STATE SPACE
and
LINEAR SYSTEMS
BY
DONALD
M. WIBERG, Ph.D.
Los Angeles
H^
SCHAIJM'S OUTLINE SERIES
McGRAWHILL BOOK COMPANY
New
York,
St. Louis,
Montreal, New Delhi, Panama, Rio de Janeiro, Singapore, Sydney, and Toronto
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 known to mathematicians for many years. As engineering, physics,
medicine, economics, and business become more cognizant of the insight that the state space
approach
offers, its
This book
popularity increases.
not only for upper division and graduate students, but for pracIt is an attempt to bridge the gap between theory and practical
use of the state space approach to the analysis and design of dynamical systems. The book
is meant to encourage the use of state space as a tool for analysis and design, in
proper
relation with other such tools. The state space approach is more general than the "classical"
Laplace and Fourier transform theory. Consequently, state space theory is applicable to all
systems that can be analyzed by integral transforms in time, and is applicable to many
systems for which transform theory breaks down. Furthermore, state space theory gives
a somewhat different insight into the time behavior of linear systems, and is worth studying
for this aspect alone.
vs^as vi^ritten
In particular, the state space approach is useful because: (1) linear systems with timevarying parameters can be analyzed in essentially the same manner as timeinvariant linear
systems, (2) problems formulated by state space methods can easily be programmed on a
computer, (3) highorder linear systems can be analyzed, (4) multiple inputmultiple output
systems can be treated almost as easily as single inputsingle 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
to contain too
is
prerequisite.
many theorems
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 highorder timeinvariant systems, and contrasts this with a detailed development of transition matrices for timevarying 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 timevarying 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 timevarying, multiple inputmultiple 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 timeinvariant problem and its associated computational techniques.
Since Chapters 6, 8, and 9 precede, relations with controllability, pole
placement, and stability properties can be explored.
originally dating
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
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
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
16
Plow Diagrams.
Chapter
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
Chapter
99
Transition Matrix.
Calculation of the Transition Matrix for TimeInvariant
Systems. Transition Matrix for TimeVarying Differential Systems. Closed
Forms for Special Cases of TimeVarying Linear Differential Systems. PeriodicallyVarying 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
&
128
CONTENTS
Page
Chapter
147
Chapter
164
Chapter
191
Functions.
Chapter
10
208
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 inputoutput pairs. An inputoutput 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
and
Example
state at time
to
to.
1.1.
Consider a black box. Fig. 11, containing a switch to one of two voltage dividers.
nition 1.1.
1.2
terminal
Fig. 11
and
Ms)
Output
many
physical objects
MEANING OP STATE
An
Definition 1.2:
abstract object
is
[CHAP.
list
The mathematical
in that
m of the time
The
Definition 1.3:
state of
an abstract object
all
is
to
to.
In essence the state parametrizes the listing of inputoutput 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.
J?
by the experiment.
RCdy/dt + y
= u
.(.tattlRC
y(to)e'
(1.1)
The
hf"
solution of (1.1) is
Tfl/RC,
m(t) dr
Fig. 12
{1.2)
This relationship explicitly exhibits the list of inputoutput 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. 13 is
two
R^C^ diy/dt^
RC
networks in
2.5RC dy/dt
Fig. 13
series.
= u
is
{1.3)
CHAP.
MEANING OF STATE
1]
with a solution
y(t)
.^[4e(totV2Rc
gCtom/BCj
rf^,^ sr (ft)/2RC
2_
_2_
[e(/2RC
3fiC,J'
S^
p(tot)2/RCl
^<.t)2/RC](^)
rf^
(^.^)
to
Here the
set of
numbers
j/(to)
and
its
pairs,
and
to
may
be chosen as 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 Kvector 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
remains
all time.
1.5.
state variable in
Example
x(t)
1.3 is
is
1.4.
The
by 2,
1.6.
Example
1.3,
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,
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
MEANING OF STATE
[CHAP.
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 TCdimensional 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
1.8.
Given
is
made by
sin
a;ia;2
and
X2(t)
plane with
cos
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
2a;2
C(2xi
+ Xg)*
where the constant C = [x^ito) + 2a;2(to)]/[2a;i(to) + a;2(*o)]* Typical trajectories in state space are shown
in Fig. 14.
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. 14
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)
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) = uitT). [See Problem 1.1.]
If a
t^
to
CHAP.
MEANING OF STATE
1]
However, we can
+ T).
If
Definition 1.6:
also
we
u{t
dynamical system
(1)
A
all
 to
t.
Outputs
(2)
is
y(t)
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:
Condition
x(ti)
>
to
>
to
i.e.
for
all
ti
^ to
{1.5)
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))
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.
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. 14, 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)
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 inputoutput
relationship resulting from the state space description to correspond with the causality of
the original dynamical system.
Example
a^(*)
1.13.
x(t)
a:(*i)e"i"/c
;^J
x(t,)
^(ye"'>''c
^J
where
(1.8)
^ t,
e<^''>^'^u(ro)dro
(1.2).
(1.9)
1.5
(1.1)
has
we will simply
Definition 1.7:
The operators
{x(io)}
is
^(t;
io,
(1)
xo, u(t))
x(i)
and
rt{t;<j>(t;to,xo,u(T)))
= y(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()
X3(to)
and
WgCf)
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.
The
state
X3{t)
x^it^
y = u cos x.
+ P'x'ziio) and
y{t)
Then
0,
ai(to)
+ I3u2(t)]
[auiit)
Xi{tg)
x^iti^
2/i(t)
cos [aXi(to)
Px^Ho)]
system
so the
0,
and
j/2(*)
'*2(*)
cos x^itQ).
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 ac voltage input (constant +
be calculated as the response to a constant voltage input plus the response to an unbiased
ac 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 timeinvariant
Definition 1.8:
lent
system
if
results.
One
Example
1.17.
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.
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 timevarying.
1.6
SYSTEMS CONSIDERED
This book will consider only timeinvariant and timevarying 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 nvector and the system is linear.
Example
1.18.
timevarying linear differential system of order n with one input and one output
,
equation
u^
+ l^n{t)u
+ a^{t)y = ^o(*)rf^ + ^
+
II
o.M^+
is
described by the
iiin\
(1.10)
MEANING OF STATE
Example
[CHAP.
1.19.
timevarying linear difference system of order n with one input and one output
the equation
y{k
The values
t
+ n) +
ai(k) y(k
+ nl) +
ajk)
y(k)
^^ik) u{k
of the process, in a
(fe)
way
+ n)+
/3(fc)
described by
is
u{k)
{1.11)
a{t)
depend on
1.7
section
Comparatively
little
design of systems
methods
is
offer great
=

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)
dyi/dt
1/2
dy2/dt
2/3
can always
Then a set
(il^)
dyni/dt
dyJdt
Example
g{yi,
2/2,
...
2/n,
u,
t)
1.20.
differential equation
d?yldt^
yi and dy/dt
dyJdt
j/2
dy2/dt
2y\
y^.
2y^ + u dy/dt =
to
two
first
Then
wj/a
Suppose a solution can be found (perhaps by computer) to equations [1.12) for some
Denote this solution as the
2/2(^0),
., yn{to) and some input w{t).
trajectory ^{t; w{t), yi{to),
., yn{to), U). Suppose now that the initial conditions are changed:
initial conditions yi{U),
y{to)
where
to u(t)
yiito)
Xi{to), X2{to)
w{t)
dy
Xl{to),
and
Xn{to)
+ v(t) where
Jfito)
y2{to)
are small.
X2{to),
...,
d" ~^v
^pr^ito)
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
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
is
. ,
i+t +  + S+S"
/>....*...') +
/j
=
l/n
M=
Now,
w.
^n ^^^^^
can be canceled to leave
'
t)
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
3/2/31*
\
\dfJdU/
dfjdyn/ \Xn/
which is, in general, a timevarying linear differential equation, so that the nonlinear equasmall
tion has been linearized. Note this procedure is valid only for Xi, X2, ...,Xn and v
matrix
The
neglected.
be
can
series
Taylor's
in
the
enough so that the higher order terms
is called the Jacobian matrix of the vector f (y, u, t).
of dfjdy. evaluated at y^ =
<^j.
Example
1.21.
at toJ= 1Consider the system of Example 1.20 with initial conditions j/(to) = 1 and y(t^ = 1
~*~^ ^^^'^^ fi  3/2.
If the particular input w(t) = 0, we obtain the trajectories ^i(t) = t~^ and ^gC*) =
= 2j/? M3/2, then dh/dyi = 6j/i, df^/dVi   and
then a/i/aj/i  0, 3/i/3j/2 = 1 and dfi/du = 0. Since
= 1 + *i(*o). ^(*o) = 1 + 2(*o) and inputs u  v(t),
=
Hence for initial conditions
5/2/aM
j/2.
we obtain
y(o)
Example
= x^t'^
1.22.
0,
we can
,/,(t)
= is
= kx + u
da;/dt
find
so that
j/(t)
+ x(t).
10
MEANING OP STATE
[CHAPi
Solved Problems
1.1.
Given a delay
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[tT,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
{u(toT), u(toT
+ e),
u(to.)}
...,
M[e_r,t,
x{to)
The
M[(_r,0
MtT), u(tT +
T
An
e),
...,
Then the
state
U(te))
line.
tot<to+T
any
1.2.
^ ^
2/(0, i)
1/(1, i)
0(.,,)
What
0.
t
to
is
is
CO
y{r,t)
where
ce^'^^"'o'
smmrr
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
(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.
_
~
dt^
Changing variables
to
2/
Ki
\y
if
idy/dty^l)^
if
{dy/dty<l)j
and dy/dt
X2 dx^ldxi
X2,
or
= x^
CHAP.
MEANING OF STATE
1]
11
The former equation can be integrated immediately to x^it) = a;2(to)> a straight line in the phase
The latter equation is solved by multiplying by dxi/2 to obtain ^3:2/2 + Xi dxi/2 = 0.
This can be integrated to obtain a;(t) + xl(t) = xlit^) + xl(to). The result is an equation of circles
The straight lines lie to the left of the line 9;2 ~ *'i ~ ^ ^^^ *h circles to
in the phase plane.
plane.
Fig. 15
(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. 16, 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. 16
To
see
u,{t).
if
the system
Then
y,{t)
ered a
MEANING' OF STATE
12
1.6.
[CHAP.
Given the electronic device of Problem 1.5 (Fig. 16), 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 inputoutput relationship is now
Vit)
[5u{t)
u(t)
sgn m(0]/12
To be
linear,
ay
^2/2
[5(aMi
+ Pu^) +
(aMi
+ pu^
aui sgn Ml
pu2 sgn Mg
(om^
1.7.
sgn {au^
+ PU2)
sgn Mj
+ Pu^)]/X2.
if
sgn (au^
sgn u^
+ pu^)
sgn (au^
+ pu^),
so that the
Is this
xoe'"'
+ f
e^''u{T)di
timevarying?
RC =
= u(tT)
u{t)
Let
r<.
a;oe'o +
= T
t^
T,
Original System
e^''u{r T) dr
J"
to+T
T:
TT
y{a)
a;oe'+'^
"
ei + 'r~'^u(i) di
j
to
Evaluating ^ at
y(t+T)
a= t+T
=
gives
a;oe*'
+ J'efMl)d{
'o
which
is identical vidth
the output
<o
y{t).
+r
Shifted System
Fig. 17
t+r
CHAP.
MEANING OF STATE
1]
13
Supplementary Problems
IX
M
1.9.
y{t)
K{t, r) m(t) dr
K{t,T)
of
Is
1.10.
differentiable function
Is the
where
system linear?
Fig. 18
+ 1) =
x{n)
+ u{n),
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.
1.13.
Draw
1.14.
+y
(i.4).
0.
Given the circuit diagram of Fig. l9(a.), where the nonlinear device NL has the voltagecurrent
shown in Fig. 19(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. 19
+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 timevarying?
1.18.
1.19.
Show equation
1.20.
Show
j
1.21.
equation
0, 1,
timevarying?
Is it linear?
Is it linear?
is
nonlinear.
(1.10) is linear.
(1.10)
is
timeinvariant
if
the coefficients
=Xi + 2,
Given dxi/dt
(a)
Does
this
(6)
What
is
(e)
Is the
dxi/dt
x^
+ u,
Xi
+ X2
system linear?
on
and ^j for
1, 2, ...,7i
and
MEANING OF STATE
14
[CHAP.
1.22.
1.23.
Is the
1.24.
What
1.25.
Is the equation
system dx/dt
is
e*u,
f(t)u
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)
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 onedimensional for t ti.
timedependent in general, it must be a family of sets for each time t. Usually
it is possible to consider a single set of inputoutput 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
1.12.
1.13.
The
It is
a dynamical system.
1.15.
1.16.
It is linear
1.17.
It is
1.21.
must be
K{t
3fc
2 m(z)
t=3
1.14.
timeinvariant.
it is
i =^ 0, 1,
only?
Mx
need not "tie" ourselves to an "initial" condition because any one of the values
state for
unless K(t,
fcl
1.10.
in
represent
y(t)
T)
Supplementary Problems
to
1.8.
du/dtl
Answers
u{t
e*x timevarying?
y(t)
<
results in
an imaginary output.
and timevarying.
(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)
etxo
have a state
e^""""
to
1.24.
If
which
X,
we
is
is started.
If
Xq
0,
the system
is
equivalent to
dx/dt
= x + u
where
timeinvariant.
define
du/dt
lim [u(t
+ e) u(t)]/e
c>0
so that
y(t(,) is
tg.
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
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
.
.,
y(t)
is
Ui{t)
U2{t)
Ui{t),U2{t),
Un{t)
Mj(<)
M2(t)
y(t)
M()
Fig. 21.
Definition 2.2:
output
is
y{t)
Um
()
Fig. 22.
Definition 2.3:
Summer
An
integrator
one output
y(t)
Scaler
is
y{t),
vit)
+ J
relationship
u{T)dr
9 v(*o)
^y(t)
u(t)
Fig. 23.
16
Integrator at Time
CHAP.
2]
Definition 2.4:
17
+ + l) ^
u{j
+ l)
for
0,1,2, ...
v{D
*y(k)
u(k)
Delayer at Time k
Fig. 24.
2.2
of the
form
the foregoing elements. Also, any given transfer function can also be represented merely
by rewriting it in terms of (1.10) or (lll). Furthermore, multiple input and/or multiple
2.1.
with
u{t)
An
ay
au
(2.i)
is
shown
<?
+
in Fig. 25.
y(to)
^yit)
Fig. 25
Since a
is
shown
if
in Fig. 26.
y(to)/a
u(t)
tO
i
y(t)
Fig. 26
is
grated to
t
y{t)
y(to)
t[j/(t)
to
+ u(r)]
dr
t,
then
(2.1)
it
can be inte
18
[CHAP. 2
= J/(*o)J
V{t)
+ u(i)]didT + tC
[y{0
[v{r)
+ u(t)]
dr
to
*o
27.
<t)
Fig. 27
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 timeinvariant systems, the diagrams are almost
identical, but the situation is not so easy for timevarying systems.
data) systems.
Example
2.2.
+ 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{
= aY{s)+aU{s)
zy{l)
aY{z)
aU{z)
25 the
in Fig. 28.
>y{l)
u{k
+ 1)
* y{k +
tO
1)
Fig. 28
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.
Note
2/(to) is
Example
To verify
equa
^t
tion {2 J.) is
the state at
to.
J/(to)
e""''
+ aj
e^>M(r)dr
2.4.
is
y{k
+ 1),
CHAP.
2]
Example
19
2.5.
From
first
Fig. 27, 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. 27 and Fig. 25 are equivalent systems.
Example
2.6.
summer
it,
is
completely determined by
the input.
2.3
Consider a general timeinvariant linear differential equation with one input and one
output, with the letter v denoting the time derivative didt. Only the differential equations
need be considered, because by Section 2.2 discrete time systems follow analogously.
+ "iP""*!/ +
p"2/
a_iP2/
a2/
i8p"M f
pjy^u
!+ i8_jPM
+ Pji
{2.3)
PHV 
+ P""'(i2/ I3^u)+
is
not true
^(^, ,!/)
Pia^_,y
if a^
 ^_itt) +
a2/
depends on time.
)8M
+ ~=iW^ua_,y) + ^(J3uaj)
drawn
{24)
"vit)
Fig. 29.
The output
Flow Diagram
of each integrator
is
X,
Form
x^
form
+ /?M
a^y
iCj
+ p^u
{2.5)
20
Using the
first
equation in
{2.5) to
X2
/a.
dt
Xnlj
\
form
\^
(xr
[CHAP.
^ /^^
^iA
X2
(2.6)
a;ni
'1
0/ \x.
^^o
nxn matrix.
(Xt
X2
{2.7)
j8tt
Xn
Note
this
directly
from the
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. 210.
relabeled.
state
now
x^
Xg
*3
x^
^n
^ A~2"'nl
{2.8)
CHAP.
2]
In matrix form,
(2.8)
may
21
be written as
1
X2
1
(2.9)
dt
3/ji
Xn
*n2
and
{2.10)
Xnl
This will be called the second canonical form, or phasevariable 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 inputoutput relation {2.3) can be verified.
The phasevariable canonical form can also be written down upon inspection of the original
differential equation {2.3).
2.4
By
{2.3)
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"! +
pj
{^,
Pi
The
^o
+
,
+ a_jip + =
 Aj)(p \^{v A)
+ ^rri^ +
,
(;2.15)
the form
Pn
''2
Pi
^^^
+ ^3r
,
{2.U)
can be calculated as
(2.15)
shown
in Fig.
211 following.
(p
{2.14^)
22
[CHAP.
M(t).
j/()
^O[>
rOv^*~r~G)
L9D^
and
pj
The
functions of time.
Fig. 211.
\^
are complexvalued
x.
state equations
^2
K^2 + ^
(;g.ie)
Consider
Jordan
now
Aa;
^o
ajP"!
++_!?)
Pi
^0^
+ ^pxtY +
The residues
Pi^'i
P2^2
pna;
2)"
+u
= {p\y{pK^{){pXn)
{2.17)
identical roots.
p^U
p^U
Py + iU
{pxi)"'
pxi
PX.+1
P kn
'
Pfc
.,
v/^^v
^[^<'^'H.=H
.V.
/vy;
,,,._,
p from
{2.12).
"=''
IV
j.,^,
.,^
{2.19)
CHAP.
2]
Fig. 212.
The
23
Xz
A.ia;2
Xa
Xv + l
=
=
=
Av + iX^ + i
PoU
Xvl
Xv
AiXyi
AiXi.
Xv
+u
{2.20)
+u
+ PiXi +
92Xz
4 PnaJn
24
The matrix
Xi
/a;i
Ai
00
Xvl
dt
Xv
...
Alio
...
...
OXiO
...
00
0...
Xv + l
\xn
[CHAP. 2
/ a;2
Xvl
...
Xv + i
A+i ...
...
l'\
(2M)
\ ccn
Xi
and
P2
(Pl
Pn)
*'
I
Xn
row
Po
A,i
2.7.
V+2y + v =
Equation
(2.2S)
can be written as
j/
T^qrjT^w
(P
Figure 213
is
+ 1)2
+M
(SS)
'
u(t)^^*Q
.p^
Xl
^E>
KE>
Fig. 213
x^,
is zero,
(0
1)
\X2
The matrix
state equations in
CHAP.
2]
25
TIMEVARYING SYSTEMS
2.5
Finding the state equations of a timevarying system is not as easy as for timeinvariant
systems. However, the procedure is somewhat analogous, and so only one method will be
given here.
is
The general timevarying differential equation of order n with one input and one output
shown again for convenience.
S + i(*)^ +
+ (%
p,{t)^ +
p^{t)u
{1.10)
{2.8).
'^i
[to
a;2
yi(t)M
)'2(*)
{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.
S+
Then by
l(*)^
l^oit)^
+ Pi(t)^ +
P^(t)u
(2.26)
(2.24)
and
2(*)3/
xi
7o(*)m
(2.27)
(2.28)
differentiating,
y =
Xi
yo(t)u
yo(t)u
= 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)
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
[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
j8oi
(2.S7)
jSoaz
u
1
...
2/
(1
{2.38)
0)
y,{t)u
"nl
2.6
Multiple inputmultiple output systems can be put in the same canonical forms as single
inputsingle 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 wvector,
u(i)
is
an mvector,
y(i)
is
A(f)x
B(t)u
C(t)x
D(t)u
{2.39)
fcvector,
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.
+ 1) =
y{n) =
A(w) x(m)
C{n) x(%)
+ B(m) u(n)
+ D(tc) u{n)
{2.i0)
CHAP.
2]
Specifically, if the
y,
27
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.
linear time
with
initial conditions
y{0)
5y
y{0)
yo,
+
=
6y
ii
+u
yo.
(2.4.1)
initial conditions
on the
state variables.
+ 5py + 6y pu + u.
Dividing by p2
= piu5y) + ^(uey)
at the right.
Fig. 214
Next, the outputs of the integrators are labeled the state variables xi and X2 as shown.
Xi
is
x^.
62/
summer on
X2
5y
Now
left:
the right:
+u
5xi
X2
exi
+u
X2
u
(242)
28
The
[CHAP. 2
o)G)
6
(l^'"
The
Xi(t)
into (2.42)
and setting
initial conditions
5j/(0)
and
Vo, the
i/(0)
2/.
052(0)
M(0)
determines
 yo+
a;2(0)
j/o
gives
y{0)
2.2.
X2
initial conditions
conditions.
y(t)
(10)
5j/o
"(0)
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. 215.
y~*
problem
is
{2..i.l)
Fig. 215
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
+ X2
5a;2 + u
Xi
6x1
(245)
X2
n/a^i
6
dt \X2
The
and
its
initial conditions
(2.U)
left integrator is
Xi
and
(243)
5/U2
(1
1)
X2
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)
2.3.
29
a;i(0)
_=
22/0
^2(0)
Byo
5a;2(0)
m{0)
(247)
initial conditions
i^o
^"(0)
+ ^yo^u(0)
Find the matrix state equations in Jordan canonical form for equation (241) of Problem 2.1, and the initial conditions on the state variables.
The transfer function
is
p^
5p
(p
p + 1
+ 2)(p + S)'
1
From
p+2
p+S
H 1
O^
+
6
X2
H
<3H
Fig. 216
The
A pi
\X2
The
equalities at each
i:)::ha
dt
From
from the
(1
2)
2a!2(0)
Vo
2i2(0)ii(0)
and
^0
a;i(0)
and
to,
2a!i(0)
0:2(0)
a;i(0)
2(0):
6x2(0)
M(0)
gives
!ri(0)
m(0)
^2(0)
i[M(0)2j/o^o]
3^0
Vo
summer:
30
2.4.
[CHAP.
llX^i'."
dt \X2
y
Find the
1)
(1
pxi
X2
px2
6xi
Xi
X2,
5py
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. 217 find a state space representation of this
closed loop system.
R{s)
()
C{s}
= TTi
W = JT3
Fig. 217
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. 218.
r).
1
1
*c(t)
4
Fig. 218
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
+ 2) +
V^{n
+ 1) +
a^y^in
Vln +
Put
this in the
1)
y^v.J,n
+ 1) +
y^yj^n)
(i^u^(n)
^^u^(n)
y^yln)
a^v^{n
+ 1) +
cc^y^{n)
^^u^n)
h^u^n)
form
+ 1) =
Ax(n)
Bu(n)
Cx(n)
Du(n)
y(w)
a^v^{n)
x(w
Where
31
= (^.),
<,
= (.
The
tions
z^Vx
Dividing by
a^zvx
a22/l
li^Vl
2/2
Yi2/2
ffsZl/i
732/2
jSiiti

SiMg
42/i
ySjMi

S2M2
2/2
3l/i
(l3/i
 722/2) +
;
(/32M1
;5(/3iMi
S2M2
2/2,
+ SiM222/i 732/2)
 7i2/2  a42/i)
drawn as shown
in Fig. 219.
2/2
Fig. 219
fn
t^^ T^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 timevarying 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)
+ 2)
4
j(n) i/(n
+ 1) +
a^{n) y{n)
w(
+ 2) +
u{n
fi^{n)
+ 1) +
pj^n) u{n)
{249)
1)
(2.52)
Stepping
y(n
(2.53)
2)
Substituting
gives
+ 1) =
a2()
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)
= aiW XaW 
x^in)
a^ivL) X2(n)
y(n)
Stepping
1)
yi(w) u(n)
Vo('^
1)
u(n
1)
1)
M(n
1)
(2.53)
(2.51) yields
iW + y2W "(w)
into (2.49)
yi(w
4
and equating
ya(n
coefficients of u(n),
+ 2) M(n + 2)
u(n
1)
(2.54)
and u(n
+ 2)
VoW = ;8o('*2)
ViW = ^i(ml) ai(n)/3o(M2)
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.
V2W
=
(1
\fxi(n)\
i(w)/ Va;2()/
0)(^'^)
"^
/yiW\
VYaW/"*"^
ro{n)u(n)
Given the timevarying second order continuous time equation with zero input,
y
a^(t)y
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)
 Tiffi 
Y2a2)i
ajj
(Yi
33
(2.56),
+ y2)a;2
(2.5S)
as before,
(2Yi
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
2yi
("I
 2a2)Y2 +
(<4
y2(*)
a2Yi
ajj
C^^^)
satisfying
~ 2i + 2  2)Y2 ~
(2.60)
This problem illustrates the utility of the given timevarying 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
Supplementary Problems
2.9.
2.10.
2.11.
(i)
y(n\2)\Zy(n\l)\2y(n)
= u(n]r\) + Zu(n),
(ii)
(iii)
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
"3
X3
t
(02
=

