LINEAR CONTROL THEORY: Structures, Robustness, and Design

LINEAR CONTROL
THEORY:
Structures, Robustness, and
Design
L. H. Keel
Tennessee State University
S. P. Bhattacharyya
Texas A&M University
PART I
LINEAR SYSTEM
THEORY
2 Contents
Chapter 1
SYSTEM
REPRESENTATION USING
STATE VARIABLES
This chapter introduces the state space representation of linear time-invariant
dynamic systems. The state spacce model is the foundation of modern optimal
control and ltering theory and has various applications in signal processing,
circuit and control theory. in this course.
1.1 INTRODUCTION
A dynamic system is a system which processes input signals and produces the
corresponding outputs. Thus, dynamic systems include not only engineering
systems, but a much wider class of systems such as population, weather forecast
systems, social and economic systems etc. If all equations describing a system
are dened for all time, the system is called a continuous system. In discrete-
time systems, some equations and variables are dened or used only at discrete
points in time. Continuous-time systems are described by dierential equations
while discrete-time systems are described by dierence equations. Furthermore,
if the input-output relationship of a dynamic system obeys the superposition
principle, such a system is called linear dynamic system. In other words, let y
i
be the corresponding output of the linear dynamic system to the input x
i
. Then
a dynamic system is linear if and only if the output of the system to the input
a
i
x
i
is

a
i
y
i
for arbitrary scalar values of a
i
. It is important to note that
the mathematical analysis of linear dynamic systems always results in linear
dierential equations.
A system is called memoryless if its output at time t
, y (t
), depends only
on the input applied at time t
, u(t
). An example of a memoryless system is

an electrical circuit consisting of resistors only. Most systems possess memory.
The instantaneous output y (t
) of a system with memory depends on its present

inputs, u(t
), as well as its past and/or future inputs. If the output y (t
)
depends only on past and present inputs, that is, u(t), t t
, the system is
3
4 SYSTEM REPRESENTATION USING STATE VARIABLES Ch. 1
called causal. State variables of a system summarize the internal status of the
system. Specically, state x(t
0
) is the information at time t
0
that determines
the entire output y(t) for all t > t
0
if the inputs are known.
1.2 STATE SPACE REPRESENTATION OF LTI SYSTEMS
Let y(t) be the output of a system with initial state x(t
0
) to the input u(t),
t t
0
. Suppose that the initial state x(t
0
) is shifted to time t(t
0
+ T) and the
same input is applied from time t
0
+ T instead of t
0
. If the new output is the
exact t
0
+T shifted version of y(t), then such a system is called a time-invariant
system. Simply speaking, if the initial state and the input are the same, the
output is always the same regardless of the time when the input is applied.
Typically, the input-output relationship of a linear time-invariant (LTI) sys-
tem is written as dierential equations, in most cases, of high order. The state
space equation is a rst order vector dierential equation which is equivalent to
the high order dierential equation. We show how to obtain state space equa-
tions from the dierential equations considering, rst, the case when there is
no derivative terms in the input. Let us consider the following time-invariant
system where u(t) is input and y(t) is output.
d
n
y(t)
dt
n
+a
n1
d
n1
y(t)
dt
n1
+ +a
1
dy(t)
dt
+a
0
= u(t). (1.1)
The dene a set of new variables as follows.
x
1
(t) := y(t)
x
2
(t) := x
1
(t) = y(t)
.
.
.
x
n
(t) := x
n1
(t) = y
(n1)
.
We now rewrite eq. (1.1) by using these variables.
x
n
(t) +a
n1
x
n
(t) +a
n2
x
n1
(t) + +a
1
x
2
(t) +a
0
x
1
(t) = u(t) (1.2)
or
dotx
n
(t) = a
n1
x
n
(t) a
n2
x
n1
(t) a
1
x
2
(t) a
0
x
1
(t) +u(t). (1.3)
This expression can be put into matrix form.
_
_
x
1
(t)
x
2
(t)
x
3
(t)
.
.
.
x
n
(t)
_
_
. .
x(t)
=
_
_
0 1 0 0
0 0 1 0
.
.
.
.
.
.
0 0 0 1
a
0
a
1
a
2
a
n1
_
_
. .
A
_
_
x
1
(t)
x
2
(t)
x
3
(t)
.
.
.
x
n
(t)
_
_
. .
x(t)
+
_
_
0
0
.
.
.
0
1
_
_
. .
B
u(t)
Sec. 1.2. STATE SPACE REPRESENTATION OF LTI SYSTEMS 5
(1.4)
y(t) =
_
1 0 0 0

. .
C
_
_
x
1
(t)
x
2
(t)
x
3
(t)
.
.
.
x
n
(t)
_
_
. .
x(t)
.
The above is called a state space representation of the system and x(t) is called
state vector. The state of a system at t
is the amount of information at t
that,
together with u
[t
,)
, that determines uniquely the behaviour of the system for
all t t
.
The general state space model of an LTI system is of the form
x(t) = Ax(t) +Bu(t)
y = Cx(t) +Du(t).
where A, B.C.D are matrices of compatible order with A being nxn and n is
the dimension of the state vector or the order of the system. To illustate, we
give some examples.
Example 1.1. Consider an electrical circuit shown in Figure 1.1.
y
+
x
1
x
2
+
x
3 L
C2
C
1
R
u(t)
+
i
Figure 1.1. An electrical circuit.
Choosing the inductor current and capacitor voltages as states and using
Kirchos laws, we have
u(t) = Ri(t) +x
1
(t), i(t) = C
1
x
1
(t) +C
2
x
2
(t), x
3
(t) = C
2
x
2
(t).
Thus,
u(t) = R(C
1
x
1
(t) +x
3
(t)) +x
1
(t)
or
x
1
(t) =
x
1
(t)
RC
1
x
3
(t)
C
1
+
u(t)
RC
1
.
and
x
2
(t) =
x
3
(t)
C
2
, x
3
(t) =
x
1
(t)
L

x
2
L
.
We write this in matrix form.
_
_
x
1
(t)
x
2
(t)
x
3
(t)
_
_
. .
x
=
_
_
1
R1C1
0
1
C1
0 0
1
C
2
1
L

1
L
0
_
_
. .
A
_
_
x
1
(t)
x
2
(t)
x
3
(t)
_
_
. .
x
+
_
_
1
RC
1
0
0
_
_
. .
B
u(t).
Since y = L x
3
(t), we write
y = L x
3
(t) = x
1
(t) x
2
(t)
= [1 1 0]
. .
C
_
_
x
1
(t)
x
2
(t)
x
3
(t)
_
_
.
This is the case of a single-input single-output (SISO) system.
Example 1.2. Let us consider the following electromechanical system shown
in Figure 1.2.
+
u
L
a
R
a
= y
B
TL
+
e
b
ia
Figure 1.2. An electromechanical system.
It represents a separately excited dc motor driving a load producing a load
torque T
L
. We denote
R
a
armature-winding resistance,
L
a
armature-winding inductance, H
i
a
armature-winding current, A
u applied armature voltage, V
e
b
back emf, V
angular displacement of motor shaft, radian
B equivalent viscous-friction coecient of the motor and load
referred to the motor shaft, ib-ft/radian/sec
J equivalent moment of inertia of the motor and load referred to
the motor shaft, slug-ft
2
K
b
back emf constnat
K
i
motor-torque constant
and suppose their numerical values are
R
a
= 0.1, L
a
= 0.01, K
b
= 5, K
t
= 6, J = 10, B = 2, K = 1
The system is represented by the following equations.
u = R
a
i
a
+L
a
di
a
dt
+K
b

K
t
i
a
= K +B

+J
+T
L
.
To write state equations, let x
1
= , x
2
=

, x
3
= i
a
, then
x
1
= x
2
x
2
=
K
J
x
1

B
J
x
2

K
t
J
x
3
+
1
J
T
L
x
3
=
K
b
L
a
x
2

R
a
L
a
x
3
+
1
L
a
u
_
_
x
1
x
2
x
3
_
_
=
_
_
0 1 0
K
J

B
J

K
t
J
0
K
b
L
a
Ra
L
a
_
_
. .
A
_
_
x
1
x
2
x
3
_
_
+
_
_
0
0
1
La
_
_
. .
B
u +
_
_
0
1
J
0
_
_
. .
E
T
L
y =
_
1 0 0
_
_
x
1
x
2
x
3
_
_
+
_
0
u +
_
0
T
L
and
A =
_
_
0 1 0
1
10

1
5

6
10
0 500 10
_
_
, B =
_
_
0
0
100
_
_
, E =
_
_
0
1
10
0
_
_
Example 1.3.
Assume that the cart and the pendulum move in only one plane; the friction
and the mass of the rod are negligible. Using the Lagrange Equation of Motion,
d
dt
L
x
i
L
x
i
= Q
i
where L = T P and
T : kinematic energy
P : potential energy
Q
i
: external force
x
i
: generalized coordinates
-
y
u
-
M
6
V
-
H
l
m
?
mg
Figure 1.3. A cart with an inverted pendulum (Example 1.3)

For this case, the kinetic energy of the system is
T =
1
2
M( y)
2
+
1
2
m
_
_
d(y +l sin)
dt
_
2
+
_
d(l cos )
dt
_
2
_
.
To linearize these equations, we let, for small ,
sin
cos = 1
2
2!
+

4
4!

. .
0
1.
Then
T
1
2
M( y)
2
+
1
2
m
_
_
d(y +l)
dt
_
2
+
_
d(l)
dt
_
2
_
=
1
2
M( y)
2
+
1
2
m
_
y +l
_
2
.
The potential energy of the system is
P = mgl cos .
Thus, we have
L = T P
=
1
2
M( y)
2
+
1
2
m( y +l
)
2
mgl cos .
For the y coordinate,
d
dt
L
y

L
y
= u
d
dt
T
y
= u
d
dt
_
M y +m( y +l
)
_
= u
(M +m) y +ml
= u. (1.5)
For the coordinate,
d
dt
L

L
= 0 (no external force)

d
dt
_
ml( y +l
)
_
mgl sin
..
= 0
y +l
g = 0. (1.6)
Now taking Laplace transform of (1.5) and (1.6), we have
(M +m)s
2
Y (s) +mls
2
(s) = U(s)
s
2
Y (s) + (ls
2
g)(s) = 0. (1.7)
If we consider two outputs, y and , the transfer function matrix of the system
becomes
_
Y (s)
(s)
_
=
_
_
ls
2
g
s
2
[Mls
2
(M +m)g]
1
Mls
2
g(M +m)
_
_
. .
G(s)
U(s). (1.8)
Recall (1.5) and (1.6), and dene state variables,
_
y y

:=
_
x
1
x
2
x
3
x
4
and the output [y, ]

T
:= [y
1
, y
2
]
T
. Then we have the following state space
description of the system.
_
_
x
1
x
2
x
3
x
4
_
_
=
_
_
0 1 0 0
0 0
mg
M
0
0 0 0 1
0 0
(M+m)g
Ml
0
_
_
. .
A
_
_
x
1
x
2
x
3
x
4
_
_
+
_
_
0
1
M
0
1
Ml
_
_
. .
B
u
_
y
1
y
2
_
=
_
1 0 0 0
0 0 1 0
_
. .
C
_
_
x
1
x
2
x
3
x
4
_
_
.
Example 1.4.
Write Kirchos equations:
v +R
1
x
1
+L
1
x
1
+x
3
+R
3
(x
1
x
2
) = 0
R
3
(x
1
x
2
) x
3
+L
2
x
2
+R
2
x
2
= 0
x
1
x
2
= C
1
x
3
.
R
1
L
1
C
1
R
3
L
2
R
2
- -
6
?
6
?
+
v
x
1
x
3
x
2
y
? ?
Figure 1.4. A circuit diagram (Example 1.4)
Then a state space representation of the above circuit using the predetermined
state variables (x
1
, x
2
, x
3
, x
4
) is
_
_
x
1
x
2
x
3
_
_
=
_
R
1
+R
3
L
1
R
3
L
1
1
L
1
R
3
L
2
R
2
R
3
L
2
1
L
2
1
C
1
1
C
1
0
_
_
_
_
x
1
x
2
x
3
_
_
+
_
_
1
L1
0
0
_
_
v
y =
_
0 R
2
0

_
_
x
1
x
2
x
3
_
_
Example 1.5. Consider the mass-spring-dashpot system shown in Figure 1.5
which illustrates a two input two output system.
M
B
u2
k
1
k2
m
u
1
x
1
x2
Figure 1.5. A mass-spring-dashpot system
We write Newtons equations for the dynamics of the system:
u
1
= = m x
1
+k
1
x
1
+k
2
(x
1
x
2
)
u
2
= M x
2
+B x
2
+k
2
(x
2
x
1
)
or
x
1
=
k
1
+k
2
m
x
1
+
k
2
m
x
2
+
1
m
u
1
x
2
=
k
2
M
x
1

k2
M
x
2

B
M
x
2
+
1
M
u
2
.
Let us dene z
1
:= x
1
, z
2
:= x
1
, z
3
:= x
2
, z
4
:= x
2
, then
z
1
= z
2
= x
1
z
2
= x
1
=
k
1
+k
2
m
z
1
+
k
2
m
z
3
+
1
m
u
1
z
3
= z
4
= x
2
z
4
= x
2
=
k
2
M
z
1

k
2
M
z
3

B
M
z
4
+
1
M
u
2
.
Thus, we have
z =
_
_
0 1 0 0
k
1
+k
2
m
0
k
2
m
0
0 0 0 1
k
2
M
0
k
2
M

B
M
_
_
_
_
z
1
z
2
z
3
z
4
_
_
+
_
_
0 0
1
m
0
0 0
0
1
M
_
_
_
u
1
u
2
_
.
The output equation becomes
_
y
1
y
2
_
=
_
x
1
x
2
_
=
_
z
1
z
3
_
= [1 0 1 0]z(t).
As shown the above example, state space representations for both SISO and
multiinput-multioutput (MIMO) systems have the same form. For a system
with n states, r inputs, and m outputs as shown in Figure 1.6.
A state space representation of such a system is as follows.
x(t) = Ax(t) +Bu(t) (1.9)
y(t) = Cx(t) +Du(t)
where
u(t) :=
_
_
u
1
(t)
u
2
(t)
.
.
.
u
r
(t)
_
_
and y(t) :=
_
_
y
1
(t)
y
2
(t)
.
.
.
y
m
(t)
_
_
.
-
-
-
-
-
-
u
1
u
2
u
m
y
1
y
2
y
r
Figure 1.6. Inputs - Outputs of a system
Consequently, with n states,
A IR
nn
, B IR
np
, C IR
qn
, D IR
qp
.
In shorthand, we call (A, B, C, D) is a state space description of a LTI system.
The above examples show how to obtain a state space representation from
the dynamic equations of the system. Another way to describe a system is by
using the transfer function. The latter is the ratio of Laplace transforms of
input and the corresponding output with zero initial conditions. Alternatively
it is simply Laplace transform of the response of the system to a unit impulse,
again, with zero initial conditions. Here, we give two examples of obtaining a
state space representation (A, B, C, D) from the given transfer function.
Example 1.6. (Transfer functions without zeros)
Consider
G(s) =
Y (s)
U(s)
=
b
0
s
2
+a
1
s +a
0
.
Write
s
2
Y (s) +a
1
sY (s) +a
0
Y (s) = b
0
U(s).
We have the equivalent dierential equation
y +a
1
ydt +a
0
y = b
0
u.
Dene
x
1
:= y,
x
2
:= x
1
= y,
then
x
2
= a
1
x
2
a
0
x
1
+b
0
u.
Thus, we have the state space representation
x =
_
0 1
a
0
a
1
_
x +
_
0
b
0
_
u
y = [1 0]x.
Sec. 1.3. SOLUTION OF STATE SPACE EQUATIONS 13
Example 1.7. (Transfer functions with zeros)
Consider
G(s) =
Y (s)
U(s)
=
b
1
s +b
0
s
2
+a
1
s +a
0
.
The equivalent dierential equation is
(y) +a
1
y +a
0
y = b
1
u +b
0
u.
Dene
x
1
:= y,
x
2
:= x
1
b
1
u = y b
1
u,
then we have
x
1
= x
2
+b
1
u
and
x
2
= y b
1
u = a
1
y a
0
y +b
0
u
= a
0
y a
1
( y b
1
u) a
1
b
1
u +b
0
u
= a
0
x
1
a
1
x
2
+ (b
0
a
1
b
1
) u.
Thus,
x =
_
0 1
a
0
a
1
_
x +
_
b
1
b
0
a
1
b
1
_
u
y = [1 0]x.
The state space representations obtained in the above two examples allows
us to implement the given transfer function by using integrators. This is the
subject that we will study in detail in the next chapter.
1.3 SOLUTION OF STATE SPACE EQUATIONS
1.3.1 Solution of State Equations
Let us consider a scalar example:
x(t) = ax(t) +bu(t) (1.10)
where a, b, c are scalar constants. To solve the dierential equation, we take
Laplace transforms, and get
sX(s) x(0) = aX(s) +bU(s), or X(s) =
1
s a
x(0) +
b
s a
U(s). (1.11)
Taking inverse Laplace transforms, we have
x(t) = e
at
x(0) +
_
t
0
e
a(t)
bu()d. (1.12)
Eq. (1.12) shows that the solution consists of two parts. The rst part is only
depending on the initial condition x(0) and the second part is only depending
on the input u(s). Therefore the rst part is called the zero input response and
the second part is called the zero state response.
The foregoing approach to the solution of the scalar dierential equation can
be extended to the case of the vector state equations. Let
x(t) = Ax(t) +Bu(t) (1.13)
where x(t) is n vector, so A IR
nn
. If a system has r inputs and m outputs,
B IR
nr
, C IR
mn
, D IR
mr
. Taking the Laplace transform of both sides
of the equation, we have
sX(s) x(0) = AX(s) +BU(s)
where X(s) = L[x(t)], U(s) = L[u(t)] and x(t) is a vector with size n. Hence
(sI A)X(s) = x(0) +BU(s)
or
X(s) = (sI A)
1
x(0) + (sI A)
1
BU(s)
Taking inverse Laplace transform
x(t) = L
1
_
(sI A)
1
x(0) +L
1
_
(sI A)
1
u(t).
Note that
(sI A)
1
=
I
s
+
A
s
2
+
A
2
s
3
+ (1.14)
Thus, we have
L
1
_
(sI A)
1
= I +At +
A
2
t
2
2!
+
A
3
t
3
3!
+
= e
At
. (1.15)
Therefore, the solution of the state equation is
x(t) = e
At
x(0) +
_
t
0
e
A(t)
Bu()d
1.3.2 State Transition Matrix
In the expression of the solution x(t) of the state equation, the term e
At
is called
state transition matrix and is commonly notated as
(t) := e
At
. (1.16)
Let us examine some properties of the state transition matrix that will be used
in later chapters. Assume zero input, u(t) = 0. Then the solution of the system
x(t) = Ax(t) (1.17)
Sec. 1.3. SOLUTION OF STATE SPACE EQUATIONS 15
becomes
x(t) = e
At
x(0) = (t)x(0). (1.18)
At t = 0, we know (0) = I. Dierentiating eq. (1.18), we have
x(t) =

(t)x(0) = Ax(t). (1.19)
At t = 0,
x(0) =

(0)x(0) = Ax(0) (1.20)
leads us to have
(0) = A. (1.21)
Therefore, the state transition matrix (t) has the following properties:
(0) = I and

(0) = A. (1.22)
In the following, we summarize the property of the state transition matrix.
(a) (0) = e
A0
= I
(b) (t) = e
At
=
_
e
At
_
1
=
1
(t)
(c)
1
(t) = (t)
(d) (a +b) = e
A(a+b)
= e
Aa
e
Ab
= (a)(b)
(e)
n
(t) =
_
e
At
_
n
= e
Ant
= (nt)
(f) (a b)(b c) = (a)(b)(b)(c) = (a)(c) = (a c)
If
A =
_
2
.
.
.
n
_
_
,
e
At
=
_
_
e
1
t
e
2
t
.
.
.
e
nt
_
_
.
If
A =
_
_
1 0 0
0 1 0
0 0 1
0 0 0
1
0
_
_
,
e
At
=
_
_
e
t
te
t 1
2
t
2
e
t 1
6
t
3
e
t
0 e
t
te
t 1
2
t
2
e
t
0 0 e
t
te
t
0 0 0 e
t
e
t
te
0 e
t
_
_
,
The state transition matrix contains all the information about the free motions
of the system in eq. (1.17). However, computing the state transition matrix is
in general not easy. To develop some approaches to computation, we rst study
functions of a square matrix.
1.4 FUNCTIONS OF A SQUARE MATRIX
1.4.1 Transfer Functions from State Space Models
Consider a state space model:
x(t) = Ax(t) +Bu(t)
y(t) = Cx(t) +Du(t).
Taking the Laplace transform, we have
sX(s) x(0) = AX(s) +BU(s)
Y (s) = CX(s) +DU(s)
and
Y (s) = C
_
(sI A)
1
BU(s) + (sI A)
1
x(0)
+DU(s)
=
_
C(sI A)
1
B +D
U(s) +C(sI A)
1
x(0).
The transfer function G(s) is obtained by letting the initial condition x(0) = 0.
Y (s) =
_
C(sI A)
1
B +D
. .
G(s)
U(s). (1.23)
This shows the relationship between a state space representation of the system
and its corresponding transfer function. The following example illustrates how
to compute the transfer function from the given state space representation of
the system.
Example 1.8. Consider a 3rd order system with 2 inputs and outputs.
x(t) =
_
_
1 2 0
1 4 1
0 0 1
_
_
x(t) +
_
_
1 0
0 1
1 0
_
_
u(t)
y(t) =
_
0 1 0
1 0 1
_
x(t) +
_
1 0
0 0
_
u(t)
Sec. 1.4. FUNCTIONS OF A SQUARE MATRIX 17
To compute the transfer function,
(sI A)
1
=
_
_
s + 4
s
2
+ 5s + 6
2
s
2
+ 5s + 6
2
s
2
+ 5s + 6
1
s
2
+ 5s + 6
s + 1
s
2
+ 5s + 6

1
s
2
+ 5s + 6
0 0
1
s + 1
_
_
=
_
_
2
s + 2

1
s + 3
2
s + 2

2
s + 3
1
s + 1

2
s + 2
+
1
s + 3
1
s + 2
+
1
s + 3

1
s + 2
+
2
s + 3

1
s + 2
+
1
s + 3
0 0
1
s + 1
_
_
.
Therefore, the state transition matrix is
e
At
=
_
_
2e
2t
e
3t
2e
2t
2e
3t
e
t
2e
2t
+e
3t
e
2t
+e
3t
e
2t
+ 2e
3t
e
2t
+e
3t
0 0 e
t
_
_
.
The transfer function is
G(s) = C(sI A)
1
B +D =
_
2
s + 2
+
2
s + 3
+ 1
1
s + 2
+
2
s + 3
2
s + 1
2
s + 2

2
s + 3
_
_
To solve for x(t) with
x(0) =
_
_
1
0
1
_
_
and u(t) =
_
U(t)
0
_
,
we have
X(s) = (sI A)
1
x(0) + (sI A)
1
BU(s)
=
_
_
1
s + 1
2
s + 2
+
2
s + 3
1
s + 1
_
_
+
_
_
1
s

1
s + 1
1
s
+
1
s + 2
+
2
3
s

2
3
s + 3
1
s

1
s + 1
_
_
so that
x(t) =
_
_
e
t
+ 1 e
t
2e
2t
+ 2e
3t
1
3
+e
2t
2
3
e
3t
e
t
+ 1 e
t
_
_
=
_
_
1
1
3
e
2t
+
4
3
e
3t
1
_
_
.
1.4.2 Computation of (sI A)
1
(sI A)
1
is the main calculation in evaluating the state vector x(t), or the
transfer function, starting from a state space representation. If the formula
(sI A)
1
=
Adj[sI A]
det[sI A]
(1.24)
is used directly, symbolic calculations are involved. Leverriers algorithm allows
us to calculate (sI A)
1
using purely numerical computations.
Leverriers Algorithm
This algorithm determines (sI A)
1
without symbolic calculation. First write
(sI A)
1
=
Adj[sI A]
det[sI A]
(1.25)
and let
a(s) = det[sI A] = s
n
+a
n1
s
n1
+ +a
1
s +a
0
T(s) = Adj[sI A] = T
n1
s
n1
+T
n2
s
n2
+ +T
1
s +T
0
, T
i
IR
nn
.
Thus, we have
(sI A)
1
=
1
a(s)
T(s) or a(s)I = (sI A)T(s) (1.26)
so that
a(s)I = Is
n
+a
n1
Is
n1
+ +a
1
Is +a
0
I
(sI A)T(s) = T
n1
s
n
+ (T
n2
AT
n1
) s
n1
+ + (T
0
AT
1
) s AT
0
.
Therefore,
T
n1
= I, T
n2
= AT
n1
+a
n1
I, T
0
= AT
1
+a
1
I, AT
0
= a
0
I (1.27)
To nd the coecients a
i
, introduce the trace function of a matrix
Trace[A] =
diagonal terms. (1.28)

