Linear System Theory

and Design
主讲：朱芳来
Part 2
 Chapter 4 State-Space Solutions
and Realizations
4.1 Introduction
™ We know in that linear systems can be
described by convolutions and, if lumped, by
state-space equations.
™ The main purpose of this chapter is making
quantitative analyses for linear system. That is
to discuss how to find the solutions for a linear
system.
™ We mainly discuss the solutions for systems
which are described by state-space
equations.
4.2 Solution of LTI State Equation
™ Consider the linear time-invariant (LTI)
state-space equation
x& (t ) = Ax(t ) + Bu (t ) (4.2)
y (t ) = Cx (t ) + Du (t ) (4.3)
where A ∈ R n×n , B ∈ R n× p , C ∈ R q×n and D ∈ R q× p are
constant matrices.
™ The purpose of this section is to find the
solution excited by the initial state x(0) and
the input u(t).
™ To do this, we need the property of

d At
e = Ae A t = e A t A
dt

Premultiplying on both sides of (4.2) yields

e − At x& (t ) − e − At Ax(t ) = e − At Bu(t )

which implies
 d − At
(e x(t )) = e − At Bu(t )
dt

Integrating the above equation from 0 to t

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

Multiplying the above equation by e At from

right side, we have
t
x(t ) = e x(0) + ∫ e A (t −τ ) Bu(τ )dτ
At
(4.5)
0
 This is the solution of (4.2). Substituting (4.5)
into (4.3) yields the solution of (4.3) as
t
y (t ) = Ce x(0) + C ∫ e A ( t −τ ) Bu (τ )dτ + Du(t )
At
(4.7)
0

Obviously, to compute (4.5) or (4.7), we

need compute the e At first.
™ The computing methods of e
At

1. Using Theorem 3.5: we are given f (λ ) and

an n × n matrix A with characteristic
polynomial
m
α (λ ) = ∏ (λ − λ i ) n
i

i =1
 m
Where n = ∑ ni . Define
i =1

h(λ ) = β 0 + β 1λ + L + β n −1λn −1

The n unknown coefficients are determined by

f (l )
(λi ) = h ( l ) (λi ) for l = 0,1, L , ni − 1 and i = 1,2, L , m

here we say that h is equal to f on the spectrum

of A,
Then we have f(A)=h(A)
If we set f (λ ) = e λt , then e At can be computed as
e At = f ( A) = h( A)
™ Example 3.7 Compute A100 with
⎛0 1 ⎞
A = ⎜⎜ ⎟⎟
⎝ −1 − 2⎠
In order to compute A100, we set f (λ ) = λ100 . The
characteristic polynomial of A is
α (λ ) = (λ + 1) 2
Let
h (λ ) = β 0 + β 1 λ
⎧ f (−1) = h(−1)
On the spectrum of A, we have ⎨
⎩ f ′(−1) = h′(−1)
That is
⎧(−1)100 = β 0 − β1
⎨
⎩ 100 ⋅ ( −1) 99
= β1

So we obtain β1 = −100 and β 0 = −99.

That is
h(λ ) = −99 − 100λ

At this time, the A100 is computed as

⎛ − 199 − 100 ⎞
100
A = f ( A) = h( A) = β 0 I + β1 A = −99 A − 100 A = ⎜⎜ ⎟⎟
⎝ 100 101 ⎠
™ Example 3.8 Let
⎛ 0 0 − 2⎞
⎜ ⎟
A1 = ⎜ 0 1 0 ⎟
⎜1 0 3 ⎟
⎝ ⎠

Compute e A t
1

The characteristic polynomial of A is

α (λ ) = (λ − 1) 2 (λ − 2)
Let
h (λ ) = β 0 + β 1 λ + β 2 λ 2
then
⎧ f (1) = h(1) : e t = β 0 + β1 + β 2
⎪
⎨ f ′(1) = h′(1) : te t = β1 + 2 β 2
⎪ f ( 2) = h ( 2) : e 2t = β 0 + 2 β1 + 4 β 2
⎩

⎧β 0 = −2te t + e 2t
⎪
Solve above equations, we get ⎨β 1 = 3te + 2e − 2e
t t 2t

⎪β = e 2t − e t − te t
⎩ 2

⎛ 2e t − e 2 t 0 2e t − 2e 2 t ⎞
⎜ ⎟
Thus e A1t = h( A1 ) = ⎜ 0 et 0 ⎟
⎜ e 2t − e t 0 2e 2t − e t ⎟⎠
⎝
 −1
2. Using Jordan form fo A: A = QAˆ Q , where  is
in Jordan form of
⎛ λ1 1 ⎞
⎜ ⎟
⎜ λ1 1 ⎟
Aˆ = ⎜ O O ⎟
⎜ ⎟
⎜ λ`1 1 ⎟
⎜ λ1 ⎟⎠ m×m
⎝
then we have
⎛ e λ1t te λ1t t 2 e λ1t / 2! L t m −1e λ1t /(m − 1)! ⎞
⎜ ⎟
⎜ e λ1t te λ1t m − 2 λ1t
L t e /(m − 2)!⎟
A = Q⎜ e λ1t m −3 λ1t
L t e /(m − 3)!⎟
⎟Q −1
⎜ (3.46)
⎜ O M ⎟
⎜ λ1t ⎟
⎝ e ⎠
™ Example 3.11 Consider
⎛ λ1 1 ⎞
⎜ ⎟
⎜ λ1 1 ⎟
A=⎜ λ1 ⎟
⎜ ⎟
⎜ λ2 1⎟
⎜ λ 2 ⎟⎠
⎝
By (3.46)
⎛ e λ1t te λ1t t 2 e λ1t / 2! 0 0 ⎞
⎜ ⎟
⎜ 0 e λ1t te λ1t 0 0 ⎟
e At = ⎜ 0 0 e λ1t 0 0 ⎟
⎟
⎜
⎜ 0 0 0 e λ2 t te λ 2t ⎟
⎜ λ2 t ⎟
⎝ 0 0 0 0 e ⎠
 [
3. Using e At = L−1 ( sI − A) −1 ]
™ The computation of the inverse of (sI-A)
z Taking the inverse of (sI-A) by definition
z Using Theorem 3.5
z Using ( sI − A) −1 = Q ( sI − Aˆ ) −1 Q −1 and (3.49)
z Using the infinite power series in (3.57)
∞
( sI − A) −1
=s −1
∑ (
k =0
s −1
A) k
(3.57)

