Complexity Results for Some Classes

of Strategic Games
Dissertation an der
Fakultät für Mathematik, Informatik und Statistik der
Ludwig-Maximilians-Universität München

vorgelegt von Felix Fischer

am 11.03.2009
Complexity Results for Some Classes
of Strategic Games
Dissertation an der
Fakultät für Mathematik, Informatik und Statistik der
Ludwig-Maximilians-Universität München

vorgelegt von Felix Fischer

am 11.03.2009

Betreuer: Dr. Felix Brandt

Ludwig-Maximilians-Universität München
Erster Berichterstatter: Prof. Martin Hofmann, Ph.D.
Ludwig-Maximilians-Universität München
Zweiter Berichterstatter: Prof. Lane A. Hemaspaandra, Ph.D.
University of Rochester, NY, USA
Datum des Rigorosums: 03.07.2009
to you
 Contents

Abstract xi

Zusammenfassung xiii

Acknowledgements xv

1 Introduction 1

2 Games, Solutions, and Complexity 5

2.1 Strategic Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Solution Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Elements of Complexity Theory . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 A Few Words on Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3 State of the Art and Our Contribution 23

4 Ranking Games 29
4.1 An Introductory Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.4 Games With Non-Pure Equilibria . . . . . . . . . . . . . . . . . . . . . . . . 35
4.5 Solving Ranking Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.6 Comparative Ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

5 Anonymous Games 57
5.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.3 Pure Nash Equilibria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.4 Iterated Weak Dominance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

vii
viii

6 Graphical Games 101

6.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.3 A Tight Hardness Result for Pure Equilibria . . . . . . . . . . . . . . . . . 105
6.4 Pure Equilibria of Graphical Games with Anonymity . . . . . . . . . . . . . 111
6.5 Interlude: Satisfiability in the Presence of a Matching . . . . . . . . . . . . 123
6.6 Mixed Equilibria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

7 Quasi-Strict Equilibria 127

7.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.3 Two-Player Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.4 A Hardness Result for Multi-Player Games . . . . . . . . . . . . . . . . . . 131
7.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

8 Shapley’s Saddles 137

8.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.3 Strict Saddles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
8.4 Weak Saddles of Confrontation Games . . . . . . . . . . . . . . . . . . . . . 141
8.5 A Hardness Result for Weak Saddles . . . . . . . . . . . . . . . . . . . . . . 146
8.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

References 149

Lebenslauf 161
 List of Figures

2.1 Alice, Bob, or Charlie? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 The Matching Pennies game . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

4.1 The game of Figure 2.1 as a single winner game . . . . . . . . . . . . . . . . 31

4.2 A ranking game form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3 A ranking game associated with the ranking game form of Figure 4.2 . . . . 34
4.4 Four-player ranking game in which all equilibria are pure . . . . . . . . . . 37
4.5 Mapping from binary two-player games to three-player single-loser games . 39
4.6 Iterated weak dominance solvability in two-player ranking games . . . . . . 43
4.7 Three-player ranking game used in the proof of Theorem 4.7 . . . . . . . . 44
4.8 Three-player ranking game used in the proof of Theorem 4.10 . . . . . . . . 48
4.9 Three-player single-winner game used in the proof of Theorem 4.11 . . . . . 48
4.10 Three-player ranking game used in the proof of Theorem 4.11 . . . . . . . . 49
4.11 Three-player ranking game used in the proof of Theorem 4.14 . . . . . . . . 52
4.12 Four-player ranking game used in the proof of Theorem 4.14 . . . . . . . . 53
4.13 Three-player ranking game Γ5 used in the proof of Theorem 4.15 . . . . . . 54
4.14 Dual linear program for computing a correlated equilibrium . . . . . . . . . 55

5.1 Inclusion relationships between anonymous, symmetric, self-anonymous,

and self-symmetric games . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.2 Relationships between the payoffs of anonymous, symmetric, self-anony-
mous, and self-symmetric games . . . . . . . . . . . . . . . . . . . . . . . . 62
5.3 Anonymous game with a unique, non-symmetric Nash equilibrium . . . . . 64
5.4 Matching problem for the game of Figure 5.3 . . . . . . . . . . . . . . . . . 65
5.5 Game used in the proof of Theorem 5.3 . . . . . . . . . . . . . . . . . . . . 66
5.6 Integer flow network for the game of Figure 5.3 . . . . . . . . . . . . . . . . 76
5.7 A matrix and a sequence of eliminations . . . . . . . . . . . . . . . . . . . . 80
5.8 Matrix Y used in the proof of Lemma 5.18 . . . . . . . . . . . . . . . . . . . 81
5.9 Overall structure of the graph used in the proof of Theorem 5.22 . . . . . . 87
5.10 Variable gadget used in the proof of Theorem 5.22 . . . . . . . . . . . . . . 88
5.11 Clause gadget used in the proof of Theorem 5.22 . . . . . . . . . . . . . . . 89
5.12 Gadget to consume remaining labels, used in the proof of Theorem 5.22 . . 90

ix
x

5.13 Labeled graph for the matrix elimination instance of Figure 5.7 . . . . . . . 91
5.14 Payoffs of a player in a self-anonymous game with three players and three
actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.15 Payoff structure of a player of the self-anonymous game used in the proof
of Theorem 5.27 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

6.1 Payoffs for input, positive literal, and negative literal players, used in the
proof of Theorem 6.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.2 Payoffs for AND and OR players, used in the proof of Theorem 6.3 . . . . . 106
6.3 Payoffs for NAND players, used in the proof of Theorem 6.11 . . . . . . . . 112
6.4 Output gadget, used in the proof of Theorem 6.11 . . . . . . . . . . . . . . 114
6.5 Equality gadget, used in the proof of Theorem 6.12 . . . . . . . . . . . . . . 115
6.6 NAND gadget, used in the proof of Theorem 6.12 . . . . . . . . . . . . . . . 116
6.7 Neighborhood graph of a graphical game with seven players, corresponding
to the hypergraph given by the lines of the Fano plane . . . . . . . . . . . . 117
6.8 Graphical game with eight players and neighborhoods of size four, used in
the proof of Theorem 6.17 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.9 NOR gadget, used in the proof of Theorem 6.17 . . . . . . . . . . . . . . . . 123

7.1 Single-winner game, repeated from Figure 4.1 . . . . . . . . . . . . . . . . . 128

7.2 Linear programs for computing minimax strategies in zero-sum games . . . 130
7.3 Linear program for computing quasi-strict equilibria in zero-sum games . . 131
7.4 Linear program for computing quasi-strict equilibria in symmetric zero-sum
games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.5 Somebody has to do the dishes. . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.6 Payoff structure of a symmetric game with two actions . . . . . . . . . . . . 133
7.7 Three-player game used in the proof of Theorem 7.5 . . . . . . . . . . . . . 135

8.1 Strict and weak saddles of a zero-sum game . . . . . . . . . . . . . . . . . . 139

8.2 Symmetric zero-sum game with multiple weak saddles . . . . . . . . . . . . 140
8.3 Payoff structure of a 4k × 4k symmetric zero-sum game with at least 5k
weak saddles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
 Abstract

Game theory is a branch of applied mathematics studying the interaction of self-interested

entities, so-called agents. Its central objects of study are games, mathematical models
of real-world interaction, and solution concepts that single out certain outcomes of a
game that are meaningful in some way. The solutions thus produced can then be viewed
both from a descriptive and from a normative perspective. The rise of the Internet as
a computational platform where a substantial part of today’s strategic interaction takes
place has spurred additional interest in game theory as an analytical tool, and has brought
it to the attention of a wider audience in computer science.
An important aspect of real-world decision-making, and one that has received only
little attention in the early days of game theory, is that agents may be subject to resource
constraints. The young field of algorithmic game theory has set out to address this short-
coming using techniques from computer science, and in particular from computational
complexity theory. One of the defining problems of algorithmic game theory concerns
the computation of solution concepts. Finding a Nash equilibrium, for example, i.e., an
outcome where no single agent can gain by changing his strategy, was considered one
of the most important problems on the boundary of P, the complexity class commonly
associated with efficient computation, until it was recently shown complete for the class
PPAD. This rather negative result for general games has not settled the question, however,
but immediately raises several new ones: First, can Nash equilibria be approximated, i.e.,
is it possible to efficiently find a solution such that the potential gain from a unilateral
deviation is small? Second, are there interesting classes of games that do allow for an
exact solution to be computed efficiently? Third, are there alternative solution concepts
that are computationally tractable, and how does the value of solutions selected by these
concepts compare to those selected by established solution concepts?
The work reported in this thesis is part of the effort to answer the latter two questions.
We study the complexity of well-known solution concepts, like Nash equilibrium and
iterated dominance, in various classes of games that are both natural and practically
relevant: ranking games, where outcomes are rankings of the players; anonymous games,
where players do not distinguish between the other players in the game; and graphical
games, where the well-being of any particular player depends only on the actions of a
small group other players. In ranking games, we further compare the payoffs obtainable
in Nash equilibrium outcomes with those of alternative solution concepts that are easy to

xi
xii

compute. We finally study, in general games, solution concepts that try to remedy some
of the shortcomings associated with Nash equilibrium, like the need for randomization to
achieve a stable outcome.
 Zusammenfassung

Die Spieltheorie ist ein Teilgebiet der angewandten Mathematik, das sich mit der Interak-
tion eigennütziger Akteure, so genannter Agenten, beschäftigt. Sie untersucht dazu Spiele,
mathematische Modelle in der realen Welt auftretender Interaktion, und Lösungskonzepte,
die bedeutsame Ergebnisse eines Spieles hervorheben. Die so erhaltenen Lösungen können
dann sowohl aus deskriptiver als auch aus normativer Sicht betrachtet werden. Der Auf-
stieg des Internet zu einer Umgebung, die für einen erheblichen Teil heutiger strategischer
Interaktion verantwortlich ist, hat das Interesse an Spieltheorie als analytischem Werkzeug
weiter vorangetrieben und ihr insbesondere zu einem erhöhten Bekanntheitsgrad in der
Informatik verholfen.
Einem wichtigen Aspekt der Entscheidungsfindung in der realen Welt wurde in den
Gründungstagen der Spieltheorie nur geringe Aufmerksamkeit zu Teil. Er betrifft die
Tatsache, dass Agenten üblicherweise Ressourcenbeschränkungen unterliegen. Das junge
Gebiet der algorithmischen Spieltheorie hat nun damit begonnen, sich diesem Defizit
mittels Techniken der Informatik, und insbesondere der Komplexitätstheorie, zu wid-
men. Ein zentrales Problem der algorithmischen Spieltheorie stellt die Berechnung von
Lösungskonzepten dar. Das Finden eines Nash-Gleichgewichts, d.h. eines Ergebnisses, in
dem kein Agent durch Änderung seiner eigenen Strategie eine Verbesserung erreichen kann,
galt beispielsweise als eines der wichtigsten Probleme an der Grenze der üblicherweise
mit effizienter Berechnung verbundenen Komplexitätsklasse P, bis kürzlich seine Voll-
ständigkeit für die Klasse PPAD gezeigt wurde. Dieses eher als negativ einzuschätzende
Resultat in Bezug auf allgemeine Spiele hat die Frage an sich jedoch keineswegs vollständig
beantwortet, sondern wirft umgehend neue Fragen auf: Können Nash-Gleichgewichte an-
genähert werden, d.h., kann effizient eine Lösung berechnet werden, die den möglichen
Gewinn durch einseitige Abweichung gering hält? Gibt es interessante Teilklassen von
Spielen, die die effiziente Berechnung exakter Lösungen erlauben? Existieren schließlich
alternative, effizient berechenbare Lösungskonzepte, und wie verhält sich ihr Nutzen zu
dem etablierter Lösungskonzepte.
Diese Arbeit beschäftigt mit den beiden letzteren Fragen. Wir untersuchen dazu die
Komplexität bekannter Lösungskonzepte, wie Nash-Gleichgewicht und iterierte Dominanz,
in verschiedenen natürlichen und praktisch relevanten Klassen von Spielen: Ranglisten-
spielen, in denen jedes Ergebnis eine Rangliste der Spieler ist; anonymen Spielen, in denen
die Spieler nicht zwischen anderen Spielern unterscheiden; und graphischen Spielen, bei

xiii
xiv

denen das Wohlergehen eines bestimmten Spielers nur von einem kleinen Teil der an-
deren Spieler abhängt. In Ranglistenspielen vergleichen wir außerdem den Nutzen von
Ergebnissen im Nash-Gleichgewicht mit denen alternativer, effizient berechenbarer Lö-
sungskonzepte. Schließlich betrachten wir in allgemeinen Spielen Lösungskonzepte, die
einige mit Nash-Gleichgewichten verbundene Schwächen zu beheben suchen, wie etwa die
Notwendigkeit von Randomisierung zum Erreichen eines stabilen Ergebnisses.
 Acknowledgements

To my supervisor Felix Brandt, for allowing me to roam freely and talk impudently, for
knowing an answer to every question and a question to every answer. I take pride in
being his first Ph.D. student, there will no doubt be many more. To Lane Hemaspaandra
and Martin Hofmann, for agreeing to act as referees for this thesis. To my coauthors and
collaborators, for teaching me everything I know about research. Obviously, this thesis
would not have been written without them. To Markus Brill, Hermann Gruber, and Paul
Harrenstein for proofreading parts of this thesis and providing valuable feedback. To Jeff
Rosenschein, for inviting me to work in his group at the Hebrew University of Jerusalem
and take part in the family Seder. To my friends in and outside academia, for the fun we
had. To my family, for supporting me in every imaginable way.

xv
 Chapter 1

Introduction

One does not have to be a follower of Machiavelli, nor a pessimist, to acknowledge that the
world is pervaded by conflict. Conflicts arise almost automatically whenever a situation
may lead to different outcomes depending on the choices of several individuals, or agents,
who disagree about some aspect of these outcomes. As a consequence, the study of
conflicts among a group of agents is at the core of many academic disciplines within the
social sciences, like sociology, economics, or political science. Ultimately, this study is
driven by two questions. From an individual’s point of view, what is the best way to act
in a given situation in order to achieve one’s goals? From the point of view of society as a
whole, how can the negative effects of conflicts be alleviated, in order to achieve outcomes
that benefit all.
In 1944, von Neumann and Morgenstern broke ground to a rigorous mathematical
treatment of the study of conflict with their work “Theory of Games and Economic Be-
havior.” Game theory studies mathematical models, called games, that abstract from
conflict situations in the real world and focus on certain aspects that seem worth study-
ing. Consider a situation in which agents move sequentially, taking turns in a particular
way that may or may not depend on previous choices or include an element of chance.
One way to represent such a situation by a game of strategy, commonly referred to as the
strategic form or normal form of a game, starts from a set of players, and a set of strate-
gies for each of the players. Each of these strategies provides a player with a complete
plan of action for any eventuality that might arise in the course of the game. A strategy
profile, consisting of exactly one strategy for each player, thus completely determines the
outcome of the game and leads to one of several possible outcomes. The force that drives
agents’ behavior in the real world is modeled by having players entertain preferences over
outcomes. Rational behavior is then characterized in terms of solution concepts that
single out certain strategy profiles. A famous example is the solution concept of Nash
equilibrium, which requires the strategies of the different players to be best responses to
each other, such that no agent can achieve a preferred outcome by unilaterally changing
his own strategy. In general, a solution concept can be viewed as prescribing a certain
behavior by each agent in a given situation, but also as describing the outcome that will

1
2 1 · Introduction

arise from the interaction of rational agents in this situation.

An important technique in the analysis of games is their classification according to
natural parameters like the number of players and actions, but also according to the
structure of players’ preferences. Well-known classes, which also contain many games
relevant in practice, are those of (two-player) zero-sum games, in which the interests of
two players are diametrically opposed, or anonymous games, in which no distinction is
made between different other players.
Like any area of applied mathematics, game theory walks the thin red line that is the
appropriate level of abstraction: abstract enough to be handled in a rigorous way, and
general enough such that the results thus obtained are sufficiently interesting and relevant
to the real world. An important aspect of real-world decision-making, and one that has
received only little attention in the “early days” of game theory, is that decision-makers
may be subject to resource constraints. Game theory avoids this issue by assuming perfect
rationality, i.e., has each agent choose an action that given a certain state of knowledge
leads to his most preferred outcome. The importance of the issue of resource bounded
reasoning, however, has by no means escaped the attention of game theorists. Nobel
laureate Robert Aumann for example expressed the following opinion in an interview
with van Damme (1998, pp. 201–202):

It is important to have an applicable model. It sounds a little like the man who
had lost his wallet and was looking under the lamppost for it. His friend asked
him: Why do you look only under the lamppost? And he answered: That’s
because there is light there, otherwise I wouldn’t be able to see anything. It
sounds crazy, but when you look at it more closely it is really important. If
you have a theory that somehow makes a lot of sense, but is not calculable, not
possible to work with, then what’s the good of it? As we were saying, there
is no “truth” out there; you have to have a theory that you can work with in
applications, be they theoretical or empirical. . . .
. . . My own viewpoint is that inter alia, a solution concept must be calcu-
lable, otherwise you are not going to use it.

Calculability as used by Aumann appears to be grounded in an informal notion of

convenience experienced by humans working with a given solution concept. It is thus not
immediately clear how it should be treated in a rigorous way. Enter another field that
was greatly influenced by work of von Neumann (1945): computer science. With merely
a year between two key publications in either field, both of which were authored or co-
authored by von Neumann, it took another half century until algorithmic game theory
set out to reveal deep connections between game theory and computer science (see, e.g.,
Nisan et al., 2007). One of these connections became obvious by the rise of the Internet
as a computational platform and spurred the interest of computer scientists in game
theory as a framework to analyze interaction between self-interested entities. Another
one concerns Aumann’s calculability, and involves the areas of algorithms and complexity
 3

theory. Complexity theory classifies problems according to the resources some idealized
computational device requires to solve them. In the case of the Turing machine, the
resulting notion of complexity is particularly meaningful: “Despite its weak and clumsy
appearance, the Turing machine can simulate arbitrary algorithms with inconsequential
loss of efficiency” (Papadimitriou, 1994a, p. 19).
Papadimitriou (2007) recasts Aumann’s statement in the framework of computational
complexity theory, and further argues that this statement still applies if one subscribes to
a purely descriptive or analytical view of solution concepts:

But why should we be interested in the issue of computational complexity in

connection to Nash equilibria? After all, a Nash equilibrium is above all a
conceptual tool, a prediction about rational strategic behavior by agents in
situations of conflict—a context that is devoid of computation.
We believe that this matter of computational complexity is one of central
importance here, and indeed that the algorithmic point of view has much to
contribute to the debate of economists about solution concepts. The reason
is simple: If an equilibrium concept is not efficiently computable, much of
its credibility as a prediction of the behavior of rational agents is lost—after
all, there is no clear reason why a group of agents cannot be simulated by
a machine. Efficient computability is an important modeling prerequisite for
solution concepts. (Papadimitriou, 2007, pp. 29–30)

For the very reasons outlined above, the computational complexity of game-theoretic
solution concepts has come under increased scrutiny. The work reported on in this thesis
is part of the endeavor to study algorithmic aspects of game-theoretic solutions. We con-
centrate on the very general class of games in normal form, and various natural subclasses.
In the following chapter, we formally define normal-form games along with several well-
known solution concepts, and lay out certain elements of the framework of computational
complexity theory. In Chapter 3, we then take a closer look at the current state of the
art and outline the contribution of this thesis. Since many problems have turned out to
be hard for general games, researchers have considered various restricted classes of games.
This is also the approach we take for the main part of this thesis. In Chapters 4 through 6,
we study the computational complexity of various game-theoretic solution concepts in four
natural classes of normal-form games. A more detailed account of the significance of each
individual class and relevant existing work will be given in the respective chapter. Finally,
in Chapters 7 and 8, we consider two solution concepts that are less well-known and try to
address two shortcomings of Nash equilibrium: the potential indifference between actions
that are played and actions that are not played, and the need for randomness in the choice
of actions.
 Chapter 2

Games, Solutions, and Complexity

In this chapter we review relevant concepts from game theory and computational com-
plexity theory. While parts of this chapter also provide a high-level overview of the objects
this thesis is concerned with, and the techniques used to analyze them, its main purpose
is to lay out in detail the formal framework for the results that appear in later chapters.
It may therefore be advisable to skip some of the technical details for the time being, and
return to this chapter later for reference. For additional details we refer to the textbooks of
Myerson (1991) and Osborne and Rubinstein (1994) on game theory, and of Papadimitriou
(1994a), Goldreich (2008), and Vollmer (1999) on complexity theory.

2.1 Strategic Games

This thesis is concerned with finite games in normal form. Such a game is given by
a finite set of players, and a finite non-empty set of actions for each player. Players
move simultaneously to select an action profile, containing exactly one action for each
player. This selection leads to a unique outcome, and each player entertains preferences
over the set of possible outcomes. A standard assumption, which we also follow in this
thesis, restricts attention to von Neumann-Morgenstern preferences over lotteries over
outcomes (von Neumann and Morgenstern, 1944). Under this assumption, the preferences
of a player can be represented by a real-valued payoff function from the set of action profiles
into the reals. Two games are equivalent if there exist bijections between their respective
sets of players and actions, and if the corresponding payoff functions can be obtained
from each other via positive affine transformations. All solution concepts we consider are
invariant under such transformations. We arrive at the following definition (e.g., Myerson,
1991).

Definition 2.1 (normal-form game). A (normal-form) game is given by a tuple Γ =

(N, (Ai )i∈N , (pi )i∈N ) where N is a finite set of players, and for each player i ∈ N, Ai is
a nonempty and finite set of actions available to i and pi : ("i∈N Ai ) → R is a function
mapping each action profile, i.e., each combination of actions, to a real-valued payoff for i.

5
6 2 · Games, Solutions, and Complexity

We write AN = "i∈N Ai for the set of action profiles and n = |N| for the number of
players in a game. Subscripts will generally be used to identify the player to which an
action belongs, superscripts to index the actions of a particular player. For example, we
write ai for a typical action of player i, and aji for the jth action of player i. In the case of
games with few players, or when we do not explicitly distinguish between specific players,
we also use lower case roman letters aj , bj , etc., for the players’ actions. An action profile
(ai )i∈N ∈ AN we abbreviate by aN .
A necessary condition for studying the computational properties of games is that these
games have a finite representation. We therefore restrict our attention to games whose
payoffs are rational numbers, and simply refer to these as “games” throughout the thesis.
We further call a game binary if pi (aN ) ∈ {0, 1} for all i ∈ N and aN ∈ AN . A two-player
game ({1, 2}, (A1 , A2 ), (p1 , p2 )) is alternatively called a bimatrix game, because it can be
represented by two matrices M1 and M2 with rows and columns indexed by A1 and A2 ,
respectively, and Mi (a1 , a2 ) = pi (a1 , a2 ) for i ∈ {1, 2} and all a1 ∈ A1 , a2 ∈ A2 . A
two-player game satisfying p1 (a, b) = −p2 (a, b) for all (a, b) ∈ A1 × A2 is called zero-
sum game or matrix game, and can be represented by a single matrix M containing
the payoffs for the first player. Since all solution concepts considered in this thesis are
invariant under positive affine transformations, the results about zero-sum games in fact
apply to the larger class of constant-sum games, in which the payoffs of the two players
always sum up to the same constant. For games with more than two players, this property
is far less interesting, as we can always add an extra player who “absorbs” the payoffs of
the others (von Neumann and Morgenstern, 1947).
The concept of an action profile can be generalized to that of a mixed strategy profile
by letting players randomize over their actions. We have Si = ∆(Ai ) denote the set of
probability distributions over player i’s actions, the mixed strategies available to player i,
and SN = "i∈N Si the set of mixed strategy profiles. Analogously to action profiles,
we abbreviate a strategy profile (si )i∈N ∈ SN by sN . In the following, A−i and S−i
respectively denote the set of action and strategy profiles for all players but i. Accordingly,
we write a−i ∈ A−i for the vector of all actions in aN but ai , and s−i ∈ S−i for the vector
of all strategies in sN but si . We further denote by si (ai ) and sN (ai ) the probability
player i assigns to action ai in strategy si or strategy profile sN . The pure strategy si
such that si (ai ) = 1 we identify with ai whenever this causes no confusion. Moreover,
we use (s−i , ti ) to refer to the strategy profile obtained from sN by replacing si by ti .
Payoff functions naturally extend to mixed strategy profiles, and we will frequently write
P Q
pi (sN ) = aN ∈AN pi (aN )( i∈N si (ai )) for the expected payoff of player i, and p(sN ) =
P
i∈N pi (sN ) for the social welfare under strategy profile sN ∈ SN . For better readability
we usually avoid double parentheses and write, e.g., p(s−i , ti ) instead of p((s−i , ti )).
To illustrate these concepts, consider a situation in which Alice, Bob, and Charlie
are to designate one of them as the winner. They do so by raising their hand or not,
simultaneously and independently of one another. Alice wins if the number of hands
raised, including her own, is odd, whereas Bob is victorious if this number equals two.
2.2 · Solution Concepts 7

c1 c2
b1 b2 b1 b2
a1 (0, 0, 1) (1, 0, 0) (1, 0, 0) (0, 1, 0)
a2 (1, 0, 0) (0, 1, 0) (0, 1, 0) (1, 0, 0)

Figure 2.1: Alice, Bob, or Charlie? Alice chooses row a1 or a2 , Bob chooses column b1
or b2 , and Charlie chooses matrix c1 or c2 . Outcomes are denoted by a vector where the
ith component is the payoff to player i.

Should nobody raise their hand, Charlie wins. The normal form of this game is shown
in Figure 2.1. Player 1, Alice, chooses between rows of the table, labeled a1 and a2 .
Action a1 corresponds to her not raising her hand, whereas a2 corresponds to her raising
her hand. Similarly, player 2, Bob, chooses between the left or right column, labeled b1
and b2 , and player 3, Charlie, between the left or right matrix, labeled c1 and c2 . Out-
comes are denoted as vectors of payoffs, the ith component corresponding to the payoff
of player i. The fact that a player wins or loses is represented by a payoff of one or zero,
respectively. For example, the top right entry in the left matrix corresponds to the action
profile (a1 , b2 , c1 ) where only Bob raises his hand, which in turn causes Alice to win.

2.2 Solution Concepts

Now that we have found a way to formalize the conflict between Alice, Bob, and Charlie,
how should they play in order to be successful? Game theory tries to answer this question
in a general way by providing a number of solution concepts. On a normative interpre-
tation, solution concepts identify reasonable, desirable, or otherwise significant strategy
profiles in games.
Perhaps the most cautious way for a player to proceed is to ensure a certain minimum
payoff even if all other players were to conspire against him.

Definition 2.2 (maximin strategy and security level). Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be
a normal-form game. A strategy s∗i ∈ Si is called a maximin strategy for player i ∈ N if

s∗i ∈ argmax min pi (si , s−i ).

si ∈Si s−i ∈S−i

The value vi = maxsi ∈Si mins−i ∈S−i pi (si , s−i ) is called the security level of player i.

Given a particular game Γ , we write vi (Γ ) for the security level of player i in Γ . In the
game of Figure 2.1, Alice can guarantee a payoff of at least 1/2 by uniformly randomizing
over her actions, i.e., by raising her hand with probability 1/2. We leave it to the reader
to verify that this is indeed her security level, and that the security level for both Bob
and Charlie is zero.
8 2 · Games, Solutions, and Complexity

A different way to identify desirable strategy profiles is to rule out those that are not
desirable. An action of a particular player in a game is said to be weakly dominated if
there exists a strategy guaranteeing him at least the same payoff for any profile of actions
of the other players, and strictly more payoff for some such action profile. A dominated
action may be discarded for the simple reason that the player will never face a situation
where he would benefit from using this action. The removal of one or more dominated
actions from the game may render additional actions dominated, and the solution concept
of iterated dominance works by removing a dominated action and applying the same
reasoning to the reduced game.

Definition 2.3 (iterated dominance). Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a game. An
action di ∈ Ai is said to be (weakly) dominated by strategy si ∈ Si if for all b ∈ "i∈N Ai ,
P
pi (b−i , di ) 6 ai ∈Ai si (ai )pi (b−i , ai ) and for at least one b̂ ∈ "i∈N Ai , pi (b̂−i , di ) <
P
ai ∈Ai si (ai )pi (b̂−i , ai ).
An elimination sequence of Γ is a finite sequence of actions in ∪i∈N Ai . For a
particular elimination sequence d = (d1 , d2 , . . . , dk ) denote by Γ (d) the induced sub-
game where the actions in d have been removed, i.e., Γ (d) = (A10 , A20 , u|A10 ×A20 ) where
A10 = A1 \ {d1 , d2 , . . . , dk } and A20 = A2 \ {d1 , d2 , . . . , dk }. Then, an elimination se-
quence d = (d1 , d2 , . . . , dm ) of Γ is called valid if either it is the empty sequence, or if
(d1 , d2 , . . . , dm−1 ) is valid in Γ and dm is weakly dominated in Γ (d1 , d2 , . . . , dm−1 ).
An action a ∈ ∪i∈N Ai is called eliminable if there exists a valid elimination sequence d
such that a is weakly dominated in Γ (d). Game Γ is called solvable if it is possible to
obtain a game where only one action remains for each player, i.e., if there is some valid
elimination sequence d such that Γ (d) = (N, (Ai0 )i∈N , (pi0 )i∈N ) with |Ai0 | = 1 for all i ∈ N.

Again consider the game of Figure 2.1. Charlie never wins by raising her hand, but
sometimes wins by not doing so, such that c1 dominates c2 . Assuming that Charlie never
raises her hand, b2 dominates b1 . Finally assuming that Bob will always raise his hand, a1
dominates a2 , and the only remaining action profile is the one where Bob alone raises his
hand.
In general, the result of iterated weak dominance elimination depends on the order in
which actions are removed, since the elimination of an action may render actions of another
player undominated (e.g., Apt, 2004). This is in contrast to iterated strict dominance,
which requires the inequality to be strict for every action profile of the other players. We
consider two problems concerning iterated dominance in this thesis. Iterated dominance
solvability (IDS) asks whether for a given game Γ there exists a sequence of eliminations
P
of length i∈N (|Ai | − 1), i.e., one that leaves only one action for each player. Iterated
dominance eliminability (IDE) is given an action ai ∈ Ai of some player i ∈ N and
asks whether it is possible to eliminate ai . Our results are often fairly robust as to the
particular way these problems are defined. For example, results about IDE can easily be
adapted to the problem of deciding whether some action of a particular player can be
eliminated.
2.2 · Solution Concepts 9

b1 b2
a1 (1, 0) (0, 1)
a2 (0, 1) (1, 0)

Figure 2.2: The Matching Pennies game, a constant-sum game without pure equilibria.
Each of two players turns a penny to heads or tails. The first player wins if both coins
show the same side, otherwise the second player wins.

A restricted variant of (iterated) dominance can be obtained by requiring that the

dominating strategy si be pure. We will frequently exploit that the two variants are
equivalent, which obviously holds for games with two actions, but also for games with only
two different payoffs (Conitzer and Sandholm, 2005a).1 Unless explicitly stated otherwise,
results hold for dominance by pure strategies and for dominance by mixed strategies. An
alternative definition of iterated dominance allows for the deletion of a set of dominated
actions in each step (e.g., Apt, 2004). A different notion of solvability merely requires
the remaining action profiles to yield a unique payoff to each of the players (e.g., Moulin,
1979). We note, but do not show explicitly, that all hardness and tractability results
extend to these definitions as well.
One of the best-known solution concept for strategic games is Nash equilibrium (Nash,
1951). Nash equilibrium requires that the strategy of each player is a best response to
the other players’ strategies, such that no player could increase his payoff by unilaterally
deviating and playing another strategy.

Definition 2.4 (Nash equilibrium). A strategy profile s∗N ∈ SN is called Nash equilib-
rium if for each player i ∈ N and every strategy si ∈ Si ,

pi (s∗N ) > pi (s∗−i , si ).

A Nash equilibrium is called pure if it is a pure strategy profile.

An equilibrium of the game in Figure 2.1 is for example attained when Alice and
Charlie do not raise their hands, and Bob raises his hand with probability at least one
half. The game thus possesses infinitely many equilibria. We leave it to the reader to
verify that the only pure equilibrium is the action profile where Bob alone raises his hand.
Nash (1951) has shown that every normal-form game possesses at least one equilib-
rium. Since the proof is not constructive, it makes sense to consider the problem of finding
an equilibrium of a given game. Pure Nash equilibria, on the other hand, are not guar-
anteed to exist, as is illustrated by the well-known Matching Pennies game depicted in
Figure 2.2. If they do exist, however, they have two distinct advantages over mixed ones.
For one, requiring randomization in order to reach a stable outcome has been criticized for
1
The game in Figure 2.1 clearly satisfies both of these properties.
10 2 · Games, Solutions, and Complexity

various reasons. In multi-player games, where action probabilities in equilibrium can be

irrational numbers, randomization is particularly questionable. Secondly, pure equilibria
as computational objects are usually much smaller in size than mixed ones. We will thus
also consider the problem of deciding, for a given game, whether this game possesses at
least one pure Nash equilibrium.
Some additional solution concepts will be introduced later in the thesis. Correlated
equilibrium generalizes Nash equilibrium by assuming the existence of a device or trusted
third party that selects actions according to some joint probability distribution, and in-
forms each player only about his own action (Aumann, 1974). In Chapter 4 we compare
the quality of Nash and correlated equilibria in a specific setting. Quasi-strict equilib-
rium, on the other hand, refines Nash equilibrium by requiring that every best response
is played with positive probability (Harsanyi, 1973). In Chapter 7 we analyze the com-
plexity of quasi-strict equilibrium in general strategic games and in some classes of games
studied in earlier chapters, and also use it to shed some light on certain peculiarities of
Nash equilibrium in the setting of Chapter 4. Finally, in Chapter 8, we consider a class
of ordinal set-valued solution concepts due to Shapley (1964). These solution concepts,
called saddles by Shapley, replace the notion of stability underlying Nash equilibrium
by a more elementary one that is based on dominance, thereby eliminating the need for
randomization as a prerequisite for the existence of a stable outcome.

2.3 Elements of Complexity Theory

The reasoning of real-world agents, both human and artificial, is often restricted by bounds
on resources, like the time available for a thought process or the capacity of their memory.
It is thus natural to study the resource requirements of game-theoretic solution concepts,
for the obvious reason that solutions that cannot be found in practice are of very limited
value. This holds irrespective of the fact whether a solution concept is to be used in an
analytical or purely descriptive way. Computational complexity theory provides a rigorous
mathematical framework to address this type question, and we introduce the necessary
concepts in this section.
Complexity theory assigns problems to different complexity classes, each of which is
characterized by several parameters: the underlying computational model, a computa-
tional paradigm, a resource, and an upper bound on this resource. The computational
model describes the basic operations that can be used in a computation. A prominent
example are Turing machines, which provide an abstract and idealized view of today’s
personal computers but are in fact able to compute any function that one would intu-
itively consider computable with only inconsequential loss of efficiency. Another example
are Boolean circuits, which formalize the type of parallel computation performed by inte-
grated circuits. The computational paradigm determines in which way the computation
is performed. In a deterministic Turing machine, every intermediate state of a compu-
tation has exactly one followup state. A nondeterministic machine, on the other hand,
2.3 · Elements of Complexity Theory 11

may investigate several followup states of each state at once, and then use the result of
one of the branches as the overall result. This latter mode can alternatively be interpreted
as verifying a given solution to a problem. The assumption that a solution was already
known is of course unrealistic, but nondeterminism has nevertheless proven very useful in
analyzing computational problems. For a Boolean circuit, the paradigm is given by the
type of the gates the circuit is composed of. Finally, a complexity class is characterized by
a particular resource, like the time or space required for the computation, or the depth
and overall number of gates of a Boolean circuit, as well as an upper bound on this
resource as a function on the size of the problem instance.
Unfortunately, the current state of complexity theory in many cases does not allow for
the separation of complexity classes, i.e., for a distinction of problems that can or cannot
be solved under certain resource constraints. Quite often, however, like in the famous
case of the classes P and NP, there is fairly strong evidence that two complexity classes
are indeed distinct. What can be done using current techniques is to identify the hardest
problems in each class, i.e., those not contained in a smaller class should the two be
distinct. Hardness is established via reductions that transform instances of an arbitrary
problem in a particular class into those of the problem in question. It is easy to see that
the reductions we employ in this thesis compose, so a reduction from a problem that is
itself hard for a particular class effectively shows that the problem in question cannot be
easier to solve than any problem in that class.
When introducing the necessary concepts from complexity theory we restrict our at-
tention to functions whose input and output are finite strings of bits. Definitions and
results then carry over to more general functions by observing that their input and out-
put can be encoded as bitstrings. Functions with several arguments, for example, can
easily be obtained from the one-argument case by introducing a new symbol, say “◦”,
using this symbol to separate the different arguments, and then encoding each of the
three values 0, 1, and ◦ by a pair of bits. Some issues related to encodings of games, and
our interpretation of these issues, will briefly be discussed in Section 2.4. In most cases,
however, it will be clear that an encoding with the desired properties exists, and we will
avoid dealing with the details of any particular encoding in these cases.
Let us define the basic concepts more formally. In the context of this thesis, an
algorithm for computing a function f will consist of a finite set of instructions describing
how f(x) can be obtained for an arbitrary input x ∈ {0, 1}∗ . The algorithm is allowed
to use a scratchpad to write down intermediate results and, finally, the output. Each
instruction starts by reading a bit of the input and a bit from the scratchpad. Based on
the values that have been read, it then writes a bit to the scratchpad, and either halts or
chooses the next instruction be to executed. Thus, while there is only a finite number of
instructions, each of them may be executed an arbitrary number of times depending on
the input. The Turing machine formalizes this idea.

Definition 2.5 (Turing machine). A (k-tape) Turing machine is given by a tuple M =

(Q, Σ, δ, b, qs , qh ), where Q is a finite set of states, Σ is a finite alphabet, δ : Q × Σk →
12 2 · Games, Solutions, and Complexity

Q × Σk−1 × {L, R}k is a transition function, b ∈ Σ is a specific blank symbol, and qs ∈ Q

and qh ∈ Q are start and halting states, respectively.

The Turing machine’s equivalent to the scratchpad are k tapes of infinitely many
cells, each of which contains a symbol in Σ. By convention, the last tape is designated
the output tape that will eventually bear the result of the computation. For each of the
tapes, a tape head determines the current position. A configuration consists of a state,
the content of each tape, and the position of the heads. The initial configuration is the one
with state qs , the input x in the first |x| cells of the first tape, and heads on the leftmost
cell of each tape. The (infinitely many) remaining cells to the right of the input are filled
with blanks.
If at some point the Turing machine is in state q ∈ Q, the symbol at the position of
the ith head is σi , and δ(q, σ1 , σ2 , . . . , σk ) = (q 0 , σ20 , . . . , σk0 , z1 , . . . , zk ), then at the next
step the entry σi on the ith tape for i > 2 will have been replaced by σi0 , the machine
will be in state q 0 , and the ith head will have moved one cell to the left if zi = L and if
this cell exists, and one cell to the right if zi = R. By convention the first tape is assumed
to be read-only. The transition function δ is further assumed never to leave the halting
state qh once it has entered it, and not to modify the content of the tapes while in qh .
Entering state qh can thus be interpreted as halting.
We are now ready to define what it means for a Turing machine to compute a function
under resource constraints. The running time of a Turing machine will be the number of
steps before it halts.

Definition 2.6 (running time). Let f : {0, 1}∗ → {0, 1}∗ and T : N → N be two functions.
Then, f can be computed in T -time if there exists a Turing machine M with the following
property: for every x ∈ {0, 1}∗ , if M is started in the initial configuration with input x,
then after at most max(1, T (|x|)) steps it halts with f(x) written on its output tape.

Bounds on the space used by a Turing machine can be defined in a similar way. Since
we will specifically be interested in computations that require less space than is needed
to store the input, we exclude the read-only input tape.

Definition 2.7 (space bound). Let f : {0, 1}∗ → {0, 1}∗ and S : N → N be two functions.
Then, f can be computed using S-space if there exists a Turing machine M with the
following property: for every x ∈ {0, 1}∗ , if M is started in the initial configuration with
input x, then it halts with f(x) written on its output tape after a finite number of steps,
and the number of cells of tapes 2 to k that differ from b at some intermediate step is at
most max(1, S(|x|)).

The exact details of Definition 2.5, like the number of tape symbols or the number of
tapes, seem rather arbitrary. It is therefore worth noting that their effect on the time and
space needed to compute a function will not be significant for the problems considered in
this thesis. In particular, the complexity classes P and L defined below are very robust
2.3 · Elements of Complexity Theory 13

to such modifications. Similar statements also apply to the much stronger properties of
modern computers, like random access to the tape cells.
Let R ⊆ {0, 1}∗ ×{0, 1}∗ be a relation. We say that R is polynomial-time recognizable if
its characteristic function, i.e., the function r : {0, 1}∗ ×{0, 1}∗ → {0, 1} such that r(x, y) = 1
if and only if (x, y) ∈ R, can be computed in polynomial time. We call R polynomially
balanced if there exists a polynomial function p : N → N such that (x, y) ∈ R implies
|y| 6 p(|x|).
Associated with a relation R are three different problems. The decision problem asks,
for a given instance x ∈ {0, 1}∗ , whether there exists a solution y ∈ {0, 1}∗ such that
(x, y) ∈ R. The search problem is to find a solution, i.e., an element y ∈ {0, 1}∗ satisfying
(x, y) ∈ R. Finally, the counting problem asks for the number |{y ∈ {0, 1}∗ : (x, y) ∈ R}| of
solutions for x. An example for a decision problem relevant in the context of this thesis
is that of deciding, for a given normal-form game x, whether this game has a pure Nash
equilibrium y, i.e., a vector of strategies that are mutual best responses to each other. To
solve the corresponding search or counting problem one would have to do more, namely
compute such a vector whenever one exists, or count the number of different vectors
satisfying the property.

2.3.1 Decision Problems

Let us focus on decision problems for a moment. The decision problem associated with a
polynomially balanced relation R ⊆ {0, 1}∗ × {0, 1}∗ can alternatively be looked at in terms
of the language LR = {x ∈ {0, 1}∗ : (x, y) ∈ R for some y ∈ {0, 1}∗ }. In the following we
say that a Turing machine decides a language L if it computes its characteristic function
fL : {0, 1}∗ → {0, 1} such that fL (x) = 1 if and only if x ∈ L. Let us define the class P
of languages that can be decided in polynomial time, which is often used synonymously
with efficient solvability.
Definition 2.8 (the class P). For a function T : N → N, let DTIME(T ) be the set of
all languages that can be decided in c · T -time for some constant c > 0. Then, P =
∪k>1 DTIME(nk ).
We proceed to define the class NP of decision problems that can be verified efficiently.
While this is exactly the class of problems associated with polynomial-time recognizable
and polynomially balanced relations, the name NP, short for nondeterministic polynomial
time, derives from the way the class has traditionally been defined. A nondeterministic
Turing machine differs from the Turing machine of Definition 2.5 in that δ is no longer a
function mapping a configuration to a follow-up configuration, but a relation between suc-
cessive configurations. A nondeterministic Turing machine is said to decide a language L
if for each x ∈ {0, 1}∗ , the following holds if and only if x ∈ L: there exists a sequence
of configurations, connected by δ, that begins with the initial configuration for input x
and ends in the halting state with 1 written on the output tape. Running time and space
requirements are defined analogously to the deterministic case.
14 2 · Games, Solutions, and Complexity

Definition 2.9 (the class NP). For a function T : N → N, let NTIME(T ) be the set of
all languages that can be decided by a nondeterministic Turing machine in c · T -time for
some constant c > 0. Then, NP = ∪c>1 NTIME(nc ).

The relationship to polynomial-time recognizable relations yields an alternative char-

acterization: a language L is in NP if there exists a polynomial function p : N → N and a
(deterministic) Turing machine M such that for all x ∈ {0, 1}∗ , x ∈ L if and only if there
exists a certificate y ∈ {0, 1}p(|x|) such that M accepts (x, y).
The relative complexity of different decision problems can be captured in terms of
reductions. Intuitively, a reduction from one problem to another transforms every instance
of the former into an equivalent instance of the latter, where equivalence means that both
of them yield the same decision. For this transformation to preserve the complexity of
the original problem, the reduction should of course have less power than is required to
actually solve the original problem. For comparing problems in NP, the type of reduction
most commonly used is the one that can itself be computed in (deterministic) polynomial
time.

Definition 2.10 (polynomial time reduction, NP-hardness). A language P ⊆ {0, 1}∗ is

called polynomial-time (many-one) reducible to a language Q ⊆ {0, 1}∗ , denoted P 6p
Q, if there exists a function f : {0, 1}∗ → {0, 1}∗ computable in polynomial time such that
for every x ∈ {0, 1}∗ , x ∈ P if and only if f(x) ∈ Q. A language Q is called NP-hard if for
every language P in NP, P 6p Q.

A problem will be called complete for a particular class if it is both hard for and
contained in this class. It is easy to see that the relation 6p and all other reducibility
relations defined below are transitive, and that membership of a hard problem from one
class in a smaller class implies that the two classes coincide. The existence of complete
problems for particular classes is less obvious, but holds for all but one of the classes
considered in this thesis.
Let us now turn to space-bounded computation, and in particular to the class L of
decision problems that require only logarithmic space. This class is highly relevant for
problems in large open systems like the Internet, where the input is often too large to be
stored locally.

Definition 2.11 (the class L). For a function S : N → N, let SPACE(S) be the set
of all languages that can be decided using c · S-space for some constant c > 0. Then,
L = SPACE(log n).

The class NL of problems with solutions verifiable in logarithmic space is obtained

by again considering nondeterministic Turing machines. An equivalent characterization
of NL in terms of certificates exists, but requires that each bit of the certificate is read
only once.
2.3 · Elements of Complexity Theory 15

Definition 2.12 (the class NL). For a function S : N → N, let NSPACE(S) be the set of
all languages that can be decided by a nondeterministic Turing machine using c · S-space
for some constant c > 0. Then, NL = NSPACE(log n).

An appropriate type of reduction for NL is one that itself requires only logarithmic
space. Since the length of the output of such a function can be at most logarithmic in the
length of its input, the following definition uses functions for which any single bit can be
computed in logarithmic space.

Definition 2.13 (log-space reduction). A function f : {0, 1}∗ → {0, 1}∗ such that f(x) 6 |x|c
for some c > 0 and every x ∈ {0, 1}∗ is called implicitly log-space computable if the
languages Lf = {(x, i) : f(x)i = 1} and Lf0 = {(x, i) : i 6 |f(x)|} are in L. A language
P ⊆ {0, 1}∗ is called log-space reducible to a language Q ⊆ {0, 1}∗ , denoted P 6` Q, if
there exists an implicitly log-space computable function f : {0, 1}∗ → {0, 1}∗ such that for
all x ∈ {0, 1}∗ , x ∈ P if and only if f(x) ∈ Q.

In some cases it is interesting to study decision problems that are the complements of
languages in a specific complexity class, and the prefix “co” is commonly used to denote
the resulting class. The class coNP, for example, can informally be described as the class of
problems for which non-existence of a solution can be verified efficiently. It is obvious that
deterministic complexity classes are closed under complementation. Immerman (1988)
and Szelepcsenyi (1988) show that this also holds for nondeterministic space complexity
classes, and for NL and coNL in particular.

