From Classical to State-Feedback-Based Controllers

By Juan I. Yuz and Mario E. Salgado
A
utomatic control teaching can be fragmented, com-
prising different streams that seemtobe disconnec t
ed or even competing. The appeal of new theories
and concepts may encourage students to forget that the key
issue in the discipline is the control design problem. If two
or more different approaches provide a good solution to
that problem, then a strong connection should exist be-
tween them. If such connections can be established, this
should help students better understand the underlying con-
cepts in the control design problem.
In this article, we consider continuous-time systems and
explore the connections betweensingle-input, single-output
(SISO) linear classical controllers (i.e., those defined by
transfer functions) and the control design approach based
on feedback of an estimatedsystemstate. The aimof this ar-
ticle is to provide insight for automatic control teaching and
control system design.
The transition fromestimated state feedback to classical
control is well known; however, to the authors’ knowledge,
the reverse transition has not been previously articulated
for the general case.
As is well known, the combination of an observer to-
gether with feedback of the observed state can be reduced
to an equivalent classical control loop in which the control-
ler is expressed in transfer function form [1]. For instance,
assume that the state-space plant model is given by the
four-tuple (A
o
, B
o
, C
o
, 0), a full-order state observer is built
with gain J, and the observed state is fed back with gain K.
Hence, the equivalent classical controller has a transfer
functionC s ( ) and an associated state-space model given by
the four-tuple (A
c
, B
c
, C
c
, D
c
), where
C s s ( ) ( ) = − + +

K I A B K JC J
o o o
1
, (1)
A A B K JC B J C K
c o o o c c
= − − = = = , , , D
c
0. (2)
This equivalent classical controller in transfer function
form has three distinctive features:
1) The controller is strictly proper since a full-order ob-
server is strictly proper (leading to D
c
= 0).
2) The complexity of the equivalent controller is tied to
the complexity of the plant since the order of the ob-
server is equal to the order of the plant model.
3) There is no guarantee that the equivalent controller
has the necessary features to yield zero steady-state
error in the presence of certain types of disturbances
(e.g., constant, sinusoid).
The last problem is the simplest to address; we will sum-
marize the main results of how that problem is solved. This
step will be instrumental in tackling the first two issues,
which pose more serious obstacles to understanding the
connection to linear classical controllers. Those difficulties
are highlighted by the following observations:
• The most common controllers, such as those belong-
ing to the proportional-integral-derivative (PID) fam-
ily, are biproper.
• Some controllers have a much simpler structure than
the model of the plant under control.
• Some controllers are more complex than the model of
the plant to be controlled.
We show how these issues can be resolved so that every
linear classical controller can be associated with a control-
ler basedonfeedback of anobservedstate. Inthis endeavor,
the key tool will be the use of reduced-order observers [2].
In addition, the technical link between both classes of con-
trollers will be established through the closed-loop charac-
teristic polynomial.
Although we will only refer to linear, time-invariant, contin-
uous-time, SISO systems, the reader will be able to appreciate
that the extensiontothe discrete-time case is straightforward.
We illustrate this latter point in the examples.
In summary, the main question addressed in this article
is: Given a biproper or strictly proper controller, what is an
equivalent observer-state feedback scheme?
Classical Control Loop
The basic classical control loop of interest here is shown in
Figure 1. The plant is described by its nominal strictly
proper transfer function, G s
o
( ). Further, we assume that the
system has no time delay. This is done to keep the model of
finite degree. Of course, plants with time delays can be eas-
ily dealt with in the discrete-time domain. The transfer func-
tion can then be written as
58 IEEE Control Systems Magazine August 2003
LECTURENOTES
0272-1708/03/$17.00©2003IEEE
Figure 1. Classical control loop.
Authorized licensed use limited to: University of Michigan Library. Downloaded on April 09,2010 at 18:24:32 UTC from IEEE Xplore. Restrictions apply.
G s
B s
A s
o
o
o
( )
( )
( )
= ,
(3)
where B s
o
( ) and A s
o
( ) are polynomials in s, with A s
o
( ) being
a monic polynomial of finite order n.
Similarly, the controller is representedby the proper (not
necessarily strictly proper) transfer function
C s
P s
L s
( )
( )
( )
= ,
(4)
where P s ( ) and L s ( ) are polynomials in s, with L s ( ) being a
monic polynomial of order n
c
.
Under these conditions, and regardless of the control de-
sign method being used, the closed-loop characteristic
polynomial A s
cl
( ) is given by
A s A s L s B s P s
cl o o
( ) ( ) ( ) ( ) ( ) = + . (5)
If pole assignment is used to synthesize the controller,
the first step is to specify A s
cl
( ) for G s
o
( ), followed by (5)
solved for P s ( ) and L s ( ). This synthesis procedure has the
following relevant features [1]:
1) To arbitrarily specify A s
cl
( ), it must be of degree
greater thanor at least equal to2 1 n − . Whenthe degree
takes the minimum value, a biproper controller of or-
der n −1is obtained.
2) If A s
cl
( ) is arbitrarily chosen with degree equal to
2 1 n n
d
− + (n
d

