THE OXFORD SERIES IN ELECTRICAL AND COMPUTER ENGINEERING
Adel S. Sedra, Series Editor, Electrical Engineering
Michael R. Lightner, Series Editor, Computer Engineering
Allen and Holberg, CMOS Analog Circuit Design
Bobrow, Elementary Linear Circuit Analysis, 2»d Ed.
Bobrow, Fundamentals of Electrical Engineering, 2nd Ed.
Campbell, The Science and Engineering of Microelectroinic Fabrication
Chen, Analog and Digital Contrl System Design
Chen, Linear System Theor and Design, 3rd Ed.
Chen, System and Signal Analysis, 2nd Ed.
Comer, Digital Logic and State Machine Design, 3rd Ed.
Cooper and McGillem, Probabilistic Methods of Signal and System Analysis, 3rd Ed.
Fortney, Principles of Electronics: Analog & Digital
Franco, Electric Circuits Fundamentals
Granzow, Digital Transmission Lines
Guru and Hiziroglu, Electric Machiner & Transforers, 2nd Ed.
Hoole and Hoole, A Modem Short Course in Engineering Electromagnetics
Jones, Introduction to Optical Fiber Communication Systems
Krein, Elements of Power Electronics
Kuo, Digital Contrl Systems, 3rd Ed.
Lathi, Modem Digital and Analog Communications Systems, 3rd Ed.
McGillem and Cooper, Continuous and Discrete Signal and System Analysis, 3rd Ed
Miner, Lines and Electromagnetic Fields for Engineers
Roberts and Sedra, SPICE, 2nd Ed.
Roulston, An Introduction to the Physics of Semiconductor Devices
Sadiku, Elements of Electromagnetics, 2nd Ed.
Santina, Stubberud, and Hostetter, Digital Contrl System Design, 2nd Ed.
Schwarz, Electromagnetics for Engineers
Schwarz and Oldham, Electrical Engineen'ng: An Introduction, 2nd Ed.
Sedra and Smith. Microelectronic Circuits. 4th Ed.
Stefani, Savant. Shahian. and Hostetter, Design of Feedback Control Systems, 3rd Ed.
Van Valkenburg. Analog Filter Design
Waer and Grung. Semiconductor Device Electronics
Wolovich. Automatic Control Systems
Yariv. Optical Electronics in Modem Communications, 5th Ed.
.llL/Ñ bYbLí
L¯ÑY /lLLLol<l
OU cUlIC
LHIÅSDBP LH0H
State University of New York at Stony Brook
New York Oxford
OXFORD UNIVERSIT PRESS
1 999
L?ÍLHUNIVERIT PRSS
Oxford New York
Athens Auckland Bangkok Bogota Buenos Aires Calcutta
Cape Town Chennai Dar es Salaam Delhi Florence Hong Kong Istanbul
Karachi Kuala Lumpur Madrid Melboure Mexico City Mumbai
Nairobi Paris Sao Paulo Singapore Taipi Tokyo Toronto Warsaw
and associated companies in
Berlin Ibadan
Copyright ß1999 by Oxford University Press. Inc.
Published by Oxford University Pr,s, Inc.
198 Madison Avenue, New York, New York 10016
Oxford is a registered trademark of Oxford University Press
All rights reserved. No part of this publication may D
reproduced, stored in a retrieval system, or tansmitted,
in any fan or by any means, electonic, mechanical,
photocopying, recording, or otherwise, without the prior
prmission of Oxford University Press.
Library of CongresCataloginginPblication Data
Chen, ChiTsong
Linear systemtheory and design Íby Chi Tsong Chen. 3rd ed.
p. cm. (The Oxford series in electical and computer engineering)
Includes bibliographical references and index.
ISBN 0195117778 (cloth).
I. Linear systems. 2. System design. Ì. Title. II. Series.
QA402.C44 1998
629.8'32<c21 9735535
Printing (last digit): 9 8 7 6 5 4 3 2 Ì
Printed in the United States of America
on acidfree paper
CI
To
BlHJAU
\COÍCOÍS
Preface xi
Chapter 1: Introduction
1.1 Introduction
1 .2 Overview 2
Chapter 2: Mathematical Descriptions of Systems 5
2.1 Introduction 5
2.1.1 Causaliit and Lumpedness Ù
2.2 Linear Systems 7
2.3 Linear TimeInvariant (LTI) Systems 11
2.3.1 OpAmp Circuit Implementation 1Ù
2.4 Linearization 17
2.5 Examples 18
2.5.1 RLC Networks ZÙ
2.6 DiscreteTime Systems 3 1
2.7 Concluding Remarks 37
Problems 38
Chapter 3: Li near Al gebra 44
3.1 Introduction 44
3.2 Basis, Representation, and Orthonormalization 45
3.3 Linear Algebraic Equations 48
3.4 Similarit Transformation 53
3.5 Diagonal Form and Jordan Form 55
3.6 Functions of a Square Matrix 61
3.7 Lyapunov Equation 70
3.8 Some Useful Formulas 71
3.9 Quadratic Form and Positive Defniteness 73
3.1 0 SingularValue Decomposition 76
3.11 Norms of Matrices 78
Problems 78
vii
viii CONTENTS
Chapter 4: StateSpace Sol uti ons and Real izations 86
4. 1 Introduction 86
4.2 Solution of LTI State Equations 87
4.2.1 Discretization VU
4.2.2 Solution of DiscreteTime Equations VZ
4.3 Equivalent State Equations 93
4.3.1 Canonical Forms V´
4.3.2 Magnitude Scaling in OpAmp Circuits VO
4.4 Realizations 700
4.5 Solution of Linear TimeVarying (LT) Equations 706
4.5.1 DiscreteTime Case !!U
4.6 Equivalent TimeVaring Equations 777
4.7 TimeVarying Realizations 77 5
Problems 77 7
I
Î
Chapter 5: Stability 121
5.1 Introduction 727
5.2 InputOutput Stability of LTI Systems 721
5.2.1 DiscreteTime Case !ZO
5.3 Internal Stability 729
5.3.1 DiscreteTime Case !J!
5.4 Lyapunov Theorem 732
5.4. 1 DiscreteTime Case !JJ
5.5 Stabilit of LTV Systems 737
Problems 740
Chapter 6: Control l abi lity and Observabi l ity 143
6.1 Introduction 743
6.2 Controllability 744
6.2.1 Controllablity Indices !JU
6.3 Observability 753
6.3.1 Observability Indices !J´
6.4 Canonical Decomposition 758
6.5 Conditions in JordanForm Equations 764
6.6 DiscreteTime State Equations 169
6.6.1 Controllabilit to the Origin and Reachability Ì¯!
6.7 Controllability After Sampling 772
6.8 LT State Equations 176
Problems 180
Chapter 7: Mi nimal Real izations and Coprime Fracti ons 184
7.1 Introduction 184
7.2 Implications of Coprimeness 185
7.2.1 Minimal Realizations !OV
7.3 Computing Coprime Fractions 192
7.3.1 QR Decomposition !VD
7.4 Balanced Realization 197
7.5 Realizations from Markov Parameters 200
7.6 Degree of Transfer Matrices 205
7.7 Minimal RealizationsMatrix Case 207
7.8 Matrix Polynomial Fractions 209
7.8.1 Column and Row Reducedness Z!Z
7.8.2 Computing Matrix Coprime Fractions Z!¬
Contents
7.9 Realizations from Matrix Coprime Fractions 220
7.10 Realizations from Matrix Markov Parameters 225
7.11 Concluding Remarks 227
Problems 228
Chapter 8: State Feedback and State Estimators 231
8.1 Introduction 231
8.2 State Feedback 232
8.2.1 Solving the Lyapunov Equation ZJV
8.3 Regulation and Tracking 242
8.3.1 Robust Tracking and Disturbance Rejection Z¬J
8.3.2 Stabilization Z+¯
8.4 State Estimator 247
8.4.1 ReducedDimensional State Estimator ZJ!
8.5 Feedback from Estimated States 253
8.6 State FeedbackMultivariable Case 255
8.6.1 Cyclic Design ZJO
8.6.2 LyapunovEquation Method ZJV
8.6.3 CanonicalForm Method ZOU
8.6.4 Effect on Transfer Matrices ZOZ
8,7 State estimatorsMultivariable Case 263
8.8 Feedback from Estimated StatesMultivariable Case 265
Problems 266
Chapter 9: Pol e Placement and Model Matchi ng 269
9.1 Introduction 269
9,1.1 Compensator EquationsClassical Method Z¯!
9.2 UnityFeedback ConfgurationPole Placement 273
9,2.1 Regulation and Tracking Z´J
9.2.2 Robust Tracking and Disturbance Rejection Z¯¯
9.2.3 Embedding Internal Models ZOU
9.3 Implementable Transfer Functions 283
9.3.1 Model MatchingTwoParameter Confguration ZOO
9.3.2 Implementation of TwoParameter Compensators ZV!
9.4 Multivariable UnityFeedback Systems 292
9.4.1 Regulation and Tracking JUZ
9.4.2 Robust Tracking and Disturbance Rejection JÕJ
9.5 Multivariable Model MatchingTwoParameter
Confguration 306
ix
x CONTENTS
9,5,1 Decoupling JJ1
9,6 Concluding Remarks 314
Problems 315
References 319
Answers to Selected Problems 321
Index 331
CÍÒCC
This text is intended for use in senior/frstyear graduate courses on linear systems and
multi variable system design in electrical, mechanical, chemical, and aeronautical departments,
It may also be useful to practicing engineers because it contains many design procedures, The
mathematical background assumed is a working knowledge of linear algebra and the Laplace
transform and an elementary knowledge of differential equations,
Linear system theory is a vast feld, In this text, we limit our discussion to the conventional
approaches of statespace equations and the polynomial fraction method of transfer matrices.
The geometrc approach, the abstract algebraic approach. rational fractions, and optimization
are not discussed.
We aim to achieve two objectives with this text. The frst objective is to use simple
and effcient methods to develop results and design procedures, Thus the presentation is
not exhaustive. For example, in introducing polynomial fractions, some polynomial theory
such as the SmithMcMillan form and Bezout identities are not discussed. The second
objective of this text is to enable the reader to employ the results to carry out design,
Thus most results are discussed with an eye toward numerical computation. All design
procedures in the text can be carried out using any software package that includes singular
value decomposition. QR decomposition, and the solution of linear algebraic equations and
the Lyapunov equation. We give examples using MATLAB®, as the packagel seems to be the
most widely available.
This edition is a complete rewriting of the book Linear System Theory and Design, which
was the expanded edition of Introduction to Linear System Theor published in 1970. Aside
from, hopefully, a clearer presentation and a more logical development, this edition difers
from the book in many ways:
• The order of Chapters 2 and 3 is reversed, In this edition. we develop mathematical
descriptions of systems before reviewing linear algebra, The chapter on stability is moved
earlier,
• This edition deals only with real numbers and foregoes the concept of felds. Thus it is
mathematically less abstract than the original book. However, all results are still stated as
theorems for easy reference.
• In Chapters 4 through 6, we discuss frst the timeinvariant case and then extend it to the
timevarying case. instead of the other way around,
l. NPÏ1PÜ is a registered trademark of the MathWorks.lnc,. 24 Prime Park Way, Natick, MA U1¯ÓU1JUU Phone:
JUbÔ4¯¯UUU. fa: 3UbÔ4¯·¯Jl . Email: info@mathworks.com, http://www,mathworks,com,
xi
xii PREFACE
• The discussion of discretetime systems is expanded.
• In statespace design, Lyapunov equations are employed extensively and multi variable
canonical forms are downplayed. This approach i s not only easier for classroom presentation
but also provides an attractive method for numerical computation.
• The presentation of the polynomial fraction method is streamlined. The method is equated
with solving linear algebraic equations. We then discuss pole placement using a onedegree
offreedom confguration, and model matching using a twodegreeoffreedom confgura
tion.
• Examples using MATLAB are given throughout this new edition.
This edition i s geared more for classroom uSe and engineering applications; therefore, many
topics in the original book are deleted, includi�g strict system equivalence, deterministic iden
tifcation, computational issues, some multi variable canonical forms, and decoupling by state
feedback. The polynomial fraction design in the input/output feedback (controller/estimator)
confguration is deleted. Instead we discuss design in the twoparameter confguration. This
confguration seems to be more suitable for practical application. The eight appendices in the
original book are either incorporated into the text or deleted.
The logical sequence of all chapters i s as follows:
Chap. 8
Chap. 15 ~
Chap. 6 ~
Chap. 7
Sec. 7. 17.3 ~ Sec. 9.19.3
 Sec. 7.67.8 ~ Sec. 9.49.5
In addition, the material in Section 7. 9 i s needed to study Section 8.6.4. However, Section 8.6.4
may be skipped without loss of continuity. Furthermore, the concepts of controllability and
observability in Chapter 6 are useful, but not essential for the material i n Chapter 7. All minimal
realizations in Chapter 7 can be checked using the concept of degrees, instead of checking
controllability and observability. Therefore Chapters 6 and 7 are essentially independent.
This text provides more than enough material for a onesemester course. A onesemester
course at Stony Brook covers Chapters 1 through 6, Sections 8. 18.5, 7. 17.2, and 9. 19.3.
Timevarying systems are not covered. Clearly, other arrangements are also possible for a
onesemester course. A solutions manual is available from the publisher.
I am indebted to many people in revising this text. Professor Imin Kao and Mr. Juan
Ochoa helped me with MATLAB. Professor Zongli Lin and Mr. T. Anantakrishnan read
the whole manuscript and made many valuable suggestions. I am grateful to Dean Yacov
Shamash, College of Engineering and Applied Sciences, SUNY at Stony Brook, for his
encouragement. The revised manuscript was reviewed by Professor Harold Broberg, EET
Department, Indiana Purdue University; Professor Peyman Givi, Department of Mechanical
and Aerospace Engineering, State University of New York at Bufalo; Professor Mustafa
Khammash, Department of Electrical and Computer Engineering, Iowa State University; and
Professor B. Ross Barmish, Department of Electrical and Computer Engineering, University
of Wisconsin. Their detailed and critical comments prompted me to restructure some sections
and to include a number of mechanical vibration problems. I thank them all.
Preface xiii
I am indebted to Mr. Bill Zobrist of Oxford University Press who persuaded me to
undertake this revision. The people at Oxford University Press, including Krysia Bebick,
Jasmine Urmeneta, Terri O
'
Prey, and Kristina Della Barolomea were most helpful in this
undertaking. Finally, I thank my wife, BihJau, for her support during this revision.
ChiTsong Chen
LINEAR SYSTEM
THEORY AND DESIGN
LH3D!CI
I ntroducti on
1 , 1 Introduction
The study and design of physical systems can be carried out using empirical methods. We can
apply various signals to a physical system and measure its responses. If the performance is not
satisfactory, we can adjust some of its parameters or connect to i t a compensator to impr
o
ve
its perforance. This approach r
�
lies heavily on past experience and is carried out by trial and
error and has succeeded in designing many physical systems.
Empirical methods may become unworkable if physical systems are complex or too
expensive or too dangerous to be experimented on. In these cases, analytical methods become
indispensable. The analytical study of physical systems consists of four parts: modeling,
development of mathematical descriptions, analysis, and design. We briefy introduce each
of these tasks.
The distinction between physical systems and models is basic in engineering. For example,
circuits or control systems studied in any textbook are models of physical systems. A resistor
with a constant resistance is a model: it will bur out if the applied voltage is over a limit.
This power limitation is often disregarded in its analytical study. An inductor with a constant
inductance is again a model: in reality, the inductance may vary with the amount of current
fowing through it. Modeling is a very important problem, for the success of the design depends
on whether the physical system is modeled properly.
A physical system may have diferent models depending on the questions asked. It may
also be modeled differently in diferent operational ranges. For example, an electronic amplifer
is modeled differently at high and low frequencies. A spaceship can be modeled as a particle
in investigating its trajectory: however, it must be modeled as a rigid body in maneuvering.
A spaceship may even be modeled as a fexible body when it is connected to a space station.
In order to develop a suitable model for a physical system, a thorough understanding of the
physical system and its operational range is essential. In this text, we will call a model of a
physical system simply a system. Thus 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 (or model) is selected for a physical system, the next step is to apply
various physical laws to develop mathematical equations to describe the system. For ex
ample, we apply Kirchhof's voltage and current laws to electrical systems and Newton's
law to mechanical systems. The equations that describe systems may assume many forms:
2 I NTRODUCTION
they may be linear equations, nonlinear equations, integral equations, difference equations,
differential equations, or others. Depending on the problem under study, one form of equation
may be preferable to another in describing the same system. In conclusion, a system may
have different mathematicalequation descriptions just as a physical system may have many
different models.
After a mathematical description is obtained, we then carry out analysesquantitative
andor qualitative. In quantitative analysis, we are interested in the responses of systems
excited by certain inputs. In qualitative analysis, we are interested in the general properties
of systems, such as stability, controllability, and observability. Qualitative analysis is very
important, because design techniques may often evolve from this study.
If the response 0f 3 system is unsatisfactory, the system must be modifed. In some
cases, this can be achieved by adjusting some parameters of the system; in other cases,
compensators must be introduced. Note th�the design is carried out on the model of the
physical system. If the model is properly chosen, then the performance of the physical system
should be improved by introducing the required adjustments or compensators. If the model is
poor, then the performance of the physical system may not improve and the design is useless.
Selecting a model that is close enough to a physical system and yet simple enough to be studied
analytically is the most diffcult and important problem in system design.
1 , 2 Overview
The study of systems consists of four parts: modeling, setting up mathematical equations,
analysis, and design. Developing models for physical systems requires knowledge of the
particular feld and some measuring devices. For example, to develop models for transistors
requires a knowledge of quantum physics and some laboratory setup. Developing models
for automobile suspension systems requires actual testing and measurements; it cannot be
achieved by use of pencil and paper. Computer simulation certainly helps but cannot replace
actual measurements. Thus the modeling problem should be studied in connection with the
specifc feld and cannot be properly covered in this text. In this text, we shall assume that
models of physical systems are available to us.
The systems to be studied in this text are limited to linear systems. Using the concept of
linearity, we develop in Chapter 2 that every linear system can be described by
,=
.
'0
( 1 . 1 )
This equation describes the relationship between the input .and output ,and is called the
inputoutput or ...description. If a linear system is lumped as well, then it can also be
described by
.= ·.+ .
,= ..+ .
( 1 .2)
( 1 .3)
Equation (1.2) is a set of frstorder differential equations and Equation (1.3) is a set of algebraic
equations. They d called the ....description of linear systems. Because the vector .is
called the ..the set of two equations is called the ....or, simply, the ..equation.
1.2 Overview 3
If a linear system has, in addition, the property of time invariance, then Equations ( 1. 1 )
through ( 1.3) reduce to
and
,= .
. = ·.+ .
, = ..+ .
( l ..l )
1 1 .5)
( 1.6)
For this class of linear timeinvariant systems, the Laplace transform is an important tool in
analysis and design. Applying the Laplace transform to ( 1.4) yields
y<s) = G(s)u(s) ( 1.7)
where a variable with a circumfex denotes the Laplace transform of the variable. The function
G(s) is called the .·.
Both ( 1 . 4) and ( 1 . 7) are inputoutput or exteral descriptions.
The former is said to be in the time domain and the latter in the frequency domian.
Equations (l.l) through ( 1 .6) are called continuoustime equations because their time
variable is a continuum defned at every time instant in (
.
C. :). If the time is derined
only at discrete instants, then the corresponding equations are called discretetime equations.
This text is devoted to the analysis and design centered around (1. 1 ) through ( 1 .7) and their
discretetime counterparts.
We briefy discuss the contents of each chapter. In Chapter 2. after introducing the
aforementioned equations from the concepts of lumpedness, linearity. and time invariance.
we show how these equations can be developed to describe systems. Chapter 3 reviews linear
algebraic equations, the Lyapunov equation, and other pertinent topics that are essential for
this text. We also introduce the Jordan form because it will be used to establish a number of
results. We study in Chapter 4 solutions of the statespace equations in ( 1 . 2) and ( 1 . 5). Different
analyses may lead to diferent state equations tliat describe the same system. Thus we introduce
the concept of equivalent state equations. The basic relationship between statespace equations
and transfer matrices is also established. Chapter 5 introduces the concepts of boundedinput
boundedoutput (BlBO) stability, marginal stability, and asymptotic stability. Every s: stem
must be designed to be stable; otherwise, it may bum out or disintegrate. Therefore stability
is a basic system concept. We also introduce the Lyapunov theorem to check asymptotic
stability.
Chapter 6 introduces the concepts of controllability and observability. They are essential
in studying the interal structure of systems. A fundamental result is that the transfer matrix
describes only the controllable and observable part of a state equation. Chapter 7 studies
minimal realizations and introduces coprime polynomial fractions. We show how to obtain
coprime fractions by solving sets of linear algebraic equations. The equivalence of controllable
and observable state equations and coprime polynomial fractions is established.
The last two chapters discuss the design of timeinvariant systems. We use controllable
and observable state equations to carry out design in Chapter 8 and use coprime polynomial
fractions in Chapter 9. We show that, under the controllability condition, all eigem'alues
of a system can be arbitrarily assigned by introducing state feedback. If a state equation
is observable, fulldimensional and reduceddimensional state estimators, with any desired
4 INTRODUCTION
eigenvalues, can be constructed to generate estimates of the state. We also establish the
separation property. In Chapter 9, we discuss pole placement, model matching, and their
applications in tracking, disturbance rejection, and decoupling. We use the unityfeedback
confguration in pole placement and the twoparameter confguration in model matching. In our
design, no control performances such as rise time, settling time, and overshoot are considered:
neither are constraints on control signals and on the degree of compensators. Therefore this
is not a control text per se. However, all results are basic and useful in designing linear time
invariant control systems.
2. 1 Introduction
LH3D!CI
�
�
Mathemati cal
Descri pti ons of Systems
The class of systems studied i n this text i s assumed to have some input terminals and output
terminals as shown in Fig. 2. 1 . We assume that if an excitation or input is applied to the input
terminals, a unique response or output signal can be measured at the output terminals. This
unique relationship between the excitation and response, input and output, or cause and effect
is essential in defning a system. A system with only one input terinal and only one output
terminal is called a singlevariable system or a singleinput singleoutput (SISO) system.
A system with two or more input terminals and/or two or more output terminals is called
a multi variable system. More specifcally, we can call a system a multiinput multioutput
(MIMO) system if it has two or more input terminals and output terminals, a singleinput
multioutput (SIMO) system if it has one input terminal and two or more output terminals.
UII)
�
Ulf)
U
�
['J
4 5
o l 2 3 k
u[k)
Figure 2.1 System.
Black
box
y(l)
y[k)
�
y[k)
o 1 234 5k
5
6 MATHEMATICAL DESCRI PTI ONS OF SYSTEMS
A system is called a continuoustime system if it accepts continuoustime signals as
its input and generates continuoustime signals as its output. The input will be denoted bv
lowercase italic II (t) for single input or by boldface n(t) for multiple inputs. If the system h;s
p input terminals, then n(t) is a p x 1 vector or n = [Ill 12
. .
, u
p
]' , where the prime denotes
the transpose. Similarly, the output will be denoted by y(t) or yet). The time t is assumed to
range from 0 to O.
A system is called a discretetime system if it accepts discretetime signals as its input
and generates discretetime signals as its output. All discretetime signals in a system will
be assumed to have the same sampling period T. The input and output will be denoted by
u[k] := lI(kT) and y[k] := y( kT) , where k denotes discretetime instant and is an integer
ranging from 0 to C. They become boldface for multiple inputs and multiple outputs.
2. 1. 1 Causalit and Lumpedness
A system is called a memol)'less system if its output y(to) depends only on the input applied
at to; it is independent of the input applied before or after to . This will be stated succinctly as
follows: current output of a memoryless system depends only on current input: it is independent
of past and future inputs. A network that consists of only resistors is a memory less system.
Most systems, however, have memory. By this we mean that the output at to depends on
net) for t < to, t = to, and t ¯ to. That is. current output of a system with memory may depend
on past, current, and future inputs.
A system is called a causal or nonanticipatory system if its current output depends on
past and curent inputs but not on future input. If a system is not causal, then its current output
will depend on future input. In other words, a noncausal system can predict or anticipate what
will be applied in the future. No physical system has such capability. Therefore every physical
system is causal and causality is a necessary condition for a system to be built or implemented
in the real world. This text studies only causal systems.
Current output of a causal system is affected by past input. How far back in time will the
past input afect the current output? Generally, the time should go all the way back to minus
infnity. In other words. the input from ¬¯ to time t has an effect on yet)· Tracking n(t)
from t = ¯ is. if not impossible, very inconvenient. The concept of state can deal with this
problem.
Defnition 2.1 The state x(to) of a system at time to is the infonnation at to that. together
with the input n(t), for t :: to. determines uniquely the output y(t) for all t :: to·
By defnition, ifwe know the state alto. there is no more need to know the inputu(t) applied
before to in determining the output y(t) after to . Thus in some sense. the state summarizes the
effect of past input on future output. For the network shown in Fig. 2.2, if we know the voltages
Xl (to) and X2 (t
O
) across the two capacitors and the current X3 (to) passing through the inductor,
then for any input applied on and after to, we can deterine uniquely the output for t :: to.
Thus the state of the network at time to is
2.2 Li near Systems 7
Figure 2.2 Network with 3 state variables.
¬
R,
It is a 3 x I vector. The entries of x are called state variables. Thus, in general, we may consider
the initial state simply as a set of initial conditions.
Using the state at to, we can express the input and output of a system as
x(to)
n( t) , t :: to
 y(t), t:: to (2. 1)
It means that the output is partly excited by the initial state at to and partly by the input applied
at and after to. In using (2. 1 ), there is no more need to know the input applied before to all the
way back to ¬¯. Thus (2. 1 ) is easier to track and will be called a stateinputoutput pair.
A system is said to be lumped if its number of state variables is fnite or its state is a
fnite vector. The network in Fig. 2. 2 is clearly a lumped system; its state consists of three
numbers. A system is called a distributed system if its state has infnitely many state variables.
The transmission line is the most well known distributed system. We give one more example.
EXAMPLE 2.1 Consider the unittime delay system defned by
yet) = u (t ¯ 1 )
The output i s simply the input delayed by one second. I n order t o determine (y(t), t :: to}
from {u(t) , t :: toJ. we need the information (u(t) , to ¯ 1 :: t < to}. Therefore the initial state
of the system is (u(t), to ¯ 1 :: t < to}. There are infnitely many points in {to ¯ I :: t < to}.
Thus the unittime delay system is a distributed system.
2. 2 Li near Systems
A system is called a linear system if for every to and any two stateinpututput pairs
for i = I, 2, we have
xJto)
.t :: to
ni( t ) , t:: to
8 MATHEMATICAL DESCRI PTIONS OF SYSTEMS
and
..
'
..
¬ ..(homogeneity)
for any real constant .The frst property is called the ..·property, the second, the
.·property. These two properties can be combined as
. . ..
'
¯ . . ..
....
for any real constants .and .and is call
�
d the ...·A system is called
a nonlinear system if the superposition property does not hold.
If the input .is identically zero for then the output will be excited exclusively
by the initial state x(to). This output is called the ...and will be denoted by
.
or
.
'
¯ .
.= 0, :
If the initial state .is zero, then the output will be excited exclusively by the input. This
output is called the ·.....and will be denoted by .or
or
.= 0
'
¬ .
.
The additivity property implies
Output due to
.
.
.
output due to
0 .± ,
.= 0
.output due to
.
·..= ...·....
Thus the response of every linear system can be decomposed into the zerostate response and
the zeroinput response. Furthermore, the two responses can be studied separately and their
sum yields the complete response. For nonlinear systems, the complete response can be very
different from the sum of the zeroinput response and zerostate response. Therefore we cannot
separate the zeroinput and zerostate responses in studying nonlinear systems.
If a system is linear, then the additivity and homogeneity properties apply to zerostate
responses. To be more specifc, if .= 0, then the output will be excited exclusively by
the input and the stateinputoutput equation can be simplifed as lUi ¬ .If the system is
linear, then we have {u ¡ Uz ¬ .¡ ..and ..¬ ..for all .and all Ui. A similar
remark applies to zeroinput responses of any linear system.
Inpututput description We develop a mathematical equation to describe the zerostate
response of linear systems. In this study, the initial state is assumed implicitly to be zero and the
2.2 L i near Systems 9
output is excited exclusively by the input. We consider first SISO linear systems. Let ,
be the pulse shown in Fig. 2.3. It has width .and height .and is located at time Then
every input u(t) can be approximated by a sequence of pulses as shown in Fig. 2.4. The pulse
in Fig. 2.3 has height .thus , .has height and the leftmost pulse in Fig. 2.4
with height .can be expressed as .¬ .. Consequently, the input H can be
expressed symbolically as
Let g"(t, til be the output at time excited by the pulse .= , applied at time
Then we have
¯
¯ . .¬ ... (homogeneity)
¯ ¬ iUi. ¯.. (additivity)
Thus the output ,excited by the input . can be approximated by
.¯ ¯. ..
Figure 2.3 Pulse at
l
u(!.}
l Î l
Figure 2.4 Approximation of input signal.
t
l t¡ ¬ D
(2. 2)
10 MATHEMATI CAL DESCRIPTIONS OF SYSTEMS
Now if t approaches zero, the pulse 0" Uti) becomes an impulse atti, denoted by oCt td, and
the corresponding output will be denoted by get=til. As t approaches zero, the approximation
in (2.2) becomes an equality, the summation becomes an integration, the discrete ti becomes
a continuum and can be replaced by r, and t can be written as dr. Thus (2.2) becomes
y
U) .
.
g(t, r)u(r) dr (2.3)
Note that g(t, r) is a function of two variables. The second variable denotes the time at which
the impulse input is applied; the frst variable denotes the time at which the output is observed.
Because gU, r) is the response excited by an impulse, it is called the impulse response.
If a system is causal, the output will not appear before an input is applied. Thus we have
Causal ¢ g�l, r) = 0 for t < T
A system is said to be relaxed at to if its initial state at to is O. In this case, the output yU),
for t :: to, is excited exclusively by the input u (t) for t :: to. Thus the lower limit of the
integration in (2.3) can be replaced by to. If the system is causal as well, then get, r) .0 for
t < r. Thus the upper limit of the integration in (2.3) can be replaced by t. In conclusion,
every linear system that is causal and relaxed at to can be described by
..gU, r)u(r) dr
!0
(2.4)
In this derivation, the condition of lumpedness is not used. Therefore any lumped or distributed
linear system has such an inpututput description. This description is developed using only
the additivity and homogeneity properties; therefore every linear system, be it an electrical
system, a mechanical system, a chemical process, or any other system, has such a description.
If a linear system has p input terminals and q output tenninals, then (2.4) can be
extended to
where
yet) = .(t« r)u(r)dr
!0
[gll(t,r)
g21(t, r)
G(t. r) .
gql(t, T)
g[2(t, T)
g22(t, T)
(2.5)
gIPU,T)]
g2p(t, T)
gqp(t. T)
and gij (t, T) is the response at time t at the i th output tenninal due to an impulse applied at
time T at the jth input tenninal, the inputs at other terminals being identically zero. That is,
gij(t, T) is the impulse response between the jth input tenninal and the ith output teninal.
Thus .is called the impulse response matrix of the system. We stress once again that if a
system is described by (2.5), the system is linear, relaxed at to, and causal.
Statespace description Every linear lumped system can be described by a set of equations
of the form
2.3 Li near Ti meI nvari ant (LI) Systems 1 1
..A(t)x(l) B(I)u(r)
Y(I) = C(t)x(t) D(t)u(t)
(2. 6)
(2. 7)
where x : = dX/dt.1 For a pinput qoutput system, u is a p x vector and y is a q x 1 vector.
If the system has n state variables, then x is an n x I vector. In order for the matrices in (2.6)
and (2. 7) to be compatible, ·B, C, and D must be n x n, n x p, q x n, and q x p matrices.
The four matrices are all functions of time or timevarying matrices. Equation (2.6) actuallv
consists of a set of n frstorder diferential equations. Equation (2. 7) consists of q algebrai�
e�uatlOns. The set of two equations will be called an ndimensional statespace equation or,
Simply, slale equation. For distributed systems, the dimension is infnity and the two equations
in (2.6) and (2.7) are not used.
The inputoutput description in (2.5) was developed from the linearity condition. The
development of the statespace equation from the linearity condition. however, is not as simple
and Will not be attempted. We will simply accept it as a fact.
2, 3 Li near Ti meI nvari ant (LTI) Systems
A system is said to be tie invariant if for every stateinputoutput pair
and any T, we have
x(to)
'
u(t). I:: to
� Y(I), I:: to
x(to T)
'
u(t  T) . t:: to + T
y(t  T)
, t:: to + T (time shifting)
It means that if the initial state is shifted to time to T and the same input waveform is applied
from to + T instead of from to. then the output waveform will be the same except that it starts
to appear from time to + T. In other words, if the initial state and the input are the same, no
matter at what time they are applied, the output waveform will always be the same. Therefore,
for timeinvariant systems, we can always assume, without loss of generality. that to = O. If a
system IS not time invariant. it is said to be time varing.
Time invariance is defned for systems, not for signals. Signals are mostly time varvina.
If a signal is time invariant such as H(t) .for all I, then it is a very simple or a tri vial si�na..
The characteristics of timeinvariant systems must be independent of time. For example, the
network in Fig. 2.2 is time invariant if R" C;, and L, are constants.
Some physical systems must be modeled as timevarying systems. For example, a buring
rocket is a timevarying system, because its mass decreases rapidly with time. Although the
performance of an automobile or a TV set may deteriorate over a long period of time, its
characteristics do not change appreciable in the frst couple of years. Thus a large number of
physical systems can be modeled as timeinvariant systems over a limited time eriod.
1. We use A .¯ Ü to denote that A. by defnition. equals Ü. We use Ó ¯. Ü to denote that Ü. by defnition, equals A.
12 MATHEMATICAL DESCRI PTI ONS OF SYSTEMS
Inputoutput description The zerostate response of a linear system can be described by
(2.4). Now if the system is time i nvariant as well, then we have'
get, r) = T, r T) = r, 0) = r)
for any T. Thus (2. 4) reduces to
y( t) = ¯ r)u( r) dr = g( r)u(r  r) dr (2.8)
where we have replaced by O. The second equality can easily be verifed by changing the
variable. The integration in ( 2. 8) is called a convolution integral. Unlike the timevarying case
where is a function of two variables, is a function of a single variable in the timeinvariant
case. By defnition = 0) i s the output at time due to an impulse input applied at
time O. The condition for a linear timeinv. iant system to be causal is = 0 for t < O.
EXAMPLE 2.2 The unittime delay system studied in Example 2. 1 is a device whose output
equals the input delayed by I second. If we apply the impulse (t) at the input terminal, the
output is  I). Thus the impulse response of the system is 8( t ¯ I).
EXAMPLE 2.3 Consider the unityfeedback system shown in Fig. 2. 5(a). It consists of a
multiplier with gain and a unittime delay element. It is a SISO system. Let r (t) be the
input of the feedback system. If r (t ) = then the output is the impulse response of the
feedback system and equals
¯
gf( l) =  1 )  2) 
. .
¨ =   i )
l=!
Let r (t) be any input with r ( t) ± 0 for t < 0; then the output is given by
y(t) =
gf(t r) r(r) dr = 8( t r  i ) r (r) dr
0
�
0
= 
�
= ·b( t  i )
Because the unittime delay system is distributed, so is the feedback system.
u)
Unittime
delay
element
Figure 2.5 Positive and negative feedback systems.
(b)
Unitlime
delay
element
(2. 9)
}!!)
2. �ote that g(r. t} and g(t " r) are two diffrent functions. However. for convenience. the same symbol g is used.
2.3 Li near Ti meI nvari ant ( LTI) Systems 13
Transferfunction matrix The Laplace transform is an important tool in the study of linear
timeinvariant ( LTI) systems. Let be the Laplace transform of y( t) , that is,
= y(t) eS'dt
Throughout this text, we use a variable with a circumfex to denote the Laplace transform of
the variable. For causal systems, we have = 0 for t < 0 or ¯ r) = 0 for r ¯ t. Thus
the upper integration limit in (2. 8) can be replaced by ¯. Substituting ( 2. 8) and interchanging
the order of integrations, we obtain
= ¯ T) U( r) dresur)eSO
dt
t=O
r
=O
=
¬ r) eSUrJdtu(r)e
sr
dr
r=O [=0
which becomes. after introducing the new variable v = t  r,
= �g (v)e Sl' dVu (r)e sr dr
Again using the causality condition to replace the lower integration limit inside the parentheses
from v = r to v = O. the integration becomes independent of r and the double integrations
become

= ,g( v)esr dv u(r) es
r
dr
or
where
= g( t) eHdt
(2.10)
is called the transfer fnction of the system. Thus the transfer function is the Laplace transform
of the impulse response and. conversely, the impulse response is the inverse Laplace transform
of the transfer function. We see that the Laplace transform transforms the convolution integral
in (2.8) into the algebraic equation in (2.10). In analysis and design. it is simpler to use algebraic
equatIOns than to use convolutions. Thus the convolution in (2. 8) will rarely be used in the
remainder of this text.
or
For a pinput qoutput system, (2.10) can be extended as
�
�
. +
« +
 «
=
��
(2.11)
14
MATHEMATICAL DESCRIPTIONS OF SYSTEMS
where guts) is the transfer function from the jth input to the ith output. The q x p matrix
(. (s) is called the transferfunction matrix or, simply, transfer matrix of the system.
EXAMPLE 2.4 Consider the unittime delay system studied in Example 2.2. Its impulse
response is o(t  1). Therefore its transfer function is
g(s) = L[o(t  I)] = 8(t  I)e"dt =
Þ¯³'
= Þ¯³
This transfer function is an irrational function of s.
EXAMPLE 2.5 Consider the feedback system shown in Fig. 2.5(a). The transfer function of
the unittime delay element is es. The transfer function from r to y can be computed directly
from the block diagram as
'
aeS
gj(s) =
�
(2.12)
This can also be obtained by taking the Laplace transform of the impulse response, which was
computed in (2.9) as
O
gj(t) = _
t
o(t  i)
t=1
Because L[o(t i)] = e
ts
the Laplace transform of gj(t) is
O
O
gj(s) = L[gjU») = ¯a'eiS = ae
s
¯(ae
s
)
,
Using
í=1 t=U
¯
I
'r ' ¯
� Ir
i=U
for Ir I < 1, we can express the infnite series in closed form as
which is the same as (2.12).
The transfer function in (2.12) is an irrational function of s. This is so because the feedback
system is a distributed system. If a linear timeinvariant system is lumped, then its transfer
function will be a rational function of s. We study mostly lumped systems; thus the transfer
functions we will encounter are mostly rational functions of s.
Every rational transfer function can be expressed as g(s) = N(s)/ D(s), where N(s) and
D(s) are polynomials of s. Let us use deg to denote the degree of a polynomial. Then g(s) can
be classifed as follows:
• g(s) proper 9 deg D(s) : deg N(s) 9 g(o) = zero or nonzero constant.
2.3 li near TimeInvari ant (IT!) Systems 1 5
• g(s) strictly proper 9 deg DCs) ¯ deg N(s) 9 g(o) = O.
• g(s) biproper 9 deg D(s) = deg N(s) 9 g(o) = nonzero constant.
• g(s) improper 9 deg D(s) < deg N(s) 9 Ig(e)1 = 0.
Improper rational transfer functions will amplify highfrequency noise, which often exists in
the real world; therefore improper rational transfer functions rarely arise in practice.
A real or complex number A is called a pole of the proper transfer function g(s) =
N(s)/ D(s) if Ig(A)1 = ¯+ a zero if g(A) = O. If N(s) and D(s) are coprime, that is, have
no common factors of degree 1 or higher, then all roots of N(s) are the zeros of g(s), and all
roots of D(s) are the poles of g(s). In terms of poles and zeros, the transfer function can be
expressed as
g(s) =
k
(s  �1)(S  (s  Zm)
(s  pIles  P2)
. .
(s PII)
This is called the zerpolegain f0r. In MATLAB, this form can be obtained from the transfer
function by calling [ z, p , kJ = tf2zp ( num, den).
A rational matrix ((s) is said to be proper if its every entry is proper or if (( 0) is a zero
or nonzero constant matrix: it is strictly proper if its every entry is strictly proper or if ((J) is
a zero matrix. If a rational matrix ((s) is square and if both (.(s) and (1 (s) are proper. then
((s) is said to be biproper. We call A a pole of ((s) if it is a pole of some entry of (.(s). Thus
every pol
�
of every entry of ((s) is a pole of ((s). There are a number of ways of defning
zeros for G(s). We call A a bloc
k
ing zero if it is a zero of every nonzero entry of ((s). A more
useful defnition is the transmission zero, which will be introduced in Chapter 9.
Statespace equation Every linear timeinvariant lumped system can be described by a set
of equations of the form
x(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + DuCt)
(2.13)
For a system with p inputs, q outputs, and n state variables, A, B, C, and D are, respectively,
n x n. n X p, q x n. and q x p constant matrices. Applying the Laplace transform to (2
.
13)
yields
which implies
sx(s) x(O) = Ax(s) + Bu(s)
hs) = Cx(s) + Du(s)
xes) = (sl  A)
l
x(O) + (sl A)IBu(s)
hs) = C(sl ¯ A)I
X
(O) + C(sI  A)IBu(s) + Du(s)
(2.14)
(2.15)
They are algebraic equations. Given x(O) and u(s). xes) and hs) can be computed algebraically
from (2.14) and (2.15). Their inverse Laplace transforms yield the time responses xU) and
y(t). The equations also reveal the fact that the response of a linear system can be decomposed
1 6 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
as the zerostate response and the zeroinput response. If the initial state x(O) is zero, then
(2.15) reduces to
Y(s) = [C(sI A) I B + D] u(s)
Comparing this with (2.11) yields
,
I
G(s) = C(sI A) B + D
This relates the inputoutput (or transfer matrix) and statespace descriptions.
(2.16)
The functions t f 2 s s and s s 2 t f in MATLAB compute one description from the other.
They compute only the SISO and SIMO cases. For example, [ num, den] ¯ ss2 tf ( a , b, C .
d , 1 ) computes the transfer matrix from the frst input to all outputs or, equivalently, the frst
column of G(s). If the last argument 1 in ss2 t f ( a , b, C , d, 1 ) is replaced by 3, then the
function generates the third column of G(s:.
To conclude this section, we mention that the Laplace transform is not used in studying
linear timevarying systems. The Laplace transform of get, r) is a function of two variables
and L[A(t)x(t)] f L[A(t) ]L[x(t)]; thus the Laplace transform does not offer any advantage
and is not used in studying timevarying systems.
2. 3. 1 OpAmp Circuit I mpl ementation
Every linear timeinvariant (LTI) statespace equation can be implemented using an operational
amplifer (opamp) circuit. Figure 2.6 shows two standard opamp circuit elements. All inputs
are connected. through resistors, to the inverting terminal. Not shown are the grounded
non inverting terminal and power supply. If the feedback branch is a resistor as shown in Fig.
2.6(a), then the output of the element is (axi +bX
2
+CX3). I the feedback branch is a capactor
with capacitance C and RC = 1 as shown in Fig. 2. 6(b), and if the output is assigned as x, then
.� = (avi +bt'2 + CV3 ) ' We call the frst element an adder; the second element, an integrtor.
Actually, the adder functions also as multipliers and the integrator functions also as multipliers
and adder. If we use only one input, say, XI . in Fig. 2.6(a), then the output equals axl , and
the element can be used as an inverter with gain a. Now we use an example to show that every
LTI statespace equation can be implemented using the two types of elements in Fig. 2.6.
Consider the state equation
·. ·
�
·
. ¬ �¬ CX3)
.
Figure 2.6 Two opamp circuit elements.
0. 3
] [
XI (t)
]
+
[
2
]
u (t )
S X2(t) 0
�·
·
· X
L`� ¬
.
¬ ¬
(2.17)
2.4 Li nearization 1 7
[
XI (t
) ] y(t) = [ 2 3] + 5u(t)
X2(t)
(2. I S)
It has dimension 2 and we need two integrators to implement it. We have the freedom in
choosing the output of each integrator as +Xi or ¬7¡ . Suppose we assign the output of the
lefthandslde (LHS) mtegrator as XI and the output of the righthandside (RHS) integrator
as X2 as shown in Fig. 2. 7
.
Then the input of the LHS integrator should be, from the frst
equation of (2.17), XI = 2xl 0 3X·+ 2u and is connected as shown. The input of the
RHS mtegrator should be .�2 = XI  SX2 and is connected as shown. If the output of the adder
l5 chosen as y, then its input should equal y = 2xI 3X2  5u, and is connected as shown.
Thus the state equation in (2.17) and (2
.
I S) can be implemented as shown in Fil. 2.7. Note that
there are m
.
any ways to implement the same equation. For example, if we as·ign the outputs
of the two mtegrators m Fig. 2.7 as XI and Xl , instead of XI and X2, then we will obtain a
different implementation.
In actual operational amplifer circuits, the range of signals is limited, usually I or : volts
below the supplied voltage. If any signal grows outside the range, the circuit will saturate or
bum out and the circuit will not behave as the equation dictates. There is. however, a way to
deal with this problem. as we will discuss in Section 4.3. 1.
2,4 Li nearization
Most physical systems are nonlinear and time varying. Some of them can be described bv the
nonlinear differential equation of the form

·
·
·
I¡ ·
u
· 3 ·.
·
· 1
\�
·
·
·
¨Ì4
Figure 2.7 Opamp implementation of (2. 1 7) and (2. 1 8).
1 8 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
x(t) ¬ b(x(t) , u(t) , t)
yet) ¬ f(x(t) , u(t) , t)
(2. 1 9)
where b and f are nonlinear functions. The behavior of such equations can be very complicated
and its study is beyond the scope of this text.
Some nonlinear equations, however,
c
an be approximated by linear equations under
certain conditions. Suppose for some input function Uo (t) and some initial state, Xo (t) is the
solution of (2. 1 9); that is,
(2.20)
Now suppose the input is perturbed slightly to become Uo (t) i(t) and tbe initial state is also
perturbed only slightly. For some nonlinexequations, the corresponding solution may difer
from xo(t) only slightly. In this case, the s
'
olution can be expressed as xo (t) + x(t) with x(t)
small for all t.3 Under this assumption, we can expand (2.19) as
xo (t) + x(t) = b(xo(t) x(t), uo(t) + i(t), t)
ab
_
ab _
= b(xo (t), Uo (t), t)
ax
x
au
u
(2.2 1 )
They are called lacobians. Because A and B are computed along the two time functions xo (t)
and uo (t), they are, i n general, functions of t. Using (2.20) and neglecting higher powers of x
and i. we can reduce (2. 21 ) to
x(t) ¬ A(t)x(t) B(t)i(t)
This is a linear statespace equation. The equation yet) ¬ f(xU). u(t) , t) can be similarly
linearized. This linearization technique i s often used in practice to obtain linear equations.
2. 5 Exampl es
In this section we use examples to illustrate how to develop transfer functions and statespace
equations for physical systems.
EXAMPLE 2.6 Consider the mechanical system shown in Fig. 2. 8. It consists of a block with
mass m connected to a wall through a spring. We consider the applied force H to b the input
3. This is not true in general. For some nonlinear equations, a very small diference in initial states will generate
completely diferent solutions, yielding the phenomenon of chaos.
2. 5 Exampl es 1 9
and displacement y from the equilibrium to be the output. The friction between the foor and the
block generally consists of three distinct parts: static friction, Coulomb friction, and viscous
fnctlOn as shown in Fig. 2.9. Note that the horizontal coordinate is velocity )' ~ dydt . The
fnctJon
·
15 clearly not a linear function of the velocity. To simplify analysis, we disregard
the statIc and Coulomb fnctlOns and consider only the viscous friction. Then the friction
becomes linear and can be expressed as k, ). (t), where kl is the viscous friction coeffcient. The
�haracteristics of the spring are shown in Fig. 2. 1 0: it i s not linear. However, if the displacement
5 IUlted to (YI , Y2
) as shown, then the spring can be considered to be linear and the spring
force
.
equals k2y, where .is the spring constant. Thus the mechanical system can be modeled
as a lInear system under linearization and simplifcation.
Figure 2.8 Mechanical system.
, = 0
Force
Force
Static
Viscous friction
o
Velocity
 Velocity
Coulomb
(a)
(b)
Figure 2.9 Mechanical system.(a) Static and Coulomb frictions. (b) Viscous friction.
Figre 2.10 Characteristic of spring.
Force
Break
f
\ì
Displacement
20
MATHEMATICAL DESCRI PTIONS OF SYSTEMS
We apply Newton's law to develop an equation to describe the system. The applied force
u must overcome the friction and the spring force. The remainder is used to accelerate the
block. Thus we have
(2.22)
where y = d
2
y(t)ldtZ and y = dy(t)ldt. Applying the Laplace transform and assuming zero
initial conditions, we obtain
which implies
I ,
)s) =
u(s)
msz + kl s + k2
í
This is the inpututput description of the
�
ystem. Its transfer function is I /(ms
z
+ kls + kl) .
If m = I , kl
= 3, kl = 2, then the impulse response of the system is
g(l) = r
l [
I
]
= £.1
[_
1
_ _ _
1
_]
= e
r
e2r
s
2
+ 3s + 2 s + I s + 2
and the convolution description of the system is
y(I) =
r
g(l  r) u (r) dr =
r
(e(
r
r)  e2(tr)
)
u
(
r
)
dr
Next we develop a statespace equation to describe the system. Let us select the displace
ment and velocity of the block as state variables; that is, XI = y, Xz = y. Then we have, using
(2.22),
They can be expressed in matrix form as
[XI (t)
] [
0
Xl (t)
=
k2 /m
I
[ XI (I)
] [
0
]
( )
kl /m X2 (t)
+
1
1m
u t
y(l) = [ I 0 ]
[
XJ Ct l ]
XlU)
This statespace equation describes the system.
EXAMPLE 2.7 Consider the system shown in Fig. 2.11. It consists of two blocks, with masses
m ¡ and m2 . connected by three springs with sprng constants k" i = I , 2, 3. To simplify the
discussion, we assume that there is no friction between the blocks and the foor. The applied
/I
� �. �
77777777777777777777
Figure 2.11 Sprngmass system.
`
2. 5 Examples 21
force u I must overcome the spring forces and the remainder i s used t o accelerate the block,
thus we have
or
mL YI + (kl + kl) YI  kZY2 = U I
For the second block. we have
They can be combined as
[
ml
O
] [
�
I
]
+
[
kl + kl

k
2 ] [ Y
I
]
=
[
u I ]
· ml Y2 k2 kl + k2 Y2 U2
(2.23)
(2.24)
This is a standard equation in studying vibration and is said to be in normal form. See Reference
[18]. Let us defne
XI :=
Y
I
Then we can readily obtain
(k' °U, )
X2
ml
X3

0
.
k,
X4 �
m2
[
),
I
]
y :
= =
)'2 0
0
0
0
0
0
X2 : = YI X3 := Y2 x� : = Y2
0
��
[
:;
]
k2
ml
0
(kl + kl)
ml m2
0 0
]
1 ·
X
This twoinput twooutput state equation describes the system in Fig. 2.11.
To develop its inputoutput description, we apply the Laplace transform to (2.23) and
(2.24) and assume zero initial conditions to yield
ml s
2
.vl (S) + (kl + kzJ.' I (s)  k2Yz (s) = Il l (s)
mls2Y
2
(S)  k2 YI (s) + (kl + k2 )Y2 (S) = 112 (S)
From these two equations, we can readily obtain
where
m2s
2
+ kl + kz
[ .�I (S) ]
=
des)
Y2 (S) k
z
des)
des) : = (m I s
2
+ kl + k2) (m2s
2
+ kl + k2)  k�
This is the transfermatrix description of the system. Thus what we will discuss in this text can
be applied directly to study vibration.
22 MATHEMATICAL DESCRI PTIONS OF SYSTEMS
EXAMPLE 2.8 Consider a cart with an inverted pendulum hinged on top of it as shown in Fig.
2. 1 2. For simplicity, the cart and the pendulum are assumed to move in only one plane, and the
friction, the mass of the stick, and the gust of wind are disregarded. The problem is to maintain
the pendulum at the vertical position. For example, if the inverted pendulum is falling in the
direction shown, the cart moves 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 vehicle on takeof.
Let H and V be, respectively, the horizontal and vertical forces exerted by the car on the
pendulum as shown. The application of Newton's law to the linear movements yields
dZy
M= u  H
dt2
dZ I "" " ¯ •
H = m(y + / si n B) = my + mlB cos B  ml(8t sm8
dtZ
dZ .. '
2
mg  V = m(/ cos 8) = ml [8 sin 8  (8) cos 8]
dtZ
The application of Newton's law to the rotational movement of the pendulum around the hinge
yields
mgl sin 8 = mle · 1 + my I cos 8
They are nonlinear equations. Because the design objective is to maintain the pendulum
at the vertical position, it is reasonable to assume 8 and e to be small. Under this assumption,
we can use the approximation sin 8 = 8 and cos 8 = By retainin
�
.
only the linear terms in 8
and e or, equivalently, dropping the terms with e2, (0)2, eo, and e8, we obtain V = mg and
which imply
My = u  my  mle
g8 = Ie + y
My = u  mg8
Mle = (M + m)g8  u
Figre 2.12 Ca with invered pendulum.
(2.25)
(2.26)
'
2. 5 Examples 23
Using these linearized equations, we now can develop the input{utput and statespace
descriptions. Applying the Laplace transform to (2.25) and (2.26) and assuming zero initial
conditions, we obtain
Ms
Z
y(s) = u(s)  mgB(s)
2
> >
Mis 8(s) = (M + m)g8(s)  u (s)
From these equations, we can readily compute the transfer function gyu (s) from u to y and the
transfer function gou (s) from u to 8 as
> s2  g
g yu (s) = 
Z
:: 
S [Ms2  (M + m)g]
>
gou (s) = �
Ms2  (M +m)g
To develop a statespace equation, we select state variables as Xl = y, X
z
= v. X} = 8.
and X4 = e. Then from this selection, (2.25), and (2.26) we can readily obtain
.
��
=
� �
m
�
/M
�
�
+
I /
O
M
u
;3 ° ° ° I X3 °
;4 ° ° (M + m)g/MI ° X4 l /MI
Y = [I ° ° O]x (2.27)
This state equation has dimension 4 and describes the system when 8 and 0 are very small.
EXAMPLE 2.9 A communication satellite of mass m orbiting around the earth is shown in
Fig. 2. 1 3. The altitude of the satellite is specifed by r(t), 8(t), and ¢(t) as shown. The orbit
can be controlled by three orthogonal thrusts u, (t), uo(t), and u<(t). The state, input. and
output of the system are chosen as
r(t)
;(t)
r(t)
8(t)
x(t) =
O(t)
uU) = uo (t) y(t) = 8(t)
¢(t)
u< (t) ¢(t)
1(t)
Then the system can be shown to be described by
x = h(x, u) =
;
r02 cos
2
¢ + r12  k/r
z
+ ur/m
o
2;0/r + 201 sin ¢/ cos ¢ + uo/mr cos ¢
1
· z
.
8 cos ¢ sin ¢  2r¢/r + u</mr
(2.28)
24 MATHEMATICAL DESCRI PTIONS OF SYSTEMS
Figure 2.13 Satellite in orbit.
One solution, which corresponds to a circular equatorial orbit, is given by
Uo = 0
with r;w; = k , a known physical constant. Once the satellite reaches the orbit, i t will remain
in the orbit as long as there are no disturbances. If the satellite deviates from the orbit, thrusts
must be applied to push it back to the orbit. Define
x(t) = xo(t) + x(t) u(t) = uo(t) + i(t) y(t) = Yo + y(t)
If the perturbation is very small, then (2. 28) can be linearized as
0 0 0 0 0
. 0 0 2woro 0 0
0 0 0 0 0
x(t) =
0
2wo
0 0 0 0
x(t)
ro
0 0 0 0 0
0 0 0 0
¯
0 ¯�
0 0 0
1
0 0
m
0 0 0
+
0
I
0
i(t)
mro
0 0 0
I
0 0
mro
2. 5 Examples 25
i'
�
0 0 0 0
'u'
0 0 0
(2. 29)
0 0 0
This sixdimensional state equation describes the system. In this equation, A, B, and C happen
to be constant.
I
f the orbit is an elliptic one. then they will be time varying. We note that the
three matrices are all block diagonal. Thus the equation can be decomposed into two uncoupled
parts. one involving r and e, the other <. Studying these two parts independently can simplify
analysis and design.
EXAMPLE 2.10 In chemical plants, it is ofen necessary to maintain the levels of liquids. A
simplifed model of a connection of two tanks is shown in Fig. 2. 1 1. I t is assumed that under
normal operation, the infows and outfows of both tanks all equal Q and their liquid levels
equal HI and · Let Il be infow perturbation of the frst tank, which will cause variations
in liquid level and outfow
as shown. These variations will cause level variation and
outfow variation y in the second tank. It is assumed that
=

·
and
=
¯
·
where ·are the fow resistances and depend on the normal height HI and H2. They can also
be controlled by the valves. Changes of liquid levels are govered by
· .= (II . and ·.= y) dt
where ·are the cross sections of the tanks. From these equations. we can readily obtain
. H
=  
· · ·
•  x.
=   ¯
¯
· · · ·
Thus the statespace description of the system is given by
�
[
=
������
= [0 · .
Figure 2.14 Hydraulic
tanks. Q ¬ tI
'
& & & ~ . ~
. . . . . . . . . .
HI + XI
Al
*
Q + )'1 Q + y
26 MATHEMATICAL DESCRI PTIONS OF SYSTEMS
�
Its transfer fnction can be computed as
. I
g(s) ¬
¯
AI A2RI R2S + (AI RI + AI R2 + A2 R2)s + I
2. 5. 1 RLC netorks
In RLC networks, capacitors and inductors can store energy and are associated with state
variables. If a capacitor voltage is assigned as a state variable x, then its curent is C X, where
C is its capacitance. If an inductor curent is assigned as a state variable x, then its voltage is
Lx, where L is its inductance. Note that resistors are memoryless elements, and their currents
or voltages should not be assigned as state variables. For most simple RLC networks, once
state variables are assigned, their state eqáations can be developed by applying Kirchhof's
current and voltage laws, as the next example illustrates.
EXAMPLE 2.1 1 Consider the network shown in Fig. 2. 1 5. We assign the Cicapacitor voltages
as Xi , i ¬ 1 , 2 and the inductor current as X3. It is important to specify their polarities. Then
their currents and voltage are, respectively, CI XI o C2X2, and LX3 with the polarities shown.
From the fgure, we see that the voltage across the resistor is H  XI with the polarity shown.
Thus its current is (u  xI l I R. Applying Kirchhof's current law at node A yields C2X2 ~ X3;
at node B it yields
Thus we have
. XI X3 U
XI ¬  +
RCI CI RCI
I
X2 " X3
C2
Appling Kirchhof's voltage law to the righthandside loop yields LX3 ¬ XI  X2 O
The output )' is given by
. XI  X2
X3
= 
Figure 2.15 Network.
'
[
They can be combined in matrix form as
o
o
I lL
y = [I  I O]x + o · u
2. 5 Examples
I/CI I IRC'
I/C2 x + 0 u
o 0
This threedimensional state equation describes the network shown in Fig. 2. 1 5.
27
The procedure used i n the preceding example can be employed t o develop state equations
to describe simple RLC networks. The procedure is fairly simple: assign state variables and
then use branch characteristics and Kirchhof's laws to develop state equations. The procedure
can be stated more systematically by using graph concepts, as we will introduce next. The
procedure and subsequent Example 2. 1 2. however, can be skipped without loss of continuity.
First we introduce briefy the concepts of tree, link, and cutset of a network. We consider
only connected networks. Every capacitor, inductor, 'resistor, voltage source, and current source
will be considered as a branch. Branches are connected at· nodes. Thus a network can be
considered to consist of only branches and nodes. A ·Ioop is a connection of branches starting
from one point and coming back to the same point without passing any point twice. The
algebraic sum ofall voltages along ever loop is zero (Kirchhoff's voltage law). The set of all
branches connect to a node is called a cutset. More generally, a cutset of a connected network
is any minimal set of branches so that the removal ohhe set causes the remaining network
to be unconnected. For example, removing all branches connected to a node leaves the node
unconnected to the remaining network. The algebraic sum of all brnch currents ill ever
cutset is zero (Kirchhoff's current law).
A tree of a network is defned as any connection of branches connecting all the nodes but
containing no loops. A branch is called a tree branch if it is in the tree, a link if it is not. With
respect to a chosen tree, every link has a unique loop, called thefulldamental loop, in which
the remaining loop branches are all tree branches. Every tree branch has a unique cutset. called
the fndamental cutset, in which the remaining cutset branches are all links. In other words, a
fundamental loop contains only one link and a fundamental cutset contains only one tree branch.
� Procedure for developing statespace equations'
1. Consider an RLC network. We frst choose a normal tree. The branches of the normal
tree are chosen in the order of voltage sources, capacitors, resistors, inductors, and current
sources.
2. Assign the capacitor voltages in the normal tree and the inductor currents in the links as
state variables. Capacitor voltages in the links and inductor curents in the normal tree are
not assigned.
3. Express the voltage and current of every branch in terms of the state variables and,
if necessary, the inputs by applying Kirchhoff's voltage law to fundamental loops and
Kirchhof's current law to fundamental cutsets.
4. The reader may skip this proedure and go directly to Example 2. 1 3.
28 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
4. Apply Kirchhof's voltage or current law to the fundamental loop or cutset of every branch
that is assigned as a state variable.
EXAMPLE 2.12 Consider the network shown in Fig. 2. 1 6. The normal tree is chosen as shown
with heavy lines; it consists of the voltage source, two capacitors, and the I Q resistor. The
capacitor voltages in the normal tree and the inductor current in the link will be assigned as
state varables. If the voltage across the 3F capacitor is assigned as ²! . then its current is 3.t l .
The voltage across the I F capacitor is assigned as X2 and its current i s X2. The current through
the 2H inductor is assigned as ²1 and its voltage is 2X3. Because the 2Q resistor is a link. we
use its fundamental loop to fnd its voltage as U I  XI . Thus its current is (U I  xI )/2. The
I Q resistor is a tree branch. We use its fundamental cutset to fnd its current as ²¸ . Thus its
voltage is I . X3 = X3. This completes Stel 3.
The 3F capacitor i s a tree branch an
d
its fundamental cutset i s as shown. The algebraic
sum of the cutset currents is 0 or
which implies
UI ¨ ²I .
 3xI + H¬  ²1 = 0
2

XI = .xI  ¹²\ + .U I + �U2
The I F capacitor is a tree branch, and from its fundamental cutset we have X2  ²\ = 0 or
The 2H inductor is a link. The voltage along its fundamental loop is 2X3 + ²1 " ²J + ²] = 0
or
(U I  xl )/2
'7 l
Figure 2.16 Network with two inputs.
/
Fundamental
cutset of @
/
Fundamental
cutset of Q
/
2. 5 Exampl es 29
They can be expressed in matrix form as
,
�
��
(2. 30)
If we consider the voltage across the 2H inductor and the current through the 2Q resistor as
the outputs, then we have
Yl = 2X3 = 7l  X2  73 = [ I  I  I lx
and
Y2 = 0. 5(u l  xt l = [ 0 5 0 O]x + [0.5 Olu
They can be written in matrix form as
y =

�
5
I
·
I 0 0
·
x +
0.5 0 u
Equations (2. 30) and (2. 3 1 ) are the statespace description of the network.
(2. 3 1 )
The transfer matrix of the network can be computed directly from the network or using
the formula in (2. 1 6) :
We will use MATLAB t o compute this equation. We type
a= [  1 / 6 0  1 ; 3 ; 0 0 1 ; 0 . 5  0 . 5  : . S ] ; b= [ 1 / 6 1 ! 3 ; C 0 ; 0 G J ;
c= [ l  1  1 ;  0 . 5 0 O ] ; d= [ O 0 ; 0 . 5 0 ] ;
[ N: , d1 ] = s s 2 t f ( a , b, c , d , 1 )
which yields
Nl=
d1=
0. 000 0. 1 667 0. 0000 0. 0000
0. 5000 0. 2500 0. 3333 0.0000
1 .0000 0.6667 0. 7500 0.0833
This is the frst column of the transfer matrix. We repeat the computation for the second input.
Thus the transfer matrix of the network is
0. 1 6675
2
G(5) =
5
1
+ 0.
�
66752 +
2
0. 755 + 0.083
0. 55 + 0.255 + 0. 33335
5
3
+ 0.666752 + 0.755 + 0. 083
0. 33335
2
5
3
+ 0.666752 + 0. 755 + 0.0833
0. 1 66752  0.08335  0.0833
5
3
+ 0.666752 + 0. 755 + 0.0833
30 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
EXAMPLE 2.13 Consider the network shown i n Fig. 2. 1 7(a), where T i s a tunnel diode with
the characteristics shown i n Fig. 2. 1 7(b). Let . be the voltage across the capacitor and .be
the current through the inductor. Then we have v = Xand
Xz (t) =
. (t)
·
i (t) =
i¡ (t) h(xi (t»
Liz(t) = · ·. X
I
(t)
They can be arranged as
. h(xI (t» X2(t)
. =
·
. XI (t) ¯ ·. ·
xz (t) =
I L
·
L
I
(2.32)
This set of nonlinear equations describes the network. Now if XI (t) is known to lie only inside
the range (a, b) shown in Fig. 2. l 7(b), then h(xi (t» can be approximated by h(.'1 (t» =
.(t)1 · In this case, the network can be reduced to the one in Fig. 2. 1 7(c) and can be
described by
.
[
=
·
[
.
[
` [
·
Xz I lL RIL Xz lL
LX¬
¬
L
³2
l /R,
t ¯ h(v)
¯
+
Ã³_
Ã
� `
0
X;
L =
J
LX¡
(a)
(b)
L X¸
L
³2
Ã
l ¯c
X¡ u
Ã;
L _
J
R
�
x, C ¡
�
t
(c) (d)
Figure 2.17 Network with a tunnel diode.
.
Ã`
'
Ì
Ì

i
í
2. 6 Di screteTime Systems 31
This is an LTI statespace equation. Now if . ( t ) is known t o lie only inside the range (c, d)
shown i n Fig. 2. l 7(b). we may introduce the variables . ( t ) = . ( t ) vo, andx
2
(t) = x2 (t ) io
and approximate h(xi (t» as io " .(t)1 ·
Substituting these into (2.32) yields
[
·
[
[
[
=
"
lL
· . ¯
I lL
E
where E =
E
¯ va  · This equation is obtained by shifting the operating point from
to (vo, io) and by linearization at (va' io) . Because the two linearized equations are identical
if ·i s replaced by ·and E by · we can readil y obtain its equivalent network shown in
Fig. 2. 1 7(d). Note that it is not obvious how to obtain the equivalent network from the original
network without frst developing the state equation.
2. 6 DiscreteTi me Systems
This section develops the discrete counterpart of continuoustime systems. Because most
concepts in continuoustime systems can be applied directly to the discretetime systems,
the discussion will be brief.
The input and output of every discretetime system will be assumed to have the same
sampling period T and will be denoted by u[k] := u(kT) , y[k] : = y(kT) , where k is an integer
ranging from ¬¯to +x. A discretetime system is causal if current output depends on current
and past inputs. The state at time ko, denoted by x[ko], is the information at time instant ko,
which together with u[k], k :: ko, determines uniquely the output y[kJ, k :: ko. The entries of
X are called state variables. If the number of state variables is fnite, the discretetime system
is lumped; otherwise. it is distributed. Every continuoustime system involving time delay,
as the ones in Examples 2. 1 and 2. 3, is a distributed system. In a discretetime system, if the
time delay is an integer multiple of the sampling period T, then the discretetime system is a
lumped system.
A discretetime system i s linear if the additivity and homogeneity properties hold. The
response of every linear discretetime system can be decomposed as
Response = zerostate response
·
zerinput response
and the zerostate responses satisfy the superposition property. So do the zeroinput responses.
Inpututput description Let o [k] be the impulse sequence defned as
o [k  m] = �
if k = m
if k f m
where both k and m are integers. denoting sampling instants. It is the discrete counterpart of
the impulse 8 (£ ¯ tt l . The impulse oCt  ti l has zero width and infnite height and cannot be
generated in practice: whereas the impulse sequence a[k  m] can easily be generated. Let
u [k] be any input sequence. Then it can be expressed as
¯
u [k] = _ u[m] 8[k  m]
O¯¬¯
Let g[k, m] be the output at time instant k excited by the impulse sequence applied at time
instant m. Then we have
32 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
··  », · .· m]
·· », .», · .· m] u[ m] (homogeneity)
S:· », . », · S.· », .»: (additivity)
Thus the output ··, excited by the input . ·, equals
¯
··, = S .· »¡ .»,
.·
(2.33)
This is the discrete counterpart of (2.3) and its derivation is considerably simpler. The sequence
.· m] is called the ·»,./·. ··,»· ·,..».
If a discretetime system is causal, no output will appear before an input is applied. Thus
we have
Causal 4 .· »i = 0, for · < »
If a system is rela'ed at ·· and causal, then (2.33) can be reduced to
.
··i = S .· »¡.»,
..
as in (2.4).
(2.34)
If a linear discretetime system is time invariant as well, then the time shifting property
holds. In this case, the initial time instant can always be chosen as ·· = 0 and (2.34) becomes
. .
··, = S .· »¡ .», = S .»i .·  »i
.. ..
This is the discrete counterpart of (2.8) and is called a e··.·· .»./.·.»
(2.35)
The �transform is û important tool in the study of LTI discretetime systems. Let S(z)
be the ztransform of ··i defned as
¯
Yez) : = ¯··,¦ : = S··,:
.
(2.36)
.�.
We frst replace the upper limit of the integration in (2.35) to C.¯ and then substitute it into
(2.36) to yield
S· (z) = __.·  m] ll [m]·
.
 .

.� �.
=
__.· ¯ m] z
<k
:. », ·

�. .�.
=
./i·
»,·

=: g( Z) u(Z)
�. �.
5. Tis is penitted under the causality assumption.
2. 6 Di screteTime Systems 33
where we have interchanged the order of summations, introduced the new variable I = · »
and then used the fact that ./· = 0 for I < 0 t o make the inner summation independent of m.
The equation
·.:· = g(z)u(z) (2.37)
is the discrete counterpart of(2.1O). The function g(z) is the ztransform of the i mpulse response
sequence .·, and is called the e··.·· ··.»·).· ).».··» Both the discrete convolution and
transfer function describe only zerostate responses.
EXAMPLE 2.14 Consider the unitsamplingtime delay system defned by
··, = .· t ,
The output equals the input delayed by one sampling period. Its impulse response sequence is
.·i = · · . , and its discrete transfer function is
g(;) = z ·· ¯ I I I = ;
1
= �
It is a rational function of z. Note that every continuoustime system involving time delay is a
distributed system. This is not so in discretetime systems.
EXAMPLE 2.15 Consider the discretetime feedback system shown in Fig. 2.18(a). It is the
discrete counterpart of Fig. 2.5(a). If the unitsamplingtime delay element is replaced by its
transfer function ZI . then the block diagram becomes the one in Fig. 2.18(b) and the transfer
function from · to V can be computed as
. .
g(z) = = ¯
 ~ .· z  .
This is a rational function of : and is similar to (2.12). The transfer function can also be
obtained by applying the ztransform to the impulse response sequence of the feedback system.
As in (2.9), the impulse response sequence is
..·, = .·· 1, + .`·· 2, + ` ` = S .

· · »¦
.
The ztransform of · Ill ] i s zm. Thus the transfer function of the feedback system is
which yields the same result .
The discrete transfer functions in the preceding two examples are all rational functions
of ¯. This may not be so in general. For example, if
for » :: 0
for k = 1 , 2 . . . .
34 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
(a)
Unittime
delay
Figre 2.18 Discretetime feedback system.
Then we have
y[kJ
g(z) = ;1 + !;
2
+ .
3
+
. . . =  In( l ¯ ;1 )
.ì
(b)
It is an irrational function of z. Such a system is a distributed system. We study in this text only
lumped discretetime systems and their discrete transfer functions are all rational functions
of z.
Discrete rational transfer functions can be proper or improper. If a transfer function is
improper such as g(z) = ( Z
2
+ 2z  1 ) /(.  0. 5), then
which implies
or
H
z )
£(.)
Z
2
+ 2z  I
z 0. 5
y[k + I ]  0. 5y[k] = u[k + 2] + 2u[k + I]  u [k]
y[k + 1 ] = 0. 5y[k] + u [k + 2] + 2u [k + I ]  u [k]
It means that the output at time instant k + I depends on the input at time instant k + 2,
a future input. Thus a discretetime system described by an improper transfer

function is
not causal. We study only causal systems. Thus all discrete rational transfer functions will
be proper. We mentioned earlier that we also study only proper rational transfer functions
of s in the continuoustime case. The reason, however, is diferent. Consider g(s) = s or
y(t) = du(t)/dt. It is a pure diferentiator. If we defne the diferenliation as
duCt) . u(t + 6)  u (t)
v(t) = ¯= hm
•
dt D�U 6
where 6 > 0, then the output yet) depends on future input u (t + 6) and the diferentiator i s
not causal. However, if we defne the differentiation as
du(t) . u (t)  u (t 6)
v(t ) = ¯= hm  .
dt ^�U 6
then the output yet) does not depend on future input and the differentiator is causal. Therefore
in continuoustime systems, it is open to argument whether an improper transfer function
represents a noncausal system. However. improper transfer functions of s will amplify high
2. 6 Di screteTi me Systems 35
frequency noise, which often exists in the real world. Therefore improper transfer functions
are avoided in practice.
Statespace equations Every linear lumped discretetime system can be described by
x[k + I] = A[k]x[k] + B[k]u[k]
y[k] = C[k]x[k] + D[k]u[k]
(2.38)
where A, B. C, and D are functions of k. If the system is time invariant as well. then (2. 38)
becomes
x [k + I ] = Ax[k] + Bu[k]
y[k] = Cx[k] + Du[k]
where A. B. C. and D are constant matrices. Let x(z) be the ztransfor of x[k] or
Then we have
¯
x(.) = Z[x[k]] := S x[k]Zk
k
=O
¯
Z[x[k + I ] ] = S x[k + I ]Z
k = Z S x[k + I ]Z
(k
+
I
)
= z _x[ l] zI + x[O]  X[O]= z(x(z)  x[O])
1=1
Applying the ztransfor to (2. 39) yields
which implies
zx(z)  zx[O] = Ax(z) + Bu(z)
Yez) = Cx(.) + Du(z)
x(z) = (zl  A) I ZX[O] + (.  A) I Bu(z)
y(z) = C(zI  A) I ZX[O] + C(¨l  A) 
I
Bu(z) + Du(z)
(2. 39)
(2.40)
( 2.41)
They are the discrete counterparts of (2. 1 4) and (2. 1 5) . Note that there is an extra ¯ in front of
x[O]. If x[O] = 0, then (2.41 ) reduces to
y(z) = [ C(d  A)
I
B + D]u(z) (2.42)
Comparing this with the MIMO case of (2. 37) yields
G(z) = C(zl  A) I B + D (2.43)
This i s the discrete counterpart of (2. 1 6) . If the Laplace transform variable s is replaced by the
ztransfor variable ¨, then the two equations are identical.
36 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
EXAMPLE 2.16 Consider a money market account i n a brokerage frm. If the i nterest rate
depends on the amount of money in the account, it is a nonlinear system. If the i nterest rate i s
the same no matter how much money i s i n the account, then i t is a linear system. The account
is a timevarying system if the interest rate changes with time; a timeinvariant system i f the
i nterest rate is fixed. We consider here only the LTI case with interest rate r = . ·· per day
and compounded daily. The input u [k] i s the amount of money deposited into the account on
the kth day and the output yk] i s the total amount of money in the account at the end of the
kth day. If we withdraw money, then . ·, i s negative.
If we deposit one dollar on the first day (that is. .a· = I) and nothing thereafter
(u[k] = a k  I . 2 . . . . ). then ·a¡ = u O)  . and . . , = I + . · = . . ·
Because the muney is compounded daily, we have
y[2) = · · ¦ + y[ I ) · º. ·  y I ] · . . ·  . . ·· `
!
and. in general,
y[k] = . . ··
Because the input ( I . a . . J is an impulse sequence, the output is. by defnition. the impulse
response sequence or
./,  . . ··
and the inputDutput description of the account is
·
k
y[· = _ .·  ¬, ¬ = _ . ·· ¬·
n=0
The discrete transfer function is the ztransform of the impulse response sequence or
¯
g(z) = Zgk, ) = _ . ·· '·
·
 _ . . ·. ·
.  . . ··  . . ·
( 2. 44)
(2. 4· )
Whenever we use (2.44) or (2. 45), the initial state must be zero, or there is initially no money
in the account.
Next we develop a statespace equation to describe the account. Suppose v[k] is the total
amount of money at the end of the kth day. Then we have
y[k + I )  y[k + . ··/, + u[k + I ] = . . ···, + . · + I )
If we defne the state variable as x[k, : = . ·, then
 [k + I ] = 1 . 000I Sx[k] + uk + I ,
yk) = xkJ
(2.46)
( 2. 47)
Because of u[k + I], (2.47) i s not i n the standard form of (2. 39). Thus we cannot select
· . yk, as a state variable. Next we select a different state variable as
2. 7 Concl udi ng Remarks 37
i
²k, ·  y k]  . k)
.'�
" Substituting y[k
,
+ I ) = k + I ) + uk + I ] and y[k) = xk, + .k] into (2. 46) yields
�:
xk + I ]  I .OOI ·xk] + . . ·.·,
yk, = x[k) + .k,
This is in the standard form and describes the money market account.
(2.48)
The l inearization discussed for the continuoustime case can also be applied to the discrete
time case with only slight modifcation. Therefore its discussion will not be repeated.
2. 7 Concl udi ng Remarks
We introduced i n this chapter the concepts of causality, lumpedness. linearity. and time invari
ance. Mathematical equations were then developed to describe causal systems, as summarized
in the following.
System type
Distributed. linear
Lumped. linear
Distributed. linear.
timeinvariant
Lumped. linear,
timeinvariant
Internal description
x = A( t )x + B(t)u
y = C(t)x + DCt)u
x = Ax +Bu
y = Cx + Du
External description
y( f ) = ¡G(t. r) u( r) dr
'U
y(t) = ¡GU. r) u(r ) dr
ÍU
y( t )  f' G(t  r)u( r) dr
10
Y( S) = (( s)u( s). ((5) irrational
yet) = G(t  r)u( r) dr
Y( 5) = ((S )UI 5) . (( 5) rational
Distributed systems cannot be described by fnitedimensional statespace equations. Exteral
description describes only zerostate responses; thus whenever we use the description. systems
are implicitly assumed to be relaxed or their initial conditions are assumed to be zero.
We study in this text mainly lumped linear timeinvariant systems. For this class of
systems, we use mostly the timedomain description (A. B. C. Dl in the interal description
and the frequencydomain (Laplacedomain) description c·· in the exteral description.
Furthermore. we wi l l express every rational transfer matrix as a fraction of two polynomial
matrices, as we wi l l develop in the text. By so doing, all designs in the SISO case can be
extended to the multi variable case.
38 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
The class of lumped linear timeinvariant systems constitutes only a very small par of
nonlinear and linear systems. For this small class of systems, we are able to give a complete
treatment of analyses and syntheses. This study will form a foundation for studying more
general systems.
2.1 Consider the memoryless systems with characteristics shown in Fig. 2. 1 9, in which .
denotes the input and · the output. Which of them is a linear system? Is it possible to
introduce a new output so that the system in Fig. 2. l 9(b) is linear?
y y y

t
I ll
(a) ()
Figure 2.19
2.2 The impulse response of an ideal lowpass flter is given by
sin 2w· · ···
. · · = 2w
2w(t ···
(e)
for all · where w and ·· are constants. Is the ideal lowpass flter causal? Is it possible to
build the flter in the real world?
2.3 Consider a system whose input . and output y are related by
·· · = r..· ·· : =
.· · for · � .
o for t > .
where U is a fxed constant. The system is called a ··.»..·» ,·.·· which chops of
the input after time . Is the system linear? Is it timeinvariant? Is it causa]?
2.4 The input and output of an initially relaxed system can be denoted by · = u... where
u is some mathematical operator. Show that if the system is causal, then
r.· = r. u. = r. ur..
where r. is the truncation operator defned in Problem 2. 3. Is it true r. u . = u r. .·
2.5 Consider a system with input H and output · Three experiments are performed on the
system using the inputs Hj ·· .···. and ., · · for · :: O. In each case, the initial state
x(O) at time t = 0 is the same. The corresponding outputs are denoted by y" ·· and Y3 .
Which of the following statements are correct if x(O) ' O?
`
2.6
1. If .· , = . ÷ .· then ·· = ·. ÷ }2 .
2. If ., = 0 5. , ÷ .··. then ·, = 0. 5(y,  ···
3. If .\ = ., .·· then Y3 = y, ·.
Which are correct if x(O) = O?
Consider a system whose input and output are related by
·.·· =
�
·
··,.· · I ) if .· I ) ' 0
if .· I ) = 0
Problems 39
for all · Show that the system satisfes the homogeneity property but not the additivity
property.
2.7 Show that if the additivity property holds, then the homogeneity property holds for all
rational numbers . Thus if a system has some "continuity" property, then additivity
implies homogeneity.
2.8 Let . · :· = ..· ÷. : ÷.· for all ·. : and . Show that .. · r) depends only on ·  :
u.»· Defne  = · + r and · = ·  r and show that .· :·, = 0. ]
2.9 Consider a system with impulse response as shown in Fig. 2.20(a). What is the zerostate
response excited by the input .··· shown in Fig. 2.20(b).
S\lI
0 j Z
Figure 2.20
• !
I4I
 ,
2.10 Consider a system described by
j + 2y  3y = U .
u(:)
IÞI
What are the transfer function and the impulse response of the system?
2.11 Let :·· be the unitstep response of a linear timeinvariant system. Show that the impulse
response of the system equals e»··,e·
2.12 Consider a twoinput and twooutput system described by
40 MATHEMATICAL DESCRI PTI ONS OF SYSTEMS
D21 (p)YI (t l + D22 ( P)Y2 (t)  N2 1 ( P)II I (t) + Nn( p)u2 ( t)
where Nij and D'i are polynomials of P .  d/dt. What i s the transfer matrix of the
system0
2.13 Consider the feedback systems shown i n Fig. 2.5. Show that the unitstep responses of
the positivefeedback system are as shown in Fig. 2. 2 1 (a) for a = I and i n Fig. 2. 2 1 (b)
for a = 0
. 5. Show also that the unitstep responses of the negativefeedback system are
as shown in Figs. 2. 2 1 (c) and 2. 2 1 (d). respectively, for a = I and a  0. 5.
y( l)
o 1 2 3 4 6
(a)
O ± I
• r
o r�
4
 1
6
( e)
Figure 2.21
2.14 Draw an opamp circuit diagram for
4
(b)
I
6
0 1 2 3 4 5 6 7
(d)
y = [3 l O]x ¯ 211
0 = 0. 5
• r
• t
2. 15 Find state equations to describe the pendulum systems in Fig. 2.22. The systems are
useful to model one or twolink robotic manipulators. If e. and e2 are very small,
can you consider the two systems as linear"
2. 16 Consider the simplifed model of an aircraft shown in Fig. 2. 23. It i s assumed that the
aircraft i s in an equilibrium state at the pitched angle eo. elevator angle H(. altitude ho.
and cruising speed ¡0. It i s assumed that small deviations of e and N from eo and lI µ
generate forces /1 = k¡ and h = k2L1 as shown in the fgure. Let m be the mass of
the aircraft, I the moment of inertia about the center of gravity P, be the aerodynamic
damping. and h the deviation of the altitude from ho. Find a state equation to describe
the system. Show also that the transfer function from H to h. by neglecting the efect of
I, is
(a)
Figure 2.22
Figure 2.23
 Dl
> h(s) k; k, l,  k,bs
g(s ) = ¯¯
~ =

! rs) ms2 (bs + kl l i l
Probl ems 41
2.17 The soft landing phase of a lunar module descending on the moon can be modeled as
shown in Fig. 2.24. The thrust generated is assumed to be proportional to In. where m is
the mass of the module. Then the system can be described by lIl
.
i' = ktiz  mg. where
g is the gravity constant on the lunar surface. Defne state variables of the system as
XI = y, X2 = 5', ²( = m, and I = m. Find a statespace equation to describe the system.
2.18 Find the transfer functions from ll to YI and from )'1 to y of the hydraulic tank system
shown in Fig. 2.25. Does the transfer function from II to y equal the product of the two
transfer functions') Is this also true for the system shown in Fig. 2. 147 [Answer: No.
because of the loading problem in the two tanks in Fig. 2. 14. The loading problem is an
important issue in developing mathematical equations to describe composite systems.
See Reference [7] . ]
42 MATHEMATICAL DESCRIPTIONS OF SYSTEMS
!Trust ¯ km
I
Lunar surface
"
¦
& . . . . &
. . & & . & & . . .
'
. & . . . .
. . . & & & & & . .
Ó2
Figure 2.24
•
Figure 2.25
Q + y
2.19 Find a state equation to describe the network shown in Fig. 2.26. Find also its trans·
fer function.
u
(current
source)
1 F Figure 2.26
2.20 Find a state equation to describe the network shown in Fig. 2.2. Compute also its transfer
matrix.
2.21 Consider the mechanical system shown in Fig. 2.27. Let I denote the moment of inertia
of the bar and block about the hinge. It is assumed that the angular displacement ( is
very small. An exteral force H is applied to the bar as shown. Let y be the displacement
t
'
Problems 43
of the block. with mass m2 . from equilibrium. Find a statespace equation to describe
the system. Find also the transfer function from H to y.
Figure 2.27
+ "
� J
_
1�12 
LH3D!CI
/
Li near Al gebra
3. 1 I ntroduction
4
This chapter reviews 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 that wi l l be used subsequently
are introduced. Most results are developed intuitively i n order for the reader to better grasp
the ideas. They are stated as theorems for easy reference i n later chapters. However. no formal
proofs are given.
As we saw i n the precedi ng chapter. all parameters that arise in the real world are real
numbers. Therefore we deal only with real numbers. unless stated otherwise. throughout this
text. Let A. B. C. and D be. respectively. n x m. m x r. l x n. and r x p real matrices. Let a,
be the ith column of ·and bj the jth row of B. Then we have
and
AB = [al
a
2
. . . aml ¨ = al b
l + a2b
2
+ · · · + ambm ( 3. 1 )
bm
b
l
b
I
D
bo boD
BD =

D =

b�
n
b,�D
(3. 2)
(3. 3)
3. 2 Basi s. Representati on. and Orthonormal ization 45
These identities can easily be verifed. Note that a, h, i s an n x r matrix; it i s the product of an
n x I colunm vector and a I x r row vector. The product bi a, i s not defined unless n = r: it
becomes a scalar if n = r.
3. 2 Basi s, Representati on, and Orthonormal ization
Consider an ndimensional real linear space. denoted by '
n
. Every vector in '
"
i s an ntuple
of real numbers such as
·¯
x =
·�
To save space, we write it as x = [X
I
²_
. . .
x,, ] , . where the.prime denotes the transpose.
The set of vectors (X
I
. Xl . . . . xm) in '
"
is said to be linearly dependent if there exist
real numbers 01 . 02 . am. not all zero. such that
(3'+)
If the only set of a, for which (3. 4) holds is a ¡ = O. 01 = O . . . . . am = O. then the set of
vectors { X
I
. Xl . . . . . xm} is said to be linearly independent.
If the set of vectors in (3.4) is l inearly dependent. then there exists at least one ai . say.
01 . that is different from zero. Then ( 3.4) implies
I
X
I
= [a
2
x2 +a3x3 +
01
+ Cmxm]
where fi = a 1 / a
I
. Such an expression is called a l i near combination.
The dimension of ü linear space can be defned as the maximum number of linearly
independent vectors i n the space. Thus i n '", we can fnd at most n linearly i ndependent
vectors.
Basis and representation A set of linearly independent vectors in '" i s called a basis if
every vector i n '
n
can be expressed as a unique linear combination of the set. In '" . any set
of n linearly independent vectors can be used as a basis. Let { q I
. q2 . . . . . q,, } be such a set.
Then every vector x can be expressed uniquely as
(3 5)
Defne the n x n square matrix
Q : = [ql ql . " q
n
I (3. 6)
Then (3. 5) can be written as
46 L I NEAR ALGEBRA
(3. 7)
We call x = [al a2 . . . a
n
]' the representation of the vector x with respect to the basis
(ql . q2 . · . . . qn ) .
We will associate with every x
n
the following orthonormal basis:
I 0 0 0
0 I 0 0
0 0 0 0
il = h = inl = in = (3. 8)
0 0 I 0
0 0 0
With respect to this basis. we have
where In is the n x n unit matrix. In other words, the representation of any vector x with respect
to the orthonormal basis in (3. 8) equals itself.
EXAMPLE 3.1 Consider the vector x = [1 3]' in X
2
as shown in Fig. 3. 1 , The two vectors
ql = [3 I ]' and q2 = [2 2]' are clearly linearly independent and can be used as a basis. If we
draw from x two lines in parallel with q2 and ql , they intersect at ql and 2q2 as shown. Thus
the representation of x with respect to {ql , q2} is [I 2]' . This can also be verifed from
To fnd the representation of x with respect to the basis [q2, i2) . we draw from x two lines
in parallel with iz and q2 . They intersect at 0.Sq2 and 2h. Thus the representation of x with
respect to [q2, iz) is [0.5 2]'. (Verify. )
Norms of vectors The concept of nor is a generalization of length or magnitude. Any
realvalued function of x. denoted by I l xl l . can be defned as a norm if it has the following
properties:
1. I l xl l ? 0 for every x and I l xl l = 0 if and only if x = O.
2. I l axl l = l al l l xl l . for any real a.
3. I l
x
l + x2 1 1 C I I xl l l + I I x2 1 1 for every XI and x2.
I
3. 2 Basi s, Representati on, and Orthonormal i zati on
Figure 3.1 Different representations of
vector x.
The last inequality is called the triangular inequalit.
X
47
Let x = [x ¡ xn] ' . Then the norm of x can be chosen as any one of the following:
I l xl l l := ¯) �, I
i=l
I l
x
l l z : = v = �
I /Z
I l xl l " : = maXi I xd
They are called. respectively. I norm. 2 or Euclidean norm. and infnitenorm. The 2norm
is the length of the vector from the origin. We use exclusively. unless stated otherwise. the
Euclidean norm and the subscript 2 will be dropped.
In MATLAB. the norms just introduced can be obtained by using the functions
norm ( x, 1 ) . norrr ( x , 2 ) �norm ( x) . and norm ' J, inf ) .
Orthonormalization A vector x is said to be normalized if its Euclidean nonn is I orx'x = I .
Note that x'x is scalar and xx' is n x n. Two vectors X t and Xl are said to be orthogonal if
x; Xl = XSXI = O. A set of vectors Xi , i = I . 2 « . « « . U. is said to be orthonormal if
7 O if i f j
x, X
j
=
I if i = j
Given a set of linearly independent vectors e
l
. el, . . . . em. we can obtain an orthonormal
set using the procedure that follows:
48 LI NEAR ALGEBRA
u] : = e]
Uz : = e2  (q'] e2 )q]
ql
: = ul / l l ul l l
q2 : = u2 /1 1 u2 1 1
um : = em  L;:i (qiem)qk qm : = umil i umi l
The frst equation normalizes the vector e
l
to have norm 1 . The vector (q; e2)ql i s the projection
of the vector e2 along q] . Its subtraction from ez yields the vertical part Uz. It is then normalized
to 1 as shown in Fig. 3. 2. Using this procedure, we can obtain an orthononnal set. This is called
the Schmidt orthonormalization procedure.
Let A = [a] a2 . . . aml be an n x m matrix with m : n. If all columns of A or
{a; , i = I , 2 . . . . , m I are orthonormal, then
[
a
;
A
'
A
�
o,; [ 0, 0, ¨
a
m
where 1m is the unit matrix of order m. Note that, in general, AA' ' In . See Problem 3.4.
3. 3 Linear Algebraic Equations
Consider the set of linear algebraic equations
Ax = Y
(3. 9)
where A and y are, respectively, m x n and m x I real matrices and x is an n x I vector. The
matrices A and y are given and x is the unknown to be solved. Thus the set actually consists of
m equations and n unknowns. The number of equations can be larger than, equal to, or smaller
than the number of unknowns.
We discuss the existence condition and general form of solutions of (3. 9). The range
space of A is defned as all possible linear combinations of all columns of A. The rank of A is
°
� . . . « . » . . . . . , . . °2
°
. ¬
Figure 3.2 Schmidt orthonormization procedure.
3. 3 Li near Algebraic Equati ons 49
defned as the dimension of the range space or, equivalently, the number oflinearly independent
columns in A. A vector x is called a null vector of A if Ax = O. The null space of A consists
of all its null vectors. The nullit is defned as the maximum number of linearly independent
null vectors of A and is related to the rank by
Nullity (A) = number of columns of A  rank (A) (3. 1 0)
.......3.2 Consider the matrix
��=: [al a2 a3 a.l (3. 1 1)
where a; denotes the ith column of A. Clearly al and az are linearly independent. The third
column is the sum of the frst two columns or a] + az  a3 = O. The last column is twice the
second column. or 2a2  a. = O. Thus A has two linearly independent columns and has rank
2. The set {a] , a2 1 can be used as a basis of the range space of A.
Equation (3. 1 0) implies that the nullity of A is 2. It can readily be verifed that the two
vectors
(3. 1 2)
meet the condition An; = O. Because the two vectors are linearly independent, they form a
basis of the null space.
The rank of A is defned as the number of linearly independent columns. It also equals
the number of linearly independent rows. Because of this fact, if A is m x n, then
rank(A) : minIm, n)
In MATLAB. the range space. null space, and rank can be obtained by calling the functions
or:h. nul l. and rank. For example, for the matrix in (3. 1 1 ), we type
a= [ 0 J 1 2 ; 1 2 3 � ; 2 0 2 0 ] ;
ranK l a )
which yields 2 . Note that MATLAB computes ranks by using singularvalue decomposition
(svd), which will be introduced later. The svd algorithm also yields the range and null spaces
of the matrix. The MATLAB function R=orth ( a ) yields'
/IS K=
0. 3782 0.3084
0 8877 ¯0+ 1 468
0. 2627 0. 9399
1. Tis is obtained using �Í.¬ÅlPÜ Version 5. Earlier versions may yield diferent results.
(3. 1 3)
50 LI NEAR ALGEBRA
The two columns of R form an orthonormal basis of the range space. To check the orthonor
mality. we type R ' *R, which yields the unity matrix of order 2. The two columns in R are not
obtained from the basis [al . a2 } in (3. 1 1 ) by using the Schmidt orthonormalization procedure;
they are a byproduct of svd. However, the two bases should span the same range space. This
can be verifed by typing
rank ( [ al a2 Rl )
which yields 2. This confrms that [ aI , a2} span the same space as the two vectors of R. We
mention that the rank of a matrix can be very sensitive to roundof errors and imprecise data.
For example, if we use the fvedigit display of R in ( 3. 1 3), the rank of [ al a2 R J is 3. The
rank is 2 if we use the R stored in MATLAB, which uses 16 digits plus exponent.
The null space of (3. 1 1 ) can be obtained by typing nul l ( a ) , which yields
/E I�
0. 3434 0. 5802
0. 8384 0. 3395
(3. 14)
0.3434 0. 5802
0.2475 0. 4598
The two columns are an orthonormal basis of the null space spanned by the two vectors [nl . n2}
i n (3. 12). All discussion for the range space applies here. That is, rank ( [ nl n2 NJ )
yields 3 if we use the fvedigit display in ( 3. 1 4) . The rank is 2 if we use the N stored i n
MATLAB.
With this background, we are ready to discuss solutions of (3. 9). We use p to denote the
rank of a matrix.
: Theorem 3. 1
1. Given an m x n matrix A and an m x I vector y, an n x I solution x exists in Ax ~ y if and only
if y lies in the range space of A or, equivalently.
peA) = p ([A y])
where [A y] is an m x (n + I) matrix with y appended to A as an additional column.
2. Given A. a solution x exists in Ax = y for every y. if and only if A has rank m (full row rank).
The frst statement follows directly from the defnition of the range space. If A has full
row rank, then the rank condition in ( I ) is always satisfed for every y. This establishes the
second statement.
> Theorem 3.2 (Parameterization of all solutions)
Given an U x n matrix A and an m x I vector y. let x
p
be a solution of Ax = y and let k := n p (A)
be the nullity of A. If A has rank n (full column rank) or k = O. then the solution x
p
is unique. If k ¯ O.
then for every real Ci , i = 1 , 2, . . . . k. the vector

I
,
I
i
I
:
>
3. 3 Li near Algebraic Equations 51
x = x
p
+ Cl nl +
. . .
+ Cknk (3. 1 5)
i s a solution of Ax = y. where (nl ' . . . , nk ! i s a basis of the null space of A.
Substituting (3. 1 5) i nto Ax = y yields
k
Axp + SOi Ani = Ax
p
+ 0 = Y
i =l
Thus, for every Oi , (3. 1 5) i s a solution. Let x be a solution or Ax = y. Subtracting this from
Ax
p
= y yields
A(x  x
p
) = 0
which implies that x  x
p
is in the null space. Thus x can be expressed as in (3. 1 5). This
establishes Theorem 3. 2.
EXAMPLE 3.3 Consider the equation
0 I I
Ax = I 2 3
2 0 2
(3. 1 6)
This y clearly lies in the range space of A and x
p
= [0  4 0 0]' is a solution. A basis of the
null space of A was shown in (3. 1 2). Thus the general solution of (3. 1 6) can be expressed as
(3. 1 7)
for any real O I and 02.
In application, we will also encounter xA = y, where the U x n matrix A and the I x n
vector y are given, and the I x m vector x is to be solved. Applying Theorems 3 . 1 and 3.2 to
the transpose of the equation, we can readily obtain the following result.
Corollary 3.2
1. Given an m x n matrix A. a solution x exists in xA = y. for any y. if and only if A has full column
rank.
2. Given an m x n matrx A and an I x n vectory, let x
p
be a solution ofxA = y and let k = m peA) .
If k = O. the solution x
p
is unique. If k ¯ O. then for any O, i = 1 . 2 . . . . . k. the vector
x = x
p
+ Ol nl +
. . .
+ Oknk
is a solution of xA = y. where niA = 0 and the set {nl , . . . , nk ! is linearly independent.
In MATLAB, the solution of Ax = y can be obtained by typing A \y. Note the use of
backslash. which denotes matrix left division. For example. for the equation in (3. 1 6) . typing
52
L I NEAR ALGEBRA
a= [ O 1 1 2 ; 1 2 3 4 ; 2 0 2 O ] ; y= [  4 ;  3 ; O ] ;
a\y
yields [ 0  4 a 0 1 ' . The solution of xA = y can be obtained by typing y / A. Here we use
slash. which denotes matrix right division.
Determinant and inverse of square matrices The rank of a matrix is defned as the number
of linearly independent columns or rows. It can also be defned using the detenninant. The
deteninant of a I x I matrix is defned as itself. For n = 2. 3 ¡ . . . . the detenninant of n x n
square matrix A = [aii ) is defned recursively as. for any chosen j.
det A = S a'
J
c'i
Í
t
( 3. 1 8)
where aij denotes the entry at the i th row and jth column of A. Equation (3. 1 8) is called
the Laplace expansion. The number c'i is the cofactor corresponding to U¡¸ and equals
( I )'+
i
det Mij . where Mij is the (n ¯ 1 ) x (n ¯ 1 ) submatrix of A by deleting its i th row
and jth column. If A is diagonal or triangular. then det A equals the product of all diagonal
entries.
The determinant of any r x r submatrix of A i s called a minor of order r. Then the rank
can be defned as the largest order of all nonzero minors of A. In other words. if .has rank r.
then there is at least one nonzero minor of order r. and every minor of order larger than r is
zero. A square matrix is said to be nonsingular if its deteninant is nonzero. Thus a nonsingular
square matrix has full rank and all its columns (rows) are linearly independent.
The inverse of a nonsingular square matrix A = [a'
i
) is denoted by A¨' . The inverse has
the property AA 
I
= A 
I
A = I and can be computed as
_¡ Adj A I ,
A = = ¯[c'i) ( 3. 1 9)
det A det A
where cii is the cofactor. If a matrix is singular. its inverse does not exist. If .is : x 2. then
we have
A
1
: =
a
l l
a2l
( 3.20)
Thus the inverse of a 2 x 2 matrix i s very simple: interchanging diagoD.l entries and changing
the sign of ofdiagonal entries (without changing position) and dividing the resulting matrix
t the deteninant of A. In general. using ( 3. 1 9) to compute the inverse is compli(uted. If A is
triangular. it is simpler to compute its inverse by solving AA I = I. :lte that the inverse of a
triangular matrix is again triangular. The MATLAB function : :s computes the inverse of A.
Theorem 3.3
Consider Ax = y with A square.
1. If A is nonsingular. then the equation has a unique solution for every y and L�e solUlion equals A 
I
y.
In particular. the only solution of Ax = 0 is x = O.
'
I
I
!
3. 4 Si mi larity Transiormation 53
2. The homogeneous equation Ax = 0 has nonzero solutions i f and only i f A is singular. The number
of linearly independent solutions equals the nullity of A.
3. 4 Si mi l arity Transformati on
Consider an n x n matrix A. It maps '
n
into itself. If we associate with '
n
the orthonormal
basis {i l . h ¡ . . . . i
n
} in (3. 8). then the ith column of A is the representation of Aii with re
spect to the orthonormal basis. Now if w� select a diferent set of basis {ql . q2 • . . . ¡ qn J. then
the matrix A has a different representation A. It turs out that the i th column ofA is the representation
ofAqi with respect to the basis {q! . q
2
• « « « ¡ qn}. This is illustrated by the example that follows.
EXAMPLE 3.4 Consider the matrix
( 3. 21 )
Let b = [ 0 0 I }' . Then we have
Ab A
'
b = A(Abl =
It can be verifed that the following relation holds:
(3. 22)
Because the three vectors b. Ab. and A
2
b are linearly independent. they can be used as a basis.
We now compute the representation of A with respect to the basis. It is clear that
A(b) = [ b Ab A
"
b)
A(Ab) = [ b Ab A
2
b)
A(A
2
b) = [b Ab A
2
b) ¯�
where the last equation is obtained from ( 3.22). Thus the representation of A with respect to
the basis { b. Ab. A"b} is
( 3 23)
5 LI NEAR ALGEBRA
The preceding discussion can be extended to the general case. Let A be an n x n matrix. If
there exists an n x 1 vector b such that the n vectors b, Ab, . . . , A
n

I
b are linearly independent
and if
then the representation of A with respect to the basis {b, Ab,
o 0
o
o
o p
o 'a
o !n1
!n
This matrix is said to be i n a companion form.
Consider the equation
Ax = y
The square matrix A maps x i n '
n
into y in '
n
. With respect to the basis {ql ,
the equation becomes
where x and y are the representations of x and y with respect to the basis {ql ,
As discussed i n (3.7), they are related by
x = Qx y = Qy
with
Q = [ql q2
. . . qn l
an n x n nonsingular matrix. Substituting these into (3. 25) yields
Comparing this with (3.26) yields
A = Q
I
AQ or A = QAQI
(3.24)
(3. 25)
q2 , . . . , qn } ,
(3.26)
q2 , . . . , qn } .
(3.27)
(3.28)
(3.29)
This is called the similarit transformation and A and A are said to be similar. We write
(3.29) as
AQ = QA
or
A[ql q2 . . . qn l = [Aql Aq2 . .
, Aqn l = [ql q2
. . .
qn lA
3. 5 Diagonal Form and Jordan Form 55
This shows that the ith column of A is indeed the representation of Aqi with respect to the
basis {ql , q2 , . . . , qn } .
3. 5 Diagonal Form and Jordan Form
A square matrix A has diferent representations with respect to diferent sets of basis. In this
section, we introduce a set of basis so that the representation will be diagonal or block diagonal.
A real or complex number A is called an eigenvalue of the n x n real matrix A i f there
exists a nonzero vector x such that Ax = AX. Any nonzero vector x satisfying Ax = AX i s
called a (right) eigenvector of A associated with eigenvalue A. In order to fnd the eigenvalue
of A, we write Ax = AX = Alx as
(A  U)x = 0 (3.30)
where I is the unit matrix of order n. This is a homogeneous equation. If the matrix (A  AI) is
nonsingular, then the only solution of (3.30) is x = 0 (Theorem 3. 3). Thus in order for (3. 30)
to have a nonzero solution x, the matrix (A  AI) must be singUlar or have detenrunant a
We defne
t(A) = det(U ¯ A)
It i s a monic polynomial of degree n with real coeffcients and is called the characteristic
polynomial of A. A polynomial is called monic if its leading coefcient is I . If ;. is a root of
the characteristic polynomial, then the detenrunant of (A  AI) is 0 and (3.30) has at least one
nonzero solution. Thus every root of t(A) is an eigenvalue of A. Because t(A) has degree n,
the n x n matrix A has n eigenvalues (not necessarily all distinct).
We mention that the matrices
and their transposes
1
o
o
all have the following characteristic polynomial:

02
o
1
o
0)
o
o
t(A) = A
4
+ 0
I
A
3
+ 02A
2
+ 03)' + 04
These matrices can easily be formed from the coefcients of t(A) and are called companion
form matrices. The companionform matrices will arise repeatedly l ater. The matrix in (3.24)
i s in such a form.
56 LINEAR ALGEBRA
Eigenvalues of A are all distinct Let Ai . i = I . 2 ¡ . . . , H. be the eigenvalues of A and
be all distinct. Let q, be an eigenvector of A associated with Ai ; that is, Aq,  Ai q, . Then
the set of eigenvectors {q  , q2 . . . . , q,, } is linearly independent and can be used as � basis.
Let A be the representation of A with respect to this basis. Then the frst column of A is the
representation of Aql = Al ql with respect to {ql , q2 , . . . , qn } . From
�
Aql = Al ql  [ql q2 .
. . qn ]
0
o
we conclude that the frst column of A/i s [AI 0 `
. .
OJ'. The second column of A is the
representation of Aq
2 = A2q
2
with respect to { ql , q2 , qn } , that is, [0 Al 0
. . .
0]' .
Proceeding forward, we can establish
0 0
A2 0
A =
0 A3 (3. 3 1 )
0 0
This is a diagonal matrix. Thus we conclude that every matrix with distinct eigenvalues has
a diagonal �atrix representation by using its eigenvectors as a basis. Diferent orderings of
eigenvectors will yield diferent diagonal matrices for the same A.
If we defne
Q = [ql q2 .
. .
qn ]
(3. 32)
then the matrix A equals
( 3. 33)
as derived in (3. 29). Computing (3. 33) by hand i s not simple because of the need to
.
compute
the inverse of Q. However, if we know A. then we can verify (3. 33) by checking Q.= AQ.
EXAMPLE 3. 5 Consider the matrix
Its characteristic polynomial is
t(A) = det(AI ¯ A) = det �l
o
A
I
o