2.3.2 Search Problems

We might ask what we have lost by looking only at decisions problems. Define FNP as the
class of search problems associated with polynomial-time recognizable and polynomially
balanced relations, and FP as the subclass of these problems that are solvable in polyno-
mial time. Call the search problem of a relation R self-reducible if it can be reduced, using
an appropriate type of reduction, to the corresponding decision problem of R. It turns
out that self-reducibility holds for a large class of natural problems, and in particular for
any search problem such that the corresponding decision problem is NP-complete. This
directly implies that FP=FNP if and only if P=NP, and means that in many cases it is
indeed enough to consider only the decision version of a problem.
An interesting subclass of FNP for which this strong relationship seems to break down
is obtained by considering search problems in which every instance is guaranteed to have a
solution. An example relevant in the context of this thesis is the problem of finding a Nash
equilibrium of a given normal-form game, the existence of which has been shown by Nash
(1951). Call TFNP, for total functions in NP, the class of search problems associated with
polynomial-time recognizable and polynomially balanced relations R ⊆ {0, 1}∗ × {0, 1}∗
such that for every x ∈ {0, 1}∗ , there exists y ∈ {0, 1}∗ with (x, y) ∈ R (Megiddo and
16 2 · Games, Solutions, and Complexity

Papadimitriou, 1991). We henceforth write R to denote both the relation and the corre-
sponding search problem. Unfortunately the mathematical lemmas that ensure existence
of solutions to problems in TFNP are very diverse, and TFNP is likely not to possess any
complete problems. It therefore makes sense to study subclasses of TFNP corresponding
to the different lemmas. These classes are most conveniently defined via complete prob-
lems, so we begin by introducing a notion of reducibility among search problems. The
appropriate type of reduction basically consists of a homomorphism between two relations,
together with a second function witnessing that this homomorphism indeed preserves the
structure of the solutions. As before, reductions will also be used to define hardness for
a class of search problems: a search problem R will be called hard for a particular class if
every problem in that class reduces to R.
Definition 2.14 (reducibility between search problems). A search problem P ⊆ {0, 1}∗ ×
{0, 1}∗ is called polynomial-time (many-one) reducible to a search problem Q ⊆ {0, 1}∗ ×
{0, 1}∗ , denoted P 6p Q, if there exist two functions f : {0, 1}∗ → {0, 1}∗ and g : {0, 1}∗ ×
{0, 1}∗ → {0, 1}∗ computable in polynomial time, such that for every x ∈ {0, 1}∗ and for
every y ∈ {0, 1}∗ such that (f(x), y) ∈ B, it also holds that (x, g(x, y)) ∈ R.
We consider two particular subclasses of TFNP in this thesis, the class PPAD of
polynomial search problems guaranteed to have a solution by a directed parity argument,
and the class PLS of polynomial local search problems. Underlying both of these classes is a
large graph, which can in turn be seen as describing a generic exhaustive search algorithm
for solving problems in the respective class. The graphs underlying the two complete
problems will be graphs of functions from the set of all bitstrings of a certain length to
itself. For this, let F ⊆ {f : {0, 1}n → {0, 1}n : n ∈ N} be a set of functions, and consider
some encoding of the members of F by elements of {0, 1}∗ . Denoting by fx : {0, 1}n → {0, 1}n
the function with encoding x ∈ {0, 1}∗ , we require that |x| is polynomial in n. We further
assume that for any n > 0, the set F contains all functions f : {0, 1}n → {0, 1}n computable
by a polynomial-size Boolean circuit as introduced in Definition 2.19 below, and that the
argument length of a function can be determined from its encoding in polynomial time.
In the case of PPAD (Papadimitriou, 1994b), the underlying graph is the graph of a
partial injective function, computable in polynomial time, whose range is strictly contained
in its codomain, i.e., a graph where the in- and outdegree of every vertex is bounded by
one and some vertex has indegree zero. Such a function can for example be defined via
a pair of functions s and p such that x is mapped to y if and only if p(s(x)) = y. A
distinguished vertex with indegree zero is provided explicitly, and the set of solutions of
the search problem is defined as the set of all vertices, apart from the distinguished vertex,
whose in- or outdegree is zero.
Definition 2.15 (the class PPAD). Let X ⊆ {0, 1}∗ be the set of encodings of functions
fx : {0, 1}2n → {0, 1}2n satisfying the following condition: there exist two functions px :
{0, 1}n → {0, 1}n and sx : {0, 1}n → {0, 1}n such that for all y, z ∈ {0, 1}n , fx (y ◦ z) =
px (y) ◦ sx (z), where ◦ denotes concatenation of bitstrings. Let R ⊆ {0, 1}∗ × {0, 1}∗ be
2.3 · Elements of Complexity Theory 17

the relation such that (x, y) ∈ R if and only if x ∈ X, sx (px (0|y| )) 6= 0|y| , and either
p(s(y)) 6= y, or s(p(y)) 6= y and y 6= 0|y| . Then, PPAD is the class of all search problems P
such that P 6p R.
A local search problem is given by a set I ⊆ {0, 1}∗ of instances and three func-
tions F, N, and c. Function F assigns to each instance x ∈ I a set F(x) ⊆ {0, 1}∗ of
feasible solutions. Function N defines, for each feasible solution y ∈ F(x), a neighbor-
hood N(y, x) ⊆ F(x). Finally, c assigns an integer cost c(y, x) to each pair of an instance
x ∈ I and a solution y ∈ F(x). The actual solutions for each instance are those feasible
solutions that have optimal cost within their neighborhood, i.e., either minimal or maxi-
mal cost depending on the exact definition of the problem. Membership in PLS (Johnson
et al., 1988, Schäffer and Yannakakis, 1991) then essentially requires that I is polynomial-
time decidable, and that for each instance, all of the following can be done in polynomial
time: finding an initial feasible solution, deciding optimality of a feasible solution, and
finding a better neighbor of one that is not optimal.
In other words, a local search problem is given by a partial order on the vertex set of
an undirected graph, with solutions corresponding to either the minimal or the maximal
elements within neighborhoods. The following definition uses a graph on the set of all
bitstrings of a certain length. The cost associated with a vertex is given by the interpre-
tation of the corresponding bitstring as a binary number, and two vertices are adjacent if
the corresponding bitstrings have Hamming distance one, i.e., differ in exactly one bit.
Definition 2.16 (the class PLS). Let c : {0, 1}∗ → N be the function such that for
P|y| i
every y ∈ {0, 1}∗ , c(y) = i=1 2 yi , where yi denotes the ith bit of y. Further let
R ⊆ {0, 1} × {0, 1} be the relation such that (x, y) ∈ R if and only if |x| = |y| and
∗ ∗

c(fx (y)) 6 min { c(fx (z)) : z ∈ {0, 1}∗ , |z| = |y|, and |{i : yi 6= zi }| = 1 }. Then, PLS is the
class of all search problems P such that P 6p R.
Implicit in the definition of PLS is a standard algorithm that is guaranteed to find
a locally optimal solution for a given instance: start with an initial feasible solution,
and repeatedly find a neighbor with strictly better cost, breaking ties in some convenient
manner. The standard algorithm problem can be phrased as follows: given x ∈ I, find
the locally optimal solution output by the standard algorithm on input x. Schäffer and
Yannakakis (1991) introduce the notion of a tight reduction and show that tight reductions
compose and preserve both hardness of the standard algorithm problem and exponential
worst-case running time of the standard algorithm.
Definition 2.17 (tight PLS reduction). Let P, Q ⊆ {0, 1}∗ × {0, 1}∗ be in PLS. Then a
reduction (f, g) from P to Q is called tight if for any instance x of P there exists a set Y
of feasible solutions of f(x) with the following properties:
(i) Y contains all local optima of f(x), i.e., Y ⊇ {y ∈ {0, 1}∗ : (f(x), y) ∈ P}.

(ii) For every feasible solution z of x, a solution y ∈ Y satisfying g(x, y) = z can be

found in polynomial time.
18 2 · Games, Solutions, and Complexity

(iii) Consider a set Y 0 = {y0 , y1 , . . . , y` } of feasible solutions of f(x) such that Y 0 ∩ Y =

{y0 , y` } and for all i < `, yi+1 is a strictly better neighbor of yi . Then either
g(x, y` ) = g(x, y0 ), or g(x, y` ) is a strictly better neighbor of g(x, y0 ).

An interesting feature of all problems in PLS is that they have a fully polynomial-time
approximation scheme, i.e., they can be approximated to any factor in time polynomial
in the size of the input and in the desired approximation factor (Orlin et al., 2004).

2.3.3 Counting Problems

As for the third type of problem, define #P as the class of counting problems associated
with polynomial-time recognizable and polynomially balanced relations. In the context of
this thesis, a problem will be called #P-hard if all problems in #P reduce to it via a type
of reduction that allows us to efficiently compute the number of solutions of one problem
from that of the other. Other notions of #P-hardness that can be found in the literature
are those based on polynomial-time many-one and polynomial-time Turing reductions.

Definition 2.18 (reducibility between counting problems, #P-hardness). A counting

problem P ⊆ {0, 1}∗ × {0, 1}∗ is called polynomial-time reducible to a counting problem
Q ⊆ {0, 1}∗ × {0, 1}∗ , denoted P 6p Q, if there exist two functions f : {0, 1}∗ → {0, 1}∗
and g : {0, 1}∗ × N → N computable in polynomial time, such that for every x ∈ {0, 1}∗ ,
|{y ∈ {0, 1}∗ : (x, y) ∈ P}| = g(x, |{y ∈ {0, 1}∗ : (f(x), y) ∈ Q}|). A counting problem Q is
called #P-hard if for every problem P in #P, P 6p Q.

Hardness for the class #L of counting problems associated with polynomially balanced
relations that can be recognized in logarithmic space is defined analogously.

2.3.4 Circuit Complexity

Let us now consider a different computational model, which captures the type of parallel
computation characteristic for the lower levels of modern computers, but also for decentral-
ized systems involving many agents: the Boolean circuit. For a directed graph G = (V, E)
and a particular vertex v ∈ V, let indeg(v) and outdeg(v) denote the in- and outdegree
of v in G, i.e., indeg(v) = |{u ∈ V : (u, v) ∈ E}| and outdeg(v) = |{u ∈ V : (v, u) ∈ E}|.

Definition 2.19 (Boolean circuit). Let B be a set of Boolean functions. Then, a Boolean
circuit over B with n inputs and m outputs is a tuple C = (V, E, α, β, ω), where (V, E)
is a directed acyclic graph, α : V → N is an injective function, ω : {1, . . . , m} → V,
and β : V → B ∪ {1, . . . , n} is a function such that the following holds for all v ∈ V: if
indeg(v) = 0, then either β(v) ∈ {1, . . . , n}, or β(v) ∈ B is a 0-ary Boolean function; if
indeg(v) = k > 0, then β(v) ∈ B is a k-ary Boolean function.

A vertex v ∈ V is also called a gate of C, and indeg(v) and outdeg(v) are respectively
referred to as its fan-in and fan-out. The function α induces an ordering on any subset
2.3 · Elements of Complexity Theory 19

of V, and β assigns a type to each gate, such that each of them either corresponds to one
of the n inputs, or to a Boolean function in B with inputs given by the gate’s predecessors
in the graph. The function ω finally identifies certain vertices that yield the output of the
circuit. For a given input, we can inductively assign a unique truth value to each gate of a
circuit, and those corresponding to the m outputs in particular, such that the value of the
gate coincides with the value of the associated Boolean function, given the truth values of
the predecessors of the gate as inputs. This is true because every vertex is reachable from
one with fan-in zero, which must either correspond to an input, or to a 0-ary function,
i.e., a constant. A circuit with n inputs and m outputs can thus be seen as computing a
function f : {0, 1}n → {0, 1}m . The following definition makes this relationship explicit.

Definition 2.20 (function computed by a circuit). Let C = (V, E, α, β, ω) be a Boolean

circuit over B with n inputs and m outputs. For input x ∈ {0, 1}n , let φx : V → {0, 1} be
the unique function such that for all v ∈ V, φx (v) = xi if β(v) = i for some i ∈ N, and
φx (v) = g(φ(z1 ), φ(z2 ), . . . , φ(zk )) if β(v) = g for some g ∈ B, such that (zi , v) ∈ E for
1 6 i 6 k and α(zi ) < α(zj ) for 1 6 i < j 6 k. Then, C is said to compute a function
f : {0, 1}n → {0, 1}m if for every x ∈ {0, 1}n , (φx (ω(1)), φx (ω(2)), . . . , φx (ω(m))) = f(x).

Apart from the allowed types of gates and their fan-in, the functions computable by
a class of circuits will depend on the size and depth of circuits in the class. The size and
depth of a Boolean circuit C = (V, E, α, β, ω) are respectively defined as size(C) = |E|
and depth(C) = max{d ∈ N : there exists a path of length d in (V, E)}.
By now, a subtle difference between Turing machines and Boolean circuits may have
become apparent. Turing machines provide a uniform computational model: an algorithm
is described by a single machine that works for every input length. To describe algorithms
by Boolean circuits, on the other hand, a different circuit has to be given for each input
length, and in particular the size of these circuits may grow with the input length. One
way to relate the two computational models to each other is to define a family of circuits,
one for every input length, and require that this infinite family has a finite description.
Such a description can for example be given in terms of a Turing machine that answers
queries about the structure of the circuit for a given input length (Ruzzo, 1981). In the
context of this thesis, we will consider log-space uniform circuit families, i.e., circuit
families that can be described by a Turing machine with logarithmic space. Henceforth,
when we talk about a complexity class defined by Boolean circuits, we mean the log-space
uniform version of this class. We further say that a family of circuits computes a function f
if for every input length, the function computed by the respective member of the family
coincides with f.
We consider two different circuit complexity classes for Boolean functions. The first
one is given by circuits of polynomial size and constant depth with three types of gates
corresponding to the logical connectives NOT, AND, and OR, where the latter two types
are allowed to have unbounded fan-in. In the following, let B0 be the set containing
functions for the logical connectives NOT, AND, and OR, i.e., B0 = {¬} ∪ {∧k , ∨k : k > 0},
20 2 · Games, Solutions, and Complexity

where ¬ : {0, 1} → {0, 1} with ¬(0) = 1 and ¬(1) = 0, and for all k > 0, ∧k : {0, 1}k → {0, 1}
with ∧k (x) = 1 if and only if x = 1k , and ∨k : {0, 1}k → {0, 1} with ∨k (x) = 0 if and only
if x = 0k . Observe that in particular, ∧0 = 1 and ∨0 = 0.

Definition 2.21 (the class AC0 ). A Boolean function f : {0, 1}∗ → {0, 1} is in the class
AC0 if there exists a log-space uniform family of circuits over B0 with polynomial size and
constant depth that computes f.

It is worth noting that inputs can be copied and used multiple times without signifi-
cantly increasing the size of a circuit, e.g., by using an AND gate and a constant. Thus,
the fact that there can only be a single edge between any pair of gates is not a restriction.
The second class of interest is obtained by adding an additional type of gate, which
outputs 1 if a majority of its inputs is 1.

Definition 2.22 (the class TC0 ). For k > 0, let gk be the majority function for input
length k, i.e., gk : {0, 1}k → {0, 1} with gk (x) = 1 if |{i : xi = 1}| > |{i : xi = 0}|. A function
f : {0, 1}∗ → {0, 1} is in the class TC0 if there exists a log-space uniform family of circuits
over B0 ∪ {gk : k > 0} with polynomial size and constant depth that computes f.

It is interesting to note that among all the complexity classes defined above, the only
known separation is between AC0 and TC0 : whether the majority of bits of a bitstring is 1
cannot be decided by a circuit with polynomial size, constant depth, and unbounded fan-
in, when using only AND, OR, and NOT gates. The same can be shown for other functions,
by reducing the majority function to them. The appropriate type of reduction itself must
of course not use the majority function. In the following, for a function fk : {0, 1}k → {0, 1}n
and for 1 6 i 6 n, let fk i : {0, 1} → {0, 1} be the function computing the ith bit of f ,
k k

i.e., fi (x) = yi whenever f(x) = y1 y2 . . . yn . For a family f = (fk )k>1 of functions, let
B(f) = {fk i : k > 1, i > 1}. Intuitively, a family f of functions will be considered reducible
to another family g if it is possible to construct circuits for members of f if one is allowed
to use “oracle” gates that compute members of g at unit cost.

Definition 2.23 (constant-depth reducibility). Let f = (fk )k>1 and g = (gk )k>1 with
fk : {0, 1}k → {0, 1} and gk : {0, 1}k → {0, 1}. Then, f is called constant-depth reducible
to g, denoted f 6cd g, if for every k > 1, there exists a Boolean circuit C = (V, E, α, β, ω)
over B0 ∪ B(g) computing fk , such that size(C) is polynomial in k and there is no directed
path between vertices u, v ∈ V in (V, E) with β(u) ∈ B(g) and β(v) ∈ B(g).

It is worth noting that the size restriction effectively restricts the use of members of g
to those that have input length polynomial in k, when computing fk .

2.4 A Few Words on Encodings

As we have seen in the previous section, complexity theory measures the complexity of
a problem relative to the size of the input defining a particular problem instance. While
2.4 · A Few Words on Encodings 21

nobody would want to argue whether this is a meaningful definition, it may sometimes
allow for results that are unsatisfactory from a practical point of view. For example, the
existence of a pure Nash equilibrium can trivially be decided in polynomial time for any
game that is given explicitly, i.e., lists the payoff for every action profile: simply check
for each of them whether it satisfies the equilibrium condition. This does not take into
account, however, that the number of different action profiles of a game, and thus the
size required for the explicit representation of its payoffs, grows exponentially in a natural
parameter, the number of players. More precisely, a general game in normal form with n
players and k actions per player comprises n · kn numbers, which means exponential
growth already in the case k = 2. Computational statements over such large objects,
and efficient algorithms for problems whose input size is already exponential in a natural
parameter are of course somewhat dubious (cf. Papadimitriou and Roughgarden, 2005).
Games emanating from real-world situations, on the other hand, would certainly be
expected to possess additional structure that allows them to be played rationally and
efficiently at the same time. In fact, it is hard to imagine how the physical world could
give rise to a game the payoffs of which cannot be represented in this world. Since the
real world is what we are ultimately interested in, we will henceforth restrict ourselves to
games that can be represented in space polynomial in their natural parameters, like the
number of players or actions. We will try to characterize the complexity of solving these
games in terms of their natural parameters, while making as few assumptions as possible
about any particular encoding.
Unless explicitly stated otherwise, we assume that the number of players of a game
is polynomial in the size of its representation. We further assume that each player can
determine efficiently whether a particular action is a best response for a given action
profile of the other players, which obviously is both necessary and sufficient for playing
a game rationally and efficiently at the same time. Tractability results then hold for any
encoding satisfying these properties. Hardness, on the other hand, is established via some
encoding which allows efficient and rational play.
 Chapter 3

State of the Art and Our Contribution

It is known since the early days of game theory that the security level of a player is
the solution to a linear program, and that the two-player case is in fact equivalent to
linear programming (Dantzig, 1951). The latter problem in turn is computationally
tractable (Khachiyan, 1979). This result also extends to Nash equilibria of two-player
zero-sum games, because Nash equilibrium strategies and maximin strategies coincide in
these games (e.g., Myerson, 1991).
Among the first to study game-theoretic solution concepts in the framework of compu-
tational complexity theory, Gilboa and Zemel (1989) show that deciding the existence of
a Nash equilibrium possessing one of several natural properties is NP-hard already in two-
player games. Examples for such properties are that a given action is played with positive
probability, or that the equilibrium is not unique. A uniform reduction given by Conitzer
and Sandholm (2008) subsumes these results, and also shows that it is already NP-hard
to find an equilibrium that is reasonably close to maximizing certain properties, like
social welfare or support size. Thus, while deciding the existence of a Nash equilibrium
is trivial due to Nash’s (1951) existence result, virtually every additional property one
might require makes the problem hard. In the case of non-uniqueness of an equilibrium,
NP-hardness already holds for binary games (Codenotti and Stefankovic, 2005). Similarly,
Gilboa et al. (1993) show NP-completeness for several decision problems concerning the
iterated removal of weakly dominated actions, like eliminability of a given action, or solv-
ability of a game in the sense that only one action remains for each player. These results
were again strengthened by Conitzer and Sandholm (2005a). Iterated strict dominance,
on the other hand, had earlier been shown tractable, and in fact P-complete, by Knuth
et al. (1988).
As we have observed earlier, the problem of finding a Nash equilibrium of a two-player
game is in the class TFNP of search problems a solution of which is guaranteed to exist and
can be verified in polynomial time. More precisely, Megiddo and Papadimitriou (1991)
show membership in the class PPAD via a reduction to the computation of Brouwer fixed
points. Finding an exact lower bound for this problem was for a long time considered
a “most important concrete open question on the boundary of P” (Papadimitriou, 2001,

23
24 3 · State of the Art and Our Contribution

p. 749), and became one of the defining problems of algorithmic game theory. A series of
papers recently established that the problem is in fact PPAD-complete in the two-player
case (Goldberg and Papadimitriou, 2006, Daskalakis et al., 2006, Chen and Deng, 2006),
i.e., as hard as finding a fixed point of any continuous function from the n-dimensional unit
ball to itself. This can be seen as fairly strong evidence that the problem cannot be solved
efficiently. By a result of Abbott et al. (2005), hardness already holds for binary two-player
games. In games with more than two players, equilibria might require randomization with
irrational probabilities, in which case they cannot be computed exactly. Daskalakis et al.
(2009a) show that the problem of finding an -equilibrium in this case, i.e., a strategy
profile where for each player the potential gain from unilateral deviation is at most , is
in PPAD. Etessami and Yannakakis (2007) show that finding an -approximation of an
equilibrium in the three-player case, i.e., a strategy profile that has distance at most  from
an equilibrium in each component, is at least as hard as the square-root sum problem, a
long standing problem about arithmetic circuits that is not known to be even in NP.
These rather negative results have not settled the question, however, but immediately
raise several new ones: First, can an -equilibrium be computed efficiently? Second,
are there interesting classes of games that do allow for an exact solution to be com-
puted efficiently? Third, are there alternative solution concepts that are computation-
ally tractable, and how does the quality of solutions selected by these concepts compare
to those selected by established solution concepts? Regarding the first question, results
by Chen et al. (2006, 2007) indicate that there might not exist a fully polynomial-time
approximation scheme for Nash equilibria, i.e., an algorithm that allows them to be ap-
proximated to any factor in time polynomial in the size of the game and in the desired
approximation factor. Limited progress has however been made regarding constant factor
approximations (see Spirakis, 2008). The work reported in this thesis is part of the effort
to answer the latter two questions. In Chapters 4 through 6, we study the complexity
of various solution concepts in four natural classes of normal-form games. In Chapters 7
and 8, we then consider two solution concepts that are less well-known, and analyze their
computational complexity in both general normal-form games and in restricted classes.

Ranking Games In two-player zero-sum games, any profile of maximin strategies forms
a Nash equilibrium, and an equilibrium can therefore be found in polynomial time. It
is natural to ask whether there is some property that captures the same degree of com-
petitiveness in the multi-player case and maintains tractability of Nash equilibria. In
Chapter 4, we consider ranking games, where each outcome is a ranking of the players
and each player entertains preferences over ranks, preferring higher ranks over lower ones.
Indeed, the outcomes of many strategic situations such as parlor games or competitive
economic scenarios are rankings of the participants.1 We investigate the computational
complexity of a variety of common game-theoretic solution concepts in ranking games, and
1
The game in Figure 2.1 on Page 7 describes a special case of such a scenario, where each participant
is only interested in being ranked first.
 25

give hardness results for iterated weak dominance and mixed Nash equilibrium when there
are more than two players, and for pure Nash equilibrium when the number of players
is unbounded but the game is described succinctly. This dashes hope that multi-player
ranking games can be solved efficiently, despite their structural restrictions. Based on
these findings, we study two alternative solution concepts, maximin strategies and cor-
related equilibrium, that are known to be efficiently computable even in general games.
In particular, we provide matching upper and lower bounds for comparative ratios that
measure the quality of outcomes selected by these solution concepts relative to that of
Nash equilibrium outcomes: the price of cautiousness, the mediation value, and the
enforcement value.
Chapter 4 is based on joint work with Felix Brandt, Paul Harrenstein, and Yoav
Shoham (Brandt et al., 2006, 2007a, 2009c).

Anonymous Games Pure Nash equilibria, if they exist, can be found easily by checking
the equilibrium condition for each action profile. A problem associated with general
games, however, is their massive input complexity in the multi-player case, making it
highly questionable that such games could even be played efficiently. By contrast, one
would certainly expect games in the real world to exhibit additional structure, and to be
given in some implicit way that allows efficient play. An interesting question thus concerns
the properties of realistic games that allow for such a compact representation. Symmetries
are an example for such a property.
Strategic games may exhibit symmetries in a variety of ways. A characteristic fea-
ture, enabling the compact representation of games even when the number of players is
unbounded, is that players cannot, or need not, distinguish between the other players.
In Chapter 5, we investigate the computational complexity of pure Nash equilibrium and
iterated weak dominance in four classes of anonymous games obtained by considering
two additional properties: identical payoff functions for all players and the ability to
distinguish oneself from the other players. In contrast to other types of compactly rep-
resentable multi-player games, the pure equilibrium problem turns out to be tractable in
all four classes when only a constant number of actions is available to each player. Identical
payoff functions make the difference between TC0 -completeness and membership in AC0 ,
while a growing number of actions renders the equilibrium problem NP-hard for three of
the classes and PLS-hard for the most restricted class for which the existence of a pure
equilibrium is guaranteed. Our results also extend to larger classes of threshold anony-
mous games where players are unable to determine the exact number of players playing a
certain action. On the other hand, we show that deciding whether a game can be solved
by means of iterated weak dominance is NP-complete for anonymous games with three
actions. For the case of two actions, this problem can be reformulated quite naturally as
an elimination problem on a matrix. While enigmatic by itself, the latter turns out to be
a special case of matching along paths in a directed graph, which we show to be compu-
tationally hard in general, but also use to identify tractable cases of matrix elimination.
26 3 · State of the Art and Our Contribution

We further identify different classes of anonymous games where iterated dominance is in

P and NP-complete, respectively.
Chapter 5 is based on joint work with Felix Brandt and Markus Holzer (Brandt et al.,
2008a, 2009d).

Graphical Games A different structural element that facilitates efficient play in games
with many players is locality. While realistic situations may involve many agents, the
weal and woe of any particular agent often depends only on the decisions made by a small
part of the overall population: neighbors, business partners, or friends. Graphical games
formalize this notion by assigning to each player a subset of the players, his neighborhood,
and defining his payoff as a function of the actions of these players. More formally, a
graphical game is given by a (directed or undirected) graph on the set of players of a
normal-form game, such that the payoff of each player depends only on the actions of
his neighbors in this graph. Any graphical game with neighborhood sizes bounded by a
constant can be represented using space polynomial in the number of players.
In Chapter 6, we first strengthen a result of Gottlob et al. (2005) concerning hardness
of the pure equilibrium problem in graphical games. To be precise, we show that two
actions per player, two-bounded neighborhood, and two-valued payoff functions suffice for
NP-completeness. This is the best possible result, because deciding the existence of a pure
Nash equilibrium becomes trivial in the case of a single action for each player and tractable
for one-bounded neighborhood. In fact, we show the latter problem to be NL-complete in
general, and thus solvable in deterministic polynomial time. Interestingly, it turns out that
the number of actions in a game with one-bounded neighborhood is a sensitive parameter:
restricting the number of actions for each player to a constant makes the problem even
easier than NL, unless L=NL. In this way, we obtain a nice alternative characterization of
the determinism-nondeterminism problem for Turing machines with logarithmic space in
terms of the number of actions for games with one-bounded neighborhood.
We then turn to graphical games that additionally satisfy one of four types of anonymi-
ty within neighborhoods. We establish that deciding the existence of a pure Nash equi-
librium is NP-hard in general for all four types. Using a characterization of games with
pure equilibria in terms of even cycles in the neighborhood graph, as well as a connection
to a generalized satisfiability problem, we identify tractable subclasses of the games satis-
fying the most restrictive type of symmetry. Hardness for a different subclass is obtained
via a satisfiability problem that remains NP-hard in the presence of a matching, a result
that may be of independent interest. Finally, games with symmetries of two of the four
types are shown to possess a symmetric mixed equilibrium which can be computed in
polynomial time. We thus obtain a class of games where the pure equilibrium problem is
computationally harder than the mixed equilibrium problem, unless P=NP.
Chapter 6 is based on joint work with Felix Brandt, Markus Holzer, and Stefan Katzen-
beisser (Fischer et al., 2006, Brandt et al., 2008b).
 27

Quasi-Strict Equilibria Despite its ubiquity, Nash equilibrium has been criticized for
various reasons. Its existence relies on the possibility of randomizing over actions, which
in many cases is deemed unsuitable, impractical, or even infeasible. Furthermore, players
may be indifferent between actions they do and do not play in equilibrium, thus calling
into question the underlying notion of stability. In Chapters 7 and 8, we consider two
solution concepts that try to address these two shortcomings. Chapter 7 is devoted to an
investigation of the computational properties of quasi-strict equilibrium, an attractive
equilibrium refinement proposed by Harsanyi and recently shown to always exist in two-
player games. Quasi-strict equilibrium strengthens the equilibrium condition by requiring
that every best response be played with positive probability. We prove that deciding the
existence of a quasi-strict equilibrium in games with more than two players is NP-hard.
We further show that unlike Nash equilibria, quasi-strict equilibria in zero-sum games have
a unique support, and propose a linear program to compute a quasi-strict equilibrium in
these games. Finally, we prove that every symmetric multi-player game where each player
has two actions at his disposal possesses an efficiently computable quasi-strict equilibrium
which may itself be asymmetric.
Chapter 7 is based on joint work with Felix Brandt (Brandt and Fischer, 2008a).

Shapley’s Saddles In work dating back to the early 1950s, Shapley proposed ordinal set-
valued solution concepts for zero-sum games that he refers to as strict and weak saddles.
These concepts are intuitively appealing, they always exist, and are unique in several
important classes of games. In Chapter 8, we study of computational aspects of Shapley’s
saddles and provide polynomial-time algorithms for computing strict saddles in general
normal-form games, and weak saddles in a subclass of symmetric zero-sum games. On
the other hand, we show that several problems associated with weak saddles are NP-hard
already in two-player games, which provides rather strong evidence that they cannot be
computed efficiently.
Chapter 8 is based on joint work with Felix Brandt, Markus Brill, and Paul Harren-
stein (Brandt and Fischer, 2008b, Brandt et al., 2009a).
 Chapter 4

Ranking Games

The situations studied by the theory of games may involve different levels of antagonism.
On the one end of the spectrum are games of pure coordination, on the other those in
which the players’ interests are diametrically opposed. In this chapter, we introduce and
study a new class of competitive multi-player games whose outcomes are rankings of the
players, i.e., orderings representing how well they have done in the game relative to one
another. We assume players to weakly prefer a higher rank over a lower one and to be
indifferent as to the other players’ ranks. Indeed, this type of situation is very common
in the real world, with examples such as parlor games, sports competitions, patent races,
competitive resource allocation, social choice settings, and other strategic situations where
players are merely interested in performing optimal relative to their opponents rather than
in absolute measures. Formally, ranking games can be defined as normal-form games in
which the payoff functions represent the players’ von Neumann-Morgenstern preferences
over lotteries over rankings.
Apart from their practical relevance, ranking games promise to be interesting also from
a computational perspective. Two-player ranking games form a subclass of constant-sum
games, such that the computationally tractable maximin solution the solution concept
of choice. With more than two players, there no longer is any inclusion relationship
between ranking games and constant-sum games. The notion of a ranking, however, is
most natural and relevant in multi-player settings, a fact that is much less true for the
requirement that the sum of payoffs in all outcomes be zero or constant. Indeed, any game
can be transformed into a zero-sum game by introducing an additional player, with only
one action at his disposal, who absorbs the payoffs of the other players (von Neumann
and Morgenstern, 1947).
The maximin solution does not unequivocally extend to general n-player games, or to
n-player ranking games. Numerous alternative solution concepts have been proposed to
cope with this type of situation, some of which we have introduced in Chapter 2. None of
them, however, seems to be as compelling as maximin is for two-player zero-sum games,
and computational intractability serves as an additional threat to many of them. One
could nevertheless hope that the notion of competitiveness captured by ranking games

29
30 4 · Ranking Games

might cause tractability of solution concepts to carry over to the multi-player case.
We will see in Section 4.5 that unfortunately this does not seem to be the case, and
solving ranking games becomes considerably more complicated as soon as more than
two players are involved, just as games in which both contrary and common interests
prevail. In particular, we give NP-hardness and PPAD-hardness results, respectively, for
iterated weak dominance and (mixed) Nash equilibria when there are more than two
players, and an NP-hardness result for pure Nash equilibria in games with an unbounded
number of players. This dashes hope that multi-player ranking games can be solved
efficiently thanks to their structural restrictions. Remarkably, all hardness results hold
for arbitrary preferences over ranks, provided they meet the minimal requirements given
above. Accordingly, even very restricted subclasses of ranking games such as single-
winner games—in which players only care about winning—or single-loser games—in
which players merely wish not to be ranked last—are computationally hard to solve.
By contrast, maximin strategies as well as correlated equilibria (Aumann, 1974) are
known to be computationally easy via linear programming even in general games. Against
the potency of these concepts, however, other objections can be brought in. Playing a
maximin strategy is extremely defensive and a player may have to forfeit a considerable
amount of payoff in order to guarantee his security level. Correlation, on the other hand,
may not be feasible in all practical applications, and may fail to provide an improvement
of social welfare in restricted classes of games (Moulin and Vial, 1978). In Section 4.6
we thus come to consider the following comparative ratios in an effort to facilitate the
quantitative analysis of solution concepts in ranking games: the price of cautiousness,
i.e., the ratio between an agent’s minimum payoff in a Nash equilibrium and his security
level; the mediation value, i.e., the ratio between the social welfare obtainable in the
best correlated equilibrium vs. the best Nash equilibrium; and the enforcement value,
i.e., the ratio between the highest obtainable social welfare and that of the best correlated
equilibrium. Very much like solution concepts themselves, these values can be interpreted
descriptively and used as tool for comparing the properties of different games, or they can
be viewed from a normative perspective as providing an index of what can be gained or
lost by following a more daring rather than a more conservative course of action. Each
of the above values obviously equals 1 in the case of two-player ranking games, as these
form a subclass of constant-sum games. An interesting question thus concerns bounds on
these values for ranking games with more than two players.

4.1 An Introductory Example

To further illustrate the issues addressed in this chapter, recall the game of Figure 2.1,
where Alice, Bob, and Charlie are to designate one of them as the winner, by simulta-
neously and independently raising their hand or not. Alice wins if the number of hands
raised, including her own, is odd, whereas Bob is victorious if this number equals two.
Should nobody raise their hand, Charlie wins. It is obvious from the description, and can
4.2 · Related Work 31

c1 c2
b1 b2 b1 b2
a1 3 1 1 2
a2 1 2 2 1

Figure 4.1: The game of Figure 2.1, involving Alice (player 1), Bob (player 2), and Charlie
(player 3). Outcomes are now denoted by the index of the winning player. The dashed
square marks the only pure Nash equilibrium. Dotted rectangles mark an equilibrium in
which Alice and Charlie randomize uniformly over their respective actions.

also be seen from Figure 2.1, that this game is a ranking game, and in fact a single-winner
game. We show the game again in Figure 4.1, now denoting each outcome by the index
of the winning player.
What course of action would you recommend to Alice? There is a Nash equilibrium
in which Alice raises her hand, another one in which she does not raise her hand, and
still another one in which she randomizes uniformly between these two options. In the
only pure equilibrium of the game, Alice does not raise her hand. For the latter to occur,
Alice would have to believe that Bob will raise his hand and Charlie will not. This
assumption is unreasonably strong, however, in that no such beliefs can be derived from
the mere description of the game. Moreover, both Bob and Charlie could in the above
case deviate from their respective strategies to any other strategy without decreasing
their chances of winning. After all, they cannot do any worse than losing. On the other
hand, by playing her maximin strategy, Alice would guarantee a particular payoff, or
winning probability, no matter which actions her opponents choose. Alice’s security level
in this particular game is 1/2 and can be obtained by randomizing uniformly between
both actions. The same expected payoff is achieved in the mixed equilibrium where Alice
and Charlie randomize uniformly and Bob invariably raises his hand, indicated by the
dotted rectangles in Figure 4.1.

4.2 Related Work

In game theory, several proposals have been made for broader classes of games that main-
tain some of the desirable properties of two-player zero-sum games. The term strict com-
petitiveness is usually reserved for two-player zero-sum games. Friedman (1983) shows
that any convex combination of equilibria in such games is again an equilibrium. In an
attempt to characterize games which can always be solved, in the sense that players have
a single optimal strategy and the outcome is strictly determined, Aumann (1961) defines
almost strictly competitive games. A two-player game is almost strictly competitive if a
pair of strategies is an equilibrium point, i.e., no player can increase his payoff by unilater-
32 4 · Ranking Games

ally changing his strategy, if and only if it is a so-called twisted equilibrium point, i.e., no
player can decrease the payoff of his opponent. These games permit a set of optimal strate-
gies for each player and a unique value that is obtained whenever a pair of such strategies
is played. Moulin and Vial (1978) call a game strategically zero-sum if it is best-response
equivalent to a zero-sum game. In the case of two players, and only in this case, one
obtains exactly the class of games for which no completely mixed equilibrium can be im-
proved upon by a correlated equilibrium. A game is unilaterally competitive, as defined
by Kats and Thisse (1992), if any deviation by a player that (weakly) increases his own
payoff must (weakly) decrease the payoffs of all other players. Unilaterally competitive
games retain several interesting properties of two-player zero-sum games in the n-player
case: all equilibria yield the same payoffs, equilibrium strategies are interchangeable, and
the set of equilibria is convex provided that some mild conditions hold. It was later shown
by Wolf (1999) that pure Nash equilibria of n-player unilaterally competitive games are
always profiles of maximin strategies. When there are only two players, all of the above
classes contain those of constant-sum games and thus two-player ranking games. Neither
is contained in the other in the n-player case. The notion of competitiveness as embodied
in ranking games is remotely related to spitefulness (Morgan et al., 2003, Brandt et al.,
2007b), where agents aim at maximizing their payoff relative to the payoff of all other
agents.
Most work on comparative ratios in game theory has been inspired by the literature on
the price of anarchy (Koutsoupias and Papadimitriou, 1999, Roughgarden, 2005), i.e., the
ratio between the highest obtainable social welfare and that of the best Nash equilibrium.
Similar ratios for correlated equilibria were introduced by Ashlagi et al. (2005): the value
of mediation, i.e., the ratio between the social welfare obtainable in the best correlated
equilibrium and the best Nash equilibrium, and the enforcement value, i.e., the ratio
between the highest obtainable social welfare and that of the best correlated equilibrium.
It is known that the mediation value of strategically zero-sum games is 1 and that of almost
strictly competitive games is greater than 1, showing that correlation can be beneficial
even in games of strict antagonism (Raghavan, 2002). To our knowledge, Tennenholtz
(2002) was the first to conduct a quantitative comparison of Nash equilibrium payoffs
and security levels. This work is inspired by an intriguing example game due to Aumann
(1985), in which the only Nash equilibrium yields each player no more than his security
level although the equilibrium strategies are different from the maximin strategies. In
other words, the equilibrium strategies yield security level payoffs without guaranteeing
them.

4.3 The Model

Intuitively, a ranking game is a normal-form game whose outcomes are rankings of the
players. A ranking indicates how well each player has done relative to the other players
in the game. Formally, a ranking r = [r1 , r2 , . . . , rn ] is an ordering of the players in N
4.3 · The Model 33

in which player r1 is ranked first, player r2 ranked second, and so forth, with player rn
ranked last. Obviously, this limits the number of possible outcomes to n! irrespective of
the number of actions the players have at their disposal. The set of rankings over a set N
of players we denote by RN .
We assume that all players weakly prefer higher ranks over lower ranks, and strictly
prefer being ranked first to being ranked last. Furthermore, each player is assumed to be
indifferent as to the ranks of the other players. Even so, a player may prefer to be ranked
second for certain to having a fifty-fifty chance of being ranked first or being ranked
third, whereas other players may judge quite differently. Accordingly, we have a rank
payoff function pi : RN → R represent player i’s von Neumann-Morgenstern preferences
over lotteries over RN . For technical convenience, we normalize the payoffs to the unit
interval [0, 1]. Formally, a rank payoff function pi over RN satisfies the following three
conditions for all rankings r, r 0 ∈ RN :

(i) pi (r) > pi (r 0 ), if rk = rm

0 = i and k 6 m,

(ii) pi (r) = 1, if i = r1 , and

(iii) pi (r) = 0, if i = rn .

It will be convenient to make the relationship between the outcomes of a game and
payoffs obtained in these outcomes explicit. To this end, let a game form be a tuple
(N, (Ai )i∈N , Ω, g) where N is a set of players, Ai is a set of actions available to player
i ∈ N, Ω is a set of outcomes, and g : "i∈N Ai → Ω is an outcome function mapping each
action profile to an outcome in Ω. A game form then is a ranking game form if the set
of outcomes is given by the set of rankings of the players, i.e., if Ω = RN . We are now in
a position to formally define the concept of a ranking game.

Definition 4.1 (ranking game). A normal-form game Γ = (N, (Ai )i∈N , (pi )i∈N ) is called
ranking game if there exists a ranking game form (N, (Ai )i∈N , RN , g) and rank payoff
functions pi0 over RN such that for all a ∈ AN and all i ∈ N, pi (aN ) = pi0 (g(aN )).

Condition (i) above implies that a player’s payoff for a ranking r only depends on
the rank assigned to him in r. Accordingly, for 1 6 k 6 n, we have pk i denote the
unique payoff player i obtains in any ranking r in which he is ranked kth. The rank
payoff function of player i can then conveniently and compactly be represented by his
rank payoff vector ~pi = (p1i , p2i , . . . , pn
i ).
In a binary ranking game, a player is completely satisfied up to a certain rank, and
not satisfied at all for any lower rank. The expected payoff of a player given a strategy
profile can then be taken as his chances of being satisfied. In this case, the use of expected
payoffs, and thus randomized strategies, is justified without relying on the von Neumann-
Morgenstern axioms (see also Aumann, 1987). An interesting subclass of binary ranking
games are so-called single-winner games, in which all players are only interested in being
ranked first. Formally, a single-winner game is a ranking game in which ~pi = (1, 0, . . . , 0)
34 4 · Ranking Games

c1 c2
b1 b2 b1 b2
a1 [1, 3, 2] [2, 1, 3] [3, 2, 1] [3, 1, 2]
a2 [2, 3, 1] [3, 2, 1] [2, 1, 3] [1, 3, 2]

Figure 4.2: A 2 × 2 × 2 ranking game form. One player chooses rows, another columns,
and a third matrices. Each combination of actions results in a ranking. For example,
action profile (a2 , b2 , c2 ) leads to the row player 1 being ranked first, the matrix player 3
second and the column player 2 third.

c1 c2
b1 b2 b1 b2
a1 (1, 0, 1) ( 21 , 1, 0) (0, 0, 1) ( 12 , 0, 1)
a2 (0, 1, 1) (0, 0, 1) ( 12 , 1, 0) (1, 0, 1)

Figure 4.3: A ranking game associated with the ranking game form of Figure 4.2. The rank
payoff for the three players are given by ~p1 = (1, 12 , 0), ~p2 = (1, 0, 0) and ~p3 = (1, 1, 0).

for all i ∈ N. When considering mixed strategies, the expected payoff in a single-winner
ranking game equals the probability of winning. Analogous to single-winner games, we can
define single-loser games as ranking games in which the players’ only concern is not to be
ranked last, as for instance in a round of musical chairs. Formally, single-loser games are
ranking games where ~pi = (1, . . . , 1, 0) for each player i. For an example illustrating the
definitions of a ranking game form and a ranking game the reader is referred to Figures 4.2
and 4.3, respectively.
At this point, a remark as to the relationship between ranking games and n-player
constant-sum games is in order. By virtue of conditions (ii) and (iii), two-player ranking
games constitute a subclass of constant-sum games. If more than two players are involved,
however, any such relation with n-person constant-sum games ceases to hold. A strategic
game can be converted to a zero-sum game via positive affine transformations only if
all outcomes of the game lie on an (n − 1)-dimensional hyperplane in the n-dimensional
outcome space. Clearly, there are ranking games, with non-identical rank payoff vectors
and more than two players, for which this is not the case, like a three-player ranking
game with rank payoff vectors ~p1 = ~p2 = (1, 0, 0) and ~p3 = (1, 1, 0) that has among its
outcomes the rankings [1, 2, 3], [2, 1, 3], [3, 1, 2], and [1, 3, 2]. As a consequence, ranking
games are no subclass of constant-sum games. It is readily appreciated that the opposite
inclusion does not hold either.
4.4 · Games With Non-Pure Equilibria 35

4.4 Games With Non-Pure Equilibria

We noticed earlier that in the only pure equilibrium of the game in Figure 4.1, two of
the three players have no reason whatsoever to stick to their equilibrium strategies. This
in fact indicates an inherent weakness of pure Nash equilibrium as a solution concept
for ranking games. In any outcome of a ranking game, some player is ranked last and
receives his minimum payoff. This means that in any pure Nash equilibrium ,some player
must be indifferent between all of his actions: no matter what he does, he will remain
a loser. Obviously, the stability of a solution like that is highly questionable, if it is to
be considered a viable solution at all. This is particularly true for single-winner games,
where in a pure equilibrium all players but the winner are indifferent over which action
to play.
On the other hand, it is very well possible that all the actions in the support of a mixed
equilibrium yield a strictly higher expected payoff than any action not in the support,
mitigating the phenomenon mentioned above. Equilibria satisfying this condition are
called quasi-strict, and will be treated in more detail in Chapter 7.
For now we concentrate on non-pure equilibria, i.e., equilibria in which at least one
player randomizes. We conjecture that every single-winner game possesses at least one
such equilibrium, and prove this claim for three subclasses.

Theorem 4.2. The following classes of ranking games always possess at least one
non-pure equilibrium:

(i) two-player ranking games,

(ii) three-player single-winner games where each player has two actions, and

(iii) n-player single-winner games where the security level of at least two players is
positive.

Proof. Statement (i) follows directly from the fact that every two-player game has a quasi-
strict equilibrium (Norde, 1999), and the above observation that quasi-strict equilibria of
ranking games are never pure. Here we give a simple alternative proof. Assume for
contradiction that there is a two-player ranking game that only possesses pure equilibria
and consider, without loss of generality, a pure equilibrium s∗N in which player 1 wins.
Since player 2 must be incapable of increasing his payoff by deviating from s∗N , player 1
has to win no matter which action the second player chooses. As a consequence, the
strategies in s∗N remain in equilibrium even if player 2’s strategy is replaced with an
arbitrary randomization among his actions.
As for (ii), consider a three-player single winner game with actions A1 = {a1 , a2 }, A2 =
{b1 , b2 }, and A3 = {c1 , c2 }. Assume for contradiction that there are only pure equilibria in
the game and consider, without loss of generality, a pure equilibrium s∗N = (a1 , b1 , c1 ) in
which player 1 wins. In the following, we say that a pure equilibrium is semi-strict if at
least one player strictly prefers his equilibrium action over all his other actions, given that
36 4 · Ranking Games

the other players play their equilibrium actions. In single-winner games, this player has to
be the winner in the pure equilibrium. We first show that if s∗N is semi-strict, i.e., if player 1
does not win in action profile (a2 , b1 , c1 ), then there must exist a non-pure equilibrium.
For this, consider the strategy profile s1N = (a1 , s12 , c1 ), where s12 is the uniform mixture
of player 2’s actions b1 and b2 , along with the strategy profile s2N = (a1 , b1 , s23 ), where s23
is the uniform mixture of actions c1 and c2 of player 3. Since player 1 does not win
in (a2 , b1 , c1 ), he has no incentive to deviate from either s1N or s2N , even if he wins in
(a2 , b2 , c1 ) and (a2 , b1 , c2 ). Consequently, player 3 must win in (a1 , b2 , c2 ) in order
for s1N not to be an equilibrium. Analogously, for s2N not to be an equilibrium, player 2
has to win in the same action profile (a1 , b2 , c2 ), contradicting the assumption that the
game is a single-winner game. The existence of a semi-strict pure equilibrium thus implies
that of a non-pure equilibrium. Now assume that s∗N is not semi-strict. When any of the
action profiles in B = {(a2 , b1 , c1 ), (a1 , b2 , c1 ), (a1 , b1 , c2 )} is a pure equilibrium, this also
yields a non-pure equilibrium because two pure equilibria that only differ by the action
of a single player can be combined into infinitely many mixed equilibria. For B not to
possess any pure equilibria, there must be (exactly) one player for every profile in B who
deviates to a profile in C = {(a2 , b2 , c1 ), (a2 , b1 , c2 ), (a1 , b2 , c2 )}, because the game is
a single-winner game and because s∗N is not semi-strict. Moreover, either player 1 or
player 2 wins in (a2 , b2 , c1 ), player 2 or player 3 in (a1 , b2 , c2 ), and player 1 or player 3
in (a2 , b1 , c2 ). This implies two facts. First, the action profile s3N = (a2 , b2 , c2 ) is a pure
equilibrium because no player will deviate from s3N to any profile in C. Second, the player
who wins in s3N strictly prefers the equilibrium outcome over the corresponding action
profile in C, implying that s3N is semi-strict. The above observation that every semi-strict
equilibrium also yields a non-pure equilibrium completes the proof.
As for (iii), recall that the payoff a player obtains in equilibrium must be at least his
security level. A positive security level for player i thus rules out all equilibria in which
player i receives payoff zero, in particular all pure equilibria in which he does not win. If
there are two players with positive security levels, both of them have to win with positive
probability in any equilibrium of the game. In single-winner games, this can only be the
case in a non-pure equilibrium.

We conjecture that this existence result in fact applies to the entire class of single-
winner games. To see this it does not extend to general ranking games, consider a 4-player
game in which the first three players have two actions. Payoffs for the case when player 4
plays his first action d1 are shown in Figure 4.4. We can in fact restrict our attention to
this case by assigning payoffs (1, 0, 0, 0) to any action profile where player 4 plays one of his
potential other actions, thereby making them strictly dominated. The resulting game is
then similar to a 3-player game by van Damme (1983) that does not have any quasi-strict
equilibria. It is furthermore straightforward to show that the game is a ranking game by
virtue of rank payoff vectors r1 = r3 = (1, 0, 0, 0), r2 = (1, 12 , 0, 0), and r4 = (1, 1, 1, 0), and
that the pure equilibria (a1 , a2 , a3 ) and (b1 , b2 , b3 ) are the only equilibria of this game.
4.5 · Solving Ranking Games 37

c1 c2
b1 b2 b1 b2
a1 (0, 0, 0, 1) (1, 0, 0, 1) (0, 12 , 0, 1) (0, 0, 1, 1)
d1
a2 (0, 0, 1, 1) (0, 1, 0, 1) (1, 0, 0, 1) (0, 0, 0, 1)

Figure 4.4: Four-player ranking game in which all equilibria are pure

It remains open whether there exists a 3-player ranking game with only pure equilibria.

4.5 Solving Ranking Games

The question we will try to answer next is whether the rather restricted payoff structure of
ranking games makes it possible to compute instances of common solution concepts more
efficiently than in general games. For this reason, we focus on solution concepts that are
known to be intractable for general games, namely (mixed) Nash equilibria (Chen and
Deng, 2006, Daskalakis et al., 2009a), iterated weak dominance (Conitzer and Sandholm,
2005a), and pure Nash equilibria in games with many players and polynomial-time com-
putable payoff functions (Schoenebeck and Vadhan, 2006). Graphical games, in which pure
Nash equilibria are also known to be intractable (Gottlob et al., 2005), are of very limited
use for representing ranking games. If two players are not connected by the neighborhood
relation, either directly or via a common player in their neighborhood, then their payoffs
are completely independent from each other. For a single-winner game with the reason-
able restriction that every player wins in at least one outcome, this implies that there
must be one designated player who alone decides which player wins the game. Similar
properties hold for arbitrary ranking games. For iterated strict dominance (Conitzer and
Sandholm, 2005a) or correlated equilibria (Papadimitriou, 2005) efficient algorithms exist
even for general games. There thus is no further need to consider these solution concepts
here. When in the following we refer to the hardness of a game we mean NP-hardness or
PPAD-hardness of solving the game using a particular solution concept.

4.5.1 Mixed Nash Equilibria

Let us first consider Nash equilibria of games with a bounded number of players. Two-
player ranking games only allow outcomes (1, 0) and (0, 1) and thus constitute a subclass
of constant-sum games. Nash equilibria of constant-sum games can be found by linear
programming (Dantzig, 1951), for which there is a polynomial time algorithm (Khachiyan,
1979).
To prove hardness for the case with more than two players, it suffices to show that
three-player ranking games are at least as hard to solve as general two-player games. To
38 4 · Ranking Games

appreciate this, observe that any n-player ranking game can be turned into an (n + 1)-
player ranking game by adding a player who has only one action at his disposal and
who is invariably ranked last, keeping relative rankings of the other players intact. Nash
equilibria of the (n + 1)-player game then naturally correspond to Nash equilibria of the
n-player game. A key concept in our proof is that of a Nash homomorphism, a notion
introduced by Abbott et al. (2005). We generalize their definition to games with more
than two players.

Definition 4.3 (Nash homomorphism). A Nash homomorphism is a mapping h from

a set of games into a set of games, such that there exists a polynomial-time computable
function f that, when given a game Γ and an equilibrium s∗N of h(Γ ), returns an equilibrium
f(s∗N ) of Γ .

Obviously, the composition of two Nash homomorphisms is again a Nash homomor-

phism. Furthermore, any sequence of polynomially many Nash homomorphisms that maps
some class of games to another class of games provides us with a polynomial-time reduction
from the problem of finding Nash equilibria in the former class to finding Nash equilibria
in the latter. Any efficient, i.e., polynomial-time, algorithm for the latter directly leads
to an efficient algorithm for the former. On the other hand, hardness of the latter implies
hardness of the former.
A very simple example of a Nash homomorphism is the one that scales the payoff
of each player by means of a positive affine transformation. It is well-known that Nash
equilibria are invariant under this kind of mapping, and f can be taken to be the identity.
We will now combine this Nash homomorphism with a more sophisticated function, which
maps payoff profiles of a two-player binary game to corresponding three-player subgames
with two actions for each player, and obtain Nash homomorphisms from two-player games
to three-player ranking games for all possible rank payoff profiles.