Pi^l
^iM2
^2"!
S2M2
^2
* V
Fig. 220
2.12.
500
p(p
+ p + 100)
6)(p2
p2
+ p + 100
+ 2p + 100
p2
rn^ig.
221
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 inputoutput relationship (2.S). Why can't the timevarying flow diagram corresponding
to equations {2.24.) be turned backwards to get another canonical form?
2.14.
with the
ai{t)y
^(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)
2.15.
Verify equation
2.16.
The
initial conditions of
{2.37)
using equation
Ri
Inertial Load:
K^i
{2.25):
motor are
Motor Armature:
di
dB
Kfjj
Lrr
+ ^
J ~Tfi
j7
to the
using a
state vector
\deldt/
2.17.
Prompt Neutrons:
In
{p{t)
Delayed Neutrons:
/Jjii
fi)n +
XjCj
2
=1
\Ci
where
/?
i=l
and
y^i
p{t) is
2.18.
Assume
Lateral Translation:
Rotation:
'4,
K^a
Angle of Attack:
+ K^^ +
K^ji
lateral translation,
I.
K^z
M{1)
and
(i)
Ka{l)a
Kb{1)IS
,,,
attitude,
= Ax +
dt
u = p
^
where
K^P =
where
K^ +
Bw,
= Cx + Dm
(ii)
(iii)
M{1)
CHAP.
2.19.
2]
35
9^000^
L,(t)
dt
= Ci(t)
dt
dCi
de,
(Ciei)
^2(<)
dLi
dii
^M0"^o
Cinr
e..
C
e,
It is
dL,
dii
eu
more convenient
La^ +
at(^2k)
Fig. 222
i^
Pi
Pi(to)
dt
p2(to)
(^12) dt
4
9i(Jo)
(aei)dt
I
to
P2
(El
Bb)
A(t)K
B(t)u
C(t)x
D(t)u
dt
y(t)
where the
'Pi
(i)
y =1
9i
ei
,P2/
^2,
'ii
X =
(ii)
ei
eb
=
I
61
.J2.
2.20.
Answers
2.9.
(i)
x(w
l)
(ii)
x(n +1)
(iii)
x( 41)
to
Supplementary Problems
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
10
oci 0:4 0:3
0:2
74
2.11.
~^**
Oil
0)2
+0)2
0)3
xi
(1
2.12.
For
Xi
V,
X2
^1
^3
^>
^i
'*''
^5
0)1
'>^>
Xs
X2
=w
1
1
1
500 105 6
100
2.13.
100 2
The timevarying flow diagram cannot be turned backwards to obtain another canonical form because the order of multiplication by a timevarying coefficient and an integration cannot be interchanged.
2.14.
(i)
?/
Fig. 223
CHAP.
2]
(li)
The values of
(iii)
XiiO)
I/O
vq, y^
YoiO) u(0)
a;2(0)
^o
(2.3S), (2.54)
and
(S.35).
/RL^
(p(f)/3)Zi
Xi
2.17.
2.18.
(i)
A =
(ii)
A
C
(iii)
A =
2.19.
(i)
A =
(ii)
=
BH
2.20.
state vector
x =
(gj
Qn
dPi
ail
dt
dt
Pi
Pn)'',
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
we suggest study
subject.
3.2
BASIC DEFINITIONS
Definition 3.1:
Example
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,)
all
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
is
mxn,
read
by
n, if it
has
rows and n
columns.
A scalar,
Definition 3.6:
Definition 3.7:
38
CHAP.
3]
Definition 3.8:
39
The diagonal
Example
3.3.
{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
The
tr
A =
is
622. ^33.
the
sum
sets
of all elements
BASIC OPERATIONS
Definition 3.10: Two matrices
3.3
means
an
Definition 3.11:
Matrix addition
Definition 3.12:
is
Definition 3.13:
is
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
is
multiplied
by the
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
^^ C.
no
is
BA = BF =
C,
"division" by
B would
A = F,
imply
division, consider
BF
Since
which
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 nvector. 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:
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
Example
and a
3.6.
X
X
3 matrix
1
matrix
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
AiiBii
A21B11
+
+
A]2B2
AllBj2
A22B21
AoiBio
+
+
A12B2
AooBo
and give
insight.
CHAP.
3]
41
SPECIAL MATRICES
3.4
Definition 3.15:
Definition 3.16:
diagonal matrix
is
0, is
all zeros.
all
zeros.
Definition 3.17:
are
I, is
all ones.
62
(ei
I
Cm).
Note lA
= AI = A, where A
is
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:
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:
Note (AB)^
B''A'r
and (AB)t
is
A'^.
BtAt.
Definition 3.21:
I,
orthogonal
3.5
AtA =
D, the matrix
is
normal.
The determinant
Definition 3.25:
otannXn
detA
^^^>
(~l)''n'i2i'2' "''
Here
Pi 2,
.,,
42
Example
[CHAP. 3
3.7.
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
(I)^ai3a22a3i
(I)2ai2a2331
nXn
given in Problem
3.2:
p, so
(I)'>aiia22a33
Theorem
from an odd
1, 2,
is
there are
possible permutations of
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) =
is
most
is
Elementary row
(i)
(ii)
Multiplication of a
(iii)
Adding a
row
(or
row
scalar times a
column) by a scalar,
(or column) to another
row
(column).
3.8.
Consider the
obtain
2X2
matrix
A=
{ay}.
'0
EA
To add
to
E.
column
AE=i
iWaii
/ V
ai2,\
021
la.2\
oii^
a,
22
**:
ai2\/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.
E^i, found
column operation of
Example
in
(ii).
3.9.
The inverse of E
Definition 3.27:
j.
The inverse
of
E = [^
Ei
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
Ei
is
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:
of the determinant of an
aijCij
A =
i=l
is
det
2)
fl^Cij
for any
row
A =
i.
3=1
w x w matrix
is
3.11.
A =
Corollary
3.4:
ai2Ci2
3 matrix
022832
~0'12('*21^33
0.^2632
~ 'f23'l3l) +
The determinant
is
'l22(*l 1^*33
of a triangular
~ *1303l)
nxn
""
032('*ll'''23
~ 13"2l)
diagonal elements.
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
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.
elements.
This
2=
is
is
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.
since
44
The Kronecker
Definition 3.28:
delta 8
1 if
and
8ij
[CHAP. 3
if i
^ j.
*Cij
2=
i=l
Definition 3.29:
of
aiciCji
SkjdetA
(3.2)
is
adj
A=
of cofactors of A.
The adjugate
Then
(3.2)
is
is
5.2.
A adj A = IdetA
Definition 3.30:
An
an
(3.3)
TO
if it
has both a
left
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
AiA = 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
Then
^"^
d^^^J^
(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
1
CHAP.
3]
1):
1
Divide the
first
45
row by 1 (detEg =
1):
11 101
1
ej,
and
all
that
3.6
Example
I is to
add the
1
all
matrix equations have solutions, and some have more than one.
3.14.
>
has no solution because no
2i
(b)
VECTOR SPACES
Not
(a)
is
is
=
=
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. 31.
in a plane.
Ij,
3).
(2
Fig. 31.
Representation of
(2
3)
3) in
the Plane
46
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
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
plane.
shown
in Fig. 32.
Fig. 32.
Definition 3.32:
Example
(a)
under addition
any
a2
S V,
then
(ai
Set of Vectors
if,
given any
in U
ai
GK
and
+ ai) S 11.
3.15.
Given
sum
of
are integers.
(6)
Given lA is the set of all 2vectors 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.
Definition 3.33:
Example
3.16.
Given
vector in
A set of  vectors li in
Definition 3.34:
space
if it is
(1)
closed
Any
(2)
one vector
is
called a vector
plication.
If a
G X/, where
ai, a2,
a,
multiplication.
Given
If is
is in
G X/
because
'U is
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
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 3vectors 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),
(y3a)(l 2
ai,a2,
..,13k
not
0)+
{2cc/3){l
0)
0(0
0)
+ ^8232 +
ySiai
if
+ ySfcak =
0.
3.18.
If
3.7:
1(1
0)
10)
1(0
(0
0)
dependent
set.
A set of vectors are linearly independent if they are not linearly dependent.
Definition 3.37:
Theorem
0)
and only
if
of a square
.,a
matrix
A = 0.
are
'
Then
1
...
AE =
p^
...
...
p^
...
...
p.
...
...
...
0\
\0
(ai
32
a)
Theorem
Example
3.19.
Consider
gives
det^g
^j
and
12.
j.
row
gives
det
(I
Jj
There
0.
is
Definition 3.38:
A
(2)
set of nvectors
ai, aa
a^
form a
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
Pi
+ ^2 =
Example
3.2U.
Any two
since
[CHAP.
0,
0)
2Pi
Mi
+ p^
10)
(0
The only
Q.
0)
solution is that p^
zero,
3.21.
Any three noncoplanar vectors in threedimensional 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 threedimensional Euclidean space
vectors in this plane form a basis for 11.
Theorem
3.8:
1)^.
Any two
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:
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.
To construct a
li,
basis
from a given
if as is
0)
a linear combination of
ai
and
a2,
combination of
They are
0).
(1
If it
ai.
82,
is,
or only ai
if
. ,
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
3.10:
bi, b2,
Note
Proof:
Then
hi.
bi
m~
with a basis
ai, a2,
set in order.
test
test
The
Then
Next
I.
set.
82
first
Delete another bvector 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 avectors are a basis, so they all must be linearly
a2, ai.
independent, hence I m. Interchanging the b and avectors in the argument gives m^l,
dent
set.
CHAP.
3]
A vector
Definition 3.39:
space
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
Consider the
matrix equation
Ax
=
are zero, x
/A
(aia2...ja)
2f
If all
Ij
(1)
Does the
x=
Yes, because
(2)
Are
Ax =
Are
Az =
if
and Ay
0,
then the
sum x = z + y
is
a solution of
0.
Yes, because
So the set of
Ax =
0,
then
Ax =
is
a solution of A(j8x)
j8x is
0.
a vector space.
of all solutions of
Ax =
is
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
is
Yes, because
(3)
element?
is
given in Problem
3.8.
Definition 3.41:
Corollary 3.12:
Theorem
3.13:
null space of
is called
the nullity of A.
The dimension
is
3.23.
,^
/I
(
3\
)
It
has one independent row vector and therefore must have only one
Definition 3.42:
Ax = y
for some
is
space of A.
It is left to the
is
a vector space.
50
Example
[CHAP. 3
3.24.
null space
may have
other vectors in
common
Consider
Then Ab
0,
so
Definition 3.43:
Theorem
3.14:
The rank
and Ac
is
b, so
is
of the
The rank of
vectors of A,
i.e.
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
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 3.44:
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
z,
objects x, y
x+
x and y
(y
+ z),
y.
V,
(a^)x,
X,
+ y) =
{a + ;8)X =
aX
ay.
aX
;8x.
of Definition 3.7 (wvectors) 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
t,
space.
Example
3.26.
A(<)
Example
The
as
is
solutions to dx/dt
A(t)
is
3.27.
w =
0, 1,
is
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
11 of
Example
3.25,
and
infinite.
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
Definition 3.45:
A
h
we have
ev with
b), is
of
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.
(6)
(distance
p(b, a)
to
For two
_

is
(ab)t(ab)
[l
+ (ab)t(ab)J
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)
iaa
aj a].
52
(1)
llall
(2)
!a
(3)
]a
=
=
(4)
ia]
Example
0,
if
andonly
llc
if
0,
(trivial),
a
in "Ug is
+ c.
ia
2)ll2
ll(i
ilaiU
VTiMla2p
3.31.
norm
Example
3.30.
norm
Example
[CHAP.
in 'V2 is
IK^i
ffl2)lli
lai
Wi\
k2l
pj
3.32.
norm
norm
in U is the Euclidean
= V^+a =
a2
il^
Any
positive
Definition 3.47:
An
Example
An
(1)
(a, a)
(2)
(a, a)
(3)
(a,b)*
(4)
(a
(a, a) is real,
(x,y), of
and only
if
tv?o vectors a
if
0,
(b,a),
+ ph,
c)
a*(a, c)
(a,
ah)
and
a(a, b)
8*(b, c).
(a, 0)
0.
3.33.
is
(a, b)
atb
Oj ^i
3.34.
of time functions of
{x,y)
Definition 3.48:
Two
Example
3.25 is
x*(t)y(t)dt
if (a,b)
0.
3.35.
any
0,
Example
to a.
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,
inner product
Note
An
Example
Theorem
if
3.33.
3.17:
Schwarz inequality
holds if and only
equality
the
furthermore
and
(a,b) the
(a,b)2
if
(a,a)(b,b)
ab or a or b
Proof:
scalar p,
If a or
is
(a
+ ;8b,
+ ;8b) =
(a, a)
if
;8*(b, a)
+ i8b =
0.
^(a, b)
Setting
+
/?
7^ 0.
iP^h, b)
(b, a)/(b, b)
and
re
CHAP.
3]
Example
53
3.36.
J
Theorem
(a
x*(t)y{t)dt
+ b,
+ b) =
=
(f''\!c*{t)\2dt\fC\y{t)\2dt
^(aTbTaTb) ^
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)
+ b,
+ b) ^
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:
Definition 3.50:
vs'here 8y is the
a2,
^^""^^
An
(b,b)
Example
shovi^s
Togeth er
inequality.
(a,a)
Kronecker
of a vector a
ai, a2,
..as
is
is
a2
ITJj]
{1
= ^A^^
''^
delta.
3.37.
e,
where
e; is
defined as
ith position
them?
To
make
its
bi
ai
length unity:
/
*Z
becomes
Because
formed from
and a2. First
unity, y^
ai^^
)'2'l
where
Now
b^
y^a^
54
[CHAP.
Here y^^
To
bj.
is
find
;77
TT
Ia2
V2lll2
bj
b.
=
la,
2(b,a,)bJ,
1=1
Example
By
3.38.
Making
this
Theorem
0),
3.19:
bj
aj
(3
2
(1
Then
1)/V2.
ai2
VI,
so
bj"
0)/\/2.
2)/v22.
Any
finite
3.11
(1
all
vectors
RECIPROCAL BASIS
Definition 3.51:
an arbitrary vector
(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
{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)
+ P^{x^,\) +
(ri,x).
Pjj^i.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 nonorthonormal basis in "Vn.
3.12
L{aa
Example
it
+ ^b) =
j8L(b)
3.39.
Example
aL(a)
whole of a vector
X^2, such that for
is
in X'^,
is
i.e.
3.40.
Consider the set lA of time functions that are linear combinations of sin nt, for n
1,2,
Then
00
any
can be expressed as
a linear operator, because
x(t) in
time
is
x{t)
"^^
+ Px2(t)]
[ax^it)
The operation
In sin nt.
dt
C Xiit) dt +
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
3.20:
{y.j}
whose
basis
ci, 02,
..,
ith
.
. ,
Cm,
dimensional Vz.
Then x =
m
L{^i)
P^^'
Furthermore since
L(bi)
is
y^j
such that
1=1
y,iCr
But
/ n
L(x)
l(x^a)
\i=l
/
^^^(b,)
i=l
2:A2:r,.c,
i
=l
j=l
n
{0^}
equals 2)
yj^p^,
i.e.
the matrix
'"^
3.42.
sin
0"^
COS0.
56
Example
An
[CHAP.
3.43.
is
as
The
null space,
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
is
For example,
First realize that a function of an mvector 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 mvectors, ai, a2,
vector space, in that it satisfies Definition 3.44, but has no components that can be arranged
z
= Ax + By
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)
afc
(3.7)
says ^^
is
alternating.
+ /8aj) A afc =
ai^a^
(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 mvectors
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
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)
(59)
aj
a aj) +
a(afc
linear in either a^ or a^
a^
Pia^
i^iO)
aj)
(bilinear)
a^).
aCaj
aj
Letbi.ba, ...,b
a^
a;
Aa^
0.
Furthermore
if
and only
if
a,
0.
beabasisof
aiAa,
Then 3;=
2^
fc*fc
Qifcbfc)A(2yibO
^"<^
^i
J^ 2
,2
T'''"
afcr,(bfc
*^^*
b.)
is
a linear com
CHAP.
Since
3]
b^
b^j
and
a bj
b^.
= b; a bj^
k<
for
Because
bfc
A b(
aj
for
a^ is
< I n form
A^V
is
n(n
 iYfc)bfc A bi
(fcYi
(3.11) is
(3J1)
Summing
tion
to
(1
2 2
1=1 fc=i
aiABj
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
<k^n
l^i<j<
_
~
nl
{nv)\p\
/n\
(^^^^
\v)
The determinant of a
Definition 3.54:
vector space
'U vi^ith
linear operator
a basis
bi, b2,
The
for
the
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.
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
Definition 3.55:
4?"
is
nxn
w^hose
b is defined as
L( MAL(b2)AAL(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 a2 a ai
Theorem
Then det A =
Let
e;
n components and
aiAa2A
3.21.
Proof:
is
in li since 'U
be the ith unit vector in '=Un and j be the jth unit vector in
Vni, i.e. e; has
., has k  1 components.
Then ai = auei + a^e^
anien so that
Aa
aiieiAa2A
Aa+
a2ie2Aa2A
Aan+
++
ttniCn
Aa2A
Aa
58
The
sum on
first
Cl
32
A an
Cl
(ai2ei
[CHAP.
+ ^2262 +
+ fti2en) A
A (ainBl
+ 0,nn^n\
82
A Hn
r\
(122
Cl
\an2/
=
detin =
Since
e,
A [a22 (^j
\a..Q^ +
a2(;;^_
J]
a(^^_
j]
detlni, then
CjA
0\
)A
/o
A ^^
l
'1/
ei
["022
]
[a22*l
+ fln2*nl]
is
A ra2 i
on the right
dnni
\
'
[a2n*l
'
+ ftnn*il]
lain\
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.
ai, as,
bi
filGiriGiits
<..
(ii)
(12,(2)
(2)
(3 4)
(g)(.IM
(V,
(iv)
(axa2)(^)
(vi)
(i)
nxm
(1X3 + 2X4) =
(
()
3.13,
82) (^^^
^^
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
+l2l'2l'2'80+l'08'0l'0'3'2
+ l'6'30l'6'l'0 + 06l00'6l'0 + 0'0'l'2
0'080 + 0l'800'l'l2 + 0'l'3'20l8'0
+ 02'8'00'2'l'2 + 0'6'l'00'6'3'0 + 2'0'30
2'0l'0 + 2'2'l'02'2'l0 + 2l'l'02'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
detA =
(6)
is
4.
0\
/I
detA
3.3.
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 op^i
'
'
"pi*
Since a determinant is all possible combinations of products of elements where only one element
taken from each row and column, the individual terms in the sum are the same. Therefore the
only question is the sign of each product. Consider a typical term from a 3 X 3 matrix: 03ia]2a23>
Permute the elements through ai2ff3iffl23 to tti2"23"3i> s*' that the row
i.e. pj = 3, P2 1' Pa 2.
numbers are in natural 1, 2, 3 order instead of the column numbers. Prom this, it can be concluded
n
in general that it takes exactly the same number of permutations to undo Pi, P2> ,Pn to 1,2
Therefore p must be the same for each
as it does to permute 1,2, ...,n to obtain Pi, P2> >Pnproduct term in the series, and so the determinants are equal.
is
60
3.4.
Vandermonde matrix
y _
1
^2
^l
nn 1
is
nonsingular
if
and only
an
/i
O2
il
Prove
[CHAP.
for
6i  Oj
if
On
i = j.
detV
(92
9l)(9392)(3ei) ( 'nl)(9nn2)(9n9l)
(ffj
" ^i)
For n = 2, det V = fig "" ^u which agrees with the hypothesis. By induction if the hypothesis
can be shown true for n given it is true for n 1, then the hypothesis holds for w  2. Note each
term of det V will contain one and only one element from the wth column, so that
70+yi9n+ +
detV
If
61,62,
Vn1;;
But Yni
is
Therefore
Tni
By assumption
...
#,
82
...
ei
,n 2
e'1
60
"2
det
1,
Yi
n1
(iOi)
detV
3.5.
Show
det/''^
r)
det
(9i)(ee2)(9n9i)
A det C,
where
nxn
and C are
and
mxm
matrices
respectively.
Either
dependent.
det
A=
or
det
A#
0.
det
If
A=
detf^
and the hypothesis
If
detA#0,
holds.
0,
are linearly
^) =
'
^0
/A
^
~
B\
c)
^A
\o
0\/I
ij[o
OWI
AiB
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
1/
Theorem
=
3.2
det
\o
det
det
CHAP.
3.6.
3]
Show
that
61
ai, a2,
li
ajc
is
expressible
by
ai,a2, ...,afc
combinations of ai,a2,
fc
i=l
and
2
=
"iSi
(^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.
.,
.,am in ^.
Kk^m,
is
afci.
// part:
combination of
If a^ is a linear
aj, ag,
.,a;j_i,
then
fci
t=i
where the
fii
are not
all
=
which
zero since
^lai
4
^2^2
a^.
is
Only
if part:
the
all
linear combination
/3j
^lai
/32a2
Ph^k
Oafc
are zero.
is
+i
Oa
y3a
all
ySj
i>
for
k.
Then the
k~l
Show
(1)8^
dependence.
a/c
3.8.
/3fciafc_i
=
where not
Then
nonzero.
that if an
X n matrix
then
columns,
the null space of
A has
A has
n columns with
dimension n r.
at
^iiai
/3i2a2
first r
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),
{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
jPnr.A
hA
Pit
kr
Pnr.r
"r
1
~lr+l
4r + 2
1
lr+2
in
\o/
\o
\:
ir+i^i
i=l
If s = 0, then the Xj do span
s is a remainder if the X( do not span the null space of A.
the null space of A. Check that the last n r elements of s are zero by writing the vector equality
as a set of scalar equations.
where
nr
Ax
~ir + i^^i
As
i=l
Ax =
Since
and
Ax;
0,
As =
0.
Writing
this out in
gives
<'iai
<
linearly independent, so
i=l
n rxi are a
3.9.
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
row vectors
aj
column vectors
be linearly independent.
a 11
112
a;
Partition
^Ir
^l,r+l
as follows:
am
]
O'rr
"^rH 1,1
<lr+ 1,2
Ctml
^m2
'xj
I
Xr
ai_r+l
Or.
^r+l,n
^mr
am
r+ 1
Xr + j
a,+ i,r+l
'^r,rl
*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
^jXi
=
=
feta,!
for
1, 2,
6ii2
^iX;
'"'
so that
^r+i)
a^+ij).
&itr
i=l
...
b'y,)
r.
2
=
b^y^
(b'yi
62
(6i
\>=1
i=l
b^'yj
ajj
(au
b^.
Therefore
and yj
a,,.)
column vectors
Then
yj
so that
ijyj
i=i
"ijb'^yj
for
2
j=l
a;
^or
"ij^i
= ^ + !,...,.
r + 1,
b^yg
(b^yi
b^y^ +
b'ryr
...
a^
b^yj =
...
fejai
6282
'
'
'
&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
a;
r
for
1,
Any y
w.
range space of
in the
them be
can be ex
= Ax =
l
^i^i
^ix^
+2
(.2
l=r+l \j=l
l
i,a,) X,
(x^
i=l y
= r+l
fc
a^ix^
a,
This shows the a; for i=l, ...,r span the range space of A, and since they are linearly independent they are a basis, so the rank of A = r.
3.11.
X n matrix A, give necessary and sufficient conditions for the existence and
uniqueness of the solutions of the matrix equations Ax = 0, Ax = b and AX = B
in terms of the column vectors of A.
For an
For
Ax =
To show the
necessity:
n
i
sufficiency:
^Ai
x =
If a^
0.
the solution
0,
is
Then there
(i
)
0.
To show
ji
J;
Sili
= Ax =
0.
For
Ax =
b,
rewrite as
2
i=l
^i^i
3.10 a necessary
and
sufficient
condition for existence of solution is that b lie in the range space of A, i.e. the space spanned by
the column vectors.
To find conditions on the uniqueness of solutions, write one solution as
n
n
n
 !>!
Then b =
In) and another as (jj fz . . . {).
=
(vi V2
Offi so that
^iVi =
('7i
<=i
1=1
i=i
The solution is unique if and only if aj, ...,a are linearly independent.
64
Whether or not b
[CHAP.
Ax =
0,
b are that b
AX
3.12.
Given an
mxn
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
that
Az =
space of A.
0,
null space of
rank A^A.
3.13.
To show rank A
y
Ay =
i.e.
0.
i.e.
Use of Theorem
k.
rank A^
is
gives rank
rank AA^".
3.11
equal to the
A=
A;
Given an
Let
mxn
rankA =
r,
so that
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
ajb?"
i=l
AB =
aj, ....a^.
aa,.br
.2
1
=1
aj bf +
\
2+
<^mK
Furthermore, use of Theorem 3.15 gives rankB = rankB^. Then use of the first part of this
 rank B. Again, Theorem 3.15
problem with BT substituted for A and A^ for B gives rankB^AT
gives rank AB = rank B^Ai", so that rank AB  rank B.
3.14.
m
i
Given n vectors
xi, X2,
Suppose det
G=
0.
vector of G.
Then
3.7, /3igi
+ /32g2 +
'
+ PnSn =
n
Pidij
i=l
/3i(Xi,Xj).
i=l
Multiplying by a constant
/?*
=
3
2
=
and summing
i8*
i
2
=
M^u^j)
still
\i=l
^i'xi,
3
2
=
/3*Xj
/
<,
where g
is
a column
CHAP.
3]
65
Use of property
(2)
2
*=i
dependence.
suppose the
Xj are linearly
3.15.
Then there
dependent.
x; gives
"
2, JiSi
which
0,
is
Now
3'
jSiXi
such that
yj
yMu^i)
Jk
3=1
exist
for any 1
yiPy
3=1
G=
VjXj
0.
Therefore
0.
Show
V.
detBA=
Let
det
det(LiL2)
detB.
have a basis
bj, bj,
b.
LiL2(b,)
LiLgCbj)
det {L1L2)
A b2 A
bj
If
L2
(S.13),
A LiL2(b)
A b
L2(bi)
C;
det (IiL
^i<=i)
^1(^2)
A Li(c)
(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
such that
hyperplane
"yP^P'*"
66
[CHAP.
Supplementary Problems
3.17.
Prove that an upper triangular matrix added to or multiplied by an upper triangular matrix results
an upper triangular matrix.
in
3.18.
Using the formula given in Definition 3.13 for operations with elements, multiply the 'ollowing
matrices
/
2
i\ /q
j
^/V'T
^0
^'^
sin *y
Next, partition in any compatible manner and verify the validity of partitioned multiplication.
3.19.
Transpose
smtj
x^
\!r
3.20.
Prove lA
= AI = A
3.21.
Prove
3.22.
Prove (AB)t
3.23.
3.24.
all
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 =
3.25.
How many
sum
xn
matrix?
3.26.
Prove (AB)"*
= B~iA~i
3.27.
Prove (Ai)i"
3.28.
3.29.
Verify
=
if
(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.
3.33.
that
Let a nonsingular matrix A be partitioned into An, A12, Aai and A22 such
have inverses. Show that
^^1
and
if
A21
0,
exist, does
then
(A
+ B)i
exist in general?
dA
Show dAVdt = Aij^Ai.
AirAi2\MiV
V"
and A22
"
'
,
(A22A2iAn*Ai2)VVA2iAii
**
/An'
(
A^
"All A12A22
^1
 A21A11
Ajj
CHAP.
3]
3.34.
3.35.
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.
3.37.
Given
3.38.
For
sum
matrix
A =
A =
3.39.
3.40.
in
(a)
x'^'y
0.
an arbitrary vector
+ y,
(6) this is
show that
Show
A'''.
j,
of two vectors,
Given
transpose of A, and
A and B
and an
and y
is in
Under what
conditions
= B?
V, where
bi, bg,
Show
that
3.41.
3.42.
Show
3.43.
3.44.
that rank (A
+ B) 
lixjla
(x,y)
y2.
(x,bi)(bi,y)
Show
(x
+ y)
is
orthogonal to
(xy).
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
GramSchmit procedure on aj, aj and ag in the order given to find a set of orthonormal basis
the
vectors.
3.45.
Show that
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,
3.47.
3.48.
3.49.
Show
A'Ti
to!
/^_\
for TCvectors.
What
is
(a,b)2
2 2
ki^i
J
^ t=l 3 = 1
 AI'
68
Answers
2a
JT
aj
a;2
sin
[CHAP.
Supplementary Problems
to
3.18.
bx^
bir
x2
/2a
a \
3.30.
(w
1)
Ai
p)
\2l3
3.25.
sint/
^j
3.24.
as
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,
linearly independent.
transformation.
bg
(0 3 4)/5
Hr)
are required.
3.48.
where a and
'U2
5
and y
3.39.
an
1,
(6
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
Ax = Ax
Note
Example
this equation
exist only if
(4.1)
det (A  Al) =
4.1.
4\/a;,
3/1x2
is
X2
Then
3
Xj
characteristic equation is
det
3
The
0.
characteristic polynomial of
polynomial
Then
is
(3
\(A
SkJ\x2J
 X)(3 X) 4 =
A is
Xj
det (A
1,
X2
x26X +
0,
5.
 Al). Note
the characteristic
Definition 4.3:
An
Definition 4.4:
Example
or
X^
Definition 4.2:
The
Vo
said to be distinct
is
if it is
not a
A there is
a nonzero
This solution
AiXi.
4.2.
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
is
easier to normalize
x
is
1.
unity.
4.3.
it is
[CHAP. 4
= =l
v/sV
],
whereas normalizing
its first
x,
1/
example
is
V1/2
4.2
A is a constant coefficient
given as x(0) = xo.
where
Example
nxn
Ax(i)
{J^.2)
The
initial
condition
is
4.4.
and the
initial conditions
dxi(t)/dt
dx2(t)/dt
dxn(t)/dt
aa;i(i)
a^^x^i^it)
a2ii()
a22a;2(*)
a^iXid)
a2i2(i)
+ a^^xJS)
+ a2nXn(t)
(()
a;2(0)
n(0)/
Now
define a
new
variable,
\y
an nvector
y{t)
y{t),
by the one
to one relationship
Mx(i)
(4,3)
Mdy{t)/dt
Multiplying on the
left
Suppose
= M'AMy(f)
(^.^)
AM.y{t)
by Mi gives
dyit)/dt
Definition 4.6:
~dr 
y^{t)
di\
/xt
\2
\yn{t)/
make MAM a
diagonal matrix A.
...
o\/i/i()\
...
0\y2{t)\
=
:
\0
...
Xn
\yn{t)
Ay(i)
Then
CHAP.
MATRIX ANALYSIS
4]
71
The transformation
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.
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
an n x % matrix are
linearly independent.
Note that
if
is
nonsingular.
^i
2 ^x.
for j
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
1=1
But the
X,,
MATRIX ANALYSIS
72
[CHAP.
det (S
 \l) =
0.
But
= det(TiATAl)
det(SAl)
= det(TiATATiIT)
=
det[Ti(AAl)T]
det
Since detT
have proved
Theorem
4,2:
Corollary
4.3:
from Problem
(detT)i
is
(S
= det(AXl).
AI)
Therefore
we
det
3.12,
given in Problem
4.1.
useful fact to note here also is that all triangular matrices B display eigenvalues on
the diagonal, because the determinant of the triangular matrix (B AI) is the product of
its
diagonal elements.
Theorem
4.4:
A matrix A can be reduced to a diagonal matrix A by a similarity transformation if and only if a set of n linearly independent eigenvectors can
be found.
Proof:
By Theorem 4.2, the diagonal matrix A must have the eigenvalues of A appearing on the diagonal. If AT = TA, by partitioned matrix multiplication it is required that
Ati = Aitj, vi^here ti are the column vectors of T. Therefore it is required that T have the
eigenvectors of A as its column vectors, and T~i exists if and only if its column vectors are
linearly independent.
It has already been shown that when the eigenvalues are distinct, T is nonsingular.
So consider what happens when the eigenvalues are not distinct. Theorem 4.4 says that
the only way we can obtain a diagonal matrix is to find n linearly independent eigenvectors.
1.
that root
is
Example
cases:
A;dimensional.
4.5.
A =
is
\l
Then
and
eigenvalue
det (A
0/
Ax =
0, 1
the
1.
o\
111).
'
j[
X2J
Xi
Xs
X3
(0 1
(1)
^2
\xJ
AA
=
0/\xJ
Xi
AA
o\
1
1
gives
=
=
1).
CHAP.
MATRIX ANALYSIS
4]
Therefore
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 TiAT
A,
where
has
0,
and
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
to
Example
2.
4.6.
A =
Since
is
1.
is
(1)
Two
form
{xi 0)1".
4.4
JORDAN FORM
The form
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 nr ones above the diagonal,
and all other elements are zero. The general form is
Lii(Ai)
L2i(Ai)
H.7)
J
_,
+I
jLjmp(Apj
MATRIX ANALYSIS
74
[CHAP.
Each Lji(Ai) is an upper triangular square matrix, called a Jordan block, on the diagonal
of the Jordan form J. Several Lii(Ai) can be associated with each value of \i, and may differ
in dimension from one another.
general Lji(Ai) looks like
Ai
Ai
M^O =
...
Ai
Example
Ai
...
4.7.
(4.5)
^0
where
Ai
..
all
Ai^
'
Because
ones
all
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
of
At2
=
=
Ati
Ati
ti
I
ti
12
equals the
.
ti
.
I
number
)
Ait2
+ Xi
+ ti
Aiti
Aiti
tii
(AiXi
+ Xi
Ait2
(^.9)
(xi ti
+ ti
12
.
1
...
I
1(
Lji(Ai).
Aiti
Then
+ tii
.)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.
/2
1^
\0 1
1.
to
CHAP.
MATRIX ANALYSIS
4]
75
What combination should be tried to start the procedure described by equations (^.9)?
expression gives
(ii
!)(::)
Then
fa
'"2
^2
{::)'(:)"
+
+
~
Tg
""s
'"3
CD'
= P
~P
1\ (
1
irVi
QUADRATIC FORMS
A quadratic
Definition 4.7:
1
4.5
form
is
li.lj,
ayljl^,
such that
^ =
.,!
2 2 %^i^}>
4.9.
Some
Theorem
4.5:
(x,Qx)
All quadratic forms Si can be expressed as the inner product
=
Q.
i.e.
Q+
matrix,
Hermitian
x
w
an
is
v/here
vice versa,
Proof:
First
to (x,Qx):
and
MATRIX ANALYSIS
76
Q=
Let
{q..}
^{a^.
Next, (x,Qx) to
+ a.^}.
(the
Then
problem
q..
and
(x,Qtx)
i=l
Then
(x,
Qx)
SS
1^6 (Q'y);j
i=l 3=1
Theorem
4.6:
= iX^iQii + Qm^i
i=l 3=1
i(x,Qx) + i(x,Qtx)
So
9*^,1,.
*
"
^ = x'^Qx.
real):
t t=
(x,Qx)
real
is
is to
= XXQiA^i
=1 =
(x,Qx)
so
q..,
[CHAP. 4
=^
The eigenvalues
and the
of an
nxn
Hermitian matrix
Q=
Qt are
real,
and the
is
Proof:
Since
of real symmetric
symmetric matrices
and
specific A;
Qxi
AiXi
Qxi
AjXj
Qtxj
AjXj
Xj
are
Hermitian,
is
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
>?'.
X^ =
\.
{i.l2)
(A*
left
by xj gives
A.)xtx,
zero, so that Aj
\. But
A*
A^.
A. ^^0,
Even
if
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
U~'^QU
= A and
'xp
/X+Xj
xjx^
^1^1
x+x^
Vu.
4^2
UtU
4
(XX
x! /
X2I
x)
xx
x+x
CHAP.
MATRIX ANALYSIS
4]
But x[x^ =
vectors of
(x.,Xj)
U are
S,j
linearly independent,
to be proven.
Therefore
= Uy
if
^ =
where the
of
77
Ai
of the
Note
+A!2/2
is
always positive
is identically
the eigenvalues
if
(x,
y)^
of
= x+Qy.
An nxn
Definition 4.8:
form
Then
Si is
An nxn
Definition 4.9:
Example
k2\y2\^
^ is a norm
ki\yi\^
4.10.
^=
Ij
2i42 + ^2 ~
The geometric
Theorem
(fi
~ ^2)^
solution of constant Si
when
fi
is
and so
^2,
is
nonnegative
positive definite is
an
definite.
ellipse in nspace.
4.9:
of Q.
hi
Proof:
Let U be the unitary matrix that diagonalizes Q. Then Q = UAIJt.
a positive diagonal element of A, defineA^'^ as the diagonal matrix of positive
is
Q =
Now let R = UAi'^Ut
are positive.
and
Uniqueness
One way
to check if a
is to
see
UA/2utUAi''2Ut
it is
is
Hermitian matrix
eigenvalues are
if its
UAi/2Ai/2Ut
all
is
Since
A/^.
its
eigenvalues
Another way
to check is to use
Sylvester's criterion.
Definition 4.10:
Theorem
4.10:
A proof is given in
Example
Q=
{g^}.
Then
< detQi =
<
IS
Problem
is
positive definite if
and only
if all
the leading
are positive.
4.6.
4.11.
Given
If
Hermitian matrix
principal minors of
g;
is
positive definite if
<
det Q^
=
is
and only
det f^J.^
\9l2
nonnegative
if
^^^V
...;
<
det
 detQ
922/
definite.
MATRIX ANALYSIS
78
Example
[CHAP.
4.12.
is
positive definite if
which
det
gn >
positive definite if
is
Q>
and either
qnQzzlfz >
and
^22
922 or
>
and
911922
^ut
0
gfg
>
The conclusion
0.
MATRIX NORMS
Definition 4.11: A norm of a
Ax[ KJIxjj
is
that
is
positive definite if
zero.
4.6
matrix A, denoted
for
A, is
the
minimum
value of
such that
all x.
defining relationship.
norm
is
Example 4.13.
To find U]2, where Ut = U"i, consider any nonzero vector
Theorem
norm
4.11:
(1)
(2)
jAxJI
A
= max jAu[j,
A
(4)
1AB^A[B[
(5)
A[ == jA
Proof:
To
show^
Since
(2),
[A[
Hull
of Auj is to be taken
B[
and only
A=
if
0.
k^^^,
=
=
maximum value
= 1.
v^^here the
+ B=^A[ +
(3)
All
are:
A [x
(6)
x.
aAu
x/a,
q:
where
Au
kx
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! ^
lAx[[
from the
j[Bx[
(jA
Bj[) x[
CHAP.
MATRIX ANALYSIS
4]
To show
Bx and
(4),
ABx
A(Bx)
79
AlBx
x:
for any x.
Ax =
Since
an eigenvector,
is
x
and
^^

