Similarity Transformation

Dr. Manas Kumar Bera

Associate Professor
Department of EE
National Institute of Technology Rourkela

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 1 / 70

Outline

1 Introduction

2 Revisit: Basis and Representation

3 Similarity Transformation

4 Diagonalization

5 Controller Canonical Form

6 Observer Canonical Form

7 References

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 2 / 70

Introduction

We have seen that, given an implementation/ wiring diagram/ block

diagram for a differential equation, we can write down a state space
representation of the system of the form

ẋ = Ax + Bu
y = Cx + Du

Converse is also true: given a state space representation we can draw

a wring diagram.
We had defined the state variables as the outputs of the integrators in
a particular implementation.
As we saw, a particular system or differential equation may have
many different realizations. Each such realization produces its own
set of state variables. Hence “states" of a system are non-unique.
So the phrase “the states of the system" is meaningless. One should
say “the states of a realization".
Dr. Manas Kumar Bera (NITRkl) February 13, 2025 3 / 70
Similarity Transformation

We have seen that there are more than one implementation/

realization for a differential equation.
Canonical representation: Controller Canonical Form(CCF), Observer
Canonical Form(OCF), Controllability Canonical Form, Observability
Canonical Form, Diagonal Canonical Form (DCF), Jordan Canonical
Form.
Can we create different realization from a given one?
One way of doing this is by using Similarity Transformation.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 4 / 70

Revisit: Basis and Representation

A set of linearly independent vectors in Rn is called a basis if every

vector in Rn can be expressed as linear combination of the set.
In Rn any set of n linearly independent vectors can be used as a basis.
Let {v1 , v2 , · · · vn } be such set.
The every vector x can be expressed as uniquely as

x = α1 v1 + α2 v2 + · · · + αn vn ; αi ∈ R (1)

Define the n × n square matrix

h i
T := v1 v2 · · · vn (2)

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 5 / 70

Revisit: Basis and Representation (contd.

Then (1) can be written as

 
α1

 α2 

x =T ..  =: T x̄ (3)
.
 
 
αn
h i>
We call x̄ = α1 α2 · · · αn the representation of vector x
h i
with respect to the basis v1 v2 · · · vn .
From (3), one can write x̄ = T −1 x .

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 6 / 70

Example 1

We will

associate

with

every

Rn the following
 
orthonormal
 
basis:
1 0 0 0
 0   1   0   0 
       
 0   0   0   0 
       
i1 =  .  , i2 =  .  , · · · in−1 =  .  , in =  . 
      
 ..   ..   ..   .. 

       
 0   0   1   0 
0 0 0 1
Withrespect
 to this basis, we have  
x1 x1

x
 2 

 x2
 

x = .  = x1 i1 + x2 i2 + · · · + xn−1 in−1 + xn in = In  . .
 .   .
 .   .


xn xn
where In is the n × n unit matrix. In other words, the representation
of any vector x with respect to the orthonoral basis equals itself.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 7 / 70

Example 2

" #
2 1
Let A = . What is the representation of A with respect to
0 3
" # " #
n 1 0.7 o
the basis vectors v1 = , v2 = .
0 0.7
" # " #
2 2.1
Step I: Act A on v1 and v2 . So Av1 = and Av2 = .
0 2.1
Step 2: Express Av1 and Av2 in terms of basis v1 and v2 .
" #
2
2v1 + 0v2 = (4)
0
" #
2.1
0v1 + 3v2 =
2.1

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 8 / 70

Example 2 (contd.)

Hence we can write

" # " #
h i 2 0 2 2.1
v1 v2 = = AT (5)
| {z } 0 3 0 2.1
T
| {z }
Ā

From above, we can write Ā = T −1 AT .