+
Z ), then the controller is strictly
proper and has degree n n
d
− + 1 .
3) If the solution L s ( ) must have a prespecified factor of
degree n
x
, and A s
cl
( ) is arbitrarily chosen with degree
equal to2 1 n n
x
− + , then the controller is biproper and
of degree n n
x
− + 1 .
4) It might happenthat for a specific choice of A s
cl
( ), with
degree n
cl
, the polynomials P s ( ) and L s ( ) have a com-
mon factor W s ( ) (stable) of degree n
w
, which is also a
factor of the chosen A s
cl
( ). In this case, the controller
has degree n n n
cl w
− − .
If we use any other design method, such as root locus,
loop shaping, or lead-lag control, we always obtain a con-
troller transfer function with structure resulting from one
(or a combination) of the four cases described above.
Consider, for instance, a plant with transfer functionG
o s ( )
having three poles (i.e., n = 3). Further assume that we want
todesigna PI controller, that is, a controller where L s ( )has a
prespecified first-order factor [i.e., n
x
=1 (case 3)]. The
reader can then verify that this corresponds to a pole place-
ment synthesis where A s
cl
( ) is a polynomial of degree
2 1 6 n n
x
− + = , leading to a biproper controller with L s ( ) of
degree 3, and where L s ( ) and P s ( ) (and A s
cl
( )) share a com-
mon stable factor W s ( ) of degree 2 [i.e., n
w
= 2 (case 4)].
Thus, any classical controller with any degree, either
biproper or strictly proper, can be thought of as the out-
come of a synthesis procedure based on the specification of
closed-loop poles.
This result will be used to forma link with controller syn-
thesis basedonfeedback of the estimatedstate. The key link
will be via the closed-loop characteristic polynomial.
Observers
We will next review full and reduced-order observers. We
consider a completely observable system having
state-space model
&
( ) ( ) ( ) x A t t u t = + x B , (6)
y t t ( ) ( ) = Cx , (7)
where A ∈
×
R
m m
, B ∈R
m
, and C
T m
∈R . Figure 2 shows a
block diagram representation of this state-space model of
the plant.
We intend to use the information provided by the system
state x( ) t to control the output y t ( ). However, we know the
assumption that all state signals are available is often unre-
alistic due to physical or economical limitations [1], [3].
August 2003 IEEE Control Systems Magazine 59
Figure 2. Block diagram of state-space model, plant G s
0
( ).
Figure 3. Block diagram for a full-order observer.
Authorized licensed use limited to: University of Michigan Library. Downloaded on April 09,2010 at 18:24:32 UTC from IEEE Xplore. Restrictions apply.
This problemis typically addressedby estimating x( ) t using
an observer, which generates the state estimate
$
( ) x t [2].
Full-Order Observers
A full-order observer has the general form
$
&
( )
$
( ) ( ) ( ( )
$
( ))
$
( )
x Ax B J Cx t t u t y t t
y t
= + + −
123
correction term
6 7 44 8 44
,
(8)
where J ∈R
m
is known as the observer gain and
$
( ) x t
m
∈R is
the estimated or observed state.
We note that the systemdefinedin (8) is a combination of
the systemmodel (6) and a correction term, which is neces-
sary due to the mismatch in the initial state, leading to an
output mismatch, since the system initial state is unknown.
This system is known as a full-order observer because it esti-
mates the mcomponents of the state x( ) t . Figure 3 shows the
full-order observer in (8), with inputs u t ( ) and y t ( ) and out-
put
$
( ) x t .
The observer gainJ plays a fundamental role indetermin-
ing the speed at which the estimated state converges to the
true system state. Given the state-space model defined in
(6)-(7) and the full-order observer (8), the dynamics of the
estimation error
~
( ) ( )
$
( ) x x x t t t = − are described by
~
&
( ) ( )
~
( ) x A JC x t t = − . (9)
Since the system is assumed to be completely observ-
able, the eigenvalues of A JC − can be placed anywhere in
the complex plane C by choosing a suitable J.
Equation (8) can also be expressed as
$
&
( )
$
( ) ( ) ( ) x(t) t t t = − + + A JC x Bu Jy ,
(10)
and the observer eigenvalues are the roots of the polyno-
mial
E s s ( ) ( ) = − + det I A JC
m
. (11)
Reduced-Order Observers
If we examine (7), we note that there is information regard-
ing the state x( ) t contained in the (measured) output, y t ( ).
Furthermore, since the state-space model of a system is not
unique and, assuming complete observability, there is al-
ways a state-space description such that (6)-(7) can be writ-
ten as [4]
&
( ) ( ) ( ) x
A
A A
x
B
12
21 22 2
t
a
t
b
u t =






+






11 1
,
(12)
y t t ( ) [ ] ( ) = 1 0 0 K x , (13)
where a b
T m
11 1
1
, ; , , ∈ ∈

R R A A B
12 21 2
, and A
22
1 1
R
( ) ( ) m m − × −
.
This description can be obtained using, for example, either
the observer or the observability canonical forms [5].
The reader can now appreciate that x t
1
( ), the first com-
ponent of the state x( ) t , is y t ( ), the system output. There-
fore, only the remaining m−1 state variables, grouped in
x
r
m
t ( ) ∈

