Controllability and Observability Part II

Chong-Jin Ong

Department of Mechanical Engineering,

National University of Singapore

1 / 23
Outline

1 Observability

2 Conditions for Observability

3 Uncontrollable and Unobservable Subsystem

4 Gilbert-Kalman Canonical Decomposition

2 / 23
Observability

Observability is a closely related concept to Controllability.

Controllability deals with the ability to steer the states using the input.
Observability studies the ability to estimate the states from the output.
Loosely, observability refers to the ability to “see” the effects of all the
states from the output.

3 / 23
Why is observability important?
The design of many controllers requires the information of the states.
However, some of the state variables are not accessible for direct measurement
⇒ Estimation of unmeasurable states is needed.
Estimation of states is possible if and only if the system observable.
If u(t) = 0 for all t ≥ 0 and x1 (0) = a 6= 0 for some unknown a, and x2 (0) = 0,
⇒ y(t) is identically zero for all time.

4 / 23
Definition

While a general definition of observability exists, the one used here is for LTI
system.
Definition: A linear time-invariant system is observable if every unknown
initial state x(0) can be determined from the knowledge of u(t) and the
observation of y(t) over a finite time interval. Otherwise, the LTI system is
said to be unobservable.
For a system to be observable, all initial states x(0) can be determined using a
finite time interval. If only a subset of the initial states can be determined or
that it takes infinite amount of time to determine all initial states, the system
is not observable.

5 / 23
The study of observability is concerned with the unforced system:

ẋ = Ax, y = Cx

as a result of the assumption of u(t) being completely known.

The time solution of the complete state-space system is
Z t
At
x(t) = e x(0) + eA(t−τ ) Bu(τ )dτ, and,
0
Z t
y(t) = CeAt x(0) + C eA(t−τ ) Bu(τ )dτ + Du(t)
0

Under complete knowledge of {A, B, C, D} and u(t), the overall system can be
restated as those above.
We first provide one condition for observability.
The others are based on a duality result.

6 / 23
Observability
Theorem 4.2 The n-dimensional LTI system or the pair (A, C) is observable if and
only if the n × n matrix (the observability grammian)
Z t
T
M (0, t) := eA τ C T CeAτ dτ
0

is non-singular for any t > 0.

Proof: (⇒) Suppose M (0, t) is non-singular for all t > 0. We have

y(t) = CeAt x(0)

T T
⇒ eA t C T y(t) = eA t C T CeAt x(0), or,
Z t Z t
T T
⇒ eA τ C T y(τ )dτ = eA τ C T CeAτ dτ x(0) = M (0, t)x(0)
0 0
This implies that Z t T
−1
x(0) = M (0, t) eA τ
C T y(τ )
0
Thus, system is observable.

7 / 23
Observability

Proof: (⇐) Suppose M (0, t) is singular for some t > 0 but the system is observable.
We want to show that this leads to a contradiction. M (0, t) is singular means that
there exists a non-zero vector α s.t.

αT M (0, t)α = 0
Z t T
⇒ αT eA t C T CeAt αdt = 0
0
Z t
⇒ kCeAt αk2 dt = 0
0

⇒ CeAt α = 0 for all t

This means that the output

y(t) = CeAt x(0) = 0

if x(0) = α. Hence, y(t) is identically zero and the system is not observable, which
leads to a contradiction. 

8 / 23
Duality
(A, B) is controllable if and only if (AT , B T ) is observable.
Rt T
(A, B) controllable ⇒ W (0, t) = 0 eAτ BB T eA τ dτ is non-singular.
Rt T
(AT , B T ) observable ⇒ M (0, t) = 0 eAτ BB T eA τ dτ is non-singular.
Using the same argument, we have
(A, C) is observable if and only if (AT , C T ) is controllable.

A, B, C, D AT,CT,BT,D

Original System Dual System

9 / 23
Conditions
The following statements are equivalent.
1 The pair (A, C) is observable.
Rt T
2 The observability Grammian M (0, t) := 0
eA τ
C T CeAτ dτ is non-singular for
all t > 0.
3 The mn × n observability matrix
 
C
 CA 
O=
 
.. 
 . 
CAn−1

is full column rank.

