Controllability and Observability

Chong-Jin Ong

Department of Mechanical Engineering,

National University of Singapore

1 / 24
Outline

1 Introduction

2 Controllability

3 Reachability

2 / 24
Introduction

This chapter deals with two concepts that are not found in classical or
frequency analysis.
Controllability and Observability are two unique features of the state
space analysis.
These concepts were first introduced by E.G. Gilbert and R.F. Kalman
in the 1960s.
Explain why cancellation of unstable poles are undesirable even if perfect
cancellation is possible.
The concepts are first illustrated via several motivating examples.

3 / 24
Example

Example 1: Consider the system

ẋ1 = u
ẋ2 = u

If x1 (0) = x2 (0), then x1 (t) = x2 (t) for all time and all control u(t).

Example 2: Consider the system

ẋ1 = x2
ẋ2 = 0
y = x2

Then y(t) = constant for all choices of x1 (0) and x2 (0). Observing y(t) does not tell
us what x1 is doing.

4 / 24
Motivating Example
A more interesting example is one given by the following diagram. It consists of two
carts coupled by a spring. A force f is applied onto both carts via some means
within the system.

Figure: The Two cart example

The equations of motion of the system are

ẋ1 = x3 , ẋ2 = x4
k f k f
ẋ3 = − (x1 − x2 ) − , ẋ4 = − (x2 − x1 ) +
m1 m1 m2 m2

5 / 24
Motivating Example

Rewritten in standard state space representation, the state equation is

0 0 1 0
   
0
 0 0 0 1   x +  01  f
 
ẋ = 
 − mk k
1 m1
0 0   − m1

k k 1
m2
− m2
0 0 m2

From law of physics, f can change the relative distance between the two carts
(x2 − x1 ) but it cannot change x1 and x2 independently.

6 / 24
Example
Example 4: Consider the system given below,

Figure: The unobservable system

The equations describing the system are;

1 1 1 1 1
v̇1 = − ( + )v1 + v2 + V
C1 R1 R2 C1 R 3 C1 R 1
1 1 1 1 1
v̇2 = v1 − ( + )v2 + V
C2 R 3 C2 R 2 R3 C2 R2

7 / 24
Example

Consider the voltage across R3 , v̄ = v1 − v2 then

   
1 1 1 1 1 1 1 1 R2 C2 − R1 C1
v̄˙ = − ( + )+ + + ( + ) + V
C1 R1 R3 C2 R3 C1 R 3 C2 R2 R3 C1 C2 R1 R2
If the bridge is balanced, then R1 C1 = R2 C2 , the coefficient of V vanishes and
R1 + R2 + R3
v̄˙ = −( )v̄
C1 R1 R3
which implies that v̄ is not influenced by V and the voltage v̄ decays from v̄(0) to
zero, i.e., v̄ is not controllable.

8 / 24
Controllability and Observability

The above shows practical examples of uncontrollable and unobservable

systems.
This chapter studies analysis tools that checks for controllability and
observability.
Given (A, B, C, D), answer the questions:
I Can we drive x(t) wherever we desire using u(t)?
I Can we track x(t) by observing y(t)?
We assume complete knowledge of (A, B, C, D) in answering these questions.

9 / 24
Controllability
Definition: A LTI system (A, B, C, D) (or more precisely (A, B)) is
controllable if there exists an input u(t), 0 ≤ t ≤ t1 that drives the system from
any initial state x(0) = x0 to any other state x(t1 ) = x1 in a finite time t1 .
Otherwise, (A, B) is said to be uncontrollable.
Possible to define the above for time-varying systems (more complicated since
reaching x1 may depend on t0 ). The above assumes t0 = 0 WLOG.
Theorem 4.1 The n-dimensional LTI system with matrices (A, B) is
controllable if and only if any of the following condition is satisfied:
1 The Controllability Grammian
Z t Z t
T T
W (0, t) = eAτ BB T eA τ dτ = eA(t−τ ) BB T eA (t−τ ) dτ
0 0
is nonsingular for all t ≥ 0.
2 The n × nr matrix
U = B AB · · · An−1 B
 

is full row rank.

