Sei sulla pagina 1di 246

SCHAUM'S OVTLIISE OF

THEORY AIVD PROBLEMS


OF

STATE SPACE
and

LINEAR SYSTEMS

BY

DONALD

M. WIBERG, Ph.D.

Associate Professor of Engineering


University of California,

Los Angeles

H^
SCHAIJM'S OUTLINE SERIES
McGRAW-HILL BOOK COMPANY
New

York,

St. Louis,

San Francisco, Diisseldorf, Johannesburg, Kuala Lumpur, London, Mexico

Montreal, New Delhi, Panama, Rio de Janeiro, Singapore, Sydney, and Toronto

1971 by McGraw-Hill, Inc. All Rights Reserved. Printed in the


Copyright
United States of America. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording, or otherwise, without the
prior written permission of the publisher.
07-070096-6
5 6 7 8 9 10 11 la 18 14 15

SH SH

7 S 4 8 2 1

6 9 8

Preface
The importance of state space analysis is recognized in fields where the time behavior
of any physical process is of interest. The concept of state is comparatively recent,
but the
methods used have been kno-wn to mathematicians for many years. As engineering, physics,
medicine, economics, and business become more cognizant of the insight that the state space
approach

offers, its

This book

popularity increases.

not only for upper division and graduate students, but for pracIt is an attempt to bridge the gap between theory and practical
use of the state space approach to the analysis and design of dynamical systems. The book
is meant to encourage the use of state space as a tool for analysis and design, in
proper
relation with other such tools. The state space approach is more general than the "classical"
Laplace and Fourier transform theory. Consequently, state space theory is applicable to all
systems that can be analyzed by integral transforms in time, and is applicable to many
systems for which transform theory breaks down. Furthermore, state space theory gives
a somewhat different insight into the time behavior of linear systems, and is worth studying
for this aspect alone.
vs^as vi^ritten

ticing professionals as well.

In particular, the state space approach is useful because: (1) linear systems with timevarying parameters can be analyzed in essentially the same manner as time-invariant linear
systems, (2) problems formulated by state space methods can easily be programmed on a
computer, (3) high-order linear systems can be analyzed, (4) multiple input-multiple output
systems can be treated almost as easily as single input-single output linear systems, and
(5) state space theory is the foundation for further studies in such areas as nonlinear
systems, stochastic systems, and optimal control. These are five of the most important
advantages obtained from the generalization and rigorousness that state space brings to
the classical transform theory.

Because state space theory describes the time behavior of physical systems in a mathematical manner, the reader is assumed to have some knowledge of differential equations and
of Laplace transform theory. Some classical control theory is needed for Chapter 8 only.

No knowledge

of matrices or complex variables

The book may appear

to contain too

is

prerequisite.

many theorems

to be comprehensible and/or useful

But the theorems have been stated and proven in a manner


suited to show the range of application of the ideas and their logical interdependence.
Space that might otherwise have been devoted to solved problems has been used instead
to the nonmathematician.

Consequently I give my strongest recommendation that the reader seek to understand the physical ideas underlying the proofs rather
than to merely memorize the theorems. Since the emphasis is on applications, the book
might not be rigorous enough for the pure mathematician, but I feel that enough information has been provided so that he can tidy up the statements and proofs himself.
to present the physical motivation of the proofs.

The book has a number of novel features. Chapter 1 gives the fundamental ideas of
from an informal, physical viewpoint, and also gives a correct statement of linearity.
Chapter 2 shows how to write transfer functions and ordinary differential equations in

state

matrix notation, thus motivating the material on matrices to follow. Chapter 3 develops
the important concepts of range space and null space in detail, for later application. Also
exterior products (Grassmann algebra) are developed, vi^hich give insight into determinants,
and which considerably shorten a number of later proofs. Chapter 4 shows how to actually
solve for the Jordan form, rather than just proving its existence. Also a detailed treatment
of pseudoinverses is given. Chapter 5 gives techniques for computation of transition
matrices for high-order time-invariant systems, and contrasts this with a detailed development of transition matrices for time-varying systems. Chapter 6 starts with giving physical
insight into controllability and observability of simple systems, and progresses to the point
of giving algebraic criteria for time-varying systems. Chapter 7 shows how to reduce a
system to its essential parameters. Chapter 8 is perhaps the most novel. Techniques from
classical control theory are extended to time-varying, multiple input-multiple output linear
systems using state space formulation. This gives practical methods for control system
design, as well as analysis. Furthermore, the pole placement and observer theory developed
can serve as an introduction to linear optimal control and to Kalman filtering. Chapter 9
considers asymptotic stability of linear systems, and the usual restriction of uniformity is
dispensed with. Chapter 10 gives motivation for the quadratic optimal control problem,
with special emphasis on the practical time-invariant problem and its associated computational techniques.
Since Chapters 6, 8, and 9 precede, relations with controllability, pole
placement, and stability properties can be explored.

The book has come from a


from 1966.

originally dating

set of notes developed for engineering course

was given to the publisher in June


the publication delay has dated some of the material. Fortunately, it
of errors to be weeded out.
It

1969.

122B at UCLA,
Unfortunately,

also enabled a

number

Now I would like to apologize because I have not included references, historical development, and credit to the originators of each idea. This was simply impossible to do because
of the outline nature of the book.
I

would

like to

express

my appreciation to

those

who

helped

me

write this book.

Chapter

was written with a great deal of help from A. V. Balakrishnan. L. M. Silverman helped
with Chapter 7 and P.K.C. Wang with Chapter 9. Interspersed throughout the book is
material from a course given by R. E. Kalman during the spring of 1961 at Caltech. J. J.
DiStefano, R. C. Erdmann, N. Levan, and K. Yao have used the notes as a text in UCLA
course 122B and have given me suggestions. I have had discussions with R. E. Mortensen,
M. M. Sholar, A. R. Stubberud, D. R. Vaughan, and many other colleagues. Improvements
in the final draft were made through the help of the control group under the
direction of
J. Ackermann at the DFVLR in Oberpfaffenhofen,
West Germany, especially by G. Grubel
and R. Sharma. Also, I want to thank those UCLA students, too numerous to mention, that
have served as guinea pigs and have caught many errors of mine. Ruthie Alperin was
very
efficient as usual while typing the text. David Beckwith,
Henry Hayden, and Daniel Schaum
helped publish the book in its present form. Finally, I want to express my appreciation
of
my wife Merideth and my children Erik and Kristin for their understanding during the
1

long hours of involvement with the book.

Donald M. Wiberg
University of California, Los Angeles
June 1971

CONTENTS
Page
Chapter

MEANING OF STATE

Introduction to State.
State of an Abstract Object.
Trajectories in State
Space.
Dynamical Systems. Linearity and Time Invariance. Systems Considered.
Linearization of Nonlinear Systems.

Chapter

METHODS FOR OBTAINING THE STATE EQUATIONS

16

Plow Diagrams.

Properties of Flow Diagrams.


Canonical Flow Diagrams
for Time-Invariant Systems.
Jordan Flow Diagram. Time- Varying Systems.
General State Equations.

Chapter

ELEMENTARY MATRIX THEORY

38

Basic Definitions. Basic Operations. Special Matrices. Determinants and Inverse Matrices. Vector Spaces.
Bases.
Solution of Sets of
Linear Algebraic Equations. Generalization of a Vector. Distance in a Vector
Space.
Reciprocal Basis. Matrix Representation of a Linear Operator.
Exterior Products.
Introduction.

Chapter

MATRIX ANALYSIS

69

Eigenvalues and Eigenvectors. Introduction to the Similarity Transformation.


Properties of Similarity Transformations.
Jordan Form.
Quadratic Forms.
Matrix Norms. Functions of a Matrix. Pseudoinverse.

Chapter

SOLUTIONS TO THE LINEAR STATE EQUATION

99

Transition Matrix.
Calculation of the Transition Matrix for Time-Invariant
Systems. Transition Matrix for Time-Varying Differential Systems. Closed
Forms for Special Cases of Time-Varying Linear Differential Systems. Periodically-Varying Linear Differential Systems.
Solution of the Linear State
Equations with Input.
Transition Matrix for Time- Varying .Difference Equations.
Impulse Response Matrices.
The Adjoint System.

Chapter

&

CONTROLLABILITY AND OBSERVABILITY


Introduction to Controllability and Observability.
Controllability in TimeInvariant Linear Systems.
Observability in Time-Invariant Linear Systems.
Direct Criteria from A, B, and C. Controllability and Observability of TimeVarying Systems. Duality.

128

CONTENTS
Page
Chapter

CANONICAL FORMS OF THE STATE EQUATION

147

Introduction to Canonical Forms.


Jordan Form for Time-Invariant Systems.
Real Jordan Form.
Controllable and Observable Forms for Time-Varying
Systems.
Canonical Forms for Time-Varying Systems.

Chapter

RELATIONS WITH CLASSICAL TECHNIQUES

164

Matrix Flow Diagrams.


Root Locus.
Steady State Errors.
Nyquist Diagrams.
State Feedback Pole Placement.
Observer Systems.
Algebraic Separation. Sensitivity, Noise Rejection, and Nonlinear Effects.
Introduction.

Chapter

STABILITY OF LINEAR SYSTEMS

191

Definitions of Stability for Nonzero Inputs.


Liapunov Techniques.
Liapunov
Functions for Linear Systems.
Equations for the Construction of Liapunov
Introduction.

Definitions of Stability for Zero-Input Linear Systems.

Functions.

Chapter

10

INTRODUCTION TO OPTIMAL CONTROL

208

The Criterion Functional.

Derivation of the Optimal Control


Law. The Matrix Riccati Equation. Time-Invariant Optimal Systems. Output Feedback.
The Servomechanism Problem.
Conclusion.
Introduction.

INDEX

233

chapter
Meaning
1.1

of State

INTRODUCTION TO STATE

To introduce the subject, let's take an informal, physical approach to the idea of state.
(An exact mathematical approach is taken in more advanced texts.) First, we make a
distinction between physical and abstract objects. A physical object is an object perceived
by our senses whose time behavior we wish to describe, and its abstraction is the mathematical relationships that give some expression for its behavior. This distinction is made
because, in making an abstraction, it is possible to lose some of the relationships that make
the abstraction behave similar to the physical object. Also, not all mathematical relationships can be realized by a physical object.
The concept of state relates to those physical objects whose behavior can change with
time, and to which a stimulus can be applied and the response observed. To predict the
future behavior of the physical object under any input, a series of experiments could be
performed by applying stimuli, or inputs, and observing the responses, or outputs. From
these experiments we could obtain a listing of these inputs and their corresponding observed
outputs, i.e. a list of input-output pairs. An input-output pair is an ordered pair of real
time functions defined for all t ^- U, where to is the time the input is first applied. Of
course segments of these input time functions must be consistent and we must agree upon

what kind of functions to consider, but in this introduction we shall not go into these
mathematical details.
Definition 1.1:

The

state of a physical object is any property of the object which relates


input to output such that knowledge of the input time function for t
to

and
Example

state at time

to

completely determines a unique output for

to.

1.1.

Consider a black box. Fig. 1-1, containing a switch to one of two voltage dividers.

box is the posiwhich agrees with DefiThis can be ascertained by the

Intuitively, the state of the

tion of the switch,

nition 1.1.

experiment of applying a voltage V to the


input terminal. Natural laws (Ohm's law)
dictate that if the switch is in the lower
position A, the output voltage is V/2, and
if the switch is in the upper position B, the
output voltage is y/4. Then the state A
determines the input-output pair to be
{V, V/2), and the state B corresponds to
(V, y/4).

1.2

terminal

Fig. 1-1

STATE OF AN ABSTRACT OBJECT


The basic

and

Ms)
Output

ideas contained in the above example can be extended to

many

physical objects

This will be done after


abstracting the properties of physical objects such as the black box. For example, the
color of the box has no effect on the experiment of applying a voltage. More subtly, the
value of resistance R is immaterial if it is greater than zero. All that is needed is a listing
of every input-output pair over all segments of time t to, and the corresponding states
at time to.
to the abstract relationships describing their time behavior.

MEANING OP STATE

An

Definition 1.2:

abstract object

is

[CHAP.

the totality of input-output pairs that describe the

behavior of a physical object.


Instead of a specific

list

of input time functions and their corresponding output time

functions, the abstract object

is usually characterized as a class of all time functions that


obey a set of mathematical equations. This is in accord with the scientific method of
hypothesizing an equation and then checking to see that the physical object behaves in a
manner similar to that predicted by the equation. Hence we can often summarize the
abstract object by using the mathematical equations representing physical laws.

relations which summarize an abstract object must be oriented,


functions that obey the relations must be designated inputs (denoted
by the vector u, having
elements tu) and k of the time functions must be designated
outputs (denoted by the vector y, having k elements yt). This need has nothing to do with
causality, in that the outputs are not "caused" by the inputs.

The mathematical

in that

m of the time

The

Definition 1.3:

state of

an abstract object

with the input


for

all

u(i) for all

is

a collection of numbers which together


uniquely determines the output y{t)

to

to.

In essence the state parametrizes the listing of input-output pairs. The state is the
to the question "Given u{t) for t to and the mathematical relationships of the
abstract object, what additional information is needed to completely specify y(t) for t^toV

answer
Example

1.2.

physical object is the resistor-capacitor network


shown in Fig. 1-2. An experiment is performed by applying
a voltage u{t), the input, and measuring a voltage y(t), the
Note that another experiment could be to apply
output.
y{t) and measure u{t), so that these choices are determined

J?

by the experiment.

The list of all input-output pairs for this example is


the class of all functions u(t),y{t) which satisfy the mathematical relationship

RCdy/dt + y

= u

This summarizes the abstract object.


yit)

.(.ta-ttlRC

y(to)e'

(1.1)

The

hf"

solution of (1.1) is

T-fl/RC,
m(t) dr

Fig. 1-2

{1.2)

This relationship explicitly exhibits the list of input-output pairs. For any input time function u{t) for
Note the
T ^ to. the output time function y{t) is uniquely determined by 2/(to). a number at time toThus the set of numbers y(to) parametrizes all inputdistinction between time functions and numbers.
output pairs, and therefore is the state of the abstract object described by {1.1). Correspondingly, a
choice of state of the RC network is the output voltage at time to-

Example 1.3.
The physical object shown

in Fig. 1-3 is

two

R^C^ diy/dt^

RC

networks in

2.5RC dy/dt

Fig. 1-3

series.

= u

The pertinent equation

is

{1.3)

CHAP.

MEANING OF STATE

1]

with a solution

y(t)

.^[4e(to-tV2Rc

gCto-m/BC-j

rf^,^ sr (f-t)/2RC

2_

_2_
[e(-/2RC
3fiC,J'

S^

p(to-t)2/RCl

^<.-t)2/RC](^)

rf^

(^.^)

to

Here the

set of

numbers

Physically, the voltage

j/(to)

and

and -^ (tp) parametrizes the input-output

its

pairs,

and

derivative across the smaller capacitor at time

to

may

be chosen as state.

correspond to the state.

state variable, denoted by the vector x(i), is the time function whose
value at any specified time is the state of the abstract object at that time.

Definition 1.4:

Note this difference in going from a set of numbers to a time function. The state can
be a set consisting of an infinity of numbers (e.g. Problems 1.1 and 1.2), in which case the
state variable is an infinite collection of time functions. However, in most cases considered
in this book, the state is a set of n numbers and correspondingly x(t) is an K-vector function
of time.

The

Definition 1,5:

Example

The
either

state variable in

or

Ebcample

for

the set of

all x(i).

Example

1.2 is

x(t)

y{t),

whereas, in Example

1.1 the state variable

remains

all time.

1.5.

state variable in

Example

x(t)

1.3 is

The state representation is not unique.


the relationship of input to output.
Example
In

is

1.4.

The

by 2,

state space, denoted

There can be many different ways of expressing

1.6.

Example

1.3,

be the voltage and


both capacitors.

instead of the voltage and its derivative across the smaller capacitor, the state could
derivative across the larger capacitor, or the state could be the voltages across

its

There can exist inputs that do not influence the state, and, conversely, there can exist
outputs that are not influenced by the state. These cases are called uncontrollable and
unobservable, respectively, about which much more will be said in Chapter 6.
Example

1.7.

state uncontrollable.
No input can make the switch change
observable. If the wire to the output were broken, it would be
unobservable.
state that is both unobservable and uncontrollable makes no physical sense, since it cannot be detected by experiment. Examples 1.2 and 1.3 are both controllable and observable.

In

Example

positions.

1.1,

the physical object

However, the switch position

is

is

One more point to note is that we consider here only deterministic abstract objects.
The problem of obtaining the state of an abstract object in which random processes are
Consequently,
inputs, etc., is beyond the scope of this book.
book are intended only for deterministic processes.

all

statements in the whole

MEANING OF STATE

[CHAP.

TRAJECTORIES IN STATE SPACE

1.3

The state variable x(*) is an explicit function of time, but also depends implicitly on the
starting time U, the initial state x(fo) = xo, and the input u(t). This functional dependency
can be written as x(t) = ^{t; to, xo, u(t)), called a trajectory. The trajectory can be plotted
in TC-dimensional state space as t increases from to, with t an implicit parameter.
Often
this plot can be

Example
a

Xi{t)

circle in the

Example
In

eliminating

from the

solutions to the state equation.

1.8.

Given
is

made by

sin

a;ia;2

and

X2(t)

plane with

cos

squaring each equation and adding gives

t,

xl

+ x^ =

1.

This

an implicit parameter.

1.9.

Example

1.3, note that equation (14) depends on t, u(t), x(to) and to. where x(to) is the vector with
components y(tg) and dy/dt(ta). Therefore the trajectories ^ depend on these quantities.
Suppose now u{t) =
and RC = 1. Let x^ = y{f) and Xz = dyldt. Then dx^/dt = x^ and d^y/dt^ =
dx2/dt.
Therefore dt = dx^/x^ and so d^y/dt^ = Xzdx^/dx^.
Substituting these relationships into {l.S)

gives

X2 dx^ldx]^

which

is

independent of

t.

2.5a;2

ajj

This has a solution


sci

2a;2

C(2xi

+ Xg)*

where the constant C = [x^ito) + 2a;2(to)]/[2a;i(to) + a;2(*o)]*- Typical trajectories in state space are shown
in Fig. 1-4.
The one passing through points a;i(to) =
and Xzitg) = 1 is drawn in bold. The arrows
point in the direction of increasing time, and all trajectories eventually reach the origin for this particular
stable system.

Fig. 1-4

1.4

DYNAMICAL SYSTEMS

In the foregoing we have assumed that an abstract object exists, and that sometimes
find a set of oriented mathematical relationships that summarizes this listing of
input and output pairs. Now suppose we are given a set of oriented mathematical relationships, do we have an abstract object? The answer to this question is not always affirmative,
because there exist mathematical equations whose solutions do not result in abstract objects.

we can

Example 1.10.
The oriented mathematical equation
input or the output must be imaginary.

y(t)

ju(t)

cannot give an abstract object, because either the

mathematical relationship always determines a real output y{t) existing for all
given any real input u(^) for all time t, then we can form an abstract object. Note
that by supposing an input u{t) for all past times as well as future times, we can form an
abstract object from the equation for a delayor y{t) = uit-T). [See Problem 1.1.]
If a

t^

to

CHAP.

MEANING OF STATE

1]

However, we can

+ T).

If

Definition 1.6:

also

form an abstract object from the equation for a predictor

we

are to restrict ourselves to mathematical relations that can be


mechanized, we must specifically rule out such relations whose present outputs depend
on future values of the input.
y{t)

u{t

dynamical system
(1)

A
all

an oriented mathematical relationship in which:

real output y{t) exists for all

- to

given a real input u(t) for

t.

Outputs

(2)

is

y(t)

do not depend on inputs

u(t) for

t>

t.

Given that we have a dynamical system relating y{t) to u(t), we would like to construct
a set of mathematical relations defining a state x(f). We shall assume that a state space
description can be found for the dynamical system of interest satisfying the following
conditions (although such a construction may take considerable thought):
unique output y(t) = ii{t, 4,{t; to,xo,u(T)), u(t)) exists for
given the state xo at time to and a real input u(t) for t ^ to.

Condition

1:

A real,

Condition

2:

Condition

3:

unique trajectory ^(t; to,xo,u(T)) exists for


time to and a real input for all t to.
unique trajectory starts from each
lim^(t; t,,x(ti),u(T))

Condition

x(ti)

>

to

>

to

given the state at

i.e.

for

all

ti

^ to

{1.5)

Trajectories satisfy the transition property

i:

^(t;

to,

x(to), u(t))

where
Condition

state,

all

all i

x(ti)

Trajectories ^(t;

5:

to,

xo, u(r))

^(t;

ti,

x(t,), u(t))

for

to

<

ti

<

(1.6)
(1.7)

^(ti; to,x(to),u(T))

do not depend on inputs

u(t) for t

>

t.

state
Condition 1 gives the functional relationship y(t) - Ti(t,x(f),u(t)) between initial
and future input such that a unique output is determined. Therefore, with a proper state
space description, it is not necessary to know inputs prior to to, but only the state at time to.
The state at the initial time completely summarizes all the past history of the input.

Example

1.11.

obtained in the past.


In Example 1.2., it does not matter how the voltage across the capacitor was
input.
future
the
state
and
the
is
output
future
unique
All that is needed to determine the

Therefore
Condition 2 insures that the state at a future time is uniquely determined.
For
output.
the
determines
uniquely
knowledge of the state at any time, not necessarily to,
exists
and
space
state
point
in
each
a given u(t), one and only one trajectory passes through
this is that the state
for all finite t ^ to. As can be verified in Fig. 1-4, one consequence of
not require the
does
condition
2
trajectories do not cross one another. Also, notice that
real.
state to be real, even though the input and output must be
Example

The

1.12.

relation

with output

y(t)

dy/dt

-jx{t)

state space description d/dt = ju{t)


obviously a dynamical system.
imaginary.
can be constructed satisfying conditions 1-5, yet the state is

u(t)

is

MEANING OF STATE

[CHAP.

Condition 3 merely requires the state space description to be consistent, in that the
starting point of the trajectory should correspond to the initial state. Condition 4 says
that the input u(t) takes the system from a state x(to) to a state x(*), and if x(ii) is on
that trajectory, then the corresponding segment of the input will take the system from
x(ii) to x(i).
Finally, condition 5 has been added to assure causality of the input-output
relationship resulting from the state space description to correspond with the causality of
the original dynamical system.
Example
a^(*)

1.13.

We can construct a state space description of equation (1.1)


= yit)- Then condition 1 is satisfied as seen by examination of

the trajectory <p{t; tg, x^, m(t)) exists and is


Also, conditions 3 and 5 are satisfied. To
then

of Example 1.2 by defining a state


the solution, equation (1.2). Clearly
unique given a specified fo, Kq and m(t), so condition 2 is satisfied.
check condition 4, given x(tg) = y(tQ) and u{t) over <<, t t,
t

x(t)

a:(*i)e"i-"/c

;^J

x(t,)

^(ye"-'>''c

^J

where

e'^^-'''''^ u(r,) dr,

(1.8)

^ t,

Substitution of (1.9) into (1.8) gives the previously obtained


a state space description satisfying conditions 1-5.

e<^-''>^'^u(ro)dro

(1.2).

(1.9)

Therefore the dynamical system

Henceforth, instead of "dynamical system with a state space description"


say "system" and the rest will be understood.

1.5

(1.1)

has

we will simply

LINEARITY AND TIME INVARIANCE


Given any two numbers a, /3; two states xi(io), X2(io); two inputs ui(t), U2(t);
and two corresponding outputs yi(T), y2(T) for t U. Then a system is

Definition 1.7:

the state X3(io) = axi(fo) + (6x2(^0), the output ysW = ayi(T)


/^yzi-r), and the input U3(t) = aUi{T) + /3u2{t) can appear in the oriented abstract object and (2) both y3(T) and X3(t) correspond to the state X3(to) and
input U3(t).
linear if

The operators
{x(io)}

is

^(t;

io,

(1)

xo, u(t))

x(i)

and

rt{t;<j>(t;to,xo,u(T)))

= y(t)

are linear on {u(t)}

an equivalent statement.

Example 1.14.
In Example

1.2,

y,(t)

x,(t)

= x.ay'"-"'^' +

2/2(0

'2(<)

a;2(ye"o-/RC

^j

e<-/ci(r)dr

+ ^j',Cx-/RC^(^)rf^
to

are the corresponding outputs j/i(t) and y2(t) to the states Xi(t) and X2(t) with inputs Mi(t) and U2(t).
Since any magnitude of voltage is permitted in this idealized system, any state x^(t) = aXi(t) + Px2(t),
any input Us(t) = aUi(t) + pu2(t), and any output ys(t) = aj/i(t) + I3y2(t) will appear in the list of inputoutput pairs that form the abstract object. Therefore part (1) of Definition 1.7 is satisfied. Furthermore,
let's look at the response generated by !K3(fo) and M3(t).
y(t)

Xs(to)e'*-""'''

e''-*'"'''u,(r)dr

to

to

=
Since ys(t)

Xg(t),

aj/l(*)

i8j/2(*)

J/3()

both the future output and state correspond to

X3(to)

and

WgCf)

and the system

is linear.

CHAP.

Example

MEANING OF STATE

1]

1.15.

Consider the system of Example 1.1. For some a and P there is no state equal to aA + pB, where A and
are the switch positions. Consequently the system violates condition (1) of Definition 1.7 and is not linear.

Example

1.16.

Given the system dx/dt

The

state

X3{t)

x^it^

y = u cos x.
+ P'x'ziio) and

y{t)

except in special cases like

Then

0,

ai(to)

+ I3u2(t)]

[auiit)

Xi{tg)

x^iti^

2/i(t)

cos [aXi(to)

Ui(t) cos x^Hq)

Px^Ho)]

system

so the

0,

and

j/2(*)

'*2(*)

cos x^itQ).

but the output

linear,

is

is

aj/i(t)

PViit)

not linear.

and also
system is linear, then superposition holds for nonzero u(t) with x(to) =
initial
=
zero
with
for nonzero x(io) with \x{t)
but not both together. In Example 1.14,
mt)
could
sin
voltage on the capacitor, the response to a biased a-c voltage input (constant +
be calculated as the response to a constant voltage input plus the response to an unbiased
a-c voltage input. Also, note from Example 1.16 that even if superposition does hold for
nonzero u(i) with x(io) =
and for nonzero x(o) with u(t) = 0, the system may still not
be linear.
If a

A system is time-invariant

Definition 1.8:

lent

system

if

the time axis can be translated and an equiva-

results.

compare the original output with the shifted output.


T seconds. Starting from the same initial state
time
function
by
First, shift the input
system equal y(i) of the original system?
y{t
the
shifted
+
T)
of
xo at time ta + T, does

One

Example

test for time-invariance is to

1.17.

Given the nonlinear differential equation

with

t;(6)

a.

Let r

= i6

so that

dr

dx

__

dt

dt

x{t

= 0)

a,

resulting in the

a;2

it2

were changed

dx
di

= t 6,

same system.

If the nonlinear equation for the state x

with the substitution t

m2

and

^
dr
where

k2

to

tx^

u^

m2

then

ra;2

6a;2

and the appearance of the last term on the right gives a different system. Therefore this is a timewill
varying nonlinear system. Equations with explicit functions of t as coefficients multiplying the state
usually be time-varying.

1.6

SYSTEMS CONSIDERED
This book will consider only time-invariant and time-varying linear dynamical systems

described by sets of differential or difference equations of finite order. We shall see in the
next chapter that in this case the state variable x(it) is an n-vector and the system is linear.
Example

1.18.

time-varying linear differential system of order n with one input and one output
,
equation
u^
+ l^n{t)u
+ a^{t)y = ^o(*)rf^ + ^
+

II

o.M^+

is

described by the
iiin\
(1.10)

MEANING OF STATE

Example

[CHAP.

1.19.

time-varying linear difference system of order n with one input and one output
the equation
y{k

The values
t

+ n) +

ai(k) y(k

+ n-l) +

of a{k) depend on the step

ajk)

y(k)

^^ik) u{k

of the process, in a

(fe)

way

+ n)+

/3(fc)

described by

is

u{k)

analogous to which the

{1.11)

a{t)

depend on

in the previous example.

1.7

LINEARIZATION OF NONLINEAR SYSTEMS


State space techniques are especially applicable to time-varying linear systems. In this
we shall find out why time-varying linear systems are of such practical importance.

section

Comparatively

little

design of systems

at present, but state space

methods

performed from the time-varying point of view


promise for the future.

is

offer great

Consider a set of n nonlinear differential equations of

=
-

dyi/dt
dy2/dt

dyjdt

first order:

.,yn,u,t)

fi{y 1,1/2,

..

/2(2/i,2/2,

...,yn,u,t)

/n (2/1,2/2, ...,yn,u,t)

nonlinear equation of nth order d^y/dt" = g{y, dy/dt,


., d''~'^y/dt''-'^, u, t)
be written in this form by defining 1/1 = V, dy/dt = 2/2,
., d''~'^yldt^-'^ - yn.
of n first order nonlinear differential equations can be obtained as
.

dyi/dt

1/2

dy2/dt

2/3

can always
Then a set

(i-l^)

dyn-i/dt

dyJdt
Example

g{yi,

2/2,

...

2/n,

u,

t)

1.20.

To reduce the second order nonlinear

differential equation

order nonlinear differential equations, define y

d?yldt^

yi and dy/dt

dyJdt

j/2

dy2/dt

2y\

y^.

2y^ + u dy/dt =

to

two

first

Then

wj/a

Suppose a solution can be found (perhaps by computer) to equations [1.12) for some
Denote this solution as the
2/2(^0),
., yn{to) and some input w{t).
trajectory ^{t; w{t), yi{to),
., yn{to), U). Suppose now that the initial conditions are changed:
initial conditions yi{U),

y{to)

where
to u(t)

yiito)

Xi{to), X2{to)

w{t)

dy
Xl{to),

and

Xn{to)

+ v(t) where

Jfito)

y2{to)

are small.

X2{to),

...,

d" ~^v
-^pr^ito)

Furthermore, suppose the input

v{t) is small.

To

yn{to)

is

+ x^/dt =

/j(^j

+ x^,

4,^

+ x^,

...,^^

+ x^,w + V,

t)

d{4,^

+ x^/dt =

f^ (^^

+ iCj,

4>2

+ X2,

.. .,<j,^

+ x^,w + v,

t)

<ii<t>n

+ ^nV^* =

fn (<Al

+ ^1'

<l>2

+ ^2'

+ ^n'

^)

'

'

'i>n

'^

^'
>

Xn(to)

changed slightly

satisfy the differential equations,

d{<j>^

CHAP.

MEANING OF STATE

1]

<^ and w using Taylor's theorem for sev...,/ can be expanded about ,j>^, <j>2,
terms
we obtain
higher
order
eral variables, then neglecting

If

/j, /j,

t+t

where each dfjdy^

is

. ,

i-+t- + - + S-+S"

/>...-.*..-.') +

the partial derivative of fiiv^.V^, ...,V^,u,

ated at 2/1 = <#.!, 2/2 = ^2'


equation, then 6,4, Jdt =

/j

=
l/n

M=

Now,

w.
^n ^^^^^
can be canceled to leave

'

t)

with respect to y^ evalu-

since each

Idfildyi

dfi/dyi

dfi/dyn\

3/2/32/1

3/2/33/2

3/2/3|/

^dfjdyi

3//32/2

...

<^j

/dfi/du\

/xi\

X2

satisfies the original

3/2/31*
\

\dfJdU/

dfjdyn/ \Xn/

which is, in general, a time-varying linear differential equation, so that the nonlinear equasmall
tion has been linearized. Note this procedure is valid only for Xi, X2, ...,Xn and v
matrix
The
neglected.
be
can
series
Taylor's
in
the
enough so that the higher order terms
is called the Jacobian matrix of the vector f (y, u, t).
of dfjdy. evaluated at y^ =
<^j.

Example

1.21.

at toJ= 1Consider the system of Example 1.20 with initial conditions j/(to) = 1 and y(t^ = -1
~*~^- ^^^'^^ fi - 3/2.
If the particular input w(t) = 0, we obtain the trajectories ^i(t) = t~^ and ^gC*) =
= 2j/? -M3/2, then dh/dyi = 6j/i, df^/dVi - - and
then a/i/aj/i - 0, 3/i/3j/2 = 1 and dfi/du = 0. Since
= 1 + *i(*o). ^(*o) = -1 + 2(*o) and inputs u - v(t),
=
Hence for initial conditions

5/2/aM
-j/2.
we obtain

y(o)

t-i and dy/dt


This linear equation gives the solution y(t) = i(t) +
small.
v
are
and
x^,
x^,
long
as
valid
as
is
equation, and

Example

= x^-t'^

for the original nonlinear

1.22.

Given the system dyldt = ky-v^ + u. Taking u(t) =


and Mt) = k. The equation for small motions x(t) about
and the equation for small motions x(t) about .^(t) = fc is

0,

we can

,/,(t)

= is
= -kx + u

da;/dt

two constant solutions <t>{t) dx/dt = kx + u so that y[t) = x(t),

find

so that

j/(t)

+ x(t).

10

MEANING OP STATE

[CHAPi

Solved Problems
1.1.

Given a delay

line whose output is a voltage input delayed T seconds. What is the


physical object, the abstract object, the state variable and the state space?
Also,
is it controllable, observable and a dynamical system with
a state space description?

The physical object is the delay line for an input u{t) and an output y(t) = u(t T). This
equation is the abstract object. Given an input time function M(t) for all
*, the output y{t) is
defined for t ^ to, so it is a dynamical system. To completely specify the output given
only u{t)
for t ^ to, the voltages already inside the delay line must be known. Therefore,
the state at
time to is x(to) = M[t-T,t) . where the notation u^^^f^y means the time function m(t) for t in the

tz-

interval

<

t^.

For

>

we

as small as

please, M[f(,_T,to)

'^^^

^^ considered as the

lin-

countably infinite set of numbers

{u(to-T), u(to-T

+ e),

u(to-.)}

...,

M[e_r,t,

x{to)

In this sense we can consider the state as consisting of an infinity of numbers.


variable is the infinite set of time functions
X(t)

The

M[(_r,0

Mt-T), u(t-T +
T

state space is the space of all time functions

breakdovra voltage of the delay

An

e),

...,

Then the

state

U(t-e))

seconds long, perhaps limited by the

line.

to-t<to+T

input u(t) for


will completely determine the state T seconds later, and.
state will be observed in the output after T seconds, so the system is both observable and controllable.
Finally, x(t) is uniquely made up of x(t t) shifted r seconds plus the input over t seconds, so that the mathematical relation
y(t) = u{t
T)
gives a system with a state space
description.

any

1.2.

^ ^

Given the uncontrollable partial differential equation (diffusion equation)

with boundary conditions


The

2/(0, i)

1/(1, i)

solution to this zero input equation for

0(.,,)

What

0.
t

to

the state variable?

is

is

CO

y{r,t)

where

y(r, to) sin nirr dr.

ce-^'^^"'o'

smmrr

All that is needed to determine the output

is v(r, tj),

so that

interval

a choice for the state at any time t. Since y{r, t) must be known for almost all r in the
r 1, the state can be considered as an infinity of numbers similar to the case of

Problem

1.1.

y{r,

1.3.

t) is

Given the mathematical equation


with a state space description?

(dy/dty

y^

+ Ou.

Is

this a

dynamical system

A real output exists for all t to for any u{f}, so it is a dynamical system. The equation
can be written as dy/dt = s{t)y, where s(t) is a member of a set of time functions that take on
the value +1 or 1 at any fixed time t. Hence knowledge of y(to) and s(t) for t - to uniquely
specify y(t) and they are the state.

1.4.

Plot the trajectories of

_
~

dt^

Changing variables

to

2/

Ki

\-y

if

idy/dt-y^l)^

if

{dy/dt-y<l)j

and dy/dt

X2 dx^ldxi

X2,

or

the mathematical relation becomes


X2 dx^ldx-^

= x^

CHAP.

MEANING OF STATE

1]

11

The former equation can be integrated immediately to x^it) = a;2(to)> a straight line in the phase
The latter equation is solved by multiplying by dxi/2 to obtain ^3:2/2 + Xi dxi/2 = 0.
This can be integrated to obtain a;|(t) + xl(t) = xlit^) + xl(to). The result is an equation of circles
The straight lines lie to the left of the line 9;2 ~ *'i ~ ^ ^^^ *h circles to
in the phase plane.

plane.

the right, as plotted in Fig. 1-5.

Fig. 1-5
(positive velocity) and decreases for X2 < 0, giving the motion
Note xi increases for aig >
of the system in the direction of the arrows as t increases. For instance, starting at the initial
conditions Xi(to) and 2(*o) corresponding to the point numbered 1, the system moves along the
outer trajectory to the point 6. Similarly point 2 moves to point 5. However, starting at either
point 3 or point 4, the system goes to point 7 where the system motion in the next instant is not
determined. At point 7 the output y(t) does not exist for future times, so that this is not a

dynamical system.

1.5.

Given the electronic device diagrammed in Fig. 1-6, with a voltage input u{t) and a
For to~t< ti, the
voltage output i/(i). The resistors R have constant values.
switch S is open; and for t^ti, S is closed. Is this system linear?

Fig. 1-6

the state. No other


Referring to Definition 1.7, it becomes apparent the first thing to do is find
is no state, i.e. the
there
so
input,
given
the
output
information is necessary to determine the
analogous to Example 1.1, except
dimension of the state space is zero. This problem is somewhat
time.
instant
of
each
at
that the position of the switch is specified
for all time, we only need assume two inputs u^)
is linear, since x{t) =
and j/i(t) = ^i(t)/3 for t ^ t,. Siniilarlyj/ajt)
=
t,^t<t,
for
u,(t)/2
and
is [au,(t) + pu2{t)]/2 for
output
The
pu^it).
+
au^it)
input
Now assume an
uJt)/2 or uM)/S.
Substituting y^{t) and y^it), the output is ai/x(t) + PV^^t),
t ^ t < t, and faM,(t) + Pu^miZ for t > t,.
the system is linear. The switch S can be considthat
and
showing that superposition does hold
zero for t ~ t^. Therefore
time-varying resistor whose resistance is infinite for t < h and

To

see

u,{t).

if

the system

Then

y,{t)

ered a

Fig. 1-6 depicts a time-varying linear device.

MEANING' OF STATE

12

1.6.

[CHAP.

Given the electronic device of Problem 1.5 (Fig. 1-6), with a voltage input u{t) and
a voltage output y{t). The resistors R have constant values. However, now the. position of the switch S depends on y(t). Whenever y{t) is positive, the switch
S is open;
and whenever y{t) is negative, the switch S is closed. Is this system linear?
Again there is no state, and only superposition for zero state but nonzero input need
be
investigated.
The input-output relationship is now
Vit)

[5u{t)

u(t)

sgn m(0]/12

where sgn m = +1 if u is positive and -1 if u is negative.


Given two inputs u^ and u^ with
resultant outputs y^ and y^ respectively, an output y with an input Mg =
au^ + pu^ is expressible as
y

To be

linear,

ay

^2/2

[5(aMi

+ Pu^) +

(aMi

+ pu^

must be equal y which would be true only

aui sgn Ml

pu2 sgn Mg

(om^

This equality holds only in special cases, such as


system is not linear.

1.7.

sgn {au^

+ PU2)

sgn Mj

+ Pu^)]/X2.
if

sgn (au^

sgn u^

+ pu^)
sgn (au^

+ pu^),

so that the

Given the abstract object characterized by


y{t)

Is this

xoe'"'

+ f

e^''u{T)di

time-varying?

This abstract object is that of Example 1.2, with


1 in equation (1.1).
By the same procedure
used in Example 1.16, it can also be shown timeinvariant. However, it can also be shown time-invariant by the test given after Definition 1.8. The input
time function u{t) is shifted by T, to become w(t).
Then as can be seen in Fig. 1-7,

RC =

= u(t-T)

u{t)

Starting from the same initial state at time


y(a)

Let

r-<.
a;oe'o +

= T-

t^

T,

Original System

e^-''u{r- T) dr
J"
to+T

T:

T-T
y{a)

a;oe'+'^

"

ei + 'r~'^u(i) di

j
to

Evaluating ^ at

y(t+T)

a= t+T
=

gives

a;oe*-'

+ J'ef-Ml)d{
'o

which

is identical vidth

the output

<o

y{t).

+r

Shifted System
Fig. 1-7

t+r

CHAP.

MEANING OF STATE

1]

13

Supplementary Problems
IX

Given the sprinir-niass system shown in Fig. 1-8. What is the


physical object, the abstract object, and the state variable?

M
1.9.

Given the hereditary system

y{t)

K{t, r) m(t) dr

is some single-valued continuously


and T. What is the state variable?
the system time-varying?

K{t,T)
of
Is

1.10.

and the state at step


1.11.

differentiable function
Is the

Given the discrete time system x{n


3, a;(3).

where

system linear?
Fig. 1-8

+ 1) =

x{n)

+ u{n),

the series of inputs m(0), m(1),

Find the state variable x{m) at any step

u(k),

m 0.

abstract object is characterized by y(t) = u(t) for t(, t< ti, and by dy/dt = du/dt for t tj.
given that this abstract object will permit discontinuities in y(t) at tj. What is the dimension
of the state space for t^ t < ti and for t ti?

An

It is

(i.jg) of equation (1.1), and then verify the solution (1.S) of equation
Finally, verify the solution Xi + Zx^ = C(2xi + W2)* of Example 1.9.

1.12.

Verify the solution

1.13.

Draw

1.14.

the trajectories in two dimensional state space of the system y

+y

(i.4).

0.

Given the circuit diagram of Fig. l-9(a.), where the nonlinear device NL has the voltage-current
shown in Fig. 1-9(6). A mathematical equation is formed using i = Cv and v = /(i) = f{Cv):

relation

where /"'

(l/C)/-i(v)

the inverse function. Also, v(to) is taken to be the initial voltage on the capacitors.
Is this mathematical relation a dynamical system with a state space description?
is

(a)

Fig. 1-9

+1 =u

a dynamical system?

1.15.

Is the

mathematical equation y^

1.16.

Is the

system dy/dt

t^y

1.17.

Is the

system dy/dt

1/y time-varying?

1.18.

Verify that the system of Example 1.16

1.19.

Show equation

1.20.

Show
j

1.21.

equation

0, 1,

time-varying?

Is it linear?

Is it linear?

is

nonlinear.

(1.10) is linear.

(1.10)

is

time-invariant

if

the coefficients

.,n, are not functions of time.

=Xi + 2,

Given dxi/dt
(a)

Does

this

(6)

What

is

(e)

Is the

dxi/dt

x^

+ u,

Xi

+ X2-

system have a state space description?

the input-output relation?

system linear?

on

and ^j for

1, 2, ...,7i

and

MEANING OF STATE

14

[CHAP.

1.22.

What is the state space description for the anticipatory system


tion 5 for dynamical systems is violated?

1.23.

Is the

1.24.

What

1.25.

Is the equation

system dx/dt

is

e*u,

f(t)u

The physical object is the spring-mass system, the abstract


and the state variable is the vector having elements x{f) and

1.9.

It is not possible to

in

which case

Vih)

toT for
< t <

K{t,

the delay

line.)

The system

t.

t) u(t) dr,

x{k)

x{S)

f(t)

for

to

*i

is

t)

K(o, t)

For a general K{t, r), the state at time


linear and time varying, unless K{t,T)

1.11.

for

fe

4, 5,

and

x(k)

x{3)

- 2

T),

m(3

- 1)

for

A;

1, 2, 3.

Note we

i=l
x{i) will

be the

.n.

Because
is zero for to t < ti, and one-dimensional for t ti.
time-dependent in general, it must be a family of sets for each time t. Usually
it is possible to consider a single set of input-output pairs over all t, i.e. the state space is timeThis
Abstract objects possessing this property are called uniform abstract objects.
invariant.
problem illustrates a nonuniform abstract object.

The dimension

of the state space


is

1.12.

Plugging the solutions into the equations will verify them.

1.13.

The

trajectories are circles.

It is

a dynamical system.

However, care must be


It is a dynamical system, because v{t) is real and defined for all t - t^.
taken in giving a state space description, because f~^(v) is not single valued. The state space
description must include a means of determining which of the lines 1-2, 2-3 or 3-4 a particular
voltage corresponds to.

1.15.

No, because the input u

1.16.

It is linear

1.17.

It is

1.21.

must be
K{t

3-fc

2 m(z)
t=3

the state space

1.14.

time-invariant.

it is

i =^ 0, 1,

only?

Mx

need not "tie" ourselves to an "initial" condition because any one of the values
state for

which only con-

object is all x obeying


+ kx = 0,
dx/dt. This system has a zero input.

unless K(t,

fc-l
1.10.

in

represent

y(t)

(This is true for

T)

Supplementary Problems

to

1.8.

taken as u(t) for

du/dtl

a dynamical system given values of

Answers

u{t

e-*x time-varying?

the state space description for the differentiator y

y(t)

<

results in

an imaginary output.

and time-varying.

nonlinear and time-invariant.

(a)

Yes

(6)

y(t)

(c)

No

e'-'[*i(*o)

+ a'2o)]+J
'"

e*-^[M(r)

+ 2]dr

MEANING OF STATE

CHAP.

1]

1.22.

No

15

additional knowledge is needed other than the input, so that state is zero dimensional and the
= u(t + T). It is not a dynamical system because it is not realizable
physically if u{t) is unknown in advance for all t to.
However, its state space description
state space description is y(t)

violates only condition 5, so that other equations besides dynamical systems can
space description if the requirement of causality is waived.

1.23.

Yes.

y{t)

e-txo

have a state

e^""""

to

on when the system

1.24.

If

which

X,

we

is

is started.

If

Xq

0,

the system

is

equivalent to

dx/dt

= x + u

where

time-invariant.

define

du/dt

lim [u(t

+ e) u(t)]/e

c->0

so that

y(t(,) is

y{t) for all t

tg.

Other definitions of du/dt

which would be the state

1.25.

may

in that case.

not defined for t > ti, so that as stated the equation is not a dynamical system.
the behavior of engineering interest lies between tg and ti, merely append y = Ou for
to the equation and a dynamical system results.

Obviously

However,

tti

and knowledge of u(t) determines


require knowledge of u(t) for fg e t to,

defined, then the state space is zero dimensional

y(t) is

if

chapter 2
Methods

Obtaining
the State Equations

2.1

for

FLOW DIAGRAMS

Flow diagrams are a simple diagrammatical means of obtaining the state equations.
Because only linear differential or difference equations are considered here, only four basic
objects are needed. The utility of flow diagrams results from the fact that no differentiating devices are permitted.
Definition 2.1:

A summer
.

.,

a diagrammatical abstract object having n inputs


Un(t) and one output y{t) that obey the relationship
is

y(t)

where the sign

is

Ui{t)

U2{t)

Ui{t),U2{t),

Un{t)

positive or negative as indicated in Fig. 2-1, for example.

Mj(<)

M2(t)

y(t)

M()-

Fig. 2-1.

Definition 2.2:

a diagrammatical abstract object having one input u{t) and one


such that the input is scaled up or down by the time function a{t)
as indicated in Fig. 2-2. The output obeys the relationship y{t) = a{t) u{t).
scalar

output

is

y{t)

-Um-

()

Fig. 2-2.

Definition 2.3:

Summer

An

integrator

one output

y(t)

Scaler

a diagrammatical abstract object having one input u{t),


and perhaps an initial condition y{to) which may be shown

is

y{t),

or not, as in Fig. 2-3.

The output obeys the


y{to)

vit)

+ J

relationship

u{T)dr

9 v(*o)

^y(t)

u(t)-

Fig. 2-3.

16

Integrator at Time

CHAP.

METHODS FOR OBTAINING THE STATE EQUATIONS

2]

Definition 2.4:

17

delayer is a diagrammatical abstract object having one input u{k), one


output y(k), and perhaps an initial condition y(l) which may be shown or
not, as in Fig. 2-4. The output obeys the relationship
y{j

+ + l) ^

u{j

+ l)

for

0,1,2, ...

v{D

*y(k)

u(k)-

Delayer at Time k

Fig. 2-4.

2.2

PROPERTIES OF FLOW DIAGRAMS


Any

of the

set of time-varying or time-invariant linear differential or difference equations


{1.10) or (1.11) can be represented diagrammatically by an interconnection of

form

the foregoing elements. Also, any given transfer function can also be represented merely
by rewriting it in terms of (1.10) or (l-ll). Furthermore, multiple input and/or multiple

output systems can be represented in an analogous manner.


Equivalent interconnections can be made to represent the same system.
Example

2.1.

Given the system


dy/dt

with

initial condition j/(to)-

u{t)-

An

ay

au

interconnection for this system

(2.i)

is

shown

<?
+

in Fig. 2-5.

y(to)

^yit)

Fig. 2-5

Since a

is

a constant function of time, the integrator and scaler can be interchanged

tion is adjusted accordingly, as

shown

if

the initial condi-

in Fig. 2-6.

y(to)/a

u(t)-

tO
-i-

y(t)

Fig. 2-6

is

This interchange could not be done if a were a general function of time.


by parts to accomplish this interchange. If a(t)

possible to use integration

grated to

-t
y{t)

y(to)

t[j/(t)
to

+ u(r)]

dr

In certain special cases

t,

then

(2.1)

it

can be inte-

METHODS FOR OBTAINING THE STATE EQUATIONS

18

[CHAP. 2

Using integration by parts,

= J/(*o)-J

V{t)

+ u(i)]didT + tC

[y{0

[v{r)

+ u(t)]

dr

to

*o

which gives the alternate flow diagram shown in Fig.

2-7.

<t)

Fig. 2-7

Integrators are used in continuous time systems, delayors in discrete time (sampled
Discrete time diagrams can be drawn by considering the analogous continuous time system, and vice versa. For time-invariant systems, the diagrams are almost
identical, but the situation is not so easy for time-varying systems.
data) systems.

Example

2.2.

Given the discrete time system


y(k

+ l+l) =

ay(k

l)

au{k

{2.S)

l)

The analogous continuous time systems is equation {2.1), where d/dt takes the
initial condition y(1).
place of a unit advance in time. This is more evident by taking the Laplace transform of {2,1),

with

sY{s)-y{to)

and the

2 transform of

{2.2),

zY{z)
zY{

Hence from Fig.

= aY{s)+aU{s)

zy{l)

aY{z)

aU{z)

diagram for {2) can be drawn immediately as

2-5 the

in Fig. 2-8.

>y{l)

u{k

+ 1)

* y{k +

tO

1)

Fig. 2-8

If the initial condition of the integrator or delayor is arbitrary, the output of that
integrator or delayor can be taken to be a state variable.

Example

The

2.3.

state variable for {2.1) is y{t), the output of the integrator.


y{t)

Note

2/(to) is

Example

To verify

this, the solution to

equa-

^t

tion {2 J.) is

the state at

to.

J/(to)

e""-''

+ aj

e-^>M(r)dr

so the state variable is y{t).

2.4.

The state variable for equation {2.2)


manner similar to the previous example.

is

y{k

+ 1),

the output of the delayor.

This can be veriiied in a

CHAP.

METHODS FOR OBTAINING THE STATE EQUATIONS

2]

Example

19

2.5.

From
first

Fig. 2-7, the state is the output of the second integrator only, because the initial condition of the
integrator is specified to be zero. This is true because Fig. 2-7 and Fig. 2-5 are equivalent systems.

Example

2.6.

summer

or a scalor has no state associated with

it,

because the output

is

completely determined by

the input.

2.3

CANONICAL PLOW DIAGRAMS FOR TIME-INVARIANT SYSTEMS

Consider a general time-invariant linear differential equation with one input and one
output, with the letter v denoting the time derivative didt. Only the differential equations

need be considered, because by Section 2.2 discrete time systems follow analogously.

+ "iP""*!/ +

p"2/

a_iP2/

a2/

i8p"M -f

pjy-^u

-!--+- i8_jPM

+ Pji

{2.3)

This can be rewritten as

PHV -

+ P""'(i2/ -I3^u)+

because a.p^-^y = 'p^~\y, which


and rearranging gives

is

not true

^(^, -,!/)

from which the flow diagram shown


at the right and working to the left.

Pia^_,y

if a^

- ^_itt) +

a2/

depends on time.

)8M

Dividing through by.p"

+ ~=iW-^u-a_,y) + ^(J3u-aj)

in Fig. 2-9 can be

drawn

{24)

starting with the output y

"-vit)

Fig. 2-9.

The output

Flow Diagram

of each integrator

The summer equations for the

is

of the First Canonical

labeled as a state variable.

state variables have the

X,

Form

x^

form

+ /?M

-a^y

iCj

+ p^u
{2.5)

METHODS FOR OBTAINING THE STATE EQUATIONS

20

Using the

first

equation in

{2.5) to

X2

/-a.

dt

Xn-lj

\-

form

\^

eliminate y, the differential equations for the state vari-

ables can be written in the canonical matrix

(xr

[CHAP.

^ /^^

^i-A

X2

(2.6)

a;n-i

'1

0/ \x.

^-^o

We will call this the first canonical form.


the first column of the
vector

nxn matrix.

Note the Is above the diagonal and the a a down


Also, the output can be written in terms of the state

(Xt
X2

{2.7)

j8tt

Xn

Note

this

form can be written down

directly

from the

original equation {2.3).

Another useful form can be obtained by turning the first canonical flow diagram "backwards." This change is accomplished by reversing all arrows and integrators, interchanging summers and connection points, and interchanging input and output. This is a heuristic
method of deriving a specific form that will be developed further in Chapter 7.

Q-^M)

Fig. 2-10.

Flow Diagram of the Second Canonical (Phase-variable) Form

Here the output of each integrator has been


variables are

relabeled.

The equations for the

state

now
x^

Xg

*3

x^

^n

^ -A~2"'n-l

{2.8)

CHAP.

METHODS FOR OBTAINING THE STATE EQUATIONS

2]

In matrix form,

(2.8)

may

21

be written as
1

X2

1
(2.9)

dt
3/ji

Xn

*n-2

and

{2.10)

Xn-l

This will be called the second canonical form, or phase-variable canonical form. Here the
Is are above the diagonal but the as go across the bottom row of the nxn matrix. By
eliminating the state variables x, the general input-output relation {2.3) can be verified.
The phase-variable canonical form can also be written down upon inspection of the original
differential equation {2.3).

JORDAN FLOW DIAGRAM

2.4

The general time-invariant linear differential equation


put can be written as

By

{2.3)

for one input and one out-

dividing once by the denominator, this becomes

the case where the denominator polynomial factors into distinct poles \i,
Distinct means \i = Xj for i - j, that is, no repeated roots. Because most
practical systems are stable, the Ai usually have negative real parts.

Consider

i=

1,2,

first

.,n.

2?"

+ ttjP"-! +

partial fraction expansion can

Here the residue

pj

{^,
Pi

The

^o

+
,

+ a_jip + =

- Aj)(p -\^---{v- A)

now be made having

+ ^rri^ +
,

(;2.15)

the form
Pn

''2

Pi

^^^

+ ^-3r
,

{2.U)

can be calculated as

- a,P,)\r' + {P, ~ ^^o)^r^ + + (^_i - -i^o)\ +(^n-n^o)


(A. - A)
(A, - A,_,)(A, - A,^
(\ - Aj)(\ - A,)
J

(2.15)

shown

in Fig.

partial fraction expansion

2-11 following.

(p

{2.14^)

gives a very simple flow diagram,

METHODS FOR OBTAINING THE STATE EQUATIONS

22

[CHAP.

M(t).

-j/()

-^O-[>

rOv^*~r~G)

L-9-D^

Note that because

Jordan Flow Diagram for Distinct Roots

and

can be complex numbers, the states


assume the simple form

pj

The

functions of time.

Fig. 2-11.

\^

are complex-valued

x.

state equations

^2

K^2 + ^
(;g.ie)

Consider

Jordan

now

Aa;

^o

extended to the general case.

ajP"-!

++_!?)

Pi

^0^

+ ^p-xtY +

The residues

Pi^'i

P2^2

pna;

For simplicity, only one multiple root (actually one


page 73) will be considered, because the results are easily
Then the denominator in {2.12) factors to

instead of {2.13). Here there are


sion for this case gives

the general case.

block, see Section 4.4,

2)"

+u

= {p-\y{p-K^{)---{p-Xn)

{2.17)

Performing the partial fraction expan-

identical roots.

p^U

p^U

Py + iU

{p-xi)"-'

p-xi

P-X.+1

P- kn

'

at the multiple roots can be evaluated as

Pfc

.,

v/^-^v
^[|^<'-^'H.=H
.V.

/vy;

,,,._,

where f{p) is the polynomial fraction in


in Fig. 2-12 following.

p from

{2.12).

"=''
IV-

j.,^,

.,^

{2.19)

This gives the flow diagram shown

CHAP.

2]

METHODS FOR OBTAINING THE STATE EQUATIONS

Fig. 2-12.

The

23

Jordan Flow Diagram with One Multiple Root

state equations are then

Xz

A.ia;2

Xa

Xv + l

=
=
=

Av + iX^ + i

PoU

Xv-l
Xv

AiXy-i
AiXi.

Xv

+u

{2.20)

+u

+ PiXi +

92Xz

4- PnaJn

METHODS FOR OBTAINING THE STATE EQUATIONS

24

The matrix

Xi

/a;i

Ai

00

Xv-l

dt

with this Jordan form are

diiferential equations associated


Xi

Xv

...

Alio

...

...

OXiO

...

00

0...

Xv + l

\xn

[CHAP. 2

/ a;2

Xv-l

...

Xv + i

A+i ...

...

l'\

(2M)

\ ccn

Xi

and

P2

(Pl

Pn)

*'
I

Xn

row

In the w X M matrix, there is a diagonal


the other X's follow on the diagonal.
Example

Po

of ones above each

A,i

on the diagonal, and then

2.7.

Derive the Jordan form of the differential system

V+2y + v =
Equation

(2.2S)

can be written as

j/

T^qrjT^w

(P

Figure 2-13

is

+ 1)2

+M

(SS)

whose partial fraction expansion gives

'

then the flow diagram.

u(t)^^*Q

.p^

Xl

^E>

KE>

Fig. 2-13

Because the scalor following


Jordan form are

x^,

is zero,

this state is unobservable.

(0

1)

\X2

The matrix

state equations in

CHAP.

METHODS FOR OBTAINING THE STATE EQUATIONS

2]

25

TIME-VARYING SYSTEMS

2.5

Finding the state equations of a time-varying system is not as easy as for time-invariant
systems. However, the procedure is somewhat analogous, and so only one method will be
given here.

is

The general time-varying differential equation of order n with one input and one output
shown again for convenience.

S + i(*)^ +

+ (%

p,{t)^ +

p^{t)u

{1.10)

Differentiability of the coefficients a suitable number of times is assumed. Proceeding in


a manner somewhat similar to the second canonical form, we shall try defining states as in

However, an amount yH)

{2.8).

'^i

be determined] of the input

[to

a;2

yi(t)M

-)'2(*)

u{t) enters into all the states.

{2.2A)

By

x^

y^{t)u

n times and using the relations for each state, each of the
In the general case.

differentiating y

can be found.

Example

unknown

y.

2Ji.

Consider the second order equation

S+
Then by

l(*)^

l^oit)-^

+ Pi(t)^ +

P^(t)u

(2.26)

(2.24)

and

2(*)3/

xi

7o(*)m

(2.27)

(2.28)

differentiating,

y =

Xi

yo(t)u

yo(t)u

Substituting: the first relation of (2.24) into (2.28) gives

= X2+

[yi(*)

+ ro(*)]M +

(2.29)

yo(*)M

Differentiating again,

y =

From

(2.24)

*2

[ri()

[7i(*)

+ 2yo(t)]u +

yo(t)

(2.30)

we have
X2

Now

+ "o(<)]m +

-ai(t)x2

a2(t)xi

y2(t)u

(2.31)

substituting (2.27), (2.29) and (2.30) into (2J1) yields

yi()

+ 'Um^ -

+ 2yo()] u - y^) M
= -ai(t){y - [7i() + yo(*)]M - 7o(t)u} [ri)

oi2(t)[y

- yo(t)u] +

y2(t)u

METHODS FOR OBTAINING THE STATE EQUATIONS

26

Equating

and

coefficients in (2.26)

(2.SS),

72

Vi

+ Vo)i +
+ 2yo + ai7o =
7o - Po

Vo

71

Substituting (2.35) into

[CHAP. 2

(7i

yo2

(2.SS)

P2

{2M)

^1

(2.S5)

(2.Si),

(2M)
and putting

(2.55)

and (2J6)

into

72

Using equation

(^.;2^),

(;2.55),

fio

-k + ^2 +

(o'lPo

+ 2;8o " ^i)i +

j8oi

(2.S7)

jSoaz

the matrix state equations become


1
1

u
1

...

2/

(1

{2.38)

0)

y,{t)u

"n-l

2.6

GENERAL STATE EQUATIONS

Multiple input-multiple output systems can be put in the same canonical forms as single
input-single output systems. Due to complexity of notation, they will not be considered here.
The input becomes a vector u(i) and the output a vector y(i). The components are the inputs
and outputs, respectively. Inspection of matrix equations {2.6), {2.9), {2.21) and {2.38)
indicates a similarity of form. Accordingly a general form for the state equations of a
inputs and k outputs is
linear differential system of order n with

dyjdt

where

x(t)

is

an w-vector,

u(i)

is

an m-vector,

y(i)

is

A(f)x

B(t)u

C(t)x

D(t)u

{2.39)

fc-vector,

A.{t) is

an X w matrix,

B(<) is

an

TO

matrix,

C{t) is a

A;

X matrix,

D(f) is a

A;

In a similar

=
=

matrix.

manner a general form for


x(

where the dimensions are

+ 1) =
y{n) =

discrete time systems is

A(w) x(m)
C{n) x(%)

+ B(m) u(n)
+ D(tc) u{n)

also similar to the continuous time case.

{2.i0)

CHAP.

METHODS FOR OBTAINING THE STATE EQUATIONS

2]

Specifically, if the

system has only one input u and one output

y,

27

the differential equa-

tions for the system are

dx/dt

+ b(t)M

A(i)x

ct(it)x

d(i)M

and similarly for discrete time systems. Here c(i) is taken to be a column vector, and
denotes the complex conjugate transpose of the column vector. Hence ct(i) is a row
vector, and c+(*)x is a scalar. Also, since u, y and d{t) are not boldface, they are scalars.

ct()

Since these state equations are matrix equations, to analyze their properties a knowledge
of matrix analysis is needed before progressing further.

Solved Problems
2.1.

Find the matrix state equations

in the first canonical

form for the

linear time-

invariant differential equation

with

initial conditions

y{0)

5y

y{0)

yo,

+
=

6y

ii

+u

Also find the

yo.

(2.4.1)

initial conditions

on the

state variables.

Using p = d/dt, equation (241) can be written as p^y


and rearranging,
y

The flow diagram

+ 5py + 6y pu + u.

Dividing by p2

= piu-5y) + -^(u-ey)

of Fig. 2-14 can be

drawn starting from the output

at the right.

Fig. 2-14

Next, the outputs of the integrators are labeled the state variables xi and X2 as shown.

an equation can be formed using the summer on the


X2

Similarly, an equation can be formed using the

Xi

Also, the output equation

is

x^.

62/

summer on

X2

5y

Now

left:

the right:

+u

Substitution of this back into the previous equations gives


Xi

5xi

X2

-exi

+u

X2

u
(242)

METHODS FOR OBTAINING THE STATE EQUATIONS

28
The

[CHAP. 2

state equations can then be written in matrix notation as

o)G)

-6

(l^'"

with the output equation

The

on the state variables must be related to


The output equation is a;i(t) = ?/(<), so that a:j(0)

Xi(t)

into (2.42)

and setting

Use of the given

initial conditions

-5j/(0)

and

Vo, the

i/(0)

2/.

given output initial


Also, substitutmg

052(0)

M(0)

determines

- yo+

a;2(0)

These relationships for the


at time t = 0.

j/o

gives

y{0)

2.2.

X2

initial conditions

conditions.
y(t)

(10)

5j/o

"(0)

can also be obtained by referring to the flow diagram

initial conditions

Find the matrix state equations in the second canonical form for the equation
of Problem 2.1, and the initial conditions on the state variables.
The flow diagram
diagram of Fig. 2-15.

y~*-

problem

(Fig. 2-14) of the previous

is

{2..i.l)

turned "backwards" to get the flow

Fig. 2-15

The outputs of the integrators are labeled x^ and X2 as shown. These state variables are different from those in Problem 2.1, but are also denoted x^ and X2 to keep the state vector x(t) notation, as is conventional. Then looking at the summers gives the equations

y
X2

Furthermore, the input to the

+ X2
5a;2 + u

Xi

6x1

(245)

X2

This gives the state equations

n/a^i
-6

dt \X2

The

and

its

initial conditions

(2.U)

left integrator is

Xi

and

(243)

-5/U2
(1

1)

X2

are found using (243),


j/o

a;i(0)

a;2(0)

Vo

a;i(0)

X2(0)

derivative

CHAP.

METHODS FOR OBTAINING THE STATE EQUATIONS

2]

Use

of (244)

and

(2.i5)

then gives
Vo

2(0)

6a!i(0)

Equations (248) and (247) can be solved for the

2.3.

29

a;i(0)

_=

-22/0

^2(0)

Byo

5a;2(0)

m{0)

(247)

initial conditions

i^o

^"(0)

+ ^yo-^u(0)

Find the matrix state equations in Jordan canonical form for equation (241) of Problem 2.1, and the initial conditions on the state variables.
The transfer function

is

p^

5p

(p

p + 1
+ 2)(p + S)'

partial fraction expansion gives

-1

From

this the flow

p+2

p+S

diagram can be drawn:

-H -1

O^
+

6-

X2

-H

<3H
Fig. 2-16

The

state equations can then be written

A pi
\X2

The

equalities at each

i-:)::ha-

dt

From

from the

(-1

2)

the output equation and its derivative at time


j/o

2a!2(0)

Vo

2i2(0)-ii(0)

and

state equation is used to eliminate i(0)

^0

Solving these equations for

a;i(0)

and

to,

2a!i(0)

0:2(0)

a;i(0)

2(0):

6x2(0)

M(0)

gives

!ri(0)

m(0)

^2(0)

i[M(0)-2j/o-^o]

3^0

Vo

summer:

METHODS FOR OBTAINING THE STATE EQUATIONS

30

2.4.

[CHAP.

Given the state equations

-l-lX^-i'."

dt \X2

y
Find the

1)

(1

differential equation relating the input to the output.

In operator notation, the state equations are

pxi

X2

px2

6xi

Xi

X2,

5py

Eliminating Xi and X2 then gives


p^y
This

is

tion in

2.5.

6y

5X2

+w

= pu + u

equation (S41) of Example 2.1 and the given state equations were derived from this equa2.2.
Therefore this is a way to check the results.

Problem

Given the feedback system of Fig. 2-17 find a state space representation of this
closed loop system.

R{s)-

()

C{s}

= TTi

W = JT3
Fig. 2-17

The transfer function diagram is almost in flow diagram form already. Using the Jordan
canonical form for the plant G(s) and the feedback H(s) separately gives the flow diagram of
Fig. 2-18.

r).
-1-

-1-

*c(t)

4-

Fig. 2-18

Note G{s) in Jordan form is enclosed by the dashed lines. Similarly the part for H(s) was drawn,
and then the transfer function diagram is used to connect the parts. From the flow diagram,

dt

Z) ^

{k

Xi

-8)(S) +

G)'(*^'

'^*^

^"^

0)

x%l\

CHAP.

2.6.

METHODS FOR OBTAINING THE STATE EQUATIONS

2]

Given the

linear, time-invariant, multiple

+ 2) +

V^{n

+ 1) +

a^y^in

Vln +
Put

this in the

1)

y^v.J,n

+ 1) +

y^yj^n)

(i^u^(n)

^^u^(n)

y^yln)

a^v^{n

+ 1) +

cc^y^{n)

^^u^n)

h^u^n)

form

+ 1) =

Ax(n)

Bu(n)

Cx(n)

Du(n)

y(w)

input-multiple output, discrete-time system

a^v^{n)

x(w

Where

31

= (^.),

<,

= (.

The
tions

first canonical form will be used.


Putting the given input-output equations into z opera(analogous to p operations of continuous time system),

z^Vx

Dividing by

a-^zvx

a22/l

li^Vl

2/2

Yi2/2

ffsZl/i

732/2

jSiiti

-|-

SiMg

42/i

ySjMi

-|-

S2M2

and respectively and solving for y^ and


2/1

2/2

-3l/i

(-l3/i

- 722/2) +
;

(/32M1

;5(/3iMi

S2M2

2/2,

+ SiM2-22/i- 732/2)

- 7i2/2 - a42/i)

Starting from the right, the flow diagram can


be

drawn as shown

in Fig. 2-19.

-2/2

Fig. 2-19

fn

t^^ T^A.

*^"

^''"^ delayers with arbitrary Initial conditions are not


needed because a

METHODS FOR OBTAINING THE STATE EQUATIONS

32

X2(m+1)
^Xgin

ri3
r33

-04

"Yl

^2W

)(

+ 1)/

\y2in)J

2.7.

\-a2 +

[CHAP.

