Chapter 18 ©

Goodwin, Graebe, Salgado , Prentice Hall 2000

Chapter 18

Synthesis via State Space

Methods
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Here, we will give a state space interpretation to

many of the results described earlier. In a sense, this
will duplicate the earlier work. Our reason for doing
so, however, is to gain additional insight into linear-
feedback systems. Also, it will turn out that the
alternative state space formulation carries over more
naturally to the multivariable case.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Results to be presented include

◆ pole assignment by state-variable feedback
◆ design of observers to reconstruct missing states from
available output measurements
◆ combining state feedback with an observer
◆ transfer-function interpretation
◆ dealing with disturbances in state-variable feedback
◆ reinterpretation of the affine parameterization of all
stabilizing controllers.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Pole Assignment by State Feedback

We begin by examining the problem of closed-loop
pole assignment. For the moment, we make a
simplifying assumption that all of the system states
are measured. We will remove this assumption later.
We will also assume that the system is completely
controllable. The following result then shows that
the closed-loop poles of the system can be arbitrarily
assigned by feeding back the state through a suitably
chosen constant-gain vector.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Lemma 18.1: Consider the state space nominal

model

Let r (t ) denote an external signal.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Then, provided that the pair (A0, B0) is completely

controllable, there exists

such that the closed-loop characteristic polynomial is

Acl (s), where Acl (s) is an arbitrary polynomial of degree n.

Proof: See the book.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Note that state feedback does not introduce additional dynamics

in the loop, because the scheme is based only on proportional
feedback of certain system variables. We can easily determine
the overall transfer function from to y(t). It is given by

r (t )
where

and Adj stands for adjoint matrices.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

We can further simplify the expression given above.

To do this, we will need to use the following results
from Linear Algebra.
Lemma 18.2: (Matrix inversion lemma). Consider
three matrices, A ∈ n× n, B ∈ n× m, C ∈ m× n.
Then, if A + BC is nonsingular, we have that

Proof: See the book.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

In the case for which B = g ∈ n and CT = h ∈ n, the

above result becomes
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Lemma 18.3: Given a matrix W ∈ n× n and a pair of

arbitrary vectors φ 1 ∈ n and φ 2 ∈ n, then
provided that W and W + φ 1φ 2 , are nonsingular,
T

Proof: See the book.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Application of Lemma 18.3 to equation

leads to

we then see that the right-hand side of the above expression

is the numerator B0(s) of the nominal model, G0(s). Hence,
state feedback assigns the closed-loop poles to a prescribed
position, while the zeros in the overall transfer function
remain the same as those of the plant model.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

State feedback encompasses the essence of many

fundamental ideas in control design and lies at the
core of many design strategies. However, this
approach requires that all states be measured. In
most cases, this is an unrealistic requirement. For
that reason, the idea of observers is introduced next,
as a mechanism for estimating the states from the
available measurements.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Observers
Consider again the state space model

A general linear observer then takes the form

where the matrix J is called the observer gain and is

xˆ ( t ) the state estimate.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The term

is known as the innovation process. For nonzero J

v(t) represents the feedback error between the
observation and the predicted model output.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The following result shows how the observer gain J

can be chosen such that the error, ~
x ( t ) defined as

can be made to decay at any desired rate.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Lemma 18.4: The estimation error ~

x ( t ) satisfies

Moreover, provided the model is completely

observable, then the eigenvalues of (A0 - JC0) can be
arbitrarily assigned by choice of J.

Proof: See the book.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Example 18.1: Tank-level estimation

As a simple application of a linear observer to
estimate states, we consider the problem of two
coupled tanks in which only the height of the liquid
in the second tank is actually measured but where we
are also interested in estimating the height of the
liquid in the first tank. We will design a virtual
sensor for this task.
A photograph is given on the next slide.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Coupled Tanks Apparatus

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 18.1: Schematic diagram of two coupled tanks

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Water flows into the first tank through pump 1 a rate fi(t)
that obviously affects the head of water in tank 1 (denoted
by h1(t)). Water flows out of tank 1 into tank 2 at a rate
f12 (t), affecting both h1(t) and h2(t). Water than flows out
of tank 2 at a rate fe controlled by pump 2.