A  I
 A[A(A  I )  2] = (A  2) ( A + I )A
3. 5 Di agonal Form and Jordan Form 57
Thus A has eigenvalues 2, I , and O. The eignevector associated with A = 2 is any nonzero
solution of
(A  2I)q= �
o
�
ql  0
I
2
Thus ql  [0 I I ]' is an eigenvector associated with A = 2. Note that the eigenvector is not
unique. [0 a a]' for any nonzero real a can also be chosen as an eigenvector. The eigenvector
associated with A = I is any nonzero solution of
.¯ ( I )I)q2 =
which yields q2 = [0  2 I ]' . Similarly, the eigenvector associated with A  0 can be
computed as q3 = [2 I  I ]' . Thus the representation of A with respect to {q ¡ , q2 . q3} is
2 0
. =
�
� ( 3. 34)
It is a diagonal matrix with eigenvalues on the diagonal. This matrix can also be obtained by
computing
with
o
2
However, it is simpler to verify QA  AQ or
o
2
�
�� � � � � �
I 0 0 0 0 I I I
o
2
The result in this example can easily be obtained using MATLAB. Typing
yields
a= [ O 0 0 ; 1 0 2 ; 0 1 1 ] ; [ q, d] =eig ( a )
q = 0. 707�
0.707 1
o
0. 8944
0.4472
0. 8 1 86
0.4082
0.4082
d = �
�
I
( 3. 35)
where d is the diagonal matrix i n ( 3. 34). The matrix q is different from the Q in (3. 35); but their
corresponding columns differ only by a constant. This is due to nonuniqueness of eigenvectors
and every column of q is normalized to have norm I in MATLAB. If we type ei g ( a ) without
the lefthandside argument. then MATLAB generates only the three eigenvalues 2 ,  1 , O .
58 LI NEAR ALGEBRA
We mention that eigenvalues in MATLAB are computed from the characterstic polynomial.
Computing the characteristic polynomial using the Laplace expansion and then computing its
roots are not numerically reliable, especially when there are repeated roots. Eigenvalues are
computed in MATLAB directly from the matrix by using similarity transformations. Once
all eigenvalues are computed, the characteristic polynomial equals n (A  Ai ) . In MATLAB,
typing r=eig ( a) ) poly ( r ) yields the characteristic polynomial.
EXAMPLE 3.6 Consider the matrix
A �
� �3
Its characteristic polynomial is (A+ 1) (A
2
4A+ 1 3) . Thus A has eigenvalues 1 , 2±3j. Note
that complex conjugate eigenvalues must appear in pairs because A has only real coefcients.
The eigenvectors associated with and2+3j are, respectively, [1 0 0]' and [j 3+2j jr.
The eigenvector associated with A 2  3j i s [

j  3  2j  j]' , the complex conjugate
of the eigenvector associated with A 2 + 3 j. Thus we have
· 1 j
Q = 0 3 + 2j
· j

j
3 2j
and A � 2 + 3j
·
The MATLAB function [ q, d] =eig ( a ) yields
and
0.2582j
q 0 0. 7746 + 0. 5 1 64j
o 0.2582j
�
·
2 + 3j
·
0.2582j
0.7746  0. 5 1 64j
0. 2582j
All discussion in the preceding e
xample applies here.
(3. 36)
Complex eigenvalues Even though the data we encounter in practice are all real numbers.
complex numbers may arise when we compute eigenvalues and eigenvectors. To deal with this
problem, we must extend real linear spaces into complex linear spaces and permit all scalars
such as Ci in (3.4) to assume complex numbers. To see the reason, we consider
Av =
+
2
j
[
v 0
1  j
(3.37)
If we restrict v to real vectors, then (3.37) has no nonzero solution and the two columns of
A are linearly independent. However, if v is permitted to assume complex numbers, then
v [2 1  j]' is a nonzero solution of (3.37). Thus the two columns of A are linearly
dependent and A has rank This is the rank obtained in MATLAB. Therefore, whenever
complex eigenvalues arise, we consider complex linear spaces and complex scalars and
3. 6 Functi ons of a Square Matrix 59
transpose is replaced by complexconjugate transpose. By so doing, all concepts and results
developed for real vectors and matrices can be applied to complex vectors and matrces.
Incidentally, the diagonal matrix with complex eigenvalues in (3. 36) can be transformed into
a very usefl real matrix as we will discuss in Section 4.3. 1 .
Eigenvalues of A are not all distinct An eigenvalue with multiplicity 2 or higher i s called a
eigenvalue. In contrast, an eigenvalue with multiplicity is called a eigenvalue.
If A has only simple eigenvalues, it always has a diagonalform representation. If .has repeated
eigenvalues, then it may not have a diagonal form representation. However, it has a block
diagonal and triangularform representation as we will discuss next.
Consider an H X H matrix A with eigenvalue A and multiplicity H. In other words, A has
only one distinct eigenvalue. To simplify the discussion, we assume H = 4. Suppose the matrix
(A  AI) has rank  1 3 or, equivalently. nullity 1 ; then the equation
(A  AI)q 0
has only one independent solution. Thus A has only one eigenvector associated with A. We
need n 3 more linearly independent vectors to form a basis for '
4
The three vectors
q2, q3 , q4 will be chosen to have the properties . Alfq2 0, (A  AI)
3
q3 0, and
(A  AI)
4
q4 = O.
and
A vector v is called a of grade if
(A  AI)"v 0
(A  AI)
n
l v � 0
If they reduce to (A AI)v 0 and v � 0 and v is an ordinary eigenvector. For H 4,
we defne
V3 . AI)V4 = (A  AI)v
V2 (A  AI)V3 . AI)2v
VI (A  AI)V2 . AI)
3
v
They are called a chain of generalized eigenvectors of length H 4 and have the properties
(A  AI)vl 0, (A  AI)
2
V2 0, (A  AI)
3
V3 0, and .¯ AI)
4
V4 O. These vectors,
as generated, are automatically linearly independent and can be used as a basis. From these
equations, we can readily obtain
AVI AVI
AV2 V
I + AV2
AV3 V2 + AV3
AV4 V3 + AV4
60 LI NEAR ALGEBRA
Then the representation of A with respect to the basis {VI , V2, V3 , V4} is
A 1 0 0
o A 1 0
J
:
=
0 0 A 1
o 0 0 A
(3. 38)
We verify this for the frst and last columns. The frst column of J is the representation of
AVI = AVI with respect to {VI , V2, V3 . V4 } . which is [A 0 0 OJ'. The last column of J is the
representation of AV4 = V3 + AV4 with respect to { VI , V2 , V3 , V4}. which is [0 0 I AJ' . This
verifes the representation in (3.38). The matrix J has eigenvalues on the diagonal and I on the
superdiagonal. If we reverse the order of the basis, then the I 's will appear on the subdiagonal.
The matrix is called a block of order = 4.
If ( A  AI) has rank H  2 or, equiv
�
ently, nullity 2, then the equation
(A  AI)q = 0
has two linearly independent solutions. Thus A has two linearly independent eigenvectors and
we need  2) generalized eigenvectors. In this case, there exist two chains of generalized
eigenvectors { VI , V2, . . . , vd and {UI , U2 . . . . , ut l with k + I = If VI and UI are linearly
independent, then the set of vectors { VI , . . . , Vk . UI , . . . , ud is linearly independent and can
be used as a basis. With respect to this basis. the representation of A is a block diagonal matrix
of form
A = diag{JI . 12 }
where JI and J2 are, respectively, Jordan blocks of order k and I .
Now we discuss a specifc example. Consider a 5 x 5 matrix A with repeated eigenvalue A ¡
with multiplicity 4 and simple eigenvalue 12. Then there exists a nonsingular matrix Q such that
.' = Q
I
AQ
assumes one of the following forms
A
' 0 0
A' 0 0
.
�
A
I 0
A
,
�
AI 0
0 AI 0 AI 0
0 0 A
I 0 0 AI
0 0 0 0 0 0
A
'
0 0
A
'
0 0
A
,
�
AI 0 0
A, = �
AI 0 0
0 AI 0 AI 0
0 0 AI 0 0 A
I
0 0 0 0 0 0
A
' 0 6 0
.
=
AI 0 0
0 AI 0 ( 3. 39)
0 0 AI
0 0 0
3. 6
3. 6 Functions of a Square Matrix 61
The frst matrix occurs when the nullity of ( A  A I I ) is I . If the nullity is 2, then A has two
Jordan blocks associated with AI ; it may assume the form in A2 or in A3. If (A Al l) has nullity
3, then A has three Jordan blocks associated with AI as shown in ., Certainly. the positions
of the Jordan blocks can be changed by changing the order of the basis. If the nullity is +. then
A is a diagonal matrix as shown in As. Al l these matrices are triangular and block diagonal
with Jordan blocks on the diagonal; they are said to be in Jordan form. A diagonal matrix is a
degenerated Jordan form: its Jordan blocks all have order I . If A can be diagonalized. we can
use [ q, d J ¯ e i g ( a ) to generate Q and A as shown in Examples 3.5 and 3. 6. If A cannot be
diagonized. A is said to be and [ q, d 1 ¯e i g ( a ) will yield an incorrect solution. In
this case. we may use the MATLAB function [ q, d 1 ~ j ordan ( a ) . However, j orca: will
yield a correct result only if A has integers or ratios of small integers as its entries.
Jordanform matrices are triangular and block diagonal and can be used to establish
many general properties of matrices. For example, because det (CD) = det C det D and
det Qdet Q
I
= det I = I, from A = QAQ
I
. we have
det A = det Qdet Adet Q
1 = det A
The determinant of A is the product of all diagonal entries or. equivalently, all eigenvalues of
A. Thus we have
det A = product of all eigenvalues of A
which implies that A .
We discuss a useful property of Jordan blocks to conclude this section. Consider the
Jordan block in (3. 38) with order 4. Then we have
0
�
0 I
0 I
(J  AI)2 =
0 0
(J  AI) =
0 0 0 0
0 0 0 0
0 0
(J  AI)'
;
=
0 0
(3.40)
0 0
0 0
and (J  U)k = 0 for k :: 4. This is called
Functi ons of a Square Matrix
This section studies functions of a square matrix. We use Jordan form extensively because
many properties of functions can almost be visualized in terms of Jordan form. We study frst
polynomials and then general functions of a square matrix.
Polynomials of a square matrix Let A be a square matrix. If k is a positive integer. we
defne
Ak := AA ·
. .
A (k terms)
62 LI NEAR ALGEBRA
and AO = I. Let be a polynomial such as =
 6 0r  3) . Then
·(A) is defned as
I(A) = A
A
 61 or I(A) = (A I. (A  31)
If A is block diagonal, such as
A =
� ·
where AI and A2 are square matrices of any order, then it is straightforward to verify
Ak =
�� :�
]
and ·(A.=
·
�
·
1(�2)
]
(3.41 )
Consider the similarity transformati�n .= oA
Q
or A = o.o·
Because
A
k
= o.o
. o.o . o.o
.= o.o
we have
I(A) = o·..o
or ·..= o
·. o 3.
A polynomial is a polynomial with I as its leading coeffi cient. The
of A is defined as the monic polynomial of least degree such that "(A) = O.
Note that the 0 is a zero matrix of the same order as A. A direct consequence of (3.42) is
that I(A) = 0 if and only if ·.,= o. Thus A and .or, more general, all similar matrices
have the same minimal polynomial. Computing the minimal polynomial directly fom A is not
simple (see Problem 3. 25); however, if the Jordanform representation of A is available, the
minimal polynomial can be read out by inspection.
Let be an eigenvalue of A with multiplicity That is, the characterstic polynomial
of A is
= det(AI  A) =  ·n
Suppose the Jordan form of A is known. Associated with each eigenvalue, there may be one or
more Jordan blocks. The .of denoted by ii , is defned as the largest order of all Jordan
blocks associated with Clearly we have ii :: H¡ . For example, the multiplicities of AI in all
fve matrices in ( 3. 39) are their indices are, respectively, 3, 2, and I . The multiplicities
and indices of in all fve matrices in (3. 39) are all I . Using the indices of all eigenvalues,
the minimal polynomial can be expressed as
with degree i = L ii :: L = = dimension of A. For example, the minimal polynomials
of the fve matrices in (3. 39) are
"I =
= (A AI )2 (A 
vs =
=
. =
3. 6 Functi ons of a Square Matrix 63
Their characteristic polynomials, however, all equal
We see that the minimal polynomial is a factor of the characteristic polynomial and has a degree
less than or equal to the degree of the characteristic polynomial. Clearly, if all eigenvalues of
A are distinct, then the minimal polynomial equals the characteristic polynomial.
Using the nilpotent property in we can show that
"(A) = 0
and that no polynomial of lesser degree meets the condition. Thus "(A) as defned is the
minimal polynomial.
´
Theorem 3.4 (CayleyHami lton theorem)
Let
be the characteristic polynomial of A. Ten
t(A) = An a
1
An
1
a
n

1
A a'll = 0 (3" +3)
In words, a matrix satisfes its own characteristic polynomial. Because ¯ iii , the
characteristic polynomial contains the minimal polynomial as a factor or 6 (i.) = "().) h().)
for some polynomial heAl. Because "(Al = 0, we have A.= ", (A)h (A) = 0 , heAl = 0,
This establishes the theorem. The CayleyHamilton theorem implies that An can be written as
a linear combination of {I. A, , . . , A,,
l
} , Multiplying (3.43) by A yields
An+
1
a
l
An
a
n
_
I
A2 a
n
A = O · A = 0
which implies that An+
1 can be written as a linear combination of { A, A2
, . . . • An } , which,
in tur, can be written as a linear combination of {I. A, . . . , AnI } . Proceeding forward. we
conclude that, for any polynomial no matter how large its degree is, I(A) can always
be expressed as
(3.44)
for some In other words, every polynomial of an x matrix A can be expressed as a linear
combination of { I. A,  , . , An
I
} . If the minimal polynomial of A with degree i is available,
then every polynomial of A can be expressed as a linear combination of { I. A. ' . . , A"
I
} . This
is a better result. However, because i may not be available, we discuss in the following only
(3.44) with the understanding that all discussion applies to i.
One way to compute (3. 44) i s t o use long division to express 1(1. ) as
= q(). . ( 3.45)
where is the quotient and heAl is the remainder with degree less than n. Then we have
I(A) = q (A)t(A) h eAl = qAO + heAl = h eAl
6 LI NEAR ALGEBRA
Long division is not convenient to carry out if the degree of f (A) is much larger than the degree
of 6(A). In this case, we may solve h eAl directly from (3.45). Let
heAl : = f0 fI A
. . .
+ fn _I An
1
where the H unknowns fi are to be solved. If all H eigenvalues of A are distinct. these fi can
be solved from the n equations
f(A, ) = q(Ai ) 6(Ai ) + h (A, ) = h(A, )
for = I . 2 . . . . . H. If A has repeated eigenvalues, then (3.45) must be diferentiated to yield
additional equations. This is stated as a theorem.
> Theorem 3.5
'
i
We are given f(A) and an n X H matrix A with characteristic polynomial
6(A) = (A  A, J"'
1 =1
where H = L;'I H, . Defne
h eAl : = f0 fI A
.
'
.
+ fn_I A
n

1
It is a polynomial of degree n I with H unknown coeffcients. These H unknowns are to be solved
from the following set of H equations:
f
(/l
(A; ) = h
(/l
(A; ) for 1 = O. 1 , . . . , ni I and i = I , 2, . . . , U
where
and h(l
l
(Ai ) is similarly defned. Then we have
f(A) = h eAl
and heAl i s said t o equal f(A) on the spectrum of A.
EXA
MPLE 3.7 Compute A l
l with
In other words, given f(A) = Al l. compute f(A). The characteristic polynomial of A is
MA) = A 2 2A + I = (A 1 )2 Let heAl = f0 + fI A. On the spectrum of A. we have
f( I ) = h( I ) :
j'( I ) = h' ( I ) :
( _ I )
I
I = f0 ¯ f1
(
_
1 )99 = f1
Thus we have f1 = f0 = I f1
h eAl =  l OA. and
A
l
l = f0I fI A =  iOOA
=
I
o I I
3 . 6 Functi ons of a Square Matrix
I
=
2
65
Clearly A
l
l can also be obtained by multiplying A times. However. it is simpler to use
Theorem 3. 5.
Functions of a square matrix Let f(A) be any function, not necessarily a polynomial. One
way to defne f (A) is to use Theorem 3. 5. Let h (A) be a polynomial of degree n = I , where H
is the order of A. We solve the coeffcients of heAl by equating f(A) = heAl on the spectrum
of A. Then f(A) is defned as h eAl ·
EX
A
MPLE 3.8 Let
¯´
AI = I
I 3
Compute e
A
1 ' . Or, equivalently. if fp.) = e)" , what is f(AJ l ?
The characteristic polynomial of AI is (A  1 )1 (A  2) . Let h eAl = f0 fI A  (
2
).
2
Then
f( l ) = h( l ) :
t( 1 ) = h' ( l ) :
f(2) = h (2) :
e' = f0 f1 + f2
te' = f
1 + 2f
2
e2' = f0 2f1 + 4f1
Note that, in the second equation, the differentiation is with respect to A, not t. Solving these
equations yields f0 = 1te' el', f1 = 3te' 2e'  2e2' , and f
2
= el'  e'  te' . Thus we
have
e
A
1 ' = heAd = (2te' + e2')I + (3te' 2e'  2e2')A
I
EX
A
MPLE 3.9 Let
2e'  el'
e
� le'
2el'
(el'  e'  te')Ar =
el' ¯ e' 2el'  e'
A
Z = ��
2
Compute e
A
: ' . The characteristic polynomial of Az is (A  1 )
2
(A ¬ 2) , which is the same as
for AI . Hence we have the same heAl as in Example 3. 8. Consequently. we have
2e'  el' 2te' 2e' 0 2e2'
eA:' = h eAl) = e'
e2
1
_ e
t
tel 2e2t  e[
66 LI NEAR ALGEBRA
'
EXAMPLE 3.10 Consider the Jordan block of order 4:
A t
, 0
A =
o
o
o
(3.46)
Its characteristic polynomial is (A Ad4. Although we can select h (A) as /0 /1 A + fhA+ fh A3,
it i s computationally simpler to select h (A) as
This selection is prmitted because h (A) has degree (n  1 ) = 3 and n = 4 independent
unknowns. The condition I(A) = hO,) on the spectrum of A yields immediately
Thus we have
,
J"
(Ad 1
(
3
)
(AI )
f
o = I(AI l . /1 = J
'
(AI l . /
2
= 
2
' , /3 = 
3

!

I(A) = I(A, )I +
1
'
I l
(A ¯ Al l) + � (.. ¬ AI I)
2
+
1
A I )
(A  AI
1
)3
Using the special forms of (A  A I I)k as discussed i n (3. 40), we can readily obtain
, 0 I(A, ) 1'(Ad/ l ! J(A d/2
!
I(AI ) /, (AI )/ I ! J(Ad/2
' 1(3) (A I )/3 !
I(A) =
0 0 I(AI l /, (A,
) / I '
If I(A) = eAt , then
o 0 0 I(AI l
teAl !
t"eA ,
t
/2!
At
0
e =
eAj t
teA, t
t"eA, t
/2 !
""/J!
0 0 eAj t
te
Al t
0 0 0 e
A 1 t
(3. 47)
(3. 48)
Because functions of A are defned through polynomials of A, Equations (3. 41 ) and (3. 42)
are applicable to functions.
EXAMPLE 3.11 Consider
0 0
A
�
�
AI 0
0 AI 0
0 0 A2
0 0 0
3. 6 Functi ons of a Square Matri x 67
It is block diagonal and contains two Jordan blocks. If I(A) = eA
t
, then (3. 41 ) and (3. 48)
imply
teA
J
t
tZeA
l
t
/2 ! 0
,
" =
�
eAJ t
te
Al t
0
0 e
At t
0
0 0 e
A2t
0 0 0 e
A�f
If I(A) = (s  A) 
I
, then ( 3.4 1 ) and (3.47) imply
I I
0 0
(5  AI ) (5  Ad
2
(5  AI l3
I I
0 0 0
(s  A I )
2
(s  Al l
(sl  A) 
I
= 0 0
I
(5  Al l
0 0 (3.49)
I
0 0 0

(5  A2) " (s  A")
I
0 0 0 0
(5  A
2
)
Using power series The function of A was defned using a polynomial of fnite degree. We
now give an altematiw defnition by using an infnite power series. Suppose 1(;) can be
expressed as the power series
¯
I(A) = ¯/i A
t=0
with the radius of convergence p. If all eigenvalues of A have magnitudes less than p, then
I (A) can be defned as
O
I(A) = ¯/i Ai
(3. 50)
t=0
Instead of proving the equivalence of this defnition and the defnition based on Theorem 3. 5,
we use ( 3. 50) to derive (3. 47).
EXAMPLE 3.12 Consider the Jordanform matrix A i n (3. 46). Let
then
, J(AI l ,
I(A) = I(AI l + I (AI l (A  Al l
�
(A  AI l  +
. .
I,
I
C )
I(A) = 1(1. 1 )1 1' (AI l (A  Al l)
.
.
.
+ A
I
(A  Al l)n1 + . . .
(n ¬  ) '
68 LI NEAR ALGEBRA
Because (A  A
I
1/ = 0 for k :: n = as discussed in ( 3.40). the infinite series reduces
immediately to ( 3. 47). Thus the two defnitions lead to the same function of a matrix.
The most important function of .is the exponential function eAr
.
Because the Taylor
series
, A
�
t� ).
n
r"
eAr = I + AI + ¯+
. . .
+ ¯+ .
converges for all fnite i. and I, we have
i
2!
(3. 5 1 )
This series involves only multiplications and additions and may converge rapidly: there
fore it is suitable for computer computation. We list in the following the program in MATLAB
that computes (3. 5 1 ) for I = I :
��nction E= exp�2 ( A )
2= z eros ( s i z e ( .. / ) ;
r=eye ( s i z e ( A) i ;
K= : ;
��i l e norm ( E+F S , l » O
E= E+ F;
l=/+l/ K;
K=K+ 1 ;
In the program, E denotes the partial sum and F is the next term to be added to E. The frst line
defnes the function. The next two lines initialize E and F Let C¡ denote the kth term of (3. 5 1 )
with = 1 . Then we have
= .k)c, for k = 1 , 2 . . . . . Thus we have F = A + F Ik. The
computation stops i f the I norm of E + F  E, denoted by norm(E + F  E. I ) . is rounded
to 0 in computers. Because the algorithm compares F and E, not F and O. the algorithm uses
norm(E + F  E, I ) instead of norm(F. I ) . Note that norm(a. l ) is the I norm discussed in
Section 3.2 and will be discussed again in Section 3.9. We see that the series can indeed be
programmed easily. To improve the computed result, the techniques of scaling and squaring can
be used. In MATLAB, the function expm2 uses ( 3. 5 1 ). The function ex;:, or eX9m1 . hO\vever,
uses the socalled Pade approximation. It yields comparable results as exp:2 but requires only
about half the computing time. Thus expI, is preferred to exp,2. The function expT.3 uses
Jordan form. but it will yield an incorrect solution if a matrix i s not diagonalizable. If a c1osed
form solution of eAr i s needed, we must use Theorem 3. 5 or Jordan form to compute e·\r .
We derive some important properties of c to conclude this section. Using (3. 5 I ), we
can readily verify the next two equalities
( 3 52)
( 3. 53)
(3.54)
3. 6 Functions of a Square Matri x 69
To show ( 3.54). we set 12 = ¬î ¡ . Then ( 3. 53) and (3. 52) imply
which implies (3. 54). Thus the inverse of eAr can be obtained by simply changing the sign of
I. Differentiating term by term of ( 3. 5 1 ) yields
Thus we have
( 3. 55)
This i s an i mportant equation. We mention that
( 3.56)
The equality holds only i f .and scommute or .s= s.This can be verifed by direct
substitution of ( 3.5 1 ).
The Laplace transform of a function ! (t) is defined as
j(s) : = .[ f(t)] = !(t)e"dt
It can be shown that
.
¸
[
= S( k+l )
Taking the Laplace transform of ( 3.5 I ) yields
¯ ¯
.[e
Ar
] =
·
.
=
5
1
2
)
sI A)k
Because the infnite series
¯
2
) s1 A)k = 1 + S I A + s
2
A
2
+
. . .
= ( l  SI A)I
k=O
converges for I s I A I < 1 . we have
¯
S  I _(s I
Al = s I I + ,
.
.+ sJ A
2
+ . . .
k=O
and
( 3.57)
70 LI NEAR ALGEBRA
(3. 58)
Although in the derivation of (3. 57) we require ·to be suffciently large so that all eigenvalues
of ·
A have magnitudes less than I, Equation (3. 58) actually holds for all ·except at the
eigenvalues of A. Equation (3. 58) can also be established from (3. 55). Because L[df(t)/dt] ¬
sL[J(t)]  frO), applying the Laplace transform to (3. 55) yields
sL[eA1
]
 eO ^ AL[e
A
1]
or
which implies (3. 58).
3. 7 Lyapunov Equati on
Consider the equation
AM+ MB = C
(3.59)
where A and B are, respectively, n x n and m x m constant matrices. In order for the equation
to be meaningful, the matrices M and C must be of order n x m. The equation is called the
Lyapunov equation.
The equation can be written as a set of standard linear algebraic equatio/ls. To see this,
we assume H = 3 and m ~ and write (3.59) explicitly as
al l aI 2 al 3
a21
a2l a2l
all al2 all
el l CI 2
= C2I cn
Cl i Cl2
ml l
mZl
ml l
m12
mi l
m22 + mIl
ml2
mll
m1 2
bl I
mZ2
b21
m32
bl 2
b22
MUltiplying them out and then equating the corresponding entries on both sides of the equality,
we obtain
al l + bl l al `
al 3
bII
0 0
a21 an + bl l a2l
0 bII
0
al l
al2
all + bl l
0 0 b21
bl2
0 0 a
l l
+ b22 an
al l
0 bI 2
0 a21
a22 + b22 a2l
0 0 bi Z
al l
al2
all +b22
3. 8 Some Useful Formul as
71
mi l CI I
m21 e2 l
m3 1 Cl i
x
mI 2 el l
m22 e22
1n32
Cl2
This is indeed a standard linear algebraic equation. The matrix on the preceding page is a
square matrix of order n x m ~ 3 x * 6.
Let us defne 5(M) . ¯ AM + MB. Then the Lyapunov equation can be written as
.
A(M) = C. It maps an nmdimensional linear space into itself. A scalar I is called an
eigenvalue of A if there exists a nonzero M such that
A(M) = IM
Because 5 can be considered as a square matrix of order nm, it has nm eigenvalues 17b for
k = . . . , nm. It turs out
Ik ¯ Ai + Jj for ~
H, j ~ . . . , m
where Ai, = . . , . , n, and Jj' j = m, are, respectively. the eigenvalues of
A and B. In other words, the eigenvalues of A are all possible sums of the eigenvalues of A
and B.
We show intuitively why this is the case. Let U be an n x right eigenvector of .associated
with Ai ; that is, Au = AiU. Let v be a 1 x m left eigenvector of B associated with Jj : that is,
vB ~ vJj . Applying A to the n x m matrix uv yields
5(uv) = Auv + uvB = AiUV + UVJj = (Ai + Jj)uv
Because both U and v are nonzero, so is the matrix uv. Thus (Ai + Jj ) is an eigenvalue of A.
The determinant of a square matrix is the product of all its eigenvalues. Thus a matrix
is nonsingular if and only if it has no zero eigenvalue. If there are no i and j such that
Ai + J j = then the square matrix in (3.60) is nonsingular and. for every C, there exists a
unique M satisfying the equation. In this case, the Lyapunov equation is said to be nonsingular.
If A, + Jj = 0 for some and j, then for a given C, solutions may or may not exist. If C lies
in the range space of jt then solutions exist and are not unique. See Problem 3. 32.
The MATLAB function m= lyap ( a, b,  c ) computes the solution of the Lyapunov
equation in (3. 59).
3. 8 Some Useful Formulas
This section discusses some formulas that will be needed later. Let A and B be m x n and
n x p constant matrices. Then we have
p(AB) ::min(p(A), pCB»�
where p denotes the rank. This can be argued as follows. Let pCB) O. Then B has O
linearly independent rows. In AB, A operates on the rows of B. Thus the rows of AB are
72 LINEAR ALGEBRA
linear combinations of the rows of B. Thus AB has at most a linearly independent rows. In
AB, B operates on the columns of A. Thus if A has l i nearly independent columns, then
AB has at most f linearly independent columns. This establishes Consequently, if
A = B
I
B2B3 . .
" then the rank of A i s equal to or smaller than the smallest rank of Bi .
Let A be m X lI and l et C and D be any 1 x 1 and m x m nonsingular matrices. Then we
have
p(AC) = peA) = p(DA)
In words, the rank of a matrix will not change after pre or postmultiplying by a nonsingular
matnx. To show we defne
P : = AC
Because peA) :: mi nim. 11) and piC) = 1.we have peA) :: piC). Thus implies
pep) :: mi n( p(A) , piC»� :: peA)
:ext we write as A = PC
I
. Usi ng the same argument, we have peA) :: pcP). Thus we
conclude p(P) = piA) . A consequence of is that the rank of a matrix wi l l not chance
by elementary operations. Elementary operations are multiplying a row or a column by
O
a
nonzero number, interchanging two rows or two columns, and adding the product of
one row (column) and a number to another row (column). These operations are the same as
muilIplY1l1g nonsingular matrices. See Reference p.
Let A be m x 1 and B be 1 x m. Then we have
det ( lm + AB) = det(l" + BA)
where 1m is the unit matrix of order U. To show let us defne
We compute
and
Q
=
�� I�
[
NP =
1m + AB 0
B I
n
A
I" + BA
A
I"
B
�
cause N and Q are block triangular. their determinants equal the products of the determinant
ot their blockdiagonal matrices or
det N = det 1m
.
det I
n
= I = det Q
Likewise. we have
det ( NP) = det(lm + AB det(QP) = det(l" + BA)
Because
3.9 Quadrati c Form and Posi tive Defi niteness 73
det(NP) = det N det P = det P
and
det(QP) = det Q det P = det P
we conclude det(lm + ABl = det(l" + BA).
I n N, Q, and P, i f I" . I"" and B are replaced. respectively, by fl". fl", . and B. then
we can readily obtain
So det(sl",  AB) = s'" deUsI
"
 BA) ( 3. 65)
which implies, for 1 = III or for 1 x 1 square matrices A and B,
det(sI"  AB) = det(sl"  BA)
They are useful formulas.
3,9 Quadrati c Form and Positive Definiteness
An 1 X 1 real matrix M is said to be symmelric if its transpose equals itself. The scalar function
x'Mx, where x is an 1 x I real vector and M' = M, is called a qlladralicjorm. We show that
all eigenvalues of symmetric M are real.
The eigenvalues and eigenvectors of real matrices can be complex as shown in Example
Therefore we must allow x to assume complex numbers for the time being and consider the
scalar function x*Mx, where x· is the complex conjugate transpose of x, Taking the complex
conjugate transpose of x'�lx yields
(x*Mx) ' = x' lI'x = x'II'x = x'�1x
where we have used the fact that the complex conjugate transpose of a real M reduces to
simply the transpose. Thus x'lIx i s real for any complex x, This assertion i s not true if l
is not symmetric. Let A be an eigenvalue of M and v be its eigenvector; that i s. Mv = t.V.
Because
v'Mv = v'Av = A( Y'V)
and because both v*Mv and v* v are real, the eigenvalue ; must be real. Thi s shows that al l
eigenvalues of symmetric :1 are real. After establishing this fact. we can retur our study to
exclusively real vector x.
We claim that every symmetric matrix can be diagonalized using a si mi l arity transfor
mation even it has repeated eigenvalue ;, To show this. we show that there is no generalized
eigenvector of grade 2 or higher. Suppose x is a generalized eigenvector of grade 2 or
Consider
(M  ;1)2x = 0
(M  ;I)x .
74 L I NEAR ALGEBRA
[ (M

)
.
I) x]
'
(M

AI)x = x
'
(M
'

AI
'
) (M

AI)X = x
'
(M

AI)
2
X
which is nonzero according to (3. 68) but is zero according to (3.67). This is a contradiction.
Therefore the Jordan form of M has no Jordan block of order 2. Similarly, we can show that
the Jordan form of 1 has no Jordan block of order 3 or higher. Thus we conclude that there
exists a nonsingular Q such that
M
= QDQ
I
(3. 69)
where
D is a diagonal matrix with real eigenvalues of
M
on the diagonal.
A square matrix A is called an .if all columns of A are orthonormal.
Clearly A is nonsingular and we have
A
'
A = I and A  I = A'
I
which imply A' = AA = I = A' A. Thus the inverse of an orthogonal matrix equals its
transpose. Consider (3.69). Because
D
'
=
D and
M
'
=
M, (3. 69) equals its own transpose or
which implies Q
I
= Q
'
and Q
'
Q = QQ
'
= I. Thus Q is û orthogonal matrix; its columns
are orthonormalized eigenvectors of
M
. This is summarized as a theorem.
> Theorem 3.6
For every real symmetric matrix
M, there exists an orthogonal matrix Q such that
M = QDQ
'
or D = Q
'
MQ
where
D is a diagonal matrix with the eigenvalues of
M
, which are all real, on the diagonal.
A symmetric matrix
M is said to be .denoted by
M ¯ 0, jf x
'
1Ix ¯ 0 for
every nonzero x. It is ..denoted by M
� 0, if x
'
M
x � 0 for every nonzero
x. If
M
¯ 0, then x
'
;1x = 0 if and only if x = O. If
M is positive semidefnite, then there
exists a nonzero x such that x
M
x = O. This property will be used repeatedly later.
Theorem 3.7
A symmetric matrix M is positive defnite (positive semidefnite) i f and only if any one of the
following conditions holds.
1. Every eigenvalue of 1 is positive (zero or positive).
2. All the leading principal minors of
M
are positive (all the principal minors of 1 are zero or positive).
3, There exists an nonsingular matrix N (an singular matix N or an .matrix N with
.< such that 1 = N'N.
Condition ( 1 ) can readily be proved by using Theorem 3.6. Next we consider Conditon (3). If
M
=
N
'
N
, then
x
'
Mx = x
'
N
'
Nx = (Nx)
'
(Nx) = I I Nx · � 0
3. 9 Quadratic Form and Posi ti ve Defi ni teness 75
for any x. I f N is nonsingular. the only x to make N
x = 0 is x = O. Thus
M
is positive defnite. If N
is singular, there exists a nonzero X to make Nx = O. Thus M is positive semidefnite. For a proof of
Condition (2), see Reference [ \ 0].
We use an example to illustrate the principal minors and leading principal minors. Consider
M
=
`� `��`�
. ..
Its principal minors are m
+ ..
det
11l :1
det
.
.
and det
M. Thus the principal minors are the determinants of al l submatrices of
M
whose diagonals
coincide with the diagonal of
M
. The leading principal minors of M
are
. det
. .
and det M
.
Thus the leading principal minors of
M
are the determinants of the submatrices of
M
obtained by deleting
the last k columns and last k rows for · = 2,
. Theorem 3,8
1. An .matrix u.with .� has rank n, if and only if the matrix uuhas rank or
det(H
'
H) o O.
2. An .matrix uwith . has rank .if and only if the ..matrix uuhas rank .or
a:.uu,o O.
The symmetric matix uuis always positive semidefnite. It becomes positive definite if uuis
nonsingular. We give a proof of this theorem. The argument in the proof will be used to establish the
main results in Chapter 6: therefore the proof is spelled out in detail.
Proof: The condition ..uu, = implies ..u, = We show this by
contradiction. Suppose ..uu, = but ..u, < Then there exists a nonzero vector
such that u = 0, which implies uu = O. This contradicts ..uu, = Thus
..uu:implies ..u,=
The condition ..u, = implies ..uu, = Suppose not, or
..uu, < then there exists a nonzero vector such that uu= 0, which implies
uu= O or
0 = uu= .u,.u,= u ·
Thus we have u·= O. This contradicts the hypotheses that Y o 0 and ..u,= Thus
..u,= implies ..uu,= This establishes the first part of Theorem 3. 8. The second
part can be established similarly. Q. E. D.
We discuss the relationship between the eigenvalues of u uand those of uu Because both uu
and uuare symmetric and positive semidefnite, their eigenvalues are real and nonnegative (zero or
76 LI NEAR ALGEBRA
positive). If H is m X n, then H'H has n eigenvalues and HH' has U eigenvalues. Let A = H and
B = H' . Then (3.65) becomes
det(slm HH') = s",n det(sln  H'H)
(3.70)
This implies that the characteristic polynomials of HH' and H' H differ only by s
m
n . Thus we
conclude that HH' and H'H have the same nonzero eigenvalues but may have different numbers of
zero eigenvalues. Furthermore. they have at most i := min (m, n) number of nonzero eigenvalues.
3 , 1 0 Si ngul arVal ue Decomposition
Let H be an nz x n real matrix. Defne M : = H'H. Clearly M is n x n, symmetric, and semidefinite.
Thus all eigenvalues ot M üI@real and nonnegative (zero or positive). Let r be the number of its positive
eigenvalues. Then the eigenvalues of M = H'� can be arranged as
_ r
AI ::
> = = . . .
=
Let i : = min(m, n) . Then the set
Aj ::
> = = . . .
=
is called the singular values of H. The singular values are usually arranged i n descending order in
magnitude.
EXAMPLE 3.13 Consider the x 3 matrix
H =
We compute
M = H'H =
and compute its characteristic polynomial as
det(AI M) =
=
Thus the eigenvalues of H'H are and and the singular values of H are =
and Note that the number of singular values equals min (n, m) .
In view of we can also compute the si ngUlar values of H from the eigenvalues of
HH'. Indeed, we have
and
lI : = HH' =
det(AI  M) = =
Thus the eigenvalues of HH' are and 0 and the singular values of H' are and
We see that the eigenvalues of H'H differ from those of HH' only in the number of zero
eigenvalues and the singular values of H equal the singular values of H'.
3 . 1 1 Norms of Matrices 77
For ,. = H' H. there exists, following Theorem an orthogonal matrix Q such that
Q'H'HQ = D =: S'S ( 71 !
v, 'here D is an n x n diagonal matrix with on the diagonal. The matrix S is 111 x n with the
singular values i., on the diagonal. Manipulation on wi l l lead eventially to the theorem
that follows.
Theorem 3.9 (Singularvalue decomposition)
Every n x H matrix H can be transformed into the form
H = RSQ'
wi th R'R = RR' = 1m, Q'Q = QQ' = In. and S being 111 x 11 with the singular values of H on the
diagonal.
The columns of Q are orthonormalized eigenvectors of H'H and the columns of R
are orthonormalized eigenvectors of HH' . Once R, S. and Q are computed. the rank of H
equals the number of nonzero singular values. If the rank of H is r, the frst r columns of
R are an orthonormal basis of the range space of H. The last (1  r) columns of Q are an
orthonormal basis of the null space of H. Although computing singularvalue decomposition is
time consuming, it i s very reliable and gives a quantitative measure of the rank. Thus it i s used in
MATLAB to compute the rank. range space. and null space. In MATLAB. the singular values
of H can be obtained by typing s � svd ( E) Typing R, S , Q J � svd ( H ) yields the three
matrices in the theorem. Typing orth ( H ) and l,ul l ( H ) yields. respectively, orthonormal
bases of the range space and null space of H. The function l".ul l wi l l be used repeatedly in
Chapter
EXAMPLE 3, 14 Consider the matrix in We type
a= � 0 l . 2 ; 1 2 3 � ; :: 0 2 0 : ;
[ r , s , q] o S'!( a )
which yield
r =
9399 0. 2182
702
0.2439
. =
0. 2475
á =
Thus the si ngular values of the matrix A in are and The matrix has
two nonzero singular values, thus its rank is and, consequently. its nullity is peA) = 2.
The frst two columns of l are the orthonormal basis in and the last two columns of (f
are the orthonormal basis i n ( 3. 14).
78 LI NEAR ALGEBRA
3, 1 1 Norms of Matrices
The concept of nons for vectors can be extended to matrices, This concept is needed in
Chapter Let A be an U x matrix. The non of A can be defned as
I I AI I = sup
I I Axl 1
= sup I I Axl 1
x±0 I l xl l I I xl l =1
(3. 72)
where sup stands for supremum or the least upper bound. This non is defned through the
nonn of x and is therefore called an For different I l xl l , we have different I I AI I ·
For example, if the I nonn I l xl l l i s used, then
I I AI I I = m·
x _lail l .= largest column absolute sum
r�l =l
where a'i is the ijth element of A. If the Euclidean non I I xl 1 2 is used, then
I I AI 1 2 = largest singular value of A
= (largest eigenvalue of A
'
A)
I /2
If the infnitenorm I l xl l " is used, then
I I AI I" = m�x l aij I = largest row absolute sum
J=I
These nons are all different for the same A. For example, i f
A =
�I �
[
then
I I AI I I = 3 + I  I I = 4, I I AI 1 2 = 3. 7, and I I Al l c = 3 + 2 = 5, as shown in Fig.
3. 3. The MATLAB functions norm ( a , 1 ) , norm( a , 2 ) =norm ( a) , and norm ( a , inf )
compute the three nonns.
The nonn of matrices has the following properties:
I I Axl l .: I I Al l l l xl 1
I I A + BI I .: I I AI I + I I BI I
I I ABI I .: I I AI I I I BI I
The reader should try frst to solve all problems involving numerical numbers by hand and
then verify the results using MATLAB or any software.
3.1 Consider Fig. 3 . 1 . What i s the representation of the vector x with respect to the basis
{ql . h) ? What i s the representation of ql with respect to {i2 ' q2 P
3.2 What are the I non, 2non, and infnitenon of the vectors
X_
II xll , 1
4
IIxll2 1 2
This magnitude gives
The sum of these two
the nonn of ¬.
magnitudes gives the
nonn of ¬.
Z
x

.:...: Ti
=4 ` 2
¬Z
(e)
This magnitude gi yes
the nonn of ¬.
Figre 3.3 Diferent norms of A.
Probl ems 7'
3.3 Find two orthononnal vectors that span the same space as the two vectors in Problem
3. 2.
3.4 Consider an n x U matrix A with :: U. If all columns of A are orthononnal, then
A' A = 1m· What can you say about AA'?
3.5 Find the ranks and nullities of the following matrices:
80 LI NEAR ALGEBRA
1
o
3
o
3.6 Find bases of the range spaces and nul l spaces of the matrices i n Problem
3.7 Consider the l inear algebraic equation
It has three equations and two unknowns. Does a solution .exist in the equationo Is the
solution unique? Does a solution exis
'
t if y =
1
3.8 Find the general solution of
o
How many parameters do you have?
3
o
3.9 Find the solution in Example 3.3 that has the smallest Euclidean norm.
3.10 Find the solution in Problem 3.8 that has the smallest Euclidean norm.
3.1 1 Consider the equation
..+ .
+ An
2
l + . . . + .
+
where A i s an matrix and i s an x 1 column vector. Under what conditions on
A and will there exist
to meet the equation for any and
.
Write the equation in the form
..
.
.
··
3.12 Given
what are the representations of A with respect to the basis {, . .
.
and the
basis {b. Ab, Alb. Alb], respectively? (Note that the representations are the same! )
3. 13 Fi nd Jordanform representations of the following matrices:
3.14
3.15
Problems 81
�
�
�
A 2 Az
4
�
A4
4 3
AJ = 1 6
Note that all except A. can be diagonalized.
Consider the companionform matrix

(
l
a
l
1
Show that its characteristic polynomial is given by
/( i, ) = A
4
+ al A
3
+ aZA2 + aJA + 0.
Show also that if Ai is an eigenvalue of A or a solution of 6(A) then
A' A; Ai
i s an eigenvector of A associated with Ai .
Show that the Valldenllollde determinant
A
3
• J
A
l
A
l 2
A
3
A
2
A� A
2
A
2
I 3
4
Al A2 A
3
A4
1 1
equals f
I
Si <jS.(Aj  A, ) . Thus we conclude that the matrix is nonsingular or. equiva
lently. the eigenvectors are l inearly independent if all eigenvalues are distinct.
3.16 Show that the companionform matrix in Problem 3 . 1 4 i s nonsingular i f and only if
04 i Under this assumption. show that its i nverse equals
3.17 Consider
A =
�
AT
A
o
o
o
with A .and T ¯ Show that is a generalized eigenvector of grade 3 and
the three columns of
82 LI NEAR ALGEBRA
¹7
¯
0
¹7 0
o 1
constitute a chain of generalized eigenvectors of length 3. Verify
1 0
Q¯' AQ = � ¹ 1
o 0 ¹
3.18 Find the char'cteristic polynomials and the minimal polynomials of the following
matrices:
Í
0
0
¹
¹i
0 ¹i
0 ¹
0 0 0 0
0
0 0
¹i
0 ¹i
0
0 ¹i
0 ¹
0 0 0 0
3.19 Show that if ¹ is an eigenvalue of A with eigenvector x, then 1(¹) is an eigenvalue of
)(A)with the same eigenvector x.
3.20 Show that an n x n matrix has the property A' = 0 for k :: i f and only if A has
eigenvalues 0 with multiplicity n and index m or less. Such a matrix i s called a nilpotent
matrix.
3.21 Given
fnd A
l¹¸ A
 ¹
:
¸and
3.22 Use two diferent methods to compute
for A and A. in Problem 3. 1 3.
3.23 Show that functions of the same matrix commute; that is,
1(A)x'A) = x(A) 1(A)
Consequently we have A
t
=
A
3.24 Let
Problems 83
Find a matrix B such that
= C. Show that if ¹. = 0 for some i, then B does not exist.
Let
¹ 1 0
C = 0 ¹ 0
o 0 ¹
Find a B such that
= C. Is it true that, for any nonsingular C, there exists a matrix B
such that
= C?
3.25 Let
!
(s¡ ~ A)¯'
= ¯Adj (s¡  A)
A(s)
and let (s)be the monic greatest common divisor of all entries of Adj (sI ¯A)
·
Verify
for the matrix A
:
in Problem 3 . 1 3 that the minimal poly nominal of Aequals ^(s)/m(s)
3.26 Defne
where
^(s) .= dct(s¡  A) .= s¹ +oi s¹¯
'
+o·¯¹¨¨ + · · · +o~
and R,are constant matrices. This defnition i s valid because the degree in sof the adjoint
of (s¡~ A)i s at most n  1 . Verify
tr(AKc)
oi = ¬
tr(ARj )
o· = ~ 
2
tr(AR
¯
)
o
'
= 
3

tr(AR
¹
.
¯
)
o
¹
~
= ¬
n ¯ 1
Kc = Í
where tr stands for the trce of a matrix and is defned as the sum of all its diagonal
entries. This process of computing Q¡ and R. is called the
3.27 Use Problem 3.26 to prove the CayleyHamilton theorem.
3.28 Use Problem 3.26 to show
1
(s¡ ~ A)' =
^(s)
¸
A¹¯'
+(s+oi )A¹¯
¯
+(s
¯
+oi
s +o;)A
¹
¯
'
8 LI NEAR ALGEBRA

¹
~
l

¨
: + ,¡¡
+
. . + s + al s · · · +an
I
3.29 Let all eigenvalues of Abe distinct and let abe a right eigenvector of A associated with
· .that is, A0. = · a
Defne
Q
= 'a. a·
.
0
¹
! and defne
P
�
Q
' � �
where ai s the ith row of
P
. Show that ais a left eigenvector of Aassociated with · .
that is, P,A = · a
3.30 Show that if all eigenvalues of A are distinct, then (sl ¬ A)¨' can be expressed as
1
(sl ¯ A)¨
i
= ' aa
where a.and p, are right and left eigenvectors of Aassociated with ·.
3.31 Find the M to meet the Lyapunov equation in (3. 59) with
What are the eigenvalues of the Lyapunov equation? Is the Lyapunov equation singular0
Is the solution unique?
3.32 Repeat Problem 3. 31 for
A
~
2
B = I
with two different C.
3.33 Check to see if the following matrices are positive defnite or semidefnite:
3
3 1
��
� 2
3.34 Compute the singular values of the following matrices:
o
 1
�
3.35 If Ais symmetric, what is the relationship between its eigenvalues and singular values0
Problems 85
3.36 Show
3.37 Show (3. 65).
3.38 Consider Ax = y,where A is m x n and has rank U. Is (N A)¨
l
Aya solution? If not,
under what condition will it be a solution? Is A'(AA)¨' y0 solution?
Chaptcr
+
StateSpace Sol uti ons
and Real i zati ons
4. 1 I ntroduction
8
We showed in Chapter 2 that linear systems can be described by convolutions and, if lumped,
by statespace equations. This chapter discusses how to fnd their solutions. First we discuss
briefy how to compute solutions of the inputoutput descrption. There is no simple analytical
way of computing the convolution
y(t) = ,�g(t, r) u(r) dr
The easiest way is to compute it numerically on a digital computer. Before doing so, the
equation must be discretized. One way is to discretize it as
k
y(kb) = S g(kt, mt)u(mt) t
(4. 1 )
m=k
where t is called the integration step size. This is basically the discrete convolution discussed
in (2.34). This discretization is the easiest but yields the least accurate result for the same
integration step size. For other integration methods, see, for example, Reference  !7.
For the linear timeinvariant (LTI) case, we can also use yes) = g(s)u(s) to compute
the solution. If a system is distributed, g(s) will not be a rational function of s. Except for
some special cases, it is simpler to compute the solution directly in the time domain as in
(4. 1 ). If the system is lumped, g(s) will be a rational function of s. In this case, if the Laplace
transform of u (t) is also a rational fnction of s , then the solution can be obtained by taking the
inverse Laplace tansform of g(s)u(s). This method requires computing poles, carying out
4. 2 Solution of LTI State Equati ons 87
partial fraction expansion, and then using a Laplace transform table. These can be carried out
using the MATLAB functions roo t s and re s idue. However, when there are repeated poles.
the computation may become very sensitive to small changes in the data, including roundof
errors; therefore computing solutions using the Laplace transform is not a viable method on
digital computers. A better method is to transform transfer functions into statespace equations
and then compute the solutions. This chapter discusses solutions of state equations, how to
transform transfer functions into state equations, and other related topics. We discuss frst the
timeinvariant case and then the timevarying case.
4.2 Sol ution of LTI State Equations
Consider the linear timeinvariant (LTI) statespace equation
x(t) = Ax(t) Bu(t)
y(t) = Cx(t) Du(t)
(4.2)
(4.3)
where A, B, C, and D are, respectively, n x n, n x p, q x n, and q x p constant matrices.
The problem is to fnd the solution excited by the initial state x(O) and the input u(t). The
solution hinges on the exponential function of A studied in Section 3.6. In particular, we need
the property in (3. 55) or
d
_eAt = AeAt ~ eAt A
dt
to develop the solution. Premultiplying eAt on both sides of (4.2) yields
eAt x(t)
 eAtAx(t) =
eAt Bu(t)
which implies
Its integration from 0 to t yields
eAr x( r) I =o = eAr Bu(r) dr
Thus we have
eAtx(t)
 eOx(O) = eArBu(r) dr (4.4)
Because the inverse of eAt is eAt and eO = I as discussed in (3.54) and ( 3.52), (4.4) implies
x(t) = eAt x(O) eA(tr)Bu(r) dr (4.5)
This is the solution of (4.2).
It is instructive to verify that (4.5) is the solution of (4.2). To verify this, we must show
that (4.5) satisfes (4.2) and the initial condition x(t) = x(O) at t = Indeed, at t = 0, (4.5)
reduces to
88 STATESPACE SOLUTI ONS AND REALI ZATI ONS
Thus · satisfes the initial condition. We need the equation
· =
· ·
·
·
to show that · satisfes · · Diferentiating · and using · we obtain
.=
· .
·
. ]
= .·.
/·· . ·.
,
= ··
.
·
.+ ·.
which becomes, after substituting ·
.= ·. .
Thus · meets · · and the initial condition .and is the solution of . ·
Substituting · into · yields the solution of · as
,= .· . .¸
·
. + .
.
·
This solution and . are computed directly in the time domain. We can also compute the
solutions by using the Laplace transform. Applying the Laplace transform to . and ·
yields, as derived in · · and ·
.. = (sl ¯ ·
. .
. = ... · . .. + ..
Once ..and Yes) are computed algebraically, their inverse Laplace transforms yield the
timedomain solutions.
We now give some remarks concering the computation of · We discussed in Section
3 .6 three methods of computing functions of a matrix. They can all be used to compute
·
1. Using Theorem First, compute the eigenvalues of ·next. fnd a polynomial of
degree n that equals
on the spectrum of ·then ·= .
2. Using Jordan form of ·Let ·= ,·, then ·= ,
·, where ·is in Jordan
form and
·
can readily be obtained by using ·
3. Using the infnite power series i n Although the series wiII not. in general. yield a
closedform solution, it is suitable for computer computation, as discussed following
In addition, we can use to compute ·that is,
·= . .. ¯ · ·
I
4. 2 Solution of LTI State Equati ons 89
The inverse of (s I ·is 3 function of ·therefore. again, we have many methods to compute
it:
1. Taking the inverse of .¯ A).
2. Using Theorem
3. Using .. . = ,.. · ,
and ··
4. Using the infnite power series in
5. Using the Leverrier algorithm discussed in Problem
EXAMPLE 4.1 We use Methods and · to compute .. ¯ A)
I
. where
·
·
Method 1: We use to compute
.. ¯ A)
I
=
� . �· '
=
�
.
. ·
=
· .
·
.
.
Method 2: The eigenvalues of ·are ¯1 + ¯ Let heAl = ·+ ·
If h equals · : =
. ¯ ,)1 on the spectrum of · then
Thus we have
and
= h
= h :
. = /0 ·
.
= /
1
(sl .
= · = .
. ·
=
·.
·
..
EXA\lPLE 4.2 Consider the equation
Its solution is
x(t) =
�
� ·
. � ` .
. = · . ¸
· .
The matrix function ·
is the inverse Laplace transform of ¯ . which was computed
in the preceding example. Thus we have
90 STATESPACE SOLUTIONS AND REALI ZATI ONS
and
=

I
(s
=
s
,� �
We discuss a general property of the zeroinput response Consider the second
matrix in (3.39). Then we have
t
ZeAl t
/2 0 0
,
"
=
Q
i�l t
0
Q
'
0
eA1 t
0
0 0
0 0 0
Every entry of and, consequently, of the zeroinput response is a linear combination of terms
These terms are dictated by the eigenvalues and their indices. In
general, if .has eigenvalue A with index i ¡ . then every entry of is a linear combination of
Every such term is analytic in the sense that it is infnitely diferentiable and can be expanded
in a Taylor series at every This is a nice property and will be used in Chapter 6.
I every eigenvalue, simple or repeated, of .has a negative real part, then every zero
input response will approach zero as t  0. If .has an eigenvalue, simple or repeated, with
a positive reai part, then most zeroinput responses will grow unbounded as t ¬ 0. If Ahas
some eigenvalues with zero real part and all with index 1 and the remaining eigenvalues all
have negative real parts, then no zeroinput response will grow unbounded. However, if the
index is 2 or higher, then some zeroinput response may become unbounded. For example, if
.has eigenvalue with index 2, then contains the terms tl . If a zeroinput response
contains the term t, then it will grow unbounded.
4. 2. 1 Discretiztion
Consider the continuoustime state equation
= .
=
(4.9)
(4. 1 0)
If the set of equations is to be computed on a digital computer, it must be discretized. Because
.
= hm
'
4.2 Solution of LTI State Equations 91
we can approximate (4.9) as
= . (4. 1 1
If we compute . only at t = for = 1 , . . . , then (4. 1 1 ) and (4.10) become
= (I .
=
This is a discretetime statespace equation and can easily be computed on a digital computer.
This discretization is the easiest to carry out but yields the least accurate results for the same
We discuss next a diferent discretization.
If an input is generated by a digital computer followed by a digitaltoanalog
converter, then will be piecewise constant. This situation ofen arises in computer control
of control systems. Let
= =: for :: t < (4. 1 2)
for = I , 2, . . . . This input changes values only at discretetime instants. For this input,
the solution of (4.9) still equals (4.5). Computing (4.5) at t = and = 1 ) T yields
and
:= =

1 : = =

Equation (4. 1 4) can be written as
1 =
(4. 13)
(4. 14)
which becomes, after substituting (4. 1 2) and (4. 13) and introducing the new variable  ª
I =
,
,
Thus, i f an input changes value only at discretetime instants and if we compute only the
responses at = then (4.9) and (4. 1 become
1 = .
=
(4. 15)
(4. 16)
92 STATESPACE SOLUTI ONS AND REALI ZATI ONS
with
(4. 1 7)
This is a discretetime statespace equation. Note that there is no approximation involved in
this derivation and (4. 1 5) yields the exact solution of (4.9) at t = if the input is piecewise
constant.
We discuss the computation of Bd. Using (3. 5 1 ), we have
¡
. . .
.
= . .. ..
¹ 3 ! 4'
Î
This power series can be computed recursively as in computing (3. 5 1 ). If .is nonsingular,
then the series can be written as, using (3. 5 1 ),
Thus we have
I
¡
I 'T
. ..... . I = .(e' 
3 '
(if .i s nonsingular)
Using this formula. we can avoid computing an infnite series.
(4. 1 8)
The MATLAB function [ ad , bd] =c2d ( a , b , T) transforms the continuoustime state
equation in (4. 9) into the discretetime state equation in (4. 1 5) .
4. 2. 2 Solution of DiscreteTime Equations
Consider the discretetime statespace equation
1 ] = ..
= .
(4. 1 9)
where the subscrpt has been dropped. It is understood that if the equation is obtained from
a continuoustime equation, then the four matrices must be computed from (4. 1 7). The two
equations in (4. 1 9) are algebraic equations. Once and = 0, I , . . . , are given, the
response can be computed recursively from the equations.
The MATLAB function ds tep computes unitstep responses of discretetime statespace
equations. It also computes unitstep responses of discrete transfer functions; interally, it frst
transforms the transfer function into a discretetime statespace equation by calling t f 2 s s,
which will be discussed later, and then uses ds t ep. The function dl s ir, an acronym for
discrete linear simulation, computes responses excited by any input. The function step
computes unitstep responses of continuoustime statespace equations. Interally, it frst uses
the function c 2 d to transform a continuoustime state equation into a discretetime equation
and then carries out the computation. If the function s tep is applied to a continuoustime
transfer function, then it frst uses t f 2 s s to transform the transfer function into a continuous
time state equation and then discretizes it by using c 2 d and then uses ds t ep to compute the
4. 3 Equi val ent State Equations 93
response. Similar remarks apply to 1 s ir, which computes responses of continuoustime state
equations or transfer functions excited by any input.
In order to discuss the general behavior of discretetime state equations. we will develop
a general form of solutions. We compute
= .
= . = ..
Proceeding forward, we can readily obtain, for ¯ 0,
kI
= .
_.
k¬l
= .
_.
.
(4. 20)
(4. 21 )
They are the discrete counterparts of (4. 5) and (·.7) . Their derivations are considerably simpler
than the continuoustime case.
We discuss a general property of the zeroinput response . Suppose .has eigen
value with multiplicity 4 and eigenvalue with multiplicity I and suppose its Jordan form
is as shown i n the second matrix i n (3. 39). In other words, has index 3 and
has index I .
Then we have
� I )
2
/2 0
Q
'
.
=
Q
�
� 0
I
0
0
I
0 0
I
0 0 0 �
which implies that every entry of the zeroinput response is a linear combination of �
� A; ) . These terms are dictated by the eigenvalues and their indices.
If every eigenvalue, simple or repeated. of .has magnitude less than I , then every zero
input response will approach zero as ¬ ¯. If .has an eigenvalue, simple or repeated, with
magnitude larger than 1 , then most zeroinput responses will grow unbounded as ¬ Ò. If .
has some eigenvalues with magnitude and all with index and the remaining eigenvalues all
have magnitudes less than I , then no zeroinput response will grow unbounded. However. if
the index is or higher, then some zerostate response may become unbounded. For example.
if .has eigenvalue I with index then .contains the terms ( I , If a zeroinput response
contains the term then it will grow unbounded as ¬ ¯.
4. 3 Equi val ent State Equati ons
The example that follows provides a motivation for studying equivalent state equations.
EXAMPLE 4.3 Consider the network shown in Fig. 4. 1 . It consists of one capacitor, one
inductor, one resistor, and one voltage source. First we select the inductor current XI and
94 STATESPACE SOLUTIONS AND REALIZATI ONS
capacitor voltage as state variables as shown. The voltage across the inductor is and
the current through the capacitor is The voltage across the resistor is thus its current is
Clearly we have and Thus the network is described
by the following state equation:
y [0 l]x (4.22)
If, instead, the loop currents and are chosen as state variables as shown, then the voltage
across the inductor is and the voltage across the resistor is
.
1 . From the lefthand
side loop, we have
or
The voltage across the capacitor is the same as the one across the resistor, which is
I
Thus the curent through the capacitor is which equals or
. + u
Thus the network is also described by the state equation
y [ l  l lx
(4.23)
The state equations in (4.22) and (4.23) describe the same network; therefore they must be
closely related. In fact, they are equivalent as will be established shortly.
Consider the ndimensional state equation
x(t) .
(4.24)
where A is an n x n constant matrix mapping an ndimensional real space R" into itself. The
state x is a vector in R" for all t; thus the real space is also called the state space. The state
equation in (4.24) can be considered to be associated with the orthonormal basis in (3.8). Now
we study the efect on the equation by choosing a diferent basis.
³2
�~ ~
Figre 4.1 Network with two different
sets of state varables.
4.3 Equi val ent State Equations
Defnition 4.1 Let P be an n x n real nons in gular matrix and let x = P. Then the state
equation,
where
x(1) = Ax(t) .Bu(t)
y(t) Cx(t) + Du(t)
C CpI
(4.25)
(4.26)
is said to be (algebraically) equivalent to (4.24) and x P is called an equivalence
transformation.
95
Equation (4.26) is obtained from (4.24) by substituting x(t) = pI X(t) and x( t) =
pI X(t) . In this substitution, we have changed, as in Equation (3. 7), the basis v
�
ctors of the
state space from the orthonormal basis to the columns ofP1 =: Q. Clearly Aand Aare similar
and A is simply a different representation of A. To be precise, let Q = pI = [ql q2 . . . qn l .
Then the ith column of A is, as discussed in Section 3.4, the representation of Aqi with respe�t
to the basis { ql , q2
.
" ' , qn } . From the equation B = PB or B = pI B = [ql ql . . . qn ]B.
we see that the i th column of B is the representation of the i th column of B with respect to the
basis { q I , q
2
.
. . . . qn } . The matrix C is to be computed from CpI . The matrix called the
direct transmission part between the input and output. has nothing to do with the state space
and is not afected by the equivalence transformation.
We show that (4.24) and (4.25) have the same set of eigenvalues and the same transfer
matrix. Indeed, we have. using det(P) det(pI ) 1 ,
and
L(A) = det(U ¯ A) = det(App1 ¯ PAPI ) = det[P(U  AlP
]
= dCt (P) det(U A) det (pI ) ¬ det(AI  A) 6(A)
c(s) = C(sI ¯ A)I B + D = CPI [P(sI  AlP I r
l
pB +
= CPIp(sI  A) 
l
pI pB .C(sI  A)I B .= c(s)
Thus equivalent state equations have the same characteristic polynomial and. consequently, the
same set of eigenvalues and same transfer matrix. In fact, all properties of (4. 24) are preserved
or invariant under any equivalence transformation.
Consider again the network shown in Fig. 4. 1 . which can be described by (4.22) and (4. 23).
We show that the two equations are equivalent. From Fig. 4. 1 , we have
I
= .r l . Because the
voltage across the resistor is X2, its curent is xzl i and equals X
I
 X2. Thus we have
or
,
rl
¦
I 0
¦
X
I
[
I 0
¦
X
I
[
I I X2
=
I ¯J 7]
(4.27)
96 STATESPACE SOLUTI ONS AND REALI ZATI ONS
Not e that, for this P, it s inverse happens t o equal i tself, I t is st raight forward t o verify that (4.22) and (4.23) are related by the equivalence transformation in (4.26).
The MATLAB function [ ab, bb , cb, db] = s s 2 s s ( a , b , c , d , p ) carries out equiv alence transformations.
Two state equations are said to be �erostate equivalent if they have the same transfer matrix or
This becomes, after substi t uting (3. 57),
D + CBSI + CABs2
+ CA2Bs3 +
. . . = 0 + CBsI + CABs2 + CA
2
Bs3 +
.
.
Thus we have the theorem tha t follows.
> Theorem 4, I
Two linear timeinvarant state equations (A, B, C_ DJ and ( A. B, C, OJ are zerostate equi valent or have the same transfer matrix if and only if D ÷ D and
U ÷
I , 2, . . .
I t is clear that (algebraic) equivalence implies zerostate equivalence. In order for two
state equations to be equivalent, they must have the same dimension. This is. however, not the
case for zerostate equivalence, as the next example shows.
EXAMPLE 4.4 Consider the two networks shown in Fig. 4.2. The capacitor is assumed to have
capacitance  I F. Such a negative capacitance can be realized using an opamp circui t. For the
circuit in Fig. 4.2(a), we have y(t) = 0.5 . u (t) or y(s) = 0.5i (s). Thus its transfer function is
0.5. To compute the transfer function of the network in Fig. 4.2(b), we may assume the initial
vol tage across the capacitor to be zero. Because of the symmetry of the four resistors, half of
the current will go through each resistor or i (t) ÷ 0.5u (f), where i (t) denotes the right upper
resistor's current. Consequently, y(t) = i (t) . I ~ 0. 5u(t) and the transfer function also equals
0.5. Thus the two networks, or more precisely their state equations, are zerostate equivalent.
This fact can also be verifed by using Theorem 4. ! . The network in Fig. 4.2(a) is described
by the zerodimensional state equation y(t) ÷ 0. 5u (t) or A ÷ B = C = 0 and D = 0. 5. To
Figure 4.2 Two zerostate
u  ,
l
equivalent networks.
0.5r
y
Î I l J
l
�
l
�
.f
 I F
o Jí
Î ! l
M ¬ l ¬ .t l
.
.t
(a)
(b)
1
i
i
·
4. 3 Equi val ent State Equations 97
.
k ' F 4 2(b) we assign the capacitor vol tage as state develop a state equatIOn for the networ lfl
.
Ig
:
f'
.
'
g from ±e ne"ative to positive polari ty
. .
h I 't shown I ts curent IS x OWlfl�
C
vanable x Wi t po 3 y
.
.
If e assi "n the right upper resistor's current as i ( t) . then because of the negative capacitance. W I
. ,
ent I' S u ¬ i and the left
. , .
. ' the left upper resistor s curr , the right lower resistor s current IS � ¬ x,
d he u er righthand loop is 0: lower resistor's current is ll ¬ i + x. The total voltage aroun t pp
i  x (u  ÷ 0 or i ÷ 0. 5(x + u)
which implies
y ~ I
.
i = i ÷ 0. 5(x + u)
The total voltage around the lower righthand loop i s 0:
x + (i x) ¬ (ll ¬ i + x) = 0
which implies
2.i ~ 2i + x lI ~ X + lI + X u = 2x
k · F' " (b) is described by the onedimensional state equation Thus the networ H Ig . .. .. .
:i' (t) = x(t)
y(t) = 0.5x(t) + 0.511 (t)
C ÷ 0. 5, and 0 ÷ 0. 5. We see that D ÷ 0 ~
.
0. 5 and CAmB =
. . . . Thus the two equations are zerostate equivalent.
with A ÷ B ÷
CAmB ÷
for U ~
, I .
4. 3. 1 Canonical Forms
.
b bb b dh p l = canon ( a , b, c , d , ' type ' . If MATLAB contains the functIOn [ a , , c , i , _
. .
' th A in the tye =companion, the function will generate an eqUlvalent �tate e��tlOn
.
WI
An
1
bt l is
.
f
:
(3 ") This function works only If Q . [bl
I compam�n .÷I
�
.
'
.he trst column ofB. This condition is the same as ( A. b
1
) controllable, nonslfgu
ll
a
d
r,
.
w ere
n
l
C
ls
hapter 6 The P that the function canOE generates equals Q

l
.
See the as we WI ISCUSS I .
discussion in Section 3.4.
.
I f Suppose A has two real eigenvalues and two We discuss a dlf
.�.�s�
a
«
°
���nl�.�1 coeffcients, the two complex eigenvalues must complex eigenvalues.
. ' .
and a ¬ I be the eigenvalues and ql , Q2. q). and be complex conjugate. Let
^],
+ J
h
�
'
q and q' are all real and q� equals be the correspondIng eigenvectors. w ere A I , 2 , a, }
1
, . q�
.
f Q  [ , q' q J. Then we have the complex conjugate of q) . De ne ¯ ql q_ : �
�
2
� �
= Q
I
AQ
J
:
=
0 0 a + jf
o 0 0 a ¬ J�
.
. ;  ,
( a ) in MATLAB as shown in Examples Note that Q and
J
can be obtaIned from [ q, J 

e �9
.
real matrix by the 3.5 and 3.6. This for is useless in practice but can be transformed Into a
.
following similarty transformation
98
STATESPACE SOLUTI ONS AND REALI ZATIONS
�
�
�
o 0 0. 5
o 0 0. 5
o
o
o
o
o
a jf
o
o
a
o
o
a

f
�
A
f
.
a
We see that this transfonnation transfonns the complex eigenvalues on the diagonal into
a block with the real part of the eigenvalues on the diagonal and the imaginary part on
the offdiagonal. This new Amatrix is slid to be in modal fonn. The MATLAB function
[ ab , bb, cb , db, P l �canon ( a , b , c ,
t
d, ' modal ' ) or canon ( a , b , c , d) with no
type specifed will yield an equivalent state equation with
A
in modal form. Note that there is
no need to transfonn Ainto a diagonal fon and then to a modal fonn. The two transfonnations
can be combined into one as
_I
. 0
P = QQ = [ql qz q3 q4
]
�
÷ [ql q2 Re(q3) Im(q3) ]
0
I
0
0
0
�
j
0
0.5
0.5 0.5j
where Re and 1m stand, respectively, for the real part and imaginary part and we have used in
the last equality the fact that q4 i s the complex conjugate of q3 . We give one more example. The
modal fonn of a matrix with real eigenvalue AI and two pairs of distinct complex conjugate
eigenvalues ai ± jfi , for i ÷ 1 , 2, i s
0 0 0
A
�
al f1 0
f
1
al 0 (4.28)
0 0 a!
0 0
f
z
az
It is block diagonal and can be obtained by the similarity transfonnation
p
I
= [ ql Re(qz) Im(q2) Re(q4) Im(q4l ]
where ql , q2, and q4 are, respectively, eigenvectors associated with
1 , 0 I + j f1 , and a2 + j f2 .
This fonn is useful in statespace design.
4.3.2 Magnitude Scaling i n OpAmp Ci rcuits
As discussed in Section · every LTI state equation can be implemented using an opamp
circuit. l In actual opamp circuits, all signals are limited by power supplies. If we use ± I Svolt
1. This subsection may be skipped without loss of continuity_
4. 3 Equi val ent State Equati ons 99
power supplies, then all signals are roughly limited to ± 13 volts. If any signal goes outside
the range, the circuit will saturate and will not behave as the state equation dictates. Therefore
saturation is an important issue in actual opamp circuit implementation.
Consider an LTI state equation and suppose all signals must be limited to ±M. For linear
systems, if the input magnitude increases by a, so do the magnitudes of all state variables
and the output. Thus there must be a limit on input magnitude. Clearly it is desirable to have
the admissible input magnitude as large as possible. One way to achieve this is to use an
equivalence transfonnation so that
for all i and for all t . The equivalence transfonnation, however, will not alter the relationship
between the input and output; therefore we can use the original state equation to fnd the input
range to achieve l y(t ) 1 :: M. In addition, we can use the same transfonnation to amplify some
state variables to increase visibility or accuracy. This is illustrated in the next example.
EXAMPLE 4.5 Consider the state equation
x =

�
·
I
�
[
. �
�
[
.
y = [O. 1 ¯ I ]x
Suppose the input is a step function of various magnitude and the equation is to be implemented
using an opamp circuit in which all signals must be limited to ± I
. First we use MATLAB
to fnd i ts unitstep response. We type
a� [  O . l 2 ; 0  l J ; b� [ 1 0 ; O . l l ; c� [ 0 . 2  l J ; d� O ;
[ y, x, t l �step ( a , b, c , d) ;
plot ( t , y , t , x)
which yields the plot in Fig. 4. 3(a) . We see that I :q Imax ÷ 1 00 > I Ylmax = 20 and
I Xl l " I yl max , The state variable X2 is hardly visible and its largest magnitude is found
to be by plotting it separately (not shown) . From the plot, we see that if i ll (t ) 1 :: 0. 5, then
the output will not saturate but XI (t) will.
Let us introduce new state variables as
20
X
I
÷ XI ÷ 0. 2xl
1 00
With this transfonnation. the maximum magnitudes of XI (t) and
.
12 will equal l y Imax · Thus
if yet) does not saturate, neither will all the state variables .ii ' The transfonnation can be
expressed as x = Px with
P =
0. 2
o 2�¦
p_1
~
0

0 0.005
Then its equivalent state equation can readily be computed from (4.26) as
�

0. 1
. =
o
x + l!
0. 002
_
2

I 20
y = [ I  0.005]x
1 00 STATESPACE SOLUTIONS AND REALIZATI ONS
I U I U
.\�
I¡
bU ¹i
\
ÓU Ó
4U
`U
`
¹!
l U /U JU 4U ! Û `U JU 4U
(al (b)
Figure ".3 Time responses.
Its step responses due to u(t ) = 0.5 are plotted in Fig. 4. 3(b). We see that all signals lie inside
the range ± 10 and occupy the full scale. Thus the equivalence state equation is better for
opamp circuit implementation or simulation.
The magnitude scaling i s important i n using opamp circuits to implement or simulate
continuoustime systems. Al though we discuss only step inputs, the idea is applicable to any
input. We mention that analog computers are essentially opamp circui ts. Before the advent of
digital computers, magnitude scaling in analog computer simulation was carried out by trial
and error. With the help of digital computer simulation, the magnitude scaling can now be
carried out easily.
4. 4 Real izations
Every l i near timeinvariant (LTI) system can be described by the inputoutput description
Yes) = G(s)i(s)
and. if the system is lumped as well. by the statespace equation description
XU) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t)
(4. 29)
If the state equation is known. the transfer matrix can be computed as G(s) = Cis I A) I B+D.
The computed transfer matrix i s uni que. Now we study the converse problem. that is. to find a
statespace equation from a given transfer matrix. This is called the realbuion problem. This
terminology is justitied by the fact that, by using the state equation. we can build an opamp
circuit for the transfer matrix.
A transfer matrix G(s) i s said to be reakahle if there exists a fni tedimensional state
equation (4.29) or, simply. ( A. B. C, D) such that
G( s) = C(sl  A)  I B + D
4. 4 Real i zati ons 1 01
and ( A. B. C, D) i s called a reali�atioll of G(s) . An LTI distributed system can be de
�
cribed
bv a transfer matrix. but not by a fnitedimensional state equation. Thus not every G(s) i s
ralizable. If G( s) i s realizable. then it has infnitely many realizations. not necessarily of
the same dimension. Thus the realization problem i s fairly complex. We study here only the
realizability condition. The other issues wi l l be studied in later chapters.
j� Theorem 4.2
A transfer matrix G(s) is realizable if and only if G(s) is a proper rational matrix.
We use (3. 1 9) to write
G (5) :
=
CCsI  A)  I B =
I
qAdj (51  A) ] B sp
det(sl  A)
(4.30)
If A is n X n, then det (51  A) has degree n. Every entry of Adj (51  A) i s the deterinant
of an (n  I) x (n  I ) submatrix of (.I  A) : thus it has at most degree (n  I ) . Their l i near
combinations again have at most degree (n  I ) . Thus we conclude that C(sl  A)  l B is a
strictly proper rational matrix. I fD is a nonzero matrix. then C(sl A) I B + D is proper. This
shows that i f G( s) is realizable. then i t i s a proper rational matrix. Note that we have
G(c) = D
Next we show the converse; that is. if G(s) is a q x p proper rational matrix. then there
exists a realization. First we decompose G(s) as
G(s) ¯ G(c) + Gsp( s)
where Gsp is t he strictly proper par of G(s) . Let
des) = sr + UI S
r
1 + . . . + Ur_
1
S + Ur
(4.3 1 )
(4. 32)
be the least common denominator of all entries of Gsp( s) . Here we require d(s ) to be monic;
that is. its leading coeffcient i s 1 . Then Gsp(s) can be expressed as
,
I 1
[
r
 I N
r'
N
_ A: ¡
G (5 ) = [N(s) ] =  Ni s + . cS  + ' " + 1 r
 I
S · + f sp
drs) dr s)
where N, are q x p constant matrices. Now we claim that t he set of equations
[ 0, 1,
Ip
X =
0
0
y = [il Nc
u]Ip
0
Ip
0
. . · Nr

1
ur_
I
lp
0
0
Ip
Nr
¡
x + G( o) u
O�I"
1 n
o
x +
0
u
0
0
(4 33)
(4.34)
1 02 STATESPACE SOLUTI ONS AND REALI ZATI ONS
is a realization of( (s) . The matrix Ip is the p x p unit matrix and every 0 is a p x p zero matrix.
The Amatrix is said to be in block companion form; it consists of r rows and r columns of
p x p matrices; thus the Amatrix has order rp x rp. The Bmatrix has order rp x p. Because
the Cmatrix consists of r number of N;, each of order q x p, the Cmatrix has order q x rp.
The realization has dimension r p and is said to be in contrllable canonical for.
We show that (4.34) is a realization of
(
(s) in (4.3 1 ) and (4.33). Let us defne
z = �
= (d A) ' B
where Zj i s p x p and Z i s rp x p. Therthe transfer matrix of (4.34) equals
C(sl  A) l B + ((0) = N1 Zl + NzZz +
. . .
+ NrZr + ((0)
We write (4.35) as (sl  A)Z = B or
sZ = AZ + B
(4.35)
(4.36)
(4.37)
Using the shifting property of the companion form of A, from the second to the last block of
equations in (4.37), we can readily obtain
which implies
Substituting these into the frst block of equations in (4.37) yields
or, using (4.32),
Thus we have
¡ 0
2
Or des)
s + 01 +  +
. . .
+  Zl = Zl = Ip
s S
r

l
S
r I
s
rl
Zl  1

des)
p
'
Substituting these into (4.36) yields
, I
l '
,
C(sl  A)
l
B + G(o) = [N1 S
r

+ Nzs
r
. +
. . .
+ Nrl + G(O)
des)
This equals ((s) in (4.31 ) and (4.33). This shows that (4. 34) is a realization of
(
(s).
'
EXAMPLE 4.6 Consider the proper rational matrix
4s  10
_
3
_
((s) =
2s
l
+ 1 s + 2
s + I
(2s + I ) (s + 2) (s + 2)
2
1 2
=
�
�
¦
+
2s l
(2s + I ) (s + 2)
4. 4 Real izations 1 03
3
s + 2
s + I
(s + 2)2
(4. 38)
where we have decomposed
(
(s) into the sum of a constant matrix and a strictly proper rational
matrix (sp(s). The monic least common denominator of (sp (s) is des) = (s +0. 5) (s + 2)2 =
s3 + 4.5s
2
+ 6s + 2. Thus we have
> 1 6(s + 2)
2
Gsp (s) =
s
3
+ 4. 5s
2
+ 6s + 2 0.5(s + 2)
3(s + 2) (s + 0.5)
[
(s + l ) (s + 0. 5)
I ¡
6
=
des) 0
·
[
s
2
24
1
+
0.5
7. 5
s +
24
1 . 5 I 0�5
[
'
and a realization of (4. 38) is
4. 5 0
6 0 2 0
0
0 4. 5 0 6 0 2
0
0 0 0 0 0 0 0
X = X +
0 0
0 0 0 0 0
0 0
0 0 0 0 0
0 0
0 0 0 0 0
6 3 24 7. 5 24
0
3
X +
�
�[
·
[
Y
=
0 0. 5 1 .5
This is a sixdimensional realization.
(4.39)
We discuss a special case of (4. 3 1 ) and (4.34) in which p = I . To save space, we assume
r = 4 and q = 2. However, the discussion applies to any positive integers r and q. Consider
the x I proper rational matrix
((s) =
d1
¦
+
� I
dz S
4
+ 01 s3 + Ozs
z
+ 03S + 0
4
.
.1 1 s+ .12s�+ .l
3
S + .14
.Z1 s + .Z2S" + .23s + .
2
4
(4.40)
1 04 STATESPACE SOLUTI ONS AND REALI ZATI ONS
Then i t s realization can be obtained directly from . ·. as
. I
· =
y
=
·
I
·1 2
03
o
.
·
¸
o
.
] ·
dl
]
¸
. ·
..
This controllablecanonicalform realization can be read out from the coeffcients of G(s) in
..
There are many ways to realize a proper transfer matrix. For example, Problem .
gi\ e< a different realization of . ·· with dimension rq. Let Gei (s) be the i th column of
G(s ) and let 1i be the ith component of the input vector . Then J(s) = G(s) u(s) can be
ex pressed as
J(s ) = Gel (5) 11 1 (5) GdS) lds) .
=: Ycl (5) Yds)
as shown i n Fig. ... Thus we can realize each column of G(s) and then combine them to
yield a realization of G(s) . Let Gri (5 ) be the i th row of G(s) and let Yi be the ith component
of the output vector
y
. Then y( s) = G(s ) u( s) can be expressed as
Yi (S ) = Gn ( s) u( s)
as shown i n Fig. 44(b). Thus we can realize each row of G(s) and then combine them to obtain
a realization of G(s) . Clearly we can also realize each entry of G(s ) and then combine them
to obtain a realization of G( s) . See Reference · pp. ·· ·
The MATLAB function [ a , b , C , d 1 ¯ t f 2 c: ( num, den) generates the controllable
canonicalform realization shown i n . . for any singleinput multipleoutput transfer matrix
G(s) In its employment, there is no need to decompose G(s) as in . · But we must compute
its least common denominator. not necessarily monic. The next example wi l l apply t f 2 s s to
each column of G(s) in . ·· and then combine them to form a realization of G(s) .
¹�
( b)
Figure 4A Realizations of G( 5) by
columns and by IOWS.
4. 4 Real i zati ons
Typing
yields the following realization for the frst column of G( S ) :
.
]
.
 ·
·
l OS
..
Similarly, the function t f2 s s can generate the fol lowing realization for the second column
of G(s):
)0 = ..
·
=
    
I
V,.· = . d'lh =
·
  
  
I
.
] .

o

·
] .
·

I

These two realizations can be combined as
�  = ª
·
]
·

·
·

]
·
·
· ·
·
.
Y =
Yc1 YL= . · · .
or
�
x =
.
· . · ·
� ] . � �  .
y
=
·
.·
. ..
This is a di fferent realization of the G(s ) i n . ·· This realization has dimension . two less
than the one i n . ·
The two state equations in . · and ... are zerostate equivalent because they have
the same transfer matrix. They are, however, not algebraical ly equi valent. More wi l l be said
1 06 STATESPACE SOLUTI ONS AND REALIZATI ONS
i n Chapter 7 regarding realizations. We mention that al l discussion i n this section, including
tf 2 ss, applies without any modification to the discretetime case
4. 5 Sol uti on of Li near Ti meVaryi ng (LTV) Equati ons
Consider the linear timevarying (LT) state equation
x(t) = A(t)x(t) B(t)u(t)
y(t) = C(t)x(t) + D(t)u(t)
(4.45)
(4.46)
It is assumed that, for every initial state
f
(to) and any input uct), the state equation has a
unique solution. A sufcient condition forisuch an assumption is that every entry of A(t) is a
continuous function of t. Before considering the general case, we frst discuss the solutions of
x(t) = A(t)x(t) and the reasons why the approach taken in the timeinvariant case cannot be
used here.
The solution of the timeinvariant equation x = Ax can be extended from the scalar
equation x = ax. The solution of .1 = ax is x(t) = ea, x(O) with d(ea')/dt = aea' =
e
aa
.
Similarly, the solution of x = Ax is x(t) = e
At
x(O) with
�eAt
= Ae
At
=
eAt
A
dt
where the commutative property i s crucial. Note that, i n general, we have AB i BA and
e(
A
+B)t i eAt eRt .
with
The solution of the scalar timevarying equation x = a(t)x due to x(O) is
( )
f' a(r
)
dr
(0) . t = eJo .
d f' a(r)dr
( )
f' a(r)dr f
o
' a(r)dr
(t) eJ
o
= a t
e
Jo =
e
JI a
dt
Extending this to the matrix case becomes
with, using (3.5 1 ),
x(t) = e'·
A
(r)dr
x(O)
e'·
A
(
r)dr
= 1 + A( r)dr + �A( r)dr·A(S)dS·+
. . .
This extension, however, is not valid because
·e'
A
(
r
)
dr
= A(t) �A(t) A(S)dS·+ �A(r)dr·A(t) +
.
. .
i A(t )e'
A
(r)dr
(4.47)
(4.48)
Thus, i n general, (4.47) is not a solution of x = A(t)x. In conclusion, we cannot extend the
4.5 Solution of Li near TimeVaryi ng (LTV) Equati ons 1 07
solution of scala timevarying equations to the matrix case and must use a diferent approach
to develop the solution.
Consider
x = A(t)x (4.49)
where A i s n x n with continuous functions of t as its entries. Then for every initial state Xi (to),
there exists a unique solution Xi (t), for i = 2, . . , H. We can arrange these H solutions as
X = [X
I
X2 . . , xn] , a square matrix of order H . Because every Xi satisfes (4.49), we have
X(!) = A(t)X(t) (4.50)
If X(to) i s nonsingular or the n initial states are linearly independent, then X(t) i s called a
fndamental matri of (4.49). Because the initial states can arbitrarily be chosen, as long as
they are linearly independent, the fundamental matrix is not unique.
EXAMPLE 4.8 Consider the homogeneous equation
¸ 0 0
x(t ) =
t 0
x(t) (4.5 1 )
or
The solution of XI (t) = o for to = 0 i s .
I
(t) = XI (O) . the solution of X2(t) = tXI (t) = tX
I
(O)
is
Thus we have
and
XI (0)
x(O) = =  x(t) = 2
X
2
(0) 0 0. 5t
.
x(O) =
X2 (0)
=
2
¯
x(t) =
0.5
t2 + 2
The two initial states are linearly independent; thus
X(t) =
'
0. 5t 0. 5t + 2
is a fundamental matrix.
(4.52)
A very imporant property of the fundamental matrix is that X (t) is nonsingular for all t .
For example, XCt) in (4.52) has determinant 0. 5t
2
+ 2 ¯ 0. 5t
2
= 2; thus it i s nonsingular for
all t. We argue intuitively why this is the case. If XCt) is singular at some II , then there exists
a nonzero vector v such that X(tl ) := X(t
I
l v = 0, which, in tr, implies xCt) := X(t)v = 0
for all t, in particular, at = to. This is a contradiction. Thus X(I) is nonsingular for all I .
•
1 08
•
STATESPACE SOLUTI ONS AND REALI ZATI ONS
Defnition 4.2 ·· ..·..·x = ..
 : = ··
..state transition matrix ·x = .......
....·
·
<(1 , = · 
·
..·= I.
·
Because ·i s nonsingular for all its inverse i s well defned, Equation ·· follows
directly from · From the defnition,
.
we have the following important properties of the
state transition matrix:
'
· = I

= · ·
÷ ··
= 
 =  ·
for every and
··
·
·
....·..4.9 Consider the homogeneous equation i n Example · Its fundamental matrix
was computed as
Its i nverse is, using
¸
· ÷ :
 +
+
· =
,
Thus the state transition matrix i s given by
� I
=
2 '
,
+
=
( ¡ �
, ¸
It is straightforward to verify that thi s transition matrix satisfes · and has the three
properties l i sted in ·· through ·
Now we claim that the solution of · · excited by the initial state .÷ .and the
input .is given by
. =  .+ '  .· ·
!
= 
x+   .
[
· ·
·
4. 5 Sol uti on of Li near Ti meVaryi ng ( LTV) Equati ons 1 09
where · is the state transition matrix of x = ..Equation · follows from ·
by using  =  (10, We show that ·· satisfes the initial condition and the
state equaion, At = we have
. = .   . = Ix0 + 0 = .
Thus · · satisfes the initial condition. Using · and · we have
· ·
. = <(1, .+   .
· · ¡¡
= ·  .'�    .
= ·  .+
·  ·  ) U( +  .
´0
÷ ·
 .+   . ¦  .
= .. +  .
Thus · · i s the solution, Substituting · · into · · yields
,= .  .+ .'   . + .
If the input i s identically zero, then Equation · · reduces to
.=  .
·
This is the zeroinput response. Thus the state transition matrix govers the unforced propa
gation of the state vector. If the initial state is zero, then · reduces to
, = .'  ·  . =  .
!
÷ .   + « · .
'0
·
This is the zerostate response. As discussed i n (2. 5), the zerostate response can be described by
,= . ·
i0
·
where · i s the impulse response matrix and is the output at time excited b y an impulse
input applied at time T. Comparing · and · yields
= .  ·  +  ·
= .··
 · +  ·
This relates the inputoutput and statespace descriptions.
1 1 0 STATESPACE SOLUTIONS AND REALIZATI ONS
The solutions in (4.57) and (4.59) hinge on solving (4.49) or (4.53). If ·is triangular
such as
·
'
·
'
·
'
·
=
· · ·
we can solve the scalar equation · =· ·and then substitute it into
Because Xl (t) has been solved, the preceding scalar equation can be solved for · This is
what we did in Example 4.8. If · such as ·diagonal or constant, has the commutative
property
for all and t, then the solution of (4.53) can be shown to be
(4.63)
)¨
\1) J1
1
 = c º =)
1
A(T) dT
k=O
!0
For ·constant, (4.63) reduces to
 r) = c
¨``´ = T)
and ·= c¯´ . Other than the preceding special cases, computing state transition matrices is
:
generally diffcult.
i
i
4. 5. 1 DiscreteTime Case
Consider the discretetime state equation
.+ 1 1 = ·.+ .
·.=..+ .
(4.64)
(4.65)
The set consists of algebraic equations and their solutions can be computed recursively once
the initial state .and the input .for are given. The situation here is much
simpler than the continuoustime case.
As in the continuoustime case, we can defne the discrete state transition matrix as the
solution of
with  = I
+ 1 , =·
for = + I , . . This is the discrete counterpart of (4.53) and its solution can be
obtained directly as
(4.66)
=· · 21 ·
. .
·
for ¯ and  = I.
We discuss a signifcant difference between the continuous and
discretetime cases. Because the fundamental matrix in the continuoustime case is nonsingular
4.6 Equi val ent Ti meVaryi ng Equati ons I I I
for all the state transition matrix  is defned for ¯
propagatIOn of the state vector in the POSI't' t'
_ and and can gover the
.
lve lme and negative tim d'
.
I
l1me case, the Amatrix can be singul
.
th h '
 e IrectlOns. Othe discrete
... .
ar, us t e m verse of 1 b
·
defned only for ¯
a d
'
0 may not e defned. Thus
. . .
_ g n govers the propagation of th t
.
poslllvetlme direction. Therefor th d' .
e s ate vector m only the
e e lscrete counterpart of (4.56) or
 = 
 •
holds only for
Using the discrete state transition mat '
as, for ¯
nx, we can express the solutions of (4. 64) and (4.65)
kI
.=  . .+ _ + .
m=ko
k"
·.= ..  .+ ._ + .+ ..
m=ko
Their derivations are similar to those of (4 '0
.
'
.
If the initial state is zero Equation ( '
'
6

7»
an
d
d (4
.
_1) and wIll not be repeated.
, ... re uces to
k¨
= .) + ....+ ..
m=ko
(4.67)
(4.68)
for ¯ This describes the zerostate re f
then (4.68) can be written as
sponse 0 (4.65). If we defne  .= 0 for
k
·.= ).. + .+  . .
m=ko
where the impulse sequence equals 1 if ¯ d O 'f
the multivariable version of (2.34). we have
¬ an 1 ' Comparing this with
= ..  + + . .
for This relates the impulse response se
counterpart of (4.62).
quence and the state equation and is the discrete
4. 6 Equi val ent Ti meVari ng Equati ons
This section extends the equivalent state e uations d'
. . .
case. Consider the ndimensional J '
q
.
lscussed m Secl10n 4.3 to the lImevarying
mear tlmevarymg state equation
x = ·.+ .
(4.69)
·= ...
1 1 2 STATESPACE SOLUTIONS AND REALIZATI ONS
Let PU) be an H X H matrix. It is assumed that pet) i s nonsingular and both P(t) and P(t) are
continuous for all t. Let s= P(t )x. Then the state equation
where
x = A( t) x + r· · a
y = C( t ) x +p(··a
A( t) = [P(t)A(t) + p(t)]pI (t)
r · · = P(t)B(t)
C(t) = C( t ) pI (t)
O(t) = 
(4. 70)
is said to be (algebraically) equivalent to (4.69) and pet) is called an (a/gebraic) equim/ellce
transformation.
Equation (4.70) is obtained from (4.69) by substituting s= P(t)x and x = P(t)x
¯
P(t)X.
Let X be a fundamental matrix of (4. 69). Then we claim that
x · · := P(t)X(t) (4. 7 1 )
i s a fundamental matrix of (4.70). By defnition, X(I) = A ( t ) X(t) and X(t) i s nonsingular for
all t. Because the rank of a matrix will not change by multiplying a nonsingular matrix, the
matrix P(t)X(t) is also nonsingular for all t. Now we show that P(t)X(t) satisfies the equation
x = A(t)x. Indeed, we have
d . . ,
Tt
[P(t)X(t) ] = P(t) X( t) + P(t)X( t ) = P( t) X(t) + P(t)A(t)X(t)
= [P(t) + P(tl A(t)] [PI (t)P(t)] X( t) = A( t ) [P(t)X(tl ]
Thus P(t)X(t) is a fundamental matrix of x(t) = A(t) x( t) .
Theorem 4.3
Let Ao b an arbitrary constant matrix. Then there exists an equivalence transformation that transfomls
(4.69) into (4.70) with A(t) = Ao.
Proof' Let XU) be a fundamental matrix of x = A(t)x. The differentiation of XI · ·
XU) = I yields
XI (t)Xul + X I X(t ) = 0
which implies
(4. 72)
Because AU) = Ao is a constant matrix, X(t) = eAQ' is a fundamental matrix of
x = A(t)s÷ Aos. Following (4. 7 \ ). we defne
(4. 73)
4. 6 Equivalent TimeVaryi ng Equati ons
and compute
A( t ) = [P( t l A( t )
¯
p(t)] pI (I )
= [eA"' XI (t)A(t) +AoeAO' XI (t) + eAo'X I U) ]X(t)eAQ'
whiG becomes, after substituting (4.72),
A(t ) = AoeAQ'X1 (t)X(t)eAo' = Ao
This establishes the theorem. Q.E. D.
If Ao is chosen as a zero matrix, then P(t) = X l et ) and (4. 70) reduces to
1 1 3
A( t ) = 0 r(t) = XI (t)B(t) C(t) = C(t)X(t) O(t) = D( t) (.4)
The block diagrams of (4.69) with A(t) o 0 and A(t) = 0 are plotted in Fig. 4. 5. The block
diagram with A(t) = 0 has no feedback and is considerably simpler. Every timevarying state
equation can be transformed into such a block diagram. However, in order to do so, we must
know its fundamental matrix.
The impulse response matrix of (4.69) is given in (4. 62). The impulse response matrix of
(4.70) is, using (4. 7 1 ) and (4. 72),
G(t, r) = t··\··x¯(t· r t·
¯
0t)·(· T )
u
(a)
u
(b)
Figur 4.5 Block daigrams with feedback and without feedback.
1 1 4 STATESPACE SOLUTI ONS AND REALI ZATI ONS
÷ C(tlP
1
(t)P(t)X(t)X
I
(r)p
l
(r)P(r)B(r) + D(t)o (t  r)
= C(t)X(t)X1 (r)B(r) + D(t)o(r  r) ÷ G(t, r)
Thus the impulse response matrix is invariant under any equivalence transfonnation. The
property of the Amatrix, however, may not be preserved in equivalence transfonnations. For
example, every Amatrix can be transfonned, as shown in Theorem 4. 3, into a constant or a
zero matrix. Clearly the zero matrix does not have any property of A( t) . In the timeinvariant
case, equivalence transfonnations will preserve all properties of the original state equation.
Thus the equivalence transfonnation in the timeinvariant case is not a special case of the
timevarying case.
Defnition 4.3 A matrix P(t) is called � Lyapunov transfonnation ifP(r) is nonsingular,
pet) and
p
et) are continuous, andP(t)
h
ndpI (t) are boundedforall t. Equations (4.69)
and (4.70) are said to be Lyapunov equivalent iP(t) is a Lyapunov transforatioll.
It is clear that if P(t) ÷ P is a constant matrix, then it is a Lyapunov transfonnation. Thus
the (algebraic) transfonnation in the timeinvariant case is a special case of the Lyapunov
transfonation. If pet) is required to be a Lyapunov transfonnation, then Theorem 4.3 does
not hold in general. In other words, not every timevarying state equation can be Lyapunov
equivalent to a state equation with a constant Amatrix. However, this is true if A(t) is periodic.
Periodic state equations Consider the linear timevarying state equation in (4.69). It is
assumed that
A(t + T) ÷ A(t)
for all t and for some positive constant T. That is, A(t) is periodic with period T. Let X(t) be
a fundamental matrix of X ÷ A(t)x or
X
(t) = A(t)X(t) with X(O) nonsingular. Then we have
X
(t + T) ÷ A(t + T)X(t + 7) = A(t)X(t + T)
Thus XU + T) is also a fundamental matrix. Furthennore, it can be expressed as
XU + T) = X(t)XI
(O)X(T) (4. 75)
This can be verifed by direct substitution. Let us defne Q = XI (O) X( T) . It is a constant
nonsingular matrix. For this Q there exists a constant matrix
A such that
/
T = Q (Problem
3.24). Thus (4.75) can be written as
X(t + T) = X(t)e
AT
Defne
We show that pet) is periodic with period T:
pet + T) ¯ eA(t+T)XI
(t + T) = e
Al iT
[
e
AT
X
 I (t)]
¯eAI X
1 (t) = per)
(4.76)
(4.77)
4.7 TimeVarying Real izations
1 1 5
.. Theorem 4,4
Consider (4.69) with �U) = A(t + T) for all t and some T ¯ 0 Let X(t) b f d t
.
of X  A(t)x L t A b h
.
a un amen matnx
 . e e t e constant matrix computed from eAT ~
X
I
Lyapunov equivalent to
m
(O)X( T) . Then (4.69) is
x
(t) ÷
A
i(t) + P(t ) B(t)u(t)
yet) ÷ C(tlP
I
(t)i(t) + D(t)u(t)
where pet) = e
A
I X
I
(r).
The matrix pet) in (4.77) satisfes all conditions in Defnition 4 3· th s ' t
.
L transfonnatlOn. The rest of the theorem follows directly from Theore�·3
u
�e
�
o
a yapunov
part of Theorem 4.4 is the theory of Floquet. It states that ifX = A(t)x a�d
'
if A(t + ��
g
`
n
��
u
,
for all t, then Its fundamental matrix is of the fonn p
I
(t )eAI
h p
_l . �
function. Furthennore x ÷ A(t)x is Lyapu Õ
.
I
q � ere (t) IS a penodlc

H v eqUiva ent to x ÷ Ai.
4.7 Ti meVaring Real izations
We studied in Section 4.4 the realization problem for linear time
. .
.
section, we study the corresponding problem for linear ti
�
mvanant systems. I thiS
trans
t
E
onn ca
l
.
nnot b
.
e used here; therefore we study the prob
��
v
¨��·
y
s
i
�
s
'�¯�'
m
��
o
�
a
��
ce
very mear tImevarymg syste b d
. .
m can e escnbed by the inpututput description
y(t) = G(t, r) u( r) dr
r0
and, if the system is lumped as well, by the state equation
x(t) ÷ A(t)x(t) + B(t ) u(t )
yet) ÷ C(t)x(t) + D(t)u(t)
(4,78)
If the state equation is available, the impulse response matrix can be computed from
G(t, r) = C(t)X(t)XI (r)B(r) + D(t ) o(t  r) for t :: r (4.79)
where X(t) is a fundamental matrix of x ÷ A(t)x T
.
.
f
. . e converse problem IS to fnd a state equatIOn rom a given impulse response matrix A ' I
b r b
. n Impu se response matrix G(t r) is said t e rea lza Ie if there exists ( ACt), B(t), C(t), D(t ) l to meet (4.79).
1 Õ
.. Theorem 4.5
A q X p impulse response matrix G(t, r) is realizable i f and only if G(t r) b d , can ecomposed ð
G(t, r) ~ M(t)N(r) + D(t) o (t  r) (4.80)
:�;e;�r
�`
r, where M, N, and D are, respectively, q x n
,
n x p, and q x p matrices for some
1 1 0 b!^JLbÍ^'L b'Í'JLlb ^^L KL^LlZ^1l L"b
Proof: If G(t, r) is realizable, there exists a realization that meets (4. 79). Identifying
1(1) = C( t) X( t) and N( r) = XI ( r) B( r) establishes the necessary pan of the theorem.
If G(t. r) can be decomposed as in (4.80), then the ,dimensional state equation
x(t) = N(t)u(t)
(4. 81 )
y(t) = 1(l) x(t) + D(t)u(t)
is a realization. Indeed, a fundamental matrix of x = O · x is X( t) = 1. Thus the impulse
response matrix of (4. 81 ) is
M(t)I · r' Ne r) ¬ D(t)S(t  r)
which equals G(t. f) . This shows the suffciency of the theorem. Q.E.D.
Although Theorem 4.5 can also be applied to timeinvariant systems. the result is not
useful i n practical implementation, as the next example i llustrates.
EXAMPLE 4.10 Consider get) = teA' or
get, r) = g( t  r) = (t  r) ewn
It is straightforward to verify
get  r ) = [e"
Thus the twodimensional timevarying state equation
x(t) =
[
� �]
x ¬
� [u(
t)
y(t) = leA' te"]x(t)
is a realization of the impulse response g(t) ¯ tei,.
The Laplace transform of the impulse response is
 (5) '[te" ]
I
g

(5  i)�
¬
52  2is + )
Using (4.41 ). we can readily obtain
[
7
x(tl = ¸
´
yet) = [
0
l ]x( )
(4.82)
(4.83)
This LTI state equation is a different realization of the same impulse response.This realization
is clearly more desirable because it can readily be implemented using an opamp circuit. The
implementation of (4.82) is much more difficult in practice.
´`` ¯� ¯¯ 4
=suw
rOUlC0S 1 1 /
4,1 An oscillation can be generated by
,
Show that its solution is
x =
[
0
[
.
1 0
[
COS t
xU) = .
 SIn l
sin
[
cos t
x(O)
Use two di fferent methods to fnd the unitstep response of
x =
[
0
l
]
x +
[
l
] u
2 2 I
y = [2 3] x
4.3 Discretize the state equation i n Problem 4.2 for T = I and T = D,
4.4 Find the companionform and modalform equivalent equations of
4,5
4.6
4,7
x = �
2
�
2
X ¬ �u
y = [ I  Í O]x
Find an equivalent state equation of the equation i n Problem 4.4 so that all state variables
have their largest magnitudes roughly equal to the largest magnitude of the output. I f
all signals are required t o lie inside J1 0 volts and i f the input is a step function with
magnitude u. what is the penissible largest aO
Consider
x =
[
� x +
[
�:
[
ll
where the overbar denotes complex conjugate. Verify that the equation can be trans
formed into
with

[
0
A = 
¬/
´
À
x = Ax ¬ bll
by using the transformation .= Qx with
Q
[

�bl
l ¬
.
ib,
Verify that the Jordanform equation
\ = ex
1 1 8 STATESPACE SOLUTIONS AND REALIZATIONS
). 0 0 0 0
0 ). 0 0 0
0 0 ). 0 0 0
�
bl
u
0 0 0 0
0 0 0 0 � b2
0 0 0 0 0 � b3
Y [cl C2 C3 CI C2 C3X
can be transformed into
12
.
'
A
[CI C2 c3]i
0
where A
,
i
,
and Ci are defned in Problem 4.6 and 12 is the unit matrix of order 2. [Hint:
Change the order of the state variables from [Xl X2 X3 X4 Xs X61' to [Xl X4 Xz Xs
X3 x61' and then apply the equivalence tansformation Qi with Q diag(QI,
Q2 , Q
.
4.8 Are the two sets of state equations
4.9
[ 1 !
and
equivalent? Ae they zerostate equivalent?
Verif that te transfer matrix in has the following realization:
al lq Iq
a2lq 0
i � [
ar

I lq 0
arlq 0
0
0
0
J
o
2
u
� �

l
This is called the and has dimension rq. It is dual
to (4.34).
4.10 Consider the I x proper rational matix
A !
[dl d2] + ¯ ¯ : ¯¯
S
4
a2s
2
+ aa4
Problems 1 1 9
X
Show that its observable canonical form realization can be reduced from Problem 4. 9 as
x ~
a
o 0
1 0 0
X

a4
o 0 1
o 0 0
4.11 Find a realization for the proper rational matrix
��
2)
 
