Solution Concepts

Jonathan Levin

April 2006

These notes discuss some of the central solution concepts for normal-
form games: Nash and correlated equilibrium, iterated deletion of strictly
dominated strategies, rationalizability, and self-confirming equilibrium.

1 Nash Equilibrium
Nash equilibrium captures the idea that players ought to do as well as they
can given the strategies chosen by the other players.

Example 1 Prisoners’ Dilemma

C D
C 1, 1 −1, 2
D 2, −1 0, 0

The unique Nash Equilibrium is (D, D).

Example 2 Battle of the Sexes

B F
B 2, 1 0, 0
F 0, 0 1, 2

There are two pure Nash equilibria (B, B) and (F, F ) and a mixed
strategy equilibrium where Row plays 23 B + 13 F and Column plays
1 2
3B + 3F.

Definition 1 A normal form game G consists of

1. A set of players i = 1, 2, ..., I.

2. Strategy sets S1 , ..., SI ; let S = S1 × ... × SI .

1
 3. Payoff functions: for each i = 1, ..., I, ui : S → R

A (mixed) strategy for player i, σ i ∈ ∆(Si ), is a probability distribution

on Si . A pure strategy places all probability weight on a single action.

Definition 2 A strategy profile (σ 1 , ..., σ I ) is a Nash equilibrium of G if

for every i, and every si ∈ Si ,

ui (σ i , σ −i ) ≥ ui (si , σ −i ).

And recall Nash’s famous result:

Proposition 1 Nash equilibria exist in finite games.

A natural question, given the wide use of Nash equilibrium, is whether

or why one should expect Nash behavior. One justification is that rational
players ought somehow to reason their way to Nash strategies. That is,
Nash equilibrium might arrive through introspection. A second justification
is that Nash equilibria are self-enforcing. If players agree on a strategy
profile before independently choosing their actions, then no player will have
reason to deviate if the agreed profile is a Nash equilibrium. On the other
hand, if the agreed profile is not a Nash equilibrium, some player can do
better by breaking the agreement. A third, and final, justification is that
Nash behavior might result from learning or evolution. In what follows, we
take up these three ideas in turn.

2 Correlated Equilibrium
2.1 Equilibria as a Self-Enforcing Agreements
Let’s start with the account of Nash equilibrium as a self-enforcing agree-
ment. Consider Battle of the Sexes (BOS). Here, it’s easy to imagine the
players jointly deciding to attend the Ballet, then playing (B, B) since nei-
ther wants to unilaterally head off to Football. However, a little imagination
suggests that Nash equilibrium might not allow the players sufficient freedom
to communicate.

Example 2, cont. Suppose in BOS, the players flip a coin and go to the
Ballet if the coin is Heads, the Football game if Tails. That is, they
just randomize between two different Nash equilibria. This coin flip
allows a payoff ( 32 , 32 ) that is not a Nash equilibrium payoff.

2
 So at the very least, one might want to allow for randomizations between
Nash equilibria under the self-enforcing agreement account of play. More-
over, the coin flip is only a primitive way to communicate prior to play. A
more general form of communication is to find a mediator who can perform
clever randomizations, as in the next example.

Example 3 This game has three Nash equilibria (U, L), (D, R) and ( 12 U +
1 1 1 5 5
2 D, 2 L + 2 R) with payoffs (5, 1), (1, 5) and ( 2 , 2 ).

L R
U 5, 1 0, 0
D 4, 4 1, 5

Suppose the players find a mediator who chooses x ∈ {1, 2, 3} with

equal probability 13 . She then sends the following messages:

• If x = 1 ⇒ tells Row to play U, Column to play L.

• If x = 2 ⇒ tells Row to play D, Column to play L.
• If x = 3 ⇒ tells Row to play D, Column to play R.

Claim. It is a Perfect Bayesian Equilibrium for the players to follow the

mediator’s advice.

Proof. We need to check the incentives of each player.

• If Row hears U , believes Column will play L ⇒ play U .

• If Row hears D, believes Column will play L, R with 12 , 12 proba-
bility ⇒ play D.
• If Column hears L, believes Row will play U, D with 12 , 12 proba-
bility ⇒ play L.
• If Column hears R, believes Row will play D ⇒ play R.

