GAME THEORY:

Analysis of Strategic Thinking

Pierpaolo Battigalli

September 2014

Abstract

These lecture notes introduce some concepts of the theory of games:

rationality, dominance, rationalizability and several notions of equilibrium
(Nash, randomized, correlated, self-confirming, subgame perfect, Bayesian
perfect equilibrium). For each of these concepts, the interpretative aspect is emphasized. Even though no advanced mathematical knowledge is
required, the reader should nonetheless be familiar with the concepts of
set, function, probability, and more generally be able to follow abstract
reasoning and formal arguments.

Milano,
P. Battigalli

Preface

These lecture notes provide an introduction to game theory, the formal analysis of strategic interaction. Game theory now pervades most
non-elementary models in microeconomic theory and many models in the
other branches of economics and in other social sciences. We introduce
the necessary analytical tools to be able to understand these models, and
illustrate them with some economic applications.
We also aim at developing an abstract analysis of strategic thinking, and a critical and open-minded attitude toward the standard gametheoretic concepts as well as new concepts.
Most of these notes rely on relatively elementary mathematics. Yet,
our approach is formal and rigorous. The reader should be familiar with
mathematical notation about sets and functions, with elementary linear
algebra and topology in Euclidean spaces, and with proofs by mathematical
induction. Elementary calculus is sometimes used in examples.

[Warning: These notes are work in progress! They probably contain some
mistakes or typos. They are not yet complete. Many parts have been translated
by assistants from Italian lecture notes and the English has to be improved.
Examples and additional explanations and clarifications have to be added. So,
the lecture notes are a poor substitute for lectures in the classroom.]

Contents
1 Introduction
1.1 Decision Theory and Game Theory . . . . . . . . . .
1.2 Why Economists Should Use Game Theory . . . . .
1.3 Abstract Game Models . . . . . . . . . . . . . . . . .
1.4 Terminology and Classification of Games . . . . . . .
1.5 Rational Behavior . . . . . . . . . . . . . . . . . . .
1.6 Assumptions and Solution Concepts in Game Theory

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

Static Games

1
1
2
4
7
11
13

22

2 Static Games: Formal Representation and Interpretation 23

3 Rationality and Dominance
28
3.1 Probabilities and Expected Payoff . . . . . . . . . . . . . . 28
3.2 Conjectures . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Best Replies and Undominated Actions . . . . . . . . . . . 36
4 Rationalizability and Iterated Dominance
4.1 Assumptions about players beliefs . . . . .
4.2 The Rationalization Operator . . . . . . . .
4.3 Rationalizability: The Powers of . . . . .
4.4 Iterated Dominance . . . . . . . . . . . . .
4.5 Rationalizability and Iterated Dominance in

. . .
. . .
. . .
. . .
Nice

. . . .
. . . .
. . . .
. . . .
Games

.
.
.
.
.

.
.
.
.
.

50
51
53
55
61
66

5 Equilibrium
72
5.1 Nash equilibrium . . . . . . . . . . . . . . . . . . . . . . . . 75
5.2 Probabilistic Equilibria . . . . . . . . . . . . . . . . . . . . . 82
v

6 Learning Dynamics, Equilibria and Rationalizability

107
6.1 Perfectly observable actions . . . . . . . . . . . . . . . . . . 108
6.2 Imperfectly observable actions . . . . . . . . . . . . . . . . . 111
7 Games with incomplete information
7.1 Games with Incomplete information . . . . . . . . . . .
7.2 Equilibrium and Beliefs: the simplest case . . . . . . . .
7.3 The General Case: Bayesian Games and Equilibrium . .
7.4 Incomplete Information and Asymmetric Information .
7.5 Rationalizability in Bayesian Games . . . . . . . . . . .
7.6 The Electronic Mail Game . . . . . . . . . . . . . . . . .
7.7 Selfconfirming Equilibrium and Incomplete Information

II

.
.
.
.
.
.
.

.
.
.
.
.
.
.

Dynamic Games

8 Multistage Games with Complete Information

8.1 Preliminary Definitions . . . . . . . . . . . . . .
8.2 Backward Induction and Subgame Perfection . .
8.3 The One-Shot-Deviation Principle . . . . . .
8.4 Some Simple Results About Repeated Games . .
8.5 Multistage Games With Chance Moves . . . . . .
8.6 Randomized Strategies . . . . . . . . . . . . . . .
8.7 Appendix . . . . . . . . . . . . . . . . . . . . . .

113
114
123
127
138
140
159
163

168
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

170
171
185
194
200
205
208
217

9 Multistage Games with Incomplete Information

9.1 Multistage game forms . . . . . . . . . . . . . . . . .
9.2 Dynamic environments with incomplete information
9.3 Dynamic Bayesian Games . . . . . . . . . . . . . . .
9.4 Bayes Rule . . . . . . . . . . . . . . . . . . . . . . .
9.5 Perfect Bayesian Equilibrium . . . . . . . . . . . . .
9.6 Signaling Games and Perfect Bayesian Equilibrium .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

222
222
224
226
228
231
233

Bibliography

.
.
.
.
.
.
.

239

Introduction
Game theory is the formal analysis of the behavior of interacting
individuals. The crucial feature of an interactive situation is that the
consequences of the actions of an individual depend (also) on the actions
of other individuals. This is typical of many games people play for fun,
such as chess or poker. Hence, interactive situations are called games
and interactive individuals are called players. If a players behavior is
intentional and he is aware of the interaction (which is not always the
case), he should try and anticipate the behavior of other players. This is
the essence of strategic thinking. In the rest of this chapter we provide a
semi-formal introduction of some key concepts in game theory.

1.1

Decision Theory and Game Theory

Decision theory is a branch of mathematics that analyzes the decision

problem of an individual (or a group of individuals acting as a single
decision unit) in isolation. The external environment is a primitive of
the decision problem. Decision theory provides simple decision criteria
characterizing an individuals preferences over different courses of action,
provided that these preferences satisfy some rationality properties, such
as completeness (any two alternatives are comparable) and transitivity (if
a is preferred to b and b is preferred to c , then a is preferred to c). These
criteria are used to find optimal (or rational) decisions.
Game theory could be more appropriately called interactive
decision theory. Indeed, game theory is a branch of mathematics

1. Introduction

that analyzes interactive decision problems: There are several individuals,

called players, each facing a decision problem whereby the external
environment (from the point of view of this particular player) is given
by the other players behavior (and possibly some random variables). In
other words, the welfare (utility, payoff, final wealth) of each player is
affected, not only by his own behavior, but also by the behavior of other
players. Therefore, in order to figure out the best course of action each
player has to guess which course of action the other players are going to
take.

1.2

Why Economists Should Use Game Theory

We are going to argue that game theory should be the main analytical
tool used to build formal economic models. More generally, game theory
should be used in all formal models in the social sciences that adhere
to methodological individualism, i.e. try to explain social phenomena as
the result of the actions of many agents, which in turn are freely chosen
according to some consistent criterion.
The reader familiar with the pervasiveness of game theory in economics
may wonder why we want to stress this point. Isnt it well known that game
theory is used in countless applications to model imperfect competition,
bargaining, contracting, political competition and, in general, all social
interactions where the action of each individual has a non-negligible effect
on the social outcome? Yes, indeed! And yet it is often explicitly or
implicitly suggested that game theory is not needed to model situations
where each individual is negligible, such as perfectly competitive markets.
We are going to explain why this is in our view incorrect: the
bottom line will be that every complete formal model of an economic (or
social) interaction must be a game; economic theory has analyzed perfect
competition by taking shortcuts that have been very fruitful, but must be
seen as such, just shortcuts.1
If we subscribe to methodological individualism, as mainstream
economists claim to do, every social or economic observable phenomena
we are interested in analyzing should be reduced to the actions of the
1
Our view is very similar to the original motivations for the study of game theory
offered by the founding fathers von Neumann and Morgenstern in their seminal book
The Theory of Games and Economic Behavior.

1.2. Why Economists Should Use Game Theory

individuals who form the social or economic system. For example, if we

want to study prices and allocations, then we should specify which actions
the individuals in the system can choose and how prices and allocations
depend on such actions: if I = {1, 2, ..., n} is the set of agents, p is the
vector of prices, y is the allocation and a = (ai )iI is the profile2 of actions,
one for each agent, then we should specify relations p = f (a) and y = g(a).
This is done in all models of auctions. For example, in a sealed-bid, firstprice, single-object auction, a = (a1 , ..., an ) is the profile of bids by the n
bidders for the object on sale, f (a1 , ..., an ) = max{a1 , ..., an } (the object is
sold for a price equal to the highest bid) and g(a) is such that the object
is allocated to the highest bidder,3 who has to pay his bid. To be more
general, we have to allow the variables of interest to depend also on some
exogenous shocks x, as in the functional forms p = f (a, x), y = g(a, x).
Furthermore, we should account for dynamics when choices and shocks
take place over time, as in y t = g t (a1 , x1 , ..., at , xt ). Of course, all the
constraints on agents choices (such as those determined by technology)
should also be explicitly specified. Finally, if we are to explain choices
according to some rationality criterion, we should include in the model the
preferences of each individual i over possible outcomes. This is what we
call a complete model of the interactive situation.4 We call variables,
such as y, that depend on actions (and exogenous shocks) endogenous.
(Actions themselves are endogenous in a trivial sense.) The rationale for
this terminology is that we try to analyze/explain actions and variables
that depend on actions.
At this point you may think that this is just a trite repetition of some
abstract methodology of economic modelling. Well, think twice! The most
standard concept in the economists toolkit, Walrasian equilibrium, is not
based on a complete model and is able to explain prices and allocations only
by taking a two-steps shortcut: (1) The modeler pretends that prices
2

Given a set of n individuals, we always call profile a list of elements from a given
set, one for each individual i. For instance, a = (a1 , ..., an ) typically denotes a profile of
actions.
3
Ties can be broken at random.
4
The model is still in some sense incomplete: we have not even addressed the issue of
what the individuals know about the situation and about each other. But the elements
sketched in the main text are sufficient for the discussion. Let us stress that what we call
complete model does not include the modelers hypotheses on how the agents choose,
which are key to provide explanations or predictions of economic/social phenomena.

1. Introduction

are observed and taken as parametrically given by all agents (including all
firms) before they act, hence before they can affect such prices; this is a
kind of logical short-circuit, but it allows to determine demand and supply
functions D(p), S(p). Next, (2) market-clearing conditions D(p) = S(p)
determine equilibrium prices. Well, this can only be seen as a (clever)
reduced-form approach; absent an explicit model of price formation (such
as an auction model), the modeler postulates that somehow the choicesprices-choices feedback process has reached a rest point and he describes
this point as a market-clearing equilibrium. In many applications of
economic theory to the study of competitive markets, this is a very
reasonable and useful shortcut, but it remains just a shortcut, forced by
the lack of what we call a complete model of the interactive situation.
So, what do we get when instead we have a complete model? As we are
going to show a bit more formally in the next section, we get what game
theorists call a game. This is why game theory should be a basic tool
in economic modelling, even if one wants to analyze perfectly competitive
situations. To illustrate this point, we will present a purely game-theoretic
analysis of a perfectly competitive market, showing not only how such an
analysis is possible, but also that it adds to our understanding of how
equilibrium can be reached.

1.3

Abstract Game Models

A completely formal definition of the mathematical object called (noncooperative) game will be given in due time. We start with a semi-formal
introduction to the key concepts, illustrated by a very simple example,
a seller-buyer mini-game. Consider two individuals, S (Seller) and B
(Buyer). Let S be the owner of an object and B a potential buyer. For
simplicity, consider the following bargaining protocol: S can ask one Euro
(1) or two Euros (2) to sell the object, B can only accept (a) or reject (r).
The monetary value of the object for individual i (i = S, B) is denoted Vi .
This situation can be analyzed as a game which can be represented with
a rooted tree with utility numbers attached to terminal nodes (leaves),
player labels attached to non terminal nodes, and action labels attached
to arcs:
The game tree represents the formal elements of the analysis: the set

1.3. Abstract Game Models

5
S

1 VS
VB 1

0
0

2 VS
VB 2

0
0

Figure 1.1: A seller-buyer mini-game tree.

of players (or roles in the game, such as seller and buyer), the actions, the
rules of interaction, the consequences of complete sequences of actions and
how players value such consequences.
Formally, a game form is given by the following elements:
I, a set of players.
For each player i I, a set Ai of actions which could conceivably
be chosen by i at some point of the game as the play unfolds.
C, a set of consequences.
E, an extensive form, that is, a mathematical representation of the
rules saying whose turn it is to move, what a player knows, i.e., his
or her information about past moves and random events, and what
are the feasible actions at each point of the game; this determines
a set Z of possible paths of play (sequences of feasible actions); a
path of play z Z may also contain some random events such as the
outcomes of throwing dice;
g : Z C, a consequence (or outcome) function which assigns
to each play z a consequence g(z) in C.

1. Introduction

The above elements represent what the layperson would call the rules
of the game. To complete the description of the actual interactive
situation (which may differ from how the players perceive it) we have
to add players preferences over consequences (and - via expected utility
calculations - lotteries over consequences):

(vi )iI , where each vi : C R is a von NeumannMorgenstern utility function representing player is preferences
over consequences; preferences over lotteries over consequences are
obtained via expected utility comparisons.

With this, we obtain the kind of mathematical structure called game

in the technical sense of game theory. This formal description does not
say how such a game would (or should) be played. The description of an
interactive decision problem is only the first step to make a prediction (or
a prescription) about players behavior.
To illustrate, in the seller-buyer mini-game: I = {S, B}; AS = {1, 2},
AB = {a, r}; C is a specification of the set of final allocations of the object
and of monetary payments (for example, C = {oS , oB } {0, 1, 2}, where
(oi , k) means object to i and k euros are given by B to S); E is represented
by the game tree, without the utility numbers at endnodes; g is the rule
stipulating that if B accepts the ask price p then B gets the object and
gives p euros to S, if B rejects then S keeps the object and B keeps the
money (g(p, a) = (oB , p), g(p, r) = (oS , 0)); vi is a risk-neutral (quasilinear) utility function normalized so that the utility of no exchange is
zero for both players (vS (oB , p) = p VS , vB (oB , p) = VB p, vi (oS , 0) = 0
for i = S, B). Game theory provides predictions on the behavior of S
and B in this game based on hypotheses about players knowledge of the
rules of the game and of each other preferences, and on hypotheses about
strategic thinking. For example, B is assumed to accept an ask price if
and only if it is below his valuation. Whether the ask price is high or low
depends on the valuation of S and what S knows about the valuation of
B. If S knows the valuation of B, he can anticipate Bs response to each
offer and how much surplus he can extract.

1.4. Terminology and Classification of Games

1.4

Terminology and Classification of Games

Games come in different varieties and are analyzed with different

methodologies. The same strategic interaction may be represented in a
detailed way, using the mathematical objects described above, or in a much
more parsimonious way. The amount of details in the formal representation
constrains the methods used in the analysis. The terminology used to refer
to different kinds of strategic situations and to different formal objects
used to represent them may be confusing: Some terms that have almost
the same meaning in the daily language, such as perfect information and
complete information have very different meaning in game theory; other
terms, such as non-cooperative game, may be misleading. Furthermore,
there is a tendency to confuse the substantive properties of the situations
of strategic interaction that game theory aims at studying and the formal
properties of the mathematical structures used in this endeavour. Here we
briefly summarize the terminology and classification of game theory doing
our best to dispel such confusion.

1.4.1

Cooperative vs Non-Cooperative Games

Suppose that the players (or at least some of them) could meet before the
game is played and sign a binding agreement specifying their course of
action (an external enforcing agency e.g. the courts system will force
each player to follow the agreement). This could be mutually beneficial
for them and if this possibility exists we should take it into account. But
how? The theory of cooperative games does not model the process
of bargaining (offers and counteroffers) which takes place before the real
game starts; this theory considers instead a simple and parsimonious
representation of the situation, that is, how much of the total surplus
each possible coalition of players can guarantee to itself by means of some
binding agreement. For example, in the seller-buyer situation discussed
above, the (normalized) surplus that each player can guarantee to itself is
zero, while the surplus that the {S, B} coalition can guarantee to itself is
VS + VB . This simplified representation is called coalitional game. For
every given coalitional game the theory tries to figure out which division
of the surplus could result, or, at least, the set of allocations that are not
excluded by strategic considerations (see, for example, Part IV of Osborne
and Rubinstein, 1994).

1. Introduction

On the other hand, the theory of non-cooperative games assumes

that either binding agreements are not feasible (e.g. in an oligopolistic
market they could be forbidden by antitrust laws), or that the bargaining
process which could lead to a binding agreement on how to play a game
G is appropriately formalized as a sequence of moves in a larger game
(G). The players cannot agree on how to bargain in (G) and this is the
game to be analyzed with the tools of non-cooperative game theory. It is
therefore argued that non-cooperative game theory is more fundamental
than cooperative game theory. For example, in the seller-buyer situation
(G) would be the game-tree displayed above to illustrate the formal
objects comprised by the mathematical representation of a game. The
analysis of this game reveals that if S knows VB he can use his first-mover
advantage to ask for the highest price below VB . Of course, the results of
the analysis depend on details of the bargaining rules that may be unknown
to an external observer and analyst. Neglecting such details, cooperative
game theory typically gives weaker but more robust predictions. But in
our view, the main advantage of non-cooperative game theory is not so
much its sharper predictions, but rather the conceptual clarity of working
with explicit assumptions and showing exactly how different assumptions
lead to different results.
Cooperative game-theory is an elegant and useful analytical toolkit.
It is especially appropriate when the analyst has little understanding of
the true rules of interaction and wishes to derive some robust results from
parsimonious information about the outcomes that coalitions of players can
implement. However, non-cooperative game theory is much more pervasive
in modern economics, and more generally in modern formal social sciences.
We will therefore focus on non-cooperative game theory.
One thing should be clear, focusing on this part of the theory does
not mean that cooperation is neglected. Non-cooperative game theory is
not the study of un-cooperative behavior, but rather a method of analysis.
Indeed we find the name non-cooperative game theory misleading, but
it is now entrenched and we will conform to it.

1.4.2

Static and Dynamic Games

A game is static if each player moves only once and all players move
simultaneously (or at least without any information about other players
moves). Examples of static games are: Matching Pennies, Stone-Scissor-

1.4. Terminology and Classification of Games

Paper, and sealed bid auctions.

A game is dynamic if some moves are sequential and some player may
observe (at least partially) the behavior of his or her co-players. Examples
of dynamic games are: Chess, Poker, and open outcry auctions.
Dynamic games can be analyzed as if the players moved only once
and simultaneously. The trick is to pretend that each player chooses in
advance his strategy, i.e. a contingent plan of action specifying how
to behave in every circumstance that may arise while playing the game.
Each profile of strategies s = (si )iI determines a particular path, viz.
(z), hence an outcome g((s)), and ultimately a profile of utilities, or
payoffs, (vi (g((s))))iI . This mapping s 7 (vi (g((s))))iI is called
the normal form, or strategic form of the game. The strategic form
of a game can be seen as a static game where players simultaneously
choose strategies in advance. For example, in the seller-buyer mini-game
the strategies of S (seller) coincide with his possible ask prices; on the
other hand, the set of strategies of B (buyer) contains four response
rules: SB = {a1 a2 , a1 r2 , r1 a2 , r1 r2 } where ap (rp ) is the instruction accept
(reject) price p. The strategic form is as follows:

S \B

a1 a2

a1 r2

r1 a2

r1 r2

p=1

1 VS , VB 1

1 VS , VB 1

0, 0

0, 0

p=2

2 VS , VB 2

0, 0

2 VS , VB 2

0, 0

Figure 1.2: Strategic form of the seller-buyer minigame

A strategy in a static game is just the plan to take a specific action, with
no possibility to make this choice contingent on the actions of others, as
such actions are being simultaneously chosen. Therefore, the mathematical
representation of a static game and its normal form must be the same.
For this reason, static games are also called normal-form games, or
strategic-form games. This is an unfortunately widespread abuse of
language. In a correct language there should not be normal-form games
; rather, looking at the normal form of a game is a method of analysis.

10

1.4.3

1. Introduction

Assumptions About Information

Perfect information and observable actions

A dynamic game has perfect information if players move one at a time
and each player when it is his or her turn to move is informed of all
the previous moves (including the realizations of chance moves). If some
moves are simultaneous, but each player can observe all past moves, we
say that the game has observable actions (or perfect monitoring,
or almost perfect information). Examples of games with perfect
information are the seller-buyer mini-game, Chess, Backgammon, Tic-TacToe. An example of a game with observable actions is the repeated Cournot
oligopoly, with simultaneous choice of outputs in every period and perfect
monitoring of past outputs. Note, perfect information is an assumption
about the rules of the game.
Asymmetric information
A game with imperfect information features asymmetric information if
different players get different pieces of information about past moves and
in particular about the realizations of chance moves. Poker and Bridge are
games with asymmetric information because players observe their cards,
but not the cards of others. Like perfect information, also asymmetric
information is entailed by the rules of the game.
Complete and incomplete information
An event E is common knowledge if everybody knows E, everybody
knows that everybody knows E, and so on for all iterations of everybody
knows that. To use a suggestive semi-formal expression: for every m =
1, 2, ... it is the case that [(everybody knows that)m E is the case].
There is complete information in an interactive decision problem
represented by a game G = hI, A, C, E, g, (vi )iI i if it is common knowledge
that G is the actual game to be played. Conversely, there is incomplete
information if for some player i either it is not the case that [i knows that
G is the actual game] or for some m = 1, 2, ... it is not the case that [i knows
that (everybody knows that)m G is the actual game]. Most economic
situations feature incomplete information because either the outcome
function, g, or players preferences, (vi )iI , are not common knowledge.

1.5. Rational Behavior

11

Note, complete (or incomplete) information is not an assumption about the

rules of the game, it is an assumption on players interactive knowledge
about the rules and preferences. For example, in the seller-buyer minigame it can be safely assumed that there is common knowledge of the
outcome function g (who gets the object and monetary transfers), but the
valuation VS , VB need not be commonly known. For some types of objects
Vi is known only to i; for other types of objects, the seller may know the
quality of the object better then the buyer, so VB could be given by an
idiosyncratic component known to B plus a quality component known to
S.
Although situations with asymmetric information about chance moves,
such as Poker, are conceptually different from situations with incomplete
information, we will see that there is a formal similarity which allows
(at least to some extent) to use the same analytical tools to study both
situations.
Examples: In parlor games like chess and poker there is presumably
complete information; indeed, the rules of the game are common
knowledge, and it may be taken for granted that it is common knowledge
that players like to win. On the other hand, auctions typically feature
incomplete information because the competitors valuations of the objects
on sale are not commonly known.

1.5

Rational Behavior

The decision problem of a single individual i can be represented as follows:

i can take an action in a set of feasible alternatives, or decisions, Di ; is
welfare (payoff, utility) is determined by his decision di and an external
state i i (which represents a vector of variables beyond is control,
e.g., the outcome of a random event and/or the decisions of other players)
according to the outcome function
g : Di i C
and the utility function
vi : C R.
We assume that is choice cannot affect i and that i does not have any
information about i when he chooses beyond the fact that i i .

12

1. Introduction

The choice is made once and for all. (We will show in the part on dynamic
games that this model is much more general than it seems.) Assume,
just for simplicity, that Di and i are both finite and let, for notational
1 , 2 , ..., n }. In order to make a rational
convenience, i = {i
i
i
choice, i has to assess the probability of the different states in i . Suppose
that his beliefs about i are represented by a probability distribution
i (i ) where
(
)
n
X
k
(i ) = i Rn+ :
i (i
)=1 .
k=1

Then is rational decision (given i ) is to take any alternative that

maximizes is expected payoff, i.e. any di such that
X
i (i )vi (g(di , i )) .
di arg max
di Di

i i

Decision di is also called a best response, or best reply, to i . The

probability distribution could be exogenously given (roulette lottery) or
simply represent is subjective beliefs about i (horse lottery, game
against an opponent). The composite function vi g is denoted ui , that is,
ui (di , i ) = vi (g(di , i )); ui is called the payoff function, because in
many games of interests the rules of the decision problem (or game) attach
monetary payoffs to the possible outcomes of the game and it is taken
for granted that the decision maker maximizes his expected monetary
payoff. In the finite case, ui is often represented as a matrix with k`
k
`
entry uk`
i = ui (di , i ):

d1i
d2i
..
.

1
i

2
i

...

u11
i

u12
i

...

u21
i
..
.

u22
i
..
.

...
..
.

Figure 1.3: Matrix representation of the payoff function

This representation of a decision makers choice, of course, fits very

well the case of static games. In this case Di = Ai is simply the set of

1.6. Assumptions and Solution Concepts in Game Theory

13

feasible actions and i may be the set Ai of the opponents feasible

action profiles (or a larger space if there is incomplete information).
However, the representation is more general: Consider a dynamic game.
Its extensive form E specifies all the circumstances under which player i
may have to choose an action and the information i would have in those
circumstances. Then player i can decide how to play this game according to
a plan of action, or strategy, specifying how he would behave in any such
circumstance given the available information. His uncertainty concerns the
strategies of others. Every possible profile of strategies induces a particular
outcome, or consequence. Therefore, the decision problem faced by a
rational agent is the problem of selecting a strategy under uncertainty
about the strategies used by the opponents (and possibly about other
unknowns). In the part of these lecture notes devoted to dynamic games
we will discuss some subtleties related to the concepts of strategy and
plan.

1.6

Assumptions and Solution Concepts in Game

Theory

Game theory provides a mathematical language to formulate assumptions

about the rules of the game and players preferences, that is, all the
elements listed in section 1.3. But such assumptions are not enough
to derive conclusions about how the players behave in a given game.
The central behavioral assumption in game theory is that players are
rational (see Section 1.5). However, without further assumptions about
what players believe about the variables which affect their payoffs, but
are beyond their control (in particular their opponents behavior), the
rationality assumption has little behavioral content: we can only say that
each players choice is undominated, i.e. is a best response to some belief.
In most games, this is too little to derive interesting results.5
How can we obtain interesting results from our assumptions about
the rules of the game? The standard approach used in game theory is
analogous to the one used in the undergraduate economics analysis of
competitive markets: we formulate assumptions about preferences and
5
There are important exceptions. In many situations where each individual in a group
decides how much to contribute to a public good, it is strictly dominant to contribute
nothing (see example 2 in chapter 2).

14

1. Introduction

technology and then we assume that economic agents plans are mutually
consistent best responses to equilibrium prices.
As explained in section 1.2, there is an important difference: the
textbook analysis of competitive markets does not specify a priceformation mechanism and uses equilibrium market-clearing as a theoretical
shortcut to overcome this problem. On the contrary, in a gametheoretic model all the observable variables we try to explain depend on
players actions (and exogenous shocks) according to an explicitly specified
function, as in auction models.
Yet, there are also similarities with the analysis of competitive markets.
One could say that the role of prices is played (sic) by players beliefs since
we assume that they are, in some sense, mutually consistent. The precise
meaning of the statement beliefs are mutually consistent is captured by
a solution concept. The simplest and most widely used solution concept
in game theory, Nash equilibrium, assumes that players beliefs about each
other strategies are correct and each player best responds to his beliefs; as
a result, each player uses a strategy that is a best response to the strategies
used by other players.
Nash equilibrium is not the only solution concept used in game theory.
Recent developments made clear that solution concepts implicitly capture
expressible6 assumptions about players rationality and beliefs, and some
assumptions that are appropriate in some contexts are too weak, or simply
inappropriate, in different contexts. Therefore, it is very important to
provide convincing motivations for the solution concept used in a specific
application.
Let us somewhat vaguely define as strategic thinking what
intelligent agents do when they are fully aware of participating in an
interactive situation and form conjectures by putting themselves in the
shoes of other intelligent agents. As the title of this book suggest,
we mostly (though not exclusively) present game theory as an analysis
of strategic thinking. We will often follow the traditional textbook
game theory and present different solution concepts providing informal
motivations for them, sometimes with a lot of hand-waving. However,
you should always be aware that a more fundamental approach is being
6
We can give a mathematical meaning to the label expressible, but, at the moment,
you should just understand something that can be expressed in a clear and precise
language.

1.6. Assumptions and Solution Concepts in Game Theory

15

developed, where game theorists formulate explicit assumptions not only

about the rules of the game and preferences, but also about players
rationality and initial beliefs, and also about how beliefs change as the play
unfolds. Implications about choices and observables can be derived directly
from such assumptions, without the mediation of solution concepts. Then,
solution concepts become mere shortcuts to characterize the behavioral
implications of assumptions about rationality and beliefs.
For example, a standard solution concept in game theory, subgame
perfect equilibrium, requires players strategies to form an equilibrium not
only for the game itself, but also in every subgame.7 In finite games
with (complete and) perfect information, if there are no relevant ties,
there is a unique subgame perfect equilibrium that can be computed
with a backward induction algorithm. For two-stage games with perfect
information, subgame perfect equilibrium can be derived from simple
assumptions about rationality and beliefs: players are rational and the first
mover believes that the second mover is rational. To illustrate, consider
the seller-buyer mini-game and assume players have complete information.
Specifically, suppose that it is common knowledge that VS < 1, VB > 2.
The assumption of rationality of the buyer, denoted with RB , implies that
he accepts every price below his valuation; since VB > 2, the only strategy
consistent with RB is a1 a2 . Rationality of the seller, denoted with RS ,
implies that he attaches subjective probabilities to the four strategies of
the buyer and chooses the ask price that maximizes his expected utility;
if the seller has deterministic beliefs, he asks for the highest prices he
expects to be accepted, e.g., he asks p = 1 if he is certain of strategy a1 r2 .
But if the seller is certain that the buyer is rational, given the complete
information assumption, he assigns probability 1 to strategy a1 a2 . Then
the rationality of the seller and the belief of the seller in the rationality of
the buyer imply that the seller asks for the highest price, i.e. he uses the
bargaining power derived from the knowledge that VB > 2 and the firstmover advantage. In symbols, write Bi (E) for i believes (with probability
one) E, where E is some event; then RS BS (RB ) implies that S asks
for p = 2; RB implies that p = 2 is accepted; RS BS (RB ) RB implies
sS = 2 and sB = a1 a2 , which is the subgame perfect equilibrium obtained
7
A precise definition of a subgame will be given in Part 2. For the time being, you
should think of a subgame as a smaller game that starts with a non-terminal node of
the original game.

16

1. Introduction

via backward induction.

Two-stage games with complete and perfect information are so simple
that the analysis above may seem trivial. On the other hand, the rigorous
analysis of more complicated dynamic games based on higher levels of
mutual belief in rationality presents conceptual difficulties that will be
addressed only in the second part of these lecture notes. Here we further
illustrate strategic thinking with a semi-formal analysis of a static game
representing a perfectly competitive market. We hope this will also have
the side effect to make the reader understand that game theory is useful to
analyze every complete model of economic interaction, including models
of perfect competition.

1.6.1

Game Theoretic Analysis of a Perfectly Competitive

Market

Let us analyze the decisions of a large number n of small, identical

agricultural firms producing a crop (say, corn). For reasons that will be
clear momentarily, it is convenient to index firms as equally spaced rational
numbers in the interval (0, 1], that is, I = { n1 , n2 , ..., n1
n , 1} (0, 1] with
generic element i I. Each firm i decides in advance how much to produce.
This quantity is denoted q(i). The output of corn q(i) will be ready to be
sold only in six months. Markets are incomplete: it is impossible to agree
in advance on a price for corn to be delivered in six months. So, each firm
i has to guess the price p for corn in six months. It is common knowledge
that the demand for corn (not explicitly micro-founded in this example)
is given by function D(p) = n max{( p), 0}, that is, D(p) = n( p)
if p and D(p) = 0 otherwise. It is also common knowledge that
each firm will fetch the highest
P price at which the market can absorb the
total output for sale, Q = ni=1 q(i).
  Thus, each firm
 will sell each unit
of output at the uniform price P
Q
n

Q
n

Q
n

if

Q
n

and p = 0

if
> (for notational convenience, we express inverse demand as a
function of the average output Q
n ). Each firm i has total cost function
q
1 2
C(q) = 2m
q , and marginal cost function M C(q) = m
. Each firm has
obvious preferences, it wants to maximize the difference between revenues
and costs. The technology (cost function) and preferences of each firm are

1.6. Assumptions and Solution Concepts in Game Theory

17

common knowledge.8
Now, we want to represent mathematically the assumption that each
firm is negligible with respect to the (maximum) size of the market .
Instead of assuming a large, but finite set of firms given by a fine grid I in
the interval (0, 1], we use a quite standard mathematical idealization and
assume that there is a continuum of firms,Rnormalized to the unit
interval
1
1 Pn
I = (0, 1]. Thus, the average output is q = 0 q(i)di instead of n i=1 q(i).
Each firm understands that its decision cannot affect the total output and
the market price.9
Given that all the above is common knowledge, we want to derive
the price implications of the following assumptions about rationality and
beliefs:
R (rationality): each firm has a conjecture about (q and) p and chooses
a best response to such conjecture, for brevity, each firm is rational;10
B(R) (mutual belief in rationality): all firms believe (with certainty)
that all firms are rational;
B2 (R) = B(B(R)) (mutual belief of order 2 in rationality): all firms
believe that all firms believe that all firms are rational;
...
Bk (R) = B(Bk1 (R)) (mutual belief of order k in rationality): [all firms
believe that]k all firms are rational
...
Note, we are using symbols for these assumptions. In a more formal
mathematical analysis, these symbols would correspond to events in a
space of states of the world. Here they are just useful abbreviations.
Conjunctions of assumptions will be denoted by the symbol , as in
R B(R). It is a good thing that you get used to these symbols, but
if you find them baffling, just ignore them and focus on the words.
What are the consequences of rationality (R) in this market? Let p(i)
8
The analysis can be generalized assuming heterogeneous firms, where each firm i is
characterized by the marginal cost parameter m(i).
P Then is is enough to assume that
each i knows m(i) and that the average m = n1 n
i=1 m(i) is common knowledge.
9
This example is borrowed from Guesnerie (1992).
10
The symbols R, B(R) etc. below should be read as abbreviations of sentences. They
can also be formally defined using mathematics, but this goes beyond the scope of these
lecture notes.

18

1. Introduction

denote the price expected by firm i.11 Then i solves



1 2
max p(i)q
q
q0
2m
which yields
q(i) = m p(i).
Since firms know the inverse demand function P (q) = max{( q)/, 0},
each firm is expectation of the price is below the upper bound p0 :=
(:= means equal by definition ): p(i) . It follows that
Z
q=

Z
mp(i)di = m

p(i)di q 1 := m ,

where q 1 denotes the upper bound on average output when firms are
rational.
By mutual belief in rationality, each firm understands that q q 1 and
p max{(m )/, 0}. If m , this is not very helpful: assuming belief
in rationality does not reduce the span of possible prices and does not yield
additional implications for the endogenous variables we are interested in. It
is not hard to understand that rationality and mutual belief in rationality
m
of every order yields the same result q m
, where because
m .
So, lets assume from now on that m < . Note, this is the so called
cobweb stability condition. Given belief in rationality, B(R), each firm
expects a price at least as high as the lower bound




m
1
1
1
p :=
q =
1
> 0.

Therefore R B(R) implies that q = m

output is above the lower bound
m
q := mp =

11

R1
0

p(i)di mp1 , that is, average

m
1

This is, in general, the mathematical expectation of p according to the subjective

probabilistic conjecture of firm i.

1.6. Assumptions and Solution Concepts in Game Theory

19

and hence the price is below the upper bound

1

p := ( q 2 ) =





2 
m
m
X
m k
1
1
=

k=0

By B(R) and B2 (R) (mutual belief of order 2 in rationality), each firm

predicts that p p2 . Rational firms with such expectations choose an
output below the upper bound
m
q := mp =





2 
m
m
mX
m k
1
1
=

k=0

Thus, R B(R) B2 (R) implies that q q 3 and the price must be above
the lower bound

3 
X
m k
1
3
3
p := ( q ) =

k=0

Can you guess the consequences for price and average output of
assuming rationality and common belief in rationality (mutual belief in
rationality of every order)? Draw the classical Marshallian cross of demandsupply, trace the upper and lower bounds found above and go on. You will
find the answer.
More formally, define the following sequences of upper bounds and
lower bounds on the price:

L 
m k
X

, L even

k=0

L 
X
m k

, L odd
: =

pL : =
pL

k=0

It can be shown by mathematical induction that R

L1
T

B k (R) (rationality

k=1

and mutual belief in rationality of order k = 1, . . . , L 1, with L 2)

implies p pL if L is even, and p pL if L is odd. Since m
< 1,


k
P
m
the sequence p2` = 2`
is decreasing in `, the sequence
k=0

20

1. Introduction


k
P
m
is increasing in `, and they both converge to
p2`+1 = 2`+1
k=0

the competitive equilibrium price p = +m

, see Figure 1.4.
p
M C(q)
p0
p2
p

p1
P (q)

q2

q3

q1

Figure 1.4: Strategic thinking in a perfectly competitive market.

A comment on rational versus rationalizable expectations.

This price p is often called the rational expectations price, because
it is the price that the competitive firms would expect if they performed
the same equilibrium analysis as the modeler. We refrain from using
such terminology. Game theory allows a much more rigorous and precise
terminology, the one that we have been using above. In these lecture notes,
rationality is a joint property of choices and beliefs and it means only
that agents best respond to their beliefs. The phrase rational beliefs, or
rational expectations was coined at a time when theoretical economists
did not have the formal tools to be precise and rigorous in the analysis of
rationality and interactive beliefs. Now that these tools exist, such phrases
as rational expectations should be avoided, at least in game theoretic
analysis. On the other hand, expectations may be consistent with common

1.6. Assumptions and Solution Concepts in Game Theory

21

belief in rationality (e.g., under common knowledge of the game), and

it makes sense to call such expectations rationalizable. Note well,
the fact that there is only one rationalizable expectations price, p , is a
conclusion of the analysis, it has not been assumed! Also, this conclusion
holds only under the cobweb stability condition m < . This contrasts
with standard equilibrium analysis, whereby it is assumed at the outset
that expectations are correct, whether or not this follows from assumptions
like common belief in rationality.
What do we learn from this example? First, it shows that certain
assumptions about rationality and beliefs (plus common knowledge of the
game) may yield interesting implications or not according to the given
model and parameter values. Here, rationality and common belief in
rationality have sharp implications for the market price and average output
if the cobweb stability condition holds. Otherwise they only yield a very
weak upper bound on average output.
Second, recall that strategic thinking is what intelligent agents do
when they are fully aware of participating in an interactive situation and
form conjectures by putting themselves in the shoes of other intelligent
agents; in this sense, the above is as good an analysis of strategic thinking
as any, and yet it is the analysis of a perfectly competitive market!
Third, the example shows how solution concepts can be interpreted as
shortcuts. Here we presented an iterated deletion of prices and conjectures,
where one deletes prices that cannot occur when firms best respond to
conjectures and delete conjectures that do not rule out deleted prices.
Note, mutual beliefs of order k do not appear explicitly in this iterative
deletion procedure. The procedure is a solution concept of a kind and it can
be used as a shortcut to obtain the implications of rationality and common
belief in rationality. Under cobweb stability, the rational expectation
equilibrium price p results. This shows that under appropriate conditions
we can use the equilibrium concept to draw the implications of rationality
and common belief in rationality. This theme will be played time and
again.

Part I

Static Games

22

Static Games: Formal

Representation and
Interpretation
A static game is an interactive situation where all players move
simultaneously and only once.1 The key features of a static game are
formally represented by a mathematical structure (a list of sets and
functions)
G = hI, C, (Ai )iI , g, (vi )iI i
where:
I = {1, 2, ..., n} is the set of individuals or players, typically denoted
by i, j;
Ai is the set of possible actions for player i; ai , ai , a0i , a
i are
alternative action labels we frequently use;
1

Sometimes static games are also called normal form games or strategic form
games. As we mentioned in the Introduction, this terminology is somewhat misleading.
The normal, or strategic form of a game has the same structure of a static game, but
the game itself may have a sequential structure. The normal form of shows the
payoffs induced by any combination of plans of actions of the players. Some game
theorists, including the founding fathers von Neumann and Morgenstern, argue that
from a theoretical point of view all strategically relevant aspects of a game are contained
in its normal form. Anyway, here by static game we specifically mean a game where
players move simultaneously.

23

24

2. Static Games: Formal Representation and Interpretation

g : A1 A2 ... An C is the outcome (or consequence)
function which captures the essence of the rules of the game, beyond
the assumption of simultaneous moves;
vi : C R is the Von Neumann-Morgenstern utility function of
player i.

The structure hI, C, (Ai )iI , gi, that is, the game without the utility
functions, is called game form. The game form represents the essential
features of the rules of the game. A game is obtained by adding to the game
form a profile of utility functions (vi )iI = (v1 , ..., vn ), which represent
players preferences over lotteries of consequences, according to expected
utility calculations.
From the consequence function g and the utility function vi of player
i, we obtain a function that assigns to each action profile a = (aj )jI the
utility vi (g(a)) for player i of consequence g(a). This function
ui = vi g : A1 A2 ... An R
is called the payoff function of player i. The reason why ui is called
the payoff function is that the early work on game theory assumed that
consequences are distributions of monetary payments, or payoffs, and that
players are risk neutral, so that it is sufficient to specify, for each player i,
the monetary payoff implied by each action profile. But in modern game
theory ui (a) = vi (g(a)) is interpreted as the von Neumann-Morgenstern
utility of outcome g(a) for player i. If there are monetary consequences,
then
g = (g1 , ..., gn ) : A Rn ,
where mi = gi (a) is the net gain of player i when a is played. Assuming
that player i a selfish, then function vi is strictly increasing in mi and
constant in each mj with j 6= i (note that selfishness is not an assumption
of game theory, it is an economic assumption that may be adopted in game
theoretic models). Then function vi captures the risk attitudes of player
i. For example, i is strictly risk averse if and only if vi is strictly concave.
Games are typically represented in the reduced form
G = hI, (Ai , ui )iI i

25
which shows only the payoff functions. We often do the same. However, it
is conceptually important to keep in mind that payoff functions are derived
from a consequence function and utility functions.
For simplicity, we often assume that all the sets Ai (i I) are finite;
when the sets of actions are infinite, we explicitly say so, otherwise they
are tacitly assumed to be finite. This simplifies the analysis of probabilistic
beliefs. A static game with finite action sets is called a finite static game
(or just a finite game, if static is clear from the context). When the
finiteness assumption is important to obtain some results, we will explicitly
mention that the game is finite in the formal statement of the result.
We call profile a list of objects (xi )iI = (x1 , ..., xn ), one object
xi from some set Xi for each player i I. In particular, an action
profile is a list a = (a1 , ..., an ) A := iI Ai .2 The payoff function
ui numerically represents the preferences of player i among the different
action profiles a, a0 , a00 ... A. The strategic interdependence is due to the
fact that the outcome function g depends on the entire action profile a and,
consequently, the utility that a generic individual i can achieve depends
not only on his choice, but also on those of other players.3 To stress how
the payoff of i depends on a variable under is control as well as on a vector
of variables controlled by other individuals, we denote by i = I\{i} the
set of individuals different from i, we define
Ai := A1 ... Ai1 Ai+1 ... An
and we write the payoff of i as a function of the two arguments ai Ai
and ai Ai , i.e. ui : Ai Ai R.
In order to be able to reach some conclusions regarding players
behavior in a game G, we impose two minimal assumptions (further
2

Recall that := means equal by definition .

Note, in our formulation, players utility depends only on the consequence
determined by action profile a through the outcome function g. Thus, we are assuming a
form of consequentialism. If you think that consequentialism is an obvious assumption,
think twice. A substantial experimental evidence suggests agents also care about other
players intentions: agent i may evaluate the same action profile differently depending on
his belief about opponents intentions. To provide a formal analysis of players intentions,
we need to specify what an agent thinks other agents will do, what an agent thinks other
agents think other agents will do, what an agent thinks other agents think other agent
think that other agents will do and so on. Although the analysis of these issues is beyond
the scope of these notes, the interested reader is referred to the growing literature on
psychological games.
3

26

2. Static Games: Formal Representation and Interpretation

assumptions will be introduced later on as necessary):

[1]
Each player i knows Ai , Ai and his own payoff function
ui : Ai Ai R.
[2]
Each player is rational (see next chapter for a more formal
definition of rationality).
It could be argued that assumption [1] (i knows ui ) is tautological.
Indeed, it seems almost tautological to assume that every individual knows
his preferences over A. However, recall that ui = vi g; thus, although it
may be reasonable to take for granted that i knows vi , i.e. his preferences
over the set of consequences C, i might not know the outcome function g.
The following example clarifies this point.
Example 1. (Knowledge of the utility and payoff functions)
Players 1 and 2 work in a team for the production of a public good. Their
action is the effort ai [0, 1]. The output, y, depends on efforts according
to a Cobb-Douglas production function
y = K(a1 )1 (a2 )2 .
The cost of effort measured in terms of public good is ci (ai ) = (ai )2 . The
outcome function g : [0, 1] [0, 1] R3+ is then
g(a1 , a2 ) = (K(a1 )1 (a2 )2 , (a1 )2 , (a2 )2 ).
The utility function of player i is vi (y, c1 , c2 ) = y ci . The payoff function
is then
ui (a1 , a2 ) = vi (g(a1 , a2 )) = K(a1 )1 (a2 )2 (ai )2 .
If i does not know all the parameters K, 1 and 2 (as well as the functional
forms above), then he does not know the payoff function ui , even if he
knows the utility function vi .
Example 1 also clarifies the difference between game in the technical
sense of game theory, and the rules of the game as understood by the
layperson (see the Introduction). In the example above, the rules of
the game specify the action set [0, 1] and the outcome function, that
is functions y = K(a1 )1 (a2 )2 and ci (ai ) = (ai )2 (i = 1, 2). In order to
complete the description of the game in the technical sense, we also have
to specify the utility functions vi (y, c1 , c2 ) = y ci (i = 1, 2).4 Also, the
4

This specification assumes that players are risk-neutral with respect to their
consumption of public good.

27
example clarifies that the outcome map a 7 g(a) may depend on personal
features of the agents playing the game: in this case, the productivity i of
each agent affects how the pair of efforts (a1 , a2 ) (say, the hours worked)
is mapped to the output. Therefore, the outcome function g is not always
fully determined by the rules of the game.
Unlike the games people play for fun (and those that experimental
subjects play in most game theoretic experiments), in many economic
games the rules of the game may not be fully known. In the above example,
for instance, it is possible that player i knows his own productivity
parameter i , but does not know K or i . Thus, assumption [1] is
substantive: i may not know the consequence function g and hence he
may not know ui = vi g.
The complete information assumption that we will consider later on
(e.g. in chapter 4) is much stronger; recall from the Introduction that there
is complete information if the rules of the game and players preferences
over (lotteries over) consequences are common knowledge. Although, as we
explained above, the rules of the game may be only partially known, there
are still many interactive situations where it is reasonable to assume that
they are not only known, but indeed commonly known. Yet, assuming
common knowledge of players preferences is often far-fetched. Thus,
complete information should be thought as an idealization that simplifies
the analysis of strategic thinking. Chapter 7 will introduce the formal tools
necessary to model the absence of complete information.

Rationality and Dominance

In this chapter we analyze static games from the perspective of decision
theory. We take as given the conjecture1 of a player about opponents
behavior. This conjecture is, in general, probabilistic, i.e. it can be
expressed by assigning subjective probabilities to the different action
profiles of the other players. A player is said to be rational if he maximizes
his expected payoff given his conjecture. The concept of dominance allows
to characterize which actions are consistent with the rationality assumption
without knowing a players conjecture.

3.1

Probabilities and Expected Payoff

The payoff function ui , which is a derived utility function, implicitly

represents the preferences of i over different probability measures, or
lotteries, over A. The preferred lotteries are those that yield a higher
expected payoff. We explain this in details below.
The set of probability measures over a generic domain X is denoted by

1
In general, we use the word conjecture to refer to a players belief about variables
that affect his payoff and are beyond his control, such as the actions of other players in
a static game.

28

3.1. Probabilities and Expected Payoff

29

(X). If X is finite2
(
(X) :=

RX
+ :

(x) = 1 .

xX

The definition above is the simplest one for finite domains. But there is
an alternative, equivalent, definition, that can be more easily generalized
to infinite domains. Since sometimes we will consider probability measures
on infinite domains, we present here also the alternative. First consider the
definition above for a finite X, and fix (X). For each event E X,
determines the probability of E as follows:
X
(E) =
(x).
xE

With this, the map E 7 (E) satisfies the following properties:

(normalization) (X) = 1,
(additivity) for all E, F 2X , E F = (E F ) = (E) + (F ).
Note that additivity implies that () = 0, because E = , hence
(E) = (E ) = (E) + (), which implies () = 0. These are the
characterizing properties of a probability measure, defined as a function
: 2X [0, 1]. Therefore, the alternative definition of (X) is
n
o
X
(X) := [0, 1]2 : (X) = 1, E, F 2X , E F = (E F ) = (E) + (F ) .
When X is infinite, we typically restrict our attention to a subcollection X 2X of subsets of X, called events (e.g., the Borel sets,
when X is a subset of a Euclidean space, or more generally a metric
space), and we strengthen the additivity property as follows:
N is a sequence of pairwise
(countable additivity) if (Ek )
k=1 X

S
Ek X , then
disjoint events and
k=1

[
k=1

!
Ek

(Ek ).

k=1

2
Recall that RX
+ is the set of non negative real valued functions defined over the
n
domain X. If X is the finite set, X = {x1 , ..., xn }, RX
+ is isomorphic to R+ , the positive
n
orthant of the Euclidean space R .

30

3. Rationality and Dominance

Whenever we consider infinite sets, we take for granted that there

is a well-defined collection of events X (such as the Borel subsets of a
metric space) and (X) is the set of functions : X [0, 1] that satisfy
normalization and countable additivity.
Note, X can be regarded as a subset of (X): a point x X
corresponds to the degenerate probability measure x , called Dirac
measure (or deterministic, or dogmatic belief), that satisfies x (E) = 1
if and only if x E. Hence, with a slight abuse of notation, we can write
X (X) (more abuses of notation will follow).
Definition 1. Consider a probability measure (X), where X is a
finite set. The subset of the elements x X to which assigns a positive
probability is called support of and is denoted by3
Supp := {x X : (x) > 0} .
Suppose that there is a function f : X C, where C is a set of
consequences, and an individual has a preference relation % over the
set of lotteries (C) represented by a Von Neumann-Morgenstern utility
function v : C R, that is, for all , 0 (C),
% 0

(c)v(c)

cC

0 (c)v(c).

cC

Then we can derive a corresponding preference relation on X as follows.

First we define the pushforward function f : (X) (C) that
determines the lottery induced, through f , by a probability measure on
X:
X


f()(c) = f 1 (c) =
(x)
xf 1 (c)

(of course, the second equality holds when X is finite). With this, we
obtain a preference relation on (X): for all , 0 (X),
% 0

X
cC

f()(c)v(c)

f(0 )(c)v(c).

cC

3
If X is a closed and measurable subset of Rm , we define the support of a probability
measure (X) as the intersection of all closed sets C such that (C) = 1. If X is
finite, this definition is equivalent to the previous one.

3.2. Conjectures

31

Consider the particular case where X = A is the set of action profiles

in a game, let f = g : A C be the consequence function, and fix any
player i. Then, for all , 0 (A),
X
X
%i 0
g()(c)vi (c)
g(0 )(c)vi (c).
cC

cC

Since g()(c) = g 1 (c) ) and we have defined the payoff function
ui := vi g, for each (A), we have
X
X X
X
g()(c)vi (c) =
(a)vi (g(a)) =
(a)ui (a)
cC

cC ag 1 (c)

aA

(to understand the second equality, note that {g 1 (c)}cC is a partition of

A, and recall that ui (a) = vi (g(a))). Therefore
X
X
%i 0
(a)ui (a)
0 (a)ui (a).
aA

aA

In other words, probability measures on A are ranked according to the

corresponding expected payoffs.

3.2

Conjectures

The reason why one needs to introduce preferences over lotteries and
expected payoffs is that individual i cannot observe other individuals
actions (ai ) before making his own choice. Hence, he needs to form
a conjecture about such actions. If i were certain of the other players
choices, then one could represent is conjecture simply with an action
profile ai Ai . However, in general, i might be uncertain about other
players actions and assign a strictly positive (subjective) probability to
several profiles ai , a0i , etc.
Definition 2. A conjecture of player i is a (subjective) probability
measure i (Ai ). A deterministic conjecture is a probability
measure i (Ai ) that assigns probability one to a particular action
profile ai Ai .
Note, we call conjecture a (probabilistic) belief about the behavior
of other players, while we use the term belief to refer to a more general
type of uncertainty.