+ 1 2
4.12 Find a realization for each column of in Problem 4. 1 1 and then connect them,
as shown in Fig. 4.4(a), to obtain a realization of What is the dimension of this
realization? Compare this dimension with the one in Problem 4. 1 1 .
4.13 Find a realization for each row of in Problem 4. 1 1 and then connect them, as shown
in Fig. 4.4(b), to obtain a realization of What is the dimension of this realization?
Compare this dimension with the ones in Problems 4. 1 1 and 4. 1 2.
4.14 Find a realization for
02S 6) 22s 23
¦ 34
4.15 Consider the ndimensional state equation
x . bu y
Let be its transfer function. Show that has m zeros or, equivalently. the
numerator of has degree m if and only if
for ~ 0, I , 2, . . . , H  m  2
and eA
n

m

I b f 0. Or, equivalently, the diference between the degrees of the denom
inator and numerator of g is  m if and only if
eA
a
I b f 0 and eAib 0
for i 0, 1 , 2, 2.
4. 16 Find fundamental matrices and state transition matrices for
. 0
¦
o
1 20 STATESPACE SOLUTI ONS AND REALIZATI ONS
and
. 1
X =
o
4.17 Show a < (to. t) /at = <{to, t) A(t) .
4.18 Given
show
4.19 Let
<l l (t. to)
<( t. to) =
<2' (t , to)
b the state transition matrix of
<dt , to)
<dt , to)
x(t) =
A'� (t) A1 2 (t)
An(t )
x(t)
Show that <2, (t . to) = 0 for all t and to and that ( a/at ) <ii ( t, to) = Ai . <, , ( ( , to) . for
i = 1 , 2.
4.20 Find the state transition matrix of
sin t
X =
0
4.21 Verfy that X(t) = eA'CeB' is the solution of
X(O) = C
4.22 Show that if A(t) = A, A( t)  A(t)A, . then
A(t) = eA" A(O)e�A"
Show also that the eigenvalues of A (t) are independent of t.
4.23 Find an equivalent timeinvariant state equation of the equation i n Problem 4. 20,
4. 24 Transform a timeinvariant (A. B. C) into (0. B(t ) . eet) by a timevarying equi\'alence
transformation,
4.25 Find a timevarying realization and a timeinvariant realization of the impulse response
g(t) = tZe
A
I ,
4.26 Find a realization of gU. r) = sin t ee�(t�r) cos T. I s i t possible to fnd a timeinvariant
state equation realization�
LH3DICI
Stabi l i ty
5. 1 I ntroducti on
Systems are designed to perform some tasks or to process signals. If a system is not stable. the
system may bum out. disintegrate. or saturate when a signal. no matter how small. is applied.
Therefore an unstable system i s useless in practice and stability i s a basic requirement for all
systems, In addition to stability. systems must meet other requirements. such as to track desired
signals and to suppress noise. to be really useful in practice.
The response of linear systems can always be decomposed as the zerostate response
and the zeroinput response. It is customary to study the stabilities of these two responses
separately. We will i ntroduce the BlBO (boundedinput boundedoutput) stability for the zero
state response and marginal and asymptotic stabilities for the zeroinput response, We study
frst the timeinvariant case and then the timevarying case.
5. 2 I nputOutput Stabi l i ty of LTI Systems
Consider a SISO linear timeinvariant (LTI) system described by
y( t) =
get  r) u(r ) dr = g( r) u(t  r ) dr (5 1 )
where ge t ) is the impulse response or the output excited by an i mpulse input applied at t = O.
Recall that in order to be describable by (5. 1 ). the system must be linear. timeinvariant. and
causal. In addition. the system must be initially relaxed at t = 0,
1 21
1 22 STABI LITY
An input u(t) is said to be bounded if U(/) does not grow to positive or negative infnity
or, equivalently, there exists a constant Ur such that
l u(t ) 1 � Ur < 0 for all I :: °
A system is said to be BIO stable (boundedinput boundedoutput stable) if every bounded
input excites a bounded output. This stability is defned for the zerostate response and is
applicable only if the system is initially relaxed.
� Theorem 5. 1
A SISO system describedby (5. 1 ) i s BlEO stable if and only if g(t) is absolutely integrable in [0, (), or
·I
g't)l dt � M < 0
for some constant M.
Proof: First we show that if g(t) is absolutely integrable, then every bounded input excites
a bounded output. Let U(/) be ûarbitrary input with ! u(t ) 1 � Ur < 0 for all t :: 0. Then
l y(t) 1 = g(r)u(t  r) dr � I g(r) l l u (t  r) l dr
� Ur I g(r) l dr � um M
Thus the output i s bounded. Next we show intuitively that if get) i s not absolutely
integrable, then the system is not BIBO stable. If get) is not absolutely integrable, then
there exists a tl such that
Let us choose
I g(r) 1 dr ~ 0
u(t\  r) =
_
�
if g(r) :: °
if g(r) < °
Clearly u is bounded. However. the output excited by this input equals
y(tt l = g\r)II(t\  r) dr = I g( r) 1 dr = 0
which is not bounded. Thus the system is not BIBO stable. This completes the proof of
Theorem 5. 1 . Q.E.D.
A function that is absolutely integrable may not be bounded or may not approach zero as
t  0. Indeed, consider the functon defned by
n + (t  n)n4
f(1  n) =
4
n ¯ (I  n)n
for n  I /n3 � I � n
for n < I � n + I /n3
for n = 2, 3, . . , and plotted in Fig. 5. 1 . The area under each triangle is l /nz. Thus the absolute
'
5. 2 InputOutput Stabi lity of LTI Systems 1 23
Figr 5.1 Function.
n
L 4 0
integration of the function equals L::
z
(llnz) < 0. This function is absolutely integrable
but is not bounded and does not approach zero a I  0.
� Theorem 5.2
•
If a system with impulse response g(t ) is BIBO stable, then, ð I  0:
1. The output excited by u(t) = a, for t :: 0, approaches g(O) . a.
2. The output excited by u ( I) = si n wot, for I :: 0, approaches
I g(jwo) 1 sin(wot+ 1:g(jwo))
where g(5) is the Laplace tansformof g(t) or
g(5) ¬ ¡g(r)esr dr
Proof: If u (t) = a for all I :: 0, then (5. 1 ) becomes
y(t) = g( r) U(1  r) dr = a ,g(r) dr
which implies
y(/)  a g( r) dr = ag(O) aS / + o
(5.2)
where we have used (5.2) with 5 = 0. This establishes the first part of Theorem 5. 2. If
u (t) = sin wol, then (5. 1 ) becomes
y(t) = g(r) sin Wo(1  r) dr
= g(r) [sin wot cos Wo r  cos wol si n war] dr
= sinwol g( r) cos wor dr  COS Wol g(r) sin wor dr
1 24 STABI LITY
Thus we have, as I  C,
yU) ¨ sin wal g(r) cos Wo r dr  cos Wol g( r) sin Wo r dr
If g(l) is absolutely integrable, we can replace s by jw in (5. 2) to yield
g(jw) =
·
g( r) [cos wr  j sin wr] dr
The impulse response g(t) i s assumed implicitly to be real : thus we have
Re[g(jw)] = g( r) cos wr dr
and
Ì
Im[g(jw)] = ¯ g(r) si n wr dr
( 5. 3)
where Re and 1m denote, respectively, the real part and imaginary pan. Substituting these
into (5. 3) yields
y(t) ¨ si n wol (Re[g(jwo) ] ) + cos wot(Im[g(jwo) ] )
= [ g( jwo ) [ si n(wol+ tg(jwo»
This completes the proof of Theorem 5.2. Q.E.D.
Theorem 5.2 i s a basic result; fl tering of signals is based essentially on this theorem.
Next we state the BlBO stabi l i ty condition in terms of proper rational transfer functions.
t. Theorem 5.3
A SISO system with proper rational transfer function g(s) is BlBO stable if and only if every pole of
g(s) has a negative real part or. equivalently, lies inside the lefthalf .plane.
If g(s) has pole Pi with multiplicity U, then its partial fraction expansion contains the
factors
.I  P, (s  p, )2 ' (
.I  p
,
)
m,
Thus the inverse Laplace transform of g (.) or the impulse response contains the factors
It is straightforward to verify that every such term is absolutely integrable if and only if P, has
a negative real part. Using this fact, we can establ i sh Theorem 5. 3.
EXAMPLE 5. 1 Consider the positive feedback system shown i n Fi g. 2. 5(a 1 . Its impulse response
was computed in (2.9) as
O
g(l) = )al 8( t 
i =!
5. 2 I nputOutput Stabi l i ty of LTI Systems 1 25
where the gain a can be positive or negative. The impulse is defned as the limit of the pulse
i n Fig. 2. 3 and can be considered to be positive. Thus we have
·
[ g( t ) 1 = )[ al
i
8( t  i )
i =l
and
if l al :: 1
i f l ai < 1
Thus we conclude that the positive feedback system in Fig. 2. 5(a) is BI BO stable if and onl y
i f the gai n a has a magnitude l ess than 1 .
The transfer function of the system was computed i n (2. 1 2) as
aeS
g(s) = 
I  aeS
It i s an irrational function of .I and Theorem 5. 3 i s not applicable. In this case, i t i s simpler to
use Theorem 5 . 1 to check its stabi l i ty.
For multi variable systems, we have the following results.
� Theorem 5. M1
A multivariable system with impulse response matrix G(I)
every gij
(! ) is absolutely integrable i n [0, :).
[g,
j
(t) ] is Bl BO stable i f and onl y if
;. Theorem 5. M3
A multivariable system with proper rational transfer matrix G(s) = [gij (.I ) ] is BlBO stable if and only
if every pole of every g'i (5) has a negative real part.
We now discuss the BlBO stability of state equations. Consider
Its transfer matrix i s
x( t ) = Ax(t) + Bu(t)
yet) = Cx(t) + Du(t )
(5.4)
Thus Equation (5.4) or. to be more precise. the zerostate response of (5.4) i s BIBO stable i f
and onl y i f every pole of G(s) has a negative real part. Recall that every pol e of every entry
of G(s) i s called a pole of G(s) .
of
We discuss the relationship between the poles of G(s) and the eigenvalues of A. Because
 1
.. = C[Adj (sI  A)]B + D
det(sI  A)
(5. 5)
1 26 STABI LITY
every pole of G(s) is an eigenvalue of A. Thus if every eigenvalue of A has a negative real
part, then every pole has a negative real part and (5.4) is BIBO stable. On the other hand,
because of possible cancellation in not every eigenvalue is a pole. Thus, even if A has
some eigenvalues with zero or positive real part, (5. 4) may still be BIBO stable, as the next
example shows.
EXAMPLE 5.2 Consider the network shown in Fig. 4.2(b). Its state equation was derived in
Example 4.4 8
~ + O · . ~ + .
The Amatrix is I and its eigenvalue is 1 . It has a positive real part. The transfer function of
the equation is
I
g(s= 1 ) 1 + =
The transfer function equals It has no pole and no condition to meet. Thus is BIBO
stable even though it has an eigenvalue with a positive real part. We mention that BIBO stability
does not say anything about the zeroinput response, which will be discussed later.
5. 2. 1 DiscreteTime Case
Consider a discretetime 8180 system described by
~ _ . ..= _. . .
. .
where is the impulse response sequence or the output sequence excited by an impulse
sequence applied at = O. Recall that in order to be describable by the discretetime
system must b linear, timeinvariant, and causal. In addition, the system must be initially
relaxed at = O.
An input sequence .is said to be ... .does not grow to positive or negative
infnity or there exists a constant Ur such that
. :: Ur < 0 for = 0, 1 , 2, . . .
A system is said to be BIBO .(boundedinput boundedoutput stable) if every bounded
input sequence excites a boundedoutput sequence. This stability is defned for the zerostate
response and is applicable only if the system is initially relaxed.
� Therem 5.01
A discretetime SISO system described by (5.7) is BIBO stable i f and only if i s absolutely summable
in 0) or
O
_ :: M < 0
for some constant M.
' 5. 2 I nputOutput Stabi lity of LTI Systems 1 27
Its proof is similar to the proof of Theorem and will not be repeated. We give a discrete
counterpart of Theorem in the following.
Theorem 5.02
If a discretetime system with impulse response sequence is BIBO stable. then. as � 0:
1. The output excited by = a. for 2 O. approaches gO) ' a.
2. Te output excited by H = sin .for 2 approaches
I g(eiw' ) 1 sin(wok
+
jg(eiw,»
where g(:) is the :transfon of or
¯
g(;) ~ _
.
.
Proof If = a for all 2 0, then becomes
which implies
k k
= = a _
o
¯
� a _. = agO) as ¬ 0
.
(5. 8)
where we have used with z = This establishes the frst part of Theorem 5.D2. If
= sin . then becomes
= _ sin . .
.
= _ . (sin . cos .. . ..
.
= sin ._. .¯ cos ._sin .
Thus we have. as � 0.
.
¯ ¯
¯ sin ._ ..¯ cos ._.sin ..
.
If is absolutely summable, we can replace Z by eiw in to yield
O >
g(eiw) = _.
= _. . j sin .
. .
1 28 STABI LI TY
Thus (5.9) becomes
y[k) ¯ sin wok(Re[g(ejwo ) ) ) + cos wok (Im[g (ejw, ) ) )
= I g(ejw" ) i sin(wok+ jg(ejw, »
This completes the proof of Theorem 5. 02. Q. E. D.
Theorem 5. 02 is a basic result in digital signal processing. Next we state the BIBO
stability in terms of discrete proper rational transfer functions.
Theorem 5.03
A discretetim
.
e S1S0 system with proper rational transfer function g<z) is BlBO stable if and only if
every pole of g(�) has a magnitude less than I o�, equivalently, lies inside the unit circle on the :plane.
If g( �) has pole p, with multiplicity U, then its partial fraction expansion contains the
factors
;  Pi (;  p
; l " , (
;
 P,
)m,
Thus the inverse ztransform of g(z) or the impulse response sequence contains the factors
p7, kp� , " ' , km, I p;
It is straightforward to verify that every such term is absolutely summable if and only if Pi has
a magnitude less than I . Using this fact, we can establish Theorem 5. 03.
In the continuoustime case, an absolutely integrable function jU) , as shown in Fig. 5 . 1 ,
may not be bounded and may not approach zero as t ¬ ¯. In the discretetime case, if g[k)
¡5 absolutely summable, then it must be bounded and approach zero as k = ¯. However. the
converse is not true as the next example shows.
EXA�IPLE 5.3 Consider a discretetime LTl system with impulse response sequence g[k) =
1 / k. for k = I . 2, . . . , and g[O] = O. We compute
�
I I I I
S : = � I g[ kl l = _ = 1 +  +  +  +
. . .
k� k=1
k 2 3
4
= I + �+ �
+
�+ �+ . . .
+ �,+ �+ ' " + � . . .
2 3 4 5 8 9 1 6 '
There are two terms, each is 'or larger. in the frst pair of parentheses; therefore their sum
is larger than ' There are four terms, each is � or larger, in the second pair of parentheses;
therefore their sum is larger than 'Proceeding forward we conclude
I I I
S > 1 +  +  +  + " ' = 0
2 2 2
This impulse response sequence is bounded and approaches 0 as k = ¯ but is not absolutelv
¡
summable. Thus the discretetime system is not Bl BO stable according to Theorem 5. 0 I . Th�
1 transfer function of the system can be shown to equal
5. 3 Internal Stabi l ity 1 29
It is not a rational function of ¯ and Theorem 5.03 is not applicable.
For multi variable discretetime systems, we have the following results.
Theorem 5. MO 1
A MIMO discretetime system with impulse response sequence matrix G[k) = [g'
j
[k]] is BlBO stable
if and only if every gi
j
[k] is absolutely summable.
Theorem 5. M03
A MIMO discretetime system with discrete proper rational transfer matrix c �· = . ,.is BlBO
stable i f and only if every pole of every gi
j
(Z) has a magnitude less than I .
We now discuss the BIBO stability of discretetime state equations. Consider
Its di screte trans fer matrix is
x[k + I ) = Ax[k] + Bu[k]
y[k] = Cx[k] + Du[k]
(5. 1 0)
Thus Equation ( 5. 10) or. to be more precise, the zerostate response of (5. 1 0) is BIBO stable
if and only if every pole of c(z) has a magnitude less than I .
We discuss the relationship between the poles of G(z) and the eigenvalues of A. Because
of
. I
GU = C[Adj (zI  A») B + D
det(:I  A)
every pole of c·· i s an eigenvalue of A. Thus if every eigenvalue of A has a negative real
part, then (5. 1 0) is BIBO stable. On the other hand, even if A has some eigenvalues with zero
or positive real part. (5. 1 0) may, as in the continuoustime case, still be BIBO stable.
5, 3 I nternal Stabi l ity
The BIBO stability is defned for the zerostate response. Now we study the stability of the
zeroinput response or the response of
x( t) = Ax(t)
excited by nonzero initial state Xo' Clearly, the solution of (5. 1 1 ) is
X( I ) = eAtxo
(5. 1 1 )
(5. 1 2)
1 30 STABI LI TY
Defnition 5.1 The :eroinput response of(5. 4) or the equation x = Ax is marginally
stable or stable in the sense of Lyapunov i ever fnite initial state Xo excites a bounded
response. It is asymptotically stable ieverfnite initial state excites a bounded response,
which, in addition, approaches 0 as t ¯ 0.
We mention that this defnition is applicable only to linear systems. The defnition that
is applicable to both linear and nonlinear systems must be defned using the concept of
equivalence states and can be found. for example, in Reference [6, pp. 401 03]. This text
studies only linear systems: therefore we use the simplifed Defnition 5 . 1 .
Theorem 5.4
I
1. The equation x ~ Ax is marginally stable ;i and only if all eigenvalues of A have zero or negative
real parts and those with zero real parts are simple roots of the minimal polynomial of A.
2. The equation X = Ax is asymptotically stable if and only if all eigenvalues of A have negative real
parts.
We frst mention that any (algebraic) equivalence transformation will not alter the stability
of a state equation. Consider x ~ P, where P is a nonsingular matrix. Then x = Ax is
equivalent to x = Ax = PAPI X. Because P is nonsingular, if x i s bounded, so is x; if x
approaches 0 as t  0, so does .Thus we may study the stability of A by using A. Note
that the eigenvalues of A and of A are the same as discussed in Section 4. 3.
The response of x = Ax excited by .equals xU) = eAt x(O). It is clear that the
response is b
.
o
unded if and only if every entry of eAt is bounded for all t :: If A is in Jordan
form, then eAt is of the form shown in (3. 48). Using (3.48), we can show that if an eigenvalue
has a negative real part, then every entry of (3. 48) is bounded and approaches 0 as t  0.
If an eigenvalue has zero real part and has no Jordan block of order 2 or higher, then the
corresponding entry in (3. 48) is a constant or is si nusoidal for all t and is, therefore. bounded.
This establishes the suffciency of the frst part of Theorem 5.4. If A has an eigenvalue with a
positive real part, then every entry in (3. 48) will grow without bound. If A has an eigenvalue
with zero real part and its Jordan block has order 2 or higher, then (3. 48) has at least one entry
that grows unbounded. This completes the proof of the frst part. To be asymptotically stable,
every entry of (3. 48) must approach zero as t  0. Thus no eigenvalue with zero real part is
permitted. This establishes the second par of the theroem.
EXAMPLE 5.4 Consider
x =
���
x
o 0  I
Its characteristic polynomial is �(A) = A
"
(A + I ) and its mimimal polynomial is >(A)
A(A + I). The matrix has eigenvalues 0, 0, and I . The eigenvlaue 0 is a simple root of the
minimal polynomial. Thus the equation is marginally stable. The equation
'
5. 3 Internal Stabi lity 1 31
x =
�
�
�
x
o 0 I
i s not marginally stable, however, because its minimal polynomial is A
2
(A + I ) and A = 0 is
not a simple root of the minimal polynomial.
As discussed earlier, every pole of the transfer matrix
((s) = C(sI  A)I B + D
is an eigenvalue of A. Thus asymptotic stability implies BIBO stability. Note that asymptotic
stability i s defned for the zeroinput response, whereas BIBO stability is defned for the
zerostate response. The system in Example 5.2 has eigenvalue and i s not asymptoti
cally stable; however, it is BIBO stable. Thus BIBO stability. in general, does not im
ply asymptotic stability. We mention that marginal stability is useful only in the design
of oscillators. Other than oscillators, every physical system is designed to be asymptoti
cally stable or BIBO stable with some additional conditions, as we will discuss in Chap
ter 7.
5.3. 1 DiscreteTime Case
This subsection studies the interal stability of discretetime systems or the stability of
x[k + = Ax[k]
excited by nonzero initial state Xo. The solution of (5. 1 3) is, as derived in (4.20).
x[k] = Akxo
(5. 1 3)
(5. 1 4)
Equation (5. 1 3) is said to be marginally stable or stable in the sense ofLyapunov if every
fnite initial state Xo excites a bounded response. It is asymptotically stable if every finite
initial state excites a bounded response. which, in addition, approaches 0 as k ¬ 0. These
defnitions are identical to the continuoustime case.
Theorem 5.04
1. The equation x[k + ~ Ax[k] is marginally stable if and only if all eigenvalues of A have
magnitudes less than or equal to I and those equal to I are simple roots of the minimal polynomial of
A.
2. The equation x[k + 1 ] = Ax[k] is asymptotically stable if and only if all eigenvalues of A have
magnitudes less than I .
As in the continuoustime case, any ( algebraic) equivalence transformation will not alter
the stability of a state equation. Thus we can use Jordan form to establish the theorem. The
proof is similar to the continuoustime case and will not be repeated. Asymptotic stability
1 32 STABI LITY
implies BI BO stability but not the converse. We mention that marginal stability is useful only
i n the design of discretetime oscillators. Other than osci l l ators, every discretetime physical
system i s designed to be asymptotical l y stable or BlBO stable with some additional conditions,
as we will discuss i n Chapter ·
5, 4 Lyapunov Theorem
This section introduces a different method of checking asymptotic stability of x = Ax. For
convenience. we call A stable if every eigenvalue of A has a negative real part.
Theorem 5.5
All eigenvalues of A have negative real parts' i f and only if for any given positive defnite symmetric
matrix N. the Lrapunol' equation
A' M + MA = N (5. 1 5)
has a unique symmetric solution 11 and NI is positive defnite.
Corollary 5. 5
All eigenvalues of an n x n matrix A have negative real parts if and onl y if for any given m x n matrix
Ilwith m < II and with the property
NA
rank 0 : = rank
NA:n 1
= n (full column rank)
where 0 is an /1111 x n matrix. the Lyapunov equation
A'M + MA = N' N =: N
has a unique symmetric solution :1 and II is positive defnite.
(5. 1 6)
(5. 1 7)
For any N, the matrix ! in · · is positive semidefnite (Theorem · · Theorem · ·
and its corollary are valid for any given N: therefore we shall use the simplest possible N.
Even so. using them t o check stability of A is HOÎ simple. It is much simpler t o compute,
using MATL\B, the eigenvalues of A and then check their real parts. Thus the importance
of Theorem · · and its corollary i s not i n checking the stability of A but rather in studying
the stability of nonlinear systems. They are essential in using the socalled second method of
Lyapunov. We mention that Corollary · · can be used to prove the RouthHurwitz test. See
Reference · pp. . ·.
Proof ofTheorem 5.5 Necessitv: Equation · · i s a special case of · · with A = A'
and B = A. Because A and A' have the same set of eigenvalues, if A is stable, A has no
two eigenvalues such that A, + A
J
= O. Thus the Lyapunov equation is nonsingular and
has a unique solution M for any N. We claim that the solution can be expressed as
M =
·
eA' f Ne\fdt
Indeed, substituting · · into · · yields
5.4 Lyapunov Theorem
A' M T MA = A'eA f NeAf dl + eA'f NeAf Adt
= eA'f Ne�f ¡
d
t =
eA
1
NeAf
= 0  N = N
1 33
· ·
·
where we have used the fact eAf = 0 at I = ¯ for stable A. This shows that t he M i n
· · i s t he solution. It i s clear that if : is symmetric. so is M. Let us decompose N as
N = N'N, where Ilis nonsingular (Theorem · · and consider
x'Mx = x'eA' f N'�eAf xdl = i : NeAf xl l � dl · .
Because both Il and eAf are nonsingular, for any nonzero x, the integrand of · . i s
positive for every I. Thus x' Mx i s positive for any x ¥ O. Thi , shows the positive
defniteness of M.
S
u
fciency: We show that i fN and :1 are positive defnite, then A is stable. Let A be an
eigenvalue of A and v ¥ 0 be a corresponding eigenvector: that is, Av = ;.v. Even though
A is a real matrix, its eigenvalue and eigenvector can be complex, as shown i n Example · ·
Taking the complexconjugate transpose of Av = AV yields v* A * = ¹¯A' = i. 'v', where
the asterisk denotes complexconjugate transpose. Premultiplying v* and postmultiplying
v to · ·. yields
v*Nv = v* A'lh + v*MAv
= (1.* + i.)v*Mv = 2Re(A)v*Mv · .
Because v*Mv and v*Nv are, as discussed in Section · both real and positive, · .
implies ReO,) < O. This shows that every eigenvalue of A has a negatiw real part.
Q. E. D.
The proof of Corollary · · fol lows the proof of Theorem ·· wi th some modi fc
�
tion. We
discuss only where the proof of Theorem · · is not applicable. Consider · . :ow N is m x n
with m < n and N = N'N is positive semidefnite. Even so. M in · · can still be positive
definite if the integrand of · . is not identically zero for ali I . Suppose the integ
�
and of · .
is identically zero or Ne'"x = O. Then its derivative with respect to î yields :AeAfx = O.
Proceeding forward, we can obtain
· ..
This equation implies that. because of · · and the nonsingularity of eAf . the only x meeting
1 3 STABI LITY
(5.22) is O. Thus the integrand of (5.20) cannot be identically zero for any x = O. Thus M is
positive defnite under the condition in (5. 1 6). This shows the necessity of Corollary 5.5. Next
we consider (5. 21 ) with N = N'N orl
2Re(A)v*Mv = v*N'Nv = I I Nvl l � (5.23)
We show that Nv is nonzero under (5. 1 6). Because of Av = AV, we have A
2
V = AAv = A
2
V,
. . . , AnI v = AnI v. Consider
IN:
=
l5
=
. t
If Nv = 0, the rightmost matrix is zero; the leftmost matrix, however, is nonzero under the
conditions of (5. 1 6) and v = O. This is a contadiction. Thus Nv is nonzero and (5. 23) implies
Re(A)< O. This completes the proof of Corollary 5. 5.
In the proof of Theorem of 5. 5, we have established the following result. For easy
reference, we state it as a theorem.
Theorem 5.6
If all eigenvalues of A have negative real parts, then the Lyapunov equation
A'M+ MA = N
has a unique solution for every N, and the solution can b expressed ð
M = ¡eA'I
NeA1
d
t
(5.24)
Because of the importance of this theorem, we give a diferent proof of the uniqueness
of the solution. Suppose there are two solutions MI and M2. Then we have
A'(MI  M2) + (MI  M2)A = 0
which implies
eA'I [A'(MI  M2) + (MI  M2)AjeAr = �[eA'r (MI  M2)eArj = 0
Its integration from 0 to 0 yields
[eA'r (MI  M2)eAr j l � = 0
or, using eAr ¬ 0 as t '" 0,
0  (M,  M2) = 0
I. Note that if x is a complex vector, then the Euclidean norm defned in Section 3.2 must be modifed 8 I l xl l� ¯ X

X.
where X

is the complex conjugate tanspose of x.
¹
5. 4 Lyapunov Theorem 1 35
This shows the uniqueness of M. Although the solution can be expressed as in (5. 24), the
integration is not used in computing the solution. It is simpler to arrange the Lyapunov equation,
after some transformations, into a standard linear algebraic equation as in (3. 60) and then solve
the equation. Note that even if A is not stable, a unique solution still exists if A has no two
eigenvalues such that Ai + Aj = O. The solution, however, cannot be expressed as in (5. 24);
the integration will diverge and is meaningless. If A is singular or, equivalently, has at least
one zero eigenvalue, then the Lyapunov equation is always singular and solutions may or may
not exist depending on whether or not N lies in the range space of the equation.
5. 4. 1 DiscreteTi me Case
Before discussing the discrete counterpart of Theorems 5. 5 and 5.6, we discuss the discrete
counterpart of the Lyapunov equation in (3. 59). Consider
M  AMB = C (5. 25)
where A and B are, respectively, n x n and m x m matrices. and M and C are H X m matrices.
As (3.60), Equation (5.25) can be expressed as Ym = c, where Y is an nm x nm matrix: m and
c are nm x I column vectors with the m columns of M and C stacked in order. Thus (5. 25) is
essentially a set of linear algebraic equations. Let 7k be an eigenvalue of Y or of (5. 25). Then
we have
7k = 1  Ai/j for i = I , 2, . . . , n; j = 1, 2, · . · . U
where Ai and Jj are, respectively, the eigenvalues of A and B. This can be established intuitively
as follows. Let us defne A(M) : = M  AMB. Then (5. 25) can be written as A(M) = C. A
scalar 7 is an eigenvalue of A if there exists a nonzero M such that A(M) = 7M. Let u be
an H X 1 right eigenvector of A associated with Ai : that is, Au = AiU. Let v be a 1 x III left
eigenvector of B associated with Jj ; that is, vB = vJj . Applying A to the n X U nonzero
matrix uv yields
A(uv) = uv  AuvB = (1  A, Jj)UV
Thus the eigenvalues of (5. 25) are I  Ai Jj ' for all i and j. If there are no i and j such that
Ai Jj = 1 , then (5.25) is nonsingular and, for any C. a unique solution M exists in (5. 25). If
Ai Jj = I for some i and j, then ( 5. 25) is singular and, for a given C, solutions may or may
not exist. The situation here is similar to what was discussed in Section 3. 7.
' Theorem 5.05
All eigenvalues of an n x n matrix A have magnitudes less than I i f and only if for any given positive
defnite symmetric matrix N or for N = N'N, where N is any given m x n matrix with U * II and
with the property in (5. 1 6), the discrete Lyapunov equation
M  A'MA = N (5.26)
has a unique symmetric solution 11 and M is positive defnite.
1 36 STABI LITY
We sketch briefy its proof for x ¯ O. If al l eigenvalues of x and, consequently, of x
have magnitudes less than 1, then we have · ·., < I for all i and j. Thus · ·
·
 and
··· is nonsingular. Therefore, for any x. a unique solution exists in · ·· We claim that
the solution can be expressed as
u ÷ ¯x· xx · ··
Because l A, I < for all i , this i nfnite series conwrges and i s well defned. Substituting · ··
into · ·· yields
�
= x + �x · xx �x · xx ÷ x
111 =1 ttt =
Thus · ·· i s the solution. l f x i s symmetric, so i s u I f � i s positive defnite, so i s u This
establishes the necessity. To show suffciency. let A be an eigenvalue of x and ·  0 be a
corresponding eigenvector: that i s, x· = ·· Then we have
··x· ÷ ··�t· ·· xxix·
= ··u· · ···xn· = ( l  · · · ··�t·
Because both ···· and ··u· are real and positiv'e, we conclude ( l  I AI
"
) ¯ or I AI
"
<
This establishes the theorem for x ¯ O. The case · � can si mi l arly be established.
Theorem 5.06
If al l eigenvalues of x have magnitudes less than I . then the discrete Lyapunov equation
u x¬x ÷ x
has ü unique solution for every N. and the solution can be expressed as
u = �x ·xx
It is important to mention that even if x has one or more eigenvalues with magnitudes
larger than a unique solution sti l l exists i n the discrete Lyapunov equation if )., ·
·
, I for all
i and j. In this case, the solution cannot be expressed as in · ·· but can be computed from
a set of linear algebraic equations.
Let us discuss the relationships between the continuoustime and discretetime Lyapunov
equations. The stability condition for continuoustime systems is that all eigenvalues l i e
inside the open lefthalf spl ane. The stability condition for discretetime systems is that all
eigenvalues lie inside the uni t circle on the :plane. These conditions can be related by the
bilinear transformation
�  l
5 = 
z + 1
+ s
1  5
· ··
5. 5 Stabi l i ty of LTV Systems 1 37
which maps the lefthalf splane into the interior of the uni t circle on the zplane and v'ice
versa. To differentiate the continuoustime and discretetime cases, we write
x u · xtx = N ( 5 . 29)
and
· ·
Fol l owi ng · ·· these two equations can be related by
Substituting the righthandside equation into · · and performing a simple manipulation.
we fnd
Comparng this with · · yields
x = ( .." + i x  I)
These relate · ·. and · ·
·
The MATLAB function lyap computes the Lyapunov equation in · : and dlyap
computes the discrete Lyapunov equation i n · · The function dlyap transforms ·
into · · by usi ng · · and then calls lyap. The result yields u = u
5. 5 Stabi l ity of LN Systems
Consider a SISO l i near timevarying ( LTV) system described by
yet ) = r)u ( r ) dr
'0
·.
The system i s said t o be BIBO stable if every bounded input excites a bounded output. The
condition for · ·· to be BIBO stable is that there exists a fnite constant M such that
! g( t . r ) 1 dr : M < ¯
Î¡
· ··
for al l t and t o wi t h t � t o. The proof i n t he timeinvariant case applies here v'ith onl y minor
modification.
For the multi variable case, · ·· becomes
ye t ) = Gu. r) u( r ) dr
to
·
The condition for · ·. to be BI BO stable i s that every entry of G(t, r ) meets the condition
i n · ·· For mul tivariable systems. we can also express the condition in terms of norms. Any
norm discussed in Section · can be used. However. the i nfni tenorm
I l ul l o = max l Ul l I I Gl l x ÷ l argest row absolute sum
,
1 38 STABI LITY
is probably the most convenient to use in stability study. For convenience, no subscript will
be attached to any norm. The necessary and suffcient condition for to be BIBO stable
is that there exists a fnite constant M such that
I I G(t, r) 1 1 dr :: M < 0
t0
for all t and to with t :: to.
is
The impulse response matrix of
x = A(t)x + B(t)u
y = C(t)x + D(t)u
G(t, r) = C( t)4(t, r)B(r) + D(t)8(t r)
and the zerostate response is
yet) = C(t ) 4( t , r)B(r) u(r) dr + D(t)u(t)
t0
Thus or, more precisely, the zerostate response of is BIBO stable if and only if
there exist constants MI and M2 such that
I I D(t) 1 1 :: MI < 0
and
I I G(t, r ) l l dr :: M2 < 0
t0
for all t and to with t :: to.
Next we study the stability of the zeroinput response of As in the timeinvariant
case, we defne the zeroinput response of or the equation x = A(t)x to be marginally
stable if every finite initial state excites a bounded response. Because the response is govered
by
xU) = 4U, to)x(to)
we conclude that the response i s marginally stable if and only if there exists a fnite constant
M such that
1 1 4(t, to) 1 1 ::
M < 0
for all to and for all t :: to. The equation x = AU)x is asymptotically stable if the response
excited by every fnite initial state is bounded and approaches zero as t  0. The asymptotic
stability conditions are the boundedness condition in and
1 1 4( t, to) 1 1 ¯
as t ¬ ¯
A great deal can be said regarding these defnitions and conditions. Does the constant M in
5. 5 Stabi l i ty of LT Systems 1 39
depend on to? Wat is the rate for the state transition matrix to approach in
The interested reader is referred to References
A timeinvariant equation x = Ax is asymptotically stable if all eigenvalues of A have
negative real parts. Is this also true for the timevarying case? The answer is negative as the
next example shows.
EXAMPLE 5.5 Consider the linear timevarying equation
x x = A(t)x = � _
I
The characteristic polynomial of A(t) is
det(AI  A(t» = det
Thus A(t) has eigenvalues and  1 for all t. It can be verifed directly that
4(t , O) =
meets and is therefore the state transition matrix of See also Problem
Because the ( l ,2)th entry of 4 grows without bound, the equation is neither asymptotically
stable nor marginally stable. This example shows that even though the eigenvalues can be
defned for AU) at every t, the concept of eigenvalues is not useful in the timevarying case.
All stability properties in the timeinvariant case are invariant under any equivalence
transformation. In the timevarying case, this is so only for BIBO stabiliry, because the
impulse response matrix is preserved. An equivalence transformation can transform, as shown
in Theorem any x = A(t)x into x = Aox, where Ao is any constant matrix; therefore, in the
timevarying case, marginal and asymptotic stabilities are not invariant under any equivalence
transformation.
. Teoem 5.7
Marginal and asymptotic stabilities of x = A(t)x are invariant under any Lyapunov transfonnation.
As discussed in Section if pet) and
P
(t) are continuous, and pet) is nonsingular
for all t, then x = PU)x is an algebraic transformation. If, in addition, P(t) and p
I
(t) are
bounded for all t , then x = P(t)x i s a Lyapunov transformation. The fundamental matrix X(t)
of x = A(t)x and the fundamental matrix
X
(t) of x = .(t)x are related by, as derived in
which implies
X
(t) = P(t)X(t)
4
(t, r) =
X
(t)
X
I
(r) = P(t)X(t)X
I
( rJP
1
(r)
= P(t)4(t, r)p
I
(r)
140 STABI LITY
PROBLEMS
Because bOlh P(t) and pl (t ) are bounded. if 1 I <( t . r ) 1 1 is bounded. so i s 1 1 4(1. r ) l l ; if
1 I <(t . r) \ 1 ¯ as t ¬ Ò. so is 1 1 4( t . r ) l l . This establishes Theorem 5. 7.
In the timeinvariant case. asymptotic stability of zeroinput responses always implies
BIBO stability of zerostate responses. This i s not necessarily so in the timevarying case. A
timevarying equation is asymptotically stable i f
1 I <(t. to) ! !  as t ¬ Ó (5. 41 )
for al l t . to with t :: to. I t i s BIBO stable if
I I C(t ) <( t , r) B(r ) 1 1 dr < O
l0
(5.42)
for al l t . to with ' :: Îµ. A function that approaches as t  Ó. may not be absolutely
integrable. Thus asymptotic stability may ,ot imply BlBO stability in the timevarying case.
However. if 1 I <( t . r) 1 1 decreases t o zero rapidly. in panicular. exponentially. and i f C( t ) and
B( t ) are bounded for all t. then asymptotic stability does imply BlBO stability. See References
[4. 6. 1 5] .
5. 1 Is t he network shown i n Fi g. 5. 2 BlBO stable� If not. fnd a bounded i nput that wi l l
excite an unbounded output.
Figure 5.2
u
5.2 Consider a system with an irrational transfer function g(5) . Show that a necessary
condition for the system to be BlBO stable is that I g( s) 1 i s fnite for all Re s ::
5.3 Is a system with implllse response g( r ) = 1 / 1 1 + I ) BlBO stable� How about g( r ) = teI
for 1 :: o�
5.4 Is a system with transfer function g( s) = e
2
J / (s + I ) BIBO stable?
5.5 Show that the negativefeedback system shown i n Fig. 2. 5(b) is BlBO stable i f and only
i f the gain O has a magnitude less than I . For U = I . fnd a bounded input r(r) that wi l l
excite an unbounded output.
5.6 Consider a system with transfer function g( s) = (s 2) /( s+ I ) . What are the steadystate
responses excited by li l t ) = 3, for 1 :: and by lI !t) = sin 21, for t :: o?
5.7 Consider
x =

�
1
� x +
�
2
11
y = [2 3] x  211
Problems 1 41
Is i t BI BO stable"
Consider a discretetime system with impulse response sequence
g[k] = k(O. S/ for k :: °
Is the system BlBO stable'}
5.9 Is the state equation in Problem 5. 7 marginally stable" Asymptotically stable�
5. 10 Is the homogeneous state equation
x x = �
marginally stable'} Asymptotically stable"
5. 11 Is t he homogeneous state equation
°
x x = �
marginally stable'} Asymptotically stable"
5. 12 Is the discretetime homogeneous state equation
°
x[k] x[k + 1 J = � I
marginally stable" Asymptotically stable'}
5.13 Is the discretetime homogeneous state equation
x[k + I J = � �
marginally stable' ) Asymptotically stable"
5.14 Use Theorem 5.5 to show that all eigenvalues of
A =
I
0. 5  I
have negative real pans.
5. 15 Use Theorem 5. 05 to show that all eigenvalues of the A in Problem 5. 1 4 have magnitudes
less than I .

5. 16 For any distinct negative real A, and any nonzero real O, . show that the matrix
1 42 STABI LITY
m ÷ [ .
� �� �':�; l
A2 + Al 2A2 A2 + A3
a3al a3aZ
aj

)3 + Al

)"3 + ).
2
2).3
is positive defnite. [Hint: Use Corollary 5.5 and A=diag(AI , A
2
, A3)' ]
5. 17 A real matrix M (not necessarily symmetric) is defned to be positive defnite ifx'Mx 7 °
for any nonzero x. Is it true that the matrix M is positive defnite if all eigenvalues of
M are real and positive or if all its leading principal minors are positive? If not, how do
you check its positive defnitene�s? [Hint: Try
�2
r
�
'
1�9 � '
5.18 Show that all eigenvalues of A have real parts less than J < ° if and only if, for any
given positive defnite symmetric matrix N. the equation
A'M+ MA + 2JM = N
has a unique symmetric solution M and M is positive defnite.
5.19 Show that all eigenvalues of A have magnitudes less than p if and only if, for any given
positive defnite symmetric matrix N, the equation
.²m
.
A'MA ÷p
2
N
has a unique symmetric solution M and M is positive defnite.
5.20 Is a system with impulse response get , T) ÷e
2
Itl

l rl
, for t :: T, BIBO stable? How
about get. T) ÷sin t (e
C
t T
)
) cos T?
5.21 Consider the timevarying equation
. = 2tx + H
Is the equation BIBO stable? Marginally stable? Asymptotically stable?
5.22 Show that the equation in Problem 5. 21 can be transformed by using x = P(t) x, with
P(t) ÷et
'
, into
» ÷ .
Is the equation BIBO stable? Marginally stable? Asymptotically stable? Is the transfor
mation a Lyapunov transformation?
5.23 Is the homogeneous equation
 1
x =
3t
¯
Þ
for to :: 0, marginally stable? Asymptotically stable?
6. 1 I ntroducti on
LH3D!CI
Control l abi l i ty
and Observabi l i ty
This chapter introduces the concepts of controllability and observability. Controllability
deals with whether or not the state of a statespace equation can be controlled from the input,
and observability deals with whether or not the initial state can be observed from the output.
These concepts can be illustrated using the network shown in Fig. 6. 1 . The network has two
state variables. Let Xi be the voltage across the capacitor with capacitance C; . for i ÷ I . 2.
The input l ! is a current source and the output y is the voltage shown. From the network. we
see that, because of the open circuit across \. the input has no efect on X2 or cannot control
X2 . The current passing through the 2Q resistor always equals the current source H. therefore
the response excited by the initial state XI will not appear in y. Thus the initial state XI cannot
be observed from the output. Thus the equation describing the network cannot be controllable
and observable.
These concepts are essential i n discussing the interal structure of linear systems. They
are also needed in studying control and fltering problems. We study first continuoustime
Figure 6.1 Network.
(curem
sen'lce)
1 >
143
1 4 CONTROLLABI LI TY AND OBSERVABI LI TY
linear timeinvariant (LTI) state equations and then discretetime LTI state equations. Finally,
we study the timevarying case.
6. 2 Control l abi l ity
Consider the ndimensional pinput state equation
x = Ax + Bu (6. 1 )
where A and B are, respectively, n and .real constant matrices. Because the output
does not play any role i n controllability. we will disregard the output equation in this study.
Defnition 6.1 ....(6. 1) .(A, B) ...be controllable ·
....x(O) = ......
... ......
..·...(6. 1) (A, B) ....be uncontrollable.
This defnition requires only that the input be capable of moving any state in the state
space to any other state in a fnite time: what trajectory the state should take is not specifed.
Furthermore, there i s no constraint imposed on the input: its magnitude can be as large as
desired. We give an example to illustrate the concept.
EXAMPLE 6.1 Consider the network shown in Fig. 6. 2(a). Its state variable ³ is the voltage
across the capacitor. If x(O) = 0, then = 0 for all :: 0 no matter what input is applied.
This i s due to the symmetry of the network, and the input has no effect on the voltage across
the capacitor. Thus the system or, more precisely, the state equation that describes the system
i s not controllable.
Next we consider the network shown i n Fig. 6.2(b). It has two state variables ¹ and
X2 as shown. The input can transfer X2 to any value; but it cannot transfer ·¡ .X2
to any values. For example, i f x, (0) = X2 (0) = 0, then no matter what input i s applied,
(I) always equals
for all :: O. Thus the equation that describes the network is not
controllable.
·
·
l í!
I í1
À
¬
.
Îf ¯
I í1 ! \1
(a) (b)
Figure 6.2 Uncontrollable networks.
6. 2 Control l abi l ity
Theorem 6. 1
The following .,tatemenb are equivalent.
1. The ndimensional pair (A, B) is controllable.
2. The Il X n matrix
'Ve r t ) =
eAt BB'eA'r dr = eA( t r ) BB'eA , , r ) dr
is nonsingular for any I > O.
3. The np controllability matrix
has rank n (full row rank).
�. The I (1/ + p) matrix [A  AI BJ has full row rank at every eigen\ alue. A. of A l
5. If. in addition. all eigenvalues of A have negative real parts. then the unique solution of
1 45
(6.2)
(6 3 )
AWe + «, A' = BB' (6. 4)
is positive definite. The solution i s called the colrollabilitv Gramiall and can be expressed as
«=
·
eAr BB' eA' r dr
(6. S)
Proof: ( I )  ( 2) : First we show the equivalence of the two forms in (6.2). Define
i := 
r
. Then we have
eAI tr )BB' eA' ( t rl dr
=
�eAiBB' eA" ( df)
· .
= ' eA'
BB' eA
"
d
i
.
It becomes the frst form of (6. 2) after replacing i by r
. Because of the foml of the
integrand, « is always positive semidefnite: it is positive definite if and on Iv if it is
nonsingular. See Section 3. 9.
'
First we show that if « is nonsingular, then (6. I ) is controllable. The response
0[ ( 6. 1 ) at time f, was derived in (4. 5) as
X(I, ) = C'x
( Q
)
+
eAl tl r ' Bu(r ) er
We claim that for any x(O) = .and any . = x" the input
u( ) = B' en'I I ) W;
I
(t, ) [eA'l xo  xd
wi l l transfer .to x, at time f l · Indeed. substituting ( 6. 7) into (6. 6) yields
( 6. 6)
( 6. 7)
Î . I f . I >compl�x. then we mu�[ use complex number� as ccalars In checking the rank. See the diSCUSSion regarding
( 3.37! .
146 CONTROLLABI LITY AND OBSERVABI LITY
,
 xd
This shows that if is nonsingular, then the pair .is controllable. We show the
converse by contradiction. Suppose the pair is controllable but is not positive
defnite for some
Then there exists an n X I nonzero vector such that
which implies
(6.8)
for all in If (6. 1 ) is controllable, there exists an input that transfers the initial state
x(o)
to 0 and (6.6) becomes
.
Its premultiplication by Y
¹
yields
which contradicts t This establishes the equivalence of ( I ) and (2).
(2) * (3): Because every entry of is, as discussed at the end of Section 4.2, an
analytical function of if is nonsingular for some f, then it is nonsingular for all
in (0, 0). See Reference (6, p. 554]. Because of the equivalence of the two forms in
(6.2), (6.8) implies that is nonsingular if and only if there exists no n x I nonzero
vector such that
for all (6.9)
Now we show that if is nonsingular. then the controllability matrix has full row
rank. Suppose does not have full row rank, then there exists an n X I nonzero vector
such that 0 or
for k = 1, 2, . . . , n I
Because can be expressed as a linear combination of . .(Theorem
3.5), we conclude Y
/
This contradicts the nonsingularity assumption of
Thus Condition (2) implies Condition (3). To show the converse, suppose has full row
rank but is singular. Then there exists a nonzero such that (6.9) holds. Setting
we have Differentiating (6. 9) and then setting = we have .= O.
Proceeding forard yields .0 for k = 1 . 2,
.
. . They can be aranged as
.
. = 0
'
6
.
2 Control l abi l ity 147
This contradicts the hypothesis that has full row rank. This shows the equivalence of
(2) and (3).
(3) + (4) : If has full row rank, then the matrix . AI has full row rank at
every eigenvalue of .If not, there exists an eigenValue
and a 1 x n vector t 0 such
that
which implies . and qB Thus q is a lef eigenvector of . We compute
... .
Proceeding forward, we have .A1q. Thus we have
. .
. . . A71qB] = 0
This contradicts the hypothesis that has full row rank.
I order to show that < n implies . AI B]) < n at some eigenvalue )' 1
of .we need Theorems 6.2 and 6.6, which will be established later. Theorem 6.2 states
that controllability is invaraint under any equivalence transformation. Therefore we may
show . AI .< n at some eigenvalue of .where ..is equivalent to .
Theorem 6.6 states that if the rank of is less than n or = n  m, for some integer
m o: l , then there exists a nonsingular matrix such that
..
.
�
o .
.
=
0
where .is m x m. Let AI be an eigenvalue of .and be a corresponding I x m
nonzero left eigenvector or . Then we have
. Now we form
the ! x n vector (0 We compute
6 1 )
which implies . AI B]) < n and, consequently, . AI B]) < n at some
eigenvalue of .This establishes the equivalence of (3) and (4).
(2) * (5) : If .is stable, then the unique solution of (6.4) can be expressed as in (6.5)
(Theorem 5.6). The Gramian is always positive semidefnite. It is positive defnite if
and only if is nonsingular. This establishes the equivalence of (2) and (5). Q.E.D.
EXAMLE 6.2 Consider the inverted pendulum studied in Example 2. 8. Its state equation was
developed in (2.27). Suppose for a given pendulum, the equation becomes
� �

(6. 1 1 )
y ( I ° ° o)x
1 48 CONTROLLABI LI TY AND OBSERVABI LITY
We compute
°
°
°
°
 l
�
This matrix can be shown to have rank thus the system i s controllable. Therefore. i f X3 = e
deviates from zero slightly, we can fnd a control to push it back to zero. In fact, a control
exists to brng X
I
= X3 , and their derivatives back to zero. This is consistent with our
experience of balancing a broom on our palm.
The MATLAB functions C crb and gram will generate the controllability matrix and
controllability Gramian. Note that the contrbllability Gramian is not computed from it is
obtained by solving a set of linear algebraic equations. Whether a state equation is controllable
can then be determined by computing the rank of the controllability matrix or Gramian by
using rank i n MATLAB.
EXAMPLE 6.3 Consider the platform system shown i n Fig. it can be used to study
suspension systems of automobiles. The system consists of one platform; both ends of the
platform are supported on the ground by means of springs and dashpots, which provide
viscous friction. The mass of the platform is assumed to be zero; thus the movements of
the two spring systems are independent and half of the force is applied to each spring
system. The spring constants of both springs are assumed to be and the viscous friction
coeffcients are assumed to be 2 and as shown. If the displacements of the two spring
systems from equilibrium are chosen as state variables XI and X2. then we have XI + =
and X2 + X2 = ll or
x = . I
. ,
°
This state equation describes the system.
Now if the initial displacements are different from zero, and if no force is applied. the
platform will retur to zero exponentially. In theory. it will take an infnite time for X, to equal
° exactly. Now we pose the problem. If XI = and X2 = can we apply a force
to bring the platform to equilibrium in seconds? The answer does not seem to be obvious
because the force i s applied to the two spring systems.
Dampmg
co�ffcienl
Spring
constant
l
�2u
Damping
coefficient
Spring
constant
1
Figure 6.3 Platfon system.
For Equation we compute
...=
6. 2 Control l abi l ity 149
. ¨
I
 
Thus the equation is controllable and, for any x(O), there exists an input that transfers x(O) to
o in 2 seconds or in any fnite time. We compute and for this system at =
and
Wc (2) =
2
°
°
°
·
°
�
¯
¸�µ '(2�l)
11 1 =
°
= +
for l in This input force will transfer x(O) = to i n seconds as shown
i n Fig. in which the input i s also plotted. It i s obtained by using the MATLAB function
1 c �:,., an acronym for The largest magnitude of the input is about
Figure plots the input U2 that transfers x(O) = to 0 i n seconds. We see
that the smaller the time interval, the larger the input magnitude. If no restriction i s imposed on
the input, we can transfer x(O) to zero in an arbitraril y small time interval; however, the input
magnitude may become very large. If some restriction i s imposed on the input magnitude, then
we cannot achieve the transfer as fast as desired. For example, i f we require 111 (1) 1 < for
all r. in Example then we cannot transfer x(O) to 0 in less than  seconds. We remark that
the input .in is called the in the sense that for any other input
i(1 ) that achieves the same transfer, we have
. . . .
'º 'u
ÚU � ! U �
¬Ü
20
�40 L=� � � =
Ú 05 1 5
(a)
o
5
U L.. �.�.��
o 4
Figure 6A Transfer x(O) ~
1 50 CONTROLLABI LITY AND OBSERVABI LITY
Its proof can be found in Reference [6, pp. 556558].
EXAMPLE 6. 4 Consider again the platform system shown in Fig. 6. 3. We now assume that the
viscous friction coefcients and spring constants of both spring systems all equal Then the
state equation that describes the system becomes
Clearly we have
1
 1
= 1
and the state equation is not controllable. If Xl (0) 1 no input can transfer to zero
in a fnite time.
I
6.2. 1 Controllabilit Indices
Let A and B be x and x constant matrices. We assume that B has rank or full column
rank. If B does not have full column rank, there is a redundancy in inputs. For example, if
the second column of B equals the frst column of B, then the efect of the second input
on the system can be generated from the frst input. Thus the second input is redundant. In
conclusion, deleting linearly dependent columns of B and the corresponding inputs will not
afect the contol of the system. Thus it is reasonable to assume that B has full column rank.
If (A, B) is controllable, its controllability matix C has rank and, consequently,
linearly independent columns. Note that there are columns in C; therefore it is possible to
fnd many sets of n linearly independent columns in C. We discuss in the following the most
important way of searching these columns; the search also happens to be most natural. Let bi
be the ith column of B. Then C can be written explicitly as
,"
(6. 13)
Let us search linearly independent columns of C fom left to right. Because of the patter of
C. if Ai bm depends on its lefthandside (LHS) columns, then Ai+
l
bm will also depend on its
LHS columns. It means that once a column associated with bm becomes linearly dependent,
then all columns associated with bm thereafter are linearly dependent. Let Jm be the number
of the linearly independent columns associated with bm in C. That is, the columns
ae linearly independent in C and Al'm+ibm are linearly dependent for = 0, « . . . . It is clea
that if C has r then
Jl + J2 +
. .
+ Jp = (6. 14)
The set {Jl , J2 , • • º , Jp} is called the and
J = max (Jl , Jz , ª ª • , Jp)
`
6. 2 Control l abi l ity 1 51
i s called the of (A, B). Or, equivalently, if (A, B) i s controllable. the
controllability index J is the least integer such that
(6. 1 5)
Now we give a range of J. If Jl = J2 =
.
. .
= Jp, then � J. If all Jm. except
one, equal ! , then J =

1 ) ; this is the largest possible J. Let n be the degree of the
minimal polynomial of A. Then, by defnition, there exist ai such that
An = al
An1 + azAn
2
+
. . + a
n
I
which implies that AnB can be written as a linear combination of {B, AB, . . . , An

l B} . Thus
we conclude
� J � min(n, + (6. 1 6)
where p(B) = Because of (6. 1 6), in checking controllability, it is unnecessary to check
the x matrix C. It is sufcient to check a matrix of lesser columns. Because the degree
of the minimal polynomial is generally not availablewhereas the rank of B can readil�be
computedwe can use the following corollary to check controllability. The second part oithe
corollary follows Theorem 3. 8.
Corollary 6. 1
The ndimensional pair (A. B) is controllable if and only if the matrix
Cn

p
+
l := [B AB ' " A
n

PB]
where p(B) = has rank n or the x matrix Cnp+1 C�_p+l is nonsingular.
(6. 1 7)
EXAMPLE 6.5 Consider the satellite system studied i n Fig. 2. 1 3. Its linearized state equation
was developed in (2.29). From the equation, we can see that the control of the frst four state
variables by the frst two inputs and the control of the last two state variables by the last input
are decoupled; therefore we can consider only the following subequation of (2. 29):
o 0 1 X
+
0 0
x =
�
�
�
� � � �
2 0 0 0 1
(6. 1 8)
1 0 0 0
'
=
0 0
where we have assumed. for simplicity, (o = m = ro = The controllability matrix of (6. 1 8)
is of order 4 x 8. If we use Corollary 6. 1 , then we can' check its controllability by using the
following 4 x 6 matrix:
[B AB A
'
Bj �
0 1 0 0
0 0 2  1
0 0 1 2
(6. 1 9)
2 0 0
1 52 CONTROLLABI LITY A,'I D OBSERVABI LI TY
It has rank 4. Thus (6. 1 8) i s controllable. From (6. 1 9) , we can readily verify that the control
lability indices are 2 and 2, and the controllability index is 2.
Theorem 6,2
The controllability property is i nvariant under any equivalence transfonation.
Proof' Consider the pair (A, B) with controllability matrix
and its equivalent pair (A, E) with A = PAP I and E = PB, where P i s a nonsingular
matri x. The coatrollability matrix of ( . .. E) is
C = [E AE .  A
n

I
E]
= [PB PAP I PB . . .
PA
nl pl
pB]
= [PB PAB . . . PAn I B]
= P[B AB . . An I B] = PC ( 6.20)
Because P i s nonsingular, we have PI C) = P I C) (see Equation ( 3. 62» . This establishes
Theorem 6. 2. Q. E. D.
Theorem 6, 3
The set of the controllability indices of (A, Bl is invariant under any equivalence transfomlation and
any reordering of the columns of B.
. Proof' Let us defne
( 6. 2 1 )
Then we have, following the proof of Theorem 6. 2.
for k = 0, I , 2, . . . . Thus the set of controllability i ndi ces i s i nvariant under any
equivalence transformation.
The rearrangement of the columns of B can be achieved by
B = BM
where ,1 is a p x p nonsingular permutation matrix. It is straightforward to veriiy
Because diag (M, M . . . . . M) is nonsingular. we have PI G) = p« (�) for k ^ O. l . . . . .
Thus the set of controllability indices is i nvariant under any reordering of the columns of
B. Q. E. D.
Because the set of the controllabi l i ty indices is invariant under any equivalence transfor
mation and any rearrangement of the inputs. it is an i ntrinsic property of the system that the
6. 3 Observabi l i ty 1 53
state equation describes. The physical signi fcance of the controllability index is not transparent
here: but it becomes obvious in the discretetime case. As we wi l l discuss in later chapters. the
controllability i ndex can also be computed from transfer matrices and dictates the minimum
degree required to achieve pole placement and model matching.
6. 3 Observabi l it
The concept of observability is dual to that of control labi l i ty. Roughly speaking. controllability
studies the possibility of steering the state from the input: observability studies the possibility
of estimating the state from the output. These two concepts are defined under the assumption
that the state equation or. equivalently. all A. B. C. and D are known. Thus the problem of
observabi lity i s different from the problem of realization or identifcation, which i s to determine
or estimate A. B. C. and D from the i nformation collected at the input and output terminals.
Consider the ndimensional pinput qoutput state equation
x ¬ Ax + Bu
( 6 22)
y = Cx + Du
where A, B. C. and D are. respectively, I/ X n, n X p, q X n, and q x p constant matri ces.
Defniton 6.01 The swte equation (6. 22) is said to be observable iffor all)' unknOlt"n
initial state x( O) , there exists a fnite tl > ° such that the knowledge ofthe input U and
the output y over [0, tl ] sufces to determine uniqueiv the initial state x(O) . Othenl"ise,
the equation is said to be unobservable.
EXAMPLE 6,6 Consider the network shown i n Fig. 6. 5. If the input is zero. no matter what the
i nitial voltage across the capacitor is, the output i s identically zero because of the symmetry
of the four resistors. We know the input and output (both are identically zero). but we cannot
determine uniquely the initial state. Thus the network or. more precisely. the state equation
that describes the network is not observable.
EXAMPLE 6.7 Consider the network shown i n Fig. 6. 6(a). The network has two state variables:
the current ·I through the inductor and the voltage '· across the capacitor. The input U is a
Figure 6,5 Unobservable network.
'
:
lÍ ' ¨
¬
í4
'.
I II I Q
1 5 CONTROLLABI LITY AND OBSERVABI LITY
u
Figr 6.6
l H l H
+
J
¡ t
�
(a) (b)
Unobservable network.

curent source. Iu = 0, the network reduces to the one shown in Fig. 6.6(b). If Xl (0) ~ a t 0
andx2 = 0, then the output is identically zero. Any x(O) = [a 0]' and u (t) = 0 yield the same
output yet) = O. Thus there is no way to determine the initial state [a 0]' uniquely and the
equation that descrbs the network is not observable.
Te response of (6.22) excited by the initial state x(O) and the input u(t) was derived in
(4.7) as
yet) = CeAlx(O) + C eAU<)Bu(r) dr + Du(t) (6.23)
In the study of observability, the output y and the input u are assumed to be known; the inital
state x(O) is the only unknown. Thus we can write (6.23) as
(6.24)
where
yet) : = yet)  C eA(H)Bu(r) dr  Du(t)
is a known function. Thus the observability problem reduces to solving x(O) from (6.24). If
u = 0, then y(t) reduces to the zeroinput response CeAlx(O). Thus Defnition 6. 01 can be
modifed as follows: Equation (6.22) is observable if and only if the initial state x(O) can be
detemtined uniquely from its zeroinput response over a fnite time interval.
Next we discuss how to solve x(O) from (6.24). For a fxed I, CeAc is a q x n constant
matrix, and y(l) is a q x 1 constant vector. Thus (6.24) is a set of linear algebraic equations
with n unknowns. Because of the way it is developed, for every fxed I, yet) is in the range
space of CeA
c
and solutions always exist in (6.24). The only question is whether the solution
is unique. I q < n, as is the case in general, the q x n matrix CeAI has rank at most q and,
consequently, has nullity n  q or larger. Thus solutions are not unique (Theorem 3. 2). In
conclusion, we cannot fnd a unique x(O) from (6.24) at an isolated I. In order to determine
x(O) uniquely from (6.24), we must use the knowledge of u(t) and y(l) over a nonzero time
interval as stated in the next theorem.
'
� Teorem 6.4
The state equation (6.22) is observable if and only if the n x n matrix
Wo(l) = eA'< C'CeMdr
is nonsingular for any I 7 O.
6. 3 Obserabi l i ty
•
Proof: We pre multiply (6.24) by eA'cC' and then integrate it over [0, 11 ] to yield
eA'C C'CeAldl, X(O) =
c,
eA'I C'y(t) dl
If Wo(ltJ is nonsingular, then
x(O) ~ W;l (tl) eA
'
I C'y(t) dt
1 55
(6.25)
(6.26)
This yields a unique x(O). This shows that if Wo(t), for any t 7 0, is nonsingular, then
(6.22) is observable. Next we show that if Wo(tt l is singular or, equivalently, psitive
semidefnite for all then (6.22) is not observable. If Wo(tl) is positive semidefnite,
there exists an n x I nonzero constant vector Y such that
which implies
for all I in [0, I t l . If u = 0, then Xl (0) = Y t 0 and X2 (0) = 0 both yield the same
yet) = CeAlx; (O) = 0
(6.27)
Two diferent initial states yield the same zeroinput response; therefore we cannot
uniquely determine x(O). Thus (6.22) is not observable. This completes the proof of
Theorem 6.4. Q.E.D.
We see from this theorem that observability depends only on A and C. This can also be
deduced from Defnition 6. 01 by choosing U(I) = O. Thus observability is a property of the
pair (A, C) and is independent ofB andD. As in the controllability part, ifWo(t) is nonsingular
for some I, then it is nonsingular for every I and the initial state can be computed from (6.26)
by using any nonzero time interval.
� Teorem 6.5 (eorem of dualit)
The pair ( A, B) is controllable if and only if the pair (A', B') is observable.
1 56 CONTROLLABI LI TY AI'D OBSERVABI LITY
Proof: The pair (A, B) i s controllable i f and only i f
We rt) =
eAr BB' eA' r dr
is nonsingular for any I , The pair (A' , B') i s observable i f and only if, by replacing A by
A' and C by B' i n (6. 25).
Wo U) = eAr BB'eA' r dr
is nonsingular for any t . The two conditions are identical and the theorem follows.
Q. E. D,
We list i n the following the observablity counterpart of Theorem 6. I. It can be proved
either directly or by applying the theorem of duality.
Theorem 6.01
The following statements are equivalent.
1. The ndimensional pair (A, C) is observable.
2. The n x n matrx
1 6. 28)
is nonsingular for any t > O.
3. The nq x 11 observabilitmatrix
(6.29)
has rank R (full column rank). This matrix can be generated by calling obsv i n MATL\B,
4. The (1 + q) x n matrix
has full column rank at every eigenvalue. A, of A.
5. If. i n addition. all ei genvalues of A have negative real parts. then the unique solution of
(6. 30)
is positive definite. The solution is called the obsen'abilir Gramian and can be expressed as
Wo = eA' r C
'CeAr dr
(6. 3 1 )
6. 3 Observabi l i ty 1 57
6, 3. 1 ObseNobilit Indices
Let A and C be n x n and q x 1 constant matrices. We assume that C has rank q (full row rank).
If C does not have full row rank, then the output at some output terminal can be expressed
as a l i near combination of other outputs. Thus the output does not offer any new information
regarding the system and the terminal can be eliminated. By deleting the corresponding row.
the reduced C will then have full row rank.
If (A, C) i s observable, its observability matrix 0 has rank 1 and. consequently, n linearly
independent rows. Let Ci be the ith row of C
. Let us search linearly independent rows of 0 in
order from top to bottom. Dual to the controllability part, if a row associated with Cm becomes
linearly dependent on its upper rows, then all rows associated with Cm thereafter will also be
dependent. Let \n be the number of the linearly independent rows associated with Cm , It is
clear that i f 0 has rank 11 . then
1} 1 + 12 + . .
. + ¹q = n (6. 32)
The set {VI , V2, . . . , 1}q } i s called the observabilit indices and
(6. 33)
is called the observabilit' index of (A, C). If ( A, C) is observable, it is the least integer
such that
Dual to the controllability part. we have
n/q � V � mi n(i, n  q 1)
where PIC) ¯ q and i is the degree of the minimal polynomial of A.
Corollary 6.01
The ndimensional pair (A. C) is observable if and onl y if the matrix
Onq+1 = �
c�
n

q
where PI C) = q, has rank 11 or the n x n matrix O�_
q
+1 O
n
q+1 is nonsingular.
: Theorem 6.02
The observability property is invariant under any equivalence transformation.
(6, 34)
( 6. 35)
1 58 CONTROLLABI LITY AND OBSERVABI LITY
� Theorem 6.03
The set of the observability indices of (A, C) is invariant under any equivalence transformation and any
reordering of the rows of C.
Before concluding this section, we discuss a diferent way of solving (6.24). Diferenti
ating (6.24) repeatedly and setting t = 0, we can obtain
or
Ovx(O) = yeO) (6.36)
where y
(i
) (I) is the ith derivative of y(t), and yeO) := W (0) v(0) (y(
v
I) ']'. Equation
(6.36) is a set of linear algebraic equations. Because of the way it is developed, y(0) must lie
in the range space of O
v
. Thus a solution x(O) exists in (6.36). If (A, C) is observable, then
Ov has full column r and, following Theorem 3.2, the solution is unique. Premultiplying
(6.36) by 0 and then using Theorem 3. 8, we can obtain the solution as
(6.37)
We mention that in order to obtain yeO), veO), " . , we need knowledge of yet) in the neigh
borhood of t = O. This is consistent with the earlier assertion that we need knowledge of Y(t)
over a nonzero time interval i n order to determine x(O) uniquely from (6.24). In conclusion,
the initial state can b computed. using (6. 26) or (6.37).
The output y(t) measured in practice is often corrupted by highfrequency noise. Because
• diferentiation will amplif highfrequency noise and
• integration will suppress or smooth highfrequency noise,
the result obtained from (6.36) or (6.37) may difer greatly from the actual initial state. Thus
(6.26) is preferable to (6.36) in computing initial states.
The physical signifcance of the observability index can be seen from (6.36). It is the
smallest integer in order to determine x(O) uniquely from (6.36) or (6.37). It also dictates the
minimum degree required to achieve pole placement and model matching, as we will discuss
in Chapter 9.
6.4 Canonical Decomposition
This section discusses canonical decomposition of state equations. This fundamental result will
be used to establish the relationship between the statespace description and the transfermatrix
description. Consider
6.4 Canoni cal Decomposi ti on
x = Ax + Bu
y = Cx + Du
Let x ~ P, where P is a nonsingular matrix. Then the state equation
x = Ax + Bu
y = Cx + Du
1 59
(6. 38)
(6.39)
with A = PAPI , B = PB. C = Cp
I
, and D = D is equivalent to (6. 38). All properties of
(6.38), including stability. controllability, and observability, are preserved in (6. 39). We also
have
;. Theorem 6.6
Consider the ndi mensional state equation i n (6. 38) with
pe e) = p([B AB
. . .
A
n
I B)) ~ n l < n
We form the n X n matix
where the first n  columns are any n I linearly independent columns of C, and the remaining columns
can arbitrarily be chosen as long as P is nonsingular. Then the equivalence transformation x ~ P or
X = p
I
X will transform (6. 38) into
(6'0)
where Ae is n  x n I and Ac i s (n ¯ n I ) X (n ¯ n do and the n I dimensional subequation of (6'0).
Xc = Aexc + Beu
(6.'1 )
i s controllable and has the same transfer matrix as (6. 38).
Proof As discussed in Section 4.3, the transforation x = p
I
X changes the basis
of the state space from the orthonormal basis in (3. 8) to the columns of Q := pl or
{q
l , . . . , qn, . . . . . qn } . The ith column of .is the representation of Aqi with respect
to {ql . " ' , qn , . " " qn } · Now the vector Aqi , for i = I , 2, . . . , n l . are linear!y
dependent on the set {q I . . . . , qn, } ; they are linearly independent of { qn , .I , . . . « qn }.
Thus the matrix A has the form shown in (6.40). The columns of B are the representation
ofthe co!umns ofB with respect to {ql • » + + , qn" . . , Q} . Now the columns ofB depend
only on {ql « • • • • qn , } : thus B has the form shown in (6.40). We mention that if the n x p
1 60 CONTROLLABI LI TY AND OBSERVBI LI TY
matrix B has rank p and if its columns are chosen as the frst p columns of pI . then the
upper part of B is the unit matrix of order p.
Let be the controllability matrix of (6. 40) . Then we have p( C) = p( = n
I
. It
is straightforward to verify
C =
�
e
AeBe
~ n¡ 
A��
I
Be
[
A
e
Be
0 0
=
�
¯ n! ¯
A;�I Be

A, Be
0
where is the controllability matrix of (Ac. Be). Because the columns of A;Bc. for
k � n
I
. are l i nearly dependent on the columns of • the condition p( C) = 11 1 implies
p( c) = n
I
. Thus the n l dimensional state equation i n (6. 4 I ) is control l able.
Next we show that (6. 41 ) has th
b
same transfer matrix as (6. 38) . Because ( 6. 38) and
(6.40) have the same transfer matrx. we need to show only that (6. 40) and ( 6.4 1 ) have
the same transfer matrix. By direct verifcation. we can show
where
.. 
0
Ae A
I'
[
 I
=
..  A, )  I M
[ sI ¬ Ai 0 ..  Ac) I
M = .. ¯ Acl I AdsI  Ac) I
Thus the transfer matrix of (6.40) i s
[ c
C ]
sI  Ae A
I
'

Be
[
D
" 0 ..  Ai 0
¯
= [Ce Cc]
..

o
Ae) 
I 1

Be
(sI  AcTI 0
¯
'
(6.42)
which i s the transfer matrix of (6. 41 ). This completes the proof of Theorem 6. 6. Q. E. D.
In the equivalence transformation x = Px. the ndi mensional state space is divided into
two subspaces. One is the n I dimensional subspace that consists of all vectors of the form
[x;. 0'] ' : the other is the (n  n
I
)dimensional subspace that consists of al l vectors of the form
[0' x:,j' . Because (6.41 ) is controllable. the input u can transfer Xc from any state to any other
state. However. the input U cannot control Xc because. as we can see from ( 6.401 . U does not
affect Xc directly. nor indirectly through the state xc . By dropping the uncontrollable state
vector. we obtain a controllable state equation of lesser dimension that is zerostate equi\'alent
to the original equation.
EXAMPLE 6.8 Consider the threedi mensional state equation
y = [ I I I ] x (6. 43)
6. 4 Canoni cal Decomposi ti on 1 61
The rank of B i s 2: therefore we can use C" = [ B AB] . instead of C = [ B AB A"B]. t o check
the controllability of (6. 43) ( Corol l ary 6. 1 ) . Because
P ( C2 J = P( [B ABl J = P
� � = 2 < 3
the state equation in (6. 43) is not control lable. Let us choose
pI = Q : =
�
The frst two columns of Q are the frst two l i nearly independent columns of C2 : the last column
is chosen arbitraril y to make Q nonsingular. Let .= Px. We compute
A = P.WI =
�
�
0
0
�
I
B = PB = 0
0
I
I
0 0
0
�
;.ote that the I x 2 submatrix A21 of A and Be are zero as expected. The 2 x I submatrix
'1 2 happens to be zero: it could be nonzero. The upper part of B i s a unit matrix because the
columns of B are the frst two columns of Q. Thus (6.43) can be reduced to
Xc = �
Xc
¯
� �

u
This equation is controllable and has the same transfer matrix as (6. 43).
The MATLAB function C trbf transforms (6. 38 ) into (6.0) except that the order of the
columns in pI is reversed. Thus the resulting equation has the form
Theorem 6. 6 is established from the controllability matrix. In actual computation. it is unnec
essary to form the controllability matrix. The result can be obtained by carrying out a sequence
1 62 CONTROLLABI LI TY AND OBSERVABI LITY
of similarity transformations to transform [B A] into a Hessenberg form. See Reference [6,
pp. 220222] . This procedure is effcient and numerically stable and should be used in actual
computation.
Dual to Theorem 6. 6, we have the following theorem for unobservable state equations.
> Theorem 6.06
Consider the ndimensional state equation in (6.38) with
We for the n X n matrix
PI
p = Pn2
pn
where the frst n2 rows are any nz linearly independent rows of 0, and the remaining rows can be chosen
arbitrarily 8 long as P is nonsingular. Then the equivalence transformation x = P will transform
(6.38) into
 xo
y = [Co 0] �o + Du (6.4)
where Ao is n2 X n2 and Ai is (n  nz) X (n ¯ 11 2 ) , and the lizdimensional subequation of (6.44).
Xo = Acxo + Bou
y = CoXo + Du
is observable and has the same transfer matrix as (6.38).
In the equivalence transformation x = P, the ndimensional state space is divided
into two subspaces. One is the lzdimensional subspace that consists of all vectors of the
form [x� 0']'; the other is the (1  n2)dimensi
o
nal subspace consisting of all vectors of the
form [0' x�]'. The state Xo can be detected from the output. However, Xi cannot be detected
from the output because, as we can see from (6. 4), it is not connected to the output either
directly, or indirectly through the state xn' By dropping the unobservable state vector. we obtain
an observable state equation of lesser dimension that is zerostate equivalent to the original
equation. The MATLAB function obsvf is the counterpart of ctrbf. Combining Theorems
6.6 and 6.06, we have the following Kalmal decomposition theorem.
6. 4 Canoni cal Decomposi tion 1 63
Theorem 6.7
Every statespace equation can be transformed. by an equivalence transformation. into the following
canonical form
(645)
Y = [ C
eo 0 C
eo O]x + Du
where the vector xeo is controllable and observable. Xci is controllable but not observable. ;'0 is
observable but not controllable. and Xci is neither controllable nor observable. Furthermore. the state
equation is zerostate equivalent to the controllable and observable state equation
XeD = Aeoxeo + Beou
(646)
and has the transfer matrix
 I 
G(s) = Ceo(sI  Aeo) Beo + D
This theorem can be illustrated symbolically as shown in Fig. 6. 7. The equation is frst
decomposed, using Theorem 6. 6, into controllable and uncontrollable subequations. We then
decompose each subequation. using Theorem 6.06, into observable and unobservable parts.
From the fgure. we see that only the controllable and observable part is connected to both
the input and output tenninals. Thus the transfer matrix describes only this part of the system.
This is the reason that the transferfunction description and the statespace description are not
necessarily equivalent. For example. if any Amatrix other than Aco has an eigenvalue with a
Figure 6.7 Kalman decomposition.
u
y
c
1 64 CO'TROLL,BI LI TY AND OBSERVABI LITY
posi ti \ e real part. then some state variable may grow without bound and the system may bum
Ollt. This phenomenon. however. cannot be detected from the transfer matri x.
The MATL.B function :�:. :eal. an acronym for minimal rea/i:ation. can reduce any
state equation to (6. 46). The reason for calling it mi nimal realization wi l l be given in the next
chapter.
E",UIPLE 6.9 Consider the network shown i n Fig. 6. S( a) . Because the input is a current source.
re'ponses due to the initial conditions i n C1 and L I wi l l not appear at the output. Thus the state
vari ables associated with C
I and L I are not observable: whether or not they are controllable
is immaterial in subsequent discussion. Si mi l arly. the state variable associated with L2 is not
controllable. Because of the symmetry of the four l Q resistors. the state variable associated
\\ ith C: is neither controllable nor observ�ble. By dropping the state variables that are either
uncontrollable or unobservable. the network in Fig. 6. S( a) can be reduced to the one in Fig.
6, S( b), The current i n each branch is un: thus the output Y equals 2 ` (11 /2) or y = ll , Thus
the transfer function of the network i n Fig. 6. 8(a) is g(s) = I .
If we assign state variables as shown. then the network can be described by
,
=
0. 5 0
0 0
0 0. 5
0 0
v = [0 0 0 l lx + ll
Because the equation is already of the form shown i n (6. 40). it can be reduced to the following
wntrol l able state equation
x, = �

�
.
5
`Xc
0
�
5
[
u
\' = [0 OJ,, + 11
The output is independent of X, : thus the equation can be further reduced to v = 11 . This is
wh,l! \\ e will obtain by using the :IATLAB function :�:lreal .
6, 5 Condi ti ons i n JordanForm Equati ons
Control l ability and observabi li ty are invariant under any equivalence transformation. If a state
equation is transformed into Jordan form. then the control lability and observabi lity conditions
become \ ery simple and can often be checked by inspection, Consider the state equation
x = ]x + Bx
(6.47)
y = Cx
'
C1 ~ 2
I¸
L· = 1
1 "
] [ L¸ = ] `*
¬
!¡ ,
L_ ¯ 2
1 11 1 Q
l ð)
6, 5 Condi ti ons i n JordanForm Equati ons 1 65
¬
I \: ] .
1 Q l Q
l C l
Figure 6.8 Networks.
where J is in Jordan form. To simpli fy discussion. we assume that J has only two distinct
eigenvalues A I and 12 and can be written as
J
=
diag (
J
I . J2)
where JI co�sists of al l Jordan blocks associated with AI and J2 consists of al l Jordan blocks
associated with i.e. Again to simplify discussion. we assume that J ¡ has three Jordan blocks
and J2 has two Jordan blocks or
The row of B corresponding to the last lV of J'l i s denoted by bill '
The column of C
corresponding to the frst column of J'l is denoted by cf'i
'
Theorem 6. 8
1. The state equation i n (6. 47) i s controllable if and onl y i f the three row vectors { bi l l . b1 l 2 . bl 1 ) } are
l i nearly independent and the two row vectors {bI2 1 . b122 } are l i nearly independent.
2. The state equation in (6.47) is observable i f and only i f the three column vectors { CI I I . Cf1 2 . Cf! 3 }
are l i nearly independent and the two column vectors {C/21 . ef22 } are linearly independent.
We discuss fi rst the implications of this theorem. If a state equation is i n Jordan form,
then the controllability of the state variables associated with one eigenvalue can be checked
independently from those associated with different eigenvalues. The controllability of the state
variables associated with the same eigenvalue depends oilly on the rows of B corresponding
to the last row of all Jordan blocks associated with the eigenvalue. All other rows of B play no
role i n determi ni ng the controllability. Similar remarks apply to the observability part except
that the columns of C corresponding to the first column of all Jordan blocks determine the
obser\abi l i ty. We use an example to i l lustrate the use of Theorem 6. 8.
1 66 CONTROLLABI LITY AND OBSERVABI LITY
EXAMPLE 6.10 Consider the Jordanform state equation
)
0 0 0 0 0 0 0 0
0 AI 0 0 0 0 0 1 0 0
0 0 AI 0 0 0 0 0 0
x = 0 0 0 AI 0 0 0 x + 1 U
0 0 0 0 A2 0 1 2 3
0 0 0 0 0 A2
0 0
(6.48)
0 0 0 0 0 0 A2
1 2 0 0 2
X
y
= 0 2 0 1
0 2 3 0
Ï.
í ¬
The matrix J has two distinct eigenvalues AI and A2 · There are three Jordan blocks, with order
2, 1 , and 1 , associated with A I . The rows of B corresponding to the last row of the three Jordan
blocks are [1 0 0], [0 1 0], and [ 1 1 1] . The three rows are linearly independent. There is only
one Jordan block, with order 3, associated with A2. The row ofB corresponding to the last row
of the Jordan block is [ 1 1 1], which is nonzero and is therefore linearly independent. Thus we
conclude that the state equation in (6.48) is controllable.
The conditions for (6.48) to be observable are that the three columns [1 1 I )' , [2 1 2)', and
[0 2 3]' are linearly independent (they are) and the one column [0 0 0]' is linearly independent
(it is not). Therefore the state equation is not observable.
æ
Before proving Theorm 6. 8, we draw a block diagram to show how the conditions in the
theorem arise. The inverse of (s I ¯ J) is of the for shown in (3.49), whose entries consist
of only 1 /(s  Ai)k. Using (3.49), we can draw a block diagram for (6.48) as shown in Fig.
6.9. Each chain of blocks corresponds to one Jordan block in the equation. Because (6.48) has
four Jordan bloks, the fgure has four chains. The output of each block can be assigned as a
state variable as shown in Fig. 6. 1 0. Let us consider the last chain in Fig. 6. 9. If bl2l = 0, the
state variable Xl21 i s not connected to the input and is not controllable no matter what values
bz21 and bl 21 assume. On the other hand, if bl2! is nonzero, then all state variables in the
chain are controllable. If there are two or more chains associated with the same eigenvalue,
then we require the linear independence of the first gain vectors of those chains. The chains
associated with different eigenvalues can be checked separately. All discussion applies to the
observability part except that the column vector L fi} plays the role of the row vector bli} .
Proof of Theorem 6.8 We prove the theorem by using the condition that the matrix
[A ¯ sl B] or [sl  A B] has full row rank at every eigenvalue of A. In order not to be
overwhelmed by notation, we assume [sl  J B] to be of the form
s  AI ¯1 0 0 0 0 0 bi l l
0 5  AI 1 0 0 0 0 b211
0 0 5 ¯ AI 0 0 0 0 bi l l
0 0 0 s  AI 1 0 0 bl l 2
(6.49)
0 0 0 0 s  AI 0 0 bl l 2
0 0 0 0 0 s  A2 1 bl 2!
0 0 0 0 0 0 s  A2 bl21
6. 5 Conditions in JordanForm Equati ons
Figure 6.9 Block diagam of (6.48).
�
�
Figure 6.10 Interal structure of 1 / (s  Ai )'
J
Ï
1
I
1 67
The Jordanform matrix J has two distinct eigenvalues AI and A2. There are two Jordan
blocks associated with AI and one associated with )'2. If 5 = A
1
, (6.49) becomes
0 1 0 0 0 0 0 bI l l
0 0 1 0 0 0 0 b2l l
0 0 0 0 0 0 0 bi l l
0 0 0 0  I 0 0 bI l l (6.50)
0 0 0 0 0 0 0 bl\
0 0 0 0 0 AI  A2 1 bl2l
0 0 0 0 0 0 AI  A2 bl2l
Ï
1 68 CONTROLLABI LITY AND OBSERVABI LITY
The rank of the matrix wi ll not change by elementary column operati ons. We add the
product of the second column of (6.50) by b i l l t o the last block col umn. Repeating the
process for the third and ffth columns. we can obtain
0 · . 0 0 0 0 0 0
0 0 . 0 0 0 0 0
0 0 0 0 0 0 0 bi l l
0 0 0 0
0 0 0
0 0 0 0 0 0 0 bi l e
0 0 0 0 0 ;·
1
 Ac  . bl e l
0 0 0 0 0 0 Al  A2 bil l
Because i' l and A2 are distinct.
A
I ¯ );2 is nonzero. We add the product of the seventh
column and bil l /0, j  A2 ) to the last tolumn and then use the sixth col umn to el i minate
its righthandside entries to yield
0  . 0 0 0 0 0 0
0 0  . 0 0 0 0 0
0 0 0 0 0 0 0 bi l l
0 0 0 0 I 0 0 0  · . ·
0 0 0 0 0 0 0 bl l 2
0 0 0 0 0 Al 1.2 0 0
0 0 0 0 0 0
AI  A2 0
It is clearthat the matrix in  · . , has ful l row rank if and only if bi l l and bl l 2 are l i nearly
independent. Proceeding similarly for each eigenvalue, we can establish Theorem  ·
Q. E. D.
Consider an ndimensional Jordanform state equation with p i nputs and q outputs. Ii
there are H, with m ` p. Jordan blocks associated wi th the same eigenvalue. then 11 / number
of I x p row vectors can never be l i nearly independent and the state equation can never be
controllable. Thus a necessary condition for the state equation to be control lable i s 11/ :: p.
Si mi larly, a necessary condition for the state equation to be observable i s 7l :: q. For the
si ngleinput or sl1gleoutput case. we then have the fol l owi ng corol laries.
Corollary 6.8
A singleinput Jordanform state equation i s controllable i f and only i f there i s only one Jordan block
a .. :�1.iatcd with each distinct eigenvalue and every entry of ecorresponding to the la�t IO\¹ of each JonJan
bl oc' k i s di fferent from zero.
Corollary 6.08
A singleoutput Jordanform state equation i s obsen able if and onl y if there i s onl y one JDrdan block
associated with each distinct eigenvalue and every entry of C corresponding to the frst col umn of each
Jord,m block is different from zero.
6. 6 Di screteTi me State Equat i ons
EX . .nIPLE 6. 1 1 Consider the state equation
�� � �
�
`
0 0 0 0
· +
0
o 0 0 . .
· . 0 0 ·i ·
1 69
 ·:·
There are two Jordan blocks. one wi th order 3 and associated wi th eigenvalue .the other with
order .and associated with eigenvalue 2. The entry of ecorresponding to the l ast row of
the frst Jordan bl ock i s zero: thus the state equation i s not controllable. The two entries of C
corresponding to the fN column of both Jordan blocks are dillerent from zero: thus the state
equation i s obser\able.
6.6 DiscreteTi me State Equations
Consider the ndimensional pinput q output state equation
··+ .a··i+ea ·
· · c· ·i
where a.e.and C are. respectively, II x n. n x p. and q X n real constant matrices.
 ·
Defnition 6.DI The discreretime state eqllalion (6.53) or rhe pair a e·is said to he
controllable ifjor any inirial state ·· X¡ and anvfnal state XI . there exist\ Ul inpllr
sequence ojfnire lengrh rhar rtansjers Xn to XI . Orhenvise rhe eqllation or \ e is
said to he uncontrollable.
Theorem 6. 01
The fol l owi ng statements are equi\' alent:
] , The ndimensional pair · ei s control l abl e.
2. The n x 11 mJtnx
tt  l
\Y,icln  I I _ \· ee ·.
ttt=0
i s nun.ingular.
3. Th� n X np cnntrollll/Jiliry Illt!Il.\
(I e ae a` e
has rank n ( ful l row rank). The matrix can be generated by cal l i ng � ¯ !. i n MATLAB.
4. The n x (n + p) matrix [A ¯ AI ehas full row rank at every eigenval ue. A. of a
5. If. i n addition. al l eigenvalues ,)f A have magnitudes iess than I . then the unique solution of
1 6.5. ))
1 6.5 5 1
1 70 CONTROLLABI LITY AND OBSERVABI LITY
(6.56)
is positive defnite. The solution is called the discrete controllabilit Gramian and can b obtained by
using the MATLAB function dgram. The discrete Gramian can be expressed as
O
... (6.57)
�
The solution of (6.53) at was derived in (4.20) as
.+ .
.
which can be written as
. .
.
(6.58)
It follows from Theorem 3. 1 that for any and x[n], an input sequence exists if and only
if the controllability matrix has full row rank. This shows the equivalence of ( 1 ) and (3). The
matrix . 1] can be written as
.
. . .
.
The equivalence of (2) and (3) then follows Theorem 3. 8. Note that . .is always positive
semidefnite. If it is nonsingular or, equivalently, positive defnite, then (6.53) is controllable.
The proof of the equivalence of (3) and (4) is identical to the continuoustime case. Condition
(5) follows Condition (2) and Theorem 5.D6. We see that establishing Theorem 6.Dl is
considerably simpler than establishing Theorem 6. 1 .
There is one important diference between the continuous and discretetime cases. I a
continuoustime state equation is controllable, the input can tansfer any state to any other
state in any nonzero time interval, no matter how small. If a discretetime state equation is
controllable, an input sequence of length can transfer any state to any other state. If we
compute the controllability index J as defned in (6. 1 5), then the transfer can be achieved
using an input sequence of length J. If an input sequence is shorter than J, it is not possible
to transfer any state to any other state.
Defnition 6.D2 .(6.53) .
observable .7 0
..~ 0 U
..
unobservable.
l
6. 6 Di screteTi me State Equati ons 1 7 1
� Theorem 6.001
The following statements dequi valent:
1. The ndimensional pair .is observable.
2. The x matrix
. 1 ] = .. (6.59)
is nonsingular or. equivalently. positive defnite.
3. The x observabilit matri
has rank (full column rank). The matrix can be generated by calling obsv in MATLAB.
4. The + x matrix
has full column rank at every eigenvalue. of .
5. If. in addition. all eigenvalues of .have magnitudes less than I . then the unique solution of
. ... (6. 61 )
i s positive defnite. The solution i s called the discrete obserabilitv Gramian and can be expressed as
¯
(6.62)
This can be proved directly or indirectly using the duality theorem. We mention that all
other propertiessuch as controllability and observability indices, Kalman decomposition,
and Jordanform controllability and observability conditionsiscussed for the continuous
time case apply to the discretetime case without any modifcation. The controllabilitv index
and observability index, however, have simple interpretations in the discretetime ca�e. The
controllability index is the shortest input sequence that can transfer any state to any other state.
The observability index is the shortest input and output sequences needed to determine the
initial state uniquely.
6. 6. 1 Controllability to the Ori gi n and Reachability
In the literature, there are three diferent controllability defnitions:
1. Transfer any state to any other state as adopted in Defnition 6.D
I
.
2. Transfer any state to the zero state, called controllability to the origin.
1 72 CONTROLLABI LITY AND OBSERVABI LI TY
3. Transfer the zero state to any state, called controllability from the origin or, more often,
reachabiiity.
In the continuoustime case. because eAr is nonsingular, the three definitions are equiva
lent. In the discretetime case, if A is nonsingular, the three defnitions are again equivalent.
But if A i s singular. then ( I ) and ·· are equivalent, but not (2) and ·· The equivalence of ( 1 )
and ·· can easily be seen from  :s· We use examples to discuss the difference between : ·
and ·· Consider
· · + 11 � �·.·¦+
�·i
 ··
Its controllability matrix has rank and the ,equation i s not controllable as defined in ( 1 ) or not
reachable as defned in ··The matrix ahas the fom1 shown in ···and has the property
Ak = 0 for ·~ ·Thus we have
for any initial state ·  Thus every state propagates to the zero state whether or not an input
sequence i s applied. Thus the equation i s controllable to the origin. A different example fol lows.
Consider
+·
Its controllability matrix
has rank 1 and the equation i s not reachable. However. for any ;= a and .· = f,
the input . . ;~ :.f transfers · it o · .  = O. Thus the equation is controllable t o the
origin. Note that the Amatrices in  ··and +·are both singular. The definition adopted
in oefi nition o.encompasses the other two definitions and makes the discussion simple.
For a thorough discussion of the three defnitions, see Reference +
6, 7 Control l abi lity After Sampl i ng
Consider a continuoustime state equation
x( t) = a···+Bu(t)
If the input i s piecewise constant or
.. · : = .. ·r· = .. · · for · rS t < ·÷. · r
then the equation can be described, as developed i n .+ . :· by
s ·+.  ~ as ·i +sa·i
 :·
 ·
¬
6. 7 Control l abi l ity After Sampl i ng 1 73
with
B = ,
T
eAr dtB =: MB

The question i s: If  · . i s controllable. wi l l i t s sampled equation in   be controllable') This
problem is i mportant i n designing socalled deadbeat sampleddata systems and in computer
control of continuoustime systems. The answer to the question depends on the sampling period
rand t
�
e location of the eigenvalues of A. Let ·and X, be. respectively. the eigenvalues of
A and A. We use Re and 1m to denote the real part and imaginary part. Then we haw the
following theorem.
Theorem 6.9
Suppose (6. 65) i s control l able . . ¬ sufficient condition for its discretized equation i n (6. 66). wi th sampling
period r to be control l able i s that I Im ·,  · i c ··· · rfor H = I . 2 . . . whene,r
e· ·
·
= For the singleinput case. the condition i s necessary as wel l .
Fi rst we remark on the conditions. If A has only real eigenvalues. then the discretized
equation with any sampling period T ` i s always controllable. Suppose A has complex
conjugate eigenvalues a = j t. If the sampling period rdoes not equal any integer multiple of
T·f, then the discretized state equatiQn is controllable. If T = mI ·f for some integer ·· then
the discretized equation lila\, ··be controllable. The reason is as follows. Because A = eAT . if
· is an eigenvalue of A. then X, := eA,
T
is an eigenvalue of A ( Problem · . ·If T = ··  c
the t wo distinct e
�
genvalues i' l = . + j f and ·
= a  j f of A become a repeated eigemalue
_eaT or ea
T
of aThis \\ ill cause the discretized equation to be uncontrollable. as \\ e \\ ill
see in the proof. We show Theorem by assuming ato be in Jordan form. This i s permirred
because controllability is in\ ariant under any equivalence transformation.
Proof of Theorem 6.9 To simplify the discussion, we assume ato be of the form
[
·
A d"" A" . A". A, , ) �
·
·
()
11
f" I
·
 ·
In other words, ahas t\\ o distinct eigenvalues · and A2 . There are two Jordan blocks.
one with order ·and one with order . associated with · and only one Jordan block of
order :associated \\ ith i. 2 . Using ·+s. we have
. . = diagl '
.
I I '
A
1
2 . ·
�
[
I
Te
;· , T
T
2
e; ' T
··
e� i 7
Te
;· , T
I
e
i. I T
( (
( e/q T
(
 
( eA
J T
( e
A;.
1 74 CONTROLLABI LITY AND OBSERVABI LITY
This is not in Jordan form. Because we will use Theorem which is also applicable to
the discretetime case without any modifcation, to prove Theorem we must transform
A in into Jordan form. It turs out that the Jordan form of A equals the one in
if Aj is replaced by >j
:
= ¿
^,
` (Problem In other words, there exists a nonsingular
triangular matrix P such that the transformation i
=
Pi will transform into
i[k + I J = PAPl i[kJ + PMBu[kJ
with PAP

I in the Jordan form in with Aj replaced by >j . Now we are ready to
establish Theorem
First we show that M in is nonsingular. If A is of the for shown in
then M is block diagonal and triangular. Its diagonal entry is of form
mjj
:
=
f
`
,
,
=
·
l )/Ai if Ai f
if Ai =
Let Aj = O¡ + jfj . The only way for mjj
=
is O¡ = and fj T
= 2rmIn this case,
jfj is also an eigenvalue and the theorem requires that 2fi ·.2rm. Thus we conclude
mjj f and M is nonsingular and triangular.
If A is of the form shown in then it is controllable if and only if the third
and fourth rows of B are linearly independent and the last row of B is nonzero (Theorem
Under the condition in Theorem the two eigenvalues >
1
= ¿´' ` and >2 =
of A are distinct. Thus is controllable if and only if the third and fourth rows of
PMB are linearly independent and the last row of PMB is nonzero. Because P and M
are both triangular and nonsingular, PMB and B have the same properties on the linear
independence of their rows. This shows the suffciency of the theorem. If the condition
in Theorem is not met, then >1 = >
2. In this case, is controllable if the third,
fourth and last rows of PMB are linearly independent. This is still possible if B has three
or mo·e columns. Thus the condition is not necessary. In the singleinput case, if >
1 =
then has two or more Jordan blocks associated with the same eigenvalue and
is, following Corollary not controllable. This establishes the theorem. Q.E.D.
In the proof of Theorem we have essentially established the theorem that follows.
: Theorem 6. 1 0
If a continuoustime linear timeinvariant state equation is nO{ controllable, then its discretized state
equation. with any sampling period, is not controllable.
This theorem is intuitively obvious. If a state equation is not controllable using any input,
it is certainly not controllable using only piecewise constant input.
EXAMPLE 6.12 Consider the system shown in Fig. Its input is sampled every ·seconds
and then kept constant using a hold circuit. The transfer function of the system is given as
s + 2 s + 2
g
(s) =
s3 + 3s2 + 7s + 5
=
(s + l ) (s + + j2) (s + j2)
U({)
�
�
r
s + 2 `
(s + I ) (s + ¡ + 2j)(s + 1  2j )
Figre 6.11 System with piecewise constant input.
6.7 Control l abi l ity After Sampl i ng
3 ~ ~! l
I
÷Z
3
Using we can readily obtain the state equation
y
=
[2Jx
1 75
to describe the system. It is a controllableform realization and is clearly controllable. The
eigenvalues of A are ± j2 and are plotted in The three eigenvalues have
the same real part: their diferences in imaginary parts are : and 4. Thus the discretized state
equation is controllable if and only if
2rm
·.
2
 = rm and
:7m
·.
= .
for m = 2, . . . . The second condition includes the frst condition. Thus we conclude that
the discretized equation of is controllable if and only if ·. .for any positive
integer m.
We use MATLAB to check the result for m = or ·= Typing
a� [  3 7  5 ; 1 0 0 ; 0 1 O I ; b� [ l ; O ; O I ;
[ ad , bdl � c2d ( a , b , pi / 2 )
yields the discretized state equation as
i[k + l ]
=
i[k] + u[k]
Its controllability matrix can be obtained by typing C trb ( ad, bd) , which yields
Cd =
Its first two rows are clearly linearly dependent. Thus Cd does not have full row rank
and is not controllable as predicted by Theorem We mention that if we type
1 76 CONTROLLABI LITY AND OBSERVABI LI TY
raC' ( c t rb ( ad, bd» , the result i s 3 and (6. 74) is controllable. This is incorrect and is due
to roundoff errors. We see once again that the rank is very sensitive to roundoff erors.
What has been discussed i s also applicable to the observability part. In other words. under
the conditions in Theorem 6.9, if a continuoustime state equation is observable. its discretized
equation i s also observable.
6. 8 LN State Equati ons
Consider the ndimensional pinput qoutput state equation
x = A(t)x + B( t ) u
y = C(t)x
(6. 75)
The state equation is said to be controllable at 10, i f there exists a finite II ` 10 such that for any
x(to) = Xo and any XI . there exists an input that transfers Xo to X¡ at time II · Othef\\ be the state
equation is uncontrollable at 10. In the timeinvariant case, if a state equation is controllable.
then it i s controllable at every 10 and for every II 7 10; thus there is no need to specify 10 and
I I . In the timevarying case, the specifi cation of 10 and II is crucial.
Theorem 6. 1 1
The ndimensional pair ( A( t ) . B( I ) ) is controllable at time 10 if and only if there exists a rinite II ` 10
such that the n X n matrix

Wc ( to . I I ) = <(11 . r) B( r) B' ( r) <' (lI . r) dr
!0
(6. 76)
where <(1. r) is the state transition matrix of x = A( I)x. is nonsingular.
Proof We frst show that if Wc (lo. IJ l i s nonsingular. then ( 6. 75) is contwllJble. The
response of ( 6. 75) at II was computed i n (4. 57) as
x(tJ l = <( t l . lolxo +

<(tl . r )B( r)u( r) dT
'0
( 6 77)
We claim that the input
u( t) = B' ( I ) <'(tI . I) W,� I (tO. IJ l [ <( t I . lo)XIJ  xi l
( 6 78)
will transfer Xo at time 10 t o XI at time II ' Indeed. substituting (6. 78) into (6. 1 � ) yields

x(tl l = <(II . lo)Xo ¯ <(11 . r ) B( r) B' (r) <' UI . r ) dr
lu
W;
I
(to. ti l [ < (tl . to)Xo  xi l
= <( tl , to)Xo ¯ Wc(to. t l ) W;
1
(10. I J l [ <(tI ' (0)Xo  XI ] = X
6. 8 LTV State Equati ons 1 77
Thus the equation is controllable at to. We show the converse by contradiction. Suppose
( 6. 75 ) i s controllable at 10 but Wc (Io. I) is singular or, positive semidefnite. for all II ` 10.
Then there exists an n x I nonzero constant vector \ such that
which implies
B' (r ) <' (lI . r)v " 0 or V' <(t I . r) B( r) " 0 ( 6 79)
for all r in [10. II j . If (6. 75) is controllable. there exists an input that transtirs the initial
state Xo = <(10. I) )V at 10 to X( II ) = O. Then (6. 77) becomes

0 = <(r l . 10) <(10. I J l v <(11 . r) B( r)uI r) dr
'0
(6. 801
Its premultiplication by ¹ yields

0 = v\· + v' <(rl . r) B( r) u( r) dr = I l vl l ' + 0
ÎU
This contradicts the hypothesis V i O. Thus if (A( t ) . B( I ) ) is controllable at 10. W, (10. I I )
must be nonsingular for some finite II ` 10. This establishes Theorem 6. 1 1 . Q. E. D.
In order to apply Theorem 6. 1 1 , we need knowledge of the state transition matri x. which.
however, may not be available. Therefore it is desirable to develop a controllability condition
without involving <( 1 . r) . This i s possible if we have additional conditions on AU) and Bi l l .
Recall that we have assumed A( I ) and B( I ) t o be continuous. Now we require them to be
(n  I ) times continuously differentiable. Defne II0 U) = B( I ) . We then defne recursively a
sequence of II x p Uatr\es M", (I) as
.
for m = O. I . . . . . l/  I . Clearly. we have
for any fixed I,. Using
i
at
' <(t,. I) = <( 1, . I ) A( I )
( Problem 4. 1 7) . we compute
( a d
;[ <1 1, . t ) B( tl ] = [ <( 1, . I ) ] B( I ) + <(I" I ) B( I )
,)1 DI <II
d
= <( h. t I [ A( I ) MoU) + Mo ( t I ] = <( h. I ) MI ( I ) 
dt

(6. 8 1 )
1 78
CONTROLLABILITY AND OBSERVABI LITY
Proceeding forward, we have
a
m

a
4Uz, t )B(t) ¨ 4 (tz , t)Mm U)
1
m
for m = . . The following theorem is suffcient but not necessary for to be
controllable.
Theorem 6. 12
Let A(I ) and B(t) b n times continuously differentiable. Then the ndimensional pair (A(t), BU))
is controllable at 10 if there exists a finite t\ ¯ to such that
(6.83)
Proof: We show that if holds, hen Wc (to , t) is nonsingular for all I : II · Suppose
not, that is, W c (to, t) is singular or positive semidefnite for some t2 : t l · Then there
exists an n x nonzero constant vector v such that
which implies
'
v'Wc(to, tz)v = v'4(lz, r) B(r)B' (r)4'(tz, r)vdr
!0
2
I I B'(r) 4'(t2, r)vl l dr ¬
'0
B'( r) 4' (tz, r)v = 0 or v'4(tz, r)B(r) = 0
for all r in [to, t2] ' Its differentiations with respect to r yield, as derived in
v'4(tz, r)M
m
(r) = 0
for m = O. . , n ¯ and all r in [to. tz], in particular, at I I · They can be arranged as
Because 4(t2 , tl ) is nonsingular, v'4(t2. II ) is nonzero. Thus contradicts
Therefore, under the condition in Wc(to, t2), for any 12 : II , is nonsingular and
(AU) , BU» is, following Theorem controllable at to· Q. E. D.
EXAMPLE 6.13 Consider
6. 8 LTV StatEquations 1 79
The determinant of the matrix
t
tZ
t
2
¯ 1
is t
2
+ which is nonzero for all t. Thus the state equation in is controllable at every t.
EXAMPLE 6. 14 Consider
and
x =
°
x +
u
Equation is a timeinvariant equation and is controllable according to Corollary
Equation is a timevarying equation; the two entries of its Bmatrix are nonzero for all
t and one might be tempted to conclude that is controllable. Let us check this by using
Theorem Its state transition matrix is
and
We compute
Wc Uo, t) ¯
'
(t  to)
(t  to)
Its determinant is identically zero for all to and t. Thus is not controllable at any to. From
this example, we see that, in applying a theorem, every condition should be checked carefully;
�
�
otherwise, we might obtain an eroneous conclusion.
We have M ~ [ and compute
MI = A(t)Mo + ·Mo =
�
t
M
¬ A(t)MI + �MI " �·
dt
t
2
¬ I
We now discuss the observability part. The linear timevarying state equation in
is observable at to if there exists a fnite tl such that for any state
x(to) ~ xo, the knowledge
of the input and output over the time interval [to, tl ] suffces to determine uniquely the initial
state Xo. Otherwise, the state equation is said to be unobservable at to .
.. Theorem 6.01 1
The pair (A(t ) , C(t» is observable at time to if and only if there exists a fnite tl ¯ to such that the
n x n matrix
1 80 CONTROLLABI LI TY AND OBSERVABI LITY
'
 . · . ·.··· .·
/0
(6.89)
where . · is the state transition matrix of X .. is nonsingular.
> Theorem 6.01 2
Let At and . be n times continuously differentiable. Then the ndimensional pair (A(t ) , .
is observable at if there exists a fni t e l 7 such that
where
with
 I
 ·
rank
�
n
H 0, 1 . .  . , n ¯ I
 .
(6.90)
We mention that the duality theorem in Theorem 6.5 for timeinvariant systems i s not
applicable to timevarying systems. It must be modifed. See Problems 6. 22 and 6. 23.
6.1 Is the state equation
x � 0 � . �
H
1 3 ¯3 0
y [ 1 2 .
controllable? Observable0
6.2 Is the state equation
controllable? Observable?
6,3 Is it true that the rank of [ B AB . . . A
n

I
BJ equals the rank of [AB A
2
B
. . . AnBP
If not, under what conditon wi l l it be true?
6.4 Show that the state equation

:
Probl ems 1 81
AU B¡
A22
X
0
.
is controllable if and only if the pair (A22 , A21 ) is controllable.
6.5 Find a state equation to describe the network shown in Fig. 6. 1 , and then check its
controllability and observability.
6.6 Find the controllability index and observability index of the state equations in Problems
6. 1 and 6. 2.
6.7 What i s the controllability index of the state equation
x .. ..
where I is the unit matrx0
6.8 Reduce the state equation
y [ 1 .
to a controllable one. Is the reduced equation observable?
6.9 Reduce the state equation in Problem 6.5 to a controllable and observable equation.
6.10
6.11
Reduce the state equation
0 0
·
.�.
,
�
A
l
0
0 A
l
0
0 0 A2
0 0 0
y = [0 I 1 1 0 ..
to a controllable and observable equation.
Consider the ndimensional state equation
x .. ..
, .. «.
The rank of its controllability matrix is assumed to be n I < n. Let , be an n x n  matrix
whose columns are any nl linearly independent columns of the controllability matrix.
Let  be an n l x n matrix such that  , I
n
1 , where Inl is the unit matrix of order
n ¡ . Show that the following n I dimensional state equation
· PI AQl xl + PI Bu
y .,¡ XI + Du
is controllable and has the same transfer matrix as the original state equation.
6.12 In Problem 6. 1 1 , the reduction procedure reduces to solving for PI in PI , 1. How
do you solve PI ?
1 82 CONTROLLABI LITY AND OBSERVABI LITY
6.13 Develop a similar statement as in Problem 6. 1 1 for an unobservable state equation.
6.14
6.15
6.16
Is the Jordanform state equation controllable and observable?
2 0 0 0 0 0 2 0
0 2 0 0 0 0 0 2
0 0 2 0 0 0 0 1
x = 0 0 0 2 0 0 0 x + 3 2 U
0 0 0 0 1 0 1 0
0 0 0 0 0 1 0 0 1
0 0 0 0 0 0 0 0
y
=
2 3 1
'
x 1 2
'
c 0
Is it possible to fnd a set of b;, and a set of c,, such that the state equation
,
=
0
1 0
0 1
0 0
0 0
c
' 
ci¯
Y = c
¯i c
¯¯
c)i
c)¯
is controllable? Observable?
Consider the state equation
0
x =
oi
hi
0
0
0
0
1
0
ci)
c
¯)
c))
0
Þi
oi
0
0
o b
¯i
c x + b)i
c b+i
1 bsi
bU
b
¯¯
b)¯
U
b+¯
bs¯
ci+
cz+
c)+
0
0
0
o
¯
h¯
cis
c
¯s
c)s
c bi i
c x + bi¯
u
h¯
b:i
o
¯
bz¯
Y =  ci c
' l
ci¯
c
¯i c
n
]
It is the modal form discussed in (4.28). It has one real eigenvalue and two pairs of
complex conjugate eigenvalues. It is assumed that they are distinct. Show that the state
equation is controllable if and only if hi = 0; b, i = 0 or b,
¯
= 0 for i = 1 . 2. It is
observable if and only if ci = 0; c,i = 0 or c,
¯
i 0 for i = 2.
6.17 Find two and threedimensional state equations to describe the network shown in Fig.
6. 1 2. Discuss their controllability and observability.
6.18 Check controllability and observability of the state equation obtained in Problem 2. 19.
Can you give a physical interpretation directly from the network?
'
Problems 1 83
Figr6.12
X2

