Project Number 001907

DELIS
Dynamically Evolving, Large-scale Information Systems

Integrated Project
Member of the FET Proactive Initiative Complex Systems

Deliverable D4.1.1

A catalogue of open problems in Game Theory

w.r.t computational efficiency
Start date of the project: January 2004

Duration: 48 months

Project Coordinator: Prof. Dr. math. Friedhelm Meyer auf der Heide
Heinz Nixdorf Institute, University of Paderborn, Germany

Due date of deliverable: September 2004

Actual submission date: Novembver 2004

Dissemination level: PU – public

Work Package 4.1: A catalogue of open problems in Game Theory

w.r.t computational efficiency

Participants: University of Cyprus (UCY), Cyprus

Authors of deliverable: Marios Mavronicolas (mavronic@turing.cs.ucy.ac.cy)

 Abstract

We survey some recent progess in several significant areas of Game Theory that are
currently under intense investigation by the Algorithms and Complexity community. We
consider areas such as Computation of Nash Equilibria, Selfish Routing Games and Network
Games. For these areas, we provide a state-of-the art survey which overly emphasizes
the numerous open problems and the major challenges now in front of the community of
Algorithms and Complexity.
1 Introduction

Much of the current research activity in Algorithmic Game Theory is centered around Nash
equilibrium [29] – the probably most important solution concept in Game Theory [31]. Nash
equilibrium may be viewed to represent a steady state of the play of a strategic game in which
each player holds an accurate opinion about the (expected) behavior of other players and acts
rationally. Despite the apparent simplicity of the concept, computation of Nash equilibria in
finite games has been long observed to be difficult (cf. [25, 40]); in fact, it is arguably one of the
few, most important algorithmic problems for which no polynomial-time algorithms are known.
Indeed, Papadimitriou [33, p. 1] (see also [34]) actively advocates the problem of computing
Nash equilibria as one of the most significant open problems in Theoretical Computer Science
today:

”The Nash equilibrium (definition omitted here) is the predominant concept of ratio-
nality in Game Theory; it is also a most fundamental computational problem whose
complexity is wide open: Is there a polynomial algorithm which, given a two-person
game with a finite strategy space, computes a mixed Nash equilibrium? Together
with factoring, the complexity of finding a Nash equilibrium is in my opinion the
most important concrete open question on the boundary of P today.”

In this work, we survey several recent results on the computational complexity of Nash
equilibria in the context of games modeling selfish routing over non-cooperatice networks. We
assume some familiarity of the reader with the basic vocabulary of Game Theory [31].

Our survey distinguishes between pure Nash equilibria, where each player chooses exactly
one strategy (with probability one), and mixed Nash equilibria, where the choices of each player
are modeled by a probability distribution over strategies. The celebrated result of Nash [29],
which follows from Kakutani’s Fixed Point Theorem [15], assures that a mixed, but not necessar-
ily pure, Nash equilibrium always exists. In a fully mixed Nash equilibrium [24], all probabilities
are strictly positive. A generalized fully mixed Nash equilibrium [6] is a generalization of the
fully mixed Nash equilibrium where there is a set of ”empty” strategies and all probabilities of
users over this set of strategies are zero.

2 Computing Nash Equilibria

In this section, we consider the problem of computing Nash equilibria for general games. (Work
on computing Nash equilibria for specific games will appear in later sections.)

1
 Extensive surveys of algorithms and techniques from the literature of Game Theory for the
computation of Nash equilibria of general bimatrix games in either strategic or extensive form
appear in [25, 40]; see also [28, Section 3.1]. All known such algorithms incur exponential run-
ning time, with the seminal algorithm of Lemke and Howson [20] being the prime example; see
also [35, 39] for still inefficient extensions. Issues of computational complexity for the computa-
tion of Nash equilibria in general games have been raised by Megiddo and Papadimitriou [26],
and Papadimitriou [32]. The N P-hardness of computing a Nash equilibrium of a general bi-
matrix game with maximum payoff has been established by Gilboa and Zemel [11]. The major
remaining open problem concerns the computation of Nash equilibria in a general game.

Open Problem 2.1 Determine the complexity of computing a Nash equilibrium for a strategic
game.

3 The KP Model for Selfish Routing

3.1 The KP Model versus the W Model

The KP model was proposed only recently in the context of studying the effects of selfish
traffic over the Internet. In contrast, the W model dates back to the 1950s, when it was used
for studying the economics of transportation networks. However, much recent work on selfish
routing (see [37] and references therein) has adopted the W model and witnessed a revival of
interest into it.

The two models differ with respect to their assumptions about the following parameters:

• The structure and topology of the underlying network they consider; however, in both
cases, two distinguished nodes, the source and the destination, are considered.

• The splittability or unsplittability of the selfish traffic; unsplittable traffic is routed all
together along a single path, while splittable traffic may split into infinitesimal pieces.

• The type of equilibria (pure or mixed) they consider.

• The definition of (Expected) Individual Cost they adopt; these are used for defining Nash
equilibria and Wardrop equilibria, respectively.

• The definitions of Social Cost and Optimum – a performance measure for (Nash or
Wardrop) equilibria, and an optimality measure for assignments (not necessarily equi-
libria), respectively.

2
 In either model, these two definitions give rise to Coordination Ratio, also known as Price
of Anarchy [34]; this is the worst-case ratio of Social Cost over Optimum. A worst-case
Nash equilibrium is one that maximizes the particular Social Cost.

We continue to describe separately the KP model and the W model.

• The KP model has considered a simple network consisting of m parallel links from source
to destination; each link bears a capacity – the rate at which the link processes traffic.
Selfish traffic is modeled as a finite collection of users. Each user bears an unsplittable
traffic, which it ships using a mixed strategy – a probability distribution over links; thus,
mixed Nash equilibria have been accomodated in the KP model. The Expected Individual
Cost of a user is its expected latency; in a Nash equilibrium, Expected Individual Cost
cannot be unilaterally decreased. The Social Cost in a Nash equilibrium, called Maximum
Social Cost, is the expectation, over all random choices of the users, of the maximum,
over all links, latency on a link; the Optimum, also called the Maximum Optimum, is
the least possible maximum, over all links, latency on a link. The Coordination Ratio is
called Maximum Coordination Ratio. Alternatively, the KP model may be viewed as a
weighted congestion game [27] equipped with Social Cost and Coordination Ratio.

• The W model has considered arbitrary networks with a source and a destination, and
a latency function for each link. (The parallel links network from the KP model is a
special case.) Selfish traffic is modeled as a splittable flow. The Individual Cost for a
path (from source to destination) is the sum of the latencies incurred on its links. In a
Wardrop equilibrium, all (used) paths have the same Individual Cost. This is like a Nash
equilibrium in the KP model, but with infinitely many users, each carrying infinitesimal
traffic. For this reason, the W model has been restricted to pure equilibria. The Social
Cost in a Wardrop equilibrium is the sum, over all paths, of Individual Costs. The
Optimum is the least possible, over all flows, Social Cost.

In the case of identical links, all link capacities are equal; the general case is known as the
case of arbitrary links or related links. In the case of identical users, all user traffics are equal;
the general case is known as the case of arbitrary users.

3.2 Algorithmic Problems

In this section, we provide a list of several important algorithmic problems (related to Nash
equilibria) that have been studied in the context of the KP model.

