Assignment 4 – ECE410F Linear Control Systems

Controllability and Stabilization of LTI Systems

Solutions

Problem 1. Consider the control system

   
0 0 0 1
ẋ =  − 1 −1 − 2 x +  1 u.
   

0 1 1 −1

This problem has three parts.

1. Is the system controllable? Justify your answer.

2. Find the reachable set RT (0), T > 0, and find a vector x ∈ R3 that cannot be reached from the origin.
⊤ ⊤
3. Set x (0) = x0 := 0 0 1 . When u(t) ≡ 0, the solution at time π/2 is x (π/2) = 0 − 2 1 . Find the
 

reachable set Rπ/2 ( x0 ). Justify your answer.

Solution. Part 1. The controllability matrix is

 
1 0 0
Qc =  1 0 0 .
 

−1 0 0

Since rank( Qc ) = 1 < 3, the system is not controllable.

Part 2. The reachable set from the origin is the image of the linear transformation Lc,T , and we have
seen in class that Im(Lc,T ) = Im( Qc ), from which we deduce that
 
 1 
 
RT (0) = span  1  .
 
 
−1
 

 ⊤
The vector 1 0 0 is not in Im( Qc ), and hence not in RT (0).

1
 Part 3. We have x (π/2) = exp( A(π/2)) x0 + Lc,π/2 (u). We are told that exp( A(π/2)) x0 = 0 −

⊤
2 1 , and therefore
   

 0 

0 3
Rπ/2 ( x ) = x ∈ R : x =  − 2 + v, v ∈ Im(Lc,π/2 )
 
 
1
 
       

 0 1 λ 

3
= x ∈ R : x =  − 2 + λ  1  =  λ − 2 , λ ∈ R
     
 
1 −1 1−λ
 

= x ∈ R3 : x1 − x2 − 2 = 0, x1 + x3 − 1 = 0 .


Problem 2. Design a state feedback controller u = Kx that asymptotically stabilizes the system
   
1 1 0 0
ẋ = 0 − 1 1 x + 0 u,
   

0 0 1 1

ensuring that the rate of decay of solutions to zero is exactly e−2t .

Solution. The controllability matrix of the system,

 
0 0 1
Q c = 0 1 0 ,
 

1 1 1

is invertible so the system is controllable. Since A is upper triangular, its eigenvalues are on the diagonal,
σ( A) = {1, 1, −1}. Without computation, we deduce that the characteristic polynomial of A is p A (λ) =
(λ − 1)2 (λ + 1) = λ3 − λ2 − λ + 1. We have ( a0 , a1 , a2 ) = (1, −1, −1). To meet the control specification,
we place the eigenvalues at {−2, −2, −2}, which gives the desired characteristic polynomial pdes (λ) =
(λ + 2)3 = λ3 + 6λ2 + 12λ + 8, so that (α0 , α1 , α2 ) = (8, 12, 6). Letting
   
−1 −1 1 1 0 0
P = Qc  − 1 1 0 =  − 1 1 0 ,
   

1 0 0 −1 0 1

The controller u = Kx, with

h i h i
−1
K = −7 − 13 −7 P = − 27 − 13 −7

solves the problem.

2
Problem 3. Consider LTI system    
0 0 0 1
ẋ = 1 − 1 0 x + 1 u.
   

1 0 0 0
This problem has three parts.

1. Is the pair ( A, B) stabilizable? Justify your answer.

2. Design a state feedback controller of the form u = Kx such that for each initial condition x (0) ∈ R3 , each
component of the solution x (t) converges asymptotically to zero at a rate of exp(−t). You may express the
controller gain K in the form K = K̂P−1 , without computing P−1 .

3. Is it possible to design a controller making each component of the generic solution x (t) converge asymptot-
ically to zero at a rate of exp(−2t)? Justify your answer.

Solution. Part 1. The controllability matrix,

 
1 0 0
Q c = 1 0 0
 

0 1 0
has rank 2, and therefore the system is not controllable. Since A is lower triangular, its eigenvalues
are its diagonal entries, σ( A) = {0, 0, −1}. The system is stabilizable if and only if the eigenvalue 0 is
controllable. We check  
h i 0 0 0 1
rank A B = rank 1 − 1 0 1 = 3
 

1 0 0 0
(because the 3 × 3 submatrix obtained by eliminating the third column is invertible), and thus the system
is stabilizable.
Part 2. We’ll place the eigenvalues of the controllable subsystem at {−1, −1}. Recall that the control-
lability matrix
 
1 0 0
Q c = 1 0 0
 

0 1 0
has rank 2. The first two columns are a basis for Im( Qc ). We set
 
1 0 1
P = 1 0 0 .
 

