Plan of the Lecture

I Review: arbitrary pole placement by full state feedback.

I Today’s topic: observer design for state estimation when
full state feedback is not implementable.

Goal: for observable systems (definition to be introduced

today), learn how to estimate the state x from output y = Cx
using an observer.

Reading: FPE, Chapter 7

Review: Pole Placement via State Feedback
Assume that the plant is controllable:

r
+ u ẋ = Ax + Bu
y
y=x

ẋ = Ax + B(−Kx + r) = (A − BK)x + Br, y=x

Transfer function from R to Y :

Y (s) = (Is − A + BK)−1 BR(s)

Closed-loop poles are the eigenvalues of A − BK!!

Review: Pole Placement in CCF
ẋ = (A − BK)x + Br, y = Cx
 
0 1 ... 0 0
 0 0 ... 0 0 
 
 .. .. .. .. .. 
A − BK = −  . . . . . 
 
 0 0 ... 0 1 
an + k1 an−1 + k2 . . . a2 + kn−1 a1 + kn
Closed-loop poles are the roots of the characteristic polynomial
det(Is − A + BK)
= sn + (a1 + kn )sn−1 + . . . + (an−1 + k2 )s + (an + k1 )
Key observation: When the system is in CCF, each control
gain affects only one of the coefficients of the characteristic
polynomial, and these coefficients can be assigned arbitrarily by
a suitable choice of k1 , . . . , kn .
Hence the name Controller Canonical Form — convenient for
control design.
Pole Placement by State Feedback

General procedure for any controllable system:

1. Convert to CCF using a suitable invertible coordinate
transformation T (such a transformation exists by
controllability).
2. Solve the pole placement problem in the new coordinates.
3. Convert back to original coordinates.
Is Full State Feedback Always Available?

In a typical system, measurements are provided by sensors:

u
plant y

controller

Full state feedback u = −Kx is not implementable!!

When Full State Feedback Is Unavailable ...

... we need an Observer!!

State Estimation Using an Observer

When full state feedback is unavailable, the observer is used to

estimate the state x:

y
u plant observer b
x
State Estimation Using an Observer
The idea is to design the observer in such a way that the state
b is asymptotically accurate:
estimate x
v
u n
uX t→∞
kbx(t) − x(t)k = t xi (t) − xi (t))2 −−−→ 0
(b
i=1

If we are successful, then we can try estimated state feedback:

y
plant observer

b
x
u= b
Kx
K
A New Concept: Observability

I Before, we saw that closed-loop poles can be assigned

arbitrarily by full state feedback when the plant is
controllable.
I Now, we will see that asymptotically accurate state
estimation will be possible when the system is observable.
I Observability is a system property which is dual to
controllability.
Observability
Consider a single-output system (y ∈ R):

ẋ = Ax + Bu, y = Cx x ∈ Rn

The Observability Matrix is defined as

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

— recall that C is 1 × n and A is n × n, so O(A, C) is n × n;

— the observability matrix only involves A and C, not B

We say that the above system is observable if its observability

matrix O(A, C) is invertible.
(This definition is only true for the single-output case; the
multiple-output case involves the rank of O(A, C).)
Example: Computing O(A, C)

 
0 −6

Let A= , C= 0 1
1 −5

Here, n = 2, C ∈ R1×2 , A ∈ R2×2 =⇒ O(A, C) ∈ R2×2 .

 
C
O(A, C) =
CA
 

0 −6

where CA = 0 1 = 1 −5
1 −5
 
0 1
∴ O(A, C) =
1 −5
det O(A, C) = −1 =⇒ the system is observable

— recall: this system is in Observer Canonical Form (OCF) ...

Observer Canonical Form
A single-output state-space model

ẋ = Ax + Bu, y = Cx

is said to be in Observer Canonical Form (OCF) if the matrices

A, C are of the form
 
0 0 ... 0 0 ∗
1 0 . . . 0 0 ∗
 

 .. ..  ,
A =  ... ... . . . ... . . C = 0 0 ... 0 1
 
0 0 . . . 1 0 ∗
0 0 ... 0 1 ∗

Fact: A system in OCF is always observable!!

(The proof of this for n > 2 uses the Jordan canonical form, we will
not worry about this.)
Coordinate Transformations and Observability
Just like controllability, observability is preserved under
invertible coordinate transformations.
T
ẋ = Ax + Bu −−−−→ x̄˙ = Āx̄ + B̄u
y = Cx y = C̄ x̄
−1 −1
where Ā = T AT , B̄ = T B, C̄ = CT

   
C̄ CT −1
  
C̄ Ā 
CT T AT −1
−1
   
O(Ā, C̄) =   =
.. 
..
  