Ax =
Ox for any
Theorem
4.12:
A
(6),
implies
A = 0.
Therefore
X.
A2
where
p^^^,
Ax
llAxIi
lAl!ixi
follows.
(5)
To show
Xx
from
The converse
p^^^ is the
^
calculate
IJAll^,
maximum
find the
AtA
rmax
At A.
Hermitian,
is
2=
pfg^
Ax2 s= o,
p2 is
in =U as
M
I1X2
11
eigenvalue of
AtAg.
(x,
M.
so that
obvious.
is
Proof:
Since
Since
0
maximum
p.
"mm
Then Ax =
Definition 4.11.
more
To
Ax.
Then
iiSr
llAxlE
112
Ifpf
\\^AS,\\1
t=l
Since
(pf)^;^
2=
the theorem.
4.7
^i
2 ihf
2
^u
=
(p?)n,ax
i=l
= p^^JM^ when x
is
the eigenvector
g.
belonging to (p^ax
FUNCTIONS OF A MATRIX
of a scalar
/(a)
a,
it
k=
where
f^
d^f(a)lda!'
Definition 4.12:
evaluated at
0.
is
/(A)
/() of
a scalar
a,
the function of
(1)'"A2'"/(2to)
annxn matrix
S/feAVA;!.
k=
Example
4.14.
Some functions
cos
of a matrix
A =
(cos 0)1
eAt
are
(sin 0)A
(eO)i
(eO)Aj
+
+
(cos 0)A2/2
{e'^)\H^I1
+ \H^lk\ +
MATRIX ANALYSIS
80
Theorem
Note
Proof:
Theorem
T^AT =
If
4.13:
A''
/(A)
From Theorem
Proof:
/(A)
where
/(X,)x,rt
is
T/(J) T.
A = TJT^TJT^
A = TAT^S
If
4.14:
= AA
then
J,
[CHAP.
TJT^
= TPT^.
Then
/(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
and
{^.13)
(Ali),
7(Ai)
/(A2)
/(A)
(xi
x)
X2
/(An),
'
/(Ai)rl
f iK)4
X2
(Xl
I
Ix.)
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
Let
/(a)
Ai
for uniqueness in
in
Theorem
Theorem
4.14
to
4.9.
A = TATi;
A = 2\x,rt
of
CIHAP.
MATRIX ANALYSIS
4]
To
is
calculate /(A)
From
/(J).
equation
LL(A)
Hence,
it is
Ixl matrix
By
calculation
it
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
81
number
i^i^)
A"
of combinations of
w elements taken
at a time.
00
Then from
/(L)
2 /fcLVA;!,
fc=0
but
The
""
by comparing
d''/(A)
(i1)!
dA'i
{i.l7)
and
{i.lS),
Therefore
//(A)
/(L)
d//dA
...
[(Zl)!]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 CayleyHamilton
theorem.
Theorem
4.15
The proof
(CayleyHamilton):
Given an arbitrary nxn matrix A with a characteristic polynomial ^(A) = det(AAl).
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)
,(A)
A2
 6A +
= C'
51
 e(l
'')
ON
+ 6('
I)
/O
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 =
(i.21) again:
Ai
For an
4.16:
nxn
y^
/(J)
if it exists,
(6IA)/5
+ Y^A"^ +
yiA"i
y_j
A + yj
y,Ji
/(J) is
301
matrix A,
/(A)
Here
in
(iSl)
51
Theorem
by
y,J2
y^j J
+ yj
{^.19),
(i.li)
if
can be
diagonalized.
This
/(J)
i y^J"*
for the
/(A)
The quantity
Also,
cos
solving
A"
'E
'^km^"'
SA
by
CayleyHamilton
the
!:/,, JA;!
= TV(A)T = Tii:y,A"*T =
=
l
^ y,TiA"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
A =
cos ,
cos
5/3
15
2\
cos
1/1
~1\
\l
\)
cos 5 /l
51
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
4.17:
~ m{sipi)'ds
/(A)
Since
Proof:
that
/(L)
/(A)
T/(J)T
^ f{s){sl sy ds\T\
= l\ ^.
is
it
suffices to
show
Since
ls\
1
...
s
A,
then
/(sA)'i
1
(^IL)i
(sAy2
...
(sA)'i
...
sA
...
(sA)'\
,,
(sy)'
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
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
/s3
(
2
3 /
_
~
/s
1
s2
6s
5 i
Then
cos
2W5'
(s
l)(s
 5)
3
s
MATRIX ANALYSIS
84
[CHAP. 4
If
1\
V1
1/
cos 1 /
/2
cosl
2\
cos
5/1
Vl
[2
5) V 2
COS.5_
i
ds
PSEUDOINVERSE
4.8
the determinant of an n x n matrix is zero, or even when the matrix is not square,
way to obtain a solution "as close as possible" to the equation Ax = y. In
x w real matrix and first examine the properties of the real
section
we
let A be an
this
respectively.
A'^A
and AA^, which are w X and
square
matrices
symmetric
When
there exists a
Theorem
mxm
4.18:
nonnegative
is
^ = y'"y,
Proof:
where B can be
m X n.
x'^B'^Bx s^ 0,
Then
definite.
which
B^B
so that
is
is
= Bx,
B=
A''
in
Theorem
Theorem
4.19:
This
is
4.18.
r^m
(1)
p?
(2)
Ag;
(3)
Ag;
(4)
A^f.
(5)
A^fj
p^i^
p,g;
for
i^l,2,...,r
for
= r+1,
for
for
= r+l, ...,m
..
.,n
l,2, ...,r
of A'^A and
Proof:
mvector
h.
as
h.
Ag./p^
for
AA''h,
1, 2,
. ,
r.
AA^Ag./p,
.,n.
AA^.
Define an
Then
A(p2g.)/p,
p^g^g/Pj
pfh,
Furthermore,
hfh.
gJA^Ag./p,p,
S..
is
Since for each i there is one normalized eigenvector, h. can be taken equal to . and (2)
and
proven. Furthermore, since there are r of the p2, these must be the eigenvalues of AA'^
and are ortho.,m such that AA^f =
Also, we can find fj for i = r+l!
(1) is proven.
basis
orthonormal
are
an
the
and
for
=U
basis
orthonormal
are
an
g^
Hence
the
fj
normal.
for =U.
.
To prove
Ag.
0.
(3),
Finally, to prove
(4),
A^f,
A^(Ag/p,)
(A^AgJ/p,
pfg,/p,
PSi
CHAP,,
Example
r
MATRIX ANALYSIS
4]
85
4.19.
.
Let
/6
"'^
0\
5
m = r = 2,
Then
w=
4.
/72
^^
73;
(^36
24
\36
The eigenvalue
''3
^'
From
it is
''4
AA^
of
are
p\
and
100
*^"
16
36^
eigenvalues of
gj
0.84
0.08
0.24
0.48)^
g2
0.24
0.12
0.64
0.72)1'
gg
(0.49
0.00
0.73
0.49)T
g4
0.04
fi
0.6
0.8)T
fa
0.8
0.6)T
A^A
are
p2
p2
100,
25,
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
The
Proof:
f J,
=l
.^m ^o^'"
Use of properties
(2)
Ax
and
(3)
of
Theorem
,Ag,
where
ligf
(=1
g^, gj,
. ,
g and
in IJ^,
= gTx
{i..22)
4.19 gives
i I^Ag,
i=l
i^
Also,
For arbitrary x
i,Ag,
i=r+l
1=1
i.p^i^
1=1
From
{i.22),
theorem
is
$^
and so
gTx
Ax =
ig/igfx.
'"^
proven.
10
,
Definition 4.14:
6/
A is
4.19,
0.48)
"^'^
5(
(0.24
V~06/
of the
mxn
real
matrix
the
is
real
Example
matrix A~^
p.igjfj'.
4.21.
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
0.11680.0576
\0.0864
0.1248/
nxm
86
MATRIX ANALYSIS
Theorem
Given an
mxn
equation
Ax =
forthose
4.21 :
In other words,
matrix
real
y.
= A"'y.
]Az y2 =
Define
xo
xo such that
[CHAP.
y2,
jjAxo
then
consider the
y,
Axo
>
jz2
y2 and
xoj2.
if
Ax =
If the solution to
Using the notation of Theorems 4.19 and 4.20, an arbitrary mvector y and an
Proof:
%vector
x can be written as
arbitrary
m
where
rj.
fjy and
.
gfx.
Ax Since the
'
i=l
2(^iPi'?/
t=l
those vectors z in U
for
r+l,
.,n
is
The
with
ivjp)'
r+l
Ij
f or
minimum norm =
+ 40:3 = 1
=
Ax y, where
6X1
norm
Then
. ,
Then using
to.
r;j
ijy from
2=
pi'/jgi
Pr*Sif^y
= A'V
Xo.
4.22.
rank of
l,2, ...,r.
i=l
Example
^f
= r + 1,
with a
= r+l
i=l
U.;25) gives z
Vf
But
arbitrary.
M\i
(4.^i)
rt
i=l
j
2
= r+l
the best
where
2 vA
t=r+l
To minimize Axy]2
4.19 gives
lAxyi!^
Theorem
of
i:iliPv)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
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/
exists,
x = A'y +
(I A~^A)z, where
li
i=r+l
i=l
where
z
C
for
S^,
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
gj
(IA'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 + (IA^A)z
Some further
nonsingular, A~'
1.
If
2.
A'A
3.
AA^A = A
4.
A^AA^ = A^
5.
(AA^)''
6.
(A'A)^
7.
A^Aw =
8.
A'x =
for
9.
is
AA^
A~'.
in general.
= AA^
= A'A
for
all
all
A^(y + z) = A'y
space of A'^.
in the
range space of
A''.
+ A^z
for
all
and
z in the null
all
10.
Properties 36 and also properties 79 completely define a unique A"^ and are
sometimes used as definitions of A^^
11.
12.
A=
diag
(Ai, A2,
H = Ht.
Let
13.
14.
=A
(A^)' = (A^^
where some
A)
is
taken to be
H = UAUt
15.
. ,
where Qi
.... A"')
may
Xi
be zero.
0.
Ui.
Then
= H^At =
(AH')+
where U+
11.
H = A+A.
Then A"^
(A')^
16.
17.
18.
r.
If
A'
>
Where
(P,
Q)
is
lim A^(AA''
+ eQ)i
:=
lim (A''A
may have
+ eP)iA''
discon
MATRIX ANALYSIS
88
[CHAP.
Solved Problems
4.1.
Show
that
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
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
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
/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
2 \
/O'
2 j Xi =
X3
1Xi/
,3
\0,
_8 2\ /4
_3 _2
3
4
1/V2
/1X
Calculation of
3.
1Xy
(b)
2\
11/
1
1X
det
10
1/ \0
\2
o\
2\ /l
/4
x,.
3/
1
\
4=0
3x/
gives
Xj
X2
X3
1.
Then
eigen
CHAP.
MATRIX ANALYSIS
4]
(A
0)"" where a
and only one vector, (a
one Jordan block Lu( 1), so
is
89
( l)I)x
gives
Therefore
arbitrary,
it
is
only
Lii(l)
Solving,
gives
ti
(y8
2a
where p
a)'''
is
arbitrary.
Finally,
from
11\
/O
Vo
4
2/
t2
(c)
Choosing
a,
p and y
to be 1,
A A
14(0
2/V
21
3/Vo
1.
Since
and
o\ /_!
1
t,
The
A(Xitix2)
How
A(X2
(Xitix2)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 2Sa
T
tl)J2
xi
(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)
)(
2/3/
0,
S are also
but in general
1
0\
)
1/
(AXI)x = (I(l)I)x
Ox
it,
since
4,3.
[CHAP.
MATRIX ANALYSIS
90
Show
The proof
fc
is
U=
matrix and
Xfc matrix;
i.e.
U^i
UjTi
Ai
Nu
Let
uIiNfc_iUfci
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
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
kP
Therefore
ai2, aig,
. ,
aifc
must
Xi
TtNfcT
all
be zero so that
...
=
*fci
and Afc_i
exists a
is
normal.
Ufc_i
= V^l^
22
2, fc1
lcl, 2
fcl, lc1
where
Since
Afc_iisA;lXfel and
such that
U^i Afc_iUfc_i
CHAP.
MATRIX ANALYSIS
4]
...
S^
91
Then S^S^
and
U,kl
o\
...
...
Xi
0\/l
...
...
X,
Si^TtNfeTSfc
uJ1
Ufc1
/
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
(TS,)t(TS,)
4.2,
unitary,
is
Nj.,
B = Bt
if
can be simultaneously
and only
AB = BA.
if
then
all
commute.
AB =
and X = (xi X2
xJ, C = XtBX = XtBtX = Ct, so C is an m X m Hermitian
Then C = U^D^U^ where D and U^ are m X m diagonal and unitary matrices
respectively.
Now A(XU) = X(XUm) since linear combinations of eigenvectors are still eigenvectors,
and D = uXxtBXU^. Therefore the set of m column vectors of XU^ together with all other
C=
{cy}
inatrix.
4.5.
Show
= (U+XtXU) =
Q,
is
unique.
is
(xi
.
I
x)
.
I
xxxt
MATRIX ANALYSIS
92
[CHAP. 4
where
t1=T'.
ym)
(xi
ym)
(Xx
(Xl
Xm)Tm
Ym.
Then
(yi
Hence
yi , y2,
and
V may
^m/'wJn
X^)
differ slightly
when
values.
4,6.
is
positive definite if
and only
If
detQm >
for
m ==
1, 2,
we
.,m,
proceed by induction.
We
0,
then
Q_i
is
qt
q+Q;"!
Qn
>
detQ
 q+Q;^iq >
0
^nn
detQ =
positive definite
0.
q+Qniq /
3.5 gives
det
Qn
Xi
>
0,
and must
gnnqQniq
for any vector (x+_i
substitution into
that
Let
Sfc
if
(x,
af).
A!
"
<
A"
1,
vfill
then (IA)i
defined
give
(y,
by
Qy)
>
0.
2 A".
n=0
and
n=0
IS
Sfc
Then
A".
= +l
fc
Acl
>
(xjila;*)
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
Afe
+ ill
so that
(IA)S =
fc^ gives
because (IA)x
(1
Theorem
of
(4)
Using property
4.11.
A^ + i =
lim
Since S exists,
I.
A)x
93
(IA)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
)xj.
112
4.8.
A =
12
1^
The
1
r;
and eigenvectors Xj
j,
1 0.5)^,
(1
The matrix
h^^i^.
i=l
and Xg
The reciprocal basis
Xa
A = 5
spectral representation of
1  J
xj
{j
has
eigenvalues
and
1 0)^
Xg
X,
1,
{j 1 0)^.
can be found as
rtj
JJ
}^'
1
\0.5
0,
Then
A ^
4.9.
Show
(l)f
the
that
(A~'A)^ = A~'A
finition
2)
j(0
(li)
(l
+ j)(
{0.5}
0.5
AA^A =
A = 2
2 pigk^tlc=l
A"^ =
and
Pitigi
i=l
(2
Pifig.n(2 p'ejtj)^^
222
T
i=l
and
Sjj
Then
It
AA'A =
Since gjgj
f ff^
=i
fc=l
PiPr'pfcfigigjfj
pAel)
ffcgfc
>
8,^,
AA'A =
A^'AA~' = A^^ and from
A'A
equation
gigf
"
^A"'
Pifig
(^.26),
(2
gigf)^
and
(AiA)T
Similarly
(AAO'^
= (?
^i^i)'^
2.
AXA = A
XAX = X
and transposing
9.
+ j)
4.14.
Represent
Similarly
solutions
3.
(AX)T
4.
(XA)r
= AX
= XA
and 5 gives
A^X^A^ = AT
10.
ATY^A^ = A^
5.
6.
AYA = A
YAY = Y
Then
7.
(AY)^
= AY
8.
(YA)r
^ YA
MATRIX ANALYSIS
94
[CHAP.
2
4
10
4
2
8
X = XAX = ATX^X = ATY^ATX^'X = A^Y^AX = A'^'Y^X = YAX
= YAYAX = YAYX^AT = YYTATXTA^ = YY^AT = YAY = Y
Therefore the four relations given form a definition for the pseudoinverse that
is
equivalent to
Definition 4.14.
4.10.
of i/ of a certain experiment is thought to depend linearly upon a parameter X, such that y = ax + (3. The experiment is repeated three times, during
which X assumes values xi = 1, xz 1 and xs  0, and the corresponding outcomes yi = 2, and 2/2 = 2 and ys = 3. If the linear relation is true,
The outcome
2 =
a(l)
= 40) +
a(l)
13
13
However, experimental uncertainties are such that the relations are not quite
in each case, so a and p are to be chosen such that
satisfied
j:{y,ax,pf
i=l
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
yAxol^
P2
and
fa
= d
/?,
as
(2/i
and
Vs,
best a and
gi
D/Vs:
 o*i  ^o)"
=
0)
(1
Now A'
and
ga
be
can
Since
minimized.
is
(0 1).
Since
calculated
from
Definition 4.14 to be
A
(1
\8
and
/3o
?
3
1
Note
4.11.
Show
C=
this procedure
that
eB, or
if
can be applied to
an n X n matrix C
B=
is
(Vi
~ ^i ~ /^*i ~ "y)^'
^t*^
In
JTi,
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]
In
[M +
 XI]
L(X)
 XI
I In
2 MM"* [^C^)
3^1]*
is
^'
0.
Then
[L(X)
= IlnX
InL(X)
Since
XI,
InL(X)
95
(iX)'[L(X)
 XI]
because Ci 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 Ci = eB.
Supplementary Problems
4.12.
Why
4.13.
4'
A =
where
1^
2
4.14.
Suppose
4.15.
all
the eigenvalues of
are zero.
A=
0?
t(
(AXil)' + 1.
4.16.
4.17.
and B.
Is
an eigenvector of (A2B)?
also
Xg,
where
Xi v^ X2.
4.18.
x,(n+l)\
..(n
4.19.
Show
that
all
/O
_

