Wiberg-StateSpaceLinearSystems Text PDF

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.
vsâs 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 = [xîto) + 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<^-''>^'û(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)
xît^
y = u cos x.
+ P'x'ziio) and
y{t)
except in special cases like
Then
0,
ai(to)
+ I3u2(t)]
[auiit)
Xi{tg)
xîti^
2/i(t)
cos [aXi(to)
Ui(t) cos x^Hq)
Px^Ho)]
system
so the
0,
and
j/2(*)
'*2(*)
cos xîtQ).
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 = ô(*)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)
^îk) 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
-^prîto)
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 î(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 xît) = 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) = î(t)/3 for t ^ t,. Siniilarlyj/ajt)
=
t,^t<t,
for
u,(t)/2
and
is [au,(t) + pu2{t)]/2 for
output
The
puît).
+
auît)
input
Now assume an
uJt)/2 or uM)/S.
Substituting y^{t) and yît), 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~'û(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)
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.
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-
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.
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-û
-!--+- i8_jPM
+ Pji
{2.3)
This can be rewritten as
PHV -
+ P""'(i2/ -I3û)+
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-û-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
+ pû
{2.5)
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.
^ /^^
î-A
X2
(2.6)
a;n-i
'1
0/ \x.
^-ô
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.
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
ô
+
,
+ 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, ~ ^ô)^r^ + + (^_i - -iô)\ +(^n-nô)

(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,
22
[CHAP.
M(t).
-j/()
-Ô-[>
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;
ô
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.
pÛ
pÛ
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]
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
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
Ê>
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.
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).
'î
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(*)^
lôit)-^
+ 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
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 " î)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.
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)
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/aî
-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.
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
iô
^"(0)
+ ^yo-û(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-ô]
3^0
Vo
summer:
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.
2]
Given the
linear, time-invariant, multiple
+ 2) +
V^{n
+ 1) +
a^yîn
Vln +
Put
this in the
1)
y^v.J,n
+ 1) +
y^yj^n)
(iû^(n)
^û^(n)
y^yln)
a^v^{n
+ 1) +
cc^y^{n)
^û^n)
hû^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^^ TÂ.
*^"
^''"^ delayers with arbitrary Initial conditions are not

needed because a
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
xîn
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 -
xîn)
aîvL) 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 = î(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.
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.
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
îM2
^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
rnîg.
2-21
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:
Kî
{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
yî
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,
Kâ
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.
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
1\
-2
-2
-3)^^'^'^
0\
"*"
/0\
ll
'
/1\
'*^"-^'
^^'^^
(1
O)x(n)
(3
l)x(n)
(-1 2)x(n)
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
^>
î
'*''
^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.
2]
(li)
The values of
(iii)
XiiO)
I/O
vq, y^
and yj are given by equations
YoiO) u(0)
a;2(0)
ô
(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.
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.
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.
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,
.,,
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)âi3a22a3i
(-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.
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 =
Eî, 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
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^TÔ.
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.
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
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
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^ + pâ^ +
Pv Pv-'Pn not all zero we get
+ /3â. 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
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 wôrd 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.
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.
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'Â = rank AA'^.
matrix A.
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
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.
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|].
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
IKî
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 î-
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
= Â^^
''^
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^
yâ^
the second vector must be broken up
54
[CHAP.
the component of a^ along bj, and a^ y^^ is the component perpendicular to

Finally, the second
72' 'ôte 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).
Pjjî.K)
CHAP.
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)]
[axît)
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{î)
P^^'-
Furthermore since
L(bi)
is
a vector in V2, determine
y^j
such that
1=1
y,iCr
But
/ n
L(x)
l(xâ)
\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 = (î 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.
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 =
aiâ^
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
^"<^
î
J^ 2
,2
T'''"
afcr,(bfc
*^^*
b.)
is
a linear com-
CHAP.
Since
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
lî<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 viîth
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 + aê^
anien so that
Aa
aiieiAa2A
Aa+
a2ie2Aa2A
Aan+
++
ttniCn
Aa2A-
Aa
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.
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-pî
'
'
"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
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
" î)
For n = 2, det V = fig "" û 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)
Â
\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.
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 +
îiai
/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
.
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+iî
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^î
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
Îr
^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.
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
îX;
'"'
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;
ôr
"ijî
= ^ + !,...,.
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
îî
îx^
+2
(.2
l=r+l \j=l
-l
i,a,) X,
(x^
i=l y
= r+l
fc
aîx^
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:
Âi
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
îî-
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
îVi =
('7i
<=i
1=1
i=i
The solution is unique if and only if aj, ...,a are linearly independent.
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^Â = 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 AÂ, note both
in the null space of A,
0,
null space of
rank AÂ.
3.13.
A and AÂ have n columns. Then consider any vector

