Controllability and Observability

Dr. Wenjie Dong

Department of Electrical and Computer Engineering,

The University of Texas Rio Grande Valley

W. Dong (UTRGV, Dept. of ECE) 1 / 94

Controllability

Consider the n-dimensional p-input state equation

ẋ = Ax + Bu (1)

where A and B are, respectively, n × n and n × p real constant matrices.

Definition 1
The state equation (1) or the pair (A, B) is said to be controllable if for
any initial state x(0) = x0 and any final state x1 . There exists an input
that transfers x0 to x1 in a finite time. Otherwise (1) or (A, B) is said to
be uncontrollable.

W. Dong (UTRGV, Dept. of ECE) 1 / 94

Example

Consider the network shown in Fig. 1(a). Its state variable x is the voltage
across the capacitor. If x(0) = 0, then x(t) = 0 for all t ≥ 0 no matter
what input is applied. The system is not controllable.

Figure 1:

W. Dong (UTRGV, Dept. of ECE) 2 / 94

Example

Consider the network shown in Fig. 1(b). It has two state variables x1 and
x2 as shown. The input can transfer x1 or x2 to any value; but it cannot
transfer x1 and x2 to any values. For example. if x1 (0) = x2 (0) = 0, then
no matter what input is applied, x1 (t) always equals x2 (t) for all t ≥ 0.
Thus the equation that describes the network is not controllable.

W. Dong (UTRGV, Dept. of ECE) 3 / 94

Controllability

Theorem: The following statements are equivalent.

1. The n-dimensional pair (A, B) is controllable.
2. The n × n matrix
Z t Z t
Aτ 0 A0 τ
Wc (t) = e BB e dτ = e A(t−τ ) BB 0 e A (t−τ ) dτ (2)
0

0 0

is nonsingular for any t > 0.

3. The n × np controllability matrix

C = [B, AB, A2 B, . . . , An−1 B] (3)

has rank n (full row rank).

4. The n × (n + p) matrix [A − λI, B] has full row rank at every
eigenvalue, λ, of A.

W. Dong (UTRGV, Dept. of ECE) 4 / 94

Controllability

5. If, in addition, all eigenvalues of A have negative real parts, then the
unique solution of

AWc + Wc A0 = −BB 0 (4)

is positive definite. The solution is called the controllability Gramian and

can be expressed as
Z ∞
Wc = e Aτ BB 0 e A τ dτ (5)
0

W. Dong (UTRGV, Dept. of ECE) 5 / 94

Controllability
Proof: The equivalence of the two terms in (2) can be proved by letting
v = t − τ.
(1) ← (2): For any time t1 ,
Z t1
At1
x(t1 ) = e x(0) + e A(t1 −τ ) Bu(τ )dτ (6)
0

choose the control input

u(t) = −B 0 e A (t1 −t) Wc−1 (t1 )[e At1 x(0) − x1 ] (7)

0

then
Z t1 
At1 A(t1 −τ ) 0 A0 (t1 −τ )
x(t1 ) = e x(0) − e BB e dτ Wc−1 (t1 )[e At1 x(0)
0
−x1 ] = x1

which means that (A, B) is controllable.

W. Dong (UTRGV, Dept. of ECE) 6 / 94
Controllability
(1) → (2): Prove by contradiction. Suppose Wc is singular at time t1 ,
then there exists a nonzero vector v such that W (t1 )v = 0. So,
Z t1 Z t1
0 A(t1 −τ ) 0 A0 (t1 −τ )
v Wc (t1 )v = ve BB e vdτ = kB 0 e A (t1 −τ ) vk2 dτ = 0
0 0

0 0

which implies that

B 0 e A (t1 −τ ) v ≡ 0
0

for all τ in [0, t1 ]. If (A, B) is controllable, there exists an input that

transfers the initial state x(0) = e −At1 v to x(t1 ) = 0 and (6) becomes
Z t1
0=v+ e A(t1 −τ ) Bu(τ )dτ
0

So Z t1
0=vv+ 0
v 0 e A(t1 −τ ) Bu(τ )dτ = v 0 v + 0 = kvk2
0
which contradicts v 6= 0. So, Wc is nonsingular.
W. Dong (UTRGV, Dept. of ECE) 7 / 94
Controllability

(2) → (3): Prove by contradiction. Suppose C is singular. Then there

exists a nonzero vector v such that v 0 C = 0 or

v 0 Ak B = 0, k = 0, 1, 2, . . . , n − 1

Because e At B can be expressed as a linear combination of

{B, AB, . . . , An−1 B}, v 0 e At B = 0. This contradicts the nonsingularity
assumption of Wc (t).
(2) ← (3): Suppose Wc (t) is singular. Then exist a nonzero vector v such
that v 0 e At B = 0 for all t. Setting t = 0, we have v 0 B = 0. Differentiating
e At B and then setting t = 0, we have v 0 AB = 0. Proceeding forward
yields v 0 Ak B = 0 for k = 0, 1, 2, . . .. So,

v 0 [B, AB, . . . , An−1 B] = v 0 C = 0

which contradicts that C has full rank.

W. Dong (UTRGV, Dept. of ECE) 8 / 94
Controllability
(3) → (4): Show by contradiction. Suppose there exists an eigenvalue λ1
and a row vector q 6= 0 such that
q[A − λ1 I, B] = 0
which implies that qA = λ1 q and qB = 0. So,
qA2 = (qA)A = (λ1 q)A = λ21 q
Proceeding forward, we have qAk = λk1 q. Thus we have
q[B, AB, . . . , An−1 B] = [qB, λ1 qB, . . . , λn−1
1 qB] = 0
This contradicts that C has full row rank.
(3) ← (4): One fact: Controllability is invariant under any equivalent
transformation. Show by contradiction. If ρ(C) = n − m, there exists a
nonsingular matrix P such that
Āc Ā12 B̄c
   
Ā = PAP =−1
, B̄ = PB =
0 Āc̄ 0
W. Dong (UTRGV, Dept. of ECE) 9 / 94
Controllability

where Āc̄ is m × m. Let λ1 be an eigenvalue of Āc̄ and q1 be a

corresponding 1 × m nonzero left eigenvector or q1 Āc̄ = λ1 q1 . Now we
form the 1 × n vector q := [0, q1 ]. Then

Āc − λ1 I Ā12 B̄c

 
q[Ā − λ1 I, B̄] = [0, q1 ] =0 (8)
0 Āc̄ − λ1 I 0

which implies ρ([Ā − λI, B̄]) < n and consequently, ρ([A − λI, B]) < n at
some eigenvalue of A, which contradicts [A − λI, B] has full row rank.
(2) ↔ (5): If A is stable, then the unique solution of (4) can be expressed
as in (5). The Gramian Wc is always positive semidefinite. It is positive
definite if and only if Wc is nonsingular. 