4 The (n + m) × n matrix  
A − λI
C
has full column rank at every eigenvalue of A.
Proof: Either directly or by invoking duality result and then followed by the
corresponding proofs given in the controllability results.
10 / 23
Observability under Similarity Transformation

Like the case of controllability, the observability matrix, under a new basis
representation is
   
C̄ CT
 C̄ Ā   CT T −1 AT 
Ō =  =  = OT
   
.. ..
 .   . 
C̄ Ān−1 CAn−1 T

Hence, observability is not affected by similarity transformation.

11 / 23
Observability Grammian

The Observability Grammian

Z ` T
M (0, `) := eA τ
C T CeAτ dτ
0

has to be full rank for all ` > 0.

Like the Controllability case, the full rank condition of M (0, `) can be shown
to be independent of `.
WLOG, consider the case where ` → ∞ and
Z ∞
T
M := lim M (0, `) = eA τ C T CeAτ dτ
`→∞ 0

12 / 23
Computation of Observability Grammian

R∞ T
The expression of M = 0
eA τ
C T CeAτ dτ is hard to compute.
Like the controllability case, there is an expression that allow easy
computation of M . In particular, M satisfies the Lyapunov Equation

AT M + M A = −C T C

such that the solution of the above yields M .

This follows because (assuming A is asymptotically stable),
Z ∞
d AT t T At
AT M + M A = (e C Ce )dt = −C T C
0 dt

13 / 23
Gilbert Decomposition

Question: What can be done with an uncontrollable (unobservable) system?

Theorem 4.3: Given a LTI system S. If the controllability matrix of S has
rank n1 (n1 < n), then there exists an equivalent transformation which
transforms the system into S̄:

x̄˙ c
     
Āc Ā12 x̄c B̄c
= + u
x̄˙ c̄ 0 Āc̄ x̄c̄ 0
 

x̄c
y = C̄c C̄c̄ + D̄u
x̄c̄
and the n1 -dimensional subsystem S̄c of S̄

x̄˙ c = Āc x̄c + B̄c u

y = C̄c x̄c + D̄u

is controllable and has the same transfer function matrix as S.

14 / 23
Proof
Proof: If S is not controllable then

rankU = rank[B AB · · · An−1 B] = n1 < n

Let p1 , · · · , pn1 be any linearly independent vectors of U . Note that

Api , i = 1, · · · , n1 can be written as a l.c. of p1 , · · · , pn1 . Hence, choose P for
x̄ = P x as
P −1 = [p1 , · · · , pn1 , pn1 +1 , · · · , pn ]
where the last (n − n1 ) columns are arbitrary so long as P −1 is non-singular. Using
this, S̄ becomes
Ā = P AP −1 , B̄ = P B, C̄ = CP −1 , D̄ + D
For columns of B, note that B is in the range space of U and can be expressed as a
l.c. of p1 , · · · , pn1 , or
 
−1 B̄c
B = P B̄ = [p1 , · · · , pn1 , pn1 +1 , · · · , pn ]
0
Similarly,
 
Āc Ā12
AP −1 = A[p1 , · · · , pn1 , pn1 +1 , · · · , pn ] = [p1 , · · · , pn1 , pn1 +1 , · · · , pn ]
0 Āc̄

Hence, S̄ has the said structure.

15 / 23
Proof

To see that S̄c is controllable, note that

n1 = rankU = rankŪ = rank[B̄ ĀB̄ · · · Ān−1 B̄]

B̄c Āc B̄c · · · Ān−1
 
B̄c
= rank c
= rank[B̄c Āc B̄c · · · Ācn−1 B̄c ]
0 0 ··· 0

To see that T.F. of S and S̄c are the same, note that T.F. of S̄ is
 −1  

sI − Āc −Ā12 B̄c
C̄c C̄c̄ + D̄
0 sI − Āc̄ 0

(sI − Āc )−1 (sI − Āc )−1 Ā12 (sI − Āc̄ )−1
  
B̄c
= C̄c C̄c̄ −1 + D̄
0 (sI − Āc̄ ) 0
= C̄c (sI − Āc )−1 B̄c + D̄ = T. F. of S̄c

16 / 23
Note that the system of S̄ can be considered as two subsystems of the form

x̄˙ c = Āc x̄c + Ā12 x̄c̄ + B̄c u

x̄˙ c̄ = Āc̄ x̄c̄

where the dynamics of x̄c is affected by x̄c̄ and u, while the dynamics of x̄c̄ is
unaffected by x̄c and u. Hence,
x̄˙ c̄ = Āc̄ x̄c̄
is an uncontrollable subsystem.