32

3. Rationality and Dominance

Remark 1. The set of deterministic conjectures of player i coincides with

the set of other players action profiles, Ai (Ai ).
One of the most interesting aspect of game theory consists in
determining players conjectures, or, at least, in narrowing down the set of
possible conjectures, combining some general assumptions about players
rationality and beliefs with specific assumptions about the given game G.
However, in this chapter we will not try to explain players conjectures.
This is left for the following chapters.
For any given conjecture i , the choice of a particular action a
i
corresponds to the choice of the lottery on A that assigns probability
i (ai ) to each action profile (
ai , ai ) (ai Ai ) and probability zero to
the profiles (ai , ai ) with ai 6= a
i . Therefore, if a player i has conjecture i
and chooses (or plans to choose) action a
i , the corresponding (subjective)
expected payoff is
X
ui (
ai , i ) :=
i (ai )ui (
ai , ai ).
ai Ai

There are many different ways to represent graphically actions /

lotteries, conjectures, and preferences when the number of available actions
is small. Let us focus our attention on the instructive case where player is
opponent (i) has only two actions, denoted l (left ) and r (right ).
Consider, for instance, the function ui represented by the following payoff
matrix for player i:
i\ i

Figure 3.1: Matrix 1

Given that i has only two actions, we can represent the conjecture

3.2. Conjectures

33

i of i about i with a single number: i (r), the subjective probability

that i assigns to r. Each action ai corresponds to a function that assigns
to each value of i (r), the expected payoff of ai . Since the expected payoff
function (1 i (r))ui (ai , l) + i (r)ui (ai , r) is linear in the probabilities,
every action is represented by a line, such as aa, bb, cc etc. in Figure 3.2.

4 a

1 f

i (r)

i (`)

Figure 3.2: Expected payoff as a function of beliefs.

From Figure 3.2 it is apparent that only actions a, b and e are

justifiable by some conjecture. In particular, if i (r) = 0, then a and e
are both optimal; if 0 < i (r) < 21 , then a is the only optimal action (the
line aa yields the maximum expected payoff). If i (r) > 12 , then b is the
only optimal action (the line bb yields the maximum expected payoff). If
i (r) = 21 , then there are two optimal actions: a and b.

3.2.1

Mixed actions

In principle, instead of choosing directly a certain action, a player could

delegate his decision to a randomizing device, such as spinning a roulette

34

3. Rationality and Dominance

wheel, or tossing a coin. In other words, a player could simply choose the
probability of playing any given action.
Definition 3. A random choice by player i, also called a mixed action,
is a probability measure i (Ai ). An action ai Ai is also called pure
action.
Remark 2. The set of pure actions can be regarded as a subset of the set
of mixed actions, i.e. Ai (Ai ).
It is assumed that (according is beliefs) the random draw of an action
of i is stochastically independent of the other players actions. For example,
the following situation is excluded : i chooses his action according to the
(random) weather and he thinks his opponents are doing the same, so that
there is correlation between ai and ai even though there is no causal link
between ai and ai (this type of correlation will be discussed in section
5.2.2 of Chapter 5). More importantly, player i knows that moves are
simultaneous and therefore by changing his actions he cannot cause any
change in the probability distribution of opponents actions. Hence, if
player i has conjecture i and chooses the mixed action i , the subjective
probability of each action profile (ai , ai ) is i (ai )i (ai ) and is expected
payoff is
X
X X
i (ai )i (ai )ui (ai , ai ) =
i (ai )ui (ai , i ).
ui (i , i ) :=
ai Ai ai Ai

ai Ai

If the opponent has only two feasible actions, it is possible to use a graph to
represent the lotteries corresponding to pure and mixed actions. For each
action ai , we consider a corresponding point in the Cartesian plane with
coordinates given by the utilities that i obtains for each of the two actions
of the opponent. If the actions of the opponents are l and r, we denote such
coordinates x = ui (, l) and y = ui (, r). Any pure action ai corresponds
to the vector (x, y) = (ui (ai , l), ui (ai , r)) (a row in the payoff matrix of
i). The same holds for the mixed actions: i corresponds to the vector
(x, y) = (ui (i , l), ui (i , r)). The set of points (vectors) corresponding to
the mixed actions is simply the convex hull of the points corresponding to
the pure actions.4 Figure 3.3 represents such a set for matrix 1.
4

The convex hull of a set of points X Rk is the smallest convex set containing
X, that is, the intersection of all the convex sets containing X.

3.2. Conjectures

35

How can conjectures be represented in such a figure? It is quite

simple. Every conjecture induces a preference relation on the space of
payoff pairs. Such preferences can be represented through a map of
iso-expected payoff curves (or indifference curves). Let (x, y) be the
generic vector of (expected) payoffs corresponding to l and r respectively.
The expected payoff induced by the conjecture i is i (l)x + i (r)y.
Therefore, the indifference curves are straight lines defined by the equation
i (l)
y = iu(r) i (r)
x, where u denotes the constant expected payoff. Every
conjecture i corresponds to a set of parallel lines with negative slope
(or, in the extreme cases, horizontal or vertical slope) determined by the
orthogonal vector (i (l), i (r)). The direction of increasing expected payoff
is given precisely by such orthogonal vector. The optimal actions (pure or
mixed) can be graphically determined considering, for any conjecture i ,
the highest line of iso-expected payoff among those that touch the set of
feasible payoff vectors (the shaded area in Figure 3.3).

(`) = 1
3

ui (, r)
b

(r) = 2
3

(r) = 2
3

} (`) =
1

1
3

ui (, `)

Figure 3.3: State contingent representation of expected payoff.

Allowing for mixed actions, the set of feasible (expected) payoffs vectors

36

3. Rationality and Dominance

is a convex polyhedron (as in Figure 2). To see this, note that if i and i
are mixed actions, then also p i + (1 p) i is a mixed action, where
p i + (1 p) i is the function that assigns to each pure action ai the
weight pi (ai )+(1p)i (ai ). Thus, all the convex combinations of feasible
payoff vectors are feasible payoff vectors, once we allow for randomization.
However, the idea that individuals use coins or roulette wheels to
randomize their choices may seem weird and unrealistic. Furthermore,
as it can be gathered from Figure 3.3, for any conjecture i and any mixed
action i , there is always a pure action ai that yields the same or a higher
expected payoff than i (check all possible slopes of the iso-expected payoff
curves and verify how the set of optimal points looks like).5 Hence, a player
cannot be strictly better off by choosing a mixed action rather than a pure
action.
The point of view we adopt in these lecture notes is that players never
randomize (although their choices might depend on extrinsic, payoffirrelevant signals, as in Chapter 5, Section 5.2.2). Nonetheless, it will
be shown that in order to evaluate the rationality of a given pure
action it is analytically convenient to introduce mixed actions. (In
Chapter 5, we discuss interpretations of mixed actions that do not involve
randomization.)

3.3

Best Replies and Undominated Actions

Definition 4. A (mixed) action i is a best reply to conjecture i if

i (Ai ), ui (i , i ) ui (i , i ),
that is
i arg max ui (i , i ).
i (Ai )

The set of pure actions that are best replies to conjecture i is denoted by6
ri (i ) := Ai arg max ui (i , i ).
i (Ai )

5
6

A formal proof of this result will be provided in the next section.

Recall: Ai is a subset of (Ai ).

3.3. Best Replies and Undominated Actions

37

The correspondence ri : (Ai ) Ai is called best reply

correspondence.7 An action ai is called justifiable if there exists a
conjecture i (Ai ) such that ai ri (i ).
Note that, even if Ai is finite, one cannot conclude without proof
that the set of pure best replies to the conjecture i is non-empty, i.e.
that ri (i ) 6= . In principle, it could happen that in order to maximize
expected payoff it is necessary to use mixed actions that assign positive
probability to more than one pure action. However, we will show that
ri (i ) 6= for every i provided that Ai is finite (or, more generally, that
Ai is compact and ui continuous in ai , two properties that trivially hold if
Ai is finite); thus it is never necessary to use non-pure actions to maximize
expected payoff.
We say that a player is rational if he chooses a best reply to the
conjecture he holds. The following result shows, as anticipated, that a
rational player does not need to use mixed actions. Therefore, it can be
assumed without loss of generality that his choice is restricted to the set
of pure actions.
Lemma 1. Fix a conjecture i (Ai ). A mixed action i is a best
reply to i if and only if every pure action in the support of i is a best
reply to i , that is
i arg max ui (i , i ) Suppi ri (i ).
i (Ai )

Proof. This is a rather trivial result. But since this is the first proof
of these notes we will go over it rather slowly.
(Only if ) We first show that if Suppi is not included in ri (i ),
then i is not a best reply to i . Let ai be a pure action such that
i (ai ) > 0 and assume that, for some i , ui (i , i ) > ui (ai , i ), so
that ai Suppi \ri (i ).8 Since ui (i , i ) is a weighted average of
the values {ui (ai , i )}ai Ai , there must be a pure action a0i such that
ui (a0i , i ) > ui (ai , i ). But then we can construct a mixed action i0 that
7

A correspondence : X Y is a multi-function that assigns to every element

x X a set of elements (x) Y . A correspondence : X Y can be equivalently be
expressed as a function with domain X and range 2Y , the power set of Y .
8
Recall that X\Y denotes the set of elements of X that do not belong to Y :
X\Y = {x X : x
/ Y }.

38

3. Rationality and Dominance

satisfies ui (i0 , i ) > ui (i , i ) by shifting probability weight from ai to

a0i , which is possible because i (ai ) > 0. Specifically, for every ai Ai , let

if ai = ai ,
0,
0

(a ) + i (ai ), if ai = a0i ,
i (ai ) =
i i
i (ai ),
if ai 6= ai , a0i .
P
P
i0 is a mixed action since ai Ai i0 (ai ) = ai Ai i (ai ) = 1. Moreover,
it can be easily checked that
ui (i0 , i ) ui (i , i ) = i (ai )[ui (a0i , i ) ui (ai , i )] > 0,
where the inequality holds by assumption. Thus ai is not a best reply to
i .
(If ) Next we show that if each pure action in the support of i is a
best reply, then i is also a best reply. It is convenient to introduce the
following notation:
u
i (i ) :=

max ui (i , i ).

i (Ai )

By definition of ri (i ),
ai ri (i ), u
i (i ) = ui (ai , i )
i

(3.3.1)

ai Ai , u
i ( ) ui (ai , ).

(3.3.2)

Assume that Suppi ri (i ). Then, for every i (Ai )

X
X
ui (i , i ) =
i (ai )ui (ai , i ) =
i (ai )ui (ai , i ) = u
i (i )
ai Suppi

= u
i (i )

X
ai Ai

ai ri (i )

i (ai ) =

X
ai Ai

i (ai )
ui (i )

i (ai )ui (ai , i ).

ai Ai

The first equality holds by definition, the second follows from Suppi
ri (i ), the third holds by eq. (3.3.1), the fourth and fifth are obvious and
the inequality follows from (3.3.2). 
In Matrix 1, for example, any mixed action that assigns positive
probability only to a and/or b is a best reply to the conjecture ( 12 , 12 ).
Clearly, the set of pure best replies to ( 12 , 12 ) is ri (( 12 , 12 )) = {a, b}.

3.3. Best Replies and Undominated Actions

39

Note that if at least one pure action is a best reply among all pure and
mixed actions, then the maximum that can be attained by constraining
the choice to pure actions is necessarily equal to what could be attained
choosing among (pure and) mixed actions, i.e.
ri (i ) 6=

max ui (i , i ) = max ui (ai , i ).

ai Ai

i (Ai )

This observation along with Lemma 1 yields the following:

Corollary 1. For every conjecture i (Ai ),
ri (i ) = arg max ui (ai , i ).
ai Ai

Hence, it is not necessary to use mixed actions to maximize expected payoff.

Moreover, if Ai is finite, the best reply correspondence has non-empty
values, that is, ri (i ) 6= for every i (Ai ).9
Proof. For every conjecture i , the expected payoff function ui (, i ) :
(Ai ) R is continuous in i and the domain (Ai ) is a compact subset
of RAi . Hence, the function has at least one maximizer i . By Lemma 1,
Suppi ri (i ), so that ri (i ) 6= . As we have seen above, this implies
that
max ui (i , i ) = max ui (ai , i )

i (Ai )

ai Ai

and therefore ri (i ) = arg maxai Ai ui (ai , i ). 

Recall that, for a given function f : X Y and a given subset C X,
f (C) := {y Y : x C, y = f (x)}
denotes the set of images of elements of C. Analogously, for a given
correspondence : X Y ,
(C) := {y Y : x C, y (x)}
denotes the set of elements y that belong to the image (x) of some
point x C. In particular, we use this notation for the best reply
9

The same conclusion holds if A is compact and ui is continuous.

40

3. Rationality and Dominance

correspondence. For example, ri ((Ai )) is the set of justifiable actions

(see Definition 4).
A question should spontaneously arise at this point: can we
characterize the set of justifiable actions with no reference to conjectures
and expected payoff maximization? In other words, can we conclude that
an action will not be chosen by a rational player without checking directly
that it is not a best reply to any conjecture? The answer comes from the
concept of dominance.
Definition 5. A mixed action i dominates a (pure or) mixed action
i if it yields a strictly higher expected payoff whatever the choices of the
other players are:
ai Ai , ui (i , ai ) > ui (i , ai ).
The set of pure actions of agent i that are not dominated by any mixed
action is denoted by N Di .
The proof of the following statement is left as an exercise for the reader.
Remark 3. If a (pure) action ai is dominated by a mixed action i
then, for every conjecture i (Ai ), ui (ai , i ) < ui (i , i ). Hence, a
dominated action is not justifiable; equivalently, a justifiable action cannot
be dominated.
In Matrix 1, for instance, the action f is dominated by c, which in turn
is dominated by the mixed action that assigns probability 21 to a and b.
Therefore, N Di {a, b, e}. Given that a, b, and e are best replies to at
least one conjecture, we have that {a, b, e} N Di . Hence, N Di = {a, b, e}.
The following lemma states that the converse of Remark 3 holds.
Therefore, it provides a complete answer to our previous question about
the characterization of the set of actions that a rational player would never
choose.
Lemma 2. (see Pearce [29]) Fix an action ai Ai in a finite game. There
exists a conjecture i (Ai ) such that ai is a best reply to i if and
only if ai is not dominated by any (pure or) mixed action. In other words,
the set of undominated (pure) actions and the set of justifiable actions
coincide:
N Di = ri ((Ai )) .

3.3. Best Replies and Undominated Actions

41

Proof. The proof of this result uses an important result known as

Farkas lemma.
Lemma 3 (Farkas Lemma). Let M be a n m matrix and c be an ndimensional vector. Then exactly one of the following statements is true:
(1) there exists a vector x Rm such that M x c;
(2) there exists a vector y Rn such that y 0, yT M = c and yT c > 0.
We will use Farkas Lemma to prove that an action is justifiable if and
only if it is not dominated. Since the game is finite, let k = |Ai | and
m = |Ai | and label elements in Ai as {1, 2, ..., k} and elements in Ai as
{1, 2, ..., m} .10 Then, construct an k m matrix U in which the (w, z)-th
coordinate is given by:
ui (ai , z) ui (w, z)
ai is justifiable if and only if we can find a conjecture i (Ai ) such
that U i 0. This condition can be rewritten as follows: ai is justifiable
if and only if we can find x Rm satisfying inequality:
M x c,
where M is a (k + m + 1) m matrix and c
vector defined as follows:

U
M = Im , c =
1Tm

(3.3.3)
is a (k + m + 1)-dimensional

0k
0m
1

(Im denotes the m-dimensional identity matrix, that is an m m matrix

having 1s along the main diagonal and 0s everywhere else; 1s and 0s denote
the s-dimensional
vectors having respectively 1s and 0s everywhere).
Pm
x
>
1, we can
substituting each
Note, if
s=1 s
 normalize the vector


x = (x1 , ..., xm ) with x0 = Pmx1 xs , ..., Pmxm xs and satisfy inequality

s=1
s=1
(3.3.3).
If ai is not justifiable, inequality (3.3.3) does not hold and Farkas lemma
implies we can find y Rk+m+1 such that:
y0
T

y c> 0
yT M = c
10

Recall that |X| denotes the cardinality of set X.

42

3. Rationality and Dominance

By construction yT c = yk+m+1 ; thus the second condition can be rewritten

as yk+m+1 > 0. Furthermore, and yT M = c implies that for every
z Ai = {1, 2, ..., m} :
k
X

ys [ui (ai , z) ui (as , z)] + yk+z + yk+m+1 = 0

s=1

Since yk+z 0, we must have ys > 0 for some s Ai . Then, we

can construct a probability distribution y,11 such that for all z Ai ,
P
k

s=1 ys ui (as , z) > ui (ai , z) . We conclude that ai is dominated by mixed

action y. 
It is possible to understand intuitively why the lemma is true by
observing Figure 2. Note that the only if part of Lemma 2 is just Remark
3. The set of justifiable actions corresponds to the extreme points (the
kinks ) of the North-East boundary of the shaded convex polyhedron
representing the set of feasible (pure and randomized) choices. An action
ai is not justifiable if and only if the corresponding point lies (strictly)
below and to the left of this boundary. But for any point (x, y) below and
to the left of the North-East boundary, there exists a corresponding point
(x , y ) on the boundary that dominates (x, y).12
We can thus conclude that a rational player always chooses
undominated actions and that each undominated action is justifiable as
a best reply to some conjecture. Note that this result also holds for games
with compact action sets and continuous payoff functions.
In some interesting situations of strategic interaction an action
dominates all others. In those cases, a rational player should choose that
action. This consideration motivates the following:
Definition 6. An action ai is dominant if it dominates every other
action, i.e. if
ai Ai \ {ai } , ai Ai , ui (ai , ai ) > ui (ai , ai ).
P
If ks=1 ys =
6 1, we can once more normalize the vector substituting each element
yr with Pkyr y .
11

12

s=1

As Figure 2 suggests, a similar result relates undominated mixed actions and mixed
best replies: the two sets are the same and correspond to the North-East boundary of
the feasible set. But since we take the perspective that players do not actually randomize,
we are only interested in characterizing justifiable pure actions.

3.3. Best Replies and Undominated Actions

43

You should try to prove the following statement as an exercise:

Remark 4. Fix an action ai arbitrarily. The following conditions are
equivalent:
(0) action ai is dominant;
(1) action ai dominates every mixed action i 6= ai ;
(2) action ai is the unique best reply to every conjecture.
The following example illustrates the notion of dominant action.
The example shows that, assuming that players are motivated by
their material self-interest, individual rationality may lead to socially
undesirable outcomes.
Example 2. (Linear Public Good Game). In a community composed
of n individuals it is possible to produce a quantity g of a public good
using an input x according to the production function y = kx. Both y and
x are measured in monetary units (say in Euros). To make the example
interesting, assume that the productivity parameter k satisfies 1 > k > n1 .
A generic individual i can freely choose how many Euros to contribute to
the community for the production of the public good. The community
members cannot sign a binding agreement on such contributions because
no authority with coercive power can enforce the agreement. Let Wi be
player is wealth. The game can be represented as follows: Ai = [0, Wi ],
ai Ai is the contribution of i; consequences are allocations of the public
and private good (y, W1 a1 , ..., Wn an ); agents are selfish,13 hence their
utility function can be written as
vi (y, W1 a1 , ..., Wn an ) = y ai ,
which yields the payoff function
ui (a1 , ..., an ) = k(

n
X

aj ) ai .

j=1

It can be easily checked that ai = 0 is dominant for any player i: just

rewrite ui as
X
ui (ai , ai ) = k
aj (1 k)ai
j6=i
13

And risk-neutral.

44

3. Rationality and Dominance

and recall that k < 1. The profile of dominant actions (0, ..., 0) is Paretodominated by any symmetric profile of positive contributions (, ..., ) (with
> 0). Indeed, ui (0, ..., 0) = 0 < (nk 1) P
= ui (, ..., ), where the
n
1
inequality holds because k > n . Let S(a) =
i=1 ui (a) be the social
surplus; the surplus maximizing profile is a
i = Wi for each i.
An action could be a best reply only to conjectures that assign zero
probability to some action profiles of the other players. For instance, action
e in matrix 1 is justifiable as a best reply only if i is certain that i does
not choose r. Let us say that a player i is cautious if his conjecture assigns
positive probability to every ai Ai . Let


o (Ai ) := i (Ai ) : ai Ai , i (ai ) > 0
be the set of such conjectures.14 A rational and cautious player chooses
actions in ri (o (Ai )). These considerations motivate the following
definition and result.
Definition 7. A mixed action i weakly dominates another (pure or)
mixed action i if it yields at least the same expected payoff for every action
profile ai of the other players and strictly more for at least one a
i :
ai Ai , ui (i , ai ) ui (i , ai ),

ai Ai , ui (i , a
i ) > ui (i , a
i ).
The set of pure actions that are not weakly dominated by any mixed action
is denoted by N W Di .
Lemma 4. (See Pearce [29]) Fix an action ai Ai in a finite game.
There exists a conjecture i o (Ai ) such that ai is a best reply to i if
and only if ai is not weakly dominated by any pure or mixed action:


N W Di = ri 0 (Ai ) .

14

o (Ai ) is the relative interior of (Ai ) (superscript o should remind you that
this is a relatively open set, i.e. the intersection between the simplex (Ai ) RAi
Ai
and the strictly positive orthant R++
, an open set in RAi ).

3.3. Best Replies and Undominated Actions

45

The intuition is similar to the one given for Lemma 2. Looking at

Figure 2 one can see that the set of (pure or mixed) actions that are not
weakly dominated corresponds to the part with strictly negative slope of
the North-East boundary of the feasible set. This coincides with the set of
points that maximize the expected payoff for at least one conjecture that
assigns strictly positive probability to all actions of i, that is, a conjecture
that is represented by a set of parallel lines with strictly negative slope
(neither horizontal, nor vertical).
Definition 8. An action ai is weakly dominant if it weakly dominates
every other (pure) action, i.e. if for every other action b
ai Ai \{ai } the
following conditions hold:
ai Ai , ui (ai , ai ) ui (b
ai , ai ),
b
ai Ai , ui (ai , b
ai ) > ui (b
ai , b
ai ).
You should prove the following statement as an exercise:
Remark 5. Fix an action ai arbitrarily. The following conditions are
equivalent:
(0) ai is weakly dominant;
(1) ai weakly dominates every mixed action i 6= ai ;
(2) ai is the unique best reply to every strictly positive conjecture i
o (Ai );
(3) ai is the only action with the property of being a best reply to every
conjecture i (note that if i is not strictly positive, i (Ai )\o (Ai ),
then there may be other best replies to i ).
If a rational and cautious player has a weakly dominant action, then
he will choose such an action. There are interesting economic examples
where individuals have weakly dominant actions.
Example 3. (Second Price Auction). An art merchant has to auction
a work of art (e.g. a painting) at the highest possible price. However, he
does not know how much such work of art is worth to the potential buyers.
The buyers are collectors who buy the artwork with the only objective to
keep it, i.e. they are not interested in what its future price might be.
The authenticity of the work is not an issue. The potential buyer i is
willing to spend at most vi > 0 to buy it. Such valuation is completely

46

3. Rationality and Dominance

subjective, meaning that if i were to know the other players valuations, he

would not change his own.15 Following the advice of W. Vickrey, winner of
Nobel Prize for Economics,16 the merchant decides to adopt the following
auction rule: the artwork will go to the player who submits the highest
offer, but the price paid will be equal to the second highest offer (in case of
a tie between the maximum offers, the work will be assigned by a random
draw). This auction rule induces a game among the buyers i = 1, ..., n
where Ai = [0, +), and

vi max ai ,
if ai > max ai ,

0,
if ai < max ai ,
ui (ai , ai ) =

1
(v

max
a
)
if ai = max ai
i
i
1+|arg max ai |
where max ai = maxj6=i (a1 , ..., ai1 , ai+1 , ..., an ) (recall that, for every
finite set X, |X| is the number of elements of X, or cardinality of X). It
turns out that offering ai = vi is the weakly dominant action (can you
prove it?). Hence, if the potential buyer i is rational and cautious, he will
offer exactly vi . Doing so, he will expect to make some profits. In fact,
being cautious, he will assign a positive probability to event [max ai < vi ].
Since by offering vi he will obtain the object only in the event that the
price paid is lower than his valuation vi , the expected payoff from this offer
is strictly positive.

3.3.1

Best replies and Undominated Actions in Nice Games

Many models used in applications of game theory to economics and other

disciplines have the following features: players choose actions in closed
intervals of the real line, their payoff functions are continuous, and the
payoff function of each player is strictly concave, or at least strictly
quasi-concave,17 in his own action. Such games are called nice (see
15

If a buyer were to take into account a potential future resale, things might be rather
different. In fact, other potential buyers could hold some relevant information that
affects the estimate of how much the artwork could be worth in the future. Similarly, if
there were doubts regarding the authenticity of the work, it would be relevant to know
the other buyers valuations.
16
See Vickrey [34].
17
A function f : X R with convex domain is strictly quasi-concave if for every
pair of distinct points x0 , x00 X (x0 6= x00 ) and every t (0, 1),
f (tx0 + (1 t)x00 ) > min{f (x0 ), f (x00 )}.

3.3. Best Replies and Undominated Actions

47

Moulin, 1984):
Definition 9. A static game G = hI, (Ai , ui )iI i is nice if, for every
i ] R, the payoff function
player i I, Ai is a compact interval [ai , a
ui : A R is continuous (that is, jointly continuous in all its arguments),
and for each ai Ai , the function ui (, ai ) : [ai , a
i ] R is strictly
quasi-concave.
Remark 6. Let Ai = [ai , a
i ] R and fix ai Ai . Function
ui (, ai ) : [ai , a
i ] R is strictly quasi-concave if and only if the following
two conditions hold: (1) there is a unique best reply ri (ai ) [ai , a
i ] and
(2) ui (, ai ) is strictly increasing on [ai , ri (ai )] and strictly decreasing on
[ri (ai ), a
i ].
Example 4. Consider Cournots oligopoly game with capacity constraints:
I is the set of firms, ai [0, a
i ] is the output of firm i and a
i is its capacity.
The payoff function of firm i is

ui (a1 , ..., an ) = ai P ai +

aj Ci (ai ),

j6=i

P
where P : [0, i a
i ] R+ is the downward-sloping inverse demand
function and Ci : [0, a
i ] R+ is is cost function. The Cournot game
is nice if P () andP
each cost function Ci () are continuous and if, for each
i and ai , P (ai + j6=i aj )ai Ci (ai ) is strictly quasi-concave. The latter
condition is easy to P
satisfy. For example, if each
P Ci is strictly increasing
and convex, and P ( jI aj ) = max{0, p jI aj }, then strict-quasi
concavity holds (prove this as an exercise).18
Nice games allow an easy computation of the sets of justifiable actions
and of dominated actions. We are going to compare best replies to
deterministic conjectures, uncorrelated conjectures, correlated conjectures,
pure actions not dominated by mixed actions and pure actions not
dominated by other pure actions. First note that, by definition, for every
18

But strict concavity

P does not hold. Can you see why? Consider the neighborhood
of a point where P ( jI aj ) becomes zero.

48

3. Rationality and Dominance

player i the following inclusions hold:19

ri (Ai ) ri (I(Ai )) ri ((Ai )) N Di
{ai Ai : bi Ai , ai Ci , ui (ai , ai ) ui (bi , ai )},
where

Y
ij
I(Ai ) := i (Ai ) : (ij )jI\{i} jI\{i} (Aj ), i =

j6=i

is the set of product probability measures on Ai , that is, the set of beliefs
that satisfy independence across opponents. In words, the set of best
replies to deterministic conjectures in Ai is contained in the set of best
replies to probabilistic independent conjecture, which is contained in the
set of best replies to probabilistic (possibly correlated) conjectures, which
is contained in the set of actions not dominated by mixed actions, which
is contained in the set of (pure) actions not dominated by other pure
actions. In each case the inclusion may hold as an equality. The following
very convenient result states that in nice games all these (weak) inclusions
indeed hold as equalities (the proof is left as a rather hard exercise):
Lemma 5. In every nice game, the set of best replies of each player i to
deterministic conjectures coincides with the set of actions not dominated
by other pure actions, and therefore it also coincides with the set of best
replies to independent or correlated probabilistic conjectures, and with the
set of actions not dominated by mixed actions:
ri (Ai ) = ri (I(Ai )) = ri ((Ai )) = N Di
= {ai Ai : bi Ai , ai Ci , ui (ai , ai ) ui (bi , ai )}.
As a preliminary step for the proof of Lemma 5, it is necessary to prove
a technical result, i.e., that in nice games (more generally, in games
with compact and convex action sets with continuous payoff functions that
are strictly quasi-concave in the own action) best replies to deterministic
19

If you do not know how to deal with probability measures on infinite sets, just
consider the simple ones, i.e., those that assign positive probability to a finite set of
points. The set of simple probability measures on an infinite domain X is a subset of
(X), but under the assumptions considered here it is possible to restrict ones attention
to this smaller set.

3.3. Best Replies and Undominated Actions

49

conjectures are well defined continuous functions. This preliminary result

will also be used in Chapter 5 to prove the existence of a symmetric
equilibrium in nice games.
Lemma 6. Suppose that, for each i I, each Ai is a compact and convex
subset of a Euclidean space, ui is continuous, and ui (, ai ) is strictly quasiconcave for each ai Ai . Then each player i has a well defined and
continuous best reply function on the domain Ai .
Proof. By convexity of Ai and strict quasi-concavity of ui , the
set arg maxai Ai ui (ai , ai ) is a singleton for each ai . To prove this,
it is sufficient to show that, if a0i 6= a00i and ui (a0i , ai ) = ui (a00i , ai ),
then ui (a0i , ai ) < maxai Ai ui (ai , ai ). The latter is easily shown. Let
a0i and a00i be as above. By convexity of Ai , for each t (0, 1),
ta0i +(1t)a00i Ai ; by strict quasi-concavity of ui , ui (ta0i +(1t)a00i , ai ) >
min{ui (a0i , ai ), ui (a00i , ai )}. Since ui (a0i , ai ) = ui (a00i , ai ) the result
follows.
Next we show that the function ri (a
i ) = arg maxai Ai ui (ai , ai ) is
i
continuous, that is, for any sequence aki kN such that limk aki = a


k
a )
the corresponding sequence of best replies ri (ai ) kN converges to ri (

i
k
k
[limk ri (ai ) = ri (
ai )]. Note first that, since the sequence ri (ai ) kN
is contained in the compact set Ai , it must have at least
 one accumulation
such that
point a
i Ai . Therefore, there exists a subsequence aki`
lim` ri (aki` ) = a
i . The definition of best reply implies

`N

ai Ai , ` 1, ui (ri (aki` ), aki` ) ui (ai , aki` ).

Taking the limit for ` on both sides for any given ai , the continuity
of ui yields
ai Ai , ` 1, ui (
ai , a
i ) ui (ai , a
i ).
Hence, a
i is a best reply to a
i . Since the best reply is unique, a
i = ri (
ai ).
This is
true for every accumulation point. Therefore, the sequence
ri (aki ) kN has only one accumulation point, ri (
ai ), which is equivalent
k
to say that limk ri (ai ) = ri (
ai ), as desired. 

Rationalizability and
Iterated Dominance
The analysis, so far, has been based on a set of minimal epistemic
assumptions:1 every player i knows the sets of possible actions A1 , A2 ,...,
An and his own payoff function ui : Ai Ai R. From this analysis we
derived a basic, decision-theoretic principle of rationality: a rational player
should not choose those actions that are dominated by mixed actions. This
principle of rationality can be interpreted in a descriptive way (assuming
that a rational player will not choose dominated actions), or in a normative
way (a player should not choose dominated actions).
The concept of dominance is sufficient to obtain interesting results
in some interactive situations: those where it is not necessary to guess
the other players actions in order to make a correct decision. Simple
social dilemmas, like the Prisoners Dilemma or contributing to a public
good, have this feature. But when we analyze strategic reasoning, the
decision-theoretic rationality principle is just a starting point: thinking
strategically means trying to anticipate the co-players moves and plan a
best response, taking into account that the also the co-players are intelligent
individuals trying to do just the same.
Strategic thinking is based on knowledge of the rules of the game
1

We call epistemic the assumptions about knowledge, conjectures and beliefs of

players.

50

4.1. Assumptions about players beliefs

51

and preferences, knowledge of the other players knowledge of rules and

preferences, and so on and so forth. As a starting point, we assume that
players have complete information, i.e. common knowledge of the rules of
the game and everybodys preferences. In Example 1 complete information
implies that the functional forms and the parameters K, 1 , 2 are common
knowledge: both players know them, both players know that both players
know them, and so on and so forth.
The complete information assumption is sufficiently realistic for some
economic situations in which the consequences of players actions are purely
monetary outcomes and players are selfish and risk neutral.2 The complete
information assumption is also a useful simplification in that it allows to
focus on other aspects of strategic thinking than players knowledge of the
game and of the knowledge of other players. Later on, we will introduce
strategic thinking in games with incomplete information and we will show
that the techniques developed for games with complete information can be
extended to the more general case.

4.1

Assumptions about players beliefs

The analysis will proceed in a series of steps. After introducing the concept
of rationality and its behavioral consequences, we will characterize which
actions are consistent not only with rationality, but also with the belief
that everybody is rational. Then we will characterize which actions are
consistent not only with the previous assumptions, but also with the
further assumption that each player believes that each other player believes
that everybody is rational; and so on. Thus, at each further step we
characterize more restrictive assumptions about players rationality and
beliefs.
Such assumptions can be formally represented as events, i.e. as subsets
of a space of states of the world where every state is a
conceivable configuration of actions, conjectures and beliefs concerning
other players beliefs. This formal representation provides an expressive
and precise language that allows to make the analysis more rigorous and
2

Recall that what matters are players preferences over lotteries of consequences,
hence also their risk attitudes. But in some cases (for instance, some oligopolistic
models) risk attitudes are irrelevant for the strategic analysis and risk neutrality can
therefore be assumed with no loss of generality.

52

4. Rationalizability and Iterated Dominance

transparent.3 However, it calls for the preliminary introduction of some

rather complex material, which is not strictly needed to understand the
basics. Therefore we opt for a compromise. On the one hand, in order to
make the presentation more transparent and to avoid verbose sentences, we
use symbols to denote the assumptions concerning players behavior and
beliefs; we refer to these assumptions as events and denote a conjunction
of assumptions by the symbol of intersection between events. On the
other hand, rather than providing an explicit mathematical definition
for such events, we only define mathematically the set of actions (and
conjectures) that are consistent with them. Of course, we can afford to
do this precisely because we know how to represent assumptions about
rationality and beliefs as events and how to derive characterizations of
behavior based on increasingly sophisticated strategic thinking by means
of rigorous mathematical methods. We ask you, the reader, to trust us
on this derivation. Since the results are quite intuitive, we hope you will
not object. (If you do object, perhaps you are ready for the complete
mathematical analysis of strategic thinking: good news for us!)
To see what we mean, consider the rationality assumption. We denote
by Ri the event player i is rational, whereas R = R1 ... Rn denotes
the event everybody (in the game) is rational. In the previous chapter
we saw that N D1 ...N Dn (where N Di is the set of undominated actions
of i) is the set of action profiles consistent with R (rationality).
The rationality assumption simply establishes a relation between
conjectures (beliefs about the behavior of other players) and actions. This
relation is represented by the best reply correspondence. Now we proceed
introducing some assumptions concerning players beliefs about each other
beliefs, also called interactive beliefs.
Denote by E a generic event about players actions and beliefs (for
example, E = R). We represent with the symbol B(E) the event
everybody believes E, meaning everybody assigns probability one
to E. Further, we write B(B(E)) = B2 (E) (everybody believes that
everybody believes E ), and more generally Bk (E) = B(Bk1 (E)) for any
integer k > 1.
Consider the conjunction of the following assumptions: R, B(R),
2
B (R), ...Bk (R), etc. Is it possible to provide a characterization in terms
3

See, for example, the survey by Battigalli and Bonanno [3].

4.2. The Rationalization Operator

53

of actions? In other words, what sets of action profiles

with
T are consistent

K
2
k
assumptions R B(R), R B(R) B (R), ..., R
k=1 B (R) , etc.? In
order to answer this question, it is useful to introduce a mapping, which
is derived from the best reply correspondence and assigns to any crossproduct subset C of A a subset of action profiles that are rationalized,
or justified by C.

4.2

The Rationalization Operator

Fix some set X and a collection C of subsets of X (C 2X ). In these

notes we call operator (on X, C) a mapping from C to itself, that is, a
mapping that associates to each subset of X in C another subset of X in
C.4
As a matter of notation, for a fixed set Xi (for instance, Xi = Ai )
and a subset Yi Xi , we denote by (Yi ) the subset of probability
measures on Xi that assign probability one to Yi :
(Yi ) := {i (Xi ) : i (Yi ) = 1}
(another slight abuse of notation).
In particular, we are going to consider the set X = A = iI Ai of
action profiles, and the collection C of all Cartesian subsets of A, i.e.
subsets of the cross-product form C = C1 ... Cn , where Ci Ai for
every i. Then, if C is finite:5
C := {C 2A : (Ci )iI iI 2Ai , C = C1 ... Cn }.
Let Ci = jI\{i} Cj . We define the following sets:


i (Ci ) := ai Ai : i (Ci ), ai ri (i ) = ri ((Ci )),
4

In mathematics, the term operator is mostly used for mappings from a space of
functions to a space of functions (possibly the same). But sets can always be represented
by functions (e.g. indicator functions). Therefore the present use of the term operator
is consistent with standard mathematical language.
5
For infinite games, we have to specify the collections of events Ai 2Ai (i I) and
A 2A , where iI Ai A, and we replace 2Ai (respectively, 2A ) with Ai (respectively,
A).For example, if Ai is a subset of a Euclidean space (more generally, a metric space),
then Ai = B(Ai ) (respectively, A = B(A)) is the collection of Borel subsets of Ai
(respectively, A). Standard regularity assumptions, such as compactness and continuity,
ensure that (C) is an event, hence (C) C, for each C C.

54

4. Rationalizability and Iterated Dominance

(C) := 1 (C1 ) ... n (Cn ).

Interpretation: (C) is the set of action profiles that could be chosen by

rational players if every i is certain that the co-players choose in Ci .
Therefore, we say that (C) is the set of action profiles rationalized (or
justified ) by C and we call the mapping : C C rationalization
operator. 6 Note that () = .
`

3, 2

0, 1

0, 0

0, 2

3, 1

0, 0

1, 1

1, 2

3, 0

Example 5. In a 3 3 game like the one above, each player has 23 = 8

subsets of actions (including the empty set). Hence, the collection of
subsets C contains 8 8 = 64 elements, 7 7 = 49 of them are nonempty
(since () = , we may ignore the empty subsets). It would be very
tedious to determine (C) for each (nonempty) C C. We give just a few
examples:
({T, M, B} {l, c, r}) = {T, M, B} {l, c},
({T, M, B} {l, c}) = {T, M } {l, c},
({T, M } {l, c}) = {T, M } {l},
({M, B} {l, c}) = {T, M } {l, c},
({T, M } {l}) = {T } {l},
({T, M } {r}) = {B} {l},
({T } {l}) = {T } {l},
({B} {r}) = {B} {c}.
To see this, check the best replies to deterministic conjectures, note that r
is strictly dominated by l and c, next note that B is strictly dominated by
1 = 12 T + 12 M in the smaller game obtained by deleting the right column
6

The rationalization operator represents an example of justification operator, a

concept which was first explicitly presented by Milgrom and Roberts (1991).

4.3. Rationalizability: The Powers of

55

r. This shows that, in general, we may have (C) C, C (C) (here

this inclusion holds as an equality for C = {T } {l}), (C)
C and
C (C).7
Remark 7. Lemma 2 implies that i (Ai ) is the set of undominated
actions for player i. Hence, (A) = N D1 ... N Dn .
Remark 8. The rationalization operator is monotone: for every pair of
subsets E, F C, if E F then (E) (F ). To see this note that,
for every i, Ei Fi implies (Ei ) (Fi ) which in turn implies
i (Ei ) = ri ((Ei )) ri ((Fi )) = i (Fi ).
Since the domain and range of coincides, it make sense to define the
k-th iteration of (the k-fold composition of with itself) recursively as
follows: For each C C, define 0 (C) = C for convenience; then for each
k 1,
k (C) = (k1 (C)).

4.3

Rationalizability: The Powers of

The set of action profiles consistent with R (rationality of all the players)
is (A). Therefore, if every player is rational and believes R, only those
action profiles that are rationalized by (A) can be chosen. It follows that
the set of action profiles consistent with R B(R) is ((A)) = 2 (A).
Iterating the procedure one more step, it is relatively easy to see that the
set of action profiles consistent with R B(R)B2 (R) is (2 (A)) = 3 (A).
The general relationship between events about rationality and beliefs and
corresponding sets of action profiles is given by the Table 4.1.
Note that, as one should expect, the sequence of subsets (k (A))kN
is weakly decreasing, i.e. k+1 (A) k (A) (k = 1, 2, ...). This fact can
be easily derived from the monotonicity of the rationalization operator:
by definition, 1 (A) = (A) A = 0 (A); if k (A) k1 (A), the
monotonicity of implies k+1 (A) = (k (A)) (k1 (A)) = k (A). By
the induction principle k+1 (A) k (A) for every k. Every monotonically
decreasing sequence of subsets has a well defined limit: the intersection of
7

Symbol means not weakly included in.

56

4. Rationalizability and Iterated Dominance

Table 4.1: Behavioral implications of assumptions on rationality and beliefs.

Assumptions about rationality and beliefs

Behavioral implications

(A)

R B(R)

2 (A)

R B(R) B2 (R)

3 (A)

...
R

T

...

K
k
k=1 B (R)

K+1 (A)

...

...

all the sets in the sequence. Therefore it makes sense to define:

\
(A) :=
k (A).
k1

Example 6. In the game of example 5, iterating starting from A we

delete one action at each step and we stop after 4 steps with one action
left for each player:
(A) = ({T, M, B} {l, c, r}) = {T, M, B} {l, c},
2 (A) = ({T, M, B} {l, c}) = {T, M } {l, c},
3 (A) = ({T, M } {l, c}) = {T, M } {l},
k (A) = ({T, M } {l}) = {T } {l}, (k 4).
It is easy to see that (A) need not be a singleton:
b

4, 3

0, 2

0, 0

0, 1

3, 4

0, 0

1, 1

1, 2

5, 0

Example 7. Here is a story for the payoff matrix above: Rowena and
Colin have to decide independently of each other whether to go to a Bach

4.3. Rationalizability: The Powers of

57

concert, a Stravinsky concert, or a Chopin concert. Rowena likes Bach

more than Stravinsky, she loves Chopin, but she prefers to go to another
concert with Colin rather than listening to Chopin alone. Colin hates
Chopin, he likes Stravinsky more than Bach, but he prefers Bach with
Rowena to Stravinsky alone. This is a variation on the Battle of the
Sexes. Iterating from A = {B, S, C} {b, s, c} we get
(A) = ({B, S, C} {b, s, c}) = {B, S, C} {b, s},
2 (A) = ({B, S, C} {b, s}) = {B, S} {b, s},
k (A) = ({B, S} {b, s}) = {B, S} {b, s}, (k 3).
Strategic thinking leads Rowena to avoid the Chopin concert, but it does
not allow to solve the Bach-Stravinsky coordination problem.
We argued informally that an action profile a is consistent with
rationality and common belief in rationality if and only if a (A).
Such action profiles are called rationalizable :
Definition 10. (Bernheim [8], Pearce [29]) An action profile a A is
rationalizable if a (A).
In finite games, the set of rationalizable profiles is obtained in a finite
number of steps (the number depends on the game). In infinite games
with compact action spaces and continuous payoff functions the set of
rationalizable profile is obtained with countably many steps (i.e. finitely
many, or a denumerable sequence of steps).
Theorem 1. (a) If, for every player i I, Ai is finite, then there exists
a positive integer K such that K+1 (A) = K (A) = (A) 6= . (b) If,
for every player i I, Ai is a compact subset of a Euclidean space8 and
the function ui is continuous, then the set of rationalizable action profiles,
(A), is non-empty, compact and satisfies (A) = ( (A)).
8

We call Euclidean space a Cartesian space Rm , endowed with the Euclidean

distance and the associated topology (collection of open sets). A subset of a Euclidean
space is compact if and only if it is closed and bounded (this is a theorem, not a definition
of compactness). An event in this space is a Borel set, that is, an element of the smallest
-algebra containing all the open sets. The conjectures of each player i are defined as
the Borel probability measures on Ai .

58

4. Rationalizability and Iterated Dominance

Proof. (a) If Ai is finite, for every conjecture i the set of best replies
ri (i ) is non-empty. Then for each non-empty Cartesian set C A, (C)
is non-empty (for every i, there exists at least a conjecture i (Ci )
with 6= ri (i ) i (Ci )). Thus, k (A) 6= implies k+1 (A) 6= . Since
A 6= , it follows by induction that k (A) 6= for every k.
The sequence of the subsets (k (A))kN is weakly decreasing. Also, if
k (A) = k+1 (A), then k (A) = ` (A) for every ` k. Given that A is
finite, the inclusion k+1 (A) P
k (A) can be strict only for a finite number
of steps K (in particular, K iI |Ai |n, where |Ai | denotes the number
of elements of Ai : when k+1 (A) ( k (A), at least one action for at least
one player i is eliminated, but at least one action for each player is never
eliminated). All the above implies that K+1 (A) = K (A) = (A) 6= .
(b) (This part contains more advanced mathematical arguments and
can be omitted.) Since ui : Ai Ai R is continuous, then also
the expected utility function ui (ai , i ) is continuous in both arguments
((Ai ) is assumed to be endowed with the topology of weak convergence
of measures).9 Given that Ai is compact, the maximum principle implies
that the best reply correspondence ri : (Ai ) Ai is upper hemicontinuous (u.h.c.) with non-empty values. For every non-empty and
compact Ci Ai , the set (Ci ) is non-empty and compact. It follows
that its image through the u.h.c. correspondence ri is also non-empty and
compact. Therefore for any non-empty and compact product set C C,
(C) = r1 ((C1 )) ... rn ((Cn )) is a non-empty, compact set. It
follows by induction that (k (A))kN is a weakly decreasing sequence of
non-empty
and compact sets. Hence, for every K = 1, 2, ..., the intersection
TK k
K
is non-empty and compact. By the finite intersection
k=1 (A) = (A) T
10

property, (A) = k1 k (A) is also non-empty and compact.

Since (A) k (A) for every k = 0, 1, ..., the monotonicity of
9

Let C Rm be a compact set and consider the probability measures defined over the
Borel subsets of C. A sequence of probability measures k converges weakly
R to a measure
R
if, for every continuous (hence bounded) function f : C R, limk f dk = f d.
10
Any collection of compact subsets {C k ; k Q} has the following property (called
finite intersection property
): if every finite sub-collection {C k ; k F } (F Q)
T
has non-empty intersection ( kF C k 6= ), then also the whole collection has non-empty
T
T
intersection ( kQ C k 6= ). Moreover, C := kQ C k is an intersection of close sets
and hence it is closed. Since C is a closed set contained in a compact set (C C k for
every k Q), C must also be compact.

4.3. Rationalizability: The Powers of

59

implies that
( (A))

(k (A)) =

k0

k (A) = (A).

k1

Let us now show that (A) ( (A)).

Let ai be a rationalizable action. Then for every k = 1, 2, ... there
exists a conjecture i,k (ki (A)) such that ai ri (i,k ), where ki (A)
denotes the projection onto Ai of k (A): ki (A) = projAi k (A).11 Since
(Ai ) is compact, it can be assumed without loss of generality that the
sequence of conjectures (i,k )kN converges to a limit i (Ai ) (any
sequence in a compact set admits a convergent subsequence, hence, even if
the original sequence does not converge, one can consider any convergent
subsequence). Such limit, i , necessarily assigns probability one to
i (A).
To see this, note that
k,K k, i,K (ki (A)) = 1,
hence
i (ki (A)) =

lim i,K (ki (A)) = 1,

\
i
i (
ki (A) = lim i (ki (A)) = 1
i (A)) =
K

k1

(the first equality is implied by the weak convergence i,K i , the

second to last equality follows from the continuity, or sigma-additivity,
of probability measures). Given that the best reply correspondence ri is
i
upper hemi-continuous, ai is also a best reply to the limit

conjecture :
i
i

ai ri ( ), with (i (A)). Hence, ai i i (A) . This holds for

every player i. Thus, for every a (A), a ( (A)). 
It is possible to give an alternative and useful characterization of
rationalizable actions. Consider the following:
Definition 11. A set C C has the best reply property if C (C).
11

Recall that, given a set C X Y , the projection of C onto X is the set {x X :

y Y, (x, y) C}.

60

4. Rationalizability and Iterated Dominance

Remark 9. By Theorem 1, the set (A) of rationalizable action profiles

has the best reply property.
We leave the proof of the following as an exercise:
Remark 10. For every pair of Cartesian subsets E, F C, if both E and
F have the best reply property then also C = iI (Ei Fi ) has the best
reply property.
Theorem 2. Under the assumptions of Theorem 1, an action profile a A
is rationalizable if and only if a C for some Cartesian subset C A with
the best reply property.
Proof. (Only if ) If a is rationalizable then a (A) = ( (A)),
where the equality follows from Theorem 1. Hence, a belongs to a set with
the best reply property, viz. (A).
(If ) Let a C (C). We will prove by
that C k (C)
T induction
k
k
(A) for each k. Therefore a C
(A) = (A), i.e. a is
kN

rationalizable.
Basis step. Since C (C), C A and is monotone, it follows that
C 1 (C) 1 (A).
Inductive step. Suppose that C k (C) k (A). By monotonicity
(C) (k (C)) (k (A)).
Since C (C) and (k ()) = k+1 (), we obtain
C k+1 (C) k+1 (A).

Remark 11. The proof of Theorem 2 clarifies that, under the assumptions
of Theorem 1, (A) has the best reply property, and for every C C with
the best reply property, C (A); therefore (A) is the largest set with
the best reply property.12
We gave a definition of rationalizability based on iterations of the
rationalization operator because it is the most intuitive. An alternative
More precisely, (A) is the maximal set with the best reply property with respect
to the partial order on C.
12

4.4. Iterated Dominance

61

definition of rationalizability is suggested by Theorem 2: a is rationalizable

if it belongs to some set C with the best reply property. This second
definition is equivalent to the first one for all games where (A) =
( (A)). But there are some badly behaved infinite games where
( (A)) (A) (where means strict inclusion). The second
definition of rationalizability is valid for those games too, while the first
one only gives a necessary condition. For the purposes of these lectures,
this technicality can be neglected.

4.4

Iterated Dominance

The equivalence between best replies and undominated actions allows us

to characterize the actions that survive the iterated dominance procedure.
In order to give the precise definition of this procedure, we first define the
concept of dominance within a subset of action profiles.
Definition 12. Fix a non-empty Cartesian subset C of A: =
6 C C.
An action ai is dominated in C if ai Ci and there is a mixed action
i (Ci ) such that
ai Ci , ui (ai , ai ) < ui (i , ai ).
We denote by N Di (C) Ai the set of actions in Ci that are not dominated
in C and let N D(C) = N D1 (C) ... N Dn (C) A denote the set of
undominated action profiles in C.
Remark 12. N D() is not monotone (can you explain why?).
Using the standard notation for iterations (k-fold composition) of
operators, we can represent the iterated dominance procedure through
the following sequence of sets N D = N D(A), N D(N D(A)) = N D2 (A),
..., N Dk (A), ... . Essentially, the idea is to eliminate all dominated
actions, thus obtaining N D(A). Then one moves on to eliminate all
dominated actions in the restricted game with set of action profiles N D(A),
thus obtaining N D2 (A), then eliminate all the dominated actions from
the restricted game with set of action profiles N D2 (A), thus obtaining
N D3 (A), and so on.
Definition 13. a TA is a profile of iteratively undominated actions
if a N D (A) := k1 N Dk (A).

62

4. Rationalizability and Iterated Dominance

Theorem 3. (Pearce [29]) Fix a finite game, for every k = 1, 2, ...,

k (A) = N Dk (A). Therefore, an action profile is rationalizable if and
only if it is iteratively undominated.
We first prove a quite simple preliminary result about the sequence
of subsets {k (A)}k1 . For any non-empty subset Ci Ai define the
constrained best reply correspondence ri (|Ci ) : (Ai ) Ai as
ri (i |Ci ) := arg max ui (ai , i ).
ai Ci

Recall that ri (i ) = arg maxai Ai ui (ai , i ). Therefore, for each i

(Ai ),
=
6 ri (i ) Ci ri (i ) = ri (i |Ci );
(4.4.1)
in words, if there is at least one best reply (as in all finite, or compactcontinuous games) and all the best replies satisfy the constraint of being in
Ci , then constrained and unconstrained best replies must coincide (check
you can prove this; more generally, prove that ri (i )Ci 6= ri (i |Ci )
ri (i )).13 For each cross-product set C C and i I, let
i (C) := ri ((Ci )|Ci ) =

[
i (Ci )

arg max ui (ai , i );

ai Ci

(C) := 1 (C) ... n (C).

Since i (C) is obtained by constrained maximization, it follows that
(C) C for every C. The rationalization operator instead does not
satisfy this property (if C, for instance, is the set of dominated action
profiles, then (C) C = ). Also, like N D(), () is not monotone,
whereas () is monotone. Yet, and coincide whenever the constraint
does not bite :
Remark 13. Suppose that ri (i ) 6= for each i I and i (Ai ).
Then,
C C, (C) C (C) = (C).
(4.4.2)
13

In some infinite games, for some i, i and Ci , ri (i ) = and ri (i |Ci ) 6= . See

Example 8.