. .
C̄ Ā n−1 −1
CT T An−1 T −1
 
C
 CA 
 
=  .  T −1 = O(A, C)T −1
 .  .
CAn−1
Coordinate Transformations and Observability
Just like controllability, observability is preserved under
invertible coordinate transformations:
T
ẋ = Ax + Bu −−−−→ x̄˙ = Āx̄ + B̄u
y = Cx y = C̄ x̄
where Ā = T AT −1 , B̄ = T B, C̄ = CT −1

If the original system is observable, then

 −1
T [O(A, C)]−1 = O(Ā, C̄)
| {z } | {z }
old new
m
 −1
T = O(Ā, C̄) [O(A, C)]
| {z } | {z }
new old
Observability and State Estimation

As we will show next:

If the system is observable, then there exists an observer (state

estimator) that provides an asymptotically convergent
estimate xb of the state x based on the observed output y.

The particular type of observer we

will construct is called the Luenberger
observer after David G. Luenberger,
who developed this idea in his 1963
Ph.D. dissertation.
David Luenberger is a Professor at
Stanford University.
The Luenberger Observer

Consider a state-space model

ẋ = Ax (for now, assume u = 0)

y = Cx

We wish to estimate the state x based on the output y.

Consider feeding the output y as input to the following system
b:
with state x

ḃ = (A − LC)b
x x + Ly.

Assumption: The output injection matrix L is chosen in such a

way that the matrix A − LC is Hurwitz (i.e., all of its
eigenvalues lie in LHP).
At this point, we do not assume anything about observability.
The Luenberger Observer

System: ẋ = Ax
y = Cx
Observer: ḃ = (A − LC)b
x x + Ly.

What happens to state estimation error e = x − x

b as t → ∞?

ė = ẋ − x
ḃ
= Ax − [(A − LC)b
x + LCx]
= (A − LC)x − (A − LC)b
x
= (A − LC)e

Does e(t) converge to zero in some sense?

Linear ODEs and Eigenvalues: A Digression

v̇ = F v, v ∈ Rn , F ∈ Rn×n
Let λ1 , . . . , λn be eigenvalues of F , i.e., roots of
det(Is − F ) = 0.
Then there exists a matrix T ∈ Rn×n , such that T −1 = T T and
 
λ1
 λ2 
 
F = T −1  .. T
 . 
λn
Consider the change of coordinates v̄ = T v. Then
 
λ1
 λ2 
 
v̄˙ = T F T −1 v̄ =  ..  v̄
 . 
λn
Linear ODEs: A Digression
 
λ1
 λ2 
 
v̄˙ = T F T −1 v̄ =  ..  v̄, (λ1 , . . . , λn ) = eig(F )
 . 
λn
m
v̄˙ i = λi v̄i , i = 1, 2, . . . , n

This system of n 1st-order ODEs has the solution

v̄i (t) = v̄i (0)eλi t , i = 1, 2, . . . , n

If all λi ’s have negative real parts, then

kv(t)k2 = v(t)T v(t) = v̄(t)T v̄(t)

≤ Ce−2σmin t , where σmin = min |Re(λi )|
1≤i≤n
The Luenberger Observer

System: ẋ = Ax
y = Cx
Observer: ḃ = (A − LC)b
x x + Ly
Error: ė = (A − LC)e

Recall our assumption that A − LC is Hurwitz (all eigenvalues

are in LHP). This implies that
n
X t→∞
b(t)k2 = ke(t)k2 =
kx(t) − x |ei (t)|2 −−−→ 0
i=1

at an exponential rate, determined by the eigenvalues of

A − LC.
For fast convergence, want eigenvalues of A − LC far into
LHP!!
The Luenberger Observer

System: ẋ = Ax
y = Cx
Observer: ḃ = (A − LC)b
x x + Ly
Error: ė = (A − LC)e

Observer transfer function:

b
sX(s) b
= (A − LC)X(s) + LY (s)
b
(Is − A + LC)X(s) = LY (s)
b
X(s) = (Is − A + LC)−1 LY (s).