R
1
, need to be estimated.
We now follow standard arguments outlined, for exam-
ple, in [6]. By defining q x J ( ) ( ) ( ) t t y t
r
= − , it is shown in [6]
that the system
[ ]
[ ]
&
( ) ( )
( )
q J I A
I
q
J I A
J
t t
y t
m
m
m
= −






+ −






+ −



1
1
1
0
1
[ ]
J I B
m− 1
u t ( )
(14)
$
( ) ( ) ( ) x
I
q
J
t t y t
m
=






+







0 1
1
(15)
is a reduced-order observer, where J ∈

R
m 1
is the observer
gain matrix. This observer is shown as a block diagram in
Figure 4.
Given the state-space model in (6)-(7) and the re-
duced-order observer (14)-(15), the dynamics of the estima-
tion error
~
( ) ( )
$
( ) x x x
r r r
t t t = − are described by
[ ]
~
&
( )
~
( ) ( )
~
( ) x J I A
I
x A JA x
r
m-1
r 22 12 r
t t t
m
= −






= −
− 1
0
,
(16)
where matrix Ais partitionedas in (12). We can also see that
the eigenvalues of the reduced-order observer (14)-(15) are
the roots of the polynomial
E s s ( ) ( ) = − + det I A JA
m-1 22 12
. (17)
60 IEEE Control Systems Magazine August 2003
Figure 4. Block diagram for a reduced-order observer.
Authorized licensed use limited to: University of Michigan Library. Downloaded on April 09,2010 at 18:24:32 UTC from IEEE Xplore. Restrictions apply.
It can be shown that complete observability of ( ) A, C
guarantees complete observability for ( , ) A A
22 12
(see the
Appendix). Thus, the observer eigenvalues can be placed
anywhere onthe complex plane Cby choosing the gainJ ap-
propriately.
In particular, the reduced-order observer
1) estimates only m−1state variables
2) has a direct link between its input y t ( ) and its output
$
( ) x t (i.e., is not strictly proper).
Although the full-order observer is preferred for applica-
tions (see, e.g., [7]) and is less sensitive to noise in y t ( ), the
second of the properties above is fundamental in under-
standing the link between biproper controllers and ob-
served-state feedback. This aspect will be explained in the
following sections.
Observed State Feedback
We now consider the control of the plant output, y t ( ), using
the estimated state provided by a state observer. The sim-
plest way to do this (and the optimal way under certain con-
ditions [1]) is to use a linear control law, where the plant
inputu t ( ) is a linear combination of estimated states in
$
( ) x t ,
through a gain matrix K. Figure 5 shows the corresponding
feedback control architecture, where r t ( ) is an external sig-
nal relatedtothe reference r t ( )inFigure 1 ina way tobe clar-
ified soon.
The feedback law is then defined by
u t r t t ( ) ( )
$
( ) = − Kx . (18)
When the state estimate is obtained using a full-order ob-
server, the closed-loop poles, which determine the control
system behavior, are the 2mzeros of the polynomial
A s s s
cl
F s E s
( ) ( ) ( )
( ) (
= − + ⋅ − + det det I A BK I A JC
m m
1 2 444 3 444
)
1 2 444 3 444
. (19)
We see that the separation principle [1] holds since the
closed-loop pole set is independently determined by the
feedback gain K and the observer gain J. Using (19), we can
choose K and J to assign the closed-loop eigenvalues at de-
sired locations in the complex plane (see [7] and [8]).
On the other hand, if we use the control scheme shown in
Figure 5 with a reduced-order observer, it can be proved [4],
[9] that the closed-loop poles are the 2 1 m− zeros of
A s s s
cl
F s
( ) ( ) (
( )
= − + ⋅ − + det det I A BK I A JA
m m-1 22
1 2 444 3 444
12
)
( ) E s
1 2 4444 3 4444
. (20)
We again see that the separation principle holds for the
set of closed-loop poles since the reduced-order observer
gain J and the feedback gain Kindependently determine the
eigenvalues of the closed-loop system.
Note that the system in Figure 5 may not be completely
controllable. In that case, the set of closed-loop poles due to
F s s ( ) ( ) = − + det I A BK
m
includes the eigenvalues of the un-
controllable subspace.
Equivalence
We next turn to an interpretation of the observer-state feed-
back strategy from the classical control standpoint.
As already notedinthe Introduction, whena full-order ob-
server is used, the equivalent controller is described by
C s s ( ) ( ) = − + +

K I A JC BK J
m
1
. (21)
Alternatively, the controller can be described in
state-space form:
&
( ) ( ) ( ) ( ( ) ( )) x A JC BK x J
c c
t t r t y t = − − + − , (22)
u t t ( ) ( ) = Kx
c
. (23)
Note that for the full-order observer case:
• The controller has m eigenvalues, which are the m
roots of
det( ) sI A JC BK
m
− + + = 0. (24)
• Zerosteady-state error is not guaranteedwhenthe ref-
erence signal is constant. This follows, since the con-
troller does not generally have poles at s = 0.
• The controller is strictly proper, since there is no direct
link between the input and the output.
The equivalence requires that r t ( ) in Figure 5 and r t ( ) in
Figure 1 be related by
R s s R s
F s
r
( ) ( ) ( )
( )
= − + ⋅

