Chapter 3

Controllability, Observability and Stability of System

3.1 Controllability of Control Systems

The concepts of controllability and observability, introduced first by Kalman, play an important role in
both theoretical and practical aspects of modern control. The conditions on controllability and
observability essentially govern the existence of a solution to an optimal control problem. This seems to
be the basic difference between optimal control theory and classical control theory. In the classical control
theory, the design techniques are dominated by trial-and-error methods so that given a set of design
specifications the designer at the outset does not know if any solution exists. Optimal control theory, on
the other hand, has criteria for determining at the outset if the design solution exists for the system
parameters and design objectives.

(a) (b)
Figure 3.1 (a) Control system with state feedback. (b) Control system with observer and state feedback.

We shall show that the condition of controllability of a system is closely related to the existence of solutions
of state feedback for assigning the values of the eigenvalues of the system arbitrarily. The concept of
observability relates to the condition of observing or estimating the state variables from the output
variables, which are generally measurable.

The block diagram shown in Figs 3.1 illustrates the motivation behind investigating controllability
and observability. Figure 3.1 (a) shows a system with the process dynamics described by

𝑑𝑥(𝑡)
= Ax(t) + Bu(t)
𝑑𝑡
The closed-loop system is formed by feeding back the state variables through a constant feedback gain
matrix K. Thus, from Fig. 3.1,

u(t) = — Kx(t) + r(t)

where K is a p x n feedback matrix with constant elements. The closed-loop system is thus described by
 𝑑𝑥(𝑡)
= (𝐀 − 𝐁𝐊)x(t) + Bu(t) (1)
𝑑𝑡

This problem is also known as the pole-placement design through state feedback. The design objective
in this case is to find the feedback matrix K such that the eigenvalues of (A — BK), or of the closed-
loop system, are of certain prescribed values. The word pole refers here to the poles of the closed-loop
transfer function, which are the same as the eigenvalues of (A — BK

Control u(t) Control x(t)

G

Figure 3.2 Linear time invariant system

General Concept of Controllability

The concept of controllability can be stated with reference to the block diagram of Fig. 3.1 (a). The process
is said to be completely controllable if every state variable of the process can be controlled to reach a
certain objective infinite time by some unconstrained control u(t), as shown in Fig. 3.2. Intuitively, it is
simple to understand that, if any one of the state variables is independent of the control u(t), there would
be no way of driving this particular state variable to a desired state in finite time by means of a control
effort. Therefore, this particular state is said to be uncontrollable, and, as long as there is at least one
uncontrollable state, the system is said to be not completely controllable or, simply, uncontrollable.
As a simple example of an uncontrollable system, Fig. 3.3 illustrates the state diagram of a linear system
with two state variables. Because the control U(t) affects only the state Xl(t), the state X2(t) is
uncontrollable. In other words, it would be impossible to drive X2(t) from an initial state X2(t0) to a desired
state X2(tf) in finite time interval tf — to by the control U(t). Therefore, the entire system is said to be
uncontrollable.

Figure 3.3 State diagram of the system that is not state controllable.
The concept of controllability given here refers to the states and is sometimes referred to as state
controllability. Controllability can also be defined for the outputs of the system, so there is a difference
between state controllability and output controllability.
 State Controllability

Consider that a linear time-invariant system is described by the following dynamic equations:
𝑑𝑥(𝑡)
= Ax(t) + Bu(t) (2)
𝑑𝑡
y(t) = Cx(t) + Du(t) (3)
where x(t) is the n x 1 state vector, u(t) is the r x 1 input vector, and y(t) is the p x 1 output vector A, B,
C, and D are coefficients of appropriate dimensions.
The state x(t) is said to be controllable at t = to if there exists a piecewise continuous input u(t) that will
drive the state to any final state a finite time (tf — to) ≥ 0. If every state x(to) of the system is controllable
in a finite time interval, the system is said to be completely state controllable or, simply, controllable.

The following theorem shows that the condition of controllability depends on the coefficient matrices
A and B of the system. The theorem also gives one method of testing for state controllability.

Theorem 3.1: For the system described by the state equation of Eq. (2) to be completely state
controllable, it is necessary and sufficient that the following n x nr controllability matrix has a rank
of n:

𝑺 = [𝑩 𝑨𝑩 𝑨𝟐 𝑩 .. . 𝑨𝒏−𝟏 𝑩] (4)

Because the matrices A and B are involved, sometimes we say that the pair [A, B] is controllable, which
implies that S is of rank n.
Although the criterion of state controllability given in Theorem 3.1 is quite straightforward,
manually, it is not very easy to test for high-order systems and/or systems with many inputs. If S is
non square, we can form the matrix SS', which is n x n; then, if SS' is nonsingular, S has rank n.

