and (Xl .
postmultiply the unim
[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
where D he is nonsingular by the column reducedness assumption ofD(s). Now if

6ciN(S) '::;6 ei D(s), N'Ie is a nonzero matrix and (;(s) is proper. Ir 6ciN(S) < 6ciD(S),
Nhc is a zero matrix and (;(s) is strictly proper.
Q.E.D.
The Hermite-form polynomial matrix shown in (G-28) is column reduced.
Since every polynomial matrix can b~ transformed nto the Hermite f0rm b)' ~'.
sequence of elementary row transformations (Theorem G-5), we conclude that
every nonsingular polynomial matrix can be transformed to be column reduced
by a sequence of elementary row operations. It turns out that the same can be
achieved by a sequence of columns operations. This will be illustrated by an
example.
where the degree of n

verified that M 1(s) is Cl
From theexample,
column degree can b,
nomial matrix can be r
the earlier statement a~
Theorem G -11
For every nonsingular

U(s) and V(s) such th
reduced.
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.
COLUMN AND ROW-REDUCED POLYNOM1AL MATRICES
Id onlJI if its row-degree
603
The column degrees are 1, 2, and O. The column-degree coefficient matrix is

1
}-9 to the fol1owing.
-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
~)]
9. We now show the
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
28) is column reduced.

the Hermite form by a
3-5), we conclude that
1to be column reduced
It that the same can be
ill be illustrated by an
-2s 2
~]
to M(s), we obtain
Hence';'e
nption ofD(s). Now if

er. IU)ci N(s) < bciD(s),
Q.E.D.
~] ~[~
3s
1
5s +1
-6s +1
-2s
~J~M'(S)
where the degree of the second column is reduced by one.

verified that MI (s) is column reduced.
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.
An algorithm is availablein Reference S137 to transform a matrix.into a

column- or row~reduced olie. In the fol1owing, we shall extend Theorem G-'1
to the matrix case.
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
N(s) = Q(s)D(s) + R(s)

R(s)D-I(s) is strictly proper
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
which can be replaced by, if D(s) is column reduced,
i = 1, 2, ... , p
Define
Proof
Consider the rational matrix (;(s) = N(s) D - I (s2. This rational matrix is not
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
Let D(s) = sl- A, and

exist unique polynomi
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
and is left as an exercis

their colurnn and row '
(G-60)
G-6
and R(s)D - I(S) is strictly proper. Then Equation (G-60) implies

[R(s) - R(s)]D-l(s) = Q(s) - Q(s).
Its right-hand side is a polynorfliaI rfiatri/~, ;vhefeas lts left-hanci siae is a stricltj
proper rational matrix. This is possible only if Q(5) =Q(s) and R(s) = R(s).
The column degree inequality follows directly from Theorem G-lO.
Q.E.D.
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'
Let A(s) and B(s) be q x q and q x p polynomial matrices. IfA(s) is nonsingular,

there exist unique q x p polynomial matrices Q(s) and R(s) such that
B(s) = A(s)Q(s) + R(s)
and
A -1(s)R(s) is striCtly proper
.Consider a q x p prop
C(s) = N(s)D - I (s). Th
.. p x p nonsingular poly
N(s)
which can be replaced by, if A(s) is row reduced,
i = 1, 2, ... ,q
Theorem G -1 3
Ir the rraction N(s)D -
- - - ---_._-----------
COPRIME FACTORIZATIONS OF PROPER RATlONAL MATRICES
;es and let O(s) be non

matrices Q(s) and R(s)
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 )
N,.(A) = N"A" + N"_IA"-1 +

N(A)=A"N" +A"-IN"_1 +
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.
N(s) = Q,.(s)(sl- A) + N..(A)

N(s) = (sI - A)Q(s) + N(A)
and
III
This corollary can be readily verified by using
)D-1(s)
matrices [R(s) = N(s)

J.d R(s) such that
R(s)
+ NoI
+IN o
Q,.(s)= N"S"-l +(N"A +N"_I)S"-2 + ... +(N"A"-l +N"_IA"-2 + ... +N 1)

and is left as an exercise. Note that N..(A) and N(A) are constant matrices and
their column and row degrees are all equal to zero.
(G-60)
G-6
O) implies
Coprime Fractiolls of Proper Rational Matrices
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
Consider a q xp proper rational matrix (;(s) with the right-coprirne fraction

(;(s) = N(s)D~ 1(s). Then for any other fraction G(s) = N(s)D - I(S) there exists a
p x p nonsingu.lar polynomial matrix T(s) such that
-.N(s) = N(s)T(s)
. and
D(s) = D(s)T(s)
If the fraction N(s)D -1(S) is also rightcoprirne, then T(s) is unimodular.
606
Proof 6
compute R - l(S) and
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)
and N(s)D - l (s) is irre<

This procedure, how(
nomial matrix and is r
we also compute Ves)
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
D- l(S) = R - 1 (s)W(s)R(s)D- l(S)
(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
Since R(s) is unimodular following the irreducibility assumption of N(s)D - l(S),

the matrix
T(s) = R - 1 (s)W(s)R(s)
(G-63)
is a polynomial matrix The substitution of(G-63) into (G-62) yields 0- l(S) =

T(s)D - I(S) or D(s) = D(s)T(s). The substitution of 0- 1 (s) = T(s)D - l(S) into
(s)D - l(S) = N(s)D- l(S) yields immediately (s) = N(s)T(s). The nonsingular
i,ty of T(s) folIo-viS fror the l10nsingularities of T{(sj, -;:;V(sj, ana 1R(s).
If (s) and D(s) are right coprime, then R(s) is unimodular. Consequently,
the T(s) in (G-63) is also unimodular: This completes the proof ofthe theorem.
Q.E.D.
From this theorem, we see that all irreducible fractions of a proper rational
matrix are related by unimodular matrices. Hence, the irreducible fraction is
unique in the sense that all irreducible fractions can be generated from a single
irreducible fraction.
Conside( a fraction (;(s) = N(s)D- l(S). lf it is not irre'ducible, we may
use the proce'dure in (G-33) to compute the gcrd R(s) of N(s) andD(s). We then
If we consider polYI
!R(s), then Equation (G
sequently, aH 1 x (p +
nomials) satisfying
is a linear space over !R(.

Following Definition 2
to (p +- q) - rank [D'(:
dimensional null space
qualifies as a basis (Th
only the pblynomial so
pan of \l. A set of q v
in (\11" !RIs]),7 if every
is in fa a. free module OVt
chaplcr and Refercncc S34.
71l
6For a differenl proof, see Problems G-12 and G-l3.
COPRIME FACTORIZATlONS OF PROPER RATIONAL MATRICES
607
compute R -I(S) and

.adjoint and det stands
)O-I(s)and O(S)O-I(S)
N(s) = N(s)R - I(S)
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')
D(s) = O(s)R -I(S)
(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)
Jmption of N(s)O- I(S),
(G-63)
(G-62) yields O-lis) =

-I(S) =T(S)O-l(S) into
f(s). The nonsingular
;), and R(s).
)dular. Consequently,
e proof of the theorem.
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)
If we consider polynomials as elements of the field of real rational functions

lR(s), thcn Equation (G-64) is a homogencous linear algcbraic cC]ualion. Con
sequently, all 1 x (p + q) vectors x(s) with elements in lR(s) (including poly
nomials) satisfying
i[})(s11
;~(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
of the q vectors by using only polynomials as coefficients. We note that every

basis of (W P' !R [s]) is a polynornial basis of (W, !R(s)); the converse however is not
true in general. 8 It turns out that A(s) and B(s) are left coprime if and only if the
set of the q rows of [-8(5) A(s)] is a basis of (W p , !R[s]) (Problern G-16).
Arnong all bases in (W P' !R[s]), sorne have the additional property that the row
degrees of the basis vectors are srnallest possible. This type of basis is called a
minimafpofynomiaf basis. It will be shown that the set of q rows of [ - B(s) A(s)]
is a rninimal polynornial basis if and only if A(s) and B(s) are left coprime and
A(s) is row reduced. In the fol1owing, we discuss a method to find a minimal
polynornial basis.
Instead of solving (G-64) directly, we shall translate it into a hornogeneous
linear algebraic equation with real nurnbers as entries. Let
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
where Di, Ni, A i , and B i are p x p, q x p, q x q, q x p constant matrices. By

substituting (G-65) into (G-64) and equating the coefficient of Si to zero yield,
similar to (G-22) and (G-24), we obtain
Do D 1 .. Dd
No NI ... N d
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
the first q linearly de:

we define
1st block {
2nd block {
(k
+ 1)th block {
The rows formed frOl

rnatrix Sk has k +1 b
Now it is assumed th::
its linearly dependent
Lemma G-1
If (;(s) = N(s)D- 1 (s) i~
pendent of their previ
This lemma will

N(s)D-I(s) be a rigl
N(s)D-I(s) is not pro
ample, a linearly depe
even though we have

Ir, instead ofSk, we
(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.
then the statement is al

consider
If we use Sk' then a lin<

Because af Lernrn
only in the N rowS. L
(i + 1)th block raw. B
609
COPRIME FACTORIZATlONS OF PROPER RATlONAL MATRICES
ltSo We note that every

converse however is not
oprime if and only if the
!R[s]) (Problem G-16).
il property that the row
type of basis is called a
qrowsof[ -B(s) A(s)]
:(s) are left coprime and
:thod to find a minimal
: it into a homogeneous
Let
(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)
ralized resultant of D(s)

an be used to solve Bi
A.(s) and B(5) are given,
)0 In Equation (G-66),
'm]. In order to have q
:>eaking, q linearly de
:sirable to have ni, the
. use, roughly speaking,
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
.0- --O; -. ~ ~ - ~ ~ ~ -D~ - -0- -.-. -. - O

2nd block { ~ __ ~_o_ ~ ~ ~ _f"!'!.-_I__f"!o__~ __. ~ ~ _~_
(k
+ 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
.. If we use Sb then ~ linearly dependent row 'w\1i appear in a D row.

.. Because of Lemma G-l, the linearly dependent rows of Sk will appear
orily in the N rows. Let r be the number .of linear1y dependent N rows in the
(i + 1)th block r6w. Because of the structure of Sk' we have
(G-G9)
610
Let
be the least integer such that r v = q or, equivalently,
matrices
(G-70)
This implies that as k increases, the total number

rows in Sk will increase monotonically. However,
increase, no matter how many more block rows
linearly independent N rows will remain the same.
n=(q-ro)+(q-rl)+'"
of linearly independent N
once the number ceases to
are added, the number of
Define
+(cl-r v -)
are left eoprime, ane

formo
Proof
In order not to be o'
(G-71 )
lt is the total number of linearly independent N rows in Sk for k;::o: v - 1. lt

turns out that n is the degree of (;(s) or the dimension of any irreducible realiza
tion of (;(s).
The number oflinearly dependent N rows in SI' is equal to ro + rl + ... + r",
which is c1early larger than q. However, there are only q primary linearly
dependent rows in S". A dependent row is called primary if the corresponding
row in the previous block is independent of its previous rows. For example, all
the ro dependent rows in the first block of SI' are primary linearly dependent
rows. However, the rl dependent rows in the second block of SI' are not all
primary because ro of the corresponding rows in the first block have already
appeared as linearly dependent rows. Hence, in the second block of SI" there
are only rl -ro primary linearly dependent rows. Similarly, there are rz -rl
primary linearly dependent rows in the third block of Sv. Proceeding in this
manner, we conclude that the number of primary linearly dependent rows in
SI' is equal to
rO+(rl-rO)+(rz-r}+'"
+(rv-r,._}=r,.=q
Consider Equation (G-66) with m replaced by v:

(G-72)
where x denotes non;

independent of their r
From (G-74), wc hay
study the structure of
These A and lB are to 1]<:': obtained by using the row-searching algo,ithm. ,~

other words, they are the q rows of l(in iK", computed as in (A-7), correspon
ding to the q primary dependent rows of SI"
Theorem G-14
Consider a q x p proper rational matrix (;(s) factored as (;(s) = N(s)D- 1 (s).

We form Sk and search its linearly dependent rows by using the row-searching
algarithm. 9 Let [ - B o Aa ,.. -B,. Av] be the q rows of I< in KS v=Sv
corresponding to the q primary de.pendent rows af SI" Then the palynomial
The primary depel

shown. Correspondi
(G-75) assume the for
SO
'-[--(~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
a~1 a~z a~3:
611
COPRIME FACTORIZATIONS OF PROPER RATIONAL MATRICES
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)
as (;(s) = N(s)D- 1 (s).

