EE61209: Linear Systems and Control

Post Mid-sem Lecture 3: Controllability and Stabilizability of LTI

System

Bhargav Jha

October 15, 2024

Bhargav Jha October 15, 2024 1 / 20

Controllability
https://www.youtube.com/watch?v=bsHryqnvyYA
Consider a linear dynamical system

ẋ(t) = Ax(t) + Bu(t)

y (t) = Cx(t) + Du(t)

The notion of controllability answers the following question:

Given any pair of initial state (x(0) := x0 ) and final state (xf ), and a finite
time tf > 0,( does there exist a u(t) for t ∈ [0, tf ] such that x(tf ) = xf ?
OR
Can the eigenvalues of A + BK can be freely assigned by using suitable K ?

if Yes, then the system (or more specifically the pair (A, B)) is
controllable.
Test method 1: A pair (A, B) is controllable
 if the controllability
2 n−1
matrix M = B AB A B · · · A B has full row rank.
Bhargav Jha October 15, 2024 2 / 20
Proof: Single Input Case

(A, B) is controllable ⇐⇒ M is full rank.

We have, Z t1
x(t1 ) = e At1
x0 + eA(t1 −τ ) Bu(τ )dτ
0
W.l.o.g consider x0 = 0. Then,
Z t1
x(t1 ) = eA(t1 −τ ) Bu(τ )dτ (1)
0

Using series expansion of eX and then by Cayley-Hamilton theorem,

eA(t1 −τ ) = g0 (t1 − τ )I + g1 (A(t1 − τ ) + · · · + gn−1 (t1 − τ )An−1

Substituting it in Eq. (1),
Z t1 Z t1 Z t1
x(t1 ) = B g0 (t1 −τ )u dτ +AB g1 (t1 −τ )u dτ +· · ·+An−1 B gn−1 (t1 −τ )u dτ
0 0 0

Bhargav Jha October 15, 2024 3 / 20

Proof (Contd..)

R t1
Let ηi = 0 gi (t1 − τ )u(τ ) dτ .
 
η0
 η1
x(t1 ) = B AB A2 B · · · An−1 B 
  
 ..
| {z }   .
M
ηn−1
if det(M) ̸= 0, then there exist a unique η such that x(tf ) = xf .
if det(M) = 0, then there exist x1 such that x1 = Mη is not solvable.

Bhargav Jha October 15, 2024 4 / 20

Controllability and Similarity Transformation

Can a similarity transformation alter the controllability of a system?

If x̃ = Tx, where det (T ) ̸= 0, then à = TAT −1 and B̃ = TB.
Hence, the new controllability matrix

M̃ = B̃ ÃB̃ · · · Ãn−1 B̃ = T B AB · · · An−1 B = T M

   

Therefore, a similarity transformation does not alter controllability.

Bhargav Jha October 15, 2024 5 / 20

Example-1

Consider a 2 mass spring damper system as shown.

u1 u2
b
m=1 m=1
k
y1 y2

Figure: 2 mass-spring-damper system

It can be represented as,

ÿ1 = u1 + k(y2 − y1 ) + b(ẏ2 − ẏ1 )

ÿ2 = u2 − k(y2 − y1 ) − b(ẏ2 − ẏ1 )

Bhargav Jha October 15, 2024 6 / 20

Example-1 (contd..)

 ⊤
One choice of states can be x = y1 y2 ẏ1 ẏ2 .
      
ẋ1 0 0 1 0 x1 0 0  
ẋ2   0 0 0 1  x2  0 0 u1
   
 =
ẋ3  −k k −b b  x3  + 1 0  u2
ẋ4 k −k b −b x4 0 1
 
  x1
1 0 0 0 x2 

y=
0 1 0 0 x3 

x4

Bhargav Jha October 15, 2024 7 / 20

Example 1 (contd..)

