ECSL

Observer-based output-feedback control

for linear systems

University of Ulsan
Prof. KIM

c
°2023-2024 All rights reserved
 ECSL

Useful Lemmas
Lemma 1 (Congruent transform). For any nonsingular (i.e.,
invertible) matrix M , the following two conditions are equivalent:

• Ω < 0 (or Ω ≤ 0) (1)

• M T ΩM < 0 (or M T ΩM ≤ 0). (2)
 ECSL

Lemma 2. For a positive scalar ² and any matrices M ∈ Rr×s and

N ∈ Rs×r , the following inequality holds:
© ª
He M N ≤ ²−1 M M T + ²N T N.

Lemma 3. For a symmetric matrix Ω < 0 and any matrices M and

N , the following inequality holds:
© T ª
M ΩM ≤ −He M N − N T Ω−1 N.
T
 ECSL

Lemma 4 (Schur complement). For any matrices S = S T ,

R = RT , and N , if the following condition is satisfied:
     
S N R NT
  < 0  or   < 0 (3)
NT R N S

then it holds that

S − N R−1 N T < 0, S < 0, R < 0. (4)

 ECSL

1 State observer design

• To begin with, let us consider the following linear state-space
model:

ẋ(t) = Ax(t) + Bu(t)
(5)
y(t) = Cx(t)

where x(t) ∈ Rn represents the state, u(t) ∈ Rm represents the

control input, and y(t) ∈ Rp represents the measurable output.
 ECSL

• In most cases, a so-called Luenberger observer is utilized for

estimating the state:
 ¡ ¢
x̃(t)
˙ = Ax̃(t) + Bu(t) − L y(t) − ỹ(t)
(6)
ỹ(t) = C x̃(t).

where x̃(t) denotes the estimated state.

• In (6), the matrix L ∈ Rn×p is called the observer gain to be
designed later.
 ECSL

• Now, let us define the estimation error as

e(t) = x(t) − x̃(t).

• Then, our aim is to design the observer gain L that achieve

lim e(t) → 0. (7)

t→∞

• To accomplish this aim, we need to derive the error dynamic

system model from (5) and (6):
¡ ¢
ė(t) = Ae(t) + L y(t) − ỹ(t)
³ ´
= A + LC e(t). (8)
 ECSL

The following theorem provides LMI-based observer design

conditions obtained in the sense of Lyapunov stability.
Theorem 1. Convergence condition (7) holds if there exist matrices
P = P T ∈ Rn×n and L̄ ∈ Rn×p , such that

P >0
n o
He P A + L̄C < 0.

Furthermore, the observer gain can be reconstructed in this manner:

L = P −1 L̄.
 ECSL

Example 1.1. Design an observer gain for

   
0 −1 0 h i
A=  , B =   , C = 0.5 0 .
1 0 0

And see if the estimated state approaches

 to the real state,
 where the
0.1 0.0
initial condition is given as x(0) =   and x̃(0) =  .
−0.1 0.0
 ECSL

2 Observer-based control design via

separation principle
• Let us consider a continuous-time linear system of the following
general form:

ẋ(t) = Ax(t) + Bu(t)
(9)
y(t) = Cx(t)

where x(t) ∈ Rn represents the state, u(t) ∈ Rm represents the

control input, and y(t) ∈ Rp represents the measurable output.
 ECSL

• In observer-based control systems,

(i) the observer is used to estimate the internal state of the system
based on the available input and output measurements.
¡ ¢
˙x̃(t) = Ax̃(t) + Bu(t) − L y(t) − C x̃(t) . (10)

(ii) This estimated state x̃(t) is then used by the controller to

generate control signals to regulate the behavior of system.

u(t) = K x̃(t). (11)

 ECSL

• To sum up, our main goal is to design the control gain K and the
observer gain L such that

(i) lim x(t) → 0

t→∞
(ii) lim e(t) → 0
t→∞

where x(t) is the state and e(t) = x(t) − x̃(t) is the estimation error.
• To accomplish this goal, we first have to obtain a dynamic system
model with respect to x(t) and e(t):

ẋ(t) = ....
ė(t) = ....
 ECSL

• Based on x̃(t) = x(t) − e(t), Eqs. (9) and (11) provide

ẋ(t) = Ax(t) + BK x̃(t)

= (A + BK)x(t) − BKe(t).

• From (9) and (10), it follows that

ė(t) = (A + LC)e(t).

• As a result, the closed-loop control system is described as

    
ẋ(t) A + BK −BK x(t)
 =  . (12)
ė(t) 0 A + LC e(t)
 ECSL

Separation principle: The separation principle refers to the ability

to design the observer and the controller independently.
In other words, the separation principle states that the design of the
observer can be carried out independently of the controller design.
• Controller design: Theorem 1 in “State-feedback control”
• Observer design: Theorem 1 in Section 1.
 ECSL

Proof: The stability of closed-loop system (12) is ensured if the

following condition holds:
 
A + BK −BK
real number of λi    < 0, ∀i = 1, 2, · · · , 2n.
0 A + LC

This condition is equivalent to

• real number of λi (A + BK) < 0, ∀i = 1, 2, · · · , n

• real number of λi (A + LC) < 0, ∀i = 1, 2, · · · , n.