W. Dong (UTRGV, Dept. of ECE) 10 / 94

Example

Consider the system

0 1 0 0 0
   
 0 0 −1 0   1
ẋ x +  u

= 
 0 0 0 1   0 
0 0 5 0 −2
y = [1, 0, 0, 0]x

is the system controllable?

W. Dong (UTRGV, Dept. of ECE) 11 / 94

Example

Solution:
0 1 0 2
 
 1 0 2 0 
C = [B, AB, A2 B, A3 B] = 
 0 −2 0

−10 
−2 0 −10 0

Its rank is 4. The system is controllable.

W. Dong (UTRGV, Dept. of ECE) 12 / 94

Controllability

MATLAB functions: ctrb and gram will generate the controllability matrix
and controllability Gramian.
Cm=ctrb(A,B);
SYS=ss(A,B,C,D);
Wc=gram(SYS,’c’);

W. Dong (UTRGV, Dept. of ECE) 13 / 94

Example

Consider the system

−0.5 0 0.5
   
ẋ = x+ u
0 −1 1

is the system controllable?

W. Dong (UTRGV, Dept. of ECE) 14 / 94

Example

Solution:
0.5 −0.25
 
C = [B, AB] =
1 −1
Its rank is 2. The system is controllable.

W. Dong (UTRGV, Dept. of ECE) 15 / 94

Example

Consider the system

−1 0 1
   
ẋ = x+ u
0 −1 1

is the system controllable?

W. Dong (UTRGV, Dept. of ECE) 16 / 94

Example

Solution:
1 −1
 
C = [B, AB] =
1 −1
Its rank is 1. The system is not controllable.

W. Dong (UTRGV, Dept. of ECE) 17 / 94

Controllability Indices

For a system (A, B), assume B ∈ R p has full column rank and
rank(C) = n. Let µm be the least integer such that column vector Aµm bm
is linearly dependent on column vectors occurring to the left of it in the
controllability matrix C = [B, AB, A2 B, . . . , An−1 B] for 1 ≤ m ≤ p, then

µ1 + µ2 + · · · + µp = rank(C) = n (9)

The set {µ1 , µ2 , . . . , µp } is called the controllability indices and

µ = max{µ1 , µ2 , . . . , µp }

is called the controllability index of (A, B). Or, equivalently, the

controllability index µ is the least integer such that

ρ(Cµ ) = ρ([B, AB, . . . , Aµ−1 B]) = n (10)

W. Dong (UTRGV, Dept. of ECE) 18 / 94

Controllability Indices

The bounds of µ are

n
≤ µ ≤ min{n̄, n − (p − 1)}
p

where n̄ is the degree of the minimal polynomial of A and p = ρ(B).

Corollary 2
The n-dimensional pair (A, B) is controllable if and only if the matrix

Cn−p+1 := [B, AB, . . . , An−p B] (11)

where ρ(B) = p, has rank n or the n × n matrix Cn−p+1 Cn−p+1

0 is
nonsingular.

W. Dong (UTRGV, Dept. of ECE) 19 / 94

Example

Consider the system

0 1 0 0 0 0
   
 3 0 0 2  1 0 
ẋ x + u
 
= 
 0 0 0 1   0 0 
0 −2 0 0 0 1
1 0 0 0
 
y = x
0 0 1 0

find the controllability indices.

W. Dong (UTRGV, Dept. of ECE) 20 / 94

Example

Solution: Its controllability matrix C is 4 × 8. If we apply Corollary 2, the

controllability can be determined by the following

0 0 1 0 0 2 −1 0
 
1 0 0 2 −1 0 0 −2 
[B, AB, A2 B, A3 B] = 

 0 0 0 1 −2 0 0 −4 


0 1 −2 0 0 4 2 0

It can be verified that µ1 = 2 and µ2 = 2. It has rank 4. The system is

controllable.

W. Dong (UTRGV, Dept. of ECE) 21 / 94

Controllability

Theorem 3
The controllability property is invariant under any equivalence
transformation.

Proof.
Assume the system is (A, B), after an equivalence transform is P, the
system is (PAP −1 , PB). So

C¯ = [B̄, B̄ Ā, . . . , B̄ Ān−1 ]

= [PB, PAP −1 PB, . . . , PAn−1 P −1 PB]
= P[B, AB, . . . , An−1 B] = PC

Since P is nonsingular, ρ(C)

¯ = ρ(C).

W. Dong (UTRGV, Dept. of ECE) 22 / 94

Controllability

Theorem 4
The set of the controllability indices of (A, B) is invariant under any
equivalence transformation and any reordering of the columns of B.

Proof: Let us define

Ck = [B, AB, . . . , Ak−1 B] (12)

By following the proof of Theorem 3, we have

ρ(Ck ) = ρ(C¯k )

for k = 0, 1, . . .. Thus the set of controllability indices is invariant under

any equivalent transformation. 

W. Dong (UTRGV, Dept. of ECE) 23 / 94

Controllability

The arrangement of the columns of B can be achieved by

B̂ = BM

where M is a p × p nonsingular permutation matrix. It is straightforward

to verify

Cˆk := [B̂, AB̂, . . . , Ak−1 B̂] = Ck diag(M, M, . . . , M)

Since diag(M, M, . . . , M) is nonsingular, we have ρ(Cˆk ) = ρ(Ck ) for

k = 0, 1, . . .. 