Given this information, the challenge is to build a virtual

sensor (or observer) to estimate the height of liquid in
tank 1 from measurements of the height of liquid in tank 2
and the flows f1(t) and f2(t).
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Before we continue with the observer design, we first make a model

of the system. The height of liquid in tank 1 can be described by the
equation

Similarly, h2(t) is described by

The flow between the two tanks can be approximated by the free-
fall velocity for the difference in height between the two tanks:
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

We can linearize this model for a nominal steady-

state height difference (or operating point). Let

This yields the following linear model:

where
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

We are assuming that h2(t) can be measured and h1(t) cannot, so we set
C = [0 1] and D = [0 0]. The resulting system is both controllable
and observable (as you can easily verify). Now we wish to design an
observer

to estimate the value of h2(t). The characteristic polynomial of the

observer is readily seen to be

so we can choose the observer poles; that choice gives us values for J1
and J2.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

If we assume that the operating point is H = 10%,

then k = 0.0411. If we wanted poles at s = -0.9291
and s = -0.0531, then we would calculate that J1 = 0.3
and J2 = 0.9. If we wanted two poles at s = -2, then J2
= 3.9178 and J1 = 93.41.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The equation for the final observer is then

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The data below has been collected from the real

system shown earlier
Set point for height in tank 2 (%)
80
Percent

70
60
50
40
0 50 100 150 200 250 300 350 400
Time (sec)
Actual height in tank 2 (%)
80
Percent

70
60
50
40
0 50 100 150 200 250 300 350 400
Time (sec)
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The performance of the observer for tank height is

compared below with the true tank height which is
actually measured on this system.
85
80
75 Actual height in tank 1 (blue),
Observed height in tank 1 (red)
Percent

70
65
60
55
50
45
40 0 50 100 150 200 250 300 350 400
Time (sec)
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Combining State Feedback with

an Observer
A reasonable conjecture arising from the last two sections is that it
would be a good idea, in the presence of unmeasurable states, to
proceed by estimating these states via an observer and then to
complete the feedback control strategy by feeding back these
estimates in lieu of the true states. Such a strategy is indeed very
appealing, because it separates the task of observer design from that
of controller design. A-priori, however, it is not clear how the
observer poles and the state feedback interact. The following
theorem shows that the resultant closed-loop poles are the
combination of the observer and the state-feedback poles.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Separation Theorem
Theorem 18.1: (Separation theorem). Consider the
state space model and assume that it is completely
controllable and completely observable. Consider
also an associated observer and state-variable
feedback, where the state estimates are used in lieu
of the true states:
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Then
(i) the closed-loop poles are the combination of the poles from the
observer and the poles that would have resulted from using the
same feedback on the true states - specifically, the closed-loop
polynomial Acl (s) is given by

(ii) The state-estimation error cannot be controlled from the

external signal . ~
x (t )
r (t )
Proof: See the book.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The above theorem makes a very compelling case for

the use of state-estimate feedback. However, the reader
is cautioned that the location of closed-loop poles is
only one among many factors that come into control-
system design. Indeed, we shall see later that state-
estimate feedback is not a panacea. Indeed it is subject
to the same issues of sensitivity to disturbances, model
errors, etc. as all feedback solutions. In particular, all of
the schemes turn out to be essentially identical.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Transfer-Function Interpretations
In the material presented above, we have developed
a seemingly different approach to SISO linear
control-systems synthesis. This could leave the
reader wondering what the connection is between
this and the transfer-function ideas presented earlier.
We next show that these two methods are actually
different ways of expressing the same result.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Transfer-Function Form of Observer

We first give a transfer-function interpretation to the
observer. We recall that the state space observer
takes the form

where J is the observer gain and xˆ ( t ) is the state

estimate.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

A transfer-function interpretation for this observer is given in

the following lemma.
Lemma 18.5: The Laplace transform of the state estimate
has the following properties:
(a) The estimate can be expressed in transfer-function form as:

where T1(s) and T2(s) are the following two stable transfer
functions:
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

(b) The estimate is related to the input and initial conditions by

where f0(s) is a polynomial vector in s with coefficients depending

linearly on the initial conditions of the error
~
x ( t ).
(c) The estimate is unbiased in the sense that

where G0(s) is the nominal plant model.

