304-501 LINEAR SYSTEMS

Lecture 22: Decomposition into Controllable and Uncontrollable Parts and

Realization Theory

5.8.4 Decomposition into Controllable and Uncontrollable Parts

For the purpose of simulating a given system's input-output behavior, we may have to look for state-
space realizations that have the same input-output behavior, but not necessarily the same state-space
matrices.

More precisely, two systems are zero-state equivalent if they have identical input-output maps
assuming initial rest. The two systems are equivalent if, for each initial state of either system, there
exists another initial state for the other such that their input-output maps are identical.

In the linear time-invariant case, equivalent systems differ only by a change of basis of the state. In fact,
it is possible to find a basis relative to which the system is decomposed into a zero-state equivalent CC
part, and a part that is completely disconnected from the input and which can be termed completely
uncontrollable.

Consider the LTI system:

x
(t ) = Ax(t ) + Bu (t )
(5.105)
y (t ) = Cx (t ) + Du (t )

Let x = Tx be a new state obtained via the similarity transformation T . Then,

x
 (t ) = TAT −1
 x (t ) + TB
 u (t )
A B
(5.106)
−1
y (t ) = CT
 x (t ) + Du (t )
C

u (t ) + x
 (t ) x (t ) y (t )
−1
B T s I T −1
C
+

T A T −1

L22- 1/9
 304-501 LINEAR SYSTEMS

We use the above form to separate the controllable part from the uncontrollable part. To find such a
−1
decomposition, we note that a change of basis mapping A into TAT via the nonsingular
 , B , C ) = (TAT
transformation T maps ( A, B , C ) into the equivalent realization ( A
−1
, TB, CT −1 ) .

Let Q be the controllability matrix, and assume rank{Q} =: q < n . Then the system is not CC.
n ×n
Choose an invertible matrix T ∈ R such that

 Q1 
TQ =: Q =   , Q1 ∈ R
q× nm
(5.107)
θ ( n − q )×nm 

e.g., T can be the row equivalent transformation that maps Q into its echelon form. The resulting
( A , B ) have the block form:

 A A12    B1  
A =  11 , =  , A11 ∈ R , A12 ∈ R
q× q q× ( n − q ) 
, A22 ∈ R ( n− q )×( n− q ) , B1 ∈ R q×m
θ ( n −q )×q A  B θ
22   ( n − q )×m 

(5.108)

If C is partitioned as

C = C1 C 2  , C1 ∈ R p×q , C 2 ∈ R p×(n −q ) . (5.109)

Then, the transformed system consists of the sum of

( A11 , B1 , C1 ) , (5.110)

 is onto by construction, and the system

which is CC as its controllability matrix Q1

 θ A12  
   ,θ , θ C 2   (5.111)
 θ A 22   


which is disconnected from the input. If the initial state of the transformed system is the zero vector,
then the subsystem of (5.111) has no effect on the output, and therefore its CC part (5.110) is zero-
state equivalent to the original system.

L22- 2/9
 304-501 LINEAR SYSTEMS

u (t ) x
1 (t ) x1 (t ) + y (t )
+ +
CC
A12 +
−1
s I C1 +
part +
+

A12 A11

x 2 ( t )
−1
s I C 2
Completely
uncontrollable part

A22

We will need the definition of an invariant subspace.

Definition: Invariant Subspace

Given:

• A vector space ( V , F )

• A subspace M ⊂ V

• A linear transformation A : V → V

Subspace M is said to be invariant under A if Am ⊂ M, ∀m ∈ M .

Lemma:

If x ∈ R {Q} , then Ax ∈ R {Q} , i.e., R {Q} is an A -invariant subspace.

Proof: Left as an exercise (use the CHT.)