W. Dong (UTRGV, Dept. of ECE) 24 / 94

Observability

Consider the n-dimensional p-input q-output state equation

ẋ = Ax + Bu (13)
y = Cx + Du (14)

where A, B, C, and D are, respectively, n × n, n × p, q × n, and q × p

constant matrices.
Definition 5
The state equation (13) is said to be observable if for any unknown initial
state x(0), there exists a finite t1 > 0 such that the knowledge of the
input u and the output y over [0, t1 ] suffices to determine uniquely the
initial state x(0). Otherwise, the equation is said to be unobservable.

W. Dong (UTRGV, Dept. of ECE) 25 / 94

Example

Consider the network shown in Fig. 2. Determine its observability.

Figure 2: Unobservable network

W. Dong (UTRGV, Dept. of ECE) 26 / 94

Example

Solution: If the input is zero, no matter what the initial voltage across the
capacitor is, the output is identically zero because of the symmetry of the
four resistors. We know the input and output (both are identically zero),
but we cannot determine uniquely the initial state. Thus the system is not
observable.

W. Dong (UTRGV, Dept. of ECE) 27 / 94

Example

Consider the network shown in Fig. 3(a). Determine its observability.

Figure 3: Unobservable network

W. Dong (UTRGV, Dept. of ECE) 28 / 94

Example

Solution: The network has two state variables: the current x1 through the
inductor and the voltage x2 , across the capacitor. The input u is a current
source. If u = 0, the network reduces to the one shown in Fig. 3(b). If
x1 (0) = a 6= 0 and x2 = 0, then the output is identically zero. Any
x(0) = [a, 0]0 and u(t) ≡ 0 yield the same output y(t) ≡ 0. Thus there is
no way to determine the initial state [a, 0]0 uniquely and the equation that
describes the network is not observable.

W. Dong (UTRGV, Dept. of ECE) 29 / 94

Observability

The response of the system (13)-(14) is

Z t
At
y(t) = Ce x(0) + C e A(t−τ ) Bu(τ )dτ + Du(t) (15)
0

Then,

Ce At x(0) = ȳ(t) (16)

where Z t
ȳ(t) = y(t) − C e A(t−τ ) Bu(τ )dτ − Du(t)
0
is a known function.
• Equation (13) is observable if and only if the initial state x(0) can be
determined uniquely from its zero-input response over a finite time interval.

W. Dong (UTRGV, Dept. of ECE) 30 / 94

Observability

Theorem 6
The state equation (13) is observable if and only if the n × n matrix
Z t
Wo (t) = e A τ C 0 Ce Aτ dτ (17)
0

is nonsingular for any t > 0.

W. Dong (UTRGV, Dept. of ECE) 31 / 94

Observability

Proof We pre-multiply (16) by e At C 0 and then integrate it over [0, t1 ] to

yield Z t1  Z t1
A0 t 0 At
e C Ce dt x(0) = e A t C 0 ȳ(t)dt
0

0 0

If Wo (t1 ) is nonsingular, then

Z t1
x(0) = Wo−1 (t1 ) e A t C 0 ȳ(t)dt (18)
0

This yields a unique x(0). This shows that if Wo (t), for any t > 0, is
nonsingular, then (13) is observable.

W. Dong (UTRGV, Dept. of ECE) 32 / 94

Observability

Next we show that if Wo (t1 ) is singular or, equivalently, positive

semidefinite for all t1 , then (13) is not observable. If Wo (t1 ) is positive
semidefinite, there exists an nx1 nonzero constant vector v such that
Z t1 Z t1
0 A0 τ 0 Aτ
v Wo (t1 )v =
0
v e C Ce vdτ = kCe Aτ vk2 dτ = 0
0 0

which implies

Ce At v ≡ 0 (19)

for all t in [0, t1 ]. If u ≡ 0, then x1 (0) = v 6= 0 and x2 (0) = 0 both yield

the same
y(t) = Ce At xi (0) ≡ 0
Two different initial states yield the same zero-input response; therefore we
cannot uniquely determine x(0). Thus (13) is not observable. 
W. Dong (UTRGV, Dept. of ECE) 33 / 94
Observability

• Observability depends only on A and C.

• If Wo (t) is nonsingular for some t, then it is nonsingular for every t and
the initial state can be computed from (18) by using any nonzero time
interval.
Theorem 7 (Theorem of Duality)
The pair (A, B) is controllable if and only if the pair (A0 , B 0 ) is observable.

W. Dong (UTRGV, Dept. of ECE) 34 / 94

Observability

Proof: The pair (A, B) is controllable if and only if

Z t
Wc (t) = e Aτ BB 0 e A τ dτ
0

is nonsingular for any t. The pair (A0 , B 0 ) is observable if and only if, by
replacing A by A0 and C by B 0 in (17),
Z t
Wo (t) = e Aτ BB 0 e A τ dτ
0

is nonsingular for any t. The two conditions are identical and the theorem
follows. 

W. Dong (UTRGV, Dept. of ECE) 35 / 94

Observability
Theorem: The following statements are equivalent.
1 The n-dimensional pair (A, C) is observable.

2 The n × n matrix

Z t
Wo (t) = e A τ C 0 Ce Aτ dτ (20)
0

is nonsingular for any t > 0.

3 The nq × n observability matrix

C
 
 CA 
O= .. (21)
 
.

 
CAn−1

has rank n (full column rank). This matrix can be generated by

calling obsv in MATLAB.
W. Dong (UTRGV, Dept. of ECE) 36 / 94
Observability
4 The (n + q) × n matrix
A − λI
 

C
has full column rank at every eigenvalue, λ, of A.
5 If, in addition, all eigenvalues of A have negative real parts, then the
unique solution of

A0 Wo + Wo A = −C 0 C (22)

is positive definite. The solution is called the observability Gramian

and can be expressed as
Z ∞
Wo = e A τ C 0 Ce Aτ dτ (23)
0

The theorem can be proved either directly or by applying the theorem of

duality.
W. Dong (UTRGV, Dept. of ECE) 37 / 94
Observability Indices
For a system (A, C), assume C ∈ R q has full column rank and
rank(O) = n. Let νm be the least integer such that row vector cm Aνm is
linearly dependent on its above rows in the observability matrix O for
1 ≤ m ≤ q, then
ν1 + ν2 + · · · + νp = rank(O) = n (24)
The set {ν1 , ν2 , . . . , νp } is called the observability indices and
ν = max{ν1 , ν2 , . . . , νp }
is called the observability index of (A, C). Or, equivalently the
observability index ν is the least integer such that
C
 
 CA 
ρ(Oν ) = ρ  ..  = n (25)
 
 . 
CAν−1
W. Dong (UTRGV, Dept. of ECE) 38 / 94
Observability Indices
The bounds of ν are
n
≤ ν ≤ min{n̄, n − (q − 1)}
q

where n̄ is the degree of the minimal polynomial of A and q = ρ(C).