Computing the controllability matrix

2b 2 − k k − 2b 2
 
0 0 1 0 −b b
 0 0 0 1 b −b k − 2b 2 2b 2 − k

M= 
 1 0 −b b 2b 2 − k k − 2b 2 4bk − 4b 3
4b b 2 − k 
0 1 b −b k − 2b 2 2b 2 − k 4b b 2 − k 4bk − 4b 3
 
1 0 0 0 −k k 2bk −2bk
 0 1 0 0 k −k −2bk 2bk 
=  RREF form
 0 0 1 0 −b b 2b − k k − 2b 2 
2

0 0 0 1 b −b k − 2b 2 2b 2 − k

rank(M) = 4
It is controllable.

Bhargav Jha October 15, 2024 8 / 20

Example 2: Coupled control u1 = βu2

      
ẋ1 0 0 1 0 x1 0
ẋ2   0 0 0 1  x2   0 
 =
ẋ3  −k k −b b  x3  +  1  u
   

ẋ4 k −k b −b x4 β

− (β − 1) 2b 2 − k




0 1 b(β − 1)
 0 β b − bβ (β − 1) 2b 2 − k

M=
 1 b(β − 1) − (β − 1) 2b 2 − k

4b(β − 1) b 2 − k





2 −4b(β − 1) b 2 − k


β b − bβ (β − 1) 2b − k
det M = −k 2 (1 − β 2 )2 .
M loses rank if β = ±1.
Physical interpretation?

Bhargav Jha October 15, 2024 9 / 20

3-Mass System

Controllability matrix test is a ”yes-no” type of test.

Cumbersome for large systems.
Does not tell us about uncontrollable modes.

Bhargav Jha October 15, 2024 10 / 20

Popov–Belevitch–Hautus Test for Controllability
 
(A, B) is controllable ⇐⇒ The matrix A − λI B has full row rank for
all λ ∈ C
 
If (A, B) is uncontrollable, A − λI B loses rank at certain λi ∈ C.
These values are called uncontrollable modes.
The uncontrollable modes remain the eigenvalues of A + BK for all K .
Let PBH test fail for at certain λi . Then there is a non-trivial vector
η ̸= 0 such that

η ′ A − λi I B = 0 ⇐⇒ η ′ A = η ′ λ and η ′ B = 0.
 

Compute η ′ (A + BK ). η ′ A + η ′ BK = η ′ λi .
Hence, λi is always the eigenvalue of A + BK and it can not be
moved.
Q. What happens if the uncontrollable mode is in OLHP?
Q. What if they are in RHP?
Bhargav Jha October 15, 2024 11 / 20
Stabilizability

Stabilizability means there exist a K such that A + BK is Hurwitz.

If all the uncontrollable modes are in OLHP, the A + BK can be
made Hurwitz by a suitable choice of K .
If an uncontrollable mode is in RHP, it can not be moved by any
choice K , and hence A + BK can not be Hurwitz.

Bhargav Jha October 15, 2024 12 / 20

Example-3

For the two-mass system in Example 1, check its controllability using PBH
test.
 
−λ 0 1 0 0 0
   0 −λ 0 1 0 0
A − λI B =  −k k −b − λ

b 1 0
k −k b −b − λ 0 1
Reminder, if the above matrix loses rank, it will be at one of the
eigenvalues of A.
Step 1. Compute all eigenvalues
√ of A. √
spec(A) = {0, 0, −b − b 2 − 2k, −b +  b 2 − 2k}
Step 2. Check if the matrix A − λI B loses rank at any of these
eigenvalues.
Step 2 (Alternative) Compute left eigenvectors of A and see if η ′ B = 0.

Bhargav Jha October 15, 2024 13 / 20

Example 4

Use PBH Test for Example 2.

Only the B-matrix differs. Checking rank at all eigenvalues of A.
Does this tell the value of β where the system becomes uncontrollable?