This says that A can be represented as a diagonal matrix by choosing
proper bases.
In fact in this example it can be verified that v1 and v2 are the
eigenvectors of A and 2, 3 are the corresponding eigenvalues.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 9 / 70

Example 3

" # " # " #

2 1 1 5
Let A = ,b= and Ab = . Represent A with
0 3 3 9
respect to {b, Ab} as basis.
Let v1 = b and v2 = Ab.
" # " #
5 19
Step I: Act A on v1 and v2 . So Av1 = and Av2 = .
9 27
Step 2: Express Av1 and Av2 in terms of basis v1 and v2 .
" #
5
0v1 + 1v2 = (6)
9
" #
19
−6v1 + 5v2 =
27

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 10 / 70

Example 3 (contd.)

Hence we can write

" # " #
h i 0 −6 5 19
v1 v2 = = AT (7)
| {z } 1 5 9 27
T
| {z }
Ā

From above, we can write Ā = T −1 AT .

This says that A can be represented as a observer canonical form by
choosing proper bases.
Here T is called the controllability matrix.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 11 / 70

Similarity Transformation

Consider the n-dimensional realization

ẋ = Ax + Bu (8)
y = Cx (9)

Let T be a n × n constant real matrix which is non-singular.

FACT: Non-singular:= det(T ) 6= 0
≡ T is full rank.
≡ T is invertible.
≡ T −1 exists.
T could be constructed from the set of basis vectors.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 12 / 70

Similarity Transformation

Define a new state variable

x̄ (t) = T −1 x (t) (10)

To find out the state equations for x̄ ,

x̄˙ (t) = T −1 ẋ (t) = T −1 [Ax + Bu] (11)

= T −1 Ax + T −1 Bu
= [T −1 AT ]x̄ + [T −1 B]u

Similarly y = Cx = [CT ]x̄ .

Define the new matrices: Ā = T −1 AT , B̄ = T −1 B and C̄ = CT .
The new realization is

x̄˙ = Āx̄ + B̄u (12)

y = C̄ x̄ (13)
Dr. Manas Kumar Bera (NITRkl) February 13, 2025 13 / 70
Similarity Transformation

Note that y and u remain the same. But they are connected through
a new “wiring diagram”.
This realization is SIMILAR to the original realization and the
transformation is called a SIMILARITY TRANSFORMATION.
Check whether realizations (8) and (12) results in the same
differential equation connecting y with u.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 14 / 70

Invariance of transfer function under Similarity
Transformation

Ḡ(s) = C̄ (sI − Ā)−1 B̄

= CT (sI − T −1 AT )−1 T −1 B
= CT (sT −1 T − T −1 AT )−1 T −1 B
= CT [T −1 (sI − A)T ]−1 T −1 B
= CT T −1 (sI − A)−1 T T −1 B
= C (sI − A)−1 B = G(s)

Hence, the transfer function remains same /invariant under similarity

transformation.The eigenvalues are invariant under similarity
transformation.
The new realization still represent the same system.
Dr. Manas Kumar Bera (NITRkl) February 13, 2025 15 / 70
Diagonalization
Simplification of Realization
Can we use similarity transformation to create realizations, which are
simpler than the original?
Can we diagonalize any realization by similarity transformation?
No, in general. Let us investigate further.
This reduces to the following question: Is there a matrix T such that
Ā = T −1 AT is diagonal?
FACT 1: (From Linear Algebra): A n × n matrix A is diagonalizable
if and only if it has n linearly independent eigenvector.
Remember: The diagonalization should be over the real numbers.
FACT 2: A n × n REAL matrix A is diagonalizable OVER THE
REALS iff it has n linearly independent REAL eigenvector.
FACT 3: an easier way to check condition: (Through only sufficient
condition): A n × n real matrix A is diagonalizable over the reals if it
has n real distinct eigenvalues.
Dr. Manas Kumar Bera (NITRkl) February 13, 2025 16 / 70
Diagonalization: Building T matrix

Assume
h A has n linearly
i independent eigenvectors
v1 v2 · · · vn .
From the definition of eigenvectors Avi = λi vi , for i = 1, 2, · · · , n,
where λi ’s are the corresponding eigenvalues.
h i
Build a matrix Tn×n with v1 v2 · · · vn as its columns.
h i
T = v1 v2 · · · vn (14)
n×n
h i
Since v1 v2 · · · vn are linearly independent, T is invertible.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 17 / 70

Diagonalization: Building T matrix (contd.)

h i
AT = A v1 v2 · · · vn
h i
= Av1 Av2 · · · Avn
h i
= λ1 v1 λ2 v2 · · · λn vn
 
λ1 0 · · · 0
h i
 0 λ2 · · · 0 

= v1 v2 · · · vn  .. .. . . .. 

