MASSACHUSETTS INSTITUTE OF TECHNOLOGY

DEPARTMENT OF MECHANICAL ENGINEERING

2.151 Advanced System Dynamics and Control

Controllability, Observability and the Transfer Function1

In this brief note we examine some additional conclusions on system controllability and
observability based on the transfer function matrix.
Consider a linear system of order n with r inputs and m outputs:
ẋ = Ax + Bu
y = Cx + Du
with distinct eigenvalues. By taking the Laplace transform of the state-equations, we obtain
X(s) = [sI − A]−1 BU(s) = H1 (s)U(s) (1)
where H1 (s) = [sI − A]−1 B is a ( n × r) transfer function matrix relating the state-variable
responses to the system inputs. Each row is the Laplace domain relationship between a
state-variable Xi (s) and the input vector U(s).
Controllability is determined by the ability to manipulate the system state x from an
initial state to any arbitrary value in finite time, through a suitable (but undefined) choice
of the input U(s). We therefore can make the following additional conclusions concerning
system controllability from H1 (s).

A system is uncontrollable if any of the following conditions are met:

• Any row of the matrix [sI − A]−1 B is 0. If a zero row exists, it implies that a
state is unaffected by any component of the input vector U(s).

• There exists a linear dependence between the rows of [sI − A]−1 B. This condi-
tion implies that the forced responses of two or more state-variables are linearly
related, and therefore cannot be independently manipulated by the input vector
U(t).
At
• There exists a linear dependence between
n the rows of
o the matrix e B. This
At −1 −1
follows directly because e B = L [sI − A] B , and since the Laplace
transform is a one-to-one operator the linear dependence in one domain trans-
lates directly to the other.

• A necessary and sufficient condition for controllability is that no single pole of

the system is cancelled by a zero in all of the elements of the transfer-function
matrix [sI − A]−1 B. If such cancellation occurs, the system cannot be con-
trolled in the direction of the cancelled mode.

• We state without proof that a system is uncontrollable if for any eigenvalue

λi , i = 1 . . . n, of A, the rank of the (n × (n + r)) matrix [(λi I − A) | B] < n.
(The proof of this is beyond the scope of this handout.)
1
D. Rowell 11/15/04

1
Observability defines the ability to determine the initial state of a system from finite obser-
vation of the output. Because the effect of the input may be subtracted out, it is sufficient to
consider the homogeneous response to determine observability. The response of the system
from initial condition x(0) is

y = CeAt x(0)
 
x1 (0)
 
 x2 (0) 
= [Q1 (t) | Q2 (t) | . . . | Qn (t)] 
 .. 
 (2)
 . 
xn (0)

where Qi (t) is the ith column of CeAt . It is easy to show that the system is unobservable if the
columns of CeAt , that is Qi (t), are linearly dependent. For example, suppose Q2 (t) = kQ1 (t),
then the output equation becomes

y(t) = Q1 (t)x1 (0) + kQ1 (t)x2 (0) + Q3 (t)x3 (0) + . . . + Qn (t)xn (0)
= Q1 (t)(x1 (0) + kx2 (0)) + Q3 (t)x3 (0) + . . . + Qn (t)xn (0)

where it can be seen that the two initial condition components x1 (0) and x2 (0) are bound
together as a weighted sum in all output equations, and therefore cannot be estimated from
measurements on y. We therefore can make the following additional conclusions concerning
system observability:

A system is unobservable if any of the following conditions are met:

• Any column of the matrix CeAt is 0. If a zero column exists, it implies that a
state does not appear in any of the output equations, and is thus unobservable.
The equivalent Laplace domain statement is that a system is unobservable if
any column of the matrix C [sI − A]−1 is 0

• There exists a linear dependence between the columns of the matrix CeAt . This
is demonstrated above in Eq. (2). The equivalent Laplace domain condition is
that a system is unobservable if there exists a linear dependence between the
columns of C [sI − A]−1 .

• A necessary and sufficient condition for observability is that no single pole of the
system is cancelled by a zero in all of the elements of the matrix C [sI − A]−1 . If
such cancellation occurs, the cancelled mode cannot be observed in the output.

• A system is unobservable if for any eigenvalue λi , i = 1 . . . n, of A the rank of

the (n + m × n)) matrix  
λi I − A
 
 ···  < n.
C
The proof of this is beyond the scope of this handout.

2
Pole/Zero Cancellation and Controllability/Observability: The previous discussion
implies that a system in which any pole is cancelled in all of the elements of the matrix
C [sI − A]−1 B will be unobservable. A system will be uncontrollable if a pole is cancelled
in all of elements the transfer function matrix H1 (s) = [sI − A]−1 B that relates the state
vector X(s) to the input U(s).

The Effect of Redundant State-Variables: While nobody would intentionally use

more state-variables than is necessary to describe the dynamics of a physical system, most
modeling methods will occasionally generate linearly dependent state equations. Consider
the system with a minimum set of state equations
ẋ = Ax + Bu
and suppose, for some reason that k more state-variables z, proportional to those present in
x, have also been defined. The linear dependence is described by
z = Fx
where F is (k × n). Then
ż = Fẋ = F(Ax + Bu)
The system model, including the superfluous variables, is then
x̂˙ = Âx̂ + B̂u
where x̂ = [x | z]T , and " # " #
A 0 B
 = B̂ =
FA 0 FB
Now consider a transformation of variables given by x̂ = Px̄ where the (n + k) × (n + k)
transformation matrix P is
" # " #
In 0 −1 In 0
P= and P =
F Ik −F Ik
where In and Ik are the (n × n) and (k × k) identity matrices, so that
" #" #" # " #
−1 In 0 A 0 In 0 A 0
Ā = P ÂP = =
−F Ik FA 0 F Ik 0 0
and " #" # " #
−1 In 0 B B
B̄ = P B̂ = =
−F Ik FB 0
Then in the transformed states
x̄˙ = Āx̄ + B̄u
and if we write x̄ = [x0 | z0 ]T , the transformed state equations are
ẋ0 = Ax0 + Bu
ż0 = 0
so that the transformation P has reduced the k excess state equations to a vector z0 containing
k integrators with no inputs driving them - these excess states cannot be controlled, and the
system is therefore uncontrollable.