Thus the players will follow the mediator’s suggestion. With the me-
diator in place, expected payoffs are ( 10 10
3 , 3 ), strictly higher than the
players could get by randomizing between Nash equilibria.

2.2 Correlated Equilibrium

The notion of correlated equilibrium builds on the mediator story.

Definition 3 A correlating mechanism (Ω, {Hi }, p) consists of:

3
 • A finite set of states Ω

• A probability distribution p on Ω.

• For each player i, a partition of Ω, denoted {Hi }. Let hi (ω) be a

function that assigns to each state ω ∈ Ω the element of i’s partition
to which it belongs.

Example 2, cont. In the BOS example with the coin flip, the states are
Ω = {Heads, Tails}, the probability measure is uniform on Ω, and Row
and Column have the same partition, {{Heads}, {Tails}}.

Example 3, cont. In this example, the set of states is Ω = {1, 2, 3}, the
probability measure is again uniform on Ω, Row’s partition is {{1}, {2, 3}},
and Column’s partition is {{1, 2}, {3}}.

Definition 4 A correlated strategy for i is a function fi : Ω → Si that

is measurable with respect to i’s information partition. That is, if hi (ω) =
hi (ω 0 ) then fi (ω) = fi (ω0 ).

Definition 5 A strategy profile (f1 , ..., fI ) is a correlated equilibrium rel-

ative to the mechanism (Ω, {Hi }, p) if for every i and every correlated strat-
egy f˜i :
X X ³ ´
ui (fi (ω), f−i (ω)) p (ω) ≥ ui f˜(ω), f−i (ω) p (ω) (1)
ω∈Ω ω∈Ω

This definition requires that fi maximize i’s ex ante payoff. That is, it
treats the strategy as a contingent plan to be implemented after learning the
partition element. Note that this is equivalent to fi maximizing i’s interim
payoff for each Hi that occurs with positive probability – that is, for all
i,ω, and every s0i ∈ Si ,
X X
ui (fi (ω), f−i (ω 0 ))p(ω 0 |hi (ω)) ≥ ui (s0i , f−i (ω 0 ))p(ω 0 |hi (ω))
ω0 ∈hi (ω) ω 0 ∈hi (ω)

Here, p(ω 0 |hi (ω)) is the conditional probability on ω 0 given that the true
state is in hi (ω). By Bayes’ Rule,

Pr (hi (ω)|ω 0 ) p(ω 0 ) p(ω0 )

p(ω0 |hi (ω)) = P 00 00
=
ω00 ∈hi (ω) Pr (hi (ω)|ω ) p(ω ) p(hi (ω))

4
 The definition of CE corresponds to the mediator story, but it’s not very
convenient. To search for all the correlated equilibria, one needs to con-
sider millions of mechanisms. Fortunately, it turns out that we can focus
on a special kind of correlating mechanism, callled a direct mechanism. We
will show that for any correlated equilibrium arising from some correlating
mechanism, there is a correlated equilibrium arising from the direct mech-
anism that is precisely equivalent in terms of behavioral outcomes. Thus
by focusing on one special class of mechanism, we can capture all possible
correlated equilibria.

Definition 6 A direct mechanism has Ω = S, hi (s) = {s0 ∈ S : s0i = si },

and some probability distribution q over pure strategy profiles.

Proposition 2 Suppose f is a correlated equilibrium relative to (Ω, {Hi }, p).

Define q(s) ≡ Pr(f (ω) = s). Then the strategy profile f˜ with f˜i (s) = si
for all i, s ∈ S is a correlated equilibrium relative to the direct mechanism
(S, {H̃i }, q).