Corollary 8
The n-dimensional pair (A, C) is observable if and only if the matrix

C
 
 CA 
On−q+1 :=  .. (26)
 
.

 
CAn−q

where ρ(C) = q, has rank n or the n × n matrix On−p+1 On−p+1

0 is
nonsingular.
W. Dong (UTRGV, Dept. of ECE) 39 / 94
Observability Indices

Theorem 9
The observability property is invariant under any equivalence
transformation.

Theorem 10
The set of the observability indices of (A, C) is invariant under any
equivalence transformation and any reordering of the columns of C.

W. Dong (UTRGV, Dept. of ECE) 40 / 94

Observability

x(0) can be solved by differentiating

Ce At x(0) = ȳ(t)

repeatedly and set t = 0. We have

C ȳ(0)
   
 CA   ȳ(0)
˙ 
..  x(0) =  ..  =: ỹ(0)
   
. .

   
CAν−1 ȳ (ν−1) (0)
or
Onu x(0) = ỹ(0)
then
x(0) = [Oν0 Oν ]−1 Oν0 ỹ(0)

W. Dong (UTRGV, Dept. of ECE) 41 / 94

Example

Consider the system

1 1 0 0 1
   

ẋ =  0 1 0  x +  1 0  u, y = [1, 1, 1]x (27)

0 1 1 0 1

is it observable or not?

W. Dong (UTRGV, Dept. of ECE) 42 / 94

Example

Solution: The observability matrix

C 1 1 1
   

O =  CA  =  1 3 1 
CA2 1 5 1

ρ(O) = 2. The system is not observable.

W. Dong (UTRGV, Dept. of ECE) 43 / 94

Example

Consider the system

1 1 0 0 0 0
   
 0 1 1 0 0   1 
ẋ =   0 0 1 0 0 0
x +   u, y = [0, 1, 1, 0, 1]x (28)
   
 0 0 0 2 1 0
  
  
0 0 0 0 2 1

is it observable or not?

W. Dong (UTRGV, Dept. of ECE) 44 / 94

Example

Solution: The eigenvalues of A are λ1 = 1 and λ2 = 2.

0 1 0 0 0
 
 0 0 1 0 0 
A − λ1 I  0 0 0 0 0 
   
=
C  0 0 0 1 1 

 0 0 0 0 1 
 

0 1 1 0 1

its column rank is 4. The system is not observable.

W. Dong (UTRGV, Dept. of ECE) 45 / 94

Conditions in Jordan-Form Equations
Consider the state equation

ẋ = Jx + Bu (29)
y = Cx (30)

where J is in Jordan form. To simplify discussion, we assume that J has

only two distinct eigenvalues λ1 and λ2 and can be written as

J = diag(J1 , J2 )

where J1 consists of all Jordan blocks associated with λ1 and J2 consists

of all Jordan blocks associated with λ2 . Again to simplify discussion, we
assume that J1 has three Jordan blocks and J2 has two Jordan blocks or

J1 = diag(J11 , J12 , J13 ), J2 = diag(J21 , J22 )

The row of B corresponding to the last row of Jij is denoted by blij , The
column of C corresponding to the first column of Jij is denoted by cfij .
W. Dong (UTRGV, Dept. of ECE) 46 / 94
Conditions in Jordan-Form Equations

Theorem 11
1. The state equation in (29)-(30) is controllable if and only if the three
row vectors {bl11 , bl12 , bl13 } are linearly independent and the two row
vectors {bl21 , bl22 } are linearly independent.
2. The state equation in (29)-(30) is observable if and only if the three
column vectors {cf 11 , cf 12 , cf 13 } are linearly independent and the two
column vectors {cf 21 , cf 22 } are linearly independent.

W. Dong (UTRGV, Dept. of ECE) 47 / 94

Conditions in Jordan-Form Equations

Proof.
The controllability can be proved by using the condition that the matrix
[A − sI, B] has full row rank at every eigenvalue of A.
The
 observability can be proved by using the condition that the matrix
A − sI

has full column rank at every eigenvalue of A.
C

W. Dong (UTRGV, Dept. of ECE) 48 / 94

Conditions in Jordan-Form Equations

Remarks:
If a state equation is in Jordan form. then the controllability of the
state variables associated with one eigenvalue can be checked
independently from those associated with different eigenvalues.
The controllability of the state variables associated with the same
eigenvalue depends only on the rows of B corresponding to the last
row of all Jordan blocks associated with the eigenvalue. All other
rows of B play no role in determining the controllability.
The observability of the state variables associated with the same
eigenvalue depends only on the columns of C corresponding to the
first column of all Jordan blocks associated with the eigenvalue. All
other columns of C play no role in determining the observability.

W. Dong (UTRGV, Dept. of ECE) 49 / 94

Example

Consider the Jordan-form state equation

λ1 1 0 0 0 0 0 0 0 0
   
 0 λ1 0 0 0 0 0   1 0 0 
 0 0 λ1 0 0 0 0 1 0
   
  
ẋ =   0 0 0 λ1 0 0 0 1 1 1
x +  u
   
 0 0 0 0 λ2 1 0 1 2 3
  
  
 0 0 0 0 0 λ2 1 0 1 0
   
  
0 0 0 0 0 0 λ2 1 1 1
1 1 2 0 0 2 1
 

y =  1 0 1 2 0 1 1 x
1 0 2 3 0 2 0

determine its controllability and observability.

