EL 625 Lecture 10 1

EL 625 Lecture 10

Pole Placement and Observer Design

Pole Placement
Consider the system

x = Ax (1)

The solution to this system is

x(t) = eAtx(0) (2)

If the eigenvalues of A all lie in the open left half plane, x(t) asymp-
totically approaches the origin as time goes to infinity (the system
is asymptotically stable). If the eigenvalues of A lie in the closed
left half plane (with some eigenvalues possibly on the imaginary ax-
is), the state x(t) remains bounded (the system is stable). If some
eigenvalues of A lie in the right half plane, the system is unstable.
Consider the system

x = Ax + Bu (3)

The state x is measured and can be used to calculate a suitable value

of u to make the system stable. The simplest feedback is a static
EL 625 Lecture 10 2

linear state feedback, u = Kx (where K is a constant matrix). With

this feedback,

x = Ax + Bu

= Ax + BKx

= (A + BK)x (4)

= A0 x (5)

A0 = A + BK (6)

If it is possible to make the eigenvalues of A0 lie in the open left

half plane by using the gains K, then, choosing u = Kx would
make the resulting closed loop system stable. It can be proved that
the eigenvalues of (A + BK) can be arbitrarily assigned through a
suitable choice of K if the pair (A, B) is controllable.
If(A,B) is a controllable pair, a similarity transformation, T can
be found such that the transformed matrices, A = T 1AT and B =
EL 625 Lecture 10 3

T 1B are in the controller canonical f orm, shown below.

0 1 0 ... ... 0






... ... ... ... ...






... ... ... ... ...



A =


(7)


... ... ... 0





... 0 1







a0 a1 . . . . . . . . . an1




0




...



B =
...
(8)







0




1

an1, an2, . . . , a1, a0 are the coefficients of the characteristic poly-

nomial (the characteristic polynomial is n + an1n1 + . . . + a1 +
a0). The similarity transformation to achieve this controller canon-
ical structure for the single-input case is described below.

Characteristic polynomial, p() = n + an1n1 + . . . + a1

+a0 (9)
EL 625 Lecture 10 4

n1 n2
= (
|
+ an1 {z
+ . . . + a1)}
p1 ()
+a0 (10)

p1() = n1 + an1n2 + . . . + a1 (11)

= |(n2 + an1{zn3 + . . . + a2)}

p2 ()
+a1 (12)

...

pn1() = + an1 (13)

pn() = 1 (14)

Define the set of vectors (v1, . . . , vn) as

v1 = p1(A)B

v2 = p2(A)B
...

vn = pn(A)B (15)

pi(A) is the polynomial pi() evaluated at = A.

To prove that this coordinate transformation achieves the desired
transformed form, we have to compute the effect of A on the coordi-
nate vectors, v1, v2, . . . , vn.

Av1 = A(An1 + an1An2 + . . . + a2A + a1I)B

EL 625 Lecture 10 5

= (An + an1An1 + . . . + a2A2 + a1A)

= (An + an1An1 + . . . + a2A2 + a1A + a0I a0I)B

= a0IB

= a0B

= a0vn (16)

(By Cayley Hamilton theorem, p(A) = An +an1An1 +. . .+a2A2 +

a1A + a0I = 0).

Av2 = A(An2 + an1An3 + . . . + a2I)B

= (An1 + an1An2 + . . . + a2A)B

= (An1 + an1An2 + . . . + a2A + a1I a1I)B

= v1 a1vn (17)
...

Avn = AB

= (A + an1I an1I)B

= (A + an1I)B an1B

= vn1 an1B

= vn1 an1vn (18)

Also, B = vn. Hence, in the new coordinates, z = T 1x where

EL 625 Lecture 10 6

T = [v1, v2, . . . , vn], the transformed matrices would have the form
shown in (7) and (8).
Once the matrices, A and B have been transformed to the con-
troller canonical form, designing the state-feedback controller is easy.
Consider




0 1 0 ... ... 0





...



0

... ... ... ...












...


... ... ... ... ...






"

#
A + B K =


+
...
k0 k1 . . . . . . kn1



... ... ... 0







0


... 0 1













1

a0 a1 . . . . . . . . . an1




0 1 0 ... ... 0






... ... ... ... ...





... ... ... ... ...



