Optimization and Algorithms
February 6, 2023
Exam
1. Deconfiicted trajectories. A trajectory T of duration T in Rd is a sequence of T points in Rd, denoted as T = {x(1), x(2), . . . , x(T)}, with x(t) ∈ Rd for 1 ≤ t ≤ T. Note that t denotes discrete-time; thus t is an integer (such as t = 0, 1, 2, 3. . . .).
Let T1 = {x1 (1), x1 (2), . . . , x1 (T)} and T2 = {x2 (1), x2 (2), . . . , x2 (T)} be two tra- jectories of duration T in Rd. We say that T1 and T2 are space-decon丑icted if Ⅱx1 (t) - x2 (s)Ⅱ2 > ∈ for 1 ≤ t, s ≤ T, where ∈ is a given positive number. We say that T1 and T2 are time-decon丑icted if Ⅱx1 (t) - x2 (t)Ⅱ2 > ∈ for 1 ≤ t ≤ T.
Consider the following two controlled dynamic linear systems. The state of system
1 at time t is denoted by x1 (t) ∈ Rd , for 1 ≤ t ≤ T and obeys the recursion
x1 (t + 1) = A1x1 (t) + B1u1 (t), 0 ≤ t ≤ T - 1,
where A1 ∈ Rd×d and B1 ∈ Rd×p are given matrices, x1 (0) ∈ Rd is a given initial state and u1 (t) ∈ Rp is the control input of system 1 at time t, for 0 ≤ t ≤ T - 1. Note that the trajectory T1 depends on the inputs u1 (t), 0 ≤ t ≤ T - 1.
Similarly, for system 2 we have
x2 (t + 1) = A2x2 (t) + B2u2 (t), 0 ≤ t ≤ T - 1.
Note that the trajectory T2 depends on the inputs u2 (t), 0 ≤ t ≤ T - 1.
Finally, let Tref = {r(1), r(2), . . . , r(T)} be a given, fixed reference trajectory of duration T in Rd.
We want to design the control inputs u1 (t) (0 ≤ t ≤ T - 1) and u2 (t) (0 ≤ t ≤ T - 1) so that:
• the final state x1 (T) of system 1 is as close as possible to a given, desired state p1 ∈ Rd;
• the final state x2 (T) of system 2 is as close as possible to a given, desired state p2 ∈ Rd;
• the trajectories T1 and T2 are time-deconflicted;
• the trajectories T1 and Tref are space-deconflicted;
• the trajectories T2 and Tref are space-deconflicted.
One of the following problem formulations is suitable for the given context.
Which one?
Write your answer (A, B, C, D, E, or F) in the box at the top of page 1
2. Unconstrained optimization. Consider the optimization problem
The point x★ = 0 is a global minimizer of (7) for one of the following choices of a:
(A) a = -2
(B) a = -1
(C) a = 0
(D) a = 1
(E) a = 2
(F) a = 3
Which one?
Write your answer (A, B, C, D, E, or F) in the box at the top of page 1
Hint: the numerical values log(2) ≈ 0.7 and log(3) ≈ 1.1 might be useful
3. Gradient descent algorithm. Consider the function f : R2 → R given by f (a, b) = 2/1a2+(a−b)
2
. Suppose we do one iteration of the gradient descent algorithm (applied to f) starting from the point
and using the stepsize 1.
Which of the following points is the next iteration x1 ?
Write your answer (A, B, C, D, E, or F) in the box at the top of page 1
4. Signal-denoising as a least-squares problem. Consider the function f : Rn → R, f (r) = rT Dr, where D is a given n × n diagonal matrix with positive diagonal entries:
with di > 0 for 1 ≤ i ≤ n.
Consider the following optimization problem
where the variables to optimize are s ∈ Rp and v ∈ Rn; the matrix A ∈ Rn×p and the vectors y ∈ Rn , and s ∈ Rp are given. This problem can be interpreted as a signal-denoising problem: we observe y and want to decompose it as the sum of a signal of interest s and noise v; we know that s should be close to the nominal signal s and that v should be close to zero (the larger the di, the more confident we are that the component vi should be close to zero).
Problem (8) can be reduced to a least-squares problem involving only the variable s, that is, it can be reduced to a problem of the form.
for some matrix A and vector β .
Give A and β in terms of the constants D , y , A, and s.
5. A simple optimization problem. Consider the function f : R2 → R, f(x) = 2/1xTMx,
where
The constants a and b satisfy 0 < a < b.
Solve in closed-form. the optimization problem
where 1 denotes the vector 1 = (1, 1).
6. A convex optimization problem. Consider the following optimization problem
where the variables to optimize are xi ∈ R, for 1 ≤ i ≤ n. The vectors a i ∈ Rp , 1 ≤ i ≤ n and b ∈ Rp are given. The constants ci, 1 ≤ i ≤ n, are also given. The function g : R → R is defined as follows:
Show that (11) is a convex optimization problem.
7. A convex function based on a worst-case representation. Show that the function f : R → R,
f (x) = max {Ⅱ(a + u)x — bⅡ2 : ⅡuⅡ2 = r} (12)
is convex, where the vectors a, b ∈ Rn and the constant r > 0 are given.
In words: f takes as input a number x and returns as output the largest value of the expression
Ⅱ(a + u)x — bⅡ2
as u ranges over the sphere centered at the origin and with radius r.