;ing the row-searching
rows of K in KS v =Sv
Then the polynomial
:ed by a numerically slable

ioyeci, lhe resu il is generally
The primary dependent rows of S3 in (G-74) are indicated by the arrows as

shown. Corresponding to these primary dependent rows, the B i and Ai in
(G-7.5) ass~me the form
SO
.".
~80
.
I
.\~~
.A
aO. .a? 2 a?3:

aL al
. 12
1\
O'
-8
a21 ti~2 a~~:
1 aL ai2
-,
aL a12
a~l a
32 aO33 ,
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:
and the v defined in (
We shaIl cal! v the re
v3::; .... then (G- 76'

The it
1. AIl elements (including - BJ on the right-hand side of:D are zeros.
ce ,
2. AIl elements, except element

of the columns corresponding to primary
linearly dependent rows of S" are zeros.
3. AIl elements ofthe column corresponding to nonprimary linearly dependent
rows of S,. are zeros.
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,
where deg det stands

we are ready to show
then there exists a po
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:
Because of deg det Ql
a32 +a32s+a32s +s '

1-
deg det A(s) =
~ number of linearly independent ith N row in S"
(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
This implies that the

than n. This is not p(
pletes the proof of thi
Proof of Lemma G -'
lf the linearly depem}(

we can see that the 1'0\
Shl.ce A\{s) is rov! red 1!'
aependen t row ol' ,. si
in the column of -B i
be larger than the con
tion that (;(s) is propt
in the N rows.
Example 1
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
COPRIME FACTORiZATlONS OF PROPER RATIONAL MATRICES
lree :1) are determined

e, the last row of the
the last col umn of the
_s obtained by the row
erties:
. and the v defined in (G-70) is

v = max {Vi, i = 1, 2, ... , q}
We shall call v the row index of C(s). Ir we rename
. " , then (G-76) and (G-77) imply that
The ith row degree of A(s) = Vi
ary linearly dependent
as '\\ such that VI
'::;"2'::;
V3 '::;
r:~e are zeros.
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)
Since A(s) is row reduced, we have

are on the right-hand
lOse columns with the
he sixth colurnn of the
are on the right-hand
l block. We note that
the rightmost column
g 2 alone. Because of
deg det A(s)
=
i
Vi
= n = total nurnber of linear independent N rows in S"
(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)
Because of deg det Q(s) > O, we have

>ivot elements. Their
Ve note that every row
Ise of Property 1, the
every right-hand-side
le degree of every left
es 2 and 3, the degree
~r element in the sarne
said to be in the row
t a polynomial matrix
1.
tt in each block row of
appears v + 1 times in
row inS.
(G-78)
. For the example in

n defined in (G-71) is
deg det A(s) < deg det A(s) = n

This implies that the nurnber of linearly independent N rows in Sv is smaller
than n. This is not possible. Hence A(s) and B(s) are left coprime. This com
pletes the proof of this theorem.
Q.E. D.
Proof of Lemma G-1
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)
\" ... A,] oblaincd by

Ice the correspondingpoly
al echclon for 111.
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.
- - - - - - - - ---= -- -6 --0- --6- -~:

-1
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
the polynomial eehel
Consider a proper
right eoprime if and
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
2, ... ,q} rather than

lish the main resulto
O ::~ I:
O
1
O
O
O - - - - -(j) - - -( - -~ i --- -0
O
8(5) = _
Corollary G-14
--------------~1---6--0-----(6-0
::-=- -(~
Note that the solutic
Combining Theo'
eorollary reduces to I
-------
O -1
**
[~.5 =~.5:
and
O
O
O
:~D O
O ~~~D
-1
1
O
Note that the pivots

/"2 = 2 = q, we have l'
the arrows shown.
the formula in (A-l1
Lemma G-2
,X:
:I~
O -\
The m x m polynom
the fie\d of rationa\ fl
- - - :l~ ---0- ---O . --0- -. -O
TO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
00
O
(i~
0- ---O . -- -- -O ---0- . - 0---
:I~
O
O
v,~;
O
O
[
.~
is offull row rank (i
615
COPRIME FACTORIZATIONS OF PROPER RATIONAL MATRICES
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
_[-0.5 -IJ [0.5 lJ [0 0J

- 1
1
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-
is of full row rank (in the field of complex numbers) for k = 0,
1,2, ....
(G-85)
616
yields the number of.

consider (G-74). Sin
we have one N row wit
with row index 3. In (
the row indices are Ul
indices are also intrim
D(s) used in the comp
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
Let G(s) = A -1(s)B(s) be a left-coprime fraction and A(s) be row reduced.

Then the row degrees of A(s) are intrinsic properties of G(s) and are independent
ofthe N(s) and D(s) used in Theorem G-14 in the computation.
Proof
Because of (G-81), it is sufficient to show that the row indices defined in (G-78)
are independent of the N(s) and D(s) used in the computation. We recall that
every right fraction N(s)D -1(S) of G(s) can be obtained from a single right
coprime fraction (s)D -1(S) by the relationship
D(S)J = [?(S)J T(s)
[ N(s)
N(s)
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)
where T(s) is a p x p nonsingular polynomial matrix (Theorem G-13). If we

write 1'(5) as T(s) = T o +ir s + ... +- Tjsj, then (Ggfi) i'Tp!ip,~. t::.c ;. . (\ ,
c.
Sk =SkV d+k
(G-87)
where Sk and Sk are defined similarly as in (G-67), d is the degree of D(s) as in

(G-65) and Vd +k is defined as in (G-85). Since Vd +k has a full row rank for every
k, (G-87) implies
rank Sk = rank Sk
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,. . .
Thus the row degrees o

[ - B(s) ~~(s)J is o. rnin
i.:OpfiiIiC di h\S) S roy
additional nice properl
quence of the row sea!
obtained by using the r
basis and in a canonice
We discss briefly 1
(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
row index i or smaller. Define
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
COPRIME FACTORIZATIONS OF PROPER RATIONAl MATRICES
617
yields the number of N rows with row index exactly equal to i. For example,
consider (G-74). Since rO-I"-I=O, rl-ro=l, 1"2-rl=0, and r 3 -r2=2,
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.
A(s) be row reduced.

:.1') and are independent
Jtation.
dices definedin (G-78)

tation. We recall that
d from a single right
( 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
L b.. A(s) > degdet A(s) =degdet A(s) = I
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
minima! polynomial basis is in an echelon formo These results are dual to

Theorems G-14 and G-15 and are stated as theorems.
Theorem G-14'
Consider a q x p proper rational matrix G(s) factored as G(5) = A -1(s)8(s) with

A(s) = I~=o ASi and B(s) = I~=o BSi. We form
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
: Aa Bo k + 1 block columns (each

block has q +p columns)
:, Al B 1
: Az B z
Transform the mal
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
into the Hermite row for
,:
,.
G-2 The polynomials [

if the square matrix S of
matrix S" _, of arder Li.
statements.
L;=o
L;=o
G-4
Find a gcrd of the I
D(
G-5
Are the following p
a. D,(s)= [
by using the column-searching algorithm corresponding to the p primary linearly

dependent columns ofT Ji' are right coprime. Furthermore, D(5) is in the col umn
polynomial echelon form, that is, the degree of every pivot element of D(s) is
larger than the degree of every other element in the same row, larger than the
degree of every lov/er el~D1en.t in ~Jv~ same columt~.
degree of every upper element in the same column. We call jJ. the column index
of G(s).
Theorem G -, 5'
+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.
in Problem O-S? Are th

notes the transpose.)
G-7. ls the matrix
Let G(s) = N(s)D-1(s) be a right-coprime fraetion and D(s) be column reduced.

Then tl).e eolumn degreesof D(s) are intrinsic properties of G(s) and the set of
the column degrees is unique. ..
I
eolumn reduced? [f not,
Problems
V(s)M(s) are column redu
G' Apply the row searching algorithm to Skin(G-20) to reduce the following rationl
functions to irreducible ones:
G-8
ls the M(s) in Probl
-------..- - - - - _ . _ -
PROBLEMS
N(s)
se results are dual to
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
(;(s) = A-1(s)B(s) with

The polynomials D(s) or degree n and N(s) or degree In < n are coprime ir and only
ir the square matrix S of order n + m in (G-12) is nonsingular or if and only ir the square
matrix S" - [ of order 2n in (G-20) is nonsngular. Show the equivalence of these t'NO
statements.
G-2
block columns (each

ck has q + P columns)
G-3
Transform the matrix
(s + 1)2
-s-1
S2
+s+
s-1
_S2
+s+1
1]
into the Hermite row formo
to right. Let r be the

,)lock, and let J1 be the
and N(s) = L.i;O N s,i
"1'
s+1
O
G-4
Find a gcrd of the fol1owing two matrices by using (G-33):
52 +2s
O(s)= [ 15 2 -5
G-5
the p primary linearly

~, D(s) is in the column
vot element of D(s) is
e row, larger than the
r han or eq ua] lO 1he
:all fJ. the column ne/ex
N(s)=[s
1]
Are the following pairs of polynomial matrices right coprime?
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
G-7. Is the matrix
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
are soll1tions of the Dio

matrix T(s). the matrice~
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
[~
are solutions of the Dior

G-15
Find a right fraction from the left fraetion
G(S)=[S2~1
and N(s), show that for ;

tine equation ifand onl
matrix F(s) such that F[
.1'2 + s J[S4 +.1'2

2.1'
.1'2 + 1
.1'2 + 1
=[.1'
G-10
----------;
s~lIs~l
~:~J
Consider
\vith ()ciR(S) < (\;D(s) = 11,

1f D(s) is column reducec
s2~lTl)
N(s)W'(s)~r
=(
Let (;(.1) = A -'(s)B(s) = N(s)O- '(s) be two eoprime fractions.

exists a unimodular matrix of the form
G-11
Show that there
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
V 12(.1) V I I (S)J[ - N(S)J = [IJ

[ A(s)
B(s)
0(.1')
O
(Hint: Ir O and N are right coprime, we have
[~:: ~::I -~J=[~J
where Nis) is a polynon

of Np(s). Define H(s) = (
Np(s)
= N(s
Q(s)~Q(s
D(s)~DI;;:
where the ith row of Qk i

compllted recursively as
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.)
whcre D[~O irl>/l. f

R(s) = N(s) - Q(s)O(s). S
Silow that ji N(.I')lJ}-l(.I') = I\J{s)D (sj are [wo right-coprime "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.)
G"hi Consicier G(s) = .

rational fUl1ction solutior
:':;;-")2
G-~3
Prove Theorem G-13 by using Problem G-12.
G-14
Consider the Diophantine polynomial equation

0c(s)O(s)
+ Nc(s)N(s) = F(s)
where O,., O, N" N, and F are, respectively, p x p, p x p, p x q, q x p, and p x" p polynomial

matrices. Let
"
"
U 11(.1) Udsf\[O(s)J~[R(S)J
[ U2\(s) udst N(s)
O
where the leftmost matrix is unimodular and R(s) is a gcrdof 0(.1) and N(s). Given O(s)
and let Wp denote the poi

nomial basis of W if ever
basis by using only pe
of [ - B(s) A(s)] is a poI
[Hint: (= Let [ - B(s)
polynomialmatrix T(s) SI
Show that T(s) is unimodu

be any vector in Wp Sho'
PROBLEMS
67), left coprime fractions
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)
Nc(s) = N~(s) +T(s)ll}}(s)
and
s right coprime?
are solu tions of the Diophantine equation (see Reference S 139).

G-15
Consider
N(s)=Q(s)O(s) +R(s)
-lJ
-2
with 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)
+ N,.(s) = Q(S)(OhcH(S) + D1c(s))H - I(S) + R(s)W I(S)

=Q(s)D"c(I +Dl~IOlcls)Wl(S)) +R(s)H-1(s)
:tions.
Show that there
where N p(s) is a polynomial matrix and N,(s) is strictly proper. Let Vi be the row degrees
of Np(s). Define H(s) = diag {s", S''', ... ,s'p} and /l = max {Jl i }, v = max (v,}. We write
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
k=I,2, ... "
1=0
he dual ofTheorem G-13,

0 22 = A.)
)rime fractions, then the

n XN+YD=I imp1ies
be a polynomial matrix.)
( 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).
Lel W (eIlOl\;; he le, luir ~fl<i;" \.)\ ,d,
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
[ - B(s) A(s)] with A(s) nonsingular and
- - [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
Ir wc write g(s) = N(s)

common factor), then
'namial matrix T(s) such

[ - 8(s) A(s)].]
adular matrix such that
r and anly ir Dei U(s) s
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
g(s) = 052 +3s +2

although it is a root of S2 + 3s + 2.
Consider a proper rational function g(s) with the following irreducible realiza
tion
X '7Ax +bu
y =cx +eu
If we write g(s) = N(s)/ D(s) and N(s) an.O D(s) are coprime (have no nontrivial
" commoti facto'i), then we have
D(s) =k det (s 1 - A)
623
624
POLES AND ZEROS
with sorne constant k. If D(s) is monic (the coefficient associated with the highest
power of s is equal to 1), then k = 1. We discuss in the following the implica
tions of poles and zeros.
not a pole of 9(s), th,

output is to apply e;'
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
Consider a single-variable system with proper transfer function 9(S) and an

irreducible realization {A, b, e, e}. Then a number 1, is a pole of 9(S) if and only
if there exists an initial state X o such that the zero-input response at the output
of the system is equal to
y(t) = re A'
for all
Theorem H-2
To prove this the
t ~O
(sI -A)-l(S
for sorne nonzero constant r.
for any 1, that is not ;

the partial fraction e;
Proof
The zero-input response of the system is given by
(s
y(s) =e(sI -A(I X (O)
Ir ), is a pole of 9(S), then it is an eigenvalue of A. Let v be an eigenvector

of A associated with A; that is, Av = AV. Then it can be verified (Problem 2-32)
that v is an eigenvector of (si - A( 1 associated with eigenvalue (s - ,,1.)-1.
Hence we have
Y(s) = c(sl - A)-I V=cv(s - ,,1.)- 1
ar
y(t)=eve AC
[Al -AJ
The response of the

as derived in (4-25),
y(s) = c(:
=c(.
The substitution of (f
y(s) =c(sl- AY
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
The identity in (H-l)

(A-I,I) and (sI-A)
~A
or
det
or
(s - A)c [Adj (sI - A)Jx(O) = r det (si - A)
A) e [Adj (sI - A)]x(O) = s
which implies det(Ai - A) = O. HenceA is an eigenvalue of A and, consequently,

a pole o[g(s). This completes the prof orthe theorem.
Q.E.D.
This theorem states that if 1, is a poIe of g(s), the mode eAC can be generated
at the output by an initial state without the application of any input. If A is
AY
y(s) = g(A)(s - A:
or
y(t) =
g(A )e A'
This completes the pr

In this theorern,
Otherwise, lhe theore
ing that the impedan
a positive real functi(
give a physicannterpl
g(s), that is,gv,) = O, t .
identically zero even :
blocked by the systerr
625
POLES AND ZEROS
,ciated with the highest

following the implica
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
function 9(S) and an

pole of 9(S) if and only
response at the output
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:
(sl-Ar1(s-A)-1 =(AI-A)-1(S-A)-1 +(sI-A)-1(A-,U)-1
(H-1)
for any A that is not an eigenvalue of A. Note the similarity of this identity to
the partial fraction expansion
1
1
1
1
1
---- =--'--+--'-
(s-a)(s-A) (A.-a) (s-),) (s-a) (a-J,)
et v be an eigenvector
'erified (Problem 2-32)
eigenvalue (s - A) - 1.
The identity in (H-l) can be readily verified by post- and premultiplication of

(A - ),1) and (si - A). (Problem H-1.)
Proot ot Theorem H-2
The response of the system due to the initial state x(O) and the input u(t) is,
as derived in (4-25),
is different fram zero.

e the matrix [sI' -A' :
ery nonzero vector, in
Y(s) = c(sI - A)-I X (O) +c(sl - A) - 1bu(s) +eu(s)

=c(sI - A)-1 X(0) +c(sI - A)-lb(s - A)-1 + e(s - ),)-1
( H-2)
The substitution of (H-l) into (H-2) yields

y(s)=c(sI - A)-I X (O) +c(AI-Ar lb(s _A)-l
+c(sl- A)-I(A - JcI)-lb + e(s-Ar 1

=c(sl- A)-I[X(O) + (A -AI)-lb] + [c(AI- A)-lb+e](s- A)-1
t would have violated
cessity of the theorem.

then A is a pole of 9(S).
Y(s)
or
= g(A)(S -
y(t) = g(A)eJ.'
A)-1
for t ~ O
This completes the proof.
-A)
,f A and, consequently,
Q.E.D.
e e;" can be generated
of any input. Ir A is
Q.E.D.
In this theorem, the assumption that J, is not a pole of g(s) is essential.