Proof. Suppose that si is recommended to i with positive probability, so

p(si , s−i ) > 0 for some s−i . We check that under the direct mechanism
(S, {H̃i }, q), player i cannot benefit from choosing another strategy s0i when
si is suggested. If si is recommended, then i’s expected payoff from playing
s0i is: X
ui (s0i , s−i )q(s−i |si ).
s−i ∈S−i

The result is trivial if there is only one information set Hi in the original
mechanism for which fi (Hi ) = si . In this case, conditioning on si is the same
as conditioning on Hi in the original. More generally, we substitute for q to
obtain: X
1
· ui (s0i , f−i (ω))p(ω).
Pr(fi (ω) = si )
ω|fi (ω)=si

Re-arranging to separate each Hi at which si is optimal:

⎡ ⎤
1 X X
· Pr (Hi ) ⎣ ui (s0i , f−i (ω))p(ω|Hi )⎦
Pr(fi (ω) = si )
Hi |fi (Hi )=si ω∈Hi

Since (Ω, {Hi }, p, f ) is a correlated equilibrium, each bracketed term for

which Pr(Hi ) > 0 is maximized at fi (Hi ) = si . So si is optimal given
recommendation si . Q.E.D.

5
 Thus what really matters in correlated equilibrium is the probability
distribution over strategy profiles. We refer to any probability distribution
q over strategy profiles that arises as the result of a correlated equilibrium
as a correlated equilibrium distribution (c.e.d.).
Example 2, cont. In the BOS example, the c.e.d. is 12 (B, B), 12 (F, F ).
Example 3, cont. In this example, the c.e.d is 13 (U, L), 13 (D, L), 13 (D, R).
The next result characterizes correlated equilibrium distributions.
Proposition 3 The distribution q ∈ ∆(S) is a correlated equilibrium dis-
tribution if and only if for all i, every si with q(si ) > 0 and every s0i ∈ Si ,
X X
ui (si , s−i ) q (s−i |si ) ≥ ui (s0i , s−i )q(s−i |si ). (2)
s−i ∈S−i s−i ∈S−i

Proof. (⇐) Suppose q satisfies (2). Then the “obedient” profile f with
fi (s) = si is a correlated equilibrium given the direct mechanism (S, {Hi }, q)
since (2) says precisely that with this mechanism si is optimal for i given
recommendation si . (⇒) Conversely, if q arises from a correlated equilib-
rium, the previous result says that the obedient profile must be a correlated
equilibrium relative to the direct mechanism (S, {Hi }, q). Thus for all i
and all recommendations si occuring with positive probability, si must be
optimal – i.e. (2) must hold. Q.E.D.

Consider a few properties of correlated equilibrium.

Property 1 Any Nash equilibrium is a correlated equilibrium
Proof. Need to ask if (2) holds for the probability distribution q over
outcomes induced by the NE. For a pure equilibrium s∗ , we have q(s∗−i |s∗i ) =
1 and q(s−i |s∗i ) = 0 for any s−i 6= s∗−i . Therefore (2) requires for all i, si :
ui (s∗i , s∗−i ) ≥ ui (si , s∗−i ).
This is precisely the definition of NE. For a mixed equilibrium, σ ∗ , we have
that for any s∗i in the support of σ ∗i , q(s−i |s∗i ) = σ −i (s−i ). This follows from
the fact that in a mixed NE, the players mix independently. Therefore (2)
requires that for all i, s∗i in the support of σ ∗i , and si ,
X X
ui (s∗i , s−i ) σ −i (s−i ) ≥ ui (si , s−i )σ −i (s−i ),
s−i ∈S−i s−i ∈S−i

again, the definition of a mixed NE. Q.E.D.

6
Property 2 Correlated equilibria exist in finite games.

Proof. Any NE is a CE, and NE exists. Hart and Schmeidler (1989)

show the existence of CE directly, exploiting the fact that a CE is just a
probability distribution q satisfying a system of linear inequalities. Their
proof does not appeal to fixed point results! Q.E.D.

Property 3 The sets of correlated equilibrium distributions and payoffs are

convex.

Proof. Left as an exercise.

2.3 Subjective Correlated Equilibrium

The definition of correlated equilibrium assumes the players share a common
prior p over the set of states (or equivalently share the same probability dis-
tribution over equilibrium play). A significantly weaker notion of equilibrium
obtains if this is relaxed. For this, let p1 , p2 , ..., pI be distinct probability
measures on Ω.