N


t
�
�
2 F
X2 Xl '
2 Q
L_�+ _

6.19 Consider the continuoustime state equation in Problem 4. 2 and its discretized equa
tions in Problem 4.3 with sampling period T = 1 and 7. Discuss controllability and
observability of the discretized equations.
6.20 Check controllability and observability of
y = 0 l ]x
6.21 Check controllability and observability of
6.22
6.23
Show that (AU), BU» is controllable al to if and only if (A' (t) , B' (t)) is observable at
t
o
·
For timeinvariant systems, show that (A, B) is controllable if and only if (A, B) is
controllable. Is this true for timevarying systems?
LH3D!CI
Mi ni mal Real i zati ons
and Copr i me Fracti ons
7 . 1 I ntroduction
1 84
This chapter stu
A
dies further the realization problem discussed in Section 4. 4. Recall that a
transfer matrix G( s) is said to be realizable if there exists a statespace equation
x = Ax + Bu
y = Cx + Du
that has G(s) as its transfer matrix. This is an important problem for the following reasons. First.
many design methods and computational algorithms are developed for state equations. In order
to apply these methods and algorithms. transfer matrices must be realized into state equations.
As an example. computing the response of a transfer function in ;lATLAB is achieved by frst
transfonning the transfer function into a state equation. Second. once a transfr function is
realized into a state equation. the transfer function can be implemented using opamp circuits.
as discussed i n Section 2. 3 . 1 .
I f a transfer function i s realizable. then i t has infnitely many realizations. not necessarily
of the same dimension. as shown i n Examples 4.6 and 4.7. An important question is then raised:
What is the smallest possible dimension? Realizations with the smallest possible dimension
are called minimaldimensional or millimal realizations. I f we use a minimal realization to
implement a transfer function. then the number of integrators used i n an opamp circuit will
be minimum. Thus minimal realizations are of practical importance.
In this chapter, we show how to obtain minimal realizations. We will show that a realization
of g( s) = N( s) / D( s) i s minimal if and only if it is controllable and observable. or if and only
7. 2 I mpl ications of Coprimeness 1 85
if its dimension equals the degree of !( s) . The degree of g( s ) is defned as the degree of D( s) i f
the two polynomials D(s ) and N(; ) are coprime or have no common factors. Thus the concept
of coprime ness i s essential here. In fact. coprimeness in the fraction N ( s) / D(s ) pl ays the same
role of controllability and observability in statespace equations.
This chapter studies only linear timeinvariant systems. We study frst SISO systems and
then MI�10 systems.
7. 2 I mpl i cati ons of Copri meness
Consider a system with proper transfer function gI S) . We decompose i t as
gl ,) = g( :) gsp(s )
where gsp ( s) is strictly proper and g (x) is a constant. The constant g (o) yields the Dmatri�
in every realization and will not play any role in what will be discussed. Therefore we consider
i n this section only strictly proper rational functions. Consider
(7. I )
To simplify the discussion. we have assumed that the denominator D( s ) has degree 4 and i s
monic ( has I as its leading coefficient). In Section 4. 4. we introduced for ( 7. 1 ) the realization
in (4.4 I ) without any discussion of its state variables. Now we will redevelop 1 4. 41 ) by frst
defning a set of state variables and then discussing the impl ications of the coprimeness of
D( s) and N(s ) .
Consider
.v( s) = NC, ) D
I
(s ) // (s )
Let u s introduce d new variable dl) defined by l� ( S) = D
1
( s) ll ( s ) . Then we ha\'e
Defne state variables as
Then we have
D(s ) v (s ) = tl ls)
\(s) = Nls ) v( s)
vi r I

or XIS) =
v i r )
vi t l
� (s )
Xc (S )
·�3 (S)
.� ( S)
(7. 2)
(7. 3 )
( 7.4)
(7. 5 )
( 7. 6)
They are independent of ( 7. 1 ) and follow directly from the defnition in (7. 5). I n order to
develop an equation for .t l . we substitute ( 7. 5) into ( 7. 3) or
1 86 MI NI MAL REALIZATIONS AND COPRIME FRACTIONS
to yield
.
which becomes, in the time domain,
.
Substituting (7. 5) into (7.4) yields
.
.
I
.
which becomes, in the time domain,
.
Equations (7.6), (7. 7), and (7.8) can be combined as
. 1
. . . �
o 0
1 0
o
. .
X
�
o
0
o 0
This is a realization of (7. 1 ) and was developed in (4.41 ) by direct verifcation.
(7.7)
(7.8)
(7.9)
Before proceeding, we mention that if in (7. 1 ) is then . and the output
and its derivatives can be chosen as state variables. However, if is a polynomial of
degree 1 or higher and if we choose the output and its derivatives as state variables, then its
realization will be of the form
X . .
.
This equation requires differentiations of and is not used. Therefore, in general, we cannot
select the output and its derivatives as state variables.' We must defne state variables by
using Thus is called a
Now we check the controllability and observability of (7.9). Its controllability matrix can
readily be computed as
1
o
o
1. Se MVExample 2.16, in paricular. (2.47).
o
1
. 10
7.2 I mpl ications of Coprimeness 1 87
Its determinant is I for any Thus the controllability matrix C has full row rank and the
state equation is always controllable. This is the reason that (7.9) is called a
.
Next we check its observability. It turs out that it depends on whether or not and
are .Two polynomials are said to be .if they have no common factor of
degree at least l . More specifcally, a polynomial is called a common factor or a common
divisor ofD(s) and N(s) if they can be expressed as . ) and =
where and are polynomials. A polynomial is called a ..
(gcd) of and if( l ) it is a common divisor of and and (2) it can be divided
without remainder by every other common divisor of and N(s) . Note that if is a
gcd. so is for any nonzero constant Thus greatest common divisors are not unique 2
In terms of the gcd. the polynomials and are coprime if their gcd is a nonzero
constant. a polynomial of degree 0; they are not coprime if their gcd has degree or higher.
Theorem 7. 1
The controllable canonical form in (7.9) is observable if and only if and in (7. 1 ) are coprime.
Proof: We frst show that if (7.9) is observable, then and are coprime. We
show this by contradiction. If and are not coprime, then there exists a such
that
. . 0
. . 0
(7. 1 1 )
(7. 12)
Let us defne . p. ' ] 1 , it is a 4 X nonzero vector. Then (7. 1 1 ) can be written
as . . O. where is defned in (7.9). Using (7. 12) and the shifting property of
the companion form. we can readily verify
�
. .
o
o
(7. 13)
Thus we have . . .. . . . and .
. We compute. using .
�
0
.
.
.
.
.
.
.
which implies that the observability matrix does not have full column rank. This contradicts
the hypothesis that (7.9) is observable. Thus if (7.9) is observable. then and N(s)
are coprime.
Next we show the converse; that is, if and are coprime, then (7.9) is
observable. We show this by contradiction. Suppose (7.9) is not observable. then Theorem
6. 01 implies that there exists an eigenvalue of . and a nonzero vector such that
2. If we require R(s) to D monic. then the gcd is unique.
1 88 MI NI MAL REALIZATI ONS AND COPRI ME FRACTI ONS
or
· .
. ·
.
·. AI . and .. .
Thus Y is an eigenvector of · associated with eigenvalue AI . From we see that
. . (x¡ I ], is an eigenvector. Substituting this . into .. . yields
· . · ·
·
Þ+
Thus is a root of ·. The eigenvalue of · is a root of its characteristic polynomial,
which, because of the companion fOIT of · equals . Thus we also have O.
and . and ·. have the same fact
o
r . This contradicts the hypothesis that .
and ·. are coprime. Thus i f . and ·. are coprime, then · is observable. This
establishes the theorem. Q. E. D.
If · is a realization of .. then we have, by defnition,
. . ... ·
.
Taking its transpose yields
·. . . . ... ·
. . . .. ·
.
Thus the state equation
.¡
• 7 Í Cz
. · . . . .
.
.
y . .. . .
·
�
·
is a diferent realization of (7. 1 ) . This state equation is always observable and is called an
observable canonical/orm. Dual to Theorem Equation · is controllable if and only
if . and ·. are coprime.
We mention that the equivalence transformation x . P with
will transform · into
p �
��
.
.
. I
( 7. 1 5)
7. 2 I mpl ications of Coprimeness 1 89
This is also called a controllable canonical form. Similarly, will transform · into
�
��
. . . .
o .
·
o . ·
y . .. . I ]x
This is a different observable canonical form.
7. 2. 1 Mi ni mal Real izations
We frst defne a degree for proper rational functions. We call ·. . apolynomial/ractioll
or, simply, a/raction. Because
» ·. ·. .
·
.
.
. .
for any polynomial . fractions are not unique. Let R(s) be a greatest common divisor
(gcd) of ·. and D(s) . That is, if we write ·. N(s) R(s) and . D(s) R(s),
then the polynomials N(s) and D(s) are coprime. Clearly every rational function .. can be
reduced to .. . N.
D
(s). Such an expression is called a coprime/raction. We call .
a characteristic polynomial of .. The degree of the characteristic polynomial is defned as
the degree of .. Note that characteristic polynomials are not unique; they may differ by a
nonzero constant. If we require the polynomial to be monic, then it is unique.
Consider the rational function
.
1
..
·. 1 )
Its numerator and denominator contain the common factor . Thus its coprime fraction is
.. . . ·.
. I ) and its characteristic polynomial is ·.
·. · Thus the rational
function has degree 2. Given a proper rational function, if its numerator and denominator are
coprimeas is often the casethen its denominator is a characteristic polynomial and the
degree of the denominator is the degree of the rational function.
� Theorem 7. 2
A state equation · . . d) is a minimal realization of a proper rational function .. if and only if
· . is controllable and (A. . is observable or if and only if
dim · . deg ..
Proof: If · . is not controllable or if · . is not observable, then the state equation
can be reduced to a lesser dimensional state equation that has the same transfer function
(Theorems 6.6 and 6.06). Thus · . . d) is not a minimal realization. This shows the
necessity of the theorem.
1 90 MI NI MAL REALIZATIONS AND COPRI ME FRACTIONS
To show the sufciency, consider the ndimensional controllable and observable state
equation
x = Ax + bu
y = ex + du
Clearly its n x n controllability matrix
C = [b Ab ' " A
n
1 b
j
and its n x n observability matrix
(7. 16)
(7. 17)
(7. 1 8)
both have rank n. We show that (7. 1 6) is a minimal realization by contradiction. Suppose
the idimensional state equation, with i < n,
X = Ax + iu
y = cx + du
is a realization of g(s). Then Theorem 4. 1 implies d = d and
Let us consider the product
cb
=
¨cA
n
b
for U = 0, 1 , 2, . . .
eA,, l b
eAnb
eAn+1 b
CA2( � I)b
Using (7.20), we can replace every cAmb by cA mi. Thus we have
OC = On Cn
(7. 1 9)
(7.20)
(7.21 )
(7.22)
where On i s defned as i n (6.21 ) for the idimensional state equation in (7. 19) and Cn is
defned similarly. Because (7. 1 6) i s controllable and observable, we have p(O) = n and
p(C) = n. Thus (3.62) implies p(OC) = n. Now On and Cn are, respectively, n x i
and i X n; thus (3.61 ) implies that the matrix On Cn has rank at most i. This contradicts
peOn Cn ) = p( OC) = n. Thus (A, b, e, d) is minimal. This establishes the frst part of
the theorem.
7. 2 Impl ications of Coprimeness 1 91
The realization in (7.9) i s controllable and observable i f and only if g( s) =
N(s) / D(s) is a coprime fraction (Theorem 7. 1 ). In this case. we have dim A = deg
D(s) = deg g (s). Because all minimal realizations are equivalent, as will be established
immediately, we conclude that every realization is minimal if and only if dim A = deg
g(s) . This establishes the theorem. Q.E.D.
To complete the proof of Theorem 7.2, we need
t
he following theorem.
� Theorem 7.3
All minimal realizations of g(s) are equivalent.
Poof: Let (A, b. e. d) and (A, i, c, d) be minimal realizations of g(s). Then we have
I = d and, following (7.22),
OC = OC
Multiplying OAC out explicitly and then using (7.20), we can show
OAC = OAC
(7.23)
(7.24)
Note that the controllability and observability matrices are all nonsingular square matrices.
Let us defne
Then (7.23) implies
P = 01 0 = CC1
and p1 = 01 0 = CC1 (7.25)
From (7.23), we have C = OI OC = Pc. The frst columns on both side of the equality
yield i = Pb. Again from (7. 23), we have a = OCCI = OpI . The frst rows on both
sides of the equality yield c = cpI . Equation (7.24) implies
A = OI OACc1 = PAPI
Thus (A. b, c d) and (A, i, c, d) meet the conditions in (4.26) and, consequently, are
equivalent. This establishes the theorem. Q. E. D.
Theorem 7. 2 has many important implications. Given a �tate equation, if we compute its
transfer function and degree. then the minimality of the state equation can readily be determined
without checking its controllability and observability. Thus the theorem provides an alterative
way of checking controllability and observability. Conversely, given a rational function, if we
compute frst its common factors and reduce it to a coprime fraction, then the state equations
obtained by using its coeffcients as shown in (7.9) and (7. 1 4) will automatically be controllable
and observable.
Consider a proper rational function g(s) = N(s)/ D(s). If the fraction is coprime, then
every root of D(s) is a pole of g(s) and vice versa. This is not true if N(s) and D(s) are not
coprime. Let (A, b.c.d) be a minimal realization of g(s) = N(s)/ D(s). Then we have
1 92 MI NI MAL REALIZATIONS AND COPRI ME F RACTI ONS
N(s) I
D( )
= c(sI  A) 
l
b + d =
d I
c [Adj(sI  A)] b + d
s et(s  A)
If N(s) and D(s) are coprime, then deg D(s ) = deg g (s) = di m A. Thus we have
D( s) = k det(sI  A)
for some nonzero constant k. Note that k = i f D( s) i s monic. This shows that i f a state
equation
,
is controllable and observable, then every eigenvalue of A is a pole of gCs) and every
pole of g( s) IS an eigenvalue of A. Thus we conclude that if (A. h, c. d) is controllable and
observable. then we have
Asymptotic stability < BIBO stability
More generally, contrllable and observale state equations and coprime fractions contain
essentially the same infonnation and either description can be lIsed H carry out analysis and
design.
. .
7. 3 Computi ng Copri me Fracti ons
The importance of coprime fractions and degrees was demonstrated in the preceding section.
In thiS section, we discuss how to compute them. Consider a proper rational function
'
( s) =
N(s )
g
D( s)
where N\s) and D(s) are polynomials. If we use the MATLAB function roo�s to compute
their roots and then to cancel their common factors, we will obtain a coprime fraction. The
MATLAB function mi"cea� can also be used to obtain coprime fractions. In this section,
we introduce a different method by solving a set of linear algebraic equations. The method
does not offer any advantages over the aforementioned methods for scalar rational functions.
However, it can readily be extended to the matrix case. More importantly, the method will be
used to carry out design in Chapter 9.
Consider N(s) /D( s) . To simplify the discussion. we assume deg N( s) : deg DI S ) =
n = 4. Let us write
which implies
N(s) N(s)
D( s) D( s)
D(s) ( N(s) ) + N( s) D( s) = 0 ( 7 26)
It is
�
lear that D(s)
_
and N( s) ar
�
not coprime if and only if there exist polynomials :(s)
and D( s) WIth deg N( s) : deg D( s) < n = 4 t o meet ( 7. 26). The condition deg  < n
is crucial; otherwise, (7. 26) has infnitely many solutions N(s) = N(s ) R(s) a�d  s =
D(s) R(s) for any polynomial R( s) . Thus the coprimeness problem can b reduced to ,olving
the polynomial equation in ( 7. 26).
Instead of solving 17. 26) directly, we will change it into solving a set of linear algebraic
equatIons. We write
7. 3 Computing Coprime Fracti ons
D( s ) = Do + DI S + D2s' + DJs3 + D.js
o
N(s ) = No + Ni s + N,s ' + NJs) + Nos.j
DI S ) = Do + D
I S + D,s' + D)S
3
F(s) =
N
o + N
l
s + N,s' + 1V3 s3
1 93
( 7. 27)
where Do ; 0 and the remaining D, . N; . D" and N, can be zero or nonzero. Substituting
these into ( 7. 26) and equating to zero the coeffcients associated with s". for k = 0, 1 . . . . . 7,
we obtain
Do No 0 0 0 0 0 0
Fo
Do
DI NI Do No
0 0 0 0
D2 N2
DI NI
Do No
0 0 II
D3 N,; D2 N, DI NI Do iVa
DI
Sm : =
= 0 ( 7.28)
D. N.j D3 N.1 D2 N2
DI NI N2
0 0 D.j N. D3 N) D, N2
D2
0 0 0 0 D.j N
.
D) N3
N3
0 0 0 0 0 0 D.j N.
D)
This is a homogeneous l i near algebraic equation. The frst block column of S consists of two
columns formed from the coeffcients of D( s) and N ( s) arranged in ascending powers of s .
The second block column i s the frst block column shifted down one position. Repeating the
process until S is a square matrix of order 2n = 8. The square matrix S is called the Sv!l'ester
resultallt. If the Sylvester resultant is singular. nonzero solutions exist in ( 7. 28) (Theorem 3. 3) .
This means that polynomials
N
( s) and D(s) of degree 3 or less exist to meet 1 7. 26) . Thus
D(s) and N( s) are not coprime. tf the Sylvester resultant is nons in gular. no nonzero solutions
exist in (7. 28) or, equivalently, no polynomials ,v (5) and D( s) of degree 3 or less exist to meet
(7. 26). Thus D(s) and N( s) are coprime. In conclusion. D( s) and N( s) are coprime ifand
ollly the Svlvester resultant is nonsingular:
If the Sylvester resultant is singular. then N ( 5) / D( s) can be reduced to
N( s ) N(s )
D(s) D( s)
where lV( S) and D(s) are coprime. We discuss how t o obtain a coprime fraction directly from
( 7.28). Let us search line:rly independent columns of S in order from left to right. We call
columns formed from D; Dcolumns and formed from N; N,columns. Then every Dcolumn
is linearly independent of its lefthandside ( LHS) columns. Indeed, because D. ; 0, the frst
Dcolumn i s li nearly independent. The second Dcolumn is also linearly independent of its
LHS columns because the LHS entries of D.j are all zero. Proceeding forward, we conclude
that all Dcolumns are line:lrly independent of their LHS columns. On the other hand. an N
column can be dependent or independent of its LHS columns. Because of the repetitive patter
1 94 MI NIMAL REALIZATI ONS AND COPRI ME F RACTI ONS
of S, if an Ncolumn becomes linearly dependent on its LHS columns, then al l subsequent
Ncolumns are l inearly dependent of their LHS columns. Let / denote the number of linearly
independent Ncolumns in S. Then the (/ + )th Ncolumn is the frst Ncolumn to become
linearly dependent on its LHS columns and will be called the primar dependent Ncolumn.
Let us use SI to denote the submatrix of S that consists of the primary dependent Ncolumn
and all its LHS columns. That is, SI consists of / + I Dcolumns (all of them are linearly
independent) and / + I Ncolumns (the last one is dependent). Thus SI has 2(/ + I) columns
but rank 2/ + J . In other words, 81 has nullity 1 and, consequently, has one independent null
vector. Note that if n is a null vector, so is an for any nonzero a. Although any null v
�
ctor can
be used, we will use exclusively the null vector with as its last entry to develop N(s) and
D(s). For convenience, we call such a null vector a monic null vector. Ifwe use the MATLAB
function null to generate a null vector, thn the null vector must b divided by its last entry
to yield a monic null vector. This is illustned in the next example.
EXAMPLE 7.1 Consider
N(s) 653 + 52 + 35  20
D( s) = 254 + 753 + 1552 + 1 65 + 1 0
(7.29)
We have n = 4 and its Sylvester resultant S is 8 x 8. The fraction is coprime if and only if S
is nonsingular or has rank 8. We use MATLAB to check the rank of S. Because it is simpler to
key in the transpose of S, we type
d= [ 1 0 1 6 1 5 7 2 ] ; n= [  2 0 3 1 6 0 1 ;
s = [ d 0 0 O ; n 0 0 0 ; 0 d 0 0 ; 0 n 0 0 ; . . .
Û 0 d O ; 0 0 n O ; 0 0 0 d; 0 0 0 nJ ' ;
m=rank ( s )
The answer is 6; thus D(s) and N(s) are not coprime. Because all four Dcolumns of 8 are
linearly independent, we conclude that S has only two linearly independent Ncolumns and
/ = 2. The third Ncolumn is the primary dependent Ncolumn and all its LHS columns are
linearly independent. Let 81 denote the frst six columns of S, an 8 x 6 matrix. The submatrix
SI has three Dcolumn (all linearly independent) and two linearly independent Ncolumns,
thus it has rank 5 and nullity I . Because all entries of the last row of SI are zero, they can be
skipped in forming SI . We type
s 1 : [ d 0 Ü , ¯ 0 0 ; 0 d 0 ; 0 n 0 ; 0 0 d ; O 0 n l ' ;
z=nul l l s l )
which yields
ans z = [ 0 . 68 6 0 0 . 3 4 3 0  0 . 5 1 4 5 0 . 3 4 3 0 0 . 00 0 0 0 . 1 7 1 5 l '
This null vector does not have as its last entry. We divide it by the l ast entry or the sixth entry
of Z by typing
zb= z / z ( 6 )
7. 3 Computing Coprime Fractions
1 95
which yields
ans zb� [ 4 2  3 2 0 1 J '
This monic null vector equals [No Do  NI DI  N
2
D2 ]' . Thus we have
N(s) = 4 + 35 + O . 52
D(s) = 2 + 25 + 52
and
35  
254 + 753 + 1 552 + 1 65 + 1 0
=
52 + 25 + 2
�ecause th
�
null vector is computed from the frst linearly dependent Ncolumn, the computed
N(s) and D(s) have the smallest possible degrees to meet (7.26) and, therefore. are coprime.
ThiS completes the reduction of N(s)/ D(s) to a coprime fraction.
The preceding procedure can be summerized as a theorem.
l Theroem 7,4
Consider g(s) = N (5)/ D(s). We use the coeffcients of D(s) and N(s) to form the Sylvester resultant
S In (7. 28) and search its linearly independent columns in order from left to rght. Then we have
deg g(s) = number of linearly independent Ncolumns
=
: /
and the coeffcients of a coprime fraction g(s) ^ N(s) / D(s) or
[ No Do  NI DI · · ·  NI" DI" J '
equals the monic null vector of the submatrix that consists of the primary dependent iVcolumn and all
Its LHS lInearly independent columns of S.
.
.
We mention that if D and Ncolumns i n S are arranged in descending powers of s. then
It ¡5 not true that all Dcolumns are l i nearly independent of their LHS columns and that the
degree of g(s) equals the number of linearly independent Ncolumns. See Problem 7.6. Thus
it is essential to arrange the D and Ncolumns in ascending powers of 5 in S.
7. 3. 1 QR Decomposition
As discussed in the preceding section, a coprime fraction can be obtained bv searchinu linearlv
independent c'olumns of the Sylvester resultant in order from left to riuht. It ;ums out te widel�
available QR decomposition can be used to achieve this searching.
"
»
Consider an n X m matrix M. Then there exists an n x n orthogonal matrix Q such that
QM = R
where R is an upper triangular matrix of the same dimensions as M
.
Because Q operates on the
rows of M, the lmear Independence of the columns of M is preserved in the columns of R In
other words, if a column of R i s linearly dependent on its lefthandside (LHS) columns, so is
1 96 MI NIMAL REALIZATI ONS AND COPRI ME F RACTI ONS
the corresponding column of M. Now because R is i n upper triangular fonn. i t s mt h column is
linearly independent of its LHS columns i f and only i f its mth entry at the diagonal position is
nonzero. Thus using R. the l i nearly i
I
dependent columns of M. in order from left to right, can
be obtained by inspection. Because Q is orthogonal, we have Q
I
" Q' =: Q and Q1 = R
becomes M " QR. This is called QR decomposition. In MATLAB, Q and R can be obtained
by typing [ q, rJ �qr ( ::) .
Let us apply QR decomposition to the resultant in Example 7. 1 . We type
do [ 1 0 1 6 1 5 7 2 ] ; no : 2 0 3
`
6 � . ;  .
S o l d Ü . ' ¦ O 0 0 ; 0 d 0 C ; C n . 0 ;
0 0 d 0 ; 0 0 ! 0 ; 0 0 0 6; 0 0 Û r. ] ¦
[ q, I J oqr ( s )
Because Q is not needed, we show onl y R:
25 . 1 3. 7 20. 6 1 0. 1  1 1 . 6 1 1 . 0 4. 1 5. 3
0 20. 7  1 0. 3 4. 3 7. 2 2 . 1 3. 6 6. 7
0 0 1 0. 2 1 5. 6 20. 3 O. S  1 6. S 9. 6
0 0 0 S. 9 3. 5 1 7. 9  1 1 . 2 7. 3
r =
0 0 0 0 5 .0 0  1 2. 0  1 5 .0
0 0 0 0 0 0 2. 0 0
0 0 0 0 0 0 4. 6 0
0 0 0 0 0 0 0 0
We see that the matrix is upper triangular. Because the sixth column has 0 as its sixth entry
(diagonal position). it is l i nearly dependent on its LHS columns. So is the last column. To
determine whether a column i s linearly dependent. we need to know only whether the diagonal
entry is zero or not. Thus the matrix can be si mplifed as
d x x x x x x x
0 n x x x x x x
0 0 d x x x x x
0 0 0 n x x x x
r =
0 0 0 0 d 0 x x
0 0 0 0 0 0 x 0
0 0 0 0 0 0 d 0
0 0 0 0 0 0 0 0
where d. n. and x denote nonzero entries and d also denotes Dcolumn and n denotes iV
column. We see that every Dcolumn is linearly independent of its LHS columns and there are
only two linearly independent iVcol umns. Thus by employing QR decomposition. we obtain
immediately / and the primary dependent iVcolumn. In scalar transfer functions. we can use
either rank or qr to fnd /. In the matrix case, using rank is very inconvenient; we wi l l use
QR decomposition.
7.4 Balanced Real i zati on 1 97
7. 4 Bal anced Real i zati on3
Every transfer function has infnitely many minimal realizations. Among these realizations, it
is of i nterest to see which realizations are more suitable for practical implementation. If we
use the controllable or observable canonical fonn, then the Amatrix and b or cvector have
many zero entries, and its implementation wi l l use a small number of components. However,
either canonical fonn is very sensitive to parameter variations; therefore both fons should
be avoided i f sensitivity is an important issue. I f all eigenvalues of A are distinct, we can
transfonn A, using an equivalence transfonation, into a diagonal fon (if all eigenvalues are
real) or into the modal fon discussed i n Section 4. 3 . 1 ( i f some eigenvalues are complex). The
diagonal or modal fon has many zero entries in A and wi l l use a small number of components
i n its implementation. More importantly, the diagonal and modal fons are least sensitive
to parameter variations among all realizations; thus they are good candidates for practical
implementation.
We discuss next a different minimal realization. called a balanced realization. However.
the discussion is applicable onl y to stable A. Consider
x = Ax + bll
(7. 30)
y = ex
It is assumed that A is stable or all its eigenvalues have negative real parts. Then the
controllability Gramian We and the observability Gramian Wo are, respectively, the unique
solutions of
(73 1 )
and
(7. 32)
They are positive defnite i f ( 7. 30) i s controllable and observable.
Di fferent mi ni mal realizations of the same transfer function have different controllability
and observability Gramians. For example, the state equation, taken from Reference [23],
X ~
4a
4/a
x .. lI
2 2a
y
=
[ 1 2/al x
(733)
for any nonzero a, has transfer function g(s) = (3.1 I S) /(s
2
+ 3s + I S) . and is controllable
and observable. Its controllability and observability Gramians can be computed as
0. 5
We =
0
and
, 0. 5
"0
=
0
(7 34)
We see that different a yields different mi ni mal realization and different controllability and
observability Gramians. Even though the controllability and observabi lity Gramians wi l l
change, their product remains the same as diag (0.25, for all a.
3. This section may be skipped without loss of continuity.
1 98 MI NIMAL REALI ZATI ONS AND COPRI ME FRACTIONS
� Theorem 7.5
Let (A. b, c) and (A, b, c) be minimal and equivalent and let We Wo and WcWo b the products of their
controllability and observability Gramians. Then We Wo and We 'Vo are similar and their eigenvalues
are all real and positive.
Proof' Let x = P, where P is a nonsingular constant matrix. Then we have
(7.35)
The controllability Gramian We and observability Gramian Wo of (A, b. c) are, respec
tively, the unique solutions of
and
A
'
Wo + WoA = C'c
Substituting A = PAPI and b = Pb into (7.36) yields
PAPI We We(P,)I A'P' = Pbb'P'
which implies
API We(pyl + PI We(p') I A' = bb'
Comparing this with (7. 3 1) yields
We = pI We(p' )
1
or We = PWeP'
Similarly, we can show
Thus we have
WeWo = pI ,Ve(p,)l p'Wop = pI WeWoP
(7.36)
(7.37)
(7. 38)
(7.39)
This shows that all We Wo are similar and. consequently, have the same set of eigenvalues.
Next we show that all eigenvalues of We Wo are real and positive. Note that both
We and Wo are symmetric, but their product may not be. Therefore Theorem 3.6 is not
directly applicable to We WOo Now we apply Theorem 3. 6 to We :
(7. 40)
where D is a diagonal matrix with the eigenvalues of We on the diagonal. Because We
is symmetric and positive defnite, all its eigenvalues are real and positive. Thus we can
express D as DI /2DI
/2, where D
I /2 is diagonal with positive square roots of the diagonal
entries of D as its diagonal entries. Note that Q is orthogonal or QI = Q'. The matrix
R = DI/2Q is not orthogonal but is nonsingular.
Consider RWoR' ; it is clearly symmetric and positive defnite. Thus its eigenvalues
are all real and positive. Using (7.40) and (3.66), we have
7. 4 Bal anced Real ization 1 99
(7.41 )
which implies that We Wo and RWoR' have the same set of eigenvalues. Thus we conclude
that all eigenvalues of We Wo are real and positive. Q.E.D.
Let us defne
(7.42)
where O¡ are positive square roots of the eigenvalues of We WOo For convenience, we arrange
them in descending order in magnitude or
These eigenvalues are called the Hankel singular values. The product We Wo of any minimal
realization is similar to 12•
� Theorem 7.6
For any ndimensional minimal state equation (A. b. c), there exists an equivalence transformation
.= P such that the controllability Gramian We and observability Gramian 'Vo of its equivalent state
equation have the property
(7.43)
This is called a balanced realization.
Poof: We frst compute We = R'R as in (7.40). We then apply Theorem 3. 6 to the real
and symmetric matrix RWoR' to yield
where U is orthogonal or U'U = I. Let
p
1
= R'U1

1/2 or P = 1
1
/2U'(R)
1
Then (7.38) and We = R'R imply
We = 1 1 /2U'(R,) I W
c
R
1
U1
1
/2 = 1
and (7.39) and RWoR' = U12U' imply
Wo = 11 /2U'RWoR' U1
I
; = 1
This establishes the theorem. Q. E. D.
By selecting a different P, it is possible to fnd an equivalent state equation with 'Ve = I
and Wo = 12• Such a state equation is called the inputnormal realization. Similarly. we
can have a state equation with We = 12 and 'Vo = I, which is called the outputnormal
realization. The balanced realization in Theorem 7. 5 can be used in system reduction. More
specifcally, suppose
200 MI NII\AL REALIZATI ONS AND COPRI I\\E F RACTI ONS
( 7, 4)
i s a balanced minimal realization of a stable .�( s) with
\Vc = \V" = diag( � I ' �2 )
where the A . b. and cmatrices are partitioned according t o the order of �, . If the Hankel
singular values of �I and �2 are disjoi nt. then the reduced state equation
XI = Al i xi + b i ll
( 7,45 )
is balanced and A I I is stabl e. If the singular values of �2 are much smaller than those of � ¡ ,
then the transfer function o f ( 7.5) wi l l be close to ,�( s) . See Reference [ 23] .
The MATLAB function ja" real wi l l transform (A. b. c) into a balanced state equation.
The reduced equation i n ( 7,45 ) can be obtained by using ba 1 red. The results i n this section are
based on the controllability and observability Gramians. Because the Gramians in the MIMO
case are square as i n the SISO case: all results i n this section apply to the MIMO case \\ ithout
any modifcation.
7. 5 Realizations from Markov Parameters'
Consider the strictly proper rational function
× f3 l s,, 1 + f2S,, 2 + . . . + f
,, 
I S + f"
g( s) =
. ^
Sll + a l s
"
1
+ 025'

 + . . . + an I S + a
"
We expand it into an infinite power series as
( 7. 47 )
If .�( s ) is strictly proper as assumed in ( 7.6) . then h (O) = O. The coefcients h em) . m =
I . 2 + .  . . are called Markm' parameters. Let g(t ) be the i nverse Laplace transform of .� ( 5 ) or.
equivalently. the impulse response of the system. Then we have
dill

I
h e m) = �( I )
d
t
lll

1 '
t=U
for m = I . 2. 3 . . . . . This method of computing Markov parameters is i mpractical because
it requires repetiti ve differentiations. and differentiations are susceptible to noise.' Equating
( 7,46) and ( 7'7) yields
. Thi: section may be :kipped without loss of continuity,
5. I n th� discretetime case, i f we apply 3n impuhe "equence I a .ystem. then the output ')eljuence di rectly yield�
'l'lrk(parame!ers. Thu. ;1<trkov parameters can t:..;il y be generated in discrere·tirne system'i.
1. 5 Real i zat i ons irom Markov Parameters
fl s
,,

1
+ f2s,,2 + . . . + f"
=
(5
"
+ 0
1
5,, 1 + 0
2
S,,2 + .  . ..0" ) I h l I )S
¡
+ h( 2) 5
2
+ .
From this equation. we can obtain the Markov parameters recursively as
h ( l ) = f
1
11 (2) = ¯ O ¡ h( l ) f2
h ( 3) =  0
1
11 ( 2)  Och ( l ) + /;
h Ill ) =  Ol h (1l  I I ¯ 0211 1 11 ¯ 2) ¯ . . .  0,, _ 1 11 ( I ) + {"
h im) =  Ol h ( m  I ) ¯ 02 h llll ¯ 2)  "  O , _ l h ( 1II  II ¯ I )
¯ 0
,,
11( 111  Il )
for m = Il + I . n + 2.
:ext we use the Markov parameters t o form the O X f matrix
h l I )
h ( 2)
T( O. () =
h
(
3 )
h (O)
h ( 2 ) h ( 3 )
h ( 3 ) 11 ( )
1( 4) 11 ( 5)
ina + l ) h (0 + 2 1
h l fi )
h( f + I )
hl f + 2)
il (a 
:
f  I )
201
( 7. 8)
( 7.9)
(7. 50)
It is cal led a Hankel matrix. It i s important to mention that even if h ( O) F O. 17 1 0) does not
appear i n the Hankel matrix.
Theorem 7.7
A stri ctly proper rational function g( s ) has degree Il if and unly if
pT(n. ll ) = pT( 1l ¬ k. 1l  l ) = Il for e\ ery k. 1 = 1 . 2 . . . . 1 7. 5 1 )
where p denotes the rank.
Proof: We first show that if deg .i(s ) = II. thenpT(ll. 1) = pT(1l + 1 . n ) ~ I,TI x. 11 ) .
[ f deg ,�( s ) = Il . then ( 7. 9) holds. and I l i s the smallest integer having the property.
Because 01' ( 7. 9) . the (II + I lth row of Ti ll + 1 . II I czn be written as a l inear combi nation
of the frst H rows. Thus we have pTI I1 . I/ ) = pTI I I + 1 . 1 ) . Again. because of (7.9). the
( n + 2)th row of T(1l + 2. 11 ) depends on its prev ious 11 rows and. consequently. on the
frst 1 rows. Proceeding forward. we can establish pTI I1 . I l ) = pT( x. I1 ) . :ow we cl ai m
pT(x. n) = n. If not. there would be an integer 11 < n having the property 1 7. 9) . This
contradicts the hypothesis that deg g( s ) = Il . Thus we hav e pT( I1 . Il ) = pT( X. I1 ) = 11 ~
Applying ( 7.49) to the columns of T yields 1 7. 5 I !.
Now we show that if ( 7. 5 1 ) holds. then g( s) = h ( I ).\ 1 + h ( 2) 5
2
+ ' ' can
be expressed as a strictly proper rational function of degree H. From the condition
202 MI NI MAL REALI ZATI ONS AND COPRI ME FRACTIONS
pT(n + 1 , 0) = pT(n, 0) = n, we can compute {a¡ , i = 2, . . . , n} to meet (7.49).
We then use (7.48) to compute {th ,
i = 1 , 2, . . . , n} . Hence we have
g(s) = h( l)sl + h(2)s2 + h(3)s3 . . .
!l s
n
1 + fhs
n
2 + . . . + !
n
I S + !
n
 sn + al snl + a
2
ss2 + . . . + a
n
I s + a
n
Because the n is the smallest integer having the property in (7. 51 ), we have deg g(s) = n.
This completes the proof of the theorem. Q.E.D.
With this preliminary, we are ready to discuss the realization problem. Consider a strictly
proper transfer function g (s) expressed as
Í
g(s) = h( l )sl
 h(2)s2 + h(3)s3 + . , .
If the triplet (A, b, c) is a realization of g(s), then
g(s) = c(sI  A)l b = c[s(I  sI AW1 b
which becomes, using (3.57),
g(s) = cbs
l
+ cAbs2 + cA2bs3 + . ,
.
Thus we conclude that (A, b, c) is a realization of g(s) if and only if
for m = 1 , 2, . . .
Substituting (7.52) into the Hankel matrix T(n, n) yields
cb
cAb
T(n, n) =
C�2b
cA
n
1 b
which implies, as shown i n (7.21),
cAb cA2b
cA
n
b
T(n, n) = oe
(7.52)
(7.53)
where 0 and e are, respectively, the n xn observability and controllability matrices of (A, b
.
c) .
Define
h(2)
h(3)
T(n, n) =
h
�
4)
hen +
h(3) h(4)
h(4) h(5)
h(5) h(6)
hen + 2) hen + 3)
hen +
hen + 2)
hen + 3)
h(2n)
(7.54)
It is the submatrix of T(n + n) by deleting the frst row or the sub matrix of T(n, n + I) by
deleting the frst column. Then as with (7.53), we can readily show
'
I
7. 5 Real izations from Markov Parameters 203
T(n , n) OAe (7. 55)
Using (7.53) and (7.55), we can obtain many different realizations. We discuss here only a
companionform and a balancedform realization.
Companion form There are many ways to decompose T(n. n) into OC. The simplest is
to select �= I or e = I. If we select 0 = I, then (7. 53) and (7. 55) imply e = T(n, n)
and �= T(n. n)T1 (n, n) .The state equation corresponding to 0 = I. e = T(n, n), and
A = T(n, nlT1 (n, n) is
0 I 0
0 0
X
�
0 0 0
an

Qn
l
an
2
y = [ I o 0 · · · 0 O]x
0
0
0
M]
0 h( l)
0 h(2)
]
<
+
[
hen I )
UI hen)
(7.56)
Indeed, the frst row of 0
_
= I and the frst column of e = T(n , n) yield the c and b i n (7.56).
Instead of showing A = Ten, n)T
1
(n, n). we show
AT(n, n) = T(n, n) (7.57)
Using the shifting property of the companionform matrix i n (7.56), we can readily verify
A
A
h (�
(7.58)
We see that the Markov parameters of a column are shifted up one position i f the column
is pre multiplied by A. Using this property, we can readily establish (7.57). Thus 0 = I,
e = T(n, n), and A = T(n. n)T1 (n. n) generate the realization in (7.56). It is a companion
form realization. Now we use (7.52) to show that (7. 56) is indeed a realization. Because of the
form of c, cAm b equals simply the top entry of Am b or
cb = h( l ) , cAb = h(2), cA"b = 1(3), . . .
Thus (7.56) is a realization of g(s). The state equation is always observable because 0 = I
has full rank. It is controllable if e = Ten, n) has rank n.
EXAMPLE 7.2 Consider
> 4s2 ¯ 25 ¬ 6
g(s) = �~ ~�~~
254 + 253 + 252 + 35 + 1
= 0 · 5
1
+ 2s2  3s3  2s4 + 255 + 3. 5s6 +
. . . (7.59)
We form T(4, 4) and compute its rank. The rank is 3; thus g(5 ) in (7.59) has degree 3 and its
204
MI NI MAL REALI ZATI ONS AND COPRI ME F RACTI ONS
numerator and denominator have a common factor of degree I. There i s no need to cancel frst
the common factor in the expansion in (7.59). From the preceding derivation. we have
and
A = [ �3 ��
�
�
2 � 3. 5 3
2
3
2
b = [ 0 2  3]'
3 I
[
0 I 0
]
2 = 0 0 I
2 0. 5  I 0
(7.60)
c = [ I 0 0]
The triplet (A. b. c) is a minimal realization of g(s) in (7. 59).
=
We mention that the matrix A in (7.60) can be obtained without computing T(n. n)
TI (n . n) . Using (7.49) we can show
T(3, 4)a := [ �
3
2 3
3 2
2 2
Thus a is a null vector of T(3, 4). The MATLAB function
t = [ O 2  3  2 ; 2  3 2 2 ;  3 2 2 3 . 5 ] ; a= :ul l ( t )
= 0
yields a= [ 0. 3333  0.6667 0. 0000  0.6667]'. We normalize the last entry of a to I by
typing a / a ( 4 ) . which yields [0. 5 I 0 I ]'. The frst three entries. wi th sign reversed. are the
last row of A.
Balanced form Next we discuss a different decomposition of T(n. n)
yield a realization with the property
cc = 0
'
0
First we use singularvalue decomposition to express T( n. n) as
oc. which will
(7. 61 )
where K and L are orthogonal matrices and A 1 . 2 i s diagonal with the singular values of T(n, n)
on the diagonal. Let us select
Then we have
0 = KAI !! and C = AI /'L'
0
1
= A 1 /2K' and C
I
= LA  I /!
For this selection of C and 0, the triplet
A = OI t(n. n) CI
b = first column of C
c = frst row of 0
(7. 62)
(7. 63)
(7.64)
(7. 65)
(7. 66)
7. 6 Degree of Transfer Matri ces 205
is a minimal realization of g( s) . For this realization. we have
cc = A 1 12L'
LA1 12
= A
and
Thus it i s called a balanced reali:ation. This balanced realization is di fferent from the balanced
realization discussed i n Section 7.4. It is not clear what the relationships between them are.
EXAMPLE 7.3 Consider the transfer function in Example 7.2. NO\\' we will find a balanced
realization from Hankel matrices. We type
t = ' C 2  3 ; 2  3  2 ;  3  2 2 J ; t c = [ 2  3  2 ;  3  2 . ¦ ~ .
[ k, s , l j s'Id ( : l ;
s l = sq�t ( 5 ) ;
O= k* s l ; C= s : * 1 ' ;
a = i �v ( O) * t : * i �v ( C) ,
D= [ C ( i , 1 ) ; C ( 2 , : ) ; C ( 3 , l ) j , c= , O ( 1 , c 0 ( 1 , 2 ) 0 ( 1 , 2 1 ]
This yields the following balanced realization:
[
0.4003 1 .0024
x = 1 . 0024 0.3 1 21
0.4805 0. 3209
0.4805
]
[
1 . 2883
0.3209 x + 0. 7303 l!
0. 0882 1 . 06 1 4
V ~ [ 1 .2883 0.7303  1 . 06 14]x + 0 · h
To check the correctness of this result. we type [ :", c 1 ¯ s s2 t f ( a , b , C , 0 ) • which yields
,
25  3
g(s) =
ó + 5 + 0. 5
Thi s equals g(s) i n 0. 59) after canceling the common factor 2(5 + I ) .
7. 6 Degree of Transfer Matrices
This section wi ll extend the concept of degree for scalar rational functions to rational matrices.
Gi ven a proper rational matrix G(s), it is assumed that every entry of G( s) is a coprime
fraction: that is. its numerator and denominator have no common factors. This will be a standing
assumption throughout the remainder of this text.
�
Defnition 7.1 The characteristic polynomial of a proper rational matrix G(s) is defned
as the least common denominator of all minors ofG(s). The degree of the characteristic
polynomial is defned as the McMillan degree or, simply, the degree of G( S) and is
denoted by 8G(s).
206 MI NIMAL REALIZATIONS AND COPRI ME FRACTI ONS
EXAMPLE 7.4 Consider the rational matrices
¨.
s +1
The ma�x G1 ( s) has 1 / (s + 1 ) , 1 / (s + 1 ) , 1 / (s + 1 ) , and 1 / (s + 1 ) as its mino
!
s of order I ,
and det G1 ( s) = 0 as its minor of order 2, Thus the characteristic polynomial of G1 (s) i s s + I
and 8G1 (s) = L The matrix G
2
(S) has 2/(s + 1 ) , 1 / (s + 1 ) . 1 /(s + 1 ) . and 1 / (s + 1 ) as its
minors of order 1 . and det Go(s) = 1 / (s + 1 ) 2 as its minor of order 2, Thus the characteristic
polynomial of G
2
(s) is (s + 1 )
2
and 8G
2
(S) = 2,
From this example, we see that the characteristic polynomial of G(s) is. in general,
diferent from the denominator of the detinant of G(s) [if G(5) is square] and different
from the least common denominator of all entries of G(s) ,
EXAMPLE 7,5 Consider the 2 x 3 rational matrix
G(s) =
s I
1
s + 1
(0 + l ) (s + 2)
1
(s + 1 ) (0 + 2)
Its minors of order 1 are the six entries of G(s), The matrix has the following three minors of
order 2:
s i
S
+ 1
(s + 1 )2(s + 2)
+
(s + 1 )
2
(s + 2) (s + 1 ) 2(s + 2) (s + I ) (s + 2)
s 1 1 s + 4
¯,  +
s + 1 s (s + l ) (s + 3) (s + I ) (5 + 3)
I 1 3
(s + l ) (s + 2)s (s + l ) (s + 2) (s + 3) s (s + l ) (s + 2)(s + 3)
The least common denominator of all these minors i s s (s + 1 ) (s + 2) (s + 3), Thus the degree
of G(s) is 4, Note that G(s) has no minors of order 3 or higher.
In computing the characteristic polynomial. every minor must be reduced to a coprime
fraction as we did in the preceding example; otherwise. we will obtain an erroneous result.
We discuss two special cases, If G(s) is 1 x p or q x l , then there are no minors of order
2 or higher. Thus the characteristic polynomial equals the least common denominator of all
entries of G(s), In particular, if G(s) is scalar. then the characteristic polynomial equals its
denominator. If every entry of q x p G(s) has poles that differ from those of aU other entries,
such as
G(s) =
(
S
+
��
+ 2)
s + 3
then its minors contain no poles with multiplicities higher than those of each entry, Thus the
characteristic polynomial equals the product of the denominators of aU entries of G(s) ,
7. 7 Mi ni mal RealizationsMatrix Case 207
To conclude this section, we rentio
�
two important properties. Let (A. B. C. 0) be a
controllable and observable realization of G( s). Then we have the followinc'
c:
• Monic least common denominator of all minors of G(s) = characteristic polynomial of A.
• Monic least common denominator of all entries of G(s) = minimal polynomial of A.
For their proofs, see Reference [4, pp. 302304].
7, 7 Mi ni mal Real izationsMatrix Case
We introduced in Section 7. 2. 1 minimal realizations for scalar transfer functions. Now we
dISCUSS the matrix case.
,. Theorem 7. M2
A state equation (A. B. C. 0) is a minimal realization of a proper rational matrix G(s) if and only if
(A, B) IS controllable and (A, C) l5observable or if and only if
dim A = deg G(s)
Proof: The proof of the frst part is similar to the proof of Theorem 7. 2. If (A. B) is not
controUable or if (A, C) is not observable, then the state equation is zerostate equivalent
to a lesserdImensIOnal state equation and is not minimal. If (A. B. C, 0) is of dimension
n and
_
Is controUable and observ
"
ble, and if the idimensional state equation (A, E. C, D),
WIth n < n. is a realization of G( s), then Theorem 4. 1 implies 0 = D and
CA
m
B = CA
m
E for m = 0, 1 , 2 . . . .
Thus. as in (7. 22) . we have
OC = On en
Note that O. C, On , and c are, respectively, nq x n . n x np, nq x i, and i x np matrices.
USIng the Sylvester inequality
p( O) + p( C)  n :: p( OC) :: min( p(O), p( C»)
which is proved in Re!er�nce [6, p. 3 1 ] ; and p( O) = p( C) = 11 , we have p( OC) = n.
SimIlarly. we have p( On Cn ) = i < n . This contradicts p( OC) = p( On Cn ) . Thus every
controllable and observable state equatIon is a minimal realization.
To show that ( A. B, C, 0) is minimal if and only if dim A = deg G(s) is much more
complex and WIU be established in the remainder of this chapter. Q.E.D.
.� Theorem 7.M3
All minimal realizations of G(s) are equivalent.
208 MI NLy\AL REALIZATI ONS AND COPRI ME F RACTIONS
Proof: The proof follows closely the proof of Theorem Let (A, B, C, D) and (A. B,
C, D) be any two ndimensional minimal realizations of a .X .proper rational matrix
c(.) As in ·and · we have
OC = OC
and
OAC = OAC
In the scalar case, O. C. O. and C are all n x n nonsingular matrices and their inverses are
well defned. Here 0 and 0 are .x matrices of rank n: C and C are x .matrices
of rank /1. They are not square and their inverses are not defned. Let us defne the 11 X nq
matrix
Because 0' is n x .and 0 i s nq X the matrix 0' 0 i s x n and is, following Theorem
nonsingular. Clearly. we have
Thus O� is called the .....or ·.of O. Note that 00+ i s .x nq and
does not equal a unit matri x. Similarly. we defne
It i s an .X n matrix and has the property
CC+ = CC' ( CC' )  I = I
Thus C� is called the ....or ..of C. In the scalar case. the
equivalence transformation is defned in · as P = 0
1
0 = CCI . Now we replace
inverses by pseudoinverses to yield
P : = 0+0 = ( 0' 0)
 1
0' 0
= Cc+ = CC( CC' ) I
·
This equality can be verifed directly by premultiplying ( 0' 0) and postmu�tiplyin
�
( CC)
and then using The inverse of P in the scalar case i s p I = 0 1 0 = C C I . In
the matrix case, it becomes
p I : = 0+ 0 = ( 0' 0)  1 0' 0
= CC' = CC' ( CC' ) I
This again can be verifed using From OC = Oc. we have
C = ( O' O) I O' OC = PC
0 = OCC' ( CC')  I = Op I
Their frst .columns and frst .rows are B = PB and C
OAC = OAC implies
CPI . The equation
7. 8 Matri x Polynomi al Fracti ons 209
This shows that all minimal realizations of the same transfer matrix are equivalent.
Q. E. D.
We see from the proof of Theorem that the results i n the scalar case can be extended
directly to the matrix case if inverses are replaced by pseudoinverses. In MATLAB, the function
pinv generates the pseudoinverse. We mention that minimal realization can be obtained from
nonminimal realizations by applying Theorems and or by calling the MATLAB
function minreal. as the next example illustrates.
�
EXAMPLE 7.6 Consider the transfer matrix in Example or
c s· =
�
.
�
·
·.+
�
.+ 2) .+ ·
Its characteristic polynomial can be computed as ·.+ l ) .+ ·
Thus the rational matrix has
degree Its sixdimensional realization in and fourdimensional realization in
.
are clearly not minimal realizations. They can be reduced to minimal realizations by calling
the MATLAB function ::r:eal . For example. for the realization i n typing
a= [ 4 . 5 0 ¬ C 0 ~. : ; 0  4 . 5 0 6 0  2 ; 1 Ü 0 0 0 0 ; . . .
0 1 0 0 0 8 ; 0 : 1 0 0 0 ; 0 0 G l 0 e i ;
b" [ 1 0 ; 0 1 ; L 0 ; 0 . ; . 0 ; 0 0 i ;
C= [

Ü ´ ¯ 2 4 7 . 5 ¬ . � ´ ¦ 0 1 O . 5 l . 5 1 O . 5 I ; d= [ 2 0 ; > G I ;
[ am, br" C¯, d] "r.l:_oeal ( a , b, C . d)
yields
·
x = ··
·
¯
·
· x +
V =
•
·
·
x
·
·
+
¯
U
Its dimension equals the degree of c(. : thus it is controllable and observable and i s a minimal
realization of G( s) in
7, 8 Matri x Pol ynomi al Fracti ons
The degree of the scalar transfer function
• ·.
.. = ¯= · .
. =
. ·.
.
i s defned as the degree of . i f ·.and . are coprime. It can also be defned as
the smallest possible denominator degree. In this section, we shall develop similar results for
21 0
MI NIMAL REALIZATIONS AND COPRI ME FRACTI ONS
transfer matrices. Because matrices are not commutative. their orders cannot be altered in our
discussion.
Every q X p proper rational matrix G(s) can be expressed as
(7.76)
where N (s) and D(s) are q x p and p X p po!
ynomial matrices. For example, the . x 3 rational
matrix in Example 7. 5 can be expressed as
• s
G(s) ~
1
o
(s + l ) (s + .
o
o _
1
sis �3)
(7.77)
The three diagonal entries of D(s) in (7.77) lare the least common denominators of the three
columns of G(s) . The fraction in (7. 76) oi (7.77) is called a right polynomial fractIOn or,
simply, right fraction. Dual to (7.76), the expression
G(s) = 01 (s)
N
(s)
where
D
(s) and
N
(s) are q X q and q X p polynomial matrices, is called a lef polynomial
fraction or, simply, leffraction.
Let R(s) be any p X p nonsingular polynomial matrix. Then we have
G(s) ~ [N(s)R(s)] [D(s)R(sW
I
~ N(s)R(s)R
1
(s)D
1
(s) = N(s)D
1
(s)
Thus right fractions are not unique. The same holds for left fractions. We introduce in the
following right coprime fractions.
. '
Consider A(s) = B(s)C(s), where A(s), B(s), and CIS) are polynomials of compatible
orders. We call C(s) a right divisor of A(s) and A(s) a left multiple of Cis) . Similarly, we call
B(s) a lef divisor of A(s) and A(s) a right multiple of B(s).
Consider two polynomial matrices D(s) and N(s) with the same number of columns p.
A p X p square polynomial maax R(s) is
.
called a common right divisor of D(s) and N(s) If
there exist polynomial matrices D(s) and N(s) such that
D(s) =
D
(s)R(s) and N(s) = N(s)R(s)
Note that D(s) and N (s) can have different numbers of rows.
Defnition 7.2 A square polynomial matrix M(s) is called a unimodular matrix i its
determinant is nonzero and independent of s.
The following polynomial matrices are all unimodular matrices:
.
.
. s2 + S + 1
. S
IO
+S + I
,
S S + I
s + 1 0 3 s  1 s
Products of unimodular matrices are elearly unimodular matrices. Consider
det M(s) det MI (s) = det [M(s)M1 (s)] = det I = 1
7.8 ,\latrix Pol Vnomi al Fractions 2 1 1
which implies that i f the determinant of M(s) is a nonzero constant, so i s the determinant of
MI (s) . Thus the inverse of a unimodular matrix .1(5) is again a unimodular matrix.
Defnition 7.3 A square polynomial matrix R\ s) is a geatest common right divisor
(gcrdJ ofD(s) and N(s) i(i) R(s) is a common right dil'isor ofD(s) and N(s) and (iiJ
R(s) is a left multiple ofevery common right dil'isor ofD(s) and N(s). I U gcrd is a
unimodular matrix, then D(s) and N (s) are sid U be right coprime.
Dual to this defnition, a square polynomial mat
R( s) is a greatest common left divisor
(geld) of
D
(s) and N(s) if (i)
R
(s) is a common left divisor of
D
(s) and
N
(s) and (ii)
R
(s)
is a right multiple of every common left divisor oi
D
(s) and
N
(s). If a geld is a unimodular
matrix, then
D(s) and
N
(s) are said to be lefcoprime.
Defnition 7.4 Consider a proper rational matrix G(s) jactored as
G(s) ~ N(s)D
1
(s)  01 (s
8
(s)
where N(s) and D(s) are right coprime, and :' 5 ) and
D
(s) are lef coprime. Ten the
characteristic polynomial ofG(s) is defned as
det D(s) or det 0( 5)
and the degree o/G(s) is defned as
deg G(s) = deg det DIS ) = deg det D(s)
Consider
G(s) = N(s)D
I
(s) ~ [N(s )R( s)] [D(s)R(sWI
(7.78)
where N(s) and D(s) are right coprime. Defne D: I s)  D(s)R(s) and N
I
(s) ~ N(s)R(s).
Then we have
det DI (s) = det[D(s)R(s l ] = det D( s) det R(s)
which implies
deg det D
I
(s) ~ deg d
�
t DI s)  deg det R(s)
Clearly we have deg det D
1
(s) :: deg det D(s) and the equality holds if and only if R(s) is
unimodular or, equivalently, N
I
(s) anc D
I
(s) are right coprime. Thus we conclude that if
N (s )D1 (s) is a coprime fraction, then D(s) has the smallest possible determinantal degree
and the degree is defned as the degree of the transfer matrix. Therefore a coprime fraction can
also be defned as a polynomial matrix fraction with the smallest denominator' s determinantal
degree. From (7.78), we can see that coprime fractions are not unique: they can difer by
unimodular matrices.
We have introduced Defnitions 7. 1 and 7.4 to defne the degree of rational matrices. Their
21 2 MI NI MAL REALI ZATI ONS AND COPRIME FRACTI ONS
equivalence can be established by using the SmithMcMillan form and will not be discussed
here. The interested reader is referred to, for example, Reference [3].
7. B. 1 Col umn and Row Reducedness
In order to apply Defnition 7.4 to determine the degree of G(s) = N( s)DI (s), we must
compute the determinant of D( s). This can be avoided if the coprime fraction has some
additional property as we will discuss next.
The degree of a polynomial vector i s defned as the highest power of s i n all entries of
the vector. Consider a polynomial matrix M(s) . We defne
8Ci M(s) = degree
?
f ith column of M(s)
8ri M(s) = degree �f ith row of M(s)
and ca1l 8e; the column degree and Ori the row degree. For example, the matrix
M(s) =
s + 1
s  1
S3  2s + 5
,
s
has Oc! = 1 , Oc2 = 3, 0c
3
= 0, Orl = 3, and Or2 = 2.
Defnition 7.5 A nonsingular polynomial matrix M( s) is column reduced if
deg det M(s) = sum of all column degrees
It is row reduced i
deg det M(s) = sum ofall rw degrees
A matrix can be column reduced but not row reduced and vice versa. For example. the
matrix
3s2 + 2s
M(s) = ,
3 s + s 
(7. 79)
has determinant s
3
s2 5s + 3. Its degree equals the sum of its column degrees 2 and I. Thus
M(s) in (7.79) is column reduced. The row degrees of M(s) are 2 and 2; their sum is larger
than 3. Thus M(s) is not row reduced. A diagonal polynomial matrix is always both column
and row reduced. If a square polynomial matrix is not column reduced, then the degree of its
determinant is less than the sum of its column degrees. Every nonsingular polynomial matrix
can be changed to a column or rowreduced matrix by pre or postmultiplying a unimodular
matrix. See Reference [6, p. 603].
Let oCi M(s) = kCi and defne Hc (s) = diag(sk" . Sk'2 , . . . ). Then the polynomial matrix
M(s) can be expressed as
M(s) = MhcHc (S) + Mlc(s) (7. 80)
For example. the M(s) in (7.79) has column degrees 2 and I and can be expressed as
M(S) =
�
7. 8 Matrix Pol ynomial F racti ons
2
[
S
2
0
[
+
2s I
`
l O s s  3 0
21 3
The constant matrix Mhe is called the columndegree coefcient matrix; its i t h column i s the
coefcients of the ith column of M(s) associated with Sk" . The polynomial matrix Mie(s)
contains the remaining terms and its ith column has degree less than k
Ci ' If M(s) is expressed
as in (7.80), then it can be verifed that M(s) is column reduced i fand only if its columndegree
coeffcient matrix Mhc is nonsingular. Dual to (7.80), we can express M(s) as
M(s) = Hr (s)Mhr + Mlr (s)
where 8" M(s) = k" and Hr (s) = diag(sk" . Sk,2 , . . . ). The matrix Mhr is called the rowdegree
coefcient matrix. Then M(s) is row reduced if and only if Mhr is nonsingular.
Using the concept of reducedness, we now can state the degree of a proper rational matrix
as follows. Consider G(s) = N(s)D I (s) = DI (s)N(s), where N(s) and D(s) are right
coprime, N(s) and D(s) are left coprime, D(s) is column reduced, and D(s) is row reduced.
Then we have
deg G(s) = sum of column degrees of D(s)
sum of row degrees of D(s)
We discuss another implication of reducedness. Consider G(s) = N(s)D1 (s). If G(s)
is strictly proper, then oei N(s) < 8ci D(s) , for i = 1 , 2, . . . , p; that is, the column degrees
of N(s) are less than the corresponding column degrees of D(s). If G(s) is proper, then
ooN(s) S 0ci D(S), for i = 1 , 2, . . +
,
p. The converse, however, is not necessarily true. For
example. consider
Although OCi N (s) < OCi D(s) for i = I , 2, N (5 )DI (5) is not strictly proper. The reason is that
D(s) is not column reduced.
Theorem 7.8
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 matrix N(s)D1 (5) is proper (strictly proper) if and only if
for i = I . 2 + . . . , p.
Proof The necessity part of the theorem follows from the preceding example. We show
the suffciency. Following (7.80), we express
D(s) = DhcHc(s) + Dlc (s) = [Dhc + Dlc(s)H;1 (s)]Hc (s)
N(s) = NhcHe (s) + NicCs) = [Nhe + N1c (s)H;I (S)]Hc(s)
Then we have
21 4
MI NIMAL REALIZATIONS AND COPRI ME F RACTI ONS
�
G(s) := N(s)D
l
(S) = [Nhe + NIc(s)H;1 (S) ] [Dhe + Dle(s)H;1 (sWI
Because Die ( s )H; 1 (s) and N/e (s )H; 1 (s) both approach zero as s ¬ 0, we have
lim G(s) = NheD;
s�>
where Dhe is nonsingular by assumption, Now if De;N(f) ::Dei D(S), then Nhe is a nonzero
constant matrix. Thus G( () is a nonzero constant and G(s) is proper. If De; N(s) < De; D(s),
then Nhe is a zero matrix. Thus G(o) = 0 and G(s) is strictly proper. This establishes
the theorem. Q.E.D.
We state the dual of Theorem 7. 8 without proof.
� Corollary 7.8
Let N (s) ad D(s) b q x p and q x q polynomial matrices, and let D(s) be row reduced. Then
the rational matix D1 (s)N(s) is proper (strictly proper) if and only if
for i = I , 2, . . . , q.
7. 8. 2 Computing Matrix Copri me Fractions
Given a right fraction N(s)D1 (s), one way to reduce it to a right coprime fraction is to
compute its gcrd. This can be achieved by applying a sequence of eltmentary opera
�
1
1
0ns.
Once a gcrd R(s) is computed, we compute N(s) = N(s)Rl (s) and D(s) = D�s)R (s) .
Then N(s)Dl (s) = N(s) Dl (s) and N(s)D1 (s) is a right coprime fraction. If D( s) is not
column reduced, then additional manipulation is needed. This procedure will not be discussed
here. The interested reader is referred to Reference [6, pp. 590591 ) .
We now extend the method of computing scalar coprime fractions in Section 7. 3 t o the
matrix case. Consider a q x p proper rational matrix G(s) expressed as
G(s) = Dl (s)N(s) = N(s)Dl (s)
(7. 81 )
I n this section, we use variables with an overbar to denote left fractions, without an overbar to
denote right fractions. Clearly (7. 81 ) implies
N(s)D(s) = D(s)N(s)
and
D(s) ( N(s» + N(s)D(s) = 0
(7.82)
We shall show that given a left fraction Dl (s)N(s). not necessarily left coprime, we can
obtain a right coprime fraction N(s)D1 (s) by solving the polynomial matnx equation m
(7.82). Instead of solving (7.82) directly, we shall change it into solving sets of Imear algebraiC
equations. As in (7.27), we express the plynomial matrices as, assuming the highest degree
to be 4 to simplify writing,
7.8 Matrix Polynomi al Fractions
D(s) = Do + DIS + D2S
2
+ D3S
3
+ D4S
4
N(s) = No + N
l
s + N2S
2
+ N3S
3
+ N4S
4
D(s) = Do + D
I
S + D2S
2
+ D3S
3
N(s) = No + Nls + N2S
2
+ N3S
3
21 5
where D; , N; , D; , and N; are, respectively, q x q, q x p, p x p, and q x p constant matrices.
The constant matices D; and N; are known, and 0; and NI are to be solved. Substituting these
into (7. 82) and equating to zero the constant matrices associated with sk, for k = 0, 1 , + . ~
we obtain
Do No 0 0 0 0 0 0

No
D
l
.
Nl Do No
Do
0 0 0 0
D2 N2 D
l
Nl Do No 0 0

Nl
D3 N3 D2
N2 Dl Nl Do No
Dl
SM : = = 0 (7. 83)
D4 N4 D3 N3 D2 N2 Dl Nl N2
0 0 D4 N4 D3 N3 D
2
N2
D2
0 0 0 0 D4 N4 D3 N3 
N3
0 0 0 0 0 0 D4 N4 D3
This equation is the matix version of (7.28) and the matix S will be called a generali�ed
res
u
ltant. Note that every Vblock column has q Vcolumns and every N block column has p
N columns. The generalized resultant 8 as shown has four pairs of V and N block columns:
thus it has a total of 4(q + p) columns. It has eight block rows; each block rows has q rows.
Thus the resultant has 8q rows.
We now discuss some general properies of8 under the assumption that linearly indepen
dent columns of 8 from left to right have been found. It turs out that every Vcolumn in every
Vblock column is linearly independent of its lefhandside (LHS) columns. The situation for
N columns, however, is diferent. Recall that there are p N columns in each N block column.
We use Ni column to denote the i th Ncolumn in each Nblock column. It turs out that if
the Ni column in some Nblock column is linearly dependent on its LHS columns, then all
subsequent Ni columns, because of the repetitive structure of8, will be linearly dependent on
its LHS columns. Let Il , i = 1 , 2, .  . , p, be the number of linearly independent Ni columns
in 8. They are called the column indices of G(s). The frst Ni column that becomes linearly
dependent on its LHS columns is called the primar dependent N i col
u
mn. It is clear that the
(I; + l )th N i column is the primary dependent column.
Corresponding to each primar dependent N i column, we compute the monic null vector
(its last entry equals 1 ) of the submatrix that consists of the primary dependent N i column and
all its LHS linearly independent columns. There are totally p such monic null vectors. Frm
these monic null vectors, we can then obtain a right fraction. This fraction is right coprime
because we use the least possible I; and the resulting D(s) has the smallest possible column
21 6 MI NIMAL REALIZATI ONS AND COPRI ME FRACTI ONS
degrees. In addition, .automatically will be column reduced. The next example illustrates
the procedure.
EXAMPLE 7.7 Find a right coprime fraction of the transfer matrix in or
. .
�
.
�
·
.
·. .·
First we must fnd a left fraction. not necessarily left coprime. Using the least common
denominator of each row. we can readily obtain
.
·
o · .: ·
· ·
¬ .¡ ¬
X
. ·.
 .·.
Thus we have
and
·
.
o ..
 :s
·
.
. 25 · .
·
.· ·.
.I
We form the generalized resultant and then use the QR decomposition discussed in Section
to search its linearly independent columns in order from left to right. Because it is simpler
to key in the transpose of S, we type
dl = [ 2 0 o 2 0 U Ü J ; d2 ¯ [ 0 ¬ 0 . U ¬ ' 2 J ;
�1 = [  2 0 2  2 l � 0 0 O ' ; �2 = [ 2 1
s = [ dl 0 0 0 0 ; d2 0 0 C O ; � 1 0 0 0 ; n2 . 0 0 0 ; . . .
o 0 dl 0 0 ; 0 Ü d2 0 0 ; 0 0 nl . 0 ; 0 \ n2 0 0 ; . . .
0 0 0 0 dl ; O . 0 0 d2 ; O 0 0 0 nl ; O C O O �2 J ' ;
[ q, r J =qr l s )
We need only r : therefore the matrix q will not be shown. As discussed in Section
we need to know whether or not entries of r are zero in determining linear independence of
columns: therefore all nonzero entries are represented by x, .and ni . The result is
7. 8 Matri x Pol ynomi al F racti ons 2 1 7
. x x x x x x x x x
.· x x x x x x x x x
n l x x x x x x x x x
x x x x x x x x
. x x x x x x x
.· x x x x x x
x x x x x
x x x
. l x x
.2
We see that all Dcolumns are l i nearly independent of their LHS columns. There are two
linearly independent
N
I columns and one linearly independent
N
2column. Thus we have
Jl ·and J2 The eighth column of S is the primary dependent
N
2column. We
compute a nul l vector of the submatrix that consists of the primary dependent
N
2column and
all its LHS linearly independent columns as
= 2 = nul l ( [ d1 0 0 ; d2 0 O ; �l 0 J ; n2 . 0 ; . . .
o 0 dl ; 0 0 d2 ; 0 0 nl ; 0 '. n2 J ' ) ;
and then nonnalize the last entry to by typing
z2b= z 2 / z 2 ( 8 )
which yields the frst monic null vector as
z2 := [ 7  1 1 2  4 0 2 1 J '
The eleventh column of S is the primary dependent
N
I column. We compute a null vector
of the submatrix that consists of the primary dependent
N
I column and all its LHS linearly
independent columns (that is, deleting the eighth column) as
¨ i ='' . 1 1 ¦ � j: L Ù . O ; d2 0 0 0 ,J ; r. l 0 0 0 0 ; :1 2 0 0
o d: 0 0 ; 0 0 d2 0 0 ; 0 0 �� 0 0 ; . . .
c e o 0 dl ; O 0 0 0 d2 ; J . ¯ �� · ¡
and then nomlal ize the last entry to I by typing
: lb= zL z : r : o )
which yields the second monic null vector as
z �b= [ l O ¬ 0 . 5 1 0 ] 0 2 . 5 2 0 1 1 '
0 ; . . .
21 8 MI NIMAL REALIZATIONS AND COPRIME FRACTI ONS
Thus we have

n61

n6
2
. ·

n5
1
·
¸
. I
·
.
0
.
0
I I
.
0
.
0
. ·
.
n¸
·
I ·
· ·
¯¯
. . ·
·.

.
.
· ·
! ' \
I
.
.
]
·

nil 
n
f
·

n,
_
n�
z
.
.
.
.
.
where we have written out Ni and  explicitly with the superscripts ij denoting the ijth entry
and the subscript denoting the degree. The two monic null vectors are aranged as shown. The
order of the two null vectors can b interchanged, as we will discuss shortly. The empty entries
are to be flled up with zeros. Note that the empty ent at the . X
I
)th location is due to
deleting the second i linearly dependent column in computing the second null vector. By
equating the corresponding enties in ·. we can readily obtain
and


� �
·
�
¯
.
� �
.
·. =
.
.
·. I
'
. ·
·
I
.
=
·
�
.
. ·
·
Thus .. in ·.· has the following right coprime fraction
•
·
. · ·. ·
·. .
..
. I .
·.
. ·
·.
The . in ·. is column reduced with column degrees
= · and
= Thus we have
deg det . = : I = · and the degree of .. in ·.· is · The degree was computed in
Example · 8 3 by using Defnition ·
7. 8 Matrix Pol ynomi al Fractions 21 9
In general, i f the generalized resultant has linearly independent Hi columns, then
. computed using the preceding procedure is column reduced with column degrees fi ·
Thus we have
deg .. = deg det . = _
= total number of linearly independent H columns in S
We next show that the order of column degrees is immaterial. In other words, the order of
the columns of · . and . can be changed. For example, consider the permutation matrix
 = �� �
. .
and . = .· and ·. = ·. Then the frst, second, and third columns of . and
· . become the third, frst, and second columns of . and N . However, we have
.. = ·.
. = ·..·.
= ·.
This shows that the columns . and ·. can arbitrarily be permutated. This is the same as
permuting the order of the null vectors in · .· Thus the set of column degrees is an intrinsic
property of a system just dthe set of controllability indices is (Theorem · What has been
discussed can be stated as a theorem. It is an extension of Theorem ·· to the matrix case.
Theorem 7.M4
Let .. = D
l
.·. be a left fraction, not necessarily left coprime. We use the coeffcient matices
of D. and ·. to form the generalized resultant S shown in . and search its linearly lndependent
columns from left to right. Let i = · , p. be the number oflinearly independent N i columns.
Then we have
deg .. =
. .
and a right coprme fraction · . 
. can le obtained by computing p monic null vectors of the p
matrices formed fromeach primary dependent N i column and all its LHS linearly independent columns.
The right coprime fraction obtained by solving the equation in ·. has one additional
important propery. After permutation, the columndegree coefcient matrix . can always
become a unit upper triangular matrix (an upper triangular matrix with I as its diagonal entries).
Such a . is said to be in column echelon form. See References [6. pp. .
· pp.
·.··.· For the . in ·. its columndegree coefficient matrix is
. =
�
�
It is unit upper triangular; thus the . is in colu� echelon form. Although we need only
column reducedness in subsequent discussions, if . is in column echelon form, then the
result in the next section will be nicer.
Dual to the preceding discussion, we can compute a left coprime fraction from a right
fraction ·. . which is not necessarily right coprime. Then similar to ·.· we form
220 MI NI MAL REALI ZATI ONS AND COPRI ME F RACTI ONS
[

No lo :  N
I
11
 N2 12 :
 N3 13J T = 0 (7. 88)
with
Do D
I
D2 D3 D4 0 0 0
N
o N
I
N2 N3 N4 0 0 0
0 Do D
I
D2 D3 D4 0 0
0 No N
I
N2
N3 N4 0 0
T
' 
(7. 89)
0 0 Do
D
I
D2 D3 D. 0
0 0 No "
N
I
Nz N3 N4 0
0 0 0 Do D
I
D2 D3 D4
0 0 0 No N
I
N2 N3 N.j
We search l i nearly independent rows in order from top to bottom. Then all Drows are linearly
independent. Let the Nirow denote the ith Nrow in each N blockrow. If an Nirow becomes
l inearly dependent, then all Nirows in subsequent N blockrows are l i nearly dependent on
their preceding rows. The frst Nirow that becomes linearly dependent i s called a primary
dependent Nirow. Let Vj , i = 1 , 2, . . . , q, be the number of linearly independent Nirows.
They are called the DW indices of G(s) . Then dual to Theorem 7.M4, we have the following
corollary.
... Corollary 7.M4
Let G(s) = N(s)D1 (5) be a right fraction. not necessarily rght coprime. We use the coeffcient
matrces of D(s) and N(s) to form the generlized resultant T shown in (7. 89) and search its linearly
independent rows in order from top to bottom. Let v, . ior i = 1 . 2 + . . . + q. be the number oi linearly
independent Nirows in T. Then we have
deg G(s) = VI + \³ + . . . + Vq
and a left coprime fraction 11 (s) N (5) can be obtained by computing q monic left null vectors of the
q matrices formed from each primary dependent Nirow and all its preceding linearly independent rows.
The polynomial matrix 1(5) obtained in Corollary 7.M4 i s row reduced with row degrees
{ Vj . i = 1 . 2 . . . . . q} . In fact, it is in row echelon form; that is, its rowdegree coeffcient
matrix, after some row permutation. i s a unit lower triangular matrix.
7 . 9 Real izations from Matrix Copri me Fractions
Not to be overhelmed by notation. we discuss a realization of a 2 x 2 strictly proper rational
matrix G(s) expressed as
G(s) = N(s)D
1
(s) (7. 90)
7. 9 Realizations from Matrix Copri me Fractions 221
where N (s) and D(s) are right coprime and D(s) is in column echelon form.6 We further assume
that the column degrees of D(s) are I
I
= 4 and 12 = 2. First we define
s
�'
H(s) : = 0
0 S
4
S�2
=
0 s�
'
(7. 91 )
and
S�, I 0
s
3
0
5
2
0
0 5 0
L(s) : =
0 S
J2

1
0
(7. 92)
0 5
0
0
The procedure for developing a realization for
yes) = G(s)u(s) = N(s)D
l
u(s)
follows closely the scalar case from (7. 3) through (7. 9). First we introduce a new variable v(t)
defned by vis) = D
I
(s)u(s). Note that vis), called a pseudo state, i s a 2 x 1 column vector.
Then we have
D(s)v(s) = u(s)
yes) = N(s)v(s)
Let us defne state variables as
o
xes) = L(s)v(s) =
o
o
0
s
3
V
I
(5) I (S)
s
2
V
I
(5) 2 (5)
SU
I
(s) X3 (S)
¬.
VI (5) X4(S)
SUl (S) 5 (5)
U2 (S) X6(S)
or, i n the time domain,
XI (t) = v�
3
) (t) X2 (1) = VI
(t)
X5 (t) ~ V2 (t) X6(t) = Vl(t)
Ú. All discussion is still applicable i f D(s) is column reduced but not in echelon form.
(7.93)
(7.94)
(7.95)
222 MI NIMAL REALIZATIONS AND COPRI ME FRACTI ONS
This state vector has dimension J1 + J2 = 6. These defnitions imply immediately
(7.96)
Next we use (7.93) to develop equations for XI and X5 . First we express D(s) as
D(s) = DhcH(S) + DlcL(s) (7.97)
where H(s) and L(s) are defned in (7.91 ) and (7.92). Note that Dhc and Dlc are constant
matrices and the columndegree coeffcient matrix Dhc is a unit upper triangular matrix.
Substituting (7.97) into (7.93) yields
or
[DhcH(S) + DlcL(s»)v(s) = u(s)
I
r
H(s)v(s) + Dh) DlcL(s)v(s) = Dhc
I
U(S)
Thus we have, using (7.95),
Let
and
H(s)v(s) = DhcDlcx(s) + Dh/u(s)
D
I
D
al i i
hc c ¬·
a21 1
al l2 al l3 a 1 1 4 a l 2l
a21 2 a2l 3 a21 4 a221
Dh
c
=
:
�
b
�2
(7.98)
(7.99)
(7. 1 00)
Note that the inv�rse .of a unit" upper triangular matrix is again a unit upper triangular matrix.
Substituting (7.99) and (7. 100) into (7.98), and using SXI
(s) = S4VI (s), and S£5(S) = S
2
V2 (s),
we fnd
S�I (S)
=
_ a l i i al l2
SX5(S)
a2ll a21 2
+
�
b
�2
u(s)
which becomes, in the time domain,
�I
=
.
al l l
X5
a21 1
al l 2 al l 3 al l4 a l 21
a21 2 a2l3 a21 4 a2Z1
a1 22 1 b
iZ
x + U
a 222 0 1
(7. 1 01 )
I f G(s) = N(s)D
I
(s) i s strictly proper, then the column degrees of N(s) are less than the
corresponding column degrees of D(s). Thus we can express N(s) as
N(s) =
.i l l
.21 1
. 1 1 2 . 1 1 3 .
1
14 . 1 2l
. I Z .21 3 .21 4 .221
(7. 1 02)
7. 9 Real izations from Matrix Copri me Fracti ons
Substituting this into (7.94) and using xes) = L(s)v(s), we have
yes) =
.i l l .l l 2 .1 I 3 .1 I 4 .l 2l .1 22 ¸
xes)
.2 1 1 .212 .Zl3 .214 .221 .222
Combining (7.96), (7. 1 01 ), and (7. 1 03) yields the following realization for G(s) :
al l a1 l2 al l 3 a1 l 4 Q1 21 I2`
0 0 0 0 0
0 0 0 0 0
x =
0 0 0 0 0
x
a21 1 a21 2 a2l 3 C214 0`21 a222
0 0 0 0 0
I b
1 2
0 0
0 0
+ 0 0 U
0 I
0 0
y
=
.
.I I I .l 2 .l l3 .I
I J .1 21 .122
X
.21 1 .212 .21 3 .21 4 .221 .222
223
(7. 1 03)
(7. 1 04)
This is a (J. 1 +J2)dimensional state ecuation. TheAmatrix has two companionform diagonal
blocks: one with order J1 = 4 and the other with order J2 = 2. The ofdiagonal blocks are
zeros except their frst rows. This state equation is a generalization of the state equation in
(7. 9) to twoinput twooutput transfer matrices. We can easily show that (7. 1 04) is always
controllable and is called a controllableform realization. Furhermore, the controllability
indices are /1 = 4 and /2 = 2. As in (7.9), the state equation in (7. 1 04) is observable if
and only if D(s) and N(s) are right coprime. For its proof, see Reference [6, p. 282]. Because
we start with a coprime fraction N(s)D
I
, the realization in (7. 1 04) is observable as well. In
conclusion, the realization in (7. 1 04) is controllable and observable and its dimension equals
/1 + /2 and. following Theorem 7.M4, the degree of G(s) . This establishes the second part
of Theorem 7.M2, namely, a state equation is minimal or controllable and observable if and
only if its dimension equals the degree of its transfer matrix.
EXAMPLE 7.8 Consider the transfer matrix in Example 7.6. We gave there a minimal realiza
tion that is obtained by reducing the non minimal realization in (4.39). Now we will develop
directly a minimal realization by using a coprme fraction. We frst write the transfer matrix as
224 MI NI MAL REALI ZATI ONS AND COPRI ME F RACTI ONS
�
`
.
.
�
�
. . .
�� [
.
l
l
��
.I ) ( s . :·
As in Example 7. 7. we can fnd a right coprime fraction for the strictly proper part of ((5) as
.
.
.
·
. .
.
Note that its denominator matrix is the same as the one in Clearly we have �l l 2 and
f 2 . We defne
Then we have
and
We compute
and
..
� [
N( s) .
¬

D
hc
L( s)
.
·
[
L( s)
Dh
c
l
D
lc
� �
.
[
.
�
�
� [
.
�
� �
Thus a minimal realization of . s ) i s
7 . 1 0 Real i zati ons from Matri x ,\\arkov Parameters 225
This Amatrix has two companionform diagonal blocks: one with order .and the other
order This threedimensional realization is a mi ni mal realization and is in controllable
canonical foml.
Dual to the preceding minimal realization. if we use ((.)
0 1 · where 0( . )
and ·are left coprime and i s i n row echelon form with row degrees { v, . i
. , q} , then we can obtain an observableform realization with observability indices
{ v" i . .. . q} . This will not be repeated.
We summarize the main results i n the followi ng. As i n the SISO case. an ndimensional
multivariable state equation is controllable and observable i f its transfer matrix has degree l . If
a proper transfer matrix is expressed as a right coprime fraction with column reducedness. then
the realization obtained by using the preceding procedure will automatically be controllable
and observable.
Let ( A. B. C. D) be a minimal realization of ((5 ) and let  ·
N( s) D
I
( s) be coprime fractions:  is row reduced. and D(s ) is column reduced. Then
we have
which implies
I
.
I
B[Adj( sI  A) ] C D = N( s) [Adj ( D(s ) ]
det ( sI  A) det D( s )
= _[Adj ( D( s) ) ] N( s)
det DI s)
Because the three polynomials det (.II  A) . det D( s) , and det 0(5) have the same degree.
they must denote the same polynomial except possibly different leading coeffcients. Thus we
conclude the following:
• deg . deg det D( ) = deg det dim A.
• The characteristic polynomial of ((.» = kl det D( s ) .k" det  k
det .
A) for
some nonzero constant k, .
• The set of column degrees of D(s) equals the set of controllability indices of ( A. B).
• The set of row degrees of 0( » equals the set of observability indices of ( A. C) .
We see that coprime fractions and controllable and observable state equations contain essen
tially the same i nformation. Thus either description can be used in analysis and design.
7. 1 0 Real izati ons from Matrix Markov Parameters
Consider a q x p strictly proper rational matrix . We expand it as
. ... . .
226 MI NIMAL REALIZATI ONS AND COPRI ME F RACTI ONS
where .are .X .constant matrices. Let be the degree of the least common denominator
of all entries of G(5) . We form
T =
. .I )
.· .
. .
.
.·
.
. ·
.·
.
.
t =
.
�
.
. .· .
.
.
. ·
.·
¯ I)
� ��
.
.·
Note that T and t have block columns and block rows and, consequently, have dimension
.x .As in and we have
T = DC and t = DAC
where 0 and C are some observability and controllability matrices of order .X n and X .
respectively. Note that is not yet determined. Because equals the degree of the minimal
polynomial of any minimal realization of G(5) and because of and ·the matrix T
is suffciently large to have rank This is stated as a theorem.
." Teorem 7.M7
A strictly proper rational matrix G(5) has degree n if and only if the matrix T in (7. 1 07) has rank
The singularvalue decomposition method discussed in Section can be applied to the
matrix case with some modifcation. This is discussed in the following. First we use singular
value decomposition to express T as
where K and L are orthogonal matrices and A = diag(il , A2, . . . . An), where Ai are the positive
square roots of the positive eigenvalues of T'T. Clearly is the rank of T. Let K denote the
frst columns of K and f: denote the frst rows of L'. Then we can write T as
where
T = KAL' = KA
1 /
2
A
1
/2
1' =; OC
0 = KA
1 /
2 and C = AI /2i'
Note that 0 is .x n and C is x .They are not square and their inverses are not defned.
However, their pseudoinverses are defned. The pseudoinverse of 0 is. as defned in
7. 1 1 Concl udi ng Remarks 227
0+
= [ (AI /
2
)'K
'
KAI /2
r
l (AI
!2
)
'
K
'
Because K is orthogonal, we have K'K = I and because A 1
/2 is symmetric. the pseudoinverse
0+ reduces to
Similarly, we have
Then, as in ·through the triplet
A = O+tC+
B = frst .columns of C
C = frst .rows of 0
is a minimal realization of G(5) . This realization has the property
and, using 1'1 = I,
CC' = A l
/
eL'LA 1 /2
= A
·
Thus the realization is a balanced realization. The MATLAB procedure in Example can
be applied directly to the matrix case if the function inverse ( �n'l) is replaced by the function
pseudoinverse (pinv). We see once again that the procedure in the scalar case can be extended
directly to the matrix case. We also mention that if we decompose T = DC in
differently, we will obtain a diferent realization. This will not be discussed.
7, 1 1 Concl udi ng Remarks
In addition to a number of minimal realizations, we introduced in this chapter coprime fractions
(right fractions with column reducedness and lef fractions with row reducednessJ. These
fractions can readily be obtained by searching linearly independent vectors of generalized
resultants and then solving monic null vectors of their submatrices. A fundamental result of
this chapter is that controllable and observable state equations are essentially equivalent to
coprime polynomial fractions, denoted as
....
.
�
...·.
Thus either description can be used to describe a system. We use the former in the next chapter
and the latter in Chapter to carry out various designs.
A great deal more can be said regarding coprime polynomial fractions. For example, it is
possible to show that all coprime fractions are related by unimodular matrices. Controllability
228 MI NIIv",L REALIZATIONS AND COPRL'\E FRACTI ONS
PROBLEMS
and observability conditions can also be expressed in terms of coprimeness conditions. See
References [4, 6, 1 3, 20]. The objectives of this chapter are to discuss a numerical method to
compute coprime fractions and to introduce just enough background to carry out designs in
Chapter 9.
In addition to polynomial fractions, it i s possible to express transfer functions as stable
rational function fractions. See References [9, 21 ] . Stable rational function fractions can be
developed without discussing polynomial fractions; however, polynomial fractions can provide
an effcient way of computing rational fractions. Thus this chapter is also useful in studying
rational fractions.
7.1 Given
7.2
7.3
5  1
g (s) 
(s"  1 ) (5 + 2)
fnd a threedimensional controllable realization. Check i ts observability.
Find a threedimensional observable realization for the transfer function i n Problem 7. 1 .
Check its controllability.
Find an uncontrollable and unobservable realization for the transfer function i n Problem
7. 1 . Find also a minimal realization.
7.4 Use the Sylvester resultant to fnd the degree of the transfer function i n Problem 7. 1 .
7.5 Use the Sylvester resultant to reduce (25  1 ) /(45"  to a coprime fraction.
7.6 Form the Sylvester resultant of g( s) = (5 + 2)/ (52 + 25) by arranging the coeffcients of
N( s) and Dis) in descending powers of 5 and then search linearly independent columns
in order from left to right. Is it true that all Dcolumns are linearly independent of their
LHS columns" Is it true that the degree of g(s ) equals the number of linearly independent
Ncolumnso
7.7 Consider
and its realization
.
0
X
+
0
.
N(s )
D( s)
y [ ,I ,2 ]X
Show that the state equation is observable if and only if the Sylvester resultant of D( s)
and N ( 5) i s nonsingular.
7.8 Repeat Problem 7.7 for a transfer function of degree 3 and its controllableform realiza
tion.
7.9 Verify Theorem 7.7 for g( S ) = 1 / (.1 + 1 )2 .
Problems 229
7.10 Use the Markov parameters of g(s) 1 /(.1 + 1 ) " to fnd an irreducible companionform
realization.
7. 11 Use the Markov parameters of g(s) = 1 /(.1 + 1 )2 to fnd an irreducible balancedform
realization.
7.12 Show that the two state equations
7. 13
and
.
2
.
.
1 2
y [2 2]x
y = [2
.
are realizations of (25 + 2)/(.12  J  2). Are they minimal realizations? Are they
algebraically equivalent?
Find the characteristic polynomials and degrees of the following proper rational matrices
and
I
×
S
GI ( s) =
_
1
_
.1 + 3
.1 + 3
·
·�
(h (s)
(. �1 )2 (5 + 1 )
1
(.1 + 2)
.1 + 1 .1 + 2 (5 + 1 ) (.1 + 2)
.1 + 3
.1 + 2
·+
5 + 4
Note that each entry of Gl(s) has different poles from other entries.
7. 14 Use the left fraction
G(s)
5
to form a generalized resultant as in (7. 83), and then search its linearly independent
columns in order from left to right. What i s the number of l inearly independent N
columns" What i s the degree of G( s) " Find a right coprime fraction of G(s). Is the given
left fraction coprimeo
7, 15 Are all Dcolumns in the generalized resultant in Problem 7. 1 4 linearly independent of
their LHS columns? Now i n forming the generalized resultant. the coeffcient matrices of
D( s) and N (5) are arranged in descending powers of 5, instead of ascending powers of
5 as in Problem 7. 1 4. Is it true that all Dcolumns are l inearly independent of their LHS
columns? Does the degree of G(s) equal the number of l inearly independent Ncolumnso
Does Theorem 7.M4 hold?
230 MI NIMAL REALIZATIONS AND COPRI ME FRACTIONS
7.16 Use the right coprime fraction of G(s) obtained in Problem 7. 1 4 to form a generalized
resultant as in (7.89), search its linearly independent rows in order from top to bottom,
and then fnd a left coprime fraction of G(s).
7.17 Find a right coprime fraction of
• 3
+ 1
G(s) =
s
s + 2
7
and then a minimal realization.
�
2
s
8. 1 Introduction
LH3D!CI
State Feedback
and State Esti mators
The concepts of controllability and observability were used in the preceding two chapters to
study the interal structure of systems and to establish the relationships between the interal
and exteral descriptions. Now we discuss their implications in the design of feedback control
systems.
Most control systems can be formulated as shown in Fig. 8 . 1 , in which the plant and
the reference signal r(t) are given. The input u(t) of the plant is called the actuating signal
or control signal. The output y(t) of the plant is called the plant output or contrlled signal.
The problem is to design an overall system so that the plant output will follow as closely as
possible the reference signal r(t). There are two types of control. If the actuating signal u(t)
depends only on the reference signal and i s independent of the plant output, the control is
called an openloop control. If the actuating signal depends on both the reference signal and
the plant output, the control is called a closedloop orfeedback control. The openloop control
is. in general, not satisfactory if there are plant parameter variations and/or there are noise
and disturbance around the system. A properly designed feedback system, on the other hand,
Figre 8.1 Design of control systems.
. . . . . . . . . . . . . . . . . . . & ~

� �

I
L & . . . . . . . . . . . . . & . & . . .
231
232 STATE FEEDBACK AiD STATE ESTI M/TORS
can reduce the effect of parameter variations and suppress noise and disturbance. Therefore
feedback control is more wi del y used in practice.
This chapter studies desi gns usi ng statespace equations. Designs using coprime fractions
wi l l be studied in the next chapter. We study frst singlevariable systems and then multi variable
systems. We study only l i near timeim' ariant systems.
8. 2 State Feedback
Consider the ndimensional si nglevariable state equation
x = Ax + bll
y = ex
( S I )
where we ha\'e assumed d = 0 to si mpl i fy discussion. In state feedback. the i nput U is given by
lI = r  kx = r  [kl k2 . . ' k" Jx = r  _k,x,
i =1
( 8. 2)
as shown i n Fig. 8. 2. Each feedback gai n k, is a real constant. This is called the constant gain
negatil'e stare feedback or. simpl y. state feedback. Substituting ( S. 2) into ( S. I ) yields
x = (A  bk)x + br
( S.3)
y = ex
Theorem 8. 1
The pair ( A  bk. bl . for any x n real constant vector k. is controllable i f and only if (A. b) is
control l abl e.
Proof: We show the theorem for n = �. Defne
and
C, = [b (A  bk)b (A  bk)2b (A  bk)'b]
They are the controllabi l i ty matrices of I S. I ) and ( S. 3) . It is straightforward to verify
, . . . . . . . . . . . . . . . . . ,
Figure 8.2 State fedback.
C, � C
0
kb
0
ki A  bk) b
kb
I
0
8. 2 State Feedback 233
··k( A ·· bkl'b
k( A  bk)b
( 8, �)
kb
I
Note that k is I x n and b is 11 X 1 . Thus kb is scalar: so is every entry in the rightmost
matrix in ( S.4) . Because the rightmost matrix i s nonsingular for any k. the rank of Ct
equals the rank of C. Thus ( 8. 3) is controllable if and only i f ( 8. 1 ) i s controllable.
This theorem can also, be establ i shed directly from the defnition of controllability.
Let Xn and X be two arbitrary states. HI S. I ) is controllable, there exists an input /I  that
transfers X( to X in a fnite time. Now i f we choose rl = I  ¬ kx, then the input rl of the
state feedback system will transfer X(}to X . Thus \\ e conclude that if ( 8. 1 ) i s controllable,
so i s ( 8. 3) .
We see from Fi g. S. 2 that the input l does not control t he state x directly: i t generates
1I to control X. Therefore. if II cannnot control x. neither can l. This establishes once again
the theorem. Q. E. D.
Although the control lability property i s i nvariant under any state feedback. the observ
ability property is not. This i s demonstrated by the example that follows.
EXAMPLE 8.1 Consider the state equation
V = [ I 2J x
The state equation can readi l y be shown t o be controllable and observable. Now we i ntroduce
the state feedback
1I = r  [3 I Jx
Then the state feedback equation becomes
Its control l abi l ity matrix is
y = [ I 2J x
C =
2
!
I 0
which is nonsi ngul ar. Thus the state feedback equation is controllable. I ts observabi l i ty matri\ i s
0,
=
I :
which is si ngul ar. Thus the state feedback equation is not observable. The reason that the
observabi li ty property may not be preserved i n state feedback wi ll be gi ven later.
234
STATE FEEDBACK AND STATE ESTI MATORS
We use an example to discuss what can be achieved by state feedback.
Consider a plant described by
x
=
[ � � } ·
�}
E
"
.
'
"
8
.
2
The Amatrix has characteristic polynomial
t(s) = (s ¯ 1 )
 9 = s2 ¬ 2s  8 = (s  4)(s 2)
. d 2 It is unstable. Let us introduce state feedback
and, consequently, eigenvalues an .
. .
u = r ¯
·
Then the state feedba
�
k system IS descnbed by
x
=
(
[ � �
�
� })
·
�}
=
[
1 �
3
�
}·
+
�
}
r
This new Amatrix has characteristic polynomial
t/(s) = (s  1 + kl ) (s  1)  3 (3 
= s
2
+  2)s (3kz   8)
. I tl the eigenvalues of the state feedback
It is clear that the roots of t/(s) or, eqUiva
l
e
�
y,
riate k and kz For example, if
be I d ' positions by se ectmg approp ]
·
system can p ace m any
I ± ' 2 th the desired characteristic polynomial is
the two eigenvalues are to be placed at  J
,
.
en
k 2 2 and 3kz ¯  8 = 5 yields
' 2) ( I 'J) ¬ s
2 + 2s + 5. Equating ) ¨ ¬
(s 1 + J s +  J . 
b k
.
.17/31 will shift the eigenvalues from
= .and k= 1 7/3. Thus the state feed ac gam
., "2 to l ± j2.
k be used to place eigenvalues in any positions.
This example shows that state feedba
�
te
�
a
¸
direct substitution. This approach, however,
Moreover the feedback gam can be comp
.
h
Y
d
'
.
nal state equations. More seriously,
.
1 d f three or hlg er ImenslO
will become very mvo ve or 
lIabili condition comes into the design. Therefore
the approach will not reveal how the contro t
d' we need the following theorem.
a more systematic approach is desirable. Befo�e procee
h���for every positive integer n.
We state the theorem for n = 4; the theorem, owever,
> Theorem 8.2
. '
. h  4 and the characteristic polynomial
Consider the state equatIOn ID (8. 1 ) Wit n 
4 ² ` _
+ a
t(s) = det (sl  A) = s + al S" + azs , a
3
s 4
f d b th transformation x = P with
If (
g ll ("
��I.b'
�
,
�
´
'
�
,
A
'�
�
a
�
(8. 5)
(8.6)
8.2 State Feedback 235
into the controllable canonical form
al
.   I
x = Ai + bu = �
o 0
1 0
o
Furherore, the transfer function of (8. 1 ) with n = 4 equals

;
4

�
o
·
+
0
u
o 0
'( )
fl s3 fzs
z
fhs + f4
g S
=
S
4
+ al s
3
+ azs2 + a
3
s + a4
(8. 7)
(8. 8)
Proof: Let C and C be the controllability matrices of (8. 1 ) and (8. 7). In the SISO case,
both C and C are square. If ( 8. 1 ) is controllable or C is nonsingular, so is C. And they
are related by C = PC (Theorem 6.2 and Equation (6.20)). Thus we have
P = CCI or Q : = pI = CC
I
The controllability matrix C of (8.7) was computed i n (7. 10). Its inverse turs out t o be
1 a
l
C
I
=
0 I
o 0
o 0
(8.9)
This can be verifed by multiplying (8.9) with (7. 10) to yield a unit matrix. Note that the
constant term a4 of (8. 5) does not appear in (8.9). Substituting (8.9) into Q = C CI yields
(8.6). As shown in Section 7.2, the state equation i n (8.7) is a realization of (8.8). Thus
the transfer function of (8.7) and, consequently, of (8. 1 ) equals (8. 8). This establishes the
theorem. Q.E.D.
With this theorem, we are ready to discuss eigenvalue assignment by state feedback.
� Theorem 8.3
If the ndimensional state equation in (8. 1 ) is controllable, then by state feedback u = r  k. where
k is a 1 X n real constant vector, the eigenvalues of A  bk can arbitrarily be assigned provided that
complex conjugate eigenvalues are assigned in pairs.
Proof: We again prove the theorem for n = 4. If (8. 1 ) is controllable, it can be transformed
into the controllable canonical form in (8.7). Let A and b denote the matrices in (8.7).
Then we have A = PAP
I
and b = Pb. SubstitUling x = P into the state feedback
yields
u = r ¬ k = r  kP
l
x =: r ¬ ki
where k : = kP
I . Because A  bk = peA  bk)pI , A  bk and A  bk have the same
set of eigenvalues. From any set of desired eigenvalues, we can readily form
236 STATE FEEDBACK AND STATE ESTIMATORS
If k i s chosen as
the state feedback equation becomes
' = (A  bk)i + b �
)'
=
[. I .
2
.3 ..]i
o
o
Because of t he companion form, t he characteristic polynomial of (A  jjk) and, conse
quently. of (A  bk) equals Thus the state feedback equation has the set of desired
eigenvalues. The feedback gain k can be computed from
k = kP = kCC
1
with k in CI in and C = [b Ab A2b A3b]. Q. E. D.
We give an alterative derivation of the formula in We compute
6f(S) = det (51  A + bk) = det ( 51 A) [I + (51  A)
l
bkJ)
= det (51 A)det [I + (51 A)I bk]
which becomes, using and ·
6j (s) = 6(5 ) [ 1 + k( sl  A) I b]
Thus we have
Let � be the output of the feedback gain shown in Fig. and let k = [kl
the transfer function from H to y i n Fig. equals
~
_
, ¡ ¬ .
1
53 + .2s2 + .ls + .4
c(s I A) b =
) u(s
the transfer function from H to Z should equal
_ _
,  kl s3 + kos
2
+ k
3
S + k.
k(sl  A) b =

6( s)
Substituting and into yields
(al  O i ls3 + (a
2
0
2
) s
2
+ (a
3
 03)S + (a4  04) = kls3 + k25
2
+ kJ5 + k4
This yields
j
8. 2 State Feedback 237
Feedback transfer function Consider a plant described by ( A. b. c) . If (A. b) is controllable.
(A. b. c) can be transformed into the controllable form in and its transfer function C0N
then be read out as. for n = 4.


I
b
.1 53 + .252 + .l s + .4
g(s) = c(sl  A) = ,
54 + 0 I 53 + 0
2
S + 0'
3
s + 04
After state feedback. the state equation becomes (A  bk. b. c) and i s still of the controllable
canonical for as shown in Thus the feedback transfer function from r to \' is
× _1 .1 S
3
+ .
2
5
2
¯ .l s + .4
g
j
(s) = c(sl  A + bkl b = _ . .
S + OI S' + 0
2
S

+ 0
3
S + a.
We see that the numerators of and are the same. I n other words, state feedback
does not affect the zeros of the plant transfer function. This is actually a general propeny of
feedback: feedhack can shift the poles of a plant but has 170 efect on the :erus. This can be used
to explain hy a state feedback may alter the observability propeny of a state quation. If one UI
more poles are shifted to coincide with zeros of g( s) . then the numerator and denominator of
gi(s) in are not coprime. Thus the state equation i n ·and. equivalently, (A bk. L¹
are not observable (Theorem
EXAMPLE 8.3 Consider the invened pendulum studied in Example · Its state equation i s.
as derived i n
x =
�
�
o
\' = [ I O O O]x
I t i s controllable; thus its eigen\'alues can b e assigned arbitrarily. Because the Amatrix i s block
triangular, its characteristic polynomial can be obtained by inspection as
6(s) = s
2
(s
2
5) = s· + s3 
+ O · s +
First we compute P that will transform into the controllable canonical form. Using
we have
I
P �
·
I
°
�
I
·
Its inverse i s
238 STATE FEEDBACK AND STATE ESTI MATORS
p � ��
Let the desired eigenvalues be 1 . S ± O.S j and I ± j. Then we have
and
!j(S) = (s + 1 . S  O. Sj) (s + 1 . 5 + O. Sj) (s + 1 ¯ j) (s + I + j)
= S4 + SS3
+ 1 0. Ss2 + l i s + S
Thus we have, using (8. 1 l),
k ÷ [S ¯ 0 10. S + S; li  0 S  0] = [S I S. S l i S]
(8. 1 9)
This state feedback gain will shift the eigenvalues of the plant from {O, c ±j�} to
{ 1 .S ± O. Sj. I ± j) .
The MATLAB function p lace computes state feedback gains for eigenvalue placement
or assignment. For the example, we type
a= [ O 1 ' 0 0 ; 0 0  1 0 ; 0 0 0 1 ; 0 0 5 0 ] ; b: [ 0 ; 1 ; 0 ;  2 ] ;
p= [  1 . 5 + 0 . 5 j  1 . 5  0 . 5 j  l + j  l  j ] ;
k=place ( a, b , p)
which yields [1 .6667 ¯ 3. 6667  8. S833  4. 3333] . This is the gain in (8. 1 9).
One may wonder at this point how to select a set of desired eigenvalues. This depends
on the performance criteria, such as rise time, settling time, and overshoot, used in the design.
Because the response of a system depends not only on poles but also on zeros, the zeros of the
plant will also afect the selection. I n addition, most physical systems will saturate or bur out if
the magnitude of the actuating signal is ver large. This will again affect the selection of desired
poles. As a guide. we may place all eigenvalues inside the region denoted by C in Fig. 8. 3(a).
The region is bounded on the right by a vertical line. The larger the distance of the vertical line
from the imaginary axis, the faster the response. The region is also bounded by two straight
lines emanating from the origin with angle a. The larger the angle. the larger the overshoot. See
Reference [7]. If we place all eigenvalues at one point or group them iIi a very small region,
then usually the response will be slow and the actuating signal will be large. Therefore it is
better to place all eigenvalues evenly around a circle with radius r inside the sector as shown.
The larger the radius. the faster the response; however. the actuating signal will also be larger.
Furthermore. the bandwidth of the feedback system will be larger and the resulting system
will be more susceptible to noise. Therefore a fnal selection may involve compromises among
many conficting requirements. One way to proceed is by computer simulation. Another way
is to fnd the state feedback gain k to minimize the quadratic performance index
J ÷ [x' (t)Qx(t) + u'(t)Ru(t)] dt
Im s
 Re S
c
(a)
Figure 8.3 Desired eigenvalue location.
1m :
1
(b)
8. 2 State Feedback 239
Re :
See Reference [ I ] . However selecting Q and R
. .
select a set of desired eigen
�
alues is not a simpl
�
e
_��·�
nal and e
r
or. In conclusion. how to
We mention that Theorems 8 I throu h 8 3
.
f
.
in this chapterapplv to the discre
'
te tl' m
g .