 . . . .


0 0 ··· λn
= T Ā
 
λ1 0 · · · 0

0 λ2 · · · 0 
Ā = T −1 AT = 
 
.. .. . . .. 

 . . . .


0 0 ··· λn
Dr. Manas Kumar Bera (NITRkl) February 13, 2025 18 / 70
Diagonalization: example

Consider the following system

 
1 0 0
ẋ =  1 2 0  x
 
1 0 −1

Diagonalize the above system by constructing suitable T matrix. Note:

Diagonal form is also known as decoupled form.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 19 / 70

Diagonalization: example (contd.)

Step I: Compute eigenvalues: λ1 = 1, λ2 = 2, λ3 = −1

Step II: Compute eigenvector to the corresponding eigenvalues; Use
formula Avi = λi vi .
To find the eigenvector for the eigenvalue 1, we have
    
1 0 0 v11 v11
1 2 0   12   v12 
v =
    

1 0 −1 v13 v13
 
2
so that  −2  is an eigenvector.
 
1
The eigenvector corresponding to the eigenvalue 2 is given by
h i> h i>
0 1 0 , while that of −1 is 0 0 1 .

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 20 / 70

Diagonalization: example (contd.)

The diagonalizing matrix T is then

 
2 0 0
T =  −2 1 0 
 
1 0 1

Step III: Compute T −1 .

 
1
2 0 0
T −1 =  1 1 0 
 
− 12 0 1

Compute T −1 AT .
 
1 0 0
Ā = T −1 AT = 0 2 0 
 
0 0 −1
Dr. Manas Kumar Bera (NITRkl) February 13, 2025 21 / 70
Diagonalization: If the matrix A is of the CCF and A has
distinct eigenvalues.

If the matrix An×n is of the CCF (controller canonical form) and A has
distinct eigenvalues, then the DCF (diagonal canonical form)
transformation matrix is the Vandermonde matrix,
 
1 1 1 ··· 1

 λ1 λ2 λ3 ··· λn 

λ21 λ22 λ23 ··· λ2n
 
T = 
 .. .. .. .. .. 
. . . . .
 
 
λ1n−1 λn−1
2 λn−1
3 ··· λnn−1

where, where λ1 , λ2 , λ3 , · · · , λn are the eigenvalues of A.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 22 / 70

Example

Diagonalize the following system

 
0 1 0
ẋ =  0 0 1 x
 
−6 −11 −6

The eigenvalues are λ1 = −1, λ2 = −2, λ3 = −3

Because A is CCF, to transform it into DCF, the transformation
matrix can be written directly as
   
1 1 1 1 1 1
T =  λ1 λ2 λ3  =  −1 −2 −3 
   
λ21 λ22 λ23 1 4 9

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 23 / 70

Example(contd.)

Thus, the DCF of A is written as

 
−1 0 0
−1
Ā = T AT =  0 −2 0 
 
0 0 −3

Observations:
Can you compute e At of this decoupled/ diagonal form?
Can you comment about the stability of the system?
Can you comment about controllability, observability properties of the
system?

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 24 / 70

Easier way to calculate e At .
Often it is time consuming to compute e At for general A.
IDEA: If Ā were diagonal, then it would be easier to compute e Āt
than it would e At .
Example:
 
1 0 0
Ā =  0 2 0 
 
0 0 −1
e Āt = L−1 ((sI − A)−1 )
1
 
s−1 0 0
= L−1  0 1
0 
 
s−2
1
0 0 s+1
 
et 0 0
∴ e Āt = 0 e 2t 0 
 
0 0 e −t

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 25 / 70

How e Āt and e At is related?

Consider what happens if the matrix A is diagonalizable, i.e. there exists a

T such that
λ1
 
−1
T AT = Ā which is diagonal, Ā = 
 .. .

