Practice Exam 1: Continuous Optimisation

Transcription

1 Practice Exam : Continuous Optimisation. Let f : R m R be a convex function and let A R m n, b R m be given. Show that the function g(x) := f(ax + b) is a convex function of x on R n. Suppose that f is strictly convex. Show that then g(x) := f(ax+b) is strictly convex if and only if A has (full) rank n. Hint: Recall that f is strictly convex if for any x y, 0 < λ < it holds: f(λy + ( λ)y 2 ) < λf(y ) + ( λ)f(y 2 ). [3 points] [ points] For x, y R n, λ [0, ] we find: g(λx + ( λ)y) = f(a(λx + ( λ)y)b) = f(λax + ( λ)ay + λb + ( λ)b) = f(λ(ax + b) + ( λ)(ay + b)) f is convex λf(ax + b) + ( λ)f(ay + b) = λg(x) + ( λ)g(y) rank(a) = n implies: x y Ax Ay. As in for x y, λ (0, ) we obtain: g(λx + ( λ)y) = f(λ(ax + b) + ( λ)(ay + b)) f is strict convex, Ax + b Ay + b < λf(ax + b) + ( λ)f(ay + b) = λg(x) + ( λ)g(y) Assume rank(a) < n. hen there exist x y with Ax = Ay and for any λ (0, ) we obtain: g(λx + ( λ)y) = f(λ(ax + b) + ( λ)(ay + b)) = f(ax + b) g(x) = g(y) = g(x) = λg(x) + ( λ)g(y). So g is not strictly convex. 2. For given S R n we define the convex hull conv(s) by { m m conv(s) = x = λ i x i : m N, λ i = ; x i S, λ i 0 i Show that conv(s) is the smallest convex set containing S: Show that the set conv(s) is convex with S conv(s). [3 points]

2 est exam: Continuous Optimisation 205 Monday th December 205 Show that for any convex set C containing S we must have conv(s) C. (Hint: You may use without proof any Lemma/heorem/Corollary from the course, except for heorem.9.) [3 points] For all x S, letting m = and λ = we see that x conv(s), and thus S conv(s). o prove convexity we consider arbitrary x, y conv(s), λ [0, ]. hen there exists p, q N and x,..., x p S and y,..., y q S and θ R p + and ψ R q + such that x = θ i x i, = θ i, y = ψ i y i, = ψ i. We then have λθ i + λx + ( λ)y = λθ i x i + ( λ)ψ i y i, λθ i 0 for all i =,..., p, ( λ)ψ i 0 for all i =,..., q, ( λ)ψ i = λ θ i + ( λ) ψ i = λ + ( λ) =. herefore λx + ( λ)y conv(s), and as x, y conv(s), λ [0, ] were arbitrary, this implies that conv(s) is a convex set. Consider an arbitrary convex set C R n such that S C. For m N let S m = { x = m λ i x i : m λ i = ; x i S, λ i > 0 i. We then have conv(s) = m= Sm. We shall show by induction that S m C for all m, and thus conv(s) = m= Sm C. For m = we have S = S C. Now suppose for the sake of induction that S m arbitrary x S m+. C and consider an We have x = m+ λ ix i for some x i S, λ i > 0 for all i and m λ i =. Letting θ R m + such that θ i = λ i /( λ m+ ) and letting y = m θ ix i we have m θ i = and thus y S m C by the induction hypothesis. We then have y C and x m+ S C and λ m+ [0, ] and x = ( λ m+ )y + λ m+ x m+. As C is a convex set, this implies that x C. 2

3 est exam: Continuous Optimisation 205 Monday th December Consider with 0 c R n the program: (P ) min x R n c x s.t. x x. Show that x = c is the minimizer of (P) with minimum value v(p ) = c. c ( x means here the Euclidian norm.) Compute the solution y of the Lagrangean dual (D) of (P). Show in this way that for the optimal values strong duality holds, i.e., v(d) = v(p ). [2 points] [ points] Either show this by a sketch. Or as follows (using Schwarz inequality): x implies: c x c x c, and c x = c holds iff x = c c So x = c c is the minimizer with v(p ) = c ( c c ) = c. (Alternatively find x by solving the KK-conditions.) he dual (D) is given by (D) max y 0 ψ(y) where ψ(y) := min L(x, y) x Rn with Lagrangean function L(x, y) = c x + y(x x ). We find for y = 0: ψ(0) =. for y > 0: he minimizer x of ψ(y) satisfies x L(x, y) = c + 2yx = 0 or x = c. So (fill in) 2y ψ(y) = 2y c c + y c c y = y c c y. he Lagrangian dual problem is thus max y>0 { y c c y. o find an (unconstrained) maximizer of ψ(y) for y > 0 we solve 0 = ψ (y) = y 2 c c with solution y = 2 c. So v(d) = ψ(y) = c = v(p ).. Consider the problem (in connection with the design of a cylindrical can with height h, radius r and volume at least 2π such that the total surface area is minimal): (P ) : min f(h, r) := 2π(r 2 + rh) s.t. πr 2 h 2π, (and h > 0, r > 0) 3

