Microeconomics II

Lecture 5: Dynamic Games of Incomplete

Information

Mohammad Vesal
Graduate School of Management and Economics
Sharif University of Technology

44706
Spring 2025

1 / 53
Dynamic games

• So far we have discussed simultaneous-move (static) games

of complete and incomplete information.
• In many cases, however, players have asymmetric
information.
• But one agent can decide to “signal” his type to the other.
Driver knows more about driving behavior than insurance
company → install monitoring device
Worker better knows her ability than the employer →
obtain degrees
Defendant knows whether he has committed a crime while
the jury does not → hire expensive lawyer
• Is there a way to credibly signal one’s type?

2 / 53
Outline

Introduction

Basic signaling game

Reputation game

Spence’s job market signaling

Reference: Fudenberg and Tirole Ch 8 (sec.1 & 2), MWG Ch 9.C (partly),
Ch 13.C (partly)

3 / 53
Outline

Introduction

Basic signaling game

Reputation game

Spence’s job market signaling

4 / 53
Basic signaling game

• Player 1 (sender)
observes θ ∈ Θ
selects a1 ∈ A1
• Player 2 (receiver)
observes a1
selects a2 ∈ A2
• Payoffs
u1 (a1 , a2 , θ) u2 (a1 , a2 , θ)
• Common knowledge about prior distribution of θ
• Upon observing action a1 , receiver asks, what type of
player 1 is more likely to take this course of action.
Update his belief about player 1’s type

5 / 53
Strategies and beliefs
• Mixed strategy for sender

σ1 (a1 | θ)

with what probability does player 1 play each of his actions

given his observed type?
• Beliefs for receiver

µ(θ | a1 ) = Pr (θ | a1 )

given the observed action of a1 what is the probability that

player 1 is of type θ?
• Mixed strategy for receiver

σ2 (a2 | a1 )

with what probability does player 2 play each of his action

given the observed action of player 1?
6 / 53
Perfect Bayesian Equilibrium (PBE) of a signaling game
A strategy profile (σ1∗ , σ2∗ ) and a belief system µ is a (weak)
PBE of the signaling game if
1. ∀θ, ∀a∗1 such that σ1∗ (a∗1 | θ) > 0
 
 X 
a∗1 ∈ arg max σ2∗ (a2 | a1 )u1 (a1 , a2 , θ)
a1  
a2 ∈A2

2. ∀a1 , ∀a∗2 (a1 ) such that σ2∗ (a∗2 (a1 ) | a1 ) > 0

( )
X
a∗2 (a1 ) ∈ arg max µ(θ | a1 )u2 (a1 , a2 , θ)
a2
θ∈Θ

3. ∀a∗1 such that 0 )σ ∗ (a∗ | θ0 ) > 0

P
θ0 ∈Θ p(θ 1 1

p(θ)σ1∗ (a∗1 | θ)
µ(θ | a∗1 ) = P 0 ∗ ∗ 0
θ0 ∈Θ p(θ )σ1 (a1 | θ )

7 / 53
PBE discussion

• Conditions 1 and 2 are the perfection conditions (sequential

rationality).
Given his type, player 1 is choosing a best response to the
equilibrium strategy played by player 2.
Given the observed action of player 1 and his updated
belief, player 2 is choosing a best response.
• Condition 3 is a consistency requirement for beliefs of
player 2.
Given the observed action of player 1 and his equilibrium
strategy, what is the probability that player 1 is of type θ.
• The consistency requirement leaves out-of-equilibrium
beliefs arbitrary

8 / 53
Perfect Bayesian Equilibrium: more general

• Beliefs: x∈H µ(x) = 1

P
∀H
• A strategy profile σ and a belief system µ constitute a
(weak) PBE if
Players’ strategies are optimal: σ is sequentially rational at
all H given µ
The belief system is consistent: µ is derived from σ using
Bayes’s rule; i.e. ∀H if Pr (H | σ) > 0 then

Pr (x | σ)
µ(x) = ∀x ∈ H
Pr (H | σ)
• Note: the consistency requirement only restricts beliefs on
the equilibrium path.
Players could have arbitrary beliefs on nodes in information
sets that are not reached on the equilibrium path.