.
λn
e λ1 t
 

Then e Āt = 
 .. .

.
e λn t

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 26 / 70

How e Āt and e At is related?

Now, A = T ĀT −1 and following can be written

1
e At = I + At + (At)2 + · · ·
2!
1
= I + T ĀT −1 t + (T ĀT −1 )2 + · · ·
2!
1
= TT −1 + T ĀT −1 t + T ĀT −1 T ĀT −1 t 2 + · · ·
2!
1
= T (I + Āt + (Āt) + · · · )T −1
2
2!
Āt −1
= Te T

Important to remember
e At = Te Āt T −1

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 27 / 70

Contd..

If A is diagonalizable, then e At may be computed from the above

formula.
Or, just compute
Z t
x̄ (t) = e Āt x0 + e Ā(t−τ ) Bu(τ )dτ
0

and then compute x (t) = T x̄ (t)

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 28 / 70

Example: Compute e At using modal
decomposition/diagonalization

Compute e At using modal decomposition/diagonalization for the following

system
 
1 0 0
ẋ =  1 2 0  x
 
1 0 −1

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 29 / 70

Solution

Step I: Compute the diagonalizing matrix T

 
2 0 0
T =  −2 1 0 
 
1 0 1

Step II: Compute T −1 .

 
1
2 0 0
T −1 =  1 1 0 
 
− 12 0 1

Step III: Compute Ā = T −1 AT .

 
1 0 0
Ā = T −1 AT = 0 2 0 
 
0 0 −1
Dr. Manas Kumar Bera (NITRkl) February 13, 2025 30 / 70
Solution (contd.)

Step IV: Compute e Āt

 
et 0 0
e Āt = 0 e 2t 0 
 
0 0 e −t

Step V: Compute e At = Te Āt T −1 .

   
1
2 0 0 et 0 0 0 0
 2
e At =  −2 1 0   0 e 2t 0  1 1 0 
  
1 0 1 0 0 e −t − 21 0 1
 
et 0 0
t 2t e 2t
=  −e + e 0 
 
1 t −t e −t
2 (e − e ) 0

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 31 / 70

Remember!

We have learnt three ways to calculate e At

Laplace Transform Technique
Using C-H Theorem
Using modal decomposition/ diagonalization

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 32 / 70

Example: Advantage of Decoupled Representation

Consider that a system is transformed in the following diagonal form

      
x̄˙ 1 λ1 0 0 x̄1 b̄1
 ˙  
 x̄2  =  0 λ2 0   x̄2  +  b̄2  u
   
x̄˙ 3 0 0 λ3 x̄3 b̄3
h i
y= c̄1 c̄2 c̄3 x̄

We can write

x̄˙ 1 = λ1 x̄1 + b̄1 u

x̄˙ 2 = λ2 x̄2 + b̄2 u
x̄˙ 3 = λ3 x̄3 + b̄3 u
y = c̄1 x̄1 + c̄2 x̄2 + c̄3 x̄3

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 33 / 70

Contd..

Each x̄i (t) is called a modal response. For x̄i (t), the solution is
Z t
e λi t x̄i (0) +
x̄i (t) = |{z} e λi (t−τ ) bi u(τ )dτ
stability? |0 {z }
can you control a particular mode?

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 34 / 70

Modal Decomposition

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 35 / 70

Additional Reading: Modal Decomposition

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 36 / 70

Additional Reading: Modal Decomposition

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 37 / 70

Modal Decomposition

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 38 / 70

Modal Decomposition

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 39 / 70

Example

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 40 / 70

Solution

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 41 / 70

Solution

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 42 / 70

Solution

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 43 / 70

Solution

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 44 / 70

Orthogonal and orthonormal vectors

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 45 / 70

Orthogonal and orthonormal vectors

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 46 / 70

Additional Reading: Diagonalization with Repeated
Eigenvalue

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 47 / 70

Additional Reading: Diagonalization with Repeated
Eigenvalue

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 48 / 70

Additional Reading: Diagonalization with Repeated
Eigenvalue

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 49 / 70

Additional Reading: Diagonalization with Repeated
Eigenvalue

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 50 / 70

Additional Reading: Diagonalization with Repeated
Eigenvalue

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 51 / 70

Additional Reading: Diagonalization with Repeated
Eigenvalue

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 52 / 70

Controller canonical Form

Problem Statement
Given a realization

ẋ = Ax + Bu
y = Cx

Can you build a T matrix such that the above realization can be transform
to controller canonical form?

x̄˙ = Āx̄ + B̄u

y = C̄ x

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 53 / 70

Contd..