W. Dong (UTRGV, Dept. of ECE) 50 / 94

Example

Solution: The matrix J has two distinct eigenvalues λ1 and λ2 . There are
three Jordan blocks, with order 2, 1 and 1, associated with λ1 . The rows
of B corresponding to the last row of the three Jordan blocks are [1, 0, 0],
[0, 1, 0], and [1, 1, 1]. The three rows are linearly independent. There is
only one Jordan block with order 3 associated with λ2 . The row of B
corresponding to the last row of the Jordan block is [1, 1, 1] which is
nonzero and is therefore linearly independent. Thus we conclude that the
state equation is controllable.
The conditions for the system to be observable are that the three columns
[1, 1, 1]0 , [2, 1, 2]0 and [0, 2, 3]0 are linearly independent (they are) and the
one column [0, 0, 0]0 is linearly independent (it is not). Therefore the state
equation is not observable.

W. Dong (UTRGV, Dept. of ECE) 51 / 94

Conditions in Jordan-Form Equations

Corollary 12
Consider an n-dimensional Jordan-form state equation with p inputs and q
outputs.
A necessary condition for the state equation to be controllable is
m ≤ p.
A necessary condition for the state equation to be observable is
m ≤ q.

W. Dong (UTRGV, Dept. of ECE) 52 / 94

Conditions in Jordan-Form Equations

Corollary 13
A single-input Jordan-form state equation is controllable if and only if
there is only one Jordan block associated with each distinct eigenvalue and
every entry of B corresponding to the last row of each Jordan block is
different from zero.

Corollary 14
A single-output Jordan-form state equation is observable if and only if
there is only one Jordan block associated with each distinct eigenvalue and
every entry of C corresponding to the first column of each Jordan block is
different from zero.

W. Dong (UTRGV, Dept. of ECE) 53 / 94

Example

Consider the state equation

0 1 0 0 10
   
 0 0 1 0
ẋ =  x +  9 u
  
 0 0 0 0   0 
0 0 0 −2 1
y = [1, 0, 0, 2]x + u

determine its controllability and observability.

W. Dong (UTRGV, Dept. of ECE) 54 / 94

Example

Solution:
There are two Jordan blocks, one with order 3 and associated with
eigenvalue 0, the other with order 1 and associated with eigenvalue −2.
The entry of B corresponding to the last row of the first Jordan block is
zero; thus the state equation is not controllable. The two entries of C
corresponding to the first column of both Jordan blocks are different from
zero; thus the state equation is observable.

W. Dong (UTRGV, Dept. of ECE) 55 / 94

Canonical Decomposition

Consider

ẋ = Ax + Bu (31)
y = Cx + Du (32)

Let x̄ = Px, where P is a nonsingular matrix. Then

x̄˙ = Āx̄ + B̄u (33)

y = C̄ x̄ + D̄u (34)

with Ā = PAP −1 , B̄ = PB, C̄ = CP −1 , and D̄ = D. The controllability

matrix and the observability matrix are

C¯ = PC, Ō = OP −1

W. Dong (UTRGV, Dept. of ECE) 56 / 94

Canonical Decomposition
Theorem 15
Consider the n-dimensional state equation in (31)-(32) with

ρ(C) = ρ([B, AB, . . . , An−1 B]) = n1 < n

Let the n × n matrix P −1 := [q1 , q2 , . . . , qn1 , . . . , qn ] where the first n1

columns are any n1 linearly independent columns of C, and the remaining
columns can arbitrarily be chosen as long as P is nonsingular. Then the
equivalence transformation x̄ = Px or x = P −1 x will transform (31)-(32)
into
x̄˙ c Āc Ā12 x̄c B̄c
      
= + u (35)
x̄˙ c̄ 0 Āc̄ x̄c̄ 0
x̄c
 
y = [C̄c , C̄c̄ ] + Du (36)
x̄c̄

W. Dong (UTRGV, Dept. of ECE) 57 / 94

Canonical Decomposition

where Āc is n1 × n1 and Āc̄ is (n − n1 ) × (n − n1 ), and the n1 -dimensional

equation in (35)-(36)

x̄˙ c = Āc x̄c + B̄c u (37)

ȳ = C̄c x̄c + Du (38)

is controllable and has the same transfer matrix as (31)-(32), i.e.,

C(sI − A)−1 B + D = C̄c (sI − Āc )−1 B̄c + D

W. Dong (UTRGV, Dept. of ECE) 58 / 94

Canonical Decomposition

Proof:The transformation x = P −1 x̄ changes the basis of the state space

from the standard orthonornal basis to the columns of Q := P −1 or
{q1 , . . . , qn1 , . . . , qn }. The ith column of A is the representation of Aqi
with respect to {q1 , . . . , qn1 , . . . , qn }. Now the vector Aqi , for
i = 1, 2, . . . , n1 are linearly dependent on the set {q1 , . . . , qn1 , . . . , qn };
they are linearly independent of {qn1 +1 , . . . , qn }. Thus the matrix A has
the form shown in (49). The columns of B̄ are the representation of the
columns of B with respect to {q1 , . . . , qn1 , . . . , qn }. Now the columns of B
depend only on {ql , . . . , qn1 }; thus B has the form shown in (49). We
mention that if the n × p matrix B has rank p and if its columns are
chosen as the first p columns of P −1 , then the upper pant of B is the unit
matrix of order p.

W. Dong (UTRGV, Dept. of ECE) 59 / 94

Canonical Decomposition
Let C¯ be the controllability matrix of (49). Then we have
ρ(C) = ρ(C)¯ = n1 . It is straightforward to verify

B̄c Āc B̄c . . . Ānc 1 B̄c . . . Ān−1 B̄c

 
c
C¯ = =
0 0 ... 0 ... 0
C¯c Ānc 1 B̄c . . . Ān−1 B̄c
 
c
0 0 ... 0

where C̄c is the controllability matrix of (Āc , B̄c ). Because the columns of
Ākc B̄c , for k ≥ n1 , are linearly dependent on the columns of C¯c , the
condition ρ(C) = n1 , implies ρ(C¯c ) = n1 . Thus the n1 -dimensional state
equation in (37) is controllable.
Next we show that (37) has the same transfer matrix as (31). Because
(31) and (49) have the same transfer matrix. we need to show only that
(49) and (37) have the same transfer matrix. By direct verification, we
can show
W. Dong (UTRGV, Dept. of ECE) 60 / 94
Canonical Decomposition

sI − Āc −Ā12 (sI − Āc )−1 M

 −1  
= (39)
0 sI − Āc̄ 0 (sI − Āc̄ )−1

where
M = (sI − Āc )−1 Ā12 (sI − Āc̄ )−1
Thus the transfer matrix of (4) is

sI − Āc −Ā12 B̄c

 −1  
[C̄c , C̄c̄ ] +D
0 sI − Āc̄ 0
(sI − Āc )−1 M B̄c
  
= [C̄c , C̄c̄ ] +D
0 (sI − Āc̄ )−1 0
= C̄c (sI − Āc )−1 B̄c + D

which is the transfer matrix of (37). 