It can be proved that
a
i
=
1
n i
Trace [AT
ni
] . (1.29)
Leverrier
s Algorithm:
T
n1
= I a
n1
= Trace [AT
n1
]
T
n2
= AT
n1
+a
n1
I a
n2
=
1
2
Trace [AT
n2
]
T
n3
= AT
n2
+a
n2
I a
n3
=
1
3
Trace [AT
n3
]
.
.
.
T
0
= AT
1
+a
1
I a
0
=
1
n
Trace [AT
0
]
(1.30)
Example 1.9.
A =
_
_
2 1 1 2
0 1 1 0
1 1 1 1
1 1 1 0
_
_
Let (sI A)
1
=
T
3
s
3
+T
2
s
2
+T
1
s +T
0
s
4
+a
3
s
3
+a
2
s
2
+a
1
s +a
0
T
3
= I a
3
= Trace[A] = 4
T
2
= AT
3
+a
3
I = A4I =
_
_
2 1 1 2
0 3 1 0
1 1 3 1
1 1 1 4
_
_
a
2
=
1
2
Trace [AT
2
] = 2
T
1
= AT
2
+a
2
I =
_
_
1 4 0 3
1 0 2 1
2 0 0 5
3 3 1 5
_
_
a
1
=
1
3
Trace [AT
1
] = 5
T
0
= AT
1
+a
1
=
_
_
0 2 0 2
1 5 2 4
1 7 2 4
0 4 2 2
_
_
a
0
=
1
4
Trace [AT
0
] = 2
Therefore,
det[sI A] = s
4
4s
3
+ 2s
2
+ 5s + 2
Adj[sI A] = Is
3
+
_
_
2 1 1 2
0 3 1 0
1 1 3 1
1 1 1 4
_
_
s
2
+
_
_
1 4 0 3
1 0 2 1
2 0 0 5
3 3 1 5
_
_
s +
_
_
0 2 0 2
1 5 2 4
1 7 2 4
0 4 2 2
_
_
and
(sI A)
1
=
1
s
4
4s
3
+ 2s
2
+ 5s + 2
_
_
s
3
2s
2
s s
2
+ 4s + 2 s
2
2s
2
3s 2
s + 1 s
3
3s
2
+ 5 s
2
2s 2 s 4
s
2
+ 2s 1 s
2
7 s
3
3s
2
+ 2 s
2
5s + 4
s
2
3s 2 s
2
3s + 4 s
2
s 2 s
3
4s
2
+ 5s 2
_
_
1.4.3 Cayley-Hamilton Theorem
An eigenvalue of the n n real matrix is a real or complex number such that
Ax = x with a nonzero vector x. Any nonzero vector x satisfying Ax = x is
called an eigenvector of A associated with eigenvalue . To nd the eigenvalues
of A, we solve
(AI)x = 0 (1.31)
In order for eq. (1.31) to have a nonzero solution x, the matrix (A I) must
be singular. Thus, the eigenvalues of A are just the solutions of the following
n
th
order polynomial equation:
() = det(I A) = 0. (1.32)
() is called the characteristic polynomial of A.
Theorem 1.1 (Cayley-Hamilton Theorem)
Let A be a n n matrix and let
() = det(I A) =
n
+a
n1
n1
+ +a
1
+a
0
be the characteristic polynomial of A. Then
(A) = A
n
+a
n1
A
n1
+ +a
1
A+a
0
I = 0.
Proof. (sI A)
1
can always be written asWcome
(sI A)
1
=
1
(s)
_
R
n1
s
n1
+R
n2
s
n2
+ +R
1
s +R
0
_
where
(s) = det(sI A) = s
n
+a
n1
s
n1
+ +a
1
s +a
0
and R
i
, i = 0, 1, , n 1 are constant matrices formed from the adjoint of
(sI A). Write
(s)I = (sI A)
_
R
n1
s
n1
+R
n2
s
n2
+ +R
1
s +R
0
_
= (sI A)R
n1
s
n1
+ (sI A)R
n2
s
n2
+
+(sI A)R
1
s + (sI A)R
0
= R
n1
s
n
+ (R
n2
AR
n1
) s
n1
+ (R
n3
AR
n2
) s
n2
+
+(R
0
AR
1
) s AR
0
.
Since
(s)I = Is
n
+a
n1
Is
n1
+ +a
1
Is +a
0
I,
we have by matching the coecients,
R
n1
= I
R
n2
= AR
n1
+a
n1
I
R
n3
= AR
n2
+a
n2
I
.
.
.
R
0
= AR
1
+a
1
I
0 = AR
0
+a
0
I.
Substituting R
i
s successively from the bottom, we have
0 = AR
0
+a
0
I
= A
2
R
1
+a
1
A+a
0
I
= A
3
R
2
+a
2
A
2
+a
1
A+a
0
I
.
.
.
= A
n
+a
n1
A
n1
+a
n2
A
n2
+ +a
1
A+a
0
I = (A).
Therefore, (A) = 0.
1.4.4 Functions of a Square Matrix
Theorem 1.2 Let A be a n n square matrix with characteristic polynomial
() =
m
i=1
(
i
)
n
i
where n =
m
i=1
n
i
. Let f() be any function and h() be a polynomial of degree
n 1
h() =
n1
n1
+
n2
n2
+ +
1
+
0
.
then
f(A) = h(A)
if the coecients,
i
s, of h() are chosen such that
d
l
f()
d
l
=i
=
d
l
h()
d
l
=i
, for l = 0, 1, , n
i
1; i = 1, 2, , m.
Example 1.10.
A =
_
_
0 2 2
0 1 0
1 1 3
_
_
To compute e
At
, we let f() = e
t
and
h() =
2
2
+
1
+
0
.
Since the eigenvalues of A are {1, 1, 2}, we evaluate the following.
f(1) = h(1) e
t
=
2
+
1
+
0
f
(1) = h
(1) te
t
= 2
2
+
1
f(2) = f(2) e
2t
= 4
2
+ 2
1
+
0
.
From these, we have
0
= 2te
t
+e
2t
1
= 3te
t
+ 2e
t
2e
2t
2
= e
2t
e
t
te
t
and
h() =
_
e
2t
e
t
te
t
_
2
+
_
3te
t
+ 2e
t
2e
2t
_
2te
t
+e
2t
.
Therefore,
e
At
= f(A) = h(A)
=
_
e
2t
e
t
te
t
_
A
2
+
_
3te
t
+ 2e
t
2e
2t
_
A+
_
2te
t
+e
2t
_
I
=
_
_
e
2t
+ 2e
t
2te
t
2e
2t
+ 2e
t
0 e
t
0
e
2t
e
t
te
t
2e
2t
e
t
_
_
.
1.4.5 Internal Stability of a System
Internal stability refers to the stability of a state space representation of a sys-
tem. Let
x(t) = Ax(t) +Bu(t)
The representation is internally stable if the zero input solution for the state
vector
x(t) = e
At
x(t
0
), for t t
0
,
tends toward zero as t goes to for arbitrary x(t
0
). It is easy to see that the
representation is internally stable if and only if real parts of all eigenvalues of
A are negative, since it is these exponentials corresponding to the eigenvalues
that appear in the state transition matrix.
1.5 STRUCTURE OF THE OPERATOR A
1.5.1 Similarity Transformation
The state variables describing a system are not unique and the state space
representation of a given system depends on the choice of state variables. By
choosing dierent bases for the underlying n ddimensional state space we ob-
tain dierent nth order representations of the same system. Such systems are
called similar. Similar systems have the same transfer function although their
state space representations are dierent. This fact can be exploited to develop
equivalent state space representations consists of matrices that display various
structural properties or simplify computation.
Now consider a state space representation
x(t) = Ax(t) +Nu(t)
Dene x(t) = Tz(t) where T is a square invertible matrix. Then
T z(t) = ATz(t) +Bu(t)
y(t) = CTz(t) +Du(t).
Sec. 1.5. STRUCTURE OF THE OPERATOR A 23
or
_
z(t) = T
1
ATz(t) +T
1
Bu(t)
y(t) = CTz(t) +Du(t)

_
z = A
n
z(t) +B
n
u(t)
y(t) = C
n
z(t) +Du(t).
The transfer function of the system may be computed
CT
_
sI T
1
AT
_
T
1
B +D = CT
_
T
1
(sI A)T
1
T
1
B +D
= CT
_
T
1
(sI A)
1
T
T
1
B +D
= CTT
1
(sI A)
1
TT
1
B +D
= C(sI A)
1
B +D
which shows that the representations in eqs. (1.33) and (1.33) have the same
transfer function.
1.5.2 Diagonalization of a Matrix
Theorem 1.3 Consider an n n matrix A with distinct eigenvalues
1
,
2
,
,
n
. Let x
i
, i = 1, 2, , n be eigenvectors associated with
i
. Dene a nn
matrix
T = [x
1
x
2
x
3
x
n
] .
Then
T
1
AT =
_
2
.
.
.
n
_
_
.
Proof. Since x
i
is an eigenvector of A associated with
i
, Ax
i
=
i
x
i
. So we
write
Ax
i
=
i
x
i
= [x
1
x
2
x
i
x
n
]
_
_
0
.
.
.
0
i
0
.
.
.
0
_
_
.
Consequently,
A[x
1
x
2
x
i
x
n
]
. .
T
= [x
1
x
2
x
i
x
n
]
. .
T
_
2
.
.
.
n
_
_
.
and
T
1
AT =
_
2
.
.
.
n
_
_
.
1.5.3 Jordan Form of A

We have seen in the previous section that an n n matrix A can be converted
through a similarity transformation into a diagonal matrix when the eigenvalues
of A are distinct. When the eigenvalues of A are not distinct, that is, some
are repeated, it is not always possible to diagonalize A, as before. In this
subsection, we deal with this case, and develop the so-called Jordan form of A
which is the maximally diagonal form that can be obtained.
The characteristic polynomial of the nn real or complex matrix A, denoted
(s), is:
(s) = det(sI A)
= s
n
+a
n1
s
n1
+ +a
1
s +a
0
. (1.33)
Write (1.33) in the factored form
(s) = (s
1
)
n
1
(s
2
)
n
2
+ + (s
p
)
n
p
(1.34)
where the
i
are distinct, that is,
i
=
j
, for i = j
and
n
1
+n
2
+ +n
p
= n.
It is easy to establish the following.
Theorem 1.4 There exists an n n nonsingular matrix T such that
T
1
AT =
_
_
A
1
A
2
.
.
.
A
p
_
_
where A
i
is n
i
n
i
and
det (sI A
i
) = (s
i
)
ni
, for i = 1, 2, , p.
In the remaining part of the section, we develop the Jordan decomposition
of the matrices A
i
. Therefore, we now consider, without loss of generality, an
n n matrix A such that
(s) = det(sI A) = (s )
n
. (1.35)
The Jordan structure of such an A matrix may be found by forming the matrices
(AI)
k
, for k = 0, 1, 2, , n. (1.36)
Let N
k
denote the null space of (AI)
k
:
N
k
=
_
x | (AI)
k
x = 0
_
(1.37)
and denote the dimension of N
k
by
k
:
dimension N
k
=
k
, for k = 0, 1, 2, , n + 1. (1.38)
We remark that
0 =
0

1

n1

n
= n =
n+1
. (1.39)
The following theorem gives the Jordan structure of A.
Theorem 1.5 Given the n n matrix A with
det(sI A) = (s )
n
, (1.40)
there exists an n n matrix T with det(T) = 0, such that
T
1
AT =
_
_
A
1
A
2
.
.
.
A
r
_
_
where
A
j
=
_
_
1 0 0
0 1 0
.
.
.
.
.
.
0 1
0 0 0
_
_
, for j = 1, 2, , r
is called a Jordan block n
j
n
j
.
The number r and sizes n
j
of the Jordan blocks A
i
can be found from the
formula given in Lemma 1.1.
Lemma 1.1 Under the assumption of Theorem 1.5, the number of Jordan
blocks of size k k is given by
2
k

k1

k+1
, for k = 1, 2, , n.
Example 1.11. Let A be an 11 11 matrix with
det(sI A) = (s )
11
.
Suppose that
1
= 6,
2
= 9,
3
= 10,
4
= 11 =
5
=
6
= = v
11
=
12
.
Then the number of Jordan blocks:
a) of size 1 1 = 2
1

0

2
= 3
b) of size 2 2 = 2
2

1

3
= 2
c) of size 3 3 = 2
3

2

4
= 0
d) of size 4 4 = 2
4

3

5
= 1
e) of size 5 5 = 2
5

4

6
= 0
.
.
.
k) of size 11 11 = 2
11

12

10
= 0.
Therefore, the Jordan form of A is as shown below:
A =
_
_
1 0 0
0 1 0
0 0 1
0 0 0
1
0
1
0
_
This result can now be applied to each of the blocks A
i
in Theorem 1.4 to
obtain the complete Jordan decomposition in the general case.
It is best to illustrate the procedure with an example.
Example 1.12. Let A be a 15 15 matrix with
det(sI A) = (s
1
)
8
(s
2
)
4
(s
3
)
3
with
i
being distinct. We compute
(A
i
I)
k
, k = 0, 1, 2, , 8, 9
and set
h
k
= dimN (A
1
I)
k
, k = 0, 1, 2, , 9.
Similarly, let
i
j
= dimN (A
2
I)
j
, j = 0, 1, 2, , 5
and
l
s
= dimN (A
3
I)
s
, s = 0, 1, 2, 3, 4.
Suppose, for example,
h
0
= 0, h
1
= 2, h
2
= 4, h
3
= 6, h
4
= 7, h
5
= h
6
= h
7
= h
8
= 8
i
0
= 0, i
1
= 3, i
2
= 4, i
3
= i
4
= i
5
= 4
l
0
= 0, l
1
= 1, l
2
= 2, l
3
= l
4
= 3.
By Theorem 1.4, the Jordan form of A has the following structure.
T
1
AT =
_
_
A
1
0 0
0 A
2
0
0 0 A
3
_
_
with A
1
IR
88
, A
2
IR
44
and A
3
IR
33
. Furthermore, the detailed
structure of A
1
, by Theorem 1.5, consists of the following numbers of Jordan
blocks.
a) of size 1 1 = 2h
1
h
0
h
2
= 0 blocks
b) of size 2 2 = 2h
2
h
1
h
3
= 0 blocks
c) of size 3 3 = 2h
3
h
2
h
4
= 1 blocks
d) of size 4 4 = 2h
4
h
3
h
5
= 0 blocks
e) of size 5 5 = 2h
5
h
4
h
6
= 1 blocks
f) of size 6 6 = 2h
6
h
5
h
7
= 0 blocks
g) of size 7 7 = 2h
7
h
6
h
8
= 0 blocks
h) of size 8 8 = 2h
8
h
7
h
9
= 0 blocks
Similarly, the numbers of Jordan blocks in A
2
are:
a) of size 1 1 = 2i
1
i
0
i
2
= 2 blocks
b) of size 2 2 = 2i
2
i
1
i
3
= 1 blocks
c) of size 3 3 = 2i
3
i
2
i
4
= 0 blocks
d) of size 4 4 = 2i
4
i
3
i
5
= 0 blocks
and the numbers of Jordan blocks in A
3
are:
a) of size 1 1 = 2l
1
l
0
l
2
= 0 blocks
b) of size 2 2 = 2l
2
l
1
l
3
= 0 blocks
c) of size 3 3 = 2l
3
l
2
l
4
= 1 blocks
With this information, we know that the Jordan form of A is given by
T
1
AT =
_
1
1 0 0 0
0
1
1 0 0
0 0
1
1 0
0 0 0
1
1
0 0 0 0
1
1
1 0
0
1
1
0 0
1
2
1
0
2
3
1 0
0
3
1
0 0
3
_
_
= A
J
Finding the Transformation Matrix T
Once the Jordan form is found, it is relatively easy to nd the transformation
matrix. We illustrate using Example 1.12. Write T in terms of its columns.
T =
_
_
t
1
t
2
t
3
t
4
t
5
. .
block 1
t
6
t
7
t
8
. .
block 2
t
9
t
10
. .
block 3
t
11
..
block 4
t
12
..
block 5
t
13
t
14
t
15
. .
block 6
.
_
_
Then we have
At
1
=
1
t
1
or (A
1
I) t
1
= 0
At
2
=
1
t
2
+t
1
or (A
1
I) t
2
= t
1
At
3
=
1
t
3
+t
2
or (A
1
I) t
3
= t
2
(1.41)
At
4
=
1
t
4
+t
3
or (A
1
I) t
4
= t
3
At
5
=
1
t
5
+t
4
or (A
1
I) t
5
= t
4
as well as
At
6
=
1
t
6
or (A
1
I) t
6
= 0
At
7
=
1
t
7
+t
6
or (A
1
I) t
7
= t
6
(1.42)
At
8
=
1
t
8
+t
7
or (A
1
I) t
8
= t
7
Therefore, we need to nd two linearly independent vectors t
1
, t
6
in N (A
1
I)
such that the chains of vectors
t
1
, t
2
, t
3
, t
4
, t
5
and t
6
, t
7
, t
8
are linearly independent. Alternatively, we can attempt to nd t
5
such that
(A
i
I)
5
t
5
= 0
and nd t
4
, t
3
, t
2
, t
1
from the set in eq. (1.41). Similarly, we can nd t
8
, such
that
(A
i
I)
3
t
8
= 0
and t
6
, t
7
, t
8
found from eq. (1.42) are linearly independent.
Moving to the blocks associated with
2
, we see that we need to nd t
9
, t
11
,
t
12
so that
(A
2
I) t
9
= 0
(A
2
I) t
11
= 0
(A
2
I) t
12
= 0
and (t
9
, t
10
, t
11
, t
12
) are independent with
At
10
=
2
t
10
+t
9
.
Likewise, we nd t
13
, such that
(A
3
I) t
13
= 0
(A
3
I) t
14
= t
13
(A
3
I) t
15
= t
14
with (t
13
, t
14
, t
15
) linearly independent. In each of the above cases, the exis-
tence of the vectors t
ij
are guaranteed by the existence of the Jordan forms.
Jordan Form of Companiopn Matrices
Frequently, the matrix A has the companion form:
A =
_
_
0 1
0 1
0
.
.
.
.
.
.
1
a
0
a
1
a
2
a
n1
_
_
.
1.5.4 Other Canonical Forms
Consider a nn real matrix with n2 real distinct eigenvalues and one pair of
complex conjugate eigenvalues, say
i
, for i = 1, , n 2 are real distinct and
(
n1
,
n
) are complex conjugate. Let t
i
be an eigenvector associated with
i
.
Then t
i
for i = 1, , n 2 are real vectors and [t
n1
, t
n
] are complex vectors.
We denote
n1
= +j and
n
= j.
Let
T := [t
1
t
2
t
n2
t
n1
t
n
] ,
then
T
1
AT =
_
2
.
.
.
n2
+j
j
_
_
:= D.
Dene
T :=
_
_
1
1
.
.
.
1
0.5 0.5j
0.5 0.5j
_
_
,
then it is easy to see that its inverse becomes
T
1
:=
_
_
1
1
.
.
.
1
1 1
j j
_
_
.
We now evaluate
T
1
D
T =
_
2
.
.
.
n2

_
_
:= A
c
.
This show that
T
1
T
1
A T

T
..
T
=

T
1
A
T = A
c
(1.43)
where
T = T

T = [t
1
t
n2
+j j ]
_
_
1
.
.
.
1
0.5 0.5j
0.5 0.5j
_
_
(1.44)
where
:= Re [t
n1
] = Re [t
n
] and := Im[t
n1
] = Im[t
n
] .
From the eq. (1.44), we have
T = T

T = [t
1
t
2
t
n2
]
= [t
1
t
2
t
n2
Re [t
n
] Im[t
n
] ] .
As seen in the form of A
c
above, A
c
is a real matrix with all distinct real (or
real parts of ) eigenvalues on the diagonal and imaginary parts of eigenvalues
on the o-diagonal. This is called the modal canonical form.
1.5.5 Elementary Divisors, Characteristic and Minimal Poly-
nomials
If () is a polynomial and A a square matrix, then () is called an annihilating
polynomial of A if (A) = 0. A monic annihilating polynomial of least degree
is called the minimal polynomial of A and it is unique.
Let
i
be an eigenvalue of a n n matrix A with multiplicity n
i
. Then the
characteristic polynomial of A is
() = det(I A) = (
i
)
n
i
(1.45)
where

i
n
i
= n. The Cayley-Hamilton theorem tells us that A satises its
own characteristic equation.
Suppose that the matrix Ais in Jordan form. Let m
i
be the largest dimension
of Jordan blocks associated with
i
. Then the minimal polynomial is dened as
() = (
i
)
m
i
. (1.46)
Note that
m :=
m
i

n
i
= n
where n is the dimension of A. The minimal polynomial is a fector of the charac-
teristic polynomial with its degree less than or equal to that of the characteristic
polynomial. It is clear that the minimal polynomial and the characteristic poly-
nomial become equal when each set of distinct eigenvalues has only one Jordan
block each.
Example 1.13. Consider the following matrices.
A
1
=
_
1
0 0 0
0
1
0 0
0 0
1
0
0 0 0
2
_
_
A
2
=
_
1
1 0 0
0
1
0 0
0 0
1
0
0 0 0
2
_
_
A
3
=
_
1
1 0 0
0
1
1 0
0 0
1
0
0 0 0
2
_
_
The characteristic polynomials of all three matrices are
() = (
1
)
3
(
2
) .
The minimal polynomials of A
1
, A
2
and A
3
are
1
() = (
1
) (
2
)
2
() = (
1
)
2
(
2
)
3
() = (
1
)
3
(
2
) .
1.6 SYSTEM STRUCTURE
1.6.1 Controllability
A dynamic system is said to be completely (state) controllable if every initial
state can be transferred to any nal prescribed state in a nite time T by some
input u(t), 0 t T. For a linear time invariant system, this condition can be
stated in terms of three equivalent rank conditions involving A and B.
Theorem 1.6 (Controllability)
The system
x = Ax +Bu (1.47)
is controllable i one of the following equivalent conditions hold:
1) rank[W
T
] = n where
W
T
:=
_
T
0
e
At
BB
T
e
A
T
t
dt, T > 0 and arbitrary
2) rank[B, AB, , A
n1
B] = n
3) rank[AI, B] = n for all eigenvalue C of A
Proof. We rst show that if W
T
is nonsingular, then eq. (1.47) is controllable.
Consider the response of the system in eq. (1.47) at time t
1
, that is
x(t
1
) = e
At1
x(0) +
_
t
1
0
e
A(t1)
Bu()d. (1.48)
We claim that for any x(0) = x
0
and any x(t
1
) = x
1
. Select
u(t) = B
T
e
A
T
(t1t)
W
1
t
1
_
e
At1
x
0
x
1
.
Then
x
1
= e
At
1
x
0
+
_
t1
0
e
A(t
1
)
Bu()d
= e
At
1
x
0

_
t
1
0
e
A(t
1
)
B
T
e
A
T
(t
1
)
W
1
t1
_
e
At
1
x
0
x
1
d
= e
At
1
x
0

_
t
1
0
e
A(t
1
)
B
T
e
A
T
(t
1
)
d
. .
Wt
1
W
1
t1
_
e
At
1
x
0
x
1
= e
At
1
x
0
e
At
1
x
0
+x
1
= x
1
.
Sec. 1.6. SYSTEM STRUCTURE 33
Since the selected u(t) transfers any x
0
to any x
1
at time t
1
, eq. (1.47) is
controllable.
Now we show the converse by contradiction. Note that the expression of W
T
shows that W
T
is positive semidenite. Suppose that eq. (1.47) is controllable
but W
T
is singular (not positive denite). It means that there exists a nonzero
vector v such that
v
T
W
T
v =
_
T
0
v
T
e
At
BB
T
e
A
T
t
vdt
=
_
T
0
_
_
_B
T
e
A
T
(t1)
v
_
_
_
2
d = 0
which implies that
B
T
e
A
T
(t
1
)
v = 0 or v
T
e
A(t
1
)
B = 0 (1.49)
for all [0, t
1
]. Now select the initial state and the nal state:
x
0
= e
At1
v and x
1
= 0.
If the system is controllable, we should be able to transfer x
0
to x
1
at a nite
time. So eq. (1.48) becomes
x
1
= e
At
1
x
0
+
_
t
1
0
e
A(t
1
)
Bu()d.
0 = e
At
1
e
At
1
v +
_
t
1
0
e
A(t
1
)
Bu()d
= v +
_
t
1
0
e
A(t1)
Bu()d.
Premultiplying by v
T
yields
0 = v
T
v +
_
t
1
0
v
T
e
A(t
1
)
B
. .
=0 from eq.(1.49)
u()d = v
2
+ 0
which contradicts v = 0. This proves (1).
To prove (2), let the initial state be x
0
and the nal state be x
1
= 0. then
0 = e
At
1
x
0
+
_
t1
0
e
A(t
1
)
Bu()d
= e
At
1
x
0
+e
At
1
_
t
1
0
e
A
Bu()d
yields
x
0
=
_
t
1
0
e
A
Bu()d.
Without loss of generality, we let
e
A
=
n1
k=0
f
k
()A
k
.
Then
x
0
=
n1
k=0
A
k
B
_
t
1
0
f
k
()u()d. (1.50)
Now let
k
:=
_
t1
0
f
k
()u()d,
then eq. (1.50) becomes
x
0
=
n1
k=0
A
k
B
k
=
_
B AB A
B
A
n1
B
1
.
.
.
n1
_
_
. (1.51)
If the system is completely controllable, eq. (1.51) must be satised for any
given unitial state x
0
. This requires that the condition (2) in the theorem.
Remark 1.1.
rank
_
B AB A
n1
B A
n
B
= rank
_
B AB A
n1
B
This is due to the Cayley Hamilton Theorem.