Consequently, it can be said that both control and observer gains can
be designed separately.
 ECSL

Example 2.1. By using the separation principle, obtain both

control and observer gains for
   
1 −2 1 h i
A= , B =   , C = 0.1 0 .
1 4 0.1

And see if the state trajectory

 on the x1 -x2 plane 
converges to the
1.0 0.0
origin, where x(0) =   and x̃(0) =  .
−1.0 0.0
 ECSL

Remark 1. The separation principle offers convenience when

designing observer-based control. However, it also presents the
following disadvantages:
• When the system contains uncertainties, it is impossible to
utilize the separation principle because these uncertainties
cannot be incorporated into the observer.
• In performance-based control designs, the separation principle
fails to yield an optimal solution that includes both control and
observer gains.
 ECSL

3 Observer-based control design via

Lyapunov stability approach
• Let us consider a continuous-time linear system of the following
general form:

ẋ(t) = Ax(t) + Bu(t)
(13)
y(t) = Cx(t)

where x(t) ∈ Rn represents the state, u(t) ∈ Rm represents the

control input, and y(t) ∈ Rp represents the output.
 ECSL

• Based on (10) and (11), the observer-based control is described as

 ¡ ¢
x̃(t)
˙ = Ax̃(t) + Bu(t) − L y(t) − C x̃(t)
(14)
u(t) = K x̃(t)

where x̃(t) ∈ Rn represents the estimated state; K ∈ Rm×n and

L ∈ Rn×p are the control gain and the observer gain, respectively, to
be designed later.
 ECSL

• As a result, based on
 
x(t)
ζ(t) =   ∈ R2n×2n
e(t)

the closed-loop system with (13) and (14) is described as follows:

ζ̇(t) = Āζ(t) (15)

where
   
x(t) A + BK −BK
ζ(t) =   ∈ R2n×2n , Ā =  .
e(t) 0 A + LC
 ECSL

Theorem 2. For a prescribed value µ, close-loop control system (15)

is stable at origin if there exist matrices P̄1 = P̄1T ∈ Rn×n ,
P2 = P2T ∈ Rn×n , K̄ ∈ Rm×n , and L̄ ∈ Rn×p such that

P̄1 > 0 (16)

P2 > 0 (17)
 © ª 
He AP̄1 + B K̄ −B K̄ 0
 © ª 
 ? −He µP̄1 µI  < 0. (18)
 
© ª
? ? He P2 A + L̄C

Furthermore, the control and observer gains are constructed as

follows:

K = K̄ P̄1−1 , L = P2−1 L̄.

 ECSL

Example 3.1. By using Theorem 2 for µ = 100, obtain both control

and observer gains for
   
1 −2 1 h i
A= , B =   , C = 0.1 0 .
1 4 0.1

And see if the state trajectory

 on the x1 -x2 plane 
converges to the
1.0 0.0
origin, where x(0) =   and x̃(0) =  . Finally,
−1.0 0.0
compare it with the state trajectory of Example 2.1.
 ECSL

4 Robust observer-based control

• Let us consider a continuous-time uncertain system of the following
general form:

ẋ(t) = (A + ∆A)x(t) + (B + ∆B)u(t)
(25)
y(t) = Cx(t)

where ∆A ∈ Rn×n and ∆B ∈ Rn×m denote the parameter

uncertainties expressed as

∆A = EΥH1 , ∆B = EΥH2 . (26)

 ECSL

• As in (14), the observer-based control is configured by using the

nominal matrices A, B, and C:
 ¡ ¢
x̃(t)
˙ = Ax̃(t) + Bu(t) − L y(t) − C x̃(t)
u(t) = K x̃(t)

where x̃(t) ∈ Rn represents the estimated state; K ∈ Rm×n and

L ∈ Rn×p are the control gain and the observer gain, respectively, to
be designed later.
 ECSL

• Hence, based on
 
x(t)
ζ(t) =   ∈ R2n×2n
e(t)

the closed-loop system with (25) and (14) is described as follows:

ζ̇(t) = (Ā + ∆Ā)ζ (27)

where
   
A + BK −BK ∆A + ∆BK −∆BK
Ā =   , ∆Ā =  .
0 A + LC ∆A + ∆BK −∆BK
 ECSL

• Furthermore, the closed-loop system can be transformed into

ζ̇(t) = (Ā + ĒΥH̄)ζ(t)

where
   
A + BK −BK E
Ā =   , Ē =  
0 A + LC E
h i
H̄ = H1 + H2 K −H2 K .