4 est exam: Continuous Optimisation 205 Monday th December 205 Compute a (the) solution (h, r) of the KK conditions of (P). Show that (P ) is not a convex optimization problem. Show that the solution (h, r) in is a local minimizer. Why is it the unique global solution? Hint: Use the sufficient optimality conditions [ points] [3 points] We first note that the functions f(h, r) = 2π(r 2 + rh) and g(h, r) := πr 2 h + 2π are not convex (for h > 0). For the objective function f, e.g., this follows by: r 0 f = 2π, 2 f = 2π 2r + h 2 By Lemma.2 we thus have that 2 f is not positive semidefinite, thus f is not a convex function and we do not have a convex optimisation problem. We now consider the KK condition: ( f = µ g, g 0, µ g = 0) So consider: 2π ( r 2r+h = µπ r 2 2rh) ( ): Case µ = 0: leads to 2π ( r 2r+h) = 0 with solution (h, r) = (0, 0) which is not feasible. Case µ > 0 and thus g = 0, or equivalently πr 2 h = 2π: he 2 equations in ( ) lead to µ = 2/r and then 2(2r+h) = 2 2rh or h = 2r. r By using the (active) constraint we find πr 2 h = 2πr 3 = 2π with solution r =. So the unique KK solution is given by (h, r) = (2, ), µ = 2. (We apply the second order sufficient conditions of h. 5. to the nonconvex program (P)). So we will show (for the set C = {d R n : f d 0, g d 0): d 2 h,rl(h, r, µ)d > 0 d C \ {0 ( ) We compute f(h, r) = 2π, g(h, r) = π, L(h, r, µ) = 2π + 2( π) = 2π and C = {d R 2 f(h, r) d 0, g(h, r) d 0 { = d R 2 d 0, d 0 { = λ λ R

5 est exam: Continuous Optimisation 205 Monday th December 205 For d = λ(, ) 0, (i.e., λ 0) we obtain (see ( )): 0 λ(, )( 2π) λ =... = 2λ 2 π6 > 0 λ 0. 2 So (h, r) = (2, ) is a local minimizer. It is the unique (global) minimizer since the point is the only KK point. Note that since the linear independency constraint qualification holds ( g = π ( r 2rh) 2 0, for r, h > 0) any local minimizer must be a KK point. Also note that for feasible (h, r) also f holds. (o show the latter fact is technically involved and was not expected to be done.) 5. Consider the closed set K = {x R 2 x + 2x 2 0 and 3x + x 2 0 Prove that K is a proper cone. [You may assume closure.] Find the dual cone to K. [5 points] [ point] In order for a set to be a proper cone it most be a closed, convex, pointed full-dimensional cone. We will assume closure and prove the rest: Convex cone: Consider an arbitrary x, y K and λ, λ 2 > 0. From heorem 7.2, if we can show that λ x + λ 2 y K then we are done. We have his implies that x + 2x 2 0, 3x + x 2 0, λ > 0, y + 2y 2 0, 3y + y 2 0, λ 2 > 0. (λ x + λ 2 y) + 2(λ x + λ 2 y) 2 = λ (x + 2x 2 ) + λ 2 (y + 2y 2 ) 0, 3(λ x + λ 2 y) + (λ x + λ 2 y) 2 = λ (3x + x 2 ) + λ 2 (3y + y 2 ) 0. herefore λ x + λ 2 y K. Full-dimensional: Using Definition 7.8, part 2, this follows from the space being two dimensional and having two linearly independent 0 vectors, K. 0 5

6 est exam: Continuous Optimisation 205 Monday th December 205 Pointed: We will consider an arbitrary x R 2 such that ±x K. Using Definition 7.7, if we can then show that x = 0 then we are done. We have (x) + 2(x) 2 0 x + 2x 2 = 0, ( x) + 2( x) 2 0 3(x) + (x) 2 0 3x + x 2 = 0. 3( x) + ( x) 2 0 herefore x = 2 (3x 5 + x 2 ) {{ (x 5 + 2x 2 ) {{ =0 =0 = 0, x 2 = (3x + x 2 ) {{ =0 3 x {{ = 0. =0 From heorem 8.7 we have { K = cl conic, 2 { 3 = conic, We will consider bounds to the optimal value of the following problem: min x 5x 2 x x 2 2x + x s.t. x 2 + 5x 2 2 x x 2 8x 2 = x R 2. (A) (c) Give a finite upper bound on the optimal value of problem (A). Formulate a positive semidefinite optimisation problem to give a lower bound on the optimal value of problem (A). Give the dual problem to the positive semidefinite optimisation problem you formulated in part of this question. [ point] [2 points] [ point] o find an upper bound we can use any feasible point, x. If we limit our search for a feasible point such that x 2 = 0 then we would have a feasible point if and only if = x x = x 2. herefore both x = (2, 0) and x = ( 2, 0) are feasible points. We only need one point to give us an upper bound, and if we consider the feasible point x = (2, 0) 6

7 est exam: Continuous Optimisation 205 Monday th December 205 then this gives us the upper bound of 5 x 2 x x 2 2 x + x = = = 8 Problem (A) is equivalent to 5 2 min, xx 2x x s.t., xx 8x = x x xx PSD 3 x R 3, A lower bound on this is then provided by solving the optimisation problem 5 2 min, X 2x x,x s.t., X 8x = x PSD 3. (c) his previous optimisation problem is equivalent to 2 0 min 5 2 x0 x, x,x s.t. 0 2 x0 x, = x0 x, =, x0 x PSD 3. We then have that the dual problem is max y s.t. y + y y 0 2 y PSD

8 est exam: Continuous Optimisation 205 Monday th December (Automatic additional points) [ points] Question: otal Points: A copy of the lecture-sheets may be used during the examination. Good luck! 8