Notes ST Lewis PDF

John Nachbar
Washington University
August 4, 2016
Game Theory Basics I:

Strategic Form Games1
1 Preliminary Remarks.
Game theory is a mathematical framework for analyzing conflict and cooperation.
Early work was motivated by gambling and recreational games such as chess, hence
the “game” in game theory. But it quickly became clear that the framework had
much broader application. Today, game theory is used for mathematical modeling in
a wide range of disciplines, including many of the social sciences, computer science,
and evolutionary biology. In my notes, I draw examples mainly from economics.
These particular notes are an introduction to a formalism called a strategic form
(also called a normal form). For the moment, think of a strategic form game as
representing an atemporal interaction: each player (in the language of game theory)
acts without knowing what the other players have done. An example is a single
instance of the two-player game Rock-Paper-Scissors (probably already familiar to
you, but discussed in the next section).
In companion notes, I develop an alternative formalism called an extensive form
game. Extensive form games explicitly capture temporal considerations, such as the
fact that in standard chess, players move in sequence, and each player knows the
prior moves in the game. As I discuss in the notes on extensive form games, there
is a natural way to give any extensive form game a strategic form representation.
Moreover, the benchmark prediction for an extensive form game, namely that be-
havior will conform to a Nash equilibrium, is defined in terms of the strategic form
representation.
There is a third formalism called a game in coalition form (also called charac-
teristic function form). The coalition form abstracts away from the details of what
players can do and focuses instead on what outcomes are physically possible for each
coalition (group) of players. I do not (yet) have notes on games in coalition form.
The first attempt at a general treatment of games, von Neumann and Morgen-
stern (1944), analyzed both strategic form and coalition form games, but merged the
analysis in a way that made it impossible to use game theory to study issues such
as individual incentives to abide by agreements. The work of John Nash, in Nash
(1951) and Nash (1953), persuaded game theorists to maintain a division between
the analysis of strategic and coalition form games. For discussions of some of the
1
cbna. This work is licensed under the Creative Commons Attribution-NonCommercial-
ShareAlike 4.0 License.
1
history of game theory, and Nash’s contribution in particular, see Myerson (1999)
and Nachbar and Weinstein (2015).
Nash (1951) called the study of strategic/extensive form games “non-cooperative”
game theory and the study of coalition form games “cooperative game theory”.
These names have stuck, but I prefer the terms strategic/extensive form game theory
and coalition form game theory, because the non-cooperative/cooperative language
sets up misleading expectations. In particular, it is not true that the study of games
in strategic/extensive form assumes away cooperation.
A nice, short introduction to the study of strategic and extensive form games
is Osborne (2008). A standard undergraduate text on game theory is Gibbons
(1992). A standard graduate game theory text is Fudenberg and Tirole (1991); I
also like Osborne and Rubinstein (1994). There are also good introductions to game
theory in graduate microeconomic theory texts such as Kreps (1990), Mas-Colell,
Whinston and Green (1995), and Jehle and Reny (2000). Finally, Luce and Raiffa
(1957) is a classic and still valuable text on game theory, especially for discussions
of interpretation and motivation.
I start my discussion of strategic form games with Rock-Paper-Scissors.
2 An example: Rock-Paper-Scissors.
The game Rock-Paper-Scissors (RPS) is represented in Figure 1 in what is called a
game box. There are two players, 1 and 2. Each player has three strategies in the
R P S
R 0, 0 −1, 1 1, −1
P 1, −1 0, 0 −1, 1
S −1, 1 1, −1 0, 0
Figure 1: A game box for Rock-Paper-Scissors (RPS).
game: R (rock), P (paper), and S (scissors). Player 1 is represented by the rows

while player 2 is represented by the columns. If player 1 chooses R and player 2
chooses P then this is represented as the pair, called a strategy profile, (R,P) and the
result is that player 1 gets a payoff of -1 and player 2 gets a payoff of +1, represented
as a payoff profile (−1, 1).
For interpretation, think of payoffs as encoding preferences over winning, losing,
or tying, with the understanding that S beats P (because scissors cut paper), P beats
R (because paper can wrap a rock . . . ), and R beats S (because a rock can smash
scissors). If both choose the same, then they tie. The interpretation of payoffs is
actually quite delicate and I discuss this issue at length in Section 3.3.
This game is called zero-sum because, for any strategy profile, the sum of payoffs
is zero. In any zero-sum game, there is a number V , called the value of the game,
2
with the property that player 1 can guarantee that she gets at least V no matter
what player 2 does and conversely player 2 can get −V no matter what player 1
does. I provide a proof of this theorem in Section 4.5. In this particular game,
V = 0 and both players can guarantee that they get 0 by randomizing evenly over
the three strategies.
Note that randomization is necessary to guarantee a payoff of at least 0. In
Season 4 Episode 16 of the Simpsons, Bart persistently plays Rock against Lisa,
and Lisa plays Paper, and wins. Bart here doesn’t even seem to understand the
game box, since he says, “Good old rock. Nothing beats that.” I discuss the
interpretation of randomization in Section 3.4.
3 Strategic Form Games.

3.1 The Strategic Form.
I restrict attention in the formalism to finite games: finite numbers of players and
finite numbers of strategies. Some of the examples, however, involve games with
infinite numbers of strategies.
A strategic form game is a tuple (I, (Si )i , (ui )i ).
• I is the finite set, assumed not empty, of players with typical element i. The
cardinality of I is N ; I sometimes refer to N -player games. To avoid triviality,
assume N ≥ 2 unless explicitly stated otherwise.
• Si is the set, assumed not empty, of player i’s strategies, often called pure
strategies to distinguish from the mixed strategies described below.
Q
S = i Si is the set of pure strategy profiles, with typical element s =
(s1 , . . . , sN ) ∈ S.
• ui is player i’s payoff function: ui : S → R. I discuss the interpretation of

payoffs in Section 3.3.
Games with two players and small strategy sets can be represented via a game box,
as in Figure 1 for Rock-Paper-Scissors. In that example, I = {1, 2}, S1 = S2 =
{R, P, S}, and payoffs are as given in the game box.
As already anticipated by the RPS example, we will be interested in random-
ization. For each i, let Σi be the set of probabilities over Si , also denoted ∆(Si ).
An element σi ∈ Σi is called a mixed strategy for player i. Under σi , the probability
that i plays pure strategy si is σi [si ].
A pure strategy si is equivalent to a degenerate mixed strategy, with σi [si ] = 1
and σi [ŝi ] = 0 for all ŝi 6= si . Abusing notation, I use the notation si for both the
pure strategy si and for the equivalent mixed strategy.
3
Assuming that the cardinality of Si is at least 2, the strategy σi is fully mixed iff
σi [si ] > 0 for every si ∈ Si . σi is partly mixed iff it is neither fully mixed nor pure
(degenerate).
Since the strategy set is assumed finite, I can represent a σi as a vector in either
R|Si | or R|Si |−1 , where |Si | is the number of elements in Si . For example, if S1 has
two elements, then I can represent σ1 as either (p, 1 − p), with p ∈ [0, 1] (probability
p on the first strategy, probability 1−p on the second), or just as p ∈ [0, 1] (since the
probabilities must sum to 1, one infers that the probability on the second strategy
must be 1 − p). And a similar construction works for any finite strategy set. Note
that under either representation, Σi is compact and convex.
I discuss the interpretation of mixed strategies in Section 3.4. For the moment,
however, suppose that players might actually randomize, perhaps by making use of
coin flip or toss of a die. In this case the true set of strategies is actually Σi rather
than Si . Q
Σ = i Σi is the set of mixed strategy profiles, with typical element σ =
(σ1 , . . . , σN ). A mixed strategy profile σ induces an independent probability dis-
tribution over S.
Example 1. Consider a two-player game in which S1 = {T, B} and S2 = {L, R}. If
σ1 [T ] = 1/4 and σ2 [L] = 1/3 then the induced distribution over S can be represented
in a game box as in Figure 2.
L R
T 1/12 2/12
B 3/12 6/12
Figure 2: An independent distribution over strategy profiles.
Abusing notation, let ui (σ) be the expected payoff under this independent dis-
tribution; that is,
X
ui (σ) = Eσ [ui (s)] = ui (s) × σ1 [s1 ] × · · · × σN [sN ].
s∈S
Finally, I frequently use the following notation. A strategy

Q profile s = (s1 , . . . , sN )
can also be represented as s = (si , s−i ), where s−i ∈ j6=i Sj is a profile of pure
strategies for players other than i.Q Similarly, σ = (σi , σ−i ) is alternative notation
for σ = (σ1 , . . . , σN ), where σ−i ∈ j6=i Σj is a profile of mixed strategies for players
other than i.
3.2 Correlation.
The notation so far builds in an assumption that any randomization is independent.
To see what is at issue, consider the following game.
4
A B
A 8, 10 0, 0
B 0, 0 10, 8
Figure 3: A game box for Battle of the Sexes.
Example 2. The game box for one version of Battle of the Sexes is in Figure 3. The
players would like to coordinate on either (A, A) or (B, B), but they disagree about
which of these is better.
If players were actually to engage in the Battle of the Sexes game depicted in
Figure 3, then one plausible outcome is that they would toss a coin prior to play and
then execute (A, A) if the coin lands heads and (B, B) if the coin lands tails. This
induces a correlated distribution over strategy profiles, which I represent in Figure
4.
A B
A 1/2 0
B 0 1/2
Figure 4: A correlated distribution over strategy profiles for Battle of the Sexes.
Note that under this correlated distribution, each player plays A half the time.
If players were instead to play A half the time independently, then the distribution
over strategy profiles would be as in Figure 5.
A B
A 1/4 1/4
B 1/4 1/4
Figure 5: An independent distribution strategy profiles for Battle of the Sexes.
The space of all probability distributions over S is Σc , also denoted

P ∆(S), with
generic element denoted σ c . Abusing notation (again), I let ui (σ c ) = s∈S ui (s)σ c [s].
For strategies for players other than i, the notation is σ−i c ∈ Σc−i = ∆(S−i ).
c
The notation (σi , σ−i ) denotes the distribution over S for which the probability
c [s ]. Any element of Σ induces an element of Σc : an
of s = (si , s−i ) is σi [si ]σ−i −i
independent distribution over S is a special form of correlated distribution over S.
Remark 1. A mixed strategy profile σ ∈ Σ is not an element of Σc , hence Σ 6⊆ Σc .
Rather, σ induces a distribution over S, and this induced distribution is an element
of Σc .
In the Battle of the Sexes of Figure 3, to take one example, it takes three numbers
to represent an arbitrary element σ c ∈ Σc (three rather than four because the four
5
numbers have to add up to 1). In contrast, each mixed strategy σi can be described
by a single number (for example, the probability that i plays A), hence σ can be
represented as a pair of numbers, which induces a distribution over S. Thus, in
this example, the set of independent strategy distributions is two dimensional while
the set of all strategy distributions is three dimensional. More generally, the set
of independent strategy distributions is a lower dimensional subset of the set of all
strategy distributions (namely Σc ).
3.3 Interpreting Payoffs.

In most applications in economics, payoffs are intended to encode choice by decision
makers, and this brings with it some important subtleties.
To make the issues more transparent, I introduce some additional structure.
Let Y be a set of possible prizes for the game and let γ(s) = (s, y) be the outcome
(recording both the strategy profile and the resulting prize) when the strategy profile
is s. Finally, let X = S × Y be the set of outcomes.2
As in standard decision theory, I assume that players have preferences over X
and that, for each i, these preferences have a utility representation, say vi . I include
s in the description of the outcome (s, y) because it is possible that players care
not only about the prizes but also about the way the game was played. The payoff
function ui is thus a composite function: ui (s) = vi (γ(s)).
Example 3. Recall Rock-Paper-Scissors from Section 2. In the game as usually
played,
Y = {(win, lose), (lose, win), (tie, tie)},
where, for example, (win, lose) means that player 1 gets “win” (whatever that
might mean) while player 2 gets “lose”. γ is given by, for example γ(R, P ) =
((R, P ), (lose, win)). The game box implicitly assumes that preferences over out-
comes depend only on the prize, and not on s directly. The game box assumes
further that for player 1, the utility representation for preferences assigns a utility
of 1 to (win, lose), -1 to (lose, win), and 0 to (tie, tie), with an analogous assignment
for player 2.
With this foundation, I can now make a number of remarks about the interpre-
tation of payoffs in decision theoretic terms.
1. In many economic applications, the prize is a profile of monetary values (prof-

its, for example). It is common practice in such applications to assume that
ui (s) equals the prize to i: if s gives i profits of $1 billion, then ui (s) = 1
billion. This assumption is substantive.
2
I can accommodate the possibility that s induces a non-degenerate distribution over Y , but I
will not pursue this complication.
6
(a) The assumption rules out phenomena such as altruism or envy. In con-
trast, the general strategic form formalism allows, for example, for ui (s)
to be the sum of the individual prizes (a form of perfect altruism).
(b) The assumption rules out risk aversion. If a player is risk averse, then
his payoff will be a concave function of his profit.
2. It can be difficult to “test” game theory predictions such as Nash equilibrium

in a lab. The experimenter can control γ, but the vi , and hence the ui , are
in the heads of the subjects and not directly observable. In particular, it is
not a violation of game theory to find that players are altruistic or spiteful.
This flexibility of the game theory formalism is a feature, not a bug: the goal
is to have a formalism that can be used to model essentially any strategic
interaction.
3. If vi represents choice over (pure) outcomes as in standard decision theory, then

payoff maximization is built into the payoffs; it is not a separate assumption.
But this logic does not carry over to lotteries. An assumption that players
maximize expected payoffs is substantive.
4. While it may be reasonable to assume that there is common knowledge of γ

(everyone knows the correct γ, everyone that everyone knows the correct γ,
and so on), there may not be even mutual knowledge of the vi and hence of the
ui (players may not know the utility functions of the other players). Games in
which there is not common knowledge of the ui are called games of incomplete
information. I discuss approaches to modeling such environments later in the
course.
The interpretation of payoffs in terms of decision theory is not the only one
possible. For example, in some applications of game theory to evolutionary biology,
the “strategies” might be alleles (alternative versions of a gene) and payoffs might
be the expected number of offspring.
3.4 Interpreting Randomization.

There are three main interpretations of randomization in games. These interpreta-
tions are not mutually exclusive.
1. Objective Randomization. Each player has access to a randomization device.

In this case, the true strategy set for player i is Σi . Si is just a concise way to
communicate Σi .
An interesting case of this is reported in Sontag and Drew (1998). Military
submarines occasionally implement hard turns in order to detect possible trail-
ing submarines; such maneuvers are called “clearing the baffles.” In order to
be as effective as possible, it is important that these turns be unpredictable.
7
Sontag and Drew (1998) reported that a captain of the USS Lapon used dice
in order to randomize. Curiously, it is a plot point in Clancy (1984), a classic
military techno-thriller, that a (fictional) top Russian submarine commander
was predictable when clearing the baffles of his submarine.
2. Empirical Randomization. From the perspective of an observer (say an exper-

imentalist), σi [si ] is the frequency with which si is played.
The observer could, for example, be seeing data from a cross section of play
by different players (think of a Rock-Paper-Scissors tournament with many
simultaneous matchings). σi [si ] is the fraction of players in role i of the game
who play si . Nash discussed this interpretation explicitly in his thesis, Nash
(1950b). Alternatively, the observer could be seeing data from a time series:
the same players play the same game over and over, and σi [si ] is the frequency
with which si is played over time.
3. Subjective Randomization. From the perspective of player j 6= i, σi [si ] is the

probability that j assigns to player i playing si .
Consider again the cross sectional interpretation of randomization, in which
many instances of the game are played by different players. If players are
matched randomly and anonymously to play the game then, from the per-
spective of an individual player, the opponents are drawn randomly, and hence
opposing play can be “as if” random even if the player knows that individual
opponents are playing pure strategies.
An important variant of the cross sectional interpretation of randomization
is the following idea, due to Harsanyi (1973). As discussed in Section 3.3,
players may know the γ of the game (giving prizes), but not the vi of the other
players (giving preferences over prizes), and hence may not know the ui (giving
payoffs) of the other players. Suppose that player j assigns a probability
distribution over possible ui , and for each ui , player j forecasts play of some
pure strategy si . Then, even though player j thinks that player i will play
a pure strategy, player i’s play is effectively random in the mind of player j
because the distribution over ui induces a distribution over si .
A distinct idea is that in many games it can be important not to be pre-
dictable. This was the case in Rock-Paper-Scissors, for example (Section 2).
A particular concern is that if the game is played repeatedly then one’s be-
havior should not follow some easily detected, and exploited, pattern, such as
R, P, S, R, P, S, R, . . . . A player can avoid predictability by literally random-
izing each period. An alternative, pursued in Hu (2014), is the idea that even
if a pure strategy exhibits a pattern that can be detected and exploited in
principle, it may be impossible to do so in practice if the pattern is sufficiently
complicated.
8
A subtlety with the subjective interpretation is that if there are three or more
players, then two players might have different subjective beliefs about what a
third player might do. This is assumed away in our notation, where σi does
not depend on anything having to do with the other players.
4 Nash Equilibrium.
4.1 The Best Response Correspondence.
Given a profile of opposing (mixed) strategies σ−i ∈ Σ−i , let BRi (σ−i ) be the set of
mixed strategies for player i that maximize player i’s expected payoff; formally,
BRi (σ−i ) = {σi ∈ Σi : ∀σ̂i ∈ Σi , ui (σi , σ−i ) ≥ ui (σ̂i , σ−i )} .
An element of BRi (σ−i ) is called a best response to σ−i .

Given a profile of strategies σ ∈ Σ, let BR(σ) be the set of mixed strategy profiles
σ̂ such that, for each i, σ̂i is a best response to σ−i . Formally,
BR(σ) = {σ̂ ∈ Σ : ∀i σ̂i ∈ BRi (σ−i )} .
BR is a correspondence on Σ. Since Σi is compact for each i, Σ is compact. For

each i, expected payoffs are continuous, which implies that for any σ ∈ Σ and any
i, BRi (σ−i ) is not empty. Thus, BR is a non-empty-valued correspondence on Σ.
4.2 Nash Equilibrium.

The single most important solution concept for games in strategic form is Nash
equilibrium (NE). Nash himself called it an “equilibrium point.” A NE is a (mixed)
strategy profile σ ∗ such that for each i, σi∗ is a best response to σ−i
∗ , hence σ ∗ ∈
BR(σ ∗ ): a NE is a fixed point of BR.
Definition 1. Fix a game. A strategy profile σ ∗ ∈ Σ is a Nash equilibrium (NE)

iff σ ∗ is a fixed point of BR.
In separate notes (currently labeled Game Theory Basics III), I survey some
motivating stories for NE. It is easy to be led astray by motivating stories, however,
and confuse what NE “ought” to be with what NE, as a formalism, actually is. For
the moment, therefore, I focus narrowly on the formalism.
A NE is pure if all the strategies are pure. It is fully mixed if every σi is fully
mixed. If a NE is neither pure nor fully mixed, then it is partly mixed. Strictly
speaking, a pure NE is a special case of a mixed NE; in practice, however, I may
sometimes (sloppily) write mixed NE when what I really mean is fully or partly
mixed.
The following result was first established in Nash (1950a).
9
Theorem 1. Every (finite) strategic form game has at least one NE.
Proof. For each i, Σi is compact and convex and hence Σ is compact and convex.
BR is a non-empty valued correspondence on Σ and it is easy to show that it has a
closed, convex graph. By the Kakutani fixed point theorem, BR has a fixed point
σ ∈ BR(σ).
Remark 2. Non-finite games may not have NE. A trivial example is the game in
which you name a number α ∈ [0, 1) and I pay you a billion dollars with probability
α (and nothing with probability 1 − α). Assuming that you prefer more (expected)
money to less, you don’t have a best response and hence there is no NE.
The most straightforward extension of Theorem 1 is to games in which each
player’s strategy set is a compact subset of some measurable topological space and
the payoff functions are continuous. In this case, the set of mixed strategies is
weak-* compact. One can take finite approximations to the strategy sets, apply
Theorem 1 to get a NE in each finite approximation, appeal to compactness to
get a convergent subsequence of these mixed strategy profiles, and then argue, via
continuity of payoffs, that the limit must be an equilibrium in the original game.
This approach will not work if the utility function is not continuous and, unfor-
tunately, games with discontinuous utility functions are fairly common in economics.
A well known example is the Bertrand duopoly game, discussed in Example 11 in
Section 4.4. For a survey of the literature on existence of NE in general and also
on existence of NE with important properties (e.g., pure strategy NE or NE with
monotonicity properties), see Reny (2008).
Remark 3. A related point is the following. The Brouwer fixed point theorem states
that if D ⊆ RN is compact and convex and f : D → D is continuous then f has a
fixed point: there is an x ∈ D such that f (x) = x. Brouwer is the underlying basis
for the Kakutani fixed point theorem, which was used in the proof Theorem 1, and
one can also use Brouwer more directly to prove NE existence (as, indeed, Nash
himself subsequently did for finite games in Nash (1951)). Thus, Brouwer implies
NE.
The converse is also true: given a compact convex set D ⊆ RN and a continuous
function f : D × D, one can construct a game in which the strategy set for player 1
is D and, if a NE exists, it must be that s1 = f (s1 ). One such construction can be
found here.
4.3 NE and randomization.