Otherwise, the theorem.does not hold. This theorem is very useful in establish
ing that the impedance of a linear time-invariant, lumped, passive network is
a positive realfuncti'on (see Reference 31). This theorem can also be used to
give a physical interpretation ofthe zeros of a -transfer function. If), is a zero of
. 9(s), that is, g(A) = 0, then' for a certain initial state, the output of the system is
identicallyzero even if the input e;" is applied. In otherwords, the input e;" is
blocked by the system.
.
626
POLES AND ZEROS
Carollary H-Z (Transmission-blocking property)
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)
Ir we multiply the 1('

become
where M(s) and Pes) 2
where DI> N, N,., and O,. are, respectively, q x q, q x p, q x p, and p x p poly

nomial matrices. Furthermore, N/(s) and O(s) are left coprime; N..(s) and O..(s)
are right coprime. Let
X=Ax +Bu
then pG(s) = p, wher

the field of rational f
p x 1 rational vector
y=Cx +Eu
for al1 possible inpu

smal1er than q. Simi
pes) to the system, th,
be an irreducible realization of (;(s). Then we have

det(sl - A) = k 1 det 01(S) = k 2 det O,(s)
where k 1 and k 2 are constants. In view of this relationship, we may give the
following definition.
Definition H-Z
A number l, real or complex, is said to be a pote of a proper rational matrix

(;(s) if and only if it is a root of det O(s) = O, where O(s) is the denominator
matrix of any right- or left-coprime fraction of C(s).
111
Hence the number of

rank, there are no red
Consider C(s) wit
Ir (;(s) is of full rank

This implies that for
a rank eq ual to min (
Definition H-3
Similar to Theorem H-l, we have the following theorem to characterize the

poles of (;(s).
Theorem H-3
Consider a multivariable system with proper transfer matrix (;(s) and an ir

n~;
1;7"t'''-[:
j}, 18\'"
~1 ~.".'.,"".,-.,.;,.,.-,.__ .'." ...1..'.-.,).h<e.
-ec1ucible
-.....
_
_ ........ _A_"-'C
,o.U ... J
t--,
.,
,-,,_ .Z '.".,_ :-'.. ..~,10-..c f':.! ..\ .
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).
fuI! rank (in Ihe field

is said lo be a transr.
in <C, the field of com
""\'J/
y(t) = reAl
for some nonzero vector r.
<;;;onsider a q x p pr
Example 1
Consider the left eopr
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
This N(s) has rank 2 I
'If G(5) does not llave a fuI

;. in C'is atransmission zen
627
POLES AND ZEROS
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)
If we multiply the least common denominator of M(s) or P(s) to (H-3), they

become
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
prime; N..(s)and D,.(s)
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)
;hip, we may give the
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)
roper' rational matrix

s) is the denominator
~m
to characterize the
latrix G(.~) ~nd an ir

E~ pole af {~\s) ir and
input response at 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
'heorem H-l and wll

l poleofsomeelement
a pole of (;(s). This
:ommon denominator
).
:ivariable case. First,
a full rank in the field
en pG(s)=q; ir q > p,
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
POLES AND ZEROS
Example 2
Case 1: q > P
Consider the Ieft coprime factorization
In this case, we have p

in ce for alI i. in ce ex,
zera, then pN(i,) <;
vector k such that
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 ,
Case 11: q < P
In this case, we have p

we have pN(,1,) sq <,
N(X); = O and
y((
Since this equation he

transmission zero of (
sider this case separat
If q < p, then A is
nonzera 1 x q constan
If, in addition, ), is ne
is also a nonzero vectl
Theorem H-4
Consider a multivariable system with q x p praper rational matrix (;(s) and an

irreducible realization [}~, R, C, lE}. ifthe input au(t) is f th [OiT ~K(/'\ v:j~i.{t;'L
is real or complex and is not a pole of (;(s) and k is an arbitrary p x 1 constant
vector, then theoutputdue to this inputand the initial state x(O) = -(A -,1,1)-1 Bk
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)
where N(s) is a q x p polynomial matrix offu\l rank. In thdollowing, we discuss

separate1y the case q ;::o:p and the case q < p. We use p to denote the rank.
"S'r
which is, because of (1

fy(t) = Ofor t ;::0:0, exist
be used to characteriz.
Corollary H-4 (Trans
Consider a multivaria
is assumed that G(s) J
andthe irreducible rea
'zera of(;(s), then ther
output of the systern
POLES AND ZEROS
629
Case 1: q"2:.p
In thiscase, we have pN(s) = p in ~(s) by assumption. Hence we have pN(}.) = P

in iC for all 1, in iC except the transmission zeros of C(s). Ir l. is a transmission
zero, then pN(A) <p. Consequently, there exists a p x 1 nonzero constant
vector k such that
NLC).)k=O
and consequently,
2 are two transmission
s).
for all t"2:.0
ned at its transmission

its transmission zero.
nomial matrices. Ir X
ifand only if pN.(A) <
~ the Smith-McMillan
uch that
( H-5)
ing N(s) or N,(s). We

os of C(s) are the roots
nant.
;mission zeros of G(s)
.case. A trans'mission
>ugh elements of C(s)
'he transmission zeros
](5), In spite of these
;ion zeros of G(s) are
e in the following.
This property is similar to the transmission-blocking property of Corollary

H-2. Note that if ), is neither a transmission zero nor a pole of (;(s), then
pN(..1.) = p and N(X)k O, for all constant k, which implies
y(t) = D- l(X)N(),)k/'/
+O
Case 11: q < p
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
lal matrix (;(s) and an

the formke At , where;i.
bitrary p x 1 constant
x(O) = -(A-XI)-lBk
( H-8)
f~l{t) == fG{.J..)~~,(!A(
hN(X)ke
-::-:
( H-10)
= hD{..1.)
i~H}J;{ ~ '?~l\- f J.)h-\!/ )\i!{ '1"l:
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.
rheorern H-2 and will
Coroltaiy H _4 (Transmission-blocking property)
ds
Consider a multivariable system with q x p proper rational matrix C(s). It

is assumedthat G(s) has a full rank, the coprime fraction G(s) = D-l(s)N(s)
and the irreducible realization{A, B, C, E}. (1) If q "2:.p, and if X is a transmission
zero of G(s), then there exists a nonzero p x l' constant vector k such that the
output of the system due to the initial state x(O) = - (A ~ AI)-l Bk and the
( H-7)
e following, we discuss
denote the rank.
630
POLES AND ZEROS
input u(t) = ke Jt is identically equal to zero. (2) Ir q < p and if }, is a transmission

zero but not a pole of G(s), then for the input u(t) = ke.l.t, where k is an arbitrary
p x 1 constant vector, there exists a nonzero 1 x q vector f such that the output
y(t) due to u(t) = ke.l. t and the initial state x(O) = - (A - i,I)- 1 Bk has the property
fy(t) = fG(}.)ke.l. t = O
for all t? O
This t heorem fol

In establishing (H-l
Hence the proof is al
this restriction, we ~
q ?p, and if A is a tr
nonzero k such that t:
x(O) =
-1
are identically zero.
Sl-A
-c
( H-12)
Ir A, B, C, and E are, respectively, n x n, n x p, q x n, and q x p matrices, then

Ses) is a (n+q) x(n+p) matrix. Though elements ofS(s) are polynomials, we
shall consider them as elements of the field of rational functions. Because of
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)~ [
zero of C(s) if and
BJ [Sl-A
E O
=[
B
J
C(sl-A)-lB+E
SI-A B
[SI-A
O
C(s) =
O
The combination of (
Since [ - x'(O) k'J' is

Now consider the
zero, for anyk there eJ
for t ?O, or fC(A)k = (
holds for every k. He
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 )
Now if s is replaced by an element in e, the field of complex numbers, say s = A,

then SeA) is a matrix with elements in e and its rank must be computed in e
which can be combin
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
POLES AND ZEROS
d if A is a transmission
{here k is an arbitrary
such that the output
- 1Bk has the property
631
zero of G(s) if and only if
pS(I,)=p
AI-A
[
: ] < n +min(p, q)
-C
(H-15 )
lO). We see that this

Beca use of this trans
Jn H -3 are ca\!ed the
I(S) and an irreducible
Jatrix
( H-12)
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
Since [-x'(O) k'J' is a nonzero vector, we have pS{l) < n +p.

Now consider the case q < p. Corollary H-4 states that if Ais a transmission
zero, for any k there exists a nonzero 1 x q vector f such that fy{t) = fG(A~eA' = O.
for t ~ O, or fG(A)k = O. From (H-ll), we can see that, for the same f, fG(A)k = O
holds for every k. Hence we have
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
The combination of (H-16) and (H-17) yields
Because of
lB
]
I,)-IB +E
(s)
(H-16)
Since y(t)=Cx(t) +Eu(t)= Cx(t) +Eke A', we have

y(O) = Cx{O)
l') are polynomials, we
(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
Ik and with an irredu

Ilen A is a transmission
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
POLES AND ZEROS
zero of (;(s) if and only if

pS(A) = P [
Proof
:J
AI-A
-c
SI-A
p[
< n +min(p, q)
and the theorem follow
Let E =
(;( co).
The concept of tral

possible ways of definin
duce one more definitic
We factor (;(s) as
(;(s)=E +D- 1 (s)(s)=D- 1 (s)[D(s)E +(S)]~D-l(S)N(s)
(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)
where L(s) is defined as in (6-129). Furthermore, we have {D(s), L(s)} left

coprime and {sI -A o, Co} right coprime. Hence, Theorems G-S and G-S'
imply that there exist polynomial matrices X(s), - Ves), X(s), and Ves) such that
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)
X(s)X(s) +Y(S)Y(S)J (i-1-22)

Iq
The right-hand-side matrix is clearly unimodular; hence the left-hand-side

matrices are also unimodular (Why? Note that the left-hand-side matrices are
polynomial matrices). We compute
~(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.
The gcd of the three nur

zero of (;(5)_ 1t can be ~
- 2 (Problem H-5).
From this example,
of blocking zero are nc
blocking zero can neve
zero can also be a polt
zero of (;(s) will appear
transmission zero. Th(
Ir (;(s) is a scalar
Similar to transmi~
blocking the transmissi
component of the inpUl
where <p(s) andf(s) are

in the output y(t), then
is blocked from y{t).
Theorem H-7
Consider a system with

form shown in (H~25).
--~--
_ .._._._-_. . ._. . ._--- - - - - - - - - - - - _ . _ - - . _...- - - -
rOLES AND ZEROS
sI -A
p [ -C
l(p, q)
BJ
= p [SI -Aa
E
-Ca
BoJ
E
633
+pN(s)
(H-24 )
and the theorem follows from the definition of transmission zero.
Q.E.D.
=11
The concept of transmission zeros introduced aboye is only one of many

possible ways of defining zeros for rational matrices. In the following, we intro
duce one more definition.
]~D-l(S)N(s)
(H-19)
nd D(s) is row reduced.

irreducible {Aa, B o, Ca}
( H-20)
(H-21 )
'e have {D(s), L(s)} left

'heorems G-S and G-S'
X(s), and Y(s) such that
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
ritten in matrix form as

X(s) +Y(S)Y(S)]
1
(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
lence the left-hand-side

t-hand-side matrices are
)8 0 +Y(S)E]
)8 0
+ D(s)E
+- Y(s)ll.!'.l
)lB\o
N(s)
)(s)E +L(s)B o in (H-19)

limodular, we have, [01'
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 .
ty of Sciences and Technology,
Consider a system with proper transfer matrix G(s). Letthe input u(s) be of the
form shown in (H-2S). Then, for any initial state, no root of cP(s) will appear as
634
POlES AND ZEROS
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
Prove the identity
H-2
Consider the dyna
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)
Find an initial state so tI"
( H-26)
Consider the egua

= ' and ti
y(t) due to ult)
j= 1
Let f3(s) be the gcd of the numerators of all gij(s). Then we can write 9ij(S) as
nij(s)
%(s) = f3(s) dij(s)
A
H-4
Prove Theorems 1-:
H-S
What are the poles
and (H-26) becomes, by using (H-25),

y(s) =
j= 1
f3(s) nij(s)h(s)
<jJ(s) d}s)
Now if every root of <jJ(s) is a blocking zero, then <jJ(s) will divide f3(s). Hence
no root of <jJ(s) will appear as a mode of y(t).
Now suppose the root, s - IX, of <jJ(s) is not a blocking zero. Then there
exists at least one nkl(s) which does not contain 's - d. as a factor; otherwise,
s - ex would be a blocking zero. We choose /;(s) = O for j = 1, 2, ... ,1 - 1,
1+ 1, ... , p andfi(s) = 1, then
Yk(S) = f3(s) nk!(s)
<jJ(s) dkl(S)
H-G Let N(s)D-1(s) = t'

set of transmission zeros
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 [
under any state feedback a

conclusion from the facts
transfer-function matrix a
function matrix?
H-S
What are the blocki
POLES AND ZEROS
f cP(s) is a blocking zero
(s) contains no root of
zero-input response of
)f cP(s). Let Yi(S) be the
:gj(s)). Then the zero
( H-26)
n we can write gij(S) as
635
Problems
H-1
Prove the identity in Equation (H-l).
H-2
x= [ -1O -2lJ x+ [O'

Ju
1
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
Prove Theorems H-3 and H-4.
H-S
What are the poles and transmission zeros of the transfer matrices:
.1'-1
G1(s)=
O
A
.viii divide f3(s). Hence
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
cing zero. Then there

as a [actor; otherwise,
) for j=l, 2, ... ,/-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
le proo[ of this theorem

Q.E.D.
then e1. wiil not appear
A is a blocking zero of
y k. These concepts of
blems in Chapter 9.
G(s) are defined fram a
om Theorem H-6 that
n of G(s) by using. the
TI the Smith-McMilIan
1S of various definiiions
utation, see Re[erences
. open left half s-plane,
fhe presentation o[ this
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,
pp. 371-378, 1966.
3. Athans, M., and P. L.Falb, Optimal Control.