Definition 7 The profile f is a subjective correlated equilibrium rela-

tive to the mechanism (Ω, {Hi }, p1 , ..., pI ) if for every i, ω and every alter-
native strategy f˜i ,
X X
ui (fi (ω), f−i (ω))pi (ω) ≥ ui (f˜(ω), f−i (ω))pi (ω)
ω∈Ω ω∈Ω

Example 3, cont. Returning to our example from above,

L R
U 5, 1 0, 0
D 4, 4 1, 5

Here, (4, 4) can be obtained as a SCE payoff. Simply consider the

direct mechanism with p1 = p2 = 13 (U, L) + 13 (D, L) + 13 (D, R). This
is a SCE, and since there is no requirement that the players have
objectively correct beliefs about play, it may be that (D, L) is played
with probability one!

7
2.4 Comments
1. The difference between mixed strategy Nash equilibria and correlated
equilibria is that mixing is independent in NE. With more than two
players, it may be important in CE that one player believes others
are correlating their strategies. Consider the following example from
Aumann (1987) with three players: Row, Column and Matrix.
0, 0, 3 0, 0, 0 2, 2, 2 0, 0, 0 0, 0, 0 0, 0, 0
1, 0, 0 0, 0, 0 0, 0, 0 2, 2, 2 0, 1, 0 0, 0, 3
No NE gives any player more than 1, but there is a CE that gives
everyone 2. Matrix picks middle, and Row and Column pick (Up,Left)
and (Down,Right) each with probability 12 . The key here is that Matrix
must expect Row to pick Up precisely when Column picks Left.
2. Note that in CE, however, each agent uses a pure strategy – he just
is uncertain about others’ strategies. So this seems a bit different than
mixed NE if one views a mixed strategy as an explicit randomization
in behavior by each agent i. However, another view of mixed NE if
that it’s not i’s actual choice that matters, but j’s beliefs about i’s
choice. On this account, we view σ i as what others expect of i, and i
as simply doing some (pure strategy) best response to σ −i . This view,
which is consistent with CE, was developed by Harsanyi (1973), who
introduced small privately observed payoff perturbations so that in
pure strategy BNE, players would be uncertain about others behavior.
His “purification theorem” showed that these pure strategy BNE are
observably equivalent to mixed NE of the unperturbed game if the
perturbations are small and independent.
3. Returning to our pre-play communication account, one might ask if a
mediator is actually needed, or if the players could just communicate
by flipping coins and talking. With two players, it should be clear
from the example above that the mediator is crucial in allowing for
messages that are not common knowledge. However, Barany (1992)
shows that if I ≥ 4, then any correlated equilibrium payoff (with ratio-
nal numbers) can be achieved as the Nash equilibrium of an extended
game where prior to play the players communicate through cheap talk.
Girardi (2001) shows the same can be done as a sequential equilibrium
provided I ≥ 5. For the case of two players, Aumann and Hart (2003)
characterize the set of attainable payoffs if players can communicate
freely, but without a mediator, prior to playing the game.

8
3 Rationalizability and Iterated Dominance
Bernheim (1984) and Pearce (1984) investigated the question of whether one
should expect rational players to introspect their way to Nash equilibrium
play. They argued that even if rationality was common knowledge, this
should not generally be expected. Their account takes a view of strategic
behavior that is deeply rooted in single-agent decision theory.
To discuss these ideas, it’s useful to explicitly define rationality.

Definition 8 A player is rational if he chooses a strategy that maximizes

his expected payoff given his belief about opponents’ strategies.

Note that assessing rationality requires defining beliefs, something that

the formal definition of Nash equilibrium does not require. Therefore, as
a matter of interpretation, if we’re talking about economic agents playing
a game, we might say that Nash equilibrium arises when each player is
rational and know his opponents’ action profile. But we could also talk
about Nash equilibrium in an evolutionary model of fish populations without
ever mentioning rationality.

3.1 (Correlated) Rationalizability

Rationalizability imposes two requirements on strategic behavior.

1. Players maximize with respect to their beliefs about what opponents

will do (i.e. are rational).

2. Beliefs cannot conflict with other players being rational, and being
aware of each other’s rationality, and so on (but they need not be
correct).