:
h
m act, all theorems to be introduced later
+
w
¯ e case wit out any mod' fe t" Th I
. .
l5 that the region in Fig. 8. 3(a) must be re laced b
.
l a IOn. e on y difference
by the transformation : = eS•
p y the one U Fig. 8. 3(b). which ¡5 obtained
8. 2. 1 Solving the Lyapunov Equation
This subsection discusses a diferent method of com
. . .
assignment. The method however ha th
.
putmg state feedback gam for eigenvalue
contain any eigenvalues �f A.
. s e restnctlOn that the selected eigenvalues cannot
> Procedure 8. 1
Consider controllable (A. bl. where A is n x n and b " I F d
has any set of desired ei2envalues that conta'
.
IS n x . :n a x n real k such that (A ¯ bk)
"
InS no eIgenvalues of A.
1. Select an n x n matrx F that has the set of desired ei
.
and will be discussed later.
genvalues. The form of F can be chosen arbitrarily
2. Select an arbitrary I x 11 vector k such that (F k

) ' b . ISO servable.
3. Solve the unique T in the Lyapunov equation AT ¯ TF = bk.
4. Compute the feedback gain k = kT
1
•
We justify
_
frst the procedure. 1fT is nonsingular then k  kT d h L
.
AT ¬ TF = bk implies
'
 an t e yapunov equauon
(A ¯ bk)T ÷ TF or A ¯ bk ¯ TFT1
Thus (A  bk) and T are similar and have the sa
.
of (A ¯ bk) can be assi2ned arbitrarily h
me
�
et of eigenvalues. Thus the eigenvalues
« except t ose ot A. As discussed in Section 3.7. if A and
240 STATE FEEDBACK AND STATE ESTI MATORS
F have no eigenvalues in common. then a solution T exists in AT  TF = bk for any k and
is unique. If 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 eigenvalues in common.
What remains to be proved i s the nonsingularity of T.
Theorem 8.4
If A and F have no eigenvalues i n common. then the unique solution T of AT TF = bk i s nonsingular
i f and only if (A. b) i s controllable and ( F. k) is observable.
Proof; We prove the theorem for n = Let the characteristic polynomial of A be
t,(s) = .1
4
+ CI �3 + C2S2 + C
3
S + C4 ·
Then we have
(CayleyHamilton theorem). Let us consider
t,(F) : = F + C I F
J
+ C
2
F2 + C
J
F + C4I ·
If
)
.j i s an eigenvalue of F. then t,()., ) i s an eigenvalue of t,(F) (Problem Because
A and F have no eigenvalues i n common. we have t,()j ) for all eigenvalues of F.
Because the determinant of a matrix equals the product of all its eigenvalues, we have
det t,(F) = t,()., )
Thus t,(F) is nonsingular.
Substituting AT = TF + bk into A2T  AF2 yields
A2T  TF2 = A( TF + bk)  TF2 = Abk + (AT  TF)F
= Abk + bkF
Proceeding forward. we can obtain the following set of equations:
IT  TI = O
AT  TF = bk
ACT  TF2 = Abk + bkF
A1T  TF
3
= A 2bk + AbkF + bkF2
A
4
T  TF4 = Albk + A 2bkF + AbkF2 + bkF
l
We multiply the first equation by C4, the second equation by C
l
, the third equation by C
2
,
the fourth equation by C I , and the last equation by 1 , and then sum them up. After some
manipulation, we fnally obtain
t,(A)T  Tt,(F) = Tt,(F)
I
J
= [b Ab A2b A3bJ
C
2
Cl
C2
Cl
Cl
8. 2 State Feedback
� o kP
241
.
where we have used �( A) = O. If ( A, b) is controllable and (F. k) i s observable. then all
three matrices after the last equality are nonsingular. Thus ·· and the nonsingularity of
t,(F) imply that T is nonsingular. If ( A. b) is uncontrollable and/or ( F. k) is unobservable.
then the product of the three matrices i s singular. Therefore T is singular. This establishes
the theorem. Q. E. D.
We now discuss the selection of F and k. Given a set of desired eigenvalues. there are
infi nitely many F that ha\ e the set of eigenvalues. If we form a polynomial from the set. we
can use its coeffcients to form a companionform matrix F as shown in For this F.
we can select k as
´ ¯ OJ and (F, k) i s observable. If the desired eigenvalues are all
distinct, we can also use the modal form discussed i n Section For example, i f II = 5,
and i f the fve distinct desired eigenvalues are selected as A I . C j ± j fI . and C
2
± j f
2
, then we
can select F as
F � �
Cl fI
fl Cl
C
2
f
2
C
2
It is a blockdiagonal matrx. For this F, if k has at least one nonzero entry associated Vith
each diagonal block such as k = k = or k = J . then (F. k) is
observable (Problem Thus the frst two steps of Procedure are very simple. Once F
and k are selected. we may use the MATLAB function lyap to solve the Lyapunov equation
in Step Thus Procedure is easy to carry out as the next example illustrates.
EXAMPLE 8.4 Consider the i nverted pendulum studied i n Example The plant state
equation is given in and
'
the desired eigenvalues were chosen as  I ± j and ± 0. 5 j.
We select F i n modal form as
and k = 1 OJ. We type
F =
o
o
1
o
o
o
o
1 . 5
a= [ 0 . 0 0 ; 0 0 < 0 ; 0 0 0 1 ; 0 0 5 0 J , b= [ 0 ; 1 ; 0 ; ¯ 2 j ;
f= :  l 1 0 0 ;  1 ¯ . Û Ü ¦ · 0  1 . 5 0 . 5 ; 0 0  0 . 5  1 . 5 ] ;
kb= [ l O 1 0 ] ; t = :: "p ( a , Q ,  b*kb) ;
k= kb* inv ( t )
242 STATE FEEDBACK AND STATE ESTIMATORS
The answer is [ 1 .6667 ¯ 3. 6667  8 . 5833 4. 3333], which is the same as the one obtained
by using function place. If we use a diferent k ~ [1 1 1 ] , we will obtain the same k. Note
that the feedback gain is unique for the SISO case.
8. 3 Regulation and Tracki ng
Consider the state feedback system shown in Fig. 8. 2. Suppose the reference signal r is zero,
and the response of the system is caused by some nonzero initial conditions. The problem is
to fnd a state feedb'ck gain so that the response will die out at a desired rate. This is called a
regulator problem. This problem may arise when an aircraft is cruising at a fxed altitude Ho·
Now, because of turbulance or other factor
1
, the aircraft may deviate from the desired altitude.
Bringing the deviation to zero is a regulator problem. This problem also arises in maintaining
the liquid level in Fig. 2. 14 at equilibrium.
A closely related problem is the tracking problem. Suppose the reference signal r is a
constant or r(t) ~ a, for t 2 O. The problem is to design an overall system so that yet)
approaches ret) = a d t approaches infnity. This is called asymptotic tracking of : step
reference input. It is clear that if ret) ~ a ~ 0, then the tracking problem reduces to the
regulator problem. Why do we then study these two problems separately? Indeed, if the same
state equation is valid for all r, designing a system to track asymptotically a step reference input
will automatically achieve regulation. However, a linear state equation is often obtained by
shifting to an operating point and linearization, and the equation is valid only for r very small
or zero; thus the study of the regulator problem is needed. We mention that a step reference
input can b set by the position of a potentiometer and is therefore often referred to as set point.
Maintaining a chamber at a desired temperature is often said to be regulating the temperature; it
is actually tracking the desired temperature. Therefore no sharp distinction is made in practice
between regulation and tracking a step reference input. Tracking a nonconstant reference signal
is called a servomechanism problem and is a much more diffcult problem.
Consider a plant described by (A. b, c) . If all eigenvalues of A lie inside the sector shown
in Fig. 8.3, then the response caused by any initial conditions will decay rapidly to zero and
no state feedback is needed. If A is stable but some eigenvalues are outside the sector, then
the decay may be slow or too oscillatory. If A is unstable. then the response excited by any
nonzero initial conditions will grow unbounded. In these situations, we may introduce state
feedback to improve the behavior of the system. Let u ~ r ¯ k'. Then the state feedback
equation becomes (A  bk, b, c) and the response caused by x(O) is
y(t) =
ce
(Abk)l x(O)
If all eigenvalues of (A ¯ bk) lie inside the sector in Fig. 8.3(a1. then the output will decay
rapidly to zero. Thus regulation can easily be achieved by introducing state feedback.
The tracking problem is slightly more complex. In general. in addition to state feedback.
we need a feedforward gain p as
u(t) ~ pr(t) ¯ kx
Then the transfer function from r to y differs from the one in (8. 1 7) only by the feedforward
gain p. Thus we have
8. 3 Regulation and Tracki ng 243
gl(s) ¯ �
(s)
= P
�
I S_
3 +
�2s
2
+ �
3
S + �4
res) S4 + £l s3 + a2s2 + a3s + a4
(8.24)
If (A, b) is controllable, all eigenvalues of (A  bk)
.
be assigned arbitrarily, in particular, assigned to lie ins���
eqUivalently, a
�
l poles of gl(
s
) can
assumption, if the reference input is a ste f
. .
the sector In Fig. S. 3(a). Under this
will approach the constant gl(O)
.
a as t
�n�·�.With magnitude a, then the output y(t)
asymptotically any step reference input, we need
orem 5.2). Thus In order for yet) to track
l = gl (0) = P
�
4
or p =
a
4
(8 "5)
£
4 �
4
.
which requires �4 f O. From (8. 1 6) and (S 1 7) we
.
of the plant transfer function. Thus �4 f 0 f ancon
see that �4 IS the numerator constant term
zero at s = O. In conclusion if g' (s) h
Iy If the plant transfer function g(s) has no
If •
, as one or more zeros at s 0 tr ki
.
g(s) has no zero at 5 = 0 we introduce df
 , ac ng IS not possible.
, a lee orward gain a
.
(8 25)
system will track asymptotically any step f '
S In . . Then the resulting
W
. .
.
re erence Input.
e summanze the preceding discussion Given (A b
. '
.
may introduce state feedback to place th
.
.
I
" c) , If (A, b) IS controllable, we
h
e elgenva ues of (A ¯bk) in any d
.
d
.
.
t e resulting system will achieve regulation If (A b)
.
eSife pOSitIOns and
no zero at s ~ 0, then after state feedback
'
' .
15 controllable and If c(sI ¯ A)
I
b has
Then the resulting system can track asymp
'
tt
e m
l
�y Introduce a feedforward gain as in (8.25).
Ica y any step reference Input.
8. 3. 1 Robust Tracking and Disturbance Rejection'
The state equation and transfer function develo ed
.
change of load, environment or aging Th I
P to descnbe a plant may change due to
T
, . us p ant parameter variations ft
.
he equation used in the design is often called the n '
.
O en occur Upractice.
p
In (S. 25), computed for the nominal plant transfer
o¯ ���
l e
q
uat
IO
n. Th
� fee
�
forward gain
nonnominal plant transfer functions Then th
.
lOn, may not Yield gl(O) = 1 for
reference input. Such a tracking is sa
'
I' d t b
e output WIll not track asymptotically anv step
.
O Ononrobust
�
I thIS subsection we discuss a diferent des"
.
disturbance rejection
.
Consider a plant described
�gn that can achieve robust tracking and
dIsturbance w with unknown magnitud
y (8. 1). We now assume that a constant
e enters at the plant input as sho
.
F' 8 4
the state equation must be modifed as
wn In Ig. . (a). Then
x = Ax + bu + bw
(8. 26)
y = ex
The problem is to design an overall system so that h
.
any step reference input even with the f
t e output y(t) WIll track asymptotically
variations. This is called robust tra
c
°
resen
�
e
¸.
a disturbance w (t) and with plant parameter
design, in addition to introducing state
n
;°"
bac
�
s
t
~
ba
�
c
e re
j
e
c
t
io
n`
In
.
order to achieve this
feedback from the output as shown in Fi S 4( ) i
e WIll Introduce an Integrator and a unity
g. . a . et the output of the Integrator be denoted by
¡. Tis section mav b skippd without loss of continuity.
2 STATE FEEDBACK AND STATE ESTI MATORS
(a)
v + 6 + �
+ +
k
()
(c)
Figure 8.4 (a) State feedback with interal model. (b) Interchange of two summers. (c) Tran;fer
function block diagram.
Xa (t). an augmented state variable. Then the system has the augmented state vector [x' xa l
'
·
From Fig. SA(a), we have
Xa = r  y = r  .
(S. 27)
u = . kal � [ (8. 28)
For convenience, the state is fed back positively to u as shown. Substituting these into (S. 26)
yields
(8. 29)
This describes the system in Fig. SAra).
Theorem 8.5
If · . is controllable and if g(.) = .. . · 
1
. has no zero at S = 0, then all eigemalues of the
Amatrix in (8.29) can be assigned arbitrarily by selecting a feedback gain . ka l ·
Proof' We show the theorem for n = 4. We assume that · . and . have been transformed
into the controllable canonical form in (8. 7) and its transfer function equals (8 8). Then
the plant transfer function has no zero at 5 = 0 if and only if f4 i O. We now show that
the pair
8. 3 Regulation and Tracki ng 245
� 
(S. 30)
is controllable if and only if f4 i O. Note that we have assumed n = 4; thus the dimension
of (S.30) is fve because of the additional augmented state variable Xa ' The controllability
matrix of (S.30) is
�
·. · . ·
. ·.
.. .·. .·
. .·
.
� �
"
a, (ai  az) + a2al  aj al aj  a2
I al
,
Ci 
.
0 I al .
0 0 ..
f1
f1 a]  f2 f1 (ai  a2
) + f
2al  fJ .
where the last column is not written out to save space. The rank of a matrix will not change
by elementary operations. Adding the second row multiplied by f1 to the last row. and
adding the third row multiplied by f2 to the last row, and adding the fourth row mUltiplied
by f3 to the last row. we obtain
�
al Cl  · 
al (ai  a2 ) + aZal  a
¯
al ai  a2
0 I al . (S. 3 1 )
0 0 1 .
0 0 0 f4
Its detenninant is f4. Thus the matrix is nons in gular if and only if f4 i O. In conclusion.
if · . is controllable and if g(5) has no zero at . = O. then the pair in (S. 30) is
controllable. It follows from Theorem S. 3 that all eigenvalues of the Amatrix in (8.29)
can be assigned arbitrarily by selecting a feedback gain . ka l . Q.E.D.
We mention that the controllability of the pair in (8. 30) can also be explained from pole
zero cancellations. If the plant transfer function has a zero at . = O. then the tandem connection
of the integrator. which has transfer function 1 / . and the plant will involve the polezero
cancellation of 5 and the state equation describing the connection will not be controllable. On
the other hand. if the plant transfer function has no zero at 5 = 0, then there is no polezero
cancellation and the connection will be controllable.
Consider again (8. 29). We assume that a set of n + 1 desired stable eigenvalues or.
equivalently. a desired polynomial ! j (5) of degree n + I has been selected and the feedback
gain . ka 1 has been found such that
.. · ..
· . = det
c
(S. 32)
Now we show that the output y will track asymptotically and robustly any step reference input
ret) = . and reject any step disturbance with unknown magnitude. Instead of establishing
the assertion directly from (S. 29). we will develop an equivalent block diagram of Fig. SA(a)
and then establish the assertion. First we interchange the two summers between U and U as
26
STATE FEEDBACK AND STATE ESTI MATORS
shown in Fig. S.4(b). This is penitted because we have u ~ U ±l ±U before and after the
interchange. The transfer function from I to y is
« Na)
. · .. .
v a) . ~ ¬.~ ca ~
La)
(8.33)
with La) ~ det a1 A  .. Thus Fig. S.4(a) can be re�rawn as shown �n Fig
:
S.4(c). We
next establish the relationship between O;a)in (S.32) and va)in (S. 33). It ¡o straIghtforward
to verify the following equality:
. .
.. · .. ¬bk,
¡ ¬ca1 · .. ! c a
.. ¬ A ..
l
bk¡
~
0 a + ca1 ~ A bk) ¯' bk¡
Taking its detenninants and using (S.32) and (S.33). we obtain
which implies
~ ¡ Ná}
l · ^, a) ~ La) a ±
¯
k¡
La)
This is a key equation.
.
From Fig. S.4(c). the tansfer function from U to y can readily be computed as
Na)
La)
g va
~ 0
k¡ Na)
l ±
a La)
aNa) aNa)
aLa) ±k¡Na)
^,a)
If the disturbance i s Uet) ~ ufor al l t :: 0, where uis an unknown constant. then ua) = u[a
and the corresponding output is given by
aNa) u uNa)
Yaa) ~
^,a)
~
tj(s)
(S.34)
Because the pole s in (S.34) is canceled. all remaining poles of Yw(s) are stable poles. Therefore
the corresponding time response. for any W. will die out as t  0. The only condltlon to
achieve the disturbance rejection is that Yw a) has only stable poles. Thus th
.
e reJectlon still
holds. even if there are plant parameter variations and variations in the feedforward gam k,
and feedback gain . as long as the overall system remains stable. Thus the dIsturbance ¡o
suppressed at the output both asymptotically and robustly.
The transfer function from r to y is
k¡ Na)
a La)
g?r (.) ~
k
¡
N
�
¯

a)
l ±
a La)
k¡Na) k¡Na)
aLa) ±k¡ Na)
=
Ô,a)
We see that
8. 4 State Estimator
>
,0) ~ ¬
k,N0)
~ ~
k¡�')
~ I g
)
o · L0) ±k¡ N0) k¡N0)
247
(S.35)
Equation (S.35) holds even when there are parameter perturbations in the plant transfer function
and the gains. Thus asymptotic tracking of any step reference input is robust. Note that this
robust tracking holds even for very large parameter perturbations as long as the overall system
remains stable.
We see that the design is achieved by inserting an integrator as shown in Fig. 8.4. The
integrator is in fact a model of the step reference input and constant disturbance. Thus it is
called the interal model principle. This will be discussed further in the next chapter.
8. 3. 2 Stabilization
If a state equation is controllable. all eigenvalues can be assigned arbitrarily by introducing
state feedback. We now discuss the case when the state equation is not controllable. Every
uncontrollable state equation can be transformed into
(8.36)
where · . is controllable (Theorem 6.6). Because the Amatrix is block triangular. the
eigenvalues of the original Amatrix are the union of the eigenvalues of · and · If we
introduce the state feedback
I¡ ~ r  l ~ r ¬ k = r · ·
·
where we have partitioned · as in X. then (8.36) becomes
·
=
·
.
.·
¸
¸
·
·
+
�
r (8. 37)
We see that · and. consequently. its eigenvalues are not afected by the state feedback. Thus
we conclude that the controllability co
n
dition of (A, . in Theorem 8.3 is not only suffcient
but also necessary to assign all eigenvalues of · .. to any desired positions.
Consider again the state equation in (8. 36). If · is stabie. and if · . i s controllable.
then (8.36) is said to be stabilizable. We mention that the conironability condition for tracking
and disturbance rejection can b replaced by the weaker condition of stabilizability. But in this
case. we do not have complete control of the rate of tracking and rejection. If the uncontrollable
stable eigenvalues have large imaginary pars or are close to the imaginary axis. then the
tracking and rejection may not be satisfactory.
8, 4 State Estimator
We introduced in the preceding sections state feedback under the implicit assumption that all
state variables are available for feedback. This assumption may not hold in practice either
248 STATE FEEDBACK AND STATE ESTIMATORS
because the state variables are not accessible for direct connection or because sensing devices
or transducers are not available or very expensive. In this case, in order to apply state feedback.
we must design a device, called a ..or .so that the output of the
device will generate an estimate of the state. I n this section, we i ntroduce fulldimensional
state estimators which have the same dimension as the original state equation. We use the
circumflex over a variable to denote an estimate of the variable. For example, x is an estimate
of .and i is an estimate of x.
Consider the ndimensional state equation
x = A.+ bll
(8. 38)
y = .
where A, b, and are given and the input 1I (t) and the output are available to us. The
state .however, is not available to us. The problem is to estimate .from h and y with the
knowledge of A, b, and If we know A and b, we can duplicate the original system as
i = Ax + bl ( 8. 39)
and as shown i n Fig. 8. 5. Note that the original system could be an electromechanical system
and the duplicated system could be an opamp circuit. The duplication will be called an .
.estimator. Now if (8. 38) and ( 8. 39) have the same initial state, then for any input. we have
x( t) = .for all t :: O. Therefore the remaining question is how to fnd the initial state of
(8. 38) and then set the initial state of ( 8. 39) to that state. If (8. 38) is observable, its initial state
.
can be computed from h and y over any time interval, say, [0, . We can then compute
the state at
and set X(t2) = X(2 ) . Then we have x) = X(I) for all :: 12. Thus if ( 8. 38) is
observable, an openloop estimator can be used to generate the state vector.
There are, however, two disadvantages in using an openloop estimator. First, the initial
state must be computed and set each time we use the estimator. This is very inconvenient.
Second, and more seriously, i f the matrix A has eigenvalues with positive real parts, then
even for a very small difference between .and x) for some l0. which may be caused by
disturbance or imperfect estimation of the initial state, the difference between X( I ) and x )
will grow with time. Therefore the openloop estimator i s, in general, not satisfactory.
We see from Fig. 8. 5 that even though the input and output of (8. 38) are available. we
¡ . . . . . . . . . . . . . . . ,
, ,
,
. .  . .  . .  .  .  . . .
Figure 8,5 Openloop state estimator.
Figure 8.6 Closedloop state estimator.
8.4 State Esti mator
, . . . . . . . . . . . . . . . ,
, ,
249
use only the input to drive the openloop estimator. Now we shall modify the estimator in Fig.
8. 5 to the one in Fig. 8.6, in which the output y(t ) = . of (8. 38) is compared with .
Their difference, passing through an n x I constant gai n vector I. i s used as a correcting term.
If the difference is zero, no correction is needed. If the difference is nonzero and i f the gain I
is properly designed. the difference will drive the estimated state to the actual state. Such an
estimator is called a ...or an ...estimator or. simply, an estimator.
The openloop estimator in ( 8.39) is now modifed as, following Fig. 8. 6.
i = Ax + bll + I(y  eX)
which can be written as
i = (A  Ie)x + bu + I I ( 8.40)
and i s shown in Fig. 8. 7. It has two inputs N and Y and its output yields an estimated state x.
Let us defne
e(t:= . x(
I t i s the error between the actual state and the estimated state. Differentiating e and then
substituting ( 8. 38) and (8.40) into it, we obtain
Figure 8.7 Closedloop state estimator.
, . . . . . . . . . . . . . .
250 STATE FEEDBACK AND STATE ESTI MATORS
or
e = i  i ¯ Ax + bu  (A  Ie)x  bu  I(cx)
= (A  Ie)x  (A  Ie)x = (A  Ie)(x  x)
e = (A  le)e
(8.41 )
This equation govers the estimation error. If all eigenvalues of ( A  Ie) can be assigned
arbitraly, then we can control the rate for e(t) to approach zero or, equivalently, for the
estimated state to approach the actual state. For example, if all eigenvalues of (A  Ie) have
negati ve real parts smaller than ¬O, then all entries of e will approach zero at rates faster
than 8
¬ºl
. Therefore, even if there is a large error between x(to) and x(to) at initial time to,
the estimated state will approach the actual !tate rapidly. Thus there is no need to compute
the initial state of the original state equation�In conclusion, if all eigenvalues of (A  Ie) are
properly assigned, a closedioop estimator is much more desirable than an openloop estimator.
As in the state feedback, what constitutes the best eigenvalues is not a simple problem.
Probably, they should be placed evenly along a circle inside the sector shown in Fig. 8. 3(a). If
an estimator is to be used in state feedback, then the estimator eigenvalues should be faster than
the desired eigenvalues of the state feedback. Again, saturation and noise problem
.
s will impose
constraints on the selection. One way to carry out the selection is by computer sImulation.
.. Teorem 8.03
Consider the pair (A, c). All eigenvalues of (A  Ie) can b assigned arbitraril y by selecting a real
constant vector I if and only if (A. c) is observable.
This theorem can be established directly or indirectly by using the duality theorem. The
pair (A, c) is observable if and only if (A', c') is controllable. If (A', c') is controllable. all
eigenvalues of (A'  c'k) can be assigned arbitrarily by selecting a constant gain vector k. The
transpose of (A'  c'k) is (A  k'c). Thus we have 1 = k'. In conclusion, the procedure for
computing state feedback gains can be used to compute the gain I in state estimators.
Solving the Lyapunov equation We discuss a different method of designing a state estimator
for the ndimensional state equation
i ¯ Ax + bu
(8.42)
y = cx
The method is dual to Procedure 8. 1 in Section 8. 2. 1 .
� Procedure 8.01
1. Select an arbitrary n X n stable matrix F that has no eigenvalues in common with those of A.
2. Select an arbitrary n x I vector I such that (F. \) is controllable.
3. Solve the unique T in the Lyapunov equation TA  I = Ie. This T is nonsingular following the
dual of Theorem 8.4.
4. Then the state equation
generates an estimate of x.
z = Fz + Tbu + Iy
x = TI z
We frst justify the procedure. Let us defne
e : = z  Tx
Then we have, replacing TA by FT + Ie,
8. 4 State Estimator
e = z  Ti = Fz + Tbu + lex  TAx  Tbu
= Fz lex  (FT + Ie)x = F(z  Tx) = Fe
251
(8.43)
(8.44)
IfF is stable. for any e(O), the error vector e(t) approaches zero as t ¬ 0. Thus z approaches
Tx or, equivalently, TI z is an estimate of x. All discussion in Section 8. 2. 1 applies here and
will not be repeated.
8. 4. 1 ReducedDi mensional State Estimator
Consider the state equation in (8.42). If it is observable, then it can be transformed, dual to
Theorem 8. 2. into the observable canonical form in (7. 14). We see that y equals XI , the frst
state variable. Therefore it is suffcient to construct an (n ¬ I )dimensional state estimator to
estimate Xi for i = 2. 3, . . . , n. This estimator with the output equation can then be used to
estimate all n state variables. This estimator has a lesser dimension than (8. 42) and is called a
reduceddimensional estimator.
Reduceddimensional estimators can be designed by transformations or by solving Lya
punov equations. The latter approach is considerably simpler and will be discussed next. For
the former approach, the interested reader is referred to Reference [6, pp. 361363].
� Procedure 8. Rl
1. Select an arbitrary (n  I ) X (n  I ) stable matrix F that has no eigenvalues in common with those
of A.
2. Select an arbitrary (n  I) x I vector I such that (F, I) is controllable.
3. Solve the unique T in the Lyapunov equation TA  FT = Ie. Note that T is an (n  I ) X n matrix.
4. Then the (n  I ) dimensional state equation
z = Fz + Tbu + Iy (8.45)
(8.46)
is an estimate of x.
We frst justify the procedure. We write (8.46) as
252 STATE FEEDBACK AND STATE ESTI MATORS
� =
T
x =: Px
which implies y = eX and z = Tx. Clearly V is an estimate of ex. We now show that z is an
estimate of Tx. Defne
e = z  Tx
Then we have
e = z  Tx = Fz + Tbll + Icx  TAx  Tbll = Fe
Clearly if F is stablt, then e(t) ~ 0 as t ¬ C. Thus z is an estimate of Tx.
� Theorem 8.6
If A and F have no common eigenvalues, then the square matrix
where T is the unique solution of TA  FT = Ie, is nonsingular if and only if (A, c) is observable and
(F. I) is controllable.
Proof' We prove the theorem for n = 4. The frst part of the proof follows closely the
proof of Theorem 8.4. Let
L(s) = det (sI ¯ A) = s· + al s3 + a
2
s
2
+ a3s + a4
Then, dual to (8.22), we have
I
o
o
(8. 47)
and L(F) is nonsingular if A and F have no common eigenvalues. Note that if A is 4 x 4.
then F is 3 x 3. The rightmost matrix in (8. 47) is the observability matrix of (A. c) and will
be denoted by O. The frst matrix after the equality is the controllability matrix of (F, I)
with one extra column and will be denoted by C. The middle matrix will be denoted by
A and is always nonsingular. Using these notations, we write T as L 
I
(F) CAO and P
becomes
P =
�
=
LI (;) CAO
[
=
� L�I (F)
[
C�O
[
(8. 48)
Note that if n = 4, then P, 0, and A are 4 x 4; T and C4 are 3 x 4 and L(F) is 3 x 3.
If (F, I) i s not controllable, C4 has rank at most 2. Thus T has rank at most 2 and P i s
singular, If (A, c) is not observable, then there exists a nonzero 4 x I vector r such that
8. 5 Feedback from Esti mated States 253
Or = 0, which implies cr = 0 and Pr = O. Thus P is singular. This shows the necessity
of the theorem.
Next we show the suffciency by contradiction. Suppose P is singular. Then there
exists a nonzero vector r such that Pr = 0, which implies
(8.49)
Defne H : = A Or = [a1 a
2
a
3
a4]' =: [a 0.]', where a represents the frst three entries of
H. Expressing it explicitly yields
I
o
o
1 cr
o cAr X
o cA
2
r
=
X
o cA
3
r cr
where X denotes entries that are not needed in subsequent discussion. Thus we have
U( = cr. Clearly (8.49) implies Uq = cr = O. Substituting UA = 0 into the lower part of
( 8.49) yields
(8.50)
where C is 3 x 3 and is the controllability matrix of (F, I) and a i s the frst three entries
of H. If (F, I) is controllable, then Ca = 0 implies a = O. In conclusion, (8.49) and the
controllability of (F, I) imply H = O.
Consider A Or ~ H = O. The matrix A is always nonsingular. If (A, c) is observable,
then 0 is nonsingular and AOr = 0 implies r = O. This contradicts the hypothesis that r
is nonzero. Thus if (A, c) is observable and (F, I) is controllable, then P is nonsingular.
This establishes Theorem 8.6. Q.E.D.
Designing state estimators by solving Lyapunov equations is convenient because the
same procedure can be used to design fulldimensional and reduceddimensional estimators.
As we shall see in a later section, the same procedure can also be used to design estimators for
multiinput multioutput systems.
8, 5 Feedback from Estimated states
Consider a plant described by the ndimensional state equation
x = Ax + bu
y = cx
(8. 5 1 )
If (A, b) i s controllable. state feedback u = r ¯ kx can place the eigenvalues of ( A  bk)
in any desired positions. If the state variables are not available for feedback, we can design a
state estimator. If (A, c) is observable, a full or reduceddimensional estimator with arbitrary
eigenvalues can be constrcted. We discuss here only fulldimensional estimators. Consider
the ndimensional state estimator
� = (A  Ie)x + bu + Iy (8.52)
25 STATE FEEDBACK AND STATE ESTIMATORS
The estimated state in (8. 52) can approach the actual state in (8. 5 1 ) with any rate by selecting
the vector I.
The state feedback is designed for the state in (8. 5 1 ) . If x is not available, it is natural to
apply the feedback gain to the estimated state as
u = r  k
( 8.53)
as shown in Fig. 8. 8. The connection is called the contrllerestimator confguration. Three
questions may be raised in this connection: ( 1 ) The eigenvalues of (A  bk) are obtained from
u = r  k. Do we still have the same set of eigenvalues in using u = r  k? (2) Will
the eigenvalues of the estimator be afected by the connection? (3) What is the efect of the
estimator on the transfer function from r to y? To answer these questions, we must develop
a state equation to describe the overall system in Fig. 8. 8. Substituting (8. 53) into (8. 5 1 ) and
(8. 52) yields
I
x = Ax  bkx + br
i = (A  Ie)x + b(r ¯ k) + lex
They can be combined as
`
[
=
� A �.�bk ,
.
[
+
´
[
r
y = [c 01
· ¦
(8.54)
This 2ndimensional state equation describes the feedback system in Fig. 8.8. It i s not easy
to answer the posed questions from this equation. Let us introduce the following equivalence
transformation:
Computing p
l
, which happens to equal p, and then using (4.26), we can obtain the following
equivalent state equation:
+
.
~
A �bk
A �
.
+
�
r
Y = [C 01 ' ¦
(8. 55)
x Figur 8.8 Controllerestimator confguration.
8.6 State FeedbackMul ti vari abl e Case 255
The Amatrix in (8. 55) is block triangular; therefore its eigenvalues are the union of those
of (A  bk) and (A  Ie). Thus inserting the state estimator does not affect the eigenvalues
of the original state feedback; nor are the eigenvalues of the state estimator affected by the
connection. Thus the design of state feedback and the design of state estimator can be carried
out independently. This is called the separation property.
The state equation in (8. 55) is of the form shown in (6.40); thus ( 8.55) is not controllable
and the transfer function of (8. 55) equals the transfer function of the reduced equation
or
x = (A  bk)x + br y = ex
(Theorem 6.6). This is the transfer function of the original state feedback system without using
a state estimator. Therefore the estimator is completely canceled in the transfer function from
r to y. This has a simple explanation. In computing transfer functions, all initial states are
assumed to be zero. Consequently, we have x(O) = x(O) = 0, which implies x(t) = x(t)
for all t. Thus, as far as the transfer function from r to y is concered, there is no difference
whether a state estimator is employed or not.
8. 6 State FeedbackMul ti vari abl e Case
This section extends state feedback to multi variable systems. Consider a plant described by
the ndimensional pinput state equation
x = Ax + Bu
y = Cx
In state feedback, the input U is given by
u = r  Kx
(8.56)
(8. 57)
where K is a p x n real constant matrix and r is a reference signal. Substituting (8. 57) into
(8. 56) yields
> Theorem 8.M1
x = (A  BK)x + Br
y = Cx
(8. 58)
Te pair (A  BK. B). for any p x n real constant matrix K, is controllable if and onl y if (A. B) is
controllable.
The proof of this theorem follows closely the proof of Theorem 8 . 1 . The only diference
is that we must modify (8.4) as
256 STATE FEEDBACK AND STATE ESTI MATORS
KB K(A  BK)B K(A  BK)
'
B
Ip KB K(A  BK) B
0 Ip KB
0 0 Ip
where Cf and C are n x np controllability matrices with n = 4 and Ip is the unit matrix of
order p. Because the rightmost 4p x 4p matrix is nonsingular. Cf has rank n if and only if C
has rank n . Thus the controllability property is preserved in any state feedback. As in the SISO
case. the observabil i ty property. however. may not be preserved. Next we extend Theorem 8. 3
to the matrix case
.
Theorem 8. M3
All eigenvalues of (A  BK) can be assigned arbitrarily (provided complex conjugate eigen\ clues are
assigned in pairs) by selecting a real constant K if and only if (A. B) is controllable.
If (A. B) i s not controllable. then (A. B) can be transformed into the form shown i n ( 8. 36)
and the eigenvalues of Ac wi l l not be affected by any state feedback. This shows the necessity
of the theorem. The suffciency will be established constructively i n the next three subsections.
8. 6. 1 Cyclic Design
In this design, we change the multiinput problem into a singleinput problem and then apply
Theorem 8. 3. A matrix A i s called cyclic if its characteristic polynomial equals its minimal
polynomial. From the discussion in Section 3. 6. we can conclude that A is cyclic if and only
if the Jordan form of A has one and onlv one Jordan block associated with each distinct
eigenvalue.
'
Theorem 8. 7
If the ndimensional pinput pai r (A, B) i s controllable and if A is cyclic. then for almost on) p X
vector Y. the singleinput pair (A. By) is control lable.
We argue intuitively the validity of this theorem. Controllabilitv is invariant under anv
equivalence transformation; thus we mav assume A to be in Jordan fo�. To see the basic ide�.
we use the following example:
.
A =
0 0 0
0 0 0
o
o
o
 I
o
There is only one Jordan block associated with each distinct eigenvalue; thus A is cyclic. The
8. 6 State FeedbackMul ti vari abl e Case 257
condition for (A. B) to be controllable is that the third and the last rows of B are nonzero
(Theorem 6. 8).
The necessary and suffcient conditions for (A. Bv) to be controllable are O 1 0 and
f 1 0 in ( 8. 59). Because C = ¡¡ + 2v� and f = VI , either O or f is zero if and only if
VI = 0 or VI / V� = 2/ 1 . Thus any y other than V I = 0 and l¡ = 2V2 will make ( A. By)
controllable. The vector ¹ can assume any value in the twodimensional real space shown in
Fig. 8.9. The conditions 1'1 = 0 and VI = 2v� constitute two straight lines as shown. The
probability for an arbitrarily selected y to lie on either straight line is zero. This establishes
Theorem 8.6. The cyclicity assumption in this theorem i s essential. For example. the pair
is controllable (Theorem 6.8). However. there i s no y such that ( A. By) is controllable
(Corollary 6. 8).
If all eigenvalues of A are distinct. then there i s only one Jordan block associated with
each eigenvalue. Thus a suffcient condition for A to be cyclic i s that all eigenvalues of A are
distinct.
Theorem 8.8
If (A. B) i s controllable. then for almost any p x n real constant matrix K, the matrix ( A  BK) hos
only distinct eigenvalues ond is. consequently. cyclic.
We show intuitively the theorem for 11 = 4. Let the characteristic polynomial of A BK be
cf ( s) = S4 + al s
3
+ a2s
�
+ a,s + a4
where the a, are functions of the entries of K. The differentiation of Sj ( s) with respect to s
yields
S. . s· = 4s
'
+ 3al s
"
+ 2a2s + a3
If cf ( s) has repeated roots. then .., ( s) and :.( s) are not coprime. The necessary and sufficient
condition for them to be not coprime is that their Sylvester resultant i s singular or
Figure 8.9 Twodimensi,)nal real space.
. Ii ¬ 0
258 STATE FEEDBACK AND STATE ESTIMATORS
a. a3 0 0 0 0 0 0
a3 2a
2
a. aJ 0 0 0 0
a
z 3a, a] 2a2 a . a] 0 0
det
a, 4 a2 3a, a3 2az a4 al
I 0 a, 4 a
2
3a, a3 2a2
= b(kij ) = 0
0 0 0 a, 4 a2 3a,
0 0 0 0 0 a, 4
0 0 0 0 0 0 0
See (7. 28). It is clear that all possible solutions of b(kij ) = 0 constitute a very small subset of
�ll real kij . Thus if we select an arbitrary K, the probability for its entries to meet b(kij) = °
!o O. Thus all eigenvalues of (A  BK) wi\ be distinct. This establishes the theorem.
With these two theorems, we can noW fnd a K to place all eigenvalues of (A  BK) in
any desire�positions. If A is not cyclic. we introduce U = w  K, x, as shown in Fig. 8. 1 0,
such that A := A  BK, in
x = (A  BK, )x + Bw =: Ax + Bw (8. 60)
is cyclic. Be
<
ause (A, B) is controllable, so is (A. B). Thus there exists a p X I real vector
v such that (A. Bv) is controllable.2 Next we introduce another state feedback w = r  K,x
with K2 = vk. where k is a I x n real vector. Then (8.60) becomes

x = (A  BKz)x + Br = (A  Bvk)x + Br
Because the singleinput pair (A, Bv) is controllable. the eigenvalues of (A  Bvk) can
Ï
Figure 8.10 State feedback by cyclic design.
Z.Te c�oices of Àj and Vare not unique. They ClnDchosen arbitrarily and the probability is l that they will meet
the reqUirements. In Theorem 1.0 of Reference [5J . a prot!dure is given K choose Àj and Y with no uncertainty.
The computation. however, is complicated.
.
¯
8. 6 State FeedbackMul tivariable Case 259
be assigned arbitrarily by selecting a k (Theorem 8. 3). Combining the two state feedback
U = w  K, x and w = r  K2x as
we obtain a K := K, + K
2
that achieves arbitrary eigenvalue assignment. This establishes
Theorem 8.M3.
8.6. 2 LyapunovEquation Method
This section will extend the procedure of computing feedback gain i n Section 8. 2. 1 to the
multivariable case. Consider an ndimensional pinput pair (A. B) . Find a p X n real constant
matrix K so that (A BK) has any set of desired eigenvalues as long as the set does not contain
any eigenvalue of A.
; Procedure 8. Ml
1. Select an n x n matrix F with a set of desired eigenvalues that contains no eigenvalues of A.
2. Select 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. If T i� singular. select a different K and repeat the process. If T is nonsingular. we compute
K = KT' ¡ and (A  BK) has the set of desired eigenvalues.
If T is nonsingular. the Lyapunov equation and KT = K imply
(A  BK)T = TF or A  BK = TFr'
Thus (A  BK) and F are similar and have the same set of eigenvalues. Unlike the SISO case
where T is always nons in gular, the T here may not be nonsingular even if (A. B) is controllable
and (F, K) is observable. In other words, the two conditions are necessary but not suffci ent
for T to be nonsingular.
; Theorem 8,M4
If A and F have no eigen\ alues in comm
_
on, then the unique solution T of AT TF = BK is nonsingular
only if (A. B) is controllable and ( F, K) is observable.
�
Proof: The proof of Theorem 8.4 applies here except that (8. 22) must be modifed as. for
n = 4.
�
o . KFJ
or
260 STATE FEEDBACK AND STATE ESTIMATORS
 Tc(F) = CEO
where 6(F) is nonsingular and C, E, and 0 are. respectively. n X np. np x IIp. and
np X n. If C or 0 has rank less than n, then T is singular following However.
the conditions that C and 0 have rank n do not imply the nonsingularity of T. Thus the
controllability of (A. B) and observability of ( F. K) are only necessary conditions for T
to be nonsingular. This establishes Theorem . .
Given a controllable (A. B). i t i s possible to construct an observable (F. K) so that the
T in Theorem S.M4 is singular. However. after selecting F. if K is selected randomly and i f
(F. K) i s observable. it is believed that the probability for T to be nonsingular is Therefore
solving the Lyapunov equation is a viable method of computing a feedback gain matrix to
achieve arbitrary eigenvalue assignment,' As in the SISO case, we may choose F in companion
form or in modal form as shown i n · If F i s chosen as in then we can select K as
K =
o
� .
or
K =
o
o o
o
�
(see Problem Once F and K are chosen, we can then use the MATLAB function lyap
to solve the Lyapunov equation. Thus the procedure can easily be carried out.
8. 6. 3 CanonicalForm Method
We i ntroduced i n the preceding subsections two methods of computing a feedback gain matrix
to achieve arbitrary eigenvalue assignment. The methods are relatively simple: howe\ er. they
will not reveal the structure of the resulting feedback system. In this subsection. we discuss
a different design that will reveal the effect of state feedback on the transfer matrix. We also
gi ve a transfer matrix interpretation of state feedback.
I n this design. we must transform (A. B) i nto a controllable canonical form. It is an
extension of Theorem ·to the multi variable case. Although the basic idea is the same, the
procedure can become very i nvolved. Therefore we wi l l skip the details and present the fnal
result. To simplify the discussion. we assume that has dimension two inputs. and two
outputs. We frst search linearly independent columns of C = [B AB
. . .
ASB] i n order from
left to right. It is assumed that its controllability indices are 11
1 = and 12 = 2. Then there
exists a nonsingular matrix P and x = P wi l l transform into the controllable canonical
for
al t t 01 1 2 al l )

a t t �
C ! :l Q122
X
=
x
al i i

all : a] 1 J
=Ô_ l + a221 C222
8. 6 State FeedbackMul tivariable Case 261
I bl 2
+ U
·
I
y =
11 1 1 11 1 2 11 1 3 11 1 . 11 21 11 22
X
121 1 121 2 121 J 121 � 1221 {h
Note that this form is identical to the one in
We now discuss how to fnd a feedback gain matrix to achieve arbitrary eigenvalue
assignment. From a given set of six desired eigenvalues. we can form
lf(S) = (5� + al l l S
3
+ al 1 2S
2
+ ai DS + (1 1 4) (s
2
+ al2 1 S + (222)
Let us select K as
K =
o
a' 1 2  C I 1 2
a21 2  eZ1 2
al 1 4  01 1 4 01 21
a214  a21 . a221  a22 1
Then i t i s straightforward to verify the following
01 1 1 01 1 2 am al l �
A 
BK =
a2 1 1 a21 2 a21 J a21 4
.
al J 3  al t J
a21 3  021 3
a 122
a222  a222
a221 a222
Because (A 
BK) is block triangular, for any a21 1 ' i = its characteristic polynomial
equals the product of the characteristic polynomials of the two diagonal blocks of orders
and ·Because the dia"onal blocks are of companion form, we conclude that the characteristic
& . ."    I
polynomial of (ABK) equals the one in I f K = KP, then ( ABK) = P(ABKJP .
Thus the feedback gain K = KP wi l l place the eigenvalues of (A BK) in the desired locations.
This establishes once again Theorem
Unlike t he singleinput case. where t he feedback gai n i s unique, t he feedback gain matrix
in the multiinput case is not unique. For example. the K in �S.6) yields a lower blocktriangular
matrix in (A  BK). It is possible to select a di ferent K to yield an upper blocktriangular
matrix or a blockdiagonal matrix. Furthermore, a different grouping of wi l l again yield
a diferent K.
262 STATE FEEDBACK AND STATE ESTIMATORS
8. 6. 4 Effect on Transfer Matrices
'
In the singlevariable case. state feedback can shift the poles of a plant transfer function g(s) to
any positions and yet has no effect on the zeros. Or. equivalently, state feedback can change the
denominator coeffcients, except the leading coeffcient 1 , to any values but has no efect on the
numerator coeffcients. Although we can establish a similar result for the multi variable case
from (8. 62) and (8.65), it is instructive to do so by using the result in Section 7. 9. Following
the notation in Section 7.9, we express ((s) = C(sI  A)
I
B as
or
yes) = N(�) D
I
(s)u(s)
where N (s) and D(s) are right coprime and D(s) is column reduced. Defne
D(s)v(s) = u(s)
as in (7. 93). Then we have
Yes) = N(s) v(s)
Let H(s) and L( s) be defned as in (7. 91 ) and (7. 92). Then the state vector in (8.62) i s
xes) = L(s)v(s)
Thus the state feedback becomes, in the Laplacetransform domain,
u(s) = res)  Kx(s) = res)  KL(s)v(s)
and can be represented as shown in Fig. 8. 1 1 .
Let us express D(s) as
D(s) = DhcH(S) + DleL(S)
Substituting (8. 71 ) and (8.70) into ( 8. 68) yields
[DhcH(s) + Dle L(S)] yes) = res) ¯ KL(s)v(s)
which implies
(8.66)
(8. 67)
(8. 68)
(8. 69)
(8.70)
(8. 71 )
Figur 8.11 Transfer matrix
interpretation of state feedback.
3. This subsection may b skipped without loss of continuity. The material in Section 7.9 is needed to study this
subsection.
8. 7 State EstimatorsMul ti vari abl e Case 263
[D;'cH(s) + (Die + K)L(s)] yes) = res)
Substituting this into (8.69) yields
yes) = N(s) [DheH(S) + (Die + K)L(sJ
I
res)
Thus the transfer matrix from r to y is
(8. 71)
The state feedback changes the plant transfer matrix N(s)D
I
(s) to the one in ( 8. 71). We see
that the numerator matrix N(s) is not affected by the state feedback. Neither are the column
degree H(s) and the columndegree coeffcient matrix Dhe afected by the state feedback.
However, all coeffcients associated with L(s) can be assigned arbitrarily by selecting a K.
This is similar to the SISO case.
It is possible to extend the robust tracking and disturbance rejection discussed in Section
8.3 to the multi variable case. It is simpler, however, to do so by using coprime fractions:
therefore it will not be discussed here.
8. 7 State EstimatorsMultivari abl e Case
All discussion for state estimators i n the singlevariable case applies to the multi variable
case; therefore the discussion will be brief. Consider the ndimensional pinput q output state
equation
x = Ax + Bu
(8. 73)
y = Cx
The problem is to use available input u and output y to drive a system whose output gives an
estimate of the state x. We extend (8.40) to the multi variable case as
i = (A  LC)x + Bu + Ly (8.74)
This is a fulldimensional state estimator. Let us defne the error vector as
eel) := x(t)  xU) (8. 75)
Then we have, as in (8. 41 ) .
e = (A  LC)e (8. 76)
If (A, C) is observable. then all eigenvalues of (ALC) can be assigned arbitrarily by choosing
an L. Thus the convergence rate for the estimated state x to approach the actual state X can be
as fast as desired. As in the SISO case, the three methods of computing state feedback gain K
in Sections 8. 6. 1 through 8. 6. 3 can be applied here to compute L.
Next we discuss reduceddimensional state estimators. The next procedure is an extension
of Procedure S.R l to the multivariable case.
264 STATE FEEDBACK AND STATE ESTI MATORS
Procedure 8. MRI
Consider the IIdimensional qoutput observable pai r · c· It i s assumed that chas rank q.
1. Select an arbitrary ( I ¯ . X . stable matrix F that has no eigenvalues i n common wi t h those
of ·
2. Select an arbitrary ( II ¯ . X . matri x such that ( F. · i s controllable.
3. Solve the unique . x n matrix 1i n the Lyapunov equation 1\ r1c
·t If t he square matrix of order l
P
�
( 8. 77)
i s singular. go back t o Step 2 and repeat t he p
�
)cess. If P i s nonsingular. then the (n . )dimensional
state equation
: r:+ 1ea+ · ( 8. 78)
( 8. 79)
generates an estimate of ·
.. .. . ( S. 79) ..
·cs.. .1s .. y . . .. c· . . . :. .
. 1· . .
. . ..
 . 1·
e : ¯ 1·r:+ 1ea+ c· 1a· 1ea
r:+ c 1a· ·r.· 1·· ¬ Fe
F . .. t) ¯ 0 U5  ¬ ¯. ... Z . . .. 1·
Theorem 8. M6
I f A and F have no common eigenvalues. then the square matrix
P : = �
where T i s the unique solution of 1a r1 cis nonsingular only if (A. c·i s observable and
( F. ·i s controllabl e.
.. . . . . . .. S' !v14 .. 8. 6.
· 8. 6, .. . ..., .. .. P ...
8. 8 Feedback from Esti mated StatesMul ti vari abl e Case 265
.. . . .... · c· . . . . .
(F. · . . P . .... . . .. F.  . .. .. .. ( F. ·
. . . . . .. P ... . I .
8, 8 Feedback from Estimated StatesMul tivari abl e Case
.. . . . .. . .... S. 5 . ..
.. . ..... .. .. . .
.. ... .. ..
. ··+ ea
(8. 80)
·c·
.. . ... .. .. ( 8. 78) .. ( S. 79) . . . .
· ( 8. 77) .. . .. o o·  ,
n X . .. , . X ( n q) : . .
. . ... .. .. ( 8.78) .. ( S.79 ) . .
: r:+ 1ea·
s o ·  o·:
( 8. 8 1 )
( S 82)
( 8. 83)
.. .. . ... .. .. .. .. .. . s
..
a ks ko · ko·.
.... . ( 8. 80) .. ( S. S2) ...
. a·+ e  ko c· ko·:·
·  eko c· · eko·:+ e
: r.+ 1e ~ ko c· ko· . = c·
c 1eko c· ·+ ( F ~ 1eko· ·. 1e
.. . . ..
.
· eko c
:
c 1eko c
eko
·
e
r 1exo .
¯
1e

· c ·
· `
( S. 84)
( 8. 85)
( 8. 86)
( 8 87)
. . (2n . . .. .. ... .. ... ... . 8. 8. ·.
.. .. ., . .. .... ...
266 STATE FEEDBACK AND STATE ESTI MATORS
After some manipulation and using TA  FT = and we can fnally obtain the
following equivalent state equation
[
�
]
=
[
A 
O
BK

B
�Q2
] [:]
[
�]
I
y = [C
01 [:]
This equation i s similar t o
for the singlevariable case. Therefore all discussion there
applies. without any modifcation. to thb multi variable case. In other words, the design of a
state feedback and the design of a state estimator can be carried out independently. This is
the separation propert. Furthermore, all eigenvalues of F are not controllable from I and the
transfer matrix from r to y equals
8.1
8.2
8.3
8.4
8.5
G(s) ~ ¯
A
BK) I B
Given
= .
fnd the state feedback gain k so that the state feedback system has and as its
eigenvalues. Compute k directly without using any equivalence transformation.
Repeat Problem
by using
Repeat Problem by solving a Lyapunov equation.
Find the state feedback gain for the state equation
x =
�
�
so that the resulting system has eigenvalues and  I ± j \ . Use the method you think
is the simplest by hand to carry out the design.
Consider a system with transfer function
,
(s
g(s) =
I) (s
Is it possible to change the transfer function to
gj(s) ~
(s
by state feedback? Is the resulting system BIBO stable? Asymptotically stable?
8.6
8.7
8.8
8.9
8.10
Consider a system with transfer function
=
(s 2)
Is It possible to change the trans"e f
.
r unctIOn to
, I
gj
(S) ¯
_
Problems
by state feedback? Is the resulting system BIBO stabl ? A
.
.
e . symptotIcally stable?
Consider the continuoustime state equation
267
Let H  p k F' d
.
 r ¯ X. In the feedforward ain
resulting system has eigenvalues ~
and _
g
l
±
P and state feedback gain k so that the
reference input.
} I and wIll track asymptotically any step
Consider the discretetime state equation
. �
.
= 0 .
Find the state feedback gain so that the I .
Show that for any initial state th
.
resu tlng system has all eigenvalues at = 0
'd
.
' e zeroInput response f h " d
� ·
I entIcally zero for
Õ t e lee back system becomes
Consider the discretetime state equation in Probl
where p is a feedforward gain For th k ' P
em Let = pr[k] ..
. e In roblem f d
.
*
WI track any step reference input Sh I
. Û a gain p so that the output
tr k "
. ow a so that
ac ng IS achieved in a fnite numb f
 r lor :: 3. Thus exact
Th"
"
er Õ sampling d '
IS IS possIble If all poles of the resultinG
peno s Instead of asymptotically.
deadbeat design.
" system are placed at : ~ This i s called the
Consider the uncontrollable state equation
.
.
0 0 0 1
Is It possible to fnd a gain k so that the
. .
eigenvalues ¯2 I I ? I '
, equatIOn with state feedback H = r ¯ ..h
• •
S It possIble to h
.
as
about ¯2 "2 ~?¯ Is th '
ave eigenvalues PH
. ¡ . & « e equatIOn stabilizable?
' . . . ow
268 STATE FEEDBACK AND STATE ESTI MATORS
8. 11 Design a fulldimensional and a reduceddimensional state estimator for the state equa
tion in Problem 8. 1 . Select the eigenvalues of the estimators from { 3, 2 ± j21 .
8. 12 Consider the state equation i n Probl em 8. 1 . Compute the transfer function from l t o y of
the state feedback system. Compute the transfer function from l to y if the feedback gain
i s applied to the estimated state of the fulldimensional estimator designed i n Problem
8. 1 1 . Compute the transfer function from r to y if the feedback gain i s applied to the
estimated state of the reduceddimensional state estimator also designed in Problem 8. 1 1 .
Are the three overall transfer functions the same?
8.13 Let
A =
Find two diferent constant matrices K such that (A ¯ BK) has eigenvalues 4 ± 3j
and 5 ± 4j.
9. 1 I ntroducti on
LH3Dl0I
Pol e Pl acement and
Model Matchi ng
We frst give reasons for introducing this chapter. Chapter 6 discusses statespace analysis
(controllability and observability) and Chapter 8 introduces statespace design (state feedback
and state estimators). In Chapter 7 coprime fractions were discussed. Therefore it is logical to
discuss in this chapter their applications in design.
One way to introduce coprime fraction design is to develop the Bezout identity and to
parameterize all stabilization compensators. See References [3. 6. 9. 1 3. 20). This approach
is important in some optimization problems but is not necessarily convenient for all designs.
See Reference [ 8) . We study i n this chapter only designs of minimumdegree compensators
to achieve pole placement and model matching. \Ve will change the problems into solving
linear algebraic equations. Using only Theorem 3.2 and its corollary. we can establish all
needed results. Therefore we can bypass the Bezout identity and some polynomial theorems
and simplify the discussion.
:lost control systems can be fonnulated as shown in Fig. 8. 1 . That i s. given a plant with
input /I and output y and a reference signal l. design an overall system so that the output V
will follow the reference signal r as closely as possible. The plant input / I is also called the
actuating signal and the plant output V. the controlled signal. If the actuating signal ll depends
only on the reference signal l as shown in Fig. 9. I ( a). it i s called an openloop control. If l!
depends on r and V. then it is called a closedloop or feedback control. The openloop control
is. in general. not sati sfactory if there are plant parameter variations due to changes of load.
environment. or aging. It is also very sensitive to noise and disturbance. which often exist in
the real world. Therefore openloop control is used less often in practice.
269
270 POLE PLACEMENT AND MODEL MATCHI NG
.
The simplest is the unityfeedback
There are many possible feedback confguratIOns.
.
and the compensator with
f
.
h " Fig 9 I (b) in which the constant gam P con guratlon s own m . '
transfer function Cis) are to be designed. Clearly we have
u (s) = C(s)[pr(s)  yes)]
(9. 1 )
.
I and the plant output V drive essentially the same
Because p is a constant, the reference signa r
.
�
.
d to have one decree
. .
al Thus the confguratIOn IS sal " compensator to generate an actuatmg sign .
d of freedom
I h i Õ confguration also has one egree .
of freedom. Clear y t e open Õ p
F ° 8 8 can be redrawn as shown
The connection of state feedback and state estlmator m Ie' .
in Fig. 9. l (c). Simple manipulation yields
1
)  ( ) res u s =
1 + C
I
(S)'
C2(S)
\' (s)
I + C
I
(s) '
(9. 2)
t cenerate a u Thus the confguration
We see that r and y drive two independent compensators Õ _ .
is said to have two degrees of freedom.
.
be obtained by modifying A more natural twodegreeoffreedom confguratIOn can
.
(9. 1 ) as
u(s ) = Cl (s)r(s)  C2(S)y(s)
(9. 3)
. .
'
.
.
ost eneral control signal because each of r and y
and is plotted m Fig. 9. l ( d).
TI
s
IS the m g
d
. .
Thus no confguration has three h· h h freedom m eSlgmng. drives a compensator, w IC we ave
.
d f freedom confgurations; see, for Th pOSSible two egreeo 
degrees of freedom. ere are many
.
F 9 l ed) the twoparameter confguration;
example, Reference [ 12]. We call the one m ig. .
. UtOllt litfeedback confguration.
the one in Fig. 9. I (c) the control
l
er

es
n
mator or p
l
an
b
tm
P
ore na
�
ral and more suitable for
f ration seems to e m Because the twoparameter con gu
thl
'
s chapter For designs using the
. ,
d 1 this conticuratlOn m .
practical apphcatlOn, we stu Y on Y C
6]
plantinputoutputfeedback confguration, see Reference [ .
�
�
�
C(s)
Cis) g(s)
(a)
(e)
Figure 9.1 Control confgurations.
¦
g(s)
(bl
(d)
9. 1 Introduction 27 1
The plants studied in this chapter wi l l be limited to those describable by strictly proper rational functions or matrices. We also assume that every transfer matrix has ful l rank in the
sense that if G(s) i s q x p, then it has a q x q or p x p submatrix with a nonzero determinant. If G(s) i s square, then its determinant i s nonzero or its inverse exists. This i s equivalent to the assumption that if (A, B. C) is a minimal realization of the transfer matrix, then B has full column rank and C has ful l row rank.
The design to be introduced in this chapter i s based on coprime polynomial fractions of rational matrices. Thus the concept of coprimeness and the method of computing coprime fractions introduced i n Sections 7. 1 through 7.3 and 7.6 through 7.8 are needed for studying this chapter. The rest of Chapter 7 and the entire Chapter 8, however, are not needed here. In this chapter, we will change the design problem into solving sets of linear algebraic equations. Thus the method i s called the linear algebraic method i n Reference [7].
For convenience, we frst introduce some terminology. Every transfer function g(s ) = N(s)j D(s) is assumed to be a coprime fraction. Then every root of D(s) is a pole and every root of N (5) i s a zero. A pole i s called a stable pole i f i t has a negative real part; an unstable pole if it has a zero or positive real part. We also defne
• Minimumphase zeros: zeros with negative real parts
• Nonminimumphase zeros: zeros with zero or positive real parts
Although some texts call them stable and unstable zeros, they have nothing to do with stability.
A transfer function with only minimumphase zeros has the smallest phase among all transfer
functions with the same amplitude characteristics. See Reference [7. pp. 284285] . Thus we
use the aforementioned terminology. A polynomial is called a Hum'it� polynomial if all its
roots have negative real parts.
9. 1 . 1 Compensator EquationsClassical Method
Consider the equation
A( s) D(s) + B(s) N(s) = F(s)
(9. 4)
where D(s) . N(s) . and F(s) are given polynomials and A(s) and B(s ) are unknown poly nomials to be solved. Mathematically speaking. this problem i s equivalent to the problem of solving integer solutions A and B in AD + BN ~ F, where D, N. and F are given integers. This i s a very old mathematical problem and has been associated with mathematicians such as Diophantine. Bezout, and Aryabhatta. ` To avoid controversy. we follow Reference [3] and call it a compensator equation. Al l design problems in this chapter can be reduced to solving compensator equations. Thus the equation i s of paramount importance.
We frst discuss the existence condition and general solutions of the equation. What will be discussed, however, i s not needed i n subsequent sections and the reader may glance through this subsection.
1. See Reference [`1. la�t page of Preface].
272 POLE PLACEMENT AND MODEL MATCHI NG
Theorem 9. 1
Given polynomials D( s) and N(s) . polynomial solutions A(s) and B( s) exi st i n (9. ) for any
polynomial F( s) i f and only if D(s ) and N(s) are coprime.
Suppose D( s) and N (5 ) are not coprime and contain the same factor 5 + a. Then the
factor 5 + a wi ll appear in F(s ) . Thus if F( s) does not contain the factor. no solutions exist in
(9
.
4). This shows the necessity of the theorem.
If D(s ) and N(s) are coprime. there exist polynomials A(s) and B ( s ) such that
A (s ) D(s ) + B (s) N(s ) = 1
(9. 5)
Its matrix version is called the Be:ollt identit i n Reference [ 1 3]. The polynomials A( s) and
B(s) can be obtained by the Euclidean algorithm and wi l l not be discussed here. See Reference
[6. pp. 578580]. For example, if D(s ) = 52  I and N(s) = s  2, then A (s ) = 1 /3 and
B(s) = (5 + 2)/3 meet (9. 5). For any polynomial F( s) . (9.5) implies
F(s ) A(s) D(s) + F( s) B(s) N(s) = F( s) (9. 6)
Thus A(s) = F(s ) A(s) and B(s ) = F(s ) B(s) are solutions. This shows the suffciency of the
theorem.
Next we discuss general solutions. For any D(s ) and N( s) . there exist two polynomials
A(s) and B( s) such that
A(s ) D(s ) + B(s ) N(s ) = 0
(9. 7)
Obviously A(s)
Q(s) .
N(s ) and B( s) = D( s) are such solutions. Then for any polynomial
A( s) = A(s ) F! s ) + Q(s ) A(s ) B( s ) = B(s) F(s) + Q(s ) B(s ) ( 9. 8)
are general solutions of ( 9. 4) . This can easily be verifed by substituting ( 9. 8) into ( 9. 4) and
using (9. 5) and (9.7) .
EXAMPLE 9,1 Given D( s) = 5"  1 . N( s) = 5  2, and F! s) = 53 + 4s" + 6s + 4. then
A(s ) ¬ � ( S3 + 4s2 + 65 + 4) + Q(s) ( s + 2)
B( s) = � (s + 2 ) (53 + 4s" + 65 + 4) + Q( s) (s"  1 ) (9 9)
for any polynomial Q( s) , are solutions of ( 9.4).
Although the classical method can yield general solutions. the solutions are not necessarily
convenient to use in design. For example, we may be interested in solving A( s) and B(s) with
least degrees to meet (9.4). For the polynomials in Example 9. 1 . after some manipulation, we
tind that if Q(s ) = (s2 + 6s + 1 5) /3, then (9.9) reduces to
A(s ) ~ s + 34/3 B(s) ~ ( 22s  23 ) /3 (9. 1 0)
9. 2 Uni tyFeedback Confi gurati onPol e Pl acement 273
They are the leastdegree solutions of Example 9. 1 . In this chapter. instead of solving the
compensator equation directly as shown, we will change it into solving a set of linear algebraic
equations a5 in Section 7. 3. By so doing, we can bypass some polynomial theorems.
9, 2 UnityFeedback ConfigurationPole Pl acement
Consider the unityfeedback system shown in Fig. 9. 1 (b). The plant transfer function g(s) is
assumed to be strictly proper and of degree 11 . The problem i s to design a proper compensator
C (s) of least possible degree m so that the resulting overall system has any set of n + m desired
poles. Because all transfer functions are required to have real coeffcients, complex conjugate
poles must be assigned in pairs. This will be a standing assumption throughout this chapter.
Let g( s) = N(s) / D( s) and Ci s) = B( 5) / A( s) . Then the overall transfer function from
r to y in Fi g. 9. 1 (b) i s
pC(s)g( s)
go es) = 
1 + C(s) g( s)
pB( s) N(s)
B( s) N(s )
p
A(s ) D( s)
B( s) N(s)
1 +  
A(s) D(s)
A(s ) D(s) + B( s) N(s)
(9. 1 1 )
In pole assignment, we are interested in assigning all poles of go es) or. equivalently. all roots
of A(s ) D(s ) + B(s) N(s)
.
In this design, nothing is said about the zeros of go (s) . As we
can see from (9. 1 1 ), the design not only has no effect on the plant zeros (roots of N (s))
but also introduces new zeros (roots of B(s» into the overall transfer function. On the other
hand, the poles of the plant and compensator are shifted from D( s) and A(s ) to the roots of
A(s ) D( s) + B( s) N(s) . Thusfeedback call shift poles bl has HO efect on :ers.
Given a set of desired poles, we can readily form a polynomial F(s ) that has the desired
poles as its roots. Then the poleplacement problem becomes one of solving the polynomial
equation
A(s ) D(s) + B(s ) N(s) = F( s) (9. 1 2)
Instead of solving (9. 1 2) directly, we will transform i t into solving a set of linear algebraic
equations. Let deg N(s) < deg D(s) = n and deg B( s) : deg A( s) = m. Then F(s ) in (9. 1 2)
has degree at most n + m. Let us write
D(s) = Do + DI S + D:s: +
. . . + D"s" D" ; 0
N( s) = No + Ni s + N:s2 +
.
.
.
+ N"s"
A(s ) = Ao + Ai s + A2s' +
.
.
.
+ Amsm
B(s) = Bo + B1 5 + B2S2 +
. . .
+ B
m
sm
F( s) ~ Fo + F1 s + F"s2 +
. .
. + F
n+
ms
n+
m
where all coeffcients are real constants. not necessarily nonzero. Substituting these into (9. 1 2)
and matching the coeffcients of like powers of s, we obtain
274
POLE PLACEMENT AND MODEL MATCHI NG
Ao Do + BoNo = Fo
AoDI + Bo NI + AI Do + BI No = FI
There are a total of (n + m + 1) equations. They can be arranged in matrix form as
9. 2 UnityFeedback ConfigurationPole Placement 275
number of rows increases by 2. Because Dn � O. the new D row is linearly independent of its
preceding rows. Thus the 2(n + I ) x (2n + 1 ) matrix Sn has rank (2n + I ) (full column rank).
Repeating the argument, we conclude that if D(s) and N(s) are coprime and if m : n  L
then the matrix Sm in (9. 1 4) has full column rank.
' Theorem 9.2
Consider the unityfeedback system shown in Fig. 9. I (b). The plant is described by a strictly proper
transfer function g(s) = N(s)/ D(s) with N( s) and D(s) coprime and deg N(s) < deg D(s) = n.
(9. 13)
Let m :: n  1. Then for any polynomial F(s) of degree (n + m), there exists a proper compensator
C(s) = 8(s)/ A(s) of degree m such that the overall transfer function equals
(9. 1 4)
o 0 0 Do Dn
o 0 0 No Nn
I we take the transpose of (9. 1 3), then it becomes the standard form studied in Theorems 3. 1
and 3.2. We use the form i n (9. 13) because it can be extended directly to the matnx case .
.
The
matrix S has 2(m + 1 ) rows and (n + m + 1) columns and is formed from the coeffcients
of D(s) �d N(s) . The frst two rows are simply the coeffcients of D(s) and N(s) arranged
i ascending powers of s. The next two rows are the frst two rows s�ifted to the nght by �ne
position. We repeat the process until we have (m + 1) sets of coeffClents. The lefthan�slde
row vector of (9. 1 3) consists of the coeffcients of the compensator C(s) to be solved. It C(s)
has degree m, then the row vector has 2(n! + 1 ) entries. The righthandslde row vector of
(9. 1 3) consists of the coeffcients of F(s) . Now solving the compensator equatIOn U (9. 1 2)
becomes solving the linear algebraic equation in (9. 13).
.
Applying Corollary 3. 2, we conclude that (9. 1 3) has a solutio
.
n for any F( s) �f and only
if Sm has full column rank. A necessary condition for Sm to have full column rank l5 that Sm
is square or has more rows than columns, that is,
2(m + 1) :: n + m + 1 or m :: II  I
If m < n  1 , then Sm does not have full column rank and solutions may exist for some F(s).
but not for every F(s). Thus if the degree of the compensator is less than n  1, lt lb not pOSSible
to achieve arbitrary pole placement.
.
If m = n ¬ 1 , SnI becomes a square matrix of order 2n. It ]5the transpose of the Sylvester
resultant in (7.28) with n ~ 4. As discussed in Section 7. 3, S,,_I is nonsillgular If and only If
D(s) and N(s) are coprime. Thus if D(s) and N(s) are coprime, then SnI has rank 2n (full
column rank). Now if m increases by 1. the number of columns increases by I but the the
go es) =
pN(s ) B(s) pN(s) B(s)
A(s ) D(s) + B(s)N(s) F(s)
Furthermore, the compensator can b obtained by solYing the linear algebraic equation in (9. 1 3).
As discussed earlier, the matrix Sm has full column rank for m :: n  I ; therefore, for
any (n + m) desired poles or. equivalently, for any F(s) of degree (n + m) , solutions exist in
(9. 1 3). Next we show that B( s)/ A(s) is proper or Am � O. I f N(s)/ D(s) is strictly proper.
then Nn = 0 and the last equation of (9. 13) reduces to
Because F(s) has degree (n + m) , we have Fn+m � 0 and, consequently, Am � O. This
establishes the theorem. If m = n ¬ 1 , the compensator is unique; if m > n  1 , compensators
are not unique and free parameters can be used to achieve other design objectives, as we will
discuss later.
9. 2. 1 Regulation and Tracking
Pole placement can be used to achieve the regulation and tracking discussed in Section 8.3. In
the regulator problem, we have r = 0 and the problem is to design a compensator C (s) so that
the response excited by any nonzero initial state will die out at a desired rate. For this problem,
if all poles of goes) are selected to have negative real parts, then for any gain p, in particular
p = I (no feedforward gain is needed). the overall system will achieve regulation.
We discuss next the tracking problem. Let the reference signal be a step function with
magnitude a. Then res) = a/s and the output yes) equals
yes) = go (s)r( s) = go(s)
�
s
If gols) is BIBO stable, the output will approach the constant gv(O)a (Theorem 5. 2). This can
also be obtained by employing the fnalvalue theorem of the Laplace transform as
lim y(t) = lim s)'(s) = go(O)a
tO x~0
Thus in order to track asymptotically any step reference input, go es) must be BIBO stable and
go(O) = L The transfer function from r to y in Fig. 9. 1 (b) is goes) ~ pN(s ) B(s) / F(s) . Thus
we have
276 POLE PLACEMENT AND MODEL MATCHI NG
*
••!
ì
·
]
1
�
j
j
Í
1
which implies
• NO BO
= p
p =
BoNo
BoNo
= p
Fa
Thus in order t o track any step reference input, we require Bo ¥ and No ¥ O. The constant
Bo is a coeffcient of the compensator and can be designed to be nonzero. The coeffcient No
is the constant ter of the plant numerator. Thus if the plant transfer function has one or more
zeros at = then No = and the plant cannot be designed to track any step reference input.
This is consistent with the discussion in Section
If the reference signal is a ramp function or = for then using a similar
argument, we can show that the overall trahsfer function goes) must be BIBO stable and has
the properties go(O) = and g�(O) = (Problems and This is summarized in the
following.
• Regulation¢ goes) BIBO stable.
• Tracking step reference input¢ go es) BIBO stable and go(O) =
• Tracking ramp reference input¢ goes) BIBO stable, go(O) = and g�(O) = O.
EXAMPLE 9.2 Given a plant with transfer function g(s) = ¯ I ), fnd a proper
compensator and a gain p in the unityfeedback confguration in Fig. 9. 1 (b) so that the
output y will track asymptotically any step reference input.
The plant transfer function has degree = Thus if we choose m = all three poles of
the overall system can be assigned arbitrarily. Let the three poles be selected as  I
±
they spread evenly in the sector shown in Fig. Then we have
= =
=
+
We use the coeffcients of D(s) =
and N =
to form
as
[Ao Bo Al Bt l
.
�
Its solution is
Al = 1 Ao =
BI =
Bo =
This solution can easily be obtained using the MATLAB function / (slash), which denotes
matrix right division. Thus we have"
2. This is the solution obtained in (9. 1 0). This process of solving the polynomial equation in (9. 1 3) is conSiderably
simpler than the procedure discussed in Section 9. 1 . 1 .
9. 2 Uni ty·Feedback ConfigurationPole Pl acement
A(s) = + = =
and the compensator
+
=
.
=
=
277
will place the three poles of the overall system at and ± If the system is designed to
achieve regulation, we set p = (no feedforward gain is needed) and the design is completed.
To design tracking. we check whether or not No ¥ O. This is the case; thus we can fnd a p
so that the overall system will track asymptotically any step reference input. We use to
compute p:
Fa
p =
BoNo
=
Thus the overall transfer function from to y is
• 2)
=
=
Because is BIBO stable and go(O) = the overall system will track any step reference
input.
9. 2. 2 Robust Tracki ng and Disturbance Rejection
Consider the design problem in Example Suppose after the design is completed, the plant
transfer function g( s) changes. due to load variations, to =
Then
the overall transfer function becomes
+
This o e is still BIBO stable, but go(O) = = ¥ If the
reference input is a unit step function, the output will approach as ¯ C. There is
a tracking error of over Thus the overall system will no longer track any step reference
input after the plant parameter variations. and the design is said to be nonrobust.
In this subsection. we discuss a design that can achieve robust tracking and disturbance
rejection. Consider the system shown in Fig. in which a disturbance enters at the plant
input as shown. The problem is to design an overall system so that the plant output y will track
asymptotically a class of reference signal even with the presence of the disturbance and with
plant parameter variations. This is called
Before proceeding. we discuss the nature of the reference signal and the disturbance
. If both and .approach zero as ¯ 0. then the design is automatically
278 POLE PLACEMENT AND MODEL MATCHI NG
/
C(s)
(b)
(a)
Figr 9.2 Robust tracking and disturbance rejection.
:.
si ned to be BIBO stable. To exclude this
achieved if the overall system m Fig. 9._ IS de g
h ero as t ¬ ¯ If we have
( ) d (t) do not approac z
.
trivial case. we assume that r t an w
( ) d (t) it is not possible to achieve
no knowledge whatsoever about the nature of
·.re�re ·e �eed some information of r(t)
asymptotic tracking and disturbance reJection.
e that the Laplace transforms of r
{t) and
and wet) before carrying out the deSign. We assum
w(t ) are given by
Nw(s)
w(s) = L[
w(t )
]
=
Dw(s)
Nr (s)
r es) = L[r(t)] =
Dr (s)
(9.20)
I
.
I
.
however Nr(s) and Nw(s) are unknown
where Dr(s) and Dw(s) are known po ynorma ·.
nknow�magnitude a, then r es) = a/so
to us. For example, if r(
�
) i s a s�p functl
�
n
(
W.
¨d); it consists of a constant biasing with
Suppose the disturbance IS w(t ) . b + Csm Wo
f
but unknown amplitude C and
.
d " d With known requency Wo
.
unknown magmtude b a
�
a smusOi
2 w
2 Let I (s) be the least common denormnator
P
hase d. Then we have w(s) = Nw(s)/s(s +
0)1'
I xcluded because they have no
,
d
'
( ) The stab e po es are e
of the unstable poles of res) an w s .
'f e real parts For the examples
effect on Y as t _ 0. Thus all roots of < (s) have zero or P
OSI IV
.
)
just discussed, we have I (s) = s (s· + w� .
��U
f .
.
. Fi 9.2(a) with a strictly proper plant transfer unctIOn
Consider the umtyfeedback system shown In g.
N ( ) . me The reference Signal l (t) and
g
A l
s) = N(s) /D(s) . It is assumed that D(s) and s a
r
d
e
,
c
(
o
P
)
n
_
N
·
(s)/D (s) .
Let l(s) be the
' ( )
_
N (s)/ D (s) an w s U ' +
disturbance w(t ) are modeled as l s /
�( ) d ' (s) If no root of l(s) is a zero of g(s),
.
f h · t ble poles of r s an w .
) h
least common denomaato.O t e uns a
II ' U t
r
ack r (t) and reject w (t , bot
then there exists a proper compnsator such that the overa system WI
asymptotically and robustly.
.
' s  N(s)/ D(s) , then D(s)l(s) and N(s) are
Proof: If no root of I(s) IS a zero of g( ) 
B( )/ A(s) such that the polynomial
coprime. Thus there exists a proper compensator s
F(s) in
A(s) D(s)l(s) + B(s) N(s) = F(s)
has any desired roots, in particular, has all roots lying inside the sector shown in Fig.
8.3(a). We claim that the compensator
9. 2 UnityFeedback ConfigurationPole Placement 279
C(s) =
B(s)
A(s)l(s)
as shown in Fig. 9.2(a) will achieve the design. Let us compute the transfer function from
w to y:
, N(s)/D(s)
g
y
w(s) =
1 + ( B(s) / A(s)l(s» ) (N(s) / D(s»
N (s) A (s)1 (s)
A(s)D(s) l(s) + B(s)N(s)
Thus the output excited by wet) equals
N(s) A(s) l(s)
F(s)
'
( )
*
( )
'
( )
_ N(s) A(s)l(s) Nw(s)
Yu· s
 gyw ð U S 
F(s) Dw(s)
(9.21 )
Because all unstable roots of Dw (s) are canceled b y I (s), all poles of Yw (s) have negative
real parts. Thus we have yw(t)  0 as t = 0. In other words, the response excited by
w(t ) is asymptotically suppressed at the output.
Next we compute the output Yr(S) excited by res) :
Thus we have
, , , . B(s)N(s) ,
Yr(S) = gyr (s)r(s) =
A(s) D(s) l(s) + B(s) N(s
·
'(s)
e(s) : = res)  Yr (s) = (I ¯ gyr (s» r (s)
A(s) D(s)l(s) Nr (s)
F(s) Dr(s)
(9.22)
Again all unstable roots of Dr(s) are canceled by I(s) in (9.22). Thus we conclude
r(t )  Yr (t) ¬ 0 as t ¬ 0. Because of linearity, we have yet ) = Yw(t) + Yr (t) and
r(t) ¬ yet ) ¬ 0 as t ¬ 0. This shows asymptotic tracking and disturbance rejection.
From (9. 21 ) and (9.22), we see that even if the parameters of D(s), N(s) , A(s), and B(s)
change, as long as the overall system remmns BIBO stable and the unstable roots of Dr (s)
and Dw(s) are canceled by I (s) , the system still achieve tracking and rejection. Tus the
design is robust Q.E.D.
This robust design consists of two steps. First fnd a model I /I(s) of the reference signal
and disturbance and then carry out poleplacement design. Inserting the model inside the loop
is referred to as the interal model principle. If the model 1 /1 (s) is not located in the forward
path from w to Y and fromr to e, then I(s) will appear in the numerators of g,w(s) and g,,(s)
(see Problem 9. 7) and cancel the unstable poles of w(s) and r es) , as shown in (9. 21 ) and
(9.22). Thus the design is achieved by unstable polezero cancellations of I(s) . It is important
to mention that there are no unstable polezero cancellations in the poleplacement design and
the resulting unityfeedback system is totally stable, which will be defned in Section 9. 3. Thus
the interal model principle can be used in practical design.
In classical control system design, if a plant transfer function or a compensator transfer
function is of type I (has one pole at s = 0), and the unityfeedback system is designed to be
280 POLE PLACE,YENT AND MODEL MATCHI NG
BIBO stable, then the overall system wi l l track asymptotically and robustly any step reference
input. This is a special case of the interal model principle.
EXAMPLE 9.3 Consider the plant in Example 9.2 or g( s) = (s  2) /(sC  I ) . Design a
unityfeedback system with a set of desired poles to track robustly any step reference input.
First we introduce the interal model < (s) = /s . Then B(s ) /A( s) in Fig. 9. 2( a) can be
solved from
A(s ) D(s) <(s ) + B(s ) N(s) = F( s )
Because D(s) := D( s) d( s) has degree 3, we may select A( s) and B(s) to have degree 2. Then
Fl. s) has degree 5. If we select five desired poles as 2. 2 : j l . and  I : j2. then we have
F(s) = (s + 2) (s
2
'+ 4s + 5 ) (s2
+ 2s + 5)
= S
5
+ Ss + 30s 3 + 66s2 + 85s + 50
Using the coeffcients of D(s) = ( sC  l )s = 0s+0· sc +s3 and M(s) = 2+.> +0· 5c +0· S3,
we form
0  I 0 I 0 0
2 0 0 0 0
[Ao Bo Al BI A
2
B
2
1
0 0  I 0 I 0
0 2 0 0 0
= [50 85 66 30 8 1 1
0 0 0  I 0 I
0 0 2 0 0
Its solution is [ 1 27. 3 25 o  I I S. 7 96. 3] . Thus we ha\'e
B(5 ) 96. 3sc  1 1 8.7s  25
and the compensator is
A( 5)
C( s) =
B(s )
A(s) <( s)
s
2
+ 1 27. 3
96. 3s2  I I S. 75  25
( 5C + 1 27. 3) 5
l'sing this compensator of degree 3 . the unityfeedback system i n Fig. 9. 2I a) will track robustly
any step reference input and has the set of desired poles.
9. 2. 3 Embedding I nternal Model s
The design i n the preceding subsection was achieved by first introducing an interal model
I N( s) and then designing a proper B(s) / A ( s) . Thus the compensator B(s) / A (5 ) <( 5 ) is always
strictly proper. In this subsection. we discuss a method of designing a biproper compensator
whose denominator will include the interal model as a factor as shown in Fig. 9. 2lb). By so
doing, the degree of compensators can be reduced.
9. 2 Uni tyFeedback Confi gurationPol e Pl acement 281
Consider
A(5 ) D(s} + B(5) N(5) = F( s)
If deg D( s) = n and i f deg A(s ) = n  1 . then the solution .( s ) and B( s) i s unique. If
we increase the degree of A(s) by one, then solutions are not unique. and there i s one free
parameter we can select. Using the free parameter, we may be able íO include an internal model
in the compensator. as the next example i l lustrates.
EXA:IPLE 9.4 Consider again the design problem in Example 9. 2. The degree of D( 5) i s 2. If
A (5 ) has degree I. then the solution i s unique. Let us select A (s) to have degree 2. Then F ( 5)
must have degree 4 and can be selected as
F( s) = (s
2
+ 45 + 5) (S2 + 2s + 5) = .> + 6s3 + 1 8 , ' + 30s + 25
We form
1 0 I 0 0
2 0 0 0
0 I 0 I 0
[ AD Bo Al BI A
2
B: l
0 2 0 0
0 0  I 0 I
0 0 2 0
In order for the proper compensator
Bo + Bl s + B,s
2
Ci s) =
 ,
Ao + Al s + A
2
s
= [25 30 1 8 6 1 1
(9. 23)
to have 1 / s as a factor. we require AD = O. There are fve equations and si x unknowns in
( 9. 23) . Thus one of the unknowns can be arbitrari ly assigned. Let us select Ao ¯ O. This is
equivalent to deleting the frst row of the 6 X 5 matrix in ( 9. 23). The remaining 5 X 5 matrix
is nonsingular, and the remaining fve unknowns can be solved uniquely. The solution i s
[Ao Bo Al BI A
2
B
2
1 = [ 0  1 2. 5 34. S  38. 7 I  18. 8]
Thus the compensator is
BI s) 28. 8s2  3S.75  1 2. 5
C(5) = = ¬�¬
A( s ) 52 + 34.S5
This biproper compensator can achieve robust tracking. This compensator has degree 2. one
less than the one obtained in Example 9. 3. Thus this is a better design.
In the preceding example, we mentioned that one of the unknowns in ( 9. 23) can be
arbitraril y assigned. This does not mean that any one of them can be arbitrarily assigned. For
example, if we assign Ac = 0 or. equivalently. delete the ffth row of the 6 x 5 matrix in (9. 23).
then the remaining square matrix is singular and no solution may exist. In Example 9. 4. i f we
282
POLE PLACEMENT AND MODEL MATCHI NG
select .and if the remaining equation in does not have a solution, then we must
increase the degree of the compensator and repeat the design. Another way to carry out the
design is to fnd the general solution of Using Corollary we can express the general
solution as
... .
with one free parameter .If we select .then .0 and we will obtain the same
compensator. . .
We give one more example and discuss a diferent method of embeddlllg U the
compensator.
EXAMPLE 9.5 Consider the unityfeedback
;
ystem in Fig. 9.2(b) with
Design a
proper compensator . so tat the system will track asymptotIcally any step
reference input and reject disturbance wet) a sin(2t f) with un�own a and f.
In order to achieve the design, the polynomial .must contalll the dIsturbance model
Note that the reference model is not needed because the plant already contallls the
factor. Consider
.
For this equation, we have deg n = I. Thus i f m n then the solution i s
unique and we have no freedom i n assigning . If m then we have two free parameters
that can be used to assign . Let
.
Defne
We write . as
Equating its coeffcients, we obtain
.
�
For this example, if we select
0
0
�
2)
then the equation becomes
¬
.
�
�� I I
I
9. 3 I mpl ementable Transfer Functi ons
Its solution is Thus the compensator is
¯= :
. I X
283
This biproper compensator will place the poles of the unityfeedback system in the assigned
positions, track any step reference input. and reject the disturbance a sin(2t f) , both
asymptotically and robustly.
9, 3 I mpl ementabl e Transfer Functions
Consider again the design problem posed in Fig. with a given plant transfer function
Now the problem is the following: given a desired overall transfer function find
a feedback confguration and compensators so that the transfer function from r to y equals
This is called the model matching problem. This problem is clearly different from the
poleplacement problem. In pole placement, we specify only poles; its design will introduce
some zeros over which we have no control. In model matching, we specify not only poles but
also zeros. Thus model matching can be considered as poleandzero placement and should
yield a better design.
Given a proper plant tansfer function we claim that I is the best possible
overall system we can design. Indeed, if I , then y(t) = r tfor t : and for any
ret). Thus the overall system can track immediately (not asymptotically) any reference input
no matter how erratic r(t) is. Note that although y(t) r(t), the power levels at the reference
input and plant output may be diferent. The reference signal may be provided by turing a
knob by hand; the plant output may be the angular position of an antenna with weight over
several tons.
Although I is the best overall system, we may not be able to match it for a given
plant. The reason is that in matching or implementation, there are some physical constraints
that every overall system should meet. These constraints are listed in the following:
1, All compensators used have proper rational transfer functions.
2. The confguration selected has no plant leakge in the sense that all forward paths from r
to y pass through the plant.
3. The closedloop transfer function of every possible inputoutput pair is proper and BlBO
stable.
Every compensator with a proper rational transfer function can be implemented using the
opamp circuit elements shown in Fig. If a compensator has an improper transfer function,
then its implementation requires the use of pure diferentiators, which are not standard op
amp circuit elements. Thus compensators used in practice are often required to have proper
transfer functions. The second constraint requires that all power passes through the plant and
no compensator be introduced in parallel with the plant. All confgurations in Fig. 9. 1 meet
this constraint. In practice. noise and disturbance may exist in every component. For example,
noise may be generated in using potentiometers because of brush jumps and wire irregularity.
The load of an antenna may change because of gusting or air turbulence. These will be modeled
as exogenous inputs entering the input and output terminals of every block as shown in Fig. 9. 3.
28 POLE PLACEME'T AND MODEL MATCHI NG
Clearly we cannot disregard the effects of these exogenous inputs on the system. Although the
plant output is the signal we want to control, we should be concered with all variables inside
the system. For example, suppose the closedloop transfer function from to is not BIBO
stable: then any wi l l excite an unbounded and the system wi l l either saturate or bur out. If
the closedloop transfer function from ¡ t o i s improper, and if  contains highfrequency
noise. then the noise will be greatly amplifed at and the amplifed noise will drive the system
crazy. Thus the closedloop transfer function of every possible input{utput pair of the overall
system should be proper and BIBO stable. An overall system is said to be if the
closedloop transfer function of every possible inputoutput pair is proper; it is
if the closedloop transfer function of every possible inputoutput pair is BIBO stable.
Total stability can readily be met in design. If the overall transfer function from to is
BIBO stable and if there is no unstable polezero cancellation in the system, then the overall
system is totally stable. For example. conside
r the system shown in Fig. 9. 3(a). The overall
transfer function from to is
» 1
=
which is BIBO stable. However, the system is not totally stable because it involves an
unstable polezero cancellation of ¯ 2) . The closedloop transfer function from to Y
is 2 + I ) , which is not BI BO stable. Thus the output will grow unbounded if noise
ewn very small, enters the system. Thus we require BIBO stability not only of but
also of every possible closedloop transfer function. Note that whether or not and
are BIBO stable is immaterial.
The condition for the unityfeedback confguration in Fig. 9.3 to be well posed is
1 (Problem 9. 9). This can readily be established by using Mason's formula.
See Reference [7, pp. 200201 ] . For example, for the unityfeedback system in Fig. 9. 3(b),
we haw = x 2 = I . Thus the system is not well posed. Indeed. the
closedloop transfer function from to is
× 2)
=
3
which is improper. The condition for the twoparameter confguration in Fig. 9. 1 (d) to be well
posed is .¯I . In the unityfeedback and twoparameter confgurations. if
is strictly proper or = 0, then = 0 f  1 for any proper and the overall
systems will automatically be well posed. In conclusion, total stability and wellposedness can
easily be met in design. Nevertheless. they do impose some restrictions on go (5 ) .
(a)
Figur 9.3 Feedback systems.
9. 3 Implementable Transfer Functi ons 285
Defnition 9.1 ..
implementable if
8. 1
If an overall transfer function is not implementable, then no matter what confgura
tion is used to implement it, the design will violate at least one of the aforementioned con.traints.
Therefore, i n model matching, the selected must be implementable; otherWIse. It IS not
possible to implement it in practice.
Theorem 9.4
Consider a plant with proper transfer function Then is implementable if and only if
and
:=
are proper and BlBO stable.
Corollar 9.4
Consider a plant with proper transfer function
implemenrable if and only if
Then
1. All roots of have negative real parts is Hurwitz).
2. Deg ¯ deg :: deg ¯ deg (polezero excess inequality).
(9.24)
is
3. All zeros of with zero or positive real parts are retained in (retainment of non minimum
phase zeros).
We frst develop Corollary 9.4 from Theorem 9.4. If = is BIBO stable.
then all roots of have negative real parts. This is condition ( I ). We write (9. 24) as
×
I (S ) =
=
The condition for r to be proper i s
deg
deg :: deg
deg
which implies (2). I n order for t o be BIBO stable, all roots of with zero or positive real
parts must be canceled by the roots of Thus must contain the nonmlillmumphase
zeros of This is condition (3). Thus Corollary 9.4 follows dIrectly Theorem 9. 4.
Now we show the necessity of Theorem 9.4. For any confguration that has no plant
leakage. if the closedloop transfer function from to y is then we have
= =
286 POLE PLACEMENT AND MODEL MATCHI NG
which implies
'
'
= = t
Thus the closedloop transfer function from to is Total stability requires every closed
loop transfer function to be BIBO stable. Thus and must be BIBO stable. Well
posedness requires every closedloop transfer function to be proper. Thus and must
be proper. This establishes the necessity of the theorem. The sufciency of the theorem will
be established constructively i n the next subsection. Note that if and are proper, then
= is proper. Thus the condition for to be proper can be dropped from
Theorem 9.4.
In pole placement, the design will always introduce some zeros over which we have no
control. In model matching, other than retaiaing nonminimumphase zeros and meeting the
polezero excess inequality, we have complete freedom in selecting poles and zeros: any pole
inside the open lefthalf splane and any zero in the entire splane. Thus model matching
can be considered as poleandzero placement and should yield a better overall system than
poleplacement design.
Given a plant transfer function how to select an implementable model is not
a simple problem. For a discussion of this problem, see Reference [7, Chapter 9].
9. 3. 1 Model MatchingTwoParameter Confguration
This section discusses the implementation of = Clearly, i f = in Fig.
9. 1 (a), then the openloop confguration has as its transfer function. This implementation
may involve unstable polezero cancellations and, consequently, may not be totally stable.
Even if it is totally stable, the confguration can be very sensitive to plant parameter variations.
Therefore the openloop confguration should not be used. The unityfeedback confguration i n
Fig. 9. l (b) can b used to achieve every pole placement; however it cannot be used to achieve
every model matching, as the next example shows.
EXMPLE 9.6 Consider a plant with transfer function =
 I ) . We can readily
show that
• (s  2)
+ 2s + 2
(9.25)
is implementable. Because = I , the plant output will track asymptotically any step
reference input. Suppose we use the unityfeedback confguration with = 1 to implement
Then from
=
1 +
we can compute the compensator as
 I )
= = . 
+ 3)
9. 3 I mplementable Transfer Functi ons 287
This compensator is proper. However, the tandem connection of and involves the
polezero cancellation of  1 ) = + I )  I ) . The cancellation of the stable pole + 1
will not cause any serious problem in the overall system. However. the cancellation of the
unstable pole  1 will make the overall system not totally stable. Thus the implementation
is not acceptable.
Model matching in general involves some polezero cancellations. The same situation
arises in statefeedback stateestimator design; all eigenvalues of the estimator are not con
trollable from the reference input and are canceled i n the overall transfer function. However,
because we have complete freedom in selecting the eigenvalues of the estimator. if we select
them properly. the cancellation will not cause any problem in design. In using the unity
feedback confguration in model matching, as we saw in the preceding example, the canceled
poles are dictated by the plant transfer function. Thus, if a plant transfer function has poles with
positive real parts, the cancellation will involve unstable poles. Therefore the unityfeedback
confguration, in general, cannot be used in model matching.
The openloop and the unityfeedback confgurations in Figs. 9. 1 (a) and 9. ] (b) have one
degree of freedom and cannot be used to achieve every model matching. The confgurations
in Figs. 9. 1 (c) and 9. I (d) both have two degrees of freedom. In using either confguration.
we have complete freedom in assigning canceled poles; therefore both can be used to achieve
every model matching. Because the twoparameter confguration in Fig. 9. 1 (d) seems to be
more natural and more suitable for practical implementation, we discuss only that confguration
here. For model matching using the confguration in Fig. 9. I (c), see Reference [6].
Consider the twoparameter confguration in Fig. 9. I (d). Let
= '
where . and are polynomials. We call the .
and
the In general, ¡ and
need not be the
same. It turs out that even if they are chosen to be the same, the confguration can still be used
to achieve any model matching. Furthermore, a simple design procedure can be developed.
Therefore we assume = = and the compensators become
= 
.
= ¬
The transfer function from t o y in Fig. 9. I (d) then becomes
+ .
.
1 + 
Thus in model matching, we search for proper and . t o meet
= ¬= ¬=¬.:
+ .
(9.26)
(9.27)
(9. 28)
288 POLE PLACE,y\ ENT AND MODEL MATCHI ,NC
Note that the twoparameter confguration has no plant leakage. If the plant transfer function
8(5) is strictly proper as assumed and if C2 ( S) = M(s) / A(s) is proper. then the overall
system is automatically well posed. The question of total stability will be discussed in the next
subsection.
Problem Given 8(5) ~ N(s) / D(s) , where N(s) and D(s ) are coprime and deg
N(s) < degD( s) ¯ n. and gil'ell an implenzentable 80 (5) = £(5) / F( s) . fnd proper
L(s) / A(s) and M( s) / A(s) to meet (9.28).
Procedure 9. 1
1. Compute
80 ( 5 )
N( s)
£(5 ) its)
F( s) N(s) F(s)
(9. 29)
where £(5) and F(s) are coprime. Since £( s) and F(s) are implicitly assumed to be coprime.
common factors may exist only between E (s) and N (s) . Cancel all common factors between them
and denote the rest as £ (s) and F( s) . Note that i f £( s) = N(s ) . then F(s ) ¯ F( s) and £( s) = I .
l'sing (9.29), we rewrite (9. 28) as
, £( s ) N(s ) L (s) N( s)
g
o (5) = = : ,
F(s) A(s) D( s) + M(s ) N(s)
( 9 30)
F!om this equation, we may be tempted to set L (s) = ir s) and solve for A( s) and M(s) from
F( s) = A(s ) D(s) + M(s ) N( s)
.
However. no proper C,(s) = M(s) / A(s ) may exist in the
equation. See Problem 9. 1 . Thus we need some additional manipulation.
2. Introduce an arbitrary Hurwitz polynomial F( s) such that the degree of F(s) F( s) is 2n  I or
higher. In other words, if deg F( s) = p. then deg F(s) :: 'n ¬ I  p. Because the polynomial
F (s) will be canceled i n the design. i ts roots should be chosen to lie inside the sector shown i n Fig.
S3I a).
3. Rewrite (9.30) as
:0\\/ we set
L(s ) N(s ) £(s ) Fc s) N(s )
80 (s) = �
F( s ) Fcs )
A( s) D( s) + M( s) ,V( s )
L (s) = £(s) i( s)
and solve A ( 5) and M ( s) from
I f we write
A(s) D( s ) + M(s ) N(s) = Fcl' ) i(s )
A(s) ° Ao + AI 5 A�s2
+
. · .
+ Ams
fll
M( s) = Ma + Bi s + ,\,s2 + . . + M", s
"
'
F( s) F( s) = Fa + F) s + Fes2 +
. .
. + Fo+", sn+m
19. 3 / )
( 9 32)
(9. 33)
9. 3 Impl ementabl e Transfer Functi ons
with m :: l = I. then A(s ) and M(s) can be obtained by solving
with
[ Ao Ma Al Ml . . ' Am MmlS", = [ Fa F
I
Fe
. . . Fn+", l
o 0
o 0
o
o
Do
Na
The computed compensators L(s) / A(s) and M ( 5) / A(s) are proper.
289
(9. 3')
We justify the procedure. By introducing F( s ) , the degree of F( s) ic s) i s 2n  I or higher
and. following Theorem 9. 2, solutions
,
A(s) and M( s) wtth deg Lf ( s) :: deg A ( s) = 111 and
H > n = I exist in (9. 34) for any F(s) F( s) . Thus the compensator M ( s) / A(s) IS proper. Note
tha�if we do not introduce F( s) . proper compensator M( s) / A(s) may not exist in (9. 34).
Next we show deg L (s) :: deg A(s) . Applying the polezero excess inequality to (9. 3 1 )
and using (9. 32). we have
deg ( F(5 ) F(s»  deg N (s)  deg L(s) :: deg D( s)  deg N(s)
which implies
deg L(s) :: deg ( F(s ) ics ) ¬ deg D( s) = deg A(s)
Thus the compensator L(s) / A (s) i s proper.
EXAMPLE 9,7 Consider the model matching problem s:udied in Example 9. 6. That is. given
8(s) = (s  2) /(52  I ) , match 8,, (s) ¯ ( s  2),(s¯ + 2s + 2) . We Implement It In the
twoparameter confguration shown in Fig. 9. 1 Cd) . Ftrst we compute
8
,,
( s)
=
(s  2)  I
¬
�( s)
N( s) (s' + 's + 2) (s  2) s2 + 2s + 2 F(s)
Because the degree of F( s) is 2, we select arbitraril y Fe s) ~ s + 4 so that the degree of
F( s) i(s) is 3 = 2n  I . Thus we have
L(s )  £( s) F( s) = (5 + .+ )
and A( s) and M(s) can be solved from
A (s) D( s) + M( s ) N(s) = Fe s) i(s) = (S2 + 's + 2) (s + 4)
= s3 + 6s2 + l Os + 8
(9. 35 )
290 POLE PLACEMENT AND MODEL MATCHI NG
or
r

1
2
[Ao Mo AI Mi l
. .
�
o
¯1
2
The solution is Ao ¯ 1 8, AI ¯ 1 , Mo = 1 3, and MI ¯ 1 2. Thus we have A(s) = 1 8 + s
and M(s) = ¯J 3  12s and the compensators are
L(5) (s + 4) M(s) _ ( 1 2s + 1 3)
C (s)   C, (s) ¯ ¯
1 8
I 
A(s)

s + 1 8 ¡
.
A(s) s +
This completes the design. Note that, becau:e go(O) ¯ 1, the output of the feedback system
will track any step reference input.
EXAMPLE 9.8 Given g(s) ¯ (s  2) /(52  1 ) , match
, (s  2) (45 + 2) _4s
2
+ 6s + 4
g
oes) =
(s2 + 2s + 2) (s + 2)  s3 + 4s2 + 6s + 4
This g (s) is BIBO stable and has the property go(O) = 1 and g�(s) = 0; thus the o�erall
syster: will track asymptotically not only any step reference input but also any ramp mput.
See Problems 9. 1 3 and 9. 14. This go es ) meets all three conditions in Corollary 9.4; thus It \b
implementable. We use the twoparameter confguration. First we compute
£(s)
goes) (s  2) (4s + 2) _ (4s + 2)
N(s)
=
(s2 + 2s + 2) (s + 2) (s  2)

s3 + 4s2 + 6s + 4

.
Fcs)
Because the degree of F(s) is 3, which equals 2n ¯ 1 ¯ 3, there i s no need t o introduce F(s )
and we set f (s) = 1 . Thus we have
L(s) = hs) £( s) ¯ (4s + 2)
and A(s) and M(s ) can be solved from
.¯
[Ao Mo AI Mi l
. .
�
o
I
2
1
o
o
as Ao ¯ 1, AI  34/3, Mo = 23/3, and MI ¯ 22/3. Thus the compensators are
(4s + 2) _ (22s + 23)
CI (5) =
S + 34/3
C
2
(5) 
3s + 34
This completes the design. Note that this design does not involve any polezero cancellation
because hs) = 1 .
9. 3 Implementable Transfer Functi ons 291
9. 3. 2 Implementation of TwoParameter Compensators
Given a plant with transfer function g (s) and an implementable model go (s) . we can implement
the model i n the twoparameter configuration shown in Fig. 9. 1 (d) and redrawn in Fig. 9.4(a).
The compensators CI (s) ¯ L(s) /A(s ) and C2 (s) = M(5) /A(5) can be obtained by using
Procedure 9. 1 . To complete the design. the compensators must be built or implemented. This
is discussed i n this subsection.
Consider the confguration i n Fig. 9.4(a). The denominator A(s) of CI (5) i s obtained by
solving the compensator equation i n ( 9. 33) and may or may not be a Hurwitz polynomial. See
Problem 9. 1 2. If it is not a Hurwitz polynomial and if we implement CI (s) as shown i n Fig.
9. 4(a). then the output of CI (s) wi l l grow without bound and the overall system i s not totally
stable. Therefore, in general. we should not implement the two compensators as shown i n Fig.
9. 4(a). If we move C
2
(s) outside the loop as shown i n Fig. 9. 4(b). then the design wi l l involve
the cancellation of M (s) . Because M (5) is also obtained by solving (9. 33). we have no direct
control of M (s) . Thus the design is in general not acceptable. If we move C 1 (5 ) inside the
loop, then the confguration becomes the one shown in Fig. 9 . .+(c). We see that the connection
involves the polezero cancellation of L(s )  Fcs) £(s) . We have freedom in selecting F(s) .
The polynomial £(s ) is part of £( s) , which. other than the nonminimumphase zeros of N(s) .
we can also select. The non minimumphase zeros, however, are completely canceled in £(s ) .
Thus L(s) can be Hurwitz3 and the implementation i n Fig. 9. 4(c) can be totally stable and is
(a)
(b)
l
L & . . . . . . . & . . . & . . . '
I e)
I d)
Figure 9.4 Twodegreesoffreedom confgurations.
3. This may not be true in the multi variable case.
292 POLE PLACEMENT AND MODEL MATCHI NG
acceptable. However. because the two compensators L ( s) / A ( s) and M ( s) / L (s) have different
denominators, their implementations require a total of 2m i ntegrators. We discuss next a better
implementation that requires only II i ntegrators and i nvolves only the cancellation of Frs) .
Consider
• • • L (s ) • M(s) •
lI (s)  CI (s)r(s)  Co (s ) v (.» = r es )  yes )
 ..
A (s ) A( s)
 AI ( s) [ L(s ) ¯ M( s) ]
�
(
S
)
y e s)
This can be plotted as shown i n Fig. 9.4(d). Thus we can consider the two compensators as a
single compensator with twa inputs and one output with transfer matrix
(9. 36)
I f we fnd a minimal realization of (9. 36), then its dimension i s m and the two compensators
can be implemented using only H i ntegrators. As we can see from (9. 3 1 ), the design involves
only the cancellation of Frs) . Thus the implementation in Fig. 9.4(d) i s superior to the one in
Fig. 9.4(c). We see that the four confgurations in Fig. 9.4 all have two degrees of freedom and
are mathematically equivalent. However, they can be di fferent i n actual implementation.
EXAMPLE 9.9 Implement the compensators i n Example 9.8 using an opamp circuit. We write
lI (s)  C
I
(s)r(s)  C2 (s) V(s) = .
• • •
(4S + 2) 7 . 33S + 7. 67
•
. + 1 1 . 33 s + 1 1 . 33 y(s)

[ 4 7. 33] +
I
[43 33
_
75. 38]
�(s)
á + 1 1 . 33 yes )
Its statespace realization i s , using the formula i n Problem 4. 1 0.
.t =  1 1 . 33x + [43. 33  75. 38]
lI = X + [4 7. 33J
,
(See Problem 4. 1 . ) This onedimensional state equation can be realized as shown in Fig. 9. 5.
Thi s completes the implementation of the compensators.
9. 4 Mul tivari abl e UnityFeedback Systems
This section extends the pole placement discussed in Section 9.2 to the multi variable case.
Consider the unityfeedback system shown in Fig. 9.6. The plant has p inputs and q outputs
and is described by a q x p strictly proper rational matrix G(S) . The compensator C(s) to be
designed must have q inputs and p outputs in order for the connection to be possible. Thus
CI s ) is required to be a p x q proper rational matrix. The matrix P is a q x q constant gain
matrix. For the time being. we assume P = Iq . Let the transfer matrix from r to y be denoted
by Gu( S) . a q x q matrix. Then we have
Figure 9.5 Opamp circuit
implementation.
9. 4 Mul ti vari abl e Un ityFeedback Systems
RC ¯ 1
ROSA ..
R
¯`
CIS) Gi s)
. . . . . . . . . . . . .
¡
~ . . . . . . . . . . . .
I
l 
l I
'ul
' l
'
~
'
L . . . . . . . . . . . . L . . . . . . . . . . . .
Figure 9.6 Multivariable unity feedback system with P  Iq .
Go(s)  [Iq + G(s) C( sWI G(s) C( s)
 G( s) C(s) [lq + G(
S
) C( sWI
 G( s) [ lp + C(s)G(s WI C(s)
293
(9. 37)
The frst equality is obtained from y(s) = G(s) C( s) [i(s)  ye s)] : th
�
second one from
e(s)  res)  G( s) C(s)e(s) : and the third one from i( s) = C(s) [r(s) 
�
( s) i(s) ] . They can
also be verifed directly. For example, pre and postmultiplying by [I" + G( s)C(s ) ] in the frst
two equations yield
G(s)C(s) [Iq + G( s) C( s) ] = [Iq + G( s) C(s ) ]G( s) C( s)
which i s an identity. This establishes the second equality. The third equality can similarly be
establis�.
. ' 
I
Let G(s) = N( s) DI (.) be a right copnme fractIOn and l et CIS)  A (s)B(s) be a left
fraction to be designed. Then (9. 37) implies
Go(s) = N(s) D1 ( s) [ I + A I (s)B(s) N(s) DI (sWI A I ( s) B( s)
1
I
 l i
B
= N(s ) DI (s) \ A
1 ( s) [A(s)D(s) + B(s) N( s) ] D (.) A (. ) ( s)
= N(s » ) [A(s) D( s) + B( s) N(s WI B( s)
 N( s) FI ( s) B( s)
(9. 38)
294 POLE PLACEMENT AND MODEL MATCHI NG
where
A(s)D(s) + 8(s) N(s) = F(s) (9.39)
It is a polynomial matrix equation. Thus the design problem can be stated as follows: given
P x P D(s) and q x p N(s) and an arbitrary p x p F(s) , fnd p x p A(s) and p x q 8(s) to
meet (9.39). This is the matrix version of the polynomial compensator equation in (9. 12).
> Theorem 9. M 1
Given polynomial matrices D(s) and N(s) . polynomial matrix solutions A(s) and 8(s) exist in (9.39)
for any polynomial matrix F(s) i f and only if D(s) and N (s) are right coprime.
Suppose D(s) and N(s) are not right doprime, then there exists a nonunimodular polyno
mial matrix R(s) such that D(s) = D(s)R(s) and N(s) = N(s)R(s) . Then F(s) in (9.39) must
be of the form F(s) R(s), for some polynomial matrix F(s). Thus if F(s) cannot be expressed
i n such a form, no solutions exist i n (9.39). This shows the necessity of the theorem. If D(s)
and N(s) are right coprime, there exist polynomial matrices A(s) and B(s) such that
A(s)D(s) + B(s)N(s) ¯ I
The polynomial matrices A(s) and B(s) can be obtained by a sequence of elementary oper
ations. See Reference [6, pp. 587595] . Thus A(s) = F(s)A(s) and 8(5) = F(s)B(s) are
solutions of (9.39) for any F(s). This establishes Theorem 9.MI . As in the scalar case, it
is possible to develop general solutions for (9. 39). However. the general solutions are not
convenient to use i n our design. Thus they will not be discussed.
Next we will change solving (9.39) into solving a set of linear algebraic equations.
Consider G(s) = N(s)D
I
(s), where D(s) and N(s) are right coprime and D(s) is column
reduced. Let Ji be the degree of the ith column of D(s). Then we have, as discussed in
Section 7. 8.2.
deg G(s) = degdet D(s) = JI + J2 +
. . .
+ Ip =: n
Let I : = max(II , 12, . . . , Ip). Then we can express D(s) and N(s) as
D(s) = Do + DI S + D:sz + . . .
+ DI
"
sl
"
N(s) = No + Nl s + Nzs2 +
. . .
+ NI
"
sl
"
(9.40)
Note that D" is singular unless III = Ilz = . . . = Ip· Note also that N" = 0, following the
strict properess assumption of G(s) . We also express A(s), 8(5), and F(s) as
A(s) = Ao + Ai s + Azs
2
+
. .
.
+ Ams
m
8(s) = 80 + 81 5 + 82sZ +
. .
+ 8msm
F(s) = Fo + Fis + Fzsz +
. . .
+ F,,+ms,
,
+
m
Substituting these into (9.39) and matching the coefficients of like powers of 5, we obtain
(9.41 )
9. 4 Mul ti vari abl e Uni tyFeedback Systems 295
where
Do DI D" 0 0 0
No NI N" 0 0 0
0 Do D,,l D" 0 0
Sm
:
= 0 No N,,

l N" 0 0 (9.42)
o o o
o o o
The matrix Sm has m + 1 block rows; each block row consists of p Drows and q Nrows.
Thus Sm has (m + l )(p + q) number of rows. Let us search linearly independent rows of Sm in
order from top to bottom. It turs out that if N (5 )D

I
(5) is proper. then all Drows are linearly
independent of their previous rows. An N row can be linearly independent of its previous rows.
However, i f an Nrow becomes linearly dependent, then, because of the structure of Sm, the
same Nrows in subsequent Nblock rows wi l l be linearly dependent. Let Vi be the number of
l inearly independent ith Nrows and let
v := max{vI , Vz, . . . , vq }
It is called the DW index of G(s) . Then all q Nrows in the last Nblock row of Sv are l inearly
dependent on their previous rows. Thus Sv

1 contains all linearly independent Nrows and its
total number equals, as discussed in Section 7. 8. 2, the degree of G(s), that is.
VI + V2 +
. . .
+ Vq = n (9.43)
Because all Drows are linearly independent and there are a total of pv Drows i n Sv1 , we
conclude that Svl has n + pv independent rows or rank n + pv.
Let us consider
It has P(1 + I) number of columns; however, it has at least a total of L;=I
(I ¯ Id zero
columns. In the matrix SI , some new zero columns wi l l appear i n the rightmost block column.
However, some zero columns in So will not b zero columns in 51 ' Thus the number of zero
columns in 51 remains as
P
U := ¯« Ild = Pl  (Il + Jz +
. . .
+ IIp) ¯ Pl n (9.44)
;=
1
L96 POLE PLACEMENT AND MODEL MATCHI NG
I n fact, this i s the number of zero columns i n S,,, , In = 2, 3, . . . . Let S,, _I be the matrix S,, _
I
after deleting the,e zer� columns. Because the number of columns in S", is p( !l + I + 111 ) . the
number of columns in S,' _
I
is
f : = p( J. + I + v  I )  ( pJ  11 ) = pv + H (9" +5)
The rank of S,L

I clearly equals the rank of S,, _I or pv + n. Thus S,L
I
has full column rank.
Now i f 111 increases by I . the rank and the number of the columns of S" both increase ·
fJ (because the
p
new Drows are all linearly independent of their previous rows) : thus S"
still has full column rank. Proceeding forward. we conclude that S"" for 111 :: 11  I . has full
column rank.
Let us define
He ( 5) := diag(s' L ] , \´ ` . . . . . s" P ) (9 .6)
and
Hr ( s) = diag(s"'] . 5
"
. . . . . S
"
e ) ( 9" +7)
Then we have the fol lowing matrix version of Theorem 9. 2.
Theorem 9. M2
Consider the unityfeedback system shown i n Fig. 9. 6 wi th P = Iq . The plant i s described by a q X p
strictly proper rational matrix G( s) Let G( s ) be factored as G( s) = N (s )D I ( s) . where D( s) and
)( 5 ) are right coprime and D( s) i s column reduced wi th col umn degrees Ji . i = 1 . 2 . . . . . p. Let v
be the ro\ index of G(.» and let m, :: v  I for i = 1 . 2 . . . . . p. Then for any I x I poly nominal
matri x F ( .1. such that
l i m H; I (s ) F( nH,�
I
( s) = Fh 1 9 . (8)
\ ¬¯
i s a nonsingular constant matri x. there exists 3 p x q proper compensator A ´ (5 ) B( s) . \here A( s ) i s
ro\\ reduced with row degrces Ill ] . such that the transfer matrix from r t o y equals
G,, ( s) = N( s) F1 (s) B( s)
Furthennore. the compensatur �Jn be obtained by �oh· ing seb of linear algebraic equations i n ( 9. Ì I .
Proof Let H = max ( 111
1
, 111: . . . . . III
P
) . Cnnsider the constant matrix
It is formed from the coeffcient matrices of F( s) and has order p X ( Ill + 11 + I ) . Clearly
F(s) has column degrees at most H + J, . Thus F has at least C number nf zero columns.
where O is given i n 1 9. 4'). Furthennore. the positions of these zero columns coincide
with those of S", . Let . be the constant matrix . after deleting these zero columns. Nnw
consider
(9" +9)
9  ,\1ul ti vari abl e Uni tyFeedback Systems 297
It is obtained from (9"+ I 1 by deleting C number of zero cnlumns in S", and the conespond
ing zero Cllumns in F. !ow because S", has full column rank i f III :: \'  I . we conclude
that for any F( s) of column degrees at most m + J, . solutions A, and B, exist i n ( 9. .+9l. Or.
equivalently, polynomial matrices A( s) and B( s) of row degree III or less exist i n ( 9"+9) .
Note that generally S", has more rows than columns: therefore solutions of (9"+9) are not
unique.
Next we show that A I ( s) B( s) i s proper. Note that D" is. i n general. singular and the
method of proving Theorem 9.2 cannot be used here. Using H,(s) and R· ( s) . we write.
as i n (7. 80).
D(s) = [Dhc + DIr ( s) H,I (s ) ] Hc Cn
N( s) = [Nhe + Nlc (.nH:
I
( s) ]He ( s)
A( s) = H, ( s) [A"r + H;I (S) A1r (s ) j
B(s l = H, ( s) [B"r + H;
I ( s)B' r ( s) ]
F(s ) = H, ( s) [Fh + H;I ( s) F, ( s) H�
I
( s) ] Hc ( s)
where Dis) H;
1
( 5) , NI( s ) H,� 1 ( S) , H;I (slA'r ( s) . H,�I (S l B,, ( s) . and H;I ( s ) F, ( s) H;' I. ·)
are all strictly proper rational functions. Substituting the above into (9.39) yields. at s = ¯.
which reduces to. because Nhc = 0 following strict properess of G( s) .
Because D\s) is column reduced. D"e i s nonsingular. The constant matrix F" i s nOl1singuiar
by assumption: thus A,,, = FhDhcl i s nonsingular and A( s) is ro\\ reduced. Therefl.re
A  1 ( s) B\s ) is proper ( Corollary 7. 8). This establishes the theorem. Q. E. D.
A polynomial matrix Fl s) meeting (9" +8) i s said to be rowCOllllllll redllced with [0\\
degrees III, and column degrees Ji . If /11 1 = U_ *
. . . = P, = 111 . then the rowcolumn
reducedness is the same as column reducedness with column degrees III + Ii , . In applicatil.n.
we can select F( s) tn be diagonal or triangular with polynomials with desired roots as its
diagonal entries. Then FI ( 5 1 and. consequently. G" (s) have the desired roots as their poles.
Consider a!win S,,  I . It is of order ( I  q)v x ( J  v) l' . It has O ~ PI  H number
nf zero column·Thus the matrix S,' _I i s of llfder ( I + q) v X [ ( Ii + \. ) p  ( 1'1.(  Il l ] or
(p + q) v x ( vp + 11 ) . The matrix S, ' _
I
contains pv li nearly independent Drows but contains
nnly VI + ' " + v'l = 11 linearly independent /rows. Thus SV
 I
contains
y := (I + q ) 1'  !' '  11 = q \  11
linearly dependent iVroWS. Let Sv
 I
be the matrix S, , _I after deleting these linearly dependent
/Vrows. Then the matrix S, _ I is of order
[ ( p ¯ q) !'  (q1J  11 ) ] X ( v! + n ) = ( vp + II ) ` ( l 'p + n)
Thus S" _I i s square and nonsingular.
298
POlE PLACEMENT AND MODEL MATCHI NG
Consider .with U .v ¬
K5v1 := (Ao Bo Al B
I
It actually consists of the following p sets of linear algebraic equations
(.
i = 1 . 2 . . . . . P
where ki and fi are the ith row of K and F. respectively. Because 5v1 has full column rank. for
any r. solutions ki exist in (9.50). Because 5v1 has more Y rows than columns. the general
solution of (9.50) contains Y free parameters (Corollary 3. 2). If m in Sm increases by 1 from
\ " to \, then the number of roWS of 5v increases by (p
q) but the rank of 5v increases
only by p. In this case. the number of free piaeters will increase from y to y + q. Thus in
the MIMO case. we have a great deal of freedom in carrying out the design.
We discuss a special case of (9. 50). The matrix 5v1 has y linearly dependent Nrows. If
we delete these linearly dependent N rowS from 5v1 and assign the corresponding columns
in Bi as zero. then becomes
.
(Ao Bo . . , AvI Bv_d5vI = F
where 5v1 is. as discussed earlier. square and nonsingular. Thus the solution is unique. This
is illustrated in the next example.
EXAMPLE 9.10 Consider a plant with the strictly proper transfer matrix
]
.
The fraction is right coprime and is column reduced with column degrees J1 ~ 2 and
J2 = 1 . We write
]
]
]
.
·
±
and
·
.
� �] ·
� �
]
� �
]
We use MATLAB to compute the row index. The QR decomposition discussed in Section
can reveal linearly independent columns. from left to right. of a matriX. Here we need linearly
independent rows. from top to bottom. of Sm; therefore we will apply QR decomposition to
the transpose of Sm. We type
d 1 = [ 0 0 0 0 1 0 1 ; d2 = l 0 0 0 1 0 0 1 ;
n1 = [ 1 1 0 0 0 0 1 ; n2 = [ 0 1 0 0 0 0 1 ;
sl= [ d1 0 0 ; d2 0 0 ; n1 0 0 ; n2 0 0 ; . · ·
o 0 d1 ; 0 0 d2 ; 0 0 nl ; O 0 n2 l ;
[ q, rl =qr ( sl ' )
which yields. as in Example
9. 4 Mul ti vari abl e Uni tyFeedback 5 ystems
dl
o d2 x x
o n l x
=
o dl
o d2
o n l
o
299
The matrix q 's d
l not nee ed and lb not typed I th
.
nonzero tr' Th
. n e matnx r we us d'
e n les. e nonzero diagonal entri f '
. . e x. l . and ni to denote
or, equivalently. linearly indepenuent ro
e
f
s
S
o �
,
�Ield the Imearly independent columns ofS'
N I
ws Õ 
·
vve see that the I'
I
rows and one linearly independent N2 Th
r, are two mearly independent
three linearly independent Nrows Th f
row. e
.
degree of G( s) is and we have found
2
. ere ore there l5 no d
VI = and V: . Thus the row index is ¯ W
nee to search further and we have
for any columnreduced ·of col d
V  
.
e select m  = m2 .U = v ¬ I = I Thus
p
umn egrees m + f ¨ 3 d
.
roper compensator such that the resulting unit feedb
 ¨ an m + J2 .2. we can fnd a
matnx. Let us choose arbitrarily
y ack system has ·as its denominator
·
.
'
�
(
+ 2
+
[
=
�
and form with m = v ¬ I =
[Ao Bo AI Bi J
I
I
I
I
=
1
¸
o 2 I
= ·
The
.
a in is for this problem. Thus S and 
see I� After deleting the zero col
I
h
·both h
�
ve one zero column as we can
any ·solutions exist in Th
umn
2
t e remammg S
I
has full column rank and fo
. . e matnx SI has order 8 x and solutions are not uni�ue
r
300 POLE PLACEMENT AND MODEL MATCHI NG
I n searching the row index, we kew that the last row of .
i s a linearly dependent row. I f
we delete the row, then we must assign the second column of B } as 0 and the solution wi l l be
unique. We type
d1 = [ 0 0 0 0 0 1 J ; d2 = [ 0
n1= [ l 1 0 0 0 O J ; n2 = [ O
dl t = [ O 0 0 0 1 J ; d2 t= [ C
s 1 t = [ d1 O ; d2 O ; n1 0 ; n2
f l t= L 1 S 0 1 7 0 1 J ;
f l t l s l t
0 0
i 0
0 0
0 ; 0
1
0
0
0 O J ;
0 O J ;
0 ] ; nl t:: [ 0 ì 0 0 O J ;
d1 t ; 0 0 d2 t ; O 0 nl tl ;
which yields Computing once again for the second row of . we
can fnally obtain
2
1 5
o
I S
o
Note that MATLAB yields the frst columns; the last column 0 is assigned by us (due to
deleting the last row of . I )' Thus we have
·
2 + s
B
o
and the proper compensator
.
�
2
will yield the overall transfer matrix
´
.
o
.
.
This transfer matrix has the desired poles. This completes the design.
(9.5.)
The design i n Theorem 9. M2 i s cared out by using a right coprime fraction of We
state next the result by using a left coprime fraction of .
> Corollary 9.M2
Consider the unityfeedback system shown in Fig. 9. 7. The plant is described by a . . strictly proper
rational matrix Let be factored as .  · where  and · are left
coprime and  is row reduced with row degrees \, i 1 , 2, . . . . Let / be the column index of
and let mi :: / ¯ 1 . Then for any . . rowcolumn reduced polynominal matrix . such that
lim diag(s"l , 5´¯ . . . . . ·.· · . . . , smq ) .
1�>
9. 4 Ilul ti vari abl e Uni tyFeedback Systems
!' º)
. . . . . . . . . . . . . , . . . . . . . . . . . . .
¦
l

l


l


L . . . . . . . . . . . .
 . . . . . . . . . . . . '
Figure 9.7 Unity feedback with ((5) ·
301
i� a nonsingul ar constant matrix. there exists a . . proper compensator .·
(5) , where
· is column reduced with column degrees m" to meet
and the transfer matrix from f to y equals
. ~ =
l
~
1 · · .
( 9.55 )
(9. 56)
Substituting · and . · into the frst equation in (9. 37)
yields
· . ·
· . ·
·  · · . · . ·
( 5)
which becomes. after substituting (9. 55) ,
· . . .·
I · .
This establishes the transfer matrix from f to y i n the theorem. The design in Corol lary 9. M2
hinges on solving (9. 55). Note that the transpose of ( 9. 55) becomes (9.39): left coprime and
row reducedness become right coprime and column reducedness. Thus the linear algebraic
equation in (9Al ) can be used to solve the transpose of (9. 55) . We can also solve (9. 55)
directly by forming
 ·
0 0 0 0
.