Proof: See the book.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Transfer-Function Form of State-

Estimate Feedback
We next give a transfer-function interpretation to the
interconnection of an observer with state-variable feedback.
The key result is described in the following lemma.
Lemma 18.6:
(a) The state-estimate feedback law

can be expressed in transfer-function form as

where E(s) is the polynomial defined previously.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

In the above feedback law

where K is the feedback gain and J is the observer gain.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

(b) The closed-loop characteristic polynomial is

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

(c) The transfer function from R ( t ) to Y(s) is given by

where B0(s) and A0(s) are the numerator and

denominator of the nominal loop respectively. P(s) and
L(s) are the polynomials defined above.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The foregoing lemma shows that polynomial pole

assignment and state-estimate feedback lead to the
same result. Thus, the only difference is in the terms
of implementation.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The combination of observer and state-estimate feedback has

some simple interpretations in terms of a standard feedback loop.
A first possible interpretation derives directly from

by expressing the controller output as

This is graphically depicted in part (a) of Figure 18.2 on the

following slide. We see that this is a two-degree-of-freedom
control loop.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 18.2: Separation theorem in standard loop forms

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

A standard one-degree-of-freedom loop can be

obtained if we generate r (t ) from the loop reference
r(t) as follows:

We then have

This corresponds to the one-degree-of-freedom loop

shown in part (b) of Figure 18.2.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Note that the feedback controller can be

implemented as a system defined, in state space
form, by the 4-tuple (A0 - JC0 - B0K, J, K, 0).
(MATLAB provides a special command, reg, to
obtain the transfer function form.)
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Transfer Function for Innovation

Process
We finally give an interpretation to the innovation process. Recall
that

This equation can also be expressed in terms of Laplace transfer

functions by using

as

We can use the above result to express the innovation process v(t)
in terms of the original plant transfer function. In particular, we
have the next lemma.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Lemma 18.7: Consider the state space model and the

associated nominal transfer function G0(s) = B0(s)/A0(s).
Then the innovations process, v(t), can be expressed as

where E(s) is the observer polynomial (called the observer

characteristic polynomial).

Proof: See the book.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Reinterpretation of the Affine

Parameterization of all Stabilizing Controllers

We recall the parameterization of all stabilizing

controllers (see Figure 15.9 below)
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

In the sequel, we take R(s) = 0. We note that the

input U(s) in Figure 15.9 satisfies

we can connect this result to state-estimate feedback

and innovations feedback from an observer by using
the results of the previous section. In particular, we
have the next lemma.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Lemma 18.8: The class of all stabilizing linear

controllers can be expressed in state space form as

where K is a state-feedback gain xˆ ( sis) a state

estimate provided by any stable linear observer, and
Ev(s) denotes the corresponding innovation process.

Proof: The result follows immediately upon using

earlier results.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

This alternative form of the class of all stabilizing controllers

is shown in Figure 18.3.

Figure 18.3: State-estimate feedback interpretation

of all stabilizing controllers
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

State-Space Interpretation of
Internal Model Principle
A generalization of the above ideas on state-estimate
feedback is the Internal Model Principle (IMP)
described in Chapter 10. We next explore the state
space form of IMP from two alternative
perspectives.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

(a) Disturbance-estimate feedback

One way that the IMP can be formulated in state space is to
assume that we have a general deterministic input disturbance
d(t) with a generating polynomial Γ d(s).
We then proceed by building an observer so as to generate a
model state estimate and a disturbance estimate,
x
ˆ 0 (t )
These estimates can then be combined in a control law of the
dˆ ( t ).
form

which cancels the estimated input disturbance from the input.

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

We will show below that the above control law

automatically ensures that the polynomial Γ d(s)
appears in the denominator, L(s), of the
corresponding transfer-function form of the
controller.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Consider a composite state description, which

includes the plant-model state

and the disturbance model state:

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

We note that the corresponding 4-tuples that define the

partial models are (A0, B0, C0, 0) and (Ad, 0, Cd, 0) for
the plant and disturbance, respectively. For the
combined state [ x T
0 ( t ) x d ( t ) ] , we have
T T

The plant-model output is given by

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Note that this composite model will, in general, be

observable but not controllable (on account of the
disturbance modes). Thus, we will only attempt to
stabilize the plant modes, by choosing K0 so that
(A0 - B0K0) is a stability matrix.
The observer and state-feedback gains can then be
partitioned as on the next slide.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

When the control law is

used, then, clearly Kd = Cd. We thus obtain
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The final control law is thus seen to correspond to the

following transfer function:

From this, we see that the denominator of the control law

in polynomial form is

We finally see that Γ d(s) is indeed a factor of L(s) as in

the polynomial form of IMP.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

(b) Forcing the Internal Model

Principle via additional dynamics
Another method of satisfying the internal Model
Principle in state space is to filter the system output
by passing it through the disturbance model. To
illustrate this, say that the system is given by
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