1966.
New York: McGraw-Hill,
4. Bass, R. W .. and I. Gura, "High order system dcsign via state-spacc con
siderations," Preprints 1965 JACC, pp. 311-318.
5. Bellman, R., Introduction to Matrix Analysis. New York: t",lcGraw-Hil1.

1960.
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.
8. - - , and M. Mesarovic, "The reproducibility of multivariable systems,"

J. Math. Anal. AppI:, vl. U,pp. 548-563, 1965. .
9. Buck, R.
c., Advar:ed Calculus.
New York: McGraw-Hill, 1956.
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
20. - - , and Deso

systems," IEEE
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
12. Chang, S. S. L., Synthesis ol Optimal Control Systems.

McGraw-Hill, 1961.
637
New York:
13. Chen, C. T., "Output controllability of composite systems," IEEE Twns.

Automatic Control, vol. AC-12, p. 201, 1967.
14. - - , "Linear independence of analytic functions," SIAM J. Appl. Math ..

vol. 15, pp. 1272-1274, 1967.
15. - - , "Representation of linear time-invariant composite systems." ! EEE
Trans. Automatic Control. vol. AC-13, pp. 277-283. 1968.
16. - - , "Stability of linear multivariable feedback systems," Proc. IEEE,
vol. 56, pp. 821-828,1968.
17. - - , "A note on pole assignment." IEEE Trans. Automatic Control, vol.
AC-13, pp. 597-598, 1968.
18. - - , "Design of feedback control systems," Proc. Natl. Electron. Cont:

vol. 57, pp. 46-51,1969.
19. - - , "Design of pole-placement compensators for multivariable systems,"
Preprints I970 J ACC.
1 in the preparation of
ion. The lists are not

ources.
:sign of multivariable
67, pp. 295.-399, 1967.
'oula, "Equivalence of
tnklin Inst., vol. 281,
'York: McGraw-Hil1,
;n via state-space con
, York: McGraw-Hili,
operties," Am. M ath
6-152,1968.
space interpretation,"
1-135,1965.
IUltivariable systems,"
'aw-Hill; 1956.
ystems/' IEEE Trons.
20. - - , and Desoer, C. A., "Controllability and observability of composite

systems," IEEE 7/-ans. Automatic Control, vol. AC-12, pp. 402-409, 1967.
21. - - , and - - , "A proof of controllability of Jordan form state equations,"
IEEE 7/-ans. Automatic Control, vol. AC-13, pp. 195-196, 1968.
22. - - , - - , and A. Niederlinski, "Simplified condition for controllability
and observability of linear time-invariant systems," IEEE 7/-ans. A utomatic
Control, vol. AC-ll, pp. 613-614,1966.
23. Chi, H. H_, and C. T. Chen, "A sensitivity study of analog computer
simulation," Proc. Allerton Conf, pp. 845-854, 1969.
24. Coddington, E. A., and N. Levinson, Theory of Ordinary Di[(erential
EquatiolJs. New York: McGraw-HilL 1955.
25. Dellon, F., and P. E. Sarachik, "Optimal control of unstable linear plants
with inaccessible states," IEEE 7/-al'ls. Automatic Control, vol. AC-13,
pp. 491 -495 19~8
26. Dertouzos, M. L., M. E. Kaliski, and K. P. Polzen, "On-line simulation of
block-diagram systems," IEEE 7/-ans. Computers, vol. C-18, pp. 333-342,
1969.
27. DeRusso, P. M., R. J. Roy, and C. M. Close, State Variablesfor Engineers.
New York: Wiley, 1965.
28. Desoer, C. A., "A general formulation of the Nyquist criterion," IEEE
7/-ans. Circuit Theory, vol. CT-12, pp. 230-234, 1965.
29. --,EECS 222 Lecture Notes, University ofCalifornia, Berkeley, Fall
1968'.
30. _."-.,
ontrollability," IEEE
1965.
and' C. T. Chen, "Controllability and observability of feedback

systems," IEEE 7/ons. Automatic Control, vol. AC-12,pp. 474-475,1967.
31. --,and E. S. Kuh, Basic Circuit Theory.

1969.
New York: McGraw-Hill,
638
REFERENCES
32. - - , and P. Varaiya, "The minimal realization of a nonanticipative

impulse response matrix," SI A M J. Appl. Math., vol. 15, pp. 754-763, 1967.
canonical (phasl
AC-9, pp. 312-3
33. - - , and M. Y. Wu, "Stability of linear time-invariant systems," IEEE

1I-ans. Circuit Theory, vol. CT-15, pp. 245-250, 1968.
34. - - , and - - , "Stability of multiple-Ioop feedback linear time-invariant
systems," 1. Math. Anal. Appl., vol. 23, pp. 121-129,'1968.
52. Joseph, P. D., a

Applications and
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
36. Falb, P. L., and W. A. Wolovich, "Decoupling in the design of multivariable

control systems," IEEE 1I-ans. Automatic Control, vol. AC-12, pp. 651-659,
1967,
37. Ferguson, J. D., and Z. V. Rekasius, "Optimallinear control systems with
incomplete state measurements," IEEE 1I'ans. Automatic Control, vol.
AC-14, pp. 135-140,1969.
38. Friedman, B., PrincipIes and Techniques of Applied Mathematics. New

York: Wiley, 1956.
39. Gantmacher, F. R., The Theory of Matrices, vols. 1 and 2. New York:
Chelsea, 1959.
40. Gilbert, E. G., "Controllability and observability in multivariable control
systems," SIAM J. Control, vol. 1, pp. 128-151,1963.
41. - - , "The decoupling of multivariable systems by state feedback," SI AM
J. Control, vol. 7, pp. 50-63, 1969.
42. Gueguen, C. J., and E. Toumire, "Comments on 'Irreducible Jordan form
realization of a rational matrix,'" IEEE 1I-ans. Automatic Control, vol.
AC-15,1970.
43. Hadley, G., Linear AIgebra. Reading, Mass.: Addison-Wesley, 1961.
44. Halmos, P. R., Finite Dimensional Vector Spaces, 2d ed. Princeton, NJ.:
Van Nostrand, 1950.
45. Herstein, I. N., Topics in AIgebra. Waltham, Mass.: Blaisdell, 1964.
46. Heymann, M., "'Comments cOn pole assignrnenl in muill-itlpul COl
trollable linear systems,'" IEEE 1I-ans. Automatic Control, vol. AC-13,
pp. 748-749, 1968.
47. Ho, B. L., and R. E. Kalman, "Effeetive construction oflinear state variable
models from inputfoutput data," Proc. Third Allerton Conf, pp. 449-459,
1965.
48. Ho, Y. c., "What constitutes a controllable system," 1RE Trans. Automatic
.
Control, vol. AC-7, p. 76, 1962.
53. J ury, E. L, Sampl

54. Kalman, R. E., "
56. - - , "Contribut
vol. 5, pp. 102-1
57. _._, "Canoniea
Acad, Sci. U.S., v
58. - - , "On the sta

Theory, vol. CT-'
59. - - , "Lyapunov
Proc, N atl. Acad.
60. - - , "Mathema
Control, vol. 1, PI
61.--, "When is a
vol. 86, pp. 51-6C
62. - - , "Irreducib1
SI A M 1., vol. 13,
63. - - , "Toward a
Proc. 4th 1BM Se
64. - - , "On structL
reprint of paper
Automatic Contr
65" ------, and J. E. 1
seco no metnod' (
1960.
66. - - , and R. S. Bu
1I-ans. ASME, ser
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,
51. Johnson, C. D., and W. M. Wonham, "A note on the transformation to
71. Kreindler, A, and
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.
canonical (phase-variable) form," IEEE 1Jans. Automatic Control, vol.
AC-9, pp. 312-313,1964.
3.riant systems," IEEE
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.
onal matrix function,"
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.
.r control systems with

utomatic Control, vol.
57. - - , "Canonical structure of linear dynamical systems," Proc. N atl.

Acad. Sci. U.S., vol. 48, no. 4, pp. 596-600, 1962.
'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:
59. - - , "Lyapunov functions for the problems of Lur'e in automatic control,"

Proc. Natl. Acad. Sci. U.S., vol. 49, pp. 201-205, 1962.
1 and 2.
1
60. - - , "Mathematical description of linear dynamical system," SI AM J.

Control, vol. 1, pp. 152-192,1963.
state feedback," SIAM
61. - - , "When is a linear control system optimal?" Trans. ASM E, ser. O,

vol. 86, pp. 51-60, 1964.
mu ltivariable control
53.
-red ucible J ordan form

.utomatic Control, vol.
son-Wesley,1961.
d ed. Princeton, N.J.:
.: Blaisdell, 1964.
62. - - , "Irreducible realizations and the degree of a rational matrix,"

SIAM J., vol. 13, pp. 520-544,1965.
63. - - , "Toward a theory of difficulty of computation in optimal control,"
Proc. 4th IBM Sci. Computo Symp., pp. 25-43, 1964.
64. - - , "On structural properties of linear constant multivariable systems,"
l'eprint of papel' 6A, Thil'd Congl'ess of the lnternational Federation of
Automatic Control, 1966.
t in muti-input con
; Control, vol. AC-13,
M. - - , and J. E. Bertram, "Control system a.nalysis and design va the

'second method' ofLyapunov," "li'ans. ,,;S!Vi E, S;;I. S', mi. 82, pp. ~./ i .:;~:,
1960.
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,
. ~ _ . _ , ~
67. - - , and T. S. Engler, A User's Manualfor Automatic Synthesis Programs,
NASA, Washington, O.e., 1966.
, IRETrans. Automatic
68. - - , P.L. Falb, and M. A. Arbib, Topics in Mathemat';al System Theory.

New York: McGraw-Hill, 1969.
multivariable feedback
69. - - , Y. e. Ho; and K. S. Narendra, "Co~tiollability of linear dynamical
systems,'; Contrib. Differential Equations, voL l~pp. 189:"213, 1961: .
erator(')=A(') +B('),"
70. Kaplan, W" Operational Methods for Linear Systems.

Addison-Wesley, 1962.
the transformation to
Reading, Mass,:
71. Kreindler, A., and P. E. Sarachik, "On the concepts of conlrollability and
640
REFERENCES
observability of linear systems," IEEE Trans. Automatic Control, vol.

AC-9, pp. 129-136, 1964.
72. Kuh, E. S., and R. A. Rohrer, "The state-variable approach to network

analysis," Proc. IEEE, vol. 53, pp. 672-686, 1965.
73. - - , and - - , Theory of Linear Active Networks.
Holden-Day,1967.
San Francisco:
74. Kumar, S., "Computer-aided design of multivariabie systems," M.S.

thesis, State University ofNew York, Stony Brook, 1969.
75. Kuo, F. F., and J. F. Kaiser, System Analysis by Digital Computer. New
York: Wiley, 1966.
76. LaSalle, l, and S. Lefschetz, Stability of Liapul1ov's Direct Method. New
York: Academic, 1961.
77. Lefschetz, S:, Differential Equations: Geometric Theory.
Interscience, 1957.
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.
vol. 42, pp. 355
"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.
97. Schwarz, R. J.,
Hill, 1965.
79. - - , "Observers for multivariable systems," IEEE Trans. Automatic

Control, vol. AC-ll, pp. 190-197, 1966.
98. Silverman, L. 1\
Ph. D. dissertati
80. - - , "Canonical forms for linear multivariable systems," IEEE Trans.

Automatic Control, vol. AC-12, pp. 290-293,1967.
99. - - , "Transfo
variable) form,"
1966.
81. Mantey, P.. E.,. "Eigenvalue sensitivity and state-variable selection,"
1EEE 1ians. Automatic Control, vol. AC-13, pp. 263-269, 1968.
82. Mayne, D. Q., "Computational procedure for the minimal realization of

transfer-function matrices," Proc. 1EE (London), vol. 115, pp. 1363-1368,
1968.
83. McMillan, B., "Introduction to formal realizability theory," Bell System
Tech. 1., vol. 31, pp. 217-279,541-600,1952.
84. Morgan, B. S., Jr., "The synthesis of linear multivariable systems by state
variable feedback," Proc. 1964 lACC, pp. 468-472.
100. - - , "Stable re:

Conv. Record, VI
101. - - , "Synthesi
passive realizati
1968.
102. - - , and B. O.
01 linear system:
103. - - , and H. E
; . ~.,..,
",
85. Narendra, K. S., and C. P. Neuman, "Stability o a class 01 ulferemial

equations with a single monotone linearity," SI AM 1. Control, vol. 4,
pp. 295-308, 1966.
86. Nering, E. D., Linear Algebraand Matrix Theory. NewYork: Wiley, 1963.
87. Newcomb, R. W., Active Integrated Circuit Synthesis. Englewood Cliffs,
N.J.: Prentice-Hall, 1968.
88. Ogata, K., State Space Analysis of Control Systems. Englewood Cliffs,
N.J.: Prentice-Hal1, 1967.
89. Panda, S. P., and C. T. Chen, "Irreducible Jordan form realization of a
rational matrix," IEEE Trans. Automatic Control, voL AC-14, pp. 66-69;
1969.
90. Parks, P.
c., "A new proof of the Routh-Hurwitz stability criterion
using
"":',
104. Truxal, J. G.,
105. Weiss, L., "The

observability," j
106. - - , "On the st
Princeton Con!
659-680, 1968.
107. ---, "Lectures
590, University ,
108. ----:-:-, and P. L. I
parametrized elf
vol. 3, pp. 67-7':;
REFERENCES

~
approach to network
'orks.
San Francisco:
lfiable systems," M.S.

(,1969.
)igital Computer. New
s Direct Method. New

Theory.
New York:
Lr system," IEEE Trans.

