AGENDA

Static Games of Incomplete Information

Dynamic Games of Incomplete Information

Key Concepts to Remember

-2-
 BAYESIAN GAMES

When agents interact strategically they often do not know everything about each
other, e.g., firms in a Cournot duopoly might not know each other’s cost functions

Approach: All players include beliefs about the other players’ “types” into their
reasoning, and maximize payoffs conditional on their beliefs about
• the other players’ types θ i ∈ Θi , AND
• the other players’ strategies σ i ∈ ∆ ( Ai )

We assume that all players’ beliefs are common knowledge, so that types are
distributed according to the known joint cdf F : Θ → [0,1] , where Θ = Θ1 × L× Θ n

Definition: A Bayesian game in normal form is the collection

ΓBN = {N , {∆( Ai )},{ui (⋅ ;θ i )}, Θ, F (⋅)}

where each player i’s payoff function ui : A × Θi → ℜ depends on his type θ i ∈ Θi

-3-
 BAYESIAN-NASH EQUILIBRIUM

Definition: For any normal-form Bayesian game ΓBN = {N ,{∆( Ai )},{ui (⋅)}, Θ, F (⋅)}
a (pure-strategy) Bayesian-Nash equilibrium (BNE) is a strategy profile
a ∗ (θ ) = (ai∗ (θ i ), a−∗i (θ −i )) , such that for every player i ∈ N :

ai∗ (θ i ) ∈ arg max ∫ ui (ai , a−∗i (θ −i );θ i ) f (θ −i | θ i ) dθ −i

ai ∈Ai
Θ −i

f (θ i , θ −i )
where f (θ −i | θ i ) = represents player i’s beliefs (in form of a pdf (1))
∫ f (θ ,θ
Θ −i
i −i )dθ −i

conditional on his own type θ i .

dF (θ i ,θ −i ) d n F (θ i , θ −i )
(1) Note that ∫
Θ −i
f (θ i ,θ −i )dθ −i =
dθ i
and f (θ i , θ −i ) =
dθ1...dθ n -4-
 PRISONER’S DILEMMA WITH INCOMPLETE INFORMATION

Consider a prisoner’s dilemma where player 2 can be of type θ 2 ∈ {L, H } = Θ 2

With probability µ ∈ (0,1) player 2 is of a “high” type H who loathes to “rat” (i.e.,
defect on his friend and doing so would cause him a psychological penalty
equivalent to 6 years in prison

Player 2 (Type L) Player 2 (Type H)

Cooperate Defect Cooperate Defect

Cooperate (-1,-1) (-10,0) Cooperate (-1,-1) (-10,-6)

Player 1
Defect (0,-10) (-5,-5) Defect (0,-10) (-5,-11)

A BNE of this game is given by the strategy profile a ∗ (θ ) = (a1∗ (θ1 ), a2∗ (θ 2 ))
D, if θ 2 = L,
where a1∗ (θ1 ) = D and a2∗ (θ 2 ) =
C, if θ 2 = H .
-5-
 COURNOT DUOPOLY WITH INCOMPLETE INFORMATION

Assume that there are two firms, 1 and 2. Each firm i knows its own (constant)
marginal costs ci , but does not know the cost of the other firm.
• For simplicity assume that beliefs about the other firm’s costs are
symmetrically i.i.d. distributed with pdf f on the compact support [0,1]
• The inverse market demand curve is given by p (Q) = a − Q , where
Q = q1 + q2 denotes the aggregate demand (where a > 1)
1
E[π i (qi , q−i ; ci ) | ci ] = ∫ (a − ci − q−i (c−i ) − qi )qi f (c−i )dc−i
0

Maximization with respect to qi yields firm i’s best response:

a − ci − q−ei 1
∗
q (q−i ; ci ) =
i where q = ∫ q−i (c−i ) f (c−i )dc−i
e
−i
2 +
0

Question: How can we determine a (symmetric) BNE?

-6-
 COURNOT DUOPOLY WITH INCOMPLETE INFORMATION
(cont’d)

In a symmetric BNE we have that Since a is assumed large enough (check!)

1
1
q−e∗i = qie∗ = ∫ q1∗ (qie∗ ; ci ) f (ci )dci = ∫ (a − ci − qie∗ ) f (ci )dci / 2 = (a − cie − qie∗ ) / 2
0
0
1

where cie = ∫ ci f (ci )dci = c e denotes firm i’s expected cost.