+ l)j
[2
2\/xi(n)\
3JU2WJ
U^W)
UtU =
of one.
4.20.
Ut/
such that
"
]U
/I
4.21.
A =
Vo
3/6
4/5
/2
/ x,(0)\
"'^'^
to
Jordan form.
I,
[1
MATRIX ANALYSIS
96
4.22.
4.23.
A =
1
4.25.
4.26.
4.27.
Show
4.28.
Show
^0^
What happens
].
A2/(A)
4.31.
^=
^^
(xj
.,x of
4.32.
4.33.
Suppose A^
4.34.
Find
4.35.
4.36.
4.37.
4.38.
4.39.
Given an
Can
0.
where
is
A =
Is it positive definite?
A" +
oiiA"
+ aI 
Does
Xj.
a.
[BaldXy daldx^
is
and
y,
its
A3(A
 Xil)(A
Xjl)
0?
grad^x^Ax
r.
for a nonsymmetric A.
Show
that
2^
Xjrj
'
a
3
/'4
f
/I
A =
are true.
listed properties 118 of the pseudoinverse
real matrix
and scalars
Agi
that only
and
be nonsingular?
J,
mXn
if
Xi
and a scalar
^n)"^
grad^a
n =
gives
fjfj
0?
[/(A)]A2?
4.30.
metric
Does
Show
.4
e^t,
as
e/2
n=
4.29.
4.40.
is
Ly(Xj).
A =
Jordan form.
its
).
4.24.
that
Find
0.
1^
o""
10
to
P2
such that
[CHAP. 4
+ 'm)X{n + m)
Sy
At,
e^i
if
gjgj
= ^r^Si
Sy.
matrix
such that
ffj
'
derive
the conclusions of
Theorems
4.19
and
4.20.
CHAP.
MATRIX ANALYSIS
4]
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
...
...
if
show
."...''!..;;.....".
\0
where
97
'
ai
a^
K..
Lj
ai
\0
4.42.
where
; is
Show
that for
Show
4.43.
that xtAxl
Show
4.44.
an arbitrary constant.
\\\\\^
\\s.\f^
= (
(A0)'
0),
).
(A^A +
eP)  1 A'',
positive definite
Show
4.45.
AJ
if
and only
Answers
Because
4.12.
det
(A X^I) =
0,
4.14.
No
4.16.
Yes
4.18.
Xi(n)
3,3,3,
and
xj
ajgCw)
4.20.
/I
4.21.
T =
0.6
0.8
0.8
0.6
Vo
4.22.
jr
a(2
any
0.
1)
is
Supplementary Problems
to
4.13.
A=
if
where P
The eigenvector
/3(0
for all n.
1 0);
xlg
AX,I
of
X; lies
y(l
1)
MATRIX ANALYSIS
98
'O
4.23.
T =
^a
4.24.
'''
+a
/2
(l
Ai
value
4.29.
Yes
4.30.
Yes
4.31.
(AT
4.33.
No
+
8et \4e +
is
eAt
eAt
eat
0\
<x^[A~^
+ a^A.^^ +
e);
].
0,
Xj
If a
XiXg.
as
all
1/
.X
Xj
Xj,
0,
Xj.
Slight perturba
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,
Pi
pj
4.43.
singular.
9e2 +
3e2t +
6e2t
/I
4.40.
+ A)x
4.37.
\0
A'^ 
exists,
zero and
is
/_4et
436.
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
x2AxJ2
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 zeroinput 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
io),
is
the
nxn
*(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 zeroinput linear state equation dx/dt = A(f)x
gives the matrix equation for *(i, to),
This
x{t)
is

^{t, io)xo
d^{t,to)/dt
Sinct! for
any
xo,
xo
x(io)
*(^o, io)xo,
A{t)^(t,to)
{5.1)
to) is
[5.2)
to a timevarying
case,
the discrete
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)
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)fli(fc)
(5.11)
separation property
(7)
determinant property
(8)
det*(A;,m)
[detA(fel)][detA(fc2)][detA(m)]
f or
>
A;
(5.12)
is
Proof of Theorem 5.1: Because we have a linear zeroinput 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.
property, partition
det
* =
row vectors
into its
^^/\^^/\
^j, ^^,
.,
Then
^.
r\^^
and
d(det ^)/dt
d^Jdt a ^^ '^
+
From
{5.1)
for *, the
a ^ + ^j a d^Jdt
^j A ^2
^
a.
A.4,^
A d^Jdt
{5.13)
related
by
d^Jdt
2 *(*)
f or
*)c
1, 2,
^jA
Ad^ydtA
A^
Aj^^a.^^^A
*lA
A^
ai\,{t)
a,^i A
is
continuous
A ^j A
A ^
{5.13),
[tr
A(t)]^i A ^2 A
.^
[tr A(t)]
det
trA{t)dt
where
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
The proof
and
is
= 1 = y,
= , the
det I
f{t)
quite similar,
CHAP.
5]
101
5.2
Theorem
The
5.2:
eA"^)
{5.U)
The Maclaurin
Proof:
series for
e*'
is
= A""
^= A.H'^/k
(5.15)
which
!,
is
uniformly convergent
fc
as
shown
in
Problem
4.27.
is
is
gives de^ydt
a solution.
from
*(t, t)
A'^+^i'^/fc!,
Furthermore, for
I.
it is
a solution,
e^''e*'
gActo+t,)
Since timeinvariant 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 RungeKutta 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
Tei'Ti
{5.17)
=
3.
^ AH^Ik
CayleyHamilton:
e^i'Xji
nl
2= y,{tW
oAf
{5.18)
nl
w'lere the
y^{t)
e"
2=
7i(^)J'
{415),
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
M^i)
'
(5.20)
I
\
...
,0
\il
then
...
4.
e'^i*
Resolvent matrix:
eAt
where
R(s)
(si
^i{R(s)}
[5.22)
A)"^
The hard part of this method is computing the inverse of (si A), since it is a polynomial in s. For matrices with many zero elements, substitution and elimination is
about the quickest method. For the general case up to about third order, Cramer's
rule can be used. Somewhat higher order systems can be handled from the flow diagram
of the Laplace transformed system. The elements nj(s) of R(s) are the response of the
For higher
ith state (integrator) to a unit impulse input of the yth state (integrator).
order systems, Leverrier's algorithm might be faster.
Theorem
5.3:
Define the
Leverrier's algorithm.
scalars 9u
62,
F2
= I
= AFi +
= AFi +
Fi
nxn
and
.,9n as follows:
02
= trAFi/1
= trAF2/2
6n
= tvAF/n
61
^iI
eiI
Then
Also,
AF +
Having
tfnl
0,
R(s), a
an
S'^'F^+S'''F2+
t
n _^nl
,n
...
^lS"l
^