EEE 7/ans. Automatic
systems," IEEE Trans.

ate-variable selection,"
53-269, 1968.
: minimal realization of
'01. 115, pp. 1363-1368,
ity theory," Bell System
ariable systems by state
2.
)f a class of differentia
AM J. Control, vol. 4,
New York: Wiley, 1963.
esis. Englewood Cliffs,
?ms. Englewood Cliffs,
form realization of a
, vol. AC-14, pp. 66-69,
Lll
stability criterion using
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.:
93. Rekasisu, Z. V., "Oecoupling of multivariable systems by means of state

variable feedback," Proc. Third Allerton Conf, pp. 439-447, 1965.
94. Sandberg, I. W., "Linear multiloop feedback systems," Bell System Tech. J.,
vol. 42, pp. 355-382, 1963.
95. - - , "On the Lz-boundness of solutions of nonlinear functional equa
tions," Bell System Tech. 1., vol. 43, pp. 1581-1599, 1964.
96. Schwartz, L., Thorie des distributions. Paris: Hermann & Cie, 1951, 1957.
97. Schwarz, R. J., and B. Friedland, Linear Systems. New York: McGraw
Hill, 1965.
98. Silverman, L. M., "Structural properties of time-variable linear systems,"
Ph.O. dissertation, Oept. of Elec. Eng., Columbia University, 1966.
99. - - , 'Transformation of time-variable systems to canonical (phase
variable) form," IEEE 7/ans. Automatic Control, vol. AC-ll, pp. 300-303,
1966.
100. - - , "Stable realization of impulse response matrices," 1967 IEEE I ntern.
Conv. Record, vol. 15, pt. 5, pp. 32-37.
un. --, "Synthesis of impulse response matrices by internally stable and
passive realizations," IEEE Trans. Circuit Theory, vol. CT-15, pp. 238-245,
1968.
102. --, and B. O. O. Anderson, "Controllability, observability and stability
of linear systems," SIAM J. Control, vol. 6, pp. 121-129,1968.
103. - - , and H. E. Meadows, "Controllability and observability in time
104. Truxal, J. G., Control System Synthesis. New York: McGraw-Hill, 1955.
105. Weiss, L., 'The concepts of differential controllability and differential
observability," J. Math. Anal. Appl., vol. 10, pp. 442-449, 1965.
106. - - , "On the structure theory of linear differential systems," Proc. Second
Princeton Conf, pp. 243-249, 1968; also SIAM 1. Control, vol. 6, pp.
659-680, 1968.
107. --, "Lectures on controllability and observability," Tech. Note BN
590, University of Maryland, Jan. 1969: .
108. --,and P. L. Falp, "Oolezal's theOl:em, linear algebra with continuously
parametriZed elements,and time-varying systems," Math. System TheorJ',
vol. 3, pp. 67-75,1969.
642
REFERENCES
109. - - , and R. E. Kalman, "Contributions to linear system theory," 1m. J.

Eng. Soc., vol. 3, pp. 141-171,1965.
of interconnec
vol. AC-28, pp
110. Wolovich, N. A., "One state estimation of observable systems," Preprints

1968 J ACC, pp. 210-220.
S9. Armstrong, E.
control. New
111. Wonham, W. M., "On pole assignment in multi-input controllable linear

systems," IEEE Trans. Automatic Control, vol. AC-12, pp. 660-665, 1967.
SlO. Astrom, K. J.,

Academic, 19T
112. - - , and A. S. Morse, "Decoupling and pole~assignment in linear multi

variable systems-a geometric approach," SI AM J. COl7trol, vol. 8, pp.
1-18,1970.
SIL -.- , "Algebn

Polytechnica S.
113. Wylie, e R., Jr., Aduanced Engineering Mathematics. New York:

McGraw-Hill,1951.
S12. Athans, M., M

Networks ami
McGraw-Hill,
114. Youla, D. e, 'The synthesis of linear dynamical systems from prescribed

weighting patterns," SI AM 1. Appl. Math., vol. 14, pp. 527-549, 1966.
S13. Balestrino, A.,

variable systen
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:
117. Zames, G., "On the input-output stability of time-varying nonlinear

feedback systems," pts. I and n, IEEE y,ans. Automatic Control, vol.
AC-ll, pp. 228-238 and 465~476, 1966.
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.
Englewood Cliffs, NJ.:
SS. Antsaklis, P. J., and J. B. Pearson, "Stabilization and regulation in

linear multivariable systems," IEEE Trans. Automatic Control, vol.
AC-23, pp. 928-930, 1978.
matic Control, .
S14. Bar-ness, Y., a

sampling," 1ni.
S15. Barnett, S., fI;.
Reinhold, 1971
S16. - - , "Regular
Proc. Cambrid[;
817. Barry, P. E., ":
unpublished ne
S18. Bartels, R. H., ~

equation AX +
826,1972.
S19. Bengtsson, G.,
dornain approa
S20. Bhattacharyya,
Automatic COl1l
of rnultivariabl,
Atherton, Ed.),
S22. Birkoff, G. and
York: Macmill
S23. Bitmead, R. R.,
common diviso
Trans. Automat
S6. Aplevich, J. D., "Direct computation of canonical forms oflinear systems

by elementary matrix operations," IEEE' y,"ans. Automatic Control,
vol. AC-19, pp. 124-126, 1974.
.
S24. Bongiorno, J. J.
system," I nI. J.
pp. 183-190, 19
S7. - - , "Tableau methods for analysis' an,d design of linear systems,"

Automatica, vol. 15, pp. 419-429, 1979.
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
;ystem theory," Int. 1.
of interconnected dynamical systems," IEEE Trans. Automatic Control,

vol. AC-28, pp. 569-577, 1983.
)le systems," Preprints
S9. Armstrong, E. S., ORACLS: A design system for linear multivariable

control. New York: Dekker, 1980.
)Ut controllable linear

12, pp. 660-665, 1967.
S10. Astrom, K. 1., Introduction to Stochastic Control Theory.

Academic, 1970.
nment in linear multi

J. Control, vol. 8, pp.
SI L - - , "Algebraic system theory as a tool for regular design," ACTA

Polytechnica Scandinavica, Ma 31, pp. 52-65, 1979.
emaUcs. New York:
SU. Athans, M., M. L. Destouzos, R. N. Spann, and S. 1. Mason, Systems,

Networks and Computational Multivariable Methods.
New York:
McGraw-Hill,1974.
,tems from prescribed

pp. 527-549,1966.
:e extraction-part 1,"
83-208.
Theory.
New York:
ne-varying nonlinear
ltomatic Control, vol.
bility of linear time

vol. 20, pp. 408-413,
'1,
:ero cancellations and

Circuits and Systems,
Syl'vesteT rD.atrices in.
ic Control, vol. AC-21,
glewood Cliffs, N.J.:

)TI and regulation in
romatic Control, vol.
N ew York:
S 13. Balestrino, A., and G. Celentano, "Pole assignment in linear multi

variable systems using observers of reduced order," IEEE ll-cms. Auto
matic Control, vol. AC-24, pp. 144-146, 1979.
S14. Bar-ness, Y., and G. Langhotz, "Preservation of controllability under
sampling," Int. J. Control, vol. 22, pp. 39-47, 1975.
S15. Barnett, S., Matrices in Control Theory. London: Van Nostrand
Reinhold,1971.
S16. - - , "Regular greatest common divisor of two polynomial matrices,"
Proc. Cambridge Philos. Soc., 72, pp. 161-165,1972.
SI7.. Barry, P. E., "State feedback, pole placement and transient response,"
unpublished notes.
S18. Bartels, R. H., and G. H. Stewart, "Algorithm 432, solution of the matrix
equation AX +XB =C," Commun. Ass. Comput Mach., vol. 15, pp. 820
826,1972.
S19. Bengtsson, G., "Output regulation and internal models-a frequency
domain approach," Automatica, vol. 13, pp. 333-345, July 1977.
S20. Bhattacharyya, S. P., "The Structure of Robust Observers," 1EEE 1l-cms.
Automatic Control, vol. AC-21 (4), pp. 581-588, 1976.
of multivariable systems," in Multivariable Technological Systems (D. P.
Atherton, Ed.), New York: Pergamon, 1979.
S22. Birkoff, G. and S. MacLane, A Survey of Modern Algebra, 3d ed. New
York: Macmillan, 1965.
S23. Bitmead,R. R., S. Y. Kung, B. D. O. Anderson, and T. Kailath, "Greatest
common divisors via generalizedSylvester and Bezout matrices," IEEE
Trans. Automatic Control, vol. AC-23, pp. 1043-1047,1978.
xms oflinear systems

. Automatic Control,
S24. Bongiorrio, J. J.; andD. C. Youla, "On observers in multivariable control

system," 1m. 1. Control, vol. 8, pp. 221-243, 1968; its discussion; Yo.1.12,
pp. 183-190, 1970.
n of linear systems,"
S25. - - , and - - , "On the design ofsingle-loop single-input-otput feed

back systems in thecomplex-frequency domain," 1EEE 1)ans. Automatic
Control, vol. AC-22, pp. 416-423,1977.
for the well-posedness
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
835. - - , and C. D. Nahum, "Necessary and sufficient conditions for the

complete controllability and observability of systems in series using the
coprime factorization of a rational matrix," IEEE 7ans. Circuits and
Systems, CAS-22, pp. 90-95, 1975.
S3~ . Chang, F. C.~ ttThe inverse of gEn~n:1.Hz~d 'O/~1.:-~c!e:'~lT1Gnd /r.L:~~:L~ ~~d;)L~6~
the partal fraction expansion," IEEE Trans. Automatic Control, vol. 19,
pp. 151-152, 1974.
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
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.
841. - - , "A generalizatian of the inertia theorem," SIAM J. Appl. Matiz.,

vol. 25, no. 2, pp. 158-161,1973.
S56. - - , and -
Trans. Automati,
858. Csaki, F. G., "S

matrices," IEEE
859. D' Alessandro, O
lordan form rea
1971.
860. Daly, K. C.,"T
elementary sim
pp. 1-15, 1976.
;nment using dynamic

1. AC-15 (l), pp. 34-43,
842. - - , "Minimization of linear sequential machines," IEEE 71a/1s. Com

puters, vol. C-23, pp. 93-95,1974.
843.
tS.
New York: Wiley,
)f realization based on
EEE 71ans. Automatie
645
"An algorithm for Jordan form transformation," J. Fra/1klin

Inst., vol. 297, pp. 449-455, 1974.
--o
S44. - - , "Inertia Theorem for general matrix equations," 1. Math. Anal.

Appl., vol. 49, pp. 207-210,1975.
845. --o "Synthesis of linear sequential machines," In[ormation Control.
vol. 32, pp. 112-127, 1976.
able systems," Kyber
S46. - - , Analysis amI Synthesis 01linear control systems. Stony Brook, N.Y.:
Pond Woods, 1978.
systems from input

lntrol, vol. AC-16, pp.
S47. - - , One-dimensional Digital Signal Proeessing. New York: Dekker,

1979.
by static and dynamic

and Control, pp. 130
transfer fu nctions for
Trans. Cireuits and
ification and c\arifica
,turbance rejection in
Con! dn Deeision and
848. - - , "Identification of linear time-invariant multivariable systems,"

Proe. 1982 1FA C Symposium on Ident Uleea ion, J une 1982.
849. - - , "A contribution to the design of linear time-invariant multivariable
systems," Proe. Am. Automatie Control Conf., June, 1982.
850. - - , "Feedback implementations of open-Ioop compensators." Tech.
Report, SUNY, Stony Brook, N.Y., 1983.
851. - - , and C. H. Hsu, "Design of dynamic compensator for multivariable
systems," Preprint J A CC, pp. 893-900, 1971.
S52. - - , and y.. S. Kao, "Identification of two-dimensional transfer functions
from finite input-output data," IEEE 71ans. Automatie Control, vol.
AC-24, pp. 748-752,1979.
New York: Springer
S53. - - , and D. P. Mital, "A simplified irreducible realization algorithm,

IEEE 71ans. Automatie Control, vol. AC-17, pp. 535-537,1972.
:nt conditions for the

:ms in series using the
'E Trans. Cireuits and
854. - - , and S. Y. Zhang, "Multivariable multipurpose controllers."

System and Control le tt as, vol. 3, pp. 73-76, 1983.
855. Cheng, L., and J. B. Pearson, Jr., "Frequency domain synthesis of multi
variable linear regulators," IEEE Trans. Automatie Control, vol. AC-23,
:londc iTiali:ix Lhroug
matie Control, vol. 19,
'witz criterion," IEEE

, 1974.
'y.
New York: Holt,
zations of sets of dif

'01, vol. AC-15, p. 131,
oe. IEEE, vol. 59, pp.
'IAM 1. Appl. Math.,
856. - - , and - - , "Synthesis of linear multivariable regulators," IEEE

71-ans. Automatie Control, vol. AC-26, pp. 194-202, 1981.
857. Cheng, V. H. L., and C. A. Desoer, "Limitations on the closed-Ioop
transfer function due to right-half plane transmission zeros of the plant,"
IEEE Trans. Automatie Control, vol. AC-25, pp. 1218-1220, 1980.
858. Csaki, F. G., "Sorne noteson the inversion of confluent Vandermonde
matrices," IEEE Trans. AutomatieControl, vol. AC-20, pp. 154-157, 1973.
859. D'Alessandro, O. P., anci S: .Guili;nelli, "A direct procedure for irreducible
'Jordan form realization," Ricnerehe Di Automatiea, vol. 2, pp. 193-198,
1971.
860. Daly, K. c., "The computationof Llienberger canonical forms using
elementary similarity transformations," Int. 1. Systems Sei., vol. 7,
pp. 1-15, 1976.
646
REFERENCES
561. Oamen, A. A. H., and A. K. Hajdasinski, "Practical tests with different

approximate realizations based on the singular value decomposition of
the Hankel matrix," Proc. 6th 1FAC Symposium on Identification,
Washington, O.C, 1982.
862. Oatta, K. B., "An algorithm to compute canonical forms in multivariable
control systems," IEEE Trans. Automatic Control, vol. AC-22, pp. 129
132, 1977.
863. - - , "Mnimal realization in companion forms," J. Frank./in Inst., vol.
309, pp. 103-123, 1980.
864. Davison, E. J., "The output control of linear time-invariant multivariable
systems with unmeasurable arbitrary disturbances," IEEE Yans. Auto
matic Control, vol. AC-17 (5), pp. 621-630, 1972; its correction, vol.
AC-20, p. 824, 1975.
865. - - , "A generalization of the output control of linear multivariable
systems with unmeasurable arbitrary disturbances," IEEE Yans. Auto
matic Control, vol. AC-20 (6), pp. 788-792, 1975.
866. - - , "The robust control of a servomechanism problem for linear time
invariant multivariable systems," IEEE Trans. Automatic Control, vol.
AC-21 (1), pp. 25-34, 1976; addendum, vol. AC-22, p. 283,1977.
867. - - , "Sorne properties of minimum phase systems and 'squared-down'
868. - - , and A. Goldenberg, "Robust control of a general servomechanism
problem: The servo compensator," Autoinati~a, vol. 11 (5), pp. 461-471,
1975.
569. - - , and S. H. Wang, "Properties of linear time-invariant multivariable
systems subject to arbitrary output and state feedback/' IEEE Trans.
Automatic Control, vol. AC-18, pp. 24-32, 1973; its correction, vol.
AC-18, p. 563, 1973.
870.
~-,
and - - , "Properties and calculation of transmission zeros oC

linear multivariable systems," Automatica, vol. 10, pp. 643-658, 1974.
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
Control, pp. 89]
884. - - , and P. V
transfer functio
55, 1982.
885. Emre, E., ''rhe
dynamic feedb~
620,1980.
886. - - , "Pole-zer
1EEE Con! De<
l
tion for relatiVt