3 The n × (n + r) matrix [A − λI B] has full row rank at every
eigenvalue λ of A.
10 / 24
Controllability
Case 1: (⇒) Suppose W (0, t) is non-singular for all t > 0 and x(t1 ) = x1 . Choose
the input as
T
u(t) = −B T eA (t1 −t) W −1 (0, t1 )(eAt1 x0 − x1 )
Using this input on the system, the state
Z t1
x(t1 ) = eAt1 x0 + eA(t1 −τ ) Bu(τ )dτ
0
Z t1
T
= eAt1 x0 − eA(t1 −τ ) BB T eA (t1 −τ ) W −1 (0, t1 )(eAt1 x0 − x1 )dτ
0

= eAt1 x0 − (eAt1 x0 − x1 ) = x1
(⇐) Suppose W (0, t) is singular for some t1 > 0, and that the system is controllable.
We show that this leads to a contradiction. Since W (0, t1 ) is singular, there exists a
non-zero α such that
Z t1
T
αT W (0, t1 )α = αT eA(t1 −τ ) BB T eA (t1 −τ ) α = 0
0
Z t1
Or, kαT eA(t1 −τ ) Bk22 dτ = 0
0
Since k · k is a non-negative function, this means that
αT eA(t1 −τ ) B = 0 for all t > τ > 0 (1)
11 / 24
Controllability

Since the system is controllable, this means that any state is reachable by an
appropriate control in finite time. Suppose we want to reach the state
x(t1 ) = α + eAt1 x0 at time t1 . Then
Z t1
x(t1 ) = α + eAt1 x0 = eAt1 x0 + eA(t1 −τ ) Bu(τ )dτ
0
Z t1
or, α = eA(t1 −τ ) Bu(τ )dτ
0
This also means that
Z t1
αT α = αT eA(t1 −τ ) Bu(τ )dτ = 0
0

from (1) which then implies that α = 0 which contradicts that α = 0.

12 / 24
Controllability
Case (2) (⇒ and ⇐): Various ways of showing this exist and this one given here is
more intuitive but less rigorous. It shows a clear connection to Caley-Hamilton
Principle. Recall that
Z t1
At1
x1 = e x0 + eA(t1 −τ ) Bu(τ )dτ
0
Z t1 Z t1
−At1 −At1 A(t1 −τ )
e x1 − x0 = e e Bu(τ )dτ = e−Aτ Bu(τ )dτ (2)
0 0

and from CH Principle,

n−1
X
e−Aτ = αk (τ )Ak
k=0

Using this into (2),

n−1
X Z t1 n−1
X
e−At1 x1 − x0 = Ak B αk (τ )u(τ )dτ = Ak Bβk
k=0 0 k=0
R t1
where βk = 0
αk (τ )u(τ )dτ .
The above
 equation can only
 be satisfied identically for any x0 , x1 and t1 if and
only if B AB · · · An−1 B is full row rank
13 / 24
Controllability

Case (3) (⇒) We want to show that

U is full rank ⇒ {[A − λI B] has full row rank at every λ of A.}

Suppose this is not true i.e., there exists a λ1 and a non-zero q s.t.

q T [A − λ1 I B] = 0

This implies that

q T A − q T λ1 = 0 and q T B = 0,

q T [B AB · · · An−1 B] = [q T B q T AB · · · q T An−1 B]
= [q T B λ1 q T B · · · λn−1
1 q T B] = 0

Hence, U is not full row rank.

14 / 24
Controllability

Case (3) (⇐) We want to show that

{[A − λI B] has full row rank at every λ of A.} ⇒ {U is full rank }

Need an additional result that controllability is invariant under similarity

transformation, a result to be shown later. Using that, and suppose rank U = n − m
for some integer m ≥ 1, (A, B) can be expressed as
   
Āc Ā12 B̄c
Ā = P AP −1 = B̄ = P B =
0 Āc̄ 0

where Āc̄ is m × m. Let λ1 be an eigenvalue and q1 be the left eigenvector of Āc̄ i.e.,
q1T Āc̄ = λ1 Āc̄ or (q1T − λ1 I)Āc̄ = 0. We form the n vector q as q T = [0 q1T ] where 0 is
a (n − m) row vector. Then, we have
 

Āc − λ1 I Ā12 B̄c
q T [Ā − λ1 I B̄] = 0 q1T =0
0 Āc̄ − λ1 I 0

This implies that [A − λI B] is not full rank at eigenvalue λ1 . (q.e.d.)

15 / 24
Controllability

The n × nr matrix U := [B AB · · · An−1 B] is commonly known as the

controllability matrix.
In the case of single-input system, r = 1 and U is a square matrix with
controllability ensured when U is non-singular.
Controllability is a concept on the ability of a system to move from one point
to another under the influence of u(t).
It does NOT take into account physical constraint.
Knowing that a system is controllable does not mean that it can be controlled
in practice.
However, knowing that a system is not controllable means one should not try
to move system state arbitrarily.