K I A JC J
m
1
1 2 444 3 444
, (25)
where F s
r
( ) is a stable reference prefilter.
Onthe other hand, if a reduced-order observer is used, the
equivalent controller takes the form
C s s D
m c
( ) ( ) = − +


C I A B
c c c 1
1
, (26)
August 2003 IEEE Control Systems Magazine 61
Figure 5. Observed-state feedback.
Authorized licensed use limited to: University of Michigan Library. Downloaded on April 09,2010 at 18:24:32 UTC from IEEE Xplore. Restrictions apply.
where
[ ]
A J I A BK
I
c m-1
m-1
− −

¸

1
]
1
( )
0
,
(27)
[ ]
B J I A BK
J
c m-1
− −

¸

1
]
1
( )
1
,
(28)
C K
I
c
m-1

¸

1
]
1
0
,
(29)
D
c


¸

1
]
1
K
J
1
.
(30)
We note that for the reduced-order observer case:
• The controller has m−1 eigenvalues, given by the
equation
[ ]
det
det
s
s
I J I A BK
I
I A
m-1 m-1
m-1
m-1
− − −

¸

1
]
1
¸
¸

_
,


( )
(
0
22 2 2 12 2
B K JA JK + + − b
1
0 )
(31)
where A and B are partitioned as in (12) and the feed-
backgainmatrixis K K
2
[ ] k
1
, k
1
∈R, andK
2
T m


R
1
.
• The controller does not necessarily have poles at the
origin, so zero steady-state error is not guaranteed
when the reference signal (or any disturbance) is con-
stant. A similar situation applies for sinusoidal refer-
ences and disturbances with respect to (nonzero)
controller poles on the imaginary axis.
• In general, this controller is biproper because the
state-space model (26) has a direct link between input
and output through the matrix D
c
. However, this con-
troller can also become strictly proper when K
T
and
[ ] 1 J
T T
are orthogonal.
• When the system under state estimate feedback con-
trol, as in Figure 5, is the same as the plant in Figure 1,
then m n and the equivalent controller, with re-
duced-order observer, will have n −1 poles. The dis-
tinction between nandmis necessary, as shown in the
next section.
The expression for the reference prefilter, analogous to
(25), is derived in [4].
The Internal Model Principle
Disturbances are always present in a real control system.
One way to reduce their deleterious effect on the con-
trolled variable [the plant output y t ( )] is to use observers.
This can be achieved by augmenting the state-space model
to include a model for the disturbance. Below we present a
brief sketch of how that is usually done. We will consider
the case showninFigure 6, where d t ( )is a disturbance pres-
ent on the plant input.
We assume that the disturbances belong to known
classes of signals. For example, we can model a disturbance
as a constant or as a sinusoidal signal of specific frequency
but unknown amplitude and phase. In these cases, we can
build an uncontrollable state-space model without input
and whose output is the disturbance
&
( ) ( ) x A x
d d d
t t , (32)
d t t ( ) ( ) C x
d d
, (33)
where A
d

×
R
n n
d d
and C
d
T n
d
∈R .
Next consider the nominal plant model
&
( ) ( ) ( ) x A x B
o o o o
t t u t
p
+ , (34)
y t t ( ) ( ) C x
o o
, (35)
where A
o

×
R
n n
and B C
o o
,
T n
∈R .
The state-space description (34)-(35) can be then com-
bined with (32)-(33) [10] to build a composite model for the
system in Figure 6, where the disturbance corresponds to a
linear combination of uncontrollable states of the aug-
mented system. This yields
&
( )
&
( )
( )
( )
x
x
A B C
0 A
x
x
o
d
o o d
d
o
d
t
t
t
t

¸

1
]
1


¸

1
]
1

¸

1
]
1
+

¸

1
]
1
B
0
o
u t ( ),
(36)
[ ]
y t
t
t
( )
( )
( )