 
0 1 0 ··· 0

 0 0 1 ··· 0 

−1

where Ā = T AT =  .. .. .. .. .. 
,
 . . . . . 
0 0 0 0 1
 
 
−an −an−1 −an−2 · · · −a1
 
0

 0 

B̄ = T −1 B =  0
 
 and C̄ = CT (do not follow any particular pattern)

.. 
.
 
 
1
and the characteristics equation of the system is
det(sI − A) = s n + a1 s n−1 + a2 s n−2 + · · · + an−1 s + an = 0.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 54 / 70

Building T matrix

Step I: Build T = CM.

Step hII: Construct controllability matrix
i
C = B AB A2 B · · · An−1 B . Note that the T matrix will
n×n
exists iff C is full rank.
Step III:
 Construct M matrix as 
an−1 an−2 · · · a1 1
 n−2 an−3 · · ·
 a 1 0 

 . . . . . . ...

M =  ..
 .. .. 

a1 1 ··· 0 0 
 

1 0 ··· 0 0 n×n
IV: Calculate Ā = T −1 AT , B̄ = T −1 B and C̄ = CT .

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 55 / 70

Example

Consider the system

ẋ = Ax + Bu
y = Cx
   
1 2 1 1 h i
with A =  0 1 3 , B =  0  and C = 1 1 0 . The realization
   
1 1 1 1
is to be transformed to CCF.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 56 / 70

Solution

Step-I: The characteristic equation of A is

det(sI − A) = s 3 − 3s 2 − s − 3 = 0
⇒s 3 + a1 s 2 + a2 s + a3 = 0

Step-II: Find the controllability matrix C and test det(C) 6= 0.

 
h i 1 2 10
C= B AB A2 B = 0 3 9 
 
1 2 7

We can show that C is nonsingular, so the system can be transformed

into the CCF. step-III: Construct M matrix.
   
a2 a 1 1 −1 −3 1
M =  a1 1 0  =  −3 1 0 
   
1 0 0 1 0 0
Dr. Manas Kumar Bera (NITRkl) February 13, 2025 57 / 70
Contd..

Step-IV: Compute T = CM.

 
3 −1 1
T = 0 3 0 
 
0 −1 1

Step-V: Calculate Ā = T −1 AT , B̄ = T −1 B and C̄ = CT .

 
0 1 0
Ā =  0 0 1 
 
3 1 3
 
0
B̄ =  0 
 
1
h i
C̄ = 3 2 1

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 58 / 70

FACT

If (A, B) is controllable there exists a similarity transformation T such that

(T −1 AT , T −1 B) is in controllable canonical form.

Proof:
Consider a third order system. The characteristics equation of the
system is s 3 + a1 s 2 + a2 s + a3 = 0.
h i
If the system is controllable, C = B AB A2 B is full rank
3×3
(rank 3).
Consider the matrix,
 
h i a2 a1 1
T = CM = B AB A2 B  a1 1 0 
 
1 0 0
h i
= v1 v2 v3

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 59 / 70

FACT
The second matrix (called a Toeplitz matrix) forming T is
non-singular. The first matrix forming T is C. Thus, controllability
ensures T −1 exists, so that its columns v1 , v2 , v3 form a basis of Rn .
Calculating vectors v1 , v2 , v3

v1 = A2 B + a1 AB + a2 B (15)
v2 = AB + a1 B
v3 = B

Now

Av1 = A3 B + a1 A2 B + a2 AB + a3 B − a3 B (16)
3 2
= (A + a1 A + a2 A + a3 I)B − a3 B
= −a3 B by the Cayley-Hamilton Theorem
= −a3 v3