The following fact says that a best response gives positive probability to a pure
strategy only if that pure strategy is, in its own right, also a best response.
Theorem 2. For any finite game and for any σ−i , if σi ∈ BRi (σ−i ) and σi (si ) > 0
then si ∈ BRi (σ−i ).
10
Proof. Since the set of pure strategies is assumed finite, one of these strategies,
call it s∗i , has highest expected payoff (against σ−i ) among all pure strategies. Let
the expected payoff of s∗i be c∗i . For any mixed strategy σi , the expected payoff is
the convex sum of the expected payoffs to i’s pure strategies. Therefore, c∗i is the
highest possible payoff to any mixed strategy for i, which implies that, in particular,
s∗i is a best response, as is any other pure strategy that has an expected payoff of
c∗i .
I now argue by contraposition. Suppose si ∈ / BRi (σ−i ), hence ui (si , σ−i ) < c∗i .
If σi [si ] > 0, then ui (σi , σ−i ) < c∗i , hence σi ∈
/ BRi (σ−i ), as was to be shown.
Theorem 2 implies that a NE mixed σi will give positive probability to two

different pure strategies only if those pure strategies each earn the same expected
payoff as the mixed strategy. That is, in a NE, a player is indifferent between all of
the pure strategies that he plays with positive probability. This provides a way to
compute mixed strategy NE, at least in principle.
Example 4 (Finding mixed NE in 2 x 2 games). Consider a general 2 x 2 game given
by the game box in Figure 6. Let q = σ2 (L). Then player 1 is indifferent between
L R
T a1 , a2 b1 , b2
B c1 , c2 d1 , d2
Figure 6: A general 2 x 2 game.
T and B iff
qa1 + (1 − q)b1 = qc1 + (1 − q)d1 ,
or
d1 − b1
q= ,
(a1 − c1 ) + (d1 − b1 )
provided, of course, that (a1 − c1 ) + (d1 − b1 ) 6= 0 and that this fraction is in [0, 1].
Similarly, if p = σ1 (T ), then player 2 is indifferent between L and R if
pa2 + (1 − p)c2 = pb2 + (1 − p)d2 ,
or
d2 − c2
p= ,
(a2 − b2 ) + (d2 − c2 )
again, provided that (a1 − c1 ) + (d1 − b1 ) 6= 0 and that this fraction is in [0, 1].
Note that the probabilities for player 1 are found by making the other player,
player 2, indifferent. A common error is to do the calculations correctly, but then
to write the NE incorrectly, by flipping the player roles.
11
Example 5 (NE in Battle of the Sexes). Battle of the Sexes (Example 2) has two
pure strategy NE: (A, B) and (B, A). There is also a mixed strategy NE with
8−0 4
σ1 (A) = =
(10 − 0) + (8 − 0) 9
and
10 − 0 5
σ2 (A) = =
(10 − 0) + (8 − 0) 9
which I can write as ((4/9, 5/9), (5/9, 4/9)); where (4/9, 5/9), for example, means
the mixed strategy in which player 1 plays A with probability 4/9 and B with
probability 5/9.
Example 6 (NE in Rock-Paper-Scissors). In Rock-Scissor-Paper (Section 2), the NE
is unique and in it each player randomizes 1/3 each across all three strategies. That
is, the unique mixed strategy NE profile is
((1/3, 1/3, 1/3), (1/3, 1/3, 1/3)).
It is easy to verify that if either player randomizes 1/3 across all strategies, then the
other payer is indifferent across all three strategies.
Example 7. Although every pure strategy that gets positive probability in a NE is
a best response, the converse is not true: in a NE, there may be best responses that
are not given positive probability. A trivial example where this occurs is a game
where the payoffs are constant, independent of the strategy profile. Then players
are always indifferent and every strategy profile is a NE.
4.4 More NE examples.

Example 8. Figure 7 provides the game box for a game called Matching Pennies,
which is simpler than Rock-Paper-Scissors but in the same spirit. This game has
H T
H 1, −1 −1, 1
T −1, 1 1, −1
Figure 7: Matching Pennies.
one equilibrium and in this equilibrium both players randomize 50:50 between the
two pure strategies.
Example 9. Figure 8 provides the game box for one of the best known games in
Game Theory, the Prisoner’s Dilemma (PD) (the PD is actually a set of similar
games; this is one example). PD is a stylized depiction of friction between joint
incentives (the sum of payoffs is maximized by (C, C)) and individual incentives
(each player has incentive to play D).
12
C D
C 4, 4 0, 6
D 6, 0 1, 1
Figure 8: A Prisoner’s Dilemma.
It is easy to verify that D is a best response no matter what the opponent does.
The unique NE is thus (D, D).
A PD-like game that shows up fairly frequently in popular culture goes some-
thing like this. A police officer tells a suspect, “We have your partner, and he’s
already confessed. It will go easier for you if you confess as well.” Although there
is usually not enough detail to be sure, the implication is that it is (individually)
payoff maximizing not to confess if the partner also does not confess; it is only if
the partner confesses that it is payoff maximizing to confess. If that is the case, the
game is not a PD but something like the game in Figure 9, where I have written
payoffs in a way that makes them easier to interpret as prison sentences (which are
bad).
Silent Confess
Silent 0, 0 −4, −2
Confess −2, −4 −3, −3
Figure 9: A non-PD game.
This game has three NE: (Silent, Silent), (Confess, Confess), and a mixed NE in
which the suspect stays silent with probability 1/3. In contrast, in the (true) PD,
each player has incentive to play D regardless of what he thinks the other player
will do.
Example 10 (Cournot Duopoly). The earliest application of game theoretic reason-
ing that I know of to an economic problem is the quantity-setting duopoly model of
Augustin Cournot, in Cournot (1838).
There are two players (firms). For each firm Si = R+ . Note that this is not
a finite game; one can work with a finite version of the game but analysis of the
continuum game is easier. For the pure strategy profile (q1 , q2 ) (qi for “quantity”),
the payoff to player i is,
(1 − q1 − q2 )qi − k,
where k > 0 is a small number (in particular, k < 1/9), provided qi > 0 and
q1 + q2 ≤ 1. If qi = 0 then the payoff to firm i is 0. If qi > 0 and q1 + q2 > 1 then
the payoff to firm i is −k.
The idea here is that both firms produce a quantity qi of a good, best thought
of as perishable, and then the market clears somehow, so that each receives a price
13
1 − q1 − q2 , provided the firms don’t flood the market (q1 + q2 > 1); if they flood the
market, then the price is zero. I have assumed that the firm’s payoff is its profit.
Ignore k for a moment. It is easy to verify that for any probability distribution
over q2 , provided E[q2 ] < 1, the best response for firm 1 is the pure strategy,
1 − E[q2 ]
.
2
If E[q2 ] ≥ 1, then the market is flooded, in which case any q1 generates the
same profit, namely 0. Because of this, if k = 0, then there are an infinite number
of “nuisance” NE in which both firms flood the market by producing at least 1 in
expectation. This is the technical reason for including k > 0 (which seems realistic
in any event): for k > 0, if E[q2 ] ≥ 1, then the best response is to shut down
(q1 = 0). Adding a small marginal cost of production would have a similar effect.
As an aside, this illustrates an issue that sometimes arises in economic modeling:
sometimes one can simplify the analysis, at least in some respects, by making the
model more complicated.3
With k ∈ (0, 1/9), the NE is unique and in it both firms produce 1/3. One
can find this by noting that, in pure strategy equilibrium, q1 = (1 − q2 )/2 and
q2 = (1 − q1 )/2; this system of equations has a unique symmetric solution with
q1 = q2 = q = (1 − q)/2, which implies q = 1/3.
Total output in NE is thus 2/3 (and price is 1/3). The NE of the Cournot quan-
tity game is intermediate between monopoly (where quantity is 1/2) and efficiency
(in the sense of maximizing total surplus; since there is zero marginal cost, efficiency
calls for total output of 1).
Example 11 (Bertrand Duopoly). Writing in the 1880s, about 50 years after Cournot,
Joseph Bertrand objected to Cournot’s quantity-setting model of duopoly on the
grounds that quantity competition is unrealistic. In Bertrand’s model, firms set
prices rather than quantities.
Assume (to make things simple) that there are no production costs, either
marginal or fixed. I allow pi , the price charged by firm i, to be any number in
R+ , so, once again, this is not a finite game.
• pi ≥ 1 then firm i gets a payoff of 0, regardless of what the other firm does.
Assume henceforth that pi ≤ 1 for both firms.
• If p1 ∈ (0, p2 ) then firm 1 receives a payoff of (1 − p1 )p1 . In words, firm 1

gets the entire demand at p1 , which is 1 − p1 . Revenue, and hence profit,
is (1 − p1 )p1 . Implicit here is the idea that the good of the two firms are
homogeneous and hence perfect substitutes: consumers buy which ever good
is cheaper; they are not interested in the name of the seller.
3
If E[q2 ] is not defined, then player 1’s best response does not exist; by definition, this cannot
happen in NE, so I ignore this case.
14
• If p1 = p2 = p, both firms receive a payoff of (1 − p)p/2. In words, the firms
split the market at price p.
• If p1 > p2 , then firm 1 gets a payoff of 0.
Payoffs for firm 2 are defined symmetrically. It is not hard to see that firm 1’s best
response is as follows.
• p2 > 1/2 then firm 1’s best response is p1 = 1/2.
• If p2 ∈ (0, 1/2] then firm 1’s best response set is empty. Intuitively, firm 1
wants to undercut p2 by some small amount, say ε > 0, but ε > 0 can be
made arbitrarily small. One can address this issue by requiring that prices lie
on a finite grid (be expressed in pennies), but this introduces other issues.
• If p2 = 0, then any p1 ∈ R+ is a best response.
It follows that the unique pure NE (and in fact the unique NE period) is the price
profile (0, 0).
The Bertrand model predicts the competitive outcome with only two firms. At
first this may seem paradoxical, but remember that I am assuming that the goods are
homogeneous. The bleak prediction of the Bertrand game (bleak from the viewpoint
of the firms; consumers do well and the outcome is efficient) suggests that firms have
strong incentive to differentiate their products, and also strong incentive to collude
(which may be possible if the game is repeated, as I discuss later in the course).
Example 12. Francis Edgworth, writing in the 1890s, about ten years after Bertrand,
argued that firms often face capacity constraints, and these constraints can affect
equilibrium analysis. Continuing with the Bertrand duopoly of Example 11, suppose
that firms can produce up to 3/4 at zero cost but then face infinite cost to produce
any additional output. This would model a situation in which, for example, each
firm had stockpiled a perishable good; this period, a firm can sell up to the amount
stockpiled, but no more, at zero opportunity cost (zero since any amount not sold
melts away in the night).
Suppose that sales of firm 1 (taking into consideration the capacity constraint
and ignoring the trivial case with p1 > 1) are given by,


0 if p1 > p2 and p1 > 1/4,

1/4 − p1 if p1 > p2 and p1 ≤ 1/4,



q1 (p1 , p2 ) = (1 − p1 )/2 if p1 = p2 ≤ 1,

1 − p1 if p1 < p2 and p1 > 1/4,





3/4 if p1 < p2 and p1 ≤ 1/4, .

The sales function for firm 2 is analogous.
15
For interpretation, suppose that there are a continuum of consumers, each want-
ing to buy at most one unit of the good. Let v` be the valuation of consumer `.
Thus a consumer with valuation v` will buy at any price less than v` . Suppose that
the valuations are uniformly distributed on [0, 1]. Then the market demand facing
a monopolist would be 1 − p, as specified.
With this interpretation, consider the case p1 > p2 and p2 < 1/4. Then demand
for firm 2 is 1 − p2 > 3/4. But because of firm 2’s capacity constraint, it can only
sell 3/4, leaving (1 − p2 ) − 3/4 > 0 potentially for firm 1, provided firm 1’s price
is not too high. Suppose that the customers who succeed at buying from the low
priced firm are the customers with the highest valuations (they are the customers
willing to wait on line, for example). Then the left over demand facing firm 1, called
residual demand, is 1/4 − p1 . It is important to understand that this specification of
residual demand is a substantive assumption about how demand to the low priced
firm gets rationed.
In contrast to the Bertrand model, it is no longer an equilibrium for both firms to
charge p1 = p2 = 0. In particular, if p2 = 0, then firm 1 is better off setting p1 = 1/8,
which is the monopoly price for residual demand. In fact, there is no pure strategy
equilibrium here. There is, however, a symmetric mixed strategy equilibrium, which
can be found as follows.
Note first that if the maximum profit from residual demand is 1/64 (= 1/8×1/8).
A firm can also get a profit of 1/64 by being the low priced firm at a price of 1/48
(and selling 3/4). There is thus no sense in charging less than 1/48. The best
response for firm 1 to pure strategies of firm 2 is then,

1/2 if p2 > 1/2,

1
BR (p2 ) = ∅ if p2 ∈ (1/48, 1/2],

1/8 if p2 ≤ 1/48.

The best response for firm 2 to pure strategies of player 1 is analogous.

Intuitively, the mixed strategy equilibrium will involve a continuous distribution,
and not put positive probability on any particular price, because of the price under-
cutting argument that is central to the Bertrand game. So, I look for a symmetric
equilibrium in which firms randomize over an interval of prices. The obvious interval
to look at is [1/48, 1/8].
Let F2 (x) be the probability that player 2 charges a price less than or equal
to x. F2 (1/48) = 0: if p1 = 1/48, then firm 1 is the low priced firm for certain.
F2 (1/8) = 1: if p1 = 1/8, then firm 1 is the high priced firm for certain.
For it to be an equilibrium for firm 1 to randomize over the interval [1/48, 1/8],
firm 1 has to get the same profit from any price in this interval.4 The profit at
either p1 = 1/48 (where the firm gets capacity constrained sales of 3/4), or p1 = 1/8
4
In a continuum game like this, this requirement can be relaxed somewhat, for measure theoretic
reasons, but this is a technicality as far as this example is concerned.
16
(where the firm sells 1/4-1/8=1/8 to residual demand) is 1/64. Thus for any other
p1 ∈ (1/48, 1/8) it must be that,
F2 (p1 )p1 (1/4 − p1 ) + (1 − F2 (p1 ))p1 3/4 = 1/64,
where the first term is profit when p2 < p1 (which occurs with probability F2 (p1 ))
and the second term is profit when p2 > p1 (which occurs with probability 1 −
F2 (p1 )); because the distribution is continuous, the probability that p2 = p1 is zero
and so this case can be ignored.
Solving for F2 yields
1 48x − 1
F2 (x) = .
32 x(1 + 2x)
The symmetric equilibrium then has F1 = F2 , where F2 is given by the above
expression.
Edgeworth argued that since there is no pure strategy equilibrium, prices would
just bounce around between 1/48 and 1/8. Nash equilibrium makes a more specific
prediction: the distribution of prices has to exhibit a specific pattern generated by
the mixed strategy equilibrium.
There is an intuition, first explored formally in Kreps and Scheinkman (1983),
that the qualitative predictions of Cournot’s quantity model can be recovered in a
more realistic, multi-period setting in which firms make output choices in period 1
and then sell that output in period 2. The Edgeworth example that I just worked
through is an example of what the period 2 analysis can look like.
4.5 The MinMax Theorem.

Recall from Section 2 that a two-player game is zero-sum iff u2 = −u1 . A game is
constant-sum iff there is a c ∈ R such that u1 + u2 = c.
Theorem 4 below, called the MinMax Theorem (or, more frequently, the Min-
iMax Theorem), was one of the first important general results in game theory. It
was established for zero-sum games in von Neumann (1928).5 Although, it is rare to
come across zero-sum, or constant-sum, games in economic applications, the Min-
Max Theorem is useful because it is sometimes possible to provide a quick proof
of an otherwise difficult result by constructing an auxiliary constant-sum game and
then applying the MinMax Theorem to this new game; an example of this is given
by the proof of Theorem 6 in Section 4.8.
The MinMax Theorem is an almost immediate corollary of the following result.
Theorem 3. In a finite constant-sum game, if σ ∗ = (σ1∗ , σ2∗ ) is a NE then,
u1 (σ ∗ ) = max min u1 (σ) = min max u1 (σ).

σ1 σ2 σ1 σ2
5
Yes, von Neumann (1928) is in German; no, I haven’t read it in the original.
17
Proof. Suppose that u1 + u2 = c. Let V = u1 (σ ∗ ). Since σ ∗ is a NE, and since, for
player 2, maximizing u2 = c − u1 is equivalent to minimizing u1 ,
V = max u1 (σ1 , σ2∗ ) = min u1 (σ1∗ , σ2 ). (1)

σ1 σ2
Then,
min max u1 (σ) ≤ max u1 (σ1 , σ2∗ ) = min u1 (σ1∗ , σ2 ) ≤ max min u1 (σ) (2)
σ2 σ1 σ1 σ2 σ1 σ2
On the other hand, for any σ̃2 ,
u1 (σ1∗ , σ̃2 ) ≤ max u1 (σ1 , σ̃2 ),

σ1
hence,
min u1 (σ1∗ , σ2 ) ≤ max u1 (σ1 , σ̃2 ).
σ2 σ1
Since the latter holds for every σ̃2 ,
min u1 (σ1∗ , σ2 ) ≤ min max u1 (σ1 , σ2 ).

σ2 σ2 σ1
Combining with expressions (1) and (2),
V = min u1 (σ1∗ , σ2 ) = min max u1 (σ).

σ2 σ2 σ1
By an analogous argument,
V = max u1 (σ1 , σ2∗ ) = max min u1 (σ).

σ1 σ1 σ2
And the result follows.
Theorem 4 (MinMax Theorem). In any finite constant-sum game, there is a num-

ber V such that
V = max min u1 (σ) = min max u1 (σ)
σ1 σ2 σ1 σ2
Proof. By Theorem 1, there is NE, σ∗. Set V = u1 (σ ∗ ) and apply Theorem 3.
Remark 4. The MinMax Theorem can be proved using a separation argument, rather
than a fixed point argument (the proof above impicitly uses a fixed point argument,
since it invokes Theorem 1, which uses a fixed point argument.)
By way of interpretation, one can think of maxσ1 minσ2 u1 (σ) as the payoff that
player 1 can guarantee for herself, no matter what player 2 does. Theorem 3 implies
that, in a constant-sum game, player 1 can guarantee an expected payoff equal to her
NE payoff. This close connection between NE and what players can guarantee for
themselves does not generalize to non-constant-sum games, as the following example
illustrates.
18
L R
T 1, 0 0, 1
B 0, 3 1, 0
Figure 10: The Aumann-Maschler Game.
Example 13. Consider the game in Figure 10, which originally appeared in Aumann
and Maschler (1972). This game is not constant sum. In this game, there is a
unique NE, with σ1∗ = (3/4, 1/4) and σ2∗ = (1/2, 1/2) and the equilibrium payoffs
are u1 (σ ∗ ) = 1/2 and u2 (σ ∗ ) = 3/4.
But in this game, σ1∗ does not guarantee player 1 a payoff of 1/2. In fact, if player
1 plays σ1∗ then player 2 can hold player 1’s payoff to 1/4 by playing σ2 = (0, 1) (i.e.,
playing R for certain); it is not payoff maximizing for player 2 to do this, but it is
feasible. Player 1 can guarantee a payoff of 1/2, but doing so requires randomizing
σ1 = (1/2, 1/2), which is not an strategy that appears in any Nash equilibrium of
this game.
4.6 Never a Best Response and Rationalizability.

Definition 2. Fix a game. A pure strategy si is never a best response (NBR)
iff there does not exist a profile of opposing mixed strategies σ−i such that si ∈
BRi (σ−i ).
If si is NBR then it cannot be part of any NE. Therefore, it can be, in ef-
fect, deleted from the game when searching for NE. This motivates the following
procedure.
Let Si1 ⊆ Si denote the strategy set for player i after deleting all strategies that
are never a best response. Si1 is non-empty since, as noted in Section 4.1, the best
response correspondence is not empty-valued. The Si1 form a new game, and for
that game we can ask whether any strategies are NBR. Deleting those strategies,
we get Si2 ⊆ Si1 ⊆ Si , which again is not empty. And so on.
Since the game is finite, this process terminates after a finite number of rounds
in the sense that there is a t such that for all i, Sit+1 = Sit . Let SiR denote this
terminal Sit and let S R be the product of the SiR . The strategies SiR are called the
rationalizable strategies for player i; S R is the set of rationalizable (pure) strategy
profiles. Informally, these are the strategies that make sense after introspective
reasoning of the form, “si would maximize my expected payoffs if my opponents
were to play σ−i , and σ−i would maximize my opponents’ payoffs (for each opponent,
independently) if they thought . . . .” Rationalizability was first introduced into game
theory in Bernheim (1984) and Pearce (1984).
Theorem 5. Fix a game. If σ is a NE and σi [si ] > 0 then si ∈ SiR .
19
Proof. Almost immediate by induction and Theorem 2.
Remark 5. The construction of S R uses maximal deletion of never a best response

strategies at each round of deletion. In fact, this does not matter: it is not hard
to show that if, whenever possible, one deletes at least one strategy for at least
one player at each round, then eventually the deletion process terminates, and the
terminal S t is S R .
Example 14. In the Prisoner’s Dilemma (Example 9, Figure 8), C is NBR. D is the
unique rationalizable strategy.
Example 15 (Rationalizability in the Cournot Duopoly). Recall the Cournot Duopoly
(Example 10).
With k > 0, no pure strategy greater than 1/2 is a best response. Therefore, in
the first round of deletion, delete all strategies in (1/2, ∞); we can actually delete
more, as discussed below, but we can certainly delete this much. In the second round
of deletion, one can delete all strategies in [0, 1/4). And so on. This generates a
nested sequence of compact intervals. There is no termination in finite time, as for
a finite game, but one can show that the intersection of all these intervals is the
point {1/3}: SiR = {1/3} for both firms. And, indeed, the NE is (as we already
saw) (1/3, 1/3).
In fact, because of k > 0, firm 1 strictly prefers to shut down if E[q2 ] is too
large; more√precisely, at the first round of deletion, firm 1√prefers to shut down if
q2 > 1 − 2 k. So we can also eliminate all q1 in (0, (1 − 2 k)/2), even at the first
round. But note that this complication does not undermine the argument above; it
only states that we could have eliminated more strategies at each round.
A natural, but wrong, intuition is that any strategy in SiR gets positive proba-
bility in at least one NE.
Example 16. Consider the game in Figure 11. SiR = {A, B, C} for either player.
A B C
A 4, 4 1, 1 1, 1
B 1, 1 2, −2 −2, 2
C 1, 1 −2, 2 2, −2
Figure 11: Rationalizability and NE.
But the unique NE is (A, A).
4.7 Correlated Rationalizability.