Example 4 In this game (from Bernheim, 1984), there is a unique Nash

equilibrium (a2 , b2 ). Nevertheless a1 , a3 , b1 , b3 can all be rationalized.

b1 b2 b3 b4
a1 0, 7 2, 5 7, 0 0, 1
a2 5, 2 3, 3 5, 2 0, 1
a3 7, 0 2, 5 0, 7 0, 1
a4 0, 0 0, −2 0, 0 10, −1

• Row will play a1 if Column plays b3

• Column will play b3 if Row plays a3

9
 • Row will play a3 if Column plays b1
• Column will play b1 if Row plays a1

This “chain of justification” rationalizes a1 , a3 , b1 , b3 . Of course a2 and

b2 rationalize each other. However, b4 cannot be rationalized, and since
no rational player would play b4 , a4 can’t be rationalized.

Definition 9 A subset B1 × ... × BI ⊂ S is a best reply set if for all i and

all si ∈ Bi , there exists σ −i ∈ ∆ (B−i ) to which si is a best reply.

• In the definition, note that σ −i can reflect correlation – it need not

be a mixed strategy profile for the opponents. This allows for more
“rationalizing” than if opponents mix independently. More on this
later.

Definition 10 The set of correlated rationalizable strategies is the

component by component union of all best reply sets:
[
R = R1 × ... × RI = B1α × ... × BIα
α

where each B α = B1α × ... × BIα is a best reply set.

Proposition 4 R is the maximal best reply set.

Proof. Suppose si ∈ Ri . Then si ∈ Biα for some α. So si is a best reply to

some σ −i ∈ ∆(B−i α ) ⊂ ∆(R ). So R is a best reply set. Since it contains
−i i
all others, it is maximal. Q.E.D.

3.2 Iterated Strict Dominance

In contrast to asking what players might do, iterated strict dominance asks
what players won’t do, and what they won’t do conditional on other players
not doing certain things, and so on. Recall that a strategy si is strictly domi-
nated if there is some mixed strategy σ i such that ui (σ i , s−i ) > ui (si , s−i ) for
all si ∈ S−i , and that iterated dominance applies this definition repeatedly.

• Let Si0 = Si
½ ¾
k si ∈ Sik−1 : There is no σ i ∈ ∆(Sik−1 ) s.t.
• Let Si = k−1
ui (σ i , s−i ) > ui (si , s−i ) for all s−i ∈ S−i
T
• Let Si∞ = ∞ k=1 Si .
k

10
 Iterated strict dominance never eliminates Nash equilibrium strategies,
or any strategy played with positive probability in a correlated equilibrium
(proof left as an exercise!). Indeed it is often quite weak. Most games,
including many games with a unique Nash equilibrium, are not dominance
solvable.
Example 4, cont. In this example, b4 is strictly dominated. Eliminating
b4 means that a4 is also strictly dominated. But no other strategy can
be eliminated.
Proposition 5 In finite games, iterated strict dominance and correlated
rationalizability give the same solution set, i.e. Si∞ = Ri .
This result is suggested by the following Lemma (proved in the last
section of these notes).
Lemma 1 A pure strategy in a finite game is a best response to some beliefs
about opponent play if and only if it is not strictly dominated.
Proof of Proposition.
R ⊂ S ∞ . If si ∈ Ri , then si is a best response to some belief over R−i .
Since R−i ⊂ S−i , Lemma 1 implies that si is not strictly dominated. Thus
Ri ⊂ Si2 for all i. Iterating this argument implies that Ri ⊂ Sik for all i, k,
so Ri ⊂ Si∞ .
S ∞ ⊂ R. It suffices to show that S ∞ is a best-reply set. By definition,
no strategy in S ∞ is strictly dominated in the game in which the set of
actions is S ∞ . Thus, any si ∈ Si∞ must be a best response to some beliefs
∞.
over S−i Q.E.D.

3.3 Comments
1. Bernheim (1984) and Pearce (1984) originally defined rationalizability
assuming that players would expect opponents to mix independently.
So B is a best reply set if ∀si ∈ Bi , there is some σ −i ∈ ×j6=i ∆(Sj ) to
which si is a best reply. For I = 2, this makes no difference, but when
I ≥ 3, their concept refines ISD (it rules out more strategies).
2. Brandenburger and Dekel (1987) relate correlated rationalizability to
subjective correlated equilibrium. While SCE is more permissive than
rationalizability, requiring players to have well defined conditional be-
liefs (and maximize accordingly) even for states ω ∈ Ω to which they
assign zero probability leads to a refinement of SCE that is the same
as correlated rationalizability.