4.4. Iterated Dominance

63

Proof. Suppose that (C) C. Since (C) = iI ri ((Ci )),

ri (i ) ri ((Ci )) Ci for each i and i (Ci ) (the first
inclusion holds by definition, the second by assumption). Under the
stated assumptions ri (i ) 6= for every i and i . Therefore, by (4.4.1),
ri (i ) = ri (i |Ci ) for every i and i (Ci ), which implies i (Ci ) =
ri ((Ci )) = ri ((Ci )|Ci ) = i (C) for each i, that is, (C) = (C). 
This implies that if and are iterated starting from the set A of all
action profiles, the same sequence of subsets obtains:
Lemma 7. Under the assumptions of Theorem 1, for every k = 1, 2, ...,
k (A) = k (A).14
Proof. By definition 0 (A) = 0 (A). Suppose by way of
that k (A) = k (A). As already shown, the monotonicity of
(k (A)) = k+1 (A) k (A); hence (k (A)) = (
k (A))
i
Under the assumptions of Theorem 1 ri ( ) 6= for each i
i (Ai ). Therefore (4.4.2) applies and yields (
k (A)) =
k+1
k
k
k+1
Thus, (A) = ( (A)) = (
(A)) = (A). 

induction
implies
k (A).
I and
(
k (A)).

The lemma can be reformulated as follows. Say that an action ai is

iteratively justifiable if (1) it is justifiable (2) it is justifiable in the
reduced game G1 obtained by elimination of all the non justifiable actions,
(3) it is justifiable in the reduced game G2 obtained by the elimination of
all non justifiable actions in G1 , and so on. The actions that are iteratively
justifiable are exactly those obtained by iterating the operator . Hence,
Lemma 7 states that, for every player, the set of rationalizable actions
coincides with the set of iteratively justifiable actions.
Example 8. Here is an example that violates the compactness-continuity
hypothesis of Theorem 1 and where, as a consequence, ri (i ) may be
empty and the thesis of Lemma 7 does not hold: I = {1, 2}, A1 = {0, 1},
A2 = {1 k1 : k = 1, 2, ...} {1}, the payoff functions are given by the
following infinite bi-matrix
14
As should be clear from the proof, the lemma holds for all games where the best
reply correspondences are non-empty valued, hence all games where A is compact and
each ui is continuous.

64

4. Rationalizability and Iterated Dominance

1
k

1\2

...

1, 0

...

1, 1

0, 0

...

0, 0

1
k

...

...

1, 0

...

0, 1

Thus, u2 1 k1 , 2 = 2 (0) 1 k1 , u2 (1, 2 ) = 2 (1) = 1 2 (0). If
2 (1) 21 the best reply is a2 = 1. If 2 (1) < 12 , then u2 (1 k1 , 2 ) >
u2 (1, 2 ) for k sufficiently large, but there is no best reply because
u2 (1 k1 , 2 ) is strictly increasing in k. Note also that a1 = 0 strictly
dominates a1 = 1. Clearly (A) = (A) = {0} {1} and 2 ({0}) = ,
2 ({0} {1}) = {1}. Therefore 2 (A) = =
6 {0} {1} = 2 (A). How
is compactness-continuty violated? A1 is finite, hence compact; A2 is a
closed and bounded subset of the real line (A2 contains the limit of the
sequence (1 k1 )
k=1 ), hence

it is also compact;15but u2 is discontinuous at
1
(0, 1): limk u2 0, 1 k = 1 6= 0 = u2 (0, 1).
Theorem 3 follows quite easily from Lemma 7: indeed Lemma 2 implies
that the iteratively undominated actions coincide with the iteratively
justifiable ones, and by Lemma 7 the iteratively justifiable actions coincide
with the rationalizable actions. The details are as follows:
Proof of Theorem 3.
Basis step. Lemma 2 implies that (A) = N D(A).
Inductive step. Assume that k1 (A) = N Dk1 (A) (inductive
hypothesis) and consider the game Gk1 where the set of actions of
each player i is ik1 (A), and the payoff functions are obtained from
the original game by restricting their domain to k1 (A). The inductive
hypothesis implies that the set of undominated action profiles in Gk1
is N D(k1 (A)) = N D(N Dk1 (A)) = N Dk (A).
By Lemma 2,
N D(N Dk1 (A)) = (N Dk1 (A)). The inductive hypothesis and Lemma
7 yield (N Dk1 (A)) = (k1 (A)) = (k1 (A)) = k (A). Hence
N Dk (A) = k (A).
So far, we have considered a procedure of iterated elimination which
is maximal, in the sense that at any step all the dominated actions of
15

For the mathematically savy: it is easy to change the example so that A2 is not
compact and u2 is trivially continuous, just endow A2 with the discrete topology.

4.4. Iterated Dominance

65

all players are eliminated (where dominance holds in the restricted game
that resulted from previous iterated eliminations). However, one can show
that to compute the set of action profiles that are iteratively undominated
(and therefore rationalizable) it is sufficient to iteratively eliminate some
actions which are dominated for some player until in the restricted game
so obtained there are no dominated actions left. Formally:
Definition 14. An iterated dominance procedure is a sequence
K+1 of non-empty Cartesian subsets of A such that (i) C 0 = A,
(C k )K
k=0 C
(ii) for each k = 1, ..., K, N D(C k1 ) C k C k1 ( is the strict
inclusion), and (iii) N D(C K ) = C K .
In words, an iterated dominance procedure is a sequence of steps
starting from the full set of action profiles (condition i) and such that
at each step k, for at least one player, at least one of the actions that
are dominated given the previous steps is eliminated (condition ii); the
elimination procedure can stop only when no further elimination is possible
(condition iii).
Theorem 4. Fix a finite game. For every iterated dominance procedure
K is the set of rationalizable action profiles.16
(C k )K
k=0 , C
Proof. Fix an iterated dominance procedure, i.e. a sequence of subsets
that satisfies (i)-(ii)-(iii) of Definition 14.
Claim. For each k = 0, 1, ..., K,

(C k )K
k=0

k+1 (A) (C k ) = N D(C k ).

The proof of the claim is by induction:
Basis step. By Lemma 2 (A) = N D(A); by (i) C 0 = A. Therefore
1
(A) (C 0 ) = N D(C 0 ) (the weak inclusion actually holds as an
equality).
Inductive step. Suppose that k+1 (A) (C k ) = N D(C k ). By
(ii), N D(C k ) C k+1 C k . By monotonicity of and the inductive
hypothesis,
k+2 (A) (N D(C k )) (C k+1 ),
16
The result does not hold for all infinite games. A version of this result does hold
if A is compact and the functions ui are continuous (of course, it may be necessary to
consider countably infinite sequences of subsets; see Dufwenberg and Stegeman, 2002).

66

4. Rationalizability and Iterated Dominance

and
(C k+1 ) (C k ) = N D(C k ) C k+1 .
Since the game is finite, there is at least one best reply to every conjecture
and (4.4.2) holds; thus,
(C k+1 ) = (C k+1 ) = N D(C k+1 ),
where the second equality follows from Lemma 2. Collecting inclusions
and equalities, k+2 (A) (C k+1 ) = N D(C k+1 ). 
The claim implies that K+1 (A) (C K ) = N D(C K ). By (iii),
K
C = N D(C K ). Therefore (A) (C K ) = C K , that is, every
rationalizable profile belongs to C K , and C K has the best reply property.
By Theorem 2, C K must be the set of rationalizable profiles. 

4.5

Rationalizability and Iterated Dominance in

Nice Games

Recall that in a nice game the action set of each player is a compact
interval in the real line, payoff functions are continuous and each ui is
strictly quasi-concave in the own action ai (see Definition 9). It turns out
that the analysis of rationalizability and iterated dominance in nice games
is nice indeed!
Let C denote the collection of closed subsets of A with a cross-product
form C = C1 ... Cn (closedness is a technical requirement; for finite
games, this coincides with the earlier definition of C).17 Define the
following operators, or maps from C to C:
(C) = iI ri (Ci ), (C) = iI ri (I(Ci )),
N Dp (C) = iI {ai Ci : bi Ci , ai Ci , ui (ai , ai ) ui (bi , ai )},
where I(Ci ) is the set of product probability measures on Ci , or
independent (uncorrelated) conjectures (see Section 3.3.1). Note that
and are monotone, but N Dp (like N D) is not monotone.
17

We could consider all the Cartesian products of Borel subsets, but this is not
necessary. We can restrict our attention to closed sets because A is closed, and for
every closed set C = C1 ... Cn both and the operators defined below give another
closed set. This follows from the continuity of best reply functions of nice games.

4.5. Rationalizability and Iterated Dominance in Nice Games

67

Iterating from A yields a solution procedure called independent

rationalizability,18 which is what obtains if (a) players are rational, (b)
they hold independent (i.e., non correlated) conjectures, and (c) there is
common belief of (a) and (b).
Iterating from A yields a solution procedure called pointrationalizability (Bernheim, 1984). This procedure has no autonomous
conceptual interest,19 because there is no good reason to assume at the
outset that players hold deterministic conjectures. Point rationalizability
is just an ancillary solution concept that turns out to be convenient in the
analysis of special games, including nice games.
Iterating N Dp from A yields the iterated deletion of (pure) actions
strictly dominated by (pure) actions. In general games with compact
action sets and continuous payoff functions, N Dk (A) (N Dp )k (A) for
all k.20 Clearly ((N Dp )k (A))kN is a much simpler procedure than
(N Dk (A))kN , but it is weaker and it does not have a satisfactory
conceptual foundation. So, in general it should be regarded as an ancillary
algorithm that allows to simplify the analysis of compact-continuous
games.
As an exercise, use monotonicity to prove the following statements:
Remark 14. For each k, k (A) k (A) k (A). Hence (A)
(A) (A).
Remark 15. For each C C, if C (C) (C (C)), then C (A)
(C (A)).
Thus, if (A) 6= , then (A) 6= and (A) 6= . Also, suppose
that C 6= is not a singleton (maybe it is infinite) and C (C). Then
the results above imply that there is a multiplicity of rationalizable (and
independent rationalizable) action profiles. These simple results show why
point rationalizability is a useful solution concept even though it does not
18

This was the solution originally called rationalizability by Bernheim (1984) and
Pearce (1984) before the epistemic analysis proved that the cleaner and more basic
concept is the correlated version (A).
19
Point rationalizability obtains if (a) players are rational, (b) they hold deterministic
conjectures, and (c) there is common belief of (a) and (b).
20
The result holds more generally for games with compact actions sets, where the
payoff functions are jointly continuous in the opponents actions and upper semicontinuous in the own action.

68

4. Rationalizability and Iterated Dominance

have an interesting conceptual foundation. One nice feature of nice games

is that Lemma 5 yields
Theorem 5. In every nice game, for every k N,
k (A) = k (A) = k (A) = N Dk (A) = (N Dp )k (A)
and each set is a product of compact intervals; hence
(A) = (A) = (A) = N D (A) = (N Dp ) (A)
and, again, each set is a product of compact intervals.
Proof. The second statement follows easily from the first, which
is trivially true for k = 0. Suppose, by way of induction, that the
first statement holds for a given k. By the inductive hypothesis the set
C = k (A) is a product of compact intervals. Then Lemma 5 (applied to
restricted action sets) yields
( k (A)) = ( k (A)) = ( k (A)) = N D( k (A)) = N Dp ( k (A)).
By the inductive hypothesis
k+1 (A) = k+1 (A) = k+1 (A) = N Dk+1 (A) = (N Dp )k+1 (A).

Example 9. As an application, consider the market for a crop analyzed
in the Introduction (section 1.6.1), but this time suppose that there is a
finite number n 2 of firms who decide how much to produce of a crop
to be sold on the market 6 months later at the highest price that allows
demand to absorb total output; thus, they compete `
a la Cournot.21 Each
firm has cost function


1
2
2
C(qi ) =
(qi ) , qi 0,
2m

22 Market demand
( 4
is a capacity constraint given by the available land).
is given by the function D(p; n) = max{0, n( p)} with 0 < < 2;
therefore price as a function of average quantity is
!
(
!)
n
n
1X
1
1X
P
qi = max 0,

qi
.
n

n
i=1

21
22

i=1

The following analysis is adapted from Boergers and Janssen (1995).

The capacity constraint is loose enough to simplify the calculations.

4.5. Rationalizability and Iterated Dominance in Nice Games

69

It is interesting to relate the analysis of rationalizable oligopolistic behavior

to the standard competitive equilibrium for these market fundamentals.
(1) First, verify that the game is nice. The payoff function of each firm i
is the profit function


(
P
qi
qi
1 P
1

q
2m
(qi )2 , if n > qi + j6=i qj ,
j
j6
=
i

n
n
i (qi , qi ) =
P
1
2m
(qi )2 ,
if n qi + j6=i qj .
P
This function is clearly continuous.23 Now, fix qi arbitrarily. If j6=i qj
1
n then i (qi , qi ) = 2m
(qi )2 , which is strictly decreasing in qi . If
P
j6=i qj < n, then i is initially increasing in qi up to the point ri (qi )
where marginal revenue equals marginal cost (see below), and then it
becomes strictly decreasing.24
(2) The best P
reply function is easily obtained from the first-order conditions
1
when > n j6=i qj , in the other case the best reply is the corner solution
qi = 0:
(
)
P
m( n1 j6=i qj )
ri (qi ) = max 0,
.
+ 2m
n
(3) The monopolistic output is
q n,M := ri (0, ..., 0) =
23
24

Note that

qi
B

qi
n

1
n

j6=i


qj

1
(q )2
2m i

m
.
+ 2m
n
1
= 2m
(qi )2 if nA = qi +

j6=i

qj .

There is a kink at qi = A j6=i qj , but this is immaterial. Note also that i is not
concave in qi : in the left neighborhood of the kink point qi the derivative is

X
X
i
q
1
q
q
1
qi
i
i
i
(qi , qi ) = P +
qj + P 0 +
qj
qi
n
n
n
n
n
m
j6=i
j6=i

qi 0 qi
1 X qi
qi

P
+
qj
<
n
n
n
m
m
j6=i

where the approximation holds because qi is close to qi and P

0

qi
n

1
n

and the inequality holds because P < 0. In a right neighborhood of qi ,

i
qmi qm
. Therefore, for > 0 small enough,

i
i
i (
qi , qi ) <
i (
qi + , qi ),
qi
qi
which implies that i is not concave.

j6=i qj
i
(qi , qi )
qi

0,
=

70

4. Rationalizability and Iterated Dominance

P
If every competitor produces at maximum capacity n1 j6=i qj > (because
< 2 and capacity is 4/), so the best reply is


4
4
= 0.
ri
, ...,


n
Since the game is nice, () = 0, q n,M . If this set has the best response
property, this is also the set of rationalizable profiles. To check this, it is
enough to verify whether the best reply to the most pessimistic conjecture
consistent with rationality
is zero, that is ri (q n,M , ..., q n,M ) = 0; in the

n
affirmative case 0, q n,M has the best reply property. The condition for
m
this is n1
n + 2m , or
n


m

n3
n


.

Therefore,
large enough so that, for each n > n
,
 n,M n if < m there is n
0, q
has the best reply property, hence there is a huge multiplicity
of rationalizable outcomes. Conversely, if > m, then
n3
)
n

n
for every n; hence, in this case the set 0, q n,M does not have the best
reply property whatever the number n of firms. Furthermore, it can be
shown that if > m there is a unique rationalizable outcome.25 By
symmetry, each firm has the same rationalizable output q n, which must
> m(

solve q =

q)
m( n1
n
,
+ 2m
n

because the singleton {(q n, , ..., q n, )} has the best

reply property. Therefore

q n, :=

m
+ m n+1
n

with corresponding rationalizable price

pn, := P
25

m
+ m n+1
n

!
=

m
n
m n+1
n

+
+

If B > m, the joint best reply function

n
(qi )n
i=1 7 (ri (qi ))i=1

is a contraction, which implies that the sequence k (A)

kN

shrinks to a point.

4.5. Rationalizability and Iterated Dominance in Nice Games

71

(of course, this is the Cournot-Nash equilibrium).

(4) Finally, the competitive equilibrium price with these market
fundamentals is

,
p =
+m
with average output
q =

m
.
+m

As pointed out in section 1.6.1 of the Introduction, > m is the cobweb

stability condition. Under this condition, the unique rationalizable
m
output is q n, = +m
n+1 . Of course,
n

m
m

n+1 = + m = q ,
+m n

+ m

n
=
= q.
= lim
n+1
n + m

+
m
n

lim q n, =

lim pn,

lim

To sum up, under the cobweb stability condition > m, the

rationalizable outcome is unique and approximates the so called rational
expectations equilibrium.

Equilibrium
In chapter 4, we analyzed the behavioral implications of rationality and
common belief in rationality when there is complete information, that is,
common knowledge of the rules of the game and of players preferences.
There are interesting strategic situations where these assumptions about
rationality, belief and knowledge imply that each player can correctly
predict the opponents behavior, e.g. many oligopoly games. But this
is not the case in general: whenever a game has multiple rationalizable
profiles there is at least one player i such that many conjectures i are
consistent with common belief in rationality, but at most one of them can
be correct. In this chapter, we analyze situations where players predictions
are, in some sense, correct, and we discuss why predictions should (or
should not) be correct. Let us give a preview.
We start with Nashs classic definition of equilibrium (in pure actions):
an action profile a = (ai )iI is an equilibrium if each action ai is a best
reply to the actions of the other players ai .
Nash equilibrium play follows from the assumptions that players are
rational and hold correct conjectures about the behavior of other players.
But why should this be the case? Can we find scenarios that make the
assumption of correct conjectures compelling, or at least plausible? We will
discuss two types of scenarios: (1) those that make Nash equilibrium an
obvious way to play the game and (2) those that make Nash equilibrium
a steady state of an adaptive process. In each case, we will point out
that Nash equilibrium is justified under rather restrictive assumptions,
72

73
and that weakening such assumptions suggests interesting generalizations
of the Nash equilibrium concept.
More specifically, Nash equilibrium may be the obvious way to
play the game under complete information either as the outcome of
sophisticated strategic reasoning, or as the result of a non binding pre-play
agreement. If sophisticated strategic reasoning is given by rationality and
common belief in rationality, then Nash equilibrium is obtained only in
those games where rationalizable actions and equilibrium actions coincide.
This coincidence is implied by some assumptions about action sets and
payoff functions that are satisfied in several economic applications, but
in general the relevant concept under this scenario is rationalizability, not
Nash equilibrium.
What about pre-play agreements? True, a non-binding agreement
to play a particular action profile a has to be a Nash equilibrium,
otherwise the agreement would be would be self-defeating. But it will be
shown that more general probabilistic agreements that make behavior
a function of some extraneous random variables may be more efficient.
Such probabilistic agreements are called correlated equilibria and
generalize the Nash equilibrium concept.
Next we turn to adaptive processes. The general scenario is that a
given game is played recurrently by agents that each time are drawn at
random from large populations corresponding to different players/roles in
the game (e.g. seller and buyer, or male and female). The agents drawn
from each population are matched, play among themselves and then are
separated. If each players conjecture about the opponents behavior in the
current play is based on his observations of the opponents actions in the
past and this player best responds, we obtain a dynamic process in which
at each point in time some agent switches from one action to another. If
the process converges so that each agent playing in a given role ends up
choosing the same action over and over, the steady state must look like a
Nash equilibrium: indeed the experience of those playing in role i is that
the opponents keep playing some profile ai and thus they keep choosing
a best reply, say ai .1 Since this is true for each i, (ai )iI must be a Nash
equilibrium.
But it is possible that the process converges to an heterogeneous
situation: a fraction qi1 of the agents in the population i choose some
1

Suppose for simplicity that there is only one best reply.

74

5. Equilibrium

action a1i , another fraction qi2 of agents choose action a2i and so on. If
the process has stabilized (qj1 , qj2 , ...) will (approximately) represent the
observed frequencies of the actions of those playing in role j, and i will
conjecture that a1j , a2j etc. are played with probabilities qj1 , qj2 etc. If each
action a1i , a2i , ... is a best reply to such conjecture, given a bit of inertia, i
will keep choosing whatever he was choosing before and the heterogeneous
behavior within each population will persist. In this case, the steady state
is described by mixed actions whose support is made of best responses to
the mixed actions of the opponents. This is called mixed equilibrium.
It turns out that a mixed equilibrium formally is a Nash equilibrium of
the extended game where each player i chooses in the set (Ai ) of mixed
actions and payoffs are computed by taking expected values (under the
assumption of independence across players). Indeed, this is the standard
definition of mixed equilibrium.
The analysis above relies on the assumption that, when a game is
played recurrently, each agent can observe the frequency of play of different
actions. But often the information feedback is poorer. For example, one
may observe only a variable determined by the players actions, such as
the number of customers of a firm in a price setting oligopoly. In such a
context, it may happen that players conjectures are wrong and yet they
are confirmed by players information feedback, so that they keep best
responding to such wrong conjectures and the system is in a steady state
which is not a Nash (or mixed Nash) equilibrium. A steady state whereby
players best respond to confirmed (non-contradicted) conjectures is called
conjectural, or self-confirming equilibrium. Since correct conjectures
(conjectures that corresponds to the actual behavior of the other players)
are necessarily confirmed, every Nash equilibrium is necessarily selfconfirming.
To sum up, a serious effort to provide reasonable justifications for
the Nash equilibrium concept leads us to think about scenarios where
strategic interaction would plausibly lead to players best responding to
correct conjectures about the opponents. But for each such scenario we
need additional assumptions (on the payoff functions, or on information
feedback) to justify Nash play. Without such additional assumptions we
are left with interesting generalizations of the Nash equilibrium concept.
The rest of the chapter is organized as follows. We analyze Nash
equilibrium in section 5.1, focusing on symmetric equilibria (subsection

5.1. Nash equilibrium

75

5.1.1), and on interpretive issues (subsection 5.1.2). This provides the

motivations to turn to probabilistic generalizations of Nash equilibrium in
pure actions: mixed equilibrium (subsection 5.2.1), correlated equilibrium
(subsection 5.2.2) and conjectural/self-confirming equilibrium (subsection
5.2.3).

5.1

Nash equilibrium

A Nash2 equilibrium is a situation in which every player is rational and

holds correct conjectures about the actions of the other players.
Definition 15. An action profile a = (ai )iI is a Nash equilibrium if,
for every i I, ai ri (ai ).
Remark 16. An action profile a is a Nash equilibrium if and only if the
singleton {a} has the best reply property. Hence, by Theorem 2, every Nash
equilibrium is rationalizable and if there is a unique action profile which
is rationalizable, then it is necessarily the unique Nash equilibrium.
By inspection simple games, such as Matching Pennies, it is obvious
that not all games have Nash equilibria. The following classic theorem
provides sufficient conditions for the existence of at least one Nash
equilibrium.
Theorem 6. Consider a game G = hI, (Ai , ui )iI i. If, for every player
i I, Ai is a (non-empty ) compact and convex subset of a Euclidean
space, the payoff function ui is continuous and, for every ai Ai , the
function ui (, ai ) is quasi-concave, then G has a Nash equilibrium.
Sketch of proof. By the maximum principle, the best reply
correspondences ri are upper hemi-continuous with non-empty values.
2

John Nash, who has been awarded the Nobel Prize for Economics (joint with John
Harsanyi e Reinhard Selten) in 1994, was the first to give a general definition of this
equilibrium concept. Nash analyzed the equilibrium concept for the mixed extension of
a game, and proved the existence of mixed equilibria for all games with a finite number
of pure actions (see Definition 12 and Theorem 7).
Almost all others equilibrium concepts used in non-cooperative game theory can be
considered as generalizations or refinements of the equilibrium proposed by Nash.
Perhaps, this is the reason why the other equilibrium concepts that have been proposed
do not take their name after one of the researchers who introduced them in the literature.

76

5. Equilibrium

Furthermore, for every ai , ri (ai ) is convex, as Ai is convex and ui (, ai )

is quasi-concave. It follows that the correspondence r : A A as defined
by r(a) = r(a1 ) ... r(an ) maps from a compact and convex domain
into itself and it is upper hemi-continuous with non-empty and convex
values. Then, by the Kakutani Theorem, r has a fixed point, i.e. there
exists an a A such that a r(a). By definition, every fixed point of r()
is a Nash equilibrium. 

5.1.1

Equilibrium in Nice Symmetric Games

As one can see from the sketch of proof of Theorem 6, rather advanced
mathematical tools are necessary to prove a relatively general existence
theorem. Here, we analyze the more specific case in which all players
are in a symmetric position and have unidimensional action sets. This
allows us to use more elementary tools to show that, under convenient
simplifying assumptions, there exists an equilibrium in which all players
choose the same action. The fact that all players choose the same action
is not part of the thesis of Theorem 6, which did not assume symmetry.
Therefore the next theorem is not a special case of Theorem 6.3
Definition 16. A static game G = hI, (Ai , ui )iI i with |I| = n players
is symmetric if the players have the same set of actions, denoted by B
(hence, for every i I, Ai = B), and if there exists a function v : B n R
which is symmetric4 with respect to the arguments 2, ..., n and such that
i I, (a1 , ..., an ) A = B n , ui (a1 , ..., an ) = v(ai , ai ).
A Nash equilibrium a of a symmetric game G is symmetric if for all
i, j I, ai = aj .
Example 10. The Cournot oligopoly game of Example 4 is symmetric if
all n firms have the same capacity constraint and cost function: for some
a
> 0, and C : [0, a
] R, a
i = a
and Ci () = C() for all i. Then
3
However, adding symmetry to the hypotheses of Theorem 6 one can prove the
existence of a symmetric Nash equilibrium.
4
Invariant to permutations.

5.1. Nash equilibrium

77

v : [0, a
]n R is the map5

(a1 , a2 , ..., an ) 7 v(a1 , a2 , ..., an ) = a1 P a1 +

n
X

aj C(a1 ),

j=2

which is symmetric in the arguments 2, ..., n.

Recall that a game is nice if the action set of each player is a compact
interval in the real line, payoff functions are continuous, and each players
payoff function is strictly quasi-concave in his own action.
Theorem 7. Every symmetric nice game has a symmetric Nash
equilibrium.
Proof. To simplify the notation, we assume without loss of generality
that common action set is the interval B = [0, 1] (this is just a
normalization). Since ui is continuous and strictly quasi-concave in ai ,
for every (deterministic) conjecture ai there exists one and only one
best reply. Let v be as in the previous definition of a symmetric game.
In this proof, we denote by r(b) the unique best reply to the symmetric
(deterministic) conjecture (b, ..., b) [0, 1]n1 .6 In other words
r(b) = arg max v(ai , b, ..., b).
ai [0,1]

Note that, by Lemma 6, the function r() is continuous. Now, let us

introduce an auxiliary function f : [0, 1] R as follows: f (b) = b r(b).
Function f is continuous (it is a sum of continuous functions) with f (0) 0
and f (1) 0. It follows that there exists a point b [0, 1] such
that f (b ) = 0, that is, b = r(b ). Such action b is a best reply
to the symmetric conjecture (b , ..., b ) [0, 1]n1 . This means that
(b , ..., b ) [0, 1]n is a symmetric Nash equilibrium. 
The reason why we considered symmetric nice games is that they are
often used in applied theory and we could use elementary results in real
5

Note that when a function f : X Y is expressed as x 7 f (x), a different

arrow-symbol (7 instead of ) is used.
6
Before we denoted with r : A A the joint best response correspondence, a formally
different mathematical object.

78

5. Equilibrium

analysis to prove the existence of a symmetric equilibrium. A key step in

the proof is that every continuous function r : [0, 1] [0, 1] has a fixed
point b = r(b ). This is an elementary example of fixed point theorem.
More general and much harder fixed point theorems are needed to prove
existence of an equilibrium in games that are not symmetric and nice.

5.1.2

Interpretations of the Nash Equilibrium Concept

Nash equilibrium is the most well known and applied equilibrium concept
in economic theory, besides the competitive equilibrium. Indeed, we have
argued in the Introduction that, in principle, any economic situation
(and more generally any social interaction) can be represented as a noncooperative game. The property according to which every action is a best
reply to the other players actions seems to be essential in order to have
an equilibrium in a non-cooperative game.
Nonetheless, we should refrain from accepting this conclusion without
further reflection.
Why does the Nash equilibrium represent an
interesting theoretical concept? When should we expect that the actions
simultaneously chosen by the players form an equilibrium? Why should
players hold correct conjectures regarding each others behavior?
We propose a few different interpretations of the Nash equilibrium
concept, each addressing the questions above. Such interpretations can
be classified in two subgroups: (1) a Nash equilibrium represents an
obvious way to play, (2) a Nash equilibrium represents a stationary
(stable) state of an adaptive process. In some cases, we will also
introduce a corresponding generalization of the equilibrium concept which
is appropriate under the given interpretation and will be analyzed in the
next section.
(1) Equilibrium as an obvious way to play : Assume complete
information, i.e. common knowledge of the game G (this assumption is
sufficient to justify the considerations that follow, but it is not strictly
necessary). It could be the case that from common knowledge of G and
shared assumptions about behavior and beliefs, or from some prior events
that occurred before the game, the players could positively conclude that
a specific action profile a represents an obvious way to play G. If a
represents an obvious way to play, every player i expects that everybody
else chooses his action in ai . Moreover, if a is to be played by rational
players, it must be the case that no i has incentives to choose an action

5.1. Nash equilibrium

79

different from ai : if i had an incentive to choose ai 6= ai , not only ai would

not be chosen, but also the other players would have no reason to believe
that a is the obvious way to play. Hence, a must be a Nash equilibrium.
What can make a an obvious way to play?
(1.a) Deductive interpretation: The players analyze G as an interactive
decision problem look for a rational solution. If such solution exists and is
unique, then it corresponds to an obvious way to play.
As an example of solution of a game, consider the case where G has
a unique rationalizable profile a (this is the case for several models in
economic theory). Then, if all players are rational and their conjectures
are derived from the assumption that there is common common belief in
rationality, they will choose exactly the actions in a . Moreover, a is
necessarily the unique Nash equilibrium of the game (Remark 16).
(1.b) Non-binding agreement: Suppose that the players are able to
communicate before playing the game and that they reach an agreement to
choose the actions specified in the action profile a . Suppose also that such
agreement is not legally binding for the parties, it is simply a gentlemen
agreement based on players honoring their words. Further, players attach
little value to honoring their words: if i believes that a different action
yields him a higher utility, he will not choose ai . All players are perfectly
aware of this. Therefore, the agreement is credible, or self-enforcing
only if no player has an incentive to deviate from the agreement, i.e. only
if a is a Nash equilibrium.
Is it really necessary that players agree on a specific action profile?
Perhaps they could agree on making the action profile actually played (for
instance, the way they coordinate in a Battle of the Sexes game) depend
on some exogenous variable, such as the weather. In some cases, this would
allow to reach, in expectation, fairer outcomes. We will go back to this
point in the next section.
(2) Equilibrium as a stationary state of an adaptive process.7 .When
introductory economic textbooks explain why in a competitive market the
price should reach the equilibrium level that equates demand and supply,
they almost inevitably rely on informal dynamic justifications. They argue,
7
To go deeper on this topic the reader can start with the survey by Battigalli et al.
(1992). In the already mentioned paper by Migrom e Roberts (1991), the connections
between the concept of rationalizability and adaptive processes are explored.

80

5. Equilibrium

for instance, that if there is an excess demand the sellers will realize
that they are able to sell their goods for a higher price; conversely, if
there is an excess supply the sellers will lower their prices to be able to
sell the residual unsold goods. Essentially, such arguments rely more or
less explicitly on the existence of a feedback effect that pushes the prices
towards the equilibrium level. These arguments cannot be formalized in
the standard competitive equilibrium model, where market prices are taken
as parametrically given by the economic agents and then determined by the
demand=supply conditions. They nonetheless provide an intuitive support
to the theory.
Similar arguments can be used to explain why the actions played in
a game should eventually reach the equilibrium position. In some sense,
the conceptual framework provided by game theory is better suited to
formalize this kind of arguments. As argued in the Introduction, in a
game model every variable that the analyst tries to explain, i.e. any
endogenous variable, is directly or indirectly determined by the players
actions, according to precise rules that are part of the game (unlike prices
in a competitive market model, where the price-formation mechanism is
not specified). Assuming that the given game represents an interactive
situation that players face recurrently, one can formulate assumptions
regarding how players modify their actions taking into account the
outcomes of previous interactions, thus representing formally the feedback
process.
One can distinguish two different types of adaptive dynamics: learning
and evolutionary dynamics. Here, we will present only a general and brief
description.8
(2.a) Learning dynamics. Assume that a given game G is played
recurrently and that players are interested in maximizing their current
expected payoff (a reason for this may be that they do not value the future
or that they believe that their current actions do not affect in any way
future payoffs). Players have access to information on previous outcomes.
Based on such information, they modify their conjectures about their
opponents behavior in the current period. Let a be a Nash equilibrium.
If in period t every player i expects his opponents to choose ai , then ai
is one of his best replies. It is therefore possible that i chooses ai . If this
8

In Chapter 6, we present an elementary analysis of learning dynamics relating their

long-run behavior to equilibrium concepts and rationalizability.

5.1. Nash equilibrium

81

happens, what players observe at the end of period t will confirm their
conjectures, which then will remain unchanged for the following period.
So even a small inertia (that is a preference, coeteris paribus, to repeat the
previous action) will induce players to repeat in period t + 1 the previous
actions a . Analogously, a will be played also in period t + 2 and so on.
Hence, the equilibrium a is a stationary state of the process.
We just argued that every Nash equilibrium is a stationary state of
plausible learning processes (we presented the argument in an informal way,
but a formalization is possible). (i) Do such processes always converge to
a steady state? (ii) Is it true that, for any plausible learning process, every
stationary state is a Nash equilibrium? It is not possible, in general, to give
affirmative answers. First, one has to be more specific about the dynamics
of the recurrent interaction. For instance, one has to specify exactly what
players are able to observe about the outcomes of previous interactions: is
it the action profile of the other players? Is it some variables that depend
on such actions? Is it only their own payoffs? Another relevant issue is
whether game G is played always by the same agents, or in every period
agents are randomly matched with strangers. An exact specification of
these assumptions shows that the process does not always converge to a
stationary state. Moreover, as we will see in the next section, there may
be stationary states that do not satisfy the Nash equilibrium condition
of Definition 15. There are two reasons for this: (1) Players conjectures
can be confirmed by observed outcomes even if they are not correct. (2) If
players are randomly drawn from large populations, then the state variable
of the dynamic process is given by the fractions of agents in the population
that choose each action. But then it is possible that in a stationary state
two different agents, playing in the same role, choose different actions.
Even though no agent actually randomizes, such situations look like mixed
equilibria, in which players choose randomly among some of the best
replies to their probabilistic conjectures, and such conjectures happen to
be correct. We analyze notions of equilibrium corresponding to situations
(1) and (2) in the next section.
(2.b) Evolutionary dynamics. In the analysis of adaptive processes in
games, the analogy with evolutionary biology has often been exploited.
Consider, for instance, a symmetric two-person game. In every period two
agents are drawn from a large population. Then they meet, they interact,
they obtain some payoff and finally they split. Individuals are compared

82

5. Equilibrium

to animals, or plants, whose behavior is determined by a genetic code

transmitted to their offsprings. If action a is more successful than b, then
the agents programmed to play a reproduce themselves faster than those
programmed to play b, and therefore the ratio between the fraction of
agents playing a and the fraction of agents playing b increases.
In evolutionary biology this theoretical approach based on game theory
has been highly successful and has allowed to explain phenomena that
appeared as paradoxical within a more naive evolutionary framework.9
In the social sciences, evolutionary dynamics are used as a metaphor to
represent, at an aggregate level, learning phenomena such as imitation, in
which agents modify the way they play not on the basis of their direct
experiences alone, but also observing the behavior of others in the same
circumstances. Behaviors that turn out to be more successful are imitated
and spread more rapidly.
As for learning dynamics, although Nash equilibria represent stationary
states of evolutionary dynamics, the process need not always converge to
a stationary state. Furthermore, there may be stationary states that do
not satisfy the best reply property of Definition 15. Indeed, in this context
a state is represented by the fractions in the population that use different
actions. It may be the case that in a stationary state distinct agents of the
same population do different things.10
These considerations motivate the definition of more general
equilibrium concepts whereby players may hold probabilistic, rather than
deterministic conjectures about the opponents behavior. For this reason
we refer to these more general concepts as probabilistic equilibria.

5.2

Probabilistic Equilibria

Up to now we considered mixed actions only with reference to dominance

relations: if a (pure) action is dominated by a mixed action, then a rational
player should not choose it. We already noticed (Lemma 1), though, that
an expected utility maximizer has no strict incentive to choose a mixed
action. If there were a small cost for tossing a coin or spinning a roulette
wheel, then no expected utility maximizer would choose a mixed action.
9
10

See the monograph by Maynard Smith (1982).

See the monographs by Weibull (1995) and Sandholm (2010).

5.2. Probabilistic Equilibria

83

We adopt the point of view that (rational) players do not actually

choose mixed actions. Nevertheless, it is possible to reconcile this point
of view with an interpretation of mixed actions that makes them relevant
independently of the equivalence results stated in Lemma 2 and Theorem
3.

5.2.1

Mixed Equilibria

Consider the following game between home owners and thieves. Each
owner has an apartment where he stores goods worth a total value of V .
Owners have the option to keep an alarm system, which costs c < V .11 The
alarm system is not detectable by thieves. Each thief can decide whether to
attempt a theft (burglary) or not. If there is no alarm system, the thieves
successfully seize the goods and resell them to some dealer, making a profit
of V /2. If there is an alarm system, the attempted burglary is detected
and the police is automatically alerted. The thieves in such event need to
leave all goods in place and try to escape. The probability they get caught
is 12 and in this case they are sanctioned with a monetary fine P and then
released.12
If the situation just described were a simple game between one owner
and one thief, it could be represented using the matrix below:
O\T

Burglary

No

Alarm

V c, P2

V c, 0

No

0, V2

V, 0

Figure 5.1: Matrix 2

It is easy to check that Matrix 2 has no equilibria according to

Definition 15.
However, the game between owners and thieves is more complex. There
are two large populations, and we assume for simplicity that they have the
11

We assume for simplicity that c represents the cost of installing an alarm system as
well as the cost of keeping it active. Furthermore, we neglect the possibility to insure
against theft.
12
Prisons do not exist. If thieves cannot pay the amount P , they are inflicted an
equivalent punishment.

84

5. Equilibrium

same size: the population of owners (each one with an apartment) and
the population of thieves. Thieves randomly distribute themselves across
apartments. For any given owner, the probability of an attempted burglary
is equal to the fraction of thieves that decide to attempt a burglary. From
the thieves perspective, the probability that an apartment has an alarm
system is given by the overall fraction of owners keeping an alarm system.
Assume that the game is played recurrently. The fractions of agents
that choose the different actions evolve according to some adaptive process
with the following features. At the end of each period it is possible to
access (reading them on the newspapers) the statistics of the numbers
of successful and unsuccessful burglaries. Players are fairly inert in that
they tend to replicate the actions chosen in the previous period. However,
they occasionally decide to revise their choices on the basis of the previous
period statistics. Since in each period only a few agents revise their choices,
such statistics change slowly.
An owner not equipped with an alarm system decides to install it if and
only if the expected benefit is larger than the cost, i.e. if the proportion of
attempted burglaries is larger than c/V . Conversely, an owner equipped
with an alarm system will decide to get rid of it if and only if the proportion
of attempted burglaries is lower than c/V . The proportion of attempted
burglaries changes only slowly and the owners equate the probability of
being robbed today with the one of the previous period. The probability
that makes an owner indifferent between his two actions is c/V . When
indifferent, an owner sticks to the previous period action.
Analogously, a thief that was active in the previous period decides not
to attempt a burglary in the current period if and only if the fraction of
apartments equipped with an alarm system (which is also the fraction of
unsuccessful burglaries) is larger than V /(V + P ). A thief that was not
active in the previous period attempts a burglary in the current one if and
only if the fraction lower than V /(V + P ).
The state variables of this process are the fraction of installed alarm
systems and the fraction of attempted burglaries. It is not hard to see
that grows (resp., decreases) if and only if > Vc (resp. < c/V ).
Similarly, grows (resp., decreases) if and only if < V /(V + P ) (resp.,
> V /(V + P )). If = V /(V + P ) and = c/V, then the state of
the system does not change, i.e. (, ) = ( V V+P , Vc ) is a stationary state,
or rest point, of the dynamic process. Whether this rest point is stable

5.2. Probabilistic Equilibria

85

depends on details that we have not specified. The mixed action pair
(1 , 2 ) = (V /(V + P ), c/V ) is said to be a mixed (Nash) equilibrium of
the matrix game above.
In this example, we have interpreted the mixed action of j, say j , both
as a statistical distribution of actions in population j and as a conjecture
of the agents of population i about the action of the opponent. In a
stable environment every conjecture about j, j , is correct in the sense
that it corresponds to the statistical distribution of actions in population
j. Furthermore, j is such that every agent in population i is indifferent
among the actions that are chosen by a positive fraction of agents.
Next, we present a general definition of mixed equilibrium and then
show that it is characterized by the aforementioned properties.
Definition 17. The mixed extension of a game G = hI, (Ai , ui )iI i is
a game G = hI, ((Ai ), ui )iI i where
X
Y
ui () =
ui (a1 , ..., an )
j (aj ),
(a1 ,...,an )A

jI

for all i and = (1 , ..., n ) (A1 ) ... (An ).

Note, the payoff functions of the mixed extension are obtained
calculating the expected payoff corresponding to a vector of mixed actions
under the assumption that the actions of different players are statistically
independent.
Definition 18. (see Nash, 1950) A mixed action profile is a mixed
equilibrium of game G if is a Nash equilibrium of the mixed extension
of G.
Note that every (pure) Nash equilibrium (ai )iI corresponds to a mixed
equilibrium (i )iI such that each i assigns probability one to ai : mixed
equilibrium is a generalization of (pure) Nash equilibrium.
Theorem 8. Every finite game has at least one mixed equilibrium.13
Proof. It is easy to verify that the mixed extension G of a finite game
G satisfies all the assumptions of Theorem 6: for each i, (Ai ) is a nonempty, convex and compact subset of R|Ai | , ui is a multi-linear function and
13

The same holds for any compact and continuous game.

86

5. Equilibrium

therefore it is both continuous in and weakly concave, indeed linear,14

in i . 
When we introduced mixed actions in Chapter 3, we said that, even
if we do not assume that players randomize, mixed actions still play
an important technical role. The space of mixed actions is convex and
expected payoff is multi-linear in the mixed actions of different players. In
Chapter 3, such convexity-linearity structure was exploited to characterize
the set of justifiable (pure) actions (Lemma 2). In Theorem 8, the
convexification-linearization allowed by the mixed extension of a game
yields existence of an equilibrium.
The following result introduces an alternative characterization of mixed
equilibria, which is not based on the mixed extension of the game. With a
small abuse of notation, we let ri (i ) denote the set of best replies to the
conjecture i (Ai ) obtained as a product of the marginal measures
j (j 6= i), that is, ri (i ) := ri (i ) where
i j6=i (Aj ), ai Ai , i (ai ) :=

j (aj ).

j6=i

Theorem 9. A mixed action profile (1 , ..., n ) is a mixed equilibrium if

and only if, for every i I, Suppi ri (i ) (every action played with
positive probability is a best reply to the mixed action profile played by the
opponents).
Proof The statement follows directly from the definition of mixed
equilibrium and from Lemma 1. 
Corollary 2. All actions played with positive probability in a mixed
equilibrium are rationalizable.
Proof. By Theorem 9, the set C =Supp1 ... Suppn has the best
reply property. Hence, Theorem 2 implies that every action played with
positive probability is rationalizable. 
14

More precisely, for each i , the section of u

i at i , u
i,i : (Ai ) R, is affine:
i , i (Ai ), [0, 1],
u
i,i (i + (1 )i ) =
ui,i (i ) + (1 )
ui,i (i ).

5.2. Probabilistic Equilibria

87

Theorem 8 and Corollary 2 provide a linear programming algorithm to

compute all mixed equilibria of finite two-person games. Let [uk`
i ] be the
payoff matrix of player i, with generic entry uk`
.
Player
1
chooses
the rows
i
(indexed by k) and player 2 the columns (indexed by `).
Step 1: Eliminate all iteratively dominated actions (by Theorem 3 and
Corollary 2 such actions are played with zero probability in equilibrium).
The order of elimination is irrelevant (Theorem 4).
Step 2: For any pair of non-empty subsets A1 A1 and A2 A2 ,
compute the set of mixed equilibria, (1 , 2 ), such that Supp1 = A1
and Supp2 = A2 . This set, which could be empty, is computed as
follows (we consider the non trivial case in which the sets contain at least
two actions). To simplify the notation, assume that A1 = {1, ..., m1 }
and A2 = {1, ..., m2 } and denote by im the generic probability that
action m of player i is played. Solve the following systems of linear
m
equations and inequalities with unknown 1 = (11 , ..., 1 1 ) (A1 ) and
m
2 = (2 , ..., 2 2 ) (A2 ):

m1
X

k
uk`
2 1

k
uk`
2 1

`
uk`
1 2

uk1
2 1 , ` = m2 + 1, ..., |A2 |;

m2
X

u1`
1 2 , k = 2, ..., m1 ,

(5.2.2)

`=1
m2

`=1
m2
`=1

(5.2.1)

k=1

k=1

m2
X

uk1
2 1 , ` = 2, ..., m2 ,

k=1
m1

k=1
m1

m1
X

`
uk`
1 2

u1`
1 2 , k = m1 + 1, ..., |A1 |.

`=1

The subset of equations in (5.2.1) determines the set of mixed actions

of player 1 that make player 2 indifferent between the actions in subset
A2 . The subset of inequalities determines the set of mixed actions of
player 1 that make the action a2 = 1 (and so all the actions in A2 ) weakly
preferred to the actions that do not belong to A2 . For any 1 that satisfies
the system (5.2.1), player 2 has no incentive to deviate from a mixed
action with support A2 . Similar considerations hold for system (5.2.2).

88

5. Equilibrium

Such system determines the set of mixed actions of player 2 that make
player 1 indifferent among all the actions in A1 and at the same time make
such actions weakly preferred to all the others. Therefore, the indifference
conditions for player 1 determine the equilibrium randomization(s) of
player 2, the indifference conditions for player 2 determine the equilibrium
randomization(s) of player 1.
In the previous example (Owners and Thieves) the equilibrium is
determined as follows: First, note that the best reply to a deterministic
conjecture is unique. However, no pair of pure actions is an equilibrium.
Hence, the equilibrium is necessarily mixed (existence follows from
Theorem 8). An equilibrium with support A1 = {Alarm, N o}, A2 =
{Robbery, N o} is given by the following (we are writing = 1 (A) and
= 2 (B)):
Indifference condition for i = 1 (Owner):
V c = V (1 ).
Indifference condition for i = 2 (Thief ):
V
P
(1 ) = 0.
2
2
Solving the system, = V V+P , = Vc , which is the stationary state
identified by the (informal) analysis of a plausible dynamic process.

5.2.2

Correlated Equilibria

Mixed equilibria can be interpreted as statistical distributions over action

profiles that are in some sense stationary. Such distributions are obtained
as a product of the marginal distributions corresponding to the equilibrium
mixed actions: if (1 , ..., n ) isQa mixed equilibrium, then the probability

of any profile (a1 , ..., an ) is

iI i (ai ); this means that actions are
statistically independent across players.
Statistical independence is
justified by the interpretation of mixed equilibrium as a stationary vector
of statistical distributions of actions in n populations whose agents are
randomly matched with the agents of the other populations to play the
game G.
Now we present a different concept of probabilistic equilibrium
that, formally, generalizes the concept of mixed equilibrium by allowing

5.2. Probabilistic Equilibria

89

correlated distributions over players action profiles. The interpretation

of this equilibrium concept given here is rather different from the one
proposed for the mixed equilibrium.15
Let us consider once again the case in which players are able to
communicate and sign a non binding agreement before the game is played.
If this agreement simply prescribes to play a certain profile a , then
as noted earlier a must be a Nash equilibrium, otherwise it would
not be self-enforcing. However, players could reach more sophisticated
self-enforcing agreements, in which the chosen actions depend on the
realizations of extraneous random variables that do not directly affect the
payoffs. We first illustrate this idea with a simple example.
B

3, 1

0, 0

0, 0

1, 3

Figure 5.2: Matrix 3.

Matrix 3 represents the classic Battle of the Sexes (BoS): Rowena (row
player) and Colin (column player) would like to coordinate and go to the
same concert, either B ach or S travinsky, but Rowena prefers Bach and
Colin prefers Stravinsky. Suppose that Rowena and Colin have to agree on
how to play the BoS the next day, when no further communication between
them will be possible. There exists two simple self-enforcing agreements,
(B, B) and (S, S). However, the first favors Rowena, the second favors
Colin, and neither player wants to give up. How to sort this out? Colin
can make the following proposal that would ensure in expected values a
fair distributions of the gains from coordination: If tomorrows weather is
bad then both of us choose B ach, if instead it is sunny then we both choose
S travinsky. Notice that the weather forecasts are uncertain: there is a
50% probability of bad weather and 50% of sunny weather. The agreement
generates an expected payoff of 2 for both players. Rowena understands
that the idea is smart. Indeed, the agreement is self-enforcing as both
players have an incentive to respect it if they expect the other to do the
same. For instance, if Rowena expects that Colin sticks to the agreement
and waking up she observes that the weather is bad, then she expects that
15

Proposition 1 of Chapter 6 offers an adaptive interpretation.

90

5. Equilibrium

Colin will go to the B ach concert and she wants to play B. Similarly, if
she observes that the weather is sunny, she expects that Colin will go to
the S travinsky concert and she wants to play S.
In other words, a sophisticated agreement can use exogenous and
not directly relevant random variables to coordinate players beliefs and
behavior. In such an agreement the conjecture of a player about the
behavior of others (and so their best replies) depends on the observed
realization of such random variables and the actions of different players
are correlated, albeit spuriously.
Clearly, it is not always possible to condition the choice on some
commonly observable random variable. Furthermore, even if this were
possible, players could still find it more convenient to base their respective
choices on random variables that are only partially correlated. Consider,
for example, the following bi-matrix game:
a

6, 6

2, 7

7, 2

0, 0

Figure 5.3: Matrix 4.

If Rowena and Colin rely on a commonly observed random variable

to design a self-enforcing agreement, they can only achieve a probability
distribution over pure Nash equilibria. Actually this is true in every game:
suppose that the agreement says if x is (commonly) observed, each j I
must take action j (x), then when i observes x he infers that each coplayer j will take action j (x), and in order to make the agreement
self-enforcing i (x) must be a best reply to i (x), thus (i (x))iI
must be a Nash equilibrium.16 The two pure equilibria of the game in
Matrix 4 yield the payoff pairs (7, 2) and (2, 7). So, the sum of the
expected payoffs attainable with self-enforcing agreements that rely on
a commonly observed random variable is (2 + 7) + (1 ) (7 + 2) = 9,
where is the probability of observing a realization x that yields (a, b)
( = P{x : 1 (x) = a, 2 (x) = b}).
If instead Rowena and Colin rely on different (but correlated) random
16

One can generalize and obtain distributions over mixed equilibria if players choose
mixed actions according to the observed realization x.

5.2. Probabilistic Equilibria

91

variables, they can do better. For example, suppose that Rowena observes
whether X = a or X = b, and Colin observes whether Y = a or Y = b,
where (X, Y ) is a pair of random variables with following joint distribution:
X\Y

1/3

1/3

1/3

They agree that each one should play a when she or he observes a,
and b when she or he observes b. It can be checked that this agreement
is self-enforcing. Suppose that Rowena observes X = a, then she assigns
the same (conditional) probability,

1
2

1
3
1
+ 13
3

, to actions a and b by Colin;

thus, the expected payoff of choosing a is 4, while the expected payoff of

choosing b is only 3.5. If she observes X = b, then she is sure that Colin
observes Y = a and plays a; so she best responds with b. A symmetric
argument applies to Colin. Therefore no player wants to deviate, whatever
she or he observes. The sum of the expected payoffs in this agreement is
1
1
1
3 (6 + 6) + 3 (2 + 7) + 3 (7 + 2) = 10.
This rather informal discussion motivates the following.
Definition 19. (see [1] and [2]) A probabilistic self-enforcing agreement,
or correlated equilibrium, is a structure h, p, (Ti , i , i )iI i, where
(, p) is a finite probability space (p ()), the sets Ti are finite17 and
the functions i : Ti and i : Ti Ai are such that,
i I, ti Ti , ai Ai , if p
X

p(|ti )ui (i (ti ), i (i ()))

0 : i ( 0 ) = ti

> 0 then

p(|ti )ui (ai , i (i ())) .

In the above definition i represents the random variable, also called

signal, observed by player i, whereas i represents the strategy, or
17

Finiteness is just a simplification without substantial loss of generality when A is

finite. If and Ti are infinite, they must be endowed with sigma-algebras of events,
probability measure p is defined on the sigma-algebra of , and all the functions i and
i must be measurable.