Then AÂy = 0, so that y is also in the null space of AÂ.
space of AÂ, i.e. AÂz = 0. Then zÂÂz = \\Az\\\ = 0, so
To show rank A
y
Ay =
i.e.
0.
z is also in the null
i.e.
AÂ, 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 rankBÂT
gives rank AB = rank BÂi", 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
=
Mû^j)
still
gives zero, so that
\i=l
î'xi,
3
2
=
/3*Xj
/
<,
where g
is
a column
CHAP.
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
yMuî)
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^Lîbi) =

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
î<=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'*"
66
[CHAP.
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-jÂ-i.
-AirAi2\MiV
V"
and A22
"
'
-,
(A22-A2iAn*Ai2)-VV-A2iAii
**
/An'
(
A^
"All A12A22
^-1
- A21A11
Ajj
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
kiî
J
^ t=l 3 = 1
- AI'
the remainder for the inner product defined as (a,b)
a*(.t) b(t) dt"!
68
Answers
2a
JT
aj
a;2
sin
[CHAP.
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 wîth 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^^xîît)
a2ii()
a22a;2(*)
aîXid)
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-ÂT 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
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.
.
î
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-ÂT 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
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
MÔ =
...
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 %î^}>
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:
= iXîQii + Qmî
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
= XXQiAî
=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-'Ût
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
<
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
which
det
gn >
is
Q>
and either
qnQzz-lfz >
and
^22
922 or
>
and
911922
ût
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
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
\\ÂS,\\1
t=l
Since
(pf)^;^
2=
the theorem.
4.7
î
2 ihf
2
û
=
(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 (pâx-
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'^)\HÎ1
+ \H^lk\ +
MATRIX ANALYSIS
80
Theorem
Note
Proof:
Theorem
T-ÂT =
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
+ YÂ"-^ +
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
(Î-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'Â
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'Â 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'Â; 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
pî^
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 AÂg. = pfg. for i = l,2, ...,r and A^'Agj =
of A'Â and
Proof:
m-vector
h.
as
h.
Ag./p^
for
AA'-'h,
1, 2,
. ,
r.
AAÂg./p,
.,n.
AA^.
Define an
Then
A(p2g.)/p,
p^g^-g/Pj
pfh,
Furthermore,
hfh.
gJAÂg./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^Âg; = 0, so that

AÂg^ =
= 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,)
(AÂgJ/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
AÂ
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 ô^'"
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 IÂg,
i=l
i^
Also,
For arbitrary x
i,Ag,
i=r+l
1=1
i.pî^
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 wîth 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(îPi-'?/
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.î)
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~Â)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-Â)z
Some further
properties of the pseudoinverse are:
nonsingular, A~'
1.
If
2.
A-'A
3.
AA-Â = A
4.
A-ÂA-^ = A-^
5.
(AA-^)''
6.
(A-'A)^
7.
A-Âw =
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-Ût where A"' can be found from
15.
. ,
where Q-i
.... A"')
may
Xi
be zero.
0.
U-i.
Then
= H-Ât =
(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 (AÂ, 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^DÛ^ 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.
normalizedêigenvectors 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
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;îq >
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^î^.
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-Â =
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
"
-Â"'-
Pifig-
(^.26),
(2
gigf)^
and
(A-iA)T
Similarly
(AA-O'^
= (?
îî)'^
To show uniqueness, assume two