¸

1
]
1
C 0
x
x
o
o
d
.
(37)
Note that this composite system is of order m n n
d
+ . A
state observer for this systemwill provide estimates for the
plant state and for the state of the disturbance model. Thus,
62 IEEE Control Systems Magazine August 2003
Figure 6. Plant with input disturbance.
Authorized licensed use limited to: University of Michigan Library. Downloaded on April 09,2010 at 18:24:32 UTC from IEEE Xplore. Restrictions apply.
wecancompensate thedisturbance effect ontheplant input
by selecting a suitable form of the feedback gain matrix
[ ]
K K C
o d
= . (38)
Therefore, the control signal is constructed as
u t r t t t
d t
( ) ( )
$
( )
$
( )
$
( )
= − − K x C x
o o d d
12 4 3 4
, (39)
leading to the plant input
u t r t t d t t
p
( ) ( )
$
( ) ( ( )
$
( )) = − + − K x C x
o o d d
. (40)
So far, the above results apply to full-order as well as re-
duced-order observers. Naturally, a different controller will
arise in each case.
If a full-order observer is used, the equivalent controller
poles are the m n n
d
= + solutions of the equation
det det ( ) ( ) s s I A J C B K I A
n o o o o o n d
d
− + + − = 0. (41)
Onthe other hand, if a reduced-order observer is used, the
controller poles are the m n n
d
− = − + 1 1 roots of
det det ( ( ) ) ( ) s b s I J A K A B K I A
n-1 o 12 2 22 2 2 n d
d
+ − − + ⋅ − =
1
0.
(42)
This decomposition is commonly called the internal
model principle [11]. We see that to
fully compensate any disturbance in a
control loop, its eigenvalues must be in-
cluded as controller poles.
In summary, we see that if input
di sturbances are model ed as
noncontrol l abl e states, and ob-
server-state feedback is used in an ap-
propriate way, then the disturbance
model eigenvalues will appear as poles
of the controller. For example, if we as-
sume a constant (or a step) input distur-
bance, the equivalent controller for the
observer-state feedback scheme will be
a controller having integral action (i.e.,
with a pole at s = 0).
The above results are valuable per
se. We have summarized a useful strat-
egy to compensate disturbances and to
track references in a control loop. How-
ever, it also sets up the necessary
framework for the main result of this ar-
ticle. This mechanism explains why the
controller might have a number of poles
that exceeds the order of the plant
model. In other words, this strategy will
allow us to interpret a classical control loop in terms of ob-
server-state feedback, even when the controller is more
complex than the plant model.
Revealing the Equivalence
Next we proceed to our core question: Given a biproper or
strictly proper controller, what is the equivalent observer-
state feedback scheme?
Our subsequent development depends on the following
facts:
F.1) The plant model has a transfer function (3) and a
completely controllable and observable state-space
description of order n, having the form (34)-(35).
F.2) The controller has a transfer functionas in(4) (possi-
bly biproper) of order n
c
.
F.3) As a consequence of F.1 and F.2, all transfer functions in
the classical closed loop have a total of n n
c
+ poles,
givenbythezerosof thepolynomial A s
cl
( )definedin(5).
We will cover all cases regarding the order of the control-
ler and its relative degree (properness); however, as we
have already seen, if a reduced-order observer is used, then
the equivalent controller can be either biproper or strictly
proper. Thus, we will be able to elucidate the equivalences
only when reduced-order observers are used. Therefore,
the cases to deal with will be classified according to the or-
der of the controller with respect to the plant order. Three
cases will be distinguished.
August 2003 IEEE Control Systems Magazine 63
Figure 7. Unit step responses for Example 1.
Figure 8. Unit step responses for Example 2.
Authorized licensed use limited to: University of Michigan Library. Downloaded on April 09,2010 at 18:24:32 UTC from IEEE Xplore. Restrictions apply.
The Prototype Case: n n
c
= − 1
In this situation, the polynomials E s ( ) and F s ( ) can be di-
rectly computed by factoring A s
cl
( ) defined in (5). Once we
have E s ( ) and F s ( ), we can compute J and K from (20).
However, we recognize that there may be more than one
state estimate feedback equivalent scheme, since there
might be different ways to assign the closed-loop poles to
E s ( ) and F s ( ) in (20). Whenever possible, the observer
poles will be assigned in such a way that the estimation er-
ror will converge faster than the state feedback modes.
Example 1
Consider a plant of order n = 3 having transfer function
G s
s
s
o
( )
( )
=
+
+
2 1
1
3
,
controlledbyastrictlyproper integratingcontroller (n
c
= 2)
C s
s
s s
( )
. ( )
( . )
=
+
+
09 1
011 1
.
In Figure 7 we show the closed-loop response to a unit
step reference and the plant open-loop unit step response.
The closed-loop characteristic polynomial is computed
using (5). This yields
A s s s s s s
cl
( ) ( )( . )( . )( . . ) = + + + + + 1 9315 04256 135 2064
2
.
This is a prototype case since n n
c
= −1. Thus, the polyno-
mials E s ( ) and F s ( ) can be evaluated immediately. The
E-polynomial is of order 2 and the F-polynomial is of order 3.
The fastest poles (at s = −1and at s = −9315 . ) are assigned to
E s ( ), and the remaining poles, to F s ( ). Thus
E s s s s
F s s
( ) . . ( ),
( ) (
= + + = − +
= +
2
2
10351 9315
1
det I A JA
2 22 12
. . )( . ) ( ). 35 2064 04256 s s s + + = − + det I A BK
3
To compute J and K we use the state-space description
(observer form [5])
A A B B C C
o o
= =