0

e∗
Hence, qi = (a − cie ) / 3, so that after substituting this into firm i’s best response
we obtain

∗ a − ci a − c e 2a − 3ci + c e
q (ci ) =
i − =
2 6 6

Remark: The equilibrium output increases with firm i’s expectation of the rival
firm’s costs. Hence, it might be in a firm’s best interest to “pretend” to have lower
costs in order to induce the competitor to produce a lower output. When/how could
this be possible?

-7-
 FIRST-PRICE AUCTIONS
Independent Private Values

Consider a single-object auction with n risk-neutral bidders. Bidder i ∈ N has a

private value vi for the object and his beliefs about the other bidders’ private
values v j , j ≠ i are given by the cdf F : V → [0,1] where V = [v, v ] and vi ∈V

• Assume that the bidders’ valuations are i.i.d.

Question: What is the symmetric BNE for the IPV first-price auction?

Idea:
• Assume that there is an increasing bidding function (bid as a function of
the private value)
• The lowest bidder bids his valuation
• An optimal bid balances off the probability of winning against the surplus
in the case of winning

-8-
 FIRST-PRICE AUCTIONS (cont’d)
Solution

Bidder i’s expected payoff as a function of his bid bi = β (vi ) can be written as

Πi (bi ; vi ) = (vi − bi ) Pr(v1 ,..., vn ≤ vi ) = (vi − bi ) F n −1 ( β −1 (bi ))

Maximization with respect to bi yields a first-order necessary optimality condition

f ( β −1 (bi )) n −2 −1
F n −1
( β (bi )) = (n − 1)(vi − β (vi ))
−1
F ( β (bi ))
β ' ( β (bi ))
−1

or equivalently (using β −1 (bi ) = vi )

f (vi )
β ′(vi ) = (n − 1)(vi − β (vi ))
F (vi )
for all vi ∈V , which is a differential equation in β (⋅) with boundary condition
β (v ) = v
-9-
 FIRST-PRICE AUCTIONS (cont’d)
Solution

This initial-value problem can be solved explicitly, and we thus obtain the
equilibrium bidding function

vi

β ∗
(v ) = v −
∫
v
F n−1 (v)dv
i i
F n −1 (vi )

Remark: The equilibrium bids are increasing in the number of bidders. Note also
that each bidder always bids below his or her true valuation (i.e., each bidder
engages in “bid shaving”)

- 10 -
 AGENDA

Static Games of Incomplete Information

Dynamic Games of Incomplete Information

Key Concepts to Remember

- 11 -
RELATION BETWEEN GAMES OF INCOMPLETE INFORMATION
AND GAMES OF IMPERFECT INFORMATION

Recall:
• A game is of perfect information, if each information set contains a single
node, otherwise it is of imperfect information.
• A game is of complete information, if all players know all relevant
information about each other, otherwise it is of incomplete information.

Theorem (Harsanyi, 1967): Any game of incomplete information can be rewritten as

a game of imperfect information.

Proof (Outline):
• Start with an arbitrary game of incomplete information
• Introduce a new player, called “Nature” (N0)
• Player N0 is the first to move, and its actions follow all other players’
beliefs: in fact, N0 randomizes over the player types in Θ
• Any move by nature corresponds to a particular type realization; however,
players cannot observe that move, and thus their respective information
sets contain all possible nodes that N0’s choice could lead to. Clearly this
is a game of imperfect information, equivalent to the game of incomplete
information we started with.
- 12 -
 EXAMPLE: PRISONER’S DILEMMA WITH INCOMPLETE
INFORMATION AS A GAME OF IMPERFECT INFORMATION

To transform the Prisoner’s Dilemma game with incomplete information discussed

earlier into a game of imperfect information, introduce nature (N0) as a player who
moves first.

Player 1 needs to have beliefs

about nature’s move

- 13 -
 EXAMPLE 1: BNE LACKS PREDICTIVE POWER

Question: Determine all (pure) BNE of this game. Answer:

L R

U 2,1 0,0

M 0,2 0,1

D 1,3 1,3

The game has two

pure-strategy BNE:

[U,L] and [D,R]

- 14 -
 EXAMPLE 1 (cont’d)

Comments on the BNEs

• Everything is fine with the BNE [U,L]
• However, the BNE [D,R] is sustained by the non-credible threat that player
2 would play R if player 1 chooses either U or M

To rule out non-credible threats in games of imperfect information, we cannot use

subgame perfection, since proper subgames have to start at a singleton
information set.(1)

The concept of Bayesian perfection strengthens the BNE by requiring that players
have beliefs about the probability that each particular node has been reached in
equilibrium. A belief is thereby a probability distribution over the set of nodes in a
given information set.(2)

(1) In example 1, there is only one subgame: the game itself.