Somewhat abusing notation, let BRi (σ−i c ) denote the set of strategies for player i
c ∈ Σc = ∆(S ).
that are best responses to the distribution σ−i −i −i
20
Definition 3. Fix a game. A pure strategy si is correlated never a best response
c such that
(CNBR) iff there does not exist a distribution over opposing strategies σ−i
c
si ∈ BRi (σ−i ).
Again, by Theorem 2, if si is never a best response, then so is any mixed strategy

that gives si positive probability.
The set of CNBR strategies is a subset of the set of NBR strategies. In particular,
any CNBR strategy is NBR (since a strategy that is not a best response to any
distribution is, in particular, not a best response to any independent distribution).
If N = 2 the CNBR and NBR sets are equal (trivially, since in this case there is
only one opponent). If N ≥ 3, the CNBR set can be a proper subset of the NBR
set. In particular, one can construct examples in which a strategy is NBR but
not CNBR (see, for instance Fudenberg and Tirole (1991)). The basic intuition is
that it is easier for a strategy to be a best response to something the larger the set
of distributions over opposing strategies, and the set of correlated distributions is
larger than the set of independent distributions.
If one iteratively deletes CNBR strategies, one gets, for each player i, the set SiCR
of correlated rationalizable strategies. Since, at each step of the iterative deletion,
the set of CNBR is a subset of the set of NBR strategies, SiR ⊆ SiCR for all i, with
equality if N = 2. Let S CR denote the set of correlated rationalizable profiles. The
set of distributions over strategies generated by Nash equilibria is a subset of the
set of distributions over S R , which in turn is a subset of the set of distributions over
S CR .
4.8 Strict Dominance.

Definition 4. Fix a game.
• Mixed strategy σi strictly dominates mixed strategy σ̂i iff for any profile of
opposing pure strategies s−i ,
ui (σi , s−i ) > ui (σ̂i , s−i ).
• Mixed strategy σ̂i is strictly dominated iff there exists a mixed strategy that
strictly dominates it.
If a mixed strategy σi strictly dominates every other strategy, pure or mixed,

of player i then it is called strictly dominant. It is easy to show that a strictly
dominant strategy must, in fact, be pure.
Definition 5. Fix a game. Pure strategy si is strictly dominant iff it strictly dom-
inates every other strategy.
Example 17. Recall the Prisoner’s Dilemma of Example 9 and Figure 8. Then, for
either player, C is strictly dominated by D; D is strictly dominant.
21
Theorem 6. Fix a game. For each player i, a strategy is strictly dominated iff it
is CNBR.
Proof. ⇒. If σ̂i strictly dominates σi then σ̂i has strictly higher payoff than σi
against any distribution over opposing strategies (since expected payoffs a are a
convex combination of payoffs against pure strategy profiles).
⇐. Suppose that σ̂i is CNBR. Construct a zero-sum game in which the two
players are i and −i, the strategy set for i is Si (unless σ̂i happens to be pure, in
which case omit that), and the strategy set for −i is S−i . In this new game, a mixed
strategy for player −i is a probability distribution over S−i , which (if there are two
or more opponents) is a correlated strategy σ−i c in the original game. In the new
game, the payoff to player i is,
ũi (s) = ui (si , s−i ) − ui (σ̂i , s−i ).
and the payoff to −i is ũ−i = −ũi .

Let
c
V = min c
max ũi (σi , σ−i ).
σ−i σi
c there is a σ for which ũ (σ , σ c ) > 0.

Since σi is CNBR, it follows that for every σ−i i i i −i
Therefore V > 0. By Theorem 4,
c
V = max min
c
ũi (σi , σ−i ).
σi σ−i
This implies that there exists a σi∗ (in fact, a NE strategy for player i in the new
game) such that for all s−i ∈ S−i , ũi (σi∗ , s−i ), hence
ui (σi∗ , s−i ) > ui (σ̂i , s−i ),
so that σi∗ strictly dominates σ̂i .
In the following examples, N = 2, so the CNBR and NBR strategy sets are
equal.
Example 18. Consider the game in Figure 12 (I’ve written payoffs only for player
1). T is the best response if the probability of L is more than 1/2. M is the best
L R
T 10 0
M 0 10
B 4 4
Figure 12: A Dominance Example.
response if the probability of L is less than 1/2. If the probability of L is exactly
22
1/2 the the set of best responses is the set of mixtures over T and M . Hence B is
NBR.
B is strictly dominated for player 1 by the mixed strategy (1/2, 1/2, 0) (which
gets a payoff of 5 no matter what player 2 does, whereas B always gets 4). No
strategy is strictly dominant.
Note that B is not strictly dominated by any pure strategy. So it is important
in this example that I allow for dominance by a mixed strategy.
Example 19. Consider the game in Figure 13. T is the best response if the probability
L R
T 10 0
M 0 10
B 6 6
Figure 13: Another Dominance Example.
of L is more than 3/5 and M is the best response if the probability of L is less than
2/5. If the probability of L is in (2/5, 3/5) then B is the best response. If the
probability of L is 3/5 then the set of best responses is the set of mixtures over T
and B. If the probability of L is 2/5 then the set of best responses is the set of
mixtures over M and B.
Note that B is not a best response to any pure strategy for player 2. So it is
important in this example that I allow for mixed strategies by player 2.
B is not strictly dominated for player 1. No strategy is strictly dominant.
It follows from Theorem 6 that one can find the correlated rationalizable strategy
sets, SiCR , by iteratively deleting strictly dominated strategies. Once again, this may
help in narrowing down the search for Nash equilibria.
4.9 How many NE are there?

Fix players and strategy sets and consider all possible assignments of payoffs. With
N players and |S| pure strategy profiles, payoffs can be represented as a point in
RN |S| . Wilson (1971) proved that there is a precise sense in which the set of Nash
equilibria is finite and odd (which implies existence, since 0 is not an odd number)
for most payoffs in RN |S| . As we have already seen, Rock-Paper-Scissors has one
equilibrium, while Battle of the Sexes has three. A practical implication of the finite
and odd result, and the main reason why I am emphasizing this topic in the first
place, is that if you are trying to find all the equilibria of the game, and so far you
have found four, then you are probably missing at least one.
Although the set of NE in finite games is typically finite and odd, there are
exceptions.
23
A B
A 1, 1 0, 0
B 0, 0 0, 0
Figure 14: A game with two Nash equilibria.
Example 20 (A game with two NE). A somewhat pathological example is the game
in Figure 14. This game has exactly two Nash equilibria, (A, A) and (B, B). In
particular, there are no mixed strategy NE: if either player puts any weight on A
than the other player wants to play A for sure.
The previous example seems somewhat artificial, and in fact it is extremely un-
usual in applications to encounter a game where the set of NE is finite and even. But
it is quite common to encounter strategic form games that have an infinite number
of NE; such games are common when the strategic form represents an extensive
form game with a non-trivial temporal structure.
Example 21 (An Entry Deterrence Game). Consider the game in Figure 15. For
A F
In 15, 15 −1, −1
O 0, 35 0, 35
Figure 15: An Entry Deterrence Game.
motivation, player 1 is considering entering a new market (staying in or going out).

Player 2 is an incumbent who, if there is entry, can either accommodate (reach some
sort of status quo intermediate between competition and monopoly) or fight (start
a price war).
There are two pure strategy NE, (In, A) and (O, F ). In addition, there are a
continuum of NE of the form, player 1 plays O and player 2 accommodates, should
entry nevertheless occur, with probability q ≤ 1/16. More formally, every profile of
the form (O, (q, 1 − q)) is a NE for any q ∈ [0, 1/16]. Informally, the statement is
that it is a NE for the entrant to stay out provided that the incumbent threatens
to fight with high enough probability (accommodate with low enough probability).
It is standard in game theory to argue that the only reasonable NE of this game
is the pure strategy NE (In, A). One justification for this claim is that if player 1
puts any weight at all on In then player 2’s best response is to play A for sure, in
which case player 1’s best response is likewise to play In for sure. The formal version
of this statement is to say that only (In, A) is an admissible NE, because the other
NE put positive weight on a weakly dominated strategy, namely F ; I discuss weak
dominance later in the course. For the moment, what I want to stress is that every
one of this infinity of NE is a NE, whether we find it reasonable or not.
24
Assuming that the number of equilibria is in fact finite and odd, how many
equilibria are typical? In 2 × 2 games, the answer is 1 or 3. What about in general?
A lower bound on how large the set of NE can possibly be in finite games is provided
by L × L games (two players, each with L strategies) in which, in the game box
representation, payoffs are (1, 1) along the diagonal and (0, 0) elsewhere. In such
games, there are L pure strategy Nash equilibria, corresponding to play along the
diagonal, and an additional 2L − (L + 1) fully or partly mixed NE. The total number
of NE is thus 2L − 1. This example is robust; payoffs can be perturbed slightly and
there will still be 2L − 1 NE.
This is extremely bad news. First, it means that the maximum number of NE
is growing exponentially in the size of the game. This establishes that the general
problem of calculating all of the equilibria is computationally intractable. Second,
it suggests that the problem of finding even one NE may, in general, be computa-
tionally intractable; for work on this, see Daskalakis, Goldberg and Papdimitriou
(2008). Note that the issue is not whether algorithms exist for finding one equi-
librium, or even all equilibria. For finite games, there exist many such algorithms.
The problem is that the time taken by these algorithms to reach a solution can grow
explosively in the size of the game. For results on the number of equilibria in finite
games, see McLennan (2005).
4.10 The structure of NE.

As in Section 4.9, if players and strategy sets are fixed, then payoff functions can
be represented as a vector in RN |S| , giving payoffs for each player and each strategy
profile. Let
N : RN |S| → P(Σ)
be the NE correspondence: for any specification of payoffs u ∈ RN |S| , N (u) is the
set of NE for the game defined by u. By Theorem 1, N is non-empty-valued.
The following result states that the limit of a sequence of NE is a NE (in par-
ticular, the set of NE for a fixed u is closed) but that for some games there are NE
for which there are no nearby NE in some nearby games.
Theorem 7. N is upper hemicontinuous but (for |S| ≥ 2) not lower hemicontinu-

ous.
Proof.
1. Upper Hemicontinuity. Since Σ is compact, it suffices to prove that N has

a closed graph, for which it suffices to prove that every point (u, σ) in the
complement of graph(N ) is interior. Take any (u, σ) in the complement of
graph(N ). Then there is an i and a pure strategy si for which
ui (si , σ−i ) − ui (σi , σ−i ) > 0.
25
By continuity of expected utility in both ui and σ, this inequality holds for all
points within a sufficiently small open ball around (u, σ), which completes the
argument.
2. Lower Hemicontinuity. Consider the game in figure 16. For any ε > 0, the
L R
A 1, ε 1, 0
Figure 16: The NE correspondence is not lower hemicontinuous.
unique NE is (A, L). But if ε = 0 then any mixed strategy profile is a NE, and
in particular (A, R) is a NE. Therefore, for any sequence of ε > 0 converging
to zero, there is no sequence of NE converging to (A, R). This example implies
a failure of lower hemicontinuity in any game with |S| ≥ 2.
The set of NE need not be connected, let alone convex. But one can prove
that for any finite game, the set of NE can be decomposed into a finite number of
connected components; see Kohlberg and Mertens (1986). In the Battle of the Sexes
(Example 5 in Section 4.3), there are three components, namely the three NE. In
the entry deterrence game of Example 21 (Section 4.9), there are two components:
the pure strategy (In, A) and the component of (O, (q, 1 − q)) for q ∈ [0, 1/16].
References
Aumann, Robert and Michael Maschler. 1972. “Some Thoughts on the Minimax
Principle.” Mangement Science 18:54–63.
Bernheim, B. Douglas. 1984. “Rationalizable Strategic Behavior.” Econometrica
52(4):1007–1028.
Clancy, Tom. 1984. The Hunt for Red October. U.S. Naval Institute.
Cournot, Augustin. 1838. Researches into the Mathematical Principles of the Theory
of Wealth. New York: Kelley. Translation from the French by Nathaniel T. Bacon.
Translation publication date: 1960.
Daskalakis, Constantinos, Paul Goldberg and Christos Papdimitriou. 2008. “The
Complexity of Computing a Nash Equilibrium.” University of California, Berkeley.
Fudenberg, Drew and Jean Tirole. 1991. Game Theory. Cambridge, MA: MIT Press.
Gibbons, Robert. 1992. Game Theory for Applied Economists. Princeton, NJ:
Princeton University Press.
26
Harsanyi, John. 1973. “Games with Randomly Disturbed Payoffs: A New Rationale
for Mixed-Strategy Eequilibrium Points.” International Journal of Game Theory
2:1–23.
Hu, Tai-Wei. 2014. “Unpredictability of Complex (Pure) Strategies.” Games and

Economic Behavior 88:1–15.
Jehle, Geoffrey A. and Phillip J. Reny. 2000. Advanced Microeconomic Theory.

Second ed. Addison Wesley.
Kohlberg, Elon and Jean-François Mertens. 1986. “On the Strategic Stability of
Equilibria.” Econometrica 54(5):1003–1037.
Kreps, David. 1990. A Course in Microeconomic Theory. Princeton, NJ: Princeton

University Press.
Kreps, David and Jose Scheinkman. 1983. “Quantity Precommitment and Bertrand
Competition Yeild Cournot Outcomes.” Bell Journal of Economics 14(2):326–337.
Luce, R. Duncan and Howard Raiffa. 1957. Games and Decisions. Wiley.
Mas-Colell, Andreu, Michael D. Whinston and Jerry R. Green. 1995. Microeconomic

Theory. New York, NY: Oxford University Press.
McLennan, Andrew. 2005. “The Expected Number of Nash Equilibria of a Normal

Form Game.” Econometrica 75:141–174.
Myerson, Roger. 1999. “Nash Equilibrium and the History of Economic Theory.”
Journal of Economic Literature 37:1067–1082.
Nachbar, John and Jonathan Weinstein. 2015. “Nash Equilibrium.” Notices of the
American Mathematical Society .
Nash, John. 1950a. “Equilibrium Points in N -Person Games.” Proceedings of the

National Academy of Sciences 36(1):48–49.
Nash, John. 1950b. Non-Cooperative Games PhD thesis Princeton.
Nash, John. 1951. “Non-Cooperative Games.” Annals of Mathematics 54(2):286–

295.
Nash, John. 1953. “Two-Person Cooperative Games.” Econometrica 21(1):128–140.
Osborne, Martin and Ariel Rubinstein. 1994. A Course in Game Theory. Cambridge,
MA: MIT Press.
Osborne, Martin J. 2008. Strategic and Extensive Form Games. In The New Palgrave
Dictionary of Economics, 2nd Edition. Palgrave MacMillan Ltd.
27
Pearce, David G. 1984. “Rationalizable Strategic Behavior and the Problem of
Perfection.” Econometrica 52(4):1029–1049.
Reny, Philip J.2̇008. Non-Cooperative Games (Equilibrium Existence). In The New

Palgrave Dictionary of Economics, ed. Lawrence E. Blume Steven N. Durlauf.
Second ed. New York: Mcmillan.
Sontag, Sherry and Christopher Drew. 1998. Blind Man’s Bluff. PublicAffairs.
von Neumann, John. 1928. “Zur Theorie der Gesellschaftsspiele.” Mathematische

Annalen 100:295–320.
von Neumann, John and Oskar Morgenstern. 1944. Theory of Games and Economic
Behavior. Princeton University Press.
Wilson, Robert. 1971. “Computing Equilibria of N -Person Games.” SIAM Journal

of Applied Mathematics 21:80–87.
28
John Nachbar
September 6, 2016
Motivating and interpreting equilibrum.

1 Introduction
There are four main stories used to justify equilibrium in game theory, and in par-
ticular Nash equilibrium (NE).
1. Pre-play Negotiation.
2. The Big Book of Game Theory.
3. Introspection.
4. Learning/evolution.
I discuss these in turn.
2 Pre-play negotiation.
2.1 Pre-play negotiation.
Suppose that players meet prior to the start of play and agree on how they will play
during the game. For the agreement to work, it must be self-enforcing: it must be
that each player is optimizing, which is to say, the agreement must constitute a NE.
I find this motivation compelling, but there are subtleties.
The first subtlety is that if we believe that players can negotiate prior to the
start of the game, then the negotiation itself should, strictly speaking, be itself part
of the analysis: the game now has two stages, a negotiation stage followed by the
original game, which begs the question how equilibrium is reached in this larger
game.
The second subtlety is that if players can meet prior to play then they can change
the game itself. A simple example of this is that in a game like Battle of the Sexes
(Figure 1), the players can first toss a coin to determine which of the two pure
A B
A 8, 10 0, 0
B 0, 0 10, 8
Figure 1: A game box for Battle of the Sexes.
1
A B
A 1/2 0
B 0 1/2
Figure 2: A correlated distribution over S for Battle of the Sexes.
strategy equilibria, (A, A) or (B, B), will be played. To an observer, the resulting
distribution will look correlated, as in Figure 2.
In effect, the players have constructed a new game in which nature moves first
and with equal probability sends a signal, either H or T , to the two players. In this
new game, it is a NE for the players to play (A, A) following signal H and to play
(B, B) following signal T (there are other NE as well; I am only saying that this is
one of them).
Note that the argument that a pre-play agreement must constitute a NE still
holds; the wrinkle is that the pre-play negotiation could change the game, and the
NE of the new game could look like correlated behavior in the original game.
2.2 Correlated equilibrium.

More broadly, pre-play negotiation is one way to motivate a solution concept called
correlated equilibrium, first introduced in Aumann (1974). Informally, a correlated
equilibrium is a distribution σ c over S with the property that for each player i and
any strategy si , if si is played with positive probability then si is a best response to
the conditional distribution σ−i |si .
The set of correlated equilibria is a convex polytope in Σc ; in contrast, the set
of NE need not even be connected. If σ is a NE mixed strategy profile then the
induced distribution over S is a correlated equilibrium in this sense. On the other
hand, as the Battle of the Sexes example illustrates, there are correlated equilibrium
distributions that cannot be generated by a NE. And even the Battle of the Sexes
example is misleading, because in it, the correlated equilibrium is randomizing over
NE distributions; formally the correlated equilibrium distribution is in the convex
hull of the NE distributions. The next example shows that correlated equilibrium
distributions can lie outside the convex hull of the set of NE distributions.
Example 1. The game in Figure 3 is inspired by “chicken,” perhaps most vividly
represented in the movie Rebel without a Cause: the players drive their cars toward
the edge of a cliff; the winner is the person who jumps last.
A B
A 6, 6 2, 7
B 7, 2 0, 0
Figure 3: Chicken
2
In the game as depicted, there are three NE, the pure NE (A, B) and (B, A) and
the mixed NE in which each player plays A with probability p = 2/3. This yields
the distribution over S given in Figure 4.
A B
A 4/9 2/9
B 2/9 1/9
Figure 4: σ c for the mixed NE of Chicken
But this game also has a correlated equilibrium distribution the distribution
given in Figure 5. This correlated equilibirium distribution cannot be generated by
A B
A 1/3 1/3
B 1/3 0
Figure 5: σ c for a correlated equilibrium in Chicken
any distribution over NE distributions.
3 The Big Book of Game Theory

The idea here is that society has settled on a correct way to play any game. The idea
is similar to pre-play negotiation, but now, in a sense, all games have been negotiated
for all players, in advance. Again, the argument is that for the prescribed play to
make sense, it must be that the prescribed play in any game is a Nash equilibrium.
The main example of a Big Book of Game Theory is Harsanyi and Selten (1988); it
is viewed more as a benchmark than as a practical description for how real games
are played.
4 Introspection
Instead of negotiating prior to play, players try to think through the game, which
includes thinking through logic of the sort, “I think that he thinks that she thinks
. . . .” Absent strong assumptions, arguments of this sort tend to support solution
concepts that are weaker than Nash equilibrium, notably rationalizability and cor-
related equilibrium. For a survey, see Brandenburger (2008).
3
5 Learning and Evolution
Players play various games with various opponents over time and gradually ad-
just their behavior. The learning literature then asks what happens to behavior
asymptotically. The set of possible learning stories is very large, encompassing both
different environments (repeated play of the same game by the same players, play
on networks, and so on) and different models of player behavior (ranging from im-
itation to sophisticated Bayesian optimization). For some games, large classes of
learning stories all predict convergence to Nash equilibrium. For other games, pre-
dictions are more heterogenous, and involve weaker equilibrium concepts, such as
correlated equilibrium or rationalizable sets. For surveys, see Nachbar (2008) and
Sandholm (2008); for a more extensive, albeit older, treatment, see Fudenberg and
Levine (1998).
References
Aumann, Robert. 1974. “Subjectivity and Correlation in Randomized Strategies.”
Journal of Mathematical Economics 1:67–96.
Brandenburger, Adam. 2008. Epistemic Game Theory: An Overview. In The New

Fudenberg, Drew and David K. Levine. 1998. Theory of Learning in Games. Cam-
bridge, MA: MIT Press.
Harsanyi, John and Reinhard Selten. 1988. A General Theory of Equilibrium Selec-
tion in Games. Cambridge: MIT.
Nachbar, John. 2008. Learning and Evolution in Games: Belief Learning. In

The New Palgrave Dictionary of Economics, 2nd Edition. New York: Palgrave
MacMillan Ltd.
Sandholm, William. 2008. Learning and Evolution in Games: An Overview. In The

New Palgrave Dictionary of Economics, ed. Lawrence E. Blume Steven N. Durlauf.
4
John Nachbar
May 2, 2017
The Revenue Equivalence Theorem1
1 Introduction.
The Revenue Equivalence Theorem gives conditions under which some very different
auctions generate the same expected revenue. I give examples in the next section.
A special case of the theorem was first stated and proved in Vickrey (1961) and
Vickrey (1962). General revenue equivalence results appeared in Myerson (1981)
and Riley and Samuelson (1981).
The argument underlying revenue equivalence plays a role in some other impor-
tant results in game theory, including the Myerson-Satterthwaite theorem, Myerson
and Satterthwaite (1983), which states that, in a broad class of settings, inefficiency
in bargaining is an unavoidable consequence of asymmetric information. For text-
book treatments of revenue equivalence, both in auctions and other settings, see
Krishna (2009), Milgrom (2004), and Borgers, Krahmer and Strausz (2015).
2 Examples.
A seller owns one unit of an indivisible object. I ≥ 2 other people, the bidders, are
interested in acquiring the object. Bidder i values the object at v i , which I also
refer to as the bidder’s value or type (I use these terms interchangeably). The v i are
distributed independently and uniformly on [0, 1]. Bidder i knows v i but not the
values of the other bidders. The distribution of the v i is common knowledge.
Consider the following four auctions. In all four auctions, each bidder i submits
a bid bi ≥ 0. And in all four auctions the high bidder wins. If there is more
than one highest bid then a referee chooses a winner at random from among the
highest bidders with equal probability. (Ties are, in fact, a zero probability event in
equilibrium in all four examples.) Where the auctions differ is in the payment rules.
1. First Price Sealed Bid Auction (1P). The winning bidder pays her bid, for a
net utility of
v i − bi .
The other bidders pay nothing, for a net utility of 0.
One can compute that the unique Nash equilibrium of 1P is for each bidder i
to bid
vi
bi = v i − .
I
1
1
Thus, the bidder bids less than the true value (“shaves” her bid), but her bid
is close to the true value if I is large.
2. Second Price Sealed Bid Auction (2P). The winner pays the second highest
bid (which equals the highest bid if there are two or more bids tied for highest
bid; but this is a zero probability event in equilibrium). The other bidders pay
nothing, for a net utility of 0.
2P has a Nash equilibrium in which each bidder bids her true value.
bi = v i .
One can check that this bid is actually weakly dominant: a deviation from
this bid never helps the bidder and sometimes hurts him.2
3. All Pay Auction. All bidders pay their bids.