0 1 0
The coordinate transformation z = T( x ) = P−1 x gives the Kalman decomposition
   
0 0 1 1
ż =  1 0 1  z +  0  u.
   

0 0 −1 0

3
We extract the controllable subsystem ( A11 , B1 ),
" # " #
1 0 0 1 1
ż = z + u,
1 0 0

and design K1 such that σ( A11 + B1 K1 ) = {−1, −1}. This can be done quickly without any computation
by noting that the matrix " #
−2 −1
1 0
has the desired eigenvalues, and noting that above matrix can be obtained through the feedback u =
− 2 − 1 z1 . The final feedback solving the problem is
 

h i
u = −2 − 1 0 P−1 x.

Part 3. The answer is no. No matter what feedback one designs, the closed-loop system will have an
eigenvalue at − 1 (the uncontrollable eigenvalue), and therefore the rate of convergence to zero of a
generic solution cannot be faster than exp(−t).

Problem 4. Consider the system    

5 − 11 5 1
ẋ =  0 −6 0 x + 0 u.
   

−5 5 −1 1

1. Show that the system is uncontrollable, but it is stabilizable.

2. Determine the controllable and uncontrollable eigenvalues of A.

3. Determine the fastest rate of exponential decay that solutions of the closed-loop system can enjoy if one
chooses suitable stabilizing feedbacks u = Kx.

4. Find a stabilizing feedback yielding the fastest rate of decay above for solutions of the closed-loop system.

Solution. Part 1. The controllability matrix of the system,

 
1 10 20
Q c = 0 0 0
 

1 −6 − 44

has rank 2, and therefore the system is uncontrollable.

To check stabilizability, we could use the PBH test, but since later in part 4 we need the Kalman
decomposition to design a stabilizing feedback, we use it now in place of the PBH rank test to find the

4
controllable and uncontrollable eigenvalues. The first two columns of Qc form a basis for Im( Qc ), and
therefore we set  
1 10 0
P = 0 0 1 .
 

1 −6 0

Letting z = P−1 x, we obtain

   
0 − 20 −1 1
ż = 1 4 − 1 z + 0 u. (1)
   

0 0 −6 0
The uncontrollable subsystem, ż3 = −6z3 , is asymptotically stable, so the pair ( A, B) is stabilizable.
Part 2. The uncontrollable eigenvalue is -6. The controllable ones are given by the spectrum of the
matrix " #
0 − 20
,
1 4
which are easily seen to be {2 ± 4i }. These eigenvalues can be arbitrarily shifted via feedback.
Part 3. Since we can’t shift the eigenvalue at − 6 while we can shift the ones at 2 ± 4i, the fastest rate
of decay that a state feedback u = Kx can induce in the closed-loop system is exp(−6t).
Part 4. We select a feedback that achieves the rate of decay above by shifting the two controllable
eigenvalues to {−7, −7}. To this end, we extract from (1) the controllable subsystem,
" # " #
0 − 20 1
ż1 = z1 + u,
1 4 0

and perform the standard single-input pole assignment procedure. The characteristic polynomial of A11
is p A11 (λ) = λ2 − 4λ + 20, from which we extract the coefficients ( a0 , a1 ) = (20, −4). The desired charac-
teristic polynomial is pdes (λ) = (λ + 7)2 = λ2 + 14λ + 49, so that (α0 , α1 ) = (49, 14). The controllability
matrix of the controllable subsystem is I2 . Letting
" # " #
−4 1 −4 1
P̂ = I2 = ,
1 0 1 0

the feedback controller for the controllable subsystem is u = K̂z1 , with

h i h i
K̂ = 20 − 49 − 4 − 14 P̂−1 = − 18 − 101 .
 
To go back to x coordinates, we express the feedback controller as u = K̂ 0 z, then use the fact that
h i
z = P−1 x, so that u = − 18 − 101 0 P−1 x.

5
Problem 5. Verify that the control system
   
0 1 0 0
ẋ =  0 0 0 x + 1 u,
   

−1 1 1 0

is not stabilizable. Find a state feedback controller with the property that solutions of the closed-loop system with
initial conditions on Im( Qc ) converge to zero at rate exp(−t).
Note: to save time, if the expression of your controller in x-coordinates involves the inversion of a 3 × 3 matrix,
you don’t need to compute the inverse.

Solution. The controllability matrix

 
0 1 0
Q c = 1 0 0
 

0 1 0
has rank 2, and therefore the system isn’t controllable. We compute the Kalman decomposition by
defining
 
0 1 0
P = 1 0 0 ,
 

0 1 1
and letting z = P−1 x. In z-coordinates, we have
   
0 0 0 1
ż =  1 0 0  z +  0  u.
   