A
n
=
0
I +
1
A+ +
n1
A
n1
.
If we let x(0) and x
be the initial condition and the desired nal state of x(t),

respectively, it is easy to show that
u() = B
T
e
A
T
(T)
W
1
T
(x
e
AT
x(0)), 0 T
transfers x(0) to x
= x(T) in T seconds. Recall the solution of the state

equation.
x(T) = e
AT
x(0) +
_
T
0
e
A(T)
Bu()d
= e
AT
x(0) +
_
T
0
e
A(T)
BB
T
e
A
T
(T)
d
. .
W
T
W
1
T
(x
e
AT
x(0))
= e
AT
x(0) + (x
e
AT
x(0))
= x
.
Sec. 1.6. SYSTEM STRUCTURE 35
1.6.2 Observability
Let
x = Ax +Bu (1.52)
y = Cx +Du.
The system in eq. (1.52) is said to be observable if for any unknown initial state
x(0), there exists a nite t
1
> 0 such that the initial state x(0) is uniquely
determined by the input u and the output y over [0, t
1
]. Similar to the case of
controllability, we have the following equivalent conditions.
Theorem 1.7 (Observability)
The system
x = Ax +Bu
y = Cx
is observable i one of the following equivalent conditions hold:
1)
rank
_
_
C
T
C
T
A
C
T
A
2
.
.
.
C
T
A
n1
_
_
= n
2)
rank
_
AI
C
T
_
= n, for all C
3) rank[
T
] = n where
T
=
_
T
0
e
A
T
t
CC
T
e
At
dt, T > 0 and arbitrary
This is a result dual to Theorem 1.6.
1.6.3 BIBO Stability or External Stability
A signal f(t) is said to be bounded if there exists a nite constant M such that
|f(t)| f
M
< .
Consider a SISO LTI system with impulse response g(t). If we assume that the
input is u(t), then the response of the system is
y(t) =
_
t
0
g(t )u()d =
_
t
0
g()u(t )d.
A LTI system is said to be BIBO (Bounded Input Bounded Output) stable if the
output y(t) is bounded for every bounded input u(t).
Theorem 1.8 A LTI system is BIBO stable if and only if
_

0
|g(t)|dt M < .
Proof. We rst show that y(t) is bounded for a bounded u(t) if the above
expression is absolutely integrable. Assume that u(t) is bounded, that is
|u(t)| u
M
< .
Then
|y(t)| =
_
t
0
g()u(t )d
_
t
0
|g()||u(t )|d
_
t
0
|g()||u
M
d
u
M
_
t
0
|g()||d
u
M
_

0
|g()||d
u
M
M < .
This shows that y(t) is bounded. Next, we show necessity. Assume that
_

0
|g(t)|dt = .
We choose a special input dened by, at a xed t
,
u() =
_
+1, when g (t
) 0
1, when g (t
) < 0
Then the output y(t) evaluated at t = t
becomes
|y (t
) | =
_
t
0
g (t
) u()d
=
_
t
0
|g (t
)| d.
It is clear that as t
goes to ,
|y()| =
_

0
|g (t
)| d = 0 from the assumption

Therefore, eq. (1.8) is necessary.
Sec. 1.7. INTERCONNECTED SYSTEMS 37
Theorem 1.9 A multivariable system with impulse response matrix g(t) is
BIBO stable if and only if every entry of the matrix g(t) is absolutely inte-
gratable in [0, ).
Theorem 1.10 A multivariable system with proper rational transfer function
matrix G(s) is BIBO stable if and only if every pole of every entry of the matrix
G(s) has a negative real part.
1.7 INTERCONNECTED SYSTEMS
The design of high performance and complex systems is always broken down
into one of suitably interconnecting simpler individual systems. The overall
interconnected system can generally perform in a far superior fashion, if it is
well designed. In particular, feedback systems can have far superior (as well as
inferior characteristics) than open loop systems.
In this section we show how the dynamics of interconnected systems can be
obtained from the models of the subsystems when the subsystems are linear
and time-invariant and the noloading assumption holds. The latter condition
essentially amounts to lack of signicant energy exchange between the systems.
We begin with the simple series, parallel and feedback connections of two sys-
tems S
1
and S
2
.
u1 y1(t) u2(t) y2(t)
S1 S2
Figure 1.7. Two systems
1.7.1 State Space Representations of Interconnected Systems
We next discuss series, parallel and feedback connections of S
1
and S
2
described
through their state space models. Let
S
1
:
_
x
1
(t) = A
1
x
1
(t) +B
1
u
1
(t)
y
1
(t) = C
1
x
1
(t) +D
1
u
1
(t)
(1.53)
S
2
:
_
x
2
(t) = A
2
x
2
(t) +B
2
u
2
(t)
y
2
(t) = C
2
x
2
(t) +D
2
u
2
(t)
(1.54)
Series Connection
This connection is specied by (see Figure 1.8): Then
x
1
(t) = A
1
x
1
(t) +B
1
u(t)
x
2
(t) = A
2
x
2
(t) +B
2
y
1
(t) = A
2
x
2
(t) +B
2
C
1
x
1
(t) +B
2
D
1
u(t)
= B
2
C
1
x
1
(t) +A
2
x
2
(t) +B
2
D
1
u(t)
u
1
(t)
S1 S2
y
2
(t) y
1
(t) = u
2
(t)
Figure 1.8. Series connection (State Space Models)
(1.55)
y(t) = C
2
x
2
(t) +D
2
y
1
(t) = C
2
x
2
(t) +D
2
C
1
x
1
(t) +D
2
D
1
u(t)
= D
2
C
1
x
1
(t) +C
2
x
2
(t) +D
2
D
1
u(t).
Therefore, the state space model of the series connection is
_
_
x
1
(t)

x
2
(t)
_
_
=
_
_
A
1
.
.
. 0

B
2
C
1
.
.
. A
2
_
_
_
x
1
(t)
x
2
(t)
_
+
_
_
B
1

B
2
D
1
_
_
u(t) (1.56)
y(t) =
_
D
2
C
1
.
.
. C
2
_
_
x
1
(t)
x
2
(t)
_
+
_
D
2
D
1
u(t) (1.57)
Parallel Connection
S
1
S
2
+
y(t)
y1(t)
y
2
(t)
u1(t)
u
1
(t)
u(t)
Figure 1.9. Parallel connection (state space model)
In the connection shown in Figure 1.9,
u(t) = U
1
(t) = u
2
(t) and y(t) = y
1
(t) +y
2
(t).
The state space model of the parallel connection is
_
_
x
1
(t)

x
2
(t)
_
_
=
_
_
A
1
.
.
. 0

0
.
.
. A
2
_
_
_
x
1
(t)
x
2
(t)
_
+
_
_
B
1

B
2
_
_
u(t) (1.58)
y(t) =
_
C
1
.
.
. C
2
_
_
x
1
(t)
x
2
(t)
_
+
_
D
1
+D
2
u(t). (1.59)
Remark 1.2. In the series and parallel connections, the eigenvalues of the
new A matrix in each case equal the union of the eigenvalues of A
1
and those
of A
2
. Thus, in particular, the interconnected system is stable if and only if
the individual systems are stable. This will not at all hold in the feedback
connection.
Feedback Connection
S
1
S
2
+
u(t) u
1
(t)
u2(t)
y
1
(t) = y(t)
y2(t)
+
Figure 1.10. Feedback connection (state space models)

In the connection shown in Figure 1.10, we have
u
1
(t) = u(t) y
2
(t) and y(t) = y
1
(t) = u
2
(t).
Therefore,
y
1
(t) = C
1
x
1
(t) +D
1
u(t) D
1
y
2
(t)
= C
1
x
1
(t) +D
1
u(t) D
1
C
2
x
2
(t) D
1
D
2
y
1
(t)
and
[I +D
1
D
2
] y
1
(t) = C
1
x
1
(t) D
1
C
2
x
2
(t) +D
1
u(t). (1.60)
Similarly,
y
2
(t) = C
2
x
2
(t) +D
2
y
1
(t) = C
2
x
2
(t) +D
2
C
1
x
1
(t) +D
2
D
1
u
1
(t)
= C
2
x
2
(t) +D
2
C
1
x
1
(t) +D
2
D
1
u(t) D
2
D
1
y
2
(t)
and
(I +D
2
D
1
) y
2
(t) = D
2
C
1
x
1
(t) +C
2
x
2
(t) +D
2
D
1
u(t). (1.61)
To proceed, bring in the wellposedness assumption. This assumption states that
y
1
(t) and y
2
(t) must be k,m, mmmAQuniquely determined by x
1
(t), x
2
(t) and
u(t). From eqs. (1.60) and (1.61), the wellposedness assumption amounts to:
(1) I +D
1
D
2
is invertible
(2) I +D
2
D
1
is invertible
Actually, the conditions (1) and (2) are equivalent, that is, one holds if and only
if the other does.
To proceed, let
E
1
= (I +D
1
D
2
)
1
and E
2
= (I +D
2
D
1
)
1
so that from eqs. (1.60) and (1.61),
y
1
(t) = E
1
C
1
x
1
(t) E
1
D
1
C
2
x
2
(t) +E
1
D
1
u(t) (1.62)
y
2
(t) = E
2
D
2
C
1
x
1
(t) +E
2
C
2
x
2
(t) +E
2
D
2
D
1
u(t). (1.63)
Similarly,
u
2
(t) = E
1
C
1
x
1
(t) E
1
D
1
C
2
x
2
(t) +E
1
D
1
u(t) (1.64)
u
1
(t) = E
2
D
2
C
1
x
1
(t) E
2
C
2
x
2
(t) +E
2
u(t). (1.65)
Therefore, in the feedback system:
x
1
(t) = A
1
x
1
(t) +B
1
u(t) B
1
y
2
(t)
= A
1
x
1
(t) +B
1
u(t) B
1
E
2
D
2
C
1
x
1
(t) B
1
E
2
C
2
x
2
(t) B
1
E
2
D
2
D
1
u(t)
= (A
1
B
1
E
2
D
2
C
1
) x
1
(t) B
1
E
2
C
2
x
2
(t) + (B
1
B
1
E
2
D
2
D
1
) u(t)
and
x
2
(t) = A
2
x
2
(t) +B
2
E
1
C
1
x
1
(t) B
2
E
1
D
1
D
2
x
2
(t) +B
2
E
1
D
1
u(t)
= B
2
E
1
C
1
x
1
(t) + (A
2
B
2
E
1
D
1
C
2
) x
2
(t) +B
2
E
1
D
1
u(t).
Therefore, the state space representation of the feedback system is:
_
_
x
1
(t)

x
2
(t)
_
_
=
_
_
A
1
B
1
E
2
D
2
C
1
.
.
. B
1
E
2
C
2

B
2
E
1
C
1
.
.
. A
2
B
2
E
1
D
1
C
2
_
_
_
_
x
1
(t)

x
2
(t)
_
_
+
_
_
B
1
B
1
E
2
D
2
D
1

B
2
E
1
D
1
_
_
u(t)
(1.66)
y(t) =
_
E
1
C
1
.
.
. E
1
D
1
C
2
_
_
_
x
1
(t)

x
2
(t)
_
_
+E
1
D
1
u(t).
Remark 1.3. The dynamics, and in particular, the stability of the feedback
connection is determined by the new A matrix. As we show later, the feedback
connection can be stable even when S
1
and S
2
are unstable. It is also possible
that the feedback system is unstable even though S
1
and S
2
are stable. Thus,
feedback systems need to be designed with caution in order to take advantage
of the potential benets of feedback while avoiding its pitfalls.
Remark 1.4. Formulas given in eq. (1.66) simplify considerably if D
1
or D
2
is zero. In this case the wellposedness assumption holds automatically.
Example 1.14. Considet the feedback connection shown in Figure 1.10 and
let
S
1
:
_
_
x
1
=
_
0 1
2 1
_
. .
A1
x
1
+
_
0
1
_
..
B1
u
1
y
1
=
_
1 1
. .
C
1
x
1
+ 1
..
D1
u
1
S
2
:
_
_
x
2
=
_
0 1
3 2
_
. .
A
2
x
2
+
_
0
1
_
..
B
2
u
2
y
2
=
_
1 2
. .
C2
x
2
+ 1
..
D
2
u
2
Using the formulas in eqs. (1.66), we have
E
1
= (I +D
1
D
2
)
1
= (1 + 1 1)
1
=
1
2
E
2
= (1 +D
2
D
1
)
1
= (1 + 1 1)
1
=
1
2
and
A
1
B
1
E
2
D
2
C
1
=
_
0 1
2 1
_
_
0
1
_
1
2
1
_
1 1
=
_
0 1
5
2

3
2
_
B
1
E
2
C
2
=
_
0
1
_
1
2

_
1 2
=
_
0 0
1
2
1
_
B
2
E
1
C
1
=
_
0
1
_
1
2

_
1 1
=
_
0 0
1
2
1
2
_
A
2
B
2
E
1
D
1
C
2
=
_
0 1
3 2
_
_
0
1
_
1
2
1
_
1 2
=
_
0 1
7
2
3
_
B
1
B
2
E
2
D
2
D
1
=
_
0
1
_
_
0
1
_
1
2
1 1 =
_
0
1
2
_
B
2
E
1
D
1
=
_
0
1
_
1
2
1 =
_
0
1
2
_
E
1
C
1
=
1
2

_
1 1
=
_
1
2
1
2
E
1
D
1
C
2
=
1
2
1
_
1 2
=
_
1
2
1
E
1
D
1
=
1
2
1 =
1
2
.
Therefore,
_
_
x
1
(t)

x
2
(t)
_
_
=
_
_
0 1
.
.
. 0 0
5
2

3
2
.
.
.
1
2
1

0 0
.
.
. 0 1
1
2
1
2
.
.
.
7
2
3
_
_
_
_
x
1
(t)

x
2
(t)
_
_
+
_
_
0
1
2

0
1
2
_
_
u(t)
y(t) =
_
1
2
1
2
.
.
.
1
2
1
_
_
_
x
1
(t)

x
2
(t)
_
_
+
1
2
u(t).
The formulas become much simpler if D
1
or D
2
is zero.
1.7.2 Cancelations in Inteconnected Transfer Functions
In obtaining the models of interconnected system transfer functions, one should
be careful of pole - zero cancellations. Specically, consider the cascade connec-
tion of S
1
and S
2
shown in Figure 1.11:
u
1
(t)
s 1
s + 2
y
1
(t) u
2
(t)
1
s 1
y
2
(t)
S
1
S
2
Figure 1.11. Two systems to be connected.
The transfer function of the series connection is shown in Figure 1.12. and
s 1
s + 2
y
1
(t) = u
2
(t) u(t) = u
1
(t)
1
s 1
y
2
(t) = y(t)
1
s + 2
u(t) y(t)
Figure 1.12. Pole-zero cancellation.
the pole at s = +1 has been cancelled. Although the overall transfer function
is in fact
1
s + 2
,
it turns out that this transfer function is an inadequate description of the inter-
connected system. To see this clearly, write the state space models for S
1
and
S
2
:
S
1
:
_
x
1
(t) = 2x
1
(t) +u
1
(t)
y
1
(t) = 3x
1
(t) +u
1
(t)
(1.67)
S
2
:
_
x
2
(t) = x
2
(t) +u
2
(t)
y
2
(t) = x
2
(t)
(1.68)
and that of the cascade connection is
S :
_
_
_
x
1
(t) = 2x
1
(t) +u(t)
x
2
(t) = 3x
1
(t) +x
2
+u(t)
y(t) = x
2
(t)
(1.69)
The A matrix of S is
A =
_
2 0
3 1
_
which has eigenvalues at 2 and +1. When nonzero initial conditions are present
as x
1
(0) and x
2
(0), we see that the free response of the system consists of
weighted combinations of exponentials e
t
and e
2t
. Since the transfer function
is
1
s+2
, it is clear that the forced response will not contain the exponential e
t
. It
is in this sense that the transfer function of the interconnection is an incomplete
description of the system. In this example, the transfer function indicates that
the system is (externally) stable, whereas in reality it is (internally) unstable.
As another example, consider the parallel connection of
S
1
1
s + 1

1
s 1
=
2
(s + 1)(s 1)
S
2
1
s 1
The overall transfer function of the parallel connection is
1
s+1
. Again, the pole
at s = 1 has been cancelled and the transfer function
1
s+1
is an inadequate
description of the system. The state space models of S
1
and S
2
are
S
1
:
_
_
_
x
11
(t) = x
11
(t) +u
1
(t)
x
12
(t) = x
12
(t) +u
1
(t)
y
1
(t) = x
11
(t) x
12
(t)
S
2
:
_
x
3
(t) = x
3
(t) +u
2
(t)
y
2
(t) = x
3
(t)
The state space model of the parallel connection is:
S
1
:
_
_
x
11
(t)
x
12
(t)
x
3
(t)
_
_
=
_
_
1 0 0
0 1 0
0 0 1
_
_
_
_
x
11
(t)
x
12
(t)
x
3
(t)
_
_
+
_
_
1
1
1
_
_
u(t)
y(t) =
_
1 1 1

_
_
x
11
(t)
x
12
(t)
x
3
(t)
_
_
With nonzero initial conditions x
11
(0), x
12
(0), x
3
(0), the system dynamics con-
tains weighted exponentials e
t
and e
t
. However, since the transfer function
of the connection is
1
s+1
, the forced response for y(t) has only the exponential
terms e
t
. In particular, the transfer function indicates the system is (externally
stable) whereas the system is (internally) unstable.
These phenomena can also occur in feedback connections and of course there-
fore in other more general connections. From the above discussion, it should be
clear that the state space model of interconnected system is a more complete
description, since no modes disappear in the state space models.
1.8 SUMMARY
1.9 EXERCISES
1.10 NOTES AND REFERENCES
Chapter 2
REALIZATION THEORY
2.1 INTRODUCTION
The realization problem is: Given a transfer function G(s), construct a cir-
cuit, using standard components, so that the transfer function of the circuit is
precisely equal to the given G(s). In the old days, the standard circuit com-
ponents used to be R, L, C elements plus transformers. The modern version of
the problem uses integrators, multipliers, and summers as standard components
for continuous time systems and delays, multipliers, and summers for discrete
time (sampled data) systems.
The basic fact on which realization theory is founded is: Any rational, proper
1
transfer function matrix can be realized by the above components. The real-
ization of improper transfer functions requires, in addition to the above com-
ponents, dierentiators (for continuous time systems) or predictors (for discrete
time systems). Throughout the book, we assume that G(s) is rational and
proper unless specically stated.
2.2 SCALAR SYSTEMS: G(s) = g(s)
Let
g(s) =
b
n
s
n
+b
n1
s
n1
+ +b
1
s +b
0
s
n
+a
n1
s
n1
+ +a
1
s +a
0
(2.1)
where a
i
, b
j
are real and n is the order of the system. One possible realization
for this system is shown in Figure 2.1:
Another possibility is to divide out the eq. (2.1) and rewrite as follows:
g(s) =
c
n1
s
n1
+c
n2
s
n2
+ +c
1
s +c
0
s
n
+a
n1
s
n1
+ +a
1
s +a
0
+d. (2.2)
This leads to the following realization (see Figure 2.2): Assigning state variables
as in Figure 2.2, we have the set of equations:
x
1
= x
2
1
a transfer function whose degree of numerator is less than that of denominator.
45
46 REALIZATION THEORY Ch. 2
u
y
1
s
1
s
1
s
1
s
b
n
b2
b
1
b
0
a0
a1
an2
an1
Figure 2.1. A Typical Realization
y
1
s
1
s
1
s
1
s
c
2
c
1
c
0
x1 xn xn xn1
an2
an1
a1
a
0
u
cn1
d
Figure 2.2. A Typical Realization
x
2
= x
3
.
.
. (2.3)
.
.
.
x
n1
= x
n
x
n
= a
0
x
1
a
1
x
2
a
n1
x
n
+u
y = c
0
x
1
+c
1
x
2
+ +c
n1
x
n
+du.
Sec. 2.2. SCALAR SYSTEMS: G(s) = g(s) 47
Therefore, in this realization, we write
x = Ax +Bu
y = C
T
x +Du
where
A =
_
_
0 1 0 0
0 0 1 0
.
.
.
.
.
.
0 0 0 1
a
0
a
1
a
2
a
n1
_
_
:= A
c
, B =
_
_
0
0
.
.
.
0
1
_
_
:= b
c
C
T
=
_
c
0
c
1
c
2
c
n1
:= c
T
c
, D = d. (2.4)
A compact notation for this is the so-called packed matrix notation:
g(s) =
_
A
c
b
c
c
T
c
d
_
(2.5)
The special form of (A
c
, b
c
) is called the controllable companion form.
Laplace transforming (2.3) assuming x
i
(0) = 0,
u(s) = sX
n
(s) +a
n1
X
n
(s) +a
n2
X
n1
+ +a
1
X
2
(s) +a
0
X
1
(s)
= s
n
X
1
(s) +a
n1
s
n1
X
1
(s) +a
n2
s
n2
X
1
(s) + +a
1
sX
1
(s) +a
0
X
1
(s)
= (s
n
+a
n1
s
n1
+a
n2
s
n2
+ +a
1
s +a
0
)X
1
(s)
and
y(s) = c
0
X
1
(s) +c
1
X
2
(s) + +c
n1
X
n
(s) +du(s)
= (c
0
+c
1
s + +c
n1
s
n1
)X
1
(s) +du(s).
Therefore,
y(s) =
_
c
0
+c
1
s + +c
n1
s
n1
s
n
+a
n1
s
n1
+a
n2
s
n2
+ +a
1
s +a
0
+d
_
u(s)
= g(s)u(s).
Equivalently,
g(s) = c
T
c
(sI A
c
)
1
b
c
+d.
Notice that
g
T
(s) = g(s) =
_
c
T
c
(sI A
c
)
1
b
c
+d
T
=
_
c
T
c
(sI A
c
)
1
b
c
T
+d
T
= b
T
c
_
(sI A
c
)
1
T
c
c
+d (d is a constant)
= b
T
c
_
(sI A
c
)
T
1
c
c
+d
= b
T
c
(sI A
T
c
)
1
c
c
+d.
Therefore, by dening
c
T
o
:= b
T
c
, A
o
:= A
T
c
, b
o
:= c
c
, d
o
= d,
we have another realization, namely
g(s) =
_
A
o
b
o
c
T
o
d
o
_
. (2.6)
Here,
A
o
=
_
_
0 0 0 a
0
1 0 0 a
1
0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 1 a
n1
_
_
b
o
=
_
_
c
0
c
1
c
2
.
.
.
c
n1
_
_
(2.7)
c
T
o
=
_
0 0 0 1

d
o
= d.
The pair (c
T
o
, A
o
) is said to be in the observable companion form. The real-
izations in eq. (2.4) and eq. (2.7) are duals of each other. There is a circuit
corresponding to the observable realization given in eq. (2.7).
Other Realizations
It is possible to get many other realizations by appropriate decompositions
of the transfer function. For instance, writing
g(s) = g
1
(s)g
2
(s) g
k
(s)
leads to the realization where each subsystem g
i
(s) should be proper or strictly
-
u
g
1
(s)
-
g
2
(s)
- -
g
k
(s)
-
y
Figure 2.3. A Cascaded Realization
proper and therefore realizable. Likewise, writing
g(s) = h
1
(s) +h
2
(s) + +h
l
(s)
where h
i
(s) is proper or strictly proper leads to the realization.
In each case, state equations can be obtained by assigning a state variable
as the output of each integrator.
Sec. 2.3. MULTIVARIABLE SYSTEMS 49
-
h
l
(s)
6
-
u
-
h
2
(s)
-
h
1
(s)
?
+
-
y
R
Figure 2.4. A Parallel Realization
2.3 MULTIVARIABLE SYSTEMS
2.3.1 Multiinput Single Output Systems
Consider a system with r inputs and 1 output:
G(s) =
_
n
1
(s)
d
1
(s)
n
2
(s)
d
2
(s)

n
r
(s)
d
r
(s)
_
. .
strictly proper

G(s)
+
_
d
1
d
2
d
r
.
Let
d(s) = LCM[d
1
(s), d
2
(s), , d
r
(s)]
and rewrite the strictly proper part as
G(s) =
1
d(s)
_
n
1
(s) n
2
(s) n
r
(s)

where
d(s) = s
n
+a
n1
s
n1
+ +a
1
s +a
0
n
i
(s) = a
i
n1
s
n1
+c
i
n2
s
n2
+ +c
i
1
s +c
i
0
, i = 1, 2, , r.
Then a possible realization is the observable companion form of order n:
A =
_
_
0 0 0 a
0
1 0 0 a
1
0 1
.
.
.
.
.
.
.
.
.
0 0 1 a
n1
_
_
B =
_
_
c
1
0
c
2
0
c
r
0
c
1
1
c
2
1
c
r
1
.
.
.
.
.
.
.
.
.
.
.
.
c
1
n1
c
2
n1
c
r
n1
_
_
C
T
=
_
0 0 0 1

D =
_
d
1
d
2
d
r
.
2.3.2 Multioutput Single Input Systems
If we have a single input multioutput system, we have
G(s) =
_
_
G
1
(s)
G
2
(s)
.
.
.
G
m
(s)
_
_
. .
strictly proper
+
_
_
d
1
d
2
.
.
.
d
m
_
_
. .
D
.
We write as before
G
i
(s) =
n
i
(s)
d(s)
=
c
i
n1
s
n1
+c
i
n2
s
n2
+ +c
i
1
s +c
i
0
s
n
+a
n1
s
n1
+ +a
1
s +a
0
, i = 1, 2, , m.
Then a possible realization is the controllable companion form:
A =
_
_
0 1 0 0
0 0 1 0
.
.
.
.
.
.
0 0 0 1
a
0
a
1
a
2
a
n1
_
_
B =
_
_
0
0
.
.
.
0
1
_
_
C
T
=
_
_
c
1
0
c
1
1
c
1
2
c
1
n1
c
2
0
c
2
1
c
2
2
c
2
n1
.
.
.
.
.
.
c
m
0
c
m
1
c
m
2
c
m
n1
_
_
D =
_
_
d
1
d
2
.
.
.
d
m
_
_
.
2.3.3 Mutiinput-Multioutput (MIMO) Systems
In the general case, write extract D at the outset
G(s) =

G(s)
..
strictly proper
+ D
..
constant
and proceed proceed with the strictly proper part. We can either place all
entries in each column over a common denominator or place each row over a
common denominator. Doing columns, rst we would have
G(s) =
_
_
n
11
(s)
d
1
(s)
n
12
(s)
d
2
(s)

n
1r
(s)
d
r
(s)
n21(s)
d
1
(s)
n22(s)
d
2
(s)

n2r(s)
d
r
(s)
.
.
.
.
.
.
n
m1
(s)
d1(s)
n
m2
(s)
d2(s)

n
mr
(s)
dr(s)
_
_
Sec. 2.3. MULTIVARIABLE SYSTEMS 51
where
d
i
(s) = s
ni
+a
i
n
i
1
s
ni1
+ +a
i
1
s +a
i
0
n
ki
(s) = c
ki
n
i
1
s
n
i
1
+ +c
ki
1
s +c
ki
0
k = 1, 2, , m; i = 1, 2, , r.
A possible realization is
A =
_
_
A
1
A
2
.
.
.
A
r
_
_
B =
_
_
b
1
b
2
.
.
.
b
r
_
_
C
T
=
_
_
c
11
c
12
c
1r
c
21
c
22
c
2r
.
.
.
.
.
.
c
m1
c
m2
c
mr
_
_
D = D
where
A
i
=
_
_
0 1 0 0
0 0 1 0
.
.
.
.
.
.
0 0 0 1
a
i
0
a
i
1
a
i
2
a
i
ni1
_
_
b
i
=
_
_
0
0
.
.
.
0
1
_
_
c
ki
=
_
c
ki
0
c
ki
1
c
ki
n
i
1
k [1, 2, , m]; i [1, 2, , r].