16 / 24
Example
Consider the system given below:

Figure: The Beam example

The equation of the system is

      
ẋ1 −0.5 0 x1 0.5
= + u
ẋ2 0 −1 x2 1

Note that the zero-input system is asymptotically stable.

17 / 24
Controllability

Questions: Is there an input u(t) such that, when applied, brings the platform
system to equilibrium in 2 seconds starting from x1 (0) = 10 and x2 (0) = −1?
 
0.5 −0.25
Answer: First check that rank[B AB]=rank( = 2.
1 1
Hence, it is possible to bring the system to the origin in 2 second. The
expression is
T
u(t) = −B T eA (t1 −t) W −1 (0, t1 )(eAt1 x0 − x1 ) with
Z 2  −0.5τ

e−0.5τ
   
e 0 0.5 0
W (0, 2) = −τ 0.5 1 −τ dτ
0 0 e 1 0 e
 
0.2162 0.3167
=
0.3167 0.4908
The corresponding u(t) = −58.82e0.5t + 27.96et .

18 / 24
Example

Consider two cases: (1) no constraint on u; and (2) −10 ≤ u(t) ≤ 10,

Figure: Plots of x1 (t), x2 (t) and u(t) for Figure: Plots of x1 (t), x2 (t) and u(t) for
transfer from x(0) = (10, −1)T in 2 seconds. transfer from x(0) = (10, −1)T in 4 seconds.

19 / 24
Controllability under Similarity Transformation

What happens to Controllability of a system when it undergoes a coordinate

change?
Consider the case of x = T x̄, the new state equations are

x̄˙ = T −1 AT x̄ + T −1 Bu
y = CT x + Du

The new controllability matrix is

Ū = T −1 B T −1 AB · · · T −1 An−1 B = T −1 B AB · · · An−1 B = T −1 U
   

Hence, rank of Ū = rank of U (why?) and Controllability is not affected by

similarity transformation.

20 / 24
Reachable States and Controllability Grammian

A closely related concept to Controllability is Reachability.

Reachability has the same definition of Controllability except that x(0) = 0.
A state x` is reachable if there exists a u(t) such that
Z `
x` = eA(`−t) Bu(t)dt
0

The set of reachable state, R, is the collection of all x` over all possible u.
Note that R is a linear subspace of Rn : for all α, β ∈ R,

xa , xb ∈ R ⇒ αxa + βxb ∈ R
R` T
Recall W (0, `) = 0
eA(`−τ ) BB T eA (`−τ )
dτ . We claim that

R = Range space of W (0, `)

Proof is omitted.

21 / 24
Controllability Grammian is time-independent

We now show that rank W is independent of the terminal time `. To do so, we

note that
rank(W (0, `)) = rank B AB · · · An−1 B = rank(U )
 

The special case when W is full rank is already established earlier.

The proof is given in the next page. However, here is an intuitive reasoning.
Since controllability requires that W (0, `) to be non-singular for all ` > 0 and
the full rank condition of U is independent of `. This provides plausible reason
to the validity of the above.

22 / 24
Controllability Grammian is time-independent
Proof: (Optional) (⇒)
Z ` T
q T W (0, `) = 0 ⇒ q T eA(`−τ ) BB T eA (`−τ )
qdτ = 0
0
T A(`−τ )
⇒q e B = 0, for all τ ∈ [0, `]. (3)
T
Let τ = ` then the above implies q B = 0. Differentiate (3) once yields
(−1)q T AeA(`−τ ) B = 0
which implies, when τ = `,
(−1)q T AB = 0
Repeating the differentiation n − 1 times yields
h i
q T B (−1)AB · · · (−1)k Ak B, · · · (−1)n−1 An−1 B = 0

which implies that q T B AB · · · An−1 B = 0. This means rank(W (0, `)) ≤ rank(U ).
 

(⇐) Let q T be in the left nullspace of U . Then

q T B AB · · · An−1 B = 0 ⇒ q T W (0, `) = 0 for any ` > 0
 

This implies rank(U ) ≤ rank(W (0, `)).

Using (⇒) and (⇐), rank(W (0, `)) = rank(U ).
Since U is independent of `, so is rank(U ) and rank(W ).
WLOG, denote W = W (0, ∞).
23 / 24
Computation of Controllability Grammian

R∞ T
The expression of W = 0
eA(t−τ ) BB T eA (t−τ )
dτ is hard to compute.
There is another expression that allows easy computation. Note that the W
satisfies the Lyapunov Equation

AW + W AT = −BB T

such that the solution of the above yields W .

This follows because (assuming A is stable)
Z ∞
d At T
AW + W AT = (e BB T eA t )dt = −BB T
0 dt

24 / 24