Discussion 8A @ 2019-10-29 09:38:28-07:00

EECS 16B Designing Information Devices and Systems II

Fall 2019 Discussion Worksheet Discussion 8A
1. Changing behavior through feedback
In this question, we discuss how feedback control can be used to change the effective behavior of a system.

(a) Consider the scalar system:

x(t + 1) = 0.9x(t) + u(t) + w(t) (1)
where u(t) is the control input we get to apply based on the current state and w(t) is the external
disturbance.
Is the system stable? If |w(t)| ≤ ε, what can you say about |x(t)| at all time if you further assume that
u(t) = 0 and the initial condition x(0) = 0? How big can |x(t)| get?
Answer: The system is stable, as λ = 0.9 → |λ | < 1. We can say that |x(t)| is bounded at all time if
the disturbance is bounded. The system is bounded-input, bounded output (BIBO) stable.
|x(t)| forms a geometric series:

i
|x(t)| ≤ ∑ ε(0.9)k
k=0
In the limit as i → ∞,
ε
|x(t)| ≤ = 10ε
1 − 0.9

(b) Suppose that we decide to choose a control law u(t) = kx(t) to apply in feedback. For what values of
λ can you get the system to behave like:
x(t + 1) = λ x(t) + w(t) (2)
vis-a-vis the disturbance w(t)? How would you pick k?
Answer: We can control the system to have any value of λ , as long as we’re not limited on the values
of k.
x(t + 1) = 0.9x(t) + kx(t) + w(t) = λ x(t) + w(t)
k = λ − 0.9

(c) For the previous part, which k would you choose to minimize how big |x(t)| can get?
Answer: In order to have the minimum bound on |x(t)|, λ = 0. To get this λ ,
k = −0.9.
In the limit as i → ∞ in this case,
ε
|x(t)| =
=ε
1−0
The minimum bound on |x(t)| = ε, the same bound as on the disturbance.

© UCB EECS 16B, Fall 2019. All Rights Reserved. This may not be publicly shared without explicit permission. 1
 Discussion 8A @ 2019-10-29 09:38:28-07:00

(d) What if instead of a 0.9, we had a 3 in the original (1). What, if anything, would change?
Answer: If our system were now,

x(t + 1) = 3x(t) + u(t) + w(t), (3)

the system would no longer be stable. However, we can still choose any eigenvalue λ using closed
loop feedback. In this case,
k = λ − 3.

(e) Now suppose that we have a vector-valued system with a vector-valued control:

~x(t + 1) = A~x(t) + B~u(t) + ~w(t) (4)

where we further assume that B is an invertible square matrix.

Suppose we decide to apply linear feedback control using a square matrix K so we choose~u(t) = K~x(t).
For what values of matrix G can you get the system to behave like:

~x(t + 1) = G~x(t) + ~w(t) (5)

vis-a-vis the disturbance ~w(t)? How would you pick K given knowledge of A, B and the desired goal
dynamics G?
Answer: Since in this case our input is the same rank as our output, we can arbitrarily choose the
matrix G. As long as B is invertible (as given), we can define:

~x(t + 1) = A~x(t) + B~u(t) + ~w(t) (6)

= A~x(t) + BK~x(t) + ~w(t) (7)
= (A + BK)~x(t) + ~w(t) (8)
= G~x(t) + ~w(t) (9)

Therefore:

K = B−1 (G − A)

2. Eigenvalues Placement in Discrete Time

Consider the following linear discrete time system
" # " #
0 1 1
~x(t + 1) = ~x(t) + u(t) + ~w(t) (10)
2 −1 0

(a) Is this system controllable from u(t)?

Answer: We calculate " #
h i 1 0
R2 = B AB =
0 2
Observe that R2 matrix is full rank and hence our system is controllable.

© UCB EECS 16B, Fall 2019. All Rights Reserved. This may not be publicly shared without explicit permission. 2
Discussion 8A @ 2019-10-29 09:38:28-07:00

(b) Is the linear discrete time system stable?