The dual procedure with common denominators over rows is as follows:
G(s) =
_
_
n11(s)
d
1
(s)
n12(s)
d
1
(s)

n1r(s)
d
1
(s)
n
21
(s)
d
2
(s)
n
22
(s)
d
2
(s)

n
2r
(s)
d
2
(s)
.
.
.
.
.
.
n
m1
(s)
dm(s)
n
m2
(s)
dm(s)

n
mr
(s)
dm(s)
_
_
where
d
i
(s) = s
n
i
+a
i
ni1
s
n
i
1
+ +a
i
1
s +a
i
0
n
ij
(s) = c
ij
n
i
1
s
n
i
1
+ +c
ij
1
s +c
ij
0
.
Then the observable realization is
A =
_
_
A
1
A
2
.
.
.
A
m
_
_
B =
_
_
b
11
b
12
b
1r
b
21
b
22
b
2r
.
.
.
.
.
.
b
m1
b
m2
b
mr
_
_
C
T
=
_
_
c
T
1
c
T
2
.
.
.
c
T
m
_
_
D = D
where
A
i
=
_
_
0 0 0 a
i
0
1 0 0 a
i
1
0 1
.
.
.
.
.
.
.
.
.
0 0 1 a
i
n
i
1
_
_
b
ij
=
_
_
c
ij
0
c
ij
1
.
.
.
.
.
.
c
ij
n
i
1
_
_
c
T
i
=
_
0 0 0 1

.
2.4 MINIMAL REALIZATIONS
Let G(s) be a rational proper matrix and {A, B, C, D} a realization. If A is
n n, we say the order of the realization is n. An important question is: Is
it possible to realize G(s) with a lower order dynamic system? If not, n is the
minimal order. Otherwise, how do we nd the minimal order?
This problem was completely solved by Kalman in a classic paper
2
. The
solution involves the concepts of controllability and observability which are also
important in other areas.
Theorem 2.1 Minimal Realization A realization {A, B, C, D} of a proper ra-
tional matrix G(s) is minimal i (A, B) is controllable and (C, A) is observable.
This is the result obtained by Kalman. It implies that if (A, B) is not control-
lable, the order can be reduced. Likewise, if (C, A) is not observable, the order
can also be reduced. This reduction proceeds as shown in the next section.
2.5 COORDINATE TRANSFORMATION AND ORDER RE-
DUCTION
If we set
x = Tz T IR
nn
where T is invertible, then we have
z = T
1
ATz +T
1
Bu
y = C
T
Tz +Du
2
R.E. Kalman, Irreducible Realizations and the Degree of a Rational Matrix, SIAM J.
Appl. Math., Vol. 13, pp. 520-544, June 1965
Sec. 2.5. COORDINATE TRANSFORMATION AND ORDER REDUCTION 53
as the new state equations in z. It can be easily veried that the new transfer
function is
C
T
T(sI T
1
AT)
1
T
1
B +D = C
T
(sI A)
1
B +D
= old transfer function
and the new state space realization is related to the old one by relationship:
{A, B, C
T
, D}
T
{T
1
AT
. .
A
new
, T
1
B
. .
B
new
, C
T
T
. .
C
T
new
, D
..
D
new
}.
This is called a similarity transformation.
The next two observations are crucial. If
T
1
AT =
_
A
1
A
3
0 A
2
_
T
1
B =
_
B
1
0
_
C
T
T =
_
C
T
1
C
T
2
D = D,
we can see that
C
T
(sI A)
1
B +D = C
T
1
(sI A
1
)
1
B
1
+D.
Similarly, if
T
1
AT =
_
A
1
A
3
0 A
2
_
T
1
B =
_
B
1
B
2
_
C
T
T =
_
0 C
T
2
D = D,
then
C
T
(sI A)
1
B +d = C
T
2
(sI A
2
)
1
B
2
+D.
In the rst case, the order is reduced from n to n
1
(size of A
1
). In the second
case, the order is reduced from n to n
2
(size of A
2
).
2.5.1 Controllability Reduction
Let us regard
A : X X
as a linear operator, and dene
R :=
_
B AB A
n1
B
,
the controllability matrix and let R denote the column span of R. In other
words if rank[R] = n
1
, then R is the n
1
dimensional subspace spanned by the
columns of R. Let {v
1
, v
2
, , v
n1
} be a set of basis vectors for R and let
{w
n1+1
, w
n1+2
, , w
n
} be n n
1
vectors such that
T :=
_
v
1
v
n1
| w
n1+1
w
n
is an n n invertible matrix. Then we have the following.

Lemma 2.1
T
1
AT =
_
A
1
A
3
0 A
2
_
T
1
B =
_
B
1
0
_
(2.8)
where A
1
is n
1
n
1
, B
1
is n
1
r, and (A
1
, B
1
) is controllable.
The proof of this lemma depends on the following fact. V X is an A-invariant
subspace if v V implies that Av V.
Lemma 2.2 R is an A-invariant subspace and
B(column span of B) R.
In fact R is the smallest A-invariant subspace containing B.
Proof. Suppose r R. Then
r = By
0
+ABy
1
+ +A
n1
By
n1
for some vectors y
0
, y
1
, , y
n1
. Then
Ar = ABy
0
+A
2
By
1
+ +A
n
By
n1
.
By the Cayley-Hamilton Theorem
A
n
=
n1
A
n1
+
n2
A
n2
+ +
1
A+
0
I.
Substituting this in the expression for Ar, we have
Ar = Bz
0
+ABz
1
+ +A
n1
Bz
n1
for some vectors z
0
, z
1
, , z
n1
. Therefore,
Ar R.
Obviously, B R. To prove that R is the smallest such subspace, let S be a
smaller subspace. Then
B S R.
Applying A to both sides, we have
AB AS S R
A
2
B AS S R
.
.
.
A
n1
B AS S R.
Therefore,
R := B +AB + +A
n1
B S R.
so that S = R.
Proof of Lemma 2.1 Eq. (2.8) is equivalent to the following
A
_
v
1
v
n
1
.
.
. w
n
1
+1
w
n
_
=
_
v
1
v
n1
.
.
. w
n1+1
w
n
_
_
_
A
1
.
.
. A
3

A
4
.
.
. A
2
_
_
_
b
1
b
2
b
r
B =
_
v
1
v
n
1
.
.
. w
n
1
+1
w
n
_
_
_
B
1

B
2
_
_
and we want to prove that A
4
= 0, B
2
= 0. This follows from the following
facts
AR R : Av
i
=
i
1
v
1
+
i
2
v
2
+ +
i
n
1
v
n
1
, i = 1, 2, , n
B R : b
j
=
j
1
v
1
+
j
2
v
2
+ +
j
n
1
v
n
1
, j = 1, 2, , n
established in Lemma 2.1.
Therefore, if a realization {A, B, C
T
, D} is given with rank[R] = n
1
< n.
We can apply
1) a coordinate transformation so that
A
n
= T
1
AT =
_
A
1
A
3
0 A
2
_
B
n
= T
1
B =
_
B
1
0
_
C
T
n
= C
T
T =
_
C
T
1
C
T
2
D
n
= D
2) use the fact
C
T
(sI A)
1
B +D = C
T
n
(sI A
n
)
1
B
n
+D
n
=
_
C
T
1
C
T
2
_
(sI A
1
)
1
(sI A
1
)
1
A
3
(sI A
2
)
1
0 (sI A
2
)
1
_ _
B
1
0
_
+D
= C
T
1
(sI A
1
)
1
B
1
+D (see Remark 2.1)
to get the lower order realization of order n
1
, {A, B, C, D} which is moreover
controllable.
Remark 2.1.
3
If A
1
and B
1
exist,
_
A D
0 B
_
1
=
_
A
1
A
1
DB
1
0 B
1
_
.
3
T. Kailath, Linear Systems, Prentice-Hall, 1980, p.656
2.5.2 Observability Reduction
Dene
O :=
_
_
C
T
C
T
A

C
T
A
n1
_
_
and let be the null space (or kernel) of O.
:= {x : Ox = 0}.
Obviously, is the subspace that is orthogonal to all the rows of O. If rank[O] =
n
2
, then has dimension n n
2
.
Lemma 2.3 is A-invariant and is contained in Kernel(C
T
). In fact is the
largest such subspace.
Proof. If v , then C
T
A
i
v = 0, i = 0,1, ,n 1. Then C
T
A
j
Av = 0,
j = 0,1, ,n 2. To complete the proof of A-invariance we need to show that
C
T
A
n1
Av = 0. This follows, as in the proof of Lemma 2.2 from the Cayley-
Hamilton Theorem. If v , then certainly C
T
v = 0 so that
Kernel(C
T
).
To prove that is the largest such subspace, suppose that it is not and
1
is a
larger subspace with the property

1
Kernel(C
T
).
Then it is possible to argue as in Lemma 2.2 and show that

1
.
Now suppose that {v

1
, , v
n1
} is a basis for and choose {w
n1+1
, , w
n
}
so that
T :=
_
v
1
v
n
1
w
n
1
+1
w
n
is an invertible n n matrix. Then we have the following:

Lemma 2.4
T
1
AT =
_
_
A
1
.
.
. A
3

0
.
.
. A
2
_
_
C
T
T =
_
0
.
.
. C
T
2
_
(2.9)
where A
2
is n
2
n
2
, C
T
2
is mn
2
, and (C
T
2
, A
2
) observable.
Proof. Again eq. (2.9 is equivalent to the following matrix equations.
A
_
v
1
v
n
1
w
n
1
+1
w
n
=
_
v
1
v
n
1
w
n
1
+1
w
n
_
A
1
A
3
A
4
A
2
_
C
T
_
v
1
v
n1
w
n1+1
w
n
=
_
C
T
1
C
T
2
and we need to show that i) A

4
= 0, ii) C
T
1
= 0. But this follows from
i) A-invariance of , Av
i
=

n
1
j

i
j
v
j
and
ii) Kernel(C
T
) which means C
T
v
i
= 0, i = 1, 2, , n
1
.
Therefore if a realization {A, B, C, D} with rank[O] = n

2
< n is given we
can
1) apply a coordinate transformation T so that
A
n
= T
1
AT =
_
A
1
A
3
0 A
2
_
B
n
= T
1
B =
_
B
1
B
2
_
C
T
n
= C
T
T =
_
0 C
2
D
n
= D
2) use the fact
C
T
(sI A)
1
B +D = C
n
(sI A
n
)
1
B
n
+D
= C
2
(sI A
2
)
1
B
2
+D
to get a realization of order nn
2
, (A
2
, B
2
, C
T
2
, D) which is moreover observable.
2.5.3 Joint Reduction
Suppose that we have a realization (A, B, C, D) with rank[R] = n
1
. Apply the
controllability reduction to get a realization of order n
1
, (A
1
, B
1
, C
T
1
, D) with
(A
1
, B
1
) controllable. If (C
T
1
, A
1
) is observable, we are through as we have a
controllable and observable realization. Otherwise carry out an observability
reduction so that
T
1
A
1
T =
_
A
11
A
13
0 A
12
_
T
1
B
1
=
_
B
11
B
12
_
C
T
T =
_
0 C
T
12
D = D
and we have a realization (A
12
, B
12
, C
T
12
, D) which is observable. The question
that arises now is: Is (A
12
, B
12
) controllable? The answer is: If (A
1
, B
1
) is
controllable, so is (A
12
, B
12
).
Remark 2.2. This shows that a two step procedure is enough to produce a
controllable and observable realization (minimal realization).
2.6 GILBERTS REALIZATION
Gilberts Realization is a particar minimal realization which can be obtained di-
rectly from a transfer function matrix G(s). However, this realization is possible
only when each entry of G(s) has simple poles.
1. Expand each entry of G(s) into partial fractions.
2. Form
G(s) =
[R
1
]
s
1
+
[R
2
]
s
2
+
[R
3
]
s
3
+ .
3. Total size of realization is
n
i
Rank [R
i
] .
4. Find B
i
and C
i
so that
C
i
B
i
= R
i
where C
i
IR
nm
, B
i
IR
mn
; m = Rank [R
i
] .
5. Form (A, B, C) where
A =
_
_
1
I
1
2
I
2
_
_
B =
_
_
B
1
B
2
.
.
.
_
_ C =
_
C
1
C
2

.
Note that I
i
is the identity matrix with dimension being equal to Rank [R
i
].
Example 2.1. Find a minimal realization of the following transfer function.
G(s) =
_
_
1
(s 1)(s 2)
1
(s 2)(s 3)
1
(s 2)(s 3)
1
(s 1)(s 2)
_
_
Since all entries of G(s) have simple poles, we can use Gilbert Realization.
G(s) =
_
_
1
s 1
+
1
s 2
1
s 2
+
1
s 3
1
s 2
+
1
s 3
1
s 1
+
1
s 2
_
_
=
1
s 1
_
1 0
0 1
_
+
1
s 2
_
1 1
1 1
_
+
1
s 3
_
0 1
1 0
_
=
1
s 1
_
1 0
0 1
_
. .
C
1
_
1 0
0 1
_
. .
B
1
+
1
s 2
_
1
1
_
. .
C
2
_
1 1

. .
B2
+
1
s 3
_
0 1
1 0
_
. .
C
3
_
1 0
0 1
_
. .
B
3
Sec. 2.7. MATRIX FRACTION DESCRIPTION (MFD) OF MIMO SYSTEMS 59
Therefore,
A =
_
_
1 0
.
.
. 0 0 0
0 1
.
.
. 0 0 0

0 0
.
.
. 2
.
.
. 0 0

0 0 0
.
.
. 3 0
0 0 0
.
.
. 0 3
_
_
B =
_
_
1 0
0 1

1 1

1 0
0 1
_
_
C =
_
_
1 0
.
.
. 1
.
.
. 0 1
0 1
.
.
. 1
.
.
. 1 0
_
_
.
2.7 MATRIX FRACTION DESCRIPTION (MFD) OF MIMO
SYSTEMS
In some cases an easier approach to nding a minimal realization is through
a factorization of the rational transfer matrix G(s) as a ratio of polynomial
matrices representing the numerator and denominatoe Any r m rational
matrix transfer function G(s) can be written as follows:
G(s)
..
rm
= N
R
(s)
. .
rm
D
1
R
(s)
. .
mm
(right MFD)
= D
1
L
(s)
. .
rr
N
L
(s)
. .
rm
(left MFD)
where N
R
(s), D
R
(s), N
L
(s), D
L
(s) are polynomial matrices.
Example 2.2.
G(s) =
_
_
n
11
(s)
d
11
(s)

n
1m
(s)
d
1m
(s)
.
.
.
.
.
.
n
r1
(s)
d
r1
(s)

n
rm
(s)
d
rm
(s)
_
_
=
_
_
n
11
(s) n
1m
(s)
.
.
.
.
.
.
n
r1
(s) n
rm
(s)
_
_
_
_
d
1
(s)
.
.
.
d
m
(s)
_
_
1
=
_
d
1
(s)
.
.
.
d
r
_
_
1
_
_
n
11
(s) n
1m
(s)
.
.
.
.
.
.
n
r1
(s) n
rm
(s)
_
_.
The right decomposition can lead to a controllable realization and the left de-
composition to an observable realization of
order = degree [det(D
R
(s))] or
= degree [det(D
L
(s))] ,
respectively as we show in the example below.
Example 2.3.
G(s) =
_
_
1
s 1
1
s + 1
1
s
1
s + 1
1
s
1
s 1
_
_
=
_
_
1 +s
s
2
1
s
s
2
+s
1 +s
s
2
s
1 +s
s
2
1
1 +s
s
2
+s
s
s
2
s
_
_
. .
A(s)
=
_
_
s +s
2
s
3
s
s +s
2
s
3
s
s
2
1
s
3
s
s +s
2
s
3
s
1 +s
2
s
3
s
s +s
2
s
3
s
_
_
. .
B(s)
.
Let us observe A(s). It can be decomposed as follows:
A(s) =
_
1 +s s 1 +s
1 +s 1 +s s
_
_
_
s
2
1
s
2
+s
s
2
s
_
_
1
.
A state space realization can be easily obtained form the above as follows:
A
c
=
_
_
0 1
.
.
.
1 0
.
.
.

.
.
. 0 1
.
.
.
.
.
. 0 1
.
.
.

.
.
. 0 1
.
.
. 0 1
_
_
B
c
=
_
_
0 0 0
1 0 0

0 0 0
0 1 0

0 0 0
0 0 1
_
_
C
T
c
=
_
_
1 1
.
.
. 0 1
.
.
. 1 1
1 1
.
.
. 1 1
.
.
. 0 1
_
_
.
As seen, the above realization is controllable of order 6 which is equal to the
degree of the determinant of D
R
(s).
Now let us observe the MFD B(s). It can be decomposed as:
B(s) =
_
s
3
s 0
0 s
3
s
_
1
_
s s
2
s +s
2
1 +s
2
s +s
2
1 +s
2
s +s
2
_
.
A state space realization may be
A
o
=
_
_
0 0 0
.
.
.
1 0 1
.
.
.
0 1 0
.
.
.

.
.
. 0 0 0
.
.
. 1 0 1
.
.
. 0 1 0
_
_
B
o
=
_
_
0
.
.
. 0
.
.
. 1
1
.
.
. 1
.
.
. 0
1
.
.
. 1
.
.
. 1

0
.
.
. 1
.
.
. 0
1
.
.
. 0
.
.
. 1
1
.
.
. 1
.
.
. 1
_
_
C
T
o
=
_
_
0 0 1
.
.
. 0 0 0

0 0 0
.
.
. 0 0 1
_
_
.
This realization is observable and of order is equal to the degree of the deter-
minant of D
L
(s).
To get minimality we must extend the concept of coprimeness of scalar polyno-
mials to polynomial matrices.
2.7.1 Coprimeness
If A(s) = B(s)C(s) (all polynomial matrices), B(s) is a left divisor of A(s) and
C(s) is a right divisor of A(s). We dene a square polynomial matrix to be
unimodular i its inverse exists and is also a polynomial.
Example 2.4.
_
1 s
0 1
_
is unimodular
_
1 s
1 1
_
is not unimodular.
Lemma 2.5 A square polynomial matrix U(s) is unimodular i its determinant
is a nonzero constant.
Proof. If U
1
(s) is polynomial, then
U(s)U
1
(s) = I
which implies that
det
_
U(s)U
1
(s)
= det [U(s)] det

_
U
1
(s)
= 1.
Also we have
det [U(s)] =
1
det [U
1
(s)]
.
Here, we can see that det [U(s)] is a polynomial, so is
1
det [U
1
(s)]
. (2.10)
The only way the eq. (2.10) becoming a polynomial is
det
_
U
1
(s)
= constant.
Conversely, det [U(s)] = c (constant),
U
1
(s) =
Adj [U(s)]
c
= a polynomial matrix.
.
If A(s), B(s) are polynomial matrices with the same number of columns, then
R(s) is a common right divisor i
A(s) =

A(s)R(s)
B(s) =

B(s)R(s)
and it is a greatest common right divisor (gcrd) i every common right divisor
of A(s) and B(s) is a right divisor of R(s). i.e., every right divisor

R(s) of A(s)
and B(s) is
R(s) = W(s)

R(s) for some W(s).
If A(s), B(s) have the same number of rows, a greatest common left divisor can
be dened similarly.
2.7.2 Computation of gcrds and gclds
We rst need the concept of Hermite form of a polynomial matrix.
Row Hermite form = Upper Triangular form
_
_
a
1
1(s) a
12
(s)
0 a
22
(s)
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 a
rr
(s) a
r,r+1
(s)
.
.
. 0 0
.
.
.
.
.
.
0 0
_
_
with a
ii
(s) of degree higher than all others in i
th
column. The construction
can be done by row elementary operations corresponding to multiplication with
unimodular matrices on the left constructed from quotients and remainders
using the Euclidean division algorithm.
Example 2.5. (Reduction to Hermite Form)
_
_
s
2
0
0 s
2
1 s + 1
_
_

_
_
1 s + 1
0 s
2
s
2
0
_
_

_
_
1 s + 1
0 s
2
0 s
2
(s + 1)
_
_

_
_
1 s + 1
0 s
2
0 0
_
_
The corresponding unimodular matrix is
_
_
1 0 0
0 1 0
0 s + 1 1
_
_
_
_
1 0 0
0 1 0
s
2
0 1
_
_
_
_
0 0 1
0 1 0
1 0 0
_
_
=
_
_
0 0 1
0 1 0
1 s + 1 s
2
_
_
.
2.7.3 Construction of GCRDs
To construct a GCRD of N(s), D(s) we nd a unimodular U(s) to carry out
the row compression
_
_
U
11
(s)
. .
mm
U
12
(s)
U
21
(s) U
22
(s)
. .
pp
_
_
. .
U(s)
_
_
D(s)
. .
mm
N(s)
. .
pm
_
_
=
_
R(s)
0
_
Then, R(s) is a GCRD.
Proof. Let
_
U
11
(s) U
12
(s)
U
21
(s) U
22
(s)
_
1
=
_
V
11
(s) V
12
(s)
V
21
(s) V
22
(s)
_
Then,
_
D(s)
N(s)
_
=
_
V
11
(s)R(s)
V
21
(s)R(s)
_
and it is easy to see that R(s) is a common right divisor. Consequently, we have
the following matrix equation.
R(s) = U
11
(s)D(s) +U
12
(s)N(s). (2.11)
If R
1
(s) is another common right divisor, then
D(s) = D
1
(s)R
1
(s) N(s) = N
1
(s)R
1
(s)
and we have
R(s) = (U
11
(s)D
1
(s) +U
12
(s)N
1
(s))R
1
(s)
so that R
1
(s) is a right divisor of R(s) and so R(s) is a GCRD.
If R
1
(s), R
2
(s) are two GCRDs, then
R
1
(s) = W
2
(s)R
2
(s) R
2
(s) = W
1
(s)R
1
(s)
R
1
(s) = W
2
(s)W
1
(s)R
1
(s).
If R
1
(s) is nonsingular, then R
2
(s) is nonsingular and R
1
(s), R
2
(s) can only
dier by a unimodular factor. If R
1
(s) is unimodular then R
2
(s) is also. If
_
D(s)
N(s)
_
has full column rank, all GCRDs are nonsingular.
Remark 2.3. (Nonuniqueness of GCRDs: Summary) Notice that by carrying
out the elementary operations to reduce to Hermite form in dierence orders we
may get dierent matrices U(s) and hence dierent gcrds. However, any two
gcrds, R
1
(s) and R
2
(s), say, must be related (by denition) as
R
1
(s) = W
2
(s)R
2
(s), R
2
(s) = W
1
(s)R
1
(s), W
i
(s) polynomial.
Since
R
1
(s) = W
2
(s)W
1
(s)R
1
(s)
it follows that
1) If R
1
(s) is nonsingular, then the W
i
(s), i = 1, 2, must be unimodular, and
hence the gcrd R
2
(s) is also nonsingular. That is, if one gcrd is nonsin-
gular, then all gcrds must be so, and they can only dier by a unimodular
(left) factor.
2) If a gcrd is unimodular, then all gcrds must be unimodular.
Remark 2.4. (Nonsingular GCRDs) If
_
D(s)
N(s)
_
has full column rank,
then all gcrds of (N(s), D(s)) must be nonsingular and can dier only by uni-
modular (left) factors.
Denition 2.1. (N
R
(s),D
R
(s)) is right coprime i all its gcrds are unimodu-
lar. Similarly, (N
L
(s),D
L
(s)) is left coprime i all its gclds are unimodular.
2.7.4 Bezout Identity
Theorem 2.2 (N(s), D(s)) is right coprime i there exist X(s) and Y (s) such
that
X(s)N(s) +Y (s)D(s) = I
Proof. From the eq. (2.11), if R(s) is a gcrd, we have
R(s) =

X(s)N(s) +

Y (s)D(s).
Furthermore, if N(s) and D(s) are coprime, then R(s) is unimodular. Thus,
I = R
1
(s)

X(s)
. .
X(s)
N(s) +R
1
(s)
Y (s)
. .
Y (s)
D(s); R
1
(s) is polynomial
= X(s)N(s) +Y (s)D(s).
Conversely, if there exist X(s) and Y (s) satisfying
I = X(s)N(s) +Y (s)D(s),
let R(s) be any gcrd. Then
N(s) =

N(s)R(s), D(s) =

D(s)R(s)
and
I = X(s)

N(s)R(s) +Y (s)

D(s)R(s)
=
_
X(s)

N(s) +Y (s)

D(s)
_
R(s).
It follows that
R
1
(s) = X(s)

N(s) +Y (s)

D(s) (polynomial)
Since R
1
(s) is unimodular, (N(s), D(s)) is coprime.
Lemma 2.6 (N(s), D(s)) is coprime i
rank
_
D(s)
N(s)
_
= m s.
Lemma 2.7
_
U
11
(s) U
12
(s)
U
21
(s) U
22
(s)
_ _
D(s)
N(s)
_
=
_
R(s)
0
_
If D(s) is nonsingular, then
a) U
22
(s) is nonsingular,
b) N(s)D
1
(s) = U
1
22
(s)U
21
(s),
c) (U
21
(s), U
22
(s)) is left coprime, and
d) deg (det [D(s)]) = deg (det [U
22
(s)]) if (N(s), D(s)) is coprime.
Proof.
a) If U
22
(s) is singular, then there exists nonzero (s) such that
(s)U
22
(s) = 0.
It follows that
(s)U
21
(s)D(s) +(s)U
22
(s)N(s) = 0
which implies that
(s)U
21
(s) = 0
(s)
_
U
21
(s) U
22
(s)

= 0.
This means that U(s) is not unimodular. Therefore, U
22
(s) is nonsingular.
b)
U
21
(s)D(s) +U
22
(s)N(s) = 0
directly implies that
U
21
(s)D(s) = U
22
(s)N(s)
U
1
22
(s)U
21
(s) = N(s)D
1
(s).
c) From Lemma 2.6, since
_
U
21
(s) U
22
(s)

has full row rank for all s,
(U
21
(s), U
22
(s)) is left coprime.
d) Notice the following two matrix identities
4
:
det
_
A D
C B
_
= det [A] det
_
B CA
1
D
(2.12)
and
_
A D
C B
_
1
=
_
A
1
E
1
F E
1
1
F
1
_
(2.13)
where
= B CA
1
D
E = A
1
D
F = CA
1
.
Recall that
_
U
11
(s) U
12
(s)
U
21
(s) U
22
(s)
_ _
D(s)
N(s)
_
=
_
R(s)
0
_
4
T. Kailath, Linear Systems, Prentice-Hall, 1980, p. 680, p. 656
and
_
V
11
(s) V
12
(s)
V
21
(s) V
22
(s)
_ _
U
11
(s) U
12
(s)
U
21
(s) U
22
(s)
_
=
_
I 0
0 I
_
,
then we have
D(s) = V
11
(s)R(s).
Thus, since deg (det [R(s)]) = 0,
deg (det [D(s)]) = deg (det [V
11
(s)]) + deg (det [R(s)])
= deg (det [V
11
(s)]) (2.14)
Using the formulae given in eq. (2.12), we have
det [V (s)] = det [V
11
(s)] det
_
V
22
(s) V
21
(s)V
1
11
(s)V
12
(s)
= det [V
11
(s)] det
_
U
1
22
(see eq. (2.13))