11
 3. One way to categorize the different solution concepts is to note that if
one starts with rationality, and common knowledge of rationality, the
concepts differ precisely in how they further restrict the beliefs of the
players about the distribution of play. Doug Bernheim suggests the
following table:

Different Priors Common Prior

Corr. Rationalizability/
Correlation Correlated Equilibrium
ISD/Refined Subj CE
Independence Rationalizability Nash Equilibrium

3.4 Appendix: Omitted Proof

For completeness, this section provides a proof of the Lemma equating dom-
inated strategies with those that are never a best response. The proof re-
quires a separation argument, so let’s first recall the Duality Theorem for
linear programming. To do this, start with the following problem:
n
X
minn cj xj (3)
x∈R
j=1
Xn
s.t. aij xj ≥ bi ∀i = 1, ..., m
j=1

This problem has the same solution as

⎛ ⎛ ⎞⎞
n
X n
X n
X
max ⎝ min cj xj + yi ⎝bi − aij xj ⎠⎠ . (4)
m
y∈R+ n
x∈R
j=1 i=1 j=1

Rearranging terms, we obtain

⎛ Ã ! ⎞
n
X m
X m
X
max ⎝ min cj − yi aij xj + yi bi ⎠ . (5)
m
y∈R+ n
x∈R
j=1 i=1 i=1

Swapping the order of optimization give us

⎛ Ã ! ⎞
Xn m
X m
X
minn ⎝ maxm
cj − yi aij xj + yi bi ⎠ , (6)
x∈R y∈R+
j=1 i=1 i=1

12
which can be related to the following “dual” problem:
m
X
max
m
yi bi (7)
y∈R+
i=1
n
à m
!
X X
s.t. cj − yi aij =0 ∀j = 1, ..., n.
j=1 i=1

Theorem 1 Suppose problems (3) and (7) are feasible (i.e. have non-empty
constraint sets). Then their solutions are the same.

We use the duality theorem to prove the desired Lemma.

Lemma 2 A pure strategy in a finite game is a best response to some beliefs

about opponent play if and only if it is not strictly dominated.

Proof (Myerson, 1991). Let si ∈ Si be given. Our proof will be based on a

comparison of two linear programming problems.

Problem I:

min δ
σ −i ,δ
s.t. σ −i (s−i ) ≥ 0 ∀s−i ∈ S−i
X X
σ −i (s−i ) ≥ 1 and − σ −i (s−i ) ≥ −1
s−i s−i
X £ ¤
δ+ σ −i (s−i ) ui (si , s−i ) − ui (s0i , s−i ) ≥ 0 ∀s0i ∈ Si
s−i

Observe that si is a best response to some beliefs over opponent play if any
only if the solution to this problem is less than or equal to zero.

Problem II:

max ε1 − ε2
η,ε1 ,ε2 ,σi
|S | |S |
s.t. ε1 , ε2 ∈ R+ , σ i ∈ R+ i , η ∈ R+ −i
X
σ i (s0i ) ≥ 1
s0i
X £ ¤
η(s−i ) + ε1 − ε2 + σ i (s0i ) ui (si , s−i ) − ui (s0i , s−i ) = 0 ∀s−i ∈ S−i
s0i

13
Observe that si is strictly dominated if and only if the solution to this
problem is strictly greater than zero – i.e. si is not strictly dominated if
and only if the solution to this problem is less than or equal to zero.

Finally, the Duality Theorem for linear programming says that so long
as these two problems are feasible (have non-empty constraint sets), their
solutions must be the same, establishing the result. Q.E.D.

4 Self-Confirming Equilibria
The third possible foundation for equilibrium is learning. We’ll look at
explicit learning processes later; for now, we ask what might happen as the
end result of a learning processes. For instance, if a learning process settles
down into steady-state play, will this be a Nash Equilibrium? Fudenberg
and Levine (1993) suggest that a natural end-result of learning is what they
call self-confirming equilibria. In a self-confirming equilibrium:

1. Players maximize with respect to their beliefs about what opponents

will do (i.e. are rational).