(2) If the information set is a singleton, then put probability one on that node. - 15 -
 (WEAK) PERFECT BAYESIAN EQUILIBRIUM

Definition: A strategy profile constitutes a (weak) perfect Bayesian equilibrium

(weak PBE) of a game of incomplete information if the following three requirements
are satisfied:

R1: At each information set, the player with the move must have a belief (i.e., a
probability distribution) about which node in his information set has been reached.

R2: Given their beliefs, all players’ strategies (=complete contingent plans) must be
sequentially rational, i.e., the actions taken at all information sets by players with
the move must be optimal.

R3: On any equilibrium path (=information sets reached with positive probability in
a given equilibrium), beliefs must be determined by Bayes’ rule and the players’
equilibrium strategies.

- 16 -
 EXAMPLE 1 (cont’d)

Introduce beliefs for player 2:

Player 2’s expected payoff from playing R versus playing L:

p ⋅ 0 + (1 − p) ⋅1 = 1 − p < 2 − p = p ⋅1 + (1 − p) ⋅ 2
Sequential rationality implies that player 2 will not choose R at his information set.

- 17 -
 EXAMPLE 2: SEQUENTIAL RATIONALITY LACKS POWER “OFF
THE EQUILIBRIUM PATH”

Consider this “slight” modification of the last game: Determine BNE

L R

U 2,0 0,0

M 0,2 0,1

D 1,3 1,3

The game has two

pure-strategy BNE:

[U,L] and [D,R]

- 18 -
 EXAMPLE 2 (cont’d)

Player 2’s expected payoff from playing R versus playing L is now:

p ⋅ 0 + (1 − p) ⋅1 = 1 − p ≤ 2 − 2 p = p ⋅ 0 + (1 − p) ⋅ 2
If p = 1, then it is a best response for player 2 to play R, i.e., the BNE [D,R] cannot
be ruled out.

In general, weakly dominated strategies cannot be ruled out merely by imposing

sequential rationality.

We therefore need to impose that the beliefs are reasonable.

Definition: An information set is on the equilibrium path, if it will be reached with

strictly positive probability if players play an equilibrium strategy. Otherwise the
information set is off the equilibrium path.

Note that in example 2, Bayes’ rule on the equilibrium path does not impose any
restriction on p following player 1’s strategy D. Thus, the weak PBE is not strong
enough and needs refinement by requiring that Bayes’ rule be used off the
equilibrium path whenever possible.(1)

- 19 - (1) If
 PERFECT BAYESIAN EQUILIBRIUM

Definition: A strategy profile constitutes a (strong) perfect Bayesian equilibrium

(weak PBE) of a game of incomplete information if the following four requirements
are satisfied:

R1: At each information set, the player with the move must have a belief (i.e., a
probability distribution) about which node in his information set has been reached.

R2: Given their beliefs, all players’ strategies (=complete contingent plans) must be
sequentially rational, i.e., the actions taken at all information sets by players with
the move must be optimal.

R3: On any equilibrium path (=information sets reached with positive probability in
a given equilibrium), beliefs must be determined by Bayes’ rule and the players’
equilibrium strategies.

R4: Off any equilibrium path, beliefs are determined by Bayes’ rule and the players’
equilibrium strategies, where possible.(1)

(1) R4 is equivalent to requiring that any strong PBE must be a weak PBE of every proper subgame,
i.e., can use backward induction. - 20 -
 EXAMPLE 3: ILLUSTRATION OF (STRONG) PBE CONCEPT

Consider the following three-player game:

- 21 -
 EXAMPLE 3 (cont’d)

The normal-form representation of this game is:

L R

U D U D

A (2,0,0) (2,0,0) A (2,0,0) (2,0,0)

B (1,2,1) (0,1,2) B (3,3,3) (0,1,1)

The pure-strategy BNE of this game are thus: [A,U,L], [A,D,L], [B,U,R], and [A,D,R]

- 22 -
 EXAMPLE 3 (cont’d)

Consider the following subgame

following a choice of B by player 1:

Player 3

L R

U (2,1) (3,3)
Player 2
D (1,2) (1,1)

The only NE of the equivalent normal-form representation is [A,U,R].