= det [V
11
(s)]
1
det [U
22
(s)]
Thus,
det [U(s)] =
det [U
22
(s)]
det [V
11
(s)]
.
Here, we know that det [U(s)] = constant (i.e., unimodular). This leads
det [V
11
(s)] =
det [U
22
(s)]
constant
deg (det [V
11
(s)]) = deg (det [U
22
])
= deg (det [D(s)]) . (see eq. (2.14))
2.7.5 Generalized Bezout Identity

Theorem 2.3 (N
R
(s),D
R
(s)) is right coprime with det [D
R
(s)] = 0, then there
exist (X(s),Y (s),X
(s),Y
(s)) and (N
L
(s),D
L
(s)) such that
_
X(s) Y (s)
D
L
(s) N
L
(s)
_ _
N
R
(s) X
(s)
D
R
(s) Y
(s)
_
=
_
I 0
0 I
_
(2.15)
and
_
N
R
(s) X
(s)
D
R
(s) Y
(s)
_ _
X(s) Y (s)
D
L
(s) N
L
(s)
_
=
_
I 0
0 I
_
. (2.16)
Moreover, the block matrices will be unimodular.
Proof. From Theorem 2.2, we know that (N
R
(s),D
R
(s)) being right coprime
implies that there exist polynomial matrices (X(s),Y (s)) such that
X(s)N
R
(s) +Y (s)D
R
(s) = I. (2.17)
From Lemma 2.7, we also know that there will exist left coprime polynomial
matrices (N
L
(s),D
L
(s)) such that
D
1
L
(s)N
L
(s) = N
R
(s)D
1
R
(s).
Furthermore, the left coprimeness of (N
L
(s),D
L
(s)) implies the existence of
polynomial matrices (

X(s),
Y (s)) such that

D
L
(s)

X(s) +N
L
(s)
Y (s) = I. (2.18)
Putting eqs. (2.17) and (2.18) together, we have
_
X(s) Y (s)
D
L
(s) N
L
(s)
_ _
N
R
(s)

X(s)
D
R
(s)

Y (s)
_
=
_
I Q(s)
0 I
_
where
Q(s) = X(s)

X(s) +Y (s)
Y (s).
Thus,
_
X(s) Y (s)
D
L
(s) N
L
(s)
_ _
N
R
(s)

X(s)
D
R
(s)

Y (s)
_ _
I Q(s)
0 I
_
=
_
X(s) Y (s)
D
L
(s) N
L
(s)
_ _
N
R
(s) N
R
(s)Q(s) +

X(s)
D
R
(s) D
R
(s)Q(s) +

Y (s)
_
=
_
I 0
0 I
_
.
If we set
X
(s) = N
R
(s)Q(s) +

X(s)
Y
(s) = D
R
(s)Q(s) +

Y (s),
we have eq. (2.15). Eq. (2.16) follows by using the fact that CD = I implies
that DC = I when (C,D) are square constant or polynomial matrices. Finally,
we need to show the block matrices will be unimodular. This part is obvious
from eqs. (2.15) and (2.16) that the inverse of each block matrix is a polynomial
matrix.
Remark 2.5. To determine properness, the same denitions in the scalar case
hold here.
Proper: lim
s
H(s) <
Strictly Proper: lim
s
H(s) = 0
Denition 2.2. The highest degree of all the entries of the vector is called the
degree of a polynomial vector (or column degree).
Lemma 2.8 If H(s) strictly proper (proper), and
H(s) = N(s)D
1
(s)
then every column of N(s) has (column) degree strictly less than (less than or
equal to) that of the corresponding column of D(s).
Proof. Since N(s) = H(s)D(s), we have
_
_
h
11
(s) h
1m
(s)
.
.
.
.
.
.
h
r1
(s) h
rm
(s)
_
_
_
_
d
j1
(s)
.
.
.
d
jm
(s)
_
_ =
_
_
n
j1
(s)
.
.
.
n
jr
(s)
_
_.
Clearly, H(s) being strictly proper (proper) implies that the column degree of
_
n
j1
(s) n
j2
(s) n
jr
(s)

T
is strictly less that (less than or equal to) the column degree of
_
d
j1
(s) d
j2
(s) d
jm
(s)

T
Example 2.6. (Converse is not true.) Let

N(s) =
_
2s
2
+ 1 2

, D(s) =
_
s
3
+s s
s
2
+s + 1 1
_
The column degrees of N(s) are 2 and 0, and the column degrees of correspond-
ing vectors in D(s) are 3 and 1. Each of the column degree of N(s) is strictly
less than the column degree of the corresponding column of D(s). So, is H(s)
strictly proper?
H(s) = N(s)D
1
(s) =
_
s
2
+s
s
2
+s 1
s
3
+s 1
s
2
+s 1
_
It shows that H(s) is improper.
To make the converse true, we need the concept of column reduced polynomial
matrices.
2.7.6 Column Reduced Matrices
Let k
i
be the degree of the i
th
column of D(s). Then
deg (det [D(s)])
m
i=1
k
i
.
We say that D(s) is column reduced if the equality holds.
Example 2.7.
D(s) =
_
s
3
+s s + 2
s
2
+s + 1 1
_
It is easy to see that k
1
= 3 and k
2
= 1, and the sum of the column degree is 4.
However,
deg
_
det
_
s
3
+s s + 2
s
2
+s + 1 1
__
= 4.
Thus, D(s) is not column reduced. This is due to the coecient matrix of the
highest terms in each column is singular.
D(s) =
_
s
3
+s s + 2
s
2
+s + 1 1
_
=
_
1 1
0 0
_ _
s
3
0
0 s
_
+
_
s 2
s
2
+s + 1 1
_
As seen above, we can always write
D(s) = D
hc
S(s) +L(s)
where
S(s) =
_
_
s
k1
s
k2
.
.
.
s
k
m
_
_
and D
hc
is the highest column degree coecient matrix. Thus,
det [D(s)] = det [D
hc
] s
i
k
i
+ terms of lower degree in s.
Lemma 2.9 If D(s) is column reduced, H(s) = N(s)D
1
(s) is strictly proper
(proper) i each column degree of N(s) is less than (less than or equal to) the
column degree of the corresponding column of D(s).
Proof. We only prove for the case of strictly proper and the case of a proper
matrix is an obvious generalization. We rst need to show that if D(s) is column
reduced and each column degree of N(s) is less than the corresponding column
degree of D(s), then H(s) is strictly proper. Using Cramers rule, we know that
ij
th
entry of H(s) can be written as
h
ij
(s) =
det
_
D
ij
(s)
det [D(s)]
(2.19)
where
D
ij
(s) =
_
_
d
11
(s) d
1m
(s)
.
.
.
.
.
.
d
j1,1
(s) d
j1,m
(s)
n
i1
(s) n
im
(s)
d
j+1,1
(s) d
j+1,m
(s)
.
.
.
.
.
.
d
m1
(s) d
mm
(s)
_
_
.
Now write
D
ij
(s) = D
ij
hc
S(s) +L
ij
(s).
Note that D
ij
hc
is identical to D
hc
except that the j
th
row is now replaced by
zero since each entry of the j
th
row is of lower degree than the corresponding
entry of the j
th
row of D(s). Hence, D
ij
hc
is singular, while D
hc
is nonsingular.
It implies that
deg
_
det
_
D
ij
(s)
_
<
m
1
k
i
= deg (det [D(s)]) = deg (det [S(s)]) .
Therefore, h
ij
(s) is strictly proper (see eq. (2.19)) and hence so is the matrix
H(s). Next, we need to prove that if H(s) is strictly proper and D(s) is column
reduced, then each column degree of N(s) is less than the corresponding column
degree of D(s). This follows directly from Lemma 2.8.
Remark 2.6. D(s) is column reduced i det [D
hc
] = 0.
Lemma 2.10 If D(s) is not column reduced there exists unimodular U(s) so
that D(s)U(s) is column reduced.
Example 2.8.
D(s) =
_
(s + 1)
2
(s + 2)
2
(s + 1)
2
(s + 2)
0 s + 2
_
k
1
= 4, k
2
= 3, and
D(s) =
_
1 1
0 0
_
. .
D
hc
_
s
4
0
0 s
3
_
+L(s)
D(s) is not column reduced since D
hc
is singular. So, there must exist U(s) to
make it column reduced. Let us rst take
U
1
(s) =
_
1 0
s 1
_
,
then we have
D
1
(s) = D(s)U
1
(s) =
_
2s
3
+ 8s
2
+ 10s + 4 s
3
+ 4s
2
+ 5s + 2
s
2
+ 2s s + 2
_
.
Now let us observe the highest degree coecient matrix of D
1
(s).
D
1,hc
=
_
2 1
0 0
_
This shows that it is still not column reduced. So we take
U
2
(s) =
_
1 0
2 1
_
and we have
D
2
(s) = D
1
(s)U
2
(s) =
_
0 s
3
+ 4s
2
+ 5s + 2
s
2
+ 4s + 4 s + 2
_
.
Since
D
2,hc
=
_
0 1
1 0
_
is nonsingular, D
2
(s) is column reduced.
2.8 MINIMAL REALIZATIONS FROM MFDS
With the results of the last section we have
Lemma 2.11 Let
G(s) = N(s)D
1
(s)
with D(s) column reduced, and let
deg [n
i
(s)] < deg [d
i
(s)]
where n
i
(s) and d
i
(s) are the i
th
columns of N(s) and D(s), respectively. Then
G(s) is strictly proper.
Using this result, we can give the following realization. Let G(s) = N(s)D
1
(s)
be strictly proper, and write
D(s) = D
hc
S(s) +D
lc
(s)
where
S(s) =
_
_
s
k1
s
k2
.
.
.
s
k
r
_
_
(s) =
_
_
1 0 0
s 0 0
.
.
.
.
.
.
.
.
.
s
k
1
1
0 0
0 1 0
.
.
. s
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. s
k21
0
.
.
.
1
s
.
.
.
0 0 s
k
r
1
_
_
Sec. 2.8. MINIMAL REALIZATIONS FROM MFDS 73
with D
hc
invertible. Let
B
o
c
=
_
_
0
.
.
.
0
1
0
.
.
.
0
1
.
.
.
0
.
.
.
0
1
_
_
A
o
c
=
_
_
0 1 0 0
0 0 1 0
.
.
.
.
.
.
1
0 0
0 1 0 0
0 0 1 0
.
.
.
.
.
. 1
0 0
.
.
.
0 1 0 0
0 0 1 0
.
.
.
.
.
.
1
0 0
_
_
Remark 2.7. Write
N(s) = N
lc
(s).
This is always possible because deg [n
i
(s)] is less than deg [d
i
(s)].
Claim:
A
c
= A
o
c
B
o
c
D
1
hl
D
lc
B
c
= B
o
c
D
1
hl
C
c
= N
lc
is a controllable realization. If (N(s), D(s)) are right coprime, it is also an
observable and therefore minimal, realization.
Example 2.9.
G(s) =
_
s 0
s s
2
_
. .
N(s)
_
0 (s
3
+ 4s
2
+ 5s + 2)
(s + 2)
2
s + 2
_
. .
D(s)
1
As seen, k
1
= 2 and k
2
= 3. Thus,
D
hc
=
_
0 1
1 0
_
.
This implies that D(s) is column reduced. Therefore,
deg [n
1
(s)] = 1 < deg [d
1
(s)] = 2, deg [n
2
(s)] = 2 < deg [d
2
(s)] = 3.
Now we write
D(s) =
_
0 1
1 0
_
. .
D
hc
_
s
2
0
0 s
3
_
. .
S(s)
+
_
0 4s
2
+ 5s + 2
4s + 4 s + 2
_
. .
L(s)
=
_
0 1
1 0
_ _
s
2
0
0 s
3
_
+
_
0 0 2 5 4
4 4 2 1 0
_
. .
D
lc
_
_
1 0
s 0
0 1
0 s
0 s
2
_
_
. .
(s)
N(s) =
_
s 0
s s
2
_
=
_
0 1 0 0 0
0 1 0 0 1
_
. .
N
lc
_
_
1 0
s 0
0 1
0 s
0 s
2
_
_
. .
(s)
.
Finally,
A
c
= A
o
c
B
o
c
D
1
hc
D
lc
=
_
_
0 1 0 0 0
4 4 2 1 0
0 0 0 1 0
0 0 0 0 1
0 0 2 5 4
_
_
B
c
=
_
_
0 0
0 1
0 0
0 0
1 0
_
_
C
c
=
_
0 1 0 0 0
0 1 0 0 1
_
.
This realization is guaranteed to be controllable. It will be observable if (N(s), D(s))
are right coprime.
Sec. 2.9. SUMMARY 75
2.9 SUMMARY
2.10 EXERCISES
PART II
STABILITY AND
STABILIZATION
Chapter 3
REGULATOR PROBLEM I
3.1 POLE ASSIGNMENT BY STATE FEEDBACK
In this chapter we begin discussion of the feedback design problem by considering
rst the state feedback pole assignment problem. Although our nal objective
is to address the multivariable system shown below we will begin with simpler
n internal states
-
-
-
-
-
-
u
1
u
2
u
r
y
1
y
2
y
m
Figure 3.1. Input - Output System
cases to build up the solution.
3.1.1 Single Input Case
x = Ax +bu
y = C
T
x
If we let u(t) = 0, then
x = Ax y = C
T
x.
Consequently,
x(t) = e
At
x(0) and y(t) = C
T
e
At
x(0).
Suppose that the output y(t) is unsatisfactory, then we need a controller to
regulate the system. The state feedback problem considers the following:
77
78 REGULATOR PROBLEM I Ch. 3
f
C
T
- - -
u x y
Figure 3.2. State Feedback Conguration
Introducing a state feedback f, we have
u(t) = fx(t) = y(t) = C
T
e
(A+bf)t
x(t)
so that y(t) may be satisfactory. For example,
1) If A is unstable, can A be stabilized by f?
2) Can we nd an f such that eigenvalues of A + bf (i.e., poles of the closed
loop system) equal to a prescribed set of eigenvalues ?
This problem is the pole assignment problem using state feedback.
Problem
Given (A, b) and a desired set of eigenvalues, nd f so that the eigenvalues
of A+bf equal the desired set. Then the feedback control law
u = fx
assigns the eigenvalues of the closed loop system to the desired location.
Consider the following special case.
A =
_
_
0 1 0 0
0 0 1
.
.
.
.
.
.
.
.
. 0
.
.
. 1
a
0
a
1
a
2
a
n1
_
_
b =
_
_
0
0
.
.
.
0
1
_
_
.
and
=
_
d
1
,
d
2
, ,
d
n
.
Sec. 3.1. POLE ASSIGNMENT BY STATE FEEDBACK 79
Let
f =
_
f
0
f
1
f
2
f
n1
,
then
A+bf =
_
_
0 1 0 0
0 0 1
.
.
.
.
.
.
.
.
. 0
.
.
. 1
a
0
+f
0
a
1
+f
1
a
2
+f
2
a
n1
+f
n1
_
_
and
det [sI (A+bf)] = s
n
(a
n1
+f
n1
)s
n1
+ + (a1 +f
1
)s + (a
0
+f
0
)
= (s
d
1
)(s
d
2
) (s
d
n
)
= s
n
a
d
n1
s
n1
+ +a
d
1
s +a
d
0
.
Now we can equate the corresponding coecients,
a
0
+f
0
= a
d
0
a
1
+f
1
= a
d
1
.
.
.
a
n1
+f
n1
= a
d
n1
and solve for f
i
s.
The solution consists of the following:
1) Taking an arbitrary system (A, b) and transforming it to controllable com-
panion form by a coordinate transformation
2) solve the easy version of pole assignment problem in this coordinate system
3) transform back to the original coordinates so that the same eigenvalues are
obtained.
When is it possible?
Lemma 3.1 If (A, b) is controllable, there exists a coordinate transformation
T such that
A
n
= T
1
AT =
_
_
0 1 0 0
0 0 1
.
.
.
.
.
.
.
.
. 0
.
.
. 1
a
0
a
1
a
2
a
n1
_
_
b
n
= T
1
b =
_
_
0
0
.
.
.
0
1
_
_
.
Theorem 3.1 Pole assignment by state feedback is possible i (A, b) is a con-
trollable pair.
Proof. Suppose that (A, b) is not controllable, then we know we can separate
controllable and uncontrollable parts as follows.
x = Ax +bu
x = Tz
z = T
1
ATz +T
1
bu
_
z
c
z
u
_
=
_
A
1
A
3
0 A
2
_
. .
A
_
z
c
z
u
_
+
_
b
1
0
_
. .
b
u
Since
u = fx = fTz =

fz =
_

f
1

f
2
_
z
c
z
u
_
,
_
z
c
z
u
_
=
_
A
1
+b
1

f
1
A
3
+b
1

f
2
0 A
2
_
. .
A+
f
_
z
c
z
u
_
.
The eigenvalues of

A+
f are the roots of the polynomial

det
_
sI (

A+
f)
_
= det
_
sI (A
1
+b
1

f
1
) (A
3
+b
1

f
2
)
0 sI A
2
_
= det
_
sI (A
1
+b
1

f
1
)
_
det [sI A
2
] .
As seen f has no eect on the uncontrollable part of eigenvalues that is, the
eigenvalues of A
2
are xed and independent of f.
How to make the controllable companion transformation The following
procedure constructs a transformation matrix that coordinate transforms an
arbitrary controllable system to the controllable companion form.
1)
L =
_
b Ab A
n1
b
2) take the last row of L

1
and call it q
T
3) Construct
T
1
=
_
_
q
T
q
T
A
.
.
.
q
T
A
n1
_
_
Sec. 3.1. POLE ASSIGNMENT BY STATE FEEDBACK 81
Proceeding, let a state feedback

f assign the eigenvalues of

A+
f to the desired
locations. The last step is to nd the solution
f =

fT
1
which is valid in the original coordinates, because from
A+bf = T(

A+
f)T
1
A+
f and A+bf have the same eigenvalues.

3.1.2 Multioutput Case via Single Input Reduction
Consider a multiinput system,
x = Ax +Bu, A IR
nn
, B IR
nm
In some cases (which cases?) we can reduce this system to a single input system
by introducing a new signal
u = gv g IR
m
and retain controllability of the system from the new input v. Then we have
the new system which has a single input,
x = Ax +Bgv := Ax +bv.
We now design a state feedback for this system. In general, we may have to use
coordinate transformation of z = Tx.
z = T
1
ATz +T
1
bv :=

Az +
bv; v =

fz
= (

A+
f)z.
Consequently, since
v =

fz =

fT
1
x = fx,
we have
x = Ax +bv = (A+b

fT
1
. .
f
)x = (A+Bg

fT
1
. .
F
)x. (3.1)
Remark 3.1. This approach of using controllable companion form can be nu-
merically unreliable, because the controllable companion form transformation
is sometimes numerically ill conditioned.
3.1.3 Direct Method
An attractive alternative method of solution is as follows. Consider the equation
X
1
(A+BF)X =

A;

A has the desired set of eigenvalues
Then,
AX +BFX = X

A
AX X

A = BFX.
This leads the following matrix equations:
AX X

A = BG; given A and

A, a choice of G (3.2)
F = GX
1
. (3.3)
The questions that arise are:
1) Does the solution of eq. (3.2) always exist? (perhaps, unique?)
3) Is the solution X invertible?
4) How to choose G?
Lemma 3.2 If (A, B) is controllable, and (G,

A) is observable, then the unique
solution X of eq. (3.2) is almost always nonsingular.
Based on this we develop the procedure:
Procedure:
1) Pick

A such that it has the desired eigenvalues.
2) Pick G with G,

A and solve eq. (3.2). If X is singular, choose a dierent G
and repeat the process.
3) If X is nonsingular, solve for F from eq. (3.3).
Extensions of this problem
1) Robust Eigenstructure Assignment Problem
2) Minimum Norm Pole Assignment Problem
3.2 STABILIZATION VIA OUTPUT FEEDBACK
Consider the closed loop control system with the plant and controller described
in their respective state space representations.
Plant : x = Ax +Bu, x n vector
y = Cx
Controller : x
c
= A
c
x
c
+B
c
u
c
, x
c
n
c
vector
y
c
= C
c
x
c
+D
c
u
c
Sec. 3.2. STABILIZATION VIA OUTPUT FEEDBACK 83
controller
plant

- -
y
c
u
u
c
y
Figure 3.3. A Closed Loop Control System
Then,
Feedback Connection : u
c
= y
u = y
c
and
Feedback System :
_
x
x
c
_
=
_
A+BD
c
C BC
c
B
c
C A
c
_
. .
A
cl
IR
(n+n
c
)(n+n
c
)
_
x
x
c
_
.
Is the closed loop system stable?
Consider the stability of
x = Ax +Bu
y = Cx.
Denition 3.1. Internal Stability: When u(t) = 0, x(t) 0.
External Stability: When u(t) = 0, y(t) 0.
Internal stability is stronger than and implies external stability, since if x(t) 0,
then clearly y(t) = Cx(t) 0. The converse is not true in general. However, if
the system is controllable and observable, then external stability implies internal
stability.
Remark 3.2. Checking Internal Stability: (A) LHP.
Checking External Stability: Poles of G(s) in LHP.
Now let us get back to the closed loop system. What we really want is
internal stability of the closed loop system.
Remark 3.3. A closed loop system is internally stable i all eigenvalues of
A
cl
lie in the open LHP.
Basic Result
If (A, B, C) corresponding to the plant is controllable and observable, then a
controller of a high enough order n
c
can always be found to assign the eigenvalues
of A
cl
arbitrarily. In particular, a controller of order n 1 always suces for a
plant of order n.
Plant: x = Ax +Bu
y = Cx
Controller: x
c
= A
c
x
c
+B
c
y
u = C
c
x
c
+D
c
y
The closed loop system becomes
_
x
x
c
_
=
_
A+BD
c
C BC
c
B
c
C A
c
_
. .
A
cl
_
x
x
c
_
Problem : Given (A, B, C) nd (A
c
, B
c
, D
c
, D
c
) so that A
cl
is stable, i.e.,
(A
cl
) I C
.
Theorem 3.2 There exists a stabilizing controller if and only if (A, B) is sta-
bilizable and (C, A) is detectable.
3.2.1 State Feedback
A quick recap of the state feedback results:
Theorem 3.3 (Pole Placement Theorem: Wonham, 1967) If (A, B) is con-
trollable, there exists F so that (A + BF) is equal to any set of n prescribed
eigenvalues (in conjugate pairs).
Proof. a) It is true for B = b (single input case). Let
L =
_
b Ab A
n1
b
and let q
be the last row of L

1
and
Q =
_
_
q
A
.
.
.
q
A
n1
_
_
.
Sec. 3.3. OBSERVER DESIGN 85
Take T = Q
1
. Then
A = T
1
AT =
_
_
0 1 0 0
0 0 1 0
.
.
.
.
.
.
0 1
a
0
a
1
a
2
a
n1
_
b
n
= T
1
b =
_
_
0
0
.
.
.
0
1
_
_
.
Let
f =
_

f
1

f
2

f
n
.
Then we have
A+
f =
_
_
0 1 0 0
0 0 1 0
.
.
.
.
.
.
0 1
a
0
+

f
1
a
1
+

f
2
a
2
+

f
3
a
n1
+

f
n
_
_
.
|sI

A
f| = s
n
+ (a
n1

f
n
)s
n1
+ + (a
0

f
1
)
= s
n
+a
d
n1
s
n1
+a
d
n2
s
n2
+ +a
d
0
(arbitrary).
Then

fT
1
= f gives A+bf the same characteristic polynomial.
b) The multiinput case can be proved by reducing to the single input case via
the following lemma. The proof of this lemma is deferred.
Lemma 3.3 If (A, B) is controllable and g = 0 is any vector, then there exists
F
0
such that (A+BF
0
, Bg) is controllable.
In practice, an arbitrary F
0
will almost always work! Let
A
0
= A+BF
0
, b
0
= Bg.
Then nd f
0
so that A
0
+b
0
f
0
has the desired eigenvalues. Then
F = F
0
+gf
0
is the state feedback so that A+BF has the desired eigenvalues.
3.3 OBSERVER DESIGN
To proceed with our construction of the feedback controller we need to bring
in the concept of an observer. Problem: Design a device which will
measure x (n dimensional state vector) from external measurements y (m
vector) and u (r vector).
F
plant
observer
-
-
-
-
u y
x
Figure 3.4. A Feedback with State Estimator
System: x = Ax +Bu
y = Cx
Observer: z = Mz +Ly +Gu
x = Pz +Qy +Ru
Requirement Design (M, L, G, P, Q, R) so that
lim
t
( x(t) x(t)) = 0 x(0), z(0), u(t).
It will turn out that this will be possible if (C, A) is detectable.
3.3.1 Full Order Observer
Let P = I
n
, Q = R = 0, then we have
x = z.
Let
e = z x.
Then
e = z x
= Mz +Ly +Gu
. .
z
(Ax +Bu)
. .
x
= Mz +LCx +Gu Ax Bu
= Me + (M A+LC)x + (GB)u.
Therefore, if we set
G = B
M = ALC
so that
e = Me = (ALC)e
and the inuence of x and u on e are cancelled. For convergence of e(t) 0,
we need that
(ALC) I C
.
If (C, A) observable, we can nd L to place (A LC) arbitrarily by the pole
placement theorem.
Remark 3.4. The eigenvalues of (A LC) is identical to the eigenvalues of
(A
T
C
T
L
T
). (C, A) is observable i (A
T
, C
T
) is controllable.
The observer designed above is sometimes called an identity observer because
each component of z estimates the corresponding component of x.
3.3.2 Reduced (Minimal) Order Observer
Since some of the n system states are measurable in the form of y, it should be
possible to estimate only the remaining (nm) unmeasurable states. Without
loss of generality (change coordinates if necessary), let
y = x
1
the rst m components of the state vector and let x
2
be the remaining n m
components which are not measurable. Then the system equations are
_
x
1
x
2
_
=
_
A
11
A
12
A
21
A
22
_ _
x
1
x
2
_
+
_
B
1
B
2
_
u
y = x
1
=
_
I 0