1.
2.
AXA = A
XAX = X
and transposing
9.
+ j)
A-ÂA-' = 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^XÂ^ = AT
10.
ATYÂ^ = 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 = ATYÂTX^'X = A^YÂX = A'^'Y^X = YAX
= YAYAX = YAYXÂT = YYTATXTA^ = YYÂT = 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 - ô)"
=
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
~ î ~ /^*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.
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^xÂx
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)"^
gradâ
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,
eî
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. ^'*''
(AÂ +
eP) - 1 A'',
positive definite
Show
4.45.
||A-J||
if
and only
Answers
Because
4.12.
det
(A- XÎ) =
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
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 =
â
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~-^
+ aÂ.^-^ +
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)
*(ô, 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)
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^â.^^Â
*lA
A^
ai\,{t)
a,î A
is
continuous
A ^j A
A ^
{5.13),
[tr
A(t)]î 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.
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'^+î'^/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.
^ AHÎk
Cayley-Hamilton:
eî'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
102
where
Ixl,
if Lji(Ai) is
[CHAP.
=1
Mî)
'
(5.20)
I
\
...
,0
\il
then
...
4.
e'î*
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
îI
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"
îs-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.
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),
Ûm)
Lî(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
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'.
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)dÂ{t)
A{t)
{5.86),
A(i)e-^-'''""*
This equation and {5.37) are equal

A{t)
=
if
A{t)
and only
^Â{r,)d^
A(i)
A(r?)d,
if
J\{r,)dr,A{t)
(5.37)
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.îo)
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^tÛ
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 xît) = 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.
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, viîth 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,
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 î,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
cîi'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
""^^"Î 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,
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)
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)
vsîth
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
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)
DÂ;)
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,
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(ô)
parts.
f\Â{t)xdt-
X
p(î),
[''^^^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.
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
aît,to)/dt ==,9Â^{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)
î(*'*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)
4ît,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
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.
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
aîo
<!(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,
'
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
aâo
= 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.
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{) acâ
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)
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
eÎ
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.
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)
''^^'''"-*'~'^
ê-it-r)(
<'t
119
e~t)
cos{e^
e-t)J
\cos(e^
e't)
sm{eê-*)
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
iô)
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 eô),
/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
aê^tx^)
{e"t
dx^/dt
aea;i)
te^^Xi
which can be rewritten as

tdê<^*x^)/dP'
d(ea;i)/df
Tiis has the same form as Bessel's equation
xît)
= dxjdt =
(5.38), so
-ax^(t)
te^tx^
the solution
e-''t[-c^Ji(t)
is
c^dYa/dt]
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-î(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 Yîf) ~ îlvi 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
AÂo.
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.
5]
Changing variables from
r to
y(t)
ij
where e-'ê-*
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
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(%)]
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.
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.
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.
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>-êM'.
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.
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
CAvÛin) = [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^ît)
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
xît)
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 =
126
Answers
=
A(k)
5.14.
If
5.15.
deteAU-T)
5.16.
A =
5.17.
(f.r)
= ^êt-rf^\(l
and only
'':
5.19.
eAt
5.20.
71
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
aiîr) aîit)
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,
5]
a-HT)dT
5.34.
where
*(to,
5.37.
/cosh T
*(t,to)
eî
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
^'^^^Â
\~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.
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 xît^) - 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
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 î 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: Mê-^^^-'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'^""'-^', eîi-^>).
Note that because

e'î"!-^)
Now we can consider what happens when A is not restricted to have
distinct eigenvalues.
CHAP.
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 wîth 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 [Lî(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
^
Therejôre
fr,z, (t)
- f^Zr, (t) =
(1)
must
// part:
To
+ fvU)
Xi\Jy)Zv
fy (Ai2;
+ f-rÛ)
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) = {xâ ~ 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.
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
/tjî'
^^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
+ \'"-û{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'û, when the system is stationary
and A has distinct eigenvalues. To avoid complexity, first we consider scalar outputs and
distinct eigenvalues.
6.1,
CHAP.
()
Theorem
133
Given a scalar output y{t) to the time-invariant system dx/dt = Ax + Bu,

y = c+x + d'û, 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(ô) = 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-ÂT
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'û, where A
and only
(1)
Theorem
matrix
Given a scalar output
6.5:
The proof
Fig. 6-5
ctx(m)
+ d'û(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.
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.
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(î)
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
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.
6]
137
Now suppose b{t) = for t^ < t ^, and we wish to reach a final state x(t^. The state xitî 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,<) vÔ 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,
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(î) 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,
êa(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.
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.
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
(iî)!_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.
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''î(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
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.
6]
143
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 iît-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
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.
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.
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
can be reached from
iit^)
dxjdt
all
|(to)
so the system is uncontrollable.
0,
states belonging to the eigenvalue 2 are unobservable