Write the matrix state equations for a general time-varying second order discrete time
equation, i.e. find matrices A{n), B{n), C{n), D{n) such that
x(n

+ 1) =

A(n) x(n)

B(n) u(n)

C(tc) x(n)

D(n) u(n)

/8(n)

y{n)

(^.^S)

given the discrete time equation


y{n

+ 2)

4-

j(n) i/(n

+ 1) +

a^{n) y{n)

w(

+ 2) +

u{n

fi^{n)

+ 1) +

pj^n) u{n)

{249)

Analogously with the continuous time equations, try


Xi(n
x^in

1)

(2.52)

up one and substituting


y(n

Stepping
y(n

(2.53)
2)

Substituting
gives

+ 1) =

a2()

(2.5;2), (^.55), (2.5Ji.)

02(11)

xi(n)

yi(n) u{n)
Xi(n)

(2.50)

+ Y^M "W

(^^^)

Yo(w) u(n)

(2.52)

(2.50) gives

X2(n)

up one and substituting

= -aiW XaW -

x^in)

a^ivL) X2(n)

y(n)

Stepping

1)

yi(w) u(n)

Vo('^

1)

u(n

1)

1)

M(n

1)

(2.53)

(2.51) yields

iW + y2W "(w)

into (2.49)

yi(w

4-

and equating

ya(n

coefficients of u(n),

+ 2) M(n + 2)

u(n

1)

(2.54)

and u(n

+ 2)

VoW = ;8o('*-2)
ViW = ^i(m-l) -ai(n)/3o(M-2)
y2W = ;82W - iW /8i(w - 1) + [ai()ai(n - 1) - ff2W]/3o(w " 2)
In matrix form this

is

xi(n

l)\

a;2(n+l);

y(n)

2.8.

V-2W
=

(1

\fxi(n)\

-i(w)/ Va;2()/

0)(^'|^|)

"^

/yiW\
VYaW/"*"^

ro{n)u(n)

Given the time-varying second order continuous time equation with zero input,
y

a^(t)y

write this in a form corresponding to the

a^{t)y
first

/-a^{t)

canonical form

-aJt)

{y,(<)

(2.55)

ymll")

(2.56)

{2.57)

CHAP.

METHODS FOR OBTAINING THE STATE EQUATIONS

2]

To do

this, differentiate

{2.57)

and substitute for i and ^2 from


(Yi

- Tiffi -

Y2a2)i

Differentiating again and substituting for x^ and

ajj

(Yi

33

(2.56),

+ y2)a;2

(2.5S)

as before,

Vl - 2aiyi otiYi a2Y2 - 2a2y2 " 2Vl + ai2y2 + a^Yi)*! +

(2Yi

- iYi - a2Y2 + Y2)2

Substituting (2.57), (2.58) and (2.59) into (2.55) and equating coefficients of
equations
"1 - aiYi + (012 - ai)Yi =

V2
In this case, yi(t)

may

(i

20:2) Y2 +

(2

"2l

+ <*2 "" 2)Y2 +

2yi

be taken to be zero, and any non-trivial


"2

("I

- 2a2)Y2 +

(<4

y2(*)

a2Yi

ajj

C^-^^)

and X2 gives the

satisfying

~ 2i + 2 - 2)Y2 ~

(2.60)

will give the desired canonical form.

This problem illustrates the utility of the given time-varying form (2.38). It may always be
found by differentiating known functions of time.
Other forms usually involve the solution of
equations such as (2.60), which may be quite difficult, or require differentiation of the ((*). Addition
of an input contributes even more difficulty.
However, in a later chapter forms analogous to the
first

canonical form will be given.

Supplementary Problems
2.9.

Given the discrete time equation


state equations in
ical

2.10.

2.11.

(i)

y(n-\-2)-\-Zy(n-\-l)-\-2y(n)

the first canonical form,

= u(n-]r\) + Zu(n),

the second canonical form,

(ii)

(iii)

find the matrix


the Jordan canon-

form.

Find a matrix state equation for the multiple input multiple output system
Vl

l^l

22/l

Y3:5'2

Y4J/2

2/2

Yl^2

liVi

03^1

o'iVi

Write the matrix state equations for the system of Fig.

"3

X3
-t-

(02

=
-

Pi^l

^iM2

^2"!

S2M2

2-20, using Fig. 2-20 directly.

^2

-* V

Fig. 2-20

2.12.

Consider the plant dynamics and feedback comAssuming


pensation illustrated in Fig. 2-21.
the initial conditions are specified in the form
^(0), v((i), V(0), T(0), w(0) and w(0), write a
state space equation for the plant plus compensation in the form x = Ax + Btt and show the
relation between the specified initial conditions
and the components of x(0).

500
p(p

+ p + 100)

6)(p2

p2

+ p + 100
+ 2p + 100

p2

rn^ig.

2-21

METHODS FOR OBTAINING THE STATE EQUATIONS

34

[CHAP.

2.13.

The procedure of turning "backwards" the flow diagram of the first canonical form to obtain the
second canonical form was never justified. Do this by verifying that equations (2.8) satisfy the
original input-output relationship (2.S). Why can't the time-varying flow diagram corresponding
to equations {2.24.) be turned backwards to get another canonical form?

2.14.

Given the linear time-varying differential equation


y

with the

ai{t)y

initial conditions 3/(0)

^(0)

j/q.

a2(t)y

I3{t)

tii{t)u

p^(t)u

Vo-

(i)

Draw

(ii)

Indicate on the flow diagram the values of the scalors at all times,

(iii)

the flow diagram and indicate the state variables,

Write down the

2.15.

Verify equation

2.16.

The

initial conditions of

{2.37)

using equation

Ri

Inertial Load:

K^i

{2.25):

motor are

simplified equations of a d-c

Motor Armature:

the state variables.

di

dB
Kf-jj

L-rr

+ ^

J ~Tfi

j7

Obtain a matrix state equation relating the input voltage

to the

output shaft angle

using a

state vector

\deldt/

2.17.

The equations describing the time behavior

of the neutrons in a nuclear reactor are


6

Prompt Neutrons:

In

{p{t)

Delayed Neutrons:

/Jjii

fi)n +
XjCj

2
=1

\Ci

where

/?

i=l

and

y^i

p{t) is

the time-varying reactivity, perhaps induced by control rod motion.

Write the matrix state equations.

2.18.

Assume

the simplified equations of motion for a missile are given by

Lateral Translation:
Rotation:

'4,

K^a

Angle of Attack:

+ K^^ +

K^ji

lateral translation,

bending moment at station

I.

K^z
M{1)

and

(i)

Ka{l)a

Kb{1)IS

,,,

attitude,

= Ax +

dt

u = p

a = angle of attack, p = engine deflection,


Obtain a state space equation in the form

^
where

K^P =

Rigid Body Bending Moment:

where

K^ +

Bw,

= Cx + Dm

taking as the state vector x

(ii)

(iii)

M{1)

CHAP.

2.19.

METHODS FOR OBTAINING THE STATE EQUATIONS

2]

Consider the time-varying electrical network of Fig. 2-22. The


voltages across the inductors and the current in the capacitors
can be expressed by the relations

35

9-^000^

L,(t)

dt

= Ci(t)

dt

dCi

de,

(Ciei)

^2(<)

dLi

dii

^M0"^-o

Ci-nr

e..

C
e,

It is

dL,

dii

eu

more convenient

La-^ +

-at(^2k)

Fig. 2-22

i^

to take as states the inductor fluxes

Pi

Pi(to)

dt

p2(to)

(^-12) dt

4-

9i(Jo)

(a-ei)dt

I
to

P2

(El

Bb)

and the capacitor charge


91

Obtain a state space equation for the network in the form


dx.

A(t)K

B(t)u

C(t)x

D(t)u

dt
y(t)

where the

state vector x, input vector u,

and output vector y are

'Pi
(i)

y =1

9i

ei

,P2/

^2,

'ii

X =

(ii)

ei

eb

=
I

61

.J2.

2.20.

= Jq'^'Vq 4- ^p^'Tp where q is a vector of n generalized coGiven the quadratic Hamiltonian


p is a vector of corresponding conjugate momenta, V and T are nX n matrices corresponding to the kinetic and potential energy, and the superscript ^ on a vector denotes transpose.
Write a set of matrix state equations to describe the state.
ordinates,

Answers
2.9.

(i)

x(w

l)

(ii)

x(n +1)

(iii)

x( 4-1)

to

Supplementary Problems

1\
-2
-2

-3)^^'^'^

0\

"*"

/0\
ll

'

/1\

'*^"-^'

^^'^^

(1

O)x(n)

(3

l)x(n)

(-1 2)x(n)

METHODS FOR OBTAINING THE STATE EQUATIONS

36

2.10.

Similar to

first

[CHAP. 2

canonical form

-1

-73
-74
-71
-72

-2

-"3
-4

+
1

0^

oy

or similar to second canonical form

10
oci 0:4 0:3

0:2

-73 -72 -71/

-74

2.11.
~^**-

-Oil

0)2

+0)2
-0)3

xi
(1

2.12.

For

Xi

V,

X2

^1

^3

^>

^i

'*''

^5

0)1

'>^>

Xs

X2

=w

the initial conditions result immediately and

1
1
1

-500 -105 -6

100

2.13.

-100 -2

The time-varying flow diagram cannot be turned backwards to obtain another canonical form because the order of multiplication by a time-varying coefficient and an integration cannot be interchanged.

2.14.

(i)

?/

Fig. 2-23

CHAP.

METHODS FOR OBTAINING THE STATE EQUATIONS

2]

(li)

The values of

(iii)

XiiO)

I/O

vq, y^

and yj are given by equations

YoiO) u(0)

a;2(0)

^o

(2.3S), (2.54)

and

(S.35).

- VoW + ri(0)) m(0) - roW m(0)


(

/-RL-^

(p(f)-/3)Z-i

Xi

2.17.

2.18.

(i)

A =

(ii)

A
C

(iii)

A =

2.19.

(i)

A =

(ii)

=
BH

2.20.

The equations of motion are


Using the

state vector

x =

(gj

Qn

dPi

ail

dt

dt
Pi

Pn)'',

the matrix state equations are

TN

37

chapter 3
Elementary Matrix Theory
3.1

INTRODUCTION

This and the following chapter present the minimum amount of matrix theory needed
comprehend the material in the rest of the book. It is recommended that even those well
versed in matrix theory glance over the text of these next two chapters, if for
no other
reason than to become familiar with the notation and philosophy. For those not
so well
versed, the presentation is terse and oriented towards later use. For a more comprehensive
to

presentation of matrix theory,

we suggest study

of textbooks solely concerned with the

subject.

3.2

BASIC DEFINITIONS

Definition 3.1:

Example

matrix, denoted by a capital boldfaced letter, such as


or *, or by the
notation {sa} or {ao}, is a rectangular array of elements that are members
of a field or ring, such as real numbers, complex numbers, polynomials,
functions, etc. The kind of element will usually be clear from the context.

3.1.

An example

of a matrix

is

A
Definition 3.2:

A
A

('
T

x^

sin

row

line

Definition 3.3:

of a matrix
can be drawn.

is

column of a matrix
can be drawn.

'

the set of

is

h^^

'^

-^

''22

<*23

all

the set of

(21

{a,)

elements through which one horizontal

all

elements through which one vertical

line

Example

3.2.

The rows

of the matrix of

Example

\x^/

\smt

Definition 3.4:

Definition 3.5:

The order

jT

3.1

are (0 2

;)

and

(tt

x^ sin

t).

The columns of

this

matrix are

square matrix has the same number of rows and columns.


of a matrix

is

mxn,

read

by

n, if it

has

rows and n

columns.

A scalar,

Definition 3.6:

denoted by a letter that is not in boldface type, is a 1 x 1 matrix.


In other words, it is one element. When part of a matrix A, the notation
Oij- means the particular element in the ith row and jth. column.

Definition 3.7:

vector, denoted by a lower case boldfaced letter, such as a, or with its


contents displayed in braces, such as {;), is a matrix with only one row or
only one column. Usually a denotes a column vector, and a^ a row vector.

38

CHAP.

ELEMENTARY MATRIX THEORY

3]

Definition 3.8:

39

The diagonal

of a square matrix is the set of all elements an of the matrix


which i = j. In other words, it is the set of all elements of a square
matrix through which can pass a diagonal line drawn from the upper left
hand corner to the lower right hand corner.
in

Example

3.3.

Given the matrix

{6i;}

hi

-^6,2

021

^022

\. &M

K.

hs

632"^ 633

The diagonal is the set of elements through which the solid line
determined by the dashed lines.
Definition 3.9:

drawn, bn,

is

trace of a square matrix A, denoted tr A,


on the diagonal of A.

The

tr

A =

is

622. ^33.

the

and not those

sum

sets

of all elements

BASIC OPERATIONS
Definition 3.10: Two matrices

3.3

means

an

are equal if their corresponding elements are equal. A = B


bn for all i and j. The matrices must be of the same order.

performed by adding corresponding elements. A + B =


C means a + bn = cn, for all i and j. Matrix subtraction is analogously
defined. The matrices must be of the same order.

Definition 3.11:

Matrix addition

Definition 3.12:

Matrix differentiation or integration means differentiate or integrate each


element, since differentiation and integration are merely limiting operations on sums, which have been defined by Definition 3.11.

is

Addition of matrices is associative and commutative, i.e. A + (B + C) = (A + B) + C


and A + B = B + A. In this and all the foregoing respects, the operations on matrices have
been natural extensions of the corresponding operations on scalars and vectors. However,
dot
recall there is no multiplication of two vectors, and the closest approximations are the

product and the cross product. Matrix multiplication


+ abn, a scalar.
product a b = aibi + a2&2 +
(scalar)

Definition 3.13:

is

an extension of the dot

multiplication, the element Cn of the product matrix C


left matrix
is found by taking the dot product of the ith row of the
=
AB,
so that
C
where
and the jth column of the right matrix B,

To perform matrix

n
Cij

2=

Cf*^fci-

columns
this definition requires the left matrix (A) to have the same number of
compatible.
to
be
said
and are
as the right matrix (B) has rows. In this case the matrices
is 1x1, i.e. a scalar.
It is undefined in other cases, excepting when one of the matrices

Note that

In this case each of the elements


Example 3.4.
The vector equation y

is

multiplied

by the

Ax, when y and x are 2 X

scalar, e.g.

matrices,

On

O12

"21

<*22

i.e.

aB =

for

{abn}

column vectors,

is

all i

and

j.

ELEMENTARY MATRIX THEORY

40

where

- ^2

j/j

for

aik^k

and

But suppose x

2.

so that

A(Bz)

Cz,

AB =

where

Bz, so that

[CHAP.

x^.

b^^Zj.

Then

'

C.

Example

3.4 can be extended to show (AB)C = A(BC), i.e. matrix multiplication is assoBut, in general, matrix multiplication is not commutative, AB BA. Also, there

ciative.

no matrix division.

is

Example

3.5.

AB 7^

To show

BA, consider

AB

2\/l
0/V2

BA

and note

Furthermore, to show there

^^ C.

no

is

BA = BF =

C,

"division" by

B would

A = F,

imply

division, consider

BF
Since

which

certainly not true.

is

Suppose we have the vector equation Ax = Bx, vv^here A and B are nXn matrices. It
can be concluded that A = B only if x is an arbitrary n-vector. For, if x is arbitrary, we
may choose x successively as ei, e2,
e and find that the column vectors ai = bi,
a2 = b2,
a = b. Here e, are unit vectors, defined after Definition 3.17, page 41.
.

. ,

Definition 3.14:

To partition a matrix, draw a vertical and/or horizontal line between two


rows or columns and consider the subset of elements formed as individual
matrices, called submatrices.

As

long as the submatrices are compatible, i.e. have the correct order so that addition
possible, the submatrices can be treated as elements in the basic

and multiplication are


operations.

Example

and a

3.6.

X
X

3 matrix
1

matrix

can be partitioned into a 2

A
A

2 matrix An, a

2 matrix A21, a 2

matrix A12,

A.^^-

similarly partitioned

3X3

matrix

ttu

*12

13

O2I

'*22

023

O3I

"32

A,

adds as

An +

A + B

Bi

A,o
Aoo

+
+

B,

Bo

and multiplies as

AB

A-iiBii
A21B11

+
+

A]2B2

AllBj2

A22B21

AoiBio

+
+

A12B2
AooBo

Facility with partitioned matrix operations will often save time

and give

insight.

CHAP.

ELEMENTARY MATRIX THEORY

3]

41

SPECIAL MATRICES

3.4

Definition 3.15:

The zero matrix, denoted

Definition 3.16:

diagonal matrix

is

a matrix whose elements are

0, is

all zeros.

a square matrix whose off-diagonal elements are

all

zeros.

The unit matrix, denoted

Definition 3.17:

are

a diagonal matrix whose diagonal elements

I, is

all ones.

Sometimes the order of I is indicated by a subscript, e.g. I is an w x w matrix. Unit vectors,


=
denoted es, have a one as the ith element and all other elements zero, so that Im
...

62

(ei
I

Cm).

Note lA

= AI = A, where A

is

any compatible matrix.

triangular matrix has all zeros below the diagonal, and a lower
triangular matrix has all zeros above the diagonal. The diagonal elements

An upper

Definition 3.18:

need not be zero.


or multiplied by an upper triangular matrix results
in an upper triangular matrix, and similarly for lower triangular matrices.

An upper triangular matrix added to

transpose matrix, denoted A'^, is the matrix resulting from an interchange of rows and columns of a given matrix A. If A = {an}, then
AJ = [an], so that the element in the rth row and ^th column of A becomes
the element in the ith row and ith column of A^.

Definition 3.19:

Definition 3.20:

The complex conjugate transpose matrix, denoted At,


elements are complex conjugates of the elements of

Note (AB)^

B''A'r

and (AB)t

is

the matrix whose

A'^.

BtAt.

A matrix A is symmetric if A = A''.


Definition 3.22: A matrix A is Hermitian if A = At.
Definition 3.23: A matrix A is normal if At A = AAt.
Definition 3.24: A matrix A is skew-symmetric if A = -A''.
-At, real
Note that for the different cases: Hermitian A - At, skew Hermitian A =
D, or
diagonal
AtA
=
symmetric A^ = A, real skew-symmetric A = -A'^, unitary

Definition 3.21:

I,

orthogonal

3.5

AtA =

D, the matrix

is

normal.

DETERMINANTS AND INVERSE MATRICES


(square) matrix {} is the sum of the signed
element
products of all possible combinations of n elements, where each
is taken from a different row and column.

The determinant

Definition 3.25:

otannXn

detA

^^^>

(~l)''n'i2i'2' "''

and the sum is taken over all possible


of
1,2,..., n into some other order such as
permutation is a rearrangement
permutations.
transposition is the interchange
transpositions.
2 n
,1, that is obtained by successive
The exponent p of -1
n.
of places of two (and only two) numbers in the list 1,2,...,
order of 1,2,
. to
natural
the
go from
(3 1) is the number of transpositions it takes to
is the
determinant
each
so
numbers,
There are n\ possible permutations of n
2,.
Pi,P2
sum of n products.

Here

Pi 2,

.,P is a permutation of 1,2,

.,,

ELEMENTARY MATRIX THEORY

42

Example

[CHAP. 3

3.7.

To find the determinant of a 3 X 3 matrix, all


forming one transposition at a time, the following
p

This table

3!

terms,

all

A =

Proof

Theorem

3.1:
is

must be found.

Per-

Vi,

P2.

Pa

1,

2,

3,

2,

2,

3,

2,

1,

3,

1,

1,

3,

Let

Ahe

an

Given two

Proof of this theorem


The proof
3.13, page 56.
Definition 3.26:

(-I)^ai3a22a3i

(-I)2ai2a2331

nXn

given in Problem

3.2:

p, so

(-I)'>aiia22a33

Theorem

= 1, possible entries are also 1, 3, 2 and 2, 1, 3. However, these


that the sign of each product in a determinant is unique. Since
Notice at each step only two
possible permutations are given in the table.
Using the table and (3.1) gives

from an odd

numbers are interchanged.


det

1, 2,

not unique in that for p

is

entries can result only

there are

possible permutations of

table can be formed.

8.3,

nXn

(-I)''ai22l033
*12'*2l"33

Then

matrix.

page

(-l)^13a2132

det(A'^)

(-l)^Ctlia2332

det (A).

59.

matrices

and B.

Then det(AB) =

(det A) (det B).

easily given using exterior products, defined in Section

is

most

is

presented in Problem 3.15, page 65.

Elementary row

(or column) operations are:

(i)

Interchange of two rows (or columns),

(ii)

Multiplication of a

(iii)

Adding a

row

(or

row

scalar times a

column) by a scalar,
(or column) to another

row

(column).

To perform an elementary row (column) operation on an m X m matrix A, calculate the


product EA (AE for columns), where E is an k X n{m X m) matrix found by performing the
elementary operation on the n x n{m x m) unit matrix I. The matrix E is called an elementary matrix.
Example

3.8.

Consider the
obtain

2X2

matrix

A=

{ay}.

'0

EA

To add

To interchange two rows, interchange

the two rows in

to

E.

6 times the second

column

AE=i

iWaii
/ V

ai2,\

021

la.2\

o-ii^

a,

22

**:

to the first, multiply


'ii

a-i2\/l
i\

21

<*22/

\6

0\
1/
.

/aii+6ai2
\

\21

aja

_icoa22

Using Theorem 3.2 on the product AE or EA, it can be found that (i) interchange of two
rows or columns changes the sign of a determinant, i.e. det (AE) = -det A, (ii) multiplication
of a row or column by a scalar a multiplies the determinant by , i.e. det (AE) = a det A,
and (iii) adding a scalar times a row to another row does not change the value of the determinant,

i.e.

det (AE)

det A.

CHAP.

ELEMENTARY MATRIX THEORY

3]

43

Taking the value of a in (ii) to be zero, it can be seen that a matrix containing a row
or column of zeros has a zero determinant. Furthermore, if two rows or columns are
identical or multiples of one another, then use of (iii) will give a zero row or column, so that
the determinant is zero.

Each elementary matrix E always has an inverse


I.
Of course an exception is =

E^i, found

column operation of
Example

in

by undoing the row or

(ii).

3.9.

The inverse of E

Definition 3.27:

j.

The inverse

of

E = [^

E-i

is

J'
(

The determinant of the matrix formed by deleting the ith row and the yth
column of the matrix A is the minor of the element a, denoted det My. The
cof actor cy

Example

E-i

is

(-!)'+ det My.

3.10.

The minor of

of a 3

0122

3 matrix

is

detM22

aii33

aisoisi-

The cofactor

C22

(1)* detM22

det M22.

Theorem

3.3:

The Laplace expansion

of the determinant of an

for any column j or det

aijCij

A =

i=l

is

det

2)

fl^Cij

for any

row

A =

i.

3=1

Proof of this theorem


Example

w x w matrix

is

presented as part of the proof of Theorem 3.21, page 57.

3.11.

The Laplace expansion of a


det

A =

Corollary

3.4:

ai2Ci2

3 matrix
022832

~0'12('*21^33

about the second column

0.^2632

~ 'f23'l3l) +

The determinant

is

'l22(*l 1^*33

of a triangular

~ *1303l)

nxn

""

032('*ll'''23

~ 13"2l)

matrix equals the product of the

diagonal elements.

Proof is by induction. The corollary is obviously true for n = 1.


For arbitrary n, the
Laplace expansion about the nth row (or column) of an tc x n upper (or lower) triangular
matrix gives det A = UnnCnn. By assumption, c = aua22- -an-hn-i, proving the corollary.

Explanation of the induction method: First the hypothesis is shown true for n = no,
Then assume the hypothesis
is a fixed number,
{no = 1 in the foregoing proof.)
true for an arbitrary n and show it is true for n + 1. Let n = no, for which it is known

where no
is

true, so that it is true for no

In this

manner

Corollary

3.5:

the hypothesis

let n n^ + l, so that
shown true for all n no.

+ 1. Then
is

The determinant

it is

true for no

+ 2,

etc.

of a diagonal matrix equals the product of the diagonal

elements.

This

2=

is

true because a diagonal matrix

is

also a triangular matrix.

The Laplace expansion for a matrix whose kth row equals the ith row is det A =
for fc =7^ i, and for a matrix whose kth column equals the jth column the Laplace
flfcjCij
n

expansion is det A
rows or columns.

'"'

^ticCij.

But these determinants are zero

since

has two identical

ELEMENTARY MATRIX THEORY

44

The Kronecker

Definition 3.28:

delta 8

1 if

and

8ij

[CHAP. 3

if i

^ j.

Using the Kronecker notation,

*Cij

2=

i=l

The adjugate matrix

Definition 3.29:

of

aiciCji

SkjdetA

(3.2)

is

adj

A=

{cuy, the transpose of the matrix

of cofactors of A.

The adjugate

Then

(3.2)

sometimes called "adjoint", but this term

is

is

saved for Definition

5.2.

can be written in matrix notation as

A adj A = IdetA
Definition 3.30:

An
an

(3.3)

m x n matrix B is a left inverse of the x m matrix A if BA = Im, and


A is said to be nonsingular
m X matrix C is a right inverse if AC =
I.

TO

if it

has both a

left

and a right inverse.

Since
left inverse B and a right inverse C, C = IC = BAC = BI = B.
and AC = In, and if C = B, then A must be square. Furthermore suppose a nonsingular matrix A has two inverses G and H. Then G = GI = GAH = IH = H so that a
nonsingular matrix A must be square and have a unique inverse denoted A"' such that
A-iA = AA-> = I. Finally, use of Theorem 3.2 gives det A det A-^ = detl = 1, so that if
det A = 0, A can have no inverse.

If A
BA = Im

has both a

Theorem

3.6:

Cramer's rule.

Given an

nxn

(square) matrix

such that detA^T^O.

Then
^"^

d^^^J^

The proof follows immediately from equation


Example

(3.3).

3.12.

The inverse

of a 2

2 matrix

is

._!

"22

~12

Another and usually faster means of obtaining the inverse of nonsingular matrix A is
row operations to reduce A in the partitioned matrix A 1 to the unit
matrix. To reduce A to a unit matrix, interchange rows until an - 0. Denote the interchange by El. Divide the first row by an, denoting this row operation by E2. Then
E2EiA has a one in the upper left hand corner. Multiply the first row by an and subtract
it from the ith row for i = 2,3,
.,n, denoting this operation by E3. The first column of
E3E2E1A is then the unit vector ei. Next, interchange rows E3E2E1A until the element in
the second row and column is nonzero. Then divide the second row by this element and
subtract from all other rows until the unit vector 62 is obtained. Continue in this manner
until EmEl A = I. Then Em- -Ei = A^S and operating on I by the same row operations
will produce A~^ Furthermore, detA~' = detEi det E2- -detEm from Theorem 3.2.
to use elementary

Example

To

3.13.

/
find the inverse of

1\
1

adjoin the unit matrix to obtain


1

-1

CHAP.

ELEMENTARY MATRIX THEORY

3]

Interchange rows (detEj

1):

-1

Divide the

first

45

row by 1 (detEg =

1):

1-1 10-1
1

turns out the first column is already


second row to the first (det E3 = 1).
It

ej,

and

all

that

3.6

Example

I is to

add the

-1

the inverse of the original matrix, which has a determinant

all

matrix equations have solutions, and some have more than one.

3.14.

The matrix equation

>
has no solution because no

exists that satisfies the

2i
(b)

necessary to reduce this matrix to

VECTOR SPACES
Not

(a)

is

is

The matrix to the right of the partition line


equal to [(-l)(-l)(l)]-i = 1.

two equations written out as

=
=

The matrix equation


/O

0\
(^)

is

satisfied for

any

J.

To find the necessary and sufficient conditions for existence and uniqueness of solutions
of matrix equations, it is necessary to extend some geometrical ideas. These ideas are
apparent for vectors of 2 and 3 dimensions, and can be extended to include vectors having
an arbitrary number of elements.
Consider the vector (2 3).
Let (i^ i^) " (^
I2 plane shown in Fig. 3-1.

in a plane.
Ij,

Since the elements are real, they can be represented as points


Then this vector can be represented as the point in the

3).

(2

Fig. 3-1.

Representation of

(2

3)

3) in

the Plane

ELEMENTARY MATRIX THEORY

46

If the real vector

were

space
could be represented as a point in (|j i^
Higher dimensions, such as (1 2 3 4), are harder
to draw but can be imagined. In fact some vectors have an infinite number of elements.
This can be included in the discussion, as can the case where the elements of the vector are
other than real numbers.

by drawing the

an

3), it

axis out of the page.

Ig

Let
be the set of all vectors with n components. Let ai and &2 be vectors
having n components, i.e. ai and ai are in "Vn. This is denoted ai S "Vn,
32 G =U. Given arbitrary scalars j8i and p^, it is seen that (/Siai + ^8282) S "Vn,
i.e. an arbitrary linear combination of ai and a2 is in 'Vn.

Definition 3.31:

"Vi is

(1

[CHAP.

and

infinite line

=^2 is

an

infinite

To represent diagrammatically these

plane.

and, in general, 'Vn for any n, one uses the area


enclosed by a closed curve. Let 'U be a set of
vectors in U.
This can be represented as

shown

in Fig. 3-2.

Fig. 3-2.

Definition 3.32:

Example
(a)

under addition

set of vectors li in "Vn is closed

any

a2

S V,

then

(ai

Set of Vectors

if,

given any

in U

ai

GK

and

+ ai) S 11.

3.15.

Given

This subset of "V^ is closed under addiwhose elements are integers.


any two 3-vectors whose elements are integers is also a 3-vector whose elements

lA is the set of all 3-vectors

sum

tion because the

of

are integers.
(6)

Given lA is the set of all 2-vectors whose first element is unity. This set is not closed under addition
because the sum of two vectors in XI must give a vector whose first element is two.

set of vectors 11 in =U is closed under scalar multiplication if, given any


vector a e Ti and any arbitrary scalar p, then pa 'U. The scalar p can be
a real or complex number.

Definition 3.33:

Example

3.16.

Given
vector in

K is the set of all 3-vectors whose


K gives another vector in li, so K

second and third elements are zero.


is

A set of - vectors li in

Definition 3.34:

space

if it is

(1)

IJn that contains at least

closed

Any

scalar times any

closed under scalar multiplication.

under addition and

(2)

one vector

is

called a vector

closed under scalar multi-

plication.
If a

G X/, where
ai, a2,

a,

a vector space, then Oa

Hence the zero vector

multiplication.

Given

If is

is in

then the set of

G X/

because

'U is

closed under scalar

every vector space.

combinations of the

all linear

a vector space

ai is

{linear manifold).

-W

3.7

for

{j/ii}

all ^,

BASES

Definition 3.35:

vector space

aic (k need not


spanned by the vectors ai, a2,
and (2) every vector in T^ is a
G'U,
., aicGlf

in =U is

aiGX/, a2
linear combination of the

equal n)

if

(1)

ai, a2,

a^.

CHAP.

Example

(1

ELEMENTARY MATRIX THEORY

3]

47

3.17.

Given a vector space V in fg to be the set of all 3-vectors whose third element is zero. Then
1 0) and (0 1 0) span
because any vector in
can be represented as (a /? 0), and

{a

Definition 3.36:

0)

Vectors
jSi,

Example

^2,

(1

0),

(y3-a)(l 2

ai,a2,

..,13k

not

0)+

{2cc-/3){l

0)

0(0

0)

.,akG'V are linearly dependent


all

zero such that

+ ^8232 +

ySiai

there exist scalars

if

+ ySfcak =

0.

3.18.

The three vectors of Example 3.17 are linearly dependent because


+1(1 2

Note that any

If

3.7:

1(1

0)

10)

1(0

(0

0)

set of vectors that contains the zero vector is a linearly

dependent

set.

A set of vectors are linearly independent if they are not linearly dependent.

Definition 3.37:

Theorem

0)

and only

if

the column vectors ai,a2,

linearly dependent, then det

of a square

.,a

matrix

A = 0.

are

Proof: If the column vectors of


are linearly dependent, from Definition 3.86 for some
= P^bl^ + p^a^ +
Pv Pv-'Pn not all zero we get
+ /3^a. Denote one nonzero B as B..

'

Then
1

...

AE =

p^

...

...

p^

...

...

p.

...

...

...

0\

(ai|a2| ... |aii ... |a)

\0

(ai

32

a)

Since a matrix with a zero column has a zero determinant, use of


the product rule of
3.2 gives detAdetE = 0.
Because detE = ;8j^0, then detA = 0.

Theorem

Next, suppose the column vectors of A are linearly independent.


Then so are the
column vectors of AE, for any elementary column operation E. Proceeding
stepwise as
on p. 44, but now adjoining I to the bottom of A, we find Ei,
E such that AEi
E = I.
Each step can be carried out since the column under consideration is not
a linear combination
of the precedmg columns. Hence (det A) (det E,)
(detE^) = 1, so that det A ^ 0.
.

Using this theorem it is possible to determine if ai,


., a.,
a2,
k^ n, are linearly dependent Calculate all the A; x fc determinants formed by deleting
all combinations oin-k
rows. If and only if all determinants are zero, the set of
vectors is linearly dependent.
.

Example

3.19.

Consider

gives

det|^g

^j

and

-12.

j.

Deleting the bottom

row

gives

det

(I

Hence the vectors are linearly independent.

Jj

There

Deleting the top row

0.

is

no need to check the

determinant formed by deleting the middle row.

Definition 3.38:

A
(2)

set of n-vectors

ai, aa

a^

form a

they are linearly independent.

basis for

X/ if

(1)

they span

li

and

ELEMENTARY MATRIX THEORY

48
Example

(1

of the three vectors given in Examples 3.17 and 3.18 form a basis of the given vector space,
as shown and (2) any two are linearly independent. To verify this for (1 2 0) and

(1)

they span

0),

set

j8i(l

This gives the equations

Pi

+ ^2 =

which violates the conditions of the

Example

3.2U.

Any two
since

[CHAP.

0,

0)

2Pi

Mi

+ p^

10)

(0

The only

Q.

0)

solution is that p^

and p^ are both

zero,

definition of linear dependence.

3.21.

Any three noncoplanar vectors in three-dimensional Euclidean space form a basis of 1)^ (not necessarily
the orthogonal vectors).
However, note that this definition has been abstracted to include vector spaces
that can be subspaces of Euclidean space. Since conditions on the solution of algebraic equations are the
goal of this section, it is best to avoid strictly geometric concepts and remain with the more abstract ideas
represented by the definitions.
Consider "U to be any infinite plane in three-dimensional Euclidean space
vectors in this plane form a basis for 11.

Theorem

3.8:

1)^.

Any two

are a basis of V, then every vector in


ai, a2, ...,afc
uniquely as a linear corobination of ai, a2,
.,Sik.
If

The key w^ord here

uniquely.

is

The proof

is

is

noncolinear

expressible

given in Problem

3.6.

To express any vector in V uniquely, a basis is needed. Suppose we are given a set of
vectors that span V. The next theorem is used in constructing a basis from this set.

Theorem

3.9:

Given nonzero vectors ai, a2,


.,am G 'Vn. The set is linearly dependent if
and only if some au, for 1< fc^ m, is a linear combination of ai, 82,
., aic-i.
.

Proof of this is given in Problem 3.7. This theorem states the given vectors need only
be considered in order for the determination of linear dependency. We need not check
and see that each aic is linearly independent from the remaining vectors.
Example

3.22.

Given the vectors (1 1 0), (2 2 0) and


2(1 1 0). We need not check whether (1
two vectors.

To construct a
li,

basis

from a given

if as is

0)

a linear combination of

ai

and

a2,

linearly dependent because (2 2 0)

can be formed as a linear combination of the

set of vectors ai,

combination of

test to see if a2 is a linear

They are

0).

(1

If it

ai.

82,

is,

or only ai

if

. ,

a that span a vector space

delete

it

from the

and in this manner delete all linearly dependent vectors from the
remaining vectors in the set form a basis of 'U.
a4, etc.,

Theorem

Given a vector space

3.10:

bi, b2,

Note

Proof:

ai, bi, b2,

Then

hi.

bi

m~

with a basis

ai, a2,

set in order.

test

test

The

.,am and with another basis

are a linearly dependent set of vectors. Using Theorem


.,bfc-i,
..K-i. Then ai.bi,
bi are a linearly depen.,
.

Then
Next

I.

3.9 delete the vector hk that is linearly dependent on ai, bi,


Next note aa, ai, bi,
.,bfc-i,bfc + i,
bfc+i,
.,bi still span V.
.

set.

has been deleted.

82

first

Delete another b-vector such that the set still spans V. Continuing in this manner
-,
.,a2, ai span V. If l<m, there is an ai+i that is a linear combination of ai,
gives a;,
But the hypothesis states the a-vectors are a basis, so they all must be linearly
a2, ai.
independent, hence I m. Interchanging the b- and a-vectors in the argument gives m^l,

dent

set.

proving the theorem.

CHAP.

ELEMENTARY MATRIX THEORY

3]

Since all bases in a vector space


the following definition.

A vector

Definition 3.39:

space

contain the same

has dimension n

49

number

and only

if

if

we can

of vectors,

a basis of

TA

give

consists of

vectors.

Note that

3.8

this extends the intuitive definition of dimension to a subspace of IJm.

SOLUTION OF SETS OF LINEAR ALGEBRAIC EQUATIONS


Now the means are at hand to examine the solution of matrix equations.

Consider the

matrix equation

Ax

=
are zero, x

/A

(ai|a2|...ja)

2f

is the trivial solution, which can be obtained in all cases. To obtain


a nontrivial solution, some of the i^ must be nonzero, which means the ai must be linearly
dependent, by definition. Consider the set of all solutions of Ax = 0. Is this a vector space?

If all

Ij

(1)

Does the

set contain at least one

x=

Yes, because
(2)

Are

Ax =
Are

Az =

if

and Ay

0,

then the

sum x = z + y

is

a solution of

0.

solutions closed under scalar multiplication?

Yes, because

So the set of

Ax =

0,

then

Ax =

is

The vector space

a solution of A(j8x)

j8x is

0.

a vector space.

of all solutions of

Ax =

is

called the null space of A.

mx

If an
matrix A has n columns with r linearly independent columns,
then the null space of A has dimension n r.

3.11:

The proof

if

solutions of

all

Definition 3.40:

Theorem

always one solution.

is

solutions closed under addition?

Yes, because

(3)

element?

is

given in Problem

3.8.

Definition 3.41:

The dimension of the

Corollary 3.12:

If A is an x % matrix with n linearly independent columns, the null space


is unique.
has dimension zero. Hence the solution x =

Theorem

3.13:

null space of

is called

the nullity of A.

of the vector space spanned by the row vectors of a matrix


equal to the dimension of the vector space spanned by the column vectors.

The dimension
is

See Problem 3.9 for the proof.


Example

3.23.

,^

Given the matrix

/I
(

3\
)

It

has one independent row vector and therefore must have only one

independent column vector.

Definition 3.42:

The vector space

of all y such that

Ax = y

for some

is

called the range

space of A.
It is left to the

reader to verify that the range space

is

a vector space.

ELEMENTARY MATRIX THEORY

50
Example

[CHAP. 3

3.24.

The range space and the

null space

may have

other vectors in

common

in addition to the zero vector.

Consider

Then Ab

0,

so

Definition 3.43:

Theorem

3.14:

The rank

and Ac

in the null space;

is

b, so

is

also in the range space.

mxn matrix A is the dimension

of the

of the range space of A.

A equals the maximum number

The rank of
vectors of A,

i.e.

of linearly independent column


the range space has dimension r.

The proof is given in Problem 3.10. Note the dimension of the range space plus the
dimension of the null space = n for anrnxn matrix. This provides a means of determining
the rank of A. Determinants can be used to check the linear dependency of the row or
column vectors.

Theorem

3,15:

Given an

mx

to

Then rank A = rank A^ = rank A'^A = rank AA'^.

matrix A.

See Problem 3.12 for the proof.

Theorem.

3.16:

mXn matrix A

Given an

rankAB ^ rankB.
is

rank AB ^ rank A,
fc matrix B, then
nonsingular, rank AB = rank A; and if

and an w x

Also, if

is

= rank B.

nonsingular, rank AB

See Problem 3.13 for the proof.

3.9

GENERALIZATION OF A VECTOR
From

we

Definition 3.7, a vector is defined as a matrix with only one


generalize the concept of a vector space and of a vector.

Definition 3.44:

A set %l of objects x, y,
X, y, z,

and

all

and z

(1)

x + yisinXf,

(2)

(3)

(x

(4)

for each

(5)

ax

(6)

a(^x)

(7)

Ix

(8)

a(x

(9)

if

in

and only

if

for

all

Xf:

x,

+ y)+z =

is in

and the objects


complex numbers a and ji

... is called a linear vector space

z,

are called vectors

objects x, y

row or column. Here

x+

x and y

(y

+ z),

in "M there is a unique

zinli such that x + z

y.

V,
(a^)x,

X,

+ y) =
{a + ;8)X =

aX

ay.

aX

;8x.

of Definition 3.7 (w-vectors) and the vector space of Definition 3.34 satisfy
this definition. Sometimes a and j8 are restricted to be real (linear vector spaces over the
field of real numbers) but for generality they are taken to be complex here.

The vectors

CHAP.

ELEMENTARY MATRIX THEORY

3]

Example
The

51

3.25.

set

of time functions that are linear comhlnations of sin

sin 2t, sin St, ... is a linear vector

t,

space.

Example

3.26.

The set of all


X + B(() u for

A(<)

Example

The
as

is

solutions to dx/dt

A(t)

is

a linear vector space, but the set of all solutions to dx/dt =


and is not a linear vector space.

fixed u(t) does not satisfy (1) or (5) of Definition 3.44,

3.27.

complex valued discrete time functions x(nT) for


the set of all complex valued continuous time functions x{t).
set of all

w =

0, 1,

is

a linear vector space,

All the concepts of linear independence, basis, null space, range space,
mediately to a general linear vector space.

Example

extend im-

3.28.

The functions

sin

so the dimension of

3.10

etc.,

t,

sin 2t,
is

sin St,

countably

form a basis for the linear vector space

11 of

Example

3.25,

and

infinite.

DISTANCE IN A VECTOR SPACE

The concept of distance can be extended to a general vector space. To compare the
from one point to another, no notion of an origin need be introduced. Furthermore, the ideas of distance involve two points (vectors), and the "yardstick" measuring
distance may very well change in length as it is moved from point to point. To abstract
distance

this concept of distance,

Definition 3.45:

A
h

we have

metric, denoted p(a,

ev with

b), is

any scalar function

of

two vectors a expand

the properties

(1)

p(a, b)

(2)

p{a, b)

(distance

if

and only

is
if

always

=b

positive),

(zero distance if

and only

if

the

points coincide),
(3)

p(a, b)

from b
(4)

Example
(a)

An

p(a, b)

from a

to b is the

same as distance

a),

p(b, c)

p(a, c)

(triangle inequality).

3.29.

example of a metric for w-vectors a and b


"^^'''^

(6)

(distance

p(b, a)

to

For two

_
-

is

(a-b)t(a-b)

[l

+ (a-b)t(a-b)J

real continuous scalar time functions x(t)

and

y(t) for

1'/^

t^

t^,

one metric

is

11/2

p{x,y)

[j\x{t)-y(t)]^dtj

Further requirements are sometimes needed. By introducing the idea of an origin and
by making a metric have constant length, we have the following definition.
Definition 3.46:

The norm, denoted ||a!|, of a vector a is a metric from the origin to the
vector a G X/, with the additional property that the "yardstick" is not a
function of a. In other words a norm satisfies requirements (l)-(4) of a
metric (Definition 3.45), with b understood to be zero, and has the additional requirement
(5)

i|aa||

|aj ||a|].

ELEMENTARY MATRIX THEORY

52

The other four properties

(1)

llall

(2)

!|a||

(3)

|]a||

=
=

(4)

i|a]|

Example

of a metric read, for a norm,

0,
if

andonly

llc||

if

0,

(trivial),

||a||

in "Ug is

+ c||.

|ia

2)ll2

ll(i

ilaiU

VTiM-la2p-

3.31.

norm

Example

3.30.

norm

Example

[CHAP.

in 'V2 is

IK^i

ffl2)lli

l|a||i

Wi\

k2l-

pj

3.32.

norm

norm

in U is the Euclidean

= V^+a =

||a||2

|il^-

In the above examples the subscripts 1 and 2 distinguish different norms.

Any

positive

Definition 3.47:

An

Example

An

(1)

(a, a)

(2)

(a, a)

(3)

(a,b)*

(4)

(a

(a, a) is real,

(x,y), of

and only

if

tv?o vectors a

if

0,

(b,a),

+ ph,

c)

a*(a, c)

(a,

ah)

and

a(a, b)

|8*(b, c).

(a, 0)

0.

3.33.
is

(a, b)

atb

Oj ^i-

3.34.

of time functions of

{x,y)

Definition 3.48:

Two

Example

3.25 is

x*(t)y(t)dt

vectors a and b are orthogonal

if (a,b)

0.

3.35.

In three-dimensional space, two vectors are perpendicular


of

any

0,

inner product in the vector space

Example

to a.

sometimes known as a scalar or dot product.

is

inner product in f

Example

a metric from

is

and h in li iscomplex scalar function of a and b such that given any complex numbers
a and j8,

The inner product, denoted

inner product

Note

An

monotone increasing concave function of a norm

Example

Theorem

if

they are orthogonal for the inner product

3.33.

3.17:

Schwarz inequality
holds if and only
equality
the
furthermore
and

For any inner product


holds,

(a,b) the

|(a,b)|2

if

(a,a)(b,b)

ab or a or b

is the zero vector.

Proof:
scalar p,

If a or

is

(a

the zero vector, the equality holds, so take b

+ ;8b,

+ ;8b) =

where the equality holds if and only


arranging gives the Schwarz inequality.

(a, a)

if

;8*(b, a)

+ i8b =

0.

^(a, b)

Setting

+
/?

7^ 0.

Then for any

iP^h, b)

-(b, a)/(b, b)

and

re-

CHAP.

ELEMENTARY MAtRIX THEORY

3]

Example

53

3.36.

Using the inner product of Example

J
Theorem

(a

x*(t)y{t)dt

+ b,

+ b) =
=

(f''\!c*{t)\2dt\fC\y{t)\2dt

^(aTbTaTb) ^

For any inner product,

3.18:

Proof:

3.34,

+ b, a + b)|
|(a,a) + (b,a) + (a,b) +

vTaTa)

VTbJbj.

|(a

(b,b)i

|(a, a)|

i(b, a)|

|(a,b)|

i(b,b)|

Use of the Schwarz inequality then gives


(a

+ b,

+ b) ^

and taking the square root of both


This theorem

2v'(a,a)(b,b)

that the square root of the inner product (a, a) satisfies the triangle
the requirements of the Definition 3.47 of an inner product, it
a) satisfies the Definition 3.46 of a norm.

yyi th

^/{a,

Definition 3.49:

The natural norm, denoted

Definition 3.50:

An orthonormal set of vectors


product

vs'here 8y is the

||a||2,

^^""^^

An

(b,b)

sides proves the theorem.

can be seen that

Example

shovi^s

Togeth er

inequality.

(a,a)

Kronecker

of a vector a
ai, a2,

..as

is

is

||a||2

a set for which the inner

ITJj]

{1

= ^A^^

''^

delta.

3.37.

orthonormal set of basis vectors

in =U is the set of unit vectors Cj, Cj,

e,

where

e; is

defined as

ith position

Given any set of k vectors,

them?

To

how can an orthonormal

illustrate the procedure,

choose either one and

make

its

bi

must have length

ai

length unity:

/
*Z

becomes

Because

formed from
and a2. First

set of basis vectors be

suppose there are only two vectors,

unity, y^

||ai||^^

into its components:

)'2'l

where

Now

b^

y^a^

the second vector must be broken up

ELEMENTARY MATRIX THEORY

54

[CHAP.

the component of a^ along bj, and a^ y^^ is the component perpendicular to


Finally, the second
72' '^ote that this is the dot, or inner, product y^ = bj+ a^.
orthonormal vector is constructed by letting &^ y^bj have unit length:

Here y^^

To

bj.

is

find

;77-

TT

I|a2-

V2lll2

This process is called the Gram-Schmit orthonormalization procedure. The orthonormal


.,/ 1 according to
is constructed from a, and the preceding bi for i = 1,

bj

b.

=
l|a,

-2(b,a,)bJ|,
1=1

Example

By

3.38.

Given the vectors a[ = (1 1


the formula, the numerator is

Making

this

Theorem

0),

have length unity gives

3.19:

bj

aj

(3

-2

(1

Then

1)/V2.

||ai||2

VI,

so

bj"

0)/\/2.

2)/v22.

dimensional linear vector space in which an inner product exists


has' an orthonormal basis.

Any

finite

Use the Gram-Schmit orthonormalization procedure on the set of


Proof:
in V, or any set that spans V, to construct the orthonormal basis.

3.11

(1

all

vectors

RECIPROCAL BASIS

Given a set of basis vectors bi.ba,


xG'Vn can be expressed uniquely as

Definition 3.51:

.,b for the vector space 'Vn,

an arbitrary vector

b for <=Un, a reciprocal basis


Given a set of basis vectors bi, b2,
inner product
the
that
such
...,r
ri,r2,
vectors
of
.

(ri,bi)

i,j

1,2, ...,n

the
Because of this relationship, defining R as the matrix made up of
have
vectors and B as the matrix with bs as column vectors, we

RtB =
which

is

is

a set

[3.5)
ri

as column

obtained by partitioned multiplication and use of

{3.5).

nonsingular so that R is
Since B has n linearly independent column vectors, B is
exists, and the set is a
uniquely determined as R = (B"')^. This demonstrates the set of r*
independent.
linearly
are
basis for =U because R-^ = Bt exists so that all n of the r*
be expressed.
Having a reciprocal basis, the coefficients p in {3.j) can conveniently
Taking the inner product with an arbitrary n on both sides of {S.Jf) gives
(r.,x)

Use

/8i(r,,bi)

of the property {3.5) then gives P^

+ P^{x^,\) +

(ri,x).

Pjj^i.K)

CHAP.

ELEMENTARY MATRIX THEORY

3]

55

Note that an orthonormal basis is its own reciprocal basis. This is what makes "breaking a vector into its components" easy for an orthonormal basis, and indicates how to go
about it for a non-orthonormal basis in "Vn.

MATRIX REPRESENTATION OF A LINEAR OPERATOR


Definition 3.52: A linear operator L is a function whose domain is the

3.12

space Vi and whose range is contained in a vector space


a and b in Vi and scalars a and /?,

L{aa
Example
it

+ ^b) =

j8L(b)

3.39.

An TO X w matrix A is a linear operator whose domain


transforms w-vectors into m-vectors.

Example

aL(a)

whole of a vector
X^2, such that for

the whole of "V^ and whose range

is

in X'^,

is

i.e.

3.40.

Consider the set lA of time functions that are linear combinations of sin nt, for n

1,2,

Then

00

any

can be expressed as
a linear operator, because

x(t) in

time

is

x{t)

"^^

+ Px2(t)]

[ax^it)

The operation

In sin nt.

dt

C Xiit) dt +

of integrating with respect to

Xzit) dt

Example

3.41.

The operation of rotating a vector in 'V2 by an angle is a linear operator of 'V2 onto L'2- Rotation of a
vector aa + /3b is the same as rotation of both a and b first and then adding a times the rotated a plus /B
times the rotated b.

Theorem

Given a vector space Vi with a basis bi, b2,


and a vector space
.,bn,
V2 with a basis ci, C2,
Then the linear operator L whose
Cm,
domain is lii and whose range is in V2 can be represented by a matrix

3.20:

{y.j}

whose

basis

ci, 02,

..,

column consists of the components of

ith
.

. ,

Cm,

L(bi) relative to the

Proof will be given for n dimensional Vi and

Consider any x in Vi.

dimensional Vz.

Then x =
m
L{^i)

P^^'-

Furthermore since

L(bi)

is

a vector in V2, determine

y^j

such that

1=1

y,iCr

But
/ n

L(x)

l(x^a)
\i=l
/

^^^(b,)

i=l

2:A2:r,.c,
i

=l

j=l
n

Hence the ith component of L(x) relative to the basis


{yj^} times a vector whose components are the j3i.
Example

{0^}

equals 2)

yj^p^,

i.e.

the matrix

'"^

3.42.

of Example 3.41 can be found as follows.


The matrix representation of the rotation by an angle
Rotating
Consider the basis Cj = (1 0) and 63 = (0 1) of "Uz- Then any x = (^i jSg) = ySiCi + /JaCagives L{ei) = (cos <p)ei (sin 0)62, and similarly Lie^) = (cos 0)62 + (sin 0)61,
Ci clockwise by an angle
can be
Therefore rotation by an angle
so that Yii = cos 0, 721 = sin 0, yi2 = sin 0, 722 = cos 0.
represented by
cos
-sin

sin

0"^

COS0.

ELEMENTARY MATRIX THEORY

56
Example

An

[CHAP.

3.43.

elementary row or column operation on a matrix

is

represented by the elementary matrix

as

given in the remarks following Definition 3.26.

The

range space, rank, etc., of linear transformation are obvious extensions


its matrix representation.

null space,

of the definitions for

EXTERIOR PRODUCTS

3.13

The advantages of exterior products have become widely recognized only recently, but
was introduced in 1844 by H. G. Grassmann. In essence it is a generalization

the concept

We shall only consider its application to determinants here, but it


part of the general framework of tensor analysis, in which there are many applications.
of the cross product.

is

For example,
First realize that a function of an m-vector can itself be a vector.
illustrates a vector function z of the vectors x and y. An exterior product 4>^
ap. However, ^p is an element in a generalized
is a vector function of p m-vectors, ai, a2,
vector space, in that it satisfies Definition 3.44, but has no components that can be arranged
z

= Ax + By

as in Definition 3.7 except in special cases.

a ak,
ap is written as ^^ = ai a aj a
The functional dependency of ip" upon ai, 82
"wedge",
read
a,
by
a
separated
where it is understood there are p of the a's. Each a is
which is Grassmann notation.

^p

ai
ai

A^V =

aj

A aj A

says ^^

(3.6)

A afc

is

9^

A aic A

and

A ai

and

sn a

ai)

+ /3(ai a afc a
ajc

a hj a

ai)

(3.6)

a ai

(5.7)

are linearly independent.

afc

(3.7)

says ^^

is

alternating.

+ /8aj) A afc =
ai^a^

Interchanging the vectors in

(3.8)

according to
afc

(3.10)

a at a

a(ai

are degenerate, in that /\Hl is the space of all complex numbers and
having dimension n. The first nontrivial example is then
%{, the original vector space of m-vectors
bilinearity property
the
become
and
(3.7)
Then equations (S.6)

and also

(aBj)

if aj, aj,

multilinear,

(aaj

By

A Hi

3.44.

The case p

AHt-

+ ;8aj) A aic A

(aai

Example

Given m-vectors as in V, where V has dimension n. An exterior product


= ai A ai A
a a/c f or p = 0,1,2, .. .,n is a vector in an abstract vector
space, denoted A"'", such that for any complex numbers a and /?,

Definition 3.53:

Equation

we

see that a^

(aa^) = a(ai

a;,

aj

(aaj

(3.9)

+ /8aj) =

a^)

-a^ a

is

a^

in (3.9) gives

p{aj

(**)

afc)

(5-9)

aj

and multiplying by -1 gives

a aj) +

a(afc

linear in either a^ or a^

a^

Pia^

i^-iO)

aj)

(bilinear)

but not both because in general

a^).

Note that setting


bination of

aCaj

aj

Letbi.ba, ...,b

a^

a;

Aa^

0.

Furthermore

if

and only

if

a,-

0.

beabasisof

aiAa,

Then 3;=

2^

fc*fc

Qifcbfc)A(2yibO

^"<^

^i

J^ 2

,2

T'''"

afcr,(bfc

*^^*

b.)

is

a linear com-

CHAP.

Since

ELEMENTARY MATRIX THEORY

3]

b^

b^j

and

a bj

b^.

= b; a bj^

k<

for

Because
bfc

A b(

aj

for

an arbitrary vector in A^V, and

a^ is

< I n form

A^V

is

n(n

- iYfc)bfc A bi

(fcYi

(3.11) is

(3J1)

a linear combination of b^ a b(, then the vectors


over all possible k and I satisfying this rela-

Summing

a basis for A^TA.

shows that the dimension of

tion

to

(-1

2 2
1=1 fc=i

aiABj

sum can be rearranged

this

I,

57

l)/2 =
(

Similar to the case A'^, if any a; is a linear combination of the other a's, the exterior
product is zero and otherwise not. Furthermore the exterior products bi a bj a
a bk for

form a basis for A^, so that the dimension of A*^ is

<k^n

l^i<j<

_
~

nl

{n-v)\p\

/n\
(^-^^^

\v)

In particular the only basis vector for A"X/ is biAb2A


aK, so A"X/ is one-dimensional.
Thus the exterior product containing n vectors of an n-dimensional space "W is a scalar.

The determinant of a

Definition 3.54:

vector space

'U vi^ith

linear operator

a basis

bi, b2,

The
for

the

is completely independent of the matrix representation of L.


If L is an
with column vectors ai, aa,
a, and since ei, ea,
e are a basis

definition

matrix

-Vn,

det

Without

Aei a Ae2 a
ei A 62 A

A =

we may

loss of generality,

det

and det I

Now

ei

a 62 a

a e

define ei a 62 a

A =

ai

A 32 A

^.

a?

0.

/v

eo

a e

a
a

/> ar.

A Cn

so that

1,

A a

ai

{3.H)

1.

For an exterior product +" =


^f'"

ci

ai a
a ap in An/ and an exterior product
ACq in A'Ti, define exterior multiplication as

^"a^" = aiA
A

a Acn
A Cn

note that the Grassmann notation has a built-in multiplication process.

Definition 3.55:

4?"

is

nxn

domain and range

w^hose

b is defined as

L( MAL(b2)A---AL(b
bl A 62 A
A bn

detL

ABpAClA

AC,

{3.15)

This definition is consistent with the Definition 3.53 for exterior products,
and so
41" is itself an exterior product.
Also,

{3.12),

if

=w

TO

then

ai

a a-2 a a-i

and must equal some vector

Theorem

Then det A =
Let

e;

n components and
aiAa2A

an n- vector since ^""' has dimension n from


and A"~*Xf must coincide.

(Laplace expansion). Given an nXn matrix A with column vectors


au
ai a 82 a
a a = af (a2 a
a an).

3.21.

Proof:

is

in li since 'U

be the ith unit vector in '=Un and j be the jth unit vector in
Vn-i, i.e. e; has
., has k - 1 components.
Then ai = auei + a^e^
anien so that

Aa

aiieiAa2A

Aa+

a2ie2Aa2A

Aan+

++

ttniCn

Aa2A-

Aa

ELEMENTARY MATRIX THEORY

58

The

sum on

exterior product in the

first

Cl

32

A an

Cl

(ai2ei

[CHAP.

the right can be written

+ ^2262 +

+ fti2en) A

A (ainBl

+ 0,nn^-n\

Using the multilinearity property gives


Cl

82

A Hn

r\
(122

Cl

\an2/

=
detin =

Since

e,

A [a22 (^j

\a..Q^ +

a2(;;^_

J]

a(^^_

j]

detln-i, then

CjA

0\
)A

/o
A ^^

-l

'1/

Performing exactly the same multilinear operations on the


hand side gives
A

ei

["022
]

[a22*l

+ fln2*n-l]

is

A ra2 i

on the right

dnni

-\-

'

[a2n*l

'

+ ftnn*i-l]

lain\

the cofactor of an.

detAfu

cu

\annl

\o.nij

Cu

side as

/a22\

vi^here

HA

dnzi

hand

left

Similarly,
n

Cj

A 82 A

A Bn

Cji

and

ai

a a2 a

a an

ftjiCji

(cn

...

C21

Cni)ai

3=1

a a = (cu C21 ... ci)'^ and Theorem 3.21 is nothing more than a statement of the Laplace expansion of the determinant about the first column. The use of column
interchanges generalizes the proof of the Laplace expansion about any column, and use of
det A = det A^ provides the proof of expansion about any row^.
so that a2 a

Solved Problems
3.1.

Multiply the following matrices, where

ai, as,

bi

and h% are column vectors with n

filGiriGiits

<..

(ii)

(12,(2)

(2)

(3 4)

(g)(-.IM

(V,

(iv)

(ax|a2)(^)

(vi)

Using the rule for multiplication from Definition


& X n matrix times an

(i)

nxm

(1X3 + 2X4) =

(-

(-)

matrix results in & k

3.13,

82) (^^^

^^

and realizing that multiplication of a


we have

in matrix,

aib

/a'^b
(11)

(a.

)(!

(iii)

'

'

(v)

a^bi
....

^"^

/(1X3)

V(2X3)

(1X4)\
(2X4))

/3

"

4\
s)

^'""^

T^

(aib;)

(a2b^)

(vi)

(Mi

\2a2)
|

'

CHAP.

3.2.

ELEMENTARY MATRIX THEORY

3]

59

Find the determinant of A (a) by direct computation, (ft) by using elementary row
and column operations, and (c) by Laplace's expansion, where
2

2
(a)

To

There are 4

P2'

Ps.

Pi

Pi,

P2.

P3,

P4

12

13

14

15

16

17

18

19

20

21

10

22

11

23

Then

+l-2l'2-l'2'80+l'0-8'0-l'0'3'2
+ l'6'30-l'6'l'0 + 06-l-0-0'6l'0 + 0'0'l'2
-0'0-8-0 + 0-l'8-0-0'l'l-2 + 0'l'3'2-0-l-8'0
+ 0-2'8'0-0'2'l'2 + 0'6'l'0-0'6'3'0 + 2'0'30
-2'0-l'0 + 2'2'l'0-2'2'l0 + 2l'l'0-2'l'3'0
4

Using elementary row and column operations, subtract 1, 3 and 4 times the bottom row from
the first, second and third rows respectively.
This reduces A to a triangular matrix, whose
determinant

(c)

Pv

24 terms in this table.

detA =

(6)

form the table

facilitate direct computation,

is

the product of its diagonal elements

4.

Using the Laplace expansion about the bottom row,

0\

/I
detA

3.3.

Prove that det (A'^)

det A,

= 2det(l
\1

where

is

01 =

2*2

an % x n matrix.
n!

If

A=

{a},

then

A^ =

{ajj

so that

det

(A''')

{~l)''fflp,i o-p^i

'
'

"pi*

Since a determinant is all possible combinations of products of elements where only one element
taken from each row and column, the individual terms in the sum are the same. Therefore the
only question is the sign of each product. Consider a typical term from a 3 X 3 matrix: 03ia]2a23>
Permute the elements through ai2ff3iffl23 to tti2"23"3i> s*' that the row
i.e. pj = 3, P2 1' Pa 2.
numbers are in natural 1, 2, 3 order instead of the column numbers. Prom this, it can be concluded
n
in general that it takes exactly the same number of permutations to undo Pi, P2>- ,Pn to 1,2
Therefore p must be the same for each
as it does to permute 1,2, ...,n to obtain Pi, P2> >Pnproduct term in the series, and so the determinants are equal.
is

ELEMENTARY MATRIX THEORY

60

3.4.

Vandermonde matrix

y _

1
^2

^l

nn 1

is

nonsingular

if

and only

an-

/i

O2

il

has the form


'l

Prove

[CHAP.

for

6i - Oj

if

On

i = j.

This will be proven by showing

detV

(92

-9l)(93-92)(3-ei)- (- 'n-l)(9n-n-2)---(9n-9l)

(ffj

" ^i)

For n = 2, det V = fig "" ^u which agrees with the hypothesis. By induction if the hypothesis
can be shown true for n given it is true for n 1, then the hypothesis holds for w - 2. Note each
term of det V will contain one and only one element from the wth column, so that

70+yi9n+ +

detV
If

61,62,

Vn-1;;

= 9; for i = 1,2, .,n-l, then detV =


because two columns are equal.
...,_ 1 are the roots of the polynomial, and
.

But Yn-i

is

Therefore

the cof actor of 9^~' by the Laplace expansion; hence


'1

Tn-i

By assumption

...

#,

82

...

e-i

,n 2
e'1

60
"2

det

that the hypothesis holds for w

1,

Y-i

n-1

(i-Oi)

Combining these relations gives

detV

3.5.

Show

det/''^

r)

det

(9-i)(e-e2)---(9n-9-i)

A det C,

where

nxn

and C are

and

mxm

matrices

respectively.
Either
dependent.

det

A=

or

det

A#

0.

det

If

Hence the column vectors of

A=

detf^
and the hypothesis
If

detA#0,

holds.

0,

then the column vectors of

are linearly

are linearly dependent, so that

^) =
'

then A""! exists and

^0

/A

^
~

B\
c)

^A

\o

0\/I
ij[o

OWI

A-iB

Cj\0

The rightmost matrix is an upper triangular matrix, so its determinant is the product of the
Furthermore, repeated use of the Laplace expansion about the
diagonal elements which is unity.
gives
of
I
diagonal elements
det

Use

of the product rule of

1/

Theorem

=
3.2

det

\o

det

then gives the proof,

det

CHAP.

3.6.

ELEMENTARY MATRIX THEORY

3]

Show

that

61

are a basis of V, then every vector in


uniquely as a linear combination of ai, aa,
ak.
if

ai, a2,

li

ajc

is

expressible

Because x is in V, and V is spanned by the basis vectors


one of uniqueness. If there are two or more linear
-t&k that represent x, they can be written as

Let X be an arbitrary vector in V.

by

ai,a2, ...,afc

definition, the question is

combinations of ai,a2,

fc

i=l

and

2
=

"iSi

Subtracting one expression from the other gives

(^1

- ai)ai +

- "2)32 +

(^2

(/3fc

- akW

Because the basis consists of linearly independent vectors, the only way this can be an equality
jGj = a,-.
Therefore all representations are the same, and the theorem is proved.

is

for

Note that both properties of a basis were used here. If a set of vectors did not span V, all
vectors would not be expressible as a linear combination of the set.
If the set did span
but
were linearly dependent, a representation of other vectors would not be unique.

3,7.

Given the set of nonzero vectors ai, 82,


dependent if and only if some ajc, for
.

.,

Show that the set is linearly


a linear combination of ai, 82,

.,am in ^.

Kk^m,

is

afc-i.

// part:

combination of

If a^ is a linear

aj, ag,

.,a;j_i,

then

fc-i

t=i

where the

fii

are not

all

=
which

zero since

^lai

4-

^2^2

a^.

is

satisfies the definition of linear

Only

if part:

the

all

linear combination

/3j

^lai

/32a2

Ph^k

Oafc

Find that nonzero ^^ such that

are zero.

is

+i

Oa

y3a

all

ySj

i>

for

k.

Then the

k~l

Show

(-1)8^

dependence.

a/c

3.8.

/3fc-iafc_i

If the set is linearly dependent, then

=
where not

Then

nonzero.

that if an
X n matrix
then
columns,
the null space of

A has
A has

/^r' 1=1 Piai

n columns with
dimension n r.

at

most r linearly independent

Because A is w X w, then a; G X)^, x G i;. Renumber the a; so that the


pendent ones. The rest of the column vectors can be written as
ar +

^iiai

/3i2a2

first r

are the inde-

Pir^r
{3.16)

because a^+i,

.,a are linearly dependent and can therefore be expressed as a linear combination
of the linearly independent column vectors.
Construct the n r vectors XjjX^,
.,x_r such that
.

ELEMENTARY MATRIX THEORY

62

^11

[CHAP.

Pnr,

/321

Pn r/2

Hn r, r

iS2r

X2

-1

w
Note that Axi
so these are

by the

nr

equation of

first

Axj

{3.16),

by the second equation of

{S.16), etc.,

solutions.

Now it will be shown that these vectors are a basis for the null space of A. First, they must
be linearly independent because of the different positions of the 1 in the bottom part of the vectors.
To show they are a basis, then, it must be shown that all solutions can be expressed as a linear
combination of the Xj, i.e. it must be shown the Xj span the null space of A. Consider an arbitrary
X

solution

Ax =

of

Then

0,

/^"\

/M

jPn-r.A

hA

Pit

kr

Pn-r.r

"r

-1

~lr+l

4r + 2

-1

lr+2

in

\o/

\o

\-:

Or, in vector notation,

ir+i^i

i=l
If s = 0, then the Xj do span
s is a remainder if the X( do not span the null space of A.
the null space of A. Check that the last n r elements of s are zero by writing the vector equality
as a set of scalar equations.

where

nr

Multiply both sides of the equation by A.

Ax

~ir + i^^i

As

i=l

Ax =

Since

and

Ax;

0,

As =

0.

Writing

this out in

terms of the column vectors of

gives

<'iai

<

But these column vectors are

linearly independent, so

i=l

n rxi are a

3.9.

basis, so the null space of

has dimension

a;

0.

Hence the

n r.

that the dimension of the vector space spanned by the row vectors of a matrix
equal to the dimension of the vector space spanned by the column vectors.

Show
is

Without
of the

loss of generality let the first r

row vectors

aj

column vectors

be linearly independent.
a 11

112

a;

Partition

^Ir

^l,r+l

be linearly independent and

as follows:

am

]
O'rr

"^rH- 1,1

<l-r+ 1,2

Ctml

^m2

'xj
I

Xr

ai_r+l

O-r.

^r+l,n

^mr

am

r+ 1

Xr + j

a,+ i,r+l

'^r,r-l

*r+l,n

^m, r + 1

let s

CHAP.