92

5. Equilibrium

decision function, imposed by the agreement on player i;


p()/p ({ 0 : i ( 0 ) = ti }) , if i () = ti
p(|ti ) =
0,
if i () 6= ti
is the probability of state conditional on observing ti (which is well
defined if the probability of ti is positive)18 and
i (i ()) := (1 (1 ()), ..., i1 (i1 ()), i+1 (i+1 ()), ..., n (n ()))
is the action profile that would be chosen by the other players (according
to the agreement) if state occurs.
The definition requires that players have no incentive to deviate from
the behavior prescribed by the agreement, whatever the realization of
their signals. Thus, conditional on every possible observation, the expected
payoff of a deviation cannot be higher than the expected payoff obtained by
sticking to the agreement. The inequalities expressing these conditions are
sometimes called incentive compatibility constraints.
Clearly, a correlated equilibrium induces a probability measure over
players action profiles, (A), where
a A, (a) = p ({ : i I, i (i ()) = ai }) .

(5.2.3)

Formally, measure (A) is the image of measure p () through

the pushforward given by the composite function : A, that is,
= p ( )1 , where = (i )iI : T and = (i )iI : T A;
more explicitly,
X
(E) = p(( )1 (E)) =
p()
:( ())E

for each subset of action profiles E A. To analyze the properties of

such induced probability measures we introduce the notion of canonical
correlated equilibrium:
18

This is the special case of the standard definition of conditional probability:

p(F ) > 0 p(E|F ) =

p(E F )
.
p(F )

If E = {} (a singleton) and F = { 0 : i ( 0 ) = ti }, then the formula in the main text

obtains.

5.2. Probabilistic Equilibria

93

Definition 20. A self-enforcing probabilistic agreement h0 , p0 , (i0 , i0 )iI i

where
(1) 0 = A,
(2) p0 (A),
(3) i I, (a1 , ..., an ) A, i0 (a1 , ..., an ) = ai (i0 = projAi is the
projection function from A onto Ai ),
(4) i I, ai Ai , i0 (ai ) = ai (i0 = IdAi is the identity function on Ai ),
is called canonical correlated equilibrium.
Although, formally, it is the whole structure h0 , p0 , (i0 , i0 )iI i that
forms a canonical correlated equilibrium, it is standard to call canonical
correlated equilibrium just the distribution p0 (A), as all the other
elements of the structure are trivially determined by the given game G.
Theorem 10. If h, p, (i , i )iI i is a correlated equilibrium, then p0 = ,
where satisfies (5.2.3), is a canonical correlated equilibrium distribution.
A canonical correlated equilibrium with distribution can be
interpreted as the following mechanism: an action profile a
is chosen at
random according to the agreed upon probability measure (A),
then a mediator observes a = (ai )iI and privately suggests to each
player i to play action a
i in the realized profile. Then player i chooses
freely his action a0i Ai , but he has no incentive to deviate from the
suggested action ai .
Proof of Theorem 10. Let h, p, (Ti , i , i )iI i be a correlated
equilibrium; fix arbitrarily a player i and an action ai with strictly positive
marginal probability:
X
(ai , ai ) = p ({ : i (i ()) = ai }) > 0,
(ai ) := (margAi )(ai ) =
ai Ai

where the second equality follows from the definition of marginal

probability, and the third one from eq. (5.2.3). We must show that i
has no incentive to deviate from the mediators suggestion ai , that is
X


a0i ,
(ai |ai ) ui (ai , ai ) ui (a0i , ai ) 0,
ai Ai

where
(ai |ai ) := (margAi (|ai ))(ai ) =

p ({ : ( ()) = (ai , ai )})

.
p ({ : i (i ()) = ai })

94

5. Equilibrium

Since (ai ) > 0, the previous system of inequalities is equivalent to

X


a0i , (ai )
(ai |ai ) ui (ai , ai ) ui (a0i , ai ) 0.
ai Ai

Since (ai , ai ) = (ai |ai )(ai ), the incentive compatibility constraints

for can be expressed as
X


a0i ,
(ai , ai ) ui (ai , ai ) ui (a0i , ai ) 0.
(5.2.4)
ai Ai

Hence it is sufficient to show that (5.2.4) holds. This system of

inequalities will be derived from the incentive compatibility constraints of
the correlated equilibrium h, p, (Ti , i , i )iI i that induces , thus proving
the result.
For all ti i1 (ai ) such that p ({ : i () = ti }) > 0 (there must be
at least one ti like this), the incentive compatibility constraints yield
X


a0i ,
p(|ti )[ui (ai , i (i ())) ui a0i , i (i ()) ] 0.
:i ()=ti

Multiplying each inequality by p ({ : i () = ti }), taking into account

that p(|ti )p ({ : i () = ti }) = p() for each i1 (ti ), and taking
the summation w.r.t. the ti s in i1 (ai ) yields
X


a0i ,
p()[ui (ai , i (i ())) ui a0i , i (i ()) ] 0.
:i (i ())=ai

(5.2.5)
Since ui depends only on actions, the terms in (5.2.5) can be regrouped
to obtain
X
X


p()[ui (ai , ai ) ui a0i , ai ] (5.2.6)
a0i ,
ai ( )1 (ai ,ai )

X


=
[ui (ai , ai ) ui a0i , ai ]
ai

p() 0,

( )1 (ai ,ai )

where = (j )jI : T A, = (j )jI : T , : A is the

composition of and , and ( )1 (ai , ai ) is the set of pre-images in
of action profile (ai , ai
P).
By definition of ,
( )1 (ai ,ai ) p() = (ai , ai ). Therefore
(5.2.6) yields (5.2.4) as desired. 
The first part of the proof yields the following:

5.2. Probabilistic Equilibria

95

Remark 17. A distribution (A) is induced by a correlated

equilibrium (hence, by Theorem 10, it is a canonical correlated equilibrium)
if and only if it satisfies the following system of linear inequalities in the
probabilities ((a))aA :
X


i I, ai , a0i Ai ,
(ai , ai ) ui (ai , ai ) ui (a0i , ai ) 0.
ai Ai

Hence, the set of measures (A) induced by a correlated equilibrium

(the set of canonical correlated equilibria) is a convex and compact
polytope.19 The same holds for the set of expected payoff profiles induced
by correlated equilibria.
The following observation, the proof of which is elementary, clarifies
that the correlated equilibrium concept generalizes the mixed equilibrium
concept:
Remark 18. Fix a mixed action
profile and let (A) be the
Q
measure defined by (a) = iI i (ai ) for every a A. Then is a
mixed equilibrium if and only if is a canonical correlated equilibrium.20
Like pure and mixed Nash equilibrium, also correlated equilibrium is a
refinement of rationalizability:
Theorem 11. If is a canonical correlated equilibrium, then every action
to which assigns a positive marginal probability is rationalizable.
n
o
P
Proof. For every i, let Ci = ai : ai (ai , ai ) > 0 be the set of is
actions to which assigns a positive marginal probability. It will be shown
that C = C1 ... Cn is a set with the best reply property. By Theorem
2, this implies that every action ai Ci is rationalizable, as desired. We
write marginal and conditional probabilities using an obvious and natural
notation: (ai ), (ai |ai ), (aj |ai ).
For every i and ai , if ai Ci , then (by definition of Ci ) (ai ) > 0, and
the conditional probability (|ai ) (Ai ) is well defined. Also, for every
19

A polytope in R2 is just a polygon, a polytope in R3 is a polyhedron. Polytope

is the generalization of these geometrical objects to Rn .
20
The word correlated here may be a bit misleading, as we are considering the case
where players actions are mutually independent. Nonetheless, this is a special case of
the general definition of correlated equilibrium.

96

5. Equilibrium

j 6= i and aj , (aj |ai ) > 0 only if (aj ) > 0. Hence, Supp(|ai ) Ci .

Finally, given that is a canonical correlated equilibrium it must be the
case that
X
X
a0i Ai ,
(ai |ai )ui (ai , ai )
(ai |ai )ui (a0i , ai ),
ai Ai

ai Ai

that is, ai ri ((|ai )). Since this is true for each player i, it follows that
C has the best reply property. 
The concept of correlated equilibrium can be generalized assuming
that different players can assign different probabilities to the states
. This generalization is called subjective correlated equilibrium
(Brandenburger e Dekel, 1987). It can be shown that an action ai is
rationalizable if and only if there exists a subjective correlated equilibrium
in which ai is chosen in at least one state (see section 7.3.5 of Chapter
7).

5.2.3

Selfconfirming Equilibrium

We conclude this section on probabilistic equilibria introducing another

generalization of the Nash equilibrium concept. In subsection 5.1.2 we
mentioned the possibility of interpreting an equilibrium as the stationary
state of a learning process and we noted that such stationary states need
not satisfy the Nash property. Consider the following simple example.
1\2

2, 0

2, 1

0, 0

3, 1

Figure 5.4: Matrix 5.

The game in Matrix 5 has a unique rationalizable outcome, and so a

unique Nash equilibrium (and a unique degenerate mixed equilibrium that
coincides with the pure strategy equilibrium). If Rowena (player 1) knew
the payoff function of Colin (player 2) and were to believe that Colin is
rational, then she would expect him to play action r and would play the
best reply b.

5.2. Probabilistic Equilibria

97

Instead, we assume the following: (i) information may be incomplete,

each player knows his payoff function, but does not necessarily know the
opponents payoff function; (ii) the game is played recurrently and, at the
end of each round, each player observes his realized payoff, but he does
not observe directly the opponents action; (iii) in every period players
form probabilistic conjectures on the opponents actions that are revised
according to previous observations; consistently with Bayesian updating,
when players observe something that was expected with probability one
they do not revise their conjectures. Also, we assume for simplicity that
(iv) players maximize the current expected payoff without worrying about
the impact of current choices on future payoffs.
The situation of Colin is very simple: he will always chooses his
dominant action r. Conversely, the situation of Rowena is not so simple.
If in any given period t she expects ` with probability larger than 13 ,
then she best responds with t. In this case Rowenas payoff is equal
to 2 independently of the choice made by Colin, hence she not able to
infer Colins choice from observing the realization u1 = 2, rather she
observes something that she was expecting with certainty (to obtain a
payoff of 2) and thus she does not revise her probabilistic conjecture.
Hence also in period t + 1 she repeats the same choice t and according
to the same reasoning her conjecture remains unaffected. This shows that
if the starting conjecture of Rowena assigns a sufficiently high probability
to ` the process is stuck in (t, r), which is therefore a stationary state, even
if it is not an equilibrium in the sense of Definitions 15 (Nash equilibrium)
or 18 (mixed equilibrium).21
The example just described illustrates a situation in which players best
respond to their conjectures and the information obtained ex post does
not induce them to change their conjectures, even if they are incorrect.
Situations of this kind are known as self-confirming, or conjectural,
equilibria.22
21

One could object to that if Rowena is patient, then she will want to experiment with
action b so as to be able to observe (indirectly) Colins behavior, even if that implies an
expected loss in the current period. The objection is only partially valid. It can be shown
that for any degree of patience (discount factor) there exists a set of initial conjectures
that induce Rowena to choose always the safe action a rather than experimenting with
b. It is true, however, that the more Rowena values future payoffs, the less is it plausible
that the process will be stuck in (t, r).
22
For an analysis of how this concept has originated and his relevance for the analysis

98

5. Equilibrium

Self-confirming Equilibria with Pure Actions

As it is apparent from the previous example, in order to verify whether a
certain situation is a self-confirming equilibrium, it is necessary to specify
what players are able to observe ex post. We represent this information
with a feedback function, fi : A Mi , where Mi is a set of messages
that i could receive at the end of each game period. Let fi,ai : Ai Mi
denote the section of fi at ai . Assuming that i remembers his choice,
what i can infer when at the end of the period he receives message
mi after having played ai is that the opponents must have played an
1
action profile from the set fi,a
(mi ) = {ai : fi (ai , ai ) = mi } (In the
i
previous example, f1 (.) = u1 (.), M1 = {0, 2, 3}, mi means You got
mi euros, {a2 : f1 (t, a2 ) = 2} = {`, r}, {a2 : f1 (b, a2 ) = 0} = {`},
{a2 : f1 (b, a2 ) = 3} = {r}). Suppose that is conjecture is i , i chooses the
best reply ai ri (i ) and then observes mi ; suppose also that i assigns
1

probability one to the set fi,a

(mi ) = {ai : fi (ai , ai ) = mi }; then the
i
conjecture of i is confirmed (he observes what he expected with probability
one), he sticks to it and keeps choosing ai .
This preliminary discussion shows that in order to ascertain the
stability of behavior under learning from personal experience, we must
enrich the standard definition of game with an ex post information
structure that specifies what players can learn after that a round of a
recurrent game has been played. We do so by adding to the static game
(in reduced form) G = hI, (Ai , ui )iI i a profile of feedback functions
f = (fi : A Mi )iI . We call the pair (G, f ) game with feedback.
With this we can define the self-confirming equilibrium concept (also called
conjectural equilibrium ), which captures with a static definition the
possible steady states of learning dynamics.23 We start with equilibria in
pure actions, which capture stability in a game recurrently played by the
same set of individuals. Then we will generalize to distributions of pure
actions.
Definition 21. Fix a game with feedback (G, f ). A profile of actions
and conjectures (ai , i )iI iI (Ai (Ai )) is a self-confirming
of adaptive processes in a repeated iteration context see the survey by Battigalli et al.
(1992) and the next chapter.
23
For brief reviews of the self-confirming/conjectural equilibrium concept see survey
on learning in games by Battigalli et al. (1992) and the discussion of the literature in
Battigalli et al. (2011).

5.2. Probabilistic Equilibria

99

equilibrium of (G, f ) if, for every player i I, the following conditions

hold:
(1) ( rationality) ai ri (i ), 

1

(2) ( confirmed conjectures) i fi,a

= 1.
(fi (a )
i

(ai )iI is a self-confirming equilibrium action profile of (G, f ) if there is

some profile of conjectures (i )iI such that (ai , i )iI is a self-confirming
equilibrium of (G, f ).
Remark 19. Clearly, every Nash equilibrium a is a self-confirming
equilibrium profile for every profile of feedback functions f (let i (ai ) = 1
for each i). If the section of fi at ai , fi,ai : Ai Mi is one-toone (injective) for every ai Ai , then the sets of Nash equilibrium and
self-confirming equilibrium profiles coincide.
It can be checked that in the game of Matrix 5 the action profile (t, r)
is a self-confirming equilibrium if each player observes only his own payoff
(fi = ui , i = 1, 2): pick any (1 , 2 ) with 1 (`) 13 and note that t
r1 (1 ), {a2 : u1 (t, a2 ) = u1 (t, r)} = A2 , {a1 : u2 (a1 , r) = u2 (t, r)} = A1 .
Self-confirming
Equilibrium with Mixed Actions: the Anonymous Interaction
Case
The Nash equilibrium concept has been generalized to allow for
randomization, thus obtaining the mixed equilibrium concept. We
motivated mixed equilibrium as a characterization of stationary patterns of
behavior in situations of recurrent anonymous interaction, i.e. situations
where a game is played recurrently and every role (e.g. the role of the
row player) is played by agents drawn at random from large populations.
Also the self-confirming equilibrium concept be generalized to allow for
randomization in order to capture stable patterns of behavior in situations
of recurrent anonymous interaction. Again, we will interpret a mixed
action i (Ai ) as representing the fractions of agents in population
i playing the different actions ai Ai . Since different agents in population
i may hold different conjectures, the extension should allow for some
heterogeneity: in a stationary state different agents of the same population
may choose different actions that are justified by different conjectures.
We will assume for the sake of simplicity that all the agents playing the

100

5. Equilibrium

same action hold the same conjecture. This restriction is without loss of
generality.
Consider an agent from population i who holds conjecture i and keeps
playing a best reply ai ri (i ). In the case of anonymous interaction, what
an agent playing in role i observes at the end of the game depends on the
behavior of the agents playing in the other roles, but these agents are drawn
at random, therefore each message mi will be observed, in the long run,
with a frequency determined by the fraction of agents in population(s) i
choosing the actions that (together with ai ) yield message mi . Conjecture
i is confirmed if the subjective probabilities that i assigns to each
1
message mi given ai , that is (i (fi,a
(mi ))mi Mi , coincide with the observed
i
long-run frequencies of these messages.
The following example clarifies this point:
1\2

2, 3

2, 1

0, 0

3, 1

Figure 5.5: Matrix 6.

Assume that 25% of the column agents choose `, so that 75% choose r.
Then the probability that a row agent receives the message You got zero
euros when she chooses b is 14 . Row agents do not know this fraction
and could, for instance, believe that ` happens with probability 12 , which
would induce them to choose t; alternatively a probability of 51 , would
induce them to play b. It may happen, for instance, that half of the row
agents believe that P(`) = 12 and the other half believe P(`) = 15 . The
former ones choose t, the latter ones b. If the fractions of column agents
playing ` and r remain 25% and 75% respectively, then those who play b
will observe Zero euros 25% of the times and Three euros 75% of the
times. They will then realize that their beliefs were not correct and will
keep on revising them until eventually they will coincide with the objective
proportions, 25%:75%. These row agents will continue to play b. The other
half of the row agents, those that believe that P(`) = 21 and choose t, do
not observe anything new and therefore keep on believing and doing the
same things.
But is it possible that the fractions of column agents playing ` and

5.2. Probabilistic Equilibria

101

r stay constant? Indeed it is. Suppose that 25% of them believe that
P(t) = 25 and the rest believe that P(t) = 51 . The former ones will choose
` (expected payoff 65 > 1), the latter ones will choose r (as 1 > 35 ). Those
choosing r do not receive any new information and keep on doing the same
thing. Those choosing ` half of the times observe Three euros, and the
other half Zero Euros. Their conjectures are not confirmed, but they
will keep revising upward the probability of t and best replying with ` ,
until their conjectures converge to the long-run frequencies 50%-50%, i.e.
the actual proportions of row agents choosing t and b.
Then there is a stable situation characterized by the following fractions,
or mixed actions: 1 (t) = 21 , 2 (`) = 41 . These fractions do not form
a mixed equilibrium: in fact, the indifference conditions for a mixed
equilibrium yield 1 (t) = 13 , 2 (`) = 13 .
It should be clear from the discussion that here, as in our interpretation
of the mixed equilibrium concept, mixed actions are interpreted as
statistical distributions of pure actions in populations of agents playing
in the same role.
Let us move to a general definition. Denote by Pfaii ,i [mi ] the probability
of message mi determined by action ai and conjecture i , given the
feedback function fi :24
X
Pfaii ,i [mi ] =
i (ai ).
ai :fi (ai ,ai )=mi

Similarly, Pfaii, i [mi ] is the probability of mi determined by ai and the

mixed action profile i , given fi :
Y
X
Pfaii ,i [mi ] =
j (aj ).
ai :fi (ai ,ai )=mi j6=i

In other words, Pfaii ,i [mi ] and Pfaii ,i [mi ] are, respectively, the pushforward
of i and i through fi,ai : Ai Mi , the section of is feedback function
at ai :
1
1
Pfaii ,i [] = i fi,a
, Pfaii ,i [] = i fi,a
.
i
i
24

In general, we write P [E|C] for the probability of E given C, computed with

probability measure on some underlying state space . If I want to emphasize that
E is obtained via some function : X, e.g. E = [x] = 1 (x), then we write
P
[E|C].

102

5. Equilibrium

Definition 22. Fix a game with feedback (G, f ). A profile of mixed actions
and conjectures (i , (iai )ai Suppi )iI is an anonymous self-confirming
(or conjectural) equilibrium of (G, f ) if for every role i I and action
ai Suppi the following conditions hold:
(1) ( rationality) ai ri (iai ),
(2) ( confirmed conjectures) mi Mi , Pfaii ,i [mi ] = Pfaii ,i [mi ].
ai

= (i )iI is an anonymous self-confirming equilibrium profile if there is

a profile of conjectures ((iai )ai Suppi )iI such that (i , (iai )ai Ai )iI is
an anonymous self-confirming equilibrium.
Remark 20. Every mixed equilibrium of G is an anonymous selfconfirming equilibrium profile for every f .
Remark 21. Every pure self-confirming equilibrium (ai )iI is a degenerate
anonymous self-confirming equilibrium, which can be interpreted as a
symmetric equilibrium where all agents of the same population i play the
same action ai .
It can be checked as an exercise that the game in Matrix 6, assuming
that the players observe ex post only their own payoff (fi = ui ), admits
three types of anonymous self-confirming equilibrium:
1. t is chosen by all the agents in population 1: 2 can take any value
since the conjecture of 1 is necessarily confirmed (1, choosing t,
does not receive new information); 1 (t) = 1, 1t (`) 31 , 0 2 (`)
1, 2` (t) = 1, 2r (t) 13 ;
2. r is chosen by all agents in population 2: 1 can take any value since
the conjecture of 2 is necessarily confirmed (2, choosing r, does
not receive new information); 0 1 (t) 1, 1t (`) 31 , 1b (`) = 0,
2 (`) = 0, 2r (t) 31 ;
3. all actions are chosen by a positive fractions of agents: the beliefs of
those who choose informative actions (b for i = 1 and ` for i = 2)
are correct; 13 1 (t) 1, 0 < 2 (`) 31 , 1t (`) 31 , 1b (`) = 2 (`),
2` (t) = 1 (t), 2r (t) 31 .
Notice that the equilibria of type 1 include the Nash equilibrium (t, `),
those of type 2 include the Nash equilibrium (b, r), and those of type 3
include the mixed equilibrium 1 (t) = 31 , 2 (`) = 31 .

5.2. Probabilistic Equilibria

103

Observability of Actions, Nash and Selfconfirming Equilibria

In simultaneous moves games, if it is possible to perfectly observe ex post
the opponents actions, then a self-confirming equilibrium is necessarily a
Nash equilibrium. Indeed, in this case, conjectures are confirmed if and
only if they are correct (Remark 19).25
However, static games are also used as reduced forms to analyze
the equilibria of corresponding dynamic games. Specifically, it will be
shown that each dynamic game admits a strategic-form representation
whereby it is assumed that players choose in advance and simultaneously
the contingent plans (strategies) that will determine their actions in every
possible situation that may arise as the play unfolds (see section 1.4.2 of
the Introduction, and section 8.1.5 of Chapter 8). In dynamic games, even
if, at the end of the game, it is possible to observe all the actions that have
actually been taken, it is not possible to observe how the opponents would
have played in circumstances that did not occur. Hence, it is impossible,
at least for some player, to observe the strategies of other players. For this
reason, in dynamic games it is easier to find self-confirming equilibria that
do not correspond to Nash equilibria of the strategic form.
Next we provide conditions on feedback that imply the equivalence of
self-confirming equilibria and mixed equilibria. First we define the ex post
information partition of Ai corresponding to a given action ai . Recall that
fi,ai = fi (ai , ) : Ai Mi is the section at ai of the feedback function fi .
This function shows how the feedback received by player i depends on the
opponents actions. For each i I and ai Ai , let
1
Fi (ai ) := {Ci 2Ai : mi Mi , Ci = fi,a
(mi )}
i

denote the partition of Ai given by the pre-images of function fi,ai . For

example, the game in Matrix 6 with fi = ui (i = 1, 2) has the following ex
post partitions, for each player and action:
F1 (t) = {A2 }, F1 (b) = {{`}, {r}},
F2 (`) = {{t}, {b}}, F2 (r) = {A1 }.
25

In the case of anonymous interaction, we can consider at least two scenarios: (a) the
individuals observe the statistical distribution of the actions in previous periods, (b) the
individuals observe the long-run frequencies of the opponents actions. In both cases we
can say that a conjecture is correct if it corresponds to the observed frequencies.

104

5. Equilibrium

Definition 23. A game with feedback (G, f ) satisfies own-action

independence of feedback if for every i I, a0i , a00i Ai ,
Fi (a0i ) = Fi (a00i ).
It is easily checked that the game in Matrix 6 with fi = ui (i = 1, 2)
does not satisfies own-action independence of feedback.
Definition 24. A game with feedback (G, f ) has observable payoffs if,
for every i I, ai Ai , and a0i , a00i Ai ,
fi (ai , a0i ) = fi (ai , a00i ) ui (ai , a0i ) = ui (ai , a00i ).
Note that the same condition can be expressed as
ui (ai , a0i ) 6= ui (ai , a00i ) fi (ai , a0 ) 6= fi (ai , a00 ).
In other words, the observable payoff condition says that for each player
i I there is a utility function vi : Mi R such that ui = vi fi . This
is the case, for example, if mi = fi (a) is the monetary payoff of i and
vi is the utility of money. More generally, f may be the consequence
function (C = iI Mi , f = g : A C) and fi may represent the
personal consequence function for player i. But even more generally fi
may specify other information accruing ex post to player i beside his
personal consequence. In all these cases the observable payoff assumption
is satisfied.
Example 11. Let G be a Cournot oligopoly (to have a finite game,
consider a version with capacity constraints and a finite grid of possible
outputs up to capacity). Suppose that firms know the market (inverse)
demand function and observe
Pex post only the price they fetch for their
n
output: fi (a1 , ..., an ) = P
j=1 aj . If P () is invertible, this game
with feedback has observable payoffs and satisfies own-action independence
of feedback. Indeed, each firm observes ex post its profit ui (ai , ai ) =
fi (a1 , ..., an )ai Ci (ai ) and also observes (indirectly) the total output of
the competitors independently of its own output choice: for each choice ai
and observed price p
X
aj = P 1 (p) ai .
j6=i

5.2. Probabilistic Equilibria

105

Yet, observability of payoffs is not a tautological assumption. Suppose,

for example, that players have other-regarding preferences. This means
that they do not only care about the consequences of interaction for
themselves, but also for (some) other players. For example, they may
care about the consumption of other players. If player i cares about
his own consumption ci and the consumption of other players ci , then
his utility is a function of the form vi (ci , ci ). The outcome function is
c = (cj )jI = (gj (a))jI = g(a). Assume that i observes only his own
consumption; then fi (a) = gi (a) Ci and ui (a) = vi (fi (a), gi (a)). Hence,
we may have fi (a0 ) = fi (a00 ) and yet ui (a0 ) 6= ui (a00 ). Intuitively, ui (a) is
not a utility experienced hence observed by i, it is a way of ranking
action profiles according to the preferences over the induced consumption
allocations. But i observes only his own component of the consumption
allocation; hence, he cannot know if the realized allocation is one he ranks
highly or not.
Theorem 12. If (G, f ) satisfies own-action independence of feedback and
has observable payoffs, then the sets of mixed equilibria and self-confirming
equilibrium mixed action profiles coincide.
Proof First note that observability of payoffs implies that, for each
player i I and action ai Ai , the section of the payoff function
of i at ai , ui,ai : Ai R, is constant on each cell of the ex post
information partition Fi (ai ): fix any Ci Fi (ai ) and a0i , a00i Ci .
Indeed, by definition of Fi (ai ), a0i and a00i yield the same message
given ai : fi (ai , a0i ) = fi (ai , a00i ); hence observability of payoffs implies
ui (ai , a0i ) = ui (ai , a00i ). With this, we write ui (ai , Ci ) for the common
value of ui (ai , ) on Ci Fi (ai ). Similarly, we write fi (ai , Ci ) for the
common feedback message on Ci Fi (ai ) given ai .26
Now fix a f -selfconfirming equilibrium (i , (iai )ai Ai )iI . We show
that, for every i I, ai Suppi and ai Ai
u
i (ai , i ) u
i (ai , i ).
By Theorem 9, this implies that (i )iI is a mixed equilibrium.
26

The message is constant on Ci Fi (ai ) by definition of Fi (ai ), whether or not

payoffs are obervable.

106

5. Equilibrium

Fix ai Ai arbitrarily (including the value ai = ai ). Observability of

payoffs implies
X Y
X
ui (ai , Ci )
u
i (ai , i ) =
j (aj ).
ai Ci j6=i

Ci Fi (ai )

Own-action independence yields Fi (ai ) = Fi (ai ); therefore

X
X Y
u
i (ai , i ) =
ui (ai , Ci )
j (aj ).
Ci Fi (ai )

ai Ci j6=i

The confirmed conjectures condition implies that for each Ci Fi (ai ),

X
X Y
iai (ai );
j (aj ) = Pfaii ,i [fi (ai , Ci )] = Pfai ,i [fi (ai , Ci )] =
i

ai Ci j6=i

ai

ai Ci

therefore
u
i (ai , i ) =

X
Ci Fi (ai )

ui (ai , Ci )

iai (ai ) = ui (ai , iai )

ai Ci

By the rationality condition

ui (ai , iai ) ui (ai , iai ).
Collecting the above equalities and inequalities we obtain, for each ai Ai
u
i (ai , i ) = ui (ai , iai ) ui (ai , iai ) = u
i (ai , i ).

Although observability of payoff is not tautological, it is common in
economic applications. Therefore the above theorem says that the key
property for the comparison between self-confirming and mixed (Nash)
equilibrium is own-action dependence/independence of feedback.

Learning Dynamics,
Equilibria and
Rationalizability
So far we avoided the details of learning dynamics. The analysis of such
dynamics requires the use of mathematical tools whose knowledge we do
not take for granted in these lecture notes (differential equations, difference
equations, stochastic processes). Nonetheless, it is possible to state some
elementary results about learning dynamics by addressing the following
question: When is it the case that a trajectory, i.e., an infinite sequence of
t
t
action profiles (at )
t=1 (with a = (ai )iI A), is consistent with adaptive
learning? We start with a qualitative answer.
Consider a finite game G that is played recurrently. Assume that all
actions are observable and consider the point of view of a player i that
observes that a certain profile ai has not been played for a very long
time, say for at least T periods. Then it is reasonable to assume that i
assigns to ai a very small probability. If T is sufficiently large, and ai was
not played in the periods t, t+ 1, ..., t+ T , ..., t+ T , then the probability of
ai in t > t+T will be so small that the best reply to is conjecture in t will
also be the best reply to a conjecture that assigns probability zero to ai .
In other words, i will choose in t > t + T only those actions that
are best
replies to conjectures i such that Suppi ai : t < t , i.e. only
107

108

6. Learning Dynamics, Equilibria and Rationalizability





actions in the set ri ai : t < t .1 Notice that this argument
assumes only that i is able to compute best replies to conjectures. The
argument is therefore consistent with a high degree of incompleteness of
information about the rules of the game (the consequence function) and
the opponents preferences.

6.1

Perfectly observable actions

Building on this intuition, one can use the rationalization operator to

define the consistency of a trajectory (at )
t=1 with adaptive learning under
the assumption that the opponents actions are perfectly observable.2
Recall that, for any given product set C1 ... Cn A, we defined the set
of profiles justified or rationalized by C, i.e. (C) = iI ri ((Ci )).
Now, the set of action profiles chosen in a given interval of time is not, in
general, a Cartesian product. For this reason, it is useful to generalize the
definition of as follows. Let C A (not necessarily a product set), then
 

(C) := iI ri projAi C ,
where projAi C = {ai : ai Ai , (ai , ai ) C} is the set of the other
players action profiles that are consistent with C. (If C is a product
set, the usual definition obtains.) Note that, even with this more general
definition, the operator is monotone: E F implies (E) (F ).
According to the reasoning developed above, suppose that players base
their conjectures on past observations and they maximize their expected
payoffs. Then, if for a very long time only action profiles in C are
observed, the current profile of best replies must be in (C). This explains
the following definition.
Definition 25. A trajectory (at )
t=1 is consistent with adaptive
learning (when opponents actions are perfectly observable)
if for

every

t there exists a T such that, for all t > t + T , at a : t < t .
1

Ai is finite and ui (ai , i ) is continuous (in fact linear) with respect to probabilities
( (ai ))ai Ai . It follows that if ai is a best reply to a conjecture that assigns a
sufficiently small probability to ai , then ai is a best reply also to a slightly different
conjecture that assigns probability zero to ai .
2
The following results are adapted from Milgrom and Robert (1991).
i

6.1. Perfectly observable actions

109

Next we define when a trajectory (at )

t=1 generates a limit
distribution (A).
Definition 26. (A) is the limit distribution of (at )
t=1 if for every
aA
| { : 1 < t, a = a} |
(i) lim
= (a)
t
t
(ii) if (a) = 0, then t : t > t, at 6= a.
This definition requires that the long-run frequency of every a is
(a), and furthermore that if (a) = 0 there exists a time starting from
which the profile a is no longer chosen. Convergence of a trajectory to
an action profile a is a special case: let (a ) = 1 in the definition, then

(at )
t=1 converges to a , written a a , if there exists a time t such that,
t

for every t > t, a = a (this is the standard notion of convergence for

discrete spaces, that are formed by isolated points). Also, notice that not
all trajectories admit a limit distribution.
It is now possible to present a few elementary results that relate
adaptive learning with the solution and equilibrium concepts introduced
earlier. The first result identifies a sufficient condition for consistency with
adaptive learning.3
Proposition 1. Let (at )
t=0 be a trajectory. If the limit distribution of
(provided
that
it
exists)
is a canonical correlated equilibrium, then
(at )
t=0
t

(a )t=0 is consistent with adaptive learning.

Proof. Let be the limit distribution of (at )
t=0 and fix any given
time t and profile a. If a is never chosen from t onwards, that is if
a
/ a : t < t for every t > t, then [Def. 26 (i)]
| { : 1 < t, a = a} |
t
lim = 0,
t
t t
t

(a) = lim

i.e. a
/ Supp. Therefore, for
every a Supp
there exists a time Ta0


such that, t > t + Ta0 , a a : t < t . Let T 0 := maxaSupp Ta0
(this
because A is finite). Then t > t + T , Supp
 is well defined

a : t < t . Moreover, if (a) = 0 there exists a Ta00 such that
3

See definition 19 and remark 10.

110

6. Learning Dynamics, Equilibria and Rationalizability

t > t + Ta00 , at 6= a [Def. 26 (ii)]. Let T 00 = maxaA\Supp Ta00 , then

t > t + T 00 , at Supp. Let T = max(T 0 , T 00 ), then


t > t + T , at Supp a : t < t .
It follows from the monotonicity of that


(Supp) ( a : t < t ).
Suppose that is a canonical correlated equilibrium. Then Supp
(Supp) (the proof is almost identical to the one of Theorem 11). Then


t > t + T , at Supp (Supp) ( a : t < t ).

Therefore
 there exists
a sufficiently large T for which t > t + T ,
t

a ( a : t < t ). 
The next two results show that, in the long run, adaptive learning
induces behavior consistent with the complete-information solution
concepts of earlier chapters, even if consistency with adaptive learning
does not require complete information.
Proposition 2. Let (at )
t=0 be a trajectory consistent with adaptive
learning. Then (1)
k 0, tk , t tk , at k (A),

(6.1.1)

so that only rationalizable actions are chosen in the long run; (2) if
at a , then a is a Nash equilibrium.
Proof.
(1) First recall that since A is finite there exists a K such that k K,
k (A) = (A). Then, (6.1.1) implies that from some time tK onwards
only rationalizable actions are chosen. The proof of (6.1.1) is by induction.
The statement trivially holds for k = 0. Suppose by way of induction that
the statement is true for a given k. By consistency with adaptive learning,
there exists a Tk such that
t > tk + Tk , at ({a : tk < t}).

6.2. Imperfectly observable actions

111

By the inductive assumption tk implies a k (A). Hence,

{a : tk < t} k (A).
By monotonicity of
({a : tk < t}) (k (A)) = k+1 (A).
Taking tk+1 = tk + Tk + 1, the above inclusions yield
t tk+1 , at ({a : tk < t}) k+1 (A)
as desired.
(2) If at a , there exists a t such that t > t, at = a . Consistency
of (at )
t=1 with adaptive learning implies that there exists T such that


t > t + T , a = at ( a : t < t ) = ({a }).
Since a ({a }), a is a Nash equilibrium. 

6.2

Imperfectly observable actions

The definition of consistency with adaptive learning needs to be modified

when players cannot perfectly observe the actions previously chosen by
their opponents. Assume that if an action profile a = (aj )jI is chosen
player i only observes a signal (or message) mi = fi (a). As in the section
on self-confirming equilibria 5.2.3, the profile of functions f = (fi : A
Mi )iI is a primitive element of the analysis.
Take a trajectory (at )
signals observed by player i from
t=1 . The set of 

time t (included) to time t (excluded) is mi : , t < t, mi = fi (a ) .

From is perspective, the set of other players action profiles that could
have
been
chosen
in


the same time interval is ai : , t < t, fi (a ) = fi (ai , ai ) . This
motivates the following definition:
Definition 27. A trajectory (at )
t=1 is f -consistent with adaptive learning

if for every
that, for every
 t there exists a T such

t > t + T and i I,
t

ai i ai : , t < t, fi (a ) = fi (ai , ai ) .

112

6. Learning Dynamics, Equilibria and Rationalizability

Remark 22. Imperfect observability implies that, in general, two players

i and j obtain different information about the past actions of a third player
k. This is the reason why the rationalization operator could not be used to
define the property of f -consistency with adaptive learning. If each feedback
function is injective, then there is perfect observability of other players
previous actions and the definition of the previous section (definition 25)
is obtained.
It is now possible to generalize the results that relate consistency with
adaptive learning and equilibrium. Given the discussion of section ??,
it should not be surprising that the relevant equilibrium concept for this
scenario is the self-confirming equilibrium.4
t
Proposition 3. Fix a trajectory (at )
t=0 . If (a )t=0 has a limit distribution
which corresponds to a f -anonymous self-confirming equilibrium, then
(at )
t=0 is f -consistent with adaptive learning.

Proof. [to be included] 

Proposition 4. If (at )
t=1 is f -consistent with adaptive learning and
t

a a then a is a self-confirming equilibrium profile.

Proof. If at a , there exists a t such that t > t, at = a . By
f -consistency of (at )
t=1 with adaptive learning, it follows that T, t >
t + T, i I,



ai = ati i ai : , t < t, fi (a ) = fi (ai , ai )



= i ai : fi (ai , ai ) = fi (ai , ai ) .
Hence, for every player i there exists

a belief i such that: (1) ai ri (i ),
i

(2)
ai : fi (ai , ai ) = fi (ai , ai ) = 1; (ai , i )iI is a self-confirming
equilibrium. 

The following results are adapted from Gilli (1999).

Games with incomplete

information
Let G = hI, (Ai , ui )iI i be a static game where the payoffs functions
ui : A R are derived from an outcome function g : A C and utility
functions vi : C R, that is, ui = vi g, i I. To make sense of the
strategic analysis of the game, it is typically assumed that G is common
knowledge, or - in other words - that there is complete information.
The reason is apparent for deductive solutions concepts, such as
rationalizability, that characterize the set of action profiles consistent
with common knowledge of the game, rationality and common belief in
rationality. Also the interpretation of the Nash equilibrium concept (or
more generally of correlated equilibrium) as a self-enforcing agreement
makes more sense under the complete information assumption. To see this,
suppose that before the game is played the players agreed to play some
Nash equilibrium profile a . Then, nobody has an incentive to deviate if
he expects that the others comply with the agreement. But why should the
players expect that the agreement is complied with? If the payoff functions
of others are unknown, a player could suspect that some other player has an
incentive to deviate and hence he may want to deviate too. In some games
the incentive to deviate when there is a small chance that the opponent
might not honor the agreement is very strong. For example, consider the
Pareto efficient agreement (a, c) of the following payoff matrix:
If Rowena does not know Colins payoff function, she may suspect for
example that d is dominant for Colin, and hence switch to her safe action
113

114

7. Games with incomplete information

c

100, 100

0, 99

99, 0

99, 99

b provided she assigns to d a probability above 1%. A similar argument

applies to Colin. Now suppose that payoff functions are mutually known,
but Rowena is not sure that Colin knows her payoff function. Then she
assigns a positive probability to Colin playing his safe action d instead of
the agreed upon action c, and again she would go for the safe action, b,
provided she assigns to d a probability above 1%. Continuing like this it
can be shown that even a high degree of mutual knowledge of the game
does not ensure that the agreement will be honored. On the other hand,
common knowledge of the game makes the agreement self-enforcing.1

7.1

Games with Incomplete information

How can we represent a game with incomplete information? The problem

can be viewed as follows: the outcome function g : A C and the
utility functions vi : C R depend on a vector of parameters which
is not commonly known. Hence, the payoff functions ui = vi g depend
on a parameters vector which is not commonly known.2 To represent this
uncertainty, payoff functions can be written in a parametrized form
ui : A R,
where is the set of values of that are not excluded by what is commonly
known among the players.
In order to represent what a player knows about , we assume for
simplicity is decomposable into subvectors j , with j {0} I, and that
each player i I knows the true value of i , whereas 0 represents residual
uncertainty that would persists even if players could credibly share their
private information. To further simplify the analysis, we will sometimes
1

Vice versa, recall that the interpretation of a Nash equilibrium (pure or mixed) as a
stationary state of an adaptive process does not require common knowledge, nor mutual
knowledge of the game.
2
The set of the opponents actions could be unknown as well. We omit this source of
uncertainty for the sake of simplicity.

7.1. Games with Incomplete information

115

assume that there is no such residual uncertainty (in many cases, this
assumption is rather innocuous). In this case 0 is a singleton and it
makes sense to omit it from the notation, writing = (1 , ..., n ).
Example 12. Consider again the team of two players producing a public
good (example 1). Now we express the production function and cost-ofeffort functions in a slightly different parametric form. The parametrized
production function is
y
y
Y = 0 (a1 )1 (a2 )2 .
The parametrized cost of the effort of player i measured in terms of output
is
1
Ci (ic , ai ) = c (ai )2 .
2i
The parametrized consequence function specifies a triple (Y, C1 , C2 ) as a
function of (, a1 , a2 ):


y
y
1
1
g(, a1 , a2 ) = 0 (a1 )1 (a2 )2 , c (a1 )2 , c (a2 )2 .
21
22
The commonly known utility function of player i is vi (Y, C1 , C2 ) = Y Ci .
The parametrized payoff function is then
y

ui (, a1 , a2 ) = vi (g(, a1 , a2 )) = 0 (a1 )1 (a2 )2

1
(ai )2 .
2ic

It is assumed to be common knowledge that 0 0 R+ and i =

(iy , ic ) i R2+ where the sets 0 , 1 and 2 are specified by the
given model. Each player i knows his efficiency parameter i = (iy , ic )
and nobody knows 0 , unless 0 is a singleton.
In example 12 the set = 0 1 2 only includes parameters that
directly affect the payoff functions of players and, for this reason, we call
them payoff-relevant (1y and 2y represent output elasticities with respect
to agents actions, 1c and 2c coefficients in the cost functions, and 0 is a
total factors productivity parameter). But, in general, the parameter
sub-vector i can also include payoff-irrelevant information, namely
information that does not directly affect payoffs. As we are going to show
in this chapter, even payoff-irrelevant parameters may affect the strategic
analysis. For instance, in example 12, we could include a description

116

7. Games with incomplete information

of agents political views; if this information is not common knowledge

between players and it is believed to affect their behavior, its inclusion may
play an important role in the analysis of the game. On the other hand,
since it is common knowledge that agents ignore the parameter capturing
residual uncertainty (0 ), it is also common knowledge that players cannot
condition their behavior on it. Therefore, residual uncertainty is relevant
only if it affects the payoff of agents and, consequently, we will assume that
0 includes only payoff-relevant parameters; formally, whenever 00 6= 000 ,
there must be some player i, information profile (i )iI , and action profile
a = (ai )iI such that ui (00 , 1 , ..., n , a) 6= ui (000 , 1 , ..., n , a).
Example 13. Consider the following Cournot oligopoly game with
incomplete information: The market inverse demand function is
P (0 , Q) = max{0, 0 Q}, with Q =

n
X

qi .

i=1

The cost function of each firm i is given by

Ci (i , qi ) =

1
(qi )2 .
2ic

Then
(
ui (, q1 , ..., qn ) = qi max 0, 0

n
X
i=1

)
qi

1
(qi )2 .
2ic

(7.1.1)

Example 14. Thus, 0 is an (unknown) parameter that captures the

market conditions in which the oligopolists operate, while ic parametrizes
the cost function of the i-th firm. Obviously, 0 and (ic )iI are payoff
relevant. Further, assume that each firm has a marketing department
which performs a market analysis and issues a score to summarize its
report; in particular, for every i I, let im = 0 + i , where i is normally
distributed with mean 0 and precision i and i , j are independent for
every i, j with i 6= j, conditional on 0 .3 The payoff function of each firm
i, given by eq. (7.1.1), is independent of (jm )jI ; thus, each im (i I) is
payoff-irrelevant; nevertheless, im may play a role in the strategic analysis:
3

0.

Assume, for simplicity, that the marketing department operational cost is equal to

7.1. Games with Incomplete information

117

(1) firm i can condition its behavior on im ; (2) firm j 6= i may believe that i
will condition its behavior on im and try to exploit the correlation between
jm , 0 and im to infer the behavior of i, and so on.

7.1.1

Private values, interdependent values and distributed

knowledge

In general, player is payoff function, ui , may depend on the whole profile

= (j )jI , and not just on i (see example 12). However, the analysis
is simpler when the consequence function g is commonly known and
uncertainty only concerns the utility functions vi : C R. In this case,
each player i knows his payoff function ui = vi g (given that g is commonly
known and vi represents is preferences over consequences and lotteries of
consequences). To represent other players ignorance of is utility (and
payoff) function, such a function can be written in parametrized form
vi (i , c), where i is known only to player i. This yields parameterized
payoff functions ui (i , a) = vi (i , g(a)), where i knows i . Whenever ui
varies only with i (and a) the game is said to have private values.
Suppose now that the consequence function g is not common
knowledge. Uncertainty about the consequence function can be expressed
by representing g in the parametrized form g(, a). Then, i identifies not
only is preferences, but also is private information about the consequence
function. Payoff functions have the general parametrized form ui (, a) =
vi (i , g(, a)). Since ui depends on the whole parameter vector , the game
is said to have interdependent values, meaning that ui may vary with
j (j 6= i).
Finally, we say that there is distributed knowledge of if parameter
can be identified by pooling the information of all the players. More
formally, let = (0 , 1 , ..., n ) and let i denote the set of possible values
of j (j {0} I). Then there is distributed knowledge of when 0
is a singleton.
Example 15. Consider again example 13. Then, there is distributed
knowledge of if the (inverse) demand function is commonly known, i.e.
0 = {0 }. In this case, the model also features private values.
Example 16. In the game of example 12 there is distributed knowledge
of if and only if there is only one possible value, say 0 , of the total
productivity parameter, i.e. 0 = {0 }. There are private values if and

118

7. Games with incomplete information

only if there is distributed knowledge of , hence common knowledge of

total productivity 0 , and moreover there is common knowledge of the
output elasticities with respect to players actions, 1y and 2y , that is,
i = {iy } ci (i = 1, 2). Otherwise there are interdependent values.

7.1.2

An elementary representation
information rationalizability

of

incomplete

A strategic environment with incomplete information can be simply

described by the following mathematical structure
= hI, 0 , (i , Ai , ui : A R)iI i ,
G
where 0 represents the set of residual uncertainty, i the set of parameter
values that is opponents consider possible, and as usual :=
iI{0} i , A := iI Ai . We call i , is private information about
parameter , the information-type of i (the term type will be used
later for a more complex object). For reasons that will be clear in a short
while, such structures are called games with payoff uncertainty (or,
sometimes, belief-free incomplete information games) rather than
games with incomplete information, although the latter terminology
would be perfectly legitimate.
It seems quite natural to assume that all the sets and functions specified
are commonly known. Otherwise, it would mean that
above by structure G
some aspects of players uncertainty have been omitted from the formal
representation.
Note, the size of j measures the ignorance of each player i 6= j
about j . It is important to specify how ignorant each player is. Even
he should still put the
if the modeler knew the true value of , say ,
sets j in the formal model. This is true even in the private value case.
Why? Because in the strategic analysis of the game each player i puts
himself in the shoes of each other player j and considers how j would
behave for each information-type j that i deems possible. Furthermore,
assuming that i deems j as strategically sophisticated as he is, i also puts
himself in the shoes of each of js opponents, say each k, considering all
the information-types k that i and j deem possible; and so on.
The specification of a game with payoff uncertainty is sufficient to
obtain some results about the outcomes of strategic interaction that are

7.1. Games with Incomplete information

119

consistent with rationality and standard assumptions about players beliefs

(for instance, some degrees of mutual belief in rationality). Consider the
following examples:
Example 17. There are two possible payoff functions for each player i,
ui (a , ) and ui (b , ). Player 1 (Rowena) knows the true payoff functions
while player 2 (Colin) does not know them. The situation can be
represented with two matrices, corresponding to the two possible values of
the parameter , assuming that Rowena knows the true payoff matrix.4
Action a dominates b in payoff matrix a , while b is justifiable (not
dominated) in matrix b .

1 :
G

4, 0

2, 1

2, 0

0, 1

3, 1

1, 0

0, 1

1, 2

Assume that the following assumptions hold: R1 , R2 , B1 (R2 ), B2 (R1 ),

B1 (B2 (R1 )).5 Obviously the implications of such hypotheses for Rowenas
choice depend on the true value of (which is known only to her). As
theorists, we do not make any assumption about , but rather we want
to explain what may happen (consistently with the above mentioned
assumptions) for each possible value .
R1 implies that Rowena chooses a if = a (a is dominant in matrix a ).
R2 B2 (R1 ) implies that Colin chooses d. Indeed, B2 (R1 ) implies that
Colin is certain that Rowena would choose a if = a . Hence, Colin assigns
probability zero to the pair (a ,b). Also,notice that d is dominant
when

the set of possible pairs is restricted to (a , a), (b , a), (b , b) .
R1 B1 (R2 ) B1 (B2 (R1 )) implies that Rowena expects d and as a
consequence chooses b if = b .
To sum up, the aforementioned assumptions imply that the outcome is
(a,d) if = a and (b,d) if = b . Note, the set of is actions consistent
with rationality and these epistemic assumptions can depend only on what
4
In this case, we can drop indexes from as only player 1 has private information
and


therefore there is a one-to-one correspondence between and 1 . Thus, 1 = 1a , 1b ,



2 = 2 , = a , b with a = (1a , 2 ) and b = (1b , 2 ).
5
The meaning of these symbols is explained in section 3.

120

7. Games with incomplete information

i knows. Then, in this case, the set of possible actions of Colin (the
singleton {d }) is independent of .
Example 18. Players 1 and 2 receive an envelope with a money prize
of k thousands Euros, with k = 1, ...K. It is possible that both players
receive the same prize. Every player knows the content of her own envelope
and can either offer to exchange her envelope with that of the other
player (action OE=Offer to Exchange), or not (action N=do Not offer to
exchange). The OE/N actions are taken simultaneously and the exchange
takes place only if it is offered by both. In order to offer an exchange a
player has to pay a small transaction cost . The players payoff is given
by the amount of money they end up with at the end of the game. For
this example, i = {1, ..., K} and ui (, a) is given by the following table:
ai \aj

OE

OE

Therefore, a necessary condition for i to offer to exchange is that he

assigns a positive probability to event [j > i ][aj = OE]. It can be shown
that the assumptions of rationality and common certainty of rationality
imply that i keeps his envelope, whatever the content. Let us consider,
for simplicity, only the case K = 3 (the general result can be obtained by
induction):
Ri implies that i keeps the envelope if i = 3, since by offering to exchange
he could obtain at most 3 .
Ri Bi (Rj ) implies that i keeps the envelope if i 2. Indeed, i is certain
that j would not offer to exchange if j = 3; hence i is certain that by
offering to exchange he would obtain at most 2 .
Ri Bi (Rj ) Bi (Bj (Ri )) implies that i keeps the envelope whatever the
value of i . Indeed i is certain that j could offer to exchange only if j = 1;
hence i is certain that by offering to exchange he could obtain at most
1 .
Intuitive, step-by-step solutions such as those of the previous examples
can be obtained by a generalization of the concept of rationalizability.
In general, we deem rationalizable any outcome consistent with the

7.1. Games with Incomplete information

121

assumptions of rationality and common belief in rationality, which we

denote using the symbol R CB(R),6 given the common background
knowledge. Rationalizability in games of complete information, for
instance, identifies the choices that are consistent with assumptions
R CB(R) given that the game is common knowledge.
In the analysis of strategic interaction with incomplete information, it is
interesting to address the following question: What behavior is consistent
with R CB(R) given that the features of strategic interaction encoded
= hI, 0 , (i , Ai , ui )iI i are common knowledge? The
by structure G
intuitive solutions of the previous examples suggest how to obtain the
answer: eliminate, step by step, the pairs (i , ai ) such that ai is not a best
reply to any probabilistic conjecture i (0 i Ai ) that assigns
probability zero to every (i , ai ) eliminated in the previous steps. By
Pearces lemma (Lemma 2), in a given step of the elimination procedure
one has to delete the pairs (i , ai ) such that ai is dominated by some mixed
action i given i and given the previous steps of elimination.
This solution procedure can be defined via a generalization of the
rationalization operator. Let i i Ai be the set of pairs for player i
that have not yet been eliminated, and let i = j6=i j . The definition
of best reply correspondence is extended in a very natural way: for every
i (0 i Ai ) and i i , let
X
i (0 , i , ai )ui (0 , i , i , ai , ai ). (7.1.2)
ri (i , i ) = arg max
ai Ai

0 ,i ,ai

As usual, we slightly abuse notation and write (0 i ) for the set of

probability measures on 0 (j6=i j Aj ) that assign probability one
to 0 i . Then define as follows: for each , =
6 = iI i
iI (i Ai ),


i (i ) = (i , ai ) i Ai : i (0 i ), ai ri (i , i ) .
() = iI i (i ).
A proper formalization of events (like rationality, R) and belief
operators (like Bi () and B()) would allow to derive the following table,
=
where it is implicitly assumed that there is common knowledge of G
6

Recall that CB(R) =

rationality.

k1

Bk (R) denotes common probability-one belief in

122

7. Games with incomplete information

hI, 0 , (i , Ai , ui )iI i [with another slight abuse of notation, write A

instead of 0 (iI (i Ai ))]:7
Assumptions about rationality and beliefs

Behavioral implications

( A)

R B(R)

2 ( A)

R B(R) B2 (R)

3 ( A)

...
R

T

K
k
k=1 B (R)

...