taT/i\l
A^l
(.J
(SlA)
S"
Proof
is
+SFnl+Fn
^^^"^^
+^_,o4^
+ dnlS + 0n
First, factor
detR(s) as
det R(s)
s"
^isi
eis
^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
(s
 A,)R(s) s=x,
{5.26)
CHAP.
5]
A,
the residue of
A)"'
(m i)
is
A^^
Ri
Then
(s
 A)'"R(s)l
\{s
{5.27)
e^' is easily
103
gAt
gA.tR,
gXj,tJJ2
Ri
Xj
+ eMR
{5.28)
{5.17),
XiTj
r;
R;.
^i{(sAi)"'"}
To
find A'',
1.
Series method:
methods similar
ESigenvalue method:
{5.30)
A''
= TPTi
A''
and for
{5.29)
A*^
2.
t"''e'''V(ml)!
{5.31)
distinct eigenvalues
= iA^x^rt
A^
i=l
3.
CayleyHamilton;
n1
A^
^lAkW
=
{5.32)
from
J'=
=
I
^
=1
7i('^)J*
{A.15),
^Um)
L^i(Ai)
{5.33)
L?,(Ai)
Ul^{k,)
..
and
if L,ii{M) is
as in equation {5.20),
'Xf
L5i(A0
4.
L^(AO
kxf'
+ !)!] '^
(fc!Af+2')[(Z2)!(A;? + 2)!]i
(fc!Af+i')[(il)!(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,
104
[CHAP.
ko
systems with
roundoff,
and
A;o
difficulties in
Otherwise
it
suffers
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 CayleyHamilton method is very similar
few more multiplications.
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
inputsingle 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.
5.3
There is NO general solution for the transition matrix of a timevarying linear system
such as there is for the timeinvariant case.
Example
We
5.1.
A = TJTi
= *(tto) =
Then
A{t)
T{t) J{t)THt),
change of variable
T(t)y
= TeJo)Ti
case,
A(<)x
t.
Attempting a
results in
dy/dt
J(t)yTHt){dT{t)/dt)y
or
We may
The behavior
Example
5.2.
Given the timevarying scalar system di/dt =  sgn (t  ti) where sgn is the signum function, so that
''
sgn(iti) = 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 timevarying system.
Also, the
whereas this
phenomenon
is
CHAi'.
5]
Exam)le
105
5.3.
d^dt
il{t
ti)2
Then
and
with
tlie
(to)
?o.
to
<
*i
?o
[(toti)~i (tti)i]
ioe
to
Fig. 51
I'hese
more
and other
difficult
mEition.
is
mati'ix.
5.4
d^/dt
a{t)^
has
Proof:
Theorem
e^
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 timevarying linear differential sys
Proof:
Use of the
series
Integrating and
a{t)dt.
is
usually
*(t, t)
met only on
final
'^
examinations.
Taking derivatives,
^g/>^'*''
A{t)
{5.86),
A(i)e^'''""*
=
if
A{t)
and only
^^A{r,)d^
A(i)
A(r?)d,
if
J\{r,)dr,A{t)
(5.37)
106
A(i)A(r)
of
Example
A{r)A{t)
r)
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).
matrix
^{t,to)
eAi"'i>$(f.^io)
for
ti^t^ti+i
Proof:
Use of the continuity property of dynamical systems, the transition property
(equation (5.5)) and the transition matrix for timeinvariant systems gives this proof.
eA"*'
for
to^t^U
etc.
Example
5J5.
Given the flow diagram of Fig. 52 with a switch S that switches from the lower position to the
upper position at time t^.
Then dx^/dt = Xi for o  * < *i and dx^/dt = 2xi for i ^ t.
The
solutions during each time interval are Xi{t) = aJiQe'"'" for t^^ t < t^ and x^it) = a;j(ti)e2"*i> for
ti
t,
where
Xiit^)
a;ioe*i'"*<'
by continuity.
Fig. 52
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.
is
d<"
nomial,
2/
(lni)'""*i^
V'.
Then a
order
a solution for
0, 1,2,
.,m 1.
A,
analogous to the
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)
t^)y 2ty +
Hermite equation:
or, vi^ith z
+ {lt^ + 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.^
+
+ {lt)y
/,
ay
j.\i\
(k + lt)y + {ak)y =
+n
,
,
Hypergeometric equation
^'''^'"^'^t{l
Confluent:
Mathieu equation:
i/
or,
with
cos^
ty
{yt)y
(a
/3
ay =
cos t)y
t,
4T{lr)y + 2{l2T)y +
[a
+ l3{2Tl)]y =
The
{5.39)
sdlutions
infinite series.
found.
5.5
Floquet theory is applicable to timevarying 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 periodicallyvarying systems.
Theorem
5.7:
(Floquet).
where
A{t)
A{t
o,).
Then
m,r) =
where
P(f, t)
is
P(i,T)eR"^'
a constant matrix.
d's./dt
A(^)x,
108
The
Proof:
Setting
[CHAP.
= t + m, and
d^{t,r)ldt
using A(i)
A{t
A(i
+ ,)$( +o,,t) =
a*(f+o,,T)/a*
with
A(i)*(i,r)
*(r,T)
gives
a.)
A(i)*(^ + a.,T)
(5.40)
was shown in Example 3.26 that the solutions to cbs./dt = A(i)x for any initial conform a generalized vector space. The column vectors ^^t, r) of *(i, t) span this vector
space, and since det *(i, r) v^ 0, the ^i,t, r) are a basis. But equation
{540) states that the
It
dition
^^{t
<o,
t)
are solutions to
dTi./dt
A(i)x,
so that
i>i(t+o>,T)
C=
z=
for
c^ii'iit.r)
1,2,
.,%
{Cji},
*(+,t)
*(i,r)C
(5M)
Then
C =
*(r,t)*(i + o>,T)
Note that C^ exists, since it is the product of two nonsingular matrices.
Problem 4.11 the logarithm of C exists and will be written in the form
C =
If P{t,
r)
it is
e"R
Therefore by
(5.^^)
P(,r)eR>
(5.^5)
t)
*(i+<o,T)eR"+"^'
*(,r)eR'
#(;t,T)e"Ke"R'^'+"^>
V{t,r)
From
hnd
fl
Example
5.6.
CHAP,
fi]
109
Fig. 53
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.
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.
5.6
Theorem
5.8:
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)
+ r
C(t)^t,T)B{T)u{r)dT
D(i)u
i5.U)
The
integral
is
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.
dy.ldt
Use of equation
dk/dt
to
ta
gives
gives
*(to, t)
^{to,t)B{t)u{t)
*(i, io)
k{to)
+ f
=U
gives x(io)
$(fo,i)x()
Multiplying by
+ f
x(io)
dr
(546)
k{to),
x(t).
dr
Substituting into
y(i)
C(i) x(i)
y{t).
x(i)
eA(tt)x(io)
and
(5.14.)
gives
eA~^)Bu(T) dr
(5.^7)
^t
y()
This
A{t)$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)
Theorem
5.9:
+ 1) =
A(A;) x{k)
C{k) x{k)
+ D{k) u{k)
y{k)
vs^ith
transition matrix
*(&,
m) obeying
*(fc
{240)
+ 1, m) =
m)
A.{k) *(&,
[equa
Then
tion {5.3)].
= nA(i)x(m)+
x(A;)
i=m
Proof:
B{k) u{k)
Stepping equation
1)
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
{2.40)
^
=m
+ 1) + B(m + 1) u(m + 1)
1) x(to
x(w +
1),
1)
1)
u(m +
1)
= A{k  1)
A(to) x(m)
+ A(A;  1)
A(fe
 1)
+ 2) B(to + 1) u(m + 1) +
+ B(A;  1) u(A;  1)
This
is
A(to
CHAP.
5.7
5]
111
equations as
Icl
 Yi Mi)
m)
*(fc,
for
f^
>
{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. 54
u(tfc)
U=
for 4
h gives
*(i)c +
^{tk+i,tk)
Bs(k)
Cs(k)
C(k)
D^A;)
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)
(2.40)
1,
f<4+i,
sampled system.
e^'^'^'^B dr, Cs
= C and Ds =
For timeD.
is (see
Since
the com
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) =
linear,
112
5.8
y{t)
y(i)
C(i) *(i,
r)
[CHAP.
y{t) is
D(i)u(i)
(5.53)
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)
^^^
0)
matrix
= C(sIA)iB+D
J:{U{t,0)]
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
5.9
k>m
l)B(m)
(5.58)
is
= C(2lA)iB + D
{5.59)
Definition 5.2:
The
We are
Lx)
for
(Lap, x)
A{t)x.
Defining
ptxdi,
equaLiuii yo.ou)
usmg migration oy
(p,Lx)
=
"
p(^o)
parts.
f\^A{t)xdt
X
p(^i),
[''^^^P
we
all
L=
is
defined
by the
x and p
A{t)
 Id/dt,
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
a^it,to)/dt ==,9^A^{t)t{t,to)
5.10:
dx/di
A(i)x
pt(f,)x(i)
Also,
Proof:
*t(i,io)
u(i)
adjoint system
its
dp/dt
pHt)B{t)u{t)dt
{5.63)
*\t,to)
{5.6i)
^i(*'*o)
(dpt/dQx
= A(i)x + B{t)u
d(ptx)/dt
Integration from
to
to
then yields
ti
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.
{5.61) gives
pt(*)x(f)
Therefore*(5.(J4)
x(io).
+ ptdx/di
Furthermore
(5.55).
pt(io)x(o)
for any
(5.62)
0,
d(ptx)/dt
From
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
Theorem
(5.60) it
defined
is
dp/dt
all
113
*(i,o)x(io)
way
*(i,T)B(r)u(T)dr
'to
The variable of integration t is the second argument of *(t, t), which sometimes poses simulation difficulties.
Since <to,T) = */(t,<o) = '+(t,o), this becomes
x(f)
4^it,to)\x{to)
J i^T,to)B{r)uir)dT']
argument of
*(t,
{5.65)
to).
final value is
5.7.
'
Givim
dx/dt
[of
2t
2f
]*
is
dp/dt
*{t,r)
/I
tl
Xi(2)
to
find
1.
3\
jp.
0.2e<^/2)(r^/2)
/ 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)
if
p(l)
(0.6ei5
+ 0.4e6)a;j(l) +
0) x(2)
[CHAP.
a;i(2)
pt(l) x(l)
(0.4e6o.4ei5)a;2(l)
Solved Problems
5.1.
A{t).
at time
(5.2) in (5.1)
to
*.
Mi) =
evaluated at
3*(*, to)/St
to
This
is
a quick
5.2.
/I
1
method, (b) eigenvalue method, (c) CayleyHamilton 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
/et
eAt
\
(6)
(l2t)e2t
te2
(l
+ 2t)e2t/
Using equation
{5.17),
eAf
A
=02
(a).
(1
0)'''
and
(0 2
1)'''.
CHAP.
5]
c)
2 and
2.
To
115
(5.18),
/I
el
Vo
0\
1
62',
/I
\0
Yi
1/
0\
02
/I
72
2/
\0
=
=
=
e~'
~^'
ie2t
yg
yj
then gives
0\
(4e3e2t2e2t)[
/I
(ete2tie2t)
(4e< 4e2t
3e2)
1/
/I
0^
12 16
Vo
(dl)
472
4e
\0
Summing
T2
 2yi +
 472
ro
/I
eAt
3e2t  2te2'
= 4e 4e2t  Ste^t
= 6 e2t ie2t
=
7i
(5.18)
Vi
7j,
7o
Using
70
4,
(a).
Xio
^(Xj)
S^{X2)
X20
^JLiXi)
sJLix^
a;ao
^(a;2)
4^(X3)
~1
2
_ r 1
 [_7+2~(7T2)2j"'20 +
^^*'2)
^^^3)
a^io
<!(a;i)
.C(x)
(g
+ 2)2
Ls +
(s
4x
430
(^
+
2)2
R(s)a;o.
/
R(s)
'+
=
s
(s
\
(s
+ 2)2
(s
+ 2)2
From
(5.22),
+
Ri(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)^
0'
R(s)
8+1
+ 2)2
(s
'
^0
The
02
1
(dl).
shown
in Fig. 55.
1*10
<:(a^3)
C(i)
1_
<3Hj
iO
i.
Fig. 55
For
xio
For
Kjo
For
a^ao
=
=
1,
.C(i)
Therefore,
\
8+1
4
R(s)
T^TW
(8
^1
\
Again a
(d4)
(7+2)2
(s
+4
+ 2)2
Using Theorem
F,
Using equation
51
81
9i
92
fls
(5.25),
74
0\
1
R(s)
+ 2)2
sf
s8
5s2
8s
CHAP.
5.3.
5]
+ l)
x(fc
(a)
(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)'^
j_.vl
+ s''^F2 +
s"'Fi
''
5lS"l
sFni
On
^fnlS
Fn
{5.23)
4.15).
Let
det(sIA) = <f>{s) = s" + 9iS"i +
+ e_is + .
Use of Cramer's rule gives
0(s)(sl A)i = F(s) where F(s) is the adjugate matrix, i.e. the matrix of signed cofactors transposed of (si A).
An intermediate result must be proven before proceeding, namely that tr F(s) = d^/ds. In the
proof of Theorem 3.21, it was shown the cofactor e^j of a general matrix B can be
represented as
Ci^ejAb2A Ab and similarly it can be shown Cy = bj a
Abi_i a e^ Ab^+i a
a b, so
that letting B = si  A and using tr F(s) = Cii(s) f Czzls) 4
+ c{s) gives
tr F(s)
cj
(sea
 ag) a
(sci
(se
 a) +
ai) A
(scj
 a{) ac^a
A(se_ia_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(se2a2) A
A(seaJ + !
(scj
 aj) a
Ae
trF(s)
is established.
Substituting the definitions for 0(s) and F(s) into the intermediate result,
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"! +
s gives
n.
{nk)$k
Rearranging
{5.23) as
AFfc
(5.66)
Ffe
I
sn2F2
+ F)
(5.67)
118
for
A;
1, 2,
.,w
1.
[CHAP.
nOk
and substituting
trAFfc
trFfc +
fcflfc
trAFfc
which are the other half of the relationships needed for the proof.
Using the
last relation
Fi
F2
ej.
Fg
e^I
9_il
eI
= AF
=
which
5.5.
is
+ AFi =
+ AF2 =
ffil
+A
82I
+ 9_2A +
i?iA
i.e.,
A2
eiA2
A"
then gives
9i A
eI
(?iA"i
+ A" =
0(A)
Ve*
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)Ti(t) and
eigenA(t) = T{t)A{t)THt) 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
A(t)A(T)
A(T)A(t).
For the given timevarying 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
fajet
rt,,
5.5,
*(t,T)
,r~i
 1
Te'r
Ti
*(t,r)
actT)
e
^
/
(
cos(e''e~')
_gijj(gr_et)
sin(e''e*)
cos(eret)
X2
= aje'K
Consequently
CHAP.
5]
To check
5'*
this,
use Problem
5.1:
\sm{e^
Setting T
sm{eret)\_^^^^_^^^/sm(eret)cos(eret)
''^^'''"*'~'^
^eitr)(
<'t
119
e~t)
cos{e^
et)J
\cos(e^
e't)
sm{e^e*)
cos(e~'o
e~h) = cos (e~'o e~*i + e~'i e*2)
To verify
i^o)
cos (e~'o
e~*i)
cos (e~'i
e~*2)
sin (e~'o
e~ti)
sin (e~*i
e~*2)
gtj) =
sin (e""*o
gti)
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
gti)
V sin(e~'o
To verify
*i(j,
fo)
*(*o.
*i).
calculate >(*!,
^'
"
cos2(e*o
equals *(*o, *i)
liUnce
To verify
e~"'i) +
*(i, *o)
sin2 (e^'o
sin^ (e~*o
gti) =
(<i)'(o).
9{t)
sin (e~'o
cos (e~*o
'i)
e'2
e~*2)
e~'i)
e^'i)
rule:
/cos (e~'o
gactoti)
by Cramer's
<o)
sin (e~'i
cos (e~'i
e~'2)
gti)
and
this
set
 et)
cos (1 e~t)
cos(le')
et/
sin (1
\^ sin (1 e~*)
Then
91(0
et
(1^"') sin(le*)^
/'=''
e*)
\sin(l
Since
cos(l
e)^
we have
cos(eoei)
cos(eol) cos(lei)
and
e2(tT)
and trA()
 ei)
verify this.
2a:,
property holds.
5.6.
dicldt
(
,1
(2a
(^2
dxjdt
find
l/t)dx^/dt
te*,
2a 
+ a)lt
+
=
[1
Xg
lit;
and so
(a2
+ a)/t]xi
trial
effuation, gives
t(e<^*
d'^xjdt^
2ae"tdxi/dt
a^e^tx^)
{e"t
dx^/dt
aea;i)
te^^Xi
d(ea;i)/df
x^it)
= dxjdt =
(5.38), so
ax^(t)
te^tx^
the solution
e''t[c^Ji(t)
is
c^dYa/dt]
120
To
so that
x(t)
and so
*(<,t)
at the point
and
[CHAP. 6
Ca,
FG(t)GMT)Fix(T),
where
0\
a
1/
^^ ^
^(t)
V^i(t)
FoCt)
dFo/d*
Admittedly this problem was contrived, and in practice a manmade system would only
However, Bessel's equation often occurs in nature, and knowledge
and Y^if) ~ ^ilvi sin(tir/4) gives great insight
that as t^ , Jo(t) ~\/27S cos{t!r/4)
accidentally have this form.
5.7.
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)
(AoAii)''2 if k is odd.
P(fe,m) = {AiAoi)i'2
must be outside the unit circle.
it is enough to find the eigenvalues
5.8.
+ 2) =
where
A^Ao.
AiAox(m).
Choose
with A(t).
(1
xi^y
= a
From
t)
This
is
the response
y(.t)
to
an input
From
5.8,
e''
u(t)
let
0)x
5.9.
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')
Sitr).
u{t)
6^"''^
y{to)^yo
and
{dy/dt){to)
5.5,
 e)vo +
"'
"'' sin
(e"^
 e'*) dr
ayo.
CHAP.
5]
r to
y(t)
ij
where e'^e*
in the integral,
efo>
cos(etoet)j/o
e*[l
gives
ii,
(e
cos
121
e*o)]
5.10.
U{s)
6/s
rr/
jl^'ind
The
easiest
way
to do this is
^{y(t)}
by using
classical techniques.
+
y
'i.
+
+ 6s
U(s)H(s)
1+34
+
+
os^
s*
The
how
= 1 + 3e~2t 4e3t.
corresponds with the classical techniques.
it
^G)
Co
(1 2)1
V^2
3)0
"'G)"
^^
=.
*(Mo)0
"
*(*.t)
+ J^\l
3e2t
2)
(^"'p'
^_3L,)(;)6(.to)rfr
4e3t
(5.e)
This integral
Instead,
we
_2
s +
1_
3
s2
5s
5.11.
t sing
We
y(t)
5s
_
~
(5.68)
is
a convolution and
+ 6/Vs
form
of control
^^'
as before.
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)]
However,
initial condition
t(r,
tf)
[t(t,
B(t) U(t) dr
tf)
x(f)
X(%)]
Supplementary Problems
5.12.
Prove equations
5.13.
5.14.
5.15.
Verify that
=:
*(t,T)
given in equations
5.16.
and
(S.12).
be found?
A(fe)
eA(fT)
and *(fc,m)
and
is
A'''"
system
5.18.
form,
/e4(tT)
e^' 
5.17.
(6)
to the
(c)
0)
dx.ldt
for
dx/df
/O
I
2\
'^
_3 j
^jx
using
reduction to Jordan
(a)
A =
0/
yo
_/3
Find eAt using the resolvent matrix and Leverrier's algorithm, where
A 
\
5.21.
eBeA.
What
5 20
(5.5)(5.12).
5.19.
is
0\
1.
3/
CHAI'.
5.22.
5]
123
5.23.
5.24.
5.25.
5.26.
A^ for
rule to find
series,
x/
\
using the matrix Laplace transform method.
5.27.
Compute
y{t)
0)x
(1
4m
solution obtained
5.28.
^ i\
+ 1) =
[^_l
y{n)
Compute
5.29.
(a)
(b)
5.30.
(1
Calculate *(,
Calculate
*(fc,
<)
dx/dt
..
/n
,,
_l)^(n) +[^lju(n)
x(A;
0)x(n)
=
1)
x(0)
/o
4M(n)
is
^
using
x(fc)
2 transforms.
How does the spectral representation for eAt extend to the case where the eigenvalues of
not distinct?
are
nl
5.31.
el*
=
i
For
2
=
5.32.
Show
5.33.
Show
5.34.
r)
A(t)x.
nXn
dx
dt
\a^t)
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
8{t)o
P(fc,
linear system
where
sin t\
x.
^
sm ,t smty
P(k,m)W<~'"
+ w).
*(k,m)
A(fc)
[CHAP.
<DH
,rN
i:_
8(t)
"Y
+
OJ/(t)
82
Fig. 56
It is
shown
yS
in Fig. 57.
Fig. 57
(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. 58 where S is a sampler that transmits the value
Find the state space representation at the sampling instants of
once a second to the hold circuit.
Use Problem 5.10.
the closed loop system.
'<..^r>^eM'.
Hold
s+
s2
5s
1
h
oy(t)
6
Fig. 58
5.39.
t^v
tv
+v
 p{t)
with
riito)
Vo
^^^
CHAP.
5.40.
5]
125
dt\xjI 
dJ\X2
dyUy+W
\c
ye
'yi\
h\fxi
(s
,
\*
3/2/
or
= Ax + Bu
""^
= Cx + Du
'"^j\'^%J
Find the transfer functions ^{2/i}/^{mi} and Jliv^lJiW^ using the relation
CAv^Uin) = [C(IsA)iB +
5.41.
x^sit)
dx,,/dt
initial condition
Xss(o)
(a)
where
is
^
an
K(7') is
Check that
u{t)
B()
K^^it)
h(t)
nXn
K{T)
t.
to
et+<'(tr) sin
dr
Hint:
(er
gt)
satisfies
T and
the system
is stable.
Find an
dr
independent of
t.
S(tT).
5.44.
1,
dx/dt
g/^
)x.
to the input
u(t)
10.0,
Xi(t)
and
x^it)
such
is
is
3/(2 \
Xiitf)
{lt^)^\^ =
A(()x
5.43.
that
time
to the initial
= f(t,^(0)f  f(t..W)^+J^f(,.)c^.
Xjt)
where
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.
126
Answers
=
A(k)
5.14.
If
5.15.
deteAUT)
5.16.
A =
5.17.
(f.r)
= ^^etrf^\(l
and only
'':
5.19.
eAt
5.20.
71
Supplementary Problems
to
*{k + l,k)
5.13.
"
[CHAP.
AB  BA
if
detel"'"'
eAeB
does
e<triSh
eBgA.
e(tT)trA
/2 2'
2 2
2)
^V1
e^<rof
2)1
9,
72
26,
yg
24
o\
/O
eAt
o.5e2t
zz
(1
+ e^t (
(6e2t8e3t
eAt
5.23.
(sIA)i
+ 3e)I+
F(s)/s(sl)2
0.5(7e2t12e3t
D/s
B/(s
Vi
^
5.24.
A"
5.25.
eAt
A;
^
~^
1
^\
j
0/
\
+ 5e4t)A+ 0.5(e2t2e3t+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)
[ll2e(tto)/2_e(tt)][7(i_y
5.28.
y(k)
5.29.
*(t,to)
[7
=
\
5.30.
 Z{V)^  4(l/2)'']/12
,
gtto
y
tj
eAt
5.31.
5.33.
Requires only
Vandermonde matrix
*(fc,m)
),
of
e'^txjrl
ai^ir) a^iit)
e'^tt^sf
Oi2() 021(7)
(^
V
and
[a22(t)
te'^tx^sf
which
is
 aii(t)]
Sj,
etc.
Then
a2i(t) [022(7)
 aii(r)].
CHAP,
5]
aHT)dT
5.34.
where
*(to,
5.37.
/cosh T
*(t,to)
e^i
cos
*o
a^T)dT
sinh
\sinh T
cos
aHr) dr
t'*
cosh t
cos
so
t,
*(t,t^)
R =
and
V(t,to)
0.
Also t(t,
!(,
to)
t).
Let
y2
j8
+ a,
S2
*(2,1)
 a.
/3
Then
\S sm
cos S
8
z{t
X^
(!,
*=^^
2 cos y cos 8
X2
det *(1, 0)
1)
where
0)
(y/8
^'^^^^A
\~y s T
eigenvalues \ of
x(t),
1)
(1.0)
X2 cos
det *(2,
2 cos 9
(2,
+ 4,r/e) =
 e2R) =
det (X/
= (2, 0) =
e^R
^''"'^^'^
'"'.'
The
a'Wdr
a:~l(t) sin
5.36.
am
aito)
'
*(t,<o)
127
e^R
X tr e2R
''"s
e*^^.
det e^R
by
2 cos y cos S
(y/S
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, JanuaryJune 1928,
pp. 1838.
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)+
^
(5e32)/3y^
\{\e^)l^
'
\^22e3y
 (eRT_)i
is
V
5.43.
y{t)
5.44.
u(t)
5.45.
5.46.
Let X
(1/4
{\2)dh/dt
^/W^n. )(t 
(sini
+ e^t)h =
(a"^ a
sini
to)/2
= ^^gs/V^ MtlO)
(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{tT)dt
chapter 6
Controllability
6.1
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:
previous time
function u(<,xo).
If all states xi are controllable, the
controllable.
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.
To determine complete
X(<i)
controllability at time
to
+ f
*(ii,*o)x(o)
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
The concept of
Definition 6.2:
x(<o) to
any
x (ii) =
to
any
x(fi).
some given
of a system
over a
is
finite
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
zeroinput 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 zeroinput
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
page
3.
Exami:le
6.1.
d^/a;i\
dt
From
the flow
diagram of Fig.
\x2j
61 it
_
~
gat
/
\0
+
13
u(t)
get
uo
Fig. 61
Fop zero
initial conditions,
u(t)
6.2
t^,
cLt
X2(ti)e~eh
so that Xzih)
is fixed
after
a;i(ti) is
chosen.
There
For timeinvariant 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:
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
130
[CHAP.
pGWc^ 0J
Af'.
0j
0J
+1
^__/rvJ
Fig. 62
Now we
assume
all /i
It is true,
to
if
the
/i
are nonzero,
1,2,
many
.,n
{6.1)
z(i).
Now we
u{t)
where the
ix^.
Prescribe u{t) as
i: M^e^^^'i'
{6.2)
{6.2) into
equation
{6.1) gives
is
A)
/ii"fc(e^'"'~'^ e^'"'""')
foralH
(6.5)
defined as
{B{T),<f>{r))
r'e*(r)</,(T)dr
/to
Equation
{6.3)
gii
912
Z2{tl)/f2
fl'21
6^22
{6.A)
Zn{tl)/fn
ffln
Qzn
by assumption, division by fi is
/i ^
linearly independent, the Gram
obviously
are
permitted. Since the time functions
can always solve equation {6.Jf)
we
Hence
Problem
3.14).
matrix {go,} is nonsingular (see
work.
always
will
control
{6.2)
for jui, 112, ..., i^n, which means the
where
gik
(e'^""'^', e^ii^>).
distinct eigenvalues.
CHAP.
6]
Theorem
6.2:
(2)
131
u{t) to
A is arbitrary.
Theorem
J,
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. 63.
The control u is connected to Zi, 22,
.,Zi\ and zi only if /i is nonzero.
It does not matter
if fufi,
and /ii are zero or not, so that the controllable system requires condition (2)
to hold.
Furthermore suppose condition (1) did not hold. Then the bottom rows of two
different Jordan blocks with the same eigenvalues [L^i(Ai) and Li(Ai)] could be written as
I
+T
^^^
dzv/dt
XiZv
dZr,ldt
XiZr,
+ fr,U
fvU
+T
+T
<^
<Hv
<7>
Fig. 63
fvZn)ldt
^
Therej^ore
fr,z, (t)
 f^Zr, (t) =
(1)
must
// part:
To
+ fvU)
Xi\Jy)Zv
fy (Ai2;
+ fr^U)
JvZn)
is
if
hold.
the sy;3tem
Example
is
{XiZv
not controllable, so
[fr,z, (0)
fr,
Then
and
(2)
of
Theorem
(6.2) to
show
6.2 hold.
6.2.
illustrate
why
condition
(1)
of
Theorem
6.2 is
d_
dt
X2
2/
(1
3)
X2
= (3^{m} + a;io)/(s  2) and j:{x2} = Uiu} + X2o)/is  2) so that y(t) = {x^a ~ Sx2o)e^t rethe action of the control.
The input is physically connected to the state and the state is
physicaJy connected to the output, but the output cannot be controlled.
Then
..zix^}
gardlesiii of
holds.
132
Theorem
6.3 :
[CHAP.
where
Proof:
Only if part: This is analogous to the only if part of Theorem 6.2, in that the
flow diagram shows the control is disconnected from at least one element of the state vector
if condition (2) does not hold, and a particular state vector with an element equal to
/tj^i'
^^t,
is
If part:
assume
uncontrollable
if
condition
(1)
Zi{m)
A.
[A'">m(0)
+ \'"^u{l) +
Jz(m)
+ iu{m), and
for simplicity
initial condition,
+ u{m  1)]/;
zi{n)lfi
Z2{n)lf2
+ 1) =
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 discretetime 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 discretetime 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. 64, 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.
Fig. 64
6.3
6.1,
CHAP.
()
Theorem
133
6.4:
From
we need
Proof:
tion 6.2
diagram for
system
this
given in Fig.
is
6.5.
*'V
=l
for giZi{U).
When written in matrix form this
set of equations gives a Vandermonde matrix
which
is
 0,
The extension
Theorem
to a general
1/
if
(2)
is
ctx
is
Theorem
if:
each eigenvalue
similar to that of
Ax{m)+Bu{m),
system
is
Theorem
is
Now we can
6.2.
Theorem
where T^AT
is
J,
distinct
/(m)
from
nonzero.
6.2.
6.6:
The proof
is
y{t)
+ d'^u, where A
and only
(1)
Theorem
matrix
6.5:
The proof
Fig. 65
ctx(m)
+ d'^u(m), where A
completely observable
if
and only
if
is
arbitrary.
conditions
(1)
x(m + l) =
Then the
and
(2)
of
6.5 hold.
similar to that of
Theorem
6.3.
= Ax + Bu,
= Cx
I
Du
Unobservable and uncontrollable elements of a state vector cancel out of the transfer
function of a single inputsingle output system.
134
Theorem
6.7:
For the
[CHAP.
c'''(sl
A)""*b
and only if
statement holds for discretetime 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)
4
<^{zi}
(s
so that
why
dz/dt
Now
Figs. 62
(2) of Theorems
and 65 to represent the element
consider condition
6.2
Zi
and
6.5.
of the bottom
Fig. 66.
Fig. 66
if
the system
is
is
p.,
f.
and/or
A,.
If and only
which occurs when
the residue of
g^
0,
Note that it is the uncontrollable and/or unobservable element of the state vector that
canceled from the transfer function.
6.4
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
controllable
way
to determine if the
system
is
completely
CHAI.
6]
Theorem
6.8:
Q =
Note that
(BAB
...
and so
and
totally controllable if
AiB)
is
135
6.2.
=
det (QQ^)  0, since rank
exist faster machine methods for
if
To reach an arbitrary
Proof:
x(^i)
we must
initial state,
find a control
such that
u(t)
x(*i)
r'eA<'i^'Bu(T)dr
{6.6)
Use
of
Theorem
gActir)
4.16 gives
^ y.(T)A"~'
i=l
x(i)
+ ABw2+
(Bwi
+A">Bw)
Wl
;
\ Wn
where
Wj,
y+i_s.(T)u(T)dT.
Hence
x(ii) lies in
must
to
have rank n
trollable,
Kbw we
time
we
assume
Therefore
if
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 mvectors 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
eActiT)B8/
eA(tit)B^^
_ f\
^^
eACtit)AB/i2
eA(tito)B
eA"i'o>A"iB^
B"'\t^)g{i)di
{6.8)
8^"^ is
d^g/dt^
and the
can be rewritten
eActoti)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).
some insight as
to
why
completely control
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).
Theorem
6.9:
proven
is
is
totally ob
CA
CA"
Theorems 6.8 and 6.9 are also true (replacing totally by completely) for the discretetime
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 continuoustime case, we give a proof for
system
6.9
from knowledge
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
+ nl) =
Cx(i
+ 1) = CAx(0
CA"ix(i)
Px(i)
y{l
6.5
exists
if
+ nl)
and only
if
P has rank
shown
n, as
in
Problem
3.11,
page
63.
In timevarying systems, the difference between totally and completely controllable becomes important.
Example
6.3.
<)
dx/dt
X
4 h{i)u
and
j/
a;,
where
6(f)
is either
zero
rO
e
Fig. 67
If
{(*)
interval
[to, to
+ At).
If
for
to
CHAP.
6]
137
Now suppose b{t) = for t^ < t ^, and we wish to reach a final state x(t^. The state xit^i can
be reached by controlling the system such that the state at time t^ is x{t^ = e(*2~*i'a;(t2). Then with zero
input the free system will coast to the desired state x(t<^ = <j>(t2. ti) (ti). Therefore if the system is totally
controllable for any time interval f,. ~ * ~ *e + ^*> it is completely controllable for all t t^.
For the timevarying system, a
Theorem
Theorem
criterion analogous to
6.10:
A(i)Qfc + dQfc/di
for k
1,2,
.,nl.
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
Construct
u(t) as
iy
such that
to
< < *i
1?
uW =
Sm^S''^V^)
(6.9)
lc=l
To
from
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
/*(ti,v)Q('?)
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 wvector z 7^
such that, for times t everywhere
dense in
U~t~ U,
zt*(to,QB(i)
By
Since
zt*(io,<) v^O such that
z+*(io,
t) Q)c(i)
0.
is
*(io, t)
(ytQi y+Q2
row vectors
is n(e.d.).
of Q.
always
y+Qn)
Since the
y^Q.
n row
2/iqi
2/2q2
yqn
138
Theorem
6.11:
[CHAP.
ation
is
Pi
C{1),
Pa
C{l
l)Ail),
...,
Pfc
C{l
+ kl)A{l + k2)
Ail)
The situation changes somewhat when only complete controllability is required. Since
any system that is totally controllable must be completely controllable, if rank Q(i) has
rank n for some t > to [not rank (e.d)] then x(io) can be transferred to x(^i) for ti ^ t.
On the other hand, systems for which rank Q(i) < n for all t might be completely controllable (but not totally controllable).
Example
6.4.
\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
row
of Q(t)
is zero,
^ea(tt)
*(t,t)B(t)
so
Mid) +
rank Q(t)
df^ldt
\/fi(t)\
ee<tH^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
<
"">
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.
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
Solved Problems
6.1.
dx
1\
/O
X +
143/
m,
2/
(1
1
l)x
\1
and un
observable states.
Following the standard procedure for transforming a matrix to Jordan form gives
1
21
0\
1
10
.14
Then
shown
= Tb=
(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. 68.
G>
s2
<7>
Z2
o
KiW
s2
<7>
^3
s
OV
0
Fig. 68
The element
Z2 is
Note Theorem
6.1 is inapplicable.
140
6.2.
[CHAP.
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 =
(s2)3
h(s)
(s
x^
(s
3^3
(6.16)
(6.13) gives
 1)2m.
2)i,
Is
6.1.
!)* (s 2)2*3.
(sl)(s2)^
it
6)^2 =
(6.16),
(sl)(s2)2
as defined in Problem
is 2
(4s
Then from
(s2)2j"'
'
(6.15)
**
^3
(i^i)!_o+^l
6.3.
2x2
and con
6.1.
totally observable
and
totally controllable?
'11
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.
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.
Y^2
vsrhere
y{ri, e,t)
modes of vibration
The
at r
of the
V'2/
u,
60
6, t)
for a circular
drumhead being
are
27rrS{rro)S{9eo)u{t)
Can
{6.17)
drum?
mode shapes
00
y(r, e,t)
is
00
2 2
COS
2njfl
a;2 +
1,
(t)
sin 2nve]
CHAP.
6]
e7(r).
141
(Px2i/dt^
Xa;2i
y cos 2jr9oW
cPxgi/dt^
X^si
y sin
first
2ire(,u
where
Pig. 69
6.6.
dx
110
=
[0
0'
0Jx + [1)mi +
hu2.
where
Ui
is totally
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
l'
0,
Q=(l01010
Vl
1
AB
ll
For case
(B
/O
(0101
\0
For b
(1)
A =
uncontrollable.
controls.
if
identical
is
uncontrollable.
is
controllable.
142
6.7.
0'
VI
o;^ + 12'^
dx
dt
The
Q(t)
6.8.
e(e
+ 2t 2).
Since
eHlt)
+ 2i =
Show
Note that
Q(t)
2(e.d.).
controllable if
and every
t>
t,
where
*(r,,)B(,7)Bt(^)*t(,,,)d,
9{t, t)
6.11.
rank
is totally
W(t,r)
Is
'e
Then detQ(0
[CHAP.
is
rj
W~i
u(r)
Then choose
exists.
Bt(T)*t(ti,r)W>(o.*l)x(ii)
(k,Wk)
=
=
(z,
kt*(T,
7;)
8(7?)
Bt()*t(r,,)kd
d,;
Suppose
is to show
is nonsingular if the system is totally controllable.
such that
Then there exists a constant nvector z
singular, to obtain a contradiction.
Wz) = 0. Define a continuous, mvector function of time f (t) = Bt(t) *t(to, t)z. But
is
is
'if(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
any nonzero
definite.
z.
CHAP.
6]
143
Supplementary Problems
6.9.
u(t)
\,i(l)
[^
Fig. 610
6.10.
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
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. 611, 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. 611
Fig. 612
6.13.
Given the electronic circuit of Fig. 612 where u(t) can be any voltage (function of time),
what conditions on R, L^ and Z2 can both i^its) and ^(ii) be arbitrarily prescribed for
given that ix(to) and t2(*o) can be any numbers?
6.14.
'$\
Under what
conditions
is
Under
144
6.15.
[CHAP.
(6.2)
z(ti).
6.16.
Prove Theorem
6.5.
6.17.
Prove Theorem
6.6.
6.18.
What
6.19.
are the conditions similar to Theorem 6.2 for which a twoinput 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.
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.
where x
(a)
Is
is
criterion
a 2vector, m
(b
is
Ab
x(fc
A =
a scalar,
1)
Ax(fc)
j,
bM(fc)
If
(6)
6.1.
iij
is
x(0)
'
^^^ *^
sequence m(0), m(1) required to drive the state to the origin in two sample periods
6.23.
J.;
X(,K r l)
x(A:
where x
(a)
is
Is the
a 2vector, y
is
a scalar,
i\x.\n,),
Ax(A;),
y\n,j
y(k)
ctx(fc)
ct
^\
1^
/9
2
(1
j/(l)
(6.7).
(fe)
x(2)
0).
2).
system observable?
(i.e.,
control
y(2)
8,
14,
find the
6.24.
6.25.
6.26.
Prove Theorem
6.8,
Hint:
See Problem
6.8.
Du,
Given the multiple inputmultiple output timeinvariant system dx/dt = Ax f Bu, y = Cx +
the Q
where y is a fcvector and u is an mvector. Find a criterion matrix somewhat similar to
matrix of Theorem 6.8 that assures complete output controllability.
Consider three linear timeinvariant systems of the form
dx</dt
Sj.
(a)
A)x("
B>u>,
y"'
C'x(
1, 2,
613 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. 613
CHAP.
6]
(6)
is
observable,
145
show that S3
is
observable.
6.27.
i
Is this
"""""'^
O''G)
(J
6.28.
Prove Theorem
6.29.
Prove a controllability theorem similar to Theorem 6.10 for the discrete timevarying case.
6.30.
continuoustime case.
Similar to Problem 6.8, show that the timevarying 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.
finite
>
17
t.
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
c(t)
(,
g dt
Supplementary Problems
iit^)
dxjdt
all
(to)
0,
6.11.
One
6.13.
Li
6.14.
Mg #
6.15.
Many
is
and those belonging to 1 are uncon3(s43)i, showing only the states belonging to 3 are both
is
La
and ^
7^
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=
iik^~^^*.
to
have an inverse.
/11/22 "~/i2/2i
0,
146
6.19.
x(i+l)
*{n,k)
w(0)\
x(n)
_'^
m(w);
lA
(l)fc(^_^
y(n)
(_2)nic
(1
0) x(w)
^2
^J
27 V 0^2(0)/
shown
is
where
in Fig. 614
Zi
and
2 ^^^ controllable,
Zj
and
Zg are observable.
+
2
+'
0j
<EH
0J
^
+
I
Fig. 614
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)(IsA(3))iB")[C(i'(IsA(i>)iBa>
D)
C(2>(IsA(2>)iBC2)]
chapter 7
Canonical Forms
of the State Equation
7.1
7.2
Section 2.4 shovs^ed hov^^ equation (2.21) in Jordan form can be found from the transfer
function of a timeinvariant system.
For singleinput systems, to go directly from the
form dx/d = Ax + bM, let x = Tz so that dz/dt = Jz + T^hu where TAT = J. The
matrix T
page 97.
For
is
arbitrary to
vi^ithin
n constants
so that
T = ToK
= 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
If the
system
is controllable, it is
Then using T = ToK, we require T<r'b= Kei = iai mi ... 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 inputoutput considerations.
Also, the n^ elements in the
matrix are transformed to the n eigenvalues that
characterize the time behavior of the system.
7.3
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
^/Rez + ilmz\
dt\B,ez
_
"
 jimz)
[CHAP.
of", this is
\/'Rez + jimz
ReL ;ImLy\^Rez  yimz
/ReL + jImL
\
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 or observable
We
dx/dt
A^()x
+ B^()u,
{T.l)
e(i)x
because
we
shall
In this chapter
n1
n1
and
it
will
Before proceeding,
Theorem
7.1:
theorems.
A;
Now Q^(t)
for fc = 2,3, ...,.
rankQ^() = rankQ^() for all t.
Ti()Q(<)
and since
T(*)
is
t,
CHAP.
7]
reassuring to
It is
by a change of
theorem
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
i ^ n
of order
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
Q^'
(i)
rii
with
Qn'i)
(7.2).
for
i.e.
all
zero rows.
Vi(S R)
is
Proof:
= (QM Q2M
uncontrollable in
Using induction,
QJ
= (^2\
?)  (f
/'^'^
and
if
QJ
obtained.
T(i) x{t)
('<")
to the
tlien for
form
of {7
1, 2,
Therefore
VOO
where
{t) is the
...0J~0
ni,ni
+ l,
...,nl.
Q^{t)
T^{t)q^{t)
(TiQ^i TiF)
(Ti(),T2(i))(^"
Ti(Q^i F)
is
an ni x mni
150
[CHAP. 7
Qn*)
{v^{t) y^it))
(^^f
^f)
where Yzit) is any set oi nni 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
and since
S{t)
equivalence transformation.
relationship for observability
The dual
x(<)
is
T^t)
z{t)
that transforms
(7.1) into
= AUt)z3
dz3/di
is
of order
Ws^m
(7.6)
C?()z3
Then
F'{t)
Theorem
The system
7.4:
and only
nsXn
if P^^
differentiable
The proof
set of
nns
is
(7.1) is
= (yJRa where R3 is an
is
a fcnsXns
n3(e.d.).
similar to that of
differentiable
if
Theorem
7.3.
Here T"
= f^']
is
'
where R4
is
any
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 wsvector for
1, 2, 3,
{7.1) is
(Cr C^
transformed into
0)w
C^^)
CHAP.
7]
d_/w,\
[ Jj
!(::)
^')
if
vwi\ ^
/Ar,
rank
of
of
/Br\
rank
the subsystem
< ,(::)
aI)(::)
^^^^^
151
m + n^ ^ n(e.d.).
,...,
Theorem
7.5:
HereRX^) is a A;n x rii matrix with rankni(e.d.), S\t) is an ^ x mn matrix with rankni(e.d.),
an ni X differentiable matrix with rank ni everywhere, Vi(i) isannxni
differentiable
matrix with rank w; everywhere, and Vi{t) V^i) = SIi. Furthermore,
the rank of R* and S'
must be such that the controllability and observability matrices of (7.8) have
the correct
Vi{t) IS
rank.
Proof:
Q^
^ ^ ^ ^
p^
{7.7).
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^
and
P^
which
is
(Pfi
PJi
Gl GJO^Ux +
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
= WS^
1.
Factor P^Q^
2.
Solve P^Vi
3.
to obtain
Ri and UiQ^
I,.
W
S>
and
S^.
152
4.
Factor P^R>Ui
5.
6.
Find V4 as any
7.
8.
Form
T{t)
columns making
set of nt differentiable
[CHAP.
S.
T{t) nonsingular.
8Ii.
(Vi V2 V3 V4).
make
7.1.
1\
/sint
Given P^
sin
Obviously rankP^
).
and
it
is
T(t)
we
Example
page
3.13,
Then
44.
fsint
[sint
 1, 0, 1,
However,
is
l\/(sint)i (sint)2\
(sin*)i/
l)\
/I
0\
\^i
0/
iir
for
if a{t)
and
Pit)
[m
^(*)
common
zeros, let
a{t))
Then
p{t))E{t)
{a{t)
(^^(f)
+ I3^{t)
0)
and
E(i) is
where
p^ is
Example
Let
Ci, ^2,
.,ik,
f2
and
p(t)
tK
common
Their
common
zero
Then
is
is
is
?i
a convergence factor.
with order
,^
Note E(0)
.,
(;S"!i:i)n(i""^A>
7.2.
a(t)
Choose
pj.
y^iX^)
_