Answer: We have to calculate the eigenvalues of matrix A. Thus,
det(λ I − A) = 0 ⇒ λ1 = 1, λ2 = −2
Since the magnitudes of the eigenvalues λ1 and λ2 are greater than 1, the system is unstable.
(c) Derive ha state space
i representation of the resulting closed loop system using state feedback of the form
u(t) = k1 k2 ~x(t)
Answer: The closed loop system using state feedback has the form
" # " #
0 1 1
~x(t + 1) = ~x(t) + u(t)
2 −1 0
" # " #
0 1 1 h i
= ~x(t) + · ( k1 k2 ~x(t))
2 −1 0
" # " #
0 1 1 h i
={ + · k1 k2 }~x(t)
2 −1 0
Thus, the closed loop system has the form
" #
k1 1 + k2
~x(t + 1) = ~x(t)
2 −1
| {z }
Acl

(d) Find the appropriate state feedback constants, k1 , k2 in order the state space representation of the re-
sulting closed loop system to place the eigenvalues at λ1 = − 21 , λ2 = 12
Answer: k1 = 1, k2 = − 11
8 Answer: From the previous part we have computed the closed loop
system as

" #
k1 1 + k2
~x(t + 1) = ~x(t)
2 −1
| {z }
Acl

Thus, finding the eigenvalues of the above system we have

" #
k1 − λ 1 + k2
0 = det(A − λ I) = det = λ 2 + (1 − k1 )λ + (−k1 − 2k2 − 2)
2 −1 − λ

We want to place the eigenvalue at λ1 = − 21 , λ2 = 12 . This means that we should choose the gains k1
and k2 so that the characteristic equation is
1 1 1
0 = (λ − )(λ + ) = λ 2 − .
2 2 4
Thus we should choose k1 and k2 satisfying the system of equations
0 = 1 − k1
1 .
− = −k1 − 2k2 − 2
4
This system has solution k1 = 1, k2 = − 11
8.

© UCB EECS 16B, Fall 2019. All Rights Reserved. This may not be publicly shared without explicit permission. 3
Discussion 8A @ 2019-10-29 09:38:28-07:00

(e) Is the system now stable?

Answer: Yes
" # " #
1 1
(f) Suppose that instead of u(t) in (10), we had u(t) as the way that the discrete-time control
0 1
acted on the system. Is this system controllable from u(t)?
Answer: We calculate " #
h i 1 1
R2 = B AB =
1 1

Observe that R2 matrix is not full rank and hence our system is not controllable.
(g) For the part above, suppose we used [k1 , k2 ] to try and control the system. What would the eigenvalues
be? Can you move all the eigenvalues to where you want? Give an intuitive explanation of what is
going on.
Answer:
" # " # 
0 1 1 h i
~x(t + 1) =  + · k1 k2 ~x(t) (11)
2 −1 1
" # " #
0 1 k k
= + 1 2 ~x(t) (12)
2 −1 k1 k2

(13)

Finding the eigenvalues λ :

" #
k −λ k2 + 1 
det  1 =0 (14)
k1 + 2 k2 − 1 − λ

= (k1 − λ )(k2 − 1 − λ ) − (k1 + 2)(k2 + 1) (15)

2
= k1 (k2 − 1) − k1 λ − λ (k2 − 1) + λ − (k1 k2 + k1 + 2k2 + 2) (16)
= k1 k2 − k1 − k1 λ − λ k2 + λ + λ 2 − k1 k2 − k1 − 2k2 − 2 (17)
2
= λ + (1 − k1 − k2 )λ − 2(1 + k1 + k2 ) (18)
= (λ + 2)(λ − (1 + k1 + k2 )) (19)

We can see that the eigenvalue at λ = −2 cannot be moved, so we cannot arbitrarily change our
eigenvalues with this control input.

Contributors:

• Anant Sahai.

• Regina Eckert.

• Ioannis Konstantakopoulos.

• John Maidens.

© UCB EECS 16B, Fall 2019. All Rights Reserved. This may not be publicly shared without explicit permission. 4