Lemma 4.4. For every rank payoff profile, there exists a Nash homomorphism from
the set of two-player games to the set of three-player ranking games.

Proof. Abbott et al. (2005) have shown that there is a Nash homomorphism from two-
player games to binary two-player games. Since a composition of Nash homomorphisms
is again a Nash homomorphism, we only need to provide a homomorphism from binary
two-player games to three-player ranking games. Furthermore, outcome (1, 1) is Pareto-
dominant and therefore constitutes a pure Nash equilibrium in any binary game, since no
player can benefit from deviating. Instances containing such an outcome are easy to solve
and need not be considered in our mapping.
In the following, we denote by (1, p2i , 0) the rank payoff vector of player i, and by
[i, j, k] the outcome where player i is ranked first, j is ranked second, and k is ranked last.
First of all, consider ranking games where p2i < 1 for some player i ∈ N, i.e., the class of
all ranking games except single-loser games.
4.5 · Solving Ranking Games 39

Outcome Constant-sum Ranking subgame

outcome

(1, 0, 1) (0, 1, 1) (0, 1, 1) (1, 0, 1)

(0, 0) 7−→ ( 12 , 12 , 1) 7−→
(0, 1, 1) (1, 0, 1) (1, 0, 1) (0, 1, 1)

(1, 0, 1) (1, 1, 0) (1, 1, 0) (1, 0, 1)

(1, 0) 7−→ (1, 12 , 21 ) 7−→
(1, 1, 0) (1, 0, 1) (1, 0, 1) (1, 1, 0)

(0, 1, 1) (1, 1, 0) (1, 1, 0) (0, 1, 1)

(0, 1) 7−→ ( 12 , 1, 21 ) 7−→
(1, 1, 0) (0, 1, 1) (0, 1, 1) (1, 1, 0)

Figure 4.5: Mapping from binary two-player games to three-player single-loser games

Without loss of generality let i = 1. Then, a Nash homomorphism from binary two-
player games to the aforementioned class of games can be obtained by first transforming
the payoffs according to

(x1 , x2 ) 7−→ (1 − p21 )x1 + p21 , x2

and then adding a third player who only has a single action and whose payoff is chosen
such that the resulting game is a ranking game (but is otherwise irrelevant). We obtain
the following mapping, which is obviously a Nash homomorphism:
(0, 0) −7 → (p21 , 0) −
7 → [3, 1, 2]
(1, 0) 7−→ (1, 0) − 7 → [1, 3, 2]
2
(0, 1) 7−→ (p1 , 1) − 7 → [2, 1, 3].
Interestingly, three-player single-loser games with only one action for some player
i ∈ N are easy to solve because either
 there is an outcome in which i is ranked last and the other two players both receive
their maximum payoff of 1, i.e., a Pareto-dominant outcome, or

 i is not ranked last in any outcome, such that the payoffs of the other two players
always sum up to 1 and the game is equivalent to a two-player zero-sum game.
As soon as the third player is able to choose between two different actions, however, binary
games can again be mapped to single-loser games. For this, consider the mapping from
binary two-player games to three-player single-loser games shown in Figure 4.5. As a first
step, binary two-player games are mapped to three-player constant-sum games according
to
1 1 1

(x1 , x2 ) 7−→ (x1 + 1), (x2 + 1), 1 − (x1 + x2 ) .
2 2 2
40 4 · Ranking Games

The first two players and their respective sets of actions are the same as in the original
game, the third player only has one action c. It is again obvious that this constitutes a
Nash homomorphism. Next, outcomes of the three-player constant-sum game are replaced
by three-player single-loser subgames. Let Γ be a binary game, and denote by Γ 0 and Γ 00
the three-player constant-sum game and the three-player single-loser game, respectively,
obtained by applying the two steps of the mapping in Figure 4.5 to Γ . We further write
pi0 , and pi00 for the payoff function of player i in Γ 0 and Γ 00 , and a1i and a2i for the two
actions of player i in Γ 00 corresponding to an action ai in Γ 0 .
The second part of the mapping in Figure 4.5 is chosen such that for all strategy
profiles sN , all players i and all actions ai ∈ Ai in Γ 0 we have

1 00 1 1
pi (ai , s−i ) + pi00 (a2i , s−i ) = pi0 (ai , f(sN )−i ), (4.1)
2 2

where for each strategy profile sN of Γ 00 , f(sN ) is the strategy profile in Γ 0 such that for
each player i ∈ {1, 2, 3} and each action ai ∈ Ai ,

f(sN )(ai ) = si (a1i ) + si (a2i ).

An important property of this construction is that each player can guarantee his payoff
in Γ 00 , for any strategy profile of the other players, to be at least as high as his payoff
under the corresponding strategy profile in Γ 0 , by distributing the weight on ai uniformly
on a1i and a2i .
Let s∗N be a Nash equilibrium in Γ 00 . We first prove that for every player i ∈ {1, 2, 3}
and each action ai of player i in Γ 0 ,

s∗i (a1i )pi00 (a1i , s∗−i ) + s∗i (a2i )pi00 (a2i , s∗−i ) = (f(s∗N )(ai ))pi0 (ai , f(s∗N )−i ). (4.2)

Recall that we write sN (ai ) for the probability of action ai in strategy profile sN , so
f(s∗N )(ai ) is the probability with which ai is played in strategy profile f(s∗N ) of Γ 0 . The
above equation thus states that the expected joint payoff from a1i and a2i in equilibrium s∗N
equals that from ai under the corresponding strategy profile f(s∗N ) of Γ 0 . To see this, first
assume for contradiction that for some player i and some action ai ∈ Ai ,

s∗i (a1i )pi00 (a1i , s∗−i ) + s∗i (a2i )pi00 (a2i , s∗−i ) < (f(s∗N )(ai ))pi0 (ai , f(s∗N )−i ),

i.e., that the expected joint payoff from a1i and a2i in Γ 00 is strictly smaller than the
expected payoff from ai in Γ 0 . Define si to be the strategy of player i in Γ 00 such
that si (a1i ) = si (a2i ) = 12 (s∗i (a1i ) + s∗i (a2i )) and si (ai0 ) = s∗i (ai0 ) for all actions ai0 ∈ Ai
4.5 · Solving Ranking Games 41

distinct from a1i and a2i . It then holds that

s∗i (a1i )pi00 (a1i , s∗−i ) + s∗i (a2i )pi00 (a2i , s∗−i )
< (f(s∗N )(ai ))pi0 (ai , f(s∗N )−i )
= (s∗i (a1i ) + s∗i (a2i ))pi0 (ai , f(s∗N )−i )
1 1
= (s∗i (a1i ) + s∗i (a2i ))( pi00 (a1i , s∗−i ) + pi00 (a2i , s∗−i ))
2 2
1 ∗ 1 1
= (si (ai ) + s∗i (a2i ))pi00 (a1i , s∗−i ) + (s∗i (a1i ) + s∗i (a2i ))pi00 (a2i , s∗−i )
2 2
= si (a1i )pi00 (a1i , s∗−i ) + si (a2i )pi00 (a2i , s∗−i ).

The second and last step respectively follow from the definition of f and si . The third
step follows from (4.1). We conclude that player i obtains a higher payoff by playing si
instead of s∗i , contradicting the assumption that s∗N is a Nash equilibrium. In particular
we have shown that for all i ∈ N and every ai ∈ Ai ,

s∗i (a1i )pi00 (a1i , s∗−i ) + s∗i (a2i )pi00 (a2i , s∗−i ) > (f(s∗N )(ai ))pi0 (ai , f(s∗N )−i ). (4.3)

Now assume, again for contradiction, that for some player i and some action ai ∈ Ai ,

s∗i (a1i )pi00 (a1i , s∗−i ) + s∗i (a2i )pi00 (a2i , s∗−i ) > (f(s∗N )(ai ))pi0 (ai , f(s∗N )−i ),

i.e., that the expected joint payoff to i from a1i and a2i in Γ 00 is strictly greater under s∗N
than the expected payoff from ai in Γ 0 . It follows from (4.3) that the expected payoff
player i receives from any action under f(s∗N ) cannot be greater than the expected joint
payoff from the corresponding pair of actions under s∗N , and thus pi00 (s∗N ) > pi0 (f(s∗N )).
Since Γ 0 and Γ 00 are both constant-sum games, there exists some player j 6= i who receives
strictly less payoff under s∗N in Γ 00 than under f(s∗N ) in Γ 0 . In particular, there has to be
an action aj ∈ Aj such that

s∗j (a1j )pj00 (a1j , s∗−j ) + s∗j (a2j )pj00 (a2j , s∗−j ) < (f(s∗N )(aj ))pj0 (aj , f(s∗N )−j ),

contradicting (4.3).
We are now ready to prove that the mapping in Figure 4.5 is indeed a Nash homomor-
phism. To this end, let s∗N be a Nash equilibrium of Γ 00 , and assume for a contradiction
that f(s∗N ) is not a Nash equilibrium of Γ 0 . Then there has to be a player i and some action
ai ∈ Ai such that pi0 (ai , f(s∗N )−i ) > pi0 (f(s∗N )). Define si to be the strategy of player i
in Γ 00 such that si (a1i ) = si (a2i ) = 12 . Then, by (4.1), pi00 (si , s−i ) = pi0 (ai , f(s∗N )−i ). It
further follows from (4.2) that for every player j, pj00 (s∗N ) = pj0 (f(s∗N )), and for j = i in
particular. Thus,

pi00 (s∗N ) = pi0 (f(s∗N )) < pi0 (ai , f(s∗N )−i ) = pi00 (si0 , s∗−i ),

contradicting the assumption that s∗N is a Nash equilibrium in Γ 00 .

42 4 · Ranking Games

The ground has now been cleared for the main result of this section.

Theorem 4.5. Computing a Nash equilibrium of a ranking game with more than two
players is PPAD-hard for any rank payoff profile. If there are only two players,
equilibria can be found in polynomial time.

Proof. According to Lemma 4.4, ranking games with more than two players are at least
as hard to solve as general two-player games. We already know that solving general games
is PPAD-hard in the two-player case (Chen and Deng, 2006).
Two-player ranking games, on the other hand, form a subclass of two-player constant-
sum games, in which Nash equilibria can be found efficiently via linear programming.

4.5.2 Iterated Weak Dominance

We now turn to iterated weak dominance. If there are only two players, the problem of
deciding whether a ranking game can be solved via iterated weak dominance is tractable.

Theorem 4.6. For two-player ranking games, iterated weak dominance solvability
can be decided in polynomial time.

Proof. First we recall that if an action in a binary game is weakly dominated by a mixed
strategy, it is also dominated by a pure strategy (Conitzer and Sandholm, 2005a). Ac-
cordingly, we only have to consider dominance by pure strategies. Now consider a path
of iterated weak dominance that ends in a single action profile (a∗1 , a∗2 ). Without loss of
generality we may assume that player 1, i.e., the row player, is the winner in this profile.
This implies that player 1 wins in (a∗1 , a2 ) for any a2 ∈ A2 , i.e., in the entire row. For
contradiction, assume the opposite and consider the particular action a12 such that player 2
wins in (a∗1 , a12 ) and a12 is eliminated last on the path that solves the game. It is easy
to see that such an elimination could only have taken place via another action a22 such
that player 2 also wins in (a∗1 , a22 ), contradicting the assumption that a12 is eliminated
last. We now claim that a ranking game with two players is solvable by iterated weak
dominance if and only if there exists a unique action a∗1 of player 1 by which he always
wins, and an action a∗2 of player 2 by which he wins for a strictly maximal set of actions
of player 1. More precisely, the latter property means that there exists a set of actions
of player 1 against which player 2 always wins when playing a∗2 and loses in at least one
case for every other action he might play. This is illustrated in Figure 4.6, and can be
verified efficiently by ordering the aforementioned sets of actions of player 1 according to
strict inclusion. If the ordering does not have a maximal element, the game cannot be
solved by means of iterated weak dominance. If it does, we can use a∗1 to eliminate all
actions a1 ∈ A1 such that player 2 does not win in (a1 , a∗2 ), whereupon a∗2 can eliminate
all other actions of player 2, until finally a∗1 eliminates player 1’s remaining actions and
solves the game.1
1
Since two-player ranking games are a subclass of constant-sum games, weak dominance and nice weak
4.5 · Solving Ranking Games 43

a∗2

2 1
.. ..
. .
2 1
1 2
..
.. .
.
2
a∗1 1 ··· 1

Figure 4.6: Iterated weak dominance solvability in two-player ranking games

As soon as there are three players, iterated weak dominance solvability becomes hard
to decide.

Theorem 4.7. For ranking games with more than two players, and for any rank
payoff profile, deciding iterated weak dominance solvability is NP-complete.

Proof. Membership in NP is immediate. We can simply guess a sequence of eliminations

and then verify in polynomial time that this sequence is valid and solves the game.
For hardness, we first reduce eliminability in binary two-player games, which asks
whether there exits a sequence of eliminations that contains a given action and has recently
been shown to be NP-hard (Conitzer and Sandholm, 2005a), to the same problem in
ranking games. A game Γ of the former class is mapped to a ranking game Γ 0 as follows:

 Γ 0 features the two players of Γ , denoted by 1 and 2, and an additional player 3.

 Players 1 and 2 have the same actions as in Γ , player 3 has two actions c1 and c2 .

 Payoffs of Γ are mapped to rankings of Γ 0 according to

(0, 0) 7−→ [3, 2, 1] [3, 1, 2] (1, 0) 7−→ [1, 2, 3] [3, 1, 2]

(0, 1) 7−→ [3, 2, 1] [2, 1, 3] (1, 1) 7−→ [1, 2, 3] [2, 1, 3] .

In the following, we write p and p 0 for the payoff functions of Γ and Γ 0 , respectively.

First observe that we can restrict our attention to dominance by pure strategies. This
property holds for binary games by Lemma 1 of Conitzer and Sandholm (2005a), and thus
dominance (Marx and Swinkels, 1997) coincide, making iterated weak dominance order independent up
to payoff-equivalent action profiles. This fact is mirrored by Figure 4.6, since there cannot be a row of 1s
and a column of 2s in the same matrix.
44 4 · Ranking Games

c1 c2
b1 ··· bk b1 ··· bk
a1 [ · , 2, · ] · · · [ · , 2, · ] [ · , 1, · ] · · · [ · , 1, · ]
.. .. .. .. .. .. ..
. . . . . . .
am [ · , 2, · ] · · · [ · , 2, · ] [ · , 1, · ] · · · [ · , 1, · ]
am+1 [3, 2, 1] · · · [3, 2, 1] [2, 1, 3] · · · [2, 1, 3]

Figure 4.7: Three-player ranking game Γ 0 used in the proof of Theorem 4.7

also for actions of player 3, who receives a payoff of either 0 or 1 in any outcome. For
players 1 and 2 we can essentially apply the same argument, because each of them can
obtain only two different payoffs for any fixed action profile of the remaining two players.
We now claim that irrespective of the rank payoffs pi = (1, p2i , 0), and for any subsets
of the actions of players 1 and 2, a particular action of these players is dominated in the
restriction of Γ 0 to these subsets if and only if the corresponding action is dominated in
the restriction of Γ to the same subsets. To see this, observe that if player 3 plays c1 ,
then for any action profile (a1 , a2 ) ∈ A1 × A2 , player 1 receives the same payoff he would
receive for the corresponding action profile in Γ , i.e., p10 (a1 , a2 , c1 ) = p1 (a1 , a2 ), whereas
player 2 receives a payoff of p22 . If on the other hand player 3 plays c2 , then player 1
obtains a payoff of p21 , and the payoff of player 2 for any action profile (a1 , a2 ) ∈ A1 × A2
is the same as that for the corresponding profile in Γ , i.e., p20 (a1 , a2 , c2 ) = p2 (a1 , a2 ).
Moreover, the implication from left to right still holds if one of the actions of player 3 is
removed, because this leaves one of players 1 and 2 indifferent between all of his remaining
actions but does not have any effect on dominance between actions of the other player.
We have thus established a direct correspondence between sequences of eliminations in Γ
and Γ 0 , which in turn implies NP-hardness of deciding whether a particular action of a
ranking game with at least three players can be eliminated.
It also follows from the above that Γ can be solved by iterated weak dominance if Γ 0
can. The implication in the other direction does not hold, however, because it may not
always be possible to eliminate an action of player 3. To this end, assume without loss of
generality that some player of Γ 0 has at least two actions, and that this player is player 1.
Otherwise both Γ and Γ 0 are trivially solvable. We augment Γ 0 by adding to the action set
A1 = {a1 , a2 , . . . , am } of player 1 an additional action am+1 such that for every action bj
of player 2, g(am+1 , bj , c1 ) = [3, 2, 1] and g(am+1 , bj , c2 ) = [2, 1, 3]. The structure of the
resulting game is shown in Figure 4.7.
It is easily verified that the above arguments about Γ 0 still apply, because player 1
never receives a higher payoff from am+1 than from any other action, and player 2 is
indifferent between all of his actions when player 1 plays am+1 . Now assume that Γ
4.5 · Solving Ranking Games 45

can be solved. Without loss of generality we may assume that (a1 , b1 ) is the remaining
action profile. Clearly, for Γ to be solvable, player 1 must be ranked first in some outcome
of Γ 0 , and it must hold that p1 (a1 , b1 ) = 1 or p2 (a1 , b1 ) = 1. We distinguish two cases.
If p1 (a1 , b1 ) = p2 (a1 , b1 ) = 1, then Γ 0 can be solved by performing the eliminations
that lead to the solution of Γ , followed by the elimination of c2 and am+1 . Otherwise
we can start by eliminating am+1 , which is dominated by the action for which player 1
is sometimes ranked first, and proceed with the eliminations that solve Γ . In the two
action profiles that then remain, player 3 is ranked first and last, respectively, and he can
eliminate one of his actions to solve Γ 0 .

4.5.3 Pure Nash Equilibria in Games with Many Players

When a game is given explicitly, pure Nash equilibria can be found efficiently by simply
checking for every action profile whether it satisfies the equilibrium condition. As the
number of players increases, however, the number of profiles to check grows exponentially,
just as the explicit representation of the game. An interesting question is whether pure
equilibria can be computed efficiently given a representation of a game that only uses
space polynomial in n.
We proceed to show that for ranking games this is most likely not the case. More
precisely, we show NP-completeness of deciding whether there is a pure Nash equilibrium
in ranking games with efficiently computable outcome functions, which is one of the
most general representations of multi-player games one might think of. It should be noted
that in contrast to Theorems 4.5 and 4.7, we now fix the number of actions and let the
number of players grow.

Theorem 4.8. For ranking games with an unbounded number of players and a
polynomial-time computable outcome function, and for any rank payoff profile, de-
ciding the existence of a pure Nash equilibrium is NP-complete. Hardness holds even
for games with two actions.

Proof. Since we can check in polynomial time whether a particular player strictly prefers
one rank over another, membership in NP is immediate. We can guess an action profile aN
and verify in polynomial time whether aN is a Nash equilibrium. For the latter, we check
for each player i ∈ N and for each action ai0 ∈ Ai whether pi (a−i , ai0 ) 6 pi (aN ).
For hardness, recall that circuit satisfiability (CSAT), i.e., deciding whether for a given
Boolean circuit C there exists an input such that the output is true, is NP-complete (e.g.,
Papadimitriou, 1994a). Given a particular Boolean circuit C with m inputs, we construct
a game Γ with players N = {1, 2, . . . , m} ∪ {x, y}, actions Ai = {0, 1} for all i ∈ N, and
payoff functions as follows:

 The outcome function of Γ is computed by a Boolean circuit that takes m + 2 bits

of input (a1 , a2 , . . . , am , ax , ay ), corresponding to the actions of the players in N,
46 4 · Ranking Games

and computes two bits of output o = (o1 , o2 ), given by o1 = C(a1 , a2 , . . . , am ) and

o2 = (o1 ∨ (ax XOR ay )).

 The possible outputs of the circuit are identified with permutations of the players
in N such that the permutation π00 corresponding to o = (0, 0) and the permutation
π11 corresponding to o = (1, 1) both rank x first and y last, the permutation π01
corresponding to o = (0, 1) ranks y first and x last, and all other players are ranked
in the same order in all three permutations. It should be noted that no matter how
permutations are actually encoded as strings of binary values, the encoding of the
above permutations can always be computed using a polynomial number of gates.

We now claim that, for arbitrary rank payoffs, Γ has a pure Nash equilibrium if and
only if C has a satisfying assignment. This can be seen as follows:

 If (a1 , a2 , . . . , am ) is a satisfying assignment of C, only a player in {1, 2, . . . , m} could

possibly change the outcome of the game by changing his action. However, these
players are ranked in the same order in all the possible outcomes, so none of them can
get a higher payoff by doing so. Thus, every action profile (a1 , a2 , . . . , am , ax , ay )
such that (a1 , a2 , . . . , am ) is a satisfying assignment of C is a Nash equilibrium.

 If in turn (a1 , a2 , . . . , am ) is not a satisfying assignment of C, both x and y are

able to switch between outcomes π00 and π01 by changing their own action. Since
further every player strictly prefers being ranked first over being ranked last, x
strictly prefers outcome π00 over π01 , while y strictly prefers π01 over π00 . Thus,
(a1 , a2 , . . . , am , ax , ay ) cannot be a Nash equilibrium, since either x or y could play
a different action to get a higher payoff.

4.6 Comparative Ratios

Despite its conceptual elegance and simplicity, Nash equilibrium has been criticized on
various grounds (see, e.g., Luce and Raiffa, 1957, for a discussion). It might require
randomization on the part of the players, and it is unclear how this randomization should
be carried out reliably. In the common case when there are multiple equilibria, it is unclear
which one should be selected. Coalitions might benefit from a joint deviation. There might
not exist an efficient algorithm for finding Nash equilibria. Finally, players may be utterly
indifferent among equilibrium and non-equilibrium strategies, which we saw is the case
particularly often in ranking games. We therefore devote the remainder of this chapter to
a comparison between Nash equilibrium and two alternative solution concepts: maximin
strategies and Aumann’s correlated equilibrium. In particular, we study how much a
player can lose by playing a maximin strategy instead of a Nash equilibrium, and how
much society can benefit from correlation.
4.6 · Comparative Ratios 47

4.6.1 The Price of Cautiousness

A compelling question is how much worse off a player can be when if he were to revert
to his most defensive course of action—his maximin strategy—instead of hoping for an
equilibrium outcome. This difference in payoff can be represented by a numerical value
which we refer to as the price of cautiousness. In the following, let G denote the class of
all normal-form games, and for Γ ∈ G, let N(Γ ) be the set of Nash equilibria of Γ . Further
recall that vi (Γ ) denotes player i’s security level in game Γ .

Definition 4.9 (price of cautiousness). Let Γ be a normal-form game with non-negative

payoffs, i ∈ N a player such that vi (Γ ) > 0. The price of cautiousness for player i in Γ
is defined as
min { pi (sN ) : sN ∈ N(Γ ) }
PC i (Γ ) = .
vi (Γ )
We further write PC i (C) = supΓ ∈C PC i (Γ ), where C ⊆ G can be any class of games
involving player i. In other words, the price of cautiousness of a player is the ratio
between his minimum payoff in a Nash equilibrium and his security level. It thus captures
the worst-case loss the player may incur by playing his maximin strategy instead of a
Nash equilibrium.2 For a player whose security level equals his minimum payoff of zero,
every strategy is a maximin strategy. Since we are mainly interested in a comparison of
normative solution concepts, we will thus only consider games where the security level of
at least one player is positive.
As we have already mentioned, Nash equilibrium and minimax strategies coincide in
two-player ranking games by virtue of the Minimax Theorem of von Neumann (1928), so
the price of cautiousness equals 1 for these games. In general ranking games, on the other
hand, the price of cautiousness is unbounded.

Theorem 4.10. Let R be the class of ranking games with more than two players that
involve player i. Then, the price of cautiousness is unbounded, i.e., PC i (R) = ∞,
even if R only contains games without weakly dominated actions.

Proof. Consider the game Γ1 of Figure 4.8, which is a ranking game for rank payoff
vectors ~p1 = (1, , 0), ~p2 = (1, 0, 0), and ~p3 = (1, 1, 0), and rankings [2, 3, 1], [1, 3, 2],
[1, 2, 3], [2, 1, 3], and [3, 1, 2]. It is easily verified that none of the actions of Γ1 is weakly
dominated and that v1 (Γ1 ) = . Let further sN = (s1 , s2 , c1 ) be the strategy profile
where s1 and s2 are uniform mixtures of a1 and a2 , and of b1 and b2 , respectively. It
is easily verified that sN is a Nash equilibrium of Γ1 , and we will argue that it is in fact
the only one. For this, consider the possible strategies of player 3. If player 3 plays c1 ,
the game reduces to the well-known Matching Pennies game for players 1 and 2, the only
Nash equilibrium being the one described above. If on the other hand player 3 plays c2 ,
2
In our context, the choice of whether to use the worst or the best equilibrium when defining the price
of cautiousness is merely a matter of taste. All results in this section still hold when the best equilibrium
is used instead of the worst one.
48 4 · Ranking Games

c1 c2
b1 b2 b1 b2
a1 (0, 1, 1) (1, 0, 0) (, 1, 0) (, 0, 1)
a2 (1, 0, 1) (0, 1, 1) (, 1, 0) (, 0, 1)

Figure 4.8: Three-player ranking game Γ1 used in the proof of Theorem 4.10

c1 c2
b1 b2 b1 b2
a1 2 1 3 1
a2 1 2 1 1

Figure 4.9: Three-player single-winner game used in the proof of Theorem 4.11. Dotted
boxes mark all Nash equilibria, one player may mix arbitrarily in boxes that span two
outcomes.

action b1 strictly dominates b2 . If b1 is played, however, player 3 will deviate to c1 to

get a higher payoff. Finally, if player 3 randomizes between actions c1 and c2 , the payoff
obtained from both of these actions must be the same. This can only be the case if either
player 1 plays a1 and player 2 randomizes between b1 and b2 , or if player 1 plays a2
and player 2 plays b2 . In the former case, player 2 will deviate to b1 . In the latter case,
player 1 will deviate to a1 . Since the payoff of player 1 in the above equilibrium is 1/2,
we have PC (Γ1 ) = 1/(2) → ∞ for  → 0.

We proceed to show that, due to their structural limitations, the price of cautiousness
in binary ranking games is bounded from above by the number of actions of the respective
player. We also derive a matching lower bound.

Theorem 4.11. Let Rb be the class of binary ranking games with more than two
players involving a player i with exactly k actions. Then, PC i (Rb ) = k, even if Rb
only contains single-winner games or games without weakly dominated actions.

Proof. By definition, the price of cautiousness takes its maximum for maximum payoff
in a Nash equilibrium, which is bounded by 1 in a ranking game, and minimum security
level. We require the security level to be strictly positive, so for every opponent action
profile s−i ∈ S−i there is some action ai ∈ Ai such that pi (ai , s−i ) > 0, i.e., pi (ai , s−i ) =
1. It is then easily verified that player i can ensure a security level of 1/k by uniform
randomization over his k actions. This results in a price of cautiousness of at most k.
For a matching lower bound, consider the single-winner game depicted in Figure 4.9.
We will argue that all Nash equilibria of this game are mixtures of the action profiles
4.6 · Comparative Ratios 49

c1 c2
b1 b2 b1 b2
a1 (0, 1, 1) (1, 0, 0) (0, 1, 0) (1, 0, 0)
a2 (1, 0, 0) (0, 1, 0) (1, 0, 1) (1, 0, 1)

Figure 4.10: Three-player ranking game Γ2 used in the proof of Theorem 4.11

(a2 , b1 , c2 ), (a2 , b2 , c2 ) and (a1 , b2 , c2 ). Each of these equilibria yields payoff 1 for
player 1, twice as much as his security level of 1/2. To appreciate this, consider the
possible strategies for player 3. If player 3 plays c1 , the game reduces to the well-known
Matching Pennies game for players 1 and 2, in which they will randomize uniformly over
both of their actions. In this case, player 3 will deviate to c2 . If player 3 plays c2 , we im-
mediately obtain the equilibria described above. Finally, if player 3 randomizes between
actions c1 and c2 , the payoff obtained from both of these actions should be the same.
This can only be the case if either player 1 plays a2 and player 2 randomizes between b1
and b2 , or if player 1 randomizes between a1 and a2 and player 2 plays b2 . In the former
case, player 2 will play b2 , causing player 1 to deviate to a1 . In the latter case, player 1
will play a1 , causing player 2 to deviate to b1 .
The above construction can be generalized to k > 2 by virtue of a single-winner game
with actions A1 = {a1 , a2 , . . . , ak }, A2 = {b1 , b2 , . . . , bk }, and A3 = {c1 , c2 }, and payoffs



(0, 1, 0) if ` = 1 and i 6= k − j + 1
i j `
p(a , b , c ) = (0, 0, 1) if ` = 2 and i = j = 1


(1, 0, 0) otherwise.

It is easily verified that the security level of player 1 in this game is 1/k, while, by similar
arguments as above, his payoff in every Nash equilibrium equals 1. This shows tightness
of the upper bound of k on the price of cautiousness for single-winner games.
Now consider the game Γ2 of Figure 4.10, which is a ranking game for rank payoff
vectors ~p1 = ~p2 = (1, 0, 0) and ~p3 = (1, 1, 0), and rankings [2, 3, 1], [1, 2, 3], [2, 1, 3], and
[1, 3, 2]. It is easily verified that none of the actions of Γ2 is weakly dominated and that
v1 (Γ2 ) = 1/2. On the other hand, we will argue that all Nash equilibria of Γ2 are mixtures
of action profiles (a2 , b1 , c2 ) and (a2 , b2 , c2 ), corresponding to a payoff of 1 for player 1.
To see this, we again look at the possible strategies for player 3. If player 3 plays c1 ,
players 1 and 2 will again randomize uniformly over both of their actions, causing player 3
to deviate to c2 . If player 3 plays c2 , we immediately obtain the equilibria described
above. Finally, assume that player 3 randomizes between actions c1 and c2 , and let α
denote the probability with which player 1 plays a1 . Again, player 3 must be indifferent
between c1 and c2 , which can only hold for 1/2 6 α 6 1. In this case, however, player 2
will deviate to b1 .
50 4 · Ranking Games

This construction can be generalized to k > 2 by virtue of a game with actions A1 =

{a1 , a2 , . . . , ak }, A2 = {b1 , b2 , . . . , bk }, and A3 = {c1 , c2 }, and payoffs


(0, 1, 1) if i = j = ` = 1





(1, 0, 0) if ` = 1 and i = k − j + 1
p(ai , bj , c` ) = or ` = 2, i = 1 and j > 1





(1, 0, 1) if ` = 2 and j > 2

(0, 1, 0) otherwise.

Again, it is easily verified that player 1 has a security level of 1/k, while his payoff is 1 in
every Nash equilibrium by similar arguments as above. Thus, the upper bound of k for the
price of cautiousness is tight as well for binary ranking games without weakly dominated
actions.

Informally, the previous theorem states that the payoff a player with k actions can
obtain in Nash equilibrium can be at most k times his security level.

4.6.2 The Value of Correlation

Nash equilibrium is based on the assumption that players select their actions indepen-
dently from each other. Aumann (1974) generalizes the notion of a strategy profile by
allowing players to coordinate their actions by means of a device or agent that randomly
selects one of several action profiles and recommends the actions of this profile to the
respective players. More formally, a correlated strategy µ ∈ ∆(AN ) is a probability dis-
tribution over the set of action profiles. The corresponding equilibrium concept is then
defined as follows.

Definition 4.12 (correlated equilibrium). A correlated strategy µ ∈ ∆(AN ) is called a

correlated equilibrium if for all i ∈ N and all a∗i , ai ∈ Ai ,
X
µ(a−i , a∗i )(pi (a−i , a∗i ) − pi (a−i , ai )) > 0.
a−i ∈A−i

In other words, a correlated equilibrium of a game is a probability distribution µ over

the set of action profiles, such that, if a particular action profile a∗N ∈ AN is chosen
according to this distribution, and every player i ∈ N is only informed about his own
action a∗i , it is optimal in expectation for every player i ∈ N to play a∗i , given that he
only knows the conditional distribution over values of a∗−i . Correlated equilibrium makes
stronger assumptions than Nash equilibrium in that it assumes the existence of a trusted
third party who can recommend behavior, but cannot enforce it. Using cryptographic
means, this requirement can essentially be reduced to the ability to carry out a distributed
computation among the players (Dodis et al., 2000).
4.6 · Comparative Ratios 51

It can easily be seen that every Nash equilibrium naturally corresponds to a corre-
lated equilibrium. Nash’s existence result thus carries over to correlated equilibria. Again
consider the game of Figure 4.1 on Page 31. The correlated strategy that assigns proba-
bility 1/4 each to action profiles (a1 , b1 , c1 ), (a1 , b2 , c1 ), (a2 , b1 , c1 ), and (a2 , b1 , c2 ) is
a correlated equilibrium, with an expected payoff of 1/2 for player 1 and 1/4 for players 2
and 3. In this particular case, the correlated equilibrium is a convex combination of Nash
equilibria, and correlation can be achieved by means of a publicly observable random vari-
able. Quite surprisingly, Aumann (1974) has shown that in general the (expected) social
welfare of a correlated equilibrium may exceed that of every Nash equilibrium, and that
correlated equilibrium payoffs may in fact be outside the convex hull of the Nash equilib-
rium payoffs. This is of course not possible if social welfare is identical in all outcomes,
as is the case in our example.
We will now turn to the question whether, and by which amount, social welfare in
a ranking game can be improved by allowing players to correlate their actions. Just as
the payoff of a player in any Nash equilibrium is at least his security level, social welfare
in the best correlated equilibrium is at least as high as social welfare in the best Nash
equilibrium. In order to quantify the value of correlation in strategic games with non-
negative payoffs, Ashlagi et al. (2005) introduce the mediation value of a game as the ratio
between the maximum social welfare in a correlated versus that in a Nash equilibrium,
and the enforcement value as the ratio between the maximum social welfare in any
outcome versus that in a correlated equilibrium. Whenever social welfare, i.e., the sum of
all players’ payoffs, is used as a measure of global satisfaction, one implicitly assumes the
inter-agent comparability of payoffs. While this assumption is controversial, social welfare
is nevertheless commonly used in the definitions of comparative ratios such as the price
of anarchy (Koutsoupias and Papadimitriou, 1999). For Γ ∈ G and X ⊆ ∆(AN ), let C(Γ )
denote the set of correlated equilibria of Γ and let vX (Γ ) = max{ p(sN ) : sN ∈ X }. Recall
that N(Γ ) denotes the set of Nash equilibria of game Γ .
Definition 4.13 (mediation value, enforcement value). Let Γ be a normal-form game
with non-negative payoffs. Then, the mediation value MV (Γ ) and the enforcement
value EV (Γ ) of Γ are defined as
vC(Γ ) (Γ ) vSN (Γ )
MV (Γ ) = and EV (Γ ) = .
vN(Γ ) (Γ ) vC(Γ ) (Γ )
If both numerator and denominator are 0 for one of the values, the respective value is
defined to be 1. If only the denominator is 0, the value is defined to be ∞. For any class
C ⊆ G of games, we further write MV (C) = supΓ ∈C MV (Γ ) and EV (C) = supΓ ∈C EV (Γ ).
Ashlagi et al. (2005) have shown that both the mediation value and the enforcement
value cannot be bounded for games with an arbitrary payoff structure, as soon as there
are more than two players, or some player has more than two actions. This holds even if
payoffs are normalized to the interval [0, 1]. Ranking games also satisfy this normalization
criterion, and here social welfare is also strictly positive for every outcome of the game.
52 4 · Ranking Games

c1 c2 c3
b1 b2 b1 b2 b1 b2
a1 (1, 1, 0) (1, 0, 0) (0, 1, 1) (0, 1, 0) (1, 0, 0) (0, 0, 1)
a2 (0, 1, 0) (0, 1, 1) (1, 0, 0) (1, 1, 0) (0, 0, 1) (1, 0, 0)

Figure 4.11: Three-player ranking game Γ3 used in the proof of Theorem 4.14

Ranking games with identical rank payoff vectors for all players, i.e., ones where pk k
i = pj
for all i, j ∈ N and 1 6 k 6 n, are constant-sum games. Hence, social welfare is the same
in every outcome so that both the mediation value and the enforcement value are 1. This
in particular concerns all ranking games with two players. In general, social welfare in
an arbitrary outcome of a ranking game is bounded by n − 1 from above and by 1 from
below. Since the Nash and correlated equilibrium payoffs must lie in the convex hull of
the feasible payoffs of the game, we obtain trivial lower and upper bounds of 1 and n − 1,
respectively, on both the mediation and the enforcement value. It turns out that the
upper bound of n − 1 is tight for both the mediation value and the enforcement value.

Theorem 4.14. Let R 0 be the class of ranking games with n > 2 players, such that in
games with only three players at least one player has more than two actions. Then,
MV (R 0 ) = n − 1.

Proof. It suffices to show that for any of the above cases there is a ranking game with
mediation value n − 1. For n = 3, consider the game Γ3 of Figure 4.11, which is a ranking
game for rank payoff vectors ~p1 = ~p3 = (1, 0, 0) and ~p2 = (1, 1, 0). First of all, we will
argue that every Nash equilibrium of this game has social welfare 1, by showing that there
are no Nash equilibria where c1 or c2 are played with positive probability. Assume for
contradiction that s∗N is such an equilibrium. The strategy played by player 3 in s∗N must
either be (i) c1 or c2 as a pure strategy, (ii) a mixture of c1 and c3 or between c2 and c3 ,
or (iii) a mixture where both c1 and c2 are played with positive probability. If player 3
plays a pure strategy, the game reduces to a two-player game for players 1 and 2. In the
case of c1 , this game has the unique equilibrium (a1 , b1 ), which in turn causes player 3
to deviate to c2 . In the case of c2 , the unique equilibrium is (a2 , b2 ), causing player 3 to
deviate to c1 . Now assume that player 3 mixes between c1 and c3 , and let α and β denote
the probabilities with which players 1 and 2 play a1 and b1 , respectively. Since player 3’s
payoff from c1 and c3 must be the same in such an equilibrium, we must either have
α = β = 1, in which case player 3 will deviate to c2 , or 0 6 α 6 1/2 and 0 6 β 6 1/2,
causing player 2 to deviate to b1 . Analogously, if player 3 mixes between c2 and c3 , we
must either have α = β = 0, in which case player 3 will deviate to c1 , or 1/2 6 α 6 1
and 1/2 6 β 6 1, causing player 2 to deviate to b2 . Finally, if both c1 and c2 are played
with positive probability, we must have α + β = 1 for player 3 to get an identical payoff
of αβ 6 1/4 from both c1 and c2 . In this case, however, player 3 can deviate to c3 for a
4.6 · Comparative Ratios 53

c1 c2
b1 b2 b1 b2
a1 (1, 1, 0, 1) (1, 0, 0, 0) (0, 1, 1, 1) (0, 1, 0, 0)
d1
a2 (0, 1, 0, 0) (0, 1, 1, 1) (1, 0, 0, 0) (1, 1, 0, 1)

a1 (0, 0, 0, 1) (0, 0, 0, 1) (0, 0, 0, 1) (0, 0, 0, 1)

d2
a2 (0, 0, 0, 1) (0, 0, 0, 1) (0, 0, 0, 1) (0, 0, 0, 1)

Figure 4.12: Four-player ranking game Γ4 used in the proof of Theorem 4.14

strictly greater payoff of 1 − 2αβ. Thus, a strategy profile s∗N as described above cannot
exist.
Now let µ∗ be the correlated strategy where action profiles (a1 , b1 , c1 ), (a2 , b2 , c1 ),
(a1 , b1 , c2 ), and (a2 , b2 , c2 ) are played with probability 1/4 each. This correlation can for
example be achieved by tossing two coins independently. Players 1 and 2 observe the first
coin toss and play a1 and b1 if the coin falls on heads, and a2 and b2 otherwise. Player
3 observes the second coin toss and plays c1 if the coin falls on heads and c2 otherwise.
The expected payoff for player 2 under µ∗ is 1, so he cannot gain by changing his action.
If player 1 observes heads, he knows that player 2 will play b1 , and that player 3 will play
c1 and c2 with probability 1/2 each. He is thus indifferent between a1 and a2 . Player 3
knows that players 1 and 2 will play (a1 , b1 ) and (a2 , b2 ) with probability 1/2 each, so
he is indifferent between c1 and c2 and strictly prefers both of them to c3 . Hence, none
of the players has an incentive to deviate, µ∗ is a correlated equilibrium. Moreover, the
social welfare under µ∗ is 2, and thus MV (Γ3 ) = 2.
Now consider the four-player game Γ4 of Figure 4.12, which is a ranking game for rank
payoffs ~p1 = ~p3 = (1, 0, 0, 0), ~p2 = (1, 1, 0, 0), and ~p4 = (1, 1, 1, 0), and rankings [1, 2, 4, 3],
[1, 3, 2, 4], [3, 2, 4, 1], [2, 3, 1, 4], and [4, 1, 2, 3]. It is easily verified that none of the action
profiles with social welfare 2 is a Nash equilibrium. Furthermore, player 4 strictly prefers
action d2 over d1 as soon as one of the remaining action profiles for players 1 to 3, i.e.,
those in the upper half of the game where the social welfare is 1, is played with positive
probability. Hence, d1 is not played with positive probability in any Nash equilibrium
of Γ4 , and every Nash equilibrium of Γ4 has social welfare 1. In turn, consider the correlated
strategy µ∗ where actions profiles (a1 , b1 , c1 , d1 ), (a2 , b2 , c1 , d1 ), (a1 , b1 , c2 , d1 ), and
(a2 , b2 , c2 , d1 ) are played with probability 1/4 each. It is easily verified that none of the
players can increase his payoff by unilaterally deviating from µ∗ . Hence, µ∗ is a correlated
equilibrium with social welfare 3, and MV (Γ4 ) = 3.
For n > 4, we can restrict our attention to games where the additional players only
have a single action. We return to the game Γ4 of Figure 4.12 and transform it into a
54 4 · Ranking Games

c1 c2
b1 b2 b1 b2
a1 (1, 0, 0) (0, 1, ) (0, 0, 1) (1, 0, )
a2 (0, 1, ) (1, 0, 1) (1, 0, ) (1, 0, )

Figure 4.13: Three-player ranking game Γ5 used in the proof of Theorem 4.15

game Γ4n with n > 4 players by assigning to players 5, 6, . . . , n a payoff of 1 in the four
action profiles (a1 , b1 , c1 , d1 ), (a2 , b2 , c1 , d1 ), (a1 , b1 , c2 , d1 ), and (a2 , b2 , c2 , d1 ) that
constitute the correlated equilibrium with maximum social welfare, and a payoff of zero in
all other action profiles. Since the additional players cannot influence the outcome of the
game, this construction does not affect the equilibria of the game. To see that the resulting
game is a ranking game, consider the rank payoff vectors ~p1 = ~p3 = (1, 0, 0, . . . , 0),
~p2 = (1, 1, 0, . . . , 0), rkm = 1 if k 6 m − 1 and 0 otherwise, for m > 4. It is easily verified
that we can retain the original payoffs of players 1 to 4 and at the same time assign a
payoff of 0 or 1, respectively, to players 5 to n by ranking the latter according to their
index and placing either no other players or exactly one other player behind them in the
overall ranking. More precisely, Γ4n is a ranking game by virtue of the above rank payoffs
and rankings [1, 2, 4, 5, . . . , n, 3], [1, 3, 2, 4, 5, . . . , n], [3, 2, 4, 5, . . . , n, 1], [2, 3, 1, 4, 5, . . . , n],
and [4, 1, 2, 3, 5, . . . , n]. Furthermore, MV (Γ4n ) = n − 1.

Theorem 4.15. Let R be the class of ranking games with n > 2 players. Then,
EV (R) = n − 1, even if R only contains games without weakly dominated actions.

Proof. It suffices to show that for any n > 3 there is a ranking game with enforcement
value n − 1 in which no action is weakly dominated. Consider the ranking game Γ5 of
Figure 4.13, which is a ranking game by virtue of rank payoff vectors ~p1 = (1, 1, 0),
~p2 = (1, 0, 0), and ~p3 = (1, , 0) and rankings [1, 2, 3], [2, 3, 1], [3, 1, 2], [3, 2, 1], and [1, 3, 2].
Obviously, all of the actions of Γ5 are undominated and vSN (Γ5 ) = 2. It remains to be
shown that the social welfare in any correlated equilibrium of Γ5 is at most (1 + ), such
that vC(Γ5 ) (Γ5 ) → 1 and EV (Γ5 ) → 2 for  → 0.
Finding a correlated equilibrium that maximizes social welfare constitutes a linear
programming problem constrained by the inequalities of Definition 4.12 and the prob-
P
ability constraints aN ∈AN µ(aN ) = 1 and µ(aN ) > 0 for all aN ∈ AN . Feasibility of
this problem is a direct consequence of Nash’s existence theorem. Boundedness follows
from boundedness of the quantity being maximized. To derive an upper bound for social
welfare in a correlated equilibrium of Γ5 , we will transform the above linear program into
its dual. Since the primal is feasible and bounded, the primal and the dual will have
the same optimal value, in our case the maximum social welfare in a correlated equi-
librium. The latter constitutes a minimization problem and finding a feasible solution
with objective value v shows that the optimal value cannot be greater than v. Since
4.7 · Discussion 55

minimize v
subject to
−x1 + y1 + z1 + v > 1,
x2 − y1 + v > 1 + ,
x1 − y2 + v > 1 + ,
−x2 + y2 + ( − 1)z1 + v > 2,
x1 − z2 + v > 1,
−x2 + v > 1 + ,
v > 1 + ,
(1 − )z2 + v > 1 + ,
x1 > 0, x2 > 0, y1 > 0, y2 > 0, z1 > 0, and z2 > 0.

Figure 4.14: Dual linear program for computing a correlated equilibrium of Γ5 , used in
the proof of Theorem 4.15

there are three players with two actions each, the primal has six constraints of the form
P
a−i ∈A−i µ(a−i , ai )(pi (a−i , ai ) − pi (a−i , ai )) > 0. For j ∈ {1, 2}, let xj , yj , and zj de-
∗ ∗

note the variable of the dual associated with the constraint for the jth action of player 1, 2,
and 3, respectively. Furthermore, let v denote the variable of the dual associated with
P
constraint aN ∈AN µ(aN ) = 1 of the primal. The dual is given in Figure 4.14.
Now let x2 = y1 = z2 = 0, x1 = y2 = ( − 1)2 /, z1 = (1 − 2)/, and v = 1 + ,
and observe that for every  > 0, this is a feasible solution with objective value 1 + .
However, the objective value of any feasible solution to the dual is an upper bound for
that of the optimal solution, which in turn equals vC(Γ5 ) (Γ5 ).
The above construction can easily be generalized to games Γ5n with n > 4 by adding
additional players that receive payoff 1 in action profile aN if a1 = a2 , a2 = b2 , and
a3 = c1 , and payoff 0 otherwise. This can for example be achieved by means of rank
payoff vectors ~p1 = (1, 0, . . . , 0), ~p2 = (1, 1, 0, . . . , 0), ~p3 = (1, , 0, . . . , 0), and ~pk
m = 1
if k 6 m − 1 and 0 otherwise for m > 4. By the same arguments as in the proof of
Theorem 4.14, this does not affect the maximum social welfare achievable in a correlated
equilibrium. It is thus easily verified that EV (Γ5k1 ×···×k4 ) → n − 1 for  → 0.

4.7 Discussion
In this chapter we proposed a new class of normal-form games, so-called ranking games,
which model settings in which players are merely interested in outperforming their op-
ponents. Arguably this class of games is very natural and relevant for many realistic
scenarios, and provides a meaningful generalization of the property of strict competitive-
ness attributed to constant-sum games in the two-player case. Despite the structural
simplicity of ranking games, however, various solution concepts turned out to be just as
hard to compute as in general normal-form games. In particular we obtained hardness
56 4 · Ranking Games

results for mixed Nash equilibria and iterated weak dominance in games with more than
two players and pure Nash equilibria in games with an unbounded number of players.
As a consequence, the mentioned solution concepts appear to be of limited use in large
instances of ranking games that do not possess additional structure. This underlines the
importance of alternative, efficiently computable, solution concepts for ranking games such
as maximin strategies or correlated equilibrium.
Based on these findings, we then quantified and bounded ratios comparing different
solution concepts in ranking games. It turned out that playing one’s maximin strategy in
binary ranking games with only few actions might be a prudent choice, not only because
this strategy guarantees a certain payoff even when playing against irrational opponents,
but also because of the limited price of cautiousness and the inherent weakness of Nash
equilibria in ranking games.
We also investigated the relationship between Nash equilibria and correlated equilibria.
While correlation can never decrease social welfare, it is an important question which
scenarios permit an increase. This question is particularly relevant for scenarios that
would intuitively be considered to be highly competitive. To this end, we showed that
in ranking games with many players and asymmetric preferences over ranks, i.e., with
non-identical rank payoff vectors, overall satisfaction can be improved substantially by
allowing players to correlate their actions.
 Chapter 5

Anonymous Games

A major obstacle when considering normal-form games with an unbounded number of

players is the exponential size of the explicit representation of the payoffs, an issue we
already touched upon in Section 4.5.3 when we investigated the computational complexity
of pure Nash equilibria in ranking games. Realistic situations, however, often possess
additional structure, and allow an agent to make rational decisions without reasoning
explicitly about all the possible ways the other agents may behave. A particular property
often found in situations with many agents is that the agents are in some sense similar
to each other, such that a particular agent does not, or need not, distinguish between the
other agents.
In this chapter, we consider four classes of anonymous games and study the com-
plexity of pure Nash equilibria and iterated dominance in these games. More formally,
an anonymous game is characterized by the fact that players do not distinguish between
other players in the game, i.e., their payoff only depends on the numbers of other players
playing the different actions, but not on their identities. Anonymous games constitute
a very natural class of multi-player games which is also highly relevant in practice (cf.
Daskalakis and Papadimitriou, 2007). A symmetric game additionally has identical pay-
off functions for all players. Two more classes, called self-anonymous and self-symmetric
games in this thesis, are obtained by assuming that a player does not distinguish himself
from the other players.
It turns out that in all four classes pure Nash equilibria can be found efficiently if
only a constant number of actions is available to each player. Moreover, identical payoff
functions for all players further reduce the computational complexity of pure equilibria,
an effect that is nullified as soon as there are two different payoff functions. The fact
that a player cannot, or does not, distinguish himself from the other players, does not
seem to offer any computational advantage. Finally, computing pure equilibria becomes
intractable in all four classes of anonymous games when the number of actions grows in
the number of players.
In the second part of the chapter we turn to iterated weak dominance. We show that
iterated dominance solvability is NP-hard for symmetric games with a growing number

57
58 5 · Anonymous Games

of actions, and tractable for symmetric games with a constant number of actions. The
only case that remains is that of anonymous games with a constant number of actions.
When restricted to two actions, it can be reformulated as a natural elimination problem
on matrices. The complexity of this problem remains open, but it turns out to be related
to a matching problem on paths of a directed graph. The latter problem, which may be of
independent interest, is intractable in general but allows us to obtain efficient algorithms
for restricted versions of matrix elimination. We finally use the matching formulation to
show NP-hardness of iterated dominance in anonymous games with three actions.