>.2
i^.
.
nonsingular.
or column operation, it is
Using repeated applications of this form of elementary row
E.(t) such that P^T = (P^^ 0), and^simUarly
obvious tlat we can Ind T(t) = E(t)E.(t)
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
As
CHAP.
7]
Consider
first
dx/dt
to a
153
form analogous
Ait)x
+ h(t)u
(7.9)
to (2.6),
la,(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:
in that
rank
Q=
Proof:
This
Q^(i)K, where
is
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
is
\(l)l
form
w = kz,
where
1
K
will give the desired
AX + dqf/(i =
0/
(7.10).
Now we
is
also
consider the transformation to the second canonical form of Section 2.3, which
as phasevariable 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)"^
Qnl.nl
and
Q^
where
qu
(l)'an for
Q'n2,2
Qn1.2
Qn2,
QnIA
l^i^n;
and for
q,^
Theorem
7.7:
(Ifani+k
An
'^
+ {If
l^k<i n,
i>'i
q+i
anjQiKHi
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
T=
Iq^Wn+i)
(qlq
QQ^',
(qi
q^ Q^Ae)
solved recursively
Therefore q^+i = Q"A'^e, which gives a set of relations that can be
for the ai{t) in terms of the ;(*) of {7.11).
Example
7.3.
Then
eta
/922
Ini
ai
al
da2/dt
and
Q==^A'e
'0
1
02
By
it is
"2
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
Zi
+ a2dy/dt+
approach is to set
For the case where u is a vector instead of a scalar in (7.9) a possible
rank
n everywhere
the resulting Q has
all elements in u except one equal to zero, and if
or desirable,
possible
not
then the methods developed previously are applicable. If this is
the form
dw
^A^i
Ar2
aV
A2"i
A^2
All
bw2
I
It
,An
may be
obtained,
Al
...
where the
ft
ATil
\0
...
h\
time, and
are in general nonzero nvector functions of
{7.13)
CHAP.
is
7]
of the
form
and for
{7.11);
155
i ^ j,
m
m
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 timevarying 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
/2
z(m
1)
0\
j
Vo
2/
/O
z(m)
2
1
3/
1/ \
V1
Substitution of
are obtained.
u{m)
2
form
is
then
156
7.2.
[CHAP.
Transform
/
dxldt
1
1
0
0'
2\
2 X +
\i.
,1/
From Problem
4.8,
page
and Problem
93,
4.41,
page
97,
dildt
,(1
Then
2
and p
1j.
Re2
(i)
)"
/I
0\
X +
01/
to a controllable system,
reduce
7.4,
dx/dt
+ j)/p*J
Im 22/
7.3.
\u
(li)//3
i^
(1
1 2)x
\2.
an observable system,
to
(ii)
2 \u,
(iii)
to a controllable
(i)
Form
Observe that
Q=
(b
Ab
and not
timeinvariant case,
elementary row operations to
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
of
is
Tz where
and
gives
CHAP.
7]
Using
this
2 1
/l\
0\
dx/dt
z
Jm
2/
(ii)
Forming the
/
(
2\
1
5/
(1 1 2)z
\0/
/
=
0\
1
)[
0/\0
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/
P^
157
OJM
i/
= (12
0)z
\2/
1/
1 1
1
P^Q*
Then
/lt>3
P^V,
1
021
)(
1
R'
gives
Vi
...
v^i
^31
and
UiQ^
(1 1 1)
SI
gives
Uj
(1 u^g
1
M13)
02/
\2
Note UiVi = 1 for any values of M13 and Vg^. The usual procedure would be to pick a value
for each, but for instructional purposes we retain M13 and v^i arbitrary.
Then
P^
RlUi
/O
(0
\0
 Mi3
+ Mi3
 Mi3
2
1
+

uia"
Mi3
Mi3
(0
1)
R2U2
Mi3^
and
/
Qx_ViSi
Choosing Vgi
2y3i V31
2'i;gi
vai
\2 + i;3i
i;3i
and M13
2i;3i\
2Vsi
2 + ^31^
\0
Using UjVj
Sy
^12
l\
^22
V32
1/
we can determine
/l
)(2ygi yg, 2
gives
/l
1'
1
I
Vo
01.
1
1
\M31
M32
M33/
 Vgi)
VgSa
158
is
[CHAP.
then
2/
0)z
(1
V3y
is controllable and observable, zi 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.
to controllableobservable form.
X +
dxjdt
2/
(0
l)x
i2
7.5,
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
(0
0)
gives
/
U2 =
S3
3(2
'
1)
(0
R2
(t
Then Vj
(0
R^Si,
Ri
Solving P^Vi
2f2
1)
and vj
3(3
\3(4
(0
form
(7.7).
1 0)
and
T = (ViiV2V3V4).
2(3
2(2
t
 12t 
24(2
i5t
It is interesting to
18/
note
CHAP.
7.5.
7]
to controllable form:
sint
j
First
we
calculate
E(f)Qa:
tcosA
_
~
coa^tj
j\(iost
/t2
to obtain
t^cost
cos2t
cos^t
is
A/
cos
\ cost
Ei(e)
t^
7.6.
\coatJ
cos t
/'
159
C0s2
cost
V cos t
=223
01/
/
x(m +
l)
0\
/1\
x(m)
\2
(i)
form
{7.10)
and
(ii)
U(m)
\2/
form
(7.12).
Note this is of the same form as the system of Problem 7.3, except that this is a discretetime
system.
Since the controllability matrix has the same form in the timeinvariant case, the procedures developed there can be used directly.
(See Problem 7.19 for timevarying discretetime
systems.)
From
part
(i)
of Problem
7.3,
/
z(m +1)
\
Therefore the best
we can
do
is
z(m)
+0
2/
u{m)
\oJ
_i)^iW + (o)W
(_2
\0
to the first canonical
zi(mll)
/A
o\
21
is
2j\l
1^2
0/
form
Xcim
+1)
z<.(m)
( 
) m()
'0
l\/0
2J\1
i\i
gu
160
 a^ =
where q^
Then
^ =
from Example
a^,
7.7.
A^Cn
 dq^'M) we
Q^ii(A^iq^i
obtain
Zp(m
By
From
7.3.
[CHAP.
+ 1)
form
^jzp(TO)
JM(m)
(i)
sini\
ft
h
dx/d
2\
+ (1 )m
)x
_TL
(ii)
i^O
for
and
(iii)
find a scalar
first
canonical form,
<^
{:,
''T%:
I)
CT'
\
from which
where
'ai(t)\
^a^it)/
2t
sint ye
6t
4i
2(2
^^
3
[1
a,(t)j
J\l
ai
1
from which
d.,/dt
{_^^\^j^^
where y
7.8.
X)^
+ (l)
to this is
ai dy/dt
{a^
+ dai/dt)y =
z^i.
Transform
to canonical
dx/dt
we
0
4
calculate the
l\
/O
1\
OxfO lu
3/
\l
1/
CHAP.
7]
'0
Q^
'1
(000
(Ml only)
4^
.13
Q^(M2only)
1
I
8.
161
2
1
4^
1
02,
two columns of
first
Q^
so that
dvildt
Also,
Q* (m2
and the
first
two columns of
only).
To transform
to a
form analogous
w = Kz
where
Then
0'
dzldt
1
4
0)m
llz + (0
0/
Vl
0>
Supplementary Problems
'1
7.9.
7.10.
7.11.
Transform
Transform
dx/dt
rfx/dt
/~^
7.12.
Prove Theorem
7.2,
page
149.
7.13.
Prove Theorem
7.4,
page
150.
Show
and
1\
OJx +
3/
/I
(
o)m
0.
i)''"*'(4)"
7.14.
= (o 1
.14
7.4,
(ii)
to
to
Jordan form.
realJordan form.
reduce
an observable system,
(iii)
to
that the factorization requirement on Theorem 7.3 can be dro^^ed if T(t) can be nonsingular
differentiable for times t everywhere dense in [to. *i]'
162
7.15.
Consider
system
the
"'^
7.16.
^
\t2Wj
\^2/
for
where
and
==
/i(t)
cos
Can
1  cos t elsewhere.
for
this
and zero
dx/dt
7.17.
*^i
/
[CHAP.
^4
8^
=12
6'"
^8
2/
(1
l)x
7.5,
{7.2),
7.18.
7.19.
Develop Theorem
7.3 for
{7.11)
systems.
t1
7.20.
dx/dt
to
7.21.
{7.2).
/I'
+ 2\
+ 2Jx + (l)M
t+1/
\0^
i
t2
gj^ in
= Ax + eM where
z = Tx where z
Show
that
terms of the
is
0^,
system.
7.23.
Example
7.3,
aj in
terms of the
for
a;
1, 2,
manner analogous
to
page 154.
11
7.24.
3 (thirdorder system) in a
~
1/2
(1/2
27/2
0]x(
7.25.
ds.ldt
7.26.
Using the
/
dx/dt
dw/di
/
into the
form
V
7.27.
1
2\
3
6\
1 1
= Tw
/l
3\^
/l
0\
\0
1/
4J
0/
systems.
CHAP.
7.28.
7]
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
()/
Answers
Supplementary Problems
to
11
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
0)
a3
7.24.
Ti
Q!i,
is
more
3/2
5/2
2
11
r43
r4i
(7.6),
for any
r^i
that
make T nonsingular.
Also, Jordan
difficult algebraically.
aj
03
1,
1
7.25.
Ti
A
ot
'1
7.26.
1
This form
(>t
2e~t
et
1,
3^
1
.0
7.28.
fi2t
is
2,
dw/dt
At(t)w
Ct(t)u
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
INTRODUCTION
Classical techniques such as block diagrams, root locus, error constants, etc., have been
used for
many
8.2
state equations.
(summer,
scalor,
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 2.12.4, pages 1617, hold with
u(i) and one
the following exceptions. A summer has n mvector inputs ui{t), U2(t),
time.
u(i).
output mvector y(i) = ui{t) U2{t)
and one output A;vector y{t) A(t)u{t), where A(i)
u(t)
A
is
. ,
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.
=
dt ^X2/
\0
Oy\a;2/
\0
(3
2)x
81.
*V(t)
Fig. 81
We
drawn
164
in a similar
CHAP.
8]
Example
165
8.2.
Example
xo
shown
8.1 is
in Fig. 82.
^W
Jliy}
Fig. 82
(6/s
(3
2)
(2s
+ 3)/s2)
['
lo
0/
Vo
^{u>:
^^{}
H(s)
Fig. 83
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 inputmultiple output devices.
8^
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
*^_B(,(i,
y(*)]
i)
I.
Further,
to infinity before
we
j,
shall
so that
y(t)
Zero output
is
asymptotically stable
A(t)
Fig. 84.
166
Theorem
8.1:
where
dvfldt
such that
on w.
lim
lim
84,
e(f)
[CHAP.
and only
if
if
C(t)w + g
t* 00
t00
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
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
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.
lime
Then
t~*
djs./dt
 vi)ldt =
lim(dy)
[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)*ABc(, 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 timeinvariant 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}.
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
C(^)
r
'o
*^(i, r) B(r) 6, dr
f or
1, 2,
CHAP.
8]
167
*/
'
*^(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.
then
is
2)x
(3
limfi'(t)
which
=\(q
^
e(t)
+ 2a!2(0)] cosV2t +
et[3a;i(0)
is
no input
Ui(t).
The zero
input,
^^s output
o)~(i)^^^M''
0,
and there
U2(t)
iz[i(0)
+ 2(0)] sinV2tl
Consequently Theorem 8.1 applies to the unity feedback system, so that the
where
8.1 in
d(t)
t+oo
d{t)
For
g{t)
0,
we
3wi(0)
2w2(0)
3fW2(0)
+ f
[S(tT)
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.
C(*)
aC*.
t)
B(t) dr
[Sit
 t) +
2] cJt
1.5*2
2t
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.
84
system
is
8.1
cannot
called
is
u^it),
The system of
lim
di
. ,
constant error and will probably follow with zero error any function
by taking g(t) = t~^.
y = 0.50:2(0)
be used.
v^ith
Use of Theorem
is
superposition integral. For classical scalar typeZ 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 timeinvariant systems with
the plant transfer function matrix H(s) = C(sl A)"'B.
Theorem
8.2:
and
lims'ip(s)[
s0
<
>.
168
assume H(s)
neighborhood of
+ P(s) where
s=C{gi}
lim sR>(s)
where
+ I]i =
(Zl)!s'[H(s)
*"*"
Choose
G(s)
s'R(s)
0.
''"
'*
Res^O.
[CHAP.
(?l)![R(s)
+ s'P(s) + s'I]>
Since [H(s) + I]~> is the asymptotically stable closed loop transfer function matrix, it is
analytic for Re s ^ 0. Then sG(s) has at most a pole of order J  1 at s =
in the region
Res ^ 0. In some neighborhood of s =
where R"Hs) exists we can expand
\s\,
analytic in
Re s
which
satisfies
+ P(s)
s'[H(s)
P(s)]
s
'R(s)
where
is still
Theorem
8.1.
where sG(s)
is
analytic in
is
(Zl)!s'I
[s'R(s)
lim
s'
Re s
'P(s)
^ 0.
<
Write H(s)
and
00
R(s)
=
=
s0
+ P(s) + I]G(s)
where
(Z
W(s)
sG(s)(I
+ W(s))i = sG(s)[IW(s)+W2(s)
+ s'iI]sG(s).
[s'ip(s)
classification of
application of
is
Example
8.4.
1
Q
^{d};
1
:>^{!/}
12s
3s2
Fig. 85
in Fig. 85
1\
"^^^
i2
0/
V
/
"^
Vl2si
S2R(s)
in which
lim sP{s)
s0
<
\
12
I
3s
'^
P(8)
in the
form
CHAP.
8]
169
and where
limsRi(s)
wo
limsf
\l
 \6s^J
so
']
\0 6/
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
S1
<
s>\
/
12
/0l\
^ \3 0/
,.(,)
p(,)
but now
s>0
lim sRi(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 typeZ system, we use block diagram manipulations on Fig. 84 to get ^{e} = [I + H(s)].<:(d}. If it exists, then
To
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'+iRHsK{d}
for any I > 0. Then an error constant matrix table can be formed for timeinvariant systems of Fig. 84.
Steady State Error Constant Matrices
System Type
Step Input
Ramp
Input
Parabolic Input
lim Ri(s)
lim [HH(s)]i
S+0
1
siO
lim
In the table
Example
all
such inputs.
8.5.
/a
of
Example
8.4
lim Ri(s)
the input were (t  to)U(t  to)e2, the steady state output would be [{ttoWit
can follow with zero steady state error an input of the form (t tQ)U(t to)ei.
8.4
RMs)
s>0
 tg) +
0\
6/'
6] 63.
"^^"^
^*
The system
ROOT LOCUS
Because root locus is useful mainly for timeinvariant systems, we shall consider only
timeinvariant 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
= CxfDu
(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 +
...
techniques.
variation,
^nlS
from both
loop gain
8.6.
k.
xo
Fig. 86
The
_
~
dx
It
The
characteristic equation is
root locus shown in Fig. 87.
s^
t
(3
I
k)s
K
1
K 3
+ 4k =
0.
4
Putting
it
form leads
Ims
Res
Fig. 87
Example 8.7.
The feedback system of Fig. 88 has an unknown
parameter a. Find the effect of variations in a upon
+ 3} sinh a
+ 3s + sinh a
(a
s2
Fig. 88
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
^Ui{e}
Fig. 89.
G(s)
J
+
+ o.
Civ}
H(s)
= Ax
I
be
Fig. 89
c+x
I
de
(8.2)
Then
Example
8.8.
The
810.
state space
form of
this
is,
in phasevariable canonical
e to v,
(hi.
\)^ + (i)^
dt
Then
ct(7<oIA)ib
=
jaija
(^
0)x
1)
ImGH
^CW
ReGH
Fig. 810
Fig. 811
About the only advantage of this over standard techniques is that it is easily mechanized
for a computer program. For multipleloop or multipleinput systems, matrix block diagram
manipulations give such a computer routine even more flexibility.
172
Example
[CHAP. 8
8.9.
shown
G(8)
in Fig. 812.
^{v}
H(s)
= c(sIAbict)ib2^{e2}
^{vg}
so that
we can ask
c(iuIA~biC+)ib2.
8.6
Fig. 812
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 wthorder scalar linear differential equation.
For timeinvariant 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 timevarying
systems. To "place the poles," the totally controllable part of the state equation is transformed via x(i) = T{t)z{t) to phasevariable 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^ +
[,(*)
 ali)\z, +
This
\ 1
an{t)/
[^{t)
grives
[(*)
 a(t)]2
...
az
a^t)
\a^{t)
aj,t)
...
...
ccj^i)
=
and each z.^^{t) for
are to be chosen, the corresponding closed loop transition matrix *^(t, t^) can be shaped accordingly. Note, however,
that x(i) = T{t)^jt, t^T.^ so that shaping of the transition matrix *^(<, t^) must be done
keeping in mind the effect of T{t).
Then
i
obeys
z^{t)
1, 2,
w
z{^'
(Q2j"^"
alt)'z^
a^{t)z^
Since the
a}^t)
This minor complication disappears when dealing with timeinvariant systems. Then
and furthermore each a.{t) is constant. In this case the time behavior of
x(i) and z(i) is essentially the same, in that both A^ and A^ have the same characteristic
equation A" + ff^A"+
ha^Atofj = 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.
^x
^
1
dt
V2 3
)x
*^'
will
0, 1 and 2.
Then the desired system
we choose u = (2 1 (t + 3))x, so
CHAP.
8]
173
For multipleinput systems, the system is transformed to the form of equation {7.13),
except that the subsystem dw./dt = Al? W; + hfu^ must be in phasevariable canonical form
{7.12) and for i  j, the ATi{t) must be all zeros except for the bottom row.
Procedures
similar to those used in Chapter 7 can usually attain this form, although general conditions
are not presently available. If this form can be attained, each control is chosen as
u^ = k?'(i)Wi elA^w. for j  i to "place the poles" of Ajj(t) and to subtract off the coupling
terms.
Why bother to transform to canonical form when trial and error can determine k?
Example
8.11.
Place the poles of the system of Problem 7.8 at Pj, p^ *nd p^.
We
calculate
det
This
is
^^
(fei3
It
+ ^21 + k^i + ^23 + 5)X2 + [fcii + 2fci3 + 3^21 + 4fe22 + S/cga + fei3(fe2i + ^22) " ^izihi + ^12) +
k^i fci3  2^21 4^22  6^23  fcll(^22 + ^23) + ^12(^21 + ^23) + ^13(^21 "" ^22)  4
trial
and error
to choose the
X3
(pi
fc's
to
8]X
match
+ P2 + P3)X2 +
(piP2
+ P2P3 + PiPs)?^ 
P1P2P3
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 highorder 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
174
To reconstruct
We
all
observable.
is totally
the states at
For
all
times,
we assume
[CHAP. 8
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
A{t)x
n.
+ k(t)[cHt)xy]+B{t)u
is
constructed as
(8.3)
where k{t) is an wvector to be chosen. Then the observer system can be incorporated into
the flow diagram as shown in Fig. 813.
Fig. 813
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
timeinvariant 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
At  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.
2/
Construct an observer system such that the error decays with poles at
First
we transform
dw
dt
/2
(1
l)x
0,
2\
w +
.
/I'
2 and
3.
CHAP.
8]
form
dz
^2 4
It
'4
where
teristic
equation
k = (1
obtained by Theorem
*.
(X
+ 2)(\ + 3) =
X^
0)t
175
+
is
7.7.
We
]v
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
{H(t)
diflferentiable
G(t))
where
is
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
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 =
dP/dt
+ PA + PLC. Then
= P(xx) = Px w
is
(F PL)j

Fig. 814
Example
8.14.
C =
(1
1),
cj
choose
8.13, construct
P =
[l
(pi
Pg)
ij
with
pi = pg
Then
P1P2VI
(H G)
I
Pi
{8.6)
dP/dt
From
+ PA
(8.5)
and
176
[CHAP.
Therefore
PiP2\1/
P2
w +
P1P2
Pi
and
PL
=
=
{dr/dt
(Pi
+ PA)H
(dP/dt
+ PA)G
/2
P2
P2
Pi
1\/P2
Pi
3pt
2yvi
P2) /'2
(Pi
Pi
SO that (pi  P2)dw/dt = (3pi
choice of Pi/p2 with 1 < pj/pg
P2)
P1P2
4p2
P2
2p,^
P?
Pi
P2
bad
The question is, can we place the poles of F by proper selection of P in a manner similar
to that of the wdimensional observer? One method is to use trial and error, which is sometimes more rapid for loworder, timeinvariant 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
12
ct
dzi/dt
Vi
ct
n
is in
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)
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 timevarying case, and an algorithm
is
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(^)
PiL;
x = Tz = T(
Then
Example
find FjLj as
where
fci,nil(*)/
z.
H.Wi
+ Gi^/i =
)w.
]yi
8.15.
2
Then equations
(8.7)
8.13.
2
dt
= {dPJdt + PiAii)Gi.
'4
from which
kii{t)
...
Also,
kOA](
177
2/
(1
l)x
Tz
where
(Tt)i
are
dz
J:4>
df
The estimate
+ (:r'"
now
2/
system
is
= 2
dw/dt
= 2w + 2y 2u.
PT
ix
w =
2xi
l)z
Z2
Then
F = 2
2 \
/w
and
PL = 2
Therefore
Tz
and the error
(0
obtained as
1
where k^
+ X2 w
3\/l
1
4JU
2\/w
l)\y
of Fig. 815.
Fig. 815
178
[CHAP.
ALGEBRAIC SEPARATION
8.8
In this section
we
where
oi'
Specifically,
we
A(f)x
C(i)x
B()u
J(i)d
(8.9)
dw/dt
F(t)wF{t)L.{t)y
H(i)w
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
diagram
is
{8.11)
as in Fig. 816.
^^>^:r0t^
d=i4;
p0if(S'
*U Feedback 
Observer
Physical system
Fig. 816
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
PBWGCPLC 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]
179
= Fe
Use of Problem
8.16.
For the system of Example 8.13, construct a onedimensional 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
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)
u
Thus the
(2 3/2)x
is
(2 3/2)
7w/2
I
5y
as in Fig. 817.
^rO
+
CMS)
Fig. 817
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, timevarying systems.
why
8.9
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
wanted
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
An
=
=
+ B{t)u
+ D(t)u
AN(t)xN
C(t)xN
{8.12)
XN(i)
drift to
*^"
=
=
dXa/dt
y
+ B(t)VL
+ D(f)u
Aa{t)Xa
C(<)xa
{8.13)
Then let 8x = XaXN, 8A = AaAs, 8y = yayN, 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)
,,,.
= An()xo + B(i)u + J{t)d
is
/+\
dxa/dt
y
{8.12)
from
8y
A{t)
drives the
rj/+\
/+\^
C{t)xaD{t)u + K{t)d
{8.15)
{8.15) to obtain
d{hx)ldt
=
=
+ 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. 818. This
small output (Fig. 819).
with a large output and a nonlinear system with a
Finally, to
Fig. 818
Fig. 819
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
and
(8.17).
possible.
Example
8.17.
dx/dt
_Mx + L jw.
Xj,(0)
equation
0,
(S.li).
then
Xf^it)
is
zero
u(t) is
dxi^/dt
If
\^
*.,/
The
1 )'''*'"*' (l
effect of small variations in a
(i_:uu:
:)(;::::
Simplifying,
d(Sx)/dt
We
_^sis.
f'^
c is
)a{lef^
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
Find the
l)x.
{1
=
^
dt
84, let
\t
+ 1/
\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/
ft
'*
is
t(leTt)
Then
1)'
+ w^{to)] + (t+l)j'^^ff(r)dr +
ll[n,,(to)
d(t)
is
JC
to
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
8.2.
0\fzi
1
3^
Place the poles of the closed loop system at 4,
Transform the system
5 and 6.
o\
1
1
i\ Ai
1 2 3 jl
1
9/
results of
jz
C2
\0
/eg/
'^*'^"
i
\ /
Ki
V
crVV
5/2
1/2^
3
4
1 )x
3/2
1/2,
Problem
7.21.
CHAP.
8]
183
and
\6 11
6/
we
u^,
set
/a/cA
"2
"3/
which gives
For a
ai
0,
1/2,
k^
we have
1,
Kg
1/2,
d^=0
\6
lx+0211
6/
\l
l/2a
l/2a
In the case a = 0, the state ^i is uncontrollable with respect to Mj and, from Theorem
7.7,
cannot be put into phase variable canonical form with respect to Mj alone.
Hence we must use'
4*2 to control Zi, and can assure a pole at 4 by choosing Mg =
3i. Then we have the singleinput
system
AA
/4
\^3/
03
/2
d y^aX
/I
0\ /^2^
ur^
s){zj ^
""
du^sy
to
give a
u^
12z2
means
+ 623.
placement.
8.3.
+ ri{t),
0.
df
_