z Using the Leverrier algorithm discussed in Problem

3.26
™ Example 4.1 Compute (sI-A)-1, where
⎡0 − 1 ⎤
A=⎢ ⎥
⎣1 − 2 ⎦
Method 1: By The definition of the inverse
of a matrix, we have
−1 adj( sI − A) 1 ⎡ s + 2 − 1⎤
( sI − A) = = 2
det( sI − A) s + 2 s + 1 ⎢⎣ 1 s ⎥⎦
⎡( s + 2) /( s + 1) 2 − 1 /( s + 1) 2 ⎤
=⎢ 2 ⎥
⎣ 1 /( s + 1) 2
s /( s + 1 ) ⎦
 Method 2: The eigenvalue of A are -1, -1.
Let h ( λ ) = β 0 + β 1 λ and f ( λ ) = ( s − λ ) −1
. From
⎧ f (−1) = h(−1) : ( s + 1) −1 = β 0 − β1
⎨
⎩ f ′( −1) = h ′( −1) : ( s + 2) −1
= β1

we obtain
⎧β 0 = ( s + 1) −1 + ( s + 1) −2
⎨
β
⎩ 1 = ( s + 1) −2

Thus we have
h(λ ) = [( s + 1) −1 + ( s + 1) −2 ] + ( s + 1) −2 λ
 and
( sI − A) −1 = h( A) = [( s + 1) −1 + ( s + 1) −2 ]I + ( s + 1) −2 A
⎡( s + 2) /( s + 1) 2 − 1 /( s + 1) 2 ⎤
=⎢ 2 ⎥
⎣ 1 /( s + 1) 2
s /( s + 1) ⎦

™ Example 4.2 Consider the equation

⎡0 − 1 ⎤ ⎡0 ⎤
x& (t ) = ⎢ ⎥ x(t ) + ⎢ ⎥u (t )
⎣1 − 2⎦ ⎣1⎦
Its solution is
t
x(t ) = e x(0) + ∫ e A(t −τ ) Bu (τ )dτ
At
0
 The matrix function eAt can be computed as
⎡ s+2 −1 ⎤
−1 −1
⎢
−1 ( s + 1)
2
( s + 1) 2 ⎥ ⎡(1 + t )e
−t
− te −t ⎤
e = L [( sI − A) ] = L ⎢
At
⎥=
⎢ 1 s ⎥ ⎢⎣ te −t −t ⎥
(1 − t )e ⎦
⎢⎣ ( s + 1) 2 ( s + 1) 2 ⎥⎦

So we have

⎡ − t (t − τ )e −(t −τ ) u (τ )dτ ⎤
⎡(1 + t )e − te ⎤ ⎢ ∫0
−t −t
x(t ) = ⎢ + ⎥
−t ⎥ x ( 0)
⎢ [1 − (t − τ )]e −(t −τ ) u (τ )dτ ⎥
(1 − t )e −t ⎦
t

⎣ ∫0
⎣ te
⎦
™ Thegeneral property of the zero-input
response eAtx(0)
Consider a simple form of eAt
⎡e λ1t te λ1t t 2 e λ1t / 2 0 0 ⎤
⎢ ⎥
⎢ 0 e λ1t te λ1t 0 0 ⎥
e At = Q⎢ 0 0 e λ1t 0 0 ⎥Q −1
⎢ ⎥
⎢ 0 0 0 e λ1t 0 ⎥
⎢ 0 e λ2t ⎥⎦
⎣ 0 0 0

So the zero-input response is a linear

combination of terms {e λ t ,te λ t , t 2 e λ t , e λ t }
1 1 1 2
™ Conclusions for zero-input response
z Ifevery eigenvalue, simple or repeated,
of A has a negative real part, then every
zero-input response will approach zero
as t → ∞
z IfA has an eigenvalue, simple or
repeated, with a positive real part, then
most zero-input response will grow
unbounded as t → ∞
z IfA has some eigenvalues with zero real
part and all with index 1 and the remaining
eigenvalues all have negative parts, then
no zero-input response will grow
unbounded.
4.2.1 Discretization
™ Consider the continuous-time state equation
x& (t ) = Ax (t ) + Bu (t ) (4.9)
y (t ) = Cx (t ) + Du (t ) (4.10)
Because
x(t + T ) − x (t )
x& (t ) = lim
T →∞ T
If T is very trivial, we get the approximating
equation
x(t + T ) − x (t )
x& (t ) ≈
T
So we can compute (4.9) approximately as
x (t + T ) = x (t ) + Ax (t )T + Bu (t )T (4.11)
If we compute x(t) and y(t) only at t=kT for
k=0,1,2,…, Equation (4.11) and (4.10)
become
x((k + 1)T ) = ( I + TA) x(kT ) + TBu ( kT )
y ( kT ) = Cx ( kT ) + Du (kT )