ELEMENTARY MATRIX THEORY

3]

yi

Yr

yr+1

Yn

^mr

^m,r+l

^n

ml

63

r+1
SO that

Xf

(a,-!

...

Oja

for some nonzero

^jXi

=
=

feta,!

for

1, 2,

6ii2

^iX;

'"'

so that

^r+i)

Since the X; are r-vectors,

a^+ij).

&itr

i=l

...

b'-y,)

r.

2
=

b^y^

(b'-yi

62

(6i

\>=1

i=l

b^'yj

ajj

(au

Let the vector b^

b^.

Therefore

and yj

a,,.)

Since the last n

column vectors

Then

yj

so that

ijyj

i=i

"ijb'^yj

for

2
j=l

a;

^or

"ij^i

= ^ + !,...,.

r + 1,

b^yg

(b^yi

b^y^ +

b'ryr

...

+ 1 of the row vectors


The same argument leads to r s,

a^

b^yj =

...

fejai

6282

are linearly dependent, so that


r = s.

that the rank of the


X n matrix
independent column vectors of A.

'

'

'

&r f la^ +

Now

r.

n.

The

O;

=
j

2
=

ijOj

consider

equals the

maximum number

A'''.

of linearly

let

r
,

.,

Hence

Let there be r linearly independent column vectors, and without loss of generality
a^.

so that

Show

!, 32, ...

3=1

Therefore r

3.10.

b^yj

are linearly dependent,

a;
r

for

1,

Any y

w.

range space of

in the

them be

can be ex-

pressed in terms of an arbitrary x as

= Ax =

-l

^i^i

^ix^

+2

(.2

l=r+l \j=l

-l

i,a,) X,

(x^

i=l y

= r+l

fc

a^ix^

a,

This shows the a; for i=l, ...,r span the range space of A, and since they are linearly independent they are a basis, so the rank of A = r.

3.11.

X n matrix A, give necessary and sufficient conditions for the existence and
uniqueness of the solutions of the matrix equations Ax = 0, Ax = b and AX = B
in terms of the column vectors of A.

For an

For

Ax =

To show the

necessity:
n

exist such that

i
sufficiency:

^Ai

x =

If a^

0.

always exists. A necessary and sufficient condition for


that the column vectors aj, ...,a are linearly independent.
are dependent, by definition of linearly dependent some nonzero 1;

the solution

0,

uniqueness of the solution

is

Then there

exists another solution

(|i

|)

0.

To show

ji

If the a; are independent, only zero

J;

exist such that

Sili

= Ax =

0.

For

Ax =

b,

rewrite as

2
i=l

^i^i-

Then from Problem

3.10 a necessary

and

sufficient

condition for existence of solution is that b lie in the range space of A, i.e. the space spanned by
the column vectors.
To find conditions on the uniqueness of solutions, write one solution as
n
n
n
- !>!
Then b =
In) and another as (jj fz . . . {).
=
(vi V2
Offi so that
^iVi =
('7i
<=i
1=1
i=i
The solution is unique if and only if aj, ...,a are linearly independent.

ELEMENTARY MATRIX THEORY

64
Whether or not b

[CHAP.

necessary and sufficient conditions for existence and uniqueness of


lie in the vector space spanned by the column vectors of A and that
the column vectors are linearly independent.
solution to

Ax =

0,

b are that b

= B can be written as Ax^ = bj for each column vector Xj of


and bj of B, by
Since
the preceding it is required that each bj lie in the range space of A and that all the column vectors
form a basis for the range space of A.

AX

3.12.

Given an

mxn

Show rank A = rank A^ = rank A^^A = rank AA''.

matrix A.

3.14, the rank of A equals the number r of linearly independent column vectors
Hence the dimension of the vector space spanned by the column vectors equals r. By
Theorem 3.13, then the dimension of the vector space spanned by the row vectors equals r. But the
row vectors of A are the column vectors of A^, so AT has rank r.

By Theorem

of A.

Now

consider any vector z in the null

that

Az =

space of A.

rank A^A, note both

in the null space of A,

0,

null space of

rank A^A.

3.13.

A and A^A have n columns. Then consider any vector


Then A^Ay = 0, so that y is also in the null space of A^A.
space of A^A, i.e. A^Az = 0. Then z^A^Az = \\Az\\\ = 0, so

To show rank A
y

Ay =

i.e.

0.

z is also in the null

i.e.

A^A, and has some dimension

Substitution of A^" for

Therefore the null space of

Use of Theorem

k.

rank A^

in this expression gives

is

gives rank
rank AA^".

3.11

equal to the

A=

A;

matrix A and an w x A; matrix B, show that rank AB ^ rank A,


rankAB^rankB. Also show that if B is nonsingular, rank AB^rankA, and
that if A is nonsingular, rankAB = rankB.

Given an

Let

mxn

rankA =

r,

so that

has r linearly indepedendent column vectors

Therefore
a,va,for i = r+l,...,n.
=1
Hence
of B, using partitioned multiplication.
aj

AB =

a^bf

i=l

+22
i

= r+lj = l

Then

where bf are the row vectors

ajb?"

i=l

AB =

aj, ....a^.

aa,.br

.2
1

=1

aj bf +
\

2+

<^mK

column vectors of AB are made up of a linear combination of the r independent


A, and therefore rank AB - r.
vectors
of
column

SO that all the

Furthermore, use of Theorem 3.15 gives rankB = rankB^. Then use of the first part of this
- rank B. Again, Theorem 3.15
problem with BT substituted for A and A^ for B gives rankB^AT
gives rank AB = rank B^Ai", so that rank AB - rank B.

^ rankAB, using A-i for A and AB for


If A is nonsingular, rankB = rank A-i(AB)
But since rank AB ^ rank B, then rankAB -rankB if
the first result of this problem.
Similar reasoning can be used to prove the remaining statement.
exists.

3.14.

m
i

Xn in a generalized vector space possessing a scalar product.


Prove that
Define the Gram matrix G as the matrix whose elements gry = (xi,Xj).
.,Xn are linearly dependent. Note that G is a matrix
if and only if xi,X2,
detG =
whose elements are scalars, and that Xi might be a function of time.

Given n vectors

xi, X2,

Suppose det

G=

0.

Then from Theorem


n

vector of G.

Then

3.7, /3igi

+ /32g2 +

'

+ PnSn =

n
Pidij

i=l

/3i(Xi,Xj).

i=l

Multiplying by a constant

/?*

=
3

2
=

and summing

i8*
i

2
=

M^u^j)

still

gives zero, so that

\i=l

^i'xi,
3

2
=

/3*Xj
/

<,

where g

is

a column

CHAP.

ELEMENTARY MATRIX THEORY

3]

65

Use of property

(2)

2
*=i

of Definition 3.47 then gives

dependence.

suppose the

Xj are linearly

Taking the inner product with any

3.15.

Then there

dependent.

x; gives

"

2, JiSi

which

0,

the definition of linear

is

Now

3'

jSiXi

and the column vectors of

such that

yj

yMu^i)

Jk
3=1

exist

for any 1

yiPy

3=1

are linearly dependent so that det

G=

VjXj

0.

Therefore

0.

Given two linear transformations Li and


in

Show

V.

detBA=
Let

det

L2, both of whose domain and range are


(detLi)(detL2) so that as a particular case detAB =

det(LiL2)

detB.

have a basis

bj, bj,

Using exterior products, from

b.

LiL2(b,)

LiLgCbj)

det {L1L2)

A b2 A

bj
If

L2

singular, the vectors

(S.13),

A LiL2(b)

A b

are linearly dependent, and so are the vectors L^L^ibi) =


Then det(LiL2) = = detLg. If Lg is nonsingular, the vectors Cj are linearly
Li(ci).
independent
and form a basis of V. Then Ci a C2 a
a c
0, so that
is

L2(bi)

C;

det (I-iL

^i<=i)

^1(^2)

A Li(c)

Using exterior products, prove that det

(I

Cj

3.16.

det

(I

+ ahi)

Use of Theorem

(ci

Ci

C2

(62

A I,2(b2) a
bi A bj A

Al,2(b)

A b

(detLi)(detL2)

+ 61a) A

A 62 A

(02

62a)

(I

+ ab^)

+ ab'') = 1 + b''a,

(e

61a

A 62 A

6eiA

6jar(e2

d=

L^Cbi)

6a)

A e +

6261

A aA

63

A e

Ae_i AB

3.21 gives

det

but

A e),

det(I

etc.,

ab^)

...

e)

628^(61

eg

e)

(-l)"-i6ar(eiA--- Ae_i)

so that

a^feiei

3^6262

aT6e

a^b

n***"/!""'"" * ^ projection matrix P = I-abT(aTb)-i


l?p-'n
P*-n ^pT^*'
"
nd the transformation y = Px leaves only the
". "' "*
^^
7
If^pointwise mvariant, i.e."' ^b^y==P'b^x
b^x
= bT[I - ab^(aTb) - i]x = (b^ - b^)x = 0.
'''

such that
hyperplane
"yP^P'*"

ELEMENTARY MATRIX THEORY

66

[CHAP.

Supplementary Problems
3.17.

Prove that an upper triangular matrix added to or multiplied by an upper triangular matrix results
an upper triangular matrix.

in

3.18.

Using the formula given in Definition 3.13 for operations with elements, multiply the 'ollowing
matrices
/
2
i\ /q
j
^/V'T

^0

^'^

sin *y

Next, partition in any compatible manner and verify the validity of partitioned multiplication.

3.19.

Transpose

smtj

x^

\-!r

and then take the complex conjugate.

3.20.

Prove lA

= AI = A

3.21.

Prove

skevs^-symmetric matrices have all their diagonal elements equal to zero.

3.22.

Prove (AB)t

3.23.

3.24.

all

for any compatible matrix A.

BtAt.

Prove that matrix addition and multiplication are associative and distributive, and that matrix
addition is commutative.

Find a nonzero matrix which, when multiplied by the matrix B of Example 3.5, page 40, results
or B = 0.
does not necessarily mean A =
Hence conclude AB =

in the zero matrix.

3.25.

How many

times does one particular element appear in the

for the determinant of an n

sum

xn

matrix?
3.26.

Prove (AB)"*

= B~iA~i

3.27.

Prove (A-i)i"

3.28.

Prove det A-i

3.29.

Verify

=-

if

the indicated inverses exist.

(AT)-i.

(detA)-i.

det

DCl

(I

f^

:)

sj^^^M

{ID2\-i

11

/-I

3/

l)

^^^

\2

'^

2\-i

0\-Y2

iMl

3,

^^^^ ^"^

3.30.

Given

3.31.

If both

3.32.

Given a matrix A(t) whose elements are functions of time.

3.33.

that
Let a nonsingular matrix A be partitioned into An, A12, Aai and A22 such
have inverses. Show that

A-i and B-i

^^-1
and

if

A21

0,

exist, does

then

(A

+ B)-i

exist in general?

dA
Show dA-Vdt = -A-i-j^A-i.

-AirAi2\MiV

V"

and A22

"
'

-,

(A22-A2iAn*Ai2)-VV-A2iAii

**

/An'
(

A^

"All A12A22

^-1

- A21A11

Ajj

ELEMENTARY MATRIX THEORY

CHAP.

3]

3.34.

Are the vectors

3.35.

Given the matrix equations

-1

(2

-3

(1

3),

and

4 0)

(1 1

-2

2)

67

linearly independent?

(an

*i2\

:::)('

::

Using algebraic manipulations on the scalar equations ajjli + ai2?2 0, find the conditions under
which no solutions exist and the conditions under which many solutions exist, and thus verify
Theorem 3.11 and the results of Problem 3.11.
3.36.

Let X be in the null space of

3.37.

Given

3.38.

For

sum

matrix

A =

A =

3.39.

3.40.

in

(a)

x'^'y

0.

an arbitrary vector

can be expressed as the

+ y,

where x is in the range space of


true for any matrix A.

(6) this is

Given nXk matrices


can we conclude A
x,

show that

Show

A'''.

Find a basis for the null space of A.

j,

of two vectors,

Given

and y be in the range space of

transpose of A, and

A and B

and an

and y

is in

m X w matrix X such that XA = XB.

the null space of the

Under what

conditions

= B?

V, where

bi, bg,

Show

b are an orthonormal basis of V-

that

3.41.

Given real vectors X and y such that

3.42.

Show

3.43.

3.44.

that rank (A

+ B) -

lixjla

(x,y)

||y||2.

(x,bi)(bi,y)

Show

(x

+ y)

is

orthogonal to

(x-y).

rank A + rank B.

Define T as the operator that multiplies every vector in "U^ by a constant a and then adds on a
translation vector to. Is T a linear operator?

Given the three vectors ai = (VIl9 -4 3), aj = (Vll9 -1 7) and ag ^ (\/ll9 -10 -5), use
Gram-Schmit procedure on aj, aj and ag in the order given to find a set of orthonormal basis

the

vectors.

3.45.

Show that

a ap satisfies Definition 3.44, i.e. that is an element


the exterior product # = Ej a
AH(, the space of all linear combinations of p-fold exterior products.

in a generalized vector space

3.46.

Show

(aiCj

+ a2e2 + "aCs) a

illustrating that a

ab

is

(PiCi

+ /Sgeg + /Sgeg) =

{cczPs

- a3^2)ei +

(i/?3

- o'3Ptl^2 +

(i^2

- 2i3i)e3,

the cross product in IJs.

3.47.

x and an n X w matrix A such that yi, y2


Given vectors Xi, Xj
Prove that Xj, X2, ..., x are linearly independent.
where y; = Ax;.

3.48.

Prove that the dimension of

3.49.

Show

A'Ti

y are linearly independent,

to!

/^_\|

that the remainder for the Schwartz inequality


(a,a)(b,b)

for TC-vectors.

What

is

|(a,b)|2

2 2

ki^i
J
^ t=l 3 = 1

- AI'

the remainder for the inner product defined as (a,b)

a*(.t) b(t) dt"!

ELEMENTARY MATRIX THEORY

68

Answers
2a

JT

aj

a;2

sin

[CHAP.

Supplementary Problems

to

3.18.

bx^

bir

x2

/-2a

a \

3.30.

(w

1)

A-i

for any a and B

p)

\-2l3

3.25.

sint/

^j

3.24.

as

most easily seen by the Laplace expansion.

is

-3/2

5/2

No

3.34.

No

3.37.

(5

3.38.

(a)

-4

and

0)^

3.43.

No, this

3.44.

bi

affine

(-/IW -4

1)^

are a basis.

3.49.

One method

r r

a(t)
I

of

must be

3)/12,

are arbitrary, and since x and y are independ-

linearly independent.

transformation.

bg

(0 3 4)/5

of proof uses induction.

Hr)

are required.

3.48.

where a and

'U2-

The n column vectors

-5

and y

3.39.

an

1,

(6

ent they span

is

l'

-2

-1

,-1
3.31.

TT

3.19.

b sin

a(r) b(t)

|2

dr dt

and

ag is coplanar with aj

and

ag so that only

and bg

chapter 4
Matrix Analysis
EIGENVALUES AND EIGENVECTORS

4.1

Definition 4.1:

An

eigenvalue of the

wxn

permit a nontrivial (x

(square) matrix

- 0)

is

one of those scalars A that

solution to the equation

Ax = Ax
Note

Example

can be rewritten as (A - Al)x = 0. Nontrivial solution vectors x


0, as otherwise (A - Al)-^ exists to determine x = 0.

this equation

exist only if

(-4.1)

det (A - Al) =

4.1.

The eigenvalue equation

Find the eigenvalues of

4\/a;,

3/1x2

is

X2

Then
3

Xj

characteristic equation is

det

3-

The

0.

characteristic polynomial of

polynomial

Then

is

(3

\(-A

S--kJ\x2J

- X)(3- X) -4 =

A is

Xj

det (A

1,

X2

x2-6X +

0,

5.

- Al). Note

the characteristic

Then there are n eigenvalues


polynomial, although some might be

Definition 4.3:

An

Definition 4.4:

Associated with each eigenvalue Ai of the nxn matrix


solution vector Xi of the eigenvalue equation Axi =
vector is called an eigenvector.

Example

or

an nth order polynomial.

An that are the roots of this


repeated roots.
Al, A2,

X^

a second-order polynomial equation whose roots are the eigenvalues

Definition 4.2:

The

Vo

eigenvalue of the square matrix


repeated root.

said to be distinct

is

if it is

not a

A there is

a nonzero
This solution

AiXi.

4.2.

In the previous example, the eigenvector assocated with the eigenvalue 1


1

Then 2*1 + 4x2 =


where the scalar

and Xi

+ ^x^,

from which Xy
x^ can be any number.
0,

4^0:1

2y\a;2

(1)

X2

= 2a;2. Thus

is

found as follows.

the eigenvector x^

is

Xj

-2
(

Note that eigenvectors have arbitrary length. This is true because for any scalar
the equation Ax = Ax has a solution vector ax since A(a:x) = aAx = aAx = A(ax).
69

x^

'/

a,

MATRIX ANALYSIS

70

Definition 4.5:

An

eigenvector

Sometimes
Example

normalized to unity if its length is unity, i.e.


x such that one of the elements

is

easier to normalize

||x||

is

1.

unity.

4.3.

The eigenvector normalized


Xi

it is

[CHAP. 4

= =l
v/sV

],

to unit length belonging to the eigenvalue 1 in the previous

whereas normalizing

element to unity gives

its first

x,

1/

example

is

V-1/2

INTRODUCTION TO THE SIMILARITY TRANSFORMATION

4.2

Consider the general time-invariant continuous-time state equation w^ith no inputs,


dx(t)/dt

A is a constant coefficient
given as x(0) = xo.

where

Example

nxn

Ax(i)

{J^.2)

matrix of real numbers.

The

initial

condition

is

4.4.

Written out, the state equations U.2) are

and the

initial conditions

dxi(t)/dt

dx2(t)/dt

dxn(t)/dt

aa;i(i)

a^^x^i^it)

a2ii()

a22a;2(*)

a^iXid)

a2i2(i)

+ a^^xJS)
+ a2nXn(t)

(()

are given, such as

a;2(0)

n(0)/

Now

define a

new

variable,

\y

an n-vector
y{t)

y{t),

by the one

to one relationship

M-x(i)

(4,3)

required that M be an n x n nonsingular constant coefficient matrix


so that the solution
X can be determined from the solution for the new variable y{t). Putting
x(t) = My(t)
It is

into the system equation gives

Mdy{t)/dt
Multiplying on the

left

Suppose

= M-'AMy(f)

(^.^)

The transformation T-'AT, where T is a nonsingular matrix, is


called a
similarity transformation on A. It is called similarity because
the problem
is similar to the original one but with
a change of variables from x to y.

M was chosen very cleverly to


/yiit)\
dyit)

AM.y{t)

by M-i gives
dyit)/dt

Definition 4.6:

~dr -

y^{t)

di\

/xt
\2

\yn{t)/

make M-AM a

diagonal matrix A.

...

o\/i/i()\

...

0\y2{t)\

=
:

\0

...

Xn

\yn{t)

Ay(i)

Then

CHAP.

MATRIX ANALYSIS

4]

71

Writing this equation out gives dyi/dt = hyi for i=l,2,


.,n.
The solution can be expressed simply as yi{t) = yi(0)e^'\ Therefore if an
such that M-'AM = A can be found,
solution of dx/dt = Ax becomes easy. Although not always, such an
can usually be found.
In cases where it cannot, a T can always be found where T-^AT is almost diagonal.
Physically it must be the case that not all differential equations can be reduced to this
simple form. Some differential equations have as solutions te^'\ and there is no way to get
this solution from the simple form.
.

The transformation

constructed upon solution of the eigenvalue problem for all


n.
1, 2
Because Axi = XiXj for i = 1, 2,
the equations
., ,
can be "stacked up" using the rules of multiplication of partitioned matrices:
the eigenvectors

xj,

is

A(Xl X2
I

X)

AX2

(AXI

A2X2

(AlXl

AXn)

...

...

AnXti)

/A.

A2

X2

(Xl

X2

(Xl
I

Therefore

When

is

singular,

cannot be found.

If all the eigenvalues of

4.1:

X)A

.
I

(xi

X2

.
|

x)

(4.5)

Under a number

be shown
is nonsingular.
One of these conditions
other conditions will be found later.

Theorem

\o
=

Xn)

is

of different conditions, it can


stated as the next theorem, and

an n x % matrix are

distinct, the eigenvectors are

linearly independent.

Note that

if

the eigenvectors are linearly independent,

is

nonsingular.

Proof: The proof is by contradiction. Let A have distinct


eigenvalues. Let xi, X2,
be the eigenvectors of A, with xi,x2,
.,x. independent and x. + i,
.,x dependent.
.

^i

2 ^x.

for j

k + l,k+2,...,n where not

all

p..

0.

Since x^

is

Then

an eigenvector,

Ax.

A.x^

A.

2 ^x,

for j

1=1

\t

= k + l,...,n

i^l

Subtracting this equation from the previous one


gives

1=1

But the

i=l,2,...,k, were assumed to be linearly independent.


Because not all B
are zero some A, = A,. This contradicts the
assumption that A had distinct eigenvalues'
and so all the eigenvectors of A must be linearly
independent.
4.3

X,,

PROPERTIES OF SIMILARITY TRANSFORMATIONS

To determine when a T can be found such that


T-^AT gives a diagonal matrix the
properties of a similarity transformation
must be examined. Define

MATRIX ANALYSIS

72

Then the eigenvalues

of S are found as the roots of

[CHAP.

det (S

- \l) =

0.

But

= det(T-iAT-Al)

det(S-Al)

= det(T-iAT-AT-iIT)
=

det[T-i(A-Al)T]

Using the product rule for determinants,

(S- XI) = detT-idet(A-Al)detT

det

Since detT-
have proved

Theorem

4,2:

Corollary

4.3:

from Problem

(detT)-i

All similar matrices

is

(S-

= det(A-Xl).

AI)

Therefore

we

have the same eigenvalues.

same traces and determinants.

All similar matrices have the

Proof of this corollary

det

3.12,

given in Problem

4.1.

useful fact to note here also is that all triangular matrices B display eigenvalues on
the diagonal, because the determinant of the triangular matrix (B AI) is the product of
its

diagonal elements.

Theorem

4.4:

A matrix A can be reduced to a diagonal matrix A by a similarity transformation if and only if a set of n linearly independent eigenvectors can
be found.

Proof:
By Theorem 4.2, the diagonal matrix A must have the eigenvalues of A appearing on the diagonal. If AT = TA, by partitioned matrix multiplication it is required that
Ati = Aitj, vi^here ti are the column vectors of T. Therefore it is required that T have the
eigenvectors of A as its column vectors, and T~i exists if and only if its column vectors are
linearly independent.
It has already been shown that when the eigenvalues are distinct, T is nonsingular.
So consider what happens when the eigenvalues are not distinct. Theorem 4.4 says that
the only way we can obtain a diagonal matrix is to find n linearly independent eigenvectors.

Then there are two


Case

1.

that root

is

Example

cases:

For each root that

repeated k times, the space of eigenvectors belonging to


In this case the matrix can still be reduced to a diagonal form.

A;-dimensional.

4.5.

Given the matrix

A =

is

\-l

Then

and

eigenvalue

det (A

XI) = X(l X)2

and the eigenvalues are

0/

Ax =

For the zero eigenvalue, solution of


problem is

0, 1

the

1.

o\

111).

'

j[

X2J

-Xi

Xs

X3

(0 1

(1)

^2

\xJ

This gives the set of equations

AA
=

0/\xJ

Xi

AA

o\

1
1

gives

=
=

1).

For the unity eigenvalue,

CHAP.

MATRIX ANALYSIS

4]

Therefore

eigenvectors belonging to the eigenvalue 1 have the form

all

X2

where

73

and X2 are arbitrary. Hence any two linearly independent vectors in the space spanned by
1) will do. The transformation matrix is then
and (1

x^

(0 1 0)

M
and T-iAT

A,

where

has

0,

and

on the diagonal in that order.

Note that the occurrence of distinct eigenvalues falls into Case 1. Every distinct eigenvalue must have at least one eigenvector associated with it, and since there are n distinct
eigenvalues there are

eigenvectors.

By Theorem

4.1 these are linearly independent.

The conditions of Case 1 do not hold,


a diagonal form by a similarity transformation.
Case

to

Example

2.

Then the matrix cannot be reduced

4.6.

Given the matrix

A =

Then the eigenvalue problem

Since

is

triangular, the eigenvalues are displayed as 1 and

1.

is
(1)

= 0. All eigenvectors belonging to 1 have the


set of equations ojj = 0,
linearly independent eigenvectors are simply not available to form M.

which gives the

Two

form

{xi 0)1".

Because in Case 2 a diagonal matrix cannot be formed by a similarity transformation,


there arises the question of what is the simplest matrix that is almost diagonal that can be
formed by a similarity transformation. This is answered in the next section.

4.4

JORDAN FORM
The form

which an arbitrary nxn matrix can be transformed


the Jordan form, denoted J. Proof of its existence in all

closest to diagonal to

by a similarity transformation is
cases can be found in standard texts. In the interest of brevity we omit the lengthy development needed to show this form can always be obtained, and merely show how to obtain it.
The Jordan form J is an upper triangular matrix and, as per the remarks of the preceding
section, the eigenvalues of the A matrix must be displayed on the diagonal. If the A matrix
has r linearly independent eigenvectors, the Jordan form has n-r ones above the diagonal,
and all other elements are zero. The general form is
Lii(Ai)

L2i(Ai)

H.7)

J
_,

+I

jLjmp(Apj

MATRIX ANALYSIS

74

[CHAP.

Each Lji(Ai) is an upper triangular square matrix, called a Jordan block, on the diagonal
of the Jordan form J. Several Lii(Ai) can be associated with each value of \i, and may differ
in dimension from one another.
general Lji(Ai) looks like

Ai

Ai

M^O =

...

Ai

Example

Ai

Consider the Jordan form

...

are on the diagonal and ones occur in

4.7.

(-4.5)

^0

where

Ai

..

all

Ai^

places just above the diagonal.

'

Because

ones

all

must occur above the

Xj

Xg,

diagonal in a Jordan block, wherever a zero above the diagonal occurs in J there must occur a boundary
between two Jordan blocks. Therefore this J contains three Jordan blocks,

I'll(^l)

^2l(^l)

'

Lj2(X2)

^1>

Xj

one and only one linearly independent eigenvector associated with each Jordan
This leads to the calculation procedure for the other column vectors
called generalized eigenvectors associated with each Jordan block Lji(Ai):

There

is

block and vice versa.


ti

of

At2

=
=

Ati

Ati

Note the number of


A(xi

ti
I

ti

12

equals the
.

ti

.
I

number
)

Ait2

+ Xi
+ ti

Aiti

Aiti

ti-i

of ones in the associated


Aiti

(AiXi

+ Xi

Ait2

(^.9)

(xi| ti

+ ti

12|

.
1

...
I

1(|

Lji(Ai).

Aiti

Then

+ ti-i

.)Lji(Aj)

This procedure for calculating the ti works very well as long as Xi is determined to within
a multiplicative constant, because then each ti is determined to within a multiplicative
However, difficulty is encountered whenever there is more than one Jordan
constant.
block associated with a single value of an eigenvalue. Considerable background in linear
algebra is required to find a construction procedure for the ti in this case, which arises so
seldom in practice that the general case will not be pursued here. If this case arises, a
trial and error procedure along the lines of the next example can be used.
Example

4.8.

Find the transformation matrix

that reduces the matrix

/2

1^

\0 -1

1.

to

Jordan form, where

CHAP.

MATRIX ANALYSIS

4]

75

The characteristic equation is (2 X)(3 X)(l \) + (2 X) = 0. A factor 2 X can be removed,


and the remaining equation can be arranged so that the characteristic equation becomes (2 X)^ = 0.
Solving for the eigenvectors belonging to the eigenvalue 2 results in

Therefore any eigenvector can be expressed in a linear combination as

What combination should be tried to start the procedure described by equations (^.9)?
expression gives

(i-i

!)(::)

Then

fa
'"2

^2

{::)'(:)"
+
+
~

Tg
""s

'"3

Trying the general

CD'

= P
~P

This gives the correct x = a(l 1 -1)1". Normalizing x by setting


satisfied if a = /?
gives t = (ti t2 1 - T^V- The transformation matrix is completed by any other linearly independent choice of x, say (0 1 -1)'^, and any choice of ri and tj such that t is linearly independent of the
and t^ = 1. This gives AT = TJ, or
choices of all x, say tj =

These equations are

1\ (

-1

irV-i

QUADRATIC FORMS
A quadratic
Definition 4.7:

-1

4.5

form

is

a real polynomial in the real variables

containing only terms of the form


where a,, is real for all i and j.
Example

li.lj,

ayljl^,

such that

^ =

.,!

2 2 %^i^}>

4.9.

Some

typical quadratic forms are

- Hi - 2Sil2 + il + Hiia - 7{2{i


^3 = lllf + 12ll?2 + a21&ll + 22S|
^4 = til + (1 - *')12 ~ ^'^3
^2

Theorem

4.5:

(x,Qx)
All quadratic forms Si can be expressed as the inner product
=
Q.
i.e.
Q+
matrix,
Hermitian
x

w
an
is
v/here
vice versa,

Proof:

First

to (x,Qx):

and

MATRIX ANALYSIS

76

Q=

Let

{q..}

^{a^.

Next, (x,Qx) to

+ a.^}.

(the

Then

problem

q..

and

(x,Qtx)

i=l

Then

(x,

Qx)

SS

1^6 (Q'y)|;|j

i=l 3=1

Theorem

4.6:

= iX^iQii + Qm^i
i=l 3=1

i(x,Qx) + i(x,Qtx)

So

9*^,1,.
*
"

^ = x'^Qx.

real):

t t=

(x,Qx)

and symmetric, and

real

is

prove the coefficients are

is to

= XXQiA^i
=1 =

(x,Qx)

so

q..,

[CHAP. 4

=^

The eigenvalues

and the

of an

nxn

coefficients are real.

Hermitian matrix

Q=

Qt are

real,

and the

eigenvectors belonging to distinct eigenvalues are orthogonal.

The most important case


set of real

is

The eigenvalue problems for

Proof:

Since

Q is included in Theorem 4.6 because the


included in the set of Hermitian matrices.

of real symmetric

symmetric matrices

and

specific A;

Qxi

AiXi

Qxi

AjXj

Qtxj

AjXj

Xj

are

Hermitian,

is

Taking the complex conjugate transpose gives

xjQ = A*xJ
Multiplying

{li:12)

on the right by

Xj

and

on the

{i.ll)

xjQx.-xjQx. =
If j

value

i,

then a/x+xJ
real.

is

so xjxj

Theorem

is

Then

a
if

norm on s.^ and cannot be


j=i,

>?'.

X^ =

\.

{i.l2)

(A*

left

by xj gives

-A.)xtx,

zero, so that Aj

\. But

A*

meaning each eigen-

for distinct eigenvalues,

A^.

A. ^^0,

and the eigenvectors are orthogonal.


4.7:

Even

if

the eigenvalues are not distinct, a set of


normal matrix N.

can be found for an

n orthonormal eigenvectors

nXn

The proof is left to the solved problems. Note both Hermitian and real symmetric
matrices are normal so that Theorem 4.6 is a special case of this theorem.
Corollary

4.8:

Hermitian

(or real

symmetric) matrix

can always be reduced to a

diagonal matrix by a unitary transformation, vi^here


u-i = U+.

U~'^QU

= A and

Since Q is Hermitian, it is also normal. Then by Theorem 4.7 there are n


Proof:
orthonormal eigenvectors and they are all independent. By Theorem 4.4 this is a necessary
and sufficient condition for diagonalization. To show a transformation matrix is unitary,
construct U with the orthonormal eigenvectors as column vectors. Then

'xp

/X+Xj

xjx^

^1^1

x+x^

Vu.

4^2

UtU

4
(XX

x! /

X2I

|x)

x|x

x+x

CHAP.

MATRIX ANALYSIS

4]

But x[x^ =
vectors of

(x.,Xj)

U are

because they are orthonormal. Then IJtU = I. Since the column


U~' exists, so multiplying on the right by U"' gives

S,j

linearly independent,

Ut = U~S which was

to be proven.

a quadratic form ^ = xtQx is given, rotating coordinates by defining


In other words, ^ can be expressed as
gives Si = y'fU+QUy = y+Ay.

Therefore

= Uy

if

^ =
where the
of

77

Ai

of the

Note

+A!2/|2
is

always positive

the zero vector.

is identically

the eigenvalues

if

Then the square root

x vector because an inner product can be defined as

(x,

y)^

of

= x+Qy.

An nxn

Definition 4.8:

form
Then

Hermitian matrix Q is positive definite if its associated quadratic


always positive except when x is identically the zero vector.
is positive definite if and only if all its eigenvalues are > 0.

Si is

An nxn

Hermitian matrix Q is nonnegative definite if its associated


quadratic form Si is never negative.
(It may be zero at times when x is
not zero.) Then Q is nonnegative if and only if all its eigenvalues are 0.

Definition 4.9:

Example

k2\y2\^

are the real eigenvalues of Q.

Q are positive, unless y, and hence x,

^ is a norm

ki\yi\^

4.10.

^=

Ij

2|i42 + ^2 ~

The geometric

Theorem

(fi

~ ^2)^

can be zero when

solution of constant Si

when

fi

is

and so

^2,

is

nonnegative

positive definite is

an

definite.

ellipse in n-space.

unique positive definite Hermitian matrix R exists such that RR = Q,


where Q is a Hermitian positive definite matrix. R is called the square root

4.9:

of Q.

hi

Proof:
Let U be the unitary matrix that diagonalizes Q. Then Q = UAIJt.
a positive diagonal element of A, defineA^'^ as the diagonal matrix of positive

is

Q =
Now let R = UAi-'^Ut
are positive.

and
Uniqueness

One way

to check if a

is to

see

UA/2utUAi''2Ut

symmetric, real and positive definite because


proved in Problem 4.5.

it is

is

Hermitian matrix

eigenvalues are

if its

UAi/2Ai/2Ut

all

is

Since
A-/^.

its

eigenvalues

positive definite (or nonnegative definite)

positive (or nonnegative).

Another way

to check is to use

Sylvester's criterion.

Definition 4.10:

The mth leading principal minor, denoted detQm, of the n x n Hermitian


matrix Q is the determinant of the matrix Qm formed by deleting the last
n m rows and columns of Q.

Theorem

4.10:

A proof is given in
Example

Q=

{g^}.

Then

< detQi =
<

IS

Problem

is

positive definite if

and only

if all

the leading

are positive.

4.6.

4.11.

Given

If

Hermitian matrix

principal minors of

g;

is

positive definite if

<

det Q^

replaced by -, vie cannot conclude

Rearrangement of the elements of

=
is

and only

det f^J.^
\9l2

nonnegative

if

^^^V

...;

<

det

- detQ

922/
definite.

sometimes leads to simpler algebraic inequalities.

MATRIX ANALYSIS

78
Example

[CHAP.

4.12.

The quadratic form

is

positive definite if

which
det

gn >

positive definite if

is

Q>

and either

qnQzz-lfz >

and

^22

922 or

>

and

911922

^ut

0-

gfg

>

can be written another way:

The conclusion

0.

g^ can be shown greater than

MATRIX NORMS
Definition 4.11: A norm of a
||Ax|[ KJIxjj

is

that

is

positive definite if

zero.

4.6

matrix A, denoted
for

||A||, is

the

minimum

value of

such that

all x.

Geometrically, multiplication by a matrix A changes the length of a vector.


Choose
the vector xo vi^hose length is increased the most. Then ||A|| is the ratio of the length of
Axo to the length of xo.

The matrix norm

understood to be the same kind of norm as the vector norm in its


Vector norms have been covered in Section 3.10. Hence the matrix
to be taken in the sense of corresponding vector norm.
is

defining relationship.

norm

is

Example 4.13.
To find ||U|]2, where Ut = U"i, consider any nonzero vector

Theorem

norm

Properties of any matrix

4.11:

(1)
(2)

jAxJI
|A||

= max j|Au[j,

|A

(4)

1AB||^||A[|||B|[

(5)

A[ == |jA||

Proof:

To

show^

Since
(2),

|[A|[

Hull

of ||Au|j is to be taken

||B[|

for any eigenvalue A of A.


if

and only

A=

if

0.

substitution into Definition 4.11 gives

k^^^,

=
=

consider the vector u


||Ax||

maximum value
= 1.

v^^here the

+ B||=^||A[| +

(3)

All

are:

||A|| |[x||

over those u such that

(6)

x.

||aAu||

x/a,
|q:|

where

||Au||

k||x||

||x]|.

(1).

Then

K[aijlu][

Division by \a\ gives |jAu|| -= k[|u||, so that only unity length vectors need be considered
instead of all x in Definition 4.11. Geometrically, since A is a linear operator its effect on
the length of ax is in direct proportion to its effect on x, so that the selection of xo should
depend only on its direction.

To show

(3),

iJ(A

+ B)x|| =

IJAx

+ Bxl! ^

for any x. The first inequality results


vector norms, as defined previously.

|lAx[[

from the

j[Bx[|

(|jA||

||Bj[) ||x|[

triangle inequality, Definition 3.46(4), for

CHAP.

MATRIX ANALYSIS

4]

To show

Bx and

use Definition 4.11 on the vectors

(4),

||ABx||

||A(Bx)||

79

||A|l||Bx||

x:

|1A|| ||B|| ||x||

for any x.

Ax =

To show (5), consider the eigenvalue problem


of vectors Definition 3.46(5),
|A|||x||

Since

an eigenvector,

is

|x|

and

^^
|

Ax =

Ox for any

Theorem

4.12:

||A||

(6),

implies

A = 0.

Therefore

X.

||A||2

where

p^^^,

||Ax||

llAxIi

|lA|l!ix|i

follows.

(5)

To show

||Xx||

from

The converse

p^^^ is the

^
calculate

IJAll^,

maximum

find the

A+Ax) = (Ax, Ax) =

AtA

rmax

At A.

Hermitian,

is

2=

pfg^

then AtA is nonnegative definite and p? ^ 0.


real and the g. can be chosen orthonormal.
Express any x

||Ax||2 s= o,

p2 is

in =U as

M
I1X||2
11

eigenvalue of

AtAg.
(x,

eigenvalue of A+A, and further-

M.

so that

obvious.

is

Consider the eigenvalue problem

Proof:

Since
Since

||0||

maximum

p.
"mm

Then Ax =

Definition 4.11.

more

To

Then using the norm property

Ax.

Then

iiSr

llAxlE
112

Ifpf

\\^AS,\\1

t=l

Since

(pf)^;^

2=

the theorem.

4.7

^i

2 ihf

2
^u
=

(p?)n,ax

i=l

taking square roots gives the inequality of

= p^^JM^ when x

Note that HAxH^

is

the eigenvector

g.

belonging to (p^ax-

FUNCTIONS OF A MATRIX

Given an analytic scalar function


convergent Maclaurin series,

of a scalar

/(a)

a,

it

can be uniquely expressed in a

k=

where

f^

d^f(a)lda!'

Definition 4.12:

evaluated at

0.

Given an analytic function

is

/(A)

/() of

a scalar

a,

the function of

(-1)'"A2'"/(2to)

annxn matrix

S/feAVA;!.

k=

Example

4.14.

Some functions
cos

of a matrix

A =

(cos 0)1

eAt

are

(-sin 0)A

(eO)i

(eO)Aj

+
+

(-cos 0)A2/2
{e'^)\H^I1

+ \H^lk\ +

MATRIX ANALYSIS

80

Theorem

Note

Proof:

Theorem

T-^AT =

If

4.13:

A''

/(A)

From Theorem

Proof:

/(A)

where

/(X,)x,rt

is

T/(J) T-.

A = TJT-^TJT-^

A = TAT^S

If

4.14:

= AA

then

J,

[CHAP.

TJT-^

= TPT-^.

Then

the diagonal matrix of eigenvalues

/(X,)x,rt

Xi,

then

/(Xjxr+

4.13,

/(A)

T/(A)T-'

(-4.15)

But
/

/(A)

S/.^J/'i^!

24^2//^!

S/.AVA;!
Jc

fc

Therefore
/(Ai)
/(A2)

(^J^)

/(A)

/(M,
Also, let

T=

X2

(xi

vector.

Then from

and T"'

x)

(ri

ra

rn)t,

.
]

where

rj

is

the reciprocal basis

and

{^.13)

(A-li),

7(Ai)
/(A2)

/(A)

(xi

|x)

X2

/(An),
'

/(Ai)rl

f iK)4

X2

(Xl
I

The theorem follows upon partitioned


The square

Ix.)

multiplication of these last two matrices.

root function f(a) a^'^ is not analytic upon substitution of any Ai. ThereR of the positive definite matrix had to be adapted by always taking

fore the square root

the positive square root of


Definition 4.13:

Let

/(a)

Ai

for uniqueness in

in

Theorem

Theorem

4.14

to

4.9.

get the spectral representation

A = TAT-i;

A = 2\x,rt

of

CIHAP.

MATRIX ANALYSIS

4]

Note that this


formation.

To

is

valid only for those

calculate /(A)

From

/(J).

last equality follows

equation

LL(A)

from partitioned matrix

Hence,

it is

Ixl matrix

only necessary to find /(L) and use {i.15).

By

calculation

it

can be found that

L,
/a''

...

(,

A"

...

(^^g)^'""""!

...

im) ~ (n~'m)\m\

*^^

'

^.)a'^-"-

fcA"-!

L-'CA)

^'^^^^

Then

multiplication.

//(Ln)

for an

(^.7),

/Ln(Ai)

'l.mn{Xr)j

where the

that can be diagonalized by a similarity trans-

T/(J)T-S we must find


/Ln(Xi)

81

number

i^-i^)

A"

of combinations of

w elements taken

at a time.

00

Then from

/(L)

2 /fcLVA;!,

the upper right hand terms are


'

fc=0

but

The

series converge since /(A) is analytic, so that

""

by comparing

d'-'/(A)

(i-1)!

dA'-i

{i.l7)

and

{-i.lS),

Therefore
//(A)

/(L)

d//dA

...

[(Z-l)!]-id'-i//dA'-i\

/(A)

...

[(;-2)!]-id'-V/dA'-M

(^.^^^

...

From

{i.l9)

and

(4.i5)

and Theorem

4.13, /(A)

/(A)

can be computed.

Another almost equivalent method that can be used comes from the Cayley-Hamilton
theorem.

Theorem

4.15

The proof

(Cayley-Hamilton):
Given an arbitrary nxn matrix A with a characteristic polynomial ^(A) = det(A-Al).
Then 9!.(A) = 0.
is

given in Problem

5.4.

MATRIX ANAI^YSIS

82

Example

4.15.

A =

Given

Then

det (A

By

[CHAP. 4

- M) =

0(X)

X2

6X

(i.SO)

the Cayley-Hamilton theorem,

,(A)

A2

- 6A +

= C'

51

- e(l

'')

ON

+ 6('

I)

/O

The Cayley-Hamilton theorem gives a means of expressing any power of a matrix in


.,nl.
terms of a linear combination of A for 'm= 0,1,
.

Example

4.16.

From Example
terms of

and

given

4.15, the

matrix

satisfies

6A + 51.

A^

= 6A -

A2
Also A3 can be found by multiplying

A3

{i.21)

6A2

by

- 5A =

and then using


6(6A

(i.21) again:

- 51) - 5A = 31A A~i

A-i

For an

4.16:

nxn

where the scalars

y^

/(J)

if it exists,

because USI) can be

(6I-A)/5

+ Y^A"-^ +

yiA"-i

y_j

A + yj

can be found from

y,J-i

found from ^.15) and

/(J) is

301

matrix A,
/(A)

Here

in

(i-Sl)

51

can be found by this method, including


Similarly any power of
multiplied by A~i to obtain

Theorem

Then A^ can be expressed

by

y,J-2

y^j J

+ yj

and very simply from

{^.19),

(i.li)

if

can be

diagonalized.

method avoids the calculation of T and T-i

This
/(J)

i y^J"-*

for the

/(A)

The quantity
Also,

cos

solving

A"

'E

'^km^"'

SA

by

Cayley-Hamilton

the

!:/,, JA;!

= T-V(A)T = T-ii:y,A"-*T =
=
l

^ y,T-iA"-T
i

^ y.J""'

i1

4.17.

For the
YiA

A=

of

4.13,

Example

expense

in brackets is y_.

from Theorem
/(J)

and

2/fcAVfc!

2/.(2:..A')/fc!

=::

the

y,.

/(A)

Since
Proof:
theorem, then

at

A given

in

+ 721 we

Example

4.15,

cos

A=

yiA +

yjl.

Here

obtain
cos Xj

yjXi

72

cos X2

7l>'2

72

has eigenvalues Xi

1,

Xg

5,

From

CHAP.

MATRIX ANALYSIS

4]

83

Solving for yj and yo gives


cos 1

A =

cos ,

cos

5/3

1-5

2\

cos

1/1

~1\

\-l

\)

cos 5 /l

5-1

COS

5/1

"^l

'

5 cos 1

123/
(

Use of complex variable theory gives a very neat representation of /(A) and leads to
other computational procedures.

Theorem

If /() is analytic in a region containing the eigenvalues Xi of A, then

4.17:

~ m{si-pi)-'ds

/(A)

where the contour integration

Since

Proof:
that

/(L)

/(A)

T/(J)T-

^ f{s){sl- sy ds\T-\

= l\ ^.

iT^ t /(s)(sl -L)-ids.

around the boundary of the region.

is

it

suffices to

show

Since

ls-\

-1

...

s-

A,

then
/(s-A)'-i
1

(^I-L)-i

(s-Ay-2

...

(s-A)'-i

...

s-A

...

(s-A)'-\

,,

(s-y)'

The upper right hand term

a^,

g7

of

all

/(s) (si

L)

'

ds

is

J_ r /(s)dg

""

Because

4)

2,r/j7(s

+ Ay

the eigenvalues are within the contour, use of the Cauchy integral formula then

gives

6}-'m

which

is

Example

identical with equation {U.19).

4.18.

Using Theorem

4.17,

cos

For

as given in

Example
(si

A =

^-^

<P

cos s(sl

A)~ids

4.15,

A) _i

_
-

-2 \-i

/s-3
(

-2

3 /

_
~

/s-

1
s2

6s

5 i

Then
cos

2W5'

(s

l)(s

- 5)

3
s

MATRIX ANALYSIS

84

[CHAP. 4

Performing a matrix partial fraction expansion gives

If

1\

V-1

1/

cos 1 /

/-2

cosl

2\

cos

5/1

Vl

[2

5) V 2

COS.5_
i

ds

PSEUDOINVERSE

4.8

the determinant of an n x n matrix is zero, or even when the matrix is not square,
way to obtain a solution "as close as possible" to the equation Ax = y. In
x w real matrix and first examine the properties of the real
section
we
let A be an
this
respectively.
A'^A
and AA^, which are w X and
square
matrices
symmetric

When

there exists a

Theorem

mxm

The matrix B'^B

4.18:

nonnegative

is

^ = y'"y,

Consider the quadratic form

Proof:

where B can be

m X n.

x'^B'^Bx s^ 0,

Then

definite.

which

B^B

so that

is

never negative. Let y


nonnegative definite.

is

= Bx,

Definition 4.9, the eigenvalues of A'^A are either zero


also true of the eigenvalues of AA^, in which case we let

From this theorem. Theorem 4.6 and


or positive and real.

B=

A''

in

Theorem

Theorem

4.19:

This

is

4.18.

r^m

and r^n; let g. be an


Let A be an m x n matrix of rank r, where
eigenvector of
orthonormal
orthonormal eigenvector of A'^A; let i be an
Then
AA'^, and let pf be the nonzero eigenvalues of A^'A.
are also the nonzero eigenvalues of AA''.

(1)

p?

(2)

Ag;

(3)

Ag;

(4)

A^f.

(5)

A^fj

p^i^

p,g;

for

i^l,2,...,r

for

= r+1,

for

for

= r+l, ...,m

..

.,n

l,2, ...,r

From Problem 3.12, there are exactly r nonzero eigenvalues


for i = r + 1,
Then A^Ag. = pfg. for i = l,2, ...,r and A^'Agj =

of A'^A and

Proof:

m-vector

h.

as

h.

Ag./p^

for

AA'-'h,

1, 2,

. ,

r.

AA^Ag./p,

.,n.

AA^.

Define an

Then

A(p2g.)/p,

p^g^-g/Pj

pfh,

Furthermore,
hfh.

gJA^Ag./p,p,

S..

is
Since for each i there is one normalized eigenvector, h. can be taken equal to . and (2)
and
proven. Furthermore, since there are r of the p2, these must be the eigenvalues of AA'^
and are ortho.,m such that AA^f =
Also, we can find fj for i = r+l!
(1) is proven.
basis
orthonormal
are
an
the
and
for
=U
basis
orthonormal
are
an
g^
Hence
the
fj
normal.
for =U.
.

for i= r+l,...,m. Then ||Ag.|i^ = gfA^^Ag; = 0, so that


A^Ag^ =
= for i^r+l,...,m, then AJi. = and (5) is proven.
AA^f^
Similarly, since

To prove
Ag.

0.

(3),

Finally, to prove

(4),

A^f,

A^(Ag/p,)

(A^AgJ/p,

pfg,/p,

PSi

CHAP,,

Example
r

MATRIX ANALYSIS

4]

85

4.19.
.

Let

/6

"'^

0\

-5

m = r = 2,

Then

w=

4.

/72

^^

73;

(^36

24

\36
The eigenvalue
''3

^'

From
it is

''4

AA^

of

are

p\

and

100

'^^ eigenvectors of Ai"A

*^"

the above, propositions

16

36^

eigenvalues of

gj

0.84

0.08

0.24

0.48)^

g2

0.24

-0.12

0.64

-0.72)1'

gg

(-0.49

0.00

0.73

0.49)T

g4

0.04

fi

0.6

0.8)T

fa

0.8

-0.6)T

A^A

are

p2

p2

100,

25,

-0.98 -0.06 -0.13)^

and (5) of Theorem 4.19 can be verified directly. Computationally


and p2 and eigenvectors f and fg of AA^, and then obtain the gj

(1), (2), (3), (4)

easiest to find the eigenvalues p^

from propositions

(4)

Theorem

Under the

and

= 25. The
AA^ are

p^

and

36^

24

(3).
r

4.20:

Theorem

conditions of

A =

4.19,

'^ p^lgT.
i

2.

mxn

matrix A is a mapping from 'V^ to =U,.


orthonormal bases for U^ and 1)^ respectively.

The

Proof:
f J,

=l

.^m ^o^'"

Use of properties

(2)

Ax

and

(3)

of

Theorem

|,Ag,

where

ligf

(=1

g^, gj,

. ,

g and
in IJ^,

= gTx

{i..22)

4.19 gives

i I^Ag,

i=l

i^

Also,

For arbitrary x

i,Ag,

i=r+l

1=1

i.p^i^

1=1

From

{i.22),

theorem

is

$^

and so

gTx

Ax =

Since this holds for arbitrary x, the

ig/igfx.

'"^

proven.

Note that the representation of Theorem 4.20 holds even when


no spectral representation.
Example 4.20.
Using the
^

matrix and the results of Example

10
,

Definition 4.14:

6/

A is

rectangular, and has

4.19,

10("^)(0.84 0.08 0.24


\^-^/

0.48)

"^'^

5(

(0.24

-0.12 0.64 -0.72)

V~0-6/

The pseudoinverse, denoted A~^

of the

mxn

real

matrix

the

is

real

Example

matrix A~^

p.-igjfj'.

4.21.

Ag:ain, for the

matrix of Example

/0.84\

^"^

0.08 \

4.19,

= 'Ao.2,W''-'^-''-\
\0.48/

0.24 \

-0.12

0.64h-'--^'

\-0.72/

/ 0.0888
-0.0144

0.0384 \
0.0208

0.1168-0.0576
\-0.0864
0.1248/

nxm

86

MATRIX ANALYSIS

Theorem

Given an

mxn

equation

Ax =

forthose

4.21 :

In other words,

matrix

real
y.

and an arbitrary w-vector


Then ||Ax y||2

= A"'y.
|]Az y||2 =

Define

xo

xo such that

[CHAP.

y||2,

jjAxo

then

consider the

y,

||Axo

>

|jz||2

y||2 and

||xoj|2.

no solution to Ax = y exists, xo gives the closest possible solution.


y is not unique, xo gives the solution w^ith the minimum norm.

if

Ax =

If the solution to

Using the notation of Theorems 4.19 and 4.20, an arbitrary m-vector y and an
Proof:
%-vector
x can be written as
arbitrary
m

where

rj.

fjy and

|.

Then use of properties

gfx.

Ax Since the

'

i=l

2(^iPi-'?/

t=l

those vectors z in U

for

r+l,

.,n

is

The

with

ivjp)'

r+l

Ij

f or

minimum norm =
+ 40:3 = 1
=
Ax y, where
6X1

This can be written

norm

Then

. ,

Then using

to.

r;j

ijy from

2=

p-i'/jgi

Pr*Sif^y

= A'V

Xo.

4.22.

Solve the equations

rank of

l,2, ...,r.

i=l

Example

^f

= r + 1,

with a

= r+l

i=l

minimum norm must have

U.;25) gives z

Vf

But

arbitrary.

M\i

(4.^i)

rt

i=l
|j

2
= r+l

the best

where

2 vA
t=r+l

we can do is choose i.-r]./-p. for


minimize
that
l|Ax y|J2 can be expressed as

To minimize ||Ax-y|]2

4.19 gives

l|Ax-yi!^

Theorem

of

i:iliP-v)ii+

are orthonormal,

(3)

"EiASi-^vA
i=l
i=l

and

(2)

is

and
y

6x1

(1

+ X2 + 6x^ = 10.
and A is the

10)t

2 and these are four columns, the solution

is

is

ly

-0.0144

If

it

0.0384 \

/1\

0.0208

0.4728 \

0.1936

-0.4592

I
I

\lOj

^.^^gg _o.0576

\-0.0864

Theorem 422:

Since the
matrix of Example 4.19.
The solution Xq with the minimum

/ 0.0888
Xo

not unique.

0.1248/

1.1616/

any solution to Ax = y can be expressed as


z is any arbitrary n-vector.

exists,

x = A-'y +

(I A~^A)z, where

.,m in equation {i.2i), and


for i = r + l,
For a solution to exist, r;j =
Proof:
= T/j/p, for i = l,2, .. .,r. Then any solution x can be written as
.

li

i=r+l

i=l

where
z

C-

for

S-^-,

= r + l, ...,n. are arbitrary scalars. Denote in "Vn


where ^j = gfz. Note from Definition 4.14 and Theorem
i

A-'A

Pr'g,iJi.SlP,
t= i
=

lc

i i PrP^S^SlB,,

fc=l 1=1

an arbitrary vector
4.20,

g,gf

'"^

{A.26)

CHAP.

MATRIX ANALYSIS

4]

87
n

Furthermore, since the

are orthonormal basis vectors for

gj

(I-A-'A)z

= r+

g.gjz

U,

= r+ 1

1=5/

S Sf-

Hence
{^.27)

g^^^

From eciuation {A.23), q^ = ffy so that substitution of (4.37) into (4.25) and use of Definition
4.14 for the pseudoinverse gives
X = A-'y + (I-A-^A)z

Some further

properties of the pseudoinverse are:

nonsingular, A~'

1.

If

2.

A-'A

3.

AA-^A = A

4.

A-^AA-^ = A-^

5.

(AA-^)''

6.

(A-'A)^

7.

A-^Aw =

8.

A-'x =

for

9.

is

AA-^

A~'.

in general.

= AA-^
= A-'A
for

all

all

A-^(y + z) = A-'y
space of A'^.

in the

range space of

A''.

in the null space of A"^.

+ A-^z

for

all

y in the range space of

and

z in the null

all

10.

Properties 3-6 and also properties 7-9 completely define a unique A"^ and are
sometimes used as definitions of A^^

11.

Given a diagonal matrix


Then A-^ = diag {x~\ \-\

12.

Given a Hermitian matrix

A=

diag

(Ai, A2,

H = Ht.

Let

13.

Given an arbitrary mxn matrix A. Let


where H"' can be computed from 12.

14.

=A
(A^)-' = (A-^^

where some

A)
is

taken to be

H = UAUt

H-' = UA-^Ut where A"' can be found from

15.

. ,

where Q-i

.... A"')

may

Xi

be zero.

0.

U-i.

Then

= H-^At =

(AH-')+

where U+

11.

H = A+A.

Then A"^

(A-')-^

16.

The rank of A, At A, AAt, A-^ A-'A and AA"' equals tr(AA-^) =

17.

A is square, there exists a unique polar decomposition A = UH, where W = At A,


U = AH-^ Ut = U-'. If and only if A is nonsingular, H is positive definite real
symmetric and U is nonsingular. When A is singular, H becomes
nonnegative
definite and U becomes singular.

18.

If A(Q is a general matrix of continuous time


functions, Pi.-\t)
tmuities in the time functions of its elements.

r.

If

A'
>

Where

(P,

Q)

is

lim A^(AA''

+ eQ)-i

:=

lim (A''A

may have

+ eP)-iA''

any symmetric positive definite matrix that commutes

Proofs of these properties are

left to the solved

discon-

with (A^A, AA^).

and supplementary problems.

MATRIX ANALYSIS

88

[CHAP.

Solved Problems
4.1.

Show

that

same determinants and

similar matrices have the

all

traces.

To show this, we show that the determinant of a matrix equals the product of its eigenvalues
and that the trace of a matrix equals the sum of its eigenvalues, and then use Theorem 4.2.
Factoring the characteristic polynomial gives

det(A-M) = (Xi-X)(X2-X)---(Xn-^) = XlX2---^n+

Setting X

det (A

A=

gives det

=
=

XI)

(ai

ai

- Xei) a

32

XiX2-

A a +

(-X)[ei

coefficients

Xl

Xa

(a

a2

a a

Hj

ai

A 62 A

eg

Ae +

+
+

aj

ACg a

ei

aj

A e]

a a]

(-X)''

- Xe)

Ae + Bi Aa2A
+ (-X)"""i[ai Ae2 A
A e
+ (-X)"ei A e2 A
X = aj a a2 a
of X again gives XiX2

Comparing

+ ^2+---+>^)(-'^)""^ +

(>^l

Furthermore,

-Xn.

- Xe2) a

(aj

ei

32

a a, and

also

ei

A 62 A

A a

However,
ai

62

and similarly

A e

ei

33

Reduce the matrix

X2

Xl

4.2.

+ agiea +

(ciiiei

to

022,

+ ani^n) a

a e

Onei a

62

a 6 = a^

Therefore

etc.

62

= an +

022

nr.

tr

Jordan form, where

/S -8 -2\
(a)

A =

(b)

A =

/8
(a)

Calculation of

det

-Xi
4
3

is

-8
-3 -4

solved for

-4

y
vector X;

-3 -2
1/
\3 -4
4

-8
-3 -4

-2
-2

(c)

(d)

gives

X^

1,

X^

2, X3

Xi

The

from the equations

-2 \
/O'
-2 j Xi =

X3

1-Xi/

,3

\0,

_8 -2\ /4
_3 _2
3
-4
1/V2

/-1-X
Calculation of

3.

1-Xy

and X3 and to two in


where the third element has been normalized to one in X2

(b)

2\

11/
1

1-X

det

10

1/ \0

\2

o\

2\ /l

/4

x,.

3/

-1

\
-4=0

-3-x/

gives

Xj

X2

X3

1.

Then

eigen-

CHAP.

MATRIX ANALYSIS

4]

(A

Solution of the eigenvalue problem

0)"" where a
and only one vector, (a
one Jordan block Lu( 1), so

is

89

( l)I)x

gives

Therefore

arbitrary,

it

can be concluded there

is

only

Lii(-l)

Solving,

gives

ti

(y8

2a

where p

a)'''

is

arbitrary.

Finally,

from

1-1\

/O

Vo

-4

-2/

t2

we find t2 = (y 2/8 a fl a)^ where y is arbitrary,


respectively gives a nonsingular (x^ tj 12), so that
|

(c)

Choosing

a,

p and y

to be 1,

-A A
1-4(0

-2/V

2-1

-3/Vo

-1.

Since

and

o\ /_!
-1

t,

is triangular, it exhibits its eigenvalues on the diagonal,


so
1.
Xj = Xg
Xg
solution of Ax
(-l)x is x - {a p 2p)T, so there are two b'liearly independent eigen0)T
vectors ad
and ^(0 1 2)t. Therefore there are two Jordan blocks Li(-l) and LgC-l).
These can form two different Jordan matrices Jj or J,.;

The

It makes no difference whether we choose Jj or Jj because we merely


reorder the eigenvectors
and generalized eigenvector in the T matrix, i.e.,

A(Xi|ti|x2)

How

A(X2

(Xi|ti|x2)Ji

/O
(A

- XI)ti

|0

from which ^
arbitrary.

^-1

so that

From Problem
2

-1

-l\A
)(

-l/\0

23

a\

-a

2/8/

The
A.

2S

- a)^

and

where y and

y, 8, etc.

fee

\/-l

/8

\0 2S-a
T

tl)J2

xi

Any choice of a, p, y and 8 such that the inverse of the


transformation to a Jordan form.
(d)

(i.l9)

P
\2P,

(a
0)^ and tj
Xi
(y &
4.41 we can always take a = 1

(X2

From

tj?

-l\
Ojti
0/

\0

|tl)

do we choose a and p to get the correct Xj to solve for

)(

2/3/

0,

S are also
but in general

-1

0\
)

-1/

matrix exists will give a similarity

matrix is already in Jordan form. Any nonsingular matrix T will transform


and T-i( I)T = I. This can also be seen from the eigenvalue problem

(A-XI)x = (-I-(-l)I)x

Ox

it,

since

so that any 3-vector x is an eigenvector. The space of eigenvectors belonging to -1 is three


dimensional, so there are three Jordan blocks Lii(X)
-1,
Li2(X) = -1 and
Li3(X) = -1
on the diagonal for X = 1.

4,3.

[CHAP.

MATRIX ANALYSIS

90

(i.e. NNt = NtN), not necessarily with distinct


that a general normal matrix
eigenvalues, can be diagonalized by a similarity transformation U such that Ut = U"'.

Show

The proof
fc

is

U=

matrix and
Xfc matrix;

i.e.

by induction. First, it is true for a 1 X 1 matrix, because it is already a diagonal


I.
Now assume it is true forafc-lXfc-1 matrix and prove it is true for a
assume that for

U^-i

UjT-i

Ai

Nu

Let

uI-iNfc_iUfc-i

to Xi, an eigenvalue of N^.


the first column vector equal to the eigenvector Xi belonging
'=0^
the Gram-Schmit process,
using
form k-1 other orthonormal vectors X2,X3,...,Xfc from

Form T with
Then

and make

T=

(x,

NuT

TtT =

Note

W)

X2

(Xl

Then

I.

and

TtNfcT

12

aifc

22

2fc

/c2

a/cfcy

nlxfe\

IllXj

XiX
1^1
BfcXa

HfcXfc

where the ay are some numbers.

But TtNfcT

is

normal, because

TtNfcT(TtNfcT)t

r=

TtNfcTTtNJ^T

= TtN^N^T =

(TtNfcT)t(T+NfcT)

Therefore

afc2

^12

22

lfc

2fc

Equating the first element of the matrix product on the left with the
on the right gives
|XiP + k2P+--- + kfcP =

first

element of the product

kP

Therefore

ai2, aig,

. ,

aifc

must

Xi

TtNfcT

all

be zero so that

...

=
*fc-i

and Afc_i
exists a

is

normal.

Ufc_i

= V^l^

22

2, fc-1

lc-l, 2

fc-l, lc-1

where

Since

normal, by the inductive hypothesis there


D, where D is a diagonal matrix.

Afc_iisA;-lXfe-l and

such that

U^-i Afc_iUfc_i

CHAP.

MATRIX ANALYSIS

4]

Define S^ such that

...

S^

91

Then S^S^

and

U,k-l

o\

...

...

Xi

0\/l

...

...

X,

Si^TtNfeTSfc

uJ-1

Ufc-1

/
Therefore the matrix TSj. diagonalizes
X2' ^3
^k of ^k on the diagonal.
Finally, to

show TS^

4.4.

and by Theorem

= sUSk =

S^Sfc

S^TtTSfc

A = UAUt

B = UDUt,

and

has the other eigenvalues

(TS,)t(TS,)

Prove that two nxn Hermitian matrices A = At and


diagonalized by an orthonormal matrix U (i.e. UtU = I)
If

4.2,

unitary,

is

Nj.,

B = Bt
if

can be simultaneously

and only

AB = BA.

if

then

AB = UAUtUDUt = UADUt = UDAUt = UDUtUAUt = BA


Therefore

all

matrices that can be simultaneously diagonalized by an orthonormal

To show the convers^, start with


Ax; - XjX;, so that ABx; - BAxj = XjBxi.

commute.

AB =

BA. Assume A has distinct eigenvalues.


Then
Hence if Xj is an eigenvector of A, so is Bx-.
For
distmct eigenvalues, the eivenvectors are proportional, so that Bx; = pjX;
where pj is a constant of
proportionality.
But then p^ is also an eigenvalue of B, and x^ is an eigenvector of B. By normalizing the Xj so that x^txi = 1, U = (xj
x) simultaneously diagonalizes A and B.
|

If neither A nor B have distinct eigenvalues, the


proof is slightly more complicated.
Let X
be an eigenvalue of A having multiplicity m.
For nondistinct eigenvalues, all eigenvectors of
X belong in the
dimensional null space of A - XI spanned by orthonormal Xj, Xg,
, x.
Therefore Bxj c^-Xj,
where the constants c^- can be determined by c,j =
Then for
xJ Bx^.

and X = (xi X2
xJ, C = XtBX = XtBtX = Ct, so C is an m X m Hermitian
Then C = U^D^U^ where D and U^ are m X m diagonal and unitary matrices
respectively.
Now A(XU) = X(XUm) since linear combinations of eigenvectors are still eigenvectors,
and D = uXxtBXU^. Therefore the set of m column vectors of XU^ together with all other

C=

{cy}

inatrix.

normalized^eigenvectors of A can diagonalize both A and B. Finally, (XU)t(XU)


(^nJmVm) Ifn, SO that the column vectors XU^ are orthonormal.

4.5.

Show

the positive definite Hermitian square root R, such that R^

= (U+XtXU) =

Q,

is

unique.

Hermitian, UAjUt = R where U is orthonormal. Also, R2 and R commute, so that


both R and Q can be simultaneously reduced to diagonal form by Problem 4.4, and Q = UDUt.
Therefore D = Aj.
Suppose another matrix S^ = Q such that S = VAjYt. By similar reasoning, D = Ag. Since a number >
has a unique positive square root, Ag = Aj and V and U are
matrices of orthonormal eigenvectors.
The normalized eigenvectors corresponding to distinct
eigenvalues are unique. For any nondistinct eigenvalue with orthonormal eigenvectors Xj.Xj,. .,x,
Since

is

(xi

.
I

x)

.
I

xxxt

MATRIX ANALYSIS

92

[CHAP. 4

and for any other linear combination of orthonormal eigenvectors,


(yi

where

t1=T'.

ym)

(xi

ym)

(Xx

(Xl

Xm)Tm

Ym.

Then

(yi

Hence

yi , y2,

and S are equal even though

and

V may

^m/'-wJ-n

X^)

differ slightly

when

has nondistinct eigen-

values.

4,6.

Prove Sylvester's theorem: A Hermitian matrix


if all principal minors det Qm > 0.

is

positive definite if

and only

Let x^ be the vector of the flrst^


for any x.
If Q is positive definite, ^ = (x, Qx) ^
n elements are zero, (x^, QmXm) = (x", Qx") - 0.
For those x" whose last
elements of x.
From Problem 4.1, the
Therefore Q^ is positive definite, and all its eigenvalues are positive.
determinant of any matrix equals the product of its eigenvalues, so det Q^ > 0.

If

detQm >

for

m ==

1, 2,

we

.,m,

proceed by induction.

Assume now that if detQi > 0, ...,detQ_i


Partition Q as
possess an inverse.

We

0,

then

Q_i

is

qt

q+Q;"-!

Qn

>

detQ

are also given

- q+Q;^iq >

0-

^nn

detQ =

positive definite

Then use of Problem


Hence

0.

-q+Qn-iq /

3.5 gives

det

Qn-

Xi

>

0,

and must

gnn-qQn-iq
for any vector (x+_i

substitution into

that

Let

Sfc

if

(x,

af).

Then for any vectors y

Qx) and use of U.28)

||A!|
"

<
A"

1,

vfill

then (I-A)-i

defined

give

(y,

by

Qy)

>

0.

2 A".

n=0

and

n=0
IS

-Sfc

Then

A".

= +l
fc

A-cl

Q = (g- qtQ_iq) detQ_i,

>

(xj-ila;*)

Show

1,

(i.28)

so that qnn

4.7.

Qn

Qn

Qn

>

For n

llAl!"

CHAP.

MATRIX ANALYSIS

4]

by properties
Note

llAfe

(3)

and

||A||fe

+ ill

so that

(I-A)S =

fc^ gives

because ||(I-A)x||

(1

Theorem

of

(4)

Using property

4.11.

A^ + i =

lim

Since S exists,

I.

||A||)||x||

93

(I-A)-i.

is

it

Since

0.

and

(6)

<

|]A||

- A)Sfc = I - A^ +

(I

This

is

gives

i,

taking

lim S^.

liniitl

as

mapping,

called a contraction

|)x|j.