9 / 53
Two examples of signaling games

• Reputation games
P1: Incumbent chooses fight/accommodate
P2: Entrant chooses to stay/exit
• Spence’s (1973) job market’s signaling
P1: worker observes productivity and chooses education
P2: employer observes education but not productivity and
chooses wage offer

10 / 53
Taxonomy of PBEs

• Separating equilibrium
understands P1’s type from observing P1’s action.
• Semi-separating equilibrium
Only sometimes P1’s action reveals his true type.
• Pooling equilibrium
All P1 types play the same strategy, no information is
revealed.

11 / 53
Outline

Introduction

Basic signaling game

Reputation game
Separating equilibrium
Semi-Separating equilibrium
Pooling equilibrium

Spence’s job market signaling

12 / 53
Reputation game
Two firms: Incumbent and entrant
1. Incumbent observes θ ∈{sane,crazy}
Chooses a1 ∈{Fight,Accommodate}
prior: Pr(sane)= p
2. Entrant observes a1
Chooses a2 ∈{Stay,Exit}
Ent. Ent.
sane S E crazy S E
F (P1 , P2 ) (M1 , 0) F (K1 , K2 ) (M1 , 0)
Inc. Inc.
A (D1 , D2 ) (M1 , 0) A (D1 , D2 ) (M1 , 0)

P1 < D1 < M1 < K1

K2 < 0 < P2 < D2

13 / 53
Reduced form of a entry deterrence game

• Incumbent monopolist faces an entrant.

• Some crazy types love to fight but the sane majority would
rather accommodate.
• Challenger enters the market not knowing the monopolist
type.
• Incumbent engages in price war or accommodates
some sane monopolists like to pretend they are crazy!
• Entrant can stay or exit the market
• If he stays the sane one will stop fighting!
• Examples: Kreps and Wilson (1982), Milgrom and Roberts
(1982)

14 / 53
What are the PBEs of this game?

• Fight is a dominant strategy for the crazy incumbent

a∗1 (crazy) = F

• What is the equilibrium strategy of the sane type?

• Let
α = Pr (a∗1 (sane) = F)

15 / 53
Taxonomy of equilibria

• Separating: actions reveal types α = 0

sane types always accommodate
• Semi-separating: α ∈ (0, 1)
sane types sometimes fight
observing accommodate reveals type
observing fight does not!
• Pooling: actions are uninformative α = 1
sane types always fight

16 / 53
Separating equilibrium

• Proposition: There exists no separating equilibria!

• Proof by contradiction:
1. Assume a separating equilibrium exists: α = 0
2. Determine entrant’s belief and optimal strategy
3. Determine incumbent’s optimal strategy
4. Show sane incumbent’s optimal strategy is not
accommodate!

17 / 53
Entrant’s belief when α = 0

• ∀a∗1 such that 0 )σ ∗ (a∗ | θ0 ) > 0

P
θ0 ∈Θ p(θ 1 1

p(θ)σ1∗ (a∗1 | θ)
µ(θ | a∗1 ) = P 0 ∗ ∗ 0
θ0 ∈Θ p(θ )σ1 (a1 | θ )

• Sane types accommodate, crazy ones fight!

µ(θ = sane | a1 = A) = 1, µ(θ = crazy | a1 = A)

µ(θ = sane | a1 = F) = 0, µ(θ = crazy | a1 = F )

• Are there any out-of-equilibrium beliefs?