·
·
�.
0 0
.
Tm
·
· 
· .
0 0
.
0 0  · 
·
·
0 0 0 0  ·
302 POLE PLACEMENT AND MODEL MATCHI NG
10
Ao
1
1
j
!
j
1
l
X
Al
1m
Am
F,+m
(9.57)
We search linearly independent columns of Tm i n order from left to right. Let f be the column
index of . or the least integer such that .
contains linearly independent ATcolumns.
Then the compensator can be solved from (9. 57) with . f ¯ J .
9. 4. 1 Regulation and Tracking ·
As in the SISO case, pole placement can be used to achieve regulation and tracking in
multi variable systems. In the regulator problem, we have r = 0 and if all poles of.the overall
system are assigned to have negative real parts, then the responses caused by any nonzero
initial state will decay to zero. Furthermore, the rate of decaying can be controlled by the
locations of the poles; the larger the negative real parts, the faster the decay.
Next we discuss tracking of any step reference input. In this design, we generally require
a feedforward constant gain matrix P. Suppose the compensator in Fig. 9.6 has been designed
by using Theorem 9.M2. Then the q x q transfer matrix from r to is given by
(9.58)
If . is BIBO stable, then the steadystate response excited by . for t > or
.
where . i s an arbitrary q x I constant vector, can be computed as, using the
fnalvalue theorem of the Laplace transform.
lim lim .. . .
r O \÷0
Thus we conclude that in order for to track asymptotically any step reference input, we
need, in addition to BIBO stability,
(9.59)
Before discussing the conditions for meeting (9.59), we need the concept of transmission zeros.
Transmission zeros Consider a q x p proper rational matrix .
where
and are right coprime. A number real or complex, is called a .. �8lU
of . if the rank of is smaller than min( p, q).
EXAMPLE 9. 11 Consider the 3 x 2 proper rational matrix
�
I
1
o
.
0
� 
1
0
1
I
9. 4 Mul ti vari abl e Uni tyFeedback Systems 303
This has rank 2 at every thus .
has no transmission zero. Consider the 2 x 2 proper
rational matrix
�
o
[
+ 2 0 
1
0
This has
�
ank
I
at 0 and 2. Thus . has two transmission zeros at 0 and
Note that . has poles also at 0 and 2.
From this example. we see that a transmission zero cannot be defined directly from .
i t must be defned from its coprime fraction. Either a right coprime or left coprime fraction
can be
,
used and each yields th� same
_
set of transmission zeros. Note that if . is square
and :.
where and are right coprime and
and are left coprime, then the transmission zeros of . are the roots of det or the
roots of det . Transmission zeros can also be defned from a minimal realization of .
�et (A, . C, be any ndimensional minimal realization of a q x p proper rational matrix
. Then the transmission zeros are those such that
[
H A .
rank
C
< min(p. q)
This i s used in the MATLAB function ¯¯cT' to compute transmission zeros. For a more
detailed discussion of transmission zeros, see Reference [6. pp. 623635] .
Now we are ready to discuss the conditions for achieving tracking or for meeting (9. 59).
Note that and . are q x p. p x p, and p X q. Because Iq has rank q, a necessary
condition for (9.59) to hold is that the q x p matrix has rank q. Necessary conditions
for q are p :: q and 0 is not a transmission zero of . Thus we conclude
that in order for the unityfeedback confguration in Fig. 9.6 to achieve asymptotic tracking,
the plant must have the following two properties:
• The plant has the same or a greater number of inputs than outputs.
• The plant transfer function has no transmission zero at
Under these conditions. 1(0) has rank q. Because we have freedom in selecting we can
select it such that . has rank q and the q x q constant matrix . is nonsingular.
Under these conditions, the constant gain matrix P can be computed as
(9 60)
Then we have . Iq, and the unityfeedback system in Fig. 9. 6 with P in (9.60) will
track asymptotically any step reference input.
9. 4. 2 Robust Tracking and Disturbance Rejection
As in the SI50 case, the asymptotic tracking design in the preceding section is not robust. In
this section. we discuss a diferent design. To simplify the discussion, we study only plants with
304
POLE PLACEMENT AND MODEL MATCHI NG

ì ¯
L . . & . . . . . . . . . . . . ÷ . . . . '
Figure 9.8 Robust tracking and disturbance rejection.
an equal number of input terinals and output ierminals or p ~ q. Consider the unityfeedback
system shown in Fig. 9. 8. The plant is described by a p x p strictly proper transfer matrix
factored in left coprime fraction as G(s) = D1 (s)N(s). It is assumed that p × I disturbance
w(t) enters the plant input as shown. The problem is to design a compensator so that the output
y(t) will track asymptotically a class of p x I reference signals r(t) even with the presence
of the disturbance w(t) and with plant parameter variations. This is called robust tracking alld
disturbance rejeclion.
As in the S1S0 case. we need some information on r( t) and w(t ) before carrying out the
design. We assume that the Laplace transforms of r(t) and w(t) are given by
res) = L[r(t )) = Nr(s)D;I (s) w(s) = L[w(t)) ~ Nw (s) D� I (s) (9 6 1 )
where Dr(s) and D�. (s) are known polynomials; however, Nr (s) and Nw( s) are unknown to
us. Let t(s) be the least common denominator of the unstable poles of res) and w( sl . The
stable poles are excluded because they have no
�
ffect on
J
U) as I ¬ C. We introduce the
interal model <I (s)lp as shown in Fig. 9. 8. If D(s) and N(s) are left coprime and irno root
of <(s) is a transmission zero of G(s) or, equivalently, < (s) and det N(s) are coprime. then
it can be shown that D(s) t( s) and N(s) are len coprime. See Reference [6. p. 4431. Then
.
¬ }
Corollary 9. M2 implies that there exists a proper compensator Cis) = B(s)A (5) such that
t(s)D( s) A( s) + N(s)B(s) = Y(S)
( 9 62)
for any F(s) meeting the condition in Corollary 9. M2. Clearly Y(s) can be chosen to be diagaonl
with the roots of its diagonal entries lying inside the sector shown in Fig. 8. 3(a). The unity
feedback system in Fig. 9.8 so designed will track asymptotically and robustly the refrence
signal r( t ) and reject the disturbance w( t). This is stated as a theorem.
Theorem 9.M3
Consider the unityfeedback system shown in Fig. 9. 8 where the plant has a p X P strictly proper transfer
matrix G(s ) = D1 (s)N(s). I t i s assumed that D(s) and N(s) are left coprime and 0( 5) is row reduced
with row degrees U;, i = 1 . 2, . . , p. The reference signal r( t) and disturbance w(I) are modeled
as res)  Nr ( s) D; I (s) and w(s) = N�. (s) D;;I (S) . Let t(s) be the least common denominator
of the unstable poles of res) and w(s). If no root of t(s) is a transmission zero of G(.I". then there
exists a proper compensator Cis) = B(s) (A(s)< ( 5) ) I such that the overall system will robu,tl' and
asymptotically track the reference signal r(1) and reject the disturbance w(t).
94 Mul ti vari abl e Uni tyFeedback Systems 305
m
Proof' First we show that the system will reject the disturbance at the output. Let us
assume r ~ 0 and compute the output YU' (s) excited by w(s) . Clearly we have
Yw (s) = Ol (s)N(s) [w(s)  B(s) • .
I
(s)ql (s)y�. (s))
whi ch implies
Yw(s) = [1 + OI (s)N(s)B(s)A

1 (s)< I (SWI D1 (s) N(s)w(s)
= OI (S)[O(S)<(S)A(S) + N(s)B(s)]A  1 (S)< I (S) ¨
x D1 (s)N(s) w(s)
= <(s)A(s)[O(s) <(s)A(s) + N(s)B(s» ) I N(s)w(s)
Thus we have, using (9. 61 ) and (9.62),
Yw(s) = A(s)y I (s)N(s)<(s)Nw(s) D;:1 (s) (9. 63)
Because all unstable roots of Du' (s) are canceled by t(s) . all poles of Y�· (s) have negative
real parts. Thus we have yw(t ) ¯ 0 as I  C and the response excited by Wet) is
suppressed asymptotically at the output.
Next we compute the eror er rs) excited by the reference signal res) :
er rs) = res)  01 (s)N(s)B(s)A
1
(s) <I (s)e, (s)
which implies
er rs) = [I + Ol (s)N(s)B( s)A \s)q I (sW1 f(s)
= <(s)A(s) [D(s)<(s)A(s) + N(s) B(sW1 D(s) r(s)
= A(s)yI (s)D(s) t(s)Nr (s) D;1 (5) (9. 64)
Because all unstable roots of Dr (s) are canceled by <(s) , the error vector er rs) has only
stable poles. Thus its time response approaches zero as I  C. Consequently, the output
y( t) will track asymptotically the reference signal r(t ) . The tracking and disturbance rejec
tion are accomplished by inserting the interal model tI (s)lp' If there is no perturbation
in the interal model. the tracking property holds for any plant and compensator paramerer
perturbations, even large ones, as long as the unityfeedback system remains BIBO stable.
Thus the design i s robust. This establishes the theorem. Q. E.D.
I n the robust design, because of the interal model, t(s) becomes zeros of every nonzero
entry of the transfer matrices fom w to y and from I to e. Such zeros are called blocking �eros.
These blocking zeros cancel all unstable poles of w( s) and r(s) ; thus the responses due to these
unstable poles are completely blocked at the output. It is clear that every blocking zero is a
transmission zero. The converse, however, is not true. To conclude this section we mention
that if we use a right coprime fraction for G(s), insert an interal model and st�bilize it, we
can show only disturbance rejection. Because of the noncommutative property of matrices. we
are not able to establish robust tracking. However, it is believed that the system still achieves
robust tracking. The design discussed in Section 9. 2. 3 can also be extended to the multi variable
case; the design, however, will be more complex and will not be discussed.
306 POLE PLACEMENT AND MODEL MATCHI NG
9. 5 Multi variable Model MatchingTwoParameter Confi guration
In this section. we extend the SISO model matching to the multi variable case
.
We study only
plants with square and nonsingular strictly proper transfer matrices. As in the SISO case,
given a plant transfer matrix G(s), a model Go(s) is said to be implementable if there exists a
noplantleakage confguration and proper compensators so that the resulting system has the
overall transfer matrix Go (s ) and is totally stable and well posed. The next theorem extends
Theorem 9.4 to the matrix case.
> Teorem 9.M4
Consider a plant with p x p strictly proper transfer matrx G(s). Then a p x p transfer matrix Go(s)
is implementable if and only if Go (s) and
'
are proper and BIO stable.'
1(s) := G
I
(s)Go(s) (9.65)
For any noplantleakage confguration, the closedloop transfer matrix from I to U is
1(s) . Thus well posedness and total stability require Go(s) and 1'(s) to be proper and BIDO
stable. This shows the necessity of Theorem 9.M4. Let us write (9.65) as
(9.66)
Then (9.66) can be implemented in the openloop confguration in Fig. 9. 1 (a) with C(s) = 1'(s) .
This design, however, i s not acceptable either because it i s not totally stable or because it i s very
sensitive to plant parameter variations. If we implement it in the unityfeedback confguration,
we have no freedom in assigning canceled poles. Thus the confguration may not be acceptable.
In the unityfeedback confguration, we have
u(s) = C(s)[r(s)  yes) ]
Now we extend it to
9. 5 Mul ti vari abl e Model Matc hi ng 307
L & & . . . & . & . . & . & . &
Figur 9.9 Twoparameter confguration.
Thus we have
u(s) = [I + A
I
(s) M(s) N(sjD
1
(sWI A I (s)L(s)r(s)
= D(s)[A(s)D(s) + M(s)N(sWI L(s)r(s)
Yes) = N(s)DI (s)u(s) = N(s) [A(s)D(s) + M(s)N(sWI L(s)r(s)
and the transfer matrix from r to y is
Go(s) = N(s) [A(s)D(s) + M(s)N(sWI L(s)
Thus model matching becomes the problem of solving A(s). M(s) and L(s) in (9.69).
(9.69)
Poblem Given a p x p strictly proper rational matrix G(s) = N(s)D
I
(s). where
N(s)
.
and D(s) are TIght copnme and D(s) is colun;.n reduced with column degrees
IL;
'I
I = 1 , 2, . . . .¬. and glven a
.
ny implementable G
2
(s). fnd prper compensators
A (s)L(s) and A Cs)�I(s) In Flg. 9. 9 to implement Go Cs).
� Procedure 9. Ml
(9.67) 1. Compute
This is a twodegreesoffreedom confguration. As in the SISO case, we may select C
I
(s) and
Cl(S) to have the same denominator matrix as
(9. 68)
Then the twoparameter confguration can be plotted as shown in Fig. 9. 9. From the fgure,
we have
ils) = A 
1
(s)[L(s)i(s)  M(s)N(s)D
1
(s)u(s)]
which implies
4. If G(s) is not square, then Go(s) is implementable if and only if Go(s) is proper, is BIBO stable. and can be
expressed as Go(s) ¯ G(s)T(s), where T(s) is proper and BIBO stable. See Reference [ 6, pp. 517523].
W
I
(s)G
o (s) = F
I
(s)£(s)
where F(s) and £(s) are left coprime and F(s) is row reduced.
(9. 70)
2. Compute the row index \ of ((s) = N(s)D1 (s). This can be achieved by using QRdecomposition.
3. Select
(9. 71 )
where Q¡(s) are arbitrary Hurwitz polynomials, such that F(s)F(s) is row<olumn reduced with
column degrees ILi and row degrees mi with
mi 2 v  I
(9.72)
for i = 1 , 2, . . . , p.
308 POLE PLACEMENT AND MODEL MATCHI NG
4. Set
L(s) = F(s)E(s) (9. 73)
and solve A(s) and M(s) from
A(s)D(s) M(s)N(s) = F(s)F(s) =: F(s) (9.74)
Then proper compensators A I (s)L(s) and A 
I (s)M(s) can be obtained to achieve the model
matching.
This procedure reduces to Procedure 9. 1 if G(s) is scalar. We frst justify the procedure.
Substituting (9. 73) and (9.74) into ( 9
.
69) yields
Go (s) = N(s) [F(s)F(s)riF(s)E(s) = N(s)yl (s)E(s)
This is (9. 70). Thus the compensators implement Go (s). Defne
He(s) = diag(sl" , sl , . . , sl'P )
By assumption, the matrix
is a nonsingular constant matrix. Thus solutions A(s) , having row degrees mi and being row
reduced, and M(s), having row degrees mi or less, exist in (9.7.) (Theorem «· Thus
A 
I
(s)M(s) is proper. To show that A  I (s) L(s) is proper, we consider
1(s) = G1 (s)Go(s) = D(s)N 1 (s)Go(s)
= D(s)F1 (s) E(s) = D(s) [F(s)F(sW
I
F(s)E(s)
= D(s) [A(s)D(s) M(s)N(sW
I
L(s)
= D(s) \ A(s)[ 1 A  1 (s)M(s)N(s)D1 (s)]D(s) rl L(s)
= [I A 
1
(s)M(s)G(sWI A 1 (s)L(s)
which implies, because G(s) = N(s)D1 (s) is strictly proper and A 
I
(5)M(s) is proper.
l i m 1'(s) = lim A1 (s)L(s)
A¬¯ 3¬·¯
Because T(o) is fnite by assumption. we conclude that A1 (s)L(s) is proper. If G(s) is
strictly proper and if all compensators are proper. then the twoparameter confguration is
automaticallY
,
well posed. The design involves the polezero cancellation of F( s). which we
can select. If F(s) is selected as diagonal with Hurwitz polynomials as its entries. then pole
zero cancellations involve only stable poles and the system is totally stable. This completes
the justifcation of the design procedure.
EX.UIPLE 9.12 Consider a plant with transfer matrix
G(S) =
1
�
2
� [
�
�
(9. 75)
9. 5 Mul ti variable Model Matchi ng
:
It has column degrees /r ~ · and Jl2=1 . Let us select a model as
309
(9.76)
"1
which is proper and BIBO stable. To check whether or not Go (s) i s implementable, we compute
Tcs ) : = G1 (s)Gv(s) =
s
_
[
Go (s) =
s2 ·. ·
which is proper and BIBO stable. Thus Go Cs) is implementable. We compute
·+
I '
I 
1
1
[
S2 +

0
�S + 2
0
N (s)Go Cs) =
s2 �s ·
_
7
=
s
2
+ �+ 2
�
s ·
s2 2s ·
=
. � ·
�S 2 [
� � '
=: F 1 (s)E\s)
For this example, the degree of N1 (s)Go(s) can easily be computed as · The deterinant
of F(s) has degree · thus the pair F(s) and E(s) are left coprime. It is clear that F(s) is row
reduced with row degrees T
I
= T2
= · The row index of G(s) was computed in Example 9. 1 0
as v = · Let us select
Then we have
F(s) = diag« s · I )
F
'
.
· · ·
(s) s) = ¯
o s· ·s ¬ 2
=
.
· 0
' · 2s s2
(9. 77)
It i s rowcolumn reduced wi th row degrees (m
1
= m2
= I = v I } and column degrees
( fl = · f
2
= I } . Note that without introducing F(s), proper compensators may not exist.
We set
^ ¬ ·
L(s) = F(s)E(s) =
1 �
=
· . ·
·
·
(9. 78)
31 0
POLE PLACEMENT AND MODEL MATCHI NG
and solve A(s) and M(s) from
A(s)D(s) M(s)N(s) = F(s)F(s) =: F(s)
From the coeffcient matrices of D(s) and N(s) and the coeffcient matrices of we can
readily form
I
[Ao Mo AI MI l
=
� ·
�
¦
· ·
As discussed in Example if we delete the last column of SI , then the remaining 8
1
has
full column rank and for any F(s), after deleting the last zero column, solutions exist in
Now if we delete the last N row of 8 , which is linearly dependent on its previous row, the set
of solutions is unique and can be obtained, using MATLAB, as
· · ·
[Ao Mo AI MI l
=
· · I
Note that the last zero column i s by assignment because of the deletion of the last Nrow in
81 , Thus we have
· s ·
A(s) =
· s
·
M(s) =
·
The two compensators A 
I
(S)M(5) and A 
I
(s)L(s) are clearly proper. This completes the
design. As a check, we compute
· ·
Go(s) = N(s)F
1
(5)L(s) =
s3 45'
·
=
� ·
s' � ·
This is the desired model. Note that the design involves the cancellation of (s · which we
can select. Thus the design is satisfactory.
9. 5 Multivariable Model Matching 31 1
Let us discuss a special case of model matching. Given G(s) = N (s )DI (s) , let us select
1(s) = D(s)D,(s), where Df(s) has the same column degrees and the same columndegree
coeffcient matrix as D(s). Then 1(s) is proper and Go(5) = G(5)1'(s) ~ N(5)D,(S) . This
is the feedback transfer matrix discussed in · Thus the state feedback design discussed
in Chapter can also be carried out by using Procedure 9.Ml .
9. 5. 1 Decoupling
Consider a p X p strictly proper rational matrix G(s) = N(5)D1 (s). We have assumed that
G(s) is nonsingular. Thus GI (s) = D(s)NI (s) is well defned; however, it is in general
improper. Let us select 1'(s) as
1(s) ~ GI (s)diag(dj
l
(s) , dil (s) , . . . , d;; l (s»
where d; (s) are Hurwitz polynomials of least degrees to make 1(s) proper. Defne
I(5) = diag(d
l
(5) , d
2
(s), . . . , dp(s» ·
Then we can write 1'(s) as
1'(s) = D(s) WI (s) I
I
(s) ~ D(s) [ I(s)N(s)r1
If all transmission zeros of G(s) or, equivalently, all roots of det N (s) have negative real pans.
then 1'(s) is proper and BIBO stable. Thus the overall transfer matrix
Go(s) ~ G(s)1'(s) = N(s)D

I (s)D(s)[I(s)N(s)r
1
is implementable. This overall transfer matrix is a diagonal matrix and is said to be decoupled.
This is the design in Example ·
If G(s) has nonminimumphase transmission zeros or transmission zeros with zero or
positive real parts, then the preceding design cannot be employed. However, with some
modifcation, it is still possible to design a decoupled overall system. Consider again G(s) =
N(s)D1 (s) . We factor N(s) as
N(s) ~ N
I
(s)N
2
(s)
with
N
I
(s) = diag(.I
I
(s) . .ds), . . . , .
l
p (5»
where .ti Cs) is the greatest common divisor of the ith row of N(s). Let us compute Nil (5),
and let .
2
; (S) be the least common denominator of the unstable poles of the ith column of
Ni' (s) . Defne
N2d · = diag(.1
1
(S) , .21 (S) , . . . , .2p (S»
Then the matrix
3 1 2 POLE PLACEMENT AND MODEL MATCHI NG
l
1
1
has no unstable poles. Now we select . as
. ·
.
·
with
· diag(dl .
.
where . are Hurwitz polynomials of least degrees to make . proper. Because · has
only stable poles, and . are Hurwitz, . is BIBO stable. Consider
. .. ·
·
 · .
· ·
. ·
diag
¡
�
I
.
,
�
�
� .
¸
. . .
where � �
.�
. I t i s proper because both . and . are proper. It is clearly
BIBO stable. Thus . i s i mplementable and i s a decoupled system.
EXA!HPLE 9.13 Consider
>
· .
.
'
+ I .
s
We compute det ·
The plant has two nonminimumphase
transmission zeros. We factor · as
·. ·
.·
.
�
. �

'
with ·
diag( l , I ) , and compute
If we select
then the rational matrix
·
¬ 1 )
�
'
·
diag\(s  I ) ,
· · · ¯
�
I
has no unstable poles. We compute
�

+
·
If we choose
I

+ .

0. 5 Mul ti \ ar i abl e Model Matchi ng
. diag« s
2
· · + . + · + · + 2))
then
i s proper. Thus the overall transfer matrix
æ » × .
· diag
+ ` ` j 2) '
.: * ·`
1)
2
¸
. · · + ·
31 3
i s implementable. This i s a decoupled system. Thi' decoupled system will not track any step
reference input. Thus we modify it as
» • ¡ ·  1 )
¸
dJag
· . . + 2) ' . · + ·
which has . I and will track any step reference input.
Next we implement in the twoparameter confi guration. We follow Procedure .
To b\ �� �' ' �
l
� �����
o
.
·. ¬ 1)
�
 1
¬ .
. .
.
. ·
. .
. .
. .
o
· ·

. · ·
It is a left coprime fraction.
and
From the plant transfer matrix. we haw
.
�
� 
� � �
.


I
·.
�
�

�  · � � 
�
�
We use QR decomposition to compute the row index of . We type
dl = [ 1 0 0 0 0 0 I ; d2 = [ 0 C O O
nl = [ O . 1 0 0 0 0 ] ; �2 = [  1 
1 0 0 I ;
Û 0 0 0 I ;
31 4 POLE PLACEMENT AND MODEL MATCHI NG
s2= [ dl 0 0 0 0 ; d2 0 0 0 O ; nl 0 0 0 0 ; n2 0 0 0 0; . . +
Û 0 dl 0 0 ; 0 0 d2 0 0 ; 0 0 nl 0 0 ; 0 0 n2 0 0 ; . . «
Û 0 0 0 dl ; O 0 0 0 d2 ; 0 0 0 0 nl ; O 0 0 0 n2 ] ;
[ q, r] =qr ( s2 ' )
From the matrix I (not shown), we can see that there are three linearly independent NI rows
and two linearly independent N2rows. Thus we have VI = 3 and I = 2 and the row index
equals v = 3. If we select
F(s) = diag (CS + 3)2
, (s + 3»)
then F(s)F(s) is rowcolurnreduced with row degrees (2. 2} and column degrees (3, 2} . We
set
í
» ~ 4(s + 3)2
L(s) = F(s)E(s) =
4(s + 3)
and solve A(s) and M(s) from
4(s + W
4s(s + 3)
A(s)D(s) + M(s)N(s) = F(s)F(s) : = F(s)
Using MATLAB, they can be solved as
S2 + lOs + 329
A(s) =
46
10
¦
s2 + 7s + 6
and
290s2  1 1 4s  36
M(s)  
4652 + 34s + 1 2
1 89s + 293
¦ 34s  46
(9.91 )
(9.92)
(9.93)
The compensators AI (5)M(s) and AI (s)L(s) are clearly proper. This completes the design.
The model matching discussed can be modifed in several ways. For example, if stable
roots of det N
2
(S) are not inside the sector shown in Fig. 8. 3(a), they can be included in th.
Then they will be retained in Go (s) and will not be canceled in the design. 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 outputs. In this case, the resulting overall transfer matrix is a blockdiagonal matrix.
These modifcations are straightforward and will not be discussed.
9.6 Concl udi ng Remarks
In this chapter, we used coprime factions to carry out designs to achieve pole placement or
model matching. For pole placement, the unityfeedback confguration shown in Fig. 9. I (a),
a onedegreeoffreedom confguration, can be used. If a plant has degree n, then any pole
placement can b achieved by using a compensator of degree n  1 or larger. If the degree of
a compensator is larger than the minimum required, the extra degrees can be used to achieve
robust tracking, disturbance rejection, or other design objectives.
Problems 31 5
Model matching generally involves polezero cancellations. Onedegreeoffreedom
confgura�s cannot be used here because we have no freedom in selecting canceled poles.
Any twodegreeoffreedom confguration can be used because we have freedom in selecting
canceled poles. This text discusses only the twoparameter confguration shown in Fig. 9" +.
All designs in this chapter are achieved by solving sets of linear algebraic equations. The
basic idea and procedure are the same for both the SISO and MIMO cases. All discussion
in this chapter can be applied, without any modifcation, to discretetime systems; the only
difference is that desired poles must be chosen to lie inside the region shown in Fig. 8.3 (b)
instead of in Fig. 8.3(a).
This chapter studies only strictly proper G(s). If G(s) is proper, the basic idea and
procedure are still applicable, but the degree of compensators must be increased to ensure
properess of compensators and well posedness of overall systems. See Reference [6]. The
model matching in Section 9. 5 can also be extended to nonsquare plant transfer matrices. See
also Reference [6].
The controllerestimator design in Chapter 8 can be carried out using polynomial fractions.
See References [6, pp. 5065 14: 7, pp. 461465]. Conversely, because of the equivalence of
controllable and observable state equations and coprime factions, we should be able to use
state equations to carry out all designs in this chapter. The statespace design, however. will
be more complex and less intuitively transparent, as we may conclude from comparing the
designs in Sections 8. 3. 1 and 9. 2. 2.
The statespace approach frst appeared in the 1 960s, and by the 1980s the concepts of
controllability and observability and controllerestimator design were integrated into most
undergraduate control texts. The polynomial fraction approach was developed in the 1 970s:
its underlying concept of coprimeness, however, is an ancient one. Even though the concepts
and design procedures of the coprime fraction approach are as simple as, if not simpler than,
the statespace approach. the approach appears to be less widely known. It is hoped that this
chapter has demonstrated its simplicity and usefulness and will help in its dissemination.
9. 1 Consider
A(s ) D(s) + 8(s) N(s) " S2 + 25 + 2
where D(s) and N(s) are given in Example 9. 1 . Do solutions A(s) and 8(5) exist in the
equation0 Can you fnd solutions with deg 8(s) : deg A( 5) in the equation0
9.2 Given a plant with transfer function g(s) = (5  1 )/(5
2
 4), fnd 0 compensator in
the unityfeedback confguration so that the overall system has desired poles at 2 and
I ± j 1 . Also fnd a feedforard gain so that the resulting system will track any step
reference input.
9.3 Suppose the plant transfer function in Problem 9. 2 changes to g(s) = (s 0. 9)/(5
2
.. 1 )
after the design is completed. Can the overall system still track asymptotically any step
reference input? If not, give two different designs, one with a compensator of degree 3
and another with degree 2. that will track asymptotically and robustly any step reference
input. Do you need additional desired poles? If yes, place them at 3.
31 6 POLE PLACEMENT AND MODEL MATCHI NG
9.4 Repeat Problem 9.2 for a plant with transfer function g(s)  (5  l ) /s (s  2) . Do
you need a feedforward gain to achieve tracking of any step reference input? Give your
reason.
9.5
9.6
9.7
Suppose the plant transfer function i n Problem 9.4 changes to g(s) = (5 0.9)/5(5 2. 1 )
after the design i s completed. Can the overall system still track any step reference input?
Is the design robust?
Consider a plant with transfer function g(s ) = 1 / (5  1 ) . Suppose a disturbance of form
wet) = a sin(2t +e) , with unknown amplitude a and phase e, enters the plant as shown
in Fig. 9. 2. Design a biproper compensator of degree 3 in the feedback system so that
the output will track asymptotically any step reference input and reject the disturbance.
Place the desired poles at  1 ± j2 and 2 ± j 1 .
Consider the unity feedback system shown i n Fig. 9. 1 0. The plant transfer function is
g(s)  2/s(s + l ) . Can the output track robustly any step reference input? Can the
output reject any step disturbance w(t) = a? Why?
�
2
I . 5
�
Figure 9.10
`
9.8 Consider the unityfeedback system shown in Fig. 9. 1 1 (a). Is the transfer function from
r to y BIBO stable? Is the system totally stable? If not, fnd an inputoutput pair whose
closedloop transfer function is not BIBO stable.
(a)
(b)
Figure 9.11
9.9 Consider the unityfeedback system shown in Fig. 9. I I (b) . ( 1 ) Show that the closedloop
transfer function of every possible inpututput pair contains the factor ( I +C(s)g
(
S ) ) 
1
.
(2) Show that ( I + C(s)g(S» 1 is proper if and only i f
1 + C(oliCo) o 0
(3) Show that if Cis) and g(s) are proper, and if C(:o)g(o) o 1 , then the system is
well posed.
9. 10 Given g(5) = (5"  1 ) / (53 + a1 5" + a
2
5 + a
3
) , which of the following g,, (s) are
i mplementable for any ai and bi .
9. 11
9.12
9.13
s  1
(s + 1 ) 2
( s2 ¯ 1 )
(s + 2)2
5 + 1
(s + 2) (5 + 3)
5"  1
(s  2)3
(s  l ) (bos + b) )
(s + 2) 2(52 + 25 + 2)
Problems 31 7
Given g(s) = (s  1 ) /s(s  2 ) , implement the model go es) = 2(s  1 ) /(s2 + 2s  2)
i n the openloop and the unityfeedback confgurations. Are they totally stable? Can the
implementations be used i n practice?
Implement Problem 9
.
1 1 i n the twoparameter confguration. Select the poles to be
canceled at s = 3. Is A(s) a Hurwitz polynomiaJ? Can you Implement the two
compensators as shown in Fig. 9.4(a)'7 Implement the two compensators in Fig. 9. '(d)
and draw its opamp circuit.
Given a BIBO stable go es). show that the steadystate response yss (t) · lim,_", y et)
excited by the ramp reference input r (t)  at for t :: 0 is given by
'ss (t)  go (O)at + g�(O)a
Thus i f the output is to track asymptotically the ramp reference input we require go( 0)  I
and g�(O) = O.
9.14 Given a BlBO stable
, bo + bl s +
.
. . + bmsm
go es) 
.
n
ao + al s ¯
. . .
+ ans
with n :: m, show that go(O)  1 and g�(O) = 0 i f and only if ao = bo and al = bl ·
9.15 Given a plant with transfer function g(s)  ( s + 3) (s  2) /(S3 + 2s  I ) . ( I ) find
conditions on bl , boo and a so that
,
bl s + bo
goe
s) =
s2 + 2s + a
is implementable; and (2) determine i f
(5  2) "bl s + bo)
g
o
(
s) =
(s + 2) (s2 + 2s + 2)
is implementable. Find conditions on bl and bo so that the overall transfer function will
track any ramp reference input.
9. 16 Consider a plant with transfer matrix
G(s) 
s
s
�
s
2
 I
Find a compensator in the unityfeedback confguration so that the poles of the
.
overall
system are located at 2.  I ± j and the rest at s = 3. Can you fnd a feedtrward
gain so that the overall system will track asymptotically any step reference mput .
31 8 POLE PLACEMENT AND MODEL MATCHI NG
9.17 Repeat Problem 9. 1 6 for a plant with transfer matrix

. =
. . ¬ I ) .
9. 18 Repeat Problem 9. 1 6 for a plant with transfer matrix
��. �
. =
I ·
~ &
. . 1
implementable? If yes, implement it.
9.20 Diagonalize a plant with transfer matrix
I
.
=
I
. ·. · . ·
Set the denominator of each diagonal entry of the resulting overall system as . ·. ·
References
We list only the references used in preparing this edition. The list does not indicate original sources of
the results presented. For more extensive lists of references, see References [2, 6, 1 3, 1 51 .
I . Anderson, B. D. 0 . . and Moore, 1. B. , Optimal ControlLinear Quadratic Methods, Englewood
Clifs, NJ: Prentice Hall. 1 990.
2. Antsaklis, A. 1., and Michel, A. N. , Linear Systems, New York: McGrawHill, 1 997.
3. Callier, E M. , and Desoer, C. A . + Multivariable Feedback Systems, New York: SpringerVerlag.
1 982.
4. Callier, F M. , and Desoer, C. A., Linear System Theor, New York: SprngerVerlag, 1 991 .
5. Chen, C. T., Introduction to Linear System Theor, New York: Holt, Rinehart & Winston, 1 970.
6. Chen, C. T., Linear System Theor and Design, New York: Oxford University Press, 1 984.
7. Chen, C. T . . Analog and Digital Control System Design: Transfer Function, StateSpace. and
Algebraic Methods. New York: Oxford University Press, 1 993.
8. Chen, C. T. , and Liu, C. S. , "Design of Control Systems: A Comparative Study," IEEE Control
System Magazine. vol. 14, pp. 475 1 . 1 994.
9. Doyle, 1. c., Francis, B. A., and Tannenbaum. A. R., Feedback Control Theory, New York:
Macmillan, 1 992.
10. Gantmacher, F R . . The Theory ofMatrices, vols. I and 2, New York: Chelsea, 1 959.
I I . Golub, G. H. , and Van Loan. C. E, Matrix Computations 3rd ed . . Baltimore: The Johns Hopkins
University Press. 1996.
1 2. Howze. J. W., and Bhattacharyya, S. P., "Robust tracking, error feedback. and twodegreeoffreedom
controllers," IEEE Trans. Automatic Control. vol. 42, pp. 980983, 1 997.
1 3. Kailath. T. . Linear Svstems, Englewood Cliffs. NJ: Prentice Hall , 1 990.
1 4. Kurcera. v, Analvsis and Design ofDiscrete Linear Control Systems. London: Prentice Hall. 1 991 .
1 5. Rugh, W, Linear SYstem Theory, 2nd ed . . Upper Saddle River, NJ: Prentice Hall. 1 996.
1 6. Strang, G . + Linear Algebra and Its Application. 3rd ed . . San Diego: Harcourt, Brace, Javanovich,
1 988.
1 7. The Student Edition ofMATLAB, Version 4. Englewood Cliffs. NJ: Prentice Hall, 1 995.
1 8. Tongue, B. H .. Principles ofVbration. New York: Oxford University Press. 1 996.
19. Tsui. C. c. , Robust COlltrol System Design. New York: Marcel Dekker. 1 996.
20. Vardulakis. A. I. G .. Linear Multivariable Control, Chichester: John Wiley. 1 99 1 .
2 1 . Vidyasagar. M., Contrl SYStem Synthesis: A Factorization Apprach. Cambrdge, �A: MIT Press.
1 985.
22. Wolovich. W. A . . Linear Multil'ariable Systems. New York: SprngerVerlag. 1 974.
23. Zhou, K . . Doyle. J. C .. and Glover. K., Robust and Optimal Contrl, Upper Saddle River. NJ:
Prentice Hall, 1 995.
31 9
Answers to Sel ected Probl ems
CHAPTER 2
2.1 (a) Linear. (b) and (cl Nonlinear. In (b). shift the operating point to (0, Yo); then (u , j) is linear.
where j = y ¯ Y
o
.
2.3 Linear. time varying. causal.
2.5 No, yes, no for x(O) i O. Yes. yes, yes for x(O) = O. The reason is that the superposition propry
must also apply to the initial states.
2.9 y(t) = 0 for t < 0 and t > 4.
0. 5t2 for 0 � t < I
y(t) = 1 . 5t2 + 4t  2 for I � t � 2
y(4 ¬ t) for 2 < t � 4
2.10 g(s) = I / (s + 3) , g(t ) = e
3
t for t :: O.
2.12
2.15 (al Defne XI = e and x2 = 8. Then XI = X2, X2 = (gl l) sin xl  (ulml) cos xl . If e is
small. then
x = x + u
. 0 0
gi l 0 l /mg
It is a linear state equation.
(b) Defne XI = el , x2 ~ 01 • X3 ¬ e2• and X4 = 82. Then
Xz " ¨ (gl I t l sin XI + (m2g1 mI L ) COS X3 sin(x3  xt l
+ ( l lml /) sin x3 sin(x3  xt l · u
This is a set of nonlinear equations. If ej " 0 and B; ¯ O. then it can b linearized as
321
322 ANSWERS TO SELECTED PROBLEMS
.
.. .
'
.1. /
1
2.16 From
.1
1
h k1 e  k2u
IB be (/1 +  II II
we can readily obtain a state equation. Assuming I and taking their Laplace transforms can
yield the transfer function.
Í
t
2.18 ~ .
1
. Yes.
2.20
CHAPTER 3
.
. .
3.1 . :
3.3
.¯
�
3.5 P .
1
) nullity .i l= I ; 3, 0; 3, 1 .
Answers to Selected Problems 323
3.7 . is a solution. Unique. No solution if
3.9 . . The solution
3.12
3.13
3.18
3.21
3.22
has the smallest 2norm.
.
`
���
.
�
.
.
�
�
�
.�
�
�
�
.
.
Þ
t
Þ
t
 I te' ¬ Þ
t
8
^ì
�
e
'
Þ
i
e
A1
324 ANSWERS TO SELECTED PROBLEMS
3.24
B =
B =
�
�
1.2
�
3.32 Eigenvalues: O. Í No solution for CI . For any m ¡ ¯ m is a solution for C,.
3.34 ). 1 . 4. 7. 1 .7.
CHAPTER 4
4.2 ,(I) =
.I
4.3 For T = ¯.
for I ::
.
=
�
.·
.= · .
.
·
.

·
4.5 �!ATLAB yields I ylmax ~ I XI I",,,.< = I x, l m".< = and I X3 1 ",,,.< = for unit
step input. Defne XI = X. .' = 0.5X2. and .\3 = Xl· Then
4.8
4. 11
4. 13
X =
�
�
X
�
.= ·
.
The largest permissible O is = ·
They are not equivalent but are zerostate equivalent.
l'sing (434). we have
. =
1
y =
·
·
.= �
·
·
·
·
�
. �
·
.
·
�
.
�
o
�
�
Il
·
·
Both have dimension 4.
4.16
4.20
co>rn
· (0) =
Answers to Selected Problems
4.23 Let x(t ) = ·.with
·
=
¸sì0 /
Then . = 0
x = O.
4.25
CHAPTER 5
325
5.1 The bounded input . = .I excites y( t ) ~ .I. which is not bounded. Thus the system
is not BIBO stable.
5.3 No. Yes.
5.6 = then y(t) ¯¯ . = .· then y(l) ~ · .· ·
5.8 Yes.
326 ANSWERS TO SELECTED PROBLEMS
5.10 Not asymptotically stable. Its minimal polynomial can be found from its Jordan fonn as t(;·) =
A(A + I ) . A = 0 is a simple eigenvalue. thus the equation is marginally stable.
5.13 Not asymptotically stable. Its minimal polynomial can be found from its Jordan fonn as teA) 
(A  1)2(A + I ). A = is a repeated eigenvalue. thus the equation is not marginally stable.
5.15 If N = I, then
M =
.
4. 8
It is positive defnite; thus al l eigenvalues have magnitudes less than I .
5.17 Because x'llx = x' [0.5 (M + M') Jx. the only way to check positive defniteness of nonsymmetric
M is to check positive definiteness of sym0tric O. S(M + M' ).
5.20 Both are Bl BO stable.
5.22 BlBO stable. marginally stable, not asymptotically stable. P (t) is not a Lyapunov transfonnation.
CHAPTER 6
6.2 Controllable, observable.
6.5
y = [0 I Jx +
Not controllable, not observable.
6.7 fi = for all i and f. =
6.9 Y = 2u.
6.14 Controllable, not observable.
6.17 Using XI and X2 yields
X =
·
O
x
. ·
.
o
··
y  [  I  I Jx
This twodimensional equation is not controllable but is observable.
Using XI , Xl, and X3 yields
X = ··
··
o 0
o 0 x
o 0
y = [0 0 I Jx
This threedimensional equation is not controllable and not observable.
6,19 For T = ¯. not controllable and not observable.
6.21 Not controllable at any t. Observable at every t .
Answers t o Selected Problems
CHAPTER 7
7.1
7.3
7.5
7.10
7.13
Not observable.
2
o U¬ 0
O
X +
0
o 0
y = [0 I Jx
For any and Cq, it is an uncontrollable and unobservable realization.
· .
¸ _
0
x +
0
. y = [0 I Jx
A controllable and observable realization.
Solve the monic null vector of
0
2
0 0
0 4
o
N
 I
Do
N
I
o DI
ð [0.5 0.5 0 I ]' . Thus 0. 5/(0.5 + 5) = +
= 0
X =
x +
0
I 1
y = [ I OJx
�1 (5) = 5(5 + 1 ) (5 + deg =
�2 (5) = (5 + + 2)
2
deg = 5
7.17 Its right coprime fraction can b computed from any lef fraction 8
(
(5) =
2. 5 5 + 0. 5
0. 55 5
2
+0. 55
1
2.5 5 + .. 5 ¯ 0. 5 ¯0. 5
or, by interchanging their two columns,
327
328 ANSWERS TO SELECTED PROBLEMS
.
=
5 + 0. 5 2. 5
]
[
52 + 0. 55 0. 55 ] 1
5 + 2. 5 2. 5 0.5 5  0.5
L'sing the latter. we can obtain
CHAPTER 8
8.1 k = [4 I ].
8.3 For
.
0. 5
X = I
0
Y = [ :
0. 5
2. 5
0. 25
0
0. 5
2. 5
] X
2. 5
0. 25
I
o X + 0
0. 5 0
F = 
I
O k = [ I I ]
o 2
we have

�
.
5
U
T =
[ 0 1 / 1 3
]
k = krl = [4 I ]
I 9/ 1 3
8.5 Yes. yes. yes.
8.7 u = pr  kx wi th k = [ 1 5 47  8] and p
=
0. 5.
8.9 u[k] = pr [k]  kx[k] with k = [ I 5 2] and p = 0. 5. The overall transfer function i s
• 0. 5(2�'  8: + 8)
gI l:)
=
_3
and the output excited by r [k] = a is )[0] = O. v [ l ] = a. v[2J = 3a. and y[k] = a for
k :: 3.
8. 11 Twodimensional estimator:
Onedimesnional estimator:
t = 32 + ( \ 3/2 1 ) u + y
x
=
[ 4 .
5 2 1 ¯
r^DSV\ers to Selected Problems 329
8.13 Select
3
F =
3

o 0
o 0
If K = �
0 I
�
[
62. 5 147 20 5 1�. 5 ]
0 0
thcn K *
0 0 0
If K = [�
0 0
�
606. 2 1 68  1 4. 2
0
then K =
1 1 9. 2 1 4. 9 2. 2
37 1 . 1
CHAPTER 9
9.2 C(s) = ( l 05 + 20)/(5  6). p = 0.2.
9.4 C(5) = (225  4)/(5  1 6). p = I. There is no need for a feed forward gain because g(s)
contains 1 /5 .
9.6 C(s) = (7s3 + 1 4s2 + 345 + 25)/(s(s' + 4» .
9.8 Yes, no. The transfer function from l to H is gu,(s) = sirs + I ) (s  2). which is not BIBO
stable.
9.10 Yes. no. no. no, yes. no (rowwise).
9.12 C1 (s) = 2(+3) /(s 2 \ ) . C2 (5) = (28s 6) /(5 2 1 ) . A(s) = 5 2 1 is not Hurwitz. Its
implementation in Fig. 9.4(a) wil l not b total l y stable. A minimal realization of [C 1 (s)  C, (s)]
t S
x = 2 1 x + [ 48  582] [ �, y
=
x + [2  28] [ :
]
from which an opamp circuit can b drawn.
9.15 ( 1 ) a > 0 and bo = 2bl . (2) Yes. bo = 2. bt = 4
.
9.16 The I x 2 compensator
1
C(s) = [3. 5s + 1 2 
2]
+ 3. 5
wi l l place the closedloop poles at 2,  1 ± j. 3. No.
9.18 If we select
F(5) = diag« s + 2) (s' + 25 + 2) (5 + 3) . (s2 + 25 + 2»
then we cannot fnd a feedforward gain to achieve tracking. If
F
[ (S + 2) (S2 + 2S + 2) (S + 3) 0 ]
=
I
S2 + 25 + 2
then the compensator
330 ANSWERS TO SELECTED PROBLEMS
·. ·
. ·
.
.
. · · . . · .. .
. · .. . ·.
and the feedforard gain matix