_
x
1
x
2
_
or since x
2
is the state to be estimated, write
x
2
= A
22
x
2
+ (B
2
u +A
21
y) (3.4)
A
12
x
2
= y A
11
y B
1
u. (3.5)
Think of (3.4) as the dynamic equation for x
2
and (3.5) as the measurement
equation for x
2
.
Lemma 3.4 If (C, A) is observable then so is (A
12
, A
22
).
Now we apply the full order observer formulas
z = (ALC)z +Ly +Bu,
to x
2
and we have
z
2
= (A
22
LA
12
)z
2
+L
2
( y A
11
y B
1
u) + (B
2
u +A
21
y) (3.6)
and the error will satisfy
e
2
= z
2
x
2
e
2
= (A
22
L
2
A
12
)e
2
.
The same as the case of full order observer, for e(t) 0, it is necessary that
(A
22
L
2
A
12
) I C
by a choice of L
2
. Now let us eliminate y to write the
reduced order observer equation. Let
w = z
2
L
2
y.
Then eq. (3.6) is rewritten as follows:
w = (A
22
L
2
A
12
)w + [(A
22
L
2
A
12
)L
2
L
2
A
11
+A
21
]y + (B
2
L
2
B
1
)u
z
2
= w +L
2
y.
3.3.3 Closing the Loop
Suppose we calculate a state feedback u = Fx to place the eigenvalues of A+BF.
Now we design an observer with eigenvalues of M chosen by us. From the
observer we obtain x. What if we close the loop with
u = F x
instead of u = Fx? What are the closed loop eigenvalues? This answer is given
by the so-called Separation Principle.
Theorem 3.4 The closed loop system under observed state feedback has eigen-
values
(A+BF) (M).
(Proof of Separation Principle) Consider the general observer
z = Mz +Ly +Gu.
Let
z Tx = e.
Then
e = Me + (MT TA+LC)x + (GTB)u.
Therefore to make z estimate Tx, we want e(t) 0. Thus, set
MT TA+LC = 0 (3.7)
GTB = 0 (3.8)
and
(M) I C
.
To estimate Fx we use the observer output
t := F x = S
1
z +S
2
y
= (S
1
T +S
2
C)x +S
1
e
and set
S
1
T +S
2
C = F. (3.9)
The loop is closed by setting u = t (instead of u = Fx). Now look at the closed
loop equations:
x = Ax +Bu
z = Mz +Ly +Gu
u = S
1
z +S
2
y.
We have
_
x
z
_
=
_
A+BS
2
C BS
1
LC +GS
2
C M +GS
1
_
. .
A
cl
_
x
z
_
. (3.10)
Now make the coordinate transformation
_
x
z
_
=
_
I 0
T I
_ _
x
e
_
_
x
e
_
=
_
I 0
T I
_ _
x
z
_
.
Then using (3.7) - (3.10), we have
_
I 0
T I
_
A
cl
_
I 0
T I
_
=
_
A+BF BS
1
0 M
_
Therefore,
(A
cl
) = (A+BF) (M).
The transfer function of the observer as controller is derived as follows:

z = Mz +Ly +Gu
t = S
1
z +S
2
y
z = (M +GS
1
)z + (L +GS
2
)y
Therefore,
t(s) =
_
S
1
(sI M GS
1
)
1
(L +GS
2
) +S
2
. .
C(s)
y(s)
Example 3.1. Consider the following single-input two-output system:
x =
_
_
0 1 0
0 0 0
1 0 0
_
_
x +
_
_
1
0
1
_
_
u
y =
_
1 0 0
0 0 1
_
x
Design a full order observer with eigenvalues at 1, 2, 3 to estimate the
state x.
Solution
ALC =
_
_
0 1 0
0 0 0
1 0 0
_
_
_
_
l
11
l
12
l
21
l
22
l
31
l
32
_
_
_
1 0 0
0 0 1
_
=
_
_
l
11
1 l
12
l
21
0 l
22
1 l
31
0 l
32
_
_
One easy choice is:
l
31
= 1, l
32
= 3, s
2
+l
11
s +l
21
= s
2
+ 3s + 2
Thus, we have
l
11
= 3, l
21
= 2, l
12
, l
22
arbitrary, say, 0.
Then
L =
_
_
3 0
2 0
1 3
_
_
and
ALC =
_
_
3 1 0
2 0 0
0 0 3
_
_
Therefore, the observer is
z =
_
_
3 1 0
2 0 0
0 0 3
_
_
z +
_
_
3 0
2 0
1 3
_
_
y +
_
_
1
0
1
_
_
u
z = x is the estimate of x.
Example 3.2. For the system given in the previous example, design a minimal
order observer with eigenvalue at 3.
Solution
y =
_
y
1
y
2
_
=
_
x
1
x
3
_
:=
_
x
1
x
2
_
Now
x
1
= x
1
= x
2
+u = x
3
+u
x
2
= x
3
= x
1
+u = x
1
+u
x
3
= x
2
= 0
_
_
x
1
x
2
x
3
_
_
=
_
_
0 0
.
.
. 1
1 0
.
.
. 0

0 0
.
.
. 0
_
_
_
_
x
1
x
2
x
3
_
_
+
_
_
1
1

0
_
_
u
_
y
1
y
2
_
=
_
x
1
x
2
_
=
_
_
1 0
.
.
. 0
0 1
.
.
. 0
_
_
_
_
x
1
x
2
x
3
_
_
Thus,
A
11
=
_
0 0
1 0
_
A
12
=
_
1
0
_
B
1
=
_
1
1
_
A
21
=
_
0 0

A
22
= 0 B
2
= 0
The observer is:
w
2
= (A
22
L
2
A
12
)w
2
+ [(A
22
L
2
A
12
)L
2
L
2
A
11
+A
21
]y + (B
2
L
2
B
1
)u
x
3
= w
2
L
2
y.
Since the eigenvalue of A
22
L
2
A
12
should be 3, we have
0
_
l
21
l
22
_
1
0
_
:= 3
Here, we choose l
21
= 3 and l
22
= 0. Therefore,
w
2
= 3w
2
9y
1
3u
x
3
= w
2
+ 3y
1
= x
2
.
Example 3.3.
A =
_
0 1
0 0
_
b =
_
0
1
_
c =
_
1 1

i) Design a minimal order state estimator with its pole at 1.
x = Ax +bu y = cx
We need to pick T so that
cV =
_
1 0

.
We choose
V =
_
1 1
0 1
_
and V
1
=
_
1 1
0 1
_
So, we have new coordinate system with V as follows:
x =

A x +
bu y = c x
where
A = V
1
AV =
_
_
0
.
.
. 1

0
.
.
. 0
_
_

b = V
1
B =
_
_
1

1
_
_
c = cV =
_
1
.
.
. 0
_
.
The minimal order state estimator equation is
w = (

A
22
L
2

A
12
)w+
_
(

A
22
L
2

A
12
)L
2
+L
2

A
11
+

A
21
y +(
b
2
L
2
b
1
)u
Let us select L
2
such that
(

A
22
L
2

A
12
) = 1.
Then we have L
2
= 1. Therefore, the minimal order state estimator
becomes
w = w +y
x
2
= w +L
2
y.
ii) Combine the state feedback and the observer into an output feedback con-
trollor, and give the state equations as well as the transfer function of this
controller. Choose a state feedback to assign poles of the original system
at all 1.
Now we need to consider the coordinate transformation.
x = V

x, y = c x, w =

x
2
L
2
y
_
y
w
_
=
_
1 0
L
2
1
_ _
x
1
x
2
_
. .
x
.
Thus,
x = V

x = V
_
1 0
L
2
1
_ _
y
w
_
=
_
1 0
1 1
_ _
y
w
_
From the control law, f = [f
1
f
2
] = [1 2] which assigns the original
poles at 1, we have
u = F x =
_
1 2

_
1 0
1 1
_ _
y
w
_
= 3y + 2w.
Therefore, the controller state equations are:
w = w +y
u = F x = 2w 3y
and the controller transfer function is:
G(s) = c(sI A)
1
b +d
= 2(s + 1)
1
1 + (3)
=
2
s + 1
3
=
3s 1
s + 1
.
3.4 LINEAR QUADRATIC PROBLEM
For the given system dynamics
x = Ax +Bu,
the LQR problem is stated as that of nding a state feedback control law of the
form
u = Kx
which minimizes a performance measure of the form
J =
_
T
0
(x
Qx +u
Ru)dt +x
(T)Mx(T)
where Q and M are typically positive semi-denite matrices, R is a positive
denite matrix.
Remark 3.5. The matrix Q is positive semi-denite if x
Qx 0 for all x. If
equality holds only when x = 0, we say that Q is positive denite.
Remark 3.6. A symmetric matrix Q is positive semi-denite (denite) i all
its eigenvalues are nonnegative (positive).
Remark 3.7. A symmetric matrix Q is positive semi-denite (denite) i the
leading principal minors (the determinants of the k k matrices in the top
left-hand corner of Q, K = 1, 2, , n) are all nonnegative (positive).
Remark 3.8. Q is positive semi-denite i it can be written in the factored
form
Q = LL
for some matrix L. L is known as a square root of Q.

Here, we consider the nal time T is xed, but the nal state x(T) is free. The
control goal is to keep the state x(t) as close to the state zero as possible. The
term x
Qx is then a measure of control accuracy, the term u
Ru a measure of
control eort, and the term x
(T)Mx(T) a measure of terminal control accuracy.

A control problem where the objective is to maintain the state close to the zero-
state is referred to as a regulator problem and hence we refer to the above
problem as the optimal linear quadratic regulator (LQR) problem. Thus, we
rewrite the problem as a minimization of control accuracy as measured by
J =
_
T
0
x
Qxdt +x
(T)Mx(T)
subject to an explicit control eort constraint of the form
_
T
0
u
Rudt = 1.
Sec. 3.4. LINEAR QUADRATIC PROBLEM 95
When the system is time-invariant and the optimization interval is innite,
i.e., T appriaches innity we refer to this special problem as the steady state
LQR problem. For simplicity, here we only consider the steady-state regulator
problem.
Solution of the steady-state regulator problem It will turn out that the
equation below will be an important ingredient in the solution of the optimiza-
tion problem. In fact let P denote the positive denite solution of
A
P +PA+QPBR
1
B
P = 0 (Algebraic Riccati Equation(ARE)

and the optimal control us given by
u
(t) = R
1
B
Px(t).
Existence and Stability of the Steady-State LQR Solution Given
the LQR problem with M = 0, R > 0, Q = D
D, where the pair (A, D) is

observable and the pair (A, B) is controllable, it follows that a solution to the
steady state LQR problem exsits, in particular that there exists a unique positive
denite solution to the ARE, P, and that the optimal closed loop system
x = (ABK)x, K = R
1
B
P
is asymptopically stable.
3.4.1 Selection of R, Q, and M Matrices
We have assumed so far that the LQR performance measure matrices, (R, Q,
M), are given. It is in general dicult to translate time and frequency domain
specications into specic values of LQR performance measure matrices. A
simple guideline is to select these matrices to be diagonal and to make the
diagonal entry positively large for any variable we would like to be small
in the time domain. This is a rather crude rule and some tuning of these
matrices may be needed depending upon the purpose. However, an important
advantage of the optimal LQR solution is that any tuning of the performance
measure matrices will preserve asymptotic stability and robustness even if LQ
optimality is sacriced.
3.4.2 Robust Stability
We model gain and phase perturbations at the input to the plant by inserting
a constant matrix transfer function L between x in the gure and the plant
input. The unperturbed case is then represented by L = I. Then we have the
following robust stability result.
Theorem 3.5 If the nominal closed loop system is stable, and the input per-
turbation matrix L is nonsingular, then the perturbed closed loop remains stable
i L satises the inequality:

_
L
1
I
_
< (I +H(j)) , (3.11)
K (sI A)
1
B +
6
- -
x
Figure 3.5. State Feedback Conguration
where () denotes the largest singular value, and () denotes the smallest sin-
gular value of a matrix (), and
H(s) = K(sI A)
1
B.
Remark 3.9. It is also known that when K is the optimal steady state LQR
feedback gain we have the (Kalman) inequality
1 (I +H(j)) , (3.12)
which then implies that the inequality

_
L
1
I
_
< 1 (3.13)
will guarantee the satisfaction of condition (3.11).
3.4.3 Gain and Phase Margins
Let the perturbation matrix L be diagonal with entries l
i
:
L =
_
_
l
1
l
2
.
.
.
.
.
.
_
_
then (L
1
I) is also diagonal and has singular values given by
i
_
L
1
I
_
=
1
l
i
1
.
Then the robust stability condition on L (see (3.13)) becomes
1
l
i
1
< 1. (3.14)
To study gain margin, let l
i
be real, then from (3.14) we have
1 <
1
l
i
1 < 1
Sec. 3.4. LINEAR QUADRATIC PROBLEM 97
which implies that
1
2
< l
i
< .
It means an increasing gain margin and a decreasing gain margin of 1/2 in each
channel.
To study the phase margin, let l
i
= e
j
i
, then (3.14) is equivalent to
e
j
i
1
2
= (cos
i
1)
2
+ (sin
i
)
2
< 1
which implies that
cos
i
>
1
2
or
60
<
i
< 60
.
It means a phase margin in each channel of 60 degrees.
Chapter 4
REGULATOR PROBLEM II
In this chapter we give some approaches to the regulator problem that are
alternatives to the state feedback observer approaches of the last chapter. In
particular these are direct approaches to output feedback stabilization without
constructing observers.
4.1 POLE PLACEMENT COMPENSATORS
Let the plant equation be
x = Ax +Bu
y = Cx measured output
Let the compensator equation be
x
c
= A
c
x
c
+B
c
y
u
c
= C
c
x
c
+D
c
y
Therefore, the plant and compensator transfer functions are
G
p
(s) = C(sI A)
1
B
G
c
(s) = C
c
(sI A
c
)
1
B
c
+D
c
.
We would like to design (A
c
, B
c
, C
c
, D
c
).
Design Problem: Choose (A
c
, B
c
, C
c
, D
c
) such that the closed loop eigenvalues
assume the desired values. (General Pole Placement Compensators)
What are the closed loop eigenvalues?
_
x
x
c
_
=
_
A+BD
c
C BC
c
B
c
C A
c
_
. .
A
cl
_
x
x
c
_
+
_
B
0
_
v
99
100 REGULATOR PROBLEM II Ch. 4
where v + u
c
= u. If the compensator order is n
c
, n + n
c
eigenvalues will be
placed in the prescribed locations.
How do we nd (A
c
, B
c
, C
c
, D
c
)?
4.1.1 Single Input System
x = Ax +bu
y = Cx
1. Let p
be the smallest integer j such that

Rank
_
_
C
CA
.
.
.
CA
j
_
_
= n(full rank)
p
exists and is unique if (c, A) is observable, which we assume. Add p

(p p
) integrators to the system as follows:

u = u
1
u
1
= u
2
u
2
= u
3
.
.
.
u
p
= v
Introduce the extended state vector
x :=
_
_
x
u
1
u
2
.
.
.
u
p
_
_
and consider the augmented system
x =

A x +
bv
where
A =
_
_
A b 0 0
0 0 1
.
.
.
.
.
. 0 1
.
.
.
.
.
.
.
.
. 0
.
.
. 1
0 0
_
b =
_
_
0
.
.
.
.
.
.
.
.
.
0
1
_
_
Sec. 4.1. POLE PLACEMENT COMPENSATORS 101
We note that (

A,
b) is controllable when (A, b) is as we assume.

2. Find the state feedback control
v =

f x
so that

A +

b
f has a prescribed set of n + p eigenvalues in the LHP.

Write
f
..
1n+p
=
_
f
0
..
1n
f
1
f
2
f
p
_
3. Introduce the feedback compensator with transfer function (input y; output
u).
C(s) =

0
+
1
s + +
p
s
p
s
p
p1
s
p1

1
s
0
where C(s) is 1 m,
i
are 1 m, and
i
are scalar. A state space
realization of this compensator is
A
c
=
_
_
0
0
1
1
0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 1
p1
_
_
B
c
=
_
0
+
0
1
+
1
p
.
.
.
.
.
.
p1
+
p1
p
_
_
c
c
=
_
0 0 0 1

D
c
=
p
.
4. Calculate the
i
from the equation
_

0

1

p
_
C
CA
.
.
.
CA
p
_
_
= f
0
.
This equation has a solution because of the observability assumption.
5. Calculate the
i
from the equations
p
cA
p1
b +
p1
cA
p2
b + +
1
cb +
0
= f
1
p
cA
p2
b +
p1
cA
p3
b + +
2
cb +
1
= f
2
.
.
.
p
cb +
p1
= f
p
This determines all the compensator parameters. It remains to be proved
that the closed loop poles (or eigenvalues) are exactly equal to the set
of n +p eigenvalues that were assigned in Step 2.
Theorem 4.1 The compensator designed above assigns the closed loop eigen-
values to the set .
Proof. We will show that the closed loop system matrix A
cl
is similar to
A+
f. Let us write out

A
cl
=
_
A+bD
c
C bc
c
B
c
C A
c
_
=
_
_
A+b
p
C
.
.
. 0 0 0 b
(
0
+
0
p
)C
.
.
. 0 0 0
0
(
1
+
1
p
)C
.
.
. 1
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
(
p1
+
p1
p
)C
.
.
. 0 0 1
p1
_
4.1.2 Single Output System

1. Let q
be the least integer j such that

Rank
_
B AB A
2
B A
j
B
= n
q
exists and is unique when (A, B) is controllable. The compensator

order can be any q q
.
2. Form the pair ( c,

A),

A IR
(n+q)(n+q)
, c IR
1(n+q)
with
A =
_
_
A 0 0
c 0 0
0 1
.
.
.
0 0 1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 1 0
_
_
c =
_
0 0 0 1

( c,

A) is observable when (c, A) is.
3. Using pole placement, nd
L =
_
_
L
0
l
1
.
.
.
l
q
_
_
so that

A +

L c has the desired set of n + q eigenvalues in the left half
of the complex plane.
4. Consider the feedback compensator with transfer function
0
+
1
s + +
q
s
q
s
q
q1
s
q1

1
s
0
i
IR
v
,
i
is scalar
A state space realization of this is
A
c
=
_
_
0 1 0 0
.
.
. 1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 1
0

1

q1
_
_
b
c
=
_
_
0
0
.
.
.
0
1
_
_
C
c
=
_

0
+
0
q

1
+
1
q

q1
+
q1
D
c
=
q
5. Calculate
i
, i = 0, 1, , q from
_
B AB A
q
B
1
.
.
.
q
_
_
= L
0
This equation has a solution because the matrix on the left has full column
rank.
6. Calculate the
i
, i = 0, 1, , q 1 from
cB
1
+cAB
2
+ +cA
q1
B
q
+
0
= l
1
cB
2
+ +cA
q2
B
q
+
1
= l
2
.
.
.
cB
q
+
q1
= l
q
This determines all the compensator parameters.
Theorem 4.2 The compensator designed by the above procedure assigns the
closed loop eigenvalues to .
Proof. BY construction we have that the eigenvalues of

A +

L c equal to .
We will show that

A+

L c is similar to the closed loop matrix A
cl
. For this write
out A
cl
:
A
cl
=
_
_
A+B
q
c B(
0
+
0
q
) B(
1
+
1
q
) B(
q1
+
q1
q
)
0 0 1 0 0
0 0 0 1
.
.
.
.
.
.
.
.
.
0
.
.
. 1
c
0

1

q1
_
A+

L c =
_
_
A 0 0 0 L
0
c 0 0 0 l
1
0 1
.
.
. l
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0
.
.
.
0 0 1 l
q
_
_
Now it can be veried that
QA
cl
= (

A+

L c)Q
where
Q =
_
_
I
n
P
1
P
2
P
q
0 l
2
l
3
l
q
1
0 l
3
l
q
0 0
.
.
.
.
.
.
.
.
. l
q
.
.
.
0 1 0 0
_
_
P
i
= B
i
AB
i+1
A
qi
B
q
, i = 1, 2, , q.
Example 4.1.
A =
_
_
0 1 1 0
1 0 0 0
0 0 1 0
0 1 0 1
_
_
b =
_
_
0
0
1
0
_
_
c =
_
0 0 0 1

Since n = 4, we need to add n 1 = 3 integrators.
u
1
= u, u
1
= u
2
, u
2
= u
3
, u
3
= v
_
_
x
u
1
u
2
u
3
_
_
=
_
_
A b 0 0
0 0 1 0
0 0 0 1
0 0 0 0
_
_
. .
A
_
_
x
u
1
u
2
u
3
_
_
+
_
_
0
0
0
1
_
_
. .
b
v
v = f
0
x +f
1
u
1
+f
2
u
2
+f
3
u
3
=
_
f
0
f
1
f
2
f
3
. .
f
_
_
x
u
1
u
2
u
3
_
_
and
f
0
=
_
f
01
f
02
f
03
f
04
.
Now suppose we assign the closed eigenvalues at 1, 2, 3, 4, 5, 6, and
7. Then
L =
_

b

A
b

A
2
b

A
6
b
L
1
=
_
_
the last row
_
_
q
T
1
=
_
_
q
q

A
.
.
.
q

A
6
_
A = T
1

AT

b = T
1
A+
f =
_
_
0 1 0 0 0 0 0
0 0 1 0 0 0 0
0 0 0 1 0 0 0
0 0 0 0 1 0 0
0 0 0 0 0 1 0
0 0 0 0 0 0 1
f
01

f
02

f
03
1 +

f
04
2

f
1

f
2
2 +

f
3
_
_
.
Once we nd

f, we have
f =

fT
1
=
_
f
0
f
1
f
2
f
3
.
Now let the controller transfer function be
C(s) =

0
+
1
s +
2
s
2
+
3
s
3
s
3
2
s
2
1
s
0
then solve the following matrix equation to get the numerator coecits.
_

0

1

2

3
_
c
cA
cA
2
cA
3
_
_
=
_
f
01
f
02
f
03
f
04
= f
0
.
The denominator coecients are obtained from the following.
0
+
1
cb +
2
cAb +
3
cA
2
b = f
1
1
+
2
cb +
2
cAb = f
2
2
+
3
cb = f
3
.
Finally, we have
C(s) =
15768 7466s 12398s
2
12154s
3
s
3
+ 30s
2
+ 382s + 2722
.
Example 4.2.
A =
_
_
0 1 0
1 0 1
0 1 0
_
_
B =
_
_
1 0
0 0
0 1
_
_
c =
_
1 0 0

Find a pole placement compensator so that all closed loop poles are assigned at
2.
Since
Rank
_
B AB
= 3,
the order of compensator is 1. Thus, we form the argument system
A =
_
_
0 1 0 0
1 0 1 0
0 1 0 0
1 0 0 0
_
_
c =
_
0 0 0 1

.
Since (c, A) is observable, ( c,

A) is onservable. Now let us nd L so that (

A+
L c) = .
A+L c =
_
_
0 1 0 0
1 0 1 0
0 1 0 0
1 0 0 0
_
_
+
_
_
l
01
l
02
l
03
l
1
_
_
_
0 0 0 1

=
_
_
0 1 0 l
01
1 0 1 l
02
0 1 0 l
03
1 0 0 l
1
_
_
We found
L =
_
_
26
24
42
_
_
_
L
0
8
_
L
1
_
_
.
Sec. 4.2. PARAMETRIZATION OF ALL STABILIZING CONTROLLERS 107
Now let the compensator transfer function be
C(s) =

0
+
1
s
s
0
.
We solve the following matrix equation
_
B AB
_

0
1
_
=
_
_
1 0 0 0
0 0 1 1
0 1 0 0
_
_
_
01
02
11
12
_
_
=
_
_
26
48
42
_
_
= L
0
.
Thus we have
0
=
_
26
42
_

1
=
_
1
47
_
.
Now we compute
0
cB
1
+
0
=
_
1 0 0

_
_
1 0
0 0
0 1
_
_
_
1
47
_
+
0
= L
1
= 8.
Thus, we have
0
= 7.
A state space representation of the compensator is
A
c
= 7, b
c
= 1, C
c
=
0
+
0
1
=
_
19
87
_
, D
c
=
_
1
47
_
.
The composite system is
_
x
x
c
_
=
_
A+B
c
D
c
C BC
c
B
c
C A
c
_ _
x
x
c
_
=
_
_
1 1 0 19
1 0 1 0
47 1 0 287
1 0 0 7
_
_
_
x
x
c
_
.
4.2 PARAMETRIZATION OF ALL STABILIZING CONTROLLERS
Given a plant with transfer function P(s), we would like to have a useful
parametrization of all feedback controllers C(s) that stabilize P(s). Such a
parametrization was obtained by Youla, Bongiorno, and Jabr (1976), and inde-
pendently by Kucera (1979), and it is called the YBJK parametrization. We
begin with the single input single output case.
4.2.1 Scalar Case: P(s) = p(s)
Let
p(s) =
n(s)
d(s)
where n(s), d(s) are polynomials. If n is the highest degree of n(s) and d(s) we
introduce
f(s) =
n(s)
(s + 1)
n
g(s) =
d(s)
(s + 1)
n
and observe that
p(s) =
f(s)
g(s)
where f(s), g(s) are stable proper transfer functions. This prompts us to intro-
duce the following set.
Notation and Preliminaries
Let IR be real numbers, IR[s] be polynomials in s with coecients in IR, IR(s)
be the eld of fractions associated with IR[s], i.e., IR(s) is the set of all rational
functions of s, namely ratios of polynomials, equivalently IR(s) is the set of
transfer functions of a linear, time-invariant systems (without delay).
S IR(s) is the set of all stable (i.e., all poles in the open left half plane
I C
), proper (bounded at innity, i.e., no poles at innity) rational function. A

rational function p(s) S if and only if it is the transfer function of a bounded
input bounded output (BIBO) stable system. Let I C
+
denote the closed right
half plane (RHP) and I C
+
e
the extended RHP, i.e., I C
+
with the point at . Then
p(s) S
if and only if p(s) has no poles in I C

+
e
.
Observe that p
1
(s) S
, p
2
(s) S
implies that
p
1
(s) +p
2
(s) S
and
p
1
(s)p
2
(s) S
and S
has the structure of a commutative ring with the units (i.e., invertible
elements) being those functions which have no zeros in I C
+
e
. Therefore, the
units are rational functions which have the same degree for numerator and
denominator with poles and zeros in the LHP. We can make S
into a Euclidean
domain by dening a degree function as follows.
If f(s) S
, let the degree of f(s) be dened by

(f) = number of zeros of f in I C
+
e
= number of zeros of f in I C
+
+relative degree of f.
So if
f(s) =
(s)
(s)
=

+
(s)
(s)
(s)
with
+
(s) containing all the unstable factors, then
(f) = degree
_
+
_
+ degree () degree ()
. .
relative degree of f
= degree
_
+ degree [] degree
_
+
_
+ degree
_
_
. .
degree ()
= degree () degree
_
_
where degree of polynomial is the usual degree.
To show that S
is a Euclidean domain, we need to show that we can carry

out division in S
, just as we do Euclidean division in IR[s] with polynomials.