18 / 53
Entrant’s optimal strategy
• ∀a1 , ∀a∗2 (a1 ) such that σ2∗ (a∗2 (a1 ) | a1 ) > 0
( )
X
∗
a2 (a1 ) ∈ arg max µ(θ | a1 )u2 (a1 , a2 , θ)
a2
θ∈Θ
• If a1 = F, entrant’s expected payoff is
X
µ(θ | a1 = F)u2 (a1 = F, a2 , θ) = u2 (F, a2 , crazy)
θ∈Θ
• Entrant’s optimal strategy is exit because
u2 (F, a2 = E, crazy) > u2 (F, a2 = S, crazy)
0 > K2
• If a1 = A, entrant’s optimal strategy is stay because
u2 (A, a2 = E, sane) < u2 (A, a2 = S, sane)
0 < D2

19 / 53
Incumbent’s optimal strategy
• Entrant’s optimal strategy was to stay if a1 = A and to exit
if a1 = F
• Now incumbent’s optimal strategy is
∀θ, ∀a∗1 such that σ1∗ (a∗1 | θ) > 0
 
 X 
a∗1 ∈ arg max σ2∗ (a2 | a1 )u1 (a1 , a2 , θ)
a1  
a2 ∈A2

• Incumbent’s expected payoff is

(
u1 (A, a2 = S, θ) if a1 = A
Eu1 =
u1 (F, a2 = E, θ) if a1 = F
• Sane type will want to Fight because
u1 (A, a2 = S, sane) < u1 (F, a2 = E, sane)
D1 < M 1
• This contradiction proves the result!
20 / 53
Discussion

• If the entrant believes that anyone who fights is a crazy

type
• the sane types would want to fight
to signal they are the crazy guy!
• But believing that anyone who fights is crazy is not
consistent because
some sane types have an incentive to fight!
• ⇒no separating equilibrium

21 / 53
Semi-separating equilibrium

• Proposition: If pP2 > −(1 − p)K2 then there exists a

unique semi-separating equilibrium in which
the sane incumbent fights with probability α = − (1−p)K
pP2
2

the entrant stays if accommodation is observed

1 −D1
the entrant exits with probability M
M1 −P1 if fight is observed
• Proof: follow the same procedure as in the separating
equilibrium. Instead of contradiction we find an
equilibrium!
Notice in any semi-separating equilibrium
α = Pr (a∗1 (sane) = F) ∈ (0, 1)

22 / 53
Entrant’s belief given α ∈ (0, 1)

• Sane types accommodate, crazy ones fight!

µ(θ = sane | a1 = A) = 1
pα
µ(θ = sane | a1 = F) =
pα + 1 − p
• Are there any out-of-equilibrium beliefs?

23 / 53
Entrant’s optimal strategy
• Entrant’s expected payoff is
X
µ(θ | a1 )u2 (a1 , a2 , θ) =
θ∈Θ
µ(sane | a1 )u2 (a1 , a2 , sane) + (1 − µ(sane | a1 ))u2 (a1 , a2 , crazy)
• If a1 = A, entrant’s optimal strategy is stay because 0 < D2
• If a1 = F and entrant stays
pα pα
P2 + (1 − )K2
pα + 1 − p pα + 1 − p
If he exits he gets zero
• Stay is optimal iff
pα pα
P2 + (1 − )K2 ≥ 0
pα + 1 − p pα + 1 − p
⇔ pαP2 ≥ −(1 − p)K2

24 / 53
Entrant’s optimal strategy

• Summarize the results so far

(1 − p)K2
exit if α<−
pP2
(1 − p)K2
indifferent if α=−
pP2
(1 − p)K2
stay if α>−
pP2
• The RHS is a valid probability as far as the assumption in
the theorem holds
(1 − p)K2
− ∈ (0, 1)
pP2
What does this assumption say?