W. Dong (UTRGV, Dept. of ECE) 61 / 94
Canonical Decomposition

Remarks:
By the state transformation x̄ = Px, the n-dimensional state space is
divided into two subspaces. One is the n1 -dimensional subspace that
consists of all vectors of the form [x̄c0 , 0]0 ; the other is the
(n − n1 )-dimensional subspace that consists of all vectors of the form
[0, x̄c̄0 ]0 .
The input u can transfer x̄c from any state to any other state.
The input u cannot control x̄c̄ because, as we can see from (49) u
does not affect x̄c̄ directly, nor indirectly through the state x̄c .
By dropping the uncontrollable state vector, we obtain a controllable
state equation of lesser dimension that is zero-state equivalent to the
original equation.

W. Dong (UTRGV, Dept. of ECE) 62 / 94

Example

Consider the system

1 1 0 0 1
   

ẋ =  0 1 0  x +  1 0  u, y = [1, 1, 1]x (40)

0 1 1 0 1

convert it into canonical form.

W. Dong (UTRGV, Dept. of ECE) 63 / 94

Example
Solution: ρ(B) = 2. So we can use C2 = [B, AB], instead of
C = [B, AB, A2 B], to check the controllability of the system.
0 1 1 1
 

ρ(C2 ) = ρ  1 0 1 0  = 2 < 3
0 1 1 1
The system is not controllable. Let
0 1 1
 

P −1 =Q= 1 0 0 
0 1 0
then
1 0 0 1 0
   

Ā = PAP −1 =  1 1 0  , B̄ = PB =  0 1  , C̄ = [1, 2, 1]
0 0 1 0 0

W. Dong (UTRGV, Dept. of ECE) 64 / 94

Example

The controllable subsystem is

1 0 1 0
   
x̄˙ c = x̄c + u, y = [1, 2]x̄c
1 1 0 1

W. Dong (UTRGV, Dept. of ECE) 65 / 94

Canonical Decomposition

The MATlAB function ctrbf transforms (31) into (49) except that the
order of the columns in P −1 is reversed. Thus the resulting equation has
the form
Āc̄ 0 0
   
,
Ā21 Āc B̄c

W. Dong (UTRGV, Dept. of ECE) 66 / 94

Canonical Decomposition
Dual to Theorem 15, for unobservable state equations we have
Theorem 16

Consider the n-dimensional state equation in (31) with

C
 
 CA 
ρ(O) = ρ  ..  = n2 < n
 
 . 
CAn−1
We form the n × n matrix
p1
 
 .. 
 . 
P = p
 
n
 2 

 .. 
 . 
pn
W. Dong (UTRGV, Dept. of ECE) 67 / 94
Canonical Decomposition
where the first n2 columns are any n2 linearly independent rows of O, and
the remaining rows can arbitrarily be chosen as long as P is nonsingular.
Then the equivalence transformation x̄ = Px will transform (31) into

x̄˙ o Āo 0 x̄o B̄o

      
= + u (41)
x̄˙ ō Ā12 Āō x̄ō B̄ō
x̄o
 
y = [C̄o , 0] + Du (42)
x̄ō

where Āo is n2 × n2 and Āō is (n − n2 ) × (n − n2 ), and the n2 -dimensional

equation in (41)-(42)

x̄˙ o = Āo x̄o + B̄o u (43)

ȳ = C̄o x̄o + Du (44)

is observable and has the same transfer matrix as (31).

W. Dong (UTRGV, Dept. of ECE) 68 / 94
Canonical Decomposition

Remarks:
By the equivalence transformation x̄ = Px, the n-dimensional state
space is divided into two subspaces. One is the n2 -dimensional
subspace that consists of all vectors of the form [x̄o , 0]0 ; the other is
the (n − n2 )-dimensional subspace consisting of all vectors of the form
[00 , x̄ō0 ]0 .
The state x̄o can be detected from the output.
x̄ō cannot be detected from the output because, as we can see from
(41), it is not connected to the output either directly, or indirectly
through the state x̄o .
By dropping the unobservable state vector, we obtain an observable
state equation of lesser dimension that is zero-state equivalent to the
original equation.

W. Dong (UTRGV, Dept. of ECE) 69 / 94

Canonical Decomposition

The MATLAB function obsvf is the counterpart of ctrbf. Combining

Theorems 15 and 16, we have the following Kalman decomposition
theorem.

W. Dong (UTRGV, Dept. of ECE) 70 / 94

Canonical Decomposition

Theorem 17

Every state-space equation can be transformed, by an equivalence

transformation, into the following canonical form

x̄˙ co Āco 0 Ā13 0 x̄co B̄co

      
 x̄˙ c ō   Ā21 Āc ō Ā23 Ā24   x̄c ō   B̄c ō 
 u (45)
 x̄˙ c̄o  =  0
+
0 Āc̄o 0   x̄c̄o   0 
   

x̄˙ c̄ ō 0 0 Ā43 Āc̄ ō x̄c̄ ō 0

y = [C̄co , 0, C̄c̄o , 0]x̄ + Du (46)

where the vector x̄co is controllable and observable, x̄c̄ ō is controllable but
not observable, x̄c̄o is observable but not controllable, and x̄c̄ ō is neither
controllable nor observable.