There is an equilibrium in which the bid is,
(v i )I
bi = (v i )I − .
I
For v i ∈ (0, 1), this is strictly smaller than the equilibrium bid in 1P, and the
difference between the two is larger the larger is I. Indeed, even for v i very
close to one, a bidder in the All Pay auction will bid close to 0 if I is large.
4. Third Price Sealed Bid Auction. Assume I ≥ 3. The winner pays the third
highest bid (which equals the highest bid if there three or more bids tie for
highest bid; but this a zero probability event in equilibrium). The other bid-
ders pay nothing, for a net utility of 0.
There is an equilibrium in which the bid is,
vi
bi = v i + .
I
This bid is larger than the bidder’s value, although it is close to the bidder’s
value if I is large.
Now for the punchline: in the given Nash equilibria of all four auctions, expected
revenue to the seller is the same, namely,
I −1
.
I +1
The claim is not that the auctions generate the same revenue for every realization
of values. For example, if I = 2 and the values are v 1 = 3/4 and v 2 = 1/2 then the 1P
2
There are other Nash equilibria. For example, it is an equilibrium for bidder 1 to bid 1 and for
all other bidders to bid 0. But I will focus on the truthful equilibrium.
2
auction generates revenue of (3/4)/2 = 3/8 while the 2P auction generates revenue
of 1/2: the 2P auction generates higher revenue. But if v 1 = 3/4 and v 2 = 1/4 then
the 1P auction still generates revenue of 3/8 while the 2P the auction generates
revenue of 1/4: the 1P auction generates higher revenue. Revenue equivalence says
that these differences wash out in expectation.
3 Auctions.
3.1 The information structure.
A seller owns one unit of an indivisible object. I other people, the bidders, are
interested in acquiring the object. Bidder i values the object at v i ≥ 0, called the
bidder’s value or type (I use these terms interchangeably).
Let F be the joint distribution over type profiles v = (v 1 , . . . , v I ). Let supp(F )
denote the support of F : supp(F ) is the smallest closed set that gets probability 1.
For each i, let V i be the set of v i that appear in supp(F ):
V i = {vi ≥ 0 : ∃ v −i s.t. (v i , v −i ) ∈ supp(F )},
where v −i is a type profile for bidders other than i and (v i , v −i ) is the type profile
determined by v i and v −i . Since v i has a lower bound (namely 0) and V i is closed,
V i has a smallest element, v i . I assume that V i is also bounded above and hence
has a largest element, v i . Thus V i ⊆ [v i , v i ]. The Revenue Equivalence Theorem
assumes that, in fact, V i = [v i , v i ]. The analysis extends easily to cases in which V i
is not bounded above, provided the required expectations are finite.
If F is independent, then supp(F ) can be written as the product supp(F ) = V 1 ×
· · · × V I . The Revenue Equivalence Theorem assumes independence but some of the
definitions below do not. Section 6.2 shows by example that Revenue Equivalence
can fail if independence is dropped.
Remark 1. Statements of the Revenue Equivalence Theorem often require that the
marginals have densities: no v i has positive probability. The density assumption is
not, in fact, needed for the theorem itself. The density assumption does, however,
sometimes plays a role in applications.
Bidder i knows v i but not the values of the other bidders. I assume, however,
that F is common knowledge (it is part of the structure of the game).
Henceforth, hold this basic information structure fixed.
3.2 An auction game form.

i i
Q ani action a ∈ A .
Now consider a specific auction. In this auction, bidder i can take
i
The set A is non-empty but otherwise unrestricted. Let A = A ; an element of
a ∈ A, a = (a1 , . . . , aI ), is an action profile .
3
In some auctions, ai could be just a number, the bid, as was the case in the
auctions described above. But in other auctions, ai could be complicated. The
auction could, for example, be an English (open-outcry ascending price) auction,
commonly used for art: bids are observable and, at each instant, any bidder can
enter a new bid, provided the new bid is higher than the highest previous bid. ai
is then the name of the function that specifies what bidder i does as a function of
time and the history of prior bids.
Since a bidder observes her own value but not those of the other bidders, a
bidder can condition her action on her own value, but she cannot condition on the
values of the other bidders. A pure strategy for bidder i in the auction is therefore
a function,
si : V i → Ai .
I denote by s the strategy profile (s1 , . . . , sI ). Abusing notation somewhat, let s(v) =
(s1 (v 1 ), . . . , sI (v I )).
3.3 Utility.
I assume that bidders are risk neutral. I discuss risk aversion briefly in Section 6.4.
Action profile a determines outcomes via functions π i : A → [0, 1] and τ i : A →
R.
π i (a) is the probability that bidder i gets the object. For any a ∈ A, I require
that X
π i (a) ≤ 1.
i
π i (a)
P
I allow i < 1: the seller could retain the object.
τ (a) is the payment made by the bidder. I do not assume that τ i (a) is positive,
i
although this will be the case in typical auctions.

Finally, utilities depend on types and on outcomes.3 In particular, if bidder i
has value v i and the bid profile is a then bidder i wins with probability q i = π i (a)
and pays ti = τ i (a); bidder i’s expected utility in this case is,
ui (v i , q i , ti ) = q i v i − ti .
This specification of τ i appears to rule out auctions in which bidder i does not
pay if she does not win. Such auctions have not, in fact, been ruled out. One merely
has to interpret the payment as an expected payment. Fix a and let q i = π i (a).
Suppose that bidder i makes the payment t̂i if she gets the object and 0 if she does
not. In expectation, therefore, she pays ti = q i t̂i . This notational trick works if, as
I assume, preferences are risk neutral.
3
In game theory, it is standard to refer to utility as the payoff. I use the term “utility”, rather
than “payoff”, to avoid confusion with the similar-sounding “payment”.
4
Note also that there are auctions where the bidder pays even if the bidder does
not win. This was true for the All Pay auction discussed in Section 2, and it is true
for raffles, discussed in the next remark.
Remark 2. As an example of an “auction” in which q i is typically strictly between
0 and 1, consider a raffle. The action ai ∈ R+ represents the purchase of ai raffle
tickets. Assuming at least one person buys at least one raffle ticket,
ai
q i = PI ,
j
j=1 a
and
ti = ai .
3.4 Nash equilibrium.

In the auction game, nature moves first and selects a value profile v according to
the joint distribution F . In this game, a strategy profile s is a (pure strategy) Nash
equilibrium (NE) iff, for each i and each v i ∈ V i , si (v i ) maximizes bidder i’s expected
utility given the strategies of the other bidders. The expectation is with respect to
the other bidders’ values, v −i , which bidder i does not know. It is customary to call
a NE in this sort of information setting a Bayesian Nash Equilibrium.4
4 Revenue Equivalence.
Fix a strategy profile s (not necessarily an equilibrium profile). Define Qi : V i →
[0, 1] by,
Qi (v i ) = Ev−i [π i (s(v i , v −i ))|v i ],
where v −i is the vector of values of the other bidders and (v i , v −i ) is the vector of
all the values, including v i . Qi (v i ) is the probability that i gets the good, given the
strategy profile s and given i’s own type v i . Similarly, define T i : V i → R by,
T i (v i ) = Ev−i [τ i (s(v i , v −i ))|v i ].

4
Strictly speaking, the equilibrium concept here is somewhat stronger than standard Nash equi-
librium, because under standard Nash equilibrium, si is required to be optimal only ex ante: si
maximizes the bidder i’s expected utility given the strategies of the other bidders, where the expec-
tation is over v. This is equivalent to requiring that si (v i ) maximize i’s expected utility for a subset
v i of probability one, where the expectation is over v −i . If, for example, the marginal distribution
over V i has a density, so that each individual value of v i has probability zero, then si could satisfy
the standard NE optimization condition even if si (v i ) is not optimal for some v i .
5
T i (v i ) is bidder i’s expected payment, given the strategy profile and given i’s own
type. Thus, if the strategy profile is s, then bidder i, with value v i , has an expected
utility of,
U i (v i ) = Qi (v i )v i − T i (v i ).
In words, a bidder’s type influences her expected utility through three channels.
There is a direct affect: v i is the value of the object if bidder i wins. There are also
two indirect effects. First, v i determines i’s action ai via the strategy si . Second,
the expectations used for Qi and T i are conditional on v i . Bidder i, if she has a
high v i , could infer that the values of the other bidders are likely to be high as well,
and this inference affects her posterior over what actions the other bidders will take.
However, the Revenue Equivalence Theorem assumes this second indirect effect
away, by assuming that F , the joint distribution over bidder values, is independent.
Section 6.2 discusses the independence assumption further.
The key fact underlying Revenue Equivalence is the following, which I prove in
Section 5.
Theorem 1 (The Integral Condition). Suppose that F is independent and that, for
each i, V i is an interval, V i = [v i , v i ]. Consider any strategy profile s. If si is
optimal for bidder i, then, for any v i ∈ V i ,
Z vi
i i i i
U (v ) = U (v ) + Qi (x) dx. (1)
vi
Equation (1) is often called the Integral Condition. Since T i (v i ) = Qi (v i )v i −

U i (v i ), the integral condition implies that,
Z vi
T i (v i ) = Qi (v i )v i − U i (v i ) − Qi (x) dx. (2)
vi
This says that bidder i’s expected payment when her value is v i depends only on
U i (v i ) and Qi . Since T i (v i ) is bidder i’s expected payment, it is the seller’s expected
revenue. Therefore, in a NE, since every bidder is optimizing, the seller’s overall
expected revenue is,
Ev T 1 (v 1 ) + · · · + T I (v I ) .

(3)
Theorem 2 (The Revenue Equivalence Theorem). Suppose that F is independent

and that, for each i, V i is an interval, V i = [v i , v i ]. Consider any two auctions and
any two NE of those auctions. If Qi and U i (v i ) are the same for each i in both NE,
then the expected revenue to the seller is the same in both NE.
Proof. Almost immediate from Equation 3 since, in NE, all bidders are optimizing,
and since, by Equation 2, expected revenue from an optimizing bidder i depends
only on Qi and U i (v i ).
6
In all four auction examples in Section 2, v i = 0 and U i (0) = 0: if you have
the lowest possible value, you expect to get nothing in the auction. And in all
four auction examples, the bidder with highest v i wins, hence, for every i, Qi is
the same.5 Since U i (v i ) and Qi are the same in all four auctions, the Revenue
Equivalence theorem implies that the expected revenue must be the same, as was
indeed the case.
5 Proof of Theorem 1, the Integral Condition.

Define Q̃i : V i × V i → [0, 1] by,
Q̃i (v i , v ∗i ) = Ev−i [π i (s(v i , v −i ))|v ∗i ].
Q̃i (v i , v ∗i ) is the probability that bidder i wins if her true value is v ∗i but she acts
as if her type were v i and takes action ai = si (v i ). Define T̃ i analogously.
The assumption that F is independent implies that conditioning on the true
value is irrelevant, in which case,
Q̃i (v i , v ∗i ) = Qi (v i ),
and,
T̃ i (v i , v ∗i ) = T i (v i ).
Therefore, given independence, bidder i’s expected utility when she is of type v ∗i
but plays as if she were of type v i is,
Ũ i (v i , v ∗i ) = Qi (v i )v ∗i − T i (v i ).
Note that U i (v ∗i ) = Ũ (v ∗i , v ∗i ). In Section 6.2, I discuss what happens when inde-

pendence is violated.
If si is optimal, then, for any v i and v ∗i , it must be that
U i (v ∗i ) = Ũ (v ∗i , v ∗i ) ≥ Ũ (v i , v ∗i ). (4)
Conversely,
U i (v i ) = Ũ (v i , v i ) ≥ Ũ (v ∗i , v i ). (5)
These inequalities are called incentive compatibility (IC) conditions. The IC in-
equalities imply,
U i (v ∗i ) − U i (v i ) ≥ Qi (v i )v ∗i − T i (v i ) − U i (v i )
= Qi (v i )v ∗i − T i (v i ) − [Qi (v i )v i − T i (v i )]
= Qi (v i )(v ∗i − v i ).
5
Explicitly, since the marginal distributions have densities, the probability that two or more v i
are tied for being highest is zero. Therefore, Qi (v i ) equals the probability that v i is strictly higher
than any v j , j 6= i.
7
Similarly,
U i (v ∗i ) − U i (v i ) ≤ U i (v ∗i ) − [Qi (v ∗i )v i − T i (v ∗i )]
= Qi (v ∗i )v ∗i − T i (v ∗i ) − [Qi (v ∗i )v i − T i (v ∗i )]
= Qi (v ∗i )(v ∗i − v i ).
Combining,
Qi (v i )(v ∗i − v i ) ≤ U i (v ∗i ) − U i (v i ) ≤ Qi (v ∗i )(v ∗i − v i ). (6)
If v ∗i > v i , this implies Qi is weakly increasing in v i : as v i increases, it is increasingly

likely that bidder i will get the object.
Since the domain of Qi , namely V i , is an interval and since Qi is weakly increas-
ing, Qi is (Riemann) integrable (Rudin (1976), Theorem 6.9). Inequality (6) implies
that the integral satisfies,
Z vi
i i i i
U (v ) − U (v ) = Qi (x) dx, (7)
vi
where v i is the minimum value of v i .6

Rearranging Equation (7) gives the Integral Condition, Equation (1), which
proves Theorem 1.
Remark 3. Assume for the moment that Qi and T i are differentiable. As in the
proof of Theorem 1, let Ũ i (v i , v ∗i ) be the expected utility to a bidder with value
v ∗i when she plays action ai = si (v i ). Thus, again as in the proof of Theorem 1,
assuming independence,
Ũ i (v i , v ∗i ) = Qi (v i )v ∗i − T i (v i ).
Consider the problem of maximizing Ũ i (v i , v ∗i ) with respect to v i (the type that

bidder i is pretending to be), holding the true type, v ∗i , fixed. If si is indeed optimal,
then the solution is to set v i = v ∗i (so that bidder i plays action a∗i = si (v ∗i )). U i is
then the value function for this maximization problem. By the Envelope Theorem,
6
Explicitly, assume that v i > v i (if not, then the argument becomes trivial). Recall that a
function is integrable iff its upper and lower integrals are equal (and finite). Consider any finite
partition of [v i , v i ] deterimined by v 1 = x1 < · · · < xK = v i . Consider any xk < xk+1 . Since Qi
is weakly increasing, the maximum value of Qi on this interval is Qi (xk+1 ), so that the associated
rectangle has area Qi (xk+1 )(xk+1 − xk ). By (6), this area is bounded below by U i (xk+1 ) − U i (xk ).
Summing over all partition intervals, all but two of the U i (xk ) cancel, yielding a total lower bound
of U i (v i ) − U i (v i ). I get this same lower bound for every partition. Hence the upper integral is
bounded below by U i (v i ) − U i (v i ). By an analogous argument, the lower integral is bounded above
by U i (v i ) − U i (v i ). Since Qi is integrable, this yields Equation (7).
8
DU i (v ∗i ) = D2 Ũ i (v ∗i , v ∗i ) (where D2 is the derivative with respect to the second
argument), hence
DU i (v ∗i ) = Qi (v ∗i ). (8)
Integrating Equation (8) this gives Equation (7).
In terms of the three channels by which v i can affect U i (see Section 4), the
Q (v ∗i ) in Equation (8) is from the direct channel. The assumption that the joint
i
distribution is independent has killed off the second indirect channel (conditional
expectation). Section 6.2 discusses what happens to Equation (8) if independence is
relaxed. Finally, the Envelope Theorem says that, because of optimization, the first
indirect channel (choice of action as a function of type), has no first-order affect on
utility.
One cannot, in fact, simply assume that Qi and T i are differentiable, since they
depend on the strategies, which are not required to be differentiable. Milgrom and
Segal (2002) shows that, despite this issue, a generalized envelope theorem argument
still applies.
Remark 4. Inequalities (4) and (5) are often called incentive compatibility conditions.
6 Qualifications and Extensions.

6.1 If the V i are not intervals.
Revenue equivalence can fail if the V i are not intervals. As an example, suppose
that I = 2 and that the v i are equally likely to be 0 or 1, independently. Hence,
V 1 = V 2 = {0, 1} (not [0, 1]).
In this environment, the 2P auction still has an equilibrium in which bidders
bid their true values; in that equilibrium, the object goes to the bidder with highest
value, and the expected utility of a bidder of lowest value is zero. Revenue in this
auction is zero except when both bidders have values v i = 1, in which case revenue
is 1. Since the probability that both bidders have value 1 is 1/4, expected revenue
is 1/4.
But the seller can do better with the following “auction”, which I call PP (posted
price). The object is given to a bidder who says “buy”. If no bidder says “buy”,
the seller keeps the object. (If both bidders say “buy”, each bidder is equally likely
to get the object.) If there is a winner, the winner pays 2/3. In this auction, the
(unique) equilibrium has each bidder say “buy” iff her value is 1.
In both equilibria of both auctions, ui (v i ) = ui (0) = 0. And both have the same
Q (for i = 1: Q1 (0) = 0; Q1 (1) = 3/4, which is the probability that v 2 = 0 plus
i
one half times the probability that v 2 = 1). Therefore, if Revenue Equivalence held,
expected revenue in both would be 1/4. But in the PP auction, expected revenue
is 1/2.
9
A partial intuition for what is going on here is that any equilibrium of any
auction must satisfy the incentive compatibility (IC) conditions, Inequality (4) and
Inequality (5). Very loosely, the more IC conditions there are, the more restricted
is equilibrium behavior. If the V i are intervals, then bidder i has, for each realized
v i , a continuum of IC conditions. In contrast, if V i = {0, 1}, then bidder i has, for
each realized v i , only a single IC condition (e.g., it cannot be profitable for a v i = 1
bidder to behave as if v i = 0, or vice versa).
In terms of the formal argument, the Integral Condition, Equation (1), assumes
that the relevant functions are all defined on an interval. And the Envelope The-
orem condition (Equation (8) in Remark 3 in Section 5) assumes that the relevant
expectations are defined for types arbitrarily near the true type.
6.2 If F , the joint type distribution, is not independent.

Revenue Equivalence can fail if the joint type distribution is not independent. As
an example, suppose that I = 2, that v 1 is distributed uniformly on [0, 1], and that
v 2 = 1 − v 1 . Here, V 1 = V 2 = [0, 1].
The 2P auction still has an equilibrium in which bidders bid their true value.
The expected revenue from this equilibrium is the expected value of the low value
bidder, namely 1/4.
But the seller can do better with the following auction. Each bidder i bids a
number bi ∈ [0, 1]. If the numbers satisfy b2 = 1 − b1 , then the high bidder gets the
object and pays 1/2. Otherwise, the seller keeps the object and no money changes
hands. This auction also has an equilibrium in which bidders bid their true values.
In both equilibria of both auctions, ui (v i ) = ui (0) = 0. And both have the same
Q (Qi (v i ) = 0 if v i < 0; Qi (v i ) = 1 if v i > 0; Qi (1/2) = 1/2, but this last case
i
has zero probability of occurring). Therefore, if Revenue Equivalence held, expected

revenue in both would be 1/4. But in the new auction, which exploits the negative
correlation in values, the expected revenue is 1/2.
In terms of the formal argument, if F is not independent then, in the notation
used in the proof of Theorem 1 in Section 5, it is no longer necessarily true that
Q̃(v i , v ∗i ) = Qi (v i ) and T̃ (v i , v ∗i ) = T i (v i ). This vitiates the subsequent analysis
because the expected payment terms may not drop out. In terms of the Envelope
Theorem approach (Remark 3 in Section 5), Equation (8) becomes,
DU i (v ∗i ) = Q̃i (v ∗i , v ∗i ) + D2 Q̃i (v ∗i , v ∗i ) − D2 T̃ i (v ∗i , v ∗i ).
Note that there is now a payment term on the right-hand side.
6.3 Asymmetric type distributions.

If the v i are drawn from different distributions then the equilibria of a 1P auction
may no longer have the property that the highest valuation bidder wins the object.
As a consequence, Revenue Equivalence between the 1P and 2P auctions can break
10
down. The ranking could go either way: the 1P auction could generate more ex-
pected revenue than the 2P auction or vice versa, depending on the distribution of
the v i .
6.4 Risk aversion.

If bidders are risk averse, meaning that utility is a concave function of v i less the
bidder’s payment, then Revenue Equivalence can fail. Consider, in particular, the
symmetric setting of Section 2. If bidders are risk averse, then the 1P auction
generates higher expected revenue than the 2P auction, even though both auctions
still assign the object to the high valuation bidder. The intuition is that it remains
an equilibrium in the 2P auction for bidders to bid their type, but risk aversion
causes bidders in a 1P auction to shave their bids by less than under risk neutrality,
and under risk neutrality, the expected revenues of the two auctions were equal.
(If we apply a concave function only to v i , then Revenue Equivalence continues to
hold.)
6.5 Interdependent values.

