Algorithmic Game Theory

Due: 12 April 2015

Homework 1
Lecturer: Michal Feldman

Assistant: Ophir Friedler

Routing Games

Problem 1 : (Non-atomic selfish routing).

In this problem we consider non-atomic selfish routing networks with one source, one sink,
one unit of selfish traffic, and affine cost functions (of the form ce (x) = ae x+be for ae , be 0).
In parts (a)-(c), we consider the objective of the maximum cost incurred by a flow f :
X
max
ce (fe )
P :fP >0

eP

The price of anarchy is then defined in the usual way, as the ratio between the maximum
cost of an equilibrium flow and that of a flow with minimum-possible maximum cost. (Of
course, in an equilibrium flow, all traffic incurs exactly the same cost; this is not generally
true in a non-equilibrium flow.)
(a) Prove that in a network of parallel links (each directly connecting the source to the
sink), the price of anarchy with respect to the maximum cost objective is 1.
(b) Prove that the price of anarchy with respect to the maximum cost objective can be as
large as 4/3 in general networks (with affine cost functions, one source and one sink).
(c) Prove that the price of anarchy with respect to the maximum cost objective is never
larger than 4/3 (in networks with affine cost functions, one source and one sink). [Hint:
try to reduce this to facts you already know.]
(d) A flow that minimizes the average cost of traffic generally routes some traffic on costlier
paths than others. Prove that the ratio between the cost of the longest used path and
that of the shortest used path in a minimum-cost flow is at most 2 (in networks with
affine cost functions, one source and one sink). Prove that this bound can be achieved.
Problem 2: (Atomic selfish routing)
(a) Consider an atomic selfish routing game in which all players have the same source vertex
and sink vertex (and each controls one unit of flow). Assume that edge cost functions

1-1

are nondecreasing, but do not assume that they are affine. Prove that a (pure-strategy)
Nash equilibrium (i.e., an equilibrium flow) can be computed in polynomial time. [Hint:
Remember the potential function. You can assume without proof that the minimumcost flow problem can be solved in polynomial time. If you havent seen the min-cost
flow problem before, you can read about it in any book on combinatorial optimization.
Be sure to discuss the issue of fractional vs. integral flows, and explain how (or if) you
use the hypothesis that edge cost functions are nondecreasing.]
(b) Prove that in an atomic selfish routing network of parallel links, every equilibrium flow
minimizes the potential function.
(c) Show by example that (b) does not hold in general networks, even when all players have
a common source and sink vertex.

Potential Games

Problem 3: (Better Replies in Potential Games)

An generalized ordinal potential function () is a function that satisfies for all i:
ui (i0 , i ) ui () > 0 (i0 , i ) () > 0
() is an ordinal potential function if for all i:
ui (i0 , i ) ui () > 0 (i0 , i ) () > 0
A better reply sequence in an n-player game G is a sequence (finite or infinite) of strategy
profiles s0 , s1 , . . . , st , . . . where s0 is arbitrary, and for each t > 0, st is obtained from st1
by a single player i switching his strategy to a better one. I.e. for each t > 0 there exists i
t1
such that sti = si
and ui (st ) > ui (st1 ). We say that all better reply dynamics converge
on G if there are no infinite better reply sequences.
(a) Show that there exists an ordinal potential game with n players, where each player has
only two strategies, that has better-reply sequences of length 2n .
(b) Prove that G has a generalized ordinal potential game if and only if all better reply
dynamics on G converge.
(c) Prove or show counter example: G has an ordinal potential game if and only if all better
reply dynamics on G converge.

1-2

Network Formation Games

Problem 4: (Network formation with path-length objectives)

In this question, we address network formation games with length objectives (this model
corresponds to real-life settings such as communication protocols that assume that a message
passes a pre-defined length before reaching its destination; examples are onion routing, and
proof-of-work protocols.)
Let G = (V, E) be a directed graph, s V a source node, and let the cost of every edge be
1. Consider a network formation game with n agents, where the objective of player i is to
form a path of length li originating at s. Each agent pays for every edge in her path a cost
proportional to her usage of the edge. Observe that every agent may use an edge multiple
times, i.e., the paths are multi-sets. Formally, given the selected
P = (P 1 , . . . , P n ),
Pn paths
i
i
i
let P (e) be number of times e appears in P , and let P (e) = i=1 P (e), then the cost of
player i is:
costi (P ) =

X P i (e)
P (e)

ePi

(a) Does this game admit a potential function? If it does, specify it, and if it doesnt prove
so. [Hint: What does the existence of a potential imply?]
(b) Establish the best lower and upper bounds you can on the PoA.
(c) Does the social optimum always correspond to a Nash equilibrium of the game?

1-3