/O
[0
l^>^
oy^ +
/fei
UJf^^")'^^"]
^2
di[ej
"
is
y ^ ,
then
may be large even
clude that it is not
/ion
can fil4y\irt
Alter ..tt4out HAin^n
some
and
from reaching
zero.
found from
\^2/j
As
it
0.
o)[eJ + (y2
(y2
 feg =
= 2y
s2
I
2ys
y2
\^
y2s
;
of 41.
the noise.
184
8.4.
[CHAP.
+ l)
(_2
J)x()
y{n)
(Jju{n),
O)x(n)
(1
plane at
(iy)/2.
We
can conan observer to generate the estimate of the state from which we
shall construct
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)
P =
(1
and w{n)
is
Find the
0.05)
and
/O
1\.
gJZ
PL
(1 0.05)(^;
''f){'T)
'''''
is
8.5.
will
it
Use
= SxJ2  2^^.
(^
w{n +1)
obtained from
^_
z(n)
where
0.05w(m)
/1\
...
i^jMn) +
^0.05
{\
2'(')
 2.1525j/() + u(n).
/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.
=
A8A
=
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.
Subtracting
SA)(Wi
A w, +
vf
A^
SWi
X?
xf
6X,
Substituting this
gives
+ SWi) =
(xf
(^^)
Xfvt
SX, vt
w,
{819)
8Xi)(wj+SWj)
gives
it is
xf vt 8w,
vt
of second order
(8X, I
8 A)
8w,
(8.18)
then
leads to
O Ai
/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 highorder systems, whereas
the
procedure just described has been applied to a 51storder system.
8.6.
Given the scalar nonlinear system of Fig. 820 with input a sin t. Should
be increased or decreased to minimize the effect of the nonlinearity?
a sin
\e{t)
/~\
_L 3
sm
^ew(<)
'
Vn
'
if
K>1
~~ *^N
Fig. 820
Fig. 821
The nominal Unear system is shown in Fig. 821. 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.
rocket engine
nommal
J)
4
Gi()
^
J
8T
Giis)
dpcr
sv
"1
+ "2/*
_!_../_
"3
Fig. 822
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
823.
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. 823
8.8.
with an
Compensate the system {s + p,)'i^ + V,)' ^o have poles at u, and ^,
effect
the
discuss
theorem, and
observer pole at ^ by using the algebraic separation
of noise
rj
at the output.
The
J/
0)x
(1
as
The feedback compensation can be found immediately
Tri^2
{PlP2
let
P=
Pl
+ P2~n~'^2)x
aa)
(ai
Then
1
1
I
(i)
from which
^^
Also
(.oiPL)f7
PL)
(F
PB = PJ =
a2
(pi
+ p2^oK
(i2)
:f)
\P1P2
and
"
~Pi
=
P2
PiP2]2
A
f
dt \a2
To construct the
^al) +
+ P2^o)PlP2]y + U +
Wo(Pl
estimate,
^P\i/wN
i
')
2y
is
shown
TTQ
Pi
^/w
 P2/\V
in Fig. 824.
+1
(s
w/ao
Pl
+ P2~''l~'^2
jro
P1P2
'^i^z
xi
'
1+
U C)*~
 Pi  P2
1/
Fig. 824
I
Pi)(s
'^0*Pl
I
P2)
+ P2  To) 
P1P2
PA
CHAP.
8j
187
Note the noise v is fed through a firstorder system and gain elements to form u. If the noise
level is high, it is better to use a secondorder 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
secondorder 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. 825.
ij
1
(S
I
+ Pi){s + P2)
o/ 4
^2
fl
*.
F2I
^
_i
^Vr2
PlV2
Fig. 825
diagram
Supplementary Problems
8.9.
Show
when
the
?Q=^
G{HHG)i
>
H
Fig. 826
8.10.
diagram of Fig.
Fig. 827
828.
=<
G3
I
+
H.
to obtain H, isolated
'
t"
H2
Fig. 828
Cz
188
8.11.
4iy
'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. 829
8.12.
8.13.
/l
2\
01
8.14.
J/
0,
l)x
(0
1 and 1.
dx
1
1.
Construct a firstorder observer system with a pole at
8.15.
:>{>
dx
2/
0)x
(1
8.16.
dx
d/T
"
/3
^1
3,
2/
l)x
(1
4.
Construct a firstorder observer system with a pole at
t^e/orm
8.17.
What is
y = C(t)x + D(t).
Given the system dx/dt = A()x + BW where
theorem when D(t)
separation
algebraic
the
is
What
0?
oblerver system when D(t) #
8.18.
dx
and
4^1
dt
(0
of
the
0?
l)x
VO
34a(t)y
8.19.
B(t),C,(t),D{t) andu(t)?
8.20.
,,
di!.ldt
+ *+
_
 P
y
f[t)
^^*^
"/fV.
^"
fi
/
variations in Rt),
will be insensitive to small
8.21.
gwen
Choose
(t)
d^.m
+ I3{t)e
=^ ^^^J^
with
fWx TolLw
atfsS
CHAP.
8.22.
8]
189
<{y<i}
F(s)
+
^{y}
=Q
^{d}:
<>=
G(8)
^{e}
H(8)
Fig. 830
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 timeinvariant system
is
N.
+ B(t)d
and
of type N, then
it is
Show
C(t)*ABc(,o)w(o)
to
C(t)w
+ g?
of type
Nk
is
that
(a)
(6)
=Q
if
nonsingular.
H(s)
Fig. 831
8.26.
Answers
8.10.
to
d(t)
sin
Supplementary Problems
This cannot be done unless the indicated inverse exists (Fig. 832). The matrices must be in the
order given.
^^0=^
>
G4Gi[IhH2(G2^G3)G4G,]i
+
H,
Fig. 832
G2 + G3
I>
190
8.11.
d{t)
ramps
(t
S'(t)
At
a;i(to)
+ J \M']^2(to) +
Wi(<o)
[CHAP.
*o)]
*o) ^2(*o)
In
tends to zero, so
for
<
> some
ti
l(e> + (i'
lit)
'dt
iji
<t>
9)x
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. 833
8.16.
8.17.
Subtract D(f)u from y entering the observer and this reduces to the given formulation.
8.18.
U:{Sym = 
8.20.
..
,';;"
Q(t)
8.21.
pit)
such that
u{t)
N1
that lim
e(*)
e{t)
G(s)(I
8.22.
lim
8.24.
Use Theorem
8.2.
8.25.
Use Theorem
8.2.
8.26.
One answer
[F(s)
det Q(t)
a(T))dn
"to
is
and
+
k is
dXfji/dt
7(*)e'T" "
(1
dx^i/dt
for all
+ l)(s + l)
"
+ t + /(t))Nl ~ u{t)xi^
f(t)Xf^i fiXffi
t.
pir)e{r)dT
~/it)K
where
an arbitrary constant,
+ G(s)H(s))i]s^{d} =
H(8)
^{"(*)>
4(s
:")".(::;,./
^Xf^i
Choose
^Mt)} =
where
:^^r^!!r"f?.
l)(s
1)
(s
unobservable.
is
8(82+ i)i^
and
is
analytic for s
0.
9(J)
is
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 timeinvariant systems. However,
The concept
we
is
quite compli
For the unforced linear timeinvariant system dx/df = Ax, we are assured that
the solution x{t) = TeJ^'o^Tixo 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 RouthHurwitz
orno 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 loworder timeinvariant systems, and
we have
seen
how
Now we
chapter.
this
we must
9.2
first
Definition 9,1:
to
Even
we can say
if
such that
ix(i)
^k
to,
if
for
the response
is
bounded.
all
to what.
Example
9.1.
Consider the scalar zeroinput 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
so that
e(l
of
Xo
>
1,
t+t,
It
nonlinear
equations in general.
Theorem
9.1:
independent of the
x(^)
191
dx/dt
A(t)x
[CHAP.
192
Proof:
Since
[x(i)!
a constant,
is
[xoi
Now we
]^(;0,xo,io)l
*(i,io)xo]
becomes unbounded as
if !x(i)
i*
it is
due to
solely
*(*,
x=
*o).
is
steady state solution (an equilibrium state) to tlie zeroinput 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 =
<
[jx
e.
>
0,
how
no matter
small, then
we
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,
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
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. 92(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
iange
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
{lXl)X2
dt
with
initial condition
x(0)
Xj
The trajectories in
shown in Fig. 91. We
Xq.
sponds to Fig.
Theorem
92(e) of
9.2:
The
all
is
Note
x(t)i is
Problem
Fig. 91
9.1.)
stable i.s.L.
bounded
if *(*,MII is
bounded.
of
dx/dt
A(t)x
CHAP.
If
9]
is
we
k(o)S
>
#(, fo)xo
8 ==
From
x()].
193
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
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.
*((,0)
_.
. )
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.
9.1,
since
any small
if
is
to,
9.6.
If 8 does not
depend on
tg in
Definition 9.2,
Example 9,7.
The stability of any timeinvariant system
The exception
we have uniform
is
stability i.s.L.
uniform.
If the linear
to
system dx/dt
x(t)
^0 as t^
fixed p
ao,
A(t)x
is
however large
is
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:
only
Example
is
given in Problem
is
positive constants
The proof
A{t)x
[CHAP.
and
all
U.
9.2.
9.8.
Given the linear timevarying 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.
= Ax
is
t,
dx.ldt
A(i)x,
has
if A(i)
mean
if
and only
all
its
if all
the
eigenvalues
the system
is
asymp
9.10.
(x\.\
_
"
dt\xj
3<ce8'<t\/xi\
4k
j\x^j
\Kei>'t
SCxio
+ a;2o)^''' 
(^'lo
+ 3a;2o)''''
and
^x^it)
(xio
3/c.
+ SiCao)*"""* "
Then
('io
if
<
0,
both
+ a'2o)~^'"
nonlinear
There are many other types of stability to consider in the general case of
9.19.4
Definitions
than
more
any
discuss
shall
not
timevarying systems. For brevity, we
next
the
of
those
and
stability,
of
definitions
for zeroinput systems. Furthermore, these
is
difference
only
The
systems.
discretetime
section, carry over in an obvious manner to
that t takes on discrete values.
9.3
its
with
Initial condition
all
x(to)
t
A(i)x
B(i)u
{9.1)
C(t)x
= 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
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.
for
we have
/3
is
>
uniformly
such that
CHAP.
9]
I95
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,
norm
C\h{t,r)\dT
Since
properties to obtain
8/3
external stability.
we
prove
shall
\h{t, T)\dT
j8
by contra
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
From
and
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
to
6*0
this larger
in such a
Xe h{e,T)u2{r)dT
way
that
dr
>
Again we
a.
Suppose we choose
a.
set M2(t)
sgn/i(6i,T)
so that
/so
Since a
y(e).
is
set
and
0
arrive at a contradiction.
Example
9.11.
system
Then
t.
is
\U{t
u.
T)\dT =
h{t, r)
tg
is
U(t
t
t),
tends to
Therefore this
stable i.s.L.
In general, external stability has no relation whatsoever with zeroinput 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.
where
(t,to)
a(f)a;
*^to
\h{t,T)\dT
<
e + 9(t.to)z{
and
air) dr.
/i(t, to)
so that
"^^
\h{t,T)\dT
e' + to.<t>(t,
1.
is
*(*, fg)
to)' + ''"'''
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
9.5:
(9.1)
with h norms,
C{t)
I,
<
To have
it satisfies
9.4
to be
an ^vector.
Theorem
9.5.
If
Theorem
is
9.5 is
proved in Problem
9.2.
LIAPUNOV TECHNIQUES
The RouthHurwitz, 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 timevarying 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.
Definition 9.6:
p(x(t),0).
This
is
all,
v(x)
(2)
v(0)
and
its
= 0;
for X ?^ 0;
v(x) >
dv/dt = (grad^ v)'^ d^/dt <
(3)
(4)
for x
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.
Theorem
9.6:
variable
X=
is
asymptotically stable.
x^
* ..
Definition 9.6 assures existence and
as
Here we shall only prove
Proof:
continuity of v and dv/dt. Now consider v{4>{t; ta, xo)), i.e. as a function of t. Since v> 6
and dv/dt
for i>^0, integration with respect to i shows v(^(ii; io,xo))> v(^(i2; io,xo))
for to<ti< tz. Although v is thus positive and monotone decreasing, its limit may not be
when
(Consider 1 + e'K) Assume v has a constant limit k > 0. Then dv/dt =
zero.
then dv/dt = {grad^vyi{Q, t) = 0. So
for ^ t^ 0, and when ^ V = K.
But dv/dt <
which implies v = 0, a contradiction. Thus v > 0, assuring x  0.
dv/dt=
implies ^ =
i5
<0
Additionally
the
if the system is linear or if we substitute "everywhere" for "in a neighborhood of
condition
(4) is
origin" in Definition 9.6, we have uniform global asymptotic stability. If
If Definition 9.6 holds for all
weakened
to dv/dt
0,
we have
to,
then
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:
v(x, t)
(1)
and
its first
x and
partial derivatives in
exist
tinuous;
v(0,i)
(3)
v(x,)^a(xj)
(4)
Note for
Theorem
all
norm
9.7:
dv/dt
v(x,
t)
to,
0;
is
tion of the
(2)
v^ dx/dt +
(gradx
must be
dv/dt
<
f or
and
a(^)
v^ 0.
xj.
Proof:
Since
dv/dt
<
>
and
>
v{x{t), t)
assumption,
for
is positive,
Now
to.
we have asymptotic
must be
the proof
stability for
some
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.
ment of
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
t^to
identically in
for any
to
Consider the SturmLiouville 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
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)xy,
dv
[CHAP.
then
2p(t) q(t)
di
dq/dt
Q'Ht)
^^
2p{t) q{t)
X2
dq/dt
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,/ci} = ;8(llx) ^^ .(x, f).
then
Definition 9.8:
we have
case,
k>
Theorem
continuous;
(1)
v(x, k) is
(2)
v(0,A;)
(3)
v(x,/c)s=a(x!)
(4)
Av(x, k)
0;
v(x(A;
>0
Xv^O;
for
for
v^ 0.
9.8:
asymptotically stable.
The proof is similar to that of Theorem 9.7. If the conditions of Definition 9.7 hold for
and if v(x, fc) ^ j8(x) where jS() is a continuous nondecreasing function of  with
all ko
j8(0)
or
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.
A(fc)x(fc),
Choose
v{k)
is
A(fc)x(fc)Ix(fc)
The major
if
the system
we
shall
diflSculty
is
A{t)x
[9.2)
CHAP.
in
9]
which there
*(t,
exists a constant
Kg
depending on
some norm.
199
such that
Using Theorem
we can show
9.3
\\^{r,t)fdT
e^'^^^^^dr
Kg(ij)
lA(t)l2
kI/2k^
that uniformly
*(f, t)1
<
k^
and where
K^e^'^i'"'''
00
this requirement.
9.18.
Theorem
9.9:
(9.2),
i'(x, i)
where
4>^(r, f )
=0
x'^P(i)x
Q(t) *(t,
is
i)
dr
which Q{t)
in
positive definite
is
(t/T)dT
to)
*.
For system
P{t)
xl2t.
0.
l<J>(T,t)l2d7
if
t^
>
liQ(*)
any continuous
is
k5(*o)
^^^ Q(*)"~I
small enough.
is
This theorem is of little direct help in the determination of stability because *(<, t) must be
known in advance to compute v(x, f). Note P(i) is positive definite, real, and symmetric.
Since
Proof:
* and
are continuous,
if
<
P(i)
for
all
*(t,
tegral exists for the systems (9.2) under discussion, then condition
obviously holds. To show condition (3),
v{K,t)
Since
Q(0  !
x(t)v^O.
is
= j'\^r,t)^{t)VQ{r)Hr,t)K{t)dr =
positive definite, then
 J)x ^
A(i)2
Kg
holds.
Condition
(1)
^ ej
Example
cZr
ex'^x
.Kgl
Kgl
X^(T)x(T)dT
{9.2),
.j
>
for any
xdr
J^"A(T)yx(T)dr
J'"x''(dx/dT)dr
eKg1
.Kglx(i)/2
f
=
X^Axdr
>
a(x)
x'^(t)
(2)
v(x,i)
dv/dt
of
j^
(1)
t)f dr.
so that x'"Qx
Therefore
v(x,t)
Since
x'^(Q
then condition
to
<
since
is
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
f1
<
t'
K3o)
and
For simplicity, we choose Q{t) = 1 so that P(t) timevarying 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,
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.
X2
and
Xi(v^^+xl)/2
dx^/dt
 x^ +
Xzix^
+ xl)/2
because
it
is
totically stable.
9.6
9.9, if
dP/dt
we
*^(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
obtain
{9.4)
CHAP.
9]
Theorem
9.11:
If
201
is
A(i)
dx/dt
{9.2) is
Theorem
where
{9.2),
then v
uniformly asymptotically stable,
Proof:
system
positive definite,
9.9,
If the
we
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 LienardChipart (RouthHurwitz) test than the
Liapunov method, however.
after
Example
9.23.
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/
P22AI
/I
<>
1
I,
1\
1
'
Using Sylvester's
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,
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 timeinvariant). A hint on how to choose Q is given in Problem 9.7.
Example
9.24.
d^y/dt^
t
As found
in
Example
9.23,
+ {2 + e*^) dy/dt+
=
2i/
dv/dt
{.:
Sxf
4)
Zx^x^
d{TsJPii)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,
>
eV2
and
e""*
>
et
stable.
Theorem
in
9.9 is
v(x,fc)
mk
of equation (9.5)
A'^PA
is
P = Q.
Solved Problems
9.1.
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. 92
We can see that if the equilibrium state is g:lobally asymptotically stable as in Fig. 92(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.
We
is
to
Since
and tends
x(t)l
i*(t, to)IMxol,
to zero as
t >
i;
17
7;
ll*(to
+ fcr,
to)ll
k such that
Choose
Theorem
9.2.
\\4>{t^
to
+ kT,t^ + (kl)T)\\'
1)T
ll*(to
and
+ 2r,
to+r)llll*(to
define
/cj
*(t, to
+ fcr)
*(to
+ fcr,
to)
+ fcr,
Kieii*(to
Kie
Kg
*(t, to
T.
to)ll
1/T
proves Theorem
9.3.
to)lI
+ kT)\\e <
^i^
Ct0+(.k+l^)Tt0^)/T
pk
Then
I*(,*o)ll
Defining
(fc+i)
<
from
CHAP.
9]
KjCfto)
If
ll*(,*o)lliie
then
H(f,T)idT
l*(,T)ilB(r)lldT
If the
system
for
it
>
"2
)/k2
is
KgKi/K2
to
/3
H(t,r)l,dT
J"to
e""^""^' dr
9.4,
*(t,T)B(T)iidr
to
*,T)B(r)lliBl(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)lidr
*(t,r)lidr
is
Bi(t)]i
9.4
/3
Since
KgKi(l
Theorem
SO that use of
*0
to
203
/3B1
<
I*(t, io)Ili
from Theorem
l*(t,T)Ii]*(T,o)lidr
S=
J"to
to
r mt,h)\Vdr
9.2.
!!*(,
Therefore
r)*(T,to)ldr
l*(f,*o)!i(to)
9.3.
system dx/dt
2i(2 sini
l)a;
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. 93). The peaks occur at ^{{2n + l)ir, 2nTr) =
gT7(47r)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
We
is
GronwallBellman 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
dr,
+ /t(t) ~
for
is true.
/((r)]
Also show
A(r)x(r)dr
>
<o(t)
T) da/dt
to
9(t, to)
>
and
lxo
(t)
u(t).
Applying' assertion
k.
,ect,to)
^ U (A(f) k),
xoe'''o"*''"^"''''
gives
\\
<o(to)
e'"''''V(T)dT
x(t)
xq
norm
{t^
"(t)
between
<o(to)
bounded for
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)
0,
in assertion
and y
ait)<p
Let
Show
ix(t)l
*i
then
0,
e^''o'du/dtaltle"""'"'
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,
Proof:
a{t)
(ii)
Assertion
a)(to)
then there
0,
Assertion
dw/dt
If
[CHAP.
A(r)x(r)dTl
^to
+J
^
Taking the
xo
f
^h
A(r)l x(r)
^(t)
dr
to obtain
9.6.
ll^oll^'""'''^'"'"'
solution
A^P +
PA = Q?
BP + PA = C where A, B and C are arbitrary real nXn
P and C as p; and q respectively, and the row vectors of A as
equation BP + PA = C can be written as BP
2 Pi*,^ = ^' '^^i<=^
We
in
for
1, 2,
. ,
w.
Then the
I
'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
unique solution to
Thereore
and only
/anl
...
ail
Vfti^l ...
al
\0
\ /ynl
T/
\i;iI
...
v^^l'"
...
rI,
aj./
det (B
+ A) =
fl
(i
+ /3j) #
0.
if
if
9.7.
Show
that
PA + AT = Q,
definite solution
This
as
if
is
lAM =
Note P is
(Mix)t(Mix)
> 0. Also, xtQx = (Mix)t(A + At)(Mix) > if and only if the real parts of the
eigenvalues of A are in the left half plane.
Furthermore, ^ = xt(Mi)tMix 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 =
v,
this
;
9.8.
_
This
is
dyr
dt^^ + ^'^% +
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,
is
[CHAP.
'_
dt
was
Supplementary Problems
9.9.
dx/dt
tx
+u
and y
x.
\h{t,T)\dT
^ V2^e'
'0
910
9 11
9.12.
9.13.
(6)
Show
(c)
Explain
the system
is
not stable
i.s.L.
why Theorem
Show
(a)
(6)
>
Explain
0.
By
Show
dx/dt
(1
+ t)x
is
asymptotically stable f or
if
given
P which
using Liapunov
theory.
9 14
and
AAAAAB(t)
L(t)
:c{t)
with
/R + {2L/RC)
P(*)
find conditions on R, L,
2/R
(^
Fig. 94
stability.
9.15.
that if >
Using the results of Example 916. show
<l</ is< unilormiy ^Jr^i^ol'^ly'st^l
asymp
ady/dt + (1 + P cos 2t/^)y +
d^y/dt^
equation
Mathieu
the
even
^Zilf
the tie
Uapunov
=^m
v,.n> e
function
((x)
x^x,
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
discretetime
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,
is
where
x)/2,
207
9.19.
^x
f~^
"^^^]js.
\a
dt
X J
*(i,T)
if
9.21.
= (A(t) + B(i))x where dx.ldt = A()x has a transition matrix with norm
Using the GronwallBellman inequality with /i = 0, show 11x11 ^ lxne'''3"''2'""'o'
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
(ttl
9.12.
Replace
9.13.
Use
<
<
<
<
9.14.
^'"'
Ki
Kg
=
*
^
^
to read
(4)
^ ^^e'^^^
e^'^ dr
/3
depends on
in
Theorem
dv/dt
>
in
x2
R{t)
Lit)
^
^
K2
<
K4
<
<
=0
00
i^always
positive definite,
eigenvalue of
Q=
and
Mi(I
9.19.
If
9.20.
Ilxll
2I,
thenar
(^
^"
xo2<'*o'
J^y
Use Problem
9.4
with
lx(*)
^.<^)
is
which
is
positive
than one.
0.
If
0,
o_
IIBWII
llxlle''^'
9.21.
The system
^
 AtA)Mi
less
9.18.
require
9.4.
in equation {9.5).
^^'^
^A.ZTX'^a'
definite if and only if each
Q=
A  al
K5 =S C{t) =^ Kg
A(t)x
t^t^Xi.
for
It)l2
is
Change condition
by
O
is
9.11.
dx/dt
Supplementary Problems
to
e'''^
xq.
//2e'o/2
(5)
00
+ J^'/'"^"""*'B(r)uWdr
'"
(t),
]x
llxll
dr
lxo!e''^'
+ .sflMdr
A(t)l
p(t),
and
BWu(r)l
,(.).
we
chapter 10
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 spiiit 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. 101 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
C{t) is
x()
y()
System
<l(f):
{lu.i)
+
+
*
Fig. 101.
Definition 10.1:
regulator
is
f'
s,
made on
x.{to)
xo
in general.
?^
We
d{t)
0.
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.
U(t
to)
Under the
restriction that
we
we
is
u(x(i),
t)
208
CHAP.
10]
209
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.
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 timevarying
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
t <C t\.
x^^x
tK
i^i^rs^i t^T'
The choice of S
Example
dictated
by the
TV,,.,,
^^^ ^^^^^^^'^^
"'
Tl ^''tent:li^:o%i:Z.T'
^Tpath
'M T"''
f flies to arrive
the missile
\ "^" makes no difference what
^'''
''^
where
10.3.
stationed 'ir
t
is
e,, e^, e,
'^"^ ^^^^^
near
at
*.
j.
210
S(i
 U)
{10.1),
[CHAP. 10
is
t>U
in the case
u=
and
p(x, 0)
its
0,
minimum
value.
We
jJ'u^(r)R(T)u(T)dr
where
R(t) is
mxm
an
definite
the designer.
10.3
is
iQ(T) is
relative to
iR(T)l,
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
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)
p'^cZx/dr
(ix^Qx
H
iu'Ru
0.
Addmg
by parts gives
v[x,u]
vjx(f,)]
"
+ B(i)udx/dt] =
Wx] =
v,[x(t,)]
ix''(i,)Sx(iJ
f ''(ix^Qx +
^
vjx]
+ V3M
x^(i^)p{i,)
3i:^Ap
+ x'^{tMK)
(^^^)
+ x^ dv/dr) dr
{10. If)
to
broken into v,, v, and v,, and heurisIntroduction of the costate p has permitted v[x, ti] to be
minimum when v,,v, and v, are each mdependently
tically we suspect that v[x,u] will be
function attains a local minimum
minimized. Recall from calculus that when a smooth
of
suspect that if v, is a minimum, then the gradient
its derivative is zero. Analogously, we
{10.3)
with respect to
x(fj) is zero:
p(ti)
and that
is
zero:
if v, is
(^0^)
SxoCii)
of the integrand of
AJ{t)v
Q(*)x
{lO.J,)
with respect to
i^^'^)
CHAP.
Here
10]
211
system
is
(10.1).
Theorem
10.1:
costate defined
= R^B^p
"
The proof
(10.8)
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.
Then
A=
2,
B=
R=
1,
1/4,
Q =
3,
x2(j)
S=
Xo,
p(l)
2a;(l).
/xHt)\
[pit))
Evaluating this at
with criterion
+ uyi)dt
solve
4\/a;op\
[S 2)[p J
5,
4N/a;op(0)\
+ ^[S
2JJ(p(0)j
_

and using
re /2
L^U
p(l)
4\
e)
2a;p(l)
^''''^
gives
10.1,
^'"'(^o.t)
Note
u (Xo,
{3x^
J
_