The above equations are easily computed

by computer but yields the least accurate
results.
™ Another discretization method
Let
u (t ) = u ( kT ) =: u[ k ], for kT ≤ t < ( k + 1)T (4.12)
for k=0,1,2,…. Computing (4.5) at t=kT and
t=(k+1)T yields
kT
x[ k ] := x (kT ) = e AkT
x (0) + ∫ e A( kT −τ ) Bu (τ ) dτ
0

and
( k +1)T
x[ k + 1] := x((k + 1)T ) = e A ( k +1)T
x ( 0) + ∫ e A(( k +1)T −τ ) Bu (τ )dτ (4.14)
0
 For (4.14), we can rewrite it as
kT ( k +1)T
x[k + 1] := e A( k +1)T
x(0) + ∫ e A (( k +1)T −τ )
Bu (τ )dτ + ∫ e A(( k +1)T −τ ) Bu (τ )dτ
0 kT
kT ( k +1)T
= e [e
AT AkT
x(0) + ∫ e A ( kT −τ )
Bu (τ )dτ ] + ∫ e A( kT +T −τ ) Bu (τ )dτ
0 kT

By substituting (4.12) and (4.13 ) and

introducing the new variable α := kT + T − τ , we
have
x[k ] + ⎜ ∫ e Aα dα ⎞⎟ Bu[k ]
⎛ T
x[k + 1] = e AT
⎝ 0 ⎠
Thus we obtain the discrete-time state-
space equation form of (4.9) and (4.10) as
follows

x[k + 1] = Ad x[k ] + Bd u[k ] (4.15)

y[k ] = C d x[k ] + Dd u[k ] (4.16)
where
, Bd = ⎜ ∫ e Aτ dτ ⎞⎟ B, C d = C and Dd = D
⎛ T
Ad = e AT
(4.17)
⎝ 0 ⎠
If A is nonsingular, the integral in Bd is
∞
τk ∞
1 T k
(∑ A )dτ = ∑ A
T T
∫ e dτ = ∫ ∫
Aτ k k
τ dτ
0 0
k =0 k! k =0 k! 0

∞
T k +1 k ∞
T k +1 k +1
=∑ A =A ∑
−1
A
k = 0 ( k + 1)! k = 0 ( k + 1)!
∞
T k
⎛ ∞
T k
⎞
= A ∑ A = A ⎜⎜ ∑ A − I ⎟⎟ = A−1 (e AT − I )
−1 k −1 k

k =1 k! ⎝ k = 0 k! ⎠

So we have
Bd = A −1 (e AT − I ) B (4.18)
4.2.2 Solution of Discrete-Time Equations
™ Consider the discrete-time state-space
equation

x[k + 1] = Ax[k ] + Bu[k ]

k = 0,1,2, L (4.19)
y[k ] = Cx[k ] + Du[k ]

The solution of above Discrete-Time

Equations is k −1
x[ k ] = A k x[0] + ∑ A k −1− m Bu[ m] (4.20)
m =0
k −1
y[k ] = CA k x[0] + ∑ CA k −1− m Bu[m] + Du[k ] (4.21)
m =0
™ Thegeneral property of the zero-input
response Akx[0].
Suppose
⎛ λ1 1 ⎞
⎜ ⎟
⎜ λ1 1 ⎟
A = Q⎜ λ1 ⎟Q −1
⎜ ⎟
⎜ λ1 ⎟
⎜ λ2 ⎟⎠
⎝

Then we have
 ⎛ λ1k kλ1k −1 k (k − 1)λ1k − 2 0 0 ⎞
⎜ ⎟
⎜0 λ1k kλ1k −1 0 0 ⎟
A k = Q⎜ 0 0 λ1k 0 0 ⎟Q −1
⎜ ⎟
⎜0 0 0 λ1k 0 ⎟
⎜ k ⎟
⎝0 0 0 0 λ2 ⎠

which implies that every entry of the zero-

input response is a linear combination of
{λ1k ,kλ1k −1 ,k 2 λ1k − 2 ,λk2 }
™ Zero-input response conclusions for
discrete-time system
z If every eigenvalue, simple or repeated, of A
has magnitude less than 1, then every zero-
input response will approach zero as k
approaches infinity.
z If A has an eigenvalue, simple or repeated,
with magnitude larger than 1, then most zero-
input response will grow unbounded as k
approaches infinity.
z IfA has some eigenvalues with magnitude 1 and
all with index 1 and the remaining eigenvalues all
have magnitudes less than 1, then no zero-input
response will grow unbounded.
4.3 Equivalent State Equations
™ An example for understanding the concept
of equivalent state equation.
™ Example 4.3 Consider the network shown in
Fig. 4.1.
™ Selecting

inductor current-----state x1
capacitor voltage-----state x2
So, we have
The voltage across the inductor---- x&1

The current through the capacitor---- x& 2

The current through the resistor---- x2 / 1 = x 2

™ Clearly we have

⎧ x1 = x 2 + x& 2
⎨
⎩ x&1 + x 2 − u

™A state equation of the network

⎡ x&1 ⎤ ⎡0 − 1⎤ ⎡ x1 ⎤ ⎡0⎤
⎢ x& ⎥ = ⎢1 − 1⎥ ⎢ x ⎥ + ⎢1⎥u
⎣ 2⎦ ⎣ ⎦⎣ 2 ⎦ ⎣ ⎦
⎡ x1 ⎤
(4.22)
y = [0 1] ⎢ ⎥
⎣ x2 ⎦
™ Another selecting of state variables
The loop current of the left side-----state x1
The loop current of the right side-----state x2

Then
The voltage across the inductor---- x&1
The voltage across the resistor---- x1 ×1 − x2 ×1 = x1 − x2