Suppose that values are interdependent in the sense that, rather than each bidder
knowing v i , each bidder receives a signal about v i , and v i could also depend on the
signals received by the other bidders. Two special cases are the following.
• Private values. Bidder i’s value is her signal; her value does not depend directly
on the signals of the other bidders. Bidder i cares about the signals of the
other bidders to the degree that these signals affect the behavior of the other
bidders, but that is a separate issue. Private values is the assumption in the
Revenue Equivalence Theorem as stated.
• Common values. All bidders value the object the same, but they have different
signals as to that value.
The classic citation for auctions with interdependent values is Milgrom and Weber
(1982).
If the value to bidder i is additively separable in the signals, and if the joint
distribution over signals is independent, then Revenue Equivalence continues to
hold. More generally, however, Revenue Equivalence may fail. In particular, in
many settings, the 2P auction can generate higher expected revenue than the 1P
auction for the following reason. In an auction with interdependent values, winning
can be bad news in the sense that the winner infers that the other bidders must
have received less promising signals of the object’s true value. This is called the
winner’s curse. In a 1P auction, anticipation of the winner’s curse gives each bidder
an additional reason to shave her bid, lowering overall expected revenue. In contrast,
the winner’s curse is less of an issue in the 2P auction precisely because bidders do
not pay their own bids.
11
6.6 Generalization.
Finally, see Milgrom and Segal (2002) for a discussion of how Revenue Equivalence
extends to more general environments (for example, environments with multiple
goods).
7 An Application: Computing the Equilibria of an Auc-

tion.
Consider, as an Section 2, a first price (1P) auction in which the bidder values are
all distributed uniformly on [0, 1]. Conjecture that there is a symmetric equilibrium
in which bids are strictly increasing in bidder value: bi (v i ) = bj (v j ) iff v i = v j . Then
Qi (v i ), the probability of winning when of type v i , is the probability that v i is the
highest value
Qi (v i ) = [v i ]I−1 .
In a 1P auction, the winning bidder pays her own bid, bi , so the expected payment
by a bidder of type v i is bi Qi (v i ), hence
T i (v i ) = bi [v i ]I−1 .
On the other hand, we know from Equation 2 that (assuming U i (v i ) = 0, as will in

fact be the case),
Z vi
T i (v i ) = Qi (v i )v i − U i (v i ) − Qi (x) dx
0
Z vi
= [v i ]I−1 v i − xI−1 dx
0
[v i ]I
= [v i ]I − .
I
Combining,
vi
bi = v i − ,
I
as stated in Section 2. (If v i = 0 then it is weakly dominant to bid bi = 0.) Note that
the bids are indeed symmetric and strictly increasing in bidder value, as conjectured.
On the other hand, for the all pay (AP) auction, the bidder always pays bi .
Conjecture again that there is a symmetric equilibrium in which bids are strictly
increasing in bidder value. Then we immediately get
[v i ]I
bi = T i (v i ) = [v i ]I − ,
I
as stated in Section 2. Again, note that the bids are indeed symmetric and strictly
increasing in bidder value, as conjectured.
12
Revenue Equivalence also provides a relatively easy way to compute the expected
revenue of any of the four auctions in Section 2: the expected revenue must be the
same as the expected revenue from the 2P auction, and that is simply the expected
value of the second highest type.
References
Borgers, Tilman, Daniel Krahmer and Roland Strausz. 2015. An Introduction to the
Theory of Mechanism Design. Oxford University Press.
Krishna, Vijay. 2009. Auction Theory. Second ed. Academic Press.
Milgrom, Paul. 2004. Putting Auction Theory To Work. Cambridge University

Press.
Milgrom, Paul and Ilya Segal. 2002. “Envelope Theorems for Arbitrary Choice
Sets.” Econometrica 70(2):583–601.
Milgrom, Paul and Robert Weber. 1982. “A Theory of Auctions and Competitive
Bidding.” Econometrica 50(5):1089–1122.
Myerson, Roger. 1981. “Optimal Auction Design.” Mathematics of Operation Re-

search 6:58–73.
Myerson, Roger and Mark Satterthwaite. 1983. “Efficient Mechanisms for Bilateral
Trading.” Journal of Economic Theory 28:265–81.
Riley, John and William Samuelson. 1981. “Optimal Auctions.” American Economic
Review 71:381–392.
Rudin, Walter. 1976. Principles of Mathematical Analysis. Third ed. New York:
McGraw-Hill.
Vickrey, William. 1961. “Counterspeculation, Auctions, and Competitive Sealed

Tenders.” Journal of Finance 16(1):8–37.
Vickrey, William. 1962. Auctions and Bidding Games. In Recent Advances in Game
Theory. Number 29 in “Princeton Conference Series” Princeton University Press
pp. 15–27.
13
John Nachbar
September 23, 2015
Refinements of Nash Equilibrium1
1 Overview
In game theory, “refinement” refers to the selection of a subset of equilibria, typically
on the grounds that the selected equilibria are more plausible than other equilibria.
These notes are a brief, largely informal, survey of some of the most heavily used
refinements. Throughout, “equilibria” means Nash equilibria (NE), unless I state
otherwise explicitly. And throughout, I assume that the game is finite. Extending
some of these concepts to more general settings can be non-trivial; see, for example,
Harris, Reny and Robson (1995).
My focus is on refinements that make an appeal to rationality arguments. This
is the traditional approach to refinement. An important alternative approach is
based on dynamic equilibration: players start out of equilibrium and in one sense or
another learn (or fail to learn) to play an equilibrium over time. For example, in 2x2
strategic form games with two pure strategy equilibria and one mixed, the mixed
equilibrium is ruled out by almost any dynamic story, even though it survives a host
of traditional refinements. I make some comments about dynamic refinements at
various points, but I do not attempt to be systematic.
For more thorough treatments of refinements, consult Fudenberg and Tirole
(1991a), van Damme (2002), and Govindan and Wilson (2008). For surveys on the
epistemic foundations of some key refinements (formalizations of arguments along
the lines, “it is optimal for me to play si because I think that my opponent will
play sj because I think that he thinks ...,”), see Brandenburger (2008) and the
introduction to Keisler and Lee (2011).
One word of warning: while it would be convenient if refinements lined up cleanly
in a hierarchy from least to most restrictive, the actual relationship between most
of them is complicated, as some of the examples below illustrate.
2 Strategic Form Refinements

2.1 Admissibility
Definition 1. Given a strategic form game G, a NE σ is admissible iff σ i (si ) > 0
implies that si is not weakly dominated.
1
1
Thus, an equilibrium is admissible iff no player plays a weakly dominated strat-
egy with positive probability. It is not hard to see that admissible equilibria always
exist in finite games. The informal rationale for admissibility is that weakly domi-
nated strategies are imprudent and that therefore players should avoid them.
In the game of Figure 1, (I, A) and (O, F ) are both pure strategy equilibria, but
A F
I 15, 15 −1, −1
O 0, 35 0, 35
Figure 1: Admissibility.
(O, F ) is not admissible since A weakly dominates F .
2.2 Iterated Admissibility

Iterated admissibility requires that players play only those strategies that survive the
iterated deletion of weakly dominated strategies. This can be an extremely powerful
refinement in some games. See, Ben-Porath and Dekel (1992) for a striking example.
On the epistemic foundation of iterated admissibility, see Brandenburger (2008) and
Keisler and Lee (2011).
With iterated strict dominance, it does not matter for predicted play whether
players delete strategies in turn or simultaneously, or whether all weakly dominated
strategies are deleted at each round, or just some of them. With weak dominance,
procedural issues of this sort become substantive. Consider the game in Figure
2. There are many equilibria here, among which are (a) an equilibrium where
A B C
A 10, 0 0, 10 3, 5
B 0, 10 10, 0 3, 5
C 5, 3 5, 3 2, 2
Figure 2: Iterated Admissibility: Order Matters.
both players randomize 50:50 between A and B, (b) an equilibrium where player A
randomizes 50:50 between A and B and player 2 plays C, and (c) an equilibrium
where player 1 plays C and player 2 randomizes 50:50 between A and B. Note that
all these equilibria have different expected payoffs. Of these, only the (a) equilibrium
is admissible, because C is weakly dominated by a 50:50 randomization between A
and B. But equilibrium (b) survives iterated admissibility if players delete strategies
in turn, with player 1 deleting first, and similarly for the (c) equilibrium. The
standard response to this issue is to consider maximal deletion, by all players, at
each round. Otherwise, as this example illustrates, iterated admissibility does not
even imply admissibility.
2
Even when deletion order is unambiguous, it is not always obvious that iterated
admissibility picks out the correct equilibrium. Consider the following game. (M, R)
L R
T 1, 1 0, 0
M 0, 100 100, 100
B 0, 0 99, 100
Figure 3: Iterated Admissibility: Other Considerations.
and (T, L) are both admissible equilibria, but (T, L) survives iterated admissibility
while (M, R) does not. (T, L) is arguably implausible. First, it is Pareto dominated
by (M, R). Second, from the perspective of many forms of learning dynamics, (M, R)
is at least as robust as (T, L).
2.3 Trembling Hand Perfection

The idea behind trembling hand perfection (THP), introduced in Selten (1975),
is that your opponents might deviate from their intended strategy (their hands
might tremble when pushing a button or setting a dial and thereby select the wrong
strategy) and you should prepare for that in choosing your strategy. THP makes
explicit the idea, implicit in much of the refinement literature, that the game is not
a complete description of the strategic situation but rather is an approximation. I
return to this issue briefly at the end of this subsection.
Formally, for any ε > 0, replace the original game with a new game in which
the set of mixed strategies is constrained to contain only those strategies that give
weight at least ε to every pure strategy. A NE of this constrained game is called an
ε-perfect equilibrium. For ε sufficiently small, an ε-perfect equilibrium exists (the
proof is nearly identical to the proof of existence of Nash equilibrium).
Definition 2. A strategy profile σ is a trembling hand perfect (THP) equilibrium iff

there is a sequence {σε } such that each σε is an ε-perfect equilibrium and limε→0 σε =
σ.
Note the quantifier: one does not need to check every sequence σε ; one must
merely show that there exists one such sequence. A compactness argument estab-
lishes that a THP equilibrium exists and continuity implies that any THP is a NE.
Hence THP is a refinement of NE. If σ is already fully mixed, then it is trivially
THP: for ε sufficiently small, take σε = σ.
THP implies admissibility and in two-player games, admissibility implies THP.
Thus, in two-player games, THP and admissibility are equivalent. In games with
three or more players, there are examples in which admissible (or even iterated
admissible) equilibria fail to be THP. Thus, in games with three or more players,
THP is stronger than admissibility.
3
To understand why, recall that while a strictly dominated strategy si is never
a best response to any distribution over opposing strategy profiles, the converse is
sometimes false in games with three or more players unless the distribution exhibits
correlation. For essentially the same reason, while a weakly dominated strategy si is
never a best response to any fully mixed distribution over opposing strategy profiles,
the converse is sometimes false unless the fully mixed distribution exhibits correla-
tion. THP implicitly assumes independent randomization (the strategy distribution
generated by σε is independent); as a consequence, THP can be more restrictive
than admissibility.
THP does not imply iterated admissibility. For example, the (M, R) equilibrium
in Figure 3 is THP but not iterated admissible. Conversely, iterated admissibility
implies THP in two player games but not in games with three or more players, for
the reason discussed in the preceding paragraph.
In applying THP, the game is fixed but players tremble away from their best
responses. Suppose that instead players best respond but that the true game is
nearby in payoff terms. One can then ask which equilibria of the original game are
limits of sequences of equilibria of nearby games. The answer is: all of the equilibria.
See Fudenberg, Levine and Kreps (1988) and also, for recent related work, Jackson,
Rodriguez-Barraquer and Tan (2011). The basic point is that refinements motivated
by approximation arguments require careful thought.
2.4 Properness
Properness, introduced in Myerson (1978), is motivated by the idea that if i does
tremble, he is more likely to tremble in directions that are least harmful to him.
Formally, say that a strategy profile σ is an ε-proper equilibrium iff (a) it is fully
mixed and (b) for any i and any pure strategies si and ŝi , if the payoff to si is greater
than that to ŝi then σ i (si ) ≥ σ(ŝi )/ε. For ε small, 1/ε is large, and hence this says
that σ i (si ) must be much larger than σ(ŝi ). One can show that ε-proper equilibria
exist, for ε sufficiently small (this requires more work than showing existence of
ε-perfect equilibria).
Definition 3. A strategy profile σ is a proper equilibrium iff there is a sequence
{σε } such that each σε is an ε-proper equilibrium and limε→0 σε = σ.
Again, proper equilibrium exist and any proper equilibrium is a NE. Hence
proper equilibrium is a refinement of Nash equilibrium. It is also almost immediate
that any proper equilibrium is THP. Hence, in particular, any proper equilibrium is
admissible, although not necessarily iteratively admissible.
As an example, consider the game in Figure 4. No strategy is weakly dominated,
hence, in particular, the NE (M, C) is THP and even survives the iterated deletion
of weakly dominated strategies (trivially, as no strategies are weakly dominated).
However, (M, C) is not proper. As long as most of player 2’s probability weight is
on C, T is a better response for player 1 than B. But if player 1 puts much more
4
L C R
T 4, 4 1, 1 2, 2
M 1, 1 1, 1 3, 0
B 2, 2 0, 3 4, 4
Figure 4: Properness
weight on T than on B, then L is a better response than C for player 2. Thus

(M, C) is not proper.
As a second example, consider again the game of Figure 3. In that game, the
NE (M, R) is proper even thought it does not satisfy iterated admissibility. (T, L)
is proper as well.
Properness is seldom invoked explicitly in applied work, possibly because it can
be difficult to verify. But it is important conceptually because it is relatively cleanly
linked to extensive form refinements. In particular, for a given strategic form game,
van Damme (1984) established that a proper equilibrium induces a sequential equi-
librium (Section 3.4) in every extensive form game consistent with the original strate-
gic form game.
2.5 Other Strategic Form Refinements

There is a bestiary of refinements and I will not be exhaustive. But let me briefly
mention some issues.
First, I would be remiss not to cite the literature on Kohlberg-Mertens Stability
(K-M Stability). The original paper, Kohlberg and Mertens (1986), is a trove of in-
teresting examples on refinements in general and on the connection between strategic
and extensive form refinements. The most restrictive K-M Stability refinement is
the one in Mertens (1989); its definition is extremely technical.
Second, σ is a strict NE iff, for every i, σ i is a strict best response to σ −i . A strict
NE is necessarily pure. A strict NE need not exist (none exists in matching pennies,
for example), in contrast to most of the other refinements discussed here. When
it does exist, however, it is iterated admissible, proper, and passes every version of
K-M Stability.
Third, another obvious refinement is to restrict attention to the NE that are
Pareto efficient relative to the other NE (they need not be Pareto efficient in an
overall sense; in games that model free riding in the provision of public goods, for
example, the NE are typically inefficient). One motivation is that if players can
engage in pre-game bargaining over which NE to play, then they will not choose a
NE that is Pareto dominated by another NE.
Application of the last two refinements is not always straightforward. For ex-
ample, in the game of Figure 3, the Pareto efficient pure NE is (M, R), which is not
strict, while the strict equilibrium is (T, L), which is Pareto dominated.
5
As another example, consider the game in Figure 5, which belongs to a class
of games sometimes called “Stag Hunt.” This game has two pure NE, (A, A) and
A B
A 100, 100 0, 98
B 98, 0 99, 99
Figure 5: Pareto Dominance
(B, B). Both are strict. (A, A) Pareto dominates (B, B), but (B, B) is arguably
more plausible. A is optimal if and only if the probability that your opponent plays
A is at least 99/101: you have to be almost certain that your opponent is playing A
in order for A to be optimal. Under learning dynamics, this translates into (B, B)
having a larger basin of attraction and under some (but not all) learning dynamics
there is a sense in which play converges to (B, B) no matter how players start out
initially; see Kandori, Mailath and Rob (1993).
3 Extensive Form Refinements

3.1 Subgame Perfection
Subgame perfection is a generalization of backward induction, one of the oldest so-
lution concepts in game theory. Whereas backward induction is typically restricted
to finite games of perfect information, subgame perfection applies to arbitrary ex-
tensive form games. Backward induction dates at least to Zermelo (1913); yes, the
article is in German. Subgame perfection was introduced in Selten (1967); yes, once
again, the article is in German. Rather than develop backward induction explicitly,
I focus on subgame perfection.
Informally, given a game Γ in extensive form, a subgame of Γ consists of a
decision node, all successor decision nodes and terminal nodes, and all associated
information sets, actions, and payoffs, provided that these all comprise a well defined
game. In particular, a node that is not in a singleton information set (an information
set containing exactly one node) cannot serve as the initial node of a subgame. For a
formal definition of subgames, see a game theory text such as Fudenberg and Tirole
(1991a).
A game always contains itself as a subgame, just as a set always contains itself
as a subset. A subgame that excludes at least one node of the original game is
called a proper subgame. Any (behavior) strategy profile induces, in a natural way,
a strategy profile in any subgame. Given a strategy profile σ and a subgame, the
subgame is on the play path iff the probability, under σ, of reaching the initial node
of the subgame is positive.
Theorem 1. σ is a NE iff it induces a NE in every subgame on the play path.
6
Proof. Since any game is a subgame of itself, the “if” direction is immediate. As
for “only if,” let σ be a strategy profile and let x be the initial node of a proper
subgame along the play path. If σ does not induce a NE in the proper subgame,
then there is some player i who can gain some amount ∆ > 0 within the sub-
game by deviating. If the probability of reaching x under σ is p > 0, then i can
gain p∆ > 0 in the overall game by deviating within the subgame, hence σ i was
not optimal for i, hence σ was not a NE. The proof then follows by contraposition.
Definition 4. σ is a subgame perfect equilibrium (SPE) iff it induces a NE in

every subgame.
In view of Theorem 1, SPE strengthens NE by requiring that players play a NE
in every subgame, and not just in every subgame on the play path. One implication
is that there is a difference between NE and SPE only if there is some subgame
(necessarily a proper subgame) that is not on the play path. In particular, if the
game has no proper subgames, which is frequently the case in Bayesian games, then
SPE has no “bite:” there is no difference between NE and SPE.
A canonical way to illustrate SPE is the entry deterrence game, a version of
which is given in Figure 6. Player 1 is a potential entrant into a market where
15, 15
A
1 2
I
O F
-1, -1
0, 35
Figure 6: A Basic Entry Deterrence Game.
player 2 is a monopolist. If player 1 stays out, then player 2 earns her monopoly
profit of 35. If player 1 enters, player 2 can either acquiesce (A), in which case both
firms earn 15, or start a price war (F), in which case both get a payoff of -1. This
game has one proper subgame, the trivial game rooted at player 2’s decision node.
The equilibrium of the proper subgame is for player 2 to choose A, and if player 1
anticipates A then player 1 will enter: the overall SPE is (I, A).
But (O,F) is also a pure strategy NE. Informally, under (O, F), player 2 threatens
a price war upon entry and player 1, believing this, stays out: player 2 is able to
deter entry. SPE captures the idea that this threat is not credible, because F is
7
not optimal should entry actually occur. This is not to say that entry deterrence
is impossible. Rather, the point is that a convincing model of entry deterrence will
have to have additional structure.
Another well known application of SPE is ultimatum bargaining. Player 1 can
propose an integer amount a ∈ {0, . . . , 100}. Player 2 can Accept or Reject. If
Player 2 Accepts then the payoffs are (a, 100 − a). If Player 2 Rejects then the
payoffs are (0, 0).
There is a different proper subgame for each of Player 1’s possible proposals.
For every proposal of 99 or less, the NE of that subgame has player 2 Accept, since
getting something is better than getting nothing. Thus, in any SPE, player 2’s
strategy must Accept any proposal of 99 or less. If the proposal is 100, then either
Accept or Reject is a NE of that subgame, since either way player 2 gets nothing.
There are, therefore, two pure SPE. In one, player 1 proposes a = 99 and player 2
plays the strategy (Accept, Accept, . . . , Accept, Reject); that is, Reject 100, Accept
otherwise. In the other pure SPE, player 1 proposes a = 100 and player 2 plays the
strategy (Accept, . . . , Accept); that is, Accept every proposal. There are also SPE
that are are mixtures over these pure SPE.
But there are many other NE. In particular, every a ∈ {0, . . . , 100} can be
supported in NE. For example, it is an equilibrium for player 1 to propose a = 50
and player 2 to play the strategy (Accept, . . . , Accept, Reject, . . . , Reject), with the
first Reject at 51; that is, player 2 rejects proposals above 50 but accepts otherwise.
In effect, player 2 says, give me half or else. And it is also an equilibrium for player
1 to propose a = 50 and player to Accept only 50, and Reject everything else. And
so on.
The ultimatum bargaining game has been heavily studied in laboratory experi-
ments. See, for example, Roth et al. (1991). The results are roughly as follows. For
subjects in the role of player 1, the modal proposal is 50, with some subjects making
higher proposals. Subjects in the role of player 2 accept 50 but frequently reject
more aggressive proposals. The behavior of subjects in the role of player 1 is broadly
consistent with maximizing monetary payoffs, given the behavior of subjects in the
role of player 2. The experimental data indicate, however, that subjects in the role
of player 2 are not maximizing monetary payoffs. For more discussion along these
lines, see Levine and Zheng (2010).
Returning to the entry deterence game of Figure 6, the strategic form was already
given in Figure 1. In the strategic form, the NE that was not SPE, namely (O,F), is
eliminated by admissibility. Similarly, if we were to write out the strategic form for
the ultimatum bargaining game, we would find that only the SPE were admissible.
These examples suggest a connection between SPE and admissibility or iterated
admissibility. A relationship exists but it is complicated. There are classes of games
in which, as in the examples above, SPE and admissibility or iterated admissibility
are equivalent. But there are also simple examples where they are not.
To see that SPE does not imply admissibility, consider the game in Figure 7.
8
The associated strategic form is in Figure 8. L weakly dominates R and hence the
L
3, 10
U
x
1 R 2, 10
L 5, 1
x'
D
2 R 1, 0
Figure 7: An SPE need not be admissible.
L R
U 3, 10 2, 10
D 5, 1 1, 0
Figure 8: The strategic form for the game in Figure 7.
unique admissible equilibrium is (D, L). But (U, R) is SPE, because there are no
proper subgames.
Perhaps more surprisingly, SPE does not imply admissibility even in games with
perfect information. Consider the game of Figure 9, which appears in van Damme
(2002). In effect, player 1 can either implement A herself or delegate the choice to
player 2. The (full) strategic form is in Figure 10. There are two pure SPE: (U A, A)
and (DA, A). But only (U A, A) is admissible: it is weakly dominant to choose A
oneself.
Conversely, admissibility, or even iterated admissibility, does not imply SPE.
Consider, for instance, the game of Figure 11. The unique SPE of this game has
player 1 choose I and both player 2 and 3 randomize 50:50. But (O, D, L) is also
a NE, and it is (iterated) admissible, for the simple reason that in this game no
strategy for any player is weakly dominated.
3.2 Weak Perfect Bayesian Equilibrium.