3
Π1 : PURE NASH EQUILIBRIUM SUPPORTS
INSTANCE: A number n of users; a number m of links; for each user i, a rational number
wi > 0, called the traffic of user i; for each link j, a rational number c j > 0, called the capacity
of link j.
OUTPUT: Indicator variables Iij ∈ {0, 1}, where 1 ≤ i ≤ n and 1 ≤ j ≤ m, that support a
Nash equilibrium for the system of the users and the links.
We continue with two complementary to each other optimization problems (with respect to
social cost).
Π2 : BEST PURE NASH EQUILIBRIUM SUPPORTS
INSTANCE: A number n of users; a number m of links; for each user i, a rational number
wi > 0, called the traffic of user i; for each link j, a rational number c j > 0, called the capacity
of link j.
OUTPUT: Indicator variables Iij ∈ {0, 1}, where 1 ≤ i ≤ n and 1 ≤ j ≤ m, that support a
Nash equilibrium with minimum social cost for the system of the users and the links.
Π3 : WORST PURE NASH EQUILIBRIUM SUPPORTS
INSTANCE: A number n of users; a number m of links; for each user i, a rational number
wi > 0, called the traffic of user i; for each link j, a rational number c j > 0, called the capacity
of link j.
OUTPUT: Indicator variables Iij ∈ {0, 1}, where 1 ≤ i ≤ n and 1 ≤ j ≤ m, that support a
Nash equilibrium with maximum cost for the system of the users and the links.
We continue with a problem of a somehow counting nature.
Π4 : NASH EQUILIBRIUM MAXIMUM SOCIAL COST
INSTANCE: A number n of users; a number m of links; for each user i, a rational number
wi > 0, called the traffic of user i; for each link j, a rational number c j > 0, called the capacity
of link j; a Nash equilibrium P for the system of the users and the links.
OUTPUT: The social cost of the Nash equilibrium P.

3.3 Pure Nash Equilibria

It has been remarked by Kurt Mehlhorn [6] that a pure Nash equilibrium always exists for the
KP model. The existence proof establishes that the lexicographically minimum expected traffic
vector represents a (pure) Nash equilibrium. Since there are exponentially many pure strategy
profiles and that many expected traffic vectors, this only provides an inefficient proof of exis-
tence of pure Nash equilibria. However, it is also shown in [6] that PURE NASH EQUILIBRIUM
SUPPORTS is in P. The proof is through a polynomial-time algorithm A pure , which is identical

4
to Graham’s algorithm [12] for minimum makespan scheduling. This algorithm runs in time
O(n max{lg n, m}).

Open Problem 3.1 What is the complexity of computing a pure Nash equilibrium? Can the
upper bound of O(n max{lg n, m}) be further improved?

It is shown [6] that both BEST PURE NASH EQUILIBRIUM SUPPORTS and WORST PURE
NASH EQUILIBRIUM SUPPORTS are N P-hard. The proofs assume identical links (but use
arbitrary users).

Open Problem 3.2 What are the complexities of BEST PURE NASH EQUILIBRIUM SUP-
PORTS and WORST PURE NASH EQUILIBRIUM SUPPORTS for the case of identical users
and arbitrary links? Do these problems remain N P-hard?

3.4 Mixed Nash Equilibria

Fotakis et al. [6] prove a polynomial upper bound on the complexity of computing a (generalized
fully) mixed Nash equilibrium for the case where all traffics are identical. The time complexity
of their algorithm is Θ(m lg m). It is natural to ask:

Open Problem 3.3 What is the complexity of computing a generalized fully mixed Nash equi-
librium? Can one do better than O(m lg m)?

Fotakis et al. [6] prove that for the case of identical links, and assuming that n = 2. the
fully mixed Nash equilibrium maximizes Social Cost. The same is proven by Lücking et al. [23]
for the case of identical users, identical links and assuming that m = 2, and for the case of
identical users, related links and assuming that n = 2. Proving the same for all cases will settle
the Fully Mixed Nash Equilibrium Conjecturs [10].

Open Problem 3.4 Prove (or disprove) the Fully Mixed Nash Equilibrium Conjecture.

Gairing et al. [10] get as close as a multiplicative factor of 6 to settling this conjecture for
the case of identical links. On the other hand, the earlier work of Fotakis et al. [6] has got as
close as a multiplicative factor of (approximately) 49 yp settling the conjecture for the case of
identical users.
It is shown in [6] that NASH EQUILIBRIUM SOCIAL COST is #P-complete [41]. The proof
considers an instance with 3 identical links and n + 1 (arbitrary) users.

5
Open Problem 3.5 What is the complexity of NASH EQUILIBRIUM SOCIAL COST when there
are only two identical links?

It is shown in [6] that for the model of identical links, there exists a fully polynomial,
randomized approximation scheme (FPRAS) for NASH EQUILIBRIUM SOCIAL COST. The idea
of the scheme is to define an efficiently samplable random variable which accurately estimates
the Social Cost of the given Nash equilibrium P on a (given) traffic vector w. An obvious open
problem is to extend this scheme to the model of arbitrary links.

Open Problem 3.6 Extend the FPRAS of Fotakis et al. [6] to work for the model of arbitrary
links.

Is randomization necessary for such a scheme?

Open Problem 3.7 Design an efficient, deterministic approximation algorithm for NASH
EQUILIBRIUM SOCIAL COST (or prove that none exists).

4 Extensions to the KP Model

In this section, we survey recent work on extending the original KP model in various ways. Such
extensions have led to new, potentially rich models which pose a new genre of mathematical
problems.

4.1 The Unrelated Links Model

This section is motivated by the work of Lücking et al. [23, Section 6].

The Unrelated Links model has been motivated from the model of unrelated machines that
has been extensively studied in Scheduling Theory (cf. Horowitz and Sahni [14]). In this model,
there are neither weights for the users nor capacities for the links. Instead, there is a cost
associated with each pair of a user and a machine. This cost determines the Individual Cost of
a user on a machine when no other user is assigned to the same machine. In the opposite case,
the Individul Cost of a user is the sum of the costs of all users assigned to the same machine
for that machine. It is very far from clear what the complexity is of computing pure Nash
equilibria in this more general model.

6
Open Problem 4.1 Determine the complexity of computing pure Nash equilibria for the model
of Unrelated Links.

Furthermore, neither the existence (and uniqueness) of the fully mixed Nash equilibrium is
clear for the model of Unrelated Links.

Open Problem 4.2 Determine a characterization for the existence and uniqueness of fully
mixed Nash equilibria in the model of Unrelated Links.

Lücking et al. [23, Theorem 6.4] prove that the fully mixed Nash equilibrium maximizes
(among all Nash equilibria) Social Cost for the case where n = 2 and m = 2. Most interestingly,
Lücking et al. [23, Theorem 6.5] present a counterexample in which the fully mixed Nash
equilibrium does not maximize Social Cost. The counterexample uses a very small instance
with n = 3 and m = 2. Even for such a small instance, it is not clear when the fully mixed
Nash equilibrium maximizes Social Cost. So, as a necessary prerequisite to understanding the
more general properties of the fully mixed Nash equilibrium in the Unrelated Links model, it
is natural to ask:

Open Problem 4.3 Characterize the instances with n = 3 and m = 2 in the model of Unre-
lated Links for which the fully mixed Nash equilibrium maximizes Social Cost.