From the left side loop, we have

u = x&1 + x1 − x2

or x&1 = − x1 + x 2 + u
 From
The voltage across the capacitor=that across the resistor

we obtain
x& 2 = x&1 − x 2 = − x1 + u

™ Another state equation for the network

⎡ x&1 ⎤ ⎡− 1 1⎤ ⎡ x1 ⎤ ⎡1⎤
⎢& ⎥ = ⎢ ⎥ ⎢ ⎥ + ⎢ ⎥u
⎣ x 2 ⎦ ⎣− 1 0⎦ ⎣ x 2 ⎦ ⎣1⎦ (4.23)
⎡ x1 ⎤
y = [1 − 1]⎢ ⎥
⎣ x2 ⎦
™ The concept of equivalent state equations
Consider the n-dimensional state equation

x& (t ) = Ax(t ) + Bu (t )
(4.24)
y (t ) = Cx(t ) + Du (t )

where x(t ) ∈ R n×n , u(t ) ∈ R m , y (t ) ∈ R p

Definition 4.1 Let P be an n × n real nonsingular
matrix and let x = Px . Then the state equation,
x& (t ) = A x (t ) + B u (t )
(4.25)
y (t ) = C x (t ) + D u (t )
where
A = PAP −1 , B = PB, C = CP −1 and D = D (4.26)
is said to be (algebraically) equivalent to (4.24) and
x = Px is called and equivalence transformation.
™ Result: The equivalence transformation
doesn’t change the eigenvalues and the
transfer function matrix of the system.
Proof
Δ (λ ) = det(λI − A ) = det(λI − PAP −1 ) = det P(λI − A) P −1
= (det P) ⋅ det(λI − A) ⋅ (det P −1 ) = det(λI − A) = Δ(λ )

and
Gˆ ( s ) = C ( sI − A ) B + D = CP −1 ( sI − PAP −1 ) PB + D
= C ( sI − A) B + D = Gˆ ( s )
™ Definition:Two state equations are said to be
zero-state equivalent if they have the same
transfer function matrix.
™ Theorem 4.1 Two linear time-invariant
equations {A, B, C, D} and { A , B , C , D } are zero-
state equivalent if and only if D = D and
CA m B = C A m B , m = 0,1,2, L

Proof: Since
 ∞ ∞
( sI − A) −1 = s −1 ( I − As −1 ) −1 = s −1 ∑ ( As −1 ) m = ∑ A m s − m −1
m =0 k =0

So we have
∞
D + C ( sI − A) −1 B = D + ∑ CA m Bs − m −1
m =0
Similarly, we still have
∞
D + C ( sI − A ) −1 B = D + ∑ C A m B s − m −1
m =0
So they are zero-state equivalent if and only if
D + C ( sI − A) −1 B = D + C ( sI − A ) −1 B

And this is equivalent to

∞ ∞
D + ∑ CA Bs m − m −1
= D + ∑ C A m B s − m −1
m =0 m=0
 ⇔ D = D and CA m B = C A m B, m = 0,1,2,L

™ It should be noted that the algebraic

equivalence implies zero-state equivalence,
but the inverse conclusion is not true.
4.4 Realizations
™ Every linear time-invariant system can be
described by input-output description
yˆ ( s ) = Gˆ ( s )uˆ ( s )
and, if the system is lumped as well, it has
state-space equation description
x& (t ) = Ax(t ) + Bu (t ) (4.29)
y (t ) = Cx (t ) + Du (t )

If the state equation is known, the transfer

function matrix can be computed by
Gˆ ( s ) = C ( sI − A) −1 B + D (4.29’)
™ Now we study the converse problem, that is,
to find a state-space equation from a given
transfer function matrix and this is called the
realization problem.
™ Definition: A transfer function matrix Gˆ ( s ) is
said to be realizable if there exists a finite-
dimensional state-equation (4.29) such that
(4.29’) holds. And the (4.29) is called a
realization of Gˆ ( s ) .
™ Remarks: Not every Gˆ ( s ) is realizable. For
example, an LTI distributed system has a
transfer function matrix, but it can not be
described by a state-equation. If Gˆ ( s ) is
realizable, the realization will not be unique.
™ Theorem 4.2 A transfer matrix Gˆ ( s ) is realizable
if and only if Gˆ ( s ) is a proper rational matrix.
Proof Suppose that {A,B,C,D} is a
realization of Gˆ ( s ), then we have
 −1 C[adj( sI − A)]B
G ( s ) = C ( sI − A) B + D =
ˆ +D
det( sI − A)

It is proper.
Next we show the converse. The Gˆ ( s ) ∈ R q× p
can be decomposed as two parts:
Gˆ ( s ) = Gˆ (∞) + Gˆ sp ( s ) (4.31)
where Gˆ sp (s) is the strictly proper part of . Let
d ( s ) = s r + α 1 s r −1 + L + α r −1 s + α r (4.32)
be the least common denominator of all
entries of Gˆ sp (s) and it is monic.
Then the Gˆ ( s ) can be expressed as
1
Gˆ sp ( s ) = N ( s)
d (s)
1 (4.33)
= [ N1s r −1 + N 2 s r − 2 + L + N r −1s + N r ]
d (s)

where N(s) is a q × p polynomial matrix of s

and Ni are q × p constant matrices.
Now we can verify that the following set of
equations
 ⎛ − α1 I p − α 2 I p L − α r −1 I p −αr I p ⎞ ⎛ I p ⎞
⎜ ⎟ ⎜ ⎟
⎜ Ip 0 L 0 0 ⎟ ⎜0⎟
x& = ⎜ 0 Ip L 0 0 ⎟ x + ⎜ 0 ⎟u
⎜ ⎟ ⎜ ⎟ (4.34)
⎜ L L L L L ⎟ ⎜M⎟
⎜ 0 0 L Ip 0 ⎟⎠ ⎜⎝ 0 ⎟⎠
⎝
y = ( N1 N 2 L N r −1 N r )x + Gˆ (∞)u

is one of the realization of Gˆ ( s ) , here we know

that dimensions of A, B and C are rp × rp, rp × p
and q × rp , respectively. Let us define
 ⎛ Z1 ⎞
⎜ ⎟
⎜ Z2 ⎟
Z := ⎜ ⎟ := ( sI − A) −1 B (4.35)
M
⎜ ⎟
⎜Z ⎟
⎝ R⎠