112
4.8.

A =

Find the spectral representation of

12

1^

The

1
r;

and eigenvectors Xj

j,

1 -0.5)^,

(1

The matrix

h^^i^.

i=l

and Xg
The reciprocal basis
Xa

A = 5

spectral representation of

1 - J

xj

{j

has

eigenvalues

and

1 0)^

Xg

X,

1,

{-j 1 0)^.

can be found as

rtj

JJ

-}^-'
1

\-0.5

0,

Then

A ^

4.9.

Show

(l)f

the

that

(A~'A)^ = A~'A
finition

-2)

j(0

(l-i)|

j(-0.5i 0.5 1-;)

(l

+ j)(

{0.5}

0.5

AA-^A =

A = 2

2 p-igk^tlc=l

A"^ =

and

Pitigi

i=l

(2

Pifig.n(2 p-'ejtj)^^

222
T

i=l

and

Sjj

Then

It

AA-'A =

Since gjgj

f ff^

=i

fc=l

PiPr'pfcfigigjfj

pAel)

ffcgfc

>

8,^,

AA-'A =
A^'AA~' = A^^ and from
A-'A

equation

gigf

"

-^A"'-

Pifig-

(^.26),

(2

gigf)^

and

(A-iA)T

Similarly

(AA-O'^

= (?

^i^i)'^

To show uniqueness, assume two


1.

2.

AXA = A
XAX = X

and transposing
9.

+ j)

A-^AA-' = A"', (AA-O'' = AA"^ and


A,
define a unique matrix A"^ that can also be expressed as in Derelations

4.14.

Represent

Similarly

solutions

3.

(AX)T

4.

(XA)r

= AX
= XA

and 5 gives

A^X^A^ = AT

10.

ATY^A^ = A^

5.
6.

satisfy the four relations.

AYA = A
YAY = Y

Then

7.

(AY)^

= AY

8.

(YA)r

^ YA

MATRIX ANALYSIS

94

[CHAP.

number above the

The following chain of equalities can be established by using the equation


equals sign as the justification for that step.

2
4
10
4
2
8
X = XAX = ATX^X = ATY^ATX^'X = A^Y^AX = A'^'Y^X = YAX
= YAYAX = YAYX^AT = YYTATXTA^ = YY^AT = YAY = Y

Therefore the four relations given form a definition for the pseudoinverse that

is

equivalent to

Definition 4.14.

4.10.

of i/ of a certain experiment is thought to depend linearly upon a parameter X, such that y = ax + (3. The experiment is repeated three times, during
which X assumes values xi = 1, xz -1 and xs - 0, and the corresponding outcomes yi = 2, and 2/2 = 2 and ys = 3. If the linear relation is true,

The outcome

-2 =

a(-l)

= 40) +

a(l)

13

13

However, experimental uncertainties are such that the relations are not quite
in each case, so a and p are to be chosen such that

satisfied

j:{y,-ax,-pf
i=l

minimum. Explain why the pseudoinverse can be used to select the


and then calculate the best a and /3 using the pseudoinverse.
is

The equations can be written

Defining

Xo

A-^y,

ATA - (^ ),
fi

Agi/pi,

then

by Theorem

then
fi

in the

pi
(1

form y

4.21,

= ^2

and

-1 0)/V2

= Ax

||y-Axo|l^
P2

and

fa

= d

/?,

as

(2/i

and

Vs,

best a and

gi

D/Vs:

- o*i - ^o)"
=

0)

(1

Now A-'

and

ga

be

can

Since

minimized.

is

(0 1).

Since

calculated

from

Definition 4.14 to be

A-

(1

\8

so that the best ao

and

/3o

?
3

1-

Note

4.11.

Show

C=

this procedure

that
eB, or

if

can be applied to

an n X n matrix C

B=

is

(Vi

~ ^i ~ /^*i ~ "y)^'

^t*^-

nonsingular, then a matrix

exists such that

In

JT-i,

In C.

Then B = In C
Reduce C to Jordan form, so that C = TJT'i.
problem is to find In L(X) where L(X) is an i X i Jordan block, because

/inLii(Xi)

InJ

.
)

lnLfcm(XJ/

=T

so

that the

CHAF.

MATRIX ANALYSIS

4]

Using the Maclaurin

In

[M +

Note all the eigenvalues | of L(X)


by the Cayley-Hamilton theorem,

- XI]

L(X)

- XI

I In

2 MM"* [^C^)-

3^1]*

are zero, so that the characteristic equation


XI]' = 0, so that

is

^'

0.

Then

[L(X)

= IlnX-

InL(X)
Since

XI,

series for the logarithm of L(X)

InL(X)

95

(-iX)-'[L(X)

- XI]

because C-i exists, InL(x) exists and can be calculated, so that In J and hence In C

can be found.

B may

Note
vi^here

given,

is

be complex, because in the 1X1 case. In (-1) = j^. Also, in the converse case
C is always nonsingular for arbitrary B because C-i = e-B.

Supplementary Problems
4.12.

Why

4.13.

Find the eigenvalues and eigenvectors of

does at least one nonzero eigenvector belong to each distinct eigenvalue?

-4'

A =

where

-1^

-2

4.14.

Suppose

4.15.

Prove by induction that the generalized eigenvector

all

the eigenvalues of

are zero.

Can we conclude that

A=

0?

of equation {4.9) lies in the null space of

t(

(A-Xil)' + 1.

4.16.

4.17.

Let X be an eigenvector of both

and B.

Is

an eigenvector of (A-2B)?

also

Let Xj and Xg be eigenvectors of a matrix A corresponding to the eigenvalues of Xj and


Show that axj + /3x2 is not an eigenvector of A if a =
and ;8 # 0.

Xg,

where

Xi v^ X2.

4.18.

Using a similarity transformation

x,(n+l)\
..(n

4.19.

Show

that

all

/O

_
-

+ l)j

to a diagonal matrix, solve the set of difference equations

[2

2\/xi(n)\

-3JU2WJ

U^W)
UtU =

the eigenvalues of the unitary matrix U, where

of one.

4.20.

Find the unitary matrix

and the diagonal matrix


2

Ut/

such that

"

]U

Check your work.

/I
4.21.

Reduce to the matrix

A =

Vo

3/6

-4/5

/2

/ x,(0)\

"'^'^

to

Jordan form.

I,

[1

have an absolute value

MATRIX ANALYSIS

96

4.22.

Given a 3 X 3 real matrix

4.23.

Given the matrix

A =

-1

4.25.

Find the square root of

4.26.

Given the quadratic form

4.27.

Show

4.28.

Show

^0^

What happens

].

A2/(A)

Let the matrix

4.31.

Given a real vector

^=

+ 2|ife + 4|2 + 4I2I3+ 2?^.

^^

have distinct eigenvalues

(xj

Show grad^xTQx = 2Qx

.,x of

4.32.

Given a basis Xi.Xj,

4.33.

Suppose A^

4.34.

Find

4.35.

Find eAt for

4.36.

Find the pseudomverse of

4.37.

Find the pseudoinverse of

4.38.

Prove that the

4.39.

Given an

Can

0.

where

is

A =

Is it positive definite?

A" +

oiiA"

+ aI -

Does

Xj.

a.

[BaldXy daldx^

symmetric, and evaluate

is

and

y,

its

A3(A

- Xil)(A-

Xjl)

0?

Define the vector


da/dx^)'''

grad^x^Ax

reciprocal basis ri,r2

r.

for a nonsymmetric A.

Show

that

2^

Xjrj

the matrix of Problem 4.2(a).

'

-a

-3

/'4
f

/I

A =

are true.
listed properties 1-18 of the pseudoinverse

real matrix

and scalars
Agi

that only

and

be nonsingular?

J,

mXn

if

Xi

and a scalar

^n)"^

grad^a

n =

gives

fjfj

Given a real r>iXn matrix A.


(n

0?

[/(A)]A2?

4.30.

metric

A"/w! converges for any A.

Does

Show

.4

that the coefficient a of I in the Cayley-Hamilton theorem


zero if and only if A is singular.

e^t,

as

e/2

n=

4.29.

4.40.

is

Ly(Xj).

A =

Find the eigenvalues and eigenvectors of

Jordan form.

its

Find the transformation matrix T that reduces

).

4.24.

that

Find

0.

1^

Jordan form, and identify the Jordan blocks

o""

10
to

P2

such that

[CHAP. 4

+ 'm)X{n + m)

Sy

At,

e^i

if

gjgj

= ^r^Si

Sy.

the real symStarting with the eigenvalues and eigenvectors of


^'^]

matrix

such that

ffj

'

derive

the conclusions of

Theorems

4.19

and

4.20.

CHAP.

MATRIX ANALYSIS

4]

the T matrix of J = T-iAT is arbitrary


which each Jordan block has distinct eigenvalues.
fixed and
/Kj

Show that

4.41.

to within

in

K^- is

an

n constants

Specifically,

is

an w X n matrix

T = T^K where Tq

of the

form

is

...

...

matrix corresponding to the jth Jordan block

if

show

."...''!..;;.....".

\0
where

97

'

ai

a^

K..

Lj-

ai

\0

4.42.

where

; is

Show

that for

Show

4.43.

any partitioned matrix (A

that |xtAxl

Show

4.44.

an arbitrary constant.

\\\\\^

\\s.\f^

= (

(A|0)-'

0),

).

that another definition of the pseudoinverse is A"' = lim


symmetric matrix that commutes vpith Ai"A. ^'*''

(A^A +

eP) - 1 A'',

positive definite

Show

4.45.

||A-J||

if

and only

Answers
Because

4.12.

det

(A- X^I) =

0,

4.14.

No

4.16.

Yes

4.18.

Xi(n)

3,3,-3,

and

xj

ajgCw)

4.20.

/I
4.21.

T =

0.6

0.8

-0.8

0.6

Vo

4.22.

jr

a(-2

any

0.

the column vectors of

1)

is

Supplementary Problems

to

space of at least one dimension.

4.13.

A=

if

where P

The eigenvector

/3(0

for all n.

1 0);

xlg

A-X,I

of

X; lies

y(l

1)

are linearly dependent, giving a null


in the null space of A-XJ.

MATRIX ANALYSIS

98

'O
4.23.

T =

^a

4.24.

'''

+a

/2

(l

Use matrix norms.


If

A-i

value

4.29.

Yes

4.30.

Yes

4.31.

(AT

4.33.

No

+
8et \-4e +

is

eAt

eAt

eat

0\

<x-^[A~-^

+ a^A.^-^ +

e);

].

0,

Xj

If a

XiXg.

as

all

one big Jordan block.

1/

.X

Xj

Xj,

0,

Xj.

Slight perturba-

then at least one eigen-

cos at

e3f

-3e

e3

6e'

e3

-3e

sin

+
+

4e2'
6e2t
2e2

+
-I-

gSt

-gt

gSt

2e

e^'

-e

+
-

2e2t

-1-

e^'

Se^t

_
+
+

gSt'
e^'

e^'^

(oi

sin ut

cos ut

1/40

7^
25 *^-3
(

0^

A-' =

\0

3^
6

0,

There are r nonzero positive eigenvalues


to the eigenvalues

Pi

pj

and r nonzero negative eigenvalues

are the eigenvectors {^^\ and to -pi are f

then gives the desired result.


I

4.43.

singular.

9e2 +
3e2t +
6e2t

/I

4.40.

+ A)x

4.37.

\0

A'^ -

exists,

zero and

is

/_4et

4-36.

and

arbitrary,

4.27.

4.35.

a, r, 6

Xj
x, = (1 0),
Xg = 1
^/2,
(1
Xi = 1 + e/2,
tions on the eigenvalues break the multiplicity.

Yes

4.34.

/l
for

4.26.

4.28.

[CHAP.

|xtAx|

||x||2||Ax|J2

by Schwartz' inequality.

^'V

-p;.

Corresponding

Spectral representation of

chapter 5
Solutions to the Linear State Equation
TRANSITION MATRIX

5.1

From

Section 1.3, a solution to a nonlinear state equation with an input u{t) and an
condition xo can be written in terms of its trajectory in state space as x{t) =
>j>(t; u{t), Xo, to).
Since the state of a zero-input system does not depend on u(t), it can be
written x(i) = ^(i; Xo, io). Furthermore, if the system is linear, then it is linear in the
initial

initial condition so

that from

The

Definition 5.1:

x(i)

Theorem

3.20

we

obtain the

transition matrix, denoted *(i,

^{t; Xo, to)

io),

is

the

nxn

matrix such that

*(f, io)xo.

true for any to, i.e. x{t) *(i, t) x(t) for t > i as well as t t. Substitution of
for arbitrary xo in the zero-input linear state equation dx/dt = A(f)x
gives the matrix equation for *(i, to),

This
x{t)

is

---

^{t, io)xo

d^{t,to)/dt
Sinct! for

any

xo,

xo

x(io)

*(^o, io)xo,

the initial condition on *(i,


^{to,to)

Noti(;e that if the transition

linear differential equation.

A{t)^(t,to)

{5.1)
to) is

[5.2)

matrix can be found, we have the solution


Also, analogous to the continuous time

to a time-varying
case,

the discrete

time transition matrix obeys


*(A;

A(A;)*(fc,m)

^{m,m)

with
so that

+ l,m) =

x(fc)

Theorem

5.1:

*(fc,

(5.3)

(5.^)

m)x(m).

Properties of the continuous time transition matrix for a linear, timevarying system are
(1)

transition property

9{t2,to)
(2)

^-\tuto)

{5.6)

9{ti)9-\to)

{5.7)

separation property

^tuto)
(4)

{5.5)

inversion property
*(fo,ii)

(3)

<S>{t2,ti)9{tl,to)

determinant property
detail,

to)

e-'f

{5.8)

and properties of the discrete time transition matrix are


(5)

transition property
*(fc,m)

99

9{k,l)^{l,m)

{5.9)

SOLUTIONS TO THE LINEAR STATE EQUATION

100

[CHAP.

inversion property

(6)

>{m,k)

*~i(fc,w)

(5.10)

*(m,fc)

fl(m)fl-i(fc)

(5.11)

separation property

(7)

determinant property

(8)

det*(A;,m)

[detA(fe-l)][detA(fc-2)]---[detA(m)]

In the continuous time case, *~'(i, fo) always exists.


cumstances, A{k) may be singular for some k, so there
in equations (5.10) and (5.11) exist.

f or

>

A;

(5.12)

However, in rather unusual cirno guarantee that the inverses

is

Proof of Theorem 5.1: Because we have a linear zero-input dynamical system, the
transition relations (1.6) and [1.7) become *(i, io)x(<o) = *(^, ti)x(ti) and x(ii) = *(<i, io)x(io).
Combining these relations gives *(i,to)x(tG) = *(t, ti)*(ti, to)x(io). Since x(to) is an arbitrary initial condition, the transition property is proven. Setting <2 U in equation (5.5)
and using {5.2) gives *(io, ti) *(ii, to) = I, so that if det *(fo, U) =
the inversion property
is proven.
Furthermore let 9{t) <b{t, 0) and set fi =
in equation {5.5) so that
*(i2, U) = B{t2) *(0, to).
Use of {5.6) gives *(0,to) = *"H*o, 0) = 9-\to) so that the separation
property is proven.

To prove the determinant

property, partition
det

* =

row vectors

into its

^^/\^^/\

^j, ^^,

.,

Then

^.

r\^^

and
d(det ^)/dt

d^Jdt a ^^ '^

+
From

the differential equation

{5.1)

for *, the

a ^ + ^j a d^Jdt

^j A ^2

-^

a.

A.4,^

A d^Jdt

row vectors are

{5.13)

related

by

d^Jdt

2 *(*)

f or

*)c

1, 2,

Because this is a linear, time-varying dynamical system, each element


and single-valued, so that this uniquely represents djtjdt for each t.

^jA

Ad^ydtA

Then from equation

A^

Aj^^a.^^^A

*lA

A^

ai\,{t)

a,^i A

is

continuous

A ^j A

A ^

{5.13),

[tr

A(t)]^i A ^2 A

.^

[tr A(t)]

det

Separating variables gives


d(det*)/det4.

trA{t)dt

Integrating and taking antilogarithms results in


det*(t,to)

where

7 is the constant of integration.

ye-*'"

Setting

to

gives

det *(to,

to)

if and only if
Since e" =
so that the determinant property is proven.
bounded.
are
A(t)
of
elements
the
because
exists
inverse of *(t, to) always

of the properties for the discrete time transition matrix


the reader is referred to the supplementary problems.

The proof
and

is

= 1 = y,
= -, the

det I

f{t)

quite similar,

CHAP.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

101

CALCULATION OF THE TRANSITION MATRIX FOR


TIME-INVARIANT SYSTEMS

5.2

Theorem

The

5.2:

transition matrix for a time-invariant linear differential system


^{t,T)

eA"-^)

{5.U)

and for a time-invariant linear difference system


it{k,m)

The Maclaurin

Proof:

series for

e*'

is

= A"-"

^= A.H'^/k

(5.15)

which

!,

is

uniformly convergent

fc

as

shown

in

Problem

4.27.

Differentiating with respect to

so substitution into equation {5.1) verifies that eA"--^'


t = r, eA(t-T) j^
gQ ^]^jg jg ^j^g unique solution starting

is

is

gives de^ydt

a solution.

from

*(t, t)

A'^+^i'^/fc!,

Furthermore, for
I.

Also, substitution of *(fe, m) = A''" into equation {5.3) verifies that


and for k = m, A"- = I. Note that eAgB ^ gBgA jn general, but e^^'e"^'" =
and Ae*' = e^'A, as is easily shown using the Maclaurin series.

it is

a solution,

e^''e*'

gActo+t,)

Since time-invariant linear systems are the most important, numerical calculation
is often necessary.
However, sometimes only x(f) for t^ta is needed. Then x(i)
can bes found by some standard differential equation routine such as Runge-Kutta or Adams,
etc., on the digital computer or by simulation on the analog computer.
of e*'

When

e^' must be found, a number of methods for numerical calculation are available.
ore method has yet been found that is the easiest in all cases. Here we present four
of the most useful, based on the methods of Section 4.7.

No

1.

Series method:

gAt

2.

{5.16)

Eigenvalue method:
gAt

and,

if

Te-i'T-i

{5.17)

the eigenvalues are distinct,

=
3.

^ AH^Ik

Cayley-Hamilton:

e^i'Xji

n-l

2= y,{tW

oAf

{5.18)

n-l

w'lere the

y^{t)

are evaluated from

e-"

2=

7i(^)J'-

Note that from

{4-15),

eL(Xi)t
gL2i(Xi)e

oJt

{5.19)

e'-"^''!''

gLi2(^2^t

gLmfc'^'^k^t

SOLUTIONS TO THE LINEAR STATE EQUATION

102

where

Ixl,

if Lji(Ai) is

[CHAP.

=1

M^i)

'

(5.20)

I
\

...

,0

\il

then

...

4.

e'^i*

Resolvent matrix:

eAt

where

R(s)

(si

^-i{R(s)}

[5.22)

A)"^

The hard part of this method is computing the inverse of (si A), since it is a polynomial in s. For matrices with many zero elements, substitution and elimination is
about the quickest method. For the general case up to about third order, Cramer's
rule can be used. Somewhat higher order systems can be handled from the flow diagram
of the Laplace transformed system. The elements nj(s) of R(s) are the response of the
For higher
ith state (integrator) to a unit impulse input of the yth state (integrator).
order systems, Leverrier's algorithm might be faster.

Theorem

5.3:

Define the

Leverrier's algorithm.
scalars 9u

62,

F2

= I
= AFi +

= AF-i +

Fi

nxn

real matrices Fi, F2, ...,Fn

and

.,9n as follows:

02

= -trAFi/1
= -trAF2/2

6n

= -tvAF/n

61

^iI

e-iI

Then

Also,

AF +

Having

tfnl

0,

R(s), a

an
S'^-'F^+S''-'F2+
t
n _^n-l
,n
...
^lS"-l

^
-

taT/i\-l
A^-l
(.J
(Sl-A)

to check the method.

S"

Proof

is

+SFn-l+Fn
^^-^"^^

+^_,o4-^
+ dn-lS + 0n

given in Problem 5.4

matrix partial fraction expansion can be performed.

First, factor

detR(s) as
det R(s)

s"

^is-i

e-is

^n

(s

- Ai)(s - A2)

(s

- Xn)

where the Ai are the eigenvalues of A and the poles of the system. Next, expand
matrix partial fractions. If the eigenvalues are distinct, this has the form
R(s)
^

where

R/c is

'

^Ri
Al

A2

R2

^Rn
S

{5.2A)

R(s) in

{5.25)

An

the matrix-valued residue


Rfc

(s

- A,)R(s) |s=x,

{5.26)

CHAP.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

For mth order roots

A,

the residue of

A)"'

(m i)

is

-A^^

Ri

Then

(s

- A)'"R(s)l

\{s

{5.27)

found by taking the inverse Laplace transform.


becomes

e^' is easily

103

In the case of distinct

roots, equation {5.25)

gAt

gA.tR,

gXj,tJJ2

Note, from the spectral representation equation

Ri

so that the eigenvectors

Xj

and their reciprocal

+ eMR

{5.28)

{5.17),

XiTj
r;

can easily be found from the

R;.

In the case of repeated roots,

^-i{(s-Ai)"'"}

To

find A'',

1.

Series method:

methods similar

ESigenvalue method:

{5.30)

A''

= TPT-i

A''

and for

{5.29)

to those discussed to find e^' are available.

A*^

2.

t"'-'e'''V(m-l)!

{5.31)

distinct eigenvalues

= iA^x^rt

A-^

i=l
3.

Cayley-Hamilton;

n-1

A-^

^lAkW
=

{5.32)

whei-e the y^{k) are evaluated

from

J'=

=
I

^
=1

7i('^)J*

where from equation

{A.15),

^Um)
L^i(Ai)

{5.33)

L?,(Ai)
Ul^{k,)

..

and

if L,ii{M) is

as in equation {5.20),
'Xf

L5i(A0

4.

L^(AO

kxf-'

+ !)!] -'^
(fc!Af+2-')[(Z-2)!(A;-? + 2)!]-i
(fc!Af+i-')[(i-l)!(fc-

...

A?

{5.3i)

Jilesolvent matrix:

M^
where

R(z)

{zl

Z-^{zn{z))

{5.35)

A)"^

Since B,{z) is exactly the same form as R(s) except with z for
given previously are exactly the same.

s,

the inversion procedures

SOLUTIONS TO THE LINEAR STATE EQUATION

104

[CHAP.

The series method is useful if A'' =


for some k = fco. Then the series truncates at
1. Because the eigenvalue problem Ax = A.x can be multiplied by A"*"' to obtain
= A''x = AA''~1x = A'^x, then X 0. Therefore the series method is useful only for

ko

systems with
roundoff,

and

poles only at the origin.

A;o

difficulties in

Otherwise

it

suffers

from slow convergence,

recognizing the resulting infinite series.

The eigenvalue method is not very fast because each eigenvector must be computed.
However, at the 1968 Joint Automatic Control Conference it was the general consensus
that this was the only method that anyone had any experience with that could compute
e** up to twentieth order.
The Cayley-Hamilton method is very similar
few more multiplications.

to the eigenvalue method,

and usually

involves a

The resolvent matrix method is usually simplest for systems of less than tenth order.
This is the extension to matrix form of the usual Laplace transform techniques for single
For very high order
input-single output that has worked so successfully in the past.
systems, Leverrier's algorithm involves very high powers of A, which makes the spread of
the eigenvalues very large unless A is scaled properly. However, it involves no matrix
inversions, and gives a means of checking the amount of roundoff in that AF + ! should
equal 0. In the case of distinct roots, R. = x^rf so that the eigenvectors can easily be
obtained. Perhaps a combination of both Leverrier's algorithm and the eigenvalue method
might be useful for very high order systems.

TRANSITION MATRIX FOR TIME-VARYING DIFFERENTIAL SYSTEMS

5.3

There is NO general solution for the transition matrix of a time-varying linear system
such as there is for the time-invariant case.
Example

We

5.1.

found that the transformation


^(t,to)

for the time-invariant case.

A = TJT-i

= *(t-to) =

For the time-varying


dx/dt

Then

A{t)

T{t) J{t)T-Ht),

change of variable

T(t)y

gave a general solution


eA(t-fo'

= TeJ-o)T-i

case,

A(<)x

where the elements of T and J must be functions of

t.

Attempting a

results in

dy/dt

which does not simplify unless dt{t)/dt

J(t)y-T-Ht){dT{t)/dt)y
or

some very fortunate combination of elements.

conclude that knowledge of the time-varying eigenvalues of a time-varying


system usually does not help.

We may

The behavior

of a time-varying system depends on the behavior of the coefficients of

the A{t) matrix.

Example

5.2.

Given the time-varying scalar system di/dt = | sgn (t - ti) where sgn is the signum function, so that
''
sgn(i-ti) = -l for t < t^ and sgn {t - ti) = +1 for t > t^. This has a solution |(i) = |(to)e "
actually
but
stable,
appears
system
<
t^,
the
=
t
For
times
>
tj.
t
l(*i)e"~'i'for
for < tj and |(0
the solution grows without bound as *-*=. We shall see in Chapter 9 that the concept of stability must
be carefully defined for a time-varying system.

Also, the
whereas this

of finite escape time can arise in a time-varying linear system,


impossible in a time-invariant linear system.

phenomenon
is

CHAi'.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

Exam|)le

105

5.3.

Gonsider the time-varying scalar system

d^dt

il{t

ti)2

Then
and

with

tlie

|(to)

?o.

to

<

*i
?o-

[(to-ti)~i- (t-ti)-i]
ioe

The solution goes

solution is represented in Fig. 5-1.

to
Fig. 5-1

infinity in a finite time.

peculiarities make the analysis of time-varying linear systems relatively


than the analysis of time-invariant linear systems. However, the analysis
For instance, a timeof time-varying systems is of considerable practical importance.
varying linear system usually results from the linearization of a nonlinear system about a
nominal trajectory (see Section 1.6). Since a control system is usually designed to keep
the ^'ariations from the nominal small, the time-varying linear system is a good approxi-

I'hese

more

and other

difficult

mEition.

no general solution for the transition matrix, what can be done? In


A computer can almost always
a closed-form solution is available.
find a numerical solution, and with the use of the properties of the transition matrix
(Theorem 5.1) this makes a powerful tool for analysis. Finally and perhaps most importantly, solutions for systems with an input can be expressed in terms of the transition
Since there

is

certsiin special cases

mati'ix.

CLOSED FORMS FOR SPECIAL CASES OF TIME-VARYING


LINEAR DIFFERENTIAL SYSTEMS
Theorem 5.4: A general scalar time-varying linear differential system

5.4

d^/dt

a{t)^

has

the scalar transition matrix


<l>it,T)

Proof:

Theorem

e^-

Separating variables in the original equation,

antilogarithms gives

taki:!ig

5.5:

If A(i) A(t)

i{t)

lo^

"

"

d^/i

is

A{t) A{t) for all t,T, the time-varying linear differential sys-

a severe requirement on A{t), and

Proof:

Use of the

series

Integrating and

a{t)dt.

tem dx/dt = A(i)x has the transition matrix


This

is

usually

*(t, t)

met only on

final

'^

examinations.

form for the exponential gives

Taking derivatives,
^g/>^'*''

But from equation

+lA{t) A{r,)drj +lj\{v)d^A{t)

A{t)

{5.86),

A(i)e-^-'''""*

This equation and {5.37) are equal


A{t)

=
if

A{t)

and only

^^A{r,)d^

A(i)

A(r?)d,

if

J\{r,)dr,A{t)

(5.37)

SOLUTIONS TO THE LINEAR STATE EQUATION

106

Differentiating with respect to r and multiplying by

A(i)A(r)

Only A{t) A(t) = G{t,


r for t and t for r

of

Example

gives the requirement

A{r)A{t)

need be multiplied in the application of this

r)

will then indicate if

G(<,t)

G(t,

test.

Substitution

t).

5.4.

Given

A{t)

/ 1^

=
(

G(^.<)

-1

[CHAP. 5

Theorem

5.6:

Then from

A(t) A(t)

G(t,T)

we

see

immediately

G{t,r).

A piecewise time-invariant system, in which A{t) = Ai for U^t^tt+i


for i = 0,1,2, ..
where each Ai is a constant matrix, has the transition
.

matrix
^{t,to)

eAi"-'i>$(f.^io)

for

ti^t^ti+i

Proof:
Use of the continuity property of dynamical systems, the transition property
(equation (5.5)) and the transition matrix for time-invariant systems gives this proof.

Successive application of this theorem gives


Ht'to)

eA"-*'

for

to^t^U

etc.

Example

5J5.

Given the flow diagram of Fig. 5-2 with a switch S that switches from the lower position to the
upper position at time t^.
Then dx^/dt = Xi for o - * < *i and dx^/dt = 2xi for i ^ t.
The
solutions during each time interval are Xi{t) = aJiQe'"'" for t^^ t < t^ and x^it) = a;j(ti)e2"-*i> for
ti

t,

where

Xiit^)

a;ioe*i'"*<'

by continuity.

Fig. 5-2

common practice to approximate slowly varying coefficients by piecewise conThis can be dangerous because errors tend to accumulate, but often suggests means
of system design that can be checked by simulation with the original system.
It is

stants.

Another special case

is

that the nth order time-varying equation

d<"

can be solved by assuming y


characteristic equation.

nomial,

2/

(lni)'""*i^

V'.

Then a

scalar polynomial results for

order

If there are multiplicities of


is

a solution for

0, 1,2,

.,m 1.

A,

analogous to the

in the solution of this poly-

CHAP.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

107

A number of "classical" second order linear equations have closed form solution in the
sense that the properties of the solutions have been investigated.
Bessel's equation:
t'y

(1

- 2a)tif +

+ (a^ - py)]v =

[l^yt^y

(5.38)

Associated Legendre equation:


(1

-t^)y- 2ty +

Hermite equation:

or, vi^ith z

+ {l-t^ + 2a)z =

Laguerre equation:

j.-

ty

with solution

L(t),
^"

with solution

d^L{t)/dt''.

or

s"
ty

+ 1) - toV(1 - f)]y =

-2ty + 2ay =

e^^'^y,

{n(n

J.^
+
+ {l-t)y
/-,

ay

j.\i\
(k + l-t)y + {a-k)y =
+n
,

-,

Hypergeometric equation
^'''^'"^'^t{l

-t)y+[y-(a + l3 + l)t]y - apy

Confluent:

Mathieu equation:
i/

or,

with

cos^

ty

{y-t)y

(a

/3

-ay =

cos t)y

t,

4T{l-r)y + 2{l-2T)y +

[a

+ l3{2T-l)]y =

and details on their behavior are available


matheinatics and mathematical physics.

The

{5.39)

sdlutions

in standard texts on engineering

time-varying case are a number of methods to give *(i, t)


Peano-Baker integration, perturbation techniques,
iteration,
Picard
as an
rapid convergence. However, even only three
give
quite
sometimes
and
etc., can be used,
complicate any sort of design procedure,
representation
greatly
or four terms in a series
texts. Use of a digital or analog
standard
left
to
so discussion of these series techniques is
form solution is not readily
a
closed
computer is recommended for those cases in which
AI(30 available in the linear

infinite series.

found.

5.5

PERIODICALLY-VARYING LINEAR DIFFERENTIAL SYSTEMS

Floquet theory is applicable to time-varying linear systems whose coefficients are constant or vary periodically. Floquet theory does not help find the solution, but instead gives
insight into the general behavior of periodically-varying systems.

Theorem

5.7:

(Floquet).

Given the dynamical linear time-varying system

where

A{t)

A{t

o,).

Then

m,r) =
where

P(f, t)

F{t + <a,T) and

is

P(i,T)eR"-^'

a constant matrix.

d's./dt

A(^)x,

SOLUTIONS TO THE LINEAR STATE EQUATION

108

The

Proof:

Setting

[CHAP.

transition matrix satisfies

= t + m, and

d^{t,r)ldt

using A(i)

A{t

A(i

+ ,)$( +o,,t) =

a*(f+o,,T)/a*

with

A(i)*(i,r)

*(r,T)

gives

a.)

A(i)*(^ + a.,T)

(5.40)

was shown in Example 3.26 that the solutions to cbs./dt = A(i)x for any initial conform a generalized vector space. The column vectors ^^t, r) of *(i, t) span this vector
space, and since det *(i, r) v^ 0, the ^i,t, r) are a basis. But equation
{540) states that the
It

dition
^^{t

<o,

t)

are solutions to

dTi./dt

A(i)x,

so that

i>i(t+o>,T)

Rewriting this in matrix form, where

C=

z=

for

c^ii'iit.r)

1,2,

.,%

{Cji},

*(+,t)

*(i,r)C

(5M)

Then

C =

*(r,t)*(i + o>,T)

Note that C-^ exists, since it is the product of two nonsingular matrices.
Problem 4.11 the logarithm of C exists and will be written in the form

C =
If P{t,

r)

can be any matrix,

it is

e"R

Therefore by

(5.^^)

merely a change of variables to write


9{t,r)

P(,r)eR-->

(5.^5)

But from equations (543), {541) and {542),


P(i+<9,

t)

*(i+<o,T)e-R"+"-^'

*(,r)e-R--'

#(;t,T)e"Ke"R'^'+"-^>

V{t,r)

From

{5 41)-{5 43), R = a.-Mn [*(r,) *(i + a.,.)] and P(,r) = *(,r)e-H"-^


so that to
and P(i, r) the solution $(*, r) must already be known. It may
be concluded that
Floquet's theorem does not give the solution, but rather
shows the form of the solution.
The matrix V{t, r) gives the periodic part of the solution, and e^"-^'

hnd

gives the envelope of


Since e<'-> is the transition matrix to dz/dt = Rz, this
is the constant
coefficient equation for the envelope x{t). If the
system dz/dt = Rz has all poles in the left
""^^"^I time-varying system x(i) is stable. If R has all eigenvalues
in the
fJ"^',
lett if
halt
plane except for some on the imaginary axis, the
steady state of z{t) is periodic
with the frequency of its imaginary eigenvalues. To
have a periodic envelope in the sense
that no element of x{t) behaves exponentially,
all the eigenvalues of R must be on the
imaginary axis. If any eigenvalues of R are in the right
half plane, then z{t) and x(t) are
unstable. In particular, if the coefficients of
the A{t) matrix are continuous functions of
some parameter a, the eigenvalues of R are also continuous
functions of , so that periodic
solutions form the stability boundaries of
the system.
the solution.

fl

Example

5.6.

Mathieu functions, which exist only for certain


combinations of a and fl.
The values of a and fi
for which these periodic solutions
exist are given by the curves in Fig. 6.3
below.
These curves tlef
form the boundary for regions of stability.

CHAP,

SOLUTIONS TO THE LINEAR STATE EQUATION

fi]

109

Fig. 5-3

Whether the regions are stable or unstable can be determined by considering the point ;8 =
and
ill region 1.
Since the curves
This is known to be unstable, so the whole region 1 is unstable.
a <
are stability boundaries, regions 2 and 6 are stable. Similarly all the odd numbered regions are unstable
represents a degenerate case,
and all the even numbered regions are stable.
The line P 0, a
which agrees with physical intuition.

from the example above that an originally unstable system


by the introduction of a periodically-varying parameter, and vice versa.

It is interesting to note

might be

stabilized

Another use of Floquet theory is in simulation of *(f, t). Only *(t, t) for one period <*
need be calculated numerically and then Floquet's theorem can be used to generate the
solution over the whole time span.

SOLUTION OF THE LINEAR STATE EQUATIONS WITH INPUT

5.6

Knowledge of the transition matrix gives the

solution to the linear state equation with

input, even in time-varying systems.

Theorem

Given the linear differential system with input

5.8:

with transition matrix


(5.1)].

dx/dt

A()x

+ B(Qu

C{t)x

+ D{t)u

*(t, t)

obeying

3*(i, T)ldt

{2.39)

A{t) *(t,

t)

[equation

Then
x(t)

yit)

*(i,o)x(to)

+ r

C()*(t,io)x(to)

*(, t) B(t) u(t) dr

+ r

C(t)^t,T)B{T)u{r)dT

D(i)u
i5.U)

The

integral

is

the superposition integral, and in the time-invariant case

it

becomes a

convolution integral.

Since the equation dx/dt = A(t)x has a solution x{t) = *(i, U), in accordance
Proof:
with the method of variation of parameters, we change variables to k(f) where
x(^)

*(i,io)k(t)

{545)

SOLUTIONS TO THE LINEAR STATE EQUATION

110

[CHAP.

Substituting into equation (2.39),

dy.ldt

Use of equation

dk/dt
to

ta

Since equation (545) evaluated at

gives

gives

*(to, t)

^{to,t)B{t)u{t)

*(i, io)

k{to)

+ f

=U

gives x(io)

$(fo,i)x()

Multiplying by

*(io, t) B(t) u(t)

+ f

x(io)

completes the proof for

dr

(546)

use of (545) in (5.^6) yields

k{to),

*(io, t) B(r) u(t)

x(t).

dr

Substituting into

y(i)

C(i) x(i)

y{t).

In the constant coefficient case, use of equation

x(i)

eA(t-t)x(io)

and

(5.14.)

gives

eA~^)Bu(T) dr

(5.^7)

^t

y()

This

A{t)$k + B(t)u

t,

k{t)

D(i) u(^)

$dk/di

and multiplication by

{5.1)

Integrating from

(a*/af)k

is

CeA"-*'x(io)

CeA<'-^'Bu(T) dr

Du(i)

(5.^5)

the vector convolution integral.

Theorem

5.9:

Given the linear difference equation


x(A;

+ 1) =

A(A;) x{k)

C{k) x{k)

+ D{k) u{k)

y{k)
vs^ith

transition matrix

*(&,

m) obeying

*(fc

{240)

+ 1, m) =

m)

A.{k) *(&,

[equa-

Then

tion {5.3)].

= nA(i)x(m)+

x(A;)

i=m

where the order of multiplication

Proof:

B{k) u{k)

Stepping equation

1)

A(z) B(i) u(y)

B(A;

- 1) u(A; - 1)

=j + l

(5.49)
i.e.

A.{k \) A{k 2)

up one gives

x(m + 2) = A(m +

x(m + 2) = A(to +

11

starts with the largest integer,

{2.40)

Substituting in equation {2.40) for

^
=m

+ 1) + B(m + 1) u(m + 1)

1) x(to

x(w +

1),

A(m) x(?n) + A(m +

1)

B(m) u(m) + B(m +

1)

u(m +

1)

Repetitive stepping and substituting gives


x{k)

= A{k - 1)

A(to) x(m)

+ A(A; - 1)

A(fe

- 1)

A(m + 1) B(m) u(m)

+ 2) B(to + 1) u(m + 1) +
+ B(A; - 1) u(A; - 1)

This

is

A(to

+ A(/(; - 1) B{k - 2) u{k - 2)

equation {5.49) with the sums and products written out.

CHAP.

5.7

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

111

TRANSITION MATRIX FOR TIME-VARYING DIFFERENCE EQUATIONS


Setting B =
in equation (549) gives the transition matrix for time-varying difference

equations as
Ic-l

- Yi Mi)

m)

*(fc,

For k = m, *(m,m) = I, and if A-^{i)


Then equation (549) can be written as
x(A;)

for

f^

exists for all

>

{5.50)

^{k,m)

i,

Yl^~K'>')

for

k<m.

'"''

*(fc,m)x(m)

^
=m

*(fe, ;

+ 1) B(i) u(i)

{5.51)

This is very similar to the corresponding equation (5.^4) for differential equations
except the integral is replaced by a sum.

Often difference equations result from periodic sampling and holding inputs to

dif-

ferential systems.

^' (h)

u(i)

dx/dt

Hold

+ B(t)u
+ D(i)u

y{t)

A(i)x

C(t)x

yfc).

<r-

Fig. 5-4

In Fig. 5-4 the output of the hold element is u(fc) =


'k+i tk-T for all fc.
Use of {5.H) at time 4 = 4+1 and

u(tfc)

U=

for 4
h gives

*(i)c +

Comparison with the difference equations


As(k)

^{tk+i,tk)

Bs(k)

Cs(k)

C(k)

D^A;)

where the subscript

ment

As

is

e^'^,

a matrix exponential,

Bs

it is

r)

B(t) u(A;) dr

where

{5.52)

results in

^*(4+i,T)B(T)dT

D(4)

s refers to the difference equations of the

invariant differential systems, As


in this case

(2.4-0)

1,

f<4+i,

sampled system.

e^'^'^-'^B dr, Cs

= C and Ds =

nonsingular no matter what

For timeD.

is (see

Since

the com-

after Problem 4.11).

Although equation (5.50) is always a representation for ^{k,m), its behavior is not
usually displayed. Techniques corresponding to the differential case can be used to show
this behavior.
For instance, Floquet's theorem becomes *(/<;, m) = P(fc,m)R''"'", where
F{k, m) = P{k + o>, m) if A(A;) = (k + a>). Also
(Il:^')ly{jc

+ n) +a,

^^'^l~^^-

y{k

+ n~l) +

a^_^(k

+ l)y(k + l) + a^k) =

has solutions of the form XVA;


Piecewise time-invariant, classical second order
and sesries solutions also have a corresponding discrete time form.
!

linear,

SOLUTIONS TO THE LINEAR STATE EQUATION

112

5.8

IMPULSE RESPONSE MATRICES


With zero initial condition x{to) = 0, from
=

y{t)

This suggests that

y(i)

C(i) *(i,

r)

{54i) the output


B(t) u(t) dr

[CHAP.

y{t) is

D(i)u(i)

(5.53)

can be written as a matrix generalization of the superposition

integral,

where
time

{5M)

H{t,r)u{T)dr

t) is the impulse response matrix, i.e. hij{t, r) is the response of the ith output at
due to an impulse at the yth input at time t. Comparison of equations (5.53) and

H{t,

{5.54-)

y{t)

gives

nit,r)

(5.55)

^^^

In the time-invariant case the Laplace trans:?ormation of H{t,

gives the transfer function

0)

matrix

= C(sI-A)-iB+D

J:{U{t,0)]

Similarly, for discrete-time systems

y(fc)

(5.56)

can be expressed as
k

2
m=

yik)

H(A;,to)u(to)

(5.5?-)

00

where
'C(A;)*(A;,m +
H(A;,

Also, the

2 transfer

m)

D(fe)

<m

function matrix in the, time-invariant case


2:{H(A;,0)}

5.9

k>m

l)B(m)

(5.58)

is

= C(2l-A)-iB + D

{5.59)

THE ADJOINT SYSTEM


The concept

Definition 5.2:

of the adjoint occurs quite frequently, especially in optimization problems.

The

adjoint, denoted La, of a linear operator


(p,

We are

Lx)

concerned with the system d^ldt

for

(Lap, x)

A{t)x.

Defining
ptxdi,

equaLiuii yo.ou)

usmg migration oy

(p,Lx)

=
"

For the case

p(^o)

parts.

f\^A{t)xdt-

X
p(^i),

[''^^^P

we

all

L=

is

defined

by the

x and p
A{t)

- Id/dt,

the adjoint system

is

f\f~dt
VHU)K{t,)

A^{t)+Id/dt.

relation
{5.60)

"'

+ ^'^lf]*''* +

find La

pt(ti)x(*i)

this

becomes

found from

CHAP.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

Since Lx
cluded Lap =

Lx) =
so that the adjoint system
for

0,

then

all x,

for

(p,

x and p. Using
by the relation

{5.61)

with

a^it,to)/dt ==,9^A^{t)t{t,to)

5.10:

Given the system

dx/di

A(i)x

pt(f,)x(i)

Also,

Proof:

*t(i,io)

u(i)

adjoint system

its

dp/dt

pHt)B{t)u{t)dt

{5.63)

*-\t,to)

{5.6i)

^i(*'*o)

Differentiate p+(f)x() to obtain

Using the system equation dx/di

(dpt/dQx

= A(i)x + B{t)u
d(ptx)/dt

Integration from

to

to

then yields

ti

the transition relation,

x(f)

p+(io)Ix(io)
p(fo)

and

and equation

ptBu

and

u{t)

if

p(<)

0,

(it,

to)

p(o)

so that

p+(M*t(i,o)*(,o)x(o)

must

hold.

The adjoint system tr^ftsition matrix gives another


of {5.U):
,
K{t)

{5.61) gives

pt(*)x(f)

*(f, to) x(fo)

Therefore*(5.(J4)

x(io).

+ ptdx/di

Furthermore

(5.55).

pt(io)x(o)

for any

(5.62)

0,

d(ptx)/dt

From

of *(, J aS* the reciprocal basis to the column vectors


then pt(t)x(f) = scalar constant for any .

i>i(t,t^)

if

*(io,io)

+ B(Qu and

pt(io)x(io)

and
The column vectors

to), i.e.,

Then

-At(i)p.

of^t,to).

can be con-

-At(i)p

Denote: the transition matrix of this adjoint system as *(t,

Theorem

(5.60) it

defined

is

dp/dt

all

113

*(i,o)x(io)

way

to express the forced solution

*(i,T)B(r)u(T)dr

'to

The variable of integration t is the second argument of *(t, t), which sometimes poses simulation difficulties.
Since <to,T) = *-/(t,<o) = '+(t,o), this becomes
x(f)

4^it,to)\x{to)

J i^T,to)B{r)uir)dT']

in whi(;h the variable of integration t is the first

argument of

Thi3 adjoint often can be used conveniently when a


motion backwards in time must be found.
Exampl*!

*(t,

{5.65)

to).

final value is

given and the system

5.7.
'

Givim

dx/dt

[of

2t
2f

]*

^^^ ^^^ adjoint system

permit the system to pass through the point

The adjoint system

is

dp/dt

*{t,r)

/I
tl

Xi(2)

to

find

the set of states

1.

3\
jp.

0.2e<^/2)-(r^/2)

This has a transition matrix

/ 3 -3\

'-22/

o.Ee^^-^'V^
I

^
3

(xi{l) X2(l))

that

SOLUTIONS TO THE LINEAR STATE EQUATION

114

Since

The

= pt(l)x(l),
+ P2(l)a;2(l). But

pt(2)x(2)

Pi(l)a;i(l)

set of states x(l) that gives

we choose pt (2) = (1 0), then (1


= (!, 2) p(2), so that
pt(l) = 0.2(3e-i-5 + 2e6 2^-2e-i-5)
Xi(2) = 1 is determined by
'

if

p(l)

(0.6e-i-5

+ 0.4e6)a;j(l) +

0) x(2)

[CHAP.

a;i(2)

pt(l) x(l)

(0.4e6-o.4e-i-5)a;2(l)

Solved Problems
5.1.

Given ^t,to), find


Using

A{t).
at time

(5.2) in (5.1)

to

*.

Mi) =

evaluated at

3*(*, to)/St

to

This

is

a quick

check on any solution.

5.2.

Find the transition matrix for the system

/-I
-1

method, (b) eigenvalue method, (c) Cayley-Hamilton method, {d) resolvent matrix method. In the resolvent matrix method, find (si A)~^ by (1) substitution and elimination, (2) Cramer's rule, (3) flow diagram,
(4) Leverrier's

by

(a) series

algorithm.
(a)

Series method.

From

(5.16),

eAt

= i + At/V. + AHy2l +

Substituting for A,

/t2/2
eAf

-4t

4t

6t2

-8t2

2t2

-2t2/

- 2t(l - 2f) +
)
-t(l - 2t + 4f2/2 +

2t

4t2/2

- 2t + 4t2/2 + )
+ 4t2/2 + 2t(l - 2t) +

4t(l

2t

Recognizing the series expression for e~' and e~2t

/e-t
eAt

\
(6)

(l-2t)e-2t
-te-2

(l

+ 2t)e-2t/

The eigenvalues of A are -1, -2 and 2, with corresponding eigenvectors


The generalized eigenvector corresponding to 2 is (0 1 1)'', so that

Using equation

{5.17),

eAf

A
=02

Multiplying out the matrices gives the answer obtained in

(a).

(1

0)'''

and

(0 2

1)'''.

CHAP.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

|c)

Again, the eigenvalues of

-2 and

are calculated to be -1,

-2.

To

115

find the yAt) in equation

(5.18),

/I

el

Vo

0\
1

6-2',

/-I

\0

Yi

1/

0\

0-2

/I

72

-2/

\0

which gives the equations

=
=
=

e~'
~^'
ie-2t

Solving for the

yg

yj

then gives

0\

(4e--3e-2t-2e-2t)[

/-I

(e-t-e-2t-ie-2t)

(4e-<- 4e-2t-

3e-2)

1/

/I

0^

12 -16

Vo

(dl)

472

4e-

\0

Summing

T2

- 2yi +
- 472

ro

/I
eAt

3e-2t - 2te-2'
= 4e- 4e-2t - Ste-^t
= 6- e-2t ie-2t
=

7i

(5.18)

Vi

7j,

7o

Using

70

-4,

these matrices again gives the answer obtained in

(a).

Taking the Laplace transform of the original equation,


SJ:(Xi)

Xio

-^(Xj)

S^{X2)

X20

-^JLiXi)

sJLix^

a;ao

-^(a;2)

4^(X3)

Solving these equations by substitution and elimination,

~1
2
_ r 1
- [_7+2~(7T2)2j"'20 +

^^*'2)

^^^3)

Putting this in matrix form

a^io

<!(a;i)

.C(x)

(g

+ 2)2

Ls +

(s

4x
430

(^
+

2)2

R(s)a;o.

/
R(s)

'+

=
s

(s
\

(s

+ 2)2

(s

+ 2)2

Inverse Laplace transformation gives eAt as found in (a)


(d2)

From

(5.22),

+
R-i(s)

10
s

-4

1s
4

+ 2)2
(s

+ 2)2

0^
4

-4
-1

-4
4,

'

SOLUTIONS TO THE LINEAR STATE EQUATION

116

Using Cramer's

[CHAP.

rule,

R(8)
(8

+ X)
+ 1)

8<8

+ l)(s +- 2)2

-(s

{s

4(8 + 1)
+ l)(s + 4)^

Performing a partial fraction expansion,


'0

0'
R(s)

8+1

+ 2)2

(s

'

^0

Addition will give R(8) as in


(d3)

The

0-2

-1

(dl).

flow diagram of the Laplace transformed system is

shown

in Fig. 5-5.

1*10
<:(a^3)

C(i)

1_

<3Hj

iO

i.

Fig. 5-5

For

xio

For

Kjo

For

a^ao

= 1/(8 + 1) and J^(x^ = Ak^^ = 0.


- 4/{ + 2)2.
1, C(*i) = 0' -^(^2) = /(* + 2)2 and J^{^
=1, ^(i) = 0, .c(a;2) = "lA* + 2)2 and .^W = ( + 4)/(s + 2^-

=
=

1,

.C(i)

Therefore,
\

8+1
4
R(s)

T^TW

(8

^1
\

Again a
(d4)

(7+2)2

(s

+4
+ 2)2

partial fraction expansion can be performed to obtain the previous result.

Using Theorem

5.3 for Leverrier's algorithm.

F,

Using equation

51

81

9i

92

fls

(5.25),

74

0\
1

R(s)

+ 2)2

sf

s8

5s2

8s

partial fraction expansion of this gives the previous result.

CHAP.

5.3.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

Using (a) the eigenvalue method and then


matrix for

+ l)

x(fc

(a)

the resolvent matrix, find the transition

(6)

117

_^)x(fc)

(-J

The eigenvalues are -1 and -2, with eigenvectors (1 0)r and (1 -1)^ respectively. The
is (1 1) and (0 -1).
Using the spectral representation equation (5.31),

re-

ciprocal basis

A"
(6)

From

equation

(-1)"

(1

1)

fj^

(-2)fc

-1)

(0

(5.35),

(' +

R(z)

-M"'

''

+ 2y

^+^

{z

+ 2)\

l)(z

so that

'

__
z

5.4.

(-2)'^

Prove Leverrier's algorithm,

j_.v-l

+ s''-^F2 +

s"-'Fi

''

5lS"-l

and the Cayley-Hamilton theorem (Theorem

sFn-i

On

^fn-lS

Fn
{5.23)

4.15).

Let
det(sI-A) = <f>{s) = s" + 9iS"-i +
+ e_is + .
Use of Cramer's rule gives
0(s)(sl- A)-i = F(s) where F(s) is the adjugate matrix, i.e. the matrix of signed cofactors transposed of (si A).

An intermediate result must be proven before proceeding, namely that tr F(s) = d^/ds. In the
proof of Theorem 3.21, it was shown the cofactor e^j of a general matrix B can be
represented as
Ci^-ejAb2A Ab and similarly it can be shown Cy = bj a
Abi_i a e^ Ab^+i a
a b, so
that letting B = si - A and using tr F(s) = Cii(s) -f- Czzls) 4
+ c{s) gives

tr F(s)

cj

(sea

- ag) a

(sci

(se

- a) +

-ai) A

(scj

- a{) ac^a

A(se_i-a_i) a

a.

(sej

- a)

But

0(s)

det

- A) =

(si

(sei

- aj) a

- ag) a

{se2

(se

- aJ

and
d<p/ds

ei

A(se2-a2) A

and so the intermediate result

A(se-aJ + -!-

(scj

- aj) a

Ae

trF(s)

is established.

Substituting the definitions for 0(s) and F(s) into the intermediate result,

= tr (s^-iFi + si-SFa-l- +sP_i + F)


= s^-itrFi + s"-2trF2 +-!- strF_i

trF(s)

d<f>/ds

Equating

like

powers of

?is"-i

+ (n

1,2, .. .,n

l,
(S"

trF

-I-

g^_^

s,

trFfe +

for

l)9iS"-2

4.

and trFj

+ ffiS"-! +

and equating the powers of

s gives

n.

{n-k)$k

Rearranging

{5.23) as

+ 9)I = (sl - A)(s-lpi


I = Fj, flI = AF, and
fffcl

-AFfc

(5.66)

Ffe

-I-

sn-2F2

+ F)

(5.67)

SOLUTIONS TO THE LINEAR STATE EQUATION

118

for

A;

1, 2,

.,w

These are half of the relationships required.

1.

[CHAP.

The other half are obtained

by taking their trace

nOk

and substituting

-trAFfc

trFfc +

into (5.66) to get

fcflfc

-trAFfc

which are the other half of the relationships needed for the proof.

To prove the Cayley-Hamilton theorem,

Using the

last relation

Fi

F2

ej.

Fg

e^I

9_il

eI

= AF
=

which

5.5.

is

+ AFi =
+ AF2 =

ffil

+A

82I

+ 9_2A +

i?iA

i.e.,

A2

eiA-2

A"-

then gives
9-i A

eI

successively substitute for F^+x,

(?iA"-i

+ A" =

0(A)

the Cayley-Hamilton theorem.

Given the time-varying system


dt

V-e-*

Find the transition matrix and verify the transition properties [equations
Note that A(t) commutes

A(t A(t

vidth A(t),

{5.5)-{5.8)].

i.e.,

.,

-ff+;J

A(r A()

can be shown similar to Problem 4.4 that two nXn matrices B and C with n independent eigenvectors can be simultaneously diagonalized by a nonsingular matrix T if and only if B commutes
with C. Identifying A() and A(t) for fixed t and r with B and C, then A(t) = T(f)A(t)T-i(t) and
eigenA(t) = T{t)A{t)T-Ht) means that T() = T(t) for all t and r. This implies the matrix of
then a general
vectors T is constant, so dT/dt = 0. Referring to Example 5.1, when d1{t)/dt =
It

solution exists for this special case of

A(t)A(T)

A(T)A(t).

For the given time-varying system, A(t) has the eigenvalues Xi = a + je-* and
1)'^.
Since A() commutes with A(r), the eigenvectors are constant, {j -1)^ and (j
)

From Theorem

i\

./.x

fa-je-t

rt,,

5.5,

*(t,T)

,r~i

- 1

Te-'-r

T-i

Substituting the numerical values, integrating and multiplying out gives

*(t,r)

act-T)
e
^

/
(

cos(e-''-e~')
_gijj(g-r_e-t)

sin(e-''-e-*)

cos(e-r-e-t)

X2

= a-je'K

Consequently

CHAP.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

To check

5'*

this,

use Problem

5.1:

\-sm{e^

Setting T

sm{e-r-e-t)\_^^^^_^^^/sm(e-r-e-t)-cos(e-r-e-t)

''^^'''"-*'~'^

^e-it-r)(

<'t

119

e~t)

cos{e^

e-t)J

\cos(e^

e't)

sm{e^e-*)

in this gives A().

= *(2. <i) *(h> *o). note


*(<2,

cos(e~'o
e~h) = cos (e~'o e~*i + e~'i e*2)

To verify

i^o)

cos (e~'o

e~*i)

cos (e~'i

e~*2)

sin (e~'o

e~ti)

sin (e~*i

e~*2)

g-tj) =

sin (e""*o

g-ti)

cos

e""'2) +

cos (e~'o

e~'i)

sin (e~'i

e~'2)

and
sin (e~'o
SIC

(e""'i

that
cos (e~*o

-^rt-,^f

e~2)

sin (e~'o

e-'2)\

cos (e~'i

e~'2)

\ sin (e~'i
cos (e~'o

g-ti)

V sin(e~'o
To verify

*-i(j,

fo)

*(*o.

*i).

calculate ->(*!,

^'

"

cos^ (e^'o e~'i)

cos2(e-*o
equals *(*o, *i)-

liUnce

To verify

e~"'i) +

*(i, *o)

sin2 (e^'o

sin^ (e~*o

g-ti) =

(<i)-'(o).

9{t)

sin (e~'o

cos (e~*o

'i)

e-'2
e~*2)
e~'i)
e^'i)

rule:

e-'i) sin (e~*o e~*i)


cos (e~'o e~*i)
sin (e~'o e-i) = sin (e-*i e^o),

/cos (e~'o

gacto-ti)

by Cramer's

<o)

sin (e~'i

cos (e~'i

e~'2)

g-ti)

and

sin (e~'o e~'i)

this

set

- e-t)
cos (1 e~t)

cos(l-e-')

et/

sin (1

\^ sin (1 e~*)

Then
9-1(0

e-t

(1-^"') -sin(l-e-*)^

/'=''

e-*)

\sin(l
Since

cos(l

e-)^

we have
cos(e-o-e-i)

and a similar formula for the


Finally, det*(t,T)

cos(e-o-l) cos(l-e-i)

and

sine, multiplication of (i)

e2(t-T)

and trA()

sin (e-<o- 1) sin (1


fl-i(<o) will

- e-i)

verify this.

so that integration shows that the determinant

2a:,

property holds.

5.6.

Find the transition matrix for the system

dicldt

(
,-1

Writing out the matrix equations, we


d^X^/df^

(2a

Multiplying by the integrating factor

(^2

dxjdt

find

l/t)dx^/dt
te*,

-2a -

+ a)lt
+

=
[1

Xg

lit;

and so
(a2

+ a)/t]xi

which was found by much

trial

and error with the

effuation, gives
t(e<^*

d'^xjdt^

2ae"tdxi/dt

a^e^tx^)

{e"t

dx^/dt

aea;i)

te^^Xi

which can be rewritten as


td^e<^*x^)/dP'

d(ea;i)/df

Tiis has the same form as Bessel's equation

x^it)

= dxjdt =

(5.38), so

-ax^(t)

te^tx^

the solution

e-''t[-c^Ji(t)

is

c^dYa/dt]

SOLUTIONS TO THE LINEAR STATE EQUATION

120

To

solve for the constants c^

so that

x(t)

and so

*(<,t)

at the point

and

[CHAP. 6

Ca,

FG(t)G-MT)F-ix(T),

where

0\

-a

1/

= FG() G-^WF"'= 0, the elements of

^^ ^

^(t)

V-^i(t)

FoCt)

dFo/d*

or t and t < 0, because


This is true only for t and t >
This accounts for Fo(0) = "
the original A(t) matrix blow up.

Admittedly this problem was contrived, and in practice a man-made system would only
However, Bessel's equation often occurs in nature, and knowledge
and Y^if) ~ ^ilvi sin(t-ir/4) gives great insight
that as t^ , Jo(t) ~\/27S cos{t-!r/4)
accidentally have this form.

into the behavior of the system.

5.7.

Given the time-varying difference equation x( + 1) = A() x(n), where A(n) = Ao


Find the fundamental matrix, analyze by
if n is even and A() = Ai if n is odd.
Floquet theory, and give the conditions for stability if Ao and Ai are nonsingular.

From

equation

(5.50),

m is even
if
is odd and m is even
is odd and m is odd
if
if k is even and m is odd

AiAqAi- -AjAo

if

*(A;,

AqAiAo-'-AiAo

m)

Ai AqAi

even,

*(fc,

m)

(AoAi-i)''2 if k is odd.
P(fe,m) = {AiAo-i)i'2
must be outside the unit circle.
it is enough to find the eigenvalues

5.8.

+ 2) =

where

A^Ao.

AiAox(m).

Find the impulse response of the system


d^y/df"

Choose
with A(t).

(1

- 2a) dy/dt +("- + e-^')y = u

which A(t) commutes


and x^ = eKdxi/dt- ax^) to find a state representation in
Then in matrix form the system is

xi^y

= a

From

equation (5.55) and *{t,

t)

This

is

the response

y(.t)

to

an input

In the system of Problem


Find the complete response.

From

5.8,

e''

u(t)

let

equations (5.44) and Problem


y{t)

0)x

obtained from Problem


H(t,T)

5.9.

P(fc.m) = I if fc is even and P(fc,m)


P(fc,m) (AoAi)(fc-"'2, ^here P(fc,m) =_I if fc is
For instability, the eigenvalues of R - (AiAo)'
if fc is even.
Since the eigenvalues of B^ are the squares of the eigenvalues of B,
This agrees with the stability analysis of the equation
of

m) (AiAo)"'-'>'2,
For m odd, *(fc,m)

P(fc,

odd and

x(n

even and

A:

AflAi

is

fe

AqAiAq- -AqAi

For

5.5,

+ '"-^> sin{e-''-e-')

Sit-r).

u{t)

6^"-''^

y{to)^yo

and

{dy/dt){to)

5.5,

e-to) cos (e-to

- e-)vo +

"'

"'' sin

(e"^

- e'*) dr

ayo.

CHAP.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

Changing variables from

r to

y(t)

ij

where e-'^e-*

in the integral,

ef-o>

cos(e-to-e-t)j/o

e-*[l

gives

ii,

(e--

cos

121

e-*o)]

Notice this problem cannot be solved by Laplace transformation in one variable.

5.10.

(Jiven a step input

U{s)

into a system with a transfer function

6/s

rr/

jl^'ind

the output y{t) assuming zero initial conditions.

The

easiest

way

to do this is

^{y(t)}

by using

classical techniques.

+
y
'i.
+
+ 6s

U(s)H(s)

1+3-4
+
+

os^

s-*

Taking the inverse Laplace transform determines y

Doing this by state space techniques shows


I'rom Problem 2.3 the state equations are

The

how

= 1 + 3e~2t 4e-3t.
corresponds with the classical techniques.

it

^G)

Co

(-1 2)1
V^2

-3)0

"'G)"

-^^

transition matrix is obviously

The response can be expressed


y(t)

=.

*(Mo)0

"

*(*.t)

directly in terms of (5.W).

+ J^\-l

3e-2t

2)

(^"'p'

^_3L,)(;)6(.-to)rfr

4e-3t

(5.e)

This integral
Instead,

we

is usually very complicated to solve analytically, although it is easy for a computer.


shall use the transfer function matrix of equation (5.56).

_2
s +

1_
3

s2

5s

This is indeed our original transfer function, and the integral


Laplace transform is
^s2

whose inverse Laplace transform gives

5.11.

t sing

We

y(t)

5s

_
~

(5.68)

is

a convolution and

+ 6/Vs

form

of control

law for use

desire to guide a vehicle to a final state x(ty), which is known.

^^'

as before.

the adjoint matrix, synthesize a

x(tf)

*{tf,t)K{t)

+ J

*{tf,T)B(r)u{T)dr

I
From

in guidance.
(S.H),

its

SOLUTIONS TO THE LINEAR STATE EQUATION

122

[CHAP.

Choose u(*) = U(t)c where V(t) is a prespecifled matrix of time functions that are easily mechThe vector c is constant, except that at intervals of time it is
anized, such as polynomials in t.
Then c can be computed as
recomputed as knowledge of x(t) becomes better.
c

rj

[*(t;,t)x(t)

^(tf,T)B{r)V{T)dr~]

y.(tj)]

dx/dt = A{i)x with x(f) as the


t) as the transition matrix of
going to x(j). Therefore *(*/, t) would have to be computed at each recomputaTo avoid this, the adjoint transition matrix
tion of c, starting with the best estimates of x(t).
(t, tf) can be found starting with the final time tf, and be stored and used for all recomputations
of c because, from equation {5M), *t(r, tf) - *(tj, r) and c is found from

However,

this involves finding *(</,

initial condition

t(r,

tf)

[t(t,

B(t) U(t) dr

tf)

x(f)

X(%)]

Supplementary Problems
5.12.

Prove equations

5.13.

Given *(k,m), how can

5.14.

Prove that AeAt

5.15.

Verify that

(5.9), (5.10), (5.11),

=:

*(t,T)

given in equations

5.16.

and

(S.12).

be found?

A(fe)

eA(f-T)

and *(fc,m)

and

such that eAgB

is

satisfy the properties of a transition matrix

A''-'"

system

Find the transition matrix

5.18.

Calculate the transition matrix *(f,

form,

/e-4(t-T)

e-^' -

the state equation corresponding to this fundamental matrix?

5.17.

(6)

to the

the Maclaurin series,

(c)

0)

dx.ldt

for

dx/df

/O
I

2\
'^

_3 j

^jx

using

reduction to Jordan

(a)

the resolvent matrix.

Find eAt by the series method, where


series method is the easiest.

A =

This shows a case where the

0/

yo

_/-3
Find eAt using the resolvent matrix and Leverrier's algorithm, where

A -

\
5.21.

eBeA.

Given the fundamental matrix

What

5 20

(5.5)-(5.12).

5.19.

gAtA and then find the conditions on

Find eAt using the Cayley-Hamilton method, where

is

0\

1.

-3/

the matrix given in Problem 5.20.

CHAI'.

5.22.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

Use the eigenvalue method

123

to find eAt for

5.23.

Use the resolvent matrix and Cramer's

5.24.

Use the resolvent matrix and Cramer's

5.25.

Find eA by using the Maclaurin

5.26.

Find the fundamental, or transition, matrix for the system

rule to find eAt for

A^ for

rule to find

series,

as given in Problem 5.22.

as given in Problem 5.22.

Cayley-Hamilton and resolvent matrix methods when

x/

\
using the matrix Laplace transform method.
5.27.

Given the continuous time system

Compute

y{t)

0)x

(1

4m

using the transition matrix if m is a unit step function.


Compare this with the
by finding the 1X1 transfer function matrix for the input to the output.

solution obtained

5.28.

Given the discrete time system


K(n

-^ -i\

+ 1) =

[^_l
y{n)

Compute

5.29.

(a)

(b)

5.30.

(1

y{n) using the transition matrix if

Calculate *(,

Calculate

*(fc,

<)

for the system

dx/dt

m) for the system

..

/n

,,

_l)^(n) +[^lju(n)

x(A;

0)x(n)

=
1)

x(0)

/o

4M(n)

the series of ones 1,1,1, ...,1

is

using Laplace transforms.

|^

using

x(fc)

2 transforms.

How does the spectral representation for eAt extend to the case where the eigenvalues of
not distinct?

are

n-l
5.31.

In the Cayley-Hamilton method of finding eAt, show that the equation

el*

=
i

be solved for the yi(t).

For

2
=

YiWAy can always

simplicity, consider only the case of distinct eigenvalues.

5.32.

Show

5.33.

Show

5.34.

Given the time-varying system

that the column vectors of *(t,

r)

span the vector space of solutions to dx/dt

A(t)x.

A(<) A(t) = A(t) A(t) when A(t) = a{t)C,


where C is a constant
matrix and a(t) is a
scalar function of t.
Also, find the conditions on ay(t) such that A(t) A(t) = A(t) A(0 for a
I X 2 A(i) matrix.

Find the transition matrix.

nXn

dx

dt

\-a-^t)

-'a{t) a-^(t) ^'

Hint: Find an integrating factor.

SOLUTIONS TO THE LINEAR STATE EQUATION

124

5.35.

Prove
P(fc

5.36.

Floquet's
<o,

m)

if

A(fc

dx/dt

/sin

5.37.

<(,)

is

The
|s(t)|

and verify

constant.

it

satisfies

Floquet's result

1.

m)

*(t, (q)

where P

P(*. <o)*"~*"'

is

periodic

and

Also find the fundamental matrix of the adjoint system.

shown in Fig. 5-6 is excited by a square wave


The system equation is y + [p + a sgn{am vt)]y = 0.

8{t)o-

P(fc,

Find the transition matrix

linear system

where

sin t\

x.
^
sm ,t smty

P(k,m)W<-~'"

+ w).

Given the time-varying periodic system


*(i,

*(k,m)

theorem for discrete time systems,

A(fc)

[CHAP.

<DH

,rN

i:_

with period 2 and amplitude

8(t)

"Y
+

OJ/(t)

82

Fig. 5-6

It is

found experimentally that the relationship between a and

solution can be plotted as

shown

yS

that permits a periodic

in Fig. 5-7.

