Massachusetts Institute of Technology

6.854J/18.415J: Advanced Algorithms Friday, March 18, 2016

Ankur Moitra

Problem Set 6

Due: Wednesday, April 6, 2016 – 7 pm

Dropbox Outside Stata G5

Collaboration policy: collaboration is strongly encouraged. However, remember that

1. You must write up your own solutions, independently.

2. You must record the name of every collaborator.

3. You must actually participate in solving all the problems. This is difficult in very large
groups, so you should keep your collaboration groups limited to 3 or 4 people in a given
week.

4. Write each problem in a separate sheet and write down your name on top of every sheet.

5. No bibles. This includes solutions posted to problems in previous years.

1 Separation Oracles
Describe efficient separation oracles for each of the following families of convex sets. Here,
“efficient” means linear time plus O(1) calls to any additional oracles provided to you.

(a) The set A ∩ B, given separation oracles for A and B.

(b) The `1 ball: kxk1 ≤ 1.

(c) Any convex set A, given a projection oracle for A. A projection oracle, given a point x,
reports whether x is in A and if it is not additionally returns

arg min kx − yk2 .

y∈A

(d) The -neighborhood of a convex set A:

{x | ∃y ∈ A, kx − yk2 ≤ }

given a projection oracle for A.

2 Problem Set 6

2 Optimization with the Ellipsoid Method

Here, we will show a way to optimize a linear function over a convex set using the ellipsoid
algorithm. Recall that in class, we showed how to find a feasible point in a convex set given a
separation oracle, and then used it to optimize a linear function by doing binary search over
the optimum value, and turning the objective function into a linear constraint. Here we will
explore an alternative method that uses the ellipsoid method directly without performing
binary search. The basic idea is to optimize over the set A ∩ {x | cT x ≤ v} without actually
knowing the function value v.
To avoid going through the analysis of the ellipsoid method again, we will give a more
black-box description of the guarantees of the ellipsoid method. In particular, we will view
it as an interactive process, producing output and reading input in each round. The process
is initialized with a radius R. At the ith round, it outputs a point xi . It then reads as input
a direction di , and then goes on to round i + 1.
Let A e be any convex set contained in B(0, R) such that A e contains some ball of radius
r. Furthermore, suppose that for all i with xi 6∈ A,
e di is a hyperplane that separates xi from
2
e Then we proved for m = O(n log(R/r)), some xi , i ≤ m must satisfy xi ∈ A.
A. e Notably,
this holds despite the ellipsoid method having no explicit knowledge of A except for the
e
conditions on the vectors di . The idea is in some sense to change A e as we go along.
Now consider the following algorithm for optimization. We will assume a set A with a
separation oracle, a bounding radius R such that A ⊆ B(0, R), and a linear objective to be
minimized, cT x.

Algorithm 1 Ellipsoid-Optimize
1: Initialize the ellipsoid method with radius R. Set vbest to ∞.
2: for i from 1 to m do
3: Get xi from the ellipsoid method.
4: Use the separation oracle on xi , finding if xi ∈ A and a separating hyperplane if not.
5: if xi ∈ A then
6: if cT xi < vbest then
7: Set vbest to cT xi .
8: Set xbest to xi .
9: end if
10: Set di = c.
11: else
12: Set di to the separating direction returned by the oracle.
13: end if
14: Send the direction di back to the ellipsoid method.
15: end for
16: return xbest .

In other words, in each step we call the separation oracle for A, using the hyperplane it
Problem Set 6 3

returns if xi 6∈ A and a hyperplane defined by the objective c otherwise (i.e. xi ∈ A). After
m iterations we simply return the best objective value out of all the points we have seen in
A.
Prove that for any objective value v, as long as A ∩ {x | cT x ≤ v} contains a ball of radius
r, Ellipsoid-Optimize will return an x ∈ A satisfying cT x ≤ v. Note that this guarantee holds
without any knowledge of the value v.