where Zi with dimension of p × p. The transfer

matrix of (4.34) can then be expressed as
C ( sI − A) −1 B + Gˆ (∞ )
= N1Z1 + N 2 Z 2 + L + N r Z r + Gˆ (∞ ) (4.36)
The (4.35) can be rewritten as (sI-A)Z=B or
sZ=AZ+B
Expanding above equation based on block
matrix, we have
sZ 1 = −α 1 Z 1 − α 2 Z 2 − L − α r Z r + I p (4.37)
sZ 2 = Z 1 , sZ 3 = Z 2 , L , sZ r = Z r −1 (4.37’)
The above last r-1 equations implies
1 1 1
Z 2 = Z 1 , Z 3 = 2 Z 1 , L Z r = r −1 Z 1
s s s
Substituting above r-1 equations into (4.37)
yields
⎛ α2 αr ⎞
sZ 1 = −⎜ α 1 + + L + r −1 ⎟ Z 1 + I p
⎝ s s ⎠
or
⎛ α2 αr ⎞
⎜ s + α1 + + L + r −1 ⎟ Z 1 = I p
⎝ s s ⎠
By using the notation of (4.32), the above
equation is turn out to be
d (s)
r −1
Z1 = I p
s
So we obtain
s r −1
Z1 = Ip
d (s)
Inserting it into (4.37’) leads to
s r −1 s r −2 1
Z1 = I p , Z2 = I p , LZ r = Ip
d (s) d (s) d (s)
 Substituting them into (4.36) yields
C ( sI − A) −1 B + Gˆ (∞)
1
= [ N1s r −1 + N 2 s r − 2 + L + N r ] + Gˆ (∞)
d (s)

This shows that (4.34) is a realization of Gˆ ( s ) .

™ The realization in Theorem 4.2 is called

controllable cannonical form.
™ Example 4.6
Consider the proper rational matrix
 ⎛ 4 s − 10 3 ⎞
⎜ ⎟
Gˆ ( s ) = ⎜ 2s + 1 s+2 ⎟
⎜ 1 s +1 ⎟
⎜ (2 s + 1)( s + 2) ( s + 2) 2 ⎟
⎝ ⎠
⎡ − 12 3 ⎤
⎡2 0⎤ ⎢ 2 s + 1 s + 2 ⎥
=⎢ ⎥ +⎢ 1 s + 1 ⎥ (4.38)
⎣0 0⎦ ⎢ ⎥
⎢⎣ (2 s + 1)( s + 2) ( s + 2) ⎥⎦
2

= Gˆ (∞) + Gˆ ( s )
sp

The monic least common denominator of Gˆ sp (s)

is
d ( s ) = ( s + 0.5)( s + 2) 2 = s 3 + 4.5s 2 + 6 s + 2
Thus we have
1 ⎡ − 6( s + 2) 2
3( s + 2)( s + 0.5)⎤
Gˆ sp ( s ) = 3 ⎢ ⎥
s + 4.5s 2 + 6s + 2 ⎣ 0.5( s + 2) ( s + 1)( s + 2) 2 ⎦

1 ⎛ ⎡− 6 3⎤ 2 ⎡− 24 7.5⎤ ⎡− 24 3 ⎤ ⎞
= ⎜⎢ ⎥ s +⎢ ⎥ s+⎢ ⎥ ⎟
⎜
d ( s ) ⎝ ⎣ 0 1⎦ 0.5⎦ ⎟⎠
⎣ 0.5 1.5 ⎦ ⎣ 1

and a realization of (4.38) is

 ⎡− 4.5 0 −6 0 −2 0 ⎤ ⎡1 0⎤
⎢ 0 − 4 .5 0 − 6 0 − 2 ⎥ ⎢0 1 ⎥
⎢ ⎥ ⎢ ⎥
⎢ 1 0 0 0 0 0⎥ ⎢0 0 ⎥
x& = ⎢ ⎥x + ⎢ ⎥u
⎢ 0 1 0 0 0 0⎥ ⎢0 0 ⎥
⎢ 0 0 1 0 0 0⎥ ⎢0 0 ⎥ (4.39)
⎢ ⎥ ⎢ ⎥
⎢⎣ 0 0 0 1 0 0 ⎥⎦ ⎢⎣0 0⎥⎦
⎡− 6 3 − 24 7.5 − 24 3 ⎤ ⎡2 0⎤
y=⎢ ⎥ x+⎢ ⎥ u
⎣ 0 1 0.5 1.5 1 0.5⎦ ⎣0 0⎦
™A special case of (4.31) and (4.34) with p=1.
To save space, we assume that r=4 and q=2.
That is
⎡ d1 ⎤ 1
G( s) = ⎢ ⎥ + 4
ˆ
⎣ 2⎦
d s + α 1 s 3
+ α 2 s 2
+ α3s + α 4
(4.40)
⎡ β11s 3 + β12 s 3 + β13 s + β14 ⎤
⋅⎢ ⎥
⎣ β 21s + β 22 s + β 23 s + β 24 ⎦
3 3

Thus its realization can be obtained directly

from (4.34) as
 ⎡− α 1 − α 2 − α 3 − α 4 ⎤ ⎡1 ⎤
⎢ 1 0 0 0 ⎥ ⎢0 ⎥
x& = ⎢ ⎥ x + ⎢ ⎥u
⎢ 0 1 0 0 ⎥ ⎢0 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0 0 1 0 ⎦ ⎣0 ⎦
⎡ β 11 β 12 β 13 β 14 ⎤ ⎡ d1 ⎤
(4.41)
y=⎢ ⎥ x + ⎢ ⎥u
⎣ β 21 β 22 β 23 β 24 ⎦ ⎣d 2 ⎦