Alternate Tests on Controllability

There are several alternate methods of testing controllability, and some of these may be more convenient
to apply than the condition in Eq. (4).

Theorem 3-2. For a single-input, single-output (SISO) system described by the state equation of Eq.
(2) with r = 1, the pair [A, B] is completely controllable if A and B are in CCF or transformable into
CCF by a similarity transformation.
 Theorem 3-3. For a system described by the state equation of Eq. (2), if A is in DCF or JCF, the pair
[A, B] is completely controllable if all the elements in the rows of B that correspond to the last row of
each Jordan block are nonzero.
EXAMPLE Consider that a third-order system has the coefficient matrices

1 2 −1 0
𝐴 = [0 1 0] 𝐵 = [0 ]
1 −4 3 1

The controllability matrix is

𝟎 −𝟏 −𝟒
𝑺 = [𝑩 𝑨𝑩 𝑨𝟐 𝑩 ] = [𝟎 𝟎 𝟎]
𝟏 𝟑 𝟖

which is singular. Thus, the system is not controllable.

Output Controllability

In the practical design of a control system, we may want to control the output rather than the state of the
system. Complete state controllability is neither necessary nor sufficient for controlling the output of the
system. For this reason, it is desirable to define separately complete output controllability.

Consider the system described by

The system described by the above equation is said to be completely output controllable if it is possible
to construct an unconstrained control vector u(t) that will transfer any given initial output to any final
output y(tf) in a finite time interval t0 ≤ t ≤ tf.

The system describe above is completely output controllable if and only if the m x (n+1) r matrix

[CB CAB CA2B … CAn-1B D] is of rank m.

 Stabilizability: For a partially controllable system, if the uncontrollable modes are stable and the
unstable modes are controllable, the system is said to be stabilizable. For example, the system defined
by

is not state controllable. The stable mode that corresponds to the eigenvalue of —1 is not controllable.
The unstable mode that corresponds to the eigenvalue of 1 is controllable. Such a system can be made
stable by the use of a suitable feedback. Thus this system is stabilizable.

3.2 OBSERVABILITY OF LINEAR SYSTEMS

A system is completely observable if every state variable of the system affects some of the
outputs. In other words, it is often desirable to obtain information on the state variables from the
measurements of the outputs and the inputs. If any one of the states cannot be observed from the
measurements of the outputs, the state is said to be unobservable, and the system is not completely
observable or, simply, unobservable. Fig. 3.4 shows the state diagram of a linear system in which the
state x2 is not connected to the output y(t) in any way. Once we have measured y(t), we can observe
the state Xl(t), since Xl(t) = y(t). However, the state X2 cannot be observed from the information on
y(t). Thus, the system is unobservable.

Observability

Given a linear time-invariant system that is described by the dynamic equations of Eq. (2) and (3),
the state x(to) is said to be observable if given any input u(t), there exists a finite time tf ≥to such that
the knowledge of u(t) for to ≤ t < tf, matrices A, B, C, and D; and the output y(t) for to ≤ t < tf are
sufficient to determine x(to). If every state of the system is observable for a finite tf, we say that the
system is completely observable, or, simply, observable.
The following theorem shows that the condition of observability depends on the matrices A and C of
the system. The theorem also gives one method of testing observability.

Figure 3.4 State diagram of a system that is not observable.

 Theorem 3.4: For the system described by Eq. (2) and (3) to be completely observable, it is necessary
and sufficient that the following n x np observability matrix has a rank of n:

(5)

The condition is also referred to as the pair [A, C] being observable. In particular, if the system has only
one output, C is a 1 x n row matrix; V is an n x n square matrix. Then the system is completely
observable if V is nonsingular.

Alternate Tests on Observability

Just as with controllability, there are several alternate methods of testing observability. These are
described in the following theorems.

Theorem 3-5: For an SISO system, described by the dynamic equations of Eq. (2) and (3) with r = 1
and p = 1, the pair [A, C] is completely observable if A and C are in OCF or transformable into OCF
by a similarity transformation,
Theorem 3-6: For a system described by the dynamic equations of Eq. (2) and (3), if Ä is in DCF or
JCF, the pair [A, C] is completely observable if all the elements in the columns of C that correspond
to the first row of each Jordan block are nonzero.
Note that this theorem is a dual of the test of controllability given in Theorem 3-3. If the system has
distinct eigenvalues, A is diagonal, then the condition on observability is that none of the columns of
C can contain all zeros.