5.1 Related Work

Symmetries in games have been investigated since the earliest days of game theory. Von
Neumann (1928) and von Neumann and Morgenstern (1947) were the first to consider
symmetries of cooperative games, calling a game in characteristic form symmetric if the
value of a coalition depends only on its size. In the context of normal-form games, a
game is usually called symmetric if the payoff functions of all players are identical and
symmetric in the other players’ actions (von Neumann and Morgenstern, 1947, Luce
and Raiffa, 1957). For two-player normal-form games, this restriction corresponds to a
skew-symmetric payoff matrix (e.g., Borel, 1921, Gale et al., 1950).
Most early research on symmetries in games has concentrated on these symmetric
games. One of the reasons for this may have been the strong focus of the early research
in non-cooperative game theory on two-player games, where anonymity as defined in
this chapter does not impose any restrictions. Gale et al. (1950) provide a (polynomial-
time) reduction from arbitrary two-player games to symmetric two-player games which
preserves equilibria. The recent PPAD-completeness result of Chen and Deng (2006) thus
also applies to symmetric games with two players and a large number of actions. An early
result by Nash (1951) shows that there always exists an equilibrium that “respects” all
symmetries of a game, which in symmetric games implies the existence of a symmetric
equilibrium, i.e., one where all players play the same (mixed) strategy. Papadimitriou and
Roughgarden (2005) capitalize on this existence result and show that a Nash equilibrium of
a symmetric game with n players and k actions can be computed in polynomial time if k =
O(log n/ log log n). While their tractability results for correlated equilibrium (Aumann,
1974) do not rely on identical payoff functions and hence apply to anonymous games
as well, this is not the case for the results about Nash equilibria. The aforementioned
existence of symmetric Nash equilibria neither extends to pure equilibria, nor does it
hold for anonymous games. For example, Figure 5.3 on Page 64 shows an anonymous
game without any symmetric equilibria.
Computational aspects of the larger class of anonymous games have recently come
under increased scrutiny due to their importance in modeling large anonymous environ-
ments like the Internet. PPAD-hardness of the Nash equilibrium problem in general
games has led to an increased interest in approximate equilibria, and anonymous games
5.1 · Related Work 59

with a constant number of actions have recently been shown to admit a polynomial-time
approximation scheme (Daskalakis and Papadimitriou, 2008).
Obviously, deciding the existence of a pure equilibrium is easy if the number of can-
didates for such an equilibrium, i.e., the number of action profiles, is polynomial in the
size of the game. This is certainly the case for the explicit representation of a game as
a multi-dimensional table of payoffs, but no longer holds if the game is represented suc-
cinctly. For example, deciding the existence of a pure equilibrium is NP-complete for
games in graphical normal form (Gottlob et al., 2005), which we consider in Chapter 6,
and games in circuit form (Schoenebeck and Vadhan, 2006). Quite a few classes of games
are related to anonymity in that they exploit some form of independence among certain
actions or players playing these actions. In congestion games (Rosenthal, 1973), the avail-
able actions consist of sets of resources, and the payoff depends on the number of other
players that have played the same action and selected the same resources. Congestion
games always have a pure equilibrium (Rosenthal, 1973), and finding one is PLS-complete
even for symmetric congestion games, but in P in the symmetric network case (Fabrikant
et al., 2004). For singleton (or simple) congestion games, where only a single resource can
be selected, there is a polynomial-time algorithm for finding a social-welfare-maximizing
equilibrium (Ieong et al., 2005). In local-effect games (Leyton-Brown and Tennenholtz,
2003), the payoff from an action may also depend on (a function of) the number of agents
playing “neighboring” actions. Dunkel and Schulz (2008) give hardness results for the
pure equilibrium problem in several classes of congestion and local-effect games. Unlike
congestion games and local-effect games, action-graph games (Bhat and Leyton-Brown,
2004) can encode arbitrary payoff functions. For action-graph games of bounded degree,
expected payoffs and the Jacobian of the payoff function can be computed in polynomial
time. The latter forms the practical bottleneck step of the algorithm of Govindan and
Wilson (2003) for finding Nash equilibria, but the algorithm may still take exponentially
many steps to converge even for bounded degree. In fact, the pure equilibrium prob-
lem is NP-complete for symmetric action-graph games with bounded degree, but becomes
tractable if the treewidth is bounded (Jiang and Leyton-Brown, 2007). In general action-
graph games, the pure equilibrium problem is NP-complete even if the action-graph is a
bounded-degree tree (Daskalakis et al., 2009b).
Deciding whether a general game can be solved by iterated weak dominance is NP-
complete already for games with two players and two different payoffs (Gilboa et al., 1993,
Conitzer and Sandholm, 2005a), even when restricted to dominance by pure strategies.
The corresponding problem for strict dominance, which requires the dominating strategy
to be strictly better under any circumstance, can be solved in polynomial time (e.g.,
Conitzer and Sandholm, 2005a). Knuth et al. (1988) provide an improved algorithm for
the case of two players and dominance by pure strategies, and show that computing the
reduced game in this case is P-complete. Apart from the results given in Section 4.5.2,
and recent work by Brandt et al. (2009b), we are not aware of any complexity results for
iterated dominance in restricted classes of games.
60 5 · Anonymous Games

5.2 The Model

Symmetry as a property of a mathematical object refers to its invariance under a certain
type of transformation. Symmetries of games usually mean invariance of the payoffs under
automorphisms of the set of action profiles induced by some group of permutations of the
set of players. Since such an automorphism preserves the number of players that play
a particular action, a characteristic feature of symmetries in games is the inability to
distinguish between other players. The most general class of games with this property
will be called anonymous. Four different classes of games are obtained by considering
two additional characteristics: identical payoff functions for all players and the ability
to distinguish oneself from the other players. The games obtained by adding the former
property will be called symmetric, and presence of the latter will be indicated by the
prefix “self ”. For the obvious reason, we will henceforth talk about games where the set
of actions is the same for all players, and write A = A1 = A2 = · · · = An and k = |A|,
respectively, to denote this set and its cardinality.
Let Γ be such a game. For any permutation π : N → N of the set of players, let π 0 :
A → AN be the permutation of the set of action profiles such that π 0 ((a1 , a2 , . . . , an )) =
N

(aπ(1) , aπ(2) , . . . , aπ(n) ). Then, Γ is anonymous if pi (aN ) = pi (π 0 (aN )) for all aN ∈ AN ,

i ∈ N and all π with π(i) = i. Similarly, Γ is symmetric if pi (aN ) = pj (π 0 (aN ))
for all aN ∈ AN , i, j ∈ N and all π with π(j) = i. Finally, Γ is self-anonymous if
pi (aN ) = pi (π 0 (aN )) for all aN ∈ AN , i ∈ N, and self-symmetric if pi (aN ) = pj (π 0 (aN ))
for all aN ∈ AN , i, j ∈ N. Since π 0 is an automorphism of the set of action profiles that
preserves the number of players who play a particular action, an intuitive way to describe
anonymous games is in terms of equivalence classes of the automorphism group of π 0 , using
a notion introduced by Parikh (1966) in the context of context-free languages. Given a
set A of actions, the commutative image of an action profile aN ∈ AN is given by
#(aN ) = (#(a, aN ))a∈A where #(a, aN ) = |{ i ∈ N : ai = a }|. In other words, #(a, aN )
denotes the number of players playing action a in action profile aN , and #(aN ) is the
vector of these numbers for all the different actions. This definition naturally extends to
action profiles for subsets of the players.

Definition 5.1 (anonymity). Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a normal-form game, A
a set of actions such that Ai = A for all i ∈ N. Γ is called

 anonymous if pi (aN ) = pi (aN

0 ) for all i ∈ N and all a , a 0 ∈ AN with a = a 0
N N i i
0 ),
and #(a−i ) = #(a−i

 symmetric if pi (aN ) = pj (aN

0 ) for all i, j ∈ N and all a , a 0 ∈ AN with a = a 0
N N i j
0
and #(a−i ) = #(a−j ),

 self-anonymous if pi (aN ) = pi (aN

0 ) for all i ∈ N and all a , a 0 ∈ AN with
N N
0 ), and
#(aN ) = #(aN
5.2 · The Model 61

anon

symm s-symm s-anon

Figure 5.1: Inclusion relationships between anonymous, symmetric, self-anonymous, and

self-symmetric games

 self-symmetric if pi (aN ) = pj (aN

0 ) for all i, j ∈ N and all a , a 0 ∈ AN with
N N
0
#(aN ) = #(aN ).

When talking about anonymous games, we write pi (ai , x−i ) to denote the payoff of
player i under any action profile aN with #(a−i ) = x−i . For self-anonymous games, pi (x)
is used to denote the payoff of player i under any profile aN with #(aN ) = x. It is easily
verified that the class of self-symmetric games equals the intersection of symmetric and
self-anonymous games, which in turn are both strictly contained in the class of anonymous
games. An illustration of these inclusions is shown in Figure 5.1. Figure 5.2 illustrates
the different payoff structures for n = 3 and k = 2.
In terms of the above characterization, a game is anonymous if the payoff pi (aN )
of player i ∈ N in action profile aN depends, besides his own action ai , only on the
number #(a, a−i ) of other players playing each of the actions a ∈ A, but not on who
plays them. If two players exchange actions, all other players’ payoffs remain the same.
For two-player games, anonymity does not impose any restrictions, because action sets of
equal size can simply be achieved by adding dummy actions. A game is symmetric if it
is anonymous and if the payoff function is the same for all players. Hence, if two players
exchange actions, their payoffs are also exchanged, while all other players’ payoffs remain
the same. Many well-known games like the Prisoner’s Dilemma, Rock-Paper-Scissors, or
Chicken are examples of symmetric (two-player) games. Simple congestion games (Ieong
et al., 2005) are an example for the multi-player case. In a self-anonymous game the
payoff of each player depends only on the number #(a, aN ) of players playing each of the
actions a ∈ A, including the player himself. If two players exchange actions, the payoffs
of all players remain the same. Matching Pennies is a self-anonymous two-player game,
and voting with identical weights can be seen as an example for the multi-player case.
Finally, in a self-symmetric game the payoff is always the same for all players and stays
the same if two players exchange actions. Self-symmetric games thus are a special case of
so-called common payoff games, in which all players get the same payoff. Obviously such
games always have a pure equilibrium, namely an action profile with maximum payoff.
Other games guaranteed to possess a pure equilibrium, and the complexity of finding an
equilibrium in these games, have been investigated by Fabrikant et al. (2004).
62 5 · Anonymous Games

(·, ·, ·) (a, g, c) (a, b, · ) ( · , e, f)

Γ1 :
( · , b, c) (d, e, · ) (d, · , f) (·, ·, ·)

(a, a, a) (b, c, b) (b, b, c) (e, d, d)

Γ2 :
(c, b, b) (d, d, e) (d, e, d) (f, f, f)

(·, ·, ·) (a, b, c) (a, b, c) (d, e, f)

Γ3 :
(a, b, c) (d, e, f) (d, e, f) (·, ·, ·)

(a, a, a) (b, b, b) (b, b, b) (c, c, c)

Γ4 :
(b, b, b) (c, c, c) (c, c, c) (d, d, d)

Figure 5.2: Relationships between the payoffs of anonymous (Γ1 ), symmetric (Γ2 ), self-
anonymous (Γ3 ), and self-symmetric (Γ4 ) games for n = 3 and k = 2. Players 1, 2,
and 3 simultaneously choose rows, columns, and tables, respectively, and obtain payoffs
according to the vector in the resulting cell. Each lower case letter stands for a payoff
value, dots denote arbitrary payoff values. As an example for the separation of the different
classes, Γ1 is not symmetric if a 6= c and not self-anonymous if b 6= g. Γ2 is not self-
anonymous if b 6= c. Γ3 is not self-symmetric if a 6= c.

There are n+k−1

k−1 distributions of n players among k actions. Since these are exactly
the equivalence classes of the set of action profiles for n−1 players under the commutative
image, an anonymous game can be represented using at most n · k · n+k−2


k−1 numbers.
In the following, we call the explicit representation of an anonymous game the one that
simply lists the payoffs for each of the above equivalence classes, and note that the explicit
representation requires space polynomial in n if and only if k is bounded by a constant.
On the other hand, its size becomes super-polynomial in n even for the slightest growth
of k. Nevertheless, space polynomial in n may still suffice to encode certain subclasses of
anonymous games with a larger number of actions if we use an implicit representation of
the payoff functions like a Boolean circuit. It is easy to see that for games with a constant
number of actions, any encoding of a game that has size at least linear in the number of
players and satisfies the basic assumptions of rational and efficient play made throughout
the thesis, is equivalent to its explicit representation under polynomial-time reductions.
Interestingly, the ability to distinguish oneself from the other players does not increase
the complexity of the pure equilibrium and iterated dominance problems when players
only have two actions.
5.3 · Pure Nash Equilibria 63

Lemma 5.2. There exists a constant-depth reduction from anonymous games with
two actions to self-anonymous games with two actions that preserves pure Nash
equilibria, dominance by pure strategies, and identical payoff functions.

Proof. Let Γ = (N, ({a1 , a2 })i∈N , (pi )i∈N ) be an anonymous game. We can define a new
game Γ 0 = (N, ({a1 , a2 })i∈N , (pi0 )i∈N ) such that for all i ∈ N and for all x ∈ {0, 1, . . . , n−1},

1. pi0 ((x, n−x)) > pi0 ((x+1, n−x−1)) if and only if pi (a1 , (x, n−x−1)) > pi (a2 , (x, n−
x − 1)),

2. pi0 ((x, n−x)) < pi0 ((x+1, n−x−1)) if and only if pi (a1 , (x, n−x−1)) < pi (a2 , (x, n−
x − 1)), and

3. pi0 ((x, n−x)) = pi0 ((x+1, n−x−1)) if and only if pi (a1 , (x, n−x−1)) = pi (a2 , (x, n−
x − 1)).

Depending on the payoff structure of Γ , it may be necessary to use up to n different payoffs

in Γ 0 , even when in Γ there are only two. It is now easily verified that Γ 0 is self-anonymous
in general, and self-symmetric if the original game Γ is symmetric.

It should be noted that the above construction cannot in general be extended to

games where players have more than two actions, because it may lead to cyclic preference
relations. The symmetric two-player game Rock-Paper-Scissors is an example for a game
that cannot be mapped to a corresponding self-symmetric game using the above technique.

5.3 Pure Nash Equilibria

For general games, simply checking the equilibrium condition for each action profile takes
time polynomial in the size of their explicit representation. Using a succinct representation
for games where the size of the explicit representation grows exponentially in the number of
players, which is the case for k = 2 already, quickly renders the problem NP-hard (Fischer
et al., 2006, Schoenebeck and Vadhan, 2006). On the other hand, the polynomial size even
of the explicit representation for anonymous games with a constant number of actions
might suggest that finding pure equilibria is easy by a similar argument as above. This
reasoning is flawed, however, since a single entry in the payoff table corresponds to an
exponential number of action profiles, and it is very well possible that only a single one of
them is an equilibrium while all others are not. The anonymous game given in Figure 5.3
illustrates this fact.

5.3.1 Games with a Constant Number of Actions

We begin by investigating games with a constant number of actions. Obviously, solving
a game cannot be easier than playing it optimally given that the opponents’ actions are
known. The most interesting upper bounds for the former problem will thus be obtained
64 5 · Anonymous Games

(0, 1, 1) (0, 0, 1) (0, 1, 0) (0, 0, 0)

(1, 1, 1) (0, 0, 0) (0, 1, 0) (1, 0, 1)

Figure 5.3: Anonymous game with a unique, non-symmetric Nash equilibrium at the
action profile with payoff (1, 1, 1). Players 1, 2, and 3 choose rows, columns, and tables,
respectively. Outcomes are denoted as a vector of payoffs for the three players. Action
profiles with the same commutative image as the equilibrium are shaded.

when the latter problem is easy. We therefore assume throughout this section that for
any action profile of his opponents, a player can compute the payoff of a particular action
in AC0 , i.e., by evaluating a Boolean circuit with constant depth and bounded fan-in.
This particularly holds if payoffs are given explicitly. It will further be obvious from the
proofs that for payoff functions that are harder to compute, the complexity of the pure
equilibrium problem exactly matches that of computing the payoff function.
As we have noted earlier, the potential hardness of finding pure equilibria in games
with succinct representation stems from the fact that the number of action profiles that
are candidates for being an equilibrium is exponential in the size of the representation of
the game. While anonymous games do satisfy this property, the pure equilibrium problem
nevertheless turns out to be tractable. The following theorem concerns games where the
number of players is polynomial in the size of the representation.
Theorem 5.3. Deciding whether an anonymous or self-anonymous game with a
constant number of actions has a pure Nash equilibrium is TC0 -complete under
constant-depth reducibility. Hardness holds even for games with three different pay-
offs and two different payoff functions.
Proof. For membership in TC0 , we propose an algorithm that decides whether there exists
a pure Nash equilibrium with a given commutative image. The theorem then follows by
observing that the number of different commutative images is polynomial in the number
of players if the number of actions is constant.
Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be an anonymous game, A = {a1 , a2 , . . . , ak } a set of
actions such that Ai = A for all i ∈ N. Given the commutative image x = (x1 , x2 , . . . , xk )
for some action profile of Γ , call an action a` ∈ A a potential best response for player i
in x if x` > 0 and
pi (a` , x−` ) > pi (am , x−` ) for all am ∈ A, (5.1)
where x−` = (x1 , . . . , x`−1 , x` − 1, x`+1 , . . . , xk ).
Fix a particular commutative image x = (x1 , x2 , . . . , xk ), and define a bipartite graph
G = (V, E) such that
V = V1 ∪ V2 , V1 = N, V2 = { (aj , `) : aj ∈ A, 1 6 ` 6 xj }, and
E = { (i, (aj , `)) : aj is a potential best response for i under x }.
5.3 · Pure Nash Equilibria 65

1 (0, 1)
{1} {(0, 1), (0, 2)}
2 (0, 2)
{2, 3} {(1, 1)}
3 (1, 1)

Figure 5.4: Matching problem for the game of Figure 5.3 (left) and representation of the
same problem by a graph with a constant number of vertices (right), as used in the proof
of Theorem 5.3.

In other words, the two sides of G respectively correspond to the players and actions of Γ ,
with action multiplicities according to x. Edges connect each player to his potential best
responses. The graph corresponding to the shaded action profiles in Figure 5.3 is shown
on the left of Figure 5.4.
It is now readily appreciated that a pure equilibrium of Γ with commutative image x
directly corresponds to a perfect matching of G, and vice versa. Furthermore, by Hall’s
Theorem, G has a perfect matching if and only if |ν(V 0 )| > |V 0 | for all V 0 ⊆ V1 , where
ν(V 0 ) = {v ∈ V2 : (u, v) ∈ E, u ∈ V1 } is the neighborhood of vertex set V 0 (e.g., Bollobás,
1998).
Observe that Hall’s condition cannot be verified efficiently in general. We will argue,
however, that this can in fact be done for G, by considering a new graph obtained from G
which possesses only a constant number of vertices. More precisely, we want to show that
Hall’s condition can be verified by a threshold circuit with unbounded fan-in, constant
depth, and a polynomial number of gates. From the description given below it is easy to
see that the constructed circuit is log-space uniform.
Assume without loss of generality that for all v ∈ V1 , ν(v) 6= ∅, and define an
equivalence relation ∼ ⊆ V × V such that for all v, v 0 ∈ V, v ∼ v 0 if and only if
ν(v) = ν(v 0 ). By construction of G, and since both the number of actions and the
number of possible subsets of actions are constant, the set V/∼ of equivalence classes
has constant size, and V/∼ = (V1 /∼) ∪ (V2 /∼). Each element of V1 /∼ corresponds to
the set of players having a particular set of actions as their potential best responses
in x. Each element of V2 /∼ corresponds to an action in A. The neighborhood func-
tion ν can naturally be extended to equivalence classes by letting for each U ∈ V1 /∼,
ν(U) = { U 0 ∈ V2 /∼ : v ∈ ν(u) for some u ∈ U, v ∈ U 0 }. This yields a bipartite graph
with vertex set V/∼, the graph corresponding to the game in Figure 5.3 is shown on the
right of Figure 5.4. It is now easily verified that G has a perfect matching, and Γ a pure
P
equilibrium, if and only if for every Y ⊆ V1 /∼, X∈Y |ν(X)| > |Y|.
Since V1 /∼ has only a constant number of subsets, we can construct a constant depth
threshold circuit which uses sub-circuits UNARY-COUNT and UNARY-TO-BINARY as
described by Chandra et al. (1984) to sum over elements of the equivalence classes, and
COMPARISON sub-circuits to verify the inequalities. The former is easily realized with
66 5 · Anonymous Games

#(1, s) 0 ... `+1 ... m+2

p0 (s) ... 1 0 1 0 2 1 0 1 0 1 ...

#(1, s) 0 ... `+1 ... m+2

p1 (s) ... 1 0 1 0 1 2 0 1 0 1 ...

Figure 5.5: Game Γ used in the proof of Theorem 5.3

the help of MAJORITY gates. It thus remains to be shown that for any X ∈ V1 /∼, |X|
and |ν(X)| can be computed in TC0 . For this, recall that a particular element of V1 /∼
corresponds to the set of players that have a certain set of actions as their set of best
responses in x. To compute the number of such players we first construct a circuit of
constant depth that uses COMPARISON sub-circuits to check whether (5.1) is satisfied
for a fixed commutative image x, a particular player i ∈ N, and a particular action a ∈ A.
To check whether C ⊆ A is the set of best responses for player i under x, we simply combine
the outputs of the above circuits for all actions a ∈ A into a single AND gate, negating the
outputs of circuits for actions a 6∈ C. The desired number of players is then obtained by
adding up the outputs of these gates for all players i ∈ N, again using UNARY-COUNT
sub-circuits. On the other hand, |ν(X)| corresponds to the number of players bound to
play an action from a certain subset in every action profile with commutative image x,
and can easily be obtained by summing over the respective elements of x.
For hardness, we reduce the problem of deciding whether exactly ` bits of a string of m
bits are 1 to that of deciding the existence of a pure equilibrium in a self-anonymous game.
Hardness of the former problem is immediate from that of MAJORITY (e.g., Chandra et al.,
1984). For a particular m-bit string b, define a game Γ with m + 2 players of two different
types 0 and 1 and actions A = {0, 1}. The ith player of Γ is of type 0 or 1 if the ith bit
of b is 0 or 1, respectively. Player m + 1 is of type 0, player m + 2 is of type 1. The
payoffs p0 and p1 for the two types are given in Figure 5.5, the column labeled j specifies
the payoff when exactly j players, including the player himself, play action 1. It is easily
verified that this yields a constant-depth reduction.
We now claim that Γ possesses a pure equilibrium if and only if exactly ` bits of b are
1, and observe the following:

 An action profile aN cannot be an equilibrium of Γ if #(1, aN ) 6= ` + 1. In this

case, the players of one of the two types get a higher payoff at both #(1, aN ) − 1
and #(1, aN ) + 1, or at one of these in case #(1, aN ) = 0 or #(1, aN ) = m + 2. By
construction there exists at least one player of each type, so there always is a player
who can change his action to get a higher payoff.

 If there are ` + 1 players of type 1, the action profile where all players of type 0
5.3 · Pure Nash Equilibria 67

play action 0 and all players of type 1 play action 1 is an equilibrium. None of the
players of type 0 can gain by changing his action to 1, and none of them can change
his action to 0, because all of them already play 0. A symmetric condition holds for
players of type 1.

 In turn, if the number of players of type 1 does not equal ` + 1, an action profile aN
with #(1, aN ) = `+1 cannot be an equilibrium. In this case, there exists (i) a player
of type 0 playing action 1 in aN , or (ii) a player of type 1 playing 0. This player
can change his action to get a higher payoff.

Hence, a pure equilibrium exists if and only if there are ` + 1 players of type 1, i.e., if and
only if b has ` 1-bits.

In contrast to anonymous games, if aN is a pure equilibrium of a symmetric game, so

0 satisfying #(a 0 ) = #(a ). This is due to the fact that the payoff functions of
are all aN N N
all players, and thus the situation of all players playing the same action a ∈ A, is identical,
as would be the situation of any other player exchanging actions with someone playing a.
We exploit this property to show that deciding the existence of a pure equilibrium in
symmetric games with a constant number of actions is strictly easier than for anonymous
and self-anonymous games.

Theorem 5.4. The problem of deciding whether a symmetric game with a constant
number of actions has a pure Nash equilibrium is in AC0 .

Proof. Like with anonymous games, an action profile aN is an equilibrium of a symmetric

game if and only if, for all i ∈ N, ai is a best response to #(a−i ), i.e., if

pi (ai , #(a−i )) > pi (ai0 , #(a−i )) for all ai0 ∈ A. (5.2)

For a particular player i ∈ N and for constant k, checking this inequality requires only a
constant number of comparisons and can be done using a circuit of constant depth and
polynomial size (e.g., Chandra et al., 1984). When it comes to checking (5.2) for the
different players, the observation about action profiles with identical commutative images
affords us a considerable computational advantage as compared to, say, anonymous or
self-anonymous games. More precisely, we only have to check if (5.2) is satisfied for a
player playing a certain action, of which there are at most k. Again, this can be done
using a circuit of constant depth and polynomial size if k is a constant.
Finally, to decide whether game Γ has a pure equilibrium, we have to verify (5.2) for the
different values of #(aN ) for aN ∈ AN . If k is constant, there are only polynomially many
of these, so the complete check requires only polynomial size and constant depth.

The reasoning in the proof of Theorem 5.4 also provides a nice illustration of the fact
that every symmetric game with two actions possesses a pure equilibrium, as shown by
Cheng et al. (2004). In the case of two actions, pi depends only on player i’s action, 0
68 5 · Anonymous Games

or 1, and on the number of other players playing action 1. A pure equilibrium exists if
for some m neither the players playing 0, who see m players playing 1, nor the players
playing 1, who see m − 1 other players playing 1, have an incentive to deviate, i.e.,
pi (0, m) > pi (1, m) and pi (1, m − 1) > pi (0, m − 1). For m = 0 and m = n, one of the
conditions is trivially satisfied, because there are no players playing 1 or 0, respectively.
It is now straightforward to show that at least one m satisfying both conditions must
exist. Alternatively, the existence of pure equilibria in symmetric games with two actions
can also be obtained as an immediate consequence of Lemma 5.2: we can transform every
symmetric game with two actions into a self-symmetric game with the same set of pure
equilibria, and every game in the latter class is guaranteed to have at least one pure
equilibrium.
As stated earlier, self-symmetric games always possess a pure equilibrium, namely an
action profile with maximum payoff for every player. We proceed to show that such an
action profile, with the additional property that it maximizes social welfare, can be found
in AC0 .

Theorem 5.5. The problem of finding a social-welfare-maximizing pure Nash equi-

librium of a self-symmetric game with a constant number of actions is in AC0 .

Proof. Since self-symmetric games belong to the class of common payoff games, any action
profile with maximum payoff for one player automatically is a social-welfare-maximizing
equilibrium, and Pareto dominates any other strategy profile. Finding such an equilibrium
is in turn equivalent to finding the maximum of n+k−2


k−1 integers. The exact number is
irrelevant as long as it is polynomial in the size of the input, which is certainly the case if
k is bounded by a constant. Chandra et al. (1984) have shown that the maximum of m
binary numbers of m bits each can be computed by an unbounded fan-in, constant-depth
Boolean circuit of size polynomial in m. Since m is of course polynomial in the size of
the input, the size of this circuit is as well.

5.3.2 Games with a Growing Number of Actions

The proofs we have seen so far in this chapter could exploit the fact that for constant
k the explicit representation of an anonymous game is equivalent, under the appropriate
type of reduction, to any kind of payoff function computable in a particular complexity
class inside P. This need no longer be the case if k is unbounded, because then the size
of the explicit representation grows exponentially in n. Such games may of course still
admit a polynomial-size representation, for example if payoff functions are encoded by
a Boolean circuit. We will now show that deciding the existence of a pure equilibrium
in anonymous, symmetric, and self-anonymous games becomes NP-hard if the number
of actions grows in n. For self-symmetric games, which always have a pure equilibrium,
the associated search problem will be shown to be PLS-hard. In particular, we show
NP-completeness and PLS-completeness, respectively, for games that have a polynomial
5.3 · Pure Nash Equilibria 69

number of players—like those covered by Theorems 5.3 and 5.4—and a number of actions
that grows linearly in the number of players. It will be obvious from the proofs that
hardness for the respective classes also holds for games with an exponential number of
players and logarithmic growth of the number of actions. The corresponding case with a
constant number of actions, on the other hand, remains open.
If the number of actions in a game is large enough, they can in principle be used to
distinguish the players playing them. We will exploit this fact and prove the following
theorems by reductions from satisfiability of a Boolean circuit. While as a matter of fact
we establish hardness via a particular encoding of a game, they nevertheless provide inter-
esting insights into the influence of restricted classes of payoff functions on the complexity
of solving a game. After all it is far from obvious that hardness results for general games
extend to anonymous and symmetric games.
Recall that circuit satisfiability (CSAT), i.e., the problem of deciding whether a Boolean
circuit has a satisfying assignment, is NP-complete (e.g., Papadimitriou, 1994a). We pro-
vide a reduction from CSAT to the problem of deciding the existence of a pure equilibrium
in a special class of games. For a particular circuit C with inputs M = {1, 2, . . . , m}, we
define a game Γ with at least m players and actions A = { aj0 , aj1 : j ∈ M }∪{b}. An action
profile aN of Γ where #(aj0 , aN ) + #(aj1 , aN ) = 1 for all j ∈ M, i.e., one where exactly
one action of each pair aj0 , aj1 is played, directly corresponds to an assignment c of C, the
jth bit of this assignment being 1 if and only if aj1 is played. Observe that in this case the
auxiliary action b has to be played by exactly n − m players. We can thus distinguish the
action profiles of Γ corresponding to a satisfying assignment of C from those corresponding
to a non-satisfying assignment and those not corresponding to an assignment at all.

Theorem 5.6. Deciding whether a self-anonymous game has a pure Nash equilibrium
is NP-complete, even if the number of actions is linear in the number of players and
there is only a constant number of different payoffs.

Proof. Membership in NP is straightforward. Since the number of players is polynomial,

we can simply guess an action profile and verify that it satisfies the equilibrium condition.
For hardness, we reduce satisfiability of a Boolean circuit C with a set M = {1, 2, . . . , m}
of inputs to the existence of a pure equilibrium in a game Γ with n > m players, actions
A = { aj0 , aj1 : j ∈ M } ∪ {b}, and payoff functions pi as follows:

 If action profile aN corresponds to a satisfying assignment of C, we let pi (aN ) = 1

for all i ∈ N.

 Otherwise we let

– p1 (aN ) = 1 and p2 (aN ) = 0 if #(b, aN ) is even,

– p1 (aN ) = 0 and p2 (aN ) = 1 if #(b, aN ) is odd, and
– pi (aN ) = 1 for all i ∈ N \ {1, 2}.
70 5 · Anonymous Games

We observe the following:

 In all of the above cases, the payoff of player i only depends on the number of
players playing certain actions. If two players exchange actions, the payoff to all
players remains the same. Hence, Γ is self-anonymous.

 Clearly, every action profile aN corresponding to a satisfying assignment of C is an

equilibrium, because in this case all players receive the maximum payoff of 1.

 For an action profile aN not corresponding to a satisfying assignment of C, either

player 1 or player 2 receives a payoff of 0. Furthermore, by choosing his own action to
be either b or some other action, this player can determine the parity of the number
of players playing b. Changing the parity strictly increases the player’s payoff. This
means that aN cannot be an equilibrium.

We have hence established a direct correspondence between satisfying assignments of C

and pure equilibria of Γ . The transformation from C to Γ essentially works by writing down
Boolean circuits that compute pi . Observing that this can be done in time polynomial in
the size of C if n 6 2k completes the proof.

As the reader may have noticed, the construction used in this proof has two distin-
guished players play Matching Pennies for any assignment that does not satisfy the
Boolean circuit. Not only is this game a well-known example for a game that does not
possess a pure equilibrium, it is also self-anonymous on its own. On the other hand, it
is easily verified that the payoffs in this game do depend on the identities of the players,
i.e., that the game is not symmetric. We will have to avail of a different construction for
the symmetric case.

Theorem 5.7. Deciding whether a symmetric game has a pure Nash equilibrium is
NP-complete, even if the number of actions is linear in the number of players and
there is only a constant number of different payoffs.

Proof. Membership in NP is again straightforward.

For hardness, we provide a reduction from CSAT, mapping a circuit C with inputs
M = {1, 2, . . . , m} to a game Γ with n > m players, actions A = { aj0 , aj1 : j ∈ M } ∪ {b},
and payoff functions pi as follows:

 If #(b, aN ) = n − m, we let

– pi (aN ) = 2 if aN corresponds to a satisfying assignment or if ai = aj1 for some

j ∈ M, #(aj0 , aN ) > 0, and #(aj1 , aN ) > 0,
– pi (aN ) = 1 if ai = aj0 for some j ∈ M, #(aj0 , aN ) > 0, and #(aj1 , aN ) = 0,
and
– pi (aN ) = 0 otherwise.
5.3 · Pure Nash Equilibria 71

 If #(b, aN ) < n − m, we let pi (aN ) = 1 if ai = b, and pi (aN ) = 0 otherwise.

 If #(b, aN ) > n − m, we let pi (aN ) = 0 if ai = b, and pi (aN ) = 1 otherwise.

We observe the following:

 For all of the above cases, the payoff of player i only depends on his own action
and on the number of players playing certain other actions. If two players exchange
actions, their payoffs are also exchanged. Hence, Γ is symmetric.

 Clearly, any action profile corresponding to a satisfying assignment of C is an equi-

librium, because in this case all players receive the maximum payoff of 2.

 If on the other hand aN does not correspond to a satisfying assignment, we have

one of three different cases, in none of which aN is an equilibrium:

– If #(b, aN ) < n − m or #(b, aN ) > n − m + 1, then there exists a player that

receives payoff 0 and can change his action to receive payoff 1.
– If #(b, aN ) = n − m and #(aj0 , aN ) = 1 for all j ∈ M, player i can change to
any a`1 such that ai 6= a`0 to increase his payoff from 1 to 2.
– Otherwise, there has to be some player i ∈ N who gets payoff 0, and, by the
pigeonhole principle, some j ∈ M such that #(aj0 , a−i ) = #(aj1 , a−i ) = 0.
Then, player i can change to aj0 to get a higher payoff.

Again, there is a direct correspondence between pure equilibria of Γ and satisfying assign-
ments of C. The transformation from C to Γ essentially works by writing down Boolean
circuits that compute pi . Observing that this can be done in time polynomial in the size
of C if n 6 2k completes the proof.

By each of the previous two theorems and by the inclusion relationships between the
different classes of games, we also have the following.

Corollary 5.8. Deciding whether an anonymous game has a pure Nash equilibrium
is NP-complete, even if the number of actions is linear in the number of players and
there is only a constant number of different payoffs.

Since the proofs of Theorems 5.6 and 5.7 work by mapping satisfying assignments of a
Boolean circuit to a certain number of pure equilibria of a strategic game, we can show
that counting the number of pure equilibria in the above classes of games is hard.

Corollary 5.9. For anonymous, symmetric, and self-anonymous games, counting

the number of pure Nash equilibria is #P-hard, even if the number of actions is
linear in the number of players and there is only a constant number of different
payoffs.
72 5 · Anonymous Games

Proof. Recall that in the proof of Theorem 5.6, actions of the game Γ are identified with
inputs of the Boolean circuit C. As a direct consequence of anonymity or symmetry, it does
not matter which player plays a particular action to assign a value to the corresponding
gate. Every satisfying assignment of C thus corresponds to n! equilibria of Γ , so the
number of satisfying assignments can be determined by counting the number of pure
equilibria, of which there are at most 2n n!, and dividing this number by n!. Division of
two m-bit binary numbers can be done using a circuit with bounded fan-in and depth
O(log m) (Beame et al., 1986). For m = log(2n n!) = O(n2 ), we have depth O(log n2 ) =
O(log n). We have thus found a reduction of the problem #SAT of counting the number
of satisfying assignments of C, which is #P-complete (e.g., Papadimitriou, 1994a), to the
problem of counting the pure equilibria of Γ . The same line of reasoning applies to the
proof of Theorem 5.7. Analogously to Corollary 5.8, #P-hardness extends to anonymous
games.

As we have already outlined above, every self-symmetric game possesses a pure equi-
librium. Theorem 5.5 states that finding even a social-welfare-maximizing equilibrium is
very easy as long as the number of actions is bounded by a constant. If now the number
of actions is growing but polynomial in the size of the input, we can start at an arbitrary
action profile and check in polynomial time whether some player can change his action
to increase the (common) payoff. If this is not the case, we have found an equilibrium.
Otherwise, we can repeat the process for the new profile, resulting in a procedure called
best-response dynamics in game theory. Since the payoff strictly increases in each step,
we are guaranteed to find an equilibrium in polynomial time if the number of different
payoffs is polynomial. Conversely we will show that, given a self-symmetric game with a
growing number of actions and an exponential number of different payoffs, finding a pure
equilibrium is PLS-complete, i.e., at least as hard as finding a locally optimal solution
to an NP-hard optimization problem. The proof of the following theorem works along
similar lines as those of Theorems 5.6 and 5.7 to give a reduction from the PLS-complete
problem FLIP.
Theorem 5.10. The problem of finding a pure Nash equilibrium in a self-symmetric
game is PLS-complete, even if the number of actions is linear in the number of
players.
Proof. Neighborhood among action profiles is given by a single player changing his action.
Since the number of players and actions is polynomial in the input size, and since the payoff
function is computable in polynomial time, membership in PLS is immediate.
For hardness, consider a Boolean circuit C with inputs M = {1, 2, . . . , m} and ` outputs.
Finding an assignment such that the output interpreted as an `-bit binary number is a
local maximum under the FLIP neighborhood, where neighbors are obtained by changing
a single input bit, is known to be PLS-complete (Johnson et al., 1988, Schäffer and Yan-
nakakis, 1991). We provide a PLS reduction to the problem of finding a pure equilibrium
in a self-symmetric game by mapping a particular circuit C as described above to a game
5.3 · Pure Nash Equilibria 73

Γ with n > m players, actions A = { aj0 , aj1 : j ∈ M }, and a (common) payoff function p
as follows:

 If action profile aN corresponds to an assignment c of C, let p(aN ) = n + C(c),

where C(c) denotes the output of C for input c, interpreted as a binary number.

 Otherwise let p(aN ) = min(#(b, aN ), n − m) + |{ j ∈ M : #(aj0 , aN ) + #(aj1 , aN ) >

0 }|. That is, the payoff is at most n − 1 and decreases in the minimum number of
players that would have to change their action in order to make aN correspond to
an assignment of C.

We observe the following:

 Obviously, Γ is a common payoff game. Since p is invariant under any permutation

of the players in both of the above cases, Γ is self-symmetric.

 If n 6 2k , a Boolean circuit that computes p can be constructed from C in time poly-

nomial in the size of C. Hence, there exists a polynomial time computable function
that maps instances of FLIP to instances of the problem under consideration.

 An action profile aN that does not correspond to an assignment of C cannot be an

equilibrium of Γ . In this case, either #(aj0 , aN ) + #(aj1 , aN ) = 0 for some j ∈ M,
or #(aj0 , aN ) + #(aj1 , aN ) > 1 for some j ∈ M and #(b, aN ) < n − m. Then there
exists a player who can increase his payoff (the payoff of all players, actually) by
changing his action, to aj0 or aj1 in the former case and to b in the latter.

 There is a direct correspondence between the FLIP neighborhood of C and a single

player changing between aj0 and aj1 for some j ∈ M. Furthermore, changing to
an action profile that does not correspond to an assignment of C will get the player
strictly less payoff. Thus, there is a direct correspondence between pure equilibria of
Γ and local maxima of C under the FLIP neighborhood. Obviously, the assignment
corresponding to an action profile can be computed in polynomial time, if such an
assignment exists. The conditions of Definition 2.16 do not require that we map
solutions of Γ that are not locally optimal to solutions of C that are not locally
optimal. This means that action profiles not corresponding to an assignment can
simply be mapped to an arbitrary assignment.

It is easily verified that this satisfies the properties of a PLS reduction.

With some extra work, we can show that the reduction used in the proof of Theo-
rem 5.10 is tight, and draw additional conclusions about the standard algorithm and the
standard algorithm problem.

Corollary 5.11. The running time of the standard algorithm for finding pure Nash
equilibria in self-symmetric games is exponential in the worst case. The standard
algorithm problem is NP-hard.
74 5 · Anonymous Games

Proof. Johnson et al. (1988) have shown that the standard algorithm for FLIP has an
exponential worst-case running time, and the standard algorithm problem is NP-hard. By
Lemma 3.3 of Schäffer and Yannakakis (1991) it thus suffices to show that the reduction in
the proof of Theorem 5.10 is tight. To this end, choose R to be the set of action profiles of
Γ that correspond to an assignment of C. Obviously, R contains all optimal solutions, and
a payoff profile corresponding to a particular assignment can be computed in polynomial
time. The third condition is trivially satisfied because the measure of any solution inside
R is strictly greater than that of any solution outside of R.

By a slight modification of the proof of Theorem 5.10, PLS-hardness, exponential

worst-case running time of the standard algorithm, and NP-hardness of the standard
algorithm problem can also be shown for general, i.e., not necessarily anonymous, common
payoff games with k = 2. This fact nicely illustrates the influence of anonymity on the
complexity of the pure equilibrium problem.

5.3.3 Threshold Anonymity

We will now extend the tractability results of Section 5.3.1 to games where the players
cannot even observe the exact number of players playing a certain action, but only whether
this number exceeds certain thresholds. Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a normal-form
game, and A a set of actions such that Ai = A for all i ∈ N. For T ⊆ {1, 2, . . . , n},
let ∼T ⊆ AN × AN be defined as follows: aN ∼T aN 0 if for all a ∈ A and all x ∈ T ,

#(a, aN ) < x if and only if #(a, aN 0 ) < x. The relation ∼ naturally extends to action
T
profiles for subsets of N. It is then easily verified that for any T ⊆ {1, 2, . . . , n}, ∼T is an
equivalence relation on the set AM for any M ⊆ N. We use ∼T to generalize Definition 5.1.
Definition 5.12 (threshold anonymity). Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a normal-form
game, A a set of actions such that Ai = A for all i ∈ N. Let T ⊆ {1, 2, . . . , n}. Γ is called
 T -anonymous if pi (aN ) = pi (aN
0 ) for all i ∈ N and all a , a 0 ∈ AN with a = a 0
N N i i
and a−i ∼T a−i
0 ,

 T -symmetric if pi (aN ) = pj (aN

0 ) for all i, j ∈ N and all a , a 0 ∈ AN with a = a 0
N N i j
and a−i ∼T a−j
0 ,

 T -self-anonymous if pi (aN ) = pi (aN

0 ) for all i ∈ N and all a , a 0 ∈ AN with
N N
aN ∼T aN , and
0

 T -self-symmetric if pi (aN ) = pj (aN

0 ) for all i, j ∈ N and all a , a 0 ∈ AN with
N N
aN ∼ T aN .
0

For T = {1, 2, . . . , n}, these classes are equivalent to those of Definition 5.1. Moreover,
for T = {1}, we obtain Boolean anonymity, where payoffs only depend on the set of actions
that are played by at least one player. In general, we call a game threshold anonymous,
for one of the above classes, if it is T -anonymous for some T and the corresponding class.
5.3 · Pure Nash Equilibria 75

Obviously, the number of payoffs that need to be written down for each player to
specify a general T -anonymous game is exactly the number of equivalence classes of ∼T
for action profiles of the other players. A T -anonymous game can be represented using at
most n · k · |An−1 /∼T | numbers, where X/∼ denotes the quotient set of set X by equivalence
relation ∼. For Boolean anonymity, the number of equivalence classes equals the number
of k-bit binary numbers where at least one bit is 1, i.e., 2k − 1. More generally, there
cannot be more than (|T | + 1)k equivalence classes if |T | is bounded by a constant, since for
every action the number of players playing this action must be between two thresholds.
For T = {n} there are as few as k + 1. Hence, any T -anonymous game with constant
|T | is representable using space polynomial in n if k = O(log n). It does not matter
if the thresholds themselves are constant or not. We are now ready to identify a class
of threshold anonymous games for which the pure equilibrium problem is tractable. It
should be noted that the proof technique is not limited to this particular class, but in fact
applies to the larger class of games for which the kernel of the best response function has
polynomial size.

Theorem 5.13. For threshold anonymous games with k = O(log n) and a constant
number of thresholds, deciding the existence of a pure Nash equilibrium is in P.

Proof. Like in the proof of Theorem 5.3, we provide an algorithm that checks whether
there is an equilibrium in a particular equivalence class X ∈ AN /∼T . Since for k = O(log n)
and |T | = O(1), the cardinality of AN /∼T is polynomial in n, it suffices to show that the
algorithm requires only polynomial time for every such set. For a particular element
X ∈ AN /∼T , the algorithm is again divided into two phases: (i) computing the set of best
responses for each player under X, and (ii) checking whether there is a particular action
profile aN ∈ X where each player plays a best response.
In the first phase, and unlike the case T = {1, 2, . . . , n} covered by Theorem 5.3, the
action a played by player i ∈ N may or may not yield a different element of AN\{i} /∼T
against which a should be a best response. Instead of just looking for best responses under
elements of T N , we thus look for best responses under those of UN , where U = { u 6 n :
u ∈ T or (u − 1) ∈ T }. Since the cardinalities of both UN and of the set of possible best
responses is polynomial if |T | = O(1) and k = O(log n), the first phase requires only
polynomial time.
As for the second phase, we show that it can be reduced to deciding the existence of
an integer flow with upper and lower bounds in a directed network with O(2k ) vertices.
Since this problem is in P if the number of vertices is polynomial (e.g., Garey and Johnson,
1979), observing that 2k is polynomial in the size of the input if k = O(log n) completes
the proof. Fix X ∈ AN /∼T and define a directed graph G = (V, E) such that

V = {s, t, t 0 } ∪ V1 ∪ V2 , V1 = 2A , V2 = A, and
E = {s} × V1 ∪ { (A 0 , a) ∈ V1 × V2 : a ∈ A 0 } ∪ V2 × {t} ∪ {(t, t 0 )}.
76 5 · Anonymous Games

{0}
(0, 3)
(2, 2)
0
(0, 3) (2, 2)
s (1, 1) {0, 1} t (3, 3) t0

(0, 3) 1 (1, 1)
(0, 0)
{1} (0, 3)

Figure 5.6: Integer flow network used in the proof of Theorem 5.13, example for the game
of Figure 5.3. Edge e is labeled (u(e), `(e)).

Further define two functions ` : E → N and u : E → N such that,

for all A 0 ∈ V1 ,
`(s, A 0 ) = u(s, A 0 ) = |{ i ∈ N : A 0 is the set of best responses of i under X }|,
for all A 0 ∈ V1 and a ∈ A 0 , `(A 0 , a) = 0 and u(A 0 , a) = n,
for all a ∈ V2 , `(a, t) = min #(a, x) and u(a, t) = max #(a, x), and
x∈X x∈X
`(t, t 0 ) = u(t, t 0 ) = n.

Figure 5.6 shows the flow network for the game in Figure 5.3. Edge capacities have
been computed by checking for each player if his action in the respective (shaded) action
profile of Figure 5.3 is a best response. Observe that since this game is not only threshold
anonymous but also anonymous, upper and lower bounds coincide for flow leaving vertices
in V2 = A.
Obviously every feasible flow from s to t 0 must have size n. Furthermore, the condi-
tions for flow leaving vertices in V1 require that there exists an assignment of actions to
players such that each player plays a best response, while those for flow leaving edges in V2
require that the resulting action profile is an element of X. It is thus readily appreciated
that a flow from s to t 0 satisfying lower bounds ` and upper bounds u directly corresponds
to a Nash equilibrium of Γ , and vice versa.

On the other hand, it is fairly straightforward to modify the games defined in the
proofs of Theorems 5.6, 5.7, and 5.10 to obey Boolean thresholds if n = k. We obtain the
following corollary.
Corollary 5.14. Deciding the existence of a pure Nash equilibrium is NP-hard for
threshold anonymous, threshold symmetric, and threshold self-anonymous games,
even if thresholds are Boolean, the number of actions is linear in the number of
players, and there is only a constant number of different payoffs. For the same
classes, counting the number of pure Nash equilibria is #P-hard.
5.4 · Iterated Weak Dominance 77

For threshold self-symmetric games, finding a pure Nash equilibrium is PLS-hard,

even if thresholds are Boolean and the number of actions is linear in the number of
players.

Proof. In all constructions, we assume n = m and remove the auxiliary action b. In

addition to that, the self-anonymous game used in the proof of Theorem 5.6 is modified
by letting players 1 and 2 play Matching Pennies on the parity of the number |{ j ∈ M :
#(aj0 , aN ) > 0 }| of 0-actions that are played by at least one player. It is easily verified
that the arguments used to show the correspondence between satisfying assignments of the
Boolean circuit and pure equilibria of the respective game still go through. Furthermore,
the payoff of a particular player in each of these games only depends on whether certain
actions are played by at least one player and, potentially, on the player’s own action.

5.4 Iterated Weak Dominance

We now turn to iterated weak dominance, and begin by looking at games with a growing
number of actions. Intuitively, a large number of actions nullifies the computational
advantage obtained from anonymity by allowing for a distinction of the players by means
of the actions they play. This intuition was confirmed in Section 5.3.2, where we saw that
the pure equilibrium problem becomes NP-hard or PLS-hard if the number of actions is
sufficiently large compared to the number of players. We derive a similar result for iterated
dominance solvability (IDS) and eliminability (IDE).

Theorem 5.15. IDS and IDE are NP-hard for all four classes of anonymous games,
even if the number of actions grows logarithmically in the number of players, if
only dominance by pure strategies is considered, and if there are only two different
payoffs.