25 / 53
Incumbent’s optimal strategy

• To get α ∈ (0, 1) the sane incumbent’s expected payoff

needs to be the same for F and A
• But if entrant plays a pure strategy they cannot be
equalized!
Entrant must also mix
Let β = Pr(a2 (F) = S)
• Incumbent’s expected payoff for a1 = A is D1 since the
challenger always stays.
• Incumbent’s expected payoff for a1 = F is

βP1 + (1 − β)M1

• If β = M 1 −D1
M1 −P1 then the sane incumbent will be indifferent
and we can assume he mixes with the α we derived above!

26 / 53
Pooling equilibrium

• Proposition: There exists no pooling equilibrium if

pP2 > −(1 − p)K2 .
• Proof by contradiction:
Assume existence: α = Pr(a∗1 (sane) = F) = 1
Determine entrants beliefs and optimal strategy
Show the sane incumbent’s optimal strategy is
accommodate.

27 / 53
Entrant’s beliefs given α = 1

• Entrant’s beliefs about the type of the incumbent are

µ(θ = sane | a∗1 = F) = p

why?
• Are there any out-of-equilibrium beliefs?
Let γ = µ(θ = sane | a∗1 = A) ∈ [0, 1]
accommodate is played with zero probability, entrant could
have any beliefs!
• We have to examine all possible γ to rule out existence of
pooling equilibrium.

28 / 53
Entrant’s optimal strategy

• If a1 =A, entrant’s payoff is D2 regardless of incumbent’s

type
prefers to stay as D2 > 0
Independent of out-of-equilibrium beliefs!
• If a1 =F, the entrant’s payoff is
pP2 + (1 − p)K2 if stay
0 if exit
entrant stays because pP2 + (1 − p)K2 > 0
• Entrant would rather stay always!

29 / 53
Incumbent’s optimal strategy

• A sane incumbent’s best response knowing the entrant

stays is to accommodate because

P1 < D1

• This is a contradiction and proves the result.

30 / 53
Discussion

• If all incumbent types fight then fight is not a credible

threat!
Entrant knows that in majority of cases the incumbent is
sane and is pretending to be crazy!
Given this entrant’s best response is to stay.
• Example was a bit simple since incumbent’s best response
was independent of out-of-equilibrium beliefs.

31 / 53
Summary of reputation game equilibria

• If pP2 > −(1 − p)K2 the reputation game has a unique

semi-separating perfect Bayesian equilibrium.
• In this equilibrium the crazy incumbent always fights and
the sane incumbent fights with probability α = − (1−p)K
pP2 .
2

• In this equilibrium the entrant stays for sure if

accommodation is played and stays with probability
1 −D1
β=M M1 −P1 if fight is played.
• What happens if pP2 = −(1 − p)K2 ?

32 / 53
Outline

Introduction

Basic signaling game

Reputation game

Spence’s job market signaling

Separating equilibria
Pooling equilibria
Intuitive criterion

33 / 53
Job market signaling

• Players: a worker and a large number of firms

1. Worker observes type θ ∈ {θL , θH }

θL < θH and Pr(θ = θH ) = γ
2. Worker chooses level of education a ∈ [0, ∞)
cost of education: a/θ
3. Firms observe a and offer w
4. Worker selects one of the firms and produces value θ

• Payoffs
a
uw = w − uf = θ − w
θ

34 / 53
Simplify the game

• Assume symmetric beliefs for firms

δ(a) = µ(θH | a)

• Competition between firms ensures that the worker gets all

of the expected surplus

w = µ(θH | a)θH + µ(θL | a)θL

= θL + δ(a) (θH − θL )

35 / 53
Separating equilibria

• Suppose high and low types choose different education

levels
a(θH ) 6= a(θL )
• On-the-equilibrium path belief

δ(a(θL )) = 0 δ(a(θH )) = 1

• What is the belief off-the-equilibrium path?

δ(a) ∈ [0, 1] ∀a ∈
/ {a(θL ), a(θH )}

• Could there be an equilibrium like this?