0 0 1 0

We see that the uncontrollable subsystem, ż2 = z2 , is unstable so ( A, B) isn’t stabilizable.

The problem asks to find a controller that is only required to work for initial conditions on Im( Qc ).
In z-coordinates, this corresponds to taking initial conditions z2 (0) = 0, and these give z2 (t) ≡ 0. The
problem, then, is to stabilize the controllable subsystem,
" # " #
0 0 1
ż1 = A11 z1 + B1 u = z1 + u,
1 0 0

whose dynamics represent precisely the dynamics on Im( Qc ) We design a feedback to place the eigen-
values of ( A11 , B1 ) at − 1, as required by the problem. The characteristic polynomial of A11 is λ2 = 0,
and so a0 = a1 = 0. The desired characteristic polynomial is λ2 + 2λ + 1, and so α0 = 1, α1 = 2. We let
" #" # " #
1 0 0 1 0 1
T= = .
0 1 1 0 1 0

The desired gain K̂ is

K̂ = [ a0 − α0 a1 − α1 ] T −1 = [−2 − 1].

6
The controller we’ve found is u = − 2 − 1 z1 = − 2 − 1 0 z. We now need to express the controller
   

in x-coordinates. Recalling that z = P−1 x, we have

− 2 − 1 0 P−1 x = − 1 − 2 0 x.
   
u=

Problem 6. Consider the system    

−1 −1 −1 1
ẋ =  3 3 − 3 x + 1 u,
   

−4 −4 2 0
 
0 −2
and the subspace V = Im 1 1.
 

1 −1

1. Show that V is A-invariant.

2. Using the representation theorem, show that the open-loop system is unstable outside V , i.e., that solutions
of ẋ = Ax diverge from V .

3. Find a state feedback controller u = Kx that asymptotically stabilizes the system outside V , i.e., such that all
solutions of the closed-loop system converge to V .

4. Using the controller you found in the previous part, examine the closed-loop system inside the subspace
V . Is the closed-loop system stable, asymptotically stable, or unstable inside V ? Do different choices of K
solving part (c) have any effect on the dynamics inside V ?

 
0 −2
Solution. Let V = 1 1 . We have
 

1 −1
 
−2 2
AV =  0 0 .
 

−2 2
You can check that the matrix [V AV ] has rank 2, so the columns of AV are linearly dependent on
the columns of V (if they were linearly independent, then [V AV ] would have rank > 2). Therefore,
Im( AV ) ⊂ Im(V ), and thus V is A-invariant.
Next, we use the representation theorem as follows. Let
 
0 −2 0
P = 1 1 0 .
 

1 −1 1

7
As usual, the first two columns span V and the third column completes the basis for R3 . The coordinate
transformation z = P−1 x, gives
   
−1 1 − 7/2 3/2
−1 −1
ż = P AP + P Bu =  1 −1 1/2  z +  − 1/2  u.
   

0 0 6 −2

In z-coordinates, the subspace V is the (z1 , z2 ) plane. Hence, the dynamics outside V are given by

ż3 = 6z3 − 2u, (2)

while dynamics inside V are found by setting z3 = 0,

" # " #" # " #
ż1 −1 1 z1 3/2
= + u. (3)
ż2 1 − 1 z2 − 1/2

The open-loop system is unstable outside V because the dynamics (2) when u = 0 are given by ż3 = 6z3 ,
an unstable system. However, the subsystem (2) is obviously controllable, so it can be stabilized. For
instance, if we set u = 4z3 = [0 0 4]z, then (2) becomes

ż3 = −2z3 .

In x-coordinates, the controller reads as u = [0 0 4] P−1 x, so K = [0 0 4] P−1 = [−4 − 4 4]. This controller
asymptotically stabilizes the system outside V .
What happens to the dynamics inside V in (3) with the above controller? The feedback we have
chosen, u = 4z3 , is 0 on V because in z coordinates V is the set {z3 = 0}, so the dynamics inside V
become " # " #" #
ż1 −1 1 z1
= . (4)
ż2 1 − 1 z2
This system has two eigenvalues at 0 and − 2, so it is stable but not asymptotically stable. We deduce
that the closed-loop system is stable, but not asymptotically stable, inside V .
Any controller that asymptotically stabilizes the system outside V or, what is the same, any controller
stabilizing (2), must be zero when z3 = 0, and thus the closed-loop dynamics inside V are always described
by (4). Different choices of K stabilizing the system outside V have no effect on the dynamics inside V ,
and the closed-loop system will always be stable, but not asymptotically stable, inside V .