Continuous optimization, an introduction

Assessment
(3rd January 2017)

Exercise I
We recall that for a convex function f : X → R ∪ {+∞},
1
proxτ f (x) = arg min f (y) + ky − xk2 .
y 2τ

Evaluate proxτ f (x) for τ > 0, and

1. X = R, f (x) = − ln x for x > 0, +∞ for x < 0.
2. f (x) = ψ(kxk) where ψ : R → R∪{+∞} is a convex, even (paire) function
with ψ(0) = 0. Show first that f is a convex function, then evaluate proxτ f
in terms of proxτ ψ .

3. f (x) = kxk3 /3.

Exercise II
We consider X a Hilbert space and a strictly convex lower-semicontinuous (lsc)
function ψ : X → R ∪ {+∞} such that the interior of dom ψ, denoted D, is not
empty, D = dom ψ, ψ ∈ C 1 (D) ∩ C 0 (D), and ∂ψ(x) = ∅ for all x 6∈ D. In other
words, ∂ψ(x) is either ∅ (if x 6∈ D), or a singleton {∇ψ(x)} (if x ∈ D). We also
assume that
lim ψ(x) = +∞.
kxk→∞

We define the “Bregman distance associated to ψ”, denoted Dψ (x, y), as,
for y ∈ D and x ∈ X,

Dψ (x, y) := ψ(x) − ψ(y) − h∇ψ(y), x − yi .

1. Show that Dψ (x, y) ≥ 0, and that Dψ (x, y) = 0 ⇒ y = x. What other

estimate can we write if in addition ψ is strongly convex? Why is Dψ not a
distance in the classical sense?

2 n n
Pn Dψ in case D = X, ψ(x) = kxk /2. In case X = R , D =]0, +∞[ ,
2. Express
ψ(x) = i=1 xi ln xi .

3. Let f : X → R ∪ {+∞} a proper, convex, lsc function. Let x̄ ∈ D. We

assume that there exists x ∈ D with f (x) < +∞. Show that there exists a
unique point x̂ ∈ X such that

f (x̂) + Dψ (x̂, x̄) ≤ f (x) + Dψ (x, x̄) ∀x ∈ X. (1)

4. Explain why ∂(f + ψ) = ∂f + ∂ψ. Write the first order optimality condition
for x̂. Deduce that x̂ ∈ D.

1
5. Show (from the first order optimality condition) that for all x ∈ X,

f (x) + Dψ (x, x̄) ≥ f (x̂) + Dψ (x̂, x̄) + Dψ (x, x̂). (2)

A “nonlinear” descent algorithm. We consider a minimisation problem

min f (x) + g(x), (P )

x∈D

for f, g convex, lsc, proper functions, where f is C 1 in D and g is “simple” in

the following sense: one assume that one knows how to solve
1
min g(x) + hp, xi + Dψ (x, y)
x τ
for any τ > 0, p ∈ X and y ∈ D. We suppose in addition that there exists
L > 0 such that for any y ∈ D, x ∈ X

Df (x, y) ≤ LDψ (x, y). (3)

(Here Df (x, y) = f (x) − f (y) − h∇f (y), x − yi.) We assume that the minimisa-
tion problem has a solution. We denote F (x) = f (x) + g(x).

6. Show that if ψ is 1-convex (strongly convex with parameter 1) and f has

L-Lipschitz gradient, then (3) is true.

Given x̄ ∈ D, τ > 0, we now define the following operator: we let x̂ = Tτ (x̄)

be the solution of the minimisation problem
1
min f (x̄) + h∇f (x̄), x − x̄i + g(x) + D(x, x̄). (4)
x∈D τ

7. Explain why this problem is easy to solve. Show that if τ is small enough,
one has the following descent rule: for all x ∈ X,
1 1
F (x) + Dψ (x, x̄) ≥ F (x̂) + Dψ (x, x̂).
τ τ

8. We define the following algorithm: we choose x0 ∈ D, and for all k ≥ 0, let

xk+1 = Tτ xk , where τ ≤ L is fixed. Show that for all k ≥ 0, F (xk+1 ) ≤ F (xk ).
If x∗ is a minimiser of F in D, show that
1
F (xk ) − F (x∗ ) ≤ Dψ (x∗ , x0 ).
kτ

9. We assume that F (x) → +∞ when kxk → +∞. Why can we find x̃ ∈ D

and extract a subsequence xkl such that xkl → x̃ as l → ∞? Why is x̃ a solution
of (P )?

2
Application: minimisation in the unit simplex. One considers the case
where X = Rd ,
( d
)
X
Σ = x ∈ X : xi ≥ 0 ∀i = 1, . . . , d; xi = 1
i=1

is the unit simplex and

(
0 if x ∈ Σ
g(x) =
+∞ else.
Pd
We choose ψ(x) = i=1 xi ln xi and D =]0, +∞[d .

10. Give the expression of Dψ (x, y) for x ∈ Σ, y ∈ Σ ∩ D.

11. Show that the algorithm described in the previous part is implementable:
express in detail the computation
P of the iterations. Hint: introduce the Lagrange
multiplier for the constraint i xi = 1.

Exercise III
We consider a maximal monotone operator A in a (real) Hilbert space X. We
consider also a “metric” M , that is, a continuous, coercive, and symmetric
operator:

kM xk ≤ kM kkxk ∀x ∈ X, hM x, xi ≥ δkxk2 , hM x, yi = hx, M yi

for all x, y ∈ X, where δ > 0.

1. Show that (x, y) 7→ hM x, yi =: hx, yiM defines a scalar product which is

equivalent to the scalar product h·, ·i. Show that for all y ∈ X, the problem
1
min kxk2M − hy, xi
x 2

has a unique solution. Deduce that M is invertible. We have denoted k.kM the
Hilbertian norm induced by the M -scalar product.

2. Show that (M −1 A) is a maximal monotone operator in the M -scalar prod-

uct. Deduce from Minty’s theorem that for any y ∈ X, there exists a unique x
such that
M (x − y) + Ax 3 0.

3. We consider A, B two maximal monotone operators and K ∈ L(X, X) a

continuous, linear operator in X. We define in X × X the metric, for τ, σ > 0,
−K ∗
 I 
M := τ .
I
−K σ

Here I ∈ L(X, X) is the identity operator. Show that if τ σ < 1/kKk2 , M is

continuous and coercive in X × X.

3
4. Deduce that (for such τ, σ) one can define the following algorithm: we let
(x0 , y 0 ) ∈ X ×X and define for each k ≥ 0 the new point (xk+1 , y k+1 ) as follows:

K∗
 k+1
− xk
  k+1  
Axk+1
  
x 0 x
M + + 3 0.
y k+1 − y k −K 0 y k+1 B −1 y k+1

Express this as a first iteration defining xk+1 from xk , y k and then an iteration
defining y k+1 from xk , xk+1 , y k .

5. In what case does (xk , y k ) converge? (and in what sense?) In this case,
what does the limit (x̄, ȳ) satisfy? Write, in particular, an equation for x̄.

6. We now consider a maximal monotone operator Cx and the new iterative

scheme:
K∗
 k+1
− xk
  k+1  
Axk+1
    k
x 0 x Cx
M + + 3 .
y k+1 − y k −K 0 y k+1 B −1 y k+1 0

Under which condition on τ, σ, C will this iterative scheme be converging? To

which limit?