For a polynomial f(s), let (f) denote the degree. If f(s), g(s) are polynomials,
then there exist polynomials q(s), r(s) such that
f(s) = g(s)q(s) +r(s)
with
(f gq) < (r).
Similarly, we need to show that for f(s) S
, g(s) S
(g(s) = 0), there exists

q(s), r(s) S
with
f(s) = g(s)q(s) +r(s)
and
(f gq) < (r).
To establish this we need the following.
Lemma 4.1 Let , , w IR[s] with , coprime. Then there exists , IR[s]
with
degree () < degree ()
such that
+ = w.
Proof. Since , are coprime, there exist
1
,
1
IR[s] such that
1
+
1
= 1.
Multiplying by w, we have
1
w
..
2
+
2
w
..
2
= w
2
+
2
= w.
We can also write
..
(
2
) +
..
(
2
+) = w
for any IR[s]. By choice of we can make the degree of
2
less than
that of .
Now suppose f,g S
(g = 0). We construct q, r S
such that
(f gq) < (g).
We take (0) = . If (g) = 0, then g is a unit and we take q = fg
1
and so
f gq = 0 and
(0) = < (g).
Now suppose
n = (g) = 0, i.e., n = (g) > 0.
Let
g(s) =

g
(s)
g
(s)
,
g
,
g
IR[s] coprime.
Factor
g
(s) =
+
g
(s)
g
(s),
with the zeros of
+
g
I C
+
,
g
I C
. Now write
g(s) =

+
g
(s)
(s + 1)
n
(s)(s + 1)
n
g
(s)
. .
u(s)
=

+
g
(s)
(s + 1)
n
u(s).
Now
(u) = (
g
) (
) n
= (
g
) (
)
_
(
g
) (
)
_
= 0
so that u(s) is a unit of S. Now let
f(s) =

f
(s)
f
(s)
with
f
,
f
coprime polynomials. Since f(s) S
the zeros ,of

f
(s) lie in
I C
, so that
f
(s) and
+
g
(s) are coprime. By the Lemma, we can nd (s),
(s) IR[s] so that
() < (
f
)
and
+
g
(s)(s) +
f
(s)(s) =
f
(s)(s + 1)
n1
.
Dividing by
f
(s)(s + 1)
n1
we have
+
g
(s)
(s + 1)
n
(s)(s + 1)
f
(s)
+
(s)
(s + 1)
n1
=

f
(s)
f
(s)
or
+
g
(s)
(s + 1)
n
u(s)
. .
g(s)
(s)(s + 1)
u(s)
f
(s)
. .
q(s)
+
(s)
(s + 1)
n1
. .
r(s)
=

f
(s)
f
(s)
. .
f(s)
.
Since
() < (
f
),
we have
((s + 1)) (
f
)
so that q(s) is proper. Also q(s) is stable so q(s) S
. Therefore, also
r = f gq S
.
Since r is proper, () n 1. Therefore,
(r) = n 1 () +(
+
)
= n 1 (
)
< (g) = n.
Therefore, we have found
r S
with
(r = f gq) < (g).
In S
, f divides g if and only if every I C

+
e
zeros of f is also a zero of g of the
same multiplicity at least. Two functions f and g in S
are coprime if and only

if they have no common zeros in the extended RHP (I C
+
e
). Equivalently, there
exists , such that
f +g = 1.
Controller Parametrization: Scalar Case
c p
- - - - -
?
6
1
e
1
y e
2
y
2
2
+
+
+
Figure 4.1. Feedback System
_
e
1
e
2
_
=
_

1
2
_
_
0 p
c 0
_ _
e
1
e
2
_
_
e
1
e
2
_
=
_
1 p
c 1
_
1
_

1
2
_
=
_
_
1
1 +pc
p
1 +pc
c
1 +pc
1
1 +pc
_
_
. .
H(p,c)
_

1
2
_
Assume that
1 +pc = 0 s.
Now c stabilizes p (or (p, c) is stable) if and only if each element of H(p, c) is in
S, i.e.,
H(p, c) S
22
.
Lemma 4.2 Let p, c IR(s) with coprime factorization
p =
n
p
d
p
c =
n
c
d
c
where
n
p
, d
p
, n
c
, d
c
S
and dene
(p, c) = d
c
d
p
+n
c
n
p
S.
Then (p, c) is stable if and only if (p, c) is a unit of S.
Proof. IF part Suppose (p, c) is a unit. Then
1
(p, c)
S.
Now
H(p, c) =
1
(p, c)
_
d
p
d
c
n
p
d
c
n
c
d
p
d
p
d
c
_
so
H(p, c) S
22
.
Only IF part Suppose H(p, c) S
22
. Then
1 +pc = 0 and d
p
= 0 and d
c
= 0.
And also
(p, c) = d
c
d
p
(1 +pc) = 0
so the above formula holds. Then
d
p
d
c
(p, c)
,
n
c
d
p
(p, c)
, and
n
p
d
c
(p, c)
, are all in S.
Also
1
d
p
d
c
(p, c)
=
n
p
n
c
(p, c)
S.
Therefore,
_
d
p
n
p
_
1
(p, c)
_
d
c
n
c
S
22
.
Since (d
p
, n
p
) is coprime and (d
c
, n
c
) is coprime, we have (x
p
, y
p
) and (x
c
, y
c
)
in S satisfying
x
p
n
p
+y
p
d
p
= 1
x
c
n
c
+y
c
d
c
= 1.
Therefore,
_
y
p
x
p
_
d
p
n
p
_
1
(p, c)
_
d
c
n
c
_
y
c
x
c
_
S
or
1
(p, c)
S
which shows that (p, c) is a unit of S.
Corollary 4.1 Suppose
p =
n
p
d
p
with n
p
, d
p
S and coprime. Then C(s) IR(s) stabilizes p if and only if
c =
n
c
d
c
for some (n
c
, d
c
) S which satisfy
n
c
n
p
+d
c
d
p
= 1.
Proof. IF part If
n
c
n
p
+d
c
d
p
= (p, c) = 1
then by the previous lemma, we have stability.
Only IF part Suppose
c =
n
1
d
1
, n
1
, d
1
S
stabilizes p. Then
n
p
n
1
+d
p
d
1
=
1
(p, c)
is a unit. Then
n
c
=
n
1
1
(p, c)
, d
c
=
d
1
1
(p, c)
S
and satisfying
n
p
n
c
+d
p
d
c
= 1.
The main result follows from this.

Theorem 4.3 Suppose
p IR(s), p =
n
p
d
p
,
(n
p
, d
p
) S and coprime. Let x, y S be such that
xn
p
+yd
p
= 1.
Then the set of all stabilizing compensators, denoted S(p), is given by
S(p) =
_
c =
x +rd
p
y rn
p
: r S, y rn
p
= 0
_
.
Proof. IF part
c =
x +rd
p
y rn
p
then
(p, c) = (x +rd
p
)n
p
+ (y rn
p
)d
p
= xn
p
+yd
p
= 1
is a unit, so c is stabilizing.
Only IF part If c is stabilizing, by the Corollary
c =
n
c
d
c
for some n
c
, d
c
S for which
n
c
n
p
+d
c
d
p
= 1. (4.1)
We now show that every solution of (4.1) is of the form
(n
c
, d
c
) = (x +rd
p
, y rn
p
).
Now
xn
p
+yd
p
= 1 (4.2)
Then
(n
c
x)n
p
+ (d
c
y)d
p
= 0
and
(n
c
x)n
p
= (y d
c
)d
p
.
Since n
p
, d
p
are coprime, it follows that d
p
divides n
c
x and n
p
divides y d
c
.
Let
n
c
x
d
p
=
y d
c
n
p
= r.
Then
n
c
= x +rd
p
d
c
= y rn
p
.
Example 4.3. Consider

p(s) =
s
(s + 1)(s 1)
=
s
(s + 1)
2
(s + 1)(s 1)
(s + 1)
2
.
Then
n
p
(s) =
s
(s + 1)
2
d
p
(s) =
s 1
s + 1
and n
p
(s), d
p
(s) S. To parametrize all stabilizing compensators, we rst need
one compensator. Let
c(s) =
2(s + 2)
s 0.5
.
Then
c(s) =
2(s + 2)
(s + 1)
(s 0.5)
(s + 1)
=
n
1
(s)
d
1
(s)
and
(s, p, c) = n
p
(s)n
1
(s) +d
p
(s)d
1
(s)
=
(s)
..
s
3
+ 1.5s
2
+ 3s + 0.5
(s + 1)
3
=
(s)
(s + 1)
3
which is a unit since the numerator and denominator are of the same degree
and Hurwitz. Now letting
x(s) =
n
1
(s)
(s, p, c)
=
2(s + 2)
(s + 1)
(s + 1)
3
(s)
=
2(s + 2)(s + 1)
2
(s)
S
y(s) =
d
1
(s)
(s, p, c)
=
(s 0.5)
(s + 1)
(s + 1)
3
(s)
=
(s 0.5)(s + 1)
2
(s)
S,
we have
x(s)n
p
(s) +y(s)d
p
(s) = 1.
Now it follows that all compensators that stabilizes p(s) are of the form
c(s) =
x(s) +r(s)d
p
(s)
y(s) r(s)n
p
(s)
, r(s) S
=
2(s + 2)(s + 1)
4
+r(s)(s 1)(s + 1)(s)
(s 0.5)(s + 1)
4
r(s)s(s)
=
2(s + 2)(s + 1) +r(s)(s 1)(s + 1)(s)
(s 0.5)(s + 1) r(s)s(s)
.
Since (s) = (s, p, c) is a unit, as r(s) varies over S so does r(s)(s). So we
replace r(s)(s) by q(s) S and
c(s) =
2(s + 2)(s + 1) + (s 1)(s + 1)q(s)
(s 0.5)(s + 1) sq(s)
, q(s) S
is the parametrization.
The closed loop transfer matrix
H(p, c) =
1
(p, c)
_
d
p
(s)d
c
(s) n
p
(s)d
c
(s)
d
p
(s)n
c
(s) d
p
(s)d
c
(s)
_
=
1
(p, c)
_
d
p
(s) (y(s) r(s)n
p
(s)) n
p
(s) (y(s) r(s)n
p
(s))
d
p
(s) (x(s) +r(s)d
p
(s)) d
p
(s) (y(s) r(s)n
p
(s))
_
and
d
p
(s) (y(s) r(s)n
p
(s)) =
(s 1)
(s + 1)
_
(s 0.5)(s + 1)
2
(s)

r(s)s
(s + 1)
2
_
=
(s 1)(s 0.5)(s + 1)
(s)

s(s)r(s)(s 1)
(s)(s + 1)
3
=
(s 1) [(s 0.5)(s + 1) q(s)s(s 1)]
(s)
where q(s) = r(s)(s) S and is arbitrary. Similarly,
h
p
(s) (y(s) r(s)n
p
(s)) =
s [2(s + 2) +q(s)(s 1)]
(s)
d
p
(s) (x(s) +r(s)d
p
(s)) =
s [2(s + 2) +q(s)(s 1)]
(s)
.
4.2.2 Parametrization of Stabilizing Controllers: Multivariable
Case
We rst extend the matrix fraction description, originally given with factoriza-
tion over the polynomial matrices to factorization with matrices whose elements
are in S (S
mn
or M(s)). Write
P(s) = N
R
(s)D
1
R
(s) = D
1
L
(s)N
L
(s)
with elements of N
R
(s), D
R
(s), N
L
(s), D
L
(s) in S. This is easily done by
dividing the polynomial factorization by (s + 1)
n
, say.
To introduce coprimeness, we note that the unimodular matrices over S
mn
are those matrices U(s) with
|U(s)| = u(s) a unit in S,
Given D
R
(s), N
R
(s), write
_
D
R
(s)
N
R
(s)
_
and perform elementary operation so that
_
U
11
(s) U
12
(s)
U
21
(s) U
22
(s)
_ _
D
R
(s)
N
R
(s)
_
=
_
R(s)
0
_
.
Then R(s) is a gcrd and N
R
(s)(s), D
R
(s)(s) are right coprime.
Now, we start with
P(s) = N
p
(s)D
1
p
(s) =

D
1
p
(s)

N
p
(s)
with (N
p
(s), D
p
(s)) right coprime, (

D
p
(s),

N
p
(s)) left coprime over M(S). Then
it is possible to show that there exist X(s), Y (s) M(s) such that
X(s)N
p
(s) +Y (s)D
p
(s) = I
and

X(s),

Y (s) M(S) such that
N
p
(s)

X(s) +

D
p
(s)
Y (s) = I
or more compactly,
_
Y (s) X(s)

N
p
(s)

D
p
(s)
_
. .
P
_
D
p
(s)

X(s)
N
p
(s)

Y (s)
_
. .
Q
=
_
I 0
0 I
_
which also shows that P and Q must, in addition, be unimodular.
Computation of N
p
(s), D
p
(s),

N
p
(s),

D
p
(s), X(s), Y (s),

X(s),

Y (s)?
Theorem 4.4 Let
P(s) = C(sI A)
1
B +E
and let (A, B) be stabilizable, (C, A) detectable. Find F and K so that
A
0
:= ABF

A
0
:= AKC
are both stable. Then
N
p
(s) = C(sI

A
0
)
1
(B KE) +E
D
p
(s) = I C(sI

A
0
)
1
K
N
p
(s) = (C EF)(sI A
0
)
1
B +E
D
p
(s) = I F(sI A
0
)
1
B
X(s) = F(sI

A
0
)
1
K
Y (s) = I +F(sI

A
0
)
1
(B KE)
X(s) = F(sI A
0
)
1
K
Y (s) = I + (C EF)(sI A
0
)
1
K
Proof. We prove fort the case E = 0 rst. First show that
P(s) = N
p
(s)D
1
p
(s) =

D
1
p
(s)

N
p
.
Use the matrix identity
T(I +UT)
1
= (I +TU)
1
T,
we have
P(s) = C(sI A)
1
B
= C(sI A
0
BF)
1
B
= C
__
I BF(sI A
0
)
1
(sI A
0
)
_
1
B
= C(sI A
0
)
1
_
I BF(sI A
0
)
1
1
B
= C(sI A
0
)
1
B
. .
N
p
(s)
_
_I F(sI A
0
)
1
B
. .
D
p
(s)
_
_
1
.
Similarly, we can prove
P(s) =

D
p
(s)

N
p
(s).
Now we show
N
p
(s)

X(s) +

D
p
(s)
Y (s) = I.
N
p
(s)

X(s) +

D
p
(s)
Y (s) = C(sI

A
0
)
1
BF(sI A
0
)
1
K
+
_
I C(sI

A
0
)
1
K
_
_
I +C(sI A
0
)
1
K
= C(sI

A
0
)
1
BF(sI A
0
)
1
K
+I C(sI

A
0
)
1
K +C(sI A
0
)
1
K
C(sI

A
0
)
1
KC(sI A
0
)
1
K.
But
C(sI

A
0
)
1
_
BF (sI A
0
) + (sI

A
0
) KC
_
. .
=0
(sI A
0
)
1
K = 0.
Therefore,
N
p
(s)

X(s) +

D
p
(s)
Y (s) = I.
Other verications are similar.
Closed Loop Stability
C P
- - - - -
?
6
1
e
1
y
1
e
2
y
2
2
+
+
+
Figure 4.2. Feedback System
_
e
1
e
2
_
=
_

1
2
_
_
0 P
C 0
_ _
e
1
e
2
_
_
y
1
y
2
_
=
_
C 0
0 P
_ _
e
1
e
2
_
or
y = Ge.
_
0 P
C 0
_
=
_
0 I
I 0
_
. .
F
_
C 0
0 P
_
. .
G
e = FGe, y = Ge
e = (I +FG)
1
.
Easily shown that
|I +FG| = |I +PC| = |I +CP|
and the system is wellposed if this determinant is nonzero.
e = (I +FG)
1
. .
H(P,C)
H(P, C) =
_
(I +PC)
1
P(I +CP)
1
C(I +PC)
1
(I +CP)
1
_
.
Using
(I +PC)
1
= I P(I +CP)
1
C (4.3)
C(I +PC)
1
= (I +CP)
1
C, (4.4)
we also get
H(P, C) =
_
I P(I +CP)
1
C P(I +CP)
1
(I +CP)
1
C (I +CP)
1
_
=
_
(I +PC)
1
(I +PC)
1
P
C(I +PC)
1
I C(I +PC)
1
P
_
.
Denition 4.1. The closed loop system is stable i H(P, C) M(S).
Coprime Representation
Consider left and right coprime factorizations:
P(s) = N
P
(s)D
1
P
(s) =

D
1
P
(s)

N
P
(s)
C(s) = N
C
(s)D
1
C
(s) =

D
1
C
(s)

N
C
(s).
Then
1
(s) =

D
C
(s)D
P
(s) +

N
C
(s)N
P
(s)
2
(s) =

D
P
(s)D
C
(s) +

N
P
(s)N
C
(s)
qualify as characteristic equations in the following sense.
Theorem 4.5 H(P, C) is stable i
1
(s) or
2
(s) is unimodular over S.
Proof.
1
(s) =

D
C
(s)D
P
(s) +

N
C
(s)N
P
(s)
=

D
C
(s) (I +C(s)P(s)) D
P
(s).
Thus, since the system is well posed,
|
1
(s)| = |
D
C
(s)| |I +C(s)P(s)| |D
P
(s)| = 0.
Also
1
1
(s) = D
1
P
(s) (I +C(s)P(s))
1
D
1
C
(s).
H(P, C) =
_
I P(I +CP)
1
C P(I +CP)
1
(I +CP)
1
C (I +CP)
1
_
=
_
I N
P
D
1
P
(I +CP)
1

D
1
C
N
C
N
P
D
1
P
(I +CP)
1
(I +CP)
1

D
1
C
N
C
(I +CP)
1
_
=
_
I N
P
1
1
N
C
N
P
1
1
D
C
D
P
1
1
N
C
D
P
1
1
D
C
_
=
_
I 0
0 0
_
+
_
N
P
D
P
_
1
1
_

N
C

D
C
. .
L(P,C)
.
Then H(P, C) is stable i L(P, C) is stable. Now multiplying L(P, C) from the
left by
_
X
P
Y
P
and from the right by

_

X
C

Y
C
we have
_
X
P
Y
P
_
N
P
D
P
_
1
1
_

N
C

D
C
_

X
C
Y
C
_
=
1
1
M(S)
Therefore,
1
is unimodular. Similarly, for
2
.
Corollary 4.2 Given
P = N
P
D
1
P
=

D
1
C
N
C
and a stabilizing controller C, then there exist N
C
, D
C
,

N
C
,

D
C
such that
C = N
C
D
1
C
=

D
1
C
N
C
and
1
=

D
C
D
P
+

N
C
N
P
= I
2
=

D
P
D
C
+

N
P
N
C
= I.
Proof. Given any left coprime factorization

D
C1
,

N
C1
, we have
D
C1
D
P
+

N
C1
N
P
= U
1
unimodular. Then
U
1
1
D
C1
. .
D
C
D
P
+U
1
1
N
C1
. .
N
C
N
P
= I
gives a new left coprime fractorization for which
1
= I. Similarly for the other
case.
Parametrization of All Stabilizing Controllers
Theorem 4.6 Let
P(s) = N
P
(s)D
1
P
(s) =

D
1
P
(s)

N
P
(s)
be a right coprime factorization and left coprime factorization of P(s) and let
C(s) be a stabilzing controller for P(s). Let X(s), Y (s),

X(s),

Y (s) be such
that
XN
P
+Y D
P
= I

N
P

X +

D
P

Y = I.
Then, the set of stabilizing controllers is given by
S(P) =
_
(Y R

N
P
)
1
(X +R

D
P
) : R M(S), |Y R

N
P
| = 0
_
=
_
(

X +D
P
S)(
Y N
P
S)
1
: S M(S), |
Y N
P
S| = 0
_
.
Proof. C stabilizes P i C has an left coprime factorization

D
1
C
N
C
such that
D
C
D
P
+

N
C
N
P
= I.
A particular solution of this equation is
Y D
P
+XN
P
= I
and it is possible to show that the general solution is
_
Y X
R
_

N
P

D
P
where R M(S) is arbitrary. Thus follows from
D
P
N
P

N
P
D
P
= 0.
Then
__
Y X
R
_

N
P

D
P
_
D
P
N
P
_
= I.
The dual parametrization is proved similarly.
PART III
TRACKING AND
DISTURBANCE
REJECTION
Chapter 5
LINEAR
SERVOMECHANISM
PROBLEM
A servomechanism is a device in which the output automatically follows or
tracks a reference signal. The design of servomechanisms is the most funda-
mental and basic problem in control theory and generally speaking most central
problems are oshoots, renements or special cases of this problem. The regu-
lator problem of the previous sections, for instance is a special case where the
desired or reference value for the system state is the zero state. We shall develop
our theory of linear servomechanisms by rst considering the simplest cases and
then generalizing.
5.1 First Order System
Let the system or plant be of rst order and let its equations be
x = ax +bu (5.1)
y = cx (5.2)
and let r() denote the reference signal which y is required to follow. Dene
e(t) = r(t) y(t)
and form the error dierential equation
e = r cax cbu
or
e = r ar cbu +ae. (5.3)
Then clearly the control law
u =
1
cb
( r ar Ke) (5.4)
125
126 LINEAR SERVOMECHANISM PROBLEM Ch. 5
causes
e = (a +K)e
and since K is a free parameter, it follows that the rate of decay of the tracking
error can be controlled by K. We have reached the following conclusion:
Remark 5.1. If r, r, e can be measured, the control law (5.4) provides asymp-
totic tracking of any signal r().
To understand this statement a little better, we calculate the transfer function
between e and r, with
u =
0
e +
1
r +
2
r.
When
y(s) = r(s) e(s)
=
cb
s a
u(s)
=
cb
s a
[(
1
+
2
s)r(s) +
0
e(s)] ,
we have
e(s)
_

0
cb
s a
+ 1
_
=
(s a)1 cb(
1
+
2
s)
s a
or
e(s) =
(1 cb
2
)s cb
1
a
s a +
0
cb
r(s). (5.5)
Now we see that the choice
2
=
1
cb
,
1
=
a
cb
,
zeros the error transfer function numerator while
0
sets its pole. This is pre-
cisely the control law (5.4) which was obtained from the consideration of forcing
the error equation (5.3) to be free of r.
The problem of implementing the control law (5.4) is that the presence of
any high frequency noise in the signal r completely wipes out the equality in
(5.3) because of the dierentiation suered by the noise. This motivates us to
reconsider the problem by restricting the control signal to be
u = K
1
r +K
2
x. (5.6)
Then the error dierential equation is given by
e = (a +bK
1
)e + [ r (a +bK
1
+cbK
2
)r] (5.7)
and it is clear that error convergence cannot be achieved for every r(). In fact,
the error can be decoupled from r only if r is known to satisfy a rst order
dierential equation. Suppose that
r = r, (5.8)
Sec. 5.1. First Order System 127
then e is decoupled from r if
= a +bK
1
+cbK
2
. (5.9)
The condition (5.9) can also be better understood by calculating the transfer
function relating e and r. This gives
e(s) =
s a bK
1
cbK
2
s a bK
1
r(s).
From (5.8)
r(s) =
r(0)
s
,
we have
e(s) =
s a bK
1
cbK
2
(s a nK
1
)
r(0)
(s )
.
Clearly, the exponent e
t
will be absent from the output response e(t) i
(s ) | (s a bK
1
cbK
2
)
i.e.,
a bK
1
cbK
2
= 0
which is the condition (5.9). To summarize:
Remark 5.2. The control law (5.6), subject to (5.9) assigns the zero of the
error transfer function to the point where the reference signal pole is, and
thereby decouples the signal mode e
t
r(0) from e(t) for all r(0). Also K
1
assigns
the pole of e(s) freely.
5.1.1 Robust Solution
Consider the plant
x = ax +bu
y = cx
e = r y, r = r(fragile solution)
and the controller
p = p +v
u = K
1
x +K
2
p +K
3
r
v = l
1
x +l
2
p +l
3
r.
Then,
_
x
p
_
=
_
a +bK
1
bK
2
l
1
+l
2
_ _
x
p
_
+
_
bK
3
l
3
_
r (5.10)
e =
_
c 0

_
x
p
_
+ 1r.
The transfer function of the system is
_
c 0

_
s a bK
1
bK
2
l
1
s l
2
_
1
_
bK
3
l
3
_
+ 1.
Requirements:
1) The denominator mist have LHP roots.
2) Tracking: the numerator must be divisible by (s ).
Thus, the numerator at s = is
2
(a+bK
1
++l
2
+cbK
3
)+(a+bK
1
)(+l
2
)bK
2
l
1
+cbK
3
(+l
2
)cbK
2
l
3
= 0.
(5.11)
Robustness: means that (5.11) remains valid despite of changes in a, b, c of
the plant.
Let us consider the case a = a + a, then from (5.11)
a( + +l
2
) = 0.
This leads the following condition:
Condition A : = +l
2
.
Applying the condition A to (5.11), we have
2
(a +bK
1
)
2
cbK
3
+ (a +bK
1
) bK
2
l
1
+cbK
3
cbK
2
l
3
= 0.
Thus,
bK
2
l
1
cbK
2
l
3
= 0
or
Condition B : bK
2
(l
1
+cl
3
) = 0.
To satisfy the condition B,
K
2
= 0 or = 0 or l
1
+cl
3
= 0.
Recall
A
cl
=
_
a +bK
1
bK
2
l
1
+l
2
_
=
_
a +bK
1
bK
2
l
1