36 / 53
Low type’s incentive compatibility
• Low type should not have an incentive to deviate

a(θL ) a
θL − ≥ θL + δ(a) (θH − θL ) − ∀a
θL θL
• This implies a(θL ) = 0 [prove by contradiction]
• Therefore
a
≥ δ(a) ∀a (1)
θL (θH − θL )
• Low type should not have an incentive to pretend as high
type

a(θH )
θL ≥ θL + (θH − θL ) −
θL
⇔ a(θH ) ≥ θL (θH − θL ) (2)

37 / 53
High type’s incentive compatibility

• High type should not have an incentive to deviate

a(θH ) a
θH − ≥ θL + δ(a) (θH − θL ) − ∀a
θH θH
a(θH ) − a
⇔1− ≥ δ(a) ∀a (3)
θH (θH − θL )
• High type should not want to pretend as low type

a(θH )
θH − ≥ θL
θH
⇔ θH (θH − θL ) ≥ a(θH ) (4)

38 / 53
Results so far

• Low type gets no education a(θL ) = 0

• Equations (1) and (3) imply
 
a(θH ) − a a
δ(a) ≤ min 1 − , ∀a (5)
θH (θH − θL ) θL (θH − θL )

this is a condition on beliefs.

• Equations (2) and (4) imply

θH (θH − θL ) ≥ a(θH ) ≥ θL (θH − θL ) (6)

this is a condition on level of education of the high type.

• The three conditions must be satisfied in any separating
equilibrium.

39 / 53
Satisfying the belief condition

• Discontinuous beliefs
(
0 if a < a(θH )
δ(a) =
1 if a ≥ a(θH )

will satisfy the belief condition (5) (when (6) is satisfied)

 
a(θH ) − a a
δ(a) ≤ min 1 − , ∀a
θH (θH − θL ) θL (θH − θL )

why?

40 / 53
Numerical example: Separating equilibria
• Let θL = 1, θH = 2, the belief and education conditions are
 
a(θH ) − a
δ(a) ≤ min 1 − ,a ∀a
2
2 ≥ a(θH ) ≥ 1
• For every â ∈ [1, 2] there exists a separating equilibrium in
which
a(θL ) = 0 a(θH ) = â
(
0 if a < â
δ(a) =
1 if a ≥ â
• There are multiple separating equilibria.
In all of them the firms have consistent beliefs
on-the-equilibrium path and specified beliefs
off-the-equilibrium.
Education level of workers are such that they do not have
an incentive to deviate to any level of education given firms
beliefs.
41 / 53
Pooling equilibria

• Suppose a(θL ) = a(θH ) = a

Both types get the same wage

w = θL + γ (θH − θL )

• What are the beliefs?

• On-the-equilibrium: δ(a) = γ (prior belief)
• Off-the-equilibrium: δ(a) ∈ [0, 1] ∀a 6= a
• Now we need to check the incentive compatibility of the
two types.

42 / 53
Low type’s incentive compatibility

• No incentive to deviate
a a
θL + γ (θH − θL ) − ≥ θL + δ(a) (θH − θL ) − ∀a
θL θL
• No incentive to deviate to zero education
a
γ (θH − θL ) − ≥ δ(0) (θH − θL )
θL
⇒ a ≤ (γ − δ(0)) θL (θH − θL )

43 / 53
High type’s incentive compatibility

• No incentive to deviate
a a
θL + γ (θH − θL ) − ≥ θL + δ(a) (θH − θL ) − ∀a
θH θH
• No incentive to deviate to zero education
a
γ (θH − θL ) − ≥ δ(0) (θH − θL )
θH
⇒ a ≤ (γ − δ(0)) θH (θH − θL )

44 / 53
Summary of conditions for pooling equilibria
• The two IC constraints for deviation to zero give

a ≤ (γ − δ(0)) (θH − θL ) · min {θH , θL } (7)