Finally, there are no bounds at all known for Coordination Ratio in the model of Unrelated
Links. It is not even clear if Coordination Ratio remains bounded in this model.

Open Problem 4.4 Investigate bounds on Coordination Ratio for the model of Unrelated
Links.

4.2 The Restricted Links Model

In this section, we follow the work of Gairing et al. [7].

The Restricted Links model is a very special case of the model of Unrelated Links, which
is defined as follows. Here, users come with weights and links come with capacities, as in the
original KP model (and unlike the Unrelated Links model). However, the cost of user i on
machine j (when it is assigned there alone) is defined to be either wcji or infinity. Another way
to look at this model is that it extends the KP model (as opoosed to restricting the Unrelated
Links model) by allowing any arbitrary combination of pairs of users and machines to have

7
an infinite entry in the matrix whose rows are the users and whose columns are the machines.
(The remaining entries are defined as in the model of Related Links.) The links with no infinite
entries for a particular user are the allowed links for the user – the user has no incentive to
switch to any link outside its allowed links since its Individual Cost there would be infinite.

Gairing et al. [7] present a very elegant polynomial-time algorithm to compute a pure Nash
equilibrium that appoximates the optimum Social Cost for the model of Restricted Links.
Their algorithm relies on a heavy use of ideas from network flows (more specifically, it employs
an extention of the generic Preflow-Push algorithm of Goldberg and Tarjan to work for
unsplittable flows). Moreover, Gairing et al. [7, Section 4] present a Nashification technique
that converts any arbitrary assignemnt to a Nash equilibrium when all links are identical.
We emphasize that Nashification algorithms have the crucial property of not increasing Social
Cost; this is significant for guaranteeing that the Nash equilibrium obtained as a result of the
Nashification algorithm approximates the optimum Social Cost if the original assignment (input
to the Nashification) does. Extending this technique to the case of related links remains a very
important open problem.

Open Problem 4.5 Develop a Nashification algorithm for the model of Restricted Links with-
out the assumption of identical links.

It is also worth investigating the optimality of the Nashification algorithm for the model of
Restricted Links and under the assumption of identical links.

Open Problem 4.6 Determine the time complexity of Nashification algorithms in the model
of Restricted Links and under the assumption of identical links.

4.3 The Quadratic Social Cost Model

This section is based on the work of Lücking et al. [22].

In the Quadratic Social Cost model, the parallel links network considered in the KP model
is adopted. Moreover, it follows the KP model to consider unsplittable traffic and mixed Nash
equilibria. The Expected Individual Cost adopted is the one used in the KP model – the
expected latency incurred to a user. Hence, the Nash equilibria for the new model are exactly
those for the KP model. However, it follows the W model to model Social Cost in such a way
that it turns out to be a certain sum of Expected Individual Costs. So, in a sense, the Quadratic
Social Cost model is the W model restricted to the parallel links network but modified to host

8
unsplittable traffic and mixed strategies; these two features are borrowed from the KP model.
This model represents the first step towards accommodating unsplittable traffic within the W
model.
We continue to describe the Quadratic Social Cost model in some more detail. For any link,
consider the latency incurred by the square of the total traffic on the link; take the expectation
of this latency and add up over all links. This results to the Social Cost for this model, which
we call Quadratic Social Cost. Naturally, Quadratic Optimum is the best possible Quadratic
Social Cost, over all pure assignments. Quadratic Coordination Ratio is the worst-case ratio,
over all Nash equilibria, of Quadratic Social Cost over Quadratic Optimum.
The results of Lücking et al. [22] are partitioned into three major groups.

4.3.1 Quadratic Social Cost

Lücking et al. [22] observe that Quadratic Social Cost can be expressed as a weighted sum of
Expected Individual Costs, where the weights are the user traffics. This observation allows for
the computation of Quadratic Social Cost for any mixed assignment in polynomial time O(mn)

Open Problem 4.7 What is the time complexity of computing Quadratic Social Cost?

For the special case of fully mixed Nash equilibria, the general analysis of Lücking et al. [22]
yields two very simple combinatorial expressions for Quadratic Social Cost in the two separate
cases of identical links. For these two cases, the two expressions immediately imply correspond-
ing polynomial time algorithms for the computation of Quadratic Social Cost.
For the case of identical users, the link capacities enter the combinatorial expression for
Quadratic Social Cost via their sum; moreover, Quadratic Social Cost is inversely proportional
to the sum. This form of dependence is significant since it excludes any possible Braess-like
paradoxes, where increasing a link capacity results to an increase of the Quadratic Social Cost
for the fully mixed Nash equilibrium. In addition, this form of dependence implies that the
Quadratic Social Cost of the fully mixed Nash equilibrium is insensitive to reallocating capacity
among the links. This motivates a very natural question.

Open Problem 4.8 Determine other models with associated functions for Social Cost that
exclude Braess-like paradoxes.

A natural question is to identify the worst-case Nash equilibrium with respect to Quadratic
Social Cost. Lücking et al. [22] address this question for the special case of identical users and

9
identical links. They use an elegant, combinatorial analysis to prove that the worst-case Nash
equilibrium is the fully mixed Nash equilibrium. Lücking et al. [22] formulate the Quadratic
Fully Mixed Nash Equilibrium Conjecture to speculate that this happens in the general case.

Open Problem 4.9 Prove (or disprove) the Quadratic Fully Mixed Nash Equilibrium Conjec-
ture.

4.3.2 Quadratic Coordination Ratio

Lücking et al. [22] present a collection of tight bounds. All their bounds are either constant or
bounded by a constant – independent of m and n. All constants are strictly less than 2.

• For pure Nash equilibria, Lücking et al. [22] establish some new properties of optimal
assignments and pure Nash equilibria to formalize the maximization of Quadratic Co-
ordination Ratio as a non-linear maximization problem, in turn, they solve it to obtain
an upper bound on Quadratic Coordination Ratio. For the case of identical users, the
Quadratic Coordination Ratio is 43 . For the case of identical links, the Quadratic Coordi-
nation Ratio is 89 . The different tight bounds of 43 and 98 imply a Quadratic Coordination
Ratio separation between the cases of identical links and of identical users, respectively,
restricted to pure Nash equilibria. The techniques of Lücking et al. [22] of expressing
the maximization of Quadratic Coordination Ratio as a non-linear optimization problem,
and subsequently solving it, may be of further interest and applicability. These bounds
motivate the following natural question:

Open Problem 4.10 What is the value of Quadratic Coordination Ratio, when restricted
to pure Nash equilibria, for the Quadratic Social Cost model? It is a very challenging task
to merge the two separate proofs for the two special cases of identical users and iden-
tical links, respectively, into a single proof for the general case (restricted to pure Nash
equilibria).

• For mixed Nash equilibria, the results of Lücking et al. [22] are as follows. For the case
of identical links, a tight bound of 2 − max n1 , m
 1
on Quadratic Social Cost for the fully
mixed Nash equilibrium is proved. For the case of identical users and identical links, a
sometimes better upper bound of 1 + min m−1 n−1

n , m , which, however, holds for all Nash
equilibria, is proved. A general bound on Quadratic Coordination Ratio that will hold
for all (mixed) Nash equilibria (and for the general case of arbitrary users and arbitrary
links) is still elusive

10
 Open Problem 4.11 What is the value of Quadratic Coordination Ratio?

Extending the study of the Quadratic Social Cost model to arbitrary networks (following
the W model) remains perhaps the most challenging direction for future research.

