First-Stage PhD Comprehensive Examination
in
                   CONTINUOUS OPTIMIZATION
     Department of Combinatorics and Optimization, Faculty of Mathematics,
                           University of Waterloo
             MC 5479, Thursday, June 13, 2019, 1pm – 4pm (3 hours)
                 Examiners: Levent Tunçel and Stephen A. Vavasis
1. Consider the univariate function
                                                                                        ex    for x < 0,
                               f (x) :=
                                              x + 1 for x ≥ 0.
  (a) Show that this function is convex. Any standard theorem that characterizes con-
  vexity of functions may be used.
  (b) Show that the gradient of this function is Lipschitz continuous, and find L, the
  Lipschitz constant of the gradient.
  (c) Since f 0 is Lipschitz continuous, Zoutendijk’s theorem for minimizing f using steep-
  est descent is applicable. State Zoutendijk’s theorem and the conclusion as it applies
  to this function f .
  (d) Suppose that xk = 0 on the kth iteration of the steepest descent method. Identify
  an interval of positive width of choices for xk+1 that satisfy the Wolfe conditions.
  (Select any reasonable values for the constants appearing in Wolfe’s conditions, and
  state what values you used.)
                                              1
2. Let n be a positive integer.
   (a) Assume f : Rn → (−∞, +∞] is positively 1-homogeneous. Show that it is subad-
       ditive if and only if it is convex.
  (b) Recall, if f : Rn → R satisfies
        (i) f (x) > 0, ∀x ∈ Rn \ {0},
       (ii) f (u + v) ≤ f (u) + f (v), ∀u, v ∈ Rn ,
      (iii) f (λx) = |λ|f (x), ∀x ∈ Rn , ∀λ ∈ R,
       then f is a norm on Rn . Let
                                   F := {f : f is a norm on Rn } ,
       and
               C := {C ⊂ Rn : C is compact, convex, 0 ∈ int (C) and C = −C} .
       Prove that (with γ being the gauge function)
                           f (·) := γ(C; ·) and C := {x ∈ Rn : f (x) ≤ 1}
       define a one-to-one correspondence between F and C.
                                              2
3. Let n be a positive integer and f : Rn → (−∞, +∞] be convex.
  (a) Define the notion of subgradient of f at x̄ ∈ Rn .
  (b) Define the notion of subdifferential of f at x̄ ∈ Rn .
  (c) Let f (x) := ||x||∞ . What is the subdifferential of f at a given x̄ ∈ Rn ? Prove your
      claims.
      For the remaining parts of this question, let m and n be positive integers such that
      n ≥ m + 1, and consider linear programming problems in the form
                                           p∗ := min c> x
                                                 s.t. Ax = 0
                                  (P)
                                                      e> x = n
                                                      x ∈ Rn+ ,
      where A is a given full row rank m-by-n matrix, c ∈ Rn is also given. e ∈ Rn is
      the vector of all ones. Assume that Ae = 0, p∗ = 0 and c> e > 0. In your answers,
      you may use the Fundamental Theorem of LP, provided you state it clearly and
      correctly. For every q ∈ R++ , define
                               (q + 1) ln c> x − ln (minj {xj }) , if x ∈ Rn++
                                                               φq (x) :=
                               +∞,                                 otherwise,
      and consider the optimization problem
                                          v(Pq ) := inf φq (x)
                                (Pq )               s.t. Ax = 0
                                                         e> x = n.
  (d) Prove that for every q ∈ R++ , v(Pq ) = −∞ (i.e., (Pq ) is unbounded). Further
      prove that for every q ∈ R++ , (P) and (Pq ) are equivalent (by stating a suitable
      definition of equivalence and proving it).
      Hint:
        • (P) has optimal solution(s),
        • (Pq ) has feasible sequences in the domain of φq which certify unboundedness
          of (Pq ).
  (e) Let f : Rn++ → R be defined by f (x) := − ln (minj {xj }). What is the subdifferen-
      tial of f at a given x̄ ∈ Rn++ ? Prove your claims.
                                           3
4. Let a1 , . . . , an ∈ Rd be given, and consider the following optimization problem in which
   x1 , . . . , xn ∈ Rd are the unknowns (i.e., nd total variables), λ > 0 is a fixed parameter,
   and all norms are Euclidean:
                                      n
                                   1X                      X
                      min f (x) :=       kxi − ai k2 + λ         kxi − xj k.
                     x1 ,...,xn    2 i=1                 1≤i<j≤n
   (Note: This formulation arises in “sum-of-norms” clustering.)
   (a) What is the definition of “strongly convex”? Suppose g, h : Rn → R are two convex
   functions such that g is strongly convex. Prove that g + h is also strongly convex.
   (b) Apply the result in (a) to the function at hand to conclude that f is strongly
   convex. Justify your answer.
   (c) By introducing auxiliary variables, rewrite the problem of minimizing f as a con-
   strained optimization in standard conic form min cT x subject to Ax = b, x ∈ K (not
   necessarily the same x), where the convex cone K is a Cartesian product of second-
   order cones. The following standard trick may help: the convex quadratic constraint
   s ≥ kxk2 /2 may be expressed in second-order cone form:                                                         
                               1                 1
                                 (s + 1) ≥ x; √ (s − 1) ,
                               2                  2
   where the notation on the right-hand side indicates concatenation of a vector and
   scalar.