= =










=
3 1 0
3 0 1
1 0 0
0
2
1
, ,
o
=










1
0
0
T
.
The above state description is used to compute J and K
(use the MATLAB command place)
[ ]
[ ]
J
K
=
= −
10315 9315
22081 11058 09876
. . ,
. . . .
T
If one uses (27)-(30) and the gains J and K obtained, the
original controller is obtained, as expected.
The Low-Complexity Controller Case: n n
c
< − 1
The gap between this case and the prototype case can be
bridged by assuming that n n
c
− −1stable cancellations arise
between poles and zeros in the controller. Thus, to build the
equivalence, we define the equivalent closed-loop charac-
teristic polynomial,
~
( ) A s
cl
, as
~
( ) ( ) ( ) A s A s W s
cl cl
= , (43)
where A s
cl
( )is definedin (5) andW s ( )is a stable polynomial
of order n n
c
− −1.
Since the choice of W s ( ) is arbitrary, there will be an infi-
nitenumber of equivalent observer-statefeedbackschemes.
Example 2
Consider a plant of order n = 3 having the transfer function
G s
s
s s s
o
( )
. .
=
+
+ + +
1
5 72 24
3 2
,
controlled by PI controller (n
c
=1)
C s
s
s
( )
.
=
+ 14 2
.
Figure 8 shows the closed-loop response to a unit step
reference and the plant open-loop unit step response.
The closed-loop polynomial can next be computed using
(5) to obtain
A s s s s s
cl
( ) . . = + + + +
4 3 2
5 86 5 8 2.
Furthermore, since n
c
=1 and n = 3, then n n
c
< −1. We
thus have to add (n n
c
− − 1 ) stable cancellations in the con-
troller. To this end, the controller is assumed to have a pole
(and a zero) at s = −5; that is, a factor W s s ( ) = + 5 is ap-
pendedtotheequivalent closed-looppolynomial leadingto
~
( ) ( ) ( ) . . A s W s A s s s s s s
cl cl
= = + + + + +
5 4 3 2
10 336 48 8 31 10,
havi ng roots at { . . , . . , } − ± − ± − 04385 04978 20615 05423 5 j j .
Also, since the plant has three states, we require that F s ( )
should be of order 3 and E s ( ) should be of order 2.
We assign two fast poles to the observer polynomial E s ( )
(at s j = − ± 20615 05423 . . ) and the remaining three to the
feedback polynomial F s ( ). This yields
E s s s s
F s s
( ) . . ( ),
( ) ( .
= + + = − +
= +
2
2
4123 4544
0
det I A JA
2 22 12
8771 04402 5 s s s + + = − + . )( ) ( ). det I A BK
3
64 IEEE Control Systems Magazine August 2003
Authorized licensed use limited to: University of Michigan Library. Downloaded on April 09,2010 at 18:24:32 UTC from IEEE Xplore. Restrictions apply.
Tocompute J andKwe use a state-space description(ob-
server form [5])
A A B B
o o
= =













= =










5 1 0
72 0 1
24 0 0
0
1
1
.
.
, , C C
o
= =










1
0
0
T
.
Theabovestatedescriptionis usedtocomputeJ andK
[ ]
[ ]
J
K
=
= − −
4129 45437
3 8215 29383 3 8154
. . ,
. . . .
T
Again, if we use these gains J and Kin (27)-(30), the origi-
nal classical controller is obtained.
The High-Complexity
Controller Case: n n
c
> − 1
When n n
c
> −1, the equivalence can be established if
n n n
d c
= − +1controller poles are assumed to originate from
the internal model principle. This implies that these poles
form a fictitious disturbance model leading to an augmented
plant of order m n n n
d c
= + = +1. Since the corresponding
eigenvalues correspond to uncontrollable but observable
states, they only yield an increase in the degree of E s ( )with-
out affecting the degree of F s ( ).
Example 3
Consider a plant of order n = 2
G s
s s
o
( ) =
+ +
10
4 13
2
controlled by a PID controller (n
c
= 2)
C s
s s
s s
( )
( . )
=
+ +
+
2
5 20
0003 1
.
Figure 9 shows the closed-loop response to a unit step
reference and the plant open-loop unit step response.
The closed-loop characteristic polynomial is computed
using (5). This yields
A s s s s s
cl
( ) ( )( . )( . . ) = + + + + 323 9 887 4399 20 87
2
.
Inthis case, n n
c
− − = ( ) 1 1. We thenuse the internal model
principle to augment the plant by n n
c
− + = 1 1 uncontrolla-
ble states. As already discussed, the pole associated with
this state must be a controller pole. Hence, the degree of
E s ( )is set equal to two, whichcorresponds to m−1, where m
is the number of states in the augmented plant. On the other
hand, the degree of F s ( ) is two, which corresponds to the
number of plant states (they are controllable states by as-
sumption).
In this case, we will associate the disturbance with a pole
at the origin; that is, we use the model (32)-(33) with A
d
= 0
and C
d
=1.
We assign the two fastest poles (at s = −323 and at
s = −9 887 . ) to the observer polynomial E s ( ) and the re-
maining poles to the feedback polynomial F s ( )
E s s s s
F s s
( ) ( )( . ) ( ),
( ) .
= + + = − +
= +
323 9 887
4
2
det I A JA
2 22 12
399 20 87 s s + = − + . ( ). det I A B K
2 o o o
To compute J and K we use the state-space description
(observer form) for the augmented system, as in (36)-(37)
A B C =