The eigenvalues of A − LC are the observer poles. We want

these poles to be stable and fast.
Observability and Estimation Error

Fact: If the system

ẋ = Ax, y = Cx

is observable, then we can arbitrarily assign eigenvalues of

A − LC by a suitable choice of the output injection matrix L.

This is similar to the fact that controllability implies arbitrary

closed-loop pole placement by state feedback.
In fact, these two facts are closely related because CCF is dual
to OCF.
Observer Pole Placement in OCF
Consider a single-output system in OCF:

ẋ = Ax
y = Cx, y∈R
 
0 0 ... 0 0 −an
1 0 ... 0 0 −an−1 
 

 ..  ,
where A =  ... .. . .
. .
.. ..
. . .  C = 0 0 ... 0 1
 
0 0 ... 1 0 −a2 
0 0 ... 0 1 −a1

Note that AT has the form of a CCF system matrix, thus:

det(Is − A) = det((Is − A)T ) = det(Is − AT )

= sn + a1 sn−1 + . . . + an−1 s + an
Now Let’s Add an Observer
 
0 0 ... 0 −an
0 1 ... 0 −an−1 
 
 .. 
A =  ... ..
.
..
.
..
. . 
 
0 0 ... 0 −a2 
0 0 ... 1 −a1
 
  0 0 ... 0 `1
`1 0 0 
 `2  ... 0 `2
 



LC =  .  0 0 ... 1 =  ... ... ..
.
..
.
..
. 
 ..   
0 0 ... 0 `n−1 
`n
0 0 ... 0 `n
 
0 0 ... 0 −(an + `1 )
0 1 ... 0 −(an−1 + `2 )
 
 
A − LC =  ... ..
.
..
.
..
.
..
. 
 
0 0 ... 0 −(a2 + `n−1 )
0 0 ... 1 −(a1 + `n )

— still in OCF!!
Observer Pole Placement in OCF
ẋ = Ax, y = Cx, ḃ = (A − LC)b
x x + Ly
 
0 0 ... 0 −(an + `1 )
1 0 ... 0 −(an−1 + `2 )
 
 
A − LC =  ... .. . .
. .
..
.
..
. 
 
0 0 ... 0 −(a2 + `n−1 )
0 0 ... 1 −(a1 + `n )
Eigenvalues of A − LC are the roots of the characteristic
polynomial
det(Is − A + LC)
= sn + (a1 + `n )sn−1 + . . . + (an−1 + `2 )s + (an + `1 )
Key observation: In OCF, each observer gain affects only one
of the coefficients of the characteristic polynomial, which can be
assigned arbitrarily by a suitable choice of `1 , . . . , `n .
Hence the name Observer Canonical Form — convenient for
observer design.
Observer Pole Placement

General procedure for any observable system:

1. Convert to OCF: T = O(Ā, C̄)−1 [O(A, C)]
| {z } | {z }
new old
2. Find L̄, such that Ā − L̄C̄ has desired eigenvalues.
3. Convert back to original coordinates: L = T −1 L̄.

The resulting observer is

ḃ = (A − T −1 L̄C)b
x x + T −1 L̄y

In fact, this procedure is not necessary because of duality

between controllability and observability!!
Controllability–Observability Duality
Claim: The system

ẋ = Ax, y = Cx

is observable if and only if the system

ẋ = AT x + C T u

is controllable.

 
Proof: C(AT , C T ) = C T | AT C T | . . . | (AT )n−1 C T
 T
C
 CA 
 
=  .  = [O(A, C)]T
 .. 
CAn−1

Thus, O(A, C) is nonsingular if and only if C(AT , C T ) is.

Observer Pole Placement, O/C Duality Version
Given an observable pair (A, C):
1. For F = AT , G = C T , consider the system ẋ = F x + Gu
(this system is controllable).
2. Use our earlier procedure to find K, such that

F − GK = AT − C T K

has desired eigenvalues.

3. Then

eig(AT − C T K) = eig(AT − C T K)T = eig(A − K T C),

so L = K T is the desired output injection matrix.

Final answer: use the observer

ḃ = (A − LC)b
x x + Ly
= (A − K T C)b
x + K T y.
Combining Full-State Feedback with an Observer
I So far, we have focused on autonomous systems (u = 0).
I What about nonzero inputs?
ẋ = Ax + Bu
y = Cx
— assume (A, B) is controllable and (A, C) is observable.
I In the next lecture, we will learn how to use an observer
together with estimated state feedback to (approximately)
place closed-loop poles.
u= b
Kx
plant y

b
x
K observer