EXAMPLE: Consider the system shown in Fig. 3.4, which was earlier defined to be unobservable. The
dynamic equations of the system are expressed in the form of Eq. (2) and (3) with

−2 0 3
𝐴= [ ] 𝐵 = [ ] 𝐶 = [1 0]
0 −1 1

Thus, the observability matrix is

𝐶 1 0
𝑉=[ ]= [ ]
𝐶𝐴 −2 0
 which is singular. Thus, the pair [A, C] is unobservable. In fact, because A is of DCF and the second
column of C is zero, this means that the state X2(t) is unobservable, as conjectured from Fig. 3.3.

3.3 RELATIONSHIP AMONG CONTROLLABILITY, OBSERVABILITY, AND TRANSFER

FUNCTIONS
In the classical analysis of control systems, transfer functions are used for modeling of linear time-
invariant systems. Although controllability and observability are concepts and tools of modern
control theory, we shall show that they are closely related to the properties of transfer functions.

Theorem 3-7: If the input-output transfer function of a linear system has pole-zero cancellation, the
system will be uncontrollable or unobservable, or both, depending on how the state variables are
defined. On the other hand, if the input-output transfer function does not have pole-zero cancellation,
the system can always be represented by dynamic equations as a completely controllable and
observable system.

3.4 INVARIANT THEOREMS ON CONTROLLABILITY AND OBSERVABILITY

We now investigate the effects of the similarity transformations on controllability and observability.
The effects of controllability and observability due to state feedback will be investigated.

Theorem 3-8: Invariant theorem on similarity transformations: Consider the system described by the
dynamic equations of Eq. (2) and (3). The similarity transformation x(t) = P𝑥̅ (t), where P is nonsingular,
transforms the dynamic equations to

𝑑𝑥̅ (𝑡)
= 𝐴̅𝑥̅ (𝑡) + 𝐵̅𝑢(𝑡) (6)
𝑑𝑡

̅ 𝑢(𝑡)
𝑦(𝑡) = 𝑐̅𝑥(𝑡) + 𝐷

𝐴̅ = 𝑃−1 𝐴𝑃 𝐵̅ = 𝑃 −1 𝐵

The controllability of [𝐴̅, 𝐵̅]and the observability of [𝐴̅, 𝐶̅ ] are not affected by the transformation.
In other words, controllability and observability are preserved through similar transformations.
The theorem is easily proven by showing that the ranks of 𝑆̅ and S and the ranks of 𝑉̅ and V are
identical, where 𝑆̅ and ̅𝑉 are the controllability and observability matrices, respectively, of the
transformed system.
 Theorem 3-9: Theorem on controllability of closed-loop systems with state feedback: If the open-loop
system
𝑑𝑥(𝑡)
= Ax(t) + Bu(t)
𝑑𝑡
is completely controllable, then the closed-loop system obtained through state feedback,

𝑢(𝑡) = r(t) − Kx(t) (7)

so that the state equation becomes

𝑑𝑥(𝑡)
= (A − BK)x(t) + Bu(t) (8)
𝑑𝑡
is also completely controllable. On the other hand, if [A, B] is uncontrollable, then there is no K that
will make the pair [A — BK, B] controllable. In other words, if an open-loop system is uncontrollable,
it cannot be made controllable through state feedback.

Theorem 3-10: Theorem on observability of closed-loop systems with state feedback: If an open-loop
system is controllable and observable, then state feedback of the form of Eq. (7) could destroy
observability. In other words, the observability of open loop and closed-loop systems due to state
feedback is unrelated.
The following example illustrates the relation between observability and state feedback.

EXAMPLE 10-14-1 Let the coefficient matrices of a linear system be

0 1 1
𝐴=[ ] 𝐵 = [ ] 𝐶 = [1 2]
−2 −3 1

We can show that the pair [A, B] is controllable and [A, C] is observable.
Let the state feedback be defined as

u(t) = r(t) – K x(t)

where

K= [k1 k2]

Then the closed-loop system is described by the state equation

𝑑𝑥(𝑡)
= (A − BK)x(t) + Bu(t)
𝑑𝑡
 −𝑘1 1 − 𝑘2
𝐴 − BK = [ ]
−2 − 𝑘1 −3 − 𝑔2

The observability matrix of the closed-loop system is

𝐶 1 4
𝑉=[ ]= [ ]
𝐶(𝐴 − 𝐵𝐾) −𝑘1 − 4 −3𝑘2 − 5
The determinant of V is |V|= 6kl - 3k2 +3

Thus, if kl and k2 are chosen so that IVI = 0, the closed-loop system would be uncontrollable.