...

K+1 (

A)

...

In Example 17 3 ( A) = {(0 , a), (00 , b)} {d}. In Example 18

K ( A) = (1 {N }) (2 {N }) (after K steps, where K is the
number of possible prizes, only the pairs (i , N ) survive).
The results about rationalizability in games with complete information
can be easily generalized.
In particular, the equivalence between
rationalizability and a procedure of iterated elimination of dominated
actions still holds. Fix = iI i with i i Ai for each i. Let
i (i ) := {ai Ai : (i , ai ) Ai } denote the i -section of i . Say that
mixed action i dominates ai given i within , written i dom|i , ai ,
if Suppi i (i ) and
(0 , i , ai ) 0 i ,
X
i (a0i )ui (0 , i , i , a0i , ai ) > ui (0 , i , i , ai , ai ).
a0i

Next define the set of undominated profiles given , N D(), as follows

N Di () = i \ {(i , ai ) i : i (i ), i dom|i , ai } ,
N D() = iI N Di ().
Within this framework, Lemma 2 can be restated as follows:
7

Note that, by finiteness of A, for each i there is at least one action justified by
some conjecture i (i ). Therefore proji i (i ) = i .

7.2. Equilibrium and Beliefs: the simplest case

123

Lemma 8. Fix a nonempty set = jI j with j j Aj for each

j I. For each i I and (i , ai ) i , the following statements are
equivalent:
(1) there is no mixed action i such that i dom|i , ai ,
(2) there is a conjecture i (0 i ) such that
ai arg max

ai i (i )

i (0 , i , ai )ui (0 , i , i , ai , ai )

0 ,i ,ai

By a straightforward extension of the methods of Chapter 4, one can

use Lemma 8 to prove the following result:
Theorem 13. For each k = 1, 2, ..., k ( A) = N Dk ( A).
Therefore a profile of information-types and actions ((i , ai ))iI is
rationalizable if and only if it survives the iterated elimination of those
pairs (i , ai ) such that ai is dominated given i within the set of profiles
that survived the previous rounds of elimination.

7.2

Equilibrium and Beliefs: the simplest case

In examples 17 and 18 the repeated elimination of dominated actions

selects a unique outcome for each . In the complete information case
we proved that if there is a unique rationalizable profile, it must be the
unique equilibrium. Also in the present incomplete-information context we
can interpret the rationalizable mapping from information-types to actions,
if it is unique, as a unique equilibrium of a kind. For every player i and
every payoff-type i there is a corresponding choice, which in general can
be denoted by i (i ) (in example 17 1 (a ) = a, 1 (b ) = b, 2 = d; in
example 18 i (i ) = N), the unique profile of rationalizable choice functions
= (i : i Ai )iI has the following property: there exists no player
i, no information-type i and no belief over 0 i , such that i has an
incentive to deviate from i (i ) provided that i expects every other player j
to act according to the choice function j (); furthermore, is the unique
profile of choice functions with this property.
A choice function of player j can be interpreted as the conjecture of
player i about how j would behave as a function of his information-type j .
Then, one can tentatively define an equilibrium in this context as a profile

124

7. Games with incomplete information

of choice functions such that every player, given his information-type,

chooses a best reply to his conjecture about other players, and furthermore
this conjecture is correct, i.e. it corresponds to how each co-player chooses
as a function of his own payoff-type.
The problem of this tentative definition is that it silent on the belief
of each player i over 0 i . But in many games it is not possible to
ascertain whether a profile of choice functions (1 , ..., n ) has the property
that no player i has incentives to unilaterally deviate from i (what the
others expect him to choose) without further assumptions about players
beliefs about . This is shown by the following example.
1 (Rowena knows
Example 19. Consider the following variation of game G
the true payoff matrix, Colin does not):

2 :
G

4, 0

2, 1

1, 1

0, 0

3, 1

1, 0

0, 1

2, 0

How can we determine whether d is a best reply to the choice function

(1 (a ), 1 (b )) = (a,b)? It all depends on the probability that 2 (Colin)
assigns to a and b . If Colin thinks that a is more likely than b ,
P2 [a ] 12 , then d is a best reply, otherwise it is not. Suppose that
P2 [a ] < 21 . In an equilibrium (1 , 2 ), we must have 1 (a ) = a, because
a is dominant for payoff-type a of Rowena.8 In matrix b Colins payoff is
independent of Rowenas action; thus, if P2 [a ] < 12 , whatever the value of
1 (b ), Colins best reply to the choice function 1 is c (c yields expected
payoff 1 P2 [a ] > 21 , while d yields expected payoff P2 [a ] < 12 ). Rowenas
best response to c in matrix b is a. Then, if P2 [0 ] < 12 , the equilibrium
profile should be 1 (a ) = 1 (b ) = a, 2 = c.
The example shows that equilibrium analysis requires a specification
of beliefs about other players information-types. In general, in order to
define equilibrium behavior, it is necessary to enrich the mathematical
with the probabilistic beliefs of every player i about the
structure G
8

Of course, equilibrium has yet to be defined in this context. The formal definition
will follow.

7.2. Equilibrium and Beliefs: the simplest case

125

residual uncertainty parameter 0 and his co-players information-types

i = (j )jI\{i} . We denote such beliefs by pi (0 i ). It is
then natural to ask: How is pi determined? What does player j (j 6= i)
know or believe about pi ? If pi is simply a subjective probability, then
it does not seem very plausible to assume that j knows pi , and it is
even less plausible to assume that the beliefs profile (pi )iI is common
knowledge. Then, even though by fixing a belief profile (pi )iI we can
determine whether each action i (i ) (i i ) is a best reply to the profile
of choice functions i = (j )j6=i , it is not clear whether this best reply
property is sufficient to ensure that there is no incentive to deviate. How
can i be confident that j will follow his prescribed choice j (j ) if i does
not know pj and therefore cannot know whether each j (j ) (j j ) is
in turn a best reply to j ?
Actually, if we are to use a rigorous language, the word knowledge
should only be used for a true belief determined by direct observation or
logical deduction. As long as we assume that players cannot peer into
other players minds, we have to exclude that they can know anything
about the beliefs of others. However, a player may know facts that,
according to some psychological hypotheses, imply that the co-players
beliefs have some features, or maybe completely pin down the co-players
beliefs about unknown parameters, beliefs about such beliefs and so on.
If the psychological hypotheses are correct, this players beliefs about the
beliefs of others are also correct. In general, we say that some event E is
transparent if E is true and there is common belief of E, in symbols, if
E CB(E) is the case. With this, the correctly posed question is whether
the belief profile (pi )iI is transparent. In the affirmative case, we can carry
out a relatively simple equilibrium analysis with incomplete information.
Next, we consider a scenario where it is plausible to assume that (pi )iI
is transparent.9 For each player/role i I there is a large population
of agents that can play in that role. Each one of them is characterized
by an information-type i (for instance, his preferences, his ability, his
strength). Let qi (i ) denote the statistical distribution of i within
population i: qi (i ) is the fraction of agents in the population i with payofftype i . Furthermore, assume that 0 is determined through a random
experiment whose probabilities are fixed and transparent (for instance,
assume that 0 depends on the color of a ball extracted from an urn
9

This is what Harsanyi (1967-68) calls the prior lottery model .

126

7. Games with incomplete information

whose composition has been publicly announced in front of all players at

the same time); let the probability of 0 according to this experiment be
q0 (0 ). If players are randomly chosen from the corresponding populations,
then the probability of meeting opponents
characterized by the profile
Q
of payoff-types i = (j )j6=i is
qj (j ).10 If these statistical
jI\{i}

distributions are commonly known, then

Q it is reasonable to assume that,
for all i and i , pi (0 , i ) =
qj (j ) and that the profile of
j{0}I\{i}
i
objective
transparent.
We can then analyze the

beliefs (p )iI is

structure I, 0 , (i , Ai , ui , pi )iI assuming that all the features of the
interactive situation it describes are transparent, and we can meaningfully
define as equilibrium a profile of choice functions (1 , ..., n ) such that,
for each i and i
X
pi (0 , i )ui (0 , i , i , ai , i (i )),
(7.2.1)
i (i ) arg max
ai Ai

0 ,i

where i (i ) = (j (j ))jI\{i} .11

Example 20. (First Price Auctions with independent private values) A
given object is put on sale by means of an auction. The set of participant
is I = {1, ..., n}, the monetary value for the object for participant i is i .
Therefore the model has private values. The set of possible valuations is
i = [0, 1] and the valuations profile (1 , ...n ) is uniformly distributed
on [0, 1]n . (This means that each bidder i believes that his competitors
valuations are mutually independent, uniform random variables.) The
object is assigned to the player that makes the highest offer. In case of
a draw, it is randomly assigned among the highest bidders. Whoever is
given the object has to pay his bid (First Price rule). The resulting payoff
function is:
(
1
(i ai ) | arg max
, if ai = maxj aj
j aj |
ui (, a) =
0,
if ai < maxj aj
10
As usual, we am assuming that every j is finite. The generalization to the infinite
case, though conceptually trivial, requires the

use of measure theoretic

 concepts.
11
Furthermore, even if the structure I, 0 , (i , Ai , ui , pi )iN obtained by the
statistical distributions were not transparent (perhaps because the statistical
distributions are not commonly known), the definition of equilibrium given in the text
could perhaps be motivated as a possible stationary state of an adaptive process. This
is certainly the case under private values.

7.3. The General Case: Bayesian Games and Equilibrium

127

It turns out that this game has a symmetric equilibrium in which each
player bids i (i ) = n1
n i .
How can we guess that these functions form an equilibrium? Consider the
following heuristic derivation. If i = 0 it does not make sense (indeed,
it is weakly dominated) to offer more than zero. Now assume i > 0. If
bidder i conjectures that each competitor bids according to the linear rule
j (j ) = kj , where k (0, 1) is a coefficient that we have to determine,
then i believes that, if he bids ai , the probability of winning the object is
ai
P[j 6= i, j (j ) < ai ] = P[j 6= i, j < ]
k
(we can neglect ties because, for each competitor j, the probability of event
[e
aj = ai ] is zero). Since we assumed independent and uniform distributions
of values, we have
 a
n1
i
ai
, if aki < 1,
k
P[j 6= i, j < ] =
k
1,
if aki 1.
Thus, if bidder i is
rational he offers the smallest of the following numbers:
n1
12
arg max0ai <k aki
(i ai ) = n1
n i and k, that is
n1
i }
n
In a symmetric equilibrium each player has a correct conjecture about the
bidding functions of the competitors and each player has the same (best
n1
reply) bidding function, therefore k = n1
n . This implies min{k, n i } =
n1
n1
min{ n1
n , n i } = n i . Therefore, the optimal bid when the symmetric
conjecture of i is that each j bids j (j ) = n1
n j is precisely ai = i (i ) =
n1
n i .
ai = min{k,

7.3

The General Case:

Equilibrium

Bayesian Games and

In order to provide a more general definition of equilibrium in games of

incomplete information it is necessary to consider the case in which the
12

To compute the maximizer in [0, 1) use the first order condition:

 a n1
(n 1)  ai n2
i
(i ai )
= 0.
k
k
k

128

7. Games with incomplete information

beliefs of a generic player i on 0 i are not known to his co-players

i. This means that i is not characterized by his private information i
alone, and the belief pi has to be separately specified. From js point of
view the pair (i , pi ) i (0 i ) is unknown. Why should j care
about pi ? The problem, from js point of view, is that js payoff depends
on is choice ai which in turn depends on i and pi . Thus, in order to
extend the equilibrium concept to this more general case it is necessary to
specify js beliefs about (i , pi ).
According to the subjective formulation, known as the Bayesian
approach, a decision maker forms probabilistic beliefs over all the relevant
variables that are unknown to him. In our example j has probabilistic
beliefs about (i , pi ).
To simplify the exposition, we first focus on the two-person case
with distributed knowledge of parameter (0 is a singleton): the only
opponent of i is j and viceversa. The beliefs of j about i and pi are
therefore a joint probability measure q j (i (j )) from which the
beliefs pj (i ) can be recovered by computing the marginal distribution
on i . The beliefs q j are called second-order beliefs, whereas the beliefs
pj are called first-order beliefs: second-order beliefs are beliefs about
the primitive uncertainty and the opponents first-order beliefs. Hence,
first-order beliefs can be derived from second-order beliefs.
For some applications there is no need to go beyond second-order
beliefs. Suppose, for example that i = {i0 , i1 }, i = 1, 2. Then we
may think of pi as is subjective probability of j1 : (j ) is isomorphic
to [0, 1]. Suppose that it is transparent that the second-order beliefs of
each player j have the form q j = pj q j where q j is given by the uniform
distribution on [0, 1], which is the set of possible values of the subjective
probability pi (j1 ) (recall: by transparent we mean that second-order
beliefs do indeed have this form and furthermore it is common belief that
they have this form). Then each first-order belief pj also pins down the
second-order belief q j = pj q j and there is no need to explicitly consider
beliefs of higher order.
However, since beliefs are subjective, we have to allow for more
general situations where second-order beliefs are not pinned down by firstorder beliefs. Since behavior may depend on second-order beliefs, it is
necessary to introduce third-order beliefs rj (i (j ) (j

7.3. The General Case: Bayesian Games and Equilibrium

129

(i ))),13 from which the first and second-order beliefs can be recovered
by marginalization. At this point, the formalism is already quite complex.
Furthermore, there is no compelling reason to stop at third-order beliefs.
How should we proceed? Is it possible to use a formal and compact
representation of strategic interactions with incomplete information which
is not too complex, but at the same time allows a representation of beliefs
of arbitrarily high order and a meaningful definition of equilibrium? The
answer is Yes.

7.3.1

Bayesian Games

The solution relies on the adoption of a more abstract and self-referential

approach proposed by John Harsanyi [22].14 Starting from the game
with payoff uncertainty hI, 0 , (i , Ai , ui )iI i, let us consider a richer
structure with a set of states of the world (assumed finite for
the sake of simplicity). A state of the world characterizes each
players knowledge and interactive beliefs. This can be formalized
mathematically introducing functions i : Ti and i : Ti i , for
each player i, a function 0 : 0 and a probability measure pi ().
It is also implicitly assumed that all these elements are common knowledge.
To grasp the meaning of the above functions we follow Harsanyi and
present a metaphor.15 Imagine an ex ante state in which all players are
equally ignorant. Each player i is endowed with a prior (subjective)
probability measure pi (). Before choosing an action, each player
i receives a signal ti about the state of the world. From signal ti player
i infers that i = i (ti ) and i1 (ti ) = { 0 : i ( 0 ) = ti } . Assume
for simplicity that pi assigns positive probability to all signals that i could
13

Clearly, the symbol rj used here is not to be confused with the symbol denoting the
best reply correspondence.
14
This fundamental contribution by lead to the award to John Harsanyi of the 1994
Nobel Prize for Economics, jointly with John Nash and Reinhard Selten.
15
This is what Harsanyi calls the random vector model of the Bayesian game. The
only difference is that here (not in his general analysis of incomplete information)
Harsany postulates a common prior, p = p1 = p2 = ... = pn , which represents an
objective probability distribution.

130

7. Games with incomplete information

conceivably receive,16 i.e.

pi [i1 (ti )] =

pi () > 0

:i ()=ti

for all ti Ti .17 Then, for each ti Ti , the corresponding conditional

probability measure pi [|ti ] is well defined:
E , pi [E|ti ] :=

pi [E i1 (ti )]
.
pi [i1 (ti )]

With this, the beliefs of player i at state of the world are given by
the distribution pi [|i ()]. Since i and pi are common knowledge, the
function 7 pi [|i ()] () is also common knowledge. It follows
that a state of the world determines a players beliefs about the parameter
and about other players beliefs about it. To verify this claim, let us focus
on the case of two players and distributed knowledge of (this is done
only to simplify the notation) and let us derive the beliefs of i about tj
given signal ti :
tj Tj , pi [tj |ti ] = pi [j1 (tj )|ti ],
where j1 (tj ) = { : j () = tj } . Next we derive the first-order beliefs
about (since i knows i , his beliefs about are determined by his beliefs
about j ):
j j , p1i [j |ti ] := pi [1
j (j )|ti ],
1
where 1
j (j ) = {tj : j (tj ) = j }. The functions tj 7 pj [|tj ] (i )
(j {1, 2}) are also common knowledge. Hence, we can derive the secondorder beliefs about , that is, the joint belief of a player about and about
the opponents first-order beliefs about :

(j , p1j ) j (i ), p2i [j , p1j |ti ] :=

pi [tj |ti ].

tj :j (tj )=j ,p1j [|tj ]=p1j

16

It can be shown that this comes without substantial loss of generality provided that
we allow the subjective priors p1 , ..., pn to be different.
17
We denote by pi [] and pi [|], respectively, the prior and conditional probabilities of
events. For instance, pi [ti |ti ] is the probability that event [ti ] = { : i () = ti }
occurs conditional on the event [ti ] = { : i () = ti } .

7.3. The General Case: Bayesian Games and Equilibrium

131

[It can be verified that

j j , p1i [j |ti ] =

p2i [j , p1j |ti ],

p1j

in words, p11 [|ti ] is the marginal on j of the joint distribution p2i [|ti ]
(j (i )).]
It should be clear by now that is possible to iterate the argument
and compute the functions that assign to each type the corresponding
third-order beliefs about , fourth-order beliefs about , and so on. To
sum up, we can conclude that the information and beliefs of all orders
of player i about are determined by ti according to the function
ti 7 (i (ti ), p1i [|ti ], p2i [|ti ], ...). Signal ti is called the type 18 of player i. We
denote the information and belief hierarchy associated with ti by pi (ti ) .
The information and beliefs of each player i in a given state of the world
are those corresponding to the type ti = i (). The information-type of
player i, i = i (ti ), is just one component of his overall type ti , which also
specifies is beliefs about all the relevant exogenous variables/parameters,
i , p1i , p2i , etc. (exogenous here means something that we are not
trying to explain as game theorists: we are not trying to explain , nor
beliefs about , nor beliefs about such beliefs, nor - more generally - any
beliefs about exogenous things).
Definition 28. A Bayesian Game is a structure
BG = hI, , 0 , 0 , (i , Ti , Ai , i , i , pi , ui )iI i
where, for each i I, pi (), 0 : 0 , i : Ti , i : Ti i ,
pi [ti ] := pi [i1 (ti )] > 0 for each ti Ti ,19 and ui : A R.
The analysis of solution concepts for Bayesian games will clarify that
only the beliefs pi [|ti ] (i I, ti Ti ) really matter. The priors pi are just
a convenient mathematical tool to represent the beliefs of each type.
In what follows, we will often use the expression type ti chooses ai
to mean that if player i were of type ti he would choose action ai .
18
19

The expression type a

` la Harsanyi is also used.
This condition implies that i : Ti is onto.

132

7.3.2

7. Games with incomplete information

Bayesian Equilibria

In general, players choices depend not only on their private information

on , but also more generally on their types. In equilibrium, each player
has a correct conjecture about the dependence of each opponents choice
on his type, and maximizes his expected payoff given his type:
Definition 29. A Bayesian equilibrium is a profile of choice functions
(i : Ti Ai )iI such that
i I, ti Ti ,
X
p[|ti ]ui (0 (), i (ti ), i (i ()), ai , i (i ())),
i (ti ) arg max
ai Ai

where, for every and ti , i () := (j ())j6=i , i (ti ) := (j (tj ))j6=i ,

i (ti ) := (j (tj ))j6=i .
Note that the expected payoff of type ti of player i choosing action ai
can be re-written as
X
p[0 , ti |ti ]ui (0 , i (ti ), i (ti ), ai , i (ti ))
0 ,ti

where
p[0 , ti |ti ]
is
the
probability
that event [0 , ti ] = { : 0 () = 0 , i () = ti } occurs conditional
on the event [ti ] = { : i () = ti } .
Example 21. We illustrate the concepts of Bayesian game and equilibrium
2 . Suppose that Rowena does not know Colins first-order
elaborating on G
beliefs. From her point of view such beliefs can assign either probability 31
or 34 to a . The two possibilities are regarded as equally likely by Rowena
and all this is common knowledge. This situation can
 be represented

 by the

following Bayesian game: = {, , , }, 1 = 1a , 1b , T1 = ta1 , tb1 ,
b
a
a
b
b
1 () = 1 () = ta1 , 1 () =
1 () = t1 , 1 (t1 ) = 1 , 1 (t1 ) = 1 ,
n o
, p1 () = 1 , 2 = 2 , T2 = {t0 , t00 }, 2 () = 2 () = t0 ,
2

t002 ,

p2 () = 83 ,
2.
are as in G

1
6,

1
8,

1
3,

2 () = 2 () =
p2 () = p2 () = p2 () = and where
the functions ui
The probabilistic structure is described in
the table below:
To verify that this Bayesian game represents the situation outlined
above we compute the following:

7.3. The General Case: Bayesian Games and Equilibrium

t02

t002

a , ta1

, 14 , 83

, 14 , 16

b , tb1

, 14 , 18

, 14 , 13

133

p2 ()
3/8
3
=
=
p2 () + p2 ()
3/8 + 1/8
4
p2 ()
1/6
1
p12 [a |t002 ] = p2 [ta1 |t002 ] =
=
=
p2 () + p2 ()
1/6 + 1/3
3
1
p1 [t02 |ta1 ] = p1 [t02 |tb1 ] = .
2
This means that in every state of the world Rowena believes that the two
events [Colin assigns probability 43 to a ] and [Colin assigns probability
1
a
4 to ] are equally likely. We now derive the Bayesian equilibria. Let
= (1 , 2 ) be an equilibrium. Since a is dominant when Rowena knows
that = a , 1 (ta1 ) = a. Hence, the equilibrium expected payoff accruing
to type t02 if he chooses c is
p12 [a |t02 ] = p2 [ta1 |t02 ] =

3
1
1
0+ 1= ;
4
4
4
0
the equilibrium expected payoff for t2 if he chooses d is
p2 [ta1 |t02 ]u2 (a , a,c) + p2 [tb1 |t02 ]u2 (b , 1 (tb1 ), c) =

3
1
3
1+ 0= .
4
4
4
0
It follows that in equilibrium 2 (t2 ) = d. The expected payoffs for the two
actions of type t002 are
p2 [ta1 |t02 ]u2 (a , a,d) + p2 [tb1 |t02 ]u2 (b , 1 (tb1 ), d) =

1
0+
3
1
p2 [ta1 |t002 ]u2 (a , a,d) + p2 [tb1 |t002 ]u2 (b , 1 (tb1 ), d) =
1+
3
Since the maximizing choice for type t002 is c, 2 (t002 ) = c.
determine the equilibrium choice for type tb1 :
p2 [ta1 |t002 ]u2 (a , a,c) + p2 [tb1 |t002 ]u2 (b , 1 (tb1 ), c) =

p1 [t02 |tb1 ]u1 (b , a,d) + p1 [t002 |tb1 ]u1 (b , a,c) =

p1 [t002 |tb1 ]u1 (b , b,d) + p1 [t002 |tb1 ]u1 (b , b,c) =

1
0+
2
1
2+
2

2
2
1= ,
3
3
2
1
0= .
3
3
We can now

1
1
1= ,
2
2
1
0 = 1.
2

134

7. Games with incomplete information

Therefore, 1 (tb1 ) = b.

7.3.3

Bayesian Equilibrium and Nash Equilibrium

The concept of Bayesian equilibrium for a game of incomplete information

BG can be restated in an equivalent fashion as a Nash equilibrium of two
games with complete information that are associated to the original game:
the ex ante strategic form and the interim strategic form.
The ex ante strategic form refers to the metaphor that was previously
introduced to explain the elements of the Bayesian game: if there is ex
ante stage in which all players are ignorant, then at such stage each
player i can make a contingent plan of action, or strategy, i : Ti Ai ,
that specifies the action to take for every possible signal ti that player i
might receive. If i believes that the other players follow the strategy profile
i , the expected payoff of strategy i is
X
Ui (i , i ) : =
pi ()ui (0 () , i (i ()), i (i ()), i (i ()), i (i ()))

X
ti Ti

pi [ti ]

pi [0 , ti |ti ]ui (0 , i (ti ), i (ti ), i (ti ), i (t(7.3.1)

i )).

ti Ti

Let i = ATi i (recall, this is the set of functions with domain Ti and range
Ai ). The ex ante strategic form of BG is the static game
AS(BG) = hI, (i , Ui )iI i ,
where Ui is defined by (7.3.1). The reader should be able to prove the
following result by inspection of the definitions and using the assumption
that each type ti is assigned positive probability by the prior pi :
Remark 23. A profile (1 , ..., n ) is a Bayesian equilibrium of BG if and
only if it is a Nash equilibrium of the game AS(BG).
The interim strategic form is based on a different metaphor. Assume
that for each role i in the game there is a set of potential players Ti . A
potential player ti is characterized by the payoff function ui (i (ti ), , , ) :
0 i A R and the beliefs pi [, |ti ] (0 Ti ). In the event
that ti is selected to play the game in the role of agent i, he will assign
probability p[0 , ti |ti ] to the event that the residual uncertainty is 0 and

7.3. The General Case: Bayesian Games and Equilibrium

135

that he is facing exactly the profile of opponents ti = (tj )j6=i . The set
of actions available to ti is Ai , that is, Ati = Ai (i I, ti Ti ). If each
potential player tj Tj chooses the action atj Atj , ti s expected payoff
is computed as follows:
X
uti ((atj )jI,tj Tj ) :=
pi [0 , ti |ti ]ui (0 , i (ti ), i (ti ), ati , (atj )j6=i ).
0 ,ti

(7.3.2)
The
interim
strategic
form
of
BG
is
the
following
static
game
with
P
|T
|
players:
i
iI
*
+
[
Ti , (Ati , uti )iI,ti Ti ,
IS(BG) =
iI

where uti is defined by (7.3.2). The reader should be able to prove the
following result by inspection of the definitions:
Remark 24. A profile (1 , ..., n ) is a Bayesian equilibrium of BG if and
only if the corresponding profile (ati )iI,ti Ti such that ati = i (ti ) (i I,
ti Ti ) is a Nash equilibrium of the game IS(BG).

7.3.4

Bayesian equilibria, common prior and correlated

equilibria

Even though it may seem counter-intuitive, the definition of Bayesian game

admits as a particular case that there is common knowledge of the payoff
functions and yet there are multiple types for some players. This is the
case when there are functions u
i RA (i I) such that ui (, ) = u
i for
each and i (when the sets 0 and i are all singletons we obtain a special
case of this). In other words, we can define a Bayesian game with many
types even starting from a game with complete information! An even more
special case can be analyzed: the one in which, besides having complete
information, all players are characterized by the same prior belief (pi = p
for each i I), which we refer to as common prior. What can we say
about Bayesian equilibria of a game with complete information?
To make the exposition more precise, we have to introduce some
terminology and notation: given a complete information game G =
hI, (Ai , u
i )iI i we call Bayesian elaboration of G any Bayesian game
BG = hI, , 0 , 0 , (i , Ti , Ai , i , i , pi , ui )iI i

136

7. Games with incomplete information

such that, 0 and every i (i I) contains a single element, say 0 and i

a) = u
(i I). Thus, for every profile a A, we can write ui (,
i (a). The
following remark follows directly from the definitions:
Remark 25. Let BG be a Bayesian elaboration of a game with complete
information G and suppose that there is a common prior. Then, every
Bayesian equilibrium of BG corresponds to a correlated equilibrium of G.20

7.3.5

Subjective
correlated
equilibrium,
equilibrium and rationalizability

Bayesian

The equivalence between the correlated equilibria of a game with complete

information G and the Bayesian equilibria of its Bayesian elaboration
BG holds under the assumption that players share a common prior.
Removing this common prior assumption, one obtains the notion of
subjective correlated equilibrium: a subjective correlated equilibrium
of a complete information game G is a Bayesian equilibrium of a Bayesian
elaboration of G.
Theorem 14. An action profile (ai )iI of the complete information game
G is rationalizable if and only if it is selected in a subjective correlated
equilibrium (more explicitly, if and only if there are a Bayesian elaboration
BG of G, an equilibrium (i )iI of BG, and a state of the world in BG
such that ai = i (i ()) for every i I).
Proof. (If ) Let be an equilibrium of a Bayesian elaboration of G.
To show that every image of is a profile of rationalizable actions, define
the sets Ci = i (Ti ) (i I). It is easy to verify that the cross-product
C = C1 ... Cn has the best reply property. Therefore, every element
in C is rationalizable (Theorem 2). By construction, C is the set of action
profiles that are selected by the equilibrium in some state of the world
.
(Only if ) We show that there exist a subjective correlated equilibrium
in which every rationalizable profile is played in some state of the world.
Let C = (A) be the set of rationalizable profiles. Then C = (C)
(Theorem 2) and for every i I and ai Ci there exists a belief
ai (Ci ) that justifies ai (i.e., such that ai ri (ai )). The Bayesian
20

In a way, one could say that Harsanyi [22] had implicitly defined the correlated
equilibrium concept before Aumann [1], but without being aware of it!

7.3. The General Case: Bayesian Games and Equilibrium

137

elaboration of G is as follows. Types are copies of rationalizable actions

(you may interpret them as suggested actions ); if the type of i is a copy
of action ai , then the belief of i is derived from the justifying conjecture
ai . Formally, = C; for every i I, Ti = Ci and i = IdCi ; for
every rationalizable action/type ti Ci = Ti , pi [|ti ] = ti , for every
t = (tj )jI C = , pi (t) = |C1i | pi [ti |ti ]. Then, the profile of identity
functions (for every i I and for every ai Ti = Ai , i (ai ) = ai ) is an
equilibrium of this Bayesian elaboration.21 
This result can be generalized. Fix a game with payoff uncertainty
= hI, 0 , (i , Ai , ui : A R)iI i ;
G
is a Bayesian game obtained by appending a
a Bayesian game based on G
that is
state space, signal functions etc. to G,
BG = hI, , 0 , 0 , (i , Ti , Ai , i , i , pi , ui )iI i .
Using arguments that generalize the proof of Theorem 14, one can
prove the following result:
Theorem 15. A profile of payoff-types and actions (i , ai )iI is
if and only if there is
rationalizable in the game with payoff uncertainty G

a Bayesian game BG based on G, an equilibrium of BG and a state of

the world in BG such that (i , ai )iI = (i (i ()), i (i ()))iI .
These results show that, without restrictions on players beliefs
about parameters, the Bayesian equilibrium assumption that players best
respond to correct conjectures about their opponents choice functions
has no behavioral implications beyond those that can be derived from
rationality and common belief in rationality.22 In a sense, rationalizability
gives the robust implications of Bayesian equilibrium analysis.
21

The type ti = ai is not to be interpreted as the action that i necessarily has to

play but rather as a specification of is belief about the opponents and of the best reply
to such belief, a best reply that i freely chooses in equilibrium.
22
See Brandenburger and Dekel [10], and Battigalli and Siniscalchi [4].

138

7.4

7. Games with incomplete information

Incomplete Information and Asymmetric

Information

Let us go back to Harsanyis metaphor, presented in section 7.3.1: the

information players have, the residual uncertainty and, potentially, players
payoffs depend on a random variable (or move by nature). Let
denote a typical realization of this random variable and assume that this
realization occurs before players make their choices. The payoff of player
i is determined by the function u
i : A R. Let pi () denote the
subjective probability measure of i over the possible moves by nature. Each
player receives a signal regarding the move by nature and then chooses an
action ai Ai (simultaneously to the other players). Let i : Ti
denote player is signal function. The prior belief and signal function of
each player i are such that pi [i1 (ti )] > 0 for all ti Ti . This is a game
with asymmetric information about an initial move by chance.
In this context, it makes sense to suppose the subjective priors pi
coincide with an objective probability measure p () and that this
is commonly knowledge. More generally, even if the priors do not coincide,
it is assumed that the profile of subjective measures (pi )iI is transparent.
An equilibrium of such game with asymmetric information is a strategy
profile (i : Ti Ai )iI such that the strategy of each player is a best
reply to the strategy profile of the other players. More explicitly, define
the payoff function in strategic form
X
Ui (1 , ..., n ) =
pi ()
ui (, 1 (1 ()), , ..., n (n ())).

An equilibrium of a given game of asymmetric information is a Nash

equilibrium of the corresponding game in strategic form hI, (i , Ui )iI i,
where i = ATi i for each i I.
At this point you may ask yourself, how is this different from a game
of incomplete information, modeled as a Bayesian game? If we only look
at the mathematical formulation, the two models are equivalent. At the
stage in which all players have received their private signals regarding
the move by nature, the strategic situation is indeed very similar to one
with incomplete information: the way payoffs depend on actions is not
commonly known; each player has private information about it, along with

7.4. Incomplete Information and Asymmetric Information

139

probabilistic beliefs regarding the private information and beliefs of other

players.
There are several ways to establish an explicit relation between the
two models. For example, we can obtain a corresponding Bayesian game
defining i = Ti , letting i be the identity function on Ti , and defining
X
ui (, a) =
p[|]
ui (, a),
(7.4.1)

where p[|] =

1 ( ,..., )]
p[{}1
n
1
0 (0 )
, 1
1 ( ,..., )]
0 (0 )
p[1
(
)
n
0
1
0
0
0
0
{ : (1 ( ) = 1 , ..., n ( ) = n }

= { 0 : 0 ( 0 ) = 0 },

1 (1 , , , .n ) =
(the value of p[|] if
p[] = 0 is irrelevant for expected payoff computations). Keeping in mind
Remark 23, it is easy to verify that a profile of functions (i )iN is a
Bayesian equilibrium of the game BG so obtained if and only if it is an
equilibrium of the asymmetric information game.
Recall that in order to introduce the constituent elements of the
model of the Bayesian game we used the metaphor of the ex ante stage.
The use of such metaphor and the mathematical-formal analogy between
Bayesian and asymmetric information games (and the corresponding
solution concepts) has induced many scholars to neglect the differences
between these two interactive decision problems. We should stress,
however, that they are indeed two different problems. In the incomplete
information problem the ex ante stage does not exist, it is just a useful
theoretical fiction: only represents a set of possible states of the world,
where by state of the world we mean a configuration of information and
subjective beliefs. The so called prior beliefs pi () are simply a useful
mathematical device to determine (along with functions 0 , i and i ) the
players interactive beliefs in a given state of the world.23 Instead, in the
23

Essentially, nothing would have changed if we had only specified the belief functions
i : Ti (0 Ti ) that for any type ti determine a probability measure over the set
of residual uncertainty and other players types. At the end of the day, we are simply
interested in the types beliefs. Moreover, it is always possible to specify a distribution
pi that yields the beliefs pi [|ti ] = i (ti ) for any type ti Ti . Actually, there exist an
infinite number of them, as it is apparent from the following construction. Let (Ti )
be any strictly positive distribution and determine pi as follows
(0 , ti , ti )

0 T , pi [0 , t] = (ti )i (ti )[0 , ti ]

pi [0 , t]
1
1
(t), pi () = 1
0 (0 )
|0 (0 ) 1 (t)|

140

7. Games with incomplete information

interactive decision problems that we just called games with asymmetric

information the ex ante stage is real and the players priors represent the
expectations they hold at that stage.
The differences in the interpretation of the formal model are not
innocuous. The interpretation (asymmetric information vs. incomplete
information) determines to what extent given assumptions are meaningful
and plausible. For instance, for the case of asymmetric information
about an initial random move, it is meaningful and plausible to assume
that there exists a common prior given by an objective distribution over
the moves by nature. On the other hand, in the case of incomplete
information the meaning of the common prior assumption is not obvious
and its plausibility is even less obvious. Furthermore, we will see that
different notions of rationalizability and self-confirming equilibrium may
be appropriate for a situation with incomplete information and for one
with asymmetric information about an initial move by nature, even though
the two situations can be represented by the same mathematical structure.

7.5

Rationalizability in Bayesian Games

A game with payoff uncertainty, i.e. the mathematical structure

= hI, 0 , (i , Ai , ui )iI i ,
G
does not specify the exogenous interactive beliefs of players, that is, their
beliefs about private information and beliefs of the opponents (here we call
such beliefs exogenous because they are not explained, nor partially
determined, by strategic reasoning). A specification of exogenous beliefs is
necessary to extend the traditional definition of equilibrium due to Nash.24
The richer structure
BG = hI, , 0 , 0 , (i , Ti , Ai , i , i , pi , ui )iI i ,
known as Bayesian game, was introduced with this purpose, allowing
the definition of Bayesian equilibrium.
(recall that () = (1 (), ..., n ()) and | | denotes the cardinality of a set). Then,
pi [0 , ti |ti ] = i (ti )[0 , ti ] for each ti , that is pi [|ti ] = i (ti ).
24
As we argue in section 7.7, this does not hold for other definitions of equilibrium
that characterize the rest points of adaptive processes.

7.5. Rationalizability in Bayesian Games

141

It was also pointed out that, for equilibrium analysis, it makes no

difference whether the given mathematical structure BG represents a game
situation where there is no common knowledge of the payoff functions
(incomplete information), or a situation where there is asymmetric
information about an initial chance move. The interpretation of BG
may make some specific assumptions about interactive beliefs more or
less plausible (think about the common prior assumption), but the
computation of equilibria is not affected.
However, the interpretation seems to matter when it comes to define the
appropriate concept of rationalizability, assuming that BG is transparent
to the players. The crucial point is the following: in an incomplete
information game, the rationality and strategic reasoning of a player need
to be evaluated according to his type. Specifically, we consider the so called
interim stage and, for each action ai and type ti , we ask: Is ai a justifiable
choice for ti ?25 On the other hand, in a game with asymmetric information
about an initial chance move, rationality and strategic reasoning need to
be assessed at the ex ante stage. For each strategy i : Ti Ai , we ask:
Is i a justifiable strategy for player i? By answering to these different
questions, game theorists obtained two different concept of rationalizability
for Bayesian games.
Definition 30. Let BG be a Bayesian game. An action ai Ai is interim
rationalizable for type ti Ti of player i I, if ai is rationalizable for ti
in the interim strategic form IS(BG). A function i : Ti Ai is ex ante
rationalizable if i is a rationalizable strategy of the ex ante strategic form
AS(BG).
Lemma 9. If a strategy i is justifiable in the ex ante strategic form,
then for every type ti the action ati = i (ti ) is justifiable in the interim
strategic form.
Proof. First observe that, for each type ti , according to eq. (7.3.2),
the actions of the other types t0i of player i are completely irrelevant
for the determination of the interim payoff. Hence, one can re-define
the conjecture of type ti in the interim strategic form as a probability
distribution over the action profiles of the types of players j 6= i. Notice
25

Recall that we say that a choice is justifiable if it is a best reply to some belief. A
choice is rationalizable if it survives the iterated elimination of non justifiable choices.

142

7. Games with incomplete information

that such action profiles coincide with the strategy profiles of the opponents
of player i in the ex ante strategic form:


(atj )j6=i,tj Tj j6=i ATj = j6=i j = i .
Hence, the set of conjectures of any type/player ti in the interim strategic
form coincides with the sets of conjectures of player i in the ex ante
strategic form. To be consistent with the notation used for games of
complete information, we denote by Ui (i , i ) and uti (ai , i ) the expected
payoffs of player i and of type ti in the corresponding strategic forms, given
conjecture i (i ).
Let us fix arbitrarily a strategy i and a conjecture i . We now prove
that i is a best reply to i if and only if, for every type ti Ti , action
i (ti ) is a best reply to i . This implies the thesis. Fix i (i ). For
every strategy i , the expected payoff of i given i is
Ui (i , i ) =

i (i )Ui (i , i )

i (i )

ti

pi [ti ]

X
i

ti

pi [ti ]

i (i )

pi [0 , ti |ti ]ui (0 , i (ti ), i (ti ), i (ti ), i (ti ))

0 ,ti

pi [0 , ti |ti ]ui (0 , i (ti ), i (ti ), i (ti ), i (ti ))

0 ,ti

pi [ti ]uti (i (ti ), i ).

ti

Recall: pi [ti ] > 0 for each ti Ti . It follows that

i arg max Ui (i , i ) = arg max
i

pi [ti ]uti (i (ti ), i )

ti

if and only if
ti Ti , i (ti ) arg max uti (ai , i ).
ai

The average of the expected payoffs for the different types of i is maximized
if and only if26 the expected payoff of every type of i is maximized. 
26

Of course, the only if part holds because we assume that pi [ti ] > 0 for each ti Ti .

7.5. Rationalizability in Bayesian Games

143

Corollary 3. If a strategy i i is rationalizable in the ex ante strategic

form, then for every type ti Ti the action i (ti ) is rationalizable for ti in
the interim strategic form.
From the previous considerations it looks like interim rationalizability is
the appropriate solution concept if we interpret the mathematical structure
BG as an interactive situation with incomplete information, where the ex
ante stage is only a useful notational device and therefore the states
represent only possible configurations of private information and beliefs.
Ex ante rationalizability is instead the appropriate solution concept if we
interpret BG as a an interactive situation with asymmetric information
over some initial move by nature (which directly affects the payoffs only
through the payoff-relevant components of ).
Corollary 3 states that ex ante rationalizability is at least as strong a
solution concept as interim rationalizability. It can be shown by example
that it is actually stronger. The formal reason may be understood already
from the proof of Lemma 9. Assume that, for every ti , action i (ti )
is justifiable. Then there exists a profile of conjectures (ti )ti Ti (with
ti (i )) such that i (ti ) is a best reply to ti for each ti . But this
does not imply that there exists a unique conjecture i (i ) such that
i (ti ) is a best reply to i for every ti . The difference between the two
solutions concept is clearly illustrated by the following numerical example.
Example 22. Let = { 0 , 00 }; player 1 (Rowena) knows the true state,
1 ( 0 ) 6= 1 ( 00 ), whereas player 2 (Colin) does not, 2 ( 0 ) = 2 ( 00 ), and
considers the two states equally likely: p2 ( 0 ) = p2 ( 00 ) = 21 . The payoff
functions are as in the matrixes below (note that the payoff of Rowena is
independent of the state, this simplifies the example):
0

00

3, 3

0, 2

3, 0

0, 2

2, 0

2, 2

2, 0

2, 2

0, 0

3, 2

0, 3

3, 2

First, observe that every action by Colin is justifiable. In particular,

c is justifiable by the conjecture that Rowena chooses a if 0 and b if

144

7. Games with incomplete information

00 . It is easy to verify that for each type of Rowena all the actions are
justifiable by some conjecture about Colin. Hence, if choices are evaluated
at the interim stage it is not possible to exclude any action, everything
is interim rationalizable. Instead if choices are evaluated at the ex ante
stage, then the strategy 1 ( 0 ) = a, 1 ( 00 ) = b, denoted by ab, can be
excluded. Indeed, since Rowenas payoff does not depend on the state, ab
could be a best reply to some conjecture 1 only if both a and b were
best replies to 1 . But if Rowena believes 1 (c) 12 , then b is not a best
reply; if Rowena believes 1 (c) 21 , then a is not a best reply. Therefore
strategy ab is not justifiable by any 1 . The same argument shows that
ba is not justifiable either. The ex ante justifiable strategies of Rowena
are: aa,am,ma,mm,bb,bm,mb.27 Given any belief 2 such that 2 (ab) = 0,
action c yields an expected payoff less or equal than 32 (check this); action
d instead yields 2 > 32 . Then, the only rationalizable action for Colin in the
ex ante strategic form is d. It follows that the only rationalizable strategy
of Rowena in the ex ante strategic form is bb.
Although the gap between ex-ante and interim rationalizability was
at first just accepted as a fact, it should be disturbing. Conceptually,
rationalizability is meant to capture the assumptions of rationality (i.e.,
expected utility maximization) and common belief in rationality given
the background transparency of the Bayesian Game BG. Since both
strategic forms are based on the same Bayesian game and since ex-ante
maximization is equivalent to interim maximization,28 why does not exante rationalizability coincide (in terms of behavioral predictions) with
interim rationalizability? Which features of the strategic form create
the gap highlighted in Example 22? Before providing an answer to
these questions, we introduce another puzzling feature of rationalizability
in Bayesian environments: the dependence of interim rationalizability
on apparently irrelevant details of the state space. To understand this
problem, consider the following example.
Example 23. [Dekel et al. 2007] Rowena and Colin are involved in a
27

Given that Rows payoff does not depend on , those strategies that select different
actions for the two states are justifiable only by beliefs that make Row indifferent between
these two actions; the strategies am and ma are among the best replies to the belief
1 (c) = 32 , the strategies bm and mb are among the best replies to the belief 1 (c) = 31 .
28
This equivalence holds under the assumptions that for every player i, all types ti
have positive probability.

7.5. Rationalizability in Bayesian Games

145

betting game. Each player can decide to bet (action B ) or not to bet
(action N ). There is an unknown parameter
0 0 = {00 , 000 } and players

have no private information (i = i for every agent i). Rowena (Colin)
wins if both players bet and 0 = 00 (0 = 000 ). The decision to bet entails
a deadweight loss of $4 independently of the opponents action; if both
agents bet, the loser gives $12 to the winner. The corresponding game
can be represented as follows:
with payoff uncertainty G

00

000

8, 16

4, 0

16, 8

4, 0

0, 4

0, 0

0, 4

0, 0

Let us assume that there is common belief that each agent assigns
probability 21 to each payoff state. Thus, each player i has only one
hierarchy of beliefs about 0 .
The simplest state space representing this situation is the following:
= { 0 , 00 } ,,0 ( 0 ) = 00 , 0 ( 00 ) = 000 and for every agent i, Ti = {ti },
functions i and i are trivially defined and pi [ 0 ] = 12 . In this case, the
ex-ante and the interim strategic forms coincide and are represented by
the following game:

4, 4

4, 0

0, 4

0, 0

It is immediate to see that betting is not rationalizable. Thus, N is

the only interim rationalizable action.


Now, consider an alternative state space: = 1 , 2 , . . . , 8 ,

0 () =

0
0

000



if 1 , 2 , 3 , 4

,
if 5 , 6 , 7 , 8

146

7. Games with incomplete information

for every agent i, Ti = {t0i , t00i }, i is trivially defined,



0
t1 if 1 , 2 , 5 , 6
1 () =
 3 4 7 8 ,
00
t1 if , , ,

2 () =

0
t2



if 1 , 3 , 5 , 7

t002



if 2 , 4 , 6 , 8

and p1 = p2 = p:
State of the world,

p []

1
4

2
0

1
4

1
4

1
4

One can easily check that this common prior induces a distribution over
residual uncertainty and types which can be represented as follows:
00

t02

t002

000

t02

t002

t01

1
4

t01

1
4

t001

1
4

t001

1
4

and that for both types of both players the following holds: (i) the player
has no private information, (ii) he assigns probability 21 to each state 0 ,
and (iii) there is common belief of this. Thus, this second state space
represents exactly the same belief hierarchies than the former one: each i
assigns probability 12 to each 0 , each i is certain that i assigns probability
1
2 to each 0 , and so on. If we write down the ex-ante strategic form derived
from this state space, we get the following game:
BB

BN

NB

NN

BB

4, 4

4, 2

4, 2

4, 0

BN

2, 4

1, 5

5, 1

2, 0

NB

2, 4

5, 1

1, 5

2, 0

NN

0, 4

0, 2

0, 2

0, 0

7.5. Rationalizability in Bayesian Games

147

Notice that, in this game, strategy NB (BN ) is a best response for

Rowena to the conjecture that Colin is playing strategy NB (BN ), while
NB (BN ) is Colins best response to the conjecture that Rowena is
playing strategy BN (NB ). Thus, the set {BN,NB}{BN,NB} has the
best response property and, consequently, betting is ex-ante and, thanks
to Corollary 3, interim rationalizable.
In Example 23, the two state spaces generate two different Bayesian
Games representing the same information and belief hierarchies over ;
thus, they seem to represent the same background common knowledge
Then, should
and beliefs concerning the game with payoff uncertainty G.
not we expect them to lead to the same interim rationalizable prediction?
Why is this intuition false? Why does the choice of one type space over
the other matter?
It turns out that these two puzzles are related: they both depend
on independence assumptions that are implicitly made through the
construction of the strategic forms. Once these assumptions are removed
and the proper (ex ante and interim) notions of correlated rationalizability
are defined, the puzzles disappear: (i) the set of interim correlated
rationalizable actions is not affected by the choice among different state
spaces representing the same information and belief hierarchies over ,
and (ii) there is no gap between interim correlated and ex-ante correlated
rationalizability.
First, let us focus on the puzzle highlighted by Example 23. Consider
the second state space and recall that in such state space each player has
two types representing the same belief hierarchy over .29 In this case,
if type t01 conjectures that t02 plays B and t002 plays N, this conjecture,
together with the common belief prior p, will lead t01 to believe that
Colin will bet only when the residual uncertainty is 00 ; this belief will
justify the decision of type t01 to bet.30 To put it differently, in this state
space, Rowenas types hold beliefs in which Colins type is correlated
with residual uncertainty and this can, in turn, introduce correlation
between Colins behavior, 2 (), and residual uncertainty, 0 . On the
contrary, this correlation is not possible with the first state space in
which each belief hierarchy is represented only by one type; this happens
because the construction of the interim strategic form imposes an implicit
29
30

In exaple 23, players have no private information.

It is easy to see that a similar reasoning holds for any other type.

148

7. Games with incomplete information

independence constraint on players beliefs: each player has to regard his

opponents behavior, 2 (), as independent of the residual uncertainty,
0 , conditional on the opponents types. If there are relatively few types
representing the same information and belief hierarchy, this independence
constraint may bind, preventing players from holding correlated beliefs
and restricting the set of rationalizable actions.31 To address this
problem, Dekel et al.
(2007) define a solution concept, Interim
Correlated Rationalizability, in which beliefs are explicitly allowed to
exhibit correlation between opponents actions and residual uncertainty.
Since the independence restriction imposed by the construction of the
interim strategic form is not related to the assumption of rationality,
we can immediately conclude that interim correlated rationalizability
represents a step forward in the search for a solution concept characterized
by Rationality and Common Belief in Rationality (RCBR) given the
background transparency of the given Bayesian game BG. Actually, it
can be shown that interim correlated rationalizability fully characterizes
the behavioral implications of these epistemic assumptions. To provide a
formal definition of this solution concept, fix a Bayesian Game BG =
hI, , 0 , 0 , (i , Ti , Ai , i , i , pi , ui )iI i . Then, for every i, ti and i
( Ai ) let
X


i (, ai ) ui (0 () , i (ti ) , i (i ()) , ai , ai ) .
ri ti , i = arg max
ai Ai

,ai

The set of interim correlated rationalizable actions is iteratively defined

as follows: for every i and ti , let ICRi0,BG (ti ) = Ai and for each k 1, let

ai Ai : i ( Ai ) , i :
0 Ti (Ai ) ,

1) ai ri ti , i
ICRik,BG (ti ) =
k1,BG

2) (0 , ti ) , i (0 , ti ) [ai ] > 0 ai ICRi

(ti )

3) (, ai ) , i (, ai ) = p [ | ti ] i (0 () , i ()) [ai ]
k1,BG
where ICRi
(ti ) = j6=i ICRjk1,BG (tj ) .
In the previous definition, function i (0 , ti ) represents the
conjecture of player i concerning the behavior of his opponents given their
31

This problem would not arise if for every player i and every information and belief
hierarchy he may hold, there were at least |Ai | types representing this information and
belief hierarchy. Stricter requirement can be provided, but this would go beyond the
scope of these notes.

7.5. Rationalizability in Bayesian Games

149

types are ti and residual uncertainty is 0 . Since such functions depend

both on 0 and on ti , representing conjectures with such functions allows
to introduce correlation between ai and 0 even after conditioning on ti ;
this correlation is not allowed by interim rationalizability.32
Definition 31. Fix a Bayesian game BG. For player i I, action
ai Ai and type tiT Ti , ai is interim correlated rationalizable for ti if
ICRik,BG (ti ) .
ai ICRiBG (ti ) =
k0

Since interim correlated rationalizability allows for correlation between

opponents behavior and residual uncertainty, it does not impose any
implicit independence restriction and it captures the assumptions of
players rationality and common belief in rationality given the transparency
of the Bayesian game BG (as usual, the k-th step in the iterative definition
of interim correlated rationalizability corresponds to the k 1-th level of
mutual beliefs in rationality). As a consequence, given a game with payoff
the set of interim correlated rationalizable actions of type
uncertainty, G,
ti depends only on the information and belief hierarchy entailed by ti .
From now on, the latter will be denoted by pi (ti ). The following Theorem
formalizes this result.