4.4 The Discrete Routing Game Model

Gairing et al. [9] introduce and study a discrete routing game as yet another hybridization of
the KP model and the W model. In their model, the latency on a link is determined by an
arbitrary, non-decreasing, convex function; in turn, the latency defines the Expected Individual
Cost for a user (in the same way as in the KP model). However, the Social Cost is taken to be
the sum of the Expected Individual Costs (as in the W model). When users are identical and
latency functions are linear, this discrete routing game reduces to the special case of identical
users for the Quadratic Social Cost model. We note that the set of Nash equilibria for the
model of Gairing et al. [9] (defined through arbitrary, convex latency functions) is (in general)
different from the set of Nash equilibria for the Quadratic Social Cost model (which coincides
with the set of Nash equilibria for the KP model.
The work of Gairing et al. [9] presents the following results:

• Gairing et al. [9, Theorem 1] prove that the fully mixed Nash equilibrium (when it exists)
is the worst-case Nash equilibrium for the case of identical users and links with arbitrary,
convex latency functions. It is very natural to ask:

Open Problem 4.12 Is the fully mixed Nash equilibrium still the worst-case Nash equi-
librium for the case of arbitrary users and links with arbitrary convex, latency functions?

• For the case of identical users and identical links with latency functions f (x) = x d , where
d ≥ 1, Gairing et al. [9, Theorem 5] prove a tight bound of B d+1 , the (d + 1)-th Bell
number, on Coordination Ratio (for their model).

• For the case of identical users and arbitrary links with polynomial latency functions of
maximum degree d, Gairing et al. [9, Corollary 1] prove an upper bound of d + 1 on
Coordination Ratio, restricted to pure Nash equilibria. However, it remains natural to
ask:

Open Problem 4.13 For the case of identical users and arbitrary links with polynomial
latency functions of maximum degree d, is the bound of d + 1 on Coordination Ratio,
restricted to pure Nash equilibria, tight?

11
4.5 The Polynomial Social Cost Model

Gairing et al. [8] introduce and study yet another hybridization of the KP model and the W
model. In their model, the latency on a link is a linear function (as in the KP model and the
Quadratic Social Cost model). However, a so called Polynomial Social Cost is adopted that
extends Quadratic Social Cost to arbitrary degrees. So assume that Polynomial Social Cost
is the expectation of the sum, over all links, of a certain polynomial evaluated at the total
latency incurred by all users choosing the link. Assume, furthermore, that this polynomial has
maximum degree d and that all of its coefficients are non-negative. In turn, Polynomial Social
Cost yields Polynomial Coordination Ratio. Gairing et al. [8] only consider the case of identical
users and identical links. For this case, they present a collection of upper bounds on Polynomial
Coordination Ratio.
More specifically, Gairing et al. [8, Section 4] consider first the subcase of two links. For
that subcase, they prove [8, Theorem 1] that the fully mixed Nash equilibrium is the worst-
case Nash equilibrium with respect to Polynomial Social Cost. This is shown using a purely
combinatorial proof. This result is crucial for proving bounds on Polynomial Coordination
Ratio, since it allows one to focus on the fully mixed Nash equilibrium and prove bounds on
Polynomial Coordination Ratio for that particular equilibrium. Using this approach, Gairing
1 d−1
et al. [8, Theorem 2] prove an exact bound of 2 (2 + 1) on Polynomial Coordination Ratio
for the case where the Polynomial Social Cost is the dth power. This immediately implies an
upper bound of 21 (2d + d − 1) on Polynomial Coordination Ratio for the general case of a degree
d polynomial with non-negative coefficients. So, it is natural to ask:

Open Problem 4.14 Find a (matching) lower bound on Polynomial Coordination Ratio for
the model of identical users and identical links, and assuming that there are only two links and
that the polynomial used for Polynomial Social Cost is a degree d polynomial with non-negative
coefficients.

For the case of an arbitrary number of links, Gairing et al. —citeGLMM04a are less success-
ful. They prove [8, Theorem 3] that the Polynomial Social Cost of any Nash equilibrium is at
1
most (1 + n−1 ) · Bd times the Social Cost of the fully mixed Nash equilibrium. They conjecture,
1
however, that the factor (1 + n−1 ) · Bd is not necessary.

Open Problem 4.15 Prove (or diprove) that for the model of identical users and identical
links, and assuming that the polynomial used for Polynomial Social Cost is a degree d polynomial
with non-negative coefficients, the Polynomial Social Cost of any (mixed) Nash equilibrium is
at most the Polynomial Social Cost of the fully mixed Nash equilibrium.

12
 The work of Gairing et al. [8] has not considered at all the case of arbitrary users and
arbitrary links. This open a wide avenue for new research.

Open Problem 4.16 Develop the theory of Polynomial Social Cost and Polynomial Coordi-
nation Ratio for the case of arbitrary users and arbitrary links.

5 Coordination Mechanisms

This section is based on the work of Christodoulou et al. [1].

Coordination Mechanisms were recently introduced by Christodoulou et al. [1] as yet another
variant of mechanism design. In contrast to general mechanism design where tolls are introduced
in order to improve the system performance, coordination mechanisms may be seen as a way
to introduce traffic lights. To define coordination mechanisms, we first define the more general
concept of Coordination Model.

Formally, a Coordination Model is a tuple hN, M, (Σ i )i∈N , (C j )j∈M i, where N = {1, 2, . . . , n}

is the set of players, M = {1, 2, . . . , m} is the set of facilities, Σ i is a set of strategies for player
i, where a strategy Ai ∈ Σi is a set of facilities, and finally C j is a family of cost functions
associated with facility j ∈ M . A specific cost function c j ∈ C j is a function that takes as input
a set of loads, one for each player that uses the facility, and outputs a cost to each participating
player. A natural property of cost functions is that players who do not use the facility incur
no cost. Consider, for example, the Coordination Model for the routing game represented by
the KP model. It is natural to assume there that the cost function for each machine and each
subset of users assigned to the machine is no less than the total load of those users. Other than
this assumption, a cost function assigns completion times to each user in an arbitrary way;
however, the way this assignment is made is fixed before the machine ”sees” the actual weights
of the users choosing it.

Christodoulou et al. [1] parallel their definition of Coordination Model to that of an On-
line Problem. Thus, a Coordination Model is best understood as a family of Coordination
Mechanisms much in the same way an Online Problem is a family of Online Algorithms. Each
particular Coordination Mechanism is evaluated on the basis of its Coordination Ratio (or
Price of Anarchy [34]) much in the same way an Online Algorithm is evaluated on the basis
of its Competitive Ratio. This nice formulation of Christodoulou et al. [1] opens up the study
and evaluation of Coordination Mechanisms, each one defined by fixing cost functions for the
facilities in a specific way.

13
 For the game in the KP model, Christodoulou et al. [1, Proposition 1] already present
a Coordination Mechanism which, for the case of two machines, already improves the tight
bound of 32 achieved by the Coordination Mechanism in the original work of Koutsoupias and
Papadimitriou [19]; that Coordination Mechanism scheduled jobs in an arbitrary order and
the cost for each user is the sum of the weights of users assigned to the machine. Instead,
the Coordination Mechanism of Christodoulou et al. [1] has one machine scheduling the jobs in
increasing order and the other machine scheduling the jobs in decreasing order; this is called the
Increasing-Decreasing Coordination Mechanism. Christodoulou et al. [1, Proposition 1] prove
that the Increasing-Decreasing Coordination Mechanism achieves Coordination Ratio 43 . Most
interestingly, Christodoulou et al. [1, Theorem 1] present a slightly more involved Coordination
1
Mechanism achieving Coordination Ratio 34 − 3m for any number of machines m. The idea is
to make the machines as asymmetric as possible by inducing them to charge different costs to
their assigned users. It is natural to ask whether one can do better.