Fig. 5-7

(Do not
Find the equation involving a and p so that these lines could be obtained analytically.
mark
and
and
all
a
for
solution
/3
form
of
the
general
attempt to solve the equations.) Also give
the regions of stability and instability on the diagram.

5.38.

of e{t)
Given the sampled data system of Fig. 5-8 where S is a sampler that transmits the value
Find the state space representation at the sampling instants of
once a second to the hold circuit.
Use Problem 5.10.
the closed loop system.

'<..-^-r>-^eM'.

Hold

s+
s2

5s

1
-h

-oy(t)
6

Fig. 5-8

5.39.

Find *(t,T) and the forced response for the system

t^v

tv

+v

-- p{t)

with

riito)

Vo

^^^

CHAP.

5.40.

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

125

Consider the system

dt\xjI -

dJ\X2
dyUy+W

\c
ye

'yi\

h\fxi

(s
,

\*

3/2/

or

= Ax + Bu

""^

= Cx + Du

'"^j\'^%J

Find the transfer functions ^{2/i}/^{mi} and Jliv^lJiW^ using the relation

CAv^Uin) = [C(Is-A)-iB +
5.41.

The steady state response


equation

x^sit)

of an asymptotically stable linear differential system satisfies the

dx,,/dt

but does not satisfy the

initial condition

Xss(o)

and has no reference

(a)

Verify, by substitution into the given equation, that

where

is

-^

Suppose A(t) = A(t + T),


expression for a periodic

an

K(7') is

Check that
u{t)

B()
K^^it)

h(t)

nXn

K{T)

t.

to-

et+<'(t-r) sin

dr

Hint:

For an arbitrary vector function

(e-r-

g-t)

and the system

*(t, t) B(t) U(r)

matrix to be found depending on

satisfies

T and

the system

is stable.

Find an

dr

independent of

t.

equation of Problem 5.8 with

S(t-T).

Fi nd in closed form the response of the system


ty/f^-l with zero initial conditions.

5.44.

Consider the scalar system


find the sign

1,

dx/dt

g/^

)x.

Show that if an nXn nonsingular matrix


known explicitly, where D(i) is an nXn
also

to the input

u(t)

10.0,

Find the relationship between

Xi(t)

and

x^it)

such

using the adjoint.

is

is

-3/(2 \

Xiitf)

{l-t^)^-\^ =

dyldt = -(1 + t)y + (1 + t)u. If the initial condition is 2/(0) =


and magnitude of the impulse in u{t) required at = 1.0 to make j/(2) = 1.0.

Given the system

A(()x

*(t, t) B(r) u(r)

= B(t + T), u(t) = u(t + T),


= x^si* + ^ in the form

5.43.

that

time

to the initial

= f(t,^(0)f - f(t..W)^+J^f(,.)c^.

Xjt)
where

the indefinite integral evaluated at t

J^Jf(t,.)..

5.46.

"o

Bit) u(t)

known.

f{t,r),

5.45.

t) is

X,,(t)

5.42.

A(t)xss

(*,

(6)

D]

known.

solution T(t) to the equation dT/dt = A{t)T TD(t)


diagonal matrix, then an explicit solution to dx/dt =

SOLUTIONS TO THE LINEAR STATE EQUATION

126

Answers
=

A(k)

5.14.

If

5.15.

deteAU-T)

5.16.

A =

5.17.

(f.r)

= ^^et-rf^\(l

and only

'':

5.19.

eAt

5.20.

71

Supplementary Problems

to

*{k + l,k)

5.13.

"

[CHAP.

AB - BA

if

detel"-'"'

eAeB

does

e<-triSh

eBgA.

e(t-T)trA

/-2 -2'

-2 -2

2)

^V-1

e^<-r-of

2)1

9,

72

26,

yg

24

o\

/O
eAt

o.5e-2t

zz

(1

+ e^t (

(6e-2t-8e-3t

eAt

5.23.

(sI-A)-i

+ 3e-)I+

F(s)/s(s-l)2

0.5(7e-2t-12e-3t

D/s

B/(s

Vi

^
5.24.

A"

5.25.

eAt

-A;

-^

~^

-1

^\
j

0/

\
+ 5e-4t)A+ 0.5(e-2t-2e-3t+e-)A2

C/(s

- 1)2

where

-i/

-fc

- 1) +

/
0.5e-

1/

\0
5.21.

2fe

=
1

e2t

e2t

1
5.27.

y{t)

|[ll-2e-(t-to)/2_e--(t-t)][7(i_y

5.28.

y(k)

5.29.

*(t,to)

[7

=
\

5.30.

- Z{-V)^ - 4(-l/2)'']/12
,
gt-to

Let the generalized eigenvector

y
tj

eAt
5.31.

5.33.

Requires only

Vandermonde matrix

*(fc,m)

),

of

e'^txjrl

ai^ir) a^iit)

have a reciprocal basis vector

e'^tt^sf

in the eigenvalues results,

Oi2() 021(7)

(^
V

and

[a22(t)

te'^tx^sf

which

is

- aii(t)]

Sj,

etc.

Then

then always invertible.


a2i(T)

a2i(t) [022(7)

- aii(r)].

CHAP,

SOLUTIONS TO THE LINEAR STATE EQUATION

5]

a-HT)dT
5.34.

where
*(to,

5.37.

/cosh T

*(t,to)

e^i

cos

*o

a-^T)dT

sinh

\sinh T

cos

a-Hr) dr

t'*

cosh t

cos

so

t,

*(t,t^)

R =

and

V(t,to)

0.

Also t(t,

-!(,

to)

t).

Let

y2

j8

+ a,

S2

*(2,1)

- a.

/3

Then

\S sm

cos S

8
z{t

X^

(!,

*=^^

2 cos y cos 8

X2

det *(1, 0)

1)

where

0)

(y/8

^'^^^^A

\~y s T

eigenvalues \ of

x(t),

stability boundaries are then determined

1)

(1.0)

X2 cos

det *(2,
2 cos 9

(2,

+ 4,r/e) =

- e2R) =

det (X/

= (2, 0) =

e^R

^''"'^^'^

'"'.'

For periodicity of the envelope

The

a-'Wdr

-a:~l(t) sin

5.36.

am

aito)

'

*(t,<o)

127

e^R

X tr e2R

''"s

e*^^.

det e^R

8/7) sin y sin 8

by

2 cos y cos S

(y/S

8/y) sin y sin 8

solution is of the form (<, r) = P(t, T)eR(-r) and the given curves form the
between unstable regions and periodic regions.

The

stability boundaries

Reference: B. Van Der Pol and M. J. 0. Strutt, On the Stability of the Solutions of Mathieu's
Equation, Philosophical Magazine, 7th series, vol. V, January-June 1928,
pp. 18-38.

5.38.

x(fc

^(l

+ l) =

5.39.

i;(t)

5.41.

K(r)

5.42.

Since h{t)

1/0

'-'-'

+ ^"')/2

cos (In

t/fo)

-I-

('-^'~yM

]^(k)+
^

(5e-3-2)/3y^

\{\-e-^)l^

'

^0*0 sin (In /to)

\^2-2e-3y

sin (In r/t) p(t)/t dr

- (e-RT_|)-i
is

an element of the transition matrix,


d^hldf^

Also, since dh/dt

-V

and h are continuous


lim

5.43.

y{t)

5.44.

u(t)

5.45.

5.46.

Let X

(1/4

{\-2)dh/dt

^/W^n. )(t -

(sin-i

+ e-^t)h =

(a"^- a

sin-i

to)/2

= ^^g-s/V^ Mt-l-O)

(3/t

+ T3)a;i(to) +

T()z.

3(1

- T*)x2{to)/tf

f or

t,

dWdt^dt

to)

in

where r

tf/to

S{t-T)dt

chapter 6

Controllability
6.1

and Observability

INTRODUCTION TO CONTROLLABILITY AND OBSERVABILITY

Can all the states of a system be controlled and/or observed? This fundamental question
arises surprisingly often in both practical and theoretical investigations and is most easily
investigated using state space techniques.
Definition 6.1:

state xi of a system is controllable if all initial conditions xo at any


to can be transferred to xi in a finite time by some control

previous time

function u(<,xo).
If all states xi are controllable, the

controllable.

system is called completely controllable or simply


depend on to, the state is said to be control-

If controllability is restricted to

If the state can be transferred from xo to xi as quickly as desired indeof in some finite time, that state is totally controllable. The system
instead
pendent of to,
Finally, we may talk about the
is totally controllable if all states are totally controllable.
for output controllable, e.g. an
definitions
similar
give
and
output y instead of the state x
output
particular
y can be attained starting from
output controllable at time to means that a

lable at time

to.

any arbitrary xo at

to.

for linear systems, it is necessary and


transsuflacient to investigate whether the zero state instead of all initial states can be
ferred to all final states. Writing the complete solution for the linear case,

To determine complete

X(<i)

controllability at time

to

+ f

*(ii,*o)x(o)

*(il, r) B(t) u(t)

dr

equivalent to starting from the zero state and going to a final state
Therefore if we can show the linear system can go from
x(ii)-*(ii,to)x(o)-

which

(ii),

is

then

it

can go from any

The concept of
Definition 6.2:

x(<o) to

any

x (ii) =
to

any

x(fi).

observability will turn out to be the dual of controllability.


state x(i) at

some given

input u(t) and output yir)


pletely determines x(t).

of a system

over a

is

finite

observable if knowledge of the


time segment to<T^t com-

system is called completely observable. If observabe observable at to. If the state can be determmed
independent of to, it is totally observable.
segment
for T in any arbitrarily small time
u(t) = 0, and give similar definitions for
when
Finally, we may talk about observability
If all states x{t) are observable, the
bility depends on to, the state is said to

zero-input observable.
to
observability for linear systems, it is necessary and sufficient
from
determined
completely
be
can
system
see if the initial state x(o) of the zero-input
permits x(*) to be calculated from the complete
y(r), because knowledge of x(o) and u(r)

To determine complete

solution equation

(5.-^^).

128

CHAP.

CONTROLLABILITY AND OBSERVABILITY

6]

129

We

have already encountered uncontrollable and unobservable states in Example 1.7,


These states were physically disconnected from the input or the output. By
physically disconnected we mean that for all time the flow diagram shows no connection,
i.e. tlie control passes through a scaler with zero gain.
Then it follows that any state
vecto:i-s having elements that are disconnected from the input or output will be uncontrollable or unobservable.
However, there exist uncontrollable and unobservable systems in
whicl: the flow diagram is not always disconnected.

page

3.

Exami:le

6.1.

Ccinsider the time-varying system

d^/a;i\
dt

From

the flow

diagram of Fig.

\x2j
6-1 it

_
~

gat

/
\0

+
13

can be seen that

u(t)

get

passes through scalers that are never zero.

uo

Fig. 6-1

Fop zero

initial conditions,

u(t)

Only those x^it^) - Xi(ti)e*''^^ "' can be reached at


fore thu system is not controllable.

6.2

t^,

cLt

X2(ti)e~eh

so that Xzih)

is fixed

after

a;i(ti) is

chosen.

There-

CONTROLLABILITY IN TIME-INVARIANT LINEAR SYSTEMS

For time-invariant systems dx/dt = Ax + Bu in the case where A has distinct eigenvalues, connectedness between the input and all elements of the state
vector becomes
equivalent to the strongest form of controllability, totally controllable. We shall first consider the case of a scalar input to avoid complexity of notation, and consider distinct eigenvalues before going on to the general case.

Theorem

6.1:

Given a scalar input u{t) to the time-invariant system dK/dt = Ax + hu,


where A has distinct eigenvalues Ai. Then the system is totally controllable
if and only if the vector = M-'b has no zero elements.
is the modal
matrix with eigenvectors of A as its column vectors.

Only if part: Change variables by x = Mz.


Then the system becomes
+ iu, where A is the diagonal matrix of distinct eigenvalues Ai. A flow diagram
system is shown in Fig. 6-2 below.

Proof:
dx/dt =: Az
of this

If any element ft of the f vector is zero, the element of the state vector Zi is disconnected
from the control. Consequently any x made up of a linear combination of the z's involving
Zi

will be uncontrollable.

/i

must be nonzero.

Therefore

if

the system

is totally

controllable, then all elements

CONTROLLABILITY AND OBSERVABILITY

130

[CHAP.

p-GWc^ 0J
Af'.

0j
0J

+1

^__/rvJ

Fig. 6-2

Now we

are nonzero, and from the remarks following Definition


6.1 we need investigate only whether the transformed system can be transferred to an
arbitrary z{U) from z(io) - 0, where ^i can be arbitrarily close to ta. To do this, note
If part:

assume

all /i

but yet unproven, that

It is true,

to

if

the

/i

are nonzero,

1,2,

many

.,n

{6.1)

different m(t) can transfer

z(i).

Now we

construct a particular u{t) that will always do the job.

u{t)

where the

ix^.

are constants to be chosen.

Prescribe u{t) as

i: M^e-^^^-'i'

{6.2)

Substituting the construction

{6.2) into

equation

{6.1) gives

where the inner product

is

A)

/ii"fc(e^'"'~'^ e^'"'""')

foralH

(6.5)

defined as

{B{T),<f>{r))

r'e*(r)</,(T)dr
/to

Equation

{6.3)

can be written in matrix notation as


'Zi{ti)/fi

gii

912

Z2{tl)/f2

fl'21

6^22

{6.A)

Zn{tl)/fn

ffln

Qzn

by assumption, division by fi is
/i ^
linearly independent, the Gram
obviously
are
permitted. Since the time functions
can always solve equation {6.Jf)
we
Hence
Problem
3.14).
matrix {go,} is nonsingular (see
work.
always
will
control
{6.2)
for jui, 112, ..., i^n, which means the

where

gik

(e'^""'-^', e^ii-^>).

Note that because


e'^i"!-^)

Now we can consider what happens when A is not restricted to have

distinct eigenvalues.

CHAP.

CONTROLLABILITY AND OBSERVABILITY

6]

Theorem

6.2:

Given a scalar input


vi^here
(1)

(2)

131

the time-invariant system dx/di = Ax + bM,


Then the system is totally controllable if and only if:

u{t) to

A is arbitrary.

each eigenvalue Ai associated with a Jordan block Lji(Xi) is distinct from


an eigenvalue associated w^ith another Jordan block, and
each element /i of f = T-^b, where T->AT
bottom row of each Jordan block is nonzero.

Note that Theorem

6.1 is a special case of

Theorem

J,

associated with the

6.2.

Only if part: The system is assumed controllable. The flow diagram for one
Proof:
X I Jordan block Lji(Ai) of the transformed system dz/dt = Jz + fw is shown in Fig. 6-3.
The control u is connected to Zi, 22,
.,Zi-\ and zi only if /i is nonzero.
It does not matter
if fufi,
and /i-i are zero or not, so that the controllable system requires condition (2)
to hold.
Furthermore suppose condition (1) did not hold. Then the bottom rows of two
different Jordan blocks with the same eigenvalues [L^i(Ai) and Li(Ai)] could be written as
I

+T

^^^

dzv/dt

XiZv

dZr,ldt

XiZr,

+ fr,U

fvU

+T

+T

<^

<Hv
<7>
Fig. 6-3

Consider the particular state having one element equal to fr^Zv-fvZ^.


d{fr, Zv

fvZn)ldt
^

Therej^ore

fr,z, (t)

- f^Zr, (t) =

(1)

must

// part:

To

+ fvU)

Xi\Jy)Zv

fy (Ai2;

+ f-r^U)

JvZn)

- f.Zr, (0)]e^'* and

is

if

independent of the control. We have


the system is controllable, condition

hold.

the sy;3tem
Example

is

{XiZv

not controllable, so

[fr,z, (0)

found a particular state that

fr,

Then

Again, a control can be constructed in similar manner to equation


is totally

controllable if conditions (1)

and

(2)

of

Theorem

(6.2) to

show

6.2 hold.

6.2.

illustrate

why

condition

(1)

of

Theorem

6.2 is

important, consider the system

d_

dt

X2

2/

(1

-3)
X2

= (3^{m} + a;io)/(s - 2) and j:{x2} = Uiu} + X2o)/is - 2) so that y(t) = {x^a ~ Sx2o)e^t rethe action of the control.
The input is physically connected to the state and the state is
physicaJy connected to the output, but the output cannot be controlled.
Then

..zix^}

gardlesiii of

For discrete-time systems, an analogous theorem

holds.

CONTROLLABILITY AND OBSERVABILITY

132

Theorem

6.3 :

[CHAP.

Given a scalar input u{m) to the time-invariant system x(m + 1) = Ax + hu{m),


A is arbitrary. Then the system is completely controllable if and
only if conditions (1) and (2) of Theorem 6.2 hold.

where

Proof:
Only if part: This is analogous to the only if part of Theorem 6.2, in that the
flow diagram shows the control is disconnected from at least one element of the state vector
if condition (2) does not hold, and a particular state vector with an element equal to
/tj^i'

^^t,

is

If part:

assume

uncontrollable

if

condition

(1)

does not hold.

Consider the transformed systems z(m

distinct roots so that

Zi{m)

A.

[A'"->m(0)

For an nth order system, the desired

+ \'"-^u{l) +

Jz(m)

+ iu{m), and

for simplicity

initial condition,

+ u{m - 1)]/;

state can be reached on the nth step because


^-2

zi{n)lfi

Z2{n)lf2

+ 1) =

Then for zero

xr'

u{0)
U{1)

\r'

{6.5)

Zn{n)/fn

u{n

1)

Note that a Vandermonde matrix with distinct elements results, and so it is nonsingular.
Therefore we can solve (6.5) for a control sequence u(0),u{l),
.,u{n l) to bring the system
to a desired state in n steps if the conditions of the theorem hold.
.

For discrete-time systems with a scalar input, it takes at least n steps to transfer to an
arbitrary desired state. The corresponding control can be found from equation {6.5), called
dead beat control. Since it takes steps, only complete controllability was stated in the
theorem. We could (but will not) change the definition of total controllability to say that
in the case of discrete-time systems, transfer in n steps is total control.
of hidden oscillations in sampled data systems deserves some mention
Given a periodic function, such as sin mt, if we sample it at a multiple of its period
Referring to Fig. 6-4, it is impossible to tell from the sample points
it will be undetectable.
whether the dashed straight line or the sine wave is being sampled. This has nothing to
do with controllability or observability, because it represents a failure of the abstract object

The phenomenon

here.

In this case, a differential-differinstants.


sampling
behavior
between
ence equation can be used to represent
(the difference equation) to represent a physical object.

Fig. 6-4

6.3

OBSERVABILITY IN TIME-INVARIANT LINEAR SYSTEMS


Analogous to Theorem

connectedness between state and output becomes equivalent


to total observability for d^ldt = Ax + Bu, y = c+x + d'^u, when the system is stationary
and A has distinct eigenvalues. To avoid complexity, first we consider scalar outputs and
distinct eigenvalues.

6.1,

CHAP.

CONTROLLABILITY AND OBSERVABILITY

()

Theorem

133

Given a scalar output y{t) to the time-invariant system dx/dt = Ax + Bu,


y = c+x + d'^u, where A has distinct eigenvalues Ai. Then the system is
totally observable if and only if the vector gt = c+M has no zero elements.
is the modal matrix with eigenvectors of A as its column vectors.

6.4:

From
we need

the remarks following Definito see if x(fo) can be recon-

Proof:
tion 6.2

structed from measurement of y(T) over to<T t


in the case where u(t) = 0. We do this by changing variables as x = Mz. Then the system becomes dz/dt Az and y c+Mz = gtz. The flow

diagram for

system

this

given in Fig.

is

6.5.

Each Zi{t) = 2i(io)e'''<'"'o' can be determined


by taking in measurements of y{t) at times t^ =
for fc = 1, 2,
w and solving
to + (ti - to)k/n
.

the set of equations

*'V

=l

for giZi{U).
When written in matrix form this
set of equations gives a Vandermonde matrix

which

always nonsingular if the Ai are distinct.


then all Zi(fo) can be found. To find
i7i
x(t), use
x(^o) = Mz(to)
and dx/dt = Ax + Bu.
Only If gi'0 is each state connected to y.
If all

is

- 0,

The extension

Theorem

to a general

1/

if

(2)

is

ctx

is

similar to the controllability

Theorem

if:

Ai associated with a Jordan block Lji(Ai) is


an eigenvalue associated with another Jordan block, and

each eigenvalue

each element gt of g+ = c+T,


top row of each Jordan block

similar to that of

Ax{m)+Bu{m),
system

is

Theorem
is

Now we can

6.2.

from the time-invariant system dx/dt = Ax + Bu,


arbitrary. Then the system is totally observable

Theorem

where T-^AT
is

J,

distinct

|/(m)

from

associated with the

nonzero.

6.2.

Given a scalar output y{m) from the time-invariant system

6.6:

The proof

is

y{t)

+ d'^u, where A

and only

(1)

Theorem

matrix

Given a scalar output

6.5:

The proof

Fig. 6-5

ctx(m)

+ d'^u(m), where A

completely observable

if

and only

if

is

arbitrary.

conditions

(1)

x(m + l) =
Then the
and

(2)

of

6.5 hold.

similar to that of

Theorem

6.3.

classify the elements of the state vectors of dx/dt

= Ax + Bu,

= Cx

-I-

Du

and of x(m + 1) = Ax(m) + Bu(to), y = Cx + Du according to whether they are controllable


or not, and whether they are observable or not. In particular, in the single input-single
output case when A has distinct eigenvalues, those elements of the state zi that have nonzero /i and gi are both controllable and observable, those elements Zj that have zero fj but
nonzero gs are uncontrollable but observable, etc. When A has repeated eigenvalues, a
glance at Fig. 6-3 shows that Zk is controllable if and only if not all of fk,fk+i,
-yfi are
zero, and the eigenvalues associated with individual Jordan blocks are distinct.
.

Unobservable and uncontrollable elements of a state vector cancel out of the transfer
function of a single input-single output system.

CONTROLLABILITY AND OBSERVABILITY

134

Theorem

6.7:

For the

[CHAP.

dKldt = Ax + hu, y = t^x. + du,


has poles that are canceled by zeros if
some states are uncontrollable and/or unobservable. A similar

single input-single output system

the transfer function

c'''(sl

A)""*b

and only if
statement holds for discrete-time systems.

Proof:
First, note that the Jordan flow diagram (see Section 2.4) to represent the
transfer function cannot be drawn with repeated eigenvalues associated with different
Jordan blocks. (Try it! The elements belonging to a particular eigenvalue must be combined.)
Furthermore, if the J matrix has repeated eigenvalues associated with different
Jordan blocks, an immediate cancellation occurs. This can be shown by considering the
bottom rows of two Jordan blocks with identical eigenvalues X.

0\fzi
A,]U2

dt \Z2

(ci

+ du

C2)

- \)-^bi^{u} and Jl{z2} = {s-\)-^b2Jl{u},


= [ci{s-\)-^bi + C2{s-k)~^b2 + d]^{u}

Then for zero initial conditions


JliV]

4-

<^{zi}

(s

so that

Combining terms gives

This is a first-order transfer function representing a second-order system, so a cancellation


has occurred.
Starting from the transfer function gives a system representation of

= \z + u, y = (ci&i + 0262)2 + du, which illustrates


not be drawn for systems with repeated eigenvalues.

why

dz/dt

Now
Figs. 6-2

(2) of Theorems
and 6-5 to represent the element

consider condition

6.2
Zi

and

the Jordan flow diagram can-

Combining the flow diagrams of


row of a Jordan block gives

6.5.

of the bottom

Fig. 6-6.

Fig. 6-6

Comparing this figure with Fig. 2-12 shows that f^g^


p.-O, a caacellation occurs, and p, =
if and only if

if

the system

is

is

p.,

f.

and/or

A,.
If and only
which occurs when

the residue of
g^

0,

uncontrollable and/or unobservable.

Note that it is the uncontrollable and/or unobservable element of the state vector that
canceled from the transfer function.

6.4

DIRECT CRITERIA FROM

A, B,

AND

If we need not determine which elements of the state vector are controllable and observable, but merely need to investigate the controllability and/or observability of the whole
system, calculation of the Jordan form is not necessary. Criteria using only A, B and C

are available and provide an easy and general


and observable.

controllable

way

to determine if the

system

is

completely

CHAI.

CONTROLLABILITY AND OBSERVABILITY

6]

Theorem

The time-invariant system dx/dt = Ax + Bu is


only if the n x nm matrix Q has rank n, where

6.8:

Q =
Note that

(B|AB|

this criterion is valid for vector u,

...

and so

and

totally controllable if

|A-iB)

is

One method of determining if rank Q = w is to see


rank QQ^ from property 16, page 87. However, there

135

more general than Theorem

6.2.

=
det (QQ^) - 0, since rank
exist faster machine methods for

if

determining the rank.

To reach an arbitrary

Proof:

from the zero

x(^i)

we must

initial state,

find a control

such that

u(t)

x(*i)

r'eA<'i-^'Bu(T)dr

{6.6)

Use

of

Theorem

gActi-r)

4.16 gives

^ y.(T)A"~'

so that substitution into (6.6) gives

i=l

x(i)

+ ABw2+

(Bwi

+A"->Bw)

Wl
;

\ Wn

where

Wj,

y+i_s.(T)u(T)dT.

Hence

x(ii) lies in

the range space of Q, so that

must

to

have rank n
trollable,

Kbw we
time

we

to reach an arbitrary vector in


has rank n.

assume

Therefore

if

the system

has rank n and show that the system

is

totally controllable.

/t^

is totally

con-

This

construct u(t) as

U(r)

whers

"U^.

Mi8(r

-g+

m,S("(t

-^ +

M,S<"-"(r

- t,)

{6.7)

are constant m-vectors to be found and 8"''(t) is the fcth derivative of the Dirac
Substituting this construction into equation {6.6) gives

delta function.

x(tj)

Jtti

eActi-T)B8/

eA(ti-t)B^^

_ f\

^^

eACti-t)AB/i2

eA(ti-to)B

eA"i-'o>A"-iB^

and the defining relation for

B"'\t-^)g{i)di

Usinj? the inversion property of transition matrices

{6.8)

8^"^ is

d^g/dt^

and the

definition of Q, equation {6.8)

can be rewritten

eActo-ti)x(fi)

(^

q(

3.11, page 63, a solution for the |t^ always exists if rankQ = n. Hence some
(perhaps not unique) always exists such that the control {6.7) will drive the system to

From Problem
ll^

X(ii).

The construction for the

control {6.7) gives

some insight as

to

why

completely control-

systems can be transferred to any desired state as quickly as possible.


No restrictions are put on the magnitude or shape of u(t). If the magnitude of the control
is bounded, the set of states to which the system can be transferred by U are called the
reachable states at ti, which has the dual concept in observability as recoverable states at
ti.
iVny further discussion of this point is beyond the scope of this text.
lable stationary linear

CONTROLLABILITY AND OBSERVABILITY

136

[CHAP.

proof can be given involving a construction for a bounded u{t) similar to equation
instead
of the unbounded u{t) of (6.7).
(6.2),
However, as ti - U, any control must become
unbounded to introduce a jump from x(io) to x(ii).

The dual theorem

Theorem

6.9:

to the one just

proven

is

The time-invariant system dx./dt=^ Ax + Bu, y = Cx + Du


servable if and only if the kn x n matrix P has rank n where

is

totally ob-

CA
CA"
Theorems 6.8 and 6.9 are also true (replacing totally by completely) for the discrete-time
x(m + 1) = Ax(m) + Bu(m), y(TO) = Cx(m) + Du(m). Since the proof of Theorem
is quite similar to that of Theorem 6.8 in the continuous-time case, we give a proof for

system
6.9

the discrete-time case.

from knowledge

the initial state

It is sufficient to see if

of y(m) for

l^m<

in the case

>,

where

u(to)

x(i)

can be reconstructed
the state equa-

From

0.

tion,

y(0

Cx(J)

yil

+ 1) =

y{l

+ n-l) =

Cx(i

+ 1) = CAx(0

CA"-ix(i)

This can be rewritten as


y{i)

Px(i)
y{l

and a unique solution

6.5

exists

if

+ n-l)

and only

if

P has rank

shown

n, as

in

Problem

3.11,

page

63.

CONTROLLABILITY AND OBSERVABILITY OF TIME-VARYING SYSTEMS

In time-varying systems, the difference between totally and completely controllable becomes important.
Example

6.3.

Consider the time-varying scalar system


or unity as shown in Fig. 6-7.

<)

dx/dt

-X

4- h{i)u

and

j/

a;,

where

6(f)

is either

zero

rO

e
Fig. 6-7

If

{(*)

for to -* < *o + '^*. the system is uncontrollable in the time


+ ^*-*-*i' *he system is totally controllable in the time

interval

[to, to

+ At).

If

interval to + A*-t-*iover the time interval


controllable
completely
but
is
controllable,
totally
not
is
system
However, the
to * *i to reach a desired final state x{t-^.
6(i)

for

to

CHAP.

CONTROLLABILITY AND OBSERVABILITY

6]

137

Now suppose b{t) = for t^ < t ^, and we wish to reach a final state x(t^. The state xit^i can
be reached by controlling the system such that the state at time t^ is x{t^ = e(*2~*i'a;(t2). Then with zero
input the free system will coast to the desired state x(t<^ = <j>(t2. ti) (ti). Therefore if the system is totally
controllable for any time interval f,. ~ * ~ *e + ^*> it is completely controllable for all t t^.
For the time-varying system, a

Theorem

Theorem

criterion analogous to

6.8 can be formed.

The time-varying system dx/dt = A{t)x + B{t)u is totally controllable if


and only if the matrix Q(t) has rank n for times t everywhere dense in
Q(i) = (Qi Q2
[to, ti], where
in which Qi = B(i) and Q^+i =
Q),

6.10:

A(i)Qfc + dQfc/di

for k

1,2,

piecewise differentiable at least %

Here A{t) and

.,n-l.

2 and n 1

B{t) are

assumed

times, respectively.

The phrase "for times t everywhere dense in [to,ti]" essentially means that there can exist
only isolated points in t in which rank Q < w. Because this concept occurs frequently, we
shall abbreviate it to "Q(f) has rank w(e.d.)".
Proof:

and show

= w in an interval containing time


First we assume rank
the system is totally controllable.

Construct

u(t) as

iy

such that

to

< < *i
1?

uW =

Sm^S'-'-^V-^)

(6.9)

lc=l

To

from

attain an arbitrary x(fi) starting

L{tr)

x(to)

we must have

t jS Wl,

r*{tur)B{r)u{r)dr

B(t)]

r)

T = 7)

M,

(6.10)

But

|:[*(t,r)B(x)]

Note
{6.11)

dl/dt

d{9^-^)/dt

= A**"i +*d*-Vdt

so that d^'^ldt

(6.11)

= -*-A and

equation

becomes

Similarly,

Therefore

(fi.iO)

[*(!, T)B(r)]

|:

[*(,, r)B(r)]

*(ti,r)Q2(r)

*(tl,r)Q.(T)

(^.1^)

becomes

x(to

Ht^.r)^ + [^*-Hr, .)] B(r)

/*(ti,v)Q('?)

solution always exists for the nk because

rank Q = w and *

is

nonsingular.

Now
6.8,

assume the system is totally controllable and show rank Q = . From Problem
page 142, there exists no constant w-vector z 7^
such that, for times t everywhere

dense in

U~t~ U,
zt*(to,QB(i)

By

differentiating h times with respect to

Since
zt*(io,<) v^O such that

z+*(io,

t) Q)c(i)

0.

is

*(io, t)

(ytQi y+Q2

where qi are the


the rank of Q(i)

row vectors
is n(e.d.).

of Q.

always

y+Qn)

Since the

and using equation (6.12), this becomes


nonsingular, there is no n-vector yt =

y^Q.

n row

2/iqi

2/2q2

yqn

vectors are then linearly independent,

CONTROLLABILITY AND OBSERVABILITY

138

Theorem

6.11:

[CHAP.

The time-varying system dx/dt = A{t)x + B{t)u, y = C(i)x + D(i)u is totally


observable if and only if the matrix P(^) has rank ^(e.d.), where F'^{t) ... \K)
in which
(PJ" PJ
Pi = C(*) and P^ + = PfcA(f) + dP^/dt for
fc = 1,2,
.,n-l. Again A(i) and B(i) are assumed piecewise differentiable
at least n 2 and n 1 times, respectively.
I

Again, the proof

is somewhat similar to Theorem 6.10 and will not be given.


The situsomewhat different for the discrete-time case, because generalizing the proof following Theorem 6.9 leads to the criterion rankP = n, where for this case

ation

is

Pi

C{1),

Pa

C{l

l)Ail),

...,

Pfc

C{l

+ k-l)A{l + k-2)-

-Ail)

The situation changes somewhat when only complete controllability is required. Since
any system that is totally controllable must be completely controllable, if rank Q(i) has
rank n for some t > to [not rank (e.d)] then x(io) can be transferred to x(^i) for ti ^ t.
On the other hand, systems for which rank Q(i) < n for all t might be completely controllable (but not totally controllable).

Example

6.4.

Given the system

\x2j

dt

-(:^)(::)^

\h{t))

where
f

sin

{2k

and f2{t)=fi{t

+ T).

t<

2k7r^

l)v

(2k

0,

for all

t.

1,2,

...

Then
\f2it)

At each

+ l),r
< 2(fc + l)ff

instant of time, one

row

of Q(t)

is zero,

^ea(t-t)

*(t,t)B(t)

so

Mid) +
rank Q(t)

df^ldt

\/fi(t)\
ee<t-H^J\f,(t)J

However,

//i()e"-to>N

V/2(*)e'-'o'

If tto>TT, the rows of *(o, *) B(t) are linearly independent time functions, and from Problem 6.30,
page 145, the system is completely controllable for tj to > tt. The system is not totally controllable because

for every

tg.

if

^2

*o

<

"">

either /i(t) or fiir) is zero for

tf,

i<2,.

B(i), it can be shown that complete conTherefore rank Q = n is necessary and sufficient
for complete controllability also. (Note /i(f) and fiif) are not analytic in Example 6.4.)
For complete controllability in a nonanalytic system with rankQ(i)<n, the rank of
*(i, t) B(t) must be found.

However, for systems with analytic A(i) and

trollability implies total controllability.

6.6

DUALITY

In this chapter we have repeatedly used the same kind of proof for observability as was
Kalman first remarked on this duality between observing a
used for controllability.
matrix of Problem
dynamic system and controlling it. He notes the determinant of the
of the optimal
The
dual
Information.
of
Shannon's
definition
analogous
page
is
to
142,
6.8,
manifested
is
by the
This
duality
Kalman
filter.
Chapter
the
of
10
is
control problem
systems:
following two

CHAP.

CONTROLLABILITY AND OBSERVABILITY

6]

System #1:

System #2:

dx/di

A(i)x

B{t)u

C(f)x

D(i)u

dw/dt
z

= -At(t)w +
=

Bt(i)w

139

Ct(t)v

Dt(f)v

Then system #1 is totally controllable (observable) if and only if system #2 is totally


observable (controllable), which can be shovi^n immediately from Theorems 6.10 and 6.11.

Solved Problems
6.1.

Given the system


1

dx

-1\

/O
X +

1-43/

|m,

2/

(1

-1

l)x

\1

and uncontrollable states and then

find the controllable

find the observable

and un-

observable states.
Following the standard procedure for transforming a matrix to Jordan form gives
1

2-1

0\

-1

10
.1-4
Then
shown

= T-b=

(0 1

3/

0)t

and

g^

c^T

(0 1

)(

2/ V
1).

0\/-l
jf

1/ V
The flow diagram

A=

TJT~i

0"

-2

10,
of the Jordan system is

in Fig. 6-8.

G>

s-2

<7>

Z2

o-

KiW

s-2

<7>

^3

s-

-OV

-0-

Fig. 6-8

The element
Z2 is

Xj of the state vector z is controllable (through Z2) but unobservable.


The element
both controllable and observable.
The elem.ent 23 is uncontrollable but observable.

Note Theorem

6.1 is inapplicable.

CONTROLLABILITY AND OBSERVABILITY

140

6.2.

Find the elements of the state vector


the system of Problem 6.1.

[CHAP.

z that are both controllable

and observable for

Taking the Laplace transform of the system with zero initial conditions gives the transfer
Using the same symbols for original and transformed variables, we have

function.

sxi

Xi

SX2

X2

sxg

y
(6.15), Xi = 4^2 + (s
Substituting this in {6.H),

From

3)3:3

~"

l)(s

(s

(6.1S)

(6.H)

Xi

4x2

Xi

a;2

Putting this in

- 2)2x3 =

(s-2)3

Thus the transfer function

h(s)

(s

trollable element of the state vector

Given the system of Problem


Forming the P and

x^

(s

3^3

(6.16)

(6.13) gives

- 1)2m.

2)-i,

Is

6.1.

!)* (s 2)2*3.

(s-l)(s-2)^

and from Theorem

it

6)^2 =
(6.16),

(s-l)(s-2)2

as defined in Problem

is 2

(4s

Then from

(s-2)2j"'

'

(6.15)

**

^3

(i^i)!_o+^l

6.3.

2x2

6.7 the only observable

and con-

6.1.

totally observable

and

totally controllable?

matrices of Theorems 6.9 and 6.8 gives

'1-1

-1

-4

Q =
1

rank P = 2, the dimension of the observable state space; and rank Q = 2, the dimension
of the controllable state space.
Hence the system is neither controllable nor observable.

Then

6.4.

Given the time-invariant system

and that u{t) = e"* and


What happens w^hen a ==

y{t)

= 2 ate~K Find

Xi{t)

and

X2{t).

Find

xi(0)

and

X2{0).

= x^, then Xi(t) = 2 ate-*. Also, dxi/dt ax^, so differentiating the output gives
= ~6~' + *"""' Then x^(<i) = 2 and a;2(0) = 1. When a = 0, this procedure does not work
because dxjdt = 0. There is no way to find X2(t), because X2 is unobservable as can be verified from
Theorem 6.5. (For a = 0, the system is in Jordan form.)
Since y

2(*)

6.5.

The normalized equations

of vertical motion y{r,

struck by a force u{t) at a point r

Y^2
vsrhere

y{ri, e,t)

modes of vibration
The

at r
of the

solution for the

V'2/

u,

60

6, t)

for a circular

drumhead being

are

27rrS{r-ro)S{9-eo)u{t)

Can

n, the edge of the drum.

{6.17)

this force excite all the

drum?

mode shapes
00

y(r, e,t)

is

00

2 2

"^n ("mr/ri) [Xin.m (*)

COS

2nj-fl

a;2 +

1,

(t)

sin 2nve]

CHAP.

CONTROLLABILITY AND OBSERVABILITY

6]

where k^ is the mth zero of the wth order Bessel function


harmonic m = 1, n = 1 into equation {6.17) gives

e7(r).

141

Substituting the motion of the

(Px2i/dt^

Xa;2i

y cos 2jr9oW

cPxgi/dt^

X^si

y sin

first

2ire(,u

^ = k^ "*" (^'^)^ ^^^ 7 *'o''^i(i''o/*"i)Using the controllability criteria, it can be found


that one particular state that is not influenced by u(t) is the first harmonic rotated so that its node
line is at angle Sq.
This is illustrated in Fig. 6-9. A noncolinear point of application of another
force is needed to excite, or damp out, this particular uncontrollable mode.

where

Pig. 6-9

6.6.

Consider the system

dx

110
=

[0

0'

0Jx + [1)mi +

hu2.

where

Ui

is totally

and U2 are two separate scalar


controllable

For each

case,

we

/I

it is

(0

0)^

(2)

(0

1)''

(3)

(1

0)^

Q=

0\

B =

and

1/

A2B)

for

\1
(1)

of

Theorem

6.2 the

system

is

(2),

1
1

1
1

/O

1
1

The first three columns are linearly independent, so


For case (3),

has rank 3 and the system


1

l'

0,

Q=(l01010
Vl
1

The bottom two rows are

AB

ll

equivalent to scalar control, and by condition

For case

(B

/O

(0101
\0

For b

(1)

investigate the controllability matrix

A =

uncontrollable.

Determine whether the system

controls.

if

identical

and so the system

is

uncontrollable.

is

controllable.

142

CONTROLLABILITY AND OBSERVABILITY

6.7.

Investigate total controllability of the time-varying system


ft

0'

VI

o;^ + 12'^

dx

dt

The

Q(t)

matrix of Theorem 6.10

6.8.

e(e

+ 2t- 2).

Since

eHl-t)

+ 2i =

only at one instant of time,

that the time-varying system dx/dt = A(i)x 4- B{t)u


and only if the matrix W(f t) is postive definite for every

Show

Note that

Q(t)

2(e.d.).

controllable if

and every

t>

t,

where

*(r,,)B(,7)Bt(^)*t(,,,)d,

this criterion depends on


Also note positive definite
of *(t, rj) B{r]) for T ~ t.

and is not as useful as Theorems 6.10 and


equivalent to linear independence of the rows

9{t, t)

6.11.

rank

is totally

W(t,r)

Is

'e

Then detQ(0

[CHAP.

is

rj

W~i

If W(t, r) is positive definite,

u(r)

Then choose

exists.

-Bt(T)*t(ti,r)W->(o.*l)x(ii)

Substitution will verify that


X(ti)

*(*!, r) B(t) u(7) dr

so that the system is totally controllable if W(t, t) is positive definite.


Now suppose the system
totally controllable and show W(t, r) is positive definite.
First note for any constant vector k,

(k,Wk)

=
=

(z,

kt*(T,

7;)

8(7?)

Bt(,) *t(T, ,)k

||Bt()*t(r,,)k|||d

d,;

Suppose
is to show
is nonsingular if the system is totally controllable.
such that
Then there exists a constant n-vector z
singular, to obtain a contradiction.
Wz) = 0. Define a continuous, m-vector function of time f (t) = Bt(t) *t(to, t)z. But

Therefore the problem

is

is

'i|f(r)||^dr

J 'zt*o,<)B()Bt(i)*t(io,f)zdi

(z,w(j,gz)

f t(i) u(t)

dt

zt*(t(

t)

=
for any u(f).

Substituting for

f(f)

Bit) u{t) dt

gives

(6.18)

'0

In particular, since the system is assumed totally controllable, take u(t) to be the control that
Then
transfers
to
*(fi, to)z - 0.

J '*(o,*)B(t)u {t)dt

Substituting this into equation (6.18) gives


Therefore no nonzero z exists for which (z, Wz)

= ztz which is impossible for


= 0, sa that W must be positive

any nonzero
definite.

z.

CHAP.

CONTROLLABILITY AND OBSERVABILITY

6]

143

Supplementary Problems
6.9.

Consider the bilinear scalar system di/dt = u(t) 4(<).


It is linear in the initial state and in the
control, but not both, so that it is not a linear system and the theorems of this chapter do not apply.
The flow diagram is shown in Fig. 6-10. Is this system completely controllable according to Definition 6.1?
Multipliei

u(t)

\,i(l)

[^

Fig. 6-10

6.10.

Given the system

dx

(-2

dt

0)x

Determine which states are observable and which are controllable, and check your
the transfer function.
6.11.

work by deriving

Given the system

dx

/3

dt

M,

= a

i)x

Classify the states according to their observability and controllability, compute the
find the transfer function.

P and Q

matrices,

and
6.12.

Six identical frictionless gears with inertia I are mounted on shafts as shown in Fig. 6-11, with
A torque u(t)
a center crossbar keeping the outer two pairs diametrically opposite each other.
is the input and a torque y(t) is the output.
Using the angular position of the two outer gearshafts
as two of the elements in a state vector, show that the system is state uncontrollable but totally
output controllable.

|2()

""

"(ofc:

u(t)

Fig. 6-11

Fig. 6-12

6.13.

Given the electronic circuit of Fig. 6-12 where u(t) can be any voltage (function of time),
what conditions on R, L^ and Z2 can both i^it-s) and ^(ii) be arbitrarily prescribed for
given that ix(to) and t2(*o) can be any numbers?

6.14.

Consider the simplified model of a rocket vehicle

'$\

Under what

conditions

is

the vehicle state controllable?

Under

CONTROLLABILITY AND OBSERVABILITY

144

6.15.

Find some other construction than equation


arbitrary

[CHAP.

that will transfer a zero initial condition to an

(6.2)

z(ti).

6.16.

Prove Theorem

6.5.

6.17.

Prove Theorem

6.6.

6.18.

What

6.19.

Given the controllable sampled data system

are the conditions similar to Theorem 6.2 for which a two-input system

iin

+ 2) +

3j(w

+ 1) +

2j(n)

u(n

+ 1) -

is

totally controllable?

u(n)

Write the state, equation, find the transition matrix in closed form, and find the control that will
(This control depends upon
force an arbitrary initial state to zero in the smallest number of steps.
these arbitrary initial conditions.)

6.20.

Given the system with nondistinct eigenvalues


-1

-r

dx

dt

-3/

M,

2/

-l)x

(0

\1,

Classify the elements of the state vector z corresponding to the Jordan form into observable/not
observable, controllable/not controllable.

Q=

6.21.

Using the

6.22.

Consider the discrete system

where x

(a)

Is

is

criterion

a 2-vector, m

(b

is

Ab

An-ib), develop the result of Theorem

x(fc

A =

a scalar,

the system controllable?

1)

Ax(fc)

j,

bM(fc)

the initial condition

If

(6)

6.1.

i-ij-

is

x(0)

'

^^^ *^

sequence m(0), m(1) required to drive the state to the origin in two sample periods
6.23.

Consider the discrete system

J.;
X(,K -r l)
x(A:

where x
(a)

is

Is the

a 2-vector, y

is

a scalar,

i\x.\n,),
Ax(A;),

y\n,j
y(k)

ctx(fc)

ct

^\
1^

/9
2

(1

j/(l)

constructing a bounded u(t) instead of equation

(6.7).

(fe)

x(2)

0).

2).

Given the observation sequence

system observable?

(i.e.,

control

y(2)

8,

14,

find the

initial state x(0).

6.24.

6.25.

6.26.

Prove Theorem

6.8,

Hint:

See Problem

6.8.

Du,
Given the multiple input-multiple output time-invariant system dx/dt = Ax -f Bu, y = Cx +
the Q
where y is a fc-vector and u is an m-vector. Find a criterion matrix somewhat similar to
matrix of Theorem 6.8 that assures complete output controllability.
Consider three linear time-invariant systems of the form
dx<-/dt

Sj.-

(a)

A)x("

B>u>,

y"'

C'x(

1, 2,

6-13 in terms of
Derive the transfer function matrix for the interconnected system of Fig.

A>,B( and C\

i=

1,2,3.
i/(0

Si
y(3)

,,(3)

S3

M c
1/(2)

Fig. 6-13

CHAP.

CONTROLLABILITY AND OBSERVABILITY

6]

If the overall interconnected system in part (a)

(6)

is

observable,

145

show that S3

is

observable.

(Note that u"' and y"' are vectors.)

6.27.

Given the time-varying system

i
Is this

"""-"-"'^

O'-'G)-

(J

system totally controllable and observable?

6.28.

Prove Theorem

6.29.

Prove a controllability theorem similar to Theorem 6.10 for the discrete time-varying case.

6.30.

continuous-time case.

Similar to Problem 6.8, show that the time-varying system dx/dt = A(t)x 4- B(t)u is completely
controllable for tj > t if and only if the matrix W(t, t) is positive definite for every r and some
finite t> T.
Also show this is equivalent to linear independence of the rows of *(t, 1;) B(i;) for

some

6.31.

6.9 for the

finite

>

17

t.

Prove that the linear time-varying system


only

if

M(ti,

to)

is positive definite

for

mtuh)

Answers
6.9.

6.10.

No

nonzero

The

to

t^

>

t^,

A(t)x,

C(t)x

is

totally observable if

and

where

J" t((, to) ct(t)

c(t)

(,

g dt

Supplementary Problems

can be reached from

iit^)

dxjdt

all

|(to)

so the system is uncontrollable.

0,

states belonging to the eigenvalue 2 are unobservable

The transfer function


controllable and observable.
trollable.

6.11.

One

6.13.

Li

6.14.

Mg #

6.15.

Many

is

and those belonging to -1 are uncon3(s4-3)-i, showing only the states belonging to -3 are both

state is observable and controllable, the other

is

neither observable nor controllable.

La

and ^

7^

choices are possible, such as


71

u{t)

2
fc=i

l^k{U[t-

to

(i

- mk - \)ln] -

U[t

to

(ti

- to)fe/i]}

where U(t

t)

is

expression for n^

6.18.

a unit step at
is different

t;

another choice

from equation

{6.4)

is

u(t)

= 2
k=

and must be shown

Two Jordan blocks with the same eigenvalue can be controlled if


the coeflScients of Mj and Mj in the last row of the Jordan blocks.

iik^~^^*.

In both cases the

to

have an inverse.

/11/22 "~/i2/2i

0,

the /'s being

CONTROLLABILITY AND OBSERVABILITY

146

6.19.

x(i+l)

*{n,k)

w(0)\

x(n)

_'^

m(w);

lA

(-l)-fc(^_^

y(n)

(_2)n-ic

(1

The flow diagram

0) x(w)

^2

^J

-27 V 0^2(0)/
shown

is

where

in Fig. 6-14

Zi

and

2 ^^^ controllable,

Zj

and

Zg are observable.

+
2

+'

0j

<EH

0J
^

+
I

Fig. 6-14

for

and p

in

T =

.0

6.22.

Yes; m(0)

6.23.

Yes; x(0)

-4, m(1)

2p

P-/

2.

G)=

6.25.

rank E

6.26.

H(s)

6i!7.

It is controllable

fe,

/13 -5\/a;i(0)\

6 1^10

M(l)y

6.20.

(_^

[CHAP.

where

R=

(CB

CAB

. . .

CA"-iB

c<3)(Is-A(3))-iB")[C(i'(Is-A(i>)-iBa>

but not observable.

D)

C(2>(Is-A(2>)-iBC2)]

chapter 7
Canonical Forms
of the State Equation
7.1

INTRODUCTION TO CANONICAL FORMS


The general state equation dx/di = A(i)x + B{t)u

appears to have all n^ elements of


the A{t) matrix determine the time behavior of x(). The object of this chapter
is to reduce
the number of states to
observable and controllable states, and then to transform the m^
elements of the corresponding- Ait) matrix to only
elements that determine the inputoutput time behavior of the system. First we look at time-invariant systems, and then
at
time-varying systems.

7.2

JORDAN FORM FOR TIME-INVARIANT SYSTEMS

Section 2.4 shovs^ed hov^^ equation (2.21) in Jordan form can be found from the transfer
function of a time-invariant system.
For single-input systems, to go directly from the
form dx/d = Ax + bM, let x = Tz so that dz/dt = Jz + T-^hu where T-AT = J. The

matrix T
page 97.

For

is

arbitrary to

vi^ithin

n constants

so that

T = ToK

as defined in Problem 4.41,

= Az + K-'T^'hu, where
is a diagonal matrix with
Defining g = To"'b, the equation for each state is dzi/dt =
\iZi + {gtU/kii).
If gi = 0, the state Zt is uncontrollable (Theorem 6.1) and does not
enter
into the transfer function (Theorem 6.7).
For controllable states, choose ku = gu Then
the canonical form of equation [2.16) is attained.
distinct eigenvalues,

dz/dt

elements ku on the diagonal.

For the case of nondistinct eigenvalues, look at the


one input, dz/dt = Lz
as in equation (2.21).

If the

system

system of one Jordan block with


desired that T-^b = ei,

is controllable, it is

Then using T = ToK, we require T<r'b= Kei = iai m-i ... m)'',
are the I arbitrary constants in the T matrix as given in Problem 4.41. In this
the canonical form of equation (2.21) can be obtained.

where the

manner

+ T-^hu.

ai

Therefore by transformation to Jordan canonical form, the uncontrollable and unobservable states can be found and perhaps omitted from further input-output considerations.
Also, the n^ elements in the
matrix are transformed to the n eigenvalues that
characterize the time behavior of the system.

7.3

REAL JORDAN FORM

Sometimes it is easier to program a computer if all the variables are real. A slight
drawback of the Jordan form is that the canonical states z{t) are complex if A has any
complex eigenvalues. This drawback is easily overcome by a change of variables. We keep
the same Zi as in Section 7.2 when Ai is real, but when Ai is complex we use the following
procedure. Since A is a real matrix, if A is an eigenvalue then its complex conjugate A*
is also an eigenvalue and if t is an eigenvector then its complex conjugate t* is
also. Without
loss of generality we can look at two Jordan blocks for the case of complex eigenvalues.

/z

dt[z*)

/L

U
147

V^

L*A/*

CANONICAL FORMS OF THE STATE EQUATION

148

If

Re means

"real part of"

^/Rez + ilmz\
dt\B,ez

and Im means "imaginary part

_
"

- jimz)

[CHAP.

of", this is

\/'Rez + jimz
ReL ;ImLy\^Rez - yimz

/ReL + jImL
\

ReLRez ImLImz + yReLImz + /ImLRez


ReLRez - ImLImz - yReLImz - jlmLRez
By

equating real and imaginary parts, the system can be rewritten in the "real" Jordan

form as

d^/Rez\
dt

7.4

/ReL -ImLN/RezN
\ImL
RelaJXImzJ

\Jmz)

CONTROLLABLE AND OBSERVABLE FORMS FOR


TIME-VARYING SYSTEMS
We can easily transform a linear time-invariant system into a

controllable or observable

done for timesubsystem by transformation to Jordan form. However,


general.
In this
form,
in
varying systems because they cannot be transformed to Jordan
subobservable
and/or
section we shall discuss a method of transformation to controllable
to
applicable
is
method
systems without solution of the transition matrix. Of course this
time-invariant systems as a subset of time-varying systems.
this cannot be

We

consider the transformation of the time-varying system

dx/dt

A^()x

+ B^()u,

{T.l)

e(i)x

and observable subsystems. The procedure for transformation can be


extended to the case y = C^(i)x + D^(*)u, but for simplicity we take D^{t) = 0. We adopt
the notation of placing a superscript on the matrices A, B and C to refer to the state variable
into controllable

because

we

shall

make many transformations

always be assumed that A{t), B() and C{t) are differentiable n- 2,


times, respectively. The transformations found in the following sections
and second ("phase-variable") canonical forms {2.6) and {2.9) when applied

In this chapter

n-1

n-1

and

of the state variable.

it

will

lead to the first


to time-invariant systems as a special case.

Before proceeding,

Theorem

7.1:

we need two preliminary

theorems.

system {7.1) has a controllability matrix Q^(i), and an equivalence


transformation x{t) = T{t) z{t) is made, where T{t) is nonsingular and difsystem
ferentiable, then the controllability matrix of the transformed
Q^() = 1-\t) Q^(*) and rankQ*(i) = rankQ^().
If the

The transformed system is dz/dt = \'{t)z + B^{t)n, where A? = T-\A.^-d1ldt)


need to
=
Q^) and Q^ is similarly partitioned, we
T~>B^ Since Q^ ^^ (Qf Qf
and B^
T 'Q?.
T-^B^
=
=
=
B^^
First
=
induction.
using
Q^
l,2, ...,%
show Qj = T->Q5^for
Then assuming Q^_i = T-Q^_i,
Q^, = -A.'q,l^,+d^l_Jdt
^ -T-i(A^T-dT/dt)(T-Q^_i) + dX-VdiQ^-i + T-^dOll-Jdt
= T-^{-A.-q,l-r + dikl-Jdt) = T-Q^
Proof:

A;

Now Q^(t)
for fc = 2,3, ...,.
rankQ^() = rankQ^() for all t.

T-i()Q-(<)

and since

T(*)

is

nonsingular for any

t,

CHAP.

CANONICAL FORMS OP THE STATE EQUATION

7]

reassuring to

It is

by a change of

theorem

that the controllability of a system cannot be altered merely


As we should expect, the same holds for observability.

system (7.1) has an observability matrix P^(f), then an equivalence


transformation x{t) = T{t) z{t) gives P^(i) = P^(i) T(f) and rankP^(t) =
If the

7.2:

rank

The proof

know

state variable.

149

P^().

Theorem

similar to that of

is

7.1.

Use of Theorem 7.1 permits construction of a T<^(i) that separates {7.1) into its conand uncontrollable states. Using the equivalence transformation x(i) = T'^(<)z(i),

trollable
(7.1)

becomes

where the subsystem


dzjdt
is

i ^ n

of order

The main problem


values of

t.

and has a

This shows

rankni(e.d.).

Also,

Theorem 73:

we

zi is

+ Bi(i)u

Aii(*)zi

matrix

controllability

controllable

(7.3)

and

Z2 is

is to keep T'^(t) differentiable


will find Q"^ such that it has n

Q^'

(i)

and nonsingular everywhere,

rii

with

Qn'i)

(7.2).

for

i.e.

all

zero rows.

The system {7.1) is reducible to {7.2) by an equivalence transformation if


and only if Q^(<) has ranki(e.d.) and can be factored as Q* = Vi(SlR)
where Vi is an nXni differentiable matrix with rankwi everywhere, S is
an wi X mni matrix with rank i(e.d.), and R is any ni x m{n - ni) matrix.

Note here we do not say how the factorization Q^

Vi(S R)

is

Proof:

= (QM Q2M

uncontrollable in

First assume {7.1) can be transformed by x(i)

Using induction,

QJ

= (^2\
?) - (f

/'^'^

and

if

QJ

obtained.

T-(i) x{t)

('<")

to the

tlien for

form

of {7

1, 2,

Therefore

VO-O
where

{t) is the

...0J~0

niXm{n-ni) matrix manufactured by

the iteration process {7.4) for


Since Q*i has rank Wi(e.d.), Q^ must also. Use of. Theorem 7.1 and
the nonsingularity of T-^ shows Q^ has ranki(e.d.). Furthermore, let T<^(i) =
(Ti() TaC*)),
so that

ni,ni

+ l,

...,n-l.

Q^{t)

T^{t)q^{t)

(TiQ^i TiF)

(Ti(),T2(i))(^"

Ti(Q^i F)

Since Ti(i) is an % x ui differentiable matrix with rankni everywhere and Q^i


matrix with rank ni(e.d.), the only if part of the theorem has been proven.

is

an ni x mni

CANONICAL FORMS OF THE STATE EQUATION

150

For the proof

[CHAP. 7

in the other direction, since Q," factors, then

Qn*)

{v^{t) y^it))

(^^f

^f)

where Yzit) is any set oi n-ni differentiable columns making \{t) = (Vi Va) nonsingular.
But what is the system corresponding to the controllability matrix on the right? From

Theorem

[^'^

6.10,

V)

= B^

Also,

-^.^

-(aI t)C')-l(?)

and

"

Therefore

Problem

3.11,

-Aliit)S{t),

Aliit)

0.

Uli

AI2A

dt[Q

has rankwi(e.d.) and Ali(fl is continuous, by


Therefore the transformation V(i) = (Vi{t) Viit)) is the required

and since

S{t)

equivalence transformation.
relationship for observability

The dual

x(<)

is

T^t)

z{t)

that transforms

(7.1) into

where the subsystem

= AUt)z3

dz3/di

is

of order

Ws^m

(7.6)

C?()z3

Then

and has an observability matrix F^t) with rankW3(e.d.).

F'{t)

has n n3 zero columns.

Theorem

The system

7.4:

and only

nsXn

if P^^

has rankw3(e.d.) and factors as P^

differentiable

matrix with rank

The proof
set of

n-ns

is

reducible to (7.5) by an equivalence transformation

(7.1) is

= (-y-JRa where R3 is an

matrix with rankws everywhere and S

is

a fcnsXns

n3(e.d.).

similar to that of

differentiable

if

Theorem

7.3.

Here T"

= f^']

rows making T nonsingular and S

is

'

where R4

is

any

the observability matrix

of the totally observable subsystem.

We

can extend this procedure to find states wi, W2, W3 and W4 that are controllable and

uncontrollable
observable, uncontrollable and observable, controllable and unobservable, and

and unobservable

The system

respectively.

y
in

which wi

is

an ws-vector for

1, 2, 3,

{7.1) is

(Cr C^

transformed into

0)w

4 and where the subsystem

C^-^)

CHAP.

CANONICAL FORMS OP THE STATE EQUATION

7]

d_/w,\

has a controllability matrix

[ Jj

!(::)

^')

if

has an observability matrix


(^^^|

vwi\ ^

/Ar,

rank

of

of

/Br\

n.+us^ w(e.d.) and where

rank

the subsystem

< ,(::)

aI)(::)
^^^^^

151

m + n^ ^ n(e.d.).

,...,

Hence these subsystems

are totally controllable and totally observable, respectively.


Clearly if such a system as
(7.7) can be found, the states wi, vi^s, ws, W4 vi^ill be as desired because
the flow diagram of
(7.7) shows W2 and W4 disconnected from the control and ws and W4
disconnected from the
output.

Theorem

The system [7.1) is reducible to {7.7) by an equivalence transformation if


and only if P^ has rank%i + n2(e.d.) and factors as P^ = RiUi + R2U2 and
Q^ has rankni + n3(e.d) and factors as Q^ = ViS' + VsS^

7.5:

HereRX^) is a A;n x rii matrix with rankni(e.d.), S\t) is an ^ x mn matrix with rankni(e.d.),
an ni X differentiable matrix with rank ni everywhere, Vi(i) isannxni
differentiable
matrix with rank w; everywhere, and Vi{t) V^i) = SIi. Furthermore,
the rank of R* and S'
must be such that the controllability and observability matrices of (7.8) have
the correct
Vi{t) IS

rank.

Proof:

First assume {7.1) can be transformed to

Q^

^ ^ ^ ^

p^

{7.7).

Q^ and P^ have rank

rii

1(t)

+ na (e.d.) and

Ui

+ Ui (e.d.),

(V, V, V,

VJ

reasoning similar to

/Pu
^-

Pl2

Gai

G32

\Gii

G42

(7.^),

P-

[
I

SO that

By

respectively.

Let

"'
(IX

Then

Q^

Vi(Qii Qi3 F12 F14)

V3(Q3i Q33 F32 F34)

and
P^

which

is

(Pfi

PJi

Gl GJO^Ux +

(P5 PJ, GJ, GJ2)^U2

the required fornl.

Now suppose P^ and Q^ factor in the required manner. Then, by reasoning similar to
the proof of Theorem 7.3, R' = (Pj; Pj Gj, Gj)^
for ^ = l,2 and S^ = (Qa Q F,2 ,,)
for i = 1,3, and Ars Ar4 A2"; A^s A^*, A4"; and Ar3 have all zero elements.
,

Theorem

7.5 leads to the following construction procedure:

= WS^

1.

Factor P^Q^

2.

Solve P^Vi

3.

Check that UiVi

to obtain

Ri and UiQ^

I,.

W
S>

and

S^.

for Vi and Ui.

CANONICAL FORMS OF THE STATE EQUATION

152

= R2U2 and Q^-ViSi =

V3S8 to obtain R^U2,V3 and

4.

Factor P^-R>Ui

5.

Find the reciprocal basis vectors of U2 to form V2.

6.

Find V4 as any

7.

Check that UiV;

8.

Form

T{t)

columns making

set of nt differentiable

[CHAP.

S.

T{t) nonsingular.

8Ii.

(Vi V2 V3 V4).

make

Unfortunately, factorization and finding sets of differentiable columns to


nonsingular is not easy in general.
Example

7.1.

1\

/sint

Given P^

sin

Obviously rankP^

).

and

it

is

T(t)

factored by inspection, but suppose

we

column operations analogous


try to mechanize the procedure for the general case by attempting elementary
to

Example

page

3.13,

Then

44.

fsint
[sint

The matrix on the right


i

- 1, 0, 1,

However,

is

l\/(sint)-i -(sint)-2\

(sin*)-i/

l)\

/I

0\

\^i

0/

perfect for P^, but the transformation matrix is not differentiable at

iir

for

if a{t)

and

Pit)

are analytic functions with no

[m

^(*)

common

zeros, let

a{t))

always nonsingular and differentiable.


analytic.
This gives us a means of attempting a factorization even if not all the elements are

Then

p{t))E{t)

{a{t)

(^^(f)

+ I3^{t)

0)

and

E(i) is

and p{t) are analytic but have common zeros


can be fixed up as
If a{t)

where

p^ is

Example
Let

the order of their

Ci, ^2,

.,ik,

f2

and

p(t)

tK

common

Their

zero C^ and y^{Q

common

zero

Then

is

is

is

?i

a convergence factor.

with order
,^

Note E(0)

the matrix E(t)

.,

(;S"!i:i)n(i-"-"^A>

7.2.

a(t)

Choose

pj.

y^iX^)

_
-

>.2

i^.

.-

nonsingular.

or column operation, it is
Using repeated applications of this form of elementary row
E.(t) such that P^T = (P^^ 0), and^simUarly
obvious tlat we can Ind T(t) = E(t)E-.(t)

for Q^ if P^ or Q^ is analytic and of fixed


then P^ factors as P'^'Ui.

75

rankr=^w.

Also, denoting

noted.
encountered if P^ or Q- changes rank should be
complicated.
or controllers in this case can become quite

The many
filters

(j^J'

Design of

difficulties

CANONICAL FORMS FOR TIME-VARYING SYSTEMS

the Jordan form exists for a general


discussed in Section 5.3, no form analogous to
study transformations to forms analogous
time-varying linear system. Therefore we shall
forms of Section 2.3.
to the first and second canonical

As

CHAP.

CANONICAL FORMS 6f THE STATE EQUATION

7]

Consider

first

the transformation of a single-input, time-varying system

dx/dt
to a

153

form analogous

Ait)x

+ h(t)u

(7.9)

to (2.6),

l-a,(t)

...

l0\

0\

-2()

dz
z

dt

Theorem

Note

-_,(i)

...

-ajt)

...

0/

(7.10)

\1|

The system (7.9) can be transformed to (7.10) by an equivalence transformation if and only if Q^(<), the controllability matrix of (7.9), is diflferentiable and has rank n everyvs'here.

7.6:

must be more than totally controllable to be put in the form (7.10)


everywhere, not w(e.d.). However, using the methods of the previous

this implies (7.9)

in that

rank

Q=

section, the totally controllable states can be


in canonical form.

Proof:
This
Q^(i)K, where

found to form a subsystem than can be put

proven by showing the needed equivalence transformation

is

K is a nonsingular constant matrix.

matrix Q^(i)

is

partitioned as
(qf

|
|

x=

First let

Q^(i)w,

where the

T(Q

nxn

Then

qj).

(-l)v

-1
dvr

-1

(-l)"-2a.

W +

(7.11)

-1
This
q?+i

true because b^ = Q^b =


qf and Q^-HA-Q--dQVdi) = A"- so that
for 1 = 1,2, ...,n-l.
Also A^e^ = Q^->(A-q- -dqj/df).
Setting

is

\(-l)-l

form

w = kz,

where

-1

K
will give the desired

-AX + dqf/(i =

0/

(7.10).

If the system can be transformed to (7.10) by an equivalence transformation, from


Theorem 7.1 Q^ = TQ^ Using Theorem 6.10 on (7.10), Q- = K-i which has rank n everywhere and T by definition has rank n everywhere, so Q^ must have rank n everywhere.

Now we
is

also

consider the transformation to the second canonical form of Section 2.3, which
as phase-variable canonical form.

known

'

'

dz

1^

'

dt

^
J

Mt)

a2(t)

a3(t)

a(t)
1

ll

(7.12)

CANONICAL FORMS OF THE STATE EQUATION

154

The

controllability

matrix

of this system

[CHAP.

is

(-1)"-^'

(-1)"-^

Qn-l.n-l

and

Q^

where

qu

(-l)'an for

Q'n-2,2

Qn-1.2

Qn-2,

Qn-IA

l^i^n;

and for

q,^

Theorem

7.7:

(-Ifan-i+k

An-

'^

+ {-If

l^k<i n,

i->'-i

q+i

an-jQi-KHi

2/ (-1)'

I.

.LI

^ff

can be transformed to (7.12) by an equivalence transdifferif Q^(t), the controllability matrix of (7.9), is
formation if
everywhere.
rank
n
entiable and has

The system

{7.9)

and only

and Q^ has rank n everywhere.


to {7.11) and
Therefore the proof proceeds similarly to that of Theorem 7.6, transforming
from
(7.11), note
then setting z = Q^w. To determine the ai{t)^ from the m{t) obtained
-A^Q^ + dQ.'^/dt = -Q^A*" when written out in its columns gives
Proof:

T=

The transformation matrix

Iq^Wn+i)

(q|lq|

Q-Q^-',

(qi

q^ -Q^Ae)

solved recursively
Therefore q^+i = -Q"A'^e, which gives a set of relations that can be
for the ai{t) in terms of the ;(*) of {7.11).

Example

7.3.

Then

Suppose we have a second-order system.

eta

/922

Ini

ai

al

da2/dt

and
-Q==^A'e

'0

-1

02

By

equating the two expressions

it is

"2

possible to find, recursively,

a^

aia2j

-a^ and a^

-a^

- da^ldt.

the b^(i) of
appears that the conditions of Theorems 7.6 and 7.7 can be relaxed if
No results
e.
of
instead
time
of
function
equation {7.10) or {7.12) is left a general vector
are available at present for this case.
It

Note that if we are given {7.9), defining y =


sponding scalar nth order equation d"2//di" = aiy

Zi

in {7.12) permits us to find a corre+and" ^y/dt ^ + u.