=


(19)


... ... ... 0





... 0 1







a0 + k0 a1 + k1 . . . . . . . . . an1 + kn1

The characteristic polynomial of (A+B K) is n+(an1kn1)n1+

. . . + (a0 k0). Thus, by choosing the gains (k0, k1, . . . , kn1) ap-
propriately, the coefficients of the characteristic polynomial can be
EL 625 Lecture 10 7

changed arbitrarily which implies that the eigenvalues can be as-

signed arbitrarily. Thus, the closed loop poles of the system can be
arbitrarily placed using a linear static state feedback, u = Kz =
KT 1x = Kx.
Example:

x = Ax + Bu (20)

where




1 3 2


A =


0 1 0


(21)



1 1 1




0


B =


1


(22)



1
The characteristic polynomial of A is

p() = 3 2 + 1 (23)

The polynomials p1(), p2() and p3() are

p1() = 2 + 1 (24)

p2() = 1 (25)

p3() = 1 (26)
EL 625 Lecture 10 8

Hence, the vectors v1, v2 and v3 are

v1 = p1(A)B

= (A2 A + I)B




1


=


1


(27)



1
v2 = p2(A)B

= (A I)B




1




=


0


(28)



1
v3 = p3(A)B

= IB




0


=


1


(29)



1

The transformation matrix, T is

" #
T = = v1 v2 v3
EL 625 Lecture 10 9

1 1 0


=


1 0 1


(30)



1 1 1

In the transformed coordinates, A and B are

1
A = T

AT



0 1 0


=


0 0 1


(31)



1 1 1
1
B = T

B



0


=


0


(32)



1

A and B are now in the controller canonical form.

The characteristic polynomial of (A + B K) is 3 + (1 k2)2 +
(1 k1) + (1 k0). If we want the closed loop poles to be
5, 1 + j and 1 j, the closed loop characteristic polynomial
must be 3 + 72 + 12 + 10. Hence, choose the gains k0, k1 and k2
as

k0 = 11 (33)
EL 625 Lecture 10 10

k1 = 11 (34)

k2 = 8 (35)

Hence, the state feedback, u = Kz achieves the required pole

placement.

u = Kz (36)

= KT 1x = Kx

= K = KT 1 (37)


1 1 1

" #




= 11 11 8


0 1 1





1 2 1
" #
= 3 16 8 (38)

To verify, the eigenvalues of (A + BK) can be computed.

1 3 2


0
" #

(A + BK) =


0 1 0


+ 1





3 16 8



1 1 1 1




1 3 2


=


3 15 8


(39)



2 17 9
EL 625 Lecture 10 11

The characteristic polynomial of (A + BK) is 3 + 72 + 12 + 10,

as desired.
State Reconstruction: Asymptotic Observers
In the previous lecture, static full state feedback was employed
to modify the eigenvalues and hence the dynamic response of a lin-
ear time-invariant system. This was possible when the realization
{A,B,C} was controllable and all state variables were available for
feedback purposes. However, measurements of all the states may not
be feasible or impossible. In some cases, the estimate of the state
variables may be desirable rather than their direct measurement s-
ince the sensor/transducer required to measure that state may be
susceptible to excessive noise or it is hard if not impossible to phys-
ically measure some particular state variables. In these cases, we
are interested in reconstructing the state variables. It will be shown
that the states may be obtained from the knowledge of the system
input and the system output. Furthermore, in a later section it will
be shown that the problem of state estimation is dual to the state
feedback problem and states can be attained if the system { A,B,C}
is observable.
The states may also be recovered through differentiation of the
EL 625 Lecture 10 12

output; however, this is undesirable. We will construct a dynami-

cal system called an asymptotic observer to estimate the states of a
system. The observer is named asymptotic since the estimates con-
verge to the true states after a transient period. Having attained an
asymptotic observer, it will be utilized to implement state feedback.
It will be shown that the overall closed-loop system will be internally
stable when an observer is utilized in the feedback loop to reconstruct
the states for feedback purposes. However, it should be noted that
utilization of an observer will deteriorate the transient response of a
system. This should be expected since the observer will provide a
good estimate of the states after a transient period.

u y
- System -

- Observer 

x
?

Figure 1: An observer.
EL 625 Lecture 10 13

1 Full Order Observers

Consider a linear time-invariant system of the form

x = Ax + Bu (40)

y = Cx . (41)

Differentiating the output map, we obtain

y = C x = CAx + CBu (42)

y = CAx + CB u = CA2x + CABu + CB u . (43)

Therefore, repeated differentiation of the output will yield

Y = n x + T U . (44)

where n is the observability matrix and

y





u






y


u

Y =


, U=


, (45)

...


...






y (n1)
u(n1)
EL 625 Lecture 10 14

0 0 0 0





CB 0 0 0







CAB CB 0 0




T =


CA2B CAB CB 0


. (46)





... ... ... ... ...







... ... ... ... ...