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 60 / 70

FACT

Now,

Av2 = A2 B + a1 AB = A2 B + a1 AB + a2 B − a2 B
= v1 − a2 v3

Now

Av3 = AB + a1 B − a1 B = v2 − a1 v3

Writing together:

Av1 = −a3 v3
Av2 = v1 − a2 v3 (17)
Av3 = v2 − a1 v3

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 61 / 70

FACT
Thus the matrix representation of A with respect to the basis
{v1 , v2 , v3 } looks like
Using (17), we can write:
 
h i 0 1 0
v1 v2 v3  0 0 1  = AT (18)
 
−a3 −a2 −a1
 
0 1 0
Ā = T −1 AT =  0 0 1 
 
−a3 −a2 −a1
Similarly, the vector B looks like
 
0
B̄ =  0 
 
1
Extend this idea for n-th order system.
Dr. Manas Kumar Bera (NITRkl) February 13, 2025 62 / 70
Observer Canonical Form

Problem Statement
Given a realization

ẋ = Ax + Bu
y = Cx

Can you build a T matrix such that the above realization can be transform
to observer canonical form?

x̄˙ = Āx̄ + B̄u

y = C̄ x

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 63 / 70

Contd..

 
0 0 ··· 0 −an

 1 0 ··· 0 −an−1 

where Ā = T −1 AT =  0 1 0 0 −an−2 , B̄ = T −1 B = not
 

.. .. . . .. .. 
. . . . .
 
 
0 0 ··· 1 −a1
h i
restricted to any form and C̄ = CT = 0 0 · · · 0 1 and the
characteristics equation of the system is
det(sI − A) = s n + a1 s n−1 + a2 s n−2 + · · · + an−1 s + an = 0.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 64 / 70

Building T matrix

Step I: Build T = (MO)−1 .

 
C

 CA 

CA2
 
Step II: Construct observability matrix O =   . Note

.. 
.
 
 
CAn−1 n×n
that the T matrix will exists iff O is full rank.
Step III:
 Construct M matrix as 
an−1 an−2 · · · a1 1
 n−2 an−3 · · ·
 a 1 0 

 . . . . . . ...

M =  ..
 .. .. 

a1 1 ··· 0 0 
 

1 0 ··· 0 0 n×n

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 65 / 70

Example

Consider the system

ẋ = Ax + Bu
y = Cx
   
1 2 1 1 h i
with A =  0 1 3 , B =  0  and C = 1 1 0 . The state
   
1 1 1 1
equation is to be transformed to OCF.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 66 / 70

Solution

Step-I: The characteristic equation of A is

det(sI − A) = s 3 − 3s 2 − s − 3 = 0
⇒s 3 + a1 s 2 + a2 s + a3 = 0

Step-II: Find the observability matrix O and test det(O) 6= 0.

   
C 1 1 0
O =  CA  =  1 3 4 
   
CA2 5 9 14

We can show that O is nonsingular, so the system can be transformed

into the OCF. step-III: Construct M matrix.
   
a2 a 1 1 −1 −3 1
M =  a1 1 0  =  −3 1 0 
   
1 0 0 1 0 0
Dr. Manas Kumar Bera (NITRkl) February 13, 2025 67 / 70
Contd..

Step-IV: Compute T = (MO)−1 .

 
0.3333 −0.1667 0.3333
T =  −0.3333 0.1667 0.6667 
 
0.1667 0.1667 0.1667

Step-V: Calculate Ā = T −1 AT , B̄ = T −1 B, C̄ = CT
 
0 0 3
Ā =  1 0 1 
 
0 1 3
 
3
B̄ =  2 
 
1
h i
C̄ = 0 0 1

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 68 / 70

References I

[1] Farid Golnaraghi and Benjamin C. Kuo, Automatic Control Systems,

John Wiley & Sons, Inc.

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 69 / 70

Thank you!!

Dr. Manas Kumar Bera (NITRkl) February 13, 2025 70 / 70