Control, vol. AC
using output feedback," IEEE Trans. Automatic Control, vol. AC-20,

pp. 516-518,1975.
872. - - , and - - , "An algorithm for the calculation of transmission
zeros of (C, A, B, O) using high-gain output feedback," IEEE Trans.
Automatic Control, vol. AC-23, pp. 738-741, 1978.
888. - - , and -
solutions, "SI A
873. Oesoer, C. A., Notesfor a Second Course on Linear Systems. Ne:w York:
Van Nostrand, 1970.
890. Fadeev, O. K.,

Algebra. San F
874. - - , F. M. Callier, and W. S. Chan; '~Robustness of stability conditions

for linear time-invarant feedback systems," 1EEE Y~ans. Automatic
.
Control, vol. AC-22, pp.. 586-590, 1977.
591. Fahmy, M. M.,

multivariable s:
pp. 690-,-693, 19:
875..- - , and W. S.Chan, "The feedback interconnection of lumped linear

time-invariantsystems," J. Frank.lin Inst., vol. 300, pp. 335-351, 1975~
892. Fairman, F. W.
889. Eykhoff, P., SY5
;al tests with different

l1ue decomposition of
um on Identification,
forms in multivariable
. vol. AC-22, pp. 129
J. Franklin Inst., vol.
lVariant multivariable
,," IEEE u-ans. Auto
2; its correction, vol.
f linear multivariable
,," IEEE Trans. Auto
:oblem for linear time
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
876. - - , and M. J. Chen, "Oesign of multivariable feedback systems with

stable plant, and its extension," IEEE T/ans. Automatic Control, vol.
AC-26, pp. 408-415, pp. 526-527,1981.
877. - - , R. W. Liu, John Murray, and Richard Sacks, "Feedback system
design: The fractional representation approach to analysis and synthesis,"
1EEE Trans. Automatic Control, vol. AC-25, pp. 399-412, 1980.
878. - - , and J. O. Schulman, "Zeros and poles of matrix transfer functions
and their dynamic interpretation," IEEE Trans.Circuits and Systems,
vol. CAS-21 (1), pp. 1-8, 1974.
879. - - , and M. Vidyasagar, Feedback Systems: Input-Output Properties.
New York: Academic, 1975.
880. - - , and Y. W. Wang, "On the minimum order of a robust servo
compensator," IEEE Trans. Automatic Control, vol. AC-23 (1), pp. 70-73,
1976.
881. - - , and Y. T. Wang, "Linear time-invariant robust servomechanism
problem: A self-contained exposition," in Control and Dynamic Systems,
vol 16 (c. T. Leondes, Ed.). New York: Academic, pp. 81-129,1980.
882. Oongarra, J. J., C. B. Moler, J. R. Bunch and G. W. Stewart, LINPACK
User's Guide. Philadelphia: SIAM, 1979.
883. Emami-Naeini, A., and G. F. Franklin, "Zero assignment in the multi
variable robust servomechanism," Proc. 21st IEEE Con! on Decision and
Control, pp.891~893, 1982.
884. - - , and P. VanDooren, "On computation of transmission zeros and
transfer functions," Proc. 21st IEEE Con! Decision and Control, pp. 51
55,1982.
885. Emre, E., "The polynomial equation QQc +RP c =<1> with application to
dynamic feedback," SIAM 1. Control & Optimization, vol. 18, pp. 611
620,1980.
886. - - , "Pole-zero cancel1ations in dynamic feedback systems," Proc. 21st
1EEE Con! Decision and Control, pp. 422-423, 1982.
tion for relatively prime polynomial matrices," IEEE Trans. Automatic
Control, vol. AC-22, pp. 239-242, 1977.
888. - - , and - - , "The equation XR +QY=<1>: A characterization of
solutions, "SIAM 1. Control & Optimization, vol. 19, pp. 32-38, 1981.
889. Eykhoff, P., System Identification. New York: Wiley, 1974.
Systems. New York:
890. Fadeev, O. K., and V. N. Faddeeva.,Computational Methods in Linear

Algebra. San Francisco: Freeman, 1963.
)f stability coriditions
EE 1Ians.. Automatic
891. Fahmy, M. M., and J. O'Reilly, "On eigenstructure assignment in linear

multivariable systems," IEEE Trans. Automatic Control, vol. AC-27,
.. pp. 690-693," 1982.
tion of lumped linear

pp. 335-351, 1975.
892. Fairman, F. W., and R. O.Gupta, "Design of multifunctional reduced
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.
SI00. - - , and - - , "The internal model principie of control theory," Auto

matica" vol. 12,pp. 457-465, 1976.
S101. Franklin, G. F., and e. R. Johnson, "A condition for full zero assignment
in linear control systems," IEEE Trans. Automatic Control, vol. AC-26,
pp. 521-523,1981.
S102. Fuhrmann, P. A., "AIgebraic system theory: An analyst's point of view,"
J. Franklin Inst., vol. 301, pp. 521-540, 1976.
8103. Garbow, B. S., J. M. Boyle, J. J. Dongarra, and e. B. Moler, "Matrix
eigensystem routines-Eispack guide extensions," Lecture Notes in
Computer Sciences, vol. 51. New York: Springer-Verlag, 1977.
~. Gentn, Y. V., ano S. Y. Kung, "A two-variabe approach to the moaei
reduction problem with Hankel norm-criterion," IEEE y,ans. Circuits
Systems, vol. CAS-28, pp. 912-924, 1981.
SI05. Gibson, J. A., and T. T. Ha, "Further to the preservation of controllability
under sampling," lnt. J. Control, vol. 31, pp. 1013-1026, 1980.
SI06. Gohberg, 1. e., and L. E. Lerer, "Resultants of matrix polynomials,"
Bull. Amer. Math. Soc., vol. 82, pp. 565-567, 1976.
S107. Golub, G. H., S. Nash ande. Van Loan, "A Hessenberg-Scheer method
for the problem AX + XB ==C," 1EEEY,ans. Automatic Control, vQ!; AC
24, pp. 909-913, 1979.
.
8108. - - , and J. H. Wilkinson, "Ill-conditioned eigensystems aI1d the com
putation oftheJordan canonical form," SIAM Rev., vol. 18, pp. 578-619.
1976.
SI09. Gopinath, B., ,

Bell Syst. Tech.
SIlO. Gupta, R. D., !
the design of si
vol. CAS-28. Pi
8111. Ham, Y. P., an
Liapunov theo
733- 734,1981
SH2. Hayton, G. E.,
32, pp. 567-57~
SI13. Heymann, M.,
D ynamical Syst
S114. - - , 'The po
Control, vol. A(
S115. Hikita, H., "De~
to output tracki
S116. Householder, J
Waltham Mass.
S117. Hsia, T. e., SYSI
S118. Hsu, e. H., and
matrices," lnt. J
SU9. Ikeda, M., H. 1
linear time-vary
Vol. 13, pp. 304
S120. Inouye, Y., "N
discrete-time sy~
S121. Johnson, e. D..
Regulator and
Control, vol. AC
amI Dynamic Sys
pp. 287-489, 19~
S 123. Jury, E. 1., lnner
1974.
S 124. Kagstrom, B.,
al
of the Jordan no
M ath. Software, .
S125. Kailath, T., Linel
S126. Kalman, R. E.. "(
forros," ACTA, J
S127. - - , "On the COI

SfAMJ. Controi
"--"'-'-'_."."---
_._--- - --
SUPJ>LEMENTARY REFERENCES
083-1094, 1980.
perturbational model
,utomatic Control, vol.
)Iocking zeros," IEEE
-259, 1977.
or spaces with applica
)I1trol, vol. 13, pp. 493

~ion
..
"--'---'----~~~ ~ ~
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
S1l3. Heymann, M., Structure and Realization Problems in the Theory oJ

Dynamical Systems. New York: Springer-Verlag, 1975.
:r observers for linear

:ontrol, vol. AC-17, pp.
Sl14. - - , "The pole shifting theorem revisited," IEEE 7i-ans. Automatc

Control, vol. AC-24, pp. 479-480, 1979.
control theory," Auto-
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.
or full zero assignment

;c Control, vol. AC-26,
S1l9. Ikeda, M., H. Malda, and S. Kodaina, "Estimation and Feedback in

linear time-varying systems: A deterministic theory," SIAM 1. Control,
Vol. 13, pp. 304-326, 1975.
nalyst's point of view,"
S120. Inouye, Y., "Notes on controllability and constructability of linear

discrete-time systems," I nt. 1. Control, vol. 35, pp. 1081-1084, 1982.
ism problem from the

Control, vol. AC-22 (3),
nission zeros in linear
(5), pp. 657-681,1975.
B. Moler, "Matrix
1S,"
Lecture Notes in
[-Verlag, 1977.
lpproach to the model
, IEEE 1i-ans. Circuts
lation of controllability
--1026,1980.
S121. Johnson, e D., "Accommodation of External Disturbances in Linear

Regulator and Servomechanism Problems," IEEE 7i-ans. Automatic
Control, vol. AC-16 (6), pp. 635-644 (1971).
SJ.2!..? --., 4'Theory o disturbance-aCCOIDJIJ.odadng controners~)) ~n ~~(}ntr(}r

and Dynamc Systems, vol. 12 (e T Leondes, ed.). New York: Academic,
pp. 287-489, 1976.
S 123. Jury, E. L, Inners and Stability of Dynamic Systems. New York: Wiley,
1974.
, 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
SUS. Kailath, T., Linear Syilems. EnglewoodCliffs,N.J.: Prentice-Hall, 1980.
lsystems and the com
vol. 18,.pp. 578-619,
'V.,
S 126. Kalman, R. E., "On partial realizations; transfer functions, and canonical
forms," ACTA', Polytechnca Scandinavca, Mil. 3i, pp. '9-32,1979.
S 127. - - , "On the computation of the reachablejobservable canonical form,';
SIAM J. Control & Optmizaton, vol. 20, pp. 408-413,1982.
_-_ __..-.
.
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.
8146. Lal, M., H. S

algorithm for
IEEE y,-ans., \
S147. Laub, A. J., "
MIT Rep. ESI
8148. - - , and B. t
techniques, Al:
S149. Lu, R., and 1.
fiability of in~
vol. AC-22, pp.
S150. Luenberger, O
matic Control, .
S151. MacFarlane, A
Proc. 2 1st lEE.
S152. MacFarlane, A
variable systen
variable theory
8153. Mansour, M.,
Lyapunov met
707- 708, 1982.
S154. Maronlas, J., a
control system
Int. J. Systems:
vol. 10, pp. 33
8155. Mayne. O. Q.,
rea lization alg<
pp. 306-307. 19
S156. Miminis, G. S.,
time-invariant l
and --
S157.
--o
"/-.,h,,
8141. - - , "A new iow-order approximation a!go!"ithr:0 "/1'"
decomposition," Proc. 12th Amer. Asilomar Con! Circuits, Systems alld
Computers, 1978.
S142. - - , and T. Kailath, "Fast projection methods for minimal design

problems in linear system theory," Automatica, vol. 16, pp. 399-403, 1980.
S143. - - , - - , and M. Morf, "A generalized resultant matrix for polynomial
matrices," Proc. IEEE Corif. on Decision and Control, pp. 892-895, 1976.
S 144. - - , - - , and - - , "Fast and stable algorithms for minimal design
problems," Proc. Fourth IFAC Int. Symposium on Multivariable Tech.
Systems (D. P. Altherton, ed). Candon: Pergamon, pp. 97-104,1977.
S145. Kuo, Y. L., "On the im;ducible Jordan form realization and the degree
of a rational matrix," IEEE Trans. Circuir Theory, vol. CT-17, pp. 322
332,1970.
pp. 62-67, 1982

S158. Mital, O. T., a
time-invariant s
8159. Moler, e B., an
the exponential
S16O. Moore, B. e, '
variable system
Trans. Atomati
vol. AC-22:pp.
1977.
'.
S161. - - , "Singular
ConI Decision e
651
for linear time-varying

r, vol. AC-24, pp. 866
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.
edback," IEEE Trans.
S147. Laub, A. 1., "Linear multivariable control: numerical considerations,"

MIT Rep. ESL-P-833, July 1978.
r EEE
Trans. Automatic
S148. - - , and B. C. Moore, "Calculation of transmission zeros using QZ

techniques, Automatica, vol. 14, pp. 557-566, 1978.
lear feedback control

~2, pp. 846-854, 1977.
I nI. J. Control, vol. 28,
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.
l.ed eigenvector assign

ic Control, vol. AC-22,
SISO. Luenberger, D. G., "An introduction to observers," IEEE Tans. Auto

matic Control, vol. AC-16, pp. 596-603, 1971.
S151. MacFarlane, A. G. J., "Gains, phase and angles in multivariable systems,"
Proc. 21st IEEE Con! on Decision and Control, pp. 944-947, 1982.
ue decomposition: Its
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.
bili ty critera," IEEE

;5, 1980.
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.
'ations," IEEE Trans.
1Jiciency. New York:
ial Equation Approach.
;ystems: Analysis and

