© Copyright F.L.

Lewis 2001
All rights reserved
Updated: Saturday, October 04, 2008

REACHABILITY AND OBSERVABILITY

Consider a linear state-space system given by

x = Ax + Bu
y = Cx + Du
with x(t ) ∈ R n the internal state, u (t ) ∈ R m the control input, and y (t ) ∈ R p the measured
output. The transfer function is given by

H ( s ) = C ( sI − A) −1 B + D .

The input-decoupling zeros are those values of s for which the n × (n + m) input-coupling
matrix
PI ( s ) = [sI − A B ]
loses rank, i.e. has rank less than n. Note that this matrix can lose rank only where ( sI − A)
loses rank, so the input-decoupling zeros must be a subset of the poles.

System (A,B) has no input decoupling zeros if and only if

rank[( sI − A) −1 B ] = n
over the complex numbers. That is, there exists no nonzero n-vector w such that
wT ( sI − A) −1 B = 0 for all s

Note that this is the right-hand portion of the transfer function. If there are input-decoupling
zeros, control effectiveness of the system is lost and we cannot fully control the system with the
given inputs. We should design systems with no input-coupling zeros, i.e. with a fully effective
set of inputs.

The output-decoupling zeros are those values of s for which the (n + p ) × n output-
coupling matrix
⎡ sI − A⎤
PO ( s ) = ⎢ ⎥
⎣ −C ⎦
loses rank, i.e. has rank less than n. Note that this matrix can lose rank only where ( sI − A)
loses rank, so the output-decoupling zeros must be a subset of the poles.

System (A,C) has no output decoupling zeros if and only if

rank[C ( sI − A) −1 ] = n
over the complex numbers. That is, there exists no nonzero n-vector v such that
 C ( sI − A) −1 v = 0 for all s

Note that this is the left-hand portion of the transfer function. If there are output-decoupling
zeros, output measurement effectiveness of the system is lost and we cannot observe the full
internal state behavior with the given outputs. We should design systems with no output-
coupling zeros.

Reachability

For either continuous-time systems or discrete-time systems, the system ( A, B, C ) is

called reachable if the control input can be selected to drive any initial state to any desired final
state at some final time. This can be done if the input coupling in the system is sufficiently
strong, which depends on the input-coupling matrix pair (A,B). Reachability greatly facilitates
control systems design. If a system is not reachable, it can be made so by adding additional
control inputs.

We can examine the discrete-time expanded state equation to determine reachability

conditions for DT systems. It is seen that the DT system is reachable if and only if the
reachability matrix

U= B [ AB A 2 B " A n−1 B . ]
has full rank n. Note that U is an n × nm matrix so that it has more columns than rows if m>1.
Such a matrix is called a flat matrix. (A matrix with more rows than columns is called a sharp
matrix.).

In fact, this rank condition on U is also necessary and sufficient for reachability of
continuous-time systems.

Reachability is equivalent to the absence of input-decoupling zeros. To understand this

connection, note that one can write

( sI − A) −1 B = Bs −1 + ABs −2 + A 2 Bs −3 + ...

To avoid investigating all powers of A, one may use the Cayley Hamilton Theorem. This
theorem states that
Δ( A) = 0 ,

that is, a matrix satisfies its own characteristic equation. If the characteristic equation is
Δ( s ) = s n + a1 s n −1 + ... + a n ,
then replace all occurrences of s by A to obtain (note the last term is a n s 0 )
Δ( A) = A n + a1 A n −1 + ... + a n I .
This is a matrix polynomial. The Cayley-Hamilton Theorem says that
 A n = − a1 A n −1 − ... − a n I ,
which states that A n can be expressed as a linear combination of lower powers of A.

Therefore, in the infinite series expansion above, one may stop at A n −1 Bs − n . Write
⎡ s −1 ⎤
⎢ −2 ⎥
2 −3 n −1 −n
[
Bs + ABs + A Bs + ... + A Bs = B AB A B ... A B ⎢ ⎥
−1 −2 2 n −1
] s
⎢ # ⎥
⎢ −n ⎥
⎣⎢ s ⎦⎥
which provides a direct relation between the input-coupling term and the reachability matrix U.
Now, it can be shown that the system is reachable iff there are no input-decoupling zeros.

The reachability matrix is an n × nm matrix. If there is only one control input (the single-
input (SI) case, where m=1), then U is square. In this case, it is easy to test whether U has rank n
by making sure the determinant U is nonzero. If m>1 one must find n linearly independent
columns of U, which may be difficult particularly if the number of inputs m is large. In this case,
define the reachability gramian
G = UU T
which is a square n × n matrix. Then the system is reachable iff G ≠ 0 .

Many design techniques (e.g. root locus) rely on trying to determine closed-loop
properties from open-loop properties. This is exactly the intent of the reachability test, which
allows one to determine in terms of the open-loop matrices A and B what can be accomplished in
the closed-loop system.

The following conditions for reachability are all equivalent and apply for CT and DT
systems:

1. (A,B) is reachable

2. U= B [ AB ]
A 2 B " A n−1 B has full rank n, i.e. has n linearly independent rows.

3. The rows of ( sI − A) −1 B are linearly independent over the complex numbers.

4. The rows of e At B are linearly independent over t ∈ [0, ∞)

5. PI ( s ) = [sI − A B ] has full row rank n over the complex numbers s.

The equivalence of 2 and 3 depends on the relation

( sI − A) −1 B = Bs −1 + ABs −2 + A 2 Bs −3 + ...