Proof. We provide a reduction from CSAT to IDS and IDE for self-symmetric games.
Hardness for the other types of anonymity follows by inclusion. For a particular Boolean
circuit C with inputs M = {1, 2, . . . , m}, we define a game Γ with n > m players and
actions A = { aj0 , aj1 : j ∈ M } ∪ {a0 , a1 }. An action profile aN of Γ where #(aj0 , aN ) +
#(aj1 , aN ) = 1 for all j ∈ M, i.e., one where exactly one action of each pair aj0 , aj1 is
played, directly corresponds to an assignment c of C, the jth bit cj of this assignment
being 1 if and only if aj1 is played. Observe that in this case the auxiliary actions a0
and a1 have to be played by exactly n−m players. We can thus also identify action profiles
of Γ that correspond to a satisfying assignment of C. Now define the (common) payoff
function p by letting p(aN ) = 1 if #(a0 , aN ) + #(a1 , aN ) > n − m, or if aN corresponds
to a satisfying assignment of C and #(a1 , aN ) = n − m. Otherwise, let p(aN ) = 0. Since
the payoff function is the same for all players, and the payoff only depends on the number
of players playing each of the different actions, Γ is self-symmetric. We will further argue
that for any A 0 ⊆ A with a1 ∈ A 0 , a1 dominates every action a ∈ A 0 \ {a0 , a1 } in the
78 5 · Anonymous Games

restriction of Γ to action set A 0 , and a1 dominates a0 in such a game if and only if C

has a satisfying assignment. These properties clearly imply that a0 is eliminable for any
player, and Γ is solvable via iterated dominance with action a1 remaining for each player,
if and only if C has a satisfying assignment. Since there are only two different payoffs, we
can restrict our attention to dominance by pure strategies (e.g., Conitzer and Sandholm,
2005a).
To see the former property, consider an action profile aN corresponding to a satisfying
assignment of C, and a player i ∈ N such that ai ∈ {a0 , a1 }. Then, 1 = p(a−i , a1 ) >
p(a−i , a 0 ) = 0 for any a 0 ∈ / {a0 , a1 }. On the other hand, consider an action profile aN
not corresponding to a satisfying assignment, and some player i ∈ N. Then, for any
a0 ∈/ {a0 , a1 }, 1 = p(a−i , a1 ) > p(a−i , a 0 ) = 0 if #(a0 , a−i ) + #(a1 , a−i ) = n − m, and
p(a−i , a1 ) = p(a−i , a 0 ) otherwise.
For the latter property, first consider an action profile aN not corresponding to a satis-
fying assignment of C, and some player i ∈ N such that ai ∈ {a0 , a1 }. Then, p(a−i , a1 ) =
p(a−i , a0 ). On the other hand, consider an action profile aN corresponding to a satisfying
assignment, and some player i ∈ N such that #(a0 , a−i ) + #(a1 , a−i ) = n − m − 1. If
#(a1 , a−i ) < n − m − 1, then p(a−i , a0 ) = p(a−i , a1 ) = 0. If #(a1 , a−i ) = n − m − 1,
then 1 = p(a−i , a1 ) > p(a−i , a0 ) = 0.
The transformation from C to Γ essentially works by writing down a Boolean circuit
that computes p. Observing that this can be done in time polynomial in the size of C if
n 6 2k completes the proof.

In the case of symmetric games, iterated dominance becomes tractable when the num-
ber of actions is bounded by a constant.

Theorem 5.16. For symmetric games with a constant number of actions, IDS and
IDE can be decided in polynomial time.

Proof. Since all players have identical payoff functions, a state of iterated dominance
elimination can be represented as a vector that counts, for each set C ⊆ A, the number of
players that have eliminated exactly the actions in C. This vector has constant dimension if
the number of actions is constant. The value of each entry is bounded by n, so the number
of different vectors is polynomial in n and thus in the size of the game. The elimination
process can then be described as a graph that has the above vectors as vertices and a
directed edge between two such vectors if the second one can be obtained from the first
by adding 1 to some component, and if the action corresponding to this component can
indeed be eliminated in the state described by the first vector. For dominance by mixed
strategies, this neighborhood relation can be computed in polynomial time via linear
programming (Conitzer and Sandholm, 2005a). This reduces the computational problems
related to iterated dominance to reachability problems in a directed graph, which in turn
can be decided in nondeterministic logarithmic space and thus in polynomial time. For
IDS, we need to find a directed path from (the vertex corresponding to) the zero vector
5.4 · Iterated Weak Dominance 79

to some vector with sum n(k − 1). For IDE, we need to find a path where the respective
action is deleted while traversing the final edge.

In light of these two results, only one interesting class remains, namely anonymous
games with a constant number of actions. To gain a better understanding of the problem,
we restrict ourselves further to games with two actions. It turns out that in this case
iterated dominance can be reformulated as a natural elimination problem on matrices.
The latter problem will be the topic of the following section.

5.4.1 A Matrix Elimination Problem

Consider an m × n matrix X with entries from the natural numbers. Call a column c
of X increasing for an interval I over the rows of X if the entries in c are monotonically
increasing in I, with a strict increase somewhere in this interval. Analogously, call c
decreasing for I if its entries are monotonically decreasing in I, with a strict decrease
somewhere in this interval. We then say that c is active for I if it is either increasing or
decreasing for this interval. Now consider a process that starts with X and successively
eliminates pairs of a row and a column. Rows will only be eliminated from the top or
bottom, such that the remaining rows always form an interval over the rows of X. A
column will only be eliminated if it is active for the remaining rows. Elimination of an
increasing column is accompanied by elimination of the top row. Similarly, a decreasing
column and the bottom row are eliminated at the same time. The process ends when no
active columns remain. In this section we study two computational problems. Matrix
elimination asks whether for a given matrix there exists a sequence of such eliminations
of length min(m − 1, n), i.e., one that eliminates all columns of the matrix or all rows
but one, depending on the dimensions of the original matrix. Eliminability of a column
asks whether a particular column can be eliminated at some point during the elimination
process.
More formally, the matrix elimination process can be described by a pair of sequences
of equal length, where the first sequence consists of column indices of X and the second
sequence of elements of {0, 1}, corresponding to elimination of the top or bottom row,
respectively. The first sequence will contain every column index at most once. The ith
element of the second sequence will be 0 or 1, respectively, if the column corresponding to
the ith element of the first sequence is increasing or decreasing in the interval described
by the number of 0s and 1s in the second sequence up to element i − 1.
Consider for example the sequence of matrices shown in Figure 5.7, obtained by start-
ing with the 5 × 4 matrix on the left and successively eliminating columns b, a, c, and d.
In this particular example, the process ends when all rows and columns of the matrix
have been eliminated. Of course, this does not always have to be the case. Again consider
the matrix on the left of Figure 5.7, with all entries in the second row from the bottom
replaced by 2. It is easy to see that in this case no column will be active after the first
elimination step, and elimination cannot continue. Since column b was the only active
80 5 · Anonymous Games

a b c d a b c d a b c d a b c d
0 1 3 2 1 1 2 1 2 1
1 0 2 2 1 0 2 1 2 1 1
2 0 2 3 0 0 3 0 3 0 0
3 0 2 3 0 0 3 0
4 3 2 3 0

Figure 5.7: A matrix and a sequence of eliminations

column in the first place, eliminating just this one column is in fact all that can be done.
A related phenomenon can be observed if we instead replace the top entry in the leftmost
column by 0, and take a closer look at the matrix obtained after one elimination. While
we could continue eliminating at this point, it is already obvious that we will not obtain a
sequence of length 4. The reason is that one of the columns not eliminated so far, namely
the leftmost one, contains the same value in every row. This column cannot become active
anymore, and, as a consequence, will never be eliminated.
Let us define the problem more formally. For a set A, v ∈ An , and a ∈ A, denote
by #(a, v) = |{ ` 6 n : v` = a }| the commutative image of a and v, and write v...k =
(c1 , c2 , . . . , ck ) for the prefix of v of length k 6 n. Further denote [n] = {1, 2, . . . , n} and
[n]0 = {0, 1, . . . , n}.

Definition 5.17 (elimination sequence). Let X ∈ Nm×n be a matrix. A column k ∈ [n]

of X is called increasing in an interval [i, j] ⊆ [m] if the sequence xik , xi+1,k , . . . , xjk is
monotonically increasing and xik < xjk , and decreasing in [i, j] ⊆ [m] if the sequence
xik , xi+1,k , . . . , xjk is monotonically decreasing and xik > xjk .
Then, an elimination sequence of length k for X is a pair (c, r) such that c ∈ [m]k ,
r ∈ {0, 1}k , and for all i, j with 1 6 i < j 6 k, ci 6= cj and

 ri = 0 and column ci is increasing in [#(0, r...i−1 ) + 1, m − #(1, r...i−1 )], or

 ri = 1 and column ci is decreasing in [#(0, r...i−1 ) + 1, m − #(1, r...i−1 )].

A column will be called active in an interval if it is either increasing or decreasing in

this interval. What really matters are not the actual matrix entries xij , but rather the
difference between successive entries xij and xi+1,j . A more intuitive way to look at the
problem may thus be in terms of a different matrix with the number of rows reduced by
one, and entries describing the relative size of xij and xi+1,j , e.g., arrows pointing upward
and downward, respectively, depending on whether xij > xi+1,j or xij < xi+1,j , and
empty cells if xij = xi+1,j . According to this representation, a column can be deleted if it
contains at least one arrow, and if all arrows in this column point in the same direction.
The corresponding row to be deleted is the one at the base of the arrows.
5.4 · Iterated Weak Dominance 81
j k l m
m−(i+j) m−(i+j)
n+ 2 n+ 2

x11 ··· x1n 0 ··· 0 1 ··· 1

.. .. .. .. .. .. .. .. ..
. . . . . . . . .
xi1 ··· xin 0 ··· 0 1 ··· 1
.. ..
. . 1 ··· 1 0 ··· 0
n
xi1 ··· xin .. .. .. .. .. ..
l
m−(i+j)
m
. . . . . . n+ 2
xi+1,1 ··· xi+1,n
.. .. 1 ··· 1 0 ··· 0
. .
0 ··· 0 1 ··· 1
xm−j−1,1 · · · xm−j−1,n .. .. .. .. .. ..
j
m−(i+j)
k
. . . . . . n+ 2
xm−j,1 ··· xm−j,n
n .. ..
. . 0 ··· 0 1 ··· 1
xm−j,1 ··· xm−j,n 1 ··· 1 0 ··· 0
.. .. .. .. .. .. .. .. ..
. . . . . . . . .
xm1 ··· xmn 1 ··· 1 0 ··· 0

Figure 5.8: Matrix Y used in the proof of Lemma 5.18

We call matrix elimination (ME) the computational problem that asks, for a given
matrix X ∈ Nm×n , whether X has an elimination sequence of length min(m − 1, n). The
problem of eliminability of a column (CE) is given k ∈ [n] and asks whether there exists
an elimination sequence (c, r) such that for some i, ci = k. Without restrictions on m
and n, ME and CE are equivalent. We prove this statement by showing equivalence to
the problem of deciding whether there exists an elimination sequence eliminating certain
numbers of rows from the top and bottom of the matrix. Several other questions, like the
one for an elimination sequence of a certain length, are equivalent as well.

Lemma 5.18. CE and ME are equivalent under disjunctive truth-table reductions.

Proof. We provide reductions between both CE and ME and the problem of matrix elim-
ination up to an interval (IE): given a matrix X and two numbers k0 and k1 , does there
exist an elimination sequence (c, r) of X such that #(0, r) = k0 and #(1, r) = k1 ?
To reduce ME to IE, observe that X is a “yes” instance of ME if and only if X and some
interval of length max(1, m − n) form a “yes” instance of IE. Analogously, to reduce CE to
IE, X and i ∈ [n] form a “yes” instance of CE if there is an interval over the rows of X in
which column i is active and which together with X forms a “yes” instance of IE.
For a reduction from IE to either ME or CE, let X ∈ Nm×n and consider the (m +
2n) × (3n + m − (i + j)) matrix Y shown in Figure 5.8. We claim that a column with
82 5 · Anonymous Games

index greater than n, and the entire matrix, can be eliminated if and only if X has an
elimination sequence (c, r) satisfying #(0, r) = i and #(1, r) = j.
For the direction from left to right, assume that (c, r) is an elimination sequence for X
as above and define (c 0 , r 0 ) by

ck if 1 6 k 6 i + j,
ck0 =
n + k − (i + j) if i + j < k 6 m + 2n, and



rk if 1 6 k 6 i + j,
0
rk = 1 if i + j < k 6 n + d m+(i+j) e,


2
0 if n + d m+(i+j)
e < k 6 m + 2n.
2

It is easily verified that (c 0 , r 0 ) is an elimination sequence of length m + 2n for Y, i.e.,

one that eliminates Y entirely.
For the direction from left to right, consider an elimination sequence (c 0 , r 0 ) of length
m + 2n for Y. Define ` to be the smallest index k for which ck0 > n, and let I =
[#(0, r...`−1 ) + 1, m − #(1, r...`−1 )]. Clearly, ` > i + m − j. Now define a sequence c
that contains the first i elements ck0 of c 0 for which rk0 = 0, and the first j elements ck0
for which rk0 = 1, in the same order in which they appear in c 0 . Define r to be sequence
of corresponding elements of r 0 . Then, (c, r) is an elimination sequence for Y, because
the set of active columns is the same for I and [i, m − j], and also for all intervals in
between. Furthermore, c only contains columns with index at most n. Thus (c, r) is
also an elimination sequence for X, and the number of rows eliminated from the top and
bottom is exactly as required.
We finally observe that the above arguments about CE still apply to the problem of
eliminability of a column in a given direction (CED), where in addition to k ∈ [n] we are
given d ∈ {0, 1} and ask for an elimination sequence (c, r) such that for some i, ci = k and
ri = d.

When restricted to the case m > n, CE is at least as hard than ME in the sense
that the latter can be reduced to the former while there is no obvious reduction in the
other direction. In general, the case of ME where m > n appears easier than the one
where m 6 n. In the former, every column has to appear somewhere in the elimination
sequence, while in the latter the set of columns effectively needs to be partitioned into
two sets of sizes m and n − m, respectively, of columns to be deleted and columns to be
discarded right away.
It will not have gone unnoticed that elimination for a matrix X is closely related to
iterated dominance in an anonymous game with two actions 0 and 1 where the payoff of
player j when exactly i−1 players play action 1 is given by matrix entry xij . Given actions
for the other players, player j can choose between two adjacent entries of column j, so one
of his two actions is dominated by the other one if the column is increasing or decreasing,
respectively. Eliminating one of two actions effectively removes a player from the game,
5.4 · Iterated Weak Dominance 83

whereas elimination of the top or bottom row of the matrix mirrors the fact that at the
same time, the number of players who can still choose between both of their actions is
reduced by one. The following result formally establishes this relationship.
Lemma 5.19. IDS and IDE in anonymous games with two actions are equivalent under
disjunctive truth-table reductions to ME and CE, respectively, restricted to instances
with m = n + 1.
Proof. By Lemma 5.2, an anonymous game with two actions can be transformed into a
self-anonymous game while preserving dominance by pure strategies. Since by a result of
Conitzer and Sandholm (2005a) dominance by a mixed strategy implies dominance by a
pure strategy when there are only two different payoffs, it suffices to prove the equivalences
for self-anonymous games. We further recall that CE is equivalent under disjunctive truth-
table reductions to the problem CED where a direction for the elimination is given as well.
We show equivalence of IDS to ME and of IDE to CED, both under polynomial-time
reductions.
Consider a self-anonymous game Γ = (N, (Ai )i∈N , (pi )i∈N ) such that for all i ∈ N,
Ai = {0, 1}, and assume without loss of generality that for all i ∈ N and all aN ∈ {0, 1}N ,
pi (aN ) ∈ N. Since in games with two actions it suffices to consider dominance by pure
strategies, we can otherwise construct a game with payoffs from the natural numbers that
is equivalent with respect to iterated dominance. Now write down the payoffs of Γ in an
(|N|+1)×|N| matrix X such that the jth column contains the payoffs of player j ∈ N for the
different numbers of players playing action 1, i.e., xij = pj (aN ) where #(1, aN ) = i. Then,
the jth column of X is increasing in an interval [k0 , k1 ] if and only if action 1 dominates
action 0 for player j given that at least k0 −1 and at most k1 −1 other players play action 1.
Analogously, the jth column is decreasing in such an interval if action 0 dominates action 1
under the same conditions. If player j eliminates action 0 or 1, respectively, this decreases
the number of players that can still play the respective action, corresponding to the
deletion of the top or bottom row of X, respectively. Furthermore, since every player has
only two actions, the corresponding column of the matrix can be ignored as soon as one
of them has been deleted. Observing that the above does not impose any restrictions on
the resulting matrix apart from its dimensions, equivalence of the corresponding problems
follows.

A natural way of obtaining restricted versions of ME is to consider special classes

of matrices, like matrices with entries in {0, 1} or with a bounded number of maximal
intervals in which a particular column is increasing or decreasing. One such restriction
is to require that all columns are increasing or decreasing in [1, m]. It is not too hard to
show that this makes the problem tractable irrespective of the dimensions of the matrix.
We will formally state this result in the next section and prove it as a corollary of a more
general result. Unfortunately, tractability of this restricted case does not tell us a lot
about the complexity of ME in general. The latter obviously becomes almost trivial if the
order of elimination for the columns is known, i.e., if we ask for a specific vector c ∈ [n]k
84 5 · Anonymous Games

whether there exists a vector r ∈ {0, 1}k such that (c, r) is an elimination sequence. This
observation directly implies membership in NP. More interestingly, deciding whether for
a given r ∈ {0, 1}k there exist c ∈ [n]k such that (c, r) is an elimination sequence is also
tractable. The reason is the specific “life cycle” of a column. Consider a matrix X, two
intervals I, J ⊆ [m] over the rows of X such that J ⊆ I, and a column c ∈ [n] that is active
in both I and J. Then, c must also be active for any interval K such that J ⊆ K ⊆ I,
and c must either be increasing for all three intervals, or decreasing for all three intervals.
Thus, r determines for every i ∈ [k] a set of possible values for ci , and leaves us with a
matching problem in a bipartite graph with edges in [n] × [k]. A simple greedy algorithm
is sufficient to solve this problem in polynomial time. Closer inspection reveals that it
can in fact be decomposed into two independent matching problems on convex bipartite
graphs, for which the best known upper bound is NC2 (Glover, 1967). As we will see in
the following section, yet another way to make the problem tractable is to provide a set
of k pairs (cj , rj ) that have to appear in corresponding places in the sequences of rows and
columns, while leaving open the ordering of these pairs.
But what if nothing about c and r is known? While we can only eliminate the top
or bottom row of the matrix in each step, this still amounts to an exponential number
of possible sequences. The best upper bound currently known for matching in convex
bipartite graphs does not allow us to construct an algorithm that determines r nonde-
terministically and computes a matching on the fly. We can nevertheless use the above
reasoning to recast the problem in the more general framework of matching on paths.
For this, we will identify intervals and pairs of intervals over the rows of X by vertices
and edges of a directed graph G, and will then label each edge (I, J) for two intervals I
and J by the identifiers of the columns of X that take I to J. An elimination sequence
of length k for X then corresponds to a path of length k in G which starts at the vertex
corresponding to the interval [1, m], such that there exists a matching of size k between
the edges on this path and the columns of X. In particular, by fixing a particular path,
we obtain the bipartite matching problem described above. A more detailed discussion of
this problem is the topic of the following section. We first study the problem on its own,
and return to matrix elimination toward the end of the section.

5.4.2 Matched Paths

Let us define the matching problem described above more formally. This problem gener-
alizes the well-studied class of matching problems between two disjoint sets, or bipartite
matching problems, by requiring that the elements of one of the two sets form a certain
sub-structure of a combinatorial structure. This problem is particularly interesting from
a computational perspective if identifying the underlying combinatorial structure can be
done in polynomial time, as for paths like in our case, or for spanning trees.

Definition 5.20 (matching, matched path). Let X be a set, Σ an alphabet, and σ : X → 2Σ

a labeling function assigning sets of labels to elements of X. Then, a matching of σ is a
5.4 · Iterated Weak Dominance 85

total function f : X → Σ such that for all x, y ∈ X, f(x) ∈ σ(x) and f(y) 6= f(x) if y 6= x.
Let G = (V, E) be a directed graph, Σ an alphabet, and σ : E → 2Σ a labeling function
for edges of G. Then, a matched path of length k in G is a sequence e1 , e2 , . . . , ek such
that

 for all i, 1 6 i < k, ei ∈ E and there exist u, v, w ∈ V such that ei = (u, v) and
ei+1 = (v, w), and

 the restriction of σ to { ei : 1 6 i 6 k } has a matching.

We call matched path (MP) the following computational problem: given the explicit
representation of a directed graph G with corresponding labeling function σ and an in-
teger k, does there exist a matched path of length k in G? Variants of this problem can
be obtained by asking for a matching that contains a certain set of labels, or a matched
path between a particular pair of vertices. These variants also have an interesting inter-
pretation in terms of sequencing with resources and multi-dimensional constraints on the
utilization of these resources: every resource can be used in certain states corresponding
to vertices of a directed graph, and their use causes transitions between states. The goal
then is to find a sequence that uses a specific set or a certain number of resources, or one
that reaches a certain state.
In the context of this thesis, we are particularly interested in instances of MP corre-
sponding to instances of ME. We will see later on that the graphs of such instances are
layered grid graphs (e.g., Allender et al., 2006), and that the labeling function satisfies
a certain convexity property. But let us first look at the general problem. Greenlaw
et al. (1995) consider the related labeled graph accessibility problem, which, given a
directed graph G with a single label attached to each edge, asks whether there exists a
path such that the concatenation of the labels along the path is a member of a context
free language L given as part of the input. This problem is P-complete in general and
LOGCFL-complete if G is acyclic. A matching, however, corresponds to a partial per-
mutation of the members of the alphabet, and Ellul et al. (2004) have shown that the
number of nonterminal symbols of any context-free grammar in Chomsky normal form
for the permutation language over Σ grows faster than any polynomial in the size of Σ.
It should thus not come as a surprise if the problem becomes harder when we ask for a
matching. Indeed, MP bears some resemblance to the NP-complete problem forbidden
pairs of finding a path in a directed or undirected graph if certain pairs of nodes or edges
may not be used together (Gabow et al., 1976). Instead of trying to reduce forbidden
pairs to MP, however, we show NP-hardness of a restricted version of MP using a slightly
more complicated construction. We will then be able to build on this construction in
Section 5.4.3.
In the following we restrict our attention to the case where G is a layered grid graph.

Definition 5.21 (layered grid graph). A directed graph G = (V, E) is an m×n grid graph
if V = [m]0 × [n]0 . An edge (u, v) ∈ E is called south edge if u = (i, j) and v = (i + 1, j)
86 5 · Anonymous Games

for some i and j, and east edge if u = (i, j) and v = (i, j + 1) for some i and j. A grid
graph is called layered if it contains only south and east edges.

Theorem 5.22. MP is NP-complete. Hardness holds even if G is a layered grid graph,

if |σ(e)| = 1 for every e ∈ E, and |{ e ∈ E : λ ∈ σ(e) }| 6 2 for every λ ∈ Σ.

Proof. Membership in NP is immediate. We can simply guess a sequence of edges of the

required length as well as an assignment of labels to these edges, and verify in polynomial
time that we have in fact obtained a path and a matching on this path.
For hardness, we provide a reduction from the NP-complete problem balanced one-
in-three 3SAT (B3SAT) (Parberry, 1991) to MP with the above restrictions. A B3SAT
instance over a set U of variables is given by a set C ⊆ U3 of clauses of length three such
that every variable occurs in exactly three clauses, i.e., |{ (x1 , x2 , x3 ) ∈ C : xi = x }| = 3
for all x ∈ U. An instance is called satisfiable if there exists an assignment to the
variables such that exactly one element of each clause is true, i.e., a set V 0 ⊆ V such that
|{ i : xi ∈ V 0 }| = 1 for all (x1 , x2 , x3 ) ∈ C. It is easily verified that |U| = |C| for every
instance of B3SAT, and |U| = 3|V 0 | for every assignment V 0 satisfying C. In particular,
satisfiable instances must have |U| divisible by three.
Given a particular B3SAT instance C, we construct an MP instance consisting of a
complete layered grid graph G = (V, E) and a labeling function σ : E → 2Σ such that
a path between two designated nodes s and t of G has a matching if and only if C is
satisfiable. For the moment, we will put aside the restrictions that all sets of labels are
singletons and every label occurs on at most two different edges. This allows us to prove
hardness for a labeled graph with special structure, which will then also be used in the
proof of Theorem 5.27. At the end of the current proof, we will see that the construction
can easily be modified to meet the above requirements for σ.
Now let m = |U|, and define G as a complete 6m × 5m layered grid graph. Figure 5.9
illustrates the overall structure of the labeling function σ. From s to t, G is composed of
gadgets for each of the variables of C, gadgets for the clauses, and a final path of 2m east
edges. We write si and ti , 1 6 i 6 2m, for the initial and final node of the ith of these
gadgets. Before we take a closer look at both types of gadgets, let us define the set Σ of
labels available for labeling edges of G. For every variable xi of C, 1 6 i 6 m, we have
six labels λij , 1 6 j 6 6, appearing on east edges only. Labels λvij , j ∈ {1, 2}, and λcij ,
j ∈ {1, 2, 3}, on the other hand, are exclusive to south edges. The labeling function σ is
defined in such a way that labels on east edges appear on every east edge in the respective
rows of the grid, and labels on south edges appear on every south edge in the respective
columns. Furthermore, for each label, there are at most two sets of subsequent rows
or columns where this label appears. Intuitively, the gadget for variable xi lies at the
intersection of columns carrying labels λvij and rows carrying labels λij , while the gadget
for clause ci lies at the intersection of columns carrying labels λij and rows carrying labels
for the variables that appear in ci .
Figures 5.10 and 5.11 illustrate the gadgets for variables and clauses of C. The labeling
5.4 · Iterated Weak Dominance 87

variable gadgets
s = s1

t1
s2

s3
...
sm

clause gadgets
sm+1

tm+1
sm+2

sm+3
...
s2m

s2m+1 Σ− ... Σ−
t

Figure 5.9: Overall structure of the layered grid graph G used in the proof of Theorem 5.22

function is defined using the following subsets of Σ:

Σi = {λi1 , λi2 , λi3 }

Σ−i = {λi4 , λi5 , λi6 }
−
Σ−
S
= 16i6m Σi
Σv0 = {λ11 }
v

Σvi = {λvi,1 , λvi,2 , λvi+1,1 } for 1 6 i 6 m − 1

Σvm = Σ0 = {λm1 , λm2 , λ11 , λ12 }
c v v c c

Σci = {λci1 , λci2 , λci3 , λci+1,1 , λci+1,2 } for 1 6 i 6 m − 1

Σcm = {λcm1 , λcm2 , λcm3 }

Labels in Σi and Σ− i correspond to a positive and negative assignment of the ith variable,
respectively. Sets Σvi and Σci contain auxiliary labels for the ith variable gadget and the ith
clause gadget. Note that while labels in sets Σ− i are marked as “negative,” assigning them
to edges in the variable gadget actually sets variable xi to true, because the selection of
labels from the corresponding set Σi will have to take place in the respective clause gadgets.
Returning to Figure 5.9, the final path of east edges from s2m+1 to t has length 2m, and
each of the edges carries all “negative” variable labels λij for 1 6 i 6 m and j ∈ {4, 5, 6}.
It is readily appreciated that G and σ can be constructed from C in polynomial time.
Two properties of G and σ will be useful in the following. First, every path from s
88 5 · Anonymous Games

.. ..
. .
Σvi−1 Σvi
Σi Σi Σi Σi Σi
... ···
si
Σvi−1 Σvi
ti
··· ···
Σ−
i
.. Σ−
i Σ−
i Σ−
i Σ−
i
.
Σvi
si+1

Figure 5.10: Gadget for variable xi used in the proof of Theorem 5.22

to t traverses exactly 6m east edges and 5m south edges, which equals the overall number
of labels for both directions. Secondly, a matched path from s to t must traverse every
edge (ti , si+1 ) for 1 6 1 6 2m. To see this, assume for contradiction that there is an edge
(v, v 0 ) 6= (ti , si+1 ) on the path such that ti is reachable from v but not from v 0 . If v is to
the west from ti , i.e., (v, v 0 ) is a south edge, then the number of south edges on the path
up to v 0 exceeds the number of labels available for these. If v is to the north from ti , i.e.,
(v, v 0 ) is an east edge, then the number of labels for south edges that do not appear on
any edge reachable from v 0 exceeds the number of south edges on the path to v 0 . In both
cases, the number of edges differs from the number of labels available for these edges, and
the path cannot have a matching.
Now assume that there exists a satisfying assignment for C. We construct a path from
s to t via all si and ti for 1 6 i 6 2m, as well as a matching for this path of size |Σ|. For
vertices si and ti with 1 6 i 6 m, i.e., the gadget for variable xi , we select the path labeled
with elements of Σi if xi = true, and the path labeled with elements of Σ− i otherwise. For
nodes si and ti with m < i 6 2m, i.e., the gadget for clause ci , we select the (unique)
path labeled with λjk for some k such that xj = true. In both cases, we arbitrarily assign
one of the available labels to each edge. By this, “positive” labels λ ∈ Σi corresponding to
variable xi are assigned to edges in clause and variable gadgets, respectively, depending
on whether or not xi = true. Every “positive” label is used exactly once on the path
from s to t2m , and none of the “negative” labels is used more than once. Since a satisfying
assignment must set exactly m/3 variables to true, and since, by construction of G, 2m
of the “negative” labels are not assigned to any edge on a labeled path from s to t2m ,
arbitrarily assigning these labels to the edges on the path from t2m to t yields a matching
for the path from s to t.
Conversely assume that there is a matched path from s to t. As observed above,
this path must traverse si and ti for all 1 6 i 6 2m. Furthermore, by construction
of G, the “positive” labels for a particular variable xi either all have to be assigned to
5.4 · Iterated Weak Dominance 89

.. ..
. .
Σci−1 Σci
{λjoji } {λjoji } {λjoji }
... ···
Σci−1 sm+i Σci
{λkoki } {λkoki } {λkoki }
··· ···
Σci−1 Σci
{λ`o`i } {λ`o`i } {λ`o`i }
··· ···
.. tm+i
. Σci

sm+i+1

Figure 5.11: Gadget for clause ci = (xj ∨ xk ∨ x` ) used in the proof of Theorem 5.22. oji
denotes the number of times variable xj occurs in clauses up to and including ci .

edges in the gadget for xi , or to edges in the gadgets for the clauses where xi appears,
but not both. It is then easily verified that setting a variable to true if and only if the
corresponding “positive” labels are assigned to edges in clause gadgets yields a satisfying
assignment. Thus, there is some path from s to t in G that has a matching if and only
if C is satisfiable.
It remains to be shown that the above construction can be simplified such that every
edge can be labeled with exactly one label and every label appears on at most two different
edges. For this, we first remove all edges that cannot be part of a path from s that has a
matching, i.e., those that are not part of any gadget. Then, for every set of labels defined
above, the number of edges labeled with this set within a particular gadget equals the
cardinality of the set, and we can assign a different singleton to each of these edges. The
path from s2m+1 to t requires some additional attention. We know that, at the time we
have found a path from s to s2m+1 that does not use any of the labels more than once,
exactly 2m labels in Σ− have not yet been assigned to an edge, but we do not know which.
To ensure that the remaining labels can be chosen in an arbitrary order, we replace the
path starting at s2m+1 by 2m/3 additional gadgets of the form shown in Figure 5.12,
which use 2m2 /3 additional labels λeji for 1 6 j 6 m and 1 6 i 6 2m/3. It is easily
verified that the modified labeling function satisfies the desired constraints.

Let us now return to matrix elimination. In light of Theorem 5.22, an efficient al-
gorithm for ME would have to exploit additional structure of MP instances induced by
instances of ME. It turns out that this structure is indeed quite restricted in that edges
carrying a particular label satisfy a “directed” convexity condition: if a particular label λ
appears on two edges e = (u, v) and e 0 = (u 0 , v 0 ), then λ must appear on all south edges
90 5 · Anonymous Games

.
s2m+i .. {λ14 } {λ15 } {λ16 }
···
{λe1i } {λe1i }
{λ24 } {λ25 } {λ26 }

.. ..
. .

{λemi } {λemi }
{λm4 } {λm5 } {λm6 }
···
..
. s2m+i+1

Figure 5.12: Gadget to consume remaining labels, used in the proof of Theorem 5.22

or on all east edges that lie on a path from u to v 0 , but not both. In particular, if there
is such a path, it cannot be that one of e and e 0 is a south edge and the other is an east
edge. This fact is illustrated in Figure 5.13, which shows the labeled graph for the ME
instance of Figure 5.7, as well as a matched path corresponding to an elimination sequence
of maximum length.
Let us formally define the above property, along with a second property which requires
the set of edges carrying a particular label to form a weakly connected subgraph of G. We
henceforth concentrate on complete layered grid graphs, i.e., ones that contain all south
and all east edges.

Definition 5.23 (directed convexity, connectedness). Let G = (V, E) be a complete lay-

ered grid graph. A labeling function σ : E → 2Σ for G is called directed convex if for
every label λ ∈ Σ and for every set of three edges e1 , e2 , e3 with ei = (ui , vi ), such that
u2 is reachable from u1 , u3 is reachable from u2 , and λ ∈ σ(e1 ) ∩ σ(e3 ), it holds that e1
and e3 have the same direction and λ ∈ σ(e2 ) if and only if e2 has the same direction as
well. A labeling function σ is called connected if for every λ ∈ Σ and every pair of edges
e1 , e2 ∈ E such that λ ∈ σ(e1 ) ∩ σ(e2 ) there exists (u, v) ∈ E such that λ ∈ σ(u, v) and
both e1 and e2 are reachable from u.

It is not too hard to see that instances corresponding to ME have a directed convex
labeling function. Connectedness is related to a restricted version of ME which we term
matrix elimination with given directions (MED): given a matrix X, a labeling function σ,
and a total function d : [n] → {0, 1}, does there exist an elimination sequence (c, r) with
directions given by d, i.e., one such that for all i, j satisfying d(i) = j there is some ` ∈ N for
which ci = i and ri = j. We will see shortly that this problem has a natural interpretation
in terms of iterated dominance.
5.4 · Iterated Weak Dominance 91

{b, d} {a, b, d} {a, b, d} {a}

(0, 0) (0, 4)
{c} {c} {c} ∅
{b, d} {b, d} {b, d}

{a, c} {c} {c} (1, 3)

∅ ∅
{a} ∅ (2, 2)
∅
{a} (3, 1)

(4, 0)

Figure 5.13: Labeled graph for the matrix elimination instance of Figure 5.7. A matched
path and its matching are shown in bold.

Lemma 5.24. ME is polynomial time reducible to MP restricted to layered grid graphs

and directed convex labeling functions. MED is polynomial time equivalent to MP
restricted to layered grid graphs and directed convex and connected labeling functions.

Proof. Consider the following reduction from ME to MP. For a matrix X ∈ Nm×n , define
a layered grid graph G = (V, E) with V = [m]0 × [n]0 and a labeling function σ : E 7→ 2[n]
such that for all λ ∈ [n], λ ∈ σ(e) if for some i, j ∈ N, e = ((i, j), (i + 1, j)) and column `
of X is increasing in [i + 1, m − j], or e = ((i, j), (i, j + 1)) and column ` of X is decreasing
in [i + 1, m − j]. Now consider k ∈ N, c ∈ [n]k , and r ∈ {0, 1}k . Let p = e1 , e2 , . . . , ek be
a path in G such that e1 = ((0, 0), v) for some v ∈ V, and ei is a south edge if and only if
ri = 0. Further define a function f : E → [n] by letting f(ei ) = ci for all i 6 k. It is not
too hard to see that (c, r) is an elimination sequence of X if and only if f is a matching
for the restriction of σ to the edges on p.
For directed convexity of σ, consider e1 , e2 , e3 ∈ E with ei = (ui , vi ), such that u2
is reachable from u1 and u3 is reachable from u2 . For ` = 1, 2, 3, define an interval
I` = [i + 1, m − j] for i, j ∈ N such that e` = ((i, j), v) for some v ∈ V. Further consider
λ ∈ σ(e1 ) ∩ σ(e3 ). By definition of σ, column λ of X must be active in both I1 and I3 .
Since I3 ⊆ I1 , λ must either be increasing in both of them, or decreasing in both of them.
Since I3 ⊆ I2 and I2 ⊆ I1 , the same must also be true for I2 .
For MED, consider a total function d : [n] → {0, 1} and define σ 0 : E 7→ 2[n] such that
for all e ∈ E and λ ∈ [n], λ ∈ σ 0 (e) if λ ∈ σ(e) and if either e is a south edge and f(λ) = 0
or e is an east edge and f(λ) = 1. It is not hard to see that σ 0 is directed convex and
connected. On the other hand consider a layered grid graph G = (V, E) and a directed
convex and connected labeling function σ. Then, for every λ ∈ Σ, there exists a unique
pair of vertices u, v ∈ V such that λ ∈ σ(e) for exactly those south edges or exactly those
92 5 · Anonymous Games

east edges e that are reachable from u but not from v. It is now possible to define a
matrix X with a column for λ that is active exactly in every interval I such that I ⊆ [i, j]
and I ∩ [i 0 , j 0 ] 6= ∅, and increasing if f(λ) = 0 and decreasing if f(λ) = 1. By the same
reasoning as above, elimination sequences of X correspond to matched paths of G and σ
with initial vertex (0, 0).

Label a in the instance of Figure 5.13 serves as an example that the labeling function
of an instance of MP corresponding to one of ME does not have to be connected, and it
even appears on both east edges and south edges. On the other hand, MP can be solved in
polynomial time when restricted to instances that do satisfy connectedness in addition to
directed convexity. This also means that we can decide in polynomial time whether there
exists an elimination sequence with a specific direction of elimination for every column of
a matrix.

Theorem 5.25. Let G be a layered grid graph, σ a directed convex and connected
labeling function for G. Then MP for G, σ and k = |Σ| is in P.

Proof. It suffices to show how to decide whether there exists a matched path from s =
(0, 0) to a particular vertex t = (ks , ke ) such that ks + ke = k. Different values for t can
then be checked sequentially.
Given a path p from some vertex v1 ∈ V to t, we define two labeling functions
p
σs : [ks ] → Σ and σp e : [ke ] → Σ, one for south edges and one for east edges of paths
from s to t. We will argue that a pair of matchings for σp p
s and σe can easily be combined
into a matching for p, while nonexistence of a matching for either of the two implies that
a large set of paths in G cannot be matched paths. The latter will ultimately provide us
with a succinct certificate that a particular pair of a graph G and a labeling function σ
does not have a matched path of length k.
More formally, consider a complete layered grid graph G = (V, E) and a labeling
function σ : E → 2Σ . For a path p = e` , e`+1 , . . . , ek , define σp p
s and σe such that for
every λ ∈ Σ,

λ ∈ σp
s (i) if there exists a path e10 , e20 , . . . , ek0 with ei0 = ei for all i > `,
and j ∈ [k], i 0 ∈ [ke ] such that
ej = ((i − 1, i 0 ), (i, i 0 )) and λ ∈ σ(ej ), and
λ ∈ σp
e (i) if there exists a path e10 , e20 , . . . , ek0 with ei0 = ei for all i > `,
and j ∈ [k], i 0 ∈ [ks ] such that
ej = ((i 0 , i − 1), (i 0 , i)) and λ ∈ σ(ej ).

In other words, σp p
s and σe provide an “optimistic” version of the matching problems
obtained by restricting σ to a path in G that contains p as a sub-path, by allowing a
certain label to be matched to the ith south edge or east edge of these paths, respectively,
if it appears on the ith edge in the respective direction of some such path. It follows from
5.4 · Iterated Weak Dominance 93

directed convexity and connectedness of σ that for every path p, σp p

s and σe are convex
functions and { a ∈ σp p
s (i) : i ∈ [ks ] } ∩ { a ∈ σe (i) : i ∈ [ke ] } = ∅. We can further assume
without loss of generality that for every path p, σp p
s and σe have images of size ks and ke ,
respectively.
Now let p be a particular path from s to t. By definition, there is a one-to-one corre-
spondence between σp s and the restriction of σ to south edges of p, and also between σe
p
p p
and the restriction of σ to east edges of p. Any pair of matchings for σs and σe thus
directly corresponds to a matching for the restriction of σ to p, and existence of the former
implies that p is a matched path.
On the other hand, consider a path p = e` , e`+1 , . . . , ek and an edge e ∈ E such that
e = (u, v) and e` = (v, w) for some u, v, w ∈ V. Denote p 0 = e, e` , e`+1 , . . . , en , and
p0 p0
assume that both σp p
s and σe have a matching while σs or σe does not. First consider
the case where e is an east edge, and where the function that does not have a matching
0 p0
is σp p
e . Let i, j ∈ N such that u = (i, j). By definition, σe and σe only differ with respect
p0
to labels λ such that λ ∈ σp 0 0
e (j ) if and only if j < j. Since σe is a convex function that
0
does not have a matching, and since the image of σp e has size ke , there has to be some
interval in [ke ] the size of which is strictly larger, and some interval the size of which
0
is strictly smaller than the number of labels σp e assigns exclusively to elements of this
interval. Furthermore, every matching f of σp e must satisfy f(j) ∈ / σ(e), since a matching
p0
with f(j) ∈ σ(e) would also be a matching for σe . This means that there actually must
exist an interval I of the second type such that I ⊆ [1, j − 1]. Now consider any path p 00
from a vertex u 0 south of u to t, i.e., a vertex u 0 = (i 0 , j) such that i 0 > i. Clearly, the
00 p0
number of labels appearing exclusively in I cannot be smaller for σp e than it is for σe .
00
This means that σp e does not have a matching, and thus that no matched path of G and σ
can traverse u . 0
0
Now assume that the function that does not have a matching is σp s . This again means
0
that there has to be an interval such that the number of different labels assigned by σp s to
elements of this interval is strictly smaller than the length of this interval. Since σp s has
p p0
a matching, since the restrictions of σs and σs to [i, ks ] are identical, and by convexity
0
of σp s , there has to exist an interval I with this property such that I ⊆ [0, i − 1]. Now
consider any path p 00 from a vertex u 0 west of u to t, i.e., a vertex u 0 = (i, j 0 ) such that
00 p0
j 0 < j. By definition, for any λ ∈ Σ, λ ∈ σp e only if λ ∈ σe , such that the number of
00 p 00
different labels assigned by σp s to elements of I is strictly smaller than |I|. Thus σs does
not have a matching, and thus no matched path of G and σ can traverse u 0 .
If e is a south edge, then by symmetrical arguments either no matched path of G and σ
can traverse any vertex north of u, or no such path can traverse any vertex east of u.
Now consider an algorithm which starts at t and tries to iteratively construct a path
from s to t by traversing edges of G backwards. Given a path p` of length ` from a
vertex v` to t, the algorithm selects p`+1 to be a path of length ` + 1 containing p` as a
sub-path such that both σs p`+1 and σs p`+1 have a matching. If the algorithm runs for k
steps, we obtain a path pk from s to t with this property, i.e., a matched path. Assume
94 5 · Anonymous Games

on the other hand that for some ` no path satisfying the above requirements exists, and
denote by P the set of paths obtainable by adding a predecessor of v` to p` . This set
contains one or two paths depending on whether v` has one or two predecessors. Then,
for every p ∈ P, one of σp p
s and σe does not have a matching, and from the above reasoning
we obtain a set of vertices such that no matched path can traverse any of these vertices.
It is easily verified that the union of these sets for the different elements of P always forms
a cut that separates s from t, implying that a matched path from s to t cannot exist.

An interesting consequence of Lemma 5.24 and Theorem 5.25 is that we can decide in
polynomial time whether an anonymous game with two actions can be solved such that
particular actions remain for the different players. Matrix elimination when all columns
are active from the beginning is a special case of Theorem 5.25. With some additional
work, we can derive a better upper bound.

Corollary 5.26. Let X ∈ Nm×n be a matrix every column of which is active in [1, m].
Then ME for X is in L.

Proof. Consider the graph G and the labeling function σ corresponding to X. For any
path p in G with final vertex t, consider the functions σp p
s and σe defined in the proof
of Theorem 5.25. Since every column of X is active in [1, m], σp p
s and σe are convex.
Moreover, for every label λ ∈ Σ, λ ∈ σp p
s (1) or λ ∈ σe (1). It is not too hard to see that for
a path p = vk , vk+1 , . . . , vn with vn = t, σp p
s and σe have a matching if and only if there
0 p0
exists a path p from vk+1 = (i, j) to t such that σp
0
s and σe have matchings fs and fe ,
and if the number of labels both in σp p
s (i − 1) \ (∪k>i fs (k)) and in σs (i − 1) \ (∪k>j fe (k))
are strictly positive. We can thus construct a path by moving backwards from t, and
storing a pointer to the current source of the path, and the numbers of labels that are
currently active but have not been assigned to edges between the current source and t.
This can clearly be done using only logarithmic space.

The complexity of ME remains open, and additional insights will be necessary to solve
this question. The proof of Theorem 5.25 hinges on connectedness of the labeling function.
On the other hand, directed convexity of the labeling function corresponding to an ME
instance means that we cannot use a construction similar to the one used in the proof of
Theorem 5.22 to show NP-hardness of MP. Another interesting question is whether the
case m > n is easier than the general case.

5.4.3 Self-Anonymous Games With a Constant Number of Actions

It is natural to ask what happens for games with more than two actions, and whether
there still exists a nice interpretation in terms of row and column eliminations in a matrix
or matrix-like structure. It turns out there is such an interpretation, but its formulation
is rather complicated. Consider a self-anonymous game with n players and k actions for
each player. As before, the payoff of a particular player i only depends on the number
5.4 · Iterated Weak Dominance 95

of players, including himself, that play each of the different actions. For a particular
P
player we thus have payoff values for each tuple (j1 , j2 , . . . , jk ) with k `
`=1 j = n. These
can be represented as entries in a discrete simplex of dimension k − 1. When writing
1 k
down the payoffs of all players, one obtains a structure X = (xji ···j )i∈N,Pk j` =n where
`=1
1 2 k
xij j ···j ∈ R denotes the payoff of player i ∈ N if for each `, j` players play action a` .
This structure has the aforementioned simplices as columns, and resembles a triangular
prism in the case k = 3.
Restricting our attention to dominance by pure strategies, action a` ∈ A weakly dom-
inates action am ∈ A for player i ∈ N if i can never decrease his payoff by playing a`
instead of am , no matter which actions the other players play, and if the payoff strictly
increases for at least one combination of actions played by the other players. This cor-
responds to the values in the ith column of X being increasing from am to a` , i.e.,
weakly increasing with a strict increase at some position (j1 , j2 , . . . , jk ). If m players have
eliminated action a` , tuples with j` > n − m are no longer reachable, corresponding to
a cut along the `th 0-face of the simplex. Eliminations of a particular action have the
same effect on the payoff simplex of every single player and thus correspond to cuts along
the respective edge of the prism in the case k = 3. Given a vector d = (di )16i6k with
1 6 di 6 n, we will write X(d) to denote the structure obtained from X by performing,
1 2 k
for each i, di eliminations in dimension i, i.e., X(d) = (xji j ···j )ji 6n−di .
Now, let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a self-anonymous game, and let X be defined
1 2 k P
by xij j ···j = pi (j1 , j2 , . . . , jk ) for all i ∈ N and j1 , j2 , . . . , jk ∈ N0 such that k `
`=1 j = n.
Then, Γ is solvable using iterated dominance by pure strategies if there exists a pair (c, r)
of sequences c ∈ N(k−1)n and r ∈ A(k−1)n such that

(i) |{ 1 6 i 6 (k − 1)n : ci = j }| = k − 1 for all j ∈ N,

(ii) ci = cj and ri = rj implies i = j for all 1 6 i, j 6 (k − 1)n, and

(iii) for each i, 1 6 i 6 (k − 1)n, there exists some r∗ ∈ A such that, for all j < i, cj 6= ci
or rj 6= r∗ , and ci is increasing from ri to r∗ in X(r1 , r2 , . . . , ri−1 ).

That is, a game is solvable if there exists a sequence of (k − 1)n pairs of a player and an
action such that (i) every player deletes exactly k − 1 times, (ii) no player deletes the same
action twice, and (iii) every action is deleted using some other action that has not itself
been deleted.
The left hand side of Figure 5.14 shows the payoffs of a particular player in a self-
anonymous game with n = 3 and k = 3. Compared to matrix elimination as introduced
in Definition 5.17 and illustrated in Figure 5.7, we notice an interesting shift. Curiously,
this shift has nothing to do with the added possibility of dominance by mixed strategies
in games with more than two actions. Rather, a particular action a ∈ A may now be
eliminated by either one of several other actions in A \ {a}, and the situations where a a
can be eliminated no longer form a convex set. Recalling the proof of Theorem 5.22, our
96 5 · Anonymous Games

1
1 1 1 1
0 0 0 0 0 0 0 0
1 0 0 1 0 0 1 0 0 1 0
3
0 1 0 1 0 1 0 1 0 1 0 3 2 1
2 3 2 3 2

Figure 5.14: Payoffs of a particular player in a self-anonymous game with three players
and three actions. Initially all actions are pairwise undominated. If one of the other
players eliminates action 1, action 3 weakly dominates action 1. Action 1 then becomes
undominated if some player deletes action 3, and dominated by action 2 if one more player
deletes action 3, and some player deletes action 2.

strategy becomes clear: try to construct a layered grid graph and a labeling of its edges
for which the existence of a matched path is NP-hard to decide, and which are induced by
a self-anonymous game with three actions for each player. It turns out that this is indeed
possible.

Theorem 5.27. IDS and IDE are NP-complete. Hardness holds even for self-anony-
mous games with three actions and two different payoffs.