3 Two Player Zero Sum Games

This problem proves the von Neumann theorem on zero-sum games mentioned in class.
In a 0-sum 2-player game, Alice has a choice of n so-called pure strategies and Bob has
a choice of m pure strategies. If Alice picks strategy i and Bob picks strategy j, then the
payoff is aij , meaning aij dollars are transfered from Alice to Bob. So Bob makes money if
aij is positive, but Alice makes money if aij is negative. Thus, Alice wants to pick a strategy
that minimizes the payoff while Bob wants a strategy that maximizes the payoff. The matrix
A = (aij ) is called the payoff matrix.
It is well known that to play these games well, you need to use a mixed strategy—a
random choice from among pure strategies. A mixed strategy is just a particular probability
distribution over pure strategies: you flip coins and then play the selected pure strategy. If
Alice has mixed strategy x, meaning he plays strategy i with probability xi , and Bob has
mixed strategy y, then it is easy to prove that the expected payoff in the resulting game is
xAy. Alice wants to minimize this expected payoff while Bob wants to maximize it. Our
goal is to understand what strategies each player should play.
We’ll start by making the pessimal assumption for Alice that whichever strategy she
picks, Bob will play best possible strategy against her. In other words, given Alice’s strategy
x, Bob will pick a strategy y that achieves maxy xAy. Thus, Alice wants to find a distribution
x that minimizes maxy xAy. Similarly, Bob wants a y to maximize minx xAy. So we are
interested in solving the following 2 problems:

Pmin Pmax xAy

xi =1 yj =1
max Pmin xAy
P
yj =1 xi =1

Unfortunately, these are nonlinear programs!

(a) Show that if Alice’s mixed strategy is known, then Bob has a pure strategy serving as
his best response.

(b) Show how to convert each program above into a linear program, and thus find an optimal
strategy for both players in polynomial time.

(c) Give a plausible explanation for the meaning of your linear program (why does it give
the optimum?)
4 Problem Set 6

(d) Use strong duality (applied to the LP you built in the previous part) to argue that the
above two quantities are equal.

The second statement shows that the strategies x and y, besides being optimal, are
in Nash Equilibrium: even if each player knows the other’s strategy, there is no point in
changing strategies. This was proven by Von Neumann and was actually one of the ideas
that led to the discovery of strong duality.

4 Weak Duality for Nonlinear Programs

Here, we will look at a simple class of second-order cone programs, which are nonlinear
convex programs. Your task will be to show that a form of weak duality still holds for these
programs.
The below convex programs are defined for an n-dimensional vector c, an m × n matrix
A, an l × n matrix B, and an m-dimensional vector b.
The primal problem is to find an n-dimensional vector x:

minimize cT x
x
subject to Ax ≥ b,
kBxk2 ≤ 1.

This is similar to a linear program except for the extra `2 ball constraint kBxk2 ≤ 1.
The dual problem is to find an m-dimensional vector y and an l-dimensional vector z:

maximize bT y − kzk2
x
subject to AT y + B T z = c,
y ≥ 0.

Prove that for any primal and any dual solution, the objective value of the primal is at
least the objective value of the dual solution (i.e. weak duality).

5 Two Definitions of Submodular

Let N be a set (the “universe”) and F a function mapping subsets of N to real numbers.
There are two different standard definitions of what it means for F to be submodular :
(i) For all A, B,
F (A ∩ B) + F (A ∪ B) ≤ F (A) + F (B).

(ii) For all A ⊆ B, j 6∈ B

F (A ∪ {j}) − F (A) ≥ F (B ∪ {j}) − F (B).

Problem Set 6 5

Your task is to prove these definitions are equivalent:

(a) Show that (i) implies (ii).

(b) Show that (ii) implies (i).

6 Fun with Lovasz Extensions

Let F (S) be a submodular function. Recall that the Lovasz extension of F , f (x) is defined
for nonnegative vectors x as
Z ∞
f (x) = F (∅) + (F ({i : xi ≥ a}) − F (∅)) da.
0

n
Note that this definition applies to all x ∈ R+ , not just x ∈ [0, 1]n .
We will be looking at an interesting optimization problem involving the Lovasz extension:
1
b = arg min f (x) + x2 .
x
x≥0 2

Recall that f is not necessarily monotonically increasing. x

b will always be unique (due to
the “strong convexity” of the objective).
The aim will be to relate xb to the minimizers of Fa (S) = F (S) + a|S|. Note that each
Fa is itself a submodular function.

(a) Show that although Fa (S) may have multiple minimizers, there is a unique set Sa that
minimizes Fa (S) and has a higher cardinality than any other minimizer.

(b) Show that for all b > a, Sb ⊆ Sa .

(c) For any positive vector x, define

Xa = {i : xi ≥ a}.

Show that Z ∞
1
f (x) + x2 = F (∅) + (F (Xa ) − F (∅) + a|Xa |) da.
2 0

(d) Show that for any a > 0, {i : x

bi ≥ a} = Sa .
This means that solving one convex optimization problem lets us minimize all the sub-
modular functions Fa simultaneously!
You may assume without proof that the optimization problem defining x
b has a unique
optimum.