· .. .
P
·..
will achieve the design.
9.20 The diagonal transfer matrx
..
,
. .
.. .
o !
.
. .. .
is implementable. The proper compensators · 
I
... and · 
I
... with
· . . · · . ·
·�
.� 
. .
o . ·
.
·
·. 
.
.
..
in the twopaameter configuration will achieve the design.
' OÒOX
Actuating signal. 231
Adder. 1 6
Additivity. 8. 32
Adjoint. 52
Asymptotic stabil ity. 1 3 1 . 1 38
As:mptotic tracking. 242. 302
of ramp input. 276
of step input. 276
Balanced realization. 204. 227
Basis. 45
change of. 53
orthonormal. 46
Bezout identity. 272
BIO stability. 1 22. 1 25. 1 38. 1 92
Black box. 5
Blocking zero. 1 5. 305
Canonicalform. companion. 97. 203
controllable. 1 02. 1 87. 235
Jordan. 61 . 164. 1 82
modal. 98. 1 82. 241
observable. 1 88
Causality. 6. 1 0. 32
CayleyHamilton theorem. 63
Characteristic polynomial. 55. 225
of constant matrix. 55
of rational function. 1 89
of rational matrix. 205
Column degree. 2 1 2. 225
Columndegreecoeffcient matrix. 21 3 . 2 1 9
Column echelon form. 2 1 9
Column indices. 2 1 5
Companionform matrix. 54. 8 1
characterstic polynomial of. 55
eigenvectors of. 8 1
Compensator. 273. 283
feedback. 287
feedforward. 287
Compensator equation. 271
Configuration. See Control configuration
Control confguration. 270
closedloop. 267
controllerestimator. 270. 287
feedback. 267
noplantleakage. 283
onedegreeoffreedom. 270. 287
openloop. 269
plantinputoutput feedback. 270
twodegreeoffreedom. 270. 287
twoparameter. 270. 287. 291 . 307
unityfeedback. 273. 284. 287
Controllability. 144. 1 69. 1 76
after sampling. 1 72
from the origin. 172
Gramian. 145. l iO
index. 1 5 1
indices. 1 50. 225
matrix. 145. 1 69
to the origin. 1 71
Convolution. integral. 1 2
discrete. 32
Coprimeness. 1 5. 1 87. 2 1 1 . 2 1 4
left. 2 1 1
right. 2 1 1
Cyclic matrix. 256
Dead beat design. 267
Decoupling. 3 1 1
Degree of rational function. 1 89
McMillan. 205
of rational matrix. 205. 2 1 1
Description of systems. 2
exteral. 2
inputoutput. 2. S. 1 2
interal. 2
statespace. 2. 10. 35
Determinant. 52
Discretization. 90
Disturbance rejection. 243. 277
Di visor. common. 1 87. 2 1 0
greatest common. 1 87. 2 1 1
left. 21 0
right. 21 0
Duality. theorem of. 1 55
331
332 I NDEX
Eigenvalue. 55
assignment of. 23'. 256
ind'x of. 62
multiplicity of. 59. 62
Eigenvector. 55
generalized. 59
Equivalence. 95, I I �. 1 9 1
algebraic. 95, 1 1 2
Lyapunov, 1 1 4
zerostate, 96
Equivalence transformation, 95. 1 1 2, 1 52
Floquet. theory of. l i S
Fraction, polynomial, 1 89
coprime. 1 89. 1 92
matrix, 209. 2 10
Fundamental cutset. 27
Fundamental loop. 27
Fundamental matri x. 1 07. 1 08
gcd. 1 87
geld. 1 1 2
gcrd. 1 1 2
General solution, 50. 272, 282
Generalized eigen vector, 59
chain of. 59
grade of, 59
Gramian. controllability, 1'5. 1 70
observability, 1 56. 1 7 1
Hankel matrix, 201 . 226
Hankel singular value, 199
Homogeneity, 8. 32
Hurwitz polynomial. 271 , 285
Implementable transfer function, 285. 306
Impulse response. 10
matrix. 10
sequence, 32
Impulse sequence. 3 1
Integrator. 1 6
Inte�al model principle. 247, 279
Invener. 1 6
Jacobian. 1 8
Jordan block, 60
Jordanform matrix. 60. 1 M
Kalman decomposition theorem. 1 62
Laplace expansion. 52
Laplace transform. 1 3
Leverrier algorithm, 83
Linear algebraic equations, 48
Linear space. 45
basi s of. 45
change of basis, 53
dimension. 45
Linearity, 7
Linearization, 1 8. 30
Link. 27
Lyapunov equation, 70. 239
discrete, 13 5
Lyapunov theorem. 1 32. 1 35
Lyapunov transformation, 1 14
Markov parameter. 200. 225
Matrix. 4
cofactor of, S 2
cyclic. 256
function of. 65
fundamental, 107
inverse of, 52
left inverse of. 208
minor of, 52. 205
nilpotent, 61
nonsingular, 5 2
norm of, 7 8
orthogonal, 74
principal minor, 75
leading. 75
pseudoinverse of. 208
rank of. 49
rght inverse of. 208
singular value of. 76
state transition, 1 08
symmetric, 73
trace of. 83
1!i nimal polynomial, 62
1linimal realization, 1 84. 1 9 1
Minimum energy control. 1 49
�linor. 52. 205
Modal form, 98, 1 82, 241
Model matching, 283. 306
Model reduction, 200
Monic polynomial. 62
1lultiple, left. 2 1 0
right. 2 1 0
'lultiplier. 1 2
Nilpotent matrix. 6 1
:ominal equation, 243
:orm. of matrix, 78
Euclidean. 47
induced, 78
of vector, 46
:ormal tree, 27
Null vector. 49
monic. 1 94
;ullity, 49
Observability, 1 53. 1 70. 1 79
Gramian. 1 56. 1 7 1
index. 1 57
indices. 1 57, 225
matrix. 1 56, 1 7 1
Observer. See State estimator
Opamp circuit. 1 6
magnitude scaling, 98
Onhogonal matrx. 7.+
Onhogonality of vectors. 47
Orthonormal set, '7. 48
Parameterzation, 50. 272. 282
Periodic state system. 1 1 .+
Pole. 1 5
Pole placement. 273. 292
Polezero cancellation, 287
unstable. 287
Polezero excess i nequality, 285
Polynomial matrix, 2 1 0
column degree. 2 1 2. 225
columndegreecoefticient matrix. 2 1 3. : 1 9
column reduced. 2 1 2
echelon form. 2 1 9. 220
left divi sor. 2 1 0
left multiple. 2 1 0
right divisor. 2 1 0
right multiple. 2 1 0
row degree. 2 1 2. 225
rowdegreecoeffcient matrix. 2 1 3
row reduced, 2 1 2
unimodular, 2 ]
Positive defnite matrix. 74
Positive semidefnite matrix. 7.+
Primary dependent column. 1 91. 2 1 5. 2 1 9
row, 220
Principal minor. 75
leading, 75
Pseudoinverse. 208
Pseudostate. 1 86. 221
QR decomposition. 1 96
Quadratic form. 73
Range space. 49
Rank. of matrix . .+9
of product of matrices. 72
Rational function. 1 4
biproper. I S
improper. 1 5. 3.+
proper. 1 4
strictly proper, I S
Rational matrix. I S
biproper. 1 5
proper. 1 5
strictly proper. 1 5
ReachabIlity. 1 7 2
Realization. 1 00
I NDEX
balanced. 1 97. 204. 227
companionform. 203
controllableform. 10 I . 1 0'. 223
inputnormal. 1 99
minimal. 1 89. 207
observableform. 1 1 8, 1 1 9
outputnormal. 1 97
timevarying. 1 1 5
Reduced. column. 2 1 2
row. 2 1 2
Regulator problem. 242. 302
Relaxedness. 1 0
Response. 8. 3 1
impulse. 1 0
zeroinput. 8 . 3 1
zerostate. 8 . 3 1
Resultant. generalized. 2 1 5
Sylvester, 1 93. 274
Robust design. 243. 30.+
Robust tracking. 243. 278
Row degree. 2 1 2. 225
Rowdegreecoeffcient matrix, 2 1 3
Row echelon form. 120
Row indices. 220
Schmidt onhonormalization. S
Separation propeny. 255. 266
Servomechanism. 242
Similar matrices. S4
Similarity transformation. 51
Singular value. 76
Singular value decomposition. 77. 204. 226
Stability. asymptotic, 1 30. 1 38
BIBO. 1 22. 1 25. 1 38
discrete. I �6. 1 29
in the sense of Lyapunov. 1 30. 1 3 I
marginal. 1 30. 1 3 1 . 1 3 8
total. 284
Stabilization. 217
State. 6
pseudo. I �6. 1 2 1
variable. 7
State equation. 2. I I . I S
discretetime. 35. 9 1 . 1 1 0
discretization of. 9 1
equi valent. 95
periodic. I l
reduction, 1 59. 1 62, 1 8 1
solution of. 87. 93, 106
timevarying. 1 06. 1 1 0
State estimator. 2'8. 263
asymptotic. 2.9
333
334 I NDEX
State estimator (continued)
closedloop, 249
flldimensional, 248, 263
openloop, 248
reduceddimensional, 25 1 , 264
State feedback, 232
Statespace equation. See State equation
State tansition matrix, 108
Superposition property, 8
Sylvester resultant, 193, 274
Sylvester's inequality, 207
System, I
causal, 6, 10
continuoustime, 6
discretetime, 6, 3 I
dist buted, 7
linear, 7
lumped, 7
memoryless, 6
MIMO, 5
multi variable, 5
nonlinear, 7
singlevariable, S
S1SO, 5
timeinvariant, 1 1
timevarying, I I
Total stability, 284
Trace of matrix, 83
Tracking. See Asymptotic tacking
Transfer function, 1 3
discrete. 33
pole of, I S
zero of. I S
Transfer matrix, 14, 1 6
blocking zero. I S. 305
pole of, I S
transmission zero. 302. 3 1 1
Transmission zero, 302, 3 1 1
Tree, 27
branch, 27
normal, 27
Truncation operator, 38
Unimodular matrix, 21 0
Unittimedelay element, 7. 1 2. 14. 33
Vanderonde determinant, 81
Vector, 45
Euclidean normof, 47
Infinite nor of, 47
Inormof, 47
2norm of, 47
Wellposedness. 284
Zero. I S
blocking, 15, 305
minimumphase zero, 3 1 1
nonminimumphase zero, 3 1 1
transmission, 302, 3 1 1
Zeroinput response, 8, 3 1
Zeropolegain form, I S
Zerostate equivalence, 96
Zerostate response, 8, 31
ztransform, 32
THE OXFORD SERIES IN ELECTRICAL AND COMPUTER ENGINEERING
Adel S. Sedra, Series Editor, Electrical Engineering Michael R. Lightner, Series Editor, Computer Engineering
Allen and Holberg, CMOS Analog Circuit Design Bobrow, Elementary Linear Circuit Analysis, 2,(d Ed.
Third Edition
LINEAR SYSTEM THEORY AND DESIGN
Bobrow, Fundamentals of Electrical Engineering, 2nd Ed. Campbell, The Science and Engineering of Microelectroinic Fabrication Chen, Analog and Digital Control System Design Chen, Linear System Theory and Design, 3rd Ed. Chen, System and Signal Analysis, 2nd Ed. Comer, Digital Logic and State Machine Design, 3rd Ed. Cooper and McGillem, Probabilistic Methods of Signal and System Analysis, 3rd Ed. Fortney, Principles of Electronics: Analog & Digital Granzow, Digital Transmission Lines Guru and Hiziroglu, Electric Machinery & Transformers, 2nd Ed.
State University of New York at Stony Brook
Chi Tsang Chen
Franco, Electric Circuits Fundamentals
Hoole and Hoole, A Modem Short Course in Engineering Electromagnetics Jones, Introduction to Optical Fiber Communication Systems Krein, Elements of Power Electronics Kuo, Digital Control Systems, 3rd Ed. Lathi, Modem Digital and Analog Communications Systems, 3rd Ed. McGillem and Cooper, Continuous and Discrete Signal and System Analysis, 3rd Ed. Miner, Lines and Electromagnetic Fields for Engineers Roberts and Sedra, SPICE, 2nd Ed. Roulston, An Introduction to the Physics of Semiconductor Devices Sadiku, Elements of Electromagnetics, 2nd Ed. Santina, Stubberud, and Hostetter, Digital Control System Design, 2nd Ed. Schwarz, Electromagnetics for Engineers Schwarz and Oldham, Electrical Engineen'ng: An Introduction, 2nd Ed. Sedra and Smith. Microelectronic Circuits. 4th Ed. Stefani, Savant. Shahian. and Hostetter, Design of Feedback Control Systems, 3rd Ed. Van Valkenburg. Analog Filter Design Warner and Grung. Semiconductor Device Electronics Wolovich. Automatic Control Systems Yariv. Optical Electronics in Modem Communications, 5th Ed.
New York Oxford
OXFORD UNIVERSITY PRESS 1 999
OXFORD Athens Karachi Auckland
Oxford Bangkok
UNIVERSITY PRESS
New York Delhi Bogota Buenos Aires Florence Tokyo Calcutta Mumbai Warsaw Hong Kong Istanbul Toronto
Cape Town Chennai Dar es Salaam Nairobi Paris
Kuala Lumpur Madrid Melbourne Sao Paulo Singapore Taipei
Mexico City
and associated companies in Berlin Ibadan
Copyright © 1999 by Oxford University Press. Inc. Published by Oxford University Pr
198 Madison Avenue, New York, New York 10016 Oxford is a registered trademark of Oxford University Press All rights reserved. No part of this publication may be in any fann or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior permission of Oxford University Press.
,;,s, Inc.
BlHJAU
To
reproduced, stored in a retrieval system, or transmitted,
Chen, ChiTsong p.
Linear system theory and design I by Chi Tsong Chen.  3rd ed. Includes bibliographical references and index. ISBN 0195117778 (cloth). 1. Title.
Library of Congress CataloginginPublication Data
cm.  (The Oxford series in electrical and computer engineering)
I.
Linear systems. 1998
2. System design.
II. Series. 9735535 CIP
QA402.C44
629.8'32<1c21
Printing (last digit): 9 8 7 6 5 4 3 2 Printed in the United States of America on acidfree paper
I
6 DiscreteTime Systems 3 1 2.1.2 Basis.3 Linear Algebraic Equations 4 8 3.8 Some Useful Formulas 71 3.9 Quadratic Form and Positive Definiteness 73 3.7 Concluding Remarks 37 Problems 38 44 11 5 16 Chapter 3: Linear Algebra 3.6 Functions of a Square Matrix 61 3.1 OpAmp Circuit Implementation 2.1 Introduction 5 2.11 Norms of Matrices 78 Problems 78 45 vii .3 Linear TimeInvariant (LTI) Systems 2.1 0 SingularValue Decomposition 76 3.5.1 Causaliity and Lumpedness 6 2.1 Introduction 44 3.2 Overview 2 Chapter 2: Mathematical Descriptions of Systems 2. Representation.5 Examples 18 2.3.7 Lyapunov Equation 70 3.Contents Preface Chapter 1: Introduction xi 1.1 Introduction 1 .5 Diagonal Form and Jordan Form 55 3. and Orthonormalization 3.1 RLC Networks 26 2.4 Similarity Transformation 53 3.4 Linearization 17 2.2 Linear Systems 7 2.
1 Discretization 4.11 Concluding Remarks 227 Problems 228 7.3 Observability 753 150 6.1 Model MatchingTwoParameter Configuration 9.8.3.1 Regulation and Tracking 302 9.5 Stability of LTV Systems Problems 740 5. Chapter 8: State Feedback and State Estimators 721 8.3.8 Feedback from Estimated StatesMultivariable Case Problems 266 8.7 State estimatorsMultivariable Case 263 8.8 LTV State Equations 176 Problems 180 6.3 CanonicalForm Method 8.4 Lyapunov Theorem 5.5.2 State Feedback 232 247 247 231 8.1 Solving the Lyapunov Equation 135 131 8.4.2.6.2.6.2.5 Conditions in JordanForm Equations 6.3.4 7.3.2 Solution of DiscreteTime Equations 4.2 Controllability 744 6.6.1 Introduction 727 5.9 Realizations from Matrix Coprime Fractions 220 7.6 7.1 QR Decomposition 195 92 98 777 706 7.4 Effect on Transfer Matrices 8.4 Multivariable UnityFeedback Systems 9.3 Equivalent State Equations 4.7 7.1 Observability Indices 6.2.1 Canonical Forms 97 90 4.4.1 Compensator EquationsClassical Method 9.2 UnityFeedback ConfigurationPole Placement 9.2 Computing Matrix Coprime Fractions Balanced Realization 197 Realizations from Markov Parameters 200 Degree of Transfer Matrices 205 Minimal RealizationsMatrix Case 207 Matrix Polynomial Fractions 209 212 214 192 Contents ix Chapter 5: Stability 5.1 DiscreteTime Case 5.4 State Estimator 8.3.6.3 Implementable Transfer Functions 9.1 Controllability to the Origin and Reachability 772 764 Chapter 9: Pole Placement and Model Matching Introduction 171 9.5 Feedback from Estimated States 253 8.7 Controllability After Sampling 6.viii CONTENTS Chapter 4: StateSpace Solutions and Realizations 4.1 Column and Row Reducedness 7.3 Regulation and Tracking 8.1 DiscreteTime Case 4.1 DiscreteTime Case 729 732 126 8.6 Equivalent TimeVarying Equations 4.8.3.4.2 Solution of LTI State Equations 4.2 Robust Tracking and Disturbance Rejection 292 286 291 .1 Introduction 184 7.1 Introduction 231 8.2 LyapunovEquation Method 260 262 259 265 157 Chapter 7: Minimal Realizations and Coprime Fractions 7.2 Robust Tracking and Disturbance Rejection 280 9.1 Minimal Realizations 189 6.1 Robust Tracking and Disturbance Rejection 251 242 239 243 737 143 8.7 TimeVarying Realizations 77 5 Problems 777 I 4.6.2 InputOutput Stability of LTI Systems 5.1 Introduction 86 4.1 Cyclic Design Chapter 6: Controllability a n d Observability 6.5 Multivariable Model MatchingTwoParameter Configuration 306 9. 1 DiscreteTime Case 121 .3 Computing Coprime Fractions 7.3.4.2 Magnitude Scaling in OpAmp Circuits 110 93 87 86 7.1 Regulation and Tracking 275 9.2.10 Realizations from Matrix Markov Parameters 225 7.2 Stabilization 5.2.3.1.2 Implications of Coprimeness 7.3 Internal Stability 5.2.4 Canonical Decomposition 758 6.6 DiscreteTime State Equations 169 6.1 Controllablity Indices 8.6 State FeedbackMultivariable Case 255 8.2.2 Implementation of TwoParameter Compensators 303 283 277 273 9.5 7.1 Introduction 743 6.2.5 Solution of Linear TimeVarying (LTV) Equations 4.1 ReducedDimensional State Estimator 256 8.3 Embedding Internal Models 9.4 Realizations 700 4.3.8 7.1 269 271 269 185 184 9.
x CONTENTS 9. the abstract algebraic approach. For example. as the packagel seems to be the most widely available. The mathematical background assumed is a working knowledge of linear algebra and the Laplace transform and an elementary knowledge of differential equations. Phone: Email: info@mathworks. we develop mathematical descriptions of systems before reviewing linear algebra.lnc. We aim to achieve two objectives with this text.5.1 Decoupling 311 9.. MA 017601500. Aside from.com. some polynomial theory such as the SmithMcMillan form and B ezout identities are not discussed. MATLAB xi . is a registered trademark of the MathWorks. The geometric approach. 24 Prime Park Way. The chapter on stability is moved earlier.6 Concluding Remarks Problems 315 References 314 319 Answers to Selected Problems Index 331 321 Preface This text is intended for use in senior/firstyear graduate courses on linear systems and multi variable system design in electrical. in introducing polynomial fractions. I. rational fractions. hopefully.com. http://www. Thus it is mathematically less abstract than the original book. Thus most results are discussed with an eye toward numerical computation. we discuss first the timeinvariant case and then extend it to the timevarying case. 5086477000. all results are still stated as theorems for easy reference. Thus the presentation is not exhaustive. However. which was the expanded edition of Introduction to Linear System Theory published in 1970. In this text. a clearer presentation and a more logical development. In Chapters 4 through 6. The first objective is to use simple and efficient methods to develop results and design procedures. fax: 5086l7700 1 . mechanical. In this edition. and the solution of linear algebraic equations and the Lyapunov equation. this edition differs from the book in many ways: • • • The order of Chapters 2 and 3 is reversed. We give examples using MATLAB®. Natick. QR decomposition. instead of the other way around. and aeronautical departments. All design procedures in the text can be carried out using any software package that includes singular value decomposition. we limit our discussion to the conventional approaches of statespace equations and the polynomial fraction method of transfer matrices. and optimization are not discussed.mathworks. chemical. Linear system theory is a vast field. This edition is a complete rewriting of the book Linear System Theory and Design. It may also be useful to practicing engineers because it contains many design procedures. The second objective of this text is to enable the reader to employ the results to carry out design. This edition deals only with real numbers and foregoes the concept of fields.
Department of Electrical and Computer Engineering.6. 7.6. and model matching using a twodegreeoffreedom configura tion. 15 => [ Chap. However. the concepts of controllability and observability in Chapter 6 are useful. The presentation of the polynomial fraction method is streamlined. and 9. In statespace design. I thank my wife. College of Engineering and Applied Sciences. therefore. 7 Sec. I thank them all. some multi variable canonical forms. other arrangements are also possible for a onesemester course. 17 . The revised manuscript was reviewed by Professor Harold Broberg. 7 .3. • I am indebted to Mr. Professor Mustafa Khammash. Ross Barmish. 18. Terri O ' Prey. for her support during this revision. This text provides more than enough material for a onesemester course. for his encouragement. ChiTsong Chen • This edition is geared more for classroom uSe and engineering applications. A onesemester course at Stony Brook covers Chapters 1 through 6. the material in Section 7.19. University of Wisconsin. The eight appendices in the original book are either incorporated into the text or deleted. The people at Oxford University Press. 19. Furthermore. includi� g strict system equivalence. State University of New York at Buffalo. Examples using MATLAB are given throughout this new edition. Clearly. but not essential for the material in Chapter 7. EET Department. 3 => { Chap. B ihJau. 17. Jasmine Urmeneta. Department of Electrical and Computer Engineering. All minimal realizations in Chapter 7 can be checked using the concept of degrees. and Professor B. This configuration seems to be more suitable for practical application. Indiana Purdue University. I am grateful to Dean Yacov Shamash.67.49.4 may be skipped without loss of continuity. Professor Imin Kao and Mr. deterministic iden tification. The method is equated with solving linear algebraic equations. . Therefore Chapters 6 and 7 are essentially independent. and Kristina Della B artolomea were most helpful in this undertaking.5. 8 Chap.4. Finally. Sections 8 . computational issues. 7. Department of Mechanical and Aerospace Engineering. Professor Zongli Lin and Mr. Timevarying systems are not covered. Professor Peyman Givi. Their detailed and critical comments prompted me to restructure some sections and to include a number of mechanical vibration problems. Bill Zobrist of Oxford University Press who persuaded me to undertake this revision. This approach is not only easier for classroom presentation but also provides an attractive method for numerical computation.2. Juan Ochoa helped me with MATLAB. A solutions manual is available from the publisher. Anantakrishnan read the whole manuscript and made many valuable suggestions. The logical sequence of all chapters is as follows: Chap. Instead we discuss design in the twoparameter configuration. I am indebted to many people in revising this text. 6 => Sec. We then discuss pole placement using a onedegree offreedom configuration. Section 8 .3 Sec. Iowa State University. SUNY at Stony Brook. The polynomial fraction design in the input/output feedback (controller/estimator) configuration is deleted. many topics in the original book are deleted. 9. and decoupling by state feedback. T. 9.xii PREFACE • • Preface xiii The discussion of discretetime systems is expanded.8 => In addition. including Krysia Bebick.9 is needed to study Section 8.5 => Sec. Lyapunov equations are employed extensively and multivariable canonical forms are downplayed. instead of checking controllability and observability.
LIN EAR SYSTE M THE O RY AND DESIGN .
A physical system may have different models depending on the questions asked. In this text. It may also be modeled differently in different operational ranges. This approach r� lies heavily on past experience and is carried out by trial and error and has succeeded in designing many physical systems. development of mathematical descriptions. In order to develop a suitable model for a physical system. 1 Introduction The study and design of physical systems can be carried out using empirical methods. An inductor with a constant inductance is again a model: in reality. it must be modeled as a rigid body in maneuvering. we can adjust some of its parameters or connect to it a compensator to impro ve its performance. a thorough understanding of the physical system and its operational range is essential. We briefly introduce each of these tasks. Modeling is a very important problem. analytical methods become indispensable. we will call a model of a physical system simply a system. A spaceship can be modeled as a particle in investigating its trajectory: however. A resistor with a constant resistance is a model: it will burn out if the applied voltage is over a limit. we apply Kirchhoff's voltage and current laws to electrical systems and Newton's law to mechanical systems. For example. We can apply various signals to a physical system and measure its responses. Empirical methods may become unworkable if physical systems are complex or too expensive or too dangerous to be experimented on. analysis. Thus 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. an electronic amplifier is modeled differently at high and low frequencies. the inductance may vary with the amount of current flowing through it. If the performance is not satisfactory. Once a system (or model) is selected for a physical system. For ex ample. In these cases. The equations that describe systems may assume many forms: . for the success of the design depends on whether the physical system is modeled properly. and design. A spaceship may even be modeled as a flexible body when it is connected to a space station. The analytical study of physical systems consists of four parts: modeling. the next step is to apply various physical laws to develop mathematical equations to describe the system. circuits or control systems studied in any textbook are models of physical systems. The distinction between physical systems and models is basic in engineering. This power limitation is often disregarded in its analytical study. For example.Chapter Introduction 1 .
variable is a continuum defined at every time instant in ( . the state equation. and other pertinent topics that are essential for this text. Equations (l. fulldimensional and reduceddimensional state estimators.1) This equation describes the relationship between the input described by inputoutp ut or external description. Chapter yet) 11 G(t. we show how these equations can be developed to describe systems. x .stem must be designed to be stable. Chapter 5 introduces the concepts of boundedinput algebraic equations. and observability. We show how to obtain coprime fractions by solving sets of linear algebraic equations. simply. The last two chapters discuss the design of timeinvariant systems. we shall assume that models of physical systems are available to us. Depending on the problem under study. In conclusion. Using the concept of linearity. Thus the modeling problem should be studied in connection with the specific field and cannot be properly covered in this text. If a linear system is lumped as well.6) y<s) = G(s)u(s) G(s) is called the ( 1.ce . and asymptotic stability. In this text. all eigem'alues of a system can be arbitrarily assigned by introducing state feedback. After a mathematical description is obtained. difference equations.2 Overview 3 y(t) = L G(t .2) is a set of firstorder differential equations and Equation (1. Because the vector is state. because design techniques may often evolve from this study.3) is a set of algebraic equations. such as stability.4) yields For this class of linear timeinvariant systems. If the model is properly chosen. we are interested in the general properties of systems.l) through ( 1 . r)u(r)dr = 10 (1. otherwise. the property of time invariance. the Laplace transform is an important tool in x(t) Ax(t) BuU) y(tl Cx(t) Du(t) + = + 1 1 . Both ( 1 . marginal stability.5). in other cases. the set of two equations is called the x(t) A(t)x(t) B(t)u(t) y(t) CU)x(t) D(t)u(t) internal statespace = + ( 1 . after introducing the aforementioned equations from the concepts of lumpedness. the system must be modified. a system may have different mathematicalequation descriptions just as a physical system may have many different models.2 I NT RODUCTION they may be linear equations. t 1 . Note th the design is carried out on the model of the and through ( 1. it cannot be achieved by use of pencil and paper. with any desired called the Equation (1. then Equations ( 1. Developing models for physical systems requires knowledge of the particular field and some measuring devices. We also introduce the Lyapunov theorem to check asymptotic stability.. We briefly discuss the contents of each chapter. then the corresponding equations are called discretetime equations. Chapter 7 studies minimal realizations and introduces coprime polynomial fractions. A fundamental result is that the transfer matrix describes only the controllable and observable part of a state equation. In some cases. integral equations. 1 ) through ( 1 . Developing models for automobile suspension systems requires actual testing and measurements.6) are called continuoustime equations because their time only at discrete instants. Computer simulation certainly helps but cannot replace actual measurements. Qualitative analysis is very important. physical system.3) = + 8 and use coprime polynomial or. we then carry out analysesquantitative and/or qualitative.3) reduce to If a linear system has. and time invariance. 1 ) 1. or others. We use controllable and observable state equations to carry out design in Chapter fractions in Chapter 9. If the time is derined transfer matrix. then it can also be u and output y 6 introduces the concepts of controllability and observability. compensators must be introduced. differential equations.4) and ( 1 . to develop models for transistors This text is devoted to the analysis and design centered around (1. Selecting a model that is close enough to a physical system and yet simple enough to be studied analytically is the most difficult and important problem in system design. If a state equation is observable. We study in Chapter 4 solutions of the statespace equations in ( 1 . The equivalence of controllable and observable state equations and coprime polynomial fractions is established. the Lyapunov equation. If the model is � analysis and design. in addition. and design.7) are inputoutput or external descriptions. setting up mathematical equations. We also introduce the Jordan form because it will be used to establish a number of requires a knowledge of quantum physics and some laboratory setup. it may bum out or disintegrate. :xl). controllability. we are interested in the responses of systems excited by certain inputs.7) where a variable with a circumflex denotes the Laplace transform of the variable. analysis. under the controllability condition. then the performance of the physical system should be improved by introducing the required adjustments or compensators. Different analyses may lead to different state equations tliat describe the same system. The basic relationship between statespace equations boundedoutput (BlBO) stability. this can be achieved by adjusting some parameters of the system..2) ( 1 . Applying the Laplace transform to ( 1. Thus we introduce the concept of equivalent state equations. In qualitative analysis.l ) poor. Every s:.r)u(r) dr = ( l . For example.2) and ( 1 . We show that. Chapter 3 reviews linear and transfer matrices is also established. They are essential and is called the in studying the internal structure of systems. linearity. In quantitative analysis. Therefore stability is a basic system concept.7) and their discretetime counterparts. one form of equation may be preferable to another in describing the same system. In Chapter 2. 2 Overview The study of systems consists of four parts: modeling. nonlinear equations. If the response 0f a system is unsatisfactory. we develop in Chapter 2 that every linear system can be described by results. They are called the description of linear systems.5) ( 1. then the performance of the physical system may not improve and the design is useless. The systems to be studied in this text are limited to linear systems. The function The former is said to be in the time domain and the latter in the frequency domian.
. In Chapter 9. .1 � I l � 2 3 45 System. a singleinput multioutput (SIMO) system if it has one input terminal and two or more output terminals. Chapter f . 1 Introduction The class of systems studied i n this text i s assumed to have some input terminals and output terminals. A system with only one input terminal and only one output A system with two or more input terminals and/or two or more output terminals is called UII) U ['J o Figure 2. In our design. terminal is called a singlevariable system or a singleinput singleoutput (SISO) system. We assume that if an excitation or input is applied to the input unique relationship between the excitation and response. we discuss pole placement. disturbance rejection. or cause and effect is essential in defining a system. and decoupling. no control performances such as rise time. More specifically. and their applications in tracking. 2 . 2. can be constructed to generate estimates of the state. We use the unityfeedback configuration in pole placement and the twoparameter configuration in model matching. we can call a system a multiinput multioutput (MIMO) system if it has two or more input terminals and output terminals. We also establish the separation property. input and output. a multi variable system. . k u[k) Ulf) B lack box y[k) y(l) y [k) o y(r) i � 1 234 5k 5 . This terminals as shown in Fig. model matching. settling time. � .�mI Mathem atical Descri ptions of Systems ).4 INTRODUCTION eigenvalues. 1 . all results are basic and useful in designing linear time invariant control systems. However. Therefore this is not a control text per se. c. . and overshoot are considered: neither are constraints on control signals and on the degree of compensators. a unique response or output signal can be measured at the output terminals.
The concept of state can deal with this problem. All discretetime signals in a system will be assumed to have the same sampling period T. Therefore the initial state  yet) = u (t  1) 1 :::: t < to}. t::: toJ. 2. The entries of x are called state variables. 2 . if we know the voltages By definition. where the prime denotes the transpose. t = to. A network that consists of only resistors is a memory less system. 2. if not impossible. u [k] := lI(kT) and y[k] := y(kT). we may consider Using the state at to.2 Linear Systems Network with 3 state variables. t::: to } . the input from :xJ to time t has an effect on yet)· Tracking n(t) from t = .2 2. 1 ) is easier to track and will be called a stateinputoutput pair. we need the information (u (t) . I n order t o determine (y(t). together with the input n(t). In other words. to. Similarly. t::: to (2. + and generates discretetime signals as its output. the output will be denoted by y (t) or yet). its state consists of three numbers. f t::: to. 2 Linear Systems A system is called a linear system if for every to and any two stateinput{)utput pairs Xl (to) and X2 (tO ) across the two capacitors and the current X3 (to) passing through the inductor. By this we mean that the output at to depends on to.6 MATHEMATICAL DESCRIPTIONS OF SYSTEMS A system is called a continuoustime system if it accepts continuoustime signals as its input and generates continuoustime signals as its output. Thus in some sense. EXAMPLE 2. to . then for any input applied on and after to. That is. u p ]'. This text studies only causal systems. No physical system has such capability. For the network shown in Fig. 2. t ::: to . If the system h. In other words. to The output i s simply the input delayed b y one second. past input affect the current output? Generally.1 Consider the unittime delay system defined by A system is said to be lumped if its number of state variables is finite or its state is a system is causal and causality is a necessary condition for a system to be built or implemented in the real world. very inconvenient. the time should go all the way back to minus De finition 2. we can express the input and output of a system as on past. 2. we have xJto) ni(t).1 :::: t < to}. Thus the state of the network at time to is for i = I.Yi (I). If a system is not causal.1.2 is clearly a lumped system. the state summarizes the Thus the unittime delay system is a distributed system. ifwe know the state alto. The network in Fig. Therefore every physical will depend on future input.y(t). and future inputs. determines uniquely the output y (t) for all t ::: to· or effect of past input on future output. t ::: to 00.1) It means that the output is partly excited by the initial state at to and partly by the input applied way back to at and after to. The time t is assumed to A system is called a discretetime system if it accepts discretetime signals as its input R. 7 range from 00 to p input terminals. < the initial state simply as a set of initial conditions. . in general. 1 ). and t > x(to) n(t).2.s DC. current. The input and output will be denoted by ranging from 00 to x. . Figure 2. it is independent of the input applied before or after to . In using (2. have memory. A system is called a causal or nonanticipatory system if its current output depends on past and current inputs but not on future input. 1 Causality and Lumpedness A system is called a memol)'less system if its output y (to) depends only on the input applied at to. How far back in time will the infinity. The input will be denoted bv lowercase italic II (t) for single input or by boldface n(t) for multiple inputs. There are infinitely many points in {to  I :::: t < to}. there is no more need to know the input applied before to all the Thus (2. t ::: to} from {u (t) . we can determine uniquely the output for t ::: to. however. A system is called a distributed system if its state has infinitely many state variables. It is a 3 x I vector.1 The state x(to) of a system at time to is the infonnation at to that. The transmission line is the most well known distributed system. We give one more example. Thus. there is no more need to know the inputu(t) applied before to in determining the output y (t) after to . then its current output will be applied in the future.:xJ is. then n(t) is a p x 1 vector or n = [Ill 112 . a noncausal system can predict or anticipate what finite vector. This will be stated succinctly as follows: current output of a memoryless system depends only on current input: it is independent of past and future inputs. of the system is (u (t). net) for t Most systems. current output of a system with memory may depend } . where k denotes discretetime instant and is an integer They become boldface for multiple inputs and multiple outputs. Current output of a causal system is affected by past input.
I.3 has height + with height expressed symbolically as 6. == Figure 2.2 L inear Systems output is excited exclusively by the input. These two properties can be combined as for any real constants If the input and and is call�d the a nonlinear system if the superposition property does not hold. the complete response can be very different from the sum of the zeroinput response and zerostate response.t. t::': to } Yzs(t). (t. t::': to superposition property. (t. then the output will be excited exclusively by the input.3. the second. and all Ui. al a2.:. Let be the pulse shown in Fig. This output is called the or == then the output will be excited exclusively and will be denoted by Thus the output 0.(t . We consider first SISO linear systems. y(t) u(t) y(t) L g. 8".)6.   i U i + + (homogeneity) (additivity) excited by the input "'" can be approximated by (2. t::': to } aYI (t). (t tl) 9 tl. 1/6. 1/6. =0 Thus the response of every linear system can be decomposed into the zerostate response and the zeroinput response..)6. If the initial state output is called the is zero.)u(t. 1 u (t.ti) g". 2. t::': to a. Furthermore. responses..)u(t. til be the output at time Then we have property. the two responses can be studied separately and their sum yields the complete response.)6.:.(t t ) (t )6. then the additivity and homogeneity properties apply to zerostate U(tl) I' I Response zeroinputresponse + zerostate response I I linear.. t::':to } alYI (t)+a2Y2(t). ti) 8". if = 0. t. )u (t. Consequently. L 8". The pulse aX I (to) aUI (t)..zeroinput response Y:i x(to) u(t) t to } Y:i (t). .4 can be expressed as . Therefore we cannot separate the zeroinput and zerostate responses in studying nonlinear systems. To be more specific.2) =0 The additivity property implies Output due to { x(to) t::': to u(t). is identically zero for A system is called by the initial state x(to).8 MATHEMATICAL DESCRIPTIONS OF SYSTEMS 2.(t. 1 I.t. 2.) u (ti)8'" (t til6.(t . L g.3 Pulse at tl' 0. In this study.til 8".4. additivity homogeneity al xl (to) + a2x2(to) alul(t)+a2u2(t). the input (t) can be property. A similar response of linear systems. = output due to + or output due to { x(to) t::': to u(t) { x(to) t::': to u(t). {u I + Uz + x(to) Y + Y2} {aui aYi} I and + for all a Yi}. thus has height  and height and is located at time u 8". then we have remark applies to zeroinput responses of any linear system.(t t. 2. It has width every input u(t) can be approximated by a sequence of pulses as shown in Fig.)6. t::': to x(to) response zerostate Yos x(to) uU).4 Approximation of input signal. the Let g". This and will be denoted by + ::': + or t u(t) 8dt .:.(t. t.(t. t::': to (homogeneity) for any real constant The first property is called the + and in Fig. excited by the pulse + = applied at time ti. uit) t::': to. For nonlinear systems. the initial state is assumed implicitly to be zero and the Inputilutput description We develop a mathematical equation to describe the zerostate Figure 2. then the output will be excited exclusively by the input and the stateinputoutput equation can be simplified as lUi + If the system is If a system is linear. Then and the leftmost pulse in Fig. + t.)lI(ti!6. g. 2.
relaxed at to.6) (2.r) g21(t. T) is the impulse response between the jth input tenninal and the ith output tenninal.1 For a pinput qoutput system. Simply.2 is time invariant if R" If a signal is time invariant such as u (t) to appear from time to + T . a mechanical system. For example. r) . In this case. r) is a function of two variables. the output will not appear before an input is applied.6) and (2. a chemical process. Thus the upper limit of the integration in (2. and D must be n x n. we have x(to) yet) = where l' G '0 G(t.T) gqp(t. if the initial state and the input are the same.T)] g2p(t.7) x(t) = A. Signals are mostly time varvina. This description is developed using only the additivity and homogeneity properties. and L. that to = O. q x n. I::: to > y(t . r) is the response excited by an impulse. it is called the impulse response. is called the impulse response matrix of the system. approaches zero.2) becomes an equality. without loss of generality. then it is a very simple or a trivial si na . the approximation by a continuum and can be replaced y r. the dimension is infinity and the two equations The inputoutput description in (2. the pulse 0". its characteristics do not change appreciable in the first couple of years. slale equation. = i: g (t.3 ) in (2. is excited exclusively by the input u (t) for t ::: to. n x p. In other words.5) was developed from the linearity condition. a burning G rocket is a timevarying system. for timeinvariant systems. C. t::: to + T (time shifting) + =1 for all I. It means that if the initial state is shifted to time to T and the same input waveform is applied from to + T instead of from to. If the system is causal as well. T) system IS not time invariant. by definition. Therefore any lumped or distributed linear system has such an inputoutput description. then g et. the output waveform will always be the same. are constants. and to. B. gij (t. We use A =: B to denote that B. the � l C. is not as simple and Will not be attempted.3) can be replaced by to. r) every linear system that is causal and relaxed at to can be described by for t ::: to.10 MATHEMATICAL DESCRIPTIONS O F SYSTEMS the corresponding output will be denoted by g et . If a linear system has p input terminals and q output tenninals. has such a description. Note that g(t.7) to be compatible. the summation becomes an integration. e uatlOns.4) A system is said to be ti e invariant if for every stateinputoutput pair m and any T. Equation (2. the output yU). then the output waveform will be the same except that it starts x(to + T) u(t). In this derivation.6) and (2. r) = [gll(t.r) = 0 for t < T If a system is causal. As to.eriod. Thus we have U) the impulse input is applied. system is described by (2. Thus (2. g[2(t. equals B.6) actuallv consists of a set of n firstorder differential equations. r)u(r)dr (2. and gij (t. The where x := dX/dt. 2.3 Linear TimeInvariant (LTI) Systems 11 (2. then (2. because its mass decreases rapidly with time. u is a p x vector and y is a q x 1 vector. I::: to u(t . We use A := B to denote that A. Some physical systems must be modeled as timevarying systems. In conclusion. 0 for development of the statespace equation from the linearity condition. T) is the response at time t at the i th output tenninal due to an impulse applied at Thus time T at the jth input tenninal. not for signals. The four matrices are all functions of time or timevarying matrices.T) . . Thus the lower limit of the = t < r. r)u(r)dr (2. by definition. denoted by oCt td.T) (t . For distributed systems. the inputs at other terminals being identically zero. approaches zero. For example. r)u(r)dr (2. The second variable denotes the time at which Because gU. 7) consists of q algebrai Y(I) = C(t)x(t) A(t)x(l) + + D(t)u(t) B(I)u(r) 1 � � A system is said to be relaxed at to if its initial state at to is O. can be written as dr. then x is an n x I vector. network in Fig. Equation (2.3) can be replaced by t. t::: to + T } } � Y(I). Causal ¢=> g�lt. and causal.5) gIPU. or any other system. integration in (2. the system is linear. T) g 22(t.T) The characteristics of timeinvariant systems must be independent of time. We stress once again that if a gql(t.. and q x p matrices. 3 Lin e a r TimeI nvariant (LTI) Systems g U . we can always assume.7) are not used. be it an electrical system. Therefore. In order for the matrices in (2. If the system has n state variables. . If a Time invariance is defined for systems. The set of two equations will be called an ndimensional statespace equation or. equals A. therefore every linear system. the discrete ti becomes 2. however. it is said to be time va rying. no matter at what time they are applied. Although the performance of an automobile or a TV set may deteriorate over a long period of time.4) can be extended to yCt) = l' '0 2 . Uti) becomes an impulse atti.5). We will simply accept it as a fact.2) becomes Now if to. Thus a large number of physical systems can be modeled as timeinvariant systems over a limited time of the form Statespace description Every linear lumped system can be described by a set of equations 1. That is.T ) . the condition of lumpedness is not used. til. the first variable denotes the time at which the output is observed. and in (2.
8) into the algebraic equation in (2. = .((5S))l l'S'q(s) gql(S) g up( qp(S) s) Y(5) G(5)I1(S) · · · = (2. r) = y ( t) = Inputoutput description The zerostate response of a linear system can be described by (2 . . then we have' by O. the impulse response is the inverse Laplace transform of the transfer function.iu"iant system to be causal is = 0 for t < O. = get + + g(t g(t l' g(t l' 10 g g get) g(t 1 g(t) 8 8(1 a 8(1). we use a variable with a circumflex to denote the Laplace transform of the variable.r) r(r) dr = a 8 ( t . Substituting (2. 8) and interchanging the order of integrations. that is. The second equality can easily be verified by changing the where we have replaced variable.Sl' dV ) ) u (r)e sr dr yes) 1: = g ( v) esr d v 1: u(r) esr dr (2.2 The unittime delay system studied in Example 2. It consists of a equals the input delayed by I second.r. 2. In analysis and design.i) Unittime delay element Unitlime delay element ylt) is called the transferfunction of the system.r. However. By definition = . �ote that g(r.10) can be extended as 2. . Now if the system is time i nvariant as well.0) is the output at time due to an impulse input applied at time O.12 MATHEMATICAL DESCRIPTIONS OF SYSTEMS for any T.5(a).I). Y�2I((SS))llg�2Il((Ss)) ggd2I (Ss)) g�2IP�((SS))llU:2l. conversely.air(r{�'_i .8) Throughout this text. the output is . Unlike the timevarying case where is a function of two variables. Thus (2. If we apply the impulse (t) at the input terminal. 8) can be replaced by 00.3 L i near TimeInvariant ( LTI) Systems 13 y (t) eS'dt T)U(r) dr ) g(t  esur)e SOdt r=O r) eSUrJdt u (r)e srdr g (v)e .10) g(s) lx = . Thus the convolutio n in (2. for convenience.  (a) (b) and g(t r) or are two different functions. r T) = . it is simpler to use algebraic equatIOns than to use convolutions.r) dr (2. . r) Figure 2. after introducing the new variable v = t . we obtain = Transferfunction matrix The Laplace transform is an important tool in the study of linear timeinvar iant ( LTI) systems.i) (2. 8) is called a convolution integral.10). If r (t ) = then the output is the impulse response of the feedback system and equals  EXAMPLE 2.. g ( t) eHdt For a pinput qoutput system. the integration becomes independent of r and the double integrations become = which becomes. . T. (2. The integration in (2. where or Again using the causality condition to replace the lower integration limit inside the parentheses from v = r to v = O. we have r) = 0 for r > t.5 Positive and negative feedback systems. s o is the feedback system.9) l=! gf(t .3 Consider the unityfeedback system shown in Fig. Let r (t) be the input of the feedback system. Thus = 0 for t < 0 or the upper integration limit in (2. We see that the Laplace transform transforms the convolution integral in (2. Thus the transfer function is the Laplace transform of the impulse response and. 0) =  r) r)u ( r) dr = g ( r)u(r . a8(t +a28(t +a38(1 3)+ 2.i ) r (r) dr = b(t . The condition for a linear timeinv. then the output is given by YO  . 4).:>i8(1 l' fai l' i�l . Let be the Laplace transform of y ( t) . is a function of a single variable in the timeinvariant case.r .4) reduces to get .2) < Because the unittime delay system is distributed. == 0 for t 0. 8) will rarely be used in the remainder of this text.11) .  Let r (t) be any input with r ( t) gf ( l) = y (t) = a multiplier with gain and a unittime delay element. the same symbol g is used.1) . 1 is a device whose output EXAMPLE 2. t=O r=O [=0 =  yeyess)) lX g(t) S(s) foo (IX g(t 1'" (fX g(t S'(s) 1: (l�r =  2. It is a SISO system. Thus the impulse response of the system is 8 ( t I). For causal systems.
(s). and q x p constant matrices.13) s. we can express the infinite series in closed form as yields which is the same as (2... this form can be obtained from the transfer Z2)'" .A)IBu(s) + Du(s)  (2. respectively. There are a number of ways of defining zeros for G(s). that is. Applying the Laplace transform to (2.s i .(s) is called the transferfunction matrix or. The transfer function from r to the unittim from the block diagram as Improper rational transfer functions will amplify highfrequency noise. which often exists in the real world.(s) is square and if both (..I)e"dt = s...9) as function by calling [ z .1 (s) are proper. Then g(s) can ials of s. In MATLAB.4 Consider the unittime delay system studied in Example 2. Given x(O) and u(s). Thus every pol� of every entry of (.(s) if it is a pole of some entry of (.12) (s .P2) . A.I)] = lC<) 8(t .. function = N(s)/ D(s). We study mostly s we will encounter are mostly rational functions of s. where N(s) and Every rational transfer function can be expressed as g(s) ial.�1)(S (s . p . The equations also reveal the fact that the response of a linear system can be decomposed .15). n X p. i=O = 00 (. If a linear timeinvariant lumped systems. g(s) biproper 9 deg D(s) = deg N(s) 9 g(oo) = nonzero constant.(JO) is A rational matrix (. This can also be obtained by taking the Laplace transform of the impulse response.(s) is said to be biproper. have no common factors of degree 1 or higher. therefore improper rational transfer functions rarely arise in practice.Zm) g(s) = k (s . kJ gj(t) = Because L[o(ti)] = e. then its transfer system is a distributed system..12) is an irrational function of system is lumped. We call A a pole of (. then (. They are algebraic equations.(s). which will be introduced in Chapter 9.2. In terms of poles and zeros. 1). and all roots of D(s) are the poles of g(s). Let us use deg to denote the degree of a polynom D(s) are polynom be classified as follows: • This is so because the feedback which implies xes) = (sl . 2. If a rational matrix or nonzero constant matrix: it is strictly proper if its every entry is strictly proper or if (.5 g(s) = L[o(t . and D are.12). . xes) and hs) can be computed algebraically from (2. g(s) improper 9 deg D(s) < deg N(s) 9 Ig(ee)1 = 00.14) and (2. n.(s) is a pole of (. den). q outputs. y can be computed directly e delay element is es. gj(t) is gj(s) = L[gjU») = Using 00 L a'eiS = aes L(aes).14 MATHEMATICAL DESCRIPTIONS OF SYSTEMS where guts) is the transfer function from the jth input to the ith output. then all roots of N(s) are the zeros of g(s)..14) (2. transfer matrix of the system.5(a). We call A a blocking zero if it is a zero of every nonzero entry of (. The transfer function in (2. A real or complex number A is called a pole of the proper transfer function g(s) = N(s)/ D(s) if Ig(A)1 = 00. and n state variables. which was computed in (2. the transfer function can be expressed as aeS gj(s) = � 00 i } (2. p inputs.3 linear TimeInvariant (IT!) Systems 15 est t=l l = es The transfer function of Consider the feedback system shown in Fig.. Its impulse • • response is o(t  • This transfer function is an irrational function of EXAMPLE 2. The q x p matrix (. 2::> o(t .15) g(s) proper 9 deg D(s) ::': deg N(s) 9 g(oo) = zero or nonzero constant.i) '" i=l the Laplace transform of i=l .  Statespace equation of equations of the form Every linear timeinvariant lumped system can be described by a set '"' r ' � i=O I Ir x(t) = Ax(t) + Bu(t) y(t) = Cx(t) + Du C t) For a system with q x for Ir I < 1. thus the transfer function will be a rational function of s. Therefore its transfer function is g(s) strictly proper 9 deg DCs) > deg N(s) 9 g(oo) = O. A more useful definition is the transmission zero. EXAMPLE 2.(s) and (. C. (s PII) This is called the zeropolegain f0rm. If N(s) and D(s) are coprime.(s) is said to be proper if its every entry is proper or if (.A)lx(O) + (slA)IBu(s) hs) = C(sl A)IX(O) + C(sI . a zero if g(A) = O.pIles ..(00) is a zero = t f 2 zp ( num.. simply. n x n. B.(s).13) sx(s)x(O) = Ax(s) + Bu(s) hs) = Cx(s) + D u(s) (2. a zero matrix. 2. Their inverse Laplace transforms yield the time responses xU) and y(t).
: C ::: � + (ax\ (a) + CX3) l'� Ric Two opamp circuit elements. an integrator.5u. the second element.6.3.3. 2 .17) RI2 RiO. 1 8).ign the outputs . (b)  x = aVI x RC = I + b V l ( V3 + R R Figure 2. then y(t) = [2 X I (t )] + 3] [ X2(t) 5u(t) Y(s) Comparing this with (2. Not shown are the grounded non inverting terminal and power supply. L[A(t)]L[x(t)]. I S) can be implemented as shown in Fil!. 1 ) computes the transfer matrix from the first input to all outputs or. X I = 2xl 0 .17) and (2. equivalently. If the initial state (2. and is connected as shown. 2. X I . 2. We have the freedom in choosing the output of each integrator as +Xi or Xi . 1 OpAmp Circuit Implementation Every linear timeinvariant (LTI) statespace equation can be implemented using an operational amplifier (opamp) circuit.16 MATHEMATICAL DESCRIPTIONS O F SYSTEMS as the zerostate response and the zeroinput response. The input of the RHS mtegrator should be . I S) x(O) is zero. from the first X equation of ( 2.6(b). Actually. then its input should equal y = 2 x 3X . 3X2 + 2u and is connected as shown.SX2 and is connected as shown. the first c. the range of signals is limited.� = . Now we use an example to show that every LTI statespace equation can be implemented using the two types of elements in Fig.6 /b :X3� bx� Ria R Ric  0. 2.I B + D . instead of X I and X2. G(s) = C (sIA). If the output of the adder IS chosen as y. 3 t� x. [ num. then the output of the element is . For example. and ax the element can be used as an inverter with gain a. Figure 2. To conclude this section. If the feedback branch is a resistor as shown in Fig. then the output equals.7. The Laplace transform of get. They compute only the SISO and SIMO cases. If the last argument 1 in s s 2 t f ( a . 1 ) is replaced by 3.15) reduces to 2. The functions t f 2 s s and s s 2 t f in MATLAB compute one description from the other.7 as XI and Xl. 1 7 ) and (2.(axi +bX2 +CX3). the adder functions also as multipliers and the integrator functions also as multipliers and adder. through resistors. in Fig. usually I or :2 volts below the supplied voltage. c . In actual operational amplifier circuits. column of G(s). 2. Suppose we assign the output of the lefthandslde (LHS) mtegrator as X I and the output of the righthandside (RHS) integrator as . if we as . 1.17 ) . 2. a way to deal with this problem. d. Consider the state equation 2. of the two mtegrators m Fig. and if the output is assigned as x. b . then .4 Linearization Most p hysical systems are nonlinear and time varying. however. For example. If we use only one input. If the feedback branch is a capactor with capacitance C and RC = 1 as shown in Fig.�2 = X I . then we will obtain a different implementation. b . thus the Laplace transform does not offer any advantage and is not used in studying timevarying systems. R RI8 RC = 1 R R /I R.6(a).3 ] [ X I (t) ] S X2(t) + [ .16) function generates the third column of G(s ) . the circuit will saturate or bum out and the circuit will not behave as the equation dictates. . Some of them can be described bv the nonlinear differential equation of the form R R C u Figure 2. then the It has dimension 2 and we need two integrators to implement it. + 2.11) yields = [C(sIA)IB + D]u(s) (2. d .2 ] u (t ) 0 (2. say.6(a).l . r) is a function of two variables and L[A(t)x(t)] f. Then the input of the LHS integrator should be.6 shows two standard opamp circuit elements. we mention that the Laplace transform is not used in studying linear timevarying systems. 2. All inputs are connected.7 Opamp implementation of (2. as we will discuss in Section 4.7 . If any signal grows outside the range. to the inverting terminal.4 L i nearization 17 (2. I 2 Thus the state equation in (2. Note that there are many ways to implement the same equation. There is.2 as shown in Fig.(a v i + b t'2 + CV3 ) ' We call the first element an adder. den ] = s s 2 t f ( a . This relates the inputoutput (or transfer matrix) and statespace descriptions.
and viscous fnctlOn as shown in Fig.20) and neglecting higher powers of x and ii.6 3. the solution can be expressed as xo (t) + x(t) with x(t) small for all t. Suppose for some input function Uo (t) and some initial state. 5 Examples x(t) yet) == == b(x(t) . u(t) . 1 9) where b and f are nonlinear functions. The behavior of such equations can be very complicated and its study is beyond the scope of this text..19) as + Figure 2. as a lInear system under linearization and simplification. uo(t) ii(t). t) f(x(t). Xo (t) is the solution of (2. Thus the mechan ical system can be modeled . The fnctJon IS clearly not a linear function of the velocity. (2. 1 0: it is not linear. Figure 2. �haracteristics of the spring are shown in Fig. yielding the phenomenon of chaos. that is. we can reduce (2. Because A and B are computed along the two time functions xo (t) and uo (t). Consider the mechanical system shown in Fig. == 19 / k2 Mechanical system.2 1 ) Force Force o Static Velocity . t) can be similarly linearized.t equations.8. however. 2 . For some nonlinear equations. 2. Y ) as shown. 2 1 ) to x(t) == A(t)x(t) + B(t)ii(t) == This is a linear statespace equation.(t). u(t) .9. In this case. It consists of a block with mass m connected to a wall through a spring. Note that the horizontal coordinate is velocity )' dy dt . ).8 and displacement y from the equilibrium to b e the output. However. t) + au ax + ab _ (2. This linearization technique is often used in practice to obtain linear equations. 2.18 MATHEMATICAL DESCRIPTIONS O F SYSTEMS 2 . Some nonlinear equations. Using (2. can be approximated by linear equations under certain conditions.(a) Static and Coulomb frictions. We consider the applied force u to be the input This is not true in general. Uo (t). functions of t. The equation yet) f(xU). if the displace The ment IS IUlll ted to (YI . the statIc and Coulomb fnctlOns and consider only the viscous friction .3 Under this assumption. 1 9). Then the friction becomes linear and can be expressed as k. For some nonlinea. . t) ab _ u+. t) (2.�_ (b) Viscous friction V elocity They are called lacobians. i n general. (b) Viscous friction. Coulom b friction .20) Now suppose the input is perturbed slightly to become Uo (t) ii(t) and tbe initial state is also perturbed only slightly. we disregard . a very small difference in initial states will generate Displacement completely different solutions. we can expand (2. where is the spring constant. The friction between the floor and the block generally consists of three distinct parts: static friction. u(t) . f ] EXAMPLE 2. 2 . 5 Exa mples Figure 2. . then 2 the spring can be considered to be linear and the spring force equals k2 y. where k is the viscous l friction coefficient.10 Characteristic of spring.. x+ b(xo (t). they are. To simplify analysis .1' = 0 xo (t) + x(t) == b(xo(t) == + x(t). Force Break )'1 In this section we use examples to illustrate how to develop transfer functions and statespace equations for physical systems.9 (a) Coulomb Mechanical system. the corresponding solution may differ ' from xo(t) only slightly.
fo r g(l .23) and (2. It consists of two blocks.2 ] [ Y ] == [ ] kl + k2 U2 Y2 and the convolution description of the system is y (I) == Next we develop a statespace equation to describe the system.r) . To develop its inputoutput description.11 [ .2r This is a standard equation in studying vibration and is said to be in normal form.24) )l(s) == m sz + kls + k2 This is the inputoutput description of the �ystem. The app lied force force u I must overcome the spring forces and the remainder i s used t o accelerate the block. Applying the Laplace transform and assuming zero initial conditions. [ m 2 s 2 + kl + kz des) kz des) 2 2 des) :== (m I s + kl + k2) (m2s + kl + k2) . Thus we have u must overcome the friction and the spring force. Let us define o O ml ] [ � I ] + [ kl + kl k2 Y2 uI k I . connected by three springs with spring constants k" i == I .kZY2 == U I For the second block. EXAMPLE 2.k2Yz (s) == Il l (s) mls2Y2 (S) . Let us select the displace ment and velocity of the block as state variables.k2 Y I (s) + (kl + k2 )Y2 (S) == 11 2 (S) From these two equations. 2.20 MATHEMATICAL DESCRIPTIONS OF SYSTEMS 2.r ) u ( r ) d r == fo r Then we can readily obtain (e(r .1 [ __ _ __ ] == e . ] m l s 2 . XI == y. we apply the Laplace transform to (2. � 0 m2 ml 0 0 0] [ ). 2.� ml � m' � 77777777777777777777 r YI r y.7 Consider the system shown in Fig.22) We apply Newton's law to develop an equation to describe the system. To simplify the discussion. . we can readily obtain This statespace equation describes the system. thus we have where y == d2 y (t)ldtZ and y == dy (t)ldt.r g(l) == r s +2 s+ I s 2 + 3s + 2 I _ e .k� This is the transfermatrix description of the system.2 (tr) ) u ( r ) dr n[ X2 X3 .11. 2. ) (k' ° k2 0 ml ml 0 0 0 . we have (2.k2 / m Xl ( t) k l /m I ] [ XI (I ) ] + [ 0 ] u ( t ) X 2 (t) 1 1m XJ C t l ] y(l) == [ I 0 ] [ XlU ) <1 This twoinput twooutput state equation describes the system in Fig. See Reference [18]..\'I (s) . that is. w ith masses m I and m2 . X4 _ XI :== Y I X2 : == Y I X3 :== Y2 x� : == Y2 They can be expressed in matrix form as [ XI (t) ] == [ 0 . kl == 2. Then we have. then the impulse response of the system is i I . The applied .5 Examples 21 block.v l (S) + (kl + kzJ. Figure 2. we obtain or m L Y I + (kl + kl)YI .24) and assume zero initial conditions to yield 0 U. The remainder is used to accelerate the (2. 1 m2 [ :. I ] == y : == X 0 0 1 )'2 [I o �WiH�' .e.22). Thus what we will discuss in this text can be applied directly to study vibration. we assume that there is no friction between the blocks and the floor. 3.(kl + kl) k. Xz == y.23) which implies (2.� I (S) ] == Y2 (S) where Springmass system. using (2. Its transfer function is I / (ms z + kls + kl) . If m == I . u (s) They can be combined as [ml 1 1 l [ ] == £.11. k l == 3.
(4 = Y = [I ° ° ° ° ° ° m /M (M + m)g/MI O]x ° � ° �l [. and the friction.25) (2.9 A communication satellite of mass m orbiting around the earth is shown in Fig.k/r z + ur/m o 2. input.l /MI My = u . For example. eo.25). through the hinge. This state equation has dimension 4 and describes the system when 8 and 0 are very small.r02 cos2 ¢ + r12 .(8) 2 cos 8 ] . it is reasonable to assume 8 and e to be small. 8 (t). and e or. For simplicity. Xz and X4 = e .� l [ 1 I X3 X4 + O I /M ° = .(t) 8 (t) O (t) ¢(t) 1(t) uU) = [Ur (t) ] uo (t) u</> (t) y (t) = [ ] r(t) 8 (t) ¢ (t) Then the system can be shown to be described by x = h(x.u (s) From these equations.u Figure 2.(M + m)g 1 . we now can develop the input{)utput and statespace descriptions.(3 .22 MATHEMATICAL DESCRIPTIONS OF SYSTEMS EXAMPLE 2. Because the design objective is to maintain the pendulum 1. and output of the system are chosen as [� l [� � .g gyu (s) = ::c:c::= S Z [Ms 2 .H dt2 dZ H = m . · 8 z cos ¢ sin ¢ .= u . we obtain Mis 28 (s) = (M + m)g8(s) .. the cart moves to the right and exerts a force. X} = 8. The altitude of the satellite is specified by r(t). .my .12 (2.8 r Ms Z y(s) = u (s) . Let H and V be.26) we can readily obtain The application of Newton's law to the rotational movement of the pendulum around the hinge yields mgl sin 8 = mle · 1 + myI cos 8 at the vertical position. we select state variables as Xl = y.2r ¢/r + u </> /mr (2. 5 Examples 23 Consider a cart with an inverted pendulum hinged on top of it as shown in Fig.v. ? • gou (s) = A  � Ms2 . 1 2 . and u</>(t). we obtain V = mg and They are nonlinear equations. we can readily compute the transfer function gyu (s) from u to y and the transfer function gou (s) from u to 8 as pendulum as shown.28) . 1 3 . and ¢ (t) as shown. 2 . the mass of the stick. u) = . to push the pendulum back to the vertical position.mle g8 = Ie + y which imply My = u .(/ cos 8) = ml [8 sin 8 dt Z ' To develop a statespace equation. (0) 2 .ml(8t sm 8 . Then from this selection. dZ = m .0/r + 201 sin ¢/ cos ¢ + uo/mr cos ¢ 1 . The application of Newton's law to the linear movements yields dZy M.mgB(s) A A 2 .27) . (t). the horizontal and vertical forces exerted by the cart on the Using these linearized equations.(M + m)g] A = my + mlB cos B . Applying the Laplace transform to (2. u (2. and the gust of wind are disregarded. uo(t). The orbit can be controlled by three orthogonal thrusts u. respectively. The problem is to maintain the pendulum at the vertical position. By retainin� only the linear terms in 8 we can use the approximation sin 8 = 8 and cos 8 = .26) x(t) = Cart with inverted pendulum. Under this assumption.25) and (2.mg8 Mle = (M + m)g8 . The state. EXAMPLE 2. This simple mechanism can be used as a model of a space vehicle on takeoff..(y mg V dtZ I + / si n B) s2 . the cart and the pendulum are assumed to move in only one plane. dropping the terms with e 2 . 2. and (2. and e 8 . r(t) . equivalently. (2. if the inverted pendulum is falling in the direction shown.26) and assuming zero initial conditions.
XI YI YI X I RI X2  X2 =  and Y X. and  y) dt Thus the sta tespace description of the system is given by (J)� . which corresponds to a circular equatorial orbit. A.2wo ro 0 0 0 0 0 m ro 0 0 0 0 0 0 0 0 0 0 0 0 I 0 2woro 0 0 0 0 0 0 0 0  0 0 0 0 x(t) 3w. They can also be controlled by the valves. A simplified model of a connection of two tanks is shown in Fig.x.10 In chemical plants. one involving r and e. In this equation. x(t) = 0 0 0 0 0 1 m XI Xi• .14 tanks. This sixdimensional state equation describes the system. we can readily ob tain 0 0 0 0 . it is often necessary to maintain the levels of liquids. S tudying these two parts independently can simplify iU' � k 0 0 0 0 0 0 0 0 0 0 :J = 'u ' 2 . Define == 0 normal operation. . From these equations. Once the satellite reaches the orbit.. thrusts must be applied to push it back to the orbit. I t is assumed that under analysis and design.24 MATH EMATICAL DESCRIPTIONS OF SYSTEMS EXAMPLE 2. i t will remain in the orbit as long as there are no dis turbances. Thus the equation can be decomposed into two uncoupled parts.. is given by Uo with r.w. We note that the three ma trices are all block diagonal..29) Figure 2. 0 Figure 2. the o ther </>. Changes of liquid levels are governed by Rj x(t) = xo(t) + x (t) u (t) = uo(t) + ii(t) y(t) = Yo + y(t) where AI dXI Aj = (II Yll dt If the perturbation is very small. a known physical constant. It is assumed that H2 .2 R = where are the flow resistances and depend on the normal height HI and H2.X2 Al AIRI Xl X. 1 . then they will be time varying. If the satellite deviates from the orbit.. =  A2 dX2 (YI X I .:� ] [ ������2 Y I/ R21x = = [0 Hydraulic Q + 11 .28) can be linearized as are the cross sections of the tanks. the inflows and outflows of both tanks all equal Q and their liquid levels equal H I and Let 1I be inflow perturbation of the first tank.13 Satellite in orbit.. [ . 2 . = k . If the orbit is an elliptic one.A2RI A2R2 = II .. and C happen to be constant... = . mro Q + )'1  Q+y . which will cause variations as shown. then (2. B. + 0 0 0 0 I ii(t) Al T HI + XI . These variations will cause level variation and in liquid level and outflow outflow variation y in the second tank.1 . 5 Examples 25 ( 2. One solution. ..
2 . Every tree branch has a unique cutset. removing all branches connected to a node leaves the node unconnected to the remaining network. From the figure. called the fulldamental loop. 1 RLC networks In RLC networks. state variables. For example. respectively. and current source will be considered as a branch. It is important to specify their polarities.I lL o o . a fundamental loop contains only one link and a fundamental cutset contains only one tree branch.I O]x +o·u . in which the remaining cutset branches are all links. 4. 3.5 Examples 27 + I IR C' 0 0 u Thus we have . we see that the voltage across the resistor is u . capacitors and inductors can store energy and are associated with state variables. The algebraic sum of all voltages along every loop is zero (Kirchhoff's voltage law).15 normal tree. . link. capacitors. at node B it yields This threedimensional state equation describes the network shown in Fig. 2 . The set of all branches connect to a node is called a cutset. Capacitor voltages in the links and inductor currents in the normal tree are not assigned. their state eqJations can be developed by applying Kirchhoff's current and voltage laws. resistors. If a capacitor voltage is assigned as a state variable x. == A I A 2 R I R2S . With respect to a chosen tree.I/CI I/C2 o ] [ ] x 2. if necessary. For most simple RLC networks. the inputs by applying Kirchhoff's voltage law to fundamental loops and Kirchhoff's current law to fundamental cutsets. The procedure and subsequent Example 2 . where L is its inductance.X2 or The output )' is given by � Procedure for developing statespace equations' 1.RC I I X3 C2 . Thus its current is (u . once state variables are assigned. and their currents or voltages should not be assigned as state variables. We consider only connected networks. The algebraic sum of all branch currents ill every cutset is zero (Kirchhoff's current law). 1 2. Every capacitor. A tree of a network is defined as any connection of branches connecting all the nodes but containing no loops. and current sources. inductor. First we introduce briefly the concepts of tree. Applying Kirchhoff's current law a t node A yields C2X2 == X3. called the fundamental cutset. Consider an RLC network. The reader may skip this procedure and go directly to Example 2. as the next example illustrates. A branch is called a tree branch if it is in the tree. Then their currents and voltage are. where C is its capacitance. then its current is C X.X I with the polarity shown. EXAMPLE 2. The procedure can be stated more systematically by using graph concepts. The procedure used i n the preceding example can be employed t o develop state equations to describe simple RLC networks.5.x I l I R. We assign the Cicapacitor voltages as Xi . 2. i == 1 . � I Its transfer function can be computed as g (s) . XI X2 == XI . Express the voltage and current of every branch in terms of the state variables and. however. in which the remaining loop branches are all tree branches. C I X I o C2X2. 2 and the inductor current as X3. A ·Ioop is a connection of branches starting from one point and coming back to the same point without passing any point twice. We first choose a Figure 2. Branches are connected at· nodes. inductors. y == [I . + (A I RI + A I R2 + A2 R2)s + I I 1 I [ They can be combined in matrix form as 2. 'resistor. a link if it is not. then its voltage is Lx . and LX3 with the polarities shown. The branches of the normal tree are chosen in the order of voltage sources. If an inductor current is assigned as a state variable x. as we will introduce next. voltage source. 1 5 . Assign the capacitor voltages in the normal tree and the inductor currents in the links as Network. a cutset of a connected network is any minimal set of branches so that the removal ohhe set causes the remaining network to be unconnected.X2 == X I . The procedure is fairly simple: assign state variables and then use branch characteristics and Kirchhoff's laws to develop state equations.26 MATHEMATICAL DESCRIPTIONS OF SYSTEMS r:. Note that resistors are memoryless elements. In other words. can be skipped without loss of continuity. X3 + X3 CI  RCI U == Appling Kirchhoff's voltage law to the righthandside loop yields LX3 == LXI .1 1 Consider the network shown in Fig. 1 5 . More generally.13. and cutset of a network. every link has a unique loop. Thus a network can be considered to consist of only branches and nodes.
28 MATHEMATICAL DESCRIPTIONS OF SYSTEMS 4.0000 This is the first column of the transfer matrix. 1 ) 1 0 1. b= [ 1 / 6 1!3 .755 + 0. This completes Stel? 3.: . [ 1 . The current through the 2H inductor is assigned as X3 and its voltage is 2X3. The normal tree is chosen as shown state variables.xI The I F capacitor is a tree branch.X 3 = 0 or .0833 3 5 + 0. it consists of the voltage source.0833 3 0. 1 6675 2 5 + 0.. 5 .5(u l y= They can be written in matrix form as . The voltage along its fundamental loop is 2X3 + X3 or i . W e type a= [ .Q resistor.5 Olu 0 0.5 O ] . 1 6. C 0. and the I . 1 6) : [ � 5 o o I ] [ x+ 0 O]x + [0.0000 0. 0 0 0 1. d= [ O 0 .2500 0.0. 66752 + 0. 1 66752 .0833 5 + 0. Because the 2Q resistor is a link. They can be expressed in matrix form as 2 .55 + 0. Thus the transfer matrix of the network is Figure 2.755 + 0. d . and from its fundamental cutset we have X2 The 2H inductor is a link.083 5 3 � 0.XI .xl )/2 I cutset / Fundamental / cutset / of @ which yields Nl= of Q) d1= 0. we use its fundamental loop to find its voltage as U I .3 . 2.0000 0.I lx U I . d1 ] = s s 2 t f ( a . b . 1 667 0. 5 I .755 + 0.X2 . 5 Examples 29 EXAMPLE 2.5 0 0 u ] (2.5 0] .! X3 + i U I + U2 � Equations (2.3 1 ) are the statespace description of the network.I .X I .30) and Y2 = 0.12 Consider the network shown in Fig. Thus its voltage is I .7500 0. The 3F capacitor i s a tree branch and its fundamental cutset i s a s shown.33335 2 1 .0833 0.666752 + 0.666752 + 0. c . Thus its current is (U I . X3 = X3. The I Q resistor is a tree branch.083 2 0. two capacitors.0 .255 + 0.08335 . 0 . GJ .1 .33335 3 + 0.0. (2.3333 0. .3 1 ) .XI W e will use MATLAB t o compute this equation. The voltage across the I F capacitor is assigned as X2 and its current i s X2.X3 = 0 which implies XI = . We repeat the computation for the second input. Fundamental Y2 (U I .3xI 2 o + I/  . � 2X3 = [ �l �l XI . then its current is 3.0.X 3 = [ I .5000 0. then we have Yl = . Apply Kirchhoff's voltage or current law to the fundamental loop or cutset of every branch that is assigned as a state variable. The transfer matrix of the network can be computed directly from the network or using the formula in (2.755 + 0.0000 0.xI )/2.0000 0.x t l = [ .0.0.16 G (5) = Network with two inputs.t l . 5 0. The capacitor voltages in the normal tree and the inductor current in the link will be assigned as If we consider the voltage across the 2H inductor and the current through the 2Q resistor as the outputs.30) and (2. S ] .666752 + 0. We use its fundamental cutset to find its current as X3 . The algebraic sum of the cutset currents is 0 or with heavy lines. If the voltage across the 3F capacitor is assigned as XI .6667 0.1 / 6 c= [ l + X2 =0 [ N : .
6 DiscreteTime Systems This section develops the discrete counterpart of continuoustime systems. t +c r L R.tt l .32) yields This equation is obtained by shifting the operating point from where E = E .l /R. io) and by linearization at (va' io) . Let be be the voltage across the capacitor and the current through the inductor. b) shown in Fig.m ] = E = R L (c) x. 2 . Because the two linearized equations are identical if is replaced by and E by we can readil y obtain its equivalent network shown in Fig. Because most concepts in continuoustime systems can be applied directly to the discretetime systems. XI XI XI R2 . Then it can be expressed as Inputoutput description Let o [k] be the impulse sequence defined as o [k . Then we have v = X and EXAMPLE 2. the network can be reduced to the one in Fig. the discretetime system is lumped. 2.1 Cx. In a discretetime system. C [ . 1 7(a). A discretetime system is linear if the additivity and homogeneity properties hold.Xl L u [m ] 8 [k . d ) shown in Fig. the discussion will be brief. Now if ( t ) is known t o lie only inside the range (c.32) 2. as the ones in Examples 2 . we may introduce the variables ( t ) = ( t ) . where T is a tunnel diode with the characteristic s shown in Fig. V � E = Y L x. k ::: ko. 0) RI E. where k is an integer ranging from . then the discretetime system is a lumped system.13 Xz (t ) = Liz(t) = They can be arranged as XI I CXI + C + E . It is the discrete counterpart of the impulse 8 (£ .17 1+ L {� if k = m if k f= m R . (a) x. L + i = h(v) . Let g [k . is a distributed system. L v Response = zerostate response (b) + zeroinput response and the zerostate responses satisfy the superposition property. Now if X I (t) is known to lie only inside the range (a. R =. l 7(b). denoting sampling instants. Let u [k] be any input sequence.X I (t) i l (t) 1 i i i I I 2 . m] be the output at time instant k excited by the impulse sequence applied at time instant m .va to (vo.(t)1 Substituting these into (2.I lL RIL Xz lL T RX' FtR LXl L X2 . (2. 1 7(d).3. l 7(b). Then we have . 1 7(b). where both k and m are integers.h (xI (t» L I L I This set of nonlinear equations describes the network.1:t2l ] [ I / C R2 IRIL ] [ X2I ] + [ 0 ] I / X =  XI lL I lL E (0. :xl u [k] = m = . andx2 (t) = x2 (t ) . 2 .m ] can easily be generated.t � x.m ] Network with a tunnel diode. XI (t) C +c RX2(t) + E X2(t)  Rz Rio. The entries of X are called state variables.30 MATH EMATICAL DESCRIPTIONS OF SYSTEMS Consider the network shown in Fig. A discretetime system is causal if current output depends on current and past inputs.oc to +x>.t i l has zero width and infinite height and cannot be generated in practice: whereas the impulse sequence a [k . The impulse o Ct . The state at time ko. 1 and 2. So do the zeroinput responses. denoted by x[ko]. If the number of state variables is finite. k ::: ko. The response of every linear discretetime system can be decomposed as XI RI · [ XI ] [ I /C RI I/ C ] [ XI ] + [ I O ] E Xz = . = . otherwise. y[k] := y(k T ) . if the time delay is an integer multiple of the sampling period T. 1 7(c) and can be described by xz (t) = .io and approximate h (xi (t» as io . it is distributed. 6 DiscreteTime Systems 31 This is a n LTI statespace equation. C 1'2 � (d) t' Figure 2. determines uniquely the output y[kJ.vo . h (x i (t» XI(t) . which together with u[k].RX2(t) (t) i (t) = Xz . 2 . Note that it is not obvious how to obtain the equivalent network from the original network without first developing the state equation. 2.'!:1 (t» = (t)1 In this case. The input and output of every discretetime system will be assumed to have the same sampling period T and will be denoted by u[k] := u(k T ) . is the information at time instant ko. then h (xi (t» can be approximated by h (. 2. Every continuoustime system involving time delay.
y[kl an If a linear discretetime system is time invariant as well. EXAMPLE 2.1 ] + a28[k . If the unitsamplingtime delay element is replaced by its transfer function ZI .37) y[k] excited by the input u [k] equals 00 y [k] = L g[k. mlu[m] m=ko It is a rational function of z .2 ] + The ztransform of ' " = We first replace the upper limit of the integration in (2. is pennitted under the causality assumption. for k<m (2.4).m .3) and its derivation is considerably simpler. 2.8) and is called a discrete convolution. then the block diagram becomes the one in Fig .35) to (2.33) can be reduced to k y [kl = L g [k.34) g[kl 8 [k . introduced the new variable I = = 0 for I < 0 t o make the inner summation independent of m .1 ] as in (2.Ill ] i s z . It is the discrete counterpart of Fig. Causal If a system is rela'led at <==> (2.9). Note that every continuoustime system involving time delay is a distributed system.36) to yield S·(z ) = = = 5 .18(a). m :::: 0 .m]u[m] + L g[k. 2.m ] + g[k.12). m]u[m] (additivity) Thus the output where we have interchanged the order of summations.m]u[m] + g[k. then the time shifting property holds. 2 . the initial time instant can always be chosen as = 0 and (2. m ] 8[k . mlu[m] m=:JQ g[k. The function g(z) is the ztransform of the impulse response sequence and is called the Both the discrete convolution and transfer function describe only zerostate responses. the impulse response sequence is z (2. The transfer function can also be obtained by applying the ztransform to the impulse response sequence of the feedback system.:) = Z[8[k  I I I = .32 MATHEMATICAL DESCRIPTIONS OF SYSTEMS 2.35) r g (z) =  important tool in the study of LTI discretetime systems.mlu[m] = L g[mlu[k . Its impulse response sequence is = and its discrete transfer function is g[k. and then used the fact that The equation g[l] k . Let defined as S ( z) aI I . The sequence m] is called the If a discretetime system is causal. Thus the transfer function of the feedback system is m]z.m ] ll [ m]  and then substitute it into L am 8[k . (2. For example.m.1 ] The output equals the input delayed b y one sampling period. ko and causal. 2 . . As in (2.6 DiscreteTime Systems 33 8[k .1O).:. . then (2. This is not so in discretetime systems.33) is the discrete counterpart of(2. This (f g[k ) z<km)zm m�O ) u[m]zm f (f g[k m�O k�O f (f g[llZ I) (m�O ll [m]zm) I �O f k�O . if for for k = 1 . Thus we have Consider the unitsamplingtime delay system defined by y[k] = u[k .14 This is the discrete counterpart of (2.5 gf[k] = a8[k .15 Consider the discretetime feedback system shown in Fig. no output will appear before an input is applied. EXAMPLE 2. ml = 0.1 = � The �transform is be the ztransform of ko k k y [k] = L g [k .18(b) and the transfer function from to y can be computed as g (. m]u[m] (homogeneity) L :)[k .<k:I) =: g(Z)u(Z) 1 which yields the same result . .acl =  z a a Y ez) := Z [y [k]l : = L y[k]zk k�O 00 This is a rational function of and is similar to (2.m l m=! 8[k . In this case.34) becomes (2. impulse response sequence. z.ml m=O m=O This is the discrete counterpart of (2.v (z ) = g(z ) u (z) g[k] discrete transfer function. .5(a).36) 00. of The discrete transfer functions in the preceding two examples are all rational functions This may not be so in general.
In ( l where A.5). improper transfer functions of s will amplify high t>�O . and D are constant matrices. however.zx[O] = Ax ( z ) + Bu(z) which implies Ye z) = Cx ( .42) v . C.x[O]) y[k + 1 ] = 0. However.4 1 ) reduces to y(z) = C(zI .0. (b) Statespace equations Every linear lumped discretetime system can be described by x [k + I] = A [k]x[k] + B[k]u[k] y[k] = C[k]x[k] + D[k]u[k] becomes (2.A).:.1 + !.37) yields G(z) = C ( z l . The reason. If we define the differenliation as • v(t) =  duCt) dt = hm where 6 > 0. it is open to argument whether an improper transfer function represents a noncausal system.A) . then (2. is different. then the two equations are identical.38) Then we have g(z) = . .: ) + Du(z) (2.I B + D (2. B.38) x [k + I ] = Ax[k] + Bu[k] _ .1 ) It is an irrational function of z. 1 4) and (2.:.39) yields [f 1=1 x[l]zI + x[O] .39) x(.40) z x(z) = (zl .5 Z[x[k + I ] ] = L x [k + I ]Zk = Z L x[k + I ] Z (k+ I ) y [k + I ] . then £1 (. . C. if we define the differentiation as t> � O .0.:.18 Discretetime feedback system. (t ) =  du(t) dt = hm  then the output yet) does not depend on future input and the differentiator is causal. 1 6) . If the system is time invariant as well. If the Laplace transform variable s is replaced by the ztransforrn variable z. However.:2 + .u (t) 6 They are the discrete counterparts of (2. Thus a discretetime system described by an improper transfer function is not causal. Thus all discrete rational transfer functions will be proper.:) = Z [x[k]] := Then we have k =O oc L x[k]Z k Hz ) which implies Z 2 + 2z . If x[O] = 0. Discrete rational transfer functions can be proper or improper.1 35 frequency noise.I Bu(z) y(t) = du(t)/dt.A).3 + .X[O] ] = z(x(z) .A ) .I ZX[O] + C( z l .A) I ZX[O] + (. u (t) . then (2. then the output yet) depends on future input u (t + 6) and the differentiator is not causal. If a transfer function is improper such as g (z) = ( Z2 + 2z .u [k] It means that the output at time instant k + I depends on the input at time instant k + 2. Consider g(s) = s or zx(z) . = . Therefore improper transfer functions are avoided in practice.34 MATHEMATICAL DESCRIPTIONS OF SYSTEMS Unittime delay y [k J 2 . Therefore in continuoustime systems.  u (t + 6) . (a) Figure 2. Such a system is a distributed system. a future input.43) This is the discrete counterpart of (2.1 ) / (. It is a pure differentiator.I Bu(z) + Du(z) y (z ) = [C( d . which often exists in the real world.:) where A.u [k] or =z Applying the ztransforrn to (2. and D are functions of k. B . We mentioned earlier that we also study only proper rational transfer functions of s in the continuoustime case.5y[k] + u [k + 2] + 2u [k + I] . Let x(z) be the ztransforrn of x[k] or oc y[k] = Cx[k] + Du[k] (2. . 15).u (t 6  6) Comparing this with the MIMO case of (2.A ) .I z  0.41) in front of (2. Note that there is an extra x[O].: . 6 DiscreteTime Systems _. We study only causal systems. We study in this text only lumped discretetime systems and their discrete transfer functions are all rational functions of z.:I .I B + D]u(z) ( 2.5y [k] = u [k + 2] + 2 u [k + I] .
By so doing. We study in this text mainly lumped linear timeinvariant systems.0001 5y[k] + u [k + I ] = 1 . .36 MATHEMATICAL DESCRIPTIONS OF SYSTEMS EXAMPLE 2. C. by definition. (2. If we withdraw money. then it is a linear system. ). then is negative. I . Thus we cannot select := y k as a state variable. Next we develop a statespace equation to describe the account. timeinvariant 5 x[k] Because of u [k + I].00015 = x 1 .00015l Distributed.1 l z .44) Lumped. or there is initially no money in the account. as summarized in the following. we have [ u [k] The discrete transfer function is the ztransform of the impulse response sequence or and the inputDutput description o f the account is Because the input ( I . all designs in the SISO case can be extended to the multivariable case. linear (2. Furthermore. J is an impulse sequence.1 .44) or (2.(5) rational yet) = 10' G(t . O.'� "� 0. If we deposit one dollar on the first day (that is. .00015y[k) + u[k + I ) If we define the state variable as x [k ] : = v[k].0001 5l We introduced i n this chapter the concepts of causality.00015 ( 1 . we use mostly the timedomain description (A. . Dl in the internal description timein variant y = Cx + Du Lumped. we will express every rational transfer matrix as a fraction of two polynomial matrices. Then we have g(z) = Z [ g [ k ) ) = y[ k k k] = L g[k . System type Internal description External description The l inearization discussed for the continuoustime case can also be applied to the discrete time case with only slight modification. linear. as we will develop in the text.m ]u[m] = L(l ·00015lmu[m] m =O g[k) = ( 1 . = Whenever we use (2. 0. �. 7 Concluding Remarks ( l . x = Ax + Bu f' G(t . Mathematical equations were then developed to describe causal systems.(S )UI5). and time invari ance.47) is not i n the standard form of (2.000 1 5 y [k ] ! = y I] · 1 u[O] = >"[1) = = I) and nothing thereafter I + = 0.00015u[k) x [k + I ] (2. in general. [ 1 . For this class of systems. r)u(r) dr G U .(s)u(s). The input u [k ] is the amount of money deposited into the account on the kth day and the output y k ] is the total amount of money in the account at the end of the kth day. lumpedness. (.46) yields . [ k) y [k] . We consider here only the LTI case with interest rate r = per day and compounded daily.r )u ( r ) dr 10 Y(S) = (. (. Suppose v [k ] is the total amount of money at the end of the kth day. systems are implicitly assumed to be relaxed or their initial conditions are assumed to be zero. . linear.7 Concluding Remarks 37 x := This is in the standard form and describes the money market account.000 1 5)kZk L( 1 . . (u[k] = k = I .01 5'!c i . The account is a timevarying system if the interest rate changes with time.u [k ) Substituting y [k + I ) = x[k + I ) + u [ k + I ] and y [k ) = x [k ) + u[ k ] into (2. Therefore its discussion will not be repeated.4 6 ) ( 2. Distributed systems cannot be described by finitedimensional statespace equations.00015.45). r ) u (r ) dr Distributed. If the i nterest rate is the same no matter how much money is i n the account. B.16 Consider a money market account i n a brokerage firm.OOO I 5x [k ] + 1 . thus whenever we use the description. then = u O) = and Because the muney is compounded daily.47 ) [) and the frequencydomain (Laplacedomain) description in the external description. If the i nterest rate depends on the amount of money in the account.00015 1 .(5) irrational y(t) = Y(5) = (.48) y [ k ) = x [k) + u[k ) 2. response sequence or and.r )u ( r ) d r G(s) . y[O] [ y[2 ) = O. Next we select a different state variable as ] 0. then x [k + I ] = 1 . External description describes only zerostate responses. the initial state must be zero.00015z.39). the impulse = y[ I J + y [ I ) · 0. linearity.00015)2 2 . linear y = C (t ) x x = A ( t )x + B(t)u y(f) = L ( 1 . the output is. .1 .000 1 5z1 (2. 4 ) + DCt)u y(t) = l' l' '" '" G(t. 2 . it is a nonlinear system. a timeinvariant system if the i nterest rate is fixed. 000 I Sx [k ] + u [k + I ] y [k ) = x [k J y [k + I ) = y [k + (2.
7 2.j. 2. g(t) Figure 2. and U3(t) for t :::: O.ll (e) 2.4 The input and output of an initially relaxed system can be denoted by Y = H1I.to) g (t) = 2w 2w (t . In each case. then Y3 = y.8 2. Show that g(t. + U2). This study will form a foundation for studying more general systems. If 113 = U. If u = U.2 The impulse response of an ideal lowpass filter is given by sin 2w (t . 2.1 Consider the memoryless systems with characteristics shown in Fig. r)/8x = 0. Which of them is a linear system? Is it possible to introduce a new output so that the system in Fig.Y2. Consider a twoinput and twooutput system described by Which of the following statements are correct if x(O) '" O ? .19 2. 2 (a) • t .. then Y3 = 0. U 2 (t).20 2. + U2 .r. 3.20(b). The corresponding outputs are denoted by y" Y2 .112. 39 2. l 9(b) is linear? y u Consider a system whose input and output are related by y y y yet) = t. 1 9.5(y. we are able to give a complete treatment of analyses and syntheses.:::=_ tI (b) r. f 2. For this small class of systems.. and a . 2.20(a). then Y3 = y.to) for all where w and are constants. { �2 (t)/u(t  I) if if u(t . Is it true Pa H u = H Pa u? 2. U3 = 0 .10 Consider a system described by What are the transfer function and the impulse response of the system? ji + 2y . 2. a.5 Consider a system with input and output y.38 MATHEMATICAL DESCRIPTIONS OF SYSTEMS The class of lumped linear timeinvariant systems constitutes only a very small part of nonlinear and linear systems. and Y3 . Three experiments are performed on the system using the inputs (t). Show that the impulse response of the system equals ji(t) dji(t)/dt. which chops off the input after time a . + }"2 . r) depends only on t . .12 Let be the unitstep response of a linear timeinvariant system.] 11(1) u(t) (a) Consider a system with impulse response as shown in Fig. to u and output y are related by a u(t) y(t) = (Pa u)(t) : = o for t � a for t > where is a fixed constant. .3 Consider a system whose input a u. Is the ideal lowpass filter causal? Is it possible to build the filter in the real world? t. then additivity implies homogeneity. (b) Figure 2. r + a ) for all t. Is the system linear? Is it timeinvariant? Is it causa]? 2. where H is some mathematical operator.r and show that 8g(t.11 2. the initial state x(O) at time t = 0 is the same. The system is called a truncation operator.I ) = 0 for all Show that the system satisfies the homogeneity property but not the additivity property.9 Show that if the additivity property holds. in which denotes the input and the output. 5 (u . r. ". What is the zerostate response excited by the input shown in Fig. Let g(t. [Hint: Define x = t + r and Y = t .I ) '" 0 u(t . If 3 Which are correct if x (O) = O? 2. + Y2 ). Show that if the system is causal.3.3y = U u 2. u { a . r) = g et + a. then the homogeneity property holds for all rational numbers Thus if a system has some "continuity" property. then PaY = Pa Hu = Pa HPau where Pa is the truncation operator defined in Problem 2.6 Problems 1. 2.
14.2 1 (a) for a = I and i n Fig.5. I the moment of inertia about the center of gravity P. The loading problem is an XI = y.bs + kllil  shown in Fig. is Find state equations to describe the pendulum systems in Fig.25. k . elevator angle lIQ. ] important issue in developing mathematical equations to describe composite systems.mg. 2. X3 = m.16 Consider the simplified model of an aircraft shown in Fig. altitude h o . l. Let m be the mass of the aircraft.40 MATHEMATICAL DESCRIPTIONS OF SYSTEMS 2. Find a state equation to describe the system.15 2. and e2 are very small.or twolink robotic manipulators. and II = m . 5 .21 1 r�4�+�L+��� 4 (e) 6 . 2 . by neglecting the effect of I.23 2. where g is the gravity constant on the lunar surface.5 Figure 2. What is the transfer matrix of the Problems 41 y (l) o o 0 = 1 1 2 3 4 (a) a = 0.23. Define state variables of the system as Find the transfer functions from II to YI and from )'1 to y of the hydraulic tank system because of the loading problem in the two tanks in Fig. Show also that the unitstep responses of the negativefeedback system are as shown in Figs. If e. 2 1 (d). 2 .17 The soft landing phase of a lunar module descending on the moon can be modeled as !i rs ) ms2 (bs g(s ) = A  h(s) = el. 2. . X2 = 5'.14 Draw an opamp circuit diagram for Figure 2.5. the mass of the module. 2. where m is k . := where Nij and D' i are polynomials of P system0 D 2 1 (p)YI (t l + D22 ( P)Y2 (t) = N2 1 (P)II I (t) + Nn (p)u2 (t) d/dt.22.18 e transfer functions') Is this also true for the system shown in Fig. The thrust generated is assumed to be proportional to In . 2 1 (c) and 2 .J c::= 0 1 2 3 4 5 6 (d) 7 • t Figure 2.24. It is assumed that the aircraft is in an equilibrium state at the pitched angle eo. 147 [Answer: No.22 (a ) Ibl 6 • r 4 (b) I 6 • r 2. See Reference [7] . be the aerodynamic damping.k. The systems are useful to model one. Does the transfer function from II to y equal the product of the two . Show also that the transfer function from It to h. It is assumed that small deviations of e and It from e o and 11 0 generate forces /1 = k l and h = k2L1 as shown in the figure. 2. shown in Fig. 2 1 (b) for a = 0 . Show that the unitstep responses of the positivefeedback system are as shown in Fig. respectively. 2. and h the deviation of the altitude from ho . 2 . can you consider the two systems as linear" y = [3 l O]x  211 2. 2 . Then the system can be described by lIl . and cruising speed Va . 2. for a = I and a = 0.13 Consider the feedback systems shown i n Fig.i' = ktiz .:�. 2. Find a statespace equation to describe the system.
An external force u is applied to the bar as shown. 2.... Find a statespace equation to describe the system. It is assumed that the angular displacement () is very small.20 Find a state equation to describe the network shown in Fig. Q+y 2..2.25 .26 (current source) u 2.42 MATHEMATICAL DESCRIPTIONS OF SYSTEMS Figure 2.26. from equilibrium.27 +" �==4====:::==l .. Az T . Let y be the displacement .27. 2. Compute also its transfer matrix. 1 F Figure 2. with mass m 2 .19 Find a state equation to describe the network shown in Fig. 2...12 � "\ Lunar surface Figure 2..J_ 1�. Find also its trans· fer function. Find also the transfer function from u to y...21 Consider the mechanical system shown in Fig. T . Let I denote the moment of inertia of the bar and block about the hinge..24 t I Thrust = km I • I I ! t Problems 43 of the block.  . 2. Figure 2.
The set of vectors (X I . x.. As we saw i n the preceding chapter. n x m. am l BD = 44 [l l [l [ bm :: = a l b l + a 2 b2 + · · · + am b m (3. Representation. unless stated otherwise. . . q 2 . X I = . Representation . Then ( 3 . . it is the product of an n x I colunm vector and a I x r row vector. and AB = [a l a 2 . . q. . 2 Basis. 3) b. throughout this text.5) can be written as Define the n x n square matrix (3 5 ) qn I where fJi = a 1 / a I . for which (3. 0' 1 . . l x n.4) holds is a I = O. C. and r x p real matrices... Let A. and only those that will be used subsequently are introduced. Then every vector x can be expressed uniquely as Then (3. Let { q I . The topics are carefully selected. respectively. not all zero. denoted by 'R n . . no formal proofs are given. be the ith column of and bj the jth row of B.. is not defined unless n = r: it becomes a scalar if n = r . . where the. B. 6) . is an n x r matrix. The dimension of a linear space can be defined as the maximum number of linearly independent vectors in the space. In 'R " . any set of n linearly independent vectors can be used as a basis.. They are stated as theorems for easy reference in later chapters. Most results are developed intuitively in order for the reader to better grasp the ideas. am = O. . The product b i a. and D be. Every vector in of real numbers such as 'R " is an ntuple 3. say. we write it as x = [X I Xl . . and Orthonormal ization 45 3 . 3. such that x= [. 0'1 = O . 4) implies (3 '+) To save space. xm ) in 'R" is said to be linearly dependent if there exist real numbers 0' 1 . Thus in 'R". 1 ) (3 . Therefore we deal only with real numbers. we can find at most n linearly i ndependent vectors. . that is different from zero.2 Basis. . However.prime denotes the transpose. Let a. m x r. 1 I ntroduction This chapter reviews a number of concepts and results in linear algebra that are essential in the study of this text. h. Then we have I f the only set of a. . Xl . a n d Orthonormalization Li n e a r Alg e b ra Consider an ndimensional real linear space. . Such an expression is called a l i near combination. ]. then there exists at least one ai .[a 2 x2 + a3x3 + 0' 1 I + C{ m xm ] bl bo  bn � D= bID boD  (3 . Xl. 2) Basis and representation A set of linearly independent vectors in 'R" is called a basis if every vector i n 'R n can be expressed as a unique linear combination of the set. D � Q : = [q l ql . If the set of vectors in (3 . Note that a. . 0'2 . xm } is said to be linearly independent.I'l l x: _\�I! A. . then the set of vectors {X I . .Chapter / These identities can easily be verified. am. . " (3 . } be such a set. . 4) is l inearly dependent. all parameters that arise in the real world are real numbers.
· . and infinitenorm. i n f ) . q2 . . I l ax l l = l a l l l x l l .8 ) 0 0 With respect to this basis. Norms of vectors 2. Any realvalued function of x. I norm.1 3 . Consider the vector x = [1 3]' in X as shown in Fig. If we draw from x two lines in parallel with q2 and ql . I i=l EXAMPLE 3. we can obtain an orthonormal set using the procedure that follows: . . Representation. an ] ' the representation of the vector x with respect to the basis ( q l .l = 0 0 il = I h = 0 0 0 in = 0 0 (3. I 3. They are called. i = I . 2. 2 . 1. 3. A set of vectors Xi . is said to be orthonormal if x. O. Note that x'x is scalar and xx' is n x n . We will associate with every xn the following orthonormal basis: I 0 0 0 in . . . ( x . they intersect at q l and 2q2 as shown. the norms just introduced can be obtained by using the functions norm ( x . Xj = { O if i f. In other words. . I l x l + x2 1 1 CS I I xl l l + I I x2 1 1 for every XI and x2. The two vectors q l = [3 I ]' and q2 = [2 2]' are clearly linearly independent and can be used as a basis. 2 Basis. and n o rm l x . 2 ) �norm ( x ) . I l x l l ?: 0 for every x and I l x l l = 0 if and only if x = Orthonormalization A vector x is said to be normalized if its Euclidean nonn is I orx'x = I . we draw from x two lines in parallel with iz and q2 . Figure 3.7) I II vector x. This can also be verified from Let x = [x t X2 xn ] ' . 1 . We use exclusively. Two vectors X t and Xl are said to be orthogonal if x . i2). . the representation of any vector x with respect to the orthonormal basis in (3. Xl = XSX I = O. we have 0 0 I 0 0 The last inequality is called the triangular inequality. The 2norm is the length of the vector from the origin. q2} is [. Then the norm of x can be chosen as any one of the following: I l x l l l := I )�.) The concept of norm is a generalization of length or magnitude. and Orthonormalization Different representations of x 47 We call x = [a l a2 .Sq2 and 2h .8) equals itself. (Verify. norrr. .I 2]'. . .1 2 I lx l l z : = v£' = I l x l l "" : = maXi I xd (�IXd2) I/Z To find the representation of x with respect to the basis [q2. unless stated otherwise. respectively. . . iz) is [0. 1 ) . can be defined as a norm if it has the following properties: In MATLAB . the Euclidean norm and the subscript 2 will be dropped .or Euclidean norm. el. for any real a. denoted by I l x l l . where In is the n x n unit matrix. Thus the representation of x with respect to {ql . Thus the representation of x with respect to [q2. . qn ) . j I if i = j Given a set of linearly independent vectors e l .5 2]'. em . m.46 L I N EAR ALGEBRA (3. They intersect at 0.
1 2 ) In MATLAB . = O. It is then normalized to 1 as shown in Fig. 0 . If all columns of A or {a.. The rank of A is Consider the set of linear algebraic equations Ax = Y u �_ e2 u.:i (qi em )qk qm : = um i l i um i l defined as the dimension of the range space or. 3 Linear Algebraic Equations 1 ' [0. we can obtain an orthononnal set. the range space.  (3.L. . n) _ _ _ _ _ _ _ _ _ _ _ _ _ ranK l a ) 2 . We discuss the existence condition and general form of solutions of (3. we type a= [ 0 1 1 rank(A) :s: min Im.2 0 2 0] .9) (3 .rank (A) EXAMPLE 3. The set {a] . am l be an n x m matrix with m :s: n . m x n and m x I real matrices and x is an n x I vector. The svd algorithm also yields the range and null spaces of the matrix. which will be introduced later. 0.2 Consider the matrix A 'A � where 1m is the unit matrix of order m. a2 1 can be used as a basis of the range space of A . or 2a2 . Because of this fact. nu l l . This is obtained using �l. Equation (3. .a. they form a basis of the null space. Because the two vectors are linearly independent. 1. 3782 0 8877  0.3084 0 . See Problem 3.9). Let A = [a] a2 .\TLAB Version 5. 0 1 46 8 (3. The last column is twice the second column. 9 3 99 . 1 2 3 �. i = I . . " ' where a. Earlier versions may yield different results. 1 1) where A and y are. The matrices A and y are given and x is the unknown to be solved. . then em . the number oflinearly independent columns in A. 1 0) implies that the nullity o f A is 2 . and rank:. The nullity is defined as the maximum number of linearly independent null vectors of A and is related to the rank by 3 . This is called the Schmidt orthonormalization procedure.2.48 LINEAR ALGEBRA u ] := e] Uz : = e2 . It can readily b e verified that the two vectors meet the condition An. The range space of A is defined as all possible linear combinations of all columns of A. The third column is the sum of the first two columns or a] + az . denotes the ith column of A. 1 1 ). equal to. 1 3) Figure 3. Thus A has two linearly independent columns and has rank 2. Clearly a l and az are linearly independent. Its subtraction from ez yields the vertical part Uz. 3 . m I are orthonormal. . Using this procedure. . which yields 2 . It also equals the number of linearly independent rows. Note that. = O. and rank can be obtained by calling the functions (3. . null space. The MATLAB function R = o r t h ( a ) yields' Ans K= 0.2 Schmidt orthonormization procedure . The vector ( q.a3 = O. in general. For example. Note that MATLAB computes ranks by using singularvalue decomposition (svd). A vector x is called a null vector of A if Ax = O. respectively. [ am a o.l (3. 2 6 2 7 . . 1 0) (3 . . The null space of A consists of all its null vectors.( q'] e2)q] um : = ql : = u l / l l u l l l q2 : = u2 / 1 1 u2 1 1 The first equation normalizes the vector el to have norm 1 . 3 Linear Algebraic Equations 49 Nullity (A) = number of columns of A . e2)q l i s the projection of the vector e2 along q ] . Thus the set actually consists of m equations and n unknowns. equivalently. or smaller than the number of unknowns. 2 . if A is m x n. AA' '" I n . The number of equations can be larger than. for the matrix in A= [� � i :J =: [a l a2 a3 a. 3. then or ::h.4. The rank of A is defined as the number of linearly independent columns.
rank ( [ n l n2 N J ) yields 3 if we use the fivedigit display in ( 3 . .2475 0. if we use the fivedigit display of R in ( 3 . 1 6) (3. let xp be a solution ofxA = If k = O. The null space of (3. 1 2). Theorem 3. In application. . let xp be a solution of Ax = y and let k := n . for every O! i . 15) into Ax = y yields i=l Axp + L O! i Ani = Axp + 0 = Y k A(x . That is. for any y. 1 5). 1 5 ) I. (3. 12). If A has full row rank.2 (Parameterization of all solutions) Given an m rank. . 1 6). Given an m x n matrix A. xp + Cl l n l + . With this background. . which yields the unity matrix of order 2. . The rank is 2 if we use the N stored in MATLAB. the solution xp is unique. x n matrix A and an m x I vector y. Subtracting this from Ax = y yields p which implies that x . Thus the general solution of (3. . a2} span the same space as the two vectors of R. This confirms that [ a I . k . n2} i n (3.2. .8384 0. . Given an m This y clearly lies in the range space of A and xp = [0 . + O!knk y and let k = m . Given A. x = 51 (3. Note the use of backslash. If k > O. the rank of [ a l a2 R J is 3. then for any O!" i = 1 .50 L I N EAR ALGE B RA The two columns of R form an orthonormal basis of the range space. where niA = 0 and the set {n l . We use p to denote the rank of a matrix. This establishes Theorem 3. which uses 16 digits plus exponent. then the solution xp is unique. . .2 1. . which denotes matrix left division. All discussion for the range space applies here. 3 Linear Algebraic Equations = y . This can be verified by typing rank ( [ a l a2 I I I : I ! 3 . 0. a solution x exists in xA = y. . Given an m x n matrix A and an I x n vectory. If k > O. For example. Thus x can be expressed as in (3. . 1 1 ) can be obtaI ned by typing nu l l ( a ) . 1 and 3. The two columns in R are not obtained from the basis [al . x n matrix A and an m x I vector y. p e A ) = p ( [A y]) = y if and only > where 2. 14) EXAMPLE 3. We mention that the rank of a matrix can be very sensitive to roundoff errors and imprecise data. 1 4). 1 1. the solution of Ax = y can be obtained by typing A \y.4 0 0]' is a solution. which yields Ans N� Rl ) Substituting (3. the two bases should span the same range space. where the m x n matrix A and the I x n vector y are given. the vector is a solution of xA = y. we are ready to discuss solutions of (3.xp ) = 0 Thus.3434 0. Theorem 3. nk ! is linearly independent. 1 1 ) by using the Schmidt orthonormalization procedure. and the I x m vector x is to be solved. a2 } in (3. For example. we type R ' * R. .xp is in the null space. 13). 2 . 2. [A y] is an m x (n + I) matrix with y appended to A as an additional column. If A has rank n (full column rank) o r k = O. for the equation in (3. . where (nl ' . 3434 0. A basis of the null space of A was shown in (3. . then for every real Cl i . i = 1 . if and only if A has rank m (full row rank). . nk ! i s a basis o f the null space o f A. t h e vector x = xp + O! l nl + .5802 0. However.p eA).3 Consider the equation 0 I 2 I I Ax = :> The two columns are an orthonormal basis of the null space spanned by the two vectors [nl . 1 6) can be expressed as [ 2 0 3 2 (3.5802 0. The rank is 2 if we use the R stored in MATLAB. In MATLAB. . Applying Theorems 3 . Corollary 3. To check the orthonor mality. . This establishes the second statement. 1 5) is a solution. if and only if A has full column > The first statement follows directly from the definition of the range space. equivalently. then the rank condition in ( I ) is always satisfied for every y.9). we can readily obtain the following result. 1 7) for any real O! I and 0!2. 2. typing . + Clknk i s a solution o f Ax which yields 2. an n x I solution x exists in Ax if y lies in the range space of A or.3395 0. Let x be a solution or Ax = y. k .4598 (3.2 to the transpose of the equation.p (A) b e the nullity of A. they are a byproduct of svd. a solution x exists in Ax = y for every y. we will also encounter x A = y.
8). then there is at least one nonzero minor of order r.I )'+ i det Mij . y= [ .19) Because the three vectors b. The inverse of a nonsingular square matrix A = [a'i ) is denoted by A . The determinant of any r x r submatrix of A is called a minor of order r. q2 • . The homogeneous equation Ax = 53 0 has nonzero solutions if and only if A is singular. Then w e have Ab A' b = A(Abl = where cii is the cofactor.20) (3.3 . It turns out that the i th column of A is the representation {q! . The number c'i is the cofactor corresponding to a'J and equals 1 ) x (n 1 ) submatrix of A by deleting its i th row ( . . L a'J c'i ( 3 . The MATLAB function :' :s computes the inverse of A. 3. O ] . Now if w� select a different set of basis {ql . A square matrix is said to be nonsingular if its detenninant is nonzero. Thus the representation of A with respect to the basis {b. then the equation has a unique solution for every y and L solUlion equals A . For n = 2 . . the detenninant of n x n square matrix A = [aii ) is defined recursively as. where Mij is the (n and jth column. Here we use slash. . = .= [c'i) det A det A  A It can be verified that the following relation holds: {f l [] (3.4 a 0 1 ' . In general. Ab. and A2 b are linearly independent. If a matrix is singular. . Determinant and inverse of square matrices The rank of a matrix is defined as the number of linearly independent columns or rows. If we associate with 'R n the orthonormal basis {i l . This is illustrated by the example that follows.4 Similarity Transformation Consider an n x n matrix A.I = A .I y. 1 8) o Aqi with respect to the basis f spect to the orthonormal basis. 1 2 a \y 3 4. 3 . . . �e In particular. The inverse has the property AA . If A is nonsingular. It maps 'Rn into itself.I A = I and can be computed as  EXAMPLE 3. h . . . The detenninant of a I x I matrix is defined as itself. . It can also be defined using the detenninant. then the ith column of A is the representation of Aii with re det A = I . . We now compute the representation of A with respect to the basis.3 Consider Ax A(Ab) = [ b Ab A2 b) A (A 2 b) = [b Ab A2 b) where the last equation is obtained from ( 3 . rI I yields [ 0 . Equation (3 . Ab.52 L I N EA R ALGEBRA a= [ O 1 1 2 . .l entries and changing the sign of offdiagonal entries (without changing position) and dividing the resulting matrix tw the detenninant of A.1 := [ a2 l al l Thus the inverse of a 2 x 2 matrix is very simple: interchanging diagoDJ. Then the rank can be defined as the largest order of all nonzero minors of A. 1 8) is called the Laplace expansion. ! 3 . Theorem 3. . they can be used as a basis. The solution of xA = y can be obtained by typing y / A. If we have A _I Adj A I . It is clear that A(b) = [b Ab A" b) A.4 Consider the matrix A (3. 1 9) to compute the inverse is compli(uted. in } in (3. . :\l'te that the inverse of a triangular matrix is again triangular. 4 S i m ilarity Transiormation 2. where aij denotes the entry at the i th row and jth column of A. which denotes matrix right division. Thus a nonsingular square matrix has full rank and all its columns (rows) are linearly independent. the only solution of Ax = y with A square. . then det A equals the product of all diagonal entries. A"b} is (3 23) [:J [ :] [ �n  1. = 0 is x = O. qn J. it is simpler to compute its inverse by solving AA I = I. In other words. . If A is triangular. using ( 3 . . for any chosen j.I . and every minor of order larger than r is zero. The number of linearly independent solutions equals the nullity of A.4 .21) Let b = [ 0 0 I } ' .22). then ( 3 . if has rank r.22) is :2 x 2. then the matrix A has a different representation A. If A is diagonal or triangular. its inverse does not exist.2 0 2 O ] . q2 • . qn}.
29) as AQ = QA or t. . we introduce a set of basis so that the representation will be diagonal or block diagonal.27) o o 1 0' ) a n n x n nonsingular matrix. The square matrix A maps x i n 'Rn into y in 'Rn .28) Comparing this with (3. The matrix in (3. . (A ) = A 4 + 3 2 0' I A + 0'2A + 0'3)' + 0' 4 A [ql q2 . . . An . .0' 2 (3. . . . We write (3. o 0 o o 3 .25) yields (3. . If there exists an n x 1 vector b such that the n vectors b. Any nonzero vector x satisfyin g Ax = AX is called a (right) eigenvector of A associated with eigenvalue A. Let A be an n x n matrix. (A) = det(U  A) x = Qx with y = Qy It is a monic polynomial of degree n with real coefficients and is called the characteristic polynomial of A . With respect to the basis {ql . the n x n matrix A has n eigenvalues (not necessarily all distinct). In this section.I AQ or A = QAQI (3.7).AI) is 0 and (3. qn l A These matrices can easily be formed from the coefficients of t. . qn } . . . . qn l . 5 Diagonal Form and Jordan Form A square matrix A has different representations with respect to different sets of basis.26) yields and their transposes o o A = Q. . A polynomial is called monic if its leading coefficient is I . A real or complex number A is called an eigenvalue of the n x n real matrix A if there exists a nonzero vector x such that Ax = AX . qn l = [Aq l Aq2 . Ab. This is a homogeneous equation. qn } . the matrix (A . 3 .24) is in such a form. q2 . .26) where x and y are the representations of x and y with respect to the basis { ql . q2 . then the only solution of (3.30) has at least one nonzero solution. (A) has degree n . 5 Diagonal Form and Jordan Form 55 then the representation of A with respect to the basis {b.30) Ax = y (3. .I b are linearly independent and if This shows that the ith column of A is indeed the representation of Aqi with respect to the basis {ql . . . Substituting these into (3. they are related by t. A s discussed i n (3.AI) must be singUlar or have detenrunant We define (3. then the detenrunant of (A . the equation becomes (3 .U)x = 0 (3. . q2 . we write Ax = AX = Alx as where I is the unit matrix of order n. The companionform matrices will arise repeatedly later. In order to find the eigenvalue of A. Because t. If . (A) and are called companion form matrices.3). Thus in order for (3. .24) This matrix is said to be in a companion form. (A) is an eigenvalue of A . .54 L I N EA R ALGE B RA The preceding discussion can be extended to the general case. qn } . Consider the equation o p o !O o !3n1 !3n (A . Thus every root of t. Ab. .25) O. is a root of the characteristic polynomial.AI) is nonsingular. Aqn l = [ql q2 . If the matrix (A . .29) o o all have the following characteristic polynomial: 1 TJ This is called the similarity transformation and A and A are said to be similar. We mention that the matrices Q = [ql q2 .30) is x = 0 (The orem 3. .30) to have a nonzero solution x. ..
.707 0. qn } . be an eigenvector of A associated with A i . } is linearly independent and can be used as � basis. the inverse of Q. . . The matrix q is different from the Q in (3.34). Thus the representation of A with respect to {q I . qn } .33) by checking Q = AQ. :' 1 0 A2 0 r)�1 1 0 o Thus q l = [0 I I ]' is an eigenvector associated with A = 2.2 I )q = Aq l = A l q l = [q l q2 . then we can verify (3. [0 A l 0 . it is simpler to verify Q A (3. . q2 .I ]'. .8 1 86 0.34) with Q = [q l q 2 .I is any nonzero solution of (A �'" � ] l[ o 2 3. i = I . Let q.32) o 2 However. that is. 0 � ] [ � �I �] [ � � �] [� AQ or = o 2 0 0 0 0 I I I 1 1] .33) o 2 I The result in this example can easily be obtained using MATLAB. . Thus we conclude that every matrix with distinct eigenvalues has a diagonal �atrix representation by using its eigenvectors as a basis. 0]'.35). Similarly.707 1 � o 0. Computing (3. q3} is = 0 0 A3 (3.2 I ]'. . qn ] then the matrix A equals (3. 2 . we can establish . Note that the eigenvector is not unique.4082 0. . This is due to nonuniqueness of eigenvectors and every column of q is normalized to have norm I in MATLAB.5 Diagonal Form and Jordan Form 57 ql I = 0  (. Then the set of eigenvectors {q I . d] =eig ( a ) d= I . = Ai q. . [ a= [ O 0 0 . 1 0 2 .4472 0. q.I . . the eigenvector associated with A = 0 can be which yields q2 computed as q3 = [2 I .33) by hand is not simple because of the need to compute . Then the first column of A is the representation of Aq l = Al q l with respect to { q l .I )I)q2 = [0 . Typing EXAMPLE 3. . [0 a a]' for any nonzero real a can also be chosen as an eigenvector. . A (3. Let A be the representation of A with respect to this basis.8944 0. . . and O.I ) .1 . . then MATLAB generates only the three eigenvalues 2 . q2 ..2] = (A . q2 . A= This is a diagonal matrix.5 Consider the matrix yields Its characteristic polynomial is t. . The eignevector associated with A = 2 is any nonzero solution of (A . qn ] we conclude that the first column of A /is [AI 0 . From . if we know A. n.4082 ] [� � n [q. Different orderings of eigenvectors will yield different diagonal matrices for the same A . that is. . .35) as derived in (3.29). . Aq. The second column of A is the representation of Aq2 = A2q 2 with respect to {ql .(A) = det(AI = A[A(A . Proceeding forward. This matrix can also be obtained by computing [ [i � 2 �I 0 (3. q2 . OJ'.3 1 ) A= 0 0 It is a diagonal matrix with eigenvalues on the diagonal. O . If we define l: . 0. be the eigenvalues of A and be all distinct. but their corresponding columns differ only by a constant. The eigenvector associated with A = . However. Thus A has eigenvalues 2. . If we type e i g ( a ) without the lefthandside argument. . .2)(A + I )A I  A) = det [� o l A o  AI '" ] q = where d is the diagonal matrix in ( 3.56 LINEAR ALGEBRA Eigenvalues of A are all distinct Let Ai .
5 164j o All discussion in the preceding example applies here. To simplify the discussion. To see the reason. Even though the data we encounter in practice are all real numbers.3j is [ .3 2 An eigenvalue with multiplicity 2 o r higher i s called a eigenvalue. By so doing.AI) q 0 1 simplAe = of grade 0 if If we restrict v to real vectors. 1 .36) can be transformed into a very useful real matrix as we will discuss in Section 4. we consider complex linear spaces and complex scalars and = [ I 1 j 1. equivalently. respectively.2j . we must extend real linear spaces into complex linear spaces and permit all scalars such as Cti in (3. Thus we have Q= The MATLAB function [ q .5 1 64j 0. Note that complex conjugate eigenvalues must appear in pairs because A has only real coefficients. We need n 3 more linearly independent vectors to form a basis for 'R4 The three vectors 3 0. These vectors.6 Consider the matrix not r 3 . then it may not have a diagonal form representation. especially when there are repeated roots. Thus A has only one eigenvector associated with A.0. To deal with this problem.37) has no nonzero solution and the two columns of A are linearly independent.58 LIN EAR ALGE B RA We mention that eigenvalues in MATLAB are computed from the characteristic polynomial.4) to assume complex numbers. The eigenvectors associated with and 2+3j are. the diagonal matrix with complex eigenvalues in (3. Thus A has eigenvalues . [1 0 0]' and [j . From these equations. typing r = e i g ( a ) . it has a block diagonal and triangularform representation as we will discuss next. if v is permitted to assume complex numbers. q3 .j]' is a nonzero solution of (3. the complex conjugate of the eigenvector associated with A 2 + 3 j. the characteristic polynomial equals n (A .AI)V3 (A .AI) q3 . we consider Complex eigenvalues d = [ 1 � Av = 0.37) They are called a chain of generalized eigenvectors of length n 4 and have the properties 3 2 (A .2582j 0 0. = (A = generalized eigenvector n = (A .AI)"v (A .3 +2j jr.36) has only one independent solution.AI) V2 O. then (3.37). However. AV I V I + AV2 V2 + AV3 V3 + AV4 .AI)2v 3 AI) v  and v is an ordinary eigenvector. Suppose the matrix 3 or. Incidentally.3 . (A . Eigenvalues are computed in MATLAB directly from the matrix by using similarity transformations. If has repeated eigenvalues. Consider an n X n matrix A with eigenvalue A and multiplicity n.AI) n .AI)v (A . d ] = e i g ( a ) yields j q [ 1 j o 0 and = [1 3 + 2j j 3 2j 1== 7 = [�1 � � ] .l v � and v � 0 . we can readily obtain AVI AV2 AV3 AV4 = V2 VI = = ::== (A := 0 (A . as generated.7746 + 0. Computing the characteristic polynomial using the Laplace expansion and then computing its roots are not numerically reliable. A vector v is called a 1 = n = and If they reduce to (A . complex numbers may arise when we compute eigenvalues and eigenvectors.Ai ) . then v [2 1 . we assume n = 4. Therefore.AI)V2 = = = = == ((AA = (A = = .2582j 0. nullity 1 . then the equation (A . 1 +j 2 ] v = 0 (3. whenever complex eigenvalues arise. ] and A 0. are automatically linearly independent and can be used as a basis. If A has only simple eigenvalues. Once all eigenvalues are computed. 6 Functions of a Square Matrix 59 A Its characteristic polynomial is (A + 1) (A 14A+ 1 3 ) . For n 0 = 4.AI) V3 0. an eigenvalue with multiplicity is called a eigenvalue.AI)V4 = (A . q4 will be chosen to have the properties (A . and AI) 4 V4 0. Thus the two columns of A are linearly dependent and A has rank This is the rank obtained in MATLAB. it always has a diagonalform representation. poly ( r ) yields the characteristic polynomial. and 0.AI) has rank .j .Alfq2 q2 .AI)v we define V3 n = 1.AI)vl 0.AI) 4 q4 = O.1 Eigenvalues o f A are not all distinct repeated transpose is replaced by complexconjugate transpose. (A . In contrast.3 . In MATLAB.j]'.2582j o 2 + 3j o ] (3.1 . In other words.7746 . However. The eigenvector associated with A 2 . A has only one distinct eigenvalue. EXAMPLE 3. (A . all concepts and results developed for real vectors and matrices can be applied to complex vectors and matrices.2582j � = [ 1 o 2 + 3j o 0. 2 ± 3j .
1 = det A The determinant of A is the product of all diagonal entries or. then the set of vectors { V I . If ( A . V3 . j orca:". we may use the MATLAB function [ q . Polynomials of a square matrix 0 0 AI 0 0 (3.39) Let A be a square matrix.38) with order 4. then A has three Jordan blocks associated with A I as shown in Certainly.6. The matrix is called a block of order = 4 . V3 .A l l) has nullity 3. I 0 0 0 0 0 (J . If A can be diagonalized. Then there exists a nonsingular matrix Q such that (n Jordan n [ 1 A 0 0 0 0 0 o 0 1 A 0 1 A o ] (3. equiv �ently. then the equation has two linearly independent solutions.AI)q = 0 n. . = 3 . AI 0 0 � AI 0 0 0 0 AI 0 0 0 0 AI 0 0 and (J . will yield a correct result only if A has integers or ratios of small integers as its entries. the representation of A is a block diagonal matrix of form A = diag{JI . In this case. which is [A 0 0 OJ'.2) generalized eigenvectors. then the I 's will appear on the subdiagonal. equivalently. .I AQ assumes one of the following forms A. Consider a 5 x 5 matrix A with repeated eigenvalue A I with multiplicity 4 and simple eigenvalue 1. . Jordanform matrices are triangular and block diagonal and can be used to establish many general properties of matrices. = l 1 /1 l l /1 � /1 l 1 Jl l l 1 jJ A' 0 0 A' 0 0 AI 0 0 0 AI 0 0 A. V2. We study first polynomials and then general functions of a square matrix. A is said to be and [ q .38 ) n (A . If k is a positive integer. We use Jordan form extensively because many properties of functions can almost be visualized in terms of Jordan form. V4} is A J := We verify this for the first and last columns. which is [0 0 I AJ'. d 1 j ordan ( a ) .AI) has rank n . 0 0 0 . then A is a diagonal matrix as shown in As. In this case. The first column of J is the representation of AVI = AVI with respect to {VI . vd and {UI . d 1 e i g ( a ) will yield an incorrect solution.38). This verifies the representation in (3.AI)'. . This is called (J . V4 } .AI)2 = 0 0 0 I 0 0 0 0 (3. A . it may assume the form in A2 or in A3. because det (CD) = det C det D and det Q det Q.60 L INEAR ALGEBRA Then the representation of A with respect to the basis {V I . For example. there exist two chains of generalized eigenvectors { V I . A diagonal matrix is a degenerated Jordan form: its Jordan blocks all have order I . ud is linearly independent and can be used as a basis.'\ = Q. Vk . . then A has two Jordan blocks associated with A I . . = 0 0 0 0 0 0 A' 0 A' 0 AI 0 0 0 0 AI 0 0 0 0 0 AI 0 A. � A. they are said to be in Jordan form. V2. . d J e i g ( a ) to generate Q and A as shown in Examples 3. If we reverse the order of the basis.40) AI 0 0 AI 0 0 AI 0 0 0 A' AI 0 0 0 6 0 0 This section studies functions of a square matrix. If the nullity is 2. . 6 Functions of a Square Matrix [l �] [1 1] [l 1]nilpotent. Then we have (J . V3 . = = det A = det Q det A det Q. Thus we have det A = product of all eigenvalues of A which implies that A We discuss a useful property of Jordan blocks to conclude this section.AI) = is lIonsingular ifand only ifit has no zero eigenvalue. The last column of J is the representation of AV4 = V3 + AV4 with respect to {VI . 2.I = det I = I. all eigenvalues of A. from A = QAQ. Consider the Jordan block in (3. If A cannot be diagonized.5 and 3.I . With respect to this basis. respectively. we can use [ q . we have 61 = 3. V2. 1 2 } where JI and J2 are. Thus A has two linearly independent eigenvectors and we need . we (k terms) AI 0 define 0 Ak := AA · . . Jordan blocks of order k and I .A I I ) is I . If (A . . the positions of the Jordan blocks can be changed by changing the order of the basis. ut l with k + I = If V I and UI are linearly independent. The matrix J has eigenvalues on the diagonal and I on the superdiagonal. V4}. .U)k = 0 for k ::: 4. If the nullity is +. . Now w e discuss a specific example. However. V2 .2 or. . UI . nullity 2. . The first matrix occurs when the nullity o f ( A . U2 . . All these matrices are triangular and block diagonal with Jordan blocks on the diagonal. . . � A.6 Functions o f a Square Matrix defective A4.
A.43) by A yields tl (A) = An "'5 = ni n (A A I )4 (A AAZZ)) ""''2 ((AA AI l3Z ((AA A22)) Ad A "'3 (A AI ) (A A ) 2 A I ) 2 (A  = dimension of A. Ak = Consider the similarity transformati�n polynomimonic al we have A = Q. . A . (3. ..44) i s t o use long division to express 1(1. .. AZ = (A = det(AI . . Thus A and or. ( QAQ. if the Jordanform representation of A is available. Then Aj I(A) Suppose the Jordan form of A is known. .. ' . is defined as the largest order of all Jordan blocks associated with Clearly we have iii ::: ni .. because ii may not be available.I . + +. . If the minimal polynomial of A with degree ii is available. Let (A) is defined as I I(A) be a polynomial such as If A is block diagonal.I )(QAQI) .44) = . every polynomial of an x matrix A can be expressed as a linear combination of {I. such as I (A) = A 3+2 2 A .39) are all I .I } .l } . then it is straightforward to verify A= [ �I :J We see that the minimal polynomial is a factor of the characteristic polynomial and has a degree less than or equal to the degree of the characteristic polynomial. A. be the characteristic polynomial of A. .) = ".25). Because characteristic polynomial contains the minimal polynomial as a factor or 6. . . which implies that An+ 1 can be written as a linear combination of {A. however.62 LIN EAR ALG E B RA and AO = I. if all e igenvalues of A are distinct. A ( 42) minimal 3.39) are their indices are.A) = n (A Ai l 4.I } . A . (A)h (A) = 0 .45) h eAl = q A O + h eAl = h eAl .I I(A)Q "' (A) ni. all similar matrices have the same minimal polynomial. Thus "' (A ) as defined is the minimal polynomial. Let be an eigenvalue of A with multiplicity That is. Clearly. (i. . Computing the minimal polynomial directly from A is not simple (see Problem 3.39) are "'I = tl(A) Ai . n . Then Their characteristic polynomials. Using the indices of all eigenvalues. The CayleyHamilton theorem implies that An can be written as a linear combination of {I. • > iii . which.1 a 1A n a'll = 0 ni ( 3 "+3) (3. Because ". Because A k = (QAQ. we conclude that. more general. 3. .I or I(A) = Q. for any polynomial no matter how large its degree is. However. nn a 1 An.3). in turn. Associated with each eigenvalue.dex Ai .in4. Using the nilpotent property in we can show that (3. The of denoted by ii i . h eAl = 0. a A=O·A=0 n An } . Proceeding forward. the minimal polynomial can be expressed as with degree ii = L iii ::: L = of the five matrices in (3. a matrix satisfies its own characteristic polynomial." + + tl( ) + + . A 2 . Theorem 3. 2. The multiplicities and indices of in all five matrices in (3. + + leA). This establishes the theorem.6 Functions of a Square Matrix 63 I(A) = (A I ( A ..44) with the understanding that all discussion applies to ii .. .I A Q o r A = QAQ. A". Multiplying (3. For example. ) for some polynomial h eA l . The of A is defined as the monic polynomial of least degree such that ". respectively. A. then every polynomial of A can be expressed as a linear combination of {I.40).42) is that I(A) = 0 if and only if = o. can be written as a linear combination of {I.31) where A I and A2 are square matrices of any order.6 0r . we discuss in the following only (3. the minimal polynomial can be read out by inspection. then the minimal polynomial equals the characteristic polynomial. I(A) can always be expressed as An+ 1 a l An a _ I A2 n . An. however. ) h ().I } . . (A) = O. the minimal polynomials for some In other words. Note that the 0 is a zero matrix of the same order as A. the multiplicities of A I in all five matrices in (3. 2. we have A = ". . For example.I ) = QAk Q _I I (A) = [ �� :� ] and I (A) = [ / (� I l 1 (�2) ] and that no polynomial of lesser degree meets the condition. all equal 3. One way to compute (3. and I .4 1 ) '>. . This is a better result. An . = where q(A) is the quotient and h eA l is the remainder with degree less than n. . (3. . the In words. (Al = 0. the characteristic polynomial of A is A Q/ (A)Q. A direct consequence of (3. ) as ' f3i .4 (CayleyHamilton theorem) Let ". Then we have I(A) = q (A)tl (A) I(A) ) tl(A) + heAl + () = q ( ). there may be one or more Jordan blocks.. . (A) = 0 polynomial is a polynomial with I as its leading coefficient.61 or I(A) A3 +2A2 (A +2)(4A +2 ) 4 = . ().
+ f3n_I A  AI = unknown coefficients.2).6 Functions of a Square Matrix .1 )1 (A . . Then we have eA 1' = h eAd = (2te' + e2')I + (3te' f (A) = h eAl and h eAl is said t o equal f (A) o n the spectrum o f A.e' . n. Thus we have and h(ll (Ai ) is similarly defined.8 matrix Let A with characteristic polynomial where n = L.5.f3 1 ( _ 1 )99 100 · is AZ = Compute eA:'.5 We are given i f (A. ) = h (/l (A. In this case.5. Theorem 3.te' . ni  I and i n Compute e A 1 ' . m t ( 1 ) = h'( l ) : f (2) = h (2) : Note that. we have h eAl = .45). One way to define f (A) is to use Theorem 3.2).64 LINEAR ALGEBRA Long division is not convenient to carry out if the degree of f (A) is much larger than the degree of 6 (A). .8.1 eigenvalues of A are distinct. These f (/l (A. . .A. for = I . 99 f (.I . these f3i can Clearly A l lJO can also be obtained by multiplying Theorem 3. . . if f p. This is stated as a theorem. . and f32 = el' . 2 . which is the same as for A I .I ) : j'( . compute f(A). Consequently. equivalently.I ) = h ( .el' e 2el' . it is simpler to use 110100 ] 65 Let f (A) b e any function. On the spectrum of A. 1 . we have I ( _ I ) IJO = f30 .l OOA. Or. . in the second equation. . . and = f31 eA:' = h eAl) = [ [: �I � ] 0 _ 2e' . . + f3n _ I An.I ) : f30 = I f3 1 + EXAMPLE 3. where n is the order of A. ) for 1 = O. . . J"' n 1 . .1 ) 2 (A . the differentiation is with respect to A. We solve the coefficients of h eAl by equating f(A) = h eAl on the spectrum of A. ) + + f3I A 1=1 A l lJO = f30I = n . not necessarily a polynomial. Hence we have the same h eAl as in Example 3. + = 99. +[ 0 ] 991 [ 0 99 . Then f (A) is defined as h eAl· Functions of a square matrix f (A) and an n X n EXAMPLE 3. The characteristic polynomial of Az is (A . 1 EXAMPLE 3. Let Then unknowns are to be solved [ 00 0 0 ] I 0 I  3 f(l) = h(l ) : = I. If all be solved from the n equations h eAl : = f30 >.e' .iOOA I I 2 A 100 ? ]. then (3.45) must be differentiated to yield additional equations.e' el' e' � _ 0 0 ] = AllJO. given f (A) 2A + I = (A + Compute A l lJO with Thus we have f31 = 100. Solving these equations yields f30 = 1te' el'.) = e)" . The characteristic polynomial of A 1 )2 Let h eAl = f30 + f3I A. However.2e2'. 2.9 Let + + e' = f30 f31 + f32 te' = f3 1 + 2f32 e2' = f30 2f31 + 4f31 + + + h eAl = f30 + f3IA +.e[ ] . If A has repeated eigenvalues. not t. . we may solve h eAl directly from (3.2e2')A I le' 2el' 2e' . what is f (A J l ? The characteristic polynomial of AI is (A . ) = h(A. = [ 199 1 00 times.':.te')Ar = [+ 0 2 2e' . Define It is a polynomial of degree n I with from the following set of n equations: where h eAl := f30 n  6 (A) = + + f3I A n (A .I ) = h ' ( .2 (el' .2e2' 0 e' tel 2e2t .el' e21 et 2te' 2e' . . I n . Let where the n unknowns f3i are to be solved.(32 ). Let h (A) be a polynomial of degree n . f3 1 = 3te' 2e' .7 i MA) = A2 In other words.100 f3I A = I o I 3 . . ) = q (A i ) 6 (Ai ) + h (A.
.A I ) 2 . ) J"��. t eA j t 0 "" " / J ! t " eA. (A I )/ I ! I (A . .A I l . If I (A ) = eAt .10 Consider the Jordan block of order A= .46).A d4 .47) Using power series The function of A was defined using a polynomial of finite degree. we can readily obtain I(A) = 1 I (A ) = I(A .4 1 ) and (3. then I (A) can be defined as I(A) = I(A) = L /3i A ex.66 L I N EA R ALG E B RA EXAMPLE 3.50) to derive (3.A l l)n1 + .5.A") (3. If I (A ) = eAt.. /33 = ! 2 If I(A) = (s . + I .. t / 2 ! t eAlt eA 1t ] ] (A .A I ) I 3 Using the special forms of (A .) C I ) (A .A l l) + + J" (A I l .4 1 ) and (3.A ) ." = Thus we have This selection is permitted because h (A) has degree (n . � I .1 ) = 3 and n = 4 independent unknowns.48) i=O 00 i Instead of proving the equivalence of this definition and the definition based on Theorem 3. Its characteristic polynomial is (A .A l l I 1(3) (A I )/3 ! ! J" (A d/ 2 /.A l l 0 I (A) = 1 (1.A ) .4 1 ) and (3.48) imply 3 . . A (n n  I)' .A d 2 I I (s .A I 1 )3 0 0 I (s .47) imply (5 . /32 = . Let ["� jJ 0 AI 0 0 0 AI 0 0 A2 0 0 0 then . (A .A 2 ) (3.I . .� Il (A _ A l l) + . then eAt = [ I (A I ) 0 0 o /.42) are applicable to functions.) on the spectrum of A yields immediately . . ( J" (Ad 1 3 ) (A I ) 'fi o = I (A I l .\ _ A I I) 2 + 1 J" (A d / 2 ' 1'(A d / l ! I (A I l (3�. then (3.+ .49 ) (5 . . ) / I ' I (A I l 0 0 0 Because functions of A are defined through polynomials of A. I(A) = I (A I l + I (A I l (A .12 Consider the Jordanform matrix A i n (3. (A . /31 = J ' (A I l . 6 Functions of a Square Matrix 67 (3. we use (3. then (3.1 1 e A�f 0 0 0 0 0 0 0 (5 .' ) can be expressed as the power series with the radius of convergence p.A I I)k as discussed in (3.40). If all eigenvalues of A have magnitudes less than p. L /3i A i i=O (3.A I l3 I (s . We now give an altematiw definition by using an infinite power series. (3. )I + ' . Although we can select h (A) as /30 it i s computationally simpler to select h (A) as [ At 4: o 0 o o + It is block diagonal and contains two Jordan blocks. 1 )1 + 1'(A I l (A .47). . t / 2 ! t eA. . AI) I (s l . Equations (3.46 ) /31 A + fh A + fh A3.50) j ! EXAMPLE 3. ) (. The condition I(A) = h O. ["" 0 0 0 � t eA l ! eA j t 0 0 t "eA . Suppose 1 ( .A l l (5 .11 Consider A EXAMPLE 3.A 2 ) " I 0 0 0 (5 .= 0 0 0 0 ["�" teA J t eA J t 0 0 0 tZ eA l t / 2 ! teA l t eAtt 0 0 0 0 0 eA 2t 0 (5 .
This series involves only multiplications and additions and may converge rapidly: there fore it is suitable for computer computation. Thus w e have F = A * F Ik. the infinite series reduces I immediately to (3.5 I ) yields It can be shown that j(s) : = .9. K = ::" .1 2 )s. and I. This can be verified by direct ! (t) e "dt ( 3 .54). E=E+F.5 1 ) yields Thus we have II ' Then ( 3 .54).I A + s . Because the Taylor  4 To show ( 3 . uses the socalled Pade approximation.54) ( 3 52) (3. It yields comparable results as exp:n2 but requires only about half the computing time. . = commute or and The equality holds only i f substitution of ( 3 . / ) .5 1 ) for I = I : j converges for all finite i. . r = eye ( s i z e ( A ) i . I ) . We see that the series can indeed be programmed easily. Note that norm(a. we must use Theorem 3. 5 1 ) which implies (3. To improve the computed result. Thus the two definitions lead to the same function of a matrix. or eX9m1 . . Thus the inverse of eAr can be obtained by simply changing the sign of I. The next two lines initialize E and F. Because the algorithm compares F and E. ��nc t i on This is an i mportant equation.5 I ). . we can readily verify the next two equalities k=k+1 .53) and (3 .2 .I A ) . not F and O.L:>(k+ll Ak = 5 . l ) is the I norm discussed in Section 3.5 1 ).5 7 ) .56) t CkI (Aj Taking the Laplace transform of ( 3 . The function e x. . We list in the following the program in MATLAB that computes (3. The = computation stops if the I norm of E + F . for k = 1 . hO\vever.L: [ f (t)] = {" .53) converges for I s . < .I A)k x '" L ( s .E. Thus expIr.2 and will be discussed again in Section 3.1 A)k = 1 + S .I 1 . We derive some important properties o f eAr to conclude this section. the algorithm uses norm (E + F .E..52) imply 3.6 Functions of a Square Matrix 69 2 = z er o s ( s i z e ( . is preferred to exp.5 1 ).40).S . but it will yield an incorrect solution if a matrix i s not diagonalizable.I A I and SI Because the infinite series 2 k=O ')C ) s .55) E = exp�2 ( A ) norm ( E + F . = ( l . the techniques of scaling and squaring can be used.:c... . we have ( 3 . +  ). Using (3. The most important function of series .3 uses Jordan form. Differentiating term by term of ( 3 . denoted by norm(E + F . l » O In the program. I ) instead of norm(F.5 1 ) with = 1 . the function expm2 uses (3.J A2 + . �� i l e F = A*F / k . we have eAr = I + A I + A is the exponential function eAr . Let Ck denote the kth term of (3. If a c1osed form solution of eAr is needed. The Laplace transform of a function ! (t) is defined as A B AB BA. Then we have k)c. In MATLAB. The first line defines the function.:. The function expT.2 A 2 + . E denotes the partial sum and F is the next term to be added to E.E. n r" n' +. I ) ..5 or Jordan form to compute e·\r . k=O :x. is rounded to 0 in computers.L: [eAr ] = .I A l = s .68 L I N EAR ALG E B RA Because (A .A 1/ = 0 for k :::: n = as discussed in ( 3 . (3.L: [ S ] = S (k+l) (3 . We mention that (3.47).S .I I + s2 A + s . we set 12 = A �t� 2! + . 2 . . .
8 Some Useful Formu las 0 bI I + bl l 0 al2 bI2 + bl l 0 0 b iZ 0 0 0 0 b2 1 all a2l all + b22 This section discusses some formulas that will be needed later. . Let p CB) where p denotes the rank. 2. Thus (Ai + J.59) [ 2 where Ai. .70 L I N EAR ALGEBRA mil CII 3 . .Lj . 8 Some Useful Formulas Although in the derivation of (3. The determinant of a square matrix is the product of all its eigenvalues. the matrices M and C must be of order n x m. The matrix on the preceding page is a square matrix of order n x m == 3 x 6. 2. (3. == Ai + J. . In other words. == . . .58) m 21 s. In AB. The MATLAB function m = lyap ( a .55). A operates on the rows of B. The equation can be written as a set of standard linear algebraic equatio/ls. that is.. To see this.Lj == (Ai + J. respectively. == A(M) == I] M n. n x n and m x m constant matrices. solutions may or may not exist. Let n x p constant matrices. We show intuitively why this is the case. applying the Laplace transform to (3.Lj) is an eigenvalue of A. Thus a matrix is nonsingular if and only if it has no zero eigenvalue.59).Lj == then the square matrix in (3. Let U be an n x right eigenvector of associated with Ai .55) yields or which implies (3. then for a given C. Because L[df(t)/dt] sL[J(t) ] . Thus linearly independent p (AB) ::: min(p(A). we assume n == 3 and m == and write (3. j it has nm eigenvalues 17b for I == I. The equation is called the Lyapunov equation. respectively. If A.Lj)uv i 3 . Then B has ct a s follows . vB == vJ. n. It turns out 3 . . we obtain a l l + bl l a21 al l b l2 an all al3 a2l all bI I [ ] a2l aI2 a l3 a2l el l al2 C2 I Cl i C I2 a ll ][ ] [ ][ m il m 12 mZ l m 22 mll m l2 + m il mI l m 12 m Z2 bl I b2 1 bl2 b22 ml l m 32 ] cn Cl 2 Because both U and v are nonzero. pCB »� (3 .59) explicitly as AM + MB == C (3.58). 5l can be considered as a square matrix of order n m . Applying A to the n x m matrix uv yields I]k i I. Then we have A and B be m x n and 0 an a22 a l l + b22 a21 al l 0 0 + b22 al2 0 ct . . If there are no i and j such that Ai + J.Lj for i I.. Au == AiU. It maps an nmdimensional linear space into itself. are.eO == AL[e A 1 ] This is indeed a standard linear algebraic equation. A scalar I] is called an eigenvalue of A if there exists a nonzero M such that == e22 Cl2 2 Because k == 1. 6 1 ) .frO).60) is nonsingular and.57) we require to be sufficiently large so that all eigenvalue s of A have magnitudes less than I. Equation (3. . there exists a unique M satisfying the equation.Lj' j == m. the eigenvalues of A are all possible sums of the eigenvalues of A and B. n m . Then the Lyapunov equation can be written as A(M) == C. If C lies in the range space of jt then solutions exist and are not unique.60) 71 ell == sL[eA1 ] . the eigenvalues of A and B. This c a n b e argued the rows of AB are rows.. Equation (3. 2. Let us define 5l(M) : == AM + MB. the Lyapunov equation is said to be nonsingular. 7 Lyapunov Equation Consider the equation where A and B are. m A al l a21 all == MUltiplying them out and then equating the corresponding entries on both sides of the equality. .c ) computes the solution o f the Lyapunov equation in (3 . I. 5l(uv) == Auv + uvB == AiUV + UVJ.32. In this case. + J.I s s x m3 1 m I2 m 22 1n 32 el l Cl i (3.Lj == 0 for some and j. In order for the equation to be meaningful. b. 2. See Problem 3.58) actually holds for all except at the eigenvalues of A. so is the matrix uv. Let v be a 1 x m left eigenvector of B associated with J. 2 0.58) can also be established from (3. . for every C. and J.Lj : that is.
Because ( x *Mx) ' = x'l\I'x = x'I\I'x = x'�1x v'Mv = v'Av = A ( Y'V) and because both v*Mv and v* v are real. s'" deUsI" . Let A be m x 11 and B be 11 x m. . This assertion i s not true if l\I is not symmetric.65) 3. Then we have f3 (3. (3. and P.6 1 ) ."I)x i= 0 0 67) ((33 68) Because . and adding the product of one row (column) and a number to another row (column).AB) = In words. To show we define Because p eA) ::: minim. the rank of a matrix will not change after pre. After establishing this fact. Q. Thus (3.BA) They are useful formulas. Taking the complex conjugate transpose of x'�lx yields An 11 X 11 real matrix M is said to be symmelric if its transpose equals itself. is called a q lladralicjorm . Let A b e m x 11 and l e t C a n d D b e any 11 x 11 and m x m nonsingular matrices.6. .AB) = det(sl" .V. p (A C ) = p eA) = p (DA) :. if AB has at most f3 linearly independent columns. we show that there is no generalized eigenvector of grade 2 or higher. we have det N = det 1 m .63) (2 ) P : = AC pep) ::: min (p(A) . These operations are the same as muilIplY1l1g nonsingular matrices. A consequence of is that the rank of a matrix will not chancre O by elementary operations. we have p eA ) ::: p cP).72 LINEAR ALGEBRA linear combinations of the rows of B. then we can readily obtain det(QP) = det Q det P = det P S ol det(sl".. interchanging two rows or two columns." .62). Thus if A has linearly independent columns. Suppose x is a generalized eigenvector of grade 2 or B cause N and Q are block triangular. and . Mv = t. B operates on the columns of A. 542].64). [ ] ] In where we have used the fact that the complex conjugate transpose of a real M reduces to simply the transpose. We claim that every symmetric matrix can be diagonalized using a similarity transfor mation even it has repeated eigenvalue . the eigenvalue . we can return our study to exclusively real vector x . for 11 = III or for 11 x 11 square matrices A and B.9 Quadratic Form and Positive Definiteness (3. .or postmultiply ing by a nonsingular matnx. Let A be an eigenvalue of M and v be its eigenvector. To show this.6 1) (3. In AB. where x· is the complex conjugate transpose of x. 3. ftl". let us define 0 (3.62) (1 ) (3) [6. The scalar function x'Mx. Thus we conclude p (P) = p iA ) . p i C» � ::: p e A) where 1m is the unit matrix of order To show Q = We compute NP = and [ �� � ] I 1m + AB B A (3."1) 2x = det(QP) = det(l" + BA) Consider (M .1ext we write as A = PC.63 ) which implies. " then the rank of A i s equal to or smaller than the smallest rank of B i . respectively. i f I" . (3. that is. Thus x'l\Ix i s real for any complex x. their determinants equal the products of the determinant � ot their blockd iagonal matrice s or Likewise. T h i s shows that all eigenvalues of symmetric :\1 are real. Then we have det ( lm + AB) = det(l" + BA) m.BA) implies (366) ( 3 .B . See Reference p. I"" and B are replaced. det(sI" . det In = I = det Q det ( NP) = det(lm + A B I" + BA ) (M .9 Quadratic Form and Positive Defin iteness 73 det(NP) = det N det P = det P and we conclude det(lm + ABl = det(l" + BA). Thus AB has at most a linearly independent rows.64) A I" ] The eigenvalues and eigenvectors of real matrices can be complex as shown in Example Therefore we must allow x to assume complex numbers for the time being and consider the scalar function x*Mx. Using the same argument. We show that all eigenvalues of symmetric M are real. 11) and p iC) = 11 1 we have p eA ) ::: p iC). then Consequently." must b e real. Elementary operations are multiplying a row or a column by a nonzero number. I n N. where x is an 11 x I real vector and M' = M.62) (3. by ftl".I .. 3. This establishes A = B I B2B3 .
This contradicts p (H'H) = Thus p (H'H) implies p (H) = The condition p (H) = implies p (H'H) = Suppose not. (3. It becomes positive definite if is nonsingular. m. Clearly A is nonsingular and we have which imply AA' = AA = I = A' A. All the leading principal minors of M are positive (all the principal minors of 1\1 are zero or positive). e. [ nm12il1 mm 2I2 ] . n. there exists an orthogonal matrix Q such that Thus the leading principal minors of M are the determinants of the submatrices of M obtained by deleting the last k columns and last k rows for . Thus Q is an orthogonal matrix. with � has rank n .I = Q ' and Q ' Q = QQ' = I. which implies H'Hv = O. Q. Similarly. The second part can be established similarly. H. we can show that the Jordan form of 1\1 has no Jordan block of order 3 or higher. This is summarized as a theorem.8 matrix 1. v'H'Hv = (Hv)'(Hv) = I I Hv l l i n. with has rank if and only if the matrix n m or or Theorem 3. or p (H'H) < then there exists a nonzero vector v such that H'Hv = 0. The leading principal minors of M are which implies Q . It is denoted by M � 0. Thus the principal minors are the determinants o f a l l submatrices o f M whose diagonals coincide with the diagonal of M .ncy: H'H n.69). its columns are orthonormalized eigenvectors of M . which implies v'H'Hv = O or Necessity: Suffi=cinen. H. mxn m n. This is a contradiction. 3. Therefore the Jordan form of M has no Jordan block of order 2.E. A square matrix A is called an if all columns of A are orthonormal.D. see Reference [ \ 0]. m33. Condition ( 1 ) can readily be proved by using Theorem 3. then ' ' ' ' x Mx = x N Nx = (Nx) (Nx) = I I Nx l l i �0 and We discuss the relationship between the eigenvalues of H'H and those of HH'.I (3.I ) x] ' (M .1x = 0 if and only if x = O. n n n n. Thus M is positive definite. A symmetric matrix M is said to be positive semidefinitpositive definite.6. Next we consider Conditon (3). Suppose p (H'H) = but p (H) < v such that Hv = 0. Thus M is positive semidefinite. m xm M = ' det 11l:! 1 det det and det M k = 2. 9 Quadratic Form and Positive Definiteness 75 which is nonzero according to (3. 1.).7 A symmetric following conditions holds. There exists an < 2.69) equals its own transpose or I A ' A = Ii I orthogonal matrix M = QDQ . An det(HH') m m m m22. which are all real. where D is a diagonal matrix with the eigenvalues of M .69) Its principal minors are m and A . Consider where D is a diagonal matrix with real eigenvalues of M on the diagonal. n. then there exists a nonzero x such that x 'Mx = O. m n) n x n nxn The symmetric matrix is always positive semidefinite. We use an example to illustrate the principal minors and leading principal minors. Thus we conclude that there exists a nonsingular Q such that for any x. denoted by M > 0. Thus the inverse of an orthogonal matrix equals its transpose. Because both H'H HH' are symmetric and positive semidefinite.67). jf x' 1\Ix > 0 for every nonzero x.AI ) x = x ' (M ' . For a proof of Condition (2). Consider (3.68) but is zero according to (3. [:�3m:1 il :�3�2 :�3: 1 [ 1IIIl 1 l [ m 31 m i l . O. This contradicts the hypotheses that \" oF 0 and p (H) = Thus p (H) = implies p(H'H) = This establishes the first part of Theorem 3 . If N is singular.74 L I N EA R ALGEBRA [ (M .. 0= The condition p (H'H) = implies p (H) = We show this by Then there exists a nonzero vector contradiction. the only x to make Nx = 0 is x = O. if x' Mx � 0 for every nonzero x.AI' ) ( M . H'H n. Proof: matrix M is positive definite (positive semidefinite) if and only if any one of the such that 1\1 = N'N. nxn mxn m :::: n. The argument in the proof will be used to establish the main results in Chapter 6: therefore the proof is spelled out in detail. nonsingular matrix N (an nxn singular matrix N or an mxn matrix N with Thus we have H\" = O.A I ) X = x' (M . if and only if the matrix H'H has rank HH' has rank matrix oF O. n.6 For every real symmetric matrix M . there exists a nonzero X to make Nx = O.\. their eigenvalues are real and nonnegative (zero or . M = QDQ' or D = Q' M Q 2. 8 . If M is positive semidefinite. Theorem 3. > Theorem 3.I = A' and det M . This property will be used repeatedly later. Every eigenvalue of 1\1 is positive (zero or positive). 1. I f N is nonsingular. If M > 0. If M = N' N . then x ' .AI ) 2 X 3 . on the diagonal. We give a proof of this theorem. Because D' = D and M ' = M . An det(H'H) oF O.
. � M = H'H = and compute its characteristic polynomial as det(AI M) = = J26. following Theorem on the diagonal. range space.14 Consider the matrix in a= [ r . Once R.83877 0.2627 0. The function l". 2.25) 5.76 L I N EAR ALGEBRA positive). 0. m ) . S.65996 0. there exists. and S being 111 x 11 with the singular values of H on the diagonal.. they have at most ii := min (m .25 O.. S .25A A(A 26.7 1) 3 . In MATLAB. Typing o r t h ( H ) and l". A. and semidefinite. Indeed.25 EXAMPLE 3.580' j 0 0 n [0. 2 1) (3.r) columns of Q are an orthonormal basis of the null space of H.2439 70 2 0 . two nonzero singular values.5255 2.14).A3 . the rank of H equals the number of nonzero singular values.26. = > = =. [0.3434(3. ::: 0 2 0 : . The matrix S is 111 x n with the v.3+3644 0. its nullity is and the last two columns of (f The first two columns of r are the orthonormal basis in are the orthonormal basis i n ( 3 . and null space. n ) . it i s very reliable and gives a quantitative measure of the rank.25. Furthermore.. A. 1 235 O. the singular values of H can be obtained by typing s svd ( E ) Typing r R . .0'295 0. n ) number of nonzero eigenvalues. 1 2 3 � . 0. If H is m X n. respectively.+S3729 ] [ 6' rS 2.0. 2 1 82 0. = x 3 matrix Theorem 3. > = =.31+08+68 0'0.0. The columns of Q are orthonormalized eigenvectors of H'H and the columns of R are orthonormalized eigenvectors of HH'. Q'H'HQ = D =: S'S 3. [ 21 10. O.65) becomes m This implies that the characteristic polynomials of HH' and H'H differ only by s n . (3. Thus all eigenvalues ot M are real and nonnegative (zero or positive). Ar 0 Ar+1 An 2 [ 4 1 2 0. Define M := H'H.9 (Singularvalue decomposition) Every nz x n matrix H can be transformed into the form H = RSQ' with R'R = RR' = 1m .8384 0. 1 568. . 0. A: ::: . Manipulation on will lead eventially to the theorem that follows.6. q ] o S'!:i ( a ) �0 1 which yield r= lYI := HH' = and det(AI  M) = = 5. orthonormal bases of the range space and null space of H..40686 00 782 0.5 10 1. 5 ] 10. 1 3 ) 40.25) 26. symmetric. The singular values are usually arranged in descending order in magnitude. Although computing singularvalue decomposition is time consuming. Thus it i s used in MATLAB to compute the rank.n det(sln .25 A2 26.5 5. 1 0 SingularValue Decomposition A I ::: A j ::: ') Let H be an nz x n real matrix.H'H) ( 3. The last (11 . on the diagonal. Let r be the number of its positive eigenvalues. Then the eigenvalues of M = H'� can be arranged as Let ii := min (m .'\. we have Thus the eigenvalues of HH' are and 0 and the singular values of H' are and We see that the eigenvalues of H'H d iffer from those of HH' only in the number of zero eigenvalues and the singular values of H equal the singular values of H'. 10.2475 9 3 99 s 1 2 . Clearly M is n x n. Then (3. det(slm  HH') = s". If the rank of H is r.4686. Q'Q = QQ' = In. .0. 0 Ar+1 An A2 ::: .0. and Q are computed. then H'H has n eigenvalues and HH' has m eigenvalues. 77 3 .70) For . 1 235 q = are and The matrix has Thus the singular values of the matrix A in p eA) = 2. consequently. s .70).4+5877 0. In view o f w e can also compute the singUlar values o f H from the eigenvalues of HH'. u l l will be used repeatedly in Chapter is called the singular values of H.13 Consider the We compute = and and the s ingular values of H are Thus the eigenvalues of H'H are and Note that the number of singular values equals min ( n .3395 3498 793 81 802 0.u l l ( H ) yields.455986.. = H'H. (3. 1 1). Thus we conclude that HH' and H'H have the same nonzero eigenvalues but may have different numbers of zero eigenvalues. H= 7. .'here D is an n x n diagonal matrix with singular values i.0..25A 2 A2 (A 26. the first r columns of R are an orthonormal basis of the range space of H.. 26. Then the set ' " 1 EXAMPLE 3.5 5 ] 2. 1 1 Norms of Matrices an orthogonal matrix Q such that ( 71 ) 3. � We type = . Q J svd ( H ) yields the three matrices in the theorem. (3. 0 0. Let A = H and B = H' . thus its rank is and.
3 . What is the representation of the vector x with respect to the basis { q l .::. n X.2 3. I I Ax l l . 2nonn.3 Consider Fig.3 Different norms of A.1 3. n . then A' A = 1m· What can you say about AA'? Find two orthononnal vectors that span the same space as the two vectors in Problem 3.3.5 Find the ranks and nullities of the following matrices: Consider an n x m matrix A with ::: m. I I A I 1 2 = 3 . i n f ) compute the three nonns. and I I A l lco = 3 + 2 = 5.72 ) =1 IIxll2 =1 2 4 lail l = largest column absolute sum The sum of these two magnitudes gives the nonn of A. if then I I A I I I = 3 + I . Problems 7' I I A I I = sup where sup stands for supremum or the least upper bound.:::: I I A l l l lxl 1 I I A + B I I . then 2 I I A I I "" = m�x These nonns are all different for the same A. 3. 1 ) . and infinitenonn of the vectors 3.. h ) ? What is the representation of q l with respect to {i2 ' q2 P What are the I nonn. then I I A I I I = mfx r l =l where a'i is the ijth element of A. For example.:::: I I A I I + I IB I I I I AB I I .:::: I I A I I I I B I I The reader should try first to solve all problems involving numerical numbers b y hand and then verify the results using MATLAB or any software. 3. This concept is needed in Chapter Let A be an m x matrix. The nonn of A can be defined as 5. The nonn of matrices has the following properties: Figure 3. if the I nonn I lx l l l i s used.I I = 4. = sup I I Axl 1 I lx l l IIxll=1 (3. ft ) � X. and norm ( a . 2 ) =norm ( a ) .2.r. as shown in Fig. This nonn is defined through the nonn of x and is therefore called an For different I l x l l .4 3.. 1 .::+t+"+=++f x 4 . we have different I I A I I · For example. If the Euclidean nonn I I x l 12 is used.78 LINEAR ALGEBRA 3 . 1 1 Norms of Matrices The concept of nonns for vectors can be extended to matrices.7.. I I A I 1 2 = largest singular value of A = (largest eigenvalue of A ' A) I/ If the infin itenorm I l x l l "" is used. If all columns of A are orthononnal. This magnitude gives the nonn of A. (e) This magnitude gi yes the nonn of A .OO I I Ax l 1 IIxll . then induced norm. The MATLAB functions norm ( a . (t ) l aij I J= I 2 = largest row absolute sum ++..3 2 A= [ � �] I 2 x. norm ( a .
. . 3..7 2 2 3 o Find bases of the range spaces and null spaces of the matrices i n Problem Consider the l inear algebraic equation 3. equiva I lently. 3.3 that has the smallest Euclidean norm.1 0 Find the solution in Problem 3.25 20 4 4 3 16 Problems 81 1 Tl ] then A T A. + + 1 A3 2 A 3 A3 • A l 2 A.6 3. AnI b] : :: = :: ] u [ O] Find the solution in Example 3.17 Consider with A and T > the three columns of i= 0 O.16 Show that the companionform matrix in Problem 3 . Ab..8 Find the general solution of 2 I 2 3 [ I I)'? x Note that all except A.(X 0 00 0 l al = [ � 0 I J 20 = [ :. the eigenvectors are l inearly independent if all eigenvalues are distinct. can be diagonalized.(Aj .8 that has the smallest Euclidean norm. show that its i nverse equals o o and the what are the representations of A with respect to the basis { .80 L I N EAR ALGEBRA 1 o 3..1 1 where A i s an matrix and i s an x 1 column vector.5. '{f' [' 1 Al A2 I Al .. respectively? (Note that the representations are the same ! ) 3 .nu [n .. u[bI]. 3. 1 4 i s nonsingular i f and only if 0'4 i= O..12 Given [ j equals f1 Si<jS. A3b} 3. ) .9 o Show that its characteristic polynomial is given by 3.I] x[n] bHint: n x n u[O]. A I AJ = It has three equations and two unknowns. basis {b. ) = A + a l A 3 + aZA2 + aJA + 0'. Ai = 0. Under what conditions on to meet the equation for any A and will there exist and Write the equation in the form x[n] = Anx[O] An. A= Show that [[�0 0 I)' AT A o is a generalized eigenvector of grade 3 and . l Under this assumption.14 Consider the companionform matrix = [ � 04 �O ] 0I [� 0 n 2 Az A4 exist in the equationoO Is the o How many parameters do you have? 3.Ibu [O] bu [ ] Abu [n 2] bu [n . Consider the equation + + Show also that if Ai is an eigenvalue of A or a solution of 6 (A) i s a n eigenvector o f A associated with Ai . Thus we conclude that the matrix is nonsingular or. [ I] ' An 2  A3 2 A� A2 l + .A. Alb].15 Show that the V 4 /'l ( i.I] x [O ] ? [ x[n]Anx[O] = [b Ab . Al b. 1 3 F i n d Jordanform representations of the following matrices: b Ab.I A4 1 3. Does a solution ' solution unique? Does a solution exist if y = 1 3. AZb. alldenllollde determinant 3.
26 to show where tr stands for the trace of a matrix and is defined as the sum of all its diagonal entries. A 1 03 . in Problem 3 .21 Given find A 1 0.19 Show that if A is an eigenvalue of A with eigenvector x.82 L I N EA R ALGEBRA constitute a chain of generalized eigenvectors of length 3.2 + (s 2 + a l s + (2)An . and R. Verify 1 QI A Q = ] ['� ] AT 2 AT o 0 1 0 Find a matrix B such that Let eB 0 1 A o 0 A 3. 3. Ak = 0 for k :::: f (A) is an eigenvalue of if and only if A has m ll (s) := det(sl . 2 3 Ro = 1  3.26 Define m e B eB = C. there exists a matrix B ] = 0 for some i. then B does not exist. Is it true that. eA eAt t = 3. are constant matrices. (sl ..25 Let Find a B such that = C? such that 3.I =  1 ll (s) Adj (sl . and eAt .I + (s + a l )An.28 Use Problem 3 .A) := sn + a l s n .cteristic polynomials and the minimal polynomials of the following matrices: 3. Show that if A. tr(AR2 ) an_ 1 = for A l and A. Verify tr(A Ro) a l = .26 to prove the CayleyHamilton theorem. that is.23 Show that functions o f the same matrix commute.I + a2S n . Verify for the matrix A 3 in Problem 3 .1 tr(ARI ) a.= . Consequently we have A 3. A 1 [ Problems 83 0 0 A C= = 0 o A 0 C. (sl . then f (A) with the same eigenvector x.. 3. for any nonsingular C.A)I = 1 ll (s) [An . . 1 3 . This process of computing Oi. .3 . 3. f (A) g ( A) = g (A) f (A) A.1 .A) .24 Let 3.2 + . tr(ARn _ 2 ) n _ 1 3. + an and R. Such a matrix is called a nilpotent matrix. is called the Leverrier algorithm.A).20 Show that an n x n matrix has the property [] n [1 n [1 n [1 n 0 0 Al Al I 0 Al 0 Al and let (s) be the monic greatest common divisor of all entries of Adj (s I .18 Find the char'l. 1 3 that the minimal poly nominal of A equals ll (s ) / m (s). This definition is valid because the degree in s of the adjoint of (sl .22 Use two different methods to compute eA t 0i 3 = .A) 0 0 0 0 0 0 0 Al 0 Al 0 0 Al 0 Al where 0 0 0 0 eigenvalues 0 with multiplicity n and index m or less.A) is at most n .27 Use Problem 3.
Show that P that is. under what condition will it be a solution? Is A'(AA').I + a l sn 2 + )1] · · · + an . .65).33 Check to see if the following matrices are positive definite or semidefinite: 3 1 3 3. Define = [q l q2 .I Problems 85 3.I y a solution? Show (3.59) with What are the eigenvalues of the Lyapunov equation? Is the Lyapunov equation singular0 Is the solution unique? 3. 2 o .+ ( sn. = Aiqi .38 Ai . then ( s l . Aq.31 Find the M to meet the Lyapunov equation in (3.= '" .36 Show 3. Is (NA). Consider Ax = y.Ai 3.29 Let all eigenvalues of A be distinct and let qi be a right eigenvector of A associated with Ai .34 Compute the singular values of the following matrices: 1 3. P i A = AiPi . 3 1 for A with two different 3.35 If A is symmetric..84 L I N EAR ALGEBRA +.37 3. where A is m x n and has rank m . . .. what is the relationship between its eigenvalues and singular values0 I0 I J [2 2] 2 0 02 [ I� 0� �] �J [ 2 B= I C.32 R epeat Problem 3 . qn l and define Q where Pi is the ith row of .qiPi qi and p.I A'y a solution? If not. that is.30 Show that if all eigenvalues of A are distinct. are right and left eigenvectors of A associated with A. (sl where  l A) .A)I can be expressed as [�:1 1 3. P � Q '� Pi is a left eigenvector of A associated with 3. � s .
54) and ( 3 . This discretization is the easiest but yields the least accurate result for the same integration step size.At is 86 where tl is called the integration step size. ) = m =ko L g(ktl. 4. how to transform transfer functions into state equations.5) reduces to + 1t I as discussed in (3 . A better method is to transform transfer functions into statespace equations and then compute the solutions. B. 1 ) eAtx(t) .eAtAx(t) which implies = eAtBu(t) y (t) = The easiest way is to compute it numerically on a digital computer. therefore computing solutions using the Laplace transform is not a viable method on digital computers. There is no simple way of computing the convolution d _ eAt = AeAt dt = eAt A to develop the solution. when there are repeated poles.2) (4. 1). g(s) will be a rational function of s. Except for some special cases. These can be carried out using the MATLAB functions roo t s and re s idue.34). The problem is to find the solution excited by the initial state x(O) and the input u(t). The solution hinges on the exponential function of A studied in Section 3 . In this case. n x n. n x p.2) and the initial condition x(t) = x(O) at t = Indeed. we can also use yes) = g (s)u(s) to compute the solution. 1 Introduction if lumped. if the Laplace transform of u (t) is also a rational function of s .2). Before doing so. (4. This is basically the discrete convolution discussed in (2. see. (4. This analytical briefly how to compute solutions of the inputoutput description. r )u (r) dr Its integration from 0 to t yields eAr x(r) I =o Thus we have : = 1t = eAr Bu(r) dr y (k b. For other integration methods. It is instructive to verify that (4. at t = 0. respectively.6. and then using a Laplace transform table. One way is to discretize it as l�to k g (t. Premultiplying eAt on both sides of (4.4 . the equation must be discretized.5) . and q x p constant matrices. To verify this. m tl) u (m tl ) tl (4.5) is the solution of (4.4) implies 1t eArBu (r) dr (4. we need the property in (3. q x n. carrying out eAt and eO = x(t) = eAt x(O) This is the solution of (4. First we discuss by statespace equations. For the linear timeinvariant (LTI) case. and other related topics. for example.3) where A. This chapter discusses solutions of state equations.52). it is simpler to compute the solution directly in the time domain as in (4. C. chapter discusses how to find their solutions. If the system is lumped. the computation may become very sensitive to small changes in the data.2) yields eAtx(t) . We discuss first the timeinvariant case and then the timevarying case. If a system is distributed.4) eA(tr )Bu(r) dr O. 2 Solution of LTI State Equations 87 Chapter StateSpace Sol utions a n d Real izations 4 partial fraction expansion. then the solution can be obtained by taking the inverse Laplace transform of g(s)u(s).eO x(O) Because the inverse of e . This method requires computing poles. However. including roundoff errors. Reference [ 17].2 Solution of LTI State E q u ations Consider the linear timeinvariant (LTI) statespace equation y(t) x(t) = Ax(t) = Cx(t) ++ Bu(t) Du(t) (4. and D are. We showed in Chapter 2 that linear systems can be described by convolutions and. (4. g (s ) will not be a rational function of s. In particular.2).5) satisfies (4. we must show that (4.55) or 4.
S8) to compute eAt LI(sl A)I _ eAI.I [X(O) + Bu(s)] yes) C(sl . then Let then (3. + Method 1: We use (3.S7). Using the infinite power series i n Although the series wiII not. w e can use (3.2). A).20)  = (sl (s  A).2) .I Q(sl . We discussed in Section We now give some remarks concerning the computation of 3 . = next. /30 /3 I A. 5. yield a closedform solution. then  to compute I 2 (sl I A = [ OI 2 ] and to compute A).A).I . Using Theorem degree n I that equals 3.I Q.48). We can also compute the solutions by using the Laplace transform. 4. as discussed following eAt can readily be obtained by using A.1 .A). Taking the inverse of Using Theorem Using = Using the infinite power series in (sl (sl . I t The matrix function is the inverse Laplace transform of in the preceding example. eAI Qe·.1 We use Methods Thus meets Substituting (4. xes) =  (4. where Let h eAl = =h =h : = + If h equals f(A) := Once and Yes) are computed algebraically.A). (4. 14) (2. find a polynomial = where eAt. T) l r=1 at at (4.S I): = (3 . (sl eAI : heAl . (4. A: A QAQ. again.2) (4.I .) 2 ] I/(s + 1) s/(S+)A I.I = heAl [(s + 1)1 + (5 + 1)2 ]12 + (s + 1)2A = + [ (5 + 2)/(s 1)11)2 1/(5 + . T) dT 1101 (i f(t.S). 1 eA(I r)Bu(T)dT] x(t) :t [eAlx(o) + 1 [' AeAlx(O) + io AeAIIr)Bu(T ) dr + eA(tr)Bu(r) l r =1 A (eAtx (O) + 11' eA(Ir )Bu(T ) dr) eA °Bu(t) = satisfies = Differentiating and using = = i (4.S) x(t) Ax(t) + Bu(t) (4.49). is a function of  A.S : A eI First. (A) A.3 ) (4 S) (4 y(t) CeAt x(O) + C it eWr ) BU(T) dr Du(t) and the initial condition and is the solution of yields the solution of into as (4.I .S) satisfies the initial condition.3 ) x(O) (4. T») dT + f(t.A) 3. 3.\IQ.6). Applying the Laplace transform to and yields. after substituting i l10l f(t.. Thus we have eAt x(t) eAtx(O) + it eArI r)BU(T)dr = which was computed .2).S. it is suitable for computer computation.S) (2.S I).88 STATESPACE SOLUTIONS AND REALIZATIONS The inverse of (s I it: 4.6) we obtain 1. their inverse Laplace transforms yield the timedomain solutions.2 Consider the equation A is in Jordan Its solution is x(t) [� =�] x(r) + [�] U(t) (51 A).1 o n the spectrum o f [ � I s � 2 r = S 2 + �s + I [ s : 2 [ (S + 2)/(s +2 1)2 1/(S + .) ] s/(s+)1/(5 + = = of EXA\lPLE 4.7) (4. They can all be used to compute xes) (sl A).  In addition. 3. compute the eigenvalues of on the spectrum of = Using Jordan form of form and 3. and Using the Leverrier algorithm discussed in Problem (3. We need the equation to show that which becomes.  EXAMPLE 4.8) that is.I . in general.). therefore. A: eAt heAl .I (3. 2. 2. f(I) ( I ) : (s + 1)1 /30 /3 1 /'(1) '(I ) (s + 1)2 /31 = are .S ) (4.26. (3.2 Solution of LTI State Equations 89 . as derived in and (4.A)I[x(O) + Bu(s)] Du(s) = + Thus we have =  and 1. we have many methods to compute Thus (4.S) (4.I = Method 2: The eigenvalues o f = + This solution and are computed directly in the time domain. (4.3) .IS).6 three methods of computing functions of a matrix.
t }. of has a negative real part. teAlAt. 1 0) Thus. = =: for t < (4. If A has some eigenvalues with zero real part and all with index 1 and the remaining eigenvalues all have negative real parts. then most zeroinput responses will grow unbounded as t > 00.. it must be discretized. We discuss next a different discretization. In has eigenvalue A with index ii I ." = Q i!�lt 0 0 t ZeAlt /2 eA1 t 0 0 Consider the second 0 0 0 0 0 0 Q ' and. . if tl. However. 4. Then we have Every entry of general.9) still equals (4. AA (s and I s = 1 = J� [ I � . then contains the term t. . Let for the solution of (4.5) at t = = := = I .90 STATESPACE SOLUTIONS A N D REALIZATIONS We discuss a general property of the zeroinput response matrix in (3.39 ). . u(t) uCt) u(t) u(kT) u[k] kT (k + I)T k 0. then no zeroinput response will grow unbounded. then it will grow unbounded.9) (4. 1 1 ) and (4. of the zeroinput response is a linear combination of terms These terms are dictated by the eigenvalues and their indices. . I eAt t. after substituting (4. 15) (4. 1 Discretization which becomes. i f an input changes value only at discretetime instants responses at = x(t + T) x(kTt) + Ax(t0. This input changes values only at discretetime instants. If has an eigenvalue.2 Solution of LTI State Equations 91 (4. kT t (k + x[k] x(kT) eAkTx(O) + tT eA(kTr)Bu(r) dr x[k + ] x«k + 1)T) eA(k+I)Tx(O) + tk+l)T eA( k+I)Tr)Bu(r) dr x[k + ] eAT [CeHAkIT)Tx(O) + lkT eACkTr)BU(r)drJ + lkT eACkT+Tr)Bu(r) dr ct kT+Tr. then (4. . if Every such term is analytic in the sense that it is infinitely differentiable and can be expanded in a Taylor series at every This is a nice property and will be used in Chapter 6. If every eigenvalue. 2. If a zeroinput response contains the terms has eigenvalue with index 2. 1 1 for = 1 . with a positive reai part. eA. . 1 4) can be written as 1 = Jo (4.5). = hm i(y(tt)) Ax(ttl) ++ Du(tt)) BuC Cx( x(t + T)T x( x(t) T. Because .10) become = (I = This is a discretetime statespace equation and can easily be computed on a digital computer. 16) . then some zeroinput response may become unbounded. If an input This discretization is the easiest to carry out but yields the least accurate results for the same is generated by a digital computer followed by a digitaltoanalog ::: converter. 1 become 1 = = (4.tCt(tr)r)e]U(tr)ur)(ur()rdr)dr ] e eAt x ( O ) . 14) . x[k + ] eATx[k] + (IT eActdct) Bu[k] t kT. 1 2) and = 1 ) T yields and Jo (4. .9) as If we compute only at t = = 4._O t)= (4. simple or repeated. simple or repeated. then will be piecewise constant. 13) and introducing the new variable Consider the continuoustime state equation = If the set of equations is to be computed on a digital computer. This situation often arises in computer control of control systems. I = and if we compute only the then (4. 13) 1 := = Equation (4 . . then every entry of is a linear combination of A 0 eAt (I. For this input. x[k + ]0) Adx(k] + BdU[k] k T y[k] Cdx[k] + Ddu[k] w e can approximate (4. consequently. t2eeA'tlt .9) and (4. teAlt teAlt l eA1l eA:2t ] ["" {eAlt. For example. 1 2) and (4.2. Computing (4. if the index is 2 or higher. + 1)2 [ (I +tet)tet (S+)2 1 X(t) [ (I +tet )tet ( I_tte)tet ] x(O) + [ _ .)T + Bu(t)T ) x(t) and y(x«k)t + I)T) + TA)k x(kT) + TBu(kT) y(kT) Cx(kT) + Du(kT) T. then every zero input response will approach zero as t > 00.
1 9) which implies that every entry of the zeroinput response is a linear combination of A. then some zerostate response may become unbounded. k (4. However.3 Equivalent State Equations The example that follows provides a motivation for studying equivalent state equations. for I i 3! 4' is nonsingular. We discuss the computation of Bd. b . Note that there is no approximation involved in this derivation and (4.).39). internally. The two equations in (4. of has magnitude less than I . 2.2 Solution o f DiscreteTime Equations Consider the discretetime statespace equation 1] = = where the subscript has been dropped. First we select the inductor current X I and . Similar remarks apply to 1 s irn. one inductor. we can avoid computing an infinite series. 1 5) yields the exact solution of (4. then it first uses t f 2 s s to transform the transfer function into a continuous time state equation and then discretizes it by using c 2 d and then uses ds t ep to compute the x[k +y[k] Ax[kk]] ++ BDu[U[kk]] Cx[ d x[O] u[k]. 1 7). . I .5 1 ). if the index is or higher. A 2 x[x[2l]] Ax[Ol]] ++ Bu[Ol]] A2x[O] + ABu[O] + Bu[l ] Ax[ Bu[ k x[k] Akx[O] + m=O AkImBu[m] y[k] CAkx[O] + m=O CAkImBu[m] + Du[k] A2 Al Akx[O]. 1 .5) and (·1. with magnitude larger than 1 . We compute 4. . . we can readily obtain. I . has index 3 and has index. then every zero input response will approach zero as > 00. The MATLAB function ds t ep computes unitstep responses of discretetime statespace equations. These terms are dictated by the eigenvalues and their indices. If has an eigenvalue. In other words. then the four matrices must be computed from (4. the response can be computed recursively from the equations. a n acronym for discrete linear simulation. It consists of one capacitor. Ak k k).I) Al The MATLAB function [ ad . . I+Ar+A22Tr2 + . k A 4. and then uses ds t ep. Internally. The function d l s irn.9) into the discretetime state equation in (4.3 Consider the network shown in Fig. I f every eigenvalue. using (3.3 Equivalent State Equations 93 Proceeding forward. If a zeroinput response if has eigenvalue I with index then contains the term then it will grow unbounded as > 00. If the function s t ep is applied to a continuoustime transfer function. . contains the terms ( I . 1 7) response. 1 9) are algebraic equations. For example. bd ] = c 2 d ( a .I (e''\T  (if i s nonsingular) (4. 3' I = . 20) L kI (4.9) at t = if the input is piecewise constant. . one resistor. .5 1 ). Then w e have . = = = = = L kI > 0. 4. and one voltage source. We discuss a general property of the zeroinput response Suppose has eigen value with multiplicity 4 and eigenvalue with multiplicity I and suppose its Jordan form is as shown i n the second matrix i n (3. simple or repeated. it first uses the function c 2 d to transform a continuoustime state equation into a discretetime equation and then carries out the computation. It also computes unitstep responses of discrete transfer functions. Their derivations are considerably simpler than the continuoustime case.92 STATESPACE SOLUTIONS AND REALIZATIONS with This is a discretetime statespace equation. Ak kA�Ak1 k(k kA�A71 Ak Ak l] A' l� A� k AA I k. If then the series can be written as. 1 5 ) . it first transforms the transfer function into a discretetime statespace equation by calling t f 2 s s . we have This power series can be computed recursively as in computing (3.. If has some eigenvalues with magnitude I and all with index and the remaining eigenvalues all have magnitudes less than I . which will be discussed later. + I ) A I) A {T o J = ( kT (4. It is understood that if the equation is obtained from a continuoustime equation. 4. 2 1 ) A A2 2 /2 0 0 = Q I 0 I 0 0 0 0 0 0 I Q ' (A�. kA71 . T ) transforms the continuoustime state equation in (4. then most zeroinput responses will grow unbounded as > 00. Using (3. Once and = 0. In order to discuss the general behavior of discretetime state equations. then no zeroinput response will grow unbounded. A T22' T3 A (TA + A2 + A3 + .5 1 ). we will develop a general form of solutions.7). EXAMPLE 4. are given. which computes responses of continuoustime state equations or transfer functions excited by any input. k2A�2. I  Thus we have Using this formula. simple or repeated. 1 8) They are the discrete counterparts o f (4. computes responses excited by any input. The function s t ep computes unitstep responses of continuoustime statespace equations.) dr T22! T3 T4 TI+ A+ _A2+ _A3 + . (4.2.
25) Cp. The voltage across the inductor is and the current through the capacitor is The voltage across the resistor is thus its current is Clearly we have and Thus the network is described by the following state equation: y If. Consider again the network shown in Fig.I X(t) . q2 .1 XI = XI +xz +U X xz.I ) = det[P(U .23). In fact. In fact.A). instead. = de t (P) det(U  A) = det(App1 (4. XI xz=.24) are preserved or invariant under any equivalence transformation.X2. all properties of (4. . we have.94 STATESPACE SOLUTIONS AND REALIZATIONS capacitor voltage as state variables as shown.23) [l direct transmission part between the input and output. G (s) = C(sI  = D. X2. The state x is a vector in R. Equation (4. Figure 4.27) . From the equation B = PB or B = pI B = [q l ql .I X (t) and x(t) = p. therefore they must be closely related.1 == Cx(tt)) ++ Du(tt)) Ax( Bu( yet) x (t) and is not affected by the equivalence transformation. Clearly A and A are similar and A is simply a different representation of A .pI pB +D = = I 6 (A) Network with two different sets of state variables.I B = CPIp(sI .I ) 1. Because the We show that the two equations are equivalent. 1 .22) and (4. " ' .l lx where A is an n x n constant matrix mapping an ndimensional real space R. =Xz xI XI (XI xz) U =XI +XI X2. . .  A) det(p. XI x2/1 = X2. q 2 . consequently. To be precise.I I +u state space from the orthonormal basis to the columns of P1 =: Q. qn } . Now we study the effect on the equation by choosing a different basis. 4.3 Equivalent State Equations 95 equation. the representation of Aqi with respe�t to the basis {ql . its current is xzl i and equals X . as in Equation (3. the same set of eigenvalues and same transfer matrix.I .AlP.1 Let P be an n x n real nonsingular matrix and let x = Px. qn } . let Q = pI = [q l q2 . using det(P) det(p.4.8). From the lefthand and the voltage across the resistor is C transformation. Indeed.7). qn ]B. . which can be described by (4. is said to be (algebraically) equivalent to (4.26) is called an equivalence (4.22) and (4. we have The voltage across the capacitor is the same as the one across the resistor.26) is obtained from (4. 1 . XI X2 X2 =XIX2 = [0 l]x or finition 4. x(1) y (t) where + = = Ax(t) Bu(t) Cx(t) + Du(t) (4. 4. Thus we have I + DD X � __ __ __ +__ __J __ 3 1F l 'lr Xz or [ ] [I ] [ ] [I rl .24) and (4.24) and Thus equivalent state equations have the same characteristic polynomial and. 1 . qn l .22) and are chosen as state variables as shown.24) and x == (4 .r l .I r l pB + l = G (s) C (s I . the basis v�ctors of the Px (4. the loop currents across the inductor is side loop.24) can be considered to be associated with the orthonormal basis in (3. . = . they are equivalent as will be established shortly. has nothing to do with the state space Li (A) = det(U we see that the i th column of B is the representation of the i th column of B with respect to the basis { q I .24) by substituting x(t) = p.A lP." into itself. we have changed.I ) = PAP. which is Thus the current through the capacitor is which equals or Thus the network is also described by the state equation y The state equations in (4. We show that (4.23) describe the same network. Then the ith column of A is. as discussed in Section 3. Consider the ndimensional state equation X2 X2. then the voltage .A)." for all t. we have I voltage across the resistor is X2. XI = X2 + Xz XI + X2 U = O. The matrix C is to be computed from Cp. In this substitution. The matrix called the .] det(AI  A) A)IB + D = CPI [P(sI .25) have the same set of eigenvalues and the same transfer matrix. . thus the real space is also called the state space. The state equation in (4. . . Then the state De 4. From Fig. I 0 I 1 X2 = XI X2 = I 0  J ][ ] XI X2 ( 4.
Figure 4. .[b l compam n f I . . To 0. D + CBS I + CAB s. O. the left upper resistor s current I S u _ i . ' type = c ompanion.i The total voltage around the lower righthand loop i s 0:  I t is clear that (algebraic) equivalence implies zeros tate equivalence. 2 .O for m 2i + x . Thus the two networks.5. This condition is the same as (A.2 MATLAB contains the functIOn [ ab . .26).5it (s). Note tha t Q and J can be obtaIned from [ q . 4. 0 the complex conjugate of q) . .f i r � i Y . I = 0.AQ .e � 9 ( a ) in MATLAB as shown in Examples AI . The P that the function c anOE generates equals Q . c .5(x + u )  (ll (A. then "n . o 0 a+ jf3 0 0 a  J� = I Q . .x . I rl W u . Let . . the right lower resistor s current IS �. 4. we have y(t) = 0. . its inverse happ ens to equa l i tself. . .I F. . 2.5x (t) + 0. d .5u(t) and the transfer function also equals 0. where i (t) denotes the right upper resistor's current. as we WIll dISCUSS In Chapter 6. . : . . " Thus the network ·III F'Ig . For the circuit in Fig. ' type ' . .i: = with A CAm B = = voltage across the capacitor to be zero. 4. If W e assi I: the righ t upper resistor's current as i (t).22) and (4. . Then we have .2(a) is described by the zerodimensional state equa tion y (t) = 0.5. The network in Fig. i j j i .5. I . are zerostate equivalent. A 2 . In order for two s ta te equations to be equivalent. B. and a q� be the correspondIng eigenvectors. Defi ne Q . not the case for zerostate equivalence.i + . .. This is. :i' (t) = x (t) = X + II + X . � �� :.. Because of the symmetry of the four resistors. . B.( I F I rl . .5u (f). + = 0 + CBs . p ) carries out equiv alence transforma tions. + J � ' and a _ I be the eigenvalues and q l .x . the two complex eigenvalues must s complex eigenvalues. . . } . To compute the transfer func tion of the ne twork in Fig.i + x. We see tha t D = 0 = 0.511 (t) � 0: � .23) are related by the equiv alence transforma tion in (4. u (t) or y(s) = 0. capacitance . An. c b . half of the current will go through each resistor or i (t) = 0. Thus its transfer function is 0. and the left ' ' lower resis tor's current is II . vanable x With poI an't y shown. This form is useless in practice but can be transformed Into a real matrix by the .4 Consider the two networks shown in Fig.5 . after subs tituti ng (3. Two zerostate equivalent networks.5.5r1 o Sri (a) t � r t y W u. Q2. Consequently. q).2 + CA2B s3 . ' nonsl flgu ar.��� �:� J := 2 0 (b ) t 3 . I Thus we have the theorem tha t follows . or more precisely their state equa tions. 1 Canonical Forms OI.5 and CAmB = . y(t) = i (t) . .2. The capacitor is assumed to have O. bb . we may assume the initial EXAMPLE 4. . ' J be complex conjuga te. which implies x + (i .3. Two state equa tions are said to be �erostate equivalent if they have the same transfer ma trix or . b .[ q l q_ q' q� J. w ere l ls he fi rst column ofB.5u (t) or A = B = C = 0 and D = 0. and q' are all real and q� equals a R . c b . This fac t can also be verified by using Theorem 4. If S uppose A has two real eigenvalues and two We discuss a dl f a ° nl 1 coefficients. .6. . 4.1 b t l is I (3 "1) This function works only If Q .57 ). they must have the same dimension. Thus the two equa tions are zeros tate equivalent.u = 2x i + x) = 0 y (t) = 0. .96 STATESPACE SOLUTIONS A N D REALIZATIONS Note tha t..4.5(x + u) = 0. . C_ DJ and or have the same transfer matri x if and only if D = D and m = Theorem 4. The MATLAB function [ a b . If v l .(u = i) which implies y = 0 or i I.3 Equivalent State Equations 97 1 This becomes. I t is straigh tforw ard to verify that (4.5 and 3. p ' = c anon ( a . The to tal voltage around the u ppe r righthand loop is 0: 4. .II B = = C = 0. b 1 ) controllable. 4. . OJ are zerostate equivalent I . following similarity transformation . A [AoI � � O� ] .(b) is described by the onedimensional state equa tion 2. . See the discussion in S ec tion 3. . d .5. ! .. where A I . I ts current IS x fl'OWlfl� from ilie ne"ative to positive polari ty g I: . . c . because of the negative capacitance. }" q 1 .x ) =i = 0. > Two linear timeinvariant state equations (A. Such a negative capacitance can be realized using an opamp circuit. .�:� � 1 . however.I + CAB s2 + CA 2 Bs3 + . the function will generate an eqUlvalent tate e tlOn WI th A in the .l . . db] = s s 2 s s ( a . Ig ' develop a state equatIOn for the network lfl F 4 2(b) ' we assign the capacitor vol tage as state .2(b) . J J . bb . for this P. C. b . as the next example shows.2(a). and 0 = 0. dh .
3 . l 2 . if the input magnitude increases by a. If we use ± I Svolt 2.3.2xl 4.1'2 if yet) does not saturate. One way to achieve this is to use an equivalence transfonnation so that a for all i and for all t . for the real part and imaginary part and we have used in the last equality the fact that q4 i s the complex conjugate of q3 . = 20 XI 1 00 = 0. From the plot. We see that I :q Imax = 1 00 > I Y lmax = 20 and IXl l « Iy l max .2 2 0. This is illustrated in the next example. d . d ) with no type specified will yield an equivalent state equation with in modal form. the maximum magnitudes of XI (t) and .5.l J . then all signals are roughly limited to ± 13 volts. A power supplies. ' modal ' ) or canon ( a .5 0. plot ( t . y . d ) .l J . [ y . : JJ n [� � o o 0 0 � A2 o o o 4. We type a� [ .005]x . q2. We give one more example. all signals are limited by power supplies.98 STATESPACE SOLUTIONS AND REALIZATIONS We see that this transfonnation transfonns the complex eigenvalues on the diagonal into a block with the real part of the eigenvalues on the diagonal and the imaginary part on the offdiagonal. 0. O . then the output will not saturate but X I (t) will. so do the magnitudes of all state variables and the output. b . and q4 are. The two transfonnations A can be combined into one as _I [1 .5 0. c . t . c � [ 0 . c . respectively. d� O .f3 a o o Ll �l f3 ' . The MATLAB function t [ ab . however. x .2 Magnitude Scaling in OpAmp Circuits P= every LTI state equation can be implemented using an opamp As discussed in Section circuit. cb .005 It (t) 0 ] Then its equivalent state equation can readily be computed from (4. therefore we can use the original state equation to find the input range to achieve l y (t ) 1 ::: M. i s = [ q l q2 R e (q3) Im (q3 ) ] A = � A[ ' ! [' � 0 I 0 0 0 0 0. respectively.5 Consider the state equation P = QQ =  [ql qz q3 q4 ] where Re and 1m stand. the circuit will saturate and will not behave as the state equation dictates. The modal fonn of a matrix with real eigenvalue A I and two pairs of distinct complex conjugate eigenvalues ai ± jf3i .O . we see that if i ll (t ) 1 ::: 0.� ] u  I ]x Suppose the input is a step function of various magnitude and the equation is to be implemented using an opamp circuit in which all signals must be limited to ± I . Note that there is no need to transfonn A into a diagonal fonn and then to a modal fonn.1. Clearly it is desirable to have the admissible input magnitude as large as possible.3(a) . This subsection may be skipped without loss of continuity_ y = [I .26) as X = [ o � 0. and a2 + jf32 . The equivalence transfonnation. [ o �] 0. 0' I + j f31 .1 Let us introduce new state variables as will equal l y Imax · Thus With this transfonnation. 2 .5 .3 Equivalent State Equations 99 a + jf3 o o o o a 0. EXAMPLE 4. Consider an LTI state equation and suppose all signals must be limited to ± M . 1 p_1 _  [5 0 0. t l � s t ep ( a . will not alter the relationship between the input and output. b� [ 1 0 .0. b . In addition. l l . x) O 0 f3 1 al al f3 1 0 0 0 0 0 0 a!  0 0 f3z az It is block diagonal and can be obtained by the similarity transfonnation pI [ql R e ( qz ) Im( q2) Re ( q4) Im( q4l ] where ql . 2. we can use the same transfonnation to amplify some state variables to increase visibility or accuracy. For linear systems. Therefore saturation is an important issue in actual opamp circuit implementation. First we use MATLAB to find its unitstep response. eigenvectors associated with This fonn is useful in statespace design. 0 . If any signal goes outside the range. This new Amatrix is sllid to be in modal fonn. for i = 1 .· [ � �1 ] X+ [ �. 1 I x= . bb .ii ' The transfonnation can be expressed as x = Px with XI 1 . The state variable X2 is hardly visible and its largest magnitude is found to be by plotting it separately (not shown) . db .5j O�5 1 j y = [O. neither will all the state variables . Thus there must be a limit on input magnitude. c . 4.002 I ] [ ] x + _ 2 20 1. l In actual opamp circuits. P l � c anon ( a . b .28 ) which yields the plot in Fig.5  0 0. 11 A (4.
I ) . by the statespace equation description X U ) = Ax(t) + Bu(t) (4. Before the advent of digital computers.t � 4 . but not by a finitedimensional state equation.30 ) G(ce) = D Next we show the converse. We see that all signals lie inside the range ± 10 and occupy the full scale.A ) : thus it has at most degree ( n . + U r _ 1 S + Ur (4.2 Figure ". Now we study the converse problem. D) i s called a reali�atioll o f G(s ) .A) .I ) submatrix of (. its leading coefficient i s 1 .u 2 Ip Ip 0 0 ur_ I lp 0 0 O I" � o G(s) = C (s l .I B + D y= [i'il N c .29) be the least common denominator of all entries of Gsp (s). Note that we have Gsp (5) : = C Cs I . 4.A ) . cS . If the state equation is known. A transfer matrix G(s) i s said to be reakahle if there exists a finitedimensional state equation (4. not necessarily of the same dimension. The magnitude scaling i s important in using opamp circuits to implement or simulate continuoustime systems. that is. 4 Realizations Every linear timeinvariant (LTI) system can be described by the inputoutput description d es ) = sr + U I Sr .29) or. This shows that i f G (s) is realizable. the magnitude scaling can now be carried out easily.l B is a strictly proper rational matrix.II .A ).4 Rea l i z ations 101 1 00 80 60 40 20 10 il 6 V and (A.I) x (n . the transfer matrix can be computed as G(s) = Cis I A)I B+D. Thus the equivalence state equation is better for opamp circuit implementation or simulation. I f D is a nonzero matrix.[ N i s .32) Yes ) = G(s)ii(s) and. C.1 Nr 1 x + G(oo)u 0 Ip 0 1n x+ 0 0 u (4.3 1 ) 4 . that is.34) . N N r' Gsp (5 ) = . .alizable. With the help of digital computer simulation. If G(s) i s realizable. are q x p constant matrices. Now w e claim that t h e set o f equations 0 . then there exists a realization. If A is n x n.1 + . B . · Nr . that is.3 Time responses. First we decompose G(s) as where Gsp is t h e strictly proper p a rt of G (s ) . the idea is applicable to any input. then i t i s a proper rational matrix. Every entry of Adj (51 .[N (s ) ] = . A n LTI distributed system c a n b e de �cribed bv a transfer matrix.+ ' " + 1 r . . Thus the realization problem i s fairly complex. if G (s) is a q x p proper rational matrix. then it has infinitely many realizations.5 are plotted in Fig. Their linear combinations again have at most degree (n .A)]B (4. Let G(s) = G( ce ) + Gsp (s) (4. by using the state equation.A) I qAdj (51 . 1. Thus not every G(s) i s r. Thus we conclude that C(sl . This terminology is justitied by the fact that. The computed transfer matrix i s unique. B.I B = det(sl . to find a statespace equation from a given transfer matrix. We use (3. 1 9 ) to write (al Its step responses due to u (t ) = 0.I S drs) d rs) . Although we discuss only step inputs. then C (sl . . (A. magnitude scaling in analog computer simulation was carried out by trial and error.1 00 STATESPACE SOLUTIONS AND REALIZATIONS x.I + . Y X2 IO 20 30 40 IO 20 (b) ji'� 30 40 Theorem 4. D) such that X = [ r 1 (4 33) Ip 0 .I B + D is proper. simply. We mention that analog computers are essentially opamp circuits. " where N. we can build an opamp circuit for the transfer matrix. C. The other issues will be studied in later chapters. Then Gsp (s) can be expressed as y(t) = Cx(t) + Du(t) 1 I r + '" .3(b).A) has degree n. This is called the realbuion problem.A) . A transfer matrix G(s) is realizable if and only if G(s) is a proper rational matrix. We study here only the realizability condition. if the system is lumped as well. then det (51 . .A) i s the determinant of an (n .I ) . Here we require d (s ) to be monic.
Thus we have and a realization of (4.35) (4. The Amatrix is said to be in block companion form. 5 1 .sp (s) is des) = (s + 0 .l B + sZ = AZ + B z = is a realization of (.33).35) as (sl .4 Real izations 1 03 (s + 2)2 s+2 s+ I 3 des) I ([ 3 ] [  0 1 [ Substituting these into the first block of equations in (4.5) s+ [� �J [ 4s .5 7. each of order q x p. 1 EXAMPLE 4.(00) [�: j . . Theruthe transfer matrix o f (4. + NrZr + (.8 1 1 s + . the Cmatrix has order q x rp.5) 24 I 05 0 ] (4.34) in which p = I . We show that (4.B + G(oo) = .8l 3 S + . Substituting these into (4.5 3 (s + 2 ) (s + 0. 5 ) (s + 2)2 = s3 + 4.des) ' (s + 0! 1 + 0! 2 + .36) matrix (. + Nr l + G(OO ) des) sr .5s 2 + 6s + 2. Consider the x I proper rational matrix This is a sixdimensional realization.3 1 ) and (4.5 0 0 0 1 Gsp (s ) = 3 s + 4. + l ) Zl = des) Z l = Ip O!r s S rSr .10 2s + 1 l [ (2s + I ) (s + 2) 2s (s + 2) 2 12 tl 3 __ s+2 s+I ] 4.5 0 0 4.' B (4. + .A) .37) yields or.I We discuss a special case of (4.37) . [ J [ : . we can readily obtain We write (4. .(s) = = Using the shifting property of the companion form of A.812 s + ..38) 24 0 3 0 X+ 2 J 0 X+ 0 0 0 0 0 ] � ]) 0 0 0 0 0 [ :J (4. it consists of r rows and r columns of p x p matrices. (s).39) [� � J [ :: J (.(s) in (4.814 . (s) . This shows that (4. Y= [ 6 0 0 0 3 0 24 0.5s 2 + 6s + 2 A where we have decomposed (.[N 1 S r .38) is X = 4. thus the Amatrix has order rp x rp.sp (s).32).34) is a realization of (.34) is a realization of (. The matrix Ip is the p x p unit matrix and every 0 is a p x p zero matrix.+ Nzs r . The monic least common denominator of (.5 2 0 0 0 7. The Bmatrix has order rp x p.. Because the Cmatrix consists of r number of N..A).l Zl .8Z2 S " + . = (d . However. To save space.5 1 . (s) in (4.(s) = . we assume r = 4 and q = 2. . .8Z1 s + . .3 1 ) and (4. 3 1 ) and (4.824 d1 dz + S 4 + 0! 1 s3 + O!zs z + 0!3S + 0! 4 _ � _ _ _ _ _ _ I � ] (4.5( s + 2) 24 0. which implies Thus we have This equals (.(00) = N 1 Z l + NzZz + .823 s + .34) equals C(sl . from the second to the last block of equations in (4. I l ' C(sl .36) yields l .40) . .5 0 0 0 6 6 0 0 0 0 2 0 0 0 0 ] [ (s + l ) (s + 0. (s) into the sum of a constant matrix and a strictly proper rational = 6 0 2 s + 6(s + 2) 2 0.A)Z = B or (.33).37) (4.A) . using (4.1p .6 Consider the proper rational matrix (2s + I ) (s + 2) + (. . the discussion applies to any positive integers r and q . Let us define where Z j i s p x p and Z i s rp x p .1 02 STATESPACE SOLUTIONS AND REALIZATIONS The realization has dimension rp and is said to be in controllable canonical form.
= : Ycl (5) + Yds) + . the function t f 2 s s can generate the fol lowing realization for the second column 12 ] XI + [ 2 ] I I 0 0.39) and (4. a different realization of = 1Ii be the ith component of the input vector (4. 158160]. For example.38).. Thus we can realize each column of G (s) and then combine them to yield a realization of G(s). (4.4 1) Typing (4.. The two state equations in (4.3I 6I ] xo. C . This realization has dimension 4. 44(b).BII . two less than the one i n (4.3 1). a2 . s. G (s ) canonicalform realization shown i n Figure 4A Realizations columns and by ro". I a. In its employment.4 1) for any singleinput multipleoutput transfer matrix .+ [ 0 ] lh.1 04 STATESPACE SOLUTIONS AND REALIZATIONS Then i t s realization can be obtained directly from y= X= 0 0 .. Clearly we can also realize each entry of G(s ) and then combine them [6.38) and then combine them to form a realization of G(s).+ b.5 ( 4.44) 0 6 12 3 � ] x + [� �] u y= [ 0 0. . These two realizations can be combined as Aox. d 1 t f 2 ss ( num .. \'� = or 0 X [ Xl ] = [AI A2 ] [ X2I ] + [ b0l b02 ] [ IIII. . .9 yields the following realization for the first column of G ( S ) : J(s ) as shown i n Fig. Then y( s ) = G (s )u(s) can be expressed as Ge l (5)11 1 (5 ) 4. .' [: [ J J [. den ) generates the controllable a realization of G(s). however..42) Gn (s)u(s) as shown i n Fig. not necessarily monic. b ..B f3i3 f31 . Then J(s) = G(s)u(s) can be + GdS)lds) + . of G(s): Similarly. The next example will apply t f 2 s s to each column of G(s) in (4. Yi (S ) = 4.= C . 4 4) are zerostate equivalent because they have the same transfer matrix.= [ I4 o4 ] x'. 4 Real izations l OS x+ 0 0 12 dl . More will be said x= [ T �l{ lJ' I 0 2 . But we must compute its least common denominator. Problem and let ex pressed as gi\ e<.33) with dimension rq. Let Gri ( 5 ) be the i th row of G(s) and let Yi be the ith component of the output vector y . Thus we can realize each row of G (s) and then combine them to obtain to obtain a realization of G ( s ) .4(a). )(0 =  (443) (4. ] X + [ d2 ] 11 0 I 0 0' 3 0 0 11 (4.B21 .5 This is a different realization of the G(s ) i n (4. ] �I I 0 = Yc1 + Y = [CI C21x + [dl d 2 ]u Y c of G(5) by ( b) 0 0 0 4 (4.40). They are. The MATLAB function [ a . there is no need to decompose G (s) as in (4. Let Gei (s) be the i th column of u.+ [ 0I ] I/o. G(s ) This controllablecanonicalform realization can be read out from the coefficients of G(s) in There are many ways to realize a proper transfer matrix.uo.+ d'lh = [ 0 v.B22 f323 f32. See Reference pp.39).x. . not algebraical ly equi valent.34) as o o 4 .
is not valid because f f.5t = 2. XCt) in (4. xn] .5 t2 + 2 • with. Consider where A is n x n with continuous functions of t as its entries. (4. We argue intuitively why this is the case. then X(t) is called a fundamental matrix of (4. Because every Xi satisfies (4. Because the initial states can arbitrarily be chosen. we have 4 . [ ] t 0 0 0 x(t) (4.46) y(t) = C(t)x(t) + D(t)u(t) X(!) = A(t)X(t) (4. we have AB i= BA and e(A +B) t i= eAt eRt . . Note that. a . thus +. then there exists a nonzero vector v such that X (t l ) := X (t I lv = 0.49) I. i n general. including t f 2 s s . 5 Solution of Linear TimeVaryin g (LTV) Equations Consider the linear timevarying (LTV) state equation x (t) = A(t)x(t) + x = A(t)x (4. X(t) = [ I? 0. A sufficient condition forisuch an assumption is that every entry of A (t) is a continuous function of t. for every initial state f (to) and any input uct). applies without any modification to the discretetime case 4.45) (4. we cannot extend the A very important property of the fundamental matrix is that X (t) is nonsingular for all t . the solution of x = Ax is x(t) = eAt x(O) with If X(to) is nonsingular or the n initial states are linearly independent.8 Consider the homogeneous equation where the commutative property is crucial. A x(O) (4.+ 2 ! A(t) + . Thus X(I) is nonsingular for all I . in particular. We can arrange these n solutions as X = [X I X2 .5t 2 0.1 06 STATESPACE SOLUTIONS AND REALIZAT I O N S i n Chapter 7 regarding realizations. using (3. Similarly.. n.5t 0. 5 2 ) A ( r )dr = A(t) i= A(t )e f. thus it is nonsingular for all t. . ef. . in turn.48) Thus. This is a contradiction. the solution of X2(t) = tX I (t) = tX I (O) with x dt x Thus we have d f' a(r)dr fo' a(r)dr f' a(r)dr a (t) . . " ' . Before considering the general case.49). .0. e l x(t) = e f. 5 t 2 + 2 . The solution of the timeinvariant equation x = Ax can be extended from the scalar equation x = ax. A +� A ( r )dr + A(t) (r)dr (l � (l A(r)dr + A(S)dS ) � (l ) (l A(S)dS A(r)dr ) ) The two initial states are linearly independent. a square matrix of order n . The solution of the scalar timevarying equation x = a (t)x due to x (O) is f' a(r ) dr (0) ( t ) = eJo � eAt = AeAt = eAt A dt or x (t ) = . at = to.5 1 ). In conclusion. The solution of .5 Solution of Linear TimeVarying (LTV) Equations 1 07 solution of scalar timevarying equations to the matrix case and must use a different approach to develop the solution. there exists a unique solution Xi (t).49). is a fundamental matrix. we first discuss the solutions of x(t) = A(t)x(t) and the reasons why the approach taken in the timeinvariant case cannot be used here.5t. the fundamental matrix is not unique.1: = ax is x (t) = ea.eJ o = eJI = a ( t ) eJo Extending this to the matrix case becomes and (r)dr x(O) = [ X I (0) X2 (0) ] [I] = 0 =} x(t) = [ I ] 0. the state equation has a unique solution. (4. for i = 2. B(t)u(t) (4. which.5 1 ) The solution of XI (t) = o for to = 0 is is x I (t) = X I (O) .52) has determinant 0 . A ( r)dr = 1+ This extension. We mention that a l l discussion i n this section. as long as they are linearly independent.50) It is assumed that.47) x(O) = [xICO) ] [ I ] X2 (0) = 2 =} x(t) = [ I ] ] (4 .47) is not a solution of x = A(t)x. x (O) with d(ea')/dt = aea' = ea .. i n general. Then for every initial state Xi (to). 2 For example. EXAMPLE 4. If X Ct ) is singular at some II . however. implies xCt) := X(t)v == 0 for all t.
60) (4. to)Xo + 1'0 ( � <I>(t.6 1) (4. to)Xa (4. state transition matrix x= = I.III u(r) dr = '0 = (4.+ If the input i s identically zero.5I t. to)Xo l ' A(t)<I>(t. t) =I = = = (4. r)B(r)u(r)dr f' [ = <I>(t. the zerostate response can be described by and has the three yet) Cit) 1'0 <I>(t. to) <I> r). We show that by using = state equaion.state response.55 ] 0.9 Consider the homogeneous equation i n Example was computed as <I>. r)B(r)u(r) dr = d X(t) = . r)B(r)u(r) dr D(t)u(t) f' + A(t)x(t) B(t)u(t) (4. tl' EXAMPLE and Its inverse is.57) (4. I) y(t) = l ' GU. G(t. Comparing and yields = = G(t.45) x + + excited by the initial state x (t) = <I>(t.I (t. then Equation y (t) C(t)<I>(r.<1>(1.5t2 0.57) <I>(t.53) (4.l ' <I>(t.I Ct) <I>(to. to) at with the initial condition <I> (ta. (4. f)B(r> m( t . to) <I>(t. we have the following important properties of the .57) i s the solution.56) Thus 4. x(to) <I>(to.5t'I 2 ] [0.57) reduces to This is the zero. (4. Equation follows from satisfies the initial condition and the (4. r) Thus (4. At = we have <I>(t.59) = + .5 t. Substituting = + = '0 (4.54) (4.59) + + (4. dt = = = satisfies the initial condition. As discussed in (2.5 Solution of Li near TimeVarying (LTV) Equations 1 09 where is the state transition matrix of x = (10.1 08 STATESPACE SOLUTIONS AND REALIZATIONS De finition 4.t5 ) �] .57) (4. 0. t)B(t)u(t) f' ot A(t)<I>(t. t to. = Let X(t) be any fundamental matrix of A(t)x. til<l>(tl.5). rlB(r)u(r)dr A(t)<I>(t. using 4.57) (4.53) (4. r) C(t)<I>(t. I)B(r) D(t)8(t .I) CU)X(t)X . to)Xo i' <I>(t.0. Equation follows directly from From the definition.56). to) := X(t)X. r)B(r)u(r) dr ] x(to) Xo = and the i s the impulse response matrix and is the output at time where input applied at time T.f) + (4 62) . Using for every t.53) .6 1) + t excited b y an impulse (4.5 ) . ' state transition matrix: of A(t)x. r)B(r») dr + <I>(t.20).25t.S.I (r)B(f) D(t)8(t . I)B(r)u(r) dr D(t)u(t) f' l' [C(t)<I>(t. . to) o l' <I>(to.54) Now we claim that the solution of input is given by u(t) (4. to.0.5S) This relates the inputoutput and statespace descriptions. a a at at to)Xo .I 2 ] = . f)B(r ) U( r) dr B(t )u(t) AU) [<I>(t. to) Xa 1'0 <I>(t.I 0.25to. Thus the state transition matrix governs the unforced propa gation of the state vector.<1> (1. Then <I>(t.57) = Ix0 + 0 = Xo + '0 + '0 + + • (3.55) (4.5S) (4. Thus the state transition matrix i s given by = It is straightforward to verify that this transition matrix satisfies properties listed in through I 25t� <I> (t.2 is calledsolution of unique the Because i s nonsingular for all its inverse i s well defined. Its fundamental matrix (4.50). to) = <I>(t. to)xo 1'0 <I>(t. to) A(t)<I>(t. to) = [ 0.60) (4. ( + 25t2 X _ I (t) [0. then reduces to I x(t) = <I>( t . to)xo CU) }to <I>(t. to) [X(t)X .0. r)u(r) dI '0 (4.46) + into yields (4. r) <I>(t. t) <I>(t. If the initial state is zero. The state transition matrix is also the a (4. r)B(r)u(r) dr ] + B(t)u(t) and we have + A(t)x.6).1 (to) x= 4. X(t) [ 0.53) (4. to) X(t) t.I (tOW I X(to)X. + This is the zeroinput response.055 ] [ 0.
67» reduces to ' m=ko k + (4.m] 8[k . ko] = A(k . The situation here for and the input the initial state case. case. has the commutative diagonal or cons such as what we did in Example 4. k < m.53 <I>(t. thus the m verse of ' ' 0 1 may not b e defined. [k..5.66) A(t)x B(t)u y C(t)x D(t)u = x= + + (4. G[k. k ko.3 to the lImevarying . _1) and wIll not be repeated. =I 4. Thus . k "" constant.49) or The solutions in (4.6 Equivalent TimeVarying Equations . k I l<l>[kj• kol = + t to t < to <I> [k k III = T) I . Beca use the <I>(ko. Equation (. ') ' If the initial state is zero. to) = e '0 J. ition matrices is ial cases. m] = 0 for Comparing this with + for This relates the impulse response sequence and the state equation and is the discrete counterpart of (4.m]  = m=ko L (C(k]<I> [k.59 such as A(t) is triangular we can solv e the scalar equation Xl (t) [ Xl(t)) } = [ all (t) a22O(t) } [ X2 (t) } X2(t a21 (t) Xl (t) = a ll (t)XI (t) A(t) . 4.65) ' as... ko]xo + L <I>[k.and and can govern the .56) or . the Amatrix can be singul ar.63) reduces to = eA' . poslllvetlme direction. . m I]B[m] D[m]8[k .63) \ k ko. m] C(k]<I>[k. This section extends the equivalent state eq uations d' lscussed m Secl10n 4. we can express the solutions of (4. kol I . > l y[k] C(k] L <I>[k. <I>(t.62). and generally difficult. kol = A(k]<I>(k. . kol <I>[k. > for all the state transition matrix is defined for .and (4 .time d' IrectlOns . ko] .64) and (4. . simp ler than the continuoustime transition matrix as the can define the disc rete state in the continuoustime case. ')0 .. Their derivations are similar to those of (4 .65) where the impulse sequence equals 1 if the multivariable version of (2.69) . kol <I>(k + 1 . ko <I>(k .. propagatIOn of the state vector in the POSIt' ' lvet' lme and negative .. for = obtained directly as inuous. If ) hinge on solving (4.67) : i i for k ko.57) and (4. we As solution of x(kol x(k A(klx(kl B(klu(kl y(kl = C(klx(k] D(klu(kl u(kl... Therefore the d'lscrete counterpart of (4. with .. kO ] IS and then substitute it into This is can be solv ed for the preceding scalar equation Because Xl (t) has been solv ed. then the solu tion of (4..6 Equivalent TimeVarying Equations t.34).53). l1me case.64) (4. m + l]B[m] + D[m]8[k . r) = e A (rr) k =O 1 (1' '0 = <I>(t  A (T) d T )k (4. tant. + 11 = + (4. for > for all to ) can be shown to be and t.68) can be written as m=ko k + (4. 1 DiscreteTime Case y(k] k ::: m. Consider the ndimensional J mear tlmevarymg state equation ' can be rt of (4. If we define then (4. In the discrete. .1 10 REALI ZAT I O N S STATE SPACE SOLU TIONS AND (4.and difference betw een the cont = We discuss a significant and for > me case is nonsingular inuousti fundamental mat rix in the cont discretetime cases. we have = + k m k 'I m .53) and its solution This is the discrete counterpa (4.68) 4.21 · A(kol k ko <I>(ko. (4.8. m + I]B[m]u[m] D[k]u[k] = This describes the zerostate response 0 f (4. x[k] <I>[k. . + + I..' Alr) dr For A(t) X(t) = L 1. m l]B[m]u[m] + D[k]u[k] = m=ko + kI k ::: kl ::: ko· k k <I>[k.llA(k . computing state trans Other than the preceding spec <I>(t . . ko]xo C(kl LI <I>[k.65). If property X2(t )· holds only for Using the discrete state transition mat nx. equation Consider the discretetime state be computed recursively once tions and their solutions can is muc h The set consists of algebraic equa are give n. A(t) .. m I]B[mlu[m] y[k] C(k]<I>[k.m] ) u[m] 'f and O 1 k <I>[k. to) . k ::: ko. . > defi ned only for _ 0 and governs the propagation of the s tate vector m only the .
. Let x = P(t )x. X(I) = A ( t) X (t) and X(t) i s nonsingular for all t. Then we claim that Dill If Ao is chosen as a zero matrix.70).I (t)P(t)]X(t) = A ( t ) [P(t)X(tl ] u Theorem 4. Proof' Let X U ) be a fundamental matrix of x = A(t)x.70) reduces to i s a fundamental matrix o f (4. whiG. the matrix P(t)X(t) is also nonsingular for all t. Equation (4.69).73) Figure 4.I (t)B(t) C(t) = C(t)X(t) O(t) = D(t) (.70) is.I (t) + eAo' X . Following (4.112 STATESPACE SOLUTIONS AND REALIZATIONS Let PU) be an n x n matrix.69) into (4.D.T ) Thus P(t)X(t) is a fundamental matrix of x (t) = A (t)x(t).70) with A(t) = Ao. E. . + Let X be a fundamental matrix of (4. The impulse response matrix of (4.I (t) C(t) = C ( t ) p . r) = C(t)X(t)X\r)B(r) + 0 ( t) 8 ( t . in order to do so.7 \ ). Every timevarying state equation can be transformed into such a block diagram. B y definition.3 Let Ao be an arbitrary constant matrix.I (t) O(t) = B(t) = P(t)B(t) This establishes the theorem.69) is given in (4.7 1 ) The block diagrams of (4.I (I) = [eA"' XI (t)A(t) + AoeAO' X .Ao' = Ao Q .l et ) and (4. = [P (t) + P(tlA(t)][P .62). then P(t) = X .70) is obtained from (4. Then there exists an equivalence transformation that transfomls (4. X (t) = eAQ' x = A (t ) x = Aox . Then the state equation and compute 4.69) by substituting x = P(t)x and x = P(t)x P(t)X.I (t)Xul + X . The impulse response matrix of (4.I U ) ]X (t)e AQ' A (t ) = AoeAQ'X1 (t)X(t)e . A(t) = 0 B (t) = X.69) and pet) is called an (a/gebraic) equim/ellce transformation. The differentiation of XI (t) X U ) = I yields X . we define (4.5 (b ) Block daigrams with feedback and without feedback. Because the rank of a matrix will not change by multiplying a nonsingular matrix.b becomes. . 0 and A(t) = 0 are plotted in Fig.5. using (4. 4.70) A ( t ) = [P( t l A ( t ) + p(t)]p. after substituting (4.7 1 ) and (4. A(t) = [P(t)A(t) + p (t)]p. However.6 Equivalent TimeVarying Equations 1 13 x where y = C ( t ) x + O(t)u = A(t)x + B(t)u (4. d [P(t)X(t) ] = P(t)X(t) + P(t)X ( t ) = P ( t ) X (t) + P(t)A(t)X(t) Tt G(t.U4) . The block diagram with A(t) = 0 has no feedback and is considerably simpler. Indeed.72).72) is a fundamental matrix of u which implies Because AU) = Ao is a constant matrix. we must know its fundamental matrix.69) with A(t) of. Now we show that P(t)X(t) satisfies the equation x = A(t)x. we have XII) := P(t)X(t) (4. is said to be (algebraically) equivalent to (4.72). It is assumed that pet) is nonsingular and both P(t) and P(t) are continuous for all t.I X (t ) = (a) 0 (4.
79 ). n x p. In other words. Furthennore. of x = A(t)x . the = A(t) x(t) + B (t ) u (t ) C(t)x(t) + D(t)u(t) (4.I (t)] Aq Theorem 4. em IS to find a state equatIOn from a given impulse response matrix ..I (O)X ( T ) .::: M(t )N(r ) + D(t) o (t .AT X .r ) for t ::: r (4. lza C(t) .3 A matrix P(t) is called � Lyapunov transfonnation ifP(r) is nonsingular. Le t A be the constant matrix comp _ uted from eA T . Thus the equivalence transfonnation in the timeinvariant case is not a special case of the timevarying case.24). r ) u ( r) dr + T) and. � . be descn bed by the inputoutput descr iption y(t) = = Periodic state equations assumed that Consider the linear timevarying state equation in (4. ant syste ms. Lyapun0v eqUivaI ent to x = Ai. For this Q there exists a constant matrix A such that /\. The matrix pet) in (4.1 (t) = per ) eA (t+T)X. this is true if A(t) is periodic. however. D (t ) l to meet (4. > 0 . The rest of the theor em follows directly from Theore 3' e o g n u part of Theorem 4.4 Consider (4. r) can be decomposed as G(t.A(t)x . In thiS section. (4. may not be preserved in equivalence transfonnations. and q x p matri ces for some .I (r)B (r) + D (t ) o (t . If pet) is required to be a Lyapunov transfonnation.77 ) . Thus the (algebraic) transfonnation in the timeinvariant case is a special case of the Lyapunov transfonnation.I (t + T) = eAl i T [ e. we study the corresponding � problem for linea r ti v s s a ce transtonn cannot be used here. all conditions in Definition 4 3· th us 't a Lyapunov transfonnatlOn.7 TimeVarying Realizations We studied in Section 4. It is A(t T) = for all t and for some positive constant T . It states that if X = A (t)x a d if A(t + for all t.75) can be written as X(t + T) = X(t)e A T (4.69) and (4.3. Then (4. r ) = C(t)X (t)X. as shown in Theorem 4. Furthennore. impulse respo nse matr ix can be comp ' (4. G(t. where M..80) q x n . The property of the Amatrix.r ) r.I (t )eAI �here _ l .r ) C(tlP.7 TimeVarying Realizations 1 15 . It is a constant nonsingular matrix. N. not every timevarying state equation can be Lyapunov equivalent to a state equation with a constant Amatrix. ation problem for linear time mvan . � p (t) IS a penodlc function. T = Q (Problem 3. pet) and p et) are continuous.r) 4.5 x p impulse response matrix G(t.70) are said to be Lyapunov equivalent ifP(t) is a Lyapunov transformatioll. 4.7 8) Define uted from G(t. That is. r) = : . It is clear that if P(t) = P is a constant matrix.X . by the state equation x (t) yet) If the state equation is available. " ' �4 � � � �� : �� .79) where X(t) is a fundamental matrix .1 (t)P(t)X(t)X.76) XU + T) = X(t)X.. and D are. . Let us define Q = X. = C (t)X(t)X1 (r)B(r) + D(t)o(r . Then we have X (t + T) Thus XU + + A(t) �� 2e��f � :�:� �� ��� T ) is also a fundamental matrix. Let X(t) be a fundamental matrix of X = A (t)x or X (t) = A(t)X(t) with X(O) nonsingular.1 14 STATESPACE SOLUTIONS A N D REALIZATIONS = Thus the impulse response matrix is invariant under any equivalence transfonnation.I (t) are boundedforall t.I (r). into a constant or a zero matrix. it can be expressed as (4. However.4 is the theory o Floq f uet. . yi m o . then it is a Lyapunov transfonnation.69) is Lyapunov equivalent to Theorem 4.I (O)X(T) = A(t + T)X(t + T) = A(t)X(t 11 10 G(t. . E very lmear tImevarymg system can . r) is realizable i f and only if G(t .. �r � e. An ' Impu Ise response matrix G(t r) is said t0 be rear b Ie if there exists (ACt) . The converse probl .69) with of X � x (t) A where pet) = e I X. For example. We show that pet) is periodic with period T: pet + T) = = eAI X. equivalence transfonnations will preserve all properties of the original state equation. Thus (4.I (t)i( t) + D(t) u(t) Definition 4.69). if the system is lumped as well. every Amatrix can be transfonned. x = A(t)x is .77 ) satisfies . and P (t) h ndp. A(t) is periodic with period T.75 ) This can be verified by direct substitution. In the timeinvariant case. therefo re we study the prob ' .l (r)P(r)B(r ) = + D(t)o (t . yet) = = Ai(t) + P(t ) B(t) u(t) C(tl P. Let X(t) be a fund amental matnx . Equations (4. Clearly the zero matrix does not have any property of A(t).4 the realiz . respectively.I (r )p. then Theorem 4. r) U) = A(t + T) for all t and some T .3 does not hold in general. then Its fundamental matr ix is of the fonn p.I (O) X ( T ) . . B(t).
6 all signals are required t o lie inside ± 1 0 volts and if the input is a step function with x = [� The Laplace transform of the impulse response is g (5) _ Using (4.10 Consider get) = teA' or 4.This realization is clearly more desirable because it can readily be implemented using an opamp circuit..5 can also be applied to timeinvariant systems.5 y = [I [ � � lJ [ � ] I 2 x + u O]x Find an equivalent state equation of the equation in Problem 4. 0  x 4. This LTI state equation is a different realization of the same impulse response.D. (5 .I ( r ) B ( r ) establishes the necessary pan of the theorem. Verify that the equation can be trans n x + [ �: ] 1I x(tl = [ yet) = [0 . 4.�.1 Problems 1 17 An oscillation can be generated by x= [ Show that its solution is 0 1 0 (4. I f x(t) = [ � �] x + y(t) = leA' te" ]x(t) is a realization of the impulse response g(t) = [ :�=" ] u ( t) (4. Indeed.4 Discretize the state equation in Problem 4.r ) ewn It is straightforward to verify get .2ies + ).r) = (t .4 so that all state variables magnitude a . there exists a realization that meets (4..E.SIn t COS t . y = [2 3]x Find the companionform and modalform equivalent equations of 2 x= get .3 xU) = [ .ie)� I 52 . we can readily obtain 'c[te" ] _ ? _ 1_ _ _  teie.2 4. then the . Verify that the Jordanform equation .2 formed into with x = Ax + bll = Qx with y = ex A= [ (4. x= [ 0 2 2 l ] l x+ [ ]u Although Theorem 4.. r) is realizable...83) by using the transformation i A . the result is not useful i n practical implementation.ieb .2 for T = I and T = Jt . where the overbar denotes complex conjugate. M(t)I · r ' N e r ) D(t)S(t .4 1 ).r ) = [e" Thus the twodimensional timevarying state equation 4. what is the pennissible largest aO Consider have their largest magnitudes roughly equal to the largest magnitude of the output.�b l . 8 1 ) l]x is a realization. . EXAMPLE 4. as the next example i llustrates.82) 4.80).r) ..dimensional state equation x(t) = N(t)u(t) y(t) = 1\1 ( t ) x (t) + D(t)u(t) + PROBLEMS". r) can be decomposed as in (4.82) is much more difficult in practice. sin cos t t] I x(O) Use two different methods to find the unitstep response of Q.1 16 STATESPACE SOLUTIONS A N D REALIZAT IONS Proof: If G(t.. 8 1 ) is which equals G(t. This shows the sufficiency of the theorem. Thus the impulse response matrix of (4.7 Q1 _  [ . l ] x(r ) . If G(t. The implementation of (4. Identifying 1\1(1) = C(t)X(t) and N ( r ) = X . a fundamental matrix of x = O · x is X(t) = 1. r) = g( t .79). f ) .t. " " .
1 18
STATESPACE SOLUTIONS AND REALIZATIONS
can be transformed into
where A ii and Ci are defined in Problem 4.6 and 12 is the unit matrix of order 2. [Hint: , , Change the order of the state variables from [Xl X2 X3 X4 Xs X61' to [X l X4 Xz Xs diag(QI, Qi with Q X3 x61' and then apply the equivalence transformation
4.8
Q2 , Q .
x= ] = = [: i J 't [i} y = '
Y [c l C2 C3 CI C2 C3 X 12
0 0 0 0 ).. 0 ).. 0 0 0 0 0 � 0 0 0 0 0 0 0 0
)..
0 0 0 0 0 0 0 � 0 �
bb2l x+ b3
bl b3 b2
Show that its observable canonical form realization can be reduced from Problem 4. 9 as
u