As already mentioned, SPE has no bite when there are no proper subgames. This
happens routinely in Bayesian games but the issue is not restricted to Bayesian
games. This issue motivates a class of refinements of which the seminal paper is
Kreps and Wilson (1982); see also Fudenberg and Tirole (1991b).
9
A 1,1
1
U
1 B 0,0
A 1,1
D
2 B 0,0
Figure 9: An SPE that is not admissible
A B
UA 1, 1 1, 1
UB 0, 0 0, 0
DA 1, 1 0, 0
DB 1, 1 0, 0
Figure 10: The strategic form for the game in Figure 9.
Given an extensive form game, define a function µ from information sets to the
interval [0, 1] with the property that for any information set h,
X
µ(x) = 1.
x∈h
If i is the player at information set h, and x ∈ h, then µ(x) is player i’s belief,
conditional on being at h, that she is at x. The pair (σ, µ) is an assessment.
Thus far, I have not said anything about how σ and µ are related. Let Prob[x|σ]
denote the probability of reaching decision node x given the strategy profile σ.
Similarly, let Prob[h|σ] denote the probability of reaching information set h given
the strategy profile σ. Thus, the decision node x is on the play path iff Prob[x|σ] > 0.
Similarly, the information set h is on the play path iff Prob[h|σ] > 0.
Definition 5. (σ, µ) is Bayes Consistent (BC) iff for any information set h on the
play path and any decision node x ∈ h,
Prob[x|σ]
µ(x) = .
Prob[h|σ]
Off the play path, Bayes Consistency imposes no restrictions on µ(x).
10
L
10,10, 0
U
1 2 R 0, 0, 10
I
L 0, 0, 10
O
D
3 R 10, 10, 0
4, 6, 6
Figure 11: An admissible NE that is not subgame perfect.
Turning now to σ, say that (σ, µ) is sequentially rational (SR) at information

set h, if, in the game fragment comprising h and all successor nodes and terminal
nodes, and all associated information sets, actions, and payoffs, no player can get
higher expected payoff by deviating at h or at any successor information set, given
the strategies of the other players. If there are two or more decision nodes in an
initial information set, then we will have to use µ to compute expected payoffs from
that point in the game looking forward. This is why µ is critical. A formal definition
of SR is slightly fussy; see Fudenberg and Tirole (1991a).
Theorem 2. σ is a NE iff there is a µ such that (σ, µ) is (a) BC and (b) SR at

every information set on the play path.
I omit the proof, which is similar to that of Theorem 1.
Definition 6. (σ, µ) is a Weak Perfect Bayesian Equilibrium (WPBE) iff it is (a)

BC and (b) SR at every information set.
In view of Theorem 2, WPBE strengthens NE by requiring that σ be SR at

every information set and not just at every information set on the play path. Given
a NE σ, say that σ is supported as a WPBE iff there exists a µ such that (σ, µ)
is WPBE. It will typically be the case that if some information sets are not on the
play path then the supporting µ is not unique. This is illustrated in some of the
examples below.
Think of WPBE as follows. Given a game in extensive form, compute the set
of NE. For a given NE σ, check whether it can be rationalized in the sense that
there is a µ, giving player beliefs at the various information sets, such that σ makes
sense (is sequentially rational) given µ. Bayes Consistency is a minimal condition
for µ to makes sense. The spirit of WPBE is that if σ cannot be supported as a
11
WPBE, then it is implausible. If σ can be supported as a WPBE, then it may still
be implausible; in particular, it may be that µ is strange, as some of the examples
below will illustrate.
The relationship between SPE and WPBE is messy. Both generalize backward
induction to games without perfect information, and both coincide with backward
induction in games with perfect information. But in games of imperfect information,
they are different.
Consider first the game in Figure 12. Here, there is a pure NE (O, R). This is
L
5, 2
U
x
1 R 1, 1
I
L 0, 3
x'
O D
2 R 2, 0
4, 10
Figure 12: SPE does not imply WPBE
SPE for the irritating reason that there are no proper subgames. But it is clearly
odd because player 2 always does better playing L rather than R at her information
set, regardless of whether she is at x or x0 . Because of this, for any µ, (O, R) is
not SR at 2’s information set. Only the NE (U, L) is supported as WPBE in this
game. In this WPBE, BC requires that since 2’s information set is on the play path,
µ(x) = 1.
Note that in the game in Figure 12, strategy L is weakly dominated by R. So
it is tempting to conjecture that WPBE is related to admissibility. In some sense
it is, but the game of Figure 9 serves as a reminder that the situation is subtle. In
that game, (DA, A) can be supported as a WPBE (trivally, since the game is one
of perfect information), but it is not admissible.
Now consider the game in Figure 13. The unique SPE is (IU, L). In particular,
in the proper subgame, U strictly dominates D, so the unique equilibrium in the
subgame is (U, L). But if player 1 chooses O, then BC does not require µ(x) = 1, as
strange as this may seem. (OU, R) is supported as a WPBE in which µ(x) = 0 (or,
more generally, µ(x) ≤ 1/2). In effect, player 2 is threatening player 1 with strange
beliefs: she is saying, “play O (which gives me my preferred payoff of 10) or else if
you enter, I will assume that you have played D, even though this contradicts the
12
L 4, 2
U
x
1 1 R 2, 1
I
L 0, 1
x'
O D
3, 10 2 R 1, 2
Figure 13: WPBE does not imply SPE.
stated strategy OU , and respond with R.”
3.3 Perfect Bayesian Equilibrium

As illustrated by the (OU, R) NE of Figure 13, implausible NE can be supported as
WPBE with implausible µ. This motivates further restrictions on µ. The restrictions
required for PBE have the flavor of, “µ should reflect Bayesian updating as much
as possible (even at information sets off the play path).” Rather than formalize
this, which is somewhat messy, let me note that if (σ, µ) is a PBE, then it induces
a WPBE in every proper subgame. In particular, if σ can be supported as a PBE,
then it is an SPE. This is enough to knock out the (OU, R) equilibrium in Figure
13.
3.4 Sequential Equilibrium

Historically, sequential equilibrium, introduced in Kreps and Wilson (1982), pre-
dates WPBE and PBE. Conceptually, however, it is a strengthening of PBE.
Definition 7. Given an assessment (σ, µ), µ is consistent iff there exists a sequence
of assessments {(σt , µt )} such that (a) limt→∞ (σt , µt ) = (σ, µ), (b) for every t, σt
is fully mixed, and (c) for every information set h and every x ∈ h,
Prob[x|σt ]
µt (x) = .
Prob[h|σt ]
Note that the definition of µt (x) makes sense because σt is fully mixed and hence
Prob[h|σt ] > 0: every information set is on the play path. Note also that consistency
does not require that one consider every possible {σt }; it only requires that there
be some such sequence {σt }.
13
For motivation, recall that, given σ, ordinary probability calculations pin down µ
at information sets on the play path. The question is how to specify µ at information
sets off the play path. Under Bayes consistency, the answer is: anyway you want.
Under consistency, one must justify µ off the play path as being the limit of the
conditional probabilities that would be generated if players trembled.
Definition 8. (σ, µ) is a sequential equilibrium (SE) iff it is (a) consistent and (b)
SR at every information set.
An SE always exists. One can show that if (σ, µ) is an SE, then it is a PBE, and
hence is a WPBE and that σ is an SPE. Let me make a few comments.
Consistency implies Bayes consistency and hence a consistent µ is pinned down
by σ at information sets on the play path; for such information sets, you don’t need
to check sequences of fully mixed σt . It is only for information sets off the play
path that consistency needs to be checked. One implication is that if an NE has
the property that all information sets are on the play path, then that NE can be
supported as an SE. In particular, in a simultaneous move game, any NE can be
supported as an SE.
The definition of consistency looks a bit like the definition of THP, and I even
mentioned trembling by way of motivation, but consistency does not assume or
imply that σti is an approximate best response to σt−i : there is no claim that σ can
be approximated by a sequence of ε-perfect equilibria. Because of this, SE does
not imply admissibility. For example, consider again the game of Figure 7. (D, R)
can be supported as an SE (trivially, since player 2’s information set is on the play
path), but it is not admissible. Similarly, in the game of Figure 9, (DA, A) can be
supported as an SE, but it is not admissible. Conversely, since THP does not imply
SPE, it does not imply SE either.
On the other hand, properness does imply SE. As noted in Section 2.4, if σ is
proper in a strategic form then it is supported as an SE in every extensive form
that generates that (reduced) strategic form; see van Damme (1984). The converse
is not true, since properness implies admissibility, while SE does not.
Last, SE is stronger than PBE in general games, but the precise way in which
this is true is subtle and I will not pursue it; see Fudenberg and Tirole (1991a) for
a discussion. In practice it is often the case that either WPBE or SPE is strong
enough or that something even stronger than SE is needed to eliminate implausible
equilibria.
3.5 Trembling Hand Perfection in the Agent Normal Form

SE is closely related to another solution concept, trembling hand perfection in the
agent normal form (THPANF), introduced, along with THP, in Selten (1975). I
mention it here briefly for completeness; think of this section as parenthetical. In
applied work, game theorists almost invariably use SE or PBE rather than THPANF.
14
Under THPANF, THP is applied to a modified strategic form in which, in ef-
fect, there is a different player for each information set. One can show that if an
equilibrium σ is THPANF, then it can be supported as a SE. Moreover, THPANF
rejects (U, R) in the game of Figure 7. Thus, THPANF is stronger than SE. But SE
and THPANF are equivalent for generic terminal node payoffs, a condition violated
in Figure 7. And SE is easier to compute (even though checking consistency can be
difficult), which is one of the reasons it, or PBE, is much more widely used.
Based on the game of Figure 7, it is tempting to conjecture that THPANF implies
admissibility. This is false. The game of Figure 9 is once again a counterexample:
the inadmissible equilibrium (DA, A) is THPANF. Informally, the reason is that,
under THPANF, player 1 is just as worried about his own trembling at his second
information set as he is about player 2 trembling. In summary, THP does not imply
THPANF (since a THP does not have to be subgame perfect) and THPANF does
not imply THP (since a THPANF does not have to be admissible).
For given ε > 0, a strategy profile that is an ε-perfect equilibrium of the modified
strategic form can be very far from any equilibrium of the original game. For
example, consider an extensive form game of perfect information with only one
player, who can choose either O immediately, for a payoff of 1000, or I, in which
case the player faces a sequence of T choices between actions A and B. If he chooses
correctly at every node, then he gets a payoff of 1001; otherwise he gets 0. There is
perfect information, so the player knows which sequence of A and B to play. The
unique Nash equilibrium, and hence the unique THPANF equilibrium, is for the
player to choose I followed by the correct sequence of A and B. But for any given
ε > 0, no matter how small, if T is sufficiently large, the ε-perfect equilibrium in
the modified strategic form has the player choose O. Similar phenomena can also
arise for THP in strategic form games.
3.6 The Cho-Kreps Intuitive Criterion

The SE belief restriction called consistency depends on the strategy profile σ and
on the extensive form of the game, but not on the game’s payoffs. Consistency does
not ask whether, given the game’s payoffs, beliefs make sense off the play path.
As illustration of why payoffs might affect beliefs off the play path, consider the
game of Figure 14, taken from Cho and Kreps (1987). Player 1 can either be Weak
or Strong. Her action is her choice of breakfast, Beer or Quiche. Player 2 observes
player 1’s breakfast, but not her type, and chooses whether to Fight player 1 or
Run away. Other things being equal, player 1 would prefer Quiche if Weak and
Beer if Strong, but she would also like player 2 to Run. She thus potentially faces
a tradeoff between having her preferred breakfast and inducing player 2 to Run.
Player 2 would like to Fight a Weak player 1 but would like to Run from a Strong
player 1. Less whimsically, one can think of this as a model of entry deterrence with
asymmetric information. Player 1 is the incumbent firm in some market and has
either low production costs or high production costs. Player 2, the potential entrant,
15
0,1 Fight Weak Fight 1,1
P1
Beer Quiche
x y
0.1
2,0 Run Run 3,0
1,0 Fight 0.9 Fight 0,0
Beer P1 Quiche
P2 Strong P2 Run
3,1 Run 2,1
Figure 14: The Beer-Quiche Game.
cannot observe player 1’s production costs directly, but he can observe player 1’s
market behavior. Beer corresponds to player 1 charging a low price while Quiche
corresponds to player 1 charging a high price.
This game has two pure strategy pooling NE: ((Beer, Beer), (Run, Fight)) and
((Quiche, Quiche), (Fight, Run)), where the first strategy profile, for example, is
read, Beer if Weak, Beer if Strong, Run if Beer, Fight if Quiche. There is no
separating NE in this particular game. The first pooling NE can be supported as a
WPBE provided µ(y) ≥ 1/2 (by Bayes Consistency, µ(x) = 1). The second can be
supported as a WPBE provided µ(x) ≥ 1/2 (by Bayes Consistency, µ(y) = 1). One
can verify easily, although I will not do so, that all of these WPBE are, in fact, SE.
Of these two classes of WPBE, the (Quiche, Quiche) WPBE are arguably im-
plausible for the following reason. The Weak type player 1 has no possible reason
to deviate from Quiche: she is making 3 in equilibrium by eating Quiche and could
get at most 2 by drinking Beer, even if she somehow convinced player 2 to Run in
response to Beer. Quiche is said to be equilibrium dominated for the Weak type in
the (Quiche, Quiche) NE. This suggests that in any (Quiche, Quiche) WPBE, player
2 should assign zero probability to player 1 being Weak if she sees player 1 drink
Beer: µ(x) = 0. But µ(x) = 0 is not compatible with any of the (Quiche, Quiche)
WPBE: if µ(x) = 0, then sequential rationality forces player 2 to Run after seeing
Beer, and if player 2 Runs after seeing player 1 drink Beer, then a Strong player 1
will deviate to drinking Beer.
16
In contrast, the (Beer, Beer) equilibrium is free from this problem. Quiche is
equilibrium dominated for Strong in the (Beer, Beer) equilibrium, suggesting that
player 2 should assign zero probability to player 1 being Strong after seeing player
1 eat Quiche: µ(y) = 1. And there is, in fact, a (Beer, Beer) WPBE with µ(y) = 1.
The Intuitive Criterion of Cho and Kreps (1987) (CKIC) formalizes the ideas
just illustrated. In contrast to the other refinements discussed here, CKIC applies
only to a special class of games, of which the game of Figure 14 was an example. In
these games, there are two players. Nature moves first and chooses a type for player
1 (e.g., Weak or Strong). Player 1 observes her type and chooses an action a ∈ A
as a function of her type (e.g., Beer or Quiche). For simplicity, assume that, as in
Figure 14, the set of allowed actions A is the same for all types; Cho and Kreps
(1987) allows the set of allowed actions to vary with type. Player 2 observes player
1’s action, but not her type, and chooses a response. The game is then over and
players get their payoffs. These games all have the property that the set of SE and
the set of WPBE coincide.
Definition 9. Given a NE σ of a game of the above form and an action a for player
1 that is not played under σ, a is equilibrium dominated for type θ iff θ’s equilibrium
payoff is higher than her payoff from deviating to action a, for any response to action
a that is sequentially rational for player 2 given beliefs µ, for any µ.
Definition 10. A NE σ of a game of the above form fails CKIC iff there is an
action a for player 1 that is not played under σ and for which the following is true.
1. There is at least one type for whom a is equilibrium dominated.
2. There is at least one other type whose payoff from deviating to action a is
higher than her equilibrium payoff, for any response by player 2 to action a
that is sequentially rational given µ, for any µ that, at the a information set,
puts probability zero on all types for which action a is equilibrium dominated.
Note that for CKIC to eliminate equilibria, it must that there is an unused action
that is equilibrium dominated for some types but not for others. If no such unused
action exists, because all unused actions are either equilibrium dominated for all
types or for no types (or because no action is unused), then the equilibrium satisfies
CKIC vacuously.
In the (Quiche, Quiche) Nash equilibrium of the example, action a was Beer. As
already noted, Beer is equilibrium dominated for Weak. If µ(x) = 0, then sequential
rationality requ ires that player 2 Run after seeing player 1 drink Beer, which gives
a Strong player 1 a payoff of 3 from drinking Beer, which is more than the 2 she
gets in equilibrium by eating Quiche. Thus, the (Quiche, Quiche) NE fails CKIC.
Conversely, the (Beer, Beer) Nash equilibrium passes CKIC. Quiche is equilibrium
dominated for Strong but if zero probability is put on Strong after seeing Quiche,
hence µ(y) = 1, then sequential rationality forces player 2 to Fight on seeing Quiche,
17
which means that a Weak player 2 has a lower payoff from deviating to Quiche than
from sticking to the equilibrium strategy of Beer.
As defined, the property of satisfying CKIC is a refinement of NE rather than of
SE. In practice, however, CKIC is used as a refinement of SE. Thus, one computes
(even if only informally) the set of SE and then checks which of these SE satisfy
CKIC. In particular, given a set of SE characterized by a strategy profile σ and
a set of associated µ, one checks whether any of these µ have the property that,
conditional on any unused action, zero probability is assigned to types for whom
that action was equilibrium dominated. If such an µ exist, then σ satisfies CKIC, if
not, then σ fails CKIC. I illustrated this in the example above.
CKIC is the best known, and most heavily used, of a large class of related
refinements in games of this general type. For more on this, see Cho and Kreps
(1987) and some of the other surveys cited. For another classic paper in this genre,
see Banks and Sobel (1987).
References
Banks, Jeffrey and Joel Sobel. 1987. “Equilibrium Selection in Signaling Games.”
Econometrica 55(3):647–61.
Ben-Porath, E. and E. Dekel. 1992. “Signaling Future Actions and the Potential for
Self-Sacrifice.” Journal of Economic Theory 57(1):36–51.
Brandenburger, Adam. 2008. Epistemic Game Theory: An Overview. In The New

Cho, In-Koo and David Kreps. 1987. “Signaling Games and Stable Equilibria.”
Quarterly Journal of Economics 102:179–221.
Fudenberg, Drew, David K. Levine and David Kreps. 1988. “On the Robustness of
Equilibrium Refinements.” Journal of Economic Theory 44:351–380.
Fudenberg, Drew and Jean Tirole. 1991a. Game Theory. Cambridge, MA: MIT
Press.
Fudenberg, Drew and Jean Tirole. 1991b. “Perfect Bayesian and Sequential Equi-
librium.” Journal of Economic Theory 53:236–260.
Govindan, Srihari and Robert Wilson. 2008. Refinements of Nash Equilibria. In The
New Palgrave Dictionary of Economics. Palgrave.
Harris, Christopher, Philip Reny and Arthur Robson. 1995. “The Existence of
Subgame-Perfect Equilibrium in Continuous Games with Almost Perfect Infor-
mation.” Econometrica 63:507–544.
18
Jackson, Matthew, Tomas Rodriguez-Barraquer and Xu Tan. 2011. “Epsilon-
Equilibria of Perturbed Games.” Department of Economics, Stanford University.
Kandori, Michihiro, George Mailath and Rafael Rob. 1993. “Learning, Mutation,
and Long Run Equilibria in Games.” Econometrica 61(1):29–56.
Keisler, H. Jerome and Byung Soo Lee. 2011. “Common Assumption of Rationality
Preliminary Report.” Department of Mathematics, University of Toronto.
Kohlberg, Elon and Jean-François Mertens. 1986. “On the Strategic Stability of
Equilibria.” Econometrica 54(5):1003–1037.
Kreps, David and John Wilson. 1982. “Sequential Equilibria.” Econometrica 50:863–
894.
Levine, David and Jie Zheng. 2010. The Relationship of Economic Theory to Ex-
periments. In The Methods of Modern Experimental Economics, ed. Guillame
Frechette and Andrew Schotter. Oxford University Press. Washington Univer-
sity, St. Louis.
Mertens, Jean-Francois. 1989. “Stable Equilibria - A Refomulation, Part I: Defini-

tion and Basic Properties.” Mathematics of Operations Research 14:575–624.
Myerson, Roger. 1978. “Refinements of the Nash Equilibrium Concept.” The Inter-
national Journal of Game Theory 15:133–154.
Roth, A, V Prasnikar, M Okuno-Fujiwara and S Zamir. 1991. “Bargaining and Mar-

ket Behavior in Jerusalem, Ljubljana, Pittsburgh, and Tokyo: An Experimental
Study.” American Economic Review 81(5):1068–1095.
Selten, Reinhard. 1967. “Spieltheoretische Behandlung eines Oligopolmodells mit

Nachfrageträgheit.” Zeitschrift für die gesamte Staatswissenschaft 12:301–324.
Selten, Reinhard. 1975. “A Reexamination of the Perfectness Concept for Equilib-

rium Points in Extensive Games.” International Journal of Game Theory 4:25–55.
van Damme, Eric. 1984. “A Relationship Between Perfect equilibria in Extensive

Form Games and Proper Equilibria in Normal Form Games.” International Jour-
nal of Game Theory 13:1–13.
van Damme, Eric. 2002. Stability and Perfection of Nash Equilibria. 2 ed. Springer.
Zermelo, Ernst. 1913. Über eine Anwendung der Mengenlehre auf die Theorie des
Schachspiels. In Proceedings of the the Fifth Congress Mathematicians. Cambridge
University Press pp. 501–504.
19
John Nachbar
April 16, 2014
Independent Private Value Auctions

