Optimization Methods: HW1 Instructor: Lei Li, INS, Shanghai Jiao Tong University; Email: leili2010@sjtu.edu.cn All “Numerical optimization” means the second edition unless explicitly stated otherwise 1 Theory 1. • Find the tangent plane of the level set of the surface F (x, y, z) = 14 at P (5, −2, 3) where F (x, y, z) = xyz + x2 − 2y 2 + z 3 . • Suppose f (x, y) = ln(x2 + y 2 ). Compute the tangent plane of the graph at (1, 1, ln 2). Can you find a normal vector of this plane? Compute the tangent line of the level set f (x, y) = ln(2). Can you find a normal vector of this line? 2. The cone z 2 = x2 + y 2 and the plane 2x + 3y + 4z + 2 = 0 intersect in an ellipse. Find the tangent line of the ellipse at P (3, 4, −5). Find the plane that is normal to the ellipse at P (3, 4, −5). 3. Explain for which direction ⃗v (a unit vector), d ϕ(x + ⃗v ∆s) − ϕ(x) ϕ(x + s⃗v )|s=0 := lim ds ∆s ∆s→0+ is maximized. If this is maximized, what is the maximum value? Using this idea, justify formally (no need to be rigorous) why Z Z f (x)δ(ϕ(x) − c)|∇ϕ(x)| dx = f (x)dS(x), Rd S where δ is the one-dimensional Dirac delta, S is the level set ϕ(x) = c and dS is the surface element of S (the Hausdorff measure). If you are interested in this, you may compare this with the coarea formula. 4. Let A ∈ Rn×n be a square real matrix. Compute the Hessian of the function f (x) = xT Ax. 5. (a). Prove the Taylor theorems in lecture notes using the Taylor theorems for one-variable functions. (b). Give an example that the claim ∇f (x + p) = ∇f (x) + ∇2 f (x + tp)p in general is not true. 1 6. Problems 2.13, 2.14, 2.16 in the book “Numerical optimization”. 7. Problem 2.6, 2.7 in the book “Numerical optimization”. 8. Suppose that f : (a, b) → R is increasing and convex. Is the inverse function convex or concave? 9. Suppose D is open, if f is convex and differentiable in D, show that f (y) ≥ f (x) + ∇f (x) · (y − x). This means that the function is above the tangent plane. 10. Suppose that f is a convex function. Then, show that the Jensen’s inequality holds X X X f( λi xi ) ≤ λi f (xi ), λi ≥ 0, λi = 1 i i i As a remark, more generally, one has f (EX) ≤ Ef (X) Using this, one can show that for two probability densities p and q, the KL divergence satisfies Z p(x) p(x) log dx ≥ 0. q(x) 11. Suppose that f : Rn → R is continuous. Show that it is convex if and only if Z 1 1 f (x + λ(y − x))dλ ≤ [f (x) + f (y)], ∀x, y. 2 0 12. Suppose that f : Rn → R is a smooth convex function, and suppose that the set of global minimizers S is nonempty. Show that S is a closed and convex set. 13. Problem 2.9 in the book “Numerical optimization” (section 2.2). 14. Let S be the set of n×n symmetric matrices. Show that X 7→ σ1 (X) := max{|λ(X)|}, the largest singular value, is a convex function for X ∈ S. [Hint: show that this is a norm.] Extra to think: what if we consider the largest eigenvalue (with sign considered)? To answer this, one property is needed: the pointwise maximum of affine functions is again convex. 2 Q 15. Show the set {x ∈ Rn+ : ni=1 xi ≥ 1} is a convex set. (You may use the inequality aθ b1−θ ≤ θa + (1 − θ)b for θ ∈ (0, 1). ) 16. Let S ⊂ Rn be a subset of Rn . Let ∥·∥ be a norm. Let a > 0. Consider the set Sa := {x : d(x, S) ≤ a} and S−a := {x : B(x, a) ⊂ S}. Show that both sets are convex if S is convex. Bonus problem This is a bonus problem. The credits here will be added to your total homework scores (the total HW scores of the whole semester), without exceeding the maximum points. set. Let xi ∈ C. Suppose θi ≥ 0 and P∞Suppose that C is a convex P∞ θ = 1. Show that if x = i=1 i i=1 θi xi converges, then x ∈ C. (Note that C may not be closed.) Does this fact hold for general Banach spaces with infinite dimension? 3