= I, 0 , (i , Ai , ui )
Theorem 16. Fix G
iI and consider two Bayesian
Games based on it: BG0 = hI, 0 , 0 , 00 , (i , Ti0 , Ai , i0 , 0i , p0i , ui )iI i and
BG00 = hI, 00 , 0 , 000 , (i , Ti00 , Ai , i00 , 00i , p00i , ui )iI i . Let t0i Ti0 and t00i
0
00
Ti00 . Then, if pi (t0i ) = pi (t00i ) , ICRiBG (t0i ) = ICRiBG (t00i ) .
Proof. Take t0i Ti0 and t00i Ti00 and suppose pi (t0i ) = pi (t00i ) . We will
0
prove by induction that for every i and for every k 0 ICRik,BG (t0i ) =
00
ICRik,BG (t00i ) . To simplify notation, we will not specify the dependence
of the set of interim correlated rationalizable actions on Bayesian Game
(notice that this is also justified by the statement of the Theorem). The
result is trivially true for k = 0. Now, suppose that ICRis (t0i ) = ICRis (t00i )
for every s k 1. We need to show that ICRik (t0i ) = ICRik (t00i ) .
Take any ai ICRik (t0i ) . By definition, we can find it0 (0 Ai )
i


0
and t0i : 0 Ti (Ai ) such that (i) ai ri t0i , it0 , (ii)
i

32
Indeed, it is possible to show that the set of interim rationalizable actions for type
ti is equivalent to the set of actions obtained through an iterative construction similar
to the one we just described, but in which i : Ti (Ai ) .

150

7. Games with incomplete information

k1
t0i (0 , ti ) [ai ] > 0 implies ai ICRi
(ti ) and (iii) for every pair


i
0
0
0 () [a ] .
, ai , t0 (, ai ) = p [ | ti ] t0i 0 () , i
i
i



k1
Let Pi
=
0j (tj ) , pk1
: tj Tj0 be the set of possible
(tj )
j
j6=i

profiles of private information and k 1-th order beliefs that agents

k1
k1
other than i can have.
For every 0 0 , i
Pi
and ai Ai , define [0 ]BG0 = {(, ai ) 0 Aih: 0 ()
i = 0 } ,




k1
0
0
0
[ai ]BG0 =
, ai Ai : ai = ai
and i
=
BG0




k1
(, ai ) 0 Ai : 0j (j ()) , pk1
= i
. We will
(j ())
j
j6=i
h
i
h
i
 
k1
k1
write 0 , i
, ai
to denote [0 ]BG0 i
a0i BG0 ,
0
0
BG
BG
h
i
h
i


k1
k1
0
0 , i
to
denote
[
]

and

,
a
to denote
0
0
0
i
BG
i
BG0
0
BG
BG0
 0 
k1,BG0
[0 ]BG0 ai BG0 . Finally, for every 0 , let
i
() =



 
j0 ()
and, similarly, for every 00 , let
0j j0 () , pk1
j





k1,BG00

i
() = 00j j00 () , pjk1 j00 () .
k1
k1
Then for every 0 0 and i
Pi
, define
h
i
X
k1
=
t0i (, ai ) .
t0i 0 , i
k1
(,ai )[0 ,i
]BG0
h
i
k1
t0i 0 , i
is the probability that type t0i assigns to the event residual
uncertainty is 0 and other agents have
private
h
i information and k 1-th
k1
k1
order beliefs given by i . If t0i 0 , i > 0, let
P
t0i (, ai )

 
k1 0
(,ai )[0 ,i
,a ]
k1
h i BG0i
;
ti 0 , i
a0i =
k1
t0i 0 , i
h
i
 
k1
0
if ti 0 , i
= 0, take any t0i =
t0j
Ti
such that
j6
=
i
  
 
k1
0j t0j , pk1
t0j
= i
and let
j
j6=i

k1
t0i 0 , i
[ai ] =

k1 0
|ICRi
(ti )|

k1 0
if ai ICRi
ti

otherwise

7.5. Rationalizability in Bayesian Games

151

(by the inductive hypothesis, the definition of t0i does not depend on the
actual choice of t0i ).
For every (, ai ) 00 Ai define:




k1,BG00
t00i (, ai ) = p | t00i t0i 000 () ,
i
() [ai ] ,
where the previous expression is well defined since BG0 and BG00 are based
and pi (t0 ) = pi (t00 ) . Moreover, since pi (t0 ) = pi (t00 ):
on the same G
i
i
i
i
h
i
X
k1
t0i 0 , i
=
t0i (, ai ) =
k1
(,ai )[0 ,i ]BG0
hn
o
i
k1,BG0
k1
= p 0 : 0 () = 0 ,
i
() = i
| t0i =
hn
o
i
k1,BG00
k1
= p 00 : 00 () = 0 ,
i
() = i
| t00i .


Then, for every 0 , a0i , we have
X

:00
0 ()=0

(,ai )[0 ,a0i ]BG00

hn



  0 


k1
00
i
()
ai =
p | t00i t0i 000 () ,
i

t00i (, ai ) =
000 ()

= 0 ,

k1,BG00

() =

k1
i

t00i


t0i

k1
0 , i



k1
k1
i
Pi

P
=

h
i
k1
t0i 0 , i

k1
k1
i
Pi

ti (, ai )
i
=

k1 0
(,ai )[0 ,i
,ai ]BG0

k1
t0i 0 , i

t0i (, ai )

0 ,a0i BG0

(,ai )[

where the definition of [0 , ai ]BG00 is analogous to the one of [0 , ai ]BG0 .

Thus, t0 and t00 have
marginal distribution over 0 Ai
 the same

implying that ai ri t00i , it00 . Furthermore we already constructed a
i

00 (A ) (namely, 0 ) such that for every

function i : 0 Ti
i
ti


00 () [a ] . Finally, by
(, ai ) , it00 (, ai ) = p [ | t00i ] i 000 () , i
i
i
construction and by the inductive hypothesis, t00i (, ai ) > 0 implies


a0i =

152

7. Games with incomplete information

k1
00 () . We conclude that a ICRk (t00 ). The statement
ai ICRi
i
i
i
i
of the Theorem follows by induction. 
Notice that, by definition, the conditional independence restriction
implicit in the construction of the interim strategic form has no bite
when there is distributed knowledge of the state (0 is a singleton); in
this case it is easy to verify that interim correlated rationalizability is
equivalent to interim rationalizability. On the contrary, suppose that there
is some residual uncertainty (|0 | > 1) and that we insist on using interim
rationalizability. A natural question arises: can we at least provide an
expressible characterization of this solution concept and, in particular, of
the independence restriction implied by it? A characterization is deemed
expressible if it can be stated in a language based on primitives (that is,
elements contained in the description of the game with payoff uncertainty,
and terms derived from them (such as, hierarchies of beliefs over these
G)
primitives).
The answer to the previous question is affirmative only in some
particular cases and the reason goes to the very heart of Harsanyis
approach. To understand why, recall that interim rationalizability requires
players to regard opponents actions as independent of the residual
uncertainty conditional on their types. But what is a type? A type is a
self-referential object: it is a private information and a belief over residual
uncertainty and other players types. Thus, unless we can establish a 1-to-1
mapping between types and agents information and belief hierarchies over
primitives, the conditional independence assumption is not expressible.
Obviously, in order to assess the existence of such a mapping, we can use
both payoff-relevant and payoff-irrelevant primitives. Thus, even though
two types represent the same private information and belief hierarchy
over payoff-relevant parameters, they can still be distinguished by the
information and beliefs hierarchies over payoff-irrelevant elements that
they capture. Whenever this is the case, the conditional independence
assumption is expressible and we can show that, given the transparency of
BG, interim rationalizability characterizes the behavioral implications of
the following epistemic assumptions: (R) players are rational, (CI) their
beliefs satisfy independence between of the opponents behavior and the
residual uncertainty conditional on the information and belief hierarchy
over primitives of the opponents, and (CB(R CI)) there is common belief
of R and CI.

7.5. Rationalizability in Bayesian Games

153

Notice that, insofar players may believe that (i) their opponents
behavior may depend on payoff-irrelevant information, and (ii) this
information is correlated with 0 , the information and belief hierarchies
over payoff-irrelevant parameters may still be strategically relevant. For
instance, in example 23, Rowena may be superstitious and believe
that the payoff-relevant state, 0 , is correlated with a particular dream
Colin may have had (payoff-irrelevant information), which could also affect
Colins decision to bet. Similarly, in example 15, a firm may believe
that the quantity produced by its competitor depends on the analysis
carried out by its marketing department and that this (payoff-irrelevant)
information may be correlated with the actual position of the demand
function (0 ).
As a special case, in which interim rationalizability admits an
expressible characterization, we can consider Bayesian Games with
information types.
Definition 32. Fix a game with payoff uncertainty
= hI, 0 , (i , Ai , ui )iI i .
G
A Bayesian Game
BG = hI, , 0 , 0 , (i , Ti , Ai , i , i , pi , ui )iI i
has information types if Ti = i .33 In this case, the
based on G
functions (i () : Ti i )iI are trivially defined (i = Idi ).
Focusing on Bayesian games with information types, one can show that
interim rationalizability characterizes the behavioral implications of the
following epistemic assumptions: (R) rationality, (CIP I) independence
between players actions and residual uncertainty conditional on players
private information, and (CB(R CIP I)) common belief of R and CIP I.
Besides being interesting per se, Bayesian games with information types
are also widely used in the study of many relevant problems in information
economics, such as adverse selection, moral hazard, mechanism design.
Now, let us turn to the other puzzling result: the gap between ex-ante
and interim rationalizability. As already pointed out, the gap highlighted
by example 22 arises because the interim strategic form regards different
33

More formally, the requirement is that Ti and i are isomorphic.

154

7. Games with incomplete information

types as different agents and, as a consequence, different types playing in

Rowenas role may hold different beliefs concerning Colins behavior (one
of Rowenas type can believe Colin will play c with probability greater
than 12 , while the other type can believe that this will happen only with
probability lower than 21 ); since Rowenas conjecture concerning Colins
behavior can vary with her own type, her beliefs can exhibit correlation
between the state of the world, (which determines residual uncertainty
and players type) and Colins action. Instead, the ex-ante strategic form
implicitly requires Rowena to regard Colins strategy as independent from
the state of the world.34 To understand this, let us analyze the strategicform expected payoff of player i when he plays strategy i and holds
conjecture i (i ) about the strategies of other players:
X
Ui (i , i ) :=
i (i )Ui (i , i )
i i

i (i )

i i

pi ()ui (0 () , i (i ()), i (i ()), i (i ()), i (i ()))

i (, i , i ),
pi ()i (i )U

i i

i (, i , i ) denote the payoff of i when the state of the world

where we let U
is and profile (i , i ) is played. As the formula shows, the expected
payoff of i is computed under the assumption that the probability of each
pair (, i ) is the product of the marginal probability of , pi () and
the marginal probability of i , i (i ). This means that i regards , the
choice of nature, as independent of i , the choice of i.
This implicit independence restriction reduces the set of Rowenas
admissible beliefs and makes ex-ante rationalizability more demanding
(hence, stronger) than interim rationalizability.
Battigalli et al.
(2011) address this problem by defining Ex-Ante Correlated
Rationalizability, a solution concept that allows correlation, according to
players subjective beliefs, between the initial move by nature and players
strategies. To introduce ex-ante correlated rationalizability, consider a
34
If we interpret as the choice made by an external and neutral player called Nature,
the independence restriction on beliefs is equivalent to requiring independence between
the strategy chosen by Colin and the one chosen by Nature.

7.5. Rationalizability in Bayesian Games

155

Bayesian game in which the state space has information types; although
such restriction is not necessary from the mathematical point of view,
the ex-ante strategic form interprets types as actual information received
by players concerning the initial move by nature and, consequently, the
assumption of Bayesian games with information types is the most, if not
the only, reasonable one. For every i and i ( i ) let
X


ri i = arg max
i (, i ) i (, i (i ()) , i (i ())) .35
i i

,i

Then we can inductively define the set of ex-ante correlated rationalizable

strategies as follows: for every i, ACRi0,BG = i and for every k 1,

i i :
i ( i ) ,

1) i ri i , 2) marg i = p,
ACRik,BG =

k1
3) i (, i ) > 0 = i ACRi
k1
where ACRi
= j6=i ACRjk1 .

Definition 33. Fix a Bayesian game with information types, BG. A

strategy for player i is ex-ante correlated rationalizable if i ACRiBG :=
T
ACRik,BG (i ) .
k0

From the iterative definition ex-ante correlated rationalizability, it is

easy to see that the justifying belief i introduces correlation between
the state of the world and opponents behavior. Insofar different can
be associated with different types ti = i () , which can hold different
beliefs about the behavior of the other players, this correlation can also
be introduced by interim correlated rationalizability. On the contrary,
such correlation is not allowed by ex-ante rationalizability as an agent
is constrained to hold the same conjecture over opponents strategies
independently of her actual type.
Thus, interim correlated rationalizability and ex-ante correlated
rationalizability avoid any independence restriction implicitly entailed by
the interim and ex-ante strategic form; as a consequence, both these
solution concepts are characterized only by the epistemic assumptions of
rationality and common belief in rationality given the background common
knowledge of BG. Then, it should not come as a surprise that, if we use
these solution concepts, the gap highlighted in example 22 disappears and
the following theorem holds.

156

7. Games with incomplete information

Theorem 17. Let BG = hI, , 0 , 0 , (i , Ti , Ai , i , i , pi , ui )iI i be a

Bayesian game with information types. Then, for every i,


ACRiBG = i i : ti Ti , i (ti ) ICRiBG (ti ) .
Proof. We will prove that a stronger result holds, namely that for
every i and every k 0
n
o
ACRik,BG = i i : ti i , i (ti ) ICRik,BG (ti )
The result is trivial for k = 0. Now suppose that the result holds for every
s k 1; we will show that it holds also for k.
Take an agent i and an arbitrary i ACRik,BG . By definition
we can find a rationalizing belief i ( i ). To simplify the
notation, let [0 ] = {(, i ) : 0 () = 0 } , [t] = {(, i ) : () = t},
[ai ] = {(, i ) : i (i ()) = ai } , [0 , t, ai ] = [0 ] [t] [ai ] and
p [0 , t] = p [{ : : 0 () = 0 , () = t}] . For every ti Ti , define a
function ti : 0 Ti (Ai ) in the following way: if p [0 , t] > 0 ,
P
ti (0 , ti ) [ai ] =

(,i )[0 ,t,ai ] (, i )

p [0 , t]

if p [0 , t] = 0,

ti (0 , ti ) [ai ] =

1




k1,BG
(ti )
ICRi

k1,BG
if ai ICRi
(ti )

k1,BG
if ai
/ ICRi
(ti )

For every ti and , ai , define

ti (, ai ) = p [ | ti ] ti (0 () , i ()) [ai ]
Observe that:
X

(, i ) = p [0 , t] ti (0 , ti ) [ai ]

(,i )[0 ,t,ai ]

= p [ti ] p [0 , ti | ti ] ti (0 , ti ) [ai ] ,

7.5. Rationalizability in Bayesian Games

157

Thus:
X

i (, i ) i (, i (i ()) , i (i ())) =

,i

(, i ) i (, i (i ()) , i (i ())) =

0 ,t,ai (,i )[0 ,t,ai ]

X
ti

p [ti ]

X
0 ,ti

p [0 , ti | ti ]

ti (0 , ti ) [ai ] ui (0 , i (ti ) , i (ti ) , i (ti ) , ai )

ai

Since i ri i and p [ti ] > 0, we conclude that i (ti ) ri (ti ,
ti ) .
Finally, if p [0 , t] = 0, by construction ti (0 () , ti ) [ai ] > 0 only if
k1,BG
ai ICRi
(ti ) . If instead, p [0 , t] > 0, the same result follows
from the definition of ex-ante correlated rationalizability and the inductive
hypothesis. Thus, for every ti , we can conclude that i (ti ) ICRik,BG (ti )
and, consequently, that:
n
o
ACRik,BG i i : ti i , i (ti ) ICRik,BG (ti ) .

Now we will prove the other inclusion. For every ti Ti , let (ati , ti ) be
a pair such that ati ICRik,BG (ti ) . Thus for every ati we can find a
rationalizing belief ati ( Ai ) and a function ati : 0 Ti
(Ai ) satisfying the properties stated in the iterative definition of interim
correlated rationalizability. Now define the belief
i ( i ) as
follows:
p () ai () (0 () , i ()) [i (i ())]
o

i (, i ) = n

k1,BG
0
0
: i (i ()) = i (i ()) 
 i ACRi
if
o
n
0 ACRk1,BG : 0 ( ()) = ( ()) 6= and
i (, i ) =
i
i i
i i
i
0 otherwise. It is immediate to see that
i (, i ) > 0 only if i
k1,BG
0
ACRi
. Now, let [, ai ] = {( , i ) : 0 = , i (i ()) = ai } .

158

7. Games with incomplete information

n
o
0 ACRk1,BG : 0 ( ()) = a
i
6=
i
i
i
i

Thus, if

i 0 , i =

( 0 ,i )[,ai ]

P
=

p () ai () (0 () , i ()) [ai ]
n
o
=
 0

k1,BG
0 ( ()) = a
: i
 i ACRi
i
i 

k1,BG
i ACRi
:i (i ())=ai

= p () ai () (0 () , i ()) [ai ] .
Also
notice
that
every
n
ofor
0 ACRk1,BG : 0 ( ()) = a
(, ai ) , if i
=
,
the
inductive
i
i i
i
k,BG
hypothesis implies that ai
/ ICRi
(i ()) and consequently
ai () (0 () , i ()) [ai ] = 0. Therefore for every (, ai ) :
X

i (, i ) = p () ai () (0 () , i ()) [ai ]
( 0 ,i )[,ai ]

and
X

i (, i ) =

i (, i ) =

ai ( 0 ,i )[,ai ]

p [] ai () (0 () , i ()) [ai ] = p []

ai

Notice that we have already proved two of the properties in the definition
of ex-ante correlated rationalizability. Finally, for every i i
X

i (, i ) i (, i (i ()) , i (i ())) =
,i

p []ai () (0 () , i ()) [ai ] ui (0 () , i (i ()) , i (i ()) , i (i ()) , ai )

,ai

Since p [ti ] > 0 for every ti , the strategy i defined by i (ti ) = ati for every


ti is such that i ri
i . We conclude that i ACRik,BG . Since pairs
(att , ti ) were chosen arbitrarily, we can write:
n
o
ACRik,BG i i : ti i , i (ti ) ICRik,BG (ti ) .
The statement of the Theorem follows by induction. 

7.6. The Electronic Mail Game

7.6

159

The Electronic Mail Game

Consider the following incomplete information game where player 1

(Rowena) knows the true payoff matrix, whereas 2 (Colin) does not know
it. The probabilities and the payoffs are represented below:

(1 ) :

M, M

1, L

L, 1

0, 0
L > M > 1,

0, 0

1, L

L, 1

M, M

< 1/2

If we had complete information, (a, a) would be the dominant strategy

equilibrium for matrix , and (b, b) would be the Pareto efficient
equilibrium for the matrix . The second matrix has also another
equilibrium, (a, a).
Notice that the game has distributed knowledge; thus, our previous
analysis implies that the set of interim rationalizable actions is equivalent
to the set of interim correlated rationalizable actions. In particular, it can
be easily verified that any Bayesian game representing this situation would
have only one interim rationalizable profile: (aa, a). As a matter of fact,
for type of Rowena a is dominant. If Colin believes that Rowena is
rational, so that type chooses a, then a2 = a yields a larger expected
utility than a2 = b. More precisely let 2 (aa, ab, ba, bb) be Colins
conjecture about Rowenas strategy and let u2 (2 , a2 ) be the expected
utility of action a2 {a, b} given conjecture 2 . Since 2 (ba) = 2 (bb) = 0
(as certainly chooses a), we obtain
u2 (2 , a) = (1 )u2 ( , a, a) + 2 (aa)u2 ( , a, a) + 2 (ab)u2 ( , b, a)
(1 )M
> (1 )L + M
(1 )u2 ( , a, b) + 2 (aa)u2 ( , a, b) + 2 (ab)u2 ( , b, b)
= u2 (2 , b).
Hence, the only rationalizable action for Colin is a. This implies that also
for type the only interim rationalizable choice is a.

160

7. Games with incomplete information

Consider now the following more complex variation of the game.36

Obviously the players would find it convenient to communicate and
coordinate on (a, a) if = and on (b, b) if = . Let us then consider
the following form of communication via electronic mail. Rowen and Colin
sit in front of their respective computer screens. If the payoff-type of
Rowena is , her computer, C1 , automatically sends a message to the
computer of Colin, C2 . Furthermore, if computer Ci receives a message,
it automatically sends a confirmation message to computer Ci . However,
any time a message is sent,it gets lost with probability and thus it does
not reach the receiver.
This information structure yields a corresponding Bayesian game. A
generic state of the world is given by a pair of numbers = (q, r) where
q = t1 is the number of messages sent by C1 and r = t2 is the number of
messages received (and therefore also sent) by C2 . Hence, the set of states
of the world is given by

= {(0, 0), (1, 0), (1, 1), (2, 1), (2, 2), (3, 2), (3, 3), ...}
= {(q, r) N N : q = r or q = r + 1} .

If the state is (q, q 1) it means that the last message (among those
sent either by C1 or by C2 ) has been sent by C1 . If the state is (r, r)
(with r > 0) it means that the last message sent by C1 has reached C2 ,
but the confirmation by C2 has not reached C1 . The signal functions are
1 (q, r) = q, 2 (q, r) = r. The function that determines is 1 (0) = ,
1 (q) = if q > 0.37 There is a common prior given by p(0, 0) = (1 ),
p(r + 1, r) = (1 )2r , p(r + 1, r + 1) = (1 )2r+1 for any r 0 (that
is, p(q, r) = (1 )q+r1 for any (q, r) \ {(0, 0)}). This information
is summed up in the following table:
The resulting beliefs, or conditional probabilities, are then determined

36
See [31] (or the textbook by Osborne and Rubinstein [28], pp 81-84). The analysis
in terms of interim rationalizability is not contained in the original work.
37
2 is a singleton, hence 2 is a constant.

7.6. The Electronic Mail Game

t1 =q \

t2 =r

161

...

0,

...

1,

(1 )

...

2,

(1 )

(1 )

...

3,

(1 )4

(1 )5

...

4,

5,

...

(1 )

(1 )
8

...
9

(1 )

(1 )

...

...

...

...

...

...

...

...

as follows:
p1 [0|0] = 1,
p1 [r|r + 1] =
p2 [0|0] =
p2 [r + 1|r + 1] =

(1 )2r
1
=
,
2r
2r+1
(1 ) + (1 )

2
1
,
1 +
(1 )2r+1
1
=
(r > 0).
(1 )2r+1 + (1 )2r+2
2

In other words, any player having received a certain number of messages

computes the probability that his confirmation message does not reach
1
the other computer, which is 2
> 12 .
If is very small, the Bayesian game is in some sense close to the
game with common knowledge of . First, notice that in all states (q, r)
with r > 0 both players know that the true state is . Moreover, for any
n and any > 0, there always exists an sufficiently small such that, given
= , the probability that there exists mutual knowledge of degree n of
the true value of is bigger than 1.38 Given that (b, b) is an equilibrium
38

For instance, consider n = 2. For any q 2, in state (q, r) (r = q, or r = q 1)

every player knows that = and every player knows that the other knows it, therefore
there is mutual knowledge of degree two of the true payoff matrix. The probability of
being in one of these states conditional on the event = is
1

p({(1, 0), (1, 1)})

= 1 2 + 2 .

To guarantee that such probability is larger than 1 we need < 1

1 .

162

7. Games with incomplete information

of the complete information game in which = , we could be induced

to think that if is very small, then action b is interim rationalizable for
some type of Rowena and Colin. But the following result shows that this
intuition is incorrect:
Proposition 5. In the Electronic Mail game there exists a unique interim
rationalizable profile: every type of every player chooses action a.
The crucial point in the proof is to realize that when a player i knows
that = , but s/he is sure that i plays a whenever s/he has not
received her/his last message, then i prefers to play a. On the other
hand, it is easy to show that when i does not know that = then a is
the only rationalizable action. It follows by induction that a is the only
rationalizable action in all states. Here is the formal argument:
Proof of Proposition 5. Clearly, the only rationalizable action for
t1 = 0 is a, as this type of Rowena knows that in the true payoff matrix
a is dominant. An argument similar to the one used for the simple case
discussed above shows that the only rationalizable action for t2 = 0 is also
a.
Consider now the rationalizable choices for the types ti = r > 0, that
is for those types for which player i knows that = . We first prove two
intuitive intermediate steps:
(i) If the only rationalizable action for type t2 = r 1 is a, then the
only rationalizable action for type t1 = r is a. Any rationalizable action
for t1 = r needs to be justified by a rationalizable conjecture about Colin
(see Theorem 1). By assumption, such conjecture must assign probability
1 to the set of profiles 2 {a, b}T2 such that 2 (r 1) = a. Hence, the
expected utility for type t1 = r if he chooses a is at least 0 (this value is
realized if type t2 = r also chooses a), whereas the expected utility from
choosing b is at most p1 [r 1|r]L + p1 [r|r]M (this value is realized if type
1
t2 = r chooses b). Since p1 [r 1|r] = 2
> 12 and L > M , it follows that
0 > p1 [r 1|r]L + p1 [r|r]M.
An analogous argument39 shows that:
(ii) If the only rationalizable action for type t1 = r is a, the only
rationalizable action for type t2 = r is also a.
39

Just reverse the roles and modify indexes accordingly.

7.7. Selfconfirming Equilibrium and Incomplete Information

163

From steps (i) and (ii) it follows that, for every r > 0, if the
only rationalizable profile in state (r 1, r 1) is (a, a) then the only
rationalizable profile in state (r, r 1) is (a, a); if the only rationalizable
profile in state (r, r 1) is (a, a) then the only rationalizable profile in
state (r, r) is (a, a). Since we have shown that the only rationalizable
profile in state (0, 0) is (a, a), it follows by induction that (a, a) is the only
rationalizable profile in every state.

7.7

Selfconfirming Equilibrium and Incomplete

Information

As previously argued, the selfconfirming (or conjectural) equilibrium

concept characterizes patterns of behavior that are stable with respect
to learning processes in situations of recurrent interaction, taking into
account players information feedback. The appropriate definition of
self-confirming equilibrium for incomplete information games depends
on the relevant scenario, and in particular on the interpretation of the
mathematical structure we use to represent incomplete information.
The first question to address is whether the set of interacting
players and their characteristics are (a) fixed once and for all (long-run
interaction), or (b) randomly determined in each period via draws from
large populations (anonymous recurrent interaction). In both cases it is
necessary to specify what information players are able to obtain ex post
about the actions and characteristics of co-players. As in Section 5.2.3, we
describe this with feedback functions. As we did earlier, we assume for the
sake of simplicity that the players goal is to maximize the expected payoff
of the current period, without worrying about future payoffs.

7.7.1

Long-run interaction

This is the easiest case to analyze. It is sufficient to consider the game

with payoff uncertainty
G = hI, 0 , (i , Ai , ui : A R)iI i .
The profile of parameters is fixed once and for all; obviously, we continue
to assume that every player i knows only i , that the sets of possible
values for is = 0 (iI i ), and that ex post i gets further

164

7. Games with incomplete information

information about 0 , i and ai according to a feedback function fi .

In general, the signal received may depend also on the parameter , that
is fi : A Mi ; for instance, the signal could be the payoff obtained
by the player.40 Every player i holds a belief i (0 i Ai )
about the residual uncertainty 0 and the opponents information types and
actions. The question is whether, for any given , a profile of actions and
beliefs (ai , i )iI iI (Ai (0 i Ai )) satisfies the conditions
of rationality and confirmed conjectures. These conditions can be obtained
as a simple generalization of definition 21. As in Section 5.2.3 we give the
definition of selfconfirming equilibrium for games with payoff uncertainty
and feedback (G, f ), where f = (fi )iI is the profile of feedback functions.
Definition 34. Fix a game with payoff uncertainty and feedback (G, f )
and a parameter value . A profile of actions and beliefs (ai , i )iI
iI (Ai (0 i Ai )) is a self-confirming equilibrium at
if for every i I:41
(1) ( rationality) ai ri (i , i ),
(2) ( confirmed conjectures):
i ({(0 , i , ai ) : fi (0 , i , i , ai , ai ) = fi ( , a )}) = 1.
In what follows, for each game with payoff uncertainty
G = hI, 0 , (i , Ai , ui : A R)iI i .
and parameter values , we let G denote the game
G = hI, (Ai , ui, : A R)iI i .
Thus, for a game with payoff uncertainty and feedback (G, f ) (where
fi : A Mi for each i I) and parameter values , (G , f ) denotes
the true game with feedback at state of nature .
Remark 26. Fix a game with payoff uncertainty G. For each
and each profile of feedback functions f , every Nash equilibrium of G
is a selfconfirming equilibrium at of the game payoff uncertainty and
feedback (G, f ).
40

However, we need to be aware that the payoff ui does not necessarily represent a
material gain which can be cashed. Therefore in the theoretical analysis we are not
forced not assume that the realization of ui is observed by i (observable payoffs).
41
Recall that ri (i , i ) is the set of actions of i that maximize her expected payoff
given i and i .

7.7. Selfconfirming Equilibrium and Incomplete Information

165

The following remark highlights the fact that in the private-value case
(when ui does not depend on 0 and i ) we get the notion of selfconfirming equilibrium already introduced in section 5.2.3, coherently with
the claim made there according to which the self-confirming equilibrium
concept only presumes that a player knows her own payoff function, not
those of the co-players.
Remark 27. Fix a game with payoff uncertainty and feedback (G, f ), a
parameter value and an action profile a , and suppose that the game
with payoff uncertainty G has private values. Then a is part of a selfconfirming equilibrium at if and only if a is part of a self-confirming
equilibrium of the game with feedback (G , f ).

7.7.2

Anonymous recurrent interaction

Consider the following scenario: There are n heterogeneous populations,

i = 1, ..., n; the fraction of agents with characteristic i is qi (i ) > 0
(assume for simplicity that i is finite); n agents are drawn at random,
one from each population i, and play a static game with payoffs determined
by the parametrized functions ui : A R (i I), where 0 represents
a set of possible values of a random shock that affect agents utility: each
time agents play the game, 0 0 is drawn with probability q0 (0 ) > 0.
This is precisely the scenario that motivated the definition of equilibrium
for a simple Bayesian game (with independent types)
G = hI, 0 , q0 , (i , qi , Ai , ui )iI i

(7.7.1)

where q0 (0 ) , qi (i ), ui : A R (i I).
In this case, however, it is only assumed that each agent in each
population i knows i and ui (i , ), hI, 0 , q0 , (i , qi , Ai , ui )iI i is not
assumed to be common knowledge. Agents play the game recurrently,
each time with a different 0 drawn from 0 and with different coplayers randomly drawn from the respective populations. If the play
stabilizes, at least in a statistical sense, for every j, j , aj , the
fraction of agents in the sub-population j with characteristic j that
choose action aj remains constant; let j (aj |j ) denote this fraction
and write i = (j (|j ))j6=i,j j to denote the profile of fractions for
the populations different from i. By random matching, the long-run
frequency of the profile of others actions and characteristics (i , ai )

166

7. Games with incomplete information

Q
is given by j6=i j (aj |j )qj (j ) and the long-run frequency of random
shock Q
0 and opponents actions and characteristics (i , ai ) is given by
q0 (0 ) j6=i j (aj |j )qj (j ).
The ex post information feedback of an agent playing in role i is
described by a feedback function fi : A M . Hence, in the long
run, an agent of population i with characteristic i that (always) chooses
ai , observes that the frequency each message mi is
X

Pfaii, i ,qi [mi |i ] =

q0 (0 )

j (aj |j )qj (j ).

j6=i

(0 ,i ,ai ):fi (0 ,i ,i ,ai ,ai )=mi

Conjecture i (i Ai ) assigns to message mi (given ai and i ) the

probability
Pfaii, i [mi |i ] =

i (i , ai ).

(0 ,i ,ai ):fi (0 ,i ,i ,ai ,ai )=mi

The conjecture is confirmed in the long run if the subjective probability

of each message is equal to the observed frequency. If ai ri (i , i ) and
i is confirmed, then the agents with payoff-type i choosing ai have no
reason to switch to another action. In the following definition we let ii ,ai
denote a conjecture that justifies action ai for an agent with characteristic
(information type) i .
Definition 35. Fix a simple Bayesian game with feedback (G, f ). A profile
of mixed actions and conjectures:


i (|i ), (i(ai ,i ) )ai Suppi (|i )


iI,i i

(a mixed action i (|i ) for each i and i , and a conjecture for each i, i e
ai Suppi (|i )) is a self-confirming equilibrium of (G, f ) if for every
i I, i i , ai Ai , the following conditions hold:
(1) ( rationality) if i (ai |i ) > 0, then ai ri (i , i ),
(2) ( confirmed conjectures) mi Mi , Pfaii, i [mi |i ] = Pfaii, i ,qi [mi |i ].
Remark 28. The definition needs to be modified and made more stringent
if the distributions qj are known. In this case equilibrium conjectures must

7.7. Selfconfirming Equilibrium and Incomplete Information

167

be consistent with q = (qj )jI0 . For instance, in the two-person case,

margj i = qj .42
Remark 29. The definition of selfconfirming equilibrium at is obtained
as a special case when q( ) = 1.
Remark 30. Every mixed equilibrium of the interim strategic form of G
is a selfconfirming equilibrium profile of mixed actions.
Adapting in the obvious way the definitions of observable payoffs
and own-action independence of feedback of Section 5.2.3, we obtain the
analog of Theorem 12: if (G, f ) satisfies observable payoffs and own-action
independence of feedback, then every selfconfirming equilibrium profile of
mixed actions (i (|i ))iI,i i is also a mixed Bayes-Nash equilibrium of
G, therefore the two equilibrium concepts coincide under these assumptions
about feedback.

42

A special case of equilibrium of this kind obtains when actions are observed
(fi (, a) = a) and conjectures are naive as they do not take into account that
opponents actions depend on their type. For example in the two-agents case without
residual uncertainty, i (j , aj ) = qj (j ) i (aj ) for some i (Aj ). This is called cursed
equilibrium. Such behavior may explain, for example, the so called winners curse in
common value auctions. This refinement of conjectural equilibrium was proposed by
Eyster and Rabin [17], although the link to the conjectural equilibrium concept was not
made explicit.

Part II

Dynamic Games

168

169
In this part we extend the analysis of strategic thinking to interactive
decision situations where players move sequentially. As the play unfolds,
players obtain new information. We consider a world inhabited by
Bayesian players. When a player receives a piece of information that
was possible (had positive probability) according to his previous beliefs,
then he simply updates his beliefs according to the standard rules of
conditional probabilities. When the new piece of information is completely
unexpected, the player forms new subjective beliefs.
The observation of some previous moves by the co-players may provides
information about the strategies they are playing and/or their type. How
such information is interpreted depends, for example, on beliefs about
the co-players rationality. This introduces a new fascinating dimension to
strategic thinking.
We focus on multistage games with observable actions. As in Part
I, we first analyze games with complete information and then move on
to games with incomplete information. Unlike Part I, here we mainly
focus on equilibrium concepts. The reason is that non-equilibrium solution
concepts of the rationalizability/iterated dominance kind are harder to
define and analyze in the context of dynamic games and, so far, such ideas
have had fewer applications to economic models than rationalizability in
static games. But we believe that these ideas are extremely important
and are going to be applied more and more often. If we managed to wet
your appetite for non-equilibrium solution concepts capturing transparent
assumptions on rationality and interactive beliefs, and you want to see
more of it in the context of dynamic games, then maybe you are ready to
start consulting the literature.43

43

You may start from reviews: Battigalli and Bonanno (1999), Brandenburger (2007),
Perea (2001).

Multistage Games with

Complete Information
We start by providing a mathematical definition of multistage games
whereby, at the beginning of each stage, all the actions taken in previous
stages are publicly observed, hence are common knowledge (Section 8.1).
In a single stage, actions are chosen simultaneously, but some players may
be inactive. Inactive players will be formally represented as players with
only one feasible action. It is possible that in a given stage all the players
but one are inactive. The set of feasible actions in a given stage may be
endogenous, that is, it may depend on the actions taken in previous stages.
The length of the game may also be endogenous. Armed with such formal
notation, we move on to define strategies and the strategic form of a
game. A first definition of equilibrium is given by the Nash equilibrium of
the strategic form. This, however, is unsatisfactory. The more demanding
concept of subgame perfect equilibrium is based on the details of
sequential representation of the dynamic game (Section 8.2). In finite
games where only one player is active at each stage and payoffs are generic,
there is only one subgame perfect equilibrium that can be computed
with the so called backward induction algorithm. In games with finite
horizon, one can use a case-by-case backward induction method to find
the subgame perfect equilibria. These backward computation methods
construct strategy profiles such that no player can improve his payoff from
any position in the game using a simple one-shot unilateral deviation.
Such one-shot deviation property is implied by subgame perfection, but
170

8.1. Preliminary Definitions

171

the converse does not hold in all games. The one-shot deviation principle
states that the one-shot deviation property implies subgame perfection in
a large class of games including all finite horizon games and all games
with discounting. This result is applied to the analysis of repeated games.
Next, we extend the analysis introducing chance moves (Section 8.5) and
two different notions of randomized strategic behavior, mixed strategies
and behavioral strategies; the latter are better suited to define subgame
perfect equilibrium in randomized strategies (Section 8.6).

8.1

Preliminary Definitions

It is useful to start with a preliminary example before we plunge into the

mathematical representation of dynamic games with observable actions.
Example 24. Ann (player 1) and Bob (player 2) play the Battle of
the Sexes (BoS) with an Outside Option. Ann first chooses between
playing the BoS with Bob (action in) or not (action out). If she chooses in,
this choice becomes common knowledge and then the simultaneous-move
subgame BoS is played. If she chooses out the game ends. Payoff are
displayed in the Figure 8.1 below.
in

out

1\2

B2

S2

B1

3, 1

0, 0

S1

0, 0

1, 3

(2, 2)
Figure 8.1: Battle of the Sexes with an Outside Option.

The set of feasible actions of a player depends on the position reached:

Anns feasible set is {in, out} at the beginning of the game, and it is
{B1 , S1 } if she moves in. At the beginning of the game, Bob can only
wait and see what Ann does; therefore his feasible set is the singleton
{wait}. If Ann moves in, Bobs feasible set is {B2 , S2 }. There are 5 possible

172

8. Multistage Games with Complete Information

plays of the game: either the pair of actions (out, wait) is played and
the game ends, or one of the following 4 sequences of actions pairs is
played: ((in, wait), (B1 , B2 )), ((in, wait), (B1 , S2 )), ((in, wait), (S1 , B2 )) and
((in, wait), (S1 , S2 )). A pair of payoffs (u1 (z), u2 (z)) is associated to each
possible play z.
A possible play of the game is called terminal history. Possible
partial plays like (in, wait) are called non-terminal (or partial) histories.
The rules of the game specify which sequences of action profiles are
terminal or non-terminal histories. For each terminal history, viz. z,1
the rules of the game specify a (collective) consequence c = g(z), e.g.,
a distribution of money among the players. Each player i assigns to
each consequence c utility vi (c). As in the first part, vi represents is
preferences over lotteries of consequences ` (C) by way of expected
utility calculations. Given the rules and the utility functions, each
terminal history z is mapped to a profile of payoffs (induced utilities)
(ui (z))iI = (vi (g(z))iI . For example, the payoff pair attached to
terminal history z = ((in, wait), (S1 , S2 )) is (u1 (z), u2 (z)) = (3, 1). Under
the complete information assumption, the rules of the game (including
the consequence function g) and players preferences over lotteries of
consequences (represented by the utility functions vi , i I) are common
knowledge. Thus, also the payoff functions z 7 ui (z) are common
knowledge. In this chapter, we assume complete information and we
directly specify the payoff functions. But, as in the analysis of static
games, it is important to keep in mind that such payoff functions are
derived from preferences and the consequence function g.
When discussing a specific example like the BoS with an Outside Option
it would be simpler to represent the possible plays as sequences like (out),
((in), (B1 , B2 )), ((in), (B1 , S2 )), etc. But we use this awkward notation
for a reason. Having each inactive player automatically choose wait
simplifies the abstract notation for general games: possible plays of the
game are just sequences of action profiles (a1 , a2 , ...) where each element
at of the sequence is a profile (ati )iI and there is no need to keep track in
the formal notation of who is active. Thus, if A denotes the set of all action
profiles, histories are just sequences of elements from set A. To allow for
the theoretical possibility of games that can go on forever, we also consider
1

Since z is the last (i.e., terminal) letter of the alphabet, it seems like a good choice
as a symbol to denote terminal histories.

8.1. Preliminary Definitions

173

infinite sequences of elements from A. The rules of the game specify which
sequences are possible, i.e., they specify the set of histories.
Since sequences of elements from a given domain are a crucial ingredient
of the formal representation, it is useful to introduce some preliminary
concepts and notation about sequences.

8.1.1

Sequences and trees

Fix a set X. The set of finite sequences of length ` of elements from

set X is the Cartesian product X ` . Such sequences can be thought as
functions from the set of the first ` positive integers, {1, ..., `} to X;
thus, X ` is isomorphic to the set of functions X {1,...,`} . Now consider
the union of all such sets, i.e. the set of all finite sequences of elements
from X, plus the singleton X 0 containing only the empty sequence (a
convenient mathematical object analogous to the empty set).2 With a
quite descriptive notation, such union is represented by the symbol X <N0
X <N0 :=

X`

`N0

where N0 = {0, 1, 2, ...} is the set of natural numbers. It may help to

think of X as an alphabet for some language . Then, the set of words in
language is a subset of X <N0 . The rules of language determine which
elements of X <N0 are words of .
Similarly, the rules of a given game determine which sequences of
actions profiles are (feasible) histories. But, unlike words in natural
languages, some histories of games may have an infinite length. This means
that the game does not necessarily end in a finite number of stages. This is
a useful abstraction. Consider, for example, a bargaining game where the
parties go on to make offers and counteroffers until they reach a binding
agreement: it is possible that they never agree and hence bargain forever.
Countably infinite sequences of elements from X are like functions from
N = {1, 2, ...} to X. Therefore the set of such sequences is denoted X N :
k
X N := {(xk )
k=1 : k N, x X}.
2

Similarly, the power set of X, contains the empty set, that is, 2X , or equivalently
{} 2X .

174

8. Multistage Games with Complete Information

The set of all finite and infinite sequences from X (empty sequence
included) is
X N0 := X <N0 X N .
Another worth noting difference between the words of natural
languages and the histories of a game is that the latter form a rooted
tree, i.e. a partially ordered subset T X N0 where (i) the order relation
x y is x is a prefix (initial subsequence) of y, (ii) every prefix of a
sequence in T (including the empty sequence) is also a sequence in T , so
that the empty sequence is the root of the tree, and (iii) (in games that
may never end) if z X N and every (finite) prefix of z is in T then also
z T.

8.1.2

Multistage games with observable actions

Consider a game that proceeds through stages. At each stage there is a

subset of active players, this set may depend on previous moves; each active
player chooses an action in a feasible set with two or more alternatives,
and the choices of the active players are simultaneous. The rules of the
game are such that at the beginning of each stage the action profiles chosen
in previous stages become public information. The rules determine what
actions are feasible for each player according to previous choices of every
player. The only feasible action of an inactive player is to wait, hence
his feasible set is a singleton. When the game ends, players have empty
feasible sets: not only they cannot choose between alternatives, they also
stop waiting. The consequences for the players implied by the rules of the
game, in general, may depend on the actions taken since the game started,
and on the order in which they were taken, that is, the consequences
depend on the sequence of action profiles that led to the end of the game.
Games forms with simultaneous moves, or static game forms, are a special
case whereby the first actions profile ends the game.
As we did in the case of simultaneous moves games, we do not
represent the rules directly, but rather we describe the players set,
their feasible sets of actions (including how they depend on previous
moves) and the consequences of their actions. We do this by providing
an abstract representation of the possible sequences of actions profiles,
called histories,
and of a map from terminal histories
to
consequences.[Change here: do in 3 steps, tree, consequences,

8.1. Preliminary Definitions

175

preferences]
A multistage game form with observable actions is a structure
hI, C, g, (Ai , Ai ())iI i
given by the following components:
For each i I, Ai is a non-empty set of potentially feasible
actions.
Let A := iI Ai and consider the set A<N0 of finite sequences
of action profiles; then, for each i I, Ai () : A<N0 Ai is a
constraint correspondence that assigns to each finite sequence of
action profiles h` = (at )`t=1 the set Ai (h` ) of actions of i that are
feasible immediately after h` . It is assumed that Ai (h0 ) 6= and, for
all h A<N0 , Ai (h) = if and only if for every j I, Aj ( h) =
(the reason for the latter assumption will be explained below).
Sequence (at )`t=1 (` = 1, 2, ..., ) is a (feasible) history if a1 A(h0 )
and at+1 A(a1 , a2 , ..., at ) for each t {1, ..., ` 1}, where A(h) :=
iI Ai (h) for each h A<N0 . Thus, a history is a sequence of action
profiles whereby each action profile is feasible given the previous
ones. By convention, the empty sequence, denoted h0 , is a history.
A<N0 denote the set of histories; a history z = (at )` H

Let H
t=1
N
is terminal if either z A or A(z) = . Let
n
o
: z AN or A(z) =
Z := z H
denote the set of terminal histories. With this
g:ZC
is the consequence function.
In order to analyze interaction between a specific group of individuals
for given rules of the game, we add their personal preferences over lotteries
of consequences, represented by utility functions:
for each i I, vi : C R is the VonNeumann-Morgenstern utility
function of player i.

176

8. Multistage Games with Complete Information

With this, we obtain a multistage game with observable actions,

that is, a structure
= hI, C, g, (Ai , Ai (), vi )iI i
given by the game form and players preferences. In this chapter, we
maintain the informal assumption of complete information: the rules of
the game and players preferences are common knowledge.
We are going to intensively use two further pieces of notation. We let

H := H\Z
denote the set of non-terminal (or partial) histories.
ui = vi g : Z R denote the payoff function of player i.
of histories has the structure of a tree (connected graph
The set H
without cycles) with a distinguished root, when endowed with the following
prefix of precedence relation:
Definition 36. Sequence h precedes sequence h0 , written h h0 , if h
is a prefix (initial subsequence) of h0 , i.e., either h = h0 and h0 6= h0 or
h = (at )kt=1 , h0 = (bt )`t=1 , k < ` and at = bt for each t {1, ..., k}.

Given
should be able to prove that
 the definitions above, the reader
0

H, is a tree with distinguished root h :

has the following properties:

Remark 31. The set of histories H
0
0 }, h0 h;
and, for each h H\{h

(1) h H
if h h0 then
(2) for each sequence h A<N and each history h0 H,

h H;

(3) for each infinite sequence z AN , if every prefix of z is in H

<N

(h A , h z implies h H), then z H.

ST At for some
Definition 37. Game has finite horizon if H
t=1
A<N and H
is finite. Game is static
T N. Game is finite if H
if Z = A(h0 ).
Sometimes the definition of games with observable actions takes the set
as primitive and postulates properties (1)-(3) of Remark 31 (cf. Osborne
H
and Rubinstein, 1994, pp 89-90, 102). The most common definition
of dynamic game takes the tree-structure as primitive, but deals with
simultaneous actions in a somewhat arbitrary and un-intuitive way (see,

8.1. Preliminary Definitions

177

e.g., Fudenberg and Tirole, 1991, pp 80-81). We will take advantage of

the tree-structure to provide graphical representations of some games, in
particular the games where at each stage t at most one player has at least
two feasible actions; such games are said to have perfect information :
Definition 38. Player i is active at history h H (that is, immediately
after h) if i has at least two feasible actions at h: |Ai (h)| 2. A multistage
game with observable actions has perfect information if, for each h H,
at most one player is active at h.
Note, perfect information is a formal property of dynamic games, and
it should not be confused with complete information, an assumption with
a different meaning that here we do not express formally.

8.1.3

Comments

According to the present notation, when Ai (h) = for some player i, it

means that the game is over. Therefore it is assumed that Ai (h) = for
some i if and only if Ai (h) = for all i. An inactive player in an ongoing
game is a player with only one feasible action, say action wait.
One could give a more elegant definition in which there is a set I(h)
of active players at each partial history h and feasible actions are specified
for active players only. This would have the advantage of simplifying the
notation for perfect information games and also for specific examples of
games. But such a definition would be more complex3 without essential
gains in generality.

8.1.4

Graphical Representation

Games of perfect information are graphically represented as game trees:

is a node; each z Z is a terminal node (or leaf); each
each h H
h H is a decision node; each terminal history/node z is associated
to the payoff vector (u1 (z), u2 (z), ..., un (z)); each partial history/decision
node h is associated to the only active player, denoted by i = (h); if
3
With this alternative definition, action
S profileshave the form aJ = (aj )jJ with 6=
J I, the set of such profiles is A =
(iJ Ai ); an active players correspondence
6=JI

I() : A<N 2I has to be introduced among the primitive elements defining the game;
the constraint correspondence Ai () is defined on the subset Hi = {h A<N : i I(h)}.

178

8. Multistage Games with Complete Information

h0 = (h, a) (the concatenation of h and a), then there is a directed arc

from h to h0 , this directed arc is associated to action a(h) , the action of the
active player in profile a = (a(h) , wait(h) ). Sometimes it is notationally
useful to distinguish between physically identical actions if they are
taken after different histories. In this case, there is a 1-1 correspondence
between arcs and actions in perfect information games.
For example, suppose the following game is played between player 1
and player 2. A referee puts a dollar on the table. Player 1 can take it or
leave it, player 2 just observes and waits. If player 1 takes the dollar the
game is over, otherwise the referee puts another dollar on the table, player
2 can take the two dollars or leave them on the table, player 1 observes
and waits. If player 2 takes the dollars the game is over, otherwise the
referee puts another dollar on the table. The game goes on like this until
the referee has exhausted his T dollars or someone has taken the dollars on
the table. If there are T dollars on the table and the active player (player
1 if T is odd) leave them, they go to the other player. It is common
knowledge that players only care about money (if uncertainty plays a role
assume for simplicity that they are risk-neutral). This game, for T = 4, is
represented as follows:
1

Take

Leave

Take
1
0

Leave

Take
0
2

Leave

Leave

4
0

Take
3
0

0
4

Figure 8.2: Game for T = 4.

Assume that Ai = {Take, Leave, Wait}. Action Wait is included for

notational convenience, to fit the general definition given above. We have
H = {h0 , (Leave, Wait), ((Leave, Wait),(Wait, Leave)) ,
((Leave, Wait),(Wait, Leave),(Leave, Wait))},
Ai (h) = {Take, Leave} if i is active at h H and Ai (h) = {Wait}
otherwise. The vertical arrows (directed arcs) are associated to the action

8.1. Preliminary Definitions

179

Take of the active player, the horizontal arrows are associated to the action
Leave of the active player. The inactive player can only Wait. We explicitly
included the pseudo-action Wait only to clarify the abstract, general
notation introduced above. But from now on we will identify histories
with sequences of actions by active players only: e.g. (Leave, Wait) will be
simply written as (Leave), ((Leave, Wait),(Wait, Leave)) will be written as
(Leave, Leave) , etc.
How would you play this game? Does your answer depend on how large
is T ?

8.1.5

Strategies, Plans, Strategic Form, Reduced Strategic

Form

A strategy is a complete contingent plan which includes instructions

contingent on every history, including histories where the opponents make
unexpected moves as well as histories inconsistent with the plan itself.
More formally:
Definition 39. A strategy for player i is an element of the product set
hH Ai (h), that is, a function si : H Ai such that si (h) Ai (h) for
each h H. The set of strategies for player i is denoted Si := hH Ai (h).
We let S := iI Si and Si := j6=i Sj .
RemarkQ
32. In a finite game the number of strategies of each player i I
is |Si | = hH |Ai (h)|.
Given S, it is possible to derive a fictitious static game G = N (),
sometimes called the normal form of , that captures some important
aspects of the game .4 The idea is the following: each player instructs a
trustworthy agent, or writes a computer program for a machine, so that the
agent/machine implements a particular strategy. Each profile of strategies
s = (si )iI induces a terminal history and the associated payoff vector. Of
course, the players choices of strategies are simultaneous (or, at least, no
player can observe anything about the strategies chosen by the opponents
before he chooses his own strategy).
The strategic form of the Take-it-or-Leave it game depicted in Figure 1
(henceforth ToL4) is represented in the following table. Strategies are
4

According to some theorists but not ourselves all the relevant aspects of a game
are captured by this derived representation.