17 / 23
Example
   
1 1 0 0 1
ẋ =  0 1 0 x +  1 0 u
0 1 1 0 1


y= 1 1 1 x
 
0 1 1 1 2 1
rankU = rank  1 0 1 0 1 0 =2<3
0 1 1 1 2 1
 
0 1 1
−1
Choose P = 1 0 0  where the last column is arbitrary so long as P is
0 1 0
non-singular. Then
     
0 1
0 1 1 0 0 1 1 1 0 0
Ā = P AP −1 = 0 0
1  0 1 0  1 0 0 = 1 1 0 
1 −1
0 0 1 1 0 1 0 0 0 1
    
0 1 0 0 1 1 0
= CP −1 =


B̄ = P B =  0 0 1  1 0  =  0 1  , C̄ 1 2 1
1 0 −1 0 1 0 0

18 / 23
An equivalent result exists for an unobservable system (not surprising in view of
duality),
Theorem 4.4: Given a LTI system S. If the observability matrix has rank
n2 (n2 < n), then there exists an equivalent transformation which transforms S
into S̄: 
x̄˙ o
     
Āo 0 x̄o B̄o
= + u
x̄˙ ō Ā12 Āō x̄ō B̄ō
 

x̄o
y = C̄o 0 + D̄u
x̄ō
and the n2 -dimensional subsystem S̄o of S̄

x̄˙ o = Āo x̄o + B̄o u

y = C̄o x̄o + D̄u

is observable and has the same transfer function matrix as S.

19 / 23
Proof

Proof: Since rankO = n2 < n. Choose (n − n2 ) l.i. vectors in the null space of the
observability matrix. Let qn2 +1 , · · · , qn be such vectors and choose non-singular Q as

Q = [q1 · · · qn2 · · · qn ]

Consider the state transformation x̄ = Q−1 x, we have

Ā = Q−1 AQ, B̄ = Q−1 B, C̄ = CQ, D̄ = D

Again, the structure of C̄ is

CQ = [C̄o 0]
Notice also that Aqn2 +1 , · · · , Aqn are in the null space of O. Hence,
 
Āo 0
A[q1 · · · qn2 · · · qn ] = Q .
Ā12 Āō

(Exercise) Note that (Āo , C̄o ) is observable and T.F. of S = T.F. of S̄o .

20 / 23
Example

   
1 2 1

A= , B= , C= −1 1
0 3 1
   
C −1 1
Note that O = = . Choose basis for null space as [1 1]T . Then
CA −1 1
   

1 1 −1 1 −1
Q= q1 q2 = and Q =
0 1 0 1

Using this Q, the transformed system becomes


x̄˙ o
     
1 0 x̄o 0
= + u, y = [−1 0]x̄
x̄˙ ō 0 3 x̄ō 1

21 / 23
Gilbert-Kalman Canonical Decomposition

Theorem 4.5 Any LTI system can be converted into the following form using an
appropriate similarity transformation.

x̄˙ cō
     
Ācō Ā12 Ā13 Ā14 B̄cō
 x̄˙ co   0 Āco 0 Ā24  +  B̄co  u
 
 x̄˙ c̄ō  =  0
  
0 Āc̄ō Ā34   0 
x̄˙ c̄o 0 0 0 Āc̄o 0


y = 0 C̄co 0 C̄c̄o + D̄u

The transfer matrix of the system is given by

G(S) = C̄co (sI − Āco )−1 B̄co + D̄

Proof: Combining proofs of Theorems 4.3 and 4.4.

22 / 23
Minimality, Controllability and Observability

From the Gilbert-Kalman decomposition, it is clear that if a LTI system is

either uncontrollable or unobservable, there exists a dynamical system of lesser
dimension that has the same transfer function matrix as the original dynamical
system.
As a result, the following definition is now given.
Definition: A LTI dynamical system, S, is said to be minimal when there is no
other LTI system of lesser dimension that has the same transfer function
matrix as S.

23 / 23