™ Other ways of realization

Let Gˆ ci (s) be the ith column of Gˆ ( s ) and let
ui be the ith component of the input vector u.
Then yˆ ( s) = Gˆ ( s)uˆ ( s) can be expressed as
 yˆ ( s) = Gˆ c1 ( s)uˆ1 ( s) + Gˆ c 2 ( s)uˆ2 ( s) + L + Gˆ cm ( s)uˆm ( s)
=: yˆ c1 ( s) + yˆ c 2 ( s) + L + yˆ cm ( s)

as shown in Fig.4.4(a)
Suppose that
⎧ x&i = Ai xi + bi ui
⎨ i = 1,2, L, m
⎩ yci = Ci xi + d i ui

a realization of
yˆ ci ( s) = Gˆ ci ( s)uˆ i ( s)
Then
⎡ A1 ⎤ ⎡b1 ⎤
⎢ A ⎥ ⎢ b ⎥
x& = ⎢ 2 ⎥x + ⎢ 2 ⎥u
⎢ O ⎥ ⎢ O ⎥
⎢ ⎥ ⎢ ⎥
⎣ Am ⎦ ⎣ bm ⎦
y = [C1 C 2 L C m ]x + [d1 d 2 L d m ]u
will be a realization of Gˆ ( s ). Thus we can realize
each column of Gˆ ( s ) and then combine them to
yield a realization of Gˆ ( s ) .
™ Example4.7
Consider the proper rational matrix
⎡ 4 s − 10 3 ⎤
⎢ 2s + 1 s+2 ⎥
G (s) = ⎢
ˆ
1 s +1 ⎥
⎢ ⎥
⎢⎣ (2 s + 1)( s + 2) ( s + 2) ⎥⎦
2

The 1th and 2th columns of Gˆ ( s ) are

 ⎡ 4 s − 10 ⎤
⎢ 2 s + 1 ⎥ ⎡ 2⎤ 1 ⎡− 6 s − 12⎤
Gc1 ( s ) = ⎢
ˆ ⎥ =⎢ ⎥+ 2 ⎢ 0.5 ⎥
⎢
1
⎥ ⎣ ⎦
0 s + 2 .5 s + 1 ⎣ ⎦
⎢⎣ (2 s + 1)( s + 2) ⎥⎦
⎡ 2⎤ 1 ⎛ ⎡− 6⎤ ⎡− 12⎤ ⎞
=⎢ ⎥+ 2 ⎜ ⎢ ⎥s + ⎢ ⎥ ⎟
⎜ ⎟
⎣0⎦ s + 2.5s + 1 ⎝ ⎣ 0 ⎦ ⎣ 0.5 ⎦ ⎠

and
⎡ 3 ⎤
⎢ s+2 ⎥ 1 ⎡3s + 6⎤
Gc 2 ( s ) = ⎢ s + 1 ⎥ = 2
ˆ
⎢ s +1 ⎥
⎢ ⎥ s + 4 s + 4 ⎣ ⎦
⎢⎣ ( s + 2) ⎥⎦
2

1 ⎛ ⎡3⎤ ⎡6⎤ ⎞
= 2 ⎜ ⎢ ⎥s + ⎢ ⎥ ⎟
s + 4 s + 4 ⎜⎝ ⎣1⎦ ⎣1⎦ ⎟⎠
The realizations of them are
⎡ A1 ⎤ ⎡b1 ⎤
x& = ⎢ ⎥ x+⎢ ⎥ u
⎣ A2 ⎦ ⎣ b2 ⎦
y = [C1 C 2 ]x + [d1 d 2 ]u

That is
⎡− 2.5 − 1 0 0⎤ ⎡1 0⎤
⎢ 1 0 0 0 ⎥ ⎢0 0 ⎥
x& = ⎢ ⎥x + ⎢ ⎥u
⎢ 0 0 − 4 − 4⎥ ⎢0 1 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0 0 1 0 ⎦ ⎣ 0 0 ⎦
⎡− 6 − 12 3 6⎤ ⎡ 2 0⎤
y=⎢ ⎥ x+⎢ ⎥ u
⎣0 0.5 1 1⎦ ⎣0 0⎦
4.5 Solution of Linear time-varying (LTV)
state equation
Consider the linear time-varying (LTV) state
equation
x& (t ) = A(t ) x (t ) + B (t )u (t ) (4.45)
y (t ) = A(t ) x (t ) + D (t )u (t ) (4.46)
where x(t ) ∈ R n, u (t ) ∈ R p
and y (t ) ∈ R q
for any t. We
assume that for any initial state and any input
u(t), the state equation has a unique solution.
We consider the solutions of
x& (t ) = A (t ) x(t ) (4.49)
first.
The solution of the scalar time-varying
equation x& = a(t ) x due to x(0) is
t

x(t ) = e ∫0
a (τ ) dτ
x(0)

Can it be extended to the matrix situation?

That is t

x(t ) = e ∫0
A (τ ) dτ
x(0)
 the solution of (4.49) due to x(0)? The answer
is no. So we have to find another approach to
develop the solution.
™ Fundamental matrix
For any n initial state xi(t0)(i=1,2,…,n) which
are linearly independent, there exist n unique
solutions xi(t) (i=1,2,…,n) corresponding to
each initial state.
 Arrange the n solutions and n initial states
as square matrices
X(t)=[x1(t) x2(t) … xn(t)]
X(t0)=[x1(t0) x2(t0) … xn(t0)]
Obviously, X(t0) is nonsingular. The X(t) is
called a fundamental matrix of (4.49).
Because the selection of the initial states is
not unique, the fundamental matrix is not
unique.
™ Example 4.8
Consider the homogeneous equation

⎡0 0 ⎤
x& (t ) = ⎢ ⎥ x(t )
⎣ t 0⎦
That is
⎧ x&1 (t ) = 0
x&1 (t ) = 0⎨
⎩ x& 2 (t ) = tx1 (t )