We then modify the system by passing the system

output through the following filter:

where observability of (Cd, Ad) implies

controllabililty of ( A T
d , C d ) . We then estimate x(t)
T

using a standard observer, ignoring the disturbance,

leading to
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The final control law is then obtained by feeding

back both xˆ ( t ) and x ′ (t ) to yield

where [K0, Kd] is chosen to stabilize the composite

system.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The results in section 17.9 establish that the cascaded

system is completely controllable, provided that the
original system does not have a zero coinciding with
any eigenvalue of Ad.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The resulting control law is finally seen to have the

following transfer function:

The denominator polynomial is thus seen to be

and we see again that Γ d(s), is a factor of L(s) as

required.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Dealing with Input Constraints in the

Context of State-Estimate Feedback
We give a state space interpretation to the anti-wind-
up schemes presented in Chapter 11.
We remind the reader of the two conditions placed
on an anti-wind-up implementation of a controller,
(i) the states of the controller should be driven by the
actual plant input;
(ii) the state should have a stable realization when
driven by the actual plant input.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The above requirements are easily met in the context

of state-variable feedback. This leads to the anti-
wind-up scheme shown in Figure 18.4.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 18.4: Anti-wind-up Scheme

Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

In the above figure, the state x̂ should also include

estimates of disturbances. Actually, to achieve a
one-degree-of-freedom architecture for reference
injection, then all one need do is subtract the
reference prior to feeding the plant output into the
observer.
We thus see that anti-wind-up has a particularly
simple interpretation in state space.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Summary
❖ We have shown that controller synthesis via pole placement can also be
presented in state space form:
Given a model in state space form, and given desired locations of the
closed-loop poles, it is possible to compute a set of constant gains, one
gain for each state, such that feeding back the states through the gains
results in a closed loop with poles in the prespecified locations.
❖ Viewed as a theoretical result, this insight complements the equivalence
of transfer function and state space models with an equivalence of
achieving pole placement by synthesizing a controller either as transfer
function via the Diophantine equation or as consntant-gain state-
variable feedback.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

❖ Viewed from a practical point of view, implementing this controller

would require sensing the value of each state. Due to physical,
chemical, and economic constraints, however, one hardly ever has
actual measurements of all system states available.
❖ This raises the question of alternatives to actual measurements and
introduces the notion of s-called observers, sometimes also called soft
sensors, virtual sensors, filter, or calculated data.
❖ The purpose of an observer is to infer the value of an unmeasured state
from other states that are correlated with it and that are being
measured.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

❖ Observers have a number of commonalities with control

systems:
◆ they are dynamical systems;
◆ they can be treated in either the frequency or the time domain;
◆ they can be analyzed, synthesized, and designed;
◆ they have performance properties, such as stability, transients, and
sensitivities;
◆ these properties are influenced by the pole-zero patterns of their
sensitivities.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

❖ State estimates produced by an observer are used for several purposes:

◆ constraint monitoring;
◆ data logging and trending;
◆ condition and performance monitoring;
◆ fault detection;
◆ feedback control.

❖ To implement a synthesized state-feedback controller as discussed

above, one can use state-variable estimates from an observer in lieu of
unavailable measurements; the emergent closed-loop behavior is due to
the interaction between the dynamical properties of system, controller,
and observer.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

❖ The interaction is quantified by the third-fundamental result presented in

this chapter: the nominal poles of the overall closed loop are the union of
the observer poles and the closed-loop poles induced by the feedback gains
if all states could be measured. This result is also known as the separation
theorem.
❖ Recall that controller synthesis is concerned with how to compute a
controller that will give the emergent closed loop a particular property, the
constructed property.
❖ The main focus of the chapter is on synthesizing controllers that place the
closed-loop poles in chosen locations; this is a particular constructed
property that allows certain design insights to be achieved.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

❖ There are, however, other useful constructed properties as

well.
❖ Examples of constructed properties for which there exist
synthesis solutions:
◆ to arrive at a specified system state in minimal time with an
energy constraint;
◆ to minimize the weighted square of the control error and energy
consumption;
◆ to achieve minimum variance control.
Chapter 18 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

❖ One approach to synthesis is to ease the constructed property

into a so-called cost-functional, objective function or criterion,
which is then minimized numerically.
◆ This approach is sometimes called optimal control, because one
optimizes a criterion.
◆ One must remember, however, that the result cannot be better than the
criterion.
◆ Optimization shifts the primary engineering task from explicit
controller design to criterion design, which then generates the
controller automatically.
◆ Both approaches have benefits, including personal preference and
experience.