L22- 3/9
 304-501 LINEAR SYSTEMS

Remarks:

(1) Matrix T can be found by adjoining an n × n identity matrix to Q and row-reducing Q [ I ] to

Q1 
obtain  T
θ 

 :=  B
(2) The controllability matrix of the CC part is Q A11 B1  A11 B 1 
n −1
(5.112)
1 1 

(3) The CC part can be written as

x
1 (t ) = A11 x1 (t ) + B11u1 (t ) + A12 x2 (t )

(5.113)
x
2 (t ) = A 22 x2 (t )

which shows that u does not affect x 2 (t ) at all.

 and the corresponding modes are called the controllable eigenvalues and
The q eigenvalues of A11

controllable modes of the pair ( A, B ) .

 and the corresponding modes are called the uncontrollable

The n − q eigenvalues of A22

eigenvalues and uncontrollable modes of the pair ( A, B ) .

In the zero-state response of the system, the controllable modes are completely absent.

In particular, in the solution

x(t ) = e At x(0) + ∫ e A(t −τ ) Bu(τ )dτ (5.114)

0

notice that

 A11 (t −τ ) B11 
e A ( t −τ )
(
B= T e−1 A ( t −τ )
)  −1 e
T T B=T  −1

θ
 (5.115)
 

which shows that the input cannot influence the uncontrollable modes.

L22- 4/9
 304-501 LINEAR SYSTEMS

Example:

 0 −1 1  1 0 
  
Let A = 1 −2 1 , B = 1 1 .

   
 0 1 −1 1 2

We wish to reduce the system to the standard form with a CC part and an uncontrollable part. Here,

1 0 0 1 0 −1
Q =  B AB A B  = 1 1 0 0 0 0 
2

1 2 0 −1 0 1 

and hence rank(Q ) = q = 2 < 3 = n .

( )
Thus, dim R {Q} = 2 and a basis {v1 , v2 } for the range is found by taking two linearly
independent columns of Q , say the first two, to obtain:

 0
T −1
= v1 v2 0 
 1 

−1
where the last column is selected to make T nonsingular. Note that with this choice we will get
Q 
T −1  1  = [v1 v2 ] Q1 = Q . We have,
0

 1 0 0  0 −1 1  1 0 0   0 1 1 
   A A12 
A = TAT −1
=  −1 1 0  1 −2 1  1 1 0  =  0 −1 0  =:  11


0 A22 
 1 −2 1  0 1 −1 1 2 1   0 0 −2 
 

 1 0 0 1 0   1 0
 B11 
B = TB =  −1 1 0  1 1  =  0 1  =:  

0
 1 −2 1  1 2  0 0  

where ( A , B ) is CC. The matrix A has three eigenvalues at 0,-1,-2, and from ( A , B ) the
11 1 11 1
eigenvalues 0,-1 are controllable, while -2 is an uncontrollable eigenvalue.

L22- 5/9
 304-501 LINEAR SYSTEMS

6 Realization Theory and Algorithms

It is easy to find the input-output relationship in the form of a convolution or a transfer function
describing the behavior of an LTI system given by an internal (state-space) description.

The inverse problem, called the realization problem, is not as straightforward:

Given an input-output description of a linear system (impulse response or transfer function), determine
a state-space model for the system that has the same input-output model.

There are infinitely many possible state-space realizations of a given system. We typically seek the
ones that have the least number of first-order differential equations.

6.1 Realizations

Consider the LTV system

x
(t ) = A(t ) x(t ) + B(t )u (t ), x(t0 ) = x0

(6.1)
y (t ) = C (t ) x (t ) + D(t )u (t )

where x(t ) ∈ R , u (t ) ∈ R , y (t ) ∈ R , and A, B , C , D are continuous over [t0 , t1 ] .

n m p

The response of the system is given by:

t
y (t ) = C (t )Φ(t , t0 ) x0 + ∫ H (t , τ )u (τ )d τ , (6.2)
t0

where H (t , τ ) is the p × m impulse response of the system, given by:

C (t )Φ (t ,τ ) B(τ ) + D (t )δ (t − τ ), t ≥ τ
H (t , τ ) =  (6.3)
 0 t <τ

In the LTI case, the impulse response of the system simplifies to:

Ce At B + Dδ (t ), t ≥ 0
H (t ) =  . (6.4)
 0, t<0

It also has a Laplace transform:

H ( s ) = C ( sI − A)−1 B + D, Re{s} > max Re{λi ( A)} . (6.5)

i =1,…, n

L22- 6/9
 304-501 LINEAR SYSTEMS

Definition: Realization of a Linear System

LTV case: A realization of H (t , τ ) is any state-space system ( A(t ), B(t ), C (t ), D(t ) ) whose
impulse response is H (t , τ ) .