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
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 shovsêd 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
viîthin
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/*
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)
(TiQî TiF)
(Ti(),T2(i))(^"
Ti(Qî F)
Since Ti(i) is an % x ui differentiable matrix with rankni everywhere and Qî

matrix with rank ni(e.d.), the only if part of the theorem has been proven.
is
an ni x mni
150
For the proof
[CHAP. 7
in the other direction, since Q," factors, then
Qn*)
{v^{t) yît))
(^^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.
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 viîll 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) Vî) = 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 GJOÛx +
(P5 PJ, GJ, GJ2)Û2
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.
152
= R2U2 and Q^-ViSi =
V3S8 to obtain RÛ2,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\
\î
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-"-"Â>
7.2.
a(t)
Choose
pj.
yîX^)
_
-
>.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 Aê^ = 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)
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
lî^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 = -QÂ*" when written out in its columns gives
Proof:
T=
The transformation matrix
Iq^Wn+i)
(q|lq|
Q-Q^-',
(qi
q^ -QÂe)
solved recursively
Therefore q^+i = -Q"A'ê, 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==Â'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
Âî
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
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
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.
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
...
-vî
^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 vî 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
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.
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)îW + (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
160
- -a^ =
where q^
Then
^ =
from Example
a^,
7.7.
A^-Cn
- dq^'M) we
Qî-i(Aîqî
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)
iÔ
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)\
âît)/
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 =
zî.
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.
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>
'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^êd if T(t) can be nonsingular
differentiable for times t everywhere dense in [to. *i]'
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.
*î
/
[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.
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.
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
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
rî
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.
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
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 tÛ 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.
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-Û{t - to)ei for all i = 1, 2,
called
is
for the inputs
uît),
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.
-vîth
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
<
>.
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.
'-'"
'-*
ResÔ.
[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.
8]
169
and where
limsR-i(s)
wo
limsf
\-l
- \6s-^J
s-o
Since lim sRî(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.
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.
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
Ûi{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.
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)
...
...
-ccjî)
=
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" + ffÂ"-+
haÂ-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.
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 + kî + ^23 + 5)X2 + [fcii + 2fci3 + 3^21 + 4fe22 + S/cga + fei3(fe2i + ^22) " îzihi + ^12) +
kî 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.
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.
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
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.
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
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.
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
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.
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
Xfît)
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).
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,
-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.
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 î 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 yâX
/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
"
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.
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.
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
êw(<)
'
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)
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,)-' ô 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-ôK
(i2)
:f)
\-P1P2
-PL = MPi + P2 - 'ô) -
and
"
-~Pi
=
P2
PiP2]2
Therefore the estimator dynamics are
A
f
dt \a2
To construct the
-âl) +
+ P2-ô)-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
'î^z
xi
'
1+
U C)*~
- Pi - P2
1/
Fig. 8-24
-I-
Pi)(s
'^0*Pl
-I-
P2)
+ P2 - To) -
P1P2
PA
CHAP.
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
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
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
tê/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.
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.
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>
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
dxî/dt
for all
+ l)(s + l)
"
+ t + /(t))Nl ~ u{t)xi^
f(t)Xfî fiXffi
t.
pir)e{r)dT
~/it)K
where
an arbitrary constant,
+ G(s)H(s))-i]s^{d} =
H(8)
-^{"(*)>
4(s
:")".(::;,./
^Xfî
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
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.
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
^xît)
(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.
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)
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
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.
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,)â(||x|j)
(4)
Note for
Theorem
all
norm
9.7:
dv/dt
v(x,
t)
to,
0;
>0 for xvÔ 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 = xê~* 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'ê-*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
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
XvÔ;
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:
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
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
Kê^'î'"-'''
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)vÔ.
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
=
XÂxdr
>
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
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.
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
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 <
î^
-Ct0+(.k+l^)T-t0^)/T
p-k
Then
I*(,*o)ll
Defining
-(fc+i)
<
from
CHAP.
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.
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.
llôll^'""'''^'"'"'-
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*,^ = ^' '^î<=^
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.
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
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
9.9.
dx/dt
Given the scalar system
tx
+u
and y
x.
\h{t,T)\dT
^ V2ê'
'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 aîC*) 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 aâo 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 ^Jrîôl'^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.
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)
^ ^ê'^^^
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
iâlways
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).
^^'-^
Â.ZTX'â'
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
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.
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.
"^ ""-' '-' '' ''' --"^'^ --' -^^^^ ^^^ô^^^
r funernTf
x^^x
t-K
iî^rsî 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.
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
+ pÂx + 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:Âp
+ 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
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
LÛ
p(l)
4\
e)
2a;p(l)
^''-''^
gives
10.1,
^'"'(ô.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.
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^-^â=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.
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 alvâys 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^
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
If
Choose a Liapunov function

nonzero x. Then
Proof:
for
all
dV/di
for
all
V = x'Ûx.
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.
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
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-Ô)
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.
10]
Q n^
^îÛA-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^ê^ât
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
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
-RîBTN^^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.
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^ê-^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
Rîb^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
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). ô)
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)-û
+ 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-iBTgradî;
Rearranging the inequality and adding ^x^Qx to both sides,