To understand the relation between 2 and 4 consider the following [p. 167, C.T. Chen,
Introduction to Linear System Theory, 1970]. A time-varying matrix F(t) which has continuous
derivatives up through order (n-1) has linearly independent rows over t ∈ [0, ∞) if and only if the
derivative matrix
⎡⎣ F (t ) F (1) (t ) " F ( n −1) (t ) ⎤⎦
has rank n for some value of t. Here the i-th derivative is denoted F ( i ) (t ) .
Set F (t ) = e At and compute the derivative matrix
⎡⎣e At B Ae At B A2 e At B " An −1e At B ⎤⎦
Now set t=0 to get the reachability matrix.

Consider condition 5, which is equivalent to the existence of a pole s0 (the input-

decoupling zero) and a nonzero n-vector w such that
wT PI ( s0 ) = wT [ s0 I − A B ] = 0
i.e.
wT ( s0 I − A) = 0
wT B = 0
The first equation says that
wT A = s0 wT
Therefore
wT B = 0
wT AB = s0 wT B = 0
wT A2 B = s0 wT AB = s0 wT B = 0
#
which is equivalent to
wT ⎡⎣ B AB A2 B " An −1 B ⎤⎦ = 0 .

Observability
For either continuous-time systems or discrete-time systems, let us define the system
( A, B, C ) to be observable if the state can be reconstructed uniquely given measurements of the
output over a time interval [0, T ] . This can be done if the output coupling in the system is
sufficiently strong, which depends on the output-coupling matrix pair (A,C). If a system is not
observable, it can be made so by adding additional measurements. It turns out that observability
means we can design a stable observer to reconstruct the internal states given the available
measurements. This is important in communications theory, navigation, and elsewhere.

Observability is equivalent to the absence of output-decoupling zeros. To find a test for

observability, note that
C ( sI − A) −1 = Cs −1 + CAs −2 + CA 2 s −3 ...
According to the Cayley-Hamilton Theorem, A n is dependent on lower powers of A. Therefore,
note that
 ⎡ C ⎤
⎢ CA ⎥
⎢ ⎥
[
Cs −1 + CAs − 2 + CA 2 s −3 + ... + CA n −1 s − n = s −1 s −2 ... s − n ] ⎢ CA 2 ⎥ .
⎢ # ⎥
⎢ n −1 ⎥
⎣CA ⎦
We call the matrix
⎡ C ⎤
⎢ CA ⎥
⎢ ⎥
V = ⎢ CA 2 ⎥
⎢ # ⎥
⎢ n−1 ⎥
⎣CA ⎦
the observability matrix. We surmise that that ( A, C ) is observable if and only if matrix V has
full rank n . This can be proved, e.g. for DT systems from the expanded state equation.

The observability matrix V has np rows and n columns, so it is called a sharp matrix if
p > 1 , for then it has more rows than columns. (Recall that U is a flat matrix.) If the number of
outputs p is one, then V is square. Otherwise, it might be quite difficult to determine if V has
n linearly independent rows. Define the observability gramian
Go = V T V
which is a square n × n matrix. This matrix has the same rank as V , but it is easy to determine if
it has full rank by simply computing its determinant.

The following conditions for observability are all equivalent and apply for CT and DT
systems:

1. (A,C) is observable

⎡ C ⎤
⎢ CA ⎥
⎢ ⎥
2. V = ⎢ CA 2 ⎥ has full rank n, i.e. has n linearly independent columns.
⎢ # ⎥
⎢ n−1 ⎥
⎣CA ⎦
3. The columns of C ( sI − A) −1 are linearly independent over the complex numbers.

4. The columns of Ce At are linearly independent over t ∈ [0, ∞)

⎡ sI − A⎤
5. PO ( s ) = ⎢ ⎥ has full column rank n over the complex numbers.
⎣ −C ⎦
Duality
Given a plant ( A, B, C ) , the plant ( AT , C T , B T ) is known as the dual system. In this
system, the effects of the inputs and outputs are effectively interchanged. We shall see that
duality provides a relation between many concepts, including, e.g., the reachable canonical form
and the observable canonical form block diagrams.

We can find a connection between reachability and observability using duality. To this
end, dualize the reachability matrix by writing
⎡ BT ⎤ ⎡ BT ⎤
⎢ ⎥ ⎢ ⎥
⎢ ( AB)T ⎥ ⎢ B T AT ⎥
[ ]
U T = B AB A 2 B " A n−1 B T = ⎢ ( A 2 B)T ⎥ = ⎢ B T ( AT ) 2 ⎥ .
⎢ # ⎥ ⎢ # ⎥
⎢ n−1 T ⎥ ⎢ T T n−1 ⎥
⎣⎢( A B) ⎦⎥ ⎣⎢ B ( A ) ⎦⎥
Now, replace ( A, B ) by the dual system ( AT , C T ) , which yields the observability matrix
⎡ C ⎤
⎢ CA ⎥
⎢ ⎥
V = ⎢ CA 2 ⎥ .
⎢ # ⎥
⎢ n−1 ⎥
⎣CA ⎦

Clearly, matrix U , with ( A, B) replaced by ( AT , C T ) , has full rank if and only matrix V ,
which is based on ( A, C ) has full rank. Indeed, it can be shown that this is the case. That is, the
system is reachable iff the dual system is observable, and vice versa.