2. Beliefs cannot conflict with the empirical evidence (i.e. must match
the empirical distribution of play).

The difference with Nash equilibrium and rationalizability lies in the

restriction on beliefs. In a Nash equilibrium, players hold correct beliefs
about opponents’ strategies and hence about their behavior. By contrast,
with rationalizability, beliefs need not be correct, they just can’t conflict
with rationality. With SCE, beliefs need to be consistent with available
data.

4.1 Examples of Self-Confirming Equilibria

In a simultaneous game, assuming actions are observed after every period,
every Nash equilibrium is self-confirming. Moroeover, any self-confirming
equilibrium is Nash.

Example 1 Consider matching pennies. It is an SCE for both players to

mix 50/50 and to both believe the other is mixing 50/50. On the other
hand, if player i believes anything else, he must play a pure strategy.
But then player j must believe i will play this strategy or else he would
eventually be proved wrong. So the only SCE is the same as the NE.

14
 In extensive form games, the situation is different, as the next example
shows.
Example 2 In the entry game below, the only Nash equilibria are (Enter,
Don’t Fight) and (Don’t Enter, Fight). These equilibria, with correct
beliefs, are also self-confirming equilibria.

Player 1

Don’t Enter Enter

Player 2
2, 0
Fight Don’t Fight

-1, 1 1, 1

An Entry Game
There is also another SCE, however, where Player 1 plays Don’t Enter
and believes Fight, while Player 2 player plays Don’t Fight and be-
lieves Don’t Enter. In this SCE, player 1 has the wrong beliefs, but
since he never enters, they’re never contradicted by the data!
In the entry game, the non-Nash SCE is indistinguishable from a Nash
equilibrium in terms of observed behavior. But even that need not be the
case, as the next example shows.
Example 3 (Fudenberg-Kreps, 1993) Consider the three player game be-
low.

1 A1 2 A2
(1,1,1)

D1 D2
3

L R L R

(3,0,0) (0,3,0) (3,0,0) (0,3,0)

SCE are different than NE

15
 In this game, there is a self-confirming equilibrium where (A1 , A2 ) is
played. In this equilibrium, player 1 expects player 3 to play R, while
player 2 expects 3 to play L. Given these beliefs, the optimal strategy
for player 1 is to play A1 , while the optimal strategy or player 2 is to
play A2 . Player 3’s beliefs and strategy can be arbitrary so long as the
strategy is optimal given beliefs.
The key point here is that there is no Nash equilibrium where (A1 , A2 )
is played. The reason is that in any Nash equilibrium, players 1 and
2 must have the same (correct) beliefs about player 3’s strategy. But
if they have the same beliefs, then at least one of them must want to
play D.

The distinction between Nash and self-confirming equilibria in extensive

form games arises because players do not get to observe all the relevant
information about their opponents’ behavior. The same issue can arise in
simultaneous-move games as well if the information feedback that players’
receive is limited.

Example 4 Consider the following two-player game. Suppose that after the
game is played, the column player observes the row player’s action, but
the row player observes only whether or not the column player chose
R, and gets no information about her payoff.

L M R
U 2, 0 0, 2 0, 0
D 0, 0 2, 0 3, 3

This game has a unique Nash equilibrium, (D, R); indeed the game is
dominance-solvable. The profile (D, R) is self-confirming too; however
that is not the only self-confirming profile. The profile (U, M ) is also
self-confirming. In this SCE, column has correct beliefs, but row be-
lieves that column will play L. This mistaken belief isn’t refuted by
the evidence because all row observes is that column does not play R.
There are also SCE where row mixes, and where both players mix.
Note that in the (U, M ) SCE, row’s beliefs do not respect column’s ra-
tionality. This suggests that one might refine SCE by further restrict-
ing beliefs to respect rationality or common knowledge of rationality
– Dekel, Fudenberg and Levine (1998) and Esponda (2006) explore
this possibility.