^xTQx + ûmu ^ ^xTQx - i(giradî;)^R-'BTgrad^v
(i 0.^5)
(gradî;)i'Ax
^(grad^v)rBR-iBTgradî;
u(x,t)
Using
r)Ru<'P(*, t)]
(W^ni-.
= Ax + Bu, we
i^^^-î)
(grad^t;)^Bu
get
+ iumu ^ -[11 + (grad^fHAx + Bu)J =
-^.,(x(t),t)
CHAP.
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(ô)>
*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'ê. 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.
xû
so that
+ xU + Zylxu + Zy^xû-\-xM
{10S5)
[CHAP. 10
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.
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-û^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^xît) 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.
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 Môrds, 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.
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)
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.
10]
227
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^ î^
(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-
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
"^^^^
Irîl)"^
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) = vÛt) 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 ô^
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.
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
î'^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
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
230
Answers
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
aî(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, pîT)
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^gê2
lim
an unstable uncontrollable
qrqêyS)
where
A =
g9
/:(*)
-a- Vo^Tp
state.
g.ffS/S
9r9eV12
CHAP.
10]
10.17.
This
10.18.
10.19.
uop
the diagram associated with Problem 10.8.
is
= Va2 +
=
62g/r
RJ"'bô'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^^'î^.
\/Kn
K,A/x(ti)
Kî)\-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 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
Iâgrange multiplier, 210

Iâguerre equation, 107
Ijaplace expansion of a determinant, 43, 57
Least squares fit, 86
Left inverse, 44
Legendre equation, 107
Iêverrier'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
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

Wiberg-StateSpaceLinearSystems Text PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Wiberg-StateSpaceLinearSystems Text PDF

Uploaded by

Copyright:

Available Formats

SCHAUM'S OVTLIISE OF

THEORY AIVD PROBLEMS

Associate Professor of Engineering

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

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

ticing professionals as well.

of matrices or complex variables

The book may appear

to be comprehensible and/or useful

But the theorems have been stated and proven in a manner

The book has come from a

set of notes developed for engineering course

was given to the publisher in June

write this book.

long hours of involvement with the book.

METHODS FOR OBTAINING THE STATE EQUATIONS

Properties of Flow Diagrams.

ELEMENTARY MATRIX THEORY

Eigenvalues and Eigenvectors. Introduction to the Similarity Transformation.

SOLUTIONS TO THE LINEAR STATE EQUATION

CONTROLLABILITY AND OBSERVABILITY

CANONICAL FORMS OF THE STATE EQUATION

Introduction to Canonical Forms.

RELATIONS WITH CLASSICAL TECHNIQUES

Matrix Flow Diagrams.

STABILITY OF LINEAR SYSTEMS

Definitions of Stability for Nonzero Inputs.

Definitions of Stability for Zero-Input Linear Systems.

INTRODUCTION TO OPTIMAL CONTROL

The Criterion Functional.

Derivation of the Optimal Control

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

completely determines a unique output for

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

Intuitively, the state of the

tion of the switch,

experiment of applying a voltage V to the

STATE OF AN ABSTRACT OBJECT

ideas contained in the above example can be extended to

This will be done after

the totality of input-output pairs that describe the

behavior of a physical object.

of input time functions and their corresponding output time

functions, the abstract object

is usually characterized as a class of all time functions that

relations which summarize an abstract object must be oriented,

with the input

u(i) for all

a collection of numbers which together

physical object is the resistor-capacitor network

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

This summarizes the abstract object.

The pertinent equation

Physically, the voltage

and -^ (tp) parametrizes the input-output

derivative across the smaller capacitor at time

correspond to the state.

1.1 the state variable

The state representation is not unique.

state space, denoted

There can be many different ways of expressing

be the voltage and

the physical object