180

8. Multistage Games with Complete Information

labelled by sequences of letters separated by dots, with the following

convention: the first letter denotes the action to be taken at the first
history where the player is active [h0 for player 1 and h = (Leave)
for player 2], the second letter denotes the action to be taken at the
second history where the player is active [h = (Leave, Leave) for player
1, h = (Leave, Leave, Leave) for player 2]. For example, s1 = T.T is the
strategy [Take; Take if (Leave, Leave)] of player 1.
1\2

T.T

T.L

L.T

L.L

T.T

1, 0

1, 0

1, 0

1, 0

T.L

1, 0

1, 0

1, 0

1, 0

L.T

0, 2

0, 2

3, 0

3, 0

L.L

0, 2

0, 2

0, 4

4, 0

Figure 8.3: Strategic form of ToL4.

Definition 40. The outcome function : S Z associates each

strategy profile s S to the induced terminal history (s) = (at )Tt=1
(T ) defined recursively as follows:


a1 = s1 (h0 ), ..., sn (h0 ) ,


at+1 = s1 ((a )t =1 ), ..., sn ((a )t =1 )
for every t with 1 t T . The strategic or normal form of is the
static game
N () = hI, (Si , Ui )iI i
where Ui = ui : S R for every i I.5 A Nash equilibrium of is
a Nash equilibrium of N ().
Consider the strategies T.T and T.L of player 1 in the Take-it-or-Leave
it game depicted in Figure 1. Both of them prescribe to take one dollar
immediately, an action that terminates the game. They differ only for
the instruction corresponding to the history h = (Leave, Leave), that is,
a history prevented by both strategies. These two strategies yield the
5

Recall that (ui )(s) = ui [(s)].

8.1. Preliminary Definitions

181

same terminal history, (Take), independently of the strategy adopted by

player 2. A similar observation holds for player 2: the two strategies
T.T and T.L differ only for the instruction corresponding to the history
h = (Leave, Leave, Leave), that both of them prevent. Which terminal
history is reached depends on the strategy of player 1, but does not depend
on which of these two strategies is played by player 2. This is the reason
why these strategies correspond to identical rows (player 1), or columns
(player 2).
These considerations motivate the following definitions. For each
strategy si there is a set of non-terminal histories H(si ) which are not
prevented by it. Formally, history h is not prevented, or allowed by si
if there is a profile of strategies of the opponents si such that (si , si )
induces h. Since (si , si ) induces h if and only if it induces a terminal
history preceded by h, we can define the set of non-terminal histories
allowed by si as follows:
H(si ) := {h H : si Si , h (si , si )} .
Definition 41. We say that two strategies si , s0i Si of player i I are
behaviorally equivalent, written si i s0i , if they allow the same nonterminal histories and prescribe the same actions at such histories, that
is, if H(si ) = H(s0i ) and si (h) = s0i (h) for each h H(si ) = H(s0i ). A
reduced strategy of i is a maximal set of behaviorally equivalent strategies
of i, that is, an element of the quotient set Si | i .6
A reduced strategy is sometimes called plan of action (see Rubinstein
[32]). The reason is that plan of action corresponds to the intuitive idea
that a player has to plan for all external contingencies, i.e. contingencies
that not depend on his behavior, but he does not have to plan for
contingences that cannot occur if he follows his plan. For example, in
the ToL4 game of Figure 8.3, T.T and T.L are realization-equivalent and
they correspond to the reduced strategy, or plan T , that is, take at the
first opportunity.
6

The quotient of a set S with respect to an equivalence relation on S is the collection

of equivalence classes:
S| := {s S : (s, t) S S, {s, t} s s t}.

182

8. Multistage Games with Complete Information

We avoid this terminology, and stick to the more neutral reduced

strategy, because there is a perfectly meaningful notion of planning
that yields the specification of a whole strategy, not just a reduced strategy.
To see this, consider the BoS with an Outside Option of Figure 0. Suppose
Ann, player 1, believes that, conditional on choosing In, the probability
that Bob chooses B2 is 1 (B2 |In) = q. Then Ann can compute a
dynamically optimal plan with the following folding back procedure.
First, she computes the values of choosing, respectively, B1 and B2 given
In:
V1q (B1 |In) = 3q,
V1q (S1 |In) = 1 q.
Next, assuming that she would maximize her expected payoff in the
subgame if she chose In, Ann computes the value of In:

3q,
if q 14 ,
q
V1 (In) = max{3q, 1 q} =
1 q, if q < 14 .
This means that Ann plans to choose B1 (resp. S1 ) in the subgame if
q > 1/4 (resp. q < 1/4).7 Finally, Ann compares V1q (In) with the value
of the outside option V1 (Out) = 2 and plans to choose In (resp. Out)
if V1q (In) > 2 (resp. V1q (In) < 2), that is, if q > 2/3 (resp. q < 2/3).
The upshot of all this is that, given her subjective beliefs q, Ann can use
the folding back procedure to form a dynamically optimal plan, which
corresponds to a full strategy, not a reduced strategy:8

Out.S1 , if q < 41 ,
Out.B1 , if 41 < q < 23 ,
s1 =

In.B1 ,
if q > 23 .
On the other hand, one may object that specifying a complete strategy
is not necessary for rational planning: if q is not high enough, i.e. if
max{3q, 1q} < 2, Ann has no need to plan ahead for the subgame, as she
can see that it is not worth her while to reach it. Therefore her best plan
is just Out, which is a reduced strategy. Since both notions of rational
7
8

If q = 1/4 she is indifferent, and the optimal plan for the subgame is arbitrary.
Again, we ignore ties.

8.1. Preliminary Definitions

183

planning are meaningful and intuitive, we are not going to endorse only
the second one of them by calling plan of action the reduced strategies.
The solution concepts presented in the next section further clarify why
the notion of strategy (as opposed to reduced strategy) is important in
game theory. Roughly, we are going to elaborate on the folding back
procedure of dynamic programming. While the folding back procedure
refers to a single player and his arbitrarily given subjective beliefs, the
standard theory of multistage games looks for collective and objective
elaborations of the folding back procedure.
[UNDER CONSTRUCTION HERE]
Definition 42. Two strategies si and s0i are realization equivalent if, for
every given strategy profile of the co-players, they induce the same history:
si Si , (si , si ) = (s0i , si ).
Lemma 10. Two strategies si and s0i are realization equivalent if and only
if they are behaviorally equivalent:


si Si , (si , si ) = (s0i , si ) si i s0i .
Definition 43. The reduced strategic (or normal) form of a game
is a static game N r () = hI, (Sri , Uir )iI i where each set Sri of each player
i I is obtained from Si by replacing each class of behaviorally equivalent
strategies with a unique representative element, that is, Sri = Si | i , and
Uir (sr ) = Ui (s) for each sr jI Srj and s sr .
Definition 44. Two strategies si and s0i are payoff-equivalent, written
si i s0i if, for each strategy profile of the co-players, they yield the same
profile of payoffs:
si Si , (si i s0i ) (j I, Uj (si , si ) = Uj (s0i , si )).
Since Uj = uj (j I), Lemma 10 implies the following:
Remark 33. If two strategies are behaviorally equivalent, then they are
payoff equivalent.
Indeed, a difference between payoff-equivalence and behavioral
equivalence may arise only if the game features some ties between payoffs

184

8. Multistage Games with Complete Information

at distinct terminal histories z 6= z 0 : (ui (z))iI = (ui (z 0 ))iI . Such ties are
structural when z and z 0 yield the same (collective) consequence, e.g.
the same allocation, g(z) = g(z 0 ); otherwise they are due to non generic
ties between utility profiles: g(z) 6= g(z 0 ) and (vi (g(z)))iI = (vi (g(z 0 )))iI .
Unfortunately, the consequence function is mostly overlooked in game
theory and very often all ties between payoffs at distinct terminal histories
are called non generic.
The reduced strategic form of the ToL4 game is represented by the
following table:
1\2

L.T

L.L

1, 0

1, 0

1, 0

L.T

0, 2

3, 0

3, 0

L.L

0, 2

0, 4

4, 0

Figure 8.4: Reduced strategic form of ToL4.

Comment: The most common definition of reduced normal form

refers to classes of payoff -equivalent strategies, rather than realizationequivalent strategies. This is fine if one is only interested in computing
equilibria (or other solutions) of the normal form. We instead are interested
in the normal form N () only as an auxiliary tool that (sometimes)
helps analyzing itself. Therefore we chose to emphasize the concept
of realization-equivalence and reduced strategy, which is based on a
meaningful notion of plan, and the corresponding concept of reduced
normal form.
Remark 34. Payoff (hence realization) equivalences induce a trivial
multiplicity of Nash equilibria: if a strategy profile s is a Nash equilibrium,
every profile s0 obtained from s by replacing some strategy si with an
equivalent strategy s0i is also a Nash equilibrium. In other words, to find
Nash equilibria we may just look at the equilibria sr of reduced strategic
form and then take into account that any profile s of the non-reduced form
that corresponds to sr is a Nash equilibrium of the dynamic game.
[Revised up to this point]

8.2. Backward Induction and Subgame Perfection

8.2

185

Backward Induction and Subgame Perfection

Any solution concept for static games can be applied to the strategic (or
normal) form N () of a multistage game and thus yields a candidate
solution for . For example, we can find the Nash equilibria or the
rationalizable strategies of N () and ask ourselves if they make sense as
solutions of .
There are many examples of rationalizable or Nash equilibrium strategy
profiles that lack obvious credibility properties. Consider the following very
stylized Entry Game : a firm, player 1, has the opportunity to enter in
a (so far) monopolistic market. The incumbent, player 2, may fight the
entry with a price war that damages both, or acquiesce. The game is
represented in Figure 8.5:
1

In

1
1

Out
0
2

1
0

Figure 8.5: An entry game.

It is easily checked that this game has two pure Nash equilibria:
[Out,(f if In)] and [In,(a if In)] (plus a continuum of randomized equilibria
where 2 chooses strategy (f if In) with probability larger than 12 ). Under
complete information, if the potential entrant believes that the incumbent
is rational, than she expects acquiescence and therefore enters. The only
reasonable equilibrium seems to be [In,(a if In)]. The problem with
the Nash equilibrium is that it implies that players maximize their payoff
on the equilibrium path, i.e. along the play induced by the equilibrium
strategy profile, but it allows players to plan non maximizing actions at
histories that are off the equilibrium path.
The reasonable equilibrium of the Entry Game can be obtained in
two ways.
(1) Backward induction : We consider the position of a player who

186

8. Multistage Games with Complete Information

is about to terminate the game (the incumbent, in the example) and we

select her best action. Next we put ourselves in the position of a player who
can either terminate the game or give the move to someone (maybe herself
in general games) who will terminate the game (the potential entrant in
the example); the payoff resulting from giving the move to another player
is what would be obtained if this player selected her best action, thus we
have to look for the action selected at the previous step. Clearly we get
[In,(a if In)] in the Entry game. In more complicated games the procedure
may take many steps (e.g., it takes fours steps in the ToL4 game).
(2) Subgame perfection : We consider strategy profiles that
induce a Nash equilibrium not only in the whole game, the one starting
with the empty history, but also in the subgames starting with all the
other non-terminal histories. In the Entry Game, [Out,(f if In)] does
not induce an equilibrium in the subgame (actually a rather trivial
decision problem) starting at history (In), whereas [In,(a if In)] induces
an equilibrium (an optimal decision) also for this subgame.

Next we give a precise formulation of these ideas and relate them to

each other: backward induction is an algorithm that computes the unique
subgame perfect equilibrium in a class of games where the algorithm
works. This class of games includes all the non-generic finite games
with perfect information. For simplicity we focus on such games. Then we
provide a critical discussion of the meaning and conceptual foundations of
this solution procedure, illustrating it with the ToL example.
The solution concepts we are going to analyze require the knowledge
of the multistage game rather than just its strategic form N (). Yet,
they look for appropriate profiles of strategies, as if players made complete
contingent plans9 before the game starts. Such plans are the object of the
analysis. However, it is not assumed that strategies are implemented by
a machine or a trustworthy agent. Rather, it is informally assumed that
a player makes a plan in advance and, as the play unfolds, he is free to
either choose actions as planned or to deviate from the plan.

Where contingencies are non-terminal histories, and hence allow for the possibility
of making mistakes, as discussed in the previous section.

8.2. Backward Induction and Subgame Perfection

8.2.1

187

Backward Induction

The backward induction solution procedure is well-defined for all the finite
games with perfect information such that different terminal histories yield
different payoffs (for every i I and for every z, z 0 Z, z 6= z 0 implies
ui (z) 6= ui (z 0 )). We refer to the latter assumption about payoffs by saying
that the game is generic. The procedure can be generalized to other perfect
information games with finite horizon10 and to some games with more than
one active player at each stage, such as the finitely Repeated Prisoners
Dilemma.
Backward Induction Procedure. Fix a generic finite game with
perfect information. Let `(h) denote the length of a history and let
d(h) = maxz:hz [`(z) `(h)] if h H, d(z) = 0 if z Z. This is the
depth of history/node h, that is the maximum length of a continuation
history following h. Recall that (h) denotes the player who is active at
R
h H. Let us define for each player i I a value function vi : H

and a strategy si as follows:

For all h H and j I, if j 6= (h) then sj (h) is obviously the only
action (wait ) in the singleton Aj (h).
If d(z) = 0 (i.e. if z Z), i I, vi (z) = ui (z).
such that
Suppose that vi (h0 ) has been defined for all h0 H
0
0
0 d(h ) k (where 0 k < d(h )); fix h such that d(h) = k + 1
(note that, by definition, this implies d ((h, a)) k for all a A(h));
if i = (h), let


si (h) = arg max vi (h, (ai , si (h))
ai Ai (h)

(si (h) is uniquely defined, see above); for all j I, let vj (h) =
vj ((h, s (h))).
10

For example, the genericity property of payoff functions rules out many games to
which the backward induction procedure can be easily applied, such as the ToL game.
Here is a simple generalization satisfied by ToL: a perfect information game has no
relevant ties if, for every pair of distinct terminal histories z 0 and z 00 (z 0 6= z 00 ), the
player who is decisive for z 0 vs z 00 (i.e. the player who is active at the last common
predecessor of z 0 and z 00 ) is not indifferent between z 0 and z 00 .

188

8. Multistage Games with Complete Information

In words, we start from the last stage of the game: h is such that all
feasible actions at h terminate the game, that is d(h) = 1.11 For example,
in the Take-it-or-Leave-it game of Figure 1 there is only one history h with
d(h) = 1, that is, h = (Leave, Leave, Leave). According to the algorithm,
the active player selects the payoff maximizing action. This determines a
profile of payoffs for all players, denoted (vi (h))iI . Now we go backward
to the second-to-last stage, or more precisely we consider histories of
depth 2: d(h) = 2. The value vi ((h, a)) has already been computed for
because such histories correspond to the last stage
all histories (h, a) H,
of the game. According to the algorithm, the active player (h) chooses
. Intuitively, the reason is that the
the feasible action that maximizes v(h)
active player expects that every following player (possibly himself) would
maximize his own payoff in the last stage. The algorithm continues to go
backward in this fashion until it reaches the first stage (h = h0 ).
Try to use the procedure to solve the Take-it-or-Leave-it game. Would
you play according to the backward induction procedure?

8.2.2

Assumptions About Beliefs Justifying Backward

Induction

In a dynamic game, such as a multi-stage game with observable actions, a

player has initial beliefs which he updates as the play unfolds. Sometimes
what he observes may be unexpected according to his previous beliefs. In
this case, he will not be able to use Bayes rule in order to update, but he
will still form new beliefs about his opponents behavior.12 For example, in
the Take-it-or-Leave-it game of Figure 1 (ToL4 for short), if Colin (player
2) believes in the backward induction solution, he initially expects that
Rowena (player 1) will immediately take one dollar. If instead Rowena
leaves the dollar on the table, Colin will revise his beliefs about Rowena,
because he now knows that she is using one of the following strategies:
LT (Leave, then Take if also Colin leaves) or LL (Leave, then Leave again
11
Recall that which stage is the last, second-to-last and so on may be endogenous.
For example, we may have a game that lasts for two stages if the first mover chooses
Left and three stages if the first mover chooses Right.
12
This subsection is quite informal, but all the arguments contained in it can be
rigorously formalized by introducing analytical tools that are beyond the scope of these
notes. For a rigorous analysis see the survey by Battigalli and Bonanno (1999) and the
references therein.

8.2. Backward Induction and Subgame Perfection

189

if also Colin leaves). The revised beliefs must assign zero probability to
every other strategy of Rowena.
We deem a player in a dynamic game is rational if he would make
expected utility maximizing choices given his (updated) beliefs for every
possible history of observed choices of her opponents. Consider ToL4. Can
one say that Rowena is irrational if she leaves three dollars on the table?
No. Rowena may hope that then Colin will be generous and leave her
four dollars. We might argue that such a belief is not very reasonable, in
fact, it is inconsistent with the rationality of Colin; yet, if Rowena had this
belief, leaving three dollars on the table would be rational, i.e. expected
utility maximizing.
In the analysis of static games, one can capture with a solution
concept the following assumptions: all players are rational and there is
common (probability-one) belief of rationality. The solution concept is
rationalizability. An action is rationalizable if and only if it is iteratively
undominated.
Is there an analog of rationalizability for dynamic games? What can
one say about generic finite games with perfect information?
Note that in generic games there is a one-to-one relationship between
terminal histories and payoff vectors. Here we make use of this to
refer to either terminal histories or the corresponding payoff vectors as
outcomes. This allows me to identify corresponding outcomes in the
given perfect information game and in its strategic form. In particular, we
are interested in the outcome induced by the backward induction solution.
It is sometimes argued that the compelling logic of the backward
induction solution rules out all the non-backward-induction outcomes as
inconsistent with common certainty of rationality. If this claim were
correct, backward induction would be the analog of rationalizability for
(generic and finite) games with perfect information.
To assess this claim, let us first restrict our attention to games with
two stages, such as the Entry Game. It is pretty clear that in such
games if the players are rational in the sense specified above and if
the first mover believes that also the opponents are rational, then the
backward induction play obtains. Furthermore, the backward induction
outcome can be obtained on the normal form of the game by first
eliminating the weakly dominated strategies and then eliminating the
(strictly) dominated strategies of the residual normal form. Thus the

190

8. Multistage Games with Complete Information

similarity between backward induction in two-stage perfect information

games and rationalizability in static games is quite strong.13
Now consider games with more than two stages. It is possible to
show that the iterative removal of weakly dominated strategies in the
strategic form yields the backward induction outcome (verify this in the
simple examples you have seen so far; note that the result refers to the
outcome not to the strategies). This seems to suggest that the similarity
between rationalizability and backward induction holds in general and that
backward induction can be justified by assuming common certainty of
rationality. But on closer inspection this intuition turns out to be incorrect.
First note that when we talk about players beliefs in a dynamic game
we must specify the point (history) of the game where the players hold
such beliefs. The most obvious reference point is the beginning of the
game (the empty history). In this case we consider initial common belief
in rationality. It turns out that the claim (rationality and) initial common
belief in rationality rules out non-backward-induction outcomes is false.
For example, it can be shown that in the ToL4 game every outcome except
(4, 0) and (0, 4) is consistent with rationality and initial common certainty
of rationality. The reason is that if a player is surprised by her/his
opponent s/he may form new beliefs inconsistent with common certainty of
rationality even if her/his initial beliefs satisfy common belief in rationality.
Of course, if there were common belief in rationality at every node
(history) of the game, then rational players would play the backward
induction outcome. But in many games it is impossible to have common
belief in rationality at every node because such assumption lead to
contradictions. Take-it-or-Leave-it is a case in point: common in
rationality at every node, if it were possible, would imply that Colin
expects Rowena to take immediately. But then there cannot be common
belief in rationality after Rowena has left one dollar on the table.
13

The reason why we start with the elimination of weakly dominated strategies is that,
if a strategy si prescribes an irrational continuation at a given history h consistent with
si itself, then this shows up in the normal form as si being weakly dominated. The
dominance relation is, in general, only weak because the opponents might behave in
such a way that h does not occur, implying that the flaw of si does not materialize. For
example, the strategy LL of Colin is irrational, because Colin leaves 4 dollars on the
table at his second move. This strategy is weakly, but not strictly, dominated by LT .
LL cannot be strictly dominated because both LL and LT are ex ante best responses to
strategy T L (or T T ) of Rowena: if she takes immediately the strategy of Colin cannot
affect his payoff.

8.2. Backward Induction and Subgame Perfection

191

Thus neither version of common belief in rationality can be

invoked to justify backward induction.
Initial common belief in
rationality is a consistent set of assumptions but allows for non-backward
induction outcomes (this is formally proved in the Appendix with
a mathematical structure similar to those used to analyze games of
incomplete information), while common belief in rationality at every node
leads to contradictions.
Then why is backward induction so compelling? Well, not everybody
finds backward induction so compelling. Furthermore, backward induction
seems more compelling in games with a small number of stages (e.g. three
or four) than in longer games. Indeed, for every generic finite game of
perfect information we can find a sufficiently long list of assumptions
about players beliefs (beside the rationality assumption) that yield the
backward induction outcome. However, the longer the game, the stronger
the assumptions justifying backward induction. Such assumptions can
be summarized by the following best rationalization principle: every
player always ascribes to his opponents the highest degree of strategic
sophistication consistent with their observed behavior.14
For example, consider ToL4.
The highest degree of strategic
sophistication that Rowena can ascribe to Colin if Colin leaves two dollars
on the table is the following: Bob is rational but believes he can trick
Rowena into leaving three dollars on the table so that he can eventually
grab four dollars. In other words, at her second decision node Rowena
may think that (a) Colin is rational, (b) Colin still believes that Rowena
is rational, but (c) Colin also believes that Rowena believes that Colin
is irrational (otherwise Rowena would grab three dollars if given the
opportunity and Colin would not leave the dollars on the table).
More generally, say that player i strongly believes event E, in
symbols SBi (R), if player i believes E conditional on every history h that
does not contradict E. Then the best rationalization principle consists of
the following assumptions:
R: every player is rational,
SB(R): every player strongly believes in rationality,
SB(R SB(R)): every player strongly believes in the conjunction of
rationality and strong belief in rationality, R SB(R),
14

The same principle can be used to justify forward induction outcomes in

multistage games with simultaneous moves and/or incomplete information.

192

8. Multistage Games with Complete Information

(...)
SB(R SB(R) SB(R SB(R))...): every player strongly believes the
conjunction of all the assumptions made above,
(...).
Once the best rationalization principle is properly formalized, it can
be shown to yield, for generic games, the iterative deletion of weakly
dominated strategies.15 Since in generic games of perfect information
iterated weak dominance yields the backward induction outcome, we
can deduce that in such games the best rationalization principle
implies the backward induction outcome.16 [Add more here: only
path equivalence. + CB in rational Planning and Continuation
Consistency]

8.2.3

Subgame Perfect Equilibrium

The backward induction procedure selects a Nash equilibrium of each

generic, finite game of perfect information. One can think of a generalized
procedure to select among Nash equilibria that applies to general
multistage games with observable actions. The main intuition is as follows:
Fix a candidate equilibrium s = (si )iI . Every player initially believes
that s is going to be played and hence terminal history (s) will occur. The
part of a strategy si which is defined at histories h H inconsistent with
si itself (i.e. such that h is not an initial sub-history of terminal history
(si , s0i ) for any s0i ) represents what every opponent j 6= i would believe
about i after i has deviated from his equilibrium behavior; furthermore, it
also represents what i believes that his opponents would believe. The part
of si defined at histories h H inconsistent with si represents what player
i threatens or promises to do if his opponents deviate. His opponents
believe these threats/promises. Thus this part represents their beliefs
about how i would react if some of them were to deviate. All these beliefs
should be reasonable. In particular, the implicit threats/promises
15

Specifically, the best rationalization principle corresponds to the assumptions of

rationality and common strong belief in rationality, which yield a solution concept called
extensive form rationalizability (Pearce, 1984). For generic values of the payoffs as
terminal nodes, extensive form rationalizability coincides with the iterated deletion of
weakly dominated strategies.
16
These issues of dynamic interactive epistemology are formally discussed in the
survey by Battigalli and Bonanno (1999), and the more focused survey by Brandenburger
(2007). The latter survey contains references to the most recent relevant literature.

8.2. Backward Induction and Subgame Perfection

193

should be credible. One way to formalize these ideas is to assume that

the continuation strategies induced by s starting from every history h H
form a Nash equilibrium of the subgame starting at h. Such a strategy
profile is called subgame perfect equilibrium. We are going to provide
two equivalent definition of subgame perfect equilibrium, the first uses the
strategic-form payoff functions Ui : S R, the second does not rely on
the strategic form.
1. For any strategy si and any non-terminal history h = (a1 , ..., at ) and
player i I, let shi denote the strategy that agrees with si at all histories h0
that do not precede h and chooses aki immediately after initial subhistory
(a1 , ..., ak1 ) (1 k t). Note that if h is consistent with si , then shi = si .
Otherwise, shi is the strategy most similar to si among those choosing
actions a1i , ..., ati in history h. For example, let s1 = T T in the ToL4 game,
and let h {(Leave), (Leave, Leave)}; then sh1 = LT .
By definition of shi , if another player j observes h and believes that
i is playing shi , than he believes that i will follow si afterward even if
i deviated before. As usual, shi = (shj )j6=i . Also, let Si (h) denote the
strategies consistent with h (h is consistent with si if h (si , s0i ) for
some s0i ). [Before you go on, you should make sure you understand this
notation using a specific example, such as the ToL game.]
Definition 45. A strategy profile s = (si )iI is a subgame perfect
(Nash) equilibrium if, for all h H, i I and s0i Si (h),
Ui (shi , shi ) Ui (s0i , shi ).
Comment. Once consistency with h is guaranteed, the only relevant
parts of shi and s0i Si (h) for the above inequality are those defined at
h and the histories h0 following h (i.e., histories h0 such that h h0 ).
The interpretation is precisely that the strategy profile s induces a Nash
equilibrium starting from every h. (You may try to make this statement
precise by providing an exact definition of subgame and of the other
terms of the statement, a very boring task.)
2. For each h H and s S, let (s|h) denote the terminal
history induced by strategy profile s starting from history h, that is,
(s|h) = (h, s(h), s(h, s(h)), ...).

194

8. Multistage Games with Complete Information

Remark 35. A strategy profile s is a subgame perfect equilibrium if and

only if, for all h H, i I, s0i S,
ui ((s|h)) ui ((s0i , si |h).
One can show that, if the backward induction procedure can be
unambiguously applied to compute a strategy profile s (e.g., in all generic
finite games with perfect information), then s is the unique subgame
prefect equilibrium of the given game. Therefore, backward induction can
be regarded as an algorithm to compute subgame perfect equilibria.

8.3

The One-Shot-Deviation Principle

The One-Shot Deviation Principle says that a strategy profile s in

a dynamic game induces an equilibrium in every subgame if no player
i, whenever it is his turn to move, has an incentive to choose an
action different from the one prescribed by si and then follow si in the
continuation (assuming he believes that the opponents stick to si in the
current stage and in the continuation). This principle is valid for a wide
class of dynamic games and it is crucial to simplify their analysis. Here
we prove this principle for a large class of games with observable actions,
which includes all finite games.
Recall that (s|h) is the terminal history induced by strategy profile
s starting from history h. Thus, (s |(h, (ai , si (h))) is the terminal
history induced by choosing action ai Ai (h) immediately after h and
continuing with si afterward, provided that the opponents stick to si at
h and afterward. The gain from a unilateral one-shot deviation ai from s
immediately after history h is:
ui ((s |(h, (ai , si (h)))) ui ((s |h))
Definition 46. A strategy profile s satisfies the One-Shot-Deviation
Property if
h H, i I, ai Ai (h), ui ((s |h)) ui ((s |(h, (ai , si (h)))).
As we did with the notion of subgame perfection, one can express the
One-Shot-Deviation Property in terms of strategic-form payoff functions.

8.3. The One-Shot-Deviation Principle

195

Recall that, for any non-terminal history h = (a1 , ..., at ) and player i I,
shi is the strategy that agrees with si at all histories h0 that do not precede
h and chooses the actions contained in h at all histories preceding h [that is,
shi selects action aki immediately after the initial sub-history (a1 , ..., ak1 )
of history h, for each k = 1, ..., t]. In particular, the equality shi (h) = si (h)
holds by definition. Now change shi at h so that it chooses action ai Ai (h).
The resulting strategy is denoted (si |h ai ). That is,
0

(si |h ai )(h ) =

ai ,
shi (h0 ),

if h = h0
otherwise

In words, si |h ai is the strategy consistent with h that chooses ai at h and

behaves as si for all the other histories not preceding h.
Remark 36. A strategy profile s satisfies the One-Shot-Deviation Property
if and only if for all h H, i I, ai Ai (h),
Ui (shi , shi ) Ui ((si |h ai ), shi )).
We use the Definition 46 to prove the one-shot-deviation principle for
games with finite horizon and the alternative definition implicitly provided
by the remark above to prove the principle for a large class of games
with infinite horizon. (This will help you to familiarize with mathematical
notation.)

8.3.1

Finite horizon

Theorem 18. ( One-Shot Deviation Principle) If has a finite horizon,

then for any strategy profile s the following conditions are equivalent:
(1) s is a subgame perfect equilibrium
(2) s satisfies the One-Shot-Deviation Property.
Proof. (1) implies (2) by definition of subgame perfect equilibrium.
To see this, let si coincide with si at all partial histories except h, where
instead we let si (h) = ai ; then (s |(h, (ai , si (h))) = (si , si |h); since s
is an SPE, we must have
ui ((s |h)) ui ((si , si |h)) = ui ((s |(h, (ai , si (h)))).

196

8. Multistage Games with Complete Information

To show that (2) implies (1) we prove the equivalent contrapositive

statement: if s is not an SPE (i.e., (1) does not hold), then (2) does not
hold either. So, suppose that s is not an SPE. Then, for some player i,
the set of histories where i has an incentive to deviate unilaterally from s
is non-empty. Let us denote this set by HiD (s ) Specifically:


HiD (s ) := h H : si , ui ((s |h)) < ui ((si , si |h)) .
Since the horizon is finite, every subset of H has at least one maximal
element with respect to the precedence relation . In particular, HiD (s )
must have at least one h such that, for every h0 having h as its prefix (for
every h0 coming after h), i has no incentive to deviate at h0 . Thus, let h
be one such maximal element, we will show that condition (2) is violated
at h. Indeed note that for any ai Ai (h) the history h0 = (h, (ai , si (h)))
(immediately) follows h, therefore i has no incentive to deviate at h0 . On
the other hand, i has an incentive to deviate at h with some strategy si .
Therefore, the following inequalities hold:
ui ((s |(h, (si (h), si (h)))) ui ((si , si |(h, (si (h), si (h))))
[no incentive to deviate at h0 = (h, (si (h), si (h)))]
and
ui ((si , si |h)) > ui [(s |h)
[incentive to deviate with si at h].
Observe that, by definition, (si , si |h) = (si , si |(h, (si (h), si (h)))
[starting from h, strategy profile (si , si ) first selects action profile
(si (h), si (h)), leading to history (h, (si (h), si (h))), and then yields
terminal history (si , si |(h, (si (h), si (h))) which must therefore be the
same as (si , si |h), the history induced by (si , si ) starting from h].
Therefore
ui ((s |(h, (si (h), si (h)))) ui ((si , si |(h, (si (h), si (h)))) > ui ((s |h))
and (2) is violated at h for player i and action ai = si (h). 
The one-shot-deviation principle shows that backward induction
computes subgame perfect equilibria.

8.3. The One-Shot-Deviation Principle

197

Proposition 6. In every generic finite game with perfect information

the backward induction strategy profile s is the unique subgame perfect
equilibrium. Every finite game with perfect information has at least one
subgame perfect equilibrium.
Proof. If the perfect information game is not generic, a generalized
backward induction procedure selects actions arbitrarily at histories where
there are multiple maximizers. Let s be the result. By construction, s
satisfies the One-Shot-Deviation Property. Hence, by Theorem 18, s is
a subgame perfect equilibrium. If the game is generic, s is the unique
profile satisfying the One-Shot-Deviation Property and, therefore, it is the
unique subgame perfect equilibrium. 
The uniqueness result can be generalized to all finite horizon games
where by backward inducting we determine a unique action profile for
each history. For example, in the finitely repeated Prisoners Dilemma,
backward induction implies that players cheat in the last stage where
cheating is dominant. Since the backward induction action of the last
stage is history-independent, backward induction implies that players also
cheat in the second-to-last stage, because this yields an immediate gain
and cannot induce a last-stage loss. The argument unravels backward.
The cheat always strategy profile is the unique subgame perfect
equilibrium.17

8.3.2

Infinite horizon and continuity at infinity (optional)

recall that `(h) denotes the length of

For any history h = (a1 , a2 , ...) H,
h, i.e. the number of elements in the sequence (a1 , a2 , ...), with `(h) = if
h is an infinite sequence (hence also a terminal history). For every positive
integer t < `(h), let ht denote the prefix (initial subsequence) of h of length
t. Finally let
Z := {h H : `(h) = }
denote the set of infinite (hence also terminal) histories.
17

It can also be shown that the finitely repeated Prisoners Dilemma has a unique
Nash equilibrium path, which of course must be a sequence of (cheat,cheat).

198

8. Multistage Games with Complete Information

Definition 47. Game is continuous at infinity if either Z = or,18

for all players i I,

oi
h
n


h) : h, e
h Z , ht = e
ht = 0.
lim sup ui (h) ui (e
t

Thus is continuous at infinity if and only if either has finite horizon,

or has infinite horizon and for all  > 0 there is some sufficiently large
positive integer T such that, for all t
of infinite histories

 T and all pairs



t
t
e
e
e
h, h Z satisfying h = h , we have ui (h) ui (h) < . This means that
what happens in the distant future has little impact on overall payoffs.
Example. At every stage all the action profiles
in
set

the Cartesian
S
t

0
A = iI Ai are feasible. Let H = {h }
t>0 A , Z = Z = A .
For every i I and t = 1, 2, ... there is a period-t (flow) payoff function
vi,t : At R . There is some positive real v such that for all i I,
sup |vi,t (ht )| < v.
ht At

Payoffs are assigned to terminal histories as follows: for each player i there
is a discount factor i (0, 1) such that for all h Z,
vi (h ) = (1 i )

X
(i )t1 vi,t (ht ).
t=1

You
can
check
that
such
games
are
continuous
at infinity. Infinitely repeated games with discounting are a special case
where vi,t (a1 , ..., at1 , at ) = vi (at ) for some stage game function vi .
Theorem 19. ( One-Shot Deviation Principle with continuity at infinity).
Suppose that is continuous at infinity. Then for any strategy profile s
the following conditions are equivalent:
(1) s is a subgame perfect equilibrium
(2) s satisfies the one-shot-deviation property.
Proof. (1) implies (2) by definition. To show that (2) implies (1), we
prove the contrapositive: if s is not a subgame perfect equilibrium, then
The case Z = is included for completeness. This is similar to stipulating that a
function with a finite domain is continuous by definition.
18

8.3. The One-Shot-Deviation Principle

199

there must be incentives for one-shot deviations. We rely on Theorem 18

and use the alternative definition of the one-shot-deviation property given
in Remark 36
Suppose that s is not a subgame perfect equilibrium, that is, there are
h H, i I, sbi Si (h) and  > 0 such that
Ui (shi , shi ) = Ui (b
si , shi ) .

(8.3.1)

For anyD positive integer TE, let us define the auxiliary truncated game
T,s = I, (ATi (), uT,s
i )iI with horizon T derived from as follows:
ATi (h) = Ai (h) if `(h) < T and ATi (h) = otherwise, thus
T = {h H
: `(h) T } and Z T = {h H
T :either `(h) = T , or
H
`(h) < T and h Z}
for all h Z Z T , uT,s
i (h) = ui (h),
h
h
for all h Z T \Z, uT,s
i (h) = ui ((s )) (recall that (s ) is the
terminal history induced by strategy profile sh , where sh is the
profile consistent with history h that behaves as s at all histories
not preceding h).

Intuitively, the game is truncated so that it has at most T stages. If

a terminal history h of the truncated game T,s is not terminal for the
original game , then the imputed payoffs are those that would obtain in
if the players followed the given strategy profile s after history h.
By continuity at infinity there is a large enough T such that, for all
e
b
h, h Z\Z T , if e
hT = b
hT (i.e., if these two histories coincide in the first T
stages) then
1
|ui (e
h) ui (b
h)| < .
2
Let sei be the strategy that behave as si after T and as sbi before T (b
si is
the profitable deviation from si postulated in 8.3.1). More precisely

si (h0 ), if `(h0 ) > T
0
sei (h ) =
.
sbi (h0 ), if `(h0 ) T
h. Then,
Let (b
shi , shi ) = b
h and ( sehi , shi ) = e
1
1
Ui (e
shi , shi ) = ui (e
h) > ui (b
h)  = Ui (b
shi , shi ) 
2
2

200

8. Multistage Games with Complete Information

(the equalities hold by definition and the inequality by choice of T and by

construction). Taking 8.3.1 into account,
1
Ui (e
shi , shi ) > Ui (sh ) + .
2
This means that the restriction of s to H T is not an SPE of the finite
horizon game T,s because player i has a profitable deviation from si at
h. Thus, by Theorem 18, there is some player i who can improve on si
with a one shot-deviation ai immediately after some history h0 in T,s . By
definition of T,s this implies that also in game player i has an incentive
to make the one-shot deviation ai immediately after history h0 . 

8.4

Some Simple
Games

Results

About

Repeated

The previous analysis can be applied to study repeated games. For any
static game G = hI, (Ai , vi )iI i where each payoff function vi : A R
is bounded, let ,T (G) denote the T -stage game with observable actions
obtained by repeating G for T times and computing payoffs ui : AT R
as discounted time averages, with common discount factor (0, 1), that
is, for every terminal history z = (a1 , a2 , ...) and each player i
1

ui (a , a , ...) =

T
X

t1 vi (at ).

(8.4.1)

t=1

Note that T may be finite or infinite (for T = , ui is well defined

because vi is assumed to be bounded and therefore the series of discounted
payoffs has a finite sum). As shown above, the one-shot-deviation principle
applies to both cases. In the analysis of finitely repeated games one can
also allow = 1 (no discounting). In the analysis of infinitely repeated
games, it may be convenient to use an equivalent representation of the
payoff associated to an infinite history:
1

u
i (a , a , ...) = (1 )

T
X

t1 vi (at ),

(8.4.2)

t=1

which is a discounted time average of the stream of payoffs {vi (at )}

t=1 .
When the stream of payoffs is constant, e.g. vi (at ) = vi (
a) for each t, then

8.4. Some Simple Results About Repeated Games

201

the average (8.4.2) is precisely this constant:

u
i (
a, a
, ...) = (1 )

T
X

t1 vi (
a) = vi (
a).

t=1

We prove three simple results relating the Nash equilibria of G to

the subgame perfect equilibria of ,T (G). We first apply the OneShot-Deviation Principle to show a rather obvious, but important result:
playing repeatedly a Nash equilibrium of G is subgame perfect. Actually,
playing any fixed sequence of Nash equilibria of G is subgame perfect (for
example, alternating between the equilibrium preferred by Rowena and
the equilibrium preferred by Colin in the repeated Battle of the Sexes is
subgame perfect).
Theorem 20. Let z = (
a1 , a
2 , ...) AT be a sequence of Nash equilibria
t in each stage t19
of G; then the strategy profile s that prescribes to play a
,T
is a SPE of ().
Proof. Let us show that there are no incentives to make one-shot
deviations from s. By the One-Shot-Deviation Principle (Theorem 19),
this implies the thesis.
Fix an arbitrary stage t and a player i. The continuation payoff of i if
he chooses ai in stage t and sticks to si afterward (assuming that everybody
else sticks to si ) is
vi (ai , a
ti )

T
X

t vi (
a ).

=t+1

Note that the second term in this expression does not depend on ai , because
the equilibria of G played in the future may depend on calendar time, but
do not depend on past actions. Since a
t is an equilibrium of G, for every
ai , vi (ai , a
ti ) vi (
ati , a
ti ). Therefore i has no incentive to deviate from
t
a
i in stage t.
By Theorem 20, any set of assumptions that guarantee the existence
of an equilibrium of G also guarantee the existence of a SPE of ,T (G).
The next result identifies situations in which a SPE must be the
repetition of the Nash equilibrium of G.
19

That is, for all t = 1, 2, ..., ht1 At1 and i I, si (ht1 ) = ati .

202

8. Multistage Games with Complete Information

Theorem 21. Suppose that G is finitely repeated (T < ) and has

a unique Nash equilibrium a , then ,T (G) has a unique SPE s that
prescribes to play a always (for every h H and for every i I,
si (h) = ai ).
For example, the finitely repeated Prisoners Dilemma has a unique
subgame perfect equilibrium: always cheat (defect).20
Proof. By Theorem 20 s is a SPE. Let s be any SPE. We will show
that s = s , that is, for every t {0, ..., T 1}, for every h At , s(h) = a
(by convention, A0 := {h0 } is the set containing the empty sequence of
action profiles). The result will be proved by induction on the number k
of stages that still have to be played.
Basis step. Suppose that a history h AT 1 has just occurred, i.e., the
game reached the last stage. Since s is subgame perfect it must prescribe
an equilibrium of G at the last stage. But there is only one equilibrium of
G, a ; thus s(h) = a .
Inductive step. Suppose, by way of induction, that s prescribes a in
the last k periods (k 1), that is, for every period t {T k + 1, ...., T }
and every history h0 of length t 1 (h0 At1 ), s(h0 ) = a . We show
that s prescribes a in period T k for every history h of length T k 1
(h AT k1 ), s(h) = a . Since s is subgame perfect, it must be immune to
deviations, in particular to one-shot deviations, in period T k. Therefore,
by the inductive hypothesis, for every h AT k1 , for every i I and for
every ai Ai ,
vi (s(h)) +

k
X
=1

vi (a ) vi (ai , si (h)) +

k
X

vi (a ).

=1

Note that the summation term is the same on both sides of the inequality
because the inductive hypothesis implies that if s is followed in the last
k periods, future payoffs are independent of current actions. Therefore
the action profile s(h) is such that for every i I, for every ai Ai ,
vi (s(h)) vi (ai , si (h)), implying that s(h) is a Nash equilibrium of G.
But the only Nash equilibrium of G is a ; thus, s(h) = a .
20

The finitely repeated PD is a very simple game. Since the PD has a dominant action
equilibrium, the subgame perfect equilibrium can be obtained by backward induction.
Another result that holds for the finitely repeated PD is that, although it has many
Nash equilibrium strategy profiles, they are all equivalent, and hence they all induce the
permanent defection path.

8.4. Some Simple Results About Repeated Games

203

Comment. There is no need to invoke the One-Shot Deviations

Principle to prove the result above. The principle states that immunity to
one-shot deviations implies immunity to all deviations, which holds under
some (fairly general) assumptions. The proof above used the converse,
immunity to all deviations implies immunity to one-shot deviations, which
is trivially true by definition.
If one of the hypotheses of the previous theorem is removed, i.e., if
either T = or G has multiple equilibria, then the thesis may fail. We
illustrate this with an example and a general result.
Suppose that T < but G has multiple equilibria. Then the Basis
Step of the proof above does not work anymore, because the last stage
equilibrium may depend on past actions. Consider, for example, the
following variation of the PD game:
1\2

4, 4

0, 5

1, 0

5, 0

2, 2

1, 0

0, 1

0, 1

0, 0

Figure 8.6: A modified Prisoners Dilemma.

It can be checked that, for sufficiently large, the following strategies

support the (4, 4) outcome in all periods but the last one:
start playing C;
if you are in period t < T and no deviation from (C, C) has occurred,
play C;
if you are in period t = T and no deviation from (C, C) has occurred,
play D;
if a deviation from (C, C) has occurred, play P in every period until
the last one.
The key observation is that in the second-to-last stage players do
not defect because the immediate gain from defection is more than

204

8. Multistage Games with Complete Information

compensated by the future loss caused by a switch to a worse last-stage

equilibrium.
Next suppose that T = and is sufficiently high, then the
repeated game may have many other equilibria beside the repetition of
the stage game equilibrium a . We show this for the case in which a is
strictly Pareto dominated by some other profile a .
Theorem 22. (Nash reversion, trigger strategy equilibria) Let G =
hI, (Ai , vi )iI i be a static game such that there are an equilibrium a and an
action profile a that strictly Pareto-dominates a . Consider the repeated
game , (G) and the strategy profile s = (si )iI defined as follows:

ai , if h = h0 or h = (a , ..., a ),
si (h) =
ai ,
otherwise.
Then s is a subgame perfect equilibrium if and only if
vi (a ) (1 ) sup vi (ai , ai ) + vi (a ),

(IC)

ai Ai

that is,

supai Ai vi (ai , ai ) vi (a )
.
supai Ai vi (ai , ai ) vi (a )

Note that by definition supai Ai vi (ai , ai ) vi (a ) and by assumption

vi (a ) > vi (a ). Therefore
0

supai Ai vi (ai , ai ) vi (a )
< 1.
supai Ai vi (ai , ai ) vi (a )

Strategies such as those defined in the theorem are called trigger

strategies because a deviation from the equilibrium path triggers some
sort of punishment.
Proof of Theorem 22. We show that there are no incentives to make
one-shot deviations if and only if condition (IC) holds. By Theorem 19
(the One-Shot Deviation Principle), this implies the result.
There are two types of histories: those for which a deviation from a
has not yet occurred and the others. For the latter histories s prescribes to
play a forever after. Since a is an equilibrium of G, Theorem 20 implies

8.5. Multistage Games With Chance Moves

205

that there cannot be incentives to make one-shot deviations. Suppose now

that no deviation from a has yet occurred. Then s prescribes to play
a . Consider the incentives of player i. His continuation payoff if he sticks
to si is vi (ai ), because i expects to get vi (ai ) forever (recall that the
continuation payoff is a discounted time average). His continuation payoff
if he deviates and plays ai 6= ai is (1 )vi (ai , ai ) + vi (a ), because i
expects that his opponents stick to ai in the current period, and that his
deviation triggers a switch to a from the following period (as prescribed
by s ). Therefore i does not have an incentive to make a one-shot deviation
if and only if condition (IC) holds.
As an application of Theorem 22 consider, for example, an infinitely
repeated Cournot oligopoly game: a is an output profile, vi (a) is the oneperiod profit of firm i, a is the Cournot equilibrium profile, a is the
joint-profit maximizing profile, i.e. the profile that would be chosen if
the firms were allowed to merge and the merged firm were not regulated.
The trigger-strategy profile s may be interpreted as a collusive agreement
to keep the price at the monopoly level. If is above the threshold
the collusive agreement is self-enforcing. Self-enforcing in this context
does not only mean that there are no incentive to deviate from the
equilibrium path (all Nash equilibria satisfy this property), but also that
the punishment or war that is supposed to take place after a deviation
is itself self-enforcing, hence credible, and this is indeed the case because
playing the Cournot profile in each period is an equilibrium (Theorem 20).

8.5

Multistage Games With Chance Moves

The definition of multistage game can be extended to allow for the

possibility that some events in the game depend on chance. Chance
is modeled as a fictitious player, denoted by 0, that chooses actions
with exogenously given conditional probabilities. This means that the
probability that a particular chance action a0 is selected immediately after
history h is a number 0 (a0 |h), which is part of the description of the
game.
We maintain the convention that at each stage there are simultaneous
moves by all players, including chance. Since the set of active players
(those with more than one feasible action) may depend on time and on

206

8. Multistage Games with Complete Information

the previous history, this assumption does not entail any loss of generality.
For example, one may model games where chance moves occur before or
after the moves of real players. To minimize on notational changes we also
ascribe a utility function u0 to the chance player 0, but we assume that it
is constant. This is just another notational trick.
A multistage game with observable actions and chance moves
is a structure

n

= Ai , Ai (), ui i=0 , 0 .
The symbols Ai , Ai (), ui have the same meaning as before. From the
constraint correspondences Ai () (i = 0, ..., n) we can derive the set of
the set of terminal histories Z and the set of nonfeasible histories H,

terminal histories H = H\Z

(see section 8.1). Functions ui are real-valued
and have domain Z. The only novelty is player 0, chance: for every z Z,
u0 (z) = 0, and 0 describes how the probabilities of chance moves depend
on the previous history, that is,
0 = (0 (|h))hH hH o (A0 (h))
where o (A0 (h)) is the set of full support probability measures on A0 (h).21
Next we define the strategic form. To avoid measure-theoretic
technicalities we will henceforth assume that all the sets A0 (h) (h H) are
finite. Let I be the set of the real players, i.e. I = {1, ..., n}. The set Si
of strategies of i I is defined as before. Although strategies of chance
are well defined mathematical objects, we disregard them, as chance is not
a strategic player. Thus, the set of strategy profiles is S := iI Si and
similarly Si := jI\{i} Sj . The presence of chance moves implies that
more than one terminal history may be possible when the players follow
a particular strategy profile s; let Z(s) denote the set of such terminal
histories. The probability of each z Z(s) depends on the probability of

the chance moves contained in z and is denoted (z|s)

(if a player, or an

external observer were uncertain about the true strategy profiles, (z|s)
would represent the probability of z conditional on the players following
the strategy profile s, which explains the notation). Therefore the outcome
function when there are chance moves has the form : S (Z). The

formal definition of Z(s) and (|s)

is as follows. Let ati (z) denote the
21

Recall that o (X) = { (X) : Supp = X}, the superscript o stands for
(relatively) open in the hyperplane of vectors whose elements sum to 1.

8.5. Multistage Games With Chance Moves

207

action of player i at stage t in history z, let ht1 (z) denote the prefix
(initial subhistory) of z of length t, and recall that `(z) is the length of z.
Then
s S, Z(s) = {z Z : t {1, ..., `(z)}, i I, ati (z) = si (ht1 (z))},
( Q
`(z)
t
t1 (z)), if z Z(s),

t=1 0 (a0 (z)|h

z Z, (z|s)
=
0,
if z
/ Z(s).

In words, (z|s)
is the product of the probabilities of the actions taken by
the chance player in history z.
A Nash equilibrium of is a strategy profile s such that for every
i I and for every si Si ,
X
X

(z|s
)ui (z)
(z|s
i , si )ui (z).
z

in other words, a Nash equilibrium of is an equilibrium of the static game

where players
P simultaneously choose strategies and have payoff functions
Ui (s) = z (z|s)u
i (z) (i I, s S).
In order to define subgame perfect equilibria, we first define the

conditional outcome function (|h,

s): for each h H, the set of terminal
histories that can be reached when s is followed starting from h (where h
may be inconsistent with s) is
Z(h, s) = {z Z : h z, t {`(h)+1, ..., `(z)}, i I, ati (z) = si (ht1 (z))},
then

z Z, (z|h;
s) =

( Q
`(z)

t
t1 (z)),
t=`(h)+1 0 (a0 (z)|h

0,

if z Z(h, s),
if z
/ Z(h, s).

A strategy profile s is a subgame perfect equilibrium if for every

h H, for every i I and for every si Si ,
X
X

(z|h,
s )ui (z)
(z|h,
si , si )ui (z).
z

Let (|h,
ai , s) (Z) denote the probability measure on Z that
results if s is followed starting from h, except that player i chooses

208

8. Multistage Games with Complete Information

ai Ai (h) at h (can you give a formal definition?). A strategy profile

s has the One-Shot-Deviation Property (OSDP) if for every h H,
for every i I and for every ai Ai (h),
X
X

(z|h,
s)ui (z)
(z|h,
ai , s)ui (z).
z

It can be shown that in every game with finite horizon or with continuity
at infinity (hence in every game with discounting) a strategy profile is a
subgame perfect equilibrium if and only if it has the OSDP.

8.6

Randomized Strategies

In this section we introduce randomization in multistage games with

observable actions. To simplify the probabilistic analysis we focus on finite
games.
In the context of multistage games (and more generally of all games
with a sequential structure) one can think of two types of randomization:
(1) player i chooses a pure strategy at random according to a
probability measure i (Si ), which is called a mixed strategy;
(2) for each non terminal history h H player i would choose an
action at random according to a probability measure i (|h) (Ai (h));
the array of probability measures i = (i (|h))hH is called behavioral
strategy.22
Note that we can always regard a pure strategy as a degenerate
randomized strategy: si can be identified with the mixed strategy i such
that i (si ) = 1 and with the behavioral strategy i such that for every
h H, i (si (h)|h) = 1.

8.6.1

Realization-equivalence
behavioral strategies

between

mixed

and

The literal interpretation of randomized strategies is that players spin

roulette wheels or toss coins and let their actions be decided by the
outcomes of such randomization devices. But this seems a bit farfetched,
22
Kuhn (1953) introduced the concept and the terminology in his analysis of general
games with a sequential structure. Obviously, i (|h) is non-trivial only at histories
where i is active.