Note that in general, it is not necessary for any H (t , τ ) to have a realization.

LTI case: A realization of H ( s ) is any state-space system ( A, B, C , D ) whose transfer function is

H ( s) .

Note again that in general, it is not necessary for any H ( s ) to have a realization. It can be shown that
a necessary condition for H ( s ) to have a realization is that all of its entries are proper rational
functions.

Definition: Markov Parameters

Suppose H ( s ) is given as a Laurent series:

H ( s ) = H 0 + H1 s −1 + H 2 s −2 + (6.6)

The constant matrices H i are called the Markov parameters of the system, and they can be computed
as follows:

s →∞ s →∞ s →∞
(
H 0 = lim H (s ), H1 = lim s ( H (s ) − H 0 ) , H 2 = lim s 2 H (s ) − H 0 − H1 s −1 ,… ) (6.7)

Theorem:

The matrices ( A, B , C , D ) form a realization of H ( s ) iff

H 0 = D, H i = CAi −1 B, i = 1, 2,… (6.8)

Proof:

H ( s ) = C ( sI − A)−1 B + D, Re{s} > max Re{λi ( A)}

i =1,…, n
−1 −1 −1
= Cs ( I − s A) B + D
 +∞ i 
( )
= Cs −1  ∑ s −1 A  B + D
 i =0 
+∞
= ∑ CAi −1 Bs − i + D
i =1

L22- 7/9
 304-501 LINEAR SYSTEMS

Remarks:

(a) The impulse response of a linear system contains no information about the initial conditions. This
implies that different state-space realizations of H (t , τ ) will yield the same zero-state response,
but may have different zero-input responses.

(b) If a realization exists for H (t ,τ ) , then infinitely many realizations exists (similarity
transformations.)

6.2 Existence of Realizations

Theorem:

The LTV system H (t , τ ) is realizable iff H (t , τ ) can be decomposed as

H (t ,τ ) = M (t ) N (τ ) + D (t )δ (t − τ ) . (6.9)

Proof:

(sufficiency) Assume (6.9) holds. Consider the realization: (θ , N (t ), M (t ), D (t ) ) …

(necessity) Let the system H (t , τ ) have a state-space realization. Then,

H (t ,τ ) = C (t )Φ(t ,τ ) B(τ ) + D(t )δ (t − τ ), t ≥ τ

= C (t )Φ(t, a) Φ( a, τ ) B(τ ) + D(t )δ (t − τ ), t ≥ τ
 
M (t ) N (τ )

= M (t ) N (τ ) + D(t )δ (t − τ ), t ≥ τ

Theorem:

The LTI system H ( s ) is realizable iff Theorem:

The LTV system H (t , τ ) is realizable iff H (t , τ ) can be decomposed as

is a matrix of proper rational functions.

Proof:

−1
(necessity) If the system is a realization of H ( s ) , then H ( s ) = C ( sI − A) B + D is rational, and
lim H ( s ) = D .
s →∞

(sufficiency) If H ( s ) is a proper rational transfer matrix, then we can use the controllable canonical
form discussed next to obtain a state-space realization.

L22- 8/9
 304-501 LINEAR SYSTEMS

Corollary:

H (t ) is realizable as the impulse response of an LTI state-space system iff all entries of H (t ) are
sums of terms of the form:

α t k eλt , βδ (t ), α , β , λ ∈ C, k ∈ {0,1, 2,…} .

L22- 9/9