Open Problem 5.1 Is there a Coordination Mechanism for the routing game of the KP model
1
with Coordination Ratio better than 34 − 3m (where m is the number of machines)?

The study of Coordination Mechanisms by Christodoulou et al. [1] has focused on the KP
model, and nothing can be said yet about Coordination Mechanisms for other models that
extend the KP model. So, a well motivated open problem is the following.

Open Problem 5.2 Develop the theory of Coordination Mechanisms for the Quadratic Social
Cost model, the Discrete Routing Game model and the Polynomial Social Cost model.

Christodoulou et al. [1, Section 4] proceed to study Coordination Mechanisms for general
congestion games [36]. The interesting observation of Christodoulou et al. is that the Coor-
dination Ratio for congestion games is unbounded if Coordination Mechanisms are ruled out.
This is true even for single-commodity congestion games where there is a single source and a
single destination for all players. On the positive side, Christodoulou et al. [1, Theorems 3 &
4] prove that there exists a Coordination Mechanism achieving Coordination Ratio n for each
specific single-commodity congestion game, and that this upper bound is tight. Whethere this
holds for all congestion games remains a very interesting open problem.

Open Problem 5.3 Find a Coordination Mechanism with Coordination Ratio n for every
congestion game with monotone positive costs, or prove that none exists.

14
6 Local-Effect Games

This section is based on the work of Leyton-Brown and Tennenholtz [21].

In congestion games [36], whenever two players affect each other’s payoff, they each do
so by the same amount. In local-effect games [21], the effects of players on each other are
potentially asymmetric. Loosely speaking, action A locally affects action B if the utility of
players taking action B depends on some function F A,B of the number of players taking action
A; note, however, that the utility of agents taking action A depends on a potentially different
function FB,A of the number of agents taking action B. Leyton-Brown and Tennenholtz [21]
provide arguments that many natural problem domains (for example, the location problem [13]
from Economics) are accurately modeled as local-effect games.

Formally, a local-effect game for n players is a pair hA, Fi where A is the ground set of
strategies available to each player in game. Let D denote the distribution of players across
strategies, and for each strategy a ∈ A, denote D(a) the number of players who choose strategy
a. For every pair of strategies a, a0 ∈ A, define the function Fa,a0 : Z+ → R+ as the cost function
expressing the cost due to the local effect from strategy a to strategy a 0 and depending on the
number of players having chosen strategy a. This definition allows to express the Individual
Cost IC of an agent who chose strategy a ∈ A as
X
IC(a) = Fa,a (D(a)) + Fa0 ,a (D(a0 )) .
a0 ∈A,a0 6=a

An additional assumption made by Leyton-Brown and Tennenholtz [21] is that of strict-

monotonicity. Formally, say that F is strictly monotonic if for all pairs of strategies a, a 0 ∈ A
with a 6= a0 , the cost function Fa,a0 (x) either increases strictly monotonically with x or is always
0. Furthermore, it is assumed that for all pairs of strategies a, a 0 ∈ A with a 6= a0 , Fa,a0 (0) = 0.

It is useful to use a directed graph representing the strategies and their local effects. We use
a vertex to represent each strategy. We put an edge from vertex a to vertex a 0 if the function
Fa,a0 is not identically zero; that is, it does not hold that for all x ∈ Z + , Fa,a0 (x) = 0. It is
assumed in [21] that all vertices in the graph are either both or neither neighbors of each other.
The resulting graph is called the local-effect graph.

Leyton-Brown and Tennenholtz [21] define two special classes of local-effect games:

• A local-effect game is a bidirectional local-effect game if for all pairs of strategies a, a 0 ∈ A

with a 6= a0 , for all x ∈ Z+ , Fa,a0 (x) = Fa0 ,a (x).

15
 Roughly speaking, for bidirectional local-effect games, cost-effect functions for a given
pair of strategies are always the same in both directions. Note, however, that this does
not imply that for any given distribution of players to strategies, the magnitude of the
local effects between a pair of strategies is necessarily the same.
Clearly, the graphical representation of a bidirectional local-effect game is simplified to
an undirected graph.

• A local-effect game is a uniform local-effect game if for each strategy a ∈ A, for all
strategies a0 , a00 ∈ A that are neighbors of a in the graph representing the game, it holds
that for all x ∈ Z+ , Fa,a0 (x) = Fa,a00 (x).

Leyton-Brown and Tennenholtz [21, Section 3.1] provide a counterexample local-effect game
to prove that not all local-effect games have a pure Nash equilibrium. Since every congestion
game has a pure Nash equilibrium [36], this counterexample implies that the class of local-
effect games is different from the class of congestion games. However, a characterization of
those local-effect games that always have a pure Nash equilibrium is still elusive.

Open Problem 6.1 Determine a characterization of local-effect games that have a pure Nash
equilibrium.

Even more so, it is not known if the class of local-effect games admits a characterization
that would lead to an efficient (polynomial-time) algorithm to decide the existence problem for
local-effect games.

Open Problem 6.2 What is the complexity of deciding whether a local-effect game has a pure
Nash equilibrium?

On the positive side, there is a subclass of local-effect games that is known to always have
a pure Nash equilibrium. This is a subclass of the bidirectional local-effect games. More
specifically, Leyton-Brown and Tennenholtz [21, Theorem] prove that a bidirectional local-
effect game has a pure Nash equilibrium if for all pairs of strategies a, a 0 ∈ A with a 6= a0
Fa,a0 (x) = ma,a0 x for some constant ma,a0 . Even for this subclass, it is not known how to find
a pure Nash equilibrium.

Open Problem 6.3 Determine the complexity of finding a pure Nash equilibrium for a bidirec-
tional local-effect game such that for all pairs of strategies a, a 0 ∈ A with a 6= a0 Fa,a0 (x) = ma,a0 x
for some constant ma,a0 .

16
 There is yet another subclass of local-effect games that is known to always have a pure
Nash equilibrium. This is a subclass of the uniform local-effect games. Leyton-Brown and
Tennenholtz [21, Theorem 2] prove that a uniform local-effect game has a pure Nash equilibrium
if its local-effect graph is a clique. This motivates the corresponding question of finding a pure
Nash equilibrium for such local-effect games.

Open Problem 6.4 Determine the complexity of finding a pure Nash equilibrium for a uni-
form local-effect game whose local-effect graph is a clique.

Leyton-Brown and Tennenholtz [21, Theorem 4] provide a sufficient condition for a class
of graphs and cost functions to have a pure Nash equilibrium. This class has the mentioned
classes of bidirectional local-effect games and uniform local-effect games as special cases. The
most interesting aspect of this result is that its proof is constructive – it gradually constructs
a pure Nash equilibrium in which players choose strategies that constitute an independent set.
This implies an exponential upper bound for the problem of finding a pure Nash equilibrium
for the subclasses of bidirectional local-effect games and uniform local-effect games in Open
Problems 6.3 and 6.4, respectively.
The study of local-effect games has been restricted to pure Nash equilibria. Investigating
the mixed Nash equilibria for local-effect games remains a tantalizing open problem.