16
 The above example is somewhat contrived, but the idea that players
might get only partial feedback, and this might affect the outcome of learn-
ing, is natural. For instance, in sealed-bid auctions it is relatively common
to announce only the winner and possibility not even the winning price, so
the information available to form beliefs is rather limited.

4.2 Formal Definition of SCE

To define self-confirming equilibrium in extensive form games, let si denote
a strategy for player i, and σ i a mixture over such strategies. Let Hi denote
the set of information sets at which i moves, and H(si , σ −i ) denote the set of
information sets that can be reached if player i plays si and opponents play
σ −i . Let π i (hi |σ i ) denote the mixture over actions that results at information
set hi , if player i is using the strategy σ i (i.e. π i is the behavior strategy
induced by the mixed strategy σ i ). Let µi denote a belief over Π−i =
×j6=i Πj the product set of other players’ behavior strategies.

Definition 11 A profile σ is a Nash equilibrium if for each si ∈support(σ i ),

there exists a belief µi such that (i) si maximizes i’s expected payoff given
beliefs µi , and (ii) player i’s beliefs are correct, for all hj ∈ H−i

µi [{π −i | π j (hj ) = π j (hj |σ j )}] = 1.

The way to read this is that for each information set at which some player
j 6= i moves, player i’s belief puts a point mass on the probability distrib-
ution that exactly coincides with the distribution induced by j’s strategy.
Thus i has correct beliefs at all opponent information sets.

Definition 12 A profile σ is a Self-Confirming equilibrium if for each si ∈support(σ i ),

there exists a belief µi such that (i) si maximizes i’s expected payoffs given
beliefs µi , and (ii) player i’s beliefs are empirically correct, for all histories
hj ∈ H(si , σ −i ) and all j 6= i

µi [{π −i | π j (hj ) = π j (hj |σ j )}] = 1.

The difference is that in self-confirming equilibrium, beliefs must be cor-

rect only for reachable histories. This definition assumes that the players
observe their opponents’ actions perfectly, in contrast to the last example
above; it’s not hard to generalize the definition. Note that this definition for-
mally encompasses games of incomplete information (where Nature moves
first); Dekel et. al (2004) study SCE in these games.

17
References
[1] Aumann, R. (1974) “Subjectivity and Correlation in Randomized
Strategies,” J. Math. Econ.

[2] Aumann, R. (1987) “Correlated Equilibrium as an Extension of

Bayesian Rationality,” Econometrica, 55, 1—18.

[3] Aumann, R. and S. Hart (2003) “Long Cheap Talk,” Econometrica.

[4] Barany, I. (1992) “Fair Distribution Protocols or How Players Replace

Fortunes,” Math. Op. Res., 327—340.

[5] Bernheim, D. (1984) “Rationalizable Strategic Behavior,” Economet-

rica, 52, 1007—1028.

[6] Brandenburger, A. and E. Dekel (1987) “Rationalizability and Corre-

lated Equilibrium,” Econometrica, 55, 1391—1402

[7] Dekel, E., D. Fudenberg and D. Levine (2004) “Learning to Play

Bayesian Games,” Games and Econ. Behav., 46, 282-303.

[8] Fudenberg, D. and David Kreps (1993) “Learning Mixed Equilibria,”

Games and Econ. Behav., 5, 320—367.

[9] Fudenberg, D. and D. Levine (1993) “Self-Confirming Equilibrium,”

Econometrica, 61, 523- 546.

[10] Gerardi, D. (2001) “Unmediated Communication in Games with Com-

plete and Incomplete Communication,” J. Econ. Theory.

[11] Harsanyi, J. (1973) “On Games with Randomly Disturbed Payoffs: A

New Rationale for Mixed Strategy Equilibrium Points,” Int. J. Game
Theory, 2, 1—23.

[12] Hart, S. and D. Schmeidler (1989) “Existence of Correlated Equilibria,”

Math. Op. Res. 14, 18—25.

[13] Holt, C. and A. Roth (2004) “The Nash Equilibrium: A Perspective,”

PNAS, 101(12), 3999—4002.

[14] Pearce, D. (1984) “Rationalizable Strategic Behavior and the Problem

of Perfection,” Econometrica, 52, 1029—1050.

18