The following notes are based on the treatment in Krishna (2009); see also
Milgrom (2004). I focus on only the simplest auction environments.
Consider I ≥ 2 bidders with values independently and identically distributed on
[0, 1] (the analysis for other intervals is similar). The distribution of a bidder’ value
has cumulative distribution function F and strictly positive density f . Assume that
each bidder has an outside option of 0 and that if bidder I wins the object and pays
ti then she gets utility
vi − ti ,
where vi is bidder I’s value and ti is her payment. Values are private: although
the values of the other bidders will affect their actions and through that channel
both the probability that bidder I wins and her payment, bidder I’s payoff does
not depend directly on the values of the other bidders. At the time of bidding, only
bidder I knows vi .
Consider anonymous direct auctions. By anonymous, I mean that the auction
treats bidders symmetrically. By direct I mean that a bidder’s action is to submit a
bid, which is simply a real number. Consider only auctions in which the high bidder
wins. In the examples below, I focus on three special cases.
1. First Price Sealed Bid Auctions. Each bidder I submits a bid bi ∈ R to a

referee. The referee gives the object to the highest bidder, who pays her bid.
All other bidders get a payoff of 0. A NE (in fact, the NE) of this auction is
computed below.
2. Second Price Sealed Bid Auctions. Each bidder I submits a bid bi ∈ R to a

referee. The referee gives the object to the highest bidder, who pays the bid
of the second highest bidder. All other bidders get a payoff of 0. This is the
VCG mechanism in an auction setting. There is a truth-telling NE: bidder I
bids bi = vi . In fact, truth telling is weakly dominant. But there are other
NE. In particular, it is a NE for bidder 1 to bid 1 and for all other bidders to
bid 0.
3. All Pay Auctions. Each bidder I submits a bid bi ∈ R to a referee. The referee
gives the object to the highest bidder. EVERY bidders pays her bid. This
auction is sometimes used to model political lobbying. A NE of this auction
is computed below.
Assume that the NE is symmetric (every bidder uses the same bid function).
Henceforth I drop the I subscript. Assume that in this NE, bids are strictly in-
creasing in type. Finally, assume that the auction has the property that the interim
1
expected utility of a bidder with value v = 0, is 0. In any given NE of any given
auction, these assumptions have to be verified.
Under these assumptions, the auction is ex-post efficient. By almost the same
argument as in MyersoI-Satterthwaite, T (v), the interim expected payment by a
bidder with value v is
Z v
T (v) = Q(v)v − U (0) − Q(x) dx, (1)
0
where Q(v) is the interim probability of winning the object when of type v and U is
the interim expected utility when of type 0. By assumption, U (0) = 0. Therefore,
integrating by parts yields
Z v
T (v) = Q0 (x)x dx.1 (2)
0
Remark 1. One way to remember equation (2) is to note that the interim expected
utility of a bidder when of type v but bidding as though of type v̂ is
Q(v̂)v − T (v̂).
If both Q and T are differentiable (which we cannot, in fact, assume, because Q and
T are endogenous), then the first order condition for an interior maximum is
Q0 (v̂)v − T 0 (v̂) = 0.
If it is in fact optimal for the bidder to bid like a type v when a type v then v̂ = v,
hence,
T 0 (v) = Q0 (v)v.
Integration gives (2).
Because bids functions are symmetric and increasing in type, Q(v) is the prob-
ability that v is the highest value, which is to say, the probability that all other
values are less than v, which is
Q(v) = [F (v)]I−1 .
Q(v) can also be interpreted as the probability that the highest value of the other
bidders is less than v. Call this highest value of the other bidders Y . Then G(v) =
[F (v)]I−1 is the cumulative distribution function for Y . Let g(Y ) = G0 (Y ) be the
density for Y . Then Z v
T (v) = g(Y )Y dY.
0
1
Strictly speaking, I cannot assume that Q, which is endogenous, is differentiable. But one can
show that the argument that leads to equation (1) implies that Q is weakly increasing. This should
be intuitive: higher v should lead to at least a weakly higher probability of winning the auction.
If Q is weakly increasing then it is differentiable almost everywhere, which is enough for present
purposes.
2
Note that
Z v Z v
g(Y ) 1
E[Y |Y < v] = Y dY = Y g(Y ) dY.
0 Prob[Y < v] Prob[Y < v] 0
Therefore
T (v) = Prob[Y < v]E[Y |Y < v]. (3)
Again, this is the interim expected payment by one bidder. The overall ex-ante
expected payment for I bidders, under the above assumptions, is
I[E[Prob[Y < v]E[Y |Y < v]]]. (4)
Equation (3) has a number of very powerful implications. First, since it gives the
interim expected payment in a form that depends only on the value distribution, and
not on any details of the auction, it implies that, in this setting, any two symmetric
equilibria of any two auctions will generate the same expected revenue if in both
equilibria (a) bids are strictly increasing in type and (b) the interim expected utility
of a bidder of type v = 0 is 0. This is true even if the auctions appear to be very
different. As may be evident by inspection, the right-hand side of (3) is the expected
revenue from the truth-telling NE of the second price auction. What (3) is saying,
then, is that a wide variety of auctions have NE with the same expected revenue as
the truth-telling NE of the second price auction.
Moreover, the ex-ante expected revenue in equation (4) is the highest ex-ante
expected revenue that can be achieved in any ex-post efficient NE that respects in-
terim individual rationality, across all possible auctions in this class of environments.
As in MyersoI-Satterthwaite, (1) implies that the seller’s ex-ante expected revenue
is entirely determined by Q and U (0). But Q is pinned down by ex-post efficiency.
And interim individual rationality holds iff U (0) ≥ 0. Since T (v) is decreasing,
U (0), the seller’s expected revenue is maximized when, as here, U (0) is set to its
minimum value, namely 0.
Example 1. Suppose that the v are uniformly distributed on [0, 1]. Then G(v) =
v I−1 , g(v) = (I − 1)v I−2 and
Z v
T (v) = (I − 1)Y I−2 Y dY
0
Z v
= (I − 1) Y I−1 dY
0
I −1 I
= v
I
It follows that ex ante expected revenue to the seller is
Z 1
I −1 I
I[E[T (v)]] = I v dv
0 I
I −1
= .
I +1
3
In particular, if I = 2 then ex ante expected revenue is 1/3. As I increases,
expected revenue converges to 1. This should make intuitive sense, since with large
I there is a high probability that there are many bidders with values close to 1.
Equation (3) can be used to compute NE in auctions. I illustrate this with
examples.
Example 2. First price sealed bid auction. Since the high bidder wins and pays her
own bid β(v), it must be that,
T (v) = Prob[Y < v]β(v).
Hence from (3),

β(v) = E[Y |Y < v].
Note that this is increasing in v and it is always less than v: bidders shave their
bids, trading off a lower probability of winning against lower payments when they
win. One can verify that there is, indeed, a NE with this bid function and it should
be obvious that in this NE, interim expected utility when v = 0 is 0.
In the particular case in which v is uniformly distributed on [0, 1],
Z v
1
β(v) = I−1 (I − 1)Y I−2 Y dY
v 0
I − 1 v I−1
Z
= I−1 Y dY
v
0
I −1 1 I
= I−1 v
v I
I −1
= v.
I
If I = 2 then β(v) = v/2. As I increases, the bid converges to v: competition from
other bidders becomes keener and therefore shaving becomes less advantageous,
although the bid is always less than v.
One can confirm by direct calculation that the ex-ante expected revenue to the
seller is as calculated in Example 1.
Example 3. All pay auction. Since each bidder always pays her bid regardless of
whether she wins,
T (v) = β(v).
Hence from (3),
β(v) = Prob[Y < v]E[Y |Y < v].
Again, note that this is increasing in v. Again, one can verify that there is, indeed, a
NE with this bid function and it should be obvious that in this NE, interim expected
utility when v = 0 is 0.
4
In the particular case in which v is uniformly distributed on [0, 1],
Z v
β(v) = (I − 1)Y I−2 Y dY
0
Z v
= (I − 1) Y I−1 dY
0
1 I
= (I − 1) v
I
I −1 I
= v .
I
If I = 2 then β(v) = v 2 /2. As I increases, the bid function becomes increasingly
convex, with bids approximately 0 unless v is very close to 1. Intuitively, a bidder
(although risk neutral) is willing to risk a noI-trivial bid only when the probability
of winning is sufficiently high.
Again, one can confirm by direct calculation that the ex-ante expected revenue
to the seller is as calculated in Example 1.
As a final remark, this analysis becomes substantially more complicated if any
of the assumptions are relaxed. It can, for example, be difficult if not impossible
to solve for the equilibrium of a first price auction if the values are not identically
distributed.
References
Krishna, V. (2009): Auction Theory. Academic Press, second edn.
Milgrom, P. (2004): Putting Auction Theory To Work. Cambridge University

Press.
5
John Nachbar
May 5, 2017
The Myerson-Satterthwaite Theorem1

1 Introduction.
The Myerson-Satterthwaite Theorem (MS) is an impossibility result for bi-
lateral bargaining. The original cite is Myerson and Satterthwaite (1983).
The proof here most closely follows Krishna (2009) and Milgrom (2004).
One person, the seller, owns one unit of an indivisible object. The seller
values the object at c, which I refer to as the seller’s type. The other person,
the buyer, values the object at v, which I refer to as the buyer’s type. Each
knows her own type but not the other person’s. More generally, one can
interpret the environment as one in which two people are bargaining over
a proposed change in the status quo, with the “buyer” gaining v from the
proposed change and the “seller” losing c.
MS says that, subject to restrictions on the joint distribution of c and v,
one cannot simultaneously have ex post efficiency, expected budget balance,
and interim individual rationality. Ex post efficiency means that trade oc-
curs when v > c (i.e., the buyer values the object more than the seller) but
not when v < c. Expected budget balance means that while a third party
(e.g., a judge) is allowed to provide a net subsidy for some type profiles (c, v)
and collect a net tax for others, the third party’s net transfer must be zero
in expectation over (c, v). A stronger requirement would be ex post budget
balance, which requires zero net transfers for any (c, v). Interim individual
rationality means that the buyer and seller can opt out of the bargaining
game after learning own type (e.g., after seeing the good to be traded).
Put differently, MS says that if one demands expected budget balance
and interim individual rationality, then bargaining cannot be ex post effi-
cient. This does not say that the outcome will be inefficient for all values of
c and v, only that it will be inefficient for some values. For many standard
bargaining models, the expected inefficiency can be large, on the order of a
loss of 25% of expected surplus. One can construct games with equilibria
that satisfy ex post efficiency and expected budget balance, and one can
construct games with equilibria that satisfy ex post efficiency and interim
cbna.
1
This work is licensed under the Creative Commons Attribution-
NonCommercial-ShareAlike 4.0 License.
1
individual rationality. So it is the combination of all three criteria that
causes a problem.
It is not surprising that bargaining is sometimes inefficient. Labor strikes,
lawsuits, and even some wars are examples of bargaining failure. MS is im-
portant because it pinpoints information asymmetry as a robust cause of
inefficiency. If instead the type profile (c, v) were known by the partici-
pants, then standard models of bargaining (e.g., Nash’s demand game; the
Rubinstein-Stähl alternating offers game) typically have equilibria that are
ex post efficient. There may also be inefficient equilibria, and in practice
inefficiency might arise if either side miscalculates. But in such cases an
arbitrator or other third party might be able to force efficient trade at a
price of, say, (c + v)/2. But MS says that if c and v are private information,
then ex post efficiency fails even with intervention by a third party.
The negative conclusion of MS is at least partly a small numbers prob-
lem. There are games, in particular auctions, for which the expected in-
efficiency, while still positive, becomes vanishingly small as the number of
participants grows large. The classic cite is Rustichini, Satterthwaite and
Williams (1994); for a more general treatment, see Cripps and Swinkels
(2006)
2 Generalized Bargaining Games.

2.1 The information structure.
The seller’s type c (cost) and the buyer’s type v (value) are jointly dis-
tributed on R2+ . Let F denote the joint distribution and let supp(F ) denote
the support of F , meaning the smallest closed subset of R2+ that gets prob-
ability 1. Let C be the set of c that appear in supp(F ),
C = {c : ∃ v s.t. (c, v) ∈ supp(F )}.
Similarly, define,
V = {v : ∃ c s.t. (c, v) ∈ supp(F )}.
• The Myerson-Satterthwaite Theorem (MS) assumes that F is inde-
pendent, in which case supp(F ) = C × V . McAfee and Reny (1992)
shows that MS fails if independence is relaxed. Section 5.1 provides
an example.
• MS assumes that C and V are non-degenerate intervals. Section
5.2 shows by example that MS can fail if the interval assumption is
dropped.
2
• MS assumes that the marginal distributions of c and v have densities;
in particular, the probability of any given c or v is zero. This can be
relaxed somewhat. See Section 5.3
Nature moves first and chooses (c, v). The seller observes c but not v,
and that the buyer observes v but not c. The joint distribution F is part of
the description of the game; it is commonly known by both players.
Henceforth, hold this information structure fixed.
2.2 A generalized bargaining game form.

The seller, having observed c, chooses an action as ∈ As . The set As is
non-empty but otherwise unrestricted. In particular, as could itself be a
complicated extensive form strategy. Similarly, the buyer, having observed
v, chooses an action ab ∈ Ab .
A pure strategy for the seller is then a function ρs : C → As . A pure
strategy for the buyer is a function ρb : V → Ab .2
2.3 Utility.
The action profile (as , ab ) determines outcomes via the functions π : As ×
Ab → [0, 1], τ s : As × Ab → R, and τ b : As × Ab → R.
π(as , ab ) = q is the probability that the buyer gets the object. In most
bargaining games of interest, q is either 0 or 1.
τ s (as , ab ) = ts is the payment received by the seller while τ b (as , ab ) = tb
is the payment made by the buyer. I do not assume that ts = tb . If ts 6= tb ,
then some third party (the government) must make up the difference by
taxing or subsidizing. This is why I refer to this game as a “generalized”
bargaining game; a standard bargaining game would require ts = tb (which
I am calling ex post budget balance).
Finally, utilities depend on types and on outcomes.3 If the type profile
is (c, v) and the outcome profile is (q, ts , tb ), then the expected utility to the
seller (net of the cost c) is
(1 − q)c + ts − c = −qc + ts ,
and the expected utility to the buyer is
qv − tb .
2
In most of my game theory notes, I use the notation s for a pure strategy. I am using
ρ for a pure strategy here to avoid confusion with “s” for “seller”.
3
In game theory, utilities are usually called payoffs; I use the term utility here to avoid
confusion with the payments, ts and tb .
3
Note that preferences are risk neutral.
Remark 1. The utility specification seems to rule out games in which a
payment is made only when the object is transferred. This would be unfor-
tunate, because many standard bargaining games have this feature. In fact,
such games have not been ruled out. One merely has to reinterpret ts and
tb as expected payments. For example, suppose that the seller receives the
payment t̂s if the object is transferred, but no payment if it is not. In ex-
pectation, therefore, she receives q t̂s . Let ts = q t̂s . This trick works because
of risk neutrality.
2.4 Nash equilibrium.

A pure strategy profile (ρs , ρb ) is a Nash equilibrium (NE) of the generalized
bargaining game iff, for each c, ρs (c) maximizes the seller’s expected utility
given ρb and, for each v, ρb (v) maximizes the buyer’s expected utility given
ρs .4 It is customary to call a NE in this sort of information setting a Bayesian
Nash Equilibrium.5
2.5 Notation for some expectations.

Fix a strategy profile (ρs , ρb ). From the perspective of a seller with cost c,
the probability of trade is given by,
Qs (c) = Ev [π(ρs (c), ρb (v))|c],
where the expectation is over v because the seller does not know the buyer’s
value. Similarly, from the perspective of a seller with cost c, the expected
payment to the seller is,
T s (c) = Ev [τ s (ρs (c), ρb (v))|c].
Thus, given the strategy profile, a seller with cost c has an expected
utility of,
U s (c) = −Qs (c)c + T s (c).
4
I can handle mixed strategy equilibria as well, but to avoid some non-essential issues,
I do not do so.
5
Strictly speaking, the equilibrium concept here is somewhat stronger than standard
Nash equilibrium, because under standard Nash equilibrium, ρs , for example, is required
to be optimal only ex ante: given ρb , ρs maximizes the seller’s expected utility, where the
expectation is over both c and v. This is equivalent to requiring that ρs (c) maximize the
seller’s expected utility, where the expectation is over v, for every c in a subset of C of
probability one. If the marginal distribution over C has a density, so that any given c ∈ C
has probability zero, then ρs could satisfy the standard NE optimization condition even
if ρs (c) is not optimal for some c.
4
The seller’s cost influences her expected utility through three channels.
There is a direct channel: c is the value of the object to the seller. There are
also two indirect channels. First, c determines the seller’s action as via the
strategy ρs . Second, the expectations are conditional on c. For general F ,
it is possible that, for example, a seller with high cost infers that the buyer
likewise has a high value, and this affects the seller’s posterior over what
action the buyer will take. However, MS assumes that F is independent,
and hence assumes this second indirect channel away. Section 5.1 discusses
the independence assumption further.
Similarly, for the buyer, define,
Qb (v) = Ec [π(ρs (c), ρb (v))|v]
T b (v) = Ec [τ b (ρs (c), ρb (v))|v]
U b (v) = Qb (v)v − T b (v).
2.6 Efficiency, Budget Balance, and Individual Rationality.

Definition 1. A Nash equilibrium of a generalized bargaining game is ex
post efficient iff, for any (c, v) ∈ supp(F ) and associated (q, ts , tb ), q = 1 if
v > c and q = 0 if v < c.
Definition 2. A Nash equilibrium of a generalized bargaining game is ex
post budget balanced iff, for any (c, v) ∈ supp(F ) and associated (q, ts , tb ),
ts = tb . A Nash equilibrium of a generalized bargaining game is expected
budget balanced iff the equilibrium expected subsidy, Ec,v [T s (c) − T b (v)], is
zero.
Definition 3. A Nash equilibrium of a generalized bargaining game is in-
terim individually rational iff, for any (c, v) ∈ supp(F ), U s (c) ≥ 0 and
U b (v) ≥ 0.
Remark 2. A sufficient condition for individual rationality in Nash equi-
librium is that the generalized bargaining game contains an action quit,
available to either participant, with the property that if either quits, then
the seller retains the object and there is no payment.
3 The Myerson-Satterthwaite Theorem: Special

Case.
The following result is the Myerson-Satterthwaite theorem for the special
case in which C = V = [0, 1]. Section 4 provides a statement of a more
5
general form of the theorem.
Theorem 1 (Baby Myerson-Satterthwaite Theorem). Suppose that F is

independent, C = V = [0, 1], and the marginal distributions have densities.
Then there is no NE of any generalized bargaining game that is ex post
efficient, expected budget balanced, and individually rational.
The proof of Theorem 1 involves the following generalized bargaining

game, which I call the VCG game in honor of Vickrey (1962), Clarke (1971),
and Groves (1973), who considered games closely related to this one.
In the VCG game, the seller’s action is to announce a type c∗ ∈ [0, 1],
which could be different from the seller’s true type, and the buyer’s action
is to announce a type v ∗ ∈ [0, 1], which could be different from the buyer’s
true type. Trade occurs iff v ∗ ≥ c∗ . If there is trade then the buyer pays c∗
and the seller receives v ∗ . If there is no trade then there are no payments.
There is a NE of the VCG game in which participants tell the truth:
ρ (c) = c and ρb (v) = v. Indeed, truth telling is weakly dominant in this
s
game.6 It is immediate that this truth-telling NE is ex post efficient.7 And

it is also immediate that it is individually rational, since the buyer always
gets either 0 (if there is no trade, which happens iff v < c) or v − c ≥ 0 (if
there is trade, which happens iff v ≥ c), and similarly for the seller.
But the VCG game isn’t even close to being expected budget balanced.
In fact, the subsidy equals v − c > 0 whenever v > c, which occurs with pos-
itive probability, and the subsidy equals zero whenever v ≤ c: the expected
subsidy equals the expected surplus from efficient trade!
Remarkably, the truth-telling equilibrium of the VCG game is the cheap-
est way to get ex post efficiency and interim individual rationality.
Theorem 2. Suppose that F is independent, C = V = [0, 1], and the

marginal distributions have densities. Then the truth-telling equilibrium of
the VCG game has the smallest expected subsidy of any ex post efficient and
interim individually rational NE of any generalized bargaining game.
6
Briefly, consider the seller. Given c, suppose the seller reports c∗ > c. The only
circumstance in which lying in this way changes the seller’s utility is when v ∗ ∈ (c, c∗ ].
In this case, there is trade if the seller tells the truth, in which case the seller gets a net
utility of v ∗ − c > 0, while there is no trade if the seller reports c∗ , in which case the seller
gets a utility of 0. So lying yields a lower utility than telling the truth. The other cases
are similar.
7
There are other NE, not in weakly dominant strategies, that are not ex post efficient.
For example, if c and v are both distributed on [0, 1], then it is a NE of the VCG game
for the buyer to announce v ∗ = 0 and the seller to announce c∗ = 1 regardless of actual
type; in this NE, there is never any trade.
6
Proof of Theorem 1. Since the VCG game has a strictly positive expected
subsidy, Theorem 2 implies that any NE of any generalized bargaining game
is not expected budget balanced if it is ex post efficient and interim individ-
ually rational.
It remains to prove Theorem 2. The key fact driving Theorem 2 is the

following.8
Theorem 3 (The Integral Conditions). Suppose that F is independent and

that C = V = [0, 1]. Consider any strategy profile (ρs , ρb ). If ρs is optimal
for the seller, given ρb , then,
Z 1
s s
U (c) = U (1) + Qs (x) dx. (1)
c
Similarly, if ρb is optimal for the buyer, given ρs , then,