=










=


4 1 0
13 0 10
0 0 0
0
10
0
1
0
0
, ,








T
.
The above state-space description can be used to com-
pute J and K
o
[ ]
[ ]
J
K
o
=
=
3329344 319 3922
06275 00399
. . ,
. . .
T
The uncontrollable pole at s = 0 can-
not be shifted by feedback and will be
preservedusing a feedback gainKof the
following form:
[ ]
K = 06275 00399 10000 . . . .
If one uses (27)-(30) and the gains J
and K obtained, the original classical
controller is obtained.
Discrete-Time Case
In the previous sections, we covered
only continuous-time systems, but the
August 2003 IEEE Control Systems Magazine 65
Figure 9. Unit step responses for Example 3.
Authorized licensed use limited to: University of Michigan Library. Downloaded on April 09,2010 at 18:24:32 UTC from IEEE Xplore. Restrictions apply.
prior results apply directly to the discrete-time case. We
present an example to illustrate this.
Example 4
Consider the third-order discrete-time plant
G z
z
z z z
o
[ ]
.
( . )( . )
=
+
− −
05
09 0 8
controlled by a discrete-time PI controller (n
c
=1):
C z
z
z
[ ]
. ( )
=


0001 2 1
1
.
The closed-loop polynomial can be computed using (5)
to obtain
A z z z z z
cl
[ ] . . . . = − + − −
4 3 2
2 7 2422 0 72 00005.
This is a low-complexity controller since n
c
=1 and n = 3
(i.e., n n
c
< −1). Therefore, we must add n n
c
− − 1 stable can-
cellations in the controller. To this end, the controller is as-
sumed to have a pole (and a zero) at z = 05 . ; that is, a factor
W z z [ ] . = −05 is appended to the equivalent closed-loop
polynomial leading to
~
[ ] [ ] [ ]
( . )( . )
( .
A z A z W z
z z
z
cl cl
=
= − +
× −
0 7701 00006928
19
2
31 09371 05 z z + − . )( . )
having zeros at − ± 00006928 09653 00729 0 7701 05 . , . . , . , . j . Also,
since the plant has three states, F z [ ] should be of order 3
and E z [ ] of order 2.
We assign two fast poles to the observer polynomial E z [ ]
(at z = −00006928 . and at z = 05 . ) and the remaining three to
the feedback polynomial F z [ ]:
E z z z z
F z z
[ ] . . ( ),
[ ]
= − − = − +
= −
2
3
04993 00003
2
det I A JA
2 22 12
. . . ( ). 7011 24242 0 7217
2
z z s + − = − + det I A BK
3
To compute J and K we use the state-space description
(observer form)
A A B B
o o
= = −