W. Dong (UTRGV, Dept. of ECE) 71 / 94

Canonical Decomposition

Furthermore, the state equation is zero-state equivalent to the controllable

and observable state equation

x̄˙ co = Āco x̄co + B̄co u (47)

y = C̄co (sI − Āco ) −1
B̄co + D (48)

The equivalence transform x̄ = Px can be chosen as follows.

P −1 = [q11 , . . . , q1,n1 , q21 , . . . , q2,n2 , q31 , . . . , q3,n3 , q41 , . . . , q4,n4 ]

where {q21 , . . . , q2,n2 } forms a basis of R(C) ∩ N (O),

{q11 , . . . , q1,n1 , q21 , . . . , q2,n2 } forms a basis of R(C),
{q21 , . . . , q2,n2 , q31 , . . . , q3,n3 } forms a basis of N (O), {q41 , . . . , q4,n4 } are
chosen such that P −1 is nonsingular.

W. Dong (UTRGV, Dept. of ECE) 72 / 94

Canonical Decomposition

The controllable and observable state equation can be obtained by

MATLAB command minreal().

W. Dong (UTRGV, Dept. of ECE) 73 / 94

Canonical Decomposition
Procedure 1 for finding P:
1 Find the equivalence transformation Pc such that the system is
converted into the canonical controllable form with the aid of
Theorem 15.
x̄˙ c Āc Ā12 x̄c B̄c
      
= + u
x̄˙ c̄ 0 Āc̄ x̄c̄ 0
x̄c
 
y = [C̄c , C̄c̄ ] + Du
x̄c̄

2 For the controllable subsystem (Āc , C̄c ), find the equivalence

transformation Pco such that it can be converted into a canonical
observe form with the aid of Theorem 16.
3 For the uncontrollable subsystem (Āc̄ , C̄c̄ ), find the equivalence
transformation Pc̄o such that it can be converted into a canonical
observe form with the aid of Theorem 16.
W. Dong (UTRGV, Dept. of ECE) 74 / 94
Canonical Decomposition

4 Combine the transformations Pco and Pc̄o as one transformation

Pco 0
 
Po =
0 Pc̄o

5 The final transformation is

P = Po Pc

W. Dong (UTRGV, Dept. of ECE) 75 / 94

Example

Consider the system

1 1 0 0 0 0
   
 0 1 0 0 0   1 
ẋ  0 0 2 0 0 0
x +  u
   
= 
 0 0 0 3 0 0
  
  
0 0 0 0 3 1
y = [0, 1, 1, 0, 1]x + u

find its Kalman decomposition.

W. Dong (UTRGV, Dept. of ECE) 76 / 94

Example

Solution: The controllability matrix is

0 1 2 3 4
 
 1 1 1 1 1 
2 3 4
C = [B, AB, A B, A B, A B] =  0 0 0 0 0
 

0 0 0 0 0
 
 
1 3 9 27 81

The column rank of C is 3. We choose

0 1 2 0 0
 
 1 1 1 0 0 
 0 0 0
Pc−1 =  1 0
 

 0 0 0 0 1


1 3 9 0 0

W. Dong (UTRGV, Dept. of ECE) 77 / 94

Example
then
−3/2 3/4 0 0 1/4
 
 2 1/2 0 0 −1/2 
Pc =  −1/2 −1/4 0 0 1/4 
 

 0 0 1 0 0 


0 0 0 1 0
0 0 3 0 0
 
 1 0 −7 0 0 
Āc A12
 
Ā = Pc APc =  0 1 5
−1
0 0 =
 
0 Āc̄
 
 0 0 0 2 0 
0 0 0 0 3
1
 
 0 
 0  , C̄ = CPc = [2, 4, 10|, 0] = [C̄c , C̄c̄ ]
B̄ = Pc B =  −1
 

 0 
0
W. Dong (UTRGV, Dept. of ECE) 78 / 94
Example
For the pair (Āc , C̄c ), the observability matrix is

C̄c 2 4 10
   

Oc =  C̄c Āc  =  4 10 28 
C̄c Ā2c 10 28 82
Its row rank is 2. We choose
2 4 10
 

Pco =  4 10 28 
0 0 1

For the pair (Āc̄ , C̄c̄ ), the observability matrix is

C̄c̄
 
1 0
 
Oc̄ = C̄c̄ Āc̄ =
2 0
 
C̄c̄ Ā2c̄
Its row rank is 1.
W. Dong (UTRGV, Dept. of ECE) 79 / 94
Example

We choose
1 0
 
Pc̄o =
0 1
Define
Pco 0
 
Po =
0 Pc̄o
The transformation for the Kalman decomposition is

0 1 0 0 1
 
 0 1 0 0 3 
P = Po Pc =  0 0 1/4
 
−1/2 −1/4 
 0 0 1 0 0
 

0 0 0 1 0

W. Dong (UTRGV, Dept. of ECE) 80 / 94

Example

The Kalman decomposition is

0 1 0 0 0 2
   
ẋco  xco
   
 −3 4 0 0 0  4 
 ẋc ō    xc ō 
 −1 1/2 1 0 0 0
   
 =  + u
 
 ẋc̄o   xc̄o  
  
 0 0 0 2 0 0
 
ẋc̄ ō xc̄ ō
  
0 0 0 0 3 0
y = [1, 0010]x̄

W. Dong (UTRGV, Dept. of ECE) 81 / 94

Example

Consider the network shown in Fig. 4(a). Find its state space model.
Determine the controllability and the observability of the system.

Figure 4: Networks

W. Dong (UTRGV, Dept. of ECE) 82 / 94

Example
If we assign state variables as shown, we have

x1 = L1 ẋ2 for L1
u = x2 + C1 ẋ1 for C1 and L1
L2 ẋ4 + x4 = 0 for L2
v − (x3 + v1 ) v1 + x3
= + C2 ẋ3 for node B
1 1
v − v1 v1
+ C2 ẋ3 = for the node C
1 1
where v is the voltage of node A and v1 is the voltage of node C. Form
the above last two equation, we have

ẋ3 = −x3 /C2 .