Z v
U b (v) = U b (0) + Qb (x) dx. (2)
0
Proof. If ρs is optimal, then no deviation by the seller, say to ρ̂s , can be

profitable. One particular deviation is for the seller, when her actual cost is
c∗ , to play as if her cost were c: she plays ρs (c) = as instead of ρs (c∗ ) = a∗s .
The probability of trade, from the perspective of a seller who has cost
c∗ but plays as if she has cost c is,
Q̃s (c, c∗ ) = Ev [π(ρs (c), ρb (v))|c∗ ].
The assumption that F is independent implies,
Q̃s (c, c∗ ) = Qs (c).
Similarly, the expected payment to the seller, from the perspective of a seller
who has cost c∗ but plays as if she has cost c, is,
T̃ s (c, c∗ ) = Ev [τ s (ρs (c), ρb (v))|c∗ ].
Independence implies,
T̃ s (c, c∗ ) = T s (c).
8
As a technical aside, note that Theorem 3 does not assume that the marginal distri-
butions have densities.
7
Therefore, the seller’s expected utility, from the perspective of a seller who
has cost c∗ but plays as if he has cost c, is,
Ũ s (c, c∗ ) = −Q̃s (c, c∗ )c∗ + T̃ s (c, c∗ ),
which, by independence, can be rewritten,
Ũ s (c, c∗ ) = −Qs (c)c∗ + T s (c).
Note that U (c∗ ) = Ũ s (c∗ , c∗ ). In Section 5.1, I discuss what happens if

independence is violated.
If ρs is optimal then, for any c and c∗ , it must be that U s (c∗ ) ≥ Ũ s (c, c∗ ),
or,
U s (c∗ ) = −Qs (c∗ )c∗ + T s (c∗ ) ≥ −Qs (c)c∗ + T s (c), (3)
and conversely U s (c) ≥ Ũ s (c∗ , c), or,
U s (c) = −Qs (c)c + T s (c) ≥ −Qs (c∗ )c + T s (c∗ ). (4)
These inequalities are called incentive compatibility (IC) conditions for the
seller. The seller’s IC conditions imply,
U s (c∗ ) − U s (c) ≥ −Qs (c)c∗ + T s (c) − U s (c)

= −Qs (c)c∗ + T s (c) − [−Qs (c)c + T s (c)]
= −Qs (c)(c∗ − c).
Similarly,
U s (c∗ ) − U s (c) ≤ U s (c∗ ) − [−Qs (c∗ )c + T s (c∗ )]

= −Qs (c∗ )c∗ + T s (c∗ ) − [−Qs (c∗ )c + T s (c∗ )]
= −Qs (c∗ )(c∗ − c).
Combining,
− Qs (c)(c∗ − c) ≤ U s (c∗ ) − U s (c) ≤ −Qs (c∗ )(c∗ − c). (5)
This implies that −Qs is weakly increasing in c, hence Qs is weakly decreas-

ing. This should be intuitive: as c increases, it is increasingly unlikely that
trade is worthwhile, so the probability of trade should decrease.
8
Since −Qs is weakly increasing, −Qs is (Riemann) integrable (Rudin
(1976), Theorem 6.9); Inequality (5) implies that the integral satisfies,9
Z 1
s s
U (1) − U (c) = −Qs (x) dx. (6)
c
Rearranging Equation (6) gives Equation (1). The argument for Equation
(2) is similar and I omit it.
Remark 3. Assume for the moment that Qs and T s are differentiable. As

in the proof of Theorem 3, the expected utility of a seller who has cost c∗
but acts as if her cost were c is,
Ũ s (c, c∗ ) = −Qs (c)c∗ + T s (c).
(Again, this notation assumes independence.) If ρs is indeed optimal, then

Ũ s (c, c∗ ) is maximized when the seller with cost c∗ actually acts as if her
cost were c∗ (so that the seller plays a∗s = ρs (c∗ )), in which case the ex-
pected utility is Ũ s (c∗ , c∗ ) = U s (c∗ ). U s is thus the value function for this
maximization problem. By the Envelope Theorem, DU s (c∗ ) = D2 Ũ s (c∗ , c∗ )
(where D2 is the derivative with respect to the second argument), hence
DU s (c∗ ) = −Qs (c∗ ). (7)
Integrating this gives Equation (6).

One cannot, in fact, simply assume that Qs and T s are differentiable,
since they depend on the strategies, which are not required to be differen-
tiable. Milgrom and Segal (2002) shows that, despite this issue, a generalized
Envelope Theorem argument still applies and establishes Equation (6).
Equation (1) and Equation (2) are often called the Integral Conditions.
Substituting U s (c) = −Qs (c)c + T s (c) and U b (v) = Qb (v)v − T b (v) into the
9
Explicitly, recall that a function is integrable iff its upper and lower integrals are equal
(and finite). Consider any finite partition of [c, 1] determined by c = x1 < · · · < xK = 1.
Consider any xk < xk+1 Since −Qs is weakly increasing, the maximum value of −Qs on
this interval is −Q(xk+1 ), so that the associated rectangle has area −Qs (xk+1 )(xk+1 −xk ).
By (5), this area is bounded below by U s (xk+1 ) − U s (xk ). Summing over such areas, all
but two of the U S (xk ) cancel, yielding a total lower bound of U s (1) − U s (c). I get the
same lower bound, namely U s (1) − U s (c), for every partition. Hence the upper integral
is bounded below by U s (1) − U s (c). By an analogous argument, the lower integral is
bounded above by U s (1) − U s (c), which yields Equation (6).
9
Integral Conditions and rearranging yields,
Z 1
T s (c) = U s (1) + Qs (c)c + Qs (x) dx, (8)
c
Z v
b b b
T (v) = −U (0) + Q (v)v − Qb (x) dx. (9)
0
These equations imply the following.
Theorem 4. Suppose that F is independent and that C = V = [0, 1]. Con-
sider any two generalized bargaining games and any two NE of these games.
If Qs and U s (1) are the same in both NE, then T s is the same in both NE.
And if Qb and U b (0) are the same in both NE, then T b is the same in both
NE.
Proof. Almost immediate from Equation 8 and Equation 9 since, in NE,
both the buyer and seller are optimizing, and since, by Equation 8, the ex-
pected payment to the seller depends only on Qs and U s (1), and similarly
for the expected payment by the buyer.
Remark 4. Theorem 4 is a variant of the Revenue Equivalence Theorem,

an important result in mechanism design theory. I have separate notes on
Revenue Equivalence in the context of auctions.
Ex post efficiency pins down Qs and Qb . If the equilibrium is ex post
efficient then Qs (c) equals the probability that v > c plus the probability
that trade occurs if v = c times the probability that v = c, all conditional
on c. But because of the assumption that the marginal distributions have
densities, the probability that v = c is zero, so that Qs (c) simply equals the
probability that v > c, given c. Similarly, Qb (v) equals the probability that
v > c, given v. Thus, Theorem 4 implies that if two NE are ex post efficient
then they have the same expected subsidy, Ec,v [T s (c) − T b (v)], iff they have
the same U s (1) and U b (0).
To complete the proof of Theorem 2, I need also to address the issue of
interim individual rationality. In words, the following theorem says that a
NE will be interim individually rational iff it is interim individually rational
for the seller and buyer types who are in the worst position to trade.
Theorem 5. Suppose that F is independent and that C = V = [0, 1].
(ρs , ρb ) satisfies interim individual rationality iff U s (1) ≥ 0 and U b (0) ≥ 0.
Proof. If U s (c) ≥ 0 for every c, then in particular U s (1) ≥ 0. So U s (1) ≥ 0
is necessary for individual rationality for the seller. To see that it is suffi-
cient, note that since Qs (c) ≥ 0 for every c, (1) implies that if U s (1) ≥ 0
10
then U s (c) ≥ 0 for every c. And a similar argument holds for U b .
I am now, finally, in a position to prove Theorem 2.

Proof of Theorem 2. Focus on ex post efficient NE of generalized bar-
gaining games. The truth-telling equilibrium of the VCG game is one such
equilibrium.
As already discussed, ex post efficiency implies that any such NE must
have the same Qs and Qb . In view of equalities (8) and (9), this means that
T s and T b are entirely determined by U s (1) and U b (0). From (8) and (9),
T s is increasing in U s (1) and T b is decreasing in U b (0). Thus, if I want
to minimize the net subsidy Ec,v [T s (c) − T b (v)], then I want to make both
U s (1) and U b (0) as small as possible. By Theorem 5, interim individual
rationality implies that the smallest possible value for either U s (1) or U b (0)
is zero.
Putting this all together, and appealing to Theorem 4, if I can find a NE
of a generalized bargaining game such that (a) the NE is ex post efficient,
and (b) U s (1) = U b (0) = 0, then this NE has the smallest expected subsidy
of any ex post efficient and interim individually rational NE of any gener-
alized bargaining game. The truth-telling NE of the VCG game has these
properties.
Remark 5. One sometimes sees MS stated in a weaker form: there does not
exist any NE in weakly dominant strategies of any generalized bargaining
game that is ex post efficient, expected budget balanced, and individually
rational. The restriction to NE in weakly dominant strategies is actually
without loss of generality, since the truth-telling NE of the VCG game is
in weakly dominant strategies. A benefit of focusing on weak dominance,
however, is that relatively easy proofs are available for MS in this special
case; Jeff Ely provides a nice graphical proof here.
4 The Myerson-Satterthwaite Theorem: General

Case.
Suppose that F is independent, so that supp(F ) = C × V . Suppose further
that C and V are non-degenerate intervals, with C = [c, c], V = [v, v], c > c,
and v > v.
If the C and V intervals do not overlap then MS fails. In particular, if
c ≤ v, then it is always ex post efficient to trade. As a bargaining “game”,
11
have the two players always trade at a price of (c + v)/2. On the other hand,
if c ≥ v, then it is always ex post efficient NOT to trade. As a bargaining
“game”, have the two players never trade.
The general statement of MS, therefore, assumes that C and V overlap,
meaning that c > v and c < v.
Theorem 6 (The Myerson-Satterthwaite Theorem). Suppose that F is in-
dependent, C and V are non-degenerate intervals, these intervals overlap,
and that the marginal distributions have densities. Then there is no NE
of any generalized bargaining game that is ex post efficient, expected budget
balanced, and individually rational.
Proof. Modify the VCG game so that if there is trade (v ∗ ≥ c∗ ) then the
buyer pays max{c∗ , v} (i.e., the buyer never pays less than v if there is trade)
and the seller receives min{v ∗ , c} (i.e., the seller never receives more than c
if there is trade).
In the modified VCG game, it is a Nash equilibrium (in weakly dominant
strategies) to report one’s true type. In this equilibrium, U s (c) = 0: if v < c
then there is no trade when c = c; otherwise, if v ≥ c, then whenever v ≥ c
(so that there is trade) and c = c, the seller receives a payment of c, for a
utility of zero. By a similar argument, U b (v) = 0. Since U s (c) = U b (v) =
0, essentially the same arguments as before imply that interim individual
rationality holds and that the expected subsidy by a third party is as small
as possible.
The assumptions on C and V guarantee that max{c, v} < min{c, v}.
There is thus an open subset of C × V for which (a) c < v (there is trade),
(b) v < c (so that the seller receives v) and (c) c > v (so that the buyer
pays c). Thus, there is a positive probability that the third party pays a
strictly positive subsidy, namely v − c > 0. Moreover, working through the
remaining cases shows that the subsidy is never negative:
• if v ≥ c and c ≤ v, then (since, by assumption, c > v), v > c (there is
trade) and the subsidy is c − v > 0;
• if v ≥ c and c > v then (since c ≤ c) v ≥ c (there is trade) and the
subsidy is c − c ≥ 0;
• if v < c and c ≤ v then (since v ≥ v) v ≥ c (there is trade) and the
subsidy is v − v ≥ 0.
The overall expected subsidy is, therefore, strictly positive, which implies
that expected budget balance fails.
12
5 Remarks on the Distribution Assumptions.
5.1 If the joint distribution is not independent.
If F is not independent then MS fails. In particular, McAfee and Reny
(1992) shows that if independence does not hold then there is a generalized
bargaining game that is ex post efficient, interim individually rational, and
satisfies expected budget balance (but possibly not ex post budget balance).
As a simple example, suppose that c is uniformly distributed on [0, 1] and
√
that v = c. The support of F is then the graph of the square root function
on [0, 1]. Consider the game in which the seller reports c∗ (which may be
different from the seller’s true type), the buyer reports v ∗√(which may be
different from the buyer’s true type), there is trade iff v ∗ = c∗ , and if trade
takes place then the seller sells to the buyer at the price (c∗ + v ∗ )/2. There
is an equilibrium in which, for any (c, v) ∈ supp(F ), both players report
their true type. This equilibrium is ex post efficient, interim individually
rational, and ex post budget balanced.
Intuitively, correlation allows a third party to use the buyer’s behavior
to infer information about the cost of the seller, and vice versa. In the
example, this takes the form of having trade take place iff the√seller and
buyer agree, meaning that they report (c∗ , v ∗ ) such that v ∗ = c∗ . More
generally, an ability to crosscheck increases the scope of what the equilibrium
of a generalized bargaining game can achieve.
In terms of the formal argument, if F is not independent then it may
not be true that, for example, Q̃s (c, c∗ ) = Qs (c∗ ). This vitiates much of the
analysis. And in the Envelope Theorem approach (Remark 3 in Section 3),
if independence is dropped then Equation (7) becomes,
DU s (c∗ ) = −Qs (c∗ ) − D2 Q̃s (c∗ , c∗ )c∗ + D2 T̃ s (c∗ , c∗ ),
which no longer implies the integral condition, Equation (6). In particular,

note that the Envelope Theorem expression for DU s now contains a term
involving the expected payment, T̃ s .
5.2 If C and V are not non-degenerate intervals.

If C and V are intervals but one is degenerate then MS can fail. For example,
suppose that C = [0, 1] but that V = {1/2} (so that v = v = 1/2). (Since V
is degenerate, this example also violates the marginal density assumption.)
Consider the game in which the seller reports c∗ ∈ C (which may be different
from the seller’s true type), there is trade iff c∗ < 1/2, and if there is trade
13
then it is at a price of 1/2. Then it is an equilibrium for both players
to report their true type, and this equilibrium is ex post efficient, interim
individually rational, and ex post budget balanced.
MS can also fail if C and V are not intervals. Suppose C = V = {0, 1},
and that the four possible outcomes are equally likely. (This example also
violates the marginal density assumption.) Consider the game in which the
seller reports c∗ ∈ C (which may be different from the actual c), the buyer
reports v ∗ ∈ V (which may be different from the actual v), there is trade iff
(c∗ , v ∗ ) = (0, 1), and if there is trade then it is at a price of 1/2. Then it is an
equilibrium for both players to report their true type, and this equilibrium is
ex post efficient, interim individually rational, and ex post budget balanced.
In the first example, the suggested game is the modified VCG game used
in the proof of Theorem 6. Here, however, because V is degenerate, there
is zero probability that (c, v) falls into a range where a third party must
provide a strictly positive subsidy.
In the second example, a partial intuition is that any equilibrium of
any generalized bargaining game must satisfy the incentive compatibility
(IC) conditions, Inequality 3 and Inequality 4. Very loosely, the more IC
conditions there are, the more restricted is equilibrium behavior. If C and
V are intervals, then the seller and buyer have, for each realized c or v, a
continuum of IC conditions. In contrast, if C = V = {0, 1} then the seller
and buyer have, for each realized c or v, only a single IC condition (e.g., it
cannot be profitable for a c = 0 seller to behave as if c = 1). In terms of
the formal argument, the integral conditions (Equations (6) and (2)) assume
that Qs and Qb are defined on intervals.
5.3 If the marginal distributions on C and V do not have

densities.
The density assumption is used to pin down Qs and Qb . Explicitly, if the
marginal distribution of v has a density then ex post efficiency implies that
Qs (c) equals the probability that v > c, given c; the case v = c (which is not
pinned down by ex post efficiency) has probability zero and can be ignored.
And a similar statement holds for Qb .
A weaker sufficient condition for this argument to hold is that the prob-
ability that v = c, given either c or v, is zero. This condition could be met
even if the marginal distributions do not admit densities, so long as any
atoms are not in the same place for both the buyer and the seller.
And this issue can be side-stepped entirely, with no restriction on the
marginal distributions (other than that their supports be non-degenerate
14
intervals) by strengthening ex post efficiency to require that trade occurs
whenever v = c, since in this case, ex post efficiency implies that Qs (c) must
equal the probability that v ≥ c, given c, and similarly for Qb (v). (The
game discussed in the second example in Section 5.2 violates this stronger
version of efficiency.)
References
Clarke, Edward. 1971. “Multipart Pricing of Public Goods.” Public Choice
2:19–33.
Cripps, Martin and Jeroen Swinkels. 2006. “Depth and Efficiency in Large
Double Auctions.” Econometrica 74(1):47–92.
Groves, Theodore. 1973. “Incentives in Teams.” Econometrica 41:617–631.
Krishna, Vijay. 2009. Auction Theory. Second ed. Academic Press.
McAfee, Preston and Philip Reny. 1992. “Correlated Information and Mech-
anism Design.” Econometrica 60(2):395–421.
Milgrom, Paul. 2004. Putting Auction Theory To Work. Cambridge Uni-

versity Press.
Milgrom, Paul and Ilya Segal. 2002. “Envelope Theorems for Arbitrary
Choice Sets.” Econometrica 70(2):583–601.
Myerson, Roger and Mark Satterthwaite. 1983. “Efficient Mechanisms for

Bilateral Trading.” Journal of Economic Theory 28:265–81.
Rudin, Walter. 1976. Principles of Mathematical Analysis. Third ed. New

York: McGraw-Hill.
Rustichini, Aldo, Mark Satterthwaite and Steven Williams. 1994. “Conver-

gence to Efficiency in a Simple Market With Incomplete Information.”
Econometrica 62:1041–63.
Vickrey, William. 1962. Auctions and Bidding Games. In Recent Advances

in Game Theory. Number 29 in “Princeton Conference Series” Princeton
University Press pp. 15–27.
15

Notes ST Lewis PDF

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Notes ST Lewis PDF

Încărcat de

Drepturi de autor:

Formate disponibile

John Nachbar

Game Theory Basics I:

Figure 1: A game box for Rock-Paper-Scissors (RPS).

game: R (rock), P (paper), and S (scissors). Player 1 is represented by the rows

3 Strategic Form Games.

• ui is player i’s payoff function: ui : S → R. I discuss the interpretation of

Figure 2: An independent distribution over strategy profiles.

Finally, I frequently use the following notation. A strategy

Figure 3: A game box for Battle of the Sexes.

Figure 5: An independent distribution strategy profiles for Battle of the Sexes.

The space of all probability distributions over S is Σc , also denoted

3.3 Interpreting Payoffs.

1. In many economic applications, the prize is a profile of monetary values (prof-

2. It can be difficult to “test” game theory predictions such as Nash equilibrium

3. If vi represents choice over (pure) outcomes as in standard decision theory, then

4. While it may be reasonable to assume that there is common knowledge of γ

3.4 Interpreting Randomization.

1. Objective Randomization. Each player has access to a randomization device.

2. Empirical Randomization. From the perspective of an observer (say an exper-

3. Subjective Randomization. From the perspective of player j 6= i, σi [si ] is the

BRi (σ−i ) = {σi ∈ Σi : ∀σ̂i ∈ Σi , ui (σi , σ−i ) ≥ ui (σ̂i , σ−i )} .

An element of BRi (σ−i ) is called a best response to σ−i .

BR(σ) = {σ̂ ∈ Σ : ∀i σ̂i ∈ BRi (σ−i )} .

BR is a correspondence on Σ. Since Σi is compact for each i, Σ is compact. For

4.2 Nash Equilibrium.

BR(σ ∗ ): a NE is a fixed point of BR.

Definition 1. Fix a game. A strategy profile σ ∗ ∈ Σ is a Nash equilibrium (NE)

4.3 NE and randomization.

Theorem 2 implies that a NE mixed σi will give positive probability to two

Figure 6: A general 2 x 2 game.

pa2 + (1 − p)c2 = pb2 + (1 − p)d2 ,

((1/3, 1/3, 1/3), (1/3, 1/3, 1/3)).

4.4 More NE examples.

Figure 7: Matching Pennies.

Figure 8: A Prisoner’s Dilemma.

Figure 9: A non-PD game.

• If p1 ∈ (0, p2 ) then firm 1 receives a payoff of (1 − p1 )p1 . In words, firm 1

• If p1 > p2 , then firm 1 gets a payoff of 0.

• p2 > 1/2 then firm 1’s best response is p1 = 1/2.

• If p2 = 0, then any p1 ∈ R+ is a best response.

The sales function for firm 2 is analogous.

The best response for firm 2 to pure strategies of player 1 is analogous.

F2 (p1 )p1 (1/4 − p1 ) + (1 − F2 (p1 ))p1 3/4 = 1/64,

4.5 The MinMax Theorem.

Theorem 3. In a finite constant-sum game, if σ ∗ = (σ1∗ , σ2∗ ) is a NE then,

u1 (σ ∗ ) = max min u1 (σ) = min max u1 (σ).

V = max u1 (σ1 , σ2∗ ) = min u1 (σ1∗ , σ2 ). (1)

On the other hand, for any σ̃2 ,

u1 (σ1∗ , σ̃2 ) ≤ max u1 (σ1 , σ̃2 ),

Since the latter holds for every σ̃2 ,

min u1 (σ1∗ , σ2 ) ≤ min max u1 (σ1 , σ2 ).

Combining with expressions (1) and (2),

V = min u1 (σ1∗ , σ2 ) = min max u1 (σ).

V = max u1 (σ1 , σ2∗ ) = max min u1 (σ).

And the result follows.

Theorem 4 (MinMax Theorem). In any finite constant-sum game, there is a num-

Proof. By Theorem 1, there is NE, σ∗. Set V = u1 (σ ∗ ) and apply Theorem 3.

Figure 10: The Aumann-Maschler Game.

4.6 Never a Best Response and Rationalizability.

Theorem 5. Fix a game. If σ is a NE and σi [si ] > 0 then si ∈ SiR .

Remark 5. The construction of S R uses maximal deletion of never a best response

Figure 11: Rationalizability and NE.

But the unique NE is (A, A).