1977.
m Vi3 singular value
Circuits, Systems and
) for minimal design

16, pp. 399-403, 1980.
matrix for polynomial
'01, pp. 892-895, 1976.
for mnimal desgn
n Multivariable Tech.
, pp. 97-104, 1977.
lS
zation and the degree

, vol. CT-17, pp. 322
S154. Maronlas, 1., and S. Barnett, "Canonical forms of time-invariant linear

control systems: a survey with extension. PartI, single-input cases,"
Int. J. Systems Sci., vol. 9, pp. 497-514,1978; Part n, multivariable case,
vol. 10, pp. 33-50, 1979.
S155. Mayne, D. Q., "An elementary derivation of Rosenbrock's minimal
realization algorithm," IEEE Trans. Automatic Control, vol. AC-18,
pp. 306-307, 1973.
S156. Miminis, G. S., and C. e Paige, "An algorithm for pole assignmenl o[
time-invariant linear systems," Int. J. Control, vo 1. 35, pp. 341-354, 1982.
S157. - - , and - - , "An algorithm for pole-assignment of timc-invariant
Inuiti-input linear systems,' Froc. st IEEE -C'onI uecision ami LoniJOi,
pp. 62-67, 1982.
SI58. Mital, D. T., and C. T. Chen, "Canonical form realization of linear
time-invariant systems," Int. J. Control, vol. 18, pp. 881-887, 1973.
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
variable systems beyond closed. loop eigenvalue assignment," IEEE
Trans. Automatic Control, vol. AC-21, pp. 689-692, 1976: its {;omments,
voL AC-22, pp. 888-889 by Porterand D'Azzo, p. 889 by Gonrishankar,
1977.'
.
S161. - - , "Singular value analysis of linear systems," in Proc. 17th / EEE
C0171 Decision Control, pp. 66-73,Jan. 1979.
652
REFERENCES
S162. - - , "Principal component analysis in linear systems: Controllability,

observability and model reduction," IEEE 7)ans. Automatic Control,
vol. AC-26, pp. 17-32,1981.
S179. Porter, B., and

time tracking
1974.
S163. Morse, A. S., "Minimal solutions to transfer function equations," IEEE

7)ans. Automatic Control, vol. AC-21, pp. 131-133,1976.
S180. Pugh, A. c.,

pp. 315-324,
SI64. - - , W. A. Wolovich and B. D. O. Anderson, "Generic pole assignment:

Preliminary results," IEEE Trans. Automatic Control, vol. AC-28, pp.
503-506, 1983.
8181. Ralson, A., a

New York: IV
S165. - - , and W. M. Wonham, "Decoupling and pole assignment com

pensation," SIAM 1. Control, vol. 8, pp. 317-337,1970.
SI66. Owens, D. R, Multivariableand Optimal Systems. New York: Academic,
1981.
S167. Pace, 1. S., and S. Barnett, "Comparison of algorithms of calculations of
g.c.d. of polynomials," Int. J. Systems Sci., vol. 4, pp. 211-226, 1973.
S182. Rice, L Matl

McGraw-Hill
S183. Rissanen, J,
Control, vol. S
8184. Roman, J. R.
observers for
Trans. AutomG
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
S169. Paige, C. c., "Properties of numerical algorithms related to computing

controllability," IEEE 7)-ans. Automatic Control, vol. AC-26, pp. 130-138,
1981.
8187. - - , "The tn
vol. 25, pp. 11
S170. Patel, R. Y., "Computation ofmatrix fraction descriptions of linear time

invariant systems," IEEE Trans. Automatic Control, vol. AC-26, pp. 148
161,1981.
S17!. - - , "Computation of numerical-order state-space realizations and
observability indices using orthogonal transformation," Int. J. Control,
vol. 33, pp. 227-246,1981.
SIn. - - , and N. Munro, Multivariable System Theoryand Design.
York: Pergamon, 1982.
New
S173. Pearson, J. B., R. W. Shields, and P. W. States, Jr., "Robust solutions to

linear multivariable control problems," IEEE Trans. AuforYw(ic "~')!'i 'I}i
vol. AC-19, pp. 508-517, 1974.
S174. - - , and P. W. Staats, Jr., "Robust controllers for linear regulators,"
IEEE Trans. Automatic Control, vol. AC-19 (3), pp. 231-234,1974.
S175. Pernebo, L., "An algebraic theory for the design of controllers for linear
multivariable system. Part 1: Structure matrices and feedforward designo
Part 11: Feedback realizations and feedbackdesign, IEEE 7)-ans. Auto
matic Control, vol. AC-26, pp: 171-172, pp. 1~3-194, 1981.
S176. - - , and L. M. Silverman,:'Model reduction via balanced state space
representation," 1EEE Trans. Automatic Control, vol. AC-2l, pp. 382-387,
1982.
.
SI77.Popov, Y. M., "Invariant description oflinear time-in~aiiantcontrol1able
systeITIs," SI AM 1. Control, vol. 10, pp. 252-264, 1972.
S178. - - , Hyperstability
01 Control Systems.
Berlin: Springer-Verlag, 1973.
S186. - - , 'The zeI
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.
stems: Control1ability,
lS. Automat ic Control,
tion equations," IEEE
" 1976.
:neric pole assignment:
,ntrol, vo\. AC-28, pp.

pole assignment com
1970.
New York: Academic,
thms of ca1culations of
>p. 211-226, 1973.
r linear system calcula
po1ynomial matrices,"
; related to computing
)1. AC-26, pp. 130-138,
riptions of linear tme
1/, voL AC-26, pp. 148
pace realizations and
lUon," Int. J. Control,
)ry and Design.
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.
l balanced state space
,\. AC-21,pp. 382-387,
invariant co ntrollable
972.
Jringer-Verlag, 1973.
653
S179. Porter, B., and A. Bradshaw, "Design of linear multivariable continuous

time tracking systems," Int. J. Syste111s Sci., vo\. 5 (12), pp. 1155-1164,
1974.
S180. Pugh, A. c., "Transmission and system zeros," Int. 1. Contro!' vol. 26,
pp. 315-324, 1979.
S181\ Ralson, A., and P. Rabinowitz, A First Cow'se in Numerical Analysis.
New York: McGraw-Hil1, 1978.
S182. Rice, J., Matrix Computation and Mathematical Software.
McGraw-Hil1,1981.
New York:
S183. Rissanen, J., "Recursive identification of linear systems," SI AM J.

Control, vol. 9, pp. 420-430,1971.
SI84. Roman, J. R., and T. E. Bullock, "Design of minimum order stab1e
observers for linear functions of the state via realization theory," IEEE
S185. Rosenbrock, H. H., State-Space and Multivariable Theory. New York:
Wiley-Interscience, 1970.
S186. - - , "The zeros ofa system," Int. J. Control, vol. 18, pp. 297-299,1973.
S187. - - , "The transformation of strict system equiva1ues," Int. J. Control,
vol. 25, pp. 11-19,1977.
S188. - - , "Comments on 'Poles and zeros of linear multivariable systems:
A survey of the a1gebraic, geometric and complex variable theory,'"
Int. 1. Corarol, vol. 26, pp. 157-161, 1977.
S189. - - , and G. E. Hayton, "The general prob1em of po1e placement,"
Int.1. Control, vol. 27, pp, 837-852,1978.
S190. --', and A. J. 1. van der Weiden, "lnverse System," Int. J. Control, vol.
25, pp. 389-392, 1977.
S191. Rozsa, P., and N. K. Sinha, "Efficient algorithm for irreducible realizatan
of a rational matrix," Int. J. Control, vol. 20, pp. 739-748, 1974. lts
comments by Y. S. Kao and C. T. Chen, Int. 1. Control, vol. 28, pp. 325
326,1978.
SI92. Saeks, R., and 1. Murray, "Feedback system design: The tracking and
disturbance rejecton prob1ems," IEEE Tians. Automatic Control, vol.
AC-26, pp. 203-217, 1981.
S193. Sain, M. K., Introduction to Algebraic System Theory. New York:
Academic, 1981.
8194. Safonov, M. G., A. 1. Laub, and G. L. Hartmann, "Feedback properties
of multivariable systems. The role and use of the return diffetenee
matrix," IEEE Trans., vol. AC-26, pp. 47-65, 1981.
8195. Schumacher, J. M., "Compensator synthesis using (C, A, B)~p'airs,"
IEEE Trans. Automatic Control, vol. AC-25; pp. 1133-1137, 1980' .
S196. Seott, R. W., and B. D. O. Anderson, "Comments on 'Conditons fora
feedback transfer matrix to be proper,''' IEEE Tians. Automatic Control,
voI.AC-21, pp. 632-634,1976.
654
REFERENCES
8197. Shieh, L. S., and Y. T. Tsay, "Transformations of a class of multivariable

control systems to block companion forms," IEEE Trans. Automatic
Control, vol. AC-27, pp. 199-203,1982.
8198. Silverman, L., "Realization of linear dynamical systems," IEEE Tians.
8199. Smith, M. e, "Matrix fractions and strict system equivalence," Int. 1.
Control, vol. 34, pp. 869-884,1981.
8200. Stewart, G. W., Introduction to Matrix Computations.
Academic,1973.
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
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
655
class of multivariable
SEE Trans. Automatie
S216. Wolovich, W. A., "On determining the zeros of state-space systems,"

IEEE 7]-ans. Automatie Control, vol. AC-18 (5), pp. 542-544, 1973.
;ystems," IEEE Trans.
S217. - - , "Multivariable system synthesis with step disturbances rejection,"

IEEE 7]-ans. Automatie Control, vol. AC-19 (1), pp. 127-130,1974.
5218.
--o
Linear Multivariable Systems. New York: Springer-Verlag, 1974.
n equivalence," Int. J.
S219. - - , "Output feedback decoupling," IEEE 7]-ans. Automatie COl1trol,

vo!. AC-20, pp. 148-149,1975.
New York:
S220. - - , "Sket prime polynomial matrices," IEEE Trans. Automatie Control,

vol. AC-23, pp. 880-887, 1978.
S221. - - , "On the stabilization of closed-Ioop stabilizable systems," IEEE
7]-ans. Automatie Control, vol. AC-23, pp. 1103-1104, 1978.
ltations.
mpanion form realiza

eory," Ph.D. Disserta
S222. - - , "Multipurpose controllers for multivariable systems," IEEE y,-ans.

:ure problem in linear

.1. AC-26, pp. 111-129,
S223. --, P. Antsaklis, and H. Elliott, "On the stability of solutions to

minimal and nonminimal design problems," IEEE Trans. Automatie
Control, vol. AC-22, pp. 88-94, 1977.
.1gorithm for transfer

, pp. 1123-1133, 1981.
S224. - - , and P. Ferreira, "Output regulation and tracking in linear multi

variable systems," IEEE 7]-ans. Automatie Control, vol. AC-24, pp. 460
465, 1979.
fer function matrix to

AC-20, pp. 570-571,
Is, N.J.: Prentice-Hall,
:rconnected systems,"
.3-421, 1980.
S225. Wonham, W. M., "Tracking and regulation in linear multivariable

systems," SIAM J. Control, vol. 11, pp. 424-437,1973.
S226. - - , "Towards an abstract internal model principIe," IEEE y,-ans.
System, Man and Cybernetics, vol. SMC-6 (11), pp. 735-740, 1976.
,n of the characteristic
. Control, vol. AC-27,
8227.. --, Linear Multivariable Control: A Geornetrie Approaeh, 2d ed.

New York: Springer-Verlag, 1979.
S228. - - , and J. B. Pearson, "Regulation and internal stabilization in linear
multivariable systems," SIAM J Control, vol. 12, pp. 5--8,1974.
?ms, Memo. No. ERL
S229. Wu, M. Y., "A note on stability of linear time-varying systems," IEEE
Trans. Automatie Control, vol. AC-19, p. 162,1974.
thm for the design of
tomatie Control, vol.
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
matic Control, vol. AC-21, pp. 3-15, pp. 319-338, 1976.
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.
le division oftwo poly
1, vol. AC-28, pp. 238

; for arbitrary denomi
1. AC-28, pp. 518-521,
!ndex
cking and disturbance

'antrol, Dec. 1983.
tlative Index, vol. AC-
Additivity, 73, 88
Adjoinl eqllation, 163, 166, 195
AIgorilhm
ElIclidean, 579
Feedback implementation, 518, 520
Levenier, 67
poIe assignment, 476
row searching, 550
stale estimator, 359, 364
state feedback, 337, 347
trianglllarizalion, 589
Analytic conlinllalion, 554
Analytic fllncton, 78, 554
Asymplotic stability, 403, 409
Canonical-form dynamical equation
controllable, 243, 284, 327, 346
Jordan, 150,209
observable, 241, 286, 329
Causality, 76
Cayley-Hamilton theorem, 49, 67
Characteristic polynomial
of constant matrix, 33, 65
of ralional matrix, 235
Characterization, complete, 433
Cholesky decomposilion, 413
Colllmn degree coeftkient matrix, 601
Companion matrix, 26, 64, 346, 349
block, 223
unifGnn!y, 403
Asymptolic trackng problem, 340, 488,
495
Basis, 14
change of, 18,22
minimal polynomial, 608
orthonormal, 17, 90
polynomial, 607,621
Bezout identity, 595
Siack box, 72
Blockng zero, 499, 633
Bounded~input bOllnded-olllpllt" (BlBO)
.
stabiiy, 385, 394
.
Canonical decomposition lheorem, 203
eigenveclors of, 64, 65
Compensator equation, 460
Composite system, 433
Controllability, 176
dilTerential, 180
grammian, 184,322
index, 188
ndices, 190
instantaneous, 181
matrx, 184, 200
output, 214
output function, 216
Popov-Belevitch-Halltlls test, 231
unifonn, 182
657
658
INDEX
Coprimeness, 579
left,592
right,592
Convolution integral, 81
Cyclic matrix, 68, 230, 342
Dead beat control, 382
Decoupling,372
static, 502
Degree,577
column,600
row,600
Degree of rational malrix, 235
computation of, 315
Delta function, 74
Description of systern
differentia1 operator, 292
external, 3, 71
input-outpul, 71, 72
internal, 3, 71
polynomial matrix, 125, 287
state-variable, 71, 83
Diophantine equation, 460
Direct sum, 110
Discrete-time, systern, 121,301
dynarnical equation, 123,226
identification of, 300
Discretization of continuous-time equa
tion, 560
Disturbance signal, 488
additive,494
rejection of, 488,495
waveform strudured, 494
Divisor, 579
cornmon, 579
greatest com11on, 579, :>92
Iefl,592
right, 592
Domain, of function, 19
of linear operator, 27
Duality, theorern of, 195
Dynamical equation, 71, 87
computer simulat.ion of, 91
equivalent, 146
input-normal, 322
internally balanced, 222
lordan-form, 150,209
of RLC network, .101
output~normal, 322
reducible, 207
(Dynamical equation)
solution of, 134
Eigenvalue,33
assignment, 336
computation of, 61, 553
geometric multiplicity of, 41,66
index of, 47, 66
multiplicity or, 41, 66
Eigenvector,33
assignment of, 351
generalized, 39
grade of, 38
Elementary matrix, 542, 587
Elementary operation, 542, 587
Equilibrium state, 401
Equivalence, 146,200,296
in the sense of Lyapunov (i.s.L.), 153
strict system, 294
transrormation, 146, 151,200
zero-inpu 1, 148
zero-state, 148
Estimator, state, 281
exp At, 55, 56
computation of, 143
Laplace transform of, 56
properties or, 55
Exponential stability, 333
Feedback syslem, 324
constanl oulput, 377,445
inpllt-outpllt, 432
outpul, 343, 377, 379,474
plant inpul-output, 432
state, 324, 341, 344, 379
unity, 432, 457
wcll-posedness or, 118
Feedback gain malrix, 341, 348
Feld,8
of complex numbers, 9
of rational functions; 9
of real numbers, 9
Floquet, theory of, 154
Fraction,599
coprime, 605
irreducible, 605
Frobenius form, 64
block,223
Function, analytic 554
(Function)
integrable, 388
linear, 19
of matrix, 45
Function reproducbilt~
Function space, 10, 391
Functional estimator, 3
Fundamental cutset, \O:
Fundamental loop, \02
Fundamental rnatrjx, 13
Gaussian elimination, 61
with complete pivolin,
wth partal pivoling, (
Generalized eigenvector,
chain of, 39
grade of, 38
Gram determinant, 172,
Grammian, controlIabili
observability, 197
Hankel matrix, 245, 257,
Hermite form, 589
Hermitian matrix, 565
nonnegative definite, 4
positive definite, 413
positive semidefinite, 4
Hessenberg form, 221,55