_
.
If we set K
2
= 0, A
cl
is unstable since
(A
cl
) = (a +bK
1
) (), > 0.
If we set = 0, the same thing happens. Therefore, the only choice is
Condition B
: l
1
+cl
3
= 0.
Therefore, the compensator must be:
p = p +v
= p +(l
1
x +l
2
p +l
3
r)
= ( +l
2
)p +l
1
x +l
3
r
= p +l
3
(r y)
= p +l
3
e.
Consider the following block diagram:
l
3
s
K
2
cb
s a
K
1
c
K
3
- - - - - -
-
?
6 6
r e u y
y
+ +
Figure 5.1. A General Servosystem
More generally, we have the following:
n
c
d
c
n
p
d
p
K
controller
- - - - -
6
-
?
plant
r e u y
y
+ +
The error transfer function becomes
e = r y = r (CPe +KPr)
e(1 +CP) = r KPr.

e(s) =
1 KP
1 +CP
r(s).
We can also rewrite
e(s) =
1 K
n
p
d
p
1 +
n
c
n
p
d
c
d
p
r(s)
=
(d
p
Kn
p
)d
c
d
c
d
p
+n
c
n
p
r(s).
Therefore,
e(s) =
(d
p
(s) Kn
p
(s)) d
c
(s)
d
c
(s)d
p
(s) +n
c
(s)n
p
(s)
r(0)
s
.
Now let us consider the following:
n
c
(s)
d
c
(s)
n
p
(s)
d
p
(s)
K
n
K
d
- - - - -
6
-
?
plant
r e u y
y
+ +
Write the transfer function
y(s) = P(s)u(s)
= P(s) (C(s)e(s) +Kr(s))
= P(s)C(s)e(s) +P(s)Kr(s) = r(s) e(s).
Thus,
e(s) =
1 P(s)K
1 +P(s)C(s)
r(s)
e(s)
r(s)
=
(d
p
(s)K
d
n
p
(s)K
n
) d
c
(s)
(d
p
(s)d
c
(s) +n
p
(s)n
c
(s)) K
d
:= E(s).
Suppose that r satises the following dierential equation:
d
p
r
dt
p
+
p1
d
p1
r
dt
p1
+
p2
d
p2
r
dt
p2
+ +
1
dr
dt
+
0
r = 0.
Taking the Laplace transform, we have
r(s) =
r
0,0
+r
1,0
s + +r
p1,0
s
p1
s
p
+
p1
s
p1
+ +
1
s +
0
=
r
0
(s)
(s)
where r
i,0
are initial conditions.
Now let us recall the error transfer function.
e(s) = E(s)r(s) = E(s)
r
0
(s)
(s)
which means that the numerator of the error transfer function E(s) must be
divided by (s). Therefore, we summarize the following.
Summary The compensator must be chosen, equivalently choose n
c
(s),
d
c
(s), K
n
(s), K
d
(s) such that
1) the denominator of E(s) must be stable
2) the numerator of E(s) must be divisible by (s).
Example 5.1. Let
P(s) =
10(s + 1)
(s + 2)(s + 3)
.
Requirements:
1) Track step input,
2) Assign the closed loop poles arbitrarily.
Assuming K = 0 (i.e., k
n
= 0, k
d
= 1),
E(s) =
dp(s)d
c
(s)
d
p
(s)d
c
(s) +n
p
(s)n
c
(s)
.
We can see that in order to track the step input, d
c
(s) must have s term.
Thus, we set the controller to be d
c
(s) = sh(s).
The closed loop characteristic polynomial is
d
c
(s)d
p
(s) +n
c
(s)n
p
(s) = s(s + 2)(s + 3) + (k
1
+k
2
s) (10(s + 1))
= s
3
+ (5 + 10k
2
)s
2
+ (6 + 10k
1
+ 10k
2
)s + 10k
1
.
k
1
+k
2
s
s
- -
e u
Figure 5.4. First Order Controller
k
1
+k
2
s +k
4
s
2
s(s +k
3
)
- -
e u
Figure 5.5. Second Order Controller
We certainly cannot assign the closed loop poles arbitrarily since it has 2 design
parameters and 3 coecients to be assigned. So, we increase the order of the
controller by one. Consider the following:
The characteristic polynomial is
d
c
(s)d
p
(s) +n
c
(s)n
p
(s) =
s
4
+(5+k
3
+10k
4
)s
3
+(6+10k
4
+10k
2
+5k
3
)s
2
+(6k
3
+10k
1
+10k
2
)s+10k
1
= 0.
Since we now have 4 design parameters and 4 coecients to assign, we can
assign the closed loop poles arbitrary.
Remark 5.3. Observer the error transfer function
E(s) =
d
c
(s)d
p
(s)
n
c
(s)n
p
(s) +d
c
(s)d
p
(s)
In order to track e
t
, d
c
(s) must have the factor (s ). Therefore, if n
p
(s)
has the factor (s ), the system cannot be made to track the signal because
the characteristic polynomial has the factor (s) and it is a unstable factor.
5.2 MIMO Servomechanism
Plant:
x = Ax +Bu +E
y
t
= Cx +Du +F. y
m
= C
m
x (measurements)
where is the disturbance vector, r is the vector of reference signals, and
e := r y
t
.
Let each component of the disturbance and each component of the reference
satisfy the dierential equation
(D)
i
(t) = 0 i = 1, 2, , q
(D)r
j
(t) = 0 j = 1, 2, , m
Sec. 5.2. MIMO Servomechanism 133
where
(D) = D
p
+
p1
D
p1
+ +
1
D +
0
.
The servocompensator is:
x
c1
= A
c1
x
c1
+b
c1
e
1
x
c2
= A
c2
x
c2
+b
c2
e
2
.
.
.
x
cm
= A
cm
x
cm
+b
cm
e
m
where
A
ci
=
_
_
0 1 0
0
.
.
.
0 1
0

1

p1
_
_
b
ci
=
_
_
0
.
.
.
0
1
_
_
, i = 1, 2, , m.
It can be written as
_
_
x
c1
x
c2
.
.
.
x
c
m
_
_
. .
x
c
=
_
_
A
c1
A
c2
.
.
.
A
c
m
_
_
. .
Ac
_
_
x
c1
x
c2
.
.
.
x
c
m
_
_
. .
x
c
+
_
_
b
c1
b
c2
.
.
.
b
c
m
_
_
. .
Bc
_
_
e
1
e
2
.
.
.
e
m
_
_
. .
e
.
Therefore,
_
x
x
c
_
. .
x
=
_
A 0
B
c
C A
c
_
. .
A
_
x
x
c
_
. .
x
+
_
B
B
c
D
_
. .
B
u +
_
E
B
c
F
_
+
_
0
B
c
_
r
e =
_
C 0

_
x
x
c
_
Du F +I
m
r.
Now let
y
m
=
_
y
m
x
c
_
=
_
C
m
0
0 I
n
c
_ _
x
x
c
_
:=

C
m
_
x
x
c
_
=

C
m
x
and consider the triple (

A,

B,

C
m
). If (

A,

B) is controllable and (

C
m
,

A) is
observable then a stabilizing compensator (A
s
, B
s
, C
s
, D
s
):
x
s
= A
s
x
s
+B
s
y
m
= A
s
x
s
+B
sm
y
m
+B
sc
x
c
u = C
s
x
s
+D
s
y
m
= C
s
x
s
+D
sm
y
m
+D
sc
x
c
can be designed so that
_
_
x

x
s
_
_
=
_
A+

BD
s
C
.
.
.

BC
s

B
s

C
.
.
. A
s
_
_
_
_
x

x
s
_
_
+
_
_
E
B
c
F

0
_
_
+
_
_
0
B
c

0
_
_
e =
_
C 0
.
.
. 0
_
_
x
x
s
_
DC
s
x
s
DD
s

C x.
Now write out
x =
_
x
x
c
_
,

A =
_
A 0
B
c
C A
c
_
,

B =
_
B
B
c
D
_
D
s
=
_
D
s
m
D
s
c
, B
s
=
_
B
s
m
B
s
c
.
Then, the composite closed loop system equations are given by
_
_
x
x
c
x
s
_
_
=
_
_
A+BD
s
m
C
m
BD
s
c
BC
s
B
c
(C +DD
s
m
C
m
) A
c
B
c
DD
c
B
c
DC
s
B
s
m
C
m
B
s
c
A
s
_
_
. .
A
cl
_
_
x
x
c
x
s
_
_
. .
x
cl
+
_
_
E
B
c
F
0
_
_
. .
B
cl
+
_
_
0
B
c
0
_
_
. .
B
clr
r
e =
_
(C +DD
sm
C
m
) DD
sc
DC
s
. .
C
cl
_
_
x
x
c
x
s
_
_
+ [F]
. .
D
cl
+ [I
m
]
..
D
clr
r
or more compactly as
x
cl
= A
cl
x
cl
+B
cl
+B
clr
r
e = C
cl
x
cl
+D
cl
+D
clr
r.
Now
e(s) =
_
C
cl
(sI A
cl
)
1
_
B
cl
]vdots B
clr
+
_
D
cl
.
.
. D
clr
__
_

r
_
and the numerator of the error transfer function has roots corresponding to
|sI A
cl
| which are all stable and at desirable locations. To ensure tracking
it is necessary that the numerators of the error transfer function cancel the
corresponding signal (disturbance or reference pole).
e
i
=
q
j=1
sI A
cl
(B
cl
)
j
(C
cl
)
i
(D
cl
)
ij
j
(s)
|sI A
cl
|
+
m
j=1
sI A
cl
(B
clr
)
j
(C
cl
)
i
(D
clr
)
ij
r
j
(s)
|sI A
cl
|
:=
q
i=1
N
ij
(s)
j
(s)
(s)
+
m
j=1
M
ij
(s)r
j
(s)
(s)
Sec. 5.2. MIMO Servomechanism 135
where (s) := |sI A
cl
|.
N
ij
(s) = det
_
_
sI ABS
sm
C
b
c1
(c
1
+d
1
D
sm
C
m
) sI A
c1
+b
c1
d
1
D
sc1
b
c
2
(c
2
+d
2
D
s
m
C
m
) 0
.
.
.
.
.
.
b
c
m
(c
m
+d
m
D
s
m
C
m
) 0
B
s
m
C
m
c
i
+d
i
D
sm
C
m
d
i
D
sc1
BD
sc
BC
s
E
j
b
c1
d
1
C
s
b
c1
f
1j
sI A
c
2
+b
c
2
d
2
D
s
c
2
0 b
c
2
d
2
C
s
b
c2
f
2j
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 sI A
c
m
+b
c
m
d
m
D
s
c
m
b
cm
d
m
C
s
b
cm
f
mj
B
s
c
sI A
s
0
d
i
D
sc2
d
i
D
scm
d
i
C
s
f
ij
_
_
By multiplying the last row by the vector b
ci
and adding to the corresponding
block of rows in the upper part of the matrix, we conclude that
N
ij
(s) = det [sI A
c
i
]
ij
(s).
M
ij
(s) = det
_
_
sI ABD
s
m
C
b
c
1
(c
1
+d
1
D
s
m
C
m
) sI A
c
1
+b
c
1
d
1
D
s
c
1
0
b
c
2
(c
2
+d
2
D
s
m
C
m
) 0 sI A
c
2
+b
c
2
d
2
D
s
c
2
.
.
.
.
.
.
b
cm
(c
m
+d
m
D
sm
C
m
) 0
B
sm
C
m
(c
i
+d
i
D
s
m
C
m
) d
i
D
s
c
1
d
i
D
s
c
2
BC
s
0
b
c1
d
1
C
s
0
b
c2
d
2
C
s
.
.
.
.
.
.
.
.
. b
j
sI A
c
m
+b
c
m
d
m
D
s
c
m
b
c
m
d
m
C
s
0
d
i
D
s
c
m
d
i
C
s
1
_
_
.
By multiplying the last row by b
c
i
and adding to the corresponding block of
rows in the upper part, we get a row to zero blocks with sI A
ci
. Therefore,
M
ij
(s) = |sI A
c
i
|
ij
(s).
Therefore each and every numerator in the transfer function of the error e
i
to
disturbance and error to reference has the factor |sI A
ci
|. Therefore, these
cancel the poles of
j
, j = 1, , q, r
j
, j = 1, , m, by the construction of A
c
i
(i.e., each A
c
i
contains all the poles of
j
as well as r
j
).
Now let us verify the controllability and observability of the regulator sub-
system
A =
_
A 0
B
c
C A
c
_

B =
_
B
B
c
D
_

C
m
=
_
C
m
0
0 I
m
_
.
Theorem 5.1 If (C
m
, A, B) is controllable and observable, and
Rank
_
I A B
C D
_
= n +m (A
c
),
then (

A,

B) is controllable and (

C
m
,

A) is observable.
Proof.
Rank
_
sI

A
C
m
_
= Rank
_
_
sI A 0
B
c
C sI A
c
C
m
0
0 I
_
_
= Rank
_
_
sI A 0
B
c
C 0
C
m
0
0 I
n
c
_
_
= n
c
+ Rank
_
_
sI A
C
m
B
c
C
_
_
= n
c
+n.
Thus, (

C
m
,

A) is observable.
Rank
_
sI

A

B
= Rank
_
sI A 0 B
B
c
C sI A
c
B
c
D
_
= Rank
_
I
n
0 0
0 B
c
sI A
c
_
_
_
sI A 0 B
C 0 D
0 I
n
c
0
_
_
= n +n
c
by Sylvesters rank inequality and the assumption. Thus, (

A,

B) is controllable.
5.3 Distrubance Rejection

Let
x = ax +bu +d
Sec. 5.3. Distrubance Rejection 137
y = cx
p = p +le
u = k
1
x +k
2
p +k
3
r.
Then,
_
x
p
_
=
_
a +bk
1
bk
2
lc
_ _
x
p
_
+
_
bk
3
l
_
r +
_
d
0
_
e =
_
c 0

_
x
p
_
+ 1r + 0
where r and are external signals. Then the transfer function between e and
(disturbance transfer function) is
e(s) =
_
c 0

_
s a bk
1
bk
2
lc s
_
1
_
d
0
_
(s)
=
cd(s )
(s)
(s)
where (s) is the characteristic polynomial of the closed loop system (i.e., sta-
ble). If (s ) is the factor of (s), the error will go to zero.
Let us consider the general description:
x = Ax +bu +g
y = cx +du +f
and
e = r y.
Suppose that and r satisfy the following dierential equations.
(D
p
+
p1
D
p1
+ +
1
D +
0
)r(t) = 0
(D
p
+
p1
D
p1
+ +
1
D +
0
)xi(t) = 0
where D denotes dierentiation. Note that this assumption is not a restriction
since if the disturbance (t) is sine wave and r(t) is ramp, we have
(D
2
+
2
0
)(t) = 0
D
2
r(t) = 0.
Then we can form a new dierential equation as follows:
(D
2
+
2
0
)D
2
r(t) = 0
D
2
(D
2
+
2
0
)(t) = 0
which are the same.
Let the proposed controller be
x
c
= A
c
x
c
+b
c
e
where
A
c
=
_
_
0 1 0
.
.
.
1
0

1

p1
_
_
b
c
=
_
_
0
.
.
.
0
1
_
_
.
Since
x
c
= A
c
x
c
+b
c
e
= A
c
x
c
+b
c
(r y)
= A
c
x
c
+b
c
(r cx du f)
= A
c
x
c
b
c
cx b
c
du b
c
f +b
c
r,
_
_
x

x
c
_
_
=
_
_
A
.
.
. 0

b
c
c
.
.
. A
c
_
_
. .
A
_
_
x

x
c
_
_
+
_
_
b

b
c
d
_
_
. .
b
u +
_
_
g

b
c
f
_
_
+
_
_
0

b
c
_
_
r
e =
_
c
.
.
. 0
_
_
x
x
c
_
+ (d)u + (f) + (1)r.
Note that the system

A is unstable because the eigenvalues of A
c
is a part of
the eigenvalues of

A. Thus, we use a state feedback to stabilize the system. Let
u = Kx +K
c
x
c
,

K := [K K
c
].
Then we have
_
_
x

x
c
_
_
=
_
_
A
.
.
. 0

b
c
c
.
.
. A
c
_
_
_
_
x

x
c
_
_
+
_
_
b

b
c
d
_
_
_
K K
c
_
_
x

x
c
_
_
+
_
_
g

b
c
f
_
_
+
_
_
0

b
c
_
_
r
_
_
x

x
c
_
_
=
_
_
A
b
K
.
.
. bK
c

b
c
c b
c
dK
.
.
. A
c
b
c
dK
c
_
_
. .
A
cl
_
_
x

x
c
_
_
+
_
_
g

b
c
f
_
_
+
_
_
0

b
c
_
_
r
e =
_
c dK
.
.
. dK
c
_
_
_
x

x
c
_
_
+ (f) + (1)r.
Now let us check this closed loop system. We need to verify the followings:
1) For disturbance rejection, the zeros of the transfer function between e(s)
and (s) should contain the roots of
(D) = D
p
+
p1
D
p1
+ +
1
D +
0
= 0.
2) For tracking, the zeros of the transfer function between e(s) and r(s) should
contain the roots of
(D) = D
p
+
p1
D
p1
+ +
1
D +
0
= 0.
3) The closed loop system (

A+
b

K) should be stable.
Let us rst consider the following.
Remark 5.4. If the system is given by
x = Ax +bu
y = cx +du,
then the transfer function may be written as:
G(s) = c(sI A)
1
b +d =
det
_
sI A b
c d
_
det[sI A]
. (5.12)
Let us rst verify 1). From the formula given in (5.12), we need to verify
det
_
_
sI AbK bK
c
.
.
. g
b
c
c +b
c
dK sI A
c
+b
c
dK
c
.
.
. b
c
f

c +dK dK
c
.
.
. f
_
_
s=
=
?
0
where is a root of (D) = 0. Since
det
_
_
sI AbK bK
c
g
b
c
c +b
c
dK sI A
c
+b
c
dK
c
b
c
f
c +dK dK
c
f
_
_
=
det
_
_
sI AbK bK
c
g
0 sI A
c
0
c +dK dK
c
f
_
_
= det[sI A
c
] [Something],
if the roots of (D) are equal to the poles of the controller, the condition is
satised. Since det[sI A
c
] = (s), 1) is shown.
2) can also be shown similarly by using
det
_
_
sI AbK bK
c
0
b
c
c +b
c
dK sI A
c
+b
c
dK
c
b
c
c +dK dK
c
1
_
_
= det
_
_
sI AbK bK
c
0
0 sI A
c
0
c +dK dK
c
1
_
_
= det[sI A
c
] [Something].
Now we need to show the closed loop stability. The main point here is
whether (

A,
b) stabilizable. If so, we are free to choose

K to stabilize the system.
To be controllable,
Rank
_
AI 0 b
b
c
c A
c
I b
c
d
_
= n +n
c
, I C.
To be stabilizable, the unstable modes must be controllable.
Example 5.2. Let the plant be
x = Ax +Bu +E
y = Cx +Du
where
A =
_
_
0 1 0
0 0 0
1 0 1
_
_
B =
_
_
1 0
0 1
0 1
_
_
E =
_
_
1
0
0
_
_
C =
_
1 0 0
0 0 1
_
D =
_
1 0
0 0
_
.
Note that (A, B, C) is controllable and observable. As seen the states x
1
and
x
3
are measurable, so we write
y
m
=
_
1 0 0
0 0 1
_
x.
Now suppose that
= sin
0
t.
Suppose that we want
1) the output y
1
to track step references
2) the output y
2
to track step and ramp references, and
3) both outputs reject a sinusoidal disturbance with
0
= 1.
Then,
disturbance:
= sin
0
t
=
0
cos
0
t
=
2
0
sin
0
t =
2
0
= D
2
+
2
0
= 0.
steps:
r
1
= u(t)
r
1
= 0 = D = 0.
ramps:
r
2
= at
r
2
= a
r
2
= 0 = D
2
= 0.
A common dierential equation which all references and disturbances satisfy
can be found:
D
2
+
2
0
= 0, D = 0, D
2
= 0
(D) = D
2
(D
2
+
2
0
) = D
4
+D
2
2
0
|
0=1
= D
4
+D
2
= 0.
Now write a servocompensator:
1) A
c
consists of 2 blocks because the tracking references are from 2 inputs,
2) A
ci
is of dimension 4 because (D) is of order 4.
From
(D) = D
4
+ 0D
3
+ 1D
2
+ 0D + 0 = 0
0
= 0,
1
= 0,
2
= 1,
3
= 0.
Thus,
x
c
=
_
_
0 1 0 0
0 0 1 0
0 0 0 1
0 0 1 0
0 1 0 0
0 0 1 0
0 0 0 1
0 0 1 0
_
_
. .
A
c
x
c
+
_
_
0
0
0
1
0
0
0
1
_
_
. .
B
c
e.
Therefore, the composite system is
_
x
x
c
_
=
_
A 0
B
c
C A
c
_
. .
A
_
x
x
c
_
+
_
B
B
c
D
_
. .
B
u +
_
E
B
c
F
_
+
_
0
B
c
_
r
e =
_
C 0

_
x
x
c
_
Du F +Ir.
To check the controllability of (

A,

B), we evaluate
Rank
_
I A B
C D
_
=?n +m, (A
c
).
So, we have
Rank
_
I A B
C D
_
= 5, for {0, 0, j}.
Thus, (

A,

B) is controllable and we have
y
m
=
_
y
m
x
c
_
=
_
C 0
0 I
8
_ _
x
x
c
_
+
_
D
0
_
u.
This means that we can measure x
1
, x
3
and all x
c
states. Then here, we set
externals to zero and design a stabilizing compensator.
x = A
c
x
c
+B
c
e
Servocompensator
Stabilizing Compensator
Plant
-
r
+
-
-
x
c
-
-
-
-
y
y
m
6
y
Figure 5.6. A servomechanism
PART IV
ROBUST AND OPTIMAL
CONTROL
Chapter 6
PARAMETRIC ROBUST
CONTROL
6.1 INTRODUCTION
6.2 BOUNDARY CROSSING THEOREM
6.3 PARAMETRIC STABILITY MARGINS
6.4 STABILITY OF LINE SEGMENTS
6.5 INTERVAL POLYNOMIALS: KHARITONOV THEOREM
6.6 EDGE THEOREM
6.7 GENERALIZED KHARITONOV THEOREM
6.8 ROBUST FREQUENCY DOMAIN DESIGN
6.9 SUMMARY
6.10 EXERCISES
145
146 PARAMETRIC ROBUST CONTROL Ch. 6
Chapter 7
ROBUSTNESS UNDER
NORM BOUNDED
PERTURBATIONS
7.1 INTRODUCTION
7.2 SMALL GAIN THEOREM
7.3 H
OPTIMAL CONTROL
7.4 ABSOLUTE STABILITY PROBLEM
7.5 ROBUSTNESS UNDER MIXED PERTURBATIONS
7.6 SUMMARY
7.7 EXERCISES
147
PART V
THREE TERM
CONTROLLER DESIGN
Chapter 8
PID CONTROLLER DESIGN
8.1 INTRODUCTION
8.2 SUMMARY
8.3 EXERCISES
149
150 PID CONTROLLER DESIGN Ch. 8
Chapter 9
FIRST ORDER
CONTROLLER DESIGN
9.1 INTRODUCTION
9.2 SUMMARY
9.3 EXERCISES
151
152 FIRST ORDER CONTROLLER DESIGN Ch. 9

LINEAR CONTROL THEORY: Structures, Robustness, and Design

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

LINEAR CONTROL THEORY: Structures, Robustness, and Design

Uploaded by

Copyright:

Available Formats

LINEAR CONTROL

). An example of a memoryless system is

) of a system with memory depends on its present

), as well as its past and/or future inputs. If the output y (t

is the amount of information at t

Figure 1.3. A cart with an inverted pendulum (Example 1.3)

= 0 (no external force)

and the output [y, ]

diagonal terms. (1.28)

1.5.3 Jordan Form of A

This is due to the Cayley Hamilton Theorem.

be the initial condition and the desired nal state of x(t),

= x(T) in T seconds. Recall the solution of the state

)| d = 0 from the assumption

Figure 1.10. Feedback connection (state space models)

k [1, 2, , m]; i [1, 2, , r].

is an n n invertible matrix. Then we have the following.

Now suppose that {v

is an invertible n n matrix. Then we have the following:

and we need to show that i) A

Therefore if a realization {A, B, C, D} with rank[O] = n

= det [U(s)] det

(see eq. (2.13))

2.7.5 Generalized Bezout Identity

Y (s)) such that

Example 2.6. (Converse is not true.) Let

f are the roots of the polynomial

2) take the last row of L

f and A+bf have the same eigenvalues.

be the last row of L

The transfer function of the observer as controller is derived as follows:

for some matrix L. L is known as a square root of Q.

Qx is then a measure of control accuracy, the term u

(T)Mx(T) a measure of terminal control accuracy.

P = 0 (Algebraic Riccati Equation(ARE)

D, where the pair (A, D) is

be the smallest integer j such that

exists and is unique if (c, A) is observable, which we assume. Add p

) integrators to the system as follows:

b) is controllable when (A, b) is as we assume.

f has a prescribed set of n + p eigenvalues in the LHP.

f. Let us write out

4.1.2 Single Output System

be the least integer j such that

exists and is unique when (A, B) is controllable. The compensator

), proper (bounded at innity, i.e., no poles at innity) rational function. A

if and only if p(s) has no poles in I C

, let the degree of f(s) be dened by

is a Euclidean domain, we need to show that we can carry

, just as we do Euclidean division in IR[s] with polynomials.

(g(s) = 0), there exists

the zeros ,of

, f divides g if and only if every I C

are coprime if and only

The main result follows from this.

Example 4.3. Consider

and from the right by

where R M(S) is arbitrary. Thus follows from

e(1 +CP) = r KPr.

5.3 Distrubance Rejection

b) stabilizable. If so, we are free to choose

You might also like