Open Problem 6.5 Develop the theory of mixed Nash equilibria for local-effect games. In
particualr, determine the complexity of finding a mixed Nash equilibrium for a local-effect game.

In particular, we advocate the study of the fully mixed Nash equilibrium [24] for local-effect
games.

Open Problem 6.6 Determine a characterization of local-effect games that always have a
fully mixed Nash equilibrium.

The next natural step would be to define some appropriate Social Cost for local-effect
games. This will lead to many natural and interesting questions in analogy to questions that
have been asked (and answered) for the KP game and its variants. Of course, the definition of
Social Cost will immediately yield a definition of Coordination Ratio for local-efect games. So,
every natural definition of Social Cost will yield a wealth of questions regarding Social Cost
and Coordination Ratio which we refrain from listing here. We cannot, however, resisting from
asking the question whether for some natural definition of Social Cost, the fully mixed Nash
equilibrium will be the Nash equilibrium that maximizes Social Cost.

17
7 Interaction Graph Games

In this section, we treat separately the model of interaction graphs, which is a special case of
the model of restricted parallel links. The work in this section is based on the recent work of
Elsässer et al. [3].

Elsässer et al. [3] use a structured and sparse representation of the relation between users
and machines that exploits the locality of their interaction in a complex scheduling system.
More specifically, they assume that each user has access (that is, finite Individual Cost) to only
two machines; its Individual Cost on other machines is infinitely large, giving it no incentive to
switch there.

The game representation used by Elsässer et al. is based on the interaction graph, whose
vertices and edges represent the machines and the users, respectively. Multiple edges are al-
lowed; however, for simplicity, the interaction multigraphs are called interaction graphs. The
model of interaction graphs is interesting because it is the simplest, non-trivial mathematical
model for selfish scheduling on restricted parallel links. Elsässer et al. [3] only consider the case
of identical users and identical links.

In addition to considering pure Nash equilibria and the fully mixed Nash equilibrium [24],
Elsässer et al. [3] also consider, in particular, the standard fully mixed Nash equilibrium, where
all probabilities are equal to 12 . It is easy to see that the standard fully mixed Nash equilibrium
exists if and only if the interaction (multi)graph is regular. In all cases, any assignment of
users to machines naturally corresponds to an orientation of the interaction graph. Thus, the
study of assignments invites Game Theory, and in particular the theory of graph orientations,
to come into play. With each (mixed) Nash equilibrium, Elsässer et al. [3] adopt the definition
of Social Cost from the work of Koutsoupias and Papadimitriou [19].

The results of Elsässer et al. [3], and the corresponding spawned open problems, may be
partitioned into three major groups.

7.1 3-regular Interaction Graphs

Elsässer et al. [3, Theorem 3.2] use an easy proof to show that the Social Cost of the standard
fully mixed Nash equilibrium for any d-regular graph is d − f (d, n), where f (d, n) is a function
that goes to 0 as n goes to infinity. This gives a general but rather rough estimation of
Social Cost for d-regular graphs; moreover, it does not say how the specific structure of each
particular d-regular graph affects Social Cost. Elsässer et al. [3] continue to prove much sharper
estimations for the special class of 3-regular graphs.

18
 More specifically, Elsässer et al. [3] pursue a thorough study of 3-regular interaction graphs;
these graphs further restrict the bounded interaction by insisting that each machine is accessible
to just three users. Specifically, Elsässer et al. focus on the standard fully mixed Nash equilib-
rium. They ask the new question of which is the best 3-regular interaction graph in this case
(among, of course, all 3-regular interaction graphs with the same number of vertices). Although
the literature has so far dealt with worst-case and best-case Nash equilibria, the graph-theoretic
model of Elsässer et al. [3] invites now the study of worst-case and best-case graphs, each with
respect to some particular type of Nash equilibrium.

This question brings into the context of selfish scheduling the problem of comparing, for a
given 3-regular graph, the expected number of 2-orientations and 3-orientations – those with
makespan 2 and 3, respectively. The manner in which these orientations outweigh each other
brings Social Cost closer to either 2 or 3. Elsässer et al. [3, Corollary 3.5] develop some
deep graph-theoretic lemmas about 2- and 3-orientations in 3-regular graphs to prove that the
simplest 3-regular parallel links graph is the best 3-regular graph in this setting. The proof
decomposes any 3-regular graph down to the parallel links graph in a way that Social Cost does
not increase. The graph theoretic lemmas about 2- and 3-orientations are proved using both
counting and mapping techniques; both the graph-theoretic lemmas and their proof techniques
appear to be of much more general interest and applicability.

This result of Elsässer et al. [3] points to a new genre of mathematical problems at the
fascinating intersection of Game Theory and Graph Theory. We mention a few of them here.
The first deals with the necessity of assuming the necessity of the standard fully mixed Nash
equilibrium for proving the optimality of the 3-regular parallel links graph.

Open Problem 7.1 Is the standard fully mixed Nash equilibrium essential for the optimality
of the 3-regular parallel links graph? Or does the optimality hold for all fully mixed Nash
equilibria?

Of course, the next major step will be to extend the analysis for 3-regular graphs to d-regular
graphs.

Open Problem 7.2 For any given integer d > 3, determine the best-case d-regular graph (with
respect to Social Cost). In particular, is the d-regular parallel links graph the best-case graph?

The work of Elsässer et al. [3] has only considered questions about best-case graphs. Cor-
responding problems about worst-case graphs are completely open.

19
Open Problem 7.3 Determine the worst-case 3-regular graph. Generalize to d-regular graphs
for any integer d > 3.

The graph-theoretic model of Elsässer et al. [3] has used the definition of Koutsoupias and
Papadimitriou [19] for Social Cost – namely, Social Cost is the expected makespan. It is not
at all clear whether the employed graph-theoretic machinery is specific to this Social Cost, or
whether the result extends to other definitions of Social Cost. In particular, this motivates a
very interesting question.

Open Problem 7.4 Determine the best-case and worst-case 3-regular graphs with respect to
Quadratic Social Cost and with respect to Polynomial Social Cost. Generalize to d-regular
graphs for any integer d > 3.

7.2 Bounds on Coordination Ratio

For pure Nash equilibria in the model of interaction graphs, an upper bound of O( logloglogn n ) on
Coordination Ratio follows from the corresponding upper bound for the more general model of
restricted parallel links[8, Theorem 5.2]. Elsässer et al. [3, Theorem 4.1] cnstruct an interaction
graph incurring Coordination Ratio Ω( logloglogn n ) to prove that this bound is tight. The construc-
tion extends an approach from [8, Lemma 5.1] that proved the same lower bound for the more
general model of restricted parallel links. The extension circumvents the additional restriction
that each user is only allowed to ship its load on just two machines.

Nothing is known about Coordination Ratio for the model of interaction graphs when all
mixed Nash equilibria are considered. Hence, it is natural to ask:

Open Problem 7.5 Determine the Coordination Ratio for the model of interaction graphs.
In particular, is it still Θ( logloglogn n ) when all mixed Nash equilibria are considered?

7.3 Fully Mixed Nash Equilibria

Elsässer et al. [3, Section 5] pursue a very thorough study of fully mixed Nash equilibria in the
model of interaction graphs.