Detectability: For a partially observable system, if the unobservable modes are stable and the observable
modes are unstable, the system is said to be detectable. Note that the concept of detectability is dual to
the concept of stabilizability.

3.5 Stability

From the studies of linear differential equations with constant coefficients of SISO systems, we learned
that the homogeneous solution that corresponds to the response of the system is governed by the roots of
the characteristic equation. Basically, the design of linear control systems may be regarded as a problem
of arranging the location of the poles and zeros of the system transfer function such that the system will
perform according to the prescribed specifications.
Among the many forms of performance specifications used in design, the most important requirement is
that the system must be stable. An unstable system is generally considered to be useless.
For nonlinear and time-varying systems, the study of stability is a complex and often difficult subject. In
this section, we will consider only LTI systems for which we have the following condition for stability:

An LTI system is said to be stable if all the roots of the transfer function denominator polynomial have
negative real parts (i.e., they are all in the left hand s-plane) and is unstable otherwise.

The transfer function is given by

𝐻(𝑠) = 𝐶(𝑠𝐼 − 𝐴)−1 𝐵 + 𝐷

𝐶[𝑎𝑑𝑗(𝑠𝐼 − 𝐴)]𝐵 𝐶[𝑎𝑑𝑗(𝑠𝐼 − 𝐴)]𝐵 + 𝐷|𝑠𝐼 − 𝐴| 𝑁𝑐 (𝑠)

= +𝐷 = =
|𝑠𝐼 − 𝐴| |𝑠𝐼 − 𝐴| ∆𝑐 (𝑠)

Where subscript ‘c’ denotes variables after pole/zero cancellation.

 The denominator of this is the characteristic polynomial
Δ(𝑠) = |𝑠𝐼 − 𝐴|
The system poles are the roots of the characteristic equation
Δ(𝑠) = |𝑠𝐼 − 𝐴| = 0
The transfer function poles are the roots of
∆𝑐 (s) = 0

Stable system
A system is stable if its initial conditions decay to zero and is unstable if they diverge. As just stated, an
LTI (constant parameter) system is stable if all the poles of the system are strictly inside the left half s-
plane (i.e., all the poles have negative real parts (s = – σ + jω, σ > 0)). If any pole of the system is in the
right half s-plane (i.e., has a positive real part, s = – σ + jω, σ < 0), then the system is unstable, as shown
in Fig. 3 With any simple pole on the jω axis (σ = 0), small initial conditions will persist. For any other
pole with σ = 0, oscillatory motion will persist. Therefore, a system is stable if its transient response decays
and unstable if it does not. Figure 3.15 shows the time response of a system due to its pole locations.

Figure 3: Time function associated with point in s-plane

Bounded Input-Bounded Output Stability

A system is said to have bounded input-bounded output (BIBO) stability if every bounded input results in
a bounded output (regardless of what goes on inside the system). A test for this property is readily available
when the system response is given by convolution. If the system has input u(t), output y(t), and impulse
response h(t), then
 If u(t) is bounded, then there is a constant M such that |u| ≤ M < ∞, and the output is bounded by

Thus the output will be bounded if is bounded

The system with impulse response h(t) is BIBO-stable if and only if the integral

Zero-Input and Asymptotic Stability of a Systems

In this section, we shall define zero-input stability and asymptotic stability and establish their relations
with BIBO stability.
Zero-input stability refers to the stability condition when the input is zero, and the system is driven only
by its initial conditions. We shall show that the zero-input stability also depends on the roots of the
characteristic equation.
If the zero-input response y(t), subject to the finite initial conditions, reaches zero as t approaches
infinity, the system is said to be zero-input stable, or stable; otherwise, the system is unstable.
𝐥𝐢𝐦 |𝒚(𝒕) = 𝟎
𝒕→∞
Because the condition in the last equation requires that the magnitude of y(t) reaches zero as time
approaches infinity, the zero-input stability is also known at the asymptotic stability.
From the preceding discussions, we see that, for linear time-invariant systems, BIBO, zero-input,
and asymptotic stability all have the same requirement that the roots of the characteristic equation
must all be located in the left-half s-plane. Thus, if a system is BIBO stable, it must also be zero-
input or asymptotically stable. For this reason, we shall simply refer to the stability condition of a
linear system as stable or unstable.
Example: The following closed-loop transfer functions and their associated stability conditions are given.