The output is
y = Rx4 + v2 − v1 = x4 + u
W. Dong (UTRGV, Dept. of ECE) 83 / 94
Example
So, the network can be described by
0 −0.5 0 0 0.5
   
 1 0 0 0
ẋ =  x +  0 u
  
 0 0 −0.5 0   0 
0 0 0 −1 0
y = [0, 0, 0, 1]x + u
It is uncontrollable and unobservable. Because the equation is already of
the form shown in (49), it can be reduced to the following controllable
state equation
0 −0.5 0.5
   
ẋc = xc + u
1 0 0
y = [0, 0]xc + u
The output is independent of xc , thus the equation can be further reduced
to y = u. This is what we will obtain by using the MATLAB function
minreal.
W. Dong (UTRGV, Dept. of ECE) 84 / 94
Discrete-Time State Equations

Consider the n-dimensional p-input q-output state equation

x[k + 1] = Ax[k] + Bu[k] (49)

y[k] = Cx[k] (50)

where A, B, and C are, respectively, n × n, n × p, and q × n real constant

matrices.
Definition 18
The discrete-time state equation (49)-(50) or the pair (A, B) is said to be
controllable if for any initial state x(0) = x0 and any final state x1 , there
exists an input sequence of finite length that transfers x0 to x1 . Otherwise
the equation or (A, B) is said to he uncontrollable.

W. Dong (UTRGV, Dept. of ECE) 85 / 94

Discrete-Time State Equations

Theorem 19

The following statements are equivalent:

1. The n-dimensional pair (A, B) is controllable.
2. The n × n matrix
n−1
Wdc [n − 1] = Am BB 0 (A0 )m (51)
X

m=0

is nonsingular.
3. The n × np controllability matrix

Cd = [B, AB, A2 B, . . . , An−1 B] (52)

has rank n (full row rank). The matrix can be generated by calling ctrb in
MATLAB.
W. Dong (UTRGV, Dept. of ECE) 86 / 94
Discrete-Time State Equations

4. The n × (n + p) matrix [A − λI, B] has full row rank at every

eigenvalue, λ, of A.
5. If, in addition, all eigenvalues of A have magnitudes less than 1, then
the unique solution of

Wdc − AWdc A0 = BB 0 (53)

is positive definite. The solution is called the discrete controllability

Gramian and can be obtained by using the MATLAB function dgram. The
discrete Gramian can be expressed as
∞
Wdc = Am B 0 (A0 )m (54)
X

m=0

W. Dong (UTRGV, Dept. of ECE) 87 / 94

Discrete-Time State Equations

Proof: The solution of (49) at k = n is

x[n) = An x[0] + n − 1m=0 An−1−m Bu[m] (55)

X

which can be written as

u[n − 1]
 

n n−1
 u[n − 2] 
x[n] − A x[0] = [B, AB, . . . , A B]  .. (56)
 
.

 
u[0]

For any x[0] and x[n], an input sequence exists if and only if the
controllability matrix has full row rank. This shows the equivalence of
Statement 1 and Statement 3.

W. Dong (UTRGV, Dept. of ECE) 88 / 94

Discrete-Time State Equations

The matrix Wdc [n − 1] can be written as

B0
 

B 0 A0
Wdc [n − 1] = [B, AB, . . . , An−1 B] 
 
..
 
.

 
B 0 (A0 )n−1

The equivalence of Statement 2 and Statement 3 then follows Theorem ??.

Note that Wdc [m] is always positive semidefinite. If it is nonsingular or,
equivalently, positive definite, then (49) is controllable. The proof of the
equivalence of Statement 3 and Statement 4 is identical to the
continuous-time case.
Condition (5) follows Condition (2) and the Theorem in the last chapter.


W. Dong (UTRGV, Dept. of ECE) 89 / 94

Discrete-Time State Equations

Remark: There is one important difference between the continuous- and

discrete-time cases. If a continuous-time state equation is controllable, the
input can transfer any state to any other state in any nonzero time interval,
no matter how small. If a discrete-time state equation is controllable, an
input sequence of length n can transfer any state to any other state. If we
compute the controllability index µ, as defined in (10), then the transfer
can be achieved using an input sequence of length µ. If an input sequence
is shorter than µ, it is not possible to transfer any state to any other state.

W. Dong (UTRGV, Dept. of ECE) 90 / 94

Discrete-Time State Equations

Definition 20
The discrete-time state equation (49) or the pair (A, C) is said to be
observable if for any unknown initial state x[0], there exists a finite integer
k1 > 0 such that the knowledge of the input sequence u[k] and output
sequence y[k] from k = 0 to k1 suffices to determine uniquely the initial
state x[0]. Otherwise, the equation is said to be unobservable.

W. Dong (UTRGV, Dept. of ECE) 91 / 94

Discrete-Time State Equations
Theorem 21
The following statements are equivalent:

1. The n-dimensional pair (A, C) is controllable.

2. The n × n matrix
n−1
Wdo [n − 1] = (A0 )m C 0 CAm (57)
X

m=0

is nonsingular or, equivalently, positive definite.

3. The nq × n observability matrix
C
 
 CA 
Od =  .. (58)
 
.

 
CAn−1

W. Dong (UTRGV, Dept. of ECE) 92 / 94

Discrete-Time State Equations
has rank n (full column rank). The matrix can be generated by calling
obsv in MATLAB.
4. The (n + q) × n matrix
A − λI
 

C
has full column rank at every eigenvalue, λ, of A.
5. If, in addition, all eigenvalues of A have magnitudes less than 1, then
the unique solution of

Wdo − A0 Wdo A = C 0 C (59)

is positive definite. The solution is called the discrete observability

Gramian. The discrete Gramian can be expressed as
∞
Wdo = (A0 )m C 0 CAm (60)
X

m=0

W. Dong (UTRGV, Dept. of ECE) 93 / 94

Discrete-Time State Equations

• The theorem can be proved directly or indirectly using the duality

theorem.
• All other properties-such as controllability and observability indices,
Kalman decomposition, and Jordan-form controllability and observability
conditions discussed for the continuous time case apply to the
discrete-time case without any modification.
• The controllability index and observability index, however, have simple
interpretations in the discrete-time case. The controllability index is the
shortest input sequence that can transfer any state to any other state.
The observability index is the shortest input and output sequences needed
to determine the initial state uniquely.

W. Dong (UTRGV, Dept. of ECE) 94 / 94