For initial state x(0)=[x1(0) x2(0)]T, the solution

is
 ⎧ x1 (t ) = x1 (0)
⎨
⎩ 2
x (t ) = 0. 5t 2
x1 (0) + x 2 (0)

Specially, for two initial state x1(0)=[1 0]T and

x2(0)= [1 2]T which are linearly independent,
the solutions are
⎡ 1 ⎤
x1 (t ) = ⎢ 2⎥
⎣ 0 . 5t ⎦
and
⎡ 1 ⎤
x 2 (t ) = ⎢ ⎥
⎣ 0.5t 2
+ 2 ⎦

respectively.
Thus
⎡ 1 1 ⎤
X(t ) = ⎢
⎣ 0.5t 2
0.5t 2 + 2⎥⎦

is a fundamental matrix
™A very important property of the fundamental
matrix is that it is nonsingular for all t.
Definition 4.2 Let X(t ) be any fundamental
matrix of (4.49). Then
Φ(t , t 0 ) := X(t ) X −1 (t 0 )

is called the state transition matrix of (4.49).

The state transition matrix is also the unique
solution of
∂
Φ(t , t0 ) = A (t )Φ(t , t0 )
∂t (4.53)
with initial condition Φ(t 0 , t 0 ) = I
™ The properties of fundamental matrix
Φ(t , t ) = I (4.54)
Φ − 1 (t , t 0 ) = Φ ( t 0 , t ) (4.55)
Φ(t , t0 ) = Φ(t , t1 )Φ(t1 , t0 ) (4.56)
The above equations hold for any t, t0, and t1.
™ Example 4.9
Consider the homogeneous equation in
Example 4.8. Its fundamental matrix is
 ⎡ 1 1 ⎤
X(t ) = ⎢
⎣ 0.5t 2
0.5t 2 + 2⎥⎦

The inverse of it is

−1 ⎡ 0.25 t 2
+ 1 − 0.5⎤
X (t ) = ⎢ ⎥
⎣ − 0. 2 t 2
0.5 ⎦

Thus the state transition matrix is given by

−1 ⎡ 1 1 ⎤ ⎡0.25t 2 + 1 − 0.5⎤
Φ(t , t 0 ) = X(t ) X (t ) = ⎢ ⎢ ⎥
⎣ 0 . 5t 2
0.5t 2 + 2⎥⎦ ⎣ − 0.2t 2 0 .5 ⎦
⎡ 1 0⎤
=⎢ ⎥
⎣ 0 . 5(t 2
− t 2
0 ) 1 ⎦
™ Now we claim that the solution of (4.45)
excited by the initial state x(t0)=x0 and the
input u(t) is given by
t
x(t ) = Φ(t , t0 ) x 0 + ∫ Φ(t ,τ )B(τ )u(τ )dτ (4.47)
t0

= Φ(t , t0 ) x 0 + ∫ Φ(t0 ,τ )B(τ )u(τ )dτ ⎤

⎡ t

⎢⎣ t0 ⎥⎦ (4.48)

where Φ(t ,τ ) is the state transition matrix of

(4.49)
™ Here we need to prove that the (4.57) or (4.58)
is the solution of (4.45).
First we show that (4.57) satisfies the initial
condition.
t0
x(t0 ) = Φ(t0 , t0 )x 0 + ∫ Φ(t ,τ )B(τ )u(τ )dτ = Ix 0 + 0 = x 0
t0

Second, we verify that (4.57) satisfying the

state equation
d ∂ ∂ t
x(t ) = Φ(t , t0 )x 0 + ∫ Φ(t ,τ )B(τ )u (τ )dτ
dt ∂t ∂t t0

t⎛∂ ⎞
= A (t )Φ(t , t0 )x 0 + ∫ ⎜ Φ(t ,τ )B(τ )u(τ ) ⎟dτ + Φ(t , t )B(t )u(t )
t 0 ∂t
⎝ ⎠
t
= A (t )Φ(t , t0 )x 0 + ∫ A (t )Φ(t ,τ )B(τ )u(τ )dτ + B(t )u(t )
t0

t
= A (t )[Φ(t , t0 )x 0 + ∫ Φ(t ,τ )B(τ )u(τ )dτ ] + B(t )u(t )
t0

= A (t ) x(t ) + B (t )u (t )

Thus (4.57) is the solution.

Substituting (4.57) into (4.46) leads to
t
y (t ) = C(t )Φ(t , t0 ) x 0 + C(t ) ∫ Φ(t ,τ )B(τ )u(τ ) dτ
t0

+ D(t )u(t ) (4.59)

If the input is identically zero, the (4.57)
becomes
x(t ) = Φ(t , t0 )x 0

Corresponding to it, the zero-input response is

y (t ) = C(t )Φ(t , t0 )x 0
If the initial state is zero, the (4.59) reduces to
t
y (t ) = C(t ) ∫ Φ(t ,τ )B(τ )u(τ )dτ + D(t )u(t ) (4.60)
t
0
If we only concern response with time between
t0 and t, we can regard the input u(t) as
⎧u(τ ) t0 ≤ τ < t
U(τ ) = ⎨
⎩0 others

So when t ∈ [t 0 , t ]
∞
D(t )u(t ) = D(t ) U (t ) = ∫ D(τ )U (τ )δ (τ − t )dτ
−∞
t
= ∫ D(τ )u(τ )δ (τ − t ) dτ
t0

So the (4.59) can also be written as

t
y (t ) = ∫ [C(t )Φ(t ,τ )B(τ ) + D(τ )δ (τ − t )]u(τ )dτ (4.60)
t0
(4.60) is the zero-state response. Comparing
it with (2.5)
t
y (t ) = ∫ G (t ,τ )u(τ )dτ (2.5)
t0

leads to
G (t ,τ ) = C(t )Φ (t ,τ )B (τ ) + D(τ )δ (τ − t ) (4.62)
The (4.62) shows a relationship between the
input-output and state-space descriptions.
™ To compute the state transition matrix Φ(t ,τ )
is generally difficult, but for linear time-
invariant system, we have
X(t ) = e At