= (γ − δ(0)) (θH − θL ) θL

(7): equil. education is such that θL does not want to

deviate to zero education.
We assumed δ(0) ≤ γ, why?
• IC constraints for deviation to any other level of education
give  
1 a−a a−a
δ(a) ≤ γ + min , (8)
θH − θL θL θH
(8): given equil. education level, beliefs are such that
θH does not have an incentive to get higher levels of
education
θL does not have an incentive to get lower levels of
education
45 / 53
Pooling equilibria
• If we take a discontinuous belief system as
(
0 if a < a
δ(a) =
γ if a ≥ a

together with
a ≤ γ (θH − θL ) θL
• You can confirm that all the conditions for a pooling
equilibrium are satisfied.
• Notice the beliefs specified above is not the only possible
beliefs system that could support the pooling equilibria.
Could you give another beliefs system that supports the
same equilibria?
• But in any pooling equilibria a must satisfy the condition
above. Why?

46 / 53
Multiplicity of equilibria

• In the Spence’s signaling game we have many equilibria

both pooling and separating!
• Is there a way to select from these?
• Intuitive Criterion: Cho and Kreps (1987)
Some out-of-equilibrium beliefs are unreasonable.

47 / 53
Intuitive Criterion

• Take a PBE and consider out-of-equilibrium action â, If

1. Type θ1 gets strictly higher payoff by deviating to â if the
receiver believes that µ(θ = θ1 | â) = 1
2. Type θ2 does not get strictly higher utility by deviating to â
for any belief that receiver might have µ(θ = θ1 | â)
• Then we say the PBE fails the intuitive criterion!
• Intuition: High type tells the receiver I will make a
deviation that cannot be possibly profitable for the low
type. Therefore when I deviate you must believe I am the
high type!

48 / 53
Applying the intuitive criterion

• Which of the many equilibria derived for the Job Signaling

game survives the intuitive criterion?
• Proposition: The only PBE that does not fail the intuitive
criterion is the separating equilibrium with
a(θH ) = θL (θH − θL ).
• Proof: take any other PBE and show there is a deviation
for one that could not be profitable by the other type.

49 / 53
Separating equilibria
• Take a separating equilibrium with

θH (θH − θL ) ≥ a(θH ) > θL (θH − θL )

Consider a deviation to â = θL (θH − θL ):

• θH gets strictly higher utility if employer has δ(â) = 1
because he incurs lower education costs and receives the
same wage as in the separating equilibrium.
• θL does not get strictly higher utility by deviating to â for
any belief δ̂(â) ∈ [0, 1] because

â
θL + δ̂(â) (θH − θL ) − ≤ θL
θL
⇔ δ̂(â) ≤ 1

50 / 53
Pooling equilibria
• Take a pooling equilibrium with a ≤ γ(θH − θL )θL and
consider a deviation to
â = a + (1 − γ) (θH − θL ) θL
• θH gets strictly higher utility if employer’s belief is δ(â) = 1
because
a â
θL + γ (θH − θL ) − < θH −
θH θH
θL
⇔ (1 − γ) (θH − θL ) > (1 − γ) (θH − θL )
θH
⇔ θH > θ L
• θL does not get strictly higher utility by deviating to â for
any belief δ̂(â) ∈ [0, 1] because
â a
θL + δ̂(â) (θH − θL ) − ≤ θL + γ (θH − θL ) −
θL θL
⇔ δ̂(â) ≤ 1
51 / 53
Two stories about education

• Increases productivity: returns to education are positive

• Works as a screening device: only high ability guys invest
in education
• What are the policy implications of each of these views?
• Could we empirically separate the two?
See Weiss (1995), Altonji (1995), Altonji and Pierret (2001)
Idea: see if wage of more experienced workers depends less
on education and more on difficult to observe ability proxies!

52 / 53
Summary

• In this lecture, we studied dynamic games of incomplete

information.
Perfect Bayesian Equilibrium
Taxonomy of equilibria in signaling games
Two examples
Intuitive criterion

53 / 53