Elsässer et al. [3, Counterexamples 5.1 & 5.2] provide counterexample interaction graphs
for which fully mixed Nash equilibria may not exist. Among them are all trees and meshes.
These counterexamples provide some insight about a possible graph-theoretic characterization

20
of interaction graphs admitting a fully mixed Nash equilibrium. 4-cycles and 1-connectivity are
factors to play a role in this characterization. However, the complete characterization is still
elusive.

Open Problem 7.6 Provide a graph-theoretic characterization of interaction graphs that ad-
mit a fully mixed Nash equilibrium.

Elsässer et al. [3] invistigate cases where infinitely many Nash equilibria may exist. In such
cases, the new concept of fully mixed Nash dimension is defined to be the dimension of the
smallest d-dimensional space that can contain all fully mixed Nash equilibria. For complete
bipartite graphs, Elsässer et al. [3, Theorem 5.2] prove a dichotomy theorem that characterizes
unique existence. The proof employs arguments from Linear Algebra. The result is that the
complete bipartite graph Kr,s with r ≤ s admits a unique fully mixed Nash equilibrium if and
only if r > 2; for r = 2, the fully mixed Nash dimension is s − 1. For the case where the fully
mixed Nash equilibrium is not unique, it is natural to ask:

Open Problem 7.7 Determine the best-case and worst-case fully mixed Nash equilibrium for
the complete bipartite graph K2,s .

For hypercubes, Elsässer et al. [3] have only been able to prove that the fully mixed Nash
dimension is the hypercube dimension for hypercubes of dimension 2 or 3. They conjecture
that this is true for all hypercubes, but they have only been able to observe that the hypercube
dimension is a lower bound on the fully mixed Nash dimension (for all hypercubes).

Open Problem 7.8 Determine the fully mixed Nash dimension for a hypercube of dimension
d > 3.

Elsässer et al. [3] were also interested in understanding whether (and when) the fully mixed
Nash equilibrium is the worst-case one for the model of interaction graphs. Elsässer et al. [3,
Counterexamples 5.3 & 5.4] present counterexample interaction graphs to show that the fully
mixed Nash equilibrium is sometimes the worst-case Nash equilibrium, but sometimes not.
More specifically, they prove that For the hypercube, there is a pure Nash equilibrium that
is worse (with respect to Social Cost) than any fully mixed one; on the other hand, for the
3-cycle, the (unique) fully mixed Nash equilibrium has worst Social Cost. We are currently
missing a graph-theoretic characterization of interaction graphs for which the fully mixed Nash
equilibrium is the worst-case Nash equilibrium.

21
Open Problem 7.9 Provide a graph-theoretic characterization of interaction graphs for which
the fully mixed Nash equilibrium is worst-case (with respect to Social Cost).

It is very natural to ask the same question when one considers Quadratic Social Cost instead
of Social Cost.

Open Problem 7.10 Provide a graph-theoretic characterization of interaction graphs for which
the fully mixed Nash equilibrium is worst-case with respect to Quadratic Social Cost.

7.4 General

The case where users have arbitrary weights and links have arbitrary capacities is left completely
untouched by the work of Elsässer et al. [3]. The study of the general case remains a major
challenge.

Open Problem 7.11 Develop the theory of Nash equilibria for interaction graphs in the gen-
eral case where users have arbitrary weights and links have arbitrary capacities.

8 Epilogue

We have surveyed a comprehensive collection of numerous, interesting open problems in the

exciting field of Algorithmic Game Theory. Many of these open problems deal with Nash equi-
libria, the dominating solution concept in Game Theory. Our syrvey is by no means complete
– for example, it has not covered combinatorial auctions [42] or market equilibria [38], two cur-
rently trendy topics in Algorithmic Game Theory. Perhaps, the current expansion of the field
already rules out the possibility of a complete survey. Nevertheless, we hope that our coverage
will stimulate further additionbal research on the topics we have surveyed and, perhaps more
important, inspire the study of new topics and new problems that we are currently not even
aware of.

22
References

[1] G. Christodoulou, E. Koutsoupias and A. Nanavati, ”Coordination Mechanisms,” Pro-

ceedings of the 31st International Colloquium on Automata, Languages and Programming
(ICALP 2004), J. Diaz, J. Karhumäki, A. Lepistö et al., pp. 345–357, Vol. 3142, Lecture
Notes in Computer Science, Springer-Verlag, Turku, Finland, July 2004.

[2] A. Czumaj and B. Vöcking, ”Tight Bounds for Worst-Case Equilibria,” Proceedings of the
13th Annual ACM Symposium on Discrete Algorithms, pp. 413–420, January 2002.

[3] R. Elsässer, M. Gairing, T. Lücking, M. Mavronicolas and B. Monien, ”A Simple Graph-

Theoretic Model for Selfish Restricted Scheduling,” submitted for publication.

[4] E. Even-Dar, A. Kesselman and Y. Mansour, “Convergence Time to Nash Equilibria,”

Proceedings of the 30th International Colloquium on Automata, Languages and Program-
ming, J. C. M. Baeten, J. K. Lenstra, J. Parrow and G. J. Woeginger eds., pp. 502–513,
Vol. 2719, Lecture Notes in Computer Science, Springer-Verlag, June/July 2003.

[5] R. Feldmann, M. Gairing, T. Lücking, B. Monien and M. Rode, ”Nashification and the
Coordination Ratio for a Selfish Routing Game,” Proceedings of the 30th International
Colloquium on Automata, Languages and Programming, J. C. M. Baeten, J. K. Lenstra,
J. Parrow and G. J. Woeginger eds., pp. 514–526, Vol. 2719, Lecture Notes in Computer
Science, Springer-Verlag, June/July 2003.

[6] D. Fotakis, S. Kontogiannis, E. Koutsoupias, M. Mavronicolas and P. Spirakis, “The Struc-

ture and Complexity of Nash Equilibria for a Selfish Routing Game,” Proceedings of the
29th International Colloquium on Automata, Languages and Programming, P. Widmayer,
F. Triguero, R. Hennessy, S. Eidenbenz and R. Conejo eds., Vol. 2380, Lecture Notes in
Computer Science, Springer-Verlag, Malaga, Spain, July 2002.

[7] M. Gairing, T. Lücking, M. Mavronicolas and B. Monien, ”Computing Nash Equilibria for
Scheduling on Restricted Parallel Links,” Proceedings of the 36th Annual ACM Symposium
on Theory of Computing, pp. 613–622, June 2004.

[8] M. Gairing, T. Lücking, M. Mavronicolas and B. Monien, ”The Price of Anarchy for Poly-
nomial Social Cost,” Proceedings of the 29th International Symposium on Mathematical
Foundations of Computer Science, J. Fiala, V. Koubek and J. Kratochvil eds., pp. 574–
585, Lecture Notes in Computer Science, Springer-Verlag, August 2004.

23
 [9] M. Gairing, T. Lücking, M. Mavronicolas, B. Monien and M. Rode, ”Nash Equilibria in
Discrete Routing Games with Convex Latency Functions,” Proceedings of the 31st Inter-
national Colloquium on Automata, Languages and Programming, J. Diaz, J. Karhumäki,
A. Lepistö and D. Sannella eds., pp. 645–657, Vol. 3142, Lecture Notes in Computer Sci-
ence, Springer-Verlag, July 2004.