Homogenety, 73,88
Householder transformat
wilh pivoting, 545
Hurwitz polynomial, 395
HUf\N!.Z sta b!1it'l !~;:::1 ~9~
Identification, determnisl
Impulse function, 74
Impulse response, 75
matrix,76
Inner product, 59
Input-output pair, 87
Internal ruodel principie, '
Inverse problem, 217, 529
Irreducible dynamical eql
irreducible polynomal.1
tion,291
Irreducible realizaton, 20
INDEX
ation)
14
36
)f, 61, 553
ltiplicity of, 41, 66
',41,66
(Function)
integrable, 388
lnear, 19
of matrix, 45
Function reproducibility, 216
Function space, 10, 391
Functonal estmator, 369
Fundamental cutset, 102
Fundamental loop, 102
Fundamental matrix, 136
351
'ix, 542, 587
alion, 542,587
~, 401
,200,296
Lyapunov (i.s.L.), 153
94
1, 146, 151,200
Gaussian elimination, 61, 543
with complete pivoting, 544
with partial pivoling, 61, 544
Generalized eigenvector, 38
chain of, 39
grade of, 38
Gram determinant, 172, 184
Grammian, controllability, 184
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
Hankel matrix, 245, 257, 265
Hermite form, 589"
Hermitian matrix, 565
nonnegative definite, 413
positive definite, 413
positive semidellnite, 413
Hessenberg form, 221, 552
Homogeneity, 73, 88
Householder transformation, 61,544
with pivoting, 545
Hurwitz polynomial, 395
Hurwitz stabilitv tesl, 39P.
Identification, deterministic, 302
Impulse funclion, 74
Impulse response, 75
matrix,76
Inner product, 59
Input-output pair, 87
Internal model principIe, 490
Inverse problem, 217, 529
Irreducible dynamical equalion, 207
Irreducible polynomial matrix dcscri p

tion,291
Irreducible realization, 208, 233
659
J ordan block, 38, 49
Jordan-canonical-form matrix, 37,41,53,
60,209
L p sta bility, 390
Laplace transform, 56, 81
Lcverrier algorithm, 67
Linear algebraic equations, 26
Linear independence
of time functons, 170
of vectors, 12, 13, 40, 41
Linear mapping, 20
Linear operator, 20
matrix rcpresentation of, 21
change of basis, 22
nullity, 29, 31
null space of, 29
range space of, 27
Linear space,9
basis of, 14
dimension of, 13
Linear system, 73, 87
Linear time-invariant system, 80, 89
Linearity, 20,73,87
Link, 102
Lyapunov equation, 572
Lyapunov function, 415
Lyapunov theorem, 414, 421
Lyapunov transformation, 153,428
Markov paral1lelerS, 245, 301
Mason's gain formula, 158
Matrix, 7, 24
;Cf~\i('.,
6P., 230. 3l2
echelon form, 307, 589
elementary, 542, 587
function of, 51
fundamental, 136
Hermite, 589
hermitian,412
nonsingular, 28, 32
norm of, 57
orthogonal, 220, 570
polynomial, 587
polynomial of, 45
" principal minar, 413
leading,413
rank or, 27
singular value of, 568
660
INDEX
(Matrix)
stale transition, 137
symmetric, 412
Sylvester, 582
system, 293, 439
trace of, 67,129
unitary, 220
Minimal design problem, 529

Minimal-dimensional realization, 233
Minimal polynomial, 46
Minimum-energy control, 178,558
Mode,145
Model,2
matching, 529
simplified,271
Module, 588, 607
Monic polynomial, 46, 577
Moment,301
Nilpotent matrix, 69
Nonanticipalive system, 76
Nonlinear system, 73, 87
Norm, of matrix, 57
Euclidean, 57
induct:d,57
of vector, 57
Normal trce, 101
Nullity, 29
Null space, 29, 284
left, 29, 69
right,29
Observability, 192, 197

direrential, 196
grammian, 197, 322
index,199
indices, 199, 264
instantaneous, 197
matrix, 198
uniform, 197
Observer (see State estimator)

Open-loop control system, 324
Orthogonal matrix, 570
Orthogonality of vectors, 566
Orthonormal set, 17,566
Output controllability, 214
Output equation, 87, 89
Output function controllability, 216
Parameter perturbation, 491

Parameterization, 540
Periodically varying system, 153
Physica\ system, 1
Pole, 441, 623
Pole-zero cancellation, 437,443,444
unstable, 444, 459
Pole-zero excess inequality, 459, 487
Polynomial matrix, .587
column reduced, 601
echelon form, 306, 612
Hermite form, 589
Popov form, 612
row reduced, 601
unimodular, 591
Popov-Belevich-Hautus test, 231

Positive definite matrix, 171, 182, 184
Positive semidefinite matrix, 413
Predictable degree properly, 541
Primary dependenl row, 247, 257, 262,
272, 306, 610
Principie minor, 413
leading, 413
Proper rational matrix, 82
characteristic polynomial of, 235
column index of, 468,618 .
cyclic, 469
degree of, 235, 315
row index of, 466, 613
Pseudoinverse, 270
Psel1dostate, 288, 293, 295
Pulse funclion, 74
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 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,
Sampled transfer funclion
QR algorithm, 553
QR faclorizatioD, 553
Ql1adratic form, 411.
~;<-iliJlillb i:::'ldlHS, .JjSJ
Range
of function, 19
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
Sampling period, 121,559

scalar,8
Sca\ar producl, 59
Schmidt orthonormalizat
Schwarz inequality, 59, 69
Separation property, 367
Sequence, input, 121
output, 121
persisten tIy exciting, 30' .
weighling, 121
z transform of, 122
Servocompensator, 504
Shifting property; 80, 89
Similar matrices, 23
INDEX
Jrbation, 491
!n,540
'ying system, 153
,1
lIation,437, 443,444
.459
; ineq uality, 459,487
:rix, 587
:ed,601
306,612
,589
512
601
i91
-Hautus test, 231
matrix, 171,182,184
'Inite matrix, 4 I3
'ee property, 541
lent row, 247, 257, 262,
O
413
matrix,82
polynomial of, 235
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
internally balanced, 271, 322
irreducible, 208
minimal-dimensional, 233
of time-varyng differential equation,
252
output-normal,322
partal, 302
reducible, 207
Reachability, 176,227
Regulator problem, 340
Relaxedness, 77, 79, 161
Representation
of linear operator, 2 I
of vector, 15
Response, 72
impulse, 75, 126
steady-state, 494, 501
step, 107, 129
transient, 494, 501
zero-input,89, 140
zero-state, 89, 140
Resultanl,582
generalized, 608
Riccati equation, 68
Ring, 9, 588
Robusl design, 491, 501
Routh-Hurwitz criterion, 398,417
Routh table, 397
Row degree coefficient matrix., 601
Row searching algorithm, 247, 257,272
Sampled transfer functon, 122
matrix, 123
S~nllplii1g insiarn.s,
550
Sampling perod, 121,559
scalar,8
Scalar product, 59
Schmdt orthonormalization, 567
Schwarz inequality, 59, 69
Separation property, 367
Sequence, input, 121
output, 121
persistently exciting, 304, 308
weightng, 121
z transform of, 122
Servocompensator, 504
Shifting property, 80, 89
Similar matrices, 23
661
Similarity transformation, 23, 60
Simulation ofdynamical equation, 91
analog computer, 92
digital computer, 93
Singular value decomposition, 61,269,569

Solution space, 135
Stability, asymptotic, 403, 409
bounded-input bounded-output, 385
exponential, 405, 409
global, 403
in the sense of Lyapunov (i.s.L.), 402
total, 407, 411
uniform, 402, 408
Stabilization, 339,490
Statc, 83 -86
definiton of, 83
equilibrium,401
initial,89
pseudo-, 288
variable, 86
vector, 86
State equation, 87, 89, 134
State estmalor, 354
algorithm of, 359
asymptolic, 356
full-dimensiollal,355
fUllctional, 369
open-Ioop,355
reduced-dimensiollal,361
State feedback, 334
Slate space, 90
State transition matrix, 137
Subspace, 11,27
Superposition principie, 73
Sylvester's inequality, 31
Svivestcr nl.rJlrix; SR2

Syslem,
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
Time invariance, 80, 89
Time varying, 80,89
662
INDEX
Total stability (T-stability), 407, 411
Trace of matrix, 67, 129
Transfer function, 82
matrix, 82, 90
sampled, 122
Transformation, 143
elementary, 542, 587
equivalence, 146, 151,200
Householder, 220, 544
Lyapunov, 153
orthogonal, 220,222
Transmission blocking property, 626, 629
Transmission zero, 443, 496
Tree,102
branch,102
normal, 102
Truncalion operator, 125
Uniform controllability, 182
Uniform observability, 197
Unimodular matrix, 591
Uniqueness of solution of differential
equation, 87, 89
Unitary matrix, 269, 568
Unit-time-delay system, 77, 82, 86
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
see also Linear space
Well posedness, 118
Weighling sequence, 121
Zero,623
blocking, 499,633
input-decoupling,292
output-decoupling, 292
transmission, 443, 627
Zero-input response, 89, 140
Zero-memory system, 72
Zero-state equivalence, 148
Zero-state response, 89, 140
z-transfer function, 122
z transform, 122
66
1
or, 1S
13
r space
118
once, 121
633
ng, 292
.ling,292
143,627
me, 89, 140
,tem,72
ilence, 148
lse, 89, 140
-n,122
l
:j
------_._._---'"_._--"----~._-~-----------_._-------~----~----_._---~--_._--------------~---------------------------~--'"._-----

Linear System Theory and Desing PDF

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Linear System Theory and Desing PDF

Transféré par

Droits d'auteur :

Formats disponibles

Linear System Theory

M. E. Van Valkenburg, Series Editor

Linear System Theory

This text is a major revision of IntroduClion lO Linear

Library of Congress Cataloging in Publication Data

2. System designo 1. Title.

*May be omitted without 10ss (

1-1 The Study of Systems

Linear Spaces and Linear Operators

Linear Spaces over a Field

Linear Independence, Bases, and Representations

Linear Operators and Their Representations

Systems ofLinear Algebraic Equations

Eigenvectors, Generalized Eigenvectors, and Jordan-Form

Representations of a Linear Operator

*Derivation of a Jordan-Form Representation

Functions of a Square Matrix

Polynomials of a Square Matrix

Functions of a Square Matrix

*May be omitted without loss of continuity.

*2-8 Nonns and Inner Product

2-9 Concluding Remarks

Mathematical Descriptions of Systems

3-2 The Input-Output Description

3-3 The State-Variable Description

The Concept of State

Analog and Digital Computer Simulations ofLinear Dyna

*Dynamical Equations for RLC Networks

3-6 Mathematical Descriptions of Composite Systems

*3-7 Discrete-Time Systems

3-8 Concluding Remarks

linear Dynamical Equations and Impulse-Response

4-2 Solutions of a Dynamical Equation

Solutions of the Dynamical Equation E

4-3 Equivalent Dynamical Eqations

Linear Time-Varying Dynamical Equation with

IUlations of Linear Dyna

Impulse-Response Matrices and Dynamical

Controllability and Observability 01 linear Dynamical

Irreducible Realizations, Strict System Equivalence, and

lical Equation with

Realization from the Hankel Matrix

*Realization of Linear Time-Varying Differential

Realizations of Vector Proper Rational Transfer

Realization from the Hankel Matrix

Irreducible Realizations of Proper Rational Matrices: Hankel

Method 1. Singular Value Decomposition

Realization ofN(s)D-1(s), Where D(s) and N(s) Are Not

Polynomial Matrix Description

Strict System Equivalence

Identification of Discrete-Time Systems from Noise-Free

Nonzero Initial Conditions

State Feedback and State Estimators

Canonical-Form Dynamical Equations

Effect on the Numerator of g(s)

Asymptotic Tracking Problem-Nonzero Set

Full-Dmensional State Estimator

*7-6 Decoupling by State Feedback

7-7 Concluding Remarks

tional Matrices: Hankel

I~e D(s) and N(s) Are Not

Stability of Linear Systems

8-2 Stability Criteria in Terms of the Input-Output