Proof. Membership in NP is immediate. We can simply guess a sequence of eliminations

and verify that all of them are valid and that they eventually leave only a single action
for each player.
For hardness of IDS, recall the construction used in the proof of Theorem 5.22. Given
a B3SAT instance C, we constructed an MP instance consisting of a layered grid graph
G = (V, E) and a labeling function σ : E → 2Σ such that a path between two designated
nodes s and t has a matching if and only if C is satisfiable. We will now show that G and σ
correspond to iterated dominance in a specific self-anonymous game Γ with k = 3 when
only actions 1 and 2 of each player are considered. Observing that, given a matched path
from s to t, all players in Γ can also eliminate action 3 at some vertex on the path without
affecting the restriction of the labeling function to the remainder of the path effectively
reduces B3SAT to IDS.
Given a particular grid graph G, a set Σ of labels, and a labeling function σ as defined
in the proof of Theorem 5.22, we construct a game Γ with players N = Σ and actions
A = {1, 2, 3}. Action 1 is associated with east edges of G, action 2 is associated with
south edges. Now consider a particular label i ∈ Σ. By construction of G, there exist two
numbers k1 and k2 such that i appears exclusively on east edges (south edges, respectively)
that can be reached from s by traversing exactly k1 or k2 south edges (east edges). In
game Γ , this is modeled by a player that can eliminate action 1 (action 2) after exactly k1
or k2 players have eliminated action 2 (action 1). Since we only use payoffs 0 and 1, it
follows from Lemma 1 of Conitzer and Sandholm (2005a) that an action is dominated by a
5.4 · Iterated Weak Dominance 97

1 (2)

···0···

1 · · · 10 · · · 0
k1
0 · · · 01 · · · 1
··· 0 ···
0 ··· 0
k2
1 ··· 1
··· 0 ···

2 (3) 3 (1)

Figure 5.15: Payoff structure of a particular player of the self-anonymous game Γ used in
the proof of Theorem 5.27. There are two types of players, eliminating action 2 and 1,
respectively, actions of the second type are shown in parentheses. The player may eliminate
action 2 (action 1, respectively) by action 3 after exactly k1 players have eliminated
action 1 (2), and by action 1 (2) after exactly k2 players have eliminated action 1 (2).

mixed strategy if and only if it is dominated by a pure strategy. We can thus concentrate
exclusively on dominance by pure strategies.
The payoff structure for players of Γ is shown in Figure 5.15. Clearly, Γ can be con-
structed from G in polynomial time. In addition to the aforementioned properties regard-
ing the elimination of action 1 or 2, it is easily verified that every player can also eliminate
action 3 after k2 eliminations of action 2 or 1, and that this has no effect whatsoever on
the ability of other players to eliminate their actions. In other words, Γ actually induces
a three-dimensional grid graph, where each layer in the third dimension is identical to G,
and transitions between different layers may take place at vertices where some player has
arrived at k2 . This means, however, that a matched path from s to t corresponds to a
sequence of eliminations of actions 1 and 2 in Γ , which can in turn be transformed into
a sequence of eliminations that solves Γ by iterated dominance by letting each player
eliminate action 3 at a certain well-defined point. On the other hand, the possible future
transitions within a particular layer of the three-dimension grid graph do not depend on
the layer, i.e., a player may not gain the ability to eliminate actions 1 or 2 by first elim-
inating action 3. Hence, if there is no matched path from s to t, then some player of Γ
will not be able to eliminate either action 1 or action 2, meaning that Γ is not solvable by
iterated dominance.
Hardness of IDE can finally be obtained by adding an additional player that can only
eliminate once the lowest level of the grid graph has been reached.
98 5 · Anonymous Games

5.5 Discussion

In this chapter, we introduced four classes of anonymous games and investigated the
computational complexity of pure Nash equilibrium and iterated weak dominance in these
classes. We established that the former solution concept is tractable for games with a
constant number of actions, but becomes intractable if the number of actions grows at
least linearly in the number of players. It is worth noting that, for games with a constant
number of actions, the pure equilibrium problem happens to lie in the complexity class
NC1 for all types of anonymity and is thus open to parallel computation. NP-hardness
also holds for games with an exponential number of players and logarithmic growth of the
number of actions. For games with an exponential number of players in which the number
of actions grows sub-logarithmically, the complexity remains open. Iterated dominance,
on the other hand, is tractable in symmetric games with any constant number of actions,
but NP-hard in anonymous and self-anonymous games with only three actions. The
complexity in anonymous and self-anonymous games with two actions remains open.
In future work, it would be interesting to extend the tractability results to larger
classes of games. For example, games with a certain number of player types, where
indistinguishability holds only for players of the same type, can be obtained by restricting
Definition 5.1 to permutations that map players from a certain subset to players of the
same set. Given a game in this class, we can construct an anonymous game with the
same set of players and an action set that is the Cartesian product of the original set of
actions and the set of player types. By assigning a unique minimum payoff to all actions
not corresponding to the type of the respective player, we can ensure that players only
play actions corresponding to their type in every equilibrium of the new game, effectively
allowing us to distinguish players of different types in the new game. For games with
a constant number of players the size of the new game is polynomial in the size of the
original game, and the tractability result of Theorem 5.3 carries over immediately. A
different notion, such that players of the same type have identical payoff functions, does
not seem to provide additional structure. As we have already shown, only two different
payoff functions suffice to make the pure equilibrium problem TC0 -hard for a constant
number of actions and NP-hard for a growing number of actions. More generally, one might
investigate games where payoffs are invariant under particular sets of permutations. For
example, von Neumann and Morgenstern (1947) regard the number of permutations under
which the payoffs of a game are invariant as a measure for the degree of anonymity. The
question is in how far the computational complexity of solving a game depends on this
degree.
With respect to iterated dominance, the most important open question concerns iter-
ated weak dominance solvability in anonymous games with two actions, and the equiva-
lent problem of matrix elimination. More generally, we looked at a problem concerning
matchings on paths in a directed graph. This problem was mainly introduced as a proxy
to matrix elimination, but appears to be interesting in their own right, with connections
5.5 · Discussion 99

to ordinary matching problems, sequencing, and planning. It will therefore be worthwhile

to investigate versions of this problem with restrictions on the graph structure or labeling
function.
 Chapter 6

Graphical Games

Another structural element commonly found in real-world interaction, besides the one
considered in the previous chapter, is locality. Often a situation involves many agents,
but the weal and woe of any particular agent depends only on the decisions made by a
select few. Graphical games (Kearns et al., 2001) formalize this notion by assigning to
each player a subset of the other players, his neighborhood, and defining his payoff as
a function of the actions of these players. More formally, a graphical game is given by
a (directed or undirected) graph on the set of players of a normal-form game, such that
the payoff of each player depends only on the actions of his neighbors in this graph. Any
graphical game with neighborhood sizes bounded by a constant can be represented using
space polynomial in the number of players.
Gottlob et al. (2005) investigate the complexity of pure Nash equilibria in graphical
games, and show that deciding the existence of a pure equilibrium is NP-complete already
for a very restricted class, namely that where each player can choose from at most three
different actions and his payoff depends on at most three other players. We begin this
chapter by strengthening this result to apply to an even more restrictive setting. To be
precise, we show that two actions per player, two-bounded neighborhood, and two-valued
payoff functions suffice for NP-completeness. This result is tight, because deciding the
existence of a pure Nash equilibrium becomes trivial in the case of a single action for
each player and tractable for one-bounded neighborhood. In fact, we show the latter
problem to be NL-complete in general, and thus solvable in deterministic polynomial
time. Interestingly, it turns out that the number of actions in a game with one-bounded
neighborhood is a sensitive parameter: restricting the number of actions for each player
to a constant makes the problem even easier than NL unless L = NL. In this way, we
obtain a nice alternative characterization of the determinism-nondeterminism problem for
Turing machines with logarithmic space in terms of the number of actions for games with
one-bounded neighborhood.
We then move on to investigate the computational complexity of the pure equilibrium
problem in graphical games which additionally satisfy one of four types of anonymity
within each neighborhood. Despite these additional restrictions, the question for tractable

101
102 6 · Graphical Games

classes of games is answered mostly in the negative. For three of the four types of
anonymity, deciding the existence of a pure equilibrium remains NP-hard for games with
two actions, two payoffs, and neighborhoods of size two. Assuming the most restricted
type of anonymity, the problem becomes NP-hard when either there are three different
payoffs, or neighborhoods of size four. On the other hand, we use interesting connections
of the latter class to the even cycle problem in directed graphs and to generalized sat-
isfiability to identify tractable classes of games. One such class for example arises from
a situation where each agent is faced with the decision of producing one of two types of
complementary goods within a regional neighborhood. In a sense, agents are not only
producers but also consumers, and thus happier when both products are available within
their neighborhood.
As a corollary, we further exhibit a satisfiability problem that remains NP-hard in
the presence of a matching, a result which may be of independent interest. Finally, we
show that mixed equilibria in games with two of four types of anonymity can be found
in polynomial time if the number of actions grows only slowly in the neighborhood size.
Quite interestingly, there exists a class of games where deciding the existence of a pure
equilibrium is likely to be harder than finding a mixed one.

6.1 Related Work

The problem of finding (mixed) Nash equilibria in graphical games with neighborhood
sizes bounded by three is equivalent to the same problem for general n-player games with
n > 4 (Goldberg and Papadimitriou, 2006), and thus complete for the complexity class
PPAD (Daskalakis et al., 2006). It is not surprising that the structure of the neighborhood
graph greatly influences the complexity of the equilibrium problem. PPAD-hardness holds
even if the underlying graph has constant pathwidth, but becomes tractable for undirected
graphs of degree two, i.e., for paths (Elkind et al., 2006). All known algorithms for the
more general case of trees have exponential worst-case running time even on trees with
bounded degree and pathwidth two, but equilibria satisfying various fairness criteria can
be computed in polynomial time if additionally there are only two actions per player and
the best response policy, a data structure representing all Nash equilibria of a game, has
polynomial size (Elkind et al., 2007).
A different line of research has investigated the problem of deciding the existence of
pure equilibria. In addition to the above-mentioned hardness result, Gottlob et al. (2005)
show tractability of the pure equilibrium problem for certain classes of games with re-
stricted graph structure, in particular for graphs with bounded treewidth. The results of
Chapter 5, and to some extent also the approximability results of Daskalakis and Papadim-
itriou (2008), fuel hope that tractability results can be obtained for larger classes of games
satisfying some kind of symmetry. In this regard, Daskalakis and Papadimitriou (2005)
consider games on a d-dimensional undirected torus or grid with payoff functions that are
identical for all players and symmetric in the actions of the players in the neighborhood,
6.2 · The Model 103

a condition we refer to as symmetry. The authors show that deciding the existence of a
pure Nash equilibrium in such games is NL-complete when d = 1 and NEXP-complete for
d > 2. In the second part of this chapter, we investigate the pure equilibrium problem in
graphical games satisfying, within each neighborhood, one of the four types of anonymity
introduced in Chapter 5. This can be seen as a refinement of the work of Gottlob et al.
(2005) and of Daskalakis and Papadimitriou (2005).

6.2 The Model

A graphical game is given by a graph on the set of players, such that the payoff of a
particular player depends only on his own action, and on the actions of his neighbors in
the graph. In the following definition, the underlying graph is directed, corresponding to
a neighborhood relation that is not necessarily symmetric.1

Definition 6.1 (graphical game). Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a normal-form game,
0 ∈A ,
ν : N → 2N . Γ is a graphical game with neighborhood ν if for all i ∈ N and aN , aN N
ν̂(i) = aν̂(i) , where ν̂(i) = ν(i) ∪ {i}.
0 ) whenever a
pi (aN ) = pi (aN 0

A game Γ is said to have k-bounded neighborhoods if there exists ν : N → 2N such

that Γ is a graphical game with neighborhood ν and for all i ∈ N, |ν(i)| 6 k.

We assume throughout the chapter that graphical games are encoded by listing the
payoffs of each player as a function of the actions of his neighbors. This encoding has
polynomial size in the number of players if and only if neighborhood sizes are bounded
by a constant.
Symmetry as a property of a mathematical object refers to its invariance under a cer-
tain type of transformation. Symmetries of games usually mean invariance of the payoffs
under automorphisms of the set of action profiles induced by some group of permutations
of the set of players. Anonymous games, for example, as considered in Chapter 5, require
the set of available actions to be the same for all players, and the payoff of a particular
player to remain the same under any permutation of the elements of an action profile.
This imposes constraints on individual payoff functions only, and can therefore directly
be applied to graphical games as well. In general, however, it does not make much sense
from a computational point of view to consider symmetries of the payoff functions without
requiring the neighborhood graph to be symmetric in an appropriate way as well. Con-
sider, for example, the class of all graphical games whose payoff functions are invariant
under automorphisms in the automorphism group of the neighborhood graph. While this
class of games is very natural, it does not impose meaningful computational restrictions.
Indeed, it is not too hard to see that any graphical game can be encoded by a game in the
1
While results can be transferred between graphical games on directed and undirected graphs, tightness
of bounds on the size of the respective neighborhoods is obviously lost in the process. Results for directed
graphs will in general be more expressive.
104 6 · Graphical Games

above class that has a neighborhood graph with a trivial automorphism group. Hardness
results for both pure and mixed equilibria thus carry over immediately.
In general, different types of restrictions on the neighborhood structure will be required
for different kinds of symmetries of the payoff functions. We take a slightly different ap-
proach by considering properties found in anonymous and symmetric games and studying
graphical games that possess these properties. A characteristic feature of symmetries in
games is the inability to distinguish between other players. As in Chapter 5, the most
general class of games with this property will be called anonymous. Four different classes
of games are again obtained by considering two additional characteristics: identical pay-
off functions for all players and the ability to distinguish oneself from the other players.
The games obtained by adding the former property will be called symmetric, and pres-
ence of the latter will be indicated by the prefix “self.” For ease of exposition, we assume
the set of actions, and possibly the payoff functions, in these games to be the same for all
players rather than just those with intersecting neighborhoods. The set of actions will be
denoted by A = A1 = A2 = · · · = An .
Again, an intuitive way to describe anonymous games is in terms of equivalence classes
of the aforementioned automorphism group. Given a set A of actions, let #(aN ) denote
the commutative image of an action profile aN ∈ AN , i.e., #(aN ) = (#(a, aN ))a∈A ,
where #(a, aN ) = |{ i ∈ N : ai = a }|. This definition naturally extends to action profiles
for subsets of the players.

Definition 6.2 (anonymity). Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a graphical game with
neighborhood ν, A a set of actions such that for all i ∈ N, Ai = A. Γ is called

 anonymous if for all i ∈ N and all aN , aN 0 ∈ AN , p (a ) = p (a 0 ) whenever

i N i N
ai = ai0 and for all a ∈ A, #(a, aν(i) ) = #(a, aν(i)
0 );

0 ∈ AN , |ν(i)| = |ν(j)| and p (a ) =

 symmetric if for all i, j ∈ N and all aN , aN i N
0 ) whenever a = a 0 and for all a ∈ A, #(a, a
pj (aN ) = #(a, a 0 );
i j ν(i) ν(j)

 self-anonymous if for all i ∈ N and all aN , aN 0 ∈ AN , p (a ) = p (a 0 ) whenever

i N i N
0
for all a ∈ A, #(a, aν̂(i) ) = #(a, aν̂(i) ); and

0 ∈ AN , |ν(i)| = |ν(j)| and p (a ) =

 self-symmetric if for all i, j ∈ N and all aN , aN i N
0 0
pj (aN ) whenever for all a ∈ A, #(a, aν̂(i) ) = #(a, aν̂(j) ).

It should be noted that a graphical game in one of the four classes does not in general
belong to the corresponding class of Chapter 5, unless the neighborhood of every player
contains all other players. When talking about self-anonymous and self-symmetric games
with two actions, we write pi (m) = pi (aN ) for any action profile aN with #(1, aν̂(i) ) = m
to denote the payoff of player i when m players in his neighborhood, including himself,
play action 1, and pi = (pi (m))06m6|ν̂(i)| for the vector of payoffs for the possible values
of m.
6.3 · A Tight Hardness Result for Pure Equilibria 105

6.3 A Tight Hardness Result for Pure Equilibria

Gottlob et al. (2005) show that deciding the existence of a pure equilibrium is NP-complete
for graphical games with three-bounded neighborhoods and at most three actions for
each player. We improve upon this result by showing that hardness holds already in
the case of two-bounded neighborhoods, two actions, and two-valued payoff functions.
Schoenebeck and Vadhan (2006) independently showed NP-hardness for three-bounded
symmetric neighborhoods. Both constructions can easily be adapted to show the respec-
tive other result.

Theorem 6.3. Deciding whether a graphical game has a pure Nash equilibrium is
NP-complete. Hardness holds even for games with two-bounded neighborhoods, two
actions for each player, and two-valued payoff functions.

Proof. Membership in NP is obvious. We can guess an action profile s and verify in

polynomial time that it satisfies the equilibrium condition.
For hardness, recall that circuit satisfiability (CSAT), i.e., deciding whether for a given
Boolean circuit C with k inputs and one output there exists an assignment such that C
evaluates to true, is NP-complete (e.g., Papadimitriou, 1994a). Assume without loss of
generality that C contains at least one input and one (internal) gate, and that NOT gates
only occur at the input layer. For an arbitrary circuit, all NOT gates can be moved to
the input layer in polynomial time by successive application of de Morgan’s law.
Given a Boolean circuit C, we define a graphical game Γ = (N, (Ai )i∈N , (pi )i∈N ), and
argue that Γ has a pure Nash equilibrium if and only if C is satisfiable. As for the players
of Γ , there is one for each input of C, one for each (positive or negative) literal, and one for
each gate of types AND and OR. We denote the respective sets of players by Ni , Nx , Nx̄ ,
N∧ , and N∨ . The output of C corresponds to a particular player o ∈ N∧ ∪ N∨ . For each
input player i ∈ Ni , let ν(i) = ∅. For each gate player i ∈ N∧ ∪ N∨ , let ν(i) be the set of
players corresponding to the inputs of the gate—whenever a gate is connected to one of
the k inputs or to a NOT-gate, ν(i) contains the player corresponding to the appropriate
positive or negative literal. Finally, for each literal player i ∈ Nx ∪ Nx̄ , let ν(i) contain the
appropriate input player and the output player o. For every player i ∈ N, let Ai = {1, 0},
where 1 and 0 can be interpreted as truth values. Finally define the payoff functions as
follows:

 For input players i ∈ Ni , let pi (aN ) = 1 for every action profile aN ∈ AN .

 For positive literal players i ∈ Nx , let pi (aN ) = 1 if in action profile aN ∈ AN , i

plays the same action as the input player j ∈ ν(i) and the output player o plays 1,
or if i plays 1 and o plays 0; pi (aN ) = 0 otherwise.

 For negative literal players i ∈ Nx̄ , let pi (aN ) = 1 if in action profile aN ∈ AN , i

plays the opposite of the action of the input player j ∈ ν(i) and the output player o
plays 1, or if i plays 1 and o plays 0; pi (aN ) = 0 otherwise.
106 6 · Graphical Games

aν(i) aν(i)
11 01 10 00 11 01 10 00
1 1 1 1 1 0 1 1 0 1 1 1
0 1 0 0 0 1 0 0 1 0 0 0

Figure 6.1: Payoffs pi (aN ) for input, positive literal, and negative literal players, used
in the proof of Theorem 6.3. The latter two depend on the actions of the corresponding
input player and the output player o.

aν(i) aν(i)
11 01 10 00 11 01 10 00
1 1 0 0 0 1 1 1 1 0
0 0 1 1 1 0 0 0 0 1

Figure 6.2: Payoffs pi (aN ) for AND and OR players, used in the proof of Theorem 6.3.
Payoffs depend on the actions of players corresponding to the inputs of the gate.

 For AND players i ∈ N∧ , let pi (aN ) = 1 if i and both players j ∈ ν(i) play 1, or if i
and at least one j ∈ ν(i) play 0; pi (aN ) = 0 otherwise.

 For OR players i ∈ N∨ , let pi (aN ) = 1 if i plays 1 and at least one player j ∈ ν(i)
plays 1, or if g and both players j ∈ ν(i) play 0; pi (aN ) = 0 otherwise.

The payoff functions for the different types of players are summarized in Figures 6.1
and 6.2. It is readily appreciated that each player of Γ has at most two neighbors and
two different actions, that all payoff functions take at most two different values, and that
for a particular Boolean circuit the neighborhood graph and the payoff functions can be
constructed in polynomial time.
Now consider a pair of a Boolean circuit C and the corresponding game Γ . We claim
that C is satisfiable if and only if Γ has a pure Nash equilibrium. For the direction from left
to right, assume that C has a satisfying assignment φ, and consider an action profile aN
of Γ where (i) each input player plays according to φ, (ii) each literal player correctly
reproduces the action of the corresponding input player, i.e., positive literals play the
same action as their input, negative ones playing the opposite action, and (iii) each gate
player correctly implements the truth function of the respective gate depending on the
inputs, i.e., actions of his neighbors. By construction of Γ , and since φ is a satisfying
assignment, the output player plays action 1 in aN , and each player receives a payoff of 1.
Since 1 is the maximum payoff, aN is a Nash equilibrium.
For the direction from right to left, we use the following properties of action profiles
of Γ :

1. A profile where all literal players play action 1 and the output player o plays 0 cannot
6.3 · A Tight Hardness Result for Pure Equilibria 107

be a pure Nash equilibrium. Since no negations occur above the literal players, any
gate player j ∈ N∧ ∪ N∨ with ν(j) ⊆ (Nx ∪ Nx̄ ) who plays 0 could increase his
payoff by playing 1. By induction over the structure of the gate this in fact holds
for all gate players, and for o in particular. This is a contradiction.

2. A profile where o plays 0 cannot be a pure Nash equilibrium. In this case, any literal
player not playing 1 could improve his payoff by playing 1, contradicting Property 1.

3. A profile where o plays 1 and some literal player i ∈ Px ∪ Px̄ plays an action that
does not correctly implement the value of the corresponding input cannot be a
Nash equilibrium. In this case, i could change his action to increase his payoff, a
contradiction.

4. A profile in which some gate player i ∈ P∧ ∪ P∨ plays an action that does not
correctly implement the Boolean function of the corresponding gate cannot be a
Nash equilibrium. In this case, i could change his action to increase his payoff, a
contradiction.

5. A profile where the input players do not play a satisfying assignment but o plays 1
cannot be a Nash equilibrium. By Property 4, all gate players would have to play ac-
cording to the Boolean function they implement. In particular, since the assignment
is not satisfying, player o would play 0, contradicting Property 2.

By combining Properties 1 to 4, we conclude that every pure Nash equilibrium of Γ

corresponds to a satisfying assignment of C, and thus there is a one-to-one correspondence
between satisfying assignments of C and pure Nash equilibria of Γ .

The proof of Theorem 6.3 also shows the following.

Corollary 6.4. Counting the pure Nash equilibria of a graphical game is #P-
complete. Hardness holds even for games with two-bounded neighborhoods, two ac-
tions for each player, and two-valued payoff functions.

Obviously, deciding the existence of a pure Nash equilibrium is trivial if players only
have one action or if payoff functions are single-valued. Hence, the only interesting case
that remains concerns games with one-bounded neighborhood.
The interaction among players of a graphical game Γ with neighborhood ν can be rep-
resented as a neighborhood graph G(Γ ) = (N, E), with vertices corresponding to players
of N and directed edges to the neighbors of each player, i.e., (i, j) ∈ E if j ∈ ν(i). The
neighborhood graph of a game with one-bounded neighborhood is a (directed) pseudo-
forest, i.e., a graph where every vertex has outdegree at most one. Each component of
such a graph can be obtained by taking a rooted tree with edges oriented towards the
root, and possibly adding an additional edge from the root to some other vertex. The
following lemma states that the existence of a pure Nash equilibrium of a game whose
108 6 · Graphical Games

neighborhood graph is a pseudoforest only depends on players corresponding to vertices

on cycles of the neighborhood graph.

Lemma 6.5. Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a graphical game with neighborhood ν
such that G(Γ ) is a pseudoforest. Denote by Γ 0 the game obtained by restricting
Γ to those players whose corresponding vertex in G(Γ ) lies on directed cycle, i.e.,
Γ 0 = (N 0 , (Ai )i∈N 0 , (pi )i∈N 0 ) where N 0 = { i ∈ N : there is a path from i to i in G(Γ ) }.
Then, Γ has a pure Nash equilibrium if and only if Γ 0 has a pure Nash equilibrium.

Proof. The implication from left to right is trivial. For the direction from right to left,
consider a pure equilibrium aN 0 of Γ 0 and define an action profile a
N of Γ as follows. For
0 0
all i ∈ N , let ai = ai . As long as there exists a player i ∈ N for which ai has not yet been
defined, find such a player such that aj is defined for all j ∈ ν(i), and define ai as an action
that maximizes the payoff of player i given aν(i) . Obviously this procedure terminates
after a finite number of steps. Furthermore, since the subgraph of G(Γ ) induced by N \ N 0
does not contain a directed cycle, it is readily appreciated that aN is a Nash equilibrium
of Γ .

We are now ready to classify the complexity of the pure Nash equilibrium problem for
games with one-bounded neighborhoods.

Theorem 6.6. Deciding whether a graphical game with one-bounded neighborhoods

has a pure Nash equilibrium is NL-complete. Hardness holds even for games with
two-valued payoff functions.

Proof. For membership in NL, let Γ be a graphical game with one-bounded neighborhood.
By the observation that G(Γ ) is a pseudoforest, and by Lemma 6.5, it is sufficient
to decide whether every game induced by the members of a cycle of G(Γ ) has a pure
equilibrium. Since every vertex of G(Γ ) has outdegree at most one, these cycles can be
found in deterministic logarithmic space.
For the players on a particular cycle, we can guess an action profile in the following way.
We start at an arbitrary deterministically chosen vertex on the cycle and guess an action
for the player corresponding to this vertex. We then traverse the cycle backwards, guessing
an action for each player and checking whether this action maximizes the player’s payoff
given the action of the next player. The computation ends when the initially chosen
vertex has been reached, accepting if and only if the initially chosen action equals the
action guessed in the last step of the traversal. Note that the traversal can be done in
deterministic logarithmic space, and that we only need to maintain a constant number
of pointers. Thus, the whole computation can be performed by a Turing machine with
logarithmic space.
For hardness, we reduce the NL-complete reachability problem for directed graphs, i.e.,
the problem of deciding whether for a given directed graph G = (V, E) and two designated
vertices s, t ∈ V there exists a path from s to t in G (e.g., Papadimitriou, 1994a), to the
6.3 · A Tight Hardness Result for Pure Equilibria 109

pure equilibrium problem for graphical games with one-bounded neighborhoods. Without
loss of generality, assume that (i) V = {1, 2, . . . , n} with n > 4, s = 1, and t = n, that
(ii) every vertex i with 1 6 i 6 n − 1 has outdegree at least one, and that (iii) the only
edge leaving n is a self-loop.
Define a graphical game Γ = (N, (Ai )i∈N , (pi )i∈N ) with players N = {1, 2, . . . , n − 1}
and neighborhoods ν(1) = {n − 1} and ν(i) = {i − 1} for 2 6 i 6 n − 1. The set of possible
actions is the same for all players and given by { aij : 1 6 i, j 6 n }, where aij can be
interpreted as selection of the (possibly non-existing) edge (i, j) in graph G. For an action
profile aN , define the payoffs as follows:

 The payoff function p1 of player 1 is such that p1 (aN ) = 1 if (i) player 1 plays
a1 = a1k for some (1, k) ∈ E and player n − 1 plays an−1 = ajn for some j, or if
(ii) player 1 plays a1 = ann and player n − 1 plays an−1 = ajk such that k 6= n;
p1 (aN ) = 0 otherwise.

 For players i ∈ N with 2 6 t 6 n − 1, the payoff function pi closely resembles the

transition matrix of G. More precisely, pi (aN ) = 1 if i plays ai = ajk for some
(j, k) ∈ E and i − 1 plays ai−1 = a`j for some `; pi (aN ) = 0 otherwise.

We claim that there exists a path from vertex 1 to vertex n in G if and only if Γ has a
pure equilibrium.
For the direction from left to right, assume that there exists a path 1 = v1 , v2 , . . . , vn =
n of length n in G. A shorter path from 1 to n can be extended to this length by virtue
of the self-loop at n. Now consider an action profile aN of Γ where ai = avi vi+1 for each
i ∈ N. It is easily verified that in this case, each player receives the maximum payoff of 1,
such that aN is a Nash equilibrium.
For the direction from right to left, we need to show that a Nash equilibrium of Γ
yields a path connecting vertices 1 and n in G. We exploit the following properties of
action profiles of Γ :

1. An action profile where some player i with 2 6 i 6 n − 1 plays ak` and player i − 1
plays aij for some j 6= k cannot be a Nash equilibrium. In this case, player i would
obtain a payoff of 0, and by construction there exists an alternative action ak` with
(k, `) ∈ E with payoff 1, because every vertex of G has outdegree at least one. This
is a contradiction.

2. An action profile where some player i with 1 6 i 6 n − 2 plays akn for some
1 6 k 6 n and some player j with i + 1 6 j 6 n − 1 does not play ann cannot be
a Nash equilibrium. Assume without loss of generality that j is the smallest such
number. Then player j could increase his payoff by playing ann , a contradiction.

3. An action profile where player n − 1 plays ak` with ` 6= n cannot be a Nash equilib-
rium. In this case the best response for player 1 would be to play ann , contradicting
Property 2.
110 6 · Graphical Games

4. An action profile where player 1 plays aij with i 6= 1 cannot be a Nash equilibrium.
We distinguish two different cases. If player n − 1 played akn for some k, player 1
could increase his payoff by playing a1j , a contradiction. If instead player n − 1
played ak` for some ` 6= n, player 1 could improve his payoff by playing ann ,
contradicting Property 2.
By combining these properties, we conclude that the only action profiles that are pure
Nash equilibria of Γ are those where (i) player 1 plays an action a1j with (1, j) ∈ E, (ii) each
pair of players i − 1 and i with 2 6 i 6 n − 2 play actions ajk and ak` , respectively, for
some 1 6 j, k, ` 6 n, with (k, `) ∈ E, and (iii) player n − 1 plays an action akn with
(k, n) ∈ E. There thus is a one-to-one correspondence of paths between vertices 1 and n
in G and pure Nash equilibria of Γ . We further observe that for a particular graph G, the
payoff functions can be constructed in logarithmic space and take two different values,
and that each player has at most one neighbor.

An immediate consequence of the above proof is given next.

Corollary 6.7. Counting the pure Nash equilibria of a game with one-bounded
neighborhoods is #L-complete.
In the hardness part of the proof of Theorem 6.6, the number of actions grows linearly
in the number of vertices of graph G. For games with one-bounded neighborhoods where
the number of actions grows only slowly, the pure equilibrium problem turns out to be
L-complete.
Theorem 6.8. Deciding whether a graphical game with one-bounded neighborhoods
and at most log log n + O(1) actions has a pure Nash equilibrium is L-complete under
constant-depth reducibility. Hardness holds even for games with two actions and
two-valued payoff functions.
Proof. The proof for membership follows similar lines as the corresponding part of the
proof for Theorem 6.6. However, deterministic logarithmic space suffices to decide the
existence of a pure equilibrium in a graphical game whose neighborhood graph is a cycle.
When traversing the cycle backwards, we can write down all best responses of a particular
player within the space bound, since the overall number of actions is small.
Hardness can be shown by a straightforward reduction from the L-complete reacha-
bility problem for directed graphs with outdegree one (Jones, 1975). Let G = (V, E) be
a graph with this property and two designated vertices s, t ∈ V. Since outgoing edges
of t have no influence on the reachability of t from s, we can assume without loss of
generality that (t, s) ∈ E. Now define a game Γ that has G as its neighborhood graph, i.e.,
Γ = (N, (Ai )i∈N , (pi )i∈N ) where N = V and j ∈ ν(i) if (i, j) ∈ E. Let Ai = {0, 1}, and
define the payoff functions as follows:
 For player t, let pt (aN ) = 1 for any action profile aN ∈ AN with at 6= as , and
pt (aN ) = 0 otherwise.
6.4 · Pure Equilibria of Graphical Games with Anonymity 111

 For any player i 6= t, let pi (aN ) = 1 if ai = aj , pi (aN ) = 0 otherwise, where

j ∈ ν(i).

It is easily verified that this reduction can be computed in constant depth. We further
claim that Γ does not have a pure equilibrium if and only if y is reachable from x in G.
The statement of the theorem then follows by recalling that deterministic logarithmic
space is closed under complementation.
First assume that t is reachable from s in G. Then s and t lie on a cycle in the
neighborhood graph of Γ , and by definition of the payoff functions pi for players i 6= t, all
players on this cycle have to play the same action in every pure equilibrium. Then, by
definition of pt , player t can increase his payoff by changing his action, a contradiction.
Now assume that yt is not reachable from s, and partition V into a set V1 of vertices
from which t is reachable, including t itself, and a set V2 of vertices from which t is not
reachable. Consider the action profile aN where all players in V1 play action 0, and all
players in V2 play action 1. Since s ∈ V2 , and since (t, s) is the only edge originating
from t, it is readily appreciated that pi (aN ) = 1 for every player i ∈ N, such that aN is
a Nash equilibrium.

By combining Theorems 6.6 and 6.8, we obtain an alternative characterization of the

determinism-nondeterminism problem for Turing machines with logarithmic space.

Corollary 6.9. The following statements are equivalent:

1. L = NL.

2. The existence of a pure Nash equilibrium in a graphical game with one-bounded

neighborhoods can be decided in deterministic logarithmic space.

3. For every graphical game Γ with one-bounded neighborhoods, we can construct

in deterministic logarithmic space a graphical game Γ 0 with one-bounded neigh-
borhoods and O(log log n) actions such that Γ 0 has a pure Nash equilibrium if
and only if Γ has a pure Nash equilibrium.

6.4 Pure Equilibria of Graphical Games with Anonymity

Let us now consider games with bounded neighborhoods that additionally satisfy anonymi-
ty within each neighborhood. For neighborhoods of size one, anonymity does not impose
any restrictions. We further know from Theorem 6.6 that the pure equilibrium problem
for such games can be decided in polynomial time. On the other hand, it is not hard
to strengthen Theorem 6.3 to apply to anonymous graphical games as well. While the
game used in the proof of that theorem is not anonymous, an anonymous game can be
obtained by adapting the construction of Schoenebeck and Vadhan (2006) mentioned
earlier to directed neighborhood graphs. Like the one in the proof of Theorem 6.3, this
112 6 · Graphical Games

#(1, aν(i) )
0 1 2 #(1, aν̂(i) ) 0 1 2 3
0 0 0 1 pi (aN ) 0 1 2 0
1 1 1 0

Figure 6.3: Payoffs pi (aN ) for NAND players in the symmetric and the self-symmetric
case, used in the proof of Theorem 6.11. Columns correspond to the different values of
the commutative image of aN with respect to ν(i) or ν̂(i). In the symmetric case, rows
correspond to the two actions of player i.

construction models gates of a Boolean circuit by players of a graphical game, but then has
two additional players play a game with or without a pure equilibrium, depending on the
output of the circuit. By observing that all payoff functions required for this construction
are anonymous, we have the following.

Theorem 6.10 (Schoenebeck and Vadhan, 2006). Deciding whether a graphical game
has a pure Nash equilibrium is NP-complete, even if every player has only two
neighbors, two actions, and two different payoffs, and when restricted to anonymous
games.

6.4.1 Symmetry and Self-Symmetry

We now turn to the more restrictive class of symmetric graphical games, where the
payoff functions of all players are identical. The proof of the following theorem again uses
a construction similar to the one of Schoenebeck and Vadhan (2006). The main difficulty
is to model the two building blocks, a Boolean circuit and games with and without pure
equilibria, using only a single payoff function.

Theorem 6.11. Deciding whether a graphical game has a pure Nash equilibrium is
NP-complete, even if every player has only two actions, and when restricted to
symmetric games with two different payoffs or to self-symmetric games with three
different payoffs.

Proof. Membership in NP is straightforward. We can simply guess an action profile and

verify that the action of each player is a best response to the actions of the players in his
neighborhood.
For hardness, we provide a reduction from CSAT. For a set N of players with appro-
priately defined neighborhoods ν, let Γ (N) = (N, {0, 1}N , (pi )i∈N ) be a graphical game
with payoffs satisfying symmetry or self-symmetry as given in Figure 6.3.2 We observe
the following properties:
2
Also recall that every anonymous or symmetric game with two actions per player can respectively be
reduced to a self-anonymous or self-symmetric game, while preserving pure equilibria.
6.4 · Pure Equilibria of Graphical Games with Anonymity 113

1. Let N be a set of players, |N| = 3, and ν̂(i) = N for all i ∈ N. Then, an action
profile aN of Γ (N) is a pure equilibrium if and only if #(1, aN ) = 2. In particular,
for every i ∈ N, there exists a pure equilibrium where player i plays action 0 and a
pure equilibrium where he plays action 1.

2. Let N and N 0 be two disjoint sets of players with neighborhoods such that for all
i ∈ N, ν(i) ⊆ N, and for all i ∈ N 0 , ν(i) ⊆ N 0 . Then, aN∪N 0 is a pure equilibrium
of Γ (N ∪ N 0 ) if and only if aN and aN 0 are pure equilibria of Γ (N) and Γ (N 0 ),
respectively.

3. Let N be a set of players such that Γ (N) has a pure equilibrium, and consider
two players 1, 2 ∈ N. Further consider an additional player 3 ∈ / N with ν(3) =
{1, 2}. Then the game Γ (N ∪ {3}) has a pure equilibrium, and in every pure equilib-
rium aN∪{c} of Γ (N ∪ {c}), a3 = 0 if a1 = a2 = 1, and a3 = 1 otherwise. In other
words, such an action profile always satisfies a3 = a1 NAND a2 .

4. Let N be a set of players and consider a particular player 1 ∈ N. Further consider

five additional players 2, 3, 4, 5, 6 ∈
/ N with neighborhoods according to Figure 6.4,
and denote N 0 = N ∪ {2, 3, 4, 5, 6}. Then Γ (N 0 ) has a pure equilibrium if and only if
Γ (N) has a pure equilibrium aN with a1 = 0. For the direction from right to left,
assume that Γ (N) has a pure equilibrium aN where a1 = 0, and extend it to an
action profile aN 0 for Γ (N 0 ) by letting a2 = 0 and a3 = a4 = a5 = a6 = 1. It is
easily verified that aN 0 is a Nash equilibrium of Γ (N 0 ). For the direction from left to
right, consider an action profile aN 0 for Γ (N 0 ) where a1 = 1. If a2 = 0, then action 1
is the unique best response for players 4 and 5, after which action 0 is the unique
best response for players 3 and 6. In this case, player 2 could change his action to
get a higher payoff. If a2 = 1, then the unique best response for players 4 and 5 is 0,
and consequently the unique best response for players 3 and 6 is 1. Again, player 2
could change his action to get a higher payoff.

5. Let N1 = {1, 2, 3} be an instance of N in Property 1, and N2 an instance of N 0 in

Property 4 with N = {1}. Let N be any set of players such that Γ (N) has a pure
equilibrium, let 4 ∈ N, and denote N 0 = N1 ∪N2 ∪N. Further consider an additional
player 5 ∈/ N 0 with ν(5) = {2, 4}. Then, Γ (N 0 ∪ {5}) has a pure equilibrium, and in
every pure equilibrium aN 0 ∪{5} of Γ (N 0 ∪ {5}), a5 = 1 − a4 . To see this, observe that
by Property 1 exactly two players in N1 must play action 1, which, by Property 4,
have to be players 2 and 3. Since φ NAND true = ¬φ, the claim follows from
Property 3.

Now consider an instance C of CSAT, and assume without loss of generality that C
consists exclusively of NAND gates and that no variable appears more than once as the
input to the same gate. The latter assumption can be made by Property 5. We construct a
game Γ = Γ (N) as follows. For every input of C we augment N by three players according to
114 6 · Graphical Games

6 3
2

5 4

Figure 6.4: Output gadget, used in the proof of Theorem 6.11. All players have payoffs
as in Figure 6.3. Player x must play action 0 in every pure equilibrium of the game.

Property 1. We then inductively define Γ by adding, for a gate with inputs corresponding
to players 1, 2 ∈ N, a player 3 as described in Property 3. Finally, we construct a player
according to Property 5 who plays the opposite action as the one corresponding to the
output of C, and identify this player with player 1 in a new instance of N 0 in Property 4.
It is now easily verified that a pure equilibrium of Γ corresponds to a computation of C
which outputs true, and that such an equilibrium exists if and only if C has a satisfying
assignment.

6.4.2 Self-Anonymity and Two Different Payoffs

Since self-symmetry implies self-anonymity, Theorem 6.11 also implies NP-hardness in
the self-anonymous case. The result is not tight, however, in that three different payoffs
are required for hardness. A natural question to ask is what happens for self-anonymous
games with only two different payoffs. In this section we prove a tight result for the most
restricted version of self-anonymity, i.e., the case with only two different payoff functions.
The problem with anonymity and the construction used in the proof of Theorem 6.11
is that two different payoffs are not enough to make a player care about his own action irre-
spective of the actions played by his neighbors. With four different values for #(1, aν̂(i) ),
there will either be an equilibrium where all players play the same action, or a situation
where a player is indifferent between both of his actions. When we want to use games
to compute a function, such indifference is clearly undesirable. The key idea that will
enable us to prove the following theorem is to isolate pure equilibria that are themselves
symmetric in the actions of a subset of the players, i.e., equilibria in which these players
all play the same action. To enforce that two particular players play the same action in
every equilibrium, we will add two additional players, each of which observes the other as
well as one of the original players. Depending on the actions of the original players, the
new players will either play a game with a unique pure equilibrium, or a game that is pro-
totypical both for self-anonymous games and for games without pure equilibria, namely
Matching Pennies.

Theorem 6.12. Deciding whether a graphical game has a pure Nash equilibrium is
NP-complete, even if every player has only two neighbors, two actions, and two
6.4 · Pure Equilibria of Graphical Games with Anonymity 115

3 4
#(1, aν̂(i) ) 0 1 2 3
pi (aN ) 0 1 0 1
1 2

Figure 6.5: Equality gadget, used in the proof of Theorem 6.12. A pure equilibrium exists
if and only if players 1 and 2 play the same action.

different payoffs, and when restricted to self-anonymous games with two different
payoff functions.

Proof. Membership in NP is again straightforward.

For hardness, we again give a reduction from CSAT. Let Γ (N) = (N, {0, 1}N , (pi )i∈N )
denote a graphical game for a set N of players with neighborhood ν and payoff functions pi
satisfying self-anonymity. We observe the following properties:

1. Let N be a set of players, 1, 2 ∈ N, and consider two additional players 3, 4 ∈ / N with

neighborhoods and payoffs according to Figure 6.5. We claim that Γ (N ∪ {3, 4}) has
a pure equilibrium if and only if Γ (N) has a pure equilibrium aN where a1 = a2 .
For the direction from right to left, assume that Γ (N) has such an equilibrium, and
extend it to an action profile aN 0 for Γ (N 0 ) by letting a3 = 0 and a4 = 1. It is easily
verified that under this action profile players 3 and 4 both receive the maximum
payoff of 1, such that the equilibrium condition is trivially satisfied. For the direction
from left to right, assume that one of players 3 or 4 observes action 0 being played
by player 1 or player 2, while the other one observes action 1. Then players 3 and 4
effectively play the well-known Matching Pennies game. More precisely, the player
observing 0 receives a payoff of 1 if and only if #(1, a{3,4} ) is odd, while the same is
true for the player observing 1 if and only if this number is even. Since both players
can change between the two outcomes by changing their own action, there cannot
be a pure equilibrium.

2. Let N = {1, 2, 3} with ν(i) = N for all i ∈ N, and let pi be defined according to
Figure 6.5. It is then easily verified that aN with a1 = a2 = a3 = 1 or with
a1 = a2 = 0 and a3 = 1 is an equilibrium of Γ (N). In particular, there exists an
equilibrium where player 1 plays 0, and one where player 1 plays 1.

3. Let N and N 0 be two disjoint sets of players with neighborhoods such that for all
i ∈ N, ν(i) ⊆ N, and for all i ∈ N 0 , ν(i) ⊆ N 0 . Again, aN∪N 0 is a pure equilibrium
of Γ (N ∪ N 0 ) if and only if aN and aN 0 are pure equilibria of Γ (N) and Γ (N 0 ),
respectively.

4. Let N = {1, 2, 3} with neighborhoods and payoffs as in Property 2, and assume

by Property 1 that every pure equilibrium aN of Γ (N) is symmetric, i.e., satisfies
116 6 · Graphical Games

= =
3 4 5 #(1, aν̂(i) ) 0 1 2 3
pi (aN ) 0 1 1 0
1 2

Figure 6.6: NAND gadget, used in the proof of Theorem 6.12. The construction of
Figure 6.5 is used to ensure that players connected by “=” play the same action in every
pure equilibrium.

a1 = a2 = a3 . Then, aN with a1 = a2 = a3 = 1 is the unique pure equilibrium

of Γ (N). Clearly, aN is an equilibrium of Γ (N), since all players receive the maximum
payoff of 1. In the only other symmetric action profile, all players play action 0 and
receive a payoff of 0, and each of them could change his action to 1 to receive a
higher payoff.

5. Let N be a set of players such that Γ (N) has a pure equilibrium, let 1, 2 ∈ N,
and consider three additional players 3, 4, 5 ∈ / N with neighborhoods and payoffs
according to Figure 6.6. Then, Γ (N ∪ {3, 4, 5}) has a pure equilibrium, and for every
pure equilibrium aN∪{3,4,5} of Γ (N ∪ {3, 4, 5}), a3 = 0 if a1 = a2 = 1, and a3 = 1
otherwise. To see this, observe that players 3, 4, and 5 get the maximum payoff
of 1 under any action profile satisfying a3 = a4 = a5 = 1 and #(1, a{1,2} ) 6 1, or
a3 = a4 = a5 = 0 and a1 = a2 = 1. On the other hand, by Property 1, an action
profile cannot be an equilibrium unless a3 = a4 = a5 . If a1 = a2 = a5 = 0 or
a1 = a2 = a5 = 1, then player 5 can change his action to receive a higher payoff. If
otherwise a1 6= a2 and a3 = a4 = 0, then there exists a player i ∈ {3, 4} such that
#(1, aν̂(i) ) = 0, who could change his action to get a higher payoff.

6. Let N be a set of players, 4 ∈ N. Let N 0 = {1, 2, 3} with neighborhoods as in

Property 4, N 00 = {5, 6} with ν(5) = {1, 6} and ν(6) = {4, 5}. Then, Γ (N ∪ N 0 ∪ N 00 )
has a pure equilibrium if and only if Γ (N) has a pure equilibrium aN with a4 = 1.
Clearly, an action profile that is not an equilibrium of Γ (N) cannot be extended
to an equilibrium of Γ (N ∪ N 0 ∪ N 00 ). Conversely assume that aN∪N 0 ∪N 00 is an
equilibrium of Γ (N ∪ N 0 ∪ N 00 ). Then, by Property 4, a1 = 1. Furthermore, by
Property 1, a1 = a4 , and thus a4 = 1.

Now consider an instance C of CSAT, and assume without loss of generality that C
consists exclusively of NAND gates. Since φ NAND true = ¬φ, and using Property 4,
we can further assume that no variable appears more than once as an input to the same
gate. We construct a game Γ = Γ (N) as follows: For every input of C, we add three players
according to Property 2. For every gate of C with inputs corresponding to players 1, 2 ∈ N,
we add three players according to Property 5. Finally, we add five players according to
Property 6, such that player 4 is the one that corresponds to the output of C. It is now
6.4 · Pure Equilibria of Graphical Games with Anonymity 117

1 2
2

7 3
4 3
7

6 4
5 1 6 5

Figure 6.7: Neighborhood graph of a graphical game with seven players (left), correspond-
ing to the three-uniform square hypergraph given by the lines of the Fano plane (right).
A directed edge from vertex i to vertex j of the neighborhood graph denotes that j ∈ ν(i).

readily appreciated that Γ has a pure equilibrium if and only if C is satisfiable.

6.4.3 Self-Symmetry and Two Different Payoffs

Let us return to self-symmetric graphical games. Self-symmetric games as studied in
Chapter 5 always possess a pure Nash equilibrium due to the fact that they are common-
payoff games. This existence result does not hold for self-symmetric graphical games, even
when there are only two different payoffs. In particular, there exists a seven-player game
in the latter class that does not have a pure equilibrium, and in which each player has
exactly two actions and two neighbors. It will be instructive to view a graphical game as a
hypergraph, with each vertex corresponding to a player and each edge to the set of players
in the neighborhood of one particular player including the player himself. Corresponding
to the set of games with neighborhoods of size m is the set of (m+1)-uniform hypergraphs
that possess a matching in the sense of Seymour (1974), i.e., a bijection from the set of
vertices to the set of edges that maps every vertex to an edge containing it. It is not
hard to see that a self-symmetric game with two actions and payoffs pi = (0, 1, 1, 0) for
all i ∈ N has a pure Nash equilibrium if and only if the corresponding hypergraph is vertex
two-colorable. Given a two-coloring, every player observes either one or two players in
his neighborhood, including himself, who play action 1, and thus obtains the maximum
payoff of 1. If on the other hand there is no two-coloring, then there is at least one player
for every action profile who plays the same action as all of his neighbors and can deviate
to obtain a higher payoff. Figure 6.7 shows the neighborhood of a graphical game with
seven players and two neighbors for each player. This graph induces the 3-uniform square
hypergraph corresponding to the lines of the Fano plane, which in turn cannot be two-
colored (e.g., Seymour, 1974). We leave it to the avid reader to verify that there is no
game with the above properties and less than seven players.
An interesting property of the neighborhood graph on the left of Figure 6.7 is that
it does not have any cycles of even length. We will begin our investigation of the pure
118 6 · Graphical Games

equilibrium problem in self-symmetric games by generalizing this observation to games

with arbitrary neighborhoods and pi = (0, 1, 1, . . . , 1, 0) for all i ∈ N. The following
lemma characterizes games with pure equilibria in the above subclass in terms of cycles
in the neighborhood graph. Seymour (1974) provides a similar characterization of the
minimal uniform square hypergraphs that do not have a two-coloring.
Lemma 6.13. Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a self-symmetric graphical game with
|Ai | = 2 and pi = (0, 1, 1, . . . , 1, 0) for all i ∈ N. Then, Γ has a pure Nash equilibrium
if and only if for all i ∈ N, there exists j ∈ N reachable from i such that j lies on a
cycle of even length.
Proof. For the implication from left to right, assume that there exists a pure equilibrium,
i.e., a two-coloring c : N → {0, 1} of the neighborhood graph such that the neighborhood
of every player contains some player playing action 0 and some player playing action 1.
Now consider an arbitrary player v1 ∈ N. Using the above property of c, we can construct
a path v1 , v2 , . . . , v|N|+1 , vi ∈ N, such that for all i, 1 6 i 6 |N|, c(vi ) = 1 − c(vi+1 ). By
the pigeonhole principle, there must exist i, j, 1 6 i < j 6 |N| + 1, such that vi = vj and
for all j 0 , i < j 0 < j, vj 0 6= vi . Then, vi , vi+1 , . . . , vj is a cycle of even length reachable
from v1 .
For the implication from right to left, let N 0 ⊆ N be a set of players such that for
every i ∈ N there exists a directed path to some j ∈ N 0 , and such that N 0 induces a
set of vertex-disjoint cycles of even length. We construct a two-coloring c : N → {0, 1},
corresponding to an assignment of actions to players, as follows. First color the members
of N 0 such that for all i ∈ N 0 and j ∈ ν(i) ∩ N 0 , c(i) = 1 − c(j). While there are uncolored
vertices left, find i, j ∈ N such that j ∈ ν(i), i is uncolored, and j is colored. Such a pair of
vertices must always exist, since for every member of N there is a directed path to some
member of N 0 , and thus to a vertex that has already been colored. Color i such that
c(i) = 1 − c(j). It is now easily verified that at any given time, and for all i ∈ N that have
already been colored, there exist j, j 0 ∈ ν̂(i) with c(j) = 0 and c(j 0 ) = 1. If all vertices have
been colored, then every neighborhood will contain at least one player playing action 0,
and at least one player playing action 1. The corresponding action profile is a pure Nash
equilibrium.