Bhargav Jha October 15, 2024 14 / 20

Example 5
Consider a system,
       
ẋp Ap ϵ xp Bp 0 δe
= + (2)
ẋr 0 Ar xr 0 Br δa

where Ap , Ar , Bp , Br are non-zero. Use PBH test to determine the

controllability of the system.
Step 1: spec(A) = {Ap , Ar }. 
Ap − λ ϵ BP 0
Step 2: rank = 2 for all λ ∈ C. Hence,
0 Ar − λ 0 Br
the system is controllable.
Q. As ϵ → 0, will the controllability hold?
Q. Is it desirable to control xp with δa in this case?
The insight about the degree of controllability is missing the
aforementioned tests of controllability.
What if the we have a LTV system (A(t), B(t), C (t), D(t))? Can we
still apply the above tests?
Bhargav Jha October 15, 2024 15 / 20
Controllability Grammian Based Test

(A, B) is controllable ⇐⇒ Controllability grammian defined as

R t Aτ ′ ′
WC (t) := 0 e BB e τ dτ is positive definite for all t > 0.
A

Note that Wc (t) is positive semi-definite irrespective of the pair (A, B).

Z t Z t
′ Aτ ′ A′ τ ′ ′
T
x WC (t)x = x e BB e xdτ = (B ′ eA τ x)′ B ′ eA τ xdτ (3)
0 0
Z t
′
= ∥B ′ eA τ x∥2 dτ ≥ 0 (4)
0

Bhargav Jha October 15, 2024 16 / 20

Wc (t) and Open Loop Minimum Energy Control
′
Let um = B ′ eA (t1 −t) Wc (t1 )−1 x1 . Show that um is the minimum energy
control input that drives the system from 0 to x1 in time t1 .

The result is valid for MIMO systems too, but we will prove here for single
input case.
Let us verify that um drives the system to x1 .

Z t1
x(t1 ) = e(t1 −τ ) Bum (τ )dτ x
0
Z t1
′
= eA(t1 −τ ) BB ′ eA (t1 −τ ) Wc (t1 )−1 x1 dτ
Z0 t1
′
= eA(t1 −τ ) BB ′ eA (t1 −τ ) dτ Wc (t1 )−1 x1
Z0 t1
′
= eA(s) BB ′ eA (s) dsWc (t1 )−1 x1 = x1 t1 − τ = s
0

Bhargav Jha October 15, 2024 17 / 20

Contd..

Assume ū = um + uδ is the R tminimum energy control.

By linearity, x(t1 ) = x1 + 0 1 eA(t1 −τ ) Buδ .
Rt
x(t1 ) = x1 ⇐⇒ 0 1 eA(t1 −τ ) Buδ = 0
The energy of the new input is
Z t1 Z t1
′
(um + uδ ) (um + uδ )dt = Eum + Euδ + 2uδ um dt
0 0
Z t1
′
= Eu m + Eu δ + 2uδ (B ′ e A (t1 −t) WC (t1 )−1 x1 )′
0
Z t1
= 2uδ x1′ WC (t1 )−1 e A(t1 −t)B uδ dt = Eu + Eδ = Eū
0
(pythagoras theorem)

Bhargav Jha October 15, 2024 18 / 20

Example 6

       
ẋp −1 ϵ xp 1 0 δe
Let = + . Compute the minimum
ẋr 0 −1 xr 0 1 δp
energycontrol andverify its structure as below?
a11 a12ϵ
um =
a21ϵ a22
u as ϵ → 0, becomes decoupled.

Bhargav Jha October 15, 2024 19 / 20

Tests for Controllability

The following statements are equivalent.

⇐⇒ The pair (A, B) is controllable. 
⇐⇒ The controllability matrix M = B AB A2 B · · · An−1 B has full


row rank. Rt ′
⇐⇒ The matrix Wc (t) := 0eAτ BB ′ eA τ dτ is positive definite for all t > 0.
⇐⇒ The matrix A − λI B has full row rank for all λ ∈ C
⇐⇒ The eigenvalues of A + BK can be freely assigned (with the obvious
restriction that complex eigenvalues are in conjugate pairs) by a
suitable choice of K .

Bhargav Jha October 15, 2024 20 / 20