+ a2dy/dt+

approach is to set
For the case where u is a vector instead of a scalar in (7.9) a possible
rank
n everywhere
the resulting Q- has
all elements in u except one equal to zero, and if
or desirable,
possible
not
then the methods developed previously are applicable. If this is
the form

dw

^A^i

Ar2

aV

A2"i

A^2

All

bw2

-I-

It
,An

may be

obtained,

Al

...

where the

ft

ATil

\0

...

h\

time, and
are in general nonzero n-vector functions of

{7.13)

CHAP.

is

CANONICAL FORMS OF THE STATE EQUATION

7]

of the

form

and for

{7.11);

155

i ^ j,

m
m

This form is obtained by calculating


Q^ matrices for each of the
columns of the B
matrix, i.e. treating the systems as
single-input systems.
Then from any I of these
single-input Q^ matrices, choose the first i columns of the first Q^ matrix, the first
columns of the second Q^ matrix,
and the first Ui columns of the Ith. Q^ matrix such that
these columns are independent and differentiable and i + W2 +
+ ni = n. In this order
these columns form the T matrix. The proof is similar to that of Theorem 7.6.

To transform from this form to a form analogous to the first canonical form {7.10),
use a constant matrix similar to the K matrix used in the proof of Theorem 7.6. However,
as yet a general theory of reduction of time-varying linear systems has not been developed.

Solved Problems
7.1.

Transform

x(m +

1)

x(m)

u{m)

to Jordan form.

From Example

4.8,

page

and Problem

74,

4.41,

page

97,

K =
Then
2
1
/?!

from which

ag

2,

-1,

ToK

aj

= -3

p^

0\ /-I

1
1

x = Tz

0\

-1

1/

/-I
-1

in the system equation then gives the canonical

/2
z(m

1)

0\
j

Vo

2/

/O
z(m)

2
1

-3/

-1/ \

V-1
Substitution of

The proper transformation

are obtained.

u{m)

-2

form

is

then

CANONICAL FORMS OF THE STATE EQUATION

156

7.2.

[CHAP.

Transform
/

dxldt

-1

1
-0

0'

2\
-2 X +

\i.

,1/

to real Jordan form.

From Problem

4.8,

page

and Problem

93,

4.41,

page

97,

Substitution into the system equation gives

dildt

,(1

Then

-2

and p

1j.

Re2

Using the methods of Section

(i)

)"

/I

0\
X +

0-1/

to a controllable system,

reduce

7.4,

dx/dt

+ j)/p*J

Putting this in real form gives

Im 22/

7.3.

\u

(l-i)//3

i^

(1

1 2)x

\-2.
an observable system,

to

(ii)

2 \u,

(iii)

to a controllable

and observable system.

(i)

Form

Observe that

Q=

(b

-Ab

and not

time-invariant case,
elementary row operations to

the controllability matrix

A^b)

Q=

(b

in the
in accord with using the form of Theorem 6.10
A^b) which is the_f orm of Theorem 6.8. Ferformmg

Ab

make the bottom row

of

Then the required transformation

is

Tz where

and

gives

CHAP.

CANONICAL FORMS OF THE STATE EQUATION

7]

Using

this

= T->(AT - dT/dt)z + T-'btt.

change of variables, dzldt

-2 -1

/l\

0\

dx/dt

|z

Jm

2/
(ii)

Forming the

/
(

2\

-1

5/

where the factorization


formation X

(-1 -1 2)z

\0/

/
=

0\

-1

)[

0/\0

made using elementary column

is

-1

0\

P^(TO)-i

1/
Using the trans-

transformations.

T^z,

-3

-1

dz/dt

/-A

o\

o)z+

-2

-2

(iii)

2/

observability matrix gives

P^

157

OJM

i/

= (12

0)z

\-2/

1/

Using the construction procedure,

-1 -1

-1

P^Q*

Then

/l-t>3
P^V,

-1

0-21

)(

-1

R'

gives

Vi

...

-v^i

^31

and

UiQ^

(Mil Mi2 "is)

(-1 -1 -1)

SI

gives

Uj

(1 u^g

-1

M13)

0-2/

\-2

Note UiVi = 1 for any values of M13 and Vg^. The usual procedure would be to pick a value
for each, but for instructional purposes we retain M13 and v^i arbitrary.
Then

P^

RlUi

/O

(0

\0

- Mi3
+ Mi3
- Mi3

2
1

+
-

uia"

Mi3

Mi3

(0

1)

R2U2

Mi3^

and

/
Qx_ViSi

Choosing Vgi

2--y3i -V31
2-'i;gi

-vai

\-2 + i;3i

i;3i

and M13

2-i;3i\

2-Vsi
-2 + ^31^

\0
Using UjVj

Sy

^12

l\

^22

V32

-1/

we can determine
/l

)(2--ygi -yg, 2

gives

/l

1'

1
I

Vo

0-1.

1
1

\M31

M32

M33/

- Vgi)

VgSa

CANONICAL FORMS OP THE STATE EQUATION

158

The reduced system

is

[CHAP.

then

2/

0)z

(1

V3y
is controllable and observable, z-i is observable and uncontrollable, and 3 is conSetting Wgi = Wij = 1 leads to the Jordan form, and setting
and unobservable.
leads to the system found in (ii), which coincidentally happened to be in the
''31 = '"\z correct controllability form in addition to the observable form.

where

trollable

7.4.

Reduce the following system

to controllable-observable form.

X +

dxjdt

Using the construction procedure following Theorem

2/

(0

-l)x

i2

7.5,

-5 -16 -2t -6t-40'

Qx

-1

-1
12t - 1
-3t3 +
24*2
+ i5t - 18
-3t* + 2t3
-3<2

PX

Then from

P^Q*

Ri and UjQ^

SI

(2

2 2 2)

gives

S^

Vi

Factoring

Ui

P^-RiUi and Q^-ViSi

(0

0)

gives

/
U2 =
S3

-3(2

'

1)

(0

R2
(t

Then Vj

-5 -16 -2i -6t-40)

(0

R^Si,

Ri

Solving P^Vi

2f2

1)

and vj

-3(3

\-3(4

(0

that the equivalence transformation

will also put the system in the

form

(7.7).

1 0)

and

T = (ViiV2|V3|V4).

2(3

2(2

-t
- 12t -

24(2

i5t

It is interesting to

18/

note

CHAP.

7.5.

CANONICAL FORMS OF THE STATE EQUATION

7]

Reduce the following system

to controllable form:

sint

j
First

we

calculate

E(f)Qa:

The required transformation

tcosA

_
~

coa^tj

j\(iost

/t2

to obtain

t^cost

cos2t

cos^t

is

A/

cos

\ cost

E-i(e)

t^

7.6.

\coatJ

cos t

Q^ and use elementary row operations

/'

159

C0s2

-cost

V cos t

Put the time-invariant system

=223
0-1/
/

x(m +

l)

0\

/1\

x(m)

\-2

(i)

into first canonical

form

{7.10)

and

(ii)

U(m)

\2/

into phase-variable canonical

form

(7.12).

Note this is of the same form as the system of Problem 7.3, except that this is a discrete-time
system.
Since the controllability matrix has the same form in the time-invariant case, the procedures developed there can be used directly.
(See Problem 7.19 for time-varying discrete-time
systems.)

From

part

(i)

of Problem

7.3,

the system can be put in the form

/
z(m +1)

\
Therefore the best

we can

do

is

z(m)

+0

2/

u{m)

\oJ

_i)^iW + (o)W

(_2

The required transformation

\0
to the first canonical

put the controllable subsystem

zi(m-l-l)

into the desired form.

/A

o\

-2-1

is

2j\-l

1^-2

0/

form
Xcim

+1)

z<.(m)

( -

) m()

To obtain the phase-variable canonical form,


'1

'0

-l\/0
2J\1

-i\-i
g-u

CANONICAL FORMS OF THE STATE EQUATION

160

- -a^ =

where q^

Then

^ =

from Example

a^,

7.7.

A^-Cn

- dq^'M) we

Q^i-i(A^iq^i

obtain

from which we obtain the phase-variable canonical form

Zp(m

By

From

7.3.

[CHAP.

+ 1)

form

chance, this also happens to be in the

^jzp(TO)

JM(m)

of the first canonical form.

Put the time-varying system

(i)

sini\

ft
h

dx/d

into first canonical form,

2\

+ (-1 )m

)x

_TL

into second canonical form,

(ii)

i^O

for

and

(iii)

find a scalar

second-order equation with the given state space representation.


To obtain the

first

canonical form,

<^

{:,

-''T%:

I)

CT'

-\

from which

where
'ai(t)\

^a^it)/

2t

sint ye

6t

4i
2(2

- 2*2 + (t-1) sini - cost


- (t + 6) sin i - 3 cos + sin2

To obtain the phase-variable form,

^^

-3

[-1

-a,(t)j

J\l

ai

-1

from which
d.,/dt

{_^^\^j^^

The second-order scalar equation corresponding


d^yldt^

where y

7.8.

X)^

+ (l)

to this is

ai dy/dt

{a^

+ dai/dt)y =

z^i.

Transform

to canonical

form the multiple-input system


2

dx/dt

To obtain the form (7.13),


of the B^ matrix separately:

we

|0

-4

calculate the

-l\

/O

1\

Ox-fO lu
3/

\l

1/

two Q^ matrices resulting from each of the columns

CHAP.

CANONICAL FORMS OF THE STATE EQUATION

7]

'0

Q^

'1

(000

(Ml only)

-4^

.1-3

Q^(M2only)

1
I

8.

161

-2
-1

Note both of these matrices are singular.


However, choosing T as the
(Ml only) and the first column of Q^ (mj only) gives

4^
1

0-2,
two columns of

first

Q^

so that

dvildt

Also,

Q* (m2

could have been chosen as the first column of Q^ (mj only)

and the

first

two columns of

only).

To transform

to a

form analogous

to the first canonical form, let

w = Kz

where

Then
0'

dzldt

-1

-4

0)m

llz + (0
0/
Vl

0>

Supplementary Problems
'1
7.9.

7.10.

7.11.

Transform

Transform

dx/dt

rfx/dt

/~^

7.12.

Prove Theorem

7.2,

page

149.

7.13.

Prove Theorem

7.4,

page

150.

Show
and

-1\

OJx +
3/

/-I
(

o)m

0.

i)''"*'(-4)"

Using the methods of Section

(i) to a controllable system,


system.

7.14.

= (o 1
.1-4

7.4,

(ii)

to

to

Jordan form.

realJordan form.

reduce

an observable system,

(iii)

to

an observable and controllable

that the factorization requirement on Theorem 7.3 can be dro^^ed if T(t) can be nonsingular
differentiable for times t everywhere dense in [to. *i]'

CANONICAL FORMS OF THE STATE EQUATION

162

7.15.

Consider

system

the

"'^

elsewhere, and /2(<) =


form of equation {7.3)1

7.16.

^
\t2Wj

\^2/

for

where

and

==

/i(t)

cos

Can

1 -- cos t elsewhere.

for

this

and zero

system be put in the

Find the observable states of the system

dx/dt

7.17.

*^i
/

[CHAP.

^4

8^

=12

6'"

^8

Check that the transformation of Problem


page 149, by calculating A^ and B^.

2/

(1

l)x

page 159, puts the system in the form of equation

7.5,

{7.2),

7.18.

7.19.

Develop Theorem

7.3 for

time-varying discrete-time systems.

Develop the transformation to a form similar to equation

{7.11)

for time-varying discrete-time

systems.

t-1
7.20.

Reduce the system

dx/dt

to

7.21.

a system of the form of equation

Given the time-invariant system dx/dt

Verify the relationship for the

{7.2).

form as given by equation {7.12). Let


Az + bM and A is a diagonal matrix.
7.22.

/I'
+ 2\
+ 2Jx + (l)M
-t+1/
\0^

-i

-t-2

gj^ in

= Ax + eM where
z = Tx where z
Show

that

terms of the

is

0^,

the system is in phase-variable canonical


the Jordan canonical form dx/dt
the Vandermonde matrix of eigenvalues.
is in

following equation {7.12) for a third-order

system.

7.23.

Solve for the

Example

7.3,

aj in

terms of the

for

a;

1, 2,

manner analogous

to

page 154.
-11

7.24.

3 (third-order system) in a

Transform the system

-~

1/2
(-1/2

-27/2

0]x-(-

to phase-variable canonical form.

7.25.

Transform the system

d-s.ldt

to phase-variable canonical form.

7.26.

Using the

results of Section 7.5, find the transf.jrmation

/
dx/dt

dw/di

/
into the

form

V
7.27.

-1

2\

-3

6\

-1 -1

= Tw

that puts the system

/l

3\^

/l

0\

\0

1/

4J

0/

directly in the case of time-invariant


Obtain explicit formulas to go to phase-variable canonical form

systems.

CHAP.

7.28.

CANONICAL FORMS OF THE STATE EQUATION

7]

Use the duality principle


into the form

to find a transformation that puts the

dz

dt

...

ai(t)\

...

02(e) \

...

03(4)

...

7.29.

Prove that

<

||T||

163

system dx/dt

A(t)x and y

C(t)x

l)z

(0

()/

for the transformation to phase-variable canonical form.

Answers

Supplementary Problems

to

1-1

0/8)

7.9.

-1

7.10.

where p

is

any number

0.

2/3.

-8

-4

-12

7.11.

There

7.15.

No.

7.16.

The transformation

is one controllable and observable state, one controllable and unobservable state, and one
uncontrollable and observable state.

Q^

Vi(l

but Vj does not have rank one everywhere.

0)

puts the system into the form of equation

form can be used but


7.23.

-a3

7.24.

T-i

Q!i,

is

more

0^2 "^" 2q:x,

-3/2

5/2

-2

-1-1

r43

r4i

(7.6),

for any

r^i

that

make T nonsingular.

Also, Jordan

difficult algebraically.

aj

03

1,

1
7.25.

T-i

A
ot

'1
7.26.

-1

This form

(>t

2e~t

-e-t

1,

3^
1

.0

7.28.

fi2t

is

2,

obtained using the same transformation that puts the system

dw/dt

At(t)w

Ct(t)u

into phase-variable canonical form.

7.29.

The elements of Q^ 1 are a linear combination of the elements of Q^, which are always
determined by the recursion relation.

finite as

chapter 8

Relations with Classical Techniques


8.1

INTRODUCTION
Classical techniques such as block diagrams, root locus, error constants, etc., have been

used for

many

years in the analysis and design of time-invariant single inputs-single output


Since this type of system is a subclass of the systems that can be analyzed by
state space methods, we should expect that these classical techniques can be formulated
within the framework already developed in this book. This formulation is the purpose of
the chapter.
systems.

8.2

MATRIX FLOW DIAGRAMS


We have already studied flow diagrams

state equations.

(summer,

scalor,

in Chapter 2 as a graphical aid to obtaining the

The flow diagrams studied in Chapter 2 used only four basic objects
integrator, delayer) whose inputs and outputs were scalar functions of

Here we consider vector inputs and outputs to these basic objects. In this chapter
these basic objects will have the same symbols, and Definitions 2.1-2.4, pages 16-17, hold with
u(i) and one
the following exceptions. A summer has n m-vector inputs ui{t), U2(t),
time.

u(i).
output m-vector y(i) = ui{t) U2{t)
and one output A;-vector y{t) A(t)u{t), where A(i)

one m-vector input

u(t)

and one output m-vector

A
is

. ,

scalor has one m-vector input u(i)

y(<)

fc

matrix.

y(io)

An

integrator has

u(t) dr.

To denote

vector (instead of purely scalar) time function flow from one basic object to another, thick

arrows
Example

will

be used.

8.1.

Consider the two input one output system

=
dt ^X2/

This can be diagrammed as in Fig.

\0

Oy\a;2/

\0

(3

2)x

8-1.

*-V(t)

Fig. 8-1

Also, flow diagrams of transfer functions (block diagrams) can be

manner for time-invariant systems.

We

drawn

denote the Laplace transform of

164

in a similar

x{t) as J!.{x}, etc.

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

Example

165

8.2.

The block diagram

of the system considered in

Example
xo

shown

8.1 is

in Fig. 8-2.

^W

Jliy}

Fig. 8-2

Using equation (5.56) or proceeding analogously from block


diagram manipulation, this can be reduced to the diagram
of Fig. 8-3 where
H(s)

(6/s

(3

2)

(2s

+ 3)/s2)

['

lo

0/

Vo

^{u>:

-^^{}

H(s)

Fig. 8-3

1,

Vector block diagram manipulations are similar to the scalar case, and are as useful
Keeping the system representation in matrix form is often helpful,
especially when analyzing multiple input-multiple output devices.

to the system designer.

8^

STEADY STATE ERRORS


Knowledge of the type of feedback system that

will follow an input with zero steady


In this section we shall investigate steady state errors
of systems in which only the output is fed back (unity feedback). The development can be
extended to nonunity feedback systems, but involves comparing the plant output with a
desired output which greatly complicates the notation (see Problem 8.22). Here we extend
the classical steady state error theory for systems with scalar unity feedback to timevarying multiple input-multiple output systems. By steady state we mean the asymptotic
behavior of a function for large t. The system considered is diagrammed in Fig. 8-4. The
plant equation is dx/dt = A{t)x + B{t)e, the output is y = C(i)x and the reference input is

state error is useful for designers.

= y{t) + e{t), where y, d and e are all m-vectors.


assumed that the zero output is asymptotically stable,
d(^)

limC(i)*^_3p(t,T)
t-*

where d^^_^c(t,T)/dt

For

this

system

it

will

always be

i.e.

CO

{A{t)-B{t)C{t)]^^_^^(t,T)

and

be concerned only with inputs d{t) that do not drive


we obtain a steady state as t tends to infinity.

*^_B(,(i,

||y(*)|]

i)

I.

Further,

to infinity before

we

j,

shall

so that

y(t)
Zero output

is

asymptotically stable

A(t)

Fig. 8-4.

Unity Feedback System with Asymptotically Stable Zero Output

RELATIONS WITH CLASSICAL TECHNIQUES

166

Theorem

For the system of Fig.

8.1:

where

dvfldt

such that
on w.

lim

lim

8-4,

e(f)

[CHAP.

and only

if

if

C(t)w + g

all t^U in which g{t) is any function


and A, B, C are unique up to a transformation

A{t)w + B{t)g for


g(t)

t-* 00
t-00

Consider two arbitrary functions


Proof:
tends to >.
If lim [f (t) - h(t)] = 0, then

and

i{t)
()

b{t)

- h(t) =

whose

limits

for

r(t)

all

may

not exist as

where

f,

is

r{t)

an

t-*00

arbitrary function such that lim


then for alH,
d(t)

y(t)

r(t)

r(i)

From

0.

this, if

C(i)*^_^^(i,gx(g +

lim

e(f)

lim [d(i)-y(t)],

C{t)4>^_^^{t,r)B{r)dir)dr

r(i)

'to

r C(f)*^_3^(i.r)BWd(r)dr +
v

g{t)

C(i) *^_3,(t,

*)

w(g

to

arbitrary constant w(to)

+ C(<)*a-bc(*, *o)[x(io)-w(io)] is one-to-one for


This Volterra integral equation
because lim C(i)*A-Bc (t, to) = 0.
the differential equations dw/dt = [A{t) - B{t) C(i)]w + B{t)d and

where the change of variables

g(i)

r(t)

equivalent to
Substituting the latter equation into the former gives the set of equations
that generate any d{t) such that lim e{t) = 0.

for

d{t) is

C(f )w

+ g.

Conversely, from Fig, 8-4,


d = C(t)w + g and subtracting
d(x

lime

Then

t~*

A(t)x +B(f)e = [A(t) - B(f ) C(^)]x


dv//dt - A(i)w + B(t)g gives

djs./dt

- vi)ldt =

lim(d-y)

[A{t)

lim

[g

t-+M

00

Bit) C()](x

C(^)(x

- w)

- w)]

limg - [limC{t)*^_c(*'*o)]W*o)-w(g]
f->-00

+ B(f)d. Assuming

t-'OO

the last part of the proof we see that e{t) = g(t)-C(i)*A-Bc(, io)[x(fo)-w(*o)]
what the function g{t) is. Therefore, the system dw/dt = A(i)w + B{t)g with
d = C{t)w + g and the system dx/dt = [A{t) - B{t) C{t)]x + B{t)d with e = d- C(i)x are
inverse systems. Another way to see this is that in the time-invariant case we have the
transfer function matrix of the open loop system H(s) = C(sl - A)-iB relating e to y. Then
for zero initial conditions, =C{d} = [H(s) + I]^{g} and =C{e} = [H(s) + I]-i=^{d} so that

From

regardless of

Consequently the case where g{t) is a constant vector forms a sort of boundthat grow with time and those that decay. Of course this neglects
functions
ary between
.sin
(like
t) that oscillate, for which we can also use Theorem 8.1.
those functions

=C{g}

-^{e}.

effect of nonzero initial conditions w(io) can be incorporated into g{t).


are
interested
in only the output characteristics of the plant, we need concern ourwe
selves only with observable states. Also, because uncontrollable but observable states of the
plant must tend to zero by the assumed asymptotic stability of the closed loop system, we
need concern ourselves only with states that are both observable and controllable. Use of
equation (6.9) shows that the response due to any Wi(io) is identical to the response due to
an input made up of delta functions and derivatives of delta functions. These are certainly
included in the class of all g(<) such that lim g(t) = 0.

Furthermore, the

Since

Since the case g{t) = constant forms a sort of boundary between increasing and decreasing functions, and since we can incorporate initial conditions into this class, we may
take g{t) as the unit vectors to give an indication of the kind of input the system can follow
with zero error. In other words, consider inputs
d{t)

at)

C
'to

*^(i, t) B(t) g(r) dr

C(^)

r
'o

*^(i, r) B(r) 6, dr

f or

1, 2,

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

167

which can be combined into the matrix function


C{t)

*/

'

*^(t, r) B(T)(ei e^

em)dT

...

C{t)

*^(t, r) B(t) dr

*^ 'o

to

Inputs of this form give unity error, and probably inputs that go to infinity any
slower than this will give zero error.
Example

little

bit

8.3.

Consider the system of Example


unity feedback system

then

is

2)x

(3

limfi'(t)

which

=\(q

-^

e(t)

+ 2a!2(0)] cosV2t +

e-t|[3a;i(0)

is

no input

Ui(t).

The zero

input,

^^s output

o)~(i)^^^M''

generate the class of inputs

0,

and there

U2(t)

-iz[i(0)

+ 2(0)] sinV2tl

Consequently Theorem 8.1 applies to the unity feedback system, so that the

tends to zero asymptotically.


equations

where

8.1 in

d(t)

that the system can follow with zero error.

t-+oo

Solving this system of equations gives

d{t)

For

g{t)

0,

we

3wi(0)

2w2(0)

3fW2(0)

+ f

[S(t-T)

2]g(T)

dr

g(t)

see that the system can follow arbitrary steps and ramps with zero error, which is in
Also, evaluating
classical conclusion that the system is of type 2.

agreement with the

C(*)

aC*.

t)

B(t) dr

[Sit

- t) +

2] cJt

1.5*2

2t

shows the system


t2-e for any e >

Now

if

we

will follow t^

This

0.

is

[3a!i(0)

Definition 8.1:

is

+ 1.5a;2(0)]e~*'

e{t)

to)

Example

in

8.1

which

~ nr^

n)

e(t)

^^ \^'

Mi(t)

'^^

and there
output

is

of

no input
this

Fig.

8-4

system

is

8.1

cannot

a type-l system (1=1,2,...) when


m.
it- toy-^U{t - to)ei for all i = 1, 2,

called

is

for the inputs

u^it),

which does not tend to zero asymptotically so that Theorem

The system of
lim

di

. ,

the unit step function starting at t = to and Ci is the tth unit


All systems that do not satisfy Definition 8.1 will be called type-0 systems.

In the definition, U{t


vector.

constant error and will probably follow with zero error any function
by taking g(t) = t~^.

consider the system of

then the closed loop system

y = 0.50:2(0)
be used.

-v^ith

in fact the case, as can be found

Use of Theorem

is

8.1 involves calculation of the transition

matrix and integration of the

superposition integral. For classical scalar type-Z systems the utility of Definition 8.1 is
that the designer can simply observe the power of s in the denominator of the plant transfer
function and know exactly what kind of input the closed loop system will follow. The following theorem is the extension of this, but is applicable only to time-invariant systems with
the plant transfer function matrix H(s) = C(sl A)"'B.

Theorem

8.2:

The time-invariant system of Fig. 8-4 is of type Z^l if and only if


H(s) = s-'R(s)+P(s) where R(s) and P(s) are any matrices such that
limsR-^(s)
s->0

and

||lims'-ip(s)|[
s-0

<

>.

RELATIONS WITH CLASSICAL TECHNIQUES

168

From Theorem 8.1, the system is of type


Proof:
{l-l)\s-n= [H(s) + I]G(s) where =C{gi}, the columns
of any functions gi{t) such that
- lim g^{t) = lim
First,

assume H(s)

neighborhood of

+ P(s) where

if and only if Jl{{Ax d2


...
d^)} =
of G(s), are the Laplace transforms

s=C{gi}

lim sR->(s)

where

+ I]-i =

(Z-l)!s-'[H(s)

s=C(gi} is analytic for

so that R-i(s) exists in a

*"*"

Choose

G(s)

s-'R(s)

0.

'-'"

'-*

Res^O.

[CHAP.

(?-l)![R(s)

+ s'P(s) + s'I]->

Since [H(s) + I]~> is the asymptotically stable closed loop transfer function matrix, it is
analytic for Re s ^ 0. Then sG(s) has at most a pole of order J - 1 at s =
in the region
Res ^ 0. In some neighborhood of s =
where R"Hs) exists we can expand

= (i-l)!s[R(s) + s'P(s)+s'I]-i = (Z-l)!sR-i(s)[I-Z(s)+Z2(s)


where Z(s) = sR-'(s)[s'->P(s) +s'-il]. Since limZ(s) = 0, this expansion is valid for
s->0
small
and lim sG(s) = 0. Consequently sG(s) has no pole at s =
and must be
sG(s)

\s\,

analytic in

Re s

which

satisfies

Conversely, assume lim sG(s)

+ P(s)

s'[H(s)
P(s)]
s

'R(s)

where
is still

Theorem

8.1.

where sG(s)

is

analytic in

any matrix such that


arbitrary.
Then
P(s)

is

(Z-l)!s-'I

[s-'R(s)

lim

s'

Re s
'P(s)||

^ 0.
<

Write H(s)

and

00

R(s)

=
=

s-0

+ P(s) + I]G(s)

can be solved for sR"i(s) as


(i-l)!sR-i(s)

where

(Z-

W(s)

sG(s)(I

+ W(s))-i = sG(s)[I-W(s)+W2(s)-

+ s'-iI]sG(s).

This expansion is valid for ||sG(s)|| small


enough, so that R"'(s) exists in some neighborhood of s 0. Taking limits then gives
limsR-i(s) = 0.
1)!

[s'-ip(s)

We should be careful in the

Theorem 8.2, however, in light of Theorem 8.1.


not as clear cut as it appears. A system with
As t tends to zero this tends to a step
H{s) = (s + e)"^ can follow inputs of the form e"".
function, so that we need only take c"^ on the order of the time of operation of the system.
The

classification of

application of

systems into type

is

Unfortunately, for time-varying systems there is no guarantee that if a system is of


k for all k 0. However, this is true for time-invariant
type N, then it is of type
systems. (See Problem 8.24.).

Example

8.4.

-1

-Q

^{d};

-1
:>^{!/}

12s

3s2

Fig. 8-5

The system shown

in Fig. 8-5

has a plant transfer function matrix H(s) that can be written

-1\
"^^^

i2

0/

-V

/
"^

Vl2s-i

S-2R(s)

in which

lim sP{s)
s-0

<
\

12

-I-

3s

'^

P(8)

in the

form

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

169

and where

limsR-i(s)

wo

limsf
\-l

- \6s-^J

s-o

Since lim sR^i(s) has a nonzero element, the system

']

\0 -6/

not of type 2 as appears to be the case upon

is

Rewrite H(s) in the form (where R(s) and P(s) are different)

first inspection.

H(s)^

Again,

l/-6- + 9-

||limP(s)j|

S-1

<

-s->\
/

12

/0-l\
^ \3 0/

,-.(,)

p(,)

but now

s->0

lim sR-i(s)
S-.0

hm

*'-*"

-6
+ 12s

9s
*

/O

Vn

Since the closed loop system has poles at 1, -1, -0.5, and -0.5, the zero output
Therefore the system is of type 1.

is

asymptotically stable.

matrix of a type-Z system, we use block diagram manipulations on Fig. 8-4 to get ^{e} = [I + H(s)]-.<:(d}. If it exists, then

To

find the error constant

lime(<)

lims[I

+ s-'R(s)-FP(s)]-=C{d}

S-+0

t-*oo

lims'+i[sl

+ R(s)-Fs'P(s)]-i^{d} =

lims'+iR-HsK{d}

for any I > 0. Then an error constant matrix table can be formed for time-invariant systems of Fig. 8-4.
Steady State Error Constant Matrices

System Type

Step Input

Ramp

Input

Parabolic Input

lim R-i(s)

lim [H-H(s)]-i
S-+0
1

s-i-O

lim

In the table

Example

means the system cannot follow

all

such inputs.

8.5.

The type-1 system

/a
of

Example

8.4

has an error constant matrix

lim R-i(s)

the input were (t - to)U(t - to)e2, the steady state output would be [{t-toWit
can follow with zero steady state error an input of the form (t tQ)U(t to)ei.

8.4

R-Ms)

s->0

- tg) +

0\

-6/'

6] 63.

"^^"^

^*

The system

ROOT LOCUS

Because root locus is useful mainly for time-invariant systems, we shall consider only
time-invariant systems in this section. Both single and multiple inputs and outputs can
be considered using vector notation, i.e. we consider
dx/dt

= Ax + Bu

= Cx-f-Du

(8.1)

Then the transfer function from u to y is H(s) = C(sl- A)-iB-(-D, with poles determined
by det (si - A) = 0. Note for the multiple input and output case these are the poles of the
whole system. The eigenvalues of A determine the time behavior of all the outputs.

RELATIONS WITH CLASSICAL TECHNIQUES

170

[CHAP.

We shall consider the case where equations (8.1) represent the closed loop system.
Suppose that the characteristic equation det (si ~ A) =
is linear in some parameter k so
that it can be written as
This can be rearranged to standard root locus form under

S"

The

+ ^,s"-2+
+ ^jS"-! +

-AiS"-i

-1

+^^_^s +

...

roots of the characteristic equation can be found as

techniques.

The assumed form

variation,

^n-lS

varies using standard root locus

of the characteristic equation results

from both

loop gain

variation and parameter variation of the open loop system.


Example

8.6.

Given the system of Fig. 8-6 with variable feedback gain

k.

xo

Fig. 8-6

The

closed loop system can be written as

_
~

dx
It

The

characteristic equation is
root locus shown in Fig. 8-7.

s^

-t-

(3

-I-

k)s

-K
1
-K -3

+ 4k =

0.

4-

Putting

it

into standard root locus

form leads

Ims

Res

Fig. 8-7

Example 8.7.
The feedback system of Fig. 8-8 has an unknown
parameter a. Find the effect of variations in a upon

+ 3} sinh a
+ 3s + sinh a

(a

s2

the closed loop roots.

The usual procedure is to set


Let sinh a = k.
the open loop transfer function equal to 1 to find
the closed loop poles of a unity feedback system.

Fig. 8-8

to the

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

-1

is

g2

+ 3)
3s +

171

This can be rearranged to form the characteristic equation of the closed loop system,
+ 3)k = 0. Further rearrangement gives the standard root locus form under k variation.

s^

+ 3s + k +

(s

s(s

+4
+ 3)

This happens to give the same root locus as in the previous example for sinh a

Q.

NYQUIST DIAGRAMS

8.5

First

we

consider the time-invariant single input-

whose block diagram is shown in


The standard Nyquist procedure is to plot
G(s)H{s) where s varies along the Nyquist path enclosing the right half s-plane.
To do this, we need
polar plots of G{jo))H{j<a) where < varies from oo to

single output system

^Ui{e}

Fig. 8-9.

G(s)

J
+

+ o.

-Civ}

H(s)

Using standard procedures, we break the closed


loop between e and v. Then setting this up in state
space form gives
dx/dt

= Ax

-I-

be

Fig. 8-9

c+x

-I-

de

(8.2)

G{jo>)H{}o>) = c^{jcoI- A)-^h + d. Usually a choice of state variable x can be found


such that the gain or parameter variation k of interest can be incorporated into the c vector
only. Digital computer computation of {jo>I-A)-'^h as varies can be most easily done by
iterative techniques, such as Gauss-Seidel.
Each succeeding evaluation of {j<o._^J. A)-^h
can be started with the initial condition {j\l-A)~% which usually gives fast convergence.

Then

Example

8.8.

Given the system shown in Fig.


form for the transfer function from

The

8-10.

state space

form of

this

is,

in phase-variable canonical

e to v,

(hi.

-\)^ + (i)^

dt

Then

ct(7<oI-A)-ib

=
jaija

(^

0)x

giving the polar plot of Fig. 8-11.

1)

ImGH
-^-CW

ReGH

Fig. 8-10

Fig. 8-11

About the only advantage of this over standard techniques is that it is easily mechanized
for a computer program. For multiple-loop or multiple-input systems, matrix block diagram
manipulations give such a computer routine even more flexibility.

RELATIONS WITH CLASSICAL TECHNIQUES

172

Example

[CHAP. 8

8.9.

Given the 2 input 2 output system with block diagram

shown

G(8)

in Fig. 8-12.

Then dx/di = Ax + bjej + b2e2 and yj = cjx and fj =


The loop connecting Vi and Cj can be closed, so that
c|x.
Then
^1 "! = cjx.

^{v}
H(s)

= c|(sI-A-bict)-ib2^{e2}

^{vg}
so that

we can ask

the computer to give us the polar plot of

c|(iuI-A~biC+)-ib2.
8.6

Fig. 8-12

STATE FEEDBACK POLE PLACEMENT

Here we discuss a way to feed back the state vector to shape the closed loop transition
matrix to correspond to that of any desired wth-order scalar linear differential equation.
For time-invariant systems in particular, the closed loop poles can be placed where desired.
This is why the method is called pole placement, though applicable to general time-varying
systems. To "place the poles," the totally controllable part of the state equation is transformed via x(i) = T{t)z{t) to phase-variable canonical form {7.12) as repeated here:
...

/O

...

dz

{7.12)

di.

...
\ai{t)

Now the

scalar control

variables as

m=

kt(i)z

u =
Each

a^{t) is

as{t)

ai{t)

- ali)\z^ +

[-,(*)

a time function to be chosen.

- ali)\z, +

This

\ 1

an{t)/

u is constructed by feeding back a


where each klt) = -aJ^t)-a.{t).

[-^{t)

grives

linear combination of the z state

[-(*)

- a(t)]2

the closed loop system


...

...

az

-a^t)

\-a^{t)

-aj,t)

...

...

-ccj^i)

=
and each z.^^{t) for
are to be chosen, the corresponding closed loop transition matrix *^(t, t^) can be shaped accordingly. Note, however,
that x(i) = T{t)^jt, t^T.^ so that shaping of the transition matrix *^(<, t^) must be done
keeping in mind the effect of T{t).

Then
i

obeys

z^{t)

1, 2,

w-

z{^'

(Q2j"^"

alt)'z^

1 is the ith derivative of z^t).

a^{t)z^

Since the

a}^t)

This minor complication disappears when dealing with time-invariant systems. Then
and furthermore each a.{t) is constant. In this case the time behavior of
x(i) and z(i) is essentially the same, in that both A^ and A^ have the same characteristic
equation A" + ff^A"-+
ha^A-t-ofj = 0.
For the closed loop system to have poles at the
desired values y^, y^
(A. - y) =
y^, comparison of coefficients of the A's in (A - yj)(A - yj)
determines the desired value of each a.
T{t) is constant,

Example

. ,

8.10.

Given the system

^x

^
1

dt

V2 -3

)x

*^'

have a time-invariant closed loop system with poles at


have a characteristic equation X^ + 3\2 + 2X = 0.
Therefore
kT = (_2 1 -(t + 3)).
It is desired to

will

0, 1 and 2.
Then the desired system
we choose u = (-2 1 -(t + 3))x, so

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

173

For multiple-input systems, the system is transformed to the form of equation {7.13),
except that the subsystem dw./dt = Al? W; + hfu^ must be in phase-variable canonical form
{7.12) and for i - j, the ATi{t) must be all zeros except for the bottom row.
Procedures
similar to those used in Chapter 7 can usually attain this form, although general conditions
are not presently available. If this form can be attained, each control is chosen as
u^ = k?'(i)Wi elA^w. for j - i to "place the poles" of Ajj(t) and to subtract off the coupling
terms.
Why bother to transform to canonical form when trial and error can determine k?
Example

8.11.

Place the poles of the system of Problem 7.8 at Pj, p^ *nd p^.

We

calculate

det

This

is

^^

(fei3

It

+ ^21 + k^i + ^23 + 5)X2 + [fcii + 2fci3 + 3^21 + 4fe22 + S/cga + fei3(fe2i + ^22) " ^izihi + ^12) +
k^i fci3 - 2^21 4^22 - 6^23 - fcll(^22 + ^23) + ^12(^21 + ^23) + ^13(^21 "" ^22) - 4

would take much

trial

and error

to choose the

(\ - Pi)(\ - p^{\ - Pa) =

Trial and error

X3

(pi

fc's

to

8]X

match

+ P2 + P3)X2 +

(piP2

+ P2P3 + PiPs)?^ -

P1P2P3

usually no good, because the algebra is nonlinear and increases greatly


with the order of the system. Also, Theorem 7.7 tells us when it is possible to "place the,
is

namely when Q(i) has rank n everywhere. Transformation to canonical form seems
the best method, as it can be programmed on a computer.
State feedback pole placement has a number of possible defects: (1) The solution appears
after transformation to canonical form, with no opportunity for obtaining an engineering
feeling for the system. (2) The compensation is in the feedback loop, and experience has
shown that cascade compensation is usually better. (3) All the state variables must be
available for measurement.
(4) The closed loop system may be quite sensitive to small
variation in plant parameters. Despite these defects state feedback pole placement may
lead to a very good system. Furthermore, it can be used for very high-order and/or timepoles",

varying systems for which any compensation may be quite difficult to find.
is to try it and then test the system, especially for sensitivity.

Perhaps the

best approach

Example 8.12.
Suppose that the system of Example 8.10 had t e instead of f in the lower right hand corner of the
A(t) matrix, where e is a small positive constant.
Then the closed loop system has a characteristic equation
X^ + 3x2 + 2X * = 0, which has an unstable root.
Therefore this system is extremely sensitive.
8.7

OBSERVER SYSTEMS

Often we need to know the state of a system, and we can measure only the output of the
system. There are many practical situations in which knowledge of the state vector is
required, but only a linear combination of its elements is known. Knowledge of the state,
not the output, determines the future output if the future input is known. Conversely

knowledge of the present state and

its derivative can be used in conjunction with the state


equation to determine the present input. Furthermore, if the state can be reconstructed
from the output, state feedback pole placement could be used in a system in which only the
output is available for measurement.
In a noise-free environment, n observable states can be reconstructed by differentiating
a single output n~l times (see Section 10.6). In a noisy environment, the optimal reconstruction of the state from the output of a linear system is given by the Kalman-Bucy filter.
A discussion of this is beyond the scope of this book. In this section, we discuss an observer
system that can be used in a noisy environment because it does not contain differentiators.
However, in general it does not reconstruct the state in an optimal manner.

RELATIONS WITH CLASSICAL TECHNIQUES

174

To reconstruct

We

all

observable.

is totally

the states at

For

all

times,

we assume

[CHAP. 8

the physical system to be observed

simplicity, at first only single-output systems will be considered.

wish to estimate the state of dx/dt = A(i)x + B(i)u, where the output y = ct(f)x. The
denoted x(i;), and here we denote the estimate of the state as x(i).

state, as usual, is

First, consider

an observer system of dimension


dx/dt

A{t)x

The observer system

n.

+ k(t)[cHt)x-y]+B{t)u

is

constructed as
(8.3)

where k{t) is an w-vector to be chosen. Then the observer system can be incorporated into
the flow diagram as shown in Fig. 8-13.

Fig. 8-13

Since the initial state x(fo), where U is the time the observer system is started, is not
known, we choose x{to) =0. Then we can investigate the conditions under which x{t) tends
Then
Define the error e(Q = x()-x(t).
to x{t).
de/dt

dx/dt

-dx/dt =

[A()

+ k()ct(i)]e

{84)

Similar to the method of the previous Section 8.6, k(<) can be chosen to "place the poles"
of the error equation (84). By duality, the (Josed loop transition matrix *(t, ) of the
adjoint equation d^/dt= -At(t)p - c(t)i; is shaped using -u = kt(t)p. Then the transition
matrix *(i,g of equation (84) is found as *(,,) = *+(*,<), from equation {5M). For
time-invariant systems, it is simpler to consider dw/dt = Atw + cv rather than the
This is because the matrix At + ckt and the matrix A + kc+ have the same
adjoint.
At + ckt, its comeigenvalues. This is easily proved by noting that if A is an eigenvalue of
A-t - ckt) = 0.
plex conjugate \* is also. Then \* satisfies the characteristic equation det (X*I Taking the complex conjugate of this equation and realizing the determinant is invariant

under matrix transposition completes the proof. Hence the poles of equations (8.3) and (84)
can be placed where desired. Consequently the error e{t) can be made to decay as quickly
system,.
as desired, and the state of the observer system tends to the state of the physical
to zero
error
tend
the
make
to
want
not
do
we
8.3,
Problem
However, as is indicated in
too quickly in a practical system.

Example

8.13.

Given the physical system


y
dt

-2/

Construct an observer system such that the error decays with poles at
First

we transform

the hypothetical system

dw
dt

/-2

(1

l)x

0,

2\

w +
.

/I'

-2 and

-3.

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

to the phase variable canonical

form
dz

^-2 -4

It
'4

where

teristic

equation

k = (1

obtained by Theorem

*.

(X

+ 2)(\ + 3) =

X^

and the observer system

0)t

175

+
is

7.7.

We

]v

desire the closed loop system to have the charac-

5\ + 6. Therefore
constructed as

choose

(-4 -l)z

Then

(-1 0)w.

Now we consider an observer system of dimension less than n. In the case of a singleoutput system we only need to estimate 1 elements of the state vector because the known
output and the n 1 estimated elements will usually give an estimate of the nth element of
the state vector. In general for a system having k independent outputs we shall construct
an observer system of dimension n k.

We

choose P(i) to be certain n

{H(t)

diflferentiable

exists at all times

G(t))

where

is

rows such that the

has n k columns.

nxn

matrix

The estimate x

constructed as
x.{t)

Analogous to equation

H{t)w

we

(8.3),

w
G{t)y

or, equivalently,

(8.5)

require

Pdx/d = P[Ax + L(Cx-y)+Bu]


where
This

nxk

an
matrix to be found. (It turns out we only need to find PL, not L.)
equivalent to constructing the following system to generate w, from equation
(8.5),

L{t) is

is

dw/dt

where F

determined from

F and PL

so that
(8.6)

is

the error

(dP/di)x

FP =

+ Pdx/d = Fw - PLy + PBu

dP/dt

+ PA + PLC. Then

are determined from

= P(x-x) = Px- w

The flow diagram

is

(F -PL)j
|

(F|-PL) = {dP/dt + PA){H\G).


de/dt^Fe.

obeys the equation

then as shown in Fig. 8-14.

Fig. 8-14

Example

8.14.

Given the system of Example


Since

C =

(1

1),

cj

choose

8.13, construct

P =

[l

(pi

Pg)

ij

a first-order observer system.

with

pi = pg-

Then

P1-P2V-I

(H G)
I

Pi

{8.6)

dP/dt

From

+ PA

(8.5)

and

RELATIONS WITH CLASSICAL TECHNIQUES

176

[CHAP.

Therefore

Pi-P2\-1/

-P2

w +

P1-P2

Pi

and

PL

=
=

-{dr/dt

(Pi

+ PA)H

(dP/dt

+ PA)G

/-2

-P2
P2

Pi

1\/-P2

Pi

-3pt

-2yv-i

P2) /'-2

(Pi

Pi
SO that (pi - P2)dw/dt = (-3pi
choice of Pi/p2 with 1 < pj/pg

P2)

P1-P2

4p2
P2
2p,^

P?
Pi

-P2

+ 4p2)w - (Pi - 2p|)j/ - pi(pi - p2)u is the first-order observer. A


< 4/3 gives an unstable observer and makes the error blow up.

bad

The question is, can we place the poles of F by proper selection of P in a manner similar
to that of the w-dimensional observer? One method is to use trial and error, which is sometimes more rapid for low-order, time-invariant systems. However, to show that the poles
of F can be placed arbitrarily, we use the transformation x = Tz to obtain the canonical

form
All
A21

+ T

dt

'Bu

\Zl

1-2

ct

where the subsystem


form

dzi/dt

= Aun + Biu and

Vi

ct

-n

is in

the dual phase variable canon-

ical

'0

ai()

a2(i)
Z:i

,0

which Bi

is

defined

from T-^B =

BiU

1, 2,

a;(i)/

l)zi

2/i

in

and

Mi is

2.

(8.7)

the dimension of the ith subsystem.

As per

the remarks following Example 8.10, the conditions under which this form can
always be obtained are not known at present for the time-varying case, and an algorithm

is

not available for the time-invariant multiple-output case.

However, assuming the subsystem {8.7) can be obtained, we construct the observer
equation (8.6) for the subsystem (8.7) by the choice of Pi = (I fc) where ki(^) is an (m - 1)vector that will set the poles of the observer. We assume ki(i) is differentiable. Then
|

Pi

ki

-ki

and

(Hi Gi)
i

(8.8)

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

We find

Fi

{dFi/dt

+ PiA)Hi =

[(0

dki/dt)

(I

...
1
Fi

...

ki2{t)
A;i3(^)

By matching coefl!icients of the "characteristic


we make the error decay as quickly as desired.
we

PiL;

x = Tz = T(

Then

Example

find FjLj as

where

fci,ni-l(*)/

equation" with "desired pole positions,"

z.-

H.-Wi

+ Gi^/i =

)w.-

]yi

8.15.

-2

To construct an observer system with a


1\

Then equations

(8.7)

8.13.

-2
dt

= -{dPJdt + PiAii)Gi.

Again, consider the system of Example

'4

from which

kii{t)

...

Also,

kOA](

177

2/

(1

l)x

pole at -2, use the transformation

Tz

where

(Tt)-i

are

dz

J:4>

df

The estimate

z according to equation {8.8) is

+ (:r'"
now

2/

system

is

= -2

sets the pole of the observer at 2.

dw/dt

= 2w + 2y 2u.

PT

ix

w =

2xi

l)z

Z2

Then

F = 2

2 \

/w

and

PL = -2

so that the observer

Therefore

Tz
and the error

(0

obtained as
1

where k^

+ X2 w

-3\/l

-1

4JU

2\/w
l)\y

decays with a time constant of 1/2.

of Fig. 8-15.

Fig. 8-15

This gives the block diagram

RELATIONS WITH CLASSICAL TECHNIQUES

178

[CHAP.

ALGEBRAIC SEPARATION

8.8

In this section

we

use the observer system

to place the closed loop poles

where

Section 8.7 to generate a feedback control

oi'

desired, as discussed in Section 8.6.

Specifically,

we

consider the physical open loop system

with an observer system

A(f)x

C(i)x

B()u

J(i)d
(8.9)

(see equation (8.6)),

dw/dt

F(t)w-F{t)L.{t)y

H(i)w

and a feedback control

tem

dx/di

u(i)

F{t)B{t)u

V{t)J{t)d

G(i)y

{8.10)

that has been formed to place the poles of the closed loop sys-

as

u = W(f)x

Then the

closed loop system block

diagram

is

{8.11)

as in Fig. 8-16.

^^>^:r0t^

d=i4;

p0if(S'

-*U- Feedback -

Observer

Physical system
Fig. 8-16

Theorem

8.3:

(Algebraic Separation). For the system [8.9) with observer {8.10) and feedback control {8.11), the characteristic equation of the closed loop system can
be factored as det (Al - A - BW) det (Al - F).

using the pole placeThis means we can set the poles of the closed loop system by choosing
of Section 8.7.
techniques
using
the
choosing
P
ment techniques of Section 8.6 and by

The equations governing the close(3 loop system are obtained by substituting
Proof:
equation {8.11) into equations {8.9) and {8.10):
_d

/x

dt[w
Changing variables
gives
_d

/x

dt\e

to

J
BWH \/x\
A + BWGC
PBWGC-PLC F + PBWhAw/ VPJ/
e = Px - w and using HP + GC = I and FP = dP/dt + PA + PLC
,

A + BW

-BWH\
F

/x\
e

/J \

O^'*

(C

0)

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

Note that the bottom equation deldt


of observer system dx/dt = (A + BW)x.

179

= Fe

generates an input -WHe to the closed loop


3.5 then shows the characteristic
equation factors as hypothesized.
Furthermore, the observer dynamics are in general
observable at the output (through coupling with x) but are uncontrollable by d and hence
cancel out of the closed loop transfer function.
Example

Use of Problem

8.16.

For the system of Example 8.13, construct a one-dimensional observer system with a pole at
generate a feedback that places both the system poles at 1.

-2

to

We employ the algebraic separation theorem to separately consider the system pole placement and
the observer pole placement.
To place the pole of

'- ^-;)" - n

dx
'

dt

using the techniques of Section 8.6

-2

we would
u

'^

(1

l)x

like

(-2 3/2)x

which gives closed loop poles at -1. However, we cannot use x to form u, but must
use x as found
from the observer system with a pole at -2, which was constructed in Example 8.15.
dw/dt

We

= -2w +

2y

2(u

+ d)

then form the control as

u
Thus the

(-2 3/2)x

closed loop system

is

(-2 3/2)

-7w/2

-I-

5y

as in Fig. 8-17.

^rO
+

CMS)
Fig. 8-17

Note that the control is still essentially in the feedback loop and that no reasons were given as to
plant poles at -1 and observer pole at -2 were selected.
However, the procedure works for highorder, multiple input multiple output, time-varying systems.

why

8.9

SENSITIVITY, NOISE REJECTION,

AND NONLINEAR EFFECTS

Three major reasons for the use of feedback

control, as opposed to open loop control,


reduce the sensitivity of the response to parameter variations, (2) to reduce the
effect of noise disturbances, and (3) to make the response of nonlinear elements more linear.
A proposed design of a feedback system should be evaluated for sensitivity, noise rejection,
and the effect of nonlinearities. Certainly any system designed using the pole placement
techniques of Sections 8.6, 8.7 and 8.8 must be evaluated in these respects because of the
cookbook nature of pole placement.

are

(1) to

RELATIONS WITH CLASSICAL TECHNIQUES

180

[CHAP.

In this section we consider these topics in a very cursory manner, mainly to show the
Consequently we consider only small
relationship with controllability and observability.
percentage changes in parameter variations, small noise compared with the signal, and

Under

nonlinearities that are almost linear.

effect produces an unwanted input into a

wanted

we will show how each


system and then how to minimize this un-

these assumptions

lineair

input.

refer to the
consider the effect of parameter variations. Let the subscript
system (the
nominal
the
Then
values.
actual
refer
to
nominal values and the subscript a
represented
by
can
variations)
be
system with zero parameter

First

we

d^N/dt
yjv

These equations determine


of the elements of

An

=
=

+ B{t)u
+ D(t)u

AN(t)xN
C(t)xN

{8.12)

and yN(<), so these quantities are assumed known. If some


some actual Aa (keeping B, C and D fixed only for simplicity),

XN(i)

drift to

*^"

=
=

dXa/dt
y

+ B(t)VL
+ D(f)u

Aa{t)Xa
C(<)xa

{8.13)

Then let 8x = Xa-XN, 8A = Aa-As, 8y = ya-yN, subtract equations {8.12) from {8.13),
and neglect the product of small quantities .SA 8x. Warning: That 8A 8x is truly small
at all times must be verified by simulation. If this is so, then
d(8x)/dt

8y

=
=

AnC*:)

8x

8A(t)

xn
{8.1^)

C{t)S%

the parameters
In these equations AN{t), C{t) and xn() are known and SA{t), the variation of
of the A{t) matrix, is the input to drive the unwanted signal 8x.
but
For the case of noise disturbances d{t), the nominal system remains equations {8.12)

the actual system

,,,.
= An()xo + B(i)u + J{t)d

is

/+\

dxa/dt
y

Then we subtract equations

{8.12)

from

8y
A{t)

drives the

rj/+\

/+\^

C{t)xa-D{t)u + K{t)d

{8.15)

{8.15) to obtain

d{hx)ldt

Here the noise

=
=

+ J(i)d
+ K(i)d

Aw{t) 8x

C(*)8x

unwanted signal

(-^^)

8x.

a scalar input
show how a nonlinearity produces an unwanted signal, consider
system
Imear
a
can be redrawn into
x^ into a nonlinearity as shown in Fig. 8-18. This
small output (Fig. 8-19).
with a large output and a nonlinear system with a
Finally, to

Fig. 8-18

Fig. 8-19

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

181

Here the unwanted signal is 8d which is generated by the nominal xn. This can be
incorporated into a block diagram containing linear elements, and the effect of the nonlinearity can be evaluated in a manner similar to that used in deriving equations
(8.16).
d(8x)/d

An() 8x

8y

C(^) 8x

j{t)

+ k{t)

Sd

Sd

(8.17)

Now

observability and controllability theory can be applied to equations


{8.U), (8.16)
We conclude that, if possible, we will choose C{t), An() and the corresponding
input matrices B{t), D{t) or J{t), K{t) such that the unwanted signal is unobservable
with
respect to the output Sy(t), or at least the elements of the state vector associated
with
the dominant poles are uncontrollable with respect to the unwanted signal. If this is impossible, the system gain with respect to the unwanted signal should be made
as low as

and

(8.17).

possible.

Example

8.17.

Consider the system

and the nominal input

dx/dt

_Mx + L jw.

Xj,(0)

equation

0,

(S.li).

then

Xf^it)

is

zero

a unit step function.

u(t) is

dxi^/dt

If

The nominal value of the parameter a

\^

*.,/

The

-1 )'''*'"*' (l
effect of small variations in a

can be evaluated from

(-i_:uu:

:)(;::::

Simplifying,
d(Sx)/dt

We

_^sis.

can eliminate the effects of a variation upon the output if


matrix (ctb ctAb) = 0.
This results in a choice ct =

f'^

c is

)a{l-e-f^
chosen such that the output observability
where y is any number.

(0 y)

Furthermore, all the analysis and synthesis techniques developed in this chapter can be
used to analyze and design systems to reduce sensitivity and nonlinear effects and to reject
noise.
This may be done by using the error constant matrix table, root locus, Nyquist,
and/or pole placement techniques on the equations {8.14^), (8.16) and (8.17).

RELATIONS WITH CLASSICAL TECHNIQUES

182

[CHAP.

Solved Problems
/
8.1.

dx

For the system of Fig.


with y
The

Find the

l)x.

{1

=
-^
dt

8-4, let

\t

+ 1/

\t

class of inputs that give a constant steady state error.

closed loop system with zero input (d(t)

= 0)

is

dx
1

dt
,t

t_) -(il)(l
\t + l/
+ l/

1)

-1

-2

'^

which has a double pole at -1. Therefore the zero output of the closed loop system is asympand the remarks following Theorem 8.1 are valid. The class of all d(t) that the
is given by
system can follow with lim e(t) =

totically stable

Vt+t
d

with

(1

l)w

g.

The

\ + i/

+ i/

transition matrix for this system


1

ft

'-*

is

t(l-eT-t)

Then

(Notice this system

1)'

+ w^{to)] + (t+l)j'-^^ff(r)dr +

ll[n,,(to)

d(t)

is

JC
to

For constant error

unobservable.)

dr
r^l
+ 1/

-c(t

l)[ln (

1)

let

g{t)

In (to

g{t)

an arbitrary constant.

k,

+ 1) +

(*o

+ l)-i "

Then

+ l)-i]

(t

Since the system is


to this.
and the system follows with error k all functions that are asymptotic
will follow all functions gomg to infinity
system
the
that
assume
can
we
behaved,
well
reasonably
slower than K(t + 1) In (t 4 1).

8.2.

Given the multiple-input system


'

0\fzi

-1

-3^
Place the poles of the closed loop system at -4,
Transform the system

-5 and -6.

to phase variable canonical


/

form using the

o\

1
1
i\ Ai
-1 -2 -3 jl
1

9/

results of

jz

-C2

\0

/eg/

'^*'^"

-i

\ /
K-i

V
-crVV

5/2

1/2^

-3

-4

-1 )x

3/2

1/2,

Problem

7.21.

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

183

and

\-6 -11

-6/

To obtain phase variable canonical form for

we

u^,

set

/a/cA
"2

"3/

which gives

For a

ai
0,

1/2,

k^

we have

-1,

Kg

1/2,

the phase variable canonical system

d^=0
\-6

lx+0-2-11

-6/

\l

To have the closed loop poles at -4, -5 and -6 we


Therefore we choose u^ - -114xi - GSajg - dx^

l/2a

l/2a

desire a charateristic polynomial X3 + 15X2 + 74^ 4- 120.


and M2 = Oaji + Oaia + 0x3.

In the case a = 0, the state ^i is uncontrollable with respect to Mj and, from Theorem
7.7,
cannot be put into phase variable canonical form with respect to Mj alone.
Hence we must use'
4*2 to control Zi, and can assure a pole at -4 by choosing Mg =
-3i. Then we have the single-input
system

AA

/-4

\^3/

0-3

whose controllable subsystem

/-2

d y^aX

/I

0\ /^2^

ur^

-s){zj ^

""

du^sy

can be transformed to phase variable canonical form, and

The above procedure can be generalized

to

give a

u^

-12z2

means

+ 623.

of obtaining multiple-input pole

placement.

8.3.

Given the system (Py/dt^


pole at -y.
y{t)

+ ri{t),

0.

Construct the observer system such that it has a double


e = x - x as a function of y, if the output really is

Then find the error


where ?;(*) is noise.

The observer system has the form


dx

df

_
-

/O
[0

l^>^

oy^ +

/fei

UJf^^")'^-^-"]
^2

The characteristic equation

of the closed loop system is X(X - fej)


the characteristic equation X^ + 2yX -h y2 = 0.
Hence set k^
equation for the error is

di[ej

"

The transfer function

is

y -^ ,

then
may be large even
clude that it is not
/ion
can fil4-y\irt
Alter .-.tt4out HAin^n
some

and

double pole at -y has


fca = -y2.
Then the

from reaching

zero.

found from

\^2/j

As

it

0.

o)[eJ + (-y2

(-y2

Note the noise drives the error and prevents

- feg =
= -2y

s2

-I-

2ys

y2

\^

y2s

Ci^r, and e^-^ d-qidt.


If >;()=: ijo cos ui, then utjo, the amplitude of d7,/d<,
though ij^ is small, because the noise may be of very high frequency. We cona good idea to set the observer system gains too high, so that the observer system
^-P

;
of 4-1.
the noise.

RELATIONS WITH CLASSICAL TECHNIQUES

184

8.4.

[CHAP.

Given the discrete-time system


x(n

+ l)

(_2

J)x()

y{n)

(Jju{n),

O)x(n)

(1

Design a feedback controller given dominant closed loop poles in the

plane at

(iy)/2.

We

can conan observer to generate the estimate of the state from which we

shall construct

The desired clos(3d loop characteristic equation is X^ - X + 1/2 - 0.


To generate $i and x^ we choose a first-order observer with a
large
hardly aflfect the response due to the dominant poles and yet will be

struct the desired control.

Hence we choose u
pole at 0.05 so that

enough

z(n

gives

ly

of variables

The transformation

l)

vM

(i "3)^(")+(o)'*<'^^'

{0

IHn)

of equation (8.8) gives

P =

(1

Then the observer

and w{n)

is

Find the

-0.05)

and

/O

1\.
gJZ

PL

-(1 -0.05)(^;

''f){'T)

'''''

is

8.5.

will

it

to filter high-frequency noise.

Use

= SxJ2 - 2^^.

(^

w{n +1)

obtained from

^_
z(n)

where

-0.05w(m)

/1\

...
i^jMn) +

^0.05

{\

2'(')

- 2.1525j/() + u(n).

sensitivity of the poles of the system

/I +

dx.

"

dt
to changes in the

parameter a where

-2/'"

small.

is

\a\

a^

A^

=
and the nominal system as dx^/dt = A^k^.
denote the actual system as dxjdt
where ||6Al| is small smce \a\ is small.
=
A-8A
=
A^
general,
,.
In
0.
wherr A^ = A when
find a corresponding eigenvector
We assume A^ has distinct eigenvalues xf so that we can always
Note the
that Aj;v; = X^ vj
so
of A^ as vf,
w. from A^w, = xf w^. Denote the eigenvectors
gives
transpose
the
Taking
and
A^
eigenvalues Xf are the same for

We

A^

v+A,
which we

shall

need

Next we

later.

let the actual eigenvalues

into the eigenvalue equation for the actual


(A.v

Subtracting

SA)(Wi

A w, +

vf

A^

SWi

X?

xf

6X,

Substituting this

gives

+ SWi) =

(xf

A^Wj = xf Wj and multiplying by v|


v+

(^^)

Xfvt

SX, vt

Neglecting the last quantity on the right since

w,

{8-19)

8Xi)(wj+SWj)

gives

it is

xf vt 8w,

vt

of second order

(8X, I

8 A)

8w,

and using equation

(8.18)

then

leads to
O Ai

Therefore for the particular system in

/-I +
Aa

a^

a\

-2J

_
"

v]

8AWi

(5.20)

ciuestion,

0\

/a2

V 2 -2^

\0

/--I

\,

^^ + g^

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

Then for A^ we

find

= -1

w,

Xf = -2

w,

Xf

Using equation

185

-2

(8.20),

a\/l
0/\2

a^

-1 +

^2

-2 + (-2

(1

0)

1)

=
=

0/Vl

-1 +

a^

-2

2a

2a

For larger values of a note we can use root locus under parameter variation to obtain exact values.
However, the root locus is difficult computationally for very high-order systems, whereas
the
procedure just described has been applied to a 51st-order system.

8.6.

Given the scalar nonlinear system of Fig. 8-20 with input a sin t. Should
be increased or decreased to minimize the effect of the nonlinearity?
a sin

-\e{t)
/~\

_L 3

sm

^ew(<)

'

Vn

'

if

K>1

~~ *^N

Fig. 8-20

Fig. 8-21

The nominal Unear system is shown in Fig. 8-21. The steady state value
of Cj, = (a sin f)/(X - 1)
approximate this as the input to the unwanted signal dHSy)/dt^ =
which gives the steady
4,
state value of Sy = 3(27 sin t - sin 3t)/[36(K - 1)3].
This approximation that dHSy)/df^ = el
instead of e must be verified by simulation.
It turns out this a good approximation for LI < 1
and we can conclude that for \a\ < 1 we increase
to make Sy/y become smaller.

We

8.7.

simplified, normalized representation of the control system of


a solid-core nuclear
is shown in Fig. 8-22 where Sn, ST, SP, Sp^^ and
87 are the changes from

rocket engine

nommal

in neutron density, core temperature, core hydrogen


pressure, control rod
and turbine power control valve setting, respectively. Also, G^{s), G^{s) and
Gg{s) are scalar transfer functions and in the compensation Kj,
k^, k^ and k^ are scalar
constants so that the control is proportional plus integral. Find a simple
means of
improving response.
setting,

J)

4-

Gi()

^
J

8T
Giis)

dpcr

sv

"1

+ "2/*

_!_../_

"3

Fig. 8-22

SP
63(8)

RELATIONS WITH CLASSICAL TECHNIQUES

186

[CHAP.

The main point is to realize this "multiple loop" system is really a multiple input- multiple
The control system shown has constrained 8pcr to be an inner loop. A means
output system.
This opens
8-23.
of improving response is to rewrite- the system in matrix form as shown in Fig.
up the possibility of "cross coupling" the feedback, such as having Spc, depend on SP as well as
input with zero
ST. Furthermore it is evident that the system is of type 1 and can follow a step
steady state error.

S^demandN

-r<>-

"

(")

C?)

Ki+^K2

G(s)

Fig. 8-23

8.8.

with an
Compensate the system {s + p,)-'i^ + V,)-' ^o have poles at -u, and -^,
effect
the
discuss
theorem, and
observer pole at -^ by using the algebraic separation
of noise

rj

at the output.

state space representation of the plant is

The

J/

0)x

(1

as
The feedback compensation can be found immediately

Tri^2

{PlP2

To construct the observer system,

let

P=

Pl

+ P2~n~'^2)x

aa)-

(ai

Then
1

1
I

(i)

from which

^^

Also

(-.oi-PL)f7

-PL)

(F

PB = PJ =

a2-

(pi

+ p2-^oK

(i2)

:f)

\-P1P2

-PL = MPi + P2 - '^o) -

and

"

-~Pi

=
P2

PiP2]2

Therefore the estimator dynamics are

A
f
dt \a2
To construct the

-^al) +

+ P2-^o)-PlP2]y + U +

Wo(Pl

estimate,

^P\-i/wN

-i

')
2y

The flow diagram of the closed loop system

is

shown

TTQ

Pi

^/w
- P2/\V

in Fig. 8-24.

+1
(s

w/ao
Pl

+ P2~''l~'^2

jro

P1P2

'^i^z

xi

'

1+
U C)*~

- Pi - P2

1/

Fig. 8-24

-I-

Pi)(s

'^0*Pl

-I-

P2)

+ P2 - To) -

P1P2

PA

CHAP.

RELATIONS WITH CLASSICAL TECHNIQUES

8j

187

Note the noise v is fed through a first-order system and gain elements to form u. If the noise
level is high, it is better to use a second-order observer or a Kalman filter because then
the noise
goes through no gain elements directly to the control, but instead is processed through first- and
second-order systems. If there is no noise whatsoever, flow diagram manipulations can be used to
show the closed loop system is equivalent to one compensated by a lead network, i.e. the above flow
diagram with
=
can be rearranged as in Fig. 8-25.
ij

1
(S

-I-

+ Pi){s + P2)

o/ 4-

^2

fl

*.

F2I

^
_i

^Vr2

-PlV2

Fig. 8-25

have cancelled out, and the closed loop system remains


secondorder %il^carZZL^T:^'""
corresponds to the conclusion the observer dynamics
are uncontrollable by d. However
anv im-nl?
'"
"" '^" ^''^'"'^ ^" '^''' " ^' ^' '^"^ " ^^^'^ f^"*" *e first flow

diagram

Supplementary Problems
8.9.

Given the matrix block diagram of Fig. 8-26.


mdicated inverse exists.

Show

that this reduces to Fig. 8-27

when

the

?Q=^
G{H-HG)-i

>

H
Fig. 8-26

8.10.

Given the matrix block


single feedback loop.

diagram of Fig.

Fig. 8-27

8-28.

Reduce the block diagram

=<

G3
I

+
H.

to obtain H, isolated

'

t"

H2

Fig. 8-28

Cz

RELATIONS WITH CLASSICAL TECHNIQUES

188

8.11.-

Determine whether the scalar feedback system of Fig.


which y{t) is related to e(t) as
\

-4-iy

'dt

d-

,8r29 in

[CHAP.

>

\2t'!^

0)x

(1

error,
(i) can follow a step input a with zero
follow a ramp input at with zero error.

(ii)

can
Fig. 8-29

8.12.

Fmd a feedback control u such


Given the time-varying system diy/dt^+ a(t) dy/dt + p(t)y = u.
<p{t)z
- 0.
dz/dt
d^z/dt^
e{t)
+
like
+
behaves
that the closed loop system

8.13.

Given the system

/l

-2\

0-1

Construct a third-order observer system with poles at

8.14.

J/

0,

-l)x

(0

-1 and -1.

Given the system

dx

-1

-1.
Construct a first-order observer system with a pole at

8.15.

Given the system

-:>-{>

dx

2/

0)x

(1

at -3 and then find a feedback control


Construct a first-order observer system with a pole
at -2.
that places both the closed loop system poles
^j = fej$j + fc22

8.16.

Given the system

dx
d/T

"

/3

^1

3,

2/

l)x

(1

-4.
Construct a first-order observer system with a pole at

t^e/orm

8.17.

What is
y = C(t)x + D(t).
Given the system dx/dt = A()x + BW where
theorem when D(t)
separation
algebraic
the
is
What
0?
oblerver system when D(t) #

8.18.

Given the system

dx

and

4^1

dt

(0

of

the

0?

l)x

VO

3-|-4a(t)y

8.19.

B(t),C,(t),D{t) andu(t)?

8.20.

Given the system

,,

di!.ldt

+ *+
_
- P
y
f[t)

-^^*^

"/fV.
^"
fi
/

variations in Rt),
will be insensitive to small

8.21.

gwen

Choose

(t)

such that at least one state

the nominal solution x^(t).

the scalar system


Given that the input d(t) is generated by
dx/dt - a(t)x
system
the
can
p(t)
on
conditions
under what
d(t) with zero error?

d^.m
+ I3{t)e

=^ ^^^J^
with

fWx TolLw

atfsS

CHAP.

8.22.

RELATIONS WITH CLASSICAL TECHNIQUES

8]

189

Given the general nonunity feedback time-invariant system of Fig. 8-30.


Under what conditions
can lim e(t) = 0?
Set F = I and H = I and derive Theorem 8.2 from these conditions.