Thomassen (1985) has shown that for every natural number m, there exists a directed
graph without even cycles where every vertex has outdegree m. On the other hand, it is
easy to construct graphs that do have even cycles. Together with Lemma 6.13, we thus
have that the pure equilibrium problem for the considered class of games is nontrivial.
Corollary 6.14. For every m ∈ N, m > 0, there exist self-symmetric graphical
games Γ and Γ 0 with two actions and |ν(i)| = m and pi = (0, 1, 1, . . . , 1, 0) for all
i ∈ N, such that Γ has a pure Nash equilibrium and Γ 0 does not.
We are now ready to identify several classes of graphical games where the existence of
a pure equilibrium can be decided in polynomial time.
6.4 · Pure Equilibria of Graphical Games with Anonymity 119

Theorem 6.15. Let Γ be a self-symmetric graphical game with payoffs pi . The pure
equilibrium problem for Γ can be decided in polynomial time if one of the following
properties holds:

(i) for all i ∈ N, pi (0) > pi (1) or for all i ∈ N, pi (|ν̂(i)|) > pi (|ν̂(i)| − 1);

(ii) for all i ∈ N and all m with 1 6 m 6 |ν(i)|, pi (m − 1) > pi (m) and pi (m + 1) >
pi (m), or pi (m − 1) < pi (m) and pi (m + 1) < pi (m);

(iii) for all i ∈ N and all m with 1 6 m < |ν(i)|, pi (m) = pi (m + 1).

Proof. It is easy to see that a game Γ satisfying (i) possesses a pure equilibrium aN such
that #(0, aN ) = 0 or #(1, aN ) = 0.
For a game Γ satisfying (ii), we observe that in every equilibrium aN , pi (aN ) = 1 for
all i ∈ N. The pure equilibrium problem for Γ thus corresponds to a variant of generalized
satisfiability, with clauses induced by neighborhoods of Γ . The constraints associated with
this particular variant require an odd number of variables in each clause to be set to true,
and can be written as a system of linear equations over GF (2). Tractability of the pure
equilibrium problem for Γ then follows from Theorem 2.1 of Schaefer (1978).
Finally, a game satisfying (iii) but not (i) can be transformed into a best response
equivalent one that satisfies the conditions of Lemma 6.13. We further claim that we
can check in polynomial time whether for every i ∈ N, there exists j ∈ N on a cycle of
even length and reachable from i. For a particular i ∈ N, this problem is equivalent to
checking whether the subgraph induced by the vertices reachable from i contains an even
cycle. The latter problem has long been open, but was recently shown to be solvable in
polynomial time (Robertson et al., 1999).

It is easily verified that every self-symmetric graphical game Γ with two different
payoffs and neighborhoods of size two or three can be transformed into a game Γ 0 with
the same set of players and the same neighborhoods, such that Γ and Γ 0 have the same
set of pure equilibria and Γ 0 satisfies one of the conditions of Theorem 6.15. We thus have
the following.

Corollary 6.16. The problem of deciding whether a self-symmetric graphical game

with two different payoffs and three-bounded neighborhoods has a pure equilibrium is
in P.

6.4.4 Self-Symmetry and Larger Neighborhoods

The question that remains is whether the pure equilibrium problem can be solved in
polynomial time for all self-symmetric graphical games with two payoffs, or whether there
is some bound on the neighborhood size where this problem again becomes hard. We will
show in this section that the latter is true, and that the correct bound is indeed four, as
suggested by Corollary 6.16.
120 6 · Graphical Games

1 2

8 3
#(1, aν̂(i) ) 0 1 2 3 4 5
pi (aN ) 0 1 0 1 1 0
7 4

6 5

Figure 6.8: Neighborhood graph and payoffs of a graphical game with eight players and
neighborhoods of size four, used in the proof of Theorem 6.17. The neighborhood graph
satisfies rotational symmetry, the neighborhood of player 1 is highlighted.

We will essentially use the same tools as in Section 6.4.2, but extract the necessary
complexity from only a single payoff function. The additional insight required for this
extraction is that “constant” players, i.e., players who play the same action in every pure
equilibrium of a game, can be used to prune a larger payoff table and effectively obtain
different payoff functions for smaller neighborhoods that can then be used to proceed with
the original proof. Constructing such players will prove a rather difficult task in its own
right.

Theorem 6.17. Deciding whether a self-symmetric graphical game with two different
payoffs has a pure Nash equilibrium is NP-complete, even if every player has exactly
four neighbors.

Proof. Membership in NP is straightforward. We can simply guess an action for each

player and then verify that no player can increase his payoff by playing a different action
instead.
For hardness, we again give a reduction from CSAT to the problem at hand. The
central idea of this proof will be to guarantee that some players in a neighborhood only
play certain well-defined actions in equilibrium. By this, the original payoff table is
effectively “pruned” to a smaller one that can then be used, like in earlier proofs, to model
the behavior of gates in a Boolean circuit.
As a first step, we will show how to construct “constants,” i.e., players who play the
same action, 0 or 1, in every equilibrium of a game. To achieve this, we will construct
a set of four players, such that in every equilibrium two of these players play action 0
and two of them play action 1. A player observing these four players can determine if
the number of players in his neighborhood, including himself, who play action 1 is two
or three. Clearly, such a player will play action 1 in every equilibrium. By a similar
argument, a player who observes four players who play action 1 in every equilibrium will
himself play action 0 in every equilibrium.
Consider the graphical game Γ with eight players and neighborhoods of size four given
6.4 · Pure Equilibria of Graphical Games with Anonymity 121

in Figure 6.8. We will argue that the pure equilibria of this game are exactly those action
profiles where two players, with distance two on the outer cycle, play action 0. More
formally, an action profile aN is an equilibrium if and only if there is one pair of distinct
players i, j ∈ N such that ai = aj = 0, and it holds that i − j = 2 (mod 4).
For the direction from right to left, we can exploit rotational symmetry of the neigh-
borhood graph and assume without loss of generality that a1 = a3 = 0. The statement
then follows by checking that the equilibrium condition is satisfied for all players. For the
direction from left to right, we exploit the following properties of the neighborhood graph:
1. For any distinct i, j ∈ N with i − j 6= 2 (mod 4), there exist a player ` ∈ N such that
ν̂(`) ∩ {i, j} = ∅. Assuming without loss of generality that i = 1, the property follows
from a case analysis.

2. For any N 0 ⊆ N, |N 0 | = 3, there exists a player i ∈ N such that N 0 ⊆ ν̂(i). Due

to rotational symmetry, we can assume without loss of generality that 1 ∈ N 0 . The
property then follows by a straightforward if somewhat tedious case analysis.

3. For any N 0 ⊆ N, |N 0 | = 3, there exists a player i ∈ N such that |N 0 ∩ ν̂(i)| = 2.

Showing this property is again straightforward by assuming without loss of generality
that 1 ∈ N 0 and showing that for any pair of distinct players, there exists a player
i ∈ N such that either ν̂(i) contains player 1 and exactly one element of the pair, or
both elements of the pair but not player 1.

4. For any N 0 ⊆ N, |N 0 | = 4, there exists a player i ∈ N such that |N 0 ∩ ν̂(i)| = 3. To

show this property, we can again assume without loss of generality that 1 ∈ N 0 , and
distinguish neighborhoods that contain player 1 from neighborhoods that do not.
The analysis is again straightforward.
Now consider an equilibrium aN of Γ , and observe that due to the structure of the payoffs,
it must be the case that pi (aN ) = 1 for all i ∈ N. If #(0, aN ) < 2 or #(1, aN ) < 2, then
there exists a player i ∈ N such that #(0, aν̂(i) ) = 0 or #(1, aν̂(i) ) = 0, contradicting the
assumption that aN is an equilibrium. If #(0, aN ) = 2, assume without loss of generality
that a1 = 0, and consider i ∈ N \ {1, 3, 7} such that ai = 0. Then, by Property 1, there
exists a player j ∈ N such that #(0, aν̂(j) ) = 0, which again leads to a contradiction. If
#(0, aN ) = 3, then by Property 2 there must exist a player i ∈ N such that #(0, aν̂(i) ) = 3
and thus #(1, aν̂(i) ) = 2, again contradicting the assumption that aN is an equilibrium.
By Property 4, the same holds if #(0, aN ) = 4. If #(0, aN ) = 5 and thus #(1, aN ) = 3,
then by Property 3 there must yet again exists a player i ∈ N such that #(1, aν̂(i) ) = 2,
a contradiction. The same trivially holds if #(1, aN ) = 2.
Now we augment Γ by a set {9, 10, . . . , 13} of five additional players such that


{1, 3, 5, 7}
 if i ∈ {9, 10}
ν(i) = {2, 4, 6, 8} if i ∈ {11, 12}


{9, 10, 11, 12} if i = 13.
122 6 · Graphical Games

By construction of the original game with eight players, every pure equilibrium has ei-
ther two or four players in the common neighborhood of players 9 and 10 play action 1.
Furthermore, if players 9 and 10 observe two players who play action 1, then players 11
and 12 will observe four players who play action 1, and vice versa. As a consequence,
either players 9 and 10 will play action 0, and players 11 and 12 will play action 1, or
the other way round. In any case, exactly two players in the neighborhood of player 13
will play action 1 in every equilibrium of the augmented game, and player 13 himself will
therefore play action 1.
In the following, we denote by 01 , 02 , 03 three players who play action 0 in every
equilibrium, and by 11 , 12 two players that constantly play action 1. Using these players
to prune the payoff table, we will then proceed to design games that simulate Boolean
circuits. We want these games to satisfy self-symmetry, and the payoff of all players will
therefore be determined by the table already used above and shown in Figure 6.8. As for
the inputs of the circuit, it is easily verified that a game with players N, |N| = 5, such that
for all i ∈ N, ν̂(i) = N, has pure equilibria aN and aN 0 such that for an arbitrary i ∈ N,

ai = 0 and ai0 = 1.
As before, we will now construct a subgame that simulates a functionally complete
Boolean gate, in this case NOR, and a subgame that has a pure equilibrium if and only if
a particular player plays action 1. For a set N of players with appropriately defined neigh-
borhoods ν, let Γ (N) = (N, {0, 1}N , (pi )i∈N ) be a graphical game with payoff functions pi
satisfying self-symmetry as in Figure 6.8. We observe the following properties:
1. Let N and N 0 be two disjoint sets of players with neighborhoods such that for all
i ∈ N, ν(i) ⊆ N, and for all i ∈ N 0 , ν(i) ⊆ N 0 . Again, aN∪N 0 is a pure equilibrium
of Γ (N ∪ N 0 ) if and only if aN and aN 0 are pure equilibria of Γ (N) and Γ (N 0 ),
respectively.

2. Let N be a set of players such that Γ (N) has a pure equilibrium, let 1, 2 ∈ N,
and consider two additional players 3, 4 ∈ / N with ν(3) = {01 , 02 , 1, 4}, and ν(4) =
{01 , 02 , 2, 3}. Then every pure equilibrium aN∪{3,4} of Γ (N ∪ {3, 4}) satisfies a1 = a2 .

3. Identifying player 2 with 11 in the previous construction, we have that Γ (N ∪ {3, 4})
has a pure equilibrium if and only if a1 = 1 in some pure equilibrium aN of Γ (N).

4. Let N be a set of players such that Γ (N) has a pure equilibrium, let 1, 2 ∈ N,
and consider two additional players 3, 4 ∈ / N with neighborhoods given by ν(3) =
{01 , 02 , 03 , 4} and ν(4) = {01 , 02 , 1, 2}. Then Γ (N ∪ {3, 4}) has a pure equilibrium,
and every pure equilibrium aN∪{3,4} of Γ (N ∪ {3, 4}) satisfies a3 = 1 whenever a1 =
a2 = 0, and a3 = 0 whenever a1 6= a2 . For every pure equilibrium aN∪{3,4} with
a1 = a2 = 1, there exists a pure equilibrium aN∪{3,4} 0 such that a3 6= a30 , and
ai = ai0 for all i ∈ N.

5. Consider an additional player 5 ∈ / N ∪ {3, 4}, and let ν(5) = {11 , 12 , 1, 2}. Then
Γ (N∪{3, 4, 5}) has a pure equilibrium, and every pure equilibrium aN∪{3,4,5} of Γ (N∪
6.5 · Interlude: Satisfiability in the Presence of a Matching 123

=
3 4 5

03 02 01 1 2 11 12

Figure 6.9: NOR gadget, used in the proof of Theorem 6.17. Payoffs are identical to those
in Figure 6.8. A construction analogous to the one shown in Figure 6.5 is used to ensure
that players 3 and 5 play the same action in every pure equilibrium.

{x, y, z}) satisfies a5 = 1 whenever a1 = a2 = 0, and a5 = 0 whenever a1 = a2 = 1.

For every pure equilibrium aN∪{3,4,5} with a1 6= a2 , there exists a pure equilibrium
0
aN∪{3,4,5} such that a5 6= a50 , and ai = ai0 for all i ∈ N.

6. By Property 2, we can assume that every equilibrium aN∪{3,4,5} of Γ (N ∪ {3, 4, 5})

satisfies a3 = a5 , and thus that a3 = 1 if and only if a1 = a2 = 0.
Properties 4 through 6 are illustrated in Figure 6.9.
Now consider an instance C of CSAT, and assume without loss of generality that C
consist exclusively of NOR gates and that no variable appears more than once as an input
to the same gate. The latter assumption can be made since φ NOR false = ¬φ, and
since there exists a self-symmetric game and a player in this game who plays action 0
in every pure equilibrium. As before, we construct a game Γ by simulating every gate
of C according to Property 6 and identifying the player that corresponds to the output of
the circuit with player 1 in Property 3. It is now readily appreciated that Γ has a pure
equilibrium if and only if C is satisfiable.

Observing that in the constructions used in the proofs of Theorems 6.11, 6.12, and 6.17
there is a one-to-one correspondence between satisfying assignments of a Boolean circuit
and pure equilibria of a game, we have that counting the number of pure equilibria in the
respective games is as hard as computing the permanent of a matrix.
Corollary 6.18. For graphical games with neighborhoods of size two, counting the
number of pure Nash equilibria is #P-hard, even when restricted to symmetric graph-
ical games with two different payoffs, to self-anonymous graphical games with two
different payoffs and two different payoff functions, or to self-symmetric graphical
games with three different payoffs. The same holds for self-symmetric graphical
games with neighborhoods of size four and two different payoffs.

6.5 Interlude: Satisfiability in the Presence of a Matching

The analysis at the end of the previous section allows us to derive a corollary that may
be of independent interest. Schaefer (1978) completely characterizes which variants of
124 6 · Graphical Games

the generalized satisfiability problem are in P and which are NP-complete. Some of the
variants become tractable if there exists a matching, i.e., a bijection from variables to
clauses that maps every variable to a clause it appears in. In the case of not-all-equal
3SAT, for example, this follows from equivalence with two-colorability of three-uniform
hypergraphs and from the work of Robertson et al. (1999). On the other hand, the proof
of Theorem 6.17 identifies a variant that is NP-complete and remains so in the presence
of a matching. We thus have the following.

Corollary 6.19. Generalized satisfiability is NP-complete, even when restricted to

instances that have a matching and clauses of size five.

We leave a complete characterization for future work. While the proof techniques
developed in this chapter will certainly be useful in this respect, it should be noted that the
equivalence between generalized satisfiability and the pure equilibrium problem covered
by Theorem 6.17 may fail to hold for instances of the latter where pi (aN ) = pi (aN 0 ) = 0
0 such that #(1, a
for action profiles aN , aN 0
ν̂(i) ) = #(1, aν̂(i) ) + 1. For example, it would
not be possible to show hardness of one-in-three 3SAT (Schaefer, 1978) using the same
approach.

6.6 Mixed Equilibria

Let us now briefly look at the problem of finding a mixed equilibrium. The following
theorem states that this problem is tractable in symmetric graphical games if the number
of actions grows slowly in the neighborhood size. The proof relies on the fact that such
games always have a symmetric equilibrium.

Theorem 6.20. Let Γ = (N, AN , (pi )i∈N ) be a symmetric graphical game such that
for all i ∈ N, |A| = O(log |ν(i)|/ log log |ν(i)|). Then, a Nash equilibrium of Γ can be
computed in polynomial time.

Proof. We show that Γ possesses a symmetric equilibrium, i.e., one where all players play
the same (mixed) strategy, and that this equilibrium can be computed efficiently. For
this, choose an arbitrary player i ∈ N and construct a game Γi = (Ni , ANi , (pi,j )j∈N )
with players Ni = ν̂(i), and payoff functions pi,j such that for all j ∈ Ni and for action
profiles aN ∈ AN and aN 0 0 ) = p (a ) if a 0 = a and for all a ∈ A,
∈ ANi , pi,j (aN
i i i N j i
0
#(a, aNi ) = #(a, aν(i) ).
Since Γ is a symmetric graphical game, it is easily verified that Γi is a symmetric game,
and must therefore possess a symmetric equilibrium, i.e., one where all the players in Ni
play the same strategy. By a result of Papadimitriou and Roughgarden (2005), one such
equilibrium sN 0
i
can be computed in polynomial time if |A| = O(log |N 0 |/ log log |N 0 |).
Moreover, due to the symmetry of Γ , all the games Γi for i ∈ N are isomorphic, and
thus sN0 is an equilibrium in each of them.
i
6.7 · Discussion 125

Now define a strategy profile sN of Γ by letting, for each j ∈ N, sj = si0 , and assume for
contradiction that sN is not an equilibrium. Then there exists a player j ∈ N and some
strategy tj ∈ ∆(A) for this player such that pj (s−i , ti ) > pi (sN ). Further, by definition
0 , t) > p (s 0 ), contradicting the assumption that s 0
of pi,j , pi,i (s−i i,i Ni Ni is an equilibrium
of Γi .

This result applies in particular to the case where both the number of actions and
the neighborhood size are bounded. Since the pure equilibrium problem in symmetric
graphical games is NP-complete even in the case of two actions, we have identified a class
of games where computing a mixed equilibrium is computationally easier than deciding the
existence of a pure one, unless P=NP. A different class of games with the same property
is implicit in Theorem 3.4 of Daskalakis and Papadimitriou (2005). It should be noted, on
the other hand, that the existence of a symmetric equilibrium does not in general extend
to games that are not anonymous or in which players have different payoff functions.

6.7 Discussion
In this chapter we have completely characterized the complexity of deciding the existence
of a pure Nash equilibrium in games with bounded neighborhoods. This problem is NP-
complete in games with neighborhoods of size two, two actions, and two-valued payoff
functions. For neighborhoods of size one it is NL-complete in general, and L-complete if
additionally the number of actions grows only very slowly. Some additional cases become
tractable for games that further satisfy the most restrictive type of anonymity considered
in Chapter 5 within each neighborhood.
For the other types of anonymity, two neighbors again suffice for NP-hardness. While
the construction used in the proof of Theorem 6.17 can be generalized to arbitrary neigh-
borhoods of even size, it is unclear what happens for odd-sized neighborhoods. The
extreme case when the neighborhood of every player consists of all other players yields or-
dinary anonymous games, in which the pure equilibrium problem is in P when the number
of actions is bounded. It remains open at which neighborhood size the transition between
membership in P and NP-hardness occurs. Another open problem concerns the complexity
of the mixed equilibrium problem in anonymous graphical games. A promising direction
for proving hardness would be to make the construction of Goldberg and Papadimitriou
(2006) anonymous. Finally, as suggested in Section 6.5, it would be interesting to study
the complexity of generalized satisfiability problems in the presence of matchings.
 Chapter 7

Quasi-Strict Equilibria

Criticism directed at Nash equilibrium has been a recurring theme in previous chapters.
In the remaining two chapters of the thesis we consider two more solution concepts, each
of them trying to address a particular shortcoming.
Consider again the single-winner game introduced in Chapter 2, in which Alice, Bob,
and Charlie select a winner among themselves using a protocol in which each of them
has to raise their hand or not. The game is repeated in Figure 7.1. As we have already
noted in Chapter 4, the only pure equilibrium of this game, in which Bob raises his hand
while Alice and Charlie do not, is particularly weak. Both Bob and Charlie could deviate
from their respective strategies to any other strategy without decreasing their chances
of winning. After all, they cannot do any worse than losing. A similar property in fact
applies to all pure equilibria of ranking games: there exists at least one player, namely the
one ranked last in the equilibrium outcome, who receives his minimum payoff regardless
of his choice of action.
To alleviate the effects of phenomena like this, Harsanyi (1973) suggests to impose
the additional restriction that every best response of a player be played with positive
probability. A Nash equilibrium satisfying this requirement is called a quasi-strict equi-
librium.1 Quasi-strict equilibrium is a refinement of Nash equilibrium in the sense that
the set of quasi-strict equilibria of every game forms a subset of the set of Nash equilibria.
This may also be beneficial with respect to another weakness of Nash equilibrium, its
potential multiplicity. It turns out that the second equilibrium of the game in Figure 7.1,
in which Bob raises his hand while Alice and Charlie randomize uniformly over their re-
spective actions, is quasi-strict. Interestingly, Charlie plays a weakly dominated action
with positive probability in this equilibrium.
Quasi-strict equilibria always exist in two-player games (Norde, 1999), but may fail to
do so in games with more than two players. The game in Figure 4.9 on Page 48 shows
that in fact, quasi-strict equilibria can already fail to exist in the very restricted class

1
Harsanyi originally referred to quasi-strict equilibrium as “quasi-strong”. However, this term has been
dropped to distinguish the concept from Aumann’s (1959) strong equilibrium.

127
128 7 · Quasi-Strict Equilibria

c1 c2
b1 b2 b1 b2
a1 3 1 1 2
a2 1 2 2 1

Figure 7.1: Single-winner game involving Alice (player 1), Bob (player 2), and Charlie
(player 3), repeated from Figure 4.1. The dashed square marks the only pure equilibrium,
dotted rectangles mark an equilibrium in which Alice and Charlie randomize uniformly
over their respective actions. The latter is the unique quasi-strict equilibrium of the game.

of single-winner games.2 In this chapter, we study the existence and the computational
properties of quasi-strict equilibrium in zero-sum games, general normal-form games, and
certain classes of anonymous and symmetric games. Section 7.3 focuses on two-player
games, and it shown that quasi-strict equilibria, unlike Nash equilibria, have a unique
support. We further give linear programs that characterize the quasi-strict equilibria in
non-symmetric and symmetric zero-sum games. In Section 7.4 we turn to games with more
than two players. We first identify new classes of games where a quasi-strict equilibrium
is guaranteed to exist, and can in fact be found efficiently. An interesting example of
such a class are symmetric games where every player has two actions. We then show that
deciding the existence of a quasi-strict equilibrium in games with more than two players is
NP-hard in general. This is in contrast to the two-player case, where the decision problem
is trivial due to the existence result by Norde (1999).

7.1 Related Work

To address various drawbacks of Nash equilibrium, a number of concepts that single
out particularly reasonable Nash equilibria—so-called equilibrium refinements—have been
proposed over the years. A result by Norde et al. (1996) has however cast doubt upon
this strand of research. The authors show that Nash equilibrium can be completely char-
acterized by payoff maximization in one-player games, consistency,3 and existence.
As a consequence, all common equilibrium refinements either violate consistency or exis-
tence. In particular, all refinements that are guaranteed to exist such as perfect, proper,
and persistent equilibria suffer from inconsistency while consistent refinements such as
2
There are only few examples in the literature for games without quasi-strict equilibria, essentially one
example by van Damme (1983) and another one by Cubitt and Sugden (1994). For this reason, the game
depicted in Figure 4.9 might be of independent interest.
3
Consistency as introduced by Peleg and Tijs (1996) is defined as follows. Let s be a solution of game Γ
and let Γ 0 be a reduced game where a subset of players are assumed to invariably play the strategies
prescribed by s. A solution concept is consistent if the solution s 0 in which all of the remaining players
still play according to s is a solution of Γ 0 .
7.2 · Preliminaries 129

quasi-strict, strong, and coalition-proof equilibria may fail to exist. Since consistency
is a very intuitive and appealing property, its failure may be considered more severe than
possible non-existence.
Harsanyi’s quasi-strict equilibrium, which refines the Nash equilibrium concept by
requiring that every action in the support yields strictly more payoff than actions not in
the support, has been shown to always exist in generic n-player games, and thus in “almost
every” game (Harsanyi, 1973), and in two-player games (Norde, 1999). Squires (1998) has
further shown that quasi-strict equilibrium satisfies a set of axioms due to Cubitt and
Sugden (1994), and is therefore very attractive from an axiomatic perspective. This result
can be interpreted in such a way that the existence of a quasi-strict equilibrium is sufficient
to justify the assumption of common knowledge of rationality. In fact, Quesada (2002)
even poses the question whether the existence of quasi-strict equilibrium is sufficient
for any reasonable justification theory. Finally, isolated quasi-strict equilibria satisfy
almost all desirable properties defined in the refinements literature. They are essential,
strongly stable, regular, proper, and strictly perfect (e.g., Jansen, 1987, van Damme, 1983,
1991). Using the framework of Peleg and Tijs (1996) and Norde et al. (1996), quasi-strict
equilibrium can be characterized by consistency and strict payoff maximization in one-
player games.

7.2 Preliminaries
The idea behind Nash equilibrium, as introduced in Definition 2.4 on Page 9, is that no
player can increase his payoff by unilaterally changing his strategy. It follows directly
from the definition that every action played with positive probability yields the same
expected payoff. Quasi-strict equilibrium strengthens the equilibrium condition, and thus
refines the Nash equilibrium concept, by requiring that the actions played with positive
probability are exactly those maximizing a player’s expected payoff, such that equilibrium
actions must yield strictly more payoff than actions played with probability zero.

Definition 7.1 (quasi-strict equilibrium). Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a normal-
form game. A Nash equilibrium s∗N ∈ SN is called quasi-strict equilibrium if for all i ∈ N
and all ai , ai0 ∈ Ai with s∗i (ai ) > 0 and s∗i (ai0 ) = 0, pi (s∗−i , ai ) > pi (s∗−i , ai0 ).

A pure equilibrium satisfying this property is usually called strict.

7.3 Two-Player Games

Using an elaborate construction based on Brouwer’s fixed point theorem, Norde (1999)
shows that quasi-strict equilibria always exists in two-player games. Since every quasi-
strict equilibrium is also a Nash equilibrium, the problem of finding a quasi-strict equi-
librium is intractable unless P=PPAD (Chen and Deng, 2006). The same is true for
symmetric two-player games, because the symmetrization of Gale et al. (1950) preserves
130 7 · Quasi-Strict Equilibria

maximize v
subject to
P
a1 ∈A1 s1 (a1 ) p(a1 , a2 ) > v ∀a2 ∈ A2
s1 (a1 ) > 0 ∀a1 ∈ A1
P
a1 ∈A1 s1 (a1 ) = 1

minimize v
subject to
P
a2 ∈A2 s2 (j) p(a1 , a2 ) 6 v ∀a1 ∈ A1
s2 (a2 ) > 0 ∀a2 ∈ A2
P
a2 ∈A2 s2 (a2 ) = 1

Figure 7.2: Primal and dual linear programs for computing minimax strategies in zero-sum
games

quasi-strictness (Jurg et al., 1992). For the restricted class of zero-sum games, however,
quasi-strict equilibria, like Nash equilibria, can be found efficiently by linear programming.
In contrast to Nash equilibria, quasi-strict equilibria have a unique support in these games.
Theorem 7.2. In two-player zero-sum games, quasi-strict equilibria have a unique
support, and one of them can be found in polynomial time.
Proof. It is known from the work of Jansen (1981) that every two-player game with a
convex equilibrium set, and thus every two-player zero-sum game, possesses a quasi-strict
equilibrium. We first establish that the support of a quasi-strict equilibrium must contain
every action that is played with positive probability in some equilibrium of the game.
Consider a quasi-strict equilibrium (s1 , s2 ) with payoff v for player 1, and assume for
contradiction that there exists an action a1 ∈ A1 played with positive probability in some
Nash equilibrium such that s1 (a1 ) = 0. Then, by Definition 7.1, p1 (a1 , s2 ) < v. Since a1
is in the support of some Nash equilibrium, and by the exchangeability of equilibrium
strategies in zero-sum games, however, it holds that p1 (a1 , s2 ) = v, a contradiction.
In order to compute quasi-strict equilibria, consider the two standard linear programs
for finding the minimax strategies for player 1 and 2, respectively, given in Figure 7.2 (e.g.,
Luce and Raiffa, 1957). It is well-known from the minimax theorem (von Neumann, 1928),
and also follows from LP duality, that the value v of the game is identical and unique in
both cases. We can thus construct a linear feasibility program that computes equilibrium
strategies for both players by simply merging the sets of constraints and omitting the
minimization and maximization objectives.
Now, quasi-strict equilibrium requires that action a yields strictly more payoff than
action a 0 of the same player if and only if a is in the support and a 0 is not. For a zero-
7.4 · A Hardness Result for Multi-Player Games 131

maximize 
subject to
P
a2 ∈A2 s2 (a2 ) p(a1 , a2 ) 6 v ∀a1 ∈ A1
s2 (a2 ) > 0 ∀a2 ∈ A2
P
a2 ∈A2 s2 (aP
2) = 1
s1 (a1 ) + v − a2 ∈A2 s2 (a2 ) p(a1 , a2 ) −  > 0 ∀a1 ∈ A1
P
a1 ∈A1 s1 (a1 ) p(a1 , a2 ) > v ∀a2 ∈ A2
s1 (a1 ) > 0 ∀a1 ∈ A1
P
a1 ∈A1 s1 (aP
1) = 1
s2 (a2 ) + v − a1 ∈A1 s1 (a1 ) p(a1 , a2 ) −  > 0 ∀a2 ∈ A2

Figure 7.3: Linear program for computing quasi-strict equilibria in zero-sum games

sum game with value v this can be ensured by requiring that for every action a1 ∈ A1 of
P
player 1, s1 (a1 ) + v > a2 ∈A2 s2 (a2 ) p(a1 , a2 ). If a1 is not in the support, i.e., s1 (a1 ) =
0, then it must yield strictly less payoff than the value of the game. If, on the other hand a1
is in the support, i.e., s1 (a1 ) > 0, these constraints do not impose any restrictions given
that the strategy profile sN is indeed an equilibrium with value v. The latter is ensured by
the remaining constraints. We finally add another variable  to be maximized to the right
hand side of the above inequality, to turn the strict inequality into a weak one. Due to
the existence of at least one quasi-strict equilibrium, we are guaranteed to find a solution
with positive . The resulting linear program is given in Figure 7.3.

We proceed by showing that every symmetric zero-sum game has a symmetric quasi-
strict equilibrium. This result stands in contrast to Theorem 7.4 in Section 7.4, which
shows that the same need not be the case for symmetric two-player games in general.
Theorem 7.3. Every symmetric two-player zero-sum game has a symmetric quasi-
strict equilibrium which can be found in polynomial time.
Proof. By Theorem 7.2, the support of any quasi-strict equilibrium contains all actions
that are played with positive probability in some equilibrium. Clearly these actions coin-
cide for both players in symmetric games, and any minimax strategy using these actions
constitutes a symmetric equilibrium. Since both players can enforce their minimax value
using the same strategy in a symmetric zero-sum game, the value of the game has to be
zero. Using this information, the linear program in Figure 7.3 can be simplified signifi-
cantly. The resulting linear program is shown in Figure 7.4.

7.4 A Hardness Result for Multi-Player Games

In games with three or more players, the existence of a quasi-strict equilibrium is no
longer guaranteed. However, there are very few examples in the literature for games
132 7 · Quasi-Strict Equilibria

maximize 
subject to
P
a2 ∈A2 s(a2 ) p(a1 , a2 ) 6 0 ∀a1 ∈ A1
s(a2 ) > 0 ∀a2 ∈ A2
P
a2 ∈A2Ps(a2 ) = 1
s(a1 ) − a2 ∈A2 s(a2 ) p(a1 , a2 ) −  > 0 ∀a1 ∈ A1

Figure 7.4: Linear program for computing quasi-strict equilibria in symmetric zero-sum
games

c1 c2
b1 b2 b1 b2
a1 (1, 1, 0) (0, 1, 1) (0, 1, 1) (1, 0, 1)
a2 (0, 1, 1) (1, 0, 1) (1, 0, 1) (1, 1, 0)

Figure 7.5: Somebody has to do the dishes, via a self-anonymous single-loser game. Play-
ers 1, 2, and 3 choose rows, columns, and matrices, respectively. In the only Nash equilib-
rium of the game player 1 plays his second action, player 2 plays his first action, and player
3 randomizes uniformly over both of his actions. The game does not have a quasi-strict
equilibrium.

without quasi-strict equilibria.4 An important question is of course which natural classes

of games always possess a quasi-strict equilibrium. We have seen in Chapter 4 that this
is not the case for the class of single-winner games, which require that all outcomes are
permutations of the payoff vector (1, 0, . . . , 0).
In the following, we look at anonymous and symmetric games. It turns out that
self-anonymous games, and thus also anonymous games, need not possess a quasi-strict
equilibrium. For this, consider the following protocol used by Alice, Bob, and Charlie to
select one among them to do the dishes. Each of them decides to raise their hand or not,
simultaneously and independently of the others. Alice loses, and has to do the dishes, if
exactly one player raises his hand, Bob loses if exactly two players raise their hands, and
Charlie loses if either all or no players raise their hand. The resulting anonymous single-
loser game, depicted in Figure 7.5, exhibits some peculiar phenomena, some of which may
be attributed to the absence of quasi-strict equilibria. For example, the security level of
all players is 1/2, and the expected payoff in the only Nash equilibrium, which has Alice
raise her hand and Charlie randomize with equal probability, is (1/2, 1/2, 1). The minimax

4
To our knowledge, there are three examples in addition to the one we will give below (van Damme,
1983, Kojima et al., 1985, Cubitt and Sugden, 1994). All of them involve three players with two actions
each, which clearly is optimal.
7.4 · A Hardness Result for Multi-Player Games 133

··· pm−1,0 pm0 pm 0 0 pm 0 +1,0 ···

···

>

<
=

=
··· pm−1,1 pm1 pm 0 1 pm 0 +1,1 ···

Figure 7.6: Payoff structure of a symmetric game with two actions

strategies of Alice and Bob, however, are different from their equilibrium strategies, i.e.,
they can guarantee their equilibrium payoff by not playing their respective equilibrium
strategies.5 Furthermore, the unique equilibrium is not quasi-strict, i.e., Alice and Bob
could as well play any other action without jeopardizing their payoff.
For symmetric and self-symmetric games, on the other hand, the picture is different.
Self-symmetric games form a subclass of common-payoff games, where the payoff of all
players is identical in every outcome. Starting from an outcome with maximum payoff, a
quasi-strict equilibrium can be found by iteratively adding other actions to the support
by which some player, and thus all players, also obtain the maximum payoff.
It follows from a theorem by Nash (1951) that every symmetric game has a symmetric
Nash equilibrium, i.e., a Nash equilibrium where all players play the same strategy. We
can use this result to show that quasi-strict equilibria are guaranteed to exist in symmetric
games with two actions for each player. Perhaps surprisingly, all quasi-strict equilibria of
such a game may themselves be asymmetric.

Theorem 7.4. Every symmetric game with two actions for each player has a quasi-
strict equilibrium. Such an equilibrium can be found in polynomial time.

Proof. Let Γ = (N, {0, 1}N , (pi )i∈N ) be a symmetric game. For an action profile aN and
an action a ∈ {0, 1}, denote by #(a, aN ) = |{ i ∈ N : ai = a }| the number of players who
play a in aN . It follows from Definition 5.1 that there exist 2n numbers pm` ∈ R for
0 6 m 6 n − 1 and ` ∈ {0, 1}, such that for all i ∈ N, pi (aN ) = pm` whenever ai = `
and #(1, a−i ) = m. We can further assume without loss of generality that p00 = p01 and
pn−1,0 > pn−1,1 , and that pm0 6= pm1 for some m. To see this, recall that Γ , being a
symmetric game, must possess a symmetric equilibrium sN (Nash, 1951), which we can
assume to be the pure strategy profile where all players play action 0 with probability 1.
If instead all players played both of their actions with positive probability, sN would itself
be quasi-strict. Now, if one of the former two equations was not satisfied, then one of
the two symmetric pure strategy profiles would be a quasi-strict equilibrium. If the latter
condition would not to hold, there would exist a quasi-strict equilibrium where all players
randomize between their actions.
It is now easily verified that there must exist m and m 0 , 0 6 m 6 m 0 6 n − 1, such
that (i) pm 00 0 = pm 00 1 for all m 00 with m 6 m 00 6 m 0 , (ii) either m = 0 or pm−1,0 <
5
Similar phenomena were also observed by Aumann (1985).
134 7 · Quasi-Strict Equilibria

pm−1,1 , and (iii) either m 0 = n − 1 or pm 0 +1,0 > pm 0 +1,1 . This situation is illustrated
in Figure 7.6. We further claim that any strategy profile sN where n − m 0 − 1 players
play action 0, m players play action 1, and the remaining m 0 − m + 1 players randomize
between both of their actions, is a quasi-strict equilibrium of Γ . First consider any player
i ∈ N who plays action 0 with probability 1. Then for any action profile aN played
with positive probability in sN , m 6 #(1, a−i ) 6 m + (m 0 − m + 1) = m 0 + 1. As a
consequence, player i weakly prefers action 0 over action 1 in any such action profile, and
the preference is strict for at least one such action profile. Now consider any player i ∈ N
who plays action 1 with probability 1. Then for any action profile aN played with positive
probability in sN , m − 1 6 #(1, a−i ) 6 (m − 1) + (m 0 − m + 1) = m 0 . Player i weakly
prefers action 1 over action 0 in any such action profile, and the preference is strict for at
least one such action profile. Finally consider any player i ∈ N who plays both actions
with positive probability. It then holds for any action profile aN played with positive
probability in sN that m 6 #(1, a−i ) 6 m + (m 0 − m) = m 0 , and player i is indifferent
between actions 0 and action 1 in any such action profile.
Since a symmetric equilibrium of a symmetric game with a constant number of actions
can be found in polynomial time (Papadimitriou and Roughgarden, 2005), and since the
proof of the first part of the theorem is constructive, the second part follows immediately.

We leave it as an open problem whether all symmetric games contain a quasi-strict

equilibrium. If the symmetrization procedure due to Gale et al. (1950) can be extended
to multi-player games while still preserving quasi-strictness, a counterexample could be
constructed from one of the known games without quasi-strict equilibria. Of course, in
light of Theorem 7.4, the number of actions per player in such a counter-example has to
be greater than two, and may very well be substantially greater than that.
We conclude this chapter by showing that the existence of a quasi-strict equilibrium
is NP-hard to decide in general.

Theorem 7.5. Deciding whether a game in normal form possesses a quasi-strict

equilibrium is NP-hard, even if there are only three players and a constant number
of payoffs.

Proof. For hardness, we give a reduction from the NP-complete problem CLIQUE (e.g.,
Papadimitriou, 1994a) reminiscent of a construction used by McLennan and Tourky (2005)
to show NP-hardness of various problems related to Nash equilibria in two-player games.
Given an undirected graph G = (V, E) and a positive integer k 6 |E|, CLIQUE asks
whether G contains a clique of size at least k, i.e., a subset C ⊆ V such that |C| > k
and for all distinct v, w ∈ C, (v, w) ∈ E. Given a particular CLIQUE instance ((V, E), k)
with V = {1, 2, . . . , m}, we construct a game Γ with three players, actions A1 = { ai :
i ∈ V } ∪ {a0 }, A2 = { bi : i ∈ V } ∪ {b0 } and A3 = {c1 , c2 }. The payoff functions pi are
defined as follows, and illustrated in Figure 7.7. If player 3 plays c1 and players 1 and 2
7.4 · A Hardness Result for Multi-Player Games 135

c1 c2
b1 ··· b|V| b0 b1 ··· b|V| b0
a1 (0, 0, 0) a1 (0, 0, K) · · · (0, 0, K) (0, 0, 0)
.. .. .. .. .. .. ..
. (mij , eij , mij )i,j∈V . . . . . .
a|V| (0, 0, 0) a|V| (0, 0, K) · · · (0, 0, K) (0, 0, 0)
a0 (0, 0, 0) · · · (0, 0, 0) (0, 1, 0) a0 (1, 0, 0) · · · (1, 0, 0) (0, 0, 0)

Figure 7.7: Three-player game Γ used in the proof of Theorem 7.5

play ai and bj , respectively, for i, j ∈ V, payoffs are given by a matrix (mij )i,j∈V defined
according to G, and by the identity matrix (eij )i,j∈V . More precisely,



 1 if (i, j) ∈ E,
mij = 0 if i = j,


−1 otherwise,

and

1 if i = j,
eij =
0 otherwise.

If player 3 instead plays c2 , he obtains a payoff of K = (2k − 3)/2k. We claim that Γ

possesses a quasi-strict equilibrium if and only if there exists a clique of size at least k
in G.
For the direction from right to left, assume that there exists a maximal clique C ⊆ V
with |C| > k, and consider the strategy profile sN with s1 (c1 ) = 1, and s1 (ai ) = s2 (bi ) =
1/|C| if i ∈ C and s1 (ai ) = s2 (bi ) = 0 otherwise. By construction of Γ , and for all
i ∈ V ∪ {0}, p2 (s−2 , bi ) < p2 (sN ) whenever s1 (ai ) = 0. Furthermore, by maximality
of C, p1 (s−1 , ai ) < p1 (sN ) for all i ∈ / C. Finally, p3 (sN ) = (k − 1)/k > (2k − 3)/2k =
p3 (s−3 , c2 ). Thus, sN is a quasi-strict equilibrium of Γ .
For the direction from left to right, consider a quasi-strict equilibrium sN of Γ , and
assume for contradiction that G does not have a clique of size at least k. In equilibrium, for
all b, b 0 ∈ A2 , we must have p2 (s−2 , b) = p2 (s−2 , b 0 ) whenever s2 (b) > 0 and s2 (b 0 ) > 0,
and thus, for all a, a 0 ∈ A1 , s1 (a) = s1 (a 0 ) whenever s1 (a) > 0 and s1 (a 0 ) > 0. As a
consequence, for sN to be quasi-strict, s2 (bi ) > 0 whenever s1 (ai ) > 0, for all i ∈ V ∪ {0}.
First consider the case when s3 (c1 ) > 0. If s1 (a0 ) = s2 (b0 ) = 1, sN would not be quasi-
strict for player 1. If on the other hand s1 (ai ) > 0 or s2 (bi ) > 0 for some i ∈ V, then
there would have to be a set C ⊆ V with |C| > k, such that for all i ∈ V with s1 (ai ) > 0
and all j ∈ C \ {i}, p1 (ai , bj , c1 ) = 1. By construction of Γ , C would be a clique of size at
136 7 · Quasi-Strict Equilibria

least k in G, a contradiction. Now consider the case when s3 (c2 ) = 1. If s1 (a0 ) = 1 or

s2 (b0 ) = 1, sN would not be quasi-strict for player 3. If, on the other hand, s1 (ai ) > 0
and s2 (bj ) > 0 for some i, j ∈ V, then player 1 could change his action to a0 to get a
higher payoff, a contradiction.

It follows that the problem of finding a quasi-strict equilibrium in games with more
than two players is NP-hard, under polynomial-time Turing reductions. It is unlikely that
the same is true for Nash equilibrium.

7.5 Discussion
In this chapter we investigated the computational properties of an attractive equilibrium
refinement known as quasi-strict equilibrium. It turned out that quasi-strict equilibria
in zero-sum games have a unique support and can be computed efficiently via linear
programming. In games with more than two players, deciding the existence of a quasi-
strict equilibrium quickly becomes NP-hard.
As pointed out in the beginning of this chapter, classes of games that always admit
a quasi-strict equilibrium, such as two-player games, are of vital importance to justify
rational play based on elementary assumptions. We specifically looked at anonymous and
symmetric games, and found that self-anonymous games, and thus also anonymous games,
need not possess a quasi-strict equilibrium, while symmetric games with two actions for
each player always do. It is an interesting question whether the latter is still true when
there are more than two actions. Other classes of multi-player games where one might hope
for the existence of a quasi-strict equilibrium include unilaterally competitive games (Kats
and Thisse, 1992), potential games (Monderer and Shapley, 1996), graphical games with
bounded neighborhood, and single-winner games where all players have a positive security
level.
 Chapter 8

Shapley’s Saddles

Another, perhaps more severe, problem of Nash equilibrium is that its existence generally
requires randomization on behalf of the players. It is not clear if and how players can
execute such randomizations in an exact and reliable way, which in turn is essential for
the notion of stability underlying Nash equilibrium. The problem becomes more severe
in games with more than two players, where randomization with irrational probabilities
may be required.
In work dating back to the early 1950s, Shapley proposed ordinal set-valued solution
concepts for zero-sum games that he refers to as saddles (Shapley, 1953a,b, 1955, 1964).
What makes these concepts intuitively appealing is that they are based on the elementary
notions of dominance and stability. Call a generalized saddle point (GSP) is a tuple of
subsets of actions, one for each player, such that every action not contained in the GSP is
dominated by some action in the GSP, given that the remaining players only play actions
from the GSP. A saddle then is an inclusion-minimal GSP, i.e., a GSP that contains no
other GSP. By varying the underlying notion of dominance, one obtains strict and weak
saddles. Shapley (1964) showed that every (two-player) zero-sum game admits a unique
strict saddle. Duggan and Le Breton (1996) proved that the same is true for the weak
saddle in a certain subclass of symmetric zero-sum games.
Despite the fact that Shapley’s saddles were devised as early as 1953 (Shapley, 1953a,b)
and are thus almost as old as Nash equilibrium (Nash, 1951), surprisingly little is known
about their computational properties. In this chapter, we provide polynomial-time algo-
rithms for computing strict saddles in normal-form games with any number of players,
and weak saddles in the subclass of symmetric zero-sum games considered by Duggan and
Le Breton (1996). We note, but do not show here, that these results extend to mixed
refinements of Shapley’s saddles introduced by Duggan and Le Breton (2001). On the
other hand, certain problems associated with (pure and mixed) weak saddles in two-player
games, such as deciding whether there exists a weak saddle with k actions for some player,
are shown to be NP-hard. We restrict our attention to games with a constant number of
players, and assume throughout the chapter that these games are given explicitly, i.e., as
a multi-dimensional table of payoffs.

137
138 8 · Shapley’s Saddles

8.1 Related Work

Shapley’s saddles are based on the notion of dominance, which has also been studied from
a computational perspective (e.g., Knuth et al., 1988, Conitzer and Sandholm, 2005a,b,
and Chapters 4 and 5 of this thesis). Iterated strict dominance, for example, is one of the
few examples of a solution concept known to be efficiently computable in general games.
Strict saddles can be seen as a refinement of iterated strict dominance in that all strict
saddles of a normal-form game are contained in the subgame that one obtains by iterated
elimination of strictly dominated actions.
A related set-valued solution concept are CURB sets (Basu and Weibull, 1991). Un-
like Shapley’s saddles, however, CURB sets are based on randomized strategies. They
are therefore not ordinal, and are subject to some of the complications associated with
randomized strategies. Every strict saddle represents the support of a CURB set, and
thus contains the support of a minimal CURB set. In confrontation games, as defined
in Section 8.4, the support of a minimal CURB set and the strict saddle trivially coin-
cide. Moreover, in this particular class of games, the strict mixed saddle is identical to
the support of the minimal CURB set when only allowing pure strategies. There appears
to be no such relationship between weak saddles and CURB sets. Benisch et al. (2006)
have recently proposed a polynomial-time algorithm for computing the CURB sets of a
two-player game.

8.2 Preliminaries
Existence of a Nash equilibrium is not guaranteed if strategies are required to be pure, as
the Matching Pennies game in Figure 2.2 on Page 9 illustrates. Requiring randomization
in order to reach a stable outcome, however, is problematic for various reasons. A possible
solution is to consider set-valued solution concepts that identify, for each player i ∈ N, a
subset Xi ⊆ Ai of his actions, such that the vector (X1 , X2 , . . . , Xn ) satisfies some notion
of stability. Shapley’s saddles generalize strict Nash equilibrium1 by requiring that for
every action ai ∈ Ai \ Xi of a player i ∈ N that is not included in Xi , there should be
some reason for its exclusion, namely an action in Xi that is strictly better than ai .
To formalize this idea, we need some notation. Henceforth, let AN = (A1 , A2 , . . . , An ).
For XN = (X1 , X2 , . . . , Xn ), write XN ⊆ AN , and say that XN is a subset of AN , if
∅ =
6 Xi ⊆ Ai for all i ∈ N. Further let X−i = (X1 , . . . , Xi−1 , Xi+1 , . . . , Xn ). Finally
consider a generalization of the notions of strict and weak dominance to specific sets of
actions of the other players. For a player i ∈ N and two actions ai , bi ∈ Ai , say that ai
strictly dominates bi with respect to X−i ⊆ A−i , denoted ai X−i bi , if p(ai , x−i ) >
p(bi , x−i ) for all x−i ∈ X−i , and that ai weakly dominates bi with respect to X−i ,
1
Recall that a pure Nash equilibrium is called strict if every player strictly loses when deviating from
his equilibrium action.
8.2 · Preliminaries 139

b1 b2 b3
a1 (3, −3) (3, −3) (4, −4)
a2 (2, −2) (3, −3) (3, −3)
a3 (1, −1) (2, −2) (3, −3)
a3 (2, −2) (1, −1) (5, −5)

Figure 8.1: Strict and weak saddles of a zero-sum game, respectively indicated by dashed
and dotted boxes

denoted ai >X−i bi , if p(ai , x−i ) > p(bi , x−i ) for all x−i ∈ X−i , with at least one strict
inequality. We are now ready to define strict and weak saddles formally.
Definition 8.1 (strict and weak saddle). Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a game,
XN = (X1 , X2 , . . . , Xn ) ⊆ AN . Then, XN is a generalized saddle point (GSP) of Γ if for
all i ∈ N and

for all ai ∈ Ai \ Xi , there exists xi ∈ Xi such that xi X−i ai . (8.1)

A GSP is called strict saddle if it does not contain a GSP as a proper subset.
Similarly, XN is a weak generalized saddle point (WGSP) of Γ if for all i ∈ N and

for all ai ∈ Ai \ Xi , there exists xi ∈ Xi such that xi >X−i ai . (8.2)

A WGSP is called weak saddle if it does not contain a WGSP as a proper subset.
The underlying intuition is that every player i ∈ N has a distinguished set Xi of actions
such that for every action ai ∈ / Xi , there is some action in Xi that dominates ai , provided
that the other players play only actions from their distinguished sets. Both strict and
weak saddle are ordinal solution concepts, i.e., they are invariant under order-preserving
transformations of the payoff functions. This is in contrast to Nash equilibrium, for which
invariance holds only under positive affine transformations.
Properties (8.1) and (8.2) are sometimes referred to as external stability. Using this
terminology, a (W)GSP is a tuple of externally stable sets. Since strict dominance im-
plies weak dominance, every strict saddle is a WGSP and thus contains a weak sad-
dle. Consider for example the two-player zero-sum game shown in Figure 8.1. The pair
XN = ({a1 , a2 }, {b1 , b2 }) is a strict saddle and a WGSP. Since a1 weakly dominates a2
with respect to {b1 , b2 }, and both b1 and b2 dominate b3 with respect to a1 , the pair
XN0 = ({a1 }, {b1 , b2 }) also is a WGSP. Indeed, X 0 does not contain a smaller WGSP, and
N
therefore is a weak saddle. Some reflection reveals that XN and XN 0 are in fact the unique

strict and weak saddles of the game.