Φ(t ,τ ) = X(t ) X −1 (τ ) = e A ( t −τ ) = Φ(t − τ )

4.5.1 Discrete-Time Case
Consider the discrete-time state equation
x[k+1]=A[k]x[k]+B[k]u[k] (4.64)
y[k]=C[k]x[k]+D[k]u[k] (4.65)
™ Definethe discrete-time state transition matrix
as the solution of
Φ[k , k0 ] = A[k ]Φ[k , k0 ] with Φ[k0 , k0 ] = I

for k=1,2,…. The solution can be obtained

directly as
 Φ[k , k0 ] = A[k − 1]A[k − 2]L A[k0 ]

for k>k0 and Φ[k0 , k0 ] = I

™ It should be pointed out that the A-matrix may
be singular, thus the inverse of Φ[k , k0 ] may not
be defined. Moreover, the transfer property
Φ[k , k0 ] = Φ[k , k1 ]Φ[k1 , k0 ]
holds only for k 0 ≤ k1 ≤ k
™ Using the discrete state transition matrix
defined by (4.66), we can express the solution
of (4.64) and (4.65) as
 k −1
x[k ] = Φ[k , k0 ]x 0 + ∑ Φ[k , m + 1]B[m]u[m]
m = k0
k −1
y[k ] = C[k ]Φ[k , k0 ]x 0 + C[k ] ∑ Φ[k , m + 1]B[m]u[m] (4.67)
m = k0

+ D[k ]u[k ]

™ The zero-state response of (4.65) is

k −1
y[k ] = C[k ] ∑ Φ[k , m + 1]B[m]u[m] + D[k ]u[k ]
m = k0
4.6 Equivalent Time-Variying Equation
Consider the n-dimensional linear time-
varying state equation
x& = A (t )x + B(t )u
(4.69)
y = C(t )x + D(t )u
make the state equivalent transformation of
x (t ) = P (t ) x(t ) ,
where P(t) is nonsingular for all t.
the (4.69) then becomes
x& = A(t ) x + B (t )u (4.70)
y = C (t ) x + D(t )u
where

A (t ) = [P (t ) A (t ) + P& (t )]P −1 (t )
B (t ) = P (t )B(t )
C (t ) = C(t )P −1 (t )
D (t ) = D(t )

The (4.70) is said to be (algebraic) equivalent

to (4.69) and P(t) is called an (algebraic)
equivalence transformation.
™ Theorem 4.3 Let A0 be an arbitrary constant
matrix. Then there exists an equivalence
transformation that transforms the (4.69) into
(4.70) with A (t ) = A 0.
Proof: Let X(t) be a fundamental matrix
of x& = A(t )x . The differential of X-1(t)X(t)=I
yields
d & (t ) = 0
[ ( X −1 (t ))]X(t ) + X −1 (t ) X
dt

That is
 d & (t ) X −1 (t )
( X −1 (t )) = − X −1 (t ) X
dt

Since X& (t ) = A(t ) X(t ), inserting it into above

equation leads to
d
( X −1 (t )) = − X −1 (t ) A (t ) (4.72)
dt
For constant matrix A (t ) = A 0 , X(t ) = e A t is a 0

fundamental matrix of x& = A (t )x = A 0 x . If we

define
 P(t ) := X(t ) X −1 (t ) = e A 0t X −1 (t )
we have
A (t ) = [P(t ) A(t ) + P& (t )]P −1 (t )
d
= [e A 0t −1
X (t ) A (t ) + A 0 e A 0t −1
X (t ) + e A 0t
( X −1 (t ))]X(t )e − A 0t
dt

= e A 0t X −1 (t ) A (t ) X(t )e − A 0t + A 0 e A 0t X −1 (t ) X(t )e − A 0t
d
+ e A 0t ( X −1 (t )) X(t )e − A 0t
dt

= e A 0t X −1 (t ) A (t ) X(t )e − A 0t + A 0
+ e A 0t [− X −1 (t ) A (t )]X(t )e − A 0t = A 0
™ Result 1: Algebraic equivalence transformation
doesn’t change the impulse response matrix of
the system.
Proof: The impulse response matrix of (4.69)
is given by (4.62)
G (t ,τ ) = C (t )Φ (t ,τ ) B (t ) + D (t )δ (t − τ )

= C (t ) X (t ) X −1 (τ ) B (τ ) + D(t )δ (t − τ ) (4.62)
Similarly, the impulse response matrix of (4.70)
is
 G (t ,τ ) = C (t ) X (t ) X −1 (τ ) B (τ ) + D (t )δ (t − τ )

Using (4.71)
X(t ) := P (t ) X(t ) (4.71)
we have
G (t ,τ ) = C (t ) P −1 (t ) P (t ) X (t ) X −1 (τ ) P −1 (τ ) P(τ ) B (τ )
+ D(t )δ (t − τ )

= C (t ) X (t ) X −1 (τ ) B (τ ) + D(t )δ (t − τ )

Compare above equation with (4.62), we

conclude that G (t ,τ ) = G (t ,τ )
™ Result 2: The equivalence transformation in the time-
invariant case is not a special case of the time-
varying case. Because at least the property of the A-
matrix may not be preserved

Definition 4.3 A matrix P(t) is called a Lyapunov

transformation if P(t) is nonsingular, P(t) and dotP(t)
are continuous, and P(t) and P-1(t) are bounded for all
t. Equations (4.69) and (4.70) are said to be Lyapunov
equivalent if P(t) is a Lyapunov transformation.