<{y<i}

F(s)

+
^{y}

=Q

^{d}:

<>=-

G(8)

^{e}

H(8)

Fig. 8-30

8.23.

In the proof of

Theorem

d(<)

equivalent to
8.24.

Show

that

that

8.25.

if

dvi/dt

J
=

why

8.1,

is

C(t)*A_Bc(,T)B(r)d(r)dT

g(i)

[A)

- B(t)

C(<)]w

a time-invariant system

is

N.

+ B(t)d

and

of type N, then

it is

Given the system of Fig. 8-31 where the constant matrix

Show

C(t)*A-Bc(,o)w(o)

to

C(t)w

+ g?

of type

N-k

for all integers k such

has been introduced as compensation.

is

that

(a)
(6)

The type number of the system cannot change


The system is of type zero if K is singular.
d(f)

=Q

if

nonsingular.

H(s)

Fig. 8-31

8.26.

Design a system that will follow

Answers
8.10.

to

d(t)

sin

with zero steady state error.

Supplementary Problems

This cannot be done unless the indicated inverse exists (Fig. 8-32). The matrices must be in the
order given.

^^0=^

>

G4Gi[I-hH2(G2-^G3)G4G,]-i

+
H,

Fig. 8-32

G2 + G3

I>

RELATIONS WITH CLASSICAL TECHNIQUES

190

8.11.

The closed loop system output


Then
Theorem 8.1 applies.

d{t)

ramps

(t

S'(t)

At

a;i(to)

+ J \-M']^2(to) +

Wi(<o)

so that the system can follow steps but not


8.12.

[CHAP.

*o)]
*o) ^2(*o)

In

tends to zero, so

for

<

> some

ti

-with zero error,

For the state equation

-l(e> + (i'

-lit)

'dt

corresponding to the given input-output relation,

iji

<t>

9)x

which shows the basic idea behind "pole placement,"


cPy/df^
1
l-2\
0-1 1

dx
8.13.

dw/dt

/By

(a

- e)

dy/dt

(^

<

$ +

-l)x -

[(0

y]

1,

.1

8.14.

a dy/dt

/1\

It

= -w +

(0

'OV
1 jw +

l)y,

/O

1-

)y

Vo

1/

0,

8.15.

Fig. 8-33
8.16.

This cannot be done because the system

8.17.

Subtract D(f)u from y entering the observer and this reduces to the given formulation.

8.18.

U:{Sym = -

8.20.

..

,'-;;"

Q(t)

8.21.

pit)

such that

u{t)

N1

that lim

e(*)

e{t)

-G(s)(I

8.22.

lim

8.24.

Use Theorem

8.2.

8.25.

Use Theorem

8.2.

8.26.

One answer

[F(s)

det Q(t)

a(T))dn

-"to

is

and

+
k is

dXfji/dt

7(*)e-'T" "

(1

dx^i/dt

for all

+ l)(s + l)

"

+ t + /(t))Nl ~ u{t)xi^
f(t)Xf^i fiXffi

t.

pir)e{r)dT

~/it)K

where

an arbitrary constant,

+ G(s)H(s))-i]s^{d} =

H(8)

-^{"(*)>

4(s

:")".(::;,./
^Xf^i

Choose

^Mt)} =

where

:^^r^!!r"f?.
l)(s
1)

(s

unobservable.

is

8(82+ i)-i^

and

is

analytic for s

0.

9(J)

is

any function such

chapter 9

Stability of Linear

Systems

INTRODUCTION

9.1

Historically, the concept of stability has been of great importance to the system designer.
of stability seems simple enough for linear time-invariant systems. However,

The concept

we

shall find that its extension to nonlinear

and/or time-varying systems

is

quite compli-

For the unforced linear time-invariant system dx/df = Ax, we are assured that
the solution x{t) = TeJ^-'o^T-ixo does not blow up if all the eigenvalues of A are in the left
half of the complex plane.
cated.

Other than transformation to Jordan form, there are many direct tests for stability.
criterion can be applied to the characteristic polynomial of A as a yes-

The Routh-Hurwitz

or-no test for the existence of poles in the right half plane. More useful techniques are
those of root locus, Nyquist, Bode, etc., which indicate the degree of stability in some sense.
These are still the best techniques for the analysis of low-order time-invariant systems, and

we have

seen

how

Now we

chapter.

them to multiple input-multiple output systems in the previous


wish to extend the idea of stability to time-varying systems, and to do
examine carefully what is meant by stability for this type of system.
to apply

this

we must

9.2

DEFINITIONS OF STABILITY FOR ZERO -INPUT LINEAR SYSTEMS

first

type of "stability" results

For every xo and every

Definition 9,1:

to

Even

we can say

if

such that

i|x(i)||

^k

to,

if

for

the response

is

bounded.

there exists a constant k depending on xo and


t ^ to,
then the response x(i) is bounded.

all

this simple definition has difficulties.

The trouble is that we must specify the response


The trajectory of x(i) = ^(^; u{T),x{to), to) depends implicitly on three quantities:
u(r), x(io) and to.
By considering only u(t) =
in this section, we have eliminated one
difficulty for a while.
But can the boundedness of the response depend on x(to) ?

to what.

Example

9.1.

Consider the scalar zero-input nonlinear equation dx/dt = -x + x^ with initial condition
x{0) = Xg.
The solution to the linearized equation dx/dt = -x is obviously bounded for all Xo- However, the solution
to the nonlinear equation is
X{t)

Xo

For

negative values of

all

time

In

a;o

In

(a;o

cbq,

1),

lim

a;(t)

- Xo)

For values

this is well behaved.

so that

e(l

of

Xo

>

1,

the denominator vanishes at a

t-+t,

It

can be concluded that boundedness depends upon the

initial conditions for

nonlinear

equations in general.

Theorem

9.1:

The boundedness of the response


is

independent of the

x(^)

of the linear system

initial condition xo.

191

dx/dt

A(t)x

[CHAP.

STABILITY OP LINEAR SYSTEMS

192

Proof:

Since

[|x(i)|!

a constant,

is

|[xo|i

Now we

]|^(;0,xo,io)|l

||*(i,io)xo]|

becomes unbounded as

if |!x(i)||

||*(t, io)|l lixo||


oo,

i-*

it is

due to

solely

First, note that

shall consider other, different types of stability.

*(*,

x=

*o).

is

steady state solution (an equilibrium state) to tlie zero-input linear system ds.ldt = A(f)x.
We shall define a region of state space by ||x|| < , and see if there exists a small region of
that give rise to a trajectory
nonzero perturbations surrounding the equilibrium state x =

which remains within


have

<

[jx||

If this is true for all

e.

>

0,

how

no matter

small, then

we

dx/dt = A(i)x is stable in the sense of


for any to and every real number e > 0,
Liapunov (for short, stable
there is some 8 > 0, as small as we please, depending on to and e such
that if |]xo|| < 8 then ||x(i)|| < for all t > to.

The equilibrium

Definition 9.2:

x=

state

of

i.s.L.) if

most

It is the

0.
also valid for nonlinear systems with an equilibrium state x
often
is
"stable
i.s.L."
literature
the
in
stability,
and
definition
of
common

This definition

is

States that are not stable i.s.L. will be called unstable. Note stability
is an
Finally, since x =
i.s.L. is a local condition, in that 8 can be as small as we please.
linear
systems
about
speaking
when
system,
obvious choice of equilibrium state for a linear
we shall not be precise but instead will say the sjstem is stable when we mean the zero state

shortened to "stable".

is stable.

Example

9.2,

Consider the nonlinear system of Example

|a;o

In Definition 9.2

we can

set

x-

et(l

>

then

1,

\Xn

kol

(t)|

If

9.1.

- a;o)1

if

"
1,

and

|1

if

(e'

>

- 1)(1 1

we

lol

Xo)!

set

to

show the zero

state of

Hence the zero state is stable i.s.L. even though the response can become
stable i.s.L.
^ >
. ^
Q course if the
(This
situation corresponds 1;o Fig. 9-2(6) of Problem 9.1.)
for
some
unbounded
oto
Another point
_
unstable.
,considered
0, the zero state would be
response became unbounded for all Xg
choice
of
the
results
there
small
where
e
is
i-ange
in
the
Definition
9.2,
of
application
in
the
note is that

Example

9.1

is

a correspondingly small

Example

S.

9.3.

Given the

Van

der Pol equation


X2

{l-Xl)X2

dt

with

initial condition

x(0)

state space can be plotted as


will call the trajectory in

Xj

The trajectories in
shown in Fig. 9-1. We
Xq.

bold the limit cycle.


Trajectories originating outside the limit cycle spiral
in towards it and trajectories originating inside the
Consider a small
limit cycle spiral out towards it.

any radius, centered at the origin but such


Call
that it lies completely within the limit cycle.
will result in a
its radius e and note only Xo =
Therefore
<
trajectory that stays within
the zero state of the Van der Pol equation is unstable
but any trajectory is bounded. (This situation correcircle of

sponds to Fig.

Theorem

9-2(e) of

9.2:

The
all
is

Note

||x(t)i| is

Problem

Fig. 9-1

9.1.)

transition matrix of a linear system is bounded as


f^to if and only if the equilibrium state x =

stable i.s.L.

bounded

if ||*(*,MII is

bounded.

mt,to)\\ <K{to) for

of

dx/dt

A(t)x

CHAP.

If

STABILITY OF LINEAR SYSTEMS

9]

First assume \\<i>[t,U)\\<K{U) where k


Proof:
are given any e > 0, then we can always find

is

we

k(o)S

>

||*(<, io)|| ||xo!|

||#(, fo)xo||

8 ==

From

||x()]|.

193

a constant depending only on U.


tUiU) such that if |!xo|| < S then

Definition 9.2

we

conclude stability

i.s.L.

Next we assume

stability i.s.L. Let us suppose *(i, to) is not bounded, so that there is at
one element $ij(*,io) that becomes large as t tends to >. If ||xoI| < 8 for a nonzero 8,
then the element Xj of xo can be nonzero, which results in a trajectory that eventually leaves

least

any region

by ||x|l < e. This results in an unstable system, so that


a contradiction and conclude that *(i, U) must be bounded.
Taken together. Theorems 9.1 and 9.2 show that boundedness of ([x(t)|| is equivalent to
stability i.s.L. for linear systems, and is independent of xo. When any form of stability is
independent of the size of the initial perturbation xo, we say the stability is global, or speak of
Stability in the large. Therefore another way of stating Theorem 9.1 is to say (local) stability i.s.L. implies global stability for linear systems.
The nonlinear system of Example
9.1 is stable i.s.L. but not globally stable i.s.L.
In practical applications we often desire the response to return eventually to the equilibrium position X =
after a small displacement. This is a stronger requirement than
stability i.s.L. which only demands that the response stay within a region ||x|| < e.
in state space defined

we have reached

The equilibrium

Definition 9.3:

i.s.L.

and

(2)

state

for any

to

x=
is asymptotically stable
and any xo sufficiently close to 0,

if

(1)

x(t)

it

^0

is

stable

as

t -* >.

This definition is also valid for nonlinear systems. It turns out that (1) must be assumed
(2), because there exist pathological systems where x(t) ->
but are not stable i.s.L.

besides

Example

9.4.

Consider the linear harmonic oscillator

*((,0)

_.

. )

with transition matrix

cos

sin

sin

cos

A matrix

has eigenvalues at jV To apply Definition 9.2, ||x(t)||2 - ||*(, to)ll2 IIX0II2 = llxoli2 < e = S since
Therefore the harmonic oscillator is stable i.s.L. However, x() never damps out to 0, so
ll*(*. *o)!l2 = !
the harmonic oscillator is not asymptotically stable.

The

Example

9.5.

The equilibrium state a; =


is asymptotically stable for the system of Example
perturbation (a;o < 1) gives rise to a trajectory that eventually returns to 0.

In all cases except one,


then the adjective uniform
Example

9.1,

since

any small

if

the conditions of the type of stability are independent of

is

added to the descriptive phrase.

to,

9.6.

If 8 does not

depend on

tg in

Definition 9.2,

Example 9,7.
The stability of any time-invariant system

The exception

we have uniform
is

stability i.s.L.

uniform.

to the usual rule of

adding "uniform" to the descriptive phrase results


because here we only consider linear systems. In the general framework of stability definitions for time-varying nonlinear systems, there is no inconsistency. To avoid the complexities of general nonlinear systems, here we give
Definition 9.4:

If the linear
to

system dx/dt

and for any

x(t)

-^0 as t-^

fixed p
ao,

A(t)x

is

however large

then the system

is

i.s.L. and if for all


rise to a response
gives
p
uniformly asymptotically stable.

uniformly stable
||xo||

<

The difference between Definitions 9.3 and 9.4 is that the conditions of 9.4 do not depend
on to, and additionally must hold for all p. If p could be as small as we please, this would
be analogous to Definition 9.3. This complication arises only because of the linearity, which
in turn implies that Definition 9.4

is

also global.

STABILITY OF LINEAR SYSTEMS

194

Theorem

The

9JS:

linear system ds./dt

if there exist two


^jg-Kact-to' for all t
to

only

Example

is

given in Problem

is

uniformly asymptotically stable if and


||*(t, to)||
ki and k2 such that

positive constants

The proof

A{t)x

[CHAP.

and

all

U.

9.2.

9.8.

Given the linear time-varying scalar system dx/dt = x/t. This has a transition matrix *(, to) = tjt.
as
the system is asymptotically stable. However, the response does not tend to
For initial times to >
the system is unstable, and the asymptotic stability is
fast as an exponential. This is because for to <
not uniform.

Example

9.9.

Any time-invariant linear system dx/dt


eigenvalues of A have negative real parts.

= Ax

is

However for the time-varying system


with negative real parts for each fixed
totically stable or even stable.
Example

t,

unii'ormly asymptotically stable

d-x.ldt

A(i)x,

has

if A(i)

this in general does not

mean

if

and only

all

its

if all

the

eigenvalues

the system

is

asymp-

9.10.

Given the system

(x\.\

_
"

dt\xj

-3<ce8'<t\/xi\

4k

j\x^j

\Ke-i>'t

The eigenvalues are Xi = k and Xg


with initial conditions x(0) = Xq.
eigenvalues have real parts less than zero. However, the exact solution is
2a;i(t)

For any nonzero

SCxio

+ a;2o)^''' -

(^'lo

+ 3a;2o)''''

and

^x^it)

(xio

3/c.

+ SiCao)*"""* "

Then

('io

if

<

0,

both

+ a'2o)~^'"

real k the system is unstable.

nonlinear
There are many other types of stability to consider in the general case of
9.1-9.4
Definitions
than
more
any
discuss
shall
not
time-varying systems. For brevity, we
next
the
of
those
and
stability,
of
definitions
for zero-input systems. Furthermore, these
is
difference
only
The
systems.
discrete-time
section, carry over in an obvious manner to
that t takes on discrete values.

DEFINITIONS OF STABILITY FOR NONZERO INPUTS

9.3

its

relationships of a system than


In some cases we are more interested in the input-output
zero-input response. Consider the system
diLldt

with

Initial condition

other constant, for

all

x(to)
t

A(i)x

B(i)u

{9.1)

C(t)x

and where ||A(t)ll<K^, a constant, and


and A{t) and B(t) are continuous.

= xo

^ U,

llB(t)il

<

k^,

an-

that
externally stable if for any to, any xo, and any u such
on
only
depends
which
e
constant
^
a
exists
o, there
S for all t
\\vL{t)\\
^
to.
t
^
all
for

||y(^)||
ta, xo and 8 such that

The system

Definition 9.5:

{9.1) is

every bounded input produces a bounded output

In other words,

if

Theorem

The system

9.4:

with xo =
externally stable if and only
(9.1),

all

^ to

where

h{t, r) is

external stability.

and single input and output,


if there exists a number ^ <
\h{t,T)\dr

for

we have

/3

the impulse response.

is
>

uniformly
such that

CHAP.

STABILITY OF LINEAR SYSTEMS

9]

I95

If /3 depends on U the system is only externally stable. If xo - 0, we additionally require


the zero-input system to be stable i.s.L. and C(*) bounded for t> to so that the output does
not become unbounded. If the systeni has multiple inputs and/or multiple outputs, the

criterion turns out to be


\M\i

l"^!!

Proof:

we show

\y(t)\

We

diction.

has bound

set ui(t)

1 for

any

norm

I.

of a vector v in

is

'L'

we

then

;8,

get external stability.

take norms on both sides and use

norm

C\h{t,r)\dT

Since

properties to obtain

8/3

external stability.

we

prove

shall

\h{t, T)\dT

j8

by contra-

sgnh{t,T) for to-r^t, where sgn is the signum function which


By the hypothesis of external stability,

t.

|/i(i,T)|dT

externally stable

is

where the

/3

4.6).

f'\hit,r)\\u(r)\dT

the system

if

if

we then have

Definition 9.5

Next,

that

we can

h{t, t) m(t) dr,

From

and

(see Sections 3.10

First

\\U{t,r%dT

\yit)\

Ht,

dr

r) Uiir)

*'to

\h(t, t)|

dr

''to

Now

suppose

\h{t,T)\dT

Then by taking

not true.

is

suitable values of

and

to

/to

we can always make


and

to

6*0

this larger

in such a

Xe h{e,T)u2{r)dT

way

that

than any preassigned number


\h{e,T)\

dr

>

Again we

a.

Suppose we choose

a.

set M2(t)

sgn/i(6i,T)

so that

/so

Since a

y(e).

is

any preassigned number, we can

set

and

-0

arrive at a contradiction.
Example

9.11.

Consider the system dy/dt


at time

system

Then

t.

is

\U{t

This has an impulse response

u.

T)\dT =

h{t, r)

which becomes unbounded as

tg

not externally stable although the zero-input system

is

U(t
t

t),

tends to

a unit step starting


!.

Therefore this

stable i.s.L.

In general, external stability has no relation whatsoever with zero-input stability concepts because external stability has to do with the time behavior of the output and zeroinput stability is concerned with the time behavior of the state.
Example

9.12.

Consider the scalar time-varying system


dx/dt

where

(t,to)

a(f)a;

*^to

\h{t,T)\dT

<

e + 9(t.to)z{

and

Then the transition matrix

air) dr.

/i(t, to)

so that

"^-^

\h{t,T)\dT

e-' + to.<t>(t,

1.

is

*(*, fg)

to)' + ''"'''

Thus the system

can be almost anything, any form of zero-input stability

is

e- + e(to.x

is

=
-

e('.<o'.

Therefore

eto~t

externally stable.

open to question.

However, since

Q:(t)

STABILITY OF LINEAR SYSTEMS

196

[CHAP.

However, if we make C{t) a constant matrix, then we can identify the time behavior of
the output with that of the state. In fact, with a few more restrictions on {9.1) we have

Theorem

For the system

9.5:

(9.1)

with h norms,

C{t)

I,

<

and nonsingular B(i) such that


if and only if it is uni-

Kg_i, the system is externally stable


formly asymptotically stable.
|jB~i(^)||j

To have

B{t) nonsingular requires

it satisfies

9.4

the requirements stated in

to be

an ^-vector.

Theorem

9.5.

If

Theorem

is

constant and nonsingular,

9.5 is

proved in Problem

9.2.

LIAPUNOV TECHNIQUES

The Routh-Hurwitz, Nyquist, root locus, et;c., techniques are valid for linear, timeinvariant systems. The method of Liapunov has achieved some popularity for dealing with
nonlinear and/or time-varying systems. Unfortunately, in most cases its practical utility
is severely limited because response characteristics other than stability are desired.
the "distance" between the state vector and
can be found such that the metric tends to
the
vector.
-
Actually,
is asymptotically stable.
the
system
that
concluded
zero as i
<,
it can be
inequality
triangle
the
this,
because
show
metric
to
need
a
Liapunov realized we do not
(Property 4 of Definition 3.45) can be dispensed! with.

Consider some metric


If

Definition 9.6:

p(x(t),0).

This

is

some metric, any metric at

all,

time-invariant Liapunov function, denoted v(x), is any scalar function


of the state variable x that satisfies the following conditions for all t^to
and all x in the neighborhood of the origin:
(1)

v(x)

(2)

v(0)

and

its

partial derivatives exist

= 0;
for X ?^ 0;
v(x) >
dv/dt = (grad^ v)'^ d^/dt <

(3)

(4)

for x

and are continuous;

v^ 0.

The problem is to find a Liapunov function for a particular system, and there
method to do this.
Example

is

no general

9.13.

= x. We shall consider the particular function -nix) x"^. Applying


the tests in Definition 9.6, (1) nix) = x^ and Sv/Sa; = 2;a; are continuous, (2) );(0) - 0, (3) t]{x) = a;2 >
for all x-Q, and (4) dt^ldt = 2x dx/dt = 2.x^ <
for all x = 0. Therefore 7)(x) is a Liapunov function.
Given the scalar system dx/dt

Theorem

9.6:

Suppose a time-invariant Liapunov function can be found for the state


x of the system dTn/dt = f (x, t) where f (0, t) = 0. Then the state

variable

X=

is

asymptotically stable.

x^

-* ..
Definition 9.6 assures existence and
as
Here we shall only prove
Proof:
continuity of v and dv/dt. Now consider v{4>{t; ta, xo)), i.e. as a function of t. Since v> 6
and dv/dt
for i>^0, integration with respect to i shows v(^(ii; io,xo))> v(^(i2; io,xo))
for to<ti< tz. Although v is thus positive and monotone decreasing, its limit may not be
when
(Consider 1 + e'K) Assume v has a constant limit k > 0. Then dv/dt =
zero.
then dv/dt = {grad^vyi{Q, t) = 0. So
for ^ t^ 0, and when ^ V = K.
But dv/dt <
which implies v = 0, a contradiction. Thus v -> 0, assuring x - 0.
dv/dt=
implies ^ =
i5

<0

Additionally
the
if the system is linear or if we substitute "everywhere" for "in a neighborhood of
condition
(4) is
origin" in Definition 9.6, we have uniform global asymptotic stability. If
If Definition 9.6 holds for all

weakened

to dv/dt

0,

we have

to,

then

we have uniform asymptotic

only stability

i.s.L.

stability.

CHAP.

STABILITY OF LINEAR SYSTEMS

9]

197

Example 9.14.
For the system

of Example 9.13, since we have found a Liapunov function v{x) = x^, we conclude
the system dx/dt = -x is uniformly asymptotically stable. Notice that we did not need to solve the state
equation to do this, which is the advantage of the technique of Liapunov.

Definition 9.7:

time-varying Liapunov function, denoted v(x, t), is any scalar function


x and time t that satisfies for all t~ to and all x
in a neighborhood of the origin:

of the state variable

v(x, t)

(1)

and

its first

x and

partial derivatives in

exist

and are con-

tinuous;
v(0,i)

(3)

v(x,)^a(||x|j)

(4)

Note for

Theorem

all

norm

9.7:

dv/dt

v(x,

t)

to,

0;

>0 for xv^O and t^to, where a(0)


a continuous nondecreasing scalar function of |;

is

tion of the

(2)

v^ dx/dt +

(gradx

must be

dv/dt

<

f or

and

a(^)

v^ 0.

a continuous nondecreasing, time-invariant func-

||xj|.

Suppose a time-varying Liapunov function can be found for the state


variable x(^) of the system dx/dt = f(x, t). Then the state x =
is asymptotically stable.

Proof:

Since

dv/dt

<
>

and

integration with respect to t shows that


altered from that of Theorem 9.6
because the time dependence of the Liapunov function could permit v to tend to zero even
though x remains nonzero.
(Consider v = x^e~* when x = t).
Therefore we require
v(x, t) ^ a(||x||), and =
implies IJxH = 0. Hence if v tends to zero with this additional
v{x{to), to)

>

v{x{t), t)

assumption,

for

is positive,

Now

to.

we have asymptotic

must be

the proof

stability for

some

If the conditions of Definition 9.7 hold for all

to.

to

and

if

v(x, t)

^ p{\\x\\)

where

^(|) is a

continuous nondecreasing scalar function of | with /3(0) = 0, then we have uniform asymptotic stability.
Additionally, if the system is linear or if we substitute "everywhere" for
"in a neighborhood of the origin" in Definition 9.7 and require a(||x]|) ^ > with ||xj| ^ =o,
then we have uniform global asymptotic stability.
Example

9.15.

Given the scalar system dx/dt = x.


The function
9.6 at any fixed t and yet the system is not stable.
This

ment of

x'^e-*t satisfies all the

because there

is

is

no

requirements of Definition
meeting the require-

a(||x||)

Definition 9.7.

often of use to weaken condition (4) of Definition 9.6 or 9.7 in the following manner.
v(x, t) to eventually decrease to zero.
However, it is permissible for v(x, t) to be
constant in a region of state space if we are assured the system trajectories will move to
a region of state space in which dv/dt is strictly less than zero. In other words, instead of
requiring dv/dt <
for all x ^
and all t ^ to, we could require dv/dt ^
for all x 9^
It is

We

need

and dv/dt does not vanish


nonzero
Example

t^to

identically in

for any

to

and any trajectory arising from a

initial condition x{to).


9.16.

Consider the Sturm-Liouville equation d'^yldt'^ + p{t) dy/dt + q(t)y = 0. We shall impose conditions
on the scalars p(t) and q(t) such that uniform asymptotic stability is guaranteed. First, call y = x^ and
dy/dt = X2, and consider the function cCx, t) = xf + xl/q{t).
Clearly conditions (1) and (2) of Definition
9.7 hold if
a(||x||)

hold,

we

q(,t)

||x|||

is

continuously differentiable, and for

min{l,Ki}.

calculate

Here min

{1,

dt

/cj}

^1 dt

if

to hold suppose

(3)

ki

-y

q(t)

and

1,
rfa-

dt

l/q{t)

Ki

min{l,Ki}
-r^

2(t) dt

kj

>

if

for all
kj

<

1.

so that

For

(4)

to

STABILITY OF LINEAR SYSTEMS

198

Since dxi/dt

2 and dx2/dt

p(t)x2

q{t)x-y,

dv

[CHAP.

then
2p(t) q(t)

di

dq/dt

Q'Ht)

^^

But even if this requirement


>
for all t tQ, since q~^ > Kj.
and
any
Condition (4) cannot hold. However, if we can show that
for
a;,.
is satisfied,
when a;2 =
and x^ # 0, the system moves to a region where x^ ^ 0, then c(x, () will still be a Liapunov
Hence when X2 = 0, we have dx^/dt q(t)xi. Therefore if q{t) 7^ 0, then x^ must become
function.
and
Thus v{-s., t) is a Liapunov function whenever, for any t t^, q(t) 7^
nonzero when x^ # 0.
Under these conditions the zero state of the given Sturm\/q(t) ^ Ki > 0, and 2p(t) q{t) + dqidt > 0.
Hence we require
dp/dt

2p{t) q{t)

X2

dq/dt

Liouville equation is asymptotically stable.

Furthermore we can show it is uniformly globally asymptotically stable if these conditions are independent of o and if q(t) - K2. The linearity of the system implies global stability, and since the previous
s: >(x, t).
If q(t) k,i.
calculations did not depend on t^ we can show uniformity if we can find a /3(||x||)
max{l,/c-i} = ;8(llx||) ^^ .(x, f).
then

For the discrete-time

Definition 9.8:

we have

case,

the analog for Definition 9.7 as

discrete-time Liapunov function, denoted v(x, A;), is any scalar function


ko and all
x and integer k that satisfies, for all

k>

of the state variable

X in a neighborhood of the origin:

Theorem

continuous;

(1)

v(x, k) is

(2)

v(0,A;)

(3)

v(x,/c)s=a(||x!|)

(4)

Av(x, k)

0;

v(x(A;

>0

Xv^O;

for

+ l),k + 1) - v(x(fc), fe) <

for

v^ 0.

Suppose a discrete-time Liapunov function can be found for the state


is
Then the state x =
variable x{k) of the system x(fe -t- 1) = f(x, k).

9.8:

asymptotically stable.

The proof is similar to that of Theorem 9.7. If the conditions of Definition 9.7 hold for
and if v(x, fc) ^ j8(||x||) where jS(|) is a continuous nondecreasing function of | with

all ko
j8(0)

or

then we have uniform asymptotic stability. Additionally, if the system is linear


substitute "everywhere" for "in a neighborhood of the origin" in Definition 9.8
require a(|Ix||) -> o with ||x|| -^ <, then we have uniform global asymptotic stability.

if

and

0,

we

Again, we can weaken condition (4) in Definition 9.8 to achieve the same results. Condition
-=
if we have assurance that the system trajectories will
(4) can be replaced by Av(x, k)
to regions in which Av < 0.
in
which Av =
space
move from regions of state
Example

9.17.

A.

where l|A(fe)|l < 1


||x||. Then
= v{K{k + l))-y{^{k)) = ||x(fc + l)||-|ix(fc)|| =
then l|A(fc) x(fc)l| ^ |lA(fc)l| ||x(fc)|| < llx(fc)|| so

Given the linear system x(A;+ 1)


as a discrete-time Liapunov function

A(fc)x(fc),

Since l|A(fc)|| < 1,


properties of n can be verified, so this system
9.5

for all k and for some norm.

Choose

v{k)

is

||A(fc)x(fc)I|-||x(fc)||

The other pertinent


that A^ < 0.
uniformly globally asymptotically stable.

LIAPUNOV FUNCTIONS FOR LINEAR SYSTEMS

with Liapunov techniques is finding a Liapunov function. Clearly,


unstable no Liapunov function ej:ists. There is indeed cause to worry that
no Liapunov function exists even if the system is stable, so that the search for such a funcHowever, for nonlineai* systems it has been shown under quite
tion would be in vain.
is uniformly globally asymptotically
general conditions that if the equilibrium state x =
The purpose of this section is to
stable, then a time-varying Liapunov function (jxists.
manufacture a Liapunov function. We shall consider only linear systems, and to be specific

The major

if

the system

we

shall

diflSculty

is

be concerned with the asymptotically stable real linear system


dK/dt

A{t)x

[9.2)

CHAP.

in

STABILITY OF LINEAR SYSTEMS

9]

which there
||*(t,

exists a constant

i)|pdT exists for

Kg

depending on

some norm.

199

such that

Using Theorem

we can show

9.3

asymptotically stable systems satisfy this last requirement because


so that

\\^{r,t)fdT

e-^'^^^^-^dr

Kg(ij)

|lA(t)|l2

kI/2k^

that uniformly

||*(f, t)1|

<

k^

However, there do exist asymptotically stable systems that do not satisfy


Example

and where

K^e^'^i'"-'''

00

this requirement.

9.18.

Consider the time-varying scalar system dx/dt


(tg/t)'^'^.
The system is asymptotically stable for t^ >

Theorem

9.9:

(9.2),

i'(x, i)

where

4>^(r, f )

=0

x'^P(i)x

Q(t) *(t,

a time-varying Liapunov function,

is
i)

dr

which Q{t)

in

symmetric matrix such that

positive definite
is

(t/T)dT

to)

*.

For system
P{t)

This has a transition matrix *((,


However, even for t t^ > 0, we find

xl2t.

0.

|l<J>(T,t)l|2d7-

Note carefully the position of the arguments of

positive definite for all

if

t^

>

liQ(*)||

any continuous

is

k5(*o)

^^^ Q(*)"~I

small enough.

is

This theorem is of little direct help in the determination of stability because *(<, t) must be
known in advance to compute v(x, f). Note P(i) is positive definite, real, and symmetric.
Since

Proof:

* and

are continuous,

if

<

P(i)

for

all

||*(t,

tegral exists for the systems (9.2) under discussion, then condition
obviously holds. To show condition (3),
v{K,t)

Since

Q(0 - !

x(t)v^O.

is

= j'\^r,t)^{t)VQ{r)Hr,t)K{t)dr =
positive definite, then

- J)x ^

||A(i)||2

Kg

holds.

Condition

(1)

^ ej

for the system

Example

cZr

ex'^x

.Kg-l

-Kg-l

X^(T)x(T)dT

{9.2),

.j

>

for any

||x|||dr

then use of Problem 4.43 gives

J^"||A(T)|y|x(T)|||dr

J'"x''(dx/dT)dr

-eKg-1

.Kg-l||x(i)|||/2

f
=

X^Axdr

>

a(||x||)

x'^(t)

(2)

v(x,i)

dv/dt

of

Since the in-

X^(t) Q(r) x(r)

j^

(1)

t)f dr.

so that x'"Qx

Therefore
v(x,t)

Since

x'^(Q

then condition

to

-x' () Q(*) x()

<

since

is

Q(t) x(t) dr,

then

positive dennite.

9.19.

Consider the scalar system of Example 9.8, dx/dt = -x/t. Then 4>(t, to) = V*can be found for this system for to > 0, when it is asymptotically stable, because
\\Mt)\\2

f-1

<

t-'

K3o)

and

For simplicity, we choose Q{t) = 1 so that P(t) time-varying Liapunov function for all t to > 0.

f||*(r,f)|12dr

(t/r)2
t

dr

t.

Liapunov function

f"(t/r)2dr

Then

>'(x, t)

=
tx^,

which

is

STABILITY OF LINEAR SYSTEMS

200
Consider the system

^ U,

,,

dK/dt

.,

+
,

A(t)x

[CHAP.

^,

,_ .,

f(x, u(i), t)

{9.3)

,,

,^,

as ||x|| ->
continuous for small ||x|| and |lu||, and ||f|| -*
- 0.
As discussed in Section 8.9 this equation often results when considering
the linearization of a nonlinear system or considering parameter variations. (Let 8x from
Section 8.9 equal x in equation {9.3).) In fact, most nonswitching control systems satisfy
this type of equation, and ||f|| is usually small because it is the job of the controller to keep
deviations from the nominal small.

where for
and as ||u||

Theorem,

(x,u,

t) is

real,

system {9.3) reduces to the asymptotically stable system {9.2) when


then the equilibrium state is asymptotically stable for some small
and {|u|{.

If the

9.10:

0,

{{x||

In other words, if the linearized system is asymptotically stable for t U, then the corresponding equilibrium state of the nonlinear system is asymptotically stable to small disturbances.
Proof:
Use the Liapunov function v(x, t) = x^P(t)x given by Theorem 9.9, in which
(t, t) is the transition matrix of the asymptotically stable system {9.2). Then conditions
(1) and (2) of Definition 9.7 hold for this as a Liapunov function for system {9.3) in the
For small ||x|| it makes little difference
same manner as in the proof of Theorem 9.9.
whether we consider equation {9.3) or {9.2), and so the lower bound a(||x||) in condition (3)
can be fixed up to hold for trajectories of {9.3). For condition (4) we investigate dv/dt =
d(x''Fx)/dt = 2x^P dK/dt + x^ {dP/dt)Ti = 2x^PAx + 2x''Pf + x'" {dF/dt)x.
But from the proof
of Theorem 9.9 we know dv/dt along motions of the system dx/dt = A(i)x satisfies dv/dt =
-x^'Qx = 2x^PAx + x'' {dP/dt)K so that -Q = 2PA + d/dt. Hence for the system {9.3),
and ||u|| ^ 0, so that for small enough
dv/dt = -x^'Qx + 2x^Pf. But ||f||^0 as ||x|| ->
{|x|{ and ||u||, the term x^Qx dominates and dv/dt is negative.
Example

9.20.

Consider the system of Example 9.1. The linearized system dx/dt = x is uniformly asymptotically
which result
stable, so Theorem 9.10 says there exist some small motions in the neighborhood of a; =
in the asymptotic stability of the nonlinear system. In fact, we know from the solution to the nonlinear
equation that initial conditions Xq < 1 cause trajectories tViat return to a; = 0.

Example

9J!1.

Consider the system dx/dt = x + u(l + x u), where u{t) = e, a small constant. Then f(,x, e, t) =
Consequently Theorem 9.10 cannot be used directly. However,
e(l + ae e) does not tend to zero as x -* 0.
note the steady state value of a; = e. Therefore redefine variables as z x e. Then dz/dt z + ez,

which

is stable

Example

for

all

1.

9.22.

Consider the system


dxi/dt

X2

and

Xi(v^^+xl)/2

dx^/dt

- -x^ +

Xzix^

+ xl)/2

the harmonic oscillator


X2 and dxz/dt
The linearized system dxi/dt
-Xi is stable i.s.L.
- t]-i- which is
=
x^^)-'^
+
+
equation
is
nonlinear
to
the
[{xIq
xf
xf
of Example 9.4. But the solution
unbounded for any scio or a!2o- The trouble is that the linearized system is only stable i.s.L., not asymp-

because

it

is

totically stable.

9.6

EQUATIONS FOR THE CONSTRUCTION OF LIAPUNOV FUNCTIONS


From Theorem

9.9, if

dP/dt

we

take the derivative of P(i)

-*^(t.

i)

+
Since

*(t, t)

and

cI*(t, t)/dt

dP/dt

Q(i) *(i,

i)

4-

J]

we

obtain

[d^{T,t)/dtYq{r)Mr,t)dT

4>'^{T,t)Ct{T)[d*{r,t)/dt]dT

= d9~\t, T)/dt = -*(t, t) A{t), we


+ AT(^) P{t) + P{t) A{t) = -Q(i)

obtain
{9.4)

CHAP.

STABILITY OF LINEAR SYSTEMS

9]

Theorem

9.11:

If

201

and only if the solution P(i) to equation {94)


and Q(i) satisfy the conditions of Theorem

is

A(i)

dx/dt

{9.2) is

Theorem

where

then the system

{9.2),

A(i)x, is uniformly asymptotically stable.

then v
uniformly asymptotically stable,

If P(t) is positive definite,

Proof:

system

positive definite,

9.9,

x^P(i)x is a Liapunov function.


obtain equation {9.4) for P.

If the

we

9.11 represents a rare occurrence in

Liapunov theory, namely a necessary and

Usually it is a difficult procedure to find a


asymptotic stability.
Liapunov function for a nonlinear System, but equation {9.4) gives a means to generate a
Liapunov function for a linear equation. Unfortunately, finding a solution to equation {9.4)
means solving n{n + 1)/2 equations (since P is symmetric), whereas it is usually easier to
However, in the constant coefficient
solve the n equations associated with dx/dt = A(t)x.
case we can take Q to be a constant positive definite matrix, so that P is the solution to

sufficient condition for

PA = -Q

A^P +

{9.5)

This equation gives a set of n{n + 1)/2 linear algebraic equations that can be solved for P
any positive definite Q has been selected. It turns out that the solution always exists
and is unique if the system is asymptotically stable. (See Problem 9.6.) Then the solution
P can be checked by Sylvester's criterion (Theorem 4.10), and if and only if P is positive
definite the system is asymptotically stable.
This procedure is equivalent to determining
if all the eigenvalues of A have negative real parts.
(See Problem 9.7.)
Experience ha;s
shown that it is usually easier to use the Lienard-Chipart (Routh-Hurwitz) test than the
Liapunov method, however.
after

Example

9.23.

Given the system

up the

cPy/dP'

+ 2dy/dt + y -

*c

arbitrarily choose

Q=
1

The
and

solution is
so

To determine the

is

I*

by Liapunov's method we

stability

set

-IVph

Pi2\

[Pn

P22J

\Pi2

/3

positive definite

{1

0)x

1\
Pu]/
-2/
P22A-I

/-I

<>

-1

I,

1\

-1

'

Using Sylvester's

and the system

immediately says the system

and solve

-2y\Pi2
(

[-1 -2

dt

We

0.

state equations

criterion

But

is stable.

it is

we

find

much

p^ =

>

a:nd

det

P=

>

0,

easier to use Routh's test which ahnQst

is stable.

We may ask at this point what practical use can be made of Theorem 9.9, other than the
somewhat comforting conclusion of Theorem 9.10. The rather unsatisfactory answer is
that we might be able to construct a Liapunov function for systems of the form dx/dt =
A{t)x + f(x, t) where A(^) has been chosen such that we can use Theorem 9.9 (or equation
{9.5) when the system is time-invariant). A hint on how to choose Q is given in Problem 9.7.
Example

9.24.

Given the system


dx/dt = Ax + f

d^y/dt^

t
As found

in

Example

9.23,

+ {2 + e-*^) dy/dt+

=
2i/

dv/dt

{.:
Sxf

4)
Zx^x^

d{TsJ-Pii)ldt

0.

We

set

state

'

Then

2isJV dx/dt

= -x^Qx + ZxTf =

up the

'

(-."O

+ |
=

-x\

2xTp(Ax

xl

+f

(xi+X2)e~tx2

equation

-in

the

form

STABILITY OF LINEAR SYSTEMS

202

[CHAP.

dv

di

Using Sylvester's

criterion,

>

e-V2

and

e""*

>

e-t

e~^*/4, so the given system is uniformly asymptotically

stable.

For discrete-time systems, the analog of the construction procedure given

Theorem

in

9.9 is

v(x,fc)

x''(A;) *''(to, ;c)

Q(w) *(w, A;) x(fc)

mk

and the analog

of equation (9.5)

A'^PA

is

P = Q.

Solved Problems
9.1.

with rolling friction rests on a deformable surface in a gravity


the surface to obtain examples of the various kinds of stability.

ball bearing

Deform

Vxi/

field.

_Q^

Global

Asymptotic

Global

Asymptotic

Stability

Stability

Stability

but Unbounded

i.s.L.

(a)

(6)

(c)

Unstable
but

Unstable

Unbounded

Bounded

Unbounded

(d)

(e)

(/)

Stability
i.s.L.

but

and

Fig. 9-2

We can see that if the equilibrium state is g:lobally asymptotically stable as in Fig. 9-2(a),
then it is also stable i.s.L., as are (6), (c) and (d), i3tc. If the shape of the surface does not vary
with time, then the adjective "uniform" can be affixed to the description under each diagram. If the
shape of the surface varies -with time, the adjective "uniform" may or may not be affixed, depending on how the shape of the surface varies with time.
9.2.

Prove Theorems

9.3

and

9.5.

wish to show dx/dt = A(t)x


-KjCt to)
for all i tQ and all

We

uniformly asymptotically stable if and only if


*(i,o)||Also wi2 wish to show dx/dt = A(t)x + B(t)u, where
-KoCt-tn)
and hence if and only
-K)|h B-i' ^^ externally stable if and only if ||*(*, to'
ll'8~K*)l!i
if the unforced system is uniformly asymptotically stable.

is

to-

||*(t, to)ll - "le-"^"-'"'


and \\xo\] ^ S, then ||x(t)|| ^ ,S
if
Next, if dx/dt = A(t)x
which proves uniform asymptotic stability.
there is a fj (independent
>
and any
is uniformly asymptotically stable, then for any p >
In fact, for p = l
for all t > to + tj.
of to by uniformity) such that if |lxo|i - p then ||x(t)|| = e-i ^ ||x(to+ DH = ||*(to + T, to)xo||. Since
= e-i, there is a h = T such that
and
= 1, from Theorem 4.11, property 2, |i*(to+ T, fo)xo]| = |l*(to+ T, to)!!- Therefore for all
llxoll
ll*(*o+ T, to)i| e-i and for any positive integer fc,
*o.

Since

and tends

||x(t)l|

i|*(t, to)IM|xol|,

to zero as

t ->

i;

17

7;

ll*(to

+ fcr,

to)ll

k such that

Choose

Theorem

9.2.

\\4>{t^

to

+ kT,t^ + (k-l)T)\\-'

+ kT^ t<to + {k+

1)T

ll*(to

and

+ 2r,

to+r)llll*(to

define

/cj

|*(t, to

+ fcr)||

||*(to

+ fcr,

to)||

+ fcr,

Kie-ii|*(to
Kie

Kg

||*(t, to

T.

to)ll

1/T

proves Theorem

9.3.

to)lI

+ kT)\\e <

^i^

-Ct0+(.k+l^)T-t0^)/T

p-k

Then

I*(,*o)ll

Defining

-(fc+i)

<

from

CHAP.

STABILITY OF LINEAR SYSTEMS

9]

-KjCf-to)

If

ll*(,*o)lli-ie

then

||H(f,T)||idT

|l*(,T)i|l|B(r)lldT

If the

system

for
it

all fo arid all

>

"2

)/k2

is

KgKi/K2

to

/3

uniformly externally stable.

||H(t,r)l|,dT

J"to

e""^""^' dr

9.4,

||*(t,T)B(T)i|idr

to

||*,T)B(r)lli||B-l(T)||idr

=S

^^_j

<

oo

Also, for the system {9.1) to be uniformly externally stable with arbitrary Xq,

t.

must be uniformly

KgKi

Kg_i, then

||*(i,r)l|idr

||*(t,r)lidr

uniformly externally stable, then from Theorem

is

||B-i(t)]|i

proves the system

9.4

/3

Since

KgKi(l

Theorem

SO that use of

*0

to

203

stable i.s.L. so that

/3B-1

<

|I*(t, io)Ili

from Theorem

l|*(t,T)|Ii|]*(T,o)l|idr

S=

J"to

to

r mt,h)\Vdr

9.2.

!!*(,

Therefore
r)*(T,to)||ldr

|l*(f,*o)!|i(-to)

T = ^K^_^Ke so that at time t = tf,+ T, iiK-^-^K |l*(to + T, *o)IIi/'b-i''^- Hence


||*(io+ T, tt^)\\x e-i, so that using the argument found in the last part of the proof just given
for Theorem 9.3 shows that |l*(i, <o)lli - Kje""""'")'''', which proves Theorem 9.5.
Define the time

9.3.

Is the scalar linear

system dx/dt

2i(2 sini

l)a;

uniformly asymptotically stable?

We

should expect something unusual because \A(t)\ increases as t increases, and A(t) alternates
in sign.
We find the transition matrix as <!>(<, to) = e'^"~*"' where e{t) 4 sin t 4t cos t P.
Since e{t) < 4: + 4\t\ t^, lim e{f) = , so that lim *(t, io) =
and the system is asymptotically stable.
This is true globally, i.e. for any initial condition Xq, because of the linearity of the
system (Theorem 9.1). Now we investigate uniformity by plotting *(t, tp) starting from three different initial times: to ^> *o == Stt- and Iq = Av (Pig. 9-3). The peaks occur at ^{{2n + l)ir, 2nTr) =
gT7(4-7r)C4i + i)
so that the vertical scale is compressed to give a reasonable plot.
We can pick an
initial time to large enough that some initial condition in Ixq] < S
will give rise to a trajectory
and 6 > 0. Therefore although the system is stable i.s.L.,
that will leave \x\ < e for any e >
it is not uniformly stable i.s.L. and so cannot be uniformly asymptotically stable.

k*(t.to)

9.4.

If

^{t)

^ K+

[p(t) |(t)

^(t)]

dr

where

p{t)

and

is

a constant,

show that

f'p(T)dT
g-'to

This is the Gronwall-Bellman lemma and


the response of linear systems.

We

is

often useful in establishing bounds for

shall establish the following chain of assertions to prove the

Gronwall-Bellman lemma.

STABILITY OF LINEAR SYSTEMS

204

Assertion

(i)

Proof:

=s t

which

If

>

lo(fi)

tj.

is

a contradiction.

and

Applying assertion

Proof:

Let

Assertion (iv):

Proof:

Set u

d<f,/dt

then
in

o)(f)

- ^-

to

<o(to)

for

(ti)

a)(t)

- a(t)u ^

that

have

finite

dy/dt

a(t)Y

and

n(t)

we have

/c

"^'o

A{t) remains

if

da/dt

escape time.

p(t) |(t)

Since

+ f
^t

||xo

then

k,

||A(T)i|

0,

we

aW dr,

then

0,

gives

then

7(to),

0.

^6(f) =^ y(t).

()

+ f

e^^^^'-'o^W dr

in assertion

e*'^*'''

(iii),

e*"'*\

because dy/dt

to

+ /j(f),

p{t) a(t)

+ f

all

||x(i:)||

and

so that the given inequality says

dr,

+ /t(t) ~

for

is true.

/((r)]

Also show

A(r)x(r)dr||

>

<o(t)

T) da/dt

to-

9(t, to)

>

and

l|xo||

(t)

u(t).

Applying' assertion

k.

,ect,to)

^ U (||A(f)|| k),

then the system cannot

for any A().

||xo||e'''o"*''"^"''''

gives

\\

<o(to)

e-'"''''V(T)dT

x(t)

xq

norm

{t^

"(t)

between

<o(to)

bounded for

Integrating dx/dt = A(t)x


of both sides gives

for

and

d{e-'^*-Va)/dt

^(to)

e"'*'^'^''o-'/iW t^r

[p{t) |(t)

and

=
<o(r) =

co(ti)

(ii).

= (+

k.

to-

^0

w(t)

and since e"*"''"'

0,

in assertion

and y

ait)<p

The Gronwall-Bellman lemma

Let

Show

i|x(t)l|

such that u(r)

*i

then

0,

e-^'-'o'du/dt-altle"""-'"'

m
9.5.

and

e~^"'''u

<p

y(fo)

Assertion (v):

Since p{t)
(iv) gives

gives

If dio/dt

and

Proof:

(i)

If

(iii):

fi(t)

is

- a{t)a ^

da/dt

If

s=

0,

Multiplying the given inequality by e"*"'*"', where

Proof:

a{t)

But using the mean value theorem gives

(ii)

Assertion

a)(to)

then there

0,

Assertion

dw/dt

If

[CHAP.

A(r)x(r)dTl|

-^to

+J
^

Taking the

A(t) x(t) dr.


'"

||xo||

f
^h

use the Gronwall-Bellman inequality of Problem 9.4 with

||A(r)l| ||x(r)||

^(t)

dr

to obtain

If I|A(t)||^/c, then ||x(t)|| ^ l|xo||e''"-'\ which shows x(t) is bounded


However, if ||A(t)|| becomes uncannot have finite escape time.
system
for finite t and so the
bounded, we cannot conclude the system has finite escape time, as shown by Problem 9.3.
ll^^.;il

9.6.

ll^oll^'""'''^'"'"'-

Under what circumstances does a unique

solution

exist for the equation

A^P +

PA = -Q?
BP + PA = C where A, B and C are arbitrary real nXn
P and C as p; and q respectively, and the row vectors of A as
equation BP + PA = C can be written as BP
2 Pi*,^ = ^' '^^i<=^

We

consider the general equation


matrices. Define the column vectors of
a?'

in

for

1, 2,

. ,

w.

Then the

-I-

turn can be written as

'b

...

...

...

0\
\

B/

/ani

a.2il

...

aigl

a.22l

\<iiI

a2l

...

ail\
an2^

anni/.

CHAP.

STABILITY OF LINEAR SYSTEMS

9]

205

Call the first matrix in brackets B and the second A, and the vectors p and c. Then this equation
can be written (B + A)p = c.
Note B has eigenvalues equal to the eigenvalues of B, call them
Also, the rows of A XI are linearly dependent if and only if the rows of
^j, repeated n times.
A''' - Xl are linearly dependent.
This happens if and only if det (A^ Xl) = 0.
So X has eigenvalues equal to the eigenvalues of A, call them a^, repeated n times. Call T the matrix that reduces
B to Jordan form J, and V = [Vij] the matrix that reduces A to Jordan form. Then

^B

...

0\

/anl

...

ail\ |/T

...

^0

...

B/

\aiI

...

al/

...

'J ...

0\

...

J/

,0

Hence the eigenvalues of

unique solution to

Thereore

and only

/anl

...

ail

Vfti^l ...

al

|\0

\ /-ynl

T/

\-i;iI

...

v^^l'"

...

rI,

aj./

B + A are a; + 13j for i,j-l,2,...,n.


(B + A)p = c exists if and only if

det (B

+ A) =

fl

(i

+ /3j) #

0.

; + /3j'^0 for all i and j, does a unique solution to BP + PA = C exist.


If B - AT then /3^ = cij, so we require aj + aj
0.
If the real parts of all aj are in the left half
plane (for asymptotic stability), then it is impossible that
+ ,. = 0.
Hence a unique solution for
P always exists if the system is asymptotically stable.

if

if

9.7.

Show

that

PA + AT = -Q,

definite solution
This
as

if

where Q is a positive definite matrix, has a positive


and only if the eigenvalues of A are in the left half plane.

obvious by Theorem 9.H, but we shall give a direct proof assuming A


can be diagonalized
A. Construct P as (M-i)t]VI-i. Obviously P is Hermitian, but not necessarily
real.
positive definite because for any nonzero vector x we have
xtPx = xt(M-i)tM-ix =

is

lAM =

Note P is
(M-ix)t(M-ix)

> 0. Also, xtQx = -(M-ix)t(A + At)(M-ix) > if and only if the real parts of the
eigenvalues of A are in the left half plane.
Furthermore, ^ = xt(M-i)tM-ix decays as fast as is
possible for a quadratic in the state.
This is because the state vector decays with a time constant
equal to one over the real part of the maximum eigenvalue
V
of A, and hence the square of the
norm of the state vector decays in the worst case with a time constant equal to
1/2);.
To investigate the time behavior of v, we find

= -xtQx =

(M-ix)t(A + At)(M-ix) ^ -2,(M-ix)t(M-ix) = -2vv


For a choice of M-ix equal to the unit vector that picks out the
eigenvalue with real part
becomes an equality and then
decays with time constant 1/2?;.
dp/dt

v,

this

;-

9.8.

In the study of passive circuits and unforced passive


vibrating mechanical systems
we ^ften encounter the time-invariant real matrix equation Fd^y/dt^
+ Gdy/dt +
Hy - where F and
are positive definite and G is at least nonnegative
definite.
Prove that the system is stable i.s.L.

Choose as a Liapunov function

_
This

is

dy-r

dt^^ + ^'^% +

positive definite since

F and

are,

J'^(G+GT)gdr +
and
^dyT

ym+my

STABILITY OF LINEAR SYSTEMS

206
which

dp

dyT

di

dt

By Theorem
chosen

is

9.6,

because

["S + S +

H^] +

[-S

Also,

a positive definite quadratic function of the state.

is

[CHAP.

'_

dt

The reason this particular Liapunov function


the system is stable i.s.L.
it is the energy of the system.

was

Supplementary Problems
9.9.

dx/dt

Given the scalar system

tx

+u

and y

x.

\h{t,T)\dT

^ V2^e'

'0

910

9 11

9.12.

9.13.

(6)

Show

(c)

Explain

the system

is

not stable

i.s.L.

9.4 does not hold.

why Theorem

Given the system dx/dt = -x/t + u and y = x.


input at time * > 0. gives an unbounded output.
though the system is asymptotically stable for io

Show

(a)
(6)

>

the response to a unit step function


why Theorem 9.5 does not hold even

Explain
0.

show how Liapunov techniques can be used to give


altering only one condition in Definition 9.6,
state is not asymptotically stable.
sufficient conditions to show the zero

By

constant matrix A are < " i* ^"^ only


Prove that the real parts of the eigenvalues of a
a symmetric. Positive definite matrix
any symmetric, positive definite matrix Q there exists
Q.
equations -2aP + ATP + PAlinear
(n
+
l)/2
of
set
is the unique solution of the

Show

dx/dt

that the scalar system

-(1

+ t)x

is

asymptotically stable f or

if

given

P which

using Liapunov

theory.

9 14

Let a^iC*) 9-4.


Consider the time-varying network of Fig.
inductor, with
the
=
in
flux
X2(t)
and
capacitor
charge on the
Then L{t) dxjdt - x^
initial values Xio and a^ao respectively.
= 0. Starting
C(i) dxa/di + L(t)xi -H fi(t) C(f)x2

and

-AAAAAB(t)

L(t)

:c{t)

with

/R + {2L/RC)
P(*)

find conditions on R, L,

2/R

(^

and C that guarantee asymptotic

Fig. 9-4

stability.

9.15.

that if >
Using the results of Example 9-16. show
<l</ is< unilormiy ^Jr^i^ol'^ly'st^l
asymp
ady/dt + (1 + P cos 2t/^)y +
d^y/dt^
equation
Mathieu
the

even

^Zilf

the tie

Uapunov

=^m

v,.n> e

function

((x)

x^x,

.ff.^^^ct^JfS^-f "rnS Z^T.^^t:"^:

show

f W(T)dT
llxoike

9.17.

What

is

the

construction

A^PA - P = -Q?

-1

similar

<-

to

+!,<

that

J^.W-^-'"^

|lx()||2

ot

,-,f

M
^. nIWUe

Problem
Frobiem

9 7

'0

for

in

the

discrete-time

case

CHAP.

9.18.

STABILITY OF LINEAR SYSTEMS

9]

Show
(x,

that

if

+ a;|)/2

Example

the system of

and

rixj/dt

-Xj

eX2

page

9.22,

(xf

200,

changed slightly to dx^/dt = x^- ex^ +


e > 0,
where the system is uniformly

is

where

x|)/2,

207

asymptotically stable (but not globally).

9.19.

Given the system

-^x

f~^

"^^^]js.

\a

dt

-X J

Construct a Liapunov function for the system in the case e(t) ~ 0.


This Liapunov function must
give the necessary and sufficient condition for stability, i.e. give the exact stability boundaries on a.
Next, use this Liapunov function on the system where e{t) is not identically zero to find a condition
on e(t) under which the system will always be stable.
9.20.

||*(i,T)||

if

9.21.

= (A(t) + B(i))x where dx.ldt = A()x has a transition matrix with norm
Using the Gronwall-Bellman inequality with /i = 0, show 11x11 ^ |lxn||e'''3"''2'""'o'

Given the system dx/df


"2"--^)

||B(t)l|^3e-''^'.

Show

that

with

B(t)u

||x()||

||x||/'l"'''^^'"'^

initial condition

x(to)

Answers
9.9.

9.10.

(a)

\h(t,r)\dr

4>(t,t,)

(c)

The system

(a)

y(t)

(6)

The system

(t-tl

9.12.

Replace

9.13.

Use

<
<
<
<

9.14.

^'"'

Ki

Kg

=
*
^
^

to read

(4)

^ ^^e'^^^

e-^'^ dr

/3

depends on

in

Theorem

dv/dt

>

a neighborhood of the origin.

in

x2
R{t)
Lit)

^
^

K2

<

K4

<
<

=0

00

+ R{L/R2 - cm + CL/RC - L/R

i^always

positive definite,

eigenvalue of

Q=

and

M-i(I

has absolute value

Use any Liapunov function constructed for the


linearized system.

9.19.

If

9.20.

Ilxll

2I,

thenar

|(^

^"

- 2a6i - 92(1 + 2/2)2 >

||xo||-2<'-*o'

J^y

Use Problem

9.4

with

|lx(*)||

^.<-^)

is

which

is

positive

than one.

stable for all a if

0.

If

0,

o_

IIBWII

llxlle''^'

9.21.

The system

^|

- AtA)M-i

less

9.18.

require

9.4.

in equation {9.5).

^^'-^
^A.ZTX'^a'
definite if and only if each

Q=

not uniformly asymptotically stable.

A - al

K5 =S C{t) =^ Kg

A(t)x

t^t^Xi.

for

It)l2

is

Change condition
by

-O

not externally stable since

is

9.11.

dx/dt

Supplementary Problems

to

e'''^

for the system

xq.

//2e-'o/2

(5)

00

+ J^'/-'"^"""*'B(r)uWdr
'"

|(t),

|]x||

llxll

dr

l|xo!|e''^'

so apply the Gronwall-Bellman inequality


to

+ .sflMdr

||A(t)l|

p(t),

and

||BWu(r)l|

,(.).

we

chapter 10

Introduction to Optimal Control


INTRODUCTION

10.1

In this chapter we shall study a particular type of optimal control system as an introduction to the subject. In keeping with the spii-it of the book, the problem is stated in
such a manner as to lead to a linear closed loop system. A general optimization problem
usually does not lead to a linear closed loop system.

Suppose it is a control system designer's task to find a feedback control for an open loop
system in which all the states are the real output variables, so that

d^ldt

B(<)u

Ix

Ou

[10.1)

are given the system shown in Fig. 10-1 and wish to design what goes
*.
Later we will c ;nsider what happens when y{t) = C{t)x{t) where
not restricted to be the unit matrix.

In other words,
into

A(i)x

we

the box marked

C{t) is

x()

y()

System

<l(f):

{lu.i)

+
+
*

Fig. 10-1.

further restriction must be

Definition 10.1:

regulator

is

f'
s,

Vector Feedbatk Control System

made on

the type of system to be controlled.

a feedback control system in which the input

For a regulator the only forcing terms are due


state variables,

x.{to)

xo

in general.

?^

We

d{t)

0.

to the nonzero initial conditions of the

shall only study regulators first because

the extension to servomechanisms (which, follow a specified d{t)) will become easier,
turns out that the solution is the same if d(f) is white noise (the proof of this result
beyond the scope of the book), and (3) many systems can be reduced to regulators.

(1) later
(2) it

is

Example

10.1.

for t < t^ and


f7(i <o) =
u(t) = aUit), where the unit step
a,
for t - to. into a scalar system with transfer function 1/(8 + /3) and initial state x^. But
no
this is equivalent to a system with a transfer function l/js(s + ;8)], with initial states a and a;o and with
input. In other words, we can add an extra integrator with initial condition a at the input to the system
flow diagram to generate the step, and the resultant system is a regulator. Note, however, the input
becomes a state that must be measured and fed back if the system is to be of the form {10.1).

Given a step input of height

U(t

to)

require d(i) = 0. Then the


the feedback control. We
input u(i) is the output of the box marked * in Fig. 10-1 and
state x(i) and time. This
present
shall assume u is then some function to be formed of the
of the state
representation
the
is no restriction, because from condition 1 of Section 1.4 upon
history
past
the
all
summarizes
of this deterministic system, the present state completely
to^T^t,
for
=
u(x(t),
t)
of the abstract object. Therefore we need not consider controls u

Under the

restriction that

we

are designing a regulator,

we

is

but can consider merely u

u(x(i),

t)

at the start of our development.

208

CHAP.

INTRODUCTION TO OPTIMAL CONTROL

10]

209

THE CRITERION FUNCTIONAL

10.2

We desire the system to be optimal, but must be very exact in the sense in which the
system is optimal. We must find a mathematical expression to measure how the
system
must be optimal in comparison with other systems. A great many factors influence
the
engineering utility of a system: cost, reliability, consumer acceptance,
etc.
The factors
mentioned are very diflScult to measure and put a single number on for
purposes of comparison. Consequently in this introductory chapter we shall
simply avoid the question by
considering optimality only in terms of system dynamic performance.
It is still left to the
art of engineering, rather than the science, to incorporate
unmeasurable quantities into
the criterion of optimality.

In terms of performance, the system response is usually


of most interest. Response is
the time behavior of the output, as the system tries to
follow the input. Since the input to
a regulator is zero, we wish to have the time behavior
of the output, which in this case
IS the state, go to zero from the initial
condition xo. In fact, for the purposes of this chapter
there exists a convenient means to assign a number
to the distance that the response [x(t)
for
U\ IS from 0.
Although the criterion need not be a metric, a metric on the
space of all time functions between U and t, will accomplish
this.
Also, if we are interested
only
how close we are to zero at time t^, a metric only on all x(ii) is desired.
To obtain
a linear closed loop system, here we shall only
consider the particular quadratic metric

U^T^
m

p2(x,0)

W{U)%^(t,)

+ \

r*'xT(r)Q(x)x(r)dr
-'to

Where S

is an n x symmetric constant
matrix and Q(r) is a x t^ symmetric time-varying
matrix. If either S or Q(.), or both, are positive
definite and the other one at least nonnegative definite then p(x,0) is a norm on the
product space {x(i,),x(r)}. It can be shown
this requirement can be weakened to
S, Q nonnegative definite if the system
d^ldt = A(^)x
with y - VQ(^)x is observable, but for simplicity
we assume one is positive definite. The
exact form of S and
is to be fixed by the designer at
the outset.
Thus a number is
assigned to the response obtained by each
control law u(x(),<), and the optimum system is
y^
:/
that whose control law gives the minimum
p(x, 0).

QM

The choice of

interval

Co

Example

10.2.

Q(r) is dictated

by the relative importance of each state over the


time

t <C t\.

^^^' P*^" ^^' ^'* ^^^ '^^'''^^ f t^te variables (i).


If the angle of
attac?"''-"?^'^'-'''''"'/*/?^'"'"
'"'"' "' ^^" ""''"'^^^ *^ '"'^^'^l ^ a^ =
where H =
1 0)'~The "o-Hm'?'
?.-" " '" nonnegative definite and we must choose S to be positive definite
/
Pnrtf.
"^ ""-' '-' '' ''' --"^'^ --' -^^^^ ^^^^o^^^
r funernTf

x^^x

t-K

i^i^rs^i t^T'

The choice of S
Example

dictated

by the

TV,,.,,

^^^ ^^^^^^^'^^
"'
Tl ^''tent:li^:o%i:Z.T'
^Tpath
'M T"''
-f flies to arrive
the missile
\ "^" makes no difference what

^'''

''^

where

relative importance of each state at the


final time, U.

10.3.

stationed 'ir
t

is

e,, e^, e,

-'^"^ ^^^^^

near

at

are small fixed positive numbers to


be chosen by trial and error airer
after nnding
finding the
^h. closed
i
^
n0, an
, e
e,.
If any
anv e;

ui
,
it
unstable
system could
result, but might not.

loop system for edcn


each nxea
fixed

*.

j.

INTRODUCTION TO OPTIMAL CONTROL

210

For the system


-B-() xo

S(i

- U)

{10.1),

[CHAP. 10

the choice of control u(x(f),t) to minimize p(x,0)


for all
where B(i) has an inverse. Then x(t) =

is

t>U

in the case

u=
and

If B(i) does not have an inverse, the optimal control is a


derivatives,
such as equation {6.7), page 135, and can drive
sum of delta functions and their
almost instantaneously. Because it is very hard to mechanize a delta function,
x(*) to
vsrhich is essentially achieving infinite closed loop gain, this solution is unacceptable.
must place a bound on the control. Again, to obt;ain the linear closed loop system, here we

p(x, 0)

its

0,

minimum

value.

We

shall only consider the particular quadratic

|jJ'u^(r)R(T)u(T)dr

where

R(t) is

mxm

an

symmetric time-varying positive

definite

matrix to be fixed at the

positive definite to Jissure each element of u is bounded. This


outset by
and will be added to p2(x,0). The relative magnitudes
"energy",
is the generalized control
to the relative values of the response and control
proportion
of ||Q(t)|| and ilR(T)j| are in

the designer.

energy. The larger


gain of the system.

10.3

is

|iQ(T)|| is

relative to

the quicker the response and the higher the

|iR(T)|l,

DERIVATION OF THE OPTIMAL CONTROL LAW

The mathematical problem statement


want to minimize
v[x,u]

ix^(*i) Sx(f i)

that

is

we

are given

+ UJx^^(t)Q(t)x(t) +

dx/dt

A(i)x

+ B(t)u and
{10.2)

Xx'{r)^{T)xx{r)]dr

defer the rigorous


shall give a heuristic derivation of ttie optimal control and
called the costate
p(i),
time
of
function
derivation to Problem 10.1. Consider a real w-vector
shall determine.
that
we
equation
multiplier, which will obey a diff(jrential

Here we

or Lagrange
Note for any x and u obeying the state equation, p''(i)[A()x
this null quantity to the criterion changes nothing.
. [x,

u]

ix^(ti) Sx{<:i)

Integrating the term

p'^cZx/dr

(ix-^Qx

H-

iu'-Ru

0.

Addmg

+ p^Ax + p^Bu -p^dx/dr) dr

by parts gives
v[x,u]

vjx(f,)]

"

+ B(i)u-dx/dt] =

Wx] =

v,[x(t,)]

ix''(i,)Sx(iJ

f ''(ix^Qx +
^

vjx]

+ V3M

x^(i^)p{i,)

3i:^Ap

+ x'^{tMK)

(^^^)

+ x^ dv/dr) dr

{10. -If)

to

broken into v,, v, and v,, and heurisIntroduction of the costate p has permitted v[x, ti] to be
minimum when v,,v, and v, are each mdependently
tically we suspect that v[x,u] will be
function attains a local minimum
minimized. Recall from calculus that when a smooth
of
suspect that if v, is a minimum, then the gradient
its derivative is zero. Analogously, we
{10.3)

with respect to

x(fj) is zero:
p(ti)

and that

is

zero:

if v, is

minimum, then the gradient


^p/^^

(^0-^)

Sxo-Cii)

of the integrand of

-AJ{t)v

Q(*)x-

{lO.J,)

with respect to
i^^-'^)

CHAP.

Here

INTRODUCTION TO OPTIMAL CONTROL

10]

211

the response x(0 using the optimal control law u''(x(t),


t) as the input to the
Consequently we define p{t) as the vector which obeys equations
(10.6) and
(10 7). It must be realized that these steps are heuristic
because the minimum of v might
not occur at the combined minima of v^,v^ and v^;
a differentiable function need not result
trom taking the gradient at each instant of time; and also
taking the gradient with respect
to X does not always lead to a minimum.
This is why the iinal step of setting the gradient
ot the integrand of (10.5) with respect to
u equal to zero (equation (10.8)) does not give a
rigorous proof of the following theorem.
x<'>(i)

system

is

(10.1).

Theorem

Given the feedback system (10.1) with the criterion


(10.2), having the
by (10.6) and (10.7). Then if a minimum exists, it is obtained
by the optimal control

10.1:

costate defined

= -R-^B^p

"-

The proof

(10.8)

given in Problem 10.1, page 220.

is

To calculate the optimal control, we must find p(^). This is done by solving
the system
equation (10.1) using the optimal control (10.8) together with
the costate equation
(10.7),

l/x-\
with

x<"'(io)

Example

and

xo

p(ti)

Sx''(ii).