8.6. Randomized Strategies

209

especially in the case of mixed strategies. Furthermore, agents who

maximize expected utility have no use for randomization, as pure choices
always do at least as well as randomized choices.
As in static games, a randomized strategy of i can be interpreted as a
representation of the opponents beliefs about the pure strategy of i. For
example, suppose that game is played by agents drawn at random from
large populations; let i (si ) be the fraction of agents in population i (the
population of agents that may play in role i) that would use strategy si . If
j knows the statistical distribution i , this will also represent the belief of
j about the strategy of i. On the other hand, i (ai |h) may be interpreted
as the conditional probability assigned by j to action ai given history h.
This belief interpretation suggests a relationship between mixed and
behavioral strategies. Suppose that h has just realized, what does anyone
learn about the pure strategy played by i? She learns that i is using a
strategy that allows (does not prevent) h. Let Si (h) be the set of such
strategies. Formally,23
Si (h) := {si Si : si Si , h (si , si )}.
Next we define the set of strategies of i that allow h and select ai at h:
Si (h, ai ) := {si Si (h) : si (h) = ai }.
Now suppose that i is derived from the statistical distribution i under
the independence-across-players assumption that what is observed about
the behavior of other players does not affect the beliefs about the strategy
of i. Then the probability of ai given h is just the fraction of agents in
population i using a strategy that allows h and selects ai at h, divided
by the fraction of agents in population i using a strategy that allows h (if
positive), that is,
h H, i (Si (h)) > 0 i (ai |h) =

i (Si (h, ai ))
i (Si (h))

(8.6.1)

P
[for any subset X Si , I write i (X) = si X i (si )]. If i (Si (h)) = 0,
i (|h) can be specified arbitrarily. Note that (8.6.1) can be written in a
more compact, but slightly less transparent form:
h H, i (ai |h)i (Si (h)) = i (Si (h, ai )).
23

If the game has chance moves the definition in the text must be adapted by including
in si also s0 , the strategy of the chance player.

210

8. Multistage Games with Complete Information

The formula above, or (8.6.1), says that i and i are mutually consistent
in the sense that they jointly satisfy a kind of chain rule for conditional
probabilities.
Definition 48. A mixed strategy i (Si ) and a behavioral strategy
i = (i (|h))hH hH (Ai (h)) are mutually consistent if they satisfy
(8.6.1).
Remark 37. If mixed strategy i is such that i (Si (h)) = 0 for some h
where player i is active, then there is a continuum of behavioral strategies
i consistent with i . If i (Si (h)) > 0 for every h where i is active, then
there is a unique i consistent with i . If i is active at more than one
history, then there is a continuum of mixed strategies i consistent with
any given behavioral strategy i .
The last statement in the remark can be understood with a countingdimensionality argument. Let Hi := {h H : |Ai (h)| 2} denote
the set of histories where
Q i is active. In a finite game, the number of
elements of SQ
i is |Si | =
hHi |Ai (h)| , thus the dimensionality of (Si )
of
is |Si | 1 = hHi |Ai (h)| 1. OnPthe other hand, the dimensionality
P
the set of behavioral strategies is hHi (|Ai (h)| 1) = hHi |Ai (h)|
|H
shown (by induction on the cardinality of Hi ) that
Q i |. It can be P
24
|A
(h)|

i
hHi
hHi |Ai (h)| because |Ai (h)| 2 for each h Hi .
Thus the dimensionality of (Si ) is higher than the dimensionality of
hH (Ai (h)) if |Hi | 2.
Example 25. In the game tree in Figure 3, Rowena (player 1) has
4 strategies, S1 = {out.u, out.d, in.u, in.d}; two of them out.u and out.d
are realization-equivalent and correspond to the plan of action out. The
figure labels terminal histories: v = (out), w = (in, (u, l)) etc. The set of
non-terminal histories is H = {h0 , (in)} (recall that h0 denotes the initial,
empty, history) and S1 (h0 ) = S1 , S1 (in) = {in.u, in.d}.
24
First note that for every integer L 1, 2L1 L. (This can be easily proved
by induction: it is trivially true for L = 1; suppose it is true for some L 1, then
2L = 2 2L1
Q
PL 2 L = L + L L + 1.) Let nk 2 for each k = 1, 2, .... I prove that
L
n

k
k=1
k=1 nk for each L = 1, 2, .... Let n = max{n1 , ..., n` }. Then
L
Y
k=1

nk n 2L1 n L

L
X
k=1

nk .

8.6. Randomized Strategies

211

in

1
out

1\2

v
Figure 8.7: A game tree.

Consider the following mixed strategy, parametrized by p [0, 1]:

1p (out.u) =
1p (in.u) =

p p
1p
, (out.d) =
,
2 1
2
1 p
2
, 1 (in.d) = .
6
6

Since no history in H is ruled out by 1 , there is only one behavioral

strategy consistent with 1 :
1 (in|h0 ) =
1 (u|in) =

1 2
1
1 (S1 (in))
= + = ,
1 (S1 )
6 6
2
1 (S1 (in, u))
=
S1 (in)

1
6
1
6

2
6

1
= .
3

Note also that 1 does not depend on the distribution p : 1 p of

the probability mass 1p (S1 (out)) = 21 between the realization-equivalent
strategies out.u and out.d. The reason is that the only difference between
these two strategies is the choice between u and d contingent on in, a
counterfactual contingency if Rowena plans to go out. Now consider the
mixed strategy p1 with p1 (out.u) = p, p1 (out.d) = 1 p, which amounts
to the deterministic plan of going out. Since p1 (S1 (in)) = 0, there is a
continuum of behavioral strategies consistent with p1 , all those in the set
{ 1 ({out, in}) ({u, d}) : 1 (out|h0 ) = 1}.
Next suppose that player i does indeed randomize according to
behavioral strategy i . Is there a method to obtain a mixed strategy i
consistent with i ? Here is an answer: It is natural, if not entirely obvious,
to consider the case where the randomization devices that i would use at

212

8. Multistage Games with Complete Information

different histories are stochastically independent. We call this assumption

independence across agents, as one can regard the contingent choice of
i at each h H as carried out by an agent (i, h) of i who only operates
under contingency h. A pure strategy si is just a collection (si (h))hH of
contingent choices of the agents (i, h) , h H. If the random choices at
different histories are mutually independent, then the probability of pure
strategy si is the product of the probabilities of the contingent choices
si (h), h H. Therefore one can derive from i the mixed strategy i that
satisfies
Y
si Si , i (si ) =
i (si (h)|h).
(8.6.2)
hH

It is boring, but rather straightforward to show that such i is consistent

with i :25
Remark 38. For any i N , i (Si ), i hH (Ai (h)), if (8.6.2)
holds, then also (8.6.1) holds, that is, i and i are mutually consistent.
Example 26. In the game tree of Figure 3, the mixed strategy
associated to 1 (in|h0 ) = 1 (out|h0 ) = 21 , 1 (u|in) = 13 , 1 (d|in) = 32
under independence-across-agents assumption is 1 :
1 (out.u) =
1 (in.u) =

1 1
1
1 2
2
= , 1 (out.d) = = ,
2 3
6
2 3
6
1
1 2
2
1 1
= , 1 (in.d) = = .
2 3
6
2 3
6

Note that 1 was derived Example 25 from an arbitrary mixed strategy

with the parametric representation 1p , whereas here I derived exactly
1/3
1 = 1 . The reason is that only 1p with p = 31 is consistent with
independence across agents.
Remark 37 says that many behavioral strategies may be consistent
with a given mixed strategy and that many mixed strategies may be
consistent with a given behavioral strategy, although only one satisfies
the assumption of independence across agents. Then why is consistency
between mixed and behavioral strategies important? Intuitively, the reason
is that mutually consistent mixed and behavioral strategies yield the same
25

This is one part of what is known as Kuhns theorem for mixed and behavioral
strategies. For a proof, see the appendix.

8.6. Randomized Strategies

213

probabilities of histories, which is all that matter to expected payoff

maximizing players. To make the intuition precise, note that a profile
of randomized strategies ( = = (i )iI or = = (i )iI ) induces a

probability measure (|)

Z on terminal histories, which is determined
by the following formulas (which implicitly assume independence across
players):26
!
X
Y

z Z, (z|)
=
i (si ) ,
s:(s)=z

iI

`(z)

z Z, (z|)
=

YY

i (ati (z)|ht1 (z)).

t=1 iI

(Kuhn)27

Theorem 23.
For any two profiles of mixed and behavioral
strategies and the following statements hold:
(a) Fix any i I, if i and i satisfy either (8.6.1) or (8.6.2) then
i , si ) = (|
i , si ).
si Si , (|
(b) If, for all i I, i and i satisfy either (8.6.1) or (8.6.2) then

(|)
= (|).
Example 27. Again, the example of Figure 3 illustrates. For Rowena
(player 1) consider the randomized strategies 1p and 1 of Example
25. Colin (player 2) is active at only one history, hence there is an
obvious isomorphism between his mixed and behavioral strategies. Let
1 , l) = (|
1 , l), (|
1 , r) = (|
1 , r) and
2 (l) = 2 (l|in) = q. Then (|

(|) = (|). In particular

(v|)
=

(y|)
=
26

= (v|),
2
q

= (y|),
3

q
1q

(w|)
= = (w|),
(x|)
=
= (x|),
6
6
1q

(z|)
=
= (z|).
3

Recall that (s) is the terminal history induced by strategy profile s. We introduced
and ht1 (z) in the previous section: ati (z) is the action played by i at stage t in
history z, ht1 (z) is the prefix of length t 1 of z.
27
In his seminal (1953) article, Harold Kuhn proved two important results about
general games with a sequential structure (so called extensive form games ), one
concerns the realization-equivalence of mixed and behavioral strategies under the
assumption of perfect recall (a generalization of the observable actions assumption),
the other concerns the existence of equilibria.

ati (z)

214

8. Multistage Games with Complete Information

8.6.2

Randomized equilibria

A technical advantage of randomization via behavioral strategies is that it

allows a simple definition and characterization of randomized subgame

perfect equilibria. Let (|h,

) denote the probability measure over
terminal histories induced by starting from h:
( Q


Q
`(z)
i (ati (z)|ht1 (z)) , if h z
iI

t=`(h)+1
z Z, (z|h, ) =
0,
otherwise

Similarly, for ai Ai (h) let (|h,

ai , ) denote the probability measure
induced from h by the profile |h ai where i (|h) in is replaced by ai
played with probability one.28 Using this notation we can define Nash
and subgame perfect equilibria in behavioral strategies, and the One-ShotDeviation Property of behavioral strategy profiles:
Definition 49. A behavioral strategy profile is a Nash equilibrium if
for every i I, for every i hH (Ai (h)),
X

(z|
)ui (z)

(z|
i , )ui (z).

A behavioral strategy profile is a subgame perfect equilibrium if for

every h H, for every i I and for every i hH (Ai (h)),
X

(z|h,
)ui (z)

(z|h,
i , )ui (z).

has the One-Shot-Deviation Property if for every h H, for every

i I and for every ai Ai (h),
X
z
28

(z|h,
)ui (z)

(z|h,
ai , )ui (z).

The formula is

Z, (z|h,
ai , ) =
Q
( Q
Q
`(h)+1
`(z)
t
t1
(z)|h`(h) (z))
(z)),
j6=i j (aj
jN j (aj (z)|h
t=`(h)+2
0,

`(h)+1

if h z, ai

(z) = ai

otherwise

8.6. Randomized Strategies

215

Let us characterize these properties and relate them to each other.

From each probability measure over Z one can derive the probability
of every non-terminal history. In particular, the probability of h H
implied by is the summation
of the probabilities of the terminal histories
P
). The following remark says that a

following h: P[h| ] := z:hz (z|

Nash equilibrium in behavioral strategies induces a Nash equilibrium
in each subgame with root h that can be reached with positive probability,
i.e. such that P[h| ] > 0. Therefore, a Nash equilibrium may be
imperfect only if P[h| ] = 0 for some h H.
Remark 39. A behavioral strategy profile is a Nash equilibrium if and
only if for every i I, for every i hH (Ai (h)) and for every h H
X
X

P[h| ] > 0
(z|h,
)ui (z)
(z|h,
i , )ui (z).
z

By standard results on expected utility maximization we can restate

the One-Shot-Deviation Property as follows:
Remark 40. A behavioral strategy profile has the One-Shot-Deviation
Property if and only if for every h H and for every i I,
X

Suppi (|h) arg max

(z|h,
ai , )ui (z).
ai Ai (h)

Adapting arguments used for pure strategy profiles one can show that
the One-Shot-Deviation Principle also holds for behavioral profiles:
Theorem 24. ( One-Shot-Deviation Principle) In every finite game a
behavioral strategy profile is a subgame perfect equilibrium if and only if
it has the One-Shot-Deviation Property.
The One-Shot-Deviation Principle allows a relatively simple proof of
existence of randomized equilibria in finite games:
Theorem 25. (Kuhn) Every finite game has at least one subgame perfect
equilibrium in behavioral strategies.
Proof. We provide only a sketch of the proof. Construct a profile
with a kind of backward induction procedure. All histories with
depth one29 define a static last stage game that has at least one mixed
29

Recall that the depth of h is d(h) = maxz:hz [`(z) `(h)]

216

8. Multistage Games with Complete Information

equilibrium. Thus, to every h of depth one we can associate a stagegame equilibrium (|h) iI (Ai (h))
payoff profile
Q and an equilibrium

P

(vi (h))iI [where vi (h) =

aA(h)
jI j (aj |h) ui (h, a)]. Suppose

that a mixed action profile (|h) and a payoff profile (vi (h))iI has been
assigned to each history with depth k or less. Then one can assign a mixed
action profile (|h) and a payoff profile (vi (h))iI to each

history h with
depth k + 1: just pick a mixed equilibrium of the game I, (Ai (h), vi )iI
where for every i I and for every a A(h), vi (a) = vi (h, a) (note that
(h, a) has depth k or less, therefore vi (h, a) is well defined). The profile
constructed in this way must satisfy the One-Shot-Deviation property
and therefore it is a subgame perfect equilibrium.

8.7. Appendix

8.7
8.7.1

217

Appendix
Epistemic Analysis Of The ToL4 Game

In this appendix we prove the following:

Claim. All terminal histories of the ToL4 game but (L, L, L, L) and
(L, L, L, T ) are consistent with rationality and initial common belief in
rationality.
In order to do this we use a mathematical structure similar to those
used to analyze incomplete information games, that is, a set of possible
worlds specifying players types. The important difference is that here
a type represents a possible way to play and think according to the
observed history. Thus a type ti specifies a plan of action (we need not
distinguish between realization-equivalent strategies such as T T and T L in
the ToL4 game) and an array of conditional probability measures i (|h),
h H. The key feature of the mathematical structure is that i (|h) is not
just a belief about the strategies of the opponent, it is also a belief about
the beliefs of the opponent; in other words, i (|h) is a belief about the
types of the opponent. From i (|h) we can recover the (conditional)
first-order belief about the strategies of i. Of course, we assume that
all the strategies of i that are inconsistent with h have zero probability
according to i (|h): he believes what he sees (h), therefore he assigns
probability one to Si (h) upon observing h. We also assume that the
standard rules for updating probabilities are applied whenever possible,
i.e. whenever the conditioning event h was not given probability zero
before being observed.
Proof (by example). The two tables below represent a structure
of the kind just described. Their purpose is to show that for each z
{(T ), (L, T ), (L, L, T )} there is a state of the world = (t1 , t2 ) consistent
with rationality and initial common belief in rationality such that the
strategies played at yield z. Thus, we want to exhibit possibilities, not
to predict how the game is played.
In the tables there are three types for each player, that is, as many
types as there are plans of action (strategies of the reduced normal form).
The row corresponding to ti specifies the plan of ti (how ti would play
at each relevant history) and the beliefs that ti would hold at the initial

218

8. Multistage Games with Complete Information

history h0 and each (other) history where i is active; these beliefs are
represented by the triple (i (t1i |h), i (t2i |h), i (t3i |h)); I am considering
deterministic beliefs just for simplicity. For example, type t11 of player 1
(Rowena), who takes immediately, starts thinking that the type of 2 (Colin)
is t12 , and hence that Colin would take immediately if given the opportunity;
but ti would be forced to change her mind at history h = (L, L). The belief
that t11 would have at this history is that the type of Colin is t22 , and thus
that Colin would take at the next round if given the opportunity. [Note
that Colin must assign probability one to strategy LL of Rowena if history
(L, L, L) occurs, as this is the only strategy of Rowena consistent with
(L, L, L). Since Colin thinks that t31 is the only type of Rowena playing
LL, he assigns probability one to t31 at h = (L, L, L). This explains the
last column of the second table.]
type of 1

plan of 1

beliefs at h0

bel. at h = (L, L)

t11

(1, 0, 0)

(0, 1, 0)

t21
t31

LT

(0, 1, 0)

(0, 1, 0)

LL

(0, 0, 1)

(0, 0, 1)

type of 2

plan of 2

beliefs at h0

bel. at h = (L)

bel. at h = (L, L, L)

t12
t22
t32

(1, 0, 0)

(0, 1, 0)

(0, 0, 1)

LT

(1, 0, 0)

(0, 0, 1)

(0, 0, 1)

LL

(0, 0, 1)

(0, 0, 1)

(0, 0, 1)

By inspection of the tables one can see that all types but t32 are
rational, in the sense they their plan of action maximizes the players
conditional expected payoff for each history consistent with the plan itself
(in this analysis we may neglect the histories inconsistent with the plan
of ti when we check for the rationality of ti , this make things slightly
simpler). Thus, the set of states where (sequential) rationality holds is
R = {t11 , t21 , t31 } {t12 , t22 }.
Now note that type t31 of Rowena is rational, but does not believe in the
rationality of Colin. On the other hand, in each state of the world in the
= {t1 , t2 } {t1 , t2 } R each player initially assigns probability
subset
1 1
2 2
Thus, at each state in
players
one to some type of the other player in .
are not only rational they are also initially certain of rationality.

8.7. Appendix

219

there is rationality
Suppose, by way of induction, that at each state in
and initial mutual belief in rationality up to order k. Since at each state in
the players initially assign probability one to (the opponents component

itself, then there is also mutual initial belief in rationality of order

of)
there is rationality
k + 1. It follows by induction that at each state in
and initial common belief in rationality.
Finally note that all states of the form (t11 , t2 ) yield history (T ) ($1 is
taken in the first period and this ends the game), state (t21 , t12 ) yields history
(L, T ), and state (t21 , t22 ) yields history (L, L, T ). Therefore we have proved
our claim: each terminal history z {(T ), (L, T ), (L, L, T )} is consistent
with rationality and initial common belief in rationality. 
It can be shown that in generic, finite games of perfect information
the set of outcomes consistent with rationality and initial common belief
in rationality is obtained by looking at the strategic form of the game
and performing one round of elimination of weakly dominated strategies,
followed by the iterated deletion of strictly dominated strategies. Let
us illustrate this result in the ToL4 game. Since realization-equivalent
strategies are not distinguishable via weak or strict dominance relations,
we may look at the slightly simpler reduced strategic form, where strategies
T L and T T are coalesced into the plan of action T :
1\2

LT

LL

1, 0

1, 0

1, 0

LT

0, 2

3, 0

3, 0

LL

0, 2

0, 4

4, 0

There is only one weakly dominated strategy: the third column,

s2 = LL. Once the third column is deleted, the third row, s1 = LL is
strictly dominated (by T ) in the reduced matrix:
1\2

LT

1, 0

1, 0

LT

0, 2

3, 0

LL

0, 2

0, 4

220

8. Multistage Games with Complete Information

In the further reduced matrix
1\2

LT

1, 0

1, 0

LT

0, 2

3, 0

no strategy is strictly dominated. The three remaining outcomes (1, 0),

(0, 2), and (3, 0) respectively correspond to the terminal histories (T ),
(L, T ), and (L, L, T ).

8.7.2

Mixed and Behavioral Strategies

H and a
(it is
Proof of Remark 38. Fix i, i , i , h
i Ai (h)
notationally convenient to put a hat on the fixed history and action).
> 0 implies
Suppose that (8.6.2) holds; it must be shown that i (Si (h))

i (
ai |h) = i (Si (h, a
i ))/i (Si (h)). For any h h, let
ai (h) denote the
from h (in other words,
action taken by i at h to reach h
a(h) = (
aj (h))jI

is the unique action profile such that (h,

a(h))  h). Then, Si (h) = {si :
si (h) =
a
=a
si (h) =
h h,
ai (h)}, Si (h,
i ) = {si : si (h)
i , h h,
ai (h)},

so that for each si Si (h),

Y
Y
Y
i (si (h)|h) =
i (
ai (h)|h)
i (si (h)|h) ,
hH

hH:hh

hH:hh

a
and for each si Si (h,
i ),

Y
Y

i (si (h)|h) =
i (
ai (h)|h)i (
ai |h)
hH

hH:hh

i (si (h)|h) .

hH:hh

Therefore,

i (Si (h))
=

i (si ) =

si Si (h)

hH:hh

i (si (h)|h)

hH
si Si (h)

i (
ai (h)|h)

hH:hh

si Si (h)

i (si (h)|h) ,

8.7. Appendix

221
X

a
i (Si (h,
i )) =

i (si ) =

i (si (h)|h)

i ) hH
si Si (h,a

i)
si Si (h,a

i (
ai (h)|h) i (
ai |h)

hH:hh

i (si (h)|h) ,

ai ) hH:hh

si Si (h,

in any order,
Count the partial histories that do not weakly precede h

e.g. {h H : h  h} = {h1 , ..., hL }. Now note that there is a canonical

and Ai (h)
L Ai (hk ) , and between Si (h,
a
bijection between Si (h)
i )
k=1
L
and k=1 Ai (hk ). Given this, we can re-order terms as follows:
X

L
X

i (si (h)|h) =

i (ai,k |hk ) = 1

k=1 ai,k Ai (hk )

ai ) hH:hh

si Si (h,

i (si (h)|h) =

hH:hh

si Si (h)

i (ai |h)

L
X

i (ai,k |hk ) = 1.

k=1 ai,k Ai (hk )

ai Ai (h)

Then
Y

i (Si (h))
=

i (
ai (h)|h),

hH:hh

a
i (Si (h,
i )) =

i (
ai (h)|h) i (
ai |h)

hH:hh

and

i (Si (h))
= i (

i (
ai |h)
ai |h)

hH:hh

a
i (
ai (h)|h) = i (Si (h,
i ))

Multistage Games with

Incomplete Information
[UNDER CONSTRUCTION!!!]
These notes (partly written for slides) extend the analysis of static
games with incomplete information to dynamic game forms with a
multistage structure. At this stage, there are redundancies and the
exposition is not homogeneous in terms of details and examples. The
chapter is just a collage of notes.
To make these notes self-contained, the definition of multistage game
form is repeated here.
These notes contain only a summary of the notation and definitions.

9.1

Multistage game forms

We consider multistage game forms with observable actions, assuming

without loss of generality that all players simultaneously take an action
at each stage:
D

E
I, Ai , Ai ()
iI

I = {1, ..., n} set of players

Ai set of actions that may become available for i at some point of the
game
222

9.1. Multistage game forms

223

!
Ai () :
A1 ... An

S t
A
t0
, At

Ai constraint correspondence of i where A :=

... A is the set of sequences of action profiles of

=A
}
| {z
t times

length t;
by convention A0 := {h!0 } where h0 is the empty sequence;
S t
sequences h
A are called histories; Ai (h) Ai is the set of
t0

actions available to i immediately after h,!it can be empty or a singleton.

S t
Assumption: for every h
A , either (for every i, Ai (h) 6= )
t0

or (for every i, Ai (h) = ) (Ai (h) = means

! game over ).
S t
Natural partial order on
A : h prefix (initial sub-history)
t0

of h0 , written h h0 , if h = (a1 , ..., ak ), h0 = (a1 , ..., ak , ak+1 , ..., a` ); by

convention the
! empty seq. is a prefix of every non-empty seq.: for every
S t
A , h0 h.
h
/
t1

the set of feasible histories, Z := {h H

: i, Ai (h) = }
DERIVE: H

the set of terminal

histories, H := H\Z the set of partial, or non-terminal
is the game tree.
histories. H,
h0 H
by convention; let A(h) := A1 (h)... A(h)n ,
[Derivation of H:
1
t

then (a , ...a ) H if and only if a1 A(h0 ) and ak+1 A(a1 , ..., ak ) for
each k = 1, ..., t 1.]
Whenever we do not say otherwise the following applies:
!
S
t

Assumption (finiteness):
For some T , H
A ,
0tT

is finite.
furthermore H

224

9.2

9. Multistage Games with Incomplete Information

Dynamic environments
information

with

incomplete

A dynamic environment with incomplete information is obtained by adding

to the previous structure a set of states of nature and payoffs that depend
on the state of nature and the terminal history (i.e. all the actions taken
until the game is over):
D

E
I, 0 , i , Ai , Ai (), ui iI ,
with ui : Z R, = 0 1 ... n .
[At the cost of more complex notation we may generalize and allow that
i affects the set of feasible actions of i: Ai (, ) : i H  Ai , Ai (i , h)
is the set of actions that are feasible for information-type i immediately
after h.]
Strategies:
in line with the genuine incomplete information
interpretation we assume that there is no ex ante stage: players are born
with private information. Thus, we do not define strategies as choice rules
that depend on private information and history, but rather as choice rules
that only depend on history:
Si := hH Ai (h)
equivalently
n
o

H
Si = si Ai
: h H, si,h Ai (h) .

9.2.1

Rationalizability

Notions of rationalizability and -rationalizability can be defined for

such an environment, extending the solution concepts presented for static
games. Intuitively we have to modify the static games definitions as
follows:
- we iteratively delete pairs (i , si ),
- player i holds beliefs i (|h) about (0 , i , si ) at each h H, beliefs
at different histories are related by the rules of conditional probabilities
whenever possible,
- the rationality condition for (i , si ) given beliefs (i (|h))hH is that
si is a best response for i to i (|h) at each h consistent with si ,

9.2. Dynamic environments with incomplete information

225

- we require at least the following: initial common belief in rationality

(at each deletion step k, the initial beliefs of each type assign zero
probability to the pairs eliminated in the previous k 1 steps),
- stronger notions of rationalizability incorporating a forward
induction principle may be obtained by requiring that every i, for each
h H, ascribes to i the highest degree of strategic sophistication
consistent with h (in particular, i believes whenever possible that i is a
conditional expected payoff maximizer).
We do not further elaborate on this topic here.

226

9.3

9. Multistage Games with Incomplete Information

Dynamic Bayesian Games

D

E
Structure I, 0 , i , Ai , Ai (), ui iI does not specify the exogenous
interactive beliefs of the players, therefore it is not rich enough to define
standard notions of equilibrium.
As in the case of static games we should add interactive beliefs
structures (type spaces) `
a la Harsanyi.
Since the dynamics involve additional complications of strategic
analysis, we simplify the interactive beliefs aspect: we assume that the
set of types `
a la Harsanyi Ti coincide with (more precisely, is isomorphic
to) i . To emphasize this assumption we use the phrase simple Bayesian
game :
Let pi (|i ) (0 i ) denote the exogenous belief of type i . A
simple dynamic Bayesian game is a structure
D

E
= I, 0 , i , Ai , Ai (), ui , (pi (|i ))i i iI
We could have also specified priors pi () with pi (i ) > 0 for all
i and let pi (0 , i |i ) = pi (0 , i , i )/pi (i ).

9.3.1

Bayesian equilibrium

Each strategy profile s S := S1 ... Sn induces a terminal history,

denoted (s). Using the mapping : S Z we can derive the interim
strategic form of the Bayesian game , that takes every i and every profile
of strategies (sj )jI,j j jeI (Sj )j (one strategy for each type of each
player) and determines the corresponding expected payoff for i :
Ui ((sj )jI,j j ) =

pi (0 , i |i )ui 0 , i , i , (si , si ) ,

0 ,i

with si

= (sj )jI\{i} Si .

DEF. A Bayesian equilibrium of is a Nash equilibrium of the interim

strategic form of .
[If we define using priors, we can equivalently characterize Bayesian
equilibrium as the Nash equilibrium of the ex ante strategic form.]

9.3. Dynamic Bayesian Games

We will consider pure and randomized equilibria.
equilibria players use behavioral strategies

227
In randomized

i = (i (|i , h))i i ,hH hH ((Ai (h)))i ,

where i (ai |i , h) is the probability of choosing ai given i and h.
Possible interpretation: i (ai |i , h) is the fraction of agents in pop. i
taking action ai at h, among those whose type is i and whose strategy
does not prevent h.

228

9.4

9. Multistage Games with Incomplete Information

Bayes Rule

Consider an agent i who is initially uncertain about two things: a

parameter , and a variable (or a vector of variables) x. In a second
stage i observes x.
Assume x X and , with X and finite. Even though i cannot
observe , he is able to assess the conditional probability of each value
x X conditional on each , P(x|).
For example, consider an urn of unknown composition. It is only known
that the urn contains between 1 and 10 balls, which may differ only in the
color, Black or White. A first ball is drawn, its color is observed and then
it is put back into the urn. Then a second ball is drawn (possibly the same
as before) and its color observed. Let denote the proportion of White
balls in the urn. The set of possible values of this parameter is finite:
=

10 
[
n=1

:=

k
for some k = 0, ..., n
n

Let Black correspond to number 0 (since black is the absence of light)

and White correspond to number 1. Then X can be identified with the
set of all ordered pairs (x1 , x2 ) with xk {0, 1}, that is, X = {0, 1}2 .
The probability of x = (x1 , x2 ) conditional on the proportion of white
balls being is P(x|) = x1 (1 )1x1 x2 (1 )1x2 . For example,
P((1, 0)|) = (1 ).
Mister i also assigns a probability P() to all the possible values of
. Note that P(x|) is well defined even if i assigns probability 0 to .
For example, for some reason i may be certain that the urn does not
1
contain more than 5 balls, and hence P( = 10
) = 0. Yet i thinks that
1
9
9
if were 10 then the probability of x = (0, 0) would be 10
10
, that is,
1
81
P((0, 0)| 10 ) = 100 .
The law of conditional probabilities says that for any two events
E, F ,
P(E F )
if P(F ) > 0, thenP(E|F ) =
.
P(F )
Note that the same condition can be written more compactly as
P(E F ) = P(E|F )P(F ).

9.4. Bayes Rule

229

Here we may assume that the space of uncertainty is = X. Each

realization x corresponds to event E = {x}, each parameter value
corresponds to event F = {}X, each pair (, x) is the singleton {(, x)}.
Then the joint probability of the pair (, x) is
P((, x)) = P(x|)P().
Note that the collection of events {F : F = {} X for some }
forms a partition of = X. This partition of induces a corresponding
partition of every event E , and in particular of events of the form E =
{x}. Then we can compute probability of any x (event {x}) starting
from the probabilities of the cells {(, x)} = ( {x}) ({} X) with
. In turn, these probabilities can be obtained from the conditional
and prior probabilities P(x|) and P(), . Thus, we obtain the
formula expressing the marginal, or total probability of x as
X
X
P(x) =
P((x, 0 )) =
P(x|0 )P(0 ).
(TotP)
0

The problem is to derive from these elements P(|x), the probability

that i would assign to each upon observing any x X. There are
two possibilities, either P(x) = 0 or P(x) > 0.
If P(x) = 0, P(|x) cannot be derived from the previous data. This
does not mean that i is unable to assess the conditional probability P(|x),
it only means that P(|x) is not determined by the other probabilistic
assessments expressed above.
If P(x) > 0, then P(|x) = P((,x))
Substituting P(x) with the
P(x) .
expression given by (TotP) we obtain
P(x|)P()
,
0
0
0 P(x| )P( )

P(|x) = P

(BFor)

(BFor) is known as Bayes Formula, that is, an equation expressing

P(|x) as a function of the conditional probabilities P(x|0 ) and the prior
probabilities P(0 ), with 0 .
We call Bayes Rule the rule that says that whenever (BFor) can be
applied, then it must be applied. Since (BFor) can be applied if and only
if P(x) > 0, we may write Bayes Rule in the following compact form:
!
X
x X, , P(|x)
P(x|0 )P(0 ) = P(x|)P().
(BRule)
0

230

9. Multistage Games with Incomplete Information

Bayes Rule is not violated if either P(x) = 0 (both sides of (BRule)

are zero), or P(x) > 0 and P(|x) is computed with (BFor). Therefore,
whenever an assessment of prior and conditional probabilities satisfies
(BRule), we say that it is consistent with Bayes rule. Note that Bayes
rule holds even if P(x) = 0 (another way to express this point is that if
the antecedent in the material implication P(x) > 0 P 0 P((,x))
P(x|0 )P(0 ) is
false, the material implication holds).1

The material implication p q is verified if either p is false, or both p and q are

true.

9.5. Perfect Bayesian Equilibrium

9.5

231

Perfect Bayesian Equilibrium

Like Nash equilibrium, Bayesian equilibrium allows for non maximizing

choices at histories that are not supposed to occur in equilibrium.
Since in multistage games with complete information this problem has
been addressed using the notion of subgame perfect equilibrium, it seems
natural to try to extend this notion to dynamic Bayesian games and define
some kind of perfect Bayesian equilibrium (PBE). [For the conceptual
reader : beware of things that seem natural!]
General ideas. We look at randomized equilibria. Beside the
candidate behavioral profile = (i )iN , we need to specify candidate
beliefs ((|i , h))iI,i i about at each h H, otherwise we cannot
compute conditional expected payoffs:
A system of beliefs is a conditional beliefs profile = (i )iI , where
i = (i (|i , h))i i ,hH hH ((0 i ))i .
i (0 , i |i , h)=prob. assigned by type i of i to (0 , i ) conditional on
h. A pair (, ) is called assessment.
Clearly, beliefs must be related to and are therefore endogenous.
Therefore a candidate equilibrium is not just a behavioral strategy profile,
but a whole assessment (, ).
Fix a candidate equilibrium assessment (, ); when can we say that
(, ) is a PBE?
First note that for each h, beliefs ((|i , h))iI,i i define a
(h, )-continuation (Bayesian) game (consider the feasible continuations
of h, the resulting -dependent payoffs and the interactive beliefs
((|i , h))iI,i i ).
Intuitively, (, ) is a PBE is it satisfies two conditions:
Bayes consistency: the initial beliefs (pi )iI,i i , the probabilities in
and the probabilities in must be related to each other via Bayes rule
whenever possible.
Continuation equilibrium (often called sequential rationality): for each
h H, induces a Bayesian equilibrium of the (h, )-continuation
Bayesian game.
However, it is not obvious how to specify Bayes consistency. Game
theorists have proposed different definitions. The reason is that, on top of
mere consistency with Bayes rule, they wanted to incorporate additional

232

9. Multistage Games with Incomplete Information

assumptions in the spirit of the Bayesian equilibrium analysis, such as that

players update in the same way (differences in beliefs are only due to
differences in information), and that beliefs satisfy independence across
opponents. Unfortunately, it was not even very clear which additional
assumptions one was trying to incorporate in the PBE concept. Appeals
to intuition and references to particular examples dominated the analysis.
Hence the mess: there is no universally accepted notion of PBE (despite
the fact that many applied papers refer to PBE as if such a universally
accepted notion existed). We consider below a minimalistic (i.e. general)
notion of PBE based only on Bayes consistency.
It is convenient to define some conditional probabilities derived from

(, ): for every (h, a) H

Y
P(a|, h; ) =
j (aj |j , h), (note: vacuous dependence on 0 )
jI

P(a|i , h; , ) =

0
0
i (00 , i
|i , h)P(a|00 , i , i
, h; ).

0
00 ,i

DEF. (, ) is Bayes consistent if for every i and for every i

(a) i (|i , h0 ) = pi (|i )
(b) for every (h, a) H, for every (0 , i ), P(a|i , h; , ) > 0 implies
i (0 , i |i , (h, a)) =

P(a|, h; )i (0 , i |i , h)
.
P(a|i , h; , )

[Recall: we assume that is finite.]

DEF. (, ) is a perfect Bayesian equilibrium (PBE) if is Bayesconsistent and, for each h H, induces a Bayesian equilibrium of the
(h, )-continuation game.
Let E[ui |i , h, ai ; , ] be the conditional expected payoff for type i of
taking action ai given h, under assessment (, ).
REMARK. By an application of the one-shot deviation principle, a
Bayes-consistent assessment (, ) is a PBE if for all i, i , h H,
suppi (|i , h) arg max E[ui |i , h, ai ; , ].
ai Ai (h)

9.6. Signaling Games and Perfect Bayesian Equilibrium

233

A precise definition of E[ui |i , h, ai ; , ] can be given as follows. For

each h H of length t and h0 = (h, at+1 , ..., at+` ) let
P(h0 |, h; ) = (at+1 |, h)

`
Y

(at+k |, h, at+1 , ..., at+k1 ).

k=2

Then
E[ui |i , h, ai ; , ]
X
0
0
=
i (00 , i
|i , h)i (ai |i
, h)E[ui |i , (h, (ai , ai )); , ]
0 ,a A (h)
00 ,i
i
i

where
E[ui |i , (h, (ai , ai )); , ]

0 , , (h, (a , a ))), if (h, (a , a )) Z
ui (00 , i
i
i i
i i
P
=
0
0
P(z|,
(h,
(a
,
a
);
)u
i i
i (0 , i , .i , z), otherwise
z:(h,(ai ,ai ))z

9.6

Signaling Games
Equilibrium

and

Perfect

Bayesian

Now assume that is known to player 1 (Ann, female), whose action2

a1 A1 is observed by player 2 (Bob, male) before he chooses a2 A2 (a1 ).
The possible parameter values are called types of player 1, the
informed player. The actions of of the informed player are also called
signals or messages
because they may reveal her private information.
S
Let A2 =
A2 (a1 ). The payoffs of players 1 (Ann) and 2 (Bob) are
a1 A1

given by the functions

u1 : A1 A2 R,
u2 : A1 A2 R.
A behavior strategy for Ann is an array of probability measures
1 = (1 (|)) (A1 ) = ((A1 )) . A behavior strategy for Bob
is an array of probability measures 2 = 2 (|a1 ))a1 A1 ai Ai (A2 (a1 )).
2

In some models the set of feasible actions of player 1 depends onS and is denoted
by A1 (). The set of potentially feasible actions of player 1 is A1 =
A1 ().

234

9. Multistage Games with Incomplete Information

Here Bob has the r

ole of agent i in the previous section on Bayes rule,
with X = A1 . Bob is initially uncertain about (, a1 ) and has a prior
probability measure (). We assume for simplicity that for every
, () > 0. This prior belief is an exogenously given element of the
model, whereas 1 and 2 are endogenous, i.e. they have to be determined
through equilibrium analysis.
In equilibrium, 1 represents the assessment, or conjecture, of Bob
about the probability of each action of Ann conditional on each possible
type (parameter value) . Thus, the probabilistic assessment of Bob is
such that
, P() = (),
, a1 A1 , P(a1 |) = 1 (a1 |), P((, a1 )) = 1 (a1 |)(),
X
a1 A1 , P(a1 ) =
1 (a1 |0 )(0 ).
0

We let (|a1 ) denote the probability that Bob would assign to upon
observing action a1 of Ann. Since Bob chooses a2 after he has observed
a1 in order to maximize the expectation of u2 (, a1 , a2 ), the system of
conditional probabilities = ((|a1 ))a1 A1 is an essential ingredient of
equilibrium analysis. In the technical language of game theory is called
system of beliefs.
P
Now, if P(a1 ) = 0 1 (a1 |0 )(0 ) > 0 then Bayes formula applies
and
P((, a1 ))
1 (a1 |)()
(|a1 ) =
=P
.
0
0
P(a1 )
0 1 (a1 | )( )
Since 1 is endogenous, also the system of beliefs is endogenous. Thus
we have to determine through equilibrium analysis the triple (1 , 2 , ).
In the technical game-theoretic language (1 , 2 , ) (a profile of behavioral
strategies plus a system of beliefs) is called assessment.
For any given assessment (1 , 2 , ) we use the following notation to
abbreviate conditional expected payoff formulas:
E[u1 |, a1 ; 2 ] : =

2 (a2 |a1 )u1 (, a1 , a2 ),

a2 A2 (a1 )

E[u2 |a1 , a2 ; ] : =

(|a1 )u2 (, a1 , a2 ).

9.6. Signaling Games and Perfect Bayesian Equilibrium

235

Thus, E[u1 |, a1 ; 2 ] is the expected payoff for Ann of choosing action a1

given that her type is and assuming that her conjecture about the
behavior of Bob is represented by 2 .3 Similarly, E[u2 |a1 , a2 ; ] is the
expected payoff for Bob of choosing action a2 given that he has observed
a1 and assuming the his conditional beliefs about are represented by .
In equilibrium, an action of player i can have positive (conditional)
probability only if it maximizes the (conditional) expected payoff of i.
Furthermore, the equilibrium assessment must be consistent with Bayes
rule. Therefore we obtain three equilibrium conditions for the three
(vectors of) unknowns (1 , 2 , ):
Definition. Assessment (1 , 2 , ) is a perfect Bayesian equilibrium
(PBE) if satisfies the following conditions:
, Supp1 (|) arg max E[u1 |, a1 ; 2 ]

(BR1 )

a1 A1

a1 A1 , Supp2 (|a1 ) arg

a1 A1 , , (|a1 )

max

a2 A2 (a1 )

E[u2 |a1 , a2 ; ]

1 (a1 |0 )(0 ) = 1 (a1 |)().

(BR2 )
(CONS)

Clearly (CONS), consistency with Bayes rule, can also be expressed as

a1
X
0

A1 , ,
1 (a1 |)()
.
0
0
0 1 (a1 | )( )

1 (a1 |0 )(0 ) > 0 (|a1 ) = P

Note that each equilibrium condition involves two out of the three
vectors of endogenous variables 1 , 2 and : (BR1 ) says that that each
mixed action 1 (|) (A1 ) ( ) is a best reply to 2 for type of
Ann, (BR2 ) says that each mixed action 2 (|a1 ) (A2 (a1 )) (a1 A1 ) is
a best reply for Bob to the conditional belief (|a1 ) (), and (CONS)
3

There is no loss of generality in representing a conjecture of player 1 as a behavioral

strategy of player 2. If player 1 had a conjecture of the form 2 (S2 ), where S2 is the
set of pure strategies of player 2, then we could derive from 2 a realization-equivalent
behavioral strategy. A similar argument holds for the conjecture of player 2 about player
1.

236

9. Multistage Games with Incomplete Information

says that 1 and (together with the exogenous prior ) are consistent
with Bayes rule.
P It should 0 be 0 emphasized that, for some action a1 , P(a1 ) =
0 1 (a1 | )( ) may be zero. For example, suppose that for each
, there is some action a1 such that E[u1 |, a1 ; 2 ] > E[u1 |, a1 ; 2 ].
Then action/message a1 must have zero probability because it is not a best
reply for any type of Ann. Yet we assume that the belief (|a1 ) is well
defined and Bob takes a best reply to this belief. This is a perfection
requirement analogous to the subgame perfection condition for games
with observable actions and complete information. A perfect Bayesian
equilibrium satisfies perfection and consistency with Bayes rule.
Furthermore, even if (|a1 ) cannot be computed with Bayes formula,
it may still be the case that the equilibrium conditions put constraints on
the possible values of (|a1 ). The following example illustrates this point.
(0, 3)

(1, 1)

{ 21 }

u
r

d
(0, 0)

(2, 0)

00
(1, 1)

1
{ 21 }

2
d
(0, 1)

The payoffs of Ann (player 1) are in bold. A2 (l) is a singleton and

therefore the action of player 2 after l is not shown. If Ann goes right (r)
then Bob can go up (u) or down (d), i.e. A2 (r) = {u, d}.
Note that action r is dominated for type 0 of Ann. Therefore
1 (r|0 ) = 0 in every PBE. Now we show that in equilibrium we also
have 1 (r|00 ) = 0. Suppose, by way of contradiction, that 1 (r|00 ) > 0 in
a PBE. Then Bayes formula applies and (00 |r) = 1. But then the best

9.6. Signaling Games and Perfect Bayesian Equilibrium

237

reply of Bob is down, i.e. 2 (d|r) = 1, and the best reply of type 00 is left,
i.e. 1 (l|00 ) = 1 1 (r|00 ) = 1, contradicting our initial assumption.
We conclude that in every PBE r is chosen with probability zero and
(|r) cannot be determined with Bayes formula. Yet the equilibrium
conditions put a constraint on (|r): in equilibrium d must be (weakly)
preferred to u (if Bob chooses u after r then type 00 chooses r and we have
just shown that this cannot happen in equilibrium). Therefore
(00 |r) 3(0 |r)
3
or (00 |r)
.
4
The set of equilibrium assessments is


3
0
00
00
(1 , 2 , ) : 1 (l| ) = 1 (l| ) = 1, 2 (d|r) = 1, ( |r) >
4


1
3
(1 , 2 , ) : 1 (l|0 ) = 1 (l|00 ) = 1, 2 (d|r) , (00 |r) =
.
2
4
These assessments are examples of pooling equilibria. A pooling
equilibrium is a PBE assessment where all types of Ann, the informed
player, choose the same pure action with probability one: there exists
a1 A1 such that for every , 1 (a1 |) = 1. In this case Bayes rule
implies that the posterior on conditional on the equilibrium action a1 is
the same as the prior: (|a1 ) = ().
The polar case is when different types choose different pure actions: a
separating equilibrium is a PBE assessment such that each type of
player 1 chooses some action a1 () with probability one (1 (a1 ()|) = 1)
and a1 (0 ) 6= a1 (00 ) for all 0 and 00 with 0 6= 00 . A separating equilibrium
may exist only if A1 has at least as many elements as . If A1 and have
the same number of elements (cardinality) then in a separating equilibrium
each action is chosen with ex ante positive probability (because () > 0
for each ) and the action of player 1 perfectly reveals her private
information (if A1 has more elements than then the actions that in
equilibrium are chosen by some type are perfectly revealing, the others
need not be revealing).
The following signaling game provides an example of separating
equilibrium (the payoffs of the informed player are in bold, call the
downward action of player 2 a and the upward action f ):

238

9. Multistage Games with Incomplete Information

(1, 0)

(0, 0)

9
}
{ 10

s
(2, 1)

(1, 1)

(0, 1)

(1, 1)

2
(1, 0)

1
1
}
{ 10

2
(2, 0)

The game can be interpreted as follows: a truck driver (player 1) enters

in a pub where an aggressive customer (player 2) has to decide whether
to start a a fight (f, upward) or acquiesce (a, downward). There are two
types of truck drivers: 90% of them are surly (s ) and like to eat sausages
(s) for breakfast; the remaining 10% are wimps (w ) and prefer a dessert
with whipped cream (w). Each type of truck driver receives a utility equal
to 2 from her favorite breakfast and a utility equal to 1 from the other
breakfast. Furthermore both types incur a loss of 1 util if they have to
fight. Player 2 prefers to fight with a wimp and to avoid the fight with a
surly driver.
This game has only one reasonable PBE and it is separating.4
Indeed, the payoffs are such that for each type of driver it is weakly
dominant to have her preferred breakfast thus iterated deletion of weakly
dominated actions yields the equilibrium 1 (s|s ) = 1 = 1 (w|w ),
2 (a|s) = 1 = 2 (f |w), (s |s) = 1 = (w |w).
The game has also two sets of pooling equilibria (meaning that one
type of player 1 chooses a weakly dominated action). In the first set of
assessments each type has sausages for breakfast and player 2 would fight if
and only he observed a whipped-cream breakfast: 1 (s|s ) = 1 = 1 (s|w ),
4
Actually, this example is a modification of a well-known game where the cost of
a fight is larger than the marginal benefit from having the preferred breakfast and all
equilibria are pooling.

9.6. Signaling Games and Perfect Bayesian Equilibrium

239

9
2 (a|s) = 1 = 2 (f |w), (s |s) = 10
, (w |w) 12 . In the second set
of assessments each type has whipped cream for breakfast and player 2
would fight if and only if he observed a sausage breakfast: 1 (w|s ) = 1 =
9
1 (w|w ), 2 (a|w) = 1 = 2 (f |s), (s |w) = 10
, (w |s) 12 .

Bibliography
[1] Aumann R.J. (1974): Subjectivity and Correlation in Randomized
Strategies, Journal of Mathematical Economics, 1, 67-96.
[2] Aumann R.J. (1987): Correlated Equilibrium as an Expression of
Bayesian Rationality, Econometrica, 55, 1-18.
[3] Battigalli P. and G. Bonanno (1999): Recent results on belief,
knowledge and the epistemic foundations of game theory, Research
in Economics, 53, 149-226.
[4] Battigalli P. and M. Siniscalchi (2003): Rationalization and
Incomplete Information, Advances in Theoretical Economics, 3 (1),
Article 3.
[5] Battigalli P., S. Cerreia-Vioglio, F. Maccheroni and
M. Marinacci (2011): Self-Confirming Equilibrium and Model
Uncertainty, IGIER working paper 428, Bocconi University.
[6] Battigalli P., M. Gilli and M.C. Molinari (1992): Learning
and Convergence to Equilibrium in Repeated Strategic Interaction,
Research in Economics (Ricerche Economiche), 46, 335-378.
[7] Battigalli P., A. Di Tillio, E. Grillo and A. Penta
(2011): Interactive Epistemology and Solution Concepts for
Games with Asymmetric Information, The B.E. Journal of
Theoretical Economics (2011), 11 : Iss. 1 (Advances), Article 6.
[http://www.bepress.com/bejte/advances/vol3/iss1/art3]
[8] Bernheim D. (1984):
Rationalizable
Econometrica, 52, 1007-1028.
240

Strategic

Behavior,

9.6. Signaling Games and Perfect Bayesian Equilibrium

241

[9] Brandenburger A. (2007): The Power of Paradox, International

Journal of Game Theory, 35, 465-492.
[10] Brandenburger, A. and E. Dekel (1987): Rationalizability and
Correlated Equilibria, Econometrica, 55, 1391-1402.
[11] Boergers T. and M. Janssen (1995): On the Dominance
Solvability of Large Cournot Games, Games and Economic
Behavior, 8, 297 321.
[12] Brandenburger A. and E. Dekel (1987): Rationalizability and
Correlated Equilibria, Econometrica, 55, 1391-1402.
[13] Dekel, E., D. Fudenberg and D. Levine (2004): Learning to
Play Bayesian Games, Games and Economic Behavior, 46, 282,303.
[14] Dekel, E., D. Fudenberg and S. Morris (2007): Interim
Correlated Rationalizability, Theoretical Economics, 2, 15-40.
[15] Dufwenberg M. and M. Stegeman (2002):
Existence
and Uniqueness of Maximal Reductions Under Iterated Strict
Dominance, Econometrica, 70, 2007-2023.
[16] Esponda I. (2011): Rationalizable Conjectural Equilibrium: A
Framework for Robust Predictions, working paper, New York
University.
[17] Eyster , E. and M. Rabin (2005):
Econometrica, 73, 1623-1672.

Cursed Equilibrium,

[18] Fudenberg D. and D. Levine (1998): The Theory of Learning in

Games. MIT Press, Cambridge MA.
[19] Fudenberg D. and J. Tirole (1991): Game Theory, Cambridge
MA: MIT Press.
[20] Gilli M. (1999): Adaptive Learning in Imperfect Monitoring
Games, Review of Economic Dynamics, 2, 472-485.
[21] Guesnerie R. (1992): An exploration on the eductive justifications
of the rational-expectations hypothesis, American Economic Review,
82, 1254-1278.

242

9. Multistage Games with Incomplete Information

[22] Harsanyi J. (1967-68): Games of incomplete information played by

Bayesian players. Parts I, II, III, Management Science, 14, 159-182,
320-334, 486-502.
[23] Harsanyi J. (1995): Games with Incomplete Information,
American Economic Review, 85, 291-303.
[24] Kuhn H.W. (1953): Extensive games and the problem of
information, in: H.W. Kuhn and A.W. Tucker (Eds.),
Contributions to the Theory of Games II, Princeton NJ: Princeton
University Press, 193-216.
[25] Milgrom P. and J. Roberts (1991): Adaptive and Sophisticated
Learning in Normal-Form Games, Games and Economic Behavior,
3,82-100.
[26] Moulin, H. (1984): Dominance-Solvability and Cournot Stability,
Mathematical Social Sciences, 7, 83-102.
[27] Nash J. (1950): Equilibrium Points in N-Person Games,
Proceedings of the National Academy of Sciences, 36, 48-49.
[28] Osborne M. and A. Rubinstein (1994): A Course in Game
Theory, Cambridge MA: MIT Press.
[29] Pearce D. (1984): Rationalizable Strategic Behavior and the
Problem of Perfection, Econometrica, 52, 1029-1050.
[30] Perea A. (2001): Rationality in Extensive Form Games, Boston,
Dordrecht: Kluwer Academic Publishers.
[31] Rubinstein A. (1989): The Electronic Mail Game: Strategic
Behavior Under Almost Common Knowledge, American Economic
Review, 79, 385-391.
[32] Rubinstein A. (1989): Comments on the Interpretation of Game
Theory, Econometrica, 59, 909-924.
[33] Sandholm, W. (2010): Population Games and Evolutionary
Dynamics, Cambridge MA: MIT Press.

9.6. Signaling Games and Perfect Bayesian Equilibrium

[34] Vickrey, W. (1961). Counterspeculation, Auctions,
Competitive Sealed Tenders, Journal of Finance, 16, 8-37.

243
and

[35] Weibull, J. (1995): Evolutionary Game Theory, Cambridge MA:

MIT Press.