[10] M. Gairing, T. Lücking, M. Mavronicolas, B. Monien and P. Spirakis, ”Extreme Nash

Equilibria,” Proceedings of the 8th Italian Conference on Theoretical Computer Science,
C. Blundo and C. Laneve eds., pp. 1–20, Vol. 2841, Lecture Notes in Computer Science,
Springer-Verlag, October 2003.
Full version: ”The Structure and Complexity of Extreme Nash Equilibria,” accepted to
Theoretical Computer Science.

[11] I. Gilboa and E. Zemel, ”Nash and Correlated Equilibria: Some Complexity Considera-
tions,” Games and Economic Behavior, Vol. 1, pp. 80–93, 1989.

[12] R. L. Graham, ”Bounds on Multiprocessing Timing Anomalies,” SIAM Journal on Applied

Mathematics, Vol. 17, pp. 416–429, 1969.

[13] H. Hotelling, ”Stability in Competition,” Economic Journal, Vol. 39, pp. 41–57, 1929.

[14] E. Horowitz and S. Sahni, ”Exact and Approximate Algorithms for Scheduling Nonidentical
Processors,” Journal of the ACM, Vol. 2, pp. 317–327, 1976.

[15] S. Kakutani, ”A Generalization of Brouwer’s Fixed Point Theorem,” Duke Mathematical

Journal, Vol. 8, pp. 457–459, 1941.

[16] R. M. Karp and M. Luby, ”Monte Carlo Approximation Algorithms for Enumeration and
Reliability Problems,” Proceedings of the 24th Annual IEEE Symposium on Foundations
of Computer Science, pp. 56–64, May 1983.

[17] R. M. Karp, M. Luby and N. Madras, ”Monte Carlo Approximation Algorithms for Enu-
meration Provblems,” Journal of Algorithms, Vol. 10, pp. 429–448, 1989.

[18] E. Koutsoupias, M. Mavronicolas and P. Spirakis, ”Approximate Equilibria and Ball Fu-
sion,” Theory of Computing Systems, Vol. 36, No. 6, pp. 683–693, November 2003.

[19] E. Koutsoupias and C. H. Papadimitriou, “Worst-case Equilibria,” Proceedings of the 16th

Annual Symposium on Theoretical Aspects of Computer Science, G. Meinel and S. Tison
eds., pp. 404–413, Vol. 1563, Lecture Notes in Computer Science, Springer-Verlag, Trier,
Germany, March 1999.

24
[20] C. E. Lemke and J. T. Howson, ”Equilibrium Points of Bimatrix Games,” Journal of the
Society for Industrial and Applied Mathematics, Vol. 12, pp. 413–423, 1964.

[21] K. Leyton-Brown and M. Tennenholtz, ”Local-Effect Games,” Proceedings of the 18th

International Joint Conference of Artificial Intelligence, pp. 772–780, August 2003.

[22] T. Lücking, M. Mavronicolas, B. Monien and M. Rode, ”A New Model for Selfish Rout-
ing,” Proceedings of the 21st International Symposium on Theoretical Aspects of Computer
Science, V. Diekert and M. Habib eds., pp. 547–558, Vol. 2996, Lecture Notes in Computer
Science, Springer-Verlag, Montpellier, France, March 2004.

[23] T. Lücking, M. Mavronicolas, B. Monien, M. Rode, P. Spirakis and I. Vrto, ”Which is

the Worst-Case Nash Equilibrium?,” Proceedings of the 28th International Symposium on
Mathematical Foundations of Computer Science, B. Rovan and P. Vojtas eds., pp. 551–561,
Vol. 2747, Lecture Notes in Computer Science, Springer-Verlag, August 2003.

[24] M. Mavronicolas and P. Spirakis, “The Price of Selfish Routing,” Proceedings of the 33rd
Annual ACM Symposium on Theory of Computing, pp. 510–519, July 2001.

[25] R. D. McKelvey and A. McLennan, ”Computation of Equilibria in Finite Games,” in

Handbook of Computational Economics, H. Amman, D. Kendrick and J. Rust eds., pp.
87–142, 1996.

[26] N. Megiddo and C. H. Papadimitriou, ”On Total Functions, Existence Theorems, and
Computational Complexity,” Theoretical Computer Science, Vol. 81, No. 2, pp. 317–324,
1991.

[27] I. Milchtaich, ”Congestion Games with Player-Specific Payoff Functions,” Games and Eco-
nomic Behavior, Vol. 13, No. 1, pp. 111–124, March 1996.

[28] R. B. Myerson, Game Theory, Harvard University Press, Cambridge, Massachusetts, 1991.

[29] J. F. Nash, “Non-cooperative Games,” Annals of Mathematics, Vol. 54, No. 2, pp. 286–295,
1951.

[30] J. F. Nash, S. Nasar (ed.) and H. Kuhn (ed.), The Essential John Nash, Princeton Uni-
versity Press, 2001.

[31] M. J. Osborne and A. Rubinstein, A Course in Game Theory, The MIT Press, 1994.

[32] C. H. Papadimitriou, ”On the Complexity of the Parity Argument and Other Inefficient
Proofs of Existence,” Journal of Computer and System Sciences, Vol. 48, No. 3, pp. 498–
532, June 1994.

25
[33] C. H. Papadimitriou, ”Algorithms, Games and the Internet,” Proceedings of the 28th Inter-
national Colloquium on Automata, Languages and Programming, F. Orejas, P. G. Spirakis
and J. von Leeuwen eds., pp. 1–3, Vol. 2076, Lecture Notes in Computer Science, Springer-
Verlag, July 2001.

[34] C. H. Papadimitriou, ”Algorithms, Games and the Internet,” Proceedings of the 33rd An-
nual ACM Symposium on Theory of Computing, pp. 749–753, July 2001.

[35] J. Rosenmuller, ”On a Generalization of the Lemke-Howson Algorithm to Non-cooperative

N -Person Games,” SIAM Journal of Applied Mathematics, Vol. 21, pp. 73–79, 1971.

[36] R. W. Rosenthal, ”A Class of Games Possessing Pure-Strategy Nash Equilibria,” Interna-

tional Journal of Game Theory, Vol. 2, pp. 65–67, 1973.

[37] T. Roughgarden, Selfish Routing, Ph. D. Thesis, Department of Computer Science, Cornell
University, May 2002.

[38] H. E. Scarf, The Computation of Economic Equilibria, Yale University Press, New Haven,
Connecticut, 1973.

[39] L. S. Shapley, ”A Note on the Lemke-Howson Algorithm,” Mathematical Programming

Study, Vol. 1, pp. 175–189, 1974.

[40] B. von Stengel, ”Computing Equlibria for Two-Person Games,” in Handbook of Game
Theory, Vol. 3, R. J. Aumann and S. Hart eds., North Holland, Amsterdam, 1998.

[41] L. G. Valiant, ”The Complexity of Computing the Permanent,” Theoretical Computer

Science, Vol. 8, pp. 189–201, 1979.

[42] W. Vickrey, ”Counterspeculation, Auctions and Competitive Sealed Tenders,” Journal of

Finance, pp. 8–37, 1961.

[43] J. G. Wardrop. ”Some Theoretical Aspects of Road Traffic Research,” Proceedings of the
of the Institute of Civil Engineers, Pt. II, Vol. 1, pp. 325–378, 1952.

26