+u
2x
Hence we
d_/xp\
iBOPCO)
dx/dt
2.
dt[p
with
B(i)Ri(^)B^(^)\/x^\
A(t)
Op + it
1
4 orxpSit
sr/s
^0
(10.m
i)
Ax^p
+ Bup(x'"'(), ^).
[CHAP.
212
Example
To
10
10.5.
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),
in the
box marked
t)
^, ,
K(i)
Y^^^a=tT
2
I
^''(*)
+ 30e8(i
]5g8(it)
is
_
"
dx"P
IT
20.38""
5e8(i
10.4
let
P(*)
P(*)
To find the timevarying gain matrix K(i) directly,
RBW^
t)
=
Then
u^lx,
an n X n Hermitian matrix to be found.
 BRB^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+
But
x'"'()
*,,(, o)xo,
dF/dt
A^Fx"'
then gives
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  PBRBT
=
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
Ut,
pendent variables by r
The equation
PA  PBR'BT where the arguments of the matrices are Ur insteadof 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.
Example
10.4, the 1
dP/dt
with P(l)
2.
Since this
**'
is
dP
f 4{P^"3/2)(P
+
l72)
P(t)
K(t)
= R^BTP =
+ 4P4P2
^

3/2
nt)
,..
8^"F(ti)3/2
_

Then
is
ti
1,
zw+i/^
*" P(ti)
1/2
gives
15 V e8""
10  2e8i'
in
4F(t), which checks with the answer obtained
Example
10.5.
CHAP.
10]
is
213
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)
U^t^
10.7.
For Example
10.4,
which reduces
We
to the
'
[2e4Cti)
answer obtained
in
6e4<ti)
Example
2(4e4"i)
4e*<ti^)]/8
10.6
Theorem 102:
If S
and
v[x. u"!
Theorem
10.3:
For
S{t), Q(i)
and
equation (10.13)
is
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 timevarying case are given in Problems
10.20 and 10.21.
10.5
U^t^
214
Theorem
system
10.4:
controllable, then
{10.1) that
lim P(fo; U)
[CHAP. 10
n{to) exists
''"
tive definite.
J,"
v\Ti,U^
Note
v[k,u^]
Xa^'sQxasdi
from Theorem
Here
x(fo).
Ps=o (*"'*)
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
It
0,
when S >
<
limPg^(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)
<
ii
must
n(to)
Since we are now dealing with a timeinvariant 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.
dP/dt
= 1  P2
with
n =
Since n
is
is
0.
lim P(to;
ti
P(ti)
=u
up
with criterion
"
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.
it satisfies
= Q + A^n + HA  nBRiB^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
controllable.
it is
Theorem
10.5:
n exists for the constant coefficient system, the closed loop system
asymptotically stable.
If
Proof:
for
all
dV/di
for
all
V = x'^Ux.
Since n
= x^n(AxBRB^nx) + (AxBRiBi'nx)rnx =
because
is
positive definite
and IlBR~iB''n
is
x'^Qx
is
positive definite,
x^'nBRiB^'nx
nonnegative
definite.
V>
<
is
CHAP.
10]
Theorem
The unique
10.6:
Proof:
solutions of (10.16),
n2)(A
definite solutions Di
helps to
it
know
that
is
the
(n,
+ l)
215
^^'
Ea.
qq
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  BRBn. for i = 1
is
as
t ^
00^
Theorem
which
10.7:
n tends
IS
true because:
to zero
8P obeys
Proof:
dSF/dt
from
(10.13))
= Fr8P + 8PF8PBRiB'^P
where F = ABR'B'"n.
we
brevity,
2F  n
zero SP
shown m
Example
10.10.
snnnLT
an^nLi
^' *""
^/T^!.^
TimpTori
^^'
'""'''^"'^
*^^
'^"^'^*'" "^"^ *
''^^'^
^'n!'"''*^^'"'
''
'^''"'"
^"
^(*o)
tanh
(*,
tanh"!
e).
i.e.
For
Theorem
10.8:
If
a^ is
(10.9),
Proof:
Denote the top
elements as gi. Then
n elements
\ii
\S^
^?)
of the
=
=
A,
Qf,
(B^^B.
ith.
eigenvector of
as
BRiB'g^
A^g,
:?)(7')
H^(7'
[CHAP. 10
216
>
Example
10.11.
corresponding to
Example
10.8
H =
is
matrix
which are symmetric with respect to the imaginary
The
Example
axis.
_^x
(_^
V
10.12.
shown
in Fig. 102.
ImX
ReX
Fig.
1(12
The
Theorem
10.9:
If Ai, A2,
.,
A are
and
p{o>^)
distinct eigenvalues of
= GFi where F =
(fi 2
real parts
is
another
way
< 0,
then
()
and G =
fr.)
>
into p''{<^)p
(gilg2
gn)
as defined
in equation (10.18).
Proof
show
first
Since n
is
GPi
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.
{ly
from which
with
H(S)
from
(4 ^".n
/m
io\
{^0.19)
FA = AFBRiB'^G
GA = QFA^G
(iO^O)
A in (10.20) gives
GFiAF  GFiBR^B'^G = QF  A^G
Next,
Hermitian.
then Hermitian.
we show GF^
GFi = FtiMFi
is
is
It suffices to
M = FtG Hermitian,
of M be mjiciigk; tor
show
is
since
j
^ k,
CHAP.
10]
Q n^
^^i^UAi
>^3r^>'5tiHV
Since the term in braces equals
Finally, to
e(<).Fe^'F
we have
!^
217
.K
I
i\hV*^
= m*
d^/di
m.'^
= (QF + A^G)e^'Fi
Then
GFi
t(o)^(o)
j'J^^e^^at
so that
(e^'F>)t(FtQF
Since the integrand
is
positive definite,
GFi
The proof
is
is
is
GF'
GtBRB''G)(e^'F)di
positive definite,
x(^)
Fe^'o^Fxo
n.
with costate
p(*)
is positive definite.
and GF^
A  BRi B^n
are
ABR'B^n
Ifi,
fz,
are Ai,X2,
fn.
Example 10.13.
The H matrix corresponding to the system of Example
Correspondmg to these eigenvalues,
1 =
X,
and
c*""
+1
rx =
X,
^2
where a and
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{tti)
For y
= Ix,
all
[CHAP. 10
218
To reconstruct the
state
C{t)K
dy/dt
differentiated
n1
times.
Ni(t)x
N,(^)dx/di
d"y/di"i
is
+ dNJdtx =
(NiF
(NiF
+ dNi/dt)x = Nax
+ dNi/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
nvector
the
conclude
results of Problem 3.11, page 63, we
R^iBTN^^Hy.
=
termined. The optimal control is then u
.
Example
10.14.
dt
012
X2
+ k^x^.
System
Since y
2/
(1
0)x
X2
x^
and dx,/dt
x^,
(iO.l)
State
estimator
R^B^
Pig. 103
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. 103.
10.7
problems.
We
wish to
find
Fig. 104.
Fig. 104.
CHAP.
10]
The
criterion to be
minimized
W{ti)
Note when
e{t)
y{t)
 d{t)
is
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.
(0
jx +
e~^<
2/
\0
62
)m
2/
(1
1)x
V63/
= z^ + Z2o( +
1) + z^^e^K
In other words,
functions and arbitrary linear combinations thereof.
d(t)
manner permits
dw/dt
A(t)w
defining
B(t)u
new
e
state variables
we can
consider as
w = x
z.
C(t)w
Then
(10.22)
Now we have the regulator problem (10.22) subject to the criterion (10.21), and solution
of the matrix Riccati equation gives the optimal control as u = R^'B'^Pw. The states
can be found from e and its w  1 derivatives as in Section 10.6, so the content of the box
marked * * in Fig. 104 is R^B'^PN'H.
A.(t)z
where
g(i) is
g(t)
0.
if
is
not
assures us the
the closed loop
Restricting the discussion to timeinvariant 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
8.2 applies.
is
integral compensation
when
10.16.
Consider a system with transfer function G{s) in which G(0) ^ , i.e. it contains no pure integrations.
We must introduce integral compensation, lis. Then the optimal servomechanism to follow a step input
R^ib^nN~iH(s). This is a linear combination
is as shown in Fig. 105 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. 105
u{t)
y{t)
10.8
[CHAP. 10
220
CONCLUSION
For timeinvariant
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=
onetoone because analysis similar to Problem 10.8 shows that a control
of this equivalence
dual
The
^
all
a,.
for
ldet(y,IAbk^)
det(>IA)l
if and only if
KalmanBucy 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,
10.2,
page 213.
(1)
(2)
(3)
We
(4)
u(t)
t).
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)
To
r[x, uP]
+
Since
we can
v(^(to),to;ti)
parameter
to)
*'
Ij
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.
+ ix^Qx 
is
v(k,
t)
to the
^(gr&d^v)'^R^Br grad^v
v(x(t,), t,)
iicT(t^) Sx(ti).
+ (erad^v)^u
+ RiBTgradxV)
t),
^(u
RiBTgrad3,iJ)''R(u
s=
and dx/dt
Jx^Qx
(10.S3)
RiBTgrad^i;
(grad^i;)i'Ax
^(grad^v)rBRiBTgrad^i;
u(x,t)
Using
r)Ru<'P(*, t)]
(W^ni.
= Ax + Bu, we
i^^^^i)
(grad^t;)^Bu
get
^.,(x(t),t)
CHAP.
10]
221
IX"
(xTQx
+ u''Ru) dt
s=
v{K(to),to)
u(x(ii), ti)
to
any
is
4x,u]
ix'r(ti) Sx(ti)
'
(x^Qx
uiRu) dt
ss
i,(x(fo), <o)
to
so that if v(x(^o)>
*o) is
nonnegative
definite,
u<>P(x, t)
then
it is
Ri(t)BT(t) gradxV(x,
control, (10.2^):
t)
To solve the HamiltonJacobi 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 HamiltonJacobi 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
and
all
nonzero
x(to),
'y(x(fo),o)
Since
10.2.
for all
we know from
Theorem
10.2.
>
'[xu""]
%<
t^
10.1.
 x/2t
= ^ C\x^ + u^)dt.
dP
dt
<t{t,r)
(r/t)i/2,
which escapes at
is
_
~
,
_ P _
P^
~fi
0.
Pit)
f?
t\
+i2
f2
in
10.3.
Given the nonlinear scalar system dy/dt = y'^e. An open loop control d(i) =
t has
been found (perhaps by some other optimization scheme) such that
the nonlinear
system dyjdt = y^d with initial condition i/(0) = 1 will follow the
nominal path
Vnit)  (1 + 1^) ^'^.
Unfortunately the initial condition is not exactly one, but
y{0) =
1 + e where is some very small, unknovra number.
Find a feedback control that is
stable and will minimize the error y{ti)  yn{ti) at some time
U.
Call the error y(t)  y^{t) = x{t) at any time. Consider the
feedback system of
The equations corresponding to this system are
dy _ dy
^x
di  'df^ It "" (Vn + =o)Hd + u)
=
Assume
the error
\x\
yld
<
+
\y\
\Zylxd
Zylxd
+ Zy^xU + xH +
and corrections
+ ylu\ >
\u\
\Zy^xU
<
\d\
j/^m
Zylxu
+ Zy^xH +
x^u
so that
+ xU + Zylxu + Zy^x^u\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. 106
Since dyjdt
yld,
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) *
dP
'
p2
etp
dt
(1
t2
t2)3B(t)
10.4.
any
+ t^W^R(m
where
do is the value of
ljl{ydoy +
{s
+ a)\ Find
Then
a compensation
{du/dtY] dt
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 Xiu 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
(e^
f
^0
then
/I
i^^^)
(0
l)w
dt.
n
where
optimal control
V^y = yJa^ + V^*^ + sfJ^^V^^ and Ai = a2 + 7a + l. The
system
is as shown
loop
closed
=
the
=
de/dt,
= A(wi + ywj). Since e w^ and Wi ae +
is
in
CHAP.
10]
step.
"'
+ a + 8)
A(y
)
/
223
f*
+ a)
8(8
Fig. 107
Figure 108
is
is
integral
plus proportional.
tO
step
I
Fig. 108
10,5.
A
dt
where Zc are the j controllable states, za are the nj uncontrollable states,
Je and Jd
are ^X; and (nj) x (nj) matrices corresponding to real Jordan form.
Section 7.3,
b is a yvector that contains no zero elements, m is a scalar control, and
the straight
lines indicate the partitioning of the vectors and matrices.
Suppose it is desired to
minimize the quadratic criterion
where
'
 f"
[z'^ Q,cZo
+ p^u^jdr.
Here
>
a scalar, and Q^ is a positive definite symmetric real matrix. Show that the
optimal
feedback control for this regulator problem is of the form
u{t) = k^x^it) where
k is a constant ^'vector, and no uncontrollable states Zd are fed back. From the results
is
which a general timeinvariant 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
A =
led
n =
Then
^cd
p(b^
OT)
^
= pb^nA
pb^HedZd
"^cd
Now
if n,.
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,
JcHed
n^dJd
pOcbbrn,,,
criterion
Then the
J.
shown for
z^
^j"
i f"(xTQx + piM2)d^
and
{z'rTTq,'Iz
T^QT 
za if
piu^)dr
Qc
(
'
10.6.
A(t)x
B(t)u,
C{t)x
{10.26)
that the closed loop response of this system approach that of an ideal or
model system dw/dt = L(i)w. In other M^ords, we have a hypothetical system
dw/dt = L(i)w and wish to adjust u such that the real system dx/dt = A(t)xf 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
Note that
scheme
in the case A,
L=
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
ds.ldt
A = ABK,
Axf
Bu
1
= i
CA  LC
/"'l
J.
A. A.
(xTQxfumu)dt
= (B^QB + R)  iBTCQIVI,
Q = CT[(MBKrQ(MBK)h K^BK]C and R = B + B^QB
dC/dt
I
(10.29)
becomes
(10.28)
^

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 = RiBTPx 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  PBRiRTp
dP/dt
with boundary condition
(ti)
The control
0.
u = RiBi^(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
10.6,
C{t)x so that
it
 v)^Q(y  v) +
[(y
u'rRu] dt
rf^(x)
and the criterion
is,
a)(J
(o
+ (b
(10^2)
^VCJ
Thus the
{10.31)
/So
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.
in Fig. 109.
OWo
Model
Mt)
>
v(t)
Feedforward
u(t)
:^
Plant
+
Feedback
Fig. 109
y(t)
x(()
>
C(t)
226
10.8.
[CHAP.
10
Given the timeinvariant controllable and observable single inputsingle output system dTildt = Ax + hu, y = c'^x. Assume that u = Ux., where k has been chosen such
that 2v
{qy"^
+ u^) dt
is
optimal,
= b^n
=
qcc'^
and g
*
where n
^ =o.
which
satisfies (10.16),
is
written here as
+ nA  nbbTn
Ai'n
(10.33)
loop eigenvalues.
sD;
from
(10.33),
 (sIAT)in 
9(sIAr)iccr(sIA)i  n(sIA)i
(sl
(si
 A)i
^ A^)inbbTn(sI A)i
(10. Si)
Jl{y(t)}
cr^{x()}
cT(sIA)lbp(8)
^{m(0} = bTn^{x(i)} =
and
is
1 to
G(s)p(s)
 A)ibp(s) =
Then
is
H(s)
p(s)
becomes
(lO.Si)
qG(s) G(s)
==
Addmg
b''n(sl
H(s)
H(s)
H(s) H(s)
+ 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
wm
nm
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(mi)
where
loop poles are asymptotic as shown in Fig. 1010.
independent of the open loop polezero conAlso, note equation (10.36) requireis the
^ 0.
closed loop poles to tend to the open loop poles as 9
This
is
figuration.
(qy^
vP) dt
is
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.
^^'
CHAP.
10]
227
Supplementary Problems
10.9.
minimize
to
2v
^0
10.10.
10.11.
+u
with
desire to find
(a)
(6)
(c)
(6)
find
(c)
verify
minimize
to
a;(0)
Xp.
^5)'' + (f'"
Hi
We
(a)
initial condition
4 ^2^ ^i^
(9x2
Ax
2c
d(.
p(t).
i).
Problem
criterion of
+ m^)
(10.9)
m(xq,
(^x\
10.10,
P(t)
p(t)
P(t) x(f)
(f^j
the
of
P(t)
(6)
matrix,
of Problem
10.10,
the results of
(6)
(a),
and that
P(t) is constant
and
positive definite,
draw
(d)
10.12.
^(r'
;)(:;)
I r
Assuming the
initial conditions
10.13.
are
What
dx/dt
10.14.
Using the
(b)
What
is
Sx^T)
lim
+ m2)
+u
r),
(c<;2
+ m2) dt
p(t)
P(t) x(t).
P(t)?
(x2
+ uyp)dt
and
find
lim K(t).
Zero input
to
10.15.
dt
(a)
(.=?
dx
,r\
+
LJ
w
.
Je
8 V a
D'^ + fJ'"
(ft
K(t)
2c
(xT:'X
+ m2) dt
Fig. 1011
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
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
[CHAP. 10
and
v(t) in
the form
where
10.17.
= T t.
Here
e is
known as
= (
criterion
"^^^^
Ir^il)"^
10.18.
r
r
_
=
= j
2v
(qxl
+ u^)dt. Draw
the
>.
,1
62/r"
L ?
a J
dx/dt
ax
bu
"
x{Q)
Show
((fx2
+ rM2) dt
p(T)
xo;
a and
h.
10.19.
Given the system (10.1) and let S = 0, Q(t) = v^Ut) and R(t) = pRo(i) in the criterion (10.2).
Prove that the optimal control law up depends only on the ratio ii/p if Qo(*) and Ko(t) are fixed.
10.20.
Show
and
10.21.
10.22.
also
det Li,
,^
/,
is
symplectic,
'
~"
i.e.
to minimize
''
r*^
^^^^^ "^
= x
dx/dt
^"^^
tan
P(t)
0,
find the
u,
This
'^^^
Find the solution P(f) to the scalar Riccati equation corresponding to the open loop system with
finite
<
for
10.24.
show
Using the
u(x,
10.23.
to
<
h,
for
(x
<
*i
+ u)/t
<
with criterion
"
<
to
<
a^d ^o^
C[{yd)^
ytj
'(6a;2
<
*i
+ u^) dt.
where
to
3/2.
+ u^]dt.
OWo
d{t)
y^
step
t
N
e(t\
e(t)
Fig. 1012
M(f)
u{t)
1^
y( t)
CHAP.
10.25.
10.26.
10]
Use the
results of
Problem 10.4
10.27.
if
229
R =
and
Q=
in equation
system as a varies.
0?
dn/dt
= Ax + Bu
with criterion
2c
(x^Qx
+ u^Ru) dt
where Q and R are both timeinvariant 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
^i'^2
'
satisfies
g2n
?^n
Sn +
.,
X2n
in
equation
(10.18),
where
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+
[(i?^
1)^^
li^]
dt
for constant
i;
Now
=
>
>
1(0),
1.
all
Jo
7;
>
1.
is
9.
with
11
is
d^/dt
ejs
asymptotically stable.
asymptotic stability of the closed loop system.
10.30.
A(t)x
f
xr(ij) x(ii)/2
Xq
'
+ r uTu dt/2
(a)
What
(6)
(c)
I.
Solve the
230
Answers
Supplementary Problems
to
= 9x
10.9.
10.10.
(a)
dxjdt
dx^ldt
=  VI X2
dpi/dt
dp2/dt
= pi +
(6)
Xiit)
(c)
(a)
X2
P2
4x1
(2xiQ
a^i(O)
x^q
2(0)
^^20
Pi()
P2()
y/5p2
+ X2o)e* 
(xio
+ X2o)e^*
P2{t)
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\/54)(a;io
= 4^2^
(6)
011
2,Si2
6
3
~ V 5 012
P()
+ a;2o)e2
2\/5
/S
012022
,622
0^2
id)
Fig. 1013
10
''''
0'
^2
di\ p.
q
01
10.13.
(a)
P(t)
[S{1
1 U
a'a
II
Pi
with
lim P(t)
t* 00
10.14.
K(t)
=
(a
10.15.
Any
10.16.
M(9)
5, XjCO)
=;
2vVTp
3, p^iT)
0,
p2(T)
0.
0/ \ P2/
[V2I +S]e2V^cr +
(6)
a;:i(0)
= 1 + V2
regardless of S.
pgyUit)  1)
 7^2+7 )e>'"i + Va^ + P +
control results in
= AHQrS +
9r9V2
+ V2 5'
where y
since there is
g^g^e2
lim
an unstable uncontrollable
qrq^eyS)
where
A =
g9
/:(*)
a Vo^Tp
state.
g.ffS/S
9r9eV12
CHAP.
10]
10.17.
This
10.18.
10.19.
uop
is
= Va2 +
=
62g/r
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,
<
i^Kji
p((i)
Let the
KiAVeA^"'!'
K22y >^
/^"
=
=
e^^'^i^.
\/Kn
K,A/x(ti)
K^i)\p(ti)
x(f).
Then
p(t)
P(t) x(f).
2m n
first
1, 2,
and
eAu(tt,)y^K2i
terms of
x(ti) in
e''''^^^
m.
f2ii
10.29.
0,
(''^^^\
for
 Fohm^^hFov/p
Vp(t)y
10.28.
where
PflA
6tJ(6tif6
P{t)
+ ATPo +
:r^]x
10.23.
E
Qo
of time in
*'r(,
as
i)/p
*T(t, <o)(HTE
10.22.
231
= (RefiImfiRef3Imf8l...Imf2,if2ml
Then F and G are real and n = GPi.
and
(a)
From Chapter
similarly.
<
i){2
10,
Vi
(,
+ 1)^2
9,
Vg
... fn)
i^^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
8211
821
INDEX
Abstract object, 26
uniform, 14
Adjoint of a linear operator, 112
Adjoint system, 112114
Adjugate matrix, 44
Aerospace vehicle, state equations of, 34
Algebraic
equation for timeinvariant 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)
128146
Basis, 47
Bessel's equation, 107
Dead beat
control, 132
Decoupling. See Pole placement of
multipleinput 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, 99100
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
Bounded
function, 191
Canonical
equations of Hamilton, 211
flow diagrams (see Flow diagrams, first
canonical form; Flow diagrams,
second canonical form)
Cauchy integral representation, 83
Causality, 26
CayleyHamilton theorem, 81
Characteristic polynomial, 69
Check on solution for transition matrices, 114
Closed
Column of a matrix, 38
Commutativity of matrices, 40
Compatible, 3940
Compensator
pole placement, 172
integral plus proportional, 219223
233
INDEX
234
Exponential stability.
See Stability, uniform asymptotic
Exterior product, 56
External stability, 194
nonnegative
Hidden
44
minimum
stability, 193
Gradient operator, 96
Gram matrix, 64
oscillations in
sampled data
Inverse
matrix, 44
A),
computation
of,
102
system, 166
Jaeobian matrix, 9
Jordan block, 74
Jordan form
flow diagram of, 2124
of a matrix, 7375
of state equations, 147
77
systems, 132
of (si
definite,
positive definite, 77
Kronecker
delta, 44
Linear
dependence, 47
independence, 47
manifold, 46
network with a switch, 11, 12
INDEX
Linear
(cont.)
235
Nonexistence
system, 67
vector space, 50
Linear operator, 55
matrix representation
null space of, 49
range of, 49
Linearization, 8, 9
55
Logarithm of a matrix, 94
Loss matrix. See Criterion functional
Norm, 51
of a matrix, 78
natural, 53
Normal matrix, 41
Normalized eigenvector, 70
eigenvalue of, 69
eigenvector of, 69
equality, 39
function of, 7984
fundamental, 99111
generalized eigenvector of, 74
Hermitian, 107
Hermitian nonnegative definite, 77
Hermitian positive definite, 77
integration of, 39
Jordan form
logarithm
of,
of,
7375
94
multiplication, 39
Null space, 49
Nullity, 49
nvectors, 38
Nyquist diagrams, 171
norm
of, 78
principal minor of, 77
rank of, 50
representation of a linear operator, 55
Riccati equation, 211231
square, 38
state equations, 26, 27
transition, 99111
unitary, 41
Minimum
energy control (Problem 10.30), 229
norm solution of a matrix equation, 86
Minor, 43
principal, 77
Motor equations of
state, 34
Multilinear, 56
Phase plane, 11
Phase variable canonical form, 20, 21
for timevarying systems, 1534
Physical object, 1
Piecewise timeinvariant system, 106
Polar decomposition of a matrix, 87
Pole placement, 172173, 220, 224
Poles of a system, 169
Polezero cancellation, 134
Positive definite, 77
Predictor, 5
Principal minor, 77
Pseudoinverse, 8487
Natural norm, 53
of a matrix, 79
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, 102103
Response function. See Trajectory
Riccati equation, 212213
negative solution to, 221
Right inverse, 44
RLC circuit, 2, 6, 11, 35
stability of,
Root
205207
locus, 169
Row of matrix,
38
Servomechanism, 208
problem, 218220
State (cont.)
reachable, 135
recoverable, 135
State equations
in matrix form, 26, 27
solutions of, with input, 109110
solutions of, with nonzero input, 99109
Steady state
errors, 165
response, 125
SturmLiouville equation, 197
stability properties of, 198
Submatrices, 40
Sufficient conditions for optimal control, 220
Summer,
16,
164
Superposition, 7
integral, 109
Symmetric matrix, 41
Symmetry
of eigenvalues of H, 215
of solution to matrix Riccati equation, 213
Symplectic, 228
Simultaneous diagonalization
of two Hermitian matrices, 91
of two arbitrary matrices, 118
Skewsymmetric matrix, 41
Systems, 7, 8
dynamical, 5
nonanticipative, 5
physical, 1
Span, 46
to go, 228
Spectral representation, 80
Time
Springmass system.
See State variable of springmass system
Square root of a matrix, 77
Timeinvariant, 7
optimal systems, 213217
Stability
asymptotic, 194
BIBO, 194
external, 194
i.s.L., 192
in the large, 193
of optimal control, 214
of solution of timeinvariant
Riccati equation, 215
of a system, 192
uniform, 193
uniform asymptotic, 193
State
Timevarying systems
flow diagram of, 18
matrix state equations
of,
2526
Trace, 39
of similar matrices, 72
Trajectory, 4, 5
Transfer function, 112
Transition
matrix, 99111
property, 5
Transpose matrix, 41
Transposition, 41
Triangle inequality, 51
Triangular matrix, 41
of a physical object, 1
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
237
Unity feedback systems, 165
Unobservable, 3
Unstable, 192
Upper triangular matrix, 41
Van
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
Unit
Wave
matrix, 41
vector, 41
Unitary matrix, 41
for similarity transformation of a
Hermitian matrix, 76
Unittime delayor. See Delayor
equation, 141
Zero matrix, 41
Zeroinput stability, 191
z transfer function, 18
Molto più che documenti.
Scopri tutto ciò che Scribd ha da offrire, inclusi libri e audiolibri dei maggiori editori.
Annulla in qualsiasi momento.