4.11
[CI C2 c3]i
Find a realization for the proper rational matrix
A
0
)]
Are the two sets of state equations
and
equivalent? Are they zerostate equivalent?
4.9
Verify that the transfer matrix in
i� a l lq a2lq
This is called the to (4.34).
4.10
Consider the I x
A
(4.3 ) O [NN' 1 Iq : x+ : � �: O] y=[obsIqerOvablO e·ca·nonixcal/onn realization 2 G(s) = + als) + 3s +
x= = y = O]x y=[I IO]x
[1  1 2
4.12
Find a realization for each column of in Problem 4 . 1 1 and then connect them, What is the dimension of this as shown in Fig. 4.4(a), to obtain a realization of realization? Compare this dimension with the one in Problem 4. 1 1 . Find a realization for each row of in Problem 4. 1 1 and then connect them, as shown in Fig. 4.4(b), to obtain a realization of What is the dimension of this realization? Compare this dimension with the ones in Problems 4. 1 1 and 4. 1 2 . Find a realization for
4.13 4.14
[
4.15
Consider the ndimensional state equation
has the following realization:
o
Iq
0
0
ar  Ilq arlq
0 0
0 0
J
u
be its transfer function. Show that Let numerator of has degree m if and only if for
=
l
m and eAn   I b f= 0. Or, equivalently, the difference between the degrees of the denom is  m if and only if inator and numerator of g eA a  I b
and has dimension rq. It is dual
for i
4.16
[dl
d2] + =:: a4 a 2s 2 + a S4
proper rational matrix
Find fundamental matrices and state transition matrices for
1
LBl s3 + f321s2 + f331s + f341 f312s3 + f32 s2 + f332s + f3421 [a3al I ] + [f3f3f3011 ao f33411 � (;(s) = [ss2I (s :�)s�s3+ ] s G(s) s+ G(s). G(s) G(s). G(s) = [ 3s + +34 3s ++ J = Ax+ g(=s)ex g(s) g(s) i (s) a = n = = . .,a x= [ tI J x
X
Problems
1 19
x
=
 a4
1 0 0 0 o o 0 1 o 0 0
X
2)
+1
2
 0 2S
6) 22s
23
34
x
bu
y
has m zeros or, equivalently. the
0, I , 2,
. . . , n
m2
f= 0 and
eAib
0
0, 1 , 2,
2.
.
0 o
1 20
STATESPACE SOLUTIONS AND REALIZATIONS
and
. X =
[
1 o
Chapter
4.17 4.18
Show a <I> (to . t)/at =  <I> {to, t)A(t) . Given
show
Sta b i l ity
<l> l l (t. to) <1>2' (t , to ) <l> d t , to ) <l>d t , to )
4.19
Let be the state transition matrix of Show that <l>2, (t . to) = i = 1 , 2.
<I>(t . to ) =
[
]
to) .
x(t) =
[ �
A ' (t)
0 for all t
and to and that (a/ a t ) <I>ii (t, to) = Ai. <1>,, ( ( ,
A 12 (t) An (t )
]
x(t)
for
5. 1 I ntro duction
Systems are designed to perform some tasks or to process signals. If a system is not stable. the system may bum out. disintegrate. or saturate when a signal. no matter how small. is applied. Therefore an unstable system is useless in practice and stability is a basic requirement for all systems, In addition to stability. systems must meet other requirements. such as to track desired signals and to suppress noise. to be really useful in practice. The response of linear systems can always be decomposed as the zerostate response and the zeroinput response. It is customary to study the stabilities of these two responses separately. We will introduce the BlBO (boundedinput boundedoutput) stability for the zero state response and marginal and asymptotic stabilities for the zeroinput response, We study first the timeinvariant case and then the timevarying case.
4.20
Find the state transition matrix of
X=
. [
sin t
0
4.21
Verify that X(t) = eA'CeB' is the solution of Show that if A(t) = A , A(t)  A(t)A , . then
4.22
4.23 4.24 4.25 4.26
Transform a timeinvariant (A. B . C) into transformation,
Find an equivalent timeinvariant state equation of the equation in Problem 4.20, Show also that the eigenvalues of A (t) are independent of t.
A(t) = eA" A (O)e�A"
X (O) = C
5 . 2 I n putOutput Stability of LTI Systems
Consider a SISO linear timeinvariant (LTI) system described by
(0. B (t ) . eet)
by a timevarying equi\'alence
Find a timevarying realization and a timeinvariant realization of the impulse response Find a realization o f g U . r ) = sin t ee�(t�r) cos state equation realization�
g (t)
=
tZe AI ,
r.
I s i t possible to find a timeinvariant
where get) is the impulse response or the output excited by an i mpulse input applied at t = O. R ecall that in order to be describable by (5 . 1 ). the system must be linear. timeinvariant. and causal. In addition. the system must be initially relaxed at t = 0,
y(t) =
l'
get  r ) u (r ) dr =
l'
g(r)u(t  r ) dr
(5 1 )
121
1 22
STA B I L ITY
An input u(t) is said to be bounded if U (/) does not grow to positive or negative infinity or, equivalently, there exists a constant Urn such that
l
5.2 InputOutput Stability of LTI Systems
Figure 5.1
1 23
�
A system is said to be BIBO stable (boundedinput boundedoutput stable) if every bounded input excites a bounded output. This stability is defined for the zerostate response and is applicable only if the system is initially relaxed.
A SISO system described b y (5 . 1 ) i s BlEO stable if and only if g(t) is absolutely integrable in [0, (0), or
lu(t)1 � Urn
<
00
for all I ::: °
Function.
T
n 2 4 5
Theorem 5.1
for some constant
M.
["
I g t)ldt � M
t
<
00
Proof: First we show that if g(t) is absolutely integrable, then every bounded input excites
a bounded output. Let U (/) be an arbitrary input with ! u (t ) 1 � Urn
ly(t) 1 =
�
Urn
Thus the output i s bounded. Next we show intuitively that if get) i s not absolutely integrable, then the system is not BIBO stable. If g et) is not absolutely integrable, then there exists a tl such that
If 1""
g(r)u(t  r) dr �
I g(r) ldr � um M
l l'
<
00 for all t ::: 0. Then
integration of the function equals L::: (llnz) < 00. This function is absolutely integrable z but is not bounded and does not approach zero as I + 00.
�
Theorem 5.2
If a system with impulse response
Ig(r) l lu (t  r)l dr
1.
The output excited by
2 . The output excited b y u (I)
u(t) = a, for t ::: 0, approaches g(O) . a. = sin wot, for I ::: 0 , approaches Ig (jwo) 1 sin(wot+ 1::g (jwo)) g (5 )
g (t ) is BIBO stable, then, as I + 00:
where
Let us choose
t 10
g(5) is the Laplace transform of g (t) or
=
I g(r ) 1 dr
=
00
u(t\  r) =
{
_
Clearly u is bounded. However. the output excited by this input equals
�
if g(r)
if g(r) ::: °
<
y(tt l =
which is not bounded. Thus the system is not BIBO stable. This completes the proof of Theorem 5. 1 . Q.E.D. A function that is absolutely integrable may not be bounded or may not approach zero as
1'1
g\r)II(t\  r) dr =
1'1
°
•
Proof: If u (t) = a for all I ::: 0, then (5. 1 ) becomes
y(t) =
which implies
l'
LX!
g(r)esr dr
(5.2)
g(r)U(1  r) dr = a
Ig(r)1 dr = 00
where we have used (5.2) with 5 = 0. This establishes the first part of Theorem 5.2. If u (t) = sin wol, then (5. 1 ) becomes
y(/) + a
100
L
g(r) dr
g(r) dr = ag(O)
aS / + oo
t + 00. Indeed, consider the function defined by f(1  n) =

{
for n = 2, 3, . . , and plotted in Fig. 5 . 1 . The area under each triangle is l/nz. Thus the absolute
n + (t  n)n4 for n  I /n3 � I � n n (I  n)n 4 for n < I � n + I /n3
y(t) = =
l' l'
g(r) sin Wo(1  r) dr g(r) [sin wot cos Wo r  cos wol sin war] dr
= sinwol
l'
g(r) cos wor dr  COS Wol
l'
g(r) sin wor dr
3. the zerostate response of (5..E. 2 .9) as g ( l) = It is straightforward to verify that every such term is absolutely integrable if and only if P.p. of L a l 8 ( t . we can replace s by jw in (5. Theorem 5 .aeS ' "" g ( r ) sin wr d r It i s an irrational function of .5(a) is B I B O stable if and only if the gain a has a magnitude less than 1 . Theorem 5.4) is BIBO stable if and only if every pole of G(s) has a negative real part.2. The transfer function of the system was computed i n (2.. In this case. Thus we have [g(t)1 = 5 ..cos Wol The impulse response g (t ) is assumed implicitly to be real : thus we have Re[g(jw)] = lx lX g ( r ) sin Wo r dr (5. equivalently.4) Its transfer matrix i s Thus Equation (5.3 and can be considered to be positive.3 is not applicable. Theorem 5. then its partial fraction expansion contains the We now discuss the BlBO stability of state equations.I . Next we state the BlBO stability condition in terms of proper rational transfer functions.i) i=! 00 G(s) = C[Adj (s I .I and Theorem 5. respectively. sin wal If g (l ) is absolutely integrable. M 1 A multivariable system with impulse response matrix sin wol (Re[g(jwo ) ] ) + cos wot(Im[g(jwo ) ] ) �# This completes the proof of Theorem 5. 2. 2. m" A multivariable system with proper rational transfer matrix G(s) = [gij (. G(I) [ g. The impulse is defined as the limit of the pulse i n Fig. the real part a n d imaginary pan. has a negative real part.3) yields y (t) = [ g ( j wo ) [ s i n (wo l + t:.i ) if la i if l a l ::: 1 < i=l x and where Re and 1 m denote. 1 2) as aeS g ( s ) = I .. Consider yet) = Cx(t) + Du ( t ) j EXAMPLE 5.j sin wr] dr and lx g ( r ) cos Wo r dr . x ( t ) = Ax(t) + Bu(t) (5. Substituting these into (5. t.D. as I > y U) > oc.4) or. we have the following results. Theorem 5. 2 InputOutput Stabil ity of LTI Systems 1 25 L [ali 8(t .Iplane. it is simpler to use Theorem 5 .2 is a basic result. filtering of signals is based essentially on this theorem.. If g(s) has pole Pi with multiplicity factors . :xl).p. For multi variable systems.I ) ] is BlBO stable if and only if every pole of every g'i (5) has a negative real part. )2 ' (. Q. Thus the inverse Laplace transform of g (.3 every gij ( ! ) is absolutely integrable in [0.A)]B + D (5.5) . lies inside the lefthalf .I . g (jwo» > Im[g(jw)] = I lx 1 1 g ( r ) cos wr dr Thus we conclude that the positive feedback system in Fig. 1 to check its stability.1 Consider the positive feedback system shown in Fig.2) to yield g (jw) = g ( r ) [cos wr ..I ) or the impulse response contains the factors (s .5(a 1 .A) of G(s) is called a pole of G (s ) . Recall that every pole of every entry We discuss the relationship between the poles of G(s) and the eigenvalues of A. Because 1 det(sI . we can establish Theorem 5.P. Using this fact.1 24 STABILITY Thus we have.M3 A SISO system with proper rational transfer function g (s ) is BlBO stable if and only if every pole of g(s) has a negative real part or. j ( t) ] is BlBO stable i f and only if . to be more precise.. ) m. Its impulse response was computed in (2.3) where the gain a can be positive or negative.
1 26
STABILITY
every pole of G(s) is an eigenvalue of A. Thus if every eigenvalue of A has a negative real part, then every pole has a negative real part and (5.4) is BIBO stable. On the other hand, because of possible cancellation in not every eigenvalue is a pole. Thus, even if A has some eigenvalues with zero or positive real part, (5.4) may still be BIBO stable, as the next example shows.
1
(5.5),
Its proof is similar to the proof of Theorem counterpart of Theorem in the following.
5.2
InputOutput Stability of LTI Systems
5.2
5.1
and will not be repeated. We give a discrete
1 27
Theorem 5.02
If a discretetime system with impulse response sequence
2. The output excited by u 1. The output excited by
EXAMPLE 5.2 Consider the network shown in Fig. 4.2(b). Its state equation was derived in
=
The Amatrix is I and its eigenvalue is 1 . It has a positive real part. The transfer function of the equation is I + = 1)1 g (s = The transfer function equals It has no pole and no condition to meet. Thus is BIBO stable even though it has an eigenvalue with a positive real part. We mention that BIBO stability does not say anything about the zeroinput response, which will be discussed later.
Example 4.4 as
(5.6) x(t) x(t) u(t) yet) 0.5x(t) 0.5u(t) ) 0.5(}  ·0 0.5 0.5 0.5. (5.6)
=
+
O·
+
where
Consider a discretetime 8180 system described by
=
5.2. 1 DiscreteTime Case
where is the impulse response sequence or the output sequence excited by an impulse sequence applied at = O. Recall that in order to be describable by the discretetime system must be linear, timeinvariant, and causal. In addition, the system must be initially relaxed at = O. An input sequence is said to be does not grow to positive or negative infinity or there exists a constant Urn such that A system is said to be BIBO (boundedinput boundedoutput stable) if every bounded input sequence excites a boundedoutput sequence. This stability is defined for the zerostate response and is applicable only if the system is initially relaxed. :::: Urn
k k y[k] m=O g[k  m]u[m] m=O g[m]u[k  m] L L g[k] k (5.7), k u[k] bounded if u[k] k l u [k ] 1 stable
= < 00
for
(5.7)
= 0, 1 , 2, . . .
�
Theorem 5.0 1
A discretetime SISO system described by (5.7) is BIBO stable i f and only if in 00) or
Thus we have. as
[0,
LO Ig [k] 1 k=
for some constant
00
g[ k ]
i s absolutely summable
::::
M < 00
If
M.
u[[kk]] wok,k k 0, g[k] L m=O g[m]:m u[k] k (5.7) y[k] Lg[m]u [k m] Lg[m) k=O y[k] m:;:::Q g[m) k L u[k) wok, k((55..87)) I. L y[k) m=O g[m ) work m) k Lg[m) wok wom coswoksinwom) m:=::::Q k k Wok m==O g[m) cos wom Wok L g[m) wom L k L y[kJ Wok m=O g[m) coswom wok L g[m) wom (5.8) g[k) Lg[ Lg[ m=O m)eiWrn m=O m] [coswm  wm)
= a. for
= sin
2': O. approaches g O ) ' a.
for
g[k]
is BIBO stable. then. as
k
I g (e iw' ) 1 sin(wok + j:g (eiw,»
or
so
2':
�
00:
approaches
g(:) is the :transfonn of
g(;:)
=
(5.8)
Proof If
= a for all
=
2': 0, then
k
becomes =a
k
o
which implies
00
� a
= agO)
as
+
00
where we have used = sin then
with z = becomes sin
This establishes the first part of Theorem 5.D2. If
=
=
(s in
cos
= sin

cos
sin
� 00.
+
x
oc
sin

cos
sin
(5.9)
is absolutely summable, we can replace
g (eiw)
=
00
=
x
z
by eiw in
to yield
j si n
1 28
STABILITY
Thus (5.9) becomes
y[k)
5 . 3 Internal Stability
>
1 29
sin wok (Re[g (ejwo ) ) ) + cos wok (Im[g (ejw, ) ) )
= I g (ejw" ) i sin(wok + j:g (ejw, »
It is not a rational function of z and Theorem 5.03 is not applicable. For multi variable discretetime systems, we have the following results.
Theorem 5.MO 1
This completes the proof of Theorem 5 . 02 .
Q.E.D.
Theorem 5.02 is a basic result in digital signal processing. Next we state the BIBO stability in terms of discrete proper rational transfer functions.
Theorem 5.03
A MIMO discretetime system with impulse response sequence matrix G [k) = [g' [k]] is BlBO stable
if and only if every gi [k] is absolutely summable.
A discretetime S1S0 system with proper rational transfer function g<z) is BlBO stable if and only if . every pole of g (�) has a magnitude less than I o , equivalently, lies inside the unit circle on the :plane.
If g ( � ) has pole p, with multiplicity factors
;,  Pi
m"
j
j
�
Theorem 5.M03
then its partial fraction expansion contains the
A MIMO discretetime system with discrete proper rational transfer matrix G(�) = stable i f and only if every pole of every gi (Z) has a magnitude less than I .
j x[k + I ) = Ax[k] + Bu[k] y[k] = Cx[k] + Du[k]
[g'i (Z)] is BlBO
( ;,  p ; l " ,
( ;:  P, ) m,
We now discuss the BIBO stability of discretetime state equations. Consider
(5 . 1 0)
Thus the inverse ztransform of g (z) or the impulse response sequence contains the factors
p7, kp� , " ' , km,  I p;
It is straightforward to verify that every such term is absolutely summable if and only if has Pi a magnitude less than I . Using this fact, we can establish Theorem 5.03 . In the continuoustime case, an absolutely integrable function j U ) , as shown in Fig. 5 . 1 , may not be bounded and may not approach zero as t > 00 . In the discretetime case, if g [k) IS absolutely summable, then it must be bounded and approach zero as k > 00 . However. the converse is not true as the next example shows.
EXA�IPLE 5.3
Its di screte transfer matrix is
Consider a discretetime LTl system with impulse response sequence g[k) = 1 / k. for k = I . 2, . . . , and g[O] = O. We compute
S : = � Ig[kll =
Thus Equation ( 5 . 10) or. to be more precise, the zerostate response of (5. 1 0) is BIBO stable if and only if every pole of ( z) has a magnitude less than I . We discuss the relationship between the poles of G(z) and the eigenvalues of A . Because of
G
k�
�
= I +
�+ �
2 3
There are two terms, each is ± or larger. in the first pair of parentheses; therefore their sum is larger than i . There are four terms, each is � or larger, in the second pair of parentheses; therefore their sum is larger than 1. Proceeding forward we conclude This impulse response sequence is bounded and approaches 0 as k > 00 but is not absolutelv summable. Thus the discretetime system is not BlBO stable according to Theorem 5 . 0 I . Th� transfer function of the system can be shown to equal
I I I S > 1 +  +  +  + " ' = 00 2 2 2
( � ) (�
+
k= 1
Lx
I
k
I I I = 1 +  +  +4 +... 2 3 5 + ...+
. GU =
I
det(:I  A )
C[Adj (zI  A » ) B + D
4
+
�) (�
8
+
9
+'" +
�) '16 '
...
is an eigenvalue of A . Thus if every eigenvalue of A has a negative real every pole of part, then (5. 1 0) is BIBO stable. On the other hand, even if A has some eigenvalues with zero or positive real part. (5. 1 0) may, as in the continuoustime case, still be BIBO stable.
G(z)
5 , 3 I nternal Stability
The BIBO stability is defined for the zerostate response. Now we study the stability of the zeroinput response or the response of excited by nonzero initial state Xo' Clearly, the solution of (5 . 1 1 ) is
x(t) = Ax(t)
X ( I ) = eAtxo
(5. 1 1 )
I _J
(5 . 1 2)
1 30
STABILITY
De finition 5.1 The :eroinput response of (5.4) or the equation x = Ax is marginally
stable or stable in the sense of Lyapunov if everyfinite initial state Xo excites a bounded response. It is asymptotically stable ifeveryfinite initial state excites a bounded response, which, in addition, approaches 0 as t + 00.
T
Theorem 5.4
We mention that this definition is applicable only to linear systems. The definition that is applicable to both linear and nonlinear systems must be defined using the concept of equivalence states and can be found. for example, in Reference [6, pp. 40 1403]. This text studies only linear systems: therefore we use the simplified Definition 5 . 1 .
I
x =
i s not marginally stable, however, because its minimal polynomial is A 2 (A + I ) and A = 0 is not a simple root of the minimal polynomial.
[� � � ]
o 0
5 . 3 Internal Stabi li ty
131
x
I
As discussed earlier, every pole of the transfer matrix
1. The equation
x = Ax is marginally stable ;i and only if all eigenvalues of A have zero or negative real parts and those with zero real parts are simple roots of the minimal polynomial of A.
I
2. The equation
X = Ax is asymptotically stable if and only if all eigenvalues of A have negative real
parts.
We first mention that any (algebraic) equivalence transformation will not alter the stability of a state equation. Consider x Px, where P is a nonsingular matrix. Then x = Ax is equivalent to x = Ax = PAP I X. Because P is nonsingular, if x is bounded, so is x; if x approaches 0 as t + 00, so does Thus we may study the stability of A by using A. Note that the eigenvalues of A and of A are the same as discussed in Section 4.3. The response of x = Ax excited by equals xU) = eAt x(O). It is clear that the response is bo unded if and only if every entry of eAt is bounded for all t :::: If A is in Jordan . form, then eAt is of the form shown in (3.48). Using (3.48), we can show that if an eigenvalue has a negative real part, then every entry of ( 3 .48) is bounded and approaches 0 as t + 00. If an eigenvalue has zero real part and has no Jordan block of order 2 or higher, then the corresponding entry in (3.48) is a constant or is sinusoidal for all t and is, therefore. bounded. This establishes the sufficiency of the first part of Theorem 5 .4. If A has an eigenvalue with a positive real part, then every entry in (3.48) will grow without bound. If A has an eigenvalue with zero real part and its Jordan block has order 2 or higher, then (3.48) has at least one entry that grows unbounded. This completes the proof of the first part. To be asymptotically stable, every entry of (3.48) must approach zero as t + 00. Thus no eigenvalue with zero real part is permitted. This establishes the second part of the theroem.
=
X.
is an eigenvalue of A. Thus asymptotic stability implies BIBO stability. Note that asymptotic stability is defined for the zeroinput response, whereas BIBO stability is defined for the zerostate response. The system in Example 5.2 has eigenvalue and is not asymptoti cally stable; however, it is BIBO stable. Thus BIBO stability. in general, does not im ply asymptotic stability. We mention that marginal stability is useful only in the design of oscillators. Other than oscillators, every physical system is designed to be asymptoti cally stable or B IB O stable with some additional conditions, as we will discuss in Chap ter 7.
(;(s) = C(sI  A )  I B + D
I
X(O)
O.
5.3. 1 DiscreteTime Case
This subsection studies the internal stability of discretetime systems or the stability of
x[k +
= Ax[k]
I]
(5 . 1 3 )
excited by nonzero initial state Xo. The solution of (5. 1 3) is, as derived in (4.20).
x[k] = Akxo (5 . 1 4)
Equation (5. 1 3) is said to be marg inally stable or stable in the sense of Lyapunov if every finite initial state Xo excites a bounded response. It is asymptotically stable if every finite initial state excites a bounded response. which, in addition, approaches 0 as k + 00. These definitions are identical to the continuoustime case.
Theorem 5.04 A.
1. The equation
x[k +
I]
=
Ax[k] is marginally stable if and only if all eigenvalues of A have
EXAMPLE 5.4
Consider
Its characteristic polynomial is � (A) = A " (A + I ) and its mimimal polynomial is >/I (A) A(A + I). The matrix has eigenvalues 0, 0, and  I . The eigenvlaue 0 is a simple root of the minimal polynomial. Thus the equation is marginally stable. The equation
x=
[� � � ]
o
0
magnitudes less than or equal to
I and those equal to I are simple roots of the minimal polynomial of
x
2. The equation
I
magnitudes less than
x[k + 1 ] = Ax[k] is asymptotically stable if and only if all eigenvalues of A have
I.
As in the continuoustime case, any (algebraic) equivalence transformation will not alter the stability of a state equation. Thus we can use Jordan form to establish the theorem. The proof is similar to the continuoustime case and will not be repeated. Asymptotic stability
we call A stable if every eigenvalue of A has a negative real part.6. (5. Thus the importance of Theorem and its corollary is not in checking the stability of A but rather in studying the stability of nonlinear systems. where I'll is nonsingular (Theorem and consider = 0 . 2 0) is x 'Mx = x'eA ' f N'�eAf x dl = i : NeAf x l l � dl LX.5 5. We mention that Corollary can be used to prove the RouthHurwitz test. the eigenvalues of A and then check their real parts. 1 7) For any N. Even so. Suppose the integ�and of is identically zero or Ne'" x == O.2 2) This equation implies that.15) implies ReO. Other than oscillators.5 has a unique symmetric solution 1\1 and NI is positive definite .) Q. 1 5) . A'M + MA = . Then its derivative with respect to I yields :"IAeAfx = O. Thus the Lyapunov equation is nonsingular and has a unique solution M for any N. A'M + MA = N All eigenvalues of A have negative real parts'if and only if for any given positive definite symmetric matrix N.7) LX. because of (5.5 (5 20) Because both I'll and eAf are nonsingular. Av = . the inte grand of (5. 5. where the asterisk denotes complexconjugate transpose. Let A be an eigenvalue of A and v ¥ 0 be a corresponding eigenvector: that is. 19) (5. the matrix !Ii in is positive semidefinite (Theorem Theorem and its corollary are valid for any given N: therefore we shall use the simplest possible N . Theorem 5.1 7) 3.. we can obtain 5.15) is a special case of (3. For convenience. Sufficiency: We show that i f N and :\1 are positive definite. Even though A is a real matrix. 'v'. where we have used the fact e A f = 0 at I = :x for stable A. Even so. L e t us decompose N as N = N'N. (5. + AJ = O.9. Premultiplying v* and postmultiplying v to yields 3. We mention that marginal stability is useful only in the design of discretetime oscillators. 2 0). as we will discuss in Chapter M= Indeed. Proceeding forward. 4 Lyapunov Theorem This section introduces a different method of checking asymptotic stability of x = Ax.E.5 [6. 1 6)  positive for every I .5 5. the Lyapunov equation has a unique symmetric solution :\ 1 and IVI is positive definite.v.20) (5. for any nonzero x. 5. l I'll 1 NA N A:n 1 = n (full column rank) (5 . It is clear that if :"I is symmetric. It is much simpler t o compute. (5. using MATL\B.N'N =: . substituting 7.1 32 STABI LITY implies B I B O stability but not the converse. if A is stable.5 5. as discussed in Section 3. shows the positive definiteness of M. This shows that t h e M i n is t h e solution. 417419]. M in can still be positive definite if the integrand of is not identically zero for ali I . A has no two eigenvalues such that A.5 Necessitv: Equation (5.15) yields = A'M T MA = 5 .)v*Mv = 2Re(A )v*Mv O.5 (5. then A is stable. Because A a n d A' have the same s e t of eigenvalues. the only x meeting . its eigenvalue and eigenvector can be complex.7).59) with A = A ' (5. 3.21) < v*Nv = v * A'l\h + v*MAv = (1. 18) I'll with m All eigenvalues of an n x n matrix A have negative real parts if and only if for any given m < II and with the property rank 0 : = rank Corollary 5.16) and the nonsingularity of eAf . s o is M..18) eA'f NeAf A d t A'eA f NeAf dl + eA' f Ne �f ) dt = eA 1 NeAf [0 1'" (5. This shows that every eigenvalue of A has a negatiw real part. Thus x'Mx is positive for any x ¥ O. (5. * + i. x n matrix where 0 is an /1111 x n matrix . Consider :J o w N is m x n with m < n and N = N'N is positive semidefinite.20) and B = A . (5. They are essential in using the socalled second method of Lyapunov.18) into (5.D. We claim that the solution can be expressed as Proof of Theorem 5. We discuss only where the proof of Theorem is not applicable. as shown in Example Taking the complexconjugate transpose of Av = AV yields v* A* = v' A ' = i. See Reference pp.21 ) Because v*Mv and v*Nv are.N = N 1"" 1x f ( 1X 5.N (5 . 18) (5.4 Lyapunov Theorem 1 33 eA'f Ne\fdt (5. both real and positive.5 The proof of Corollary fol lows the proof of Theorem with some modific� tion. using them t o check stability of A is n o t simple. every discretetime physical system is designed to be asymptotically stable or B lB O stable with some additional conditions. the Lrapunol' equation (5. Thi. (5.
m where A and B are. Because of Av = AV.Lj = I for some i and j. Let v be a 1 x III left eigenvector of B associated with J. 1 6).25) are I . Let us define A(M) : = M . I I N:J l5J l5J This shows the uniqueness of M. j = 1. Thus (5. .M2) = 0 '> Theorem 5.Ai J.7. after some transformations. the leftmost matrix. .59).2 must be modified as where x'" is the complex conjugate transpose of x. the eigenvalues of A and B. It is simpler to arrange the Lyapunov equation. . then the Euclidean norm defined in Section 3 . the rightmost matrix is zero.A. vB = vJ. A scalar 7] is an eigenvalue of A if there exists a nonzero M such that A(M) = 7]M. . the integration will diverge and is meaningless. 2. 2 .25) can be expressed as Ym = c. 4 Lyapunov Theorem 1 35 5.23) implies Re(A)< O. the integration is not used in computing the solution. is nonzero under the conditions of (5. Then (5. As (3. we have established the following result.A u vB = (1 . This is a contradiction. Thus M is positive definite under the condition in (5.60). .(M. for any C. where Y is an nm x nm matrix: m and c are nm x I column vectors with the m columns of M and C stacked in order.Lj .05 All eigenvalues of an n xn 0 Ilxl l � = definite symmetric matrix x' x.25) can be written as A(M) = C. Applying A to the n x III nonzero matrix u v yields n. .M2) eAr j l � = 0 Thus the eigenvalues of (5. Next we consider (5. a unique solution M exists in (5 2 5 ) . solutions may or may not exist.20) cannot be identically zero for any x =f.24) 0 or. A u = AiU. M . .J.2 1 ) with N = N'N orl 2Re(A)v*Mv = v*N'Nv = . For easy reference. and the solution can be expressed as M= A'M + MA = N 100 eA'I NeA1 dt (5 . 0 . Then we have which implies eA'I [A'(MI .M2) + (MI . An .Lj ' for all i and j . .5. has at least one zero eigenvalue. This shows the necessity of Corollary 5.25) is singular and.25) is essentially a set of linear algebraic equations.25) is nonsingular and. Note that even if A is not stable. Then we have Before discussing the discrete counterpart of Theorems 5. . using eAr Its integration from 0 to 00 yields '""* [ eA'r (M I . we discuss the discrete counterpart of the Lyapunov equation in (3. respectively. The solution. 1 6). This can be established intuitively as follows. into a standard linear algebraic equation as in (3. that is.26) .5 and 5. 1 6).A ' MA = N (5. however.5.23) T Theorem 5.24). we state it as a theorem. Equation (5.1 34 STABILITY 2 We show that Nv is nonzero under (5. with the property in (5. O. The situation here is similar to what was discussed in Section 3. then (5. . for a given C. M is positive definite.25). Suppose there are two solutions MI and M2. cannot be expressed as in (5.M2) eArj = A'(MI .Ai/. n x n and m x m matrices. This completes the proof of Corollary 5.60) and then solve the equation. then the Lyapunov equation If Nv = 0. Consider M . a unique solution still exists if A has no two eigenvalues such that Ai + Aj = O. In the proof of Theorem of 5.M2)A = N.AMB. equivalently. however. I f A is singular or. 1 DiscreteTime Case has a unique solution for every Because of the importance of this theorem.4. A(uv) = u v .5. If Ai J.24).6.Lj)UV 0 as t '""* 00 . the discrete Lyapunov equation has a unique symmetric solution 1\'1 and matrix A have magnitudes less than I i f and only if for any given positive N or for N = N'N. we give a different proof of the uniqueness of the solution. 5 . and M and C are n x m matrices. Let u be an n x 1 right eigenvector of A associated with Ai : that is. 1 6) and v =f.Lj for i = I .AMB = C (5. Let 7]k be an eigenvalue of Y or of (5. respectively.Lj are.M2)A jeAr = � [eA'r (M I . Although the solution can be expressed as in (5.Lj . where N is any given m x n matrix with m < 11 and I.6 If all eigenvalues of A have negative real parts. where Ai and J.22) is O. . Note that if x is a complex vector. . then (5. we have A V = AAv = A V.25) 7]k = 1 . Thus Nv is nonzero and (5.M2) + (M I .I v = AnI v. Consider (5. . If there are no i and j such that AiJ.Lj = 1 .I I Nv l l � = = 2 (5. Thus the integrand of (5. . then the Lyapunov equation is always singular and solutions may or may not exist depending on whether or not N lies in the range space of the equation. O.
27) M = I )A')"'NA'" = � and A'M t. If all eigenvalues of and. (5. we conclude ( l .I A (S. the infinitenorm (S. let A be an eigenvalue of and corresponding eigenvector: that i s . for any a unique solution exists in We claim that the solution can be expressed as 5. The stability condition for continuoustime systems is that all eigenvalues lie inside the open lefthalf splane.34) yet) = 11 II) G u . (5. the solution cannot be expressed as in but can be computed from a set of linear algebraic equations.30).2 6) (S. consequently.26). v*Nv = v*�lv .28) The condition for to be BIBO stable i s that every entry of G(t. These conditions can be related by the bilinear transformation M = � (A')"'NAm A The system i s said t o b e B IB O stable if every bounded input excites a bounded output. I for all i and j. The proof in t h e timeinvariant case applies here vv'ith o n l y minor modification.3 0) A Av Av. we write Because lA.5 Stability of LTV Systems 1 37 N (5.I A I " ) > 0 or I A I " This establishes the theorem for N > O. Thus and is nonsingular. Any in norm discussed in Section can be used. 5. M. To differentiate the continuoustime and discretetime cases.32) � .:. Therefore.1 1 I IGllx = l argest row absolute sum . 3.29) Following (5.N ( 5 .33). then we have < I for all i and j. Substituting (S. and the solution can be expressed as Consider a SISO linear timevarying ( LTV) system described by y et ) = 11 10 gU.v* A'l\IAv = v*Mv .5 Stability of LN Systems These relate and The MATLAB function lyap computes the Lyapunov equation in and d l yap computes the discrete Lyapunov equation i n The function dl yap transforms into by using and then calls l yap. r ) 1 dr ::: M < x 1.2 7) 111 = 1 Substituting the righthandside equation into we find 15.l 5 = z + 1 1 5 1 +s (5.I) (5.29) (S. I < into yields 15.\IA = . these two equations can be related by (5.06 If all eigenvalues o f Thus i s the solution. (S. = Then we have N + � (A')"' NA'" . (5. of have magnitudes less than 1.1 36 STA B I LITY We sketch briefly its proof for > O. In this case. Theorem 5. r )u ( r ) dr (532) (5 33) It is important to mention that even if has one or more eigenvalues with magnitudes larger than a unique solution still exists in the discrete Lyapunov equation if ).3D) has a unique solution for every A have magnitudes less than I .29) (5. To show sufficiency. The stability condition for discretetime systems is that all eigenvalues lie inside the unit circle on the :plane.3D) and performing a simple manipulation. we can also express the condition in terms of norms.3'+ ) I lul l oe = max lUl l ..A'MA = N N. Let us discuss the relationships between the continuoustime and discretetime Lyapunov equations.29) M Md . s o i s I f � i s positive definite.32) 11 I" ! g ( t . .\" + nI (A" . then the d iscrete Lyapunov equation M .31) (S. = ( l . r ) u( r ) dr (5. s o i s establishes the necessity. Ai AJ 1= I A' which maps the lefthalf splane into the interior of the unit circle on the zplane and v'ice versa. The case N � 0 can similarly be established.29) yields (531 ) (5. The to be BIBO stable is that there exists a fi nite constant M such that condition for (S. This v 1= 0 be a Comparing this with < 1.27) m=O 1 for all i . this i n fi nite series conwrges and i s well defined. l f i s symmetric.26) N..'. (S.28).30). The result yields = A = ( . M. r ) meets the condition For multivariable systems. For the multi variable case. becomes (5.A *v*l\h.IAI")v*�lv Because both v*Nv and v*Mv are real and positiv'e.�(A')"' NA'" = N N 1Il = ! (5. However. IAi.27) AJ t= for a l l t a n d to w i t h t � to.
7 Marginal and asymptotic stabilities of x = A(t)x are invariant under any Lyapunov transfonnation. This example shows that even though the eigenvalues can be defined for A U ) at every t.I (rJP. and pet) is nonsingular for all t..36) (5.38)? (5 .2)th entry of 4> grows without bound. more precisely. (5. the concept of eigenvalues is not useful in the timevarying case.34) for all t and to with t ::: to. 0 (5.39) EXAMPLE 5. The equation x = A U)x is asymptotically stable if the response excited by every finite initial state is bounded and approaches zero as t + 00.37) [4. All stability properties in the timeinvariant case are invariant under any equivalence transformation. . therefore. r)B(r) + D(t)8(t r) 1 and . then x = P(t)x is a Lyapunov transformation. e') ] 0 e (5. Because the response is governed by l' '0 I I G(t. [A+ 1 4.35).37) As discussed in Section if pet) and P (t) are continuous.(r) (5.5 Stabi l i ty of LTV Systems 1 39 depend on to? What is the rate for the state transition matrix to approach in The interested reader is referred to References A timeinvariant equation x = Ax is asymptotically stable if all eigenvalues of A have negative real parts. to) 1 1 :::: M < 00 we conclude that the response is marginally stable if and only if there exists a finite constant M such that (5. (5.35)  The characteristic polynomial of A(t) is det(AI . we define the zeroinput response of stable if every finite initial state excites a bounded response. r ) = C(t)4> (t. An equivalence transformation can transform. this is so only for BIBO stabiliry.40) . and for all t and to with t ::: to.35) l' '0 4> (t .39).35) (5.1 (r) I = P(t)4>(t. In the timevarying case. to)x(to) 1 1 4> (t.5(e' _. Theorem 5. r)p.3. r) = X (t) X . P(t) and p .35) is BIBO stable if and only if meets and is therefore the state transition matrix of See also Problem Because the ( l . 15]. in the timevarying case. The necessary and sufficient condition for to be BIBO stable is that there exists a finite constant M such that (5. r ) 1 1 dr :::: M < 00 1 5. where Ao is any constant matrix. r)B(r)u(r) dr + D(t)u(t) (5.53) 0 [ e' 0.1 38 STA B I LITY is probably the most convenient to use in stability study. Does the constant M in + oc which implies I 4> (t. The impulse response matrix of x = A(t)x + B (t)u y = C(t)x + D(t)u l' '0 I I G(t.1 for all t. If. the zerostate response of there exist constants MI and M2 such that I I D(t) 1 1 :::: MI < 00 (5. to) 1 1 + X (t) = P(t)X(t) (5. the equation is neither asymptotically stable nor marginally stable. because the impulse response matrix is preserved. O) = C (t ) 4> ( t . no subscript will be attached to any norm. as shown in Theorem any x = A(t)x into x = Aox. as derived in 4. xU) = 4>U .37) 0 as t A great deal can be said regarding these definitions and conditions.6.(r) = P(t)X(t)X. then x = PU)x is an algebraic transformation.A (t» = det Thus A (t) has eigenvalues [ �1 _ e2' ] I x is G(t. 16. (4. It can be verified directly that and the zerostate response is yet) = Thus or. The asymptotic stability conditions are the boundedness condition in and 1 1 4>(t . in addition.38) (5. For convenience. A for all to and for all t ::: to .5 Consider the linear timevarying equation x = A(t)x = (5. Next we study the stability of the zeroinput response of As in the timeinvariant or the equation x = A(t)x to be marginally case.I (t) are bounded for all t . The fundamental matrix X(t) of x = A (t)x and the fundamental matrix X (t) of x = (t) x are related by.7 1). marginal and asymptotic stabilities are not invariant under any equivalence transformation. Is this also true for the timevarying case? The answer is negative as the next example shows. (4. r ) l l dr :::: M2 < 00 4.
to with t ::: to . then asymptotic stability does imply BlBO stability.211 [ � �] [ � ] 1 x+ 2 0.9� °� 0. in panicular. 5 . 1 I <I> (t . This establishes Theorem 5 . Consider a system with transfer function g ( s ) = (s . 1 1 Is the homogeneous state equation marginally stable'} Asymptotically stable" x = u 5. and i f C ( t ) and B ( t ) are bounded for all t. For a = I . A function that approaches as t > oc .2 0.42) 5 . 05 to show that all eigenvalues of the A in Problem 5 .7 . and any nonzero real a . to) ! ! > 10 ° as t Consider a discretetime system with impulse response sequence 5. for 1 ::: and by lI !t) = sin 21. This is not necessarily so in the timevarying case. 2 B lBO stable� I f not.12 Is the discretetime homogeneous state equation marginally stable'} Asymptotically stable" x = [I� ° [0.2J / (s + I ) B IB O stable? 5. 1 4 have magnitudes  less than I . have negative real pans. may not be absolutely integrable. 5.3 5.140 STA B I L ITY Because bOlh P(t) and p. 1 5 ] .2 ) / ( s + I ) .4