10.4.

Consider the scalar time-invariant system

Then

A=

2,

B=

R=

1,

1/4,

Q =

3,

x2(j)

S=

Xo,

p(l)

2a;(l).

/xHt)\

[pit))
Evaluating this at

Then from Theorem

with criterion

+ uyi)dt

solve

-4\/a;op\

[-S -2)[p J

Using the methods of Chapter

5,

-4N-|/a;op(0)\

+ -^[-S

2JJ(p(0)j

_
-

and using

re- /2

L^U
p(l)

4\

e)

2a;p(l)

^''-''^

gives

10.1,
^'"'(^o.t)

Note
u (Xo,

{3x^

-J

_
-

+u

2x

Hence we

d_/xp\

iBOPCO)

dx/dt

2.

dt[p
with

-B(i)R-i(^)B^(^)\/x^\

A(t)

Op + it
1

4- orxpS-it

sr/s

^0

(10.m

this procedure has generated

u- as a function of t and x. If xo is known, storage


as a function of time only, in the memory of a
computer controller, permits open
loop control; i.e. starting at U the computer
generates an input time-function for d^^/dt =
Axp + Bu'"'(xo, t) and no feedback is needed.
ot

i)

However, in most regulator systems the initial conditions


xo are not known.
By the
introduction of feedback we can find a control law such
that the system is optimum for
arbitrary xo. For purposes of illustration we will
now give one method for finding the
feedback, although this is not the most efficient way to
proceed for the problem studied in
this chapter.
We can eliminate xo from dxV(i^ = Ax"" + Bu'>(xo, i) by solving for x''(i)
terms of xo and t, i.e. x-Ci) =: #(;xo,*o) where
^ is the trajectory of the optimal system.
liien x^(t) - ^(f;xo, to) can be solved for
xo in terms of x'">(i) and t, xo = xo(x'>(i), t). Substituting for xo
the system equation then gives the feedback
control system d^^^/dt =

Ax^p

+ Bup(x'"'(), ^).

[CHAP.

INTRODUCTION TO OPTIMAL CONTROL

212
Example

To

10

10.5.

find the feedback control for

Example

from

10.4,

(10.10)

and

(10.11)

feedback control
Solving this for Xo and substituting into (10.12) gives the
4- '50g8(l t)
Mop(xP(t),

Hence what goes

in the

box marked

t)

* in Fig. 10-1 is the

^, ,
K(i)

and the overall closed loop system

-Y^-^^a=tT

2
I

^''(*)

time-varying gain element K(t) where

+ 30e8(i-
]5g8(i-t)

is

_
"

dx"P

IT

20.38"-"
5e8(i-

THE MATRIX RICCATI EQUATION

10.4

let

P(*)

P(*)
To find the time-varying gain matrix K(i) directly,
-R-BW^
t)
=
Then
u-^lx,
an n X n Hermitian matrix to be found.
- BR-B^P)xo^
=
The closed loop system then becomes d.^^/dt (A

matrix

Substituting p

*(*'-)

Vdx'-'-/dt

(dP/dt

= -Qx^ -

AT + PA - PBR-'B''P)xp

+Q+

and since xo is arbitrary and


matrix Riccati equation
the
satisfy
matrix P must

But

x'"'()

*,,(, o)xo,

-dF/dt

A^Fx"-'

then gives

(10.9) for dTc'>/dt

Using the top equation of

gives
into the bottom equation of {10.9)

= Px-

{dF/dt)K.^

^"""-,^S.^'

't!^*)-

'"f
^.T
r.L.Hion
its transition
^ call

= Q+

*,, is

nonsingular,

we

find the

AT + PA - PBR-BT
=

nxn

(10.13)

Changing inde-

P(i)
This has the "final" condition P(x) = S since :p(f 1)
becomes dP/dr = Q + A P +
equation
Riccati
=
matrix
the
U-t,
pendent variables by r
The equation
PA - PBR-'BT where the arguments of the matrices are U-r instead-of t.starting at the
to,
to r = i,
=
can then be solved numerically on a computer Jirom r
also be solved
can
equation
Riccati
matrix
the
=
Occasionally
S
P(0)
initial condition

x(tx)

Sx(iO.

analytically.

Example

10.6.

For the system of

Example

10.4, the 1

matrix Riccati equation

-dP/dt
with P(l)

2.

Since this

*-*'

is

dP
f 4{P^"3/2)(P
+

l72)

P(t)

K(t)

= -R-^BTP =

+ 4P-4P2

separable for this example,

^
-

3/2
nt)
,..
8^"F(ti)-3/2

_
-

Taking antilogarithms and rearranging, after setting

Then

is

ti

1,

zw+i/^

*" P(ti)

1/2

gives

15 -V e8"-"
10 - 2e8-i'

in
-4F(t), which checks with the answer obtained

Example

10.5.

CHAP.

INTRODUCTION TO OPTIMAL CONTROL

10]

Another method of solution of the matrix Riccati equation


Partition the transition matrix as

is

213

to use the transition

o (10.9) directly.

matrix

Then

Eliminating

so that

x(ij)

from

V{t)

this gives

P(i)x(^)

[*2i(i,g

*,,(,

<j)S][*,^(f,g+*^^(^gS]-ix()

P is

(i0.i5)

the product of the two bracketed matrices.


sufficient condition (not necessary)
for the existence of the solution V{t) of the Riccati equation is
that the open loop transition
matrix does not become unbounded in
U. Therefore the inverse in equation (10.15)

U^t^

can alv^ays be found under this condition.


Example

10.7.

For Example

10.4,

use of equation [10.10) gives P(t) from (10.15) as

which reduces

We

to the

'

[2e4Ct-i)

answer obtained

in

6e-4<t-i)

Example

2(4e4"-i)

4e-*<t-i^)]/8

10.6-

have a useful check on the solution of the matrix


Riccati equation.

Theorem 102:

If S
and

is positive definite and


Q() at least nonnegative definite, or vice versa,
R() is positive definite, then an optimum v[x,
u^p] exists if and only if
the solution P(^) to the matrix Riccati equation
{10.13) exists and is bounded
and positive definite for all t < U. Under these conditions
=

v[x. u"-!

Proof is given in Problem 10.1, page 220. It is evident


from the proof that if both
() are nonnegative definite, P(f) can be nonnegative definite if v[x,
u""] exists.

Theorem

10.3:

For

S{t), Q(i)

and

equation (10.13)

is

R{t) symmetric, the solution P(^) to the

S and

matrix Riccati

symmetric.

Proof:
Take the transpose of (10.13), note it is identical to
(10.13), and recall that
there is only one solution F(t) that is equal to S at time
tu

Note this means for an nXn P(t) that only (n + l)/2 equations
need be solved on the
computer, because S, Q and R can always be taken symmetric.
Further aids in obtaining
a solution for the time-varying case are given in Problems
10.20 and 10.21.

10.5

TIME-INVARIANT OPTIMAL SYSTEMS

So far we have obtained only time-varying feedback gain elements.


The most important
engineering application is for time-invariant closed loop
systems.
Consequently in this
section we shall consider A, B, R and Q to be constant,
S =
and ii -> oc to obtain a constant feedback element K.
Because the existence of the solution of the Riccati equation is guaranteed if
the open
loop transition matrix does not become unbounded in
U, in the limit as ii -> =o we
should expect no trouble from asymptotically stable open loop
systems. However, we wish
to incorporate unstable systems in the following
development, and need the following existence theorem for the limit as <i -> =o of the solution to
the Riccati equation, denoted by n.

U^t^

INTRODUCTION TO OPTIMAL CONTROL

214

Theorem

system

If the states of the

10.4:

controllable, then

are not asymptotically stable are


and n{to) is constant and posi-

{10.1) that

lim P(fo; U)

[CHAP. 10

n{to) exists

''"

tive definite.

Define a control u1(t) = -B''(T)*^(i2,T) W->(o,t2)x(*o) for U^r^tz, similar


Proof:
drives all the controllable states to zero at
to that used in Problem 6.8, page 142. Note
=
the response of the asymptotically
Defining
>
for
t
ti.
the time t2<. Let u'(t)

J,"

v\Ti,U^

Note

v[k,u^]

Xa^'sQxasdi

from Theorem

Here

x(fo).

Ps=o (*"'*)

after carrying out the integration, since both x(*) and


some bounded scalar function of U and tz. Then

10.2,

so
\\n(to)\\2 - 2a(to, tz) <
~ Ps=o(*<';*i) ^^ U^t^ ti,
'X'

v[x,u'']

n(io) is

bounded.

and

also that

bound

=0

can be shown that for S


0, then

It

0,

when S >

Therefore, lim F{to; h) must be a constant because for


P(*o;*i) until it hits the

<

(io, t2) x^(to) x(to)

lim||Pg^(io;i)-Ps.(o;<)||

be to increase

<

Xa'^QXasd*

a{to, ^2) is

ix''(io)n(<o)x(to)

Therefore

Therefore

o.

r'(xTQx + ul^Rul)di +

^ a{to,t2)-x.'^{to)7^{to)

u\t) are linear in

<

large any change in P(fo; U)

ii

must

n(to)-

Since we are now dealing with a time-invariant closed loop system, by Definition 1.8 the
time axis can be translated and an equivalent system results. Hence we can send U -^ >,
and integrate numerically backwards in time
start the Riccati equation at P(ii) = 8 =
until the steady state constant U is reached.
Example

10.8.

Consider the system dx/dt

-dP/dt

= 1 - P2

with

n =

Since n

is

is

0.

lim P(to;

ti

The optimal control

P(ti)

=u

up

with criterion

"

This has a solution


*i)

CO

lim tanh
tj

= R~W^Jlx =

-+

(fj

{x^

P(to)

<o) =

tanh

lim
fQ -+

00

+ u^)

CO

Then the

dt.

(tj

tanh

is

Therefore

t).

(i

Riccati equation

to)

x.

a positive definite constant solution, of the matrix Riccati equation,

it satisfies

the quadratic algebraic equation

= Q + A^n + HA - nBR-iB^n
Example

(10.16)

10.9.

For the system of Example 10.8, the quadratic algebraic equation satisfied by
has solutions 1, so that n = 1 which is the only positive definite solution.

is

n?.

This

A very useful engineering property of the

quadratic optimal system is that it is always


Since a linear, constant coefficient, stable closed loop system
always results, often the quadratic criterion is chosen solely to attain this desirable feature.
stable if

controllable.

it is

Theorem

10.5:

n exists for the constant coefficient system, the closed loop system
asymptotically stable.

If

Choose a Liapunov function


nonzero x. Then

Proof:
for

all

dV/di
for

all

V = x'^Ux.

Since n

= x^n(Ax-BR-B^nx) + (Ax-BR-iBi'nx)rnx =

because

is

positive definite

and IlBR~iB''n

is

-x'^Qx
is

positive definite,

x^'nBR-iB^'nx

nonnegative

definite.

V>
<

is

CHAP.

INTRODUCTION TO OPTIMAL CONTROL

10]

Since there are in general n{n


only positive definite solution.

Theorem

The unique

10.6:

Proof:

solutions of (10.16),

n2)(A

definite solutions Di

- BR- wn.) + (A - BR->B^n2r(ni -

helps to

it

know

that

is

the

positive definite solution of (10.16) is n.

Suppose there exist two positive

(n,

+ l)

215

^^'

*^^* *^^ ^^"^^^"

Ea.

q-q

n^)

and

Then

(lo.i?)

XF + GX = K

f7?^f
ever^w^l^
A,(F) +
A,(G) '''n for any i and j, where A,(F)
^0
eigenvalue of G. But A - BR-B-n. for i = 1

is

has a unique solution X whenan eigenvalue of F and A,(G) is an

and 2 are stability matrices, from Theorem


Therefore the sum of the real parts of any combination
of their eigenvalues must
be less than zero.
Hence (10.17) has the unique solution Hi-Hz = 0.
Generally the equation (JO.iff) can be solved
for very high-order systems, n ^ 50
This
gives another advantage over classical
frequency domain techniques, which are not so easily
adapted to computer solution. Algebraic solution
of (10.16) for large n is difficult, because
there are n(n +
1) solutions that must be checked.
However, if a good initial guess F(t,)
IS available, i.e. P(U) =
8F(U) + n where 8P(i.) is small, numerical solution of
(10.13) backwar from anyj. to a steady state gives n. In other
words, 8F(t) = P(t) 10.5.

as

t -^

00^

Theorem

which

10.7:

n tends

IS

true because:

to zero

If A, B, R and Q are constant matrices,


and R, Q and the initial state F(t2)
are positive definite, then equation
(10.13) is globally asymptotically stable
as ^ ^ -co relative to the equilibrium state
n.

8P obeys

Proof:

the equation (subtracting (10.16)

-dSF/dt

from

(10.13))

= Fr8P + 8PF-8PBR-iB'^P

where F = A-BR-'B'"n.

Choose as a Liapunov function 2F = tr[8P^{nn^)-i8P1


For
investigate here only the real scalar case.
Then 2F = -Qn-^ - B^R-m and
2 8P^,
so that dV/dt = n-^ 8PdSP/dt = n-^ SP^Qn-^
+ B^R-^P) >
for all nonsince Q, R, P and n are all > 0. Hence 7 as -> -=c, and 8P - 0.
It can be
the vector case that dV/dt >
also.

we

brevity,

2F - n

zero SP
shown m
Example

10.10.

snnnLT

an^nLi

^'- *""

^/T^!.^

TimpTori

^-^'

'""'''^"'^

*^^

'^"^'^*'" "^"^ *

''^^'^

^'n!'"''*^^'"'

''

'^''"'"

^"

^(*o)

'<=^^t i"ial condition,

tanh

(*,

tanh"!

e).

i.e.

For

Experience has shown that for very high-order systems,


the approximations inherent
numerical solution (truncation and roundoff) often lead
to an unstable solution for bad
initial guesses, in spite of Theorem
10.7.
Another method of computation of n will now
be investigated.

Theorem

10.8:

If

a^ is

(10.9),

an eigenvalue of H, the constant 2n x 2n matrix corresponding


to
then -Ai is also an eigenvalue of H.

Proof:
Denote the top
elements as gi. Then

n elements

\ii

\S^

-^?)

of the

=
=

A,

-Qf,

(-B^-^B.

ith.

eigenvector of

as

BR-iB'-g^

A^g,

:?)(7')

H^(7'

and the bottom n

[CHAP. 10

INTRODUCTION TO OPTIMAL CONTROL

216

= (detM^)* for any matrix


This Shows -\i is an eigenvalue of H^. Also, since detM
we find that det (H- - |I) = 0. This shows that
M, then for any | such that det (H - |I) =
Since -Xi is an eigenvalue of
of H.
is an eigenvalue of H^ then ^ is an eigenvalue
if
of H.
H'', it is concluded that -xt is an eigenvalue

symmetrically with regard


This theorem shows that the eigenvalues of H are placed
eigenvalues with real parts < 0,
axis in the complex plane. Then H has at most n
to the
and will have exactly n unless some are purely imaginary.

>

Example

10.11.

corresponding to

Example

10.8

H =

is

matrix
which are symmetric with respect to the imaginary

The

Example

axis.

_^x

This has eigenvalues +1 and -1,

(_^
V

10.12.

fourth-order system might give rise to an 8

8 IH having eigenvalues placed as

shown

in Fig. 10-2.

ImX

ReX

Fig.

1(1-2

factorization into eigenvalues with real parts <


of looking at the Wiener factorization of a polynomial

The

Theorem

10.9:

If Ai, A2,

.,

A are

and
p{o>^)

distinct eigenvalues of

= GF-i where F =

(fi |2|

real parts

is

another

way

< 0,

then

()

having real parts

and G =

|fr.)

>

into p''{<^)p

(gilg2|

|gn)

as defined

in equation (10.18).

Proof-

show

first

Since n

is

GP-i

is

that

we will
the unique Hermitian positive definite solution of (10.16),
it is
that
and finally
a solution to (10.16), th<m that it is Hermitian,

positive definite.

Define annxn diagonal matrix


the eigenvalue equation (10.18),

{ly
from which

with

H(S)

Ai, \2, ....

An its diagonal elements so that

from

(4 -^".n
/m
io\
{^0.19)

FA = AF-BR-iB'^G
GA = -QF-A^G

(iO-^O)

A in (10.20) gives
GF-iAF - GF-iBR-^B'^G = -QF - A^G

Premultiplying (10.19) by F"

and substituting for

Postmultiplying by F-^ shows GF-i satisfies (10.16).

Next,

Hermitian.
then Hermitian.

we show GF-^

GF-i = Ft-iMF-i

is

is

It suffices to

M = FtG Hermitian,
of M be mjic-iigk; tor

show

Let the elements

is

since
j

^ k,

CHAP.

INTRODUCTION TO OPTIMAL CONTROL

10]

Q n^

^^i^UA-i
>^3r^>'5-tiHV
Since the term in braces equals
Finally, to

e(<)-.Fe^'F

we have

!^

217

.K

-I

i\hV*^

= m*

and thus GP-i is Hermitian.


show GF-i is positive definite, define two w x w matrix functions of time
as
and #() = Ge^'P-'. Then fl(c) = = ^(oc). Using (J0J9) and
(iO.^0),
de/dt = (AF - BR-iB^G)e^'F-i
0,

d^/di

m.'^

= -(QF + A^G)e^'F-i

Then

GF-i

t(o)^(o)

-j'J^^e^^at

so that

(e^'F->)t(FtQF
Since the integrand

is

positive definite,

GF-i

Corollary 10.10: The closed loop response

The proof

is

similar to the proof that

is

is

GF-'

GtBR-B''G)(e^'F-)di

positive definite,
x(^)

Fe-^'-o^F-xo

n.

with costate

p(*)

is positive definite.

Corollary 10.11: The eigenvalues of the closed loop


matrix

and the eigenvectors of

and GF-^

A - BR-i B^n

are

A-BR-'B^n
Ifi,

fz,

are Ai,X2,

fn.

The proof follows immediately from equation (10.19).


Furthermore, since \^, X2
\
are assumed distinct, then from Theorem 4.1 we
know f., f
f are linearly independent,
so that F-i always exists.
Furthermore, Theorem 10.5 assures ReAi <
so no A, can be
imaginary.
.

Example 10.13.
The H matrix corresponding to the system of Example
Correspondmg to these eigenvalues,

10.8 has eigenvalues


^

-1 =

X,

and
c*""

+1
-rx =

X,
^2-

where a and

;8 are any nonzero constants.


Usually, we would merely set a and fi equal to one hut for
purposes of instruction we do not here. We discard X^ and its
eigenvector since it has a real part > 0,
and form F - /i - a and G = g^ = a. Then n = GP-^ =1
because the a's cancel. From Problem 4.41,
the vector case F = FoK and G = GoK where K is the
diagonal matrix of arbitrary constants associated with each eigenvector, but still n = GF-i = GoK(FoK)-i =
GoFq.

Use of an eigenvalue and eigenvector routine to calculate n from Theorem


10 9 has
given results for systems of order n ^ 50. Perhaps the
best procedure is to calculate an
approximate Ho = GF 1 using eigenvectors, and next use Re(no +
nJ)/2 as an initial guess
to the Riccati equation (10.13).
Then the Riccati equation stability properties (Theorem
10.7) will reduce any errors in Do, as well as provide a check on the eigenvector
calculation.

10.6

OUTPUT FEEDBACK
Until

now we have

considered only systems in which the output was the state, y = Ix.
the states are available for measurement. In the general case y =
C(<)x,
the states must be reconstructed from the output. Therefore we must
assume the observability of the closed loop system dx/d^ = F(i)x where F(t) =
A(t)~B{t)R-\t)Bm)F{t-ti)

For y

= Ix,

all

[CHAP. 10

INTRODUCTION TO OPTIMAL CONTROL

218

To reconstruct the

state

C{t)K

dy/dt

from the output, the output

differentiated

n-1

times.

Ni(t)x

N,(^)dx/di

d"-y/di"-i

is

+ dNJdtx =

(N-iF

(NiF

+ dNi/dt)x = Nax

+ dN-i/dQx = Nx

Define a
NI) is the observability matrix defined in Theorem 6.11. ^
where N'' = (Nr|
|Id" Vdi" y.
{I\Id/dt\
=
H^(d/di)
ll{d/dt)
by
operators
nk X k matrix of differentiation
rank n. From
Then x = N-'{t) H{d/dt)y. Since the closed loop system is observable, N has
Using the
n.
rank
has
N"^
inverse
generalized
the
Property 16 of Section 4.8, page 87,
deuniquely
is
and
exists
x
n-vector
the
conclude
results of Problem 3.11, page 63, we
-R^iBTN^^Hy.
=
termined. The optimal control is then u
.

Example

10.14.

Given the system


1

dt

-012

X2

with optimal control u = k,x,


output is M = k^y + k^ dy/dt.

+ k^x^.

System

Since y

2/

(1

0)x

X2

x^

and dx,/dt

x^,

the optimal control in terms of the

(iO.l)

State
estimator

-R-^B^
Pig. 10-3

This
is not very practical
involves derivatives of y, it may appear that this
there
which
deterministic system
mathematical result arises because we are controlling a
feasible.
is
is no noise, i.e. in which differentiation
system must
thai; the probabilistic nature of the
such
is
noise
However, in most cases the
that under certain cirresult of stochastic optimization theory is
be taken into account.
the optimal
in place of the state
cumstances the best estimate of the state can be used
An estimate of each
(the "separation theorem"
control and still an optimum is obtained
is as shown
so that structure of the optimal controller
state can be obtained from the output,
Since

in Fig. 10-3.

10.7

THE SERVOMECHANISM PROBLEM


Here we

problems.

servomechanism problems that can be reduced to regulator


the box marked * * in
the optimum compensation to go into

shall discuss only

We

wish to

find

Fig. 10-4.

Fig. 10-4.

The Servomechanism Problem

CHAP.

INTRODUCTION TO OPTIMAL CONTROL

10]

The

criterion to be

minimized

W{ti)
Note when

e{t)

y{t)

- d{t)

is

[e^(T) Q(t) e(T) +


+ ^
J''

Se(i)

minimized,

is

219

follow

y(i) will

u'-Cr)

R(t) u(t)] dr

(10.21)

d{t) closely.

To reduce this problem to a regulator, we consider only those d{t) that can be generated
by arbitrary z{to) in the equation dz/dt = A()z, d = C()z. The coefficients A() and C(t)
are identical to those of the open loop system
Example

(10.1).

10.15.

Given the open loop system

(0

jx +

Then we consider only those inputs


inputs only ramps, steps and

e~^<-

Restricting d(t) in this

-2/

\0

62

)m

2/

(1

1)x

V63/

= z^ + Z2o( +

1) + z^^e-^K
In other words,
functions and arbitrary linear combinations thereof.

d(t)

manner permits

dw/dt

A(t)w

defining

B(t)u

new
e

state variables

we can

consider as

w = x-

z.

C(t)w

Then
(10.22)

Now we have the regulator problem (10.22) subject to the criterion (10.21), and solution
of the matrix Riccati equation gives the optimal control as u = R^'B'^Pw. The states
can be found from e and its w - 1 derivatives as in Section 10.6, so the content of the box
marked * * in Fig. 10-4 is R-^B'^PN-'H.

d(t) be generated by the zero-input equation dzldt =


has significance in relation to the error constants discussed in Section 8.3. Theorem
8.1 states that for unity feedback systems, such as that of Fig. 10-4, if the zero output is
asymptotically stable then the class of inputs d(t) that the system can follow (such that
lim e(t) 0) is generated by the equations dw/dt A(i)w + B(f)g and d(t) C(i)w + g

The requirement that the input

A.(t)z

where

any function such that lim

g(i) is

g(t)

0.

Unfortunately, such an input

reducible to the regulator problem in general.


However, taking g =
system can follow the inputs that are reducible to the regulator problem
zero output is asymptotically stable.

if

is

not

assures us the
the closed loop

Restricting the discussion to time-invariant systems gives us the assurance that the
closed loop zero output is asymptotically stable, from Theorem 10.5. If we further restrict
discussion to inputs of the type di = (t io)'"^C/(*-<o)ei as in Definition 8.1, then Theorem

Then we must introduce

8.2 applies.

function matrix H(s)


Example

is

integral compensation

not of the correct type

when

the open loop transfer

to follow the desired input.

10.16.

Consider a system with transfer function G{s) in which G(0) ^ , i.e. it contains no pure integrations.
We must introduce integral compensation, lis. Then the optimal servomechanism to follow a step input
R^ib^nN~iH(s). This is a linear combination
is as shown in Fig. 10-5 where the box marked ** contains
Thus we can write the contents of **
of l,s,
.,s" since the overall system G{s)/s is of order m + 1.
+ fc^s""!. In a noisy
as fcj + &2S +
+ A;s". The compensation for G{s) is then fej/s + A;2 + ^38 +
environment the differentiations cannot be realized and are approximated by a filter, so that the compensation takes the form of integral plus proportional plus a filter.
.

OJ/o

d(t)

step

^
I

-e(t)

K
Fig. 10-5

u{t)

y{t)

10.8

[CHAP. 10

INTRODUCTION TO OPTIMAL CONTROL

220

CONCLUSION

For time-invariant
In this chapter we have studied the linear optimal control problem.
We can take
Section
8.6.
systems, we note a similarity to the pole placement technique of
we select
Either
problem.
our choice as to how to approach the feedback control design
equivaThe
criterion.
the
the pole positions or we select the weighting matrices Q and R in
is not
equivalence
although the
lence of the two methods is manifested by Corollary 10.11,
optimal
is
k'-x
u=
one-to-one because analysis similar to Problem 10.8 shows that a control
of this equivalence
dual
The
^
all
a,.
for
ldet(y,IA-bk^)|
|det(>IA)l
if and only if
Kalman-Bucy filter and the algebraic separais that Section 8.7 on observers is similar to the
theorem of stochastic optimal control.
tion theorem of Section 8.8 is similar to the separation

Solved Problems
10.1.

Prove Theorem
The

10.1,

page 211, and Theorem


was

heuristic proof given previously

10.2,

page 213.

not rigorous for the following six reasons.

fixed, the resulting


By minimizing under the integral sign at each instant of time, i.e. with t
optimal functions
resulting
the
Therefore
minimum is not guaranteed continuous in time.
sense.
make
no
would
(10.7)
equation
and
derivative
have
a
x<>P(t) and p(t) may not

(1)

was found, and only

(2)

The open loop time function

(3)

We

(4)

Taking derivatives of smooth functions gives


We wish to guarantee a global minimum.

u(t)

later related to the feedback control u(x,

t).

wish to take piecewise continuous controls into account.


local

maxima, minima, or

inflection

points.

the minimum of v.
supposed the minimum of each of the three terms n, V2, and v^ gave
minimum, then equations (10.6), (10.7)
said in the heuristic proof that if the function were a
We wish to give sufiicient condiminimum.
a
for
conditions
necessary
were
and (10.8) held, i.e.
certain quantity v(x, t) obeys a
a
that
tions for a minimum: we will start with the assumption
is attained.
minimum,
that
a
show
then
and
equation,
partial differential

We
We

(5)
(6)

corresponding to the optimal system


start the proof, call the trajectory x(t) =: *(t; x(to).
Note ^ does not depend on u
condition
x(to).
dx/dt = Ax + BuP(x, t) starting from to and initial
v[x, u<'p(x, t)] can be evaluated
Then
t.
and
x
function
of
specified
as
a
chosen
since u has been
parameters <i, to and x(to). Symbolically,
is known, simply by integrating out * and leaving the
if

To

r[x, uP]

= i^nh; x(*o), <o)S*(*i; x(y

+
Since

we can

v(^(to),to;ti)

parameter

to)

*'

Ij

[i.^T; x(to). to) Q(r) *(r; x(to). ^o)

dr

"[^.u"^]
from any initial conditions x(<o) and initial time to, we can consider
dependmg on the fixed
where v is an explicit function otn + 1 variables x(to) and to.

start

tj.

Suppose we can find a solution

+ ix^Qx -

with the boundary condition


iu^TBu
=:

where the equality

is

v(k,

t)

to the

Hamilton-Jacobi) partial differential equation

^(gr&d^v)'^R-^B-r grad^v
v(x(t,), t,)

iicT(t^) Sx(ti).

+ (erad^v)-^u
+ R-iBTgradxV)

t),

^(u

R-iBTgrad3,iJ)''R(u

s=

attained only for

and dx/dt

Jx^Qx

(10.S3)

Note that for any control u(x,

-R-iBTgrad^i;

Rearranging the inequality and adding ^x^Qx to both sides,


^xTQx + ^umu ^ ^xTQx - i(girad^i;)^R-'BTgrad^v
(i 0.^5)

(grad^i;)i'Ax

^(grad^v)rBR-iBTgrad^i;

u(x,t)

Using

r)Ru<'P(*, t)]

(W^ni-.

= Ax + Bu, we

i^^^-^i)

(grad^t;)^Bu

get

+ iumu ^ -[11 + (grad^fHAx + Bu)J =

-^.,(x(t),t)

CHAP.

INTRODUCTION TO OPTIMAL CONTROL

10]

221

Integration preserves the inequality, except for a set of measure zero:

IX"

(xTQx

+ u''Ru) dt

s=

v{K(to),to)

u(x(ii), ti)

to

Given the boundary condition on the Hamilton-Jacobi equation


control

(10.23), the criterion ^[x, u] for

any

is

4x,u]

ix'r(ti) Sx(ti)

'

(x^Qx

ui-Ru) dt

ss

i,(x(fo), <o)

to

so that if v(x(^o)>

*o) is

nonnegative

definite,

u<>P(x, t)

then

the cost due to the optimal

it is

-R-i(t)BT(t) gradxV(x,

control, (10.2^):

t)

To solve the Hamilton-Jacobi equation {10.2S), set v{Ti, t) = lxT(t)x where P(t) is a timevarying Hermitian matrix to be found. Then grad,. v = P(t)x and the Hamilton-Jacobi equation
reduces to the matrix Riccati equation [10.13).
Therefore the optimal control problem has been
reduced to the existence of positive definite solutions for t < ij of the matrix Riccati equation.
Using more advanced techniques, it can be shown that a unique positive definite solution to the
matrix Riccati equation always exists if ||A(t)]| < < for t < t^. We say positive definite because

|xr(to) P(to) x(to)

and

all

nonzero

x(to),

'y(x(fo),o)

which proves Theorem

Since

10.2.

for all

we know from

Theorem

Riccati equation is equivalent to (10.9), this proves

10.2.

>

-'[x-u""]

%<

t^

Section 10.4 that the matrix

10.1.

Find the optimal feedback control for the scalar


time-varying system dx/dt
u/t to minimize

- x/2t

= ^ C\x^ + u^)dt.

Note the open loop system has a transition matrix

The corresponding Riccati equation

dP
dt

with boundary condition P{j)

<t-{t,r)

(r/t)i/2,

which escapes at

is

_
~

-,

_ P _

P^
~fi

This has a solution

0.

Pit)

f?

t\

+i2

f2

^""^ ^""^ * < *i - 0* ^ *'


However, for -t, < t < Q < t the interval
\- ^^1
which the open loop system escapes, P{t) < 0. Hence we do
not have a nonnegative solution of
the Riccati equation, and the control is not the optimal
one in this interval. This can only happen
when \\A{t)\\ is not bounded in the interval considered.

in

10.3.

Given the nonlinear scalar system dy/dt = y'^e. An open loop control d(i) =
-t has
been found (perhaps by some other optimization scheme) such that
the nonlinear
system dyjdt = y^d with initial condition i/(0) = 1 will follow the
nominal path
Vnit) - (1 + 1^) ^'^.
Unfortunately the initial condition is not exactly one, but
y{0) =
1 + e where is some very small, unknovra number.
Find a feedback control that is
stable and will minimize the error y{ti) - yn{ti) at some time
U.
Call the error y(t) - y^{t) = x{t) at any time. Consider the
feedback system of
The equations corresponding to this system are
dy _ dy
^x
di - 'df^ It "" (Vn + =o)Hd + u)

=
Assume

the error

\x\

yld

<

+
\y\

\Zylxd

Zylxd

+ Zy^xU + xH +

and corrections

+ ylu\ >

\u\

\Zy^xU

<

\d\

j/^m

Zylxu

+ Zy^xH +

Pig. 10-6 below.

x^u

so that

+ xU + Zylxu + Zy^x^u-\-xM

{10S5)

[CHAP. 10

INTRODUCTION TO OPTIMAL CONTROL

222

V(0)

d{t).

tO

+.
y{t)

e(t)

Multiplier

Cuber

u{t)

x{t)

K{t)

l/n()

Fig. 10-6

Since dyjdt

yld,

the approximate relation

we have

dx

Sylxd

,3

3t

:=

ylu

dt

We

:X ^
+
(2*

(l

f2)3/2

choose to minimize

= uHtj) + I f

time

R{t) uHt) dt

some appropriately chosen weightinj? function such that neither x{t) nor
Then K{t) = -P(t)/[a
gets so large that the inequality (10.25) is violated.

where R(t)
t

where

'

is

u(t) *

P(t) is the solution to the Riccati equation

dP
'

p2

etp

dt

(1

t2

t2)3B(t)

This Riccati equation is solved numerically on a computer from tj to 0.


with P(ti) = 1.
the proposed feedback system.
the time-varying gain function K(t) can be found and used in

10.4.

any

+ t^W^R(m

Given the open loop system with a transfer function


to minimize

where

do is the value of

ljl{y-doy +

{s

+ a)-\ Find

Then

a compensation

{du/dtY] dt

an arbitrary step input.

we introduce
Since the closed loop system must follow a step, and has no pure integrations,
described by
is
integration
pure
with
a
series
in
system
open
loop
The
integral compensation.
that
dx2/dt = -aX2 + u, and defining Xi-u and du/dt = n gives dx^/dt = du/dt = /< so
dx

:-!)-fi-

dt

(0

l)x

Since an arbitrary step can be generated by this, the equations for the error become

dw

;-!)-(::"

dt

subject to minimization of

Using the method of Theorem

(e^

f
^0

then

/I

i^^^)

(0

l)w

The corresponding matrix Riccati equation

dt.

n-(:->--(;-)-"G o10.9 gives

n
where

optimal control
V^y = -yJa^ + V^*^ + sfJ^^V^^ and A-i = a2 + 7a + l. The
system
is as shown
loop
closed
=
the
=
de/dt,
= A(wi + ywj). Since e w^ and Wi ae +

Fig. 10-7 below.

is

in

CHAP.

INTRODUCTION TO OPTIMAL CONTROL

10]

step.

"'

+ a + 8)

A(y

)
/

223

f*

+ a)

8(8

Fig. 10-7

Figure 10-8

is

equivalent to that of Fig. 10-7 so that the compensation introduced

is

integral

plus proportional.

tO

step

-I-

Fig. 10-8

10,5.

Given the single-input time-invariant system in partitioned Jordan canonical


form

A
dt

where Zc are the j controllable states, za are the n-j uncontrollable states,
Je and Jd
are ^X; and (n-j) x (n-j) matrices corresponding to real Jordan form.
Section 7.3,
b is a y-vector that contains no zero elements, m is a scalar control, and
the straight
lines indicate the partitioning of the vectors and matrices.
Suppose it is desired to
minimize the quadratic criterion

where

'

| f"

[z'^ Q,cZo

+ p-^u^jdr.

Here

>

a scalar, and Q^ is a positive definite symmetric real matrix. Show that the
optimal
feedback control for this regulator problem is of the form
u{t) = -k^x^it) where
k is a constant ^'-vector, and no uncontrollable states Zd are fed back. From the results
is

of this, indicate briefly the conditions under

which a general time-invariant singleinput system that is in Jordan form will have only controllable .states fed
back; i.e.
the conditions under which the optimal closed loop system can be separated
into controllable and uncontrollable parts.
The matrix n

Q =

satisfies

equation (10.16).

and partition

For the case given,

A =

led

n =

Then
^cd

-p(b^

OT)

-^

= -pb^nA

pb^HedZd

"^cd

Now

if n,.

can be shown to be constant and n^d can be shown

0,

then

k^

pb^n..

(10.16),
lcd
-t-

Jd
n,
''cd

Dcd

Ucd

-"'^(^^)-(^

But from

[CHAP.

INTRODUCTION TO OPTIMAL CONTROL

224

10

or

JcDc

JcHcd

^d"-cd

JriO
d"d

HcdJd

HdJd

nL'^c

HcbbTHed

n>brn.

Qc

nJdbbTne

partition,

Within the upper left-hand

system, which has a constant positive definite


the matrix Riccati equation for the controllable
corner,
right-hand
Within the upper
solution He-

JcHed

n^dJd

pOcbbrn,,,

= by inspection, and thus u = -pb^n^z,.


This is a linear equation in n,d and has a solution n,d
Jordan form by x = Tz, where
Now a general system i - Ax + Bu can be transformed to real
T-iAT =

criterion

Then the

J.

gives the matrix Riccati equation

shown for

z^

^j"

i f"(xTQx + p-iM2)d^

and

{z'rTTq,'Iz

T^QT -

za if

p-iu^)dr

Qc
(

'

^^- ^^^ ^^^ '^^'

Otherwise uncontrollable states must be fed back^


troUable states are weighted in the criterion.
but if they are all stable
if any are unstable,
diverges
.
If the uncontrollable states are included,
and v remains finite.
them
by
influenced
is
states
controllable
the action of the

10.6.

Given the controllable and observable system


dx/dt

A(t)x

B(t)u,

C{t)x

{10.26)

that the closed loop response of this system approach that of an ideal or
model system dw/dt = L(i)w. In other M^ords, we have a hypothetical system
dw/dt = L(i)w and wish to adjust u such that the real system dx/dt = A(t)x-f B(t)u
behaves in a manner similar to the hypothetical system. Find a control u such that
the error between the model and plant output vector derivatives becomes small, i.e.
It is desired

minimize

\t

^t,!-//ir

/^-.r

constants and B = C = I, R = 0, this is equivalent to


asking that the closed loop system dx/dt = (A + BK)x be identical to dx/dt = Lx
and hence have the same poles. Therefoire in this case it reduces to a pole placement

Note that

scheme

in the case A,

L=

(see Section 8.6).

Substituting the plant equation (10.26) into the performance index (10.27),

= 1 r

''

([(c

+ CA - LC)x

CBu]TQ[(C

-t-

+ CA - LC)x +

CBu]

u^-Ru} dt

(10.28)

cross products of x
This performance index is not of the same form as criterion (10.2) because
Since R is positive
'+
LC)x.
BrQB)-iBrc^Q(dC/dt
CA
+
and u appear. However, let S = u + (R
z^Rz -I- zTB^QEz definite,
<
nonnegative
B'^QB
is
Since
z.
nonzero
for
any
definite, z^Rz >
always exists.
inverse
its
hence
and
zT(B -I- B'^QB)z so that B = R -i- B^QB is positive definite
system
(10.26) becomes
the
Then
x.
u
and
terms
of
in
found
be
can
always
control
u
Therefore the
d-s.ldt

and the performance index

A = A-BK,

Ax-f

Bu

1
= i

CA - LC

/"'l

J.

A. A.

(xTQx-fumu)dt

= (B^QB + R) - iBTCQIVI,
Q = CT[(M-BKrQ(M-BK)-h K^BK]C and R = B + B^QB

dC/dt

-I-

(10.29)

becomes

(10.28)
^

in which, defining M(t)

--

and K()

(10.30)

CHAP.

INTRODUCTION TO OPTIMAL CONTROL

10]

225

Since R has been shown positive definite and Q is nonnegative definite by a similar argument,
the
regulator problem (10.29) and (10.30) is in standard form.
Then u = -R-iBTPx is the optimal
solution to the regulator problem (10.29) and (10.30), where P is the positive definite
solution to

- q + A^p + PA - PBR-iRTp

-dP/dt
with boundary condition

(ti)

The control

0.

to minimize (10.27) is then

u = -R-iBi^(P + BrcrQM)x
Note that cases in which Q is not positive definite or the system (10.29) is not controllable
give no solution P to the matrix Riccati equation.
Even though the conditions under which
procedure works have not been clearly defined, the engineering approach would be
to try
particular problem and see if a satisfactory answer could be obtained.

10.7.

In

Problem

it

may
this

on a

feedback was placed around the plant dx/dt = A(i)x + B{t)u,


would behave similar to a model that might be hypothetical. In
this problem we consider actually building a model dw/dt = 'L{t)w with output
V = J{t)w and using it as a prefilter ahead of the plant. Find a control u to the plant
such that the error between the plant and model output becomes small, i.e. minimize

10.6,

C{t)x so that

it

- v)^Q(y - v) +

[(y

u'rRu] dt

Again, we reduce this to an equivalent regulator problem.


be written as

rf^(x)
and the criterion

is,

a)(J

(o

The plant and model equations can

+ (b

(10^2)

in terms of the (w x) vector.

^VC-J
Thus the

{10.31)

/So

solution proceeds as usual and

is

C)rQ(-J

o^^

u'TRu Idt

a linear combination of

and x

However, note that the system (10.82) is uncontrollable with respect to the respect
to the w
In fact, merely by setting L =
and using 3(t) to generate the input, a form of general
servomechanism problem results. The conditions under which the general solution
to the servomechanism problem exists are not known. Hence we should expect the solution
to this problem
to exist only under a more restrictive set of circumstances than
Problem 10.6. But again, if a
positive definite solution of the corresponding matrix Riccati
equation can be found, this procedure
can be useful in the design of a particular engineering system.
variables.

The resulting system can be diagrammed as shown

in Fig. 10-9.

OWo
Model

Mt)

>

v(t)

Feedforward

u(t)

:^

Plant

+
Feedback

Fig. 10-9

y(t)

x(()

>

C(t)

INTRODUCTION TO OPTIMAL CONTROL

226

10.8.

[CHAP.

10

Given the time-invariant controllable and observable single input-single output system dTi-ldt = Ax + hu, y = c'^x. Assume that u = U-x., where k has been chosen such
that 2v

{qy"^

+ u^) dt

the closed loop system eigenvalues as g


Since the system

is

Find the asymptotic behavior of

has been minimized.

optimal,

= -b^n
=

qcc'^

and g

-*

where n

-^ =o.

which

satisfies (10.16),

is

written here as

+ nA - nbbTn

Ai'n

(10.33)

is the unique nonnegative definite solution of the matrix Riccati equation.


If q = 0, then n =
Then u = 0, which corresponds to the intuitive feeling that if the criterion is independent of
Therefore as q -' 0, the closed loop eigenvalues tend to the open
response, mal<e this control zero.

loop eigenvalues.

To examine the case g -* , add and subtract


and on the left by (si Ai")"! to obtain

sD;

from

(10.33),

multiply on the right hy

- (-sI-AT)-in -

9(-sI-Ar)-iccr(sI-A)-i - n(sI-A)-i

(-sl

(si

- A)-i

^ A^)-inbbTn(sI- A)-i
(10. Si)

cr(sX- A)-ib and the


Multiply on the left by b^ and the right by b, and call the scalar G(s)
given
an input with Laplace
that,
is
notation
=
this
A)-ib.
reason
for
The
+bTn(sI
scalar H(s)
transform p(s) to the open loop system, x(i) = ^-'{(sl - A)-ibp(s)} so that

Jl{y(t)}

cr^{x()}

cT(sI-A)-lbp(8)

-^{m(0} = bTn^{x(i)} =

and

In other words, G(s)


input to the control.

is

1 to

G(s)p(s)

- A)-ibp(s) =

the open loop transfer function and H(s)

Then

is

H(s)

p(s)

the transfer function from the

becomes

(lO.Si)

qG(-s) G(s)

==

Addmg

b''-n(sl

H(s)

H(-s)

H(s) H(-s)

each side and rearranging gives


!1

+ H(s)|2 =

(10.35)

q\G(sYf

numerator of H(s)
has been shown that only optimal systems obey this relationship. Denote the
But d(G) G(s) as d(G).
of
d(H)
and
H(s)
as
of
denominator
the
and
n(G),
G(s)
as
and
of
as n(H)
Multiplying (10.35) by ld(G)|2 gives
det (si - A) = d(H).
It

\d(G)

+ n(H)\2 =

\d(G)\^

(10.36)

q\n(G)\^

zeros of n(G), the open loop system, then 2m zeros of (10.36) tend to the
The remaining 2(n -m) zeros tend to and are asymptotic to the zeros of the
zeros of \n(G)\^.
of \d(G) +
equation s2(n-"5 = q. Since the closed loop eigenvalues are the left half plane zeros
remainthe
and
zeros
loop
open
-*
the

to
poles
tend
loop
closed
w(i?)|2, we conclude that as q
s2C-"i> = q.
closed loop poles tend to the left half plane zeros of the equation
ing

As

-^ =, if

there are

w-m

n-m

and radius q^
order
In other words, they tend to a stable Butterworth configuration of
open
loop zero, then 2 open
and
one
poles
loop
open
=
3
y-'.
has
system
the
If
2(m-i)
where
loop poles are asymptotic as shown in Fig. 10-10.
independent of the open loop pole-zero conAlso, note equation (10.36) requireis the
-^ 0.
closed loop poles to tend to the open loop poles as 9
This

is

figuration.

Furthermore, the criterion

(qy^

vP) dt

is

general for scalar controls since c can be chosen


designer (also see Problem 10.19).

quite
loy

the

results of
criterion
particular
this
for
only
are
valid
this analysis
involving the output and are not valid for a general
quadratic in the state.

Of course we should remember that the

^^'

CHAP.

INTRODUCTION TO OPTIMAL CONTROL

10]

227

Supplementary Problems
10.9.

Given the scalar system dxidt

minimize

to

2v

^0
10.10.

10.11.

+u

with

Consider the plant

desire to find

(a)

Write the canonical equations

(6)

Solve the canonical equations for x(t) and

(c)

Find the open loop control

For the system and


write

(6)

find

(c)

verify

minimize

to

a;(0)

Find a feedback control

Xp.

-^5)'' + (f'"

Hi

We

(a)

initial condition

4- ^2^ ^i^

(9x2

Ax

2c

d(.

p(t).

i).

Problem

criterion of

+ m^)

and their boundary conditions.

(10.9)

m(xq,

(^x\

10.10,

three coupled nonlinear scalar equations for the elements

using the relationship

P(t)

p(t)

P(t) x(f)

(f^j

the

of

and the results of part

P(t)

(6)

matrix,

of Problem

10.10,

the results of

satisfy the equations found in

(6)

(a),

and that

P(t) is constant

and

positive definite,

draw

(d)

10.12.

a flow diagram of the closed loop system.

Consider the plant


dt

^(r'

;)(:;)

and performance index

I r
Assuming the

initial conditions

10.13.

are

What

will give the control (10.8).

dx/dt

and the performance index

10.14.

Using the

(b)

What

is

Sx^T)

state-costate transition matrix *({,

lim

+ m2)

+u

r),

(c<;2

+ m2) dt

find P{t) such that

p(t)

P(t) x(t).

P(t)?

Given the system of Fig. 10-11, find K(t) to minimize


2v

(x2

+ uyp)dt

and

find

lim K(t).
Zero input

to

10.15.

dt

Xi{0) = 5, KgiO) = -3, find the equations (10.9) whose solutions


are the boundary conditions for these equations (10.9)'!

Consider the scalar plant

(a)

(.=?

Given the system

dx

,r\
+

LJ

w
.

Je

8 -V a

D'^ + fJ'"

(ft

K(t)

Find the control that minimizes


-too

2c

(xT:'X

+ m2) dt

Fig. 10-11

10.16.

line.
Let r = v and v = u, where r is the position
the acceleration due to the control force.
Hence the state equation

Consider the motion of a missile in a straight

and V

is

the velocity.

d fr
di

Also,

is

;)(:)-(:'-

It is desired to

minimize
T

2v

where

and

q^ are scalars

>

0.

q^vHT)

q,rKT)

+ T

uHt)

d-

INTRODUCTION TO OPTIMAL CONTROL

228

Find the feedback control matrix M{e) relating the control

[CHAP. 10

u(t) to the states r{t)

and

v(t) in

the form

where

10.17.

= T t.

Here

Given the system

e is

known as

= (

the "time to go".

criterion

"^^^^

Ir^il)"^

root locus of the closed loop roots as q varies frora zero to

10.18.

Consider the matrix

r
r

_
=

= j

2v

(qxl

+ u^)dt. Draw

the

>.

,1

-62/r"|

L ?

a J

associated with the optimization problem

dx/dt

ax

bu

"

x{Q)

Show

that the roots of


location of the roots of

((fx2

+ rM2) dt

p(T)

xo;

H are symmetrically placed with respect to


H depends only upon the ratio g/r for fixed

the ju axis and show that the

a and

h.

10.19.

Given the system (10.1) and let S = 0, Q(t) = v^Ut) and R(t) = pRo(i) in the criterion (10.2).
Prove that the optimal control law up depends only on the ratio ii/p if Qo(*) and Ko(t) are fixed.

10.20.

Show

and

10.21.

10.22.

also

det Li,

,^

/,

is

symplectic,

'

~"

i.e.

Problem 10.20, show that for the case


provides a check on numerical computations.
results of

For the scalar time-varying system


t)

to minimize

''

r*^

^^^^^ "^

= -x

dx/dt

^"^^

tan

P(t)

0,

find the

u,

<l>~^ (t, t) *2i(t, r).

This

optimal feedback control

'^^^

Find the solution P(f) to the scalar Riccati equation corresponding to the open loop system with
finite

escape time dx/dt

<

for

10.24.

show

Using the

u(x,

10.23.

that the transition matrix (lO.li)

to

<

h,

for

-(x

<

*i

+ u)/t

<

with criterion

"

<

to

<

a^d ^o^

Given the system shown in Fig. 10-12.


2u

C[{y-d)^

-ytj

'(6a;2

<

*i

+ u^) dt.

where

Find the compensation

to

Consider the behavior

3/2.

optimize the criterion

+ u^]dt.
OWo

d{t)

y-^
step

-t-

-N

-e(t\
-e(t)

Fig. 10-12

M(f)
u{t)

1^

y( t)

CHAP.

10.25.

10.26.

INTRODUCTION TO OPTIMAL CONTROL

10]

In Problem 10.6, show that if C = B = I in equation {10.26) and


(10.27), then the system becomes identical to the model.

Use the

results of

Problem 10.4

What happens when

10.27.

if

229

R =

and

to generate a root locus for the closed loop

Q=

in equation

system as a varies.

0?

Given the linear time-invariant system

dn/dt

= Ax + Bu

with criterion

2c

(x^Qx

+ u^Ru) dt

where Q and R are both time-invariant positive definite symmetric matrices, but Q is to X w
and R is m X m. Find the feedback control u(x, t\ <j) that minimizes v in terms of the constant
M X n matrices Kn, K12, Kji, K22 where

Sn
in

which the 2M-vector

^i'^2

'

satisfies

g2n

the eigenvalue problem given

have negative real parts and X + i, X+2,

?^n

Sn +

.,

X2n

in

equation

(10.18),

where

have positive real parts and are

distinct.

10.28.

Given only the matrices G and F as defined in Theorem 10.9, note that G and F are in general
complex since they contain partitions of eigenvectors. Find a method for calculating n = GF"!
using only the real number field, not complex numbers.

10.29.

(a)

- i+

Given the nominal scalar system d^/dt


"^

[(i?^

1)^^

li^]

dt

for constant

i;

optimal closed loop system that are valid for


given in Chapter 10 and another in Chapter
(6)

Now

=
>

>

1(0),

and the performance criterion

Construct two Liapunov functions for the

1.

all

Jo

7;

>

1.

Hint: One method of construction

is

9.

+ ? + ^. Using both Liapunov functions obtained


can be such that the closed loop system remains
In other words, find a function /(,, i) such that e < f(ri, |) implies

suppose the actual system

in (o), give estimates

with

11

is

d^/dt

upon how large

ejs

asymptotically stable.
asymptotic stability of the closed loop system.

10.30.

Given the system dx/d

A(t)x

-f-

B(t)u with x(to)

xr(ij) x(ii)/2

and with criterion function

Xq
'

+ r uTu dt/2

(a)

What

(6)

Given the transition matrix *(*,


*(<, t) where B
5*(t, T)/et - A(t)*(t, t) and *(t,
canonical equations in terms of *(*,
9(t, r)
t) and find
f
a feedback control u().

(c)

Write and solve the matrix Riccati equation

are the canonical equations and their boundary conditions?


t)

I.

Solve the

INTRODUCTION TO OPTIMAL CONTROL

230

Answers

Supplementary Problems

to

= -9x

10.9.

10.10.

(a)

dxjdt

dx^ldt

= - VI X2

dpi/dt

dp2/dt

= -pi +

(6)

Xiit)

(c)

(a)

X2
P2

4x1

(2xiQ

a^i(O)

x-^q

2(0)

^^20

Pi()

P2()

y/5p2

+ X2o)e-* -

(xio

+ X2o)e-^*

P2{t)

= -&Xio + 2o)~' + 2(a;io + a;2o)e-2t


= (^/5-l){2xlo + >2o)e-*+{4-2^/5)(XM + X2o)e-^*

Piit)

Xiit)

10.11.

[CHAP. 10

M(t)

4(2a;jo

+ X2o)e-* -

(l-\/5)(2a;io

2(xio

+ a;2o)6~2'

+ a;2o)e- +

(2\/5-4)(a;io

= 4-^2^

(6)

011

2,Si2

6
3

~ V 5 012

P()

+ a;2o)e-2

2\/5

-/S

012022

,622

0^2

id)

Fig. 10-13

10
''''

0'

^2

di\ p.

-q

0-1
10.13.

(a)

P(t)

[S{1

-1 U

a'a

II

Pi

with

lim P(t)
t-* 00

10.14.

K(t)

=
(a

10.15.

Any

10.16.

M(9)

5, XjCO)

=;

-2vVTp

-3, p^iT)

0,

p2(T)

0.

0/ \ P2/

+ V2 + l]e2 v^(T-t) -i-S{l-y^)


)

[V2-I +S]e2V^cr- +
(6)

a;:i(0)

= -1 + V2

regardless of S.

pgyUi-t) - 1)
- 7^2+7 )e>'"i- + Va^ + P +

control results in

= -A-HQrS +

9r9V2

+ V2 -5'

where y

since there is

g^g^e2

lim

an unstable uncontrollable

qrq^eyS)

where

A =

g9

/:(*)

-a- Vo^Tp

state.

g.ffS/S

9r9eV12

CHAP.

INTRODUCTION TO OPTIMAL CONTROL

10]

10.17.

This

10.18.

10.19.

uop

the diagram associated with Problem 10.8.

is

= Va2 +
=

62g/r

RJ"'b^o'j/p where Pq depends only on the ratio

-dPo/dt

10.20.

Let

#(t,to)

- 4

(tant

to)E*{t, to)

tan

g^5

and tends

10.24.

10.26.

10 27

K=

_|_

to

At a =

= (_i

''''*^"

q)-

*(<o. to)

Note

For

2^5

as

<

to

<

and

fj

to

<

tj

<

0,

P(t)

>

0.

For -ytj

10.30.

and

d#/dt

<

to

<

<

P{t)

ti,

<

tends to yt^ from the right.

i^Kji

p((i)

Let the

KiA-VeA^"-'!'
K22y >^

/^"

=
=

and solve for

e^^'^i^.

\/Kn

K,A/x(ti)
K^i)\-p(ti)

x(f).

Then

p(t)

P(t) x(f).

2m n

first
1, 2,

and

eAu(t-t,)y^K2i

terms of

x(ti) in

e''''^^^

eigenvectors be complex conjugates, so that


Define

m.

f2i-i

10.29.

the closed loop system poles are at

0,

(''^^^\

for

- Fohm^^hFov/p

Vp(t)y

10.28.

where

PflA

6tJ(-6tif6

P{t)

+ ATPo +

:r-^]x

10.23.

-E

Qo

EH)*(t, to) = 0, so *(t, to) = 0.


This shows that for every increasing function
there is a corresponding decreasing function of time.

of time in

*'r(,

as

i)/p

*T(t, <o)(HTE

10.22.

231

= (Refi|Imfi|Ref3|Imf8l...|Imf2,-i|f2ml
Then F and G are real and n = GP-i.

and

(a)

From Chapter

similarly.

<

i){-2

10,

Vi

(,

+ 1)^2

and from Chapter

9,

Vg

... |fn)

i^^V-

for both V^ and V^.

(6)

(a)

dx/dt

(c)

obvious from

- Ax - BBTp

(6)

and dp/df

above.

= -A^p

with

x(to)

xo and p(fi)

f2i

x(fi).

aiid

821-1

821

INDEX
Abstract object, 2-6
uniform, 14
Adjoint of a linear operator, 112
Adjoint system, 112-114
Adjugate matrix, 44
Aerospace vehicle, state equations of, 34
Algebraic
equation for time-invariant optimal
control, 214
separation, 178
Alternating property, 56
Analytic function, 79
Anticipatory system. See State variable of
anticipatory system; Predictor
Associated Legendre equation, 107
Asymptotic
behavior of optimal closed loop poles, 226
stability, 193
state estimator (see Observer systems;
Steady state errors)

Component. See Element


Contour integral representation of /(A), 83
Contraction mapping, 92-93
Control
energy, 210
law, 208-211
Controllable form of time-varying
systems, 148-152
Controllability,

128-146

of discrete-time systems, 132, 136


of the output, 144
for time-invariant systems with distinct
eigenvalues, 129-130
for time-invariant systems with nondistinct
eigenvalues, 131, 135
of time-varying systems, 136-138, 142

Convergence of the matrix exponential, 96


Convolution integral, 109
Costate, 210
Cramer's rule, 44
Criterion functional, 209

Basis, 47
Bessel's equation, 107

Dead beat

control, 132
Decoupling. See Pole placement of
multiple-input systems
Delay line, state variables of.
See State variable of a delay line
Delayer, 4, 17, 164
Delta function
derivatives of Dirac, 135
Dirac, 135
Kronecker, 44
Derivative
formula for integrals, 125
of a matrix, 39
Determinant, 41
by exterior products, 57
of the product of two matrices, 42
of similar matrices, 72
of a transition matrix, 99-100
of a transpose matrix, 42
of a triangular matrix, 43
Diagonal matrix, 41
Diagonal of a matrix, 39
Difference equations, 7
from sampling. 111
Differential equations, 7

BIBO

stability. See External stability


Block diagrams, 164-169

Bounded
function, 191

input-bounded output stability


(see External stability)
Boundedness of transition matrix, 192
Butterworth configuration, 226
Cancellations in a transfer function, 133-134

Canonical
equations of Hamilton, 211
flow diagrams (see Flow diagrams, first
canonical form; Flow diagrams,
second canonical form)
Cauchy integral representation, 83
Causality, 2-6
Cayley-Hamilton theorem, 81

Characteristic polynomial, 69
Check on solution for transition matrices, 114

Closed

forms for time-varying systems, 105


loop eigenvalues, 217
loop optimal response, 217
under addition, 46
under scalar multiplication, 46
Cofactor of a matrix element, 43

Differentiator, state of.

See State variable of a differentiator


Diffusion equation, 10
Dimension, 49
of null space, 49
of range space, 50
Disconnected flow diagram, 129
Discrete-time systems, 7

Column of a matrix, 38
Commutativity of matrices, 40
Compatible, 39-40
Compensator
pole placement, 172
integral plus proportional, 219-223

from sampling. 111

Complex conjugate transpose matrix, 41

233

INDEX

234

Distinct eigenvalues, 69-71


Dot product. See Inner product
Duality, 138-139

(Jram-Schmidt process, 53-54


(Jrassmann algebra. See Exterior product
CJronwall-Bellman lemma, 203
CJuidance control law, 121-122

Dynamical systems, 4-6


Eigenvalue, 69
of a nonnegative definite matrix, 77
of a positive definite matrix, 77
of similar matrices, 72
of a time-varying system, 194
of a triangular matrix, 72
of a unitary matrix, 95
Eigenvector, 69
generalized, 74
expression for H = GF-i, 216, 229
of a normal matrix, 76
Element, 38
Elementary operations, 42
Ellipse in n-space, 77
Equilibrium state, 192
Equivalence of pole placement and
optimal control, 220
Equivalence transformation, 148
Estimation of state. See Observer systems
Existence of solution
to matrix equations, 63
to Riccati equation, 213
to time-invariant optimal control
problem, 214
Exp At, 101-111

Ilamilton-Jacobi equation, 220


Ilamiltonian matrix, 211
Harmonic oscillator, 193
Hereditary system.
See State variable of hereditary system
Ilermite equation, 107
Ilermitian matrix, 41

Exponential stability.
See Stability, uniform asymptotic
Exterior product, 56
External stability, 194

Integral plus proportional control, 219-223


Integrator, 16, 164

nonnegative

Hidden

Flow diagrams, 16-24

of Jordan form, 21-24


of second canonical (phase variable) form, 20
for time-varying systems, 18
Frequency domain characterization of
optimality, 220, 226
Function of a matrix, 79-84
cosine, 82
exponential, 96-104
logarithm, 94
square root, 80, 91
Fundamental matrix. See Transition matrix
3.13),

44

Gauss-Seidel method, 171


Global

minimum

of criterion functional, 220

stability, 193
Gradient operator, 96
Gram matrix, 64

oscillations in

sampled data

Hypergeometric equation, 107


Impulse response matrices, 112
Induced norm. See Matrix, norm of
Induction, 43
Inner product, 52
for quadratic forms, 75

Input-output pair, 1-6


Input solution to the state equation, 109
Instantaneous controllability and observability.
See Totally controllable and observable
Integral
of a matrix, 39
representation of /(A), 83

Inverse
matrix, 44

A),

computation

of,

102

system, 166

Jaeobian matrix, 9
Jordan block, 74

Jordan form
flow diagram of, 21-24
of a matrix, 73-75
of state equations, 147

of first canonical form, 19


interchange of elements in, 17

Gauss elimination (Example

77

systems, 132

of (si

Factorization of time-varying matrices, 152


Feedback optimal control, 211-231
Finite escape time, 104-105, 204
First canonical form, 19, 20
for time-varying systems, 153
Floquet theory, 107-109

definite,

positive definite, 77

Kronecker

delta, 44

I^agrange multiplier, 210


I^aguerre equation, 107
Ijaplace expansion of a determinant, 43, 57
Least squares fit, 86
Left inverse, 44
Legendre equation, 107
I^everrier's algorithm, 102
Ijapunov function
construction of, for linear systems, 199-202
discrete-time, 198
time-invariant, 196
time-varying, 197

Linear
dependence, 47
independence, 47
manifold, 46
network with a switch, 11, 12

INDEX
Linear

(cont.)

235
Nonexistence

system, 6-7
vector space, 50
Linear operator, 55
matrix representation
null space of, 49
range of, 49
Linearization, 8, 9

of solution to the matrix Riccati


equation, 221, 228
of trajectories, 10
of,

55

Logarithm of a matrix, 94
Loss matrix. See Criterion functional

Lower triangular matrix, 41

Nonlinear effects, 179-181


Nonnegative definite, 77
Nonsingular, 44

Norm, 51
of a matrix, 78

natural, 53

Normal matrix, 41
Normalized eigenvector, 70

Mathieu equation, 107-109


Matrix, 38. See also Linear operator
addition and subtraction, 39
characteristic polynomial of, 69
compatibility, 39
differentiation of, 39

eigenvalue of, 69
eigenvector of, 69
equality, 39
function of, 79-84
fundamental, 99-111
generalized eigenvector of, 74
Hermitian, 107
Hermitian nonnegative definite, 77
Hermitian positive definite, 77
integration of, 39

Jordan form
logarithm

of,

of,

73-75

94

multiplication, 39

Null space, 49
Nullity, 49
n-vectors, 38
Nyquist diagrams, 171

Observable form of time-varying


systems, 148-152
Observability, 128-146
of discrete-time systems, 133-138
of time-invariant systems, 133-136
of time-varying systems, 136-138, 146
Observer systems, 173-177
Open loop optimal control, 211
Optimal control law, 210
Order of a matrix, 38
Oriented mathematical relations, 2-4
Orthogonal vectors, 52
Orthonormal vectors, 53
Output feedback, 217-218

norm

of, 78
principal minor of, 77
rank of, 50
representation of a linear operator, 55
Riccati equation, 211-231

square, 38
state equations, 26, 27
transition, 99-111

unitary, 41

Matrizant. See Transition matrix


Mean square estimate, 94
Metric, 51
Minimal order observer, 175-177

Minimum
energy control (Problem 10.30), 229
norm solution of a matrix equation, 86
Minor, 43
principal, 77

Missile equations of state, 34

Model-following systems, 224-225

Motor equations of

state, 34

Multilinear, 56

Multiple loop system, 185-186

Parameter variations, 170, 171


Partial differential equation of state
diffusion, 10
wave, 141
Partial fraction expansion, 21, 22
of a matrix, 102
Partition of a matrix, 40

Passive system stability, 205


Penalty function. See Criterion functional
Periodically varying system, 107-109
Permutation, 41

Phase plane, 11
Phase variable canonical form, 20, 21
for time-varying systems, 153-4
Physical object, 1
Piecewise time-invariant system, 106
Polar decomposition of a matrix, 87
Pole placement, 172-173, 220, 224
Poles of a system, 169
Pole-zero cancellation, 134
Positive definite, 77
Predictor, 5
Principal minor, 77

Pseudoinverse, 84-87

Natural norm, 53
of a matrix, 79

Neighboring optimal control, 221


Noise rejection, 179-181

Nominal optimal control, 221


Nonanticipative system. See Causality

Quadratic form, 75

Random

processes, 3

Rank, 50
n(e.d.),

Range

137

space, 49

INDEX

236
Reachable states, 135
Reciprocal basis, 54
Recoverable states, 135
Regulator, 208
Representation
of a linear operator, 55
spectral, 80
for rectangular matrices, 85
Residue
of poles, 21,22
matrix, 102-103
Response function. See Trajectory
Riccati equation, 212-213
negative solution to, 221
Right inverse, 44
RLC circuit, 2, 6, 11, 35
stability of,

Root

205-207

locus, 169

Row of matrix,

38

Sampled systems. 111


Scalar, 38

equation representation for time-varying


systems, 153-154, 160
product (see Inner product)
transition matrix, 105
Scalor, 16, 164
Schmidt orthonormalization.
See Gram-Schmidt process
Schwarz inequality, 52
Second canonical form, 20-21
for time-varying systems, 153-154
Sensitivity, 179-181

Servomechanism, 208
problem, 218-220

State (cont.)
reachable, 135
recoverable, 135
State equations
in matrix form, 26, 27
solutions of, with input, 109-110
solutions of, with nonzero input, 99-109

State feedback pole placement, 172173


State space, 3
conditions on description by, 5
State transition matrix, 99-111
State variable, 3
of anticipatory system, 14
of delay line, 10
of differentiator, 14
of diffusion equation, 10
of hereditary system, 10
of spring-mass system, 10

Steady state
errors, 165
response, 125
Sturm-Liouville equation, 197
stability properties of, 198

Submatrices, 40
Sufficient conditions for optimal control, 220

Summer,

16,

164

Superposition, 7
integral, 109

Switch. See Linear network with a switch


Sylvester's criterion, 77

Symmetric matrix, 41

Symmetry
of eigenvalues of H, 215
of solution to matrix Riccati equation, 213

Similarity transformation, 70-72

Symplectic, 228

Simultaneous diagonalization
of two Hermitian matrices, 91
of two arbitrary matrices, 118
Skew-symmetric matrix, 41

Systems, 7, 8
dynamical, 5
nonanticipative, 5
physical, 1

Span, 46
to go, 228

Spectral representation, 80

Time

Spring-mass system.
See State variable of spring-mass system
Square root of a matrix, 77

Time-invariant, 7
optimal systems, 213-217

Stability

asymptotic, 194
BIBO, 194
external, 194
i.s.L., 192
in the large, 193
of optimal control, 214
of solution of time-invariant
Riccati equation, 215
of a system, 192
uniform, 193
uniform asymptotic, 193
State

Time-varying systems
flow diagram of, 18
matrix state equations

of,

25-26

transition matrices for, 104-111


Totally controllable and observable, 128

Trace, 39
of similar matrices, 72
Trajectory, 4, 5
Transfer function, 112

Transition
matrix, 99-111
property, 5
Transpose matrix, 41
Transposition, 41

of an abstract object, 1-3


controllable, 128
observable, 128

Triangle inequality, 51
Triangular matrix, 41

of a physical object, 1

Type-l systems, 167

INDEX
Undetermined coefficients, method of.
See Variation of parameters, method of
Uniform abstract object.
See Abstract object, uniform
Uniform stability, 193
asymptotic, 193
Uncontrollable, 3
states in optimal control, 223

Uniqueness of solution
to
to

ATp + PA = -Q, 204


matrix equations, 63

to the time-invariant optimal control, 215

237
Unity feedback systems, 165
Unobservable, 3
Unstable, 192
Upper triangular matrix, 41

Van

der Pol equation, 192

Vandermonde matrix, 60
Variation of parameters, method of, 109
Variational equations. See Linearization
Vector, 38, 46, 50
Vector flow diagrams, 164
Voltage divider, 1, 3

for the square root matrix, 91

Unit

Wave

matrix, 41
vector, 41
Unitary matrix, 41
for similarity transformation of a
Hermitian matrix, 76
Unit-time delayor. See Delayor

White noise input, 208


Wiener factorization, 216

equation, 141

Zero matrix, 41
Zero-input stability, 191
z transfer function, 18