It is easy to see that every game has both a strict and a weak saddle. By definition,
the tuple AN is a GSP. Furthermore every GSP that is not a saddle must contain a GSP
140 8 · Shapley’s Saddles

a1 a2 a3 a4
a1 (0, 0) (1, −1) (0, 0) (0, 0)
a2 (−1, 1) (0, 0) (0, 0) (0, 0)
a3 (0, 0) (0, 0) (0, 0) (1, −1)
a4 (0, 0) (0, 0) (−1, 1) (0, 0)

Figure 8.2: Symmetric zero-sum game with multiple weak saddles

that is strictly smaller. Finiteness of AN implies the existence of a minimal GSP, i.e.,
a strict saddle. An analogous argument applies to the weak saddle. Strict saddles are
unique in two-player zero-sum games but not in general games, whereas weak saddles are
not even unique in the former class. To see this, consider the symmetric zero-sum game
of Figure 8.2. It is easily verified that each of the following pairs is a weak saddle of this
game: ({a1 , a2 }, {a1 , a2 }), ({a3 , a4 }, {a3 , a4 }), ({a1 , a3 }, {a1 , a3 }), ({a2 , a3 }, {a1 , a4 }), and
({a1 , a4 }, {a2 , a3 }).

8.3 Strict Saddles

Shapley (1964) shows that every two-player zero-sum game possesses a unique strict sad-
dle, because the set of GSPs in such games is closed under intersection. He further
describes an algorithm, attributed to Harlan Mills, to compute this saddle. The idea be-
hind this algorithm is that given a subset of the saddle, the saddle itself can be computed
by iteratively adding actions that are maximal, i.e., not dominated with respect to the
current subset of actions of the other player. Shapley further points out that the strict
saddle must contain all actions possibly leading to a minimax or maximin point. Note,
however, that being able to find a subset of the saddle is not crucial for the above method.
Starting the algorithm from singleton sets of actions, and invoking it for every combination
of such sets for the different players, yields a number of candidates for the strict saddle.
The strict saddle itself can then easily be identified as the inclusion-minimal set among
these candidates. Correctness of this procedure follows from the fact that every candidate
set is a GSP, and that the unique strict saddle is contained in every GSP. Furthermore,
Mills’ iterative procedure is invoked only a polynomial number of times.
In general games, strict saddles are no longer unique. For example, consider the two-
player coordination game in which both players have two actions and receive a payoff
of one if they play the same action, and a payoff of zero otherwise. Obviously, this game
has two strict saddles, one where both players play their first action, and one where both
of them play their second action. From a computational point of view, the existence of
multiple strict saddles turns out not to have any serious consequences.
8.4 · Weak Saddles of Confrontation Games 141

Theorem 8.2. All strict saddles of an n-player game can be computed in polynomial
time.

Proof. We generalize Mills’ algorithm and show that the inclusion-minimal GSP contain-
ing a given input set X0N can be computed in polynomial time. The statement of the
theorem then follows by similar arguments as before: we can invoke the algorithm a poly-
nomial number of times, namely for each vector containing for each player a singleton
set of actions, and then select the inclusion minimal elements from the resulting set of
candidate GSPs.
Let Γ = (N, (Ai )i∈N , (pi )i∈N ) be a game, AN = (Ai )i∈N , and X0N ⊆ AN , and consider
the following algorithm:

1. Let XN = X0N .

2. While there exists a player i ∈ N and an action ai ∈ Ai \ Xi that is not dominated

with respect to X−i , i.e., one for which there is no bi ∈ Ai with bi X−i ai , add ai
to Xi .

3. Return XN .

Clearly, we can check in polynomial time whether a given action is dominated with respect
to particular subsets of actions of the other players. Since there is only a polynomial
number of actions, and one of them is added to XN in every iteration, the algorithm runs
in polynomial time.
It remains to be shown that the algorithm returns the inclusion minimal GSP con-
taining X0N . To see this, let Xmin
N be the minimal GSP containing X0N . We show that
XN ⊆ XminN always holds during the execution of the algorithm. When the algorithm
terminates, all elements outside XN are dominated, meaning that XN is a GSP. At
the beginning of the algorithm, XN = X0N ⊆ Xmin N by definition of Xmin
N . Now assume
that XN ⊆ XN holds at the beginning of a particular iteration, and let Ai0 be the set
min

of actions of player i not dominated with respect to X−i , i.e., Ai0 = {ai ∈ Ai \ Xi :
there is no bi ∈ Ai such that bi X−i ai }. Let ai ∈ Ai0 be the element added to Xi , and
assume for contradiction that ai ∈/ Xmin
i . Since Xmin
N is a GSP, there must exist an action
bi ∈ Xi with bi Xmin ai . By the induction hypothesis, X−i ⊆ Xmin
min
−i , which in turn
−i
implies bi X−i ai . This contradicts the assumption that ai ∈ Ai .0

A similar result can be obtained for the mixed strict saddle, where external stability
is defined in terms of dominance by mixed strategies.

8.4 Weak Saddles of Confrontation Games

The computation of weak saddles turns out be significantly more complicated than that
of strict saddles. Somewhat surprisingly, this is the case even in two-player zero-sum
142 8 · Shapley’s Saddles

games. In particular, Mills’ algorithm does not easily generalize to weak saddles. When it
is invoked on the game of Figure 8.1, for example, and initialized with X0N = ({a1 }, {b2 }),
it might add action a2 in the next step, which is not contained in a minimal WGSP
containing X0N . In this section we consider a subclass of symmetric zero-sum games that
is guaranteed to possess a unique weak saddle. The complexity of the problem for general
zero-sum games remains open.
It will be instructive to view a symmetric zero-sum game from a slightly different
perspective, namely that of comparison functions as introduced by Dutta and Laslier
(1999). Let A be a finite set of alternatives. A comparison function on A is a function
g : A × A → R such that for all x, y ∈ A, g(x, y) = −g(y, x). This obviously implies
that g(x, x) = 0 for all x ∈ X. For two alternatives x, y ∈ A, the value g(x, y) can be
interpreted as the extent to which x is “better” than y.
The subclass of comparison functions g satisfying g(x, y) ∈ {−1, 1} for all x, y ∈ A
with x 6= y corresponds to tournaments, i.e., complete and asymmetric2 relations, on A.
Tournaments for example arise if an odd number of agents with strict preferences conduct
a majority vote for each pair of alternatives, or from coalitional games with the property
that either a coalition or its complement is winning. A very active subarea of social choice
theory is concerned with the study of tournament solutions that map each tournament
to a nonempty set of “desirable” alternatives (e.g., Laslier, 1997). A particularly appealing
tournament solution is the minimal covering set of Dutta (1988), which was subsequently
generalized to comparison functions by Dutta and Laslier (1999).
An element x is said to cover another element y if it is better according to g, and if
for every third element z, x compares at least as well against z as y does.

Definition 8.3 (covering). Let A be a set of alternatives, g a comparison function on A,

and B ⊆ A. Then, for x, y ∈ B, x covers y in B if g(x, y) > 0 and g(x, z) > g(y, z) for all
z ∈ B \ {x, y}.

We will call the uncovered set of a comparison function g on a set B, denoted

UC (B, g), the set of alternatives in B not covered by any alternative in B. A covering
set then is a set of alternatives that is internally and externally stable, in the appropriate
sense, with respect to the covering relation.

Definition 8.4 (covering set). Let A be a set of alternatives, g a comparison function

on A, and B ⊆ A. Then, B is a covering set for g in A if

(i) UC (B, g) = B and

(ii) for all x ∈ A \ B, x ∈

/ UC (B ∪ {x}, g).

Dutta and Laslier (1999) show that the intersection of any two externally stable sets is
nonempty and itself externally stable, and that the unique inclusion minimal such set also
2
A relation R ⊆ A × A is called asymmetric if for x, y ∈ A with x 6= y, xRy implies that not yRx.
8.4 · Weak Saddles of Confrontation Games 143

satisfies internal stability. For any set A of alternatives and any comparison function g
on A there thus exists a unique minimal covering set MC (A, g).
We will now show that the minimal covering set of a comparison function can be
computed efficiently, and then use a result about the relationship between the minimal
covering set and the weak saddle to obtain a tractability result for the latter in a restricted
class of games. Our algorithm for computing the minimal covering set proceeds in a similar
fashion as that for the strict saddle: start with a subset of the minimal covering set and
then iteratively add elements outside the current set that are still uncovered. There are
two problems, however. For one, it is not clear how to find a subset of the minimal
covering set. Secondly, it is important only to add elements that may not be covered in a
later iteration, and it is not obvious which elements these should be.
A solution known to yield a subset of the minimal covering set is the essential set of
Dutta and Laslier (1999), which generalizes the bipartisan set of Laffond et al. (1993). The
essential set makes a connection back to game theory by considering, for a comparison
function g on a set A of alternatives, a symmetric zero-sum game ({1, 2}, (A, A), (g, −g)).
The intuition underlying this so-called tournament game is that each of two parties
proposes an alternative for adoption, and the better of the two, according to the com-
parison function, gets selected. The essential set itself is defined as the set of actions
that are played with positive probability in some Nash equilibrium of the tournament
game. Since the tournament game is a symmetric zero-sum game, it suffices to consider
symmetric equilibria.
Definition 8.5 (essential set). Let A be a set of alternatives, g a comparison function
on A. Then, the essential set of g is given by
ES (A, g) = { x ∈ A : s1 (x) > 0, (s1 , s1 ) is a Nash equilibrium of
the game ({1, 2}, (A, A), (g, −g)) }.
By Theorem 7.2, the essential set equals the support of the unique quasi-strict equi-
librium of the tournament game and can be computed in polynomial time by solving a
linear program.
The key observation that will allow us to overcome the second problem above is that
any element in the minimal covering set of the set of elements not covered by the
members of a subset of the minimal covering set must itself be in the minimal covering
set.
Lemma 8.6. Let A be a set of alternatives, g a comparison function on A. Further let
B ⊆ MC (A, g) and A 0 = a∈A\B (UC (B ∪ {a}, g) ∩ {a}). Then, MC (A 0 , g) ⊆ MC (A, g).
S

Proof. Partition A 0 , the set of alternatives not covered by B, into two sets C and C 0
of elements contained in MC (A, g) and elements not contained in MC (A, g), i.e., C =
A 0 ∩ MC (A, g) and C 0 = A 0 \ MC (A, g). We will show that C is externally stable for A 0 .
Since MC (A 0 , g) must lie in the intersection of all sets that are externally stable for A 0 ,
this means that MC (A 0 , g) ⊆ MC (A, g).
144 8 · Shapley’s Saddles

In the following, we will use an easy consequence of the definition of the covering
relation: for two sets X, X 0 with X ⊆ X 0 ⊆ A and two alternatives x, y ∈ X, if y covers x
in X 0 , then y also covers x in X. We will refer to this property as “covering in subsets”.
Let x ∈ C 0 . Since x ∈ / MC (A, g), there has to be some y ∈ MC (A, g) that covers x
in MC (A, g) ∪ {x}. It is easy to see that y ∈ / B. Otherwise, since B ⊆ MC (A, g) and by
covering in subsets, y would cover x in B ∪ {x}, contradicting the assumption that x ∈ A 0 .
On the other hand, assume that y ∈ MC (A, g)\(B∪C). Since y covers x in MC (A, g)∪{x},
since B ⊆ MC (A, g), and by covering in subsets, we have that g(y, x) > 0 and for all
w ∈ B, g(y, w) > g(x, w). Furthermore, since y ∈ / A 0 , there has to be some z ∈ B
covering y in B ∪ {y}, i.e., one such that g(z, y) > 0 and for all w ∈ B, g(z, w) > g(y, w).
Combining the two, we get g(z, x) > 0 and for all w ∈ B, g(z, w) > g(x, w), i.e., z covers x
in B ∪ {x}. This again contradicts the assumption that x ∈ A 0 . It thus has to be the case
that y ∈ C. Since C ⊆ MC (A, g), it follows from covering in subsets that y also covers x
in C ∪ {x}. We have shown that for every x ∈ C 0 , there exists y ∈ C such that y covers x
in C ∪ {x}, i.e., C is externally stable for A 0 .

Note that this lemma directly yields a recursive algorithm for computing the minimal
covering set. Some reflection reveals, however, that this algorithm requires exponentially
many steps in the worst case. Since the sets B and A 0 in the statement of Lemma 8.6
are always disjoint, the lemma also tells us how to find, for every proper subset of the
minimal covering set, another disjoint and non-empty subset. Again using the insight
that a non-empty subset of the minimal covering set can be found efficiently, we finally
obtain a polynomial-time algorithm.

Theorem 8.7. Let A be a set of alternatives, g a comparison function on A. Then,

MC (A, g) can be computed in time polynomial in |A|.

Proof. Consider the following simple algorithm:

1. Let B = ES (A, g).

2. While A 0 = a∈A\B (UC (B ∪ {a}, g) ∩ {a}) is nonempty, add ES (A 0 , g) to B.

S

3. Return B.

In each iteration of the second step, at least one element is added to set B, so the algorithm
is guaranteed to terminate after a linear number of iterations. In each iteration the
algorithm first identifies a subset A 0 of alternatives that are not yet covered by elements
of B. This can be done in polynomial time by computing, for each element a outside B,
the covering relation for B ∪ {a}, and checking whether a itself is a maximal element of
this relation. Then the elements in the essential set of A 0 , which by Theorem 7.2 can be
computed in polynomial time, are added to B.
As for correctness, a simple inductive argument shows that B ⊆ MC (A, g) holds at
any time. The base case follows directly from the fact that ES (A, g) ⊆ MC (A, g) (Dutta
8.4 · Weak Saddles of Confrontation Games 145

and Laslier, 1999, Theorem 4.3), the induction step from Lemma 8.6. When the algorithm
terminates, B is a covering set for A, so we must actually have B = MC (A, g).

Duggan and Le Breton (1996) consider an interesting class of symmetric zero-sum

games, and show that every game in this class has a unique weak saddle XN = (X1 , X2 ),
which is itself symmetric, i.e., satisfies X1 = X2 , and coincides with the minimal covering
set of the corresponding comparison function. Games in this class, which we will term
confrontation games, are characterized by the fact that the two players get the same
payoff if and only if they play the same action. In any other case, one of the players
strictly wins and the other strictly loses, and the outcome would be reversed if players
were to exchange their actions. Duggan and Le Breton (1996) call this property the
off-diagonal property in the context of comparison functions.

Definition 8.8 (confrontation game). Let Γ = (N, AN , (pi )i∈N ) be a symmetric zero-sum
game. Γ is called confrontation game if for all a, a 0 ∈ A, p1 (a, a 0 ) = p2 (a, a 0 ) = 0 if and
only if a = a 0 .

We have the following corollary of Theorem 8.7.

Corollary 8.9. The unique weak saddle of a confrontation game can be computed
in polynomial time.

In the remainder of this section, we present a family of symmetric zero-sum games

that are not confrontation games and have an exponential number of weak saddles. An
immediate consequence of this is that the computation of all weak saddles of a game
requires exponential time in the worst case, even for zero-sum games. For an odd integer
k > 1, define a two-player game Γk with action set A = A1 ∪ A2 ∪ · · · ∪ Ak for both players,
where Aj = {aj1 , aj2 , aj3 , aj4 } for 1 6 j 6 k. Let p10 denote the payoff function of player 1
in the game of Figure 8.2, and define the payoff function p1 of player 1 such that for all
j1 , j2 with 1 6 j1 , j2 6 k and all m1 , m2 with 1 6 m1 , m2 6 4,


 0 m1 m2 if j1 = j2
p1 (a , a )
p1 (aj1 m1 , aj2 m2 ) = (−1)j1 +j2 if j1 < j2


(−1)j1 +j2 +1 otherwise.

Further let p2 (sN ) = −p1 (sN ) for all sN ∈ A × A. The overall structure of this game is
shown in Figure 8.3.
Now, consider a pair (X1 , X2 ) ⊆ (A, A) such that for all j with 1 6 j 6 k, the pair
(Y1 , Y2 ) with Y1 = {am : ajm ∈ X1 ∩ Aj } and Y2 = {am : ajm ∈ X2 ∩ Aj } is a weak saddle
of the game in Figure 8.2. It is not hard to see that (X1 , X2 ) is a weak saddle of Γk , which
means that the total number of weak saddles of Γk is at least 5k .
146 8 · Shapley’s Saddles

A1 A2 A3 A4 Ak
A1 Γ (−1, 1) (1, −1) (−1, 1) · · · (1, −1)
A2 (1, −1) Γ (−1, 1) (1, −1) (−1, 1)
A3 (−1, 1) (1, −1) Γ (−1, 1) (1, −1)
A4 (1, −1) (−1, 1) (1, −1) Γ (−1, 1)
.. .. ..
. . .
Ak (−1, 1) (1, −1) (−1, 1) (1, −1) · · · Γ

Figure 8.3: Payoff structure of a 4k×4k symmetric zero-sum game Γk for an odd integer k.
The game has at least 5k weak saddles.

8.5 A Hardness Result for Weak Saddles

In this section, we establish a relationship between weak saddles of two-player games and
inclusion-maximal cliques of undirected graphs. Our construction is similar to one used
by McLennan and Tourky (2005) to show hardness of various problems concerning Nash
equilibria, and directly yields a result concerning the computational hardness of weak
saddles in general two-player games. This in turn leaves little hope that weak saddles can
be found efficiently in general games.

Theorem 8.10. Deciding whether there exists a weak saddle with a given minimum
number of actions, or one with a given average payoff, is NP-hard, even when re-
stricted to two-player games.

Proof. We provide a reduction from the NP-complete problem CLIQUE (e.g., Papadim-
itriou, 1994a). A clique of an undirected graph G is a subset C ⊆ V such that (i, j) ∈ E
for all distinct i, j ∈ C. For a given graph G and k ∈ N, CLIQUE asks whether G has a
clique of size at least k.
Let G = (V, E) be an undirected graph, and define a two-player game ΓG where both
players have V as their set of actions, and payoffs are given by



 0 if i = j
p1 (i, j) = 1 if (i, j) ∈ E


−1 otherwise,

and

1 if i = j
p2 (i, j) =
0 otherwise.
8.5 · A Hardness Result for Weak Saddles 147

Obviously ΓG can be computed from G in polynomial time. We claim that a pair (X1 , X2 )
is a weak saddle of ΓG if and only if X1 = X2 and X1 is an inclusion-maximal clique of G.

Recall that a WGSP is a pair of subsets of V that is externally stable for both players.
For v ∈ V and U ⊆ V, denote by pi (v, U) = (pi (v, u))u∈U the vector of payoffs for player i
if he plays v and the other player plays some u ∈ U.

We begin by showing that (X1 , X2 ) ⊆ (V, V) is externally stable for player 2 if and
only if ∅ 6= X1 ⊆ X2 . Since this effectively means that player 1 will only play actions he
thinks player 2 might play, we will refer to this property as imitation. For the direction
from left to right, assume that (X1 , X2 ) is externally stable for player 2. Obviously,
X1 6= ∅. Consider any x ∈ X1 and assume x ∈ / X2 . Then there exists x 0 ∈ X2 with
x 0 >X1 x, contradicting the fact that p2 (x 0 , x) = 0 < 1 = p2 (x, x). For the direction
from right to left, consider (X1 , X2 ) ⊆ (V, V) such that ∅ 6= X1 ⊆ X2 . We have to show
that for all x ∈ V \ X2 , there exists x 0 ∈ X2 with x 0 >X1 x. Let x ∈ V \ X2 . Since
X1 ⊆ X2 , it follows that x ∈ / X1 and thus p2 (x, X1 ) = (0, . . . , 0). Now let x 0 ∈ X1 . Then
p2 (x , X1 ) = (0, . . . , 0, 1, 0, . . . , 0), with entry 1 at x 0 . This implies that x 0 >X1 x.
0

We are now ready to prove that a pair (S1 , S2 ) is a weak saddle of ΓG if and only if
S1 = S2 and S1 is an inclusion-maximal clique in G. The statement of the theorem then
follows immediately. For the direction from right to left, consider a maximal clique C
of G. We have to show that (C, C) is a WGSP of ΓG , i.e., externally stable for both
players, and does not contain a WGSP as a proper subset. External stability for player 2
follows directly from imitation. For external stability for player 1, consider any v ∈ V \ C.
Since C is a maximal clique, there must exist some x ∈ C with (v, x) ∈ / E or, equivalently,
p1 (v, x) = −1. Then, x >C v because p1 (x, C) = (1, . . . , 1, 0, 1, . . . , 1), with entry 0
at x. Now assume for contradiction that there exists a WGSP (X1 , X2 ) with X1 ⊆ C and
X2 ⊆ C, such that at least one of the inclusions is strict. By imitation, X1 ⊆ X2 . This
in fact means that X1 must be a strict subset of C, because otherwise (X1 , X2 ) = (C, C).
Consider some x ∈ C \ X1 . Since (X1 , X2 ) is a WGSP, there must exist some x 0 ∈ X1 with
x 0 >X2 x. This is a contradiction, since x 0 ∈ X2 and p1 (x 0 , x 0 ) = 0 < 1 = p1 (x, x 0 ), where
the last equality is due to the fact that both x and x 0 are in the clique C.

For the direction from left to right, let (X1 , X2 ) be a weak saddle of ΓG and observe
that X1 ⊆ X2 by imitation. Further let C be an inclusion-maximal clique in the induced
subgraph G|X1 of G with vertex set X1 . We claim that C is also an inclusion-maximal
clique in G. Then, by the above, (C, C) is a weak saddle of ΓG . Furthermore, C ⊆ X1 ⊆ X2 ,
and thus (X1 , X2 ) = (C, C) by inclusion minimality of (X1 , X2 ). Assume for contradiction
that C is not an inclusion-maximal clique of G, i.e., that there exists some v ∈ V \ C
connected to every vertex in C, such that p1 (v, C) = (1, . . . , 1). Since (X1 , X2 ) is a weak
saddle, there exists x ∈ X1 with x >X2 v. In particular, p1 (x, C) = (1, . . . , 1), implying
that x ∈ / C and that x is connected to all vertices in C. This obviously contradicts
maximality of C in G|X1 .
148 8 · Shapley’s Saddles

8.6 Discussion
In this chapter we have studied computational aspects of Shapley’s saddles—ordinal set-
valued solution concepts dating back to the early 1950s—by proposing polynomial-time
algorithms for computing strict saddles in general normal-form games and pure weak
saddles in a subclass of symmetric two-player zero-sum games. The algorithm for the
latter class of games is highly non-trivial and relies on linear programs that determine the
support of Nash equilibria in certain subgames of the original game. We have also seen
that, in general two-player games, natural problems associated with weak saddles, such as
deciding the existence of a weak saddle of a certain size or one containing a given action,
are NP-hard. Several interesting questions concerning weak saddles, however, remain
open. In particular, it is not known whether weak saddles can be computed efficiently in
general two-player zero-sum games. Furthermore, the aforementioned NP-completeness
results do not imply that finding an arbitrary weak saddle is NP-hard. Finally, gaps
remain between the known upper and lower bounds for different problems in two-player
games, like membership in some weak saddle or uniqueness of a weak saddle.
All of the above results apply to games with a constant number of players and many
actions. It is an interesting question whether strict saddles can still be computed efficiently
in certain classes of games that allow for a compact representation when the number of
players is growing. Similarly, one might ask for classes of games where weak saddles
become tractable. A natural candidate for such a class are anonymous games, studied in
Chapter 5.
 References

T. Abbott, D. Kane, and P. Valiant. On the complexity of two-player win-lose games. In

Proceedings of the 46th Symposium on Foundations of Computer Science (FOCS),
pages 113–122. IEEE Computer Society Press, 2005. [24, 38]

E. Allender, D. A. Mix Barrington, T. Chakraborty, S. Datta, and S. Roy. Grid graph

reachability problems. In Proceedings of the 21st Annual IEEE Conference on Com-
putational Complexity (CCC), pages 299–313, 2006. [85]

K. R. Apt. Uniform proofs of order independence for various strategy elimination proce-
dures. Contributions to Theoretical Economics, 4(1), 2004. [8, 9]

I. Ashlagi, D. Monderer, and M. Tennenholtz. On the value of correlation. In Proceedings

of the 21st Annual Conference on Uncertainty in Artificial Intelligence (UAI),
pages 34–41. AUAI Press, 2005. [32, 51]

R. J. Aumann. Acceptable points in general n-person games. In A. W. Tucker and R. D.

Luce, editors, Contributions to the Theory of Games IV, volume 40 of Annals of
Mathematics Studies, pages 287–324. Princeton University Press, 1959. [127]

R. J. Aumann. Almost strictly competitive games. Journal of the Society of Industrial

and Applied Mathematics, 9(4):544–550, 1961. [31]

R. J. Aumann. Subjectivity and correlation in randomized strategies. Journal of Math-

ematical Economics, 1:67–96, 1974. [10, 30, 46, 50, 51, 58]

R. J. Aumann. On the non-transferable utility value: A comment on the Roth-Shafer

examples. Econometrica, 53(3):667–678, 1985. [32, 133]

R. J. Aumann. Game theory. In J. Eatwell, M. Milgate, and P. Newman, editors, The

New Palgrave: A Dictionary of Economics, volume 2, pages 460–482. MacMillan,
1987. [33]

K. Basu and J. Weibull. Strategy subsets closed under rational behavior. Economics
Letters, 36:141–146, 1991. [138]

149
150 8 · Shapley’s Saddles

P. Beame, S. A. Cook, and H. J. Hoover. Log depth circuits for division and related
problems. SIAM Journal on Computing, 15(4):994–1003, 1986. [72]

M. Benisch, G. B. Davis, and T. Sandholm. Algorithms for rationalizability and CURB

sets. In Proceedings of the 21st National Conference on Artificial Intelligence
(AAAI), pages 598–604. AAAI Press, 2006. [138]

N. Bhat and K. Leyton-Brown. Computing Nash equilibria of action-graph games. In

Proceedings of the 20th Annual Conference on Uncertainty in Artificial Intelligence
(UAI), pages 35–42. AUAI Press, 2004. [59]

B. Bollobás. Modern Graph Theory. Springer-Verlag, 1998. [65]

E. Borel. La théorie du jeu et les équations intégrales à noyau symétrique. Comptes

Rendus de l’Académie des Sciences, 173:1304–1308, 1921. [58]

F. Brandt and F. Fischer. On the hardness and existence of quasi-strict equilibria. In

B. Monien and U.-P. Schroeder, editors, Proceedings of the 1st International Sym-
posium on Algorithmic Game Theory (SAGT), volume 4997 of Lecture Notes in
Computer Science (LNCS), pages 291–302. Springer-Verlag, 2008a. [27]

F. Brandt and F. Fischer. Computing the minimal covering set. Mathematical Social
Sciences, 56(2):254–268, 2008b. [27]

F. Brandt, F. Fischer, and Y. Shoham. On strictly competitive multi-player games. In

Y. Gil and R. Mooney, editors, Proceedings of the 21st National Conference on
Artificial Intelligence (AAAI), pages 605–612. AAAI Press, 2006. [25]

F. Brandt, F. Fischer, P. Harrenstein, and Y. Shoham. A game-theoretic analysis of

strictly competitive multiagent scenarios. In M. Veloso, editor, Proceedings of the 20th
International Joint Conference on Artificial Intelligence (IJCAI), pages 1199–1206,
2007a. [25]

F. Brandt, T. Sandholm, and Y. Shoham. Spiteful bidding in sealed-bid auctions. In

M. Veloso, editor, Proceedings of the 20th International Joint Conference on Arti-
ficial Intelligence (IJCAI), pages 1207–1214, 2007b. [32]

F. Brandt, F. Fischer, and M. Holzer. On iterated dominance, matrix elimination, and

matched paths. ECCC Report TR08-077, Electronic Colloquium on Computational
Complexity (ECCC), 2008a. Working paper. [26]

F. Brandt, F. Fischer, and M. Holzer. Equilibria of graphical games with symmetries.

In C. H. Papadimitriou and S. Zhang, editors, Proceedings of the 4th International
Workshop on Internet and Network Economics (WINE), Lecture Notes in Computer
Science (LNCS), pages 198–209. Springer-Verlag, 2008b. [26]
8.6 · Discussion 151

F. Brandt, M. Brill, F. Fischer, and P. Harrenstein. Computational aspects of Shap-

ley’s saddles. In K. S. Decker and J. S. Sichman, editors, Proceedings of the 8th
International Joint Conference on Autonomous Agents and Multi-Agent Systems
(AAMAS), 2009a. [27]

F. Brandt, M. Brill, F. Fischer, and P. Harrenstein. On the complexity of iterated weak

dominance in constant-sum games. In M. Mavronicolas, editor, Proceedings of the
2nd International Symposium on Algorithmic Game Theory (SAGT), 2009b. To
appear. [59]

F. Brandt, F. Fischer, P. Harrenstein, and Y. Shoham. Ranking games. Artificial Intel-

ligence, 173(2):221–239, 2009c. [25]

F. Brandt, F. Fischer, and M. Holzer. Symmetries and the complexity of pure Nash
equilibrium. Journal of Computer and System Sciences, 75(3):163–177, 2009d. [26]

A. K. Chandra, L. Stockmeyer, and U. Vishkin. Constant depth reducibility. SIAM

Journal on Computing, 13(2):423–439, 1984. [65, 66, 67, 68]

X. Chen and X. Deng. Settling the complexity of 2-player Nash-equilibrium. In Proceed-

ings of the 47th Symposium on Foundations of Computer Science (FOCS), pages
261–272. IEEE Press, 2006. [24, 37, 42, 58, 129]

X. Chen, X. Deng, and S.-H. Teng. Computing Nash equilibria: Approximation and
smoothed complexity. In Proceedings of the 47th Symposium on Foundations of
Computer Science (FOCS), pages 603–612. IEEE Computer Society Press, 2006. [24]

X. Chen, S.-H. Teng, and P. Valiant. The approximation complexity of win-lose games. In
Proceedings of the 18th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA), pages 159–168. SIAM, 2007. [24]

S.-F. Cheng, D. M. Reeves, Y. Vorobeychik, and M. P. Wellman. Notes on equilibria

in symmetric games. In Proceedings of the 6th International Workshop On Game
Theoretic And Decision Theoretic Agents (GTDT), 2004. [67]

B. Codenotti and D. Stefankovic. On the computational complexity of Nash equilibria for

(0, 1) bimatrix games. Information Processing Letters, 94(3):145–150, 2005. [23]

V. Conitzer and T. Sandholm. Complexity of (iterated) dominance. In Proceedings of

the 6th ACM Conference on Electronic Commerce (ACM-EC), pages 88–97. ACM
Press, 2005a. [9, 23, 37, 42, 43, 59, 78, 83, 96, 138]

V. Conitzer and T. Sandholm. A generalized strategy eliminability criterion and compu-

tational methods for applying it. In Proceedings of the 20th National Conference on
Artificial Intelligence (AAAI), pages 483–488. AAAI Press, 2005b. [138]
152 8 · Shapley’s Saddles

V. Conitzer and T. Sandholm. New complexity results about Nash equilibria. Games
and Economic Behavior, 63:621–641, 2008. [23]

R. Cubitt and R. Sugden. Rationally justifiable play and the theory of non-cooperative
games. Economic Journal, 104(425):798–803, 1994. [128, 129, 132]

G. B. Dantzig. A proof of the equivalence of the programming problem and the game
problem. In T. C. Koopmans, editor, Activity Analysis of Production and Allocation,
pages 330–335. John Wiley & Sons Inc., 1951. [23, 37]

C. Daskalakis and C. H. Papadimitriou. The complexity of games on highly regular

graphs. In Proceedings of the 13th Annual European Symposium on Algorithms
(ESA), volume 3669 of Lecture Notes in Computer Science (LNCS), pages 71–82.
Springer-Verlag, 2005. [102, 103, 125]

C. Daskalakis and C. H. Papadimitriou. Computing equilibria in anonymous games. In

Proceedings of the 48th Symposium on Foundations of Computer Science (FOCS),
pages 83–93. IEEE Computer Society Press, 2007. [57]

C. Daskalakis and C. H. Papadimitriou. Discretized multinomial distributions and Nash

equilibria in anonymous games. In Proceedings of the 49th Symposium on Founda-
tions of Computer Science (FOCS). IEEE Computer Society Press, 2008. [59, 102]

C. Daskalakis, P. Goldberg, and C. Papadimitriou. The complexity of computing a Nash

equilibrium. In Proceedings of the 38th Annual ACM Symposium on the Theory of
Computing (STOC), pages 71–78. ACM Press, 2006. [24, 102]

C. Daskalakis, P. Goldberg, and C. Papadimitriou. The complexity of computing a Nash

equilibrium. SIAM Journal on Computing, 2009a. To appear. [24, 37]

C. Daskalakis, G. Schoenebeck, G. Valiant, and P. Valiant. On the complexity of Nash

equilibria of action-graph games. In Proceedings of the 20th Annual ACM-SIAM
Symposium on Discrete Algorithms (SODA), 2009b. [59]

Y. Dodis, S. Halevi, and T. Rabin. A cryptographic solution to a game theoretic prob-

lem. In Proceedings of the 20th Annual International Cryptology Conference
(CRYPTO), volume 1880 of Lecture Notes in Computer Science (LNCS), pages
112–130. Springer-Verlag, 2000. [50]

J. Duggan and M. Le Breton. Dutta’s minimal covering set and Shapley’s saddles. Journal
of Economic Theory, 70:257–265, 1996. [137, 145]

J. Duggan and M. Le Breton. Mixed refinements of Shapley’s saddles and weak tourna-
ments. Social Choice and Welfare, 18(1):65–78, 2001. [137]
8.6 · Discussion 153

J. Dunkel and A. S. Schulz. On the complexity of pure-strategy Nash equilibria in con-

gestion and local-effect games. Mathematics of Operations Research, 33(4):851–868,
2008. [59]

B. Dutta. Covering sets and a new Condorcet choice correspondence. Journal of Eco-
nomic Theory, 44:63–80, 1988. [142]

B. Dutta and J.-F. Laslier. Comparison functions and choice correspondences. Social
Choice and Welfare, 16(4):513–532, 1999. [142, 143, 144]

E. Elkind, L. A. Goldberg, and P. W. Goldberg. Equilibria in graphical games on trees

revisited. In Proceedings of the 7th ACM Conference on Electronic Commerce
(ACM-EC), pages 100–109. ACM Press, 2006. [102]

E. Elkind, L. A. Goldberg, and P. W. Goldberg. Computing good Nash equilibria in graph-

ical games. In Proceedings of the 8th ACM Conference on Electronic Commerce
(ACM-EC), pages 162–171. ACM Press, 2007. [102]

K. Ellul, B. Krawetz, J. Shallit, and M.-W. Wang. Regular expressions: New results
and open problems. Journal of Automata, Languages and Combinatorics, 9(2–3):
233–256, 2004. [85]

K. Etessami and M. Yannakakis. On the complexity of Nash equilibria and other fixed
points. In Proceedings of the 48th Symposium on Foundations of Computer Science
(FOCS), pages 113–123. IEEE Computer Society Press, 2007. [24]

A. Fabrikant, C. H. Papadimitriou, and K. Talwar. The complexity of pure Nash equilibria.

In Proceedings of the 36th Annual ACM Symposium on the Theory of Computing
(STOC), pages 604–612. ACM Press, 2004. [59, 61]

F. Fischer, M. Holzer, and S. Katzenbeisser. The influence of neighbourhood and choice

on the complexity of finding pure Nash equilibria. Information Processing Letters, 99
(6):239–245, 2006. [26, 63]

J. W. Friedman. On characterizing equilibrium points in two person strictly competitive

games. International Journal of Game Theory, 12:245–247, 1983. [31]

H. N. Gabow, S. N. Maheshwari, and L.J. Osterweil. On two problems in the generation

of program test paths. IEEE Transactions on Software Engineering, 2(3):227–231,
1976. [85]

D. Gale, H. W. Kuhn, and A. W. Tucker. On symmetric games. In H. W. Kuhn and

A. W. Tucker, editors, Contributions to the Theory of Games, volume 1, pages 81–
87. Princeton University Press, 1950. [58, 129, 134]
154 8 · Shapley’s Saddles

M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory

of NP-Completeness. W. H. Freeman, 1979. [75]

I. Gilboa and E. Zemel. Nash and correlated equilibria: Some complexity considerations.
Games and Economic Behavior, 1(1):80–93, 1989. [23]

I. Gilboa, E. Kalai, and E. Zemel. The complexity of eliminating dominated strategies.

Mathematics of Operations Research, 18(3):553–565, 1993. [23, 59]

F. Glover. Maximum matching in convex bipartite graphs. Naval Research Logistics

Quarterly, 14:313–316, 1967. [84]

P. W. Goldberg and C. H. Papadimitriou. Reducibility among equilibrium problems. In

Proceedings of the 38th Annual ACM Symposium on the Theory of Computing
(STOC), pages 61–70. ACM Press, 2006. [24, 102, 125]

O. Goldreich. Computational Complexity: A Conceptual Perspective. Cambridge

University Press, 2008. [5]

G. Gottlob, G. Greco, and F. Scarcello. Pure Nash equilibria: Hard and easy games.
Journal of Artificial Intelligence Research, 24:195–220, 2005. [26, 37, 59, 101, 102, 103, 105]

S. Govindan and R. Wilson. A global Newton method to compute Nash equilibria. Journal
of Economic Theory, 110(1):65–86, 2003. [59]

R. Greenlaw, H. J. Hoover, and W. L. Ruzzo. Limits to Parallel Computation. Oxford

University Press, 1995. [85]

J. C. Harsanyi. Oddness of the number of equilibrium points: A new proof. International

Journal of Game Theory, 2:235–250, 1973. [10, 27, 127, 129]

S. Ieong, R. McGrew, E. Nudelman, Y. Shoham, and Q. Sun. Fast and compact: A

simple class of congestion games. In Proceedings of the 20th National Conference on
Artificial Intelligence (AAAI), pages 489–494. AAAI Press, 2005. [59, 61]

N. Immerman. Nondeterministic space is closed under complementation. SIAM Journal

on Computing, 17(5):935–938, 1988. [15]

M. J. M. Jansen. Regularity and stability of equilibrium points of bimatrix games. Math-

ematics of Operations Research, 6(4):530–550, 1981. [130]

M. J. M. Jansen. Regular equilibrium points of bimatrix games. OR Spektrum, 9(2):

82–92, 1987. [129]

A. X. Jiang and K. Leyton-Brown. Computing pure Nash equilibria in symmetric action-

graph games. In Proceedings of the 22nd AAAI Conference on Artificial Intelligence
(AAAI), pages 79–85. AAAI Press, 2007. [59]
8.6 · Discussion 155

D. S. Johnson, C. H. Papadimitriou, and M. Yannakakis. How easy is local search?

Journal of Computer and System Sciences, 37:79–100, 1988. [17, 72, 74]

N. D. Jones. Space-bounded reducibility among combinatorial problems. Journal of

Computer and System Sciences, 11:68–85, 1975. [110]

A. P. Jurg, M. J. M. Jansen, J. A. M. Potters, and S. H. Tijs. A symmetrization for

finite two-person games. ZOR – Methods and Models of Operations Research, 36(2):
111–123, 1992. [130]

A. Kats and J.-F. Thisse. Unilaterally competitive games. International Journal of

Game Theory, 21:291–299, 1992. [32, 136]

M. J. Kearns, M. L. Littman, and S. P. Singh. Graphical models for game theory. In

Proceedings of the 17th Annual Conference on Uncertainty in Artificial Intelligence
(UAI), pages 253–260. Morgan Kaufmann, 2001. [101]

L. Khachiyan. A polynomial algorithm in linear programming. Soviet Mathematics

Doklady, 20:191–194, 1979. [23, 37]

D. E. Knuth, C. H. Papadimitriou, and J. N. Tsitsiklis. A note on strategy elimination in

bimatrix games. Operations Research Letters, 7:103–107, 1988. [23, 59, 138]

M. Kojima, A. Okada, and S. Shindoh. Strongly stable equilibrium points of N-person

noncooperative games. Mathematics of Operations Research, 10(4):650–663, 1985.
[132]

E. Koutsoupias and C. Papadimitriou. Worst-case equilibria. In Proceedings of the 16th

International Symposium on Theoretical Aspects of Computer Science (STACS),
volume 1563 of Lecture Notes in Computer Science (LNCS), pages 404–413. Springer-
Verlag, 1999. [32, 51]

G. Laffond, J.-F. Laslier, and M. Le Breton. More on the tournament equilibrium set.
Mathématiques et sciences humaines, 31(123):37–44, 1993. [143]

J.-F. Laslier. Tournament Solutions and Majority Voting. Springer-Verlag, 1997. [142]

K. Leyton-Brown and M. Tennenholtz. Local-effect games. In Proceedings of the 18th

International Joint Conference on Artificial Intelligence (IJCAI), pages 772–777,
2003. [59]

R. D. Luce and H. Raiffa. Games and Decisions: Introduction and Critical Survey.
John Wiley & Sons Inc., 1957. [46, 58, 130]

L. M. Marx and J. M. Swinkels. Order independence for iterated weak dominance. Games
and Economic Behavior, 18:219–245, 1997. [43]
156 8 · Shapley’s Saddles

A. McLennan and R. Tourky. Simple complexity from imitation games. Unpublished

manuscript, 2005. [134, 146]

N. Megiddo and C. H. Papadimitriou. On total functions, existence theorems and com-

putational complexity. Theoretical Computer Science, 81:317–324, 1991. [15, 23]

D. Monderer and L. S. Shapley. Potential games. Games and Economic Behavior, 14

(1):124–143, 1996. [136]

J. Morgan, K. Steiglitz, and G. Reis. The spite motive and equilibrium behavior in
auctions. Contributions to Economic Analysis & Policy, 2(1):1102–1127, 2003. [32]

H. Moulin. Dominance solvable voting schemes. Econometrica, 47:1337–1351, 1979. [9]

H. Moulin and J.-P. Vial. Strategically zero-sum games: The class of games whose com-
pletely mixed equilibria cannot be improved upon. International Journal of Game
Theory, 7(3–4):201–221, 1978. [30, 32]

R. B. Myerson. Game Theory: Analysis of Conflict. Harvard University Press, 1991.

[5, 23]

J. F. Nash. Non-cooperative games. Annals of Mathematics, 54(2):286–295, 1951. [9, 15,

23, 58, 133, 137]

N. Nisan, T. Roughgarden, É. Tardos, and V. Vazirani. Algorithmic Game Theory.

Cambridge University Press, 2007. [2]

H. Norde. Bimatrix games have quasi-strict equilibria. Mathematical Programming, 85:

35–49, 1999. [35, 127, 128, 129]

H. Norde, J. Potters, H. Reijnierse, and D. Vermeulen. Equilibrium selection and consis-

tency. Games and Economic Behavior, 12(2):219–225, 1996. [128, 129]

J. B. Orlin, A. P. Punnen, and A. S. Schulz. Approximate local search in combinatorial

optimization. SIAM Journal on Computing, 33(5):1201–1214, 2004. [18]

M. J. Osborne and A. Rubinstein. A Course in Game Theory. MIT Press, 1994. [5]

C. H. Papadimitriou. Computational Complexity. Addison-Wesley, 1994a. [3, 5, 45, 69, 72,

105, 108, 134, 146]

C. H. Papadimitriou. On the complexity of the parity argument and other inefficient proofs
of existence. Journal of Computer and System Sciences, 48(3):498–532, 1994b. [16]

C. H. Papadimitriou. Algorithms, games, and the internet. In Proceedings of the 33rd

Annual ACM Symposium on the Theory of Computing (STOC), pages 749–753.
ACM Press, 2001. [23]
8.6 · Discussion 157

C. H. Papadimitriou. Computing correlated equilibria in multi-player games. In Proceed-

ings of the 37th Annual ACM Symposium on the Theory of Computing (STOC),
pages 49–56. ACM Press, 2005. [37]

C. H. Papadimitriou. The complexity of finding Nash equilibria. In N. Nisan, T. Rough-

garden, É. Tardos, and V. Vazirani, editors, Algorithmic Game Theory, chapter 2,
pages 29–78. Cambridge University Press, 2007. [3]

C. H. Papadimitriou and T. Roughgarden. Computing equilibria in multi-player games.

In Proceedings of the 16th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA), pages 82–91. SIAM, 2005. [21, 58, 124, 134]

I. Parberry. On the computational complexity of optimal sorting network verification.

In Proceedings of the Conference on Parallel Architectures and Languages Europe
(PARLE), volume 505 of Lecture Notes in Computer Science (LNCS), pages 252–
269. Springer-Verlag, 1991. [86]

R. Parikh. On context-free languages. Journal of the ACM, 13(4):570–581, 1966. [60]

B. Peleg and S. Tijs. The consistency principle for games in strategic form. International
Journal of Game Theory, 25:13–34, 1996. [128, 129]

A. Quesada. Another impossibility result for normal form games. Theory and Decision,
52:73–80, 2002. [129]

T. E. S. Raghavan. Non-zero-sum two-person games. In R. J. Aumann and S. Hart, editors,

Handbook of Game Theory with Economic Applications, volume III, chapter 17,
pages 1687–1721. North-Holland, 2002. [32]

N. Robertson, P. D. Seymour, and R. Thomas. Permanents, Pfaffian orientations, and

even directed circuits. Annals of Mathematics, 150:929–975, 1999. [119, 124]

R. W. Rosenthal. A class of games possessing pure-strategy Nash equilibria. International

Journal of Game Theory, 2:65–67, 1973. [59]

T. Roughgarden. Selfish Routing and the Price of Anarchy. MIT Press, 2005. [32]

W. Ruzzo. On uniform circuit complexity. Journal of Computer and System Sciences,

22(3):365–383, 1981. [19]

T. J. Schaefer. The complexity of satisfiability problems. In Proceedings of the 10th

Annual ACM Symposium on the Theory of Computing (STOC), pages 216–226.
ACM Press, 1978. [119, 123, 124]

A. A. Schäffer and M. Yannakakis. Simple local search problems that are hard to solve.
SIAM Journal on Computing, 20(1):56–87, 1991. [17, 72, 74]
158 8 · Shapley’s Saddles

G. Schoenebeck and S. Vadhan. The computational complexity of Nash equilibria in con-

cisely represented games. In Proceedings of the 7th ACM Conference on Electronic
Commerce (ACM-EC), pages 270–279. ACM Press, 2006. [37, 59, 63, 105, 111, 112]

P. D. Seymour. On the two-colouring of hypergraphs. The Quarterly Journal of Math-

ematics, 25:303–312, 1974. [117, 118]

L. Shapley. Order matrices. I. Technical Report RM-1142, The RAND Corporation, 1953a.
[27, 137, 138]

L. Shapley. Order matrices. II. Technical Report RM-1145, The RAND Corporation,
1953b. [137]

L. Shapley. A condition for the existence of saddle-points. Technical Report RM-1598,

The RAND Corporation, 1955. [137]

L. Shapley. Some topics in two-person games. In M. Dresher, L. S. Shapley, and A. W.

Tucker, editors, Advances in Game Theory, volume 52 of Annals of Mathematics
Studies, pages 1–29. Princeton University Press, 1964. [10, 137, 140]

P. Spirakis. Approximate equilibria for strategic two person games. In Proceedings of the
1st International Symposium on Algorithmic Game Theory (SAGT), pages 5–21,
2008. [24]

D. Squires. Impossibility theorems for normal form games. Theory and Decision, 44:
67–81, 1998. [129]

R. Szelepcsenyi. The method of forced enumeration for nondeterministic automata. Acta

Informatica, 26(3):279–284, 1988. [15]

M. Tennenholtz. Competitive safety analysis: Robust decision-making in multi-agent

systems. Journal of Artificial Intelligence Research, 17:363–378, 2002. [32]

C. Thomassen. Even cycles in directed graphs. European Journal of Combinatorics, 6:

85–89, 1985. [118]

E. van Damme. Refinements of the Nash Equilibrium Concept. Springer-Verlag, 1983.

[36, 128, 129, 132]

E. van Damme. Stability and Perfection of Nash Equilibria. Springer-Verlag, 2nd

edition, 1991. [129]

E. van Damme. On the state of the art in game theory: An interview with Robert Aumann.
Games and Economic Behavior, 24:181–210, 1998. [2]

H. Vollmer. Introduction to Circuit Complexity. Springer-Verlag, 1999. [5]

8.6 · Discussion 159

J. von Neumann. Zur Theorie der Gesellschaftspiele. Mathematische Annalen, 100:

295–320, 1928. [47, 58, 130]

J. von Neumann. First draft of a report on the EDVAC. Technical report, Moore School
of Electrical Engineering, University of Pennsylvania, 1945. [2]

J. von Neumann and O. Morgenstern. Theory of Games and Economic Behavior.

Princeton University Press, 1944. [1, 5]

J. von Neumann and O. Morgenstern. Theory of Games and Economic Behavior.

Princeton University Press, 2nd edition, 1947. [6, 29, 58, 98]

O. De Wolf. Optimal strategies in n-person unilaterally competitive games. Discussion

Paper 9949, Center for Operations Research and Econometrics, Université catholique
de Louvain, 1999. [32]
 Lebenslauf

Felix Fischer wurde am 11. März 1978 in München geboren. Von 1984 bis 1997 besuchte er
Grundschule und Gymnasium in Garching bei München. Nach Erfüllung der allgemeinen
Wehrpflicht begann er 1998 ein Studium der Informatik mit Nebenfach Elektrotechnik
an der Technischen Universität München, das er im Jahr 2003 mit dem Abschluß zum
Diplom-Informatiker beendete. Nach einer gut einjährigen Tätigkeit als wissenschaftlicher
Mitarbeiter am Lehrstuhl für Theoretische Informatik und Grundlagen der Künstlichen
Intelligenz der Technischen Universität München begann er im Jahr 2005 schließlich mit
der Promotion in der Forschungsgruppe Präferenzbündelung in Multiagentensystemen an
der Lehr- und Forschungseinheit für Theoretische Informatik der Ludwig-Maximilians-
Universität München. Im Jahr 2007 war er über einen Zeitraum von insgesamt fünf
Monaten als Gastwissenschaftler an der Hebrew University of Jerusalem in Israel tätig.

161