CAn2B CAn3B CB 0
The matrix T is called a lower triangular Toeplitz matrix.
Given the realization of a system, i.e., {A,B,C}, and its input and
output signals, the states may be recovered through the differenti-
ation of the input and output signals. Of course, this is possible if
the system is observable. This can be easily verified from (44). The
set of equations derived in (44) is linear and has a solution if the
vector Y T U is contained in the range space of the linear trans-
formation n. Therefore, if n has full rank, then this system of
equations may be solved for the states. As shown before, n hav-
ing full rank corresponds to the observability of the system. This
process for determining states is undesirable since derivative of input
and output signals are necessary. In the next sections, dynamical
systems, namely observers, are introduced for state estimation.
EL 625 Lecture 10 15

1.1 Open-Loop Observers

Consider the following dynamical system

x = Ax + Bu ; x(0) = x0 (47)

y = C x. (48)

The above system is an observer for the system given by (40)-(41).

This observer is a replica of the original system (Figure (2)) and uti-
lizes the knowledge of the system dynamics (i.e., the triple {A,B,C}),
and the input to the system. However, the initial conditions for the
observer are the crucial information needed. If the initial conditions
of the original system, namely x(0) = x0, are exactly known, then
the observer will produce the state variables exactly. However, due
to disturbances and in many cases only a partial knowledge of the
initial conditions and dynamics of the system at hand, there exists
a discrepancy between the initial conditions of the original system
and the observer. Therefore, it is important to study the effect of an
error in the knowledge of the initial conditions.
Define the error between the exact value of the state and its esti-
mate at time t as e(t), i.e.,

e(t) = x(t) x(t). (49)

EL 625 Lecture 10 16

Plant

H  x H
u - BHH
+ P
- -
R
- CHH -y
 
  
+ 6

A
HH
H

H  x H
- BHH
+ P
- -
R
- CH
H - y

   
+ 6

HHA
H

Observer ?
x

Figure 2: Open-loop Observer.

The dynamics of the error are

e = x x = A(x x) = Ae ; e(0) = x0 x0 . (50)

From (50), it is easy to see that if the initial conditions are exactly
known, then the error is identically zero. However, if the initial
condition on the error is nonzero, then the behavior of the error is
governed by (50). Therefore, if the open-loop system is unstable, the
error will grow unbounded and the estimate of the states will not
converge to the true state variables. Even if the open-loop system
is asymptotically stable, the behavior of the error dynamics will be
EL 625 Lecture 10 17

sluggish if the open-loop system has eigenvalues close to the j-

axis. Since we are interested in using the estimates of the states
in a full state feedback design, if the error dynamics are sluggish
and the error persists for a long time, then the overall controlled
system performance will be unsatisfactory. In the next section, we
will introduce a scheme to alleviate this problem.

1.2 Closed-Loop Observers

Recall that the original problem was to estimate the state variables
of a linear time-invariant system knowing its dynamics, and its input
and output signals. In the proposed observer scheme in the previous
section, we utilized all the given information excluding the knowl-
edge of the output measurements. This is the reason why the above
scheme is called an open-loop observer. Therefore, we propose to use
a modified version of the open-loop observer that makes use of the
output measurements.
Consider the following dynamical system

x = F x + G1u + G2y . (51)

Define the error as before (i.e., given as in (49)). Then, the error
EL 625 Lecture 10 18

dynamics in this case are

e = x x = F x + G1u + G2y Ax Bu

= F (x + e) + (G1 B)u + G2Cx Ax

= F e + (F + G2C A)x + (G1 B)u . (52)

Now, the matrices F and G1 may be chosen as

F = A G2C (53)

G1 = B . (54)

Substituting (53) and (54) in (51) and (52), we obtain

x = (A G2C)x + Bu + G2y (55)

and
e = F e = (A G2C)e . (56)

Equations (55) and (56) describe the dynamics of the observer

and the error respectively. First we will examine the error dynamics
and then we will discuss the observer topology. As derived, the error
dynamics depend on the system dynamics through the pair {A,C}
and the unknown gain G2. Since the unknown gain G2 is chosen by
the designer, there is a chance to modify the dynamics of the error.
Therefore, the question is: given a pair {A,C}, is it possible to choose
 EL 625 Lecture 10 19

a matrix G2 such that the eigenvalues of the matrix A G2C lies in

the open left-half plane of the complex plane?
The answer to the above question is yes if the pair {A,C} is an
observable pair (or the original system is observable). The above
question is similar to the pole placement by full state feedback en-
countered in the previous chapter. There we considered placing the
eigenvalues of A + BK when the pair {A,B} were given and the ma-
trix K needed to be chosen. The pole placement was indeed solvable
if and only if the system was controllable. The current problem turns
out to be a dual problem of the pole placement, and as we should ex-
pect the controllability condition in the case of pole placement should
be changed to the observability in the observer problem.
Knowing that the eigenvalues of a matrix and its transpose are the
same, we may consider pole placement for the transpose of the error
dynamics matrix (i.e., F = A G2C) instead of pole placement for
A G2C, i.e.,

det(IF ) = det[I(AG2C)] = det[I(AT C T G2T )] = det(IF T ) .