Hence, [B,U,R; p=1] is a strong PBE (note: there are no information sets off the
equilibrium path)

- 23 -
 EXAMPLE 3 (cont’d)
1. Consider the BNE [A,U,L]

• Note that player 3’s information set is

off the equilibrium path!
• [A,U,L;p=0] satisfies R1 through R3 and is
thus a weak PBE.
• However, p=0 is inconsistent with player 2’s
strategy U, and thus R4 is violated!

Thus [A,U,L;p=0] is not a (strong) PBE (the

only belief consistent with U is p=1).(1)

2. Consider the BNE [A,D,L]

• p=0 is only belief consistent with Bayes’
rule, and L is player 3’s best response
• Player 2 anticipating this, would thus not choose D
but rather U (i.e., [A,D,L; p=0] is a weak PBE but not a strong PBE)

3. Consider the BNE [A,D,R]

• p=0 is only belief consistent with Bayes’ rule, but then L is player 3’s best response
and not R (i.e., [A,D,R; p] with p<1 is a weak PBE but not a strong PBE)

M: (1) Note that [A,U,L,p] is a weak PBE for any p ≤ 1 / 3 . - 24 -

 SIGNALING GAMES

In a signaling game, one player (sender) has private information (corresponding to

his type θ ∈ Θ ) and tries to communicate this information to the other player
(receiver)

• First, the sender (S) sends a message m∈M

• Second, the receiver (R) takes an action a ∈ A
• Finally, the players’ payoffs uS (m, a;θ ) and u R (a, m) are realized

Definition: Consider a (strong) PBE of a signaling game.

• If R is able to distinguish between all the different sender types the
equilibrium is called separating;
• If R is not able to distinguish between any types, then the equilibrium is
called pooling;
• Otherwise, the equilibrium is called partially pooling or semi-separating.

- 25 -
 STRUCTURE OF A SIMPLE SIGNALING GAME

Consider a signaling game with Θ = {θ1 , θ 2 }, M = {m1 , m2 }, A = {a1 , a2 }

- 26 -
 (STRONG) PBEs IN SIGNALING GAMES

From the definition of a strong PBE in games of incomplete information we can

derive the following three requirements for (strong) PBEs in the simple signaling
game:

R1S: After observing any message m ∈ M must have a belief p(θ | m) about which
types could have sent m .

R2S: For each possible message m ∈ M , the receiver’s action a(m) must
maximize his payoffs, given his beliefs about the types who could have sent m.

R3S: For any possible type θ ∈ Θ , a sender’s message m(θ ) must maximize her
payoffs, given the receiver’s strategy a(⋅).

( )
The tuple m(⋅), a(⋅), p(⋅ | ⋅) is a PBE of the simple signaling game, if its elements
satisfy requirements R1S, R2S, and R3S.

- 27 -
EXAMPLE 4: A SIMPLE SIGNALING GAME

- 28 -
 EXAMPLE 4 (cont’d)

I. Introduce R’s beliefs

[ p] [q]

[1 − p] [1 − q]

- 29 -
 EXAMPLE 4 (cont’d)

II. Check for Pooling Equilibria

• Pooling on m = L: This implies that p = .5. The belief q lies off the
equilibrium path.
- Receiver always plays a = U as a response tom=L
- When would Receiver play a = D as a response to m = R ?
Answer: Comparing payoffs we find the condition q ≤ 2(1 − q)
or equivalently: q ≤ 2 / 3

- Thus, [( L, L), (U , D), p = .5, q ≤ 2 / 3] is a pooling equilibrium

• Pooling on m = R: This implies that q = .5. The belief p lies off the
equilibrium path.
- Receiver always plays a = U as a response to m=L
- Receiver best response to m = R is a = D
- But this implies that type θ1 would not like to play m = R
- Hence, there is no pooling equilibrium on m=R

- 30 -
 EXAMPLE 4 (cont’d)

III. Check for Separating Equilibria

• Separating (L,R): This implies that p = 1 and q = 0.

- Type θ 2 would not like to follow this strategy
- Hence, there is no separating (L,R) equilibrium

• Separating (R,L): This implies that p = 0 and q =1.

- This is compatible with both types’ best responses
- Hence, [( R, L), (U ,U ), p = 0, q = 1] is a separating equilibrium

- 31 -
 AGENDA

Static Games of Incomplete Information

Dynamic Games of Incomplete Information

Key Concepts to Remember

- -
 KEY CONCEPTS TO REMEMBER

• Bayes-Nash Equilibrium
• Reasonable Beliefs
• Bayes’ Rule
• Weak/Strong Perfect Bayesian Equilibrium
• Signaling; Spence Condition
• Separating/Pooling Equilibria