= =









17 1 0
0 72 0 1
0 0 0
0
1
05
.
. ,
. 
= =










, C C
o
1
0
0
T
.
The above state description can be used to compute J
and K:
[ ]
[ ]
J
K
= − −
= −
0493 00003
00017 00011 0688
. . ,
. . . .
T
Using these gains J and K in (27)-(30), the original classi-
cal controller is obtained.
Conclusions
We have shown that the use of reduced-order observers al-
lows any SISO classical linear controller to be translated
into a control architecture based on feedback of an esti-
mated system state. This result allows shows that classical
controllers add dynamics to the control loop that originate
from a plant state observer plus possible application of the
internal model principle.
It is alsointeresting tonote that anobserver-state feedback
control law leads to a unique equivalent classical controller
(except for a reference feedforward block). However, the con-
verse is not true ingeneral: givena classical control loopanda
fixed plant state-space description, there might be many (in-
deed, an infinite number of) equivalent observer-state feed-
back control systems. This applies particularly when we deal
with low-complexity controllers such as PIDcontrollers. In ad-
dition, the classical loop performance regarding disturbance
compensation, noise attenuation, and robustness is com-
pletely determined by the (classical) controller transfer func-
tion. Therefore, there might exist an infinite number of
estimated state-based controllers that yield the same
closed-loop performance. This leads to an interesting related
question: Given that the equivalent observer-based controller
is not unique, howcansensible performance specifications be
introduced in the design of this control law? Perhaps numeri-
cal or other issues may be relevant in this regard.
This article does not advocate that classical control only
be taught using the state-space synthesis methodology, but
it shows that, at a more advanced level, both approaches
can and should be brought together for the benefit of a
deeper understanding of the fundamental issues in control
system design.
The extension of these ideas to the SISO discrete-time
case is straightforward. In particular, sampled-data mod-
els allow one to deal with systems with pure time delays.
However, as investigated in [4], the extension to multi-
ple-input, multiple-output (MIMO) systems (either con-
tinuous-time or discrete-time) is not trivial and involves
more complex issues.
Acknowledgments
The authors gratefully acknowledge Prof. G.C. Goodwin of
the School of Electrical Engineering and Computer Science,
the University of Newcastle, for his valuable comments and
suggestions.
References
[1] G.C. Goodwin, S. Graebe, and M.E. Salgado, Control System Design.
Englewood Cliffs, NJ: Prentice-Hall, 2001.
[2] D. Luenberger, “An introduction to observers,” IEEE Trans. Automat.
Contr., vol. 16, no. 6, pp. 596-602, June 1971.
[3] K. Ogata, State Space Analysis of Control Systems. Englewood Cliffs, NJ:
Prentice-Hall, 1967.
66 IEEE Control Systems Magazine August 2003
Authorized licensed use limited to: University of Michigan Library. Downloaded on April 09,2010 at 18:24:32 UTC from IEEE Xplore. Restrictions apply.
[4] J.I. Yuz, “Equivalencia entre controladores clásicos y control del estado
observado,” Master’s thesis, Departamento de Electrónica, UTFSM,
Valparaiso, Chile, Dec. 2001.
[5] T. Kailath, Linear Systems. Englewood Cliffs, NJ: Prentice-Hall, 1980.
[6] H. Kwakernaak and R. Sivan, Linear Optimal Control Systems. New York:
Wiley-Interscience,1972.
[7] G. Duan, S. Thompson, and GLiu, “Separation principle for robust pole as-
signment-an advantage of full-order state observers,” in Proc. 38th IEEE Conf.
Decision and Control, 1999, vol. 41, pp. 76-78.
[8] J. Kautsky, N.K. Nichols, andP. vanDooren, “Robust pole assignment inlin-
ear state feedback,” Int. J. Contr., vol. 41, pp. 1129-1155, 1985.
[9] G.F. Franklin, J.D. Powell, and A. Emami-Naeini, Feedback Control of Dy-
namic Systems. Reading, MA: Addison-Wesley, 1991.
[10] K. Zhou, J. Doyle, and K. Glover, Robust and Optimal Control. Englewood
Cliffs, NJ: Prentice-Hall, 1996.
[11] B.A. Francis and W. Wonham, “The internal model principle of control
theory,” Automatica, vol. 12, pp. 457-465, 1976.
Juan I. Yuz received his B.Eng. and M.Sc. degrees in elec-
tronics engineering from the Universidad Técnica
Federico Santa María, Valparaíso, Chile, in 2001. He is cur-
rently a Ph.D. student in electrical engineering at the Uni-
versity of Newcastle, Australia. His research areas include
performance limitations, control with constraints, and
system identification.
Mario E. Salgado received his professional title of
Ingeniero Civil Electrónico from the Universidad Técnica
Federico Santa María, Valparaíso, Chile, in 1974. He re-
ceived the M.Sc. degree from Imperial College, London, in
1979 and the Ph.D. degree in electrical engineering from
the University of Newcastle, Australia, in 1989. He is cur-
rently an academic with the Department of Electronic En-
gineering, Universidad Técnica Federico Santa María. His
research areas include control systemdesign and system
identification. He can be reached at the Department of
Electronic Engineering, Universidad Técnica Federico
Santa María (UTFSM), Casilla 110-V, Valparaíso, Chile,
mario.salgado@elo.utfsm.cl.
Appendix
Complete Observability
One way to test complete controllability and observability
of a systemis providedby the following lemma knownas the
Popov-Belevitch-Hautus (PBH) test, whichwe recall without
proof (for a complete proof see [5]).
Lemma 1
Consider a state-space model given by the matrices
( , , ) A B C . Then:
i) The system is not completely observable if and only if
there exists a nonzero vector x ∈C
n
and a scalar
λ ∈C such that
Ax x = λ , (44)
Cx = 0. (45)
ii) The systemis not completely controllable if andonly if
there exists a nonzero vector x ∈C
n
and a scalar
λ ∈C such that
x A x
T T
= λ , (46)
x B
T
= 0. (47)
We will use the PBH test to prove that the complete
observability of ( , ) A C implies the complete observability
for ( , ) A A
22 12
.
Lemma 2
If the pair ( , ) A C is completely observable, then the pair
( , ) A A
22 12
is completely observable too, where the matrices
A
22
and A
12
are defined in (12).
Proof
We use contradiction. We first assume that the pair ( , ) A C is
completely observable and the pair ( , ) A A
22 12
is not com-
pletely observable. Then the PBHtest implies that there is a
nonzero vector x
2


C
n 1
and a scalar λ
2
∈C such that
A x x
22 2 2
= λ
2
(48)
A x
12 2
= 0. (49)
If we next consider the vector x x
2
= [ , ] 0
T T
, we have that
Ax
A
A A x
A x
A x
12
21 22 2
12 2
22 2
=












=






=
a
11
0 0
λ
2
2
x
x
2






= λ .
(50)
And if we write the matrix C as in (13), we have that
[ ]
Cx
x
2
=






= 1 0 0
0
0 K .
(51)
The results givenin(50) and(51) showthat there is a vec-
tor x ∈C
n
anda scalar λ
2
∈Csatisfying (44), (45) sothe pair
( , ) A C cannot be completely observable. This contradicts
the initial assumption. Thus, ( , ) A A
22 12
must be completely
observable.
August 2003 IEEE Control Systems Magazine 67
Authorized licensed use limited to: University of Michigan Library. Downloaded on April 09,2010 at 18:24:32 UTC from IEEE Xplore. Restrictions apply.