From the above observation, ensuring that the eigenvalues of A

G2C are in the open left-half plane is equivalent to pole placement
for the pair {AT , C T }, i.e., find a gain K such that the eigenvalues
EL 625 Lecture 10 20

of AT + C T K are arbitrarily placed in the open-left half plane. After

finding this gain through the techniques described in the previous
chapter for pole placement through state feedback, the observer gain
is given by
G2 = K T . (57)

To further study the structure of the observer, rewrite the observer

dynamics given by (55) as

x = Ax + Bu + G2y G2C x . (58)

Now define an output for the observer of the form y = C x. There-

fore, (58) may be rewritten as

x = Ax + Bu + G2(y y) (59)

y = C x . (60)

It is interesting to note that the above observer is similar to the

open-loop observer (or dynamics of the original plant) with the d-
ifference that the observer dynamics contain a term corresponding
to the output error. For this reason the observer has been named
a closed-loop observer since the dynamics of the observer is being
driven by the error between the actual output of the system and the
estimated output by the observer. This term corresponding to the
EL 625 Lecture 10 21

output error is also known as the innovation process (especially in the

Kalman Filtering problem). Henceforth, through the proper choice
of the error gain, the error dynamics may be stabilized.
The choice of the observer poles is an important one since the
observer should generate the estimates of the state variables in real-
time for control purposes. Therefore, the observer poles should be
selected to be faster than the closed-loop poles of the overall system.
A figure of merit is that the observer time constants should be three
to five times faster than the closed-loop system time constants. The
reasons for not choosing the observer poles far in the left-half complex
plane are: 1) larger gains in the observer and hence saturation, and 2)
increased bandwidth which makes the observer more susceptible to
sensor noise occurring at high frequencies. Another important issue
is the robustness of the observer and hence the closed-loop system to
parameter variations in the plant dynamics.

Example .1 Consider a voltage driven DC motor where its shaft

position is being measured and we are interested in designing an
observer to estimate the shaft velocity of the motor. Ignoring the
electrical time-constant of the motor, the dynamics are:
EL 625 Lecture 10 22

KmKb Km
Jm + (Bm + ) = V. (61)
R R
Choosing the shaft position and velocity as the state variables, i.e.,
x1 = and x2 = , the state variable representation of the dynamics
is

x1








0 1


x1





0



=




+


u (62)

x2 0 x2 1
where = (Bm + KmRKb )/Jm and u = Km
RJm V .
In this case, the output is the shaft position, i.e.,

" #


x1


y= 1 0


.

x2

Therefore, the full-order observer is

x1








0 1


x1





0


g1





=




+


u+


(y y) (63)

x2 0 x2 1 g2

" #


x1


y = 1 0


.

x2

The gain vector G = [g1 g2]T needs to be chosen such that the
error dynamics are asymptotically stable. Lets choose the eigenval-
ues of the error dynamics at 5 and 6 . Then, the gain matrix
EL 625 Lecture 10 23

may be found. Consider

T T




0 0





1



{A , C } =


,



.



1 0

Then, pole placement techniques may be utilized to place the poles

at 5 and 6 and obtain the gain as G = K T . Since the system
has only two states, it is rather easy to proceed directly to place the
poles of A GC, i.e.,




0 0





g1 "


#


g1 1


A GC =





1 0 =




1 g2 g2

(AGC) = {|2 +(g1 + )+g1 +g2 = 0} = {5, 6 }.

Therefore,






g1 + = 11

g1 = 10




2


g1 + g2 = 30

g2 = 20 2
The observer is given as follows:


x 0 1 x1 0 10


1



=




+


u+


(y y) (64)
2
x2 0 x2 1 20

" #


x1


y = 1 0


.

x2
EL 625 Lecture 10 24

2 Reduced Order Observers

The state space dimension of the observers mentioned in the previ-

ous section is the same as the original plant. For this reason, these
observers are named full order observers. However, the output mea-
surements may contain some of the state variables directly, if not
linear combinations of the states. Therefore, it is natural to wonder
if the dimension of the observer may be reduced by an intelligent
use of the output vector. The answer to this question is yes! To
show this, we first assume that through an appropriate change of
coordinates the output matrix is rewritten in the following form:

C = [In1 | 0] (65)

where In1 represents an n1 by n1 identity matrix and n1 is the row

rank of the original output matrix.
Partition the original plant dynamics as follows:

x1 A11 A12 x1 B1


= + u . (66)


x2 A21 A22 x2
B2
where x1 Rn1n1 .
The choice of the output equation guarantees that the first n1
states are measured directly. Therefore, the states x2 need to be
EL 625 Lecture 10 25

estimated. Hence, a reduced order observer of dimension n n1

should be sufficient where n is the dimension of the state space of
the plant. From the previous analysis on the full order observers,
the following choice for the dynamics of the reduced order observer
is taken
x2 = F x2 + G1u + G2y + G3y . (67)

From the above dynamics, it seems that the derivative of the out-
put is also an input to the reduced order observer and as discussed
before this is not desirable. However, the term containing the deriva-
tive of the output may be eliminated by an appropriate change of
coordinates. These coordinates are

z = x2 G3y . (68)

Utilizing this new choice of state variables, the observer dynamics

may be rewritten as

z = F z + G1u + (G2 + F G3)y . (69)

At this point, the error dynamics need to be examined. Define the

error as before, i.e.,
e = x2 x2 .
EL 625 Lecture 10 26

Then

e = F x2 + G1u + G2y + G3y (A21x1 + A22x2 + B2u)

= F e + (G1 B2 + G3B1)u + (G2 + G3A11 A21)x1

+(F + G3A12 A22)x2. (70)

Therefore, the following choice of the gain matrices will make the
error dynamics independent of the other signals in the system (i.e.,
e = F e)

G1 = B2 G3B1 (71)

G2 = A21 G3A11 (72)

F = A22 G3A12 . (73)

The observer design is reduced to finding a gain, namely G3, such

that the error dynamics are asymptotically stable, i.e., (F ) C .
It is worthwhile to note that this problem is very similar to the find-
ing the gains for the full order observer which in turn is dual to the
pole placement problem. In the case of full order observer design, it
was proved that the observability of the original system guaranteed
that the eigenvalues of the error dynamics may be placed arbitrarily.
For the problem at hand, it seems that the pair {A22, A12} need to be
an observable pair. However, it can be proved that the original plant
EL 625 Lecture 10 27

dynamics are observable if and only if the pair {A22, A12} is observ-
able. Therefore, the observability of the original system is necessary
and sufficient to attain an asymptotic reduced order observer.

3 Observer-Based State Feedback: Separation Princi-

ple

In the previous sections, observers were studied to reconstruct the

states of a dynamical system. Having attained the full states, state
feedback may be utilized to shape the dynamic response of the system
(Figure (3)). However, the overall stability of the closed-loop system
need to be examined when an observer (full-order or reduced-order)
is injected in the feedback loop. In this section, we will show that the
observer dynamics may be realized independent of the gains for the
state feedback. Therefore, the state feedback gains may be calculated
as if the full states were accessible. Furthermore, the observer gain
may be chosen independent of the selection of controller gains. This
result is known as the separation theorem since the gains for the
observer and the controller may be calculated independently. We
will prove this result in the analysis to follow.
Consider the state feedback resulting from the use of the estimates
EL 625 Lecture 10 28

u y
- System -

+
6

- Observer 

x
 ?
HHK
H
Figure 3: An observer-based state feedback.

of the states generated by the observer, i.e.,

u = K x (74)

where x is the states of the observer.

The closed-loop dynamics are as follows:

x = Ax + BK x

x = Ax + Bu + G(y y) = (A + BK GC)x + GCx

or

x A BK x






=






(75)

x A + BK GC GC x
The above system can be rewritten in a triangular structure by
EL 625 Lecture 10 29

utilizing the following similarity transformation:

I 0


T =


. (76)

I I
Transform the system given by (75) into its new coordinates, name-
ly Z = T X where X = [xT xT ]T , by the the transformation given
in (76). Then,

Z = AZ (77)

where

1



I 0


A BK


I 0


A = T AT =








I I GC A + BK GC I I




A + BK BK


=


.

0 A GC

Since A has a triangular structure, its eigenvalues are

[
(A) = (A + BK) (A GC). (78)

From (78), we observe that the eigenvalues of the overall closed-

loop system is composed of the eigenvalues of the open-loop system
with a state feedback (as if the true states are available) and the
eigenvalues of the error dynamics for the observer. This is a very
EL 625 Lecture 10 30

nice result since the implication is that the pole-placement and the
observer design can be performed independent of each other. This
result is known as the separation principle or separation property.